Datasets:

timaeus
/

pile-arxiv

Dataset card Data Studio Files Files and versions Community

Split (1)

train · 100k rows

text stringlengths 4 2.78M	meta dict
--- bibliography: - 'biblioCH.bib' --- The Consistent Histories Formalism and the Measurement Problem\ Elias Okon\ Instituto de Investigaciones Filosóficas, Universidad Nacional Autónoma de México, Mexico City, Mexico.\ E-mail: `[email protected]`\ Daniel Sudarsky\ Instituto de Ciencias Nucleares, Universidad Nacional Autónoma de México, Mexico City, Mexico.\ E-mail: `[email protected]`\ Abstract: In response to a recent rebuttal of [@Oko.Sud:14b] presented in [@Gri:15], we defend the claim that the Consistent Histories formulation of quantum mechanics does not solve the measurement problem. In order to do so, we argue that satisfactory solutions to the problem must not only not contain anthropomorphic terms (such as measurement or observer) at the fundamental level, but also that applications of the formalism to concrete situations (e.g., measurements) should not require any input not contained in the description of the situation at hand at the fundamental level. Our assertion is that the Consistent Histories formalism does not meet the second criterion. We also argue that the so-called second measurement problem, i.e., the inability to explain how an experimental result is related to a property possessed by the measured system before the measurement took place, is only a pseudo-problem. As a result, we reject the claim, defended in [@Gri:15], that the capacity of the Consistent Histories formalism to solve it should count as an advantage over other interpretations. Introduction {#Int} ============ The Consistent Histories (CH) framework provides a formulation of quantum mechanics that assigns probabilities to histories of all kinds of systems, microscopic and macroscopic, using a single universal machinery and without “Heisenberg cuts” or references to measurements or observers. As a result, proponents of CH maintain that the formalism overcomes the measurement problem of the standard interpretation (as well as all other standard quantum paradoxes). In [@Oko.Sud:14b] we have disputed such an assertion by displaying an array of conceptual problems with the way the formalism is deployed in measurement situations.[^1] In [@Gri:15], however, arguments against our objections are presented, and so the main objective of this article is to respond to such a challenge. We hope that, by doing so, we will not only be able to adequately defend our position, but also to shed light on the root of the disagreement. In this regard, we believe that the origin of the dispute arises from a difference on what CH proponents and us take the measurement problem to be, and, more importantly, on what CH proponents and us regard as a satisfactory solution to the problem. In short, we believe that such a solution must not only avoid making any reference to measurements or cuts at the fundamental level, but also that successful applications of the formalism must not depend on input not present in the fundamental theory. Our claim, in a nutshell, is that CH accomplishes the first but not the second. In the rest of the paper we develop these ideas. To do so, we briefly review the CH formalism in section \[CH\] and in section \[MP\] we discuss what it takes to solve the measurement problem. Then, in section \[MCH\] we summarize our arguments in [@Oko.Sud:14b] and in section \[Rep\] we evaluate and respond to the challenges raised in [@Gri:15]. Finally, in section \[C\] we present our conclusions. A brief presentation of the Consistent Histories formalism {#CH} ========================================================== Before getting down to business, we will briefly describe the CH formalism (see [@Gri:03] for a comprehensive presentation). As we said above, CH assigns probabilities for all systems, microscopic or macroscopic, using the same machinery and without any reference to measurements or cuts. More specifically, the most general objective of CH is the prediction of probabilities for time histories of systems, where histories are defined as sequences of properties and are represented by projection operators at successive times. CH, then, introduces the notion of sets of histories and specifies rules that assign probabilities to the various elements of each set. However, according to CH, not all sets of histories allow for probabilities to be assigned. This is possible only when: i) the sum of probabilities of all members of a set equals one, and ii) all pairs of histories within the set are orthogonal. Families satisfying these two conditions are called frameworks, or realms. A natural consequence of the CH formalism is that, given a system, multiple incompatible frameworks can be constructed (i.e., different frameworks that assign incompatible properties to the system). Therefore, in order to avoid inconsistencies, CH requires the imposition of the following rules: - Single-Framework Rule: probabilistic reasoning is invalid unless it is carried out using a single framework. - Principle of Liberty: one can use whatever framework one chooses in order to describe a system. - Principle of Equality: all frameworks are equally acceptable in terms of fundamental quantum mechanics. - Principle of Utility: not all frameworks are equally useful in answering particular questions of physical interest. This, however, comes with a price because the enforcement of these rules leads to the violation of the following principle: - Principle of Unicity: alternative descriptions of physical systems always can be combined into a single unified one, from which all views can be derived as partial characterizations. Whether this is too high a price to pay is an interesting question. However, what we would like to point out for now is that, as we will see in section \[Rep\], and contrary to what is claimed in [@Gri:15], none of the objections that were presented in [@Oko.Sud:14b] are based on the fact that the Principle of Unicity is not valid within CH. Solving the measurement problem {#MP} =============================== A lot has been written about the measurement problem of quantum mechanics. A popular way to describe it, among many, is the following: even though quantum mechanics depends crucially on the notion of measurement, such notion is never formally defined within the theory. Then, in order to use quantum mechanics, one needs to know, by means external to the theory, what constitutes a measurement. Of course, the measurement problem is a problem of a theoretical framework and so, in order to state it, one needs to first specify in detail the theoretical framework in question.[^2] This, given the proliferation of views regarding quantum mechanics, leads to a proliferation of ways to state the problem. For example, the description given above is suitable for Dirac’s or von Neumann’s formulation but does not apply to Bohr’s, where the problem manifests as an ambiguity regarding where the classical-quantum cut should be drawn. It also does not apply to a formulation with a purely unitary evolution, where the problem manifests as a mismatch between experience and some predictions of the theory. At any rate, for the purposes of this paper it is sufficient to note that both us, and the author of [@Gri:15], agree on the fact that the standard or orthodox interpretation suffers from the measurement problem (see e.g. [@Gri:03 p. 214]). What we consider more important, given the objective of this work (i.e., evaluating whether CH solves or not the measurement problem), is a discussion of what constitutes a valid solution to the problem. In this regard, [@Gri:15] offers the following: > If quantum mechanics applies not only to the microscopic world of nuclei and atoms, but also to macroscopic objects and things that are even larger - from the quarks to the quasars - then the measurement process in which an earlier microscopic property is revealed in a macroscopic outcome should itself be describable, at least in principle, in fully quantum mechanical terms. Applied equally to the system being measured and to the macroscopic apparatus, and without the evasion and equivocation ridiculed by Bell [@Bel:90]. ([@Gri:15 p. 3]) Namely, if quantum mechanics applies to everything - from quarks to quasars - then measurements must be fully describable in purely quantum terms. Of course, one could hold that quantum mechanics does not apply to everything, but in such a case one would need to clearly establish where to draw the line (i.e., where to insert the “Heisenberg cut”) - something that no one has been able to achieve. In any case, the CH formalism assumes that quantum mechanics does apply to everything so we will stick to such a premise. The quote also mentions that the application of the quantum formalism to measurement scenarios must not involve “the evasion and equivocation ridiculed by Bell” in [@Bel:90]. So what does Bell say in [@Bel:90] regarding a satisfactory quantum formalism (i.e., one that solves the measurement problem)? He concisely states the following: > The theory should be fully formulated in mathematical terms, with nothing left to the discretion of the theoretical physicist. ([@Bel:90 p. 33]) Then, according to Bell, there are two main components required by a valid solution for the measurement problem: 1. The theory should be fully formulated in mathematical terms: i.e., concepts such as measurement, measuring apparatus, observer or macroscopic should not be part of the fundamental language of the theory. 2. Nothing should be left to the discretion of the theoretical physicist: i.e., successful applications of the theory must not require any input not contained in the description of the situation at hand at the fundamental level. The point, then, is that in order to solve the measurement problem it is not enough to construct a formalism fully written in precise terms. One must also make sure that successful applications of the formalism do not require the introduction of information that is not already contained in the fundamental description given by the theory of the situation one wants to consider. That is, once a complete quantum description of the measurement scenario is given, including the quantum state of the apparatus and the full Hamiltonian (and remember that we are assuming that, at least in principle, that is always possible because we are assuming that quantum mechanics applies to everything), then, with that information alone, one must be able to use the theory to make concrete predictions regarding the possible final outcomes of the experiment. It is important to emphasize that the above mentioned restriction to introduce “further information not contained in the description at the fundamental level” does not preclude the full specification of the physical situation characterizing the experiment. On the contrary, one is expected to provide the full fledged quantum state of the complete system (which consists of the sub-system of interest together with all the devices and apparatuses involved), as well as the full Hamiltonian characterizing the behavior of the sub-system, all devices and their interactions with the sub-system. It is clear that without such detailed characterization of the situation it is impossible to make concrete prediction. The important issue is whether a certain approach, provided with all the elements mentioned above, is capable or not to produce the specific predictions (even if these are probabilistic in nature) that correspond to what is found in actual experiments. Let’s illustrate this issue with a simple example, first from the perspective of the standard interpretation. It consists of a free spin-$\frac{1}{2}$ particle, with initial state $${\|\psi (0)\rangle} = {\|+_x\rangle} = \frac{1}{\sqrt{2}} \left ({\|+_z\rangle}+{\|-_z\rangle} \right ) ,$$ to be measured by a suitable apparatus. The apparatus contains a macroscopic pointer, whose center of mass $y$ has an initial (ready) state given by a narrow wavefunction $\varphi (y)$ centered at $y=0$. For simplicity, we take the pointer’s free Hamiltonian and the one corresponding to the spin degree of freedom to be zero. We also ignore the degree of freedom associated to the position of the particle.[^3] The spin and the apparatus interact via the interaction Hamiltonian: $$\hat{H}_{I}=2 i \hbar \lambda \left( \frac{\partial}{\partial y} \right) \otimes S_z$$ with $\lambda$ a constant. Given our assumptions, the interaction Hamiltonian also represents the complete Hamiltonian $\hat{H}$ of the whole system. Thus, putting everything together, we have that the initial state of the complete system is given by $$\left\vert \Psi(0)\right\rangle = \varphi(y)\otimes {\|\psi (0)\rangle} = \varphi(y) \otimes \frac{1}{\sqrt{2}} \left ({\|+_z\rangle}+{\|-_z\rangle} \right ) ,$$ and that the total Hamiltonian is $$\hat{H}=2 i \hbar \lambda \left( \frac{\partial}{\partial y} \right) \otimes S_z .$$ The situation is thus, according to the theory, described in full. We have provided the entire Hamiltonian of the complete system, including the measuring apparatus, and the initial state of the whole system as well. The next step is to use the dynamical law of the theory (i.e., the Schrödinger equation), from which it is easy to show that the state of the complete system at time $t$ is given by $${\|\Psi(t)\rangle} =e^{-\frac{i}{\hbar}\hat{H}t}\left\vert \Psi(0)\right\rangle = \frac{1}{\sqrt{2}} \left [ \varphi(y-\lambda t) \otimes {\|+_z\rangle}+ \varphi(y+\lambda t) \otimes ({\|-_z\rangle} \right ] . \label{sup}$$ Therefore, if $\lambda t$ is big enough, one can infer the value of the spin along $z$ by looking at the position of the pointer. Note however, that according to standard quantum mechanics, the final state can perfectly well be written as $${\|\Psi(t)\rangle} = \frac{1}{\sqrt{2}} \left [ \varphi_+(y,t) \otimes {\|+_x\rangle}+ \varphi_-(y,t) \otimes ({\|-_x\rangle} \right ]$$ with $$\begin{aligned} \varphi_+(y,t) & \equiv \frac{1}{\sqrt{2}} \left[ \varphi(y-\lambda t) +\varphi(y+\lambda t) \right] , \nonumber\\ \varphi_-(y,t) & \equiv \frac{1}{\sqrt{2}} \left[ \varphi(y-\lambda t) - \varphi(y+\lambda t) \right] , \nonumber\end{aligned}$$ so it seems that one could also use the system to measure the spin along $x$ by projecting the state of the center of mass of the pointer unto $\varphi_+(y,t)$ and $\varphi_-(y,t)$. We of course know that if we perform the experiment in the laboratory we will end up with either $\varphi(y-\lambda t)$ or $\varphi(y+\lambda t)$ and not with either $\varphi_+(y,t)$ or $\varphi_-(y,t)$. The first two are perfectly sensible, well-localized states for a macroscopic object but the latter two represent bizarre “Schrödinger cat” states. The issue of course is how do we know this? Is the standard interpretation capable of predicting it? Certainly not! The truth is that our knowledge of which one is the appropriate basis comes from the experience we have with macroscopic objects: we know that they always possess well-defined positions. The problem is that standard quantum mechanics is unable to account for this because it is impossible to derive such a result from the standard formalism alone, even if as above, the complete system is completely described to the extent required by the theory. The natural question, then, is how does standard quantum mechanics manage to be as successful as it is, in spite of such a glaring deficiency? The answer is that, in order to make such accurate predictions, it requires to be implicitly supplemented by external information regarding what it is that the apparatus one uses actually measures (the spin along $z$ in the above example, or, more generally, the fact that macroscopic objects always possess well-localized states). Given that such information is not contained, codified or accounted for in the standard fundamental description of the situation (given by the complete quantum state and the total Hamiltonian), we conclude that the standard interpretation does not satisfy Bell’s criteria, and so it does not offer a satisfactory solution to the measurement problem. It is important to point out, though, that what the above example illustrates is the inability of the standard interpretation to solve the so-called basis problem (i.e., the impossibility, given the full quantum description of a system, to pinpoint the appropriate basis to describe actual experimental results); and that solving the basis problem represents a necessary but not a sufficient condition in order to solve the measurement problem. What also needs to be ensured is that the formalism in question is able to accommodate, again, without any external input, the fact that at the end of the experiment what obtains, or at least what we perceive, is only one of the terms of the final state written as a superposition in the appropriate basis (e.g., either one or the other of the terms in equation (\[sup\])). How does the situation changes from the perspective of alternatives to the standard formalism, like objective collapse models (e.g., GRW or CSL) or de Broglie-Bohm mechanics? In the first case, given that the spontaneous collapses occur into highly localized states, and that the efficiency of the collapse process increases rapidly with the number of elementary constituents of the system, the theory straightforwardly predicts that the macroscopic apparatus will necessarily end-up in a state with well-defined pointer position (i.e., one of the terms in the first basis and not the second); and that is enough to determine, given the quantum description, what it is that the apparatus actually measures. Regarding Bohmian mechanics, a similar thing happens but for a different reason. In such case the apparatus also ends-up in a state with well-defined pointer position, but this time because the fundamental Bohmian description contains, beside the quantum state, the Bohmian particles, which always posses well-defined positions. As a result, the theory also unambiguously dictates that the first basis is the appropriate one to describe the system of our example. What about CH? Well, the fundamental description of the CH formulation, regarding the situation described above, certainly contains frameworks with final projections into states of the first or the second basis. How does it manage, then, to make predictions? Does it require, as does the standard interpretation, some extra input not given at the fundamental level? Indeed, much of the analysis in [@Oko.Sud:14b], which we review below, aims at showing that this is precisely what occurs for standard measuring scenarios, and that in spite of what proponents of the formalism sometimes claim, CH in fact requires such external input in order to be successful. We conclude, then, that the CH formalism cannot be considered as providing a viable solution to the measurement problem. At any rate, before reviewing in detail our arguments in this respect, we will say a few words about the so-called second measurement problem considered in [@Gri:15]. The second measurement problem ------------------------------ An important component of [@Gri:15] is committed to argue that CH is preferable with respect to other formulations of quantum mechanics (e.g., objective collapse models, or de Broglie-Bohm mechanics) because, besides solving the standard measurement problem, is unique in solving as well the so-called second measurement problem. This second measurement problem is described as the inability of a formalism to explain how an actual experimental result (i.e., an actual pointer direction of an apparatus) is related to the corresponding property of the measured system at a time before the measurement took place. The problem, according to the author of [@Gri:15], is that experimental physicists routinely use measurement results in order to infer which properties measured systems possessed before measurements, and so, he claims, a satisfactory quantum formalism must accommodate this common practice. The author also points out that this second measurement problem has received little attention in the literature but that solving it is as important as solving the “first” measurement problem. In spite of these opinions, it seems to us that the reason why almost nobody is concerned with the second measurement problem is because it, unlike the standard measurement problem, does not represent a conceptual problem but merely an unfamiliar feature present in some formulations of the theory (not unlike the validity of the uncertainty principle or the existence of entanglement). We believe that the author of [@Gri:15] finds this aspect to be problematic only because it clashes with our intuitions, common practices and beliefs but that neither the theory nor its usage forces us to assume that experiments must determine preexisting values of measured properties. That is, nothing compels us to ensure that the inference that experimental physicists routinely draw, regarding properties possessed by measured systems, is sound. Therefore, the absence of a “solution” to the second measurement problem does not render a theory inconsistent, ambiguous or vague, in contrast to what one faces in the absence of a solution to the standard measurement problem. It seems to us, furthermore, that the type of arguments the author uses in order to defend the breakdown of unicity within CH, i.e., arguments to the effect that such feature seems problematic only because it clashes with our imperfect classical intuitions, can well be used in order to defend a theory that does not address the second measurement problem, i.e., a theory incorporating the notion that “measurements routinely perturb measured systems.” Such a feature is only strange from the point of view of classical physics, but it does not really represent a conceptual or internal problem for any formalism which contains it. We conclude, then, that the inability of other solutions to the measurement problem to “solve” the second measurement problem does not constitute a relevant complication and that the alleged capacity of CH to solve it does not render it superior in any way. “Measurements according to Consistent Histories” in a nutshell {#MCH} ============================================================== As we show in [@Oko.Sud:14b], the application of the CH formalism to measurement scenarios requires the (often implicitly) introduction of input not contained in the formalism at the fundamental level. For example, in order to successfully apply the formalism to a concrete measurement situation, one needs to know in advance (or as Bell would put it, “using discretion”) what it is that the apparatus one is using actually measures. Only then one is able to choose the appropriate framework, i.e., the one that contains histories for which the apparatus is in a well-defined pointer position when the measurement is completed. But what about the Principles of Utility and Equality? Do not they imply that one does not actually need to know in advance what is it that one is measuring? That any choice of framework is as valid as any other, and that one takes the one that, for whatever reason, considers to be more useful? We do not believe that this way of reasoning is valid. That is because, as a matter of fact, given a concrete measurement scenario, it is not the case that all frameworks are equally valid and that one is more useful or informative than the others. The truth is that there is only one framework which contains the history that correctly describes the actual experimental results observed by the experimentalist. And the problem is that CH, in the absence of external input, is incapable of predicting which one is the framework that will do the job. There are two common responses to our claims. These assert that, for measurement scenarios, frameworks must be chosen either: 1. To model the experimental situation at hand. 2. According to the questions one is interested in answering. The problem with the first option is that, as we saw in the example of section \[MP\], the fact that a given apparatus actually measures some property is something that cannot be deduced from the CH formalism, even if the full quantum description of the situation at hand is provided. Instead, such knowledge must be discovered by means external to CH (e.g., empirically) in order to be used in the selection of the framework that actually “models the experimental situation” at hand. That is, if one is given the description of an apparatus and the subsystem of interest in purely quantum terms, then it is impossible to infer, using only the CH formalism, which is the property that will be actually measured in the laboratory, and so it is impossible, without external input, to select the appropriate framework. We conclude, then, that in order to apply the first recipe one requires input that goes beyond that provided by the theory. The problem with the second option, i.e., choosing the framework according to the questions one is interested in answering, is that it is not clear what is supposed to be the relation between the framework one chooses, which presumably reflects what one is interested in, and what one in fact observes when the experiment is performed. For example, suppose that we are interested in the spin along $z$ of a particle but that we only have at our disposal an apparatus that measures spin along $x$ (of course, as we just saw, we know what it actually measures through experience, and not because the CH formalism is capable of predicting it). Then, according to the second option above, we must choose a framework that reflects what we are interested in, i.e., a framework with projections onto spin-up and spin-down along $z$. However, it is clear that such a choice of framework will make absolutely no difference regarding what we will in fact observe when we perform the experiment with the equipment provided! The choice of framework, then, must be done not according to the questions one is interested in answering, as the second recipe above suggests, but according to the actual experimental set-up. The problem, of course, is that such conclusion takes us back to the first option above which we already saw is unsatisfactory. Therefore, for measurement scenarios, information not provided by the CH formalism is essential in order to select the framework. We conclude that the CH formalism does not meet the second of Bell’s criteria described above from which it follows that it does not solve the measurement problem. Reply to “Consistent Quantum Measurements” {#Rep} ========================================== Section 7 in [@Gri:15] directly addresses the criticisms we raised in [@Oko.Sud:14b]; in this section we present our response to such a rebuttal. In order to evaluate our objections, the author of [@Gri:15] begins by presenting four quotes from [@Oko.Sud:14b] (which he labels Q1-Q4) that, he believes (and we agree), summarize our complains. Then he proceeds to comment on them one by one. We will follow a similar procedure. Q1 asserts, in essence, that in order to be useful when dealing with measurements, CH requires the introduction of external input. The author of [@Gri:15] accepts that that in fact is the case but defends it by stating the following: > \[O\]ne would anticipate that any plausible fundamental theory of quantum mechanics would contain no reference to \[measurements\], and therefore additional concepts must be employed in order to describe them. The proper question is not whether elements external to the formalism are used, but instead whether these “extras” are needed because one is discussing a measurement and not some other physical process. ([@Gri:15 p. 12]) We find the first sentence above, in the light of Bell quote of section \[MP\], very revealing. Let us recall that, according to Bell, a proper solution to the measurement problem contains two aspects: i) measurements do not appear at the fundamental level and ii) successful applications do not require external input. Then, if what is said in the quote is true about CH, then it is clear that it does not satisfy Bell’s requirements and so it does not solve the measurement problem. As we explained in section \[MP\], the point is that it is true that satisfactory solutions should not contain concepts such as measurement as part of their fundamental language (as the first part of the first sentence of the quote suggests). But that is not enough; it must also be the case for satisfactory solutions to be capable of describing all situations, including measurements, in terms of the fundamental language of the theory. Therefore, the fact that a formalism requires the introduction of external elements in order to deal with measurements, as CH does, signals that the formalism in question fails to solve the measurement problem. The author of [@Gri:15] then proceeds to explain in detail how external input (of the kind explicitly not allowed by Bell’s criteria) is introduced while using CH for a specific application. All he manages to show, though, is that once such an external input is brought in, the theory manages to deal satisfactorily with the situation. Note however that the same thing would happen if one where using the orthodox interpretation: if one allows the introduction of external input, in the form of knowledge regarding which interactions count as measurements, then the standard quantum mechanical formalism works beautifully. The issue of course is that, as the author of [@Gri:15] agrees, the need to introduce this extraneous information renders the orthodox theory problematic. The question is: why does he believe that the situation regarding CH is any different? The quote Q2 starts by claiming that “\[A\]mong all the possible frameworks, only one is suitable to describe what in fact we perceive and experience...” It then continues to stress that CH is incapable of selecting the framework that will do the job. The author of [@Gri:15] points out that the topic of perception is subtle and controversial and, taking that as a general remark regarding perceptions, we agree. Nevertheless, the use of the concept of perception we intended should not be controversial or problematic. All we had in mind is the idea that one can meaningfully talk about actual measurement outcomes; and it is clear, from many passages in [@Gri:15] and elsewhere, that the author of [@Gri:15] agrees with that (for example, in [@Gri:15] he describes a measurement as “a physical process with some specific macroscopic outcome”). Our point, then, is that CH is incapable of selecting (in advance, and without inappropriate external input) the framework suitable to describe the actual experimental outcomes - a fact that, as we saw, the author of [@Gri:15] explicitly agrees with. Q3 elaborates on the idea that “the fact that a given measuring apparatus actually measures some property is something that cannot be deduced from the CH formalism...” and concludes that, while using such an apparatus, “CH is incapable of predicting which framework one must choose.” As a response, the author of [@Gri:15] offers the following: > The choice of a framework is one made by the physicist constructing a quantum description, and there is no constraint among the principles of CH that prescribes which one must be used. However, the consistent historian will add that if the “given measuring apparatus actually measures some property,” that fact alone is enough to constrain the choice of an appropriate framework..., contrary to the claim in the final sentence in Q3. ([@Gri:15 p. 12]) Once again we wonder how is CH supposed to improve the situation regarding the orthodox interpretation. That is, if one is allowed to use the extra-theoretical knowledge that a “given measuring apparatus actually measures some property,” then why would one seek to overcome the standard interpretation? Finally, quote Q4 contains the statement “the CH formalism is incapable of picking out the right framework.” As a response, the author of [@Gri:15] argues that our use of the notion of the “right framework” exhibits a classical prejudice to the effect that “at any given time there is a single true state of the world.” He further claims that, at the end of the day, such a prejudice is what lies behind our criticisms of CH. That is, he believes that what truly bothers us about CH is the fact that it does not obey the Principle of Unicity. This, of course, is not correct. While we do find the breakdown of unicity inconvenient, such worry is independent from the objections we present in [@Oko.Sud:14b]. The confusion, at least in this particular case, arises because the author of [@Gri:15] attaches too much metaphysical baggage to the idea of “the right framework.” All we have in mind with such a phrase, as with the idea of “the framework that correspond with what we perceive” in Q2, is the following: the framework suitable to describe the actual experimental outcomes. Given that the notion of actual experimental outcomes is accepted and commonly employed by the author of [@Gri:15], we conclude that his critique of our statements is unfounded. The upshot, once more, is that CH is unable to make correct predictions in the absence of external knowledge. Conclusions {#C} =========== We have carefully considered the response offered in [@Gri:15] to the criticisms of the CH formalism we presented in [@Oko.Sud:14b]; we have found that such response does not manage to refute the basis of our critiques. We do believe, though, that the dispute has served to expose the core of our disagreement with the author of [@Gri:15], as well as with other advocates of the CH approach. In this regard, we have learned that in order to have a sufficiently precise analysis of the way in which the advocates of a scheme such as CH intend to address the measurement problem, one needs a very clear characterization of the problem itself, as well as of what would constitute a satisfactory solution thereof. In order to achieve this clear characterization we have relied upon John Bell’s views on the matter and we have used those to pinpoint the shortcomings of the CH approach in dealing with the measurement problem. The measurement problem in quantum theory has been one of the most highly debated topics in the foundations of physics. Therefore, a critical analysis of the various proposals to address it represents an essential aspect of our search for a better understanding of the workings of nature. The CH approach is one of the most well-developed proposals attempting to resolve the problem, while preserving almost intact the essence of the standard formulation of quantum theory. It must be said that, as such, it has served to highlight the difficulties one must face when attempting such a task. Unfortunately, as we have been able to shown in previous works [@Oko.Sud:14a; @Oko.Sud:14b] and has been further clarified in this manuscript, the conceptual difficulties that afflict quantum theory are such that they cannot be overcome by a relatively conservative proposal such as CH. Acknowledgments {#acknowledgments .unnumbered} =============== We acknowledge partial financial support from DGAPA-UNAM project IA400114 (EO) and CONACyT project 220738 (DS). [^1]: Other objections against CH can be found in [@Esp:87; @Dow.Ken:96; @Ken:97; @Bar:99; @Ken:10; @Oko.Sud:14a]. [^2]: The formulations of the measurement problem developed in [@Mau:95], instead of specifying in detail the theoretical framework to be dealt with, imposes general restrictions that all satisfactory formulations must obey. [^3]: None of these simplifying assumptions impinges on the conceptual validity of the example.	{ "pile_set_name": "ArXiv" }
--- abstract: 'For transversely homogeneous foliations on compact manifolds whose global holonomy group has connected closure, it is shown that either all holonomy covers of the leaves have polynomial growth with degree bounded by a common constant, or all holonomy covers of the leaves have exponential growth. This is an extension of a recent answer given by Breuillard and Gelander to a question of Carrière. Examples of transversely projective foliations satisfying the above condition were constructed by Chihi and ben Ramdane.' address: - \| Departamento de Xeometría e Topoloxía\ Facultade de Matemáticas\ Universidade de Santiago de Compostela\ 15782 Santiago de Compostela\ Spain - \| Jagiellonian University\ Faculty of Mathematics and Computer Science\ Institute of Mathematics\ ul. prof. Stanisława Łojasiewicza 6\ 30-348 Kraków\ Poland author: - 'Jesús A. Álvarez López' - Robert Wolak title: Growth of some transversely homogeneous foliations --- Introduction {#s: intro} ============ Let $G$ be a Lie group, and $P\subset G$ a closed subgroup. The concept of (transversely homogeneous) $(G,G/P)$-foliation was introduced by Blumenthal [@Blumenthal1979]: it is a smooth foliation with the structure given by a defining cocycle with values in the $G$-manifold $G/P$. Assume that the $G$-action on $G/P$ is faithful, and $G/P$ is connected. Let ${{\mathcal{F}}}$ be a $(G,G/P)$-foliation on a compact connected manifold $M$. Then Blumenthal proved that there a homomorphism ${\mathbf{h}}:\pi_1(M)\to G$, well-defined up to conjugation, whose image is denoted by $\Gamma$, such that the lift $\widetilde{{\mathcal{F}}}$ to the cover $\widetilde M$ of $M$ associated to $\ker{\mathbf{h}}$ is given by a $\Gamma$-equivariant submersion $D:\widetilde M\to G/P$, where $\Gamma$ acts on $\widetilde M$ via deck transformations [@Blumenthal1979 Theorem 1]. It is said that $\Gamma$ is the ([global]{}) [holonomy group]{} of ${{\mathcal{F}}}$. Moreover Blumenthal proved that, if $\pi_1(M)$ has non-exponential growth (respectively, polynomial growth of degree $d$), then all leaves of ${{\mathcal{F}}}$ have non-exponential growth (respectively, polynomial growth of degree $d$) [@Blumenthal1979 Theorem 1]. (Recall that the growth is one of the classical invariants of leaves of foliations on compact manifolds [@Plante1973; @Plante1975].) When ${{\mathcal{F}}}$ is a Lie $G$-foliation ($P=\{1\}$), Carrière has shown that [@Carriere1988]: - the leaves of ${{\mathcal{F}}}$ are Følner if and only if $G$ is solvable; - the leaves of ${{\mathcal{F}}}$ have polynomial growth if and only if $G$ is nilpotent; and, - in the last case, the leaves of ${{\mathcal{F}}}$ have polynomial growth with degree less or equal than the degree of nilpotence of $G$. Carrière asked in [@Carriere1988] about the existence of a Lie $G$-foliation on a compact manifold whose leaves have neither polynomial nor exponential growth. This question was recently answered by Breuillard and Gelander [@BreuillardGelander2007 Theorem 10.1], obtaining the following dichotomy as consequence of their study of a topological Tits alternative: - either all leaves of ${{\mathcal{F}}}$ have polynomial growth with degree bounded by a common constant; - or all leaves of ${{\mathcal{F}}}$ have exponential growth. Indeed, Carrière, and Breuillard and Gelander stated these results for Riemannian foliations on compact manifolds, either by considering the residual set of leaves without holonomy, or by considering the holonomy covers of all leaves. This kind of extension is straightforward by Molino’s theory [@Molino1988]. In this paper, we show that the arguments of Breuillard and Gelander about growth of Lie foliations can be extended to $(G,G/P)$-foliations assuming that $\overline\Gamma$ is connected. Let us remark that a $(G,G/P)$-foliation may not be Riemannian (it is Riemannian if $P$ is compact). Precisely, we prove the following complement of Blumenthal’s observations about growth of transversely homogeneous foliations. \[t:growth transv homog folns\] Let $G$ be a Lie group, and $P\subset G$ a closed subgroup such that $G/P$ is connected and the $G$-action on $G/P$ is faithful. Let ${{\mathcal{F}}}$ be a $(G,G/P)$-foliation on a compact connected manifold with holonomy group $\Gamma\subset G$. If $\overline\Gamma$ is connected, then: - either all holonomy covers of the leaves of ${{\mathcal{F}}}$ have polynomial growth with degree bounded by a common constant; - or all holonomy covers of the leaves have exponential growth. As a particular case, $({\operatorname{PSL}}(2;{{\mathbb R}}),{{\mathbb R}}P^1)$-foliations are called [transversely projective]{}. The first example of a codimension $1$ foliation on a compact $3$-manifold with nonzero Godbillon-Vey invariant, due to Roussarie, was transversely projective, and its holonomy group $\Gamma$ is discrete and uniform in ${\operatorname{PSL}}(2;{{\mathbb R}})$ (see e.g. [@CandelConlon2000-I Example 1.3.14]). Chihi and ben Ramdane [@ChihiRamdane2008] have shown that, for any transversely projective foliation on a compact manifold, if the Godbillon-Vey invariant is nonzero, then $\Gamma$ is either discrete or dense in ${\operatorname{PSL}}(2;{{\mathbb R}})$. Moreover they constructed examples satisfying the second alternative on compact manifolds of dimension $\ge5$, and therefore satisfying the conditions of Theorem \[t:growth transv homog folns\]. Preliminaries {#s:prelim} ============= Coarse quasi-isometries and growth of metric spaces {#ss:growth metric} --------------------------------------------------- A [net]{} in a metric space $M$, with metric $d$, is a subset $A\subset M$ that satisfies $d(x,A)\le C$ for some $C>0$ and all $x\in M$; the term [$C$-net]{} is also used. A [coarse quasi-isometry]{} between $M$ and another metric space $M'$ is a bi-Lipschitz bijection between nets of $M$ and $M'$; in this case, $M$ and $M'$ are said to be [coarsely quasi-isometric]{} (in the sense of Gromov) [@Gromov1993]. If such a bi-Lipschitz bijection, as well as its inverse, has dilation $\le\lambda$, and it is defined between $C$-nets, then it will be said that the coarse quasi-isometry has [distortion]{} $(C,\lambda)$. A family of coarse quasi-isometries with a common distortion will be called [uniform]{}, and the corresponding metric spaces are called [uniformly]{} coarsely quasi-isometric. The version of growth for metric spaces given here is taken from [@AlvCandel:gcgol]. Since [@AlvCandel:gcgol] is not finished yet, some short proofs are included. Recall that, given non-decreasing functions[^1] $u,v:[0,\infty)\to[0,\infty)$, it is said that $u$ is [dominated]{} by $v$, written $u\preccurlyeq v$, when there are $a,b,c,d>0$ such that $u(r)\le a\,v(br+c)+d$ for all $r$. If $u\preccurlyeq v\preccurlyeq u$, then it is said that $u$ and $v$ represent the same [growth type]{}; this is an equivalence relation and “$\preccurlyeq$” defines a partial order relation between growth types called [domination]{}. For a family of pairs of non-decreasing functions $[0,\infty)\to[0,\infty)$, [uniform]{} domination means that those pairs satisfy the above condition of domination with the same constants $a,b,c$. A family of non-decreasing functions $[0,\infty)\to[0,\infty)$ will be said to have [uniformly]{} the same growth type if they uniformly dominate one another. For a complete connected Riemannian manifold $L$, the growth type of each mapping $r\mapsto{\operatorname{vol}}B(x,r)$ is independent of $x$ and is called the [growth type]{} of $L$. Another definition of [growth type]{} can be similarly given for metric spaces whose bounded sets are finite, where the number of points is used instead of the volume. Let $M$ be a metric space with metric $d$. A [quasi-lattice]{} $\Gamma$ of $M$ is a $C$-net of $M$ for some $C\ge0$ such that, for every $r\ge0$, there is some $K_r\ge0$ such that ${\operatorname{card}}(\Gamma\cap B(x,r))\le K_r$ for every $x\in M$. It is said that $M$ is of [coarse bounded geometry]{} if it has a quasi-lattice. In this case, the [growth type]{} of $M$ can be defined as the growth type of any quasi-lattice $\Gamma$ of $M$; i.e., it is the growth type of the [growth function]{} $r\mapsto v_\Gamma(x,r)={\operatorname{card}}(B(x,r)\cap\Gamma)$ for any $x\in\Gamma$. This definition can be proved to be independent of $\Gamma$ as follows. Let $\Gamma'$ be another quasi-lattice in $M$. So $\Gamma$ and $\Gamma'$ are $C$-nets in $M$ for some $C\ge0$, and there is some $K_r\ge0$ for each $r\ge0$ such that ${\operatorname{card}}(B(x,r)\cap\Gamma)\le K_r$ and ${\operatorname{card}}(B(x,r)\cap\Gamma')\le K_r$ for all $x\in M$. Fix points $x\in\Gamma$ and $x'\in\Gamma'$, and let $\delta=d(x,x')$. Because $B(x,r)\subset B(x',r+\delta)$ and $\Gamma'$ is a $C$-net, it follows that $$B(x,r)\cap\Gamma\subset\bigcup_{y'\in B(x',r+\delta+C)\cap\Gamma'}B(y',C)\cap\Gamma'\;,$$ yielding $$v_\Gamma(x,r)\le K_C\,v_{\Gamma'}(x',r+\delta+C)\le K_C\,v_{\Gamma'}(x',(1+\delta+C)r)$$ for all $r\ge1$. Hence the growth type of $r\mapsto v_\Gamma(x,r)$ is dominated by the growth type of $r\mapsto v_{\Gamma'}(x',r)$. For a family of metric spaces, if they satisfy the above condition of coarse bounded geometry with the same constants $C$ and $K_r$, then they are said to have [uniformly]{} coarse bounded geometry. If moreover the lattices involved in this condition have growth functions defining uniformly the same growth type, then these metric spaces are said to have [uniformly]{} the same growth type. The condition of coarse bounded geometry is satisfied by complete connected Riemannian manifolds of bounded geometry, and by discrete metric spaces with a uniform upper bound on the number of points in all balls of each given radius [@BlockWeinberger1997]. In those cases, the two given definitions of growth type are equal. \[l:growth type\] Two coarsely quasi-isometric metric spaces of coarse bounded geometry have the same growth type. Moreover, if a family of metric spaces are uniformly coarsely quasi-isometric to each other, then they have uniformly the same growth type. Let $\phi:A\to A'$ be a coarse quasi-isometry between metric spaces $M$ and $M'$ of coarse bounded geometry. Then $A$ is of coarse bounded geometry too, and thus it has some lattice $\Gamma$, which is also a lattice in $M$ because $A$ is a net. Since $\phi$ is a bi-lipschitz bijection, it easily follows that $\Gamma$ and $\phi(\Gamma)$ have the same growth type, and that $\phi(\Gamma)$ is a lattice in $A'$, and thus in $M'$ too because $A'$ is a net. This argument has an obvious uniform version for a family of metric spaces. Pseudogroups ------------ A [pseudogroup]{} ([of local transformations]{}) on a topological space $T$ is a collection ${{\mathcal{H}}}$ of homeomorphisms between open subsets of $T$ that contains the identity map and is closed under composition (wherever defined), inversion, restriction and combination (union) of maps. Such a pseudogroup ${{\mathcal{H}}}$ is [generated]{} by a set $E\subset{{\mathcal{H}}}$ if every element of ${{\mathcal{H}}}$ can be obtained from $E$ by using the above pseudogroup operations; usually, $E$ will be symmetric ($h^{-1}\in E$ if $h\in E$). The [orbit]{} of a point $x\in T$ is the set ${{\mathcal{H}}}(x)=\{\,h(x)\mid h\in{{\mathcal{H}}},\ x\in{\operatorname{dom}}h\,\}$. Many other basic dynamical concepts can be generalized to pseudogroups because they are natural generalizations of dynamical systems (each group action on a topological space generates a pseudogroup). The [restriction]{} of ${{\mathcal{H}}}$ to an open subset $U\subset T$ is the pseudogroup ${{\mathcal{H}}}\|_U=\{\,h\in{{\mathcal{H}}}\mid{\operatorname{dom}}h\cup{\operatorname{im}}h\subset U\,\}$. It is said that ${{\mathcal{H}}}$ is [quasi-analytic]{} when any $h\in{{\mathcal{H}}}$ is the identity around any $x\in{\operatorname{dom}}h$ if $h$ is the identity on some open set whose closure contains $x$; [quasi-analytic]{} group actions are similarly defined, which means that they generate quasi-analytic pseudogroups. Pseudogroups that are equivalent in the following sense should be considered to have the same dynamics. Let ${{\mathcal{H}}}$ and ${{\mathcal{H}}}'$ be pseudogroups on respective topological spaces $T$ and $T'$. An [equivalence]{} $\Phi:{{\mathcal{H}}}\to{{\mathcal{H}}}'$ is a maximal collection $\Phi$ of homeomorphisms of open subsets of $T$ to open subsets of $T'$ such that: - If $\phi\in\Phi$, $h\in{{\mathcal{H}}}$ and $h'\in{{\mathcal{H}}}'$, then $h'\phi h\in\Phi$; and - ${{\mathcal{H}}}$ and ${{\mathcal{H}}}'$ are generated by the maps of the form $\psi^{-1}\phi$ and $\psi\phi^{-1}$, respectively, with $\phi,\psi\in\Phi$. In this case, ${{\mathcal{H}}}$ and ${{\mathcal{H}}}'$ are said to be [equivalent]{}. If $\Phi:{{\mathcal{H}}}\to{{\mathcal{H}}}'$ is an equivalence, then $\Phi^{-1}=\{\phi^{-1}\ \|\ \phi\in\Phi\}$ is an equivalence ${{\mathcal{H}}}'\to{{\mathcal{H}}}$, called the [inverse]{} of $\Phi$. An equivalence $\Phi:{{\mathcal{H}}}\to{{\mathcal{H}}}'$ is [generated]{} by a subset $\Phi_0\subset\Phi$ if all of the elements of $\Phi$ can be obtained by restriction and combination of composites $h'\phi h$ with $h\in{{\mathcal{H}}}$, $\phi\in\Phi_0$ and $h'\in{{\mathcal{H}}}'$. If $\Phi:{{\mathcal{H}}}\to{{\mathcal{H}}}'$ and $\Psi:{{\mathcal{H}}}'\to{{\mathcal{H}}}''$ are equivalences, then the maps $\psi\phi$, for $\phi\in\Phi$ and $\psi\in\Psi$, generate an equivalence $\Psi\Phi:{{\mathcal{H}}}\to{{\mathcal{H}}}''$, called the [composite]{} of $\Phi$ and $\Psi$. Thus the equivalence of pseudogroups is an equivalence relation. An equivalence $\Phi:{{\mathcal{H}}}\to{{\mathcal{H}}}'$ induces a homeomorphism between the corresponding orbit spaces, $\bar\Phi:T/{{\mathcal{H}}}\to T'/{{\mathcal{H}}}'$. A basic example of a pseudogroup equivalence is the following. Let ${{\mathcal{H}}}$ be a pseudogroup of local transformations of a space $T$, let $U\subset T$ be an open subset that meets every ${{\mathcal{H}}}$-orbit. Then the inclusion map $U\hookrightarrow T$ generates an equivalence ${{\mathcal{H}}}\|_U\to{{\mathcal{H}}}$. In fact, this example can be used to describe any pseudogroup equivalence in the following way. Let ${{\mathcal{H}}}$ and ${{\mathcal{H}}}'$ be pseudogroups of local transformations of respective spaces $T$ and $T'$, and let $\Phi:{{\mathcal{H}}}\to{{\mathcal{H}}}'$ be an equivalence. Let ${{\mathcal{H}}}''$ be the pseudogroup of local transformations of $T''=T\sqcup T'$ generated by ${{\mathcal{H}}}\cup{{\mathcal{H}}}'\cup\Phi$. Then the inclusions of $T$ and $T'$ in $T''$ generate equivalences $\Psi_1:{{\mathcal{H}}}\to{{\mathcal{H}}}''$ and $\Psi_2:{{\mathcal{H}}}'\to{{\mathcal{H}}}''$ so that $\Phi=\Psi_2^{-1}\Psi_1$. For a pseudogroup ${{\mathcal{H}}}$ on a locally compact space $T$, the orbit space $T/{{\mathcal{H}}}$ is compact if and only if there exists a relatively compact open subset of $T$ that meets every ${{\mathcal{H}}}$-orbit. The following is a stronger “compactness” condition on a pseudogroup. \[d:compactly generated\] Let ${{\mathcal{H}}}$ be a pseudogroup on a locally compact space $T$. It is said that ${{\mathcal{H}}}$ is [compactly generated]{} if there is a relatively compact open set $U$ in $T$ meeting each ${{\mathcal{H}}}$-orbit, and such that ${{\mathcal{H}}}\|_U$ is generated by a finite symmetric collection $E$ so that each $g\in E$ has an extension $\bar g\in{{\mathcal{H}}}$ with $\overline{{\operatorname{dom}}g}\subset{\operatorname{dom}}\bar g$. It was observed in [@Haefliger1985] that this notion is invariant by equivalences, and that the relatively compact open set $U$ meeting each orbit can be chosen arbitrarily. If $E$ satisfies the conditions of Definition \[d:compactly generated\], it will be called a [system of compact generation]{} of ${{\mathcal{H}}}$ on $U$. For $h\in{{\mathcal{H}}}$ and $x\in{\operatorname{dom}}h$, let ${\operatorname{germ}}(h,x)$ denote the germ of $h$ at $x$. Then $${{\mathfrak{H}}}=\{\,{\operatorname{germ}}(h,x)\mid h\in{{\mathcal{H}}},\ x\in{\operatorname{dom}}h\,\}$$ becomes a topological groupoid, with object space $T$, equipped with the étale topology, the operation induced by composition, and the source and target projections to $T$. For $x,y\in T$, let ${{\mathfrak{H}}}_x$ (respectively, ${{\mathfrak{H}}}^y$) denote the set of elements in ${{\mathfrak{H}}}$ with source $x$ (respectively, with target $y$), and let ${{\mathfrak{H}}}_x^y={{\mathfrak{H}}}_x\cap{{\mathfrak{H}}}^y$; in particular, the group ${{\mathfrak{H}}}_x^x$ will be called the [germ group]{} of ${{\mathcal{H}}}$ at $x$. Points in the same ${{\mathcal{H}}}$-orbit have isomorphic germ groups (if $y\in{{\mathcal{H}}}(x)$, an isomorphism ${{\mathfrak{H}}}_y^y\to{{\mathfrak{H}}}_x^x$ is given by conjugation with any element in ${{\mathfrak{H}}}_x^y$); hence the germ groups of the orbits make sense up to isomorphism. Under pseudogroup equivalences, corresponding orbits have isomorphic germ groups. The set ${{\mathfrak{H}}}_x$ will be called the [germ cover]{} of the orbit ${{\mathcal{H}}}(x)$ with base point $x$. The target map restricts to a surjective map ${{\mathfrak{H}}}_x\to{{\mathcal{H}}}(x)$ whose fibers are bijective to ${{\mathfrak{H}}}_x^x$ (if $y\in{{\mathcal{H}}}(x)$, a bijection ${{\mathfrak{H}}}_x^x\to{{\mathfrak{H}}}_x^y$ is given by left product with any element in ${{\mathfrak{H}}}_x^y$); thus ${{\mathfrak{H}}}_x$ is finite if and only if both ${{\mathfrak{H}}}_x^x$ and ${{\mathcal{H}}}(x)$ are finite. Moreover germ covers based on points in the same orbit are also bijective (if $y\in{{\mathcal{H}}}(x)$, a bijection ${{\mathfrak{H}}}_y\to{{\mathfrak{H}}}_x$ is given by right product with any element in ${{\mathfrak{H}}}_x^y$); therefore the germ covers of the orbits make sense up to bijections. Quasi-isometry type of orbits {#ss:q-i} ----------------------------- Let ${{\mathcal{H}}}$ be a pseudogroup on a space $T$, and $E$ a symmetric set of generators of ${{\mathcal{H}}}$. For each $h\in{{\mathcal{H}}}$ and $x\in{\operatorname{dom}}h$, let $\|h\|_{E,x}$ be the length of the shortest expression of ${\operatorname{germ}}(h,x)$ as product of germs of maps in $E$ (being $0$ if ${\operatorname{germ}}(h,x)={\operatorname{germ}}({\operatorname{id}}_T,x)$). For each $x\in T$, define metrics $d_E$ on ${{\mathcal{H}}}(x)$ and ${{\mathfrak{H}}}_x$ by $$\begin{gathered} d_E(y,z)=\min\{\,\|h\|_{E,y}\mid h\in{{\mathcal{H}}},\ y\in{\operatorname{dom}}h,\ h(y)=z\,\}\;,\\ d_E({\operatorname{germ}}(f,x),{\operatorname{germ}}(g,x))=\|fg^{-1}\|_{E,g(x)}\;.\end{gathered}$$ Notice that $$d_E(f(x),g(x))\le d_E({\operatorname{germ}}(f,x),{\operatorname{germ}}(g,x))\;.$$ Moreover, on the germ covers, $d_E$ is right invariant in the sense that, if $y\in{{\mathcal{H}}}(x)$, the bijection ${{\mathfrak{H}}}_y\to{{\mathfrak{H}}}_x$, given by right multiplication with any element in ${{\mathfrak{H}}}_x^y$, is isometric; so the isometry types of the germ covers of the orbits make sense without any reference to base points. In fact, the definition of $d_E$ on ${{\mathfrak{H}}}_x$ is analogous to the right invariant metric $d_S$ on a group $\Gamma$ defined by a symmetric system of generators $S$: $d_S(\gamma,\delta)=\|\gamma\delta^{-1}\|$ for $\gamma,\delta\in\Gamma$, where $\|\gamma\|$ is the length of the shortest expression of $\gamma$ as product of elements of $S$ (being $0$ if $\gamma=1$). Assume that ${{\mathcal{H}}}$ is compactly generated and $T$ locally compact. Let $U\subset T$ be a relatively compact open subset that meets all ${{\mathcal{H}}}$-orbits, let ${{\mathcal{G}}}={{\mathcal{H}}}\|_U$, and let $E$ be a symmetric system of compact generation of ${{\mathcal{H}}}$ on $U$. With this conditions, the quasi-isometry type of the ${{\mathcal{G}}}$-orbits with $d_E$ may depend on $E$ [@AlvCandel2009 Section 6]. So the following additional condition on $E$ is considered. \[d:recurrent finite symmetric family of generators\] With the above notation, it is said that $E$ is [recurrent]{} if, for any relatively compact open subset $V\subset U$ that meets all ${{\mathcal{G}}}$-orbits, there exists some $R>0$ such that ${{\mathcal{G}}}(x)\cap V$ is an $R$-net in ${{\mathcal{G}}}(x)$ with $d_E$ for all $x\in U$. The role played by $V$ in Definition \[d:recurrent finite symmetric family of generators\] can actually be played by any relatively compact open subset that meets all orbits [@AlvCandel2009 Lemma 4.3]. Furthermore there always exists a recurrent system of compact generation on $U$ [@AlvCandel2009 Corollary 4.5]. In other words, the recurrence of $E$ means that there is some $N\in{{\mathbb N}}$ such that $$\label{E^N} U=\bigcup_{h\in E^N}h^{-1}(V\cap{\operatorname{im}}h)\;,$$ where $E^N$ is the family of compositions of at most $N$ elements of $E$. \[t:quasi-isometric orbits\] Let ${{\mathcal{H}}}$ and ${{\mathcal{H}}}'$ be compactly generated pseudogroups on locally compact spaces $T$ and $T'$, let $U$ and $U'$ be relatively compact open subsets of $T$ and $T'$ that meet all orbits of ${{\mathcal{H}}}$ and ${{\mathcal{H}}}'$, let ${{\mathcal{G}}}$ and ${{\mathcal{G}}}'$ denote the restrictions of ${{\mathcal{H}}}$ and ${{\mathcal{H}}}'$ to $U$ and $U'$, and let $E$ and $E'$ be recurrent symmetric systems of compact generation of ${{\mathcal{H}}}$ and ${{\mathcal{H}}}'$ on $U$ and $U'$, respectively. Suppose that there exists an equivalence ${{\mathcal{H}}}\to{{\mathcal{H}}}'$, and consider the induced equivalence ${{\mathcal{G}}}\to{{\mathcal{G}}}'$ and homeomorphism $U/{{\mathcal{G}}}\to U'/{{\mathcal{G}}}'$. Then the ${{\mathcal{G}}}$-orbits with $d_E$ are uniformly quasi-isometric to the corresponding ${{\mathcal{G}}}'$-orbits with $d_{E'}$. An obvious modification of the arguments of the proof of [@AlvCandel2009 Theorem 4.6] gives the following. \[t:quasi-isometric germ covers of orbits\] With the notation and conditions of Theorem \[t:quasi-isometric orbits\], the germ covers of the ${{\mathcal{G}}}$-orbits with $d_E$ are uniformly quasi-isometric to the germ covers of the corresponding ${{\mathcal{G}}}'$-orbits with $d_{E'}$. \[c:growth orbits and their germ covers\] With the notation and conditions of Theorems \[t:quasi-isometric orbits\] and \[t:quasi-isometric germ covers of orbits\], the corresponding orbits of ${{\mathcal{G}}}$ and ${{\mathcal{G}}}'$, as well as their germ covers, have the same growth type, uniformly. This follows from Lemma \[l:growth type\] and Theorems \[t:quasi-isometric orbits\] and \[t:quasi-isometric germ covers of orbits\]. Growth of leaves {#ss:growth leaves} ---------------- Let us recall some basic concepts of foliation theory (see e.g. [@HectorHirsch1981-A; @HectorHirsch1983-B; @Godbillon1991; @CandelConlon2000-I; @Walczak2004]). Let ${{\mathcal{F}}}$ be a smooth foliation on a manifold $M$ given by a defining cocycle $(U_i,p_i,h_{ij})$ [@Haefliger1985; @Haefliger1988], where $p_i:U_i\to T_i$, and $h_{ij}:p_i(U_i\cap U_j)\to p_j(U_i\cap U_j)$ is determined by the condition $p_j=h_{ij}p_i$ on $U_i\cap U_j$. We can assume that $(U_i,p_i,h_{ij})$ is induced by a regular foliation atlas: the sets $U_i$ are the domains of the foliation charts, the maps $p_i$ are the local projections whose fibers are the plaques, and the maps $h_{ij}$ are the transverse components of the changes of coordinates. The equivalence class of the pseudogroup ${{\mathcal{H}}}$ on $T=\bigsqcup T_i$ generated by the maps $h_{ij}$ is independent of the choice of defining cocycle, and is called the [holonomy pseudogroup]{} of ${{\mathcal{F}}}$. There is a canonical identity between the space of leaves and the space of ${{\mathcal{H}}}$-orbits, $M/{{\mathcal{F}}}\equiv T/{{\mathcal{H}}}$. The [holonomy group]{} of each leaf $L$ is defined as the germ group of the corresponding orbit. It can be considered as a quotient of the fundamental group of $L$ by taking “chains” of sets $U_i$ along loops in $L$, and the corresponding covering space is called the [holonomy cover]{} $\widetilde{L}$ of $L$. If ${{\mathcal{F}}}$ admits a countable defining cocycle, then the leaves in some saturated residual subset of $M$ have trivial holonomy groups [@HectorHirsch1981-A; @HectorHirsch1983-B], and therefore they can be identified to their holonomy covers. Suppose that $M$ is compact. Then, given any Riemannian metric $g$ on $M$, for each leaf $L$, the differentiable (and coarse) quasi-isometry types of $g\|_L$ and its lift to $\widetilde{L}$ are independent of the choice of $g$; they depend only on ${{\mathcal{F}}}$ and $L$. On the other hand, ${{\mathcal{H}}}$ is compactly generated [@Haefliger1985], which can be seen as follows. There is some defining cocycle $(U'_i,p'_i,h'_{ij})$, with $p'_i:U'_i\to T'_i$, such that $\overline{U_i}\subset U'_i$, $T_i\subset T'_i$ and $p'_i$ extends $p_i$. Therefore each $h'_{ij}$ is an extension of $h_{ij}$ so that $\overline{{\operatorname{dom}}h_{ij}}\subset{\operatorname{dom}}h'_{ij}$. Moreover ${{\mathcal{H}}}$ is the restriction to $T$ of the pseudogroup ${{\mathcal{H}}}'$ on $T'=\bigsqcup_iT'_i$ generated by the maps $h'_{ij}$, and $T$ is a relatively compact open subset of $T'$ that meets all ${{\mathcal{H}}}'$-orbits. Let $E$ denote the collection of maps $h_{ij}$. According to Theorems \[t:quasi-isometric orbits\] and \[t:quasi-isometric germ covers of orbits\], by using the above $T'$ and ${{\mathcal{H}}}'$, it follows that the quasi-isometry type of the ${{\mathcal{H}}}$-orbits and their germ covers with $d_E$ are independent of the choice of $(U_i,p_i,h_{ij})$ under the above conditions; in fact, they are coarsely quasi-isometric to the corresponding leaves, and therefore they have the same growth type [@Carriere1988] (this is an easy consequence of the existence of a uniform bound of the diameter of the plaques). Similarly, the germ covers of the ${{\mathcal{H}}}$-orbits are also quasi-isometric to the holonomy covers of the corresponding leaves. Growth of homogeneous pseudogroups {#s:growth homog pseudogroups} ================================== Let $G$ be a Lie group, and $P\subset G$ a closed subgroup such that the $G$-action on $G/P$ is faithful; thus this action is also quasi-analytic by the analyticity of $G/P$. Let us define a [[(]{}homogeneous[)]{} $(G,G/P)$-pseudogroup]{} as a pseudogroup equivalent to the pseudogroup ${{\mathcal{H}}}$ generated by the action of some subgroup $\Gamma\subset G$ on some $\Gamma$-invariant open subset $T\subset G/P$. Suppose that ${{\mathcal{H}}}$ is compactly generated, and let ${{\mathcal{G}}}={{\mathcal{H}}}\|_U$ for some relatively compact open subset $U\subset T$ that meets all ${{\mathcal{H}}}$-orbits. For every $\gamma\in\Gamma$ with $\gamma\cdot U\cap U\ne\emptyset$, let $h_\gamma$ denote the restriction $U\cap\gamma^{-1}\cdot U\to\gamma\cdot U\cap U$ of the left translation $\gamma\cdot:G\to G$. There is a finite symmetric set $S=\{s_1,\dots,s_k\}\subset\Gamma$ such that $E=\{h_{s_1},\dots,h_{s_k}\}$ is a recurrent system of compact generation of ${{\mathcal{H}}}$ on $U$; in fact, by reducing $\Gamma$ if necessary, we can assume that $S$ generates $\Gamma$. For each $x\in U$, let $$\Gamma_{U,x}=\{\,\gamma\in\Gamma\mid \gamma\cdot x\in U\,\}\;.$$ Let ${{\mathfrak{G}}}$ denote the topological groupoid of germs of ${{\mathcal{G}}}$. Since the $G$-action on $G/P$ is faithful and quasi-analytic, and $G/P$ is connected, we get a bijection $\Gamma_{U,x}\to{{\mathfrak{G}}}_x$, $\gamma\mapsto{\operatorname{germ}}(h_\gamma,x)$. For $\gamma\in\Gamma_{U,x}$, let $\|\gamma\|_{S,U,x}:=\|h_\gamma\|_{E,x}$. Thus $\|1\|_{S,U,x}=0$, and, if $\gamma\ne1$, then $\|\gamma\|_{S,U,x}$ equals the minimum $n\in{{\mathbb N}}$ such that there are $i_1,\dots,i_n\in\{1,\dots,k\}$ with $\gamma=s_{i_n}\cdots s_{i_1}$ and $s_{i_m}\cdots s_{i_1}\cdot x\in U$ for all $m\in\{1,\dots,n\}$. Moreover $d_E$ on ${{\mathfrak{G}}}_x$ corresponds to the metric $d_{S,U,x}$ on $\Gamma_{U,x}$ given by $$d_{S,U,x}(\gamma,\delta)=\|\delta\gamma^{-1}\|_{S,U,\gamma(x)}\;.$$ Observe that, for all $\gamma\in\Gamma_{U,x}$ and $\delta\in\Gamma_{U,\gamma\cdot x}$, $$\begin{aligned} {2} \delta\gamma&\in\Gamma_{U,x}\;,&\quad \|\delta\gamma\|_{S,U,x}&\le\|\gamma\|_{S,U,x}+\|\delta\|_{S,U,\gamma\cdot x}\;,\label{delta gamma}\\ \gamma^{-1}&\in\Gamma_{U,\gamma\cdot x}\;,&\quad \|\gamma\|_{S,U,x}&=\|\gamma^{-1}\|_{S,U,\gamma\cdot x}\;.\label{gamma^-1} \end{aligned}$$ \[t:growth homog pseudogroup\] With the above notation and conditions, if $\overline\Gamma$ is connected, then - either all germ covers of the ${{\mathcal{G}}}$-orbits have polynomial growth with degree bounded by some common constant; - or all infinite germ covers of the ${{\mathcal{G}}}$-orbits have exponential growth. We can suppose that $\Gamma_{U,x}$ is infinite for all $x\in U$, otherwise $\Gamma=\{1\}$ because $\overline\Gamma$ is connected. Fix any open set $V$ that meets all orbits such that $\overline V\subset U$. \[cl:B\] For each finite subset $F\subset\Gamma$, we have $$U\subset\bigcup_{\gamma\in\Gamma\setminus F}\gamma\cdot V\;.$$ Suppose that Claim \[cl:B\] is false. Then there is some finite symmetric subset $F\subset\Gamma$ and some $x\in U$ such that $((\Gamma\setminus F)\cdot x)\cap V=\emptyset$. By the recurrence of $E$, there is some $N\in{{\mathbb N}}$ satisfying . Since $\Gamma_{U,x}$ is infinite, it follows that there is some $\gamma\in\Gamma_{U,x}\setminus F$ such that $$\label{\|gamma\|_S,U,x} \|\gamma\|_{S,U,x}>N+\max\{\,\|\epsilon\|_{S,U,x}\mid\epsilon\in F\cap\Gamma_{U,x}\,\}\;.$$ By , there is some $h\in E^N$ such that $$\gamma\cdot x\in h^{-1}(V\cap{\operatorname{im}}h)\;.$$ We have $h=h_{\delta}$ for some $\delta\in\Gamma$. Note that $\delta\in\Gamma_{U,\gamma\cdot x}$ and $\|\delta\|_{S,U,\gamma\cdot x}\le N$. Hence $$\|\gamma\|_{S,U,x}\le\|\delta\gamma\|_{S,U,x}+\|\delta^{-1}\|_{S,U,\delta\gamma\cdot x}=\|\delta\gamma\|_{S,U,x}+\|\delta\|_{S,U,\gamma\cdot x}\le\|\delta\gamma\|_{S,U,x}+N$$ by and , obtaining that $\delta\gamma\not\in F$ by . However $\delta\gamma\cdot x\in V$, obtaining a contradiction, which completes the proof of Claim \[cl:B\]. \[cl:overline B\] For each finite subset $F\subset\Gamma$, we have $$\overline U\subset\bigcup_{\gamma\in \Gamma\setminus F}\gamma\cdot V\;.$$ Take a relatively compact open subset $U'\subset T$ such that $\overline{U}\subset U'$, and let ${{\mathcal{G}}}'={{\mathcal{H}}}\|_{U'}$. Then Claim \[cl:overline B\] follows by applying Claim \[cl:B\] to ${{\mathcal{G}}}'$. According to Claim \[cl:overline B\], by increasing $S$ if necessary, we can suppose that $$\label{s_i} \overline{U}\subset\bigcup_{i<j}(s_i\cdot V\cap s_j\cdot V)=\bigcup_{i<j}(s_i^{-1}\cdot V\cap s_j^{-1}\cdot V)\;.$$ Assume first that $\Gamma$ is not nilpotent. Thus $\overline\Gamma$ is a non-nilpotent connected Lie group, and therefore we can apply [@BreuillardGelander2007 Proposition 10.5] to its finitely generated dense subgroup $\Gamma$, obtaining that there are elements $t_1,\dots,t_k$ in $\Gamma$, as close as desired to $s_1,\dots,s_k$, respectively, which are free generators of a free semi-group. By the compactness of $\overline U$, if $t_1,\dots,t_k$ are close enough to $s_1,\dots,s_k$, then gives $$\label{t_i} \overline U\subset\bigcup_{i<j}(t_i^{-1}\cdot V\cap t_j^{-1}\cdot V)\;.$$ Now, we adapt the argument of the proof of [@BreuillardGelander2007 Lemma 10.6]. Let $\Gamma'\subset\Gamma$ be the subgroup generated by $t_1,\dots,t_k$; thus $S'=\{t_1^{\pm1},\dots,t_k^{\pm1}\}$ is a symmetric set of generators of $\Gamma'$, and $S\cup S'$ is a symmetric set of generators of $\Gamma$. With $E'=\{h_{t_1}^{\pm1},\dots,h_{t_k}^{\pm1}\}$, observe that $E\cup E'$ is a recurrent system of compact generation of ${{\mathcal{H}}}$ on $U$. Given $x\in U$, let $\mathsf{S}(n)$ be the sphere with center the identity element and radius $n\in{{\mathbb N}}$ in $\Gamma'_{U,x}$ with $d_{S',U,x}$. By , for each $\gamma\in\mathsf{S}(n)$, we have $\gamma\cdot x\in t_i^{-1}\cdot V\cap t_j^{-1}\cdot V$ for some indices $i<j$. So the points $t_i\gamma\cdot x$ and $t_j\gamma\cdot x$ are in $V$, obtaining that $t_i\gamma,t_j\gamma\in\mathsf{S}(n+1)$. Moreover all elements obtained in this way from elements of $\mathsf{S}(n)$ are pairwise distinct because $S'$ freely generate a free semigroup. Hence ${\operatorname{card}}(\mathsf{S}(n+1))\ge2{\operatorname{card}}(\mathsf{S}(n))$, giving ${\operatorname{card}}(\mathsf{S}(n))\ge2^n$. So $\Gamma'_{U,x}$ has exponential growth with $d_{S',U,x}$. Since $\Gamma'_{U,x}\subset\Gamma_{U,x}$ and $d_{S\cup S',U,x}\le d_{S',U,x}$ on $\Gamma'_{U,x}$, it follows that $\Gamma_{U,x}$ also has exponential growth with $d_{S\cup S',U,x}$. So ${{\mathfrak{G}}}_x$ has exponential growth with $d_{E\cup E'}$, obtaining that ${{\mathfrak{G}}}_x$ has exponential growth with $d_E$ by Corollary \[c:growth orbits and their germ covers\]. If $\Gamma$ is nilpotent, then it has polynomial growth with respect to $d_S$, and this growth type dominates the growth type of $\Gamma_{U,x}$ with the metric $d_{S,U,x}$ because $d_S\le d_{S,U,x}$ on $\Gamma_{U,x}$. Now, according to [@Blumenthal1979 Theorem 1] (see Section \[s: intro\]), Theorem \[t:growth transv homog folns\] is a direct consequence of Theorem \[t:growth homog pseudogroup\] and the observations of Section \[ss:growth leaves\]. [10]{} J.A. Álvarez López and A. Candel, Generic coarse geometry of leaves, arXiv:1406.1746. [to3em]{}, Equicontinuous foliated spaces, Math. Z. 263 (2009), 725–774. J. Block and S. Weinberger, Large scale homology theories and geometry, Geometric Topology (Athens, GA, 1993) (RI), AMS/IP Stud. Adv. Math., no. 2.1, Amer. Math. Soc., 1997, pp. 522–569. R.A. Blumenthal, Transversely homogeneous foliations, Ann. Inst. Fourier (Grenoble) 29 (1979), 143–158. E. Breuillard and T. Gelander, A topological [Tits]{} alternative, Ann. of Math. 166 (2007), 427–474. A. Candel and L. Conlon, Foliations [I]{}, Graduate Studies in Mathematics, Amer. Math. Soc., Providence, RI, 2000. Y. Carrière, Feuilletages riemanniens à croissance polynomiale, Comment. Math. Helv. 63 (1988), 1–20. S. Chihi and S. ben Ramdane, On the [Godbillon-Vey]{} invariant and global holonomy of [$\mathbb{R}P^1$]{}-foliations, Balkan J. Geom. Appl. 13 (2008), 24–34. C. Godbillon, Feuilletages: Études géométriques, Progress in Math., vol. 98, Birkhäuser, Boston, Basel and Stuttgart, 1991. M. Gromov, Asymptotic invariants of infinite groups, Geometric Group Theory (Cambridge) (G.A. Niblo and M.A. Roller, eds.), vol. 2, Cambridge University Press, 1993. A. Haefliger, Pseudogroups of local isometries, Differential Geometry, Santiago de Compostela, 1984 (L.A. Cordero, ed.), Pitman, 1985, pp. 174–197. [to3em]{}, Leaf closures in [Riemannian]{} foliations, A Fête on Topology (New York), Academic Press, 1988, pp. 3–32. G. Hector and U. Hirsch, Introduction to the geometry of foliations, part [A]{}, Aspects of Mathematics, vol. E1, Friedr. Vieweg and Sohn, Braunschweig, 1981. [to3em]{}, Introduction to the geometry of foliations, part [B]{}, Aspects of Mathematics, vol. E3, Friedr. Vieweg and Sohn, Braunschweig, 1983. P. Molino, Riemannian foliations, Progress in Mathematics, no. 73, Birkh[ä]{}user Boston, Inc., Boston, MA, 1988. A. Phillips and D. Sullivan, Geometry of leaves, Topology 20 (1981), 209–218. J.F. Plante, A generalization of the [P]{}oincaré-[B]{}endixson theorem for foliations of codimension one, Topology 12 (1973), 177–181. [to3em]{}, Foliations with measure preserving holonomy, Ann. of Math. (2) 102 (1975), 327–361. P. Walczak, Dynamics of foliations, groups and pseudogroups, Instytut Matematyczny Polskiej Akademii Nauk., Monografie Matematyczne (New Series), no. 64, Birkh[ä]{}user Verlag, Basel, 2004. [^1]: Usually, growth types are defined by using non-decresing functions ${{\mathbb Z}}^+\to[0,\infty)$, but this gives rise to an equivalent concept (see [@PhillipsSullivan1981]).	{ "pile_set_name": "ArXiv" }
--- abstract: 'Motivated by practical needs such as large-scale learning, we study the impact of adaptivity constraints to linear contextual bandits, a central problem in online active learning. We consider two popular limited adaptivity models in literature: batch learning and rare policy switches. We show that, when the context vectors are adversarially chosen in $d$-dimensional linear contextual bandits, the learner needs $O(d \log d \log T)$ policy switches to achieve the minimax-optimal regret, and this is optimal up to $\mathrm{poly}(\log d, \log \log T)$ factors; for stochastic context vectors, even in the more restricted batch learning model, only $O(\log \log T)$ batches are needed to achieve the optimal regret. Together with the known results in literature, our results present a complete picture about the adaptivity constraints in linear contextual bandits. Along the way, we propose the distributional optimal design, a natural extension of the optimal experiment design, and provide a both statistically and computationally efficient learning algorithm for the problem, which may be of independent interest.' author: - 'Yufei Ruan[^1]' - 'Jiaqi Yang[^2]' - 'Yuan Zhou[^3]' bibliography: - 'library.bib' title: \| Linear Bandits with Limited Adaptivity and\ Learning Distributional Optimal Design --- Introduction ============ Online active learning is a fundamental research direction in machine learning where the learner conducts sequential interactions, once per time step, with the environment in order to learn the optimal policies and maximize the total reward. To achieve optimal learning performance, the learner must seek a balance between exploration and exploitation, which is usually done by adaptively selecting actions based on all historical observations. However, full adaptivity at a per-time-step scale significantly sacrifices parallelism and hinders the large-scale deployment of learning algorithms. To facilitate scalable learning, it is worthwhile to study the following question: > What is the minimum amount of adaptivity needed to achieve optimal performance in online active learning? In this paper, we address the above question through studying the impact of two popular types of adaptivity constraints to the linear contextual bandits, a central problem in online learning literature. We prove tight adaptivity-regret trade-offs for two natural settings of the problem. Along the way, we make a new connection to optimal experiment design: we propose the natural distributional optimal design problem, prove the existence of parametric forms for the optimal design, and present sample-efficient algorithms to learn the parameters. Our proposed framework contributes a novel learning component to the classical field of experiment design in statistics, and may be of independent interest. #### Linear Contextual Bandits. The linear contextual bandits (or linear bandits for short), also known as “associative reinforcement learning” [@abe1999associative; @auer2002using], are a generalization of the ordinary multi-armed bandits. While also encapsulating the fundamental dilemma of “exploration vs. exploitation” in online active learning, linear contextual bandits highlight the guidance of contextual information for decisions, enabling personalized treatments and recommendations in real-world applications such as clinical trial, recommendation systems, and advertisement selection. In a bandit game, there are $T$ time steps in total. At each time step $t \in [T]$, the learner has to make a decision among $K$ candidate actions (a.k.a. arms in bandit literature). While in ordinary multi-armed bandits, the mean rewards of the actions have to be completely independent from each other, linear bandits allow a linear model for the mean rewards. More specifically, at time step $t$, each action $i \in [K]$ is associated with a $d$-dimensional context vector $\vx_{ti}$ (a.k.a., the feature vector), and the context vectors are presented to the learner. The expected reward for the $i$-th action is $\vtheta^\top \vx_{ti}$, where $\vtheta \in \sR^d$ is hidden from the learner. The goal is to gradually learn $\vtheta$ and maximize the cumulative expected reward, or equivalently, minimize the expected regret (i.e., the difference between the received rewards and the rewards of the best actions in hindsight, as later defined in ). For example, in clinical trial, the candidate actions correspond to the $K$ involved treatments. At time step $t$, an individual patient arrives with the context vectors $\{\vx_{ti}\}_{i=1}^{k}$ characterizing his/her response to the candidate treatments, and the recovery probability given treatment $i$ is modeled by the linear function $\vtheta^\top \vx_{ti}$, which corresponds to the expected reward in linear bandits. There are two natural settings of the linear bandits: adversarial and stochastic contexts. The first setting is harder for the learner, as the context vectors are chosen by an adversary and the learner has to minimize the regret in the worst case. In the second setting, in contrast, the sets of context vectors are independently drawn from an unknown distribution $\gD$ (while correlation may still exist among the contexts during the same time step), and the learner aims at minimizing the expected regret over $\gD$. Note that in the clinical trial example, the individual patients can often be viewed as independent samples from the population which is characterized by $\gD$. #### Limited Adaptivity Models: Batch Learning and Rare Policy Switches. We consider two popular models of adaptivity constraints. The first model is batch learning, where the time steps are grouped into pre-defined batches. Within a batch, the same (possibly randomized) policy is used to select actions for all data and the rewards are observed only at the end of the batch. The amount of adaptivity is measured by the number of batches, which is expected to be as small as possible. A notable example is designing clinical trials, where each phase (batch) of the trial involves simultaneously applying medical treatments to a batch of patients. The outcomes are observed at the end of the phase, and may be used for designing experiments in future phases. Finding the correct number and sizes of the batches may achieve optimal efficiency for the trial by creating sufficient intra-batch parallelism while still providing sufficient adaptivity at the inter-batch scale. The other model is learning with rare policy switches, where the amount of adaptivity is measured by the number of times allowed for the learner to change the action-selection policy. For the same amount of adaptivity measure, this model can be viewed as a relaxation of the batch learning model, because the learner in the batch learning model can only change the policy at the pre-defined time steps. Both of the above models are closely connected to parallel learning, as we will discuss at the end of . We also note that another natural limited adaptivity model is “batch learning with adaptive grid” [@gao2019batched]. This model allows the learner to adaptively decide the size of a batch at the beginning of the batch, which is a more relaxed constraint than batch learning with pre-defined batches (a.k.a., the static grid model) but more restricted than the rare policy switch model, given the same amount of adaptivity measure.[^4] Simple arguments will show that the bounds for the adaptive grid model are the same as the static grid model in both linear bandit settings. Therefore, for succinct exposition, we omit further discussions about the adaptive grid model.[^5] #### Optimal Experiment Design. Optimal experiment design seeks to minimize the estimation variances of parameters via intelligently choosing queries to the given set of data points. Among the multiple optimization criteria, the one most related to linear bandits is the G-optimality criterion which seeks to minimize the maximum estimation variance among the given data points. More precisely, given a set of data points $X \subseteq \sR^d$ that spans the full dimension, the goal is to find a distribution $\gK$ supported on $X$, such that $\max_{\vx \in X} \vx^\top ({\mathop{\mathbb{E}}}_{\vy \sim \gK} \vy\vy^\top)^{-1} \vx$ is minimized. Here, $\info(\gK) = {\mathop{\mathbb{E}}}_{\vy \sim \gK} \vy\vy^\top$ is the information matrix of the design $\gK$, and $\vx^\top \info(\gK)^{-1} \vx$ is the variance of the estimate for data point $\vx$. The General Equivalence Theorem of @kiefer1960equivalence implies that there always exists a design $\gK$ such that $\max_{\vx \in X} \vx^\top \info(\gK)^{-1} \vx \leq d$ and such designs have been used for linear bandits with fixed candidate action set (see Chapter 22 of [@lattimore2020bandit], and [@esfandiari2019batched]). However, to the best of our knowledge, traditional optimal design does not address the problem when the candidate action set $X$ is stochastic. In this work, motivated by the algorithmic needs from batch linear bandits, we address this problem and develop a framework named distributional optimal design that runs at the core of our algorithm. We will introduce this framework in the next subsection. Our Contributions {#sec:our-contribution} ----------------- Adaptivity constraints in online active learning have attracted much attention recently. It has been shown that multi-armed bandits only need $O(\log \log T)$ batches to achieve asymptotically minimax-optimal regret [@perchet2016batched; @gao2019batched]. For linear contextual bandits with adversarial contexts, when $\ln K \geq \Omega(d)$, @abbasi2011improved showed an optimal-regret algorithm with $O(d \log T)$ policy switches. In contrast, for the batch model, @han2020sequential recently showed that as many as $\Omega(\sqrt{T})$ batches are needed to achieve the optimal regret bound, implying that batch learning is significantly more restrictive than policy switch constraints for adversarial contexts. In light of these partial results, quite a few questions are intriguing and remain to be explored – What makes the adaptivity requirements of linear contextual bandits fundamentally different from multi-armed bandits? What is the limitation for algorithms with rare policy switches, or in other words, can we extend the algorithm by [@abbasi2011improved] to the full parameter range of $K$, and further improve the number of policy switches to $O(\log \log T)$? Do linear bandits with stochastic contexts require substantially less adaptivity than the adversarial setting? We address these questions and summarize our answers as follows. : (Contribution \#1, informal statements of and ) For linear bandits with adversarial contexts, we show that $d \log T$ (up to $\mathrm{poly}(\log d, \log \log T)$ factors) is the tight amount of policy switches needed to achieve the minimax-optimal regret. To this end, we first extend the algorithm by [@abbasi2011improved] to the case where $\ln K \leq o(d)$. Our algorithm achieves the asymptotically minimax-optimal regret with $O(d \log d \log T)$ policy switches. We then prove that our algorithm and the one by [@abbasi2011improved] achieve the near-optimal policy switch vs. regret trade-off. In particular, $\Omega(d\log T / \log (d \log T))$ policy switches are needed to achieve any $\sqrt{T}$-type regret. : (Contribution \#2, an informal statement of ) For linear bandits with stochastic contexts, even in the more restricted batch learning model, it is possible to achieve the asymptotically minimax-optimal regret using only $O(\log \log T)$ batches. Our algorithm can be easily adapted to use $M$ batches and achieve $\sqrt{d\log K} T^{\frac{1}{2(1 - 2^{-M})}} \times \mathrm{poly}\log T$ regret, for any $M$. -- --------------------------------------------- ------------------------------------------------------------------- UB: $O(\sqrt{dT})$ [@han2020sequential] LB: $\Omega(\sqrt{T})$ [@han2020sequential] LB: $\Omega(\frac{d \log T}{\log (d \log T)})$ (by \[contrib:1\]) UB: $O(\log \log T)$ (by \[contrib:2\]) UB: $O(\log \log T)$ (implied by \[contrib:2\]) LB: $\Omega(\log \log T)$ [@gao2019batched] LB: $\Omega(\log \log T)$ [@simchi2019phase] -- --------------------------------------------- ------------------------------------------------------------------- : Amount of adaptivity needed in various models and settings for linear bandits. \[tab:results\] Together with the known results in literature, we are able to present an almost complete picture about the adaptivity constraints for linear bandits in . Most interestingly, compared to ordinary multi-armed bandits, linear bandits exhibit a richer set of adaptivity requirements, and strong separations among different models and settings. We also find that adversarially chosen context vectors are the main source of difficulty for reducing adaptivity requirements. #### Comparison of \[contrib:2\] and [@han2020sequential]. Compared to \[contrib:1\], our result in \[contrib:2\] requires substantially more technical effort and is also the main motivation for us to develop the framework of distributional optimal design (which will be elaborated soon). We note that @han2020sequential also studied batch learning for linear bandits with stochastic contexts and showed an algorithm with $O(\log \log T)$ batches. However, their results are for a special case of the problem with the following assumptions: the context vectors are drawn from a Gaussian distribution, the ratio between the maximum and minimum eigenvalues of the Gaussian co-variance matrix should be $O(1)$, and the number of candidate actions $K$ cannot be greater than a polynomial of $d$. The design and analysis of their algorithm crucially rely on these three assumptions and it seems not obvious that their result can be directly extended to the general context set distribution. Indeed, their algorithm can safely choose the action to maximize the estimated mean reward, thanks to the isotropic Gaussian assumption ensuring sufficient exploration towards other directions. In contrast, without these assumptions, much effort in our algorithm is spent on the careful design of the exploration policy using many candidate actions, which motivates the problem of distributional optimal design. #### Distributional Optimal Design. As mentioned above, to facilitate the algorithm for stochastic contexts, we have to extend the traditional experiment design results to the regime where the set $X$ of contexts/data points is stochastic. Suppose that $X$ follows the distribution $\gD$, the goal of our proposed distributional optimal design problem is to find a sample policy $\pi$ that maps any set $X$ to a probability distribution supported on $X$, so as to minimize the distributional G-variation, defined as ${\mathop{\mathbb{E}}}_{X \sim \gD} \max_{\vx \sim X} \vx^\top \info_{\gD}(\pi)^{-1} \vx$, where $\info_{\gD}(\pi) = {\mathop{\mathbb{E}}}_{X \sim \gD} {\mathop{\mathbb{E}}}_{\vy \sim \pi(X)} \vy \vy^\top$ is the information matrix of sample policy $\pi$ over $\gD$.[^6] Note that the traditional G-optimal design is the special case of our problem when $\gD$ is deterministic. The first natural question about our proposed problem is on the existence of a good sample policy. Regarding this, we prove the following result. : (Contribution \#3, an informal statement of ) For any $\gD$, there exists a sample policy $\pi$ such that the distributional G-variance is bounded by $O(d \log d)$.[^7] Moreover, we can construct such a policy from the class of so-called mixed-softmax policies, which admits a succinct description using $O(d^3 \log d)$ real-valued parameters. Since $\gD$ is not known beforehand in linear bandits, we have to learn a good sample policy $\pi$ via finite samples from $\gD$. Since even the input of $\pi$ lie in a continuous space with $dK$ dimensions, proving the existence of the succinct parametric form of $\pi$ in \[contrib:3\] is a good news to learning. However, we find that directly constructing a policy based on the uniform distribution over empirical samples does not generalize to the true distribution $\gD$. We will come up with a more careful learning procedure to achieve the following goal. : (Contribution \#4, an informal statement of ) For any $\gD$, we design an algorithm to learn a good mixed-softmax policy $\pi$ using only $\mathrm{poly}(d)$ independent samples from $\gD$.[^8] We remark that the introduction of the distribution $\gD$ brings a unique learning challenge to optimal experiment design. It is hopeful that our results and the future study on other criteria in distributional optimal design may lead to broader applications in machine learning and statistics. #### Implications for Collaborative and Concurrent Learning. The idea of letting multiple learning agents learn in parallel so as to save overall running time has been studied a lot recently in online active learning, which is also the main motivation of this study (as mentioned in the very beginning of the paper). Below we discuss the implications of our algorithmic results for a few parallel learning models. The first implication is for the collaborative learning with limited interaction model, which was recently studied for pure exploration (i.e., top arm(s) identification) in multi-armed bandits [@hillel2013distributed; @tao2019collaborative; @karpov2020collaborative]. In this model, there are $\mathfrak{K}$ learning agents, and the learning process is partitioned into rounds of pre-defined time intervals. During each round (which is also referred to as the communication round), each of the $\mathfrak{K}$ agents learns individually like in the centralized model – image that there is a global buffer of the context vectors, and the agents repeatedly draw a set of context vectors from the buffer and make corresponding decisions. Each play of an arm takes one time step, and the agents may choose to skip a few time steps without playing. The agents can only communicate at the end of each round. The collective regret is defined to be the sum of the regret incurred by each agent. Suppose there are $T$ sets of context vectors in the global buffer, the goal is to finish the game in $O(\lceil T/\mathfrak{K}\rceil)$ time (i.e., achieving the full speedup), while minimizing the collective regret and the number of communication rounds $R$. Observe that a batch learning algorithm with $M$ batches can be easily transformed to a collaborative algorithm with $R = M$ communication rounds, where in each round $i$, each agent uses the policy for the $i$-th batch to play for $\lfloor \gT_i/\mathfrak{K}\rfloor$ or $\lceil \gT_i/\mathfrak{K}\rceil$ times, where $\gT_i$ is the size of the $i$-th batch. The total running time for collaborative learning is at most $T /\mathfrak{K} + M$, achieving the full speedup when $M \cdot \mathfrak{K} \leq O(T)$. Therefore, when $\mathfrak{K} \leq O(T/\log\log T)$, our algorithmic result \[contrib:2\] implies a collaborative algorithm for stochastic-context linear bandits with full speedup and minimax-optimal collective regret, using only $O(\log \log T)$ communication rounds. The second implication is for the concurrent learning model which was recently studied in [@guo2015concurrent; @bai2019provably; @zhang2020almost]. In this model, there is no limit on the number of communication rounds and the $\mathfrak{K}$ learning agents may communicate at the end of every time step. By a simple reduction described in [@bai2019provably], any algorithm with at most $M$ policy switches can be transformed to a $\mathfrak{K}$-agent concurrent learning algorithm with full speedup, and the collective regret is at most $M \cdot \mathfrak{K}$ plus the original regret bound. Therefore, our algorithmic result in \[contrib:1\] implies a concurrent learning algorithm for adversarial-context linear bandits with full speedup and minimax-optimal collective regret, as long as $\mathfrak{K} \leq O(\sqrt{(T \log K)/d})$. Additional Related Works ------------------------ The linear contextual bandit problem is a central question in online active learning, and its regret minimization task has been studied during the past decades [@auer2002using; @abe2003reinforcement; @dani2008stochastic; @rusmevichientong2010linearly; @chu2011contextual; @abbasi2011improved; @li2019nearly]. The minimax-optimal regret is proved to be $\sqrt{d T \min\{\log K, d\}}$ up to $\mathrm{poly} \log T$ factors, which is also the target regret for our algorithms with limited adaptivity. When the candidate action set is fixed, the task of identifying the best action has also been studied [@soare2014best; @tao2018best; @xu2018fully], and many of these works borrow the idea of G-optimal design. Batch regret minimization for multi-armed bandits was introduced by @perchet2016batched with 2 arms, and the $K$-arm general setting was recently studied by @gao2019batched. @simchi2019phase studied the $K$-arm setting with the rare policy switch constraint and achieved comparable results. For batch linear bandits, @esfandiari2019batched and @han2020sequential recently studied the problem with aforementioned additional assumptions. We note that another usage of batch learning (mainly in reinforcement learning) refers to learning from a fixed set of a priori-known samples with no adaptivity allowed, which is very different from the definition in our work. For the rare policy switch model, [@abbasi2011improved] showed a rarely switching algorithm for linear bandits. Rare policy switch constraints have also been studied for a broader class of online active learning problems, such as multinomial logit bandits [@dong2020multi] and Q-learning [@bai2019provably]. Under the broader definition of adaptivity constraints including batch learning and learning with low switching cost (which might not exactly align with the models defined in this work), many other online learning problems are studied, such as adversarial multi-armed bandits [@cesa2013online; @dekel2014bandits], the best (multiple-)arm identification problem [@jun2016top; @agarwal2017learning], and convex optimization [@duchi2018minimax]. The optimal design of experiments is a fundamental problem in statistics, with various optimality criteria proposed and many statistical models studied (see, e.g., [@pukelsheim2006optimal; @atkinson2007optimum]). When the sample budget is finite, finding the exact solutions to certain optimality criteria is NP-Hard [@welch1982algorithmic; @ccivril2009selecting; @summa2014largest], thus a sequence of recent works have studied approximation algorithms for the problem [@wang2017computationally; @singh2018approximate; @madan2019combinatorial; @nikolov2019proportional; @allen2020near]. However, to the best of our knowledge, all previous works have considered the fixed set of all possible experiments. In contrast, we propose and study the distributional optimal design problem where the set of candidate experiments might be stochastic. Technical Overview ================== Batch Algorithms for Stochastic Contexts ---------------------------------------- As the main technical contribution, we first describe the techniques developed in , and for proving our algorithmic result \[contrib:2\]. Along the way, the proof techniques for \[contrib:3\] and \[contrib:4\] are also explained. In , we combine all these technical components and prove the main theorem. #### The Batch Elimination Framework. All our algorithms are elimination-based: at each time step, the confidence intervals are estimated for each candidate action, and the actions whose confidence intervals completely fall below those of other actions are eliminated. All survived actions are likely to be the optimal one, and the learner has to design an intelligent sample policy $\pi$ to select the action from the survived set. In such a way, the incurred regret can be bounded by the order of the length of the longest confidence interval in the survived set. We note that this elimination-based approach is not new: it is adopted by the batch algorithms for multi-armed bandit (e.g., [@gao2019batched]) as well as the recent batch algorithm for linear bandits with fixed action set [@esfandiari2019batched]. However, thanks to the simple structures of the two problems, during each batch, both of their algorithms are able to construct confidence intervals for survived actions with a uniform length, so that the regret can be relatively more easily bounded. Indeed, although the algorithm by @han2020sequential does not explicitly eliminate actions, their analysis relies on the uniform estimation confidence for the actions (which requires the isotropic Gaussian assumption for context vectors). In contrast, we have to deal with confidence intervals with wildly different lengths because of the inherent non-uniformity of the probability mass assigned to each context direction in the general distribution $\gD$. To deal with such non-uniformity, in , we provide an analysis framework to relate the regret bound to the distributional G-variation of $\pi$ over $\gD$, as introduced in . In particular, we show that if we let $\pi(X) = \gopt(X)$, which returns the G-optimal design of the input context set $X$ (regardless of $\gD$), its distributional G-variation can be bounded by $d^2$ (for all $\gD$), leading to $O(d \sqrt{T \log K}) \times \mathrm{poly} \log T$ regret with $O(\log \log T)$ batches. This regret is $\sqrt{d}$ times greater than the minimax-optimal target. To achieve optimality, we need to improve the distributional G-variation to $O(d)$ (up to logarithmic factors), which requires to optimize $\pi$ specifically according to $\gD$. #### Existence of Distributional Optimal Design and its Parametric Form. In , we show that, given $\gD$, there exists a sample policy $\pi$ whose distributional G-variation is $O(d \log d)$. Our proof is constructive and the algorithm involves an innovative application of the rarely switching linear bandit algorithm [@abbasi2011improved]. We consider a long enough sequence of independent samples from $\gD$: $X_1, X_2, \dots, X_N$, and sequentially feed the context vector sets to the rarely switching algorithm. Instead of minimizing the regret (as the reward is undefined), the rarely switching algorithm selects the context vector $\vx$ that maximizes the variance according to the delayed information matrix, and update the total information matrix by adding $\vx \vx^\top$ to it. Borrowing the regret analysis techniques in linear bandits literature, and together with an adapted form of the celebrated Elliptical Potential Lemma, we are able to prove that, with the proper configuration of the initial information matrix, the average maximum confidence interval length throughout the $N$ time steps is $O(d\log d)$. Moreover, the rarely switching trick makes sure that the delayed information matrix switches for at most $O(d \log d)$ times. This allows us to extract $O(d \log d)$ (deterministic) sample policies $\{\pi_j\}$ from the execution trajectory of the algorithm, each of which chooses the variance maximizer according to a delayed information matrix in the trajectory. We also associate each $\pi_j$ with a probability mass $p_j$, which is proportional to the number of time steps when the corresponding delayed information matrix is used in the trajectory. We can then construct a so-called mixed-argmax policy $\pi$ as follows: with probability $1/2$, $\pi$ acts the same as $\gopt$; otherwise, $\pi$ acts the same as $\pi_j$ with probability $p_j$. We are then able to prove that the distributional G-variance of $\pi$ over $\gD$ is $O(d \log d)$. This is done mainly by showing that $\info_{\gD}(\pi)$ is comparable to the final information matrix in the trajectory, so that the distributional G-variance of $\pi$ can be bounded by the empirical average of the maximum confidence interval lengths. To lower bound $\info_\gD(\pi)$ using the total information matrix in the trajectory, while the portion corresponding to the larger switching window (i.e., greater $p_j$) in the trajectory can be directly compared, the smaller switching window will be handled by the $\gopt$ component in $\pi$. We note that the $\gopt$ component is also crucial to configuring the “proper” initial information matrix in the rarely switching algorithm. We finally observe that $\pi$ can be characterized by $O(d^3 \log d)$ parameters, because each $\pi_j$ is parameterized by a $d \times d$ information matrix. Since the $\argmax$ operator could be very sensitive to noise when the top input elements are close, to facilitate learning, we will also work on the mixed-softmax policy where each $\pi_j$ uses the $\softmax$ operator instead. #### [CoreLearning]{} for Distributional Optimal Design. It is tempting to build the natural learning algorithm that computes the distributional optimal design from the empirical samples, hope that the Lipschitz-continuity of the softmax policies provides a small covering of the policy space, which leads to uniform concentration results, and finally prove that the learned policy generalizes to the true distribution $\gD$. However, in , we construct an example to show that such an approach requires much higher sample complexity than we can afford. To enable sample-efficient learning, we propose a new algorithm, [CoreLearning]{}, that first identifies a core set, which is a subset of the empirical samples, and then computes a mixed-softmax policy from the core. To identify the core, we develop a novel procedure to iteratively prune away the sets that contain less explored directions among the empirical samples, so that the set of the remaining samples at the end of the procedure becomes the core. Via a volumetric argument, we show that the directions in the core can be sufficient explored even if only using the sets in the core, and the core is still overwhelmingly large. Both properties are crucially used in the [CoreLearning]{} algorithm. The high-level idea behind [CoreLearning]{} is that, on one hand, we can prove fast uniform concentration for the information matrix if all directions are sufficiently explored, so that the directions spanned by the core can be handled. On the other hand, the directions not included in the core are infrequent in $\gD$ (because the core is large enough), and can be dealt with by the $\gopt$ component in the mixed-softmax policy. Much technical effort is devoted to the analysis of [CoreLearning]{} because (1) it seems not quite obvious whether a core with the desired properties even exists, and (2) a careful analysis is needed when combining the analysis for sufficiently explored directions and infrequent directions, since the (possible) directions of the context vectors are continuous, and the boundary between the two types of directions may not be always clear. Please refer to for more detailed explanation. Policy Switch Bounds for Adversarial Contexts --------------------------------------------- #### The Algorithm with Rare Policy Switches. We first recall that @abbasi2011improved proposed a determinant-based doubling trick that only updates the policy when the determinant of the associated information matrix doubles. When applying to the [OFUL]{} algorithm in [@abbasi2011improved], the doubling trick leads to $O(d \log T)$ policy switches. However, due to technical difficulty, the state-of-the-art analysis for the [OFUL]{} algorithm shows the asymptotically minimax-optimal regret only for $\ln K \geq \Omega(d)$. While it is still an open question whether a simple adaptation of [OFUL]{} (such as [LinUCB]{} proposed in [@chu2011contextual]) also achieves the asymptotically minimax-optimal regret for $\ln K \leq o(d)$, the only known technique in literature to achieve the optimality is via building a more sophisticated “super algorithm” based on the idea of [LinUCB]{} (e.g., [SupLinUCB]{} [@chu2011contextual] and [SupLinRel]{} [@auer2002using]). There are $\Theta(\log T)$ information matrices maintained in these super algorithms, and therefore a naïve application of the determinant-based doubling trick to these super algorithms leads to $O(d \log^2 T)$ policy switches. To improve the number of policy switches, we adopt a simple combination of [OFUL]{} and [SupLinUCB]{}, so that only $O(\log d)$ information matrices are maintained, leading to $O(d \log d \log T)$ policy switches. Please refer to for more detailed explanation. #### The Lower Bound. In , we prove that to achieve any $\sqrt{T}$-type regret in the adversarial context setting, the algorithm has to switch the policy for at least $\Omega(d \log T / \log (d \log T))$ times. We first observe that the classical hard instances for linear bandits in [@dani2008stochastic] cannot work for our goal since their context vector set does not change with time and therefore their instances can be solved by our algorithm for stochastic contexts using $O(\log \log T)$ batches. Instead, we divide the $T$ time steps into stages of consecutive time periods, and design different context vectors for different stages. We will design a class of specially structured hard instances, where the hidden vector $\vtheta$ delicately matches with the context vectors in each instance. We then lower bound the average-case regret over the class for any rarely switching learner, which implies the worst-case regret lower bound. At a higher level, our construction is more similar to the recent work by @li2019nearly. However, the difference is that, in their construction, the regret that can be incurred by the worst learner is no more than $\sqrt{T}$ (up to polynomial factors in $d$ and $\log T$). In contrast, in our task, we need to show that the learner could easily incur $T^{1/2 + \Omega(1)}$ regret when using few policy switches. To achieve this, we need to design a class of hidden vectors $\vtheta$ and context vectors $\{\vx_{ti}\}$ so that the mean rewards of the candidate actions are much more separated from each other, while we still have to make sure that a rarely switching learner cannot learn enough information. Preliminaries ============= #### Notations. Throughout the paper, we denote $[N] {\stackrel{\mathrm{def}}{=}}\{1, 2, \dots, N\}$ for any integer $N$. We define $\log x {\stackrel{\mathrm{def}}{=}}\log_2 x$ and $\ln x {\stackrel{\mathrm{def}}{=}}\log_e x$. We use $\ind[\cdot]$ to denote the indicator variable for a given event (i.e., the value of the variable is $1$ if the event happens, and $0$ otherwise). We use $\norm{\cdot}$ to denote the 2-norm of matrices and vectors. Matrix and vector variables are displayed in bold letters. For any discrete set $X$, we use $\triangle_X$ to denote the set of all probability distributions supported on $X$. #### Linear Contextual Bandits. There is a hidden vector $\vtheta$ ($\norm{\vtheta} \leq 1$). For a given time horizon $T$, the context vectors $\{\{\vx_{ti}\}_{i=1}^K\}_{t=1}^T$ are drawn from the product distribution $\mathcal{D}_1 \otimes \mathcal{D}_2 \otimes \dots \otimes \mathcal{D}_T$, where $\mathcal{D}_t$ is the distribution for the context vectors at time step $t$. We assume $\norm{\vx_{ti}} \leq 1$ for all $i$ and $t$ almost surely. Before the game starts, the learner only knows $T$. At each time step of the game $t = 1, 2,\dots, T$, the learner has to first decide a policy $\chi_t$ that maps any set of context vectors $X$ to a distribution in $\triangle_X$. The learner then observes $X_t = \{\vx_{ti}\}_{i=1}^K$, samples an action $i_t$ from $\chi_t(X_t)$,[^9] plays arm $i_t$, and finally receives the reward $r_t = \vtheta^\top \vx_{t, i_t} + \varepsilon_{t}$, where $\varepsilon_{t}$ is an independent sub-Gaussian noise with variance proxy at most $1$. The goal of the learner is to minimize the expected regret $$\begin{aligned} R^T {\stackrel{\mathrm{def}}{=}}{\mathop{\mathbb{E}}}\left[\sum_{t = 1}^T \max_{i \in [K]} \vx_{ti}^\top \vtheta - \vx_{t,i_t}^\top \vtheta\right], \label{eq:regret}\end{aligned}$$ where the expectation is taken over $\mathcal{D}_1 \otimes \mathcal{D}_2 \otimes \dots \otimes \mathcal{D}_T$, the noises, and the internal randomness of the learner. In our algorithmic results, we also prove $(1-\delta)$-high probability expected regret, which is defined as $\sup_{A} {\mathop{\mathbb{E}}}\left[\ind[A] \cdot \sum_{t = 1}^T \max_{i \in [K]} \vx_{ti}^\top \vtheta - \vx_{t,i_t}^\top \vtheta\right]$ where the supremum is taken over all events $A$ such that $\Pr[A] \geq 1 - \delta$. In this definition, setting $\delta = O(1/T)$ recovers the usual expected regret up to an additive error of $O(1)$. #### Settings of Adversarial and Stochastic Contexts. In the setting of adversarial contexts, there are no additional constraints for the distributions $\{\gD_t\}$. Note that this corresponds to the oblivious adversary in bandit literature, meaning that the adversary has to choose all context vectors beforehand. In contrast, the stronger non-oblivious adversary may adaptively choose context vectors for any time step according to all game history before that time. Since we only prove lower bounds for the adversarial context setting in this work, dealing with a weaker adversary actually means a stronger lower bound result. In the setting of stochastic contexts, we have the additional assumption that $\gD = \gD_1 = \dots = \gD_T$. However, correlation may still exist among the contexts at the same time step. #### Models for Limited Adaptivity. In the batch learning model, the learner has to first choose a grid $\gT = \{\gT_0, \gT_1, \dots, \gT_M\}$ where $1 = \gT_0 < \gT_1 < \gT_2 < \cdots < \gT_{M-1} < \gT_M = T$. For any $i \in [M]$, the $i$-th batch consists of the time steps $t = \gT_{i-1} + 1, \dots, \gT_i$. During the $i$-th batch, the learner must choose the policy $\chi^{(i)}$ at the beginning of the batch, and the same policy will be used throughout the batch. The amount of adaptivity is measured by $M$, the number of batches. In the rare policy switch model, a policy switch occurs at time step $t > 1$ if $\chi_t \neq \chi_{t-1}$, and there is always a policy switch at time step $1$. The amount of adaptivity is measured by the number of policy switches. As mentioned before, the goal for the learner is to achieve the target minimax-optimal regret $\sqrt{d T \min\{\log K, d\}}$ (up to $\mathrm{poly}\log T$ factors) with as little adaptivity (measured in each specific model) as possible. We also remark that the rare policy switch model is a relaxation of the batch learning model, because the learner can decide whether to change the policy at any time step. Therefore, the amount of adaptivity needed in the rare policy switch model is always less than or equal to the batch model. Batch Elimination Framework and the G-Optimal Design {#sec:blinucb} ==================================================== $M = \lceil \log \log T \rceil, \alpha \gets 10\sqrt{\ln \frac{2dKT}{\delta}}$, $\gT = \{\gT_1, \gT_2, \dots, \gT_M\}, \gT_0 = 0, \gT_M = T, \forall i \in [M - 1]: \gT_i = T^{1 - 2^{-i}}$ As a warm-up, in this section, we first present [[BatchLinUCB]{}]{}() to illustrate the batch elimination framework for the linear bandit problem with stochastic contexts. Later in , we will introduce the G-optimal experiment design and show how it helps to reduce the regret bound of the algorithm. While the regret bound in is improved, it still has an extra $\sqrt{d}$ factor compared to the optimal minimax regret bound (without adaptivity constraints). The quest for optimal regret will be addressed in the later sections. We now introduce our first algorithm. [[BatchLinUCB]{}]{}() uses $M = O(\log \log T)$ batches and a pre-defined static grid $\gT = \{\gT_1, \gT_2, \dots, \gT_M\}$. For each batch $k$, [[BatchLinUCB]{}]{}keeps an estimate $\hat{\vtheta}_k$ for the hidden vector $\vtheta$, which is learned using the samples obtained in the batch. To decide an arm during any time $t$ in the $k$-th batch, the algorithm first performs an elimination procedure that is based on the estimate $\hat{\vtheta}_{\kappa}$ and the corresponding confidence region for each previous batch $\kappa \in \{1, 2, \dots, k-1\}$. Let $A_t$ be the set of survived arms after the elimination. The algorithm then plays a uniformly random arm from $A_t$. The following theorem upper bounds the regret of [[BatchLinUCB]{}]{}. \[thm:BatchedLinUCB\] With probability at least $(1 - \delta)$, the expected regret of [[BatchLinUCB]{}]{}is $$\begin{aligned} R^T_{\text{{{\sc BatchLinUCB}\xspace}}} \leq O(\sqrt{dKT \log (dKT/\delta)} \times \log \log T).\end{aligned}$$ To prove , we first introduce the following lemma that constructs the confidence intervals of the estimated rewards. \[lem:conf\] Fix any batch $k$, for each time step $t$ in batch $k$, with probability at least $(1 - \delta/T^2)$, for all $\kappa \in \{1, 2, \dots, k-1\}$ and all $i \in A_t$, we have that $$\begin{aligned} \abs{\vx_{ti}^\top \hat\vtheta_{\kappa} - \vx_{ti}^\top \vtheta} \le \omega_{ti}^{(\kappa)}.\end{aligned}$$ The proof of can be found in many papers in linear bandit literature (e.g., [@chu2011contextual; @li2019nearly]), and is included in for completeness. We now start proving . Fix any batch $k$ such that $k \geq 2$, when conditioned on the first $(k-1)$ batches, we let $\gD_k$ be the distribution of the survived candidate arms $X = \{\vx_{ti} : i \in A_t\}$ at any time $t$ during the $k$-th batch. We also let $\gD_0 = \{\vx_{ti}\}$ be the distribution of all candidate arms at any time $t$. Suppose that the desired event in happens for every time step during the $k$-th batch (which happens with probability at least $(1 - \delta \gT_k/T^2)$ by a union bound), it is straightforward to verify that for each time $t$ during the $k$-th batch, the optimal arm is not eliminated by the elimination procedure ( to ) in [[BatchLinUCB]{}]{}. In other words, we have that $i^_t {\stackrel{\mathrm{def}}{=}}\argmax_{i\in [K]} \vx_{ti}^\top \vtheta \in A_t$ for each time step $t$ in the $k$-th batch. Therefore, we can now upper bound the expected regret incurred during batch $k$ as $$\begin{aligned} R_k & = {\mathop{\mathbb{E}}}\sum_{t \text{~in batch $k$}} (\max_{i \in [K]} \vx_{ti}^\top \vtheta - \vx_{t, i_t}^\top \vtheta) \leq {\mathop{\mathbb{E}}}\sum_{t \text{~in batch $k$}} (\vx_{t,i^_t}^\top \hat{\vtheta}_{k-1} - \vx_{t,i_t}^\top \hat{\vtheta}_{k-1} + \omega_{t, i^_t}^{(k-1)} + \omega_{t, i_t}^{(k-1)}) \label{eq:thm-blinucb-10}\\ & \leq {\mathop{\mathbb{E}}}\sum_{t \text{~in batch $k$}} 2 \cdot (\omega_{t, i^_t}^{(k-1)} + \omega_{t, i_t}^{(k-1)}) \leq 4 {\mathop{\mathbb{E}}}\sum_{t \text{~in batch $k$}} \max_{i \in A_t} \omega_{ti}^{(k-1)} , \label{eq:thm-blinucb-20}\end{aligned}$$ where is due to the successful events of , the both inequalities in are due to the elimination process and that $i^_t \in A_t$. By the definition of $\omega_{ti}^{(k-1)}$ and the definition of $\gD_k$, we further have that $$\begin{aligned} R_k\leq 4\alpha {\mathop{\mathbb{E}}}\sum_{t \text{~in batch $k$}} \max_{i \in A_t} \sqrt{\vx_{ti}^\top \mLambda_{k-1}^{-1} \vx_{ti}} \leq 4\alpha \times \sum_{t \text{~in batch $k$}} {\mathop{\mathbb{E}}}_{X \sim \gD_{k}}\max_{\vx \in X} \sqrt{\vx^\top \mLambda_{k-1}^{-1} \vx}. \label{eq:thm-blinucb-30}\end{aligned}$$ We finally observe that $X \sim \gD_k$ can be sampled by drawing an $X' \sim \gD_{k-1}$ and performing an elimination process using $\hat\vtheta_{k-1}$ as well as the corresponding confidence region for $X'$. We note that $X \subseteq X'$. Therefore, continuing with , we have that $$\begin{aligned} R_k \leq 4\alpha \times \sum_{t \text{~in batch $k$}} {\mathop{\mathbb{E}}}_{X \sim \gD_{k-1}}\max_{\vx \in X} \sqrt{\vx^\top \mLambda_{k-1}^{-1} \vx} = 4\alpha \gT_k \times {\mathop{\mathbb{E}}}_{X \sim \gD_{k-1}}\max_{\vx \in X} \sqrt{\vx^\top \mLambda_{k-1}^{-1} \vx}. \label{eq:thm-blinucb-40}\end{aligned}$$ Now the goal is to upper bound ${\mathop{\mathbb{E}}}_{X \sim \gD_{k-1}}\max_{\vx \in X} \sqrt{\vx^\top \mLambda_{k-1}^{-1} \vx}$. The following lemma is a direct application of in . \[lem:exploration\] For each batch $k$ ($k < M$), with probability $(1 - \delta/T^2)$, we have that $$\begin{aligned} \mLambda_k \succcurlyeq \frac{\gT_{k}}{16} \left(\frac{ \ln T}{ \gT_{k}} \mI + {\mathop{\mathbb{E}}}_{X \sim \gD_{k-1}} {\mathop{\mathbb{E}}}_{\vx \sim \Unif(X)}[\vx \vx^\top]\right). \label{eq:lem-exploration}\end{aligned}$$ Assuming that holds for batch $(k-1)$, we have that $$\begin{aligned} & \quad {\mathop{\mathbb{E}}}_{X \sim \gD_{k-1}}\max_{\vx \in X} \sqrt{\vx^\top \mLambda_{k-1}^{-1} \vx} \leq {\mathop{\mathbb{E}}}_{X \sim \gD_{k-1}}\sum_{\vx \in X} \sqrt{\vx^\top \mLambda_{k-1}^{-1} \vx} \\ &\leq \frac{4}{\sqrt{\gT_{k-1}}} \sqrt{ {\mathop{\mathbb{E}}}_{X \sim \gD_{k-1}}\sum_{\vx \in X}\vx^\top \left(\frac{ \ln T}{ \gT_{k}} \cdot \mI + {\mathop{\mathbb{E}}}_{Y \sim \gD_{k-1}} \|Y\|^{-1} \cdot \sum_{\vy \in Y} \vy \vy^\top\right)^{-1} \vx} \\ &\leq \frac{4}{\sqrt{\gT_{k-1}}}\sqrt{ \Tr \left( \left(\frac{ \ln T}{ \gT_{k}} \cdot \mI + {\mathop{\mathbb{E}}}_{X \sim \gD_{k-1}} K^{-1} \cdot \sum_{\vy \in Y} \vy \vy^\top\right)^{-1} {\mathop{\mathbb{E}}}_{X \sim \gD_{k-1}}\sum_{\vx \in X} \vx \vx^\top \right)} \\ &\leq 4 \sqrt{dK/\gT_{k-1}}.\end{aligned}$$ Together with , and collecting the probabilities, we have that with probability at least $(1 - \delta \gT_k/T^2 - \delta/T^2)$, the expected regret incurred during batch $k$ ($k \geq 2$) is $$\begin{aligned} R_k \leq 16\alpha \gT_k \cdot \sqrt{dK/\gT_{k-1}} \leq 16\alpha \sqrt{dKT}. \label{eq:thm-blinucb-100}\end{aligned}$$ Note that also holds for $k=1$ almost surely, because $\gT_1 \leq \sqrt{dT}$ and the maximum regret incurred per time step is at most $1$. Finally, summing up the expected regret incurred across all batches and collecting the probabilities, we have that, with probability at least $(1 - \delta)$, the expected regret is bounded by $$\begin{aligned} R^T \leq M \times 16\alpha \sqrt{dKT} = O(\sqrt{dKT \log (dKT/\delta)} \times \log \log T). \notag \end{aligned}$$ This concludes the proof of . Improved Regret via the G-Optimal Design {#sec:blinucbkw} ---------------------------------------- In this subsection, we show how a simple application of the G-optimal design can help to replace the $K$ factor in by (the usually smaller quantity) $d$. To achieve this, we first introduce the following lemma on G-optimal design, which is a direct corollary of the General Equivalence Theorem of @kiefer1960equivalence. For any subset $X \subseteq \R^d$, there exists a distribution $\gK_X$ supported on $X$, such that for any $\varepsilon > 0$, it holds that $$\begin{aligned} \max_{\vx \in X} \vx^\top \left(\varepsilon \mI + {\mathop{\mathbb{E}}}_{\vy \sim \gK_X} \vy \vy^\top \right)^{-1} \vx \leq d . \label{eq:thm-KW}\end{aligned}$$ Furthermore, if $X$ has a finite support, one can find a distribution such that the right-hand side of is relaxed to $2d$ in time $\mathrm{poly}(\abs{\supp(X)})$. We now describe the new [[BatchLinUCB-KW]{}]{}algorithm. It is almost the same as [[BatchLinUCB]{}]{}, while the only difference is that at of , letting $X = \{\vx_{t i} : i \in A_t\}$, we compute a distribution $\gK_X$ satisfying (up to the factor $2$ relaxation) and randomly select the action $$\begin{aligned} i_t \sim \gopt(X) {\stackrel{\mathrm{def}}{=}}\gK_X. \label{eq:def-gopt}\end{aligned}$$ For completeness, a full description of [[BatchLinUCB-KW]{}]{}is provided in . We now prove the expected regret of [[BatchLinUCB-KW]{}]{}as follows. \[thm:BatchedLinUCBKW\] With probability at least $(1-\delta)$, the expected regret of [[BatchLinUCB-KW]{}]{}is $$\begin{aligned} R^T_{\text{{{\sc BatchLinUCB-KW}\xspace}}} \leq O(d\sqrt{T \log (dKT/\delta)} \times \log \log T). \notag\end{aligned}$$ We now prove . Note that the analysis for [[BatchLinUCB]{}]{}also applies to [[BatchLinUCB-KW]{}]{}up to . Thus, we will focus on bounding ${\mathop{\mathbb{E}}}_{X \sim \gD_{k-1}}\max_{\vx \in X} \sqrt{\vx^\top \mLambda_{k-1}^{-1} \vx}$ while keeping in mind that $\mLambda_{k-1}^{-1}$ is a different quantity due to $\gopt$. Similarly to , for each For each batch $k$ ($k < M$), with probability $(1 - \delta/T^2)$, we have that $$\begin{aligned} \mLambda_k \succcurlyeq \frac{\gT_{k}}{16} \left(\frac{ \ln T}{ \gT_{k}} \mI + {\mathop{\mathbb{E}}}_{X \sim \gD_{k-1}} {\mathop{\mathbb{E}}}_{\vx \sim \gopt(X)}[\vx \vx^\top]\right). \label{eq:lem-exploration-KW}\end{aligned}$$ Assuming that holds for batch $(k-1)$, letting $\vx^(X) = \argmax_{\vx \in X} \vx^\top \mLambda_{k-1}^{-1} \vx$, we have that $$\begin{aligned} & {\mathop{\mathbb{E}}}_{X \sim \gD_{k-1}}\max_{\vx \in X} \sqrt{\vx^\top \mLambda_{k-1}^{-1} \vx} = {\mathop{\mathbb{E}}}_{X \sim \gD_{k-1}} \sqrt{(\vx^(X))^\top \mLambda_{k-1}^{-1} \vx^(X)} \nonumber \\ &\qquad \qquad \leq \sqrt{{\mathop{\mathbb{E}}}_{X \sim \gD_{k-1}} (\vx^(X))^\top \mLambda_{k-1}^{-1} \vx^(X)} = \sqrt{\Tr(\mLambda_{k-1}^{-1} {\mathop{\mathbb{E}}}_{X \sim \gD_{k-1}} \vx^(X) (\vx^(X))^\top)}, \label{eq:blinucbkw-1800}\end{aligned}$$ where the inequality is Jensen. By and (up to the factor $2$ relaxation), we have that $$\begin{aligned} \vx^(X) (\vx^(X))^\top \preccurlyeq 2d \times {\mathop{\mathbb{E}}}_{\vy \in \gopt(X)} \vy\vy^\top . \label{eq:blinucbkw-1900}\end{aligned}$$ Combining and , we have that $$\begin{aligned} {\mathop{\mathbb{E}}}_{X \sim \gD_{k-1}}\max_{\vx \in X} \sqrt{\vx^\top \mLambda_{k-1}^{-1} \vx} \leq \sqrt{2d \times \Tr(\mLambda_{k-1}^{-1} {\mathop{\mathbb{E}}}_{X \sim \gD_{k-1}} {\mathop{\mathbb{E}}}_{\vy \in \gopt(X)} \vy\vy^\top )} \leq 4\sqrt{2} d/ \sqrt{\gT_{k-1}},\label{eq:blinucbkw-2000}\end{aligned}$$ where the last inequality is due to . Combining and , we have that with probability at least $(1 - \delta \gT_k/T^2 - \delta/T^2)$, the expected regret incurred during batch $k$ ($k \geq 2$) is $$\begin{aligned} R_k \leq 4\alpha \gT_k \cdot 4\sqrt{2}d / \sqrt{\gT_{k-1}} \leq 16\sqrt{2} \alpha d\sqrt{T}. $$ Using the similar argument as the analysis for , we have that with probability at least $(1-\delta)$, the expected regret of [[BatchLinUCB-KW]{}]{}is at most $$\begin{aligned} R^T \leq O(d\sqrt{T \log (dKT/\delta)} \times \log \log T),\end{aligned}$$ proving . Distributional G-Optimal Design: Existence & Parametric Forms {#sec:dist-opt-design} ============================================================= We now work towards removing the extra $\sqrt{d}$ factor in the regret of , so as to achieve the optimal $\sqrt{dT}$-type regret. The high level idea is to use a difference sample policy other than uniform sampling over all (survived) candidate arms or the G-optimal-design-based $\gopt$. Given a sample policy $\pi$ that maps any set of arms ($X \subseteq \R^d$) to a distribution in $\triangle_X$, we will be interested in its performance, defined as follows. \[defn:distributional-g-opt-design\] For any distribution $\gD$ of the set of arms $X \subseteq \R^d$ and any sample policy $\pi$, we define the $\lambda$-distributional G-variation, or $\lambda$-variation for short ($\lambda > 0$), of $\pi$ over $\gD$ as $$\begin{aligned} \val_\gD^{(\lambda)}(\pi){\stackrel{\mathrm{def}}{=}}{\mathop{\mathbb{E}}}_{X \sim \gD}\max_{\vx \in X} \vx^\top \left(\lambda \mI + \info_\gD(\pi)\right)^{-1} \vx, \end{aligned}$$ where we define the information matrix by $$\begin{aligned} \info_\gD(\pi) {\stackrel{\mathrm{def}}{=}}{\mathop{\mathbb{E}}}_{X \sim \gD} \info_X(\pi), \qquad \text{where}~ \info_X(\pi) {\stackrel{\mathrm{def}}{=}}{\mathop{\mathbb{E}}}_{\vx \sim \pi(X)} \vx \vx^\top . \end{aligned}$$ Since $\val_\gD^{(\lambda)}$ is non-increasing as $\lambda$ grows, when the limit exists, we also define $$\begin{aligned} \val_\gD^{(0)}(\pi) {\stackrel{\mathrm{def}}{=}}\lim_{\lambda \to 0^+} \val_\gD^{(\lambda)}(\pi), \label{val-gopt}\end{aligned}$$ and set $\val_\gD^{(0)}(\pi) = + \infty$ otherwise. Indeed, the arguments in imply the following lemma. \[lem:variation-unif-optg\] For any distribution $\gD$ on the context vectors of the $K$ arms, we have that $$\begin{aligned} \val_\gD^{(0)}(\Unif) \leq O(dK), ~~\text{and}~~ \val_\gD^{(0)}(\gopt) \leq O(d^2) . \label{eq:util-unif-gopt}\end{aligned}$$ In light of , the question whether the regret of our algorithms can be improved to $O(\sqrt{dT \mathrm{poly}\log (KT/\delta)})$ boils down to whether one can find a sample policy $\pi$ such that the bounds in are improved to $O(d) \times \mathrm{poly} \log d$. In this section, we will show that such policies not only exist, but also admit a succinct parametric form so that we can later study how to efficiently learn the relevant parameters. To better explain our results, we first define the following class of parameterized sample policies. Suppose we are given a positive semi-definite matrix $\mV \succcurlyeq \vzero$. We define the associated argmax policy by $$\begin{aligned} \pap_\mV(X) = \argmax_{\vx \in X} \vx^\top \mV \vx,\end{aligned}$$ where in the $\argmax$ operator, ties are broken in a deterministic manner. In this subsection, we use $\gopt$ to denote a fixed policy with respect to and satisfying (up to the factor $2$ relaxation). Suppose we are given a set $\gV = \{(p_i, \mV_i)\}_{i = 1}^n$ such that $p_i \ge 0$ and $p_1 + \cdots + p_n \le 1$. We define the associated mixed-argmax policy by $$\begin{aligned} \pmap_\gV(X) = \begin{cases} \gopt(X), & \text{with probability } 1/2, \\ \pap_{\mV_i}(X), & \text{with probability } p_i / 2. \end{cases}\end{aligned}$$ The following theorem states that for any $\gD$, there exists a good mixed-argmax policy with only $O(d \log d)$ argmax policies in the mixture.[^10] \[thm:argmax-policy\] Fix any distribution $\gD = \Unif(S)$ where $S = \{X_1, X_2, \dots, X_{\Gamma}\}$ (which may be a multi-set) and any $\lambda \in (0, 1)$. There exists a mixed-argmax policy with parameters $\gV = \{(p_i, \mV_i)\}_{i = 1}^n$ such that 1. $n \leq 4 d \log d$; 2. for all $i \in [n]$, $p_i \geq 1/d^3$ and $ d^{-1} \mI \preccurlyeq \mV_i \preccurlyeq \lambda^{-1} \mI $; 3. $\val_\gD^{(\lambda)}(\pmap_{\gV}) \leq O(d \log d)$. We will assume $\Gamma > \lambda^{-1}$ without loss of generality, as the properties to be proved do not depend of $\Gamma$ and $S$ is a multi-set so that we can always duplicate the elements by finitely many times. We prove the theorem constructively. We consider , which is very similar to the linear bandits algorithms in literature. For $N = \Theta(d^2 \log d)$, the algorithm creates $\Gamma N$ times steps, which includes $N$ blocks, each of which contains $\Gamma$ consecutive time steps. In each block, the $\Gamma$ sets of arms $X_1, \dots, X_\Gamma$ are sequentially presented. The algorithm then simulates the linear bandit algorithms, where at each time step, the arm with the maximum variance (according to the information matrix $\mW_n$) is selected. Inspired by the rarely switching algorithm for linear bandits [@abbasi2011improved], the information matrix $\mW_n$ is only updated when its determinant doubles. This significantly reduces the number of updates and is crucial to upper bounding the number of individual argmax policies in the returned mixed-argmax policy. We refer to the consecutive time steps between two neighboring updates as a stage. Each of the information matrices in a stage corresponds to an individual argmax policy in the returned policy, and the corresponding probability weight is proportional to the length of the stage. The only exception is that we discard the stages that contain less than $\Gamma$ time steps (i.e., the ones that are shorter than a block). $N \gets d^2 \log d, \forall (i, j) \in [N] \times [\Gamma]: X_{(i-1)\Gamma + j} \gets X_j$ $\mU_0 \gets \lambda N\Gamma \mI + \frac{N}{2}\sum_{i = 1}^\Gamma {\mathop{\mathbb{E}}}_{\vx \sim \gopt(X_i)}[\vx \vx^\top] \succcurlyeq \mI, n \gets 1, \tau_n \gets \emptyset, \mW_n = \mU_0$ for all $i \in [n]$, \[line:rarelysw-9\] for all $i \in [n]$, set $p_i = \abs{\tau_i} / \sum_j \abs{\tau_j}$ #### Proof of Item (a). Note that $$\begin{aligned} \mU_{N\Gamma} = \mU_0 + \sum_{t=1}^{N\Gamma} \vx_t \vx_t^\top = \lambda N\Gamma \mI + \frac{N\Gamma}{2} \info_\gD(\gopt) + \sum_{t=1}^{N\Gamma} \vx_t \vx_t^\top. \label{eq:thm-ap-2700}\end{aligned}$$ By and (up to the factor $2$ relaxation), for all $t$, we have that $$\begin{aligned} \vx_t \vx_t^\top \preccurlyeq 2d \times {\mathop{\mathbb{E}}}_{\vy \in \gopt(X_t)} \vy\vy^\top. \label{eq:thm-ap-2800}\end{aligned}$$ Combining and , we have that $$\begin{aligned} \mU_{N\Gamma} \preccurlyeq \lambda N\Gamma \mI + (1/2 + 2d) N\Gamma \times \info_\gD(\gopt) \preccurlyeq 4d \mU_0.\end{aligned}$$ Therefore, we have $$\begin{aligned} \det \mU_{N\Gamma} \leq \det (4d \mU_{0}) = d^{4d} \det\mU_0,\label{eq:thm-ap-3000}\end{aligned}$$ and $n \leq \log (d^{4d}) = 4d \log d$. #### Proof of Item (b). Because we discard the stages whose lengths are less than $\Gamma$, for $p_i > 0$, we have that $$\begin{aligned} p_i \geq \frac{\Gamma}{N\Gamma} \geq \frac{1}{d^3}\end{aligned}$$ for large enough $d$. For each $\mW_i$, we have $\mW_i \succcurlyeq \mU_0 \succcurlyeq \lambda N \Gamma \mI$, and $\mW_i \preccurlyeq 2N\Gamma \mI$. Since $\mV_i = N \Gamma \mW_i^{-1}$, we have that $ d^{-1} \mI \preccurlyeq \mV_i \preccurlyeq \lambda^{-1} \mI$. #### Proof of Item (c). We finally upper bound the $\lambda$-variation of the returned policy $\pi = \pmap_{\gV}$. Note that $$\begin{aligned} \val^{(\lambda)}_\gD(\pi) &= {\mathop{\mathbb{E}}}_{X \sim \gD}[\max_{\vx \in X} \vx^\top (\lambda\mI + {\mathop{\mathbb{E}}}_{X \sim \gD} {\mathop{\mathbb{E}}}_{\vx \sim \pi(X)} \vx \vx^\top)^{-1} \vx] \notag \\ &= \sum_{t = 1}^{N\Gamma} \max_{\vx \in X_t} \vx^\top (N\Gamma (\lambda \mI + \info_\gD(\pi)))^{-1} \vx \notag \\ &= \sum_{i = 1}^n \sum_{t \in \tau_i} \max_{\vx \in X_t} \vx^\top (N\Gamma (\lambda \mI + \info_\gD(\pi)))^{-1} \vx + \sum_{t \in \gB} \max_{\vx \in X_t} \vx^\top (N\Gamma (\lambda \mI +\info_\gD(\pi)))^{-1} \vx, \label{eq:thm-ap-3300}\end{aligned}$$ where we let $\gB$ be the set of time steps that are discarded in of . It remains to show that both terms are $O(d \log d)$. For the second term, we have $$\begin{aligned} \sum_{t \in \gB} \max_{\vx \in X_t} \vx^\top (N\Gamma (\lambda\mI +\info_\gD(\pi)))^{-1} \vx &= \frac{1}{N\Gamma} \sum_{t \in \gB} \max_{\vx \in X_t} \vx^\top (\lambda \mI + \info_\gD(\pi))^{-1} \vx \nonumber \\ &\le \frac{2}{N\Gamma} \sum_{t \in \gB} \max_{\vx \in X_t} \vx^\top (\lambda\mI + \info_\gD(\gopt))^{-1} \vx, \label{eq:thm-ap-3100}\end{aligned}$$ where the inequality is because by definition of a mixed-argmax policy, with probability $1/2$, $\gopt$ is invoked, and therefore $$\begin{aligned} \info_\gD(\pi) = {\mathop{\mathbb{E}}}_{X \sim \gD, \vx \sim \pi(X)}\vx \vx^\top \succcurlyeq {\mathop{\mathbb{E}}}_{X \sim \gD} \frac{1}{2} \times{\mathop{\mathbb{E}}}_{\vx \sim \gopt(X)} \vx \vx^\top. \notag \end{aligned}$$ Continuing with , since $\gB$ contains at most $n$ stages that are shorter than a block, therefore, we have that $$\begin{aligned} &\frac{2}{N\Gamma} \sum_{t \in \gB} \max_{\vx \in X_t} \vx^\top (\lambda\mI + \info_\gD^{-1}(\gopt)) \vx \leq \frac{2}{N\Gamma} \times n \times \sum_{t=1}^{\Gamma} \max_{\vx \in X_t} \vx^\top (\lambda\mI + \info_\gD(\gopt))^{-1} \vx \nonumber\\ &\qquad = \frac{2n}{N} {\mathop{\mathbb{E}}}_{X \sim \gD} \max_{\vx \in X_t} \vx^\top (\lambda\mI + \info_\gD(\gopt))^{-1} \vx = \frac{2n}{N} \val_{\gD}^{(\lambda)}(\gopt) \leq \frac{2n}{N} \times O(d^2) \leq O(d \log d), \end{aligned}$$ where the second inequality is due to and the monotonicity of $\val_{\gD}^{(\lambda)}$. For the first term in , we claim that $$\begin{aligned} \info_\gD(\pi) \succcurlyeq \frac{1}{4N\Gamma} \sum_{t = 1}^{N\Gamma} \vx_t \vx_t^\top, \label{eq:thm-ap-4000}\end{aligned}$$ which will be established at the end of this proof. Once we have , also noting that $ \info_\gD(\pi) \succcurlyeq (1/2) \info_\gD(\gopt)$ because of the $1/2$ portion of $\gopt$ in the definition of the mixed-argmax policy, we get that $$\begin{aligned} \lambda \mI + \info_\gD(\pi) &\succcurlyeq \frac{1}{2}( \frac{1}{2} \info_\gD(\gopt) + \frac{1}{4N\Gamma} \sum_{t = 1}^{N\Gamma} \vx_t \vx_t^\top) \succcurlyeq \frac{1}{8 N\Gamma} \mU_{N\Gamma} \succcurlyeq \frac{1}{8 N\Gamma} \mW_n. \label{eq:thm-ap-4200}\end{aligned}$$ Therefore, $$\begin{aligned} & \sum_{i = 1}^n \sum_{t \in \tau_i} \max_{\vx \in X_t} \vx^\top N\Gamma (\lambda \mI + \info_\gD(\pi))^{-1} \vx \le 8 \sum_{i = 1}^n \sum_{t \in \tau_i} \max_{\vx \in X_t}\vx^\top \mW_n^{-1} \vx \notag \\ & \qquad \qquad \qquad \qquad \le 8 \sum_{i = 1}^n \sum_{t \in \tau_i} \max_{\vx \in X_t}\vx^\top \mW_i^{-1} \vx = 8 \sum_{i = 1}^n \sum_{t \in \tau_i} \vx_t^\top \mW_i^{-1} \vx_t \notag \\ & \qquad \qquad \qquad \qquad \le 16 \sum_{i = 1}^n \sum_{t \in \tau_i} \vx_t^\top \mU_t^{-1} \vx_t \le 16 \sum_{t = 1}^{N\Gamma} \vx_t^\top \mU_t^{-1} \vx_t \label{eq:thm-ap-4500} \\ & \qquad \qquad \qquad \qquad \le 32 \ln \frac{\det \mU_{N\Gamma}}{\det \mU_0} \le O(d \log d). \label{eq:thm-ap-4600}\end{aligned}$$ where the first inequality in is by , the first inequality in is by the elliptical potential lemma (),[^11] and the second inequality in is due to . It remains to establish . Note that $$\begin{aligned} \info_\gD(\pi) &\succcurlyeq \frac{1}{2} \info_\gD(\gopt) + \frac{1}{2} \sum_{i = 1}^n \frac{\abs{\tau_i}}{\abs{\tau_1} + \cdots + \abs{\tau_n}} \frac{1}{2\abs{\tau_i}}\sum_{t \in \tau_i} \vx_t \vx_t^\top \nonumber \\ &= \frac{1}{2} \info_\gD(\gopt) + \frac{1}{4} \frac{1}{N\Gamma - \abs{\gB}} \sum_{i = 1}^n \sum_{t \in \tau_i} \vx_t \vx_t^\top. \label{eq:thm-ap-4800}\end{aligned}$$ By , we have $$\begin{aligned} \info_\gD(\gopt) = \frac{1}{\abs{\gB}\Gamma} \sum_{t=1}^{\Gamma} \abs{\gB} \times {\mathop{\mathbb{E}}}_{\vx \sim \gopt(X_t)} \vx \vx^\top \succcurlyeq \frac{1}{\abs{\gB}\Gamma} \sum_{t\in \gB} \frac{1}{2d} \vx_t \vx_t^\top. \notag\end{aligned}$$ Therefore, continuing with , we have that $$\begin{aligned} \info_\gD(\pi) &\succcurlyeq \frac{1}{2\abs{\gB}d \Gamma} \sum_{t\in \gB} \vx_t \vx_t^\top + \frac{1}{4 N \Gamma} \sum_{i = 1}^n \sum_{t \in \tau_i} \vx_t \vx_t^\top \succcurlyeq \frac{1}{2 n d\Gamma} \sum_{t\in \gB} \vx_t \vx_t^\top + \frac{1}{4 N \Gamma} \sum_{i = 1}^n \sum_{t \in \tau_i} \vx_t \vx_t^\top \nonumber \\ &\succcurlyeq \frac{1}{4 N \Gamma} \sum_{t\in \gB} \vx_t \vx_t^\top + \frac{1}{4 N \Gamma} \sum_{i = 1}^n \sum_{t \in \tau_i} \vx_t \vx_t^\top = \frac{1}{4 N\Gamma} \sum_{t = 1}^{N \Gamma} \vx_t \vx_t^\top, \label{eq:thm-ap-5100}\end{aligned}$$ which concludes the proof of the theorem. The Mixed-Softmax Policies with More Robustness ----------------------------------------------- To make the sample policy learnable, instead of the mixed-argmax policies, we will deal with the more robust mixed-softmax policies. To define this class of policies, we first define the softmax function as a distribution such that $$\begin{aligned} \softmax_\alpha(s_1, \dots, s_k) = i \quad \text{with probability} \quad \frac{s_i^{\alpha}}{s_1^\alpha + \cdots + s_k^\alpha}, \notag\end{aligned}$$ where we assume that $s_i \geq 0$ for all $i \in [k]$. It is easy to check the following fact. \[fact:softmax\] Suppose $\alpha \geq \log k$, then $$\begin{aligned} {\mathop{\mathbb{E}}}_{i \sim \softmax_\alpha(s_1, \dots, s_k)} [s_i] \geq \frac14 \times \max\{s_1, \dots, s_k\}. \notag\end{aligned}$$ Let $i^$ be an index that maximizes $s_{i}$. Note that for all $j$ such that $s_j \leq (1/2) \times s_{i^}$, the probability mass that softmax put for $j$ is at most $(1/k)$ of that for $i^$. Therefore, $$\begin{aligned} \Pr_{i \sim \softmax_\alpha(s_1, \dots, s_k)} [s_i \geq \frac12 \times s_{i^}] \geq \frac{1}{2}, \notag\end{aligned}$$ and the fact follows. We now define the class of mixed-softmax policies. \[defn:sp\] Fix $\alpha = \log K$ (where $K$ is the number of arms per time step). Suppose we are given a positive semi-definite matrix $\mM \succcurlyeq \vzero$. We define the softmax policy $$\begin{aligned} \psp_\mM(X) = \vx_i, \qquad \text{where} \quad X = \{\vx_1, \dots, \vx_k\}, k \leq K, \text{~and~} i \sim \softmax_\alpha(\vx_1^\top \mM \vx_1, \dots, \vx_k^\top \mM \vx_k). \end{aligned}$$ Suppose we are given a set $\gM = \{(p_i, \mM_i)\}_{i = 1}^n$ such that $p_i \geq 0$ and $p_1 + \cdots + p_n = 1$. We define the mixed-softmax policy $$\begin{aligned} \pmsp_\gM(X) = \begin{cases} \gopt(X), & \text{with probability } 1/2, \\ \psp_{\mM_i}(X), & \text{with probability } p_i / 2. \end{cases}\end{aligned}$$ Similarly to , we prove the following theorem on the existence of good mixed-softmax policies. \[thm:ms-policy\] Fix any distribution $\gD = \Unif(S)$ where $S = \{X_1, X_2, \dots, X_{\Gamma}\}$ (which may be a multi-set) and any $\lambda \in (0, 1)$. There exists a mixed-softmax policy $\pmsp_\gM$ with parameters $\gM = \{(p_i, \mM_i)\}_{i = 1}^n$ such that 1. $n \leq 4 d \log d$; 2. for all $i \in [n]$, $p_i \geq 1/d^3$ and $ d^{-1} \mI \preccurlyeq \mM_i \preccurlyeq \lambda^{-1} \mI $; 3. $\val_\gD^{(\lambda)}(\pmsp_{\gM}) \leq O(d \log d)$. The proof of is very similar to that of . Here we only point out the differences as follows. First, at of , we let $\mX_t \gets {\mathop{\mathbb{E}}}_{\vx \sim \psp_{W_n^{-1}}(X_t)} \vx \vx^\top$, and at Line \[line:alg-rarelysw-6\], we let $\mU_{t} \gets \mU_{t- 1} + \mX_t$. Note that $\Tr(\mX_t) \leq 1$. Let $\gM$ be the output of the algorithm. The proof of Items (a) and (b) remains the same except for the occurrences of $\vx_t \vx_t^\top$ are replaced by $\mX_t$ in and . For the proof of Item (c), let $\pi = \pmsp_{\gM}$, we still get , and the second term of is bounded by the same way. For the first term, replacing $\vx_t\vx_t^\top$ by $\mX_t$ in (and its proof from to ), we still get . Therefore, $$\begin{aligned} & \sum_{i = 1}^n \sum_{t \in \tau_i} \max_{\vx \in X_t} \vx^\top ( N\Gamma (\lambda \mI + \info_\gD(\pi)))^{-1} \vx \le 8 \sum_{i = 1}^n \sum_{t \in \tau_i} \max_{\vx \in X_t}\vx^\top \mW_n^{-1} \vx \nonumber \\ & \quad \le 8 \sum_{i = 1}^n \sum_{t \in \tau_i} \max_{\vx \in X_t}\vx^\top \mW_i^{-1} \vx \leq 32 \sum_{i = 1}^n \sum_{t \in \tau_i} {\mathop{\mathbb{E}}}_{\vx \sim \pmsp_{\mW_i^{-1}}(X_t)}\vx^\top \mW_i^{-1} \vx = 32 \sum_{i = 1}^n \sum_{t \in \tau_i} \Tr(\mW_i^{-1} \mX_t), \label{eq:thm-sp-5700} \end{aligned}$$ where the third inequality in is by . Again, by and the elliptical potential lemma (), we have that $$\begin{aligned} & 32 \sum_{i = 1}^n \sum_{t \in \tau_i} \Tr(\mW_i^{-1} \mX_t) \leq 64 \sum_{i = 1}^n \sum_{t \in \tau_i} \Tr(\mU_t^{-1} \mX_t) \nonumber \\ & \qquad \qquad \qquad \qquad \qquad\leq 64 \sum_{t = 1}^{N\Gamma} \Tr(\mU_t^{-1} \mX_t) \leq 128 \ln \frac{\det \mU_{N\Gamma}}{\det \mU_0} \le O(d \log d). \notag\end{aligned}$$ Combining the bounds on both terms of , we prove the theorem. Learning the Distributional G-Optimal Design {#sec:learn-design} ============================================ In this section, we present an algorithm to learn a good mixed-softmax policy using only $\mathrm{poly}(d) \log \delta^{-1}$ samples with success probability at least $(1 - \delta)$. #### The Natural Idea and its Counterexample. The most natural idea is to first draw $\gamma$ independent samples $X_1, \dots, X_\gamma \sim \gD$ and form an empirical distribution $\gS = \Unif\{X_1, \dots, X_\gamma\}$, learn a good policy $\pi$ for $\gS$ according to , and hope that $\pi$ also works well for $\gD$ (i.e., $\pi$ generalizes to the true distribution). Unfortunately, such an approach is unlikely to work. Below we illustrate an example where, even when the number of samples $\gamma$ is very large, a good policy for $\gS$ still fails to generalize to $\gD$ with significant probability. Let $\{\ve_i\}_{i=1}^d$ be the set of canonical basis, and $\varepsilon > 0$ be a parameter to be determined later. Let $Y_1 = \{\ve_1\}$ and $Y_i = \{\sqrt{1-\varepsilon^2} \ve_i + \varepsilon \ve_1, \ve_i\}$ for $i \in \{2, 3, \dots, d\}$. Consider $\gD$ supported on $\{Y_1, \dots, Y_d\}$ the probability mass for $Y_1$ is $1/(d\gamma)$ and the probability for $Y_i$ ($i \geq 2$) is $q = (1 - 1/(d\gamma))/(d-1)$. If we make $\gamma$ independent samples $X_1, \dots, X_\gamma \sim \gD$, with probability $\Omega(1/d)$, we will see $Y_1$ once among the samples, and the probability mass of $Y_1$ in $\gS$ becomes $1/\gamma$, which is $d$ times its true probability mass. Due to this discrepancy, we will show that a good sample policy for the empirical distribution $\gS$ does not work as well on true distribution $\gD$. We consider the sample policy $\pi$ such that $\pi(X) = \ve_i$ when $X = Y_i$. When the event above happens, we have that $\info_\gS(\pi)= \mathrm{diag}(1/\gamma, p_2, \dots, p_d)$ where $p_i$ is the probability mass for $Y_i$ in $\gS$ (for $i \geq 2$). When $\varepsilon = \sqrt{d/\gamma}$, we can verify that $\pi$ is a good policy for the empirical distribution $\gS$ since $$\begin{aligned} \val_\gS^{(0)}(\pi) = {\mathop{\mathbb{E}}}_{X \sim \gS} \max_{\vx \in X} \vx^\top \info_\gS(\pi)^{-1} \vx = \frac1\gamma \cdot \gamma + \sum_{i=2}^d p_i \cdot \max\{\varepsilon^2 \gamma + (1 - \varepsilon^2) \cdot \frac{1}{p_i}, \frac{1}{p_i}\} \leq O(d). \notag \end{aligned}$$ However, for the true distribution $\gD$, we have that $\info_\gD(\pi)= \mathrm{diag}(1/(d\gamma), q, \dots, q)$, and for any $\lambda \in [0, 1/(d\gamma))$, it holds that $$\begin{aligned} & \val_\gD^{(\lambda)}(\pi) = {\mathop{\mathbb{E}}}_{X \sim \gD} \max_{\vx \in X} \vx^\top (\lambda \mI + \info_\gD(\pi))^{-1} \vx \notag \\ &\quad = \frac{1}{d\gamma} \cdot \frac{1}{\lambda + 1/(d\gamma)} + (1 - \frac{1}{d\gamma}) \cdot \max\{\varepsilon^2 \cdot \frac{1}{\lambda + 1/(d\gamma)} + (1-\varepsilon^2) \cdot \frac1{\lambda + q}, \frac1{\lambda +q}\} \geq \Omega(d^2). \notag \end{aligned}$$ Note that in this example, the only constraint for $\gamma$ is that $1/(d\gamma) > \lambda \Leftrightarrow \gamma < 1/(d\lambda)$. Therefore, we have illustrated that, even when $\gamma$ is greater than an arbitrary polynomial of $d$, with probability $\Omega(1/d)$, a good policy for the empirical distribution $\gS$ does not generalize to the true distribution $\gD$.[^12] By adding more dimensions, we can even strengthen this counterexample so that the failure probability becomes $(1 - o(1))$. Using similar tricks, we can also show that a good mixed-softmax policy does not generalize well. #### Our Algorithm: [CoreLearning]{}. The key message from the counterexample above is that if a context direction in $\sR^d$ appears with tiny probability in $\gD$, a limited amount of samples might greatly change its probability in the empirical distribution $\gS$, and fail the generalization argument. To address this issue, the idea of our new algorithm is to prune these infrequent context directions, learn a mixed-softmax policy over the remaining “core” directions, and finally argue that the infrequent directions can be properly handled by the $\gopt$ component in the mixed-softmax policy. In light of this idea, we propose [CoreLearning]{} (). In this algorithm, instead of directly learn the policy from the whole set of samples, we first find a large enough core set $C$ at , and then learn the mixed-softmax policy only using the samples in $C$. The key property of the core is specified by , which is a technical realization of our pruning idea. The property requires that every direction in $C$ should be well explored by the $\gopt$ policy and only the context vectors within $C$. To see how the core set helps to resolve the issue in our counterexample, we note that the infrequent set $Y_1$ is the main trouble-maker. However, even if $Y_1$ happens to appear among the samples $\{X_1, \dots, X_\gamma\}$, it will not be included in the core since its corresponding variation $\max_{\vy \in Y_1}\vy^\top (\lambda \mI + \info_{\Unif(C)}(\gopt))^{-1} \vy \geq (\lambda + 1/\gamma)^{-1} > d^c$ when $\lambda$ is sufficiently small and $\gamma \gg d^c$. Therefore, [CoreLearning]{} will learn a sample policy with $Y_1$ pruned away, and void our counterexample. While the core set property is much desirable, even whether such a core set with cardinality constraint exists is not obvious. In , we prove to show the its existence, and provide an efficient algorithm [CoreIdentification]{} to find one. The analysis of also relies on a few uniform concentration inequalities ( and ) which are proved later in . Set constant $c = 6$ Find a core $C \subseteq S = \{X_1, \dots, X_\gamma\}$ (using [CoreIdentification]{} (), see ) such that \[loc:learn-2\] $$\begin{aligned} & \max_{X \in C} \max_{\vx \in X} \{\vx^\top (\lambda \mI + \info_{\Unif(C)}(\gopt))^{-1} \vx\} \leq d^{c}, \qquad\qquad\qquad\qquad \label{eq:core1}\\ \text{and}\qquad\qquad\qquad\qquad & \qquad \qquad \frac{\abs{C}}{\gamma} \geq 1 - O(d^{3-c} \log \lambda^{-1}), \label{eq:core2}\end{aligned}$$ which is at least $1/2$ for sufficiently large $d$ Compute the mixed-softmax policy $\pi$ for the samples in $C$ (according to ) and return $\pi$ For now, assuming the lemmas introduced above, we prove the following main theorem of this section (the guarantee for ). \[thm:learn-design\] Suppose that $\lambda \in (\exp(-d), 1)$. Let $X_1, \dots, X_\gamma \sim \gD$ be [i.i.d.]{} drawn from the distribution $\gD$. Let $\pi$ be the returned policy of . We have that $$\begin{aligned} \Pr[\widetilde{\val}_\gD^{(\lambda)}(\pi) \leq O(\sqrt{d\log d})] &\geq 1 - \exp(O(d^4 \log^2 d) - \gamma d^{-2c} \cdot 2^{-16}) \notag \\ &= 1 - \exp(O(d^4 \log^2 d) - \gamma d^{-12} \cdot 2^{-16}), \notag \end{aligned}$$ where we define the $\lambda$-deviation of $\pi$ over $\gD$ by $$\begin{aligned} \widetilde{\val}_\gD^{(\lambda)}(\pi) {\stackrel{\mathrm{def}}{=}}{\mathop{\mathbb{E}}}_{X \sim \gD} \sqrt{\max_{\vx \in X} \{\vx^\top (\lambda \mI + \info_\gD(\pi)) \vx\}}. \label{eq:def-root-variation}\end{aligned}$$ Note that we are only able to provide the upper bound for $\widetilde{\val}_\gD^{(\lambda)}(\pi)$ instead of $\val_\gD^{(\lambda)}(\pi)$. However, this is still enough for our linear bandit application. We now prove . For notation convenience, we define $\gS {\stackrel{\mathrm{def}}{=}}\Unif(S)$, $\gC {\stackrel{\mathrm{def}}{=}}\Unif(C)$, and we define the mollifier $$\begin{aligned} \varphi_\beta(x) {\stackrel{\mathrm{def}}{=}}\begin{cases} 1, & \text{when~} x \le \beta, \\ \frac{2 \beta - x}{\beta} , & \text{when~} \beta \le x \le 2 \beta, \\ 0, & \text{when~} x > 2 \beta. \end{cases} \notag \end{aligned}$$ which is a continuous surrogate of the indicator function $\ind[x \leq \beta]$. We now condition on the successful events of the uniform convergence lemmas ( and ), which, by a union bound, happens with probability $$1 - \exp(O(d^3 \log d \log (d \lambda^{-1})) - \gamma d^{-2c} \cdot 2^{-16}) \geq 1 - \exp(O(d^4 \log^2 d) - \gamma d^{-2c} \cdot 2^{-16}).$$ Then, the proof of consists of the following four steps. #### Step : Lower Bounding the Information Matrix. The goal of this step is to establish . Let $\mU = \lambda \mI + \info_{\gC}(\pi)$. Note that $\lambda \mI \preccurlyeq \mU \preccurlyeq (1 + \lambda) \mI$. By the successful event in of (letting $\mW = \mU$), we have that $$\begin{aligned} {\mathop{\mathbb{E}}}_{X \sim \gD} \varphi_{2d^c} (\max_{\vx \in X}\{\vx^\top \mU^{-1} \vx\}) \cdot \info_X(\pi) \succcurlyeq \frac{1}{\gamma}\sum_{i=1}^\gamma \varphi_{2d^c}(\max_{\vx \in X_i} \{\vx^\top \mU^{-1} \vx\}) \cdot \info_{X_i}(\pi) -\frac14 \mU. \label{eq:sketch-1a}\end{aligned}$$ Since $\info_{\gD}(\pi) \succcurlyeq {\mathop{\mathbb{E}}}_{X \sim \gD} \varphi_{2d^c} (\max_{\vx \in X}\{\vx^\top \mU^{-1} \vx\}) \cdot \info_X(\pi)$ and $\frac{1}{\gamma}\sum_{i=1}^\gamma \varphi_{2d^c}(\max_{\vx \in X_i} \{\vx^\top \mU^{-1} \vx\}) \cdot \info_{X_i}(\pi) \succcurlyeq \frac12 \info_\gC(\pi)$, implies that $$\begin{aligned} \lambda \mI + \info_{\gD}(\pi)& \succcurlyeq \lambda \mI + \frac12 \info_\gC(\pi) - \frac14 \mU \succcurlyeq \frac14 (\lambda \mI + \info_\gC (\pi)) \notag \\ & \succcurlyeq \frac14 \left(\lambda \mI + \info_\gS (\pi) - \frac1\gamma \sum_{X_i \in S \setminus C} \info_{X_i}(\pi)\right) \succcurlyeq \frac14 \left(\lambda \mI + \info_\gS (\pi) - \frac{d}{\gamma} \sum_{X_i \in S \setminus C} \info_{X_i}(\gopt)\right) \notag \\ & \succcurlyeq \frac18 (\lambda \mI + \info_\gS (\pi)), \label{eq:sketch-1}\end{aligned}$$ where the last inequality is for $c \geq 6$. #### Step : Upper Bounding the Variation in the “Core Directions”. Let $\mW = \lambda \mI + \info_{\gS}(\pi) \succcurlyeq \frac12 (\lambda \mI + \info_{\gC}(\pi) )$. The goal of this step is to establish . By the successful event in of , we have that $$\begin{aligned} & {\mathop{\mathbb{E}}}_{X \sim \gD} \varphi_{4d^c}(\max_{\vx \in X}\{\vx^\top \mW^{-1} \vx\}) \cdot \sqrt{ \max_{\vx \in X}\{\vx^\top \mW^{-1} \vx\}} \notag \\ & \qquad\qquad\qquad\qquad\qquad\qquad \leq d + \frac1\gamma \sum_{i=1}^{\gamma} \varphi_{4d^c}(\max_{\vx \in X_i}\{\vx^\top \mW^{-1} \vx\}) \cdot \sqrt{\max_{\vx \in X_i}\{\vx^\top \mW^{-1} \vx\} }. \notag\end{aligned}$$ This implies that $$\begin{aligned} {\mathop{\mathbb{E}}}_{X \sim \gD} \varphi_{4d^c}(\max_{\vx \in X}\{\vx^\top \mW^{-1} \vx\}) \cdot \sqrt{\max_{\vx \in X}\{\vx^\top \mW^{-1} \vx\}} \leq d + \frac1\gamma \sum_{i=1}^{\gamma} \sqrt{\max_{\vx \in X_i}\{\vx^\top \mW^{-1} \vx\}} . \label{eq:sketch-2aa}\end{aligned}$$ Let $\zeta = 1 - \abs{C}/\abs{S} = 1 - \abs{C}/\gamma \leq O(d^{3-c} \log (1/\lambda))$. Note that $$\begin{aligned} &\quad \frac1\gamma \sum_{i=1}^{\gamma} \sqrt{\max_{\vx \in X_i}\{\vx^\top \mW^{-1} \vx\}} \nonumber \\ &\le \frac1\gamma \sum_{X_i \in C} \sqrt{\max_{\vx \in X_i}\{2 \vx^\top (\lambda \mI + \info_{\gC}(\pi))^{-1} \vx\}} + \frac1\gamma \sum_{X_i \in S \setminus C} \sqrt{\max_{\vx \in X_i}\{\vx^\top (\lambda \mI + (\zeta / 2) \info_{\Unif(S \setminus C)}(\gopt))^{-1} \vx\}}. \label{eq:sketch-2a}\end{aligned}$$ For the first term in , by the guarantee of , we have $$\begin{aligned} \frac1\gamma \sum_{X_i \in C} \sqrt{\max_{\vx \in X_i}\{2 \vx^\top (\lambda \mI + \info_{\gC}(\pi))^{-1} \vx\}} \le \sqrt{2 \rval_\gC(\pi)} \le O( \sqrt{d \log d} ). \label{eq:sketch-2b}\end{aligned}$$ For the second term in , by the variation bound for $\gopt$ (), we have that $$\begin{aligned} &\frac1\gamma \sum_{X_i \in S \setminus C} \sqrt{\max_{\vx \in X_i}\{\vx^\top (\lambda \mI + (\zeta / 2) \info_{\Unif(S \setminus C)}(\gopt))^{-1} \vx\}} \nonumber \\ & \qquad \leq \zeta \sqrt{\frac{1}{\zeta\gamma} \sum_{X_i \in S \setminus C} {\max_{\vx \in X_i}\{\vx^\top (\lambda \mI + (\zeta / 2) \info_{\Unif(S \setminus C)}(\gopt))^{-1} \vx\}}} \leq O(\sqrt{\zeta d^2}) \leq O(\sqrt{d}), \label{eq:sketch-2c}\end{aligned}$$ where the first inequality is Jensen and the last inequality is for $c \geq 5$. Combining , , , , we have that $$\begin{aligned} {\mathop{\mathbb{E}}}_{X \sim \gD} \varphi_{4d^c}(\max_{\vx \in X}\{\vx^\top \mW^{-1} \vx\}) \cdot \sqrt{\max_{\vx \in X}\{\vx^\top \mW^{-1} \vx\}} \leq O(\sqrt{d \log d}). \label{eq:sketch-2}\end{aligned}$$ #### Step : Upper Bounding the Variation in the “Infrequent Directions”. The goal of this step is to establish . By the successful event in of , we have that $$\begin{aligned} {\mathop{\mathbb{E}}}_{X \sim \gD}\varphi_{4d^c}(\max_{\vx \in X}\{\vx^\top \mW^{-1} \vx\}) \geq -d^{-1} + \frac1\gamma \sum_{i=1}^{\gamma}\varphi_{4d^c}(\max_{\vx \in X_i}\{\vx^\top \mW^{-1} \vx\}). \notag \end{aligned}$$ This implies that $$\begin{aligned} {\mathop{\mathbb{E}}}_{X \sim \gD}\varphi_{4d^c}(\max_{\vx \in X}\{\vx^\top \mW^{-1} \vx\}) & \geq - d^{-1} + \frac1\gamma \sum_{i=1}^{\gamma} \varphi_{2d^c}(\max_{\vx \in X_i}\{\vx^\top \mU^{-1} \vx\}) \notag \\ &\geq 1 - d^{-1} - O(d^{3-c} \log (1/\lambda)) \geq 1- O(d^{-1}), \notag\end{aligned}$$ where the last inequality is for $c \geq 5$. Let $\tau_X = 1 -\varphi_{4d^c}(\max_{\vx \in X}\{\vx^\top \mW^{-1} \vx\})$. We have that ${\mathop{\mathbb{E}}}_{X \sim \gD} \tau_X \leq O(d^{-1})$. Note that, $$\begin{aligned} &\quad {\mathop{\mathbb{E}}}_{X \sim \gD} \tau_X \sqrt{\max_{\vx \in X}\{\vx^\top \mW^{-1} \vx\}} = {\mathop{\mathbb{E}}}_{X \sim \gD} \tau_X \sqrt{\max_{\vx \in X}\{2 \vx^\top (\lambda \mI + {\mathop{\mathbb{E}}}_{X \sim \gD} \tau_X \info_X(\gopt))^{-1} \vx\}} \notag \\ &= {\mathop{\mathbb{E}}}_{X \sim \gD} \sqrt{\tau_X} \cdot \sqrt{\tau_X} \sqrt{\frac{1}{{\mathop{\mathbb{E}}}_X \tau_X} \cdot \max_{\vx \in X}\{2 \vx^\top (\frac{\lambda}{{\mathop{\mathbb{E}}}_X \tau_X} \mI + {\mathop{\mathbb{E}}}_{X \sim \gD} \frac{\tau_X}{{\mathop{\mathbb{E}}}_X \tau_X} \info_X(\gopt))^{-1} \vx\}} \notag \\ &\leq \sqrt{{\mathop{\mathbb{E}}}_{X \sim \gD} \tau_X} \cdot \sqrt{ {\mathop{\mathbb{E}}}_{X \sim \gD} \frac{\tau_X}{{\mathop{\mathbb{E}}}_X \tau_X} \cdot \max_{\vx \in X}\{2 \vx^\top (\frac{\lambda}{{\mathop{\mathbb{E}}}_X \tau_X} \mI + {\mathop{\mathbb{E}}}_{X \sim \gD} \frac{\tau_X}{{\mathop{\mathbb{E}}}_X \tau_X} \info_X(\gopt))^{-1} \vx\}} \label{eq:sketch-3a}\\ &\leq \sqrt{O(d^{-1})} \cdot \sqrt{O(d^2)} = O(\sqrt{d}). \label{eq:sketch-3b}\end{aligned}$$ Here, is due to Cauchy-Schwarz and the first inequality in is by the variation bound for $\gopt$ (). Altogether, we have that $$\begin{aligned} {\mathop{\mathbb{E}}}_{X \sim \gD}(1 -\varphi_{4d^c}(\max_{\vx \in X}\{\vx^\top \mW^{-1} \vx\})) \cdot \max_{\vx \in X}\{\vx^\top \mW^{-1} \vx\}\leq O(\sqrt{d}). \label{eq:sketch-3}\end{aligned}$$ #### Step : Putting Things Together. Combining and , we have $$\begin{aligned} {\mathop{\mathbb{E}}}_{X \sim \gD} \sqrt{\max_{\vx \in X}\{\vx^\top \mW^{-1} \vx\}} \leq O(\sqrt{d \log d}). \notag \end{aligned}$$ By the definition of $\mW$, and together with , we have that $$\begin{aligned} \widetilde{\val}_\gD^{(\lambda)}(\pi) = {\mathop{\mathbb{E}}}_{X \sim \gD} \sqrt{\max_{\vx \in X}\{\vx^\top (\lambda \mI +\info_{\gD}(\pi))^{-1} \vx\}} \leq O(\sqrt{d \log d}), \notag\end{aligned}$$ proving . Finding the Core {#sec:core} ---------------- We now present our algorithm ([CoreIdentification]{}, ) to find the core, and prove the following lemma on its guarantee. $C_1 = S$ \[lem:core\] Let $S = \{X_1, \dots, X_\gamma\}$ be a sequence/multi-set of context sets. finds a cor set $C \subseteq S$ in $O(d \log \lambda^{-1})$ iterations that satisfies and $$\begin{aligned} \max_{X_i \in C} \max_{\vx \in X_i} \vx^\top (\lambda \mI + \frac{1}{\gamma}\sum_{X_i \in C}\info_{X_i}(\gopt))^{-1} \vx \le d^c. \label{eq:core3}\end{aligned}$$ We remark that implies , because $\frac{1}{\gamma}\sum_{X_i \in C}\info_{X_i}(\gopt) \preccurlyeq \info_{\Unif(C)}(\gopt)$. For any iteration $\xi$, we denote $$\begin{aligned} \mJ_\xi = (\lambda \mI +\frac{1}{\gamma}\sum_{X_i \in C_\xi}\info_{X_i}(\gopt))^{-1}. \notag \end{aligned}$$ We first claim that, for each $\xi$, either (a) $C_{\xi + 1}$ satisfies (and thus the algorithm returns), or (b) $\det \mJ_{\xi + 1} \ge 2 \det \mJ_{\xi}$. To see this, suppose that (a) does not hold. In this case, we have that there exists $X_i \in C_{\xi + 1}$ and $\vx_i \in X_i$, such that $$\begin{aligned} \vx_i^\top \mJ_{\xi + 1} \vx > d^c. \label{eq:core-7000}\end{aligned}$$ Since $X_i \in C_{\xi + 1}$, by of , we know that $$\begin{aligned} \vx_i^\top \mJ_\xi \vx \le \frac{1}{2}d^c. \label{eq:core-7100}\end{aligned}$$ Dividing by , together with , we find that $$\begin{aligned} \frac{\det \mJ_{\xi + 1}}{\det \mJ_{\xi}} \ge \frac{\vx_i^\top \mJ_{\xi + 1} \vx}{\vx_i^\top \mJ_\xi \vx} > 2, \label{eq:core-7200}\end{aligned}$$ proving the claim. Now we prove the lemma. First, we prove that the algorithm returns after at most $O(d \log \lambda^{-1})$ iterations. Note that $\mJ_1 \succcurlyeq (1 + \lambda)^{-1}\mI$. Furthermore, for every iteration $\xi$, we have that $\mJ_\xi \preccurlyeq \lambda^{-1} \mI$. Together with $\lambda < 1$, we have that $\det \mJ_\xi \le (2\lambda)^{-d} \det \mJ_1$. By the claim established above in , we have that $\det \mJ_\xi \ge 2^{\xi - 1} \det \mJ_1$ so long as the algorithm does not return at iteration $\xi$. Thus we conclude that $\xi \le O(d \log \lambda^{-1})$ when the algorithm returns. Let $C = C_\xi$ be the returned set. We also need to show that $C$ satisfies . We claim that for each iteration $j$, $$\begin{aligned} \abs{C_j \setminus C_{j + 1}} \le d^{2-c} \gamma, \label{eq:core-7300}\end{aligned}$$ which implies , because $$\begin{aligned} \frac{\abs{C}}{\gamma} = \frac{\gamma - \abs{C_1 \setminus C_\xi}}{\gamma} = 1 - \sum_{j = 1}^{\xi - 1} \frac{\abs{C_j \setminus C_{j + 1}}}{\gamma} \ge 1 - \sum_{j = 1}^{\xi - 1} d^{2-c} \ge 1 - d^{2-c} \cdot O(d \log \lambda^{-1}), \notag \end{aligned}$$ where the first inequality uses and the second inequality uses $\xi \le O(d \log \lambda^{-1})$. Finally, we prove . We have that $$\begin{aligned} \abs{C_j \setminus C_{j + 1}} &= \abs{\{X_i \in C_j : \max\limits_{\vx \in X_i} \vx^\top \mJ_j \vx > (1/2)d^c\}} \notag \\ &= \gamma \cdot \Pr_{X_i \sim \Unif(S)}[\max_{\vx \in X_i} \{\ind[X_i \in C]\vx^\top \mJ_j \vx \}> (1/2)d^c] \notag \\ &\le \gamma \cdot 2 d^{-c} {\mathop{\mathbb{E}}}_{X_i \sim \Unif(S)} \max_{\vx \in X_i} \{ \ind[X_i \in C]\vx^\top \mJ_j \vx\}\le 2 \gamma d^{2-c}, \notag $$ where the first inequality is by Markov’s inequality and the second inequality uses the variation bound for $\gopt$ () on the distribution $\gD = \ind[X \in C] \cdot X$, where $X \sim \Unif(S)$. Uniform Concentration Lemmas {#sec:uniform-con} ---------------------------- Fix $\lambda < 1$. We define the following set of positive semi-definite matrices $$\begin{aligned} \mathfrak{W} {\stackrel{\mathrm{def}}{=}}\{\mW \in \sR^{d \times d} \mid \lambda \mI \preccurlyeq \mW \preccurlyeq (1 + \lambda) \mI\}. \label{eq:core-psd-1}\end{aligned}$$ Let $X_1, \dots, X_\gamma$ be a sequence of sets of context vectors with norm at most $1$. For any positive definite matrix $\mW \in \sR^{d \times d}$, we define the following functions. $$\begin{aligned} f(\mW) &{\stackrel{\mathrm{def}}{=}}\frac{1}{\gamma} \sum_{i = 1}^\gamma \varphi_{4d^c}(\max_{\vx \in X_i}\{\vx^\top \mW^{-1} \vx\}) \cdot \max_{\vx \in X_i}\{\vx^\top \mW^{-1} \vx\}, \notag \\ g(\mW) &{\stackrel{\mathrm{def}}{=}}\frac1\gamma \sum_{i=1}^{\gamma}\varphi_{4d^c}(\max_{\vx \in X_i}\{\vx^\top \mW^{-1} \vx\}). \notag\end{aligned}$$ \[lem:func-lip-1\] For any positive $\lambda < 1$, $f(\mW)$ and $g(\mW)$ are $2\lambda^{-3}$-Lipschitz (in terms of $2$-norm $\\|\cdot\\|$) in the range $\mathfrak{W}$. By (and that the context vectors have norm at most $1$), for any $X_i$, the function $\max_{\vx \in X_i}\{\vx^\top \mW^{-1} \vx\}$ is $\lambda^{-2}$-Lipschitz with respect to $\mW \in \mathfrak{W}$. Therefore, $\varphi_{4d^c}(\max_{\vx \in X_i}\{\vx^\top \mW^{-1} \vx\})$ is also $\lambda^{-2}$-Lipschitz with respect to $\mW$. Since $\phi_{4d^c}(\cdot) \in (0, 1)$ and $\max_{\vx \in X_i}\{\vx^\top \mW^{-1} \vx\} \in (0, \lambda^{-1})$, we have that $\varphi_{4d^c}(\max_{\vx \in X_i}\{\vx^\top \mW^{-1} \vx\}) \cdot \max_{\vx \in X_i}\{\vx^\top \mW^{-1} \vx\}$ is $(\lambda^{-3} + \lambda^{-2})$-Lipschitz, and this proves the lemma. We now present our first uniform concentration lemma. \[lem:uniform-con-1\] Let $X_1, \dots, X_\gamma \sim \gD$ be [i.i.d.]{} drawn from the distribution $\gD$. We have the following concentration properties, $$\begin{aligned} \Pr[\sup_{\mW \in \mathfrak{W}} \{{\mathop{\mathbb{E}}}f(\mW) - f(\mW)\}\le d] &\ge 1 - \exp(O(d^2 \log (d\lambda^{-1})) - \gamma d^{2-2c}/128), \label{eq:core-eve2} \\ \Pr[\sup_{\mW \in \mathfrak{W}} \{{\mathop{\mathbb{E}}}g(\mW) - g(\mW)\} \le d^{-1}] &\ge 1 - \exp(O(d^2 \log (d\lambda^{-1})) - \gamma d^{-2}/2). \label{eq:core-eve3}\end{aligned}$$ Let $\mathfrak{V}_\varepsilon \subseteq \mathfrak{W}$ be an $\varepsilon$-covering of $\mathfrak{W}$ so that for any $\mM \in \mathfrak{W}$, there exists $\mN = \mN_\varepsilon(\mM) \in \mathfrak{V}_\varepsilon$ satisfying $\norm{\mM - \mN} \le \varepsilon$. For , we first consider a fixed matrix $\mN \in \mathfrak{V}_\varepsilon$. For $i \in [\gamma]$, let $$\begin{aligned} Y_i = \varphi_{4d^c}(\max_{\vx \in X_i}\{\vx^\top \mW^{-1} \vx\}) \cdot \max_{\vx \in X_i}\{\vx^\top \mW^{-1} \vx\}. \notag \end{aligned}$$ Then $\{Y_i\}$ are independent and bounded as $\abs{Y_i} \le 8 d^c$ almost surely. Using with $\delta = \frac{d}{2}$ and $R = 8 d^c$, we have $$\begin{aligned} \Pr[{\mathop{\mathbb{E}}}f(\mN) - f(\mN) \le d/2] \ge 1 - 2\exp(-\frac{2 \gamma \delta^2}{R^2}) = 1 - 2\exp(-\gamma d^{2-2c}/128)). \notag $$ Next we consider all $\mN \in \mathfrak{V}_\varepsilon$. Using a union bound, we have that $$\begin{aligned} \Pr[\max_{\mN \in \mathfrak{V}_\varepsilon}\{{\mathop{\mathbb{E}}}f(\mN) - f(\mN)\} \le d/2] \ge 1 - \abs{\mathfrak{V}_\varepsilon} \cdot 2 \exp(-\gamma d^{2-2c}/128). \label{eq:uniform-con-1-b}\end{aligned}$$ Finally, we choose $\varepsilon = \lambda^3 d/ 4$. By the Lipschitzness of $f(\mW)$ in , we have that $$\begin{aligned} \abs{f(\mW) - f(\mN_\varepsilon(\mW))} \le 2\lambda^{-3} \norm{\mW - \mN_{\varepsilon}(\mW)} \le 2\lambda^{-3} \varepsilon = d / 2. \notag \end{aligned}$$ Therefore, using , we have that $$\begin{aligned} \Pr[\sup_{\mW \in \mathfrak{W}} \{{\mathop{\mathbb{E}}}f(\mW) - f(\mW)\}\le d] &\ge \Pr[d/2 + \sup_{\mW \in \mathfrak{V}_\varepsilon} \{{\mathop{\mathbb{E}}}f(\mW) - f(\mW)\}\le d] \notag \\ &\ge 1 - \abs{\mathfrak{V}_\varepsilon} \cdot 2 \exp(-\gamma d^{2-2c}/128)) \notag \\ &\ge 1 - \exp(O(d^2 \log (d\lambda^{-1})) - \gamma d^{2-2c}/128), \notag \end{aligned}$$ where the last inequality uses the covering number bound in . For , we can prove it similarly as . The only differences are that 1) we need to apply with $\delta = 1/(2d)$ and $R = 1$, and 2) we need to choose $\varepsilon = \lambda^3 / (4d)$. We define the policy class $\Pi$ by $$\begin{aligned} \Pi &{\stackrel{\mathrm{def}}{=}}\{\pmsp_\gM \mid \gM = \{(p_j, \mM_j)\}_{j = 1}^n, p_j \ge 0, p_1 + \cdots + p_n = 1, \mM_j \in \mathfrak{M}, n \leq 4 d \log d\}, \notag \\ \text{where}~\mathfrak{M} &{\stackrel{\mathrm{def}}{=}}\{\mM \in \sR^{d \times d} \mid d^{-1} \mI \preccurlyeq \mM \preccurlyeq \lambda^{-1} \mI \}, \notag $$ and we define the following (matrix-valued) function on $\mW \in \mathfrak{W}$ and $\pi = \pmsp_\gM\in \Pi$, $$\begin{aligned} \mF(\mW, \pmsp_\gM) = \mF(\mW, \gM) {\stackrel{\mathrm{def}}{=}}\frac{1}{\gamma}\sum_{i = 1}^\gamma \varphi_{2d^c}(\max_{\vx \in X_i} \{\vx^\top \mW^{-1} \vx\}) \cdot \mW^{-1/2} \info_{X_i}(\pmsp_\gM) \mW^{-1/2}.\label{eq:def-mF}\end{aligned}$$ \[lem:func-mF-smooth\] We claim the following smoothness properties of the function $\mF(\mW, \gM)$ on its parameters, 1. $\mF(\cdot,\cdot)$ is $3\lambda^{-3}$-Lipschitz with respect to $\mW$; 2. $\mF(\cdot,\cdot)$ is $\lambda^{-2}$-Lipschitz with respect to each $p_j$; 3. for any two parameters $$\begin{aligned} \gM = \{(p_j, \mM_j)\}_{j = 1}^n, \gM' = \{(p_j, \mM'_j)\}_{i = 1}^n, \quad \mathrm{such~that} ~ \max_{1 \le j \le n} \norm{\mM_j - \mM'_j} \le 1/R, \notag \end{aligned}$$ where $R \ge 100\lambda^{-1} \cdot d \log K$, we have $$\begin{aligned} (1-\lambda/30)\mF(\mW, \gM') \preccurlyeq \mF(\mW, \gM) \preccurlyeq (1-\lambda/30)^{-1} \mF(\mW, \gM'), \label{eq:func-mF-3} \end{aligned}$$ which further implies (since $ \mF(\mW, \gM')\preccurlyeq \lambda^{-1} \mI$), $$\begin{aligned} \mF(\mW, \gM') - \frac1{30} \cdot \mI \preccurlyeq \mF(\mW, \gM) \preccurlyeq \mF(\mW, \gM') + \frac1{20} \cdot \mI. \notag \end{aligned}$$ For item (a), we note that $\norm{\mW^{-1/2}} \le \lambda^{-1/2}$, that $\mW^{-1/2}$ is $\lambda^{-3/2}$-Lipschitz with respect to $\mW$ by , and that $\norm{\info_{X_i}(\pi)} \le \Tr(\info_{X_i}(\pi)) \le 1$. Also, by the proof of , we have that $\varphi_{2d^c}(\max_{\vx \in X_i}\{\vx^\top \mW^{-1} \vx\})$ is $\lambda^{-2}$-Lipschitz with respect to $\mW$, that $\phi_{2d^c}(\cdot) \in (0, 1)$, and that $\max_{\vx \in X_i}\{\vx^\top \mW^{-1} \vx\} \in (0, \lambda^{-1})$. Therefore, we can prove this item. For item (b), we note that $\info_{X_i}(\pmsp_\gM)$ is $1$-Lipschitz in each $p_j$ and we conclude by noting that $\norm{\mW^{-1/2}}^2 \le \lambda^{-1}$ and $\max_{\vx \in X_i}\{\vx^\top \mW^{-1} \vx\} \in (0, \lambda^{-1})$. For item (c), since a mixed-softmax policy is a mixture of softmax policies, in the remaining proof, we first analyze the information matrix of the softmax policies $\psp_{\mM_j}, \psp_{\mM'_j}$, and then analyze that of the mixed-softmax policies $\pmsp_\gM, \pmsp_{\gM'}$. Since $\norm{\mM_j - \mM'_j} \le 1/R$, we have $$\begin{aligned} \mM'_j = \mM_j + (\mM'_j - \mM_j) \succcurlyeq \mM_j - \frac{\mI}{R} \succcurlyeq (1 - \frac{d}{R}) \mM_j, \label{eq:rounding-3500}\end{aligned}$$ where the second inequality in uses $\mM_j \succcurlyeq d^{-1}\mI$. Similarly, we can show $$\begin{aligned} \mM'_j = \mM_j + (\mM'_j - \mM_j) \preccurlyeq \mM_j + \frac{\mI}{R} \preccurlyeq (1 + \frac{d}{R}) \mM_j. \notag \end{aligned}$$ Recall that $\alpha = \ln K$. For the softmax policy $\psp_{\mM'_j}$ and any vector $\vx \in \sR^d$, we have that $$\begin{aligned} (\vx^\top \mM'_j \vx)^\alpha &\ge (1 - \frac{d}{R})^\alpha (\vx^\top \mM_j \vx)^\alpha = (1 - \frac{d}{R})^{\ln K} (\vx^\top \mM_j \vx)^\alpha \nonumber \\ &\ge (1 - \frac{1}{100 \lambda^{-1} \cdot \ln K})^{\ln K} (\vx^\top \mM_j \vx)^\alpha \ge (1-\lambda/100)(\vx^\top \mM_j \vx)^\alpha. \label{eq:rounding-4500}\end{aligned}$$ Similarly, we have $$\begin{aligned} (\vx^\top \mM'_j \vx)^\alpha &\le (1 + \frac{d}{R})^\alpha (\vx^\top \mM_j \vx)^\alpha = (1 + \frac{d}{R})^{\ln K} (\vx^\top \mM_j \vx)^\alpha \nonumber \\ &\le (1 + \frac{1}{100\lambda^{-1} \cdot \ln K})^{\ln K} (\vx^\top \mM_j \vx)^\alpha \le (1 + \lambda/50) (\vx^\top \mM_j \vx)^\alpha. \label{eq:rounding-5500}\end{aligned}$$ Therefore, for any context set $X$, we have $$\begin{aligned} \Pr[\psp_{\mM'_j}(X) = \vx]& = \frac{(\vx^\top \mM'_j \vx)^\alpha}{\sum_{\vx \in X} (\vx^\top \mM'_j \vx)^\alpha} \nonumber \\ &\ge \frac{(1 - \lambda / 100)(\vx^\top \mM_j \vx)^\alpha}{\sum_{\vx \in X} (1 + \lambda / 50) (\vx^\top \mM_j \vx)^\alpha} \geq (1 - \lambda / 30)\Pr[\psp_{\mM_j}(X) = \vx], \notag\end{aligned}$$ where the first inequality uses and . As a direct corollary, for any context set $X$, we have $$\begin{aligned} \info_X(\psp_{\mM'_j}) &= \sum_{\vx \in X} \vx \vx^\top \cdot \Pr[\psp_{\mM'_j}(X) = \vx] \nonumber \\ &\succcurlyeq (1 - \lambda / 30)\sum_{\vx \in X} \vx \vx^\top \cdot \Pr[\psp_{\mM_j}(X) = \vx] = (1 - \lambda / 30) \info_{X}(\psp_{\mM_j}). \label{eq:rounding-6000}\end{aligned}$$ Therefore, for the mixed-softmax policy, we have $$\begin{aligned} &\info_X(\pmsp_{\gM'}) = \frac{1}{2} \info_X(\gopt) + \sum_{i = 1}^n \frac{p_j}{2} \info_X(\psp_{\mM'_j}) \nonumber \\ &\qquad\qquad\qquad\qquad\succcurlyeq (1 - \lambda / 30)\left(\frac{1}{2} \info_X(\gopt) + \sum_{i = 1}^n \frac{p_j}{2} \info_X(\psp_{\mM_j}) \right) = (1 - \lambda / 30)\info_{X}(\pmsp_{\gM}), \label{eq:rounding-6500}\end{aligned}$$ where the inequality uses . Swapping $\mM_j$ and $\mM'_j$, we find $$\begin{aligned} \info_X(\pmsp_{\gM'}) \preccurlyeq (1-\lambda/30)^{-1}\info_{X}(\pmsp_{\gM}). \label{eq:rounding-7000}\end{aligned}$$ Setting $X = X_i$ in and , we establish . We now present our second uniform concentration lemma. \[lem:uniform-con-2\] Fix $\lambda <1$. For any mixed-softmax policy $\pmsp_{\gM}$ with $\gM \in \mathfrak{M}$, and any positive semi-definite matrix $\mW \in \mathfrak{W}$ (defined in ), we define the random function $$\begin{aligned} \mG(\mW, \gM) {\stackrel{\mathrm{def}}{=}}\mF(\mW, \gM) - {\mathop{\mathbb{E}}}\mF(\mW, \gM), \notag\end{aligned}$$ where the random function $\mF$ is defined in , and the randomness is from the independent samples $X_1, \dots, X_\gamma \sim \gD$. We then have that $$\begin{aligned} \Pr[\sup_{\mW \in \mathfrak{W}} \sup_{\pi \in \Pi} \norm{\mG(\mW, \gM)} \le \frac14] &\ge 1 - \exp(O(d^3 \log d \log (d \lambda^{-1})) - \gamma d^{-2c} \cdot 2^{-16}), \label{eq:core-eve1}\end{aligned}$$ First, we consider fixed $\mW, \gM$. Using with $\delta = 1/32$ and $R = 2 d^c$, we have $$\begin{aligned} \Pr[\norm{\mG(\mW, \gM)} \le \frac{1}{32}] \ge 1 - 2d \exp(- \frac{\gamma \delta^2}{8 R^2 + 4 \delta R / 3}) \ge 1 - 2d \exp(-\gamma d^{-2c}\cdot 2^{-16}). \notag \end{aligned}$$ Second, we define the covering. Let $$\begin{aligned} \mathfrak{V} = \gN(\mathfrak{W}, \lambda^{3} / 120, \norm{\cdot}), ~~ \mathfrak{P} = \gN([0, 1], \lambda^2 / (40 \cdot 4d\log d), \abs{\cdot}), ~~ \mathfrak{N} = \gN(\mathfrak{M}, 1 / (100 \lambda^{-1}\cdot d \ln K), \norm{\cdot}), \notag \end{aligned}$$ and let $$\begin{aligned} \mathfrak{S} = \{\pmsp_{\gM} \in \Pi \mid \gM = \{(p_j = q_j/(q_1 + \cdots + q_n), \mM_i)\}_{j = 1}^n, q_j \in \mathfrak{P}, \mM_i \in \mathfrak{N}\}. \notag \end{aligned}$$ We have $$\begin{aligned} &\quad\Pr[\max_{\mW \in \mathfrak{V}} \max_{\pi \in \mathfrak{S}} \norm{\mF(\mW, \pi) - {\mathop{\mathbb{E}}}\mF(\mW, \pi)} \le \frac{1}{32}] \label{eq:core-8000} \\ &\ge 1 - 2d \abs{\mathfrak{V}}\abs{\mathfrak{S}} \exp(-\gamma d^{-2c} \cdot 2^{-16}) \notag \\ &\ge 1 - 2d \abs{\mathfrak{V}}\abs{\mathfrak{P}}^n \abs{\mathfrak{N}}^n \exp(-\gamma d^{-2c} \cdot 2^{-16}) \notag \\ &\ge 1 - \exp(\log d + O(d^2 \log(d\lambda^{-1})) + n \cdot \left[ O(d \log \lambda^{-1}) + O\left(d^2 \log(d\lambda^{-1}d \log K) \right) \right] - \gamma d^{-2c} \cdot 2^{-16}) \notag \\ &\ge 1 - \exp(O(d^3 \log d \log (d \lambda^{-1})) - \gamma d^{-2c} \cdot 2^{-16}), \notag \end{aligned}$$ where the last inequality uses that $\log K \le O(d)$. Finally, we invoke the smoothness results from . Note that when the event in holds, by item (c) of , we have $$\begin{aligned} \max_{\mW \in \mathfrak{V}} \max_{\{p_j = q_j / (q_1+\dots + q_n) : q_j \in \mathfrak{P}\}} \sup_{\{\mM_i \in \mathfrak{M}\}} \norm{\mG(\mW, \pi)} \le \frac{1}{32} + \frac{1}{20} \cdot 2 \leq \frac{1}{7}. \notag \end{aligned}$$ Note that the Lipschitz constant of $\mG(\cdot,\cdot)$ is at most double of that of $\mF(\cdot,\cdot)$. By item (b) of , we have $$\begin{aligned} \max_{\mW \in \mathfrak{V}} \sup_{\pi \in \Pi} \norm{\mG(\mW, \pi)} \le \frac{1}{7} + 2\lambda^{-2} \cdot 4d\log d \cdot \frac{\lambda^2}{40 \cdot 4d\log d} = \frac{1}{5}. \notag \end{aligned}$$ Finally, by item (a) of , we have that $$\sup_{\mW \in \mathfrak{W}} \sup_{\pi \in \Pi} \norm{\mG(\mW, \pi)} \le \frac{1}{5} + 2 \cdot 3\lambda^{-3} \cdot \frac{\lambda^{3}}{120} = \frac{1}{4}. \qedhere$$ Putting Everything Together: the Optimal Batch Algorithm {#sec:blinucbdg} ======================================================== $M = \lceil \log \log T \rceil + 1, \alpha \gets 10\sqrt{ \ln \frac{2dKT}{\delta}}, \pi^0 = \gopt$, $\gT = \{\gT_1, \gT_2, \dots, \gT_M\}$, where $\gT_0 = 0$, $\gT_1 = \sqrt{T}$, $\gT_2 = 2\sqrt{T}$, and $\gT_i = T^{1 - 2^{-(i-1)}}$ for $i \in \{3, \dots, M-1\}, \gT_M = T$ Our final algorithm with $O(\log \log T)$ static-grid batches and optimal minimax expected regret (up to $\mathrm{poly} \log T$ factors) is presented in . Compared with [[BatchLinUCB]{}]{}and [[BatchLinUCB-KW]{}]{}, the main difference here is the addition of from to , which not only learns the new estimate $\hat\vtheta_k$, but also the new sample policy $\pi_k$. Learning of the two objects are done through disjoint sets of samples ($\gA$ and $\gB$). This is because that $\gD_k$ depends on $\hat\vtheta_k$ (which is learned from $\gA$) and we have to make $\gB$ disjoint from $\gA$ so as to ensure elements in $S$ are independently sampled from $\gD_k$. The following theorem bounds the expected regret of . \[thm:blinucbdg\] Assume that $T \leq \exp(d)$ and $T \geq \Omega(d^{32} \log^4 d \log^2 (\delta^{-1}))$. With probability at least $(1 - \delta)$, the expected regret of is bounded as $$\begin{aligned} R^T_{\text{{{\sc BatchLinUCB-DG}\xspace}}} \le O(\sqrt{d T \log d \log(dKT/\delta)} \times \log \log T). \notag \end{aligned}$$ Note that the assumption that $T \leq \exp(d)$ is not restrictive since otherwise we have $\log d \geq \Omega(T)$ and [[BatchLinUCB-KW]{}]{}() already achieves the minimax optimal regret up to $\mathrm{poly}\log T$ factors. We also note that the $K$ in the regret bound can be replaced by $\min\{K, d \log T\}$ by a simple $\varepsilon$-net argument, so that our regret bound becomes minimax-optimal for all $K$ (up to $\mathrm{poly}\log T$ factors). We finally remark that we make no effort in optimizing the exponent in the constraint that $T \geq d^{O(1)}$. Some simple tricks may significantly reduce this exponent constant. For example, first running a revised version of [[BatchLinUCB-KW]{}]{}till time $\sqrt{T/d}$ and then switch to [[BatchLinUCB-DG]{}]{}would reduce the exponent to $17$. A more careful analysis in the concentration lemmas in may further substantially optimize the constant. We now provide the proof of . We adopt the notations in . Conditioned on the batches $1, 2, \dots, k-1$, we can bound the expected regret incurred in batch $k$ similarly as , and have that with probability at least $(1 - \delta \gT_k/T^2)$, $$\begin{aligned} R_k \leq 4\alpha \gT_k \times {\mathop{\mathbb{E}}}_{X \sim \gD_{k-1}}\max_{\vx \in X} \sqrt{\vx^\top \mLambda_{k-1}^{-1} \vx}. \label{eq:blinucbdg-rk}\end{aligned}$$ Furthermore, similar to , we can show that for each batch $k$ ($k < M$), with probability $(1 - \delta / T^2)$, we have that $$\begin{aligned} \mLambda_k \succcurlyeq \frac{\gT_{k}}{32} \left(\frac{ \ln T}{ \gT_{k}} \mI + {\mathop{\mathbb{E}}}_{X \sim \gD_{k-1}} {\mathop{\mathbb{E}}}_{\vx \sim \pi_{k-1}(X)}[\vx \vx^\top]\right) \succcurlyeq \frac{\gT_k}{32} \left(T^{-1} \cdot \mI + \info_{\gD_{k-1}}(\pi_{k-1})\right). \label{eq:lem-exploration-new}\end{aligned}$$ Note that compared with , has a worse constant $32$ since $\gA$ only contains half of the samples. For each $k < M$, note that at , $S = \{\{\vx_{\tau, a} \mid a \in A_\tau^{(k)}\}\}_{\tau \in \gB}$ contains [i.i.d.]{} samples from $\gD_{k}$, and $\|S\| \geq \|\gT_k - \gT_{k-1}\| /2 \geq \sqrt{T} / 4$. By , we have that with probability $(1 - \exp(O(d^4 \log^2 d) - \sqrt{T} d^{-12} \cdot 2^{-18}) \geq 1 - \delta/T^2$ (since $T \geq \Omega(d^{32} \log^4 d \log^2 (\delta^{-1}))$, it holds that $$\begin{aligned} \widetilde{\val}_{\gD_k}^{(1/T)}(\pi_k) \leq O(\sqrt{d\log d}). \label{eq:blinucbdg-val}\end{aligned}$$ The expected regret incurred during batch $1$ and batch $2$ is at most $2\sqrt{T}$. For any $k\geq 3$, assuming holds for batch $k$, and and hold for batch $(k-1)$, we have that $$\begin{aligned} R_k &\le 4\alpha \gT_k {\mathop{\mathbb{E}}}_{X \sim \gD_{k-1}}\max_{\vx \in X} \sqrt{\vx^\top \mLambda_{k-1}^{-1} \vx} \leq \frac{4\sqrt{32}\alpha \gT_k }{\sqrt{\gT_{k-1}}} {\mathop{\mathbb{E}}}_{X \sim \gD_{k-1}}\max_{\vx \in X} \sqrt{\vx^\top \left(T^{-1} \mI + \info_{\gD_{k-2}}(\pi_{k-2})\right)^{-1} \vx} \nonumber \\ &\leq 32 \alpha \sqrt{T} \cdot {\mathop{\mathbb{E}}}_{X \sim \gD_{k-2}}\max_{\vx \in X} \sqrt{\vx^\top \left(T^{-1} \mI + \info_{\gD_{k-2}}(\pi_{k-2})\right)^{-1} \vx} \label{eq:blinucbdg-10} \\ &\leq 32 \alpha \sqrt{T} \cdot \widetilde{\val}^{(1/T)}_{\gD_{k-1}}(\pi_{k-1}) \leq O(\sqrt{dT \log d \log (dKT/\delta)}), \notag \end{aligned}$$ where is because that $X\sim \gD_{k-1}$ can be sampled via first drawing $X' \sim \gD_{k-2}$, then performing one-step elimination on $X'$, and getting $X \subseteq X'$. Finally, collecting the failure probabilities for all $O(\log \log T)$ batches, we prove the desired regret bound. Rarely Switching Algorithm for Adversarial Contexts and $\ln K \leq o(d)$ {#sec:rarely-switch-suplinucb} ========================================================================= @abbasi2011improved showed an algorithm for adversarial contexts that achieves $d\sqrt{T} \times \mathrm{poly}\log T$ regret for any $K$. The authors also propose a special doubling trick that only updates the policy when the determinant of the corresponding information matrix (i.e., $\lambda \mI + \sum_{t} \vx_{t, i_t} \vx_{t, i_t}^\top$) doubles. Using this trick, their algorithm only uses $O(d \log T)$ policy switches, while still achieving the same order of regret. However, when $\log K \ll d$, there is a gap between the regret of their algorithm and the target minimax-optimal regret $\sqrt{T \min\{d, \log K\}} \times \mathrm{poly}\log T$. In this section, we propose an algorithm to close this gap, while still maintaining a small number of policy switches. #### The Natural Approach and its Limitation. The most natural approach is to apply the determinant-based doubling trick to the minimax-optimal algorithms for fewer number of arms, such as [SupLinUCB]{} [@chu2011contextual] and [SupLinRel]{} [@auer2002using]. However, a direct implementation of such an approach would lead to $O(d \log^2 T)$ policy switches. The reason is that, to replace an $\sqrt{d}$ factor by the $\sqrt{\log K}$ factor in the algorithm by @abbasi2011improved, the state-of-the-art concentration inequalities have to crucially rely on the statistical independence between the noises and the context vectors of the played arms, which is not true in the plain [LinUCB]{} algorithm (and the [OFUL]{} algorithm in [@abbasi2011improved]). In contrast, the concentration inequality used in [@abbasi2011improved] does not require such strong independence, but loses a $\sqrt{d}$ factor when $K$ is small. To ensure the independence, @auer2002finite and @chu2011contextual came up with a more sophisticated layering trick, where each time step is assigned to one of the layers. The layers form a hierarchy and the observations from (the time steps in) each layer give more and more accurate estimates for the mean rewards, as the level of the layer increases. Meanwhile, it is possible to ensure the independence between the observations and the context vectors within each layer, so that the more accurate concentration inequality (e.g., in this paper) can be applied. There are $\Theta(\log T)$ layers in [SupLinUCB]{} and [SupLinRel]{}, where each layer maintains a separate information matrix for the estimation. Therefore, if we directly apply the determinant-based doubling trick, there will be $O(d \log T)$ updates in each layer, leading to $O(d \log^2 T)$ policy updates in total. #### Our Approach. Our approach is a simple combination of both types of algorithms mentioned above. Note that the estimation accuracy of the layers in [SupLinUCB]{} and [SupLinRel]{} starts from $\Omega(1)$ for the first layer, and halves as the level of the layer increases. Therefore, it takes $\Theta(\log T)$ levels to reach the sufficient accuracy of $\sqrt{d/T}$. We also note that by the detailed analysis, if the accuracy provided by a layer is $\varpi$, the regret incurred by the layer can be roughly bounded by $d/\varpi$ (up to poly-logarithmic factors). To reduce the number of layers, in our algorithm, we introduce a special layer, namely layer $0$, which helps to bootstrap the accuracy parameters. More precisely, at layer $0$, we use the concentration inequality by @abbasi2011improved (). Since such an inequality is not as efficient as , the regret incurred by layer $0$ can only be bounded by $d^2/\varpi_0$ (up to poly-logarithmic factors), where $\varpi_0$ is the accuracy parameter for layer $0$. However, since the inequality does not rely on the strong independence assumption, instead of starting from the $\Omega(1)$ accuracy, we may directly set $\varpi_0 = d^{1.5}/\sqrt{T}$, a much smaller value, while the incurred regret is still as desired. From layer $1$, we go back to the normal layer settings as [SupLinUCB]{} and [SupLinRel]{}, and set $\varpi_{\kappa} = \varpi_{\kappa}/2$ for $\kappa = 1, 2, \ldots$. Since the target accuracy is $d/\sqrt{T}$, we now only need $\kappa_0 = O(\mathrm{log}(\varpi_0 / (d/\sqrt T))) = O(\log d)$ layers to achieve the minimax-optimal regret. Together with the determinant-based doubling trick, our algorithm uses only $O(d \log d \log T)$ policy switches. Our [[RarelySwitch-SupLinUCB]{}]{}algorithm is formally presented in . Note that the key difference from [SupLinUCB]{} is at and , where $\alpha_0$ and $\varpi_0$ are specially set. Also, at , a special elimination rule for layer $0$ is implemented, which is different from the elimination rules for the rest of the layers at . We next formally analyze the algorithm. $\kappa_0 \gets \lceil \log d \rceil, \alpha_0 \gets 2 \sqrt{d \ln(2T / \delta)}, \forall \kappa \in \{1, 2, \dots, \kappa_0\}: \alpha_\kappa \gets 10\sqrt{ \ln(2dKT /\delta)}$\[loc:suplinucb-1\] $\forall \kappa \in \{0, 1, 2, \dots, \kappa_0\}: \mLambda_{\kappa 0} \gets \mI, \vxi_{\kappa 0} \gets \vzero, \zeta_\kappa \gets 0$ $ \varpi_0 \gets d^{1.5} / \sqrt{T}, \forall \kappa \in \{1, 2, \dots, \kappa_0\}: \varpi_\kappa \gets \varpi_{\kappa - 1} / 2$\[loc:suplinucb-3\] \[thm:suplinucb\] For any $C \ge 2$, the number of policy switches made by is at most $O(d \log d \log T / \log C)$; with probability $(1-2\delta)$, the expected regret of the algorithm is at most ${O}(C \sqrt{d T} \log d \log T \log(d K T / \delta))$. We first upper bound the number of policy switches. Note that for each $\kappa$, we have that $\ln \det \mLambda_{\kappa, T} \le O(d \log T)$, and $\ln \det \mLambda_{\kappa, 0} = 0$; therefore, $\zeta_\kappa$ is updated at by at most $O(d \log T / \log C)$ times. Since the learning policy is completely decided by $\{\mLambda_{\kappa, \zeta_\kappa}, \vxi_{\kappa, \zeta_\kappa}\}_{\kappa=0}^{\kappa_0}$, we conclude that the policy changes by at most $(\kappa_0 + 1) \cdot O(d \log T / \log C) = O(d \log d \log T /\log C)$ times. We next prove the regret of . Note that when the event specified in holds (which happens with probability at least $1 - \delta$), for any $i \in [K]$ and any time step $t \in [T]$, we have $$\begin{aligned} \abs{\vx_{ti}^\top \hat{\vtheta}_0 - \vx_{ti}^\top \vtheta} \le \omega_{ti}^{(0)},\end{aligned}$$ where we use and that $\lambda = 1$ in . We define $\Psi_{t,\kappa} {\stackrel{\mathrm{def}}{=}}\{\tau \leq t \mid \kappa_\tau = \kappa\}$ to be the set of the time steps assigned to layer $\kappa$ at or before time step $t$. Similar to Lemma 14 in [@auer2002using] and Lemma 4 in [@chu2011contextual], we claim that for each $\kappa \geq 1$ and each time $t$, conditioned on any fixed $\Psi_{t-1, \kappa}$, the corresponding noises $\{r_{\tau} - \vx_{\tau, i_\tau}^\top \vtheta \mid \tau \in \Psi_{t - 1, \kappa}\}$ are independent sub-Gaussian random variables with variance proxy $1$. This is because for $\kappa \geq 1$, $\Psi_{t - 1, \kappa}$ only depends on $\{\omega_{\tau,i}^{(\kappa')}\mid \tau < t, \kappa' \leq \kappa, i \in [K]\}$ and $\{\hat{r}_{\tau,i}^{(\kappa')}\mid \tau < t, \kappa' < \kappa, i \in [K]\}$. While $\{\omega_{\tau,i}^{(\kappa')}\mid \tau < t, \kappa' \leq \kappa, i \in [K]\}$ only depends on the context vectors which are independent from the noises, $\{\hat{r}_{\tau,i}^{(\kappa')}\mid \tau < t, \kappa' < \kappa, i \in [K]\}$ depends on the context vectors and the noises generated from time steps in $\Psi_{t-1, 0} \cup \dots \cup \Psi_{t-1,\kappa-1}$, which is disjoint from $\Psi_{t-1, \kappa}$. Thus, the procedure for generating $\Psi_{t-1,\kappa}$ does not use the noises in the time steps in $\Psi_{t-1,\kappa}$, and therefore the noises are independent sub-Gaussian random variables even when conditioned on $\Psi_{t-1,\kappa}$. Given this statistical independence property, by , we have that with probability at least $1 - \delta$, for any $i \in [K]$, any time step $t \in [T]$ and any $\kappa \in \{1, 2, \dots, \kappa_0\}$, it holds that $$\begin{aligned} \abs{\vx_{ti}^\top \hat{\vtheta}_\kappa - \vx_{ti}^\top \vtheta} \le \omega_{ti}^{(\kappa)}.\end{aligned}$$ Now, summarizing the discussion above, we define the desired event $$\begin{aligned} E {\stackrel{\mathrm{def}}{=}}\{\forall i \in [K], t \in [T], \kappa \in \{0\} \cup [\kappa_0]: \abs{\vx_{ti}^\top \hat{\vtheta}_\kappa - \vx_{ti}^\top \vtheta} \le \omega_{ti}^{(\kappa)}\}, $$ and have that $\Pr[E] \ge 1 - 2 \delta$. Below we will upper bound the expected regret incurred by the algorithm when conditioned on $E$. For each layer $\kappa \in \{0, 1, 2, \dots, \kappa_0\}$, we define the regret incurred during time steps that are assigned to layer $\kappa$ as $$\begin{aligned} R_\kappa {\stackrel{\mathrm{def}}{=}}\sum_{t =1}^T \ind[\kappa_t = \kappa] \cdot (\max_{i \in [K]} \vx_{ti}^\top \vtheta - \vx_{t, i_t}^\top \vtheta),\end{aligned}$$ Since each time step will be assigned to exactly one layer $\kappa \in \{0, 1, \dots, \kappa_0\}$, the total regret is $$\begin{aligned} R^T = \sum_{\kappa = 0}^{\kappa_0} {\mathop{\mathbb{E}}}[R_\kappa]. \label{eq:suplinucb-reg-500}\end{aligned}$$ We will use the following lemmas. \[lem:layersize\] We have the following bounds for the size of each layer, $$\begin{aligned} \abs{\Psi_{T,\kappa}} \le \begin{cases} 8 C (T / d) \ln T \ln(2 T / \delta), & \kappa = 0, \\ 200 C \cdot 4^{\kappa} (T / d^2) \ln T \ln(2 d K T / \delta), & 1 \le \kappa \le \kappa_0 - 1, \\ T, & \kappa = \kappa_0. \end{cases}\end{aligned}$$ \[lem:suplinucb-layer-regret\] When the event $E$ happens, for any $t \in [T], \kappa \in \{1, 2, \dots, \kappa_0\}$, it holds that $$\begin{aligned} \ind[\kappa_t = 0] \cdot (\max_{i \in [K]} \vx_{ti}^\top \vtheta - \vx_{t, i_t}^\top \vtheta)&\le 4 \omega_{t, i_t}^{(0)}, \\ \ind[\kappa_t = \kappa] \cdot (\max_{i \in [K]} \vx_{ti}^\top \vtheta - \vx_{t, i_t}^\top \vtheta) &\le 8 \varpi_{\kappa}.\end{aligned}$$ Now, in light of , we upper bound each $R_\kappa$ (conditioned on $E$). For $R_0$, we have that $$\begin{aligned} R_0 &= \sum_{t \in \Psi_{T,0}} (\max_{i \in [K]} \vx_{ti}^\top \vtheta - \vx_{t, i_t}^\top \vtheta) \le \sum_{t \in \Psi_{T,0}} 4 \omega^{(0)}_{t, i_t} = \sum_{t \in \Psi_{T,0}} 4 \alpha_0 \sqrt{\vx_{t, i_t}^\top \mLambda_{0, \zeta_0}^{-1} \vx_{t, i_t}} \label{eq:suplinucb-reg-1000} \\ & \le \sum_{t \in \Psi_{T,0}} 4 \alpha_0 \sqrt{C \vx_{t, i_t}^\top \mLambda_{0, t - 1}^{-1} \vx_{t, i_t}} \le 4 \alpha_0 \sqrt{C \abs{\Psi_{T,0}} \sum_{t \in \Psi_{T,0}} \vx_{t, i_t}^\top \mLambda_{0, t - 1}^{-1} \vx_{t, i_t}} \label{eq:suplinucb-reg-1100} \\ &\le 4 \alpha_0 \sqrt{2 C d \abs{\Psi_{T,0}} \ln T} \le 16 C \sqrt{dT} \ln T \ln(2 T / \delta), \label{eq:suplinucb-reg-1200}\end{aligned}$$ where the inequality in uses , the first inequality in is due to and the update rule at , the second inequality in uses Cauchy-Schwarz, the first inequality in uses the elliptical potential lemma (), and the second inequality in uses . For $\kappa \in \{1, 2, \dots \kappa_0 - 1\}$, we have that $$\begin{aligned} R_\kappa &= \sum_{t \in \Psi_{T,\kappa}} (\max_{i \in [K]} \vx_{ti}^\top \vtheta - \vx_{t, i_t}^\top \vtheta) \leq \sum_{t \in \Psi_{T,\kappa}} 8 \varpi_{\kappa} = 8 \varpi_{\kappa} \abs{\Psi_{T,\kappa}} \label{eq:suplinucb-reg-2100} \\ &\le 1600 C\cdot 2^{\kappa} \cdot \frac{d^{1.5}}{\sqrt T} \cdot \frac{T}{d^2} \ln T \ln(2 d KT / \delta) \le 3200 C \sqrt{dT} \ln T \ln(2 d KT / \delta), \label{eq:suplinucb-reg-2200}\end{aligned}$$ where the inequality in uses , the first inequality in uses and that $\varpi_{\kappa} = 2^{-\kappa} d^{1.5} / \sqrt T$. For $k = \kappa_0$, we have that $$\begin{aligned} R_{\kappa_0} =\sum_{t \in \Psi_{T,\kappa_0}} (\max_{i \in [K]} \vx_{ti}^\top \vtheta - \vx_{t, i_t}^\top \vtheta) \le 8 \varpi_{\kappa_0} \abs{\Psi_{\kappa_0}} \le 8 \varpi_{\kappa_0} T \le 8 \times \frac{d^{0.5}}{\sqrt T} \cdot T = 8 \sqrt{dT}, \label{eq:suplinucb-reg-3000}\end{aligned}$$ where the first inequality uses , the second inequality uses , the third inequality uses that $\varpi_{\kappa_0} = 2^{-\kappa_0} d^{1.5} / \sqrt T$ and that $2^{\kappa_0} \ge d$. We prove the theorem by plugging , , and back to . Proof of --------- The third bound $\abs{\Psi_{T,\kappa_0}} \le T$ is self-evident, so we only prove the first two bounds. By the elliptical potential lemma (), for every $\kappa \in \{0,1,2,\dots, \kappa_0\}$, we have that $$\begin{aligned} \sum_{t \in \Psi_{T,\kappa}} \vx_{t, i_t}^\top \mLambda_{\kappa, t - 1}^{-1} \vx_{t, i_t} \le 2 d \ln T, \end{aligned}$$ which, together with the Cauchy-Schwarz inequality, implies that $$\begin{aligned} \sum_{t \in \Psi_{T, \kappa}} \sqrt{\vx_{t, i_t}^\top \mLambda_{\kappa, t - 1}^{-1} \vx_{t, i_t}} \le \sqrt{2 d \abs{\Psi_{T,\kappa}} \ln T}. $$ In the following, we use $\zeta_{\kappa,t}$ to denote the value of $\zeta_\kappa$ at of during time step $t$. Note that we have $\zeta_{\kappa,t} \le t - 1$ for every $\kappa$. By our update rule (), we have that $\det \mLambda_{\kappa,t - 1} \le C\det \mLambda_{\kappa,\zeta_{\kappa,t}} $ for every $t \in [T]$. Therefore, for each $t \in [T]$ and $\kappa \in \{0,1, \dots \kappa_0 - 1\}$ such that $\kappa_t = \kappa$, together with , we have that $$\begin{aligned} \sqrt{C \vx_{t, i_t}^\top \mLambda_{\kappa, t - 1}^{-1} \vx_{t, i_t}} \ge \sqrt{\vx_{t, i_t}^\top \mLambda_{\kappa, \zeta_{\kappa,t}}^{-1} \vx_{t, i_t}} \ge \frac{\varpi_\kappa}{\alpha_\kappa}, $$ where the last inequality is by . Therefore, for each $\kappa \in \{0,1, \dots \kappa_0 - 1\}$, we have that $$\begin{aligned} \sqrt{2 C d \abs{\Psi_{T, \kappa}} \ln T} \ge \abs{\Psi_{T,\kappa}} \cdot \frac{\varpi_\kappa}{\alpha_\kappa},\end{aligned}$$ which implies that (for $\kappa \in \{1, 2, \dots, \kappa_0 -1\}$) $$\begin{aligned} \abs{\Psi_{T,\kappa}} \le (\frac{\alpha_{\kappa} \sqrt{2 C d \ln T}}{\varpi_\kappa})^2 \le \frac{200 C d \ln T \ln (2 d K T / \delta)}{4^{-\kappa} d^3 / T } =200 C \cdot 4^{\kappa} (T / d^2) \ln T \ln(2 d K T / \delta), $$ and $$\begin{aligned} \abs{\Psi_{T,0}} \le (\frac{\alpha_0 \sqrt{2 C d \ln T}}{\varpi_0})^2 \le \frac{8 C d^2 \ln T \ln(2 T / \delta)}{d^3 / T } = 8 C (T / d) \ln T \ln(2 T / \delta). $$ Proof of --------- is a direct corollary of the following two lemmas, which we prove separately in this subsection. \[lem:layer\] When the event $E$ happens, for any $t \in [T], \kappa \in \{1, 2, \dots, \kappa_0\}$, it holds that $$\begin{aligned} \ind[\kappa_t = 0] \cdot (\max_{i \in A_t^{(0)}} \vx_{ti}^\top \vtheta - \vx_{t, i_t}^\top \vtheta) &\le 4 \omega_{t, i_t}^{(0)}, \\ \ind[\kappa_t = \kappa] \cdot (\max_{i \in A_t^{(\kappa)}} \vx_{ti}^\top \vtheta - \vx_{t, i_t}^\top \vtheta) &\le 8 \varpi_{\kappa}.\end{aligned}$$ \[lem:suplinucb-elim\] When the event $E$ happens, for any $t \in [T]$, and all $\kappa$ such that $\kappa \leq \kappa_t$, we have that $$\begin{aligned} \max_{i \in [K]} \vx_{ti}^\top \vtheta = \max_{i \in A_t^{(\kappa)}} \vx_{ti}^\top \vtheta. \end{aligned}$$ For the inequality, assuming that $\kappa_t = 0$, we have $i_t = \argmax_{i \in A_t^{(0)}} \omega_{t i}^{(0)}$. Then we have that $$\begin{aligned} \max_{i \in A_t^{(0)}} \vx_{ti}^\top \vtheta - \vx_{t, i_t}^\top \vtheta &\le \max_{i \in A_t^{(0)}} \vx_{ti}^\top \vtheta - \min_{i \in A_t^{(0)}} \vx_{t i}^\top \vtheta \\ &\le \max_{i \in A_t^{(0)}}\{\vx_{ti}^\top \hat{\vtheta}_{0} + \omega_{ti}^{(0)}\} - \min_{i \in A_t^{(0)}}\{\vx_{t i}^\top \hat{\vtheta}_0 - \omega_{ti}^{(0)}\} \\ &\le 4 \max_{i \in A_t^{(0)}} \omega_{ti}^{(0)} = 4 \omega_{t, i_t}^{(0)},\end{aligned}$$ where the second inequality is because of event $E$ and the third inequality follows from the elimination rule at , and the last equality is due to . For the second inequality, assuming that $\kappa_t = \kappa \geq 1$, we have that $$\begin{aligned} \max_{i \in A_t^{(\kappa)}} \vx_{ti}^\top \vtheta - \vx_{t, i_t}^\top \vtheta &\le \max_{i \in A_t^{(\kappa-1)}}\{\vx_{ti}^\top \hat{\vtheta}_{\kappa-1} + \omega_{ti}^{(\kappa-1)}\} - \min_{i \in A_t^{(\kappa-1)}}\{\vx_{t i}^\top \hat{\vtheta}_{\kappa-1} - \omega_{ti}^{(\kappa-1)}\} \\ &\le2 \max_{i \in A_t^{(\kappa-1)}} \omega_{ti}^{(\kappa-1)} + \max_{i \in A_{t}^{(\kappa-1)}} \hat r_{ti}^{(\kappa-1)} - \min_{i \in A_t^{(\kappa-1)}} \hat r_{ti}^{(\kappa-1)} \\ &\le2 \max_{i \in A_t^{(\kappa-1)}} \omega_{ti}^{(\kappa-1)} + 2 \varpi_{\kappa-1} \leq 4\varpi_{\kappa - 1} = 8 \varpi_\kappa,\end{aligned}$$ where the first inequality is by the event $E$, the third inequality follows from that $\min_{i \in A_{t}^{(\kappa-1)}} \hat{r}_{ti}^{(\kappa-1)} \ge \max_{i \in A_{t}^{(\kappa-1)}} \hat r_{ti}^{(\kappa-1)} - 2 \varpi_{\kappa-1}$ as implied by of , the last inequality is because the condition at was met at iteration $\kappa - 1$ (since otherwise the loop should have terminated at iteration $\kappa - 1$). For any time step $t \in [T]$, note that when the event $E$ holds, by the elimination rule at , we have that $$\begin{aligned} \max_{i \in [K]} \vx_{ti}^\top \vtheta = \max_{i \in A_t^{(0)}} \vx_{ti}^\top \vtheta.\end{aligned}$$ Also, for each $\kappa < \kappa_t$, by the elimination rule at , we have that $$\begin{aligned} \max_{i \in A_t^{(\kappa)}} \vx_{ti}^\top \vtheta = \max_{i \in A_t^{(\kappa+1)}} \vx_{ti}^\top \vtheta,\end{aligned}$$ Applying the equality iteratively for $\kappa = 0, 1, 2, \dots, \kappa_t - 1$, and we prove the lemma. Lower Bounds for Adversarial Contexts {#sec:lb} ===================================== In this section, we prove the following lower bound for the number of policy switches in the adversarial context setting. \[thm:lb\] Let $K = 2$, for any even number of dimensions $d \geq 2$, and $T$ greater than a sufficiently large constant times $d$, suppose the expected number of policy switches made by the learner is at most $M$ ($20d \leq M \leq (d \ln T)/ 48$), then there exists a bandit instance such that the learner’s expected regret on the instance is at least $\sqrt{dT} \times \frac{1}{32d}\left(\frac{2T}{d}\right)^{1/(16M/d + 2)}$. shows that, even for $K = 2$, when $T \geq d^2$, in order to achieve $\sqrt{dT} \times \mathrm{poly}\log T$ regret, $M$ has to be $\Omega(d \log T/\log (d\log T))$. Note that on the upper bound side, our and the algorithm by [@abbasi2011improved] achieves $C$ times the target minimax-optimal regret (up to $\mathrm{poly} \log T$ factors) with $O((d \log d\log T) / \log C)$ policy switches, and our shows that $\Omega(\frac{d\log T}{\log C + \log (d \log T)})$ policy switches are needed, almost matching the upper bound for every $C$. To prove , we first prove the lower bound in the special case of $d = 2$ in . Then, in , we prove the theorem for general $d$ using the special case as a building block. Lower Bound for Constant-Dimension Special Case {#sec:lb-2-d} ----------------------------------------------- \[lem:lb\] When $K=d=2$, for sufficiently large $T$, suppose the expected number of policy switches made by the learner is at most $M$ ($40 \leq M \leq (\ln T)/24$), then there exist a bandit instance such that the learner’s expected regret on the instance is at least $T^{1/2+1/(8M+2)}/32$. To prove , we will construct a class of bandit problem instances $\mathfrak{B} = \{B^{(\vu)}\}$, where each instance $B^{(\vu)}$ is parameterized by $\vu \in \{\pm 1\}^{L}$ and $L = 4M$. For any fixed learner with no more than $M$ policy switches, we will show that the regret averaged over the $2^L$ instances in $\mathfrak{B}$ is large, and therefore there exists at least one instance in $\mathfrak{B}$ that is bad for the learner. For each $B^{(\vu)}$, we assume that the noises are independent centered Gaussian with variance $1$. We also need to define the hidden vector $\vtheta^{(\vu)}$ and the context vectors $\{\vx_{t,1}^{(\vu)}, \vx_{t,2}^{(\vu)}\}_{t = 1}^{T}$ (where, in our formal definition of linear bandits, $\gD_t$ is the deterministic distribution supported on $\{\vx_{t,1}^{(\vu)}, \vx_{t,2}^{(\vu)}\}$ for every $t$). Before defining $\mathfrak{B}$, we first divide the time steps into stages, and define a few helpful notations. #### Stages. We uniformly divide the $T$ time steps into $L$ stages. Let $t_j {\stackrel{\mathrm{def}}{=}}\lceil jT/L \rceil$ for all $j \in \{0, 1, 2, \dots, L\}$. The $j$-th stage consists of the time steps in the range $(t_{j-1}, t_j]$. #### Additional Notations. Let $\upsilon {\stackrel{\mathrm{def}}{=}}\sqrt{T}^{-1/(L+1)}$. Note that $\upsilon \leq 1/10$ since $L \leq (\ln T) / 6$. For each $\vu = (u_1, u_2, \dots, u_L) \in \{\pm 1\}^L$ and each $j \in \{ 1, 2, \dots, L\}$, we define the map $\psi_j(\vu) {\stackrel{\mathrm{def}}{=}}1/2 + \sum_{i=1}^{j} u_i \cdot \upsilon^i$ that sends the sequence to the decimal. We have that $\psi_j(\vu) \leq 2/3$ since $\upsilon \leq 1/10$. For convenience, we also define $\psi(\vu) {\stackrel{\mathrm{def}}{=}}\psi_L(\vu)$. For each $j$, we also define $z_j {\stackrel{\mathrm{def}}{=}}\upsilon^{-(j+1)}/\sqrt{T} \leq 1$. #### Bandit Instances. We now define $B^{(\vu)}$ for each $\vu \in \{\pm 1\}^{L}$. For the hidden vector, we let $\vtheta^{(\vu)} {\stackrel{\mathrm{def}}{=}}(\psi(\vu), \frac23 )^\top$. For every stage $j$, and every time step $t$ during stage $j$, we set the context vectors by $\vx_{t, 1}^{(\vu)} {\stackrel{\mathrm{def}}{=}}(z_j, 0)^\top$ and $\vx_{t, 2}^{(\vu)} {\stackrel{\mathrm{def}}{=}}(0, \frac{3}{2} z_j \cdot \psi_{j-1}(\vu))^\top$. One can easily verify that the norms of all vectors are upper bounded by $1$. We now start analyzing the constructed instances. #### Suboptimal Action and its Regret. Since there are only two candidate actions during each time step, we refer to the one with smaller expected reward as the suboptimal action. The following lemma lower bounds the expected regret incurred by playing a suboptimal action. \[lem:subopt-action\] For any instance $B^{(\vu)}$, and any time step $t$, the regret incurred by playing the suboptimal action at time step $t$ is at least $ \upsilon^{-1}/(2\sqrt{T})$. Suppose that time step $t$ is in stage $j$. The regret incurred by the suboptimal action is $$\begin{aligned} \abs{(z_j,0)^\top \vtheta ^{(\vu)}-(0,\frac32 z_j \cdot \psi_{j-1}(\vu) )^\top \vtheta ^{(\vu)}} &= z_j \cdot \abs{\psi(\vu)-\psi_{j-1}(\vu)} \nonumber \\ &\ge z_j\cdot\left(\upsilon^j - \sum_{i=j+1}^{+\infty} \upsilon^i\right) \geq \frac{ z_j \upsilon^j}{2} \geq \frac{\upsilon^{-1}}{2\sqrt{T}}.\nonumber \qedhere\end{aligned}$$ #### The Regret of a Rarely Switching Learner. For any learner who switches the policy for at most $M$ times, let $F_j$ be the event that the policy is not switched during stage $j$. Let $E_j$ be the event that the learner’s policy $\chi_t$ places greater or equal to $1/2$ probability mass on the suboptimal action at time $t$, where $t = t_{j-1}+1$ is the first time step of stage $j$. By , the expected regret of the learner for bandit instance $B^{(\vu)}$ can be lower bounded by $$\begin{aligned} \label{eq:lb-regret-rarely-switch-first-step} {\mathop{\mathbb{E}}}_{B^{(\vu)}}[R^T] \geq \sum_{j=1}^L \sum_{t=t_{j-1}+1}^{t_j} \Pr_{B^{(\vu)}} [E_j \cap F_j] \cdot \frac{1}{2} \cdot \frac{\upsilon^{-1}}{2\sqrt{T}} = \frac{1}{4\upsilon\sqrt{T}} \cdot (T/L) \cdot \sum_{j=1}^L \Pr_{B^{(\vu)}} [E_j \cap F_j],\end{aligned}$$ where ${\mathop{\mathbb{E}}}_{B^{(\vu)}}[\cdot]$ denotes the expectation taken over the probability distribution induced by the learner and the bandit instance $B^{(\vu)}$ (and we similarly define $\Pr_{B^{(\vu)}}[\cdot]$). Let $p_j^{(\vu)} {\stackrel{\mathrm{def}}{=}}\Pr_{B^{(\vu)}} [E_j]$, continuing with , we have $$\begin{aligned} {\mathop{\mathbb{E}}}_{B^{(\vu)}}[R^T] \geq \frac{\sqrt{T}}{4\upsilon L} \sum_{j=1}^L \left(\Pr_{B^{(\vu)}} [E_j] - \Pr_{B^{(\vu)}} [\overline{F_j}]\right) \geq \frac{\sqrt{T}}{4\upsilon L}\left( \sum_{j=1}^L p_j^{(\vu)} - M\right), \label{eq:lb-regret-rarely-switch-first-step-a}\end{aligned}$$ where $\overline{F_j}$ denotes the complement event of $F_j$ and the last inequality is because that the learner can switch in at most $M$ stages (in expectation). #### Probability of Playing a Suboptimal Action. By the discussion above, to lower bound the regret, we need to lower bound $p_j^{(\vu)}$. We first prove the following lemma. \[lem:lb-prob-difference-bound\] Consider any $\vu = (u_1, u_2, \dots, u_L) \in \{\pm 1\}^L$ and $\vu' = (u_1', u_2', \dots, u_L') \in \{\pm 1\}^L$. Suppose $\vu \neq \vu'$, let $j$ be the smallest index such that $u_j \neq u_j'$. For any event $E$, we have $$\begin{aligned} \abs{\Pr_{B^{(\vu)}}[E] - \Pr_{B^{(\vu')}}[E] } \leq 0.25 .\end{aligned}$$ Let $t = t_{j-1}$ be the last time step before stage $j$. We will consider the sample space $\Omega_t$ that consists of the trajectories $(i_1, r_1, \dots, i_t, r_t)$ and the internal randomness source $\vs$ used by the learner. Now consider two probability distributions $D$ and $D'$ over $\Omega_t$, where $D$ is induced by the learner and the instance $B^{(\vu)}$, and $D'$ is induced by the learner and $B^{(\vu')}$. We will show that $$\begin{aligned} \label{eq:lb-KL-single-stage-bound} \mathrm{KL}(D \Vert D') \leq 0.1,\end{aligned}$$ where $\mathrm{KL}(\cdot \Vert \cdot)$ denotes the Kullback–Leibler (KL) divergence between the two distributions, so that we can prove the lemma by invoking Pinsker’s inequality (). We now prove . Fix any $\vs$, let $D_\vs$ be $D$ conditioned on $\vs$ and let $D'_\vs$ be $D'$ conditioned on $\vs$. Since $D$ and $D'$ share the same marginal distribution on $\vs$, to prove , we only need to show $$\begin{aligned} \label{eq-lb-prob-difference-bound-1} \mathrm{KL}(D_\vs \Vert D'_\vs) \leq 0.1.\end{aligned}$$ Let $q(i_1, r_1, \dots, i_t, r_t)$ and $q'(i_1, r_1, \dots, i_t, r_t)$ be the probability density functions for $D$ and $D'$ respectively. We have that $$\begin{aligned} q(i_1, r_1, \dots, i_t, r_t) &= \prod_{\tau=1}^{t} \ind[i_\tau = i_\tau(i_1, r_1, \dots, i_{\tau-1}, r_{\tau-1}; \vs)] \cdot q^{(\vu)}(r_\tau \| i_\tau), \label{eq-lb-prob-difference-bound-10} \\ \text{and}\qquad q'(i_1, r_1, \dots, i_t, r_t) &= \prod_{\tau=1}^{t} \ind[i_\tau = i_\tau(i_1, r_1, \dots, i_{\tau-1}, r_{\tau-1}; \vs)] \cdot q^{(\vu')}(r_\tau \| i_\tau), \label{eq-lb-prob-difference-bound-20}\end{aligned}$$ where $i_\tau(i_1, r_1, \dots, i_{\tau-1}, r_{\tau-1}; \vs)$ is the deterministic decision of the learner at time $\tau$ given the trajectory $(i_1, r_1, \dots, i_{\tau-1}, r_{\tau-1})$ and $\vs$, and $q^{(\vu)}(r_\tau \vert i_\tau)$ is the probability density function for the reward at time $t$, if playing action $i_\tau$ in instance $B^{(\vu)}$. Since the second dimensions of $\vtheta^{(\vu)}$ and $\vtheta^{(\vu')}$ are the same, the difference of the mean reward at any time step in stage $j' < j$ for the same action in $B^{(\vu)}$ and $B^{(\vu')}$ is either $\|\psi(\vu) z_j' - \psi(\vu') z_j'\|$ (if the first action is played) or $0$ (if the second action is played, since $\psi_{j-1}(\vu) = \psi_{j-1}(\vu')$). Since the rewards are Gaussian with variance $1$, and the KL divergence between two variance-1 Gaussian variables with means $\mu_1$ and $\mu_2$ is $(\mu_1-\mu_2)^2/2$, for any $\tau \leq t$ that is in stage $j'$ and any $i_\tau \in \{1, 2\}$, we have that $$\begin{aligned} \mathrm{KL}\left(q^{(\vu)}(\cdot\vert i_\tau) ~\Vert~ q^{(\vu')}(\cdot\vert i_\tau)\right) \leq z_{j'}^2 \cdot \frac{ (\psi(\vu) - \psi(\vu'))^2}{2} \leq z_{j'}^2 \cdot \frac{(4 \upsilon^{j})^2}{2} = 8 z_{j'}^2 \upsilon^{2j}, \label{eq-lb-prob-difference-bound-30}\end{aligned}$$ where the last inequality is because that $j$ is the first index where $\vu$ and $\vu'$ differ and that $\upsilon \leq 1/10$. By , , , and the Chain Rule for KL divergence, we have that $$\begin{aligned} \mathrm{KL}(D_\vs \Vert D_{\vs'}) \leq \sum_{j'=1}^{j-1} \sum_{\tau=t_{j'-1}+1}^{t_{j'}} 8 z_{j'}^2 \upsilon^{2(j+1)} & = \sum_{j'=1}^{j-1} \sum_{\tau=t_{j'-1}+1}^{t_{j'}} 8\cdot \frac{\upsilon^{-2j'-2}}{T} \cdot \upsilon^{2j}\\ &\leq \frac{8(T/L)}{T} \cdot 2 \upsilon^{-2j} \cdot \upsilon^{2j} = \frac{16}{L} \leq 0.1,\end{aligned}$$ proving . We now bound $p_j^{(\vu)}$ by the following lemma. \[lem:lb-paired-lowerbound\] Consider any $\vu = (u_1, u_2, \dots, u_L) \in \{\pm 1\}^L$ and $\vu' = (u_1', u_2', \dots, u_L') \in \{\pm 1\}^L$. Suppose $\vu \neq \vu'$, let $j$ be the smallest index such that $u_j \neq u_j'$. Then we have $p_j^{(\vu)}+p_j^{(\vu')}\ge 0.75$. Since $j$ is the smallest index such that $u_j \neq u_j'$, by our construction, exactly one of $\psi_j(\vu)$ and $\psi_j(\vu')$ is greater than $\psi_{j-1}(\vu)$, which means, at stage $j$, any action that is suboptimal for instance $B^{(\vu)}$ is optimal for instance $B^{(\vu')}$, and vice versa. Let $t = t_{j-1}+1$ be the first time step in stage $j$. Let $E$ be the event that the learner’s policy $\chi_t$ for time step $t$ assigns at least $1/2$ probability mass to the suboptimal action for $B^{(\vu)}$. Since $B^{(\vu)}$ and $B^{(\vu')}$ share the same context vector set at time step $t$ (because $\psi_{j-1}(\vu) = \psi_{j-1}(\vu')$), the complement event $\overline{E}$ is that $\chi_t$ assigns at least $1/2$ probability mass to the suboptimal action for $B^{(\vu')}$. Invoking , we have that $$p_t^{(\vu)}+p_t^{(\vu')} = \Pr_{B^{(\vu)}}[E] + \Pr_{B^{(\vu')}}[\overline{E}] = 1 + \Pr_{B^{(\vu)}}[E] - \Pr_{B^{(\vu')}}[E] \geq 0.75. \qedhere$$ #### Putting Things Together and the Average Case Analysis. Note that $$\begin{aligned} &\frac{1}{2^L} \sum_{\vu \in \{\pm 1\}^L} \sum_{j=1}^{L} p_j^{(\vu)} = \frac{1}{2^L} \sum_{j=1}^L \sum_{\vu \in \{\pm 1\}^L} p_j^{(\vu)} \nonumber \\ &\qquad\qquad\qquad\qquad\qquad\qquad = \frac{1}{2^L} \sum_{j=1}^L \sum_{\vu \in \{\pm 1\}^L} \frac{p_j^{(\vu)} + p_j^{(\vu^{\oplus j})}}{2} \geq \frac{1}{2^L} \sum_{j=1}^{L} 2^L \cdot \frac{0.75}{2} = 0.375 L, \label{eq:lb-avg-case}\end{aligned}$$ where $\vu^{\oplus j}$ is the $\pm 1$ sequence derived by flipping the sign of the $j$-th element of $\vu$, and the inequality is due to . Therefore, there exists $\vu^$ such that $\sum_{j=1}^{L} p_j^{(\vu^)} \geq 0.375 L$. Together with , we have $$\begin{aligned} {\mathop{\mathbb{E}}}_{B^{(\vu^)}}[R^T] \geq \frac{\sqrt{T}}{4 \upsilon L} \times (0.375L - M) = \frac{\sqrt{T}}{4 \upsilon \cdot 4M} \times (0.375 \cdot 4M - M) = \frac{1}{32} \cdot T^{1/2 + 1/(8M+2)},\end{aligned}$$ proving . Proof of for General Dimensions {#sec:lb-general} ------------------------------- We equally divide the $T$ time steps into $d/2$ intervals. We construct the class of bandit instances, $\widetilde{\mathfrak{B}}$ from the class $\mathfrak{B}$ constructed in the proof of as follows. For each $\ell \in \{1, 2, \dots, d/2\}$, we choose a bandit instance $B_\ell$ from $\mathfrak{B}$, and construct the $d$-dimensional instance $\widetilde{B}$. The hidden vector $\vtheta$ of $\widetilde{B}$ is derived by concatenating the hidden vectors of the $d/2$ smaller instances. During the $\ell$-th interval of time, we use the context vectors in $B_\ell$ in order: for each time step in the $\ell$-th interval, we put the $2$-dimensional context vectors in the corresponding time step in $B_\ell$ at the $(2\ell - 1)$-th and $2\ell$-th entries, while filling other entries with $0$. $\widetilde{\mathfrak{B}}$ will consist of all possible instances that can be constructed in this way, and we have $\abs{\widetilde{\mathfrak{B}}} = \abs{\mathfrak{B}}^{d/2}$. By our construction, the rewards from different time intervals are completely independent. Since the length of an interval is $T/(d/2)$, if we let $L = 4M/(d/2)$ and $\upsilon = \sqrt{T/(d/2)}^{-1/(L+1)}$, we can prove similarly as that $$\begin{aligned} \frac{1}{\abs{\widetilde{\mathfrak{B}}}} \sum_{\widetilde{B} \in \widetilde{\mathfrak{B}}} {\mathop{\mathbb{E}}}_{\widetilde{B}}[R_\ell] \geq \frac{\sqrt{T/(d/2)}}{4\upsilon L} \times \left(0.375L - \frac{1}{\abs{\widetilde{\mathfrak{B}}}} \sum_{\widetilde{B} \in \widetilde{\mathfrak{B}}} M_{\widetilde{B},\ell}\right), \nonumber\end{aligned}$$ where $R_\ell$ is the regret incurred during the $\ell$-th interval, and $M_{\widetilde{B},\ell}$ is the expected number of policy switches made during the $\ell$-th interval, when given instance $\widetilde{B}$. Therefore, we have that $$\begin{aligned} \frac{1}{\abs{\widetilde{\mathfrak{B}}}} \sum_{\widetilde{B} \in \widetilde{\mathfrak{B}}} {\mathop{\mathbb{E}}}_{\widetilde{B}}[R^T] & = \sum_{\ell=1}^{d/2} \frac{1}{\abs{\widetilde{\mathfrak{B}}}} \sum_{\widetilde{B} \in \widetilde{\mathfrak{B}}} {\mathop{\mathbb{E}}}_{\widetilde{B}}[R_\ell] \geq \frac{ \sqrt{T/(d/2)}}{4\upsilon L} \times \left(\frac{3dL}{16} -\frac{1}{\abs{\widetilde{\mathfrak{B}}}}\sum_{\widetilde{B} \in \widetilde{\mathfrak{B}}} \sum_{\ell=1}^{d/2} M_{\widetilde{B},\ell}\right) \nonumber \\ & \geq \frac{ \sqrt{T/(d/2)}}{4\upsilon L} \times \left(\frac{3dL}{16} - M\right) \geq \frac{\sqrt{T/(d/2)}}{4\upsilon \cdot 4M} \times \frac{M}{2} \geq \frac{\sqrt{T/d}}{32} \times \left(\frac{2T}{d}\right)^{1/(16M/d + 2)} , \nonumber\end{aligned}$$ which means that there exists at least one instance from $\widetilde{\mathfrak{B}}$ such that the learner incurs at least $\sqrt{dT} \times \frac{1}{32d}\left(\frac{2T}{d}\right)^{1/(16M/d + 2)}$ expected regret, proving the theorem. Acknowledgments {#acknowledgments .unnumbered} =============== We thank Yanjun Han, Zhengyuan Zhou, and Zhengqing Zhou for their valuable comments. Technical Lemmas ================ Concentration Inequalities {#app:concentration} -------------------------- \[lem:hoeffding\] Let $X_1, \dots, X_n \in [0, R]$ be independent bounded random variables. Let $\overline{X} = \frac{1}{n}\sum_{i = 1}^n X_i$ be their average. Then $$\begin{aligned} \Pr[\abs{\overline{X} - {\mathop{\mathbb{E}}}\overline{X}} \ge \delta] \le 2\exp(- \frac{2 n \delta^2}{R^2}).\end{aligned}$$ \[lem:mchernoff\] Let $\mX_1, \dots, \mX_n$ be a sequence of independent positive semi-definite random matrices in dimension $d$ such that $\norm{\mX_i} \le R$ almost surely (where $\norm{\cdot}$ denotes the operator norm). Let $\mX = \sum_{i = 1}^n \mX_i$ be their sum. Let $\mu_{\min} = \lambda_{\min}({\mathop{\mathbb{E}}}\mX)$ and $\mu_{\max} = \lambda_{\max}({\mathop{\mathbb{E}}}\mX)$. Then we have $$\begin{aligned} \Pr[\lambda_{\min}(\mX) \le (1 - \delta) \mu_{\min}] \le d \left( \frac{e^{-\delta}}{(1 - \delta)^{1 - \delta}}\right)^{\mu_{\min} / R}, &\qquad \mathrm{when}~\delta \in [0, 1], \\ \Pr[\lambda_{\max}(\mX) \ge (1 + \delta) \mu_{\max}] \le d \left( \frac{e^{\delta}}{(1 + \delta)^{1 + \delta}}\right)^{\mu_{\max} / R}, &\qquad \mathrm{when}~\delta \ge 0.\end{aligned}$$ \[lem:covcon\] Suppose $\vx_1, \dots, \vx_{n} \sim \gD$ are [i.i.d.]{} drawn from a distribution $\gD$ and $\vx_i^\top \vx_i \leq 1$ almost surely. Let $\lambda = \lambda_{\min}({\mathop{\mathbb{E}}}_{\vx \sim \gD}{\vx \vx^\top}) > 0$ be the smallest eigenvalue of the co-variance matrix. We have that $$\begin{aligned} \Pr[\frac{1}{n} \sum_{i = 1}^n \vx_i \vx_i^\top \succcurlyeq \frac{1}{2} {\mathop{\mathbb{E}}}_{\vx \sim \gD} \vx \vx^\top] \ge 1 - d \exp(-\frac{\lambda n}{8}). \end{aligned}$$ Let $\mSigma = {\mathop{\mathbb{E}}}_{\vx \sim \gD}{\vx \vx^\top}$ and $\vy_i = \mSigma^{-1/2} \vx_i$ for all $i \in [n]$. Note that $\norm{\vy_i \vy_i^\top} \leq \lambda^{-1}$ almost surely, and that ${\mathop{\mathbb{E}}}_{\vy_i} \vy_i \vy_i^\top = \mI$. Therefore, by , we have that $$\begin{aligned} 1 - d\exp(-\frac{\lambda n}{8}) &\leq \Pr[\frac{1}{n}\sum_{i=1}^{n} \vy_i\vy_i^\top \succcurlyeq \frac{1}{2} \mI] = \Pr[\frac{1}{n}\sum_{i=1}^{n} \vy_i\vy_i^\top \succcurlyeq \frac{1}{2} {\mathop{\mathbb{E}}}\mSigma^{-1/2} \vx\vx^\top \mSigma^{-1/2}] \\ &= \Pr[\frac{1}{n} \sum_{i = 1}^n \vx_i \vx_i^\top \succcurlyeq \frac{1}{2} {\mathop{\mathbb{E}}}_{\vx \sim \gD} \vx \vx^\top]. \qedhere\end{aligned}$$ \[lem:conconcut\] Suppose $\vx_1, \dots, \vx_{n} \sim \gD$ are [i.i.d.]{} drawn from a distribution $\gD$ and $\vx_i^\top \vx_i \leq 1$ almost surely. For any cutoff level $\lambda > 0$, with probability at least $(1 - 2d \exp(- \frac{\lambda n}{8}))$, we have that $$3 \lambda \mI + \frac{1}{n} \sum_{i = 1}^n \vx_i \vx_i^\top \succcurlyeq \frac{1}{8}{\mathop{\mathbb{E}}}_{\vx \sim \gD} \vx \vx^\top.$$ Suppose ${\mathop{\mathbb{E}}}_{\vx \sim \gD} \vx \vx^\top = \sum_{i=1}^d \lambda_i \vv_i \vv_i^\top$ where $\{\vv_i\}_{i=1}^d$ is a set of orthonormal basis. Let $\mP_{+} = \sum_{i=1}^d \vv_i \vv_i^\top \ind[\lambda_i \geq\lambda]$ and $\mP_{-} = \sum_{i=1}^d \vv_i \vv_i^\top \ind[\lambda_i <\lambda]$, so that $\mI = \mP_+ \mP_-$. Observe that the eigenvalues of ${\mathop{\mathbb{E}}}_{\vx \sim \gD} \mP_{+} \vx \vx^\top \mP_+^\top$ are greater or equal to $\lambda$ when restricted to the space spanned by the $\mP_+$. Therefore, by , we have that with probability at least $(1 - d \exp(-\lambda n/8))$, $$\begin{aligned} \frac{1}{n} \sum_{i=1}^n (\mP_+\vx_i) (\mP_+\vx_i)^\top \succcurlyeq \frac{1}{2} {\mathop{\mathbb{E}}}_{\vx \sim \gD} \mP_{+} \vx \vx^\top \mP_+^\top. \label{eq:lem-conconcut-1}\end{aligned}$$ Note that $$\begin{aligned} &\quad \frac{1}{n} \sum_{i = 1}^n \vx_i \vx_i^\top = \frac{1}{n} (\mP_+ \mP_-) \sum_{i = 1}^n \vx_i \vx_i^\top (\mP_+ \mP_-)^\top \notag \\ &= \frac{1}{2n}\sum_{i=1}^n \mP_{+} \vx_i \vx_i^\top \mP_+^\top + \frac{1}{n} \sum_{i=1}^n \left(\frac{1}{2} \mP_{+} \vx_i \vx_i^\top \mP_+^\top + \mP_{+} \vx_i \vx_i^\top \mP_-^\top + \mP_{-} \vx_i \vx_i^\top \mP_+^\top + 2 \mP_{-} \vx_i \vx_i^\top \mP_-^\top\right) \nonumber \\ & \qquad - \frac{1}{n} \mP_{-} \vx_i \vx_i^\top \mP_-^\top, \label{eq:lem-conconcut-2}\end{aligned}$$ where the first term is $\succcurlyeq \frac{1}{4} {\mathop{\mathbb{E}}}_{\vx \sim \gD} \mP_{+} \vx \vx^\top \mP_+^\top$ by , the second term is a sum of positive semi-definite matrices, and for the third term, by , with probability is at least $(1 - d \exp(-\lambda n /3)$, we have that $$\begin{aligned} \frac{1}{n} \sum_{i=1}^n \mP_{-} \vx_i \vx_i^\top \mP_-^\top \preccurlyeq 2 \lambda \mI. \end{aligned}$$ Therefore, continuing with , and collecting probabilities, we have that with probability at least $(1 - 2d \exp(-\lambda n /8))$, $$\begin{aligned} \frac{1}{n} \sum_{i = 1}^n \vx_i \vx_i^\top& \succcurlyeq \frac{1}{4} {\mathop{\mathbb{E}}}_{\vx \sim \gD} \mP_{+} \vx \vx^\top \mP_+^\top - 2 \lambda \mI \\ &= \frac{1}{8} {\mathop{\mathbb{E}}}_{\vx \sim \gD} \vx \vx^\top + \frac{1}{8} {\mathop{\mathbb{E}}}_{\vx \sim \gD}\left(\mP_{+} \vx \vx^\top \mP_+^\top - \mP_{+} \vx \vx^\top \mP_-^\top - \mP_{-} \vx \vx^\top \mP_+^\top + \mP_{-} \vx \vx^\top \mP_-^\top \right) \nonumber \\ & \qquad \qquad - \frac{1}{4} {\mathop{\mathbb{E}}}_{\vx \sim \gD} \mP_{-} \vx \vx^\top \mP_-^\top - 2 \lambda \mI \\ &\succcurlyeq \frac{1}{8} {\mathop{\mathbb{E}}}_{\vx \sim \gD} \vx \vx^\top - 3\lambda \mI,\end{aligned}$$ proving the lemma. \[lem:mbernstein\] Let $\mX_1, \dots, \mX_n \in \sR^{d \times d}$ be independent symmetric random matrices, such that ${\mathop{\mathbb{E}}}\mX_i = \vzero$ and $\norm{\mX_i} \le R$ almost surely. Then $$\begin{aligned} \Pr[\norm{\sum_{i = 1}^n \mX_i} \ge \delta] \le 2 n \exp(-\frac{\delta^2 / 2}{\sigma^2 + n R /3 }), \qquad \text{where} \quad \sigma^2 = \norm{\sum_{i = 1}^n {\mathop{\mathbb{E}}}\mX_i^2}. \end{aligned}$$ \[lem:ourmhoeffding\] Let $\mX_1, \dots, \mX_n \in \sR^{d \times d}$ be a sequence of i.i.d. positive semi-definite random matrices such that $\norm{\mX_i} \le R$ almost surely. Let $\overline{\mX} = \frac{1}{n}\sum_{i = 1}^n \mX_i$ be their average. Then we have $$\begin{aligned} \Pr[\norm{\overline{\mX} - {\mathop{\mathbb{E}}}\overline{\mX}} > \delta] \le 2 d \exp(- \frac{n\delta^2 }{8 R^2 + 4\delta R / 3})\end{aligned}$$ Define $\mY_i = \frac{\mX_i - {\mathop{\mathbb{E}}}\overline{\mX}}{n}$. Note that $\norm{\mX_i}, \norm{\overline{\mX}} \le R$ almost surely, so $\norm{\mY_i} \le 2 R / n$ almost surely. Furthermore, we have $$\begin{aligned} \sigma^{2} = \norm{\sum_{i = 1}^n {\mathop{\mathbb{E}}}\mY_i^2} \le n \cdot \frac{4 R^2}{n^2} = \frac{4 R^2}{n}. \end{aligned}$$ By , we have $$\begin{aligned} \Pr[\norm{\overline{\mX} - {\mathop{\mathbb{E}}}\overline{\mX}} \ge \delta] &= \Pr[\norm{\sum_{i = 1}^n \mY_i} \ge \delta] \\ &\le 2 d \exp(- \frac{\delta^2 / 2}{\sigma^2 + 2\delta R / (3n)}) \le 2 d \exp(- \frac{n\delta^2 }{8 R^2 + 4\delta R / 3}).\end{aligned}$$ \[lem:abbasi-ball\] Let $\{\gF_{i}\}_{i = 0}^\infty$ be a filtration. Let $\{\vx_{i}\}_{i = 1}^\infty$ be an $\sR^d$-valued stochastic process such that $\vx_i$ is $\gF_{i-1}$-measurable and $\norm{\vx_i} \le 1$ almost surely. Let $\{\epsilon_i\}_{i = 1}^\infty$ be a real-valued stochastic process such that $\varepsilon_i$ is $\gF_{i-1}$-measurable and is sub-Gaussian with variance proxy $1$ when conditioned on $\gF_{i-1}$. Fix $\vtheta \in \sR^d$ such that $\norm{\vtheta} \le 1$. Let $\mLambda_n = \lambda \mI + \sum_{i = 1}^n \vx_i \vx_i^\top, y_i = \vx_i^\top \vtheta + \epsilon_i$, and $\hat \vtheta_n = \mLambda_n^{-1} \sum_{i = 1}^n y_i \vx_i$. For every $\delta > 0$, we have that $$\begin{aligned} \Pr[\forall n \ge 0: \norm{\hat \vtheta_n - \vtheta}_{\mLambda_n} \le \sqrt{\lambda} + \sqrt{d \ln(\frac{1 + n / \lambda}{ \delta})} ] \ge 1 - \delta, \label{eq:abbasi-event}\end{aligned}$$ where we define $\norm{\vx}_{\mLambda} {\stackrel{\mathrm{def}}{=}}\sqrt{\vx^\top \mLambda \vx}$. Furthermore, when the event specified in holds, we have for every $n \ge 0$ and any vector $\vx \in \sR^d$ that $$\begin{aligned} \abs{\vx^\top (\hat \vtheta_n - \vtheta)} \le \left(\sqrt{\lambda} + \sqrt{d \ln(\frac{1 + n / \lambda}{ \delta})}\right) \sqrt{\vx^\top \mLambda_n^{-1} \vx}. \label{eq:abbasi-event-2}\end{aligned}$$ Tools for Matrix Operations --------------------------- \[lem:comp\] Given two positive semi-definite matrices $\mA$ and $\mB$. Suppose that $\mA \succcurlyeq \mB$. Then we have that $$\begin{aligned} \sup_{\vx \ne \vzero} \frac{\vx^\top \mA \vx}{\vx^\top \mB \vx} \le \frac{\det \mA}{\det \mB}.\end{aligned}$$ \[lem:unif\] Let $\mA \succcurlyeq \vzero$ be a positive semi-definite matrix. Suppose we are given a vector $\vx \in \sR^d$ such that $\vx^\top \mA^{-1} \vx \le z$, then we have $z \mA \succcurlyeq \vx \vx^\top$. Without loss of generality, assume that $\mA = \mathrm{diag}(\lambda_1, \dots, \lambda_d)$ is diagonal. Let $\vx = (x_1, \dots, x_d)^\top \in \sB^d$. For any vector $\vy = (y_1, \dots, y_d)^\top$, we have that $$\begin{aligned} \vy^\top (z \mA) \vy \ge (\vy^\top \mA \vy) (\vx^\top \mA^{-1} \vx^\top ) = \left( \sum_{i = 1}^d \lambda_i y_i^2 \right) \left( \sum_{i = 1}^d \lambda_i^{-1} x_i^2 \right) \ge \left(\sum_{i = 1}^d x_i y_i \right)^2 = \vy^\top (\vx \vx^\top) \vy,\end{aligned}$$ where the last inequality is by Cauchy-Schwarz. \[lem:matcover\] The covering number of the matrix set $$\begin{aligned} \rmM =\{\mA \in \sR^{d \times d} \mid R_1 \mI \preccurlyeq \mA \preccurlyeq R_2 \mI\}\end{aligned}$$ is bounded by $$\begin{aligned} \log \gN(\rmM, \varepsilon) \le O(d^2 \log(\max\{\abs{R_1}, \abs{R_2}\}d/\varepsilon)). \end{aligned}$$ Let $\widetilde{\mathfrak{M}} = \{\mA : \mA_{ij} \in [-\max\{R_1, R_2\}, \max\{\abs{R_1}, \abs{R_2}\}] \cap \{k \epsilon/(10d^2) : k \in \sZ\}\}$, and let $\mathfrak{M}$ be the projection of $\widetilde{\mathfrak{M}}$ onto $\rmM$. One can show that $\mathfrak{M}$ is an $\varepsilon$-cover of $\rmM$, and $\log \abs{\mathfrak{M}} \leq O(d^2 \log (\max\{\abs{R_1}, \abs{R_2}\}d/\epsilon))$. \[lem:matinvlip\] For any two positive semi-definite matrices $\mA, \mB \succcurlyeq \lambda \mI$, we have that $$\begin{aligned} \norm{\mA^{-1} - \mB^{-1}} \le \lambda^{-2} \norm{\mA - \mB} \qquad \mathrm{and} \qquad \norm{\mA^{-1/2} - \mB^{-1/2}} \le \lambda^{-3/2} \norm{\mA - \mB}.\end{aligned}$$ Note that $\norm{\mA^{-1}}, \norm{\mB^{-1}} \le \lambda^{-1}$. We have that $$\begin{aligned} \norm{\mA^{-1} - \mB^{-1}} &\le \norm{\mA^{-1}} \norm{\mI - \mA \mB^{-1}} \le \lambda^{-1} \norm{\mI - (\mA - \mB + \mB) \mB^{-1}} \\ &\le \lambda^{-1} \norm{\mB^{-1}} \norm{\mA - \mB} \le \lambda^{-2} \norm{\mA - \mB}. \end{aligned}$$ It remains to show that $\norm{\mA^{-1/2} - \mB^{-1/2}} \le \lambda^{-3/2} \norm{\mA - \mB}$. Since $\mA^{-1/2}, \mB^{-1/2} \succcurlyeq \lambda^{1/2} \mI$, we have that $$\begin{aligned} \norm{\mA^{-1/2} - \mB^{-1/2}} \le \lambda^{-1} \norm{\mA^{1/2} - \mB^{1/2}}. \label{eq:matinvlip-1}\end{aligned}$$ To complete, we assume that $\norm{\mA - \mB} \le \varepsilon (<\lambda^{-1})$. For any unit vector $\vx \in \sS^{d-1}$, we have $$\begin{aligned} \vx^\top (\mA^{1/2} - \mB^{1/2}) \vx &= \sqrt{\vx^\top \mA \vx} - \sqrt{\vx^\top \mB \vx} \\ &= \sqrt{\vx^\top \mB \vx + \vx^\top (\mA - \mB) \vx} - \sqrt{\vx^\top \mB \vx} \\ &\le \sqrt{\vx^\top \mB \vx + \varepsilon} - \sqrt{\vx^\top \mB \vx} \\ &\le \varepsilon / (2 \sqrt{\vx^\top \mB \vx}) \\ &\le \varepsilon / (2 \sqrt \lambda).\end{aligned}$$ By swapping $\mA$ and $\mB$, we can show $$\begin{aligned} \vx^\top (\mB^{1/2} - \mA^{1/2}) \vx \le \varepsilon / (2 \sqrt \lambda).\end{aligned}$$ Therefore, we have $$\begin{aligned} \abs{\vx^\top (\mA^{1/2} - \mB^{1/2}) \vx} \le \varepsilon / (2 \sqrt \lambda),\end{aligned}$$ which implies $\norm{\mA^{1/2} - \mB^{1/2}} \le \varepsilon / (2 \sqrt \lambda) \le \norm{\mA - \mB} / (2 \sqrt \lambda)$ by the definition of the matrix norm. We conclude with . The Generalized Elliptical Potential Lemma ------------------------------------------ Below we prove a generalized version of the elliptical potential lemma. Compared to the usual version in literature (e.g., [@abbasi2011improved]), our versions works for positive semi-definite matrices $\mX_1, \dots, \mX_n$ with traces upper bounded by $1$ instead of just rank-$1$ positive semi-definite matrices. However, we also need the extra assumption that $\Tr(\mX_i \mV_0^{-1}) \le 1$ for all $i \in [n]$. \[lem:ellip\] Suppose we are given a sequence positive semi-definite matrices $\mX_1, \dots, \mX_n$ such that $\Tr(\mX_i) \le 1$ for every $i \in [n]$. Let $\mLambda_0$ be a positive semi-definite matrix and let $\mLambda_i = \mLambda_{i - 1} + \mX_i$ for $i \in [n]$. When $\Tr(\mX_i \mV_0^{-1}) \le 1$ for $i \in [n]$, we have $$\begin{aligned} \sum_{i = 1}^n \Tr(\mX_i \mLambda_{i-1}^{-1}) \le 2 \ln \frac{\det \mLambda_n}{\det \mLambda_0}.\end{aligned}$$ Note that $$\begin{aligned} \mLambda_i = \mLambda_{i - 1} + \mX_i = \mLambda_{i - 1}^{1/2}(\mI + \mLambda_{i - 1}^{-1/2} \mX_i \mLambda_{i-1}^{-1/2}) \mLambda_{i - 1}^{1/2},\end{aligned}$$ so we have $$\begin{aligned} \det \mLambda_i &= \det(\mLambda_{i - 1}) \times \det(\mI + \mLambda_{i - 1}^{-1/2} \mX_i \mLambda_{i-1}^{-1/2}) \\ &\ge \det(\mLambda_{i - 1}) \times \left(1 + \Tr(\mLambda_{i - 1}^{-1/2} \mX_i \mLambda_{i-1}^{-1/2}) \right) \\ &= \det(\mLambda_{i - 1}) \times \left(1 + \Tr(\mLambda_{i - 1}^{-1} \mX_i) \right),\end{aligned}$$ where the inequality follows from that $$\begin{aligned} \det(\mI + \mA) = \prod_{j = 1}^d (1 + \lambda_j) \ge 1 + \sum_{j = 1}^d \lambda_j = 1 + \Tr(\mA), \end{aligned}$$ where $\lambda_j \ge 0$ is the $j$-th eigenvalue of $\mA = \mLambda_{i - 1}^{-1/2} \mX_i \mLambda_{i-1}^{-1/2}$. Together with the fact that $x \le 2 \ln(1 + x)$ for $x \in [0, 1]$, we have $$\sum_{i = 1}^n \Tr(\mX_i \mLambda_{i - 1}^{-1}) \le \sum_{i = 1}^n 2\ln(1 + \Tr(\mX_i \mLambda_{i - 1}^{-1})) \le 2 \sum_{i = 1}^n \ln \frac{\det \mLambda_i}{\det \mLambda_{i - 1}} = 2 \ln \frac{\det \mLambda_n}{\det \mLambda_0}. \qedhere$$ Pinsker’s Inequality -------------------- \[lem:pinsker\] If $P$ and $Q$ are two probability distributions on a measurable space $(X, \Sigma)$, then for any event $A \in \Sigma$, it holds that $$\left\| P(A) - Q(A) \right\| \leq \sqrt{\frac{1}{2} \mathrm{KL}(P \\| Q)},$$ where $$\mathrm{KL}(P \Vert Q) = \int_X \left(\ln \dv{P}{Q}\right) \dd{P}$$ is the Kullback–Leibler divergence. Omitted Algorithms, Lemmas and Proofs in ========================================= Full Description of [[BatchLinUCB-KW]{}]{} {#app:blinucbkw} ------------------------------------------ The algorithm is presented in . $M = \lceil \log \log T \rceil, \alpha \gets 10\sqrt{ \ln \frac{2dKT}{\delta}}$, $\gT = \{\gT_1, \gT_2, \dots, \gT_M\}, \gT_0 = 0, \gT_M = T, \forall i \in [M - 1]: \gT_i = T^{1 - 2^{-i}}$ Proof of (Analysis of Linear Regression) {#app:analysis-lr} ---------------------------------------- can be proved by a straightforward union bound over all stages and candidate arms, and the application of the following lemma. \[lem:analysis-lr\] Given $\vtheta, \vx_1, \vx_2, \dots, \vx_n \in \R^d$ such that $\norm{\vtheta} \leq 1$, for all $i \in [n]$, let $y_i = \vx_i^\top \vtheta + \epsilon_i$ where $\epsilon_i$ is an independent sub-Gaussian random variable with variance proxy $1$. Let $\mLambda = \lambda \mI + \sum_{i=1}^n \vx_i \vx_i^\top$, and $\hat\vtheta = \mLambda^{-1} \sum_{i=1}^n y_i \vx_i$. For any $\vx \in \R^d$ and any $\gamma > 0$, we have that $$\begin{aligned} \Pr[\abs{\vx^\top (\vtheta - \hat\vtheta)} > (\gamma + \sqrt{\lambda}) \sqrt{\vx^\top \Lambda^{-1} \vx}] \leq 2\exp(-\gamma^2/2).\end{aligned}$$ Note that $$\begin{aligned} & \abs{\vx^\top (\vtheta - \hat\vtheta)} =\abs{ \vx^\top \left(\mLambda^{-1} \sum_{i=1}^n \vx_i (\vx_i^\top \vtheta + \epsilon_i) - \vtheta\right)} \notag \\ & \qquad = \abs{\vx^\top \left(\mLambda^{-1} \sum_{i=1}^n \vx_i \epsilon_i + \mLambda^{-1}(\mLambda - \lambda \mI) \vtheta - \vtheta\right)} = \abs{\vx^\top \mLambda^{-1} \left(\sum_{i=1}^n \vx_i \epsilon_i - \lambda \vtheta\right)} \notag \\ & \qquad \leq \lambda \abs{\vx^\top \mLambda^{-1} \vtheta} + \abs{\sum_{i=1}^n \vx^\top \mLambda^{-1}\vx_i \epsilon_i}. \label{eq:lemma-lr-0}\end{aligned}$$ For the first term, since $\\|\vtheta\\| \leq 1$ and $\mLambda \succcurlyeq \lambda \mI$, we have that $$\begin{aligned} \lambda \abs{\vx^\top \mLambda^{-1} \vtheta} \leq \lambda \sqrt{\vx^\top \mLambda^{-2} \vx} \leq \sqrt{ \lambda \vx^\top \mLambda^{-1} \vx}. \label{eq:lemma-lr-part1}\end{aligned}$$ For the second term, since $\sum_{i=1}^n \vx^\top \mLambda^{-1}\vx_i \epsilon_i$ is independent sub-Gaussian with variance proxy $$\begin{aligned} \vx^\top \mLambda^{-1} \left( \sum_{i=1}^{n} \vx_i \vx_i^\top \right) \mLambda^{-1} \vx \leq \vx^\top \mLambda^{-1} \vx,\end{aligned}$$ by sub-Gaussian concentration inequalities, we have $$\begin{aligned} \Pr[\abs{\sum_{i=1}^n \vx^\top \mLambda^{-1}\vx_i \epsilon_i} > \gamma \sqrt{\vx^\top \mLambda^{-1} \vx}] \leq 2 \exp(-\gamma^2/2). \label{eq:lemma-lr-part2}\end{aligned}$$ Combining , , and , we prove the lemma. [^1]: Department of Industrial & Enterprise Systems Engineering, University of Illinois at Urbana-Champaign. Email: `[email protected]`. [^2]: Institute for Interdisciplinary Information Sciences, Tsinghua University. Work done while visiting the University of Illinois at Urbana-Champaign. Email: `[email protected]`. [^3]: Department of Industrial & Enterprise Systems Engineering, University of Illinois at Urbana-Champaign. Email: `[email protected]`. [^4]: Indeed, in the adaptive grid model, the time for a policy switch has to be decided when the previous policy switch happens, while in the rare policy switch model, the learner can freely switch the policy, as long as the total number of switches is limited. [^5]: A simple argument will prove the $\Omega(\sqrt{T})$ batch lower bound for achieving the asymptotically minimax-optimal regret for the adaptive grid model with adversarial contexts, and the rest bounds can be derived by direct corollaries of this work and the existing results in [@gao2019batched; @han2020sequential]. [^6]: For simplicity of presentation, we assume that the vectors in the sets of $\gD$ span the full dimension, so that there always exists a sample policy with invertible information matrix. Please refer to for the general definition. [^7]: This bound can be improved to $O(d)$ with additional techniques, which will be included in the full version of the paper. [^8]: More precisely, the good policy here is defined by the distributional G-deviation. Please refer to for more details. [^9]: When clear from the context, we interchangeably use the arm indices and their corresponding context vectors. [^10]: Note that although the theorem only works for the uniform distribution over a multi-set, since the properties to be proved in the theorem statement do not truly depend on $\Gamma$, the theorem can be generalized to any distribution via a simple discretization argument. [^11]: This is a generalized version and we invoke the lemma by letting $\mX_t$ in the lemma statement be $\vx_t \vx_t^\top$ and letting $\mLambda_t$ in the lemma statement be $\mU_t$. Note that $\mLambda_0 = \mU_0 \succcurlyeq \mI$ so that $\Tr(\mX_t \mLambda_0^{-1}) \leq 1$ is satisfied. [^12]: Although in our later algorithm, we only learn a policy with small $\lambda$-deviation* as defined in , however, one can also verify that the $\lambda$-deviation of $\pi$ over $\gD$ in this counterexample is also high.	{ "pile_set_name": "ArXiv" }
--- author: - Emanuele Massaro - Daniel Kondor - Carlo Ratti bibliography: - 'scibib\_DK.bib' title: Assessing the interplay between human mobility and mosquito borne diseases in urban environments --- Introduction {#Sec:intro .unnumbered} ============ Rapid urbanization and increased mobility brings new challenges for epidemics [@neiderud2015urbanization]. Estimates show that more than half of the world’s population already lives in cities, while further big increases are expected especially in Asia and Africa. Challenges presented by new megacities include the rapid spread of new epidemics, which can become worldwide threats due to increased global connectivity [@Brockmann2013; @gvr2017; @massaro2018resilience]. Poor housing conditions in rapidly growing cities can exacerbate epidemic threats, especially in the case of insect and rodent vector diseases and geohelminths [@world2010hidden; @ajelli2017modeling]. Governments need to look for innovative solutions for monitoring and controlling epidemics [@lindsay2017improving]. An important part of these considerations is understanding the relationship between disease spread and human mobility, which have been previously linked on global scales [@Brockmann2013; @gvr2017]. In this paper we explore the effectiveness of pervasive technologies, specifically mobile phone data, in predicting and understating the emergence of mosquito borne disease outbreaks in urban environments. Cell phone data has been shown to be valuable in monitoring mobility patterns in near real-time [@kung2014exploring]. Such information has a large potential in epidemiological modeling and control [@wesolowski2014quantifying], yet it has been often unreliable and difficult to obtain with traditional methods, especially in developing countries with rapidly changing urban environments and limited resources to conduct travel surveys. We study the influence of human mobility on the spread of the mosquito-borne dengue virus, as inferred from a large-scale mobile communication dataset in the city-state of Singapore. Contrary to previous studies that either focused on this problem at the scale of countries or regions [@cummings2004travelling; @lourencco20142012; @wesolowski2012quantifying; @teurlai2012can; @Brockmann2013; @wesolowski2014quantifying; @WesolowskiPNAS2015], essentially treating cities as well-mixed nodes in a larger travel network, or used small-scale data of human movement inside cities collected through surveys [@stoddard2013house] or only use theoretical models and aggregate on intra-city human mobility [@stoddard2009role; @kong2018modeling; @karl2014spatial; @zellweger2017socioeconomic; @telle2016spread], we now employ a large-scale dataset of human mobility to study the connection between intra-city mobility and dengue spread. We focus on comparing a dengue transmission model based on people’s real commuting patterns (as inferred from the mobile phone dataset) with the observed dengue cases and with simulations employing random mobility models. This allows us to measure the impact of mobility model on the accuracy of modeling the spatial distribution of dengue cases. We especially focus on comparing random mobility that results in perfect mixing of population with more structured mobility models, effectively evaluating the importance of intra-city human mobility in dengue spread. Dengue fever is a mosquito-borne viral infection, transmitted by female mosquitoes of the species Aedes aegypti and Aedes albopictus when biting humans. The infection causes flu-like symptoms with occasional complications that can be fatal. There are four strains of the virus and the infection with one strain produces lifelong immunity to that type. However, a second infection with a different type increases the risk of severe complications. Dengue continues to be a global threat, with about half the world’s population being at risk of infection [@guzman2010dengue]. Worldwide, there are more than 50 million infections every year, leading to half a million hospitalizations and up to 25 thousands deaths. Dengue is prevalent in tropical and sub-tropical climates worldwide, mostly in urban and semi-urban areas. The prevention and control solely depends on controlling the mosquito populations. There is active development for vaccines, with a first-generation vaccine becoming available recently [@rothman2016dengue]. Dengue affects Singapore in particular and two major outbreaks were observed in 2013 and 2014 ( \[fig:tempCases\](a)). The modeling of dengue outbreaks has attracted the interest of many researchers in many disciplines from physics to computational biology. Presented models investigate, for example, the variability of the mosquito population [@esteva1998analysis], the variability of the human population [@esteva1999model], the vertical transmission between mosquitos (that is, the transmission between mosquito generations) [@esteva2000influence] as well as seasonal patterns [@hartley2002seasonal]. Otero at al. [@otero2006stochastic; @otero2008stochastic] presented a dengue model, which takes into account the evolution of the mosquito population. Another study shows that dengue appears to travel in waves [@cummings2004travelling]. As the flight range of mosquitoes is limited to a few hundred meters [@stoddard2013house], it is generally assumed that humans carry the dengue virus to previously dengue-free areas and infect local mosquitoes. There is evidence that the spread of mosquito-borne diseases is related to human mobility [@stoddard2009role]. Various agent-based simulations suggest that the mobility of humans could be the main driving force behind the spread of the dengue virus [@barmak2011dengue; @de2011modeling]. Teurlai et al. [@teurlai2012can] showed that the human mobility, estimated from the road network, influences the spread at a national scale in Cambodia. Especially house-to-house human movements seem to play a key role in Iquitos, Peru [@stoddard2009role]. Related malaria studies show that human mobility, which is estimated from cell-phone networks, drives the dissemination of malaria parasites as well [@wesolowski2012quantifying]. Recently Wesolowski et al. studied the impact of human mobility on the emergence of dengue epidemics in Pakistan [@WesolowskiPNAS2015] using mobile phone-based mobility. Considering the threat presented by dengue especially in cities, many authors studied the effect of dengue fever in urban environments [@kong2018modeling; @karl2014spatial; @lourencco20142012; @zellweger2017socioeconomic; @telle2016spread; @lindsay2017improving]. While these work generally assume that intra-city mobility is an important factor for dengue epidemics, a direct quantification of this effect is still lacking. For the first time, we analyze the effectiveness of mobile phone useage data to predict the dengue spreading in an urban environment, such as Singapore. In doing so, we compare random mobility patterns with the real one estimated from anonymized mobile phone usage records in an agent-based model of dengue transmission adapted from previous studies [@lourencco20142012; @WesolowskiPNAS2015]. This way, we are able to characterize the effect of human mobility on urban scales in the spread of vector-borne diseases and the effectiveness to use mobile phone data to estimate disease epidemics on this scale as well. Results {#Sec:res .unnumbered} ======= ![Temperature Dependency of the Dengue cases and Schematic representation of the Human-Vectors interactions in the epidemiological model. (a) Weekly observed dengue cases and average temperature in Singapore from January 2013 to December 2014. Two outbreaks took place during those two years during the summer. It is possible to observe a correlation between temperature and number of reported cases of people affected by the disease. (b) Compartmental classification for DENGUE disease. Humans can occupy one the four top compartments: susceptible, which can acquire the infection through contacts (bites) with infectious mosquitoes; exposed, where individuals are infected but are not able yet to transmit the virus; infectious, where individuals are infected and can transmit the disease to susceptible mosquitoes; and recovered or removed, where individuals are no longer infectious. The density of mosquitoes changes according to the seasonal transition from Aquatic (A) to Adult Mosquitoes (V). Similar to the humans case, Mosquitoes can occupy three different compartments and they can die with a given rate depending on the temperature. []{data-label="fig:tempCases"}](Figure1.pdf){width="1\linewidth"} ![Commuting flows from home to work locations aggregated at the 55 planning areas. The location of the nodes corresponds to the centroid of the areas and their size corresponds to the incoming degree which corresponds the total amount of agents that commutes everyday to that area. In this figure we report only the most significant nodes in terms of incoming flow (i.e. greater than 95th percentile the distribution). (a) We can observe that major hub in the mobile phone data mobility model corresponds to the Central Business District where the majority of the jobs are located. (b) The random mobility mobility has different hubs randomly distributed in the space. (c) The Levy-distribution and (d) the radiation model show similar patterns, with an homogeneous distribution on the territory without significant hubs: however the mobility derived from the radiation model is more aggregated in the central part of the city.[]{data-label="fig:flowsMap"}](Figure2.pdf){width="1\linewidth"} We propose an agent-based dengue transmission model in which humans and mosquitoes are represented as agents and humans go through the epidemic states of dengue [@esteva1998analysis; @esteva1999model; @esteva2000influence; @barmak2011dengue]. To model dengue dynamics, we use a stochastic population model based on the ordinary differential equation (ODE) framework employed by Lourenco and Recker to describe a dengue outbreak in Madeira, Portugal [@lourencco20142012] and then used by Wesolowski and colleagues to model the dengue outbreak in Pakistan [@WesolowskiPNAS2015]. The epidemiological model depends on both temperature-dependent and constant parameters as described in the Methods section and reported in Table \[table:tempPar\] and Table \[table:constPar\]. We employ an agent-based approach for humans, while we model localized mosquito subpopulations stochastically. As a necessary simplification, we only consider one serotype of dengue; in this case, individuals can only be infected once. The physical environment in which the epidemic takes place is a regular grid, composed of $320\,\mathrm{m}\times 320\,\mathrm{m}$ cells, overlaid the city of Singapore. The model is composed of two phases: (i) the phase of reaction, defined by the epidemioloigcal model (see the section Materials and Methods for details and \[fig:tempCases\](b) for a schematic overview), where disease transmission takes place in each grid cell; (ii) the phase of diffusion where agents are moved from one grid cell to another according to the mobility model under consideration: the mobility flows aggregated at census district level for the different mobility models are reported in \[fig:flowsMap\]. Each day consists of two reaction phases, corresponding to day and night, and two diffusion phases, corresponding to people’s morning and evening commute. In this work, we consider four different mobility models (see \[fig:flowsMap\]) and compare their predictive power about the dengue outbreaks of 2013 and 2014 in Singapore. In each mobility model, we assign a home and work location (grid cell) to each agent who are assumed to commute between these two daily. The models differ in the way how this assignment is made: (1) mobile phone data: we use anonymized call detail records of one mobile phone operator in Singapore, collected in a two month period in 2011 that allows us to estimate home and work cells for $2.3$ million agents; (2) random work location: in this case, we still use the home cells estimated from the mobile phone data, but work locations are assigned randomly; (3) Levy-distribution: each agent is assigned a random home location based on the mobile phone data and a work cell is chosen such that the commuting distance follows a truncated Levy-distribution; (4) radiation model: we use census data to distribute the home locations of agents [@singapore2010census] and then we choose work cell locations according to the radiation model of Simini et al. [@simini2012universal]. In total, there are 2,598 grid cells with either a home or work location in them. More detailed description of the mobility models is given in the Materials and Methods section, while we present a comparion between the mobility models in the Supplementary Material, in Figure S1 to Figure S5.. Most notably, flows of people on the district level are highly correlated among the mobile phone data and the radiation model ($r = 0.938$), somewhat less correlated with the Levy-distribution model ($r = 0.901$) and significantly less correlated among mobile phone data and random mobility ($r = 0.304$). This way, we conclude that the radiation, Levy-distribution and random mobility models give successively worse approximations of real mobility patters. Beside the mobility model, we have two main variable parameters, the number of mosquitoes per human, $x_v$ and average bite rate of mosquitoes, $a$ (more thorough definitions of these and a discussion on model parameters are given in the Materials and Methods section). We perform a sensitivity analysis on these, by exploring the phase space $x_v \in [0.004, 0.1]$ and $a \in [0.14, 0.26]$. This allows us to calibrate our model to the population of agents in Singapore; this is a necessary step since exact estimation of these parameters is especially challenging in real-world condition, while several parameters in the epidemiological model cannot be reliably measured in real-world conditions, only in controlled laboratory experiments [@liu2014vectorial]. In our approach, we use best available estimates from the literature for most parameters, while allow variation of $x_v$ and $a$ to deal with this inherent uncertainty. We select the best parameter combination for each mobility model to evaluate our results. We start our simulations with initial conditions for infected human agents based on the observed number and distribution of cases in January 2013, while we obtain the initial mosquito populations by running the population dynamic model for an initial warm-up period as described in the Materials and Methods section. To account for the stochastic nature of the simulation, for each parameter value, we ran the simulation 100 times and report the median and average values in the following. Temporal Analysis {#temporal-analysis .unnumbered} ----------------- ![Temporal analysis. We report the comparison between the best simulated scenario and the observed number of dengue cases during the 2013-2014 outbreaks. Parameter values for $x_v$ (average number of mosquitoes per human) and $a$ (mosquito bite rate) are displayed in the figure legends for each case.[]{data-label="fig:simCases2013"}](temporalAnalysis.pdf){width="1\linewidth"} [l l l l]{} Mobility Model & $R_2$ & $x_v$ & $a$\ \ Mobile Phone & 0.65 & 0.006 & 0.16\ Random & 0.51 & 0.006 & 0.2\ Levy Distribution & 0.62 & 0.009 & 0.26\ Radial Model & 0.56 & 0.005 & 0.24\ In \[fig:simCases2013\] we report the comparison of the number of observed cases and the median of the simulated infected cases estimated from our simulations during the epidemiological weeks in 2013 and 2014 for the four different mobility models. In particular for each mobility model we report the pair of parameters $x_v$-$a$ that maximize the $R^2$ between the simulations and observed number of cases. Each mobility model is able to predict quite well the temporal evolution of the dengue outbreaks, since the epidemiological dynamics mainly depends on the value of the temperature. Each mobility model optimizes the prediction for different values of the parameter $x_v$-$a$ as reported in the legends of \[fig:simCases2013\]. In order to find the best pair of parameter values, we compute the $R^2$ between the observed and the predicted number of cases between the 12th and 26th epidemiological weeks when the epidemic peaked during the study period. We show optimal $R^2$ values in Table \[tab:rmodels\] and display variation of $\log R^2$ in the phase space in Figure S13 in the Supplementary Information. The Mobile Phone Data and the Levy Distribution mobility models have the better accuracy with a value of $R^2$ of $0.65$ and $0.61$ respectively while the Random and the Radiation mobility models tend to overestimate the number of cases and with $R^2$ of $0.52$ and $0.57$ respectively. Nevertheless, we still conclude that all models reproduce the main trends in the epidemic well. Spatial Analysis {#spatial-analysis .unnumbered} ---------------- In this section we show the results of our simulations and we compare it with the spatial distribution of number of reported cases in 2013-2014 in Singapore. The distribution of Ae. aegypti expanded during the decade from 2003 to 2013 and the percentage of houses with mosquito breeding in 2013 and 2014 was significantly higher than in previous years [@hapuarachchi2016epidemic]. As expected, the dengue case distribution pattern in 2013 and 2014 was in line with the geographical spread of A. aegypti in the country [@hapuarachchi2016epidemic]. The biggest clusters remain in Tampines in the eastern part of the island, however more are now in the west and north. In order to quantify the effect of human mobility on the spatial propagation the dengue in Singapore, we compare the results of our model with observed case scenarios by considering the four different mobility patterns: (1) mobile phone data; (2) random; (3) Levy; (4) radiation. We show the cumulative spatial distribution of observed cases in \[fig:spatCaseReal\] and simulated cases in the four models in \[fig:spatCase\] with with the $x_v$ and $a$ parameters that give the best estimate for the temporal patterns (as reported in \[fig:simCases2013\]). We can see that the mobility plays an important role for predicting the spatial distribution of the number of cases. Indeed the spatial distribution of the number of cases predicted by the random mobility model is uniformly distributed among the city, while the other mobility models allow us to detect key hotspots of the outbreaks similar to the observed scenario. ![Observed dengue cases. Cumulative spatial distribution of observed dengue cases during the 2013 and 2014 outbreaks.[]{data-label="fig:spatCaseReal"}](realMap.pdf){width=".5\textwidth"} ![Spatial analysis. We report the heatmap of the cumulative number of cases for the four mobility models. For each simulated scenario we report the results with the best parameter values, as shown in the figures.[]{data-label="fig:spatCase"}](Figure5a.png){width="100.00000%"} To better distinguish between the predictive power of different mobility models, we computed structural similariy (SSIM) scores [@Wang2004; @wang2009mean] (see the Supplementary Information for a description) for each case in each epidemiological weeks, with the best parameters $x_v$-$a$ and compare their distribution in \[fig:ssim\]. We can observe that the mobile phone mobility model and the radiation model perform in a similar way, consistently well approximating the observed spatial distribution during the time period of our study. The Levy-distribution model is performing slightly worse, while the random mobility model gives significantly worse results. Looking at the results in \[fig:spatCase\], we find that the overall distribution of the infected cases for the random mobility model corresponds well to the average population density (i.e. average of work and home locations in each cell). This means that our random mobility model achieves a good mixing among the population. The difference from the real distribution of dengue cases and the other mobility models highlights that uniform mixing among the population does not account for a spread of dengue in Singapore, thus mobility patterns are an important factor. While previous large-scale epidemiological studies often treat cities as well-mixed nodes in a larger travel network [@cummings2004travelling; @lourencco20142012; @wesolowski2012quantifying; @teurlai2012can; @Brockmann2013; @wesolowski2014quantifying; @WesolowskiPNAS2015], our results show that disease spreading can exhibit important localized patterns inside cities as well, in line with studies done previously on smaller samples of the population or aggregate models of human mobility [@stoddard2009role; @stoddard2013house; @kong2018modeling; @karl2014spatial; @zellweger2017socioeconomic; @telle2016spread]. It is unclear yet, how the intra-city and inter-city epidemiological models are best reconcilied; we note that frequent travelers are often a non-uniform sample of the total population of any city, thus local and long-range spread of infectious diseases can have complex intervowen patterns. The further difference between the Levy and radiation mobility model is consistent with previous work which found the radiation model to best reproduce the statistical properties of human commuting [@simini2012universal]. Furthermore, the good results obtained from the mobile phone data show that the home-work commuting estimated from this dataset indeed accounts for the most important factors in human mobility in Singapore. ![Spatial analysis. Boxplot of the value of the SSIM Index for each weeks during the 2013-2014 outbreaks using the best parameter $x_v$-$a$ shown in \[fig:simCases2013\]. SSIM index values were calculated for each epidemiological week during the outbreak for each of the 100 simulation runs. The distributions of these values are shown as boxplots for each mobility model in this figure. The boxplots show the minimum, first quartile, median, third quartile and maximum among the SSIM values observed. We see that in all cases, the range of data is quite small; the mobile phone data and radiation model results are clearly distinguished from the random mobility and Levy-distribution results.[]{data-label="fig:ssim"}](ssim1.pdf){width="0.75\linewidth"} Discussion {#Sec: disc .unnumbered} ========== More than $80\%$ of the world’s population is at risk from at least one vector-borne disease [@gvr2017]. The populations most at risk are those living in poverty in the tropical and subtropical areas, but as the case of Singapore shows, highly developed cities and countries still need continued efforts to prevent outbreaks [@hapuarachchi2016epidemic]. The rapid urbanization, the increase in international travel and trade, the modification of agriculture and environmental changes have increased the spread of vector populations, putting more and more people at risk. Mobile phone data can give real-time mobility information that can be combined with infectious disease surveillance data and seasonally varying environmental data to map these changing patterns of vulnerability in cities that are changing everyday. In this paper we proposed an agent based model in order to explicitly simulate the epidemic spread of the disease as governed by the transmission dynamics of the dengue virus through human-mosquito interactions and promoted by the population movements across the city state of Singapore. In this methodology humans and mosquitoes are represented as agents and humans go through the epidemic states of dengue. We modelled four different mobility patterns: 1) mobility estimated from mobile phone data, 2) random mobility patterns, 3) mobility estimated from census data following a Levy distribution model and 4) mobility estimated from census data following a radiation model. We were able to reproduce the main temporal and spatial patterns of the dengue outbreak in 2013 and 2014. Our results show that human mobility is a very important factor in the spread of vector-borne diseases such as dengue even on the short scale corresponding to intra-city distances. This is evidenced by the large difference found between the observed spatial pattern of dengue cases and the ones obtained by the completely random mobility model which corresponds to a “perfect mixing” among the population. This extends the results obtained from the previous work of Wesolowski et al. [@WesolowskiPNAS2015] who showed how human mobility determines the spread of dengue on the scale of a country and studies that investigated the relationship between human mobility and spread of vector diseases on different spatial scales [@kong2018modeling; @wesolowski2014quantifying; @wesolowski2012quantifying; @lindsay2017improving; @telle2016spread; @zellweger2017socioeconomic; @maneerat2017agent]. We believe that our main contribution is showing that human mobility patterns are important for the spread of vector-borne diseases even on intra-city scales; this is in contrast to previous studies which often assume cities to be a well-mixed environment for the purpose of epidemiology and study disease spreading between cities and regions [@Brockmann2013; @wesolowski2014quantifying; @lourencco20142012]. It is an interesting question for future work to what extent this result applies to other types of diseases, e.g. airborne infections that require only shorter co-location of people to spread, thus are able to exploit mixing of population in a more rapid way. Furthermore, we found that more sophisticated models of intra-city mobility can give good estimates of the spatial spread of dengue, opening up the possiblity to incorporate these into modeling and control of vector diseases in urban environments. The proposed methods could be integrated into urban planning in near real time. Mobile phone data is an obvious candidate for this purpose, giving real-time information on people’s movements. A major limitation of mobile phone data generated by national operators is the difficulty in capturing cross-border travel patterns and it is not possible to monitor with high accuracy the flux of people travelling to the city. On the other hand, we found that the radiation model of people’s commuting behavior performs similarly well, opening up the possibility to improve prediction of disease spread if accurate census data is available. Thus, we believe our methods can be readily used in other cities where these mobility models can be estimated, while accuracy will be affected by overall predictability of human movements and regularity of commuting patterns [@gonzalez2008understanding; @returners; @simini2012universal; @Song1018; @kung2014exploring]. Concluding, we note that the methods we presented here can be readily generalized to consider different mosquito-borne diseases such as dengue, chikungunya, malaria, yellow fever and different sources and models of human mobility, having a large potential usability for better understanding, control and prevention of vector disease epidemics in urbanized areas. Materials and Methods {#materials-and-methods .unnumbered} ===================== The code used for our simulations is available online [@CODE]. \[Sec: method\] Mobile phone data {#mobile-phone-data .unnumbered} ----------------- Anonymized call detail records (CDRs) were collected over a two month period in 2011 by one of the mobile phone operators in Singapore with a significant market share (a statistical analysis is reported in Figure S9 in the supplementary materials.). The data includes more than $2$ billion records in total and includes the approximate time and location of events (including phone calls and text messages). Locations are collected at the cell tower level with further noise applied for privacy reasons. We use this data to assign two “favorite” locations to each user: (i) home and (ii) work. In Singapore, according to a study by the Land Transport Authority, about $80\%$ of all trips go to either a work or a home location[@holleczek2014detecting]. This implies that the infection with the dengue virus in Singapore very likely happens either at home or at work, thus we focus on commuting between these two locations when modeling human mobility in this paper. To estimate home and work locations, we perform a spatial clustering of the CDRs, creating overlapping clusters of events which are spatially close to each other (a threshold of $500\,\mathrm{m}$ was used so as to account for the potential uncertainty regarding which one of nearby antennas a phone connects to). To be able to distinguish between home and work locations, we performed this clustering procedure separately for records generated between 8pm and 6am on weekdays and during weekend (for home locations) and records generated between 10am and 4pm on weekdays (for work locations). After this procedure, we selected the largest clusters for both cases and filtered the list of users who had at least 10 events in both clusters. Following this procedure, there are $2,307,230$ users to whom we can assign a home and a work location. We then assign users’ home and work locations into a $320m \times 320m$ grid overlay $G$ which we use as the basis of the epidemic simulation. We display the distribution of these home and work locations in Figure S2, while we show the nonempty grid cells in Figure S12 in the Supplementary Material. To show that the cellphone dataset is representative of Singapore it is possible to compare the distribution of the home locations identified by our clustering procedure with official census data from 2010 [@singapore2010census] (See Figure S10 and Figure S11 in SI). Singapore is divided into 55 urban planning areas[@wikipedia] and we compare the number of home locations identified in each of them with the 2010 census data[@singapore2010census]. With a correlation coefficient of 0.96, the two spatial distributions are highly linearly correlated as shown in Figure S10 in the supplementary materials. Furthermore, we note that penetration of mobile phones (number of active mobile phone subscriptions compared to the total population) in Singapore was above 140% at the time of our study [@mobilepenetration], one of the highest rates in the world. This way, we expect that almost all of the population has a mobile phone and many people have more than subscription. As the flight range of mosquitoes is limited to a few hundred meters [@stoddard2009role], it is generally assumed that humans carry the dengue virus to previously dengue-free areas and infect local mosquitoes. For this reason in our model mosquitoes don’t travel among different cells. Therefore, in the computational implementation each day is represented by two steps: daytime, during which population stay at work, and nightime during which poluation stay at home. Mobility Models {#mobility-models .unnumbered} --------------- We use four different models to estimate mobility of people and assign home and work locations to our agents. The first one is the mobility model defined according to the real estimation of mobility patterns from CDR data as described above. The second mobility model is a model in which for each agent we take the home location from the mobile phone data while the work location is assigned randomly (according to a uniform distribution) among the $2598$ cells. The third mobility model is defined in the following way: for each agent we choose a random home cell of the grid, while the work location is choose with a distance ($d$) that follows a truncated Levy distribution [@gonzalez2008understanding] as distribution of the mobility patterns, such as $P(d) \sim (d+d_0)^{-\beta}\exp(-d/k)$, where $P(d)$ is the probability to have of distance $d$ between home and work location, $d_0 (m) = 100$, $\beta=2$ and $k(m)=1500$. The fourth mobility pattern has been generated according to the radiation law of human mobility [@simini2012universal]. According to this we generated a mobility pattern considering the following: i) we assigned to each cell a number of inhabitants randomly distributed (normal distribution) according to the census data. ii) for each cell we consider that the percentage of commuters is the $80\%$ while the other $20\%$ work and live in the same cell. iii) For all the other inhabitants we computed that distance between home and work location based the radiation laws that reads $\langle T_{ij}\rangle = T_i \frac{m_i n_j} {(m_i+s_{ij})(m_i + n_{j} + s_{ij})}$, where $T_i$ is the total number of commuters from county $i$, $m_i$ and $n_j$ are the population in county $i$ and $j$ respectively, and $s_{ij}$ is the total population in the circle centered at $i$ and touching $j$ excluding the source and the destination population. The displacement of the agents for the different mobility models are reported in Figure S2 to Figure S5 in the supplementary materials. The generated mobility models show that the radiation model the model generated with the mobile phone are the most similar while there is almost no correlation with the random mobility model as shown in Figure S1 in Supplementary materials. Epidemiological Data {#epidemiological-data .unnumbered} -------------------- Information about the weekly number of reported Dengue cases in Singapore was collected from the official Singapore’s government’s one-stop portal[@weekly]. In the 2013 dengue outbreak in Singapore, a significant rise in the number of dengue fever cases was reported in Singapore and caused 8 victims and a total of 22318 cases. In the week of 16–22 June 2013, there was a record of 842 dengue cases in Singapore in a single week. This figure was far beyond the highest number of cases per week in the years 2010, 2011 and 2012. The number of weekly dengue fever cases has exceeded the epidemic threshold of 237. Similarly high number of cases were reported over the course of 2014, with the maximum number of weekly reported cases having a peak of 891. In the following years, the number of dengue cases were significantly lower due to increased efforts to control the mosquito population. We show the total number of dengue cases during 2013 and 2014 in \[fig:tempCases\](a). For the spatial analysis of Dengue outbreaks, we use a dataset that is a collection of data from the NEA. Data was collected twice a week since May 2013 (except for a gap in October 2013) from SGCharts Charting Singapore’s Data[@SGC]. The data provide information of the number of dengue cases in local spatial clusters that were established dynamically based on the location of recent cases. Spatial clusters are typically a few hundred meters in size, encompassing multiple city blocks. This allows us to have a good representation of the spatial spread of dengue, while still protecting to privacy of people affected. We display the cumulative spatial distribution of dengue cases in \[fig:spatCase\](a). Climate Data {#climate-data .unnumbered} ------------ We collected data about climate conditions in Singapore during years 2013 and 2014, during which two outbreaks during the respective summers took place. In \[fig:tempCases\](a) we show the number of dengue cases during the epidemiological weeks in 2013 and 2014 comparing it with the average temperature. The impact of daily temperature fluctuations on dengue virus transmission by the A.aegypti mosquitoes have been extensively studied and the results indicate that the weekly mean temperature is statistically significant relative to the increases in dengue incidence in Singapore and signifies the hazardous impacts of climatic factors on the increase in intensity and magnitude of dengue cases [@hii2009climate]. This reflection can be observed in the outbreaks of 2013 and 2014 where the comparison between reported cases and temperature has been reported in \[fig:tempCases\](a). Weather data including Mean temperature (MeanT, $^{\circ}C$), Minimum temperature (MinT, $^{\circ}C$), Maximum temperature (MaxT, $^{\circ}C$), Rainfall (Rain, mm), Relative humidity (RH, $\%$) and Wind speed (WindS, m/s) were obtained from the National Environment Agency, Singapore (NEA)[@rainy]. Epidemiological Model {#epidemiological-model .unnumbered} --------------------- The epidemiological model can be described schematically as shown in \[fig:tempCases\](b). Motivated by research that shows that mosquitoes have a very limited flight range and infection is carried by human movement [@stoddard2009role; @barmak2011dengue; @de2011modeling; @stoddard2013house], we assume mosquitoes to have a fixed location, i.e. there is no interaction between mosquito populations in distinct grid cells. For this reason, humans are treated as distinct agents, while the values for mosquitoes are aggregated at the cell level. The transitions on the proposed epidemiological model depend on temperature dependent parameters as reported in Table \[table:tempPar\] and described in the Supplementary materials (see also Figure S6 and Figure S7 in the supplementary materials). The constant parameters are described in Table \[table:constPar\]. ------------------------------------------------------------- ----------------------------------------------------------------- ------------------------- Notation Description Reference \[0.5ex\] $\dot{\epsilon_A^v} = \epsilon_A^v (T)$ transition rate from aquatic to adult mosquito life-stages [@yang2009assessing] $\dot{\mu_A^v} = \mu_A^v (T)$ mortality rate of aquatic mosquito life-stages [@yang2009assessing] $\dot{\mu_V^v} = \mu_V^v (T)$ mortality rate of adult mosquito life-stage [@yang2009assessing] $\dot{\theta_V^v} = \theta_V^v (T)$ intrinsic oviposition rate of adult mosquito life-stage [@yang2009assessing] $\dot{\gamma_V^v} = \gamma_V^v (T)$ extrinsic incubation period of adult mosquito life-stage [@yang2009assessing] $\dot{\phi}^{h \rightarrow v} = \phi^{h \rightarrow v} (T)$ human-to-vector probability of transmission per infectious bite [@lambrechts2011impact] $\dot{\phi}^{v \rightarrow h} = \phi^{v \rightarrow h} (T)$ vector-to-human probability of transmission per infectious bite [@lambrechts2011impact] \[1ex\] ------------------------------------------------------------- ----------------------------------------------------------------- ------------------------- : Temperature-dependent parameters. \[table:tempPar\] Notation Description Value Reference ------------ ----------------------------------------------------------------------------------- ---------------------------- ------------------------------------------- $\gamma^h$ transition rate from exposed (E) to infected (I) for humans $0.5\,\mathrm{days}^{-1}$ [@WesolowskiPNAS2015; @lourencco20142012] $\sigma^h$ recovery rate, i.e. transition rate from infected (I) to recovered (R) for humans $0.25\,\mathrm{days}^{-1}$ [@WesolowskiPNAS2015; @lourencco20142012] $c$ mosquite eggs hatching to larvae $1$ [@yang2009assessing] $f$ female mosquitoes hatched from all eggs $1$ [@yang2009assessing] : Constant parameters. \[table:constPar\] ### Humans {#humans .unnumbered} In the stochastic framework, we represent each human as an agent $i$, who at each timestep $t$ can be described by a pair $(N,c)_{t,i}$, where $N = S,E,I,R$ is the epidemic state (susceptible, exposed, infected and recovered, respectively), and $c$ denotes the grid cell where the agent resides. In our mobility models, $c$ alternates between a set home and work location, either inferred from the mobile phone usage data in the realistic scenario or generated randomly. We denote by $S_{t,c}$, $E_{t,c}$, $I_{t,c}$ and $R_{t,c}$ the total number of susceptible, exposed, infected and recovered humans in cell $c$ for timestep $t$. We further use $N_{t,c} \equiv S_{t,c} + E_{t,c} + I_{t,c} + R_{t,c}$ for the total number of humans. We assume each agent to be susceptible to the virus initially ($S$). Upon challenge with infectious mosquito bites ($\lambda^{v \rightarrow h}$), individuals enter the incubation phase ($E$) with mean duration of days $1/\gamma^h$, later becoming infectious ($I$) for days $1/\sigma^h$ and finally recovering ($R$) with life-long immunity. In each timestep, transition of the states can occur with probabilities $\lambda^{v\rightarrow h}_{t,c} / 2$, $\gamma^h / 2$ and $\sigma^h / 2$ for the case of the $S\rightarrow E$, $E\rightarrow I$ and $I\rightarrow R$ transitions respectively (as each reaction timestep takes half day, we obtain the transition probabilities by halving the daily transition rates). We evaluate the transitions individually for each human agent as a Bernoulli-process, and update the state accordingly. While the $\gamma^h=0.5 \quad days^{-1}$ and $\sigma^h= 0.25 \quad days^{-1}$ rates are constants[@lourencco20142012; @WesolowskiPNAS2015], the $\lambda^{v\rightarrow h}_{t,c}$ rate is related to the mosquito population of the grid cell where the human agent is currently residing: $$\lambda^{v\rightarrow h}_{t,c} = a \dot{\phi}^{v \rightarrow h} \frac{I^v_{t,c}}{N_{t,c}} = a \dot{\phi}^{v \rightarrow h} \frac{V_{t,c}}{N_{t,c}} \rho^I_{t,c} \propto V \rho^I$$ where $a$ is the biting rate (i.e. how many humans a mosquito bites on average per day), $\dot{\phi}^{v \rightarrow h}$ is the disease transmission rate per bite, while $I^v_{t,c}$ is the total number of infected mosquitos in cell $c$ at time $t$ (i.e. $a \frac{I^v_{t,c}}{N_{t,c}}$ gives the probability that an infected mosquito bites the given human agent during this timestep), while $V_{t,c}$ is the total number of mosquitos in cell $c$ and $\rho^I_{t,c} = I_{t,c} / N_{t,c}$ represents the fraction of infected individuals in that cell. The change in compartments of human agents is then expressed by the following equations: $$\begin{aligned} t^{S\rightarrow E}_{t,c} = BD(S_{t,c},\lambda^{v\rightarrow h}_{t,c} / 2) \\ t^{E\rightarrow I}_{t,c} = BD(E_{t,c},\gamma^h / 2) \\ t^{I\rightarrow R}_{t,c} = BD(I_{t,c},\sigma^h / 2) \\ S_{t+1,c} = S_{t,c} - t^{S\rightarrow E}_{t,c} \\ E_{t+1,c} = E_{t,c} + t^{S\rightarrow E}_{t,c} - t^{E\rightarrow I}_{t,c} \\ I_{t+1,c} = I_{t,c} + t^{E\rightarrow I}_{t,c} - t^{I\rightarrow R}_{t,c} \\ R_{t+1,c} = R_{t,c} + t^{I\rightarrow R}_{t,c}\end{aligned}$$ where $BD(n,p)$ represents a sample taken from a binomial distribution with $n$ samples and $p$ success probability. We note that during the simulation, the $t$ transition numbers are not calculated by sampling a binomial distribution, but by performing an independent trial for each human agent with the appropriate transition probabilities and recording the number of successes. While the resulting $t$ values are equivalent to sampling a binomial distribution directly, performing the individual trials allow us to track the state of each agent individually. This is necessary to update the populations in the next step based on the movement of agents determined by the mobility model used. ### Mosquitoes {#mosquitoes .unnumbered} We model the vector population in each grid cell stochastically, where mosquitos have two pertinent life-stages: aquatic (eggs, larvae and pupae, $A$) and adult females ($V$) [@yang2009assessing]. We keep track of the number of mosquitoes for each grid cell and calculate the transmission between the classes stochastically based on the rates calculated from the parameters of the model, some of them being dependent on the temperature. For this, we denote the total number of mosquitoes in each class by $A_{t,c}$ and $V_{t,c}$ respectively for timestep $t$ and cell $c$. We then calculate the changes in mosquito numbers of each mosquito class in each cell according to the following rules. $$\begin{aligned} d^A &= BD( A_{t,c}, \dot{\mu}^v_A/2 ) \\ t^{A\rightarrow V} &= BD( A_{t,c} - d^A, \dot{\epsilon}^v_A/2 ) \\ d^V &= B( V_{t,c}, \dot{\mu}^v_V /2) \\ t^{V\rightarrow A} &= PD\left[ cf \dot{\theta}^v_A/2 \left(1 - \frac{A}{K_{t,c}} \right) V \right] \label{eq:mosqchanges1}\end{aligned}$$ and then update the mosquito populations accordingly $$\begin{aligned} A_{t+1,c} &= A_{t,c} - d^A - t^{A\rightarrow V} + t^{V\rightarrow A} \\ V_{t+1,c} &= V_{t,c} - d^V + t^{A\rightarrow V} \label{eq:mosqchanges2}\end{aligned}$$ Here $PD(x)$ represents a sample taken from a Poisson distribution with a mean of $x$. The coefficients $c$ and $f$ are the fraction of eggs hatching to larvae and the fraction of female mosquitoes hatched from all eggs, respectively. For simplicity and lack of quantifications for the local mosquito population, we assume these to be 1 [@yang2009assessing]. Moreover, $\dot{\epsilon}^v_A$ denotes the rate of transition from aquatic to adults, $\dot{\mu}^v_A$ and $\dot{\mu}^v_V$ are the mortality rates for aquatic and adult mosquitoes, $\dot{\theta}^v_A$ is the intrinsic oviposition rates. The logistic term $\left(1 - \frac{A}{K_{t,c}} \right)$ can be understood as the physical/ecological available capacity to receive eggs, scaled by the carrying capacity term $K_{t,c}$ in each cell. The effective carrying capacity $K_{t,c}$ is defined as: $$K_{t,c} = x_v \frac{W_c + H_c}{2}$$ where $x_v$ is the average number of mosquitos per human, $W_c$ and $H_c$ are respectively the number people whose works or home location is in the cell $c$. This form assumes that the number of mosquitos in a cell scales with the average number of people found there, i.e. the mean of the nighttime population (defined by the number of home locations in that cell) and daytime population (defined by the number of work locations). Depending on the efficiency of vector control mechanisms, the number of female Aedes mosquitoes per residence varies greatly between countries. In Puerto Rico, the number of mosquitoes per home appears to be between 5 and 10 per home [@newton1992model], whereas in Singapore, this number is estimated as slightly greater than 0.2 [@ooi2006dengue]. This means that the average number of mosquitoes per human in Singapore should be in the range from $0.004$ to $0.01$. Note that such incorporation of aquatic mosquitoes in our models assumes that every cell contains some breeding sites, which is necessary to sustain a mosquito population if we do not allow mosquitoes to travel between cells. All the aquatic mosquitoes ($A^V_t$) that become adult mosquitoes at time $t$ are susceptible ($S^V_t$) and they can eventually become exposed ($E^V_t$) if they a bite an infected human and they become infected ($I^V_t$) after an incubation time as shown in \[fig:tempCases\](b). Both the aquatic and the adult mosquitoes can die with given probabilities ($\mu_A$ and $\mu_V$ respectively). Similarly to the human epidemiological models, the equations describing the vector dynamics are: $$\begin{aligned} t^{S^V\rightarrow E^V}_{t,c} = BD(S^V_{t,c},\lambda^{h\rightarrow v}_{t,c} / 2) \\ t^{E^V\rightarrow I^V}_{t,c} = BD(E^V_{t,c}, \dot{\gamma}^v / 2) \\ S^V_{t+1,c} = S^V_{t,c} - t^{S^V\rightarrow E^V}_{t,c} \\ E^V_{t+1,c} = E^V_{t,c} + t^{S^V\rightarrow E^V}_{t,c} - t^{E^V\rightarrow I^V}_{t,c} \\ I^V_{t+1,c} = I^V_{t,c} + t^{E^V\rightarrow I^V}_{t,c} \\\end{aligned}$$ where the transition rate human-to-vector $\lambda^{h\rightarrow v}_{t,c}$ is defined as[@WesolowskiPNAS2015]: $$\lambda^{h\rightarrow v}_{t,c} = a \dot{\phi}^{h \rightarrow v} S^V_{t,c} \frac{I^v_{t,c}}{N_{t,c}}.$$ These transitions are function on two temperature dependent parameters such as $\dot{\gamma}^v$ and $\dot{\phi}^{h \rightarrow v}$. ### Summary {#summary .unnumbered} Using these equations, running the model means iterating the following two steps: 1) Evaluate change of states for every human using individual Bernoulli-trials, and the change in mosquito populations in each cell using Eqs. (\[eq:mosqchanges1\]) and (\[eq:mosqchanges2\]); 2) Update the locations of human agents based on the mobility model and recalculate the number of humans of each class in each cell accordingly. We can characterize the mosquito population dynamics and the epidemics based on the ODE representation of the previous model (see SI for the corresponding equations). Using these, we can derive the basic offspring number ($Q$), that is, the mean number of viable female offspring produced by one female adult during its entire time of survival (and in the absence of any density-dependent regulation) as: $$Q = \frac{\dot{\epsilon}^v_A}{\dot{\epsilon}^v_A + \dot{\mu}^v_A} \frac{cf \dot{\theta}}{\dot{\mu}^v_V}$$ All parameters defining $Q$ are temperature-dependent (see below). For a fixed temperature $T_0$ it is possible to derive expressions for the expected population sizes of each mosquito life-stage modelled. These are used to initialize the system, given the temperature present at the initial timepoint: $$A(T_0) = K \left(1 - \frac{1}{Q(T_0)} \right) \\ V(T_0) = K \left(1 - \frac{1}{Q(T_0)} \right) \frac{\dot{\epsilon}^v_A(T_0)}{\dot{\mu}^v_V(T_0)}$$ Including the humans, the expression for dengue’s basic reproductive number is defined similarly to previous modeling approaches [@wearing2006ecological; @lourencco2013natural] but without human mortality: $$\dot{R_0} = \frac{V}{N} \frac{ a^2 \dot{\phi}^{v \rightarrow h}}{\sigma^h \dot{\mu}^v_V} \label{eq:repnumber}$$ We note that as necessary, our model includes some simplifications. Most importantly, parameter values for mosquito population modeling come from controlled experiments performed in laboratory studies [@yang2009assessing]. Clearly, it seems prohibitably challenging to directly estimate these parameters in the wild, as tracking individual mosquitoes is infeasible; studies can test the applicability of the models by comparing predictions to estimates of observed mosquito population sizes. Furthermore, accurately measuring mosquito populations itself present difficulties in real-world conditions. We note that uncertainties in parameters are inherently linked in our model; e.g. a shorter mosquito lifespan could be offset by higher bite rate as evident from Eq. \[eq:repnumber\]. This way, any calibration process among the parameter values will likely be degenerate. Another main limitation in our dataset is that we have no estimate of any existing immunity to dengue in the population. While dengue has mulitple strains, and partial or full immunity can be acquired after being infected with a specific strain, the picture is quite complex. Similarly to uncertainty of parameters for mosquitoes, uncertainty in the size of susceptible population is linked to any variations in other parameters. For this reason, we do not perform a scaling of the population size, but use the sample obtained from the mobile phone data which covers a large part of Singapore’s population. We deal with these issues by using established values and temperature-dependent forms from the literature for most parameters [@WesolowskiPNAS2015; @lourencco20142012; @yang2009assessing], while exploring a phase space determined by variations in a small number of parameters, namely the bite rate ($a$) and average number of mosquitoes per human ($x_v$). Finding an ideal combination in for this pair of parameters allows us to calibrate the model for Singapore, while avoiding overfitting. In summary, as initial conditions for the simulations setting we consider $N = 2,307,230$ agents derived from the mobile phone data and described above. At the beginning of the simulations, i.e. January 1st 2013, we set the $I_{init}$ number of initial infected agents as retrieved from the official Singapore’s government’s portal [@weekly]. In particular $I_{init} = 242$ infected individuals in 93 different cells of the grid $G$. In order to keep the outbreaks alive we ensured that the number of infected individuals in the systems always $I >=100$ as visible in \[fig:simCases2013\]. The number of initial aquatic and adult mosquito have been computed for each values of the parameter $x_v$ from January 1st 2011. For each day from January 1st 2011 to December 31st 2012 we collected the temperature and we simulated the dynamics of aquatic and adult mosquitoes in each cell given the population estimated from the mobile phone and following the Eq \[eq:mosqchanges1\] and Eq \[eq:mosqchanges2\]. In this way, for each value of the parameter $x_v$ it has been possible to set a stable number of aquatic and adult mosquitoes the first day of the simulation. Acknowledgements {#acknowledgements .unnumbered} ================ E.M. would like to thank the HERUS Lab at the École polytechnique fédérale de Lausanne, the Swiss Mobiliar insurance company, the ENAC Exploratory Grant 2018 (Preparatory Funding Scheme) with the project entitled “Risk evaluation of mosquito-borne disease transmission through urban commutes pathways” and the“Healthy Cities Towards a One Health agenda for urban space” from the Habitat Research Center at EPFL for partially founding this research . The authors would like to thank Prof. Alessandro Vespignani for the helpful suggestions regarding the epidemic model. The authors thank all sponsors and partners of the MIT Senseable City Laboratory including Allianz, the Amsterdam Institute for Advanced Metropolitan Solutions, the Fraunhofer Institute, Kuwait-MIT Center for Natural Resources and the Environment, Singapore-MIT Alliance for Research and Technology (SMART) and all the members of the Consortium. Author contributions statement {#author-contributions-statement .unnumbered} ============================== E.M. and D.K. conceived the experiments, E.M. conducted the experiments, E.M. analysed the results, D.K. analyzed the mobile phone data. E.M. and D.K. wrote the paper. All authors reviewed the manuscript. Additional information {#additional-information .unnumbered} ====================== The authors declare that there are not any competing financial and/or non-financial interests in relation to the work described. Supplementary materials\ \ Assessing the interplay between human mobility and mosquito borne diseases in urban environments {#supplementary-materials-assessing-the-interplay-between-human-mobility-and-mosquito-borne-diseases-in-urban-environments .unnumbered} ================================================================================================ Emanuele Massaro$^$, Daniel Kondor, Carlo Ratti\ \[email protected]\ Mobility Models {#mobility-models-1 .unnumbered} --------------- We report the comparison between the four mobility models used in this research: i) derived from mobile phone data, ii) random, iii) derived from a Levy flight distribution and iv) derived from the radiation model. In \[fig:flowsCorr\] we report the Pearson correlation ($P_c$) coefficient between the 3 models. In the scatterplots each point correspond to the flow (i.e. total number of commuters) from a location $w$ to a location $h$. mobility models. ![Correlation of amount of people travelling between the census areas: each point corresponds to the flow between two census areas.[]{data-label="fig:flowsCorr"}](flowsCorrAll.pdf){width="0.7\linewidth"} \ \[fig:hw\], \[fig:hwrand\], \[fig:hwlevy\] and \[fig:hwradial\] show displacement of the agents in their home and work locations for the different mobility models. ![Home and work locations from mobile phone data. Count of users in the home and work locations respectively determined from the mobile phone dataset in each cell. The majority of jobs are located in the Central Business District, whereas the home locations are more equally distributed.[]{data-label="fig:hw"}](hwlocationmobile){width="100.00000%"} ![Home and work locations from the random model. The home locations are taken from the mobile phone data model while the work locations are randomly assigned.[]{data-label="fig:hwrand"}](randModel){width="100.00000%"} ![Home and work locations from the Levy model. The home locations are randomly assigned while the work locations are given with a distance from a Levy flight distribution.[]{data-label="fig:hwlevy"}](hwlocationlevy){width="100.00000%"} ![Home and work locations from the radiation model.. The home locations are assigned from the census data while the work locations are assigned with a distance following the radiation model.[]{data-label="fig:hwradial"}](hwlocationradial){width="100.00000%"} Temperature dependent parameters {#temperature-dependent-parameters .unnumbered} -------------------------------- Most of the parameters (as reported in Table 1 in the main text) used in our methodological approach depend on the temperature. The equations governing those parameters are the following:\ \ $\begin{array}{rcl} \epsilon_A^v(T) & = & 0.131 - 0.05723T + 0.01164T^2 - 0.001341T^3 + \\ & & + 0.00008723T^4 - 3.017\cdot10^{-6}T^5 + 5.153\cdot10^{-8}T^6 - 3.42\cdot10^{-10}T^7 \\ \mu_A^v(T) & = & 2.13 - 0.3787T + 0.02457T^2 - 6.778\cdot10^{-4}T^3 + 6.794\cdot10^{-6} T^4\\ \mu_V^v(T) & = & RHF * (0.8692 - 0.1599T + 0.01116T^2- 3.408\cdot10^{-4})T^3 + 3.809\cdot10^{-6}T^4) \\ \theta_A^v(T) & = & -5.4 + 1.8T - 0.2124T^2 + 0.01015T^3 - 1.515\cdot10^{-4}T^4 \\ \gamma_V^v(T) & = & \frac{(3.3589 \cdot 10^{-3} * Tk)/298exp( (1500/R)(1/298-1/Tk))}{1 + exp( ( 6.203\cdot10^{21} )/R * (1/(-2.176 \cdot 10^{30})) - 1/Tk )} \end{array} $\ \ where $Tk$ is the degrees in kelvin.\ \ $\begin{array}{rcl} \phi^{h \rightarrow v} (T) & = & 1.004\cdot 10^{-3} T (T - 12.286) \cdot (32.461-T)^{1/2} \\ \phi^{v \rightarrow h} (T) & = & 0.0729T - 0.97. \end{array} $\ \ We have also included an adult mortality factor based on relative humidity[@WesolowskiPNAS2015]. Temperature and relative humidity are converted to a vapor pressure measure, $VP= 6.11 \cdot 10^{(7.5T/273.3 + T)/ 10}$. This value is converted to a relative humidity factor RHF) based on the following rules: If $10<VP<30$, $RHF= 1.2 - 0.2 \cdot VP$, and if $VP \geq 30$, $RHF= 0.5$.\ In \[fig:partemp\] we show the values of the described parameters for temperature between -10C and 40C while in \[fig:tempParameters\] we show the values of the parameters during the years 2013 and 2014. ![The temperature-dependent parameters as function of the temperature.[]{data-label="fig:partemp"}](valParameters.pdf){width="1\linewidth"} ![The temperature-dependent parameters used in the ento-epidemiological framework for Singapore in 2013-2014.[]{data-label="fig:tempParameters"}](temperatureParameters.pdf){width="1\linewidth"} Structural Similarity Index {#structural-similarity-index .unnumbered} --------------------------- In order to compare and quantify the spatial prediction of the simulations with the real case scenario, we use the structural similarity index. The Structural SIMilarity (SSIM) index is a method originally proposed for measuring the similarity between two images, but is applicable when comparing structural properties of 2-dimensional data, i.e. the spatial distribution of dengue cases in our case. The SSIM index can be viewed as a quality measure of one of the images being compared, provided the other image is regarded as of perfect quality. It is an improved version of the universal image quality index proposed before [@Wang2004; @wang2009mean] and is computed as: $$SSIM(x,y) = \frac{(2\mu_x \mu_y + c_1)(2\sigma_{xy}+c_2)}{(\mu^2_x\mu^2_y + c_1 )(\sigma^2_x + \sigma^2_y +c_2) }$$ where $x$ and $y$ are appropriate-sized windows of the images to compare, where $\mu_x$ and $\mu_y$ are the average of $x$ and $y$, $\sigma^2_x$ and $\sigma^2_y$ are the variances of $x$ and $y$ while $\sigma_{xy}$ is the covariance of $x$ and $y$. The parameters $c_{1}=(k_{1}L)^{2}, c_{2}=(k_{2}L)^{2}$ are two variables to stabilize the division with a weak denominator, where $L$ is the dynamic range of the discrete pixel values. The two additional parameters are $k_{1}=0.01$ and $k_{2}=0.03$ by default. To obtain a similarity metric between two images, the SSIM values are averaged over all possible subsections of the images, defined by sliding windows of size $7\times 7$ pixels. The range of the value of the SSIM index is between $0$ and $1$: when two images are nearly identical, their SSIM is close to $1$. For each epidemiological week in the period we compute the SSIM between the real case and the three simulated scenarios using the the `Python` function `structural_similarity` from the package `skimage`[^1]. An example of SSIM in a toy grid is reported in \[fig:SSIMEx\]. ![Structural similarity index (SSIM) illustration. We generated a $10 \times 10$ grid in which in each cell we assign a random number between $-1$ and $1$. This grid is mathematically described by a matrix $B^{10 \times 10}$ in which each element of the matrix $B_{i,j}$ is a random number between $-1$ and $1$. This matrix represents our benchmark to test the SSIM index. We then generate other three grids (from left to right bottom) starting from the benchmark in which for each $B_{i,j}$ we add or subtract random number between $0$ and $1$ times $0.25$, $0.5$ and $0.75$ respectively. In this way we are able to compare three different scenarios with the benchmark with different degree of difference from the original one. As we can observe similar images generate greater SSIM if compared with the benchmark.[]{data-label="fig:SSIMEx"}](ssim.png){width="0.8\linewidth"} [0.5]{} ![Statistical analysis of mobile phone data. A) Distribution of number of locations (antennas). B) events (calls / texts) per user and distribution of visitation frequencies of user locations as a function of location ranks (here the locations still refer to antennas). C) Distribution of number of locations (antennas) and (b) events (calls / texts) per user and distribution of visitation frequencies of user locations as a function of location ranks (here the locations still refer to antennas). D) Distribution of the commute matrix elements (i.e. the number of users who commute between any two cells).[]{data-label="fig:s1"}](userpfreq.pdf "fig:"){width="1\linewidth"} [0.5]{} ![Statistical analysis of mobile phone data. A) Distribution of number of locations (antennas). B) events (calls / texts) per user and distribution of visitation frequencies of user locations as a function of location ranks (here the locations still refer to antennas). C) Distribution of number of locations (antennas) and (b) events (calls / texts) per user and distribution of visitation frequencies of user locations as a function of location ranks (here the locations still refer to antennas). D) Distribution of the commute matrix elements (i.e. the number of users who commute between any two cells).[]{data-label="fig:s1"}](snum.pdf "fig:"){width="1\linewidth"} \ [0.5]{} ![Statistical analysis of mobile phone data. A) Distribution of number of locations (antennas). B) events (calls / texts) per user and distribution of visitation frequencies of user locations as a function of location ranks (here the locations still refer to antennas). C) Distribution of number of locations (antennas) and (b) events (calls / texts) per user and distribution of visitation frequencies of user locations as a function of location ranks (here the locations still refer to antennas). D) Distribution of the commute matrix elements (i.e. the number of users who commute between any two cells).[]{data-label="fig:s1"}](matrix500wfreq.pdf "fig:"){width="1\linewidth"} [0.5]{} ![Statistical analysis of mobile phone data. A) Distribution of number of locations (antennas). B) events (calls / texts) per user and distribution of visitation frequencies of user locations as a function of location ranks (here the locations still refer to antennas). C) Distribution of number of locations (antennas) and (b) events (calls / texts) per user and distribution of visitation frequencies of user locations as a function of location ranks (here the locations still refer to antennas). D) Distribution of the commute matrix elements (i.e. the number of users who commute between any two cells).[]{data-label="fig:s1"}](matrix500cellcnts.pdf "fig:"){width="1\linewidth"} ![Fraction of the population in Singapore’s districts according to the 2010 census versus the home locations determined from the mobile phone dataset. With a correlation coefficient of 0.96, the two spatial distributions are highly linearly correlated.[]{data-label="fig:singcensus"}](censusData.pdf){width="0.5\linewidth"} ![Basic demographic characteristics of the Singapore resident population by their registered place of address from the Census of Population 2010. The Singapore resident population comprises Singapore citizens and permanent residents. Of the $3.77$ million Singapore residents as at end-June 2010, about $57\%$ were concentrated in ten planning areas. There were five planning areas with more than $200,000$ Singapore residents. Bedok, Jurong West and Tampines each had more than $250,000$ Singapore residents, with Bedok having the most number at $294,500$ in 2010. The other two planning areas with more than $200,000$ Singapore residents in 2010 were Woodlands ($245,100$) and Hougang ($216,700$). Shapefile data are downloaded from the Singapore open data portal https://data.gov.sg, while population data are downloaded from https://www.worldpop.org/[]{data-label="fig:s2"}](populationShape.pdf){width="\linewidth"} ![The 2598 cells used in this research.[]{data-label="fig:grids"}](grids){width="\linewidth"} [0.5]{} ![Logarithmic values of the $R^2$ of the temporal estimation for each couple of parameters in the different mobility models where the white spaces correspond to the negative value of the $R^2$.[]{data-label="fig:sIM"}](mobilePhoneIm.pdf "fig:"){width="1\linewidth"} [0.5]{} ![Logarithmic values of the $R^2$ of the temporal estimation for each couple of parameters in the different mobility models where the white spaces correspond to the negative value of the $R^2$.[]{data-label="fig:sIM"}](randomIm.pdf "fig:"){width="1\linewidth"} \ [0.5]{} ![Logarithmic values of the $R^2$ of the temporal estimation for each couple of parameters in the different mobility models where the white spaces correspond to the negative value of the $R^2$.[]{data-label="fig:sIM"}](LevyIm.pdf "fig:"){width="1\linewidth"} [0.5]{} ![Logarithmic values of the $R^2$ of the temporal estimation for each couple of parameters in the different mobility models where the white spaces correspond to the negative value of the $R^2$.[]{data-label="fig:sIM"}](radialIm.pdf "fig:"){width="1\linewidth"} [^1]: <http://scikit-image.org/docs/dev/auto_examples/transform/plot_ssim.html>	{ "pile_set_name": "ArXiv" }
CERN-PH-TH/2004-107\ hep-ph/0406156 [The transverse momentum distribution\ of the Higgs boson at the LHC]{} [^1] [Massimiliano Grazzini]{}\ Theory Division, CERN, CH-1211 Geneva 23, Switzerland [Abstract]{} > 10000 > > We present perturbative predictions for the transverse momentum ($q_T$) distribution of the Higgs boson at the LHC. At small $q_T$ the logarithmically-enhanced terms are resummed to all orders up to next-to-next-to-leading logarithmic accuracy. The resummed component is consistently matched to the next-to-leading order calculation valid at large $q_T$. The results, which implement the most advanced perturbative information that is available at present for this observable, show a good stability with respect to perturbative QCD uncertainties. hep-ph/0406156\ June 2004 The search for the Higgs boson is among the major issues in the LHC physics program [@atlascms]. In recent years much effort has been devoted to refining the theoretical predictions for the various Higgs production channels and the corresponding backgrounds, which are now known to next-to-leading order accuracy (NLO) in most of the cases [@leshouches]. For the main Standard Model (SM) Higgs production channel, gluon–gluon fusion, even next-to-next-to leading (NNLO) QCD corrections to the total rate have been computed [@NNLOtotal]. Nonetheless, predictions for less inclusive observables are definitely required to perform realistic studies. In particular, an accurate knowledge of the transverse-momentum distribution of the Higgs boson can be important to enhance the statistical significance of the signal over the background. In this contribution we focus on the dominant SM Higgs production channel, gluon–gluon fusion. When the transverse momentum $q_T$ of the Higgs boson is of the order of its mass $M_H$, the perturbative series is controlled by a small expansion parameter, ${\alpha_{\mathrm{S}}}(M_H^2)$, and the fixed order prediction is reliable. The leading order (LO) calculation [@Ellis:1987xu] shows that the large-$M_t$ approximation ($M_t$ being the mass of the top quark) works well as long as both $M_H$ and $q_T$ are smaller than $M_t$. In the framework of this approximation, the NLO QCD corrections have been computed [@deFlorian:1999zd; @Ravindran:2002dc; @Glosser:2002gm]. The small-$q_T$ region ($q_T\ll M_H$) is the most important, because it is here that the bulk of events is expected. In this region the convergence of the fixed-order expansion is spoiled, since the coefficients of the perturbative series in ${\alpha_{\mathrm{S}}}(M_H^2)$ are enhanced by powers of large logarithmic terms, $\ln^m (M_H^2/q_T^2)$. To obtain reliable perturbative predictions, these terms have to be systematically resummed to all orders in ${\alpha_{\mathrm{S}}}$ [@Dokshitzer:hw] (see also the list of references in Sect. 5 of Ref. [@Catani:2000jh]). To correctly enforce transverse-momentum conservation, the resummation has to be carried out in $b$ space, where the impact parameter $b$ is the variable conjugate to $q_T$ through a Fourier transformation. In the case of the Higgs boson, the resummation has been explicitly worked out at leading logarithmic (LL), next-to-leading logarithmic (NLL) [@Catani:vd], [@Kauffman:cx] and next-to-next-to-leading logarithmic (NNLL) [@deFlorian:2000pr] level. The fixed-order and resummed approaches then have to be consistently matched at intermediate values of $q_T$, so as to avoid double counting. In the following we present predictions for the Higgs boson $q_T$ distribution at the LHC within the formalism described in Ref. [@Catani:2000vq]. In particular, we include the best perturbative information that is available at present: NNLL resummation at small $q_T$ and NLO calculation at large $q_T$. An important feature of our formalism is that a unitarity constraint on the total cross section is automatically enforced, such that the integral of the spectrum reproduces the known results at NLO [@Dawson:1991zj] and NNLO [@NNLOtotal]. More details can be found in Ref. [@Bozzi:2003jy]. Other phenomenological studies are presented in [@recent]. We are going to present quantitative results at NLL+LO and NNLL+NLO accuracy. At NLL+LO (NNLL+NLO) accuracy the NLL (NNLL) resummed result is matched to the LO (NLO) perturbative calculation valid at large $q_T$. As for the evaluation of the fixed-order results, the Monte Carlo program of Ref. [@deFlorian:1999zd] has been used. The numerical results are obtained by choosing $M_H=125$ GeV and using the MRST2002 set of parton distributions [@Martin:2003es]. At NLL+LO, LO parton densities and 1-loop ${\alpha_{\mathrm{S}}}$ have been used, whereas at NNLL+NLO we use NLO parton densities and 2-loop ${\alpha_{\mathrm{S}}}$. ----------- =12truecm ----------- The NLL+LO results at the LHC are shown in Fig. \[fig1\]. In the left panel, the full NLL+LO result (solid line) is compared with the LO one (dashed line) at the default scales $\mu_F=\mu_R=M_H$. We see that the LO calculation diverges to $+\infty$ as $q_T\to 0$. The effect of the resummation starts to be relevant below $q_T\sim 100$ GeV. In the right panel we show the NLL+LO band obtained by varying $\mu_F=\mu_R$ between $1/2 M_H$ and $2M_H$. ----------- =12truecm ----------- The corresponding NNLL+NLO results are shown in Fig. \[fig2\]. In the left panel, the full result (solid line) is compared with the NLO one (dashed line) at the default scales $\mu_F=\mu_R=M_H$. The NLO result diverges to $-\infty$ as $q_T\to 0$ and, at small values of $q_T$, it has an unphysical peak that is produced by the numerical compensation of negative leading and positive subleading logarithmic contributions. Notice that at $q_T \sim 50$ GeV, the $q_T$ distribution sizeably increases when going from LO to NLO and from NLO to NLL+LO. This implies that in the intermediate-$q_T$ region there are important contributions that have to be resummed to all orders rather than simply evaluated at the next perturbative order. The resummation effect starts to be visible below $q_T\sim 100$ GeV, and it increases the NLO result by about $40\%$ at $q_T=50$ GeV. The right panel of Fig. \[fig2\] shows the scale dependence computed as in Fig. \[fig1\]. Comparing Figs. 1 and 2, we see that the NNLL+NLO band is smaller than the NLL+LO one and overlaps with the latter at $q_T \ltap 100$ GeV. This suggests a good convergence of the resummed perturbative expansion. ----------- =10truecm ----------- A practical application of the results discussed in this contribution has recently been presented [@dddgp] in the context of the Higgs search in the channel $pp\to H\to WW\to l\nu l\nu$. In this channel a jet veto is necessary to cut events with high-$q_T$ $b$ jets from the $t{\bar t}$ background. It is known that the effect of higher order corrections depends on the actual value of this cut [@Catani:2001cr]. As an approximate way to include higher order effects in the analysis, the QCD-corrected Higgs spectrum presented above has been used in Ref. [@dddgp] to reweight signal events generated with the PYTHIA Monte Carlo [@pythia]. The same method has been applied to correct the main $WW$ background, by using a NLL+LO calculation performed within the same resummation formalism [@Catani:2000vq]. The resulting integrated luminosity needed to discover the Higgs is shown in Fig. \[fig3\] and compared with the other channels. We see that, if systematical errors are under control, around the $WW$ threshold the Higgs can be discovered with an integrated luminosity of about half fb$^{-1}$. [99]{} CMS Coll., [Technical Proposal]{}, report CERN/LHCC/94-38 (1994); ATLAS Coll., [ATLAS Detector and Physics Performance: Technical Design Report]{}, Vol. 2, report CERN/LHCC/99-15 (1999). , to appear in the Proceedings of the Workshop on Physics at TeV Colliders, Les Houches, France, 2003, hep-ph/0406152. S. Catani, D. de Florian and M. Grazzini, JHEP [0105]{} (2001) 025; R. V. Harlander and W. B. Kilgore, Phys. Rev. D [64]{} (2001) 013015, Phys. Rev. Lett. [88]{} (2002) 201801; C. Anastasiou and K. Melnikov, Nucl. Phys. B [646]{} (2002) 220; V. Ravindran, J. Smith, W. L. van Neerven, Nucl. Phys. B [665]{} (2003) 325. R. K. Ellis, I. Hinchliffe, M. Soldate and J. J. van der Bij, Nucl. Phys. B [297]{} (1988) 221; U. Baur and E. W. Glover, Nucl. Phys. B [339]{} (1990) 38. D. de Florian, M. Grazzini and Z. Kunszt, Phys. Rev. Lett. [82]{} (1999) 5209. V. Ravindran, J. Smith and W. L. Van Neerven, Nucl. Phys. B [634]{} (2002) 247. C. J. Glosser and C. R. Schmidt, JHEP [0212]{} (2002) 016. G. Parisi and R. Petronzio, Nucl. Phys. B [154]{} (1979) 427; Y. L. Dokshitzer, D. Diakonov and S. I. Troian, Phys. Rep. [58]{} (1980) 269; J. C. Collins, D. E. Soper and G. Sterman, Nucl. Phys. B [250]{} (1985) 199. S. Catani et al., hep-ph/0005025, in the Proceedings of the CERN Workshop on [Standard Model Physics (and more) at the LHC]{}, eds. G. Altarelli and M.L. Mangano (CERN 2000-04, Geneva, 2000), p. 1. S. Catani, E. D’Emilio and L. Trentadue, Phys. Lett. B [211]{} (1988) 335. R. P. Kauffman, Phys. Rev. D [45]{} (1992) 1512. D. de Florian and M. Grazzini, Phys. Rev. Lett. [85]{} (2000) 4678, Nucl. Phys. B [616]{} (2001) 247. S. Catani, D. de Florian and M. Grazzini, Nucl. Phys. B [596]{} (2001) 299. S. Dawson, Nucl. Phys. B [359]{} (1991) 283; A. Djouadi, M. Spira and P. M. Zerwas, Phys. Lett. B [264]{} (1991) 440; M. Spira, A. Djouadi, D. Graudenz and P. M. Zerwas, Nucl. Phys. B [453]{} (1995) 17. G. Bozzi, S. Catani, D. de Florian and M. Grazzini, Phys. Lett. B [564]{} (2003) 65. C. Balazs and C. P. Yuan, Phys. Lett. B [478]{} (2000) 192; E. L. Berger and J. w. Qiu, Phys. Rev. D [67]{} (2003) 034026; A. Kulesza, G. Sterman, W. Vogelsang, preprint BNL-HET-03/20 \[hep-ph/0309264\]. A. D. Martin, R. G. Roberts, W. J. Stirling and R. S. Thorne, Eur. Phys. J. C [28]{} (2003) 455. G. Davatz, G. Dissertori, M. Dittmar, M. Grazzini and F. Pauss, JHEP [0405]{} (2004) 009. S. Catani, D. de Florian and M. Grazzini, JHEP [0201]{} (2002) 015. T. Sjostrand, P. Eden, C. Friberg, L. Lonnblad, G. Miu, S. Mrenna and E. Norrbin, Comput. Phys. Commun. [135]{} (2001) 238. [^1]: Talk given at the XXXIXth Rencontres de Moriond, La Thuile, Italy, 28 March–4 April 2004.	{ "pile_set_name": "ArXiv" }
--- abstract: 'Recent additions to the population of Long-duration Gamma Ray Burst (LGRB) host galaxies with measured metallicities and host masses allow us to investigate how the distributions of both these properties change with redshift. We form a sample out to $z \lesssim 2.5$ which we show does not have strong redshift dependent populations biases in mass and metallicity measurements. Using this sample, we find a surprising lack of evolution in the LGRB metallicity distribution across different redshifts and in particular the fraction of LGRB hosts with relatively high-metallicity, that is those with 12+log(O/H) $\geq 8.4$, remains essentially constant out to $z=2.5$. This result is at odds with the evolution in the mass metallicity relation of typical galaxies, which become progressively more metal poor with increasing redshift. By converting the measured LGRB host masses and redshifts to expected metallicities using redshift appropriate mass-metallicity relations, we further find that the increase in LGRB host galaxy mass distribution with redshift seen in the [@Perley_shoals_masses] SHOALS sample is consistent with that needed to preserve a non-evolving LGRB metallicity distribution. However, the estimated LGRB host metallicity distribution is at least a quarter dex higher at all redshifts than the measured metallicity distribution. This corresponds to about a factor of two in raw metallicity and resolves much of the difference between the LGRB host metallicity cutoffs determined by [@diff_rate_letter] and [@Perley_shoals_masses]. As LGRB hosts do not follow the general mass metallicity relations, there is no substitute for actually measuring their metallicities.' author: - 'J. F. Graham$^{\hyperref[jfg]{1}}$' - 'P. Schady$^{\hyperref[ps]{2},\hyperref[ops]{3}}$' - 'A. S. Fruchter$^{\hyperref[af]{4}}$' bibliography: - 'XS\_paper.bib' title: A Surprising Lack of LGRB Metallicity Evolution with Redshift --- Introduction ============ Upon collecting the first samples of the galaxies hosting Long-soft Gamma-Ray Bursts (LGRBs) it was apparent that LGRBs occur in blue, highly starforming, and often irregular galaxies with a preponderance that clearly separated them from the general galaxy population [@Fruchter1999; @Fruchter; @LeFlochblue; @LeFlochblue2002; @Christensen; @LeFloch2006; @Savaglio]. The seminal work of [@Fruchter] compared the hosts of LGRBs with those of Core-Collapse Supernovae (CCSNe) found in the Great Observatories Origins Deep Survey (GOODS) and found that while half of the GOODs CCSNe occurred in grand design spirals (with the other half in irregulars), only one out of the 18 LGRB host galaxies of a comparable redshift distribution was in a grand design spiral. This comparison with SNe allowed [@Fruchter] to establish that this difference was beyond that expected from a sample of galaxies drawn randomly according to their rate of star formation. As the Initial Mass Function (IMF) of blue irregular and spiral galaxies are thought to be largely similar [@Bastian], massive stellar progenitors should be just as available per unit star-formation in both galaxy types. However, the much smaller size of blue irregulars would suggest, due to the galaxy mass-metallicity relation [@Tremonti2004], that blue irregular galaxies are typically metal poor in comparison with grand design spirals. Thus [@Fruchter] postulated that LGRB formation is affected by the metallicity of their environments. [@Stanek2007] showed that the very nearest LGRB hosts all have low metallicity when compared to similar magnitude galaxies in the Sloan Digital Sky Survey (SDSS). Furthermore, [@Kewley2007] found the LGRB host sample to be comparable to extremely metal-poor galaxies in luminosity-metallicity relation, star-formation rate (SFR), and internal extinction. [@Modjaz2008] dramatically strengthened this result by taking advantage of the fact that a broad-lined Type Ic (Ic-bl) supernova has been found underlying the light of nearly [@Thone2014] every LGRB in which a deep spectroscopic search was performed [@Cano2014; @HjorthBloom2012book; @Hjorth2013]. The fact that LGRBs nearly always are associated with a Type Ic-bl SNe would suggest that LGRB progenitors probably have similar masses to those of regular Type Ic-bl SNe, thus largely eliminating the possibility that the observed LGRB metallicity bias is somehow a byproduct of a difference in the initial stellar mass functions. [@Modjaz2008] showed that Ic-bl SNe with associated LGRBs are observed to occur in host galaxies with much lower metallicities than the hosts of Type Ic-bl SNe without associated LGRBs or the bulk of the star-forming galaxies in the SDSS. More recently [@Modjaz2019] has shown that Ic-bl in the local universe typically have low host metallicities essentially identical to LGRBs. (The [-@Modjaz2008] sample had a large number of SN found in directed surveys of nearby massive galaxies that may have produced this biased result.) However, beginning with LGRB 051022 [@conference_proceedings], some LGRBs were found to be located in host galaxies with a high metallicity (near-solar and above) but this accounted for only a small fraction of the LGRB population. The dramatic metallicity difference between both the star-formation weighted SDSS and non-engine driven SNe verses LGRB samples (including normal Type Ic, c.f. [@Modjaz2019] suggests a metallicity dependent step in either the formation of the gamma-ray jet or in its ability to escape the progenitor which has either burned or lost its outer hydrogen and helium layers [@Woosley1993; @Woosley; @Langer]. To understand how LGRBs seem to have a metallicity aversion despite the existence of counterexamples, [@stats_paper] compared the metallicity distribution of the hosts of LGRBs with that of the hosts of several similar indicators of star-formation: LGRBs, Type Ic-bl, and Type II SNe as well as with the metallicity distribution of star-formation in the general galaxy population of the local universe. [@stats_paper] found that three quarters of the LGRB host population have metallicities below 12+log(O/H) $<$ 8.6 (in the @KobulnickyKewley KK04 metallicity scale), while less than a tenth of local star-formation is at similarly low metallicities. However, the supernovae were statistically consistent with the metallicity distribution of star-formation in the general galaxy population. While the LGRB sample did extend to higher redshifts than the other populations the observed metallicity difference was far too great to be a product of metallicity evolution. [@stats_paper] concluded that a low metallicity environment must be a fundamental component of the evolutionary process that forms LGRBs. [@diff_rate_letter] expanded on the analysis of [@stats_paper] by comparing the metallicity distribution of LGRB host galaxies to the that of star forming galaxies in the local universe. By effectively dividing one distribution by the other, [@diff_rate_letter] was able to directly determine the relative rate of LGRB formation as a function of metallicity in the low-redshift universe. They found a dramatic cutoff in the LGRB formation rate per unit star-formation above a metallicity of $ {\rm log(O/H)}+12 \approx 8.3 $ (in the @KobulnickyKewley scale), with LGRBs forming between ten and fifty times more frequently per unit star-formation below this cutoff than above. [@xshooter_survey] performed a detailed analysis of 96 LGRB host galaxies with ESO Very Large Telescope (VLT) X-Shooter emission-line spectroscopy, providing the largest spectroscopic sample of LGRB host galaxies and comprising most of the publicly available data at the time. [@xshooter_survey] finds that at z $<$ 1 $\sim$20% of their LGRB host galaxies have super-solar metallicities a result comparable with [@stats_paper]. [@xshooter_survey] concludes that the properties of LGRB hosts can be explained by the tendency of LGRB events to avoid metal-rich environments. [@Vergani2015] estimated the host stellar masses of the 14 z $<$ 1 BAT6 [@BAT6] host galaxies via spectral energy distribution (SED) fitting and found that those LGRBs tend to avoid massive galaxies in preference for faint low-mass star-forming galaxies typically below galaxy survey completeness limits. [@Vergani2015] estimates that with a metallicity threshold in LGRB production (of between $\frac{1}{3}$ to $\frac{1}{2}$ solar) the typical galaxy mass distribution would then reproduce that of the LGRB host galaxy population (i.e. primarily a reduction in the high-mass end of the distribution). In a similar analysis, [@Perley_shoals_masses] used Spitzer rest-frame near-IR (NIR) luminosity observations to calculate masses for 82 LGRB host galaxies from the Swift GRB Host Galaxy Legacy Survey (SHOALS) and also use the distribution of these masses to estimate a metallicity threshold. However [@Perley_shoals_masses] estimates the metallicity threshold to be much higher at “approximately the solar value." They also find that dust-obscured LGRBs dominate the massive host population while little dust is seen in low-mass hosts and that host metallicity has little impact on LGRB production at z $>$ 1.5 while preventing most LGRB events at lower redshifts. Here (in [first@refmzd,@]{}) we will fully exploit the new larger emission-line metallicity LGRB host galaxy sample of [@xshooter_survey] (in concert with the @stats_paper sample) to examine how the LGRB metallicity distribution changes with redshift. First however (in [first@refzd,@]{}) we will analyze and address the redshift biases of the different LGRB populations as this is necessary to validate our samples. Also (in [first@refemzd,@]{}), in order to test the validity in the general approach of using galaxy mass to estimate LGRB host metallicities, we will compare our observed LGRB metallicity distributions (as a function of redshift) with the expected metallicity distribution for galaxies with the mass and redshift the LGRB population. Our results are summarized in [first@refResults\_Summary,@]{}, and in [first@refDiscussion,@]{} we discuss some possible explanations for what we see. Methods ======= Sample selection ---------------- ### Host Metallicities {#Host_Metallicities} The vast majority of LGRB host galaxy emission-line metallicity measurements are contained in either the [@stats_paper] sample or the X-Shooter observations compiled in [@xshooter_survey]. The objects of the later sample are heavily dominated by observations of [@TOUGH]: The Optically Unbiased Gamma-Ray Burst Host Survey (TOUGHS) during the time period when TOUGHS was ongoing. TOUGHS is a homogeneous sample of 69 [Swift]{} GRB hosts selected via their bursts high-energy properties and locations on the sky and not dependent on the optical properties of the galaxies observed. [@xshooter_survey] was able to obtain hosts metallically measurements for about 3 times as many objects as the previous sample of [@stats_paper] yielding a combined sample (after removing duplicates and constantly recalculating the metallicities) of 45 objects (see [first@refxsd\_data\_table,@]{}). Furthermore X-Shooter’s integrated IR spectroscopic channel extends the redshift overage of the LGRB host metallically sample out to [z]{} = 2.47[^1]. In combination [@xshooter_survey] and [@stats_paper] provide a sample of sufficient size and redshift range to allow division into different redshift bins as necessary to study the evolution of the LGRB metallically distribution with redshift as we will do here. Two objects overlap between the samples (LGRBs 050824 and 051022) whose measured metallicity values differ by 0.23 and 0.31 dex between the [@stats_paper] and [@xshooter_survey] estimates respectively. In the combined sample the duplication has been removed and the flux measurements from the [@xshooter_survey] spectroscopy used. \[dupout\] We standardize on the [@xshooter_survey] spectroscopy because they publish errors on their line flux values whereas the LGRB 050824 line fluxes of [@Sollerman2007] (used in @stats_paper) and the LGRB 051022 line fluxes of [@Levesque2] do not. Furthermore the [@stats_paper] metallically for LGRB 051022 was determined from equivalent width observations [@conference_proceedings], not line flux values, which while an established technique [@KobulnickyPhillips], introduces an difference in method that we can easily avoid here. Note that the metallicity values and relative measured line fluxes for the host of LGRB 051022 have previously been noted as being significantly different between the various observations mentioned above, potentially due to as yet unresolved differences in the metallicity across the galaxy [@obs_paper]. We have removed LGRB 020819B from our combined sample due to the revised host galaxy association of [@Perley020819B] and the poor metallicity constraint on the now correctly associated host galaxy. ### Host Masses While neither the [@stats_paper] or [@xshooter_survey] provide host galaxy masses or even have the required IR observations required for such mass measurements, the [@Perley_shoals_masses] SHOALS sample does. Unfortunately the [@Perley_shoals_masses] SHOALS sample has a paucity of objects at low redshift (only 3 at z $<$ 0.5) so we supplement it with the [@Svensson] population which is predominately at z $<$ 1 and also has mass values based on deep IR observations. Two objects overlap between the samples (LGRBs 060218 and 080319B) whose log masses differ by 0.24 and -0.43 between the [@Svensson] and [@Perley_shoals_masses] estimates respectively. In our combined sample the duplication has been removed and the [@Perley_shoals_masses] values used. Metallicity Determination {#Metallicity_Determination} ------------------------- Here we use metallicities determined via the R$_{23}$ diagnostic in the [@KobulnickyKewley] (KK04) scale. The R$_{23}$ method is the primary metallicity diagnostic for galaxies at $z>0.3$ (especially those where the faint \[O III\] 4363 [Å]{} line is not measurable). The R$_{23}$ diagnostic is based on the electron temperature sensitivity of the oxygen spectral lines, achieved via using the ratio of the oxygen line strength to a spectral feature independent of metallicity. R$_{23}$ requires measurement of the 3727 [Å]{} \[O II\], 5007 [Å]{} \[O III\], and H$\beta$ lines.\[metal\_lines\] However as the Oxygen line strength initially increases with and then (due to infrared fine-structure lines inducing a cooling effect) decreases with increasing metallicity this method alone is only sufficient to determine a pair of degenerate upper and lower branch metallicity values. Therefore the R$_{23}$ diagnostic is typically coupled with a second metallicity diagnostic to resolve this degeneracy. Typically, and specifically for all the objects in this paper, this is done using the \[N II\] / H$\alpha$ diagnostic requiring spectral coverage of the 6584 [Å]{} \[N II\] to H$\alpha$ line ratio (or its limit). While the \[N II\] / H$\alpha$ diagnostic is not nearly as precise as R$_{23}$ it is sufficient to exclude one of the R$_{23}$ branches except when the branches are near their intersection point where the metallicity values of the two branches converge anyway. Due to differences between metallicity diagnostics and their various calibrations, a comparison of metallicity values requires that they be determined using as consistent a method as possible. Therefore we recalculate the [@xshooter_survey] metallicities, from their published flux values, to give R$_{23}$ metallicity values in the KK04 scale, using the exact same code as used to determine the metallicities given in [@stats_paper]. The measured metallicity values used in this work are given in [first@refxsd\_data\_table,@]{}. Specifically to calculate metallicities we, as in [@stats_paper], employ an improved version of the idl code outlined in [@kd2002] which has been updated to the KK04 scale developed in [@KobulnickyKewley]. In order to correlate the flux of a line belonging to an individual atomic ionization level with the total abundance of that element it is necessary to know what fraction of the element is ionized to the level in question. This is achieved by measuring the flux ratio between the \[O II\] and \[O III\] line strength, giving the relative population in the O II and O III ionization states, and fitting the metallicity for that specific ionization state ratio. In the classical application of this diagnostic, an R$_{23}$ value would be calculated from the measured line ratios (using $R_{23} = {I_{3727} + I_{4959} + I_{5007} \over I_{H\beta}}$) and then compared along an \[O III\] to \[O II\] line ratio contour against the best calibration data available. This classical application however treats ionization as a parameter independent of metallicity and ignores the feedback the latter has on the former. [@kd2002] solve this issue by using iterative fitting to dynamically factor the effects of the metallicity on the ionization parameter.[^2] Distribution Analysis --------------------- One of the most powerful tools for understanding the behavior of GRBs (and SNe) is to study the distribution of such events with respect to a physical property. [@Fruchter] compared the location distribution of LGRB and ccSNe events on their host galaxies with the distribution of the blue light in their hosts. The result was that, while core collapse supernovae follow the blue light distribution of their host galaxies, LGRBs showed a strong preference for occurring in the brightest regions of their hosts. This suggested that, while ccSNe are unbiased tracers of the available star-formation, LGRBs likely favor very massive progenitors which are concentrated in the most star forming regions of their host galaxies. [@Stanek2007] compared the metallicity distribution of LGRBs and SDSS galaxies to show that nearby LGRB hosts have low metallicities when compared to similar magnitude Sloan galaxies. These methods employ normalized cumulative evenly spaced step-functions to track the distributions of the objects. This approach works well for objects such as LGRBs and SNe whose detection is proportionate to their event rate. Galaxy samples however have massive Malmquist biases and the individual galaxies are themselves dissimilar. To allow for direct meaningful comparison between the metallicity distribution of LGRBs and galaxies [@stats_paper] first converted the SDSS to a volume limited sample with a tight redshift cut and then weighted the galaxies by their star-formation rate. The result was a weighted step-function for the galaxy population with a varying step height. The LGRBs showed a strong preference for low metallicities while the SNe populations tracked the star-formation weighted volume limited galaxy sample. Here we are tasked with the need to compare a variety of different LGRB populations (composted of LGRB samples with different measured physical properties) to exclude redshift dependent selection biases from contaminating our further analysis. Normalized cumulative distribution plots would ordinarily be ideal for such a task except that the differing physical properties have different redshift ranges where they can be observed and different rates of observation (e.g. metallicity measurement requires spectroscopy of certain lines thus having redshift limitations and spectroscopy is an expensive use of telescope time). To examine the extent to which the samples we consider in this paper suffer from different selection effects we therefore employ a cumulative distribution plot with a manual weighting for each population so as to scale the populations to an arbitrary level at a designated redshift. In this work, we will adopt [z]{} of 2.5 for such a designated redshift as it is the approximate end of the redshift range for which we have metallicity measurements. For populations that do not reach the designated redshift we scale the weighting to achieve the best congruence with the other populations. We also adjust the weightings of all the populations as needed to favor a common congruence. This is a manual process, as the goal is to produce a distribution plot where deviation of congruence is reflective of redshift dependent selection biases. While a program to estimate a scaling values that minimizes the area between two curves could be used instead, the advantage of doing the scaling by hand it that when the samples agree to a point and then diverge this can be shown in a physically meaningful way. This human touch allows for choosing scaling values that work sensibly, and where a best fit area minimizing value is not ideal for comparison, the author can adjust accordingly. The the graphical manually scaled approach allows us to know where biases are present in our data, and after achieving an understanding of their causes we can estimate their effects on our analysis. One critical limitation of this process is it requires having redshift data and is therefore insensitive to biases in measuring the redshifts of our targets. To attempt to correct for this we also include a similar scaling of a theoretical LGRB distribution using the results of [@form_rate_letter]. Since we cannot exclude a redshift bias in the [@form_rate_letter] results (i.e. inducing an error in the metallicity evolution history of the universe) this is not a perfect solution (see Footnote ). We then also use normalized cumulative distribution plots with a uniform step height to track the metallicity distribution of LGRB hosts within various redshift bins. This ability to break the sample into different redshift bins is enabled by the enlarged LGRB host metallicity galaxy population, however this analytical method has been used previously. Analysis of Sample Redshift Distributions {#zd} ========================================= Before we can explore the metallicity distribution of LGRBs as a function of redshift we need to account for completeness of our sample population. Given the nature of this sample as a completion of observations from many different observing programs, each with different sample criteria, luminosity biases, available per object observing time, and science goals, an a priori assessment is not practical. We can however annualize the effect of redshift on the completeness of the differing LGRB populations. To achieve this we cumulatively plot the redshifts of the different LGRB populations. In a typical cumulative distributions plot we would normalize the populations so as to allow for comparison of populations with different numbers of objects. Here however we are faced with some samples with a redshift range that ends far before others (and the end of the range plotted). We also have other samples with additional selection effects apparent only at high redshifts. Therefore, instead of normalizing the distributions to unity we scale each population by a unique arbitrary factor so as to produce distributions consistent with each other. As we will later focus on analyzing the LGRB metallicity distribution we thus chose to scale our populations to the [@xshooter_survey] metallicity distribution (thus using the larger of our metallicity samples as a baseline). The degree of scaling is not of particular interest as it merely indicates the relative number of objects in the populations (out to their redshift limit) which tracks general trends in the degree of observational followup (i.e. $\sim$$\frac{1}{5}$ of LGRBs with spectroscopic redshifts have measured metallicities). Thus were a population of LGRBs to exist with exactly the same distribution as [@xshooter_survey] but with half the number of objects it would be given a scaling factor of 2 (as @xshooter_survey has an a priori scaling factor of 1). What is of interest in this analysis are any differences in the distributional shape between the different populations. Differences in slope indicate redshift dependent biases between the populations, sharp changes within a population’s slope reveal discontinuities in that population’s rate of observation with redshift, and gradual changes show the differences in that population’s rate of observation due to the expanding number of objects out to that redshift balanced by their increasing difficulty of observation. Thus by comparing populations of galaxies with different observed properties we can thus identify redshift biases in how samples of those properties are collected and measured. To investigate the potential of redshift dependent sample biases on our metallicity distributions we therefore compare the following populations for the described reasons: - We begin by plotting the redshift distributions of the LGRB host galaxy populations for which we have host metallicity measurements. Since [@xshooter_survey] has approximately three times as many objects, we use this as the numeric baseline and scale the sample of [@stats_paper] by a factor of 0.7 to match. - To address concerns about a metallicity dependent redshift bias we also plot the subset of [@xshooter_survey] hosts with metallicities above solar. As this is only about a third of the total [@xshooter_survey] host population, it is scaled by a factor of three. We also plot the subset of [@xshooter_survey] hosts with metallicities above $ {\rm log(O/H)}+12 > 8.4 $ (KK04 scale) so as to sample only objects clearly above the $ {\rm log(O/H)}+12 \approx 8.3 $ metallicity cutoff noted in [@diff_rate_letter]. Similar high metallicity subsets of the [@stats_paper] sample is omitted due to its small size of only two and four objects (after the removal of LGRB 020819B). - To address concerns about a bias in which objects have the lines needed for metallicity measurement we plot the entire [@xshooter_survey] host population including the objects without a metallicity. Since metallicity measurement requires a specific set of lines whereas redshift measurement requires only a single identified line (in practice any two lines is used to identify both the lines and the redshift) this provides an immediate crosscheck on observational biases in line measurement. As including [@xshooter_survey] objects with redshift but not metallicity makes an approximate 50% increase in the sample size (at z $<$ 2.5 before the slope noticeably changes) we scale it down by a factor of $\frac{2}{3}$. As the slope of this, and some other distributions, noticeably changes at z $\approx$ 2.5, and since z $\approx$ 2.5 is the limit of the redshift range for which we have measured metallicities, we adopt this redshift, along with the matching ordinate of this distribution (a value of $\sim$60), as the reference point for scaling the remaining distributions. - Next we compare it with the general burst population by plotting all known LGRB host galaxy redshifts as conveniently complied in the GRB table of Jochen Greiner[^3] (excluding for this and all other samples those objects with no or only photometric redshifts), as of UTC noon March 15^th^ 2019. As this sample is undoubtably subject to a variety of complex selection effects, its use provides a reasonable assessment of the degree that such unmitigated effects have on our analysis. Again and subsequently the sample is scaled to match the other distributions. - To address biases in host redshift measurement as a function of redshift itself, we plot the [@Perley_shoals_masses] SHOALS and [@BAT6] BAT6 samples, which claim to be unbiased in their selections. We also plot the Optically Unbiased Gamma-Ray Burst Host Survey (TOUGHS) sample updated using all currently known spectroscopically determined redshift values for objects from that sample. TOUGHS is an attempt to provide a host sample which is not biased by host absorption of the optical afterglow of the LGRB. Of the 69 [Swift]{} GRB hosts we find 58 of them have exact redshift values. We note that one of the missing redshifts is due to a bright foreground star and can thus be excluded from the sample without introducing any biases, giving the TOUGHS redshift sample an 85% effective completeness. - To also sample bias in host galaxy mass measurement we plot the subset of [@Perley_shoals_masses] SHOALS galaxies with mass values (excluding limits) as well as the sample of [@Svensson]. We also consider a theoretical curve for the LGRB event rate. [@form_rate_letter] estimated the LGRB progenitor rate using the [@diff_rate_letter] results in concert with SNe statistics via an approach patterned loosely off the Drake equation. Beginning with the cosmic star-formation history, [@form_rate_letter] took the expected number of broad-line Type Ic events that are in low metallicity host environments as potential LGRB progenitor candidates. They then adjusted this number by adding the contribution of broad-line Type Ic SNe in high metallicity host environments at a much reduced weighting of $\sim$$\frac{1}{30}$ (the @diff_rate_letter estimate on the relative suppression of LGRB formation in high metallicity environments). A comparison of this estimate of potential LGRB progenitor candidates to the observed LGRB rate corrected for instrumental selection effects (estimated by @Lien and @Graff), provided a combined estimate of the fraction of broad-line Type Ic SNe residing in metal poor environments that produce an LGRB and the fraction of such LGRBs that are beamed in our direction. Thus [@form_rate_letter] estimated that, in a low metallically environment, an aligned LGRB occurs out of approximately every 4000 $\pm$ 2000 broad-lined Type Ic Supernovae. Therefore if one assumes a semi-nominal beaming factor of 100 then 1 out of $\sim$40 low metallicity Ic-bl SNe give rise to an LGRB. Using the process described above, we derive an expected LGRB event rate as follows: Beginning with the cosmic star-formation rate history of [@hb] and the [@form_rate_letter] estimates of the metallicity distribution of the universe as a function of redshift we obtain an estimate of the amount of star-formation above and below the [@diff_rate_letter] metallicity cutoff. The expected number of LGRBs formed is thus estimated after applying the [@diff_rate_letter] estimate of the effect of metallicity on the LGRB formulation rate per unit of available star-formation. The result is an expected LGRB event rate as a function of redshift which can then be scaled in the same manner as the observed LGRB populations described previously. ![\[redshift\_distribution\] The redshift distribution of various LGRB samples scaled as labeled such that the distributions are approximately 60 at redshift 2.5 which provides reasonable agreement across the lower redshift range. The purpose of this plot is to compare the redshift distributions of different LGRB samples and find any discrepancies beyond that different samples have a different total number of objects. Note that the @stats_paper (GF13) sample (thick black line in the inset plot) has a lower rate of observations beyond [z]{} of $\sim$0.5 because an IR spectrum is required with a different instrument in addition to the optical spectrum. This is not seen in the [@xshooter_survey] sample because X-shooter obtains optical and IR spectra in unison. However a gap is seen in the X-shooter metallicity sample (both the total sample and its high metallicity sub-sample) from 1.7 $<$ [z]{} $<$ 2.1 due to observational difficulties in obtaining all the lines needed for metallicity measurement (see text). As measuring the redshift does not require observing specific lines this gap is not seen in the [@xshooter_survey] redshift sample. We also compare these LGRB samples with our predicted LGRB rate estimate from @form_rate_letter (thick grey line) and find it to be in good agreement with the samples at [z]{} $<$ 2.5 and with the [@Perley_shoals_masses] SHOALS sample in particular at higher redshifts. KS test results for the lines in this figure are shown in [first@refLGRB\_num\_stat\_KS\_table,@]{} in the appendix. Note that the super solar metallicity LGRB sample follows the same distribution as the general LGRB sample, which suggests both that the metallicity of the LGRB hosts does not bias the likelihood of their metallicity being measured and that evolution LGRB host metallicity has a negligible effect up to z $\sim$ 2 which is consistent with the results of [first@refmzd,@]{}.](LGRB_num_stat_single.pdf){width=".49\textwidth"} Rather surprisingly, as shown in [first@refredshift\_distribution,@]{}, we find that there does not appear to be any particularly significant differences in the redshift distributions of the different LGRB samples aside from the following: 1. [@stats_paper] sample shows a considerably reduced rate of LGRB observation at redshifts high enough ([z]{} $\gtrsim$ 0.5) to require a separate IR spectrograph to observe some of the lines required for determining a metallicity. As X-Shooter contains an integrated IR spectrograph, the [@xshooter_survey] sample is not subject to this effect. 2. The [@xshooter_survey] metallicity sample has a gap from 1.7 $\lesssim$ [z]{} $\lesssim$ 2.1 which is due to a confluence of observational limitations. The \[O II\] 3727 [Å]{} is redshifted to the optical IR transition at about 1 micron, the 4861 [Å]{} H$\beta$ and 4959 & 5007 [Å]{} \[O III\] lines are redshifted to a region of poor sky transmittance between J and H bands, and 6563 [Å]{} H$\alpha$ is redshifted into a similar low transmittance region between the H and K bands. We also find a smaller gap from 1.0 $\lesssim$ [z]{} $\lesssim$ 1.2 due to similar effects. Both the full metallicity sample and high metallicity sub-sample from [@xshooter_survey] are affected by these redshift gaps since both of these samples share the same redshift distributions as expected since the cause of this gap is not dependent on the metallicity of the objects. However this gap in the [@xshooter_survey] metallicity sample is not seen in the [@xshooter_survey] redshift sample as any distinct combination of emission lines is sufficient to identify the redshift. After correcting for this gap (i.e. artificially scaling the post gap metallicity sample lines back up to the other distributions) the metallicity populations continue to track the other distributions for the (limited) remainder of their redshift ranges. Strangely we also find similar gaps in the TOUGHS redshift sample from 1.06 $\lesssim$ [z]{} $\lesssim$ 1.26 and 1.56 $\lesssim$ [z]{} $\lesssim$ 1.88 with an approximate respective 2 and 4 sigma significance above the mean z $<$ 2.5 redshift spacing. Given the construction of the TOUGHS sample these gaps are unexpected and puzzling. It is conceivable that this gaps are a result of the same difficulties in observing certain lines as in the metallicity samples, however as any two lines are sufficient for determining a redshift it remains odd that this would be the explanation here. 3. The samples diverge at [z]{} $\gtrsim$ 2.5 due to different completeness rates for high redshift objects. As this is beyond the redshift range of our host metallicity sample, it does not impact our subsequently metallicity distribution analysis. It is worth noting that the [@Perley_shoals_masses] SHOALS sample does not show a change of slope in these regions unlike the other samples and instead agrees quite closely with our predicted LGRB event rate curve from the [@form_rate_letter] formulation. Since the BAT6 sample is flux limited, as they go higher in redshift, their sensitively to the lower end of the GRB luminosity function decreases. This results in an underrepresentation of high redshift objects, which is consistent with the underrepresentation in the sample of all LGRBs with spectroscopically measured redshifts (private communications). Given that the BAT6 sample tracks the Greiner table sample of all known LGRB redshifts after the [z]{} $\sim$ 2.5 diverge point with the other samples, this suggests that the “variety of complex selection effects” in the full LGRB redshift sample is dominated by the same flux limited selection effect as the BAT6 sample. The difference in slope between the [@xshooter_survey] redshift sample and the BAT6 and Greiner table samples reflects the increasing fractional usage of absorption line redshift measurements at high redshifts while [@xshooter_survey] is strictly a sample of objects with emission lines. While the enumerated effects noted above are of some interest they appear relatively minor and should not introduce any significant biases in the overall focus of this paper: the comparison of how the LGRB host metallicity distribution varies as a function of redshift. However, if the LGRB host metallicity distribution where to extend past [z]{} $\sim$ 2.5 then biases introduced via effect 3 would require special attention. KS Analysis of Redshift Distributions ------------------------------------- To analyze the differences in the redshift distributions of our samples more rigorously we calculate Kolmogorov–Smirnov (KS) probabilities comparing all of our samples with each other in [first@refLGRB\_num\_stat\_KS\_table,@]{}. To address the issues noted above we introduce some additional subsets of our samples as follows. (1) We subdivide the [@stats_paper] sample into objects with z $<$ 0.5 and z $>$ 0.5 so as to assess the fit of the sample with and without the application of a separate IR spectrograph. (2) We subdivide the [@xshooter_survey] metallicity sample and the $ {\rm log(O/H)}+12 > 8.4 $ subset into objects with z $<$ 2 and z $>$ 2 so as to allow us to assess the fit of the sample without the gap at z $\approx$ 2. For the same reason, we also add a z $<$ 2 [@xshooter_survey] super solar metallicity sample but can not add a matching super solar metallicity z $>$ 2 sample as we have fewer than 4 such objects. (3) For all samples that extend to z $>$ 2.5 we add a subset of values with z $<$ 2.5 to compare only those regions before the z $>$ 2.5 divergence. To compare with the our [@form_rate_letter] based theoretical curve for the LGRB event rate, we create a simulated population of about 3000 objects using an appropriately weighted random generation approach. Specifically we create a redshift array corresponding to million year increments since [z]{} = 4. We then calculate an LGRB rate estimate for each redshift normalized such that the maximum value at any redshift is 1. Then we exclude bins where the rate estimate is lower than a random number, unique for each redshift, uniformly distributed in the 0 to 1 range. The result is a set of redshifts with the same distribution as expected for the theoretical LGRB event rate. This simulated population is then treated like any other population for the KS analysis. Calculating KS values for comparison has the notable advantage that the values do not need to be scaled but can be compared directly (therefore the scaling factors from [first@refredshift\_distribution,@]{} are not used in the KS analysis). Critical to our KS analysis is that samples must be matched across a common redshift range. To achieve this, for any two samples being compared, we take the sample with the smallest redshift range, set a redshift cutoff at 0.05 [z]{} higher than the highest redshift in that sample, and remove all objects (in the other sample) with redshifts above this cutoff. (This additional 0.05 [z]{} redshift grace allows us to utilize both samples in full if their maximum redshifts are similar, otherwise one sample would always have its highest redshift object removed.) Samples are compared without a lower redshift cut, unless a lower redshift range is specified for the sample, in which case the same process is used but in reverse (again with a 0.05 [z]{} grace). Since some samples are redshift subsets of others, and the redshift ranges are matched, some comparisons will be of a set of objects with itself, those are thus marked as having a KS probability of “1" (values of “1.0000" are the result of rounding). Some comparisons have fewer than the 4 objects in each sample needed for the KS test and are marked as NaN. The KS values calculated for the populations shown in [first@refredshift\_distribution,@]{} are given in [first@refLGRB\_num\_stat\_KS\_table,@]{} in the appendix. This consistency of redshift distributions between different LGRB samples, aside from the perviously enumerated effects, suggests that there are no unknown significant deviations in the redshift completeness of our sample (at least that correlate with redshift) and that we have a good understanding of the causes for the deviations that are present. The consistency with our theoretical curve further suggests that the observed populations are unbiased and that there are unlikely to be effects caused by systematic biasing in population of all LGRB with spectroscopically measured redshifts.\[rate\_feedback\_bias\][^4] Metallicity Distribution as a Function of Redshift {#mzd} ================================================== ![image](Measured_2_pane.pdf){width="100.00000%"} We begin, in [first@refmeasured,@]{} left, with a standard normalized cumulative metallicity distribution plot of LGRBs with measured host metallicities divided across five redshift bins with a width of 0.5 [z]{}. This gives a series of 5 distributions for the [@xshooter_survey] and 2 for the [@stats_paper] samples (since all objects in the latter sample are [z]{} $<$ 1). A distribution for both the @stats_paper & @xshooter_survey samples across all redshifts is also provided for reference. The [@xshooter_survey] [z]{} $<$ 0.5 and both the [@stats_paper] 0 $<$ [z]{} $<$ 0.5 and 0.5 $<$ [z]{} $<$ 1 bins seem to deviate slightly towards a more metal poor distribution than the remaining [z]{} $>$ 0.5 [@xshooter_survey] populations, however the statistical significance of this is tenuous. The [z]{} $>$ 2.0 [@xshooter_survey] population also appears to be a little more metal rich than the 0.5 $<$ [z]{} $<$ 2.0 [@xshooter_survey] populations but this is not statistically significant and is counter intuitive, as due to both the typically smaller masses of star-forming galaxies and the less time available for metal enrichment, one would expect the metallicity of any typical galaxy sample to decrease with redshift. In [first@refmeasured,@]{} right we combine the [@stats_paper] and [@xshooter_survey] observations into a single combined sample, while removing duplicates (as described in [first@refHost\_Metallicities,@]{}), which we use subsequently. The overall result is a markedly constant metallicity distribution across the different redshift bins, indicative of far less metallicity evolution than is present in the typical star-forming galaxy population across the same redshift range [@Zahid2013]. Since LGRBs are formed in star-forming galaxies this discrepancy is perplexing. We note that while we addressed redshift biases in [first@refzd,@]{} it is not possible to account for all the biases that could potentially be affecting the selection of which objects were observed with spectroscopy. In particular luminosity biases, i.e. selecting only galaxies bright enough for spectroscopy in a reasonable amount of telescope time, would presumably be prevalent. However the ratio of LGRB host galaxies observed at low, intermediate, and super solar metallicities seems to remain relatively consistent, particularly at [z]{} $<$ 2, (see [first@refmeasured,@]{}, [first@refhist\_table,@]{}, and the blue, purple and dark green lines in [first@refredshift\_distribution,@]{}) whereas the general expectation for luminosity biases would be to observe an increasingly large fraction of higher metallicity galaxies with redshift. A more contrived scenario, where the lower metallicity LGRB hosts are being selected out via a luminosity bias and the higher metallicity LGRB hosts become intrinsically less common with redshift yielding both populations to drop off at the same rate and thus maintain a constant ratio, remains possible but seems unlikely. Ideally we would next compare the metallicity distribution of LGRB hosts with a sample of typical star-forming galaxies at the same redshifts selected in a star-formation weighted manner (i.e. the sample methodology of the [@stats_paper] SDSS population but at higher redshifts). Unfortunately, a suitable galaxy sample does not exist. The SDSS metallicity sample only extends (with a large sample size) out to [z]{} $<$ 0.3 and even out to that redshift range there are completeness issues with faint galaxies. While using the [z]{} = $\sim$0 SDSS population and subtracting the expected metallicity shift from the mass metallicity relation would provide a crude metallicity distribution, this would not provide a suitable enough estimate for the star-formation weighting. Therefore our ability to quantify the expected metallicity distribution evolution for the LGRB hosts from the typical star-forming galaxy population is lacking. Still, the absence of any apparent evolution can be reasonably excluded from being due to mass metallicity relation evolution across the redshift range in question. Although we cannot create a suitable comparison sample of star-forming galaxies (i.e. galaxies of the same redshift distribution selected in a star-formation weighted manner), we can compare our sample of LGRB host metallicities with a sample of typical metallicities for galaxies with the masses and redshifts of an actual LGRB host galaxy population. In [first@refzd,@]{} we already assessed the redshift distribution of our LGRB host galaxy mass samples and found them to be consistent with that of our host metallicity samples. \[1\][>m[\#1]{}]{} [L[0.3cm]{}m[6.7cm]{}L[2.2cm]{}L[2.2cm]{}L[1.8cm]{}L[1.8cm]{}L[1.8cm]{}L[1.8cm]{}]{} line & & 0.0 $<$ z $<$ 1.0 & 1.2 $<$ z $<$ 1.7[^5] & 2.1 $<$ z $<$ 2.5\ 1 & Number of All LGRBs with Redshifts[^6] & 155 & 78 & 55\ 2& LGRBs with Measured Metallicites & 32 & 14 & 8\ 3& LGRBs with Metallicites Z $>$ 8.4 & 17 & 9 & 7\ 4& Measured Metallicites / All LGRBs & 0.206 $\pm$ 0.040 & 0.179 $\pm$ 0.052 & 0.145 $\pm$ 0.055\ 5& Metallicites Z $>$ 8.4 / All LGRBs & 0.110 $\pm$ 0.039 & 0.115 $\pm$ 0.041 & 0.127 $\pm$ 0.051\ 6 & Dark Burst Adj. Z $>$ 8.4 / All LGRBs & 0.087 $\pm$ 0.031 & 0.081 $\pm$ 0.028 & 0.078 $\pm$ 0.031\ The Fraction of High Metallicity LGRBs {#Z_frac} -------------------------------------- While the agreement of the cumulative distributions reported in the previous section is impressive, there are likely selection effects for which we can not fully exclude, for example the cutoff in absolute luminosity below which we cannot practically obtain a host galaxy spectrum differer for each of the redshift bins To investigate this, we consider the cutoff in the LGRB formation rate per unit star-formation above a metallicity of $ {\rm log(O/H)}+12 \approx 8.3 $. Here we look at the fraction of hosts that lie above this cutoff as a function of redshift. In the first three lines of [first@refcalc\_table,@]{} we present the number of all LGRBs with redshifts (as complied in the GRB table of Jochen Greiner), the number of LGRBs with measured metallicities in our combined sample of bursts from [@stats_paper] and [@xshooter_survey] sample and the number of LGRBs with metallicities of $ {\rm log(O/H)}+12 > 8.4 $ in this sample for three redshift bins: $0<z<1$, $1.2<z<1.7$ and $2.1<z<2.5$. As discussed in [first@refzd,@]{}, the redshift range of the middle bin (and starting the last bin at [z]{} of 2.1) are forced on us by observational effects that greatly reduce the efficiency of the metallicity observations immediately above a redshift of one and around a redshift of two. In Line 4 of [first@refcalc\_table,@]{}, we give the fraction of LGRBs with redshifts which also have metallicity measurements in our sample. As one would expect, the fraction declines continuously with increasing readshift, from about one-fifth in the first redshift bin to about one-tenth in the last. However, in Line 5 we give the fraction of all GRBs which have a measured metallicity in our sample of $ {\rm log(O/H)}+12 > 8.4 $. Strikingly, this fraction [does not decline]{} with redshift. However, [@xshooter_survey] realized that their sample, which was a combination of several small surveys, had an overabundance of “dusty” GRBs, that is burst with $A^{GRB}_{V} > 1$. Comparing their sample with more representative surveys, they determined this overabunce as a function of redshift (typically a factor of $\sim$30%) anc corrected their sample for this effect. In Line 6 we have recalculated the ratios in Line 5 after reducing the number of bursts in our combined sample from [@xshooter_survey] with $ {\rm log(O/H)}+12 > 8.4 $ by the appropriate overabundance factor. (We have not adjusted the part of our sample from [@diff_rate_letter], because, as discussed by those authors, that sample had a typical number of dusty bursts). Although the ratios all go down somewhat, the primary result remains, the fraction of bursts with metallicities $ {\rm log(O/H)}+12 > 8.4 $ remains essentially constant independent of redshift. Expected Metallicity Distribution From Mass & Redshift {#emzd} ------------------------------------------------------ In order to further explore potential selection effects, we use the mass-metallicity relation to estimate the metallicity distribution from the comparatively less biased sample of host galaxy masses. This metallicity distribution is only an estimate since it reflects only the typical galaxy metallicity expected for a galaxy of a given LGRB host mass at its specific redshift. However, the mass metallicity relation itself varies with redshift. Therefore, we estimate the mass metallicity relation for the redshift of the galaxy in question by interpolating from the mass metallicity relations measured at different redshifts. While [@Mannucci] claim a fundamental mass, metallicity, & star-formation rate relation that remains independent of redshift, we found (for the local universe) this to be mostly a mass metallicity relation, with a small but statistically significant star-formation rate perturbation, that is not independent of redshift. The applicably of the [@Mannucci] relation beyond the local universe is dubious at best with [@Wuyts] and [@Sanders2014] finding no SFR dependance on the mass-metallicity relation, and [@Yates2012; @Yates2014] finding a slight positive correlation for high mass objects, all contrary to the [@Mannucci] claim of a strong anti-correlation. As this clearly invalidates the fundamental relation being invariant with redshift, we therefore do not use the [@Mannucci] relation here. While [@Zahid2013] provide a series of mass metallicity relation fits across the redshift range of interest, we apply our own 2-dimensional fit to the [@Zahid2013] Figure 1 data so such that we can estimate the metallicity for a galaxy of any given mass and redshift. Care was taken to avoid over-fitting the [@Zahid2013] data and a number of fitting procedures were trialed with fitting a minimum curvature spline surface using the MIN\_CURVE\_SURF procedure adopted. Our estimated metallicity results for a continuous range of galaxy masses are plotted against the individual redshift fits of [@Zahid2013] in [first@refZahid\_polyfit,@]{}. ![\[Zahid\_polyfit\] The mass metallicity relation fits for different redshifts of [@Zahid2013] Figure 1 (thick dashed lines) with our estimated metallicity results for a continuous range of galaxy masses and matching redshift bins (thin continuous lines) overplotted. The fitting is deliberately kept approximate to avoid over-fitting.](Zahid2013_fitting.pdf){width=".49\textwidth"} We therefore apply this methodology to estimate the expected metallicities for the [@Perley_shoals_masses] SHOALS sample. We show the estimated metallicity distribution of this sample in [first@refmeasured\_and\_estimated,@]{} right, using the same redshift binning as in [first@refmeasured,@]{}. For comparison we show a simplified version of our measured LGRB host metallicities on the left side of this figure ([first@refmeasured\_and\_estimated,@]{}). The upper half of the expected metallicity distributions shows a lack of metallicity evolution with redshift consistent with that seen in the measured metallicity distributions. Since for the same given mass, the expected metallicity is lower at higher redshift, to therefore maintain a consistent metallicity distribution the LGRB mass distribution would have to increase with redshift. [@Perley_shoals_masses] finds that this is indeed the case. To illustrate this we plot the combined [@Perley_shoals_masses] and [@Svensson] mass distributions with the same redshift binning and colors in an inset of [first@refmeasured\_and\_estimated,@]{}. While an increase in observed host galaxy masses with redshift could potentially be due to Malmquist bias it seems unlikely that such a bias would exactly match that needed to correct for mass metallicity evolution so as to maintain a consistent metallicity distribution. Moreover, sample of LGRBs in [@Perley_shoals_masses] is unbiased in its selection and therefore their observed mass distribution increase is not caused by observational biases in their sample (we show this explicitly in [first@refshoals\_mass\_step\_plot\_5,@]{}). ![image](Measured_and_estimated.pdf){width="100.00000%"} Comparing between the right and left sides of [first@refmeasured\_and\_estimated,@]{}, the estimated metallicity distributions have higher metallicities for a given redshift than the measured metallicity distributions. We overplot these metallicity distributions (simplified to show only the full 0 $<$ [z]{} $<$ 2.5 redshift range) in [first@refmeasured\_vs\_estimated,@]{}. This distributional offset of, on average, approximately a quarter dex is roughly constant across the distributions interquartile range. Therefore, the LGRB host galaxy population is systematically lower in metallicity than typical galaxies of comparable mass and redshift. This suggests that the LGRBs are biased towards the lowest metallicities within any galaxy population and can not be correctly modeled using the general mass metallicity relation. In [first@refmetal\_comp\_table,@]{} we explicitly compare the 29 objects for which we have both measured and estimated metallicity values. We find a mean difference between the estimated and measured metallicities of 0.23 dex with a median difference of 0.30 dex but with a standard deviation on the difference of 0.46 dex. In raw (i.e. non-log) metallicities we find a median ratio of 0.496 between the measured and estimated metallicities. Hence we conclude that while the estimated metallicities are typically higher than the measured values the offset is too random to be useful for correcting the estimated metallicities to match the measured values. ![\[measured\_vs\_estimated\] Measured (black) and estimated (grey) metallicities from [first@refmeasured\_and\_estimated,@]{} left and right respectively (i.e. the black lines in each side showing the metallicities across the entire [z]{} $<$ 2.5 redshift range). A green box with a width of 0.25 dex is overplotted; this corresponds to a difference factor of 1.8 in raw metal content. This resolves much of the difference between the LGRB formation metallicity cutoff of about a third solar in [@diff_rate_letter] with the cutoff value of approximately solar claimed in [@Perley_shoals_masses] in favor of the former. As LGRB hosts do not follow the general mass metallicity relation, there is no substitute for actually measuring their metallicities!](Measured_vs_estimated.pdf){width=".49\textwidth"} Selection Effects in the Mass & Expected Metallicity Distributions ------------------------------------------------------------------ ![\[shoals\_mass\_step\_plot\_5\] Cumulative distribution of [@Perley_shoals_masses] SHOALS measured LGRB host masses binned by redshift. Mass upper limits are shown as dashed lines, mass measurements are shown as solid lines. We also plot the mass measurements without the limits, while tracking the number of limits in the normalization, causing the distributions do not extend down to zero. These two approaches define the maximum and minimum range possible for the distributions. A shaded region between them indicates where, due to the upper limits, the true distribution must lie. Where the shaded region is bounded by lines (dotted or solid) the true distribution is constrained between those bounds. If the color fades out to the left, the distribution could, in principle, extend to an arbitrarily low mass.](shoals_w_limits_mass_step_plot_5_v2.pdf){width=".49\textwidth"} Key to understanding the validity of [first@refmeasured\_and\_estimated,@]{} is the completeness of the samples. While this is difficult for metallicity measurements (i.e. [first@refmeasured\_and\_estimated,@]{} left) due to the sample being a completion of many different observing programs, it is possible for the estimated metallicities if we use a mass sample with suitably unbiased selection criteria such as those used in the [@Perley_shoals_masses] SHOALS sample. However the [@Perley_shoals_masses] SHOALS mass sample is not perfectly complete. Seven out of 119 objects are missing clear specific redshift values and some objects with redshifts have only upper limits on their host galaxy mass values. The later however we can address by constraining the maximum and minimum possible distributions. Ideally the upper limit values would all be below the data values, however (while most of the SHOALS upper limit values are below most of the SHOALS mass values) this is not true in all cases. The maximum possible distribution is set by plotting the data with measured mass values and mass upper limits together. We indicate which values are limits by using a dashed line. The minimum possible distribution is set by assuming all the upper limit values are below the measured values. As there are no lower bounds on the upper limits, these are not plotted however we do count the number of such objects and normalize the distributions accordingly. This results in distributions that do not extend all the way to zero but instead stop where the limits are unconstrained. Where both sides are constrained, we shade the region between the maximum and minimum possible distributions, as the true distribution must lie between these bounds. When the minimum distribution is not constrained we use a gradient shading where the color fades out to the left as the distribution could, in principle, rise at an arbitrarily low mass. The resulting plot, [first@refshoals\_mass\_step\_plot\_5,@]{}, is similar to [@Perley_shoals_masses] Figure 4 except for the more detailed plotting system described above. We also use the same color scheme as in our other figures and we explicitly use log mass on the abscissa. Of particular note in [first@refshoals\_mass\_step\_plot\_5,@]{} is that the host galaxy mass values continue to increase with redshift. This is in agreement with the behavior of our combined host mass sample plotted in [first@refmeasured\_and\_estimated,@]{} inset and shows that the trend of increasing host galaxy masses with redshift is not due to selection and measurement biases. For the estimated metallicities, upper limit mass values create upper limit metallicity estimate values. Similarly, for the estimated metallicity values, the upper limits are generally below the measured values. In [first@refshoals\_metal\_est\_step\_plot\_5,@]{} we plot the estimated metallicity cumulative distribution redshift bins of the [@Perley_shoals_masses] SHOALS sample in the same manner as [first@refshoals\_mass\_step\_plot\_5,@]{}. We discard the z $<$ 0.5 bin from SHOALS due to it having too few objects to be useful and replace it with z $<$ 0.5 bin from [first@refmeasured\_and\_estimated,@]{} right (as the sample biases within the lowest redshift bin are likely to be minimal). ![\[shoals\_metal\_est\_step\_plot\_5\] Cumulative distribution of estimated metallicities from the [@Perley_shoals_masses] SHOALS mass and redshift values binned by redshift. Estimated metallicity upper limits are shown as dashed lines, estimated metallicities are shown as solid lines. As in [first@refshoals\_mass\_step\_plot\_5,@]{}, a shaded region indicates where, due to the upper limits, the true distribution must lie. If the color fades out to the left, the distribution could, in principle, extend to an arbitrarily low metallicity. Where the shaded region is bounded by lines (dotted or solid) the true distribution goes between the bounds. Due to the limited number of objects in the SHOALS z $<$ 0.5 bin we omit plotting this bin and instead replot the combined z $<$ 0.5 bin from [first@refmeasured\_and\_estimated,@]{} inset (purple dotted line). For reference and comparison with our larger estimated metallicity sample from [first@refmeasured\_and\_estimated,@]{} right we also replot the combined z $<$ 2.5 distribution from there (dotted black line). KS test results for the upper half of the lines in this figure are shown in [first@refshoals\_w\_limits\_metal\_est\_step\_plot\_5\_KS\_table\_up\_half,@]{} in the appendix. Fortunately all the upper limits are contained in the lower half of the distributions and thus do not disrupt this analysis.](shoals_w_limits_metal_est_step_plot_5.pdf){width=".49\textwidth"} In [first@refshoals\_metal\_est\_step\_plot\_5,@]{}, while not as apparent as in [first@refmeasured\_and\_estimated,@]{}, we still see evidence of a constant metallicity distribution, at least across upper half of the distributions where the redshift bins are relatively complete. (Compare the KS results given in [first@refshoals\_w\_limits\_metal\_est\_step\_plot\_5\_KS\_table\_up\_half,@]{} for the upper half of [first@refshoals\_metal\_est\_step\_plot\_5,@]{} with those given in [first@refmetal\_est\_comb\_step\_plot\_5\_KS\_table\_up\_half,@]{} for the upper half of [first@refmeasured\_and\_estimated,@]{} right.) Results Summary {#Results_Summary} =============== We find an absence of large differences in the LGRB redshift distribution out to [z]{} of 2.5 between different samples studying different physical properties of bursts. The minor exceptions are that the [@stats_paper] LGRB metallicity sample shows an expected bias against LGRB metallicities where a separate IR instrument is required to observe the reddest required lines. This bias is not repeated in the [@xshooter_survey] X-Shooter survey since X-Shooter can observe all required spectral lines for determining metallicity out to [z]{} = 2.5, however a gap in the X-Shooter metallicity redshift coverage is seen between approximately 1.7 $<$ [z]{} $<$ 2.1 due to some of the required lines falling out of observable spectral regions. We also find that the observed LGRB redshift distribution agrees well with the distribution predicted in [@form_rate_letter] and that the [@Perley_shoals_masses] SHOALS redshift distribution in particular matches predictions at redshifts above [z]{} of 2.5 where the redshift distributions of the different LGRB samples (which extend to these high redshifts) diverge. Of particular note, while we cannot account for all the biases that went into the selection of LGRB hosts for spectroscopy, we find that the ratio of low, intermediate, and super solar LGRB hosts does not change dramatically across our sample. In particular host luminosity biases would expectedly over observe higher metallicity, presumably brighter, objects at higher redshifts. Given the unconstrained nature of the LGRB host metallicity sample it is likely subject to innumerable complex and subtle selection biases that are not practical to quantify, however we do not see evidence that selection biases are driving our results. We then proceed to look for evolution in the metallicity distribution as a function of redshift and see surprisingly little. There appears to be curious consistency in the metallicity distribution across different redshifts. This is at odds with the general evolution in the mass metallicity relation, which becomes progressively more metal poor with increasing redshift. To explore this further with a sample more a priori controlled in selection effects we turn to simulated metallicities. As the average mass, redshift, and metallicity are related (by an evolving mass metallicity relation) we can thus use any two properties to estimate the third. By converting the measured LGRB host masses and redshifts from [@Perley_shoals_masses] to expected metallicities, using a fitting of the mass-metallicity-redshift relation of typical galaxies from [@Zahid2013], we further find that the LGRB host galaxy mass distribution increase with redshift seen in the [@Perley_shoals_masses] SHOALS sample is consistent with that needed to preserve the LGRB metallicity distribution because the mass metallicity relation decreases with redshift. Furthermore we find that these estimated LGRB host metallicities consistently overestimate the actual measured host metallicities by approximately a quarter dex. This corresponds to about a factor of two in raw metallicity and resolves much of the difference between the LGRB formation metallicity cutoff of about a third solar in [@diff_rate_letter] with the cutoff value of just under solar claimed in [@Perley_shoals_masses] in favor of the former. As LGRB hosts do not follow the general mass metallicity relation, there is no substitute for actually measuring their metallicities! While one cannot use the mass-metallicity relation for estimating LGRB metallicities for the simple reason that the majority of LGRB host galaxies have metallicities below that expected from the relation, the difficulties in getting spectra for frequently low mass galaxies at high redshift are considerable and likely will prohibit obtaining the type of ideal unbiased spectroscopic sample that this work requires. It may be possible to use cosmological multi-galaxy simulations to approach the problem, however great care would be required in both ensuring that the population of high redshift low metallicity galaxies is correctly rendered (given the limited observational constraints on this part of the galaxy sample) and that the inevitable extrapolation with increasing redshift is sufficiently constraining as to be useful (i.e. that there are not so many free parameters to arrive at a wide multitude of possible high redshift galaxy metallicity distributions). Still the possibility this could be practical suggests it should be attempted. Speculative Discussion of Possible Causes {#Discussion} ========================================= The absence of evolution in the LGRB metallicity distribution is quite puzzling and does not seem to conform to our general expectations of galaxy evolution. While we cannot fully constrain the impact of selection effects on the sample of host galaxies for which we have spectroscopy, that the absence of evolution is also seen in (the upper half of the distributions of) metallicities estimated from host galaxy masses and redshifts, of which very complete samples exist, makes it unlikely that the observed lack of evolution is the product of selection biases. Nor would the product of selection biases have reason to particularly favor a metallicity distribution that remains constant with redshift. Moreover while we cannot fully account for how selection effects are impacting our measured metallicity results, we find that the redshift evolution in the metallicity distributions estimated from host galaxy masses taken from the unbiased SHOALS sample give results that are consistent with our directly measured metallicity distributions. Therefore we proceed assuming the observed constant metallicity distribution is true and address its origins and implications. The simplest explanation of the known LGRB preference for low metallicity environments [@stats_paper; @diff_rate_letter; @Perley_shoals_masses; @Levesque051022; @Levesque2] is that this effect is caused by a difference in the LGRB formation rate per unit star-formation at different metallicities (c.f. @diff_rate_letter). However, this explanation would not produce an LGRB metallicity distribution with no redshift evolution, as seen here, since the underlying star-formation from which the LGRBs are formed does evolve and thus would be reflected in the LGRB metallicity distribution. One possible explanation might be that the high metallicity LGRB hosts are not representative of the general galaxy population and may have recently been low metallicity galaxies that have undergone a sudden burst of enrichment. This could also explain the existence of high metallicity LGRBs as the gas that formed the progenitor may have been segregated from this enrichment process and thus allowing the LGRB to form in a low metallicity environment. The recent discovery of host association confusion leading to the spurious high metallicity host measurement of LGRB 020819B [@Perley020819B] and potentially also LGRB 050826 (C. Th[ö]{}ne private communication) demands a closer vetting of the high metallicity LGRB host population. Perhaps then claims of high metallicity LGRB hosts being dominated by merger induced starbursts could be validated and the disproportionate (with respect to the underlaying star-formation) association of high metallicity LGRB hosts with dynamical systems shown. At the moment this remains speculative as, while the high metallicity LGRB 051022 is clearly such a system [@conference_proceedings; @obs_paper], an extensive study of a well vetted representative population will be needed to establish this trend beyond mere anecdotal cases. Were merger induced starbursts (of low metallicity galaxies) the cause of the high metallicity LGRB host population then the rate of such mergers would likely not depend on the fraction of high metallicity galaxies and their high metallicity star-formation. Therefore the general galaxy metallicity distribution (and its evolution) would largely be irrelevant, as it is thus not producing LGRBs, and therefore the LGRB metallicity distribution (at least for all but the low metallicity end) would be a function of the distribution in the metallicity enrichment timescale and the timescale distribution between LGRB progenitor gas segregation and the LGRB explosion. Essentially this reduces the problem to a race between how fast the host galaxy metallicity can increase and how long the LGRB can wait to explode. Such a race condition would be expected to produce a static LGRB metallicity distribution. While LGRB progenitors are certainly massive stars it remains an open question of whether they are the most massive and thus if, in a starburst scenario, they are the first stars to explode. stars would actually be problematic as we need to enrich the host galaxy ISM without enriching the LGRB progenitor itself. Assuming it is possible to enrich the host galaxy ISM in the interval between when the gas that will form the LGRB progenitor becomes segregated from that of the host to when the LGRB progenitors lifespan ends in an LGRB event, this would explain both how the high (host) metallicity LGRBs form and why the LGRB host galaxy metallicity distribution doesn’t evolve. Unfortunately, the required understanding of rapid, perhaps collisionally induced, starburst metallicity enrichment is lacking. The limitations of rapid enrichment are not observationally constrained and the simulations not yet sufficient to address the issue directly. A detailed study of high metallicity LGRB host galaxies to compare their properties in detail against a matching typical star forming galaxy population (i.e. investigating anecdotal claims of merger over representation) and extending our study of the LGRB host metallicity distribution out to redshifts where the metallicity of typical star forming galaxy population is consistent with or lower than the typical LGRB population (perhaps at z $\sim$ 4) is essential to continuing this investigation. While some difference in the LGRB metallicity distribution would still be expected because the average initial (i.e. pre merger enrichment) metallicites of the future LGBR host galaxies would still be expected to vary slightly (due to normal metallicity enrichment in these sub typical metallicity galaxies), this effect is likely to be small as long as the (star-formation weighted) post merger (rapidly enriched) mean host galaxy metallicity is much higher than the typical LGRB metallicity formation range. A better understanding of the metallicity evolution for the sub typical metallicity galaxy population is needed, but conceptually it is not hard to view galaxies which have failed to keep pace with the typical galaxy evolution and enrichment as being non-evolved / primordial galaxies. At a high enough redshift even galaxies following the typical galaxy evolution track will eventually have to be metal poor enough so as to be in within the typical low metallicity range conducive to LGRB formation. However this redshift is (likely far) beyond the redshift range of the [@xshooter_survey] host galaxy metallicity sample (0 $\lesssim$ [z]{} $\lesssim$ 2.5). We may already have some very limited evidence of this effect in the behavior of the [z]{} $>$ 2.5 line in [first@refmeasured\_and\_estimated,@]{} right however as this is beyond the redshift range of the [@Zahid2013] galaxy populations this could also simply be caused by uncertainty in the requisite extrapolation. It is worth noting that the highest redshift galaxy population in [@Zahid2013], that of [@Erb2006] at a redshift of [z]{} = 2.26, still expects a mean metallicity of only log(O/H)+12 $\approx$ 8.5 (after conversion into the KK04 scale) for a billon solar mass galaxy at that redshift whereas [@diff_rate_letter] found the metallicity cutoff to be log(O/H)+12 $\approx$ 8.3 for the LGRB formation per unit star-formation rate. Even if such an extrapolation of the @Zahid2013 fit could be trusted outside of its observed redshift range, it would still be unwise to draw too heavily on inferring host metallicity behavior from host masses as one of the significant results of this paper is that the former are not a typical representation of the latter. We must stress again that there is no substitute for actually measuring galaxy metallicities. We note however that this explanation is at odds with some recent theoretical [@Bustamante2018; @Torrey2019] and observational [@Wilson2018] work suggesting merger induced decreases in gas-phase metallicity. However it has been observed in LGRB 980425 that at least this, and probably many, LGRB hosts are subject to very recent inflows of low-metallicity gas in localized regions [@inflow980425]. These inflows could be providing primordial gas from which LGRB progenitors are formed even when the typical HII metallicity of the galaxy is too high. High resolution spectroscopy of the LGRB 980425 progenitor region does not show a lower metallicity than typical for the galaxy [@Christensen980425], despite the galaxy having a favorable face on geometry, (nor do similar observations of LGRB 111005A however this galaxy was edge on — @111005Tanga) so this possibility is less than compelling. Much more work is needed to probe the effect of mergers on the galaxy’s metallicity change over time and across a full range of different initial metallicities. To be relevant to transient population rates such work would also need to apply a star-formation weighted analysis (i.e. look at the average effect of mergers in terms of star-formation and not just on typical numerous small galaxies) and consider galaxies beyond when they appear as classical merging systems. In general merger induced star-formation likely contributes significantly to the metallicity enrichment over cosmic time and thus any such merger induced metallicity decreases are likely temporary. Still this discussion is rather speculative and a much better understanding of galaxies is needed to understand LGRB rates. There may be a simpler, and more direct (though still tentative) explanation. [@Hakobyan2014] found evidence that the relative rate of Type Ibc to Type II SNe is significantly higher in merging galaxies than in undisturbed star-forming galaxies. As the merger rate (particularly the major merger rate) increases with redshift [@Lotz2011; @Snyder2017] the formation of Type Ic SNe (and thus potentially LGRBs) may be enhanced as well. How mergers produce this change in the type of SN formed is not known. While there is now some growing evidence that the stellar initial mass function may vary [@Conroy2013; @Navarro2015], we have no evidence that this is the cause. Most recently, [@Schneider2018] have argued that the 30 Doradus starbust in the Large Magellanic Cloud has more massive (15 to 200 $M_\sun$) stars than would be expected from the Galactic IMF. However, if this change is due to metallicity, it might help explain why LGRBs prefer low-metallicity galaxies, but would not make them more likely in high-metallicity galaxies at high redshift. Conclusions =========== The ratio of low, intermediate (i.e. below and above the @diff_rate_letter log(O/H)+12 $\approx$ 8.3 metallicity cutoff), and super solar LGRB hosts does not change dramatically across our sample as would be expected for luminosity biases in host spectroscopy. [In particular, LGRBs with host metallicities $log(O/H)+12 >8.4$ remain a constant fraction of the entire known LGRB population between redshifts of $z=0$ and $z=2.5$ even as the fraction of LGRBs with metallicities drops steeply.]{} The measured metallicity distribution of the LGRB host galaxy population does not seem to evolve with redshift as would be expected given the cosmological metallicity enrichment of the universe, in particular the lower average metallicity of star-forming environments at high redshift. [At the same time, the metallicity of typical galaxies with the same mass and redshift as the LGRB hosts is systematically approximately twice that of the measured metallicities, independent of redshift. The LGRB host galaxy mass distribution slowly increases with redshift as might be expected to maintain a constant metallicity distribution.]{} While we cannot fully exclude the possibility of Malmquist bias contributing to our results, we have analyzed in detail possible redshift biases between our samples and find no such unexplained effects. We also would not expect Malmquist bias to reproduce either consistent observed metallicity distributions or the mass increase with redshift needed to exactly maintain consistent estimated metallicity distributions. Malmquist bias certainly would not give us a constant fraction of GRBs with metallicities above $log(O/H)+12 > 8.4$ (see [first@refZ\_frac,@]{}). Were Malmquist effects biasing us towards lower metallicity for a given mass (due to lower UV stellar opacities resulting in higher equivalent widths)[^7] then the constant metallicity distribution we observe would require a true metallicity distribution that increases with redshift contrary to typical galaxy evolution trends. Therefore we conclude that something more complicated is occurring than the simple relative rate difference in LGRB formation as a function of metallicity proposed in [@diff_rate_letter]. One potential explanation is that the LGRB events seen in high metallicity environments do not actually originate from high metallicity progenitors. Instead these events may originate from low metallicity star formation and the metallicity of their environments is otherwise enriched (e.g. the enrichment happening after the progenitor was formed). Alternatetively, there is the possible that conditions in galaxies at higher redshifts are changing in ways that make the production of LGRBs more likely at high, or apparently high, metallicity. The latter could be through the localized infall of low-metallicity gas, but there could be more fundamental changes in the nature of star formation in galaxies at high redshift, such as a change in the IMF, which could itself be due to processes such as mergers, which at low-redshifts have been shown to change the ratio of SNe types formed [@Hakobyan2014]. However, our result that the LGRB host metallicity distribution remains largely unchanged as we go to higher redshifts, when we expect quite the opposite, is quite possibly an important clue to the nature of massive star formation at high redshifts. Ideally we would extend our analysis of the LGRB host metallicity distribution until the mass metallicity relation of typical galaxies descended into the typical low metallicity, $log(O/H)+12 < 8.3$, LGRB metallicity range. However, actually measuring galaxy metallicities in emission beyond [z]{} $>$ 2 becomes increasing difficult due to the required spectral lines being redshifted out of the observable range. X-shooter with the K band blocking filter in place is limited to observations at [z]{} $<$ 2.15, and without it, observations are only practical out to [z]{} $<$ 2.6 which is also the limit of what can be reasonably expected from the ground. JWST will be able to brute force these observations at high redshifts however a more efficient approach would be to use absorption metallicities at such higher redshifts. Fortuitously, LGRB host galaxies themselves are uniquely suited to absorption metallicity measurement as the GRB afterglow itself provides a bright background source clean of intrinsic spectral features (as present in QSOs) and LGRB hosts are also star-forming galaxies with typically robust emission lines. Absorption metallicities however are practically limited to galaxies at [z]{} $>$ 1.6 due to requiring observations of the Ly$\alpha$ line. The lack of much overlap between these redshift ranges has deterred a direct comparison of measured values leaving great uncertainty in their respective cross-calibration. Once such a cross-calibration between emission and absorption metallicities (similar to the cross-calibration between different emission line diagnostics of @KewleyEllison) is achieved, extending our analysis out to a higher redshifts in absorption will allow us to directly probe the critical region where the mass metallicity relation of the typical galaxy population transits the optimal metallicity range for LGRB formation. Such analysis is critical to understand how metallicity shapes the formation process of the LGRBs seen in high metallicity host galaxies and thus how LGRBs form in general. More generally we find that even though estimating the metallicity of LGRB hosts from their mass and redshift broadly follows observed trends, the resulting distribution is strongly biased to higher metallicities than the observations. Even a specific LGRB-calibrated mass-metallicity relation is not practical since there is no universal correspondence between LGRB host mass and metallicity, regardless of redshift, at the low metallicity end of the population. To say it again: as LGRB hosts do not follow the general mass metallicity relation and thus there is no substitute for actually measuring their metallicities! We thank Thomas Kr[ü]{}hler for many useful discussions and assorted assistance. His omission as a coauthor is due solely to his personal preference since leaving academia. His presence in astronomy is missed and we wish him well in his new career. We thank the BAT6 team for a detailed explanation of their [z]{} $\gtrsim$ 2.5 selection effects. John Graham acknowledges support through the National Science Foundation of China (NSFC) under grant 11750110418. APPENDIX A {#appendix-a .unnumbered} ========== Data and KS Tables {#data-and-ks-tables .unnumbered} ================== Here we provide a data table of our combined LGRB sample ([first@refxsd\_data\_table,@]{}) and KS test result tables for Figures \[redshift\_distribution\] ([first@refLGRB\_num\_stat\_KS\_table,@]{}) and \[measured\_and\_estimated\]. [first@refmeasured\_and\_estimated,@]{} has KS test results for the left (measured host metallicity – [first@refcomb\_step\_plot\_5\_KS\_table,@]{}), inset (host mass – [first@refmass\_comb\_step\_plot\_5\_KS\_table,@]{}), and right (estimated host metallicity – [first@refmetal\_est\_comb\_step\_plot\_5\_KS\_table,@]{}) plots as well as another KS test ([first@refmetal\_est\_comb\_step\_plot\_5\_KS\_table\_up\_half,@]{} and [first@refshoals\_w\_limits\_metal\_est\_step\_plot\_5\_KS\_table\_up\_half,@]{}) on only the upper half of the values in each redshift bin of [first@refmeasured\_and\_estimated,@]{} left and [first@refshoals\_metal\_est\_step\_plot\_5,@]{} respectively. These KS test result tables are computed (for each figure) by calculating the KS probabilities of every line in the figure against every other line. [first@refLGRB\_num\_stat\_KS\_table,@]{} shows that the redshift distribution of the different observed LGRB populations are in reasonable agreement with each other except as noted in section \[enumerated\_pop\_differences\]. By further dividing the populations (into subpopulations not separately shown in \[redshift\_distribution\]) we can validate that these exceptions are due the reasons claimed in the text. Specifically we divide the [@stats_paper] sample into objects with redshift above and below z = 0.5 and find that both of these subpopulations are much more consistent with the other lines, while the combined population is not, due to a lower observing rate at [z]{} $>$ 0.5 where a separate IR spectrograph is required. We also divide the [@xshooter_survey] metallicity into objects with redshift above and below [z]{} = 2 and again find that both of these subpopulations are much more consistent that the combined population, due to a gap in the [@xshooter_survey] metallicity sample from 1.7 $\lesssim$ [z]{} $\lesssim$ 2.1 caused by a line measurement difficulty. (Note that the sample redshift matching process results in these this redshift gap being trimmed from the comparison samples as well.) Finally for those samples who’s redshift range exceed [z]{} $<$ 2.5 we create a [z]{} $<$ 2.5 subpopulation so as to compare them without being subject to the gamma-ray flux limited selection effects present at higher redshifts and find that these populations are indeed more consistent at lower redshifts. [first@refcomb\_step\_plot\_5\_KS\_table,@]{} shows that the different LGRB redshift bins indeed have the same metallicity distribution as appears in [first@refredshift\_distribution,@]{} left. [first@refmass\_comb\_step\_plot\_5\_KS\_table,@]{} shows however that there is a strong difference in the host mass distribution of the different LGRB redshift bins (see [first@refredshift\_distribution,@]{} inset). It is thus surprising that when these mass and redshift values are used to estimate a host metallicity values for these objects the LGRB redshift bins have a much more consistent metallicity distribution ([first@refmetal\_est\_comb\_step\_plot\_5\_KS\_table,@]{}), particularly on the upper half of the distributions ([first@refmetal\_est\_comb\_step\_plot\_5\_KS\_table\_up\_half,@]{}). To exclude possible biases in the host mass distribution, in [first@refshoals\_w\_limits\_metal\_est\_step\_plot\_5\_KS\_table\_up\_half,@]{}, we repeat this analysis, on the upper half of the distributions, using only the [@Perley_shoals_masses] sample while counting objects with only upper mass limits in the normalization (since all the upper mass limits are contained in the lower half of the distributions they do not disrupt this analysis) with comparable results. \[1\][>m[\#1]{}]{} [m[1.9cm]{}L[1.0cm]{}L[1.8cm]{}L[0.9cm]{}]{} \[1\][>m[\#1]{}]{} / GRAHAM ET AL. [m[4.1cm]{}\|L[1.2cm]{}L[1.5cm]{}L[1.5cm]{}L[1.5cm]{}L[1.2cm]{}L[1.2cm]{}L[1.2cm]{}L[1.2cm]{}L[1.2cm]{}L[1.2cm]{}L[1.2cm]{}L[1.2cm]{}L[1.2cm]{}]{} [Computed Kolmogorov–Smirnov (KS) probabilities comparing the lines in the [first@refredshift\_distribution,@]{} redshift distributions (and some additional subsets thereof). This was affected by, for each paring, removing the objects of each line outside the redshift range of the other (with a grace of 0.05 $z$) and then running a normal KS test on the remaining values. NaN values indicate KS test failure due to having fewer than 4 objects in a redshift matched comparison sample. Values of “1" are exact and indicate a sample which, due to redshift range matching cuts, is being evaluated with itself, values of “1.0000" are the result of rounding. Note: The KS values in this table are not dependent on the line scaling factors used in [first@refredshift\_distribution,@]{}.]{} [m[4.1cm]{}\|L[1.5cm]{}L[1.5cm]{}L[1.5cm]{}L[1.5cm]{}L[1.5cm]{}L[1.5cm]{}L[1.5cm]{}L[1.5cm]{}L[1.5cm]{}L[1.5cm]{}]{} [The number of objects from the [first@refredshift\_distribution,@]{} samples. Note the irregular and overlapping bins from 1.2 $<$ [z]{} $<$ 1.7 and 2.1 $<$ [z]{} $<$ 2.5 to avoid the redshift gaps (from 1.0 $\lesssim$ [z]{} $\lesssim$ 1.2 and 1.7 $\lesssim$ [z]{} $\lesssim$ 2.1) in the [@xshooter_survey] metallicity sample. Also note that the number of simulated objects in the GS16 [@form_rate_letter] Prediction is arbitrary, however the ratio of the number of objects between the different redshift bins is meaningful.]{} \[1\][>m[\#1]{}]{} [m[2.8cm]{}\|L[1.8cm]{}L[1.8cm]{}L[1.8cm]{}L[1.8cm]{}L[1.8cm]{}L[1.8cm]{}L[1.8cm]{}]{} [Computed Kolmogorov–Smirnov probabilities comparing the lines in the [first@refmeasured\_and\_estimated,@]{} left measured host metallicity distributions.]{} [m[2.8cm]{}\|L[1.8cm]{}L[1.8cm]{}L[1.8cm]{}L[1.8cm]{}L[1.8cm]{}L[1.8cm]{}L[1.8cm]{}]{} [Computed Kolmogorov–Smirnov probabilities comparing the lines in the [first@refmeasured\_and\_estimated,@]{} inset mass distributions.]{} [m[2.8cm]{}\|L[1.8cm]{}L[1.8cm]{}L[1.8cm]{}L[1.8cm]{}L[1.8cm]{}L[1.8cm]{}L[1.8cm]{}]{} [Computed Kolmogorov–Smirnov probabilities comparing the lines in the [first@refmeasured\_and\_estimated,@]{} right estimated metallicity distributions. The metallicities are estimated based on the mass metallicity relation for their redshifts.]{} [m[2.8cm]{}\|L[1.8cm]{}L[1.8cm]{}L[1.8cm]{}L[1.8cm]{}L[1.8cm]{}L[1.8cm]{}L[1.8cm]{}]{} [Computed KS probabilities comparing the upper half of the lines in [first@refmeasured\_and\_estimated,@]{} right estimated metallicity distributions. This was affected by removing the lower half of the metallicity values of each line and then running a normal KS test on the remaining values.]{} [m[2.8cm]{}\|L[1.8cm]{}L[1.8cm]{}L[1.8cm]{}L[1.8cm]{}L[1.8cm]{}L[1.8cm]{}L[1.8cm]{}]{} [Computed KS probabilities comparing the upper half of the lines in the [first@refshoals\_metal\_est\_step\_plot\_5,@]{} estimated metallicity distributions. This was affected by removing the lower half of the metallicity values of each line and then running a normal KS test on the remaining values. Fortunately all the upper limits are contained in the lower half of the distributions and thus do not disrupt this analysis.]{} \[1\][>m[\#1]{}]{} [m[1.9cm]{}L[1.8cm]{}L[1.8cm]{}L[2.1cm]{}L[2.4cm]{}L[0.9cm]{}L[1.1cm]{}]{} [^1]: This is discounting some higher redshift objects in [@xshooter_survey] due to use of the \[Ne III\] method for breaking R$_{23}$ metallically degeneracy which should never be applied to LGRB hosts [@stats_paper]. [^2]: For a more detailed description of our metallicity calculation methodology we refer the reader to [@stats_paper]. For a more detailed description of the advantages and accuracy of the iterative fitting approach we refer the reader to [@kd2002]. For readers who wish to calculate their own metallicities we refer them to the excellent newer metallicity code of [@new_metal_code] which is capable of determining metallicities using a range of different diagnostics and scales (note that this code is not used in this paper due solely to our desire to retain full constancy with @stats_paper). For the solar metallicity value in the KK04 scale we refer the reader to the [@solar] estimated of log(O/H)+12 = 8.69 $\pm$ 0.05 based on solar 6300 [Å]{} \[O I\] line measurement. It should be noted that while the emission line diagnostics can be cross calibrated due to the large number of H II regions upon which multiple diagnostics can be applied [@KewleyEllison] this is not true of the density-sensitive 6300 [Å]{} \[O I\] line measurement where the line strength is insufficient for widespread application. Thus any absolute reference to the solar value should be considered very approximate and detailed scientific comparisons to emission line metallicities in terms of solar fractions should be avoided. For conversions to other metallicity scales and a discussion of associated issues we refer the reader to the transforms of [@KewleyEllison]. For a general introduction to the R$_{23}$ diagnostic we refer the reader to [@Pagel1979; @Pagel1980]. [^3]: <http://www.mpe.mpg.de/~jcg/grbgen.html> [^4]: \[rate\_feedback\_bias\_footnote\]It should be noted that we cannot completely exclude a bias in LGRB detectability with redshift within the theoretical curve, since the theoretical curve is based on an estimate of the evolution in the metallicity distribution of star-formation across the universe as a function of redshift, which is in turn based on observational estimates of the LGRB rate as a function of redshift. While these estimates are not fully decoupled, we believe it unlikely (but possible) that a LGRB detectability bias would match one of the metallicity evolution cases evaluated in [@form_rate_letter] Table 3. In particular none of these models shows a change in shape at [z]{} $\sim$ 2.5 as seen in many of the [first@refzd,@]{} populations. [^5]: Irregular bin spacing to avoid redshifts gaps in measured metallicites [^6]: Errors on object numbers assumed to be square root of the number of objects. [^7]: High metallicity creates a greater UV opacity in the atmospheres of the galaxies massive stars, reducing the ionizing photon flux, and thus the amount of ionized gas. Also higher metallicity galaxies can support more dust which also reduces the UV photon density. As higher equivalent widths makes it more likely we can measure a metallicity, we may be biasing ourselves towards lower metallicity at a given luminosity.	{ "pile_set_name": "ArXiv" }
--- abstract: 'In the two-dimensional electron gas (2DEG) emerging at the transition metal oxide surface and interface, it has been pointed out that the Rashba spin-orbit interaction, the momentum-dependent spin splitting due to broken inversion symmetry and atomic spin-orbit coupling, can have profound effects on electronic ordering in the spin, orbit, and charge channels, and may help give rise to exotic phenomena such as ferromagnetism-superconductivity coexistence and topological superconductivity. Although a large Rashba splitting is expected to improve experimental accessibility of such phenomena, it has not been understood how we can maximally enhance this splitting. Here, we present a promising route to realize significant Rashba-type band splitting using a thin film heterostructure. Based on first-principles methods and analytic model analyses, a tantalate monolayer on BaHfO$_3$ is shown to host two-dimensional bands originating from Ta $t_{2g}$ states with strong Rashba spin splittings - up to nearly 10% of the bandwidth - at both the band minima and saddle points due to the maximal breaking of the inversion symmetry. Such 2DEG band structure makes this oxide heterostructure a promising platform for realizing both a topological superconductor which hosts Majorana fermions and the electron correlation physics with strong spin-orbit coupling.' author: - Minsung - Jisoon - Suk Bum bibliography: - 'rashbaoxide.bib' date: - - title: 'Strongly enhanced Rashba splittings in oxide heterostructure: a tantalate monolayer on BaHfO$_3$' --- Recently, the spin-orbit interaction of the two-dimensional electron gas (2DEG) at the surfaces and interfaces of the perovskite transition metal (TM) oxide [-@Ohtomo2004; -@Takagi2010; -@Mannhart2010; -@Santander2011; -@Meevasana2011] has been much investigated experimentally [-@BenShalom2010; -@Caviglia2010; -@Santander2014; -@Santander2012; -@King2012; -@Reyren2012; -@MKim2012]. However, definite understanding on how its magnitude might be maximized has not been well established. It is the combination of the broken inversion symmetry and the atomic spin-orbit coupling (SOC) of the TM that gives rise to a non-zero spin splitting in the form of the Rashba spin-orbit interaction [-@Bychkov1984; -@Dresselhaus1955; -@Winkler2003]. But this origin implies that the magnitude of the spin-orbit interaction is intrinsically limited by the TM atomic SOC. The limitation should be apparent in the best-studied perovskite 2DEGs — the SrTiO$_3$ (STO) surface and the LaAlO$_3$/SrTiO$_3$ (LAO/STO) heterostructure interface — as the atomic SOC strength of the 3$d$ TM Ti is relatively small [-@Zhong2013; -@Khalsa2013; -@PKim2014]. Experimental evidences have been mixed, with the claims of large magnitude stemming from the magnetoresistance measurements [-@BenShalom2010; -@Caviglia2010; -@Santander2012; -@King2012; -@Santander2014] contradicted by the Hanle effect measurement [-@Reyren2012] as well as the measurement of similar magnetoresistance in the $\delta$-doped STO heterostructure where the inversion symmetry breaking is hardly present [-@MKim2012]. Meanwhile, theoretical calculations show the splitting near the $\Gamma$ point to be two orders of magnitude smaller than the bandwidth at the best [-@Zhong2013; -@YKim2013; -@PKim2014]. One natural way to overcome this limitation is adopting 5$d$ TM oxides, such as tantalate, with a stronger atomic SOC. This has motivated the recent experiments on the 2DEG at the surface of KTaO$_3$ (KTO) [-@Santander2012; -@King2012]. However, the experiments on KTO have suggested that another important condition for enhancing the surface 2DEG Rashba spin-orbit interaction is to have the density profile of the surface state concentrated to the surface-terminating layer, which maximizes the effect of the broken inversion symmetry. The ARPES measurements on the KTO surface have seen no measurable spin splitting [-@Santander2012; -@King2012], in spite of not only the stronger SOC of Ta but also the polar nature of KTO (001) surface. According to a density functional theory calculation [-@Shanavas2014], the surface state penetrates deeply into the bulk as the surface confinement potential is made shallow by the atomic relaxation near the surface layer. This suppresses the effect of the inversion symmetry breaking (ISB) on the surface state (which can be quantified by various parameters, [e.g.]{} the chiral orbital angular momentum coefficient [-@SPark2011; -@BKim2012; -@CPark2012; -@PKim2014]), and hence significantly reduces the Rashba spin-orbit interaction. ![\[fig:atomicstr\]Atomic structure of a tantalate layer on BaHfO$_3$ (001). (a) Schematic illustration of TaO$_2$/KO on HfO$_2$-terminated BaHfO$_3$, and TaO$_2$ on BaO-terminated BaHfO$_3$. (b) Atomic structure of TaO$_2$/KO on BaHfO$_3$ from first-principles calculations; note the height difference between Ta and O at the top layer. (c) Wave function weight projected on $d_{xz/yz}$ of TMs for the lowest $d_{xz/yz}$ Rashba band at $\Gamma$. Monolayer (TaO$_2$/KO on HfO$_2$-terminated BaHfO$_3$) case is to be compared with bilayer (TaO$_2$/KO/TaO$_2$ on BaO-terminated BaHfO$_3$) case. The TM-O$_2$ layers are numbered starting from the outermost layer. ](fig1.pdf){width="40.00000%"} ![\[fig:bandpdos\]Electronic structure of TaO$_2$/KO on BaHfO$_3$ from first-principles calculations. (a) Calculated band structure along high symmetry points. Insets show the magnified view of the upper and lower $d_{xz/yz}$ bands near $\Gamma$, and the $d_{yz}$ bands near $\mathrm{X}$. (b) Projected weights of Ta $d$ states. The Fermi level is set to the valence band maximum. (c) Schematic illustration of energy levels at $\Gamma$ without SOC. The crystal field splitting of the monolayer-substrate heterostructure is different from that of the cubic (octahedral) case.](fig2.pdf){width="45.00000%"} In this study, we have theoretically constructed a realistic oxide heterostructure that has a surface 2DEG with a strong Rashba spin-orbit interaction. Our idea is to consider a 5$d$ TM oxide monolayer on a substrate, where 2DEG predominantly lies in the outermost monolayer film, maximizing the effect of the broken inversion symmetry from the substrate. Specifically, we attempt to replace the outer layers of perovskite oxide (001) surface with another perovskite thin film layers, in which the electronic bands of the substrate need to lie sufficiently far from the conduction band minimum (CBM) to make all essential low-energy physics originate from the thin film states near CBM. After calculations of a number of candidate perovskite oxides for the heterostructure, we find that TaO$_2$/KO or TaO$_2$/BaO layer on BaHfO$_3$ (001) surface (Fig. \[fig:atomicstr\]) is a promising candidate structure possessing Ta $t_{2g}$-originated two-dimensional (2D) bands with a strong Rashba spin-orbit interaction. BaHfO$_3$ (BHO) is suitable as a substrate because its lattice structure (cubic at room temperature) matches with that of KTO, the only stable perovskite material containing TaO$_2$ [-@Zhurova2000; -@Maekawa2006], and the alignment of its conduction bands and the Ta $t_{2g}$ bands enables minimal hybridization. It should be emphasized that the concentration of the surface state in the outermost layer is an important condition to maximize the ISB effect. For instance, if we consider a TaO$_{2}$ bilayer as opposed to the monolayer (Fig. \[fig:atomicstr\]c), the ISB effect is weakened as the surface state wave function no longer peaks at the outermost layer. We will show that the coupling of the $t_{2g}$ surface bands to the Ta $e_g$ bands, which comes from the local asymmetric environment of the surface Ta atoms, plays a key role in the enhanced splitting. The band splitting from intra-$t_{2g}$ coupling is smaller, due to specific orbital symmetry of Ta $d$ and O $p$ states that we will discuss. We will further show that the $t_{2g}$-$e_g$ coupling gives rise to the enhanced Rashba-Dresselhaus splitting not only at the band bottom at $\Gamma$ but also at the band saddle points at $\mathrm{X}$. Finally, we will consider the substitution of Ba atoms for K in the KO layer for electron doping of the system. Results [Rashba splitting near the $\Gamma$ point.]{} The electronic structures of TaO$_2$/KO monolayer on BHO (lattice constant $\approx 4.155$ Å) from our first-principles calculations are presented in Fig. \[fig:bandpdos\]. The bands near the CBM consist of $t_{2g}$ ($d_{xy}$, $d_{xz}$, $d_{yz}$) states of Ta in the outermost layer, with the calculated bandwidth of $\approx$ 1.7 eV for the $d_{xz/yz}$ bands. These bands being 2D, the triple degeneracy (excluding spin) of the $t_{2g}$ bands at $\Gamma$ is lifted, splitting the $d_{xy}$ and the $d_{xz/yz}$ manifolds; the Ta atomic SOC further splits the $d_{xz/yz}$ bands into upper and lower $d_{xz/yz}$ states. Finally, when the ISB at the surface is accounted for, the Rasha-type band splitting lifts spin degeneracies in the entire Brillouin zone (BZ) except at the time-reversal invariant momenta $\Gamma$, $\mathrm{X}$ and $\mathrm{M}$. We note that this Rashba-type band splitting of the $d_{xz/yz}$ bands is strikingly larger in magnitude than that of the $d_{xy}$ bands, contrary to the prediction of the $t_{2g}$-only model [-@Khalsa2013; -@YKim2013; -@Scheurer2015]. Moreover, our calculation gives the Rashba coefficient of the lower $d_{xz/yz}$ bands at $\Gamma$ of $\alpha_R \approx 0.3~\mathrm{eV\AA}$, which is an order of magnitude larger than that of LAO/STO heterostructure deduced from the experimental magnetoresistance data [-@Caviglia2010], and the Rashba energy of $E_R \gtrsim 15~\mathrm{meV}$; these values are also significantly larger than $\alpha_R \approx 0.1~\mathrm{eV\AA}$, $E_R \approx 1~\mathrm{meV}$ for the bilayer case of Fig. \[fig:atomicstr\]c. The Rashba-Dresselhaus effect along the BZ boundary is even more pronounced, with a giant splitting ($\approx 180$ meV), which is nearly twice the maximum reported value [-@Santander2014] in the perovskite oxide 2DEG, occurring near $\mathrm{X}$ along $k_{x/y} = \pi$. ----------- ------------------- --------------------------------------------------------------------------------------------------------- -- -- -- $\vec{k}$ Manifold Splitting terms in $\mathcal{H}_{\mathrm{eff}}$ upper $d_{xz/yz}$ $\left[\frac{-2\gamma_3 \xi}{\Delta_{uxz/yz,x^2-y^2}}+\frac{-2\gamma_1 \xi}{\Delta_{uxz/yz,xy}}\right] (\vec{\sigma}\times \vec{k})\cdot \hat{z}$ lower $d_{xz/yz}$ $\frac{2\sqrt{3}\gamma_2 \xi}{\Delta_{lxz/yz,z^2}} (\vec{\sigma}\times \vec{k})\cdot \hat{z}$ $d_{xy}$ $\frac{-2\gamma_1 \xi}{\Delta_{xy,uxz/yz}} (\vec{\sigma}\times \vec{k})\cdot \hat{z}$ $d_{yz}$ $\left[\frac{-2\sqrt{3}\gamma_2 \xi}{\tilde{\Delta}_{yz,z^2}} +\frac{2\gamma_3 \xi}{\tilde{\Delta}_{yz,x^2-y^2}}\right] \sigma_x k_y -\frac{2\gamma_1 \xi}{\tilde{\Delta}_{yz,xy}}\sigma_y k_x$ ----------- ------------------- --------------------------------------------------------------------------------------------------------- -- -- -- : \[tab:heff\]Splitting terms of the effective Hamiltonian $\mathcal{H}_{eff}$ for Ta $t_{2g}$ manifolds. $\vec{k}$ denotes the reference point of the effective Hamiltonian with $\Gamma=(0,0)$ and $\mathrm{X}=(\pi,0)$. $\Delta (\tilde{\Delta})$ represents the energy difference between the states in the subscript at $\Gamma (\mathrm{X})$, where $uxz/yz$ and $lxz/yz$ mean the upper $xz/yz$ and the lower $xz/yz$, respectively. An analysis that includes all Ta $d$-orbitals – not only the $t_{2g}$ orbitals but also the $e_g$ orbitals – is required in order to understand the two conspicuous features of Fig. \[fig:bandpdos\], the discrepancy between the Rashba splitting of the $d_{xy}$ and the $d_{xz/yz}$ bands, and the giant band splitting along $k_{x/y} = \pi$. We employ an analytic TB model for a qualitative analysis and supplement it with quantitative results from maximally localized Wannier functions (MLWFs). In the TB model, we consider a Hamiltonian for all Ta $d$-orbitals, including the $e_g$ orbitals ($d_{z^2}$, $d_{x^2-y^2}$) in a square lattice [-@Shanavas2014; -@Shanavas2014a] to describe the TaO$_2$ layer 2D bands, $$\begin{aligned} \mathcal{H}=\mathcal{H}_\mathrm{hop}+\mathcal{H}_\mathrm{SOC}+\mathcal{H}_\mathrm{E}+\mathcal{V}_\mathrm{sf},\end{aligned}$$ where the first term $\mathcal{H}_\mathrm{hop}$ describes the nearest-neighbor hopping, and the second term $\mathcal{H}_\mathrm{SOC} = \xi {\bf L} \cdot {\bf S}$ is the atomic SOC, with $\xi \approx 0.26$ eV for Ta. The third term $\mathcal{H}_\mathrm{E}$ includes the additional hoppings that would have been forbidden if not for the ISB: $$\begin{aligned} &\gamma_1&=\langle d_{xy}\| \mathcal{H}_\mathrm{E} \|d_{xz} \rangle_{\hat{y}} =\langle d_{xy}\| \mathcal{H}_\mathrm{E} \|d_{yz} \rangle_{\hat{x}} \nonumber \\ &\gamma_2&=\langle d_{xz}\| \mathcal{H}_\mathrm{E} \|d_{z^2} \rangle_{\hat{x}} =\langle d_{yz}\| \mathcal{H}_\mathrm{E} \|d_{z^2} \rangle_{\hat{y}} \\ &\gamma_3&=\langle d_{x^2-y^2}\| \mathcal{H}_\mathrm{E} \|d_{yz} \rangle_{\hat{y}} =\langle d_{xz}\| \mathcal{H}_\mathrm{E} \|d_{x^2-y^2} \rangle_{\hat{x}} \nonumber,\end{aligned}$$ in which the vectors in the subscripts denote the relative position of the two orbitals with the lattice constant set to 1 for convenience (these ISB hoppings play a role analogous to the chiral orbital angular momentum effect in the $p$-orbital bands [-@SPark2011; -@BKim2012; -@CPark2012]). Here, $\gamma_1$ is intra-$t_{2g}$ ISB hopping while $\gamma_2$ and $\gamma_3$ describe $t_{2g}$-$e_g$ ISB hoppings. The fourth term $\mathcal{V}_\mathrm{sf}$ describes the potential difference originating from the surface field. By deriving the effective Hamiltonian $\mathcal{H}_{eff}$ that acts on each two-fold degenerate band in the weak SOC limit where $\mathcal{H}_\mathrm{hop}+\mathcal{V}_\mathrm{sf}$ is dominant over $\mathcal{H}_\mathrm{SOC}$, we obtain Rashba-type band splitting terms near $\Gamma$ and $\mathrm{X}$ as summarized in Table \[tab:heff\] (see Appendix A for details). Table \[tab:heff\] shows the Rashba coupling to be linear in the ISB hopping $\gamma$ divided by the energy difference between two relevant states $\Delta$. ![\[fig:wanhopping\]Hopping strengths (in eV) between Wannier functions for Ta $d$ states at one site and the neighboring site in $x$ direction. Ta $d$ states and O $p$ states are used for the Wannier function construction. The terms relevant to the ISB $\gamma_1$, $\gamma_2$, $\gamma_3$ are presented. Both direct (horizontal arrows) and indirect (oblique arrows, via O $p$) paths are depicted. Empty arrows indicate terms that would be absent without ISB.](fig3.pdf){width="35.00000%"} One reason why the $e_g$ orbital contribution is crucial for the Rashba splitting in the $t_{2g}$ bands is that the $t_{2g}$-$e_g$ ISB hoppings $\gamma_{2,3}$ are significantly larger than the intra-$t_{2g}$ hopping $\gamma_1$: $\gamma_1 \approx -0.04$ eV, $\gamma_2 \approx -0.25$ eV, $\gamma_3 \approx 0.30$ eV. This is necessary condition for the effective Hamiltonian of Table \[tab:heff\] to give larger Rashba splittings in the $d_{xz/yz}$ bands than $d_{xy}$ as shown in Fig. \[fig:bandpdos\], given that the $d_{xz/yz}$ bands are closer in energy to the $d_{xy}$ band than the $e_g$ bands (albeit within an order of magnitude). The intra-$t_{2g}$ ISB hopping $\gamma_1$ remains relatively small due to the orbital symmetry of Ta $d$ and O $p$ states, which we can see from a quantitative analysis with MLWFs that includes not only the Ta $d$ states but also the neighboring O $p$ states. Examining the hopping parameters relevant to $\gamma_1$, the particularly small ISB hopping between O $p_y$ and Ta $d_{yz}$ along $x$ direction (Fig. \[fig:wanhopping\]) can be attributed to the relative positions and shapes of the two orbitals; the lobes of the two orbitals lie on the $yz$ plane that is perpendicular to the hopping direction ($\hat{x}$), and $p_y$ has maximum amplitude along $y$ direction whereas $d_{yz}$ has a node along it. Thus, we have the negligible Rashba splitting of the $d_{xy}$ band as shown in Fig. \[fig:bandpdos\], despite the smaller energy difference with the $d_{xz/yz}$ bands. The other reason why the $e_g$ orbital contribution is crucial for the Rashba splitting in the $t_{2g}$ bands is the reduced $t_{2g}$-$e_g$ energy splitting. Indeed, when the $t_{2g}$-$e_g$ energy splitting is set to be infinite in Table \[tab:heff\], all the results from the $t_{2g}$-only TB models [-@Khalsa2013; -@YKim2013; -@PKim2014; -@Scheurer2015] are recovered, including the absence of $k$-linear Rashba in the lower $d_{xz/yz}$ band near $\Gamma$. In the case of 3D cubic KTO, the $t_{2g}$-$e_g$ energy separation at $\Gamma$ is calculated to be $\approx 4.6~\mathrm{eV}$, which is larger than that of our system (Fig. \[fig:bandpdos\]a, b). Compared with the 3D cubic bulk case, Figure \[fig:bandpdos\]b shows considerable portion of $d_{z^2}$ states close in energy, [i.e.]{}, less than bandwidth, to the $t_{2g}$ bands; this is due to the absence of an O atom in one of the octahedral points surrounding Ta. Hence, as shown in Fig. \[fig:bandpdos\]c, the local atomic configuration for the Ta atom is close to a square pyramid, where the $d_{z^2}$ and lower $d_{xz/yz}$ are close in energy. This can be taken as a generic result for the case where the 2DEG wave function is confined almost entirely to the outermost layer. The height difference of Ta and O atoms ($\approx 0.20$ Å) in the TaO$_2$ layer enhances the $t_{2g}$-$e_g$ coupling in both ways; the larger effect being the enhancement of the ISB hopping $\gamma_3$, but there is also noticeable lowering of the $d_{x^2-y^2}$ orbital energy level. ![\[fig:am\] Angular momentum texture from first-principles calculations in close vicinity of the high-symmetry points in BZ. Orbital and spin angular momenta are presented for the lower Rashba band of the (a) $d_{xy}$ and (b) lower $d_{xz/yz}$ (c) upper $d_{xz/yz}$ bands near $\Gamma$, and the lower Rashba-Dresselhaus band of the (d) $d_{yz}$ bands near $\mathrm{X}=(\pi,0)$. ](fig4.pdf){width="45.00000%"} The $t_{2g}$-$e_g$ coupling also plays a key role in determining the angular momentum (AM) texture of the $t_{2g}$ bands in close vicinity of $\Gamma$ (Fig. \[fig:am\]). The tetragonal crystal field and the SOC determine the spin-orbital entanglement of the band manifolds; the spin-up and spin-down are in nearly the same orbital state for the $d_{xy}$ bands while they are in nearly orthogonal orbital states for the $d_{xz/yz}$ bands. This, in turn, affects the AM character; the spin AM is dominant in the $d_{xy}$ bands whereas the orbital AM is dominant [-@SPark2011; -@BKim2012; -@PKim2014] in the $d_{xz/yz}$ bands (see Appendix A). The $t_{2g}$-$e_g$ coupling is important in that it gives rise to finite AM in the lower $d_{xz/yz}$ bands and quantitatively changes the AM in the upper $d_{xz/yz}$. In the $t_{2g}$-only TB model, the AM texture of the lower $d_{xz/yz}$ band is completely missing and that of the upper $d_{xz/yz}$ is not quantitatively correct. [Band splitting near the $\mathrm{X}$ point.]{} As shown in Table \[tab:heff\], the lowered symmetry $C_{2v}$ at $\mathrm{X}$ allows the mixture of Rashba and linear Dresselhaus terms in general (see Appendix A), with the linear Dresselhaus larger in magnitude, as shown in Fig. \[fig:am\]d. Due to the anisotropic dispersion of $d_{xz/yz}$ bands, the lowest conduction band at $\mathrm{X}=(\pi,0)$ mainly consists of $d_{yz}$ state. We find that the larger band splitting along $\mathrm{X}$—$\mathrm{M}$ comes from the $t_{2g}$-$e_g$ coupling whereas the smaller splitting along $\mathrm{X}$—$\Gamma$ is due to the intra-$t_{2g}$ coupling. Hence, the giant Rashba-Dresselhaus splitting in vicinity of $\mathrm{X}$ ($\approx 180$ meV) is due to the $e_g$ contribution. It has been recently pointed out [-@Chung2015] that this Rashba-Dresselhaus splitting along $\mathrm{X}$—$\mathrm{M}$ is necessary for weak topological superconductivity, which gives rise to dislocation Majorana zero modes. The Rashba-Dresselhaus splitting near $\mathrm{X}$ also affects the superconducting instability, as it splits the logarithmic van Hove singularity (VHS) of the $d_{xz/yz}$ band saddle point and shifts them away from $\mathrm{X}$ (see Appendix B for the logarithmic VHS splitting). Given that the splitting results in the lower and upper Rashba-Dresselhaus bands reaching VHS separately, the shifted VHSs do not have spin degeneracy. While it has been long recognized that the logarithmic VHS at $\mathrm{X}$ enhances the superconducting instability in the spin-singlet channel [-@Schulz1987; -@Dzyaloshinskii1987; -@Furukawa1998; -@Gonzalez2008; -@Nandkishore2012], it was recently shown [-@Meng2015; -@Yao2015; -@Cheng2015] that the logarithmic VHS away from $\mathrm{X}$ enhances the instability to the spin-triplet $p$-wave superconductivity. The physics at $\mathrm{X}$ should be experimentally accessible, as the VHS at $\mathrm{X}$ is not too far from the band bottom of the $d_{xz/yz}$ in energy ($\approx 0.23$ eV), and we will show in the next section how our heterostructure can be chemically doped all the way to the VHS at $\mathrm{X}$. Discussions ![\[fig:tao2bho\]Electronic structure of a TaO$_2$ layer on BaO-terminated BaHfO$_3$ (001) from first-principles calculations. (a) Band structure along high-symmetry points in BZ. The dotted horizontal line denotes the energy level of the VHS in absence of the Rashba-Dresselhaus splitting, $E_{vH}$, with the Fermi level set to 0. (b) AM texture at $E_{vH}$ with constant energy lines. The red and blue arrows correspond to the orbital and spin AM, respectively. ](fig5.pdf){width="40.00000%"} To actually realize the 2DEG in the TaO$_2$ layer, electron doping is needed because the nominal charge of the TaO$_2$ layer is +1 and that of the KO layer is -1 (BaO and HfO$_2$ layers are neutral.). One possible way would be substituting K atoms with Ba in the KO layer. In this case, the Rashba strength remains still large ($\alpha_R \approx 0.2$ eVÅ) in the lower $d_{xz/yz}$ bands (Fig. \[fig:tao2bho\]a), while the Rashba splitting in the upper $d_{xz/yz}$ bands is greatly suppressed largely because the Ba substitution reduces by 70% the height difference of Ta and O atoms in the TaO$_2$ layer. Once again, the considerable band splitting along $\mathrm{X}$–$\mathrm{M}$ ($\approx 24$ meV) and AM texture at $E_{vH}$ in Fig \[fig:tao2bho\] are contrary to the prediction of the $t_{2g}$-only TB model (see Appendix C and Fig. A1), and hence demonstrate that inclusion of the $e_g$ manifold is essential in understanding the Rashba-Dresselhaus splitting. We expect that partial chemical substitution (TaO$_2$/K$_{1-x}$Ba$_{x}$O layer on BHO) could induce the 2DEG in the TaO$_2$ layer in experiments. At $x=1$, [i.e.]{}, 100% Ba substitution, the Fermi level lies slightly above the VHS (Fig. \[fig:tao2bho\]a), which means we can access the VHS at $x \lesssim 1$. We expect the qualitative features of our Rashba-Dresselhaus splitting to be generic for the (001) perovskite transition metal oxide 2DEG with maximal ISB, where the 2DEG wave function profile is required to be concentrated on the surface-terminating TM-O$_2$ layer. In such an environment, the effective crystal field on the 2DEG $d$ orbitals should be quite different from that of the cubic perovskite, with much lower energy level at least for the $d_{z^2}$ orbital. Even if we use alternative materials for our heterostructure — viability of substituting BHO by BaSnO$_3$ or TaO$_2$ by WO$_2$ still remains to be investigated — we expect a significant role of the transition metal $e_g$ orbitals if the Rashba-Dresselhaus splitting is comparably large. It is also found that compressive strain on the BHO substrate, which might be needed for the feasible deposition of the thin film taking into account relatively large lattice mismatch between KTO and BHO, does not substantially affect the band splitting (see Appendix D and Fig. A2). Considering that 2DEG in an artificial film-substrate system is realized experimentally in SrVO$_3$ thin films on Nb-doped STO [-@Yoshimatsu2011], we expect that our system can be realized in experiments using the state-of-the-art layer-by-layer growth control of perovskite oxide thin films [-@Thiel2006]. Methods [Theoretical approach.]{} We performed density functional theory calculations as implemented in VASP [-@Kresse1993; -@Kresse1996]. Projector augmented-wave method was used [-@Blochl1994]. A plane-wave basis set with the cutoff energy 520 eV was employed, and PBEsol (Perdew-Burke-Ernzerhof revised for solids) exchange-correlation functional was adoped [-@Perdew2008]. We used the lattice constant optimized in bulk calculations of BaHfO$_3$ and the internal atomic positions were fully relaxed until the force became less than 0.01 eV/Å. Details of the analytic tight-binding approximation and effective Hamiltonian description were presented in Appendix A. We employed maximally localized Wannier functions [-@Marzari1997; -@Souza2001; -@Mostofi2008] to further analyze the results of the first-principles calculations. The Wannier functions were constructed for $d$ orbitals of Ta in one set, and $p$ orbitals of three neighboring O as well as $d$ orbitals of Ta in the other set. Acknowledgments We thank Jung Hoon Han, Changyoung Kim, Choong-Hyun Kim, Hyeong-Do Kim, Minu Kim, Hyun-Woo Lee, Hosub Jin, Seung Ryong Park, Cai-Zhuang Wang, Hong Yao, and Jaejun Yu for fruitful discussions and comments. This work was supported by the NRF of the MSIP of the Korean government Grant No. 2006-0093853 (M.K., J.I.) and IBS-R009-Y1 (S.B.C.). Research at Ames laboratory (M.K.) was supported by the U.S. Department of Energy, Office of Basic Energy Sciences, Division of Materials Sciences and Engineering under Contract No. DE-AC02-07CH11358. Computations were performed through the support of the Korea Institute of Science and Technology Information (KISTI) and the National Energy Research Scientific Computing Center (NERSC) in Berkeley, CA. Author contributions All authors contributed extensively to the work presented in this paper. Tight-binding model for the TaO$_2$ film ======================================== We consider the tight-binding Hamiltonian for a TaO$_2$ film ($d$-orbitals in a square lattice) [-@Shanavas2014a], $$\begin{aligned} \mathcal{H}=\mathcal{H}_\mathrm{hop}+\mathcal{H}_\mathrm{SOC}+\mathcal{H}_\mathrm{E}+\mathcal{V}_\mathrm{sf},\end{aligned}$$ where $\mathcal{H}_\mathrm{hop}$ describes the hopping between the nearest neighbors, $\mathcal{H}_\mathrm{SOC}$ is the atomic spin-orbit coupling of Ta, $\mathcal{H}_\mathrm{E}$ describes the orbital mixing due to the inversion symmetry breaking field near the surface, and $\mathcal{V}_\mathrm{sf}$ describes onsite potential changes due to the surface field. Specifically, the hopping term is given by $$\begin{aligned} \mathcal{H}_\mathrm{hop}= { \left( \begin{matrix} \frac{t_{\sigma}+3t_{\delta}}{2}(c_x+c_y) & -\frac{\sqrt{3}}{2}(t_{\sigma}-t_{\delta})(c_x-c_y) & 0 & 0 & 0 \\ -\frac{\sqrt{3}}{2}(t_{\sigma}-t_{\delta})(c_x-c_y) & \frac{3t_{\sigma}+t_{\delta}}{2}(c_x+c_y) & 0 & 0 & 0 \\ 0 & 0 & 2t_{\pi}(c_x+c_y) & 0 & 0 \\ 0 & 0 & 0 & 2(t_{\pi} c_x +t_{\delta} c_y) & 0 \\ 0 & 0 & 0 & 0 & 2(t_{\delta} c_x +t_{\pi} c_y) \\ \end{matrix} \right), } \label{eq:ham_hopping} \nonumber\end{aligned}$$ where the basis is $\{ \|d_{z^2}\rangle, \|d_{x^2-y^2}\rangle, \|d_{xy}\rangle, \|d_{xz}\rangle, \|d_{yz}\rangle \}$, and $t_{\sigma}$, $t_{\pi}$, $t_{\delta}$ are hopping parameters between $d$-orbitals. $c_x$ means $\cos k_x$. The lattice constant is set to 1. The spin-orbit coupling term is $$\begin{aligned} \mathcal{H}_\mathrm{SOC}= { \left( \begin{matrix} 0 & 0 & 0 & 0 & 0 & 0 & 0 & -\frac{\sqrt{3}}{2}\xi & 0 & \frac{\sqrt{3}}{2}\xi i \\ 0 & 0 & 0 & 0 & 0 & 0 & \frac{\sqrt{3}}{2}\xi & 0 & \frac{\sqrt{3}}{2}\xi i & 0 \\ 0 & 0 & 0 & 0 & -\xi i & 0 & 0 & \frac{1}{2}\xi & 0 & \frac{1}{2}\xi i \\ 0 & 0 & 0 & 0 & 0 & \xi i & -\frac{1}{2}\xi & 0 & \frac{1}{2}\xi i & 0 \\ 0 & 0 & \xi i & 0 & 0 & 0 & 0 & -\frac{1}{2}\xi i & 0 & \frac{1}{2}\xi \\ 0 & 0 & 0 & -\xi i & 0 & 0 & -\frac{1}{2}\xi i & 0 & -\frac{1}{2}\xi & 0 \\ 0 & \frac{\sqrt{3}}{2}\xi & 0 & -\frac{1}{2}\xi & 0 & \frac{1}{2}\xi i & 0 & 0 & -\frac{1}{2}\xi i & 0 \\ -\frac{\sqrt{3}}{2}\xi & 0 & \frac{1}{2}\xi & 0 & \frac{1}{2}\xi i & 0 & 0 & 0 & 0 & \frac{1}{2}\xi i \\ 0 & -\frac{\sqrt{3}}{2}\xi i & 0 & -\frac{1}{2}\xi i & 0 & -\frac{1}{2}\xi & \frac{1}{2}\xi i & 0 & 0 & 0 \\ -\frac{\sqrt{3}}{2}\xi i & 0 & -\frac{1}{2}\xi i & 0 & \frac{1}{2}\xi & 0 & 0 & -\frac{1}{2}\xi i & 0 & 0 \end{matrix} \right) } \label{eq:ham_soc}. \nonumber\end{aligned}$$ The inversion symmetry breaking field terms are given by $$\begin{aligned} \mathcal{H}_\mathrm{E}+\mathcal{V}_\mathrm{sf}= { \left( \begin{matrix} \delta_2 & 0 & 0 & -2i\gamma_2 \sin k_x & -2i\gamma_2 \sin k_y \\ 0 & \delta_3 & 0 & -2i\gamma_3 \sin k_x & 2i\gamma_3 \sin k_y \\ 0 & 0 & \delta_1 & 2i\gamma_1 \sin k_y & 2i\gamma_1 \sin k_x \\ 2i\gamma_2 \sin k_x & 2i\gamma_3 \sin k_x & -2i\gamma_1 \sin k_y & 0 & 0 \\ 2i\gamma_2 \sin k_y & -2i\gamma_3 \sin k_y & -2i\gamma_1 \sin k_x & 0 & 0 \\ \end{matrix} \right), } \label{eq:ham_efield} \nonumber\end{aligned}$$ where $$\begin{aligned} &\delta_1&=\varepsilon(d_{xy})-\varepsilon(d_{xz/yz}) \\ &\delta_2&=\varepsilon(d_{z^2})-\varepsilon(d_{xz/yz}) \\ &\delta_3&=\varepsilon(d_{x^2-y^2})-\varepsilon(d_{xz/yz})\\ &\gamma_1&=\langle d_{xy}\| \mathcal{H}_\mathrm{E} \|d_{xz} \rangle_{\hat{y}} =\langle d_{xy}\| \mathcal{H}_\mathrm{E} \|d_{yz} \rangle_{\hat{x}} \\ &\gamma_2&=\langle d_{xz}\| \mathcal{H}_\mathrm{E} \|d_{z^2} \rangle_{\hat{x}} =\langle d_{yz}\| \mathcal{H}_\mathrm{E} \|d_{z^2} \rangle_{\hat{y}} \\ &\gamma_3&=\langle d_{x^2-y^2}\| \mathcal{H}_\mathrm{E} \|d_{yz} \rangle_{\hat{y}} =\langle d_{xz}\| \mathcal{H}_\mathrm{E} \|d_{x^2-y^2} \rangle_{\hat{x}} .\end{aligned}$$ The Hamiltonian near the $\Gamma$ point can be written as $$\begin{aligned} \mathcal{H}(\vec{k})\approx \hspace{9cm} \\ {\scalebox{0.65}{\mbox{\ensuremath{\displaystyle \left( \begin{matrix} \frac{t_{\sigma}+3t_{\delta}}{2} C+\delta_2 & 0 & -\frac{\sqrt{3}}{2}(t_{\sigma}-t_{\delta}) D & 0 & 0 & 0 & 0 & -\sqrt{2}\gamma_2 (i k_x-k_y) & -\sqrt{\frac{3}{2}} \xi & -\sqrt{2}\gamma_2 (i k_x+k_y) \\ 0 & \frac{t_{\sigma}+3t_{\delta}}{2} C+\delta_2 & 0 & -\frac{\sqrt{3}}{2}(t_{\sigma}-t_{\delta}) D & 0 & 0 & -\sqrt{2}\gamma_2 (i k_x+k_y) & 0 & -\sqrt{2}\gamma_2 (i k_x-k_y) & \sqrt{\frac{3}{2}} \xi \\ -\frac{\sqrt{3}}{2}(t_{\sigma}-t_{\delta}) D & 0 & \frac{3t_{\sigma}+t_{\delta}}{2} C+\delta_3 & 0 & -\xi i & 0 & \frac{1}{\sqrt{2}} \xi & -\sqrt{2}\gamma_3 (i k_x+k_y) & 0 & -\sqrt{2}\gamma_3 (i k_x-k_y) \\ 0 & -\frac{\sqrt{3}}{2}(t_{\sigma}-t_{\delta}) D & 0 & \frac{3t_{\sigma}+t_{\delta}}{2} C+\delta_3 & 0 & \xi i & -\sqrt{2}\gamma_3 (i k_x-k_y) & -\frac{1}{\sqrt{2}} \xi & -\sqrt{2}\gamma_3 (i k_x+k_y) & 0 \\ 0 & 0 & \xi i & 0 & 2t_{\pi} C+\delta_1 & 0 & -\frac{1}{\sqrt{2}} i \xi & \sqrt{2} i \gamma_1 (i k_x+k_y) & 0 & -\sqrt{2} i \gamma_1 (i k_x-k_y) \\ 0 & 0 & 0 & -\xi i & 0 & 2t_{\pi} C+\delta_1 & -\sqrt{2} i \gamma_1 (i k_x-k_y) & -\frac{1}{\sqrt{2}} i \xi & \sqrt{2} i \gamma_1 (i k_x+k_y) & 0 \\ 0 & \sqrt{2}\gamma_2 (i k_x-k_y) & \frac{1}{\sqrt{2}} \xi & \sqrt{2}\gamma_3 (i k_x+k_y) & \frac{1}{\sqrt{2}} i \xi & -\sqrt{2} i \gamma_1 (i k_x+k_y) & (t_{\pi} +t_{\delta}) C+\frac{\xi}{2} & 0 & (t_{\pi} -t_{\delta}) D & 0 \\ \sqrt{2}\gamma_2 (i k_x+k_y) & 0 & \sqrt{2}\gamma_3 (i k_x-k_y) & -\frac{1}{\sqrt{2}} \xi & \sqrt{2} i \gamma_1 (i k_x-k_y) & \frac{1}{\sqrt{2}} i \xi & 0 & (t_{\pi} +t_{\delta}) C+\frac{\xi}{2} & 0 & (t_{\pi} -t_{\delta}) D \\ -\sqrt{\frac{3}{2}} \xi & \sqrt{2}\gamma_2 (i k_x+k_y) & 0 & \sqrt{2}\gamma_3 (i k_x-k_y) & 0 & \sqrt{2} i \gamma_1 (i k_x-k_y) & (t_{\pi} -t_{\delta}) D & 0 & (t_{\pi} +t_{\delta}) C-\frac{\xi}{2} & 0 \\ \sqrt{2}\gamma_2 (i k_x-k_y) & \sqrt{\frac{3}{2}} \xi & \sqrt{2}\gamma_3 (i k_x+k_y) & 0 & -\sqrt{2} i \gamma_1 (i k_x+k_y) & 0 & 0 & (t_{\pi} -t_{\delta}) D & 0 & (t_{\pi} +t_{\delta}) C-\frac{\xi}{2} \end{matrix} \right), }}}} \label{eq:ham_k_neargamma} \nonumber\end{aligned}$$ with $C=\cos k_x + \cos k_y \approx 2-\frac{k_x^2}{2}-\frac{k_y^2}{2}$, $D=\cos k_x - \cos k_y \approx -\frac{k_x^2}{2}+\frac{k_y^2}{2}$, where we performed a unitary transformation to diagonalize the $d_{xz/yz}$ subspace in the limit that the $d_{xz/yz}$ states are sufficiently far from other manifolds and $\vec{k} \rightarrow 0$. The effective Hamiltonian can be obtained by projection onto the concerned manifold $$\begin{aligned} \mathcal{H}_{\mathrm{eff}}=\mathcal{P}\mathcal{H}\mathcal{P}+\mathcal{P}\mathcal{H}\mathcal{Q} \frac{1}{\epsilon-\mathcal{Q}\mathcal{H}\mathcal{Q}}\mathcal{Q}\mathcal{H}\mathcal{P}, \label{eq:ham_eff}\end{aligned}$$ where $\mathcal{P}$ is the projection operator onto the relevant manifold and $\mathcal{Q}=1-\mathcal{P}$. For the $d_{xy}$ bands, the effective Hamiltonian is $$\begin{aligned} \mathcal{H}_{\mathrm{eff}} \approx h_{xy}(\vec{k})I_{2\times 2}+\frac{-2\gamma_1 \xi}{\Delta_{xy,uxz/yz}} (\vec{\sigma}\times \vec{k})\cdot \hat{z}, \label{eq:ham_eff_gamma_dxy}\end{aligned}$$ where $\Delta_{xy,uxz/yz}=4t_{\pi}+\delta_1-\{2(t_{\pi}+t_{\sigma})+\frac{\xi}{2}\}$, and the Pauli matrices describe the subspace defined by $\{ \|d_{xy}\uparrow\rangle, \|d_{xy}\downarrow\rangle \}$. For the lower $d_{xz/yz}$ bands, $$\begin{aligned} \mathcal{H}_{\mathrm{eff}} \approx h_{lxz/yz}(\vec{k})I_{2\times 2} +\frac{2\sqrt{3}\gamma_2 \xi}{\Delta_{lxz/yz,z^2}}(\vec{\sigma}\times \vec{k})\cdot \hat{z}, \label{eq:ham_eff_gamma_lxz/yz}\end{aligned}$$ where $\Delta_{lxz/yz,z^2}=2(t_{\pi}+t_{\delta})-\frac{\xi}{2}-\{t_{\sigma}+3t_{\delta}+\delta_2\}$, and the Pauli matrices describe the subspace defined by $\{ \frac{1}{\sqrt{2}}(\|d_{xz}\downarrow\rangle +i \|d_{yz}\downarrow\rangle), \frac{1}{\sqrt{2}}(\|d_{xz}\uparrow\rangle -i \|d_{yz}\uparrow\rangle) \}$. For the upper $d_{xz/yz}$ bands, $$\begin{aligned} \mathcal{H}_{\mathrm{eff}} \approx h_{uxz/yz}(\vec{k})I_{2\times 2} +\left[\frac{-2\gamma_3 \xi}{\Delta_{uxz/yz,x^2-y^2}}+\frac{-2\gamma_1 \xi}{\Delta_{uxz/yz,xy}}\right] (\vec{\sigma}\times \vec{k})\cdot \hat{z}, \label{eq:ham_eff_gamma_uxz/yz}\end{aligned}$$ where $\Delta_{uxz/yz,x^2-y^2}=2(t_{\pi}+t_{\delta})+\frac{\xi}{2}-\{3t_{\sigma}+t_{\delta}+\delta_3\}$, $\Delta_{uxz/yz,xy}=2(t_{\pi}+t_{\delta})+\frac{\xi}{2}-\{4t_{\pi}+\delta_1\}$, and the Pauli matrices describe the subspace defined by $\{ \frac{1}{\sqrt{2}}(\|d_{xz}\downarrow\rangle -i\|d_{yz}\downarrow\rangle), \frac{1}{\sqrt{2}}(\|d_{xz}\uparrow\rangle +i\|d_{yz}\uparrow\rangle) \}$. The angular momentum (AM) texture can be calculated using the eigenstates with the lowest perturbative correction in $\xi$. For the $d_{xy}$ manifold in close vicinity of the $\Gamma$ point, the dominant spin AM expectation value $\langle S_y \rangle \approx \hbar/2$ for an eigenstate in $x$ direction comes from the original $d_{xy}$ manifold. The remnant orbital AM $\langle L_y \rangle \approx\hbar \xi/\Delta_{xy,uxz/yz}$ is due to the inter-band coupling to the upper $d_{xz/yz}$, which can be calculated using the eigenstate with the first-order correction in $\xi$ that hybridizes the $d_{xy}$ manifold with the upper $d_{xz/yz}$ and the $d_{x^2-y^2}$ manifolds. As for the lower $d_{xz/yz}$ manifold, both the orbital and spin AM expectation values $$\begin{aligned} \langle L_y \rangle &\approx& \frac{-3 \hbar \xi}{\Delta_{lxz/yz,z^2}} \\ \langle S_y \rangle &\approx& -\frac{3}{4} \hbar \left( \frac{\xi}{\Delta_{lxz/yz,z^2}} \right)^2. \label{eq:ham_R_lxz/yz_correction_lysy} \end{aligned}$$ for an eigenstate in $x$ direction can be obtained only from the eigenstates with the first-order correction in $\xi$ which leads to hybridization with the $d_{z^2}$ manifold. Thus, the orbital dominant AM texture in the lower $d_{xz/yz}$ bands comes from the inter-band coupling to the $d_{z^2}$. Similarly, we find that the orbital dominant AM texture in the upper $d_{xz/yz}$ bands originates from the inter-band coupling to the $d_{xy}$ and the $d_{x^2-y^2}$ manifolds. Near the $\mathrm{X}=(\pi,0)$ point, the effective Hamiltonian is $$\begin{aligned} \mathcal{H}_{\mathrm{eff}} &=&h_{yz}(\vec{k})I_{2\times 2} +\left[\frac{-2\sqrt{3}\gamma_2 \xi}{\tilde{\Delta}_{yz,z^2}}+\frac{2\gamma_3 \xi}{\tilde{\Delta}_{yz,x^2-y^2}}\right] \sigma_x k_y -\frac{2\gamma_1 \xi}{\tilde{\Delta}_{yz,xy}}\sigma_y k_x, \label{eq:ham_eff_X_yz}\end{aligned}$$ where $\tilde{\Delta}_{yz,z^2}=2(t_{\pi}-t_{\delta})-\delta_2$, $\tilde{\Delta}_{yz,x^2-y^2}=2(t_{\pi}-t_{\delta})-\delta_3$, $\tilde{\Delta}_{yz,xy}=2(t_{\pi}-t_{\delta})-\delta_1$, and the Pauli matrices describe the subspace defined by $\{ \| d_{yz}\uparrow\rangle, \| d_{yz}\downarrow\rangle \}$, and $(k_x, k_y)$ is a local coordinate with respect to $(\pi,0)$. Here, the splitting terms are mixture of Rashba and linear Dresselhaus terms, which are of the form $$\begin{aligned} \mathcal{H}_{\mathrm{splitting}}=A\sigma_x k_y -B\sigma_y k_x,\end{aligned}$$ with $A=\frac{-2\sqrt{3}\gamma_2 \xi}{\tilde{\Delta}_{yz,z^2}}+\frac{2\gamma_3 \xi}{\tilde{\Delta}_{yz,x^2-y^2}}$ and $B=\frac{2\gamma_1 \xi}{\tilde{\Delta}_{yz,xy}}$. If we rotate the local coordinate by $\pi/4$ about $k_z$ axis, the splitting terms become $$\begin{aligned} \mathcal{H}_{\mathrm{splitting}}&=&\frac{A+B}{2}(\sigma_x k_y -\sigma_y k_x) +\frac{A-B}{2}(\sigma_x k_x -\sigma_y k_y) \\ &=&\alpha_R(\sigma_x k_y -\sigma_y k_x) +\alpha_D(\sigma_x k_x -\sigma_y k_y).\end{aligned}$$ In our case, we have $\|A\|\gg\|B\|$, thus both Rashba and linear Dresselhaus terms are present with similar strength. Due to the symmetry, only Rashba term is allowed for $C_{4v}$ at $\mathrm{\Gamma}$ (where we should have $A=B$), and both Rashba and linear Dresselhaus terms are allowed for $C_{2v}$ at $\mathrm{X}$ [-@Stroppa2014]. ![\[fig:tb\_am\] Angular momentum texture from the tight-binding model. The angular momentum textures are calculated (a) at $E_{vH}$ and (b) near $\mathrm{X}$ using both $t_{2g}$ and $e_g$, and (c) at $E_{vH}$ and (d) near $\mathrm{X}$ using only $t_{2g}$. The red and blue arrows represent the orbital and spin AM, respectively. ](figS1.pdf){width="90.00000%"} The splitting of the log van Hove singularity at $\mathrm{X}$ ============================================================= We show here that the Rashba-Dresselhaus splitting removes the spin degeneracy of the logarithmic van Hove singularity at $\mathrm{X}$, resulting in the two separate logarithmic van Hove singularities for the upper and lower Rashba-Dresselhaus bands. This implies that there will be a very large density of state change between the upper and lower Rashba-Dresselhaus bands, which would have a significant effect on the phase competition, [e.g.]{} the relative magnitude of the pairing susceptibilities with different symmetries. It is well-known that there is a logarithmic van Hove singularity at $\mathrm{X}$ in absence of the Rashba-Dresselhaus splitting. The dispersion of the lowest energy band near $\mathrm{X}=(\pi,0)$ approximately follows the dispersion of the $d_{yz}$ band, $$\xi = 2(t_\delta \cos k_x + t_\pi \cos k_y) \approx t_\delta (k_x -\pi)^2 - t_\pi k_y^2 + 2(t_\pi - t_\delta);$$ it is well-understood that there is a logarithmic van Hove singularity at the saddle point of a quadratic Hamiltonian in 2D [-@Yu2010]. The addition of the Rashba-Dresselhaus term near $\mathrm{X}$, $\mathcal{H}_{R-D} = A\sigma_x k_y - B \sigma_y (k_x - \pi)$, leads to the spin splitting of this saddle point, which modifies the dispersion to $$\xi_\pm \approx t_\delta (k_x -\pi)^2 - t_\pi k_y^2 \pm \sqrt{A^2 k_y^2 + B^2 (k_x -\pi)^2} + 2(t_\pi - t_\delta).$$ Using the fact that the Fermi velocity vanishes when the van Hove singularity occurs, we can see that the van Hove singularity at $\mathrm{X} = (\pi, 0)$ is shifted to $(\pi \pm B/2t_\delta, 0)$ for the upper Rashba-Dresselhaus band, with the dispersion in its vicinity $$\xi_+ \approx t_\delta \left(k_x - \pi \mp \frac{B}{2t_\delta}\right)^2 - \left(t_\pi + t_\delta \frac{A^2}{B^2}\right) k_y^2 + 2(t_\pi - t_\delta) - \frac{B^2}{4t_\delta}$$ and $(\pi, \pm A/2t_\pi)$ for the lower Rashba-Dresselhaus band, with the dispersion in its vicinity $$\xi_- \approx \left(t_\delta + t_\pi \frac{B^2}{A^2}\right)(k_x - \pi)^2 - t_\pi (k_y \mp \frac{A}{2t_\pi})^2 + 2(t_\pi - t_\delta) + \frac{A^2}{4t_\pi}.$$ We see here that when we raise the chemical potential so that the Fermi surface passes through the ${\rm X}$ point, the Fermi level first passes through the logarithmic van Hove singularity of the lower Rashba-Dresselhaus band, and then that of the upper Rashba-Dresselhaus band. The importance of $e_g$ manifold in the angular momentum texture ================================================================ Because the $e_g$ manifold affects the Rashba-Dresselhaus splitting, the inclusion of the $e_g$ manifold is important to correctly describe the AM texture. By numerically solving the tight-binding model, we obtained the AM expectation values with and without $e_g$ manifold (Fig. \[fig:tb\_am\]). For the $t_{2g}$-only limit, we set $\delta_2 \approx \delta_3 \approx 10^3 \mathrm{eV}$. We find considerable differences in view of the direction and magnitude of the AM. Notably, the coupling to $e_g$ manifold has significant effects in the direction of the AM near $\mathrm{X}$ and in the intermediate region. The effect of compressive strain in the substrate ================================================= Due to the large lattice constant of BaHfO$_3$, it might be helpful to apply compressive strain to the substrate for the deposition of the tantalate thin film. The electronic band structure of TaO$_2$/KO layer on HfO$_2$-terminated BaHfO$_3$ with the lattice constant reduced by 2% is presented in Fig. \[fig:strain\_band\]. We find that the Rashba coefficient remains still large (for example, $\alpha_R \approx 0.3~\mathrm{eV\AA}$ in the lower $d_{xz/yz}$ bands at $\Gamma$). The relation between the band effective mass and Rashba-related parameters ========================================================================== Here, we show that both the momentum offset $k_R$ and the Rashba energy $E_R$ are proportional to the effective mass of the Rashba bands for a given Rashba strength $\alpha_R$. We consider the Hamiltonian $$\begin{aligned} \mathcal{H} =\frac{\hbar^2 k^2}{2m^} I_{2\times2}+\alpha_R (\vec{\sigma}\times \vec{k})\cdot \hat{z},\end{aligned}$$ with $k=\sqrt{k_x^2+k_y^2}$, where $m^$ is the effective mass of the band and $I_{2\times2}$ is the $2\times2$ identity matrix. The energy dispersion of the lower Rashba band is given by $$\begin{aligned} E(k) &=&\frac{\hbar^2 k^2}{2m^} - \|\alpha_R\| k \\ &=&\frac{\hbar^2}{2m^} (k-\frac{m^* \|\alpha_R\|}{\hbar^2})^2 - \frac{m^* \|\alpha_R\|^2}{2\hbar^2} \\ &\equiv&\frac{\hbar^2}{2m^} (k-k_R)^2 - E_R.\end{aligned}$$ We find that the momentum offset $k_R = \frac{m^ \|\alpha_R\|}{\hbar^2}$ and the Rashba energy $E_R = \frac{m^* \|\alpha_R\|^2}{2\hbar^2}$, which are principal measures of the band splitting size when one sees a band structure figure, are proportional to the effective mass $m^*$ for a given Rashba parameter $\alpha_R$. Thus, the Rashba splitting of the $d_{xz/yz}$ bands would look more pronounced due to the heavier effective mass compared with the $d_{xy}$ band even if they had the same Rashba strength. ![\[fig:strain\_band\] Band structure of TaO$_2$/KO on HfO$_2$-terminated BaHfO$_3$ with the lattice constant reduced by 2%. ](figS2.pdf){width="45.00000%"}	{ "pile_set_name": "ArXiv" }
--- abstract: 'Let $f$ be an analytic function mapping the unit disk ${{\mathbb D}}$ to itself. We give necessary and sufficient conditions on the local behavior of $f$ near a finite set of boundary points that requires $f$ to be a finite Blaschke product.' address: \| Department of Mathematics\ College William and Mary\ Williamsburg, Virginia 23187-8795, U. S. A. author: - Vladimir Bolotnikov title: A uniqueness result on boundary interpolation --- \[theorem\][Definition]{} \[theorem\][Proposition]{} \[theorem\][Corollary]{} \[theorem\][Remark]{} \[theorem\][Lemma]{} \[theorem\][Problem]{} \[theorem\][Example]{} \[section\] Introduction ============ The following boundary uniqueness result was presented in [@bk] as an intermediate step to obtain a similar result in the multivariable setting of the unit ball. Let $f\in{{\mathcal S}}$ and let $f(z)=z+O((z-1)^4)$ as $z\to 1$. Then $f(z)\equiv z$. \[T:1.1\] Here and in what follows, ${{\mathcal S}}$ denotes the Schur class of functions analytic and bounded by one in modulus on the unit disk ${{\mathbb D}}$. In [@ch], Theorem \[T:1.1\] was generalized in the following way. Let $f\in{{\mathcal S}}$ and let $b$ be a finite Blaschke product. Let $\tau$ be a unimodular number and let $A_{b,\tau}=b^{-1}(\tau)=\{t_1,\ldots,t_d\}$ (since $b$ is a finite Blaschke product, $A_{b,\tau}$ is a finite subset of the unit circle ${\mathbb T}$). If 1. $f(z)=b(z)+O((z-t_1)^4)$ as $z\to t_1$ and 2. $f(z)=b(z)+O((z-t_i)^{\ell_i})$ for some $\ell_i\ge 2$ as $z\to t_i$ for $i=2,\ldots,d$, then $f(z)\equiv b(z)$ on ${{\mathbb D}}$. \[T:1.2\] Thus, conditions in Theorem \[T:1.2\] are sufficient to guarantee $f(z)\equiv b(z)$. The question raised in [@ch] was to find necessary (in a sense) and sufficient conditions. The answer is given below. For a given real $x$, $[x]$ denotes the largest integer that does not exceed $x$. Let $f\in{{\mathcal S}}$ and let $b$ be a finite Blaschke product of degree $d$. Let $t_1,\ldots,t_n$ be points on ${\mathbb T}$ and let $$f(z)=b(z)+o((z-t_i)^{m_i})\quad\mbox{for}\quad i=1,\ldots,n \label{1.1}$$ as $z$ tends to $t_i$ nontangentially and where $m_1,\ldots,m_n$ are positive integers. If $$\left[\frac{m_1+1}{2}\right]+\ldots +\left[\frac{m_n+1}{2}\right]>d= \deg \, b, \label{1.2}$$ then $f(z)\equiv b(z)$ on ${{\mathbb D}}$. Otherwise, the uniqueness result fails. \[T:1.3\] In other words, the points $t_i\in{\mathbb T}$ can be chosen arbitrarily (regardless $b$) as well as degrees of convergence. To derive Theorem \[T:1.2\] from Theorem \[T:1.3\], note that if $\deg \, b=d$, the set $A_{b,\tau}$ consists of exactly $d$ points on ${\mathbb T}$. The assumptions in Theorem \[T:1.2\] mean that holds for $m_1=3$ and $m_i\ge 1$ for $i=2,\ldots,d$. Then the sum on the left hand side in is not less than $2+(d-1)=d+1$ which is greater than $d$ and the result follows. The proof of Theorem \[T:1.3\] is given in Section 4. It relies on some recent results on boundary interpolation [@bk] that we recall in Section 2 and Section 3. Boundary Schwarz-Pick matrices ============================== Let $w$ be a Schur function. Then for every choice of $n\in{\mathbb N}$ and of $n$-tuples ${\bf z}=(z_1,\ldots,z_n)\in{{\mathbb D}}^n$ and ${\bf k}=(k_1,\ldots,k_n)\in{\mathbb N}^n$, the [Schwarz-Pick matrix]{} $P^w_{{\bf k}}({\bf z})$ defined as $$P^w_{{\bf k}}({\bf z})=\left[P^w_{k_i,k_j}(z_i,z_j)\right]_{i,j=1}^n \label{2.1}$$ where $$P^w_{k_i,k_j}(z_i,z_j)=\left[\left.\frac{1}{\ell !r!} \, \frac{\partial^{\ell+r}}{\partial z^\ell\partial\bar{\zeta}^r} \, \frac{1-w(z)\overline{w(\zeta)}}{1-z\bar{\zeta}} \right\vert_{{\scriptsize\begin{array}{c} z=z_i\\ \zeta=\overline{z}_j\end{array}}} \right]_{\ell=0,\ldots,k_i-1}^{r=0,\ldots,k_j-1}, \label{2.2}$$ is positive semidefinite. Indeed, every Schur function $w$ admits a de Branges–Rovnyak realization $$w(z)=w(0)+zC(I_{{{\mathcal H}}}-zA)^{-1}B\quad (z\in{{\mathbb D}}), \label{2.3}$$ (see [@dbr]) with an operator $A$ acting on an auxiliary Hilbert space ${{\mathcal H}}$ and operators $B: \, {{\mathbb C}}\to {{\mathcal H}}$ and $C: \, {{\mathcal H}}\to{{\mathbb C}}$ such that the block operator ${\bf U}=\begin{bmatrix} A & B \\ C & w(0)\end{bmatrix}$ is a coisometry on ${{\mathcal H}}\oplus {{\mathbb C}}$ (if ${\bf U}$ is unitary, representation is called a [unitary realization]{} of $w$). A consequence of equality ${\bf U}{\bf U}^=I$ is that $$\frac{1-w(z)\overline{w(\zeta)}}{1-z\bar{\zeta}}= C(I-zA)^{-1} (I-\bar{\zeta}A^)^{-1}C^.$$ Differentiating both parts in the latter identity gives $$\frac{1}{\ell!r!} \, \frac{\partial^{\ell+r}}{\partial z^\ell\partial\bar{\zeta}^r} \, \frac{1-w(z)\overline{w(\zeta)}}{1-z\bar{\zeta}} =CA^\ell(I-zA)^{-\ell-1}(I-\bar{\zeta}A^)^{-r-1}A^{r}C^$$ which allows us to represent the matrix in as $$P^w_{{\bf k}}({\bf z})=R_{{\bf k}}({\bf z})R_{{\bf k}}({\bf z})^, \label{2.4}$$ where $$R_{\bf k}({\bf z})=\left[\begin{array}{c}R_{k_1}(z_1)\\ \vdots \\ R_{k_n}(z_n)\end{array}\right]\quad\mbox{and}\quad R_{k_i}(z_i)=\left[\begin{array}{c}C(I-z_iA)^{-1} \\ CA(I-z_iA)^{-2}\\ \vdots \\ CA^{k_i-1}(I-z_iA)^{-k_i}\end{array}\right], \label{2.5}$$ and to conclude that $P^w_{{\bf k}}({\bf z})\ge 0$. In case when $w$ is a finite Blaschke product, the above realization arguments are more informative. In what follows, we will write ${\mathcal{BF}}$ for the class of all finite Blaschke products and more specifically, ${\mathcal{BF}}_d$ for the set of all Blaschke products of degree $d$. The symbol ${\rm Dom}(w)$ will stand for the domain of holomorphy of $w$. Apperently, the next result is well known. Let $w\in{\mathcal{BF}}_d$ and let ${\bf k}=(k_1,\ldots,k_n)\in{\mathbb N}^n$. Then 1. The function $P^w_{{\bf k}}({\bf z})$ defined on ${{\mathbb D}}^n$ by formulas and , can be extended continuously to $({\rm Dom} (w))^n$. 2. For every ${\bf z}\in({\rm Dom} (w))^n$, the matrix $P^w_{{\bf k}}({\bf z})$ is positive semidefinite and ${\rm rank} \, P^w_{{\bf k}}({\bf z})=\min \{\|{\bf k}\|, \, d\}$ where we have set $\|{\bf k}\|:=k_1+\ldots+k_n$. \[L:2.1\] [Proof:]{} Since $w$ is a rational function of McMillan degree $d$, it admits ([@BGR Chapter 4]) a minimal realization $$w(z)=w(0)+zC(I_d-zA)^{-1}B\quad (z\in{\rm Dom} (w)), \label{2.6}$$ holding for all $z\in{\rm Dom} (w)$, with ${{\mathcal H}}={{\mathbb C}}^d$ and matrices $A\in{{\mathbb C}}^{d\times d}$, $B\in{{\mathbb C}}^{d\times 1}$, $C\in{{\mathbb C}}^{1\times d}$ such that $$\bigcap_{j=0}^{d-1}{\rm Ker} \, CA^j=\{0\}\quad\mbox{and}\quad \det \, (I-zA)\neq 0 \; \; (z\in{\rm Dom} (w)). \label{2.7}$$ Furthermore, if $w$ inner, then the matrices $A$, $B$ and $C$ can be chosen so that the minimal realization will be unitary [@brkl]. The same result comes out of the de Branges–Rovnyak model: if $w$ is inner, the de Branges–Rovnyak realization is unitary (not just coisometric) with the state space ${{\mathcal H}}=H^2\ominus wH^2$; if $w\in {\mathcal{BF}}_d$, then $\dim\, {{\mathcal H}}=d$ and is obtained upon identifying ${{\mathcal H}}$ with ${{\mathbb C}}^d$. Since realization is unitary, formulas and hold. By , $R_{\bf k}({\bf z})$ is analytic on (more precisely, can be extended analytically to) $({\rm Dom} (w))^n$ and then formula gives the desired extension of $P^w_{{\bf k}}({\bf z})$ to the all of $({\rm Dom}(w))^n$. By , $R_{{\bf k}}({\bf z})\in{{\mathbb C}}^{\|{\bf k}\|\times d}$, and therefore we have from $$P^w_{{\bf k}}({\bf z})\ge 0\quad\mbox{and}\quad {\rm rank} \, P^w_{{\bf k}}({\bf z})\le \min \{\|{\bf k}\|, \, d\}. \label{2.8}$$ On the other hand, if $\|{\bf k}\|=d$, the square matrix $R_{\bf k}({\bf z})$ is not singular. Indeed, assuming that it is singular, we take a nonzero vector $x\in{{\mathbb C}}^d$ such that $$R_{\bf k}({\bf z})\prod_{j=1}^n (I-z_jA)^{k_j}x=0.$$ By , the latter matrix equation reduces to the system of $d=\|{\bf k}\|$ equalities $$CA^{\ell}(I-z_iA)^{k_i-\ell-1}\prod_{j\neq i} (I-z_jA)^{k_j}x=0$$ for $\ell=0,\ldots,k_i-1$ and $i=1,\ldots,n$. Expanding polynomials leads to a homogeneous liner system (with respect to $Cx$, $CAx$,…$CA^{d-1}x$) with the nonzero Vandermonde-like determinant from which it follows that $GA^{\ell}x=0$ for $\ell=0,\ldots,d-1$. Then $x=0$, by the first relation in , and thus, $\det \, R_{\bf k}({\bf z})\neq 0$. By , $P^w_{{\bf k}}({\bf z})>0$ whenever ${\bf z}\in ({\rm Dom} (w))^n$ and $\|{\bf k}\|=d$. Finally if ${\bf k}=(k_1,\ldots,k_n)$ is any tuple with $\|{\bf k}\|=\widetilde{d}<d$, let $\widetilde{\bf k}=(k_1,\ldots, k_{n-1},k_n+d-\widetilde{d})$ so that $\|\widetilde{\bf k}\|=d$. Since $P^w_{{\bf k}}({\bf z})$ is the top $\widetilde{d}\times\widetilde{d}$ principal submatrix in $P^w_{\widetilde{\bf k}}({\bf z})$ and since the latter matrix is positive definite by the preceding analysis, we have $$P^w_{{\bf k}}({\bf z})>0\quad\mbox{whenever} \; \; {\bf z}\in ({\rm Dom}(w))^n \; \; \mbox{and}\; \; \|{\bf k}\|<d. \label{2.9}$$ Combining and gives the second assertion of the lemma and completes the proof. Given $w\in{\mathcal{BF}}$ and a “boundary” tuple ${\bf t}=(t_1,\ldots,t_n)\in{{\mathbb T}}^n$, Lemma \[2.1\] allows us to define the [boundary Schwarz-Pick matrix]{} $P^w_{{\bf k}}({\bf t})= R_{{\bf k}}({\bf t})R_{{\bf k}}({\bf t})^$ via factorization formula for every ${\bf k}\in{\mathbb N}^n$. However, we are more interested in boundary Schwarz-Pick matrices for more general Schur functions. The following definition looks appropriate: Given $w\in{{\mathcal S}}$, ${\bf k}=(k_1,\ldots,k_n)\in{\mathbb N}^n$ and ${\bf t}=(t_1,\ldots,t_n)\in{{\mathbb T}}^n$, the boundary Schwarz-Pick matrix is defined by $$P^w_{{\bf k}}({\bf t})=\lim_{{\bf z}\to{\bf t}}P^w_{{\bf k}}({\bf z}) \label{2.10}$$ as $z_i\in{{\mathbb D}}$ tends to $t_i$ nontangentially for $i=1,\ldots,n$, provided the limit in exists. \[D:2.2\] Here and in what follows, “the limit exists” means also that it is finite. By and , $P^w_{{\bf k}}({\bf t})$ is of the form $$P^w_{{\bf k}}({\bf t})=\left[P^w_{k_i,k_j}(t_i,t_j)\right]_{i,j=1}^n \label{2.11}$$ where $$P^w_{k_i,k_j}(t_i,t_j)=\lim_{{\scriptsize\begin{array}{c}z\to t_i\\ \zeta\to t_j\end{array}}}\left[\frac{1}{\ell !r!} \, \frac{\partial^{\ell+r}} {\partial z^\ell\partial\bar{\zeta}^r} \, \frac{1-w(z)\overline{w(\zeta)}}{1-z\bar{\zeta}} \right]_{\ell=0,\ldots,k_i-1}^{r=0,\ldots,k_j-1}. \label{2.12}$$ A necessary and sufficient condition for the limits to exist is that $$\liminf_{z\to t_i} \frac{\partial^{2k_i-2}}{\partial z^{k_i-1}\partial\bar{z}^{k_i-1}} \, \frac{1-\|w(z)\|^2}{1-\|z\|^2}<\infty \quad\mbox{for}\; \; i=1,\ldots,n, \label{2.13}$$ where $z\in{{\mathbb D}}$ tends to $t_i$ arbitrarily (not necessarily nontangentially). Necessity is self-evident since the bottom diagonal entries in the diagonal blocks $P^w_{k_i,k_i}(t_i,t_i)$ are the nontangential (angular) limits $$\lim_{z,\zeta\to t_i}\frac{1}{((k_i-1)!)^2} \, \frac{\partial^{2k_i-2}} {\partial z^{k_i-1}\partial\bar{\zeta}^{k_i-1}} \, \frac{1-w(z)\overline{w(\zeta)}}{1-z\bar{\zeta}}$$ and their existence clearly implies . The sufficiency part was proved in [@bk] along with some other important consequences of conditions that are recalled in the following theorem. Let $t_1,\ldots,t_n\in{\mathbb T}$, $k_1,\ldots,k_n\in{\mathbb N}$, $w\in{{\mathcal S}}$ and let us assume that conditions are met. Then 1. The following nontangential boundary limits exist $$w_j(t_i):=\lim_{z\to t_i}\frac{w^{(j)}(z)}{j!}\quad\mbox{for} \; \; j=0,\ldots,2k_i-1; \; i=1,\ldots,n. \label{2.14}$$ 2. The nontangential boundary limit exists (or equivalently all the limits in ) exist) and can be expressed in terms of the limits as follows: $$P^w_{k_i,k_j}(t_i,t_j)={\bf H}^w_{k_i,k_j}(t_i,t_j) {\bf \Psi}_{k_j}(t_j){\bf T}^w_{k_j}(t_j)^* \label{2.15}$$ where ${\bf \Psi}_{k_j}(t_j)$ is the $k_j\times k_j$ upper triangular matrix with the entries $$\psi_{r\ell}=\left\{\begin{array}{ccl} 0, & \mbox{if} & r>\ell \\ (-1)^\ell{\scriptsize\left(\begin{array}{c} \ell \\ r \end{array}\right)}t_0^{\ell+r+1}, & \mbox{if} & r\leq\ell \end{array}\right.\quad (r,\ell=0,\ldots,k_j-1), \label{2.16}$$ where ${\bf T}^w_{k_j}(t_j)$ is the lower triangular Toeplitz matrix: $${\bf T}^w_{k_j}(t_j)=\left[\begin{array}{cccc} w_0(t_j) & 0 & \ldots & 0\\ w_1(t_j) & w_0(t_j) & \ddots&\vdots \\ \vdots &\ddots&\ddots& 0\\ w_{k_j-1}(t_j) & \ldots & w_1(t_j) & w_0(t_j)\end{array}\right],$$ and where ${\bf H}^w_{k_i,k_j}(t_i,t_j)$ is defined for $i=j$ as the Hankel matrix $${\bf H}^w_{k_i,k_i}(t_i,t_i)=\left[\begin{array}{cccc} w_1(t_i) & w_2(t_i) & \ldots & w_{k_i}(t_j) \\ w_2(t_i) & w_3(t_i) & \ldots & w_{k_i+1}(t_i) \\ \vdots & \vdots && \vdots \\ w_{k_i}(t_i) & w_{k_i+1}(t_i) & \ldots & w_{2k_i-1}(t_i) \end{array}\right] \label{2.17}$$ and entrywise (if $i\neq j$) by $$\begin{aligned} \left[{\bf H}(t_i,t_j)\right]_{r, \ell} &=&\sum_{\alpha=0}^{r} (-1)^{r-\alpha} \left(\begin{array}{c}\ell+r-\alpha \\ \ell\end{array}\right)\frac{w_{\alpha}(t_i)} {(t_i-t_j)^{\ell+r-\alpha+1}}\nonumber \\ &&-\sum_{\beta=0}^{\ell} (-1)^{r}\left(\begin{array}{c}\ell+r-\beta \\ r\end{array}\right)\frac{w_{\beta}(t_j)}{(t_i-t_j)^{\ell+r-\beta+1}} \label{2.18}\end{aligned}$$ for $r=0,\ldots,k_i-1$ and $\ell=0,\ldots,k_j-1$. $$(3) \; \mbox{It holds that} \; \; \|w_0(t_i)\|=1\quad(i=1,\ldots,n)\quad\mbox{and}\quad P^w_{{\bf k}}({\bf t})\ge 0.\qquad \qquad\quad \label{2.19}$$ \[T:2.3\] [Once the two first statements in Theorem \[T:2.3\] are proved, the third statement is immediate. Inequality $P^w_{{\bf k}}({\bf t})\ge 0$ follows from and the fact that $P^w_{{\bf k}}({\bf z})\ge 0$ for every $z\in{{\mathbb D}}$. Furthermore, existence of the limits implies in particular that the nontangential boundary limits ${\displaystyle\lim_{z\to t_i}\frac{1-\|w(z)\|^2}{1-\|z\|^2}}$ exist for $i=1,\ldots,n$ (and are finite) which together with existence of the nontangential limits $w_0(t_i)$ in implies that $\|w_0(t_i)\|=1$]{}. \[R:2.4\] [In case $n=1$ and $k_1=1$, Theorem \[T:2.3\] reduces to the classical Carathéodory-Julia theorem [@cara] on angular derivatives.]{} \[R:2.5\] [In [@Kov], I. Kovalishina considered the single point case ($n=1$ and $k_1>1$) under an additional assumption that $w$ satisfies the symmetry relation $w(z)\overline{w(1/\bar{z})}\equiv 1$ in some neighborhood of $t_1$. A remarkable “Hankel-${\bf \Psi}$-Toeplitz” structure of $P^w_{k_1,k_1}(t_1,t_1)$ has been observed there]{}. \[R:2.6\] Carathéodory-Julia type conditions are worth a formal definition. Given $n$-tuples ${\bf t}=(t_1,\ldots,t_n)\in{{\mathbb T}}^n$ and ${\bf k}=(k_1,\ldots,k_n)\in{\mathbb N}^n$, a Schur function $w$ is said to belong to the class ${{\mathcal S}}_{\bf k}({\bf t})$ if it meets conditions . \[D:2.7\] Statement (1) in Theorem \[T:2.3\] shows that the definition of the boundary Schwarz-Pick matrix $P^w_{{\bf k}}({\bf t})$ makes sense if and only if $w\in{{\mathcal S}}_{\bf k}({\bf t})$. Statement (2) expresses $P^w_{{\bf k}}({\bf t})$ in terms of boundary limits of $w$ and of its derivatives. An interesting point in is that the block matrix $P^w_{{\bf k}}({\bf t})$ of the form constructed via structured formulas – (rather than by the limits ) does not look like a Hermitian matrix and nevertheless, it turns out to be Hermitian (and even positive semidefinite) due to conditions . The next theorem (see [@bk] for the proof) shows that relations are characteristic for the class ${{\mathcal S}}_{\bf k}({\bf t})$. Let $w$ be a Schur function, let ${\bf t}\in{{\mathbb T}}^n$, ${\bf k}\in{\mathbb N}^n$ and let us assume that the nontangential limits exist and satisfy conditions where $P^w_{{\bf k}}({\bf t})$ is the matrix constructed from the limits via formulas –. Then $w\in{{\mathcal S}}_{\bf k}({\bf t})$. \[T:2.7\] From the computational point of view, it is much easier to construct the boundary Schwarz-Pick matrix $P^w_{{\bf k}}({\bf t})$ via formulas –, than by (for example, if $w$ is a rational function, the boundary limits $w_i(t_j)$ are just the Taylor coefficients from the expansion of $w$ around $t_i$). However, as follows from Theorems \[T:2.3\] and \[T:2.7\], the matrix constructed in this way will be indeed the boundary Schwarz-Pick matrix if and only if conditions are satisfied. Boundary interpolation for classes ${{\mathcal S}}_{\bf k}({\bf t})$ ==================================================================== The following interpolation problem has been studied in [@bk]. Given ${\bf t}=(t_1,\ldots,t_n)\in{{\mathbb T}}^n$, ${\bf k}=(k_1,\ldots,k_n)\in{\mathbb N}^n$ and numbers $b_{ij}\in{{\mathbb C}}$ $(j=0,\ldots,k_i-1; \; i=1,\ldots,n)$, find all functions $f\in{{\mathcal S}}_{\bf k}({\bf t})$ such that $$f_j(t_i):=\lim_{z\to t_i}\frac{f^{(j)}(z)}{j!}=b_{ij}\quad (j=0,\ldots,2k_i-1; \; i=1,\ldots,n). \label{3.1}$$ where all the limits are nontangential. \[P:3.1\] This interpolation problem makes perfect sense: if $f$ belongs to ${{\mathcal S}}_{\bf k}({\bf t})$, then the nontangential limits in exist by Theorem \[T:2.3\]; we just want them to be equal to the preassigned numbers. Let define the $\|{\bf k}\|\times \|{\bf k}\|$ matrix $P$ (the [Pick matrix]{} of the problem) by formulas similar to –, but with $w_j(t_i)$ replaced by $b_{ij}$: $$P=\left[P_{ij}\right]_{i,j=1}^n\quad\mbox{with}\quad P_{ij} =H_{ij}\cdot {\bf \Psi}_{k_j}(t_j)\cdot T_j^, \label{3.3}$$ where ${\bf \Psi}_{k_j}(t_j)$ is the upper triangular matrix with the entries given in , where $T_i$ is the lower triangular Toeplitz matrix and $H_{ii}$ is the Hankel matrix defined by $$T_i=\left[\begin{array}{ccc} b_{i,0} & & 0 \\ \vdots & \ddots & \\ b_{i,k_j-1} & \ldots & b_{i,0}\end{array}\right],\quad H_{ii}=\left[\begin{array}{ccc} b_{i,1} & \cdots & b_{i,k_i}\\ \vdots & & \vdots\\ b_{i,k_i} & \cdots & b_{i,2k_i-1}\end{array}\right] \label{3.4}$$ for $i=1,\ldots,n$ and where the matrices $H_{ij}$ (for $i\neq j$) are defined entrywise by $$\begin{aligned} \left[H_{ij}\right]_{r, \ell} &=&\sum_{\alpha=0}^{r} (-1)^{r-\alpha} \left(\begin{array}{c}\ell+r-\alpha \\ \ell\end{array}\right)\frac{b_{i,\alpha}} {(t_i-t_j)^{\ell+r-\alpha+1}}\nonumber \\ &&-\sum_{\beta=0}^{\ell} (-1)^{r}\left(\begin{array}{c}\ell+r-\beta \\ r\end{array}\right)\frac{b_{j,\beta}}{(t_i-t_j)^{\ell+r-\beta+1}}. \label{3.2}\end{aligned}$$ The purpose of this construction is clear: the matrix $P$ constructed above depends on the interpolation data only; on the other hand, for every solution $f$ of Problem \[P:3.1\], the boundary Schwarz-Pick matrix $P^f_{{\bf k}}({\bf t})$ must be equal to $P$, by the very construction. Let $P$ be the matrix defined in . Then 1. If Problem \[P:3.1\] has a solution, then $$\|b_{i,0}\|=1\quad(i=1,\ldots,n)\quad\mbox{and}\quad P\ge 0. \label{3.5}$$ 2. If holds and $P>0$, then Problem \[P:3.1\] has infinitely many solutions. 3. If holds and $P$ is singular, then Problem \[P:3.1\] has at most one solution. 4. If holds and $f$ is a Schur function satisfying conditions , then necessarily $f\in{{\mathcal S}}_{\bf k}({\bf t})$. \[T:3.2\] The first statement follows from Statement (3) in Theorem \[T:2.3\], since $b_{i,0}=f_0(t_i)$ and $P^f_{{\bf k}}({\bf t})=P$ for every solution $f$ of Problem \[P:3.1\]. The last statement follows from Theorem \[T:2.7\]. The second statement is proved in [@bk] where moreover, a linear fractional parametrization of all solutions of Problem \[P:3.1\] (in case $P>0$) is given. The third statement also was proved in [@bk]. The proof of Theorem \[T:1.3\] will rest on Theorem \[T:3.2\] and the following simple observation. Let $\widetilde{P}=[p_{ij}]$ be an $r\times r$ Hermitian matrix and let us assume that its principal submatrix $P=[p_{i_\alpha,i_\beta}]_{\alpha,\beta=1}^{\ell}$ is positive definite. Then $\widetilde{P}$ can be turned into a positive definite matrix upon an appropriate modification of the $r-\ell$ diagonal entries $p_{ii}$ for $i\not\in\{i_1,\ldots,i_{\ell}\}$ (we will call these entries the diagonal entries of $\widetilde{P}$ complementary to the principal submatrix $P$). \[P:3.3\] [Proof:]{} Without loss of generality we can assume that $P$ is the leading principal submatrix of $\widetilde{P}$ so that $\widetilde{P}=\left[\begin{array}{cc} P & R^ \\ R & D\end{array}\right]$. Let us modify the diagonal entries in $D$ as follows: $$\widetilde{P}^\prime=\left[\begin{array}{cc} P & R^* \\ R & D^\prime\end{array}\right] \quad \mbox{where} \; \; D^\prime= D+\rho I_{r-\ell} \; \; (\rho >0).$$ Since $P>0$, the factorization formula $$\left[\begin{array}{cc} P & R^* \\ R & D^\prime\end{array}\right]= \left[\begin{array}{cc} I_\ell & 0 \\ RP^{-1} & I_{r-\ell}\end{array}\right] \left[\begin{array}{cc} P & 0 \\ 0 & D^\prime-RP^{-1}R^\end{array}\right] \left[\begin{array}{cc} I_\ell & P^{-1}R^ \\ 0 & I_{r-\ell}\end{array}\right]$$ shows that $\widetilde{P}^\prime>0$ if and only if $D^\prime-RP^{-1}R^* =\rho I_{r-\ell}+D-RP^{-1}R^>0$ and the latter inequality indeed can be achieved if $\rho$ is large enough. Proof of Theorem \[T:1.3\] ========================== Let us assume for a moment that the Schur function $f$ in is not given and let us consider the following interpolation problem. Given $t_1,\ldots,t_n\in{\mathbb T}$, $m_1,\ldots,m_n\in{\mathbb N}$ and $b\in{\mathcal{BF}}_d$, find all Schur functions $f$ satisfying asymptotic equations . \[P:4.1\] Note that conditions can be reformulated equivalently (see e.g., [@boldym1 Corollary 7.9] for the proof) as follows: [the nontangential boundary limits $f_j(t_i)$ exist and satisfy $$f_j(t_i):=\lim_{z\to t_i}\frac{f^{(j)}(z)}{j!}=\frac{b^{(j)}(t_i)}{j!}=:b_{ij}\quad \mbox{for} \; \; j=0,\ldots,m_i; \; i=1,\ldots,n. \label{3.6}$$]{} Define the integers $k_i:=\left[\frac{m_i+1}{2}\right]$ for $i=1,\ldots,n$ so that $m_i=2k_i-1$ or $m_i=2k_i$. Reindexing if necessary, we can assume without loss of generality that the first $\ell$ integers $m_1,\ldots,m_\ell$ are odd while the remaining ones (if any) are even. Now we split conditions into two parts: $$f_j(t_i)=\frac{b^{(j)}(t_i)}{j!}=:b_{ij}\quad \mbox{for} \; \; j=0,\ldots,2k_i-1; \; i=1,\ldots,n \label{3.7}$$ and $$f_{2k_i}(t_i)=\frac{b^{(2k_i)}(t_i)}{(2k_i)!}=:b_{i,2k_i}\quad \mbox{for} \; \; i=\ell+1,\ldots,n. \label{3.8}$$ First we consider the interpolation problem with interpolation conditions (this problem is “truncated” with respect to Problem \[P:4.1\]). This problem looks like Problem \[P:3.1\]; however, it is more special, since that data $\{b_{ij}\}$ comes from certain $b\in{\mathcal {BF}}_d$. In other words, the Pick matrix $P$ of the problem coincides with the boundary Schwarz-Pick matrix $P^b_{\bf k}({\bf t})$. Then we may conclude by Lemma \[L:2.1\] that $P\ge 0$ and $${\rm rank} \, P=\min \{\|{\bf k}\|, \, d\}. \label{3.9}$$ Thus, the second condition in is met, while the first condition holds since $b_{i,0}=b(t_i)$ and $b\in{\mathcal {BF}}$. Assuming that inequality is in force, i.e., that $$d<\sum_{i=1}^n\left[\frac{m_i+1}{2}\right]=\sum_{i=1}^nk_i=\|{\bf k}\|$$ we conclude from that $P$ is singular and then by Statement (3) in Theorem \[T:3.2\], there is at most one $f\in{{\mathcal S}}$ satisfying conditions . Therefore (since is just part of ), there is at most one $f\in{{\mathcal S}}$ satisfying conditions . A self-evident observation that the Schur function $b$ does satisfy (this information is contained in ) gives the desired uniqueness: there are no functions $f$ in ${{\mathcal S}}$ different from $b$ that satisfy interpolation conditions or, equivalently, asymptotic equalities . Thus, once is in force and $f$ is subject to , we have necessarily $f(z)\equiv b(z)$. This completes the proof of the first statement in Theorem \[T:1.3\]. It remains to show that the uniqueness result fails whenever $\|{\bf k}\|\le d$. In this case we conclude from that the $\|{\bf k}\|\times \|{\bf k}\|$ matrix $P$ is positive definite and then, by Statement (2) in Theorem \[T:3.2\], there are infinitely many Schur functions $f$ satisfying conditions . In case all $m_i$’s are odd, this completes the proof: conditions are identical with and thus, there are infinitely many Schur functions satisfying asymptotic . The general case (when the set of conditions is not empty) requires one step more. Assuming that $\|{\bf k}\|\le d$ so that the Pick matrix $P=P^b_{\bf k}({\bf t})$ corresponding to interpolation problem is positive definite and that $\ell<n$ in , let us attach interpolation conditions $$f_{2k_i+1}(t_i)=\frac{b^{(2k_i+1)}(t_i)}{(2k_i+1)!}=:b_{i,2k_i+1}\quad \mbox{for} \; \; i=\ell+1,\ldots,n \label{3.10}$$ to and let us consider the extended interpolation problem (for Schur class functions) with interpolation conditions , and . The collection of $b_{ij}$’s appearing in and will be called the [original data]{}, the collection $\{b_{i,2k_i+1}\}$ from will be called the [supplementary data]{} whereas their union will be referred to as to the [extended data]{}. For the extended interpolation problem we have an even number of conditions for each interpolating point $t_i$ which allows us to construct the corresponding extended Pick matrix $\widetilde{P}$ via formulas : $$\widetilde{P}=\left[\widetilde{P}_{ij}\right]_{i,j=1}^n \quad\mbox{where}\quad \widetilde{P}_{ij} =\widetilde{H}_{ij}\cdot {\bf \Psi}_{\widetilde{k}_j}(t_j)\cdot \widetilde{T}_j^ \label{3.11}$$ and where $\widetilde{H}_{ij}$ and $\widetilde{T}_j$ are defined by formulas , with $k_i$ replaced by $\widetilde{k}_i$. It is clear that $\widetilde{P}$ coincides with the boundary Schwarz-Pick matrix $P^b_{\widetilde{\bf k}}({\bf t})$ based on the same $b\in{\mathcal {BF}}_d$, the same ${\bf t}=(t_1,\ldots,t_n)\in{{\mathbb T}}^n$ and $$\widetilde{\bf k}=(\widetilde{k}_1,\ldots,\widetilde{k}_n)= (k_1,\ldots,k_\ell,k_{\ell+1}+1,\ldots,k_n+1)\in{\mathbb N}^n.$$ Of course, all the entries in $\widetilde{P}$ are expressed in terms of the extended data. However, it turns out that all its entries but $\ell$ diagonal ones are uniquely determined from the [original data]{}. Indeed, if $i\neq j$, then $\widetilde{H}_{ij}$ and $\widetilde{T}_j$ (and therefore, $\widetilde{P}_{ij}$) are expressed via formulas , in terms the numbers $b_{i,0},\ldots,b_{i,\widetilde{k}_i-1}$ and $b_{j,0},\ldots,b_{j,\widetilde{k}_j-1}$ all of which are contained in the original data, since $\widetilde{k}_i-1\le k_i\le 2k_i-1$. Now we examine the diagonal blocks $\widetilde{P}_{ii}$ for $i>\ell$ (if $i\le \ell$, then $\widetilde{P}_{ii}=P_{ii}$ is completely determined by the original data). By and , $$\widetilde{P}_{ii}=\left[\begin{array}{cccc} b_{i,1} & \cdots & b_{i,k_i}& b_{i,k_i+1}\\ \vdots & & \vdots &\vdots\\ b_{i,k_i} & \ldots & b_{i,2k_i-1}& b_{i,2k_i}\\ b_{i,k_i+1} & \ldots & b_{i,2k_i} & b_{i,2k_i+1}\end{array}\right] {\bf \Psi}_{k_i+1}(t_i)\left[\begin{array}{ccc}\overline{b}_{i,0} & \ldots & \overline{b}_{i,k_i}\\ & \ddots & \vdots \\ 0 & & \overline{b}_{i,0} \end{array}\right]. \label{3.12}$$ It is readily seen from that the only entry in $\widetilde{P}_{ii}$ that depends on the supplementary data is the the bottom diagonal entry $$\gamma_i:=\left[\widetilde{P}_{ii}\right]_{k_i,k_i} =\left[\begin{array}{ccc}b_{i,k_i+1} & \cdots & b_{i,2k_i+1}\end{array}\right]{\bf \Psi}_{k_i+1}(t_i) \left[\begin{array}{ccc} b_{i,k_i}&\cdots & b_{i,0}\end{array}\right]^* \label{3.13}$$ which, on account of , can be written as $$\begin{aligned} \gamma_i&=&(-1)^{k_i}t_i^{2k_i+1}b_{i,2k_i+1}\overline{b}_{i,0}\label{3.14}\\ &&+\sum_{r=0}^{k_i-1}b_{i,k_i+r+1}\sum_{j=k_i+r}^{k_i} (-1)^{j}t_i^{k_i+r+j+1}\left(\begin{array}{c} j \\ k_i+r \end{array}\right)\overline{b}_{i,n_i-j}.\nonumber\end{aligned}$$ Since $\widetilde{P}$ coincides with the boundary Schwarz-Pick matrix $P^b_{\widetilde{\bf k}}({\bf t})$, it is positive semidefinite (by Lemma \[L:2.1\]) and Hermitian, in particular. Furthermore, the Pick matrix $P=P^b_{{\bf k}}({\bf t})$ of the problem is a positive definite principal submatrix of $\widetilde{P}$. The diagonal entries in $\widetilde{P}$ complementary to $P$ are exactly $\gamma_i$’s from , the bottom diagonal entries in the blocks $\widetilde{P}_{ii}$ of $\widetilde{P}$ for $i=\ell+1,\ldots,n$. By Proposition \[P:3.3\], upon replacing $\gamma_i$ in $\widetilde{P}$ by appropriately chosen (sufficiently large) positive numbers $\gamma_i^\prime$ (for $i=\ell+1,\ldots,n$) and keeping all the other entries the same, one gets a positive definite matrix $\widetilde{P}^\prime$. Furthermore, for each chosen $\gamma_i^\prime$, there exists (the unique) $b_{i,2k_i+1}^\prime$ such that $$\begin{aligned} \gamma_i^\prime&=&(-1)^{k_i}t_i^{2k_i+1}b_{i,2k_i+1}^\prime\overline{b}_{i,0} \nonumber\\ &&+\sum_{r=0}^{k_i-1}b_{i,k_i+r+1}\sum_{j=k_i+r}^{k_i} (-1)^{j}t_i^{k_i+r+j+1}\left(\begin{array}{c} j \\ k_i+r \end{array}\right)\overline{b}_{i,n_i-j}\nonumber\end{aligned}$$ (since ${b}_{i,0}\neq 0$, the latter equality can be solved for $b^\prime_{i,2k_i+1}$). Now we replace the supplementary interpolation conditions by $$f_{2k_i+1}(t_i)=b^\prime_{i,2k_i+1}\quad \mbox{for} \; \; i=\ell+1,\ldots,n \label{3.15}$$ where the numbers on the right have nothing to do with the finite Blaschke product $b$ anymore. It is easily seen that the Pick matrix of the modified extended interpolation problem with interpolation conditions , and is $\widetilde{P}^\prime$. Since it is positive definite, there are (by Statement (2) in Theorem \[T:3.2\]) infinitely many Schur functions $f$ satisfying these interpolation conditions. Thus, there are infinitely many Schur functions satisfying , (that is, ) or equivalently, the asymptotic equalities . Thus, the uniqueness conclusion in Theorem \[T:1.3\] fails which completes the proof. [10]{} J.A. Ball, I. Gohberg and L. Rodman, [Interpolation of Rational Matrix Functions]{}, OT45, Birkhäuser-Verlag, Basel-Boston, 1990. V. Bolotnikov and H. Dym, On boundary interpolation for matrix valued Schur functions, Mem. Amer. Math. Soc., to appear. V. Bolotnikov and A. Kheifets, A higher multiplicity analogue of the Carathéodory–Julia theorem and related boundary interpolation, Preprint. V. Bolotnikov and L. Rodman, Krein–Langer factorizations via pole triples, Integral Equations and Operator Theory [47]{} (2003), no. 2, 169–195. L. de Branges and J. Rovnyak, Square summable power series, Holt, Rinehart and Winston, New York, 1966. D. Burns and S. G. Krantz, Rigidity of holomorphic mappings and a new Schwarz lemma at the boundary, J. Amer. Math. Soc. [7]{} (1994), no. 3, 661–676. D. Chelst, A generalized Schwarz lemma at the boundary, Proc. Amer. Math. Soc. [129]{} (2001), no. 11, 3275–3278. C. Carathéodory, , Sitzungber. Preuss. Akad. Wiss. (1929), 39–52. I. V. Kovalishina, multiple boundary interpolation problem for contractive matrix–valued functions in the unit circle, J. Soviet Math. [52(6)]{} (1990), 3467–3481.	{ "pile_set_name": "ArXiv" }
--- abstract: 'Let $ \widetilde{W} = \Lambda \rtimes W_{\circ} $ be an Iwahori-Weyl group of a connected reductive group $ G $ over a non-archimedean local field. The subgroup $ W_{\circ} $ is a finite Weyl group and the subgroup $ \Lambda $ is a finitely-generated abelian group (possibly containing torsion) which acts on a certain real affine space by translations. I prove that if $ w \in \widetilde{W} $ and $ w \notin \Lambda $ then one can apply to $ w $ a sequence of conjugations by simple reflections, each of which is length-preserving, resulting in an element $ w^{\prime} $ for which there exists a simple reflection $ s $ such that $ \ell ( s w^{\prime} ), \ell ( w^{\prime} s ) > \ell ( w^{\prime} ) $ and $ s w^{\prime} s \neq w^{\prime} $. Even for affine Weyl groups, a special case of Iwahori-Weyl groups and also an important subclass of Coxeter groups, this is a new fact about conjugacy classes. Further, there are implications for Iwahori-Hecke algebras $ \mathcal{H} $ of $ G $: one can use this fact to give dimension bounds on the “length-filtration” of the center $ Z ( \mathcal{H} ) $, which can in turn be used to prove that suitable linearly-independent subsets of $ Z ( \mathcal{H} ) $ are a basis.' address: \| University of Wisconsin\ Department of Mathematics\ 480 Lincoln Dr.\ Madison, WI 53706-1325\ United States author: - Sean Rostami title: 'Conjugacy Classes of Non-Translations in Affine Weyl Groups and Applications to Hecke Algebras' --- Introduction ============ A Coxeter group is a pair $ ( W, S ) $ consisting of a group $ W $ and a generating set $ S $ which is presented using the relations $ s^2 = 1 $ for all $ s \in S $ and relations of the form $ ( s t )^{m(s,t)} = 1 $ for some, but not necessarily all, pairs $ s, t \in S $. The most common examples of infinite Coxeter groups are affine Weyl groups, which are groups generated by the reflections of an affine space across special collections of hyperplanes coming from root systems. The theory of Coxeter groups is both complicated and, especially in the case of affine Weyl groups, highly-developed. Affine Weyl groups are ubiquitous in the subject of smooth representations of algebraic groups over non-archimedean local fields due to their connection with Hecke algebras of reductive groups, and many questions about Hecke algebras can be reduced to questions about affine Weyl groups. Let $ F $ be a non-archimedean local field and let $ G $ be a connected reductive affine algebraic $ F $-group. If $ J \subset G(F) $ is a compact-open subgroup and $ ( \rho, V ) $ is a smooth complex representation of $ J $ then the Hecke algebra $ \heckefont{H} ( G ; J, \rho ) $ is the convolution algebra of all compactly-supported functions $ f : G(F) \rightarrow \operatorname{End}_{\C} ( V ) $ satisfying $ f ( \jmath \cdot g \cdot \jmath^{\prime} ) = \rho ( \jmath ) \circ f ( g ) \circ \rho ( \jmath^{\prime} ) $ for all $ g \in G(F) $ and $ \jmath, \jmath^{\prime} \in J $. Many of the simple subcategories in the Bernstein decomposition of the category of smooth representations of $ G(F) $ are equivalent to the category of modules over a Hecke algebra of this form. The center of such a Hecke algebra is important because it consists of the (functorial) $ G(F) $-linear endomorphisms of the representations in the subcategory. The case that $ J $ is an Iwahori subgroup and $ \rho $ is the trivial $ 1 $-dimensional representation yields one particularly important Hecke algebra: the Iwahori-Hecke algebra. In general, an Iwahori subgroup is the group of $ \mathcal{O}_F $-points of a certain connected model $ \mathfrak{G} $ of $ G $ defined in general by Bruhat and Tits, although in nice cases like $ G = \operatorname{GL}_n $ there is a much more straightforward description: an Iwahori subgroup is the inverse image in $ G ( \mathcal{O}_F ) $ of a Borel subgroup in $ G ( \mathbf{k}_F ) $ under the reduction-mod-$ \pi $ map $ \mathcal{O}_F \rightarrow \mathbf{k}_F = \mathcal{O}_F / ( \pi_F ) $. It can be shown that any Iwahori-Hecke algebra $ \heckefont{H} $ has a presentation, called the Iwahori-Matsumoto presentation, consisting of a basis of characteristic functions of the double-cosets $ \mathfrak{G} ( \mathcal{O}_F ) \backslash G ( F ) / \mathfrak{G} ( \mathcal{O}_F ) $ together with a certain pair of relations which depend on some numerical parameters coming from $ G $. It turns out that a group called the Iwahori-Weyl group serves as a system of representatives for these double-cosets. The Iwahori-Weyl group is in general merely a semidirect extension of an affine Weyl group but its behavior is nonetheless extremely similar to that of a true Coxeter group. Taken together with the numerical parameters, the group-theoretic structure of this “quasi-Coxeter group” completely controls the ring-theoretic structure of $ \heckefont{H} $. In this paper, I prove a group-theoretic property of conjugacy classes in Iwahori-Weyl groups, which is described precisely in the next subsection of this introduction. The class of all Iwahori-Weyl groups properly contains the class of all affine Weyl groups (since affine Weyl groups arise as the Iwahori-Weyl groups of semisimple and simply-connected $ G $) and this property is new even in this narrower context. Additionally, there are implications for Iwahori-Hecke algebras $ \heckefont{H} $. For example, this property can be used to show that suitable linearly-independent collections of functions in the center $ Z ( \heckefont{H} ) $ also span the center. The prototypical example of such a collection is the Bernstein basis, although the existence and precise definition of such a basis do not appear in the literature for Iwahori-Hecke algebras of completely general connected reductive groups. The article [@roro] fills this gap, which I explain in more detail next, and also serves as a sample application of the main theorem of this paper. In unpublished work, concerning essentially only the case of split $ G $, Bernstein introduced a particularly important basis for $ Z ( \heckefont{H} ) $ whose elements can be effectively calculated (by computer, if desired) and which simultaneously have a straightforward representation-theoretic interpretation: the basis elements are indexed by $ \finiteweylgroup $-orbits of cocharacters valued in a certain maximal torus $ T \subset G $, and if $ \mathcal{O} = \{ \mu_1, \mu_2, \ldots, \mu_r \} $ is such an orbit then the corresponding basis element $ z_{\mathcal{O}} $ acts on the Iwahori-fixed subspace of the (normalized) induced representation $ \operatorname{ind}( \chi ) $ of an unramified character $ \chi : T ( F ) \rightarrow \C^{\times} $ by the scalar $ \chi ( \mu_1 ) + \chi ( \mu_2 ) + \cdots + \chi ( \mu_r ) $. This work was extended to the affine Hecke algebra on any reduced root datum with any parameter system by Lusztig in [@lusztig]. An affine Hecke algebra on a reduced root datum $ \Psi = ( X, \Phi, X^{\vee}, \Phi^{\vee} ) $ is constructed by using the extended affine Weyl group $ W^{\prime} = X^{\vee} \rtimes \finiteweylgroup ( \Phi ) $ as a vector space basis, choosing parameters, and mimicking the Iwahori-Matsumoto presentation abstractly (see §3.2 of [@lusztig] and §7.1 of [@humphreys] for details). These abstractly defined affine Hecke algebras are useful in the study of reductive groups for the following reason: in some cases, for example if $ G $ is unramified, any Iwahori-Hecke algebra of $ G $ is naturally isomorphic to an affine Hecke algebra for appropriate choice of root datum and parameters. Unfortunately, many Iwahori-Weyl groups are not the extended affine Weyl group of any root datum. A specific example of such a $ G $ is a ramified even-dimensional special orthogonal group (see §1.16 in [@tits] for the definition of such a group): such a group has torsion in the translation subgroup of its Iwahori-Weyl groups, which simply cannot happen in the extended affine Weyl group of a root datum because its translation subgroup $ X^{\vee} $ is free by definition. Consequently, many Iwahori-Hecke algebras are not affine Hecke algebras and so are not within the scope of the Bernstein/Lusztig work. In the article [@roro], I extend the Bernstein/Lusztig results to Iwahori-Hecke algebras of all connected reductive $ F $-groups. After establishing, analogous to [@lusztig], a Bernstein presentation for an Iwahori-Hecke algebra $ \heckefont{H} $ of a general connected reductive $ F $-group, I determine, again following [@lusztig], a Bernstein basis for $ Z ( \heckefont{H} ) $. To prove that these functions indeed span the center, I apply the main theorem of this paper as explained in §\[Sapplicationtoheckealgebras\] below. Statement of results -------------------- More precise definitions of everything here are given in §\[Snotation\] and §\[Siwahoriweylgroups\]. Let $ F $ be a non-archimedean local field and $ G $ a connected reductive affine algebraic $ F $-group. Fix a maximal $ F $-split torus $ A \subset G $ and let $ \widetilde{W} $ be the corresponding Iwahori-Weyl group, which acts on the vector space $ \enlargedapartment \defeq \cochargroup ( A ) \otimes_{\Z} \R $. It is known that $ \widetilde{W} $ contains as a normal subgroup the affine Weyl group $ W_{\operatorname{aff}} ( \scaledrootsystem ) $ of a reduced root system $ \scaledrootsystem $, and that if $ \translationsubgroup \subset \widetilde{W} $ is the subgroup of elements which act on $ \enlargedapartment $ by translations then $ \widetilde{W} $ splits as $ \widetilde{W} = \translationsubgroup \rtimes \finiteweylgroup ( \scaledrootsystem ) $ (here $ \finiteweylgroup $ denotes the finite Weyl group). Further, there are sections for $ \Omega \defeq \widetilde{W} / W_{\operatorname{aff}} ( \scaledrootsystem ) $, so that $ \widetilde{W} $ also splits as $ \widetilde{W} = W_{\operatorname{aff}} ( \scaledrootsystem ) \rtimes \Omega $. If $ \Delta_{\operatorname{aff}} $ is a Coxeter generating set for $ W_{\operatorname{aff}} ( \scaledrootsystem ) $ and $ \ell $ is the corresponding length function, then $ \ell $ extends to $ \widetilde{W} $ by inflation. Denote by $ \basereflections \subset \Delta_{\operatorname{aff}} $ a Coxeter generating set for $ \finiteweylgroup ( \scaledrootsystem ) $. The main result of the paper is the following: Fix $ w \in \widetilde{W} $. If $ w \notin \Lambda $ then there exists $ s \in \Delta_{\operatorname{aff}} $ and (if necessary) $ s_1, \ldots, s_n \in \Delta_{\operatorname{aff}} $ such that, setting $ w^{\prime} \defeq s_n \cdots s_1 w s_1 \cdots s_n $, - $ \ell ( s_i \cdots s_1 w s_1 \cdots s_i ) = \ell ( w ) $ for all $ i $, - both $ \ell ( s w^{\prime} ) > \ell ( w^{\prime} ) $ and $ \ell ( w^{\prime} s ) > \ell ( w^{\prime} ) $, and - $ s w^{\prime} s \neq w^{\prime} $. Note that the last two properties asserted by the Main Theorem can be unified into “$ \ell ( s w^{\prime} s ) > \ell ( w^{\prime} ) $”. A related result appears in the preprint [@HN]: that for any element $ w $ in the extended affine Weyl group of a (reduced) root datum there is a sequence of conjugations by simple reflections, each of which preserves or decreases length, resulting in an element $ w^{\prime} $ which is minimal length in its conjugacy class. Note however that the [@HN] result is also true for finite Weyl groups, originally proven by [@GP], whereas the Main Theorem here is special to infinite groups (length is bounded on a finite group!). Denote by $ \translationsubgroup / \finiteweylgroup $ the set of $ \finiteweylgroup ( \Sigma ) $-conjugacy classes in $ \translationsubgroup $ and recall that $ \ell $ is constant on any $ \mathcal{O} \in \translationsubgroup / \finiteweylgroup $. Denote by $ \Omega ( w ) $ the $ \Omega $-coordinate of any $ w \in \widetilde{W} $. Let $ \heckefont{H} $ be an Iwahori-Hecke algebra for $ G $. By analyzing the equations that define the center $ Z ( \heckefont{H} ) $, the Main Theorem can be used to prove dimension bounds for a certain filtration/partition of $ Z ( \heckefont{H} ) $: If $ Z_{L,\tau} ( \heckefont{H} ) \subset Z ( \heckefont{H} ) $ is the $ \C $-subspace of functions supported only on those $ w $ for which $ \ell ( w ) \leq L $ and $ \Omega ( w ) = \tau $, and if $ N_{L,\tau} $ is the total number of $ \mathcal{O} \in \translationsubgroup / \finiteweylgroup $ such that $ \ell ( \mathcal{O} ) \leq L $ and $ \Omega ( t ) = \tau $ for all $ t \in \mathcal{O} $ then $$\dim_{\C} ( Z_{L,\tau} ( \heckefont{H} ) ) \leq N_{L,\tau}$$ It follows that if $ \{ z_{ \mathcal{O} } \}_{ \mathcal{O} \in \translationsubgroup / \finiteweylgroup } $ is a linearly-independent subset of $ Z ( \heckefont{H} ) $ such that $ z_{ \mathcal{O} } $ is supported only on those $ w $ for which $ \ell ( w ) \leq \ell ( \mathcal{O} ) $ and $ \Omega ( w ) $ is the same for all $ w $ supporting $ z_{ \mathcal{O} } $ then $ \{ z_{ \mathcal{O} } \}_{ \mathcal{O} \in \translationsubgroup / \finiteweylgroup } $ is a basis. Outline of paper ---------------- In §\[Snotation\], I set some notation and define most of the objects that will be used throughout the paper. I use several non-standard but convenient notations, and almost all of them can be found here. Two exceptions are a “quasi-Coxeter group” and the Hecke algebra on such a group, which are treated in §\[Siwahoriweylgroups\] and §\[SSdefofheckealgebra\], respectively. In §\[Siwahoriweylgroups\], I recall the notion of Iwahori-Weyl group and several of its most important properties, mostly to make explicit the scope of the Main Theorem. Reductive groups do not appear after this section. In §\[Smarkedalcoves\], I define what is a marked alcove. This is just a straightforward generalization of the notion of the type of a face of an alcove. This extension is necessary to include Iwahori-Weyl groups, rather than just affine Weyl groups, in the scope of the paper. In §\[Sdiamonds\], I precisely define the Diamond Property for an element of a Coxeter group. In short, the Diamond Property is the property asserted by the Main Theorem. I then define an equivalent property that refers only to pairs of alcoves and verify the equivalence of the two definitions. In §\[Slemmas\], I prove some simple geometric lemmas about Weyl chambers, alcoves, hyperplanes, etc. that will be used throughout §\[Sproof\]. In §\[Sproof\], I prove the Main Theorem, divided into three cases. In §\[SSdominantcase\], I prove the “dominant case”, Proposition \[PdominantnontranslationshaveDCP\], of the Main Theorem: if $ w \notin \Lambda $ and $ w $ sends the base alcove $ \basealcove $ into the dominant chamber $ \basechamber $ then $ w $ has the Diamond Property. It is obvious from the hypothesis that $ \ell ( s w ) > \ell ( w ) $ for all $ s \in \basereflections $, and it is not hard to visualize why there also exists $ s \in \basereflections $ such that $ \ell ( w s ) > \ell ( w ) $. In §\[SSantidominantcase\], I prove the “anti-dominant case”, Proposition \[PantidominantnontranslationshaveCP\], of the Main Theorem: if $ w \notin \Lambda $ and $ w ( \basevertex ) \in \closedbasechamber^{\operatorname{opp}} $ then $ w $ has the Diamond Property. This is the most difficult case and the general case can be reduced to this one (the complexity of the anti-dominant case is in some sense maximal while that of the dominant case is minimal–this can be quantified somewhat by noting that $ w ( \basealcove ) \subset \basechamber^{\operatorname{opp}} \Rightarrow \ell ( s w ) < \ell ( w ) $ for all $ s \in \basereflections $). The basic idea is that, by carefully inspecting the relative position and orientation of the alcoves $ \basealcove $ and $ w ( \basealcove ) $, one can perform an infinite sequence of conjugations by $ \Delta_{\operatorname{aff}} $ which do not decrease length and which continually move the alcoves in “different directions”, guaranteeing an eventual length-increasing conjugation by $ \Delta_{\operatorname{aff}} $. In §\[SSintermediatecase\], I finish proving the Main Theorem, showing that for an arbitrary $ w \notin \Lambda $ one can repeatedly perform conjugations by $ \Delta_{\operatorname{aff}} $ which do not decrease length and such that eventually the situation qualifies for the anti-dominant case. In §\[Sapplicationtoheckealgebras\], I use the Main Theorem to give dimension bounds, Proposition \[Pdimensionbounds\], for every subspace in the “length filtration” of $ Z ( \heckefont{H} ) $. The term “length filtration” is a slight abuse, since it is necessary to first filter $ Z ( \heckefont{H} ) $ by the lengths of its supporting elements and then partition each of those subspaces by the $ \Omega $-components of its supporting elements (if $ \Omega $ is infinite then the subspaces in the length filtration are infinite dimensional, and without refining it further most Iwahori-Hecke algebras would be outside the scope of the paper). In §\[Sfigures\], I include several color pictures that illustrate the iterative arguments that occur in the proof of the Main Theorem. I use the affine Weyl group and apartment of the exceptional type $ \Gaff{2} $ because there are too many coincidences for extremely symmetric types like $ \Aaff{2} $ to correctly explain things, and because I can suppress hyperplane labels for $ \Gaff{2} $ since its labelings are, unlike $ \Aaff{2} $ and $ \Caff{2} $, unambiguous (the alcoves are $ 30^{\circ} $-$ 60^{\circ} $-$ 90^{\circ} $ triangles). Acknowledgements ---------------- I thank Thomas Haines for carefully reading earlier drafts of this paper and suggesting a very large number of changes to the exposition, and especially for noticing that the original proof occurring in §\[SSintermediatecase\] was needlessly labyrinthine. I also thank my postdoctoral mentor, Tonghai Yang, for bringing me to the wonderful University of Wisconsin at Madison. Finally, I thank the referee for suggesting many improvements to the introduction, and also for the care and speed with which the report was composed and returned. Notation and Setup {#Snotation} ================== The symbols $ \N $, $ \Z $, $ \R $, and $ \C $ refer to the natural numbers (including $ 0 $), the integers, the real numbers, and the complex numbers. Root systems and affine Weyl groups {#SSweylgroups} ----------------------------------- Let $ \rootsystem $ be a reduced and irreducible root system. Let $ \finiteweylgroup $ be the finite Weyl group of $ \rootsystem $, let $ \basereflections $ be a simple system for $ \finiteweylgroup $. Let $ \apartmentsymbol $ be the $ \R $-vector space spanned by the dual root system $ \rootsystem^{\vee} $. Let $ \langle -, - \rangle $ be the natural pairing $ \rootsystem^{\vee} \times \rootsystem \rightarrow \Z $ and $ \langle -, - \rangle_{\R} $ the extension to $ \apartmentsymbol \times \apartmentsymbol^{\vee} \rightarrow \R $. Let $ \rootsystem_{\operatorname{aff}} $ be the affine root system associated to $ \rootsystem $, i.e. the set $ \rootsystem + \Z $ of affine functionals on $ \apartmentsymbol $. The term hyperplane always means the null-set of an element of $ \rootsystem_{\operatorname{aff}} $, not just an arbitrary codimension-$ 1 $ affine subspace. The term root hyperplane means the null-set of an element of $ \rootsystem $. Let $ W_{\operatorname{aff}} $ be the affine Weyl group of $ \rootsystem_{\operatorname{aff}} $ and $ \Delta_{\operatorname{aff}} $ the simple system for $ W_{\operatorname{aff}} $ extended from $ \basereflections $. Denote by $ s_{\operatorname{aff}} $ the single element of $ \Delta_{\operatorname{aff}} \setminus \basereflections $ and by $ H_{\operatorname{aff}} $ the hyperplane fixed pointwise by $ s_{\operatorname{aff}} $. Denote by $ \ell : W_{\operatorname{aff}} \rightarrow \N $ the usual length function relative to the generating set $ \Delta_{\operatorname{aff}} $. Let $ \corootlattice \subset W_{\operatorname{aff}} $ be the subgroup of translations by elements of $ \rootsystem^{\vee} $, so that $ W_{\operatorname{aff}} = \corootlattice \rtimes \finiteweylgroup $. Choose a special vertex $ \basevertex \in \apartmentsymbol $ and identify $ \finiteweylgroup $ to the finite Weyl group at $ \basevertex $. Let $ \basechamber $ be the Weyl chamber at $ \basevertex $ corresponding to $ \basereflections $ and let $ \basealcove \subset \basechamber $ be the alcove for which $ \basevertex \in \closedbasealcove $. If $ \weylchambersymbol $ is a Weyl chamber (at some arbitrary special vertex), denote by $ \weylchambersymbol^{\operatorname{opp}} $ the opposite chamber. Simplices and topology ---------------------- The term face always means a codimension-$ 1 $ facet of an alcove (or Weyl chamber). The term wall always means the unique hyperplane containing some face of some alcove (or Weyl chamber). If $ \alcovefont{A} $ and $ \alcovefont{B} $ are (distinct) adjacent alcoves then denote by $ \alcovefont{A} \vert \alcovefont{B} $ the wall separating them. The term half-space refers to one of the connected components of the complement in $ \apartmentsymbol $ of a hyperplane (in particular, half-spaces are open, and therefore also Weyl chambers and alcoves). A subset $ \mathfrak{R} \subset \apartmentsymbol $ is called simplicial iff there exists a set $ \mathbb{S} $ of alcoves in $ \apartmentsymbol $ such that $ \cup_{\alcovefont{A} \in \mathbb{S}} \alcovefont{A} \subset \mathfrak{R} \subset \cup_{\alcovefont{A} \in \mathbb{S}} \overline{\alcovefont{A}} $. For example, half-spaces and Weyl chambers are simplicial. Galleries and distances {#SSgalleries} ----------------------- The length of a gallery $ ( \alcovefont{B}_0, \alcovefont{B}_1, \ldots, \alcovefont{B}_n ) $ is defined to be $ n $ (for consistency with the length function $ \ell $). To say that a hyperplane $ H $ is an intermediate wall of a (non-stuttering) gallery $ ( \alcovefont{B}_0, \alcovefont{B}_1, \ldots, \alcovefont{B}_n ) $ is the same as to say that there exists an index $ 0 \leq j < n $ such that $ H $ is the unique hyperplane containing the face shared by $ \alcovefont{B}_j $ and $ \alcovefont{B}_{j+1} $. I sometimes use the fact that if $ \mathfrak{R} $ is simplicial and (topologically) convex then it is convex in the combinatorial sense, i.e. if $ \alcovefont{A}, \alcovefont{A}^{\prime} \subset \mathfrak{R} $ are alcoves and $ \mathcal{G} $ is a minimal gallery from $ \alcovefont{A} $ to $ \alcovefont{A}^{\prime} $ then $ \alcovefont{A}^{\prime \prime} \subset \mathfrak{R} $ for all $ \alcovefont{A}^{\prime \prime} \in \mathcal{G} $ (see Theorem 5.11.4 in [@BGW]). I sometimes refer to this property as simplicial convexity. For example, half-spaces and Weyl chambers are simplicially-convex. By infinite gallery I mean an infinite sequence $ ( \alcovefont{B}_0, \alcovefont{B}_1, \ldots ) $ of alcoves such that for all $ i \geq 0 $ the alcoves $ \alcovefont{B}_i $ and $ \alcovefont{B}_{i+1} $ are adjacent. An infinite gallery is minimal iff for all $ 0 \leq i < j $ the finite sub-gallery $ ( \alcovefont{B}_i, \ldots, \alcovefont{B}_j ) $ is minimal in the usual sense. Let $ d $ be the usual $ \N $-valued metric on the set of all alcoves: $ d ( \alcovefont{A}, \alcovefont{B} ) $ is defined to be the minimum length among all galleries from $ \alcovefont{A} $ and $ \alcovefont{B} $. Equivalently, $ d ( \alcovefont{A}, \alcovefont{B} ) $ is the total number of hyperplanes separating $ \alcovefont{A} $ from $ \alcovefont{B} $–see Theorem 5.1.4 in [@BGW]. I also use two extensions of this distance function $ d $: If $ \mathfrak{R} \neq \emptyset $ is a simplicial subset and $ \alcovefont{A} \subset \apartmentsymbol $ is an alcove then define $$d ( \alcovefont{A}, \mathfrak{R} ) \defeq \min ( d ( \alcovefont{A}, \alcovefont{B} ) \suchthat \text{all alcoves } \alcovefont{B} \subset \mathfrak{R} )$$ If $ \vertexfont{w} $ is a vertex, and $ \mathfrak{R} $ is as before, then I define $$d ( \vertexfont{w}, \mathfrak{R} ) \defeq \min ( d ( \alcovefont{A}, \mathfrak{R} ) \suchthat \text{all alcoves $ \alcovefont{A} \subset \apartmentsymbol $ such that } \vertexfont{w} \in \overline{\alcovefont{A}} )$$ In applications, $ \mathfrak{R} $ will be either a single half-space or a Weyl chamber. Iwahori-Weyl Groups {#Siwahoriweylgroups} =================== In this section, I define the object which is the main focus of this paper: the Iwahori-Weyl group. Definition and key properties {#SSdefofiwahoriweylgroup} ----------------------------- In this subsection, I briefly explain what is an Iwahori-Weyl group and isolate its key properties. The purpose here is merely to explain the scope of the Main Theorem, so all proofs are omitted, although I give references whenever possible. Let $ F $ be a non-archimedean local field and let $ G $ be a connected reductive affine algebraic $ F $-group. Let $ A \subset G $ be a maximal $ F $-split torus and set $ M \defeq C_G ( A ) $, a minimal $ F $-Levi subgroup. Certain group homomorphisms called Kottwitz homomorphisms are very useful to understand the theory of parahoric subgroups, and in particular Iwahori subgroups. The Kottwitz homomorphism of a connected reductive affine algebraic $ F $-group $ H $ is a surjective group homomorphism $ \kappa_H : H(F) \twoheadrightarrow \Omega_H $, where $ \Omega_H $ is a finitely-generated abelian group whose precise definition is not relevant to this paper–see §7 of [@kottwitz] for the definitions of $ \kappa_H $ and $ \Omega_H $ (the map $ \kappa_H $ occurring here is (7.7.1) in [@kottwitz]). The kernel of the Kottwitz homomorphism is denoted by $ H(F)_1 \defeq \operatorname{ker}( \kappa_H ) $. One may define the Iwahori-Weyl group of $ ( G, A ) $ to be the quotient $ \widetilde{W} \defeq N_G ( A ) ( F ) / M(F)_1 $. Note that this is seemingly different from the quotient occurring in Remark 9 of the Appendix to [@PRH], but it can be shown that there is a tautological isomorphism between the two quotients. Proposition 8 combined with Remark 9 of the Appendix to [@PRH] shows if $ I \subset G(F) $ is an Iwahori subgroup then the double-cosets modulo $ I $ are naturally represented by $ \widetilde{W} $. The Iwahori-Weyl group $ \widetilde{W} $ acts on the vector space $ \enlargedapartment \defeq \cochargroup ( A ) \otimes_{\Z} \R $. There are two extremely important ways to express the Iwahori-Weyl group $ \widetilde{W} $ as a semidirect product. By the work of Bruhat and Tits, it is known that there exists a reduced root system $ \scaledrootsystem $ such that the affine Weyl group $ W_{\operatorname{aff}} ( \scaledrootsystem ) $, in the sense of Ch VI §2 no. 1 of [@bourbaki], is a subgroup of $ \widetilde{W} $ (this root system is called an échelonnage in §1.4 of [@BT1]; see §4 of [@tits] for an extremely nice table listing $ \scaledrootsystem $ for every almost-simple group, and much more). Denoting by $ W_{\circ} ( \scaledrootsystem ) $ the finite Weyl group of $ \scaledrootsystem $, it can be shown that $ \widetilde{W} = \Omega_M \rtimes W_{\circ} ( \scaledrootsystem ) $ and that $ \widetilde{W} = W_{\operatorname{aff}} ( \scaledrootsystem ) \rtimes \Omega_G $. Further, the subgroup $ \Omega_M $ acts on $ \enlargedapartment $ by translations and the subgroup $ \Omega_G $ acts on $ \enlargedapartment $ by invertible affine transformations that stabilize any prescribed base alcove in $ \enlargedapartment $. For more details of all these semidirect products, consult [@PRH] and [@HRo]. A few comments are necessary to emphasize the small but important difference between the notion of an Iwahori-Weyl group and the possibly more familiar notion of an extended affine Weyl group of a (reduced) root datum. First, it is possible that both $ \Omega_M $ and $ \Omega_G $ have torsion elements (in fact, it can be shown that the torsion of the former is contained in the torsion of the latter). Second, the elements of $ \Omega_M $ are not actually translations, but merely act by translations on $ \enlargedapartment $. Third, it is possible that some non-identity elements in $ \Omega_M $ act by the identity on $ \enlargedapartment $. Axiomatization -------------- Using the previous discussion §\[SSdefofiwahoriweylgroup\] as a guide, I now isolate the relevant properties of the Iwahori-Weyl group and present them axiomatically for clarity. Let $ \mathcal{N} $ be the group of invertible affine transformations of $ \apartmentsymbol $ which normalize $ W_{\operatorname{aff}} $. Fix a finitely-generated abelian group $ \Omega $, a group homomorphism $ \psi : \Omega \rightarrow \mathcal{N} $, and act by $ \Omega $ on $ \apartmentsymbol $ via this $ \psi $. The Quasi-Coxeter Group $ \widetilde{W} $ extended from $ W_{\operatorname{aff}} $ by $ \Omega \stackrel{\psi}{\rightarrow} \mathcal{N} $ is the semidirect product $ W_{\operatorname{aff}} \rtimes \Omega $ and acts on $ \apartmentsymbol $ in the obvious way: $ ( w, \tau ) ( a ) \defeq w ( \tau ( a ) ) $ for all $ ( w, \tau ) \in \widetilde{W} $ and $ a \in \apartmentsymbol $. Denote by $ \Omega ( w ) $ the projection of $ w \in \widetilde{W} $ into $ \Omega $. Note that if $ w, w^{\prime} \in \widetilde{W} $ are conjugate then $ \Omega ( w ) = \Omega ( w^{\prime} ) $, since $ \Omega $ is abelian. Strictly speaking, the space $ \apartmentsymbol $ on which the quasi-Coxeter group acts is only a proper subspace of the space $ \enlargedapartment $ on which the Iwahori-Weyl group acts when $ G $ is not semisimple. But due to the way that affine root hyperplanes in $ \enlargedapartment $ are defined, the details of which I omit in this paper, the difference is totally irrelevant from a group-theory perspective. The setup that I use is essentially the same as that used in Ch VI §2 no. 3 of [@bourbaki]. Let $ \translationsubgroup \subset \widetilde{W} $ be the subgroup consisting of all elements that act by translations on $ \apartmentsymbol $, and note that $ \translationsubgroup $ is obviously normalized by $ \finiteweylgroup \subset W_{\operatorname{aff}} $. Extend the length function $ \ell : W_{\operatorname{aff}} \rightarrow \N $ to $ \widetilde{W} $ by inflation along the projection $ W_{\operatorname{aff}} \rtimes \Omega \rightarrow W_{\operatorname{aff}} $. I impose the following hypotheses: - Assume that $ \tau ( \basealcove ) = \basealcove $ for all $ \tau \in \Omega $. - Assume that $ \translationsubgroup $ is a semidirect complement, i.e. that $ \widetilde{W} = \translationsubgroup \rtimes \finiteweylgroup $. - Assume that $ \ell $ is constant on each $ \finiteweylgroup $-conjugacy class in $ \translationsubgroup $. - Assume that $ \translationsubgroup $ is finitely-generated and abelian. Note that by choice of $ \mathcal{N} $, the action by $ \Omega $ on $ \apartmentsymbol $ permutes the set of hyperplanes in $ \apartmentsymbol $. Therefore, hypothesis is equivalent to the hypothesis that $ \tau ( \Delta_{\operatorname{aff}} ) = \Delta_{\operatorname{aff}} $ for all $ \tau \in \Omega $. Marked Alcoves {#Smarkedalcoves} ============== The definitions in this section, which are mostly just a variant on the notion of the type of a face, will be used heavily in §\[SSantidominantcase\] and §\[SSintermediatecase\]. A Labeling of an alcove $ \alcovefont{A} \subset \apartmentsymbol $ is a bijection from $ \Delta_{\operatorname{aff}} $ to the set of walls of $ \alcovefont{A} $. A Marked Alcove is a triple $ ( \alcovefont{A}, \vertexsymbol, \labelingsymbol ) $ such that $ \alcovefont{A} $ is an alcove, $ \vertexsymbol \in \overline{\alcovefont{A}} $ is a special vertex and $ \labelingsymbol $ is a labeling of $ \alcovefont{A} $. The Weyl Chamber of a marked alcove $ ( \alcovefont{A}, \vertexsymbol, \labelingsymbol ) $ is the unique Weyl chamber at $ \vertexsymbol $ containing $ \alcovefont{A} $. Whenever the special vertex $ \vertexsymbol $ and labeling $ \labelingsymbol $ of a marked alcove $ ( \alcovefont{A}, \vertexsymbol, \labelingsymbol ) $ are understood and there is no danger of confusion, I abuse notation and refer to $ \alcovefont{A} $ as the marked alcove. Accordingly, if $ \alcovefont{A} $ represents a marked alcove then its special vertex is denoted by $ \vertexsymbol_{ \alcovefont{A} } $, its labeling by $ \labelingsymbol_{ \alcovefont{A} } $, its Weyl chamber by $ \weylchambersymbol_{ \alcovefont{A} } $, and the hyperplane $ \labelingsymbol_{\alcovefont{A}} ( s ) $ is called simply “the wall of $ \alcovefont{A} $ labeled by $ s $”. Two marked alcoves $ ( \alcovefont{A}, \vertexfont{v}, \labelingfont{t} ) $ and $ ( \alcovefont{B}, \vertexfont{w}, \labelingfont{s} ) $ are called Compatible iff there exists $ w \in \widetilde{W} $ such that $ \alcovefont{B} = w ( \alcovefont{A} ) $, $ \vertexfont{w} = w ( \vertexfont{v} ) $ and $ \labelingfont{s} = w \circ \labelingfont{t} $. The marked alcoves are called NT-Compatible iff $ w \notin \translationsubgroup $. Finally, the base alcove $ \basealcove $ is given the tautological labeling, and all other alcoves inherit (in general, multiple) labelings via the action of $ \widetilde{W} $ in the obvious way: The Base Labeling is the bijection $ \baselabeling $ from $ \Delta_{\operatorname{aff}} $ to the set of walls of the base alcove $ \basealcove $ defined by assigning to $ s $ the unique wall of $ \basealcove $ that is fixed pointwise by $ s $. The Base Marking is the marked alcove $ ( \basealcove, \basevertex, \baselabeling ) $. For each $ w \in \widetilde{W} $, the $ w $-Labeling is defined to be the bijection $ \labelingsymbol_w \defeq w \circ \baselabeling $ from $ \Delta_{\operatorname{aff}} $ to the set of walls of the alcove $ w ( \basealcove ) $. The $ w $-Marked Alcove is by definition the triple $ ( w ( \basealcove ), w ( \basevertex ), \labelingsymbol_w ) $. As before, I sometimes abuse notation by using $ w ( \basealcove ) $ to refer to the $ w $-marked alcove. Note that $ w ( \basealcove ) $ is compatible with $ \basealcove $ and it is NT-compatible with $ \basealcove $ if and only if $ w \notin \translationsubgroup $. When $ \Omega = \{ 1 \} $, alcoves are in bijection with $ w $-marked alcoves (due to simple-transitivity of affine Weyl groups on alcoves) and a labeling is essentially just the assignment to every face of every alcove its type in the usual way. The following operation will be used frequently in the limiting/inductive arguments of §\[SSantidominantcase\] and §\[SSintermediatecase\]: For any marked alcove $ ( \alcovefont{A}, \vertexfont{v}, \labelingfont{t} ) $ and any $ s \in \Delta_{\operatorname{aff}} $, the marked alcove $ ( \alcovefont{A}, \vertexfont{v}, \labelingfont{t} )^s $ is by definition the triple $ ( s_H ( \alcovefont{A} ), s_H ( \vertexfont{v} ), s_H \circ \labelingfont{t} ) $, where $ H \defeq \labelingfont{t} ( s ) $ is the wall of $ \alcovefont{A} $ labeled by $ s $. Note that if two marked alcoves $ \alcovefont{A} $ and $ \alcovefont{B} $ are NT-compatible then $ \alcovefont{A}^s $ and $ \alcovefont{B}^s $ are also NT-compatible for all $ s \in \Delta_{\operatorname{aff}} $. I frequently use the fact that applying a sequence of various $ * \mapsto ^s $ operations to a single marked alcove results in a sequence of marked alcoves whose (un-marked) alcoves form a gallery. Conversely, if $ \alcovefont{B}_0 $ is a marked alcove and $ ( \alcovefont{B}_0, \ldots, \alcovefont{B}_n ) $ is a gallery with no repeated alcoves, then each $ \alcovefont{B}_i $ becomes a marked alcove in a unique way, by iteratively applying $ \mapsto ^s $ operations across each intermediate wall of the gallery. In this situation, the “label” of $ \alcovefont{B}_i \vert \alcovefont{B}_{i+1} $ is understood to refer to the element of $ \Delta_{\operatorname{aff}} $ corresponding to the wall $ \alcovefont{B}_i \vert \alcovefont{B}_{i+1} $ relative to the labeling of $ \alcovefont{B}_i $ (or $ \alcovefont{B}_{i+1} $) inherited from $ \alcovefont{B}_0 $. Diamond Properties {#Sdiamonds} ================== Lateral-conjugacy and the diamond property in the group ------------------------------------------------------- $ w, w^{\prime} \in \widetilde{W} $ are Laterally-Conjugate* iff there exist $ s_1, \ldots, s_n \in \Delta_{\operatorname{aff}} $ such that $ w^{\prime} = s_n \cdots s_1 w s_1 \cdots s_n $ and $ \ell ( s_i \cdots s_1 w s_1 \cdots s_i ) = \ell ( w ) $ for all $ i $. Any $ w \in \widetilde{W} $ is always considered to be laterally-conjugate to itself. $ w \in \widetilde{W} $ has the Direct Diamond Property iff there exists $ s \in \Delta_{\operatorname{aff}} $ such that - $ s w s \neq w $, - $ \ell ( s w ) > \ell ( w ) $, and - $ \ell ( w s ) > \ell ( w ) $. By using the well-known Lemma \[Lhumphreys\], these three properties could be replaced by the single property “$ \ell ( s w s ) > \ell ( w ) $”, but this formulation is not as convenient for me. $ w \in \widetilde{W} $ has the Diamond Property iff it is laterally-conjugate to an element with the Direct Diamond Property. I frequently use the following geometric characterization of length: \[Lgeometriclength\] Let $ w \in \widetilde{W} $ and $ s \in \Delta_{\operatorname{aff}} $ be arbitrary. If $ H \defeq \baselabeling ( s ) $ and $ K \defeq \labelingsymbol_w ( s ) $ then - $ \ell ( s w ) > \ell ( w ) $ if and only if $ \basealcove $ and $ w ( \basealcove ) $ are on the same side of $ H $, - $ \ell ( w s ) > \ell ( w ) $ if and only if $ \basealcove $ and $ w ( \basealcove ) $ are on the same side of $ K $, and - $ \ell ( w ) = d ( \basealcove, w ( \basealcove ) ) $. When $ \Omega = \{ 1 \} $ this is all well-known: see Proposition (c) in §4.4 and Theorem (b) in §4.5 of [@humphreys]. The more general statement is immediate by definition of the labeling $ \labelingsymbol_w $ because $ \ell $ factors through $ W_{\operatorname{aff}} $ and $ \Omega $ stabilizes $ \basealcove $. Lateral-conjugacy and the diamond property in the apartment ----------------------------------------------------------- Here are the gallery-theoretic versions of the above 3 definitions: An ordered pair $ ( \alcovefont{A}, \alcovefont{B} ) $ of marked alcoves is Laterally-Conjugate to another pair $ ( \alcovefont{A}^{\prime}, \alcovefont{B}^{\prime} ) $ iff there exists a gallery $ ( \alcovefont{A}_0, \alcovefont{A}_1, \ldots, \alcovefont{A}_n ) $ from $ \alcovefont{A} $ to $ \alcovefont{A}^{\prime} $ and a gallery $ ( \alcovefont{B}_0, \alcovefont{B}_1, \ldots, \alcovefont{B}_n ) $ from $ \alcovefont{B} $ to $ \alcovefont{B}^{\prime} $ such that - $ d ( \alcovefont{A}_i, \alcovefont{B}_i ) = d ( \alcovefont{A}, \alcovefont{B} ) $, and - both $ \alcovefont{A}_i \vert \alcovefont{A}_{i+1} $ and $ \alcovefont{B}_i \vert \alcovefont{B}_{i+1} $ have the same label. for all $ i $. Note the symmetry in the definition of lateral-conjugacy: $ ( \alcovefont{A}, \alcovefont{B} ) $ is laterally-conjugate to $ ( \alcovefont{A}^{\prime}, \alcovefont{B}^{\prime} ) $ if and only if $ ( \alcovefont{B}, \alcovefont{A} ) $ is laterally-conjugate to $ ( \alcovefont{B}^{\prime}, \alcovefont{A}^{\prime} ) $. A pair $ \{ \alcovefont{A}, \alcovefont{B} \} $ of marked alcoves has the Direct Diamond Property iff there exists $ s \in \Delta_{\operatorname{aff}} $ such that - $ \mathbf{t}_{\alcovefont{A}} ( s ) \neq \mathbf{t}_{\alcovefont{B}} ( s ) $, - both alcoves are on the same side of $ \mathbf{t}_{\alcovefont{A}} ( s ) $, and - both alcoves are also on the same side of $ \mathbf{t}_{\alcovefont{B}} ( s ) $. Note that, due to the symmetry in the definition, the Direct Diamond Property refers to unordered pairs of alcoves. A pair $ \{ \alcovefont{A}, \alcovefont{B} \} $ of marked alcoves has the Diamond Property iff it is laterally-conjugate to a pair with the direct diamond property. Equivalence ----------- It is easy to show using the previous lemma that the two notions of Diamond Property coincide: \[Lequivalentdiamonds\] The element $ w \in \widetilde{W} $ has the Diamond Property if and only if the pair $ \{ \basealcove, w ( \basealcove ) \} $ of marked alcoves has the Diamond Property. By Lemma \[Lgeometriclength\], $ \ell ( s w s ) = d ( \basealcove, s w s ( \basealcove ) ) $. Since $ d $ is invariant under the diagonal action of $ W_{\operatorname{aff}} $, $ d ( \basealcove, s w s ( \basealcove ) ) = d ( s ( \basealcove ), w s ( \basealcove ) ) $. By definition of the labelings $ \baselabeling $ and $ \labelingsymbol_w $, $ d ( s ( \basealcove ), w s ( \basealcove ) ) = d ( \basealcove^s, w ( \basealcove )^s ) $. Altogether, $ \ell ( s w s ) = d ( \basealcove^s, w ( \basealcove )^s ) $. Since the operation $ * \mapsto ^s $ always creates galleries, this shows that the two notions of “lateral-conjugacy” are equivalent. Since $ s_H w s_H = w $ if and only if $ w ( H ) = H $, it follows that the condition $ s w s \neq w $ is equivalent to the condition $ \labelingsymbol_w ( s ) \neq \baselabeling ( s ) $. Finally, the fact that the remaining two statements in both Direct Diamond Properties are equivalent follows directly from Lemma \[Lgeometriclength\]. Although I will have no direct use for this in the remainder of the paper, note that the statement of Lemma \[Lequivalentdiamonds\] could be made much more specific: if $ w $ is laterally-conjugate to $ w^{\prime} $ via the sequence $ s_1, \ldots, s_r $ then $ ( \basealcove, w ( \basealcove ) ) $ is laterally-conjugate to $ ( s_1 \cdots s_r ( \basealcove ), w s_1 \cdots s_r ( \basealcove ) ) $ via a gallery whose intermediate walls are labeled (in tandem) by the same sequence $ s_1, \ldots, s_r $, etc. Basic Lemmas {#Slemmas} ============ \[Laffinehyperplanesarenotbipolar\] Let $ H $ be a hyperplane in $ \apartmentsymbol $. If both* $ \closedbasechamber \cap H \neq \emptyset $ and $ \closedbasechamber^{\operatorname{opp}} \cap H \neq \emptyset $ then in fact both $ \closedbasechamber \cap H \subset \partial \basechamber $ and $ \closedbasechamber^{\operatorname{opp}} \cap H \subset \partial \basechamber^{\operatorname{opp}} $. I usually apply Lemma \[Laffinehyperplanesarenotbipolar\] in the following way: if $ H \cap \basechamber \neq \emptyset $ then $ H \cap \closedbasechamber^{\operatorname{opp}} = \emptyset $. Suppose $ x \in H \cap \closedbasechamber $ and $ y \in H \cap \closedbasechamber^{\operatorname{opp}} $. Since $ H $ is the null-set of an affine root, there exists $ \beta \in \rootsystem $ such that $ x - y \in H_{\beta} $. On the other hand, $ x \in \closedbasechamber $ and $ y \in \closedbasechamber^{\operatorname{opp}} $ implies $ x - y \in \closedbasechamber $. It is easy to check from the definitions of $ \basechamber $ and $ \basechamber^{\operatorname{opp}} $ that if $ x \in \basechamber $ or $ y \in \basechamber^{\operatorname{opp}} $ (or both) then necessarily $ x - y \in \basechamber $. But $ H_{\beta} \cap \basechamber = \emptyset $ since a root hyperplane can never intersect a Weyl chamber at $ \basevertex $, so this is impossible and therefore both $ x \notin \basechamber $ and $ y \notin \basechamber^{\operatorname{opp}} $. \[Lnegativechambercontainment\] Let $ \alcovefont{A} \subset \basechamber $ be an alcove and $ H $ a wall of $ \alcovefont{A} $ that is not a wall of the Weyl chamber $ \basechamber $. Let $ \alcovefont{B} $ be any alcove and $ \vertexsymbol_{ \alcovefont{B} } \in \overline{\alcovefont{B}} $ a vertex. If both $ \alcovefont{A} $ and $ \basealcove $ are on the same side of $ H $ and $ \vertexsymbol_{ \alcovefont{B} } \in \closedbasechamber^{\operatorname{opp}} $ then $ \alcovefont{B} $ is also on the same side of $ H $ as $ \alcovefont{A} $. Let $ \alcovefont{f} \subset \overline{\alcovefont{A}} $ be the face supported by $ H $. If it were true that $ \alcovefont{f} \subset \partial \basechamber $ then necessarily $ H $ would be a wall of $ \basechamber $. This is prohibited by hypothesis on $ H $, so $ H \cap \basechamber \neq \emptyset $. Suppose for contradiction that $ H $ separated $ \alcovefont{B} $ from $ \alcovefont{A} $. By hypothesis on $ \alcovefont{A} $, $ H $ must separate $ \alcovefont{B} $ from $ \basealcove $. By hypothesis on $ \vertexsymbol_{ \alcovefont{B} } $, the set $ \alcovefont{B} \cup \{ \vertexsymbol_{ \alcovefont{B} } \} \cup \basechamber^{\operatorname{opp}} \cup \{ \basevertex \} \cup \basealcove $ is path-connected and obviously $ H \cap ( \alcovefont{B} \cup \basealcove ) = \emptyset $, so necessarily $ H \cap \closedbasechamber^{\operatorname{opp}} \neq \emptyset $. But this contradicts Lemma \[Laffinehyperplanesarenotbipolar\] since $ H \cap \basechamber \neq \emptyset $ is known already. \[Ldoublechamberlemma\] Let $ \weylchambersymbol $ be a Weyl chamber at some (arbitrary) special vertex $ \vertexsymbol $. If $ \vertexsymbol \in \closedbasechamber^{\operatorname{opp}} $ then either $ \weylchambersymbol \cap \basechamber = \emptyset $ or $ \weylchambersymbol \supset \basechamber $. Suppose for contradiction that both $ \weylchambersymbol \cap \basechamber \neq \emptyset $ and $ \weylchambersymbol \not\supset \basechamber $. Choose an alcove $ \alcovefont{A} \subset \basechamber \setminus \weylchambersymbol $ and an alcove $ \alcovefont{A}^{\prime} \subset \basechamber \cap \weylchambersymbol $. Let $ ( \alcovefont{A}_1, \ldots, \alcovefont{A}_n ) $ be a minimal gallery from $ \alcovefont{A} $ to $ \alcovefont{A}^{\prime} $. By choice of $ \alcovefont{A}, \alcovefont{A}^{\prime} $, there exists a wall $ H $ of $ \weylchambersymbol $ separating $ \alcovefont{A} $ from $ \alcovefont{A}^{\prime} $. Such an $ H $ must be an intermediate wall of the gallery (see Lemma 5.1.5 of [@BGW]), say $ H = \alcovefont{A}_j \vert \alcovefont{A}_{j+1} $. Since the gallery is minimal and both $ \alcovefont{A}, \alcovefont{A}^{\prime} \subset \basechamber $, simplicial convexity of $ \basechamber $ forces $ \alcovefont{A}_i \subset \basechamber $ for all $ i $. This means that $ \basechamber \cap H \neq \emptyset $, (for example, if $ \alcovefont{f} $ is the common face of $ \alcovefont{A}_j $ and $ \alcovefont{A}_{j+1} $ then $ \alcovefont{f} \subset \basechamber \cap H $). But by hypothesis $ \vertexsymbol \in \closedbasechamber^{\operatorname{opp}} $ and obviously $ \vertexsymbol \in H $, so $ \closedbasechamber^{\operatorname{opp}} \cap H \neq \emptyset $ also. This contradicts Lemma \[Laffinehyperplanesarenotbipolar\]. \[Lgallerytoinfinity\] Let $ \weylchambersymbol $ be a Weyl chamber at some (arbitrary) special vertex. If $ \basechamber \cap \weylchambersymbol = \emptyset $ then there exists a minimal infinite gallery $ ( \alcovefont{A}_0 = \basealcove, \alcovefont{A}_1, \alcovefont{A}_2, \ldots ) $ within $ \basechamber $ such that $$\lim_{i \rightarrow \infty} d ( \alcovefont{A}_i, \weylchambersymbol ) = \infty.$$ (see §\[SSgalleries\] for the notions of distance $ d ( * , \weylchambersymbol ) $ and infinite gallery) It is not always possible to have a sequence of alcoves for which the sequence of distances is monotone increasing. Choose some translation $ t \in \corootlattice $ such that $ t ( \basevertex ) \in \basechamber $ (e.g. translation by $ 2 \rho^{\vee} $ where $ \rho^{\vee} \defeq \omega_1 + \cdots + \omega_r $ and $ \omega_i $ are the fundamental coweights). I claim that the required gallery can be constructed by iterating $ t $. Let $ t = s_0 s_1 \cdots s_{k-1} $ ($ s_i \in \Delta_{\operatorname{aff}} $) be a reduced expression and let $ [i] $ be the remainder of $ i $ mod $ k $. Give $ \alcovefont{A}_0 \defeq \basealcove $ the base marking as usual and define a sequence of marked alcoves inductively by $ \alcovefont{A}_{i+1} \defeq \alcovefont{A}_i^{s_{[i]}} $ (so the finite subgallery $ ( \alcovefont{A}_0, \alcovefont{A}_1, \ldots, \alcovefont{A}_k ) $ is just the usual gallery associated to the word $ s_0 s_1 \cdots s_{k-1} $). By construction, this sequence is an infinite gallery in $ \basechamber $. Since $ \ell ( t^N ) = N \ell ( t ) $, a general property of dominant translations in an affine Weyl group, it follows that the infinite gallery $ ( \alcovefont{A}_0, \alcovefont{A}_1, \alcovefont{A}_2, \ldots ) $ is minimal. I first show that the infinite subsequence $ \alcovefont{A}_{k N} = t^N ( \basealcove ) $, $ N = 0, 1, 2, \ldots $, diverges from $ \weylchambersymbol $, and then I use the triangle-inequality to prove the full limit property. For any alcove $ \alcovefont{A} \subset \apartmentsymbol $ and radius $ R \in \R $, let $ \mathbb{B} ( \alcovefont{A}, R ) $ be the set of alcoves $ \alcovefont{B} \subset \apartmentsymbol $ such that $ d ( \alcovefont{A}, \alcovefont{B} ) \leq R $. It is clear from the “cone” property of Weyl chambers and the boundedness of alcoves that for any $ R \in \R $, there exists $ n_R \in \N $ such that $ \mathbb{B} ( t^N ( \basealcove ), R ) \subset \basechamber $ for all $ N \geq n_R $. It is also clear that $ d ( t^N ( \basealcove ), \weylchambersymbol ) > R $ for all $ N \geq n_R $ because otherwise there would be some alcove $ \alcovefont{B} \subset \weylchambersymbol $ such that $ d ( t^N ( \basealcove ), \alcovefont{B} ) \leq R $, but this would imply $ \alcovefont{B} \subset \mathbb{B} ( t^N ( \basealcove ), R ) \subset \basechamber $, which contradicts the hypothesis $ \basechamber \cap \weylchambersymbol = \emptyset $. This establishes the claim for the subsequence. Let radius $ R > 0 $ be arbitrary. Fix $ n \defeq k \cdot n_{R+k} $ (recall $ k = \ell ( t ) $). For any $ N \in \N $, let $ \lfloor N \rfloor $ be the largest $ m \in \N $ such that $ k m \leq N $. Observe that if $ N \geq n $ then $ \lfloor N \rfloor \geq n_{R+k} $. Altogether, if $ \alcovefont{B} \subset \weylchambersymbol $ is an arbitrary alcove and $ N \geq n $ then $$R + k < d ( t^{ \lfloor N \rfloor } ( \basealcove ), \alcovefont{B} ) \leq d ( t^{ \lfloor N \rfloor } ( \basealcove ), \alcovefont{A}_N ) + d ( \alcovefont{A}_N, \alcovefont{B} )$$ Since $ d ( t^{ \lfloor N \rfloor } ( \basealcove ), \alcovefont{A}_N ) \leq \ell ( t ) = k $ by definition of $ \lfloor N \rfloor $, it follows that $ d ( \alcovefont{A}_N, \alcovefont{B} ) > R $. Proof of Main Theorem {#Sproof} ===================== Case: the dominant chamber {#SSdominantcase} -------------------------- \[PdominantnontranslationshaveDCP\] Fix $ w \in \widetilde{W} $. If $ w \notin \translationsubgroup $ and $ w ( \basealcove ) \subset \basechamber $ then $ w $ has the Direct Diamond Property realized by some $ s \in \basereflections $. Let $ u \in \finiteweylgroup $ and $ t \in \translationsubgroup $ be such that $ w = t \circ u $. By hypothesis, $ u \neq 1 $. Let $ H^{\prime} $ be a wall of $ u ( \basechamber ) $ which separates $ u ( \basealcove ) $ from $ \basealcove $, and note that $ \basevertex \in H^{\prime} $. Let $ H $ be the wall of $ \basealcove $ such that $ u ( H ) = H^{\prime} $ and let $ s \in \basereflections $ be the element fixing $ H $ pointwise (in other words, $ H^{\prime} = \mathbf{t}_u ( s ) $). I claim that $ s $ realizes the Direct Diamond Property for $ w $. I first claim that $ t ( H^{\prime} ) \neq H^{\prime} $. Suppose for contradiction that $ t ( H^{\prime} ) = H^{\prime} $. Because $ t $ is a translation, $ u ( \basealcove ) $ and $ t ( u ( \basealcove ) ) $ are on the same side of $ t ( H^{\prime} ) = H^{\prime} $. On the other hand, $ H^{\prime} $ separates $ \basealcove $ from $ u ( \basealcove ) $ by choice. Together, $ H^{\prime} $ separates $ \basealcove $ from $ t ( u ( \basealcove ) ) = w ( \basealcove ) $. But $ w ( \basealcove ) \subset \basechamber $ by hypothesis, so it is impossible for the root hyperplane $ H^{\prime} $ to separate $ \basealcove $ from $ w ( \basealcove ) $. By hypothesis that $ w ( \basealcove ) \subset \basechamber $, it is automatic that $ \ell ( s w ) > \ell ( w ) $. To show that $ \ell ( w s ) > \ell ( w ) $, it suffices by Lemma \[Lgeometriclength\] to show that both alcoves $ \basealcove $ and $ w ( \basealcove ) $ are on the same side of $ w ( H ) $. Let $ \alpha \in \rootsystem $ be the positive root whose null-set is $ H^{\prime} $. By choice of $ H^{\prime} $, $ \langle \alpha, x \rangle_{\R} < 0 $ for all $ x \in u ( \basealcove ) $. Since $ u ( \basevertex ) = \basevertex $ and $ w ( \basevertex ) \in \closedbasechamber $, it must be true that $ t ( \basevertex ) \in \closedbasechamber $. Since $ t $ is a translation, this implies that there exists $ n \in \N $ such that $ t ( H^{\prime} ) = w ( H ) $ is the null-set of $ \alpha - n $ and $ t ( u ( \basealcove ) ) = w ( \basealcove ) $ consists of points $ x \in \apartmentsymbol $ such that $ \langle \alpha, x \rangle_{\R} < n $. Since $ t ( H^{\prime} ) \neq H^{\prime} $, it must be true that $ n \geq 1 $. But $ 0 < \langle \alpha, x \rangle_{\R} < 1 \leq n $ for all $ x \in \basealcove $ so $ \basealcove $ and $ w ( \basealcove ) $ are on the same side of $ w ( H ) $, as desired. I now show that $ s w s \neq w $. Suppose for contradiction that $ s w s = w $. Then $ w ( H ) = H $, and since $ u ( \basevertex ) = \basevertex $, it follows that $ t ( \basevertex ) \in H $. Since $ t $ is a translation and $ \basevertex \in H $, $ t ( H ) = H $. Combining with $ w ( H ) = H $ implies $ u ( H ) = H $ also. But $ H^{\prime} \defeq u ( H ) $ so this contradicts $ t ( H^{\prime} ) \neq H^{\prime} $. \[Rindependenceofchamber\] Note that it is not important which of the two alcoves is considered the “base” alcove, nor is it important which chamber of the base alcove is considered “dominant”. In other words, if $ \alcovefont{A} $ and $ \alcovefont{B} $ are NT-compatible marked alcoves and $ \alcovefont{B} \subset \weylchambersymbol_{\alcovefont{A}} $ then $ \{ \alcovefont{A}, \alcovefont{B} \} $ has the Direct Diamond Property. It is plausible that one might be able to prove the Diamond Property for general $ w \notin \translationsubgroup $ by proving that there always exists a lateral conjugate $ w^{\prime} $ of $ w $ such that one of $ w^{\prime} ( \basealcove ) $ or $ \basealcove $ is contained in some Weyl chamber of the other. This latter statement is false. See Figure \[example-g2-annoyingorbit\] for an example in the case of the exceptional affine Weyl group $ \Gaff{2} $. Case: the anti-dominant chamber {#SSantidominantcase} ------------------------------- Let $ ( \alcovefont{B}_0, \ldots, \alcovefont{B}_n ) $ be a gallery, $ \alcovefont{A} $ an alcove, and $ H $ a wall of $ \alcovefont{A} $. The triple $ ( ( \alcovefont{B}_0, \ldots, \alcovefont{B}_n ), \alcovefont{A}, H ) $ is an Umbrella iff 1. \[UMBsamehalfspace\] all alcoves $ \alcovefont{B}_i $ are on the same side of $ H $ as $ \alcovefont{A} $, and 2. \[UMBextendstominimal\] $ ( \alcovefont{B}_0, \ldots, \alcovefont{B}_n ) $ can be extended to a minimal gallery from $ \alcovefont{B}_0 $ to $ \alcovefont{A} $. Observe that to say $ ( \alcovefont{B}_0, \ldots, \alcovefont{B}_n ) $ can be extended to a minimal gallery from $ \alcovefont{B}_0 $ to $ \alcovefont{A} $ is the same as to say both that $ ( \alcovefont{B}_0, \ldots, \alcovefont{B}_n ) $ is a minimal gallery itself and that each intermediate wall $ \alcovefont{B}_i \vert \alcovefont{B}_{i+1} $ ($ 0 \leq i < n $) separates $ \alcovefont{B}_i $ from $ \alcovefont{A} $ (I use this observation in the proof of Induction Lemma). Let $ ( \alcovefont{B}_0, \ldots, \alcovefont{B}_n ) $ be a gallery and $ \vertexsymbol_{\alcovefont{B}_0} \in \overline{\alcovefont{B}}_0 $ a special vertex. Let $ \alcovefont{A}, \alcovefont{A}^{\prime} \subset \basechamber $ be (distinct) adjacent alcoves, separated by a wall $ H $. Let $ H^{\prime} $ be a wall of $ \alcovefont{A}^{\prime} $. Assume that 1. \[ENUMinductionhypoth\] $ ( ( \alcovefont{B}_0, \ldots, \alcovefont{B}_{n-1} ), \alcovefont{A}, H ) $ is an Umbrella, 2. \[ENUMvertexinchamber\] $ \vertexsymbol_{\alcovefont{B}_0} \in \closedbasechamber^{\operatorname{opp}} $, 3. \[ENUMawayfrombase\] the base alcove $ \basealcove $ is on the same side of $ H^{\prime} $ as $ \alcovefont{A}^{\prime} $, 4. \[ENUMnotbasewall\] $ H^{\prime} $ is not a wall of the Weyl chamber $ \basechamber $, and 5. \[ENUMcontradictionhypoth\] the wall $ \alcovefont{B}_{n-1} \vert \alcovefont{B}_n $ separates $ \alcovefont{B}_{n-1} $ from $ \alcovefont{A} $. Then $ ( ( \alcovefont{B}_0, \ldots, \alcovefont{B}_n ), \alcovefont{A}^{\prime}, H^{\prime} ) $ is an Umbrella. By hypotheses (\[ENUMvertexinchamber\]), (\[ENUMawayfrombase\]), and (\[ENUMnotbasewall\]), Lemma \[Lnegativechambercontainment\] implies that $ \alcovefont{B}_0 $ is contained on the same side of $ H^{\prime} $ as $ \alcovefont{A}^{\prime} $. Since half-spaces are simplicially-convex, it therefore suffices to show only Umbrella Property (\[UMBextendstominimal\]), i.e. that $ ( \alcovefont{B}_0, \ldots, \alcovefont{B}_n ) $ can be extended to a minimal gallery connecting $ \alcovefont{B}_0 $ to $ \alcovefont{A}^{\prime} $ (because then both endpoints of the gallery, and therefore the whole gallery, must be contained in that half-space). Let $ H_i \defeq \alcovefont{B}_i \vert \alcovefont{B}_{i+1} $ ($ i = 0, \ldots, n-1 $) be all the intermediate walls of the gallery $ ( \alcovefont{B}_0, \ldots, \alcovefont{B}_n ) $. Note that $ H_i $ separates $ \alcovefont{B}_i $ from $ \alcovefont{A} $ for all $ 0 \leq i < n-1 $ by hypothesis (\[ENUMinductionhypoth\]) (more specifically, Umbrella Property (\[UMBextendstominimal\])) and for $ i = n-1 $ by hypothesis (\[ENUMcontradictionhypoth\]). By the observation preceding this proof, it therefore suffices to show that the alcoves $ \alcovefont{A} $ and $ \alcovefont{A}^{\prime} $ are on the same side of $ H_i $ for all $ 0 \leq i \leq n - 1 $. But this is obviously true: if the claim were false for $ H_i $, then necessarily $ H_i = H $, the only hyperplane separating $ \alcovefont{A} $ from $ \alcovefont{A}^{\prime} $, which would mean that $ H $ separated $ \alcovefont{B}_i $ from $ \alcovefont{A} $, a contradiction to hypothesis (\[ENUMinductionhypoth\]) (more specifically, Umbrella Property (\[UMBsamehalfspace\])). \[PantidominantnontranslationshaveCP\] Fix $ w \in \widetilde{W} $. If $ w \notin \translationsubgroup $ and $ w ( \basevertex ) \in \closedbasechamber^{\operatorname{opp}} $ then $ w $ has the Diamond Property. View (sequentially!) Figures \[example-g2-inductionlemma00\] to \[example-g2-inductionlemma04\] for a picture of the use of Induction Lemma in this proof. Let $ \alcovefont{B} $ be the $ w $-marked alcove $ w ( \basealcove ) $. By Lemma \[Lequivalentdiamonds\], it suffices to show that $ \{ \basealcove, \alcovefont{B} \} $ has the Diamond Property. By Lemma \[Ldoublechamberlemma\], either $ \basechamber \cap \weylchambersymbol_{\alcovefont{B}} = \emptyset $ or $ \basechamber \subset \weylchambersymbol_{\alcovefont{B}} $. If $ \basechamber \subset \weylchambersymbol_{\alcovefont{B}} $ then the claim follows from Proposition \[PdominantnontranslationshaveDCP\]. (using origin $ \vertexsymbol_{\alcovefont{B}} $ and dominant chamber $ \weylchambersymbol_{\alcovefont{B}} $; see Remark \[Rindependenceofchamber\]). So, assume that $ \basechamber \cap \weylchambersymbol_{\alcovefont{B}} = \emptyset $. Applying Lemma \[Lgallerytoinfinity\] to the chambers $ \basechamber $ and $ \weylchambersymbol_{\alcovefont{B}} $ yields a certain infinite minimal gallery $ ( \alcovefont{A}_0 = \basealcove, \alcovefont{A}_1, \alcovefont{A}_2, \ldots ) $ within $ \basechamber $. As usual, give $ \basealcove $ the base marking and let $ ( s_0, s_1, s_2, \ldots ) $ be the infinite sequence in $ \Delta_{\operatorname{aff}} $ such that $ \alcovefont{A}_1 = \alcovefont{A}_0^{s_0} $, $ \alcovefont{A}_2 = \alcovefont{A}_1^{s_1} $, etc. Let $ H_i $ be the wall of the marked alcove $ \alcovefont{A}_i $ labeled by $ s_i $. Similarly, use the sequence $ ( s_0, s_1, s_2, \ldots ) $ to define, relative to the prescribed labeling of $ \alcovefont{B} $, a corresponding infinite gallery: $$( \alcovefont{B}_0, \alcovefont{B}_1, \alcovefont{B}_2, \ldots ) \defeq ( \alcovefont{B}, \alcovefont{B}^{s_0}, ( \alcovefont{B}^{s_0} )^{s_1}, \ldots )$$ As before, each $ \alcovefont{B}_i $ here represents a marked alcove. Note that by definition of the labeling $ \labelingsymbol_{\alcovefont{B}} = \labelingsymbol_w $, the gallery $ ( \alcovefont{B}_0, \alcovefont{B}_1, \alcovefont{B}_2, \ldots ) $ is simply the image under $ w $ of the gallery $ ( \alcovefont{A}_0, \alcovefont{A}_1, \alcovefont{A}_2, \ldots ) $. In particular, $ ( \alcovefont{B}_0, \alcovefont{B}_1, \alcovefont{B}_2, \ldots ) $ is minimal and $ \alcovefont{B}_i \subset \weylchambersymbol_{ \alcovefont{B} } $ for all $ i $. Because the gallery $ ( \alcovefont{A}_0, \alcovefont{A}_1, \alcovefont{A}_2, \ldots ) $ starts at $ \basealcove $ and is contained completely within $ \basechamber $, necessarily $ s_0 = s_{\operatorname{aff}} $. Because of this and the hypothesis on $ \vertexsymbol_{ \alcovefont{B} } $, Lemma \[Lnegativechambercontainment\] says that alcoves $ \alcovefont{B}_0 $ and $ \alcovefont{A}_0 $ are on the same side of $ H_0 $, i.e. $ d ( \alcovefont{B}_0, \alcovefont{A}_1 ) = d ( \alcovefont{B}_0, \alcovefont{A}_0 ) + 1 $. Let $ K \defeq \labelingsymbol_{\alcovefont{B}_0} ( s_0 ) $ be the wall of $ \alcovefont{B}_0 $ labeled by $ s_0 $. If $ \alcovefont{B}_0 $ and $ \alcovefont{A}_1 $ are on the same side of $ K $ then necessarily $ K \neq H_0 $ (because $ H_0 $ separates $ \alcovefont{B}_0 $ from $ \alcovefont{A}_1 $) and both $ \alcovefont{B}_0 $ and $ \alcovefont{A}_0 $ are on the same side of $ K $ (because $ H_0 $ is the unique hyperplane separating $ \alcovefont{A}_0 $ from $ \alcovefont{A}_1 $ and $ H_0 \neq K $). It is then immediate from the definition that $ \{ \alcovefont{A}_0, \alcovefont{B}_0 \} = \{ \basealcove, \alcovefont{B} \} $ has the Direct Diamond Property (realized by $ s_0 $). Otherwise, $ K $ separates $ \alcovefont{B}_0 $ from $ \alcovefont{A}_1 $ and by Lemma \[Lgeometriclength\], $ d ( \alcovefont{B}_1, \alcovefont{A}_1 ) = d ( \alcovefont{B}_0, \alcovefont{A}_1 ) - 1 = d ( \alcovefont{B}_0, \alcovefont{A}_0 ) + 1 - 1 = d ( \alcovefont{B}_0, \alcovefont{A}_0 ) $, i.e. $ ( \alcovefont{A}_1, \alcovefont{B}_1 ) $ is laterally-conjugate to $ ( \basealcove, \alcovefont{B} ) $ via $ s_0 $. In these circumstances, Induction Lemma implies that $ ( ( \alcovefont{B}_0, \alcovefont{B}_1 ), \alcovefont{A}_1, H_1 ) $ is an Umbrella: - the non-numbered hypotheses of Induction Lemma are true by choice, - hypothesis (\[ENUMinductionhypoth\]) is true because $ ( ( \alcovefont{B}_0 ), \alcovefont{A}_0, H_0 ) $ is trivially an Umbrella, - hypothesis (\[ENUMvertexinchamber\]) is true by hypothesis on $ w $, - hypothesis (\[ENUMawayfrombase\]) is true by choice of $ H_1 $ because $ ( \alcovefont{A}_0, \alcovefont{A}_1, \alcovefont{A}_2, \ldots ) $ is minimal, - hypothesis (\[ENUMnotbasewall\]) is true by choice of $ H_1 $ because $ \alcovefont{A}_1, \alcovefont{A}_2 \subset \basechamber $, and - hypothesis (\[ENUMcontradictionhypoth\]) is true by the assumption that $ \{ \alcovefont{A}_0, \alcovefont{B}_0 \} $ did not have the Direct Diamond Property for $ s_0 $ (see previous paragraph: by choice $ K = \alcovefont{B}_0 \vert \alcovefont{B}_1 $). So, the triple $ ( ( \alcovefont{B}_0, \alcovefont{B}_1 ), \alcovefont{A}_1, H_1 ) $ is an Umbrella by Induction Lemma. In particular, $ d ( \alcovefont{B}_1, \alcovefont{A}_2 ) = d ( \alcovefont{B}_1, \alcovefont{A}_1 ) + 1 $ by Umbrella Property (\[UMBsamehalfspace\]). I now iterate this process. Let $ K \defeq \labelingsymbol_{\alcovefont{B}_1} ( s_1 ) $ be the wall of the marked alcove $ \alcovefont{B}_1 $ labeled by $ s_1 $. If $ \alcovefont{B}_1 $ and $ \alcovefont{A}_2 $ are on the same side of $ K $ then $ K \neq H_1 $ (because $ H_1 $ separates $ \alcovefont{B}_1 $ from $ \alcovefont{A}_2 $) and both $ \alcovefont{B}_1 $ and $ \alcovefont{A}_1 $ are on the same side of $ K $ (because $ H_1 $ is the unique hyperplane separating $ \alcovefont{A}_1 $ from $ \alcovefont{A}_2 $ and $ H_1 \neq K $). It is then immediate from the definition that $ \{ \alcovefont{A}_1, \alcovefont{B}_1 \} $ has the Direct Diamond Property realized by $ s_1 $ and therefore $ \{ \basealcove, \alcovefont{B} \} $, being laterally-conjugate to it, has the Diamond Property. Otherwise, $ K $ separates $ \alcovefont{B}_1 $ from $ \alcovefont{A}_2 $ and by Lemma \[Lgeometriclength\], $ d ( \alcovefont{B}_2, \alcovefont{A}_2 ) = d ( \alcovefont{B}_1, \alcovefont{A}_2 ) - 1 = d ( \alcovefont{B}_1, \alcovefont{A}_1 ) + 1 - 1 = d ( \alcovefont{B}_1, \alcovefont{A}_1 ) $, i.e. $ ( \alcovefont{A}_2, \alcovefont{B}_2 ) $ is laterally-conjugate to $ ( \alcovefont{A}_1, \alcovefont{B}_1 ) $ via $ s_1 $, and therefore also laterally-conjugate to $ ( \basealcove, \alcovefont{B} ) $. In these circumstances, Induction Lemma implies that $ ( ( \alcovefont{B}_0, \alcovefont{B}_1, \alcovefont{B}_2 ), \alcovefont{A}_2, H_2 ) $ is an Umbrella: - the non-numbered hypotheses are again true by choice, the status of hypothesis (\[ENUMvertexinchamber\]) has not changed, and hypothesis (\[ENUMawayfrombase\]) is true by choice of $ H_2 $ for the same reason as before, - hypothesis (\[ENUMinductionhypoth\]) is known by the previous iteration, - hypothesis (\[ENUMnotbasewall\]) is true by choice of $ H_2 $ because $ \alcovefont{A}_2, \alcovefont{A}_3 \subset \basechamber $, and - hypothesis (\[ENUMcontradictionhypoth\]) is supplied by the assumption that $ \{ \alcovefont{A}_1, \alcovefont{B}_1 \} $ did not have the Direct Diamond Property for $ s_1 $ (see previous paragraph: by choice $ K = \alcovefont{B}_1 \vert \alcovefont{B}_2 $). The above induction shows that if $ n \in \N $ and $ \{ \alcovefont{A}_i, \alcovefont{B}_i \} $ does not have the Direct Diamond Property for all $ i \leq n $ then $ ( \alcovefont{A}_i, \alcovefont{B}_i ) $ is laterally-conjugate to $ ( \basealcove, \alcovefont{B} ) $ for all $ i \leq n $, and in particular, $ d ( \alcovefont{A}_i, \alcovefont{B}_i ) = ( \basealcove, \alcovefont{B} ) = \ell ( w ) $ for all $ i \leq n $. But Lemma \[Lgallerytoinfinity\] says that $ d ( \alcovefont{A}_i, \alcovefont{B}_i ) \rightarrow \infty $, so there must exist $ i \in \N $ such that $ \{ \alcovefont{A}_i, \alcovefont{B}_i \} $ has the Direct Diamond Property. If $ i \in \N $ is the smallest such index then $ ( \basealcove, \alcovefont{B} ) $ is laterally-conjugate to a pair with the Direct Diamond Property, as desired. \[Rindependenceofchamber2\] Similar to the “dominant case” Proposition \[PdominantnontranslationshaveDCP\], the choices of vertex, alcove, and chamber are notationally convenient but otherwise totally unimportant to the conclusion of Proposition \[PantidominantnontranslationshaveCP\]. In other words, if $ \alcovefont{A} $ and $ \alcovefont{B} $ are NT-compatible marked alcoves and $ \vertexsymbol_{\alcovefont{B}} \in \closedweylchambersymbol_{\alcovefont{A}}^{\operatorname{opp}} $ then $ \{ \alcovefont{A}, \alcovefont{B} \} $ has the Diamond Property. In type A, there is very simple proof that if $ w ( \basealcove ) \subset \basechamber^{\operatorname{opp}} $ then $ w $ has the Diamond Property using at most one lateral-conjugation. Because of the extreme symmetry of type A, the inverse image $ s_{\operatorname{aff}}^{-1} ( \basechamber ) $ consists of $ \basechamber^{\operatorname{opp}} $ together with all alcoves $ \alcovefont{B} $ such that $ \overline{\alcovefont{B}} \cap \overline{\weylchambersymbol}^{\operatorname{opp}} \neq \emptyset $. Because of this, if $ w $ does not already have the Direct Diamond Property, $ s_{\operatorname{aff}} $ laterally conjugates $ w $ into the dominant chamber, in which case Proposition \[PdominantnontranslationshaveDCP\] applies. Of course, this idea fails in (almost?) every other type. Case: the intermediate chambers {#SSintermediatecase} ------------------------------- I now prove that the general case can, at worst, be reduced to the anti-dominant case, Proposition \[PantidominantnontranslationshaveCP\]: Suppose $ \alcovefont{A} $ and $ \alcovefont{B} $ are marked alcoves. If $ \alcovefont{A} $ and $ \alcovefont{B} $ are NT-compatible then $ \{ \alcovefont{A}, \alcovefont{B} \} $ has the Diamond Property. In particular, if $ w \in \widetilde{W} $ and $ w \notin \translationsubgroup $ then $ w $ has the Diamond Property. View (sequentially!) Figures \[example-g2-2ndinductionlemma-00\] to \[example-g2-2ndinductionlemma-01\] for a picture of the iteration used in this proof. Let $ \mathcal{S} $ be the set of all $ s \in \Delta_{\operatorname{aff}} $ such that if $ H \defeq \labelingsymbol_{\alcovefont{A}} ( s ) $ then the following three properties are true simultaneously: $ \vertexsymbol_{\alcovefont{A}} \in H $, both $ \alcovefont{A} $ and $ \alcovefont{B} $ are on the same side of $ H $, and $ \vertexsymbol_{\alcovefont{B}} \notin H $. If $ \mathcal{S} = \emptyset $ then by definition for every wall $ H $ of $ \alcovefont{A} $ containing $ \vertexsymbol_{\alcovefont{A}} $ either $ H $ separates $ \alcovefont{A} $ from $ \alcovefont{B} $ or $ \vertexsymbol_{\alcovefont{B}} \in H $. In this case, $ \vertexsymbol_{\alcovefont{B}} \in \overline{\weylchambersymbol}_{\alcovefont{A}}^{\operatorname{opp}} $ and Proposition \[PantidominantnontranslationshaveCP\] applies. So assume that $ \mathcal{S} \neq \emptyset $ and let $ s \in \mathcal{S} $ be arbitrary. Let $ ( \alcovefont{B}_0, \ldots, \alcovefont{B}_d ) $ be a gallery realizing the distance $ d ( \vertexsymbol_{\alcovefont{B}}, \weylchambersymbol_{\alcovefont{A}}^{\operatorname{opp}} ) $ (see §\[SSgalleries\] for this notion of distance). By definition, this means that $ \vertexsymbol_{\alcovefont{B}} \in \overline{\alcovefont{B}}_0 $, the gallery is minimal, and $ \alcovefont{B}_d \subset \weylchambersymbol_{\alcovefont{A}}^{\operatorname{opp}} $ (note that $ \alcovefont{B}_0 \neq \alcovefont{B} $ is possible). Set $ H \defeq \labelingsymbol_{\alcovefont{A}} ( s ) $. By definition of $ \mathcal{S} $, $ H $ separates $ \alcovefont{B} $ from every alcove in $ \weylchambersymbol_{\alcovefont{A}}^{\operatorname{opp}} $ and $ \vertexsymbol_{\alcovefont{B}} \notin H $ so $ H $ also separates $ \alcovefont{B}_0 $ from every alcove in $ \weylchambersymbol_{\alcovefont{A}}^{\operatorname{opp}} $. Therefore, $ H $ must be an intermediate wall of the gallery $ ( \alcovefont{B}_0, \ldots, \alcovefont{B}_d ) $ (see Lemma 5.1.5 of [@BGW]). Let $ 0 \leq j < d $ be the index such that $ H = \alcovefont{B}_j \vert \alcovefont{B}_{j+1} $. Then the sequence of alcoves $ ( \alcovefont{B}_0, \ldots, \alcovefont{B}_j, s_H ( \alcovefont{B}_{j+2} ), \ldots, s_H ( \alcovefont{B}_d ) ) $ is a gallery. This is obviously a gallery “from $ \vertexsymbol_{\alcovefont{B}} $ to $ \weylchambersymbol_{\alcovefont{A}^s}^{\operatorname{opp}} $” and has fewer than $ d $ alcoves. By choice of $ s $ and the compatibility hypothesis, $ \vertexsymbol_{\alcovefont{B}^s} = \vertexsymbol_{\alcovefont{B}} $. Altogether, $ d ( \vertexsymbol_{\alcovefont{B}^s}, \weylchambersymbol_{\alcovefont{A}^s}^{\operatorname{opp}} ) < d ( \vertexsymbol_{\alcovefont{B}}, \weylchambersymbol_{\alcovefont{A}}^{\operatorname{opp}} ) $. On the other hand, by choice of $ s $, Lemma \[Lgeometriclength\] implies that $ d ( \alcovefont{A}^s, \alcovefont{B}^s ) \geq d ( \alcovefont{A}, \alcovefont{B} ) $. Since these distances $ d ( \vertexsymbol_{}, \weylchambersymbol_{}^{\operatorname{opp}} ) $ are $ \N $-valued, this means that one may iterate the previous process until a pair of alcoves $ ( \alcovefont{A}^{\prime}, \alcovefont{B}^{\prime} ) $ is constructed which is laterally-conjugate to $ ( \alcovefont{A}, \alcovefont{B} ) $ and such that either $ ( \alcovefont{A}^{\prime}, \alcovefont{B}^{\prime} ) $ has the Direct Diamond Property or $ \vertexsymbol_{\alcovefont{B}^{\prime}} \in \overline{\weylchambersymbol}_{\alcovefont{A}^{\prime}}^{\operatorname{opp}} $, in which case Proposition \[PantidominantnontranslationshaveCP\] applies. If $ w \in \finiteweylgroup \subset \widetilde{W} $ then $ w \notin \Lambda $ if and only if $ w \neq 1 $, and it is easy to show that both $ \ell ( w s_{\operatorname{aff}} ) > \ell ( w ) $ and $ \ell ( s_{\operatorname{aff}} w ) > \ell ( w ) $ directly: combine the Exchange Property of the Coxeter group $ ( W_{\operatorname{aff}}, \Delta_{\operatorname{aff}} ) $ with the fact that all reduced expressions for a single element must use the same subset of $ \Delta_{\operatorname{aff}} $ (see Proposition 7 in Ch IV §1 no. 8 of [@bourbaki]) to conclude that neither length can decrease. It is tempting to think that such $ w $ always have the Direct Diamond Property realized by $ s_{\operatorname{aff}} $ but this is not always true: in the affine Weyl group $ \Caff{2} $ (or $ \Gaff{2} $), there exists $ s \in \basereflections $ such that $ s \cdot s_{\operatorname{aff}} = s_{\operatorname{aff}} \cdot s $, so $ s_{\operatorname{aff}} w s_{\operatorname{aff}} = w $ for $ w \defeq s \in \finiteweylgroup $. Application to Hecke Algebras {#Sapplicationtoheckealgebras} ============================= Hecke algebras on quasi-Coxeter groups {#SSdefofheckealgebra} -------------------------------------- Fix a function $ q : \Delta_{\operatorname{aff}} \rightarrow \N $ which is invariant under conjugation by $ \widetilde{W} $. In the rest of this section §\[Sapplicationtoheckealgebras\], I assume given a $ \C $-algebra $ \heckefont{H} $ which, as a $ \C $-vector space, has a basis of elements $ \heckefont{T}_w $ indexed by all $ w \in \widetilde{W} $. Further, denoting the ring operation by $ * $, I assume that the following Iwahori-Matsumoto identities are true in $ \heckefont{H} $: for all $ w \in \widetilde{W} $ and $ s \in \Delta_{\operatorname{aff}} $, $$\begin{aligned} \heckefont{T}_s * \heckefont{T}_w &= \begin{cases} \heckefont{T}_{sw} &\text{ if } \ell ( s w ) > \ell ( w ) \\ ( q(s) - 1 ) \heckefont{T}_w + q(s) \heckefont{T}_{sw} &\text{ if } \ell ( s w ) < \ell ( w ) \end{cases} \text{ (left-handed)} \\ \heckefont{T}_w * \heckefont{T}_s &= \begin{cases} \heckefont{T}_{ws} &\text{ if } \ell ( w s ) > \ell ( w ) \\ ( q(s) - 1 ) \heckefont{T}_w + q(s) \heckefont{T}_{ws} &\text{ if } \ell ( w s ) < \ell ( w ) \end{cases} \text{ (right-handed)}\end{aligned}$$ Note that because of the way that the length function $ \ell $ was extended to $ \widetilde{W} $, if $ \tau \in \Omega $ and $ w \in \widetilde{W} $ then $ \heckefont{T}_{w \tau} = \heckefont{T}_w * \heckefont{T}_{\tau} $. If $ h \in \heckefont{H} $ then denote by $ h_w $ the coefficient of $ \heckefont{T}_w $ in the linear combination of $ h $ with respect to this basis. If $ h \in \heckefont{H} $ and $ h_w \neq 0 $ then $ w $ is said to support $ h $. It is not difficult to show, and I do so in the article [@roro] using ingredients from the Appendix to [@PRH], that any Iwahori-Hecke algebra $ \heckefont{H} $ of any connected reductive affine algebraic $ F $-group is of the form described above. Therefore, the results of this section §\[Sapplicationtoheckealgebras\] apply to Iwahori-Hecke algebras. If greater generality is desired, one can use a pair $ a, b : \Delta_{\operatorname{aff}} \rightarrow \C $ of parameter systems and a “generic algebra” as in §7.1 of [@humphreys]. I will need the following slight extension of a well-known property of Coxeter groups: \[Lhumphreys\] Fix $ w \in \widetilde{W} $ and $ s, t \in \Delta_{\operatorname{aff}} $. If $ \ell ( s w t ) = \ell ( w ) $ and $ \ell ( s w ) = \ell ( w t ) $ then $ s w t = w $. When $ \Omega = \{ 1 \} $, this is exactly Lemma in §7.2 of [@humphreys]. The general case follows immediately from this since $ \Omega $ permutes $ \Delta_{\operatorname{aff}} $ and $ \ell $ factors through $ W_{\operatorname{aff}} $. Equations defining the center {#SScentralityequations} ----------------------------- Denote by $ Z ( \heckefont{H} ) $ the center of the ring $ \heckefont{H} $. Fix $ h \in \heckefont{H} $. It is clear from the Iwahori-Matsumoto relations that $ h \in Z ( \heckefont{H} ) $ if and only if $ h * \heckefont{T}_s = \heckefont{T}_s * h $ and $ h * \heckefont{T}_{\tau} = \heckefont{T}_{\tau} * h $ for all $ s \in \Delta_{\operatorname{aff}} $ and $ \tau \in \Omega $. Fix $ s \in \Delta_{\operatorname{aff}} $. For each $ x \in \widetilde{W} $, one can use the left-handed Iwahori-Matsumoto relation to compute that the coefficient of $ \heckefont{T}_x $ in $ \heckefont{T}_s * h $ is $$\begin{aligned} q(s) h_{ s x } &\text{ if } \ell ( s x ) > \ell ( x ) \\ h_{ s x } + ( q(s) - 1 ) h_x &\text{ if } \ell ( s x ) < \ell ( x )\end{aligned}$$ Similarly, one can use the right-handed Iwahori-Matsumoto relation to compute that the coefficient of $ \heckefont{T}_x $ in $ h * \heckefont{T}_s $ is $$\begin{aligned} q(s) h_{ x s } &\text{ if } \ell ( x s ) > \ell ( x ) \\ h_{ x s } + ( q(s) - 1 ) h_x &\text{ if } \ell ( x s ) < \ell ( x )\end{aligned}$$ It is obvious from the Iwahori-Matsumoto identities that $ h * \heckefont{T}_{\tau} = \heckefont{T}_{\tau} * h $ if and only if $ h_{x \tau} = h_{\tau x} $ for all $ x \in \widetilde{W} $. It follows that the center $ Z ( \heckefont{H} ) $ is the $ \C $-subspace of vectors $ h \in \heckefont{H} $ whose Iwahori-Matsumoto coefficients $ h_x $ solve the (infinite) linear system consisting of the equation $ h_{x \tau} = h_{\tau x} $ for each pair $ ( x, \tau ) \in \widetilde{W} \times \Omega $ together with the appropriate equation from $$\begin{aligned} \nonumber q(s) h_{ s x } = q(s) h_{ x s } &\text{ if } \ell ( s x ), \ell ( x s ) > \ell ( x ) \\ \label{Ecenterequations_us_lowest} q(s) h_{ s x } = h_{ x s } + ( q(s) - 1 ) h_x &\text{ if } \ell ( s x ) > \ell ( x ) > \ell ( x s ) \\ \nonumber h_{ s x } + ( q(s) - 1 ) h_x = q(s) h_{ x s } &\text{ if } \ell ( s x ) < \ell ( x ) < \ell ( x s ) \\ \label{Ecenterequations_bothdecrease} h_{ s x } = h_{ x s } &\text{ if } \ell ( s x ), \ell ( x s ) < \ell ( x )\end{aligned}$$ for each pair $ ( x, s ) \in \widetilde{W} \times \Delta_{\operatorname{aff}} $. These equations appeared already in §3 of [@haines] for affine Hecke algebras on reduced root data. Length-filtration and dimensions {#SSdimensionbounds} -------------------------------- Recall that $ \Omega ( w ) $ denotes the projection of $ w \in \widetilde{W} $ into $ \Omega $, and that the set of $ \finiteweylgroup $-conjugacy classes in $ \translationsubgroup $ is denoted by $ \translationsubgroup / \finiteweylgroup $. Fix $ L \in \N $ and $ \tau \in \Omega $. Define $ Z_{L,\tau} ( \heckefont{H} ) $ to be the set of all $ z \in Z ( \heckefont{H} ) $ such that $ z_w = 0 $ if either $ \ell ( w ) > L $ or $ \Omega ( w ) \neq \tau $. Note that each $ Z_{L,\tau} ( \heckefont{H} ) $ is a finite-dimensional $ \C $-subspace of $ Z ( \heckefont{H} ) $ and that $ Z ( \heckefont{H} ) $ is the union of all $ Z_{L,\tau} ( \heckefont{H} ) $. Recall that if $ \mathcal{O} \in \translationsubgroup / \finiteweylgroup $ then $ \ell $ is constant on $ \mathcal{O} $ and define $ \ell ( \mathcal{O} ) $ to be this constant length. It follows that any two $ t, t^{\prime} \in \mathcal{O} $ are laterally-conjugate. Fix $ L \in \N $ and $ \tau \in \Omega $. Define $ N_{L, \tau} $ to be the total number of conjugacy classes $ \mathcal{O} \in \translationsubgroup / \finiteweylgroup $ such that $ \ell ( \mathcal{O} ) \leq L $ and $ \Omega ( t ) = \tau $ for all $ t \in \mathcal{O} $. The following two lemmas show how lateral-conjugacy and the diamond property are related to centers of Hecke algebras: \[Llateralconjugacyimpliesdependence\] Suppose $ z \in Z ( \heckefont{H} ) $. If $ w \in \widetilde{W} $ is laterally-conjugate to $ w^{\prime} $ then $ z_w = z_{w^{\prime}} $. If $ s \in \Delta_{\operatorname{aff}} $ is such that $ \ell ( s w s ) = \ell ( w ) $ then either $ \ell ( s w ) < \ell ( w ) < \ell ( w s ) $ or $ \ell ( w s ) < \ell ( w ) < \ell ( s w ) $. Choosing $ x \defeq w s $ in the former case and $ x \defeq s w $ in the latter case, centrality equation (\[Ecenterequations\_bothdecrease\]) implies that $ z_{w} = z_{s w s} $, and the claim follows immediately from this. \[Ldiamondpropertyimpliesdependence\] Suppose $ z \in Z ( \heckefont{H} ) $. If $ w \in \widetilde{W} $ has the Diamond Property then there exist $ u, v \in \widetilde{W} $ satisfying $ \ell ( u ) > \ell ( v ) > \ell ( w ) $ such that $ z_w $ is a $ \C $-linear combination of $ z_u $ and $ z_v $. By Lemma \[Llateralconjugacyimpliesdependence\], I may assume that $ w $ has the Direct Diamond Property. Let $ s \in \Delta_{\operatorname{aff}} $ be the element realizing the property. By basic Coxeter theory, $ \ell ( s w ) = \ell ( w ) + 1 = \ell ( w s ) $ and it is true that either $ \ell ( s w s ) = \ell ( w ) + 2 $ or $ \ell ( s w s ) = \ell ( w ) $. If it were true that $ \ell ( s w s ) = \ell ( w ) $ then by Lemma \[Lhumphreys\] it would be true that $ s w s = w $, but this is explicitly prohibited by the Direct Diamond Property. Therefore, $ \ell ( s w s ) > \ell ( s w ) = \ell ( w s ) > \ell ( w ) $. Choosing $ x \defeq w s $ and applying centrality equation (\[Ecenterequations\_us\_lowest\]) proves that $ z_w $ is a linear combination of $ z_{ws} $ and $ z_{sws} $, as desired. Materially, both lemmas \[Llateralconjugacyimpliesdependence\] and \[Ldiamondpropertyimpliesdependence\] appeared already as Lemma 3.1 of [@haines]. \[Pdimensionbounds\] Fix $ L \in \N $ and $ \tau \in \Omega $. $$\dim_{\C} ( Z_{L,\tau} ( \heckefont{H} ) ) \leq N_{L,\tau}$$ Suppose $ w \in \widetilde{W} $ and $ w \notin \translationsubgroup $. Consider the linear system defining $ Z ( \heckefont{H} ) $ as a $ \C $-subspace of $ \heckefont{H} $. By applying the Main Theorem and Lemma \[Ldiamondpropertyimpliesdependence\] repeatedly, one can express the variable $ h_w $ as a $ \C $-linear combination of variables $ h_x $ such that either $ \ell ( x ) > L $ or $ x \in \translationsubgroup $. Since $ Z_{L,\tau} ( \heckefont{H} ) $ is the $ \C $-subspace of $ Z ( \heckefont{H} ) $ defined by the additional equations $ h_x = 0 $ for all $ x \in \widetilde{W} $ such that either $ \ell ( x ) > L $ or $ \Omega ( x ) \neq \tau $, it follows that $ \dim_{\C} ( Z_{L,\tau} ( \heckefont{H} ) ) $ is at most the total number of $ t \in \translationsubgroup $ such that $ \ell ( t ) \leq L $ and $ \Omega ( t ) = \tau $. On the other hand, if $ \mathcal{O} \in \translationsubgroup / \finiteweylgroup $ and $ t, t^{\prime} \in \mathcal{O} $ then $ t $ is laterally-conjugate to $ t^{\prime} $ and $ \Omega ( t ) = \Omega ( t^{\prime} ) $, so the dimension bound now follows from Lemma \[Llateralconjugacyimpliesdependence\]. Note one extra detail from the proof: if $ z \in Z_{L,\tau} ( \heckefont{H} ) $ and both $ w \notin \translationsubgroup $ and $ \ell ( w ) = L $ then $ z_w = 0 $. Suppose that for each conjugacy class $ \mathcal{O} \in \translationsubgroup / \finiteweylgroup $ there is an element $ z_{ \mathcal{O} } \in Z ( \heckefont{H} ) $ such that $ \ell ( w ) \leq \ell ( \mathcal{O} ) $ for all $ w \in \widetilde{W} $ supporting $ z_{ \mathcal{O} } $ and such that $ \Omega ( w ) $ is the same for all $ w \in \widetilde{W} $ supporting $ z_{ \mathcal{O} } $. If $ \{ z_{ \mathcal{O} } \}_{ \mathcal{O} \in \translationsubgroup / \finiteweylgroup } $ is linearly-independent then it is a basis for $ Z ( \heckefont{H} ) $. Fix $ L \in \N $ and $ \tau \in \Omega $. Consider only those $ z_{ \mathcal{O} } $ for which $ \ell ( \mathcal{O} ) \leq L $ and for which the uniform $ \Omega $-component of those $ w $ supporting $ z_{ \mathcal{O} } $ is $ \tau $. By hypothesis, the set of all such $ z_{ \mathcal{O} } $ is a linearly-independent subset of $ N_{L,\tau} $ vectors in the subspace $ Z_{L,\tau} ( \heckefont{H} ) $. By Proposition \[Pdimensionbounds\], this set must also span $ Z_{L,\tau} ( \heckefont{H} ) $. Since $ Z ( \heckefont{H} ) $ is the union over all pairs $ ( L, \tau ) $ of the subspaces $ Z_{L,\tau} ( \heckefont{H} ) $, the claim follows. Pictures {#Sfigures} ======== ![The Main Theorem does not follow directly from Proposition \[PdominantnontranslationshaveDCP\] (the dominant case). The blue alcove in the center is the base alcove $ \basealcove $ and the red alcoves surrounding it constitute a full lateral-conjugacy class. The light blue/pink cones are the unique dominant Weyl chamber containing the various alcoves (uniqueness is due to the fact that in this $ \Gaff{2} $ example each alcove contains only one special vertex in its closure).[]{data-label="example-g2-annoyingorbit"}](figure1) ![Sample initial situation in the proof of Proposition \[PantidominantnontranslationshaveCP\]. The purple alcove is $ \basealcove $ and the red alcove is $ \alcovefont{B} $. The light blue cone is the dominant Weyl chamber $ \basechamber $ and the pink cone is the chamber $ \weylchambersymbol_{\alcovefont{B}} $.[]{data-label="example-g2-inductionlemma00"}](figure2a) ![After 3 iterations of Induction Lemma. The purple alcove is $ \alcovefont{A}^{\prime} $ and all possibilities for $ H^{\prime} $ (only one in this case) are also purple. All other alcoves in the gallery $ \alcovefont{A}_{\bullet} $ are blue, in particular the alcove $ \alcovefont{A} $ adjacent to $ \alcovefont{A}^{\prime} $. The blue wall is $ H $. The red alcoves constitute the gallery $ \alcovefont{B}_{\bullet} $. Observe that the current $ H $ is always (one of) the previous $ H^{\prime} $.[]{data-label="example-g2-inductionlemma03"}](figure2b) ![After 4 iterations of Induction Lemma. Observe that (for the first time) the purple alcove and the nearest red alcove (its lateral conjugate) have the Direct Diamond Property. Nonetheless, the conclusion of Induction Lemma remains true for several more iterations. An important observation is that the conclusion of Induction Lemma must eventually fail because the Diamond Property is true.[]{data-label="example-g2-inductionlemma04"}](figure2c) ![Situation in which $ \mathcal{S} \neq \emptyset $ and $ d ( \vertexsymbol_{ \alcovefont{B} }, \weylchambersymbol_{\alcovefont{A}}^{\operatorname{opp}} ) > 0 $ in the proof of the Main Theorem (although in this particular example, the pair of alcoves already has the Direct Diamond Property realized by $ s_{\operatorname{aff}} $ so no action is necessary). The blue alcove is $ \alcovefont{A} $, the red alcove is $ \alcovefont{B} $, and the grey cone is $ \weylchambersymbol_{\alcovefont{A}}^{\operatorname{opp}} $. The black outline is merely a visual aid.[]{data-label="example-g2-2ndinductionlemma-00"}](figure3a) ![After 1 iteration of the process described in the proof of the Main Theorem. The value $ d ( \vertexsymbol_{ \alcovefont{B} }, \weylchambersymbol_{\alcovefont{A}}^{\operatorname{opp}} ) $ is now $ 0 $ and $ \vertexsymbol_{ \alcovefont{B} } \in \overline{ \weylchambersymbol }_{\alcovefont{A}}^{\operatorname{opp}} $. In this particular example, it is possible to laterally conjugate once more to arrange $ \alcovefont{B} \subset \weylchambersymbol_{\alcovefont{A}}^{\operatorname{opp}} $, but this is not always possible. The light blue/pink alcove merely represents the previous position of the blue/red alcove.[]{data-label="example-g2-2ndinductionlemma-01"}](figure3b)	{ "pile_set_name": "ArXiv" }
--- abstract: 'We studied the critical state stability in a large cubic sample of a single crystalline [La$_{1.85}$Sr$_{0.15}$CuO$_4$]{} for different sample orientations with respect to the external magnetic field as well as the dynamics of the flux jumps. It is shown that thermomagnetic avalanches develop in dynamic conditions characterized by significantly lower magnetic diffusivity than the thermal one. In this case, critical state stability depends strongly on cooling conditions. We compared predictions of the isothermal model and of the model for the weakly cooled sample with experimental results. In both models, the field of the first flux jump decreases with an increase of sweep rate of the external magnetic field. We also investigated the influence of external magnetic field on the dynamics of the following stages of the thermomagnetic avalanche. It is shown that the dynamics of the flux jumps is correlated with the magnetic diffusivity proportional to the flux flow resistivity.' address: - '$^1$Institute of Physics, Polish Academy of Sciences, Al. Lotników 32/46, 02-668 Warsaw, Poland.' - '$^2$Institute for Physics and Engineering, NASU, 72 ul. R. Luxemburg, 83114, Donetsk, Ukraine.' - '$^3$Instituto de Física, Benemérita Universidad Autónoma de Puebla, Apdo. Post. J-48 Puebla, 72570 Pue., Mexico.' author: - 'A Nabia[ł]{}ek$^1$, A Wiśniewski$^1$, V V Chabanenko$^2$, S V Vasiliev$^2$, T V Tsvetkov$^2$ and F Pérez-Rodríguez$^3$' title: 'Influence of crystal anisotropy on the critical state stability and flux jumps dynamics in a single crystal of La$_{1.85}$Sr$_{0.15}$CuO$_4$' --- INTRODUCTION ============ At certain conditions, the critical state of a type-II superconductor may become unstable [@Mints81], then thermomagnetic avalanche may develop. During the thermomagnetic avalanche, temperature of the superconducting sample increases rapidly and large amount of the magnetic flux enters into the sample volume. Thermomagnetic avalanches, called flux jumps, are problematic from a viewpoint of practical applications of superconductors, because they may drive a superconducting sample into a normal or, at least, into a resistive state. The stability of critical state depends strongly on a relation between the thermal ($D_\mathrm{t}$) and magnetic ($D_\mathrm{m}$) diffusivity. $D_\mathrm{t}=\frac{\kappa}{c_\mathrm{V}}$, where $\kappa$ is the thermal conductivity, and $c_\mathrm{V}$ is the specific heat at constant volume. $D_\mathrm{m}=\frac{\rho}{\mu_0}$. In the flux flow regime, $\rho=\rho_\mathrm{ff}=\rho_\mathrm{n}(\frac{H}{H_\mathrm{c2}})$, where $\rho_\mathrm{ff}$ is the flux flow resistivity, $\rho_\mathrm{n}$ is the normal state resistivity, $H_\mathrm{c2}$ is the second critical field and $\mu_0$ is the magnetic permeability of vacuum. In the case of hard conventional superconductors, usually $\tau=\frac{D_\mathrm{t}}{D_\mathrm{m}}<<1$. Such conditions are called “locally adiabatic”. The most commonly studied parameter, which determines the stability of the critical state is a field of the first (after cooling the sample in zero magnetic field) flux jump, $H_\mathrm{fj1}$. This parameter is closely correlated with the critical penetration depth, $L_\mathrm{c}$. If the critical current density, $j_\mathrm{c}$, does not depend on the magnetic field, $L_\mathrm{c}=\frac{H_\mathrm{fj1}}{j_\mathrm{c}}$. This parameter is very important from a viewpoint of the practical applications of the superconductors, because it is possible to avoid the thermomagnetic avalanches, if the width of the superconducting sample is smaller than $2L_\mathrm{c}$. We can also express the condition of the stability of the critical state in a form: $H_\mathrm{fj1}>H_\mathrm{p}$, where $H_\mathrm{p}$ is the field of full penetration, $H_\mathrm{p}=j_\mathrm{c}a$, and $a$ is half-width of the investigated sample. In the simplest approximations, the sample is usually assumed to be of a shape of an infinite slab, of width $2a$, and an external magnetic field is applied parallel to its surface. In locally adiabatic conditions, the field of the first flux jump can be expressed by a formula $$\begin{aligned} H_\mathrm{fj1} &=& \sqrt{ \frac {2c_\mathrm{V}j_{c}} {\mu_0 \|\frac{\partial j_{c}}{\partial T}\|} } \label{eq:one}\end{aligned}$$ The thermal parameter, which determines the critical state stability in locally adiabatic conditions, is the specific heat $c_\mathrm{V}$. In locally adiabatic conditions, the stability of the critical state is not influenced by the cooling conditions at the surface of the superconducting sample. The situation is different in a case when $\tau >> 1$. Such conditions usually occur in superconducting composites and are called “dynamic”. In dynamic conditions, the thermal parameter which determines the stability of the critical state is the thermal conductivity, $\kappa$, or the thermal boundary conductivity, $h$ [@Swarz89]. Some problems connected with the critical state stability in high temperature superconductors (HTSC) are not clarified yet. The problems are connected with the complex structure of these materials. The analysis of the critical state stability in HTSC is usually performed within approximations developed for conventional superconductors [@Mints81], which sometimes lead to wrong results. Because of the increasing number of applications of HTSC, it is very important to understand all aspects of the thermomagnetic avalanche development in these materials. One of the specific features of HTSC is their strong crystallographic anisotropy. Experimental data show that high temperature superconductors are usually more stable against flux jumping than conventional superconductors. The larger stability of the high temperature superconductors against flux jumping can be correlated with the flux creep phenomenon, which is usually strong in these materials [@McHenry92; @Gerber93]. In studies of the magnetization of superconductors, flux creep manifests itself as a relaxation of the magnetic moment. The process of magnetic relaxation reflects strongly nonlinear current-voltage characteristics of superconducting sample. These characteristics can be approximated by the formula $$\begin{aligned} j(E) &=& j_\mathrm{c}+\frac{j_\mathrm{c}}{n} \ln \left( \frac{E}{E_0} \right), \label{eq:two}\end{aligned}$$ where $E_0$ is the voltage criterion at which the critical current density, $j_\mathrm{c}$, is defined, and $n$ is a dimensionless parameter. One usually assumes $E_0=10^{-4}$ V/m. In this case $n>>1$. The influence of nonlinear current-voltage characteristics on the critical state stability, in the case of a weakly cooled sample (for $Bi=\frac{ah}{\kappa} \ll 1$, $Bi$ is so-called Biot number), was analyzed in Ref.[@Mints96]. It was shown that in such conditions the critical state stability depends on the external magnetic sweep rate,$\frac{\partial{H_\mathrm{ext}}}{\partial{t}}$. $$\begin{aligned} H_\mathrm{fj1}^\mathrm{w} &=& j_\mathrm{c}\sqrt{\frac{2\mu_0 h}{n\left\|\frac{\partial j_\mathrm{c}}{\partial T}\right\| \frac{\partial H_\mathrm{ext}}{\partial t}}} \label{eq:three}\end{aligned}$$ In previous works [@Nab03; @Nab06], we studied the influence of external magnetic sweep rate on the critical state stability in crystals of high temperature [Bi$_2$Sr$_2$CaCu$_2$O$_{8+\delta}$]{} superconductor. It is extremely difficult to obtain thick crystals of [Bi$_2$Sr$_2$CaCu$_2$O$_{8+\delta}$]{}, and we have shown [@Nab06] that the thickness of the crystal strongly influences the critical state stability. On the other hand, it is relatively easy to obtain large cubic samples of [La$_{1.85}$Sr$_{0.15}$CuO$_4$]{}. For this reason, we have chosen this system to study the influence of crystal anisotropy on the critical state stability. A strong influence of cooling condition and a role of crystal anisotropy on the critical state stability in the case of textured YBaCuO were considered in Ref. [@Guillot; @Watanabe]. The effects of thermal insulation on the critical state stability in conventional NbTi superconductor were investigated in Ref. [@Chikaba]. We paid special attention to the dynamics of the flux jumps, which remains poorly understood up to date. This dynamics is also important from a viewpoint of superconductor applications, because a large power of the heat released during short avalanches may, at certain conditions, lead to the destruction of superconducting devices. In the present work, we also studied the dynamics of the magnetic flux changes in a single crystal of [La$_{1.85}$Sr$_{0.15}$CuO$_4$]{} during thermomagnetic avalanches. In particular, we focused on the influence of external magnetic field on dynamics of the following stages of the thermomagnetic avalanche. EXPERIMENT ========== In our investigations, we used a large 3 x 3 x 3 $\mathrm{mm}^3$ single crystal of [La$_{1.85}$Sr$_{0.15}$CuO$_4$]{}, obtained by the floating zone technique. In order to study the influence of the external magnetic sweep rate on the critical state stability, we used the Quantum Design PPMS system with the extraction magnetometer option and the maximal external magnetic field attainable of 9 tesla. After cooling the sample in zero magnetic field, external magnetic field was swept with constant rate from 0 to 9 T. In the following experiments, we varied the sweep rate from the minimal ($10^{-3}$ T/s) to the maximal ($2 \cdot 10^{-2}$ T/s) one attainable in our system. The distance between the following data points taken in each magnetization curve was limited by the sweep rate and the time needed to take each data point. At the maximal sweep rate ($2 \cdot 10^{-2}$ T/s), we were not able to obtain the distance smaller than about 0.1 T. This distance has limited an accuracy of $H_\mathrm{fj1}$, determined in our experiments. In order to determine the current-voltage characteristics of our sample, we performed studies of magnetic moment relaxation. After cooling in zero external magnetic field, external magnetic field was first swept, with minimal sweep rate, up to 0.5 T, and next the relaxation of magnetic moment was registered at constant external magnetic field during approximately 1 hour. The studies of $H_\mathrm{fj1}$, as well as of magnetic relaxation, were performed both for external magnetic field parallel to the $c$-axis and for external magnetic field parallel to the $ab$-plane. The dynamics of the flux jumps was studied in the Cryogenics 12 tesla magnet system with variable temperature insert for external magnetic field parallel to the $c$-axis. A pick-up coil consisted of 6 turns of copper wire wound around the investigated sample and connected to data acquisition board (DAQ) in the computer. Because of small inductance of the applied coil and large internal resistance of DAQ, we were able to register the changes of the voltage with time resolution better that 10$^{-6}$ s. The time resolution of DAQ was about 10$^{-7}$ s. We registered time dependence of the coil voltage during the following flux jumps. Additionally, we used two miniature Hall probes to register magnetic field dependence of the surface self-field, $H_\mathrm{self-surf}$, of the investigated sample. One of the probes was put in the centre of the surface of the investigated sample. The second probe measured external magnetic field. The differential signal from the two probes was proportional to $H_\mathrm{self-surf}$. The cryogenic Hall probes were made from tin doped InSb films. The probe was supplied with the current of 10 mA. The sensitivity of the Hall probe was about 5 mV/T. The rate of external magnetic field sweep was approximately 1 T/min. In the 12 tesla magnet system, we also performed magnetization measurements using vibrating sample magnetometer (VSM). It is important to note that the heat exchange conditions in both systems used in our experiments were very different. In the PPMS system, the sample was put into a standard PPMS holder, which was surrounded by helium gas with low pressure (about 0.5 Torr). In the 12 tesla magnet system, during the experiments at 2 K and 4.2 K, the sample was immersed in liquid helium. During experiments at higher temperatures, the sample was surrounded by flowing helium gas with approximately atmospheric pressure. The sample was either glued to a quartz plate (studies of the flux jumps dynamics) or put into Teflon holder (VSM). RESULTS ======= ![Virgin magnetization curves taken by extraction magnetometer (PPMS) with the sweep rate of $2 \cdot 10^{-2}$ T/s at different temperatures for both sample orientations studied in our experiments.](Fig1.eps){width="0.95\columnwidth"} ![The changes of the field of the first flux jump, $H_\mathrm{fj1}$, with the sweep rate at different temperatures and for different sample orientations with respect to the external magnetic field, in a double-logarithmic scale, for the 3 x 3 x 3 $\mathrm{mm}^3$ [La$_{1.85}$Sr$_{0.15}$CuO$_4$]{} single crystal surrounded by helium gas (PPMS experiments). Dashed lines show the predictions of Eq. 3, where $H_\mathrm{fj1}$ is inversely proportional to the square root of the sweep rate. ](Fig2.eps){width="0.95\columnwidth"} The range of flux jumps occurrence, the field of the first flux jump as well as the distance between the following flux jumps depends on the sample orientation, temperature, sweep rate and the experimental system. Figure 1 presents some of the virgin magnetization curves taken by extraction magnetometer (PPMS) with the sweep rate of $2 \cdot 10^{-2}$ T/s at different temperatures for the both sample orientations studied in our experiments. Magnetization curves on the left side in Fig. 1 are cut-off for the signal amplitude above 20 emu, because of a limitation of the signal in extraction magnetometer. Nevertheless, in all our PPMS experiments, we were able to determine the value of $H_\mathrm{fj1}$. Temperature range of the flux jumps occurence depends on the sample orientation. For the external magnetic field parallel to the $c$-axis, flux jumps occur even at 9.5 K (see Fig. 1c). For the external magnetic field parallel to the $ab$-palne, flux jumps disappear already at 5 K (see Fig. 1f). With the increase of temperature, we observed an increase of $H_\mathrm{fj1}$ and of the distance between the following flux jumps. We also observed a decrease of $H_\mathrm{fj1}$ with an increase of external magnetic field sweep rate (see. Fig. 2) Hence, we observed the flux jumps in the largest temperature range, for the highest sweep rate ($2 \cdot 10^{-2}$ T/s). At the lowest sweep rate, $10^{-3}$ T/s, we did not observe any flux jumps at the temperatures higher than 4 K, for the external magnetic field parallel to the $c$-axis, and higher than 2.2 K, for the external magnetic field parallel to the $ab$-plane, respectively. Figure 2 shows the changes of $H_\mathrm{fj1}$ with the sweep rate at different temperatures and for different sample orientations - the data were taken from PPMS experiments. The results are shown in a double-logarithmic scale. If $H_\mathrm{fj1}$ changes with the sweep rate according to the power law, we expect a straight line in such plot. Dashed lines show the predictions of Eq. 3, where $H_\mathrm{fj1}$ is inversely proportional to the square root of the sweep rate. The experiments in the 12 tesla magnet (with Hall probes or VSM) were all performed with the external magnetic field sweep rate approximately the same as the highest sweep rate in the PPMS system. However, we observed here the flux jumps only at the lowest temperature of about 2.2 K and for the external magnetic field parallel to the $c$-axis (see Fig. 3a - experiment with the Hall probes). No jumps were present neither at 4.2 K nor at higher temperatures. ![a) Hysteresis loops of the surface self-field, $H_\mathrm{self-surf}$, measured by the Hall probes in 12 tesla magnet system at 2.2 K and 4.2 K - sample immersed in liquid helium. b,c) Time dependence of the voltage, per one turn of the coil, registered during the following flux jumps. The numbers correspond to the numbers denoting successive flux jumps presented in part (a). Initial voltages of the following flux jumps were shifted, in order to present all data in one figure. $t_\mathrm{bm}$ and $t_\mathrm{am}$ describe duration of the two characteristic stages of the flux jump. ](Fig3.eps){width="0.95\columnwidth"} The relaxation of the magnetic moment, measured in external magnetic field $\mu_0 H_\mathrm{ext}=0.5$ T, was approximately logarithmic in time, which was consistent with the model of thermally activated motion of bundles of vortices [@Anderson]. In Figure 3a, on can see the hysteresis loops of the surface self-field, $H_\mathrm{self-surf}$, measured by the Hall probes in 12 tesla magnet system at the temperature of about 2.2 K (solid line) and 4.2 K (dotted line), for the external magnetic field parallel to the $c$-axis. At 2.2 K, one can see giant jumps of the surface self-field caused by the thermomagnetic avalanches. No jumps were present at the temperature of 4.2 K or at higher temperatures. In Figure 3b and 3c, one can see time dependence of the voltage per one turn of the coil wounded around the investigated sample, registered during the succeeding flux jumps. The numbers at the curves correspond to the numbers denoting successive flux jumps presented in Fig. 3a. Typical duration of the flux jumps was of several microseconds. The dependence shown in Fig. 3b corresponds to the lower (ascending) branch of the hysteresis loop. The dependence shown in Fig. 3c corresponds to the upper (descending) branch of the hysteresis loop, shown in Fig. 3a. Before and after each flux jump, the voltage of the coil was close to zero. However, in order to present all the data in one figure, the initial voltages of the following flux jumps were shifted. The voltage of the coil is proportional to the time derivative of the magnetic flux in the investigated sample. We can recognize two characteristic stages of the flux jump. During the first stage, the rate of the magnetic flux changes increases. During the second stage, the rate of the magnetic flux changes decreases. We have denoted duration of these two stages as $t_\mathrm{bm}$ (“before maximum”) and $t_\mathrm{am}$ (“after maximum”), respectively (see Fig. 3b). We have determined duration of the following stages of the flux jump at a noise level, which in our system was about 30 mV. One can see that the voltage of the observed impulses for the ascending branch of the hysteresis loop (Fig. 3b) has the sign opposite to the descending branch (Fig. 3c). This is because, for the ascending branch, the magnetic flux enters into the investigated sample during the succeeding flux jumps, while for the descending branch, we observe flux exit. ANALYSIS ======== The current-voltage characteristics ----------------------------------- In order to compare the models of the critical state stability with our experimental results, it is necessary to know the current-voltage characteristics of the investigated sample [@Mints96]. These characteristics can be determined using the data of the magnetic moment relaxation. In order to obtain these characteristics, we applied the following procedure. We used the model of an infinite slab sample. Additionally, we assumed the screening current density to be the same in the whole sample volume penetrated by the magnetic flux. The time dependent screening current density can be in this case correlated with the magnetization. The electric field at the surface of the slab is proportional to the time derivative of the screening current density. Treating the time as a parameter, one can obtain the current-voltage dependence, $j(E)$. The obtained by the described above procedure current-voltage characteristics, were fitted according to Eq. 2. We applied this procedure for both sample orientations studied in our experiments. Our fitting procedure gave the critical current density of an order of $10^8$ A/$\mathrm{m}^2$ and $10^9$ A/$\mathrm{m}^2$ and the $n$-parameter (see Eq. 2) of about 40 and 30 for the external magnetic field parallel to the $ab$-plane and parallel to the $c$-axis, respectively. In the range of the flux jumps occurrence, the $n$-parameter was only weakly dependent on temperature. Hence, we did not consider this dependence in our further analysis. The temperature dependence of the critical current density was estimated using the widths of the magnetization hysteresis loops in the external magnetic field of 1 T, which was close to the field of the first flux jump, and the formulas from Ref. [@Gyorgy]. We used here the loops taken by VSM in 12 tesla magnet system. In VSM system, we did not observe any flux jumps neither at 4.2 K nor at higher temperatures, and there was no limitation of the signal above 20 emu like in PPMS extraction magnetometer. Limitation of the signal as well as a large number of the flux jumps observed in PPMS experiments made determination of the critical current density (using PPMS data) impossible. The experimentally obtained temperature dependence of the critical current density was fitted according to the exponential formula: $j_\mathrm{c}(T)=j_\mathrm{c0} \exp \left( -\frac{T^{}}{T_0} \right)$, where $T^{}=\frac{T}{g(T)}$, $g(T)=1- \left( \frac{T}{T_\mathrm{c}} \right)^2 $ and $T_\mathrm{c}=~$35.5 K. We have found the following fitting parameters. For the external magnetic field parallel to the $c$-axis, $j_\mathrm{c0} = 5.6 \cdot 10^9$ A/$\mathrm{m}^2$ and $T_0$ = 7.4 K (the critical current density in the direction parallel to the $ab$-plane). For the external magnetic field parallel to the $ab$-plane, $j_\mathrm{c0} = 7.3 \cdot 10^8$ A/$\mathrm{m}^2$ and $T_0$ = 4.3 K (the critical current density in the direction perpendicular to the $ab$-plane). Our fitting procedure was performed only in the range 4 - 15 K, which was most interesting from the viewpoint of the flux jumps occurrence. The $\tau$-parameter and the Biot number ---------------------------------------- In order to determine the conditions of the critical state stability and the thermomagnetic avalanche development, it is necessary to find the relation between the thermal and magnetic diffusivity or the parameter $\tau = \frac{D_\mathrm{t}}{D_\mathrm{m}} = \mu_0 \frac{\kappa}{c_\mathrm{V}} \sigma $. In the case of the flux creep model: $\sigma = \frac{j_\mathrm{c}}{n E}$. In the case of an infinite slab model, the electric field at the surface of the slab: $E=\mu_0 a \left( \frac{H}{H_\mathrm{p}} \right) \frac{\partial H_\mathrm{ext}}{\partial t}$ for $H<H_\mathrm{p}$ and $E=\mu_0 a \frac{\partial H_\mathrm{ext}}{\partial t}$ for $H>H_\mathrm{p}$. The maximal value of the electric field which can be, at given conditions, induced in superconducting sample $E_\mathrm{max}=\mu_0 a \frac{\partial H_\mathrm{ext}}{\partial t}$. Using the maximal value of the electric field we can calculate the minimal value of $\tau$, $\tau_\mathrm{min}=\frac{\kappa j_\mathrm{c}}{{c_\mathrm{V}} n a \left( \frac{\partial H_\mathrm{ext}}{\partial t} \right)}$. In our analysis, we used the temperature dependence of the thermal parameters $\kappa$ and $c_\mathrm{V}$ from Ref. [@Morelli] and Ref. [@Chen], respectively. In all our calculations, we used the value of in-plane thermal conductivity, which is approximately one order of magnitude higher than out-of-plane thermal conductivity [@Morelli]. As a result, the ability of investigated sample to remove thermal fluctuations is limited by the in-plane thermal conductivity. We have found that in the case of all our experiments $\tau>\tau_\mathrm{min}>10^2>>1$. Hence, we can assume the dynamic conditions of the thermomagnetic avalanche development. In dynamic conditions, we expect the critical state stability to be dependent on the cooling conditions. This is consistent with the results of our experiments. In order to analyze this problem, we calculated the Biot number, $Bi=\frac{a \ h}{\kappa}$. The Biot number depends on the thermal boundary conductivity, $h$. This parameter is very difficult to evaluate experimentally [@Swarz89]. For the sample immersed in liquid helium at normal pressure and in the absence of the boiling crisis, (12 tesla magnet experiments), we expect $h$ to be of an order of $10^4$ W/$\mathrm{m}^2$K [@Mints81]. In the case of PPMS experiments, we expect this parameter to be of several orders of magnitude lower. The criteria of the critical state stability can be easily derived in two limiting cases. 1) $Bi<<1$ - weakly cooled sample (this case was considered in Ref. [@Mints96] ), 2) $Bi>>1$ - so-called isothermal conditions. We have found that the approximation of the weakly cooled sample cannot be applied in the case of our sample immersed in liquid helium (in this case $Bi\sim$ 10). However, it seems to be reasonable to use this approximation in the case of PPMS experiments, where we expect $Bi$ to be several orders of magnitude lower. The weakly cooled sample - PPMS experiment ------------------------------------------ We assumed the dynamic model for the weakly cooled sample ($Bi<<1$) to describe the critical state stability in the case of our experiments performed in PPMS. Hence, we used Eq. 3, and our experimental data, to calculate the thermal boundary conductivity, $h$. If the boundary conductivity is governed by phonon processes, we expect this dependence to be described by a power function with the power of about 3 [@Swarz89]. Hence, we assumed $h=c \cdot T^3$. For the highest sweep rate $2 \cdot 10^{-2}$ T/s, we have found, from the fit to experimental data, $c=0.25$ W/$\mathrm{m}^2$K$^4$ or $c=2.8$ W/$\mathrm{m}^2$K$^4$ for the external magnetic field parallel to the $c$-axis or to the $ab$-plane, respectively. At 4.2 K the calculated, according to our procedure, thermal boundary conductivity was 2-3 orders of magnitude lower than that for the sample immersed in liquid helium at normal pressure. Thus our assumption about a low value of the Biot number seems to be correct. However, it is necessary to notice that Eq. 3 does not describe precisely the experimentally observed dependence of $H_\mathrm{fj1}$ on the sweep rate (see Fig. 2). If $H_\mathrm{fj1}$ changes with the sweep rate according to Eq. 3, we expect the dashed lines in Fig. 2 to be parallel to the experimental curves. We observed the larges disagreement for the external magnetic field parallel to the $ab$-plane. The critical state stability and flux creep in the isothermal approximation --------------------------------------------------------------------------- In order to derive the critical state stability criterion in the isothermal conditions ($Bi>>1$), we used the following approximations. Let us assume the infinite slab to be penetrated by the magnetic flux to the depth $d$. To simplify further calculations, we assume the space coordinate $x=d$ at the sample border. According to the arguments presented in Ref. [@Mints96] in dynamic conditions for an infinite slab model, we can correlate the fluctuation of the electric field $\delta E(x)$ with the fluctuation of the temperature $\delta T(x)$ according to the formula: $$\begin{aligned} \delta E(x)=\frac{n E(x)}{j_\mathrm{c}} \left\| \frac{\partial j_\mathrm{c}}{\partial T} \right\| \delta T(x) \label{eq:ten}\end{aligned}$$ The background electric field, $E(x)$, is induced by the sweep of the external magnetic field: $E(x)=\mu_0 \frac{\partial H_\mathrm{ext}}{\partial t} x$. In the case of weakly cooled sample ($Bi<<1$), one can assume the thermal fluctuation to be independent on the space coordinate. The stability criterion (Eq. 3) can be in this case derived from the inequality $\int_0^d j \delta E dx \leq h \delta T(d) $, where the left side of the inequality is the heat generated by the fluctuation in the sample volume and the right side is the heat removed from the sample through its border [@Mints96]. In isothermal conditions ($Bi>>1$), because of strong cooling of the sample, we can assume the fluctuation of the temperature at the sample border to be equal to zero, $\delta T(d)=0$. In our approximation, we assumed the space distribution of the thermal fluctuation to be described by a square function $\delta T(x)=\delta T_0 \left( 1-\frac{1}{d^2}x^2 \right)$ where $\delta T_0=\delta T(x=0)$ is the maximal temperature fluctuation at the penetration depth. The maximal heat power, generated in the sample by the fluctuation, which can be removed from the sample volume, is in this case limited by the thermal conductivity, and the critical state stability criterion can be described by the inequality $j\delta E(x)\leq -\kappa \frac{\partial^2}{\partial x^2} \delta T(x)$. Hence, $j \delta E(x)=n \mu_0 \frac{\partial H_\mathrm{ext}}{\partial t} \left\| \frac{\partial j_\mathrm{c}}{\partial T} \right\|\delta T_0 x \left( 1-\frac{1}{d^2} x^2 \right)\leq - \kappa \frac{\partial^2}{\partial x^2} \left( \delta T_0 \left( 1-\frac{1}{d^2} x^2 \right) \right) = \frac{2 \kappa}{d^2} \delta T_0$. The left side of this inequality has a maximum for $x=\frac{d}{\sqrt{3}}$. As a result, the critical state stability criterion is given by the condition $d\leq L_\mathrm{c} = \sqrt[3]{\frac{\gamma \mu_0 \kappa}{n \left\| \frac{\partial j_\mathrm{c}}{\partial T} \right\| \frac{\partial H_\mathrm{ext}}{\partial t}}}$, or $$\begin{aligned} H_\mathrm{fj1}^{\mathrm{is}} = j_\mathrm{c} \sqrt[3]{\frac{\gamma \mu_0 \kappa}{n \left\| \frac{\partial j_\mathrm{c}}{\partial T} \right\| \frac{\partial H_\mathrm{ext}}{\partial t}}} \label{eq:eleven}\end{aligned}$$ where $\gamma = \frac{3}{13} \left( 9+ \sqrt{3} \right) \approx 2.48$. It is important to note that, also in the case of isothermal approximation, we expect the field of the first flux jump to decrease with an increase of the sweep rate, $ \frac{\partial H_\mathrm{ext}}{\partial t}$. Comparison of theoretical models with the experimental results -------------------------------------------------------------- ![Temperature dependence of the field of the first flux jump, $H_\mathrm{fj1}$ at the sweep rate of $2 \cdot 10^{-2}$ T/s and in the external magnetic field parallel to the $c$-axis (a) and parallel to the $ab$-plane (b) of the 3 x 3 x 3 $\mathrm{mm}^3$ single crystal of [La$_{1.85}$Sr$_{0.15}$CuO$_4$]{}. The predictions of the approximated models of the critical state stability - see Eqs. 1, 3 and 5 are compared with the experimental data. Dashed line shows temperature dependence of the field of full penetration, $H_\mathrm{p}(T)$. ](Fig4.eps){width="0.95\columnwidth"} In Figure 4, we present the comparison of the predictions of the three theoretical approximations discussed in the present work, i.e. the local adiabatic approximation and two dynamic approximations (of weakly cooled sample and isothermal) with the experimentally obtained temperature dependence of $H_\mathrm{fj1}$. The experimental data were taken for both sample orientations studied in our experiments: in the external magnetic field parallel to the $c$-axis (Fig. 4a) or parallel to the $ab$-plane (Fig. 4b), taken at the sweep rate of $2 \cdot 10^{-2}$ T/s. For the external magnetic field parallel to the $c$-axis, we present the results both in the case when the sample is cooled by gaseous He (the PPMS experiment) and in the case when the sample is cooled by liquid He (the experiment in 12 tesla magnet with the Hall probe - see Fig. 3a, in this case $\mu_0H_\mathrm{fj1}$(2.2 K) = 3.5 T). In the both figures, we also present temperature dependence of the field of full penetration: $H_\mathrm{p} (T)$. This parameter is very important, because in the infinite slab model of the critical state stability, we expect the flux jumps to vanish when $H_\mathrm{fj1}$ exceeds $H_\mathrm{p}$. For the model of weak cooling, we assumed $h=c \cdot T^3$, where $c=0.25$ W/$\mathrm{m}^2$K$^4$ or $c=2.8$ W/$\mathrm{m}^2$K$^4$ for the external magnetic field parallel to the $c$-axis or to the $ab$-plane, respectively. It is clearly seen that our experimental results cannot be explained by the local adiabatic approximation. The best fit of the PPMS results was found in the dynamic model of weak cooling, while the results obtained with the 12 tesla magnet are better described by the isothermal approximation. It is worth to note that the isothermal curve in Fig. 4a crosses the $H_\mathrm{p}(T)$ curve at the temperature of about 3 K. This result can explain the fact that, in the case of 12 tesla magnet experiment, we observed the flux jumps only at the temperature of 2 K and not at temperatures higher than 4.2 K. On the other hand, in the case of PPMS experiments, we observed the flux jumps also above the $H_\mathrm{p}(T)$ curve. It is worth to note that also in other experiments, one sometimes observes flux jumps in the external magnetic field higher than the field of full penetration [@Gerber00]. Such behavior of the flux jumps is not fully clarified yet. Dynamics of the flux jumps -------------------------- ![a) The magnetic field dependence of the characteristic times $t_\mathrm{bm}$ and $t_\mathrm{am}$, before and after maximum of the coil voltage. b) The magnetic field dependence of the magnetic diffusivity, estimated form the exponential fit of the coil voltage at the begin and at the end of the flux jumps, respectively. ](Fig5.eps){width="0.95\columnwidth"} Duration of the two characteristic stages of the flux jumps denoted in Fig. 3b as $t_\mathrm{bm}$ and $t_\mathrm{am}$ decreases with an increase of the external magnetic field, and $t_\mathrm{bm}$ was approximately three times shorter than $t_\mathrm{am}$ (see Fig. 5a). The exponential character of the final stage of the flux jump suggests that it can be analyzed in terms of the magnetic diffusion. The process of the magnetic diffusion depends on the initial magnetic field distribution in the sample. It is possible to decompose this distribution into a series of functions characteristic for the sample geometry (e.g. a series of cosines in the case of infinite slab geometry or a series of Bessel functions in the case of a sample with cylindrical symmetry). Each harmonic of this decomposition has specific time dependence. From the sum of the time dependent harmonics, one can obtain the time dependent distribution of the magnetic flux in the investigated sample. The voltage of the coil, wound up around the investigated sample, is proportional to the time derivative of the magnetic flux. We assumed the sample to be of a shape of infinite slab. For sufficiently long times in comparison to the parameter $$\begin{aligned} t_0 = \frac{4 l^2}{\pi^2 D_\mathrm{m}}, \label{eq:fourteen}\end{aligned}$$ we can assume the voltage of the coil - $U(t)\sim \exp(-t/t_0)$. The parameter $l$ is a characteristic length of the diffusion process. If the slab is fully penetrated by magnetic flux, $l=a$. Hence, when one fits the time dependence of the coil voltage at the final stage of the flux jump, one can find the coefficient $D_\mathrm{m}$. The magnetic field dependence of this parameter is shown in Fig. 5b. We have found an increase of this parameter with increasing external magnetic field. The initial exponential increase of the voltage also depends on the magnetic diffusivity of the superconducting sample [@Mints81]. We have fitted this dependence by an exponential function $(U(t)\sim \exp(t/t_0)$). However, it is necessary to discuss the meaning of the length parameter (the parameter $l$ in Eq. 6) at the begin and at the end of the flux jump. During the flux jumps observed in our experiment, the sample was wholly penetrated by the magnetic flux (see Fig. 3a). Hence, as a characteristic length of the diffusion process (in Eq. 6), at the end of the flux jump, we assumed the parameter $a$, it means half of the sample width. On the other hand, at the begin of the flux jump, only a part of the sample is penetrated by the magnetic flux. Hence at this stage, one should assume the characteristic length parameter to be equal to the penetration depth, which is smaller than $a$. The magnetic field dependence of the parameter $D_\mathrm{m}$ for the begin of the flux jump, we have found to be similar as that for the end of the jump. These two parameters coincide very well, if we assume in Eq. 6 $l=0.5 a$ (see Fig. 5b) for the begin of the flux jump. The characteristic diffusivity found in our experiment was of an order of 1 $\mathrm{m}^2$/s (see Fig. 5b). Such an order of diffusivity is characteristic for the magnetic diffusion in our system, providing the sample is in the normal state or in the flux flow regime. The resistivity of [La$_{1.85}$Sr$_{0.15}$CuO$_4$]{} is anisotropic. However, we studied dynamics of the flux jumps only for the external magnetic field parallel to the $c$-axis of the investigated crystal. In such sample orientation, screening currents flow only in the $ab$-plane. If we assume the in-plane normal state resistivity $\rho_\mathrm{n}> 0.25 \ \mathrm{m} \Omega \mathrm{cm}$ [@Fuji], then $D_\mathrm{m} = \rho_\mathrm{n} / \mu_0 > 2 \ \mathrm{m}^2/\mathrm{s} $. On the other hand, if one takes data of the thermal conductivity ($\kappa$) [@Morelli] and of the specific heat ($c_\mathrm{V}$) [@Chen], one can estimate the in-plane thermal diffusivity ($D_\mathrm{th}=\kappa/c_\mathrm{V}$) to be of an order of $10^{-3} - 10^{-2} \ \mathrm{m}^2/\mathrm{s}$. The out-of-plane thermal diffusivity is approximately one order of magnitude lower than the in-plane thermal diffusivity [@Morelli]. Discussion ========== The critical state stability in a single crystalline [La$_{1.85}$Sr$_{0.15}$CuO$_4$]{} is influenced by a large number of parameters. All of them must be considered in order to determine the critical state stability in the framework of existing theoretical models. In [La$_{1.85}$Sr$_{0.15}$CuO$_4$]{} crystal, most of these parameters are anisotropic. Hence, the stability of the critical state depends strongly on the sample orientation with respect to the external magnetic field. In our analysis, we consider both anisotropy of critical current density and the anisotropy of the current-voltage characteristic (the anisotropy of the $n$-parameter). In the present analysis, we assumed the thermal conductivity of the investigated sample to be limited only by the in-plane thermal conductivity, which is approximately one order of magnitude higher that the out-of-plane thermal conductivity. According to our results in the case of [La$_{1.85}$Sr$_{0.15}$CuO$_4$]{} crystal, the anisotropy of the $n$-parameter is relatively weak. Hence, the anisotropy of the critical state stability is connected mainly with the anisotropy of screening currents. The difference between the in-plane and out-of-plane screening current, at given temperature, is about one order of magnitude. In both sample orientations studied in our experiments, the thermomagnetic avalanche develops in dynamic conditions. Our sample is characterized by a strong flux creep phenomenon. Hence, the critical state stability is influenced both by the external magnetic field sweep rate and the cooling conditions. These cooling conditions were very different in two cryostats used in our experiments. In PPMS experiments, where the sample is surrounded by helium gas of low pressure, the critical state stability can be analyzed in the approximation of weak cooling or low Biot number. In such condition, the thermal parameter that influences the critical state stability is the thermal boundary conductivity. This parameter is difficult to determine experimentally. It depends e.g. on the roughness of the surface of the investigated sample. In the case of PPMS experiments, the investigated sample is put additionally into a holder which makes an analysis of the heat exchange conditions even more difficult. Nevertheless, it seems to be slightly surprising that, according to our analysis, the thermal boundary conductivity changes by an order of magnitude after the change of the sample orientation. It is possible that such results are, to some extent, connected with the approximations assumed by the models used in our analysis. In our analysis we used models of an isotropic infinite slab sample. In order to determine (in the framework of these models) the critical state stability criterion, it is necessary to know a large number of parameters. Each of these parameters must be determined experimentally and all parameters have some experimental errors. As the result, the uncertainty of the obtained critical state stability criterion can be relatively large. The models used in our analysis can be applied only for some limiting cases ($Bi<<1$ or $Bi>>1$). In order to determine the critical state stability in an intermediate case, numerical calculations are necessary. Such calculations were performed e.g. in Ref. [@Zhou]. In order to determine the current-voltage characteristics, we study the relaxation of the magnetic moment. Although Eq. 2 can be applied to describe the the current-voltage characteristics in a wide range of the electric field, one must bear in mind that the electric field induced in the superconducting sample during magnetic relaxation experiments is significantly lower than the electric field induced by the external magnetic field sweep. One should also bear in mind that the accuracy of determination of $H_\mathrm{fj1}$ for the highest sweep rate ($2 \cdot 10^{-2}$ T) in our experiment was only about $\pm 0.1$ T. This experimental error has a special significance for the sample orientation with the external magnetic field parallel to the $ab$-plane, because for this orientation all observed values of $H_\mathrm{fj1}$ were in the range between 0.5 T and 1 T (see Fig. 2). Finally, more accurate analysis of the critical state stability in an anisotropic material should also take into account the value of the out-of-plane thermal conductivity. Our experimental results, as well as the data for textured YBaCuO [@Guillot], show that by the improvement of the cooling conditions, i.e. by immersing the sample in liquid helium, the critical state stability increases significantly. Both dynamic models used in our analysis predict the dependence of $H_\mathrm{fj1}$ on the external magnetic sweep rate (see Eq. 3 and Eq. 5). These equations were derived assuming the critical current density to be independent on the magnetic field and predict this dependence to be described by a power function with the power of $-\frac{1}{2}$ and $-\frac{1}{3}$ for the approximation of weak cooling and the isothermal approximation, respectively. However, if the critical current density is strongly dependent on the magnetic field, the dependence of $H_\mathrm{fj1}$ on the sweep rate can be modified [@Mints96]. The comparison of the predictions of Eq. 3 with the experimental results, one can see in Fig. 2 and in Fig. 4. One can observe the largest inconsistency for the lowest temperatures and for the external magnetic field parallel to the $ab$-plane. The influence of the heat exchange conditions on critical state stability as well as a strong dependence of $H_\mathrm{fj1}$ on the sweep rate show that thermomagnetic instabilities are initiated in dynamic conditions, in which the magnetic diffusivity is smaller than the thermal one. On the other hand, the effective diffusivity describing the dynamics of the flux jumps (see Fig. 5b) is very high, it means comparable to the magnetic diffusivity in the normal state. The magnetic field dependence of this effective diffusivity suggests that it can be correlated with the flux flow resistivity. In order to understand this discrepancy, one should explain that the initial stage of the avalanche develops at the voltage level that is induced by the external magnetic field sweep. We estimate that the sweep rates used in our experiments induced in the coil the voltage of an order of $10^{-8}-10^{-6}$ V, which is six or four orders of the magnitude lower than the noise level registered by the acquisition board. For this reason, we were not able to register this very initial stage of the avalanche. Our experimental results show that after this initial stage of the avalanche the resistivity of the sample rapidly increases and the dynamics of the flux jumps is governed by the magnetic diffusivity, which depends of the flux flow resistivity. Conclusions =========== The stability of the critical state in a large crystal of [La$_{1.85}$Sr$_{0.15}$CuO$_4$]{} is influenced by its anisotropy, and it depends on the sample orientation with respect to the external magnetic field. The most relevant is the anisotropy of the critical current. At relatively slow magnetic field sweep rates, thermomagnetic avalanches are initiated in the dynamic conditions. Hence, one observes a strong influence of the heat exchange conditions as well as of the sweep rate on the critical state stability. In the case of the experiments, where the investigated sample is surrounded by helium gas with low pressure (PPMS), the critical state stability criterion can be derived in the approximation of a weakly cooled sample. However, this approximation cannot be applied in the case of large crystals immersed in liquid helium. Flux jumps in the single crystalline [La$_{1.85}$Sr$_{0.15}$CuO$_4$]{} are very short, typical duration is of several microseconds. The dynamics of the jumps is governed by the magnetic diffusivity, proportional to the flux flow resistivity. This work was partly supported by Polish National Science Centre under research project for years 2011-2012 (grant N N202 1663 40). References {#references .unnumbered} ========== [99]{} Mints R G and Rakhmanov A L 1981 [Rev. Mod. Phys.]{} [53]{} 551 Swarz E T and Pohl R O 1989 [Rev. Mod. Phys.]{} [61]{} 605 McHenry M E, Lessure H S, Maley M P, Coulter J Y, Tanaka I and Kojima H 1992 [Physica C]{} [190]{} 403 Gerber A, Tarnawski Z and Franse J J 1993 [Physica C]{} [209]{} 147 Mints R G 1996 [Phys. Rev. B]{} [53]{} 12311 Nabialek A, Niewczas M, Dabkowska H, Dabkowski A, Castellan J P and Gaulin B D 2003 [Phys. Rev. B]{} [67]{} 024518 Nabialek A and Niewczas M 2006 [Physica C]{} [436]{} 43 Guillot M, Potel M, Gougeon P, Noel H, Levet J C, Chouteau G and Tholence J L 1988 [Phys. Lett. A]{} [127]{} 363 Watanabe K, Kobayashi N, Awaji S, Kido G, Nimori S, Kimura K, Sawano K and Muto Y 1991 [Japan Journ. of Appl. Phys.]{} [30]{}, L 1638 Chikaba J 1970 Cryogenics [10]{} 306 Anderson P W 1962 [Phys. Rev. Lett.]{} [9]{} 309 Gyorgy E M, van Dover R B, Jackson K A, Schneemeyer L F and Waszczak J V 1989 [Appl. Phys. Lett.]{} [55]{} 283 Morelli D T, Doll G L and Heremans J 1991 [Solid State Commun.]{} [77]{} 773 Chen S J, Chang C F, Tsay H L, Yang H D and Lin J Y 1998 [Phys. Rev. B]{} [58]{} R14 753 Gerber A and Milner A 2000 [Phys. Rev. B]{} [62]{} 9753 Fujishiro H, Ikebe M and Shibazaki Y 1999 [J. Low Temp. Phys.]{} [117]{} 1181 Zhou Y H and Yang X 2006 [Phys Rev. B]{} [74]{} 054507	{ "pile_set_name": "ArXiv" }
--- author: - \| \ NIC, DESY, Platanenallee 6, 15738 Zeuthen, Germany\ E-mail: - '(for the ALPHA collaboration)' title: 'Fixing the parameters of Lattice HQET including $1/m_B$ terms' --- Heavy Quark Effective Theory [@eichten_hill] in its basic formulation provides an effective description of QCD with $N_f-1$ light quarks and a single heavy quark whose mass is much larger than the QCD energy scale $\Lambda_{{\textrm{QCD}\,}}$. The heavy quark is treated non-relativistically and processes are described in its reference frame. In order to avoid the ambiguities of the perturbative expansion of HQET in the strong coupling $g$ [@grozin] one may employ non-perturbative techniques such as lattice QCD and consequently lattice HQET [@leshouches]. As any effective theory HQET containes several low energy constants which need to be determined in order to match it to QCD and grant it a predictive power. This step called ’matching’ should also be performed non-perturbatively [@sachrajda]. The ALPHA collaboration has set up a non-perturbative matching strategy to determine the needed HQET parameters at order $1/m_b$ [@heitger_sommer; @1; @2]. It relies on a set of carefully chosen observables which are precisely computable in lattice QCD as well as in lattice HQET. We describe the results of a one-loop computation which tests the quality of some of these observables. In order to estimate the $1/m_b^2$ contributions we define a quantity $R$ which measures the ratio of the one-loop corrections to their tree-level value of $1/m_b$ terms. The paper is organized as follows: in section \[sec1\] we introduce the lattice HQET Lagrangian and the currents as well as higher dimensional operators needed to account for $1/m_b$ corrections, then in section \[sec2\] we briefly describe the framework in which the matching observables are constructed and finally in section \[sec3\] we discuss the one-loop results and conclude. HQET at next-to-leading order in $m_b$ {#sec1} ====================================== The formulation of lattice HQET was thoroughly discussed in [@leshouches] and therefore we only quote the relevant formulae. The Lagrangian is a sum of the static part and two $1/m_b$ corrections $$\mathcal{L}_{{\textrm{HQET}\,}} = \mathcal{L}_{{\textrm{stat}\,}} - \Big( {\color{black} \omega_{{\textrm{kin}\,}}} \mathcal{L}_{{\textrm{kin}\,}} + {\color{black} \omega_{{\textrm{spin}\,}}} \mathcal{L}_{{\textrm{spin}\,}} \Big) + \mathcal{O}(1/m_b^2),$$ It is part of the definition of HQET that the kinetic and chromomagnetic operators enter only as insertions in the static vacuum expectation values, namely for some operator $\mathcal{O}$ we have $$\langle \mathcal{O} \rangle_{{\textrm{HQET}\,}} = \langle \mathcal{O} \rangle_{\textrm{stat}} + {\color{black}\omega_{\textrm{kin}}} \sum_{x} \langle \mathcal{O} \mathcal{L}_{\textrm{kin}}(x) \rangle_{\textrm{stat}} + {\color{black}\omega_{\textrm{spin}}} \sum_{x} \langle \mathcal{O} \mathcal{L}_{\textrm{spin}}(x) \rangle_{\textrm{stat}}$$ The HQET operators themselves are also expanded in $1/m_b$. For the axial current we have $$\begin{aligned} \big(A_0\big)_R &= {\color{black}Z_{A_0}^{{\textrm{HQET}\,}}} \big\{ \bar{\psi}_l \gamma_0 \gamma_5 \psi_h + {\color{black} c_{A_{0,1}}} \bar{\psi}_l \frac{1}{2} \gamma_5 \gamma_i ( \nabla_i^S - \overleftarrow{\nabla}_i^S) \psi_h + {\color{black} c_{A_{0,2}}} \bar{\psi}_l \frac{1}{2} \gamma_5 \gamma_i ( \nabla_i^S + \overleftarrow{\nabla}_i^S) \psi_h \big\}, \nonumber \\ \big(A_k\big)_R &= {\color{black}Z_{A_k}^{{\textrm{HQET}\,}}} \big\{ \bar{\psi}_l \gamma_k \gamma_5 \psi_h + {\color{black} c_{A_{k,1}}} \bar{\psi}_l \frac{1}{2} (\nabla_i^S - \overleftarrow{\nabla}_i^S) \gamma_i \gamma_5 \gamma_k \psi_h + {\color{black} c_{A_{k,2}}} \bar{\psi}_l \frac{1}{2} (\nabla_k^S - \overleftarrow{\nabla}_k^S) \gamma_5 \psi_h \nonumber \\ &\ \qquad \qquad + {\color{black} c_{A_{k,3}}} \bar{\psi}_l \frac{1}{2} (\nabla_i^S + \overleftarrow{\nabla}_i^S) \gamma_i \gamma_5 \gamma_k \psi_h + {\color{black} c_{A_{k,4}}} \bar{\psi}_l \frac{1}{2} (\nabla_k^S + \overleftarrow{\nabla}_k^S) \gamma_5 \psi_h \big\}, \nonumber\end{aligned}$$ and similarly for the vector current (we use the notation from Ref.[@leshouches]). $\psi_l$ denotes a relativistic, massless fermion, whereas $\psi_h$ is a nonrelativistic heavy fermion. In order to define HQET and the currents at the next-to-leading order one has to fix 3 parameters in $\mathcal{L}_{{\textrm{HQET}\,}}$ and 2 $\times$ 3 parameters in $A_0(x)$ and $V_0(x)$ and 2 $\times$ 5 in $A_k(x)$ and $V_k(x)$ giving in total 19 parameters. They are usualy denoted collectively by $\omega_i$, with $i=1, \dots, 19$. In this work we concentrate on the parameters $\omega_{{\textrm{kin}\,}}$ and $c_{A_{0,1}}$ and the corresponding matching conditions. How to determine the HQET parameters? {#sec2} ===================================== The HQET parameters are determined by considering observables $\phi_i$ which can be reliably calculated in lattice QCD and in lattice HQET. The matching condition reads $$\phi_{i,{\textrm{QCD}\,}}(L,z,a = 0) \stackrel{!}{=} \phi_{i,{\textrm{HQET}\,}}(L,z,a, \{ {\color{black}\omega(z,a)} \}) = \phi_{i,{\textrm{stat}\,}}(L,a) + \phi_{ij,1/m}(L,a) \ {\color{black}\omega_j(z,a)}, \label{matching}$$ where $L$ is the size of the finite volume in which the observables $\phi_i$ are defined, $a$ is the lattice spacing and $z$ is a dimensionless parameter used to fix the heavy quark mass $m$ given by $z = \bar{m}(L) L$, where $\bar{m}(L)$ is the mass defined in the lattice minimal subtraction scheme [@leshouches2]. In Ref.[@3] it was proposed to use observables defined in the Schrödinger functional framework which differs from the usual one by the boundary conditions that are imposed on the fields at time 0 and $T$ (for a more detailed discussion of the Schrödinger functional framework see [@leshouches2]). Apart from the usual fields in the bulk one has boundary fields which can be used to construct correlation functions. The observables analyzed in this work are constructed from boundary-to-boundary or boundary-to-bulk correlation functions, e.g. $$\begin{aligned} F_1(\theta) &= -\frac{a^{12}}{2L^6} \sum_{\textbf{u}, \textbf{v}, \textbf{y}, \textbf{z}}\langle \bar{\zeta}'_l(\textbf{u}) \gamma_5 \zeta'_h(\textbf{v}) \bar{\zeta}_h(\textbf{y}) \gamma_5 \zeta_l(\textbf{z}) \rangle, \\ K_1(\theta) &= -\frac{a^{12}}{6L^6} \sum_i \sum_{\textbf{u}, \textbf{v}, \textbf{y}, \textbf{z}}\langle \bar{\zeta}'_l(\textbf{u}) \gamma_i \zeta'_h(\textbf{v}) \bar{\zeta}_h(\textbf{y}) \gamma_i \zeta_l(\textbf{z}) \rangle, \\ f_{A_{0}}(\theta, x_0) &= - \frac{a^6}{2} \sum_{\textbf{u}, \textbf{v}} \langle \bar{\zeta}_h(\textbf{u}) \gamma_5 \zeta_l(\textbf{v}) \big(A_{0}\big)_I(x_0) \rangle \end{aligned}$$ where $\zeta$ and $\bar{\zeta}$ denote fermionic fields living on the boundary. The $\theta$ angles are additional kinematic parameters corresponding to the momenta of quark fields in the bulk. The observables are defined in such a way as to cancel all renormalization factors and the angles can be tuned such as to minimize cut-off effects [@3] $$\phi_2(\theta_1, \theta_2) = \frac{1}{4} \log \frac{F_1(\theta_1)}{F_1(\theta_2)}+\frac{3}{4} \log \frac{K_1(\theta_1)}{K_1(\theta_2)}, \quad \phi_4(\theta_1, \theta_2) = \log \frac{f_{A_0}(\theta_1,x_0=T/2)}{f_{A_0}(\theta_2,x_0=T/2)}. \label{obs} $$ In Ref.[@3] the proposed set of matching conditions was solved at tree-level yielding the classical HQET parameters and it was checked that the $1/m_b^2$ corrections are small. The purpose of the present study is to confirm these conclusions by a one-loop computation similar to the one performed in [@dirk]. The one-loop contributions to the observables Eq. were calculated with `pastor`, an automatic tool for generation and calculation of lattice Feynman diagrams [@pastor]. It is a flexible package, which takes as input the discretized action, the definition of the correlation function, and parameters such as $L/a$ and the dimensionless heavy quark mass $z$. Then, `pastor` automatically generates the Feynman rules corresponding to the specified action, all Feynman diagrams corresponding to the requested correlation function and a numerical contribution of each diagram. The calculations were performed for the Wilson plaquette gauge action and $\mathcal{O}(a)$-improved Wilson fermions with two light quarks and one massive. One-loop contributions were evaluated for $\phi_{i,{\textrm{QCD}\,}}$ and $\phi_{i,{\textrm{stat}\,}}$. How to estimate the quality of the observables? {#sec3} =============================================== To analyze the $1/z^{2}$ corrections to an observable $\phi$ at one-loop order in the coupling constant we expand the matching condition Eq. in $g^2$ and get (dropping terms of order $g^4$ and $1/z^{2}$) $${\color{black}\phi_{{\textrm{QCD}\,}}^{(0)}(z)} + g^2 \phi_{{\textrm{QCD}\,}}^{(1)}(z) = {\color{black}\phi_{{\textrm{stat}\,}}^{(0)}} + g^2 \phi_{{\textrm{stat}\,}}^{(1)} + z^{-1} \sum_t \Big( {\color{black}\hat{\omega}_{t}^{(0)} \hat{\phi}_{t}^{(0)}} + g^2 \hat{\omega}_{t}^{(1)}(z) \hat{\phi}_{t}^{(0)} + g^2 \hat{\omega}_{t}^{(0)} \hat{\phi}_{t}^{(1)} \Big) \label{exp}$$ where the sum over $t$ refers to different subleading contributions. The parameters $\hat{\omega}_t$ differ from the HQET parameters in Eq. by an explicit factor $1/\bar{m}(L)$ which was factored out, whereas $\hat{\phi}_t$ have an explicit factor $L$. In this notation the kinetic contribution is $\hat{\omega}_{{\textrm{kin}\,}}^{(0)} = \frac{1}{2}$, the spin contribution vanishes at tree level ($\phi^{(0)}_{{\textrm{spin}\,}}=0$), and the remaining contributions correspond to corrections to the current operators proportional to the coefficients $c_X$. To quantify the $1/z^2$ corrections we define a ration $R$ by extracting the one-loop contribution from Eq. by dividing by $( {\color{black}\phi_{{\textrm{QCD}\,}}^{(0)}(z) - \phi_{{\textrm{stat}\,}}^{(0)}} ) g^2$ $$\begin{aligned} R(\theta_1, \theta_2) = \frac{\phi_{{\textrm{QCD}\,}}^{(1)}(z) - \phi_{{\textrm{stat}\,}}^{(1)}}{{\color{black}\phi_{{\textrm{QCD}\,}}^{(0)}(z) - \phi_{{\textrm{stat}\,}}^{(0)}}} &= \frac{ \sum_t {\color{black}\hat{\omega}_{t}^{(0)}} \hat{\phi}_{t}^{(1)}(\theta_1, \theta_2)}{ \sum_t {\color{black}\hat{\omega}_{t}^{(0)}} {\color{black} \hat{\phi}_{t}^{(0)}(\theta_1, \theta_2)}} + \frac{ \sum_t \hat{\omega}_{t}^{(1)}(z) {\color{black}\hat{\phi}_{t}^{(0)}(\theta_1, \theta_2)} }{\sum_t {\color{black}\hat{\omega}_{t}^{(0)} {\color{black}\hat{\phi}_{t}^{(0)}(\theta_1, \theta_2)}}} \\ &= {\color{black}\alpha(\theta_1, \theta_2) + \beta(\theta_1, \theta_2) +\gamma(\theta_1, \theta_2)\log(z)} \label{def_r}\end{aligned}$$ Since the left hand side of Eq. has a well defined continuum limit, the right hand side can be also considered in the continuum (the right hand side of Eq. must be considered as a entity since the particular terms in the sum may be divergent as $a \rightarrow 0$). Note that the explicit $1/z$-dependence cancels. The only $z$ dependence remains in $\hat{\omega}_t^{(1)}(z)$ and can be parametrized as $\hat{\omega}_t^{(1)}(z) = \beta_t + \gamma_t \log(z)$. Hence, when $R$ is plotted on a linear-log plot, the ratio $R$ measures simultaneously: - $1/z^{2}$ corrections: deviations from a linear behaviour signal $1/z^{2}$ contributions, - slope: the coefficient of the subleading logarithm. The slope gives information about a specific linear combination of the anomalous dimensions of the $1/m_b$ operators, which at one-loop order are universal (independent of the scheme) and in some cases can be predicted analytically. We will now present the ratios $R$ for two representative matching conditions. The simplest matching condition is the one for $\omega_{{\textrm{kin}\,}}$ [@3]. The $z$ dependence of the one-loop contribution to the observable $\phi_2$ is shown on figure \[r1z\]. The ratio $R_2$ is particularly simple since there is only one subleading contribution, namely $\hat{\omega}_{{\textrm{kin}\,}}$, $$\begin{aligned} R_{2}(\theta_1, \theta_2) \equiv \frac{\phi_{2,{\textrm{QCD}\,}}^{(1)}(z) - \phi_{2,{\textrm{stat}\,}}^{(1)}}{{\color{black}\phi_{2,{\textrm{QCD}\,}}^{(0)}(z) - \phi_{2,{\textrm{stat}\,}}^{(0)}}} &= \frac{\hat{\phi}_{2,{\textrm{kin}\,}}^{(1)}(\theta_1, \theta_2)}{{\color{black}\hat{\phi}_{2,{\textrm{kin}\,}}^{(0)}(\theta_1, \theta_2)}} + \frac{\hat{\omega}_{{\textrm{kin}\,}}^{(1)}(z)}{{\color{black}\hat{\omega}_{{\textrm{kin}\,}}^{(0)}}} = {\color{black}\alpha(\theta_1, \theta_2) + \beta + \gamma \log(z)} \label{rf1} $$ Using the fact that in the continuum $\bar{m}(L) \omega_{{\textrm{kin}\,}}^{(0)} = \frac{1}{2}$ we have $\hat{\omega}_{{\textrm{kin}\,}}^{(1)}(z)/{\color{black}\hat{\omega}_{{\textrm{kin}\,}}^{(0)}} = 2 \bar{m}(L)\omega_{{\textrm{kin}\,}}^{(1)}(z)$. From continuum HQET we know (see for example [@grozin]) that reparametrization invariance fixes the renormalization factor for the kinetic operator to its classical value to all orders of perturbation theory. This is true if the quark mass used to define $\omega_{{\textrm{kin}\,}}$ is given as the pole mass. In our computation we use the $\bar{m}(L)$ mass, therefore a conversion factor needs to be included. The one-loop conversion between the pole mass and the $\overline{\textrm{MS}}$ scheme can be taken for example from [@msbar_mass] where its 3-loop version was derived, whereas the relation between the $\overline{\textrm{MS}}$ mass and the $\textrm{MS}_{\textrm{lat}}$ was given in [@sommer_kurth]. Hence, we obtain the one-loop correction to $\omega_{{\textrm{kin}\,}}$ as $$\bar{m}(L) \omega_{{\textrm{kin}\,}}^{(1)}(z) = -\frac{1}{6\pi^2} -\frac{1}{2}0.122282 \ C_F +\frac{1}{4\pi^2} \log z,$$ from which the parameters in Eq. can be obtained, namely, $\beta = -\frac{1}{3\pi^2} - 0.122282 \ C_F$ and $\gamma = \frac{1}{2\pi^2}$. The data shown on figure \[r1\] exhibits a slope compatible with the predicted one. We can also conclude that $1/z^{2}$ corrections are equally small for all sets of $\theta$ angles, hence the best setting can be choosen by Monte Carlo precision and tree-level considerations. Figure \[r8\] shows the $R$ ratio for the observable $\phi_4$, where several terms contribute to the sums on the right hand side of Eq. and the $\theta$-dependence of the coefficient of the logarithm doesn’t cancel any more. Hence, a seperate fit was performed for each set of $\theta$ angles. Again, one concludes that all data lie on straight lines and therefore the $1/z^{2}$ correction are indeed small. Conclusions =========== Lattice HQET is a prototype of an effective theory where one can perform a non-perturbative matching. We have studied quantitatively the contamination of the matching conditions by $1/m_b^2$ contributions and confirmed the tree-level conclusion that such corrections are negligible. The ratios $R$ introduced in Eq.\[def\_r\] proved to be usefull as they provide a handle for the various parts of the $1/m_b$ contributions. Complete results for the remaining matching conditions will be presented elsewhere [@4]. The author would like to thank especially D. Hesse for the help with `pastor` and R. Sommer and H. Simma for many useful discussions. [99]{} E. Eichten, B. Hill, Phys. Lett. B 234 (1990) 511, Phys. Lett. B 243 (1990) 425, A. Grozin, Heavy quark effective theory, Springer (2004), R. Sommer, in Modern perspectives in lattice QCD, Springer (2010), C. Sachrajda, Nucl. Phys. Proc. Suppl. 185 (2008) 62, J. Heitger, R. Sommer, JHEP 02 (2004) 022, B. Blossier, M. Della Morte, N. Garron, R. Sommer, JHEP 06 (2010) 002, B. Blossier, M. Della Morte, P. Fritzsch, N. Garron, J. Heitger, H. Simma, R. Sommer, N. Tantalo, JHEP 09 (2012) 132, P. Weisz, in Modern perspectives in lattice QCD, Springer (2010), M. Della Morte, S. Dooling, J. Heitger, D. Hesse, H. Simma, in preparation, D. Hesse, R. Sommer, JHEP 1302 (2013) 115, D. Hesse, `pastor`, in preparation, N. Gray, D. Briadhurst, W. Grafe, K. Schilcher, Z. Phys. C 48 (1990) 673, M. Kurth, R. Sommer, Nucl. Phys. B 623 (2002) 271, P. Korcyl, in preparation.	{ "pile_set_name": "ArXiv" }
--- abstract: 'We deal with a dynamical mechanism in which a large cosmological constant, as suggested by inflationary scenarios, decays due to expansion of the universe. This mechanism has its origin in the gravitational coupling of the vacuum density. We assume that the vacuum couples anomalously to gravity that is the metric tensor that appears the gravitational part is not the same as that appears the matter part as suggested by weak equivalence principle. Instead, the two metric tensors are taken to be conformally related. We show that this provides a dynamical mechanism which works during expansion of the universe. We also consider some observational consequences of such a gravitational model.' --- .5cm -26pt -.85in \ ${\bf Yousef~Bisabr}$[^1]\ \ [Keywords:]{} Gravity, Modified Gravity, The Cosmological Constant.\ \ Introduction ============ There is a large discrepancy between observations and theoretical estimates on the vacuum energy density. This problem is known as the first cosmological constant problem (or the fine tuning problem) [@1] [@a1]. The second one known as the coincidence problem [@2] deals with the question that why the vacuum energy is the same order of magnitude of the matter density.\ There have been many attempts trying to resolve the fine tuning problem [@a1]. Most of them are based on the belief that the cosmological constant may not have such an extremely small value at all times. In fact, it has a large value at early times as suggested by inflationary models and there should exist a dynamical mechanism working during evolution of the universe which provides a cancelation of the vacuum energy density at late times [@bis].\ In this work, we will consider the possibility that this cancelation mechanism is related to gravitational coupling of matter in a gravitational system. The matter part of a gravitational system is usually taken to be coupled with the metric which describes the gravitational part. This is a normal coupling which has its origin in the weak equivalence principle. Firstly, it should be pointed out that even though, observational evidences are fully consistent with weak equivalence principle, the experiments that support the principle are recent and it is quite possible that the principle is violated during evolution of the universe. Secondly, the weak equivalence principle is a statement about the gravitational coupling of normal matter and there is no experimental constrains on gravitational coupling of the other forms of energy densities such as vacuum.\ We will assume that the vacuum energy density does not couple normally to gravity but rather they couple abnormally or anomalously, that is the metrics that appear in the matter and the gravitational parts are different but conformally related [@bis]. As illustrations, we will consider such an anomalous gravitational coupling in $f(R)$ gravity [@f1] [@f2] and Brans-Dicke (BD) theory [@bd] and discuss the conditions that lead to alleviation of the cosmological constant problem. Field Equations =============== We consider the action functional $$S= \int d^{4}x \sqrt{-g}~\{\frac{M_p^2}{2} R -\frac{1}{2}g^{\mu\nu} \nabla_{\mu} \varphi~ \nabla_{\nu} \varphi -V(\varphi)\}+ S_{m}(A^2(\varphi)g_{\mu\nu} , \psi) \label{b4}$$ where $M_p^{-2}\equiv 8\pi G$, $G$ is the gravitational constant and $S_{m}$ is the action of some matter field $\psi$. The function $A(\varphi)$ is a coupling function that characterizes coupling of the scalar field $\varphi$ with the matter sector. It defines anomalous gravitational coupling of matter characterized by the field $\psi$. In general, such a matter system does not respect the weak equivalence principle. However, the scalar field $\varphi$ can hide the anomalous coupling via chameleon mechanism [@cham] and pass local gravity experiments [@bc] [@ch].\ Variation of this action with respect to the metric tensor $g^{\mu\nu}$, gives $$G_{\mu\nu}=M_p^{-2}(T^{\varphi}_{\mu\nu}+ T^{m}_{\mu\nu}) \label{b6} \label{b7}$$ where $$T^{\varphi}_{\mu\nu}=\nabla_{\mu}\varphi \nabla_{\nu}\varphi-\frac{1}{2}g_{\mu\nu}\nabla^{\gamma}\varphi \nabla_{\gamma}\varphi-V(\varphi)g_{\mu\nu} \label{b8}$$ $$T^m_{\mu\nu}=\frac{-2}{\sqrt{-g}}\frac{\delta S_{m}}{\delta g^{\mu\nu}} \label{b9}$$ are stress-tensors of the scalar field and the matter system. If we vary the action with respect to the scalar field $\varphi$, we obtain $$\Box\varphi-\frac{dV(\varphi)}{d\varphi}=-\frac{\beta(\varphi)}{M_p} T^{m} \label{b11}$$ or, equivalently, $$\nabla^{\mu}T^{\varphi}_{\mu\nu}=-\frac{\beta(\varphi)}{M_p} \nabla_{\nu}\varphi T^m \label{b8-10}$$ where $$\beta(\varphi)=M_p \frac{d\ln A(\varphi)}{d\varphi} \label{b8-1}$$ and $T^m\equiv g^{\mu\nu}T^m_{\mu\nu}$. The two stress tensors $T^m_{\mu\nu}$ and $T^{\varphi}_{\mu\nu}$ are not separately conserved due to coupling of the scalar field $\varphi$ with matter. This can be seen by applying the Bianchi identities $\nabla^{\mu} G_{\mu\nu}=0$ to (\[b6\]) which results in $$\nabla^{\mu}T^{m}_{\mu\nu}=-\nabla^{\mu}T^{\varphi}_{\mu\nu}= \frac{\beta(\varphi)}{M_p}\nabla_{\nu}\varphi~T^{m}\label{b13}$$ The parameter $\beta$ is generally a function of time. We will, however, restrict ourselves to the case that it can be regarded as a constant parameter. There are two important cases in which the function $\beta(\varphi)$ takes a constant configuration; $f(R)$ gravity and Brans-Dicke theory. $f(R)$ Gravity -------------- The action for an $f(R)$ gravity theory in Jordan frame is given by [@f2] $$S_{JF}= \frac{1}{2}M_p^2\int d^{4}x \sqrt{-\bar{g}} f(\bar{R}) +S_{m}(\bar{g}_{\mu\nu}, \psi)\label{b1}$$ where $\bar{g}_{\mu\nu}$ is the metric in Jordan frame. A conformal transformation $$g_{\mu\nu} =A^{-2}(\varphi)~ \bar{g}_{\mu\nu} \label{b2}$$ with $A^{-2}(\varphi)\equiv\frac{df}{dR}=f^{'}(R)$ together with $$\varphi = \frac{M_p}{\beta } \ln A(\varphi) \label{b3}$$ and $\beta=-\sqrt{\frac{1}{6}}$, transforms (\[b1\]) into the action (\[b4\]) with a potential [@soko] [@w] $$V(\varphi(R))=\frac{M_p^2}{2}(\frac{R}{f'(R)}-\frac{f(R)}{f'^2(R)}) \label{b5}$$ Scalar-Tensor Gravity --------------------- The general action of a scalar-tensor gravity is given by [@pol] $$S_{JF}=\frac{1}{16\pi G} \int d^{4}x \sqrt{-\bar{g}}~\{F(\phi)\bar{R}-Z(\phi)\bar{g}^{\mu\nu} \bar{\nabla}_{\mu} \phi~ \bar{\nabla}_{\nu} \phi -2U(\phi)\}+ S_{m}( \bar{g}_{\mu\nu}, \psi_m) \label{d1}$$ where $F(\phi)$, $Z(\phi)$ and $U(\phi)$ are some functions[^2]. This action is reduced to the action (\[b4\]) by the conformal transformation (\[b2\]) with $A(\varphi)=F^{-1/2}(\phi)$ and $$(\frac{d\varphi}{d\phi})^2=2M_p^2[\frac{3}{4}(\frac{d\ln F(\phi)}{d\phi})^2+\frac{Z(\phi)}{2F(\phi)}] \label{d2}$$ $$V(\varphi)=M_p^2 U(\phi)F^{-2}(\phi)\label{d3}$$ It is cleat from (\[d2\]) that the coupling function $A(\varphi)$ depends on the functions $F(\phi)$, $Z(\phi)$ and $U(\phi)$ through the relation (\[d2\]). For some particular choices of these functions, $\beta$ takes a constant configuration and then, as a result of (\[b8-1\]), $A(\varphi)$ takes an exponential form. This defines a class of scalar-tensor theories and we restrict ourselves to this class. One important theory in this class is given by BD parameterization in which $F(\phi)=16\pi G\phi$, $Z(\phi)=16\pi G\omega_{BD}/\phi$ and $U(\phi)=8\pi G W(\phi)$, and then $$S_{JF}= \int d^{4}x \sqrt{-\bar{g}} (\phi \bar{R} -\frac{\omega_{BD}}{\phi}\bar{g}^{\mu\nu}\bar{\nabla}_{\mu}\phi \bar{\nabla}_{\nu}\phi-W(\phi))+S_{m}(\bar{g}_{\mu\nu}, \psi)\label{d5}$$ with $\omega_{BD}$ and $W(\phi)$ being BD parameter and the potential in Jordan frame, respectively. This action is reduced to (\[b4\]) by [@bis] [@sc] $$A(\varphi)=e^{\beta_{BD}\varphi/M_p}\label{d6}$$ $$\varphi(\phi)/M_p=\sqrt{\omega_{BD}+3/2}\ln (\frac{\phi}{\phi_0}) \label{a3}$$ $$V(\varphi)=W(\phi(\varphi))~e^{8\beta_{BD}\varphi/M_p} \label{d6-1}$$ where $\phi_0\sim G^{-1}$ and $\beta_{BD}=-1/2\sqrt{\omega_{BD}+3/2}$. When $\omega_{BD}\rightarrow 0$, then $\beta_{BD}\rightarrow \frac{-1}{\sqrt{6}}$ and Einstein frame representations of BD model and $f(R)$ gravity are the same. Cosmological Setting ==================== We use a spatially flat homogeneous and isotropic cosmology described by Friedman-Robertson-Walker spacetime $$ds^2=-dt^2+a^2(t)(dx^2+dy^2+dz^2)$$ where $a(t)$ is the scale factor. We also take $T^m_{\mu\nu}$ as the stress-tensor of a perfect fluid with energy density $\rho_{m}$ and pressure $p_m$. The Friedman equation is $$3H^2=M_p^{-2}(\rho_{m}+\rho_{\varphi}) \label{b14}$$ where $H=\frac{\dot{a}}{a}$, $\rho_{\varphi}=\frac{1}{2}\dot{\varphi}^2+V(\varphi)$ and overdot indicates differentiation with respect to time $t$. From (\[b11\]) and (\[b13\]), we obtain $$\ddot{\varphi}+3\frac{\dot{a}}{a}\dot{\varphi}+\frac{dV(\varphi)}{d\varphi}=-\frac{\beta(\varphi)}{M_p} (\rho_{m}-3p_m) \label{b15}$$ $$\dot{\rho}_{m}+3\frac{\dot{a}}{a}(\omega_m+1)\rho_{m}=Q \label{b17}$$ $$\dot{\rho}_{\varphi}+3\frac{\dot{a}}{a}(\omega_{\varphi}+1)\rho_{\varphi}=-Q \label{b18}$$ where $$Q=\frac{\beta(\varphi)}{M_p} \dot{\varphi}(\rho_{m}-3p_m) \label{b-18}$$ is the interaction term, $\omega_m\equiv p_m/\rho_m$ and $\omega_{\varphi}\equiv p_{\varphi}/\rho_{\varphi}$. The direction of energy transfer depends on the sign of $Q$. For $Q>0$, the energy transfer is from the scalar field to the matter system and for $Q<0$, the reverse is true. The solution of equation (\[b17\]) is $$\rho_m=\rho_{0m}~a^{-3(\omega_m+1)}~e^{\frac{(1-3\omega_m)}{M_p}\int \beta d\varphi} \label{c1}$$ in which $\rho_{0m}$ is an integration constant. This solution indicates that the evolution of energy density is modified due to interaction of $\varphi$ with matter. The expression (\[c1\]) can be written as $$\rho_m=\rho_{0m}~a^{-3(\omega_m+1)+\epsilon} \label{c3}$$ with $\epsilon$ being defined by $$\epsilon=\frac{(1-3\omega_m)\int\beta d\varphi}{M_p\ln a} \label{c2-1}$$ For constant $\beta$, (\[c2-1\]) reduces to $$\varphi=\sigma M_p\ln a \label{c2}$$ with $\sigma$ being a constant defined by the relation $\epsilon=\beta \sigma(1-3\omega_m)$. Here the parameter $\beta$ takes $\beta_{R}$ and $\beta_{BD}$ in $f(R)$ gravity and BD theory, respectively.\ The expression (\[c3\]) states that when $\epsilon>0$, matter is created and energy is constantly injecting into the matter so that the latter will dilute more slowly compared to its standard evolution $\rho_m\propto a^{-3(\omega_m+1)}$. Similarly, when $\epsilon<0$ the reverse is true, namely that matter is annihilated and the direction of energy transfer is outside of the matter system so that the rate of dilution is faster than the standard one.\ Let us first take the matter system to be vacuum energy density characterized by a perfect fluid with equation of state parameter $\omega_m=-1$. Thus the vacuum energy density is anomalously coupled to gravity. In this case, (\[c3\]) becomes $\rho_{\Lambda}\equiv\rho_m=\rho_{0\Lambda}~a^{\epsilon}$ where we have set $\rho_{0m}=\rho_{0\Lambda}$. In an expanding universe, the requirement that a large vacuum energy density $\rho_{\Lambda 0}$ reduces during the expansion needs $\varepsilon<0$. Note that in this case the direction of energy transfer is out of vacuum. We can use Friedman equation (\[b14\]) and the equations (\[b17\]) and (\[b18\]) to write the deceleration parameter $$q(z)=-1-\frac{\dot{H}}{H^2}=-1-\frac{\frac{1}{2}[\epsilon\Omega_{\Lambda}-\frac{Q}{H\rho_c}-3(\omega_{\varphi}+1)\omega_{\varphi}]} {(\Omega_{\Lambda}+\Omega_{\varphi})}\label{cd7}$$ where $\Omega_{\Lambda}=\rho_{\Lambda}/\rho_c$, $\Omega_{\varphi}=\rho_{\varphi}/\rho_c$ and $\rho_c=3H^2_0 M_p^2$ is the critical density. On can use (\[b-18\]) to show that $\frac{Q}{H\rho_c}=\epsilon\Omega_{\Lambda}$. This together with spatial flatness, which requires that $\Omega_{\Lambda}+\Omega_{\varphi}=1$, reduce (\[cd7\]) to $$q(z)=-1+\frac{3}{2}(\omega_{\varphi}+1)(1-\Omega_{\Lambda})\label{dd7}$$ which accelerating expansion requires that $\omega_{\varphi}+1<\frac{2}{3(1-\Omega_{\Lambda})}$.\ The coincidence problem concerns with the constancy of the ratio $r=\frac{\rho_m}{\rho_c}$. Using (\[b17\]) and (\[b18\]), we can write $$\dot{r}=rH[\epsilon(r+1)+(\omega_{\varphi}-\omega_m)]\label{dd7}$$ where we have used $Q=\epsilon H \rho_m$. If we set $\omega_m=0$, the constancy of $r$ requires that $\omega_{\varphi}= -\frac{1}{3}\epsilon (r+1)$ which $\omega_{\varphi}>1$ for a vacuum decay ($\epsilon<0$). Conclusion ========== There is no observational constraints on gravitational coupling of matter systems during expansion of the universe such as baryons, radiation, vacuum, dark matter or dark energy. The weak equivalence principle which is supported by recent observations [@will] only constrains the baryons. Thus it is quite possible that other kinds of matter systems such as vacuum couples differently with gravity.\ We consider a dynamical mechanism which works with expansion of the universe and has its origin in gravitational coupling of the vacuum. We assume that vacuum couples anomalously with gravity in the sense that the metrics of the gravitational and matter parts are not the same but conformally related. The conformal factor is then controlled by a dynamical scalar field. We provide some examples from $f(R)$ gravity and BD theories. We have shown that the cosmological constant problem may be alleviated in such theories due to anomalous gravitational coupling of vacuum. [99]{} S. Weinberg, Rev. Mod. Phys. 61, 1 (1989) S. M. Carroll, Living Rev. Rel. 4, 1 (2001)\ T. Padmanabhan, Phys. Rept. 380, 235 (2003) Y. Bisabr, Phys. Rev. D 82 124041, (2010) H. Salehi and Y. Bisabr, Class. Quant. Grav. 19, 2369 (2002)\ Y. Bisabr, Gen. Rel. Grav. 42, 1211 (2010) J. Khoury and A. Weltman, Phys. Rev. D 69, 044026 (2004)\ J. Khoury and A. Weltman, Phys. Rev. Lett. 93, 171104 (2004) Y. Bisabr, Phys. Lett. B 683, 96 (2010) Ph. Brax, C. van de Bruck, A. Davis, J. Khoury and A. Weltman, Phys. Rev. D 70, 123518 (2004 )\ A. Upadhye, S. S. Gubser and J. Khoury, Phys. Rev. D 74, 04024 (2006)\ Ph. Brax, C. van de Bruck, A. Davis and D. J. Shaw , Phys. Rev. D 78, 104021 (2008 )\ Ph. Brax and K. Zioutas, Phys. Rev. D 82, 043007 (2010 )\ S. M. Carroll, V. Duvvuri, M. Trodden and M. S. Turner, Phys. Rev. D 70, 043528, (2004) S. M. Carroll, A. De Felice, V. Duvvuri, D. A. Easson, M. Trodden and M. S. Turner, Phys. Rev. D 71, 063513, (2005)\ G. Allemandi, A. Browiec and M. Francaviglia, Phys. Rev. D 70, 103503 (2004)\ X. Meng and P. Wang, Class. Quant. Grav. 21, 951 (2004)\ M. E. soussa and R. P. Woodard, Gen. Rel. Grav. 36, 855 (2004)\ S. Nojiri and S. D. Odintsov, Phys. Rev. D 68, 123512, (2003) C. Brans and R. H. Dicke, Phys. Rev. 124, 925 (1961) G. Magnano and L. M. Sokolowski, Phys. Rev. D 50, 5039 (1994) K. Maeda, Phys. Rev. D 39, 3159 (1989)\ D. Wands, Class. Quant. Grav. 11, 269 (1994) G. Esposito-Farese and D. Polarski, Phys. Rev. D 63, 063504 (2001) V. Faraoni, Cosmology in Scalar-Tensor Gravity. Kluwer Academic, Norwell (2004) A. G. Riess, et al., Astrophys. J. 560, 49 (2001)\ T. Padmanabhan and T. Roy Choudhury, Mon. Not. Roy. Astron. Soc.344, 823 (2003)\ L. Amendola, Mon. Not. R. Astron. Soc. 342, 221 (2003) C. M. Will, Theory and Experiment in Gravitational Physics, (Cambridge University Press, 3rd edition, Cambridge, 1993)\ C. M. Will, Liv. Rev. Rel. 9, 3 (2005) [^1]: e-mail: [email protected]. [^2]: One can always redefine the scalar field to reduce $F(\phi)$ and $Z(\phi)$ to only one unknown function.	{ "pile_set_name": "ArXiv" }
--- abstract: 'We report on the potentialities offered by a fixed-target experiment at the LHC using the proton and ion LHC beams (AFTER@LHC project) regarding the study of $J/\psi$ exclusive-photoproduction in $pA$ and $AA$ collisions. The foreseen usage of polarised targets (hydrogen, deuteron, helium) gives access to measurements of the Single-Transverse-Spin Asymmetries of this exclusive process, therefore allowing one to access the helicity-flip Generalised Parton Distribution (GPD) $E_{g}$. We detail the expected yields of photoproduced $J/\psi$ in proton-hydrogen and lead-hydrogen collisions and discuss the statistical uncertainties on the asymmetry measurement for one year of data taking at the LHC.' author: - 'L. Massacrier$^{1}$, J.P. Lansberg$^{1}$, L. Szymanowski$^{2}$ and J. Wagner$^{2}$' bibliography: - 'cernrepexa.bib' title: 'Quarkonium-photoproduction prospects at a fixed-target experiment at the LHC (AFTER@LHC)' --- The AFTER@LHC project ===================== AFTER@LHC is a proposal for a multi-purpose fixed-target programme using the multi-TeV proton or heavy ion beams of the LHC to perform the most energetic fixed-target collisions ever performed so far [@AFTERweb]. If used in a fixed-target mode, the LHC beams offer the possibility to study with high precision pH or $pA$ collisions at $\sqrt{s_{NN}}$ = 115 GeV and PbA collisions at $\sqrt{s_{NN}}$ = 72 GeV, where $A$ is the atomic mass of the target. The fixed-target mode offers several unique assets compared to the collider mode: outstanding luminosities can be reached thanks to the high density of the target; standard detectors can access the far backward center-of-mass (c.m) region thanks to the boost between the colliding-nucleon c.m system and the laboratory system (this region remains completely uncharted with hard reactions); the target species can easily be changed; polarised target can be used for spin studies. The physics opportunities offered by a fixed-target programme at LHC have been published in [@Brodsky:2012vg; @Lansberg:2016urh; @Lansberg:2012kf; @Lansberg:2012wj; @Rakotozafindrabe:2012ei; @Lorce:2012rn; @Lansberg:2012sq; @Rakotozafindrabe:2013au; @Lansberg:2013wpx; @Rakotozafindrabe:2013cmt; @Lansberg:2014myg; @Massacrier:2015nsm; @Massacrier:2015qba; @Lansberg:2016gwm; @Kikola:2017hnp; @Trzeciak:2017csa] and are summarised below in three main objectives. First, whereas the need for precise measurements of the partonic structure of nucleons and nuclei at small momentum fraction $x$ is usually highlighted as a strong motivation for new large-scale experimental facilities, the structure of nucleon and nuclei at high $x$ is probably as poorly known with long-standing puzzles such as the EMC effect [@Aubert:1983xm] in the nuclei or a possible non-perturbative source of charm or beauty quarks in the proton which would carry a relevant fraction of its momentum [@Brodsky:1980pb]. With an extensive coverage of the backward region corresponding to high $x$ in the target, AFTER@LHC is very well placed for performing this physics with a hadron beam. The second objective of AFTER@LHC is the search and characterisation of the Quark-Gluon Plasma (QGP), a deconfined state of nuclear matter, which was prevailing in the universe few microseconds after the Big Bang. QGP is expected to be formed when the surrounding hadronic matter is extremely compressed or heated. These conditions can be achieved in ultra-relativistic Heavy-Ion collisions (HI). AFTER@LHC with a c.m. energy of 72 GeV provides a complementary coverage to the RHIC- and SPS- based experiments, in the region of high temperatures and low baryon-chemical potentials, where the QGP is expected to be produced. AFTER@LHC will provide crucial information about the phase transition by: (i) scanning the longitudinal extension of the hot medium, (ii) colliding systems of different sizes, (iii) analysing the centrality dependence of these collisions. Together they should provide a measurement of the temperature dependence of the system viscosity both as a QGP or a hadron gas. Additionally, measurements of production of various quarkonia states in HI collisions can provide insight into thermodynamic properties of the QGP. Their sequential suppression was predicted to occur in the deconfined partonic medium due to Debye screening of the quark-antiquark potential [@Matsui:1986dk]. However, other effects (Cold Nuclear Matter effects (CNM), feed-down, secondary production via coalescence...) can also alter the observed yields. AFTER@LHC will provide a complete set of quarkonia measurements (together with open heavy flavours) allowing one to access the temperature of the formed medium in $AA$ collisions, and cold nuclear matter effects in $pA$ ($AA$) collisions. Thanks to the large statistics expected, a golden probe will be the measurement of $\Upsilon$(nS) production in $pp$, $pA$ and $AA$ collisions, allowing one to calibrate the quarkonium thermometer and to search for the phase transition by looking at $\Upsilon$(nS) suppression (and other observables like charged hadron $v_{2}$) as a function of rapidity and the system size. Finally, despite decades of efforts, the internal structure of the nucleon and the distribution and dynamics of its constituents are still largely unknown. One of the most significant issues is our limited understanding of the spin structure of the nucleon, especially how its elementary constituents (quarks and gluons) bind into a spin-half object. Among others, Single Transverse Spin Asymmetries (STSA) in different hard-scattering processes are powerful observables to further explore the dynamics of the partons inside hadrons [@DAlesio:2007bjf]. Thanks to the large yields expected with AFTER@LHC, STSA of heavy-flavours and quarkonia –which are currently poorly known– could be measured with high accuracy, if a polarised target can be installed [@Kikola:2017hnp]. We show here that AFTER@LHC can also rely on quarkonium exclusive-photoproduction processes to explore the three-dimensional tomography of hadrons via Generalised Parton Distributions (GPDs) [@Diehl:2003ny]. Photoproduction is indeed accessible in Ultra-Peripheral Collisions (UPCs) and the quarkonium mass presumably sets the hard scale to use collinear factorisation in terms of (gluon) GPDs, which are directly related to the total angular momentum carried by quarks and gluons. In fact, exclusive $J/\psi$ production [@Ivanov:2004vd] drew much attention in the recent years due to the fact that it is sensitive, even at leading order, to gluon GPDs. Beside, with the addition of a polarised hydrogen[^1] target, AFTER@LHC opens a unique possibility to study STSA of such process, which is sensitive to yet unknown GPD $E_{g}$ [@Koempel:2011rc]. In this contribution, we report on such studies through the collisions of proton and lead beams onto a polarised hydrogen target at AFTER@LHC energies. Possible technical implementations at the LHC and projected luminosities ======================================================================== Several promising technical solutions exist in order to perform fixed-target collisions at the LHC. One can either use an internal (solid or gaseous) target coupled to an already existing LHC detector or build a new beam line together with a new detector. The first solution can be achieved in a shorter time scale, at limited cost and civil engineering. Moreover the fixed-target programme can be simultaneously run with the current LHC collider experiments, without affecting the LHC performances. The direct injection of noble gases in the LHC beam pipe is currently being used by the LHCb collaboration with the SMOG device [@Aaij:2014ida]. However, this system has some limitations, in particular: (i) the gas density achieved inside the Vertex Locator of LHCb is small (of the order of $10^{-7}$ mbar); (ii) there is no possibility to inject polarised gas; (iii) there is no dedicated pumping system close to the target; (iv) the data taking time has so far been limited to few days in a row. The use of more complex systems with higher density gaseous and polarised target inside one of the existing LHC experiment is under study. For instance, an hydrogen gas jet is currently used at RHIC to measure the proton beam polarisation [@Zelenski:2005mz]. The H-jet system consists of a polarised free atomic beam source cooled down to 80K, providing an hydrogen inlet flux of 1.3 $\times 10^{7}$ H/s. With such a device, the gas density can be increased by about one order of magnitude with respect to the SMOG device and probably be continuously run. Another promising alternative solution is the use of an openable storage cell placed inside the LHC beam pipe. Such a system was developed for the HERMES experiment [@Barschel:2015mka; @Steffens:2015kvp]. Polarised hydrogen, deuteron and Helium-3 at densities about 200 times larger than the ones of the H-jet system can be injected, as well as heavier unpolarised gases. Fixed-target collisions can also be obtained with a solid target (wire, foil) interacting with the LHC beams. There are two ways of doing so: either thanks to a system which permit to move directly the target inside the beam halo [@Kurepin:2011zz]; or by using a bent crystal (see work by UA9 collaboration [@Scandale:2010zzb]) upstream of an existing experiment ($\sim$ 100 m) to deflect the beam halo onto the fixed target. In both cases, the target (or an assembly of several targets) can be placed at a few centimetres from the nominal interaction point, allowing one to fully exploit the performances of an existing detector. The usage of the bent crystal offers the additional advantage to better control the flux[^2] of particles sent on the target, and therefore the luminosity determination. Most probably such simple solid targets could not be polarised. The spin physics part of the AFTER@LHC programme could naturally not be conducted with such an option. Tab. \[Tab1\] summarises the target areal density, the beam flux intercepting the target, the expected instantaneous and yearly integrated luminosities for $p$H and PbH collisions, for the possible technical solutions described above. Luminosities as large as 10 fb$^{-1}$ can be collected in $p$H collisions in one LHC year with a storage cell. In PbH collisions, similar luminosities ($\sim$ 100 nb$^{-1}$) can be reached with a gas-jet or a storage cell since the gas density has to be levelled in order to avoid a too large beam consumption[^3]. We stress that these are annual numbers and they can be cumulated over different runs. \[Tab1\] Prospects for $J/\psi$ photoproduction studies with AFTER@LHC ============================================================= Let us now quickly summarise our feasibility study: 100 000 photoproduced $J/\psi$ have been generated in the dimuon decay channel, using STARLIGHT Monte Carlo (MC) generator [@Klein:1999qj; @Baltz:2002pp; @Klein:2003vd; @Baltz:2009jk] in $p$H$^{\uparrow}$ ($\sqrt{s}$ = 115 GeV) and Pb+H$^{\uparrow}$ collisions ($\sqrt{s_{NN}}$ = 72 GeV). In $p$H$^{\uparrow}$ collisions, both protons can be photon emitters, while in Pb+H$^{\uparrow}$ only the Pb nuclei was considered as photon emitter (dominant contribution). The $J/\psi$ photoproduction cross sections given by STARLIGHT MC are summarised in Tab. \[Tab2\]. In order to mimic an LHCb-like forward detector, the following kinematical cuts have been applied at the single muon level: 2 $< \eta_{\rm lab}^{\mu} <$ 5 and $p_{\rm T}^{\mu} >$ 0.4 GeV/c. Fig. \[fig-1\] shows the rapidity-differential (left) and $p_{\rm T}$-differential (right) cross sections of the photoproduced $J/\psi$ in the dimuon decay channel in $p$H$^{\uparrow}$ collisions, generated with STARLIGHT MC generator. The blue curves have been produced without applying any kinematical cuts, while the red curves are produced by applying the cuts: 2 $< \eta_{\rm lab}^{\mu} <$ 5 and $p_{\rm T}^{\mu} >$ 0.4 GeV/c at the single muon level. On the left panel is also indicated the photon-proton c.m energy (W$_{\gamma p}$), calculated as: $$W_{\gamma p} = \sqrt{M_{J/\psi}^{2} + M_{p}^{2} + 2 \times M_{p} \times M_{J/\psi} \times \cosh (\rm y_{\rm lab})},$$ with $M_{J/\psi}$ and $M_{p}$, being respectively the $J/\psi$ and proton masses, and $y_{\rm lab}$ the $J/\psi$ rapidity in the laboratory frame. On the vertical axis is shown as well the photoproduction $J/\psi$ yield per year per 0.1 $y_{\rm lab}$ unit (left) and per 0.1 GeV/c unit (right). An integrated yearly luminosity of $\cal{L_{\rm int}}$ = 10 fb$^{-1}$, corresponding to the storage cell solution, has been assumed in $p$H$^{\uparrow}$ collisions in order to calculate the $J/\psi$ yearly photoproduction yield. About 200 000 photoproduced $J/\psi$ are expected to be detected in an LHCb-like acceptance per year with AFTER@LHC[^4]. Similarly to Fig. \[fig-1\], Fig. \[fig-2\] shows the rapidity-differential (left) and $p_{\rm T}$-differential (right) cross sections of the photoproduced $J/\psi$ in the dimuon decay channel in PbH$^{\uparrow}$ collisions, generated with STARLIGHT MC generator. The collection of an integrated luminosity of 100 nb$^{-1}$ per year is expected at AFTER@LHC with the storage cell option. This would result in about 1 000 photoproduced $J/\psi$ per year emitted in an LHCb-like acceptance. \[Tab2\] ![image](Fig1procleft){width="7.9cm"} ![image](Fig1procright){width="7.9cm"} ![image](Fig2procleft){width="7.9cm"} ![image](Fig2procright){width="7.9cm"} In a forthcoming publication [@forthcoming], we will report on the evaluation of the uncertainty on the STSA, $A_{N}$, from the photoproduced-$J/\psi$ yields obtained with STARLIGHT MC and the expected modulation for $E_g$ following [@Koempel:2011rc]. Indeed, $A_{N}$, the amplitude of the spin-correlated azimuthal modulation of the produced particles, is defined as $A_{N} = \frac{1}{P}\frac{N^{\uparrow} - N^{\downarrow}}{N^{\uparrow} + N^{\downarrow}}$, where $N^{\uparrow}$ ($N^{\downarrow}$) are the photoproduced-$J/\psi$ yields for an up (down) target-polarisation orientation, and $P$ is the effective polarisation of the target. The statistical uncertainty $u_{A_{N}}$ on $A_{N}$ can then be derived using $ u_{A_{N}} = \frac{2}{P(N^{\uparrow} + N^{\downarrow})^{2}}\sqrt{N^{\downarrow2}u^{\uparrow2} + N^{\uparrow2}u^{\downarrow2}}$ with $u^\uparrow$ ($u\downarrow$) the relative uncertainties on the $J/\psi$ yields with up (down) polarisation orientation. Dividing the sample into two $J/\psi$ $p_{\rm T}$ ranges relevant for GPDs extraction: 0.4 $< p_{\rm T} <$ 0.6 GeV/c and 0.6 $< p_{\rm T} <$ 0.8 GeV/c, the expected statistical precision on $A_{N}$ is already expected to be better than 10$\%$ given the integrated yearly yield of 200 000 photoproduced $J/\psi$ in $pH^{\uparrow}$ collisions at AFTER@LHC. This will also allow for an extraction of A$_{N}$ as a function Feynman-$x$, $x_{\rm F}$ [^5]. Conclusion ========== We have presented projections for $J/\psi$ photoproduction measurements in polarised $p$H$^{\uparrow}$ and Pb+H$^{\uparrow}$ collisions after one year of data taking at AFTER@LHC energies and assuming a storage cell technology coupled to an LHCb-like forward detector. In $p$H$^{\uparrow}$ collisions, a yearly yield of about 200 000 photoproduced $J/\psi$ is expected, allowing one to reach a very competitive statistical accuracy on the $A_{N}$ measurement differential in $x_{\rm F}$. A non-zero asymmetry would be the first signature of a non-zero GPD $E_g$ for gluons. Acknowledgements {#acknowledgements .unnumbered} ================ We thank D. Kikola, S. Klein, A. Metz, J. Nystrand and B. Trzeciak for useful discussions. This work is partly supported by the COPIN-IN2P3 Agreement, the grant of National Science Center, Poland, No. 2015/17/B/ST2/01838, by French-Polish scientific agreement POLONIUM, by the French P2IO Excellence Laboratory and by the French CNRS-IN2P3 (project TMD@NLO). [^1]: Measurements with deuteron and helium targets are also considered. [^2]: The deflected halo beam flux is considered to be about 5 $\times$ 10$^{8}$ p/s and 10$^{5}$ Pb/s. [^3]: Assuming a total cross section $ \sigma_{\rm PbH}$ = 3 barn, 15$\%$ of the beam is used over a fill. [^4]: The detector efficiency has been assumed to be 100$\%$. [^5]: $x_{\rm F} = 2 \times M_{J/\psi} \times \sinh ({y_{\rm cms}}/{\sqrt{s}})$, with $y_{cms}$ being the $J/\psi$ rapidity in the c.m. system frame and $\sqrt{s}$ the c.m energy	{ "pile_set_name": "ArXiv" }
--- abstract: 'We propose a technique aimed at cooling a harmonically oscillating mirror mechanically coupled to another vibrating mirror to its quantum mechanical ground state. Our method involves optmechanical coupling between two optical cavities. We show that the cooling can be controlled by the mechanical coupling strength between the two movable mirrors, the phase difference between the mechanical modes of the two oscillating mirrors and the photon number in each cavity. We also show that both mechanical and optical cooling can be achieved by transferring energy from one cavity to the other. We also analyze the occurrence of normal-mode splitting (NMS). We find that a hybridization of the two oscillating mirrors with the fluctuations of the two driving optical fields occurs and leads to a splitting of the mechanical and optical fluctuation spectra.' address: - '$^{1}$Department of Physics and Astrophysics, University of Delhi, Delhi-110007, India' - '$^{2}$Department of Physics, ARSD College, University of Delhi (South Campus), New Delhi-110021, India' author: - 'Tarun Kumar$^{1}$, Aranya B. Bhattacherjee$^{2}$ and ManMohan$^{1}$' title: 'Optomechanical coupling between two optical cavities: cooling of a micro-mirror and parametric normal mode splitting' --- Introduction ============ In recent years mechanical and optical degrees of freedom have become entangled experimentally by underlying mechanism of radiation pressure forces. This field known as cavity optomechanics has played a vital role in the conceptual exploration of the boundaries between classical and quantum mechanical systems. The coupling of mechanical and optical degrees of freedom via radiation pressure has been a subject of early research in the context of laser cooling [@hansch; @wineland; @chu] and gravitational-wave detectors [@caves]. Recently there has been a great surge of interest in the application of radiation forces to manipulate the center-of-mass motion of mechanical oscillators covering a huge range of scales from macroscopic mirrors in the Laser Interferometer Gravitational Wave Observatory (LIGO) project [@corbitt1; @corbitt2] to nano-mechanical cantilevers[@hohberger; @gigan; @arcizet; @kleckner; @favero; @regal], vibrating microtoroids[@carmon; @schliesser], membranes[@thompson] and Bose-Einstein condensates [@brennecke; @murch; @bhattacherjee; @bhattacherjee2]. The central accomplishment of the field of cavity optomechanics is the investigation of radiation pressure forces which allow one to manipulate the motional state of micromechanical oscillators. In particular, it has become possible to substantially cool the thermal excitation of a single mechanical mode, down to a few tens of remaining phonons [@schliesser_2]. With these developments, micro- and nanomechanical resonators now represent an important model system with the prospect of demonstrating quantum effects on a macroscopic scale. Theoretical work has also proposed to use the radiation-pressure coupling for quantum non-demolition measurements of the light field [@braginsky]. It has been shown that ground state cooling of micro-mechanical mirror is possible only in the resolved side band regime (RSB) where the mechanical resonance frequency exceeds the bandwidth of the driving resonator [@marquardt; @braginsky]. The cooling of mechanical oscillators in the RSB regime at high driving power can entail the appearance of normal mode splitting (NMS) [@dobrindt]. Recently, it was shown that an optical parametric amplifier inside a cavity considerably improves the cooling of a micro-mechanical mirror by radiation pressure [@huang]. Recently, dynamics of a micro mirror was studied in the presence of a nonlinear kerr medium placed inside an optical cavity [@tarun]. It was demonstrated that due to the photon blockade mechanism, as the Kerr nonlinearity is increased, the NMS progressively decreases. The Kerr medium was found to be a new handle to efficiently control the micro-mirror dynamics and this suggests a possible quantum device [@tarun]. In this work, we propose a technique aimed at cooling a harmonically oscillating mirror (mechanically coupled to another vibrating mirror) to its quantum mechanical ground state. Our method involves optmechanical coupling between two optical cavities. We show that the cooling can be controlled by the mechanical coupling strength between the two movable mirrors and the phase difference between the mechanical modes of the two oscillating mirrors. We also analyze the occurrence of normal-mode splitting (NMS). We find that a hybridization of the two oscillating mirrors with the fluctuations of the two driving optical fields occurs and leads to a splitting of the mechanical and optical fluctuation spectra. The continuous variable entanglement between two mechanical modes could be used to improve the detection of weak classical forces in optomechanical devices as atomic force microscopes or gravitational wave detectors. Optomechanically coupled mirrors has been investigated earlier [@mancini; @ludwig]. A continuous variable entanglement between the two mirrors was maintained by the light bouncing between the mirrors and was found to be robust against thermal noise [@mancini]. Entanglement between two mechanical oscillators coupled to a nonequilibrium environment showed that there is an optimal dissipation strength for which the entanglement between two coupled oscillators is maximized [@ludwig]. A new cooling method which involves the two-sided irradiation of the vibrating mirror inside an optical cavity has been proposed recently [@bhattacharya]. This method provides a stiffer trap for cooling the mirror and has several advantages over conventional methods of optomechanical cooling. Theoretical Framework ===================== We consider two Fabry-Perot cavities connected with each other through their movable mirrors as shown in fig. 1. Here mirror M1 and mirror M4 are fixed and are partially transmitting whereas mirrors M2 and M3 are movable and totally reflecting. The two mirrors M2 and M3 can both oscillate under the effect of the radiation pressure. The motion of each mirror is the result of the excitation of many oscillation modes which can be either external or internal. The external modes corresponds to pendulum modes which leads to global displacements of the mirror while the internal modes corresponds to deformations of the mirror surface due to excitation of internal acoustic modes of the mirror surface. These various degrees of freedom have different resonance frequencies and experimentally it is possible to select the mechanical response of a single mode by using a bandpass filter in the detection circuit. Consequently, we will consider a single mechanical mode for each movable mirror, which will be therefore described as a simple harmonic oscillator. The system under consideration is in contact with thermal bath in equilibrium with thermal bath at temperature T. The movable mirrors are treated as quantum mechanical oscillators with masses m1 and m2, frequencies $\nu1$ and $\nu2$, and energy decay rates $\gamma_{m1}$ and $\gamma_{m2}$ respectively of the mirrors M1 and M4. The system is coherently driven by two laser fields ($a_{1,in}$ and $a_{2,in}$) with frequencies $\omega_{1}$ and $\omega_{2}$ as shown in fig. 1. It is well known that high Q-optical cavities can significantly isolate the system from its environment, thus strongly reducing decoherence and ensuring that the light field remains quantum mechanical for the duration of the experiment. We also assume that the induced cavity resonance frequency shift of each cavity is much smaller than the longitudinal spacing, so that we restrict the model to a single longitudinal mode for each cavity. Let $\epsilon_{1}$ and $\epsilon_{2}$ be the amplitudes of the two laser fields. As we know in a Fabry-perot cavity, when a photon collides with the surface of the movable mirror, it exerts radiation pressure on the mirror and the force that mirror will experience is proportional to the photon number inside the cavity. But in our system here, the force experienced by one of the movable mirror, say M2 not only depends on the number on photons of the corresponding cavity but also depends on the number of photons of the second cavity. This is because the two mirrors are coupled, therefor position of one mirror is influenced by the position of the other mirror. We also assume that $\omega_{1}$, $\omega_{2}<<$ $\pi$c/L (adiabatic limit); c is the speed of light in vacuum and L is the cavity length in the absence of the cavity field. (assuming same length for the two cavities). The Hamiltonian of the system can be written as ![Schematic description of the system under study.Two Fabry-Perot cavities are connected with each other through their movable mirrors. Here mirror M1 and mirror M4 are fixed and are partially transmitting whereas mirrors M2 and M3 are movable and totally reflecting. The system is coherently driven by two laser fields $a_{1,in}$ and $a_{2,in}$. $a_{1,out}$ and $a_{2,out}$ are the output fields.[]{data-label="1"}](fig1.eps) $$\begin{aligned} \label{equation1} H&=& \hbar \omega_{1}a_{1}^{\dagger}a_{1} + \hbar \omega_{2}a_{2}^{\dagger}a_{2} + \hbar \nu_{1}(b_{1}^{\dagger}b_{1}+1/2) + \hbar \nu_{2}(b_{2}^{\dagger}b_{2}+1/2) + \hbar \nu_{12}(b_{1}^{\dagger}e^{-i \theta_{1}} + b_{1}e^{i \theta_{1}})(b_{2}^{\dagger}e^{-i \theta_{2}} + b_{2}e^{i \theta_{2}}) \nonumber \\ &-& \hbar g_{1}a_{1}^{\dagger}a_{1} (b_{1}^{\dagger}e^{-i \theta_{1}} + b_{1}e^{ i \theta_{1}}) + i \hbar \epsilon_{1}(a_{1}^{\dagger}-a_{1}) - \hbar g_{2}a_{2}^{\dagger}a_{2} (b_{2}^{\dagger}e^{-i \theta_{2}} + b_{2}e^{i \theta_{2}}) + i \hbar \epsilon_{2}(a_{2}^{\dagger}-a_{2})\end{aligned}$$ Here $a_{1}$ ($a_{1}^{\dagger}$) and $a_{2}$ ($a_{2}^{\dagger}$) are the annihilation (creation) operators of the two cavity fields, $b_{1}$ ($b_{1}^{\dagger}$) and $b_{2}$ ($b_{2}^{\dagger}$) are the phonon annihilation (creation) operators of the two movable mirrors M2 and M3 respectively. The parameters $g_{1}$ and $g_{2}$ are the coupling parameters between the cavity fields and fixed mirrors M1 and M4 respectively, $\nu_{12}$ is the coupling frequency of the two movable mirrors and $\theta_{1}$ and $\theta_{2}$ are the phases of the two movable mirrors. The phases $\theta_{1}$ and $\theta_{2}$ can be thought of as arising from the complex mirror-photon coupling strengths $g_{i} (i=1,2)$. The system we are considering here is intrinsically open as the cavity fields are damped by the photon leakage through the massive coupling mirrors. In the absence of the radiation pressure coupling, the cantilevers would undergo pure Brownian motion, driven by their contact with the thermal environment. The motion of the system can be described by the following Quantum Langevin equations- $$\label{equation1(a)} \dot{q_{1}}=\nu_{1}p_{1}+Ap_{2}+Bq_{2}+2D_{2}\|a_1\|^{2}-\gamma_{m1}q_{1}$$ $$\label{equation1(b)} \dot{q_{2}}=\nu_{2}p_{2}+Ap_{1}-Bq_{1}+2D_{4}\|a_2\|^{2}-\gamma_{m2}q_{2}$$ $$\label{equation1(c)} \dot{p_{1}}=-\nu_{1}q_{1}-Aq_{2}+Bp_{2}+2D_{1}\|a_1\|^{2}-\gamma_{m1}p_{1}+\sqrt{2\gamma_{m1}}p_{in}^{1}$$ $$\label{equation1(d)} \dot{p_{2}}=-\nu_{2}q_{2}-Aq_{1}-Bp_{1}+2D_{3}\|a_2\|^{2}-\gamma_{m2}p_{2}+\sqrt{2\gamma_{m2}}p_{in}^{2}$$ $$\label{equation1(e)} \dot{a_{1}}=-i \omega_{1}a_{1}+ i D_{1}a_{1}q_{1}-i D_{2}a_{1}p_{1}-\kappa_{1}a_{1}+\sqrt{2\kappa_{1}}C_{in}^{1}+\epsilon_{1}$$ $$\label{equation1(f)} \dot{a_{2}}=-i \omega_{2}a_{2}+ i D_{3}a_{2}q_{2}-i D_{4}a_{2}p_{2}-\kappa_{2}a_{2}+\sqrt{2\kappa_{2}}C_{in}^{2}+\epsilon_{2},$$ where we have defined ($b_{i}$+$b_{i}^{\dagger}$)=$q_{i}$ and $i$($b_{i}^{\dagger}$-$b_{i}$)=$p_{i}$ ; $i=1,2$. Also $D_{1}$ = $g_{1}\cos(\theta_{1})$, $D_{2}$ = $g_{1}\sin(\theta_{1})$, $D_{3}$ = $g_{2}\cos(\theta_{2})$, $D_{4}$ = $g_{2}\sin(\theta_{2})$, $A$ = $\nu_{12}\cos(\theta_{2}-\theta{1})$, $B$ = $\nu_{12}\sin(\theta_{2}-\theta{1})$. $C_{in}^1$ and $C_{in}^2$ are input noise operators with zero mean value and obeys following commutation relation $<\delta C_{in}^{i}\delta C_{in}^{j\dagger}>$ = $\delta_{ij}(t-t^{'})$, $<\delta C_{in}^{i}\delta C_{in}^{j}>$, $<\delta C_{in}^{i\dagger}\delta C_{in}^{j\dagger}> = 0$. Also $p_{in}^i$ = $ i(\xi^{i\dagger}-\xi^{i})$, $\xi$ is the Brownian noise operator, arising due to the thermal bath. Brownian noise operator has zero mean value and obeys following correlation at temperature $T$: $<\xi^{i}(t)\xi^{j\dagger}(t^{'})> = 2\gamma_{mi}(1+2n_{T})\delta_{ij}(t-t^{'})$ and $<\xi^{i}(t)\xi^{j}(t^{'})> = <\xi^{i\dagger}(t)\xi^{j\dagger}(t^{'})> = <\xi^{i\dagger}(t)\xi^{j}(t^{'})>=0$, $n_T = coth(\frac{\hbar\omega}{2k_BT})$, where $k_B$ is the Boltzmann constant and $T$ is the temperature of the thermal bath. Small Fluctuations Dynamics: Normal Mode Splitting and Cooling of a micro mirror ================================================================================ Here we show that the coupling of the two mechanical oscillators and the two cavity field fluctuations leads to the splitting of the normal mode into two modes (Normal Mode Splitting(NMS)) for each cavity depending on the system parameters. The optomechanical NMS however involves driving four parametrically coupled nondegenerate modes out of equilibrium. The NMS does not appear in the steady state spectra but rather manifests itself in the fluctuation spectra of the mirror displacement. In order to study the dynamics of the coupled mirror, we need to find out the fluctuations in the mirror’s position. As is clear from the equations 2-7 that the problem involved here is non-linear. We assume that this non-linearity is small. Therefor we study the dynamics of fluctuations around the steady state of the system. We write each canonical operator of the system as a sum of its steady state mean value and a small fluctuation with zero mean value, $q_{1}\rightarrow q_{1s}+\delta q_{1}$, $q_{2}\rightarrow q_{2s}+\delta q_{2}$, $p_{1}\rightarrow p_{1s}+\delta p_{2}$, $p_{2}\rightarrow p_{2s}+\delta p_{2}$, $a_{1}\rightarrow a_{1s}+\delta a_{1}$, $a_{2}\rightarrow a_{2s}+\delta a_{2}$. The steady state values are obtained by putting the left hand side of Eqns.(2)-(7) to zero. In order to achieve ground state cooling, we will always take $\gamma_{mi}$ $<<$ $\kappa_{i} $, $g_{i}$ $<$ $\nu_{i}$ and $\nu_{i}$ $>$ $\kappa_{i}$ (with $i=1,2$) The last condition is the resolved side band regime necessary for ground state cooling. Note that these conditions necessary for cooling also implies that the system is stable. Linearizing equation 2 to 7 to obtain following Heisenberg - Langevin equations for the fluctuation operators : $$\dot{\delta q_{1}}= \nu_{1}\delta p_{1} + A\delta p_{2} + B\delta q_{2} + 2D_{2}a_{1s}\delta a_{1}^{\dagger} + 2D_{2}a_{1s}^{}\delta a_{1}$$ $$\dot{\delta q_{2}}= \nu_{2}\delta p_{1} + A\delta p_{1} - B\delta q_{1} + 2D_{4}a_{2s}\delta a_{2}^{\dagger} + 2D_{4}a_{2s}^{}\delta a_{2}$$ $$\dot{\delta p_{1}}= -\nu_{1}\delta q_{1} - A\delta q_{2} + B\delta p_{2} + 2D_{1}a_{1s}\delta a_{1}^{\dagger} + 2D_{1}a_{1s}^{}\delta a_{1}+ \sqrt{2\gamma_{m2}}\delta p_{in}^1-\gamma_{m1} \delta p_{1}$$ $$\dot{\delta p_{2}}= -\nu_{2}\delta q_{2} - A\delta q_{1} - B\delta p_{1} + 2D_{3}a_{2s}\delta a_{2}^{\dagger} + 2D_{3}a_{2s}^{}\delta a_{2} + \sqrt{2\gamma_{m2}}\delta p_{in}^2-\gamma_{m2}\delta p_{2}$$ $$\dot{\delta a_{1}}= - i \omega_{1}\delta a_{1} + i D_{1}(q_{1s}\delta a_{1} + a_{1s}\delta q_{1}) - i D_{2}(a_{1s}\delta p_{1} + p_{1s}\delta a_{1}) - \kappa_{1}\delta a_{1} + \sqrt{2\kappa_{1}}\delta C_{in}^1$$ $$\dot{\delta a_{2}}= -i \omega_{2}\delta a_{2} + i D_{3}(q_{2s}\delta a_{2} + a_{2s}\delta q_{2}) - i D_{4}(a_{2s}\delta p_{2} + p_{2s}\delta a_{2}) - \kappa_{2}\delta a_{2} + \sqrt{2\kappa_{2}}\delta C_{in}^2$$ On fourier transforming all operators and noise sources of equation 9 to 14 and solving in the frequency domain, the position fluctuations $\delta q_{1}(\omega)$ of the movable mirror M2 is obtained as - $$\begin{aligned} \delta q_{1}(\omega) &=& \frac{1}{d(\omega)}\{X(\omega)(\frac{a_{1s}^}{\kappa_1- i (\omega - \Delta_1)}\delta C_{in}^1 + \frac{a_{1s}}{\kappa_1-i (\omega + \Delta_1)}\delta C_{in}^{1\dagger})2\sqrt{2\kappa_1} \nonumber \\ &+& Y(\omega)(\frac{a_{2s}^}{\kappa_2- i (\omega - \Delta_2)}\delta C_{in}^2 + \frac{a_{2s}}{\kappa_2- i (\omega + \Delta_2)}\delta C_{in}^{1\dagger})2\sqrt{2\kappa_2} \nonumber \\ &+& (\sqrt{2 \gamma_{m1}}Z(\omega)\delta p_{in}^1 + \sqrt{2 \gamma_{m2}}T(\omega) \delta p_{in}^2)\}\end{aligned}$$ All the variables are defined in the appendix. In equation 14, the first two terms corresponding to $X(\omega)$ and $Y(\omega)$ gives rise the effect of radiation pressure whereas last two terms corresponding to $Z(\omega)$ and $T(\omega)$ originate because of the thermal noise. The coupling to the mirror shifts the cavity resonance frequency and changes the field inside the cavity in a way to induce a new stationary intensity. The shift in the cavity resonance is seen in the renormalized detunings $\Delta_{1}$ and $\Delta_{2}$. The change occurs after a transient time depending on the response of the cavity and the strength of the coupling to the mirrors. Now the spectrum of fluctuation of mirror can be defined as - $$S_{q}(\omega) = \frac{1}{4\pi}\int d\Omega e^{- i(\omega+\Omega)t}<\delta q(\omega)\delta q(\Omega) + \delta q(\Omega)\delta q(\omega)>$$ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ×![Contour plot of the displacement spectrum $S_{q1}$ as a function of normalized effective detuning $\Delta_{1}$ for the following parameters: $\theta_1 = 0$, $\theta_2 = \pi/2$,$\Delta_{2}/\nu=0$, $\gamma_1/\nu$ $=$ $\gamma_2/\nu$ $=$ $0.01$, $g_1/\nu$ $=$ $g_2/\nu$ $=$ $0.2$, $\nu_{12}/\nu$ $=$ $0.03$, $\kappa_1/\nu$ $=$ $\kappa_2/\nu$ $=$ $0.1$. (a): $\|a_{1s}\|^{2}=0.1$ and $\|a_{2s}\|^{2}=0.1$, (b):$\|a_{1s}\|^{2}=0.1$ and $\|a_{2s}\|^{2}=0.25$, (c): $\|a_{1s}\|^{2}=0.25$ and $\|a_{2s}\|^{2}=0.1$, (d):$\|a_{1s}\|^{2}=0.25$ and $\|a_{2s}\|^{2}=0.25$. The values of the color scheme is shown in Fig.5 in the appendix. []{data-label="figure2"}](s1.eps "fig:") ![Contour plot of the displacement spectrum $S_{q1}$ as a function of normalized effective detuning $\Delta_{1}$ for the following parameters: $\theta_1 = 0$, $\theta_2 = \pi/2$,$\Delta_{2}/\nu=0$, $\gamma_1/\nu$ $=$ $\gamma_2/\nu$ $=$ $0.01$, $g_1/\nu$ $=$ $g_2/\nu$ $=$ $0.2$, $\nu_{12}/\nu$ $=$ $0.03$, $\kappa_1/\nu$ $=$ $\kappa_2/\nu$ $=$ $0.1$. (a): $\|a_{1s}\|^{2}=0.1$ and $\|a_{2s}\|^{2}=0.1$, (b):$\|a_{1s}\|^{2}=0.1$ and $\|a_{2s}\|^{2}=0.25$, (c): $\|a_{1s}\|^{2}=0.25$ and $\|a_{2s}\|^{2}=0.1$, (d):$\|a_{1s}\|^{2}=0.25$ and $\|a_{2s}\|^{2}=0.25$. The values of the color scheme is shown in Fig.5 in the appendix. []{data-label="figure2"}](s2.eps "fig:") ![Contour plot of the displacement spectrum $S_{q1}$ as a function of normalized effective detuning $\Delta_{1}$ for the following parameters: $\theta_1 = 0$, $\theta_2 = \pi/2$,$\Delta_{2}/\nu=0$, $\gamma_1/\nu$ $=$ $\gamma_2/\nu$ $=$ $0.01$, $g_1/\nu$ $=$ $g_2/\nu$ $=$ $0.2$, $\nu_{12}/\nu$ $=$ $0.03$, $\kappa_1/\nu$ $=$ $\kappa_2/\nu$ $=$ $0.1$. (a): $\|a_{1s}\|^{2}=0.1$ and $\|a_{2s}\|^{2}=0.1$, (b):$\|a_{1s}\|^{2}=0.1$ and $\|a_{2s}\|^{2}=0.25$, (c): $\|a_{1s}\|^{2}=0.25$ and $\|a_{2s}\|^{2}=0.1$, (d):$\|a_{1s}\|^{2}=0.25$ and $\|a_{2s}\|^{2}=0.25$. The values of the color scheme is shown in Fig.5 in the appendix. []{data-label="figure2"}](s3.eps "fig:") ![Contour plot of the displacement spectrum $S_{q1}$ as a function of normalized effective detuning $\Delta_{1}$ for the following parameters: $\theta_1 = 0$, $\theta_2 = \pi/2$,$\Delta_{2}/\nu=0$, $\gamma_1/\nu$ $=$ $\gamma_2/\nu$ $=$ $0.01$, $g_1/\nu$ $=$ $g_2/\nu$ $=$ $0.2$, $\nu_{12}/\nu$ $=$ $0.03$, $\kappa_1/\nu$ $=$ $\kappa_2/\nu$ $=$ $0.1$. (a): $\|a_{1s}\|^{2}=0.1$ and $\|a_{2s}\|^{2}=0.1$, (b):$\|a_{1s}\|^{2}=0.1$ and $\|a_{2s}\|^{2}=0.25$, (c): $\|a_{1s}\|^{2}=0.25$ and $\|a_{2s}\|^{2}=0.1$, (d):$\|a_{1s}\|^{2}=0.25$ and $\|a_{2s}\|^{2}=0.25$. The values of the color scheme is shown in Fig.5 in the appendix. []{data-label="figure2"}](s4.eps "fig:") ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ![ Contour plot of the displacement spectrum $S_{q1}$ as a function of dimensionless effective detuning $\Delta_{1}/\nu$ for the following parameters: $\Delta_{2}/\nu=0$, $\gamma_1/\nu$ $=$ $\gamma_2/\nu$ $=$ $0.01$, $g_1/\nu$ $=$ $g_2/\nu$ $=$ $0.2$, $\nu_{12}/\nu$ $=$ $0.03$, $\kappa_1/\nu$ $=$ $\kappa_2/\nu$ $=$ $0.1$, $\|a_{1s}\|^{2}=0.1$, $\|a_{2s}\|^{2}=0.1$, $\theta_1 = 0$, $\theta_2 = \pi/4$ (left plot), $\theta_1 = 0$, $\theta_2 = 3\pi/4$ (right plot).[]{data-label="figure3"}](oscl_2.eps "fig:") ![ Contour plot of the displacement spectrum $S_{q1}$ as a function of dimensionless effective detuning $\Delta_{1}/\nu$ for the following parameters: $\Delta_{2}/\nu=0$, $\gamma_1/\nu$ $=$ $\gamma_2/\nu$ $=$ $0.01$, $g_1/\nu$ $=$ $g_2/\nu$ $=$ $0.2$, $\nu_{12}/\nu$ $=$ $0.03$, $\kappa_1/\nu$ $=$ $\kappa_2/\nu$ $=$ $0.1$, $\|a_{1s}\|^{2}=0.1$, $\|a_{2s}\|^{2}=0.1$, $\theta_1 = 0$, $\theta_2 = \pi/4$ (left plot), $\theta_1 = 0$, $\theta_2 = 3\pi/4$ (right plot).[]{data-label="figure3"}](oscl_1.eps "fig:") ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- The displacement spectrum of mirror M2 i.e, $S_{q1}(\omega)$ is finally obtained as - $$\begin{aligned} S_{q1}(\omega) &=& \frac{1}{d(\omega)d(-\omega)}[(\frac{4\|a_{1s}\|^2\kappa_1}{\kappa_{1}^2+(\omega - \Delta_1)^2}+ \frac{4\|a_{1s}\|^2\kappa_1}{\kappa_{1}^2+(\omega +\Delta_1)^2})X(\omega)X(-\omega) + (\frac{4\|a_{2s}\|^2\kappa_2}{\kappa_{2}^2+(\omega -\Delta_2)^2} \nonumber \\ &+& \frac{4\|a_{2s}\|^2\kappa_2}{\kappa_{2}^2+(\omega +\Delta_2)^2})Y(\omega)Y(-\omega)+ (1+2n_T)(\gamma_{m1}^{2}Z(\omega)Z(-\omega) + \gamma_{m2}^{2}T(\omega)T(-\omega))];\end{aligned}$$ Here we have used the commutation relation for $\xi$. In equation 16, the first two term due to radiation pressure contribution of the optical modes in the two cavities, whereas the last term is due to the thermal noise contribution from the two cavities. In Fig. 2, we show the contour plot of the displacement spectrum $S_{q1}$ as a function of dimensionless effective detuning $\Delta_{1}/\nu$ for $\theta_1 = 0$, $\theta_2 = \pi/2$,$\Delta_{2}/\nu=0$, $\gamma_1/\nu$ $=$ $\gamma_2/\nu$ $=$ $0.01$, $g_1/\nu$ $=$ $g_2/\nu$ $=$ $0.6$, $\nu_{12}/\nu$ $=$ $0.3$, $\kappa_1/\nu$ $=$ $\kappa_2/\nu$ $=$ $0.1$ for different values of the photon numbers in the two cavities, (a): $\|a_{1s}\|^{2}=0.1$ and $\|a_{2s}\|^{2}=0.1$, (b):$\|a_{1s}\|^{2}=0.1$ and $\|a_{2s}\|^{2}=0.25$, (c): $\|a_{1s}\|^{2}=0.25$ and $\|a_{2s}\|^{2}=0.1$, (d):$\|a_{1s}\|^{2}=0.25$ and $\|a_{2s}\|^{2}=0.25$. The values of the color scheme is shown in Fig.5 in the appendix. Clearly, four modes are visible corresponding to the two mechanical and two optical modes. The coupling of the cavity field fluctuations and the mirror fluctuations leads to splitting of the normal mode of each cavity into two modes (NMS). The NMS is associated with the mixing between the fluctuation of the cavity field around the steady state and the fluctuations of the mirror mode around the mean field. The origin of the fluctuations of the cavity field is the beat of the pump photons with the photons scattered from the mirrors. We observe from the displacement spectra that NMS is observed only in plots (a) and (b) where the photon number in first cavity is $\|a_{1s}\|^{2}=0.1$. Increasing the photon number in the first cavity destroys the NMS. We also note by comparing plots (b) and (c) that decreasing the photon number in the second cavity to $\|a_{2s}\|^{2}=0.1$ does not restore the NMS. An important point to note is that in order to observe the NMS, the energy exchange between the modes should take place on a time scale faster than the decoherence of each mode. The parameter regime in which NMS appears implies cooling. For other values of the system parameters, the observation of NMS is prevented by the onset of the parametric instability. Therefore, a presence of NMS cannot be decoupled from the associated cooling which we discuss next where we calculate the effective temperature. Energy exchange between the modes of the two cavities depends on the two phases $\theta_{1}$ and $\theta_{2}$. In Fig.3 we show the contour plot of the displacement spectrum $S_{q1}(\omega)$ as a function of dimensionless effective detuning for $\theta_1 = 0$, $\theta_2 = \pi/4$ (left plot) and $\theta_1 = 0$, $\theta_2 = 3\pi/4$ (right plot). Clearly we observe energy exchange between the modes as we go from left to the right plot. Such energy exchange also implies that we can selectively cool one mirror at the expense of the other. In general it is known from basic physics that energy exchange between two mechanical oscillators takes place only for the anti-symmetric mode i.e when each mechanical oscillator is initially displaced from its position in opposite direction. In our case, such energy exchange between the two cavities can be achieved by tuning the phases $\theta_{1}$ and $\theta_{2}$. We now calculate the effective temperature of the mirror $M2$. In order to calculate the effective temperature, we need the spectrum of the momentum of the mirror $M2$ in fourier space. In a similar manner as above, we can calculate the momentum spectrum of the mirror $M2$, which is found as: $$\begin{aligned} S_{p1}(\omega) &=& \frac{1}{t_{15}(\omega)t_{15}(-\omega)}[(\frac{4\|a_{1s}\|^2\kappa_1}{\kappa_{1}^2+(\omega - \Delta_1)^2}+ \frac{4\|a_{1s}\|^2\kappa_1}{\kappa_{1}^2+(\omega +\Delta_1)^2})t_{16}(\omega)t_{16}(-\omega) + (\frac{4\|a_{2s}\|^2\kappa_2}{\kappa_{2}^2+(\omega -\Delta_2)^2} \nonumber \\ &+& \frac{4\|a_{2s}\|^2\kappa_2}{\kappa_{2}^2+(\omega +\Delta_2)^2})t_{17}(\omega)t_{17}(-\omega)+ (1+2n_T)t_{18}(\omega)t_{18}(-\omega)(\gamma_{m1}^{2} + \gamma_{m2}^{2})];\end{aligned}$$ For a driven system, effective temperature can be defined as [@huang]: $$T_{eff}=\dfrac{< \delta p^{2}>+<\delta q^{2}>}{2},$$ where the variances are calculated as, $$<\delta q^{2}>=\dfrac{1}{2 \pi}\int_{-\infty}^{\infty} S_{q}(\omega) d\omega$$ $$<\delta p^{2}>=\dfrac{1}{2 \pi}\int_{-\infty}^{\infty} S_{p}(\omega) d\omega$$ The equation for the effective temperature is one of our key results which tells how the temperature of one mirror depends on the various system parameters. Note that in general $\delta q^{2}$ $\neq$ $<\delta p^{2}>$. This implies that one does not have energy equipartition. This means that the steady state of the system is not, strictly speaking , a thermal equilibrium state. However, in order to get to the quantum ground state, both variances have to tend to $1/2$ and therefore energy equipartition has to be satisfied in the optimal regime close to the ground state. ![ A plot of the effective temperature $T_{eff}$ of the mirror $M_{2}$ as a function of mirror-mirror coupling strength $\nu_{12}$ for four different values of $\|a_{1s}\|^{2}$ and $\|a_{2s}\|^{2}$. (a): $\|a_{1s}\|^{2}=0.1$ and $\|a_{2s}\|^{2}=0.1$, (b):$\|a_{1s}\|^{2}=0.1$ and $\|a_{2s}\|^{2}=0.25$, (c): $\|a_{1s}\|^{2}=0.25$ and $\|a_{2s}\|^{2}=0.1$, (d):$\|a_{1s}\|^{2}=0.25$ and $\|a_{2s}\|^{2}=0.25$. Other parameters are same as in Fig.2. []{data-label="4"}](teff.eps) ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ×![ Plot of $T_{eff}$ of the mirror $M_{2}$ as a function of detuning $\Delta_{1}/\nu$ corresponding to the cases shown in Fig.2.Other parameters are same as in Fig.2. []{data-label="figure5"}](t1.eps "fig:") ![ Plot of $T_{eff}$ of the mirror $M_{2}$ as a function of detuning $\Delta_{1}/\nu$ corresponding to the cases shown in Fig.2.Other parameters are same as in Fig.2. []{data-label="figure5"}](t2.eps "fig:") ![ Plot of $T_{eff}$ of the mirror $M_{2}$ as a function of detuning $\Delta_{1}/\nu$ corresponding to the cases shown in Fig.2.Other parameters are same as in Fig.2. []{data-label="figure5"}](t3.eps "fig:") ![ Plot of $T_{eff}$ of the mirror $M_{2}$ as a function of detuning $\Delta_{1}/\nu$ corresponding to the cases shown in Fig.2.Other parameters are same as in Fig.2. []{data-label="figure5"}](t4.eps "fig:") ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- A plot of the effective temperature $T_{eff}$ of the mirror $M_{2}$ as a function of dimensionless mirror-mirror coupling strength $\nu_{12}/\nu$ for four different combinations of $\|a_{1s}\|^{2}$ and $\|a_{2s}\|^{2}$. (a): $\|a_{1s}\|^{2}=0.1$ and $\|a_{2s}\|^{2}=0.1$, (b):$\|a_{1s}\|^{2}=0.1$ and $\|a_{2s}\|^{2}=0.25$, (c): $\|a_{1s}\|^{2}=0.25$ and $\|a_{2s}\|^{2}=0.1$, (d):$\|a_{1s}\|^{2}=0.25$ and $\|a_{2s}\|^{2}=0.25$ is shown in Fig. 4. We clearly observe that minimum temperature is attained when both the cavities have low photon numbers. Increasing the photon number in any one of the cavity increases the temperature and the influence of $\|a_{1s}\|^{2}$ is more profound than $\|a_{2s}\|^{2}$. This is consistent with our earlier result on the NMS where we mentioned that the presence of the NMS also indicates cooling. Plots (a) and (b) in Fig.4 which show minimum temperature corresponds to plots (a) and (b) in Fig.2 which show NMS. Cooling of the mechanical mode of the mirror by the radiation pressure can be understood in thermodynamical sense. Radiation pressure couples the mirror to the optical cavity mode, which behaves as an effective additional reservoir for the mechanical oscillator. As a consequence, the effective temperature of the mirror mode will be intermediate between the initial thermal reservoir temperature and that of the optical reservoir, which is practically zero due to the condition that the mean number of photons is extremely small. Therefore one can approach the mechanical ground state of the mirror when the number of photons is small. In our case, the mechanical mode of mirror $M_{2}$ not only couples to optical mode of first cavity but also to the optical mode of the second cavity via the mechanical mode of the mirror $M_{3}$. This explains why significant mechanical cooling of the mechanical mode is obtained when number of photons in both the cavities is low. Fig.5 shows the $T_{eff}$ of the mirror $M_{2}$ as a function of detuning $\Delta_{1}/\nu$ corresponding to the cases shown in Fig.2. Interestingly we observe that the peaks in $T_{eff}$ corresponds exactly to the points in Fig.2 where the mirror displacement is maximum. These plots also illustrates energy exchange between the various modes as we vary mean photon numbers in the two cavities. In this system, the presence of the additional modes of the second cavity allows one to transfer energy from the mechanical mode of the mirror $M_{2}$ and the optical mode $a_{1}$ to the mechanical mode of the mirror $M_{3}$ and the optical mode $a_{2}$. This shows that both mechanical and optical cooling can be achieved by transferring energy from one cavity to the other. From the experimental point of view, the mirror’s position can be measured by means of a phase-sensitive detection of the cavity output, which is then fed back to the mirror by applying a force whose intensity is proportional to the time derivative of the output signal, and therefore to the mirror’s velocity. Conclusion ========== In this work, we have proposed a new technique to cool a harmonically oscillating mirror of an optical cavity mechanically coupled to another movable mirror of a second optical cavity. The system behaves as four coupled oscillators exchanging energy. Energy exchange can be coherently controlled by the phases of the opto-mechanical coupling strength. We find that a hybridization of the two oscillating mirrors with the fluctuations of the two driving optical fields occurs and leads to a splitting of the mechanical and optical fluctuation spectra. We also showed that normal mode splitting (NMS) leads to mechanical cooling. Significant mechanical cooling can be achieved by controlling the photon number in the two cavities. In addition, we demonstrate for the first time that by coupling two cavities, we can cool one cavity (both in the mechanical and optical sense) by transferring energy to the other. A continuous variable entanglement between the two mechanical modes could be used to improve the detection of weak classical forces in optomechanical devices as atomic force microscopes or gravitational wave detectors. Acknowledgements ================ One of the authors Tarun Kumar thanks the University Grants Commission, New Delhi for the Junior Research Fellowship. Appendix ======== ![Colour scheme for the contour plots []{data-label="6"}](colorbar_1.eps) where, $d(w) = M'-\dfrac{R P}{K'}$; $X (\omega) = [\dfrac{R}{K'}(\dfrac{A}{C'}-\dfrac{B L}{F' C}) + (\dfrac{N}{C'}-\dfrac{A B}{F' C})] D_1 + D_2$; $Y (\omega)= [\dfrac{R}{K'}(\dfrac{A B}{C' F}+\dfrac{L}{F'}) + (\dfrac{A}{F'}+\dfrac{N B}{C' F})] D_3 + D_4$; $Z (\omega)= \dfrac{R}{K'}(\dfrac{A}{C'}-\dfrac{L B}{F' C}) + \dfrac{N}{C'}-\dfrac{A B}{F' C} $; $T (\omega) = \dfrac{R}{K'}(\dfrac{A B}{C' F}+\dfrac{L}{F'}) + \dfrac{A}{F'}+\dfrac{B N}{C' F} $; $K' (\omega)= K-\dfrac{A}{C'}(\dfrac{B G}{F}-A)-\dfrac{L}{F'}(G+\dfrac{A B}{C})$; $C' (\omega)= C+\dfrac{B^2}{F}; F' = F+\dfrac{B^2}{C}$; $P (\omega)= -B-\dfrac{L}{F'}(\dfrac{E B}{F'}+A)+\dfrac{A}{C'}(E-\dfrac{A B}{F})$; $M' (\omega)= M+\dfrac{A}{F'}(\dfrac{E B}{C}+A)-\dfrac{N}{C'}(E-\dfrac{A B}{F})$; $R (\omega)= B+\dfrac{A}{F'}(\dfrac{A B}{C}+G)+\dfrac{N}{C'}(-A+\dfrac{B G}{F}) $; $C (\omega)= \gamma_{m1}- i \omega + \dfrac{4 D_1D_2\|a_{1s}\|^2\Delta_1}{(\kappa_1- i \omega)^2+\Delta_{1}^2}$; $E (\omega) =-\nu_1+ \dfrac{4 D_{1}^2\|a_{1s}\|^2\Delta_1}{(\kappa_1- i \omega)^2+\Delta_{1}^2}$; $F(\omega) = \gamma_{m2}- i \omega + \dfrac{4 D_3D_4\|a_{2s}\|^2\Delta_2}{(\kappa_2- i \omega)^2+\Delta_{2}^2}$ ; $G (\omega) =-\nu_2+ \dfrac{4 D_{3}^2\|a_{2s}\|^2\Delta_2}{(\kappa_2- i \omega)^2+\Delta_{2}^2}$; $M (\omega)= - i \omega + \dfrac{4 D_1D_2\|a_{1s}\|^2\Delta_1}{(\kappa_1- i \omega)^2+\Delta_{1}^2}$ ; $N(\omega) =\nu_1+ \dfrac{4 D_{2}^2\|a_{1s}\|^2 \Delta_1}{(\kappa_1- i \omega)^2+\Delta_{1}^2}$; $K (\omega)= - i \omega + \dfrac{4 D_3D_4\|a_{2s}\|^2\Delta_2}{(\kappa_2- i \omega)^2+\Delta_{2}^2}$ ; $L (\omega)=\nu_2+ \dfrac{4 D_{4}^2\|a_{2s}\|^2\Delta_2}{(\kappa_2- i \omega)^2+\Delta_{2}^2}$; $\Delta_1 = \omega_1-q_{1s}D_1+p_{1s}D_2$ ; $\Delta_2 = \omega_2-q_{2s}D_3+p_{2s}D_4$; $t_{15}(\omega)=t_{7}-\dfrac{t_{8}t_{12}}{t_{11}}$; $t_{16}(\omega)=t_{9}+\dfrac{t_{8}t_{13}}{t_{11}}$; $t_{17}(\omega)=t_{10}+\dfrac{t_{8}t_{14}}{t_{11}}$; $t_{18}(\omega)=\dfrac{t_{8}}{t_{11}}$; $t_{11}(\omega)=F-\dfrac{G t_{6}}{t_{4}}+\dfrac{A t_{3}}{t_{1}}$; $t_{13}(\omega)=-\dfrac{D_{2}B G}{M t_{4}}-\dfrac{A D_{2}}{t_{1}}$; $t_{12}(\omega)=\dfrac{G t_{5}}{t_{4}}-\dfrac{A t_{2}}{t_{1}}-B$; $t_{14}(\omega)=D_{3}+\dfrac{D_{4} G}{t_{4}}-\dfrac{A D_{2}}{t_{1}}$; $t_{7}(\omega)=C-\dfrac{E t_{2}}{t_{1}}+\dfrac{A t_{5}}{t_{4}}$; $t_{8}(\omega)=B+\dfrac{E t_{2}}{t_{1}}-\dfrac{A t_{6}}{t_{4}}$; $t_{9}(\omega)=D_{1}+\dfrac{D_{2} F}{t_{1}}+\dfrac{D_{2} A B}{M t_{4}}$; $t_{10}(\omega)=\dfrac{D_{4} F B}{t_{1} K}-\dfrac{A D_{4}}{t_{4}}$; $t_{1}(\omega)=M+\dfrac{B^{2}}{K}$; $t_{2}(\omega)=N+ \dfrac{A B}{K}$; $t_{3}(\omega)=A+ \dfrac{L B}{K}$; $t_{4}(\omega)=K+\dfrac{B^{2}}{N}$; $t_{5}(\omega)=A-\dfrac{B N}{M}$; $t_{6}(\omega)=L-\dfrac{A B}{M}$ [plain]{} T. W. Hansch and A. L. Schawlow, Optics Comm. [13]{}, 68 (1975). D. J. Wineland, R. E. Drullinger and F. L. Walls, Phy. Rev. Letts. [40]{}, 1639 (1978). S. Chu, L. Hollberg, J. E. Bjorkholm, A. Cable and A. Ashkin, Phy. Rev. Letts. [55]{}, 48 (1985). C M Caves, Phys. Rev. Letts, [45]{}, 75 1980, R. Loudon, Phys. Rev. Letts, [47]{}, 815 (1981). T. Corbitt and N. Mavalvala, J. Opt. B: Quantum Semi-class. Opt. [6]{}, S675 (2004). T. Corbitt et al., Phy. Rev. Letts. [98]{}, 150802 (2007). C. Höhberger-Metzger and K. Karrai, Nature [432]{}, 1002 (2004). S. Gigan et al., Nature [444]{}, 67 (2006). O. Arcizet et al., Nature [444]{}, 71 (2006). D. Kleckner and D. Bouwmeester, Nature [444]{}, 75 (2006). I. Favero et al., Appl. Phys. Lett. [90]{}, 104101 (2007). C. Regal, J. D. Teufel and K. Lehnert, Nature Physics, [4]{}, 555 (2008). T. Carmon et al., Phys. Rev. Letts. [94]{}, 223902 (2005). A. Schliesser et al., Phys. Rev. Letts. [97]{}, 243905 (2006). J. D. Thompson et al., Nature [452]{}, 72 (2008). F. Brennecke et al., Science [322]{}, 235 (2008). K. W. Murch et al., Nature Physics [4]{}, 561 (2008). A. B Bhattacherjee, Phys. Rev. A [80]{}, 043607,(2009). A. B. Bhattacherjee, J. Phys. B [43]{}, 205301,(2010). A. Schliesser, O. Arcizet, R. Riviére, G. Anetsberger, and T. J. Kippenberg, Nature Physics [5]{},509 (2009), Y.-S. Park and H. Wang, , Nature Physics [5]{}, 489 (2009), S. Gröblacher, J. B. Hertzberg, M. R. Vanner, G. D. Cole, S. Gigan, K. C. Schwab, and M. Aspelmeyer, Nature Physics [5]{}, 485 (2009). V.B. Braginsky and F. Khalili, Quantum Measurements (Cambridge University Press, 1992). F. Marquardt, J. G. E. Harris and S. Girvin, Appl. Phys. Lett., [96]{}, 103901 (2006). J. M. Dobrindt, et. al., Phys. Rev. Letts ,[101]{}, 263602 (2008). S. Huang and G. S. Agarwal, Phys. Rev. A, [79]{}, 013821 (2009). T. Kumar, A. Bhattacherjee and ManMohan, Phys. Rev. A, [81]{}, 013835 (2010). S. Mancini, D. Vitali, V. Giovannetti and P. Tombesi, Eur. Phys. J. D, [22]{}, 417 (2003); S. Mancini, V. Giovannetti, D. Vitali and P. Tombesi, Phys. Rev. Letts ,[88]{}, 120401 (2002). M. Ludwig , K. Hammerer and F. Marquardt, Phys. Rev. A, [82]{}, 012333 (2010). M. Bhattacharya and P. Meystre, Phys. Rev. Letts ,[99]{}, 073601 (2007).	{ "pile_set_name": "ArXiv" }
--- abstract: 'Planets and minor bodies such as asteroids, Kuiper-belt objects and comets are integral components of a planetary system. Interactions among them leave clues about the formation process of a planetary system. The signature of such interactions is most prominent through observations of its debris disk at millimeter wavelengths where emission is dominated by the population of large grains that stay close to their parent bodies. Here we present ALMA 1.3 mm observations of HD 95086, a young early-type star that hosts a directly imaged giant planet b and a massive debris disk with both asteroid- and Kuiper-belt analogs. The location of the Kuiper-belt analog is resolved for the first time. The system can be depicted as a broad ($\Delta R/R \sim$0.84), inclined (30$\pm$3) ring with millimeter emission peaked at 200$\pm$6 au from the star. The 1.3 mm disk emission is consistent with a broad disk with sharp boundaries from 106$\pm$6 to 320$\pm$20 au with a surface density distribution described by a power law with an index of –0.5$\pm$0.2. Our deep ALMA map also reveals a bright source located near the edge of the ring, whose brightness at 1.3 mm and potential spectral energy distribution are consistent with it being a luminous star-forming galaxy at high redshift. We set constraints on the orbital properties of planet b assuming co-planarity with the observed disk.' author: - 'Kate Y. L. Su, Meredith A. Macgregor, Mark Booth, David J. Wilner, Kevin Flaherty, A. Meredith Hughes, Neil M. Phillips, Renu Malhotra, Antonio S. Hales$^{8,9}$, Sarah Morrison, Steve Ertel, Brenda C. Matthews, William R. F. Dent$^{8}$, Simon Casassus$^{12,13}$' title: 'ALMA 1.3 Millimeter Map of the HD 95086 System' --- Introduction ============ Debris disks were discovered by [IRAS]{} [@aumann84] as infrared excess emission from dust orbiting stars and sustained by collisions of leftover planetesimals and cometary activity. They often have a structure analogous to that of minor body belts in the solar system, with asteroid- or Kuiper-belt components. The majority of the known debris disks are massive, Kuiper-belt analogs not only because the collisional evolution proceeds more slowly at large orbital distances but also stars are faint in the far-infrared, making positive identifications of excess from cold debris much easier. It is interesting to note that the first Kuiper Belt Objects in our own solar system were not discovered until 1992 [@jewitt93], eight years later than the [IRAS ]{} discovery. Planets, minor bodies, and leftover planetesimals all form as a consequence of agglomeration processes that occur within the protoplanetary disk. Interactions between them during the formation and subsequent evolution leave signs in the disk that can be used to study the current state and past history of a planetary system. Therefore, these faint dusty disks are excellent tools to understand the outer zones of exoplanetary systems including our own. With sensitive infrared surveys, hundreds of debris disks are known [@matthews14], providing a rich resource to study planetary system evolution and architecture. Although thousands of exoplanets and candidates have been discovered through radial velocity and transit measurements, this breakthrough is currently biased toward the inner zones of systems, not sensitive to planets like Jupiter and Saturn beyond 5 au. Recent improvements in high contrast imaging have enabled us to find planets out at the same stellocentric distance scales as the debris disks. Fomalhaut [@kalas08], HR 8799 [@marois08; @marois10], $\beta$ Pic [@lagrange09], HD 95086 [@rameau13], HD 106906 [@bailey14], and 51 Eri [@machintosh15] are prominent examples of such systems known to host both debris disks and directly imaged planets. From the observed dust temperatures derived from disk spectral energy distributions (SEDs) of $\sim$200 debris disks, @ballering13 report a weak trend that the inner edge of the cold planetesimal zone appears to depend on the luminosity/temperature of the star, indicating a signpost for planetary migration and/or shepherding. However, disk extents estimated from SEDs are degenerate. Any inferred radii depend strongly on the assumed composition and the particle size distribution. Due to this degeneracy, it is very difficult to translate SED measurements into physical sizes directly. Even when there are resolved images available (mostly in the far-infrared), the exact location of the parent bodies is still uncertain due to the effect of non-gravitational forces (radiation and drag) on small grains. The true parent-body distribution in debris disks can be provided by resolved submillimeter/millimeter images which probe large (mm-size) grains that stay close to their parent bodies. Disk morphologies suggestive of influences from unseen planets, such as resonance clumps [@wyatt03] and/or apo-center glow [@pan16], are also best observed at submillimeter/millimeter wavelengths (e.g., @ertel12 [@lohne17]). Existing ALMA data on debris disks show a large variety of Kuiper-belt analogs: some systems have very narrow rings of parent bodies (e.g., Fomalhaut, @boley12 [@macgregor17], and $\epsilon$ Eri, @booth17), and some have either multiple rings (HD 107146, @ricci15a) or broad disks (HR 8799, @booth16; $\tau$ Ceti, @macgregor16b; 61 Vir, @marino17). The parent body distributions are therefore giving insights to the possible overall structure of the planetary systems. HD 95086 is a young (17$\pm$4 Myr, @meshkat13) A8 star that possesses a large infrared excess, indicative of a massive debris disk [@chen12], and a $\sim$5 $M_J$ planet at the projected distance of $\sim$56 AU [@rameau13; @rameau16]. Compared to the Hipparcos catalog, the Gaia DR1 catalog gives a slightly closer distance, 83.8$\pm$1.9 pc [@gaia16], which we adopt throughout the paper. Its disk was marginally resolved by [Herschel ]{} and found to be inclined at $\sim$25$^{\circ}$ from face-on [@moor13]. Analysis of its detailed infrared SED and re-analysis of the resolved images suggest that the debris structure around HD 95086 is very similar to that of HR 8799: a warm ($\sim$170 K) belt, a cold ($\sim$60 K) disk, and an extended disk halo (up to $\sim$800 AU) [@su15]. Modeling the disk surface brightness distribution at 70 and 160 $\mu$m suggests that the extended emission seen in the far-infrared is largely from the small grains produced by frequent collisions due to dynamical stirring of planetesimals and launched from the system by stellar radiation in the form of a disk halo. Therefore, the inclination derived from the [Herschel]{} images might not be a true representative of the planetesimal disk or be subject to a large error. It is then crucial to measure the intrinsic distribution of the planetesimal population, as traced by millimeter emission from large grains, in order to properly characterize the possible perturbers, HD 95086 b and any unseen planet(s) interior to the cold disk. Here we present the first millimeter observations of the HD 95086 system, obtained by the Atacama Large Millimeter/submillimeter Array (ALMA). Our observations reveal the location of the cold Kuiper-belt analog for the first time. The paper is organized as follows. Details about the observations and general data reduction are given in Section 2. In Section 3, we first present the dust continuum map of the system which can be described as an inclined ring plus a bright point source near the outer edge of the ring. We then determine the properties of the disk (flux and geometry) and those of the bright source (flux and position) using both visibilities and imaging model approaches. In Section 4, we revise the disk SED based on the new observations, discuss the ring’s width and possible asymmetry, the likely nature of the bright source, and obtain new constraints on HD 95086 b. Conclusions are given in Section 5. [lccccrccc]{} & & & & & &\ 2015-01-28 & X1beb & 38 & 15.1 – 348.5 & 1.27 & 83.7 & 45.86 & +02:25 & J1107\ 2015-04-04 & Xba2 & 39 & 15.1 – 327.8 & 1.02 & 73.8 & 45.86 & –00:38 & Callisto\ 2015-04-05 & X267e & 39 & 15.1 – 327.8 & 1.28 & 77.2 & 45.86 & –00:09 & Ganymede\ 2015-04-05 & X2a9e & 39 & 15.1 – 327.8 & 1.27 & 77.3 & 45.86 & +01:27 & J1107\ 2015-04-05 & X2e6d & 39 & 15.1 – 327.8 & 1.25 & 80.7 & 45.86 & +02:55 & Titan\ 2015-04-06 & X14f2 & 36 & 15.1 – 327.8 & 1.23 & 76.5 & 45.86 & +01:47 & J1107\ \ & & & & & &\ 2015-04-10 & X1412 & 35 & 15.3 – 348.5 & 2.14 & 101.0 & 45.36 & +00:14 & Ganymede\ 2015-04-10 & X1d34 & 35 & 15.3 – 348.5 & 2.38 & 108.9 & 45.36 & +02:16 & J1107\ 2015-04-14 & Xbb6 & 36 & 15.3 – 348.5 & 3.65 & 144.6 & 45.36 & –01:18 & Ganymede\ 2015-04-23 & X1462 & 39 & 15.1 – 348.5 & 1.88 & 103.3 & 45.36 & +03:12 & Titan\ 2015-05-01 & X883 & 37 & 15.1 – 348.5 & 1.95 & 100.7 & 45.36 & +00:55 & Ganymede\ 2015-05-02 & Xd15 & 37 & 15.1 – 348.5 & 1.18 & 85.0 & 45.36 & +00:15 & Ganymede\ \[alma\_images\] Observations ============ We observed HD 95086 with ALMA in Band 6 (1.3 mm) under two projects: \#2013.1.00773.S, PI: Su (referred to as data set A) and \#2013.1.00612.S, PI: Booth (referred to as data set B). The observations consist of 12 single pointing block executions centered at HD 95086 (phase center RA: 10:57:02.91 Dec: –68:40:02.27 (J2000)). The majority of the observations were obtained in April/May 2015, while one was done in January 2015. The proper motion of the star (pmra = –41.11$\pm$0.03 mas/yr and pmdec = 12.91$\pm$0.03 mas/yr) gives 11 mas offset for the three month time span, i.e., there is no significant pointing difference in these observations. Table \[tab1\_obs\] lists the details about these observations including dates, block id, number of antennae used, projected baselines, weather conditions, on-source integration time and flux calibrators. The correlator set-up was designed to optimize the continuum sensitivity, but also covered the $^{12}$CO J=2-1 transition at 230.538 GHz with 3840 channels over a bandwidth of 1.875 GHz. The set-up was slightly different between the two projects. The four basebands were centered at 215, 217, 230 and 232.5 GHz for data set A, but at 231.87, 232.55, 245, 247 GHz for data set B. The raw data were processed by the ALMA Regional Centers using the CASA package (ver.4.2.2 for data set A and ver. 4.3.1 for data set B). Nearby quasars and solar system objects (Callisto and Ganymede) were used for flux calibration, resulting in an absolute flux uncertainty $\lesssim$10 % (the Technical Handbook for cycle 2). The total on-source integration time is 4.58 hours for data set A, and 4.54 hours for data set B. No CO detection was reported in the pipeline reduced product. Details regarding the CO gas in the system will be reported in another publication (Booth et al. in prep.). Results and Analysis ==================== Continuum Emission ------------------ We generated the calibrated measurement sets using the scripts provided by the ALMA project for each of the data sets. We then split the observations into different fields (pointing) and spectral windows by binning the time sampling to 30 s and averaging the spectral channels with a width of 128 channels. These averaged, binned $uv$ visibilities were then exported to FITS format for further analysis using the MIRIAD software [@sault95]. Visibilities were then inverted with natural weighting, deconvolved, and restored to generate a final synthesized map using the standard procedures in MIRIAD. For the data set A, the synthesized continuum image is shown in Figure \[alma\_images\]a with a synthesized beam of 128$\times$103 and a position angle (P.A., measured from North toward the East) of 75.1, and a rms of 8.7 $\mu$Jy beam$^{-1}$. For the data set B, the image is shown in Figure \[alma\_images\]b with a synthesized beam of 120$\times$104 and a P.A. of 79.4, and a rms of 11.0 $\mu$Jy beam$^{-1}$. In both images, a ring-like structure is clearly seen with a very bright point-like source offset from the star $\sim$3 away at a P.A. of 293 (–67). Since the quality (rms and beam) of both data sets was similar, we then combined both data sets and generated a slightly deeper continuum map (shown in Figure \[alma\_images\]c). The combined continuum map has a synthesized beam of 122$\times$103 and a P.A. of 77.4, and a rms of 7.5 $\mu$Jy beam$^{-1}$. The ring’s circumference is detected at signal-to-noise (S/N) $\gtrsim$15 $\sigma$ per beam, and is slightly inclined from face-on. We estimate the pointing accuracy of the data, $\sim$resolution/signal-to-noise, to be 013 since the main ring is detected at S/N$\gtrsim$10. We adopt two approaches to explore the best-fit parameters for the HD 95086 system: (1) visibilities fitting and (2) image plane fitting. In both approaches, we assume the millimeter emission can be described by an optically and geometrically thin (no scale height) model plus a point source offset from the center. We explore two simple axi-symmetric models to describe the disk: (1) a two-boundary disk confined in a radial span of $R_{in}$ and $R_{out}$ with a surface density power law of $\Sigma(r) \propto r^{p}$ where $r$ is the stellocentric distance, and (2) a Gaussian ring defined by the peak ($R_p$) and the width (FWHM) of the ring ($R_w$). For the millimeter emission of the disk (i.e., dominated by large grains), we expect the dust temperatures follow $T_d(r) = 278.3 L_{\ast}^{0.25} r^{-0.5}$ where $L_{\ast}$ is the stellar luminosity in units of the solar luminosity (6 $L_{\sun}$ for HD 95086 using the new distance) and $r$ is in au. The disk has a total flux, $F_{tot}$, at 1.3 mm, and its mid-plane is assumed to incline by an angle of $i$ from face-on (i.e., $i=0$) with the major axis along a position angle (P.A.). There are three parameters describing the bright source: the total flux ($F_{pt}$) and offset from the star ($\Delta x$ and $\Delta y$). We discuss the results in the following subsections for both approaches, and synthesize the final best-fit model parameters in Section 3.4. \[mcmc\_model\] [llccccccccccc]{} & & & & & & & &\ & & Value & $\pm$1$\sigma$ & & Value & $\pm$1$\sigma$ & & Value & $\pm$1$\sigma$ & & Value & $\pm$1$\sigma$\ \ $R_{in}$ \[au\] & inner belt radius & 107 & +6 –5 && 110 & +3 –4 &&& &&&\ $R_{out}$ \[au\] & outer belt radius & 327 & +6 –7 && 328 & +7 –6 &&& &&&\ $p$ & surface density index & –0.48 & +0.34 –0.38&& –0.42 & +0.13 –0.12&&& &&&\ $R_{p}$ \[au\] & peak radius && &&& && 204 & +7 –7 && 208 & +4 –3\ $R_{w}$ \[au\] & width (FWHM) && &&& && 176 & +6 –6 && 179 & +6 –6\ $F_{tot}$ \[mJy\] & total belt flux density & 2.87 & +0.10 –0.11&& 2.89 & +0.08 –0.08&& 2.91 & +0.10 –0.18&& 3.07 & +0.09 –0.09\ $F_{pt}$ \[mJy\] & flux of Pt.$^2$ & 0.88 & +0.05 –0.05 && 0.92 & +0.01 –0.01 && 0.87 & +0.06 –0.06 && 0.92 & +0.01 –0.01\ $\Delta$x \[\] & RA offset of Pt. & –3.06 & +0.04 –0.04 & & –3.07 & +0.01 –0.01 && –3.05 & +0.05 –0.05 && –3.07 & +0.01 –0.01\ $\Delta$y \[\] & Dec offset of Pt. & 0.85 & +0.05 –0.05 && 0.85 & +0.01 –0.01 && 0.85 & +0.05 –0.05 && 0.85 & +0.01 –0.01\ $i$ \[\] & inclination & 36 & +3 –2 && 35 & +2 –2 && 36 & +2 –2 && 34 & +2 –2\ P.A. \[\] & position angle & 98 & +3 –3 && 98 & +3 –3 && 98 & +3 –4 && 96 & +4 –3\ Visibilities Modeling Approach ------------------------------ We model the visibilities for both data sets simultaneously. To minimize the free parameters, we assume no offset between the center of the disk and the star. Therefore, there are a total of eight/nine free parameters to describe the system in both axi-symmetric disk models: two/three parameters for the disk density distribution ($R_{in}$, $R_{out}$ and $p$ for the two-boundary disk, or $R_{p}$ and $R_{w}$ for the Gaussian ring), two parameters for the disk viewing geometry ($i$ and P.A.), the total flux of the disk ($F_{tot}$), and three parameters for the point source ($\Delta x$, $\Delta y$ and the total flux $F_{pt}$). We determine the best-fit values for these free parameters independently by adopting the MCMC approach outlined in @macgregor13. For all parameters, we assume uniform priors and require that the model be physically plausible (flux greater than zero and the outer radius larger than the inner one). The best-fit parameters and their $\pm$1$\sigma$ uncertainties are given in Table \[tab2\_mcmc\]. For each set of the best-fit parameters, we generated a high resolution model image and transformed it to the visibility domain according to the observation. We then constructed the residual map by subtracting the model from the data in the visibility domain and imaging the residual using the same procedures in MIRIAD. The residual maps are shown in Figure \[mcmc\_model\]. Overall, the residuals are within $\pm$3$\sigma$ for the main disk. The subtraction of the bright source is not perfect, and creates an over-subtraction at the center of the bright source, and positive residuals in the area around it, suggesting the source might be extended . In all residual maps, there appears to be another faint (S/N$\sim$9) source $\sim$25 south of the bright one. The two axi-symmetric models yield very similar parameters in terms of the disk flux, viewing geometry and point source parameters. However, the residual in the main disk is slightly smaller in the Gaussian ring model. Although the residuals in the main disk tend to be more negative in the east side of the disk, unfortunately the bright source is along the west side of the major-axis, making it difficult to assess any asymmetric structure present in the disk (more detail is given in Section 4.2). Imaging Plane Modeling Approach ------------------------------- Given the good S/N detection of the main disk, we also try to derive the best-fit parameters for the two models by fitting in the image plane. Details about this approach can be found in @booth16. We use the combined synthesized map for the MCMC search. For experiments, two more free parameters are included in this part of the fitting. For both models, the center of the ring is not fixed at the star position. Although a small offset[^1] is preferred for both models, these values are within one pixel of the reconstructed maps (02 per pixel) and within 2 $\sigma$ of the pointing accuracy ($\sigma\sim$ 013), therefore not significant. The final best-fit parameters and the residual maps are also given in Table \[tab2\_mcmc\] and Figure \[mcmc\_model\]. Overall, the best-fit parameters agree with the ones derived from the visibilities method within the uncertainties, except for the total fluxes of the ring and the point source where the derived flux using the imaging approach is consistently larger. We also note that the estimated uncertainties are also smaller using the image plane approach. This is because the MCMC uncertainty depends strongly on the weightings of the data. The noise within the beam is highly correlated, and the image deconvolution (the “CLEAN” procedure) treats noise non-linearly, both resulting in smaller uncertainties for the MCMC image fitting that might not be statistically robust. A factor equivalent to the square root of the beam size in pixels is included to mitigate the correlated noise, but this is only an approximation as it assumes a Gaussian beam whereas the dirty beam has some low level, non-Gaussian structure that this factor cannot account for. Fitting in the image domain is computationally faster, and can achieve the same result in terms of geometric parameters for high S/N data; however, we caution against relying on the robustness of the uncertainties using imaging plane fitting. [clrr]{} $F_{pt}$ \[mJy\] & flux density & 0.81 $\pm$0.03 & 0.10$\pm$0.02\ $\Delta$x \[\] & RA offset & –3.08 $\pm$0.04 & –2.80$\pm$0.03\ $\Delta$y \[\] & Dec offset & 0.83 $\pm$0.05 & –1.61$\pm$0.04 \[pts\_ring\] [llccccccccccc]{} & && & & & & &\ & & Value & $\pm$1$\sigma$ & & Value & $\pm$1$\sigma$ & & Value & $\pm$1$\sigma$ & & Value & $\pm$1$\sigma$\ \ $R_{in}$ \[au\] & inner belt radius & 105 & +5 –5 & & 106 & +5 –4 & && & &&\ $R_{out}$ \[au\] & outer belt radius & 331 & +4 –6 & & 312 & +7 –8 & && & &&\ $p$ & surface density index & -0.5 & +0.3 –0.3 & & –0.5 & +0.2 –0.2 & && & &&\ $R_{p}$ \[au\] & peak radius && & && & & 199 & +6 –6 & & 200 & +4 –4\ $R_{w}$ \[au\] & width (FWHM) && & && & & 169 & +5 –6 & & 167 & +7 –7\ $F_{tot}$ \[mJy\] & total belt flux density & 2.77 & +0.19 –0.07 & & 2.57 & +0.09 –0.08 & & 2.72 & +0.12 –0.08 & & 2.74 & +0.09 –0.10\ $\Delta$x$_c$ \[\] & RA offset of ring & 0.15 & +0.04 –0.04 & & 0.09 & +0.03 –0.03 & & 0.15 & +0.05 –0.04 & & 0.12 & +0.04 –0.04\ $\Delta$y$_c$ \[\] & Dec offset of ring & 0.05 & +0.04 –0.04 & & 0.05 & +0.03 –0.03& & 0.06 & +0.04 –0.04 & & 0.08 & +0.03 –0.03\ $i$ \[\] & inclination & 31 & +3 –3 & & 28 & +3 –3 & & 31 & +2 –3 & & 30 & +3 –3\ P.A. \[\] & position angle & 98 & +4 –3 & & 97 & +6 –6 & & 98 & +4 –4 & & 95 & +5 –5 Best-Fit Disk Parameters by Minimizing Other Contamination ---------------------------------------------------------- Since ALMA is sampling the sky with many different baselines (i.e., spatial scales) through interferometry, we can better assess the properties of the bright source by generating a map with only the long baseline data ($>$80 $k\lambda$ = 31) where any extended structure with sizes $>$31 (i.e., the disk emission) is filtered out. The long baseline map is shown in Figure \[pts\_ring\]a. Within 35 radius from the star, there are two sources (detected above 8 $\sigma$) that appear in the long baseline map. The measured FWHM of the faint source is 079$\times$064, the same as the synthesized beam in the long baseline map, while the FWHM of the bright source is $\sim$8% broader, 085$\times$070. To evaluate whether the threshold defining the long baseline data affects the FWHM of the bright source, we also generated the long baseline maps with different thresholds between 60 to 80 $k\lambda$. The FWHM of the bright source is consistently broader than the synthesized beam by 8%. Furthermore, we generated individual, long baseline maps per data set to see if there exists a flux difference between the two data sets for the bright source. Taken at face value, the bright source is about 5% brighter in data set B. Given the typical absolute flux uncertainty (10%) in the ALMA data, the flux difference is not significant. The properties (offset and flux density) of the two point sources are determined using ’uvmodelfit’ in CASA, and given in Table \[tab3\_twops\]. Using the long baseline data, the flux of the bright source is 9% lower than the derived flux by fitting the disk and point source simultaneously. \[residual\_disk\] We then generated a “disk-only” map by subtracting the best-fit point sources in the visibility domain. The “disk-only” map is shown in Figure \[pts\_ring\]b. The subtraction of the bright source is far from perfect. There is still significant flux ($\gtrsim$25 $\sigma$) near the east side of the bright source, which could be part of the disk structure or the bright source has a non-symmetric, extended shape. If the bright source is a dusty galaxy in nature (see Section 4.3), it is very likely to have an irregular shape, making it challenging to separate it from the disk without high angular resolution observations. The subtraction of the faint source is better. The emission near the position of the faint source is more smooth, but it does appear that the disk flux extends toward the faint source, explaining the residual in Figure \[mcmc\_model\]. Similarly, the observed resolution prevents further assessment. To evaluate the impact of the two sources on the derived disk parameters, we searched for the best-fit disk parameters using the visibility approach by fixing the properties of the two point sources. The best-fit disk parameters are basically the same as the ones without fixing the two point sources (Section 3.2). The residual maps are shown in Figures \[residual\_disk\]a and \[residual\_disk\]b. Compared to Figures \[mcmc\_model\]a and \[mcmc\_model\]b, the fits with the fixed point source properties have no over subtraction at the position of the bright source, but the residual around the bright source is higher. We performed similar searches using the image plane approach by masking out the pixels that have fluxes $>$0.1 mJy beam$^{-1}$ (the region around the bright source and the center of the faint one). To explore whether allowing a slight offset between the star and the disk center can improve the results, we also allow offsets in the MCMC parameters. The best-fit disk parameters are given in Table \[tab4\_disk\]. Interestingly, the disk size parameters are slightly smaller than the ones derived in Section 3.3, but still within uncertainties. The residual maps are shown in Figures \[residual\_disk\]c and \[residual\_disk\]d. The derived disk fluxes is also smaller, reducing the over-subtraction in the east side of the disk. The presence of the bright source is undoubtedly affecting the derived disk parameters. Unfortunately, there is no easy way to mask out the contribution of the bright source in the visibility domain since the bright source contributes to all baselines. We tried to minimize the impact of the bright source by fitting it as a Gaussian profile in visibilities, allowing for some sort of extension. The results are also given in Table \[tab4\_disk\]. Similar to the image plane approach, the disk size parameters are also slightly smaller than the ones derived in Section 3.2. The Gaussian parameters for the bright source are basically the same as the synthesized beam, but with much less flux compared to the fits derived from the long baseline data (Table \[tab3\_twops\]). The residual maps (not shown) are neither better than the ones in Figures \[residual\_disk\]a and \[residual\_disk\]b nor than those in Figures \[mcmc\_model\]a and \[mcmc\_model\]b as a result. We synthesized the various fitting results as follows. The disk has a total flux density of 2.79$\pm$0.1 mJy at 1.3 mm, and is inclined by 30$\pm$3 from face-on with a P.A. of 97$\pm$3. It is interesting to note that the best-fit models (Table \[tab4\_disk\]) all prefer to have the ring center east of the star by $\sim$01. Since the star is not detected in the ALMA data, the “translated” pointing accuracy cannot confirm such an offset. The disk is very broad ($\Delta R/R\sim$0.84) in the millimeter continuum, and its width is resolved by $\sim$1.7 beam widths. As a result, we cannot determine the exact disk density distribution nor the offset between the star and disk center. For the two-boundary model, the disk can be described as having sharp boundaries at $R_{in}=$106$\pm$5 au and $R_{out}=$320$\pm$10 au with a surface density power index of –0.5$\pm$0.3. For the Gaussian ring, the disk is peaked at $R_p=$200$\pm$6 au with a width of $R_w=$168$\pm$7 au. The reduced chi-squared ($\chi^2$) is 1.20 for the two-boundary model, but 1.15 for the Gaussian ring model with the total number of visibilities (574358) in the fitting; similar $\chi^2$ numbers are also found using the image plane approach. Based on the $\chi^2$, the Gaussian ring model gives a slightly better fit; however, the number of free parameters is different (7 vs. 8), and some of the parameters are correlated. Discussion ========== Revised 3-Component SED Model ----------------------------- \[sed\] @su15 examined the resolved disk images of HD 95086 from [Herschel]{} and argued, by a detailed SED analysis, that the system is likely to possess three debris structures. The three debris components are very similar to the ones of the HR 8799 system – (1) an inner warm emission representing the dust in an asteroid-belt analog, (2) an outer cold emission representing the dust in a Kuiper-belt analog, and (3) an extended disk halo surrounding the aforementioned two components and composed of small grains. Since only the extended disk halo is resolved in the far infrared, the exact boundaries of the different components are very uncertain and not well constrained by the SED model. The presence of an asteroid-belt analog is only constrained by the excess emission detected in the [Spitzer]{} IRS spectrum and unresolved MIPS 24 [${\rm \mu m}$ ]{}photometry; therefore, its location is set by the observed dust temperature ($\sim$175 K, i.e., 7–10 au). Given the warm temperature and large distance to the system, this asteroid-belt component is not expected to be detected nor resolved by the ALMA observation. Since we now have the measured size for the cold disk ($R_{in}$=106 au and $R_{out}$=320 au compared to the old SED 63–189 au value), the 3-component SED model needs revision. Furthermore, a re-reduction of the archive APEX/LABOCA 870 [${\rm \mu m}$ ]{}data on HD 95086 using the techniques described in @phillips11 finds a total flux of 19.4$\pm$11.0 mJy, much lower than the one published by @nilsson10. The peak emission in the 870 [${\rm \mu m}$ ]{}map is offset by 13 (i.e., 2/3 of the beam diameter) from the expected star position; therefore, the quoted flux is estimated as an unresolved source at the position of the star. Due to the large offset, we consider it as a non-detection and use 33 mJy as the 3 $\sigma$ upper limit. We adopted this value for further SED analysis. Using similar approaches and grain parameters as in @su15 (minimum and maximum grain sizes of $\sim$1.8 [${\rm \mu m}$ ]{}and 1000 [${\rm \mu m}$]{}, and a particle size distribution in a power law index of $-$3.5), a geometrically thin, constant surface density disk with a radial span of 106-320 au provides a good fit to the ALMA 1.3 mm flux and maximum allowable fluxes in the far infrared (gray squares in Figure \[sed\]). This cold disk SED model also agrees with the observed 7 mm flux (not shown in Figure \[sed\]) obtained by @ricci15b within the uncertainty. Compared to the cold component model presented in @su15, this revised planetesimal disk model contributes much more flux shortward of 60 [${\rm \mu m}$]{}, and is the dominant component (compared to the disk halo) at 20–30 [${\rm \mu m}$]{}. With a much larger cold planetesimal disk component, the inner radius of the disk halo is more distant from the star (from old $\sim$190 au to $\sim$300 au), and contributes much less flux at mid- and far-infrared wavelengths using the same grain parameters as in @su15. Although we now have resolved the cold disk at millimeter wavelengths, revealing the placement of the large grain population, there are still a wide range of parameters that are not constrained in the SED models, especially for the grain parameters in the disk halo. The SED models shown in Figure \[sed\] are not unique. Future resolved images of the various components at crucial mid-infrared wavelengths will shed light on this. Ring Width and Azimuthal Asymmetry ---------------------------------- We computed the azimuthally averaged radial profile for the disk using the synthesized image data without the two point sources (the right panel of Figure 3). Assuming the disk is inclined by 30 with the major axis along P.A. of 97, we first created a series of elliptical rings with a width of 2 pixels (04) centered at the star, and computed the average value of all pixels that fall in each ring. Since the pixels are highly correlated within the area of each synthesized beam, the noise in each ring can be approximated with the standard deviation in the ring divided by the number of beams in that ring. The background noise per ring is computed in a similar fashion. The total error in the average flux measurement per ring, therefore, is the nominal error and the background noise added in quadrature. The resultant disk surface brightness profile is shown in the top panel of Figure \[diskprofiles\]. For reference, the profile using the original data is also shown, and the contamination from the bright source obviously results in extra flux in the radii at 2–4 from the star. The point-source subtracted ring profile is centered at $r\sim$23 from the star with symmetric profiles inside and outside the peak within uncertainties. The millimeter emission of the disk is surprisingly broad ($\Delta R/R\sim$0.84). To assess the degree of the asymmetry in the millimeter disk emission, we also computed the radially averaged surface brightness profile along the disk circumference. To minimize the contamination of the bright source, we picked a radial span of 14–26 from the star, and computed the average disk brightness within an incremental angle of 27 azimuthally. Similarly, the uncertainty includes the standard deviation and background noise in each of the wedges. The azimuthal profile is shown in the bottom panel of Figure \[diskprofiles\]. The azimuthal profile agrees within 1 $\sigma$ to the average disk surface brightness (horizontal grey line in Figure \[diskprofiles\]), except at P.A. of $\sim$300, the direction toward the bright source (although the difference is still within 2 $\sigma$). Given the contamination by the bright source, plus the modest resolution of the ring ($\sim$6 beam widths in circumference and $\lesssim$2 beam widths in width), the apparent asymmetry is not significant. \[diskprofiles\] Possible Nature of the Bright Source ------------------------------------ As demonstrated in Section 3.4, the bright source near the edge of the disk, at 32 (a projected distance of 268 au) from the star, is slightly more extended than the synthesized beam, and roughly along the major axis of the disk (3 off). Although the disk emission can be traced up to $\sim$320 au from the star, the peak millimeter emission is within 200 au, i.e., this bright structure is quite far away from the main location of the colliding planetesimals. We explore various possibilities for the nature of the bright source either physically associated with the HD 95086 system or due to chance alignment of a background source. ### Debris Phase of a Circumplanetary Disk? It is challenging to form planets at large orbital distances through the usual route of km-sized planetesimal merger. However, there are multiple ways to bypass this hurdle like pebble accretion in conjunction with planet scattering (e.g., @lambrechts12 [@kenyon15; @bromley16]). Therefore, it may be possible to have a newly formed planet at $\sim$270 au from this 17 Myr old star. Given the fact that there is no sign of this bright source in the deep $K/L'$ data (J. Rameau priv. comm.) and in the mid-infrared photometry of the system [@moor13; @su15], it is unlikely that the bright source is the direct detection of a newly formed planet[^2]. Assuming the bright source is at the distance of 83.8 pc, the bulk of the 1.3 mm flux suggests a bolometric luminosity of 2$\times10^{-3}$ to 7$\times10^{-5} L_{\sun}$ assuming it has a temperature of 30–100 K, which translates to a radius of 0.7–1.4 au for the optically thick emitting area. It is interesting to note that the Hill radius for a 10 $M_{\oplus}$ planet at 270 au around 1.6 $M_{\sun}$ star is $\sim$5 au. Therefore, the mm flux of the bright source could come from the dust emission of a circumplanetary disk (CPD) whose typical size is expected to be one third of the Hill radius [@martin11; @zhu15]. The CPD is expected to be gas-rich around a planet in formation, like a scaled-down version of a protoplanetary disk around a young star. With a total flux of 0.81 mJy at 1.3 mm, the estimated dust mass is $\sim$0.2–0.4 $M_{\oplus}$ assuming a typical dust opacity, $\kappa_{1.3mm}$= 2.3 cm$^2$g$^{-1}$ [@beckwith90] and dust temperatures of 30–60 K. We do not detect any CO gas emission from the bright source. The noise level in the integrated CO (2-1) line flux is 1.5$\times10^{-23}$ W m$^{-2}$ assuming a velocity dispersion of 4.6 km s$^{-1}$ (twice the Keplerian velocity at 270 au) (Booth et al. in prep.). The CO gas mass for the hypothesized CPD is less than 2.3$\times10^{-6} M_{\oplus}$ (1 $\sigma$, details see Booth et al.in prep.), suggesting an extremely low gas-to-dust mass ratio. If the dust emission did come from the CPD of a newly formed planet, the CPD might be in the “debris” phase as moons/satellites are being formed. However, the mass fraction between the hypothesized CPD and newly formed planet is uncomfortably high ($\sim$10$^{-2}$ for a Neptune-size planet) in comparison to the typical mass fraction of $10^{-4}$ between the satellites and the giant planets in the solar system, making this “debris CPD” hypothesis unlikely. ### A Dust Clump due to A Giant Impact? Alternatively, a recent giant impact in the disk can create a bright, concentrated region in the disk [@telesco05]. Depending on the impact velocity, the disk morphology could remain in the clump-dominated phase that lasts for a few orbital periods after a giant impact [@jackson14]. Given the system’s young age, observing such a large impact at 270 au is not impossible. Next, we estimate whether the brightness/mass in the clump is consistent with such a scenario. Using the parameters derived in Section 3.4, the bright source contributes $\sim$25% of the total disk flux at 1.3 mm, which is significantly larger than the clump in the $\beta$ Pic disk ($<$ 4% of the disk flux). At 270 au, the orbital velocities are so low that most collisions between large objects would be mergers[^3] and not produce much debris (with masses $\sim $ a few % of the impactors, @jackson14). To produce the amount of dust observed in the bright source, objects with masses $\gtrsim$10–20 $M_{\oplus}$ (Neptune mass) are required. Impacts involving such large objects are likely to be very explosive, i.e., the resultant clump is expected to smear and spread very rapidly after the impact (a few orbital periods). The fact that the bright source is relatively compact ($\sim$8% broader than the synthesized beam in the long baseline data) disfavors the origin of a giant impact. ### A Dust Clump due to Planetesimals trapped by an Unseen Planet? In addition, a concentrated dust clump can also be created by the intense collision among the planetesimals trapped in the resonance with an unseen planet [@wyatt03], as one of the proposed origins for the dust clump in the $\beta$ Pic disk. The dust clump in the $\beta$ Pic disk is also found to be very bright in CO gas emission, probably released by the icy planetesimals [@matra17]. Therefore, we might also expect to detect a significant amount of CO gas associated with the bright source in HD 95086 if all the planetary systems have a similar composition. The upper limit on the integrated CO (2-1) line flux is $\sim$100 times fainter than the integrated line flux of CO (2-1) line in the $\beta$ Pic disk after scaling by the distance difference, while the dust flux in the clump is much brighter in the HD 95086 disk. Given these comparisons, it seems unlikely the bright source has a similar nature as the clump in the $\beta$ Pic disk. ### Alignment of a Background Galaxy An alternative explanation is that the source is a background galaxy. In fact, @su15 suspected that the integrated submillimeter flux of the system is likely contaminated by background galaxies due to the excess emission detected at [Herschel]{}/SPIRE bands and APEX 870 [${\rm \mu m}$ ]{}compared to the disk SED model. Using the parameters of the Schechter function in @carniani15, the probability of a galaxy with an 1.3 mm flux of $>$0.81 mJy within 4 of the star is $\sim$0.5%, but increases to 5% and 14% chance within the FWHM and 10% of the primary beam, respectively. For the fainter source, the probability of having a source with an 1.3 mm flux of $>$0.1 mJy within 4 of the star is 11.4%. Assuming the two sources are physically not related, the chance of both within 4 of the star is 0.06%. The contamination from multiple background point sources is also seen around $\epsilon$ Eri [@chavez16]. Therefore, it is possible that the sources in HD 95086 are related background galaxies, and the 0.06% probability is then a lower limit. These values are a statistical assessment given an ensemble of observations, and the application to one single observation is not one-to-one correspondence. Based on these probabilities, the faint source is very likely to be a background galaxy; the bright source could be a background galaxy, a rare case but not impossible. In general, a background galaxy might have a steep dust spectrum, like $\nu^{3.5}$, and the debris disk is likely to be shallower, like $\nu^{2.6}$ [@macgregor16a; @holland17]. Therefore, a background galaxy is likely to be brighter in the data set B than in A due to the frequency difference if the absolute flux calibration in both data sets is consistent. As estimated in Section 4.1, the flux difference of the bright source between the two data sets is within the uncertainty of absolute flux calibration, i.e., not significant. Alternatively, we can compare the spectral indexes between the bright source and the disk. Spectral index maps were generated with the data sets combined and separately using all baselines and long baselines only. Figure \[fig\_spindex\] shows the spectral index map of the combined data set. The bright source has a spectral index of 3.0$\sim$4.0 derived from the combined data. Due to the signal-to-noise in the extended emission, the spectral index across the whole disk varies, however, the index of the disk is generally shallower than the one of the bright source. \[fig\_spindex\] A steeper spectral index does not necessarily mean the bright source is indeed a background galaxy because an impact produced clump may also have a steep particle size distribution, resulting in a high spectral index. Assuming this is indeed the case, the spectral slope suggests that the clump should have a total flux of a few hundred mJy at 200 $\mu$m from extrapolating the measured flux of 0.81 mJy at 1.3 mm. The flux of such a clump at 70–100 $\mu$m range would have been even brighter, i.e., comparable to the total disk flux in the far infrared. Given the measured disk SED (Figure \[sed\]), it would be very difficult to have such a component co-exist with other components (planetesimal disk and disk halo), corroborating our early assessment. As a sanity check, we can also construct the SED of the bright source using the measured 1.3 mm flux and the revised 3-component disk SED presented in Section 4.2. By comparing the photometric measurements and the model disk SED, the “excess” emission, presumably from the bright source, is 31.4$\pm$18.9 mJy, 30.9$\pm$10.6 mJy, 24.8$\pm$10.4 and 10.1$\pm$11.0 mJy at 250, 350, 500 and 870 [${\rm \mu m}$]{}, respectively (the quoted errors include 10% uncertainty from the SED model). With the 1.3 mm flux measured by ALMA, the SED of the bright source is shown in Figure \[galaxysed\], and the SED is consistent with the one of a dusty star forming galaxy at $z$=2 [@casey14]. The angular size of the bright source (Section 3.4) is in the range of angular sizes for luminous infrared and submillimeter galaxies at $z\sim$2 measured in the radio [@gurvits99; @rujopakarn16]. It seems very plausible that this bright source is due to the chance alignment of a background galaxy. \[galaxysed\] Constraints for the HD 95086 b from the Disk Perspective -------------------------------------------------------- HD 95086 b was discovered by @rameau13 at an angular separation of 060–063 from the star [@rameau16]. At a distance of 83.8 pc, this translates to a projected distance of 50.3–52.3 au. Assuming the planet b and the disk are co-planar, the inclination of the disk (30$\pm$3) implies that planet b has a stellocentric distance of 56–63 au ($\sim$semimajor distance if the planet b is on a circular orbit). The bright source makes it difficult to assess the asymmetry in the disk, and no “significant” asymmetry is present in the millimeter disk emission. Alternatively, we can also put some constraint on the eccentricity of the shepherding planet orbit, presumably the planet b, by determining the offset of the ring and the star (i.e., the offset is $\sim a e$ where $a$ is the semi-major distance and $e$ is the eccentricity). As discussed earlier, we did not detect a significant offset between the ring center and the star. The expected pointing accuracy, $\sim$ resolution/signal-to-noise, is 013 since the main ring is detected at S/N$\gtrsim$10. The non-detection of an offset suggests $e<0.17$ for the shepherding planet (presumably planet b) with a semi-major distance of 63 au. The most recent orbital parameters for HD 95086 b are from @rameau16 where a small angular movement is detected using the data obtained by Gemini/GPI between 2013 and 2016: a semi-major distnce of 61.7$^{+20.7}_{-8.4}$ au, an inclination of 27$^{+10}_{-13}$ and an eccentricity less than 0.2. With the revised distance of 83.8 pc to HD 95086, we revise the semi-major distance of HD 95086 b to 57.2$^{+19.2}_{-7.8}$ au. These orbital parameters are all consistent with the ones derived from the disk geometry (assumed co-planar). The mass of the planet is estimated to be 4.4$\pm$0.8 $M_{Jup}$ [@derosa16]. The 5-$\sigma$ detection limits from VLT/NaCo and Gemini/GPI observations suggest that our current high contrast capability is not sensitive to planet masses less than 1.5 $M_{Jup}$ in the 60–800 au region from the star [@rameau16]. Assuming the mass of the star is 1.6$\pm$0.16 $M_{\sun}$, the mass ratio between the planet b and the star, $\mu \equiv M_p/M_{\ast}$, is 1.95–3.45$\times10^{-3}$. For this mass ratio, the timescale to clear the planet’s chaotic zone is $\lesssim 1$ Myr [@morrison15], much shorter than the estimated $\sim$17 Myr age of the system. We adopt the numerically derived formula from @morrison15 to compute the size of the planet’s chaotic zone. Assuming the planet b is at a circular orbit with a semi-major axis of $a_p$, the interior chaotic zone width is $\Delta a_{int}=1.2 \mu^{0.28} a_p$ and the exterior chaotic zone width is $\Delta a_{ext}=1.7 \mu^{0.31} a_p$. With the range of $\mu$ and $a_p$ (56–63 au), the width of the exterior chaotic zone is 14–19 au, suggesting the outer boundary of the chaotic zone from the planet b is 70–82 au. If the inner edge of the disk is at 106 au as derived from the two boundary model in Section 3.4, an eccentricity of $e\sim$0.29 is needed to extend its chaotic zone to the inner edge of the disk if planet b is the shepherding planet and coplanar with the disk. This eccentricity is marginally consistent ($<$2$\sigma$) with no significant offset detected between the ring center and the star. If we relax the assumption that the disk and the planet are co-planar, HD 95086 b’s chaotic zone can reach 98 au assuming a semi-major distance of 76 au (the maximum allowable range from high contrast imaging, @rameau16). A small eccentricity ($e\sim$0.08) can extend the planet’s influence to 106 au (the inner boundary of the disk). Therefore, the planet b can be the shepherding planet to maintain the inner edge of the Kuiper-belt analog. In summary, the values listed above are all within the allowable ranges for the system. If the system hosts an additional planet outside the orbit of HD 95086 b that is shepherding the inner edge of the cold disk, it would have to be less than 1.5 $M_{Jup}$ to have eluded detection [@rameau16]. From the dynamical stability criteria of @gladman93 and @morrison16, the estimated separation between planet b and the disk is sufficient for another planet below this mass threshold to reside there while remaining long-term dynamically stable with respect to planet b if both possess low orbital eccentricities. However, the youth of this system ($\sim$17 Myr) and long dynamical timescales at these large orbital distances place a lower limit on planets that could have cleared debris from their chaotic zones over the system’s lifetime. From the clearing timescales estimated in @morrison15, a putative low eccentricity, coplanar outer planet orbiting in the region beyond planet b’s orbit and interior to the cold disk, would have to be $\gtrsim$0.2 $M_{Jup}$ ($\sim$4 Neptune masses) to have cleared debris from that region over the system’s lifetime. In summary, the shepherding planet between the planet b and the inner edge of the cold disk would have a mass of 0.2–1.5 $M_{Jup}$ with the assumption of coplanarity and low eccentricity. Conclusion ========== We obtained an ALMA 1.3 mm image of HD 95086, a young ($\sim$17 Myr) star hosting a directly imaged planet and a debris disk composed of dust generated in massive asteroid- and Kuiper-belt analogs. The high angular resolution (a beam of 11) and sensitivity (rms of 7.5 $\mu$Jy beam$^{-1}$) provided by ALMA enable us to resolve the Kuiper-belt analog for the first time. The sensitive ALMA millimeter image reveals an inclined ring centered at the star and a bright source near the edge of the ring along the major axis of the ring. Our observations also covered the the $^{12}$CO J=2-1 transition at 230.538 GHz, and no CO emission above 3 $\sigma$ per beam was found in the pipeline produced CO channel maps in the region of the ring and the bright source. To access the properties of the bright source, we also generated the continuum map using data with baselines longer than 60–80 $k\lambda$ where an extended structure like the planetesimal disk is filtered out. The long baseline data reveal two sources within 35 of the star (a much fainter source south of the bright one). The FWHM of the sources is consistent with being a point source although the bright one is slightly broader (8%) than the synthesized beam. We determined the best fit parameters (total fluxes and positions) of the two sources in the long baseline data using point source fitting in the visibility domain. The faint source has a 1.3 mm flux of 0.1 mJy (S/N $\gtrsim$9 in the long baseline map), and is most likely a background galaxy, similar to other faint sources in the long baseline map. The bright source is detected at high S/N with a total 1.3 mm flux of 0.81 mJy and located –308 E and 083 N of the star. We explored the possible nature of the bright source including (1) a debris phase of the circumplanetary disk, (2) a dust clump produced by a giant impact, (3) a dust concentration due to planetesimals trapped by an unseen planet, and (4) a background dusty galaxy. The source’s brightness in dust continuum and non-detection of the CO emission suggest that it is unlikely to have resulted from a structure physically assocated with the system (the first three scenarios), and more likely due to chance alignment of a background source. We further constructed the SED of the bright source using the 1.3 mm flux and the “excess” emission by comparing the unresolved photometry and the expected disk emission, and found it is consistent with the expected SED from a $z$=2 dusty galaxy. The slight extension of the bright source and the steeper spectral index compared to the spectral index of the disk are both consistent with the bright source being a luminous, high redshift galaxy. We used the MCMC approach to determine the best fit parameters of the disk. We assumed the Kuiper-belt analog can be described by simple, parametric models. We explored two axi-symmetric, geometric thin models for the disk surface density profiles: (1) a two-boundary, power-law disk with sharp inner and outer boundaries, and (2) a Gaussian ring. We further assumed that the millimeter emission comes from large grains whose temperatures follow a $r^{-0.5}$ power law. The best-fit parameters and associated uncertainties are derived by fitting the visibilities and image plane data. We found that fittings in the visibilities and image plane domain give consistent results within $\pm$1 $\sigma$ uncertainties in terms of the geometric parameters (disk extent, position and inclination angles and the offset of the bright source); however, the derived fluxes are consistently larger when using the fits in the image plane. Although within the uncertainties, the flux of the bright source is also brighter than the one derived from the long baseline data, implying a difficulty of separating the disk from the bright source. We then synthesized the disk parameters by (1) fixing the point source parameters in the visibilities fit, (2) masking out the pixels affected by the bright source in the imaging fit, and (3) fitting the bright source as a Gaussian profile. The final parameters for the Kuiper-belt analog are all very similar. The major axis of the disk is along P.A. of 97$\pm$3 with the mid-plane inclined by 30$\pm$3 from face-on. The width of the disk is very broad and resolved by $\lesssim$2 beam widths. The disk density profile is consistent with either (1) a broad, Gaussian ring peaked at 200$\pm$6 au with a FWHM width of 168$\pm$7 au or (2) an $r^{-0.5\pm0.3}$ power-law profile with an inner radius of 106$\pm$5 au and outer radius of 320$\pm$10 au. Although the residual maps (data – model) are very similar between the two models, the Gaussian ring model gives a slightly better reduced $\chi^2$. In all residual maps, the east side of the disk has more negative residuals compared to the opposite side, suggesting an apparent disk asymmetry. However, the residual in the west side of the disk is also contaminated by the imperfect subtraction of the bright source. Judging from the azimuthal profile along the disk circumference, the apparent asymmetry is not significant. We also explored whether allowing an offset between the ring and the star would produce a better residual map. Although a small offset (within 015, $\sim$one eighth of the synthesized beam) is preferred in $\chi^2$ statistics, it produces no significant difference in the residual map. Based on the estimated pointing accuracy of 013, the non-detection of an offset suggests that the orbit of the shepherding planet has an eccentricity $<$0.17 if it has a semi-major axis of 63 au. Given the observed projected separation between HD 95086 b and the star, the semi-major axis of the planet b orbit is 56–63 au if the planet and the disk share the same orbital plane and the planet is on a circular orbit. The estimated major axis is consistent with the apparent motion of the planet detected within three years. However, in the co-planar case for a planet in a circular orbit, the expected chaotic zone of the planet b (maximum of 82 au), does not reach the inner boundary of the disk ($\sim$106 au). If the planet b is the shepherding planet to maintain the inner edge of the cold disk, an eccentricity $\gtrsim$0.29 is needed to extend its influence. Such a planet would create an offset of 02 between the star and the ring center, marginally consistent with the observed data. It is also possible that the planet b is not the shepherding planet for the cold disk given the large separation between its chaotic zone and the inner disk edge. An additional unseen low-mass planet (0.2–1.5 $M_{Jup}$) on a circular orbit can also occupy the dust-free zone between the planet b and the cold disk and have eluded detection. Alternatively, relaxing the co-planarity assumption, a larger semimajor axis of planet b (76.4 au, still within the allowable range) and a small eccentricity ($\sim$0.08) would extend its chaotic zone to $\sim$106 au. We thank Alan Jackson for the discussion on the evolution of impact produced clumps, and Benjamin Weiner for the discussion on the properties of background galaxies. We are grateful to George Rieke for his thorough proofreading of the manuscript. We also thank Julien Rameau providing additional inforamtion about the background source, and Scott Kenyon and Ya-Lin Wu for their comments. M.B. is grateful to Bruce Sibthorpe and Andrés Jordán for the help during proposal preparation. K.Y.L.S. acknowledges the partial support from the NASA grant \# NNX15AI86G. M.A.M. acknowledges support from the National Science Foundation under Award No. 1701406. M.B. acknowledges support from the Deutsche Forschungsgemeinschaft (DFG) through project Kr 2164/15-1. R.M. acknowledges research support from NASA (grant \# NNX14AG93G) and NSF (grant \#AST-1312498). This paper makes use of the following ALMA data: ADS/JAO.ALMA \#2013.1.00773.S and \#2013.1.00612.S. ALMA is a partnership of ESO (representing its member states), NSF (USA) and NINS (Japan), together with NRC (Canada) and NSC and ASIAA (Taiwan) and KASI (Republic of Korea), in cooperation with the Republic of Chile. The Joint ALMA Observatory is operated by ESO, AUI/NRAO and NAOJ. The National Radio Astronomy Observatory is a facility of the National Science Foundation operated under cooperative agreement by Associated Universities, Inc. This work has made use of data from the European Space Agency (ESA) mission [Gaia]{} (<https://www.cosmos.esa.int/gaia>), processed by the [Gaia]{} Data Processing and Analysis Consortium (DPAC, <https://www.cosmos.esa.int/web/gaia/dpac/consortium>). Funding for the DPAC has been provided by national institutions, in particular the institutions participating in the [Gaia]{} Multilateral Agreement. Facilities: . Aumann, H. H., Beichman, C. A., Gillett, F. C., et al. 1984, , 278, L23 Beckwith, S. V. W., Sargent, A. I., Chini, R. S., & Guesten, R. 1990, , 99, 924 Bailey, V., Meshkat, T., Reiter, M., et al. 2014, , 780, L4 Ballering, N. P., Rieke, G. H., Su, K. Y. L., & Montiel, E. 2013, , 775, 55 Boley, A. C., Payne, M. J., Corder, S., et al. 2012, , 750, LL21 Booth, M., Jord[á]{}n, A., Casassus, S., et al. 2016, , 460, L10 Booth, M., Dent, W. R. F., Jord[á]{}n, A., et al. 2017, , 469, 3200 Bromley, B. C., & Kenyon, S. J. 2016, , 826, 64 Gaia Collaboration, Prusti, T., de Bruijne, J. H. J., et al. 2016, , 595, A1 Carniani, S., Maiolino, R., De Zotti, G., et al. 2015, , 584, A78 Casey, C. M., Narayanan, D., & Cooray, A. 2014, , 541, 45 Chavez-Dagostino, M., Bertone, E., Cruz-Saenz de Miera, F., et al. 2016, , 462, 2285 Chen, C. H., Pecaut, M., Mamajek, E. E., Su, K. Y. L., & Bitner, M. 2012, , 756, 133 Dent, W. R. F., Wyatt, M. C., Roberge, A., et al. 2014, Science, 343, 1490 De Rosa, R. J., Rameau, J., Patience, J., et al. 2016, , 824, 121 Eisner, J. A. 2015, , 803, L4 Ertel, S., Wolf, S., & Rodmann, J. 2012, , 544, A61 Gladman, B. 1993, [Icarus]{}, 106, 247 Gurvits, L. I., Kellermann, K. I., & Frey, S. 1999, , 342, 378 Holland, W. S., Matthews, B. C., Kennedy, G. M., et al. 2017, arXiv:1706.01218 Jackson, A. P., Wyatt, M. C., Bonsor, A., & Veras, D. 2014, , 440, 3757 Jewitt, D., & Luu, J. 1993, , 362, 730 Kalas, P., Graham, J. R., Chiang, E., et al. 2008, Science, 322, 1345 Kenyon, S. J., & Bromley, B. C. 2015, , 806, 42 Lagrange, A.-M., Gratadour, D., Chauvin, G., et al. 2009, , 493, L21 Lambrechts, M., & Johansen, A. 2012, , 544, A32 Leinhardt, Z. M., & Stewart, S. T. 2012, , 745, 79 L[ö]{}hne, T., Krivov, A. V., Kirchschlager, F., Sende, J. A., & Wolf, S. 2017, arXiv:1704.08085 Rameau, J., Nielsen, E. L., De Rosa, R. J., et al. 2016, , 822, L29 MacGregor, M. A., Wilner, D. J., Rosenfeld, K. A., et al. 2013, , 762, L21 MacGregor, M. A., Wilner, D. J., Chandler, C., et al. 2016, , 823, 79 MacGregor, M. A., Lawler, S. M., Wilner, D. J., et al. 2016, , 828, 113 MacGregor, M. A., Matr[à]{}, L., Kalas, P., et al. 2017, , 842, 8 Macintosh, B., Graham, J. R., Barman, T., et al. 2015, Science, 350, 64 Marino, S., Wyatt, M. C., Kennedy, G. M., et al. 2017, , 469, 3518 Marois, C., Macintosh, B., Barman, T., et al. 2008, Science, 322, 1348 Marois, C., Zuckerman, B., Konopacky, Q. M., Macintosh, B., & Barman, T. 2010, , 468, 1080 Martin, R. G., & Lubow, S. H. 2011, , 413, 1447 Matthews, B. C., Krivov, A. V., Wyatt, M. C., Bryden, G., & Eiroa, C. 2014, Protostars and Planets VI, 521 Matr[à]{}, L., Dent, W. R. F., Wyatt, M. C., et al. 2017, , 464, 1415 Meshkat, T., Bailey, V., Rameau, J., et al. 2013, , 775, L40 Mo[ó]{}r, A., [Á]{}brah[á]{}m, P., K[ó]{}sp[á]{}l, [Á]{}., et al. 2013, , 775, L51 Morrison, S., & Malhotra, R. 2015, , 799, 41 Morrison, S. J., & Kratter, K. M. 2016, , 823, 118 Nilsson, R., Liseau, R., Brandeker, A., et al. 2010, , 518, A40 Pan, M., Nesvold, E. R., & Kuchner, M. J. 2016, , 832, 81 Phillips, N. M. 2011, Ph.D. Thesis, Pope, A., Chary, R.-R., Alexander, D. M., et al. 2008, , 675, 1171-1193 Ricci, L., Carpenter, J. M., Fu, B., et al. 2015a, , 798, 124 Ricci, L., Maddison, S. T., Wilner, D., et al. 2015b, , 813, 138 Rameau, J., Chauvin, G., Lagrange, A.-M., et al. 2013a, , 772, L15 Rujopakarn, W., Dunlop, J. S., Rieke, G. H., et al. 2016, , 833, 12 Sault, R. J., Teuben, P. J., & Wright, M. C. H. 1995, Astronomical Data Analysis Software and Systems IV, 77, 433 Su, K. Y. L., Rieke, G. H., Stapelfeldt, K. R., et al. 2009, , 705, 314 Su, K. Y. L., Rieke, G. H., Malhotra, R., et al. 2013, , 763, 118 Su, K. Y. L., & Rieke, G. H. 2014, IAU Symposium, 299, 318 Su, K. Y. L., Morrison, S., Malhotra, R., et al. 2015, , 799, 146 Telesco, C. M., Fisher, R. S., Wyatt, M. C., et al. 2005, , 433, 133 Wyatt, M. C. 2003, , 598, 1321 Zhu, Z. 2015, , 799, 16 [^1]: 016 and 006 for the two boundary model, and 021 and 008 for the Gaussian ring [^2]: The typical temperature for a protoplanet is expected to be a few 100 to a few 1000 K [@zhu15; @eisner15]. [^3]: The typical impact velocity is roughly the orbital velocity, which is $\sim$2.3 $km s^{-1}$ at 270 au around HD 95086. Given the escape velocity of 25 $km s^{-1}$ for a Neptune-like planet, a typical impact between two Neptune-like objects at 270 au belongs to the merging collision outcome based on the work by @leinhardt12.	{ "pile_set_name": "ArXiv" }
--- abstract: \| SuperCDMS is the next phase of the Cryogenic Dark Matter Search experiment, which measures both phonon and charge signals generated by particle recoils within a germanium target mass. Charge signals are employed both in the definition of a fiducial volume and in the rejection of electron recoil background events. Alternatively, phonons generated by the charge carriers can also be used for the same two goals. This paper describes preliminary efforts to observe and quantify these contributions to the phonon signal and then use them to reject background events. A simple analysis using only one pulse shape parameter shows bulk electron recoil vs. bulk nuclear recoil discrimination to the level of 1:10$^3$ (limited by the statistics of the data), with little degradation in discrimination ability down to at least 7 keV recoil energy. Such phonon-only discrimination can provide a useful cross-check to the standard discrimination methods, and it also points towards the potential of a device optimized for a phonon-only measurement. PACS numbers: 29.40.Wk 63.20.kd author: - 'S.A. Hertel$^1$ and M. Pyle$^2$ for the SuperCDMS Collaboration' date: '07.08.2011' title: Phonon Pulse Shape Discrimination in SuperCDMS Soudan --- Phonons in a SuperCDMS Detector =============================== SuperCDMS (Cryogenic Dark Matter Search) Soudan is in preparation for installation at the Soudan Underground Laboratory in northern Minnesota. The target mass consists of $\sim$15 large roughly cylindrical (3.81 cm radius, 2.54 cm thickness, $\sim$600 g) ultra-pure germanium crystals. These target masses are instrumented on the flat surfaces with interleaved arrays of phonon-sensing Quasiparticle-trap-assisted Electrothermal-feedback Transition-edge-sensors (QETs) and charge-sensing electrodes.[@luke1994; @brink2006] The measurement of both the phonon energy and charge energy of each event enables the ratio of these two energies to be used as a powerful event-by-event discriminator between low-charge-production nuclear recoils (a possible dark matter signature) and high-charge-production electron recoils (the vast majority of background events). The interleaving of biased electrodes (typically +2 V on one side, -2 V on the opposite side) with QET arrays (essentially at 0 V) defines a $\sim$0.5 V/cm drift field in the bulk of the target mass. Within $\sim$1 mm of the top and bottom surfaces, however, the field created by the interleaved structures is both much stronger than the bulk ($\sim$20 V/cm), and also largely parallel to the surface. This field shape prevents carriers created in this region from propagating to the opposite side electrodes, and the resulting side-asymmetric charge signal tags such events as near-surface events[@pyle2009] (near-surface electron recoils would otherwise form a dangerous background[@bailey2009]). Much of the essential information content of the charge measurement is duplicated in the phonon measurement, through the production of secondary phonon populations as charges propagate through the crystal and then enter aluminum structures at the surface. These different phonon populations are produced with different initial energy distributions. In Germanium, high energy phonons exhibit high isotopic scattering rates ($\tau_{i}^{-1}~=~[36.7\times10^{-42}]\nu^4$ for Ge)[@tamura1985] meaning that they have short mean free paths and propagate diffusively ($\ell~=~[5.4~km/s]\tau_{i}~=~[1.5\times10^{44}~m/s^4]\nu^{-4}$ at long wavelengths). Only 6.1$\%$ of the top and bottom crystal surfaces is covered with phonon-absorbing aluminum (largely in the form of QETs), meaning that, per surface interaction, a phonon is far more likely to reflect off a polished Ge surface than be absorbed in a QET. A phonon’s timescale of surface absorption then is roughly proportional to its rate of surface interactions, and is therefor highly dependent on mean free path. Phonons with a mean free path on order of the detector size (ie, which propagate ballistically) are absorbed with the maximally slow decay constant of 755 $\mu$s, and higher energy phonons are absorbed with significantly faster decay constants. Phonons anharmonically decay in energy ($\tau_{a}^{-1} = [1.61\times10^{-55}]\nu^5$)[@msall1997], meaning that mean free path increases with time. This energy-dependent diffusion and decay process is typically termed “quasidiffusion”. Here we summarize the relevant characteristics of the major phonon populations in a SuperCDMS detector, emphasizing where they begin in their quasidiffusive evolution: - Primary phonons produced by the recoil event itself are initially highly energetic ($\nu > $1 THz, ie $\ell~<~\sim$ 100$\mu$m). If the event occurs far from a sensor surface, the diffusive behavior slows the arrival at the surface and lengthens the eventual mean free path, slowing the absorption rate at that surface. If the event occurs near a sensor surface, there is little delay in arrival, and the comparatively short mean free path increases a the rate of absorption at that surface. - Neganov-Luke phonons[@neganov1978; @luke1988] are created as charge carriers are drifted by the electric field. Such phonons are created very soon after the event time (the charge drift time is $\sim$ 1 $\mu$s). In the low field of the detector bulk, Neganov-Luke phonons are emitted with low energies and correspondingly ballistic free paths, whereas the majority of Neganov-Luke phonons are produced in the strong near-surface fields at higher frequencies ($\sim$300 GHz and $\sim$700 GHz for electrons and holes) with correspondingly short mean free paths ($\sim$1.8 cm and $\sim$600 $\mu$m)[@wang2010; @leman2010] and fast absorption rates. - Relaxation phonons are emitted when charge carriers cross from the Ge into the superconducting Al surface structures, converting their Ge gap energy (0.75 eV) into Al quasiparticles (of energy 2$\Delta_{Al}$ = 200 $\mu$eV)[@court2008] and a population of initially high energy phonons. In the presence of superconducting Al, such phonons experience rapid downconversion (at time scales of $\sim$ 1 $\mu$s)[@kaplan1976; @kozorezov2007], meaning that relaxation phonons are ballistic at the time scales considered here. These three phonon populations combine together to produce the observed phonon pulses and their associated rising and falling timing characteristics. We will from now on ignore relaxation phonons, since they are initially ballistic (their absorption time constant is a slow $\sim$755 $\mu$s, and a minimal contribution to pulse shape). A recoil occurring in the crystal bulk will produce a roughly side-symmetric signal of near-surface Neganov-Luke phonons, which will arrive at both the top and bottom surfaces at fast times ($\sim$1 $<~t~<~\sim$5 $\mu$s) and are absorbed rapidly by these surfaces. Primary phonons will reach the top and bottom surfaces more gradually, as quasi-diffusive propagation allows ( $\sim$1 $<$ t $< \sim$15 $\mu$s). These primary phonons will have shorter mean free paths (and will be more quickly absorbed) at the closer surface, and will have longer mean free paths (and will be more slowly absorbed) at the further surface. When an event occurs within the near-surface electric field region, it will be distinguished by a boosted population of Neganov-Luke phonons at one surface (with fast absorption times) and a lack of any distinct Neganov-Luke sharp rising edge in the opposite side pulse. Quantifying Phonon Pulse Characteristics in SuperCDMS Detectors =============================================================== The understanding of phonon populations given in the previous section predicts certain shape characteristics as a function of both proximity to the absorbing surface and quantity of charge carriers (and resulting Neganov-Luke production). Looking at recent calibration data (Figure 1), we do see such behaviors, confirming our understanding. ![Example phonon pulses from calibration runs with a SuperCDMS Soudan detector. The four side 1 QET channels are summed and shown in blue (dark grey); the side 2 channels are summed and shown in red (light grey). All eight events shown are of similar total phonon energy ($\sim$50 keV), and are located at different positions along the central axis of the cylindrical crystal (as determined by relative weighting among the eight phonon channels). Surface 1 and surface 2 events were selected using an additional a requirement that the charge signals from the two sides be strongly asymmetric.\[fig:pulses\]](pulses.png) To make use of these phonon pulse shape variations as event type discriminators(distinguishing electron recoil and near-surface recoils from nuclear recoils), we first reduce the raw traces to a small number of simple quantities, and then combine these quantities in a way that accommodates the large position-dependent variation. First, each of the raw traces from the detector’s eight QET channels were fit using the following functional form: $$\begin{aligned} P(t) = A_{fast} \bigg(1-e^{\frac{-(t-T_{fast})}{R_{fast} }} \bigg) e^{\frac{-(t-T_{fast})}{F_{fast} }} + A_{slow}\bigg(1-e^{\frac{-(t-T_{slow})}{R_{slow} }} \bigg) e^{\frac{-(t-T_{slow})}{F_{slow} }} \nonumber\end{aligned}$$ where the total pulse is treated as the sum of a “fast” pulse and a “slow” pulse, with rising and falling time constants $R_{fast}$, $R_{slow}$, $F_{fast}$, and $F_{slow}$, and start time offsets $T_{fast}$ and $T_{slow}$. The fast pulse can be thought of as the contribution of highly-diffusive quickly-absorbing phonons, while the slow signal can be thought of as the contribution of more ballistic slowly-absorbing phonons, but the quasi-diffusive nature means that in reality there is not a simple two-category distinction. The falling time constant $F_{slow}$ of the slow pulse was observed to be nearly identical for every event, and was set to 755 $\mu$s, the observed rate of absorption of the late-time uniform bath of low energy phonons.[@kevin2011] Each event’s resulting 7 fit parameters for each of 8 phonon channels were further reduced by summing the fits for each side and then finding key points (10$\%$, 20$\%$, 30$\%$, etc.) along the rising and falling edges of these side-summed fits. The partition of energy between the eight channels also contains significant discrimination information (in addition to position information). The amplitude of each channel’s pulse was obtained using an optimal filter, in which a template pulse of fixed shape was scaled to best match the amplitude of the (variable shape) phonon pulse. These measured amplitudes, then, are shape-dependent. Distributions of some example partition and pulse shape quantities are shown in Figure 2. ![Example phonon discrimination quantities. On the left, an example of partitioning between channels is shown. The vertical axis is an axial partition quantity (side1 - side2)/(side1 + side2), and the x axis is a similarly defined radial partition quantity. On the right, the time difference (in $\mu$s) between the 40$\%$ and 70$\%$ points on the rising edge of the side 1 summed pulse (x axis) and side 2 summed pulse (y axis) are shown. In both plots, calibration events are colored using a charge-signal-based categorization as either bulk electron recoils (blue or dark grey), bulk nuclear recoils (green or light grey), or side 1 electron recoils (black). The total phonon energy for these events are between 7 and 20 keV recoil energy, where recoil energy has been scaled from total phonon energy using a nuclear recoil assumption for all events.\[fig:rqdistributions\]](rqdistributions.png) Discrimination Based on Phonon Pulse Quantities =============================================== As a very first look at discrimination based on phonon pulse shape characteristics, we here show a simple example of bulk electron recoil (ER) vs bulk nuclear recoil (NR) discrimination using only one pulse shape characteristic: the 40$\%$-to-70$\%$ risetimes. This simple discrimination example serves as a lower bound on the abilities one might expect when using a combined analysis of all phonon pulse quantities. We first construct ER and NR populations. Identical charge-based fiducial volume cuts are enforced on both populations, and measured \[charge:phonon\] energy ratios are used to categorize events as either ER or NR. After these two populations have been defined, a new pulse shape discrimination quantity is constructed using the 40$\%$-to-70$\%$ risetimes (for side 1 and side 2) for each event: “radius” = (\[40$\%$-to-70$\%$ side 1\]$^2$ + \[40$\%$-to-70$\%$ side 2\]$^2$)$^{1/2}$. One can see in Figure 2 that this combined quantity is largely position-independent. Figure 3 shows a histogram of this timing quantity, plotting only the lowest energies inspected in this analysis (7 $<$ E$_{recoil}$ $<$ 20 keV, defined by assuming that all events are nuclear recoils and scaling the total phonon energy accordingly). Although the ER and NR distributions overlap somewhat, there are no slow ER outliers to the statistics available in the calibration dataset. The right panel of Figure 3 shows the NR acceptance fraction vs ER leakage fraction as one varies a 1D cut threshold. Discrimination better than 1:10$^3$ is seen, and it is further seen that this discrimination shows no degradation with energy down to at least 7 keV E$_{recoil}$. ![A simple look at pulse shape discrimination between bulk nuclear recoils and bulk electron recoils, using 40$\%$-to-70$\%$ risetimes. Nuclear recoils are green (light grey); electron recoils are blue (dark grey). For discussion, see text.\[fig:discrimination\]](discrimination.png) Conclusions =========== Although significant pulse shape discrimination ability has been demonstrated, phonon-only discrimination is only beginning to show its promise. The simple analysis performed here can be improved in many ways, and looking further on the horizon, this rough analysis using a detector optimized for charge-electrode-based discrimination suggests that a detector optimized specifically for phonon-only discrimination could take advantage of the extreme sensitivity of transition edge sensors to extend event-by-event discrimination well below the typical CDMS low energy threshold of $\sim$10 keV (set by charge readout noise). Such phonon-only discrimination capabilities should be possible if detectors are specifically optimized for this goal, through the reduction of TES internal thermal fluctuation noise[@adam2011], and also through an increase in pulse shape differences themselves by increasing the total phonon-absorbing Al area of the QETs. We would like to thank E. Figueroa-Feliciano, S.W. Leman, and Adam Anderson for valuable discussions. This work is supported by the United State Department of Energy under grand DE-AC02-76SF00515 and by the United States National Science Foundation under grants PHY-0847342, 0705052, 0902182, and 1004714. [99]{} P. Luke, [Appl. Phys. Lett.]{} 65, 2884 (1994). P.L Brink et al, [Nucl. Instr. Meth. A]{} 559(12) 414 (2006). J. Lindhard, M.S.V. Nielsen, and P. Thomsen, [Mt. Fys. Medd. Dan. Vid. Selsk.]{} 33(10), 1 (1963). M. Pyle et al, [AIP Conf. Proc., LTD 13]{} 1185, 223 (2009). C.N. Bailey et al, [AIP Conf. Proc., LTD 13]{} 1185, 643 (2009). S. Tamura, [Phys. Rev. B]{} 31, 2574 (1985). M. E. Msall and J. P. Wolfe, [Phys. Rev. B]{} 56, 9557 (1997). B.S. Neganov and V.N. Trofimov, [J. Exp. Theor. Phys.]{} 28, 328 (1978). P.N. Luke, [Journal of Applied Physics]{} 64(12), 6858 (1988). G. Wang, [Journal of Applied Physics]{} 107, 094504 (2010). S.W. Leman et al, [Chinese Journal of Physics Research A]{} 49, 349 (2011). N. A. Court, A. J. Ferguson, and R. G. Clark, [Supercond. Sci. Technol.]{} 21, 015013 (2008). S.B. Kaplan et al, [Phys. Rev. B]{} 14, 4854 (1976). A. G. Kozorezov, J. K. Widmore, D. Martin, P. Verhoeve, and A. Peacock, [Phys. Rev. B]{} 75, 094513 (2007). K.A. McCarthy et al, [these proceedings.]{} A.J. Anderson et al, [these proceedings.]{}	{ "pile_set_name": "ArXiv" }
--- abstract: 'If $d$ is even, the resonances of the Schrödinger operator $-\Delta +V$ on $\Real^d$ with $V$ bounded and compactly supported are points on $\Lambda$, the logarithmic cover of $\Complex \setminus \{0\}$. We show that for fixed sign potentials $V$ and for $m\in \Integers \setminus \{0\}$, the resonance counting function for the $m$th sheet of $\Lambda$ has maximal order of growth.' address: 'Department of Mathematics, University of Missouri, Columbia, Missouri 65211, USA' author: - 'T.J. Christiansen' title: Lower bounds for resonance counting functions for Schrödinger operators with fixed sign potentials in even dimensions --- Introduction ============ The purpose of this paper is to prove some optimal lower bounds on the growth rate of resonance-counting functions for certain Schrödinger operators in even-dimensional Euclidean space. The resonances associated to the Schrödinger operator $-\Delta +V$, with potential $V\in L^{\infty}_{\comp}(\Real^d)$, lie on $\Lambda$, the logarithmic cover of $\Complex \setminus \{ 0 \}$, if $d$ is even. The main result of this paper is that for scattering by a fixed sign, compactly supported potential $V$ the resonance counting function for the $m$th sheet of $\Lambda$ has maximal order of growth for any $m\in \Integers \setminus \{ 0\}$. Though the results of [@ch-hi2] show that there are many potentials with resonance counting functions for the $m$th sheet having maximal order of growth, the technique of [@ch-hi2] does not give a way of identifying them other than those which are scalar multiples of the characteristic function of a ball. In comparison, in odd dimensions $d\geq 3$ the only specific real-valued potentials $V\in L^\infty_{\comp}(\Real^d)$ which are known to have resonance-counting function with optimal order of growth are certain radial potentials [@zwradpot], though in that case asymptotics are known (see [@zwradpot] and [@dinh-vu]). Let $V\in L^\infty_{\comp}(\Real^d)$ and let $\Delta\leq 0$ denote the Laplacian on $\Real^d$. We define the resolvent $R_V(\lambda)=(-\Delta +V-\lambda^2)^{-1}$ for $\lambda$ in the “physical space”, $0<\arg \lambda <\pi$. With at most a finite number of exceptional values of $\lambda$, $R_V(\lambda)$ is bounded on $L^2(\Real^d)$ for $\lambda$ in this region. It is well known that for $\chi \in L^{\infty}_{\comp}(\Real^d)$, $\chi R_V(\lambda)\chi$ has a meromorphic continuation to $\Complex$ when $d$ is odd and to $\Lambda$, the logarithmic cover of $\Complex \setminus \{ 0\}$, when $d$ is even (e.g. [@lrb Chapter 2]). In either case, the resonances are defined to be the poles of $\chi R_V(\lambda)\chi$ when $\chi$ is chosen to satisfy $\chi V\equiv V$. The fact that when $d$ is even the resonances lie on $\Lambda$ makes them generally more difficult to study in the even-dimensional case than in the odd-dimensional case. A point on $\Lambda$ can be described by its modulus and argument, where we do not identify points which have arguments differing by nonzero integral multiples of $2\pi$. Thus the physical half plane corresponds to $\Lambda_0 {\stackrel{\rm{def}}{=}}\{ \lambda \in \Lambda: 0<\arg \lambda <\pi\}$. Likewise, for $m\in \Integers$ we may define the $m$th sheet to be $$\Lambda_m {\stackrel{\rm{def}}{=}}\{ \lambda \in \Lambda: m\pi <\arg \lambda<(m+1)\pi\}$$ which is homeomorphic with the physical region and can be identified with the upper half plane when convenient. Vodev [@vodeveven; @vodev2], following earlier work of Intissar [@intissar] studied the resonance counting function $n_V(r,a)$, defined to be the number of resonances (counted with multiplicity, here and everywhere) with norm at most $r$ and argument between $-a$ and $a$. He showed that there is a constant $C$ which depends on $V$ but not on $r$ or $a$ so that $$n_V(r,a) \leq C a( r^d + (\log a)^d)),\; \text{for}\; r,\; a>1.$$ The most general lower bound known is due to SáBarreto ([@SaB1], $d\geq 4$) and Chen ([@chen], d=2): $$\lim \sup_{r \rightarrow \infty} \frac{\# \{ \lambda_j:\; \text{pole of $R_V(\lambda)$}\; \text{with} \;\frac{1}{r}\leq \|\lambda_j\|\leq r,\; \|\arg \lambda_j\| \leq \log r \}}{(\log r)( \log \log r)^{-p}} =\infty \; \forall p>1$$ for any nontrivial $V\in C_c^{\infty}(\Real^d;\Real)$. This follows the earlier work of [@s-t]. We note that the assumption that the potential is real-valued is crucial here. There are explicit examples of nontrivial complex-valued potentials $V\in L^{\infty}_{\comp}(\Real^d)$ which can be chosen to be smooth so that the corresponding Schrödinger operator $-\Delta +V$ has neither eigenvalues nor resonances [@autin; @chex; @iso]. For $m\in \Integers$, let $n_m(r)=n_{m,V}(r)$ be the number of resonances of $-\Delta + V$ which both lie on $\Lambda_m$ and have norm at most $r$. We call this the resonance counting function for the $m$th sheet. It follows from Vodev’s result that $n_m(r)= O(r^d)$ as $r\rightarrow \infty$. On the other hand, lower bounds have proved more elusive. The results of [@ch-hi2 Theorem 1.1] show that “generically” for potentials $V\in L^{\infty}_{\comp}(\Real^d)$, $m\in \Integers \setminus \{ 0\}$, $$\label{eq:limsup} \lim \sup _{r\rightarrow \infty}\frac{\log n_{m,V}(r)}{\log r}=d.$$ However, the result of [@ch-hi2] is nonconstructive in the sense that other than potentials which are nonzero positive scalar multiples of the characteristic function of a ball and those complex-valued potentials which are isoresonant with them [@iso], that paper does not give a way of identifying the particular potentials for which (\[eq:limsup\]) holds. The main result of this paper is the following theorem. \[thm:lbd\] Let $d$ be even. Suppose $V\in L^{\infty}_{\comp}(\Real^d)$ with $V$ bounded below by $\epsilon \chi_B$, where $\epsilon>0$ and $\chi_B$ is the characteristic function of a nontrivial ball $B$. Then for any nonzero $m\in \Integers$, $$\lim \sup _{r \rightarrow \infty} \frac{\log n_{m, \pm V}(r)}{\log r} =d.$$ We note that by Vodev’s result $d$ is the maximum value this limit can obtain. When this limit is $d$, we say that the $m$th counting function has maximal order of growth. Theorem \[thm:lbd\], when combined with [@ch-hi2 Theorem 3.8], has the following theorem as an immediate corollary. Let $d$ be even, and $K\subset \Real^d$ be a compact set with nonempty interior. Let $F$ denote either $\Real $ or $\Complex$. Then for $m\in \Integers$, $m\not = 0$, the set $$\{ V\in C^{\infty}(K;F): \lim\sup_{r\rightarrow \infty } \frac{ \log n_{m,V}(r)}{\log r}=d\}$$ is dense in $C^{\infty}(K;F)$. For the case of [odd]{} dimension $d\geq 3$, the analog of this theorem was proved in [@scv]. A stronger result holds in dimension $d=1$, see [@zworski1d] or [@froese; @regge; @simon]. Theorem \[thm:lbd\] may be compared with other results for fixed-sign potentials. In the odd-dimensional case Lax-Phillips [@l-p] and Vasy [@vasy] proved lower bounds on the number of pure imaginary resonances for potentials of fixed sign. In [@ch-hi4] it is shown that in even dimensions there are no “pure imaginary” resonances for positive potentials, and on each sheet $\Lambda_m$ of $\Lambda$, only finitely many for negative potentials. Both [@l-p] and [@vasy] use a monotonicity property for potentials of fixed sign. This paper also uses a monotonicity property, though it is more closely related to one used in [@scv]. Also important here are some results from one-dimensional complex analysis, more delicate than the corresponding complex-analytic arguments from [@scv]. [Acknowledgments.]{} The author gratefully acknowledges the partial support of the NSF under grant DMS 1001156. Some Complex Analysis ===================== The main result of this section is Proposition \[p:complex\], which, roughly speaking, controls the growth of a function $f$ analytic in a half plane in terms of the growth of the counting function for the zeros of $f$ in the half plane and the behavior of $f$ on the boundary of the half plane. Both the statement and the proof of the following lemma bear some resemblance to those for Carathéodory’s inequality for the disk. The estimate we obtain here is likely a crude one, but suffices for our purposes. \[l:caratheodorytype\] Let $f$ be analytic in a neighborhood of $$\Omega_R{\stackrel{\rm{def}}{=}}\{ z\in \Complex: 1\leq \|z\| \leq R,\; {\operatorname{\rm Im}\nolimits}z\geq 0\},$$ $\rho>0$, and for $x\in \Real \cap \Omega_R$, $\|f(x)\|\leq C_0\|x\|^{\rho}$ for some constant $C_0>0$. Set $$M=\max_{\|z\|=1, \; z\in \Omega_R}\|f(z)\|$$ and define $$A(R)= \max\left( C_0 R^\rho, M R^\rho , \max_{ z\in \Omega_R }{\operatorname{\rm Re}\nolimits}f(z) \right).$$ Then if $1<r<R$ and $z\in \Omega_R$ with $\|z\|=r$, then $\|f(z)\|\leq \frac{2r^{\rho}}{R^\rho- r^\rho} A(R).$ Set $$g(z)= \frac{1}{z^\rho} \frac{f(z)}{2A(R)-f(z)}$$ which is analytic in a neighborhood of $\Omega_R$. We bound $\|g\|$ on the boundary of $\Omega_R$. If $z\in \Omega_R$ has $\|z\|=R$, then $$\|g(z)\|\leq \frac{1}{R^\rho} \frac{\|f(z)\|}{\|2A(R)-f(z)\|}\leq \frac{ \|f(z)\|}{R^\rho \|f(z)\|}=\frac{1}{R^\rho}.$$ Notice that if $x\in \Omega_R\cap \Real$, $$\|g(x)\|\leq \frac{1}{\|x\|^\rho}\frac{C_0 \|x\|^\rho}{R^\rho C_0} \leq \frac{1}{R^\rho}.$$ Moreover, if $z\in \Omega_R$ has $\|z\|=1$, since $A(R)\geq \|f(z)\|R^\rho$, $\|g(z)\|\leq 1/R^\rho$. Thus, by the maximum principle $\|g(z)\|\leq 1/R^\rho$ for all $z\in \Omega_R$. Suppose $z\in \Omega_R$ with $\|z\|=r$, $1<r<R$. Then $$\|f(z)\|\leq \|z\|^\rho \|2 A(R)-f(z)\| R^{-\rho}\leq r^\rho (2A(R)+\|f(z)\|) R^{-\rho}.$$ Rearranging, we find $$(R^\rho -r^\rho )\|f(z)\|\leq 2 r^\rho A(R),$$ or $$\|f(z)\|\leq \frac{ 2 r^\rho}{R^\rho -r^\rho}A(R).$$ \[l:p-lapp\] Let $\Omega=\{ z:\; {\operatorname{\rm Im}\nolimits}z\geq 0,\;\|z\|\geq 1\}$ and suppose $f$ is analytic in a neighborhood of $\Omega$, and there are constants $\rho_0$, $C_0$, so that $\|f(z)\|\leq C_0 \exp (C_0 \|z\|^{\rho_0})$ for all $z\in \Omega$. Suppose there are constants $C_1$, $\rho>0$ so that for all $x>1$, $\|\int_1^x f'(t)/f(t)dt\|\leq C_1 \|x\|^\rho$ and for all $x<-1$, $\|\int_x^{-1} f'(t)/f(t)dt\|\leq C_1 \|x\|^\rho$. If, in addition, $f$ does not vanish in $\Omega$, then there is a constant $C_3$ so that $\|f(z)\|\leq C_3 \exp(C_3\|z\|^\rho)$ for all $z\in \Omega$. In the proof we shall denote by $C$ a constant the value of which may change from line to line without comment. Since $f$ is nonvanishing in $\Omega$, there is a function $g$ analytic on $\Omega$ so that $\exp g(z)= f(z)$. Since $g'(z)= f'(z)/f(z)$, $$g(x)-g(1)= \int_1^x \frac {f'(t)}{f(t)} dt\; \text{if}\; x>1$$ so that $\|g(x)\|\leq C \|x\|^\rho+\|g(1)\|$ when $x\geq 1$ for some constant $C$. A similar argument gives a similar bound for $x\leq -1$. We now assume $\rho < \rho_0$ since otherwise there is nothing to prove. We give a bound on the growth of $g$ at infinity which is more than adequate to allow us to apply a version of the Phragmén-Lindellöf theorem as we will below. Since ${\operatorname{\rm Re}\nolimits}g(z)=\log \|f(z)\|$, for all $z\in \Omega$, ${\operatorname{\rm Re}\nolimits}g(z) \leq C (1+ \|z\|^{\rho_0})$. Applying Lemma \[l:caratheodorytype\], we find that $\|g(z)\|\leq C (1+\|z\|^{\rho_0})$ for all $z\in \Omega$. Consider the function $h(z)= g(z) / (i+z)^{\rho}$. This is an analytic function in a neighborhood of $\Omega$ and is bounded on $\partial \Omega$. Thus, by a version of the Phragmén-Lindellöf theorem (proved, for example, by an easy modification of the proof of [@conway Corollary VI.4.2] using [@conway Theorem VI.4.1]), $h$ is bounded in $\Omega$. This implies that for $z\in \Omega$, $\|f(z)\| = \exp ({\operatorname{\rm Re}\nolimits}g(z)) \leq \exp (C(1+\|z\|)^{\rho})$ for some constant $C$. We shall use the notation $E_0(z) =1-z$ and $E_p(z) = (1-z)\exp(z+z^2/2+...+ z^p/p)$ for $p\in \Natural$ for a canonical factor. The proof of the following lemma bears many similarities to proofs for estimates of canonical products; see for example [@levin Lemma I.4.3]. \[l:cproductreals\] Let $\{a_j\}\subset \Complex$ be a set of not necessarily distinct points in the open upper half plane, with $\|a_1\|\leq \|a_2\|\leq ...$ and suppose for some constants $C_0$ and $\rho$ $$n(r){\stackrel{\rm{def}}{=}}\# \{ j :\; \|a_j\|\leq r \} \leq C_0 r^\rho\; \text{ when} \; r\geq 1.$$ Suppose $\rho > 0$ is not an integer, let $p$ be the greatest integer less than $\rho$, and set $$f(z)=\prod_{n=1}^{\infty} \frac{E_p(z/\overline{a_n})}{ E_p(z/a_n)}.$$ Then for $x\in \Real$ $$\left\|\int_0^x \frac{f'(t)}{f(t)} dt\right\| = O(\|x\|^\rho)$$ as $\|x\|\rightarrow \infty$. We note first that our assumption on $n(r)$ ensures that the canonical products converge, so that $f$ is a meromorphic function on $\Complex$. Moreover, by assumption $f$ has neither poles nor zeros on the real line. A computation shows that $E_p'(z)/E_p(z)= -z^p/(1-z)$. Thus $$\label{eq:logderiv} \frac{f'(x)}{f(x)}= \sum_{n=1}^{\infty}\left( \frac{(x/\overline{a_n})^p}{x-\overline{a_n}}-\frac{(x/a_n)^p}{x-a_n}\right).$$ Let $a\in \Complex$, ${\operatorname{\rm Im}\nolimits}a >0 $, $t\in \Real$. Then $$\begin{aligned} \label{eq:basic} \frac{(t/\overline{a})^p}{t-\overline{a}}- \frac{(t/a)^p}{t-a} & = \frac{1}{\|a\|^{2p}} t^p \left(\frac{ a^p(t-a)-\overline{a}^p (t-\overline{a})}{\|t-a\|^2}\right) \nonumber \\ & = 2i \frac{1}{\|a\|^{2p}} t^p \left( \frac{ t {\operatorname{\rm Im}\nolimits}(a^p) - {\operatorname{\rm Im}\nolimits}(a^{p+1})}{\|t-a\|^2}\right).\end{aligned}$$ Now set $a=\alpha + i\beta$, $\beta>0$, and note that $\|{\operatorname{\rm Im}\nolimits}a^p\| \leq p\beta \|a\|^{p-1}$. Thus for $x\in \Real$ $$\begin{aligned} \label{eq:intbd} \left\| \int_0^x \left( \frac{(t/\overline{a})^p}{t-\overline{a}}- \frac{(t/a)^p}{t-a}\right) dt\right\| & \leq \frac{2}{\|a\|^{p+1}} \left\| \int_0^x \frac{ p\|t\|^{p+1}\beta + (p+1) \|a\|\|t\|^p\beta}{(t-\alpha)^2+ \beta^2} dt\right\|.\end{aligned}$$ Now for $q>0$ $$\int_0^x \frac{ t^{q}\beta }{(t-\alpha)^2+ \beta^2} dt = x^{q} \arctan((x-\alpha)/\beta)- q\int_0^x t^{q-1} \arctan((t-\alpha)/\beta) dt.$$ Using that for $s\in \Real$, $\|\arctan s\| <\pi/2$, we find that for $\|x\|>1$ $$\label{eq:est1} \left\|\int_0^x \frac{ t^{q}\beta }{(t-\alpha)^2+ \beta^2} dt \right\| \leq C \|x\|^{q}$$ for some constant $C$, independent of $\alpha$ and $\beta$. To prove the lemma, we will split $\{a_j\}$ into two sets, depending on the relative size of $\|a_j\|$ and $2\|x\|$. For $\|a_j\|\leq 2\|x\|$, we first note that $$p\int _0^x \frac{ \beta t^{p+1}}{(t-\alpha)^2+\beta^2}dt = p\beta \int _0^x t^{p-1}\left( 1+ \frac{ 2\alpha t -(\alpha^2+\beta^2) } {(t-\alpha)^2+\beta^2}\right)dt$$ and use (\[eq:intbd\]) and (\[eq:est1\]) to get $$\begin{aligned} \label{eq:smalla} \sum_{\|a_j\|\leq 2\|x\| } \left\| \int_0^x \left( \frac{(t/\overline{a_j})^p}{t-\overline{a_j}}- \frac{(t/a_j)^p}{t-a_j}\right) dt\right\| & \leq C \sum_{\|a_j\|\leq 2\|x\|} \left( \|x\|^p \|a_j\|^{-p} + \|x\|^{p-1}\|a_j\|^{-p+1} \right).\end{aligned}$$ Since $$\label{eq:apowersum} \sum_{1\leq \|a_j\|\leq r}\|a_j\|^{q} = \int_1^r t^{q}dn(t)= r^qn(r)-n(1)-\int_1^r qt^{q-1} n(t)dt$$ applying our upper bound on $n(r)$ we get from (\[eq:smalla\]) $$\label{eq:asmallbound} \sum_{\|a_j\|\leq 2\|x\| } \left\| \int_0^x\left( \frac{(t/\overline{a_j})^p}{t-\overline{a_j}}- \frac{(t/a_j)^p}{t-a_j}\right) dt\right\|\leq C(\|x\|^{\rho}+1).$$ Now we bound the contribution of the $a_j$ with $\|a_j\|> 2 \|x\|$. For this we use (\[eq:basic\]) more directly. Here $$\begin{aligned} & \sum_{\|a_j\|>2\|x\|} \left\| \int_0^x \left( \frac{(t/\overline{a_j})^p}{t-\overline{a_j}}-\frac{(t/a_j)^p}{t-a_j}\right) dt \right\| \\ & = 2 \sum_{\|a_j\|>2\|x\|} \left\| \int_0^x \frac{1}{\|a_j\|^{2p}} t^p \left( \frac{ t {\operatorname{\rm Im}\nolimits}(a_j^p) - {\operatorname{\rm Im}\nolimits}(a_j^{p+1})}{\|t-a_j\|^2}\right)dt\right\|\\ & \leq C \sum_{\|a_j\|>2\|x\|} \left\| \int_0^x \frac{1}{\|a_j\|^{2p}} \left( \frac{ \|t\|^{p+1} \|a_j\|^p + \|t\|^p\|a_j\|^{p+1})}{\|a_j\|^2}\right) dt \right\| \\ & \leq C \sum_{\|a_j\|>2\|x\|} (\|x\|^{p+2} \|a_j\|^{-p-2}+\|x\|^{p+1}\|a_j\|^{-p-1}).\end{aligned}$$ Applying the analog of (\[eq:apowersum\]) and using the upper bound on $n(r)$ we obtain $$\sum_{\|a_j\|>2\|x\|} \|a_j\|^{-q} \leq C\|x\|^{\rho-q}$$ provided $q>\rho$, giving us $$\sum_{\|a_j\|>2\|x\|} \left\| \int_0^x \left( \frac{(t/\overline{a_j})^p}{t-\overline{a_j}}-\frac{(t/a_j)^p}{t-a_j}\right) dt \right\| \leq C\|x\|^\rho.$$ Combined with (\[eq:asmallbound\]), this completes the proof of the lemma. \[p:complex\] Let $f$ be a function analytic in a neighborhood of $ \Omega=\{ z:\: {\operatorname{\rm Im}\nolimits}z \geq 0, \; \|z\|\geq 1 \}$. Suppose $f$ does not vanish on $\Real \cap \Omega$, and let $n(r)$ be the number of zeros of $f$ in $\{ z:\; {\operatorname{\rm Im}\nolimits}z \geq 0, \; 1\leq \|z\|\leq r\}$ counted with multiplicity. Suppose that that there are constants $C_0$ and $\rho$, $\rho$ not an integer, so that $$n(r) \leq C_0(1+r^\rho )$$ and $$\left\| \frac{f'(x)}{f(x)}\right\| \leq C_0 (1+\|x\|^{\rho-1}) \; \text{ for all } x\in \Real \; \text{with}\; \|x\|\geq 1.$$ Suppose in addition that there are some constants $\rho_1$, $C_1$ so that $\log \|f(z)\|\leq C_1(1+\|z\|^{\rho_1}) $ for all $z\in \Omega$. Then there is a constant $C$ so that $\|f(z)\|\leq Ce^{\|z\|^\rho} $ for $z\in \Omega$. We will assume $\rho_1>\rho$ as otherwise there is nothing to prove. To aid in notation, we set $$\Omega_R= \{ z\in \Complex: \; {\operatorname{\rm Im}\nolimits}z \geq 1\; \text{and}\; 1\leq \|z\|\leq R\}.$$ We prove this proposition by constructing a function to which we can apply Lemma \[l:p-lapp\]. Let $p$ denote the greatest integer less than $\rho$, and $\{ a_j\}$ the zeros of $f$ in $\Omega$, repeated according to multiplicity, with $\|a_1\|\leq \|a_2\|\leq ...$. Set $$h(z)=\frac{f(z)g_1(z)}{g_2(z)},$$ where $$g_1(z)= \prod_{n=1}^{\infty} E_p(z/\overline{a_n}) \; \text{and}\; g_2(z)= \prod_{n=1}^{\infty} E_p(z/a_n).$$ Note that $h$ is analytic in $\Omega$ and does not vanish there. As an intermediate step we show that $\log \|h(z)\|\leq C\|z\|^{\rho_1} $ for all $z\in \Omega$. Recall we have assumed $\rho_1>\rho$. Here and below $C$ is a finite constant which may change from line to line. If $x\in \Real$, $1\leq \|x\|$, then $\log\|h(x)\|=\log\|f(x)\|\leq C_0(1+\|x\|^{\rho_1})$. Moreover, from estimates on canonical products, $$\label{eq:gjbound} \log \|g_j(z)\| \leq C(1+\|z\|^\rho), \; j=1,2$$ for some constant $C$, see [@levin Lemma I.4.3]. To aid in notation, we set $\Omega_R= \{ z\in \Complex: \; {\operatorname{\rm Im}\nolimits}z \geq 1\; \text{and}\; 1\leq \|z\|\leq R\}$. As is shown in the proof of [@levin Theorem I.12], given $R>0$ and $0<\delta<1$ there is an $r_j\in [R, R(1-\delta)^{-1}]$ so that for all $z\in \Complex$ with $\|z\|=r_j$, $$\begin{aligned} \log \|g_j(z)\| & \geq -\left(2+\log \frac{12 e}{\delta}\right) \log \max_{\|z\|=2 e R(1-\delta)^{-1}} \|g_j(z)\|,\; \text{for $j=1,\;j$}.\end{aligned}$$ Using (\[eq:gjbound\]), this gives $$\begin{aligned} \label{eq:g2lb} \log \|g_j(z)\| & \geq -C_{ \delta} (1+(R(1-\delta)^{-1})^\rho ),\; \|z\|=r_j,\; j=1,\;2.\end{aligned}$$ Now fix a $\delta>0$, $\delta<1$. Given any $R>1$, we can find an $r_2\in [R,R(1-\delta)^{-1}]$ as above so that (\[eq:g2lb\]) holds for $j=2$. Then using that $$\max_{z\in \Omega_R } \log \|h(z)\| \leq \max_{ z\in \Omega_{r_2}} \log \|h(z)\|= \max_{z\in \partial \Omega_{r_2}} \log \|h(z)\|,$$ our assumptions on $f$, and (\[eq:g2lb\]) we find $$\max_{ z\in \Omega_R}\log \|h(z)\| \leq C_\delta (1+ (R(1-\delta)^{-1})^{\rho_1})+ C(1+R^\rho)\leq \tilde{C}_\delta (1+R^{\rho_1}).$$ For $x\in \Real$, $\|x\|\geq 1$, $$\begin{aligned} \frac{h'(x)}{h(x)}& =\frac{f'(x)}{f(x)}+ \frac{(g_1/g_2)'(x)}{g_1(x)/g_2(x)}.\end{aligned}$$ By applying our assumptions on $f$ and Lemma \[l:cproductreals\], we find that for $x>1$, $\|\int_1^x h'(t)/h(t) dt\| = O(x^\rho)$, and likewise for $x<-1$, $\|\int _{x}^{-1} h'(t)/h(t) dt\|=O(\|x\|^{\rho})$. By Lemma \[l:p-lapp\], there is a constant $C$ so that $$\label{eq:hbd} \log \|h(z)\|\leq C(1+ \|z\|^\rho),\; \text{when}\; z\in \Omega.$$ Now we write $f(z) = g_2(z)h(z)/g_1(z)$, holomorphic in a neighborhood of $\Omega$. Given $R>1$ and $\delta$ satisfying $0<\delta<1$, as above we choose $r_1\in [R,R(1-\delta)^{-1}]$ so that (\[eq:g2lb\]) holds for $g_1$. Using in addition (\[eq:gjbound\]) and (\[eq:hbd\]), we find there is a constant so that $$\log \|f(z)\|\leq C(1+(R(1-\delta )^{-1})^{\rho})\; \text{for}\; \|z\|=r_1, \; {\operatorname{\rm Im}\nolimits}z\geq 0.$$ As in the proof of the bound on $h$, since $\|h(x)\|=\|f(x)\|$ for $x\in \Real \cap \Omega_R$, we find then that there is a constant $C$ so that $$\max_{z\in \Omega_R}\log \|f(z)\|\leq C(1+R^{\rho}).$$ A scalar function having zeros at the poles of the resolvent {#s:defF} ============================================================ We recall the derivation of some identities commonly used in the study of resonances for Schrödinger operators. Let $V\in L^{\infty}_{\comp}(\Real^d)$ and let $d\geq 2$ be even. There is no need to make an assumption on the sign of $V$ here. We recall the notation $R_V(\lambda)=(-\Delta +V-\lambda^2)^{-1}$ when $\lambda \in \Lambda_0$. For such $\lambda$, $(-\Delta +V-\lambda^2)R_0(\lambda)= I+V R_0(\lambda)$ and by meromorphic continuation, $$R_0(\lambda) = R_V(\lambda)(I+VR_0(\lambda)), \; \lambda \in \Lambda.$$ Thus $R_V(\lambda)$ has a pole if and only if $I+VR_0(\lambda) $ has a zero, and multiplicities agree. Writing $V^{1/2}=V/\|V\|^{1/2}$ with the convention that $V^{1/2}=0$ outside the support of $V$, we see that $I+VR_0(\lambda)$ has a zero if and only if $I+V^{1/2}R_0(\lambda)\|V\|^{1/2}$ has a zero. Consequently, $I+V^{1/2}R_0(\lambda)\|V\|^{1/2}$ is invertible for all but a finite number of points in $\overline{\Lambda_0}$. Thus, if $m\in \Integers$, $\lambda\in \overline{\lambda_0}$, $$\begin{gathered} I+V^{1/2}R_0(e^{im\pi}\lambda)\|V\|^{1/2}\\ = (I+V^{1/2}R_0(\lambda)\|V\|^{1/2})\left(I+(I+V^{1/2}R_0(\lambda)\|V\|^{1/2})^{-1} V^{1/2}\left(R_0(e^{im\pi}\lambda)-R_0(\lambda)\right)\|V\|^{1/2}\right).\end{gathered}$$ But when $d$ is even $$R_0(e^{im\pi}\lambda)-R_0(\lambda)= imT(\lambda)$$ with $$(T(\lambda)f)(x)=\alpha_d \lambda^{d-2}\int_{\Real^d} \int_{\Sphere^{d-1}} e^{i\lambda (x-y)\cdot \omega} f(y) d\omega \; dy$$ for $f\in L^{2}_{\comp}(\Real^d)$, with $\alpha_d =(2\pi)^{1-d}/2$, [@lrb (1.32)]. Moreover, $V^{1/2}T(\lambda) \|V\|^{1/2}$ is trace class. Thus, with at most a finite number of exceptions, the poles of $R_V(e^{im\pi}\lambda)$ with $\lambda \in \Lambda_0$ correspond, with multiplicity, to the zeros of $$\label{eq:FmV} F_{m,V}(\lambda) {\stackrel{\rm{def}}{=}}\det \left(I+im(I+V^{1/2}R_0(\lambda)\|V\|^{1/2})^{-1} V^{1/2}T(\lambda) \|V\|^{1/2}\right)$$ in $\Lambda_0$. Lower bounds on $F_{m,\pm V}(i\sigma)$ when $V$ has fixed sign ============================================================== In the remainder of this paper we assume $d\geq 2$ is even. Let $V\geq 0$, $V\in L^{\infty}_{\comp}(\Real^d)$. In this section we study the function $F_{m,\pm V}$ from (\[eq:FmV\]). For $\sigma \in \Real_+$, we shall use the shorthand $i\sigma$ to denote the point in the physical region with norm $\sigma$ and argument $\pi/2$. Taking the positive sign, $I+V^{1/2}R_0(i\sigma)\|V\|^{1/2}=I+V^{1/2}R_0(i\sigma)V^{1/2} $ is a positive operator for $\sigma>0$. When we choose the negative sign, we will additionally assume that $\sigma$ is chosen large enough that $I- \vhalf R_0(i\sigma) \vhalf $ is a positive invertible operator; this is possible by, for example, insisting $\sigma>2(\\| V \\|_{\infty}+1)$ since $\\|R_0(i\sigma)\\| \leq 1/\sigma^2$. With these assumptions on $\sigma$, using the properties of the determinant and the fact that $V\geq 0$ we may rewrite the function $F_{m,\pm V}(i\sigma)$ from (\[eq:FmV\]) as $$\begin{aligned} \label{eq:Fmrewrite} F_{m}(i\sigma)& =F_{ m,\pm V}(i\sigma)\\ &= \det \left( I \pm im (I\pm \vhalf R_0(i\sigma) \vhalf )^{-1/2} \vhalf T(i\sigma) \vhalf (I\pm \vhalf R_0(i\sigma) \vhalf )^{-1/2} \right). \nonumber\end{aligned}$$ We shall obtain a lower bound on $F_m(i\sigma)$ as $\sigma \rightarrow \infty$. The following proposition is central to the proof of Theorem \[thm:lbd\] and is the main result of this section. Related results were obtained in odd dimensions in [@scv Section 5]. \[p:lb\] Let $V\in L^{\infty}_{\comp}(\Real^d)$, $V\geq 0$, and let $V$ be bounded below by $\epsilon \chi_{B}$ where $\epsilon>0$ and $\chi_B$ is the characteristic function of a nontrivial open ball. Let $m\in \Integers$, $m\not =0$. Then there is a constant $c_0>0$ so that $\|F_{m,\pm V}(i\sigma)\|\geq c_0\exp(c_0 \sigma ^d)$ for all sufficiently large $\sigma>0$. The proof is similar to the proofs of some results of [@l-p; @vasy] in that it uses both a property of monotonicity in $V$ and the fact that for potentials which are positive multiples of the characteristic function of a ball much can be said by using a decomposition into spherical harmonics and special functions. However, the implementation of these underlying ideas is rather different here. The proof of Proposition \[p:lb\] uses the following lemma, a monotonicity result reminiscent of results of [@l-p; @vasy]. In fact, the proof of this lemma uses a result from [@vasy]. \[l:normbd\] Let $V_1,\; V_2\in L^{\infty}(\Real^d)$ and suppose the support of $V_j$ is contained in $\overline{B}(R)= \{ x\in \Real^d: \|x\| \leq R\}$ for $j=1,\;2$. Suppose $V_2(x)\geq V_1(x) \geq 0$ for all $x\in \Real^d$. We use the convention that $\vhalf_1/\vhalf_2$ is $0$ outside the support of $V_1$. Then $$\left\\| (I + \vhalf_1 R_0(i\sigma) \vhalf_1)^{-1/2}\frac{ \vhalf_1}{\vhalf_2} (I + \vhalf_2 R_0(i\sigma) \vhalf_2)^{1/2} \right\\| \leq 1.$$ If $\sigma \geq 2(\\|V_2\\|_{\infty}+1)$, then $$\left\\| (I - \vhalf_1 R_0(i\sigma) \vhalf_1)^{-1/2} \frac{\vhalf_1}{\vhalf_2} (I - \vhalf_2 R_0(i\sigma) \vhalf_2)^{1/2} \right\\| \leq 1.$$ When $\sigma>0$ is sufficiently large that $I\pm \vhalf_j R_0(i\sigma) \vhalf_j$ is a positive operator, $$\begin{aligned} & (I\pm V_j R_0(i\sigma) )\vhalf_j (I\pm \vhalf_j R_0(i\sigma)\vhalf_j)^{-1}\\ & = \vhalf_j (I\pm \vhalf_j R_0(i\sigma)\vhalf_j) (I\pm \vhalf_j R_0(i\sigma)\vhalf_j)^{-1} \\& = \vhalf_j.\end{aligned}$$ Thus $$\vhalf_j(I\pm \vhalf_j R_0(i\sigma) \vhalf_j)^{-1} \vhalf_j = (I \pm V_j R_0(i\sigma))^{-1}V_j,\; j=1,2$$ for $\sigma>0 $ sufficiently large. Applying [@vasy Lemma 2.2], and using that $V_2\geq V_1$, we get $$(I + V_2 R_0(i\sigma))^{-1}V_2\geq (I + V_1 R_0(i\sigma))^{-1}V_1.$$ When we take the “$-$” sign, again applying [@vasy Lemma 2.2], $$(I - V_2 R_0(i\sigma))^{-1}V_2\geq (I - V_1 R_0(i\sigma))^{-1}V_1$$ when $\sigma>2( \\|V_2\\|_\infty +1)$. Here we note our convention differs somewhat from [@vasy], in that we take $V_j\geq 0$. Summarizing, $$\vhalf_2(I \pm \vhalf_2 R_0(i\sigma) \vhalf_2)^{-1} \vhalf_2 \geq \vhalf_1(I \pm \vhalf_1 R_0(i\sigma) \vhalf_1)^{-1} \vhalf_1$$ when $\sigma>0$ (for the “$+$” sign) or $\sigma>2(\\|V\\|+1)$ (for the “$-$” sign). For the remainder of the proof, we shall assume $\sigma>0$ satisfies these requirements and suppress the argument $i\sigma$. Now let $\chi_{V_2}$ be the characteristic function of the support of $V_2$ and recall $\vhalf_1 \chi_{V_2}=\vhalf_1$ and note that $\chi_{V_2}(I\pm \vhalf_2 R_0\vhalf_2)= (I\pm \vhalf_2 R_0\vhalf_2)\chi_{V_2}$. Then $$\chi_{V_2} (I \pm \vhalf_2 R_0 \vhalf_2)^{-1} \chi_{V_2} \geq \frac{ \vhalf_1}{\vhalf_2}(I \pm \vhalf_1 R_0 \vhalf_1)^{-1} \frac{\vhalf_1} {\vhalf_2}.$$ This implies $$\chi_{V_2} \geq (I \pm \vhalf_2 R_0 \vhalf_2)^{1/2} \frac{ \vhalf_1}{\vhalf_2}(I \pm \vhalf_1 R_0 \vhalf_1)^{-1} \frac{\vhalf_1} {\vhalf_2}(I \pm \vhalf_2 R_0 \vhalf_2)^{1/2}.$$ This proves the lemma, since the norm of the right hand side is the square of the norm of the operator in question. \[l:eigenvaluebd\] Let ${\mathcal H}$ be an infinite dimensional complex separable Hilbert space, $A,\; B\in {\mathcal L }({\mathcal H})$, with $B=B^$, and $\\|A\\| \leq 1$. Let $\|\lambda_1\|\geq \|\lambda_2\|\geq...$ be the norms of the eigenvalues of $A^BA$, and $\|\mu_1\|\geq \|\mu_2\|\geq...$ be the norms of the eigenvalues of $B$. In both cases we repeat according to multiplicity. Then $\|\mu_j\|\geq \|\lambda_j\|$ for all $j$. One way to prove this it that by noting that since $B$ and $A^BA$ are self-adjoint, the norms of the the eigenvalues are the characteristic values. Then this lemma is an immediate application of the bound for the characteristic values of a product found, for example, in [@simonti Theorem 1.6]. The next lemma shows that $F_{m,\pm V}(i\sigma)$ depends monotonically on $V$ in some sense. \[l:mono\] Let $V_1,\; V_2\in L^{\infty}(\Real^d)$ and suppose the support of $V_j$ is contained in $\overline{B}(R)$ for $j=1,\;2$. Suppose $V_2(x)\geq V_1(x)\geq 0$ for all $x\in \Real^d$. Then $\|F_{m,V_1}(i\sigma)\|\leq \|F_{m,V_2}(i\sigma)\|$ for all $\sigma \in \Real_+$. Moreover, if $\sigma \geq 2(\\|V_2\\|_{\infty}+1)$, then $\|F_{m,-V_1}(i\sigma)\|\leq \|F_{m,-V_2}(i\sigma)\|$. For any compactly supported $V\geq 0$, set $$\label{eq:B1} B_{1,\pm, V}(i\sigma ) = (I\pm \vhalf R_0(i\sigma)\vhalf)^{-1/2} \vhalf T(i\sigma) \vhalf (I\pm \vhalf R_0(i\sigma)\vhalf)^{-1/2}$$ and notice that if $\sigma >0$ (for the “$+$” sign) or $\sigma > 2(\\| V\\|_{\infty} +1)$ (for the “$-$” sign), $B_{1,\pm V}(i\sigma )$ is a self-adjoint trace class operator. Comparing (\[eq:Fmrewrite\]), we see that $$F_{m,\pm V}(i\sigma)= \det(I \pm i m B_{1,\pm, V}(i\sigma)).$$ Hence for sufficiently large $\sigma$ $$\begin{aligned} \label{eq:detprod} \|F_{m,\pm V}(i\sigma)\| & = \left\|\prod (I+ im \lambda_j(B_{1,\pm, V}(i\sigma))) \right\| \nonumber \\ & = \prod \left\|(I+ im \lambda_j(B_{1,\pm, V}(i\sigma)))\right\| \nonumber \\ & = \prod \sqrt{1+ m^2 \lambda_j^2(B_{1,\pm, V}(i\sigma))}\end{aligned}$$ where $\lambda_j(B_{1,\pm, V})$ are the nonzero eigenvalues of $B_{1,\pm, V}$, repeated according to multiplicity and arranged in decreasing order of magnitude: $\|\lambda_1(B_{1,\pm, V})\| \geq \|\lambda_2(B_{1,\pm, V})\| \geq ...$. Now we turn to $V_1$ and $V_2$, and $\sigma$ as in the statement of the lemma. Note that $$\begin{gathered} B_{1,\pm, V_1}(i\sigma)= (I \pm \vhalf_1 R_0(i\sigma) \vhalf_1)^{-1/2} \frac{\vhalf_1}{\vhalf_2} (I \pm \vhalf_2 R_0(i\sigma) \vhalf_2)^{1/2} B_{1, \pm , V_2}(i\sigma) \\ \times (I \pm \vhalf_2 R_0(i\sigma) \vhalf_2)^{1/2}\frac{\vhalf_1}{\vhalf_2}(I \pm \vhalf_1 R_0(i\sigma) \vhalf_1)^{-1/2}.\end{gathered}$$ Again we use the convention that $\vhalf_1/\vhalf_2$ is $0$ outside the support of $V_1$. The lemma now follows from (\[eq:detprod\]) and Lemmas \[l:normbd\] and \[l:eigenvaluebd\]. In order to obtain the lower bounds of Proposition \[p:lb\], we shall need a special case of that proposition, in which the potential is of the form $V(x)=\epsilon\chi_B(x)$, and $\chi_B(x)$ is the characteristic function of a ball centered at the origin. To study such a special case, we will introduce spherical coordinates in $\Real^d$ (polar coordinates in the case $d=2$). In spherical coordinates, $$-\Delta = -\frac{\partial^2}{\partial r^2}-\frac{d-1}{r}\frac{\partial }{\partial r} +\frac{1}{r^2}\Delta_{\Sphere^{d-1}}.$$ The eigenvalues of of the Laplacian on $\Sphere^{d-1}$, $\Delta_{\Sphere^{d-1}}$, are $l(l+d-2)$, $l\in \Natural_0$ with multiplicity $$\mu(l)=\frac{2l +d-2}{d-2}\left( \begin{array}{c} l+d-3\\ d-3 \end{array}\right) = \frac{2l^{d-2}}{(d-2)!}(1+O(l^{-1})).$$ Denote by $Y_l^\mu$, $1\leq \mu \leq \mu(l)$, $l=0,1,2,...$ a complete orthonormal set of spherical harmonics on $\Sphere^{d-1}$ with eigenvalue $l(l+d-2)$. We denote by $P_l$ projection onto the span of $$\{ h(\|x\|)Y^\mu_l(x/\|x\|): \; 1\leq \mu \leq \mu(l),\; h(\|x\|)\in L^2(\Real^d; r^{d-1}dr)\}.$$ Thus writing $x=r\theta$, with $r>0$ and $\theta \in \Sphere^{d-1}$ $$\label{eq:Pl} (P_l g)(r\theta)=\sum_{\mu =1}^{\mu(l)} \int_{\Sphere^{d-1}} g(r\omega) Y^{\mu}_l(\theta)\overline{Y}^\mu_l(\omega ) d S_{\omega}.$$ \[l:B\_1approx\] Let $V\geq 0$, $V\in L^{\infty}_{\comp}(\Real^d)$ be a radial function, so that $V(x)=f(\|x\|)$ for some function $f\in L^{\infty}_{\comp}([0,\infty))$. Then for $\sigma>0$ sufficiently large, with $B_1=B_{1,\pm, V}$ the operator defined in (\[eq:B1\]), $$\left\\| \large(\vhalf T(i\sigma)\vhalf -B_{1,\pm ,V}(i\sigma)\large) P_l \right\\| \leq \frac{C}{\sigma^2}\\| \vhalf T(i\sigma) \vhalf P_l\\|$$ where $C$ depends on $V$ but not $\sigma$ or $l$. To simplify the notation, we write $A(i\sigma)=A_{\pm, V}(i\sigma)=I\pm \vhalf R_0(i\sigma)\vhalf$, and note that for $\sigma>0$ sufficiently large, $$\label{eq:Abd} \\|A^{-1}(i\sigma)-I\\|=O(1/\sigma^2),\; \\|A^{-1/2}(i\sigma)-I\\|=O(1/\sigma^2).$$ Now with $B_1$ the operator defined in (\[eq:B1\]), $$\begin{aligned} \label{eq:diffB1} B_1-\vhalf T \vhalf & = (A^{-1/2}-I)\vhalf T \vhalf A^{-1/2} + \vhalf T \vhalf (A^{-1/2}-I).\end{aligned}$$ Because $V$ is radial, multiplication by either $V$ or $\vhalf$ commutes with $P_l$. Since $R_0$ commutes with $P_l$, so do $A$, $A^{-1}$, and $A^{-1/2}$. Thus $$\begin{gathered} \\| (B_1-\vhalf T \vhalf)P_l\\| \\ \leq \\|(A^{-1/2}-I)\\| \\| \vhalf T \vhalf P_l\\| \\| A^{-1/2}\\| + \\|\vhalf T \vhalf P_l\\| \\| (A^{-1/2}-I) \\|.\end{gathered}$$ Thus using (\[eq:Abd\]) we are done. Using the notation of [@ab-st], let $J_\nu$ and $Y_\nu$ denote the Bessel functions of the first and second kinds, respectively, and recall that $H^{(1)}_\nu(z)= J_\nu(z)+iY_{\nu(z)}$. For $l\in \Natural_0$, set $\nu_l= l+(d-2)/2$ and notice that $\nu_l$ is an integer since $d$ is even. We can now expand $R_0(\lambda)$ using spherical harmonics. When $0<\arg \lambda<\pi$ and $g\in L^2(\Real^d)$, $$\label{eq:R0exp} (R_0(\lambda)g)(r\theta)= \sum_{l=0}^{\infty} \sum_{\mu=1}^{\mu(l)} \int_0^\infty \int_{\Sphere^{d-1}} G_{\nu_l}(r,r';\lambda) Y^{\mu}_l( \theta) \overline{Y}^\mu_l(\omega) g(r'\omega) (r')^{d-1}dS_\omega dr'$$ with $$\label{eq:Gnu}G_{\nu_l}(r,r';\lambda)= \left\{ \begin{array}{ll}\frac{\pi}{2i}(r r')^{-(d-2)/2} J_{\nu_l}(\lambda r)H_{\nu_l}^{(1)}(\lambda r'),\; & \text{if } r<r'\\\frac{\pi}{2i}(rr')^{-(d-2)/2} H_{\nu_l}^{(1)}(\lambda r)J_{\nu_l}(\lambda r'), & \text{if}\; r\geq r' \end{array} \right.$$ As noted earlier, for compactly supported, bounded $\chi$, $\chi R_0(\lambda)\chi$ has an analytic continuation to $\Lambda$, and $G_{\nu_l}(r,r';\lambda)$ does as well. Now we use [@ab-st 9.1.35, 9.1.36] to obtain $$J_\nu(e^{i\pi}z)= e^{i\pi \nu}J_{\nu}(z).$$ Specializing [@ab-st 9.1.36] to the case of $\nu$ an integer we have $$Y_{\nu_l}(e^{i\pi}z)= e^{-\nu_l \pi i}(Y_{\nu_l}(z)+2i J_{\nu_l}(z))$$ giving $$H_{\nu_l}^{(1)}(e^{i\pi }z)= e^{i\nu_l \pi}(-J_{\nu_l} (z)+i Y_{\nu_l}(z)).$$ Thus $$\label{eq:gnudiff} \tilde{G}_{\nu_l}(r,r';\lambda){\stackrel{\rm{def}}{=}}G_{\nu_l}(r,r';e^{i\pi} \lambda)-G_{\nu_l}(r,r';\lambda) = i\pi (rr')^{-(d-2)/2}J_{\nu_l}(\lambda r)J_{\nu_l}(\lambda r').$$ Together, (\[eq:R0exp\]) and (\[eq:gnudiff\]) give us an expression for the Schwartz kernel of $R_0(e^{i\pi}\lambda)-R_0(\lambda)$ in spherical coordinates: with $r, \; r'>0$, $\theta \in \Sphere^{d-1}$, $$\begin{gathered} \label{eq:shexp12} \left((R_0(e^{i\pi \lambda})-R_0(\lambda))g\right)(r\theta) \\= \sum_{l=0}^{\infty} \sum_{\mu =1}^{\mu(l)} \int_0^\infty \int_{\Sphere^{d-1}} \tilde{G}_{\nu_l}(r,r';\lambda) Y^{\mu}_l( \theta) \overline{Y}^\mu_l(\omega) g(r'\omega) (r')^{d-1}dS_\omega dr'.\end{gathered}$$ We continue to denote by $P_l$ the operator given in (\[eq:Pl\]). \[l:B1Pl\] Let $B_1$ be the operator defined in (\[eq:B1\]). Let $V_0=\epsilon \chi_a$, where $\epsilon,\; a>0$ and $\chi_a$ is the characteristic function of the ball of radius $a$ and center $0$. Fix a constant $M>3$. Then there is a constant $c>0$ independent of $\sigma$ so that $$\\| B_{1,\pm ,V_0}(i\sigma)P_l\\| \geq c \frac{e^{c\nu_l}}{\nu_l}$$ for all $l\in \Natural$ which satisfy $a\sigma/6>\nu_l> a \sigma /M$ for all sufficiently large $\sigma>0$. Before beginning the proof, we note that the constant $c$ does depend on $\epsilon$ and on $a$. From Lemma \[l:B\_1approx\] it suffices to prove an analogous lower bound for $\\| \vhalf_0 T (i\sigma)\vhalf_0 P_l\\|.$ Recall $iT(i\sigma)= R_0(e^{i\pi }i\sigma)-R_0(i\sigma)$. Set $$\psi_l(r\theta)= \chi_a(r\theta)Y^\mu_l(\theta)r^{-(d-2)/2} J_{\nu_l}(i\sigma r)$$ for any $\mu\in \{ 1,...,\mu(l)\},$ and note that $$\\| \vhalf_0 T \vhalf_0 P_l\\| \geq \frac{\left\| \langle \vhalf_0 T \vhalf_0 \psi_l, \psi_l \rangle \right\| }{\\|\psi_l\\|^2} .$$ By (\[eq:gnudiff\]) and (\[eq:shexp12\]), $$\begin{aligned} \label{eq:lb1} \frac{\|\langle \vhalf_0 T \vhalf_0 \psi_l, \psi_l \rangle \|} {\\|\psi_l\\|^2} & = \frac{ \pi \left( \int_0^a \epsilon^{1/2} \|J_{\nu_l}(i\sigma r)\|^2 r^{-(d-2)}r^{d-1}dr\right)^{2}} { \int_0^a \|J_{\nu_l}(i\sigma r)\|^2 r^{-(d-2)}r^{d-1}dr} \nonumber \\ & = \pi \epsilon \int_0^a \|J_{\nu_l}(i\sigma r)\|^2 r dr \nonumber \\ & \geq \pi \epsilon \int_{a/2}^a \|J_{\nu_l}(i\sigma r)\|^2 r dr.\end{aligned}$$ As in [@ab-st 9.6.3], setting $$I_\nu(z){\stackrel{\rm{def}}{=}}e^{-\nu \pi i/2}J_{\nu}(z e^{i\pi/2}),\; -\pi <\arg z \leq \pi/2,$$ from [@ab-st 9.7.7] there is a constant $c>0$ so that for $\nu$ sufficiently large $$\|I_{\nu}(\nu s)\| \geq c \frac{e^{c\nu}}{\sqrt{\nu}},\; 3\leq s \leq M.$$ Here and below we denote by $c$ a positive constant, independent of $\nu$, $l$, and $\sigma$, which may change from line to line. Now we use that $\| J_{\nu_l}(i\sigma z)\|= \| I_{\nu_l}(\sigma z)\|$ and apply these to (\[eq:lb1\]). We find that if $3\leq \sigma r/\nu_l\leq M$ for all $r$ with $a/2\leq r \leq a$, then $$\left\| \langle \vhalf_0 T \vhalf_0 \psi_l, \psi_l \rangle \right\| \geq c \int_{a/2}^a \frac{e^{2\nu_l c}}{\nu_l} dr \geq c \frac{e^{2\nu_l c}}{\nu_l}$$ for all sufficiently large $\sigma$. Thus, this holds for $l$ satisfying $a\sigma/6>\nu_l> a \sigma /M$ if $\sigma $ is sufficiently large, providing a lower bound on $\\| \vhalf_0 T \vhalf_0 P_l\\|$, and thus on $\\|B_{1,\pm V_0}(i\sigma)P_l\\|$. \[l:spcaselb\] Let $V_0=\epsilon \chi_a$, where $\epsilon, \; a>0$ and $\chi_a$ is the characteristic function of the ball of radius $a$ and center $0$. Then for $m_0\not =0$, $m_0\in \Integers$, there is a $c>0$ so that for $\sigma>0$ sufficiently large $$F_{m_0,\pm V_0}(i\sigma)\geq c\exp(c \sigma^d).$$ The constant $c$ depends on $a,\;\epsilon$, and $m_0$. Recall that $$\|F_{m_0,\pm V_0}(i\sigma)\|= \|\det( I\pm i m_0 B_{1, \pm, V_0}(i\sigma))\|$$ and that for sufficiently large $\sigma>0$ $B_1(i\sigma)$ is a self-adjoint operator. Thus for sufficiently large $\sigma$ $$\label{eq:Fmeigenvalues} \|F_{m_0,\pm V_0}(i\sigma)\|= \prod_{j=1}^{\infty} \sqrt{1+m_0^2 \lambda_j^2}$$ where $\lambda_j$ are the nonzero eigenvalues of $B_{1,\pm, V_0}(i\sigma)$. The $\lambda_j$ of course depend on $\sigma$, but we omit this in our notation. A decomposition of $B_{1,\pm, V_0}$ using spherical harmonics shows that $B_{1,\pm, V_0}$ has eigenvalue $\\| B_{1,\pm, V_0} P_l\\|$ with multiplicity (at least) $\mu(l)$. Thus using (\[eq:Fmeigenvalues\]) and the fact that $\lambda_j^2> 0$, we get $$\|F_{m_0}(i\sigma)\|^2 \geq \prod_{l=1}^{\infty} (1+m_0^2\\| B_{1,\pm, V_0} P_l\\|^2)^{\mu(l)}$$ for sufficiently large $\sigma$. From Lemma \[l:B1Pl\], we see $$\begin{aligned} \|F_{m_0}(i\sigma)\|^2 & \geq \prod_{a\sigma/6>\nu_l> a \sigma /M} (1+ cm_{0}^2 \frac{e^{c\nu_l}}{\nu_l^2})^{\mu(l)}\\ & =\exp\left( \sum_{a\sigma/6>\nu_l> a \sigma /M}\mu(l) \log \left(1+ c m_0^2 \frac{e^{c\nu_l}}{\nu_l^2}\right)\right) \\ & \geq \exp\left( \sum_{a\sigma/6-(d-2)/2>l> a \sigma /M -(d-2)/2} \mu(l)\left(cl -c(d-2)/2+\log (c /\nu_l^2) \right) \right)\end{aligned}$$ Now for $l$ sufficiently large, $\mu(l) \geq l^{d-2}/(d-2)!$ so we get $$\|F_m(i\sigma)\|^2 \geq \exp( c \sigma^d -C)$$ for some constants $C$ and $c>0$ for all sufficiently large $\sigma$. [Proof of Proposition \[p:lb\].]{} We are now ready to give the proof of Proposition \[p:lb\]. Since if $W$ is a translate of $V$, $F_{m,\pm,V}= F_{m,\pm, W}$, we may assume $V$ can be bounded below by $V_0=\epsilon \chi_{B_a}$, where $\chi_{B_a}$ is the characteristic function of the ball of radius $a>0$ and center at the origin. Then using Lemmas \[l:mono\] and \[l:spcaselb\] proves the proposition immediately. Proof of Theorem \[thm:lbd\] ============================ Let $V\in L^{\infty}_{\comp}(\Real^d),$ $V\geq 0$. We continue to assume $d$ is even and to use the function $$F_m(\lambda)=F_{m,\pm V}(\lambda)=\det( I\pm im (1\pm \vhalf R_0(\lambda)\vhalf )^{-1} \vhalf T(\lambda) \vhalf)$$ defined first by (\[eq:FmV\]). Note that since $(I \pm \vhalf R_0(\lambda)\vhalf )^{-1}$ is a meromorphic function on $\Lambda$, $F_{m,\pm V} (\lambda) $ is meromorphic on $\Lambda$. We shall be most interested in the behavior of $F_{m,\pm V}(\lambda)$ in $\overline{\Lambda}_0$, since the zeros of $F_{m,\pm V}$ in $\Lambda_0$ correspond to the poles of $R_{\pm V}$ in $\Lambda_m$. In the proof of Theorem \[thm:lbd\] we shall apply Proposition \[p:complex\] to a function obtained by multiplying $F_{m, \pm V}$ by a rational function. Thus we begin this section by checking properties of $F_{m,\pm V}$. \[l:gprop1\] The function $F_{m,\pm V}(\lambda)$ has only finitely many poles in $\{ \lambda \in \Lambda: 0\leq \arg \Lambda \leq \pi\}$ and only finitely many zeros with argument $0$ or $\pi$. We recall first the well-known estimate $$\label{eq:phyregbd} \\| \vhalf R_0(\lambda) \vhalf \\|\leq C/\|\lambda\| \; \text{for}\; \lambda \in \Lambda, \; 0\leq \arg \lambda \leq \pi$$ (e.g. [@agmon; @vodeveven; @vodev2]). Thus for $\|\lambda\|\geq 2/C$, $I \pm \vhalf R_0(\lambda)\vhalf $ is invertible, with norm of the inverse bounded by $2$. Since the function $F_{m, \pm V}$ cannot have a pole at $\lambda_0$ unless $(I\pm \vhalf R_0(\lambda)\vhalf )^{-1}$ has a pole at $\lambda_0$, we see $F_{m,\pm V}(\lambda)$ has no poles in the region $\{ \lambda \in \overline{\Lambda}_0,\; \|\lambda\|\geq r_0\}$ for some constant $r_0$ depending on $V$. Moreover, from (\[eq:phyregbd\]) $\\|\vhalf T(\lambda) \vhalf \\| \leq C/\|\lambda\|$ for $\lambda \in \partial \overline{\Lambda}_0.$ Thus, there is an $r_0\geq 0$ so that $F_{m,\pm V}(\lambda)$ has no zeros in $\{ \lambda \in \partial \overline{\Lambda}_0,\; \|\lambda \|\geq r_0\}$. The bounds of Vodev [@vodeveven; @vodev2] ensure that there are only finitely many poles of $R_{\pm V}(\lambda)$ in $\{ \lambda \in \overline{\Lambda_{m}} : \|\lambda\|\leq r \}$ for any $r$. Since $F_{m,\pm V}$ has a pole at $ \lambda \in \overline{\Lambda_0}$ only if $R_{\pm V}$ has a pole there, and has a zero at $z\in \partial\overline{\Lambda_0}$ only if $R_{\pm V}$ has a pole at $e^{im \pi} \lambda $, this finishes the proof of the claim. \[l:bdonreal\] Let $t\in \Lambda $ have $\arg t=0 $ or $\arg t=\pi$. Then there are constants $C, \; r_0 >0$ depending on $V$ and $m$ so that $$\left\| \frac{\frac{d}{dt} F_{m,\pm V}(t)}{F_{m, \pm V}(t)} \right\| \leq C\|t\|^{d-2}\; \text{for}\; \|t\|\geq r_0.$$ Note that $$\label{eq:detderiv} \frac{\frac{d}{dt}F_{m,\pm V}(t)}{F_{m, \pm V}(t)} = \tr \left( \pm i m (I\pm im W(t))^{-1}\frac{d}{dt}W(t) \right)$$ where $$W(t)=W_{\pm V}(t)= (I\pm V^{1/2}R_0(t) V^{1/2})^{-1} V^{1/2}T(t) V^{1/2}.$$ Using (\[eq:phyregbd\]) we see that that there is an $r_0 >0$ so that $$\label{eq:invbd} \\| (I \pm \vhalf R_0(t)\vhalf)^{-1} \\|\leq 2 \; \text{ for $\|t\|>r_0$}.$$ For the values of $t$ in question (on the boundary of the physical region), for any $\chi \in C_c^{\infty}(\Real^d)$ and any $j\in \Natural_0$ there are constants $C_j$ depending on $\chi$ so that $$\left\\| \frac{d^j}{dt^j} \chi R_0(t) \chi \right\\| \leq C_j \|t\|^{-1-j}, \; \|t\|\geq 1,$$ see e.g. [@j-k Section 8] or [@k-k Section 16]. This implies that for $\|t\|$ sufficiently large with $\arg t=0,\; \pi$, $\\| \frac{d^j}{dt^j}W(t)\\|\leq C_j$, $j=0,\; 1$, for some new constant $C_j$ depending on $V$. Now we use an argument as in [@froeseodd Lemma 3.3] to bound $\\| W(t)\\|_1$ and $\\| \frac{d}{dt}W(t)\\|_1$, where $\\| \cdot \\|_1$ is the trace class norm. We write, for $\chi \in L^{\infty}_{\comp}(\Real^d)$ $$\label{eq:Tascomp} \chi T(\lambda) \chi = \alpha_d \lambda^{d-2} {\mathbb E}_\chi^t(e^{i\pi}\lambda) {\mathbb E}_\chi(\lambda)$$ where $${\mathbb E}_\chi( \lambda):L^2(\Real^d)\rightarrow L^2(\Sphere^{d-1}),\; {\mathbb E}_{\chi}(\lambda)(\theta,x) =\chi(x) e^{i\lambda x \cdot \theta},\; x\in \Real^d,\; \theta \in \Sphere^{d-1}.$$ Then, just as in [@froeseodd], we note that with $\\| \cdot \\|_2$ denoting the Hilbert-Schmidt norm, $$\\| {\mathbb E}_\chi( t ) \\|^2_2=\int_{\Sphere^{d-1}} \int_{\Real^d}\|e^{it\omega\cdot x} \chi(x)\|^2dxd\omega \leq C_\chi, \; \text{for}\; (\arg t)/\pi \in \Integers$$ and $$\left\\|\frac{d}{dt} {\mathbb E}_\chi( t ) \right\\|^2_2 = \int_{\Sphere^{d-1}} \int_{\Real^d}\left\|i (\omega\cdot x)e^{it\omega\cdot x} \chi(x)\right\|^2dxd\omega \leq C_\chi, \; \text{for}\; (\arg t)/\pi \in \Integers.$$ The same estimate holds for $\\| {\mathbb E}_\chi^t(e^{i\pi}t)\\|_2^2$ and $\\| \frac{d}{dt}{\mathbb E}_\chi^t(e^{i\pi}t)\\|_2^2$. Putting this all together and using that $\\| AB\\|_1\leq \\|A\\|_2\\| B\\|_2$, we see that $$\left \\| \frac{d^j}{dt^j} W(t) \right\\|_1 \leq C,\; \text{for}\; j=0,\; 1.$$ Thus $$\begin{aligned} \left\| \frac{\frac{d}{dt}F_{m,\pm V}(t)}{F_{m, \pm V}(t)}\right\| & = \left\| \tr \left( \pm i m (I\pm im W(t))^{-1}\frac{d}{dt}W(t) \right)\right\|\\ & \leq \left\\| m (I\pm im W(t))^{-1}\frac{d}{dt}W(t)\right\\|_1 \leq C\|t\|^{d-2}\end{aligned}$$ when $\|t\|$ is sufficiently large. The next lemma gives a bound on $F_{m,\pm V}(z)$, $z\in \Lambda_0$, which is of a type which has been repeatedly used in proofs of upper bounds on the number of resonances. Closely related results can be found in [@melrosepb; @zworskiodd; @froeseodd], among others. We include the proof for the convenience of the reader, although it is essentially a minor modification of arguments used in, for example, [@zworskiodd; @froeseodd] to, in the odd-dimensional case, bound something like the determinant of the scattering matrix in the physical half-plane. There are constants $C$, $r_0>0$ depending on $V$ and $m$ so that $$\|F_{m,\pm, V}(\lambda)\|\leq C \exp(C\|\lambda\|^d),\; \text{for all } \lambda \in \overline{\Lambda_0}, \; \|\lambda\|>r_0.$$ \[l:upperbd\] Using (\[eq:Tascomp\]) and that $\det(I+AB)=\det(I+BA)$ when both $AB$ and $BA$ are trace class, $$F_{m,\pm,V}(\lambda)= \det( I+ K(\lambda))$$ where $K(\lambda):L^2(\Sphere^d)\rightarrow L^2(\Sphere^d)$ is given by $$K(\lambda)= \pm i m \alpha_d \lambda^{d-2} {\mathbb E}_{\vhalf}(\lambda) (I\pm \vhalf R_0(\lambda)\vhalf)^{-1} {\mathbb E}_{\vhalf}^t(e^{i\pi}\lambda).$$ Choose $r_0\geq 0$ so that $$\\| (I\pm \vhalf R_0(\lambda)\vhalf)^{-1}\\| \leq 2 \; \text{for }\; \lambda\in \Lambda_0,\; \|\lambda\|\geq r_0.$$ By slight abuse of notation, we denote the Schwartz kernel of $K$ by $K$ as well. Then there is some constant $C$ so that for each $j\in \Natural$, $$\|\Delta_{\Sphere^{d-1},\theta}^j K(\lambda)(\theta,\omega) \|\leq C^{2j+1}(\|\lambda\|^{2j}+ (2j)!)e^{C\|\lambda\|}\; \text{for}\; \lambda\in \overline{\Lambda_0},\; \|\lambda\|\geq r_0$$ since $$\|\Delta_{\Sphere^{d-1}}^k e^{i \lambda x\cdot \theta} \vhalf (x) \| \leq C^k (\|\lambda\|^{2k} + (2k)!) )e^{C\|\lambda\|}$$ and $\| (I\pm \vhalf R_0(\lambda)\vhalf)^{-1} {\mathbb E}_{\vhalf}(e^{i\pi}\lambda)^t \| \leq C \exp( C \|\lambda\|), $ when $\|\lambda\|\geq r_0$. Thus by [@zworskiodd Proposition 2], $$\|\det (I+K(\lambda))\|\leq C' e^{C'\|\lambda\|^{d}},\; \lambda\in \overline{\Lambda_0},\; \|\lambda\|>r_0.$$ We are now ready to give the proof of Theorem \[thm:lbd\]. The proof is by contradiction. So suppose for some fixed potential $V$ satisfying the hypotheses of the theorem and for some value of $m\in \Integers \setminus \{ 0\}$ and for choice of sign (positive or negative) $$\label{eq:contassumpt} \lim \sup_{r \rightarrow \infty} \frac{\log n_{m, \pm V}(r)}{\log r}<d.$$ We work with this fixed value of $m$ and fixed choice of sign for the remainder of this proof. For this choice of $m$ and sign consider the function $$F_{m,\pm V}(\lambda)= \det( I\pm im (1\pm \vhalf R_0(\lambda)\vhalf )^{-1} \vhalf T(\lambda) \vhalf).$$ We denote by $\tilde{n}(r) $ the number of zeros, counted with multiplicity, of $F_{m,\pm V}$ in $\Lambda_0$ of norm at most $r$. The assumption (\[eq:contassumpt\]) means that there is a constant $d'<d$ so that $n_{m,\pm V}(r) =O(r^{d'})$ for $r\rightarrow \infty$. Since with at most finitely many exceptions the zeros of $F_{m,\pm V}$ in $\Lambda_0$ correspond, with multiplicity, to the poles of $R_{\pm V}$ in $\Lambda_m$ (see Section \[s:defF\]), $\tilde{n}(r)=n_{m,\pm V}(r)+O(1)\leq C(1+r^{d'})$ for some constant $C$. We identify $\Lambda_0$ with the upper half plane and use the variable $z$ there. Thus we may think of $F_{m,\pm V}$ as function meromorphic in a neighborhood of $${\Omega}= \{ z \in \Complex: \|z\|\geq 1,\; 0\leq \arg z \leq \pi\}.$$ Let $a_1,...,a_{m_p}$ be the poles of $F_{m,\pm V}$ in ${\Omega}$, and let $b_1,...,b_{m_z}$ be the zeros of $ g$ in $\partial {\Omega}$, in both cases repeated according to multiplicity. Recall we know there are only finitely many by Lemma \[l:gprop1\]. Now set $$h(z){\stackrel{\rm{def}}{=}}\frac{\prod_{j=1}^{m_p}(z-a_j) }{\prod _{j=1}^{m_z}(z-b_j)} F_{m,\pm V}(z).$$ If there are no poles or no real zeros, the corresponding product is omitted. By applying Lemmas \[l:bdonreal\] and \[l:upperbd\], we see that $h$ satisfies the hypotheses of Proposition \[p:complex\] with $\rho=\max(d',d-1+\epsilon)$ for any $\epsilon>0$. Thus for some constant $C$, $\|F_{m,\pm V}(z)\| \leq C\exp(C\|z\|^\rho)$ for $z\in {\Omega}$ and $\rho<d$. But this contradicts Proposition \[p:lb\]. [99]{} S. Agmon, [Spectral properties of Schrödinger operators and scattering theory]{}. Ann. Scuola Norm. Sup. Pisa Cl. Sci. (4) [2]{} (1975), no. 2, 151-218. A. Autin, [Isoresonant complex-valued potentials and symmetries]{}. Canad. J. Math. [63]{} (2011), no. 4, 721-754. L.-H. Chen, [A sub-logarithmic lower bound for resonance counting function in two-dimensional potential scattering.]{} Rep. Math. Phys. [65]{} (2010), no. 2, 157-164. T. Christiansen, [Several complex variables and the distribution of resonances in potential scattering]{}. Comm. Math. Phys. [259]{} (2005), no. 3, 711-728. T. Christiansen, [Schrödinger operators with complex-valued potentials and no resonances.* ]{} Duke Math. J. [133]{} (2006), no. 2, 313-323. T. Christiansen, [Isophasal, isopolar, and isospectral Schrödinger operators and elementary complex analysis.]{} Amer. J. Math. [130]{} (2008), no. 1, 49-58. T.J. Christiansen and P.D. Hislop, [Maximal order of growth for the resonance counting functions for generic potentials in even dimensions.]{} Indiana Univ. Math. J. [59]{} (2010), no. 2, 621-660. T.J. Christiansen and P.D. Hislop, [Some remarks on resonances in even-dimensional Euclidean scattering]{}. Preprint: arXiv:1307.5822 J. Conway, Functions of one complex variable. Second edition. Graduate Texts in Mathematics, 11. Springer-Verlag, New York-Berlin, 1978. T.-C. Dinh and D.-V. Vu, [Asymptotic number of scattering resonances for generic Schrödinger operators]{}, preprint arXiv 1207.4273. R. Froese, [Asymptotic distribution of resonances in one dimension]{}, J. Differential Equations [137]{} (1997), no. 2, 251–272. R. Froese, [Upper bounds for the resonance counting function of Schrödinger operators in odd dimensions]{}, Canad. J. Math. [50]{} (1998), no. 3, 538–546. A. Intissar, [A polynomial bound on the number of the scattering poles for a potential in even-dimensional spaces $\Real^n$.]{} Comm. Partial Differential Equations [11]{} (1986), no. 4, 367-396. A. Jensen and T. Kato, [Spectral properties of Schrödinger operators and time-decay of the wave functions.]{} Duke Math. J. [46]{} (1979), no. 3, 583-611. A. Komech and E. Kopylova, Dispersion decay and scattering theory. John Wiley & Sons, Inc., Hoboken, NJ, 2012. P.D. Lax and R. S. Phillips, [Decaying modes for the wave equation in the exterior of an obstacle.]{} Comm. Pure Appl. Math. [22]{} (1969) 737-787. B. Ja. Levin, [Distribution of zeros of entire functions]{}, American Mathematical Society, Providence, R.I. 1964. R.B. Melrose, [Polynomial bound on the number of scattering poles]{}. J. Funct. Anal. [53]{} (1983), no. 3, 287-303. R.B. Melrose, Geometric scattering theory. Stanford Lectures. Cambridge University Press, Cambridge, 1995. F.W.J. Olver, [Bessel functions of integer order]{}, in Handbook of mathematical functions with formulas, graphs, and mathematical tables. Ed. M. Abramowitz and I. Stegun. National Bureau of Standards Applied Mathematics Series, 55. Government Printing Office, Washington, DC, 1964. 355-434. T. Regge, [Analytic properties of the scattering matrix]{}, Nuovo Cimento [8]{} (5), (1958), 671–679. A. Sá Barreto, [Lower bounds for the number of resonances in even dimensional potential scattering]{}, J. Funct. Anal. [169]{} (1999), 314–323. A. Sá Barreto and S.-H. Tang, [Existence of resonances in even dimensional potential scattering.]{} Comm. Partial Differential Equations [25]{} (2000), no. 5-6, 1143-1151. B. Simon, [ Resonances in one dimension and Fredholm determinants]{}, J. Funct. Anal. [178]{} (2000), no. 2, 396–420. B. Simon, Trace ideals and their applications. Second edition. Mathematical Surveys and Monographs, 120. American Mathematical Society, Providence, RI, 2005. A. Vasy, [Scattering poles for negative potentials]{}. Comm. Partial Differential Equations [22]{} (1997), no. 1-2, 185–194. G. Vodev, [Sharp bounds on the number of scattering poles in even-dimensional spaces.]{} Duke Math. J. [74]{} (1994), no. 1, 1-17. G. Vodev, [Sharp bounds on the number of scattering poles in the two-dimensional case.]{} Math. Nachr. [170]{} (1994), 287-297. M. Zworski, [Distribution of poles for scattering on the real line. ]{} J. Funct. Anal. [73]{} (1987), no. 2, 277-296. M. Zworski, [Sharp polynomial bounds on the number of scattering poles of radial potentials]{}, J. Funct. Anal. [82]{} (1989), 370-403. M. Zworski, [Sharp polynomial bounds on the number of scattering poles]{}. Duke Math. J. [59]{} no. 2 (1989), 311-323.	{ "pile_set_name": "ArXiv" }
--- abstract: 'In the $\Lambda$CDM paradigm the Galactic stellar halo is predicted to harbor the accreted debris of smaller systems. To identify these systems, the H3 Spectroscopic Survey, combined with Gaia, is gathering 6D phase-space and chemical information in the distant Galaxy. Here we present a comprehensive inventory of structure within 50 kpc from the Galactic center using a sample of 5684 giants at $\|b\|>40^{\circ}$ and $\|Z\|>2$ kpc. We identify known structures including the high-$\alpha$ disk, the in-situ halo (disk stars heated to eccentric orbits), Sagittarius (Sgr), Gaia-Sausage-Enceladus (GSE), the Helmi Streams, Sequoia, and Thamnos. Additionally, we identify the following new structures: (i) Aleph (\[Fe/H\]$=-0.5$), a low eccentricity structure that rises a surprising 10 kpc off the plane, (ii, iii) Arjuna (\[Fe/H\]$=-1.2$) and I’itoi (\[Fe/H\]$<-2$), which comprise the high-energy retrograde halo along with Sequoia, and (iv) Wukong (\[Fe/H\]$=-1.6$), a prograde phase-space overdensity chemically distinct from GSE. For each structure we provide \[Fe/H\], \[$\alpha$/Fe\], and orbital parameters. Stars born within the Galaxy are a major component at $\|Z\|\sim$2 kpc ($\approx$60$\%$), but their relative fraction declines sharply to $\lesssim$5$\%$ past 15 kpc. Beyond 15 kpc, $>$80$\%$ of the halo is built by two massive ($M_{\star}\sim10^{8}-10^{9}M_{\odot}$) accreted dwarfs: GSE (\[Fe/H\]$=-1.2$) within 25 kpc, and Sgr (\[Fe/H\]$=-1.0$) beyond 25 kpc. This explains the relatively high overall metallicity of the halo (\[Fe/H\]$\approx-1.2$). We attribute $\gtrsim$95$\%$ of the sample to one of the listed structures, pointing to a halo built entirely from accreted dwarfs and heating of the disk.' author: - 'Rohan P. Naidu' - Charlie Conroy - Ana Bonaca - 'Benjamin D. Johnson' - 'Yuan-Sen Ting (丁源森)' - Nelson Caldwell - Dennis Zaritsky - 'Phillip A. Cargile' bibliography: - 'MasterBiblio.bib' title: Evidence from the H3 Survey that the Stellar Halo is Entirely Comprised of Substructure --- [UTF8]{}[gbsn]{} Introduction {#sec:introduction} ============ The Milky Way’s stellar halo comprises only $\sim1\%$ of its stellar mass [e.g., @Deason19; @Mackereth20], and yet it is an object of intense interest because it acts as a time capsule, preserving memory of the Galaxy’s assembly history with high fidelity. As early as @Woolley57 it was realized that “the time of relaxation of stellar motions in this part of the galaxy is at least $10^{12}$ years, whereas the stars themselves have not existed in their present form for much more than $10^{10}$ years". In detail, halo stars belonging to the same structure, even when they are scattered across the sky, retain similar coordinates in their integrals of motion (e.g., energy, angular momenta, actions). Further, stars belonging to the same structure share similar chemical abundance patterns [e.g., @Freeman02; @Venn04; @Lee15]. This expected clustering of halo stars in both integrals of motion and chemistry opens the door to “reconstruct the galactic past" [@Eggen62]. Thanks to large stellar spectroscopic surveys, e.g., RAVE [@RAVE], SEGUE [@SEGUE], LAMOST [@LAMOST], GALAH [@GALAH], APOGEE [@APOGEE], and the Gaia mission [@gaia], integrals of motion and chemical abundances have become available for millions of stars in the solar neighborhood. Strikingly, more than half of the \[Fe/H\]$<-1$ stars in the local halo[^1] appear to originate from a single system, the accreted Gaia-Sausage-Enceladus (GSE) dwarf galaxy [e.g., @Belokurov18; @Koppelman18; @Myeong18; @Haywood18; @Helmi18; @Mackereth19]. However, the most substantive component ($\gtrsim50\%$) of the local halo, the “in-situ halo"/“Splash", likely arose from the heating of the primordial high-$\alpha$ disk by early mergers [e.g., @Bonaca17; @Bonaca20; @Haywood18; @DiMatteo19; @Belokurov20]. While the local halo has provided these vital insights into the Galaxy’s assembly, a complete census of accretion events requires going beyond the solar neighborhood. A number of simulations show that debris from minor mergers, higher-mass but recently accreted galaxies, and galaxies accreted along particular inclinations (e.g., at high angular momentum) are under-represented in the local halo [e.g., @bj05_1; @Amorisco17; @Fattahi20; @Pfeffer20]. As a consequence, studies that rely on local high-energy orbits to deduce the nature of the distant halo are biased against these populations. The disrupting Sagittarius dwarf galaxy [e.g., @Ibata94; @Majewski03] is a prime example of a massive structure that is completely absent from the local halo. Fully characterizing the global extent of structures discovered in local samples also demands pushing farther into the halo. Debris from low-mass structures like Thamnos ($M_{\star}<5\times10^{6}M_{\odot}$, @Koppelman19), which is barely discernible in local samples, might be more apparent at larger distances due to “apocenter pile-up" [@Deason18] or due to higher contrast once the density of GSE and disk-like stars falls off. Studying massive accreted structures like GSE and the Helmi Streams [@Helmi99] – e.g., the presence/absence of metallicity gradients, robust stellar masses from star counts, if they even are a single contiguous structure – will also become more tractable with samples spanning their full extent. Ranging beyond the local halo is also necessary to settle long-standing debates about the origin and nature of the halo. Is the halo largely formed in-situ or ex-situ (e.g., @Eggen62 vs. @Searle78)? To what radius does the recently discovered in-situ component of the halo dominate the halo mass function? Some simulations [e.g., @Monachesi19] show disk stars, heated by mergers, comprising $\sim20\%$ of the halo even beyond 50 kpc. Consequently, the extent and relative fraction of the in-situ halo should provide an independent constraint on the Galaxy’s accretion history [e.g., @Zolotov09; @Purcell10]. More generally, the fraction of in-situ halo stars (not only the heated disk, but also stars formed from stripped gas from satellites or through cosmological accretion) varies widely across simulations, ranging from negligible to comparable to the accreted mass, and could act as a sensitive constraint on sub-grid physics like star-formation and feedback prescriptions [e.g., @Cooper15; @Pillepich15; @Fattahi20]. Intertwined questions about the ex-situ component persist. Is it built from a handful of massive galaxies ($M_{\star}\sim10^{8}-10^{9} M_{\odot}$), or a multitude of metal-poor ultra-faints ($M_{\star}\lesssim10^{5} M_{\odot}$) [e.g., @bj05_2; @Frebel10; @Deason15; @Deason16; @dsouza18]? How does the metallicity of the halo change as a function of radius? Does the halo transition into a metal-poor (\[Fe/H\]$\sim-2.2$), spherical structure beyond 20 kpc as predicted by local energetic orbits in the popular “dual halo" scenario [e.g., @Carollo07; @Carollo10; @Beers12]? Are different accreted galaxies responsible for this shift? Or could this be due to a smooth component from dissolved ancient globular clusters [e.g., @Martell11; @Carretta16; @Koch19]? Is the traditional conception of the distant halo as a metal-poor structure a selection artifact, arising from color cuts designed to avoid the disk, and from metallicity-biased standard candles [e.g., @Conroy19b]? Studying the stellar halo also enables new forms of near-field cosmology. For instance, accreted debris from $M_{\star} = 10^{6}-10^{7} M_{\odot}$ galaxies gives us access on a star-by-star level to high-redshift galaxies whose evolution was frozen at the time of infall. This provides a complementary view on issues of the distant universe – for instance, the evolution of the ISM [e.g., @Steidel16; @Bian20], the interplay between reionization and low-mass galaxies [e.g., @Barkana99; @Naidu20], the shape of high-$z$ star-formation histories [e.g., @Carnall19; @Leja19] – at a resolution and mass-limit even beyond the reach of upcoming ELTs (Extremely Large Telescopes) and the James Webb Space Telescope [e.g., @Boylan-Kolchin15; @Boylan-Kolchin16; @Weisz14]. Previous efforts to directly probe the distant halo have had to overcome the challenge of targeting rare, distant stars without the benefit of Gaia parallaxes to filter out nearby contaminants. One common solution has been to use color cuts that implicitly or explicitly select for low-metallicities to avoid the disk [e.g., @Chiba00; @Carollo07; @Ivezic08; @Sesar11; @Xue15; @Zuo17]. Another common choice is to rely on rare, standardizable candles like RR Lyrae and blue horizontal branch stars (BHBs) that are inherently metal-poor and more abundant in older populations [e.g., @Deason11; @Kafle12; @Janesh16; @Cohen17; @Iorio19]. Studies based on these tracers have collectively shown the distant Galaxy to display a high degree of substructure, which has been interpreted as support for an accretion origin of the halo [e.g., @Bell08; @Starkenburg09; @Xue11; @Schlaufman12; @Deason18]. In order to make further progress, and ask more detailed questions – which accreted structure dominates at what radius? how far does the in-situ halo extend? what is the mass function of accreted material? – we require a homogeneously selected, metallicity-unbiased sample with full 6D phase-space coordinates, chemical information, and an easily interpretable selection function. The H3 (“Hectochelle in the Halo at High Resolution") Survey [@Conroy19] is fulfilling this need. H3 is a spectroscopic survey of 200,000 stars in high-latitude fields designed to study the distant Galaxy. A defining feature of H3 is a simple, Gaia-based selection function (parallaxes implying $d_{\rm{helio}}>2$ kpc) that, critically, is unbiased in metallicity. With this survey we aim to search for new structure in the distant halo, trace known structures out to their apocenters, clarify long-standing debates about the nature of the halo, and explore promising avenues for near-field cosmology. ![image](Fig1.pdf){width="\linewidth"} ![Correcting for the magnitude selection in the H3 selection function. Top: An \[Fe/H\]$=-1$, 10 Gyr isochrone at different distances, with giants ($\log{g}<3.5$) highlighted with solid lines. At $d_{\rm{helio}}\sim4-35$ kpc the silver band representing the survey magnitude limit ($15<r<18$) almost completely contains the sections of the red giant branch which have a high number density. Therefore, the stars at these distances require little correction for the magnitude limit. Bottom: Correction weights as a function of heliocentric distance (dashed line) overplotted on the distance distribution for stars in this work. The weights are derived using isochrones and assuming a @Kroupa01 IMF (see §\[subsec:magcorr\]), and are remarkably flat for the bulk of the sample. The rise at 35 kpc coincides with the red clump moving out of our magnitude range.[]{data-label="fig:magcorr"}](Fig2.pdf){width="0.9\linewidth"} ![image](Fig3.pdf){width="\linewidth"} In this work we present a census of substructure, previously known and unknown, out to 50 kpc and link the results to the questions outlined in this section. In §\[sec:data\_methods\] we provide details of H3 pertinent to this study (§\[subsec:H3\]), outline how we compute dynamical quantities (§\[subsec:pot\]), and correct for the survey selection function (§\[subsec:magcorr\]). In §\[subsec:results\_overview\] we present an overview of our sample in integrals of motion and chemistry, revealing a high degree of substructure. §\[subsec:inventory\] forms the bulk of the paper – here we identify and define individual structures, and remark on their chemodynamical properties. §\[subsec:structuresummary\] provides a synopsis of all the structures identified. In §\[sec:discussion\] we discuss the implications of the inventory – we chart the relative fractions of structures with distance (§\[subsec:relfrac\]), interpret what this means for the origin of the halo (§\[subsec:origin\]), evaluate the net rotation of the halo (§\[subsec:protretro\]), dissect the halo in chemical space (§\[subsec:chemistry\]), and discuss caveats (§\[subsec:caveats\]). A summary of our results is provided in §\[sec:summary\]. To describe central values of distributions we generally report the median, along with 16th and 84th percentiles. We use $\langle x\rangle$ to denote the mean of the quantity $x$, and report the corresponding error on the mean as 16th and 84th percentiles estimated via bootstrapping. We use $r_{\rm{gal}}$ to denote 3D Galactocentric distance, $R_{\rm{gal}}$ to denote axial distance in Galactocentric cylindrical coordinates, $Z_{\rm{gal}}$ to denote distance from the plane, and $d_{\rm{helio}}$ to refer to 3D heliocentric distance. We use $V_{r}$, $V_{\rm{\phi}}$, $V_{\rm{\theta}}$ to represent velocities in a right-handed spherical coordinate system with origin at the Galactic center. That is, prograde stars have negative $V_{\rm{\phi}}$ and $L_{\rm{z}}$. In the context of photometric magnitudes“$r$" refers to the Pan-STARRS $r$-band [@psack1; @psack2] that is used in the H3 selection function. Magnitudes are in the AB system [@Oke83]. When converting between redshifts and ages we use a cosmology with $\Omega_M=0.3, \Omega_\Lambda=0.7, H_0=70$ km s$^{-1}$ Mpc$^{-1}$, i.e., $h=0.7$. Unless mentioned otherwise, total orbital energy ($E_{\rm{tot}}$) is always reported in units of $10^{5}\ \rm{km^{2}\ s^{-2}}$ and angular momenta ($L_{\rm{x}}$, $L_{\rm{y}}$, $L_{\rm{z}}$) are reported in units of $10^{3}\ \rm{kpc}\ \rm{km\ s^{-1}}$. ![image](Fig4.pdf){width=".95\linewidth"} Data and Methods {#sec:data_methods} ================ The H3 Survey {#subsec:H3} ------------- H3 [@Conroy19] is the first spectroscopic survey to leverage Gaia parallaxes, $\pi$, in its selection of targets. The selection function of the primary sample is composed of the following conditions: (i) $15<r<18$, (ii) $\pi-2\sigma_{\pi}<0.5$, implying $d_{\rm{helio}}>2$ kpc, (iii) $\|b\|>30^{\circ}$, to avoid the disk, and (iv) Dec.$>-20^{\circ}$, observable from the MMT located in Arizona, USA. This simple selection function ensures a view of the halo that is free from metallicity biases due to color cuts or metal-poor stellar tracers (e.g., BHBs, RR Lyrae). While H3 will eventually cover $\|b\|>30^{\circ}$ and the survey selection function requires Gaia parallaxes consistent with $d_{\rm{helio}}>2$ kpc, the data presented in this paper is at $\|b\|>40^{\circ}$, and also limited to $d_{\rm{helio}}>3$ kpc (for reasons outlined in §\[subsec:magcorr\]). Complementing the primary selection, we target a small number ($\approx6\%$ of the final sample used in this work) of color-selected K giants ($\approx5\%$, cuts from @Conroy18), BHBs ($\approx1\%$, cuts from @Deason14), and RR Lyrae (7 in number, sourced from @Sesar17rrl). We take care to appropriately weight these specially targeted stars while accounting for the selection function in §\[subsec:magcorr\]. Inspection of stellar parameters of the BHBs reveals that while their distances and radial velocities are robust, their abundances are not reliable, so they are omitted from plots featuring \[Fe/H\] and \[$\alpha$/Fe\]. The key outputs from the survey are radial velocities precise to $\lesssim$1 km $\rm{s^{-1}}$, \[Fe/H\] and \[$\alpha$/Fe\] abundances precise to $\lesssim$0.1 dex, and spectrophotometric distances precise to $\lesssim$10$\%$ (see @Cargile19 for details on the stellar parameter pipeline). Combined with Gaia proper motions (SNR$>$3 for $>$90$\%$ of the sample), H3 thus provides the full 6D phase-space and 2D chemical-space for the sample stars. The survey is ongoing – $\approx$125,000 targets have been observed as of March 2020 and they form the basis of this work. In this paper we focus on an SNR$>3$ sub-sample whose stellar parameters are deemed robust (“`flag`$=0$" in `v2.4` of the survey catalogs, but also allowing for BHBs and RR Lyrae). We work only with the primary parallax-selected and secondary color-selected K giant/BHB/RR Lyrae samples described earlier (`xfit_rank`$=$ 1 or 2), leaving out the fainter and higher parallax filler targets. We restrict our sample to the 6799 giants ($\log{g}<3.5$) to ensure a relatively uniform view of the halo. The dwarfs, while numerous, are complete only out to $d_{\rm{helio}}\sim10$ kpc, and would require significant selection function corrections (see §\[subsec:magcorr\] for details) to be interpreted on the same footing as the giants used in this study. Visual inspection of the spectra and corner-plots of the stellar parameters suggest metallicities below $-3$ are less reliable at SNR$\approx3$ so we remove the 23 stars that would have otherwise made it into our sample. We further limit this sample as per the considerations in §\[subsec:pot\] and §\[subsec:magcorr\]. ![image](Fig5.pdf){width="\linewidth"} ![image](Fig6.pdf){width="\linewidth"} ![image](Fig7.pdf){width="1.0\linewidth"} ![image](Fig8.pdf){width="\linewidth"} Computing Phase-Space Quantities {#subsec:pot} -------------------------------- We adopt the Galactocentric frame implemented in `Astropy v4.0` [@astropy1; @astropy2] which has the following parameters: $R_{0}=8.122$ kpc [@Gravity19], $[V_{\rm{R,\odot}}, V_{\rm{\phi,\odot}}, V_{\rm{Z,\odot}}]=[-12.9, 245.6, 7.78]$ km s$^{-1}$ [@Drimmel18], $Z_{\rm{\odot}}=20.8$ pc [@BennettBovy19]. This frame is right-handed, i.e., prograde (retrograde) orbits have $L_{\rm{z}}<0$ ($L_{\rm{z}}>0$). Potential-related quantities (actions, eccentricities, energies) are computed using `gala v1.1` [@gala1; @gala2] with its default `MilkyWayPotential`. This potential, based on @Bovy15, is composed of a @Hernquist90 nucleus ($m=1.7\times10^{9}\ \rm{M}_{\rm{\odot}}$, $a=1$ kpc) and bulge ($m=5\times10^{9}\ \rm{M}_{\rm{\odot}}$, $a=1$ kpc), a @MiyamotoNagai75 disk ($m=6.8\times10^{10}\ \rm{M}_{\rm{\odot}}$, $a=3$ kpc, $b=0.28$ kpc), and a spherical @Navarro97 dark matter halo ($m=5.4\times10^{11}\ \rm{M}_{\rm{\odot}}$, $a=15.62$ kpc), where $m$, $a$, $b$ are the characteristic mass and scale radii of these models respectively. The mass enclosed within 200 kpc is $9.9\times10^{11}$ M$_{\rm{\odot}}$ consistent with recent estimates [e.g., @Zaritsky20 and references therein]. We also show the final summary plots in the @McMillan17 potential in Appendix \[appendix:pot\], both to ease comparison with studies that use this potential [e.g. @Myeong19; @Koppelman19], and to demonstrate that the features described in this paper are not specific to our choice of the Galactic potential. Orbits are computed using the @DormandPrince explicit integration scheme, which belongs to the Range-Kutta family of ordinary differential equation solvers, with time-steps of 1 Myr and a total integration time of 25 Gyrs. Eccentricities are computed from these orbits as $e=\frac{r_{\rm{apo}}-r_{\rm{peri}}}{r_{\rm{apo}}+r_{\rm{peri}}}$ where $r_{\rm{apo}}$ and $r_{\rm{peri}}$ are the orbital apocenter and pericenter respectively. Actions ($J_{\mathrm{R}}, J_{\mathrm{\phi}}, J_{\mathrm{z}}, J_{\rm{tot}}=\sqrt{J_{\rm{\phi}}^{2}+J_{\rm{z}}^{2}+J_{\rm{R}}^{2}}$) are estimated from the computed orbits as per the torus-mapping method described in @Sanders16. We test the robustness of the computed actions by checking that (a) $J_{\mathrm{\phi}}$ converges to $L_{\mathrm{z}}$ within $5\%$, as expected in an axisymmetric potential like the one adopted here, and (b) actions calculated using $75\%$ of the orbit and $100\%$ of the orbit differ by no more than $1\%$. In the few cases ($\approx5\%$) where these conditions are not met (typically long-period orbits for stars at $>20$ kpc), we recompute the actions by extending the 25 Gyr integration period by $2\times$, up to 200 Gyrs (this is a choice made purely for numerical stability to collect a statistical number of orbital periods for long-period orbits). After this, only a small number of bound stars ($\approx50$), mostly with $L_{\mathrm{z}}\approx0$ fail our tests (as expected for very eccentric orbits with $J_{\rm{R}}\sim J_{\rm{tot}}$, Figure 3 of @Sanders16) and we exclude them from any analysis involving actions. The error-budget on phase-space quantities is dominated by uncertainties in spectrophotometric distances and Gaia PMs. For an illustration of how measurement errors distort substructure in [$E-L_{\rm{z}}\ $]{}we point readers to Appendix \[appendix:errorvec\]. Since [$E-L_{\rm{z}}\ $]{}is a key diagram in the analysis to follow, we limit the sample to stars that satisfy (i) $(\|E_{\rm{tot}}\|/\sigma_{E_{\rm{tot}}}>3) \land (\|L_{\rm{z}}\|/\sigma_{L_{\rm{z}}}>3)$, or (ii) $(\sigma_{E_{\rm{tot}}}<0.1 \times 10^{5} \rm{km}^{2}) \land (\sigma_{L_{\rm{z}}}<0.5\times 10^{3} \rm{kpc\ km\ s^{-1}})$, where “$\land$" stands for the Boolean “and" operator. Condition (i) is a relative error cut, and condition (ii) ensures we do not discriminate against low $\|L_{\rm{z}}\|$ stars that comprise a large fraction of the halo. 1024 stars, a majority of which have uncertain Gaia PMs (SNR$<3$) are excised due to these cuts. The excised stars lie at larger distances, including some of our most distant giants, and $\approx300$ of them judging by their PMs are likely members of the Sagittarius stream. The excised stars have an MDF similar to the sample used in this work, except they have fewer metal-rich disk stars (since the excised stars lie at larger distances). We expect improvement in the PM SNR for these stars from future data releases of the Gaia mission. This leaves us with a current sample of 5752 giants. Correcting for the H3 Selection Function {#subsec:magcorr} ---------------------------------------- Every spectroscopic survey has a selection function that can be thought of as the conditional probability $p\ (obs.\|\ \theta)$ that a star with parameters $\theta$ ($l$, $b$, $d_{\rm{helio}}$, age, \[Fe/H\], \[$\alpha$/Fe\]...) will be observed by the survey [for a comprehensive overview see @Everall20]. In general, the survey selection function represents a biased view of the underlying population. For instance, in a purely magnitude-limited survey, $p\ (obs.)$ is higher for nearby stars, so if one compares the fractions of two accreted structures with different mean heliocentric distances, raw star counts would provide a biased picture. In order to obtain an unbiased view, one must correct for the selection function by using weights that are proportional to $p\ (obs.\|\ \theta)^{-1}$. In what follows, we outline how these weights are computed for our sample. The H3 selection function can be decomposed into three independent components: (i) where we point the telescope ($\|b\|>30^{\circ}$, Dec$>-20^{\circ}$), i.e., the “window selection", (ii) the sample definition ($15<r<18$, $\pi-2\sigma_{\pi}<0.5$), i.e., the “magnitude selection" and (iii) the fraction of stars from the input sample that end up with spectra, which leads to a “targeting selection". Window Selection: H3 is limited to $\|b\|>30^{\circ}$ by design and to Dec.$>-20^{\circ}$ by geography. Any structures that are anisotropically distributed on the sky will require some correction for the survey window function. However, correction for the window is difficult as it requires a model for the underlying anisotropy. In this work we limit ourselves to demonstrating the existence of various substructures, and commenting on their relative contribution to the high-latitude Galaxy sampled within our survey fields. Magnitude Selection: The H3 selection function imposes a magnitude cut ($15<r<18$), which introduces a bias against distant and intrinsically less-luminous sources. We also limit the sample in this work to SNR$>3$, which further discriminates against fainter sources. Further, by restricting the sample only to giants ($\log{g}<3.5$) we are excluding bright, nearby dwarfs that satisfy the magnitude and parallax selections. These effects are illustrated in the top-panel of Figure \[fig:magcorr\] for an example MIST `v2.0` isochrone (10 Gyr, \[Fe/H\]$=-1$, \[$\alpha$/Fe\]$=0$, @Choi16, Dotter et al. in prep). To correct for this, we sample from an isochrone matched to each star’s derived parameters (age, \[Fe/H\], \[$\alpha$/Fe\], $A_{\rm{V}}$) using a @Kroupa01 IMF and calculate $f_{\rm{mag}}\ (d_{\rm{helio}})$, the fraction of $\log{g}<3.5$ stars at the star’s distance that fall at $15<r<18$. We shrink the magnitude range according to the SNR$>$3 cut-off for each field, which varies with observing conditions. For the subsample of color-selected rare stars (K giants and BHBs, $\approx6\%$ of the sample), instead of calculating $f_{\rm{mag}}\ (d_{\rm{helio}})$ using $15<r<18$ we use the appropriate color cuts and magnitude limits. The correction weights ($1/f_{\rm{mag}}$) for the example isochrone are plotted as a function of $d_{\rm{helio}}$ in the bottom panel of Figure \[fig:magcorr\]. The curve has a “U" shape with a steep rise below 4 kpc and above 35 kpc. The number density of stars predicted by the IMF is high close to the main-sequence turnoff and falls off precipitously as one goes up the red giant branch. At $d_{\rm{helio}}=4-35$ kpc the sections of the red giant branch with the highest number density as well as the red clump are almost entirely contained within $15<r<18$, and so at these distances the correction factor is fairly flat. Importantly, this distance range is where the bulk of our sample ($>90\%$) lies. This means the H3 giants (even without any applied corrections) provide a relatively unbiased view of the halo at these distances. Targeting Selection: Of all the stars that satisfy our selection function, we assign fibers to $\approx200$ per field. In fields closer to the dense galactic plane, the fraction of stars that are assigned a fiber is lower compared to higher latitudes – that is, at low $\|b\|$ the stars in our sample represent a larger underlying population. Another targeting bias arises from the higher fiber assignment rank we award to the small number ($<$1 per field on average) of rare, color-selected BHBs and K giants that we complement our main parallax-selected sample with. The higher rank means that fibers are assigned to all possible BHBs and K giants in a field before the other sources. As a result, the median fiber assignment probability for these stars is slightly higher than the main parallax-selected sample ($\approx85\%$ vs $\approx65\%$). Correcting for both these effects is straightforward. For a given field we compute $f_{\rm{target}} \rm{(rank)}$, the fraction of stars of a given rank that ended up with SNR$>$3 spectra out of all the stars of that rank that satisfied our selection function. For stars of the same rank, $f_{\rm{target}}$ is completely independent of stellar properties and hence can be simply multiplied with $f_{\rm{mag}}$ to produce the total weight. We note that our approach here is conceptually similar to previous work [e.g., @Bovy14; @Stonkute16; @Das16; @Vickers18; @Everall20]. We excise all 68 stars at $d_{\rm{helio}}<3$ kpc from our sample due to their very high weights (see Fig. \[fig:magcorr\]), and because we are interested in the distant Galaxy, leaving us with a final sample of 5684 stars. In the summary plots where we interpret the relative fractions of various substructures (Figs. \[fig:summary1\], \[fig:accretedinsitu\], \[fig:FeHaFe\]), and in Table \[table:summary\] we employ weights equal to $(f_{\rm{mag}}f_{\rm{target}})^{-1}$. In all other figures we display raw counts. The distinction between raw and weighted quantities is made explicit throughout the text. Results {#sec:results} ======= Overview of the High-Latitude Galaxy {#subsec:results_overview} ------------------------------------ We begin with a general overview of the data in chemistry and integrals of motion ([$E-L_{\rm{z}}$]{}, actions, eccentricity) to motivate the selection criteria for various structures in the sections to come. Figure \[fig:data\] introduces the projections of phase-space and chemistry we will use frequently. Figure \[fig:elz\_opener\] provides an [$E-L_{\rm{z}}\ $]{}“map" that identifies structure that will be presented in subsequent sections. We do this so readers can see the entire landscape at once, which will be helpful as we discuss individual structures in depth. Figures \[fig:feh\_slices\], \[fig:feh\_orbit\_slices\], \[fig:confusogram\_feh\] present a high-level overview of features in [$E-L_{\rm{z}}$]{}, actions and chemistry. In Figure \[fig:data\], the top-left panel shows [$E-L_{\rm{z}}$]{}, which we use as our primary workspace. It has long been recognized that groups of stars accreted together display coherence in their energies, and in the $z$-component of their angular momentum, even when they are thoroughly dispersed in configuration space [e.g., @Helmi00; @Brown05; @Gomez10; @Gomez13; @Simpson19]. In the second panel we display $V_{\rm{r}}$ vs. $V_{\rm{\phi}}$ – in this space stars on disk-like orbits intuitively occupy the region around the assumed rotation velocity of the Sun. This is also the space in which GSE was discovered by @Belokurov18 as an overdensity of stars around $V_{\rm{\phi}}=0$ that is also prominent in our data. In the third panel we depict a summary of actions in the form of $(J_{\rm{z}}-J_{\rm{R}})/J_{\rm{tot}}$ vs. $J_{\phi}/J_{\rm{tot}}$ following @BT [@Vasiliev19]. Generally, stars on very radial or eccentric orbits are confined to the bottom half of this diagram while stars on polar or circular ($J_{\rm{R}}=0$) orbits occupy the top half of this diagram. Circular, in-plane disk orbits have $J_{\rm{z}}=J_{\rm{R}}=0$, and $J_{\rm{\phi}}/J_{\rm{tot}}=-1$. Purely planar ($J_{\rm{z}}=0$) orbits, that would also occupy the bottom half of this diagram, are under-represented in our $\|Z_{\rm{gal}}\|>2$ kpc sample. A similar diagram was used by @Myeong19 to discover the retrograde, accreted Sequoia structure, and is a useful way to isolate GSE since it is largely confined to $J_{\rm{z}}<J_{\rm{R}}$ orbits. Eccentricities (bottom-left panel of Figure \[fig:data\]) are similarly useful, in that GSE is almost completely confined to $e>0.7$. Plotting eccentricity vs $r_{\rm{gal}}$ also shows abrupt changes in the density of stars around the pericenter/apocenter of various structures. While actions are useful, we favor [$E-L_{\rm{z}}\ $]{}while defining selections in part because this space is simpler to understand. Further, a large body of local halo studies has primarily deployed energy, eccentricities, and angular momenta, and we seek to draw direct connections and build on it [e.g., @Helmi17; @Belokurov18; @Koppelman19]. A high degree of overlap of multiple accreted structures is expected in [$E-L_{\rm{z}}\ $]{}and other projections of phase-space [e.g., @bj05_3; @Jean-Baptiste17; @Pfeffer20], which we resolve when possible using chemistry (bottom-center and bottom-right panels of Figure \[fig:data\]). Stars belonging to the same structure are expected to show coherent MDFs and distinct chemical evolutionary tracks in the \[Fe/H\] vs. \[$\alpha$/Fe\] plane that are a function of their mass, star-formation history and formation redshift (discussed further in §\[subsec:chemistry\]). In Figure \[fig:feh\_slices\] we show [$E-L_{\rm{z}}\ $]{}in bins of metallicity and color-coded by \[$\alpha$/Fe\]. The most metal-rich stars define two sequences: one at higher energy, lower \[$\alpha$/Fe\] (Aleph), and the other at lower energy, higher \[$\alpha$/Fe\] (the high-$\alpha$ disk) that extends to $L_{\rm{z}}\sim0$ orbits (the in-situ halo). At $-0.75<$\[Fe/H\]$<-0.5$ two structures appear, one centered at $L_{\rm{z}}\sim0$ (GSE), and the other at high energy, $L_{\rm{z}}\sim-2$, $E_{\rm{tot}}\sim-0.75$ (Sgr). The density of the $L_{\rm{z}}\sim0$ population (GSE) peaks in the $-1.25<$\[Fe/H\]$<-1$ panel. High-energy retrograde stars only begin to appear at \[Fe/H\]$<-0.75$, and are almost entirely absent from the higher metallicity bins. Several smaller clumps appear in various \[Fe/H\] intervals. From this figure it is already clear that a very small fraction of the halo within 50 kpc is metal poor (\[Fe/H\]$\leq-1.75$). Figure \[fig:feh\_orbit\_slices\] is similar to Figure \[fig:feh\_slices\], but here we separate the stars by actions. Stars with $J_{\rm{z}}>J_{\rm{R}}$, on polar or circular orbits are limited to the top row and stars with $J_{\rm{z}}<J_{\rm{R}}$, on radial or eccentric orbits are limited to the bottom row. The most prograde structure ($L_{\rm{z}}<-2$) is $\alpha$-poor and confined to the first two panels (Aleph). The high-energy population at $L_{\rm{z}}\sim-2$, $E_{\rm{tot}}\sim-0.75$ is also confined to the top-row (Sgr). A prominent structure centered at $L_{\rm{z}}\sim0$ appears largely at $J_{\rm{z}}<J_{\rm{R}}$ and dominates the bottom row (GSE). The other disk-like prograde population that extends to $L_{\rm{z}}\sim0$ is spread across the top and bottom rows (high-$\alpha$ disk and in-situ halo), as are the high-energy retrograde halo stars. Figure \[fig:confusogram\_feh\] depicts \[$\alpha$/Fe\] vs \[Fe/H\], binned by actions and angular momenta, and color-coded by eccentricity. This figure is particularly rich in structure, and underscores the power of combining chemistry with dynamics. We highlight a few prominent populations apparent in this figure. There is a highly circular (shaded black), metal-rich (\[Fe/H\]$<-0.5$), $J_{\rm{z}}>J_{\rm{R}}$ population completely contained within the top-left panel (Aleph). Adjacent to it, at lower \[Fe/H\] and \[$\alpha$/Fe\], appears an agglomeration of more eccentric stars that is also confined purely to the top-left panel (Sgr). The most $\alpha$-rich population is dispersed across the first two columns, and has orbits ranging from highly eccentric to circular, and extends from prograde to radial $L_{\rm{z}}$ (high-$\alpha$ disk + in-situ halo). Among the $J_{\rm{z}}<J_{\rm{R}}$ orbits we see a well-populated sequence largely contained within the bottom-center panel (GSE). Through these figures we have demonstrated the distant halo to be highly structured in chemodynamical space, with various populations appearing preferentially in certain regions of metallicity and orbital space. We now proceed to define and characterize these individual structures in detail. Substructure Inventory {#subsec:inventory} ---------------------- In what follows, we provide a detailed inventory of the $\|b\|>40^{\circ}, d_{\rm{helio}}>3$ kpc Milky Way, one component at a time. We provide relevant background on each component, justify our selection, and comment on any noteworthy features. We support this discussion with a corresponding 6-panel figure for each component that situates it in chemodynamical space. Each 6-panel figure follows the layout introduced in Figure \[fig:data\], with the top-right panel changing across figures to highlight a particular projection of chemodynamical space most relevant for the structure under discussion. We emphasize that the primary goal of this work is a high-level inventory. This means we focus on cleanly selecting various components rather than on a thorough characterization and analysis of their nature, which we defer to forthcoming work. We first outline our overall strategy. We begin by selecting the most coherent, well-defined structures in chemodynamical space – Sagittarius, Aleph, the high-$\alpha$ disk $\&$ and the in-situ halo. Having accounted for the eccentric stars of the in-situ halo and Sgr, we assign the remaining highly eccentric ($e>0.7$) stars to GSE. Next, we isolate other known halo structures in the literature (the Helmi Streams, Thamnos, Sequoia). While investigating Sequoia in the high-energy retrograde halo we identify a relatively metal-poor (I’itoi) and metal-rich population (Arjuna) in the same [$E-L_{\rm{z}}\ $]{}region. After subtracting out all these structures, we turn to a remaining prograde [$E-L_{\rm{z}}\ $]{}overdensity (Wukong). We also highlight a metal-poor, $\alpha$-rich, rotationally supported population that we identify as the metal-weak thick disk. The remaining stars are labeled unclassified debris. While selecting a structure we exclude all the previously defined structures. This ensures that new structures (Arjuna, I’itoi, Wukong) have minimal overlap with previously identified ones. We often rely on chemistry in our selections due to the high degree of overlap expected for accreted structures in integrals of motion [e.g., @bj05_3; @Jean-Baptiste17; @Pfeffer20]. For instance, the low eccentricity tail of GSE ($e<0.7$) is a major contaminant in purely phase-space selections of lower-mass objects, but we are able to exclude it by appealing to chemistry. As much as possible, we incorporate insights from the existing literature in our selections – for instance, for the Helmi Streams and Thamnos we use literature definitions as our starting point, and for Sequoia we are guided by previous studies of its chemistry. Instead of using clustering algorithms [e.g., @Yuan18; @Yuan20; @Mackereth19; @Koppelman19], we take an artisanal approach, making simple, physically motivated, easily reproducible selections. Through extensive experimentation we have found that clustering algorithms (e.g., DBSCAN, HDBSCAN, k-means) either fracture the space into too many clusters, or assign the entire sample to Sagittarius, GSE, the high-$\alpha$ disk, and Aleph (i.e., the structures apparent by eye in Figures \[fig:feh\_slices\], \[fig:feh\_orbit\_slices\], \[fig:confusogram\_feh\]). In the case of a high degree of fracturing, we then had to consider one at a time the nature of each mini-cluster, akin to @Yuan20 whose algorithm applied to \[Fe/H\]$<-1.8$ stars in LAMOST yielded 57 distinct groups, almost all of which they coalesced back into GSE and Sequoia. A downside of our approach compared to clustering methods is the deterministic assignment of every star as belonging to one structure or another instead of assigning membership probabilities and marginalizing over them (also discussed in §\[subsec:caveats\]). ### Sagittarius {#subsec:sgr} The stream of debris associated with the Sagittarius (Sgr) dwarf spheroidal galaxy [@Ibata94] provides the clearest demonstration of the hierarchical build-up of the stellar halo. In recent years Sgr debris has been traced out to $\sim100$ kpc, showing surprising features [e.g., @Belokurov14; @Hernitschek17; @Sesar17; @Li19], and inspiring a new generation of numerical models [e.g., @Dierickx17; @Fardal19; @Laporte18] that builds on earlier work [e.g., @Johnston95; @LM05; @LM10]. In tandem, its chemistry is beginning to be resolved in ever greater detail by large spectroscopic efforts [e.g., @Alfaro-Cuello19; @Li19; @Hayes20] that are building on earlier efforts [e.g., @Bellazzini06; @Chou07; @Moncao07; @Carlin12; @Gibbons17]. Before Gaia, studies made the best of incomplete phase-space data to select Sgr stream stars, typically relying on heuristics such as distance from the orbital plane [e.g., @Newberg03; @Belokurov14; @Lancaster19]. However, with full phase-space information, clean selections that fully exploit the highly coherent Sgr features are now possible [e.g., @Li19; @Yang19; @Hayes20]. Sgr is a high-energy, prograde overdensity in $E-L_{\mathrm{z}}$, and owing to its polar orbit, its angular momentum is concentrated in $L_{\rm{y}}$. Capitalizing on this, we define Sgr stars as those which satisfy: $$\label{eq:sgr} \begin{aligned} L_{\rm{y}} < -0.3L_{\rm{z}}-2.5\times10^{3}\ \rm{kpc\ km\ s^{-1}}. \end{aligned}$$ We verify that this simple criterion selects $>99.5\%$ of the star particles in a version of the @LM10 model that is matched to the current H3 footprint with a $10\%$ distance uncertainty (see @Johnson20 for more detailed comparisons with models). The selected 612 stars are shown in Figure \[fig:sgr\] – Sgr comprises the majority of our distant stars. Additionally, we identified 63 stars that do not satisfy Eq. \[eq:sgr\] but have PMs highly aligned with stars selected by Eq. \[eq:sgr\]. Closer inspection revealed that the distances to these stars are incorrect, due to confusion between the red clump and red giant branch [@Masseron17; @Mackereth17]. When the distances to these stars are doubled, they satisfy Eq. \[eq:sgr\]. In our pipeline this confusion arises for low-$[\alpha\rm{/Fe}]$ stars ($<0.1$) where Sgr is the dominant structure. These stars have a similar magnitude distribution to those selected by Eq. \[eq:sgr\], so when computing relative fractions (e.g., in Table \[table:summary\] and Figure \[fig:summary2\]) we adjust our Sgr numbers upwards by $10\%$ (i.e., 63/612) but do not show these stars in projections of phase-space. A detailed characterization of Sgr in H3 is forthcoming [@Johnson20]. Here we only remark on prominent features in the MDF, which displays two peaks in \[Fe/H\] separated by $\approx 0.4$ dex. This is consistent with the picture in @Hayes20 [their Fig. 7], who find different mean metallicities in the leading and trailing arms and a similarly multi-peaked MDF. There may also be a link to the complex star-formation history and distinct chemical populations recently shown to exist in the core of the Sgr dwarf [@Alfaro-Cuello19]. Stars in the metal poor (\[Fe/H\]$\lesssim-2$) tail are highly aligned with Sgr in angular momenta as well as in Gaia proper motions (that are independent of the measured distances and radial velocities) and will be a point of focus of @Johnson20. ### Aleph {#subsec:aleph} ![image](Fig9.pdf){width="\linewidth"} ![image](Fig10.pdf){width="\linewidth"} ![image](Fig11.pdf){width="\linewidth"} Aleph[^2] is a hitherto unknown prograde substructure. We discovered Aleph in [\[$\alpha$/Fe\]]{}$-$\[Fe/H\] as a sequence below the high-$\alpha$ disk at similar \[Fe/H\]. Examining the dynamics of these stars, we found them to be highly coherent and on circular orbits (Figure \[fig:confusogram\_feh\]), comprising the most prograde stars of our sample at higher-energy than the high-$\alpha$ disk in [$E-L_{\rm{z}}\ $]{}(Fig. \[fig:feh\_orbit\_slices\]). Another characteristic feature of Aleph that clearly differentiates it from the canonical disk populations is its significant vertical action, which is seen prominently in Fig. \[fig:confusogram\_feh\], where Aleph is completely confined to the top-left panel depicting prograde, $J_{\rm{z}}>J_{\rm{R}}$ stars. The classic $\alpha$-rich and $\alpha$-poor disk sequences typically have $J_{\rm{z}}<J_{\rm{R}}$ [e.g., @Sanders16; @Beane19; @Ting19]. We define Aleph stars as follows: $$\begin{aligned} (V_{\rm{\phi}} < -175\ \rm{km\ s^{-1}})\ \land\ (V_{\rm{\phi}} > -300\ \rm{km\ s^{-1}})\\ \land\ (\|V_{\rm{r}}\|<75\ \rm{km\ s^{-1}})\\ \land\ ([\rm{Fe/H}]>-0.8) \land ([\alpha/\rm{Fe}]<0.27)\\ \land\ (\rm{excluding\ all\ previously\ defined\ structures}). \end{aligned}$$ The resulting population is shown in Figure \[fig:aleph\]. In our sample Aleph is localized spatially ($r_{\rm{gal}}=11.1^{+5.7}_{-1.6}$ kpc), with stars extending to 25 kpc. It is a metal-rich (\[Fe/H\]$=-0.51$), relatively alpha-poor (\[$\alpha$/Fe\]$=0.19$), rapidly rotating ($V_{\rm{\phi}}\approx-210$ km s$^{-1}$) structure on a highly circular orbit ($e=0.13\pm0.06$) with a strong vertical action ($\langle J_{z}\rangle\approx190$ kpc km s$^{-1}$) and orbits that rise to $\|Z_{\rm{gal}}\|\approx10$ kpc. All quoted values have been weighted by the selection function. The low eccentricity and chemistry of Aleph suggest an origin within the Galactic disk. Interestingly, in our sample Aleph is mostly confined to the Galactic anti-center, where several overdensities linked to the excitation of the outer disk (e.g., Monoceros, A13, TriAnd1, TriAnd2) have been observed [e.g., @Newberg02; @Ivezic08; @Price-Whelan15; @Li17; @Bergemann18], though our sample is at $\|b\|>40^{\circ}$, at slightly higher latitudes than these features. Several of Aleph’s properties – the radial extent, chemical nature, rotational velocity – are also similar to recently reported features of outer disk stars in @Lian20. It is possible that the @Lian20 APOGEE sample is the in-plane view of Aleph, while we are sampling it at higher latitudes. A detailed exploration of Aleph’s nature is the subject of ongoing work. Aleph is coincident with the enigmatic GC Palomar 1 (Pal 1, red star in Fig \[fig:aleph\]) in integrals of motion, is at very similar elevation ($Z=3.6$ kpc) and similar metallicity (\[Fe/H\]$\approx-0.5$), but is less $\alpha$-enhanced (Pal 1: [\[$\alpha$/Fe\]]{}$\approx0$, Aleph: [\[$\alpha$/Fe\]]{}$\approx0.2$). We adopt Pal 1 phase-space coordinates from @Baumgardt19 and abundances from @Sakari11. Since its discovery Pal 1 has been recognized as a curiosity – its high elevation resembles halo GCs but its young age and high metallicity have proven puzzling ($4-7$ Gyrs, and among the youngest, most metal rich, and faintest of MW GCs, e.g., @vanderbergh04 [@Sakari11; @Sarajedini11]). For decades authors have speculated about its origin, wondering whether it may be an unusually old open cluster, may have a peculiar IMF, or may have been accreted with a dwarf galaxy [e.g., @Rosenberg98; @Rosenberg98b; @Niederste-Ostholt10]. In recent years the accretion origin has gained currency. Other young ($5-8$ Gyrs), low surface brightness GCs (e.g,. Terzan 7, Pal 12, and Whiting 1) have been associated to Sgr, i.e, they are of extragalactic origin [e.g., @Carraro07; @Koposov07; @LM10; @Johnson20]. @Sakari11 analyzed neutron capture elements in four stars in Pal 1, and found them to be distinct from MW field stars, which led them to argue Pal 1 was accreted along with a dwarf galaxy. Pal 1 also does not lie on the in-situ branch of MW GCs in the age-metallicity relation (@Forbes10 [@Forbes20]; @Kruijssen20; but see @Massari19). Whether or not Pal 1 was accreted or born in-situ, the similarity between Pal 1 and Aleph in chemodynamical space suggests a common origin. ### High-$\alpha$ Disk and In-situ Halo {#subsec:insitu} It has long been known that stars on disk-like orbits lie on one of two chemical sequences characterized by low or high values of [\[$\alpha$/Fe\]]{} [e.g., @Edvardsson93; @Fuhrmann98; @Chen00; @Bensby03; @Adibekyan12]. With Gaia data it was realized that the high-$\alpha$ population extends to higher eccentricities than a conventional disk-like population. This high-$\alpha$, high eccentricity population has been dubbed the “in-situ halo" and later as the “Splash" [e.g., @Bonaca17; @Bonaca20; @Haywood18; @DiMatteo19; @Amarante20; @Belokurov20]. Simultaneously, a link between the accretion of GSE, the formation of the high-$\alpha$ disk, and the creation of the in-situ halo has been proposed [e.g., @Helmi18; @Gallart19; @Belokurov20; @Bonaca20]. Characterizing this component of the halo is thus critical to understanding the Galaxy’s earliest epoch. We define the high-$\alpha$ disk and in-situ halo stars relying purely on chemistry: $$\label{eq:td} \begin{aligned} \mathrm{[\alpha/Fe]} > 0.25-0.5\ (\mathrm{[Fe/H]}+0.7)\\ \land\ (\rm{excluding\ all\ previously\ defined\ structures}). \end{aligned}$$ In Figure \[fig:insitu\] we see these stars form a kinematic population that extends continuously from rotationally supported orbits (forming a locus at lower $\|V_{\rm{\phi}}\|$ than Aleph) to highly eccentric ones. In [$E-L_{\rm{z}}\ $]{}the high-$\alpha$ disk forms a more diffuse track slightly steeper than Aleph, which extends into high-eccentricity orbits with $L_{\rm{z}}\sim0$. The continuity of the distribution in phase-space supports scenarios in which the ancient, rotationally supported high-$\alpha$ disk was dynamically heated, perhaps by a merger. The in-situ halo extends to $r_{\rm{gal}}\approx25$ kpc, but we caution that the stars at $\|Z\|>15$ kpc lie very close to the selection boundary in chemistry, and may belong to other structures. We discuss the physical origin of the in-situ halo further in §\[subsec:origin\]. This selection excludes the metal-poor tail of the high-$\alpha$ disk and the in-situ halo [e.g., @Carollo19] which lies in a region of \[Fe/H\] vs \[$\alpha$/Fe\] coincident with GSE and other accreted structures. We will return to these stars in the sections dealing with the metal-weak thick disk (§\[subsec:mwtd\]) and the unclassified debris (§\[subsec:leftovers\]). ### Gaia-Sausage-Enceladus (GSE) {#subsec:ge} We define “GSE” as the highly radial population that comprises the bulk of the accreted local halo. This population was identified in various ways by @Belokurov18 [@Koppelman18; @Myeong18; @Haywood18; @Helmi18; @Mackereth19; @Koppelman19; @Helmi20]. Different selections result in differing degrees of contamination with overlapping structures [see discussion in @Evans20]. We select GSE stars by excluding the previously defined structures and requiring $e>0.7$. The eccentricity selection is motivated by the dense cloud of stars at $e>0.7$ seen in eccentricity vs $r_{\rm{gal}}$, whose density sharply drops off at $\approx30$ kpc (corresponding to the proposed apocenter of GSE; @Deason18 [@Lancaster19]). Our GSE selection is therefore simply: $$\begin{aligned} (e>0.7)\\ \land\ (\rm{excluding\ all\ previously\ defined\ structures}). \end{aligned}$$ This selection is very similar in spirit to the $V_{\rm{r}}-V_{\rm{\phi}}$ selection in @Belokurov18, where this structure was discovered, as borne out by the second panel of Figure \[fig:ge\]. This selection is by no means perfect – it is incomplete in that it misses the low-eccentricity tail of GSE at $e<0.7$ that manifests as a strong peak at \[Fe/H\]$\approx-1.2$ in subsequent plots. And it is impure, as suggested by the structure along the margins of GSE in [$E-L_{\rm{z}}\ $]{}– for instance, $e>0.7$ stars from Wukong (discussed in §\[subsec:wukong\]) are apparent at $L_{\rm{z}}/[\rm{10^{3} kpc\ km\ s^{-1}}]\sim-0.5$. A subtle sequence corresponding to residue from Wukong also appears under the GSE sequence in \[Fe/H\] vs \[$\alpha/\rm{Fe}$\]. However, the very well-behaved, unimodal MDF inspires confidence that this selection is overwhelmingly comprised of GSE stars. The MDF is narrow – corrected for the selection function, $85\%$ of stars are contained within 0.9 dex in \[Fe/H\] – and reminiscent of some local dwarfs (e.g., Leo I and Fornax, @Kirby13). Like Leo I and Fornax, the GSE MDF is well-fit by a simple, analytical, chemical evolution model, namely the “Best Accretion Model" [@Lynden-Bell75] used in @Kirby11 [@Kirby13], that explains all features, including the extended metal-poor tail (dotted line in MDF panel of Figure \[fig:ge\]). This model is a generalization of traditional leaky box models, allowing for the accretion of fresh gas, and has two parameters – $M$, the ratio between the final mass and initial gas mass of the system, and $p$, the effective yield (i.e., a measure of the fraction of metals produced by stars the system retains) – for which we find best-fit values of $p=0.085, M=3.28$ (slightly different from @Conroy19b who found $p=0.08, M=2.1$ for a differently selected, SNR$>5$ kinematic halo sample at $-0.5<L_{\rm{z}}/[10^{3}\ \rm{kpc\ km\ s^{-1}}]<1$). As in the case of both Fornax and Leo I in @Kirby13, the data falls off more steeply than the model on the metal-rich side of the MDF. Our estimate of GSE’s metallicity (\[Fe/H\]$=-1.15^{+0.24}_{-0.33}$, weighted) is $\approx0.1-0.2$ dex higher than most of the literature [e.g., @Helmi18; @Matsuno19; @Sahlholdt19; @Vincenzo19; @Mackereth19] and more in line with the recent \[Fe/H\]$=-1.17\pm0.34$ estimate of @Feuillet20. To convert \[Fe/H\]$=-1.15$ to a mass estimate, we use the mass-metallicity relation from local dwarfs [@Kirby13] and account for the redshift evolution of the relation – i.e., higher masses at higher redshift at fixed \[Fe/H\] [e.g., @Zahid14; @Steidel14; @Sanders15mosdef; @Ma16MZR; @Torrey19]. Assuming the trend from the FIRE simulations @Ma16, which agrees well with observations out to $z\sim3$, produces $M_{\star}=4-7\times10^{8}\,M_{\odot}$ for accretion redshifts between $z=1.3$ [@Kruijssen20] and $z=2$ [@Bonaca20]. This is in excellent agreement with recent estimates from GSE’s GC age-metallicity relation ($\approx2-4\times10^{8}\,M_{\odot}$, @Kruijssen20), star counts of \[Fe/H\]$<-1$, $e>0.7$ APOGEE red giants ($\approx2-5\times10^{8}\,M_{\odot}$, @Mackereth20), and the integrated SFR of a chemical evolution model ($\approx6\times10^{8}\,M_{\odot}$, @Helmi18 [@Fernandez-Alvar18]). We are also in a position to address the mean rotational velocity of GSE, which is of interest because it informs the initial configuration of the merger. We find $\langle V_{\rm{\phi}}\rangle = 1.04^{+1.26}_{-1.25}\ \rm{km\ s^{-1}}$, $\langle L_{\rm{z}}\rangle = 4.7^{+20.1}_{-10.5}\ \rm{kpc\ km\ s^{-1}}$ with errors estimated via bootstrap resampling including fully propagated errors from distance and PM samples, and weighting for the selection function. This measurement places a very strong constraint on the lack of net rotation of GSE. This conclusion is in excellent agreement with @Belokurov20 who report $\langle V_{\rm{\phi}}\rangle\sim0$ for the “Sausage" component in their velocity ellipsoid fits for a local sample drawn from @Sanders18 with $6.5<R_{\rm{gal}}<10$ kpc. The magnitude of rotation we measure is much lower than @Mackereth19 who find $L_{\rm{z}}= 176\ \rm{kpc\ km\ s^{-1}}$ using 673 APOGEE stars at $\|Z\|<10$ kpc with spectrophotometric distances uncertain on the $\sim15\%$ level, and @Helmi20 who report $\langle V_{\rm{\phi}}\rangle (d_{\rm{helio}}<1\ \rm{kpc}) = 21.1\pm1.8\ \rm{km\ s^{-1}}$ using 6 stars and $\langle V_{\rm{\phi}}\rangle (d_{\rm{helio}}<2\ \rm{kpc}) = 16.1\pm2.8\ \rm{km\ s^{-1}}$ using 23 stars from the Gaia RVS sample cross-matched with APOGEE \[Fe/H\]$\geq-1.3$ stars. We caution that this measurement is sensitive to the assumed solar motion, and that the mentioned GSE samples have all been selected differently. We also observe that including Sequoia, Arjuna, or Arjuna and Sequoia (retrograde structures discussed in §\[subsec:retrograde\]) in GSE results in $\langle V_{\rm{\phi}}\rangle = [3.7^{+1.3}_{-1.3}, 7.0^{+1.6}_{-1.5}, 9.4^{+1.7}_{-1.6}]\ \rm{km\ s^{-1}}$ respectively. Despite these caveats, it is clear that GSE is far from highly retrograde. We further discuss the rotation of the halo in §\[subsec:protretro\]. ### Helmi Streams {#subsec:hs} ![image](Fig12.pdf){width="\linewidth"} ![image](Fig13.pdf){width="\linewidth"} The Helmi Streams were among the first bona fide accreted substructures discovered in the halo via integrals of motion as opposed to on-sky streams [@Helmi99]. @Koppelman19 [@Koppelman19HS; @Koppelman20] provide an updated view of these streams using Gaia DR2 data. While a prominent [$E-L_{\rm{z}}\ $]{}overdensity corresponding to the Helmi Streams appears among the H3 dwarfs, it is not as readily apparent in the giants, though there is a hint of a vertical spur at $L_{\rm{z}}/[\rm{10^{3} kpc\ km\ s^{-1}}]\sim-1.5$ in e.g., Figure \[fig:elz\_opener\]. To select the Helmi Streams we rely on the $L_{\rm{z}}-L_{\rm{\perp}}$ selection in : $$\label{eq:HS} \begin{aligned} (-1.7<L_{\rm{z}}/[\rm{10^{3} kpc\ km\ s^{-1}}]<-0.75)\\ \land\ (1.6<L_{\rm{\perp}}/[\rm{10^{3} kpc\ km\ s^{-1}}]<3.2)\\\land\ (\rm{excluding\ all\ previously\ defined\ structures}). \end{aligned}$$ The selected stars are shown in Figure \[fig:helmi\]. Orbits of the Helmi Streams in local samples [@Helmi20 their Figure 12] rise to high latitudes and extend out to $R_{\rm{gal}}\approx25$ kpc – this is borne out in Figure \[fig:helmi\]. The large spread in eccentricity mirrors the large spread in eccentricties of GCs attributed to HS from considerations of the GC age-metallicity relation [@Massari19; @Kruijssen20; @Forbes20]. In the MDF we see a broad distribution, consistent with the complex population with an extended star-formation history modeled in @Koppelman19HS. The distribution in \[Fe/H\] vs \[$\alpha$/Fe\] traces the typical trend of decreasing \[$\alpha$/Fe\] with increasing \[Fe/H\] expected in halo populations. Assuming accretion redshifts of $z\sim0.5-1.1$, i.e., 5-8 Gyrs ago [@Koppelman19HS], and a (weighted) mean metallicity of \[Fe/H\]$\approx-1.3$, we estimate the Helmi Streams to have a stellar mass of $M_{\star}\approx0.5-1\times10^{8}M_{\rm{\odot}}$ via the @Kirby13 MZR and its expected evolution to higher redshifts [@Ma16], in excellent agreement with @Koppelman19HS. ### Thamnos {#subsec:thamnos} The Thamnos structure was recently discovered in @Koppelman19. These authors found two overdensities in chemodynamical space (“Thamnos 1" and “Thamnos 2") that they attribute to the same progenitor. An overdensity corresponding to their proposed structure appears in our [$E-L_{\rm{z}}\ $]{}diagrams as a jagged ridge along the retrograde edge of GSE (at $E_{\rm{tot}}/[10^{5}\ \rm{km}^{2} \ \rm{s}^{-2}]\approx-1.4$). This ridge resembles the corrugations of a single massive satellite producing multiple over-densities in [$E-L_{\rm{z}}\ $]{}and other phase-space diagrams [@Jean-Baptiste17]. To distinguish whether Thamnos is a remnant of a distinct satellite, or a part of GSE, it is important to verify that the chemistry of Thamnos is distinct from GSE, particularly because of the small sample ($\sim$20) of Thamnos stars with abundances in @Koppelman19, the majority of which overlap with GSE within error-bars (their Figure 4). To define our Thamnos selection we begin by selecting all stars at $(L_{\rm{z}}>0.2) \land (-1.5<E_{\rm{tot}}<-1.3)$. The [$E-L_{\rm{z}}\ $]{}selection is motivated by the contours for Thamnos provided in @Koppelman19 as well as the overdensity we see in that region. In this energy range we expect high contamination primarily from GSE. The resulting MDF depicted with a dashed line in Figure \[fig:thamnos\] shows a strong peak at \[Fe/H\]$=-1.9$ that we attribute to Thamnos as well as a second peak corresponding to GSE’s metallicity of \[Fe/H\]$\approx-1.2$. We further refine the Thamnos selection informed by this MDF by restricting the selection to \[Fe/H\]$<-1.6$, where we see a break. This results in a final selection of 32 stars that produces a clean \[Fe/H\]-\[$\alpha$/Fe\] sequence. To summarize, the Thamnos selection is: $$\begin{aligned} (0.2<L_{\rm{z}}/[\rm{10^{3}\ kpc\ km\ s^{-1}}]<1.5) \\ \land\ (-1.5<E_{\rm{tot}}/[10^{5}\ \rm{km}^{2}\ \rm{s}^{-2}]<-1.3)\\ \land\ (\rm{[Fe/H]}<-1.6)\\\land\ (\rm{excluding\ all\ previously\ defined\ structures}). \end{aligned}$$ @Koppelman19 estimate Thamnos’ stellar mass to be $<5\times10^{6}M_{\odot}$ by comparing its extent in [$E-L_{\rm{z}}\ $]{}against a suite of simulations. Thamnos lies at lower energy than GSE, which is unexpected for a system of such low stellar mass if it were accreted at $z\sim0$, as its would be shredded in the outer reaches of the halo [e.g., @Amorisco17; @Pfeffer20]. This suggests that it was accreted very early when the Galaxy was not very massive, or perhaps simultaneously with GSE at $z\sim1.3-2$ [@Kruijssen20; @Bonaca20]. Now that we have a good handle on the metallicity (\[Fe/H\]=$-1.9$) we can provide a complementary mass-metallicity relation (MZR) constraint on the mass. Using the $z=0$ @Kirby13 relation for \[Fe/H\] produces a mass $2\times10^{5}M_{\odot}$, but this is a strict lower limit. This is because we must account for the redshift evolution of the MZR. Assuming the trend from the FIRE simulations @Ma16, which agree well with observations out to $z\sim3$, and predict a $\sim1$ dex increase in mass for an \[Fe/H\]$=-1.9$ galaxy between $z=0$ and $z=1.5$, we find $M_{\star}\approx2\times10^{6}M_{\odot}$ for Thamnos, in good agreement with @Koppelman19. Thamnos is potentially a very exciting object because of its very low stellar mass that we estimate here, and because it lies so deep in the potential. Put another way, the debris from Thamnos lies only at $d_{\rm{helio}}\approx6$ kpc (weighted), can be easily targeted using our clean sample, and thus offers a unique view of galaxy evolution (e.g., stellar abundances) in a mass regime that will be out of reach even for the James Webb Space Telescope at high redshift [e.g., @Boylan-Kolchin15; @Boylan-Kolchin16; @Weisz14]. ### The High-Energy Retrograde Halo: Arjuna, Sequoia, and I’itoi {#subsec:retrograde} ![image](Fig14.pdf){width="\linewidth"} . \[fig:retrograde\] ![image](Fig15.pdf){width="\linewidth"} Studies of the local stellar halo have found a wealth of retrograde substructure [e.g., @Helmi17; @Myeong18d; @Myeong18b; @Myeong18c; @Myeong19; @Matsuno19; @Koppelman19; @Yuan20]. Debris of the Sequoia dwarf galaxy [@Myeong18b; @Myeong19; @Matsuno19] dominates the local retrograde halo at higher energies while Thamnos [@Koppelman19] resides at lower energy. In this section we turn our attention to the high-energy retrograde halo (i.e., at higher energies than Thamnos) that we select as follows: $$\begin{aligned} \label{eq:retro} (\eta>0.15) \land (L_{\rm{z}}/[\rm{10^{3} kpc\ km\ s^{-1}}]>0.7)\\ \land\ (E_{\rm{tot}}/[10^{5} \rm{km}^{2}\ \rm{s}^{-2}]>-1.25)\\\land\ (\rm{excluding\ all\ previously\ defined\ structures}). \end{aligned}$$ The circularity, $\eta=L_{\rm{z}}/\|L_{\rm{z,max}}(E_{\rm{tot}})\|$, where $L_{\rm{z,max}}(E_{\rm{tot}})$ is the maximum $L_{\rm{z}}$ achievable for an orbit of energy $E_{\rm{tot}}$. We compute $L_{\rm{z,max}}(E_{\rm{tot}})$ by assuming a perfectly circular orbit with the star’s $r_{\rm{gal}}$ and total 3D velocity. The circularity condition, $\eta>0.15$, ensures a generous selection of retrograde orbits, $L_{\rm{z}}/[\rm{10^{3}\ kpc\ km\ s^{-1}}]>0.7$ reduces contamination from GSE, and the energy limit avoids Thamnos. The stars satisfying this selection are shown in Figure \[fig:retrograde\]. A prominent peak in the high-energy retrograde MDF appears exactly where expected for Sequoia at \[Fe/H\]$\approx-1.6$ [@Matsuno19; @Myeong19; @Monty19]. More surprisingly, two other distributions are apparent in the MDF – one centerd at \[Fe/H\]$\approx-1.2$, and another spanning very low metallicity at \[Fe/H\]$<-2$. Furthermore, these stars occupy a complex distribution in \[Fe/H\]-\[$\alpha$/Fe\] that is suggestive of multiple populations. We name the metal-rich population “Arjuna"[^3] and the metal-poor sequence “I’itoi"[^4]. Based on the peaks and breaks in the MDF we define Arjuna stars as those with \[Fe/H\]$>-1.5$, Sequoia stars as those with $-2<$\[Fe/H\]$<-1.5$, and the remaining stars at \[Fe/H\]$<-2$ as belonging to I’itoi. Contrary to expectations from local studies [e.g., @Myeong19; @Koppelman19], Sequoia is not the dominant component of the high-energy retrograde halo – it has fewer than half as many stars as Arjuna. This raises the question as to why Arjuna was missed in the local studies that found Sequoia. The answer may lie in its spatial extent – Arjuna lies at larger distances compared to Sequoia (median $r_{\rm{gal}}\sim25$ kpc vs. $\sim15$ kpc, weighted), and stars with apocenters of 25 kpc are relatively rare in the solar neighborhood compared to those with apocenters of 15 kpc. Given that Arjuna may prove to be a massive component of the halo, and because of its similarity in \[Fe/H\] to GSE, it is important to discuss possible connections to GSE. Recent work has also cast doubt on the status of Sequoia as a dwarf galaxy and argued that it may be debris from the outer reaches of GSE [@Koppelman19; @Helmi20]. We explore this issue in Figure \[fig:retrograde\]. We show that even when restricting the retrograde selection to very high $L_{\rm{z}}/\rm{[10^{3}\ kpc\ km\ s^{-1}]}>2$, far from the $L_{\rm{z}}\approx0$ overdensity defined by GSE, the peaks in the MDF associated with Arjuna and Sequoia remain. None of the existing studies of GSE show a significant \[Fe/H\]$\approx-1.2$ population at such high $L_{\rm{z}}$, with stars outnumbering those from Sequoia [e.g., @Belokurov18; @Mackereth19; @Myeong19; @Koppelman19; @Helmi20]. So Arjuna may either be a hitherto unknown extension of GSE to highly retrograde orbits (such high $L_{\rm{z}}$ extensions exist for the largely eccentric debris of massive, $M_{\rm{\star}}>10^{8}M_{\rm{\odot}}$, accreted galaxies in the @bj05_1 halos), or it may be the debris of a distinct dwarf galaxy. As for Sequoia, any attempt to tie both Arjuna and Sequoia to GSE must account for $L_{\rm{z}}/\rm{[10^{3} kpc\ km\ s^{-1}]}>2$ components of GSE as well as the spread in abundances (e.g., by appealing to a steep metallicity gradient or multiple populations). More detailed modeling of a GSE-like merger [in the vein of @Bignone19; @Vincenzo19; @Elias20] is required to understand if it is possible for a single progenitor to simultaneously populate such disparate regions of phase-space as well as chemistry. The proximity of I’itoi and Thamnos in \[Fe/H\] vs \[$\alpha$/Fe\] may indicate these structures are related – however, I’itoi is prominent at $L_{\rm{z}}/\rm{[10^{3} kpc\ km\ s^{-1}]}>2$ and Thamnos’ mass argues against such a wide extent in [$E-L_{\rm{z}}$]{}[@Koppelman19 their Fig. 5]. This is because low-mass structures are typically compact, and experience similar dynamical friction across all their stars, compared to larger structures like GES. I’itoi may also be the metal-poor tail of Arjuna and/or Sequoia – more work needs to be done to differentiate these three structures in phase space as well. This will be particularly challenging due to the error-vector in [$E-L_{\rm{z}}\ $]{}and similar spaces that dramatically scatters structures with high angular momentum (see Appendix \[appendix:errorvec\]), highlighting the importance of leveraging chemistry in this region of phase-space. We defer detailed characterization of these structures to forthcoming work. ### Wukong {#subsec:wukong} ![image](Fig16.pdf){width="\linewidth"} ![image](Fig17.pdf){width="\linewidth"} Here we present Wukong[^5], a hitherto unknown prograde structure, that appears as a pair of overdensities in [$E-L_{\rm{z}}\ $]{}($E_{\rm{tot}}/[10^{5} \rm{km}^{2}/\rm{s}^{2}]$ = -1.1, -1.3), lining the prograde margin of GSE. In Appendix \[appendix:pot\] we show these clumps to be even more pronounced in the @McMillan17 potential. We select Wukong as follows: $$\label{eq:wukong} \begin{aligned} (-1<L_{\rm{z}}/[\rm{10^{3} kpc\ km\ s^{-1}}]<-0.2) \\ \land\ (-1.35<E_{\rm{tot}}/[10^{5} \rm{km}^{2}/\rm{s}^{2}]<-0.9)\\ \land\ (\rm{[Fe/H]}<-1.45)\\ \land\ (\rm{excluding\ all\ previously\ defined\ structures}). \end{aligned}$$ The conditions in [$E-L_{\rm{z}}\ $]{}draw a box around the overdensities and extend it inwards towards GSE. The \[Fe/H\] selection is motivated by the MDF resulting from the [$E-L_{\rm{z}}\ $]{}cut (dashed histogram in MDF panel of Figure \[fig:wukong\]) that shows multiple peaks at \[Fe/H\]$=-1.2$ (GSE),$-1.6$,$-1.9$, and a break at $\approx-1.45$. This leaves us with a sample of 111 Wukong stars that form a sequence in \[Fe/H\] vs \[$\alpha$/Fe\]. Three metal-poor GCs – NGC 5024 (\[Fe/H\]$=-2.1$), NGC 5053 (\[Fe/H\]$=-2.5$), ESO 280-SC06 (\[Fe/H\]$=-2.5$) – satisfy all the selection criteria in Eq. \[eq:wukong\], and may have been accreted along with Wukong (phase-space parameters from @Baumgardt19 and abundances from @Boberg15 [@Boberg16; @Simpson19b]). @Massari19 attribute NGC 5024, NGC 5053 to the Helmi Streams and ESO 280-SC06 to GSE. We note however, that none of these GCs satisfy the Helmi Streams selection from @Koppelman19HS that we also use in this work (Eq. \[eq:HS\]), and that ESO 280-SC06 (\[Fe/H\]$=-2.5$, $e=0.66$) has properties only marginally consistent with the GSE stars in the sample. That multiple GCs with metallicities consistent with Wukong are aligned with it in phase-space is a promising sign that it is a genuine structure. ### Metal-Weak Thick Disk {#subsec:mwtd} The metal-poor tail of the high-$\alpha$ disk was not included in our earlier selection of the high-$\alpha$ disk and in-situ halo in chemical space (Figure \[fig:insitu\]). Metal-weak thick disk (MWTD)[^6] stars are expected to fall right next to the high-$\alpha$ disk in \[Fe/H\] vs \[$\alpha$/Fe\]. They are more metal poor than the high-$\alpha$ disk but are at similar $\alpha$ and are rotationally supported – i.e., prograde and with strong $J_{\rm{\phi}}$ (top-left panel of Figure \[fig:confusogram\_feh\]). This motivates the MWTD selection: $$\begin{aligned} (-2.5<\rm{[Fe/H]}<-0.8) \land (0.25<\rm{[\alpha/Fe]}<0.45)\\ \land\ (J_{\phi}/J_{\rm{tot}}<-0.5)\\ \land\ (\rm{excluding\ all\ previously\ defined\ structures}). \end{aligned}$$ The stars that satisfy these cuts are shown in Figure \[fig:MWTD\]. They are mostly clustered very close to the high-$\alpha$ disk in \[Fe/H\] vs \[$\alpha$/Fe\] at \[Fe/H\]$\approx-0.8$. These stars support the finding of @Carollo19 that while the MWTD may be a prominent component of the $\|Z_{\rm{gal}}\|<3$ kpc Galaxy, it is only a minor component at larger distances (weighted fraction of $<5\%$ at $\|Z_{\rm{gal}}\|>3$ kpc). ### Unclassified Debris {#subsec:leftovers} We have assigned $92\%$ (weighted) of our sample to the aforementioned structures. This leaves us with $8\%$ (weighted) that we designate as “unclassified debris". The unclassified debris is depicted in Figure \[fig:leftovers\]. A prograde population that closely follows the contours of the high-$\alpha$ disk/in-situ halo in [$E-L_{\rm{z}}\ $]{}is evident at $E_{\rm{tot}}<-1.3, \|L_{\rm{z}}\|<1$ (compare with Figure \[fig:insitu\]). There is also an overlapping population, extending to higher energies ($E_{\rm{tot}}\approx-0.8$) with a high degree of rotational support ($J_{\rm{}\phi}/J_{\rm{tot}}<-0.75$). These disk-like stars did not meet the high-$\alpha$ disk/in-situ halo and MWTD cuts. These stars are either (i) the eccentric, metal-poor tail of the high-$\alpha$ disk (or the low eccentricity tail of GSE) that did not meet the rotational support criteria of the MWTD, or ii) rotationally supported stars that fall outside our high-$\alpha$ disk and MWTD chemistry selection boxes. We designate these stars as “disk-like" unclassified debris and they constitute $2\%$ (weighted) of the total sample. Then there are “halo-like" stars at higher-energy clustered around various selection boxes. Most of these stars have eccentricities between 0.6 and 0.7 ($5\%$ of the total sample, weighted). This concentration in eccentricity is noteworthy since we select GSE stars with a hard cut at $e>0.7$. This, and the prominent peak in the MDF at GSE’s metallicity strongly suggests these stars are $e<0.7$ members of GSE. The clumps of unclassified debris stars that appear at the locations of Thamnos and Wukong in [$E-L_{\rm{z}}\ $]{}bear this out: these stars satisfied the phase-space selections for these structures, but had GSE-like metallicity and were excluded via cuts on the MDF (see dashed MDFs in Figures \[fig:thamnos\], \[fig:wukong\]). This aspect of our work may be improved with more probabilistic methods of assigning membership that do not impose discontinuous selection boxes as we have done here [e.g., @Yuan20]. Some of the prograde, high-energy “halo-like" stars ($<1\%$ of the entire sample, weighted) are also clustered in a selection box corresponding to the Helmi Streams. These stars have similar $L_{\rm{z}}$ and energies but do not satisfy the $L_{\rm{\perp}}$ condition we imposed. These stars are plausible members of the Helmi Streams. Taking into account these likely associations, we are left with $\approx1\%$ (weighted) of the total sample as being either unclassified or unassociated. These stars may belong to low-mass structures that we sample too few stars from to detect coherent features. Or these stars may simply have bad stellar or orbital parameters. Either way, it is clear that we have identified the vast majority of structure in the halo as viewed by H3. ![image](Fig18.pdf){width="0.85\linewidth"} [lrrrrrrrrrr]{} \[table:summary\]\ Gaia-Sausage-Enceladus & 2684 & 0.42 & -1.15 & 0.21 & 0.84 & 17.72 & 12.88 & -0.50 & -1.04 & -0.01\ Sagittarius & 675 & 0.24 & -0.96 & 0.12 & 0.54 & 32.31 & 24.13 & 0.51 & -0.67 & -1.51\ High-$\alpha$ Disk + In-Situ Halo & 950 & 0.15 & -0.54 & 0.34 & 0.48 & 9.00 & 3.53 & -0.03 & -1.34 & -0.94\ Arjuna & 139 & 0.02 & -1.20 & 0.24 & 0.55 & 22.91 & 16.66 & 0.16 & -0.91 & 1.73\ Metal-Weak Thick Disk & 144 & 0.02 & -1.12 & 0.32 & 0.47 & 8.60 & 4.25 & 0.04 & -1.38 & -0.9\ Aleph & 122 & 0.02 & -0.51 & 0.19 & 0.13 & 11.06 & 3.51 & 0.07 & -1.13 & -2.36\ Wukong & 111 & 0.01 & -1.58 & 0.24 & 0.56 & 12.75 & 9.55 & 0.42 & -1.18 & -0.59\ Helmi Streams & 91 & 0.01 & -1.28 & 0.15 & 0.46 & 17.17 & 13.55 & 0.47 & -1.03 & -1.14\ Sequoia & 72 & 0.01 & -1.59 & 0.14 & 0.56 & 15.55 & 11.02 & 0.16 & -1.02 & 1.31\ I’itoi & 65 & 0.01 & -2.39 & 0.38 & 0.47 & 12.37 & 7.46 & 0.09 & -1.04 & 1.35\ Thamnos & 32 & 0.01 & -1.90 & 0.29 & 0.46 & 8.68 & 5.11 & 0.41 & -1.35 & 0.46\ Unclassified Debris (disk-like) & 208 & 0.02 & -1.16 & 0.22 & 0.53 & 8.63 & 4.72 & 0.13 & -1.36 & -0.52\ Unclassified Debris (halo-like) & 463 & 0.06 & -1.20 & 0.19 & 0.60 & 18.25 & 14.19 & 0.40 & -1.04 & -0.03\ ![image](Fig19.pdf){width="\linewidth"} ![image](Fig20.pdf){width="0.95\linewidth"} ![image](Fig21.pdf){width="0.95\linewidth"} Summary of Structure {#subsec:structuresummary} -------------------- In this section we present a synopsis of the structures identified in this work. 1. Gaia-Sausage-Enceladus ($f=0.42$[^7]): The radial, head-on merger that dominates the metal-poor local halo, GSE is largely contained within $\approx30$ kpc, and displays a narrow MDF (\[Fe/H\]$=-1.15^{+0.24}_{-0.33}$, weighted) reminiscent of the local dwarfs Leo I and Fornax. Its metallicity and proposed accretion redshifts imply a stellar mass of $4-7\times10^{8}M_{\rm{\odot}}$. Its mean rotation is consistent with zero: $\langle V_{\rm{\phi}}\rangle = 1.04^{+1.26}_{-1.25}\ \rm{km\ s^{-1}}$, $\langle L_{\rm{z}}\rangle = 4.7^{+20.1}_{-10.5}\ \rm{kpc\ km\ s^{-1}}$. \[§\[subsec:ge\], Figure \[fig:ge\]\] 2. Sagittarius ($f=0.24$): One of the first known streams, Sgr displays a uniquely high $\|L_{\rm{y}}\|$ that allows for a clean selection leveraging full 6D phase-space. Its MDF is multi-peaked with a pronounced metal-poor tail. \[§\[subsec:sgr\], Figure \[fig:sgr\]\] 3. High-$\alpha$ Disk and In-situ Halo ($f=0.15$): A major component of the local Galaxy, the high-$\alpha$ disk, and its high-eccentricity tail (the “in-situ" halo) extend out to $\|Z_{\rm{gal}}\|\approx15$ kpc. Their eccentricity distribution is continuous, ranging from very circular to highly eccentric, supporting scenarios in which the primordial high-$\alpha$ disk was disturbed by a merger event (likely GSE). \[§\[subsec:insitu\], Figure \[fig:insitu\]\] 4. Arjuna, Sequoia, I’itoi ($f=0.02,\ 0.01,\ 0.01$): The constituents of the high-energy retrograde halo – Arjuna (\[Fe/H\] $= -1.2$), Sequoia (\[Fe/H\] $= -1.6$), and I’itoi (\[Fe/H\] $<-2$) – are eccentric ($e\approx0.5-0.6$) and extend to highly retrograde orbits ($L_{\rm{z}}/[10^{3}\rm{\ kpc\ km\ s^{-1}}]>2$). Arjuna is the dominant component with $\gtrsim$2$\times$ the stars as Sequoia, and may be a distinct accreted structure or a hitherto unknown, highly retrograde extension of GSE. \[§\[subsec:retrograde\], Figure \[fig:retrograde\]\] 5. Metal-weak thick disk ($f=0.02$): The metal-poor extension of the high-$\alpha$ disk is only a minor component of the $\|Z\|>3$ kpc halo ($\lesssim5\%$) as suggested by local studies. \[§\[subsec:mwtd\], Figure \[fig:MWTD\]\] 6. Aleph ($f=0.02$): A highly circular structure ($e=0.13\pm0.06$) that rises $\approx$10 kpc off the plane. It is significantly enriched compared to typical halo structures (\[Fe/H\]$=-0.5$, \[$\alpha$/Fe\]$=0.2$), and may be associated with the enigmatic globular cluster Palomar 1. Whether it is an in-situ or ex-situ structure is under investigation. \[§\[subsec:aleph\], Figure \[fig:aleph\]\] 7. Wukong ($f=0.01$): Comprising the “prograde shards" of the halo, Wukong (\[Fe/H$=-1.7$\]) spans a wide range in energy and eccentricity reminiscent of massive structures like Sequoia and the Helmi Streams and is likely associated with the metal-poor GCs NGC 5024, NGC 5053, and ESO 280-SC06. \[§\[subsec:wukong\], Figure \[fig:wukong\]\] 8. Helmi Streams ($f=0.01$): Among the first halo structures discovered in integrals of motion, the Helmi Streams show a complex chemical population consistent with an extended star-formation history, have a stellar mass of $\approx0.5-1\times10^{8}M_{\rm{\odot}}$, and rise $\approx25$ kpc off the plane, as expected from local samples. \[§\[subsec:hs\], Figure \[fig:helmi\]\] 9. Thamnos ($f=0.01$): A recently discovered structure whose existence we confirm, Thamnos is among the most metal-poor structures in the halo (\[Fe/H\]=$-1.9$, \[$\alpha$/Fe\]=$0.3$). With a stellar mass that we estimate to be $\approx2\times10^{6}M_{\rm{\odot}}$, it is remarkable that Thamnos lies so deep in the potential ($r_{\rm{gal}}=9$ kpc), which makes it an exciting and accessible ($d_{\rm{helio}}=6$ kpc) target for near-field cosmology. \[§\[subsec:thamnos\], Figure \[fig:thamnos\]\] We have assigned $92\%$ (weighted) of our sample to these structures. Their properties are summarized in Table \[table:summary\]. In Figure \[fig:summary2\] we depict the various structures we have identified in [$E-L_{\rm{z}}\ $]{}, and in [$E-L_{\rm{z}}\ $]{}split by actions. As we have shown in this section, these populations that are clumped in [$E-L_{\rm{z}}\ $]{}not only occupy similar regions of phase-space, but also often define distinct chemical populations. Almost the entire halo can be accounted for as the superposition of these populations. Examining the remaining $8\%$, the “unclassified debris" (§\[subsec:leftovers\], Figure \[fig:leftovers\]), $2\%$ are “disk-like" and likely eccentric, metal-poor members of the high-$\alpha$ disk and in-situ halo. The remaining $5\%$ have higher-energy “halo-like" orbits. A large fraction of these ($\approx4\%$ of the sample) are plausible members of GSE and the Helmi Streams. Within the survey footprint, any remaining unidentified systems must comprise, in aggregate, no more than $\approx1\%$ of the high-latitude Galaxy within 50 kpc. Discussion {#sec:discussion} ========== Relative Fractions of Substructure and the Mass Function of Accreted Structure {#subsec:relfrac} ------------------------------------------------------------------------------ We are now in a position to examine which components of the halo are dominant at different distances. We depict the relative fractions of substructure, corrected for the selection function, as a function of Galactocentric distance and distance from the plane in Figure \[fig:summary1\]. Fractions are computed in running 5 kpc bins and smoothed with a Savitzky-Golay filter (10 kpc window, second-order polynomial) for clarity. The unclassified debris are included as grey bands on the top – note that we argued in §\[subsec:leftovers\] that a majority of these stars can be reasonably attributed to GSE. In agreement with local studies [e.g., @Bonaca17; @Haywood18; @DiMatteo19; @Amarante20; @Belokurov20] we find the high-$\alpha$ disk and its heated high-eccentricity tail (referred to as the “in-situ halo" in this work, and “Splash" in @Belokurov20) contribute the majority of stars at $\|Z_{\rm{gal}}\|\approx2$ kpc. We separate the high-$\alpha$ disk from the in-situ halo based on eccentricity ($e>0.5$) – this is an arbitrary cut, since these populations define a continuous distribution in eccentricity (Figure \[fig:insitu\]). The eccentric in-situ halo (blue hatched region) extends farther and rises to larger elevation than high-$\alpha$ disk. The high-$\alpha$ disk and in-situ halo fraction falls rapidly from $\sim50\%$ at $\|Z_{\rm{gal}}\|=2$ kpc to $<5\%$ beyond $\|Z_{\rm{gal}}\|=15$ kpc. At $\|Z_{\rm{gal}}\|\approx10$ kpc, GSE takes over and comprises $>50\%$ of the stars, and at farther distances, between $\|Z_{\rm{gal}}\|\approx25-50$ kpc, the majority of stars belong to Sagittarius. Similar trends are observed in the relative fractions as a function of $r_{\rm{gal}}$ as well. The span of GSE, largely contained within $35$ kpc, is in excellent agreement with observational estimates of its spatial extent [e.g., @Deason18; @Lancaster19]. Figure \[fig:summary1\] shows an uptick in the GSE relative fraction past 35 kpc – this is mostly due to the fractions of all non-Sgr structures falling off, and the large Poisson noise at these distances (depicted in Figure \[fig:summary1\] as a golden envelope). At all distances the other structures comprise $<25\%$ of the sample. Our fractional budget clearly confirms the prediction of various simulations that at $r_{\rm{gal}}<50$ kpc the accreted halo is built by a handful of massive ($M_{\rm{star}}=10^{8}-10^{9} M_{\rm{\odot}}$) progenitors (GSE and Sgr in the MW’s case) with a subdominant contribution from lower-mass galaxies and ultra-faints ($M_{\rm{star}}<10^{5} M_{\rm{\odot}}$) [e.g., @Deason16; @Santistevan20; @Fattahi20]. The three structures that together comprise $\gtrsim75\%$ of the sample at all distances (the high-$\alpha$ disk + in-situ halo, GSE, Sgr) have particularly secure selections. Sgr due to its recent accretion is highly coherent in phase-space, the high-$\alpha$ disk and in-situ halo occupy a unique location in \[Fe/H\] vs \[$\alpha$/Fe\], and the GSE MDF shows one clear component that is well-fit by a simple analytical model and is reminiscent of the narrow MDFs of local dwarfs like Fornax and Leo I [@Kirby13]. The robustness of these selections inspire confidence in our conclusion that the mass function of accreted material is indeed “top-heavy". In fact, the dominance of these three components is even more pronounced ($>80\%$) if a large fraction of the “unclassified debris", as we have argued in §\[subsec:leftovers\], is allocated to GSE. The Origin of the Stellar Halo {#subsec:origin} ------------------------------ As per our accounting of structure, the halo is built almost entirely by the accretion of dwarf galaxies (e.g., GSE, Sgr, Arjuna), and the response of the Galaxy to their accretion (e.g., the heating of the high-$\alpha$ disk). The in-situ halo is an important component at $r_{\rm{gal}}<10$ kpc, but its relative fraction rapidly falls off at larger radii (Figure \[fig:summary1\]). This can also be seen in Figure \[fig:accretedinsitu\] where we combine all the components that likely originated in the Galaxy (high-$\alpha$ disk and in-situ halo, the metal-weak thick disk, Aleph, the unclassified disk debris), and compare their extent to the accreted components. As we detail in the remainder of this section, our inventory of structure leaves little room for other proposed in-situ (e.g., a smooth “collapsed halo", outflows that deposit stars in the halo) or ex-situ components (e.g., dissolved globular clusters, an \[Fe/H\]$\sim-2.2$ spherical “outer halo"). ### The In-Situ Halo {#subsubsec:insitu} The mass budget and origin of the in-situ halo – not just the heated disk, but also stars forming from gas that is smoothly accreted, stripped from infalling galaxies, or ejected in outflows – is debated across simulations and may help constrain sub-grid physics such as star-formation and feedback prescriptions [e.g., @Font11; @Cooper15; @Pillepich18; @Fattahi20; @Yu20; @Font20]. The extent and relative fraction of the in-situ halo (in particular, the heated disk) are also sensitive probes of the accretion history of the MW [e.g., @Zolotov09; @Monachesi19]. We find that other than the heated high-$\alpha$ disk, the $r_{\rm{gal}}<50$ kpc halo does not contain any other in-situ populations (Figures \[fig:summary1\], \[fig:accretedinsitu\]). Aleph, which may have been heated or kicked from the disk may be an exception, but more work needs to be done to ascertain its nature. In any case, the high-$\alpha$ disk + in-situ halo and Aleph combined form a significant fraction of the halo only at $r_{\rm{gal}}\lesssim15$ kpc. We do not see a substantial fraction of eccentric stars with low-$\alpha$ or high-$\alpha$ disk-like chemistry in the distant halo (see hatched blue band in Figure \[fig:summary2\]) as might be expected if stars in outflows formed a significant component of the halo [e.g., @Yu20]. Nor do we see a significant smoothly accreted component built out of cooling gas from the circumgalactic medium and cosmological inflows whose relative fraction is comparable to accreted material [e.g., @Cooper15]. These stars would resemble a smooth, isotropic, relatively metal-poor halo from monolithic collapse and not show the cogent structure associated with chemical evolution in dwarf galaxies in \[Fe/H\] vs \[$\alpha$/Fe\]. The final category of in-situ halo stars, stars formed from stripped gas from a dwarf galaxy, would be very difficult to tell apart from the debris of the dwarf, since they are likely to be aligned in phase-space as well as chemically – progress could be made with precise ages to isolate stars that formed after the satellite’s infall, as has been done for Sgr [e.g., @Siegel07; @deboer15; @Alfaro-Cuello19]. The in-situ halo can also be used to constrain the MW’s merger history. @Monachesi19 found that galaxies in the Auriga simulation suite with high in-situ fractions ($>50\%$) beyond their optical radii ($\gtrsim30$ kpc) either underwent a recent violent merger, or a very early ($>8$ Gyr ago) merger that ejected disk stars to large radii. Similarly @Zolotov09 noted that simulated galaxies with quiescent merger histories (most of the mass in their halos was in place at $\sim9$ Gyrs) have a higher fraction of in-situ stars in their inner halo ($\sim20-50\%$) that rapidly tapers off by $\sim30$ kpc. This is very similar to what we find (Figure \[fig:accretedinsitu\]), suggesting the bulk of the halo was already in place at early times and that the MW’s recent growth has largely been quiescent (modulo Sgr, which is an important perturber of the disk, but due to its polar orbit, not a major contributor to the in-situ halo). This is a completely complementary way of accessing the MW’s merger history and is in excellent agreement with the picture of a quiet merger history at later times inferred from GCs [@Kruijssen19; @Kruijssen20], precise ages of the MW’s various components [@Bonaca20], and the presence of a prominent break in the density profile [@Deason13]. This qualitative finding can be further refined through more detailed comparisons with simulations after accounting for the simple H3 selection function. ### The Ex-Situ Halo The fact that the distant halo is clumpy and highly structured has long been interpreted as strong evidence for an accretion-origin of the halo [e.g., @Newberg02; @Bell08; @Starkenburg09; @Xue11; @Schlaufman12]. Here we confirm this picture, and further refine this finding by quantifying the proportions of various in-situ and ex-situ components. As described in §\[subsec:relfrac\] and Figure \[fig:summary1\] we find the accreted component almost entirely arises from a handful of massive ($M_{\rm{\star}}\sim10^{8}-10^{9}M_{\rm{\odot}}$) dwarfs. In the rest of this section we examine two popular scenarios about the nature of the accreted component in the context of our findings (the “dual halo" and disrupting GCs). @Carollo07 [@Carollo10] and @Beers12 used a local sample ($d_{\rm{helio}}<4$ kpc) and integrated orbits to infer that the halo was best described as a “dual halo" [but see @Schonrich11]. In their picture the dual halo is comprised of an inner in-situ halo ($r_{\rm{gal}}\lesssim15$ kpc, \[Fe/H\]$=-1.6$, small net prograde motion, high eccentricity) and an outer accreted halo ($r_{\rm{gal}}\sim20-50$ kpc,\[Fe/H\]$=-2.2$, mean retrograde motion, wide range of eccentricities) that are two fundamentally different populations. With the benefit of a post-Gaia perspective, @Helmi20 interpret the outer retrograde halo as GSE with a steep metallicity gradient and the inner halo as the heated high-$\alpha$ disk. This is partially motivated by the Gaia color-magnitude diagram in the local halo which shows two prominent sequences that have been attributed to GSE and the heated high-$\alpha$ disk. @Belokurov20 on the other hand argue the “inner halo" is in fact GSE. Figures \[fig:summary2\] and \[fig:accretedinsitu\] help clarify this debate. The “inner halo" ($r_{\rm{gal}}\lesssim15$ kpc) is predominantly built by GSE and the heated high-$\alpha$ disk, with GSE contributing a larger relative fraction. This is exactly as expected from simulations that find the inner halo to be a mixture of heated disk stars and accreted material, with the proportion varying with details of the accretion history of the galaxy [e.g., @Zolotov09; @Tissera14; @Monachesi19]. As for the @Carollo10 “outer halo" ($r_{\rm{gal}}\sim20-50$ kpc), after setting Sagittarius aside since it does not pass through the local halo, GSE is still a major component, but the retrograde Arjuna, Sequoia, and I’itoi contribute a significant fraction too, and perhaps explain the finding of a net retrograde motion. Though we note that considering only these structures at $r_{\rm{gal}}\sim20-50$ kpc produces only a mildly retrograde $\langle V_{\rm{\phi}}\rangle=12.5^{+6}_{-6}\ \rm{km}\ \rm{s}^{-1}$ (weighted). Also, at no distance does a very metal-poor (e.g., \[Fe/H\]$<-2$) component comprise a significant fraction of the Galaxy. We conclude by noting that no single population – neither the in-situ halo nor GSE – neatly maps onto either component of the dual halo, and more work needs to be done to understand the effects of extrapolating the nature of the distant Galaxy from energetic local halo orbits. Several authors have hypothesized that stars born in GCs might contribute significantly to the stellar halo mass budget – with estimates ranging from $\lesssim10\%$ to $\lesssim50\%$ [e.g., @Gnedin97; @Schaerer11; @Martell11; @Martell16; @Carretta16; @Schiavon17; @Koch19]. GCs are attractive candidates for building up at least some fraction of the halo because several of them are in the process of being tidally disrupted [e.g., @Grillmair06; @Myeong17; @Shipp18; @Malhan18]. Further, several popular scenarios for GC formation assume they underwent drastic mass-loss at some point in their early history [see @Bastian18 for a recent review]. Post Gaia-DR2 almost all halo GCs (those at high-energy, that are not on disk-like orbits) have been associated with phase-space structures seen in stars, strongly suggesting they were accreted along with some dwarf Galaxy [@Massari19; @Kruijssen19; @Myeong19; @Kruijssen20; @Forbes20]. This complicates the evaluation of the halo fraction arising from GCs, since in their phase-space coordinates these accreted GCs resemble field stars from their parent dwarf galaxies. However, there may still be a contribution from GCs associated with low-mass accreted dwarfs or ancient in-situ GCs that may have dissolved in the MW halo in the distant past. We find that at least within 50 kpc such GCs play a very limited role in building the halo. Being very conservative and allowing all the unclassified “halo-like" debris (§\[subsec:leftovers\]) to emanate from GCs limits their contribution to $<6\%$ (weighted) at all distances. This upper limit is in excellent agreement with high-resolution simulations [@Reina-Campos20] that find similarly low fractions ($2-5\%$), with recent searches for second-generation GC stars in the halo [@Koch19; @Hanke20] that find a low observed fraction of $2.6\pm0.2\%$ that they adjust to $<11\pm1\%$ to account for first-generation stars, and arguments based on BHB to blue-straggler ratios that found the halo ratio resembled dwarf galaxies and not GCs [@Deason15]. Prograde vs. Retrograde, and the Net Rotation of the Halo {#subsec:protretro} --------------------------------------------------------- In local halo studies there appears to be an asymmetry in the distribution of accreted stars, with more retrograde than prograde structure. For instance, @Helmi17 find $58-73\%$ of high-energy halo stars are retrograde [see also @Myeong18b; @Myeong18c]. More recently, @Yuan20 recovered six new retrograde “dynamically tagged groups" compared to two prograde groups in LAMOST DR3. These findings bear echoes of the @Carollo07 [@Carollo10] “dual halo", whose “outer halo" is retrograde. This asymmetry may simply be a selection effect, since it is easier to avoid contamination from the disk and in-situ halo on the retrograde side, and because some structures, such as Sgr, are not represented in the local halo. It might also be physical – some models predict that the disk is more efficient at mixing structure accreted on prograde orbits compared to retrograde orbits [e.g., @Quinn86; @Byrd86; @Norris89; @donghia10]. We observe no significant asymmetry in the distant halo. Setting the disk populations and unclassified debris aside, we find three prograde (Sagittarius, Helmi Streams, Wukong) and four retrograde (Arjuna, Sequoia, I’itoi, Thamnos) accreted structures. GSE shows net rotation consistent with zero ( $\langle V_{\rm{\phi}}\rangle = 1.04^{+1.26}_{-1.25}\ \rm{km\ s^{-1}}$). If Sequoia and Arjuna are indeed associated with GSE, the net rotation combining these two components is weakly retrograde $\langle V_{\rm{\phi}}\rangle = 9.4^{+1.7}_{-1.6}\ \rm{km\ s^{-1}}$. In terms of relative fractions, after setting the $\langle V_{\rm{\phi}}\rangle \sim0$ GSE aside, prograde stars outnumber retrograde stars $\approx$3:2 (mostly due to Sgr, excluding it results in $\approx$1:1). The net rotation of accreted material (also counting the halo-like unclassified debris) is $\langle V_{\rm{\phi}}\rangle = -25.23^{+2.54}_{-2.71}\ \rm{km\ s^{-1}}$ with Sgr, and $\langle V_{\rm{\phi}}\rangle = 5.7^{+1.7}_{-1.6}\ \rm{km\ s^{-1}}$ when Sgr is excluded. All numbers quoted here have been weighted to correct for the selection function. Assuming an isotropic distribution of infalling satellites, the numbers in this section suggest that the prograde satellites are not more efficiently disrupted than retrograde satellites, and that there is no strong mean rotation signal. Interpreting the Halo in Chemical Space {#subsec:chemistry} --------------------------------------- The locations of various structures in \[Fe/H\] vs \[$\alpha$/Fe\], as listed in Table \[table:summary\], are depicted in Figure \[fig:FeHaFe\]. The markers representing the structures are sized linearly as per their weighted relative fractions. The markers are all square, and not intended to reproduce the spread in abundances – since we make hard cuts on the MDF to select some structures we are not well-positioned to make fair estimates of the spread. Broadly, in the \[Fe/H\]-\[$\alpha$/Fe\] plane, galaxies are expected to start off in the top-left ([\[$\alpha$/Fe\]]{}-rich and [\[Fe/H\]]{}-poor), and end up towards the bottom right ([\[$\alpha$/Fe\]]{}-poor and [\[Fe/H\]]{}-rich) as they evolve and Type Ia supernovae take over from Type II supernovae as the chief pollutants of the ISM [e.g., @Tinsley80; @Matteucci86; @Maiolino19]. This journey is interrupted when a galaxy is accreted and shredded by the Milky Way, and to first order its abundances are frozen in place. We highlight some features of this space, while noting that detailed modeling [e.g., @Fernandez-Alvar18; @Vincenzo19; @Lian20] is required to deduce finer details. Thamnos and I’itoi are the most $\alpha$-rich and metal-poor structures. Based on its depth in the potential, we argued Thamnos was accreted at high-redshift ($z\approx1.5$), i.e., early in its chemical evolution. This may be the case for I’itoi as well, though it occurs at higher energy, which may mean it was accreted relatively recently but is simply very low-mass and formed stars inefficiently. At the other extreme of the plot, Sagittarius is the most $\alpha$-poor and metal-rich of all the structures in the halo, likely because it was accreted very recently ($z<1$, e.g., @Laporte18 [@Kruijssen20; @Ruiz-Lara20]), after undergoing significant enrichment by Type Ia supernovae. GSE, with mass comparable to Sagittarius, is relatively $\alpha$-enhanced and metal-poor, which fits with the recent finding that it began interacting with the Milky Way at $z\approx2$ [@Bonaca20], when we expect its chemical evolution to have been interrupted. The Helmi Streams, accreted at $z\sim0.5-1$ [@Koppelman19HS], i.e., at a similar epoch as Sgr, have \[$\alpha$/Fe\] similar to Sgr, but are about 0.3 dex more metal-poor, exactly as expected for a structure $\sim10\times$ less massive [@Lee15 their Figure 2]. Wukong, Sequoia, and Arjuna are at intermediate locations between the two extremes of I’itoi and Sagittarius, and finer estimates of their masses and accretion redshifts are required to further interpret their \[Fe/H\] vs \[$\alpha$/Fe\] locations. This figure is another way to see that the halo, at least out to 50 kpc, as seen in our sample, is relatively metal-rich compared to several earlier studies [e.g., @Carollo07; @Carollo10; @Xue15; @Das16b; @Liu18]. The three main components – GSE, Sgr, the high-$\alpha$ disk and in-situ halo (those with the largest marker sizes in Figure \[fig:FeHaFe\]) all lie almost entirely at \[Fe/H\]$>-1.5$. Only a small fraction of material can be attributed to the debris of metal-poor structures like I’itoi and Thamnos (\[Fe/H\]$\lesssim-2$). Even setting our definitions of various structures aside, we compute $\langle\rm{[Fe/H]}\rangle=-1.18^{+0.01}_{-0.01}$ (weighted) for all the accreted material taken together – i.e., Sgr, GSE, Arjuna, Helmi Streams, Sequoia, Wukong, I’itoi, Thamnos, the unclassified halo-like debris [see also @Conroy19b]. Caveats and Limitations {#subsec:caveats} ----------------------- Our census of the stellar halo is incomplete owing to the H3 Survey field locations, which are currently restricted to $\|b\|>40^\circ$ and Dec.$>-20^\circ$ (see Figure \[fig:data-1\]). An accreted structure completely confined to in-plane orbits or the bulge, e.g., the “ex-situ disk" [@Gomez17], “Kraken"/“Koala" [@Kruijssen19; @Kruijssen20; @Forbes20], would be missed. Furthermore, we systematically under-count stars from structures on orbits that spend most of their time close to the plane or in the Southern Hemisphere. Finally, recently accreted structures that have a strong on-sky coherence may be missed or biased in our existing fields. Assessing and correcting these biases will be the subject of future work. The number of halo stars in our current sample sets an effective limit on the lowest stellar mass system we could plausibly detect. We can estimate the mass-completeness as follows: if the halo has $\sim10^{9}$ stars [e.g., @Deason19; @Mackereth20], and we are tracing it with 5684 stars, for every structure with $\sim175,000$ stars we find 1 star in the survey (assuming all structures are isotropic and completely mixed – in detail we observe more stars from nearby structures). That is, from a $10^{6} M_{\odot}$ accreted galaxy we expect $\sim10$ stars in the sample. The detectability of such a structure would depend strongly on its location in phase-space/chemistry, e.g., the $\sim10^{6}\, M_{\odot}$ Thamnos stands out due to being very metal-poor/$\alpha$-rich in a region of phase-space that is populated by metal-rich GSE stars. Even if we are currently unable to identify some low-mass structures as distinct components of the halo, it is clear that taken together they play only a subdominant role in the overall mass budget (Figures \[fig:summary1\], \[fig:FeHaFe\], §\[subsec:relfrac\]). Another limitation of this work is our decision to apply hard cuts to select various structures. Due to this choice we miss the tails of various distributions. This is particularly evident for GSE, whose low-eccentricity ($e<0.7$) tail appears as a contaminant in e.g., the initial phase-space selections for Thamnos and Wukong (see dashed histogram in the MDF panels of Figure \[fig:thamnos\], \[fig:wukong\]). These GSE stars end up classed as “unclassified halo debris". On the other hand, because we attribute all $e>0.7$ stars to GSE (after excluding the in-situ halo and Sgr) we miss the high-eccentricity tails of all the structures that follow it in our inventory. These stars are likely a very insignificant fraction of the GSE sample, as our GSE MDF is smooth, well-behaved, and modeled well as a single population. However, what is a small fraction for GSE might be a significant fraction of the other low-mass structures. This aspect of our work may be improved with probabilistic methods (discussed in §\[subsec:inventory\]), but for now we list it as a caveat. Finally, we caution that individual structures identified in this work may not necessarily correspond to unique, accreted dwarf galaxies. It is possible that e.g., Wukong is comprised of multiple sub-populations corresponding to the modes in its MDF, or that GSE and Arjuna are linked. Simulations show the same accreted structure can deposit stars in surprisingly disparate regions of phase-space (e.g., @Jean-Baptiste17 [@Lilleengen20; @Elias20]), though this is typically not the case [e.g., @bj05_1; @bj05_3; @Pfeffer20]. Further analysis of these individual structures is necessary in order to link them to unique accreted systems. Summary {#sec:summary} ======= We have used the H3 Survey in combination with Gaia data to conduct a detailed census of substructure beyond the solar neighborhood. Our sample extends to $50$ kpc, is unbiased in metallicity, arises from a simple selection function, and has full 6D phase-space coordinates along with \[Fe/H\] and \[$\alpha$/Fe\]. We find the following: - The distant Galaxy displays a high degree of structure in integrals of motion (energy, actions, angular momenta) and chemistry (\[Fe/H\], \[$\alpha$/Fe\]) – spaces in which co-eval stars are expected to cluster for timescales longer than the age of the universe. \[Figures \[fig:feh\_slices\], \[fig:feh\_orbit\_slices\], \[fig:confusogram\_feh\]\] - $92\%$ of our sample can be assigned to one of the following structures: Sagittarius, Aleph, the high-$\alpha$ disk + in-situ halo (the heated high-$\alpha$ disk), the Helmi Streams, Thamnos, Arjuna, Sequoia, I’itoi, Wukong, Gaia-Sausage-Enceladus, and the metal-weak thick disk – our key findings on each structure are distilled in §\[subsec:structuresummary\]. This leaves us with $8\%$ of the sample (“unclassified debris", $2\%$ disk-like, and $6\%$ halo-like) that can be largely accounted for as artifacts of our sharp selection boundaries. \[§\[subsec:structuresummary\], Table \[table:summary\]\] - The high-$\alpha$ disk, the in-situ halo, GSE, and Sgr account for $\gtrsim75\%$ of all stars at all distances. The high-$\alpha$ disk and in-situ halo are a major component at $\lesssim10$ kpc ($\approx50\%$), but their relative fraction rapidly declines to $\lesssim10\%$ beyond 15 kpc. GSE dominates between $\approx15-25$ kpc and Sgr forms the bulk of the halo beyond $30$ kpc. The accreted halo within 50 kpc is therefore mainly built out of a small number of $10^{8}-10^{9}M_{\odot}$ galaxies (GSE, Sgr). That is, the mass function of accreted material is “top-heavy". This explains the metallicity of the halo (\[Fe/H\]$\approx-1.2$, see also @Conroy19b) that we find to be more metal-rich than several previous studies. \[§\[subsec:relfrac\], Figures \[fig:summary1\], \[fig:accretedinsitu\], \[fig:FeHaFe\]\] - This inventory of substructure leaves very limited room for other proposed constituents of the halo including a spherical, retrograde, \[Fe/H\]$\sim-2.2$ “dual halo" beyond 25 kpc, dissolved globular clusters, stars deposited by outflows, or stars born from smoothly accreted gas. \[§\[subsec:origin\], Figure \[fig:summary1\]\] - There is no preference for retrograde orbits in the distant Galaxy as has been observed in the local halo. GSE shows net rotation consistent with zero ($V_{\rm{\phi}}=1.0\pm1.3\ \rm{km\,s^{-1}}$). In fact, setting the disk populations aside, prograde stars outnumber retrograde stars $\approx3:2$. \[§\[subsec:protretro\], Table \[table:summary\]\] It has long been recognized that the distant halo is highly structured, and that this likely indicates an accretion-origin. Here, we have confirmed this picture, and further refined it by quantifying the exact proportions and extents of various in-situ and ex-situ components. In forthcoming work we will present detailed characterizations of the identified structures. With future Gaia data releases we will extend this work even further into the halo using the $\sim1000$ H3 giants extending out to $100$ kpc that were excluded from this work due to uncertain proper motions. By resolving the stellar halo into its constituent pieces we are delivering on the promise of Galactic Archaeology as a powerful tool to determine the assembly history of our Galaxy. Error Propagation in Phase Space {#appendix:errorvec} ================================ Here we explore how measurement errors – the $\lesssim10\%$ distance uncertainty, and the error on Gaia PMs – distort substructure in phase-space. We use stellar halos built through hierarchical accretion from the @bj05_1 [@bj05_2; @bj05_3] simulations. These halos feature a realistic, evolving potential, including a disk component. Using the default settings of `Galaxia`, a code to generate synthetic surveys of the Milky Way from analytical and N-body models [@Sharma11], we generate an H3-like survey – with a $10\%$ error on distances, and errors on PMs as per the Gaia DR2 error model. Potential-dependent phase-space quantities are computed using the $z=0$ potential described in @bj05_1. Figure \[fig:bj05\_02\] shows [$E-L_{\rm{z}}\ $]{}diagrams for three halos, both for the noiseless mock catalogs (top panels) and noisy mocks (center and bottom panels). We also highlight GSE-like, Sgr-like, and high-energy retrograde halo-like progenitors. While GSE-like and Sagittarius-like progenitors, which have comparatively lower $L_{\mathrm{z}}$, retain their general morphology, the strongly retrograde progenitors are significantly smeared out along diagonal tracks in [$E-L_{\rm{z}}$]{}. This is likely why we find it difficult to differentiate between Arjuna, Sequoia, and I’itoi in phase-space even though they are chemically distinct. By comparing the center panels (noisy PMs and noisy distances) and bottom panels (perfect PMs and noisy distances) we see the distance errors are the most significant component of the error budget. ![Comparison of mocks from three different halos in the @bj05_1 suite with no errors (top), with distance and PM errors (center), and with only distance errors (bottom). Sgr-like (navy blue), high-energy retrograde halo-like (brown), and GSE-like (golden) progenitors are highlighted here. While the Sgr-like and GSE-like locii retain their morphology to first order across all three rows, the retrograde progenitors with high $L_{\rm{z}}$ are dispersed dramatically along a diagonal track in [$E-L_{\rm{z}}$]{}. This is likely why disambiguating the various components of the high-energy retrograde halo (Arjuna+Sequoia+I’itoi) purely in phase-space without relying on chemistry is challenging. The center and bottom panels are virtually indistinguishable, emphasizing that the $10\%$ distance error is the dominant piece of the error budget.[]{data-label="fig:bj05_02"}](Fig22.pdf){width="0.8\linewidth"} Comparison with alternate potential {#appendix:pot} =================================== Here we provide a comparison against the @McMillan17 potential which is widely employed in the halo literature [e.g., @Myeong19; @Koppelman19] and features a more massive Milky Way than in the adopted fiducial potential ($1.3\times10^{12}M_{\rm{\odot}}$ versus $9.9\times10^{11}M_{\rm{\odot}}$ within 200 kpc), with several differences in how the potential is parametrized (thick and thin disks, gas disks, a different form for the bulge). Figure \[fig:M17\] allows for a straightforward visual conversion between the locations of various substructures across these two potentials. $L_{\rm{z}}$ is independent of the potential and so is the same in the left and right panels. It is encouraging that the structures identified in our fiducial potential remain coherent and well-defined in the alternative potential. ![[$E-L_{\rm{z}}\ $]{}diagrams in the fiducial potential (left) compared against the @McMillan17 potential (right). In the top row we show the data as is, and in the bottom row colored as per the substructure we have assigned these stars to (same as Figure \[fig:summary2\]). [$E-L_{\rm{z}}\ $]{}appears clumpier in the @McMillan17 potential due to the larger virial mass [e.g., @Sanderson15]. Importantly, all the proposed structures (e.g., Wukong, Thamnos, the retrograde shards) correspond to clear clumps and overdensities in this potential.[]{data-label="fig:M17"}](Fig23.pdf){width="0.9\linewidth"} [^1]: By “local halo" we mean the portion of the kinematic halo within a few kpc from the Sun that is typically selected using 3D Galactocentric velocity (e.g., $\|V-V_{\rm{LSR}}\|>210\ \rm{km\ s^{-1}}$, @Helmi18) with a view to avoid the disk. [^2]: Named for its prominence in [\[$\alpha$/Fe\]]{}$-$[\[Fe/H\]]{}; see Figure \[fig:confusogram\_feh\]. [^3]: Arjuna is named for the legendary archer from the Indian epic, the Mahabharata. Arjuna extends to high-energy, mirroring Sagittarius (Latin for archer) on the prograde side. [^4]: I’itoi (pronounced “ee ee thoy") is named for the “man in the maze" who features in creation legends of the Tohono O’odham people. Our survey telescope, the MMT Observatory, stands on the ancestral lands of the Tohono O’odham. Further, I’itoi is said to reside in a cave adjacent to a mountain, paralleling the location of I’itoi in [$E-L_{\rm{z}}\ $]{}with respect to GSE (Enceladus is entombed within Mt. Etna in Sicily). [^5]: Named for Sun Wukong, the celestial Monkey King from the Journey to the West. Sun Wukong is imprisoned under a mountain by the Buddha for his uprising against Heaven and is later set free by the scholar Tripitaka. We play the role of the scholar here, setting Wukong free from underneath Gaia-Sausage-Enceladus (Enceladus is entombed within Mt. Etna in Sicily). [^6]: Following previous work, we refer to this population as the metal-weak thick disk. However, given our selection it might be more appropriate to refer to this structure as the “metal-weak high-$\alpha$ disk". [^7]: Fraction of stars assigned to the structure, weighted to account for the selection function.	{ "pile_set_name": "ArXiv" }
--- abstract: 'In this paper, we consider the solutions of Einstein gravity in the presence of a generalized Maxwell theory, namely power Maxwell invariant. First, we investigate the analogy of nonlinear charged black hole solutions with the Van der Waals liquid–gas system in the extended phase space where the cosmological constant appear as pressure. Then, we plot isotherm $P$–$V$ diagram and study the thermodynamics of AdS black hole in the (grand canonical) canonical ensemble in which (potential) charge is fixed at infinity. Interestingly, we find the phase transition occurs in the both of canonical and grand canonical ensembles in contrast to RN black hole in Maxwell theory which only admits canonical ensemble phase transition. Moreover, we calculate the critical exponents and find their values are the same as those in mean field theory. Besides, we find in the grand canonical ensembles universal ratio $\frac{P_{c}v_{c}}{T_{c}}$ is independent of spacetime dimensions.' author: - 'S. H. Hendi$^{1,2}$[^1] and M. H. Vahidinia$^{1,3}$[^2]' title: \| Extended phase space thermodynamics and\ $P$–$V$ criticality of black holes with a nonlinear source --- Introduction ============ Theoretically one may be expect the cosmological constant term to arise from the vacuum expectation value of a quantum field and hence can vary. Therefore, it may be considered in the first law of thermodynamics with its conjugate [@Gibbons1; @BrownPLB1987; @CaldarelliCQG2000]. By this generalization, the cosmological constant and its conjugate can be interpreted as geometrical pressure and volume of a black object system, respectively. Moreover, this approach leads to an interesting conjecture on reverse isoperimetric inequality for black holes in contrast to a Euclidean version of isoperimetric inequality. Regarding the inequality conjecture, some of the black hole processes may be restricted [@RayCQG2009; @Gibbons2; @PVpapers]. Furthermore, the extension of thermodynamic phase space has dramatic effects on the studying of famous phase transition of black holes in AdS space [MyersPRD1999,Banerjee:2011au,Wu2012]{} and improves the analogy between small/large black hole with the Van der Waals liquid/gas phase transitions. Indeed, the AdS charged black holes exhibit an interesting phase transition with the same critical behavior as Van der Waals model, qualitatively [PVpapers]{}. Taking into account the above statements, we should note that the charge of the black hole plays a crucial role in this phase transition. Therefore it is important to know effects of any modification in the electromagnetic field. Indeed, some characteristic features of universality class of phase transitions such as the value of critical exponents or universal ratio $\frac{% P_{c}v_{c}}{T_{c}}$ may depend on electromagnetic source or spacetime dimension. On the other hand, considering strong electromagnetic field in regions near to point-like charges, Dirac suggested that one may have to use generalized nonlinear Maxwell theory in those regions [@Dirac]. Similar behavior may occur in the vicinity of neutron stars and black objects and so it is expected to consider nonlinear electromagnetic fields with an astrophysical motive [@Bialynicka]. In addition, within the framework of quantum electrodynamics, it was shown that quantum corrections lead to nonlinear properties of vacuum which affect the photon propagation [@Heisenberg; @Delphenich; @Schwinger; @Stehle]. Moreover, the effects of Born–Infeld (BI) source in the thermodynamics and phase transition of black hole [@ThermoBI] have been studied. Besides, in context AdS/CFT some authors consider roles of BI source on shear viscosity [@Sun08] and holographic superconductors [@AdSCFTBI]. By this observations one may find it is worthwhile to study the effects of nonlinear electrodynamics (NLEDs) on phase transition of black holes in the extended phase space. In this direction, the effects of nonlinear electromagnetic field of static and rotating AdS black holes in the extended phase space have been analyzed [@PVnonlinear]. It has been shown that for the BI black holes, one may obtain the same qualitative behavior as RN black holes. Indeed, BI electromagnetic field does not have any effect on the values of critical exponents, but it changes the universal ratio $\rho_{c}=% \frac{P_{c}v_{c}}{T_{c}}$ [@PVnonlinear]. Although BI theory is a specific model in the context of NLEDs, the recent interest on the NLEDs theories is mainly due to their emergence in the context of low-energy limit of heterotic string theory or as an effective action for the consideration of effects loop corrections in QED where quartic corrections of Maxwell field strength appear [@Kats]. In the last five years, a class of NLEDs has been introduced, the so-called power Maxwell invariant (PMI) field (for more details, see [@PMIpapers1; @PMIpapers2]). The PMI field is significantly richer than that of the Maxwell field, and in the special case ($s=1$) it reduces to linear electromagnetic source. The black hole solutions of the Einstein-PMI theory and their interesting thermodynamics and geometric properties have been examined before [@PMIpapers1; @PMIpapers2]. In addition, in the context of AdS/CFT correspondence, the effects of PMI source on strongly coupled dual gauge theory have been investigated [@AdSCFTPMI]. The bulk action of Einstein-PMI gravity has the following form [PMIpapers2]{} $$I_{b}=-\frac{1}{16\pi }\int_{M}d^{n+1}x\sqrt{-g}\left( R+\frac{n(n-1)}{l^{2}}% +\mathcal{L}_{PMI}\right) , \label{Action}$$where $\mathcal{L}_{PMI}=(-\mathcal{F})^{s}$ and $\mathcal{F}=F_{\mu \nu }F^{\mu \nu }$. Before we proceed, we provide some of reasonable motivation for considering this form of NLEDs. First, between NLEDs theories, the PMI theory is a toy model to generalize Maxwell theory which reduces to it for $s=1$. One of the most important properties of the PMI model in $(n+1)$-dimensions occurs for $% s=(n+1)/4$ where the PMI theory becomes conformally invariant and so the trace of energy-momentum tensor vanishes, the same as Maxwell theory in four-dimensions [@PMIpapers1]. Considering this value for the nonlinearity parameter, $s$, one can obtain inverse square law for the electric field of charged pointlike objects in arbitrary dimensions (the same as Coulomb’s field in four-dimensions). Furthermore, it has been shown that there is an interesting relation between the solutions of a class of pure $F(R)$ gravity and those of conformally invariant Maxwell source ($% s=(n+1)/4$) in Einstein gravity [@CIMFR]. Second, we should note that considering the $E_{8}\times E_{8}$ heterotic string theory, the $SO(32)$ gauge group has a $U(1)$ subgroup. It has been shown that [@StrinNL] taking into account a constant dilaton, the effective Lagrangian has Gauss-Bonnet term as well as a quadratic Maxwell invariant in addition to the Einstein-Maxwell Lagrangian. Since, unlike the quadratic Maxwell invariant, the Gauss-Bonnet term becomes a topological invariant and does not give any contribution in four-dimensions, it is natural to investigate Einstein-NLEDs in four dimensions. Taking into account a PMI theory as a NLEDs Lagrangian and expanding it for $\mathcal{F\longrightarrow F}_{0}$ (where we considered $\mathcal{F}_{0}$ as a unknown constant which we should fix it.), we find $$\mathcal{L}_{PMI}\simeq -a_{1}\mathcal{F}+(s-1)\left[ a_{0}+a_{2}(-\mathcal{F% })^{2}+a_{3}(-\mathcal{F})^{3}+...\right] . \label{Expand}$$In other words, one can consider series expansion of $\mathcal{L}_{PMI}$ near a constant $\mathcal{F}_{0}$ and obtain Eq. ([Expand]{}), in which the constants $a_{i}$’s are depend on $s$ and $\mathcal{F% }_{0}$. In order to obtain $\mathcal{F}_{0}$ and also have a consistent series expansion with linear Maxwell Lagrangian ($s=1$), one should set $a_{1}=1$. Taking into account $a_{1}=1$ and obtaining $\mathcal{F}_{0}$, we are in a position to get a new series expansion for $\mathcal{L}_{PMI}$ $$\mathcal{L}_{PMI}\simeq -\mathcal{F}+(s-1)\left[ b_{0}+b_{2}(-\mathcal{F}% )^{2}+b_{3}(-\mathcal{F})^{3}+...\right] , \label{Expand2}$$where $b_{i}$’s are only depend on $s$. Although, $\mathcal{L}_{PMI}=(-% \mathcal{F})^{s}$ can lead to Eq. (\[Expand2\]) by a series expansion, working with Eq. (\[Expand2\]) is more complicated and we postpone the study of this scenario to another paper. Third, taking into account the applications of the AdS/CFT correspondence to superconductivity, it has been shown that the PMI theory makes crucial effects on the condensation as well as the critical temperature of the superconductor and its energy gap [@AdSCFTPMI]. Motivated by the recent results mentioned above, we consider the PMI theory to investigate the effects of nonlinearity on the extended phase space thermodynamics and $P$–$V$ criticality of the solutions. Moreover, to better understand the role of nonlinearity, we relax the conformally invariant constraint and take $s$ as an arbitrary constant. It helps us to have a deep perspective to study the universal behavior of large/small black hole phase transitions. In particular, we are keen on understanding sensitivity of the critical exponents, universal ratio and other thermodynamic properties to nonlinearity parameter, $s$. Outline of this paper is as follows: In Sec. \[BH\], we consider spherically symmetric black hole solutions of Einstein gravity in the presence of the PMI source. Regarding the cosmological constant as thermodynamic pressure, we study thermodynamic properties and obtain Smarr’s mass relation. In Sec. \[Cano\], we investigate the analogy of black holes with Van der Waals liquid–gas system in the grand canonical ensemble by fixing charge at infinity. In this ensemble we find the free energy and plot the coexistence curve of a small/large black hole. Then, we calculate the critical exponent and find they match to mean field value (same as a Van der Waals liquid). Moreover, we consider the special case $s=n/2$ as BTZ-like solution, study its phase transition and show the critical exponents are the same as former case. In Sec. \[GCano\], we consider the possibility of the phase transition in the grand canonical ensemble and find that in contrast to RN black holes, the phase transition occurs for $s \neq 1$. Finally, we finish this work with some concluding remarks. Extended phase-space thermodynamics of black holes with PMI source {#BH} ================================================================== We consider a spherically symmetric spacetime as $$ds^{2}=-f(r)dt^{2}+\frac{dr^{2}}{f(r)}+r^{2}d\Omega _{d-2}^{2}, \label{Metric}$$ where $d\Omega _{d}^{2}$ stands for the standard element on $S^{d}$. Considering the field equations following from the variation of the bulk action with Eq. (\[Metric\]), one can show that the metric function $f(r)$, gauge potential one–form $A$ and electromagnetic field two–form $F$ are given by [@PMIpapers2] $$\begin{aligned} f(r) &=&1+\frac{r^{2}}{l^{2}}-\frac{m}{r^{n-2}}+\frac{(2s-1)^{2}\left( \frac{% (n-1)(2s-n)^{2}q^{2}}{(n-2)(2s-1)^{2}}\right) ^{s}}{% (n-1)(n-2s)r^{2(ns-3s+1)/(2s-1)}}, \label{metfunction} \\ A &=&-\sqrt{\frac{n-1}{2(n-2)}}qr^{(2s-n)/(2s-1)}dt, \label{A} \\ F&=&dA. \label{dA}\end{aligned}$$ The power $s \neq n/2$ denotes the nonlinearity parameter of the source which is restricted to $s>1/2$ [@PMIpapers2], and the parameters $m$ and $q$ are, respectively, related to the ADM mass $M$ and the electric charge $% Q $ of the black hole $$\begin{aligned} M &=&\frac{\omega _{n-1}}{16\pi }(n-1)m, \label{Mass} \\ Q &=&\frac{\sqrt{2} (2s-1)s\; \omega _{n-1}}{8\pi }\left( \frac{n-1}{n-2}% \right) ^{s-1/2}\left( \frac{\left( n-2s\right) q}{2s-1}\right) ^{2s-1}, \label{Charge}\end{aligned}$$ where $\omega _{n-1}$ is given by $$\omega _{n-1}=\frac{2\pi ^{\frac{n}{2}}}{\Gamma \left( \frac{n}{2}\right) }. \label{Omega}$$ It has been shown that [@PMIpapers2] Eqs. (\[Metric\]) and ([metfunction]{}) describe a black hole with a cauchy horizon ($r_{-}$) and an event horizon ($r_{+}$). The event horizon radius of this black hole can be calculated numerically by finding the largest real positive root of $% f(r=r_{+})=0$. Using the surface gravity relation, we can obtain the temperature of the black hole solutions as $$T=\frac{f^{\prime }(r_{+})}{4\pi }=\frac{n-2}{4\pi r_{+}}\left( 1+\frac{n}{% n-2}\frac{r_{+}^{2}}{l^{2}}-\frac{(2s-1)\left( \frac{(n-1)(2s-n)^{2}q^{2}}{% (n-2)(2s-1)^{2}}\right) ^{s}}{(n-1)(n-2)r_{+}^{2(ns-3s+1)/(2s-1)}}\right). \label{T}$$ The electric potential $\Phi $, measured at infinity with respect to the horizon while the black hole entropy $S$, was determined from the area law. It is easy to show that $$\begin{aligned} \Phi &=&\sqrt{\frac{n-1}{2(n-2)}}\frac{q}{r_{+}^{(n-2s)/(2s-1)}}, \label{Phi} \\ S &=&\frac{\omega _{n-1}r_{+}^{n-1}}{4}. \label{S}\end{aligned}$$ Now, as it was considered before [@PVpapers], we interpret $\Lambda $ as a thermodynamic pressure $P$, $$P=-\frac{1}{8\pi }\Lambda =\frac{n(n-1)}{16\pi l^{2}}, \label{PLambda}$$ where its corresponding conjugate quantity is the thermodynamic volume [Gibbons2]{} $$V=\frac{\omega _{n-1}{r_{+}}^{n}}{n}. \label{volrp}$$ Considering obtained quantities, one can show that they satisfy the following Smarr formula $$M=\frac{n-1}{n-2}TS+\frac{ns-3s+1}{s(2s-1)(n-2)}\Phi Q-\frac{2}{n-2}VP. \label{Smarr}$$ It has been shown that Eq. (\[Smarr\]) may be obtained by a scaling dimensional argument [@scaling; @RayCQG2009]. In addition, the (extended phase-space) first law of thermodynamics can be written as $$dM=TdS+\Phi dQ+VdP. \label{firstLaw}$$ In what follows, we shall study the analogy of the liquid–gas phase transition of the Van der Waals fluid with the phase transition in black hole solutions in the presence of PMI source. Canonical ensemble {#Cano} ================== In order to study the phase transition, one can select an ensemble in which black hole charge is fixed at infinity. Considering the fixed charge as an extensive parameter, the corresponding ensemble is called a canonical ensemble. Equation of state ----------------- Using the Eqs. (\[PLambda\]) and (\[T\]) for a fixed charge $Q$, one may obtain the equation of state, $P(V,T)$ $$P=\frac{(n-1)}{4r_{+}}T-\frac{(n-1)(n-2)}{16\pi r_{+}^{2}}+\frac{1}{16\pi }% \frac{(2s-1)\left( \frac{(n-1)(2s-n)^{2}q^{2}}{(n-2)(2s-1)^{2}}\right) ^{s}}{% r_{+}^{2s(n-1)/(2s-1)}}, \label{state}$$where $r_{+}$ is a function of the thermodynamic volume, $V$ \[see Eq. ([volrp]{})\]. Following [@PVpapers], we identify the geometric quantities $% P $ and $T$ with physical pressure and temperature of system by using dimensional analysis and $l_{P}^{n-1}=G_{n+1}\hbar /c^{3}$ as $$\lbrack \mbox{Press}]=\frac{\hbar c}{l_{p}^{n-1}}[P],\quad \lbrack % \mbox{Temp}]=\frac{\hbar c}{k}[T]. \label{dimless1}$$Therefore, the physical pressure and physical temperature are given by $$\begin{aligned} \mbox{Press} &=&\frac{\hbar c}{l_{p}^{n-1}}P=\frac{\hbar c}{l_{p}^{n-1}}% \frac{(n-1)T}{4r_{+}}+\dots \notag \\ &=&\frac{k\mbox{Temp}(n-1)}{4l_{p}^{n-1}r_{+}}+\dots \;. \label{dimless2}\end{aligned}$$Now, one could compare them with the Van der Waals equation [@PVpapers], and identify the specific volume $v$ of the fluid with the horizon radius as $v=\frac{4r_{+}l_{P}^{n-1}}{n-1}$, and in geometric units ($l_{P}=1$, $% r_{+}=\left( n-1\right) v/4$), the equation of state (\[state\]) can be written in the following form $$\begin{aligned} P &=&\frac{T}{v}-\frac{(n-2)}{\pi (n-1)v^{2}}+\frac{1}{16\pi }\frac{\kappa q^{2s}}{v^{2s(n-1)/(2s-1)}}, \label{StateV} \\ \kappa &=&\frac{4^{2s(n-1)/(2s-1)}(2s-1)\left( \frac{(n-1)(2s-n)^{2}}{% (n-2)(2s-1)^{2}}\right) ^{s}}{(n-1)^{2s(n-1)/(2s-1)}}, \label{kappa}\end{aligned}$$Considering Eq. (\[StateV\]), we plot the $P-V$ isotherm diagram in Fig. \[FPV\]. This figure shows that, similar to Van der Waals gas, there is a critical point which is a point of inflection on the critical isotherm. The pressure and volume at the critical point are known as the critical pressure and the critical volume, respectively. Above the critical point and for large volumes and low pressures, the isotherms lose their inflection points and approach equilateral hyperbolas, the so-called the isotherms of an ideal gas. It is shown that the slope of the isotherm $P-V$ diagram passing through the critical point is zero. Furthermore, as we mentioned before, the critical point is a point of inflection on the critical isotherm, hence $$\begin{aligned} \frac{\partial P}{\partial v} &=&0, \label{dpdv} \\ \quad \frac{\partial ^{2}P}{\partial v^{2}} &=&0. \label{d2pdv2}\end{aligned}$$Using Eqs. (\[dpdv\]) and (\[d2pdv2\]) with the equation of state ([StateV]{}), we will be able to calculate the critical parameters $$\begin{aligned} v_{c} &=&\left[ \frac{\kappa s(n-1)^{2}(2ns-4s+1)q^{2s}}{16(n-2)(2s-1)^{2}}% \right] ^{(2s-1)/[2(ns-3s+1)]}, \label{Vc} \\ T_{c} &=&\frac{4(n-2)(ns-3s+1)\left[ \frac{\kappa s(n-1)^{2}(2ns-4s+1)q^{2s}% }{16(n-2)(2s-1)^{2}}\right] ^{(1-2s)/[2(ns-3s+1)]}}{\pi (n-1)(2ns-4s+1)}, \label{Tc} \\ P_{c} &=&\frac{(n-2)(ns-3s+1)}{\pi s(n-1)^{2}\left[ \frac{\kappa s(n-1)^{2}(2ns-4s+1)q^{2s}}{16(n-2)(2s-1)^{2}}\right] ^{(2s-1)/(ns-3s+1)}}. \label{Pc}\end{aligned}$$These relations lead us to obtain the following universal ratio $${\rho }_{c}=\frac{P_{c}v_{c}}{T_{c}}=\frac{2ns-4s+1}{4s(n-1)}. \label{UniversalRatio}$$Note that for $s=-2/(n-5)$ with arbitrary spacetime dimensions, one can recover the ratio $\rho _{c}=3/8$, characteristic for a Van der Waals gas. $% \begin{array}{cc} \epsfxsize=7cm \epsffile{PVs34n3.eps} & \epsfxsize=7cm \epsffile{PVs2n5.eps}% \end{array} $ Free energy ------------ Thermodynamic behavior of a system may be governed by the thermodynamic potentials such as the free energy. It is known that the free energy of a gravitational system may be obtained by evaluating the Euclidean on-shell action. In order to calculate it, we use the counterterm method for cancelling of divergences. Furthermore, to make an action well-defined, one should add the Gibbons-Hawking boundary term to the bulk action. In addition, in order to fix charge on the boundary (working in canonical ensemble) we should consider a boundary term for electromagnetic field. So the total action is [@DEhShakVah] $$I=I_{b}+I_{ct}-\frac{1}{8\pi }\int_{\partial M}d^{n}x~\sqrt{\gamma }~K-\frac{% s}{4\pi }\int_{\partial M}d^{n}x~\sqrt{\gamma }(-\mathcal{F})^{s-1}~n_{\mu }F^{\mu \nu }A_{\nu }, \label{FullAction}$$where $I_{ct}$ is the counterterm action, and $\gamma _{ij}$ and $K$ denote the induced metric and extrinsic curvature of the boundary. Using Eq. ([FullAction]{}), it is straightforward to calculate the on-shell value of the total action $$I=\frac{\beta \omega _{n-1}}{16\pi }\left( 1-{\frac{r_{+}^{2}}{{l}^{2}}}+{% \frac{\left( 2s-1\right) (2sn-4s+1)\Psi ^{s}r_{+}^{2}}{\left( n-1\right) \left( n-2s\right) }}\right) r_{+}^{n-2}, \label{Onshell}$$where $$\Psi =\left( \frac{n-1}{n-2}\right) \left( \frac{2s-n}{2s-1}\right) ^{2}q^{2}r_{+}^{-\frac{2(n-1)}{2s-1}}.$$and $\beta $ is the periodic Euclidean time which is related to the inverse of Hawking temperature. Using the fact that $G=I\beta ^{-1}$ with Eq. (\[PLambda\]), the (fixed charge) free energy in the extended phase space may be written as $$G(T,P)=\frac{\omega _{n-1}}{16\pi }\left( {1}-\frac{16\pi Pr_{+}^{2}}{n(n-1)}% +{\frac{\left( 2s-1\right) (2sn-4s+1)\Psi ^{s}r_{+}^{2}}{\left( n-1\right) \left( n-2s\right) }}\right) {r}_{+}^{n-2}.$$ $% \begin{array}{cc} \epsfxsize=7cm \epsffile{GS65n4.eps} & \epsfxsize=7cm \epsffile{GS34n3.eps}% \end{array} $ $% \begin{array}{cc} \epsfxsize=7cm \epsffile{COPTplot.eps} & \epsfxsize=7cm % \epsffile{COPconformal345.eps}% \end{array} $ The behavior of the free energy is displayed in Fig. \[FG\]. In this figure the characteristic swallowtail behavior of the free energy shows the first order phase transition happen between large and small charged black holes. Using the fact that the free energy, temperature and the pressure of the system are constant during the phase transition, one can plot the coexistence curve of two phases large and small charged black holes in the PMI theory (see Fig. \[PT\]). Along this curve, small and large black holes have alike temperature (horizon radii) and pressure. ### Critical exponents One of the most important characteristics of the phase transition is the value of its critical exponents. So, following the approach of [@PVnonlinear], we calculate the critical exponents $\alpha $, $\beta $, $\gamma $, $\delta $ for the phase transition of $(n+1)$-dimensional charged black holes with an arbitrary $s$. In order to obtain the critical exponent $\alpha $, we consider the entropy of horizon $S$ and rewrite it in terms of $T$ and $V$. So we have $$S=S(T,V)=\Bigl[\omega _{n-1}\Bigl(nV\Bigr)^{n-1}\Bigr]^{\frac{1}{n}}. \label{Ent}$$Obviously, this is independent of $T$ and then the specific heat vanishes, ($% C_{V}=0$), and hence $\alpha =0$. To obtain other exponents, we study equation of state (\[StateV\]) in terms of reduced thermodynamic variables $$p=\frac{P}{P_{c}},\quad \nu =\frac{v}{v_{c}},\quad \tau =\frac{T}{T_{c}}. \label{Reduced}$$So, Eq. (\[StateV\]) translates into the following reduced equation of state $$p=\frac{4(n-1)s\tau }{(2ns-4s+1)\nu }-\frac{n-1}{(ns-3s+1)\nu ^{2}}+\frac{% (2s-1)^{2}}{(2ns-4s+1)(ns-3s+1)\nu ^{\frac{2s(n-1)}{2s-1}}}. \label{statesd}$$To study the recent equation, we will slightly generalize the argument of [@PVnonlinear] for nonlinear Maxwell theory. Indeed, we can rewrite the equation of state (\[statesd\]) as $$p=\frac{1}{\rho _{c}}\frac{\tau }{\nu }+f(\nu ,s), \label{general}$$where $\rho _{c}$ stands for the critical ratio and $$f(\nu ,s)=\frac{1}{s\left( 1-4{\rho }_{c}\right) }\left( \frac{1}{\nu ^{2}}-% \frac{\left( \frac{2s-1}{n-1}\right) ^{2}}{4s{\rho }_{c}\nu ^{\frac{2s(n-1)}{% 2s-1}}}\right) .$$The function $f(v,s)$ depends on $v$ and $s$ compared to [@PVnonlinear] where it is independent of $s$. But as we will see the nonlinearity parameter $s$ does not play any dramatic role and does not change critical exponents. Following the method of Ref. [@PVnonlinear], one may define two new parameters $t$ and $\omega $ $$\tau =t+1,\quad \nu =(\omega +1)^{1/\epsilon }, \label{omegat}$$where $\epsilon $ is a positive parameter. Now we can expand (\[statesd\]) near the critical point to obtain $$p=1+At-Bt\omega -C\omega ^{3}+O(t\omega ^{2},\omega ^{4}), \label{generalexpansion}$$with $$A=\frac{1}{\rho _{c}},\quad B=\frac{1}{\epsilon \rho _{c}},\quad C=\frac{% 2s(n-1)}{3\epsilon ^{3}(2s-1)}. \label{ABC}$$We consider a fixed $t<0$ and differentiate the Eq. (\[generalexpansion\]) to obtain $$dP=-P_{c}(Bt+3 C \omega ^{2})d\omega . \label{dPgeneral}$$Now, we denote the volume of small and large black holes with $\omega _{s}$ and $\omega _{l}$, respectively, and apply the Maxwell’s equal area law. One obtains $$\begin{aligned} p &=&1+At-Bt\omega _{l}-C\omega _{l}^{3}=1+At-Bt\omega _{s}-C\omega _{s}^{3} \notag \\ 0 &=&\int_{\omega _{l}}^{\omega _{s}}\omega dP.\end{aligned}$$ This equation leads to a unique non-trivial solution $$\omega _{s}=-\omega _{l}=\sqrt{\frac{-Bt}{C}},$$ and therefore we can find $$\eta =V_{c}(\omega _{l}-\omega _{s})=2V_{c}\omega _{l}\propto \sqrt{-t}\quad \Rightarrow \quad \beta =\frac{1}{2}.$$ Now, we should calculate the next exponent, $\gamma $. In order to obtain it, one should consider Eq. (\[generalexpansion\]). After some manipulation one can obtain $$\kappa _{T}=-\frac{1}{V}\frac{\partial V}{\partial P}\Big \|_{T}\propto \frac{% 1}{P_{c}}\frac{1}{Bt}\quad \Rightarrow \quad \gamma =1.$$Next, we calculate the final exponent, $\delta $. To do this, we should obtain the shape of the critical isotherm $t=0$ (\[generalexpansion\]), i.e., $$p-1=-C\omega ^{3}\quad \Rightarrow \quad \delta =3.$$We conclude that the thermodynamic exponents associated with the nonlinear charged black holes in any dimension $n\geq 3$ with arbitrary nonlinearity parameter, $s\neq n/2$, coincide with those of the Van der Waals fluid (the same as critical exponents of the linear Maxwell case). Equation of state for the BTZ-like black holes ---------------------------------------------- So far, we have investigated the phase transition of black holes in the presence of nonlinear PMI source with the nonlinearity $s\neq n/2$. Interestingly, for $s=n/2$, the solutions (the so-called BTZ-like black holes) have different properties. In other words, the solutions for $s=n/2$ are not the special limit of the solutions for general $s$. In fact, the solutions of $s=n/2$ are completely special and differ from the solutions of other values of $s$. As we will see, for $s=n/2$ the charge term in metric function is logarithmic and the electromagnetic field is proportional to $% r^{-1}$ (logarithmic gauge potential). In other words, in spite of some differences, this special higher dimensional solution has some similarity with the charged BTZ solution and reduces to the original BTZ black hole for $% n=2 $. Considering the metric (\[Metric\]) and the field equations of the bulk action (\[Action\]) with $s=n/2$, we can find that the metric function $% f(r)$ and the gauge potential may be written as $$\begin{aligned} f(r) &=&1+\frac{r^{2}}{l^{2}}-\frac{m}{r^{n-2}}-\frac{2^{n/2}q^{n}}{r^{n-2}}% \ln \left( \frac{r}{l}\right) , \label{Vbtz} \\ A &=&q\ln \left( \frac{r}{l}\right) dt, \label{Abtz}\end{aligned}$$ Straightforward calculations show that BTZ-like spacetime has a curvature singularity located at $r=0$, which is covered with an event horizon. The temperature of this black hole can be obtained as [@BTZlike] $$T=\frac{n-2}{4\pi r_{+}}\left( 1+\frac{n}{n-2}\frac{r_{+}^{2}}{l^{2}}-\frac{% 2^{n/2}q^{n}}{(n-2)r_{+}^{n-2}}\right) , \label{TBTZ}$$ In this section, we will investigate the analogy of the liquid–gas phase transition of the Van der Waals fluid with the phase transition in BTZ-like black hole solutions [@BTZlike]. Following the same approach and using Eqs. (\[PLambda\]) and (\[TBTZ\]) for a fixed charge $Q$, we obtain $$P=\frac{(n-1)}{4r_{+}}T-\frac{(n-1)(n-2)}{16\pi r_{+}^{2}}+\frac{1}{16\pi }% \frac{2^{n/2}(n-1)q^{n}}{r_{+}^{n}}. \label{P1}$$ Using Eqs. (\[dimless1\]) and (\[dimless2\]) with the fact that in geometric units $v=\frac{4r_{+}}{n-1}$, Eq. (\[P1\]) may be rewritten as $$\begin{aligned} P &=&\frac{T}{v}-\frac{(n-2)}{\pi (n-1)v^{2}}+\frac{1}{16\pi }\frac{\kappa ^{\prime }q^{2s}}{v^{n}}, \label{P2} \\ \kappa ^{\prime } &=&\frac{2^{5n/2}}{(n-1)^{n-1}}.\end{aligned}$$ Now, we plot the isotherm $P-V$ diagram in Fig. \[FPVbtz\]. The behavior of these plots is the same as the Van der Waals gas. In order to find the critical quantities, one may use Eqs. (\[dpdv\]) and (\[d2pdv2\]). These relations help us to obtain $$\begin{aligned} v_{c} &=&\left[ \frac{\kappa ^{\prime }n(n-1)^{2}q^{n}}{32(n-2)}\right] ^{1/(n-2)}, \label{Vc2} \\ T_{c} &=&\frac{2(n-2)^{2}\left[ \frac{\kappa ^{\prime }n(n-1)^{2}q^{n}}{% 32(n-2)}\right] ^{-1/(n-2)}}{\pi (n-1)^{2}}, \label{Tc2} \\ P_{c} &=&\frac{(n-2)^{2}}{\pi n(n-1)\left[ \frac{\kappa ^{\prime }n(n-1)^{2}q^{n}}{32(n-2)}\right] ^{2/(n-2)}}. \label{Pc2}\end{aligned}$$Having the critical quantities at hand, we are in a position to obtain the following universal ratio $${\rho }_{c}=\frac{P_{c}v_{c}}{T_{c}}=\frac{n-1}{2n}. \label{ratio}$$It is notable that only for $n=4$ ($5$-dimensional BTZ-like black holes), one can recover the ratio $\rho _{c}=3/8$, characteristic for a Van der Waals gas, where for higher dimensional Reissner–Nordström black holes, this ratio has been recovered only for $4$-dimensions [@PVnonlinear]. In addition, considering $s=n/2$ in Eq. (\[UniversalRatio\]), one can obtain $% {\rho }_{c}$ of the BTZ-like black holes. $% \begin{array}{cc} \epsfxsize=7cm \epsffile{PVbtzn4.eps} & \epsfxsize=7cm \epsffile{PVbtzn5.eps}% \end{array} $ ### Critical exponents of BTZ-like black holes In order to obtain the critical exponents of the phase transition, we follow the same procedure of [@PVnonlinear]. The entropy of horizon $S(T,V)$ is the same as Eq. (\[Ent\]), which is independent of $T$ and hence $\alpha =0 $. In addition we can use Eq. (\[Reduced\]) and rewrite Eq. (\[P2\]) in the following form $$p=\frac{2n\tau }{(n-1)\nu }-\frac{n}{(n-2)\nu ^{2}}+\frac{2}{(n-1)(n-2)\nu ^{n}}. \label{statesd2}$$ It is straightforward to show that the thermodynamic exponents associated with the BTZ-like black holes in arbitrary dimension, coincide with those of the Van der Waals fluid (the same as critical exponents of the PMI case). Grand canonical ensemble {#GCano} ======================== In addition to the canonical ensemble, one can work with a fixed electric potential at infinity. The ensemble of this fixed intensive quantity translates into the grand canonical ensemble. It is worthwhile to note that, for linear Maxwell field, the criticality cannot happen in the grand canonical ensemble [@PVpapers]. Equation of state ----------------- In this section, we study the critical behavior of charged black holes in the grand canonical (fixed $\Phi $) ensemble. We take $q=\Phi r_{+}^{(n-2s)/(2s-1)}$ with $v=\frac{4r_{+}}{n-1}$ to rewrite Eq. ([state]{}) in the following form $$P=\frac{T}{v}\mathbf{-}\frac{(n-2)}{(n-1)\pi v^{2}}\mathbf{+}\frac{2s-1}{% 16\pi }\left( \frac{4\sqrt{2}(n-2s)\Phi }{(2s-1)(n-1)v}\right) ^{2s}\mathbf{,% } \label{State Grand}$$ In the Maxwell theory ($s=1$) Eq. (\[State Grand\]) reduces to $$Pv^{2}=Tv-\frac{(n-2)}{(n-1)\pi }+\frac{2}{\pi }\frac{(n-2)^{2}}{(n-1)^{2}}% \Phi ^{2}\mathbf{.}$$ Clearly, this is a quadratic equation and does not show any criticality. Interestingly, in contrast to the Maxwell field ($s=1$), the PMI theory admits phase transition in this ensemble and one can study the fixed potential $P-V$ phase transition of the black holes in extended phase space. However, the general behavior of the isotherm $P-V$ diagram for fixed potential is same as fixed charge ensemble as displayed in Fig \[PotPV\]. Applying Eqs. (\[dpdv\]) and (\[d2pdv2\]) to the equation of state, it is easy to calculate the critical point in the grand canonical ensemble $$\begin{aligned} v_{c} &=&\frac{4\sqrt{2}(n-2s)}{(2s-1)(n-1)}\left[ \frac{2s(2s-n)^{2}}{% (n-2)(n-1)}\right] ^{\frac{1}{2(s-1)}}\Phi ^{\frac{s}{s-1}}, \\ T_{c} &=&\frac{(s-1)(n-2)}{\pi (n-2s)}\left[ \frac{(n-2)(n-1)}{% s2^{s}(n-2s)^{2}}\right] ^{\frac{1}{2(s-1)}}\Phi ^{\frac{-s}{s-1}}, \\ P_{c} &=&\frac{(s-1)(2s-1)^{2}2^{(4-5s)/(s-1)}}{s\pi }\left[ \frac{% (n-2s)^{2}s^{\frac{1}{s}}}{(n-1)(n-2)}\right] ^{\frac{-s}{s-1}}\Phi ^{\frac{% -2s}{s-1}}.\end{aligned}$$One must consider that in contrast to the canonical ensemble for $s>\frac{n}{2}$ or $s<1~$the $v_{c}$, $T_{c}$ or $P_{c}~$ take the negative value so there is not any physical phase transition in these cases. Using the values of $v_{c},T_{c\text{ }}$and $P_{c}$ we will be able to obtain the following universal ratio$${\rho }_{c}=\frac{P_{c}v_{c}}{T_{c}}=\frac{2s-1}{4s}. \label{ratio2}$$Note that here this universal ratio is independent of $n$ and for $s=2$ one can recover the ratio $\rho _{c}=3/8$, characteristic for a Van der Waals gas. Although Eq. (\[ratio2\]) does not depend on the spacetime dimensions, but one can take $s=n/2$ to recover Eq. (\[ratio\]) of BTZ-like black holes. Furthermore, it is interesting to mention that one can set $n=2s$ in Eq. (\[UniversalRatio\]) to obtain Eq. (\[ratio2\]). $% \begin{array}{cc} \epsfxsize=7cm \epsffile{PotPS65n3.eps} & \epsfxsize=7cm % \epsffile{PotPS54n4.eps}% \end{array} $ Free energy ------------ By ignoring the surface term of PMI and fixing the potential on the boundary $% \delta A_{\mu }\|_{\partial M}=0$, one can find the on-shell action correspondence to free energy in the grand canonical ensemble. So, we take the action as follows $$I=I_{b}+I_{ct}-\frac{1}{8\pi }\int_{\partial M}d^{n}x~\sqrt{\gamma }~K.$$ Now we can find the free energy $G=I\beta ^{-1}$ as $$G(T,P)=\frac{\omega _{n-1}}{16\pi }\left( {1}-\frac{16\pi Pr_{+}^{2}}{n(n-1)}% +{\frac{2^{s}\left( 2s-1\right) ^{2-2s}\Phi ^{2s}r_{+}^{2-2s}}{\left( n-1\right) \left( 2s-n\right) ^{1-2s}}}\right) {r}_{+}^{n-2}.$$ Now, we are looking for the phase transition. We plot figure \[PotG\] and find that, in contrast to Maxwell case, there is a first order phase transition. In other words, the nonlinearity parameter, $s$, affects the existence of the phase transition in the grand canonical ensemble. $% \begin{array}{cc} \epsfxsize=7cm \epsffile{PotGS65n3.eps} & \epsfxsize=7cm % \epsffile{PotGS54n4.eps}% \end{array} $ Concluding Remarks ================== In this paper, we have considered the cosmological constant and its conjugate quantity as thermodynamic variables and investigated the thermodynamic properties of a class of charged black hole solutions. At the first step, we have introduced the black hole solutions of the Einstein-$% \Lambda $ gravity in the presence of the PMI source. Then, we have used the Hawking temperature as an equation of state and calculated the the critical parameters, $T_{c}$, $v_{c}$ and $P_{c}$. We have plotted the isotherm diagram ($P$–$V$) of charged black holes in PMI theory and found that the total behavior of this diagram is the same as that of the Van der Waals gas. Also, we have obtained the free energy of a gravitational system through the use of Euclidean on-shell action to investigate its thermodynamics behavior. Furthermore, we have calculated the critical exponents of the phase transition and concluded that the thermodynamic exponents associated with the nonlinear charged black holes in arbitrary dimension coincide with those of the Van der Waals fluid (the mean field theory). Also, we have applied the same procedure for the BTZ-like black holes to obtain their phase transition. Calculations showed that thermodynamic behaviors of BTZ-like black holes are the same as PMI ones. Moreover, we have studied the grand canonical ensemble in which the potential, instead of charge, should be fixed on the boundary. In contrast to the Maxwell case [@PVpapers], here one sees a phase transition. We have also computed the universal ratio $\frac{P_{c}v_{c}}{T_{c}}=\frac{2s-1}{4s}$ and found that it does not depend on the spacetime dimensions. Finally, we have found that, $v_{c}$, $T_{c}$ and $P_{c}$ have different dependencies of $n$ and $s$ for PMI and BTZ-like in the canonical ensemble, but when $n=2s$ both of them reduce to the universal ratio $\rho_{c}=\frac{P_{c}v_{c}}{T_{c}}=\frac{2s-1}{4s}$ which we have found in the grand canonical ensemble. It is interesting to investigate underlining reasons for the mentioned universality and figure out why this ratio in the grand canonical ensembles is independent of spacetime dimensions. Moreover, in the statistical physics, it is known that a universality class of criticality is characterized by dimensions of space, order parameters and fluctuations [@Gold]. However, it is not clear which features of gravitational theories and black holes determine the universality class of phase transition or modify critical exponents. Clearly as we found in this paper and have been shown in [@PVnonlinear], the critical exponents do not change by crucial modifications of matter fields (such as PMI modification) or changing the space-time dimensions. In addition, it seems the geometry of spacetime is also irrelevant to critical exponents as one can see in the slowly rotating black holes [@PVnonlinear]. Interestingly, these critical exponents remain unchanged in the mean field class even when one considers some corrections to gravity action [@GB]. Besides, it is worthwhile to think about whether there is any holographic interpretation for the extended phase space thermodynamics and the universal classification. Perhaps a holographic approach helps us to have a better understanding of this problem. We leave the study of these interesting questions for future studies. Acknowledgement =============== We thank an anonymous referee for useful comments. M. H. Vahidinia would like to thank A. Montakhab, S. Jalali, P. Manshour and A. Moosavi for useful discussions. S. H. Hendi wishes to thank Shiraz University Research Council. This work has been supported financially by Research Institute for Astronomy & Astrophysics of Maragha (RIAAM), Iran. [99]{} G. Gibbons, R. Kallosh and Barak Kol, Phys. Rev. Lett. 77, 4992 (1996). J.D. Brown and C. Teitelboim, Phys. Lett. B 195, 177 (1987). M. M. Caldarelli, G. Cognola and D. Klemm, Class. Quantum Gravit. 17, 399 (2000). D. Kastor, S. Ray and J. Traschen, Class. Quantum Gravit. 26, 195011 (2009). M. Cvetic, G. Gibbons, D. Kubiznak and C. Pope, Phys. Rev. D 84, 024037 (2011). B. P. Dolan, Class. Quantum Gravit. 28, 125020 (2011); B. P. Dolan, Class. Quantum Gravit. 28, 235017 (2011); D. Kubiznak and R. B. Mann, JHEP 07, 033 (2012); B. P. Dolan, \[arXiv:1209.1272\]. C. Niu, Y. Tian and X. Wu, Phys. Rev. D 85, 024017 (2012). A. Chamblin, R. Emparan, C. Johnson and R. Myers, Phys. Rev. D 60, 064018 (1999); A. Chamblin, R. Emparan, C. Johnson and R. Myers, Phys. Rev. D 60, 104026 (1999). R. Banerjee and D. Roychowdhury, JHEP 1111, 004 (2011). P. A. M. Dirac, Lectures on Quantum Mechanics, Yeshiva University, Belfer Graduate School of Science, New York (1964). Z. Bialynicka-Birula and I. Bialynicka-Birula, Phys. Rev. D 2, 2341 (1970). W. Heisenberg and H. Euler, Z. Phys. 98, 714 (1936); Translation by: W. Korolevski and H. Kleinert, Consequences of Dirac’s Theory of the Positron, \[physics/0605038\]; H. Yajima and T. Tamaki, Phys. Rev. D 63, 064007 (2001). D. H. Delphenich, Nonlinear electrodynamics and QED, \[arXiv: hep-th/0309108\]; D. H. Delphenich, Nonlinear optical analogies in quantum electrodynamics, \[arXiv: hep-th/0610088\]. J. Schwinger, Phys. Rev. 82, 664 (1951). P. Stehle and P. G. DeBaryshe, Phys. Rev. 152, 1135 (1966). Y. S. Myung, Y. W. Kim and Y. J. Park, Phys. Rev. D 78, 044020 (2008); S. Fernando, Phys. Rev. D 74, 104032 (2006); O. Miskovic and R. Olea, Phys. Rev. D 77, 124048 (2008); Y. S. Myung, Y. W. Kim and Y. J. Park, Phys. Rev. D 78, 084002 (2008); R. Banerjee and D. Roychowdhury, Phys. Rev. D 85, 044040 (2012); R. Banerjee and D. Roychowdhury, Phys. Rev. D 85, 104043 (2012). R. G. Cai and Y. W. Sun, JHEP 09, 115 (2008); X. H. Ge, Y. Matsuo, F. W. Shu, S. J. Sin and T. Tsukioka, JHEP 10, 009 (2008). S. Gangopadhyay and D. Roychowdhury, JHEP 1205, 002 (2012); D. Roychowdhury, Phys. Rev. D 86, 106009 (2012); S. Gangopadhyay and D. Roychowdhury, JHEP 1205, 156 (2012). S. Gunasekaran, R. B. Mann and D. Kubiznak, \[arXiv:1208.6251\]. Y. Kats, L. Motl and M. Padi, JHEP 0712, 068 (2007); D. Anninos and G. Pastras, JHEP `0907`, 030 (2009); R. G. Cai, Z. Y. Nie and Y. W. Sun, Phys. Rev. D 78, 126007 (2008); N. Seiberg and E. Witten, JHEP 09, 032 (1999); E. Fradkin and A. Tseytlin, Phys. Lett. B 163, 123 (1985); R. Matsaev, M. Rahmanov and A. Tseytlin, Phys. Lett. B 193, 205 (1987); E. Bergshoff, E. Sezgin, C. Pope and P. Townsend, Phys. Lett. B 188, 70 (1987); A. Tseytlin, Nucl. Phys. B 276, 391 (1986); D.J. Gross and J. H. Sloan, Nucl. Phys. B 291, 41 (1987). M. Hassaine and C. Martinez, Phys. Rev. D 75, 027502 (2007); S. H. Hendi and H. R. Rastegar-Sedehi, Gen. Relativ. Gravit. 41, 1355 (2009); S. H. Hendi, Phys. Lett. B 677, 123 (2009); M. Hassaine and C. Martinez, Class. Quantum Gravit. 25, 195023 (2008); H. Maeda, M. Hassaine and C. Martinez, Phys. Rev. D 79, 044012 (2009); S. H. Hendi and B. Eslam Panah, Phys. Lett. B 684, 77 (2010); S. H. Hendi, Prog. Theor. Phys. 124, 493 (2010); S. H. Hendi, Eur. Phys. J. C 69, 281 (2010); S. H. Hendi, Phys. Rev. D 82, 064040 (2010). J. Jing, Q. Pan and S. Chen, JHEP 1111, 045 (2011); D. Roychowdhury, Phys. Lett. B 718, 1089 (2013). S. H. Hendi, Phys. Lett. B 690, 220 (2010); D. J. Gross and J. H. Sloan, Nucl. Phys. B 291, 41 (1987); W. A. Chemissany, M. de Roo and S. Panda, JHEP 08, 037 (2007). J. P. Gauntlett, R. C. Myers and P. K. Townsend, Class. Quantum Gravit. 16, 1 (1999). M. H. Dehghani, C. Shakouri and M. H. Vahidinia, Phys. Rev. D. 87, 084013 (2013). S. H. Hendi, Eur. Phys. J. C 71, 1551 (2011). N. Goldenfeld, “Lectures on Phase Transitions and the Renormalization Group,” (Westview Press, New York, 1992). S. W. Wei and Y. X. Liu, Phys. Rev. D 87, 044014 (2013); R. G. Cai, L. M. Cao, L. Li and R. Q. Yang, \[arXiv:1306.6233\]. [^1]: email address: [email protected] [^2]: email address: [email protected]	{ "pile_set_name": "ArXiv" }
--- abstract: \| A comprehensive investigation of the electronic and magnetic properties of NaOsO$_{3}$ has been made using the first principle calculations, in order to understand the importance of Coulomb interaction, spin–orbit coupling and magnetic order in its temperature–induced and magnetic–related metal–insulator transition. It is found that its electronic structure near the Fermi energy is dominated by strongly hybridized Os 5d and O 2p states. Despite of the large strength of spin–orbit coupling, it has only small effect on the electronic and magnetic properties of NaOsO$_{3} $. On the other hand, the on–site Coulomb repulsion affects the band structure significantly, but, a reasonable U alone cannot open a band gap. Its magnetism is itinerant, and the magnetic configuration plays an important role in determining the electronic structure. Its ground state is of a G–type antiferromagnet, and it is the combined effect of U and magnetic configuration that results in the insulating behavior of NaOsO$_{3}$. author: - Yongping Du - Xiangang Wan - Li Sheng - Jinming Dong - 'Sergey Y. Savrasov' title: 'Electronic structure and magnetic properties of NaOsO$_{3}$' --- [^1] INTRODUCTION ============ It is well known that the Coulomb interaction among 3d electrons in transition–metal oxides (TMO) is substantially important, which induces peculiar properties, such as metal–insulator transition[@Review; @M-I], colossal magnetoresistance[@CMR] and high critical temperature superconductivity[@Pickett-HTC]. Because the 5d orbitals are highly extended compared to those in the 3d systems, it is natural to expect that the electronic correlations are weak and have only negligible effect in the 5d compounds. However, recent theoretical and experimental works have given the evidence on the importance of Coulomb interactions here[@Sr2IrO4-1; @Sr2IrO4-2]. On the other hand, the spin–orbit coupling (SOC) in the 5$d$ transition metal elements is expected to be strong due to the large atomic number[@SOC]. Hence, due to the interplay of electron correlations and strong spin–orbit interactions, various anomalous electronic properties have been observed/proposed in the 5d transition oxides, such as J$_{eff}$=1/2 Mott state[Sr2IrO4-1,Sr2IrO4-2]{}, giant magnetoelectric effect[@GME], high T$_{c}$ superconductivity[@HTC], Weyl semimetal with Fermi arcs[Weyl-semimetal]{}, Axion insulator with large magnetoelectric coupling[Axion insulator]{}, topological insulator[@Pesin-Balents; @Kim], correlated metal[@Bi2Ir2O7], Kitaev mode[@Kitav], etc. One class of the well studied 5d compounds are the osmates [Pickett,NaOsO3, Cd2Os2O7-1,Cd2Os2O7, Singh LSDA,RbOs2O6,BaNiOsO,KOs2O6 DFT,Cd2Os2O7-3,Cd2Os2O7-4,Cd2Os2O7-5]{}. For example, the physical properties of Cd$_{2}$Os$_{2}$O$_{7}$ are quite intriguing. It has been found that Cd$% _{2}$Os$_{2}$O$_{7}$ is metallic at room temperatures, while undergoing a metal–insulator transition (MIT) at about 230 K[@Cd2Os2O7-1]. Experiments reveal that this MIT is continuous and purely electronic. Moreover, it is coincident with a magnetic transition of antiferromagnetic (AFM) character[@Cd2Os2O7-1; @Cd2Os2O7]. Therefore, experimentalists argue that Cd$_{2}$Os$_{2}$O$_{7}$ is the first well–documented example of a pure Slater transition[@Cd2Os2O7; @Slater; @state]. However, despite of the vast efforts devoted[@Cd2Os2O7-1; @Cd2Os2O7; @Cd2Os2O7-3; @Cd2Os2O7-4; @Singh; @LSDA; @Cd2Os2O7-5], its exact magnetic ground state configuration is still unknown due to the strong geometric frustration of the pyrochlore lattice. Therefore, theoretical evidence of the Slater transition in this compound is still lacking. Recently, using high pressure technique, Shi et al.[@NaOsO3] synthesized another osmate: NaOsO$_{3}$. Similar to Cd$_{2}$Os$_{2}$O$_{7}$, NaOsO$_{3}$ also exhibits a temperature–induced MIT, which is again accompanied by a magnetic ordering without any lattice distortion[NaOsO3]{}. However, better than Cd$_{2}$Os$_{2}$O$_{7}$, NaOsO$_{3}$ has a simple perovskite structure, consequently being free from the complication induced by magnetic frustration. Therefore, NaOsO$_{3}$ provides a unique platform to understand the temperature–induced and magnetic–related MIT. Based on the experimental crystal structure, Shi et al.[@NaOsO3] also perform the band-structure calculation for this compound. They[NaOsO3]{} find that both LDA and LDA+SO calculation give the paramagnetic solution. Their numerical results[@NaOsO3] show that Coulomb U is not efficient, and antiferromagnetic correlation is essential to open the band gap[@NaOsO3]. Recently, there are experimental and theoretical evidences of the importance of electronic correlation and spin-orbital coupling in 5d transition-metal compounds. Therefore a comprehensive investigation of the effect of the Coulomb interaction, SOC and magnetic order on its electronic structure and MIT is still an interesting problem which we address in the present work. METHOD ====== The electronic band structure calculations have been carried out by using the full potential linearized augmented plane wave method as implemented in WIEN2K package[@WIEN]. Local spin density approximation (LSDA) is widely used for various 4$d$ and 5$d$ transition metal oxides[Sr2IrO4-1,Sr2IrO4-2,LSDA-good-1,LSDA-good-2,KOs2O6 DFT]{}, and we therefore adopt it as the exchange–correlation potential. The muffin–tin radii for Na, Os and O are set to 1.13, 1.02, and 0.90 Å, respectively. The basic functions are expanded to R$_{mt}$K$_{max}$=7 (where R$_{mt}$ is the smallest of the muffin-tin sphere radii and K$_{max}$ is the largest reciprocal lattice vector used in the plane wave expansion), corresponding to 1915 LAPW functions at the $\Gamma $ point. Using the second–order variational procedure[@SOC; @in; @WIEN2k], we include the spin–orbital coupling interaction (SOC), which has been found to play an important role in the 5d system[@Sr2IrO4-1; @Sr2IrO4-2; @Weyl-semimetal; @Axion; @insulator]. A 10$\times $6$\times $10 mesh is used for the Brillouin zone integral. The self–consistent calculations are considered to be converged when the difference in the total energy of the crystal does not exceed 0.1 mRy and that in the total electronic charge does not exceed 10$^{-3}$ electronic charge at consecutive steps. ------ -- -- -- -- -- -- -- atom Na O1 O2 ------ -- -- -- -- -- -- -- : Numerical and experimental internal coordinates of NaOsO$_{3}$. NaOsO$_{3}$ has an orthorhombic perovskite structure with space group of Pnma[@NaOsO3]. There are four formula units (f.u.) per unit cell, and the 20 atoms in the unit cell can be classified as four nonequivalent crystallographic sites: Na, Os, O1 and O2 according to the symmetry. They are located at 4c, 4b, 4c and 8d sites, respectively and result in seven internal coordinates. From the X–ray diffraction experiment[@NaOsO3], the lattice constants of NaOsO$_{3}$ are determined to be a=5.384 Å, b=7.580 Å and c=5.328 Å, respectively. Based on the experimental lattice parameters, we optimize all independent internal atomic coordinates until the corresponding forces are less than 1 mRy/a.u. We confirm that the Coulomb $U$ and SOC have only small effect on the crystal structure and list in Table I the internal atomic coordinates by LDA calculations. Our numerical internal coordinates are in good agreement with the experimental result, as shown in Table 1. ![Electronic density of states (DOS) from LDA calculation. Fermi energy E$_{f}$ is set to zero. (a) TDOS, (b) Os 5d PDOS, (c) O1 2p PDOS, (d) O2 2p PDOS, (e) Na PDOS.](Fig-1.eps) ![Band structure of NaOsO$_{3}$, shown along the high symmetry directions. (a) LDA, (b) LDA+SO.](Fig-2.eps) RESULTS AND DISCUSSIONS ======================= Using the experimental lattice constants and the numerical internal coordinates, we first perform non–magnetic LDA calculation. The total density of states (TDOS), Os 5d partial density of states (PDOS), O1 2p PDOS, O2 2p PDOS, Na 2s, 2p PDOS has been plotted in Fig.1(a)-(e), respectively. Our TDOS is very similar to that obtained based on the experimental crystal structure (See Fig.4a of Ref.[@NaOsO3]). The energy range, -9.0 to -2.4 eV is dominated by O1 2p and O2 2p bands with a small contribution from Os 5d state. Both Na 2s and 2p states, appearing mainly above 4 eV, have also considerable distribution between -9.0 to -2.4 eV, where O 2p state is mainly located, indicating the non–negligible hybridization between Na and O states despite that Na is highly ionic. The Os atom is octahedrally coordinated by six O atoms, making the Os 5d band to split into the t$_{2g}$ and e$_{g}$ states, and the 12 t$_{2g}$ bands are located from -2.8 to 1.2 eV, as shown in Fig. 2a. Due to the extended nature of 5d states, the crystal splitting between t$_{2g}$ and e$_{g}$ states is large, and the e$_{g}$ states are located about 2.0 eV higher than the Fermi energy (E$_{f}$) and disperse widely. While providing the basic features of the electronic structure, LDA produces a metallic state due to partially occupied Os 5d t$_{2g}$ band. SOC of 5d electrons is about 0.5 eV[@SOC], which is one order of magnitude larger than that of 3d electrons. Therefore, SOC usually changes the 5d band dispersion significantly and plays an essential role in the gap opening of Sr$_{2}$IrO$_{4}$ as well as of pyrochlore iridates[@Sr2IrO4-1; @Sr2IrO4-2; @Weyl-semimetal]. In order to investigate the effect of SOC on the electronic structure, we compare the results obtained in the presence and absence of SOC, which are given in Fig. 2. The difference between the bands with and without SOC is small, as demonstrated in Fig.2. For 5d$^{5}$ electronic configuration of Sr$% _{2}$IrO$_{4}$[@Sr2IrO4-1], and A$_{2}$Ir$_{2}$O$_{7}$ (A=Y or rare earth)[@Weyl-semimetal], where J$_{eff}=1/2$ picture is valid, SOC has a dramatic effect on the band structure. In NaOsO$_{3}$, Os occurs in its 5$^{+}$ valence and there are 3 electrons in its t$_{2g}$ band. Since t$% _{2g}$ band is half filled, it is natural to expect the effect of SOC to be small. As shown in Fig. 1a, the Fermi level is located near a sharp peak in the DOS. The relatively high density of states at the Fermi energy (N(E$_{f}$)) suggests the possibility of a Stoner instability against ferromagnetism (FM). Therefore, we perform a spin polarized calculation, however our LSDA+SO calculation with initial FM setup converges to the non–magnetic state. Thus consistent with Shi et al.[@NaOsO3], the FM state is not stable in LSDA+SO calculation. ![Electronic density of states (DOS) from LDA+SO+U (U=2.0 eV) calculation. Fermi energy E$_{f}$ is set to zero. (a) FM configuration, (b) G-AFM (c) C-AFM, (d) A-AFM (e) D-AFM. The black/red line is for up/down-spin, respectively. For AFM configurations, spin–up is the same as spin–down, thus only one spin channel is plotted.](Fig-3.eps) Although the 5d orbitals are spatially extended, it has been found that the electronic correlations are important for 5d TMO[Sr2IrO4-1,Sr2IrO4-2]{}. Moreover, the experiment reveal that NaOsO$_{3}$ has a long-range magnetic order at low temperature[@NaOsO3]. We therefore utilize LSDA+U scheme[@+U], which is adequate for searching for magnetically ordered insulating ground states [@LDA+DMFT]. Although, the accurate value of U is not known for perovskite osmates, the estimates of the values of $U$ have been recently obtained between 1.4 and 2.4 eV in layered Sr$_{2}$IrO$_{4}$/Ba$_{2}$IrO$_{4}$ [@cRPA]. We generally expect the screening to be larger in three dimensional (3D) systems than in two dimensional (2D) ones, and one can image that U in NaOsO$_{3}$ should be smaller than that in Sr$_{2}$IrO$_{4}$/Ba$_{2}$IrO$% _{4}$. We therefore perform LDA+U+SO calculation and vary parameter $U$between 0.5 and 2.0 eV. Numerical results show that the electronic correlations can stabilize the FM configuration and narrow the Os t$_{2g}$ band. However, as shown in Fig.3(a), our LDA+U+SO calculation with $U$= 2.0 eV still gives a metallic solution. Naively, one may expect that using larger Coulomb $U$ will result in an insulating state. However, consistent with Shi et al.[@NaOsO3], our additional calculations show that increasing $U$ cannot solve this problem, and even a quite large U (=6.0 eV) cannot open the band gap. Therefore, electronic correlations alone cannot explain the insulating behavior, and the MIT is not of a Mott–type. After studying the effect of SOC and U, we subsequently investigate the effect of various magnetic orders. We considered four antiferromagnetic (AFM) configurations besides the FM state: A–type AFM state (A–AFM) with layers of Os ions coupled ferromagnetically in a given set of (001) planes but with alternate planes having opposite spin orientation; C–type AFM state (C–AFM) with lines of Os ions coupled ferromagnetically in a given direction (001) but with alternate lines having opposite spin orientation; G–type AFM state (G–AFM) with Os ions coupled antiferromagnetically with all of their nearest neighbors; D–type AFM state (D–AFM) where Os ions lying within alternating planes perpendicular to \[001\] direction are coupled ferromagnetically along either \[010\] or \[100\] directions while different lines are coupled antiferromagnetically. Same as with the FM setup, the LSDA+SO calculation with U=0 for all considered AFM setups converges to the nonmagnetic metallic state. Configuration G–AFM FM C–AFM A–AFM D–AFM --------------------- ------- ------- ------- ------- ------- E$_{tot}$ 0 0.243 0.186 0.282 0.205 $\langle S\rangle $ 0.94 0.22 0.54 0.29 0.20 $\langle O\rangle $ -0.11 -0.01 -0.04 -0.03 -0.03 : Spin $\langle S\rangle $ and orbital $\langle O\rangle $ moment (in $\protect\mu _{B}$) as well as the total energy E$_{tot}$ per unit cell (in eV) for several magnetic configurations, as calculated using LDA+U+SO method with U=2.0 eV. (E$_{tot}$ is defined relative to the G–AFM configuration.). On the other hand, the non–zero Coulomb interaction U of Os 5d is found to stabilize the AFM configuration. Our calculation confirms that the magnetic order has a significant effect, and for a reasonable U ($\leq $2.0 eV), G–AFM configuration is the only insulating solution as shown in Fig.3. Moreover, regardless the value of U, the G–AFM configuration always has the lowest total energy. Thus we believe G–AFM configuration is the magnetic ordering state observed by the experiment [@NaOsO3] . With increasing U the band structure will change, but only when U is larger than 1.0 eV, the G–AFM solution becomes insulating. The DOS from U=1.0 eV (see the Fig.4c of Ref.21) is similar with that from U=2.0 eV (see Fig.3b of present work), which again indicates that the Coulomb U is not efficient to open the band gap. It is found that the magnetic moment is mainly located at Os site, and despite of strong hybridization between Os 5d and O* 2p, O site is basically non–magnetic (less than 0.003 $\mu _{B}$). The numerical data for U=2.0 eV are given in Table II. For the 5d$^{5}$ electronic systems such as BaIrO$_{3}$, Sr$_{2}$IrO$_{4}$, pyrochlore iridates etc, it has been found that due to the strong spin–orbit entanglement in 5d* states, the magnetic orbital moment is about twice larger than the spin moment[Sr2IrO4-1,Sr2IrO4-2,Weyl-semimetal,Orb moment in BaIrO3]{}, even in the presence of strong crystal field and band effects. Contrary to 5d$% ^{5}$ systems, the obtained orbital moment for NaOsO$_{3}$ is much smaller than its spin moment, showing again that SOC effect is small for this 5d$^{3}$ electronic configuration case. As shown in Table II, the magnitude of magnetic moment is sensitive to the magnetic configuration, indicating the itinerant nature of magnetism. For the same U value, the G–AFM configuration always has the largest magnetic moment among the considered states. However, as shown in Table II, our numerical magnetic moment (0.83 $\mu _{B}$) is much smaller than the experimental one[NaOsO3]{}. For an itinerant magnet, one may still fit the $\chi (T)$ curve by the Curie–Weiss law, but cannot estimate the magnetic moment accurately based on the Curie–Weiss constant[@Moriya; @book]. Thus, the experimental magnetic moment may not be reliable. The energy difference between various magnetic configurations is large, which is consistent with the observed high magnetic transition temperature (about 410 K)[@NaOsO3] although here one cannot estimate the interatomic exchange interaction and T$_{N}$ based on the difference between total energies accurately as in the local moment systems[@Wan]. Since the G–AFM configuration is the only insulating state, it is easy to understand that both magnetic and electronic phase transitions occur at the same temperature and our calculation indeed confirms that the MIT of NaOsO$_{3}$ is a Slater–type transition. SUMMARY ======= In summary, we have investigated the detailed electronic structure and magnetic properties of NaOsO$_{3}$ using full potential linearized augmented plane wave method. Our results show that the electronic structure near the Fermi energy E$_{f}$ is dominated by strongly hybridized Os 5d and O 2p states. Despite of its big value the SOC has only weak effect on the band structure and magnetic moment. The electronic correlations alone cannot open the band gap, and the low temperature phase of NaOsO$_{3}$ is not a Mott–type insulator. The magnetic configuration has an important effect on the conductivity, and the ground state is a G–type AFM insulator. It is the interplay of the Coulomb interaction and magnetic ordering that result in the insulating behavior of NaOsO$_{3}$. ACKNOWLEDGMENTS =============== The work was supported by the National Key Project for Basic Research of China (Grant no. 2011CB922101 and 2010CB923404), NSFC under Grant no. 91122035, 11174124 and 10974082. The project also funded by Priority Academic Program Development of Jiangsu Higher Education Institutions. S.Y.S was supported by DOE Computational Material Science Network (CMSN) and DOE SciDAC Grant No. SE-FC02-06ER25793 [99]{} M. Imada, A. Fujimori, and Y. Tokura, Rev. Mod. Phys. 70, 1039 (1998). S. Jin, T. H. Tiefel, M. McCormack, R. A. Fastnacht, R. Ramesh and L. H. Chen, Science 264, 413 (1994); P. Schiffer, A. P. Ramirez, W. Bao, and S-W. Cheong, Phys. Rev. Lett. 75, 3336 (1995). W. Pickett, Rev. Mod. Phys. 61, 433 (1989). B.J. Kim, Hosub Jin, S. J. Moon, J.-Y. Kim, B.-G. Park, C. S. Leem, Jaejun Yu, T. W. Noh, C. Kim, S.-J. Oh, J.-H. Park, V. Durairaj, G. Cao, and E. Rotenberg, Phys. Rev. Lett. 101, 076402 (2008); B.J. Kim, H. Ohsumi, T. Komesu, S. Sakai, T. Morita, H. Takagi, and T. Arima, Science 323, 1329 (2009). H. Jin, H. Jeong, T. Ozaki and J. Yu, Phys. Rev. B 80, 075112 (2009). L.F. Mattheiss, Phys. Rev. B 13, 2433 (1976). S. Chikara, O. Korneta, W. P. Crummett, L. E. DeLong, P. Schlottmann and G. Cao, Phys. Rev. B 80, 140407 (R) (2009). F. Wang and T. Senthil, Phys. Rev. Lett. 106, 136402 (2011). X. Wan, A.M. Turner, A. Vishwanath, and S.Y. Savrasov, Phys. Rev. B 83, 205101 (2011). X. Wan, A. Vishwanath, and S. Y. Savrasov, Phys. Rev. Lett 108, 146601 (2012). D.A. Pesin and L. Balents, Nature Physics 6, 376 (2010). H.-M. Guo and M. Franz, Phys. Rev. Lett 103, 206805 (2009). B. J. Yang, Y. B. Kim, Phys. Rev. B 82, 085111 (2010). M. Kargarian, J. Wen, G. A. Fiete, Phys. Rev. B 83, 165112 (2011). T. F. Qi, O. B. Korneta, X. Wan, G. Cao, arXiv:1201.0538 (2012). G. Jackeli and G. Khaliulin, Phys. Rev. Lett. 102, 017205 (2009). W. Sleight, J. L. Gillson, J. F. Weiher, and W. Bindloss, Solid State Commun. 14, 357 (1974). D. Mandrus, J.R. Thompson, R. Gaal, L. Forro, J.C. Bryan, B.C. Chakoumakos, L.M. Woods, B.C. Sales, R.S. Fishman, and V. Keppens, Phys. Rev. B 63, 195104 (2001); W. J. Padilla, D. Mandrus and D. N. Basov, Phys. Rev. B 66, 035120 (2002). Y. H. Matsuda, J. L. Her, S. Michimura, T. Inami, M. Suzuki, N. Kawamura, M. Mizumaki, K. Kindo, J. Yamauara, and Z. Hiroi, Phys. Rev. B 84, 174431 (2011). A. Koda, R. Kadono, K. Ohishi, S. R. Saha, W. Higemoto, S. Yonezawa, Y. Muraoka, Z. Hiroi, J. Phys. Soc. Japan 76, 063703 (2007). D. J. Singh, P. Blaha, K. Schwarz and J. O. Sofo, Phys. Rev. B 65, 155109 (2002); H. Harima, J. Phys. Chem. Solids 63, 1035 (2002). H. Shinaoka, T. Miyake, S. Ishibashi, arXiv:1111.6347 (2011). Y.G. Shi, Y.F. Guo, S. Yu, M. Arai, A.A. Belik, A. Sato, K. Yamaura, E. Takayama-Muromachi, H.F. Tian, H.X. Yang, J.Q. Li, T. Varga, J.F. Mitchell, and S. Okamoto, Phys. Rev. B 80, 161104 (2009). A. S. Erickson, S. Misra, G. J. Miller, R. R. Gupta, Z. Schlesinger, W. A. Harrison, J. M. Kim, and I. R. Fisher, Phys. Rev. Lett. 99, 016404 (2007). K.-W. Lee, W. E. Pickett, EPL 80, 37008 (2007). Z. Hiroi, S. Yonezawa, Y. Muraoka, J. Phys. Soc. Jpn. 73, 1651 (2004); , R. Saniz, J. E. Medvedeva, L.-H. Ye, T. Shishidou, and A. J. Freeman, Phys. Rev. B 70, 100505 (2004). J. Kuneš, T. Jeong and W.E. Pickett, Phys. Rev. B 70, 174510 (2004). J. C. Slater, Phys. Rev. 82, 538 (1951). P. Blaha, K. Schwarz, G. K. H. Madsen, D. Kvasnicka, and J. Luitz, WIEN2k, An Augmented Plane Wave + Local Orbitals Program for Calculating Crystal Properties (Karlheinz Schwarz, Technische Universität Wien, Austria), 2001, ISBN3–9501031–1–2. K. Maiti, Solid State Commun. 149, 1351 (2009); K. Maiti, Phys. Rev. B 73, 115119 (2006). D. J. Singh, J. Appl. Phys. 79, 4818 (1996); A. T. Zayak, X. Huang, J. B. Neaton, and K. M. Rabe, Phys. Rev. B 77, 214410 (2008); X. Wan, J. Zhou, and J. Dong, Europhys. Lett. 92, 57007 (2010). D.D. Koelling, B.N. Harmon, J. Phys. C 10, 3107 (1977). V.I. Anisimov, F. Aryasetiawan, and A.I. Lichtenstein, J. Phys.: Condens. Matter 9, 767 (1997). G. Kotliar, S. Y. Savrasov, K. Haule and V. S. Oudovenko, Rev. Mod. Phys. 78, 865 (2006). R. Arita, J. Kuneš, A.V. Kozhevnikov, A.G. Eguiluz, M. Imada, arXiv:1107.0835 (2011). M. A. Laguna-Marco1, D. Haskel, N. Souza-Neto, J. C. Lang, V. V. Krishnamurthy, S. Chikara, G. Cao, and M. van Veenendaal, Phys. Rev. Lett. 105, 216407 (2010). T. Moriya, Spin fluctuations in itinerant electron magnetism, (Springer-Verlag, 1985). X. Wan, M. Kohno and X. Hu, Phys. Rev. Lett. 94, 087205 (2005). [^1]: Corresponding author: [email protected]	{ "pile_set_name": "ArXiv" }
--- abstract: 'The strong decays of the $\pi(2070)$, $\eta(2010)$, $\eta(2100)$, $\eta(2190)$, and $\eta(2225)$ as the $4\,^1S_0$ quark-antiquark states are investigated in the framework of the $^3P_0$ meson decay model. It is found that the $\pi(2070)$, $\eta(2100)$, and $\eta(2225)$ appear to be the convincing $4\,^1S_0$ $q\bar{q}$ states while the assignment of the $\eta(2010)$ and $\eta(2190)$ as the $4\,^1S_0$ isoscalar states is not favored by their widths. In the presence of the $\pi(2070)$, $\eta(2100)$, and $\eta(2225)$ being the members of the $4\,^1S_0$ meson nonet, the $4\,^1S_0$ kaon is phenomenologically determined to has a mass of about 2153 MeV. The width of this unobserved kaon is expected to be about 197 MeV in the $^3P_0$ decay model.' author: - \| De-Min Li[^1] and Shan Zhou\ Department of Physics, Zhengzhou University, Zhengzhou, Henan 450052, People’s Republic of China title: 'Towards the assignment for the $4\,^1S_0$ meson nonet ' --- -1cm mesons, $^3P_0$ model [PACS numbers:]{}14.40. Cs, 12.39.Jh Introduction ============ From PDG2006[@pdg2006], the $1\,^1S_0$ meson nonet ($\pi$, $\eta$, $\eta^\prime$, and $K$) as well as the $2\,^1S_0$ members \[$\pi(1300)$, $\eta(1295)$, and $\eta(1475)$\] has been well established. In Ref.[@eta1835], we suggested that the $\pi(1800)$, $K(1830)$, together with the $X(1835)$ and $\eta(1760)$ observed by BES Collaboration[@x1835; @BES17602] constitute the $3\,^1S_0$ meson nonet. More recently, we argued that the $\eta(2225)$ with a mass of $(2240^{+30+30}_{-20-20})$ MeV and a width of $(190\pm 30^{+40}_{-60})$ MeV observed by the BES Collaboraton[@BES08] could be the $4\,^1S_0$ $s\bar{s}$ in Ref. [@eta2225] where the other members of the $4\,^1S_0$ meson nonet were not discussed. In the present work, we shall address the possible assignment for the $4\,^1S_0$ meson nonet. With the assignment of the $\eta(2225)$ as the $s\bar{s}$ member of the $4\,^1S_0$ meson nonet, one can expect that both the isovector and another isoscalar members of the $4\,^1S_0$ meson nonet should be lighter than the $\eta(2225)$. Experimentally, in the mass region 2000-2225 MeV, the pseudoscalar states $\pi(2070)$ \[Mass: $2070\pm 35$ MeV, Width: $310^{+100}_{-50}$ MeV\], $\eta(2010)$ \[Mass: $2010^{+35}_{-60}$ MeV, Width: $270\pm 60$ MeV\], $\eta(2100)$ \[Mass: $2103\pm 50$ MeV, Width: $187\pm 75$ MeV\], and $\eta(2190)$ \[Mass: $2190\pm 50$ MeV, Width: $850\pm 100$ MeV\] are reported[@pdg2006]. Theoretically, some predicted values for the $\pi(4\,^1S_0)$ mass are 2.15 GeV by QCD sum rules[@qcdsump1; @qcdsump2], 2.009 GeV by the spectrum integral equation[@spectrum] , 2.193 GeV by a covariant quark model[@covqm], 2.039 GeV by a relativistic independent quark model[@riqm], and 2.07 GeV by Regge phenomenology[@reggeph]. In addition, the mass of the third radial excitation of the $\eta$ is predicted to be about 2.267 GeV by a covariant quark model[@covqm] or 2.1 GeV by Regge phenomenology[@reggeph]. The $\pi(2070)$ mass is similar to the predicted $\pi(4\,^1S_0)$ mass, and all the masses of the $\eta(2010)$, $\eta(2100)$, and $\eta(2190)$ are close to the predicted mass range of the third radial excitation of the $\eta$. Only the mass information of these states is insufficient to classify them. The main purpose of this work is to discuss whether these reported pseudoscalar states can be assigned as the members of the $4\,^1S_0$ meson nonet or not by investigating their decay properties in the $^3P_0$ meson decay model. The organization of this paper is as follows. In section 2, the brief review of the $^3P_0$ decay model is given (for the detailed review see [e.g.]{} Refs.[@3p0rev1; @3p0rev2; @3p0rev3; @3p0rev4].) In sections 3 and 4, the decay widths of the $\pi(2070)$, $\eta(2010)$, $\eta(2100)$, $\eta(2190)$, and $\eta(2225)$ as the $4\,^1S_0$ $q\bar{q}$ state are presented. The decay widths of the $4\,^1S_0$ kaon are predicted in section 5, and the summary and conclusion are given in section 6. The $^3P_0$ meson decay model ============================== The $^3P_0$ decay model, also known as the quark-pair creation model, was originally introduced by Micu[@micu] and further developed by Le Yaouanc et al.[@3p0rev1]. The $^3P_0$ decay model which ( in several variants) is the standard model for strong decays at least for mesons in the initial state, has been widely used to evaluate the strong decays of hadrons[@3p00; @3p0y; @3p0x; @3p0x1; @3p0x2; @3p01; @3p02; @3p03; @quarkmass; @3p04], since it gives a good description of many of the observed decay amplitudes and partial widths of the hadrons. The main assumption of the $^3P_0$ decay model is that strong decays take place via the creation of a $^3P_0$ quark-antiquark pair from the vacuum. The new produced quark-antiquark pair, together with the $q\bar{q}$ within the initial meson regroups into two outgoing mesons in all possible quark rearrangement ways, which corresponds to the two decay diagrams as shown in Fig.1 for the meson decay process $A\rightarrow B+C$. The transition operator $T$ of the decay $A\rightarrow BC$ in the $^3P_0$ model is given by $$\begin{aligned} T=-3\gamma\sum_m\langle 1m1-m\|00\rangle\int d^3\vec{p}_3d^3\vec{p}_4\delta^3(\vec{p}_3+\vec{p}_4){\cal{Y}}^m_1(\frac{\vec{p}_3-\vec{p}_4}{2})\chi^{34}_{1-m}\phi^{34}_0\omega^{34}_0b^\dagger_3(\vec{p}_3)d^\dagger_4(\vec{p}_4),\end{aligned}$$ where $\gamma$ is a dimensionless parameter representing the probability of the quark-antiquark pair $q_3\bar{q}_4$ with $J^{PC}=0^{++}$ creation from the vacuum, $\vec{p}_3$ and $\vec{p}_4$ are the momenta of the created quark $q_3$ and antiquark $\bar{q}_4$, respectively. $\phi^{34}_{0}$, $\omega^{34}_0$, and $\chi_{{1,-m}}^{34}$ are the flavor, color, and spin wave functions of the $q_3\bar{q}_4$, respectively. The solid harmonic polynomial ${\cal{Y}}^m_1(\vec{p})\equiv\|p\|^1Y^m_1(\theta_p,\phi_p)$ reflects the momentum-space distribution of the $q_3\bar{q}_4$ . For the meson wave function, we adopt the mock meson $\|A(n^{2S_A+1}_AL_{A}\,\mbox{}_{J_A M_{J_A}})(\vec{P}_A)\rangle$ defined by[@mock] $$\begin{aligned} \|A(n^{2S_A+1}_AL_{A}\,\mbox{}_{J_A M_{J_A}})(\vec{P}_A)\rangle &\equiv& \sqrt{2E_A}\sum_{M_{L_A},M_{S_A}}\langle L_A M_{L_A} S_A M_{S_A}\|J_A M_{J_A}\rangle\nonumber\\ &&\times \int d^3\vec{p}_A\psi_{n_AL_AM_{L_A}}(\vec{p}_A)\chi^{12}_{S_AM_{S_A}}\phi^{12}_A\omega^{12}_A\nonumber\\ &&\times \|q_1({\scriptstyle \frac{m_1}{m_1+m_2}}\vec{P}_A+\vec{p}_A)\bar{q}_2 ({\scriptstyle\frac{m_2}{m_1+m_2}}\vec{P}_A-\vec{p}_A)\rangle,\end{aligned}$$ where $m_1$ and $m_2$ are the masses of the quark $q_1$ with a momentum of $\vec{p}_1$ and the antiquark $\bar{q}_2$ with a momentum of $\vec{p}_2$, respectively. $n_A$ is the radial quantum number of the meson $A$ composed of $q_1\bar{q}_2$. $\vec{S}_A=\vec{s}_{q_1}+\vec{s}_{q_2}$, $\vec{J}_A=\vec{L}_A+\vec{S}_A$, $\vec{s}_{q_1}$ ($\vec{s}_{q_2}$) is the spin of $q_1$ ($q_2$), $\vec{L}_A$ is the relative orbital angular momentum between $q_1$ and $q_2$. $\vec{P}_A=\vec{p}_1+\vec{p}_2$, $\vec{p}_A=\frac{m_1\vec{p}_1-m_1\vec{p}_2}{m_1+m_2}$. $\langle L_A M_{L_A} S_A M_{S_A}\|J_A M_{J_A}\rangle$ is a Clebsch-Gordan coefficient, and $E_A$ is the total energy of the meson $A$. $\chi^{12}_{S_AM_{S_A}}$, $\phi^{12}_A$, $\omega^{12}_A$, and $\psi_{n_AL_AM_{L_A}}(\vec{p}_A)$ are the spin, flavor, color, and space wave functions of the meson $A$, respectively. The mock meson satisfies the normalization condition $$\begin{aligned} \langle A(n^{2S_A+1}_AL_{A}\,\mbox{}_{J_A M_{J_A}})(\vec{P}_A) \|A(n^{2S_A+1}_AL_{A}\,\mbox{}_{J_A M_{J_A}})(\vec{P}^\prime_A)\rangle=2E_A\delta^3(\vec{P}_A-\vec{P}^\prime_A).\end{aligned}$$ The $S$-matrix of the process $A\rightarrow BC$ is defined by $$\begin{aligned} \langle BC\|S\|A\rangle=I-2\pi i\delta(E_A-E_B-E_C)\langle BC\|T\|A\rangle,\end{aligned}$$ with $$\begin{aligned} \langle BC\|T\|A\rangle=\delta^3(\vec{P}_A-\vec{P}_B-\vec{P}_C){\cal{M}}^{M_{J_A}M_{J_B}M_{J_C}},\end{aligned}$$ where ${\cal{M}}^{M_{J_A}M_{J_B}M_{J_C}}$ is the helicity amplitude of $A\rightarrow BC$. In the center of mass frame of meson $A$, ${\cal{M}}^{M_{J_A}M_{J_B}M_{J_C}}$ can be written as $$\begin{aligned} {\cal{M}}^{M_{J_A}M_{J_B}M_{J_C}}(\vec{P})&=&\gamma\sqrt{8E_AE_BE_C} \sum_{\renewcommand{\arraystretch}{.5}\begin{array}[t]{l} \scriptstyle M_{L_A},M_{S_A},\\\scriptstyle M_{L_B},M_{S_B},\\ \scriptstyle M_{L_C},M_{S_C},m \end{array}}\renewcommand{\arraystretch}{1}\!\! \langle L_AM_{L_A}S_AM_{S_A}\|J_AM_{J_A}\rangle\nonumber\\ &&\times\langle L_BM_{L_B}S_BM_{S_B}\|J_BM_{J_B}\rangle\langle L_CM_{L_C}S_CM_{S_C}\|J_CM_{J_C}\rangle\nonumber\\ &&\times\langle 1m1-m\|00\rangle\langle \chi^{14}_{S_BM_{S_B}}\chi^{32}_{S_CM_{S_C}}\|\chi^{12}_{S_AM_{S_A}}\chi^{34}_{1-m}\rangle\nonumber\\ &&\times[f_1I(\vec{P},m_1,m_2,m_3)+(-1)^{1+S_A+S_B+S_C}f_2I(-\vec{P},m_2,m_1,m_3)],\end{aligned}$$ with $f_1=\langle \phi^{14}_B\phi^{32}_C\|\phi^{12}_A\phi^{34}_0\rangle$ and $f_2= \langle\phi^{32}_B\phi^{14}_C\|\phi^{12}_A\phi^{34}_0\rangle$, corresponding to the contributions from Figs. 1 (a) and 1 (b), respectively, and $$\begin{aligned} I(\vec{P},m_1,m_2,m_3)&=&\int d^3\vec{p}\,\mbox{}\psi^\ast_{n_BL_BM_{L_B}} ({\scriptstyle\frac{m_3}{m_1+m_2}}\vec{P}_B+\vec{p})\psi^\ast_{n_CL_CM_{L_C}} ({\scriptstyle\frac{m_3}{m_2+m_3}}\vec{P}_B+\vec{p})\nonumber\\ &&\times\psi_{n_AL_AM_{L_A}} (\vec{P}_B+\vec{p}){\cal{Y}}^m_1(\vec{p}),\end{aligned}$$ where $\vec{P}=\vec{P}_B=-\vec{P}_C$, $\vec{p}=\vec{p}_3$, $m_3$ is the mass of the created quark $q_3$. The spin overlap in terms of Winger’s $9j$ symbol can be given by $$\begin{aligned} &&\langle \chi^{14}_{S_BM_{S_B}}\chi^{32}_{S_CM_{S_C}}\|\chi^{12}_{S_AM_{S_A}}\chi^{34}_{1-m}\rangle=\nonumber\\ &&\sum_{S,M_S}\langle S_BM_{S_B}S_CM_{S_C}\|SM_S\rangle\langle S_AM_{S_A}1-m\|SM_S\rangle\nonumber\\ &&\times(-1)^{S_C+1}\sqrt{3(2S_A+1)(2S_B+1)(2S_C+1)}\left\{\begin{array}{ccc} \frac{1}{2}&\frac{1}{2}&S_A\\ \frac{1}{2}&\frac{1}{2}&1\\ S_B&S_C&S \end{array}\right\}.\end{aligned}$$ In order to compare with the experiment conventionally, ${\cal{M}}^{M_{J_A}M_{J_B}M_{J_C}}(\vec{P})$ can be converted into the partial amplitude by a recoupling calculation[@recp] $$\begin{aligned} {\cal{M}}^{LS}(\vec{P})&=& \sum_{\renewcommand{\arraystretch}{.5}\begin{array}[t]{l} \scriptstyle M_{J_B},M_{J_C},\\\scriptstyle M_S,M_L \end{array}}\renewcommand{\arraystretch}{1}\!\! \langle LM_LSM_S\|J_AM_{J_A}\rangle\langle J_BM_{J_B}J_CM_{J_C}\|SM_S\rangle\nonumber\\ &&\times\int d\Omega\,\mbox{}Y^\ast_{LM_L}{\cal{M}}^{M_{J_A}M_{J_B}M_{J_C}} (\vec{P}).\end{aligned}$$ If we consider the relativistic phase space, the decay width $\Gamma(A\rightarrow BC)$ in terms of the partial wave amplitudes is $$\begin{aligned} \Gamma(A\rightarrow BC)= \frac{\pi P}{4M^2_A}\sum_{LS}\|{\cal{M}}^{LS}\|^2. \label{width1}\end{aligned}$$ Here $P=\|\vec{P}\|$=$\frac{\sqrt{[M^2_A-(M_B+M_C)^2][M^2_A-(M_B-M_C)^2]}}{2M_A}$, $M_A$, $M_B$, and $M_C$ are the masses of the meson $A$, $B$, and $C$, respectively. The decay width can be derived analytically if the simple harmonic oscillator (SHO) approximation for the meson space wave functions is used. In momentum-space, the SHO wave function is $$\begin{aligned} \psi_{nLM_L}(\vec{p})=R^{\mbox{\tiny SHO}}_{nL}(p)Y_{LM_L}(\Omega_p),\end{aligned}$$ where the radial wave function is given by $$\begin{aligned} R^{\mbox{\tiny SHO}}_{nL}=\frac{(-1)^n(-i)^L}{\beta^{\frac{3}{2}}} \sqrt{\frac{2n!}{\Gamma(n+L+\frac{3}{2})}}\left(\frac{p}{\beta}\right )^L e^{-\frac{p^2}{2\beta^2}}L^{L+\frac{1}{2}}_n({\scriptstyle \frac{p^2}{\beta^2}}).\end{aligned}$$ Here $\beta$ is the SHO wave function scale parameter, and $L^{L+\frac{1}{2}}_n({\scriptstyle \frac{p^2}{\beta^2}})$ is an associated Laguerre polynomial. The SHO wave functions cannot be regarded as realistic, however, they are a [[de facto]{}]{} standard for many nonrelativistic quark model calculations. Moreover, the more realistic space wave functions such as those obtained from Coulomb, plus the linear potential model do not always result in systematic improvements due to the inherent uncertainties of the $^3P_0$ decay model itself[@3p0y; @3p0x; @3p0x2]. The SHO wave function approximation is commonly employed in the $^3P_0$ decay model in literature. In the present work, the SHO wave function approximation for the meson space wave functions is taken. Under the SHO wave function approximation, the parameters used in the $^3P_0$ decay model involve the $q\bar{q}$ pair production strength parameter $\gamma$, the SHO wave function scale parameter $\beta$, and the masses of the constituent quarks. In the present work, we take $\gamma=8.77$ and $\beta_A=\beta_B=\beta_C=\beta=0.4$ GeV, the values recently obtained by fitting 32 experimentally well-determined decay rates with the $^3P_0$ decay model[^2], and $m_u=m_d=0.33$ GeV, $m_s=0.55$ GeV[@quarkmass]. The meson masses used to determine the phase space and final state momenta are[^3] $M_{\pi}=138$ MeV , $M_K=496$ MeV, $M_{\eta}=548$ MeV, $M_{\eta^\prime}=958$ MeV, $M_{\rho}=776$ MeV, $M_{K^\ast}=894$ MeV, $M_{\omega}=783$ MeV, $M_{\phi}=1019$ MeV, $M_{a_2(1320)}=1318$ MeV, $M_{K^\ast_2(1430)}=1429$ MeV, $M_{f_2(1270)}=1275$ MeV, $M_{f^\prime_2(1525)}=1525$ MeV, $M_{\pi(1300)}=1240$ MeV, $M_{a_0(1450)}=1474$ MeV, $M_{K^\ast_0(1430)}=1414$ MeV, $M_{f_0(1370)}=1370$ MeV, $M_{\rho(1450)}=1459$ MeV, $M_{\omega(1420)}=1420$ MeV, $M_{K^\ast(1580)}=1580$ MeV, $M_{\rho(1700)}=1720$ MeV, $M_{K^\ast(1680)}=1717$ MeV, and $M_{K^\ast_3(1780)}=1776$ MeV. Decays of the $\pi(2070)$ ========================= From (\[width1\]), the numerical values of the partial decay widths of the $\pi(2070)$ as the $4\,^1S_0$ isovector state are listed in Table 1. The initial state mass is set to $2070$ MeV. [cc\|cc]{}Mode & $\Gamma_i$ (MeV)& Mode & $\Gamma_i$ (MeV)\ $\rho\omega$ & 3.1& $\rho\pi$ & 5.9\ $\pi(1300)\rho$& 52.0& $\rho(1700)\pi$& 3.6\ $\rho(1450)\pi$& 112.5& $ f_2(1270)\pi$ & 48.0\ $f_0(1370)\pi $ & 0.3& $ a_2(1320)\eta$ & 8.8\ $ a_0(1450)\eta$ & 5.1& $K K^\ast$ & 8.5\ $ K^\ast K^\ast$ & 14.9& $ K^\ast_0(1430)K$& 10.3\ $ K^\ast_2(1430)K$ & 4.6&\ \ Table 1 indicates that the total width of the $\pi(2070)$ as the $4\,^1S_0$ isovector state predicted by the $^3P_0$ decay model is about $278$ MeV, consistent with the observation $(310^{+100}_{-50})$ MeV within errors, and the dominant decay modes are expected to be $\pi(1300)\rho$, $\rho(1450)\pi$ and $ f_2(1270)\pi$. Also, in order to check the dependence of the theoretical result on the initial state mass, the predicted total width of the $\pi(2070)$ is shown in Fig. 2 as the function of the initial state mass. Fig. 2 shows that when the initial state mass varies from 2035 to 2105 MeV, the total width of the $4\,^1S_0$ isovector state varies from about 230 to 320 MeV, generally in accord with the width range of the $\pi(2070)$. Both the mass and width of the $\pi(2070)$ are consistent with the predicted $4\,^1S_0$ isovector state, which therefore suggests that the assignment of the $\pi(2070)$ as the $4\,^1S_0$ isovector state seams plausible. Decays of the $\eta(2010)$, $\eta(2100)$, $\eta(2190)$, and $\eta(2225)$ ======================================================================== In the presence of the $\eta(2225)$ being one isoscalar member of the $4\,^1S_0$ meson nonet, the $\eta(2010)$, $\eta(2100)$, and $\eta(2190)$ would complete another isoscalar member. It is well known that in a meson nonet, the two physical isoscalar states can mix. The mixing of the two isoscalar states can be parameterized as $$\begin{aligned} &&\eta(x)=\cos\phi~ n\bar{n}-\sin\phi~ s\bar{s},\\ &&\eta(2225)=\sin\phi~ n\bar{n}+\cos\phi~ s\bar{s},\end{aligned}$$ where $n\bar{n}=(u\bar{u}+d\bar{d})/\sqrt{2}$ and $s\bar{s}$ are the pure $4\,^1S_0$ nonstrange and strange states, respectively, and $\eta(x)$ denotes the $\eta(2010)$, $\eta(2100)$, or $\eta(2190)$. According to (\[width1\]), the partial widths of $\eta(x)$ and $\eta(2225)$ become with mixing $$\begin{aligned} &&\Gamma(\eta(x)\rightarrow BC)=\frac{\pi~P}{4M^2_{\eta(x)}}\sum_{LS}\|\cos\phi {\cal{M}}^{LS}_{n\bar{n}\rightarrow BC}-\sin\phi {\cal{M}}^{LS}_{s\bar{s}\rightarrow BC}\|^2, \label{w1}\\ &&\Gamma(\eta(2225)\rightarrow BC)=\frac{\pi~P}{4M^2_{\eta(2225)}}\sum_{LS}\|\sin\phi {\cal{M}}^{LS}_{n\bar{n}\rightarrow BC}+\cos\phi {\cal{M}}^{LS}_{s\bar{s}\rightarrow BC}\|^2. \label{w2}\end{aligned}$$ Based on (\[w1\]) and (\[w2\]), the predicted total widths of the $\eta(2010)$, $\eta(2100)$, $\eta(2190)$, and $\eta(2225)$ are shown in Fig. 3 as functions of the initial state mass and the mixing angle $\phi$. From Fig. 3, one can see that with the variations of the initial state mass and $\phi$, only the measured widths of the $\eta(2100)$ and $\eta(2225)$ are possible to be reasonably reproduced in the $^3P_0$ model. We therefore suggest that the assignment of the $\eta(2010)$ and $\eta(2190))$ as the $4\,^1S_0$ isoscalar states seems unfavorable. We shall focus on the possibility of the $\eta(2100)$ being the partner of the $\eta(2225)$. Taking $M_{\eta(2100)}=2103$ MeV and $M_{\eta(2225)}=2240$ MeV, we list the numerical values of the partial decay widths of the $\eta(2100)$ and $\eta(2225)$ in Table 2. The variation of the theoretical total widths of the $\eta(2100)$ and $\eta(2225)$ with the mixing angle $\phi$ is shown in Fig. 4. -------------------- ------------------------- --------------------------- Mode $\Gamma_i$(MeV) $\Gamma_i$(MeV) $\rho\rho$ $2.1c^2$ $1.4s^2$ $\omega\omega$ $0.9c^2$ $0.3s^2$ $\phi\phi$ $9.7s^2$ $20.1c^2$ $a_2(1320)\pi$ $143.7c^2$ $158.8s^2$ $a_0(1450)\pi$ $0.1c^2$ $4.5s^2$ $KK^\ast$ $7.4c^2-15.0cs+7.6s^2$ $14.4c^2+12.7cs+2.8s^2$ $K^\ast K^\ast$ $15.4c^2-13.4cs+2.9s^2$ $0.8c^2-6.6cs+13.6s^2$ $KK^\ast_0(1430)$ $8.8c^2-7.9cs+1.8s^2$ $2.5c^2-4.7cs+2.2s^2$ $K K^\ast_2(1430)$ $7.7c^2-31.1cs+31.3s^2$ $69.4c^2+91.9cs+30.4s^2$ $KK^\ast(1580)$ $5.4c^2+17.2cs+13.6s^2$ $90.7c^2-158.1cs+68.9s^2$ $K K^\ast(1680)$ $1.6c^2-1.4cs+0.3s^2$ -------------------- ------------------------- --------------------------- : Decays of the $\eta(2100)$ and $\eta(2225)$ as the $4\,^1S_0$ isoscalar states in the $^3P_0$ model. $c\equiv\cos\phi$, $s\equiv\sin\phi$. From Fig. 4, we find that if the $\eta(2100)-\eta(2225)$ mixing angle $\phi$ lying in the range from about $-0.6$ to $+0.7$ radians, both the measured widths of the $\eta(2100)$ and $\eta(2225)$ can be reasonably reproduced. In order to check whether the possibility of $-0.6 \leq \phi \leq +0.7$ radians exists or not, below we shall estimate the $\eta(2100)$-$\eta(2225)$ mixing angle phenomenologically. In the $n\bar{n}$ and $s\bar{s}$ basis, the mass-squared matrix describing the $\eta(2100)$ and $\eta(2225)$ mixing can be written as[@jpglidm; @epja-1] $$\begin{aligned} M^2=\left(\begin{array}{cc} M^2_{n\bar{n}}+2A_m&\sqrt{2}A_mX\\ \sqrt{2}A_mX&M_{s\bar{s}}^2+A_mX^2 \end{array}\right), \label{matrix}\end{aligned}$$ where $M_{n\bar{n}}$ and $M_{s\bar{s}}$ are the masses of the states $n\bar{n}$ and $s\bar{s}$, respectively, $A_m$ denotes the total annihilation strength of the $q\bar{q}$ pair for the light flavors $u$ and $d$, $X$ describes the $SU(3)$-breaking ratio of the nonstrange and strange quark masses via the constituent quark mass ratio $m_u/m_s$. The masses of the two physical states $\eta(2100)$ and $\eta(2225)$ can be related to the matrix $M^2$ by the unitary matrix $U=\left(\begin{array}{cc} \cos\phi&-\sin\phi\\ \sin\phi&\cos\phi\end{array}\right)$ $$\begin{aligned} U M^2 U^\dagger=\left(\begin{array}{cc} M^2_{\eta(2100)}&0\\ 0&M^2_{\eta(2225)}\end{array}\right). \label{diag}\end{aligned}$$ $n\bar{n}$ is the orthogonal partner of the $\pi(4^1S_0)$, the isovector state of $4\,^1S_0$ meson nonet, and one can expect that $n\bar{n}$ degenerates with $\pi(4\,^1S_0)$ in effective quark masses, here we take $M_{n\bar{n}}=M_{\pi(4^1S_0)}=M_{\pi(2070)}$. With the help of the Gell-Mann-Okubo mass formula $M^2_{s\bar{s}}=2M^2_{ K(4^1S_0)}-M^2_{n\bar{n}}$[@okubo], the following relations can be derived from (\[diag\]) $$\begin{aligned} 8X^2(M^2_{K(4^1S_0)}-M^2_{\pi(2070)})^2&=&[4M^2_{K(4^1S_0)}-(2-X^2)M^2_{\pi(2070)}-(2+X^2)M^2_{\eta(2100)}]\nonumber\\ &&\times[(2-X^2)M^2_{\pi(2070)}+(2+X^2)M^2_{\eta(2225)}-4M^2_{K(4^1S_0)}], \label{msch}\end{aligned}$$ $$\begin{aligned} A_m=\frac{(M^2_{\eta(2225)}-2M^2_{K(4^1S_0)}+M^2_{\pi(2070)})(M^2_{\eta(2100)}-2M^2_{K(4^1S_0)}+M^2_{\pi(2070)})}{2(M^2_{\pi(2070)}-M^2_{K(4^1S_0)})X^2}. \label{am}\end{aligned}$$ If the $SU(3)$-breaking effect is not considered, i.e., $X=1$, relation (\[msch\]) can be reduced to Schwinger’s original nonet mass formula[@schwinger]. Taking $X=m_u/m_s=0.33/0.55=0.6$, from (\[msch\]) and (\[am\]) we have $$\begin{aligned} M_{K(4\,^1S_0)}=2.153~ \mbox{GeV},~ A_m=0.07~ {\mbox{GeV}}^2.\end{aligned}$$ Based on the values of the above parameters involved in (\[matrix\]), the unitary matrix $U$ can be given by $$\begin{aligned} U=\left(\begin{array}{cc} \cos\phi&-\sin\phi\\ \sin\phi&\cos\phi\end{array}\right)=\left(\begin{array}{cc} +0.995&-0.104\\ +0.104&+0.995 \end{array}\right), \label{mixangle}\end{aligned}$$ which gives $\phi= +0.1$ radians, just lying in the range from about $-0.6$ to $+0.7$ radians. From Table 2, this estimated mixing angle leads to $\Gamma_{\mbox{thy}}(\eta(2100))=185.2$ MeV and $\Gamma_{\mbox{thy}}(\eta(2225))=193.7$ MeV, both in good agreement with the experimental results. The $\eta(2100)$ and $\eta(2225)$, together with the $\pi(2070)$ therefore appear to be the convincing $4\,^1S_0$ states. Decays of the $4\,^1S_0$ kaon ============================== The above two sections show that in the presence of the $\pi(2070)$, $\eta(2100)$, and $\eta(2225)$ belonging to the $4\,^1S_0$ meson nonet, the total widths of these three states can be naturally accounted for in the $^3P_0$ decay model, and the $4\,^1S_0$ kaon is expected to have a mass of about 2153 MeV by the mass formula (\[msch\]). Below the $K(2150)$ denotes the $4\,^1S_0$ kaon. We note that the $K$, $K(1460)$[^4], $K(1830)$, and $K(2150)$ approximately populate a common trajectory as shown in Fig. 5. The quasi-linear trajectories at the (n, Mass-squared)-plots turned out to be able to describe the light mesons with a good accuracy[@reggeph]. Fig. 5 therefore indicates that the $K(1460)$, $K(1830)$, and $K(2150)$ could be the good candidates for the $2\,^1S_0$, $3\,^1S_0$, and $4\,^1S_0$ kaons, respectively. The $K(2150)$ is not related any current experimental candidate. The predicted decay widths of the $K(2150)$ are listed in Table 3. The initial state mass is set to 2153 MeV. The total width of the $K(2150)$ is predicted to be about 197 MeV, and the dominant decay modes of the $K(2150)$ are expected to be $ K^\ast_2(1430)\pi$, $ K^\ast(1580)\pi$, $\rho(1459)K$ and $a_2(1320)K$. These results could be of use in searching the candidate for the $4\,^1S_0$ kaon experimentally. [cc\|cc]{}Mode & $\Gamma_i$ (MeV)&Mode & $\Gamma_i$ (MeV)\ $\rho K$ & 3.9& $\omega K$ & 1.3\ $\phi K$ & 1.3& $\rho K^\ast$ & 0.08\ $\omega K^\ast$& 0.04& $\phi K^\ast$ & 12.0\ $\pi K^\ast$ & 4.5& $\eta K^\ast$& 1.5\ $\eta^\prime K^\ast$& 0.09& $ K^\ast_0(1430)\pi$& 1.6\ $ K^\ast_2(1430)\pi$& 29.7& $ K^\ast(1580)\pi$& 34.5\ $ K^\ast(1680)\pi$& 2.5& $ K^\ast_3(1780)\pi$& 1.0\ $\pi(1300) K^\ast$& 6.3& $\rho(1450) K$& 42.8\ $\omega(1420) K$& 14.6& $a_2(1320) K$& 26.7\ $f_2(1270)K$ & 9.9& $ f^\prime_2(1525) K$& 2.9\ \ Summary and conclusion ====================== With the assignment of the $\eta(2225)$ recently observed by the BES Collaboration as the $s\bar{s}$ member of the $4\,^1S_0$ meson nonet, the possibility of the $\pi(2070)$, $\eta(2010)$, $\eta(2100)$, and $\eta(2190)$ being the $4\,^1S_0$ $q\bar{q}$ states is discussed. With respect to the $\pi(2070)$, its assignment to the $4\,^1S_0$ isovector state is not only favored by its mass, but also by its width. The assignment of the $\eta(2010)$ and $\eta(2190)$ as the $4\,^1S_0$ isoscalar states is not favored by their widths. Both the widths of the $\eta(2100)$ and $\eta(2225)$ can be reasonably reproduced with the mixing angle lying in the range from about $-0.6$ to $+0.7$ radians. The assignment of the $\pi(2070)$, $\eta(2100)$, and $\eta(2225)$ as the members of the $4\,^1S_0$ meson nonet not only leads to that the $\eta(2100)$-$\eta(2225)$ mixing angle is about $+0.1$ radians which naturally accounts for the widths of the $\eta(2100)$ and $\eta(2225)$, but also gives that the $4\,^1S_0$ kaon has a mass of about 2153 MeV. The $K$, $K(1460)$, $K(1830)$, and $K(2150)$ approximately populate a common (n, Mass-squared)-trajectory. We tend to conclude that the observed pseudoscalar states $\pi(2070)$, $\eta(2100)$, $\eta(2225)$, together with the unobserved $K(2150)$ appear to be the good candidates for the members of the $4\,^1S_0$ meson nonet. The $K(2150)$ width is predicted to be about 197 MeV, and the dominant decay modes of the $K(2150)$ are expected to be $ K^\ast_2(1430)\pi$, $ K^\ast(1580)\pi$, $\rho(1459)K$ and $a_2(1320)K$. These results could be of use in searching the candidate for the $4\,^1S_0$ kaon experimentally. Acknowledgments {#acknowledgments .unnumbered} =============== This work is supported in part by HANCET under Contract No. 2006HANCET-02, and the Program for Youthful Teachers in University of Henan Province. [99]{} W.-M. Yao et al., J. Phys. G [33]{}, 1 (2006). De-Min Li and Bing Ma, Phys. Rev. D [77]{}, 074004 (2008). M. Ablikim et al., (BES Collaboration), Phys. Rev. Lett. [95]{}, 262001 (2005). M. Ablikim et al., (BES Collaboration), Phys. Rev. D [73]{}, 112007 (2006). M. Ablikim et al., (BES Collaboration), Phys. Lett. B [662]{}, 330 (2008). De-Min Li and Bing Ma, Phys. Rev. D [77]{}, 094021 (2008). N. V. Karsnikov and A. A. Pivovarov, Phys. Lett. B [112]{}, 397 (1982); A. L. Kataev, N. V. Karsnikov, and A. A. Pivovarov, Phys. Lett. B [123]{}, 93 (1983); S. G. Gorishny, A. L. kataev, and S. A. Larin, Phys. lett. B [135]{}, 457 (1984); A. L. Kataev, Phys. Atom. Nucl. [68]{}, 567 (2005). S. S. Afonin, A. A. Andrianov, A. V. Andrianov, and D. Espriu, JHEP, [0404]{}, 039 (2004). V. V. Anisovich, L. G. Dakhno, M. A. Matveev, V. A. Nikonov, and A. V. Sarantsev, Phys. Atom. Nucl. [70]{}, 450 (2007). R. Ricken, M. Koll, D. Merten, B. C. Metsch, and H. R. Petry, Eur. Phys. J. A [9]{}, 221 (2000). V. V. Khruschev, Yad. Fiz. [46]{}, 219 (1987); V. V. Khruschev, Yad. Fiz.[55]{}, 773 (1992); S. V. Semenov and V. V. Khruschev, Yad. Fiz. [56]{}, 218 (1993); V. I. Savrin and S. V. Semenov, Phys. Lett. B [374]{}, 159 (1996); V. V. Khruschev, hep-ph/0504077. A. V. Anisovich, V. V. Anisovich, and A. V. Sarantsev, Phys. Rev. D [62]{}, 051502(R) (2000). A. Le Yaouanc, L. Oliver, O. Pene, and J-C. Raynal, Phys. Rev. D [8]{}, 2223 (1973); Phys. Rev. D [9]{}, 1415 (1974); Phys. Rev. D [11]{}, 1272 (1975); Phys. Lett. B [71]{}, 397 (1977); Phys. Lett. B [72]{},57 (1977). A. Le Yaouanc, L. Oliver, O. Pene, and J-C. Raynal, Hadron transitons in the quark model ( Gordon and Breach Science Publishers, New York, 1988). W. Roberts and B. Silvestr-Brac, Few-Body Syst. [11]{}, 171 (1992). H. G. Blundel, hep-ph/9608473. L. Micu, Nucl. Phys. B [10]{}, 521 (1969). S. Capstick and N. Isgur, Phys. Rev. D [34]{}, 2809 (1986); S. Capstick and W. Roberts, Phys. Rev. D [49]{} 4570 (1994). P. Geiger and E. S. Swanson, Phys. Rev. D [50]{}, 6855 (1994). H.G. Blundell and S. Godfrey, Phys. Rev. D [53]{},3700 (1996). H. G. Blundell, S. Godfrey, and B. Phelps, Phys. Rev. D [53]{}, 3712 (1996). R. Kokoski and N. Isgur, Phys. Rev. D [35]{}, 907 (1987). E. S. Ackleh, T. Barnes and E. S. Swanson, Phys. Rev. D [54]{}, 6811 (1996). T. Barnes, F. E. Close, P. R. Page and E. S. Swanson, Phys. Rev. D [55]{}, 4157 (1997). T. Barnes, N. Black and P. R. Page, Phys. Rev. D [68]{}, 054014 (2003). F. E. Close and E. S. Swanson, Phys. Rev. D [72]{}, 094004 (2005). L. Burakovsky and P. R. Page, Phys. Rev. D [62]{}, 014011 (2000); H. Q. Zhou, R. G. Ping, and B. S. Zou, Phys. Lett. B [611]{}, 123 (2005); T. Barnes, S. Godfrey and E. S. Swanson, Phys. Rev. D [72]{}, 054026 (2005); J. Lu, W. Z. Deng, X. L. Chen, and S. L. Zhu, Phys. Rev. D [73]{}, 054012 (2006); B. Zhang, X. Liu, W. Z. Deng, and S. L. Zhu, Eur. Phys. J. C [50]{}, 617 (2007); F. E. Close, C. E. Thomas, O. Lakhina, and E. S. Swanson, Phys. Lett. B [647]{}, 159 (2007); O. Lakhina and E. S. Swanson, Phys. Lett. B [650]{}, 159 (2007); C. Chen, X. L. Chen, X. Liu, W. Z. Deng, and S. L. Zhu, Phys. Rev. D [75]{}, 094017 (2007); G. J. Ding and M. L. Yan, Phys. Lett. B [657]{}, 49 (2007). C. Hayne and N. Isgur, Phys. Rev. D [25]{}, 1944 (1982). M. Jacob and G. C. Wick, Ann. Phys. [7]{}, 404 (1959). J. Vijande, F. Fernandez and A. Valcarce, J. Phys. G [31]{}, 481 (2005). S. Godfrey and N. Isgur, Phys. Rev. D [32]{}, 189 (1985). De-Min Li, Bing Ma, Xue-Chao Feng, and Hong Yu, Mod. Phys. Lett. A [20]{}, 2497 (2005). D. M. Li, H. Yu and Q. X. Shen, J. Phys. G [27]{}, 807 (2001). De-Min Li, Ke-Wei Wei and Hong Yu, Eur. Phys. J. A [25]{}, 263 (2005). S. Okubo, Prog. Theor. Phys. [27]{}, 949 (1962). J. Schwinger, Phys. Rev. Lett. [12]{}, 237 (1964). G. W. Brandenberg et al., (LASS Collaboration), Phys. Rev. Lett. [36]{}, 1239 (1976). [^1]: E-mail: [email protected] [^2]: Our value of $\gamma$ is higher than that used by Ref.[@quarkmass] (0.505) by a factor of $\sqrt{96\pi}$ due to different field conventions, constant factor in $T$, etc. The calculated results of the widths are, of course, unaffected. [^3]: The assignment the $K^\ast(1410)$ as the $2\,^3S_1$ kaon is problematic[@3p03; @vij]. Quark model[@flavorfun] and other phenomenological approaches[@mpla] consistently suggest the $2\,^3S_1$ kaon has a mass about 1580 MeV, here we take 1580 MeV as the mass of the $2\,^3S_1$ kaon \[$K^\ast(1580)$\]. Also, we assume that the $a_0(1450)$, $K^\ast_0(1430)$, and $f_0(1370)$ are the ground scalar mesons as Refs.[@3p01; @3p02; @3p03]. [^4]: The $K(1460)$ mass is taken 1400 MeV reported by[@k1460].	{ "pile_set_name": "ArXiv" }
--- abstract: 'We study a hairy black hole solution in the dilatonic Einstein-Gauss-Bonnet theory of gravitation, in which the Gauss-Bonnet term is nonminimally coupled to the dilaton field. Hairy black holes with spherical symmetry seem to be easily constructed with a positive Gauss-Bonnet (GB) coefficient $\alpha$ within the coupling function, $f(\phi) = \alpha e^{\gamma \phi}$, in an asymptotically flat spacetime, i.e., no-hair theorem seems to be easily evaded in this theory. Therefore, it is natural to ask whether this construction can be expanded into the case with the negative coefficient $\alpha$. In this paper, we numerically present the dilaton black hole solutions with a negative $\alpha$, and we analyze the properties of GB term through the aspects of the black hole mass. We construct the new integral constraint allowing the existence of the hairy solutions with the negative $\alpha$. Through this procedure, we expand the evasion of the no-hair theorem for hairy black hole solutions.' author: - 'Bum-Hoon Lee' - Wonwoo Lee - Daeho Ro bibliography: - 'ref.bib' title: 'Expanded evasion of the black hole no-hair theorem in dilatonic Einstein-Gauss-Bonnet theory' --- Introduction {#sec:1} ============ The first astrophysical black hole is Cygnus A, which was later recognized as a black hole [@Reber]. The black hole is now a real thing, a most fascinating object, and worth exploring more deeply in the Universe. It was extensively investigated both observationally and theoretically. At the same time, various theories of gravitation, inspired by string theory or astrophysics were also developed. Based on these backgrounds, a variety of black hole solutions, such as a dilaton black hole [@Gibbons:1987ps; @Horne:1992zy; @Cai:1997ii] and Gauss-Bonnet (GB) black hole [@Boulware:1985wk; @Cai:2001dz; @Cai:2003gr], have been studied. Furthermore, recent observations of a gravitational wave coming from the mergers of compact binary sources [@Abbott:2016nmj; @Abbott:2017vtc] have opened new horizons in astrophysics as well as cosmology, in which it could be very interesting to test which theory of gravitation describes our Universe and the existence of the hairy black hole [@Pani:2009wy; @TheLIGOScientific:2016src; @Berti:2015itd]. The existence of the hairy black hole solution with the GB term has been constructed and extensively studied over the past two decades [@Kanti:1995vq; @Torii:1996yi; @Kanti:1997br; @Torii:1998gm; @Guo:2008hf; @Rogatko:2014maa; @Herdeiro:2015waa; @Myung:2018iyq], in which the black hole has an exponentially decaying dilaton hair. Because of the motivation coming from string theory in general [@Boulware:1986dr; @Callan:1988hs; @Campbell:1991kz; @Mignemi:1992pm], the GB coefficient $\alpha$ is related to the Regge slope $\alpha'(=16\alpha)$ [@Torii:1996yi; @Wiltshire:1988uq], and it is always treated to be a positive constant on those works. From extensive studies, we noticed that there is the lower bound for a black hole mass, and that mass of a black hole increases when the dilaton coupling $\gamma$ increases [@Ahn:2014fwa; @Gwak:2017fea]. We thought that the GB term seems to provide a repulsive property, which makes the formation of the dilaton black hole harder, and the lower bound increases as a result. What would happen if we change the sign of the GB coefficient? In this perspective, it is interesting to consider it as a kind of modified theory of gravitation, even though the motivation from string theory would not valid. For this reason, we more deeply investigate the dilaton black hole with the negative GB coefficient. The no-hair theorem for black hole solutions was conjectured [@Ruffini:1971bza] to summarize the progress in black hole physics [@Oppenheimer:1939ue; @Kerr:1963ud; @Newman:1965my; @Israel:1967wq; @Israel:1967za; @Carter:1971zc], and developed [@Bekenstein:1972ny; @Bekenstein:1995un] in Einstein-Maxwell theory, in which the solutions are associated with Gauss’s law. In [@Bekenstein:1972ny], the author used an integral constraint obtained from the equation of motion for the scalar field. Later, it was further developed into the novel no-hair theorem through the analysis of an energy-momentum tensor, especially the $T_r^{\ r}$ component [@Bekenstein:1995un]. If a black hole has the dilaton hair in the dilatonic Einstein-Gauss-Bonnet (DEGB) theory, the no-hair theorem should be avoided. Recently, we have seen that the no-hair theorem is bypassed for the black holes with a dilaton hair in DEGB theory [@Antoniou:2017acq; @Antoniou:2017hxj]; by presenting both, the old no-hair theorem is easily evaded and the novel no-hair theorem is not applicable for DEGB theory. However, the GB coupling functions were positive definite in their analysis. For your interests, the no-hair theorem can also be evaded in the extended scalar-tensor-Gauss-Bonnet gravity [@Doneva:2017bvd; @Doneva:2018rou] In this paper, we numerically present the dilaton black holes with a negative $\alpha$, and we analyze the properties of the GB term through the aspects of the lower bound for a black hole mass in more detail. The purpose of this paper is to provide the expanded evasion of the no-hair theorem for hairy black holes by constructing the new integral constraint to allow the existence of the dilaton black hole with arbitrary GB coefficients. The paper is organized as follows: In Sec. \[sec:2\], we review and calculate the numerical setup. We analyze the energy momentum tensor and construct the new integral constraint. In Sec. \[sec:3\], we present a dilaton black hole solution with the negative GB coefficient , and we analyze the black hole properties with respect to the dilaton coupling and GB coefficient. In Sec. \[sec:4\], we summarize our results and discuss the role of the GB term with the difference between both cases. DEGB black hole {#sec:2} =============== Let us consider the action with the GB term: $$S = \int d^4x \sqrt{-g} \left[\dfrac{R}{2} - \dfrac{1}{2} \nabla_\mu \phi \nabla^\mu \phi + f(\phi) R_{\rm GB}^2 \right] + S_{\rm b},$$ where $g=\det g_{\mu\nu}$, the coupling function with the GB term is given by $f(\phi) = \alpha e^{\gamma \phi}$, and $\phi$ is a dilaton field. The scalar curvature of the spacetime is denoted by $R$, and the GB curvature term is given by $R_{\text{GB}}^2 = R^2 - 4R_{\mu\nu}R^{\mu\nu} + R_{\mu\nu\rho\sigma} R^{\mu\nu\rho\sigma}$. In this work, the boundary term $S_{\rm b}$ [@York:1972sj; @Gibbons:1976ue; @Myers:1987yn; @Davis:2002gn] is not important, so it is abbreviated. The Einstein constant $\kappa = 8\pi G$ is set to unity for simplicity. The dilaton field equation is $$\label{eq:scalar} \dfrac{1}{\sqrt{-g}} \partial_\mu(\sqrt{-g} \partial^\mu \phi) + \dot{f}(\phi) R_{\text{GB}}^2 = 0,$$ where the dot notation denotes the derivative with respect to $\phi$, and Einstein’s equation is $$R_{\mu\nu} - \dfrac{1}{2} g_{\mu\nu} R = T_{\mu\nu} = \partial_\mu \phi \partial_\nu \phi - \dfrac{1}{2} g_{\mu\nu} \partial_\rho \phi \partial^\rho \phi + T_{\mu\nu}^{\text{GB}},$$ where $T_{\mu\nu}^{\rm GB}$ is the energy momentum tensor contributed from the GB term [@Nojiri:2005vv] as follows: $$\begin{aligned} \nonumber T_{\mu\nu}^{\rm GB} &=& 4 (\nabla_\mu \nabla_\nu f(\phi)) R - 4 g_{\mu\nu} (\nabla^2 f(\phi)) R \\ \nonumber && - 8 (\nabla_\rho \nabla_\mu f(\phi)) R_\nu{}^\rho - 8 (\nabla_\rho \nabla_\nu f(\phi)) R_\mu{}^\rho \\ \nonumber && + 8 (\nabla^2 f(\phi)) R_{\mu\nu} + 8 g_{\mu\nu} (\nabla_\rho \nabla_\sigma f(\phi)) R^{\rho\sigma} \\ && - 8 (\nabla^\rho \nabla^\sigma f(\phi)) R_{\mu\rho\nu\sigma}.\end{aligned}$$ The equation only has the derivative terms of $f(\phi)$ because the minimally coupled terms in four-dimensions are cancelled identically [@deWitt:1964], and the contribution of the GB term with the equations of motion is coming from the nonminimally coupled terms only. Let us consider the spherically symmetric static metric in an asymptotically flat spacetime as follows: $$ds^2 = - e^{X(r)} dt^2 + e^{Y(r)} dr^2 + r^2 (d\theta^2 + \sin^2 \theta d\varphi^2),$$ where the metric functions $X$ and $Y$ depend only on $r$. Then, the dilaton field and Einstein’s equations turn out to be [@Kanti:1995vq] $$\begin{aligned} \label{eq:df} 0 &=& \phi'' + \phi' \left( \dfrac{X'- Y'}{2} + \dfrac{2}{r} \right) - \dfrac{4\dot{f}}{r^2} \left( X' Y' e^{-Y} + (1-e^{-Y})\left(X'' + \dfrac{X'}{2}(X'-Y')\right)\right), \\ \label{eq:tt} 0 &=& \dfrac{r}{2}\phi'^2 + \dfrac{1-e^Y}{r} - Y' \left(1+\dfrac{4\dot{f} \phi'}{r}(1-3e^{-Y})\right) + \dfrac{8\dot{f}}{r} \left(\phi'' + \dfrac{\ddot{f}}{\dot{f}}\phi'^2 \right) (1-e^{-Y}), \\ \label{eq:rr} 0 &=& \dfrac{r}{2}\phi'^2 + \dfrac{1 - e^{Y}}{r} - X' \left(1 + \dfrac{4\dot{f} \phi'}{r}(1 - 3e^{-Y})\right), \\ \nonumber 0 &=& X'' + \left(\dfrac{X'}{2} + \dfrac{1}{2} \right)(X' - Y') + \phi'^2 - \dfrac{8\dot{f}e^{-Y}}{r} \left(\phi' X'' + \left(\phi'' + \dfrac{\ddot{f}}{\dot{f}}\phi'^2 \right) X' + \dfrac{\phi'X'}{2}(X' - 3Y') \right), \\ \label{eq:qq}\end{aligned}$$ where the prime notation denotes the derivatives with respect to $r$. Equation can be solved in terms of $Y$ as follows: $$\begin{gathered} \label{eq:ey} e^{Y(r)} = \dfrac{1}{4}\bigg(- r^2 \phi'^2 + 2r X' + 8 \dot{f} X' \phi' + 2 \\ \pm \sqrt{(- r^2 \phi'^2 + 2r X' + 8 \dot{f} X' \phi' + 2)^2 - 192 \dot{f} X' \phi' } \bigg).\end{gathered}$$ We should take the positive sign from the above equation to be valid the near the horizon limit. In terms of the above equation, $Y$ and $Y'$ can be eliminated from the equations of motion. Thus, we use the Eqs. and mainly for numerical calculation and the remaining one for constraint. For later use, it is better to calculate the GB term $R_{\rm GB}^2$, which is given by $$R_{\rm GB}^2 = \dfrac{2e^{-Y}}{r^2}[(1 - 3e^{-Y})X'Y'- (1 - e^{-Y})(X'^2+ 2X'')].$$ To perform the numerical computation, we impose the boundary conditions at the black hole horizon $r_h$ and asymptotically flat region $r \gg 1$. At the black hole horizon, the metric components should be zero such as $g_{tt}(r_h) = e^{X(r_h)} = 0$ and $g^{rr}(r_h) = e^{-Y(r_h)} = 0$. Then, the metric components and the dilaton field can be expanded in the near horizon limit by using the length parameter from the horizon, $\delta r = r - r_h$ as follows: $$\begin{aligned} \label{eq:nhexpand1} e^{X(r)} &=& 0 + x_1 \delta r + x_2 \delta r^2 + {\cal O}(\delta r^3), \\ e^{-Y(r)} &=& 0 + y_1 \delta r + y_2 \delta r^2 + {\cal O}(\delta r^3), \\ \phi(r) &=& \phi_h + \phi'_h \delta r + \phi''_h \delta r^2 + {\cal O}(\delta r^3).\end{aligned}$$ The above equations provide the boundary conditions for $X$ and $\phi$, but the value $\phi_h$ is not determined yet. To do so, we expand the Eq. in the near horizon limit: $$\begin{gathered} \label{eq:eynr} e^{Y(r)} = (r + 4\dot{f} \phi') X' \\ + \dfrac{2r - r^3 \phi'^2 - 4\dot{f} \phi'(4 + r^2 \phi'^2)}{2(r + 4\dot{f} \phi')} + {\cal O} \left(\dfrac{1}{X'}\right).\end{gathered}$$ Now, we differentiate the equation with respect to $r$ and substitute the result from Eq. into Eqs. and , as we discussed earlier. It eliminates $Y'$, and we only need to solve $X$ and $\phi$, not $Y$, when we solve the equation numerically. It is also possible to diagonalize the equations in terms of $X''$ and $\phi''$, but the result is not simple [@Kanti:1995vq; @Ahn:2014fwa; @Gwak:2017fea]. Finally, an expansion of the result in the near horizon limit gives the results [$$\begin{aligned} \nonumber X''(r) &=& - \dfrac{r^4 + 8r^3 \dot{f} \phi' + 16r^2 \dot{f}^2 \phi'^2 - 48 \dot{f}^2}{r^4 + 4r^3 \dot{f}\phi' - 96\dot{f}^2} X'^2 + {\cal O} \left(X'\right), \\ \nonumber \phi''(r) &=& - \dfrac{(r + 4\dot{f} \phi')(r^3\phi' + 4r^2 \dot{f}\phi'^2 + 12 \dot{f})}{r^4 + 4r^3 \dot{f}\phi' - 96\dot{f}^2} X'+ {\cal O} \left(1 \right). \\ \label{eq:pp}\end{aligned}$$]{} One might notice that $\phi''$ will diverge at the horizon because of $X'$ diverges. Thus, the numerator should be zero, e.g. $r^3\phi' + 4r^2 \dot{f}\phi'^2 + 12 \dot{f} = 0$ not the other factor to be consistent with $e^{Y(r)}$, which indicates the value of $\phi'_h$ that is $$\label{eq:phiph} \phi'_h = - \dfrac{r_h^2 \pm \sqrt{r_h^4 - 192 \dot{f}_h^2}}{8r_h \dot{f}_h}.$$ We take the negative sign to get the appropriate solution. It is easy to see that the positive sign will not give a solution that we want [@Kanti:1995vq]. To obtain the real value, there exists a restriction for $\phi_h$ from the square root term of $\phi'_h$, however, it is highly model dependent. Let us assume that $\phi_h$ is an arbitrary constant but it satisfies $r^4_h \geq 192 \dot{f}^2_h$. The substitution of $\phi'_h$ into Eqs. at the horizon reduces the equations as follows: $$\phi''(r) \approx 0, \qquad \text{and} \qquad X''(r) \approx -X'^2.$$ Then, the derivative of metric function $X'$ is obtained, $$X'(r) = \dfrac{1}{\delta r} + {\cal O} \left(1 \right),$$ which recovers that which was originally assumed near the horizon limit, Eq. . As a result, the GB term at the horizon is also obtained by $$\label{eq:gbh} R_{\rm GB}^2 \approx \dfrac{4 e^{-2Y}}{r^2}X'^2.$$ We also can expand the metric components and dilaton field in the asymptotically flat region in terms of the Arnowitt-Deser-Misner (ADM) mass $M$ [@Arnowitt:1959ah; @Arnowitt:1962hi] and dilaton charge $D$ as follows: $$\begin{aligned} e^{X(r)} &=& 1 - \dfrac{2M}{r} + {\cal O}(r^{-2}), \\ e^{Y(r)} &=& 1 + \dfrac{2M}{r} + {\cal O}(r^{-2}), \\ \phi(r) &=& \phi_\infty + \dfrac{D}{r} + {\cal O}(r^{-2}).\end{aligned}$$ The GB term in the asymptotically flat region is then, $$\label{eq:gba} R_{\rm GB}^2 = \dfrac{48M^2}{r^6} + {\cal O}(r^{-7}).$$ In order to focus on the numerical analysis, we will not show the relation between the parameters in the near horizon limit and asymptotically flat region in more detail (see the Refs. [@Antoniou:2017acq; @Antoniou:2017hxj]). The ADM mass is represented as follows: $$M = M(r_h) + M_{\rm hair},$$ where the first term is the mass inside the horizon, $M(r_h) = r_h / 2$, and the second term is the mass of the dilaton hair. Once the metric is obtained numerically by the shooting method from the horizon, it is possible to obtain the mass and charge of the black hole by matching the behavior of the metric in the asymptotically flat region. Since the dilaton black hole mass has the contribution coming from the existence of a scalar hair, it seems to evade the no-hair theorem. For this reason, we analyze whether or not there is a contradiction in the equations of motion with the energy-momentum tensor as in [@Antoniou:2017acq]. This is an important procedure to both compare and evade the novel no-hair theorem in [@Bekenstein:1995un]. The $(tt)$ and $(rr)$ components of the energy-momentum tensor are given by $$\begin{aligned} \nonumber T_t^{\ t} &=& - \dfrac{e^{-Y}}{2r^2} \bigg( r^2 \phi'^2 - 8 \dot{f} \phi' Y' (1 - 3 e^{-Y}) \\ && + 16 (\dot{f} \phi'' + \ddot{f} \phi'^2) (1 - e^{-Y}) \bigg), \\ T_r^{\ r} &=& \dfrac{e^{-Y}}{2r^2} \left( r^2 \phi'^2 - 8\dot{f} \phi' X' (1 - 3e^{-Y}) \right), \\ \nonumber (T_r^{\ r})' &=& \dfrac{e^{-Y}}{2r^2} \bigg( 2 r^2 \phi' \phi'' - r^2 \phi'^2 Y' - 8 \dot{f} \phi' X'' (1 - 3e^{-Y}) \\ \nonumber && - 8 (\dot{f} \phi'' + \ddot{f} \phi'^2 ) X' (1 - 3e^{-Y}) \\ \nonumber && + 8 \dot{f} \phi' X' Y' (1 - 6e^{-Y}) + \dfrac{16}{r} \dot{f} \phi' X' (1 - 3e^{-Y}) \bigg). \\\end{aligned}$$ In the near horizon limit, $T_t^{\ t}$, $T_r^{\ r}$ and $(T_r^{\ r})'$ are reduced to $$\begin{aligned} T_t^{\ t} &=& \dfrac{4 r^3 \dot{f} \phi' + 96 \dot{f}^2}{r^2(r^4 + 4 r^3 \dot{f} \phi' - 96 \dot{f}^2)} + {\cal O}(\delta r), \\ T_r^{\ r} &=& -\dfrac{4 \dot{f}\phi'}{r^2 (r + 4 \dot{f}\phi')} + {\cal O}(\delta r), \\ (T_r^{\ r})' &=& 0 \times X' + {\cal O}(1) + {\cal O}(\delta r).\end{aligned}$$ Since $\dot{f}\phi'$ is negative definite from Eq. at the horizon, $T_r^{\ r}$ is positive definite in the near horizon limit. However, for $(T_r^{\ r})'$, the first order, which depends on $X'$, is identically zero, and so we should consider the next order. But, it is very hard to find and difficult to express, so we calculate the sign by putting whole horizon values into the second order, and we obtain the negative value. The sign is same as shown in [@Antoniou:2017acq], even though we were not able to reproduce the results that they obtained. Similar to the case in the near horizon limit, those are given in the asymptotically flat region $$\begin{gathered} - T_t^{\ t} = T_r^{\ r} = \dfrac{1}{2} \phi'^2 + {\cal O}(r^{-5}), \quad {\rm and} \\ (T_r^{\ r})' = \phi' \phi'' + {\cal O}(r^{-6}) = -\dfrac{2}{r} \phi'^2 + {\cal O}(r^{-6}),\end{gathered}$$ where we used the asymptotic relation $\phi'' = -(2/r) \phi' + {\cal O}(r^{-4})$ from Eq. . As a result, the tendency of $T_r^{\ r}$ and $(T_r^{\ r})'$ are summarized in Table \[tab:nohair\]. We also plot the numerical results in Sec. \[sec:3.1\] which correspond with our description. Thus, there is no contradiction in the equations of motion with the energy-momentum tensor, and we argue that the dilaton black hole evades the novel no-hair theorem even for the negative $\alpha$. --------------------------- --------------------- ---------------------------- \[-0.5em\] Near horizon region Asymptotically flat region \[0.5em\] \[-0.5em\] $T_t^{\ t}$ $> 0$ $< 0$ \[0.7em\] \[-0.5em\] $T_r^{\ r}$ $> 0$ $> 0$ \[0.7em\] \[-0.5em\] $(T_r^{\ r})'$ $< 0$ $< 0$ \[0.7em\] --------------------------- --------------------- ---------------------------- : Behavior summary of $T_r^{\ r}$ and $(T_r^{\ r})'$[]{data-label="tab:nohair"} To be sure of our result, we also checked the old no-hair theorem, as shown in [@Antoniou:2017acq; @Antoniou:2017hxj], in which they developed the integral constraint equation with the positive definite coupling function $f(\phi)$ to show the evasion of the no-hair theorem. We construct the new integral constraint allowing the existence of the hairy solutions with arbitrary coupling functions. Starting with Eq. , it is possible to obtain the integral constraint, $$\begin{gathered} \int f(\phi) \left( \nabla^2 \phi + \dot{f}(\phi) R_{\rm GB}^2 \right) \\ = - \int \dot{f}(\phi) \left( \phi'^2 - f(\phi) R_{\rm GB}^2 \right) = 0,\end{gathered}$$ where they used the integration in part only for the first term. The boundary term vanishes at the horizon and infinity due to the exponential factor of the metric and the derivative of the dilaton field, respectively [@Antoniou:2017acq]. Simply $\phi'^2$ is positive definite. The GB term is positive definite both on the horizon and in the asymptotically flat region. Thus, one can guess that the GB term is positive definite for all regions and is monotonically decreasing with respect to the radial length. Indeed, this really happens, and we will show the result in Sec. \[sec:3.1\]. In order to avoid the no-hair theorem, the only condition for $f(\phi)$ is positive definite. In our study, we consider the negative $\alpha$, which makes the coupling function negative definite, in which the no-hair theorem seems to valid in this analysis. Therefore, we should find another way of treating the integral constraint and expand that, which covers all signs of definite cases of $f(\phi)$. As a result, we construct the integral constraint equation as follows: $$\begin{gathered} \int e^{f(\phi)} \left( \nabla^2 \phi + \dot{f}(\phi) R_{\rm GB}^2 \right) \\ = - \int e^{f(\phi)} \dot{f}(\phi) \left( \phi'^2 - R_{\rm GB}^2 \right) = 0,\end{gathered}$$ where we also used the integration in part only for the first term. In the above equation, $\phi'^2$ and $R_{\rm GB}^2$ are positive definite and the coupling function $f(\phi)$ can be arbitrary. Thus, it is shown that the dilaton black hole solutions with arbitrary coupling functions evade the old no-hair theorem. In order to find the dilaton black hole solution, we used the Dormand-Prince method [@DORMAND198019], which is one of the Runge-Kutta methods with specific parameters. Since the metric function diverges at the horizon, we start our calculation at $\delta r = \epsilon = 10^{-8}$, and we also set the infinity as $r_{\rm max} = 10^5$. Let us define the subscript $h$ and $\infty$ by means of the value at the initial point $r_h + \epsilon$ and the final point $r_{\rm max}$, respectively. Then, the initial conditions of the metric functions and field with the given coupling function $f(\phi) = \alpha e^{\gamma \phi}$ are $$\begin{gathered} \label{eq:horizon} X_h = \log(x_1 \epsilon), \quad X'_h = \dfrac{1}{\epsilon}, \quad \phi_h \leq \dfrac{1}{2\gamma} \log\left(\dfrac{r_h^4}{192\alpha^2 \gamma^2} \right), \\ \quad {\rm and} \quad \phi'_h = - \dfrac{r_h^2 - \sqrt{r_h^4 - 192 \dot{f}_h^2}}{8r_h \dot{f}_h}.\end{gathered}$$ One can see that the equations of motion are invariant under the shift of a dilaton field $\phi \rightarrow \phi + \phi_0$ for a constant $\phi_0$ with the rescaling of $r \rightarrow r e^{\gamma \phi_0 / 2}$ [@Kanti:1995vq]. Thus, we fix $r_h = 1$, vary $\phi_h$ to get a different black hole solution as a free parameter, and rescale the result. Until now, the parameter $x_1$ is arbitrary. On the other side, the boundary conditions at infinity are given by $$X_\infty = 0, \quad X'_\infty = 0, \quad \phi = 0, \quad {\rm and} \quad \phi'_\infty = 0.$$ To make $X_\infty = 0$, $x_1$ should be chosen properly, because the equations of motion depend only on $X'$ so the non-zero remaining constant $X_\infty$ can exist. In our calculation, we obtain the value by setting $x_1 = 1$, and we do the same procedure again with $\log x_1 = - X_\infty$, which is a way of fixing the parameter $x_1$. We also want to make the dilaton field vanish in the asymptotically flat region. For the dilaton field, the value $\phi_\infty$ also exists. The value can be absorbed by using the symmetry between $r$ and $\phi$ such as the rescaling of $r_h$ as $r_h \rightarrow r_h e^{-\gamma \phi_\infty / 2}$. Thus, the numerical calculation starts with same $r_h$ but can vary with the dilaton field value $\phi_h$, and the solutions form an one parameter family. Finally, we obtain $X$ and $\phi$ from the equations of motion and it is possible to obtain $Y$ by using Eq. . The ADM mass $M$ and dilaton charge $D$ are obtained by fitting the equation in the asymptotically flat region. Results {#sec:3} ======= In this section, we present a hairy black hole solution in DEGB theory. We set the dilaton coupling function $f(\phi) = \alpha e^{\gamma \phi}$, where the GB coefficient $\alpha$ has the negative value, not the usual positive one. Since the DEGB theory has the rescaling invariance under the $r \rightarrow r / \sqrt{\|\alpha\|}$, we choose $\alpha = -1$ for all of our data. Furthermore, one can see that the theory is invariant under the changes of $\gamma \rightarrow -\gamma$ and $\phi \rightarrow -\phi$. Thus, we always choose the positive dilaton coupling $\gamma$, which is enough to obtain the solutions. Even in this unusual negative coefficient set up, we obtained the dilaton black hole solution and the different tendency of a minimum black hole mass depending on the $\gamma$. Dilaton black hole {#sec:3.1} ------------------ This is an example of a hairy black hole solution with the negative $\alpha$. In order to get and present the dilaton black hole in this section, we set $\gamma = 1$ and $\phi_h = \log(r_h^4 / 192\alpha^2 \gamma^2) / 2\gamma$, which is the maximum value of the given range in the initial condition, Eq. . Figure \[fig:metric\] represents the metric functions and the profile of the dilaton field for a black hole solution with respect to $r$. In Fig. \[fig:metric\] (a), the black and red lines indicate the metric components $-g_{tt}(r)$ and $g_{rr}(r)$, which converge or diverge at the horizon, respectively. Both metric components converge to unity at infinity. In Fig. \[fig:metric\] (b), the dilaton field $\phi(r)$ always has a negative value and the derivative of the dilaton field has a positive value. Both quantities also become zero at infinity. The blue dashed line in each figure indicates the value of the horizon radius, which is not unity. Originally, we set the horizon radius $r_h = 1$ but it is modified by the factor $e^{-\gamma \phi_\infty / 2}$, as we explained in the previous section. Figure \[fig:emtensor\] shows the positivity of the $(rr)$ component of the energy-momentum tensor, $T_r^{\ r}(r)$, and the negative value of its derivative, $-(T_r^{\ r})'(r)$, with respect to $r$. Those quantities have positive values at the horizon and diminish when $r$ goes to infinity, but the signs never change. Therefore, there is no contradiction in the equations of motion with the energy-momentum tensor, which show that the novel no-hair theorem is not applicable and finally is not valid for the DEGB thoery, as we claimed before. Figure \[fig:rgb\] (a) illustrates the GB term $R_{\rm GB}^2 (r)$ with respect to $r$. The GB term is positive definite on all regions of $r$, and it shows the monotonically decreasing behavior when $r$ increases. The result corresponds well with our expectation about the old no-hair theorem, and the no-hair theorem is again evaded. In Fig. \[fig:rgb\] (b), the black and red lines depict the $(tt)$ and $(rr)$ components of the energy-momentum tensor, $-T_t^{\ t}(r)$ and $T_r^{\ r}(r)$, respectively. The energy density $-T_t^{\ t}(r)$ has the negative value only for the near horizon region and the positive value for all the other regions of $r$. One of key assumptions in the novel no-hair theorem is related to the energy condition. The energy density is non-negative everywhere for any timelike observer. Thus, the existence of the negative value in some region shows the violation of the key assumptions satisfied in the novel no-hair theorem. Spectrum of dilaton black holes {#sec:3.2} ------------------------------- Now, we have the dilaton black hole solutions with the negative GB coefficient $\alpha$. In order to investigate the properties of the dilaton black holes with negative $\alpha$, we obtain the dilaton black hole solutions with same boundary conditions and compare them, but we just change the sign of $\alpha$ with respect to $\gamma$. ![The lower bound for the black hole mass vs. $\alpha$ with several $\gamma$ values.[]{data-label="fig:a_vs_m_wrt_gamma"}](a_vs_m_wrt_gamma){width="47.50000%"} Figure \[fig:a\_vs\_m\_wrt\_gamma\] represents the lower bound for the dilaton black hole mass with respect to $\alpha$ for several selected values of $\gamma$. The $\phi_h$ is also chosen by the maximum value. It clearly shows that there exists the $\sqrt{\|\alpha\|}$ dependency of the black hole mass. The $\alpha$ dependency can be absorbed by the radial coordinate transformation, $r \rightarrow r / \sqrt{\|\alpha\|}$, as we discussed earlier. Therefore, we focused on the $\gamma$ dependency of the black hole mass for each sign of $\alpha$. The lower bound is increased when $\alpha$ has the positive value, but the lower bound is increased up to some specific $\gamma$ and decreased when $\alpha$ has the negative value as $\gamma$ is increased. Therefore, we expect that there exists some maximum $\gamma$ value, which restricts the dilaton black hole for the negative $\alpha$, and this is the most different behavior between the dilaton black holes with different signs of $\alpha$. Figures \[fig:minmass1\], \[fig:minmass2\], and \[fig:minmass3\] illustrate the mass of dilaton black holes with respect to $\gamma$ for each of the $\alpha$ signs. The black line denotes the black hole mass with the maximum value of $\phi_h$. The red dashed line indicates the maximum $\gamma$, which can have the dilaton black hole when $\alpha$ is negative. The blue dashed line represents the dilaton black hole, which has the minimum mass under the variations of $\phi_h$ when $\alpha$ is positive. It is already a well-known result that the maximum $\phi_h$ does not always give the minimum mass of the dilaton black hole [@Kanti:1997br; @Torii:1998gm; @Ahn:2014fwa; @Gwak:2017fea]. The black and blue lines show the dilaton black hole with maximum $\phi_h$ or minimum mass. By changing $\phi_h$, we obtain the dilaton black holes with a higher mass than the ones represented by black or blue lines. Therefore, there also exists the dilaton black holes over the black or blue lines but not under the lines. Figures \[fig:minmass1\] (a) and \[fig:minmass1\] (b) show the dilaton black hole mass increases and decreases when $\alpha$ is negative but it keep increasing even for the black holes with minimum masses when $\alpha$ is positive by increasing $\gamma$, as we have shown in Fig. \[fig:a\_vs\_m\_wrt\_gamma\]. When $\alpha$ is negative, we cannot obtain the dilaton black hole solution with the $\gamma$ value which exceeds the red dashed line. In order to get the contribution coming from the GB term in more detail, we need to investigate the large $\gamma$ region, because the GB term is exponentially dependent on the $\gamma$. In this region, it seems that the GB term decreases its own repulsive property and relatively easily assists in making the dilaton black hole. As a result, we argue that the minimum mass of the dilaton black hole decreases in the large $\gamma$ region. However, when $\alpha$ is positive, the GB term appears to demonstrate a dispersive behavior, and it disturbs the creation of the the dilaton black hole. Thus, we also argue that the minimum mass of the dilaton black hole increases depending on $\gamma$. Figures \[fig:minmass2\] (a) and \[fig:minmass2\] (b) show the hairy mass of the black hole, and Figs. \[fig:minmass3\] (a) and \[fig:minmass3\] (b) show the ratio between the hairy mass and the total mass with respect to $\gamma$. The hairy mass increases in general, except the near maximum $\gamma$ region of the case with the negative $\alpha$ and the minimum mass black hole of the case with the positive $\alpha$. However, the ratio always increases or decreases when $\alpha$ is negative or positive, respectively. Especially for the case with the negative $\alpha$, the ratio increases even higher than $> 0.5$ near the maximum $\gamma$; the red line and the behavior are really strange. We wonder whether or not the value will keep growing until it reaches unity, which means that the black hole horizon would disappear and have the mass only as the dilaton hair. This does not happen when $\alpha$ is positive. The maximum and restricted value of $\gamma$ seems to be motivated from this reason. However, we cannot do an exact numerical calculation on the limit of maximum $\gamma$ due to the difficulties of error control, thus our argument remains as a reasonable but open question. Conclusion {#sec:4} ========== We have investigated the hairy black hole solutions in DEGB theory, in particular with the negative GB coefficient $\alpha$. In Refs. [@Antoniou:2017acq; @Antoniou:2017hxj], the authors showed that the no-hair theorems are easily evaded by the hairy black hole solutions in DEGB theory. They considered the black hole solutions with only the positive $\alpha$ and many scalar couplings, and they constructed the integral constraint in the theory. In this paper, we tried to expand the description about the dilaton black hole into negative GB coefficients by changing the sign of $\alpha$. We constructed the new integral constraint equation allowing the existence of the hairy black hole solution with the arbitrary signature of $\alpha$. Through this procedure, we have expanded the evasion of the no-hair theorem for hairy black hole solutions. As a consequence of our analysis, we have numerically obtained the dilaton black hole solutions with the negative $\alpha$. The dilaton black holes have more hair, in general, than the case for the positive $\alpha$. We restricted our calculation to the dilaton black holes that have the maximum values of dilaton fields at the horizon $\phi_h$ or have minimum masses. It is enough to investigate the properties of the dilaton black hole. The minimum mass of a dilaton black hole with a positive $\alpha$ is obtained from the maximum $\phi_h$ in the small dilaton coupling $\gamma$ region. When $\gamma$ is increased, the cases of dilaton black holes with the minimum mass and the maximum $\phi_h$ are divided. The mass of a dilaton black hole increases in both cases. We think that the GB term seems to provide a repulsive effect, and it disturbs the formation the dilaton black hole. However, those two cases are not divided with the negative $\alpha$ and the minimum mass decreases for large $\gamma$. Since the minimum mass decreases, it seems that the GB term decreases its own repulsive property, and the black hole forms relatively easy. Furthermore, there exists a maximum value $\gamma$ that limits the existence of the dilaton black hole solution. Until the maximum $\gamma$, the minimum mass decreases but the hairy mass increases and decreases again. Interestingly, the hairy mass ratio of the total mass always increases. The results give us an expectation that the dilaton black hole solution with the maximum $\gamma$ would have no horizon, and the mass of the black hole is composed by the dilaton field only. Even though this expectation has yet ti be fulfilled, due to the difficulties of the numerical calculation, it is worthwhile to investigate the properties and implications of such behaviors for large $\gamma$ in more detail, and we postpone further analysis for future work. It is important to investigate the stability for the maximally symmetric background as well as the hairy black hole solutions. For the maximally symmetric background, there is nonperturbative instability due to the tunneling for a nucleation of a vacuum bubble when the initial vacuum state is in the metastable vacuum and the true vacuum state exists [@Coleman:1977py; @Coleman:1980aw; @Lee:2008hz; @Charmousis:2008ce]. For black hole solutions, there is no instability in Einstein theory when making use of the Regge-Wheeler prescription [@Regge:1957td], while the stability issue is nontrivial in DEGB theory. There exists the positive lower-bound for the black hole mass in that theory. There are two black holes for a given mass above the lower-bound in which the smaller one is unstable and the larger one is stable under perturbations [@Torii:1996yi; @Torii:1998gm]. The equations governing the perturbations of the metric are decoupled from the equation governing the perturbation of the scalar field [@Doneva:2017bvd]. Recently, it has been reported that the black hole without hair becomes unstable against scalar perturbations, and a new black hole solution with scalar hair bifurcates from the one without hair in DEGB [@Doneva:2017bvd; @Silva:2017uqg] and Einstein-Maxwell scalar theory [@Herdeiro:2018wub]. The higher-dimensional black hole in EGB theory was first discovered in Ref. [@Boulware:1985wk] and in Einstein-Maxwell-Gauss-Bonnet theory [@Wiltshire:1985us], in which the lower bound for a black hole mass is proportional to $\alpha$, not $\sqrt{\alpha}$. However, there is no bound for the mass in six and higher dimensions [@Guo:2008hf; @Sahabandu:2005ma]. It will be interesting to investigate our results in comparison with those in higher-dimensional DEGB theory or EGB theory. The stability of the maximally symmetric vacuum in DEGB theory is not complicated. Let us consider the maximally symmetric vacuum spacetime, flat Minkowski. One can consider linear perturbations around this background. Then the quadratic curvature terms in the field equations should not contribute to the perturbation equations as in the higher-dimensional EGB theory [@Boulware:1985wk]. Therefore, the perturbation equations are the same as those in general relativity with a massless scalar field, and hence, the flat background is stable against linear perturbations for any value of $\alpha$ and $\gamma$. This can be confirmed for spherically symmetric linear perturbations [@Torii:1998gm]. When the background is flat Minkowski, the perturbation equation for the dilaton perturbation reduces to the Klein-Gordon equation in a two-dimensional flat spacetime that does not contain $\alpha$ and $\gamma$ as well as any higher-curvature terms. Thus, the flat background is stable against these spherical perturbations [^1]. We postpone any possible application for the evasion of the no-hair theorem and the applications for black hole solutions including the stability issue in higher-dimensional DEGB theory for our future work. B.-H. Lee(NRF-2018R1D1A1B07048657), W. Lee (NRF-2016R1D1A1B01010234) and D. Ro (NRF-2017R1D1A1B03029430) were supported by Basic Science Research Program through the National Research Foundation of Korea funded by the Ministry of Education. D. Ro was also supported by the Korea Ministry of Education, Science and Technology, Gyeongsangbuk-Do and Pohang City. We would like to thank Hyeong-Chan Kim and Youngone Lee for their hospitality during our visit to Korea National University of Transportation, Seoktae Koh to Jeju National University, and Yun Soo Myung, Hyung Won Lee, Kyoung Yee Kim, Jeongcho Kim and De-Cheng Zou to Inje University. [^1]: We would like to thank an anonymous referee for useful comments and pointing this out.	{ "pile_set_name": "ArXiv" }
--- abstract: 'In an earlier paper (A. N. Kochubei, [Pacif. J. Math.]{} 269 (2014), 355–369), the author considered a restriction of Vladimirov’s fractional differentiation operator $D^\alpha$, $\alpha >0$, to radial functions on a non-Archimedean field. In particular, it was found to possess such a right inverse $I^\alpha$ that the appropriate change of variables reduces equations with $D^\alpha$ (for radial functions) to integral equations whose properties resemble those of classical Volterra equations. In other words, we found, in the framework of non-Archimedean pseudo-differential operators, a counterpart of ordinary differential equations. In the present paper, we begin an operator-theoretic investigation of the operator $I^\alpha$, and study a related analog of the Laplace transform.' author: - \| Anatoly N. Kochubei\ Institute of Mathematics,\ National Academy of Sciences of Ukraine,\ Tereshchenkivska 3, Kiev, 01024 Ukraine\ E-mail: [email protected] title: 'Non-Archimedean Radial Calculus: Volterra Operator and Laplace Transform' --- fractional differentiation operator; non-Archimedean local field; radial functions; Volterra operator; Laplace transform . Primary: 47G10. Secondary: 11S80, 35S10, 43A32. Introduction ============ The basic linear operator defined on real- or complex-valued functions on a non-Archimedean local field $K$ (such as $K={\mathbb Q_p}$, the field of $p$-adic numbers) is the Vladimirov pseudo-differential operator ${D^\alpha}$, $\alpha >0$, of fractional differentiation [@VVZ]; for further development of this subject see [@AKS; @KKZ; @K2001; @Z2016]. It was found in [@K2014] that properties of ${D^\alpha}$ become much simpler on radial functions. Moreover, in this case it was found to possess a right inverse ${I^\alpha}$, which can be seen as a $p$-adic counterpart of the Riemann-Liouville fractional integral or, for $\alpha =1$, the classical anti-derivative. The change of an unknown function $u={I^\alpha}v$ reduces the Cauchy problem for an equation with the radial restriction of ${D^\alpha}$ to an integral equation with properties resembling those of classical Volterra equations. In other words, we found, in the framework of non-Archimedean pseudo-differential operators, a counterpart of ordinary differential equations. In [@K2014], we studied linear equations of this kind; nonlinear ones were investigated in [@K2020]. In this paper we study the operator ${I^1}$ on the ring of integers $O\subset K$ as an object of operator theory. The operator ${I^1}$ on $L^2(O)$ happens to be a sum of a bounded selfadjoint operator and a simple Volterra operator ${I_0^1}$ with a rank two imaginary part $J$, such that $\operatorname{tr} J=0$. The characteristic matrix-function $W(z)$ of ${I_0^1}$ is such that $W(z^{-1})$ is, in contrast to classical examples, an entire matrix function of zero order. While the theory of Volterra operators and their characteristic functions is well-developed (see [@GK; @Br; @GT; @Ma; @Zo]), properties of the above operator are very different from those known for operators of classical analysis and their generalizations. Therefore, while ${I^1}$ and ${I_0^1}$ are just specific examples, they create a framework for future studies in this area. Another subject touched in this paper is a version of the Laplace transform. The classical Laplace transform is based on the function $x\mapsto e^{-\lambda x}$ satisfying an obvious differential equation. A similar equation involving ${D^\alpha}$ has a unique radial solution [@K2008]. This leads to a definition of the Laplace type transform in the above framework. We prove a uniqueness theorem and the inversion formula for this transform. Preliminaries ============= [2.1. Local fields.]{} Let $K$ be a non-Archimedean local field, that is a non-discrete totally disconnected locally compact topological field. It is well known that $K$ is isomorphic either to a finite extension of the field $\mathbb Q_p$ of $p$-adic numbers (if $K$ has characteristic 0), or to the field of formal Laurent series with coefficients from a finite field, if $K$ has a positive characteristic. For a summary of main notions and results regarding local fields see, for example, [@K2001]. Any local field $K$ is endowed with an absolute value $\|\cdot \|_K$, such that $\|x\|_K=0$ if and only if $x=0$, $\|xy\|_K=\|x\|_K\cdot \|y\|_K$, $\|x+y\|_K\le \max (\|x\|_K,\|y\|_K)$. Denote $O=\{ x\in K:\ \|x\|_K\le 1\}$, $P=\{ x\in K:\ \|x\|_K<1\}$. $O$ is a subring of $K$, and $P$ is an ideal in $O$ containing such an element $\beta$ that $P=\beta O$. The quotient ring $O/P$ is actually a finite field; denote by $q$ its cardinality. We will always assume that the absolute value is normalized, that is $\|\beta \|_K=q^{-1}$. The normalized absolute value takes the values $q^N$, $N\in \mathbb Z$. Note that for $K=\mathbb Q_p$ we have $\beta =p$ and $q=p$; the $p$-adic absolute value is normalized. The additive group of any local field is self-dual, that is if $\chi$ is a fixed non-constant complex-valued additive character of $K$, then any other additive character can be written as $\chi_a(x)=\chi (ax)$, $x\in K$, for some $a\in K$. Below we assume that $\chi$ is a rank zero character, that is $\chi (x)\equiv 1$ for $x\in O$, while there exists such an element $x_0\in K$ that $\|x_0\|_K=q$ and $\chi (x_0)\ne 1$. The above duality is used in the definition of the Fourier transform over $K$. Denoting by $dx$ the Haar measure on the additive group of $K$ (normalized in such a way that the measure of $O$ equals 1) we write $$\widetilde{f}(\xi )=\int\limits_K\chi (x\xi )f(x)\,dx,\quad \xi \in K,$$ where $f$ is a complex-valued function from $L_1(K)$. As usual, the Fourier transform $\mathcal F$ can be extended from $L_1(K)\cap L_2(K)$ to a unitary operator on $L_2(K)$. If $\mathcal F f=\widetilde{f}\in L_1(K)$, we have the inversion formula $$f(x)=\int\limits_K\chi (-x\xi )\widetilde{f}(\xi )\,d\xi .$$ Working with functions on $K$ and operators upon them we often use standard integration formulas; see [@K2001; @VVZ]. The simplest of them are as follows: $$\int\limits_{\|x\|_K\le q^n}dx=q^n;\quad \int\limits_{\|x\|_K=q^n}dx=\left( 1-\frac1q \right)q^n.$$ $$\int\limits_{\|x\|_K\le q^n}\|x\|_K^{\alpha -1}\,dx=\frac{1-q^{-1}}{1-q^{-\alpha }}q^{\alpha n};\quad \text{here and above $n\in \mathbb Z,\alpha >0$}.$$ A function $f:\ K\to \mathbb C$ is said to be locally constant, if there exists such an integer $l$ that for any $x\in K$ $$f(x+x')=f(x), \quad \text{whenever $\|x'\|\le q^{-l}$}.$$ The vector space $\mathcal D(K)$ of all locally constant functions with compact supports is used as a space of test functions in analysis on $K$. Note that the Fourier transform preserves $\mathcal D(K)$. There exists a well-developed theory of distributions on local fields; see [@AKS; @K2001; @VVZ]. . On a test function $\varphi \in \mathcal D(K)$, the fractional differentiation operator $D^\alpha$, $\alpha >0$, is defined as $$\label{2.1} \left( D^\alpha \varphi \right) (x)=\mathcal F^{-1}\left[ \|\xi \|_K^\alpha (\mathcal F (\varphi ))(\xi )\right] (x).$$ Note that ${D^\alpha}$ does not preserve $\mathcal D(K)$; see [@AKS] regarding the spaces of test functions and distributions preserved by this operator. The operator $D^\alpha$ can also be represented as a hypersingular integral operator: $$\label{2.2} \left( D^\alpha \varphi \right) (x)=\frac{1-q^\alpha }{1-q^{-\alpha -1}}\int\limits_K \|y\|_K^{-\alpha -1}[\varphi (x-y)-\varphi (x)]\,dy$$ [@K2001; @VVZ]. In contrast to (\[2.1\]), the expression in the right of (\[2.2\]) makes sense for wider classes of functions. In particular, ${D^\alpha}$ is defined on constant functions and annihilates them. Denote for brevity $\theta_\alpha =\dfrac{1-q^\alpha }{1-q^{-\alpha -1}}$. Below we consider the operator ${D^\alpha}$ on a radial function $u=u(\|x\|_K)$; here we identify the function $x\mapsto u(\|x\|_K)$ on $K$ with the function $\|x\|_K\mapsto u(\|x\|_K)$ on $q^{\mathbb Z}$. This abuse of notation does not lead to confusion. The explicit expression of ${D^\alpha}u$ for a radial function $u$ satisfying some growth restrictions near the origin and infinity was found in [@K2014]. If $u=u(\|x\|_K)$ is such that $$\label{2.3} \sum\limits_{k=-\infty}^m q^k\left\| u(q^k)\right\| <\infty ,\quad \sum\limits_{l=m}^\infty q^{-\alpha l}\left\| u(q^l)\right\| <\infty,$$ for some $m\in \mathbb Z$, then for each $n\in \mathbb Z$ the expression in the right-hand side of (\[2.2\]) with $\varphi (x)=u(\|x\|_K)$ exists for $\|x\|_K=q^n$, depends only on $\|x\|_K$, and $$\begin{gathered} \label{2.4} (D^\alpha u)(q^n)=\theta_\alpha \left(1-\frac1q \right)q^{-(\alpha +1)n}\sum\limits_{k=-\infty}^{n-1} q^ku(q^k) +q^{-\alpha n-1}\frac{q^\alpha +q-2}{1-q^{-\alpha -1}}u(q^n)\\ +\theta_\alpha \left(1-\frac1q \right)\sum\limits_{l=n+1}^\infty q^{-\alpha l}u(q^l).\end{gathered}$$ Under the conditions (\[2.3\]), the expression (\[2.4\]) agrees also with the definition of ${D^\alpha}$ in terms of Bruhat-Schwartz distributions (see Chapter 2 of [@VVZ]). . The fractional integral mentioned in Introduction, was defined in [@K2014] initially for $\varphi \in \mathcal D(K)$ as follows: $$(I^\alpha \varphi )(x)=(D^{-\alpha}\varphi )(x)-(D^{-\alpha}\varphi )(0) \tag{$$}$$ where $D^{-\alpha}$ is the right inverse of ${D^\alpha}$ introduced by Vladimirov [@VVZ]: $$\left( D^{-\alpha} \varphi \right) (x)=(f_\alpha \varphi )(x)=\frac{1-q^{-\alpha} }{1-q^{\alpha -1}}\int\limits_K \|x-y\|_K^{\alpha -1}\varphi (y)\,dy,\quad \alpha \ne 1,$$ $$\left( D^{-1}\varphi \right) (x)=\frac{1-q}{q\log q}\int\limits_K \log \|x-y\|_K\varphi (y)\,dy.$$ $D^{-1}$ is a right inverse to $D^1$ only on such functions $\varphi$ that $$\int\limits_K\varphi (x)\,dx=0.$$ On such a function $\varphi$ we have also $D^{-1}D^1\varphi =\varphi$. The above definition $()$ leads to explicit expressions $$(I^\alpha \varphi )(x)=\frac{1-q^{-\alpha}}{1-q^{\alpha -1}}\int\limits_{\|y\|_K\le \|x\|_K}\left( \|x-y\|_K^{\alpha -1}-\|y\|_K^{\alpha -1}\right) \varphi (y)\,dy,\quad \alpha \ne 1,$$ and $$(I^1\varphi )(x)=\frac{1-q}{q\log q}\int\limits_{\|y\|_K\le \|x\|_K}\left( \log \|x-y\|_K-\log \|y\|_K\right) \varphi (y)\,dy.$$ Note that the integrals are taken, for each fixed $x\in K$, over bounded sets, and $(I^\alpha \varphi )(0)=0$. These properties are different from those of the anti-derivatives $D^{-\alpha }$ studied in [@VVZ]. Let $u=u(\|x\|_K)$ be a radial function, such that $$\sum\limits_{k=-\infty}^m \max \left( q^k,q^{\alpha k}\right) \left\| u(q^k)\right\| <\infty ,\quad \text{if $\alpha \ne 1$},$$ and $$\sum\limits_{k=-\infty}^m \|k\|q^k \left\| u(q^k)\right\| <\infty ,\quad \text{if $\alpha =1$},$$ for some $m\in \mathbb Z$. Then [@K2014] $I^\alpha u$ exists, it is a radial function, and for any $x\ne 0$, $$(I^\alpha u)(\|x\|_K)=q^{-\alpha}\|x\|_K^\alpha u(\|x\|_K)+\frac{1-q^{-\alpha}}{1-q^{\alpha -1}}\int\limits_{\|y\|_K< \|x\|_K}\left( \|x\|_K^{\alpha -1}-\|y\|_K^{\alpha -1}\right) u(\|y\|_K)\,dy,\quad \alpha \ne 1,$$ and $$\label{2.5} (I^1 u)(\|x\|_K)=q^{-1}\|x\|_K u(\|x\|_K)+\frac{1-q}{q\log q}\int\limits_{\|y\|_K<\|x\|_K}\left( \log \|x\|_K-\log \|y\|_K\right) u(\|y\|_K)\,dy.$$ On an appropriate class of radial functions, ${I^\alpha}$ is a right inverse to ${D^\alpha}$ [@K2014].An important difference between $D^{-\alpha }$ and ${I^\alpha}$ is the bounded integration domain in the integral formulas for ${I^\alpha}$. The operator ${D^\alpha}$ defined initially on $\mathcal D(K)$ is, after its closure in $L^2(K)$, a selfadjoint operator with a pure point spectrum $\{ q^{\alpha N},N\in \mathbb Z\}$ of infinite multiplicity and a single limit point zero. It was shown in [@K2008] that for each $N\in \mathbb Z$, there exists a unique (up to the multiplication by a constant) radial eigenfunction $$\label{2.6} v_N(\|x\|_K)=\begin{cases} 1, & \text{if $\|x\|_K\le q^{-N}$,}\\ -\frac1{q-1}, & \text{if $\|x\|_K=q^{-N+1}$,}\\ 0, & \text{if $\|x\|_K\ge q^{-N+2}$,}\end{cases}$$ corresponding to the eigenvalue $\lambda=q^{\alpha N}$. Below we interpret this function as an analog of the classical exponential function $x\mapsto e^{-\lambda x}$. Note that $v_N\in \mathcal D(K)$; this is a purely non-Archimedean phenomenon reflecting the unusual topological property of $K$, its total disconnectedness. The operator ${D_O^\alpha}$ in the space $L^2(O)$ on the ring of integers (unit ball) $O$ is defined as follows. Extend a function $\varphi \in \mathcal D(O)$ (that is a function $\varphi \in \mathcal D(K)$ supported in $O$) onto $K$ by zero. Apply ${D^\alpha}$ and consider the resulting function on $O$. After the closure in $L^2(O)$ we obtain a selfadjoint operator ${D_O^\alpha}$ with a discrete spectrum [@K2001; @VVZ] (here we do not touch different definitions from [@BGPW] and [@K2018]). Denote by ${\mathcal H}$ the subspace in $L^2(O)$ consisting of radial functions. The functions $v_N$, $N=1,2,\ldots$ belong to ${\mathcal H}$, as well as the function $$v_0(\|x\|_K)\equiv 1,\quad \|x\|_K\le 1.$$ By the definition of ${D_O^\alpha}$, the functions $v_N$ are its eigenfunctions corresponding to the eigenvalues $q^{\alpha N}$. As for $v_0$, it is also an eigenfunction, with the eigenvalue $\mu_0=\dfrac{q-1}{q^{\alpha +1}-1}q^\alpha$ [@K2001; @VVZ]. Therefore $\{ v_N\}_{N\ge 0}$ is an orthonormal system in $L^2(O)$, hence in ${\mathcal H}$. We have $\\| v_0\\| =1$ ($\\| \cdot \\|$ is the norm in ${\mathcal H}$), $$\\|v_N\\|^2=\int\limits_{\|x\|_K\le q^{-N}}dx+(q-1)^{-2} \int\limits_{\|x\|_K=q^{-N+1}}dx=q^{-N}+(q-1)^{-2}q^{-N+1}(1-\frac1q)=(q-1)^{-1}q^{-N},$$ $$\int\limits_{\|x\|_K\le 1}v_N(\|x\|_K)\,dx=0,\quad N\ge 1.$$ Therefore the functions $$\label{2.7} e_0( \|x\|_K)\equiv 1;\quad e_N( \|x\|_K)=(q-1)^{1/2}q^{N/2}v_N(\|x\|_K),\quad N\ge 1,$$ form an orthonormal system in ${\mathcal H}$. The sytem $\{ e_N\}_{N\ge 0}$ is an orthonormal basis in ${\mathcal H}$. . Let $u\in {\mathcal H}$ be orthogonal to all the functions $e_N$. Then $$\int\limits_{\|x\|_K\le 1}u(\|x\|_K)\,dx=0,$$ so that $$\label{2.8} \sum\limits_{j=-\infty}^0 u(q^j)q^j=0$$ and $$\int\limits_{\|x\|_K\le q^{-N}}u(\|x\|_K)\,dx -(q-1)^{-1}\int\limits_{\|x\|_K=q^{-N+1}}u(\|x\|_K)\,dx=0,$$ so that $$\label{2.9} \sum\limits_{j=-\infty}^{-N} u(q^j)q^j-(q-1)^{-1}u(q^{-N+1})=0,\quad N=1,2,\ldots .$$ Subtracting from (\[2.8\]) the equality (\[2.9\]) with $N=1$, we find that $u(1)=0$. Now the equality (\[2.9\]) with $N=1$ takes the form $$\sum\limits_{j=-\infty}^{-1} u(q^j)q^j=0,$$ while (\[2.9\]) with $N=2$ yields $$\sum\limits_{j=-\infty}^{-2} u(q^j)q^j-(q-1)^{-1}u(q^{-1})=0.$$ Subtracting we obtain that $u(q^{-1})=0$. Repeating the above reasoning we find that $u=0$.$\blacksquare$ Another (obvious) orthonormal basis in ${\mathcal H}$ is $$\label{2.10} f_n( \|x\|_K)=\begin{cases} (1-\frac1q)^{-1/2}q^{n/2}, & \text{if $\|x\|_K=q^{-n}$;}\\ 0, & \text{elsewhere,}\end{cases} \quad n=0,1,2,\ldots .$$ The next result is of some independent interest. The set of “polynomials” $$\label{2.11} u(\|x\|_K)=\sum\limits_{n=1}^N a_n\|x\|_K^n,\quad a_n\in \mathbb C,\ N\ge 1,$$ is dense in ${\mathcal H}$. . Suppose that a function $F\in {\mathcal H}$ is orthogonal to all the functions $X_l(\|x\|_K)=\|x\|_K^l$, $l\ge 1$. Using the basis (\[2.10\]), write $$F=\sum\limits_{n=0}^\infty c_nf_n,\quad \{c_n\} \in l^2.$$ We have $$\langle X_l,f_n\rangle =(1-\frac1q)^{-1/2}q^{n/2}\int\limits_{\|x\|_K=q^{-n}}\|x\|_K^l\,dx=(1-\frac1q)^{1/2}q^{-n/2-nl},$$ so that $$\langle F,X_l\rangle = (1-\frac1q)^{1/2} \sum\limits_{n=0}^\infty c_nq^{-n/2-nl}=0,\quad l=1,2,\ldots .$$ Denoting $\beta =q^{-1}$, $b_n=c_nq^{-n/2}$, we see that the vector $(b_0,b_1,b_2,\ldots )\in l^2$ is orthogonal in $l^2$ to each vector $(1,\beta^l,\beta^{2l},\ldots )$, $l\ge 1$. It is known ([@Ha], Problem 6) that the set of all these vectors is total in $l^2$, so that $F=0$. $\blacksquare$ In fact, the above reasoning proves the density of polynomials (\[2.11\]) in a wider weighted space determined by the condition $\{ c_nq^{-n/2}\} \in l^2$. Integration Operators ===================== [3.1. The operator ${I^1}$.]{} Let us study ${I^1}$ as an operator in ${\mathcal H}$, find its matrix representation with respect to the basis $\{e_N\}$ and investigate the spectrum of ${I^1}$. The operator ${I^1}$ has the matrix representation $${I^1}=\begin{pmatrix} 0 & -(q-1)^{1/2}q^{-1/2} & -(q-1)^{1/2}q^{-1} & \ldots & -(q-1)^{1/2}q^{-n/2} & \ldots \\ 0 & q^{-1} & 0 & \ldots & 0 & \ldots \\ 0 & 0 & q^{-2} & \ldots & 0 & \ldots \\ \hdotsfor{6}\\ 0 & 0 & 0 & \ldots & q^{-n} & \ldots \\ \hdotsfor{6} \end{pmatrix}$$ (only the first row and the principal diagonal have nonzero elements). ${I^1}$ is a Hilbert-Schmidt operator. Apart from being a point of essential spectrum, $\lambda =0$ is a simple eigenvalue. In addition, $I^1$ has simple eigenvalues $\lambda_m=q^{-m}$, $m=1,2,\ldots$. . Since the integral of each function $e_N$, $N\ge 1$, equals zero, we have $D^{-1}D^1 e_N=e_N$. On the other hand, $D^1e_N=q^Ne_N$, so that $D^{-1}e_N=q^{-N}e_N$, and by the definition $()$ of ${I^1}$, $$\label{3.1} {I^1}e_N=q^{-N}e_N-(q-1)^{1/2}q^{-N/2}e_0,\quad N=1,2,\ldots .$$ Next, $({I^1}e_0)(\|x\|_K)$, $\|x\|_K\le 1$, depends only on the values of $e_0$ for $\|x\|_K\le 1$. Let $f(x)\equiv 1$, $x\in K$. Then ${I^1}f=0$ [@K2014], so that $$\label{3.2} {I^1}e_0=0\quad \text{in ${\mathcal H}$}.$$ The equalities (\[3.1\]) and (\[3.2\]) imply the required matrix representation, which implies the Hilbert-Schmidt property. Let us find the eigenvalues of $I^1$. As we have seen, ${I^1}e_0=0$. Suppose that $$u=\sum\limits_{n=0}^\infty c_ne_n,\quad \{c_n\}\in l^2,\quad {I^1}u=\lambda u.$$ By (\[3.1\]) and (\[3.2\]), $${I^1}u=\sum\limits_{n=1}^\infty q^{-n}c_ne_n-\left[ \sum\limits_{n=1}^\infty (q-1)^{1/2}q^{-n/2}c_n\right] e_0,$$ and we find that $$\label{3.3} \left\{ \begin{aligned} \lambda c_0 & =-(q-1)^{1/2}\sum\limits_{n=1}^\infty q^{-n/2}c_n;\\ \lambda c_n & =q^{-n}c_n,\quad n\ge 1. \end{aligned} \right.$$ A nonzero value of $c_n (n\ge 1)$ is possible only for a single index $n=m$, and in this case $\lambda =q^{-m}$. Then the first equation in (\[3.3\]) gives $c_m=-(q-1)^{-1/2}q^{-m/2}c_0$, so that $$u=c_0e_0-(q-1)^{-1/2}q^{-m/2}c_0e_m$$ is the unique (up to the multiplication by a constant) eigenfunction. $\blacksquare$ The definition $()$ of the operator ${I^\alpha}$ involves operators in $L^2(K)$; then we make restrictions to $L^2(O)$ and ${\mathcal H}$. In this section we show, for the case where $\alpha =1$, that a similar representation containing only operators in $L^2(O)$ is also possible. If $u\in L^2(O)$, then $$\label{3.4} ({I^1}u)(x)=\left( \left( {D_O^1}\right)^{-1}u\right) (x)-\left( \left( {D_O^1}\right)^{-1}u\right) (0).$$ . In [@K2018], we found the resolvent $\left( {D_O^1}-\mu_0 +\mu\right)^{-1}$ where $\mu_0=\dfrac{q}{q+1}$ (the first eigenvalue of ${D_O^1}$), $\mu >0$. In [@K2018], in connection with nonlinear equations, we considered operators in $L^1(O)$, but the result is valid for $L^2(O)$ too. For $\mu =\mu_0$, $$\label{3.5} \left( {D_O^1}\right)^{-1}u (x)=\int\limits_{\|\xi \|_K\le 1}\mathcal K (x-\xi )u(\xi )\,d\xi+\mu_0^{-1}\int\limits_{\|\xi \|_K\le 1}u(\xi )\,d\xi ,$$ where for $\|x\|_K=q^m$, $m\le 0$, $$\mathcal K(x)=\int\limits_{q\le \|\eta \|_K\le q^{-m+1}}\|\eta \|_K^{-1}\chi (\eta x)\,d\eta .$$ Using the well-known integration formula (see, for example, Section 1.5 in [@K2001]), we get $$\begin{gathered} \mathcal K(x)=\sum\limits_{j=1}^{-m+1}q^{-j}\int\limits_{\|\eta \|_K=q^j}\chi (\eta x)\,d\eta =(1-\frac1q)\sum\limits_{j=1}^{-m}1-q^{-1}\\ =-(1-\frac1q)m-q^{-1}=\frac{1-q}{q\log q}\log \|x\|_K-q^{-1}.\end{gathered}$$ By (\[3.5\]), $$\left( {D_O^1}\right)^{-1}u (x)=\frac{1-q}{q\log q}\int\limits_{\|\xi \|_K\le 1}\log \|x-\xi \|_Ku(\xi )\,d\xi +\int\limits_{\|\xi \|_K\le 1}u(\xi )\,d\xi .$$ Comparing with the expression for ${I^1}$ and noticing that $\|x-\xi \|_K-\|\xi\|_K=0$, if $\|\xi \|_K>\|x\|_K$, we obtain (\[3.4\]).$\blacksquare$ Let us consider the integral part of (\[2.5\]), the operator $$\left( {I_0^1}u\right) (x)=\frac{1-q}{q\log q}\int\limits_{\|y\|_K<\|x\|_K}\left( \log \|x\|_K-\log \|y\|_K\right) u(\|y\|_K)\,dy.$$ Recall [@GK] that a compact operator is called a Volterra operator, if its spectrum consists of the unique point $\lambda =0$. An operator $A$ is called simple, if $A$ and $A^$ have no common nontrivial invariant subspace, on which these operators coincide. It is known [@GK] that a Volterra operator $A$ is simple, if and only if the equations $Af=0$ and $A^f=0$ have no common nontrivial solutions. The main technical tool in the study of ${I_0^1}$ is the identity [@K2014] $$\label{3.6} \int\limits_{\|y\|_K<\|x\|_K}\left( \log \|x\|_K-\log \|y\|_K\right)\|y\|_K^m\,dy=d_m\|x\|_K^{m+1},\quad m=0,1,2,\ldots ,$$ where $0<d_m\le Aq^{-m}$, $A>0$ does not depend on $m$. The operator ${I_0^1}$ in ${\mathcal H}$ is a simple Volterra operator with a rank 2 imaginary part $J=\frac1{2i}(A-A^)$, such that $\operatorname{tr}J=0$. . 1) Suppose that ${I_0^1}u=\lambda u,$ $u\in {\mathcal H}$, $\lambda \in \mathbb C$, $\lambda \ne 0$. Then for $\|x\|_K\le 1$, $$\begin{gathered} \|u(\|x\|_K)\|\le \frac{c}{\|\lambda\|}\\|u\\|_{L^2(O)}\left[ \int\limits_{\|y\|_K<\|x\|_K}\left( \log \|x\|_K-\log \|y\|_K\right)^2\,dy\right]^{1/2}\\ \le \frac{c}{\|\lambda\|}\\|u\\|_{L^2(O)}\left[ q^{-1}\|x\|_K(\log \|x\|_K)^2\right]^{1/2}\le H\end{gathered}$$ where $c=\dfrac{q-1}{q\log q}$, $H$ is a positive constant. This implies the estimate $$\|u(\|x\|_K)\|\le \frac{cH}{\|\lambda\|}\int\limits_{\|y\|_K<\|x\|_K}\left( \log \|x\|_K-\log \|y\|_K\right)\,dy,$$ and by the identity (\[3.6\]) with $m=0$, $$\|u(\|x\|_K)\|\le \frac{cHA}{\|\lambda\|}\|x\|_K.$$ Similarly, the identity (\[3.6\]) with $m=1$ gives $$\|u(\|x\|_K)\|\le \frac{c^2HA^2}{\|\lambda\|^2}q^{-1}\|x\|_K^2,$$ and we find by induction that $$\label{3.7} \|u(\|x\|_K)\|\le \frac{c^{m+1}HA^{m+1}}{\|\lambda\|^{m+1}}q^{-1}q^{-2}\cdots q^{-m+1}\|x\|_K^{m+1},$$ for an arbitrary natural number $m$. Note that $$q^{-1}q^{-2}\cdots q^{-m+1}=\left( \frac1q\right)^{m(m-1)/2}.$$ Together with (\[3.7\]), this shows that $u\equiv 0$. 2\) It follows from the definition of ${I_0^1}$ that $\lambda =0$ is an eigenvalue corresponding to the eigenfunction $$\label{3.8} u_0(\|x\|_K)=\begin{cases} 1, & \text{if $\|x\|_K=1$};\\ 0, & \text{if $\|x\|_K<1$}.\end{cases}$$ Let us show that $\lambda =0$ does not correspond to other eigenfunctions. Suppose that ${I_0^1}\varphi =0$ for some $\varphi \in {\mathcal H}$, so that $$\label{3.9} \sum\limits_{j=-\infty}^{n-1}(n-j)q^j\varphi (q^j)=0,\quad n=0,-1,-2,\ldots .$$ Together with (\[3.9\]), consider a similar equality with $n-1$ substituted for $n$, that is Suppose that ${I_0^1}\varphi =0$ for some $\varphi \in {\mathcal H}$, so that $$\label{3.10} \sum\limits_{j=-\infty}^{n-2}(n-1-j)q^j\varphi (q^j)=0,\quad n=0,-1,-2,\ldots .$$ Subtracting (\[3.9\]) from (\[3.10\]) we find that $$\sum\limits_{j=-\infty}^{n-1}q^j\varphi (q^j)=0,\quad n=0,-1,-2,\ldots ,$$ that is, in particular, $$q^{n-1}\varphi (q^{n-1})+q^{n-2}\varphi (q^{n-2})+\cdots =0,$$ $$q^{n-2}\varphi (q^{n-2})+q^{n-3}\varphi (q^{n-3})+\cdots =0,$$ Subtracting the second equality from the first one, we find that $\varphi (q^{-1})=\varphi (q^{-2})=\ldots =0$, so that $\varphi$ is proportional to the eigenfunction (\[3.8\]). 3\) The imaginary part $J$ has the following matrix representation with respect to the basis $\{ e_N\}$: $$\label{3.11} J=\frac{(q-1)^{1/2}}{2i}\begin{pmatrix} 0 & -q^{-1/2} & -q^{-1} & \ldots & -q^{-N/2} & \ldots \\ q^{-1/2} & 0 & 0 & \ldots & 0 & \ldots \\ q^{-1} & 0 & 0 & \ldots & 0 & \ldots \\ \hdotsfor{6}\\ q^{-N/2} & 0 & 0 & \ldots & 0 & \ldots \\ \hdotsfor{6} \end{pmatrix}.$$ It is easy to write an integral representation $$\label{3.12} (Ju)(\|x\|_K)=\frac{1-q}{2iq\log q}\int\limits_{\|y\|_k\le 1}(\log \|x\|_K-\log \|y\|_K)u(\|y\|_K\,dy ,$$ that is $$\label{3.13} (Ju)(\|x\|_K)=\frac{1-q}{2iq\log q}\left[ \langle u,1\rangle \log \|\cdot \|_K-\langle u,\log \|\cdot \|_K\rangle 1\right]$$ hence $J$ is a rank 2 operator. We see from (\[3.11\]) that $\operatorname{tr}J=0$. 4\) The only solution in ${\mathcal H}$ (up to the multiplication by a constant) of the equation ${I_0^1}u=0$ is the eigenfunction $u_0$ given by (\[3.8\]). Suppose that $\left( {I_0^1}\right)^u_0=0$. Then $Ju_0=0$. However by (\[3.12\]), $$(Ju_0)(\|x\|_K)=\frac{1-q}{2iq\log q}\log \|x\|_K\int\limits_{\|y\|_K=1}dy=-\frac{(q-1)^2}{2iq^2\log q}\log \|x\|_K,$$ so that $Ju_0\not\equiv 0$, and we have come to a contradiction. This proves that ${I_0^1}$ is a simple Volterra operator. $\blacksquare$ Let us calculate the action of ${I_0^1}$ upon the basis $\{f_n\}$ defined in (\[2.10\]). We find for $\|x\|_K=q^{-j},j\ge 0$, that $$\begin{gathered} \left( {I_0^1}f_n\right) (\|x\|_K)=-\frac{(1-q^{-1})^{1/2}}{\log q}q^{n/2}\int\limits_{\|y\|_K<q^{-j},\|y\|_K=q^{-n}}(\log \|x\|_K-\log \|y\|_K)\,dy\\ =(1-q^{-1})^{1/2}q^{n/2}(j-n)\int\limits_{\|y\|_K<q^{-j},\|y\|_K=q^{-n}}dy =\begin{cases} (1-q^{-1})^{3/2}q^{-n/2}(j-n), & \text{if $n>j$;}\\ 0, & \text{if $n\le j$}. \end{cases}\end{gathered}$$ This implies the equality $$\langle {I_0^1}f_n,f_j\rangle =0 \text{ for $n\le j$},$$ meaning that $\{ f_n\}$ [is a basis of triangular representation for the operator]{} ${I_0^1}$. The operator ${I_0^1}$ is $S$-real with respect to the involution $S$ in ${\mathcal H}$ given by the complex conjugation. Therefore it is $S$-unicellular ([@GK], Appendix, Theorem 5.5). It is not clear whether it is unicellular in the usual (complex) sense. However it is unicellular in a smaller space ${\mathcal H}^p$ defined as a completion of the set of all “polynomials” $\varphi (\|x\|_K)=\sum\limits_{j=0}^N c_j\|x\|_K^j$ with respect to the norm $\\|\varphi \\|=\left\{ \sum \|c_j\|^p\right\}^{1/p}$,$1\le p<\infty$. By virtue of (\[3.6\]), ${I_0^1}$ acts on the space ${\mathcal H}^p$ (isomorphic to $l^p$) as a weighted shift, for which the unicellularity was proved by Yakubovich [@Ya]. Following the notation in [@GT], let us write (\[3.13\]) in the form $$\frac1i \left( {I_0^1}-\left( {I_0^1}\right)^\right) u=\sum\limits_{\alpha,\beta =1}^2\langle u,h_\alpha \rangle j_{\alpha \beta}h_\beta$$ where $h_1(\|x\|_K)=\dfrac{q-1}{iq\log q}(=\operatorname{const})$, $h_2(\|x\|_K)=-\log \|x\|_K$, $x\in O$, $j=\left( \begin{smallmatrix} 0 & 1\\ 1 & 0 \end{smallmatrix} \right)$. For the operator ${I_0^1}$, we consider the $2\times 2$ characteristic matrix-function of inverse argument $$W(z^{-1})=E+izj\left[ \langle \left( E-z{I_0^1}\right)^{-1}h_\alpha ,h_\beta\rangle \right]_{\alpha,\beta =1}^2$$ where $E$ denotes both the unit operator in ${\mathcal H}$ and the unit matrix. For the Volterra operator ${I_0^1}$, $W(z^{-1})$ is an entire matrix-function. Matrix elements of $W(z^{-1})$ are entire functions of zero order. . For small values of $\|z\|$, the Fredholm resolvent $\left( E-z{I_0^1}\right)^{-1}$ is given by the Neumann series $$\left( E-z{I_0^1}\right)^{-1}f=\sum\limits_{n=0}^\infty \left( z{I_0^1}\right)^nf,\quad f\in {\mathcal H}.$$ In order to calculate the characteristic function,we have to compute the functions $\left({I_0^1}\right)^n 1$ and $\left({I_0^1}\right)^n\log \|\cdot \|_K$. The first of them is obtained easily from (\[3.6\]): $$\left( \left( {I_0^1}\right)^n 1\right)(\|x\|_K)=c^n\prod\limits_{m=0}^{n-1}d_m\cdot \|x\|_K^n,\quad \|x\|_K\le 1,$$ where $c=\dfrac{1-q}{q\log q}$, $0<d_m\le Aq^{-m}$. Summing the progression we find that $$\label{3.14} \left( \left( E-z{I_0^1}\right)^{-1}1\right) (\|x\|_K)=\sum\limits_{n=0}^\infty\rho_nz^n\|x\|_K^n,\quad \|\rho_n\|\le C^nq^{-n^2/2},$$ where $C>0$ is a constant. Let us consider $\left({I_0^1}\right)^n\log \|\cdot \|_K$. We have $$\left( {I_0^1}\log\|\cdot \|_K\right) (\|x\|_K)=c\int\limits_{\|y\|_K<\|x\|_K}\left( \log \|x\|_K-\log \|y\|_K\right) \log \|y\|_K\,dy.$$ Setting $y=xt$, $\|t\|_K<1$, we obtain $$\begin{gathered} \left( {I_0^1}\log\|\cdot \|_K\right) (\|x\|_K)=-c\|x\|_K\int\limits_{\|t\|_K<1}\log \|t\|_K(\log \|x\|_K+\log \|t\|_K)\,dt=-ca_0\|x\|_K\log \|x\|_k-cb_0\|x\|_K\\ \stackrel{\text{def}}{=}\sigma_1\|x\|_K\log \|x\|_K-\eta_1\|x\|_K\end{gathered}$$ where $$a_0=\int\limits_{\|t\|_K<1}\log \|t\|_K\,dt,\quad b_0=\int\limits_{\|t\|_K<1}\log^2 \|t\|_K\,dt.$$ A similar calculation yields the expression $$\left( {I_0^1}(\|\cdot \|_K\log \|\cdot \|_K)\right) (\|x\|_K)=-ca_1\|x\|^2_K\log \|x\|_k-cb_1\|x\|_K^2$$ where $$a_1=\int\limits_{\|t\|_K<1}\|t\|_K\log \|t\|_K\,dt,\quad b_1=\int\limits_{\|t\|_K<1}\|t\|_K\log^2 \|t\|_K\,dt.$$ Together with (\[3.6\]), this implies the formula $$\left( \left({I_0^1}\right)^2 \log\|\cdot \|_K\right) (\|x\|_K)=c^2a_0a_1\|x\|^2_K\log \|x\|_k+c^2a_0b_1\|x\|_K^2-cb_0d_1\|x\|_K^2\stackrel{\text{def}}{=}\sigma_2\|x\|_K^2\log \|x\|_K+\eta_2\|x\|_K^2.$$ Introducing similar constants for the next iterations, $$a_n=\int\limits_{\|t\|_K<1}\|t\|^n_K\log \|t\|_K\,dt,\quad b_1=\int\limits_{\|t\|_K<1}\|t\|^n_K\log^2 \|t\|_K\,dt,$$ and noticing that $\|a_n\|,\|b\|_n\le Mq^{-n}$, we prove by induction that $$\label{3.15} \left({I_0^1}\right)^n\log \|\cdot \|_K=\sigma_n\|x\|_K^n\log \|x\|_K+\eta_n\|x\|_K^n$$ where $\|\sigma_n\|,\|\eta \|_n\le C^nq^{-1}q^{-2}\cdots q^{-n+1}=C^nq^{-n(n-1)/2}$. It follows from (\[3.15\]) that $$\left( \left( E-z{I_0^1}\right)^{-1}\log \|\cdot \|_K\right) (\|x\|_K)=\sum\limits_{n=0}^\infty\sigma_nz^n\|x\|_K^n\log \|x\|_K+\sum\limits_{n=0}^\infty\eta_nz^n\|x\|_K^n$$ where $\|\sigma_n\|,\|\eta_n\|\le C_1^nq^{-n^2/2}$. Now we can compute the matrix-function $W(z^{-1})$. By (\[3.14\]), $$\langle \left( E-z{I_0^1}\right)^{-1}h_1,h_1\rangle =\operatorname{const}\cdot \sum\limits_{n=0}^\infty \rho_nz^n\int\limits_{\|x\|_K^n\le 1}\,dx=\sum\limits_{n=0}^\infty \gamma_nz^n$$ where $\|\gamma_n\|\le C_2^nq^{-n^2/2}$, so that this matrix element is an entire function of zero order. Other matrix elements are estimated similarly on the basis of (\[3.15\]), by inserting 1 as an upper bound of $\|x\|_K$ and taking into account the convergence of the integrals of $\log \|x\|_K$ and $\log^2 \|x\|_K$. $\blacksquare$ The Laplace Type Transform ========================== [4.1. Definition and Properties.]{} Our definition of a Laplace type transform is based on the function $v_N$ given by (\[2.6\]). It is essential that $v_N\in \mathcal D(K)$. As we know, ${D^\alpha}v_N=q^{\alpha N}v_N$ ($\alpha >0$). Let $\xi \in K$, $\|\xi \|_K=q^N$. Then for any $x\in K$, $v_N(\|x\|_K)=v_0(\|x\xi \|_K)$, $$D_x^\alpha v_0(\|x\xi \|_K)=D_x^\alpha v_N(\|x\|_K)=q^{\alpha N}v_N(\|x\|_K)=\|\xi\|_K^\alpha v_0(\|x\xi \|_K).$$ We call the function $$\widetilde{\varphi}(\|\xi\|_K)=\int\limits_Kv_0(\|x\xi \|_K)\varphi (\|x\|_K)\,dx$$ the Laplace type transform of a radial function $\varphi\in L^1_{\text{loc}}(K)$. By the dominated convergence theorem, $\widetilde{\varphi}$ is continuous, bounded, and $\widetilde{\varphi}(\|\xi\|_K)\to 0$, $\|\xi\|_K\to \infty$. As a simple computation shows, if $\varphi (\|x\|_K)\equiv \operatorname{const}$, then $\widetilde{\varphi}(\|\xi\|_K)\equiv 0$. The above calculations, together with the selfadjointness of ${D^\alpha}$ in $L^2(K)$, show that $$\widetilde{{D^\alpha}\varphi} (\|\xi\|_K)=\int\limits_K\left( D_x^\alpha v_0(\|x\xi \|_K)\right) (\|x\|_K)\varphi (\|x\|_K)\,dx=\|\xi\|_K^\alpha\widetilde{\varphi}(\|\xi\|_K),\quad \xi\in K.$$ If $\widetilde{\varphi}(\|\xi\|_K)\equiv 0$, then $\varphi (\|x\|_K)\equiv \operatorname{const}$. By the definition, $$\widetilde{\varphi}(\|\xi\|_K)=\int\limits_{\|x\|_K\le \|\xi\|_K^{-1}}\varphi (\|x\|_K)\,dx-\frac1{q-1}\int\limits_{\|x\|_K=q\|\xi\|_K^{-1}}\varphi (\|x\|_K)\,dx.$$ Let $\|\xi\|_K=q^n$, $n\in \mathbb Z$. Then $$\widetilde{\varphi}(q^n)=(1-\frac1q)\sum\limits_{j=-\infty}^{-n}\varphi (q^j)q^j-\varphi (q^{-n+1})q^{-n}.$$ If we denote $\widetilde{\varphi}(q^n)=f_n$, then $$f_{n+1}=(1-\frac1q)\sum\limits_{j=-\infty}^{-n-1}\varphi (q^j)q^j-\varphi (q^{-n})q^{-n-1},$$ so that $$\label{4.1} f_n-f_{n+1}=q^{-n}\left[ \varphi (q^{-n})-\varphi (q^{-n+1})\right].$$ If $\widetilde{\varphi}(q^n)=0$ for all $n$, then, by (\[4.1\]), $ \varphi (q^{-n})=\varphi (q^{-n+1})$ for all $n$, so that $\varphi (q^{-n})\equiv \operatorname{const}$. $\blacksquare$ The identity (\[4.1\]) is of some independent interest, and we formulate it as a corollary. For all $n\in \mathbb Z$, $$\label{4.2} \widetilde{\varphi}(q^{n})-\widetilde{\varphi}(q^{n+1})=q^{-n}\left[ \varphi (q^{-n})-\varphi (q^{-n+1})\right].$$ A function $\varphi$ is (strictly) monotone, if and only if $\widetilde{\varphi}$ is (strictly) monotone. For each $n=1,2,\ldots $, $$\label{4.3} \varphi (q^m)=\varphi (1)+\sum\limits_{j=0}^{m-1}q^{-j}\left[ \widetilde{\varphi}(q^{-j+1})-\widetilde{\varphi}(q^{-j})\right],$$ $$\label{4.4} \varphi (q^{-m})=\varphi (1)+\sum\limits_{j=1}^m q^j\left[ \widetilde{\varphi}(q^j)-\widetilde{\varphi}(q^{j+1})\right],$$ . According to (\[4.2\]), $$\begin{gathered} \varphi (1)-\varphi (q)=\widetilde{\varphi}(1)-\widetilde{\varphi}(q),\\ \varphi (q)-\varphi (q^2)=q^{-1}\left[ \widetilde{\varphi}(q^{-1})-\widetilde{\varphi}(1)\right],\\ \varphi (q^2)-\varphi (q^3)=q^{-2}\left[ \widetilde{\varphi}(q^{-2})-\widetilde{\varphi}(q^{-1})\right],\end{gathered}$$ etc. Summing up the first $m$ equalities we obtain (\[4.3\]). Similarly, by (\[4.2\]), $$\begin{gathered} \varphi (q^{-1})-\varphi (1)=q\left[ \widetilde{\varphi}(q)-\widetilde{\varphi}(q^2)\right],\\ \varphi (q^{-2})-\varphi (q^{-1})=q^2\left[ \widetilde{\varphi}(q^2)-\widetilde{\varphi}(q^3)\right],\end{gathered}$$ etc, and the summation yields (\[4.4\]). $\blacksquare$ Acknowledgement {#acknowledgement .unnumbered} =============== This work was funded in part under the research project “Markov evolutions in real and $p$-adic spaces” of the Dragomanov National Pedagogic University of Ukraine. [999]{} S. Albeverio, A. Yu. Khrennikov, and V. M. Shelkovich, [Theory of $p$-Adic Distributions]{}, Cambridge University Press, 2010. A. D. Bendikov, A. A. Grigor’yan, Ch. Pittet, W. Woess, Isotropic Markov semigroups on ultra-metric spaces, [Russian Math. Surveys]{}, [69]{} (2014), 589–680. M. S. Brodskii, [Triangular and Jordan representations of Linear Operators]{}. American Mathematical Society, Providence, R.I., 1971. I. C. Gohberg and M. G. Krein, [Theory and Applications of Volterra Operators in Hilbert Space]{}. American Mathematical Society, Providence, R.I. 1970. G. M. Gubreev and A. A. Tarasenko, Spectral decomposition of model operators in de Branges spaces, [Sbornik Math.]{} [201]{} (2010), 1599–1634. P. R. Halmos, [Hilbert Space Problem Book]{}, D. Van Nostrand, Princeton, 1967. A. Yu. Khrennikov, S. V. Kozyrev and W. A. Zúñiga-Galindo, [Ultrametric Pseudodifferential Equations with Applications]{}, Cambridge University Press, 2018. A. N. Kochubei, [Pseudo-Differential Equations and Stochastics over Non-Archimedean Fields]{}, Marcel Dekker, New York, 2001. A. N. Kochubei, A non-Archimedean wave equation, [Pacif. J. Math.]{} [235]{} (2008), 245–261. A. N. Kochubei, Radial solutions of non-Archimedean pseudodifferential equations, [Pacif. J. Math.]{} [269]{} (2014), 355–369. A. N. Kochubei, Linear and nonlinear heat equations on a $p$-adic ball, [Ukrainian Math. J.]{} [70]{} (2018), 217–231. A. N. Kochubei, Nonlinear pseudo-differential equations for radial real functions on a non-Archimedean field. [J. Math. Anal. Appl.]{} [483]{} (2020), no. 1, Article 123609, 11 pp. M. M. Malamud, Spectral theory of fractional order integration operators, their direct sums, and similarity problem to these operators of their weak perturbations. In: [Handbook of fractional calculus with applications. Vol. 1]{} (A. N. Kochubei and Yu. Luchko, Eds), De Gruyter, Berlin, 2019, pp. 427–460. V. S. Vladimirov, I. V. Volovich and E. I. Zelenov, [$p$-Adic Analysis and Mathematical Physics]{}, World Scientific, Singapore, 1994. D. V. Yakubovich, Invariant subspaces of weighted shift operators, [J. Soviet Math.]{} [37]{} (1987), 1323–1346. V. A. Zolotarev, L. de Branges spaces and functional models of nondissipative operators, [Mat. Fiz. Anal. Geom.]{} [9]{} (2002), 622–641 (Russian). W. A. Zúñiga-Galindo, [Pseudodifferential Equations over Non-Archimedean Spaces]{}, Lect. Notes Math. 2174, Springer, Cham, 2016.	{ "pile_set_name": "ArXiv" }
--- abstract: 'After a brief survey of the definition and the properties of $\Lambda$-symmetries in the general context of dynamical systems, the notion of “$\Lambda$-constant of motion” for Hamiltonian equations is introduced. If the Hamiltonian problem is derived from a $\Lambda$-invariant Lagrangian, it is shown how the Lagrangian $\Lambda$-invariance can be transferred into the Hamiltonian context and shown that the Hamiltonian equations turn out to be $\Lambda$-symmetric. Finally, the “partial” (Lagrangian) reduction of the Euler-Lagrange equations is compared with the reduction obtained for the corresponding Hamiltonian equations.' author: - \| Giampaolo Cicogna\ [Dipartimento di Fisica, Università di Pisa]{}\ and [INFN, Sezione di Pisa,]{}\ [Largo B. Pontecorvo 3, 50127 Pisa (Italy)]{}\ [[email protected]]{} title: \| $\Lambda$-symmetries of Dynamical Systems,\ Hamiltonian and Lagrangian equations --- Introduction ($\lambda$-symmetries) =================================== Let me briefly recall for the reader’s convenience the basic definition of $\lambda$-symmetry (with lower case $\lambda$), originally introduced by C. Muriel and J.L. Romero in 2001 [@Cic:MR1; @Cic:MR2]. Let me consider the simplest case of a single ODE $\Delta(t,u(t),\dot u,\ddot u\ldots)=0$ for the unknown function $u\,=\,u(t)$ (I will denote by $t$ the independent variable, with the only exception of the final Section 4, because the applications I am going to propose will concern the case of Dynamical Systems (DS), where the independent variable is precisely the time $t$, and $\dot u={\rm d}u/{\rm d} t$, etc.). Given a vector field $$X\,=\,\varphi(u,t){\frac {\partial} {\partial u}}+ \tau(u,t){\frac \partial {\partial t}}$$ the idea is to suitably [modify]{} its prolongation rules. The first $\lambda$-prolongation $X^{(1)}_\lambda$ is the defined by $$\label{Cic:la1}X^{(1)}_\lambda\,=\, X^{(1)}+ \lambda (\varphi-\tau\dot u){\frac \partial {\partial \dot u}}$$ where $\lambda=\lambda(u,\dot u,t)$ is a $C^\infty$ function, and $X^{(1)}$ is the standard first prolongation. Other modifications have to be introduced for higher prolongations, but in the present paper I need only just the first one. An $n$-th order ODE $\Delta=0$ is said to be $\lambda$-invariant under $X^{(n)}_\lambda$ if $$X^{(n)}_\lambda\Delta\Big\|_{\Delta=0}\,=\,0$$ where $X^{(n)}_\lambda$ is the $n$-th $\lambda$-prolongation of $X$. It should be emphasized that $\lambda$-symmetries are not properly symmetries, because they do not transform in general solutions of a $\lambda$-invariant equation into solutions, nevertheless they share with standard Lie point-symmetries some important properties, namely: if an equation is $\lambda$-invariant, then $\bullet$ the order of the equation can be lowered by one $\bullet$ invariant solutions can be found (notice that conditional symmetries do the same, but $\lambda$-symmetries are clearly [not]{} conditional symmetries) $\bullet$ convenient new (“symmetry adapted”) variables can be suggested. In the context of DS, which is the main object of this paper, the first two properties are not effective, the third one is instead one of my starting points. Before considering the role of $\lambda$-symmetries in DS, let me recall that many applications and extensions of this notion have been proposed in these 10 years: these include extensions to systems of ODE’s, to PDE’s, applications to variational principles and Noether-type theorems, the analysis of their connections with nonlocal symmetries, with symmetries of exponential type, with hidden, or “lost” symmetries, with potential, telescopic symmetries as well. Other investigations concern their deep geometrical interpretation, with the introduction of a suitable notion of deformed Lie differential operators, the study of their dynamical effects in terms of changes of reference frames, and so on. Only the papers more directly involved with the argument considered in this paper will be quoted; for a fairly complete list of references see e.g. [@Cic:CHam; @Cic:Gatw; @Cic:GC09]. A very recent application concerns discrete difference equations [@Cic:LRdde]. $\Lambda$-symmetries for DS =========================== I am going to consider the case of dynamical systems, i.e. systems of first-order ODE’s $$\dot u_a\,=\,f_a(u,t)\quad\quad\ (a=1,\ldots, m)$$ for the $m>1$ unknowns $u_a\,=\,u_a(t)$. Let me start with a trivial (but significant) case: if the DS admits a rotation symmetry, then it is completely natural to introduce as new variables the radius $r$ and the angle $\theta$, and the DS immediately takes a simplified form, as well known. However, in general, symmetries of DS may be very singular, and/or difficult to detect. An example can be useful: the DS $$\dot u_1\,=\,u_1u_2 \quad\quad\ \dot u_2\,=\,-u_1^2$$ admits the (not very useful or illuminating) symmetry generated by (with $r^2=u_1^2+u_2^2$) $$X\,=\,\Big({\frac {2u_1} {r^2}}- {\frac {u_1u_2} {r^3}}\log{\frac {u_2-r} {{u_2+r}}}\Big) {\frac {\partial} {\partial u_1}}+ \Big({\frac {2u_2}{r^2}-{\frac{u_1^2}{r^3}} \log{\frac {u_2-r} {{u_2+r}}}\Big) {\frac {\partial} {\partial u_2}}}\ .$$ In this example the rotation (with a commonly accepted abuse of language, the same symbol $X$ denotes both the symmetry and its Lie generator) $$X\,=\,u_2{\frac {\partial} {\partial u_1}}-u_1{\frac \partial {\partial u_2}}$$ is a [$\lambda$-symmetry]{} (its precise definition will be given in a moment), and [not]{} a symmetry in the “standard” sense; [nevertheless]{}, still introducing the variables as before, i.e. $r$ and $\theta$, the DS takes the very simple form $$\dot r\,=\,0\quad\quad\ \dot\theta\,=\,-r\cos\,\theta\ .$$ This is just a first, simple example of the possible role of $\lambda$-symmetries in the context of DS. $\Lambda$-symmetries of general DS ---------------------------------- The natural way to extend the definition (\[Cic:la1\]) of the first $\lambda$-prolongation of the vector field $$X\,=\,\varphi_a(u,t){\frac \partial {\partial u_a}}+ \tau(u,t){\frac\partial {\partial t}}\,=\, \varphi\cdot{\nabla_u}+\tau\partial_t$$ to the case of $m>1$ variables $u_a$ is the following (sum over repeated indices) $$X_\Lambda^{(1)}\,=\, X^{(1)}+\Lambda_{ab}(\varphi_b-\tau\dot u_b) \cdot\nabla_{\dot u_a}$$ where now $\Lambda=\Lambda(t,u_a,\dot u_a)$ is a $(m\times m)$ matrix; accordingly, I denote by the upper case $\Lambda$ these symmetries in this context. To simplify, let me assume from now on $\tau\,=\,0$ (or use evolutionary vector field, it is not restrictive). Then the given DS is $\Lambda$-invariant under $X$ (or $X$ is a $\Lambda$-symmetry for the DS), i.e. $X_\Lambda^{(1)}(\dot u-f)\|_{\dot u=f}=0$, if and only if $$[\,f,\varphi\,]\,_a+\partial_t\varphi_a\,=\,-(\Lambda\varphi)_a \quad\qquad (a=1,\ldots,m)\,$$ where $$[\,f, \varphi\,]\,_a\equiv f_b\nabla_{u_b}\varphi_a-\varphi_b \nabla_{u_b}f_a \ .$$ Given $X$, we now introduce the following new $m+1$ “canonical” (or [symmetry-adapted]{}) variables ([notice that they are independent of]{} $\Lambda$): precisely, $m-1$ variables $w_j=w_j(u)$ which, together with the time $t$, are $X$- invariant: $$X\,w_j\,=\,X\,t\,=\,0\quad\quad (j=1,\ldots,m-1)$$ and the coordinate $z$, “rectifying” the action of $X$, i.e. $$X\,=\,{\frac \partial {\partial z}}\ .$$ Writing the given DS in these new variables, we obtain a “reduced” form of the DS, as stated by the following theorem [@Cic:Olv; @Cic:MRVi; @Cic:CLa]. Let $X$ be a $\Lambda$-symmetry for a given DS; once the DS is written in terms of the new variables $w_j,z,t$, i.e. $$\dot w_j \,=\,W_j(w,z,t) \quad\quad\ \dot z \,=\,Z(w,z,t)$$ the dependence on $z$ of the r.h.s. $W_j\,,Z$ is controlled by the formulas $${\frac{\partial W_j}{\partial z}}\,=\, {\frac{\partial w_j}{\partial u_a}}(\Lambda\varphi)_a\equiv M_j \quad\quad\ {\frac{\partial Z}{\partial z}}\,=\, {\frac{\partial z}{\partial u_a}}(\Lambda\varphi)_a\ \equiv M_{m}\ .$$ One has: $$\bullet\ \ If \ \Lambda=0\ \ then\ M_j=M_m=0 \ \ and \ \ W_j\,,Z \ are\ independent\ of \ z \hskip 7.6cm$$ $$\bullet\ \ If \ \Lambda\,=\,\lambda I\ \ then\ only\ Z\ depends \ on\ z \hskip8.7cm$$ .05cm$\bullet$ Otherwise, a “partial” reduction is obtained:If some $M_k=0$, then $W_k$ is independent of $z$. In terms of the new variables, the $\Lambda$-prolongation becomes $$X_{\Lambda}^{(1)}\,=\,{\frac\partial {\partial z}}+M_{j}{\frac\partial {\partial\dot w_j}}+M_{m}{\frac\partial {\partial\dot z}}\ .$$ The first case ($\Lambda=0$) clearly means that $X$ is an [exact]{}, or standard Lie point-symmetry [@Cic:Olv]; the second one has been considered in detail by Muriel and Romero [@Cic:MRVi] (notice that actually it would be enough to require $\Lambda\varphi=\lambda\varphi$); the last case has been dealt with in [@Cic:CLa]: several situations can be met, depending on the number of vanishing $M_j$ (e.g., one may obtain triangular DS, or similar). Hamiltonian DS -------------- I now consider the special case in which the DS is a [Hamiltonian]{} DS. Obvious changes in the notations can be introduced: the $m$ variables $u=u_a(t)$ are replaced by the $m=2n$ variables $q_\alpha(t),p_\alpha(t)\ (\alpha=1,\ldots,n)$, and the DS is now the system of the Hamiltonian equations of motions for the given Hamiltonian $H=H(q,p,t)$: $$\dot u\,=\,J\nabla H\equiv F(u,t)\quad\,,\quad \ \nabla\equiv \nabla_u\equiv(\nabla_q,\nabla_p)$$ where $J$ is the standard symplectic matrix $$J\,=\,\pmatrix{0&I \cr -I&0}\ .$$ A vector field $X$ can be written accordingly (with $a=1,\ldots,2n\,;\, \alpha=1,\ldots,n$) $$X=\varphi_\alpha(u,t){\frac\partial {\partial q_\alpha}}+ \psi_\alpha (u,t){\frac\partial {\partial p_\alpha}}\equiv{\bf \Phi}\cdot\nabla_u\quad\,,\quad\ {\bf \Phi}\equiv(\varphi_\alpha,\psi_\alpha)$$ and all the above discussion clearly holds if $X$ is a $\Lambda$-symmetry for an Hamiltonian DS. Clearly, here $\Lambda$ is a $(2n\times 2n)$ matrix. But Hamiltonian problems possess certainly a [richer]{} structure with respect to general DS, which deserves to be exploited; a first instance is clearly provided by the notion of conservation rules, with its related topics. Let me then distinguish two cases: $(i$) $X$ admits a [generating function]{} $G(u,t)$ (then $X$ is often called a “Hamiltonian symmetry”): $$\label{Cic:Phi}{\bf \Phi}\,=\,J\nabla G\quad{\rm i.e.}\quad \varphi\,=\,\nabla_pG \ ,\ \psi\,=\,-\nabla_qG$$ this implies $\nabla D_tG=0$, where $D_t$ is the total derivative, i.e. $G$ is a constant of motion, $D_tG=0$, possibly apart from an additional time-dependent term, as well known. $(ii$) $X$ does not admit a generating function: also in this case, defining $$\label{Cic:defS}S(u,t)\equiv \nabla\cdot{\bf \Phi}\quad one \ has \ that\ \quad D_tS\,=\,0$$ and therefore, if $S\not=$ const, then $S$ is a first integral (the examples known to me of first integrals of this form are rather tricky, being usually obtained multiplying symmetries by first integrals; but they “in principle” exist, and their presence will be important for the following discussion, see subsect. 3.4). Direct calculations can show the following: If the Hamiltonian equations of motion admit a $\,\Lambda$-symmetry $X$ with a matrix $\Lambda$, then: in case $(i)$ $$\nabla(D_tG)\,=\,J\,\Lambda\,{\bf \Phi}\,=\,J\,\Lambda\,J\,\nabla\,G$$ in case $(ii)$ $$D_tS\,=\,-\nabla(\Lambda\,{\bf \Phi})\ .$$ When this happens, $G$ (resp. $S$) will be called a “[$\Lambda$-constant of motion]{}”. If $\Lambda\!=\!0$, i.e. when $X$ is a “standard” (or “exact”) symmetry, the above equations become clearly the usual conservation rules; $\Lambda$-symmetries can then be viewed as “perturbations” of the exact symmetries. More explicitly, the equations in Theorem 2 state the precise “deviation” from the conservation of $G$ (resp. of $S$) due to the fact that the invariance under $X$ is “broken” by the presence of a nonzero matrix $\Lambda$. As a special case for case $(i)$, the following Corollary may be of interest: Under mild assumptions ($\Lambda\,{\bf \Phi}\!=\!\lambda\,{\bf \Phi}$, $\lambda\!=\!\lambda(G)$), then the $\Lambda$-constant of motion $G$ satisfies a “completely [separated]{} equation”, involving only $G(t)$: $$\dot G\,=\,\gamma(t,G)\ .$$ This equation expresses how much the conservation of $G(t)$ is “violated” along the time evolution. If $\Lambda$ is in some sense “small”, then $G$ is “almost” conserved. When a $\Lambda$-symmetry of the Hamiltonian equations is inherited by a $\Lambda$-invariant Lagrangian ======================================================================================================= $\Lambda$-invariant Lagrangians, Noether theorem and\ $\Lambda$-conservation rules ----------------------------------------------------- Let me consider (for simplicity) only first-order Lagrangians: $${{\cal L}}\,=\,{{\cal L}}(q_\alpha,\dot q_\alpha,t)\quad\quad\quad (\alpha=1,\ldots,n)$$ Such a Lagrangian is said to be $\Lambda^{({\cal L})}$-invariant[@Cic:MRO; @Cic:CGNoe] under $$X^{({{\cal L}})}\,=\, \varphi_\alpha(q,t){\frac\partial {\partial q_\alpha}}\,=\, \varphi\cdot\nabla_q$$ if there is an $(n\times n)$ matrix $$\Lambda^{({{\cal L}})}\,=\,\Lambda^{({{\cal L}})}(q,\dot q,t)$$ such that $$\Big(X_\Lambda^{({{\cal L}})}\Big)^{(1)}({{\cal L}})\,=\,0$$ where $\Big(X_\Lambda^{({{\cal L}})}\Big)^{(1)}$ is the first $\Lambda^{({{\cal L}})}$-prolongation of $X^{({{\cal L}})}$ (the notation is rather heavy, to carefully distinguish the Lagrangian case from the Hamiltonian one, to be considered in the next subsection). We then have [@Cic:CGNoe] If the Lagrangian ${{\cal L}}$ is $\Lambda^{({{\cal L}})}$-invariant under $X^{({{\cal L}})}$ then, putting $${{\cal P}}_{\alpha\beta}\,=\,\varphi_\alpha p_\beta \quad\quad with \quad\quad p_\beta\,=\, {\frac{\partial{{\cal L}}}{\partial \dot q_\beta}}$$ one has $$D_t{\bf P}= -\Lambda^{({{\cal L}})}_{\alpha\beta}\varphi_\beta{\frac{\partial{{\cal L}}}{\partial \dot q_\alpha}} \,=\, -\big(\Lambda^{({{\cal L}})}\varphi\big)_\alpha p_\alpha$$ where ${\bf P}={{\tt Tr}}({{\cal P}})\,=\,\varphi_\alpha p_\alpha$; or also, introducing a “deformed derivative” ${\widehat D}_{t}$ $$({\widehat D}_{t})_{\alpha\beta}\equiv D_t\delta_{\alpha\beta}+ \Lambda^{({{\cal L}})}_{\alpha\beta} \quad then \quad {{\tt Tr}}({\widehat D}_t{{\cal P}})\,=\,0\ .$$ This result can be called the [“Noether $\Lambda^{({{\cal L}})}$-conservation rule”]{}. Indeed, if $\Lambda^{({{\cal L}})}\!=\!0$, the standard Noether theorem is recovered. In the special case $\Lambda^{({{\cal L}})}\varphi=\lambda \varphi$, the above result becomes $${\widehat D}_t{\bf P}=0\quad\quad {\rm where}\quad\quad {\widehat D}_t=D_t+\lambda\ .$$ Theorem 3 can be extended [@Cic:CGNoe] to divergence symmetries and to generalized symmetries as well. Also, higher-order Lagrangians can be included: the $\Lambda^{({{\cal L}})}$-conservation rule has the same form, but ${{\cal P}}_{\alpha\beta}$ is different: for instance, for second-order Lagrangians one has $${{\cal P}}_{\alpha\beta}\,=\, \varphi_\alpha{\frac{\partial {{\cal L}}}{\partial \dot q_\beta}} + (({\widehat D}_t)_{\alpha\gamma} \varphi_\gamma) {\frac{\partial {{\cal L}}}{\partial \ddot q_\beta}} - \varphi_\alpha D_t{\frac{\partial {{\cal L}}} {\partial\ddot q_\beta}} \ .$$ From Lagrangians to Hamiltonians -------------------------------- Assume to have a Lagrangian which is $\Lambda^{({{\cal L}})}$-invariant under a vector field $$X^{({{\cal L}})}\,=\,\varphi_\alpha{\frac\partial {\partial q_\alpha}}$$ and introduce the corresponding Hamiltonian $H$ with its Hamiltonian equations of motion. The natural question is whether the $\Lambda^{({{\cal L}})}$-symmetry $X^{({{\cal L}})}$ of the Lagrangian is transferred to some $\Lambda^{(H)}$-symmetry $X^{(H)}$ of the Hamiltonian equations of motion. Two problems then arise: $i)$ to extend the vector field $X^{({{\cal L}})}$ to a suitable vector field $X^{(H)}$, and $ii)$ to extend the $(n\times n)$ matrix $\Lambda^{({{\cal L}})}$ to a suitable $(2n\times 2n)$ matrix $\Lambda^{(H)}$. First, the vector field $X^{(H)}$ is expected to have the form $$\label{Cic:Lapq} X\equiv X^{(H)}\,=\, \varphi_\alpha{\frac\partial {\partial q_\alpha}}+ \psi_\alpha{\frac\partial {\partial p_\alpha}}$$ where the coefficient functions $\psi$ must be determined. This can be done observing that the variables $p$ are related to $\dot q$ (and then the first $\Lambda^{({{\cal L}})}$-prolongation of $X^{({{\cal L}})}$ is needed, where the “effect” of $\Lambda^{({{\cal L}})}$ is present). One finds, after some explicit calculations, $$\label{Cic:XH} \psi_\alpha\,=\,{\frac\partial {\partial \dot q_\alpha}}\Big(D_t {\bf P}+\Lambda^{({{\cal L}})}_{\beta\gamma}\varphi_\gamma{\frac{\partial {{\cal L}}}{\partial\dot q_\beta}}\Big)-{\frac{\partial\Lambda^{({{\cal L}})}_ {\beta\gamma}}{\partial\dot q_\alpha}}\varphi_\gamma{\frac{\partial{{\cal L}}} {\partial\dot q_\beta}}- p_\beta{\frac{\partial\varphi_\beta}{\partial q_\alpha}}\ .$$ But the term in parenthesis vanishes if the Lagrangian is $\Lambda^{({{\cal L}})}$-invariant, thanks to Theorem 3; in addition, if $\Lambda^{({{\cal L}})}$ does not depend on $\dot q$ (as happens in most cases, otherwise a separate treatment is needed, see subsect. 3.4), then we are left with $$\label{Cic:psiH} \psi_\alpha=-p_\beta{\frac{\partial\varphi_\beta}{\partial q_\alpha}}\ .$$ This implies that $X$ admits a generating function, which is just $$G\,=\,\varphi_\alpha p_\alpha\equiv{\bf P}$$ using the notations introduced in Theorem 3. Second, let me now introduce the following $(2n\times 2n)$ matrix $$\Lambda\equiv\Lambda^{(H)}\,=\,\pmatrix{\Lambda^{({{\cal L}})} & 0\cr - {\frac{\partial\Lambda^{({{\cal L}})}}{\partial q_\alpha}}p_\gamma & \Lambda^{(2)}}$$ where $\Lambda^{(2)}$ must satisfy ($\Lambda$ is not uniquely defined, as well known) $$\Lambda^{(2)}_{\alpha\beta}\,{\frac{\partial\varphi_\gamma}{\partial q_\beta} }\,=\,\Lambda^{({{\cal L}})}_{\gamma\beta} {\frac{\partial\varphi_\beta} {\partial q_\alpha}}\ .$$ It is well known that Euler-Lagrange equations coming from a $\Lambda^{({{\cal L}})}$-invariant Lagrangian do [not exhibit in general $\Lambda$-symmetry]{}. In contrast with this, it is not difficult to verify explicitly that the Hamiltonian equations of motion turn out to be $\Lambda^{({H)}}$-symmetric under the vector field $X^{({H)}}$ obtained according to the above prescriptions. In conclusion, I have shown the following If ${{\cal L}}$ is a $\Lambda$-invariant Lagrangian under a vector field $X^{({{\cal L}})}$ with a matrix $\Lambda^{({{\cal L}})}$ (not depending on $\dot q$), one can extend $X^{({{\cal L}})}$ to a vector field $X\equiv X^{(H)}$ and the $(n\times n)$ matrix $\Lambda^{({{\cal L}})}$ to a $(2n\times 2n)$ matrix $\Lambda\equiv \Lambda^{(H)}$ in such a way that the resulting Hamiltonian equations of motion are $\Lambda$-symmetric under $X$; in addition, $G=\varphi_\alpha p_\alpha$ is a $\Lambda$-constant of motion. The Lagrangian (with $n=2$) $${{\cal L}}={\frac 1 2}\Big({\frac{\dot q_1} {q_1}}-q_1\Big)^2+{\frac 1 2}(\dot q_1-q_1\dot q_2)^2\exp(-2q_2)+q_1\exp(-q_2)$$ is $\Lambda^{({{\cal L}})}$-invariant under $$X^{({{\cal L}})}\,=\,q_1{\frac\partial {\partial q_1}}+{\frac\partial {\partial q_2}}$$ with $$\Lambda^{({{\cal L}})}\,=\,{\rm diag}\ (q_1,q_1)\ .$$ It is easy to write the Hamiltonian equations of motion and to check that they are indeed $\Lambda$-symmetric under $$X\,=\,q_1{\frac\partial {\partial q_1}}+ {\frac\partial {\partial q_2}}-p_1{\frac\partial {\partial p_1}}$$ with $$\Lambda\,=\,\Lambda^{(H)}\,=\,\pmatrix {q_1&0&0&0\cr 0&q_1&0&0\cr -p_1&-p_2&q_1&0\cr 0&0&0&0}\ .$$ $X$-invariant coordinates are $w_1=q_1\exp(-q_2),\,w_2=q_1p_1,\,w_3=p_2$, and, as expected, the generating function $G=w_2+w_3$ satisfies the $\Lambda$-conservation rule $$\nabla_uD_tG\,=\,J\Lambda{\bf \Phi} \quad\ {\rm or}\quad D_tG\,=\,-q_1G\ .$$ A special, but rather common, case is described by the following: If $$\Lambda^{({{\cal L}})}{\bf \varphi}=c\, \varphi$$ where $c$ is a constant, then also $\Lambda{\bf \Phi}=c\,{\bf \Phi}$ and the “most complete” reduction of the Hamiltonian equations of motion is obtained: $$\dot G=\gamma(G,t)\quad\quad \dot w_j=W_j(w,G,t)\quad\quad \dot z=Z(w,G,z,t)$$ Reduction of the Euler-Lagrange equations versus\ the Hamiltonian equations ------------------------------------------------- In this section I want to compare the reduction procedure which is provided by the presence of a $\Lambda$-symmetry of a Lagrangian (i.e. the reduction of Euler-Lagrange equations) with the analogous reduction of the Hamiltonian equations of motion. Let me start recalling that any vector field$X=\varphi_\alpha\partial/\partial q_\alpha$ admits $n$ (0-order) invariants (as already said, see subsect. 2.1) $$w_j=w_j(q,t)\quad\ (j=1,\ldots,n-1) \quad {\rm and \ the\ time}\ t$$ and $n$ first-order differential invariants $\eta_\alpha=\eta_\alpha(q,t,\dot q)$ under the first prolongation $X^{(1)}$ $$X^{(1)}\eta_\alpha\,=\,0 \quad\quad\quad\ (\alpha=1,\ldots,n)\ .$$ [Both]{} if $X^{(1)}$ is standard and if it is a $\Lambda$ prolongation ([under the condition]{} $\Lambda\,\varphi=\lambda\,\varphi$), it is well known that $\dot w_j$ are $n-1$ first-order differential invariants (notice that this is an “algebraic” property, not related to dynamics). If one now chooses another independent first-order differential invariant $\zeta=\zeta(q,t,\dot q)$, then one has that any first-order $\Lambda^{({{\cal L}})}$-invariant Lagrangian is a function of the above $2n$ invariants $$t,w_j,\dot w_j\quad{\rm and}\quad \zeta \ .$$ Writing the Lagrangian in terms of these variables, the Euler-Lagrange equation for $\zeta$ is then simply $${\frac{\partial {{\cal L}}}{\partial\zeta}}\,=\,0\ .$$ This first-order equation provides in general a “partial” reduction, i.e., it produces only [particular solutions]{}, even considering the Euler-Lagrange equations for the other variables [@Cic:MRO; @Cic:CHam] (notice that this is true both for exactly invariant and for $\Lambda^{({{\cal L}})}$-invariant Lagrangians). I want to emphasize that, introducing $\Lambda$-symmetric Hamiltonian equations of motion along the lines stated in Theorem 4, then a “better” reduction is obtained, and no solution is lost. The following example clarifies this point. The Lagrangian ($n=2$) $${{\cal L}}\,=\,{\frac 1 2}\Big({\frac{\dot q_1}{q_1}}-\log q_1\Big)^2+{\frac 1 2}\Big({\frac{\dot q_1}{q_1}}+{\frac{\dot q_2}{q_2}}\Big)^2\quad\quad (q_1>0)$$ is $\Lambda^{({{\cal L}})}$-invariant under $$X^{({{\cal L}})}\,=\,q_1{\frac\partial {\partial q_1}}-q_2{\frac\partial {\partial q_2}}$$ with $\Lambda^{({{\cal L}})}={\rm diag}\ (1,1)$. With $$w\,=\,q_1q_2,\ \dot w\,=\,\dot q_1q_2+q_1\dot q_2,\ \zeta\,=\,{\frac{\dot q_1}{ q_1}}-\log q_1$$ the Lagrangian becomes $${{\cal L}}\,=\,{\frac 1 2}\zeta^2+{\frac1 2}{\frac{\dot w^2}{w^2}}$$ and the Euler-Lagrange equation for $\zeta$ is $${\partial {{\cal L}}/{\partial \zeta}}=\zeta=0\quad\quad\ {\rm or}\quad\ \dot q_1\,=\,q_1\log q_1$$ with the particular solution $$q_1(t)\,=\,\exp(c\, e^t)\ .$$ The corresponding Hamiltonian equations of motion are $\Lambda$-symmetric under $$X\,=\, q_1{\frac\partial {\partial q_1}}- q_2{\frac\partial {\partial q_2}}-p_1{\frac\partial {\partial p_1}}+p_2{\frac\partial {\partial p_2}}$$ with $\Lambda={\rm diag}\ (1,1,1,1)$. Invariants under this $X$ are $$w_1\,=\,q_1q_2,\ w_2\,=\, q_1p_1,\ w_3\,=\,q_2p_2$$ and $X$ is generated by $G=w_2-w_3$. A “complete” reduction is obtained: with $z=\log q_1$, we get $$\dot w_1\,=\,w_1w_3 \quad\quad \dot w_2\,=\,w_3-w_2$$ $$\dot G\,=\,-G \quad\quad\ \dot z\,=\,z+w_2-w_3$$ The above “partial” (Lagrangian) solution $\zeta=0$ corresponds to $$\dot z\,=\,z,\quad w_2=w_3=c= {\rm const},\quad \dot w_1=cw_1\ .$$ From the Hamiltonian equations, instead, e.g.: $$q_1(t)\,=\,\exp(c\, e^t)+c_1\exp(-t)\quad\quad {\rm etc.}$$ The reader can easily complete the calculations. When $\Lambda^{({{\cal L}})}$ depends on $\dot q$ ------------------------------------------------- If $\Lambda$ depends also on $\dot q$ (see eq.s (\[Cic:Lapq\],\[Cic:XH\])), the calculations performed in subsect. 3.2 cannot be repeated, the coefficient functions $\psi_\alpha$ cannot be expressed in the simple form (\[Cic:psiH\]) and the vector field $X$ does not admit a generating function $G$. In this case one can resort to the other quantity $S$, introduced in (\[Cic:defS\]), which provides a $\Lambda$-constant of motion. An example can completely illustrate this situation. ($n=1$) $${{\cal L}}\,=\,{\frac 1 2}\Big({\frac{\dot q}{q}}+1\Big)^2\exp(-2q)$$ is $\Lambda^{({{\cal L}})}$-invariant under $$X^{({{\cal L}})}\,=\,q{\frac\partial {\partial q}}\quad\quad\ {\rm with}\quad\quad \Lambda^{({{\cal L}})}\,=\,q+\dot q\ .$$ One finds $\psi=-qp-p$ and the resulting vector field $$X\,=\,q{\frac\partial {\partial q}}-(qp+p){\frac\partial {\partial p}}$$ does [not]{} admit a generating function. Nevertheless, the Hamiltonian equations of motion are $\Lambda$-symmetric under $X$ with $$\Lambda\,=\,\pmatrix{ q+\dot q & 0 \cr -p & q+\dot q}\ .$$ Here $$S\,=\,-q$$ satisfies $D_tS=-\nabla(\Lambda\Phi)$ and is a $\Lambda$-constant of motion. A digression: general $\Lambda$-invariant Lagrangians ===================================================== The $\Lambda$-invariance of a Lagrangian ${{\cal L}}={{\cal L}}(q,\dot q,t)$ considered in subsect. 3.1 is a special case of a much more general situation. Instead of $n$ time-dependent quantities $q_\alpha(t)$, let me consider now $n$ “fields” $$u_\alpha(x_i)\quad (\alpha=1,\ldots,n\,;\,i=1,\ldots,s)$$ depending on $s>1$ real variables $x_i$. Now, the Euler-Lagrange equations become a system of PDE’s, and the notion of $\mu$-symmetry [@Cic:GM; @Cic:CGMmu] extends and replaces that of $\lambda$-symmetry (or $\Lambda$-symmetry if $n>1$). In this case, there are $s>1$ matrices $\Lambda_i$ ($n\times n$), which must satisfy the compatibility condition $$\label{Cic:LaLa}D_i\Lambda_j-D_j\Lambda_i+[\Lambda_i,\, \Lambda_j]\,=\,0\quad \quad (D_i\equiv D_{x_i})$$ which can be rewritten putting ${\widehat D}_i\,=\,D_i\,\delta+\Lambda_i$ (or, in explicit form: $({\widehat D}_i)_{\alpha\beta}\,=\,D_i\delta_{\alpha\beta}+ (\Lambda_i)_{\alpha\beta}$, with a notation extending the one introduced in Theorem 3), $$[ {\widehat D}_i , {\widehat D}_j ] \ = \ 0 \ .$$ Then one has [@Cic:CGNoe; @Cic:GM; @Cic:CGMmu]: Given $s>1$ matrices $\Lambda_i$ satisfying (\[Cic:LaLa\]), there exists (locally) a $(n\times n)$ nonsingular matrix $\Gamma$ such that $$\Lambda_i\,=\,\Gamma^{-1}(D_i\Gamma)\ .$$ If a Lagrangian ${{\cal L}}$ is $\Lambda$-invariant under a vector field $$X\,=\,\varphi_\alpha{\frac {\partial} {\partial u_\alpha}}$$ then there is a matrix-valued vector $${{\cal P}}_i\equiv({{\cal P}}_i)_{\alpha\beta}$$ which is $\Lambda$-conserved; this $\Lambda$-conservation law holds in the form $${\tt {Tr}}\, \big[ \Gamma^{-1}D_i\big( \Gamma\,{{\cal P}}_i \big)\big] \,=\, 0$$ or in the equivalent forms $$D_i{{\bf P}}_i\,=\,-(\Lambda_i)_{\alpha\beta}({{\cal P}}_i)_{\beta\alpha} \,=\,- {\tt {Tr}}(\Lambda_i{{\cal P}}_i)\ ,\ \ where \quad {{\bf P}}_i=({{\cal P}}_i)_{\alpha\alpha}\,=\,{\tt Tr}\,{{\cal P}}_i \ ,$$ $${\tt {Tr}}({\widehat D}_i\,{{\cal P}}_i)\,=\,0\ .$$ For first-order Lagrangians the $\Lambda$-conserved “current density vector” ${\cal P}_i$ is given by $$({{\cal P}}_i)_{\alpha\beta}\,=\,\varphi_\alpha{\frac{\partial {\cal L}} {\partial u_{\beta,i}}} \quad\quad \quad {\rm where}\quad \quad u_{\beta,i}\,=\,{\frac{\partial u_\beta} {\partial x_i}}$$ and for second-order Lagrangians by $$({{\cal P}}_i)_{\alpha\beta}\,=\, \varphi_\beta{\frac {\partial{{\cal L}}}{\partial u_{\alpha,i}}}+ (({\widehat D_j)_{\beta \gamma}\varphi_\gamma){\frac{\partial {{\cal L}}} {\partial u_{\alpha,ij}}}-\varphi_\beta D_j {\frac{\partial {{\cal L}}} {\partial u_{\alpha,ij}}}}\ .$$ Let $n=s=2$. Writing for ease of notation, $x,y$ instead of $x_1,\, x_2$, and $u=u(x,y), \,v=v(x,y)$ instead of $u_1,\,u_2$, consider the vector field $$\label{Cic:exe4}X\,=\,u{\frac \partial {\partial u}}+ {\frac \partial {\partial v}}$$ and the two matrices $$\Lambda_1\,=\,\pmatrix{0 & 0\cr u_x & 0} \quad\quad \Lambda_2\,= \pmatrix { 0 & 0\cr u_y & 0 }$$ and then $$\Gamma\,=\,\pmatrix{1 & 0 \cr u & 1}\ .$$ It is easy to check that the Lagrangian $${\cal L}\,=\, {\frac 1 2}\Big( u_x^2+u_y^2\Big)-{\frac 1 u} \big( u_xv_x+u_yv_y \big) + u^2 \exp(-2v)$$ is $\Lambda$-invariant (or better, in this context, $\mu$-invariant) but not invariant under the above vector field $X$. The $\mu$-conservation law ${\tt Tr} (\widehat D_i {{\cal P}}_i)=0$ takes here the form $$D_i {\bf P}^i \equiv D_x \big( uu_x-v_x-{\frac {u_x} u} \big)+ D_y\big(uu_y-v_y-{\frac {u_y} u}\big)\,=\, u_x^2+u_y^2\ .$$ In agreement with Theorem 5, the r.h.s. of this expression is precisely equal to $$-{\tt Tr}(\Lambda_i{\cal P}_i)= -(\Lambda_i\varphi)_\alpha{\frac {\partial {\cal L}} {\partial u_{\alpha,i}}}\ .$$ Notice that in this case the quantity $u_x^2+u_y^2$ is just the “symmetry-breaking term", i.e. the term which prevents the above Lagrangian from being exactly symmetric under the vector field (\[Cic:exe4\]). It should be remarked that $\mu$-symmetries are actually strictly related to [standard]{} symmetries, or – more precisely – are [locally gauge-equivalent]{} to them (see for details [@Cic:CGNoe; @Cic:Gatw; @Cic:Ggau]). Given indeed the vector field $X=\varphi_\alpha\partial/\partial u_\alpha$ and the $s$ matrices $\Lambda_i$, let me denote by $$X_\Lambda^{(\infty)}\,=\,\sum_J\Psi^{(J)}_ \alpha {\frac \partial {\partial u_{\alpha,J}}}$$ the infinite $\Lambda$-prolongation of $X$, where the sum is over all multi-indices $J$ as usual, and $\Psi^{(0)}_\alpha=\varphi_\alpha$. Introducing now the other vector field $\widetilde X$ $$\widetilde X\equiv\widetilde\phi_\alpha{\frac\partial{\partial u_\alpha}}\ \quad {\rm with}\quad \ \widetilde\varphi_\alpha\equiv(\Gamma\,\varphi)_\alpha$$ where $\Gamma$ is assigned in Theorem 5, and denoting by $$\widetilde X^{(\infty)}\,=\, \sum_J\widetilde\varphi^{(J)}_\alpha{\frac \partial {\partial u_{\alpha,J}}}$$ the [standard]{} prolongation of $\widetilde X$, one has [@Cic:CGMmu; @Cic:CGNoe] that the coefficient functions $\Psi^{(J)}_\alpha$ of the $\Lambda$ prolongation of $X$ are connected to the coefficient functions $\widetilde\varphi^{(J)}_\alpha$ of the standard prolongation of $\widetilde X$ by the relation $$\Psi^{(J)}_\alpha\,=\,\Gamma^{-1}\widetilde\varphi^{(J)}_\alpha \ .$$ In the particularly simple case $n=1$ (i.e., a single “field" $u(x_i)$), then the $s>1$ matrices $\Lambda_i$, and the matrix $\Gamma$ as well, become (scalar) functions $\lambda_i$ and $\gamma$; in this case, if a Lagrangian is $\mu$-invariant under the vector field $X$, then it is also invariant under the [standard]{} symmetry $\widetilde X=\gamma X$. In addition, the $\mu$ conservation law can be also expressed as a [standard]{} conservation rule $$D_i{\bf\widetilde P}^i\,=\,0$$ where ${\bf\widetilde P}^i=\gamma\,\varphi_\alpha\partial{\cal L}/\partial u_{\alpha,i}$ is the “current density vector” determined by the vector field $\widetilde X=\gamma X$. Let now $n=1,\,s=2$, and let me introduce for convenience as independent variables the polar coordinates $r,\theta$. I am considering a single “field" $u=u(r,\theta)$ and the rotation vector field $X=\partial/\partial\theta$. The Lagrangian $${\cal L}\,=\, {\frac 1 2} r^2\exp(-\epsilon\theta)u_r^2+{\frac 1 2} \exp(\epsilon\theta)u_\theta^2$$ is clearly not invariant under rotation symmetry (if $\epsilon\not=0$), but is $\mu$-invariant with $\lambda_1=0,\,\lambda_2=\epsilon$. The above Lagrangian is the Lagrangian of a perturbed Laplace equation, indeed the Euler-Lagrange equation is the PDE $$r^2u_{rr}+2ru_r+\exp(2\epsilon\theta)(u_{\theta\theta}+\epsilon\, u_\theta)\,=\,0 \ .$$ It is easy to check that the current density vector $${\bf P}\equiv\big(-r^2\exp(-\epsilon\theta)u_ru_\theta\, , \, {\frac 1 2} r^2\exp(-\epsilon\theta)u_r^2-{\frac 1 2} \exp(\epsilon\theta)u_\theta^2\big)$$ satisfies the $\mu$-conservation law $$D_i{\bf P}_i\,=\,-\epsilon {\bf P}_2 \ .$$ According to the above remark on the (local) equivalence of the $\mu$-symmetry $X$ to the standard symmetry $\widetilde X\,=\,\gamma\,X\,=\,\exp(\epsilon\theta)\,{\partial/\partial\theta}$, also the (standard) conservation law $D_i\widetilde{{\bf P}}^i=0$ holds, with $$\widetilde{{\bf P}}\equiv \Big(-r^2 u_ru_\theta\, , \, {\frac 1 2} r^2 u_r^2- {\frac 1 2}\exp(2\epsilon\theta)u_\theta^2\Big)\ .$$ Conclusions =========== I have shown that the notion of $\lambda$-symmetry, and the related procedures for studying differential equations, can be conveniently extended to the case of dynamical systems. The use and the interpretation of this notion becomes particularly relevant when the DS is a Hamiltonian system, and even more if the symmetry is inherited by an invariant Lagrangian: in this context indeed it is possible to introduce in a natural way and to draw a comparison between the notions of $\Lambda$-constant of motion and of Noether $\Lambda$-conservation rule. Similarly, the symmetry properties of Euler-Lagrange equations and of the Hamiltonian ones can be compared, and some reduction techniques for the equations can be conveniently introduced. Finally, I have shown that the $\Lambda$-invariance of the Lagrangians in the context of the DS is a special case of a more general and richer situation, where several independent variables are present and a $\Lambda$-conservation rule of very general form is true. Another interesting problem is the nontrivial relationship between $\lambda$ (or $\Lambda$, or $\mu$) symmetries with the standard ones. An aspect of this problem has been mentioned in the above section of this paper. In different situations, this may involve the introduction of nonlocal symmetries and other concepts in differential geometry, as briefly indicated in the Introduction, which clearly go beyond the scope of the present contribution. [99]{} Muriel C. and Romero J.L., New method of reduction for ordinary differential equations [IMA J. Appl. Math.]{}, 2001, V.66, 111–125 Muriel C. and Romero J.L., $C^\infty$-symmetries and nonsolvable symmetry algebras [IMA J. Appl. Math.]{}, 2001, V.66, 477–498 Cicogna G., Symmetries of Hamiltonian equations and $\Lambda$-constants of motion [J. Nonlin. Math. Phys.]{}, 2009, V.16, 43–60 Gaeta G., Twisted symmetries of differential equations [J. Nonlin. Math. Phys.]{}, 2009, V.16, 107–136 Gaeta G. and Cicogna G., Twisted symmetries and integrable systems [Int. J. Geom. Meth. Math. Phys.]{}, 2009, V.6, 1306–1321 Levi D. and Rodrìguez M.A., lambda-symmetries for discrete equations [J. Phys. A: Math. Theor.]{}, 2010, V.43, 292001 (9pp) Olver P.J., [Application of Lie Groups to Differential Equations]{}, Springer, Berlin, 1986 and [Romero J.L.]{}, $C^\infty$ symmetries and integrability of ordinary differential equations, in [Proceedings of the First Colloquium on Lie Theory and Applications]{}, Editors: [Bajo I.]{} and [Sanmartin E.]{}, Publicacións da Universidade de Vigo, Vigo, 2002 Cicogna G., Reduction of systems of first-order differential equations via $\Lambda$-symmetries [Phys. Lett. A]{}, 2008, V.372, 3672–3677 , [Romero J.L.]{} and [Olver P.J.]{}, Variational $C^\infty$ symmetries and Euler-Lagrange equations [J. Diff. Eqs.]{}, 2006, V.222, 164–184 and [Gaeta G.]{}, Noether theorem for $\mu$-symmetries [J. Phys. A: Math. Theor.]{}, 2007, V.40, 11899–11921 and [Morando P.]{}, On the geometry of lambda-symmetries and PDEs reduction [J. Phys. A: Math. Gen.]{}, 2004, V.37, 6955–6975 and [Morando P.]{}, On the relation between standard and $\mu$-symmetries for PDEs [J. Phys. A: Math. Gen.]{}, 2004, V.37, 9467–9486 , A gauge-theoretic description of $\mu$-prolongations, and $\mu$-symmetries of differential equations [ J. Geom. Phys.]{}, 2009, V.59, 519–539	{ "pile_set_name": "ArXiv" }
--- abstract: \| The logarithmically enhanced $\alpha^3\ln(1/\alpha)$ corrections to the para- and orthopositronium decay widths are calculated in the framework of dimensionally regularized nonrelativistic quantum electrodynamics. In the case of parapositronium, the correction is negative, approximately doubles the effect of the leading logarithmic $\alpha^3\ln^2(1/\alpha)$ one, and is comparable to the nonlogarithmic $O(\alpha^2)$ one. As for orthopositronium, the correction is positive and almost cancels the $\alpha^3\ln^2(1/\alpha)$ one. The uncertainties in the theoretical predictions for the decay widths are reduced. PACS numbers: 12.20.Ds, 31.30.Jv, 36.10.Dr author: - \| [Bernd A. Kniehl]{} and [Alexander A. Penin]{}[^1]\ [II. Institut für Theoretische Physik, Universität Hamburg,]{}\ [Luruper Chaussee 149, 22761 Hamburg, Germany]{} title: \| -3cm DESY 00-068ISSN 0418-9833 hep-ph/0004267 April 2000 1.5cm Order $\alpha^3\ln(1/\alpha)$ Corrections to Positronium Decays --- 16.cm 22.cm Positronium (Ps), which is an electromagnetic bound state of the electron $e^-$ and the positron $e^+$, is the lightest known atom. Since its theoretical description is not plagued by strong-interaction uncertainties, thanks to the smallness of the electron mass $m_e$ relative to typical hadronic mass scales, its properties can be calculated perturbatively in quantum electrodynamics (QED), as an expansion in Sommerfeld’s fine-structure constant $\alpha$, with very high precision. Ps is thus a unique laboratory for testing the QED theory of weakly bound systems. The decay widths of the $^1S_0$ parapositronium (p-Ps) and $^3S_1$ orthopositronium (o-Ps) ground states to two and three photons, respectively, have been the subject of a vast number of theoretical and experimental investigations. The present theoretical knowledge may be summarized as: $$\begin{aligned} \Gamma_p^{\rm th}&=&\Gamma^{(0)}_{p} \left[1+A_p{\alpha\over\pi}+2\alpha^2\ln{1\over\alpha} +B_p\left({\alpha\over\pi}\right)^2 \right.\nonumber\\ &&{}-\left. {3\alpha^3\over2\pi}\ln^2{1\over\alpha} +C_p{\alpha^3\over\pi}\ln{1\over\alpha}+D_p\left({\alpha\over\pi}\right)^3 +\ldots\right], \label{spar}\\ \Gamma_o^{\rm th}&=&\Gamma^{(0)}_{o} \left[1+A_o{\alpha\over\pi}-{\alpha^2\over3}\ln{1\over\alpha} +B_o\left({\alpha\over\pi}\right)^2 \right.\nonumber\\ &&{}-\left. {3\alpha^3\over2\pi}\ln^2{1\over\alpha} +C_o{\alpha^3\over\pi}\ln{1\over\alpha}+D_o\left({\alpha\over\pi}\right)^3 +\ldots\right], \label{sort}\end{aligned}$$ where $$\begin{aligned} \Gamma^{(0)}_{p}&=&{\alpha^5m_e\over 2},\nonumber\\ \Gamma^{(0)}_{o}&=&{2(\pi^2-9)\alpha^6m_e\over 9\pi},\end{aligned}$$ are the lowest-order results. The $O(\alpha)$ coefficients in Eqs. (\[spar\]) [@HarBra] and (\[sort\]) [@CLS] read $$\begin{aligned} A_p&=&{\pi^2\over4}-5,\nonumber\\ A_o&=&-10.286\,606(10).\end{aligned}$$ The logarithmically enhanced $\alpha^2\ln(1/\alpha)$ terms in Eqs. (\[spar\]) and (\[sort\]) have been obtained in Refs. [@KhrYel; @CasLep1], respectively. Recently, the nonlogarithmic $O(\alpha^2)$ coefficients in Eqs. (\[spar\]) [@CMY2] and (\[sort\]) [@AFS] have been found to be $$\begin{aligned} B_p&=&5.14(30),\nonumber\\ B_o&=&44.52(26). \label{bpbo}\end{aligned}$$ Note that the light-by-light scattering diagrams have been omitted in Ref. [@AFS]. In the p-Ps calculation [@CMY2] the diagrams of this type increase the coefficient $B_p$ by $1.28$, and, therefore, their contribution to the coefficient $B_o$ is assumed to be relatively small. The p-Ps (o-Ps) decays into four (five) photons, which are not included in Eq. (\[bpbo\]), lead to an increase of the coefficient $B_p$ ($B_o$) by $0.274(1)$ ($0.19(1)$) [@LMSZ]. In $O(\alpha^3)$, only the leading logarithmic $\alpha^3\ln^2(1/\alpha)$ terms are known [@Kar]. Including all the terms known so far, we obtain for the p-Ps and o-Ps total decay widths $$\begin{aligned} \Gamma_p^{\rm th}&=&7989.512(13)~\mu{\rm s}^{-1}, \label{thpar}\\ \Gamma_o^{\rm th}&=&7.039943(10)~\mu{\rm s}^{-1}, \label{thort}\end{aligned}$$ where the given errors stem only from the coefficients $B_p$ and $B_o$ respectively and we postpone the discussion of total uncertainty of theoretical estimates to the end of the paper. The purpose of this letter is complete our knowledge of the logarithmically enhanced terms of $O(\alpha^3)$ by providing the coefficients $C_p$ and $C_o$, in analytic form. We also give order-of-magnitude estimates of the unknown coefficients $D_p$ and $D_o$. On the experimental side, the present situation is not entirely clear. Recently, the Ann Arbor group measured the p-Ps width to be [@Annp] $$\Gamma_p^{\rm exp}=7990.9(1.7)~\mu{\rm s}^{-1}, \label{exannp}$$ which agrees with Eq. (\[thpar\]) within the experimental error. However, in the case of o-Ps, their measurements [@Ann1o; @Ann2o], $$\begin{aligned} \Gamma_o^{\rm exp}(\mbox{gas})&=&7.0514(14)~\mu{\rm s}^{-1}, \nonumber\\ \Gamma_o^{\rm exp}(\mbox{vacuum})&=&7.0482(16)~\mu{\rm s}^{-1}, \label{exanno} \end{aligned}$$ exceed Eq. (\[thort\]) by 8 and 5 experimental standard deviations, respectively. This apparent contradiction is known as the [o-Ps lifetime puzzle]{}. On the other hand, the Tokyo group found [@Tok] $$\begin{aligned} \Gamma_o^{\rm exp}(\mbox{SiO$_2$})&=&7.0398(29)~\mu{\rm s}^{-1}, \label{extoko}\end{aligned}$$ which agrees with Eq. (\[thort\]) within the experimental error. Leaving this aside, the o-Ps results from Ann Arbor could be considered as a signal of new physics beyond the standard model. However, a large number of exotic decay modes have already been ruled out [@CzaKar]. No conclusion on the o-Ps lifetime puzzle can be drawn until the experimental precision increases and the data become unambiguous. On the theoretical side, it is an urgent matter to improve the predictions of the Ps lifetimes as much as possible. Thus, one is faced with the task of analyzing the $O(\alpha^3)$ corrections, which is extremely difficult, especially for o-Ps. However, there is a special subclass of the $O(\alpha^3)$ corrections which can be analyzed separately, namely those which are enhanced by powers of $\ln(1/\alpha)\approx5$. They may reasonably be expected to provide an essential part of the $O(\alpha^3)$ corrections. This may be substantiated by considering Eqs. (\[spar\]) and (\[sort\]) in $O(\alpha^2)$, where logarithmic terms enter for the first time. In the case of p-Ps, 98% of the $O(\alpha^2)$ correction stems from the logarithmic term. In the case of o-Ps the logarithmic term is not so dominant but still gives about $1/4$ of the total $O(\alpha^2)$ correction. The origin of the logarithmic corrections is the presence of several scales in the bound-state problem. The dynamics of the nonrelativistic (NR) $e^+e^-$ pair near threshold involves four different scales [@BenSmi]: (i) the hard scale (energy and momentum scale like $m_e$); (ii) the soft scale (energy and momentum scale like $\beta m_e$); (iii) the potential scale (energy scales like $\beta^2m_e$, while momentum scales like $\beta m_e$); and (iv) the ultrasoft (US) scale (energy and momentum scale like $\beta^2m_e$). Here $\beta$ denotes the electron velocity in the center-of-mass frame. The logarithmic integration over a loop momentum between different scales yields a power of $\ln(1/\beta)$. Since Ps is approximately a Coulomb system, we have $\beta\propto\alpha$. This explains the appearance of powers of $\ln(1/\alpha)$ in Eqs. (\[spar\]) and (\[sort\]). The leading logarithmic corrections may be obtained straightforwardly by identifying the regions of logarithmic integration [@KhrYel; @CasLep1; @Kar]. The calculation of the subleading logarithms is much more involved because certain loop integrations must be performed exactly beyond the logarithmic accuracy. In the following, we briefly outline the main features of our analysis. We work in NR QED (NRQED) [@CasLep2], which is the effective field theory that emerges by expanding the QED Lagrangian in $\beta$ and integrating out the hard modes. If we also integrate out the soft modes and the potential photons, we arrive at the effective theory of potential NRQED (pNRQED) [@PinSot1], which contains potential electrons and US photons as active particles. Thus, the dynamics of the NR $e^+e^-$ pair is governed by the effective Schrödinger equation and by its multipole interaction with the US photons. The corrections from harder scales are contained in the higher-dimensional operators of the NR Hamiltonian, corresponding to an expansion in $\beta$, and in the Wilson coefficients, which are expanded in $\alpha$. In the process of scale separation, spurious infrared (IR) and ultraviolet (UV) divergences arise, which endow the operators in the NR Hamiltonian with anomalous dimensions. We use dimensional regularization (DR), with $d=4-2\epsilon$ space-time dimensions, to handle these divergences [@PinSot1; @PinSot2; @CMY1]. This has the advantage that contributions from different scales are matched automatically. The logarithmic corrections are closely related to the anomalous dimensions and can be found by analyzing the divergences of the NR effective theory. In this way, we have obtained the leading logarithmic third-order corrections to the energy levels and wave functions at the origin of heavy quark-antiquark bound states [@KniPen2], which includes the QED result [@KhrYel; @CasLep1; @Kar] as a special case. Here, we extend this approach to the subleading logarithms in QED. Note that the NRQED approach, endowed with an explicit momentum cutoff and a fictitious photon mass to regulate the UV and IR divergences, has also been applied to find the third-order correction, including subleading logarithms, to the hyperfine splitting in muonium [@NioKin]. The annihilation of Ps is the hard process which gives rise to imaginary parts in the local operators of the NR Hamiltonian [@LLM]. The decay width can be obtained by averaging these operators over the bound-state wave function. The hard-scale corrections, which require fully relativistic QED calculations and are most difficult to find, do not depend on $\beta$ and do not lead to logarithmic contributions by themselves. However, they can interfere with the logarithmic corrections from the softer scales. Thus, the only results from relativistic perturbation theory that enter our analysis are (i) the one-loop hard renormalizations of the imaginary parts of the leading four-fermion operators, [i.e.]{}, the Born decay amplitudes, which are given by the coefficients $A_p$ and $A_o$, and (ii) the hard parts of the one-loop $O(\alpha\beta^2)$ operators [@MPS]. The missing ingredients can all be obtained in the NR approximation. These include (i) the correction to the Ps ground-state wave function at the origin due to the $O(\alpha\beta^2)$ terms in the NR Hamiltonian, (ii) the $O(\alpha\beta^2)$ and $O(\alpha^2\beta)$ corrections to the leading four-fermion operators, and (iii) the correction due to the emission and absorption of US photons by the Ps bound state. The value of the ground-state ($n=1$) wave function at the origin $\psi_1(0)$ may be extracted from the NR Green function $G({\bf x},{\bf y},E)$, which satisfies the equation $$\left({\cal H}_C+\Delta{\cal H}-E\right)G({\bf x},{\bf y},E) =\delta^{(3)}({\bf x}-{\bf y}), \label{Sch}$$ where ${\cal H}_C$ is the Coulomb Hamiltonian and $\Delta{\cal H}$ stands for the terms of higher orders in $\alpha$ and $\beta$. The solution of Eq. (\[Sch\]) can be found in time-independent perturbation theory as an expansion in $\alpha$ around the leading-order Coulomb Green function. We thus obtain the correction $\Delta\psi_1^2$ in the relationship $\|\psi_1(0)\|^2=\left\|\psi_1^C(0)\right\|^2\left(1+\Delta\psi_1^2\right)$, where $\psi_1^C(0)$ is the ground-state wave function at the origin in the Coulomb approximation. As mentioned above, this analysis may be enormously simplified by the use of DR. Proceeding along the lines of Ref. [@KniPen2], we thus recover with ease the well-known $\alpha^2\ln(1/\alpha)$ terms in Eqs. (\[spar\]) and (\[sort\]) [@KhrYel; @CasLep1], $$\Delta^\prime\psi_1^2={\alpha^2}\ln{1\over\alpha} \left[2-{7\over 6}S(S+1)\right], \label{wfcorp}$$ where $S$ is the eigenvalue of the total-spin operator ${\bf S}$. As mentioned above, $\Delta^\prime\psi_1^2$ interferes with the one-loop hard renormalization of the Born amplitudes, $A_p$ and $A_o$, to produce $\alpha^3\ln(1/\alpha)$ terms. The resulting contributions to the coefficients $C_p$ and $C_o$ read $2A_p$ and $-A_o/3$, respectively. The generic logarithmic $O(\alpha^3)$ correction $\Delta^{\prime\prime}\psi_1^2$ to $\|\psi_1(0)\|^2$ is generated by the following one-loop operators, given in the momentum representation with $O(\epsilon)$ accuracy, $$\begin{aligned} \Delta_{\rm h}{\cal H}&=&-{1\over3}\,{\alpha^2\over m_e^2} \left[{1\over\hat\epsilon}\left({\mu^2\over m_e^2}\right)^{\epsilon} +{39\over5}-12\ln2 \right.\nonumber\\ &&{}+\left. \left({32\over3}+6\ln2\right){\bf S}^2\right], \label{hamhard}\\ \Delta_{\rm s}{\cal H}&=&-{7\over3}\,{\alpha^2\over m_e^2} \left[{1\over\hat\epsilon}\left({\mu^2\over{\bf q}^2}\right)^{\epsilon} -{1\over7}\right], \label{hamsoft}\\ \Delta_{\rm us}{\cal H}&=&{8\over3}\,{\alpha^2\over m_e^2} \left[{1\over\hat\epsilon}\left({\mu\over{\bf p}^2/m_e-E_1^C} \right)^{2\epsilon}+{5\over3}-2\ln2\right], \label{hamus}\end{aligned}$$ where $1/\hat\epsilon=1/\epsilon-\gamma_E+\ln(4\pi)$, with $\gamma_E$ being Euler’s constant, $\mu$ is the ’t Hooft mass scale of DR, ${\bf q}$ is the three-momentum transfer, and $E_1^C=-\alpha^2m_e/4$ is the Coulomb ground-state energy. Equations (\[hamhard\]) and (\[hamsoft\]) give the hard [@MPS] and soft [@PinSot2] $O(\alpha\beta^2)$ contributions to the NR Hamiltonian, respectively. The US contribution given in Eq. (\[hamus\]) arises from the emission and absorption of an US photon, which converts the on-shell Ps ground state into some off-shell state of the Coulomb spectrum, with three-momentum ${\bf p}$, before it decays. It is the only US contribution which can be represented by an operator of instantaneous interaction and thus give rise to logarithmic corrections. It has been found with the help of the method developed for the more complicated case of quantum chromodynamics in Ref. [@KniPen1], where it was applied to the on-shell renormalization of the heavy-quarkonium wave function at the origin. The singularities of the operators in Eqs. (\[hamhard\])–(\[hamus\]) yield the logarithmic corrections which we are interested in. Up to their logarithmic dependences on ${\bf q}^2$ and ${\bf p}^2$, these operators are of the $\delta$-function type in coordinate space and, therefore, lead to additional singularities in the Coulomb Green function at the origin [@KniPen2]. As a consequence, in the evaluation of the Green function in time-independent perturbation theory from Eq. (\[Sch\]) with $\Delta_{\rm h}{\cal H}$, $\Delta_{\rm s}{\cal H}$, and $\Delta_{\rm us}{\cal H}$, overlapping logarithmic divergences appear in the part of the first-order term which corresponds to the interference of the one-photon contribution to the Coulomb Green function and the first terms of Eqs. (\[hamhard\])–(\[hamus\]). This results in the double-logarithmic contribution, which can be directly extracted from the coefficient of the leading double-pole singularity [@KniPen2]. Since we are interested in the single-logarithmic contribution, we also have to keep the subleading terms in this analysis. The logarithmic corrections which are generated by the non-overlapping singularities can be obtained by putting $\mu=m_e$ in the Coulomb Green function at the origin and ${\bf q}^2={\bf p}^2=-m_eE_1^C$ in Eqs. (\[hamhard\])–(\[hamus\]), and proceeding as in the evaluation of Eq. (\[wfcorp\]). We thus obtain $$\begin{aligned} \Delta^{\prime\prime}\psi_1^2&=&{\alpha^3\over\pi} \left\{-{3\over2}\ln^2{1\over\alpha}+\left[-{184\over45}+{2\over3}\ln2 \right.\right.\nonumber\\ &&{}+\left.\left. \left({16\over9}+\ln2\right)S(S+1)\right]\ln{1\over\alpha}\right\}.\end{aligned}$$ The first term herein agrees with the corresponding terms in Eqs. (\[spar\]) and (\[sort\]) [@Kar], while the second one represents a new result. Another source of $\alpha^3\ln(1/\alpha)$ terms is the $O(\alpha\beta^2)$ corrections to the leading four-fermion operators. Since they do not involve the singular Coulomb Green function at the origin, there are no overlapping divergences, and we may simply read off the resulting $\alpha^3\ln(1/\alpha)$ terms from the poles of their US parts, which are given by the operator $$-{1\over\epsilon}\,{2\alpha\over3\pi}\, {{\bf p}^2+{\bf p^\prime}^2\over m_e^2}V_4({\bf p},{\bf p^\prime},{\bf S}). \label{power}$$ Here $V_4({\bf p},{\bf p^\prime},{\bf S})$ is the local four-fermion operator which generates the leading-order decay widths. Taking the expectation value of Eq. (\[power\]) w.r.t. the ground-state wave function, one encounters power-divergent integrals [@LLM]. They can be consistently treated within DR [@CMY2]. This leads to the substitution ${\bf p}^2,{\bf p^\prime}^2\to m_eE_1^C$ in the matrix element. The UV-pole contribution of Eq. (\[power\]) is then canceled by the IR pole of the hard contribution [@Kur]. This implies that the logarithmic integration ranges from the US scale $\alpha^2m_e$ up to the hard scale $m_e$, so that $1/\epsilon$ should be replaced by $4\ln(1/\alpha)$ [@KniPen2]. The resulting $\alpha^3\ln(1/\alpha)$ corrections to the decay widths are spin-independent and read $$\Delta\Gamma_{p,o}=\Gamma^{(0)}_{p,o}{4\alpha^3\over3\pi}\ln{1\over\alpha}.$$ The last source of $\alpha^3\ln(1/\alpha)$ terms is the $O(\alpha^2\beta)$ corrections to the leading four-fermion operators. These corrections are non-analytic in ${\bf p}^2$ and of the form $$\left[-{7\over 6}+\left(8-{32\over3}\ln2\right)\right]{\alpha^2\over4}\, {\|{\bf p}\|+\|{\bf p^\prime}\|\over m_e}V_4({\bf p},{\bf p^\prime},{\bf S}), \label{powernonan}$$ where the first and second terms contained within the square brackets are the soft and US contributions, respectively. Although the coefficient of $V_4({\bf p},{\bf p^\prime},{\bf S})$ in Eq. (\[powernonan\]) is finite, the matrix element of Eq. (\[powernonan\]) between Ps bound-state wave functions is logarithmically divergent in the UV region. In DR, the divergent part is given by the matrix element of Eq. (\[powernonan\]) with $\|{\bf p}\|,\|{\bf p^\prime}\|$ replaced by $m_e\alpha/(\pi\epsilon)$. In contrast to Eq. (\[power\]), the logarithmic integration now ranges from the soft scale up to the hard one, and $1/\epsilon$ should be replaced by $2\ln(1/\alpha)$. The resulting contributions to the decay widths read $$\Delta\Gamma_{p,o}=\Gamma^{(0)}_{p,o}{\alpha^3\over\pi}\ln{1\over\alpha} \left({41\over 6}-{32\over 3}\ln2\right).$$ Summing up the various $\alpha^3\ln(1/\alpha)$ terms derived above, we obtain $$\begin{aligned} C_p&=&2A_p+{367\over90}-10\ln2\approx-7.919,\nonumber\\ C_o&=&-{A_o\over 3}+{229\over 30}-8\ln2\approx5.517.\end{aligned}$$ In the case of p-Ps, the new $\alpha^3\ln(1/\alpha)$ term has the same sign as the $\alpha^3\ln^2(1/\alpha)$ one and exceeds the latter in magnitude. The sum of these two terms compensates approximately 1/3 of the positive contribution from the nonlogarithmic $O(\alpha^2)$ term. As for o-Ps, the new $\alpha^3\ln(1/\alpha)$ term cancels approximately 3/4 of the $\alpha^3\ln^2(1/\alpha)$ contribution. Our final predictions for the p-Ps and o-Ps total decay widths, including the multi-photon channels, read $$\begin{aligned} \Gamma_p^{\rm th}&=&7989.620(13)~\mu{\rm s}^{-1}, \label{finpar}\\ \Gamma_o^{\rm th}&=&7.039968(10)~\mu{\rm s}^{-1}, \label{finort}\end{aligned}$$ which has to be compared with Eqs. (\[thpar\]) and (\[thort\]). As before, Eq. (\[finpar\]) agrees with the Ann Arbor [@Annp] measurement (\[exannp\]), and Eq. (\[finort\]) favours the Tokyo [@Tok] measurement (\[extoko\]), while it significantly undershoots the Ann Arbor [@Ann1o; @Ann2o] measurements (\[exanno\]). The missing nonlogarithmic $O(\alpha^3)$ corrections in Eqs. (\[spar\]) and (\[sort\]) receive contributions from three-loop QED diagrams with a considerable number of external lines, which are far beyond the reach of presently available computational techniques. In this sense, we expect Eqs. (\[finpar\]) and (\[finort\]) to remain the best predictions for the forseeable future. However, we may speculate about the magnitudes of the coefficients $D_p$ and $D_o$. Two powers of $\alpha$ in these terms can be of NR origin. Each of them should be accompanied by the characteristic factor $\pi$, which happens for the logarithmic terms. Thus, we estimate the coefficients $D_p$ and $D_o$ to be a few units times $\pi^2$. This rule of thumb is in reasonable agreement with the situation at $O(\alpha^2)$, where we have $B_p\approx\pi^2/2$ and $B_o\approx4\pi^2$. If the coefficients $D_p$ and $D_o$ do not have magnitudes in excess of 100, then the uncertainties due the lack of their knowledge falls within the errors quoted in Eqs. (\[finpar\]) and (\[finort\]). Then, our new results reduce the uncertainties in the predicted p-Ps decay width to $10^{-2}\mu{\rm s}^{-1}$. The main remaining theoretical uncertainty in o-Ps decay width is related to the unknown $O(\alpha^2)$ contribution of the light-by-light scattering diagrams which can be estimated as a few units times $10^{-5}\mu{\rm s}^{-1}$ on the basis of p-Ps result. Calculation of this contribution along with our present result will reduce the uncertainties in the predicted o-Ps decay width to $10^{-5}\mu{\rm s}^{-1}$. Further progress in our understanding the Ps lifetime problem crucially depends also on the reduction of the experimental errors, which now greatly exceed the theoretical ones. Finally, we note that the technique developed in this letter can also be applied to the calculation of the subleading logarithmic $\alpha^7\ln(1/\alpha)$ terms for the Ps hyperfine splitting. This problem is of special interest because of the apparent discrepancy between the latest experimental data [@HFS] and the best theoretical predictions, which include the $O(\alpha^6)$ corrections (see Ref. [@CMY1] and references cited therein) and the leading logarithmic $\alpha^7\ln^2\alpha$ term [@Kar]. We are indebted to R. Hill and G. P. Lepage for useful discussions and to K. Melnikov and A. Yelkhovsky for pointing out the relevance of the $O(\alpha^2\beta)$ contribution from Eq. (\[powernonan\]). This work was supported in part by the Deutsche Forschungsgemeinschaft under Contract No. KN 365/1-1, by the Bundesministerium für Bildung und Forschung under Contract No. 05 HT9GUA 3, and by the European Commission through the Research Training Network [Quantum Chromodynamics and the Deep Structure of Elementary Particles]{} under Contract No. ERBFMRXCT980194. The work of A.A.P. was supported in part by the Volkswagen Foundation under Contract No. I/73611. [Note added]{} After including the $O(\alpha^2\beta)$ contribution from Eq. (\[powernonan\]), which was missed in the previous version of this letter, we find agreement with the analytical results for the p-Ps and o-Ps decay widths of Ref. [@MY] and with the numerical result for the o-Ps decay width of Ref. [@HL]. However, this correction is of insignificant size and immaterial for our numerical estimates of the Ps decay widths. [99]{} I. Harris and L. M. Brown, Phys. Rev. [105]{}, 1656 (1957). W. E. Caswell, G. P. Lepage, and J. Sapirstein, Phys. Rev. Lett. [38]{}, 488 (1977); G. S. Adkins, Phys. Rev. Lett. [76]{}, 4903 (1996). I. B. Khriplovich and A. S. Yelkhovsky, Phys. Lett. B [246]{}, 520 (1990). W. E. Caswell and G. P. Lepage, Phys. Rev. A [20]{}, 36 (1979). A. Czarnecki, K. Melnikov, and A. Yelkhovsky, Phys. Rev. Lett. [83]{}, 1135 (1999); Erratum [ibid.]{}, to be published. G. S. Adkins, R. N. Fell, and J. Sapirstein, Phys. Rev. Lett. [84]{}, 5086 (2000). G. S. Adkins and F. R. Brown, Phys. Rev. A [28]{}, 1164 (1983); G. P. Lepage, P. B. Mackenzie, K. H. Streng, and P. M. Zerwas, Phys. Rev. A [28]{}, 3090 (1983). S. G. Karshenboim, Sov. Phys. JETP [76]{}, 541 (1993) \[Zh. Eksp. Teor. Fiz. [103]{}, 1105 (1993)\]. A. H. Al-Ramadhan and D. W. Gidley, Phys. Rev. Lett. [72]{}, 1632 (1994). C. I. Westbrook, D. W. Gidley, R. S. Conti, and A. Rich, Phys. Rev. A [40]{}, 5489 (1989). J. S. Nico, D. W. Gidley, A. Rich, and P. W. Zitzewitz, Phys. Rev. Lett. [65]{}, 1344 (1990). S. Asai, S. Orito, and N. Shinohara, Phys. Lett. B [357]{}, 475 (1995). A. Czarnecki and S. G. Karshenboim, Report No. BNL-HET-99/37 and hep-ph/9911410. M. Beneke and V. A. Smirnov, Nucl. Phys. [B522]{}, 321 (1998). W. E. Caswell and G. P. Lepage, Phys. Lett. [167B]{}, 437 (1986). A. Pineda and J. Soto, Nucl. Phys. Proc. Suppl. [64]{}, 428 (1998). A. Pineda and J. Soto, Phys. Lett. B [420]{}, 391 (1998); Phys. Rev. D [59]{}, 016005 (1999). A. Czarnecki, K. Melnikov, and A. Yelkhovsky, Phys. Rev. Lett. [82]{}, 311 (1999); Phys. Rev. A [59]{}, 4316 (1999). B. A. Kniehl and A. A. Penin, Nucl. Phys. [B577]{}, 197 (2000). M. Nio and T. Kinoshita, Phys. Rev. D [55]{}, 7267 (1997). P. Labelle, G. P. Lepage, and U. Magnea, Phys. Rev. Lett. [72]{}, 2006 (1994). A. V. Manohar, Phys. Rev. D [56]{}, 230 (1997); A. Pineda and J. Soto, Phys. Rev. D [58]{}, 114011 (1998). B. A. Kniehl and A. A. Penin, Nucl. Phys. [B563]{}, 200 (1999). E. A. Kuraev, T. V. Kukhto, and Z. K. Silagadze, Sov. J. Nucl. Phys. [51]{}, 1036 (1990) \[Yad. Fiz. [51]{}, 1638 (1990)\]. A. P. Mills, Jr., Phys. Rev. A [27]{}, 262 (1983); M. W. Ritter, P. O. Egan, V. W. Hughes, and K. A. Woodle, Phys. Rev. A [30]{}, 1331 (1984). K. Melnikov and A. Yelkhovsky, Report No. SLAC-PUB-8557 and hep-ph/0008099. R. Hill and G. P. Lepage, Report No. hep-ph/0003277 v2 (May 2000). [^1]: Permanent address: Institute for Nuclear Research, Russian Academy of Sciences, 60th October Anniversary Prospect 7a, Moscow 117312, Russia.	{ "pile_set_name": "ArXiv" }
--- abstract: 'In this paper, we investigate exceptional sets in the Waring-Goldbach problem for unlike powers. For example, estimates are obtained for sufficiently large integers below a parameter subject to the necessary local conditions that do not have a representation as the sum of a square of prime, a cube of prime and a sixth power of prime and a $k$-th power of prime. These results improve the recent result due to Brüdern in the order of magnitude. Furthermore, the method can be also applied to the similar estimates for the exceptional sets for Waring-Goldbach problem for unlike powers.' address: - 'School of Mathematics and Statistics, Minnan Normal University, Zhangzhou 363000, P. R. China' - 'School of Mathematics, Jilin University, Changchun 130012, P. R. China' author: - Zhenzhen Feng - 'Jing Ma\' title: 'Waring-Goldbach problem for unlike powers' --- [^1] Introduction ============ Let $N$, $k_1$, $k_2$, $\cdots$, $k_r$ be natural numbers such that $2 \leq k_1\leq k_2 \leq \cdots \leq k_r$. The Waring-Goldbach problem for unlike powers concerns the representation of $N$ as the form $$N=p_1^{k_1}+p_2^{k_2}+\cdots+p_r^{k_r}.$$ Not very much is known about results of this kind. However, these topics have attracted mathematicians’ attentions. Schwarz [@Schwarz] considered the exceptional set of expressing an positive even number as the sum of a square of prime, a cube of prime, a sixth power of prime and a $k$-th power of prime, i.e. $$\begin{aligned} \label{236k} n=p_1^2+p_2^3+p_3^6+p_4^k,\end{aligned}$$ where $p_1$, $p_2$, $p_3$, $p_4$ are primes. Let $E_1(k, N)$ be the number of positive even integers $n$ up to $N$ which can not be written in the form (\[236k\]). Exactly, Schwarz [@Schwarz] showed that $E_1(k, N)\ll N(\log N)^{-A}$ for any fixed $A>0$. Recently, Brüdern [@Brudern] improved this result and established that $E_1(k, N)\ll N^{1-\frac{1}{8k^2}+\varepsilon}$. In this paper, we further improve the result of Brüdern by giving \[Thm236k\] Let $E_1(k, N)$ be defined as above. We have $$E_1(k, N)\ll N^{1-\theta_1(k)+\varepsilon},$$ among which, $$\begin{aligned} \theta_1(k)= \begin{cases} \frac{1}{54}, & k=6,\\ \frac{1}{81}, & k=7,\\ \frac{1}{54x}, & k\ge 8, \end{cases}\end{aligned}$$ where $$\begin{aligned} \label{236kx} x= \begin{cases} \left\lceil(\frac{k}{6}+1-[\frac{k}{6}])2^{[\frac{k}{6}]-1}\right\rceil, & 8\le k\le 23,\\ \left\lceil \frac{7k}{6}-20 \right\rceil, & 24\le k\le 29,\\ \left\lceil (\frac{k}{6}-\frac{1}{2}[\frac{k}{6}])([\frac{k}{6}]+1)\right\rceil, & k\ge 30. \end{cases}\end{aligned}$$ Here $\lceil a \rceil$ means the smallest integer no smaller than $a$ and $[a]$ means the biggest integer no bigger than $a$. Remark:* We can compare the results of Theorem 1.1 with those of Brüdern [@Brudern]. For example, we obtain $E_1(6, N)\ll N^{1-\frac{1}{54}+\varepsilon}$ and $E_1(7, N)\ll N^{1-\frac{1}{81}+\varepsilon}$. Meanwhile, Brüdern’s results indicated that $E_1(6, N)\ll N^{1-\frac{1}{288}+\varepsilon}$ and $E_1(7, N)\ll N^{1-\frac{1}{392}+\varepsilon}$. In addition, for large value $k$, Theorem 1.1 gives that $E_1(k, N)\ll N^{1-\frac{1}{\frac{3}{4}k^2+O(k)}+\varepsilon}$, whereas Brüdern’s result [@Brudern] showed that $E_1(k, N)\ll N^{1-\frac{1}{8k^2}+\varepsilon}$. In the same paper [@Schwarz], Schwarz also considered the problem of representing a large even integer $n$ in the form $$\begin{aligned} \label{244k} n=p_1^2+p_2^4+p_3^4+p_4^k,\end{aligned}$$ where $p_1, p_2, p_3, p_4$ are primes. Let $E_2(k, N)$ denote the number of positive even integers $n$ up to $N$ which can not be written in the form (\[244k\]). In fact, Schwarz [@Schwarz] proved that $E_2(k, N)\ll N(\log N)^{-A}$ for any fixed $A>0$. Using the similar method to treat Theorem \[Thm236k\], we obtain the following result: \[Thm244k\] Let $E_2(k, N)$ be defined as above. We have $$E_2(k, N)\ll N^{1-\theta_2(k)+\varepsilon},$$ here $$\begin{aligned} \theta_2(k)= \begin{cases} \frac{1}{32}, & k=4, \\ \frac{1}{48}, & k=5, \\ \frac{1}{64}, & 6\le k\le 8, \\ \frac{1}{48x}, & k\ge 9, \end{cases}\end{aligned}$$ where $$\begin{aligned} x= \begin{cases} \left\lceil (\frac{k}{4}+1-[\frac{k}{4}])2^{[\frac{k}{4}]-1}\right\rceil, & 9\le k\le 19,\\ \left\lceil (\frac{k}{4}-\frac{1}{2}[\frac{k}{4}])([\frac{k}{4}]+1)\right\rceil, & k\ge 20. \end{cases}\end{aligned}$$ Remark: For example, we obtain that $E_2(4, N)\ll N^{1-\frac{1}{32}+\varepsilon}$ and $E_2(6, N)\ll N^{1-\frac{1}{64}+\varepsilon}$. Meanwhile, Brüdern’s method in [@Brudern] indicated that $E_2(4, N)\ll N^{1-\frac{1}{128}+\varepsilon}$ and $E_2(6, N)\ll N^{1-\frac{1}{288}+\varepsilon}$. In addition, for large value $k$, Theorem \[Thm244k\] gives that $E_2(k, N)\ll N^{1-\frac{1}{\frac{3}{2}k^2+O(k)}+\varepsilon}$, whereas Brüdern’s method in [@Brudern] showed that $E_2(k, N)\ll N^{1-\frac{1}{8k^2}+\varepsilon}$. Another related problem is to study for the diophantine equation $$\begin{aligned} \label{235k} n=p_1^2+p_2^3+p_3^5+p_4^k\end{aligned}$$ where $p_1$, $p_2$, $p_3$, $p_4$ are primes. Let $E_3(k, N)$ be the number of even integers $n\le N$ that can not be represented in the form (\[235k\]). In 1953, Prachar [@Prachar] proved that $E_3(4, N)\ll N(\log N)^{-\frac{30}{47}+\varepsilon}$. This has been improved by a number of authors (c. f.[@Bauer1; @Bauer2; @RenTsang1; @RenTsang2]). The latest result is $$\begin{aligned} E_3(4, N)\ll N^{1-\frac{1}{16}+\varepsilon}\end{aligned}$$ given by Zhao [@Zhao3]. For general $k\ge 5$, Lu and Shan [@LuShan] proved that $E_3(k, N)\ll N(\log N)^{-c}$ for some $c>0$. Lately, it was improved to $E_3(k, N)\ll N^{1-\frac{1}{3k\times 2^{k-2}}+\varepsilon}$ by Liu [@LiuZ]. The current best result was given by Hoffman and Yu [@HoffmanYu] which is $$\begin{aligned} \label{E3HY} E_3(k, N)\ll N^{1-\frac{47}{420\cdot 2^s}+\varepsilon}\end{aligned}$$ where $s=[\frac{k+1}{2}]$. In this paper, we established the following result which improves (\[E3HY\]). \[Thm235k\] Let $E_3(k, N)$ be defined as above. We have $$E_3(k, N)\ll N^{1-\theta_3(k)+\varepsilon},$$ here $$\begin{aligned} \theta_3(k)= \begin{cases} \frac{1}{24}, & k=5, \\ \frac{2}{81}, & k=6, \\ \frac{1}{36x}, & k\ge 7, \end{cases}\end{aligned}$$ where $$\begin{aligned} x= \begin{cases} \left\lceil (\frac{k}{6}+1-[\frac{k}{6}])2^{[\frac{k}{6}]-1}\right\rceil, & 7\le k\le 23, \\ \left\lceil \frac{7k}{6}-20 \right\rceil, & 24\le k\le 29, \\ \left\lceil (\frac{k}{6}-\frac{1}{2}[\frac{k}{6}])([\frac{k}{6}]+1)\right\rceil, & k\ge 30. \end{cases}\end{aligned}$$ Remark: Our results indeed improve the result of Hoffman and Yu [@HoffmanYu]. For example, we obtain that $E_3(5, N)\ll N^{1-\frac{1}{24}+\varepsilon}$ and $E_3(7, N)\ll N^{1-\frac{1}{72}+\varepsilon}$. Meanwhile, Hoffman and Yu’s results indicated that $E_3(5, N)\ll N^{1-\frac{47}{3360}+\varepsilon}$ and $E_3(7, N)\ll N^{1-\frac{47}{6720}+\varepsilon}$. In addition, for large value $k$, $E_3(k, N)\ll N^{1-\frac{1}{\theta(k)}+\varepsilon}$, Hoffman and Yu [@HoffmanYu] showed that $\theta(k)$ grows exponentially, whereas, Theorem \[Thm235k\] implicates that $\theta(k)=\frac{k^2}{2}+O(k)$ with polynomial growth. Finally, we consider the problem of representing a large odd integer $n$ in the form $$\begin{aligned} \label{3333k} n=p_1^3+p_2^3+p_3^3+p_4^3+p_5^k,\end{aligned}$$ where $p_1$, $p_2$, $p_3$, $p_4$ and $p_5$ are primes. Let $E_4(k, N)$ denote the number of positive odd integers $n$ up to $N$ which can not be written in the form (\[3333k\]). In the following result, we will give a up bound for $E_4(k, N)$ for $k\ge 4$. \[Thm3333k\] Let $E_4(k, N)$ be defined as above. We have $$E_4(k, N)\ll N^{1-\theta_4(k)+\varepsilon},$$ here $$\begin{aligned} \theta_4(k)= \begin{cases} \frac{1}{24}, & k=4, \\ \frac{1}{54}, & k=5, \\ \frac{1}{9x}, & k\ge 6, \end{cases}\end{aligned}$$ where $$\begin{aligned} x=\begin{cases} \left\lceil \frac{14k}{3}-20 \right\rceil, & k=6,7, \\ \left\lceil (\frac{2k}{3}-\frac{1}{2}[\frac{2k}{3}])([\frac{2k}{3}]+1)\right\rceil, & k\ge 8. \end{cases}\end{aligned}$$ As usual, we abbreviate $e^{2\pi i\alpha}$ to $e(\alpha)$. The letter $p$, with or without indices, is prime number. The letter $\varepsilon$ denotes a sufficiently small positive real number, and the value of $\varepsilon$ may change from statement to statement. Let $N$ be a real number sufficiently large in terms of $\varepsilon$ and $k$. We use $\ll$ and $\gg$ to denote Vinogradov’s well-know notation, while implied constant may depend on $\varepsilon$ and $k$. Preliminaries and lemmas ======================== We will prove Theorems 1.1-1.4 by using circle method. Now the treatment for major arcs of Hardy-Littlewood method are standard nowadays, for example Liu and Zhan [@LiuZhan]. We need the following lemmas to control the minor arcs of circle method. \[LemHuaBou\] Let $$S_k(\alpha)=\sum_{N/4<p^k\le N}(\log p)e(\alpha p^k).$$ Then for $1\le j\le k$, we have $$\int_0^1\big\|S_k^{2^j}(\alpha)\big\|d\alpha\ll N^{\frac{1}{k}(2^j-j)+\varepsilon}$$ and $$\int_0^1\big\|S_k^{j(j+1)}(\alpha)\big\|d\alpha\ll N^{\frac{j^2}{k}+\varepsilon}.$$ In fact, Lemma \[LemHuaBou\] is the classical result of Hua [@Hua] and the recent work of Bourgain [@Bourgain]. The next lemma is a generalization of Lemma [\[LemHuaBou\]]{}. \[lemmain\] Let $S_k(\alpha)$ be defined as Lemma \[LemHuaBou\]. For $0<\delta\le 1$, $$\begin{aligned} \int_0^1\big\|S_k^{2x}(\alpha)\big\|d\alpha\ll N^{\frac{2x}{k}-\delta+\varepsilon},\end{aligned}$$ where $$\begin{aligned} \label{mainx} x= \begin{cases} \left\lceil (k\delta+1-[k\delta])2^{[k\delta]-1}\right\rceil, & [k\delta]\le 3, \\ \left\lceil 7k\delta-20 \right\rceil, & [k\delta]=4, \\ \left\lceil (k\delta-\frac{1}{2}[k\delta])([k\delta]+1)\right\rceil, & [k\delta]\ge 5. \end{cases}\end{aligned}$$ For $\delta=1$, this is Lemma \[LemHuaBou\]. Next, we consider the case $0<\delta<1$:\ for $[k\delta]\le 3$, clearly by (\[mainx\]) we have $$2^{[k\delta]}\le 2x\le 2^{[k\delta]+1}.$$ Applying Hölder’s inequality and Hua’s lemma, one has $$\begin{aligned} \int_0^1\big\|S_k^{2x}(\alpha)\big\|d\alpha\ll & \Big(\int_0^1\big\|S_k^{2^{[k\delta]}}(\alpha)\big\|d\alpha\Big)^a \Big(\int_0^1\big\|S_k^{2^{[k\delta]+1}}(\alpha)\big\|d\alpha\Big)^b \notag \\ \ll & N^{\frac{2x}{k}-c+\varepsilon},\end{aligned}$$ where $$a=2-\frac{x}{2^{[k\delta]-1}}, \quad b=\frac{x}{2^{[k\delta]-1}}-1, \quad c=\frac{[k\delta]+\frac{2x}{2^{[k\delta]}}-1}{k}.$$ Recall that $x\ge (k\delta+1-[k\delta])2^{[k\delta]-1}$, so we have $c\ge \delta$. Thus this lemma holds for $[k\delta]\le 3$. For $[k\delta]=4$, obviously by (\[mainx\]) we have $$16 <2x\le 30.$$ Applying Hölder’s inequality and Lemma \[LemHuaBou\], one has $$\begin{aligned} \int_0^1\big\|S_k^{2x}(\alpha)\big\|d\alpha\ll & \Big(\int_0^1\big\|S_k^{16}(\alpha)\big\|d\alpha\Big)^{\frac{15}{7}-\frac{x}{7}} \Big(\int_0^1\big\|S_k^{30}(\alpha)\big\|d\alpha\Big)^{\frac{x}{7}-\frac{8}{7}} \notag \\ \ll & N^{\frac{2x}{k}-\frac{x+20}{7k}+\varepsilon}.\end{aligned}$$ This combining with $x\ge 7k\delta-20$ gives $\frac{x+20}{7k}\ge \delta.$ Thus this lemma holds for $[k\delta]=4$. For $[k\delta]\ge 5$, by (\[mainx\]) we have $$[k\delta]([k\delta]+1)\le 2x\le ([k\delta]+1)([k\delta]+2).$$ Applying Hölder’s inequality and Lemma \[LemHuaBou\], one has $$\begin{aligned} \int_0^1\big\|S_k^{2x}(\alpha)\big\|d\alpha\ll & \Big(\int_0^1\big\|S_k^{[k\delta]([k\delta]+1)}(\alpha)\big\|d\alpha\Big)^{a} \Big(\int_0^1\big\|S_k^{([k\delta]+1)([k\delta]+2)}(\alpha)\big\|d\alpha\Big)^{b} \notag \\ \ll & N^{\frac{2x}{k}-c+\varepsilon},\end{aligned}$$ where $$a=1-\frac{x}{[k\delta]+1}+\frac{[k\delta]}{2}, \quad b=\frac{x}{[k\delta]+1}-\frac{[k\delta]}{2},\quad c=\frac{\frac{x}{[k\delta]+1}+\frac{[k\delta]}{2}}{k}.$$ Then we have $c\ge \delta$ because of $x\ge (k\delta-\frac{[k\delta]}{2})([k\delta]+1)$. This lemma holds for $[k\delta]\ge 5$. Hence, this lemma holds for $0<\delta\le 1$. \[lemS2x\] For $k\ge 3$, we have $$\begin{aligned} \int_0^1\|S_2^2(\alpha)S_k^{2x}(\alpha)\|d\alpha\ll N^{\frac{2x}{k}+\varepsilon},\end{aligned}$$ where $$\begin{aligned} x= \begin{cases} \left\lceil (\frac{k}{2}+1-[\frac{k}{2}])2^{[\frac{k}{2}]-1}\right\rceil, & 3\le k\le 9, \\ \left\lceil (\frac{k}{2}-\frac{1}{2}[\frac{k}{2}])([\frac{k}{2}]+1)\right\rceil, & k\ge 10. \end{cases}\end{aligned}$$ $\int_0^1\|S_2^2(\alpha)S_k^{2x}(\alpha)\|d\alpha$ is no more than $N^{\varepsilon}$ times the number of solutions of the equation $$t_1^2-t_2^2=y_1^k+y_2^k+\cdot\cdot\cdot+y_x^k-y_{x+1}^k-\cdot\cdot\cdot-y_{2x}^k$$ with $N^{\frac{1}{2}}<t_1,t_2\le 2N^{\frac{1}{2}}$ and $N^{\frac{1}{k}}<y_1,y_2,\cdot\cdot\cdot,y_{2x}\le 2N^{\frac{1}{k}}$. If $t_1\neq t_2$, the contribution is bounded by $N^{\frac{2x}{k}+\varepsilon}$. If $t_1=t_2$, the contribution is bounded by $N^{\frac{1}{2}+\varepsilon}\int_0^1\|S_k^{2x}\|d\alpha$. Thus $$\int_0^1\|S_2^2(\alpha)S_k^{2x}(\alpha)\|d\alpha\ll N^{\frac{2x}{k}+\varepsilon}+N^{\frac{1}{2}+\varepsilon}\int_0^1\|S_k^{2x}(\alpha)\|d\alpha.$$ What we need is $$\int_0^1\|S_k^{2x}(\alpha)\|d\alpha\ll N^{\frac{2x}{k}-\frac{1}{2}+\varepsilon}.$$ Hence this lemma holds by Lemma \[lemmain\] with $\delta=\frac{1}{2}$. \[lemS24x\] For $k\ge 4$, we have $$\begin{aligned} \int_0^1\|S_2^2(\alpha)S_4^2(\alpha)S_k^{2x}(\alpha)\|d\alpha\ll N^{\frac{2x}{k}+\frac{1}{2}+\varepsilon},\end{aligned}$$ where $$\begin{aligned} x=\begin{cases} \left\lceil (\frac{k}{4}+1-[\frac{k}{4}])2^{[\frac{k}{4}]-1}\right\rceil, & 4\le k\le 19, \\ \left\lceil (\frac{k}{4}-\frac{1}{2}[\frac{k}{4}])([\frac{k}{4}]+1)\right\rceil, & k\ge 20. \end{cases}\end{aligned}$$ $\int_0^1\|S_2^2(\alpha)S_4^2(\alpha)S_k^{2x}(\alpha)\|d\alpha$ is no more than $N^{\varepsilon}$ times the number of solutions for the equation $$t_1^2-t_2^2=y_1^4-y_2^4+z_1^k+z_2^k+\cdot\cdot\cdot+z_x^k-z_{x+1}^k-\cdot\cdot\cdot-z_{2x}^k$$ with $P_2< t_1,t_2\le 2P_2$, $P_4< y_1,y_2\le 2P_4$ and $P_k< z_1,z_2,\cdot\cdot\cdot,z_{2x}\le 2P_k$, where $N/4<P_2^2, P_4^4, P_k^k\le N.$ If $t_1\neq t_2$, the contribution is bounded by $P_4^{2+\varepsilon}P_k^{2x}$. If $t_1=t_2, y_1\neq y_2$, the contribution is bounded by $P_2P_k^{2x+\varepsilon}$. If $t_1=t_2, y_1= y_2$, the contribution is bounded by $P_2^{1+\varepsilon}P_4\int_0^1\|S_k(\alpha)\|^{2x}d\alpha$. Thus $$\begin{aligned} \int_0^1\|S_2^2(\alpha)S_4^2(\alpha)S_k^{2x}(\alpha)\|d\alpha\ll N^{\varepsilon}P_4^{2+\varepsilon}P_k^{2x}+N^{\varepsilon}P_2P_4\int_0^1\|S_k(\alpha)\|^{2x}d\alpha.\end{aligned}$$ What we need is $$\int_0^1\|S_k^{2x}(\alpha)\|d\alpha\ll N^{\frac{2x}{k}-\frac{1}{4}+\varepsilon}.$$ Hence this lemma holds by Lemma \[lemmain\] with $\delta=\frac{1}{4}$. \[lemS23k\] For $k\ge 3$, we have $$\begin{aligned} \int_0^1\|S_2^2(\alpha)S_3^2(\alpha)S_k^{2x}(\alpha)\|d\alpha\ll N^{\frac{2x}{k}+\frac{2}{3}+\varepsilon},\end{aligned}$$ where $$\begin{aligned} x=\begin{cases} \left\lceil (\frac{k}{6}+1-[\frac{k}{6}])2^{[\frac{k}{6}]-1}\right\rceil, & 3\le k\le 23,\\ \left\lceil \frac{7k}{6}-20 \right\rceil, & 24\le k\le 29,\\ \left\lceil (\frac{k}{6}-\frac{1}{2}[\frac{k}{6}])([\frac{k}{6}]+1)\right\rceil, & k\ge 30. \end{cases}\end{aligned}$$ The proof is similar as the proof of Lemma \[lemS24x\] with $\delta=\frac{1}{6}$. \[8233\] For $k\ge 3$, we have $$\begin{aligned} \int_0^1\|S_3^4(\alpha)S_k^{2x}(\alpha)\|d\alpha\ll N^{\frac{2x}{k}+\frac{1}{3}+\varepsilon},\end{aligned}$$ where $$\begin{aligned} x=\begin{cases} \left\lceil (\frac{2k}{3}+1-[\frac{2k}{3}])2^{[\frac{2k}{3}]-1} \right\rceil, & 3\leq k \leq 5, \\ \left\lceil \frac{14k}{3}-20 \right\rceil, & k=6,7,\\ \left\lceil (\frac{2k}{3}-\frac{1}{2}[\frac{2k}{3}])([\frac{2k}{3}]+1)\right\rceil, & k\ge 8. \end{cases}\end{aligned}$$ We have $$\begin{aligned} \label{4z1} \int_0^1\|S_3^4(\alpha)S_k^{2x}(\alpha)\|d\alpha\ll N^{\varepsilon}\int_0^1\|f_3^4(\alpha)S_k^{2x}(\alpha)\|d\alpha,\end{aligned}$$ where $$f_3(\alpha)=\sum_{t\sim P_3}e(\alpha t^3).$$ By Lemma 2.3 in [@Vaughan], one has $$\begin{aligned} \big\|f_3(\alpha)\big\|^4\ll P_3\sum_{\|h_1\|< P_3}\sum_{\|h_2\|< P_3}\sum_{x\in \mathcal{J}}e(\alpha\Delta(t^3; {\bf h})),\end{aligned}$$ where $\mathcal{J}=\mathcal{J}({\bf h})$ is a subinterval of $[P_3, 2P_3)$ and $\Delta(t^3; {\bf h})$ is the second-order forward difference of the function $t\rightarrow t^3$ with steps $h_1, h_2$, that is, $$\Delta(t^3; {\bf h})=3h_1h_2(2t+h_1+h_2).$$ Thus, we deduce from (\[4z1\]) that $$\begin{aligned} \int_0^1\|S_3^4(\alpha)S_k^{2x}(\alpha)\|d\alpha\ll P_3J(P_3),\end{aligned}$$ where $J(P_3)$ is the number of solutions of the diophantine equation $$\begin{aligned} \label{4z2} \Delta(t^3; {\bf h})=3h_1h_2(2t+h_1+h_2)=p_1^k+p_2^k+\cdot\cdot\cdot+p_x^k-q_1^k-q_2^k-\cdot\cdot\cdot-q_x^k\end{aligned}$$ subject to $$\begin{aligned} \label{4z3} P_3\le t\le 2P_3, \quad \|h_i\|<P_3, \quad P_k<p_1,\cdot\cdot\cdot, p_x,q_1,\cdot\cdot\cdot, q_x\le 2P_k, \quad N/4<P_3^3, P_k^k\le N.\end{aligned}$$ The number of solutions of (\[4z2\]), (\[4z3\]) with $\Delta(t^3; {\bf h})=0$ is bounded by\ $P_3^{2+\varepsilon}\int_0^1\big\|S_k^{2x}(\alpha)\big\|d\alpha$. The number of solutions of (\[4z2\]), (\[4z3\]) with $\Delta(t^3; {\bf h})\neq 0$ is bounded by $N^{\frac{2x}{k}+\varepsilon}$.Thus, $$\begin{aligned} \int_0^1\|S_3^4(\alpha)S_k^{2x}(\alpha)\|d\alpha\ll N^{1+\varepsilon}\int_0^1\big\|S_k^{2x}(\alpha)\big\|d\alpha+N^{\frac{1}{3}+\frac{2x}{k}+\varepsilon}.\end{aligned}$$ Thus, we just need $$\int_0^1\|S_k^{2x}(\alpha)\|d\alpha\ll N^{\frac{2x}{k}-\frac{2}{3}+\varepsilon}.$$ Hence it establishes this lemma by Lemma \[lemmain\] with $\delta=\frac{2}{3}$. Proof of Theorem \[Thm236k\] ============================ The purpose of this section is to concentrate on proving Theorem \[Thm236k\]. We establish Theorem \[Thm236k\] by means of the Hardy-Littlewood method. We will give the proof of Theorem \[Thm244k\] in Sect. 4 and will describe the straight forward modifications needed for Theorem \[Thm235k\] in Sect. 5. In Sect. 6, we will give the outline of proof of Theorem \[Thm3333k\]. Let $S_k(\alpha)$ be defined as in Lemma \[LemHuaBou\]. We denote $$\begin{aligned} \label{r236k} r(n)=\sum_{\substack{p_1^2+p_2^3+p_3^6+p_4^k=n\\ N/4<p_1^2,p_2^3,p_3^6,p_4^k\le N}}(\log p_1)(\log p_2) (\log p_3)(\log p_4),\end{aligned}$$ where $p_1$, $p_2$, $p_3$, $p_4$ are primes. Let $Q=N^{\frac{2}{5k}}$, and write $\mathfrak{M}(Q)$ for the union of the intervals $$\{\alpha\in [0,1]: \|q\alpha-a\|\le QN^{-1}\}$$ with $1\le a\le q, (a,q)=1$ and $1\le q\le Q$. We define $ \mathfrak{M}=\mathfrak{M}(Q), \textrm{ }\ \mathfrak{m}=[0,1]\backslash\mathfrak{M}$. Thus the formula (\[r236k\]) becomes $$r(n)=\Big\{\int_{\mathfrak{M}}+\int_{\mathfrak{m}}\Big\}S_2(\alpha)S_3(\alpha)S_6(\alpha)S_k(\alpha)e(-n\alpha)d\alpha.$$ Whenever $\mathfrak{B}\subset [0,1]$ is measurable, we put $$r_{\mathfrak{B}}(n, N)=\int_{\mathfrak{B}}S_2(\alpha)S_3(\alpha)S_6(\alpha)S_k(\alpha)e(-n\alpha)d\alpha.$$ Then we have $$r(n)=r_{[0,1]}(n, N)=\int_ 0^1S_2(\alpha)S_3(\alpha)S_6(\alpha)S_k(\alpha)e(-n\alpha)d\alpha.$$ Next we will deal with the integral of major arcs and minor arcs respectively. Applying the now standard methods of enlarging major arcs (c.f. [@LiuZhan]), we can get the following result: For all even integer $n$ with $N<n \le 2N$, one has $r_{\mathfrak{M}}(n, N)\gg N^{\frac{1}{k}}$. To estimate the integral of the minor arcs, we split the minor arcs in two part. Let $1\le Y\le N^{\frac{1}{8}}$, and denote $\mathfrak{N}$ the union of the pairwise disjoint intervals $$\mathfrak{N}_{q,a}(Y)=\{\alpha\in[0,1]: \|q\alpha-a\|\le Y/N\}$$ with $1\le a\le q, (a,q)=1$ and $1\le q\le Y$. We write $\mathfrak{N}=\mathfrak{N}(N^{\frac{1}{8}})$ and $\mathfrak{n}=\mathfrak{m}\backslash \mathfrak{N}.$ \[nS2S3S5\] For $\alpha\in \mathfrak{n}$, we have $$S_2(\alpha)\ll N^{\frac{1}{2}-\frac{1}{16}+\varepsilon}, \quad S_3(\alpha)\ll N^{\frac{1}{3}-\frac{1}{36}+\varepsilon}, \quad S_4(\alpha)\ll N^{\frac{1}{4}-\frac{1}{96}+\varepsilon}.$$ For any given $\alpha\in \mathbb{N}$, by Dirichlet’s approximation theorem, there exist $a\in \mathbb{Z}$ and $q\in \mathbb{N}$ with $$(a,q)=1, 1\le q\le N^{\frac{5}{12}} \textrm{ and }\ \|q\alpha-a\|\le N^{-\frac{5}{12}}.$$ Then by Theorem 1 in [@Kumchev], one has $$S_2(\alpha)\ll N^{\frac{1}{2}-\frac{1}{16}+\varepsilon}+\frac{N^{\frac{1}{2}+\varepsilon}}{(q+N\|q\alpha-a\|)^{\frac{1}{2}}},$$ and $$S_4(\alpha)\ll N^{\frac{1}{4}-\frac{1}{96}+\varepsilon}+\frac{N^{\frac{1}{4}+\varepsilon}}{(q+N\|q\alpha-a\|)^{\frac{1}{2}}},$$ and by Lemma 2.3 in [@Zhao2], one has $$S_3(\alpha)\ll N^{\frac{1}{3}-\frac{1}{36}+\varepsilon}+\frac{N^{\frac{1}{3}+\varepsilon}}{(q+N\|q\alpha-a\|)^{\frac{1}{2}}}.$$ If $\alpha\in \mathfrak{n}$, then $$q> N^{\frac{1}{8}} \textrm{ or }\ q\le N^{\frac{1}{8}}, \textrm{ }\ N^{-\frac{7}{8}}\le \|q\alpha-a\|<N^{-\frac{5}{12}}.$$ In any case, we have $$q+\|q\alpha-a\|\gg N^{\frac{1}{8}},$$ then this lemma clearly holds. Proof of Theorem 1.1. By Bessel’s inequality, we have $$\begin{aligned} \sum_{N< n \le 2N} & \Big\|\int_{\mathfrak{m}}S_2(\alpha)S_3(\alpha)S_6(\alpha)S_k(\alpha)e(-n\alpha)d\alpha\Big\|^2 \le \int_{\mathfrak{m}}\big\|S_2^2(\alpha)S_3^2(\alpha)S_6^2(\alpha)S_k^2(\alpha)\big\|d\alpha.\end{aligned}$$ To prove Theorem \[Thm236k\], it suffices to show that $$\begin{aligned} \label{thm1236k} \int_{\mathfrak{m}}\big\|S_2^2(\alpha)S_3^2(\alpha)S_6^2(\alpha)S_k^2(\alpha)\big\|d\alpha\ll N^{1+\frac{2}{k}-\theta_1(k)+\varepsilon},\end{aligned}$$ where $\theta_1(k)$ is defined in Theorem \[Thm236k\]. Obviously, we know that $$\begin{aligned} \int_{\mathfrak{m}}\big\|S_2^2(\alpha)S_3^2(\alpha)S_6^2(\alpha)S_k^2(\alpha)\big\|d\alpha\ll & \int_{\mathfrak{N}\backslash\mathfrak{M}}\big\|S_2^2(\alpha)S_3^2(\alpha)S_6^2(\alpha)S_k^2(\alpha)\big\|d\alpha \\ &+ \int_{\mathfrak{n}}\big\|S_2^2(\alpha)S_3^2(\alpha)S_6^2(\alpha)S_k^2(\alpha)\big\|d\alpha.\end{aligned}$$ By the estimate on Page 80 in [@Brudern], one has $$\begin{aligned} \label{thm1236km} \int_{\mathfrak{N}\backslash\mathfrak{M}}\big\|S_2^2(\alpha)S_3^2(\alpha)S_6^2(\alpha)S_k^2(\alpha)\big\|d\alpha\ll N^{1+\frac{2}{k}-\frac{1}{4k}+\varepsilon} \ll N^{1+\frac{2}{k}-\theta_1(k)+\varepsilon}\end{aligned}$$ since $\frac{1}{4k}>\theta_1(k)$ for all $k\ge 6$. Next we estimate $\int_{\mathfrak{n}}\big\|S_2^2(\alpha)S_3^2(\alpha)S_6^2(\alpha)S_k^2(\alpha)\big\|d\alpha$. For $k=6$, by Lemma \[lemS2x\] and Lemma \[nS2S3S5\], one has $$\begin{aligned} \label{thm1k6} \int_{\mathfrak{n}} & \big\|S_2^2(\alpha)S_3^2(\alpha)S_6^4(\alpha)\big\|d\alpha \notag \\ \ll & \sup_{\alpha\in\mathfrak{n}}\|S_3(\alpha)\|^{\frac{2}{3}} \Big(\int_0^1\big\|S_2^2(\alpha)S_6^8(\alpha)\big\|d\alpha\Big)^{\frac{1}{3}} \Big(\int_0^1\big\|S_2^2(\alpha))S_3^2(\alpha)S_6^2(\alpha)\big\|d\alpha\Big)^{\frac{2}{3}} \notag \\ \ll & N^{1+\frac{1}{3}-\frac{1}{54}+\varepsilon}.\end{aligned}$$ For $k=7$, by Lemma \[lemS2x\] and Lemma \[nS2S3S5\], we have $$\begin{aligned} \label{thm1k7} \int_{\mathfrak{n}} & \big\|S_2^2(\alpha)S_3^2(\alpha)S_6^2(\alpha)S_7^2(\alpha)\big\|d\alpha \notag \\ \ll & \sup_{\alpha\in\mathfrak{n}}\|S_3(\alpha)\|^{\frac{4}{9}} \Big(\int_0^1\big\|S_2^2(\alpha)S_6^8(\alpha)\big\|d\alpha\Big)^{\frac{1}{18}} \Big(\int_0^1\big\|S_2^2(\alpha)S_7^{12}(\alpha)\big\|d\alpha\Big)^{\frac{1}{6}} \notag \\ &\times \Big(\int_0^1\big\|S_2^2(\alpha)S_3^2(\alpha)S_6^2(\alpha)\big\|d\alpha\Big)^{\frac{7}{9}} \notag \\ \ll & N^{1+\frac{2}{7}-\frac{1}{81}+\varepsilon}.\end{aligned}$$ For $k\ge 8$ and $x$ in the form (\[236kx\]), by Lemma \[lemS2x\] and Lemma \[nS2S3S5\], we have $$\begin{aligned} \label{thm1k8} \int_{\mathfrak{n}} & \big\|S_2^2(\alpha)S_3^2(\alpha)S_6^2(\alpha)S_k^2(\alpha)\big\|d\alpha \notag \\ \ll & \sup_{\alpha\in\mathfrak{n}}\|S_3(\alpha)\|^{\frac{2}{3x}} \Big(\int_0^1\big\|S_2^2(\alpha)S_6^8(\alpha)\big\|d\alpha\Big)^{\frac{1}{3x}} \Big(\int_0^1\big\|S_2^2(\alpha)S_3^{2}(\alpha)S_k^{2x}(\alpha)\big\|d\alpha\Big)^{\frac{1}{x}} \notag \\ &\times \Big(\int_0^1\big\|S_2^2(\alpha)S_3^2(\alpha)S_6^2(\alpha)\big\|d\alpha\Big)^{1-\frac{4}{3x}} \notag \\ \ll & N^{1+\frac{2}{k}-\frac{1}{54x}+\varepsilon}.\end{aligned}$$ By (\[thm1k6\]), (\[thm1k7\]) and (\[thm1k8\]), we have $$\begin{aligned} \label{thm1236kn} \int_{\mathfrak{n}} & \big\|S_2^2(\alpha)S_3^2(\alpha)S_6^2(\alpha)S_k^2(\alpha)\big\|d\alpha\ll N^{1+\frac{2}{k}-\theta_1(k)+\varepsilon}.\end{aligned}$$ Thus, it establishes (\[thm1236k\]) by (\[thm1236km\]) and (\[thm1236kn\]). Hence, Theorem \[Thm236k\] holds. Proof of Theorem 1.2 ==================== Suppose that $N$ is a large positive integer. Let $S_k(\alpha)$ be defined as in Lemma \[LemHuaBou\]. Let $$r(n)=\sum_{\substack{p_1^2+p_2^4+p_3^4+p_4^k=n\\ N/4< p_1^2, p_2^4, p_3^4, p_4^k\le N}} (\log p_1)(\log p_2)(\log p_3)(\log p_4),$$ where $p_1$, $p_2$, $p_3$, $p_4$ are primes and the major arcs $\mathfrak{M}$, minor arcs $\mathfrak{m}$, $\mathfrak{N}$ and $\mathfrak{n}$ be defined as in section 3. Then the weighted number of representations of $n$ in the form of (\[244k\]) equals $$r(n)=\int_{0}^{1}S_2(\alpha)S_4^2(\alpha)S_k(\alpha)e(-n\alpha)d\alpha=\int_{\mathfrak{M}}+\int_{\mathfrak{m}}.$$ Whenever $\mathfrak{B}\subset [0,1]$ is measurable, we put $$r_{\mathfrak{B}}(n, N)=\int_{\mathfrak{B}}S_2(\alpha)S_4^2(\alpha)S_k(\alpha)e(-n\alpha)d\alpha.$$ Then we have $$r(n)=r_{[0,1]}(n, N)=\int_ 0^1S_2(\alpha)S_4^2(\alpha)S_k(\alpha)e(-n\alpha)d\alpha.$$ Next we will deal with the integral of major arcs and minor arcs respectively. Applying the now standard methods of enlarging major arcs (c.f. [@LiuZhan]), we can get the following result: \[888q\] For all even integer $n$ with $N<n \le 2N$, one has $r_{\mathfrak{M}}(n, N)\gg N^{\frac{1}{k}}$. \[l1501\] For $\alpha\in \mathfrak{M}(2K)\backslash\mathfrak{M}(K)$, $N^{\frac{2}{5k}}\ll K\ll N^{\frac{1}{8}}$, one has $$\begin{aligned} &S_2(\alpha)\ll N^{\frac{1}{2}+\varepsilon}K^{-\frac{1}{2}},\\ &S_3(\alpha)\ll N^{\frac{1}{3}+\varepsilon}K^{-\frac{1}{2}},\\ &S_4(\alpha)\ll N^{\frac{11}{80}+\varepsilon}K^{\frac{1}{2}}+N^{\frac{1}{4}+\varepsilon}K^{-\frac{1}{2}},\\ &S_5(\alpha)\ll N^{\frac{11}{100}+\varepsilon}K^{\frac{1}{2}}+N^{\frac{1}{5}+\varepsilon}K^{-\frac{1}{2}}.\end{aligned}$$ The Theorem 2 in [@Kumchev] implies that, if $1\le q\le H, \quad (a,q)=1, \quad \|q\alpha-a\|<HN^{-1}$ with $H\ll N^{\frac{1}{k}}$, then $$\begin{aligned} \label{lemKumchev} \sum_{p\sim N^{\frac{1}{k}}}e(\alpha p^k)\ll H^{\frac{1}{2}}N^{\frac{11}{20k}+\varepsilon}+\frac{N^{\frac{1}{k}+\varepsilon}}{(q+N\|q\alpha-a\|)^{\frac{1}{2}}}.\end{aligned}$$ If $\alpha\in\mathfrak{M}(2K)\backslash \mathfrak{M}(K),\textrm{ }\ N^{\frac{2}{5k}}\ll K\ll N^{\frac{1}{8}}$, then this lemma clearly follows by (\[lemKumchev\]). Proof of Theorem \[Thm244k\]. By Bessel’s inequality, we have $$\begin{aligned} \sum_{N< n \le 2N} & \Big\|\int_{\mathfrak{m}}S_2(\alpha)S_4^2(\alpha)S_k(\alpha)e(-n\alpha)d\alpha\Big\|^2 \le \int_{\mathfrak{m}}\big\|S_2^2(\alpha)S_4^4(\alpha)S_k^2(\alpha)\big\|d\alpha.\end{aligned}$$ Thus, to prove Theorem \[Thm244k\], it suffices to show that $$\begin{aligned} \label{thm1244k} \int_{\mathfrak{m}}\big\|S_2^2(\alpha)S_4^4(\alpha)S_k^2(\alpha)\big\|d\alpha \ll N^{1+\frac{2}{k}-\theta_2(k)+\varepsilon},\end{aligned}$$ where $\theta_2(k)$ is defined in Theorem \[Thm244k\]. Obviously, we know that $$\begin{aligned} \int_{\mathfrak{m}}\big\|S_2^2(\alpha)S_4^4(\alpha)S_k^2(\alpha)\big\|d\alpha\ll & \int_{\mathfrak{N}\backslash\mathfrak{M}}\big\|S_2^2(\alpha)S_4^4(\alpha)S_k^2(\alpha)\big\|d\alpha \\ &+ \int_{\mathfrak{n}}\big\|S_2^2(\alpha)S_4^4(\alpha)S_k^2(\alpha)\big\|d\alpha.\end{aligned}$$ First, we show that $$\begin{aligned} \label{s1506} \int_{\mathfrak{N}\backslash\mathfrak{M}}\big\|S_2^2(\alpha)S_4^4(\alpha)S_k^2(\alpha)\big\|d\alpha\ll N^{1+\frac{2}{k}-\theta_2(k)+\varepsilon}.\end{aligned}$$ It suffices to prove that $$\int_{\mathfrak{M}(2K)\backslash\mathfrak{M}(K)}\big\|S_2^2(\alpha)S_4^4(\alpha)S_k^2(\alpha)\big\|d\alpha\ll N^{1+\frac{2}{k}-\theta_2(k)+\varepsilon}$$ for $N^{\frac{2}{5k}}\ll K\ll N^{\frac{1}{8}}$. By Lemma \[l1501\] and Lemma 5.2 in [@HoffmanYu], we have $$\begin{aligned} \label{} \int_{\mathfrak{M}(2K)\backslash \mathfrak{M}(K)} & \| S_2^2(\alpha)S_4^4(\alpha)S_k^{2}(\alpha)\|d\alpha \notag \\ \ll & \sup_{\alpha\in \mathfrak{M}(2K)\backslash \mathfrak{M}(K)}\|S_2^2(\alpha)S_4^4(\alpha)\| \int_{\mathfrak{M}(2K)}\big\|S_k^2(\alpha)\big\|d\alpha \notag \\ \ll & \frac{N^{1+\varepsilon}}{K}\Big(K^2N^{\frac{11}{20}+\varepsilon}+N^{1+\varepsilon}K^{-2}\Big) \Big(N^{-1}K(N^{\frac{1}{k}}K+N^{\frac{2}{k}})\Big) \notag \\ \ll & N^{\frac{11}{20}+\frac{1}{k}+\varepsilon}K^3+N^{\frac{11}{20}+\frac{2}{k}+\varepsilon}K^2+N^{1+\frac{1}{k}+\varepsilon}K^{-1}+N^{1+\frac{2}{k}+\varepsilon}K^{-2} \notag \\ \ll & N^{1+\frac{2}{k}-\frac{4}{5k}+\varepsilon} \notag \\ \ll & N^{1+\frac{2}{k}-\theta_2(k)+\varepsilon},\end{aligned}$$ since $\frac{4}{5k}>\theta_2(k)$ for all $k\ge 4$. Next we show that $$\begin{aligned} \label{s1505} \int_{\mathfrak{n}}\big\|S_2^2(\alpha)S_4^4(\alpha)S_k^2(\alpha)\big\|d\alpha\ll N^{1+\frac{2}{k}-\theta_2(k)+\varepsilon}.\end{aligned}$$ For $k\ge 9$ and $x$ in the form in Theorem \[Thm244k\], by Lemma \[lemS2x\] and Lemma \[lemS24x\], one has $$\begin{aligned} \label{} \int_{\mathfrak{n}} & \big\|S_2^2(\alpha)S_4^4(\alpha)S_k^2(\alpha)\big\|d\alpha \notag \\ \ll & \Big(\int_0^1\|S_2^2(\alpha)S_4^{4}(\alpha)\|d\alpha\Big)^{1-\frac{1}{x}} \Big(\int_0^1\|S_2^2(\alpha)S_4^2(\alpha)S_k^{2x}\|d\alpha\Big)^{\frac{1}{x}}\sup_{\alpha\in \mathfrak{m}}\|S_4(\alpha)\|^{\frac{2}{x}} \notag \\ \ll & N^{1+\frac{2}{k}-\frac{1}{48x}+\varepsilon}.\end{aligned}$$ We use Lemma 3.1 of Zhao [@Zhao1] to prove (\[s1505\]) for $4\le k\le 8$, since the methods are same, we only give the proof for $k=5$ for simplicity. For $k=5$, by Lemma 3.1 in [@Zhao1] with $g(\alpha)=S_4(\alpha)$ and $h(\alpha)=S_5(\alpha)$, one has $$\begin{aligned} \label{871} \int_{\mathfrak{n}} & \big\|S_2^2(\alpha)S_4^4(\alpha)S_5^2(\alpha)\big\|d\alpha \notag \\ \ll & N^{\frac{1}{4}}J_0^{\frac{1}{4}}\Big(\int_{\mathfrak{n}}\big\|S_2^4(\alpha)S_4^6(\alpha)S_5^2(\alpha)\big\|d\alpha\Big)^{\frac{1}{4}} \Big(\int_{\mathfrak{n}}\big\|S_2^2(\alpha)S_4^3(\alpha)S_5^2(\alpha)\big\|d\alpha\Big)^{\frac{1}{2}} \notag \\ &+ N^{\frac{1}{4}(1-2^{-4})+\varepsilon} \int_{\mathfrak{n}}\big\|S_2^2(\alpha)S_4^3(\alpha)S_5^2(\alpha)\big\|d\alpha,\end{aligned}$$ where $$J_0\ll N^{-\frac{3}{5}+\varepsilon}$$ by Lemma 2.2 in [@Zhao1]. By Hölder’s inequality, Lemmas \[lemS2x\] and \[lemS23k\], one has $$\begin{aligned} \label{872} \int_{\mathfrak{n}} &\big\|S_2^2(\alpha)S_4^3(\alpha)S_5^2(\alpha)\big\|d\alpha \notag \\ \ll & \Big(\int_{\mathfrak{n}}\big\|S_2^2(\alpha)S_4^4(\alpha)S_5^2(\alpha)\big\|d\alpha\Big)^{\frac{1}{4}} \Big(\int_0^1\big\|S_2^2(\alpha)S_4^4(\alpha)\big\|d\alpha\Big)^{\frac{1}{2}} \Big(\int_0^1\big\|S_2^2(\alpha)S_5^6(\alpha)\big\|d\alpha\Big)^{\frac{1}{4}} \notag \\ \ll & N^{\frac{4}{5}+\varepsilon} \Big(\int_{\mathfrak{n}}\big\|S_2^2(\alpha)S_4^4(\alpha)S_5^2(\alpha)\big\|d\alpha\Big)^{\frac{1}{4}}.\end{aligned}$$ By Lemma \[nS2S3S5\], $$\begin{aligned} \label{873} \int_{\mathfrak{n}}\big\|S_2^4(\alpha)S_4^6(\alpha)S_5^2(\alpha)\big\|d\alpha\ll & \sup_{\alpha\in \mathfrak{n}}\big\|S_2(\alpha)S_4(\alpha)\big\|^2\int_{\mathfrak{n}}\big\|S_2^2(\alpha)S_4^4(\alpha)S_5^2(\alpha)\big\|d\alpha \notag \\ \ll & N^{\frac{65}{48}+\varepsilon} \int_{\mathfrak{n}}\big\|S_2^2(\alpha)S_4^4(\alpha)S_5^2(\alpha)\big\|d\alpha.\end{aligned}$$ By (\[871\]),(\[872\]) and (\[873\]), we have $$\begin{aligned} \int_{\mathfrak{n}}\big\|S_2^2(\alpha)S_4^4(\alpha)S_5^2(\alpha)\big\|d\alpha\ll N^{1+\frac{2}{5}-\frac{1}{48}+\varepsilon}.\end{aligned}$$ Thus, it establishes (\[thm1244k\]) by (\[s1506\]) and (\[s1505\]). Proof of Theorem 1.3 ==================== Suppose that $N$ is a large positive integer. Let $S_k(\alpha)$ be defined as Lemma \[LemHuaBou\]. Denote $$r(n)=\sum_{\substack{p_1^2+p_2^3+p_3^5+p_4^k=n\\ N/4<p_1^2, p_2^3, p_3^5, p_4^k\le N}} (\log p_1)(\log p_2)(\log p_3)(\log p_4),$$ and the major arcs $\mathfrak{M}$, minor arcs $\mathfrak{m}$, $\mathfrak{N}$ and $\mathfrak{n}$ be defined as in section 3. Then the weighted number of representations of $n$ in the form of (\[235k\]) equals $$r(n)=\int_{0}^{1}S_2(\alpha)S_3(\alpha)S_5(\alpha)S_k(\alpha)e(-n\alpha)d\alpha=\int_{\mathfrak{M}}+\int_{\mathfrak{m}}.$$ Whenever $\mathfrak{B}\subset [0,1]$ is measurable, we put $$r_{\mathfrak{B}}(n, N)=\int_{\mathfrak{B}}S_2(\alpha)S_3(\alpha)S_5(\alpha)S_k(\alpha)e(-n\alpha)d\alpha.$$ Then we have $$r(n)=r_{[0,1]}(n, N)=\int_ 0^1S_2(\alpha)S_3(\alpha)S_5(\alpha)S_k(\alpha)e(-n\alpha)d\alpha.$$ Next we will deal with the integral of major arcs and minor arcs respectively. Applying the now standard methods of enlarging major arcs (c.f. [@LiuZhan]), we can get the following result: For all even integer $n$ with $N<n \le 2N$, one has $r_{\mathfrak{M}}(n, N)\gg N^{\frac{1}{30}+\frac{1}{k}}$. Proof of Theorem \[Thm235k\]. By Bessel’s inequality, we have $$\begin{aligned} \sum_{N< n \le 2N} & \Big\|\int_{\mathfrak{m}}S_2(\alpha)S_3(\alpha)S_5(\alpha)S_k(\alpha)e(-n\alpha)d\alpha\Big\|^2 \le \int_{\mathfrak{m}}\big\|S_2^2(\alpha)S_3^2(\alpha)S_5^2(\alpha)S_k^2(\alpha)\big\|d\alpha.\end{aligned}$$ Thus, to prove Theorem \[Thm235k\], it suffices to show that $$\begin{aligned} \label{s1507} \int_{\mathfrak{m}}\big\|S_2^2(\alpha)S_3^2(\alpha)S_5^2(\alpha)S_k^2(\alpha)\big\|d\alpha\ll N^{\frac{16}{15}+\frac{2}{k}-\theta_3(k)+\varepsilon},\end{aligned}$$ where $\theta_3(k)$ is defined in Theorem \[Thm235k\]. Obviously, one has $$\begin{aligned} \int_{\mathfrak{m}}\big\|S_2^2(\alpha)S_3^2(\alpha)S_5^2(\alpha)S_k^2(\alpha)\big\|d\alpha\ll & \int_{\mathfrak{N}\backslash\mathfrak{M}}\big\|S_2^2(\alpha)S_3^2(\alpha)S_5^2(\alpha)S_k^2(\alpha)\big\|d\alpha \\ & +\int_{\mathfrak{n}}\big\|S_2^2(\alpha)S_3^2(\alpha)S_5^2(\alpha)S_k^2(\alpha)\big\|d\alpha.\end{aligned}$$ First, we show that $$\begin{aligned} \label{s1508} \int_{\mathfrak{N}\backslash\mathfrak{M}}\big\|S_2^2(\alpha)S_3^2(\alpha)S_5^2(\alpha)S_k^2(\alpha)\big\|d\alpha \ll N^{\frac{16}{15}+\frac{2}{k}-\theta_3(k)+\varepsilon}.\end{aligned}$$ It suffices to prove that $$\int_{\mathfrak{M}(2K)\backslash\mathfrak{M}(K)}\big\|S_2^2(\alpha)S_3^2(\alpha)S_5^2(\alpha)S_k^2(\alpha)\big\|d\alpha \ll N^{\frac{16}{15}+\frac{2}{k}-\theta_3(k)+\varepsilon}$$ with $N^{\frac{1}{5k}}\ll K\ll N^{\frac{1}{8}}$. By Lemma \[l1501\] and Lemma 5.2 in [@HoffmanYu], one has $$\begin{aligned} \int_{\mathfrak{M}(2K)\backslash\mathfrak{M}(K)} & \big\|S_2^2(\alpha)S_3^2(\alpha)S_5^2(\alpha)S_k^2(\alpha)\big\|d\alpha \\ \ll &\frac{ N^{1+\varepsilon}}{K}\frac{N^{\frac{2}{3}+\varepsilon}}{K}(N^{\frac{11}{50}+\varepsilon}K+N^{\frac{2}{5}+\varepsilon}K^{-1}) \Big(N^{-1}K(N^{\frac{1}{k}}K+N^{\frac{2}{k}})\Big) \\ \ll & N^{\frac{16}{15}+\frac{2}{k}-\frac{4}{5k}+\varepsilon} \\ \ll & N^{\frac{16}{15}+\frac{2}{k}-\theta_3(k)+\varepsilon},\end{aligned}$$ since $\frac{4}{5k}>\theta_3(k)$ for all $k\ge 5$. This establishes (\[s1508\]). Next, we show that $$\begin{aligned} \label{s1509} \int_{\mathfrak{n}} & \big\|S_2^2(\alpha)S_3^2(\alpha)S_5^2(\alpha)S_k^2(\alpha)\big\|d\alpha \ll N^{\frac{16}{15}+\frac{2}{k}-\theta_3(k)+\varepsilon}.\end{aligned}$$ For $k=5$, by Lemma 3.1 in [@Zhao1] with $g(\alpha)=S_3(\alpha)$ and $h(\alpha)=S_5(\alpha)$, one has $$\begin{aligned} \label{x01} \int_{\mathfrak{n}} & \big\|S_2^2(\alpha)S_3^2(\alpha)S_5^4(\alpha)\big\|d\alpha \notag \\ \ll & N^{\frac{1}{3}}J_1^{\frac{1}{4}}\Big(\int_0^1\big\|S_2^4(\alpha)S_3^2(\alpha)S_5^6(\alpha)\big\|d\alpha\Big)^{\frac{1}{4}} \Big(\int_0^1\big\|S_2^2(\alpha)S_3(\alpha)S_5^4(\alpha)\big\|d\alpha\Big)^{\frac{1}{2}} \notag \\ & + N^{\frac{1}{3}(1-2^{-3})}\int_0^1\big\|S_2^2(\alpha)S_3(\alpha)S_5^4(\alpha)\big\|d\alpha,\end{aligned}$$ where $$J_1\ll N^{-\frac{3}{5}+\varepsilon}$$ by Lemma 2.2 in [@Zhao1]. By Hölder’s inequality, Lemmas \[lemS2x\] and \[lemS23k\], one has $$\begin{aligned} \label{x02} \int_{\mathfrak{n}} \big\|S_2^2(\alpha)S_3(\alpha)S_5^4(\alpha)\big\|d\alpha \notag \ll & \Big(\int_0^1\big\|S_2^2(\alpha)S_3^2(\alpha)S_5^2(\alpha)\big\|d\alpha\Big)^{\frac{1}{2}} \Big(\int_0^1\big\|S_2^2(\alpha)S_5^6(\alpha)\big\|d\alpha\Big)^{\frac{1}{2}} \notag \\ \ll & N^{\frac{17}{15}+\varepsilon}.\end{aligned}$$ Also, $$\begin{aligned} \label{x03} \int_{\mathfrak{n}} \big\|S_2^4(\alpha)S_3^2(\alpha)S_5^6(\alpha)\big\|d\alpha \notag \ll & \sup_{\alpha\in \mathfrak{n}}\big\|S_2^2(\alpha)S_5^2(\alpha)\big\|\int_0^1\big\|S_2^2(\alpha)S_3^2(\alpha)S_5^4(\alpha)\big\|d\alpha \notag \\ \ll & N^{\frac{19}{15}+\varepsilon}\int_0^1\big\|S_2^2(\alpha)S_3^2(\alpha)S_5^4(\alpha)\big\|d\alpha.\end{aligned}$$ Thus,by (\[x01\])-(\[x03\]), we obtain that $$\begin{aligned} \int_{\mathfrak{n}} \big\|S_2^2(\alpha)S_3^2(\alpha)S_5^4(\alpha)\big\|d\alpha \ll N^{1+\frac{7}{15}-\frac{1}{24}+\varepsilon}.\end{aligned}$$ It establishes (\[s1509\]) for $k=5$. For $k=6$, applying Hölder’s inequality, Lemmas \[lemS2x\] and \[lemS23k\], one has $$\begin{aligned} \label{} \int_{\mathfrak{n}} & \big\|S_2^2(\alpha)S_3^2(\alpha)S_5^{2}(\alpha)S_6^2(\alpha)\big\|d\alpha \notag \\ \ll & \sup_{\alpha\in \mathfrak{n}}\big\|S_3(\alpha)\big\|^{\frac{8}{9}}\Big(\int_0^1\big\|S_2^2(\alpha)S_5^6(\alpha)\big\|d\alpha\Big)^{\frac{1}{3}} \Big(\int_0^1\big\|S_2^2(\alpha)S_6^8(\alpha)\big\|d\alpha\Big)^{\frac{1}{9}} \notag \\ & \times\Big(\int_0^1\big\|S_2^2(\alpha)S_3^2(\alpha)S_6^2(\alpha)\big\|d\alpha\Big)^{\frac{5}{9}} \notag \\ \ll & N^{1+\frac{2}{5}-\frac{2}{81}+\varepsilon}.\end{aligned}$$ It establishes (\[s1509\]) for $k=6$. For $k\ge 7$ and $x$ in the form in Theorem \[Thm235k\], by Lemmas \[lemS2x\], \[lemS23k\] and \[nS2S3S5\], one has $$\begin{aligned} \int_{\mathfrak{n}} & \big\|S_2^2(\alpha)S_3^2(\alpha)S_5^{2}(\alpha)S_k^2(\alpha)\big\|d\alpha \notag \\ \ll & \sup_{\alpha\in \mathfrak{n}}\big\|S_3(\alpha)\big\|^{\frac{1}{x}}\Big(\int_0^1\big\|S_2^2(\alpha)S_5^6(\alpha)\big\|d\alpha\Big)^{\frac{1}{2x}} \Big(\int_0^1\big\|S_2^2(\alpha)S_3^2(\alpha)S_5^2(\alpha)\big\|d\alpha\Big)^{1-\frac{3}{2x}} \notag \\ \times & \Big(\int_0^1\big\|S_2^2(\alpha)S_3^2(\alpha)S_k^{2x}(\alpha)\big\|d\alpha\Big)^{\frac{1}{x}} \notag \\ \ll & N^{\frac{16}{15}+\frac{2}{k}-\frac{1}{36x}+\varepsilon}.\end{aligned}$$ It establishes (\[s1509\]) for $k\ge 7$. Hence, (\[s1507\]) holds by (\[s1508\]) and (\[s1509\]), and it establishes Theorem \[Thm235k\]. Proof of Theorem \[Thm3333k\] ============================= Suppose that $N$ is a large positive integer. Let $S_k(\alpha)$ be defined as in Lemma \[LemHuaBou\]. Let $$r(n)=\sum_{\substack{p_1^3+p_2^3+p_3^3+p_4^3+p_5^k=n\\ N/5< p_1^3, p_2^3, p_3^3, p_4^3, p_5^k\le N}} (\log p_1)(\log p_2)(\log p_3)(\log p_4)(\log p_5),$$ and the major arcs $\mathfrak{M}=\mathfrak{M}(Q)$, minor arcs $\mathfrak{m}$, $\mathfrak{N}$ and $\mathfrak{n}$ be defined as in section 3 with $Q=N^{\frac{1}{2k}}$. Then the weighted number of representations of $n$ in the form of (\[3333k\]) equals $$r(n)=\int_{0}^{1}S_3^4(\alpha)S_k(\alpha)e(-n\alpha)d\alpha=\int_{\mathfrak{M}}+\int_{\mathfrak{m}}.$$ Whenever $\mathfrak{B}\subset [0,1]$ is measurable, we put $$r_{\mathfrak{B}}(n, N)=\int_{\mathfrak{B}}S_3^4(\alpha)S_k(\alpha)e(-n\alpha)d\alpha.$$ Then we have $$r(n)=r_{[0,1]}(n, N)=\int_ 0^1S_3^4(\alpha)S_k(\alpha)e(-n\alpha)d\alpha.$$ Next we will deal with the integral of major arcs and minor arcs respectively. Applying the now standard methods of enlarging major arcs (c.f. [@LiuZhan]), we can get the following result: For all odd integer $n$ with $N<n\le 2N$, one has $r_{\mathfrak{M}}(n, N)\gg N^{\frac{1}{k}+\frac{1}{3}}$. Proof of Theorem \[Thm3333k\]. By Bessel’s inequality, we have $$\begin{aligned} \sum_{N< n \le 2N} & \Big\|\int_{\mathfrak{m}}S_3^4(\alpha)S_k(\alpha)e(-n\alpha)d\alpha\Big\|^2 \le \int_{\mathfrak{m}}\big\|S_3^8(\alpha)S_k^2(\alpha)\big\|d\alpha.\end{aligned}$$ Thus, to prove Theorem \[Thm3333k\], it suffices to show that $$\begin{aligned} \label{s1701} \int_{\mathfrak{m}}\big\|S_3^8(\alpha)S_k^2(\alpha)\big\|d\alpha\ll N^{\frac{5}{3}+\frac{2}{k}-\theta_4(k)+\varepsilon},\end{aligned}$$ where $\theta_4(k)$ is defined in Theorem \[Thm3333k\]. Obviously, one has $$\begin{aligned} \int_{\mathfrak{m}}\big\|S_3^8(\alpha)S_k^2(\alpha)\big\|d\alpha\ll \int_{\mathfrak{N}\backslash\mathfrak{M}}\big\|S_3^8(\alpha)S_k^2(\alpha)\big\|d\alpha + \int_{\mathfrak{n}}\big\|S_3^8(\alpha)S_k^2(\alpha)\big\|d\alpha.\end{aligned}$$ First, we show that $$\begin{aligned} \label{s1702} \int_{\mathfrak{N}\backslash\mathfrak{M}}\big\|S_3^8(\alpha)S_k^2(\alpha)\big\|d\alpha \ll N^{\frac{5}{3}+\frac{2}{k}-\theta_4(k)+\varepsilon}.\end{aligned}$$ It suffices to prove that $$\int_{\mathfrak{M}(2K)\backslash\mathfrak{M}(K)}\big\|S_3^8(\alpha)S_k^2(\alpha)\big\|d\alpha \ll N^{\frac{5}{3}+\frac{2}{k}-\theta_4(k)+\varepsilon}$$ with $N^{\frac{1}{2k}}\ll K\ll N^{\frac{1}{8}}$. By Lemma \[l1501\] and Lemma 5.2 in [@HoffmanYu], one has $$\begin{aligned} \int_{\mathfrak{M}(2K)\backslash\mathfrak{M}(K)}\big\|S_3^8(\alpha)S_k^2(\alpha)\big\|d\alpha \ll & N^{\frac{8}{3}+\varepsilon}K^{-4} \Big(N^{-1}K(N^{\frac{1}{k}}K+N^{\frac{2}{k}})\Big) \\ \ll & N^{\frac{5}{3}+\frac{2}{k}-\frac{3}{2k}+\varepsilon} \\ \ll & N^{\frac{5}{3}+\frac{2}{k}-\theta_4(k)+\varepsilon},\end{aligned}$$ since $\frac{3}{2k}>\theta_4(k)$ for all $k\ge 4$. This establishes (\[s1702\]). Next, we show that $$\begin{aligned} \label{s1703} \int_{\mathfrak{n}} & \big\|S_3^8(\alpha)S_k^2(\alpha)\big\|d\alpha \ll N^{\frac{5}{3}+\frac{2}{k}-\theta_4(k)+\varepsilon}.\end{aligned}$$ For $k=4$, by Lemma 3.1 in [@Zhao1] with $g(\alpha)=h(\alpha)=S_3(\alpha)$, one has $$\begin{aligned} \label{10310} \int_{\mathfrak{n}}\big\|S_3^8(\alpha)S_4^2(\alpha)\big\|d\alpha\ll & N^{\frac{1}{3}}J^{\frac{1}{4}}\Big(\int_{\mathfrak{n}}\big\|S_3^{12}(\alpha)S_4^4(\alpha)\big\|d\alpha\Big)^{\frac{1}{4}} \Big(\int_{\mathfrak{n}}\big\|S_3^{7}(\alpha)S_4^2(\alpha)\big\|d\alpha\Big)^{\frac{1}{2}} \notag \\ &+ N^{\frac{1}{3}(1-2^{-3})}\int_{\mathfrak{n}}\big\|S_3^{7}(\alpha)S_4^2(\alpha)\big\|d\alpha\end{aligned}$$ where $$J\ll N^{-\frac{1}{3}+\varepsilon}$$ by Lemma 2.2 in [@Zhao1]. By Hölder’s inequality and Lemma \[8233\], one has $$\begin{aligned} \label{10311} \int_{\mathfrak{n}}\big\|S_3^{7}(\alpha)S_4^2(\alpha)\big\|d\alpha &\ll \Big(\int_0^1\big\|S_3^{8}(\alpha)\big\|d\alpha\Big)^{\frac{3}{4}} \Big(\int_0^1\big\|S_3^{4}(\alpha)S_4^8(\alpha)\big\|d\alpha\Big)^{\frac{1}{4}}\notag\\ &\ll N^{\frac{11}{6}+\varepsilon}.\end{aligned}$$ Also, $$\begin{aligned} \label{10312} \int_{\mathfrak{n}}\big\|S_3^{12}(\alpha)S_4^4(\alpha)\big\|d\alpha &\ll \sup_{\alpha\in \mathfrak{n}}\|S_3(\alpha)\|^6\Big(\int_0^1\big\|S_3^{8}(\alpha)\big\|d\alpha\Big)^{\frac{1}{2}} \Big(\int_0^1\big\|S_3^{4}(\alpha)S_4^8(\alpha)\big\|d\alpha\Big)^{\frac{1}{2}}\notag\\ &\ll N^{\frac{23}{6}+\varepsilon}.\end{aligned}$$ Thus, by (\[10310\])-(\[10312\]), we obtain that $$\int_{\mathfrak{n}}\big\|S_3^8(\alpha)S_4^2(\alpha)\big\|d\alpha\ll N^{\frac{5}{3}+\frac{1}{2}-\frac{1}{24}+\varepsilon}.$$ It establishes (\[s1703\]) for $k=4$. For $k\ge 5$ and $x$ in the form in Theorem \[Thm3333k\], by Hölder’s inequality and Lemmas \[8233\] and \[nS2S3S5\], one has $$\begin{aligned} \int_{\mathfrak{n}}\big\|S_3^8(\alpha)S_k^2(\alpha)\big\|d\alpha \ll & \sup_{\alpha\in\mathfrak{n}}\big\|S_3(\alpha)\big\|^{\frac{4}{x}}\big(\int_0^1\big\|S_3^8(\alpha)\big\|d\alpha\big)^{1-\frac{1}{x}} \big(\int_0^1\big\|S_3^4(\alpha)S_k^{2x}(\alpha)\big\|d\alpha\big)^{\frac{1}{x}} \notag \\ \ll & N^{\frac{5}{3}+\frac{2}{k}-\theta_4(k)+\varepsilon}.\end{aligned}$$ This establishes (\[s1703\]) for $k\ge 5$. Hence, (\[s1701\]) holds by (\[s1702\]) and (\[s1703\]).Thus it establishes Theorem \[Thm3333k\]. Acknowlegements. This work is supported by National Natural Foundations of China (No. 11771176) . [2010]{} C. Bauer, [An improvement on a theorem of the Goldbach-Waring type]{}, Rocky Mount. J. Math. [31]{} (2001), 1151–1170. C. Bauer, [A remark on a theorem of the Goldbach Waring type]{}, Studia Sci. Math. Hungar. [41]{} (2004), 309–324. J. Bourgain, [On the Vinogradov mean value,]{}, Trudy Mat. Inst. Steklova. [296]{} (2017), 36–46. J. Brüdern, [A ternary problem in addditive prime number theory]{}, J., Steuding, R. (eds.)From Arithmetic to Zeta-Functions. Springer, Cham (2016). J. W. Hoffman and G. Yu, [A ternary addtive problem]{}, Monatsh Math. [172]{} (2013), 293–321. L.K. Hua, [Additive theory of prime numbers]{}, Science press, Beijing 1957; English version, American Mathematical Society, Providence 1965. A. Kumchev, [On weyl sums over primes and almost primes]{}, Michigan Math. J. [54]{} (2006), 243–268. J.Y. Liu and T. Zhan, [Sums of five almost equal prime squares (II)]{}, Sci. in China. [41]{} (1998), 710–722. Z.X. Liu, [An improvement on Waring-Goldbach problem for unlike powers]{}, Acta Math. Hungar. [130]{} (2011), 118–139. M.G. Lu and Z. Shan, [A problem of Waring-Goldbach type]{}, J. China Univ. Sci. Tech. Suppl. I (1982), 1–8 (in Chinese). K. Prachar, [Über ein Problem vom Waring-Goldbach’schen Typ]{}, Monatsh. Math. [57]{} (1953), 66–74. X.M. Ren and Kai-Man Tsang, [Goldbach-Waring problem for unlike powers]{}, Acta Math. Sinica, English Series (2). [23]{} (2007), 265–280. X. M. Ren and Kai-Man Tsang, [Goldbach-Waring problem for unlike powers (II)]{}, Acta Math. Sinica (Chinese Series). [50]{} (2007), 175–182. W. Schwarz, [Zur Darstellung von Zahlen durch Summen von Primzahlpotenzen]{}, II. J. Reine Angew. Math. [206]{} (1961), 78–112. R. C. Vaughan, [The Hardy-Littlewood method]{}, 2nd ed. Cambridge University Press, Cambridge 1997. L. Zhao, [On the Waring-Goldbach problem for fourth and sixth powers]{}, Proc. London Math. Soc. [108]{} (2014), 1593–1622. L. Zhao, [The additive problem with one cube and three cubes of primes]{}, Michigan Math. J. [63]{} (2014), 763–779. L. Zhao, [The exceptional set for sums of unlike powers of primes]{}, Acta Math. Sinica, English Series. [30]{} (2014), 1897–1904. [^1]: \*corresponding author	{ "pile_set_name": "ArXiv" }
--- abstract: 'Hot electron temperatures and electron energy spectra in the course of interaction between intense laser pulse and overdense plasmas are reexamined from a viewpoint of the difference in laser wavelength. The hot electron temperature measured by a particle-in-cell simulation is scaled by $I$ rather than $I\lambda^2$ at the interaction with overdense plasmas with fixed ions, where $I$ and $\lambda$ are the laser intensity and wavelength, respectively.' address: 'National Institute of Advanced Industrial Science and Technology (AIST), Tsukuba, Ibaraki 305-8568, Japan' author: - \| Susumu Kato, Eiichi Takahashi, Tatsuya Aota,\ Yuji Matsumoto, Isao Okuda, and Yoshiro Owadano title: 'Effects of Laser Wavelength and Density Scalelength on Absorption of Ultrashort Intense Lasers on Solid-Density Targets' --- Introduction ============ The interaction of intense laser pulses with overdense plasmas has attracted much interest for the fast ignitor concept in inertial fusion energy [@Tabak1994]. The interaction of ultrashort intense laser pulses with thin solid targets have also been of great interest for the application to high energy ion sources [@ionsource]. Ultraintense irradiation experiments using an infrared subpicosecond laser, e.g., Nd:glass ($\lambda =$ 1,053 nm) or Ti:sapphire ($\lambda =$800 nm) lasers, whose powers and focused intensities exceed 100 TW and $10^{20}$ W/cm$^2$, are possible using chirped pulse amplification techniques [@Maine:1988]. In these experiments, the classical normalized momentum of electrons $a_0\equiv P_{\rm osc}/mc=({I\lambda_\mu^2}/{1.37\times10^{18}})^{1/2} \geq 1$, where $m$ is the electron mass, $c$ is the speed of light, $I$ is the laser intensity in W/cm$^2$, and $\lambda_\mu$ is the wavelength in $\mu$m. On the other hand, a KrF laser ($\lambda =$ 248 nm) has an advantage as the fast ignitor in that the critical density is close to the core, and hot electron energies are suitable since the critical density of the KrF laser is ten times greater than that of an infrared laser [@Shaw:1999]. The peak intensities of KrF laser systems were only the order of $10^{18}$ W/cm$^2$, namely $a_0 < 1$ [@Teubner:1996]. Therefore, the dependence of the laser plasma interactions on the laser wavelength was not investigated in $a_0 \geq 1$. Recently, the laser absorption and hot electron generation have been studied by the high intensity KrF laser system of which focused intensity is greater than $10^{19}$ W/cm$^2$ [@takahashi2004]. However, the production of hot electrons by the high intensity KrF laser has not been fully understood yet. Namely, it has been not clear that the effects of laser wavelength on hot electrons produced by ultrashort intense laser pulse on solid-density targets. The absorption, electron energy spectrum, and hot electron temperature have usually been investigated and scaled using the parameters $I\lambda^2$, $n_{\rm{e}}/n_{\rm{c}}$, and $L/\lambda$ [@Lefebvre:1997; @Wilks:1997], where $n_{\rm{e}}$, $n_{\rm{c}}$, and $L$ are the electron density, critical density, and density scale length, respectively. Critical density absorption of the laser light converts laser energy into hot electrons having a suprathermal temperature $T_{\rm {h}}$ approximately proportional to $\sqrt{I\lambda^2}$ for $a_0 > 1$, and $T_{\rm{h}}\sim [(1+a_0^2)^{1/2}-1]mc^2$ at moderate densities [@Wilks:1992], where $mc^2=511$ keV, $m$ is an electron rest mass. The scaling of the hot electron temperature has been supported by experiments of Nd:glass and Ti:sapphire lasers [@Malka:1996]. On the other hand, the results of one-dimensional simulation for normal incidence in the density region $4 < n_{\rm{e}}/n_{\rm{c}} < 100$ and the normalized intensity $4 < a_0^2 < 30$ have shown that $T_{\rm{h}} \sim \eta \left(n_{\rm{e}}/n_{\rm{c}}\right)^{\alpha} [(1+a_{\rm s}^2)^{1/2}-1] mc^2$, where $a_{\rm s}= \beta a_0\left(n_{\rm{c}}/n_{\rm{e}}\right)^{1/2}$ is the electromagnetic fields at the surface of the overdense plasma, $\eta = 0.5\sim1.1$ and $\alpha = 1/2$, which depend weakly on $I\lambda^2$ and $n_{\rm{e}}/n_{\rm{c}}$ [@Wilks:1997]. $\beta$ is weakly depend on the angle of incidence, absorption rate, and $n_{\rm{e0}}/n_{\rm{c}}$[@Lichters1996]. The hot electron temperature is scaled by the amplitude of electromagnetic fields at the plasma surface rather than that in vacuum; namely, the hot electron temperature is slightly dependent on the wavelength. In addition, at the interaction of intense laser pulses with solid density plasma which has a sharp density gradient the hot electron temperature is scaled by $I\lambda^{-1}$ rather than $I\lambda^2$ [@kato2002]. In the present paper, we study the absorption of ultrashort intense laser pulses on overdense plasmas for different laser wavelengths ($\lambda$ = 0.25, 0.5, and 1 $\mu$m) using a particle-in-cell (PIC) simulation. PIC simulation ============== In order to investigate hot electron generation for oblique incidence, we use the relativistic 1 and $2/2$ dimensional PIC simulation with the boost frame moving with $c \sin \theta$ parallel to the target surface, where $c$ and $\theta$ are the speed of light and an angle of incidence[@Gibbon:1992]. In the simulation, the target is the fully ionized plastic and the electron density $n_{\rm{e0}} \sim 3.5 \times10^{23}{\rm cm^{-3}}$. The density correspond to $n_{\rm{e0}}/n_{\rm{c}}=20$, $78$, and $310$ for $\lambda =$ 0.25, 0.5, and $\lambda = 1 \mu$m, respectively. The density profile has a sharp density gradient, $n_{\rm{e}}(x)= n_{\rm{e0}}$ for $x \ge 0$ and $n_{\rm{e}}(x)= 0$ for $x<0$. In order to clarify the boundary effect, ions are fixed, namely, the boundary does not move all the time. The laser pulse starts at $x < 0$ and propagates towards $x > 0$. The laser intensity rises in 5 fs and remains constant after that. The irradiated intensity $I=5\times10^{19}$ W/cm$^2$ and the angle of incidence $\theta=30^\circ$ and $45^\circ$ (p-polarized), respectively. $a_0^2 =$ 2.3, 9.2, and 36 for $\lambda =$ 0.25, 0.5, and 1.0 $\mu$m, respectively. However, $a_s^2=0.12\beta$ for all wavelength. Normalized electron energy distributions after 50 fs are shown in Fig.1(a) and 1(b) for $\theta=30^\circ$ and $45^\circ$, respectively. The hot electron temperatures are 140 and 340 keV for $\theta=30^\circ$ and $45^\circ$, respectively. The hot electron temperatures does not depend on the laser wavelength. The result is well agreement with that of a simple sharp boundary theory. On the other hand, the absorption depends on the laser wavelength, $A(\theta=30^\circ)=$ 0.9-1.8%, 2.2-3.0%, and 3.6-4.3% and $A(\theta=45^\circ)=$ 2.6-4.1%, 5.3-6.7%, and 7.8-9.0% for $\lambda =$ 1.0, 0.5, and 0.25 $\mu$m, respectively. Concluding Remarks ================== The effects of laser wavelength on hot electrons produced by ultrashort intense laser pulse on solid-density targets are studied by the use of a PIC simulation. As a result, the dependence to the wavelength of hot electron temperature strongly depend on the boundary condition, even in the one dimensional case, namely, all are not determined only by $I\lambda^2$. The density profiles of both preformed plasma [@Yu:2000] and multi-dimensional effects such as surface deformation [@Wilks:1992] are very important in the actual experiments. References {#references .unnumbered} ========== [99]{} M. Tabak [et al.]{}, Phys. Plasmas [1]{}, 1626 (1994); R. Kodama [et al.]{}, Nature [412]{}, 798 (2001). D. Umstadter, J. Phys. D [36]{} R151 (2003); S. P. Hatchett [et al.]{}, Phys. Plasmas [7]{}, 2076 (2000), P. Maine [et al.]{}, IEEE J. Quantum Electron. 24, 398 (1988); M. D. Perry [et al.]{}, Opt. Lett. 24, 160 (1999). M. J. Shaw [et al.]{}, Fusion Eng. Des. 44, 209 (1999). U. Teubner [et al.]{}, Phys. Rev. E 54, 4167 (1996); M. Borghesi [et al.]{}, Phys. Rev. E 60, 7374 (1999). E.Takahashi [et al.]{}, [Proceedings of the Third International Conference on Inertial Fusion Sciences and Applications (IFSA2003)]{}, Editors: B. A. Hammel D. D. Meyerhofer J. Meyer-ter-Vehn H. Azechi, p.406 (American Nuclear Society, Inc., 2004). E. Lefebvre and G. Bonnaud, Phys. Rev. E 55, 1011 (1997). S. C. Wilks and W. L. Kruer, IEEE J. Quantum Electron. 33, 1954 (1997) S. C. Wilks, Phys. Fluids B 5, 2603 (1993); S. C. Wilks [et al.]{}, Phys. Rev. Lett. 69, 1383 (1992). G. Malka and J. L. Miquel, Phys. Rev. Lett. 77, 75 (1996); Y. Oishi [et al.]{}, Appl. Phys. Lett. 79, 1234 (2001). A. Lichters [et al.]{}, Phys. Plasmas 3, 3425 (1996) S. Kato [et al.]{}, J. Plasma Fusion Res. 78, 717 (2002). A. Bourdier, Phys. Fluids 26, 1804 (1983); P. Gibbon and A. R. Bell, Phys. Rev. Lett. 68, 1535 (1992). W. L. Kruer and K. G. Estabrook, Phys. Fluids 28, 430 (1985). S. Kato [et al.]{}, [Proc. of the Seventh International Symposium of the Graduate University for Advanced Studies on Science of Super-Strong Field Interactions, Hayama, JAPAN, 2002]{}, Editors: K. Nakajima and M. Deguchi, p.290 (American Institute of Physics, 2002) . Wei Yu [et al.]{}, Phys. Rev. Lett. 85, 570 (2000); T. E. Cowan [et al.]{}, Phys. Rev. Lett. 84, 903 (2000). ![image](Fig1a.eps){width="80mm"} ![image](Fig1b.eps){width="80mm"}	{ "pile_set_name": "ArXiv" }
--- abstract: 'This paper deals with dynamical networks for which the relations between node signals are described by proper transfer functions and external signals can influence each of the node signals. We are interested in graph-theoretic conditions for identifiability of such dynamical networks, where we assume that only a subset of nodes is measured but the underlying graph structure of the network is known. This problem has recently been investigated from a generic viewpoint. Roughly speaking, generic identifiability means that the transfer functions in the network can be identified for “almost all" network matrices associated with the graph. In this paper, we investigate the stronger notion of identifiability for all network matrices. To this end, we introduce a new graph-theoretic concept called the graph simplification process. Based on this process, we provide necessary and sufficient topological conditions for identifiability. Notably, we also show that these conditions can be verified by polynomial time algorithms. Finally, we explain how our results generalize existing sufficient conditions for identifiability.' author: - 'Henk J. van Waarde, Pietro Tesi, and M. Kanat Camlibel [^1]' bibliography: - 'MyRef.bib' title: Necessary and Sufficient Topological Conditions for Identifiability of Dynamical Networks --- Network Analysis and Control, System Identification, Linear Systems. Introduction ============ of dynamical systems appear in a variety of domains, including power systems, robotic networks, and aerospace systems [@Mesbahi2010]. In this paper, we consider a dynamical network model in which the relations between node signals are modelled by proper transfer functions. Such network models have received much attention in recent years, see e.g. [@vandenHof2013; @Dankers2014; @Hendrickx2018; @Weerts2018; @vanWaarde2018]. The interconnection structure of a dynamical network can be represented by a directed graph, where vertices (or nodes) represent scalar signals, and edges correspond to transfer functions connecting different node signals. We will assume that the underlying graph (i.e., the topology) of the dynamical network is known. We remark that the related problem of topology identification has also been studied, see e.g. [@Goncalves2008; @Yuan2011; @Nabi-Abdolyousefi2012; @Materassi2012; @Shahrampour2015; @vanWaarde2017a]. We are interested in conditions for identifiability of dynamical networks. Identifiability is a fundamental property of a model set that guarantees that a unique (network) model can be identified, given informative data. Thus identifiability can be thought of as a prerequisite for identification: if identifiability does not hold then it is impossible to uniquely determine a network model, irrespective of the particular identification method and the experimental conditions. In the literature, several methods have been proposed for network identification [@vandenHof2013; @Dankers2014; @Weerts2018b; @Haber2014], these methods all exploit the structure of the network. For instance, a prediction error method was considered in [@Weerts2018b], where consistency and minimum variance properties were proven under the assumption that the network is identifiable, the disturbances are filtered white noise, and the inputs are persistently exciting and uncorrelated with the disturbances. Another work [@Haber2014] considers subspace identification of networks with a path graph topology. As we will see, the structure of the network plays a fundamental role also with respect to the question of identifiability. We follow the setup of [@Hendrickx2018], where all network nodes can be externally excited, but only a subset of nodes can be measured. Within this setup, we are interested in two identifiability problems. Firstly, we want to find conditions under which the transfer functions from a given node to its out-neighbours are identifiable. Secondly, we wonder under which conditions the transfer functions of all edges in the network are identifiable. In particular, our aim is to find graph-theoretic conditions for the above problems, that is, conditions in terms of the network structure and the locations of measured nodes. Such conditions based on the network topology are desirable since they give insight on the types of network structures that allow unique identification, and in addition may aid in the selection of measured nodes. Graph-theoretic methods have also been succesfully applied to assess other system-theoretic properties like structural controllability [@Liu2011; @Chapman2014; @Monshizadeh2014; @vanWaarde2017b; @Jia2019] and fault detection [@Rapisarda2015; @deRoo2015]. Identifiability of dynamical networks is an active research area, see e.g. [@Hendrickx2018; @Weerts2018; @vanWaarde2018; @Weerts2016; @Adebayo2012; @Hayden2017; @Nabavi2016; @vanWaarde2018b] and the references therein. The papers that are most closely related to the work presented here are [@Nabavi2016], [@vanWaarde2018b], [@Hendrickx2018], and [@vanWaarde2018], in which identifiability is also considered from graph-theoretic perspective. In [@Nabavi2016] and [@vanWaarde2018b], sufficient graph-theoretic conditions for identifiability have been presented for a class of state-space systems. In [@Hendrickx2018], graph-theoretic conditions have been established for generic identifiability. That is, conditions were given under which transfer functions in the network can be identified for “almost all" network matrices associated with the graph. The authors of [@Hendrickx2018] show that generic identifiability is equivalent to the existence of certain vertex-disjoint paths, which yields elegant conditions for generic identifiability. Inspired by the work in [@Hendrickx2018], we are interested in graph-theoretic conditions for a stronger notion, namely identifiability for all network matrices associated with the graph, a notion often referred to as global identifiability. This problem is motivated by the fact that, although generic identifiability guarantees identifiability for almost all network matrices, there are meaningful examples of network matrices that are not contained in this set of almost all network matrices. As a consequence, a situation may arise in which the system under consideration is not identifiable, even though the conditions for generic identifiability are satisfied. For an example of such a situation, we refer to Section \[sectionproblem\]. On the other hand, if the conditions derived in this paper are satisfied, then it is guaranteed that the network is identifiable for all network matrices associated with the graph. The contributions of this paper are the following. 1. We introduce the so-called graph simplification process. Based on this process, we provide necessary and sufficient conditions for the left-invertibility of certain network-related transfer matrices. 2. Using the fact that identifiability is characterized by the left-invertibility of transfer matrices [@Hendrickx2018], [@vanWaarde2018], we provide necessary and sufficient graph-theoretic conditions for identifiability based on graph simplification. We also show that these conditions can be verified by polynomial time algorithms. 3. We compare our results with the sufficient topological conditions for identifiability based on constrained vertex-disjoint paths [@vanWaarde2018]. In particular, we show that the results obtained in this paper generalize those in [@vanWaarde2018]. This paper is organized as follows. In Section \[sectionpreliminaries\] we discuss the preliminaries that are used throughout this paper. Subsequently, in Section \[sectionproblem\] we state and motivate the problem. Next, in Section \[sectionrank\] we recall rank conditions for identifiability. Sections \[sectiongraphsimplification\] and \[sectionidentifiabilitysimplification\] contain our main results. In Section \[sectiongraphsimplification\] we introduce the graph simplification process and show its relation to the left-invertibility of transfer matrices. Subsequently, in Section \[sectionidentifiabilitysimplification\] we provide graph-theoretic conditions for identifiability. Our main results are compared to previous work in Section \[sectionresultscvdp\]. Finally, Section \[sectionconclusions\] contains our conclusions. Preliminaries {#sectionpreliminaries} ============= We denote the set of natural numbers by $\mathbb{N}$, real numbers by $\mathbb{R}$, and complex numbers by $\mathbb{C}$. The set of real $m \times n$ matrices is denoted by $\mathbb{R}^{m \times n}$. The $n \times n$ identity matrix is denoted by $I_n$. If its dimension is clear, we simply write $I$. Rational functions and rational matrices ---------------------------------------- Consider a scalar variable $z$ and a rational function $f(z) = \frac{p(z)}{q(z)}$, where $p(z)$ and $q(z)$ are real polynomials and $q$ is nonzero. The function $f$ is proper if the degree of $p(z)$ is less than or equal to the degree of $q(z)$. We say $f$ is strictly proper if the degree of $p(z)$ is less than the degree of $q(z)$. An $m \times n$ matrix $A(z)$ is called rational if its entries are rational functions in the variable $z$. In addition, $A(z)$ is proper if its entries are proper rational functions. We omit the argument $z$ whenever the dependency of $A$ on $z$ is clear from the context. The normal rank of $A(z)$ is defined as $\max_{\lambda \in \mathbb{C}} \operatorname{rank}A(\lambda)$ and denoted by $\operatorname{rank}A(z)$, with slight abuse of notation. We say $A(z)$ is left-invertible if $\operatorname{rank}A(z) = n$. We denote the $(i,j)$-th entry of a matrix $A$ by $A_{ij}$. Moreover, the $j$-th column of $A$ is given by $A_{\bullet j}$. More generally, let $\mathcal{M} \subseteq \{1,2,\dots,m\}$ and $\mathcal{N} \subseteq \{1,2,\dots,n\}$. Then, $A_{\mathcal{M},\mathcal{N}}$ denotes the submatrix of $A$ containing the rows of $A$ indexed by $\mathcal{M}$ and the columns of $A$ indexed by $\mathcal{N}$. Next, consider the case that $A$ is square, i.e., $m = n$. The determinant of $A$ is denoted by $\det A$, while the adjugate of $A$ is denoted by ${\operatorname{adj}}A$. A principal submatrix of $A$ is a submatrix $A_{\mathcal{M},\mathcal{M}}$, where $\mathcal{M} \subseteq \{1,2,\dots,m\}$. The determinant of $A_{\mathcal{M},\mathcal{M}}$ is called a principal minor of $A$. Graph theory {#sectiongraphtheory} ------------ Let $\mathcal{G} = (\mathcal{V},\mathcal{E})$ be a directed graph, with vertex (or node) set $\mathcal{V} = \{1,2,\dots,n\}$ and edge set $\mathcal{E} \subseteq \mathcal{V} \times \mathcal{V}$. The graphs considered in this paper are simple, i.e., without self-loops and with at most one edge from one node to another. Consider an edge $(i,j) \in \mathcal{E}$. Then $(i,j)$ is called an outgoing edge of node $i \in \mathcal{V}$ and $j$ is called an out-neighbour of $i \in \mathcal{V}$. The set of out-neighbours of $i$ is denoted by $\mathcal{N}^+_i$. Similarly, $(i,j)$ is called an incoming edge of $j \in \mathcal{V}$ and node $i$ is called an in-neighbour of $j$. The set of in-neighbours of node $j$ is denoted by $\mathcal{N}_j^-$. For any subset $\mathcal{S} = \{v_1,v_2,\dots,v_s\} \subseteq \mathcal{V}$ we define the $s \times n$ matrix $P(\mathcal{V};\mathcal{S})$ as $P_{ij} := 1$ if $j = v_i$, and $P_{ij} := 0$ otherwise. The complement of $\mathcal{S}$ in $\mathcal{V}$ is defined as $\mathcal{S}^c := \mathcal{V} \setminus \mathcal{S}$. Moreover, the cardinality of $\mathcal{S}$ is denoted by $\|\mathcal{S}\|$. A path $\mathcal{P}$ is a set of edges in $\mathcal{G}$ of the form $\mathcal{P} = \{(v_i,v_{i+1}) \mid i = 1,2,\dots,k \} \subseteq \mathcal{E}$, where the vertices $v_1,v_2,\dots,v_{k+1}$ are distinct. The vertex $v_1$ is called a starting node of $\mathcal{P}$, while $v_{k+1}$ is the end node. The cardinality of $\mathcal{P}$ is called the length of the path. A collection of paths $\mathcal{P}_1,\mathcal{P}_2,\dots,\mathcal{P}_l$ is called vertex-disjoint if the paths have no vertex in common, that is, if for all distinct $i,j \in \{1,2,\dots,l\}$, we have that $$(u_i,w_i) \in \mathcal{P}_i, (u_j,w_j) \in \mathcal{P}_j \implies u_i,w_i,u_j,w_j \text{ are distinct}.$$ Let $\mathcal{U}, \mathcal{W} \subseteq \mathcal{V}$ be disjoint. We say there exists a path from $\mathcal{U}$ to $\mathcal{W}$ if there exist vertices $u \in \mathcal{U}$ and $w \in \mathcal{W}$ such that there exists a path in $\mathcal{G}$ with starting node $u$ and end node $w$. Similarly, we say there are $m$ vertex-disjoint paths from $\mathcal{U}$ to $\mathcal{W}$ if there exist $m$ vertex-disjoint paths[^2] in $\mathcal{G}$ with starting nodes in $\mathcal{U}$ and end nodes in $\mathcal{W}$. In the case that $\mathcal{U} \cap \mathcal{W} \neq \emptyset$, we say there exist $m$ vertex-disjoint paths from $\mathcal{U}$ to $\mathcal{W}$ if there are $\max\{0,m - \abs{\mathcal{U} \cap \mathcal{W}}\}$ vertex-disjoint paths from $\mathcal{U} \setminus \mathcal{W}$ to $\mathcal{W} \setminus \mathcal{U}$. Roughly speaking, this means that we count paths of “length zero" from every node in $\mathcal{U} \cap \mathcal{W}$ to itself. Problem statement and motivation {#sectionproblem} ================================ Let $\mathcal{G} = (\mathcal{V},\mathcal{E})$ be a simple directed graph with vertex set $\mathcal{V} = \{1,2,\dots,n\}$ and edge set $\mathcal{E} \subseteq \mathcal{V} \times \mathcal{V}$. We associate with each node $i \in \mathcal{V}$ a scalar node signal $w_i(t)$, an external excitation signal $r_i(t)$ and a disturbance signal $v_i(t)$. Then, we consider the following discrete-time dynamics: $$w_i(t) = \sum_{j \in \mathcal{N}_i^-} G_{ij}(q) w_j(t) + r_i(t) + v_i(t),$$ where $G_{ij}(z)$ is a scalar transfer function and $q$ denotes the forward shift operator defined by $q w_i(t) = w_i(t+1)$. By concatenation of the node signals, excitation signals and disturbance signals, we can write the dynamics of all nodes compactly as $w(t) = G(q) w(t) + r(t) + v(t)$, where $w, r$, and $v$ are $n$-dimensional vectors and $G(z)$ is a $n \times n$ rational matrix. In addition, we consider a measured output vector $y(t)$ of dimension $p$ that consists of the node signals of a subset $\mathcal{C} \subseteq \mathcal{V}$ of so-called measured nodes. By defining an associated binary matrix $C$ as $C := P(\mathcal{V},\mathcal{C})$, we can write this output as $y(t) = C w(t)$. Finally, by combining the equations for $w$ and $y$, we obtain the networked system $$\label{system} \begin{aligned} w(t) &= G(q) w(t) + r(t) + v(t) \\ y(t) &= C w(t). \end{aligned}$$ We call the matrix $G(z)$ the network matrix and assume that it satisfies the following properties: 1. For all $i,j \in \mathcal{V}$, the entry $G_{ij}(z)$ is a proper rational (transfer) function. 2. The function $G_{ij}(z)$ is nonzero if and only if $(j,i) \in \mathcal{E}$. A matrix $G(z)$ that satisfies this property is said to be consistent with the graph $\mathcal{G}$. 3. Every principal minor of $\lim_{z \to \infty} (I-G(z))$ is nonzero. This implies that the network model is well-posed in the sense of Definition 2.11 of [@Dankers2014]. A network matrix $G(z)$ satisfying Properties P1, P2, and P3 is called admissible. The set of all admissible network matrices is denoted by $\mathcal{A}(\mathcal{G})$. A continuous-time counterpart of can be obtained by replacing $q$ by the differential operator, hence our results will also be applicable to continuous-time systems. Besides the model , also state-space network models have received much attention (see, e.g., [@Goncalves2008; @Yuan2011; @Nabi-Abdolyousefi2012; @Hayden2017; @vanWaarde2017a]). A state-space model with scalar node dynamics can be obtained from by choosing the nonzero entries of $G$ as first-order functions [@Kivits2018]. Also more general state-space models can be found in the literature, where the node dynamics are described by general linear systems, see e.g [@Fuhrmann2015]. The model cannot capture these dynamics since the node signals $w_i$ are assumed to be scalar. The extension to the non-scalar case is therefore of interest, and will be considered for future work. For the development of this paper, it is important to distinguish between the following two concepts: - Identifiability: this is a fundamental property of the set of models of the form that captures under what conditions identification is conceptually possible. If this property is not satisfied, one cannot uniquely identify the dynamics, no matter which identification algorithm is used. Identifiability does not involve any use of data. - Identification: this involves the development of numerical algorithms for identifying the system dynamics from data. If identifiability holds then identification can be successfully performed in different ways under hypotheses on the noise and the informativity of the data [@Ljung1999]. This paper focuses on characterizations of identifiability. To explain what identifiability means in a network context, we first write in input/output form as $$y(t) = C(I-G(q)){^{-1}}r(t) + \bar{v}(t),$$ where $\bar{v}(t) := C(I-G(q)){^{-1}}v(t)$. It is well-known that the transfer matrix $C(I-G(z)){^{-1}}$ from $r$ to $y$ can be obtained from $\{r(t),y(t)\}$-data, under suitable assumptions on $r$ and $v$ [@Ljung1999]. The question of network identifiability is then the following: which entries of $G(z)$ can be uniquely reconstructed from $C(I-G(z)){^{-1}}$? In this paper we restrict our attention to the identifiability of the transfer functions outgoing a given node $i$ (i.e., identifiability of a column of $G(z)$), and to the identifiability of the entire matrix $G(z)$. A standing assumption in our work is that we know the graph structure $\mathcal{G}$ underlying the dynamical network. In recent work [@Hendrickx2018], [@Bazanella2017] the problem of identifiability has been considered from generic viewpoint. Graph-theoretic conditions were given under which certain entries of $G(z)$ can be uniquely reconstructed from $C(I-G(z)){^{-1}}$ for almost all network matrices $G$ consistent with the graph. For a formal definition of generic identifiability we refer to Definition 1 of [@Hendrickx2018]. Here, we will informally illustrate the approach of [@Hendrickx2018]. We will use the shorthand notation $T(z;G) := (I - G(z)){^{-1}}$. This means that the transfer matrix from $r$ to $y$ equals $CT$. \[example1\] Consider the graph $\mathcal{G} = (\mathcal{V},\mathcal{E})$ in Figure \[fig:graph1\]. We assume that the node signals of nodes $4$ and $5$ can be measured, that is, $\mathcal{C} = \{4,5\}$. Suppose that we want to identify the transfer functions from node $1$ to its out-neighbours, i.e., the transfer functions $G_{21}(z)$ and $G_{31}(z)$. According to Corollary 5.1 of [@Hendrickx2018], this is possible if and only if there exist two vertex-disjoint paths from $\mathcal{N}^+_1$ to $\mathcal{C}$. Note that this is the case in this example, since the edges $(2,4)$ and $(3,5)$ are two vertex-disjoint paths. To see why we can generically identify the transfer functions $G_{21}$ and $G_{31}$, we compute $CT$ as: $$CT = \begin{pmatrix} G_{42} G_{21} + G_{43} G_{31} & G_{42} & G_{43} & 1 & 0 \\ G_{52} G_{21} + G_{53} G_{31} & G_{52} & G_{53} & 0 & 1 \end{pmatrix},$$ where we omit the argument $z$. Clearly, we can uniquely obtain the transfer functions $G_{42}, G_{43}, G_{52}$, and $G_{53}$ from $CT$. Moreover, the transfer matrices $G_{21}$ and $G_{31}$ satisfy $$\label{equationex1} \begin{pmatrix} G_{42} & G_{43} \\ G_{52} & G_{53} \end{pmatrix} \begin{pmatrix} G_{21} \\ G_{31} \end{pmatrix} = \begin{pmatrix} T_{41} \\ T_{51} \end{pmatrix}.$$ Equation has a unique solution in the unknowns $G_{21}$ and $G_{31}$ if $G_{42} G_{53} - G_{43} G_{52} \neq 0$, which means that we can identify $G_{21}$ and $G_{31}$ for “almost all" $G$ consistent with $\mathcal{G}$. As mentioned before, the approach based on vertex-disjoint paths [@Hendrickx2018] gives necessary and sufficient conditions for generic identifiability. This implies that for some network matrices $G$, it might be impossible to identify the transfer functions, even though the path-based conditions are satisfied. For instance, in Example \[example1\] we cannot identify $G_{21}$ and $G_{31}$ if the network matrix $G$ is such that $G_{42} = G_{43} = G_{52} = G_{53}$. Nonetheless, scenarios in which some (or all) transfer functions in the network are equal occur frequently, for example in the study of undirected (electrical) networks [@dorfler2018], in unweighted consensus networks [@Olfati-Saber2007], and in the study of Cartesian products of graphs [@Chapman2013]. Therefore, instead of generic identifiability, we are interested in graph-theoretic conditions that guarantee identifiability for all admissible network matrices. Such a problem might seem like a simple extension of the work on generic identifiability [@Hendrickx2018]. However, to analyze strong structural network properties (for all network matrices), we typically need completely different graph-theoretic tools than the ones used in the analysis of generic network properties. For instance, in the literature on controllability, weak structural controllability is related to maximal matchings [@Liu2011], while strong structural controllability is related to zero forcing sets [@Monshizadeh2014] and constrained matchings [@Chapman2013]. To make the problem of this paper more precise, we state a few definitions. First, we are interested in conditions under which all transfer functions from a node $i$ to its out-neighbours $\mathcal{N}^+_i$ are identifiable (for any admissible network matrix $G \in \mathcal{A}(\mathcal{G})$). If this is the case, we say $(i,\mathcal{N}^+_i)$ is globally identifiable, or simply $(i,\mathcal{N}^+_i)$ is identifiable for short. \[def1\] Consider a directed graph $\mathcal{G} = (\mathcal{V},\mathcal{E})$ and let $i \in \mathcal{V}$ and $\mathcal{C} \subseteq \mathcal{V}$. Moreover, define $C = P(\mathcal{V},\mathcal{C})$. We say $(i,\mathcal{N}^+_i)$ is (globally) identifiable from $\mathcal{C}$ if the implication $$CT(z;G) = CT(z;\bar{G}) \implies G_{\bullet i}(z) = \bar{G}_{\bullet i}(z)$$ holds for all $G(z), \bar{G}(z) \in \mathcal{A}(\mathcal{G})$. In addition, we are interested in conditions under which the entire network matrix $G$ can be identified. If this is the case, we say the graph $\mathcal{G}$ is (globally) identifiable. Consider a directed graph $\mathcal{G} = (\mathcal{V},\mathcal{E})$ and let $\mathcal{C} \subseteq \mathcal{V}$ and $C = P(\mathcal{V},\mathcal{C})$. We say $\mathcal{G}$ is (globally) identifiable from $\mathcal{C}$ if the implication $$CT(z;G) = CT(z;\bar{G}) \implies G(z) = \bar{G}(z)$$ holds for all $G(z), \bar{G}(z) \in \mathcal{A}(\mathcal{G})$. The main goals of this paper are to find graph-theoretic conditions for identifiability of $(i,\mathcal{N}_i^+)$ and $\mathcal{G}$. \[problem1\] Consider a directed graph $\mathcal{G} = (\mathcal{V},\mathcal{E})$ with measured nodes $\mathcal{C} \subseteq \mathcal{V}$. Provide necessary and sufficient graph-theoretic conditions under which, respectively, $(i,\mathcal{N}_i^+)$ and $\mathcal{G}$ are identifiable from $\mathcal{C}$. Graph-theoretic conditions for global identifiability are attractive for two reasons. First, such conditions will give insight on the types of graph structures that allow identification. Secondly, they allow us to select measured nodes guaranteeing identifiability before collecting data. To deal with Problem \[problem1\], we make use of rank conditions for identifiability which we will recall in Section \[sectionrank\]. To verify such rank conditions, we introduce a novel graph-theoretic concept called the graph simplification process in Section \[sectiongraphsimplification\]. Rank conditions for identifiability {#sectionrank} =================================== First, we review some of the conditions for identifiability in terms of the normal rank of transfer matrices. For the proofs of all results in this section, we refer to [@vanWaarde2018]. Recall from Section \[sectionpreliminaries\] that $T_{\mathcal{C}, \mathcal{N}^+_i}(z;G)$ denotes the submatrix of $T$ formed by taking the rows of $T$ indexed by $\mathcal{C}$ and the columns of $T$ corresponding to $\mathcal{N}^+_i$. This means that $T_{\mathcal{C}, \mathcal{N}^+_i}(z;G)$ is a submatrix of the transfer matrix $CT(z;G)$ from $r$ to $y$, obtained by selecting the columns corresponding to $\mathcal{N}_i^+$. The following lemma (Lemma 5 of [@vanWaarde2018]) asserts that identifiability of $(i,\mathcal{N}^+_i)$ is equivalent to a rank condition on the matrix $T_{\mathcal{C}, \mathcal{N}^+_i}(z;G)$. \[lemma1\] Consider a directed graph $\mathcal{G} = (\mathcal{V},\mathcal{E})$, let $i \in \mathcal{V}$, and $\mathcal{C} \subseteq \mathcal{V}$. Then, $(i,\mathcal{N}^+_i)$ is identifiable from $\mathcal{C}$ if and only if $\operatorname{rank}T_{\mathcal{C}, \mathcal{N}^+_i}(z;G) = \| \mathcal{N}^+_i \|$ for all $G(z) \in \mathcal{A}(\mathcal{G})$. As an immediate consequence of Lemma \[lemma1\], we find conditions for the identifiability of $\mathcal{G}$ based on the normal rank of transfer matrices. This is stated in the following corollary. \[cor1\] Consider a directed graph $\mathcal{G} = (\mathcal{V},\mathcal{E})$ and let $\mathcal{C} \subseteq \mathcal{V}$. Then, $\mathcal{G}$ is identifiable from $\mathcal{C}$ if and only if $\operatorname{rank}T_{\mathcal{C}, \mathcal{N}^+_i}(z;G) = \| \mathcal{N}^+_i \|$ for all $i \in \mathcal{V}$ and all $G(z) \in \mathcal{A}(\mathcal{G})$. Although Lemma \[lemma1\] and Corollary \[cor1\] give necessary and sufficient conditions for the identifiability of respectively $(i,\mathcal{N}_i^+)$ and $\mathcal{G}$, these conditions are limited since there is no obvious method to check left-invertibility of $T_{\mathcal{C}, \mathcal{N}^+_i}(z;G)$ for an infinite number of matrices $G$. Therefore, one of the main results of this paper will be graph-theoretic conditions for the left-invertibility of $T_{\mathcal{W}, \mathcal{U}}(z;G)$, where $\mathcal{U}, \mathcal{W} \subseteq \mathcal{V}$ are any two subsets of vertices. These conditions will be introduced in the next section. The graph simplification process {#sectiongraphsimplification} ================================ In this section we provide necessary and sufficient conditions for left-invertibility of $T_{\mathcal{W},\mathcal{U}}(z;G)$ for all $G(z) \in \mathcal{A}(\mathcal{G})$, where $\mathcal{U}, \mathcal{W} \subseteq \mathcal{V}$. Loosely speaking, the idea is to simplify the graph $\mathcal{G}$ and nodes $\mathcal{W}$ in such a way that checking left-invertibility becomes easy. To give the reader some intuition for the approach, we start with the following basic lemma, which asserts that $T_{\mathcal{W},\mathcal{U}}(z;G)$ is left-invertible if $\mathcal{U} \subseteq \mathcal{W}$. \[lemmanecsys\] Consider a directed graph $\mathcal{G} = (\mathcal{V},\mathcal{E})$ and let $\mathcal{U}, \mathcal{W} \subseteq \mathcal{V}$. If $\mathcal{U} \subseteq \mathcal{W}$ then $\operatorname{rank}T_{\mathcal{W},\mathcal{U}}(z;G) = \abs{\mathcal{U}}$ for all $G(z) \in \mathcal{A}(\mathcal{G})$. The proof of Lemma \[lemmanecsys\] is postponed to Appendix \[appendixlemmanecsys\]. The condition $\mathcal{U} \subseteq \mathcal{W}$ considered in Lemma \[lemmanecsys\] is clearly not necessary for left-invertibility. One can show this using the example $\mathcal{G} = (\mathcal{V},\mathcal{E})$, where $\mathcal{V} = \{1,2\}$, $\mathcal{E} = \{(1,2)\}$, and the subsets $\mathcal{U}$ and $\mathcal{W}$ are chosen as $\mathcal{U} = \{1\}$ and $\mathcal{W} = \{2\}$. However, the main idea of the graph simplification process is to simplify $\mathcal{G}$ and to ‘move’ the nodes in $\mathcal{W}$ closer to the nodes in $\mathcal{U}$ such that the condition $\mathcal{U} \subseteq \mathcal{W}$ possibly holds after applying these operations. Of course, we cannot blindly modify the graph $\mathcal{G}$ since this would affect the left-invertibility of $T_{\mathcal{W},\mathcal{U}}(z;G)$. Instead, we will now state two lemmas in which we consider two different operations on $\mathcal{G}$ and $\mathcal{W}$ that preserve left-invertibility of $T_{\mathcal{W},\mathcal{U}}(z;G)$. We emphasize that the graph operations are introduced for analysis purposes only. Indeed, since the condition of Lemma \[lemmanecsys\] is simple to check, the graph operations should be seen as a tool to check left-invertibility of the transfer matrix of a given fixed graph $\mathcal{G}$. First, we state Lemma \[lemmaprocess1\] which asserts that left-invertibility of $T_{\mathcal{W},\mathcal{U}}(z;G)$ is unaffected by the removal of the outgoing edges of $\mathcal{W}$. \[lemmaprocess1\] Consider a directed graph $\mathcal{G} = (\mathcal{V},\mathcal{E})$ and let $\mathcal{U}, \mathcal{W} \subseteq \mathcal{V}$. Moreover, let $\bar{\mathcal{G}} = (\mathcal{V},\bar{\mathcal{E}})$ be the graph obtained from $\mathcal{G}$ by removing all outgoing edges of the nodes in $\mathcal{W}$. Then $\operatorname{rank}T_{\mathcal{W},\mathcal{U}}(z;G) = \abs{\mathcal{U}}$ for all $G(z) \in \mathcal{A}(\mathcal{G})$ if and only if $\operatorname{rank}T_{\mathcal{W},\mathcal{U}}(z;\bar{G}) = \abs{\mathcal{U}}$ for all $\bar{G}(z) \in \mathcal{A}(\bar{\mathcal{G}})$. Let $G(z) \in \mathcal{A}(\mathcal{G})$. Relabel the nodes in $\mathcal{V}$ such that $$\label{matrixG} G = \begin{pmatrix} G_{\mathcal{R},\mathcal{R}} & G_{\mathcal{R},\mathcal{W}} \\ G_{\mathcal{W},\mathcal{R}} & G_{\mathcal{W},\mathcal{W}} \end{pmatrix},$$ where $\mathcal{R} := \mathcal{V} \setminus \mathcal{W}$ and the argument $z$ has been omitted. Define the matrix $\bar{G}$ as $$\label{matrixGbar} \bar{G} = \begin{pmatrix} G_{\mathcal{R},\mathcal{R}} & 0 \\ G_{\mathcal{W},\mathcal{R}} & 0 \end{pmatrix}.$$ The matrix $\bar{G}$ is an admissible matrix consistent with $\bar{\mathcal{G}}$, i.e., $\bar{G} \in \mathcal{A}(\bar{\mathcal{G}})$. To see this, note that $\bar{G}$ satisfies Property P1. Moreover, since all outgoing edges of nodes in $\mathcal{W}$ are removed in the graph $\bar{\mathcal{G}}$, the matrix $\bar{G}$ is consistent with $\bar{\mathcal{G}}$. Hence, $\bar{G}$ satisfies property P2. Finally, to see that $\bar{G}$ satisfies Property P3, note that any principal minor of $$\label{eqlim} \lim_{z \to \infty} \begin{pmatrix} I-G_{\mathcal{R},\mathcal{R}}(z) & 0 \\ -G_{\mathcal{W},\mathcal{R}}(z) & I \end{pmatrix}$$ is either 1 or equal to a principal minor of $\lim_{z \to \infty} (I-G_{\mathcal{R},\mathcal{R}}(z))$, which is nonzero by the assumption that $G$ is admissible. We conclude that $\bar{G} \in \mathcal{A}(\bar{\mathcal{G}})$. Next, by Proposition 2.8.7 of [@Bernstein2011], the inverse of $I - G$ can be written as $$T = (I - G){^{-1}}= \begin{pmatrix} \ast & \ast \\ S(G) G_{\mathcal{W},\mathcal{R}} (I-G_{\mathcal{R},\mathcal{R}}){^{-1}}& S(G) \end{pmatrix},$$ where $S(G) := (I-G_{\mathcal{W},\mathcal{W}} - G_{\mathcal{W},\mathcal{R}}(I - G_{\mathcal{R},\mathcal{R}}){^{-1}}G_{\mathcal{R},\mathcal{W}})){^{-1}}$ denotes the inverse Schur complement of $I - G$. Using the same formula to compute the inverse of $I - \bar{G}$, we find $$\bar{T} := (I - \bar{G}){^{-1}}= \begin{pmatrix} \ast & \ast \\ G_{\mathcal{W},\mathcal{R}} (I-G_{\mathcal{R},\mathcal{R}}){^{-1}}& I \end{pmatrix}.$$ The above expressions for $T$ and $\bar{T}$ imply that $$T_{\mathcal{W},\mathcal{U}} = S(G) \bar{T}_{\mathcal{W},\mathcal{U}},$$ and because $S(G)$ has full normal rank, we obtain $$\label{rankequal} \operatorname{rank}T_{\mathcal{W},\mathcal{U}} = \operatorname{rank}\bar{T}_{\mathcal{W},\mathcal{U}}.$$ Next, we use to prove the lemma. First, to prove the ‘if’ statement, suppose that $\operatorname{rank}T_{\mathcal{W},\mathcal{U}}(z;\bar{G}) = \abs{\mathcal{U}}$ for all matrices $\bar{G} \in \mathcal{A}(\bar{\mathcal{G}})$. Let $G \in \mathcal{A}(\mathcal{G})$. Using $G$, construct the matrix $\bar{G} \in \mathcal{A}(\bar{\mathcal{G}})$ in . By hypothesis, $\operatorname{rank}T_{\mathcal{W},\mathcal{U}}(z;\bar{G}) = \abs{\mathcal{U}}$ and therefore we conclude from that $\operatorname{rank}T_{\mathcal{W},\mathcal{U}}(z;G) = \abs{\mathcal{U}}$. Subsequently, to prove the ‘only if’ statement, suppose that $\operatorname{rank}T_{\mathcal{W},\mathcal{U}}(z;G) = \abs{\mathcal{U}}$ for all $G(z) \in \mathcal{A}(\mathcal{G})$. Consider any matrix $\bar{G}(z) \in \mathcal{A}(\bar{\mathcal{G}})$ and note that $\bar{G}$ can be written in the form . Next, we choose the matrices $G_{\mathcal{R},\mathcal{W}}$ and $G_{\mathcal{W},\mathcal{W}}$ such that the matrix $G$ in is consistent with the graph $\mathcal{G}$, and such that the nonzero entries of $G_{\mathcal{R},\mathcal{W}}$ and $G_{\mathcal{W},\mathcal{W}}$ are strictly proper rational functions. This means that $G$ readily satisfies Properties P1 and P2 (see Section \[sectionproblem\]). In fact, $G$ also satisfies P3. Indeed, since $\lim_{z\to \infty} (I-G(z))$ is given by , it follows that every principal minor of $\lim_{z\to \infty} (I-G(z))$ is either 1 or equal to a principal minor of $\lim_{z \to \infty} (I-G_{\mathcal{R},\mathcal{R}})$, which is nonzero by the hypothesis that $\bar{G}(z) \in \mathcal{A}(\bar{\mathcal{G}})$. We conclude that $G$ satisfies Properties P1, P2, and P3, equivalently, $G \in \mathcal{A}(\mathcal{G})$. By hypothesis, $\operatorname{rank}T_{\mathcal{W},\mathcal{U}}(z;G) = \abs{\mathcal{U}}$ and consequently, by we conclude that $\operatorname{rank}T_{\mathcal{W},\mathcal{U}}(z;\bar{G}) = \abs{\mathcal{U}}$. This proves the lemma. \[remarkUincoming\] In similar fashion as in the proof of Lemma \[lemmaprocess1\], we can prove that all incoming edges of nodes in $\mathcal{U}$ can be removed without affecting the left-invertibility of $T_{\mathcal{W},\mathcal{U}}(z;G)$. Inspired by Lemma \[lemmaprocess1\], we wonder what type of operations we can further perform on the graph $\mathcal{G}$ and nodes $\mathcal{W}$ without affecting left-invertibility of $T_{\mathcal{W},\mathcal{U}}(z;G)$. In what follows we will show that under suitable conditions it is possible to ‘move’ the nodes in $\mathcal{W}$ closer to the nodes in $\mathcal{U}$. Here the notion of reachability in graphs will play an important role. For a subset $\mathcal{U} \subseteq \mathcal{V}$ and a node $j \in \mathcal{V} \setminus \mathcal{U}$, we say $j$ is reachable from $\mathcal{U}$ if there exists at least one path from $\mathcal{U}$ to $j$. By convention, if $j \in \mathcal{U}$ then $j$ is reachable from $\mathcal{U}$. In the following lemma, we will show that the rank of $T_{\mathcal{W},\mathcal{U}}(z;G)$ is unaffected if we replace a node $k \in \mathcal{W} \setminus \mathcal{U}$ by $j$, provided that $j$ is the only in-neighbour of $k$ that is reachable from $\mathcal{U}$. \[lemmaprocess2\] Consider a directed graph $\mathcal{G} = (\mathcal{V},\mathcal{E})$ and let $\mathcal{U},\mathcal{W} \subseteq \mathcal{V}$. Suppose that $k \in \mathcal{W} \setminus \mathcal{U}$ has exactly one in-neighbour $j \in \mathcal{N}_k^-$ that is reachable from $\mathcal{U}$. Then for all $G(z) \in \mathcal{A}(\mathcal{G})$, we have $$\operatorname{rank}T_{\mathcal{W},\mathcal{U}}(z;G) = \operatorname{rank}T_{\mathcal{\bar{W}},\mathcal{U}}(z;G),$$ where $\bar{\mathcal{W}} := (\mathcal{W} \setminus \{k\}) \cup \{j\}$. We emphasize that Lemma \[lemmaprocess2\] does not require node $k$ to have exactly one in-neighbour. In general, node $k$ may have multiple in-neighbours, but if exactly one of such neighbours is reachable from $\mathcal{U}$, we can apply Lemma \[lemmaprocess2\]. The intuition of Lemma \[lemmaprocess2\] is as follows: under the assumptions, all information from the nodes in $\mathcal{U}$ enters node $k$ via node $j$. Therefore, choosing node $k$ or node $j$ as a node in $\mathcal{W}$ does not make any difference. An interesting special case is obtained when both nodes $j$ and $k$ are contained in $\mathcal{W}$. In this case, we obtain $\bar{\mathcal{W}} = \mathcal{W} \setminus \{k\}$, that is, node $k$ can be removed from $\mathcal{W}$ without affecting the rank of $T_{\mathcal{W},\mathcal{U}}(z;G)$. By Lemma \[lemmaprocess1\], we can assume without loss of generality that the nodes in $\mathcal{W}$ have no outgoing edges. Let $G(z) \in \mathcal{A}(\mathcal{G})$. In what follows we omit the dependence of $G$ on $z$ and the dependence of $T(z;G)$ on both $z$ and $G$. Consider a vertex $v \in \mathcal{U}$. Note that $$\begin{aligned} (I - G) T &= I \\ \sum_{l = 1}^n (I - G)_{k l} T_{l v} &= 0, \label{sumGT} \end{aligned}$$ where $n := \|\mathcal{V}\|$ and follows from the fact that $k \in \mathcal{W} \setminus \mathcal{U}$ and $v \in \mathcal{U}$ are distinct. Equation implies that $$\label{eqTkv} T_{kv} = \sum_{l \in \mathcal{N}_k^{-}} G_{k l} T_{l v}.$$ Note that $j \in \mathcal{N}_k^{-}$, but possibly $\mathcal{N}_k^-$ contains other vertices. We will now prove that for all these other vertices, the corresponding transfer function $T_{lv}$ equals zero. That is, $T_{lv} = 0$ for all $l \in \mathcal{N}_k^- \setminus \{j\}$. To see this, we first observe that there does not exist a path in $\mathcal{G}$ from $v$ to $l \in \mathcal{N}_k^- \setminus \{j\}$. Indeed, suppose that there is a path $\mathcal{P}$ from $v$ to $l$. Then this path cannot contain the edge $(j,k)$, since node $k \in \mathcal{W} \setminus \mathcal{U}$ does not have any outgoing edges. This implies that there exists a path $\mathcal{P} \cup (l,k)$ from $v$ to $k$ via node $l$. This is a contradiction since by hypothesis $j$ is the only in-neighbour of $k$ that is reachable from $\mathcal{U}$. Therefore, we conclude that there does not exist a path from $v$ to $l$. By Lemma 3 of [@vandenHof2013] we conclude that $T_{lv} = 0$. This means that can be simplified as $$T_{kv} = G_{k j} T_{j v}.$$ Since $v \in \mathcal{U}$ is arbitrary, it follows that $$T_{k,\mathcal{U}} = G_{kj} T_{j,\mathcal{U}}.$$ As $G_{kj} \neq 0$, we conclude that $$\operatorname{rank}T_{\mathcal{W},\mathcal{U}} = \operatorname{rank}T_{\bar{\mathcal{W}},\mathcal{U}},$$ where $\bar{\mathcal{W}} := (\mathcal{W} \setminus \{k\}) \cup \{j\}$. This proves the lemma. From Lemma \[lemmaprocess1\] and Lemma \[lemmaprocess2\], we see that (i) we can always remove the outgoing edges of nodes in $\mathcal{W}$ and (ii) we can ‘move’ nodes in $\mathcal{W}$ closer to $\mathcal{U}$ under suitable conditions. Of course, since both operations do not affect left-invertibility of $T_{\mathcal{W},\mathcal{U}}$, we can also apply these operations multiple times consecutively. Therefore, we introduce the following process to simplify the graph $\mathcal{G}$ and move the nodes in $\mathcal{W}$. The idea of this process is to apply the above operations to the graph until no more changes are possible.\ \ \ Clearly, the graph simplification process terminates after a finite number of applications of operations 1 and 2. Indeed, operation 1 can only be applied once in a row, and a node in $\mathcal{W} \setminus \mathcal{U}$ can be ‘moved’ at most $\abs{\mathcal{V}}-1$ times which means that operation 2 can be applied only a finite number of times. In fact, it is attractive to apply the operations 1 and 2 in alternating fashion since the process will then terminate within $\abs{\mathcal{V}}$ operations of both types. This is due to the fact that if the outgoing edges of a node $j \in \mathcal{V}$ are removed, then we cannot apply operation 2 to replace a node $k$ by $j$. A graph obtained by applying the graph simplification process to $\mathcal{G}$ is called a derived graph, which we denote by $\mathcal{D}(\mathcal{G})$. Similarly, we call a vertex set obtained by applying the graph simplification process to $\mathcal{W}$ a derived vertex set, denoted by $\mathcal{D}(\mathcal{W})$. To stress the fact that $\mathcal{D}(\mathcal{G})$ and $\mathcal{D}(\mathcal{W})$ do not only depend on the graph $\mathcal{G}$ and set $\mathcal{W}$, but also on the set $\mathcal{U}$, we say that $\mathcal{D}(\mathcal{G})$ and $\mathcal{D}(\mathcal{W})$ are a derived graph of $\mathcal{G}$ and derived vertex set of $\mathcal{W}$ with respect to the set $\mathcal{U}$. We emphasize that derived graphs and derived vertex sets are not necessarily unique. In general, the derived graph and derived vertex set that are obtained from the graph simplification process depend on the order in which the operations 1 and 2 are applied, and on the order in which operation 2 is applied to the nodes in $\mathcal{W}$. However, it turns out that the non-uniqueness of derived graphs and derived vertex sets is not a problem for the application (left-invertibility) we have in mind. In fact, we will show in Theorem \[maintheorem\] that any derived graph and derived vertex set will lead to the same conclusions about left-invertibility. \[remarkcomputation\] In step 2 of the graph simplification process, we have to decide whether there exists a node $k \in \mathcal{W} \setminus \mathcal{U}$ that has exactly one in-neighbour $j \in \mathcal{N}_k^-$ which is reachable from $\mathcal{U}$. Therefore, we want to find which in-neighbours of $k$ are reachable from $\mathcal{U}$. One of the ways to do this, is to use Dijkstra’s single source shortest path (SSSP) algorithm [@Dijkstra1959], [@Orlin2010]. This algorithm computes the shortest paths (i.e., paths of minimum length) from a given source node $s$ to every other node in the graph, and returns an ‘infinite’ distance for each node which is not reachable from $s$. If we apply the SSSP algorithm to each node in $\mathcal{U}$, we obtain all nodes in $\mathcal{V}$ that are reachable from $\mathcal{U}$. Dijkstra’s SSSP algorithm has time complexity $\mathcal{O}(n + e)$, where $n = \abs{\mathcal{V}}$ and $e = \abs{\mathcal{E}}$ [@Orlin2010], and therefore we can find all nodes reachable from $\mathcal{U}$ in time complexity $\mathcal{O}(un + ue)$, where $u = \abs{\mathcal{U}}$. Once we know the nodes in $\mathcal{V}$ that are reachable from $\mathcal{U}$, we can simply check whether there exists exactly one $j \in \mathcal{N}_k^-$ that is reachable from $\mathcal{U}$. In particular, this shows that the graph simplification process can be implemented in polynomial time since both operations 1 and 2 can be implemented in polynomial time, and the graph simplification process executes at most $n$ operations of type 1 and 2 (if applied in this order). \[example5\] Consider the graph $\mathcal{G} = (\mathcal{V},\mathcal{E})$ in Figure \[fig:graph5\] and define $\mathcal{U} := \{2\}$ and $\mathcal{W} := \{5,6\}$. The goal of this example is to apply the graph simplification process to obtain a derived graph and derived vertex set. After this simplification, it will be easy to check left-invertibility of $T_{\mathcal{W},\mathcal{U}}(z;G)$. First, note that both nodes $5$ and $6$ do not have outgoing edges, so at the moment we cannot apply operation 1. However, we observe that node $6$ has exactly one in-neighbour (node $4$) that is reachable from $\mathcal{U}$. Consequently, we can replace node $6$ by node $4$ in $\mathcal{W}$ (see Figure \[fig:graph6\]). To follow up, we see that node $4$ has outgoing edges, which we can remove by applying operation 1, see Figure \[fig:graph7\]. Subsequently, node $5$ has exactly one in-neighbour that is (trivially) reachable from $\mathcal{U}$. Therefore, we replace vertex $5$ by $2$ in $\mathcal{W}$. Next, we can remove all outgoing edges of node $2$ using operation 1. These result of these two operations is depicted in Figure \[fig:graph8\]. Note that nodes $2$ and $4$ do not have any outgoing edges. Moreover, the in-neighbour $3$ of node $4$ is not reachable from node $2$, so we cannot use operation 2 to node $4$. In addition, operation 2 cannot be applied to node $2$ since $2 \in \mathcal{U}$. Therefore, the graph simplification process terminates. We conclude that the graph $\mathcal{D}(\mathcal{G})$ in Figure \[fig:graph8\] is a derived graph of $\mathcal{G}$, whereas the vertex set $\mathcal{D}(\mathcal{W}) = \{2,4\}$ is a derived vertex set of $\mathcal{W}$ (with respect to $\mathcal{U}$). This example shows the strength of the graph simplification process in the following way: since $\mathcal{U} \subseteq \mathcal{D}(\mathcal{W})$, we conclude by Lemma \[lemmanecsys\] that $T_{\mathcal{D}(\mathcal{W}),\mathcal{U}}(z;G)$ is left-invertible for all $G(z) \in \mathcal{A}(\mathcal{D}(\mathcal{G}))$. However, by Lemma \[lemmaprocess1\] and Lemma \[lemmaprocess2\], we immediately see that $T_{\mathcal{W},\mathcal{U}}(z;G)$ is left-invertible for all $G(z) \in \mathcal{A}(\mathcal{G})$. This suggests that the graph simplification process is a promising tool to study left-invertibility of transfer matrices (and hence, to study identifiability of dynamical networks). To summarize, we have seen that it is possible to remove the outgoing edges of nodes in $\mathcal{W}$ and to ‘move’ the nodes in $\mathcal{W}$ closer to $\mathcal{U}$ if certain conditions are satisfied. Since left-invertibility is preserved by both operations due to Lemmas \[lemmaprocess1\] and \[lemmaprocess2\], we see that left-invertibility of $T_{\mathcal{W},\mathcal{U}}(z;G)$ for all $G(z) \in \mathcal{A}(\mathcal{G})$ is equivalent to the left-invertibility of $T_{\mathcal{D}(\mathcal{W}),\mathcal{U}}(z;G)$ for all $G(z) \in \mathcal{A}(\mathcal{D}(\mathcal{G}))$. Using Lemma \[lemmanecsys\], this shows that the condition $\mathcal{U} \subseteq \mathcal{D}(\mathcal{W})$ is sufficient for the left-invertibility of $T_{\mathcal{W},\mathcal{U}}(z;G)$. Remarkably, the condition $\mathcal{U} \subseteq \mathcal{D}(\mathcal{W})$ turns out to be also necessary for left-invertibility of $T_{\mathcal{W},\mathcal{U}}(z;G)$. This is stated more formally in the following theorem, which is one of the main results of this paper. \[maintheorem\] Consider a directed graph $\mathcal{G} = (\mathcal{V},\mathcal{E})$ and let $\mathcal{U}, \mathcal{W} \subseteq \mathcal{V}$. Let $\mathcal{D}(\mathcal{W})$ be any derived vertex set of $\mathcal{W}$ with respect to $\mathcal{U}$. Then $\operatorname{rank}T_{\mathcal{W},\mathcal{U}}(z;G) = \abs{\mathcal{U}}$ for all matrices $G(z) \in \mathcal{A}(\mathcal{G})$ if and only if $\mathcal{U} \subseteq \mathcal{D}(\mathcal{W})$. Before we prove Theorem \[maintheorem\], we need some auxiliary results. Consider a directed graph $\mathcal{G} = (\mathcal{V},\mathcal{E})$, let $n = \abs{\mathcal{V}}$, $s = \abs{\mathcal{E}}$, and index the edges as $\mathcal{E} = \{e_1,e_2,\dots,e_s\}$. We associate with each edge $e \in \mathcal{E}$ an indeterminate $\mathsf{g}_{e}$. Moreover, we define the s-dimensional vector $$\mathsf{g} := \begin{pmatrix} \mathsf{g}_{e_1} & \mathsf{g}_{e_2} & \dots & \mathsf{g}_{e_s} \end{pmatrix}^\top,$$ which we call the indeterminate vector of $\mathcal{G}$. Next, we define the $n \times n$ matrix $\mathsf{G}$ as $$\mathsf{G}_{ji} = \begin{cases} \mathsf{g}_{e_k} & \text{if} $e_k = (i,j)$ for some $k$ \\ 0 & otherwise. \end{cases}$$ We emphasize that not all entries of $\mathsf{G}$ are indeterminates, but some are fixed zeros. Note that we write $\mathsf{G}$ in sans-serif font, to clearly distinguish between $\mathsf{G}$ and a fixed rational matrix $G(z)$. It is clear that the determinants of square submatrices of $I - \mathsf{G}$ are real polynomials in the indeterminate entries of $\mathsf{G}$, i.e., in the indeterminate vector $\mathsf{g}$. Hence, the entries of the adjugate of $I- \mathsf{G}$ are real polynomials in $\mathsf{g}$. We state the following basic lemma, which gives conditions under which an entry of ${\operatorname{adj}}(I- \mathsf{G})$ is a nonzero polynomial. \[lemmapolynomial\] Consider a directed graph ${\mathcal}{G} = ({\mathcal}{V},{\mathcal}{E})$ and let $i,j \in \mathcal{V}$. Let $\mathsf{g}$ and $\mathsf{G}$ be the indeterminate vector and matrix of ${\mathcal}{G}$, respectively, and define $\mathsf{A} := {\operatorname{adj}}(I-\mathsf{G})$. Then $\mathsf{A}_{ji}$ is a nonzero polynomial in $\mathsf{g}$ if and only if there exists a path from $i$ to $j$. Lemma \[lemmapolynomial\] follows from Proposition 5.1 of [@Hendrickx2018]. Next, we state the following basic result on polynomials. \[propositionpolynomials\] Consider $k$ nonzero real polynomials $p_i(x)$, where $i = 1,2,...,k$ and $x = (x_1,x_2,\dots,x_n)$. There exists an $\bar{x} \in \mathbb{R}^n$ such that $p_i(\bar{x}) \neq 0$ for all $i = 1,2,...,k$. \[remarkpolynomials\] Without loss of generality, we can assume that $\bar{x}$ in Proposition \[propositionpolynomials\] has only nonzero coordinates. Indeed, by continuity, if $p_i(\bar{x}) \neq 0$ for $i = 1,2,...,k$, there exists an open ball $B(\bar{x})$ around $\bar{x}$ in which $p_i(x) \neq 0$ for all $i = 1,2,...,k$ and all $x \in B(\bar{x})$. Clearly, this open ball contains a point with only nonzero coordinates. Finally, we require a proposition on rational matrices. \[propositionrationalmatrix\] Let $A(z)$ be an $m \times n$ rational matrix and assume that each row of $A(z)$ contains at least one nonzero entry. There exists a vector $b \in \mathbb{R}^n$ such that each entry of $A(z)b$ is a nonzero rational function. The proof of Proposition \[propositionrationalmatrix\] follows simply from induction on the number of rows of $A(z)$ and is therefore omitted. With these results in place, we are ready to prove Theorem \[maintheorem\]. Let $\mathcal{D}(\mathcal{G})$ and $\mathcal{D}(\mathcal{W})$ be a derived graph and derived vertex set with respect to $\mathcal{U}$ obtained from the graph simplification process. To prove the ‘if’ statement, suppose that $\mathcal{U} \subseteq \mathcal{D}(\mathcal{W})$. By Corollary \[lemmanecsys\] we find that $\operatorname{rank}T_{\mathcal{D}(\mathcal{W}),\mathcal{U}}(z;G) = \abs{\mathcal{U}}$ for all $G(z) \in \mathcal{A}(\mathcal{D}(\mathcal{G}))$. By consecutive application of Lemmas \[lemmaprocess1\] and \[lemmaprocess2\], we conclude that $\operatorname{rank}T_{\mathcal{W},\mathcal{U}}(z;G) = \abs{\mathcal{U}}$ for all $G(z) \in \mathcal{A}(\mathcal{G})$. Conversely, to prove the ‘only if’ statement, suppose that $\mathcal{U} \not\subseteq \mathcal{D}(\mathcal{W})$. We want to show that $$\operatorname{rank}T_{\mathcal{D}(\mathcal{W}),\mathcal{U}}(z;G) < \abs{\mathcal{U}} \text{ for some } G(z) \in \mathcal{A}(\mathcal{D}(\mathcal{G})).$$ Since $\mathcal{U} \not\subseteq \mathcal{D}(\mathcal{W})$, the set $\bar{\mathcal{U}} := \mathcal{U} \setminus \mathcal{D}(\mathcal{W})$ is nonempty. Furthermore, as $\mathcal{D}(\mathcal{G})$ and $\mathcal{D}(\mathcal{W})$ result from the graph simplification process, it is clear that nodes in $\mathcal{D}(\mathcal{W})$ do not have outgoing edges. In addition, each node in the set $\bar{\mathcal{W}} := \mathcal{D}(\mathcal{W}) \setminus \mathcal{U}$ has either zero or at least two in-neighbours that are reachable from $\mathcal{U}$. As nodes in $\mathcal{D}(\mathcal{W}) \cap \mathcal{U}$ have no outgoing edges, this means that each node in $\bar{\mathcal{W}}$ has either zero or at least two in-neighbours that are reachable from $\bar{\mathcal{U}}$. Finally, we assume that the nodes in $\mathcal{U}$ do not have any incoming edges, which is without loss of generality by Remark \[remarkUincoming\]. The idea of the proof is to show that $T_{\mathcal{D}(\mathcal{W}),\bar{\mathcal{U}}}(z;G) b = 0$, for some to-be-determined network matrix $G(z) \in \mathcal{A}(\mathcal{D}(\mathcal{G}))$ and nonzero vector $b$. Hence, $\operatorname{rank}T_{\mathcal{D}(\mathcal{W}),\bar{\mathcal{U}}}(z;G) < \|\bar{\mathcal{U}}\|$ and since $T_{\mathcal{D}(\mathcal{W}),\bar{\mathcal{U}}}$ is a submatrix of $T_{\mathcal{D}(\mathcal{W}),\mathcal{U}}$, it will then immediately follow that $\operatorname{rank}T_{\mathcal{D}(\mathcal{W}),\mathcal{U}}(z;G) < \abs{\mathcal{U}}$. We investigate a row $T_{w,\bar{\mathcal{U}}}(z;G)$ of the transfer matrix $T_{\mathcal{D}(\mathcal{W}),\bar{\mathcal{U}}}(z;G)$ and we distinguish two cases, namely the case that $w \in \mathcal{D}(\mathcal{W}) \cap \mathcal{U}$ and the case that $w \in \bar{\mathcal{W}}$. First, suppose that $w \in \mathcal{D}(\mathcal{W}) \cap \mathcal{U}$. This implies that $w \in \mathcal{U}$. Recall that the nodes in $\mathcal{U}$ do not have any incoming edges. Consequently, there are no paths from $v$ to $w$ for any $v \in \bar{\mathcal{U}}$. We conclude from Lemma 3 of [@vandenHof2013] that $T_{wv}(z;G) = 0$ for all $G(z) \in \mathcal{A}(\mathcal{D}(\mathcal{G}))$. Therefore, $T_{w,\bar{\mathcal{U}}}(z;G) = 0$ for all $G(z) \in \mathcal{A}(\mathcal{D}(\mathcal{G}))$. Obviously, this implies that $T_{w,\bar{\mathcal{U}}}(z;G) b = 0$ for all $G(z) \in \mathcal{A}(\mathcal{D}(\mathcal{G}))$ and all real vectors $b$. Next, we consider the second case in which $w \in \bar{\mathcal{W}}$. Let $\mathsf{G}$ denote the indeterminate matrix of $\mathcal{D}(\mathcal{G})$. In addition, define $\mathsf{A} := {\operatorname{adj}}(I - \mathsf{G})$. Then, we have $$\begin{aligned} (I - \mathsf{G}) \mathsf{A} &= \det(I-\mathsf{G})I \\ (I - \mathsf{G})_{\bar{\mathcal{W}},\mathcal{V}} \mathsf{A}_{\mathcal{V},\bar{\mathcal{U}}} &= 0, \label{varGT} \end{aligned}$$ where follows from the fact that $\bar{\mathcal{U}}$ and $\bar{\mathcal{W}}$ are disjoint. Recall that nodes in $\bar{\mathcal{W}}$ do not have any outgoing edges, and therefore $(I - \mathsf{G})_{\bar{\mathcal{W}},\bar{\mathcal{W}}} = I$. This means that we can rewrite as $$\label{varGT2} \mathsf{A}_{\bar{\mathcal{W}},\bar{\mathcal{U}}} = \mathsf{G}_{\bar{\mathcal{W}},\bar{\mathcal{W}}^c} \mathsf{A}_{\bar{\mathcal{W}}^c,\bar{\mathcal{U}}},$$ where we recall that $\bar{\mathcal{W}}^c := \mathcal{V} \setminus \bar{\mathcal{W}}$. Note that for $j \in \bar{\mathcal{W}}^c$, the column $\mathsf{G}_{\bar{\mathcal{W}},j}$ is equal to $0$ if $j$ is not an in-neighbour of any node in $\bar{\mathcal{W}}$. In addition, for any $j \in \bar{\mathcal{W}}^c$, the row $\mathsf{A}_{j,\bar{\mathcal{U}}}$ equals $0$ if there is no path from $\bar{\mathcal{U}}$ to $j$ (by Lemma \[lemmapolynomial\]). Therefore, we can rewrite as $$\label{relationA} \mathsf{A}_{\bar{\mathcal{W}},\bar{\mathcal{U}}} = \mathsf{G}_{\bar{\mathcal{W}},\mathcal{N}} \mathsf{A}_{\mathcal{N},\bar{\mathcal{U}}},$$ where $\mathcal{N} \subseteq \bar{\mathcal{W}}^c$ is characterized by the following property: we have $j \in \mathcal{N}$ if and only if $j$ is an in-neighbour of a node in $\bar{\mathcal{W}}$ and there is a path from $\bar{\mathcal{U}}$ to $j$. By definition of the adjugate, the entries of $\mathsf{A}_{\mathcal{N},\bar{\mathcal{U}}}$ are polynomials in the indeterminate entries of $\mathsf{G}$. We claim that the indeterminate entries of $\mathsf{G}_{\bar{\mathcal{W}},\mathcal{N}}$ do not appear in any entry of $\mathsf{A}_{\mathcal{N},\bar{\mathcal{U}}}$, that is, $\mathsf{A}_{\mathcal{N},\bar{\mathcal{U}}}$ is independent of the indeterminate entries of $\mathsf{G}_{\bar{\mathcal{W}},\mathcal{N}}$. For the sake of clarity, we postpone the proof of this claim to the end. For now, we assume that $\mathsf{A}_{\mathcal{N},\bar{\mathcal{U}}}$ is independent of the indeterminate entries of $\mathsf{G}_{\bar{\mathcal{W}},\mathcal{N}}$. By definition, there is a path from $\bar{\mathcal{U}}$ to each node in $\mathcal{N}$. Let $\mathcal{N} = \{n_1,n_2,\dots, n_{r}\}$, where $r = \abs{\mathcal{N}}$. Then, for each node $n_i \in \mathcal{N}$, there exists a node $u_i \in \bar{\mathcal{U}}$ such that $\mathsf{A}_{n_i,u_i}$ is a nonzero polynomial in the indeterminate entries of $\mathsf{G}$ (by Lemma \[lemmapolynomial\]). We emphasize that $u_i$ and $u_j$ are not necessarily distinct. We focus on the $r$ nonzero polynomials $$\label{rpolynomials} \mathsf{A}_{n_1,u_1}, \mathsf{A}_{n_2,u_2}, \dots, \mathsf{A}_{n_r,u_r}.$$ The idea is to apply Proposition \[propositionpolynomials\] and Remark \[remarkpolynomials\] to these $r$ polynomials. By Remark \[remarkpolynomials\], we can substitute nonzero real numbers for the indeterminate entries of $\mathsf{G}$ such that all $r$ polynomials evaluate to nonzero real numbers. Since the polynomials are independent of the indeterminate entries of $\mathsf{G}_{\bar{\mathcal{W}},\mathcal{N}}$, we do not have to fix the entries of $\mathsf{G}_{\bar{\mathcal{W}},\mathcal{N}}$. In addition, it is possible to substitute strictly proper functions in $z$ for the indeterminate entries of $\mathsf{G}$ (except for entries of $\mathsf{G}_{\bar{\mathcal{W}},\mathcal{N}}$) such that the polynomials evaluate to nonzero rational functions. Indeed, one can simply choose all indeterminate entries of $\mathsf{G}$ as nonzero real numbers as before, and then divide all of these real numbers by $z$. To summarize the progress so far, we have substituted strictly proper functions for the indeterminate entries of $\mathsf{G}$ (except for the entries of $\mathsf{G}_{\bar{\mathcal{W}},\mathcal{N}}$) such that the polynomials evaluate to nonzero rational functions. Note that this implies that the matrix $\mathsf{A}_{\mathcal{N},\bar{\mathcal{U}}}$ evaluates to a rational matrix, which we denote by $A_{\mathcal{N},\bar{\mathcal{U}}}(z)$ from now on. Since each row of $A_{\mathcal{N},\bar{\mathcal{U}}}(z)$ contains a nonzero rational function, by Proposition \[propositionrationalmatrix\] there exists a nonzero real vector $b$ such that $A_{\mathcal{N},\bar{\mathcal{U}}}(z) b$ has only nonzero rational entries. Subsequently, we will choose the indeterminate entries of $\mathsf{G}_{\bar{\mathcal{W}},\mathcal{N}}$ such that $\mathsf{G}_{\bar{\mathcal{W}},\mathcal{N}} A_{\mathcal{N},\bar{\mathcal{U}}}(z) b = 0$. Recall that the nodes in $\bar{\mathcal{W}}$ either have zero or at least two in-neighbours from the set $\mathcal{N}$. If a node $w \in \bar{\mathcal{W}}$ has no in-neighbours, then $\mathsf{G}_{w,\mathcal{\mathcal{N}}} = 0$, and therefore clearly $\mathsf{G}_{w,\mathcal{N}} A_{\mathcal{N},\bar{\mathcal{U}}}(z) b = 0$. If a node $w \in \bar{\mathcal{W}}$ has at least two in-neighbours, say $n_1,n_2,\dots,n_p \in \mathcal{N}$, then we substitute strictly proper functions for the indeterminate entries $\mathsf{G}_{w,n_1}, \mathsf{G}_{w,n_2}, \dots, \mathsf{G}_{w,n_p}$ so that $\mathsf{G}_{w,\mathcal{N}} A_{\mathcal{N},\bar{\mathcal{U}}}(z) b = 0$. Note that this is possible since the vector $A_{\mathcal{N},\bar{\mathcal{U}}}(z) b$ has only nonzero rational entries. To conclude, we have substituted strictly proper functions for the indeterminate entries of $\mathsf{G}$ which yields a matrix which we denote by $G(z)$. The adjugate of $I-G(z)$ is denoted by $A(z) = {\operatorname{adj}}(I-G(z))$. We have shown that $G_{\bar{\mathcal{W}},\mathcal{N}}(z) A_{\mathcal{N},\bar{\mathcal{U}}}(z) b = 0$. By , this yields $A_{\bar{\mathcal{W}},\bar{\mathcal{U}}}(z) b = 0$. Note that $\det (I - G(z))$ is nonzero since all nonzero entries of $G$ are strictly proper functions. Therefore, $$T(z;G) = \frac{1}{\det(I-G(z))} A(z),$$ from which we find that $T_{\bar{\mathcal{W}},\bar{\mathcal{U}}}(z;G) b = 0$. Consequently, $T_{\mathcal{D}(\mathcal{W}),\bar{\mathcal{U}}}(z;G) b = 0$, and $\operatorname{rank}T_{\mathcal{D}(\mathcal{W}),\bar{\mathcal{U}}}(z;G) < \abs{\bar{\mathcal{U}}}$. Therefore, we conclude that $\operatorname{rank}T_{\mathcal{D}(\mathcal{W}),\mathcal{U}}(z;G) < \abs{\mathcal{U}}$. We still have to show that $G(z)$ is admissible, i.e., $G(z) \in \mathcal{A}(\mathcal{D}(\mathcal{G}))$. Since the indeterminate matrix $\mathsf{G}$ is consistent with the graph $\mathcal{D}(\mathcal{G})$ and we substituted (nonzero) strictly proper functions for each indeterminate entry of $\mathsf{G}$, the matrix $G(z)$ readily satisfies Properties P1 and P2. In addition, since all nonzero entries of $G(z)$ are strictly proper, we obtain $ \lim_{z \to \infty} I - G(z) = I$, and hence, $G(z)$ also satisfies Property P3. We conclude that $\operatorname{rank}T_{\mathcal{D}(\mathcal{W}),\mathcal{U}}(z;G) < \abs{\mathcal{U}}$ for some $G(z) \in \mathcal{A}(\mathcal{D}(\mathcal{G}))$. Finally, by consecutive application of Lemmas \[lemmaprocess1\] and \[lemmaprocess2\], we conclude that $\operatorname{rank}T_{\mathcal{W},\mathcal{U}}(z;G) < \abs{\mathcal{U}}$ for some $G(z) \in \mathcal{A}(\mathcal{G})$. Recall that we have so far assumed that $\mathsf{A}_{\mathcal{N},\bar{\mathcal{U}}}$ is independent of the indeterminate entries of $\mathsf{G}_{\bar{\mathcal{W}},\mathcal{N}}$. It remains to be shown that this is true. To this end, label the nodes in $\mathcal{V}$ such that $\mathsf{G}$ can be written as $$\begin{aligned} \mathsf{G} &= \begin{pmatrix} \mathsf{G}_{\bar{\mathcal{W}}^c,\bar{\mathcal{W}}^c} & \mathsf{G}_{\bar{\mathcal{W}}^c,\bar{\mathcal{W}}} \\ \mathsf{G}_{\bar{\mathcal{W}},\bar{\mathcal{W}}^c} & \mathsf{G}_{\bar{\mathcal{W}},\bar{\mathcal{W}}} \end{pmatrix} \\ &= \begin{pmatrix} \mathsf{G}_{\bar{\mathcal{W}}^c,\bar{\mathcal{W}}^c} & 0 \\ \mathsf{G}_{\bar{\mathcal{W}},\bar{\mathcal{W}}^c} & 0 \end{pmatrix}, \label{eqWc2} \end{aligned}$$ where follows from the fact that nodes in $\bar{\mathcal{W}}$ have no outgoing edges. This implies that $$I - \mathsf{G} = \begin{pmatrix} I - \mathsf{G}_{\bar{\mathcal{W}}^c,\bar{\mathcal{W}}^c} & 0 \\ -\mathsf{G}_{\bar{\mathcal{W}},\bar{\mathcal{W}}^c} & I \end{pmatrix},$$ and therefore $$\label{formulaadj} \mathsf{A} = {\operatorname{adj}}(I - \mathsf{G}) = \begin{pmatrix} {\operatorname{adj}}(I - \mathsf{G}_{\bar{\mathcal{W}}^c,\bar{\mathcal{W}}^c}) & 0 \\ * & * \end{pmatrix}.$$ Since the entries of $\mathsf{G}_{\bar{\mathcal{W}}^c,\bar{\mathcal{W}}^c}$ are independent of the indeterminate entries of $\mathsf{G}_{\bar{\mathcal{W}},\bar{\mathcal{W}}^c}$, we conclude from that the matrix $\mathsf{A}_{\bar{\mathcal{W}}^c,\bar{\mathcal{W}}^c} = {\operatorname{adj}}(I - \mathsf{G}_{\bar{\mathcal{W}}^c,\bar{\mathcal{W}}^c})$ is independent of the indeterminate entries of $\mathsf{G}_{\bar{\mathcal{W}},\bar{\mathcal{W}}^c}$. Now, to prove the claim, note that $\bar{\mathcal{U}}$ and $\bar{\mathcal{W}}$ are disjoint by definition, and therefore $\bar{\mathcal{U}} \subseteq \bar{\mathcal{W}}^c$. In addition, we have $\mathcal{N} \subseteq \bar{\mathcal{W}}^c$. Therefore, the matrix $\mathsf{A}_{\mathcal{N},\bar{\mathcal{U}}}$ is a submatrix of $\mathsf{A}_{\bar{\mathcal{W}}^c,\bar{\mathcal{W}}^c}$. Furthermore, we see that $\mathsf{G}_{\bar{\mathcal{W}},\mathcal{N}}$ is a submatrix of $\mathsf{G}_{\bar{\mathcal{W}},\bar{\mathcal{W}}^c}$ by using the fact that $\mathcal{N} \subseteq \bar{\mathcal{W}}^c$. We conclude that the entries of $\mathsf{A}_{\mathcal{N},\bar{\mathcal{U}}}$ are independent of the indeterminate entries of $\mathsf{G}_{\bar{\mathcal{W}},\mathcal{N}}$. Identifiability and graph simplification {#sectionidentifiabilitysimplification} ======================================== In this section we use Theorem \[maintheorem\] to provide solutions to the identifiability problems introduced in Section \[sectionproblem\]. Specifically, the following theorem follows from Theorem \[maintheorem\] and Lemma \[lemma1\] and states necessary and sufficient graph-theoretic conditions for identifiability of $(i,\mathcal{N}^+_i)$. \[mainresultiNi\] Consider a directed graph $\mathcal{G} = (\mathcal{V},\mathcal{E})$, let $i \in \mathcal{V}$ and $\mathcal{C} \subseteq \mathcal{V}$. Moreover, let $\mathcal{D}(\mathcal{C})$ be any derived vertex set of $\mathcal{C}$ with respect to $\mathcal{N}^+_i$. Then $(i,\mathcal{N}^+_i)$ is identifiable from $\mathcal{C}$ in $\mathcal{G}$ if and only if $\mathcal{N}^+_i \subseteq \mathcal{D}(\mathcal{C})$. Consider the graph in Figure \[fig:graph5\]. We wonder whether $(1,\mathcal{N}^+_1)$ is identifiable. Note that we have $\mathcal{N}_1^+ = \{2\}$. The set of measured nodes is $\mathcal{C} = \{5,6\}$. As shown in Example \[example5\], a derived vertex set of $\mathcal{C}$ with respect to $\mathcal{N}_1^+$ is given by $\mathcal{D}(\mathcal{C}) = \{2,4\}$. Since $\{2\} \subseteq \mathcal{D}(\mathcal{C})$, we conclude by Theorem \[mainresultiNi\] that $(1,\mathcal{N}_1^+)$ is identifiable. In other words, we can uniquely reconstruct $G_{21}(z)$ from the transfer matrix $CT(z;G)$. This approach shows the strength of our approach. Indeed, note that to check identifiability, we do not have to verify Definition \[def1\] directly. Also, we do not have to compute $CT(z;G) = C(I-G(z))^{-1}$ and verify its rank for all $G(z) \in \mathcal{A}(\mathcal{G})$, which is required to check the condition of Lemma \[lemma1\]. By definition of the graph simplification process, we have that $\|\mathcal{D}(\mathcal{C})\| \leq \|\mathcal{C}\|$. Hence, it follows from Theorem \[mainresultiNi\] that identifiability of $(i,\mathcal{N}_i^+)$ implies that the number of measured nodes is greater or equal to the number of out-neighbours of node $i$. \[cormainNi\] Consider a directed graph $\mathcal{G} = (\mathcal{V},\mathcal{E})$, let $i \in \mathcal{V}$ and $\mathcal{C} \subseteq \mathcal{V}$. If $(i,\mathcal{N}^+_i)$ is identifiable from $\mathcal{C}$ in $\mathcal{G}$ then $\|\mathcal{N}^+_i\| \leq \|\mathcal{C}\|$. The next result gives necessary and sufficient graph-theoretic conditions under which the entire graph $\mathcal{G}$ is identifiable. This result is a corollary of Theorem \[mainresultiNi\] but is stated as a theorem due to its importance. \[mainresultG\] Consider a directed graph $\mathcal{G} = (\mathcal{V},\mathcal{E})$ and let $\mathcal{C} \subseteq \mathcal{V}$. Then $\mathcal{G}$ is identifiable from $\mathcal{C}$ if and only if for all $i \in \mathcal{V}$, we have $\mathcal{N}^+_i \subseteq \mathcal{D}(\mathcal{C})$, where $\mathcal{D}(\mathcal{C})$ is any derived vertex set of $\mathcal{C}$ with respect to $\mathcal{N}^+_i$. We emphasize that the derived set $\mathcal{D}(\mathcal{C})$ of $\mathcal{C}$ depends on the choice of neighbour set $\mathcal{N}_i^+$, and hence, for each node $i \in \mathcal{V}$ we have to compute the derived set of $\mathcal{C}$ with respect to $\mathcal{N}_i^+$. Comparison to results based on constrained vertex-disjoint paths {#sectionresultscvdp} ================================================================ In the previous section we established necessary and sufficient graph-theoretic conditions for the identifiability of respectively $(i,\mathcal{N}_i^+)$ and $\mathcal{G}$. The purpose of the current section is to compare these results to the ones based on so-called constrained vertex-disjoint paths [@vanWaarde2018]. We first recall the definition in what follows. Let $\mathcal{G} = (\mathcal{V},\mathcal{E})$ be a directed graph. Consider a set of $m$ vertex-disjoint paths in $\mathcal{G}$ with starting nodes $\bar{\mathcal{U}} \subseteq \mathcal{V}$ and end nodes $\bar{\mathcal{W}} \subseteq \mathcal{V}$. We say that the set of vertex-disjoint paths is constrained if it is the only set of $m$ vertex-disjoint paths from $\bar{\mathcal{U}}$ to $\bar{\mathcal{W}}$. Next, let $\mathcal{U}, \mathcal{W} \subseteq \mathcal{V}$ be disjoint subsets of vertices. We say that there exists a constrained set of $m$ vertex-disjoint paths from $\mathcal{U}$ to $\mathcal{W}$ if there exists a constrained set of $m$ vertex-disjoint paths in $\mathcal{G}$ with starting nodes $\bar{\mathcal{U}} \subseteq \mathcal{U}$ and end nodes $\bar{\mathcal{W}} \subseteq \mathcal{W}$. In the case that $\mathcal{U} \cap \mathcal{W} \neq \emptyset$, we say that there is a constrained set of $m$ vertex-disjoint paths from $\mathcal{U}$ to $\mathcal{W}$ if there exists a constrained set of $\max\{0,m - \abs{\mathcal{U} \cap \mathcal{W}}\}$ vertex-disjoint paths from $\mathcal{U} \setminus \mathcal{W}$ to $\mathcal{W} \setminus \mathcal{U}$. Note that for a set of $m$ vertex-disjoint paths from $\mathcal{U}$ to $\mathcal{W}$ to be constrained, we do not require the existence of a unique set of $m$ vertex-disjoint paths from $\mathcal{U}$ to $\mathcal{W}$. In fact, we only require the existence of a unique set of vertex-disjoint paths between the starting nodes $\bar{\mathcal{U}}$ of the paths and the end nodes $\bar{\mathcal{W}}$. We will illustrate the definition of constrained vertex-disjoint paths in Example \[example2\]. The notion of constrained vertex-disjoint paths is strongly related to the notion of constrained matchings in bipartite graphs [@Hershkowitz1993]. In fact, a constrained matching can be seen as a special case of a constrained set of vertex-disjoint paths where all paths are of length one. \[example2\] Consider the graph $\mathcal{G} = (\mathcal{V},\mathcal{E})$ in Figure \[fig:graph2\]. Moreover, consider the subsets of vertices $\mathcal{U} := \{2,3\}$ and $\mathcal{W} := \{6,7,8\}$. Clearly, the paths $\{(2,4), (4,6)\}$ and $\{(3,5), (5,7)\}$ form a set of two vertex-disjoint paths from $\mathcal{U}$ to $\mathcal{W}$. In fact, this set of vertex-disjoint paths is constrained since there does not exist another set of two vertex-disjoint paths from $\bar{\mathcal{U}} = \{2,3\}$ to $\bar{\mathcal{W}} = \{6,7\}$. Therefore, there exists a constrained set of two vertex-disjoint paths from $\mathcal{U}$ to $\mathcal{W}$. Note that there are also other sets of vertex-disjoint paths from $\mathcal{U}$ to $\mathcal{W}$. For example, the paths $\{(2,4),(4,7)\}$ and $\{(3,5),(5,8)\}$ also form a set of two vertex-disjoint paths. However, this set of vertex-disjoint paths is not constrained. To see this, note that we have another set of vertex-disjoint paths from $\bar{\mathcal{U}} = \{2,3\}$ to $\bar{\mathcal{W}} = \{7,8\}$, namely the set consisting of the paths $\{(2,4),(4,8)\}$ and $\{(3,5),(5,7)\}$. In the following theorem, we recall the main result presented in [@vanWaarde2018], which relates the notion of constrained vertex-disjoint paths and identifiability of $(i,\mathcal{N}_i^+)$. \[theoremnecsys\] Consider a directed graph $\mathcal{G} = (\mathcal{V},\mathcal{E})$, let $i \in \mathcal{V}$ and $\mathcal{C} \subseteq \mathcal{V}$. If there exists a constrained set of $\| \mathcal{N}_i^+ \|$ vertex-disjoint paths from $\mathcal{N}_i^+$ to $\mathcal{C}$ then $(i,\mathcal{N}_i^+)$ is identifiable from $\mathcal{C}$. The proof of Theorem \[theoremnecsys\] can be found in [@vanWaarde2018] (see Theorem 13). A natural question to ask is whether the condition given in Theorem \[theoremnecsys\] is also necessary for identifiability. It turns out that this is not the case, as demonstrated next. \[example4\] In this example, we revisit the graph $\mathcal{G} = (\mathcal{V},\mathcal{E})$ in Figure \[fig:graph5\]. Suppose that we are interested in the identifiability of $(1,\mathcal{N}_1^+)$, i.e., in the identifiability of the transfer function corresponding to the edge $(1,2)$. The set of measured nodes is given by $\mathcal{C} = \{5,6\}$. The purpose of this example is to show that Theorem \[theoremnecsys\] is not necessary, i.e., we have to show that $(1,\mathcal{N}_1^+)$ is identifiable even though there does not exist a constrained set of one (vertex-disjoint) path from $\mathcal{N}_1^+$ to $\mathcal{C}$. Note that $\mathcal{N}_1^+ = \{2\}$ and that there are three different paths from $2$ to $5$. In addition, there are two different paths from node $2$ to node $6$. This implies that there does not exist a constrained set of one (vertex-disjoint) path from $\mathcal{N}_1^+$ to $\mathcal{C}$. Nonetheless, we can show that $(1,\mathcal{N}_1^+)$ is identifiable. The easiest way to show this is by noting that we already proved in Example \[example5\] that $\mathcal{N}_1^+ \subseteq \mathcal{D}(\mathcal{C})$, where $\mathcal{D}(\mathcal{C})$ is a derived vertex set of $\mathcal{C}$. Hence, by Theorem \[mainresultiNi\] we conclude that $(1,\mathcal{N}_1^+)$ is identifiable. However, to gain a bit more insight we will prove identifiability of $(1,\mathcal{N}_1^+)$ by inspection of the transfer matrix $T_{\mathcal{C},\mathcal{N}_1^+}(z;G)$. For any $G(z) \in \mathcal{A}(\mathcal{G})$, we obtain $$\label{TCN} T_{\mathcal{C},\mathcal{N}_1^+} = \begin{pmatrix} G_{52} + G_{54}(G_{42} + G_{43}G_{32}) \\ G_{64}(G_{42} + G_{43}G_{32}) \end{pmatrix},$$ where we omitted the argument $z$. If $G_{42} + G_{43}G_{32} \neq 0$ then $G_{64}(G_{42} + G_{43}G_{32}) \neq 0$ and therefore $\operatorname{rank}T_{\mathcal{C},\mathcal{N}_1^+} = 1$. If $G_{42} + G_{43}G_{32} = 0$, we see that $G_{52} + G_{54}(G_{42} + G_{43}G_{32}) = G_{52} \neq 0$ so also in this case $\operatorname{rank}T_{\mathcal{C},\mathcal{N}_1^+} = 1$. We conclude that $\operatorname{rank}T_{\mathcal{C},\mathcal{N}_1^+} = 1$ for all admissible network matrices, which means that $(1,\mathcal{N}_1^+)$ is identifiable by Lemma \[lemma1\]. Example \[example4\] also gives some intuition for the fact that Theorem \[theoremnecsys\] is not necessary for identifiability. Indeed, the condition based on constrained vertex-disjoint paths guarantees that a square submatrix of $T_{\mathcal{C},\mathcal{N}_i^+}(z;G)$ is invertible for all admissible $G$, where the columns and rows of this submatrix are indexed by the starting nodes and end nodes of the paths, respectively [@vanWaarde2018]. However, as can be seen from , the matrix $T_{\mathcal{C},\mathcal{N}_i^+}(z;G)$ might be left-invertible for all admissible $G$, even though there does not exist a square $\abs{\mathcal{N}_i^+} \times \abs{\mathcal{N}_i^+}$ submatrix of $T_{\mathcal{C},\mathcal{N}_i^+}(z;G)$ that is invertible for all admissible $G$. In general, the particular square submatrix of $T_{\mathcal{C},\mathcal{N}_i^+}(z;G)$ that is invertible depends on the network matrix $G$. Interestingly, we can use the general theory developed in this paper to show that the condition of Theorem \[theoremnecsys\] is necessary and sufficient in the special case that $T_{\mathcal{C},\mathcal{N}_i^+}(z;G)$ is square itself (a proof is given in Appendix \[appendixtheoremnecsuf\]). \[theoremnecsuf\] Consider a directed graph $\mathcal{G} = (\mathcal{V},\mathcal{E})$. Let $i \in \mathcal{V}$ and $\mathcal{C} \subseteq \mathcal{V}$ be such that $\abs{\mathcal{C}} = \abs{\mathcal{N}_i^+}$. Then, $(i,\mathcal{N}_i^+)$ is identifiable if and only if there exists a constrained set of $\abs{\mathcal{N}_i^+}$ vertex-disjoint paths from $\mathcal{N}_i^+$ to $\mathcal{C}$. The main message of this section is that the conditions in terms of constrained vertex-disjoint paths [@vanWaarde2018] are only necessary and sufficient in the special case that $\abs{\mathcal{N}_i^+} = \abs{\mathcal{C}}$. This case is quite particular, especially if one is interested in identifiability of the entire network. In the latter situation, Theorem \[theoremnecsuf\] can only be applied if the number of out-neighbours of each node is equal to the number of measured nodes, which is very restrictive. Therefore, we conclude that the necessary and sufficient conditions for identifiability based on graph simplification are much more general. Additional advantages of the conditions based on the graph simplification process are that they are conceptually simpler and appealing from computational point of view, see Remark \[remarkcomputation\]. Conclusions {#sectionconclusions} =========== In this paper we have considered the problem of identifiability of dynamical networks for which interactions between nodes are modelled by transfer functions. We have been interested in graph-theoretic conditions for two identifiability problems. First, we wanted to find conditions under which the transfer functions of all outgoing edges of a given node are identifiable. Secondly, we have been interested in conditions under which all transfer functions in the network are identifiable. It is known that these problems are equivalent to the left-invertibility of certain transfer matrices for all networked matrices associated with the graph [@Hendrickx2018], [@vanWaarde2018]. However, the downside of such rank conditions is that it is not clear how to check the rank of a transfer matrix for an infinite number of network matrices. Therefore, as our first contribution, we have provided a necessary and sufficient graph-theoretic condition under which a transfer matrix has full column rank for all network matrices. To this end, we have introduced a new concept called the graph simplification process. The idea of this process is to apply simplifying operations to the graph, after which left-invertibility can be verified by simply checking a set inclusion. Based on the graph simplification process, we have given necessary and sufficient conditions for identifiability. Notably, we have shown that our conditions can be verified by polynomial time algorithms. Finally, we have shown that our results generalize existing sufficient conditions based on constrained vertex-disjoint paths [@vanWaarde2018]. It is interesting to observe that our topological conditions for global identifiability are quite different from the path-based conditions for generic identifiability [@Hendrickx2018]. This is analogous to the controllability literature, where it was shown that weak structural controllability can be characterized in terms of maximal matchings [@Liu2011], while strong structural controllability was characterized using a (different) graph-theoretic concept called zero forcing [@Monshizadeh2014]. For future work, it would be interesting to consider a minimum sensor placement problem. The goal in such a problem is to find sets of measured nodes of minimum cardinality such that the entire network is identifiable. Proof of Lemma \[lemmanecsys\] {#appendixlemmanecsys} ------------------------------ Without loss of generality, we assume that the nodes in $\mathcal{W}$ do not have outgoing edges (see Lemma \[lemmaprocess1\]). Since $\mathcal{U} \subseteq \mathcal{W}$, the nodes in $\mathcal{U}$ do not have outgoing edges. We now relabel the nodes in $\mathcal{G}$ such that $G(z) \in \mathcal{A}(\mathcal{G})$ can be written as $$G = \begin{pmatrix} G_{\mathcal{U},\mathcal{U}} & G_{\mathcal{U},\mathcal{U}^c} \\ G_{\mathcal{U}^c,\mathcal{U}} & G_{\mathcal{U}^c,\mathcal{U}^c} \end{pmatrix} = \begin{pmatrix} 0 & G_{\mathcal{U},\mathcal{U}^c} \\ 0 & G_{\mathcal{U}^c,\mathcal{U}^c} \end{pmatrix},$$ where we omitted the argument $z$, and where the zeros are present due to the fact that nodes in $\mathcal{U}$ do not have outgoing edges. Consequently, we obtain $$T = (I - G){^{-1}}= \begin{pmatrix} I & -G_{\mathcal{U},\mathcal{U}^c} \\ 0 & I-G_{\mathcal{U}^c,\mathcal{U}^c} \end{pmatrix}{^{-1}}= \begin{pmatrix} I & * \\ 0 & * \end{pmatrix},$$ and therefore, $T_{\mathcal{U},\mathcal{U}} = I$. Hence, $T_{\mathcal{U},\mathcal{U}}$ has full rank for all $G(z) \in \mathcal{A}(\mathcal{G})$ and we conclude that $T_{\mathcal{W},\mathcal{U}}$ has rank $\abs{\mathcal{U}}$ for all $G(z) \in \mathcal{A}(\mathcal{G})$. Proof of Theorem \[theoremnecsuf\] {#appendixtheoremnecsuf} ---------------------------------- To prove Theorem \[theoremnecsuf\], we will first state two lemmas. Under the assumption that $\abs{\mathcal{U}} = \abs{\mathcal{W}}$, the following lemma asserts that the existence of a set of constrained vertex-disjoint paths from $\mathcal{U}$ to $\mathcal{W}$ is preserved by operation 1. \[lemmacp1\] Let $\mathcal{G} = (\mathcal{V},\mathcal{E})$ be a directed graph and consider $\mathcal{U}, \mathcal{W} \subseteq \mathcal{V}$ such that $\abs{\mathcal{U}} = \abs{\mathcal{W}}$. Moreover, let $\bar{\mathcal{G}} = (\mathcal{V},\bar{\mathcal{E}})$ be the graph obtained from $\mathcal{G}$ by removing all outgoing edges of the nodes in $\mathcal{W}$. There exists a constrained set of $\abs{\mathcal{U}}$ vertex-disjoint paths from $\mathcal{U}$ to $\mathcal{W}$ in $\mathcal{G}$ if and only if there exists a constrained set of $\abs{\mathcal{U}}$ vertex-disjoint paths from $\mathcal{U}$ to $\mathcal{W}$ in $\bar{\mathcal{G}}$. The lemma follows from the following important observation: if $\abs{\mathcal{U}} = \abs{\mathcal{W}}$, then a set of $\abs{\mathcal{U}}$ vertex-disjoint paths from $\mathcal{U}$ to $\mathcal{W}$ does not contain any outgoing edge of a node in $\mathcal{W}$. Indeed, if a path $\mathcal{P}$ from $\mathcal{U}$ to $\mathcal{W}$ in such a set of vertex-disjoint paths contains an edge $(w,v)$, where $w \in \mathcal{W}$ and $v \in \mathcal{V}$, then the path $\mathcal{P}$ contains at least two vertices in $\mathcal{W}$ (namely $w$ and the end node). This means that $\mathcal{P}$ is contained in a set of at most $\abs{\mathcal{U}} - 1$ vertex disjoint paths from $\mathcal{U}$ to $\mathcal{W}$. However, this is a contradiction since we assumed that $\mathcal{P}$ was contained in a set of $\abs{\mathcal{U}}$ vertex-disjoint paths from $\mathcal{U}$ to $\mathcal{W}$. Next, we prove the ‘if’ statement. Suppose that there exists a constrained set $\mathcal{S}$ of $\abs{\mathcal{U}}$ vertex-disjoint paths from $\mathcal{U}$ to $\mathcal{W}$ in $\bar{\mathcal{G}}$. Then $\mathcal{S}$ is also a set of $\abs{\mathcal{U}}$ vertex-disjoint paths from $\mathcal{U}$ to $\mathcal{W}$ in $\mathcal{G}$. We want to prove that $\mathcal{S}$ is constrained (in the graph $\mathcal{G}$). Therefore, suppose on the contrary that there exists another set $\bar{\mathcal{S}}$ of $\abs{\mathcal{U}}$ vertex-disjoint paths from $\mathcal{U}$ to $\mathcal{W}$ in $\mathcal{G}$. By the above discussion, we know that no path in $\bar{\mathcal{S}}$ contains an outgoing edge of a node in $\mathcal{W}$. Therefore, $\bar{\mathcal{S}}$ is a set of $\abs{\mathcal{U}}$ vertex-disjoint paths from $\mathcal{U}$ to $\mathcal{W}$ in $\bar{\mathcal{G}}$. As such, we conclude that $\bar{\mathcal{S}} = \mathcal{S}$. In other words, $\mathcal{S}$ is a constrained set of $\abs{\mathcal{U}}$ vertex-disjoint paths from $\mathcal{U}$ to $\mathcal{W}$ in $\mathcal{G}$. Conversely, to prove the ‘only if’ statement, suppose that there exists a constrained set $\mathcal{S}$ of $\abs{\mathcal{U}}$ vertex-disjoint paths from $\mathcal{U}$ to $\mathcal{W}$ in $\mathcal{G}$. Again, by the previous discussion we know that no path in $\mathcal{S}$ contains an outgoing edge of a node in $\mathcal{W}$. Therefore, $\mathcal{S}$ is also a constrained set of $\abs{\mathcal{U}}$ vertex-disjoint paths from $\mathcal{U}$ to $\mathcal{W}$ in $\bar{\mathcal{G}}$. This proves the lemma. The following lemma relates the existence of a constrained set of vertex-disjoint paths and the second graph operation. \[lemmacp2\] Consider a directed graph $\mathcal{G} = (\mathcal{V},\mathcal{E})$ and let $\mathcal{U},\mathcal{W} \subseteq \mathcal{V}$. Suppose that $k \in \mathcal{W} \setminus \mathcal{U}$ has exactly one in-neighbour $j \in \mathcal{N}_k^-$ that is reachable from $\mathcal{U}$. Then there exists a constrained set of $\abs{\mathcal{U}}$ vertex-disjoint paths from $\mathcal{U}$ to $\mathcal{W}$ if and only if there exists a constrained set of $\abs{\mathcal{U}}$ vertex-disjoint paths from $\mathcal{U}$ to $\bar{\mathcal{W}}$ with $\bar{\mathcal{W}} := (\mathcal{W} \setminus \{k\}) \cup \{j\}$. We will first show that $\mathcal{S}$ is a set of $\abs{\mathcal{U}}$ vertex-disjoint paths from $\mathcal{U}$ to $\mathcal{W}$ if and only if $\bar{\mathcal{S}}$ is a set of $\abs{\mathcal{U}}$ vertex-disjoint paths from $\mathcal{U}$ to $\bar{\mathcal{W}}$, where $\bar{\mathcal{S}}$ will be specified. Suppose that $\mathcal{S}$ is a set of $\abs{\mathcal{U}}$ vertex disjoint paths from $\mathcal{U}$ to $\mathcal{W}$. Consider the path $\mathcal{P} \in \mathcal{S}$ that goes from $\mathcal{U}$ to $k$. Since $j \in \mathcal{N}_k^-$ is the only in-neighbour of $k$ that is reachable from $\mathcal{U}$, we obtain $(j,k) \in \mathcal{P}$. This means that $\bar{\mathcal{P}} := \mathcal{P} \setminus (j,k)$ is a path from $\mathcal{U}$ to $j$. Clearly, $\bar{\mathcal{S}} := (\mathcal{S} \setminus \mathcal{P}) \cup \bar{\mathcal{P}}$ is a set of $\abs{\mathcal{U}}$ vertex-disjoint paths from $\mathcal{U}$ to $\bar{\mathcal{W}}$. Conversely, suppose that $\bar{\mathcal{S}}$ is a set of $\abs{\mathcal{U}}$ vertex-disjoint paths from $\mathcal{U}$ to $\bar{\mathcal{W}}$. Consider the path $\bar{\mathcal{P}} \in \bar{\mathcal{S}}$ that goes from $\mathcal{U}$ to $j \in \bar{\mathcal{W}}$. Since $j \in \mathcal{N}_k^-$ is the only in-neighbour of $k$ that is reachable from $\mathcal{U}$, the path $\bar{\mathcal{P}}$ does not pass through the vertex $k$. Consequently, $\mathcal{P} := \bar{\mathcal{P}} \cup (j,k)$ is a path from $\mathcal{U}$ to $k$. Again using the fact that $j$ is the only in-neighbour of $k$ that is reachable from $\mathcal{U}$, we see that no path in $\bar{\mathcal{S}}$ passes through the vertex $k$. This implies that $\mathcal{S} := (\bar{\mathcal{S}} \setminus \bar{\mathcal{P}}) \cup \mathcal{P}$ is a set of $\abs{\mathcal{U}}$ vertex-disjoint paths from $\mathcal{U}$ to $\mathcal{W}$. To conclude, we have shown that $\mathcal{S}$ is a set of $\abs{\mathcal{U}}$ vertex-disjoint paths from $\mathcal{U}$ to $\mathcal{W}$ if and only if $\bar{\mathcal{S}}$ is a set of $\abs{\mathcal{U}}$ vertex-disjoint paths from $\mathcal{U}$ to $\bar{\mathcal{W}}$, where the set $\bar{\mathcal{S}}$ is defined as $\bar{\mathcal{S}}:= (\mathcal{S} \setminus \mathcal{P}) \cup \bar{\mathcal{P}}$. This implies that $\mathcal{S}$ is a constrained set of $\abs{\mathcal{U}}$ vertex-disjoint paths from $\mathcal{U}$ to $\mathcal{W}$ if and only if $\bar{\mathcal{S}}$ is a constrained set of $\abs{\mathcal{U}}$ vertex-disjoint paths from $\mathcal{U}$ to $\bar{\mathcal{W}}$. The ‘if’ statement follows from Theorem \[theoremnecsys\]. To prove the ‘only if’ part, suppose that $(i,\mathcal{N}_i^+)$ is identifiable. By Theorem \[mainresultiNi\], $\mathcal{N}_i^+ \subseteq \mathcal{D}(\mathcal{C})$, where $\mathcal{D}(\mathcal{C})$ is a derived vertex set of $\mathcal{C}$ with respect to $\mathcal{N}_i^+$. In fact, we obtain $\mathcal{N}_i^+ = \mathcal{D}(\mathcal{C})$ as $\abs{\mathcal{N}_i^+} = \abs{\mathcal{C}}$. Let $\mathcal{D}(\mathcal{G})$ denote the associated derived graph of $\mathcal{G}$. By definition, there exists a constrained set of $\abs{\mathcal{N}_i^+}$ vertex-disjoint paths from $\mathcal{N}_i^+$ to $\mathcal{D}(\mathcal{C})$ in $\mathcal{D}(\mathcal{G})$ (see Section \[sectionresultscvdp\]). By consecutive application of Lemmas \[lemmacp1\] and \[lemmacp2\], we conclude that there exists a constrained set of $\abs{\mathcal{N}_i^+}$ vertex-disjoint paths from $\mathcal{N}_i^+$ to $\mathcal{C}$ in the graph $\mathcal{G}$. [^1]: Henk van Waarde and Kanat Camlibel are with the Bernoulli Institute for Mathematics, Computer Science and Artificial Intelligence, Faculty of Science and Engineering, University of Groningen, P.O. Box 407, 9700 AK Groningen, The Netherlands. Henk van Waarde is also with the Engineering and Technology Institute Groningen, Faculty of Science and Engineering, University of Groningen, 9747 AG Groningen, The Netherlands. Pietro Tesi is with the Engineering and Technology Institute Groningen and also with the Department of Information Engineering, University of Florence, 50139 Florence, Italy. Email: [[email protected], [email protected], [email protected], [email protected]]{}. [^2]: Such sets of vertex-disjoint paths have been studied in detail in [@Murota1987], where they were called linkings.	{ "pile_set_name": "ArXiv" }
--- abstract: 'Topological boundary and interface modes are generated in an acoustic waveguide by simple quasi-periodic patterning of the walls. The procedure opens many topological gaps in the resonant spectrum and qualitative as well as quantitative assessments of their topological character are supplied. In particular, computations of the bulk invariant for the continuum wave equation are performed. The experimental measurements reproduce the theoretical predictions with high fidelity. In particular, acoustic modes with high Q-factors localized in the middle of a breathable waveguide are engineered by a simple patterning of the walls.' author: - 'David J. Apigo' - Wenting Cheng - 'Kyle F. Dobiszewski' - Emil Prodan - Camelia Prodan title: 'Observation of topological edge modes in a quasi-periodic acoustic waveguide' --- The ideas based on topological concepts [@ThoulessPRL1982; @HaldanePRL1988] have revolutionized the field of condensed matter physics and led to the discovery of topological insulators and superconductors. The latter have been classified at the end of the previous decade [@SRFL2008; @QiPRB2008; @Kit2009; @RSFL2010] and a table of strong topological phases has been conjectured. One of their common characteristics is the emergence of disorder-immune boundary modes whenever a sample is halved. Physics akin to that of topological condensed matter systems has been also predicted in classical wave-supporting materials [@HaldaneRaghu2008; @PP2009] and many examples of topological metamaterials have been reported in the literature [@WangNature2009; @NashPNAS2015; @HafeziNatPhot2013; @WuPRL2015; @SusstrunkScience2015; @KaneNatPhys2013; @PauloseNatPhys2015; @ProdanNatComm2017; @KhanikaevNatPhot2017; @MousaviNatComm2015; @RuzzeneArxiv2017; @ChaunsaliPRB2018; @ChernArxiv2018; @PalJAP2016]. At the same time, it has been pointed out that the periodic table of topological systems is highly enhanced if more complex systems are considered, such as the quasi-periodic or quasi-crystalline ones [@KLR2012; @VZK2013; @ProdanPRB2015; @BabouxPRB2017]. In [@Apigo2], $K$-theoretic arguments [@Bel86; @PS] were applied for quasi-periodically coupled discrete mechanical resonators. The finding was that, if these are single-mode resonators, then every gap in the bulk resonant spectrum is topological, in the sense that it will be completely filled by boundary spectrum under any boundary condition. The practical value of the finding is that the quasi-periodic Hamiltonians display a large number of topological gaps, hence one can generate localized wave-modes in both space and energy by simply halving the system. In this work, we put these general principles to the test in a completely different regime and we implement them for the first time using sound waves. Acoustic setups have been successfully used in the past to generate topological edge modes [@He; @Xiao; @NiArxiv2018] and even to map the Hofstadter butterfly [@RichouxEPL2002]. In particular, [@NiArxiv2018; @RichouxEPL2002] introduced re-configurable acoustic resonant structures where the building blocks are sealed acoustic chambers connected via thin bridges. They have isolated resonant modes, hence these structures fall under the umbrella of patterned resonators introduced in [@Apigo2] and they can be analyzed by similar methods. However, these types of acoustic structures are not breathable, which is a key requirement for many practical applications. As such, here we ask the question: Can one generate topological edge and interface modes by patterning the walls of an acoustic waveguide without impeding the air flow? As we shall see, the answer is yes, but the methods of analysis are very different from those introduced in [@Apigo2]. Indeed, the picture of coupled discrete resonators is no longer applicable and a full continuum medium treatment must be employed for the theoretical analysis. Furthermore, the topological character of the spectral gaps cannot be taken for granted because the waveguide supports many overlapping modes. As such, a new assessment of the topological character is introduced based on the continuum version of the lattice non-commutative Chern number proposed in [@ProdanPRB2015], achieved in [@BourneMPAG2018]. This invariant is here evaluated numerically using the methods developed in [@ProdanSpringer2017; @ProdanJPA2018]. Let us recall from [@Apigo2] that the role of aperiodicity in this type of applications is to generate virtual dimensions and, as we shall see [@Suppl], the Chern number mentioned above is defined on a 3-dimensional non-commutative manifold, while for discrete patterns is on a 2-dimensional manifold. At the experimental level, challenges exist because some of the spectral bands are very narrow and this, together with the aperiodicity, can lead to irregular mode profiles, although the bulk states are extended. As such, the only way to accurately map the bulk spectrum is to collect data from a large number of points along the waveguide. Following this protocol, we map not only the frequecy but also the spatial profile of the bulk modes. Furthermore, inside the topological bulk gaps, we were able to detect sharp edge modes, which flow with the phason degree of freedom in a manner consistent with the computed Chern numbers. ![image](Experimental_setup.jpg){width="\linewidth"} The quasi-periodic acoustic waveguide consists of a uniform cylindrical tube decorated with walls. The parts were 3D-printed out of polylactic acid (PLA) using an Ultimaker 3 and then assembled as in Fig. \[Fig:Setup\].The walls have identical thickness but the spacings between adjacent walls are modulated according to the algorithm: $$\label{Eq:Algorithm} L_n = L_{\rm avg} +\Delta L \,\sin(n \theta + \phi ), \quad n \in {{\mathbb Z}}.$$ The geometric parameters used in the experiments are supplied in Fig. \[Fig:Setup\]. To make the above labels meaningful, we assume that the waveguide is centered at a point inside $L_0$. In Equation , $\theta$ is an angle incommensurate with $2\pi$, which will be kept fixed during the measurements, and $\phi$ is the phason, which should be let to vary. For example, a simple relabeling $n \rightarrow n+m$, which corresponds to re-centering the waveguide, will change $\phi$ into $(\phi + m \theta) {\rm mod}\, 2\pi$. Since $\theta$ is incommensurate, these relabelings alone will sample the phason densely in the $[0,2\pi]$ interval. $L_{\rm avg}$ in Equation is the average distance between the walls and $\Delta L$ sets the magnitude of the fluctuations in $L_n$. In the inset of Fig. \[Fig:Setup\], we show a front view of the waveguide, confirming that air can flow freely through the structure. It is then somewhat striking that, with the proposed patterning, we can stop sound propagation over several intervals of frequencies and, furthermore, we can generate, very much on demand, topological sound modes localized at any desired location along the tube. As opposed to an ordinary resonant mode produced in a fully sealed acoustic chamber, the interface modes produced in the present work have less contact with the boundary, hence they are expected to have very high Q-factors, a much desired characteristic for practical applications. To understand the effect of the patterning, we report in Fig. \[Fig:BandStructure\] the dispersion of the acoustic modes for clean and periodically ($L_n = L_{\rm avg}$) patterned waveguides, as well as the resonant spectrum of the aperiodically patterned waveguide ($L_n$ set by ). As expected for quasi 1-dimensional wave propagation, the periodic pattern opens spectral gaps in the gapless spectrum of the clean tube. These gaps, however, are not topological. The role of aperiodicity is to open additional gaps in the spectrum that, as one can see, resemble quite closely the Hofstadter butterfly [@HofstadterPRB1976], when mapped as function of $\theta$. As we shall see, these are the gaps that carry non-trivial bulk topological invariants prompting the topological edge and interface modes. Let us mention that the spectra in Fig. \[Fig:BandStructure\] were produced with an in-house Fortran code, which diagonalizes the Laplace operator expressed in the cylindrical coordinates $(\rho,z)$ and resolved over the azimuthal symmetry sectors. In appropriate units, the operator reads: $$\Delta_m = -\frac{1}{\rho} \frac{\partial}{\partial \rho}\rho \frac{\partial }{\partial \rho} + \frac{m^2}{\rho^2} - \frac{\partial^2 }{\partial z^2}, \quad m=0,\pm 1,\ldots,$$ and von Neumann condition is considered at the boundary. Recall that the latter is set by $\theta$ and $\phi$, hence $\Delta_m$ depends in a fundamental way on these parameters. The Laplace operator was discretized using finite differences. ![a) Dispersion of the acoustic modes for the un-patterned waveguide, for $m=0$ sector. b) The band structure of a periodically patterned waveguide ([i.e.]{} $\theta=0$), for $m=0$ sector. c,d) Resonant spectrum of a patterned waveguide as function of $\theta$, for $m=0,1$ sectors, respectively.[]{data-label="Fig:BandStructure"}](BandStructure.jpg){width="\linewidth"} The protocol for acoustic data acquisition was as follows. Sinusoidal signals of duration 1 s and amplitude of 0.5 V were produced by a Rigol DG1022 function generator, amplified by a Crown XLS 2502 power amplifier with the gain set to 6, and then applied on a CUI Inc. GF0501 speaker, placed at one of the portholes. A PCB Piezotronics Model-378C10 microphone and a PCB Piezotronics Model-485B12 power conditioner acquired the acoustic signals at a porthole opposite the speaker (see Fig. \[Fig:Setup\]). To account for the frequency-dependent response of the components, a separate measurement is performed with the waveguide removed but speaker and microphone kept in the same positions. All readings are normalized by the output of these measurements. The outputs were read by a custom LabVIEW code via a National Instruments USB-6112 data acquisition box and the ratio of the two measurements is stored on a computer for graphic renderings. For the bulk measurements, the protocol was repeated for all 48 chambers of a patterned waveguide, with frequency scans from 500 to 6000 Hz in 25 Hz steps. The results are reported in Fig. \[Fig:BulkSpectrum\]. When the data is rendered as function of frequency and chamber index, clear extended acoustic modes can be identified. Furthermore, when the data is collapsed on the frequency axis, clear spectral gaps can be identified, two of which are predicted to be topological. Unfortunately, the $m=0,1$ spectra overlap above the non-topological gap (see Fig. \[Fig:BandStructure\]) and the higher frequency topological gaps could not experimentally resolved. Let us note that the agreement between experiment and theory in Fig. \[Fig:BulkSpectrum\] is less than 5%. ![Bulk resonant spectrum for the geometry described in Fig. \[Fig:Setup\]. Left: Theoretical resonant spectrum reproduced from Fig. \[Fig:BandStructure\](c), with arrows indicating the topological gaps. The vertical marking identifies $\theta = \frac{2 \pi}{\sqrt{117}}$, used in experiments. Center: Normalized microphone readings from the center of 48 chambers, recorded over a wide frequency interval. Right: Collapse on the frequency axis of the intensity plot reported in the mid panel. Three spectral gaps can be clearly identified in the experimental data and seen to be well aligned with the theoretical calculations. The values of the Chern numbers for the two topological gaps are also indicated.[]{data-label="Fig:BulkSpectrum"}](Bulk_spectrum.jpg){width="\linewidth"} To assess the topological character of the gaps, we employ the bulk-boundary correspondence for continuum models established in [@BourneMPAG2018]. The bulk-topological invariant is supplied by the non-commutative Chern number of the gap projection $P_G = \chi_{(-\infty,G]}\big (\Delta_m(\phi) - G\big )$: $$\label{Eq:BulkInvariant} {\rm Ch}(P_G) = {\rm Tr}_L\big ( P_G[\partial_\phi P_G,[Z,P_G]] \big ),$$ where $Z$ is the position operator parallel to the tube and ${\rm Tr}_L$ is the trace per length. The invariant can be computed at any arbitrary but fixed phason value, which is a consequence of Birkhoff ergodic theorem [@Bir]. With the Laplacian discretized on a lattice via finite differences, Eq. was evaluated using methods which are by now standard [@ProdanSpringer2017; @ProdanJPA2018]. The results are reported in Fig \[Fig:BulkSpectrum\], confirming that the smaller gaps are topological. Furthermore, [@BourneMPAG2018] established the existence of a boundary topological invariant which counts the number of chiral boundary bands, as well as the equality between the bulk and boundary invariants. ![Topological edge spectrum. a) Theoretical prediction of the spectral flow against the phason parameter $\phi$, demonstrating the existence of chiral bands. The red/blue marks relate to the left/right edge of the waveguide, respectively. b) Experimental mapping of the spectral flow, confirming the existence of chiral bands. c) The measurements for bulk spectrum, reproduced from Fig. \[Fig:BulkSpectrum\], indicating the position of the bulk gap edges.[]{data-label="Fig:Edge"}](Edge_spectrum.jpg){width="\linewidth"} The presence of chiral modes, in accordance to the above bulk-boundary principle, is confirmed by our numerical simulations reported in Fig. \[Fig:Edge\]a. To map the boundary modes experimentally, the acquisition protocol was applied on the second chamber from the left physical edge, which was plugged. The frequency was swept from 2.0 to 2.6 kHz in steps of 25 Hz and the value of the phason was modified by moving the physical edge sequentially to the right, hence from $L_0$ to $L_n$, $n=1,2,\ldots$. The results are presented in Fig. \[Fig:Edge\]b and they indeed confirm the existence of one chiral band in the upper topological gap and two such bands in the lower topological gap. For reference, we reproduced in panel c) the experimental data from Fig. \[Fig:BulkSpectrum\], from where the exact position of the bulk edges can be inferred. As one can see, the boundary resonances occur inside the bulk gaps and the dispersion with $\phi$ is consistent with the theoretical prediction. We now demonstrate that a localized topological edge mode can be created without the assistance of any plug. For this, we consider a domain wall configuration: $$\ldots \|L_{31}\|L_{30}\|L_{29}\|L_{29}\|L_{30}\|L_{31}\|\ldots$$ where the waveguide is mirror-reflected relative to left edge of $L_{29}$ chamber. This particular index was chosen because moving the origin to that chamber generates a phason $\phi = (29 \theta){\rm mod}\, 2\pi$, which coincides with the value where strong mid-gap edge modes were observed in the first topological gap. Since Eq. \[Eq:BulkInvariant\] is odd under reflection, with this patterning, an interface between topological systems with opposite Chern numbers is created As such, the bulk-boundary principle predicts the emergence of $2\times {\rm Ch}$ acoustic modes localized at the interface. The experimental measurements are reported in Fig. \[Fig:Localized\_sound\](a). The frequencies were swept as in Fig. \[Fig:Edge\] and, in order to probe the localization of the acoustic modes, the speaker and microphone were placed at several portholes at and away from the interface. A strong and sharp resonance was detected in the first topological gap (${\rm Ch}=1$), when the measurements were performed one and two chambers away from the interface. The resonance was not detectable further away from the interface or at the interface itself. A similar resonance can be detected at the other side of the interface, leading to a full confirmation of the topological bulk-boundary prediction. The interface mode is also observed in a standard COMSOL simulation, as shown in Fig. \[Fig:Localized\_sound\](b). ![a) Topological interface mode, measured for a waveguide configuration similar to that in Fig. \[Fig:Setup\]. The spatial localization of the interface mode was mapped by moving the speaker and microphone incrementally away from the domain wall. b) The topological interface mode is also observed in COMSOL simulations. Red, blue and green colors represent high, low and zero pressure variations, respectively. []{data-label="Fig:Localized_sound"}](Interface-Closed_Measurement.jpg){width="\linewidth"} In conclusion, we have demonstrated that topological edge and interface modes can be created by a simple quasi-periodic patterning of an acoustic waveguide. The topological gaps can be easily identified when the resonant spectrum is mapped as function of modulation parameter $\theta$. Furthermore, a topological invariant was computed and shown to be in agreement with the number of observed topological chiral edge modes. As we have seen, quasi-periodicity opens topological gaps inside the bands of the periodic structure, which resemble the Hofstadter butterfly when mapped as function of $\theta$. Optimization over $\Delta L$ in Eq. and the geometric parameters of the tube, as well as improvements in materials ([e.g.]{} by replacing the polymer with metal), can highly enhance these topological gaps and the Q-factors of the topological boundary and interface modes. Other than that, the procedure requires no further fine tuning and, due to its simplicity, we believe it can be easily incorporated in practical applications. The present analysis can also serve as a model for acoustic implementations of many other promising aperiodic structures [@ProdanJGP2018].\ Supplemental ============ ![image](Pattern.pdf){width="0.8\linewidth"} The aperiodic continuum systems are quite different from the discrete aperiodic ones for the following reasons: - They can be halved at any point of their axis. Hence, the phason $\phi$ alone does not specify completely the configuration of the waveguide, for we also need to know where the origin of the Euclidean space is located relative to the walls of the patterned waveguide. This is so because, by convention, it is at this origin where the cut is made and the edge modes emerge. - The bulk topological invariant, while still a formal Chern number, is defined on a different algebra of observables. - The proof [@BourneMPAG2018] of the quantization and stability of the topological invariants for continuum models also proceeds quite differently from the one for discrete models [@PS]. The aim of this note is to walk the reader through [@BourneMPAG2018], as adopted to the acoustic waveguide analyzed in the main text. [The continuous hull.]{} The continuous hull $\Omega$ of the patterned waveguide is the topological space traced by the pattern when one continuously translates the waveguide along the axis [@Bel86]. Here we show that $\Omega$ is a 2-torus. With the elements introduced in diagram \[Fig:Pattern\], let $\omega \in {{\mathbb T}}^2$ be an arbitrary point and imagine the blue line as being a physical rope wound around the torus. Let the red line, whose equation is supplied in the diagram, be soaked with ink so that, every time when the rope crosses the red line, a mark is imprinted. Let us label these marks as shown in the diagram. Then, after we unwind the rope and lay it flat and parallel to axis of the tube, one will find that: $$\begin{aligned} x_{n+1} -x_n & = L_{\rm avg}+y\big ((n+1)\theta + \phi_\omega \big) -y(n\theta+\phi_\omega) \\ \nonumber &= L_{\rm avg} + \Delta L \sin(n\theta + \phi_\omega).\end{aligned}$$ The marks $x_n$ will overlay perfectly over the centers of the walls if the origin of the Euclidean space is fixed at $x_\omega$ (the position of $\omega$ on the rope)! The conclusion is that every rigidly translated waveguide configuration can be uniquely characterized by a point $\omega \in {{\mathbb T}}^2$, hence the continuous hull is the 2-torus. Furthermore, the group of translations parallel to the waveguide’s axis induces an action $\tau$ of ${{\mathbb R}}$ on ${{\mathbb T}}^2$, which amounts to shifting $\omega$ along the winded rope. As such, the hull becomes a topological dynamical system $(\Omega,\tau,{{\mathbb R}})$. [Algebra of physical observables.]{} When defining a topological invariant for aperiodic systems, the first task is to determine the operator algebra which supplies the associated physical observables. As it is now well known [@Bel86], for continuous 1-dimensional models, this algebra is the crossed product ${{\mathcal A}}=C(\Omega) \rtimes_\tau {{\mathbb R}}$. The elements of this algebra belong to a certain class of complex valued functions over ${{\mathbb R}}\times \Omega$ and the multiplication rule is: $$(f_1f_2)(z,\omega)= \int_{{\mathbb R}}d \xi \, f_1(\xi,\tau_{\xi-z}\omega) f_2(z-\xi,\omega).$$ The algebra accepts a canonical representation on $L^2({{\mathbb R}})$: $$\big [ (\pi_\omega f) \psi \big ](z) = \int_{{\mathbb R}}d \xi \, f(z-\xi,\tau_{-\xi} \omega) \psi(\xi).$$ Here, $z$ is the coordinate along the axis of the waveguide. The dispersion equation for our waveguide is defined over $L^2([0,R]) \otimes L^2({{\mathbb R}})$, where $[0,R]$ is the interval where the radial coordinate $\rho$ takes values. The transversal modes, however, from topology point of view, brings nothing significant because all spectral projectors of the dispersion operator can be generated from the algebra ${{\mathbb K}}\otimes {{\mathcal A}}$ (via the above representation), where ${{\mathbb K}}$ is the algebra of compact operators over $L^2([0,R])$. This is the case because the resolvent of the radial part of the Laplace operator is compact when $\rho$ is restricted to a finite interval. [Topological Invariant.]{} We now can specify the input for the machinery developed in [@BourneMPAG2018]: - If $(\omega_1,\omega_2)$ are the coordinates of $\omega \in {{\mathbb T}}^2={{\mathbb S}}\times {{\mathbb S}}$, then we have the derivations $\partial_{\omega_1}$ and $\partial_{\omega_2}$ ($=\partial_\phi$), as well as: $$(\partial_z f)(z,\omega)= z f(z,\omega), \quad f \in {{\mathcal A}}.$$ - The trace ${\rm Tr} \otimes {{\mathcal T}}$ on ${{\mathbb K}}\otimes {{\mathcal A}}$ with: $${{\mathcal T}}(f) = \int_{{{\mathbb T}}^2} \frac{d\omega}{(2\pi)^2} \, f(0,\omega)$$ Then, for a projection $p \in {{\mathbb K}}\otimes {{\mathcal A}}$, [@BourneMPAG2018] showed that: $${\rm Ch}(p) = 2 \pi \, {\rm Tr}\otimes {{\mathcal T}}\big ( p [\partial_{\omega_2} p, \partial_z p]\big )$$ equals the index of a certain Fredholm operator, which ensures the quantization and stability of this Chern number. We can use the physical representation to write this invariant. Indeed, if $P_\omega=\pi_\omega(p)$, then: $$\label{Eq:Chern} {\rm Ch}(P) = \int d \rho \int_{{{\mathbb T}}^2} d \omega \, \langle 0,\rho\| P_\omega[\partial_\phi P_\omega,[Z,P_\omega]] \|0,\rho \rangle,$$ where $Z$ is the position operator parallel to the tube. Lastly, since $\tau$ acts ergodically on ${{\mathbb T}}^2$, we have from Birkhoff theorem [@Bir]: $$\int d \rho \int_{{{\mathbb T}}^2} d\omega \, \langle 0,\rho\| \ldots \|0,\rho \rangle = {\rm Tr}_L( \ldots ),$$ the latter being the trace per length introduced in the main text. With this simplification, Eq. becomes identical with the one supplied in the main text. [Bulk-boundary correspondence.]{} According to [@BourneMPAG2018], the topological class of $P_\omega$ is mapped into the $K_1$-class of the torus ${{\mathbb T}}^2$ generated by the function $e^{\imath \phi}$. Since it involves only the vertical coordinate of the torus \[Fig:Pattern\], the horizontal coordinate plays no role in the bulk-boundary correspondence treated in our work. [9]{} D. Thouless, M. Kohmoto, M. Nightingale, M. den Nijs, [Quantized Hall Conductance in a Two-Dimensional Periodic Potential]{}, Phys. Rev. Lett. [49]{}, 405-409 (1982). F. D. M. Haldane, [Model for a Quantum Hall-Effect without Landau levels: Condensed-matter realization of the parity anomaly]{}, Phys. Rev. Lett. [61]{}, 2015-2018 (1988). A. P. Schnyder, S. Ryu, A. Furusaki, A. W. W. Ludwig, [Classification of topological insulators and superconductors in three spatial dimensions]{}, Phys. Rev. [B 78]{}, 195125 (2008). X.-L. Qi, T. L. Hughes, Shou-Cheng Zhang, [Topological field theory of time-reversal invariant insulators]{}, Phys. Rev. B [78]{}, 195424 (2008). A. Kitaev, [Periodic table for topological insulators and superconductors]{}, (Advances in Theoretical Physics: Landau Memorial Conference) AIP Conference Proceedings [1134]{}, 22-30 (2009). S. Ryu, A. P. Schnyder, A. Furusaki, A. W. W. Ludwig, [Topological insulators and superconductors: tenfold way and dimensional hierarchy]{}, New J. Phys. [12]{}, 065010 (2010). F. D. M. Haldane, S. Raghu, [Possible realization of directional optical waveguides in photonic crystals with broken time-reversal symmetry]{}, Phys. Rev. Lett. Lett. [100]{}, 013904 (2008). E. Prodan, C. Prodan, [Topological phonon modes and their role in dynamic instability of microtubules]{}, Phys. Rev. Lett. [103]{}, 248101 (2009). Z. Wang, Y. Chong, J. D. Joannopoulos, M. Soljacic, [Observation of unidirectional backscattering-immune topological electromagnetic states]{}, Nature [461]{}, 772–775 (2009). L. M. Nash, D. Kleckner, A. Read, V. Vitelli, A. M. Turner, W. T. M. Irvine, [Topological mechanics of gyroscopic metamaterials]{}, Proc. Nat. Acad. Sci. [112]{}, 14495-14500 (2015). M. Hafezi, S. Mittal, J. Fan, A. Migdall, J. M. Taylor, [Imaging topological edge states in silicon photonics]{}, Nature Photonics [7]{}, 1001-1005 (2013). L.-H. Wu, X. Hu, [Scheme for Achieving a Topological Photonic Crystal by Using Dielectric Material]{}, Phys. Rev. Lett. [114]{}, 223901 (2015). R. S[ü]{}sstrunk, S. Huber, [Observation of phononic helical edge states in a mechanical topological insulator]{}, Science [349]{}, 47-50 (2015). C. Kane, T. Lubensky, [Topological boundary modes in isostatic lattices]{}, Nature Physics [10]{}, 39-45 (2013). J. Paulose, B. G. Chen, V. Vitelli, [Topological modes bound to dislocations in mechanical metamaterials]{}, Nature Physics [11]{}, 153-156 (2015). E. Prodan, K. Dobiszewski, A. Kanwal, J. Palmieri, C. Prodan, [Dynamical Majorana edge modes in a broad class of topological mechanical systems]{}, Nature Communications [8]{}, 14587 (2017). A. Slobozhanyuk, S. H. Mousavi, X. Ni, D. Smirnova, Y. S. Kivshar, A. B. Khanikaev, [Three-dimensional all-dielectric photonic topological insulator]{}, Nature Photonics [11]{}, 130-136 (2017). S. H. Mousavi, A. B. Khanikaev, Z. Wang, [Topologically protected elastic waves in phononic metamaterials]{}, Nature Communications [6]{}, 8682 (2015). M. Miniaci, R. K. Pal, B. Morvan, M. Ruzzene, [Observation of topologically protected helical edge modes in Kagome elastic plates]{}, arXiv:1710.11556 (2017). R. Chaunsali, C.-W. Chen, J. Yang, [Subwavelength and directional control of flexural waves in zone-folding induced topological plates]{}, Phys. Rev. B [97]{}, 054307 (2018). H. Chen, H. Nassar, G. Huang, [Topological mechanics of edge waves in Kagome lattices]{}, arXiv:1802.04404 (2018). R. K. Pal, M. Schaeffer, M. Ruzzene, [Helical edge states and topological phase transitions in phononic systems using bilayered lattices]{}, J. Appl. Phys. [119]{}, 084305 (2016). Y. E. Kraus, Y. Lahini, Z. Ringel, M. Verbin, O. Zilberberg, [Topological states and adiabatic pumping in quasicrystals]{}, Phys. Rev. Lett. [109]{}, 106402 (2012). M. Verbin, O. Zilberberg, Y. E. Kraus, Y. Lahini, Y. Silberberg, [Observation of topological phase transitions in photonic quasicrystals]{}, Phys. Rev. Lett. [110]{}, 076403 (2013). E. Prodan, [Virtual topological insulators with real quantized physics]{}, Phys. Rev. B [91]{}, 245104 (2015). F. Baboux, E. Levy, A. Lemaitre, C. Gomez, E. Galopin, L. Le Gratiet, I. Sagnes, A. Amo, J. Bloch, E. Akkermans, [Measuring topological invariants from generalized edge states in polaritonic quasicrystals]{}, Phys. Rev. B 95, 161114 (2017). D.J. Apigo, K. Qian, C. Prodan, and E. Prodan, [Topological edge modes by smart patterning]{}, Phys. Rev. Materials [2]{}, 124203 (2018). J. Bellissard, [K-theory of $C^$-Algebras in solid state physics]{}, Statistical mechanics and field theory: mathematical aspects, (Springer, Berlin, Heidelberg, 1986, 99-156). E. Prodan, H. Schulz-Baldes, [Bulk and boundary invariants for complex topological insulators: From $K$-theory to physics]{}, (Springer, Berlin, 2016). M. Xiao, G. Ma, Z. Yang, P. Sheng, Z.Q. Zhang, C.T. Chan, [Geometric phase and band inversion in periodic acoustic systems]{}, Nature Physics, [11]{}:240, (2015). C. He, X. Ni, H. Ge, X. Sun, Y. Chen, M. Lu, X. Liu, Y. Chen, [Acoustic topological insulator and robust one-way sound transport]{}, Nature Physics, [12]{}:1124, (2016). X. Ni, M. Weiner, A. Alù, A. B. Khanikaev, [Observation of bulk polarization transitions and higher-order embedded topological eigenstates for sound]{}, arXiv:1807.00896 (2018). O. Richoux, V. Pagneux, [Acoustic characterization of the Hofstadter butterfly with resonant scatterers]{}, EPL (Europhysics Letters), Vol 59, 34 (2002). C. Bourne, A. Rennie, [Chern numbers, localisation and the bulk-edge correspondence for continuous models of topological phases]{}, Math. Phys. Anal. Geom. 21: 16. https://doi.org/10.1007/s11040-018-9274-4 (2018). E. Prodan, [A computational non-commutative geometry program for disordered topological insulators]{}, Springer Briefs in Mathematical Physics, Springer, 2017. C. Bourne, E. Prodan, [Non-commutative Chern numbers for generic aperiodic discrete systems]{}, J. Phys. A: Math. Theor. [51]{}, 235202 (2018). See Supplemental Material at \[URL\] for a computation of the continuous hull of the pattern as well as for the definition of the bulk topological invariant. D. R. Hofstadter, [Energy levels and wave functions of Bloch electrons in rational and irrational magnetic fields]{}, Phys. Rev. B [14]{}, 2239-2249 (1976). G. D. Birkhoff, [Proof of the ergodic theorem]{}, Proc. Natl. Acad. Sci. USA [17]{}, 656-660 (1931). E. Prodan, Y. Shmalo, [The K-Theoretic Bulk-Boundary Principle for Dynamically Patterned Resonators]{}, Journal of Geometry and Physics [135]{}, 135-171 (2019).	{ "pile_set_name": "ArXiv" }
--- abstract: 'We investigate the quantum breathing mode (monopole oscillation) of trapped fermionic particles with Coulomb and dipole interaction in one and two dimensions. This collective oscillation has been shown to reveal detailed information on the many-particle state of interacting trapped systems and is thus a sensitive diagnostics for a variety of finite systems, including cold atomic and molecular gases in traps and optical lattics, electrons in metal clusters and in quantum confined semiconductor structures or nanoplasmas. An improved sum rule formalism allows us to accurately determine the breathing frequencies from the ground state of the system, avoiding complicated time-dependent simulations. In combination with the Hartree-Fock and the Thomas-Fermi approximations this enables us to extend the calculations to large particle numbers $N$ on the order of several million. Tracing the breathing frequency to large $N$ as a function of the coupling parameter of the system reveals a surprising difference of the asymptotic behavior of one-dimensional and two-dimensional harmonically trapped Coulomb systems.' address: - 'Institut für Theoretische Physik und Astrophysik, Christian-Albrechts-Universität zu Kiel, Leibnizstraße 15, D-24098 Kiel, Germany ' - 'Department of Physics, University of Ottawa, Ottawa, ON K1N6N5, Canada' author: - Jan Willem Abraham and Michael Bonitz - 'Chris McDonald, Gianfranco Orlando, and Thomas Brabec' title: Quantum Breathing Mode of Trapped Systems in One and Two Dimensions --- Trapped systems are of major interest in many fields of research. Prominent examples are correlated electrons in metal clusters, e.g. [@baletto2005], confined plasmas [@amiranashvili2003], ultracold quantum gases in traps or optical lattices [@giorgini2008; @dalfovo99; @bloch2005], electrons in quantum dots [@filinov2001; @filinov2000; @reimann2002; @ashoori96; @filinov_prb08] (“artificial atoms”), excitons in bilayers and quantum wells, e.g. [@filinov_cpp01; @hartmann05; @ludwig_njp08; @boening_prb11; @schleede_cpp12] trapped ions [@schweigert95], and colloidal particles [@tatarkova2002]. Although these systems may differ in many physical details, their theoretical descriptions are often similar. A key property of finite systems in traps are the low-lying collective oscillations since they serve as a valuable diagnostic tool for the investigation of time-dependent and static features, e.g. [@giorgini2008; @moritz2003; @string]. The importance of these collective modes is comparable to that of spectroscopy in atomic systems. On the other hand, from the theoretical point of view, the calculation of normal modes is interesting as the results can be used to check the quality of models and of nonequilibrium simulations. In this work, we concentrate on the breathing mode – the (uniform) radial expansion and contraction of the system. This collective mode can easily be excited in experiments, e.g. by modulation of the confinement frequency or by a rapid compression or expansion [@moritz2003]. For classical systems, the breathing mode is well understood [@schweigert95; @henning2008; @olivetti2009]. However, if one incorporates quantum effects, the description of the mode becomes more complex. In our recent works [@bauch09; @bauch10; @Abrahamprb], we reported results from time-dependent simulations and presented some unique properties of the quantum breathing mode. Among others, we discovered that the breathing mode comprises a superposition of at least two oscillations, which is a pure quantum effect. While one of the corresponding frequencies, $\omega_\mathrm{cm}$, has a universal value that equals twice the trap frequency, the other one, $\omega_\mathrm{rel}$, is strongly dependent on various system parameters. The goal of the present work is to determine how the frequency $\omega_\mathrm{rel}$ is influenced by, e.g., the particle number, the coupling parameter of the system, the dimension of the system and the type of binary interactions. These results are crucial for exploiting the above mentioned diagnostic potential of the breathing mode. An analysis for bosons with contact interaction in a one-dimensional trap that is complementary to ours is given in Ref. [@schmelcher13]. Most of the numerical results in our previous works are based on time-dependent approaches. The high numerical effort of these simulations poses strong restrictions on the accessibility of several physical parameters. In particular, the particle number usually does not exceed the values $N\approx 20$ in one-dimensional (1D) systems [@Abrahamprb] and $N\approx 6$ in two-dimensional (2D) systems [@Brabec13]. To overcome this problem, a simple semi-analytical estimator for the breathing frequency has been presented in our recent work [@Brabec13]. As this estimator is solely based on equilibrium quantities, it enables one to save the computational effort of the time propagation in computer simulations, and, additionally, to use the breathing frequency as an experimental tool to determine the kinetic, trap and interaction energies of trapped systems. In [@Brabec13], we validated the accuracy of the estimator, comparing its predictions with the results from correlated time-dependent calculations. In this work, we make use of it to describe physical situations for which no time-dependent calculations are available. We systematically study one- and two-dimensional systems of charged fermionic particles in the range from small finite systems to gases with many particles. Such a joint analysis of the dimensionality and the particle number has not been performed so far. To work out how the equilibrium approach is connected to existing theories for the collective motions of many-body systems, we further extend the theoretical foundations of the breathing mode, providing a systematic description in terms of time-dependent perturbation theory. This allows one to conclude that the breathing mode is characterized by the spectrum of the initial Hamiltonian. To avoid an exact diagonalization of this Hamiltonian, we follow the well-known sum rule formalism [@lipparini89; @stringari82; @bohigas79; @bohigas76] to extract an approximation for the lowest excitation energy from the ground state. Furthermore, we show how to improve the sum rule formulas, which turns out to be important, especially for small systems. The sum rules allow us to apply ground-state methods with strongly reduced computational costs. Hence, we are able to considerably extend our previous results for fermionic particles with Coulomb interaction [@Abrahamprb; @Brabec13]. The calculations are performed in the framework of the Hartree-Fock approximation, for small weakly coupled systems with significant finite-size effects, and the Thomas-Fermi approximation, for the transition to large systems. One of the major results from our analysis is that the dimension of the system has a strong influence on the qualitative behavior of large systems: With growing particle number, a one-dimensional system behaves more like an ideal quantum system (see also Ref. [@Abrahamprb]), while a two-dimensional system, by contrast, approaches the classical limit. We show that this unexpected behavior is indicated by the lowest frequency of the breathing oscillation as well as a localization parameter which relates the average extension of the system to that of an ideal quantum system. Discussing how the total energies scale in non-interacting and strongly coupled (classical) systems, we further provide an explanation for the behavior in terms of simple quantities. We start our presentation with a brief review of the quantum mechanical description of the trapped system. Expressing the excitation of collective modes in terms of time-dependent perturbation theory, we lay the foundation for the application of the sum rules. We briefly recapitulate the most important sum rule formulas and show how their accuracy can be improved. Using this formalism, we analyze how the breathing frequency depends on the particle number, the coupling parameter and the dimensionality of the system. Theory ====== Time-dependent Schrödinger equation ----------------------------------- We briefly recall the physical setting which has already been described in our previous works [@bauch09; @bauch10; @Abrahamprb]. We aim at the quantum mechanical description of $N$ identical particles in a $d$-dimensional space. The time-evolution of the corresponding wave function is governed by the time-dependent Schrödinger equation (TDSE) $$\label{eq:tdse} {\mathrm{i}}\frac{\mathrm{d}}{\mathrm{d}t} \| \Psi(t) \rangle = \hat H(t) \| \Psi(t) \rangle \;,$$ where we set $\hbar=1$. We assume that the Hamiltonian, $$\hat{H}(t) = \hat{H}_0 + \hat{H}_1(t) \;,$$ consists of the stationary part, $$\hat{H}_0 = \hat{T} + \hat{V} + \hat{W} \;,$$ and an additional perturbation term, $\hat{H}_1(t)$. The explicit form of $\hat{H}_0$ in the spatial coordinates ${\mathbf}{r} \equiv ({\mathbf}{r}_1, \dots, {\mathbf}{r}_N)$ is given by the kinetic energy $$T({\mathbf}{r}) = \sum_{i=1}^N - \frac{1}{2m} \frac{\partial^2}{\partial {\mathbf}{r}_i^2} \;,$$ the trap energy $$V({\mathbf}{r}) = \sum_{i=1}^N \frac{1}{2}m\Omega^2 {\mathbf}{r}_i^2 \;,$$ and the interaction energy $$W({\mathbf}{r}) = \sum_{i<j} \frac{K_\alpha}{\|{\mathbf}{r}_i-{\mathbf}{r}_j\|^\alpha} \;.$$ If it is not further specified, we assume that the interaction is characterized by a repulsive power-law potential, $w(r) \propto 1/r^\alpha$, with the proportionality constant $K_\alpha$. Below we will concentrate on the important cases of Coulomb interaction ($\alpha = 1$) and dipole interaction ($\alpha=3$). After rescaling all lengths, times and energies in terms of $l_0= \left(1 / (m\Omega) \right)^{1/2}$, $\Omega^{-1}$ and $\Omega$, respectively, $\hat{H}_0$ takes the form $${H}_0({\mathbf}{r}) = \frac{1}{2}\sum_{i=1}^N \left \{-\frac{\partial^2}{\partial {\mathbf}{r}_i^2} + {\mathbf}{r}_i^2 \right \} +\sum_{i<j} \frac{\lambda}{\|{\mathbf}{r}_i-{\mathbf}{r}_j\|^\alpha} \;.$$ The dimensionless coupling parameter $\lambda>0$ determines the relative strength of the interaction energy [@Abrahamprb]. It is given by $$\lambda = \frac{1}{\hbar\Omega} \frac{K_\alpha}{l_0^\alpha} \;,$$ with, e.g., $K_1 = q^2 / (4\pi\epsilon)$, for the Coulomb interaction of particles with charge $q$ and the permittivity $\epsilon$, and $K_3 = C_\mathrm{dd} / (4\pi)$ for the interaction of polarized magnetic or electric dipoles, with the corresponding proportionality constant $C_\mathrm{dd}$ [@Lahaye2009]. While, in general, the interaction of two dipoles depends on their orientation, here we consider only the case of parallel dipoles [@filinov_2012_coll; @odell_2008]. In experiments, parallel alignment perpendicular to the plane of the dipoles is typically realized with external fields. Mode excitation --------------- To formally include a weak mode excitation by an arbitrary operator $\hat{Q}$, acting only at the time $t=0$, we specify $$\hat{H}_1(t) = \eta \delta(t) \hat{Q} \;,$$ where $\eta$ is a small real parameter. In the following, we assume that the eigenstates of $\hat{H}_0$ are given by $\left\{ \|0\rangle, \|1\rangle, \dots \right\}$ with the corresponding eigenvalues $ \left \{ E_0, E_1, \dots \right \}$. Furthermore, the system is supposed to be initially in the ground state $\|0\rangle$. With the help of first-order time-dependent perturbation theory, it can be shown that the time-dependent expectation value of an operator $\hat{A}$ without explicit time-dependence takes the form [@Abrahamprb] $$\label{eq:pert_expect} \langle \hat A \rangle(t) = \sum_{ij} c^_i c_j \exp \left\{{\mathrm{i}}\left(E_i - E_j \right) t \right\} \langle i\| \hat A\| j\rangle$$ with the constant coefficients $$\label{eq:pert_expans} c_k = \delta_{k,0} - {\mathrm{i}}\eta \langle k\| \hat Q \| 0\rangle \;.$$ The breathing mode is excited by the monopole operator $\hat{Q} = \hat{{\mathbf}{r}}^2 = \sum_i \hat{{\mathbf}{r}}_i^2$. For the typical observable $\hat{A}=\hat{{\mathbf}{r}}^2$, one can reduce Eq. (\[eq:pert\_expect\]) to the expression $$\label{eq:pert_ev_final} \langle \hat{\mathbf{r}}^2\rangle(t) = \langle 0 \| \hat{\mathbf{r}}^2\|0\rangle - {2\eta} \sum_i \| \langle 0 \| \hat{\mathbf{r}}^2\|i\rangle\|^2 \sin(\omega_{i,0} t) \;,$$ where $\omega_{i,0} = E_i - E_0$ are the mode frequencies. Equation (\[eq:pert\_ev\_final\]) reveals that the quantum breathing mode is characterized by a superposition of sinusoids with different frequencies. For a full characterization of the breathing mode, one has to calculate the eigenvalues of $\hat{H}_0$ and the matrix elements $ \langle 0 \| \hat{\mathbf{r}}^2\|i\rangle$. Separation ansatz ----------------- Although the breathing motion comprises a variety of possible frequencies, it was shown in recent works [@bauch09; @bauch10; @Abrahamprb] that the breathing mode is dominated by just two frequencies. One of these has the universal value $2\,\Omega$. This value can be explained by a formal decoupling of the wave function into a center-of-mass (CM) and a relative part, $$\|\Psi (t)\rangle = \|\Psi_\mathrm{cm}(t) \rangle \otimes \|\Psi_\mathrm{rel}(t) \rangle \;.$$ Such a decoupling is induced by the splitting of the Hamiltonian [@Kim2001], $$\hat{H}(t) = \hat{H}_\mathrm{cm}(t) + \hat{H}_\mathrm{rel}(t) \;,$$ where the contributions read $$\begin{aligned} \hat{H}_\mathrm{cm}(t) &= \frac{N}{2} {\hat{{\mathbf}{P}}}^2 + \frac{N}{2} {\hat{{\mathbf}{R}}}^2 + \eta \delta(t) N {\hat{{\mathbf}{R}}}^2 \;, \label{eq:hcm}\\ \hat{H}_\mathrm{rel}(t) &= {\sum_{i<j}}\Big\{ \frac{1}{2N} {\hat{{\mathbf}{p}}}_{ij}^2 + \frac{1}{2N} {\hat{{\mathbf}{r}}}_{ij}^2 + \hat w(\|{\hat{{\mathbf}{r}}}_{ij}\|) + \eta \delta(t) \frac{1}{N}{\hat{{\mathbf}{r}}}_{ij}^2 \Big\} \;. \label{eq:hrel}\end{aligned}$$ Here, we have introduced the CM and relative contributions, according to $$\begin{aligned} {\hat{{\mathbf}{O}}} &= \frac{1}{N} {\sum_{i=1}^{N}}{\hat{{\mathbf}{o}}}_i \;, \\ {\hat{{\mathbf}{o}}}_{ij} &= {\hat{{\mathbf}{o}}}_i - {\hat{{\mathbf}{o}}}_j \;.\end{aligned}$$ Since Eq. (\[eq:hcm\]) describes a non-interacting $d$-dimensional oscillator problem for the observables $ \hat{T}_\mathrm{cm} = N \, {\hat{{\mathbf}{P}}}^2 / 2$ and $ \hat{V}_\mathrm{cm} = N \,{\hat{{\mathbf}{R}}}^2 / 2$, one can conclude that the quantities $\langle \hat{V} \rangle(t)$ and $\langle \hat{T}\rangle(t)$ always contain oscillations with the frequency $\omega_\mathrm{cm}=2\,\Omega$. The contributions from the relative system, however, are non-trivial. Depending on the coupling parameter $\lambda$, the dominating frequency obtains the values $ \sqrt{3} \, \Omega \leq \omega_\mathrm{rel} \leq 2 \, \Omega $, for Coulomb interaction, and $ 2 \, \Omega \leq \omega_\mathrm{rel} \leq \sqrt{5} \, \Omega $, for dipole interaction [@bauch09; @bauch10; @henningphd]. This frequency corresponds to the first monopole excitation in the relative system. Sum rules ========= As has been shown, the value of $\omega_\mathrm{rel}$ can be extracted from the spectrum of $\hat{H}_0$. For more than just a few particles, however, an exact computation of the spectrum is impossible. Nevertheless, if it is possible to calculate the ground state of the system, one can make use of the quantum mechanical sum rules to gain some insight into the spectral properties. In this section, we give a brief review of the sum rules. Comprehensive overviews of the theory were presented for the study of collective resonances in nuclear physics [@lipparini89; @stringari82; @bohigas79; @bohigas76]. While the application to quantum gases has been subject of many recent investigations [@astrakharchik2005; @string; @pedri2008; @stringari96], we use the sum rules for the description of few-particle systems. We show how the conventional sum rule formulas can be modified to achieve very accurate results even for such small systems. Calculation of weighted moments ------------------------------- The weighted moments are defined by $$\label{eq:moments} m_k = \sum_{i=1}^\infty (E_i - E_0)^k \, \|\langle 0 \|\hat{Q}\|i\rangle\|^2 \;,$$ for any operator $\hat Q$ and any integer number $k \in \mathbb{Z} $. Containing the exact excitation energies, the moments can be used to define average excitation energies $$E_{k,l} = \left(\frac{m_k}{m_{k-l}} \right)^{1/l} \;,$$ for positive integer numbers $l$. In the literature [@lipparini89], one often finds the quantities $E_{k,2}$ and $ E_{k,1}$. The average excitation energies fulfill the relation $$\dots \geq E_{k+2,1} \geq E_{k+2,2} \geq E_{k+1,1} \geq E_{k+1,2} \geq \dots$$ and, especially, [@bohigas79] $$\lim_{k \to -\infty} E_{k,1} = \omega_{a,0} \;,$$ where the index $a$ corresponds to the lowest state excited by the operator $\hat{Q}$. Instead of directly evaluating the sum in Eq. (\[eq:moments\]), one can make use of the sum rules to simplify selected moments. In the following, we will concentrate on the calculation of the moments $m_3$, $m_1$ and $m_{-1}$ for the monopole operator $\hat{Q} = \hat{{\mathbf}{r}}^2$. For $m_1$ and $m_3$, one finds the sum rules $$\begin{aligned} m_1 &= \frac{1}{2} \langle 0 \| [\hat{Q}, [\hat{H}_0, \hat{Q} ]] \|0\rangle \\ &= 2 \ \langle 0 \| \hat{\mathbf{r}}^2 \|0\rangle \;.\end{aligned}$$ and $$\begin{aligned} \label{eq:sumrule_m3} m_3 &= \frac{1}{2} \langle 0 \| [ [\hat{Q},\hat{H}_0] , [\hat{H}_0, [\hat{H}_0, \hat{Q} ]]]\|0 \rangle \\ &= {8 \langle \hat{T} \rangle + 8 \langle \hat{V} \rangle + 2 \alpha^2 \langle\hat{W}\rangle} \;.\end{aligned}$$ The latter result can be obtained, reducing Eq. (\[eq:sumrule\_m3\]) to the evaluation of commutators with the types $ [\hat{{\mathbf}{p}}_i, \hat{{\mathbf}{r}}_j]$ and $ [\hat{{\mathbf}{p}}_i, 1 / \| \hat{{\mathbf}{r}}_j - \hat{{\mathbf}{r}}_k\|^\alpha ]$ for all occurring indices $i$, $j$, $k$. The moment $m_{-1}$ can be calculated by $$\label{eq:inverse_mono} m_{-1} = - \frac{\partial}{\partial \gamma}\langle \hat{\mathbf{r}}^2 \rangle_{\gamma=1} \;,$$ where the expectation value refers to the ground state of the Hamiltonian $$\hat{H}_0(\gamma) := \hat{T} + \gamma \hat{V} + \hat{W} \;.$$ This becomes clear if one writes the derivative in Eq. (\[eq:inverse\_mono\]) as $$\frac{\partial}{\partial \gamma}\langle \hat{\mathbf{r}}^2 \rangle_{\gamma=1} = \lim_{\epsilon \to 0} \frac{\langle\hat{\mathbf{r}}^2 \rangle_{\gamma=1+\epsilon} - \langle\hat{\mathbf{r}}^2 \rangle_{\gamma=1}}{\epsilon}$$ and evaluates $\langle\hat{\mathbf{r}}^2 \rangle_{\gamma=1+\epsilon}$ with stationary perturbation theory [@bohigas79]. Henceforth, we will use the notation $\mathrm{sr}(k,k-l):=E_{k,l}$. With the moments $m_3$ and $m_1$, one obtains the convenient sum rule formula $$\begin{aligned} \label{eq:brabecfinalfull} \mathrm{sr}(3,1)&= \left\{\left( 2+\alpha \right) + \left( 2-\alpha \right) \frac{\langle \hat{T} \rangle}{\langle \hat{V} \rangle} \right\}^{1/2} \nonumber \\ &= \left\{\left( 2+\alpha \right) + \left( 2-\alpha \right) \left( 1- \frac{\alpha\langle \hat{W} \rangle }{2\langle \hat{V} \rangle} \right) \right\}^{1/2}\;.\end{aligned}$$ The two different representations of $\mathrm{sr}(3,1)$ in Eq. (\[eq:brabecfinalfull\]) are due to the virial theorem $$2 \langle \hat{T}\rangle - 2 \langle \hat{V}\rangle + \alpha \langle \hat{W}\rangle = 0 \;.$$ For the special case of Coulomb interaction ($\alpha=1$), Eq. (\[eq:brabecfinalfull\]) has already been used by Sinha [@sinha2000]. A similar result for contact interaction was derived by Stringari [@stringari96]. Furthermore, we mention that Eq. (\[eq:brabecfinalfull\]) is a quantum generalization of the formula by Olivetti et al.* for classical systems [@olivetti2009]. A revealing property of Eq. (\[eq:brabecfinalfull\]) is the fact that it allows one to interpret the behavior of the breathing frequency with the ratios of the contributions to the total energy. Especially the classical limit, with $\langle\hat{T}\rangle/\langle\hat{V}\rangle=0$, and the ideal quantum limit, with $\langle\hat{W}\rangle/\langle\hat{V}\rangle=0$, are included. Another sum rule formula we will use in this paper is given by $$\label{eq:bound_dimensionless} \mathrm{sr}(1,-1)= \left\{ -2 \frac{\langle \mathbf{r}^2 \rangle}{ \left[ \partial \langle \mathbf{r}^2 \rangle/\partial \gamma \right]_{\gamma=1}} \right\}^{1/2} \;.$$ It has been presented as a rigorous upper bound of the breathing frequency by Menotti and Stringari [@string]. ![Comparison of the exact breathing frequencies, the conventional sum rule formulas and the improved sum rule formulas for a spin-polarized two-particle system in one dimension with Coulomb interaction (top) and dipole interaction (bottom).[]{data-label="fig:twopart_sr"}](fig1.eps) Improved sum rule formulas -------------------------- So far, we expressed the moments with the eigenstates and the eigenvalues of the full Hamiltonian $\hat{H} = \hat{H}_\mathrm{rel}+\hat{H}_\mathrm{cm}$. This may be disadvantageous for the estimation of $\omega_\mathrm{rel}$ if the weights of the center-of-mass terms are comparable to those of the relative terms. With the notations $$\hat{\mathbf{R}}_\mathrm{cm}^2 := N \hat{\mathbf{R}}^2$$ and $$\hat{\mathbf{r}}_\mathrm{rel}^2 := \frac{1}{N} \sum_{q<r} \hat{\mathbf{r}}^2_{qr} \;,$$ this can be understood more formally, expressing the moments as $$\begin{aligned} \label{eq:momentssplit} m_l &=\sum_{i=1}^\infty \left( E_{\mathrm{rel},i} - E_{\mathrm{rel},0} \right)^l \|\langle 0_\mathrm{rel} \| \hat{\mathbf{r}}_\mathrm{rel}^2 \| i_\mathrm{rel}\rangle \|^2 \nonumber \\ &\;\;\; + \sum_{k=1}^\infty \left( E_{\mathrm{cm},k} - E_{\mathrm{cm},0} \right)^l \|\langle 0_\mathrm{cm} \| \hat{\mathbf{R}}_\mathrm{cm}^2 \| k_\mathrm{cm}\rangle \|^2 \;.\end{aligned}$$ Using the formulas $\mathrm{sr}(k,j)$, the contributions from the second sum in Eq. (\[eq:momentssplit\]) may cause considerable inaccuracies of the estimator for $\omega_\mathrm{rel}=E_{\mathrm{rel},1} - E_{\mathrm{rel},0}$. Furthermore, it cannot be guaranteed that the sum rule formulas are upper bounds of the frequency $\omega_\mathrm{rel}$ if there exist non-vanishing contributions $E_{\mathrm{cm},k} - E_{\mathrm{cm},0} < \omega_\mathrm{rel}$. For example, this is the case for dipole interaction, where $\omega_\mathrm{cm} \leq \omega_\mathrm{rel}$ is valid for all coupling parameters. A simple solution to this problem is to eliminate the second sum in Eq. (\[eq:momentssplit\]). As all terms are known analytically, this can simply be accomplished by introducing the corrected moments $$\label{eq:momentscorr} m^_k := m_k - 2^{k-1}d \;.$$ Using these moments, we introduce the following improved sum rule formulas, $$\mathrm{sr}^(k,k-l) := \left( \frac{m^_k}{m^_{k-l}} \right)^{1/l} \;.$$ These formulas are not only expected to yield more accurate results than the conventional formulas, they also restore the character of the approximation as an upper bound for all $\alpha$. The special case $\mathrm{sr}^(3,1)$ is equivalent to the equilibrium formula presented in Ref. [@Brabec13]. To demonstrate the difference between the formulas $\mathrm{sr}$ and $\mathrm{sr}^$, we exactly determine the spectrum of the two-particle system. Expressing the relative vector ${\mathbf}{r}_{1,2}$ in the hyperspherical coordinates $\left(\rho, \phi_1, \ldots, \phi_{d-1} \right)$ [@montgomery08], one can reduce the relative problem to the equation $$\label{eq:final2part} \hspace{-1cm} \bigg\{-\frac{1}{2}\frac{\mathrm{d^2}}{\mathrm{d}\rho^2} + \frac{\rho^2}{2} + \frac{\left(l+(d-2)/2 \right)^2 - 1/4}{2\rho^2} + \frac{\lambda}{2^{\alpha/2}\rho^\alpha} - E_\mathrm{rel}\bigg\}u_l(\rho) = 0\;.$$ We obtain the spectrum by solving the corresponding eigenvalue problem in the matrix representation that arises from the expansion of $u_l$ in terms of FEDVR basis functions (see Sec. \[sec:hf\]). As the spectrum of the Hamiltonian $\hat{H}_\mathrm{cm}$ is known analytically, we can directly reconstruct the moments. In Fig. \[fig:twopart\_sr\], we compare the exact breathing frequencies with the sum rule estimators for Coulomb and dipole interaction. In both cases, the improvement of the sum rules leads to a higher accuracy. For dipole interaction, it is shown that, in fact, the frequencies from the estimators $\mathrm{sr}(3,1)$ and $\mathrm{sr}(1,-1)$ are below the exact values. At the same time, it can be seen that this problem is solved by the improved formulas. To conclude, we remark that the subtractive correction of the moments in Eq. (\[eq:momentscorr\]) does not depend on the particle number. Therefore, it can be expected that the improvement of the sum rules is only important for small systems, where the eigenvalues of $\hat{H}_\mathrm{rel}$ and $\hat{H}_\mathrm{cm}$ are of the same order. Nevertheless, we stress the difference between the results, because the influence of the CM subsystem was not mentioned in some other works, e.g., Ref. [@pedri2008] for the case of dipole interaction. Numerical methods ================= In order to obtain the breathing frequencies of one- and two-dimensional systems, we calculate the ground-state energies and apply the sum rule formulas. We use the Hartree-Fock (HF) and the Thomas-Fermi (TF) approximation to cover the range from small to large systems. For comparison, we also show some exact configuration interaction (CI) and time-dependent Hartree-Fock (TDHF) results, which were obtained in our previous work [@Abrahamprb]. In the following, some details of the methods are mentioned. Hartree-Fock approximation {#sec:hf} -------------------------- The Hartree-Fock approximation [@szabo96; @bonitzteub] reduces the $N$-body problem to an effective one-body problem, where the interactions are taken into account on the mean-field level. The central idea of the HF method is to assume that the solution of the Schrödinger equation is a single Slater determinant $$\|\Psi\rangle = \|\phi_1 \dots \phi_N \rangle \;.$$ Requiring that the set of single-particle spin orbitals $\|\phi_1\rangle, \dots \|\phi_N\rangle$ minimizes the total energy $E=\langle \Psi\|\hat{H}_0\|\Psi\rangle$, one can derive the Hartree-Fock equations $$\hat{F} \|\phi_k\rangle = \epsilon \| \phi_k\rangle \;,$$ where $$\hat{F} = \hat{h} + \sum_{i=1}^N \hat{J}_i - \hat{K}_i$$ is the Fock operator. For the trapped system, $\hat{h}$ takes the form $$h(\mathbf{r}) = -\frac{1}{2}\frac{\partial^2}{\partial {\mathbf}{r}^2} + \frac{1}{2}{\mathbf}{r}^2 \;.$$ The interaction is incorporated by the operators $\hat{J}_i$ and $\hat{K}_i$, which are defined by their actions $$\begin{aligned} \hat{J}_i \|\phi_k\rangle &= \lambda \langle \phi_i \|\hat{w}\|\phi_i\rangle\|\phi_k\rangle \;, \\ \hat{K}_i \|\phi_k\rangle &= \lambda \langle \phi_i \| \hat{w} \|\phi_k\rangle \| \phi_i \rangle \;.\end{aligned}$$ Expanding the spin orbitals in terms of a single-particle basis, we transfer the HF equations to a matrix equation. For one-dimensional systems, the utilized basis is a finite-element discrete variable representation [@rescigno2000; @balzerfedvr2010] (FEDVR). To avoid divergences, the Coulomb potential is regularized, according to the standard formula [@bauch09; @bauch10] $$w(\|r-r'\|)=\frac{1}{\left[ (r-r')^2+ \kappa^2 \right]^{1/2}} \;.$$ with the screening parameter $\kappa$. In our recent work [@Abrahamprb], we checked that converged results are obtained with $\kappa =0.1$. In two-dimensional systems, by contrast, such a screening is not necessary, as the basis is formed by the eigenfunctions of the harmonic oscillator in spherical coordinates. Our implementation for the calculation of the matrix elements is based on the code OpenFCI by Kvaal [@kvaal2008]. ![image](fig2.eps) Thomas-Fermi approximation -------------------------- For the extension of the HF results, we use the well-known Thomas-Fermi approximation [@thomas27; @fermi27; @spruch91; @march57]. It is expected to yield the correct trend for large particle numbers, because the oscillations of the single-particle density – which are not shown by the density in TF approximation – become negligible [@Abrahamprb]. In the one-dimensional case, the TF equation for spin-polarized particles reads $$\label{eq:tf1d} \frac{\pi^2}{2} n(r)^2 + \frac{1}{2}r^2 + \lambda \int \frac{n(r')}{\left[ (r-r')^2+ \kappa^2 \right]^{1/2}} \mathrm{d}r' = \mu\;,$$ where $\mu$ is the chemical potential. Fixing the particle number $N$ and requiring the normalization $$\int n(r) \mathrm{d}r = N \;,$$ one has to find the density $n(r)$ and the corresponding chemical potential which solve Eq. (\[eq:tf1d\]). We obtained our results on a grid, according to the following scheme: For each considered particle number, we start our calculation with a very small coupling parameter $\lambda$. At this, we choose the initial trial density $$n(r) = \frac{1}{\pi} \sqrt{2N-r^2} \;,$$ which is exact for a non-interacting system [@astrakharchik2011]. Updating the chemical potential and the density in a self-consistent procedure [@Guelveren2012], we finally obtain the density that solves Eq. (\[eq:tf1d\]). After that, we slightly increase $\lambda$ and start a new calculation, where the initial density is the final density from the previous calculation. This procedure is repeated until some final value of $\lambda$ is reached. For two-dimensional systems, we follow a simpler approach by Sinha [@sinha2000]. Making the ansatz $$\label{eq:tf2dansatz} n({\mathbf}{r}) = n(r) = \frac{1}{2\pi \gamma} (r_0^2 - r^2)$$ for the density, we determine the variational parameter $\gamma$ that minimizes the total energy $$\label{eq:tf_energy} E = \frac{1}{3}N^{3/2} \frac{1}{\gamma^{1/2}} +\frac{1}{3}N^{3/2} \gamma^{1/2} + \lambda \frac{512}{315} \frac{\sqrt{2}}{\pi \gamma^{1/4}} N^{7/4} \;.$$ In this equation, the single terms correspond to the kinetic energy, the trap energy and the interaction energy in this order. The method for the 2D case has the advantage that the parabolic ansatz for the density allows us to express the interaction energy analytically. In the 1D case, however, the calculations are more complex, because the integral $$\lambda \int \frac{n(r')}{\left[ (r-r')^2+ \kappa^2 \right]^{1/2}} \mathrm{d}r'$$ has to be calculated for each grid point $r$ in each iteration step. Results ======= We report the breathing frequencies for one- and two-dimensional systems. While only spin-polarized systems are treated in the 1D case, the occupation of the spin orbitals follows Hund’s rules in 2D systems. In the following, we consider the finite-size effects of small systems and the transition to large systems separately. In the end of this section, we explain the calculated breathing frequencies by the characteristics of the ground states. ![$N$-dependent breathing frequencies for Coulomb-interacting particles in two dimensions. The minima correspond to closed energy shells. For 2 and 3 particles, the improvement of the sum rule formulas is crucial for the correct qualitative description.[]{data-label="fig:2d_coulomb"}](fig3.eps) Small systems ------------- ### 1D $\,$\ We start with an analysis of small one-dimensional systems. In Fig. \[fig:sr\_1d\_omegaN\], we compare the breathing frequencies from different methods for the coupling parameters $\lambda=0.3$ and $\lambda=1$. On the one hand, we show the time-dependent HF and CI results from our previous work [@Abrahamprb]. On the other hand, we show the results obtained with static HF calculations in combination with the conventional and the improved sum rule formulas, respectively. One can draw the following conclusions: Comparing with the exact CI results, the results from the improved sum rule formulas are more accurate than the TDHF results. The improved accuracy of the sum rules can be explained by the fact that the static HF calculations could be performed with a larger basis than the time-dependent HF calculations. Furthermore, the error induced by the sum rules appears to be less important than the error induced by the HF approximation. Finally, one notices that the improvement of the sum rule formulas does not only lead to a higher accuracy for small particles, it also enables one to find the frequency minimum – which was already discovered in our previous work [@Abrahamprb] – at the correct position $N=6$, instead of $N=7$. However, the figure already reveals that the difference between both formulas tends to vanish for large particle numbers. ### 2D $\,$\ The step from 1D to 2D systems is usually numerically demanding, because – roughly estimated – the number of single-particle basis functions has to be squared. This is especially challenging for time-dependent calculations. However, using the sum rules and the HF method, we are able to investigate the finite-size effects of the breathing mode in 2D Coulomb systems for the first time. In Fig. \[fig:2d\_coulomb\], we show the frequencies for small ($N\leq22$) systems with an intermediate coupling parameter $\lambda=0.5$. Compared to the one-dimensional systems, we observe that the non-monotonic behavior of the frequency has increased. More precisely, the frequency has local minima for 2, 6, 12 and 20 particles. These “magic” particle numbers are well-known from various experimental and theoretical studies of quantum dots [@tarucha96; @reimann99; @reimann2002]. The corresponding configurations are very stable, because they are characterized by closed energy shells. Typically, experimental evidence for the occurrence of the magic configurations is given by the measurement of $N$-dependent addition energies. Apparently, the breathing mode provides an alternative tool for the diagnostics of these properties. Furthermore, it has been shown recently that important system properties, such as the kinetic and potential energy, can be obtained from the results of the breathing frequency [@Brabec13]. Using the HF approximation, we observe the same non-monotonic behavior for all coupling parameters $\lambda \leq 1$. Another important result is that – especially for 2 and 3 particles – the improvement of the sum rules is crucial to reproduce the correct behavior. ![Behavior of the breathing frequencies for large particle numbers in 1D (top) and 2D (bottom).[]{data-label="fig:largeN"}](fig4.eps) ![ Breathing frequency of a Coulomb system in a one-dimensional harmonic trap in the $(\lambda,N)$-plane. The data were produced with the Thomas-Fermi approximation ($\mathrm{sr}^(3,1)$, $N > 100$) and the Hartree-Fock approximation ($\mathrm{sr}^(1,-1)$, $N\leq 100$). The step around $N=100$ particles is due to the different approximation methods.[]{data-label="fig:2dsurvey"}](fig5.eps) Large systems ------------- ### 1D $\,$\ In our recent work [@Abrahamprb], we found several hints that an increase of the particle number in one-dimensional systems leads to a steady increase of quantum effects. This means that the breathing frequency slowly transitions to the quantum limit $\omega_\mathrm{rel}=2\,\Omega$. However, the calculations of the frequencies were restricted to 20 particles with coupling parameters $\lambda \leq 1$. In this work, we extend the results, reporting the frequencies for up to 10000 particles with $\lambda \lesssim 2$. To start the analysis for large systems, we compare the results from 1D HF calculations and the corresponding TF calculations for several hundred particles in the top plot of Fig. \[fig:largeN\]. As one can see, the difference between both approximations tends to vanish for large $N$. Hence, both approximations confirm the trend $\omega_\mathrm{rel} \to 2 \, \Omega$ in the limit $N\to\infty$. For an overview, the contour plot in Fig. \[fig:2dsurvey\] summarizes the behavior of the breathing frequency in the $(\lambda,N)$-plane. For all coupling parameters, the contours have a small step around $N=100$, where the results from HF and TF calculations have been joined. Nevertheless, the plot is suitable to trace the finite size effects, with the frequency minima for $N=6$, as well as monotonic behavior of large systems. As the frequencies approach the ideal limit only very slowly, we conclude with a remark that one can give a numerical proof for the large-$N$ behavior in the TF approximation. For that purpose, one uses the ideal density to calculate the kinetic energy, the trap energy and the interaction energy (with arbitrary fixed $\lambda$) as a function of the particle number. Such a calculation reveals that, for each coupling parameter $\lambda$, there exists a large particle number for which the interaction energy becomes negligible compared to the other two contributions. Consequently, the ideal density becomes a correct solution of Eq. (\[eq:tf1d\]) in the limit $N\to\infty$. In practice, that implies that an increase of $N$ leads to a faster convergence of the density if the initial guess in the self-consistent procedure is the ideal density. ### 2D $\,$\ In two-dimensional systems, the $N$-dependent behavior of the breathing frequency is contrary to that of one-dimensional systems. In the bottom plot of Fig. \[fig:largeN\], one can see that the frequency always reaches the value $\sqrt{3}\,\Omega$, in the limit $N\to\infty$. This behavior is confirmed for a broad range of coupling parameters in Fig. \[fig:2d\_2dsurvey\]. Hence, 2D systems always approach the classical limit if either the particle number or the coupling parameter is increased. ![ Breathing frequency of a Coulomb system in a two-dimensional harmonic trap in the $(\lambda,N)$-plane. The data were produced with the Thomas-Fermi approximation ($\mathrm{sr}^(3,1)$).[]{data-label="fig:2d_2dsurvey"}](fig6.eps) Compared to the results for the 1D systems, the ranges of numerically accessable $\lambda$ and $N$ are much broader for 2D systems. This can be explained by the fact that the parabolic ansatz for the density profile in the Thomas-Fermi approximation, Eq. (\[eq:tf2dansatz\]), is a drastic simplification of the problem. Hence, we can treat all coupling parameters with the same computational effort of a simple minimization procedure. Despite the restrictiveness of the parabolic ansatz, we can show that the results are fairly accurate. On the one hand, the good agreement of the TF and the HF results shown in Fig. \[fig:largeN\] justifies the parabolic ansatz, at least for weakly coupled systems. On the other hand, to check that the ansatz can also be used to reproduce the correct energies of strongly coupled systems, we show the total energies for a broad range of coupling parameters and different particle numbers in Fig. \[fig:2d\_classical\]. It turns out that – for sufficiently large particle numbers – the TF results converge to the results from classical molecular dynamics (MD) simulations in the limit of large $\lambda$. ![ $\lambda$-dependent total energies of a 2D Coulomb system for different particle numbers. The results from the Thomas-Fermi approximation are compared to the results from classical molecular dynamics simulations and the analytical values of an ideal Fermi gas. An analogous figure with exact results for the 1D case is to be found in the work of Astrakharchik and Girardeau [@astrakharchik2011].[]{data-label="fig:2d_classical"}](fig7.eps) Explanation of the asymptotic behavior with ground-state properties ------------------------------------------------------------------- In the following, we provide some supportive explanation for the different behaviors of one- and two-dimensional systems. First, we trace the characteristics with the help of a localization parameter that measures the extension of the system. Second, we introduce a simple estimator for the distinction between quantum and classical systems. ### Localization parameter $\,$\ Being inspired by the degeneracy parameter of a macroscopic homogeneous electron gas, one can define a localization parameter for the trap. This parameter is meant to express the nonideality of the system with geometric quantities. An estimator for the mean extension of the system is given by $$\begin{aligned} \sigma &= \left \{ 2 \langle \hat{V} \rangle \right \}^{1/2} \;.\end{aligned}$$ For the non-interaction systems, it has the value $$\begin{aligned} \sigma_\mathrm{ideal} &=\frac{1}{\sqrt{2}} N \;, &\qquad \mathrm{(1D)} \;,\\ \sigma_\mathrm{ideal} &= \sqrt{2/3} N^{3/4} \;, &\qquad \mathrm{(2D)} \;.\end{aligned}$$ With this, one can define the localization parameter $$\chi = \frac{\sigma_\mathrm{ideal}}{\sigma} \;,$$ which measures how much the extension of the system deviates from that of an ideal non-interacting quantum system. Starting with the value $\chi=1$ in the ideal system, the localization parameter decreases to zero in the course of the transition to the strongly coupled regime. In Fig. \[fig:join\], we compare the localization parameter with the breathing frequency for one- and two-dimensional systems in the $(\lambda,N)$-plane. The data were produced with the Thomas-Fermi approximation and the frequencies were estimated with the sum rule formula $\mathrm{sr}^(3,1)$. We concentrate on particle numbers $N\geq 100$ to make sure that finite-size effects are reduced. Furthermore, since the 1D TF method is reliable for small coupling parameters, we restrict the illustration to the regime $\lambda \lesssim 1$. The figure shows that the localization parameter is well-suited to track the qualitative behavior of the breathing frequency. Equal values of $\chi$ correspond to equal values of the breathing frequency. To illustrate this, the plot shows exemplary dotted lines which indicate equal values of each quantity. In the logarithmic plot, these are straight lines. Remarkably, the lines for 1D systems have a positive slope, while the lines for 2D systems have a negative slope. The corresponding best fits in the ranges $100\leq N \leq 10000$ and $0.1\leq\lambda \leq 1$ read $$\begin{aligned} \label{eq:geradengleichung1d} N &= (1.3 \times 10^4) \lambda^{2.45} \;, &\qquad \mathrm{(1D)} \;, \\ \label{eq:geradengleichung2d} N &= 100 \lambda^{-4} \;, &\qquad \mathrm{(2D)} \;.\end{aligned}$$ ![image](fig8.eps) ### Estimator for intermediate couplings $\,$\ Finally, we provide another rough estimator for a supportive understanding of the observed behavior. For the two-dimensional system, the straight lines in Fig. \[fig:2d\_classical\] demonstrate that the total energies can be approximated by those from non-interacting Fermi gases for weak couplings, $$E_\mathrm{ideal}^\mathrm{2D}= \frac{2}{3} N^{3/2} \;,$$ and those from purely classical systems for strong couplings, $$E_\mathrm{classical}^\mathrm{2D}= K \lambda^{2/3} N^{5/3} \;.$$ The last equation is derived by setting the kinetic energy to zero and minimizing the remaining terms in Eq. (\[eq:tf\_energy\]). The constant $K$ has the value $$K= \left( \frac{256\sqrt{2}}{315\pi} \right)^{2/3} + \frac{512\sqrt{2}}{315\pi} \left( \frac{256\sqrt{2}}{105\pi} \right)^{-1/3} \;.$$ As has been shown by Astrakharchik and Girardeau [@astrakharchik2011], one can also derive the corresponding terms for one-dimensional systems, where one obtains $$E_\mathrm{ideal}^\mathrm{1D}= \frac{1}{2} N^2$$ for non-interacting systems and $$E_\mathrm{classical}^\mathrm{1D}= \frac{3}{10} (3 \lambda N \ln N)^{2/3}N$$ for classical systems. In both 1D and 2D, we use the quantities $E_\mathrm{ideal}$ and $E_\mathrm{classical}$ to define a regime of intermediate coupling. Taking account of the $\lambda$-dependence of the total energy shown in Fig. \[fig:2d\_classical\], we mark this region by the coupling parameter $\tilde{\lambda}$ for which the classical estimator $E_\mathrm{classical}$ is equal to the ideal estimator $E_\mathrm{ideal}$. One obtains $$\label{eq:lambdaIntersect2D} \tilde{\lambda} = \left( \frac{2}{3K} \right)^{3/2} N^{-1/4}$$ for 2D systems, which is in agreement with the functional form of Eq. (\[eq:geradengleichung2d\]). For the 1D systems, one obtains $$\label{eq:lambdaIntersect1D} \tilde{\lambda} =\frac{1}{3} \left( \frac{5}{3} \right)^{3/2} \frac{N^{1/2}}{\ln N} \;.$$ In this equation, the relation between $N$ and $\tilde\lambda$ slightly differs from that in Eq. (\[eq:geradengleichung1d\]). However, we expect that both definitions come to agreement if one covers a larger area of the ($\lambda,N$)-plane shown in Fig. \[fig:join\]. The coupling parameter $\tilde{\lambda}$ roughly divides the systems into the ones with dominating quantum-like behavior ($\lambda < \tilde{\lambda}$) and the other ones with dominating classical behavior ($\lambda > \tilde{\lambda}$). Apparently, the trend for large systems is as follows: An increase of $N$ leads to a growing classical regime in 2D and a growing quantum regime in 1D. In Fig. \[fig:intersect\], Eqs. (\[eq:lambdaIntersect2D\]) and (\[eq:lambdaIntersect1D\]) are plotted. Remarkably, for one-dimensional systems, $\tilde{\lambda}$ has a minimum for $N=7$ and thus even reproduces the non-monotonic behavior. ![$N$-dependent behavior of the intermediate coupling parameter $\tilde \lambda$, for one- and two-dimensional systems.[]{data-label="fig:intersect"}](fig9.eps) Discussion ========== In this work, we presented an improvement of the conventional sum rule formulas for the calculation of the monopole excitation spectrum. The main idea was the elimination of the analytically known center-of-mass contributions to the weighted moments. It turned out that the improvement yields very accurate results for the frequencies of the quantum breathing mode. In our previous work, we discussed that the breathing mode is an indicator for the nonideality of the system [@Abrahamprb]. With Eq. (\[eq:brabecfinalfull\]), we can express this statement more precisely, mapping the breathing frequency to a fixed relation between the kinetic energy, the trap energy and the interaction energy. Having analyzed the breathing mode for various configurations, we can summarize the characteristics as follows. If the particle number of the system is fixed and the coupling parameter $\lambda$ is increased, the frequency always reaches the classical limit $\sqrt{3}\,\Omega$. This behavior corresponds to the well-known Wigner crystallization [@wigner34; @filinov2001]. The $N$-dependence of the breathing frequency is more complicated. [One-dimensional systems]{} are characterized by a frequency minimum for $N=6$, followed by a monotonic increase of the frequencies until the ideal limit $2 \,\Omega$ is reached. In [two-dimensional systems]{}, the breathing frequency of small systems reflects the shell structure of the configuration. For weakly interacting systems, we could observe minimum frequencies for configurations with closed energy shells. We expect that this $N$-dependent behavior will be replaced by the characteristics of strongly coupled clusters, for large coupling parameters. The most remarkable difference from one-dimensional systems is that the frequency reaches the classical value in the limit $N\to\infty$ with fixed coupling parameter $\lambda$. To summarize, we state that the sole knowledge of the parameter $\lambda$ is not sufficient to characterize the state of an interacting trapped quantum system, in particular, to decide whether the breathing mode is dominated by quantum effects or classical effects. Instead, one also has to take into account the particle number and the dimensionality of the system. We provided several estimators for the behavior of large one- and two-dimensional systems. The influence of the explicit form of the pairwise interactions on this behavior remains to be analyzed in future works. Furthermore, the presented improved sum rules should allow to investigate in more detail the influence of the spin statistics and the effects of strong correlations [@schmelcher13] e.g., with path integral Monte Carlo methods [@filinov2010; @schoof2011]. We expect that our results will be of interest for a variety of quantum many-body systems confined in harmonic potential wells. As an example, we mention ultracold ions and neutral atoms or combinations of them [@peotta_13] in traps. In these systems collective modes, including the breathing mode are easily excited [@peotta_13]. In Ref. [@harlander_11] it was suggested to use the sloshing mode of a trapped small ion ensemble to transmit information to a second ensemble. A similar concept might also be possible by using the quantum breathing mode. Acknowledgements ================ We thank Hauke Thomsen for providing the results for classical systems. This work is supported by the Deutsche Forschungsgemeinschaft via SFB-TR24, project A5 and a grant for CPU time at the HLRN. References {#references .unnumbered} ========== [48]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{} , **, (). , , , , (). , , , , (). , , , , , (). , , (). , , , , (). , , , , (). , , (). , , (). A. Filinov, J. B[öning]{}, M. Bonitz, and Yu.E. Lozovik, Yu. E., Phys. Rev. B [77]{}, 214527 (2008). A. Filinov, M. Bonitz, and Yu.E. Lozovik, Contrib. Plasma Phys. [41]{}, 357 (2001). P. Harmann, Z. Donko, and G.J. Kalman, Eur. Phys. Lett. [72]{}, 396-402 (2005). P. Ludwig, K. Balzer, A. Filinov, H. Stolz, and M. Bonitz, New J. Phys. [10]{}, 083031 (2008). J. B[öning]{}, A. Filinov, and M. Bonitz, Phys. Rev. B [84]{}, 075130 (2011) J. Schleede, A. Filinov, M. Bonitz, and H. Fehske, Contrib. Plasma Phys. [52]{}, 819-826, (2012) , , (). , , , , (). , , , , , (). , , (). , , , , , , , (). , , , , , , (). , , , , , (). , , , , , (). , , , , , (). , , , , , (). , , , , , , (). , , (). , , (). , , , , (). , , , , , (). , , , , , , (). , , (). , , (). , , (). , , , (). , **, (). , , , , , , (). , , (). , , (). , , , (), . , , Dover Books on Chemistry Series (, ). , (, ). , , (). , , , , (). , (), . , , (). , , (). , , (). , , (). , , (). , , (). , , , , , , (). , , , , , , , (). , , (). , , , , (). , , , , , , (). , , , , , , (). , , , , , **, ().	{ "pile_set_name": "ArXiv" }
--- abstract: 'The continuous development of superconducting electronics is encouraging several studies on hybrid Josephson junctions (JJs) based on superconductor/ferromagnet/superconductor (SFS) heterostructures, as either spintronic devices or switchable elements in quantum and classical circuits. Recent experimental evidence of macroscopic quantum tunneling and of an incomplete $0$-$\pi$ transition in tunnel-ferromagnetic spin-filter JJs could enhance the capabilities of SFS JJs also as active elements. Here, we provide a self-consistent electrodynamic characterization of // spin-filter JJs as a function of the barrier thickness, disentangling the high-frequency dissipation effects due to the environment from the intrinsic low-frequency dissipation processes. The fitting of the $I-V$ characteristics at $\SI{4.2}{\K}$ and at $\SI{300}{\milli\K}$ by using the Tunnel Junction Microscopic model allows us to determine the subgap resistance $R{_{\text{sg}}}$, the quality factor $Q$ and the junction capacitance $C$. These results provide the scaling behavior of the electrodynamic parameters as a function of the barrier thickness, which represents a fundamental step for the feasibility of tunnel-ferromagnetic JJs as active elements in classical and quantum circuits, and are of general interest for tunnel junctions other than conventional SIS JJs.' author: - 'H. G. Ahmad' - 'R. Caruso' - 'A. Pal' - 'G. Rotoli' - 'G. P. Pepe' - 'M. G. Blamire' - 'F. Tafuri' - 'D. Massarotti' bibliography: - 'refTJM.bib' title: 'Electrodynamics of highly spin-polarized tunnel Josephson junctions' --- \[sec:level1\]Introduction ========================== Ferromagnetic Josephson junctions (SFS JJs) have attracted considerable attention in the emerging fields of superconducting spintronics [@Bergeret2005; @Robinson2010; @Khaire2010; @Eschrig2011; @Linder2015] and as quantum and classical devices, since they have been proposed as energy-efficient memories [@Ryazanov2001; @Larkin2012; @Goldobin2013; @Niedzielski2018] and as passive $\pi$ shifters (phase inverters) in quantum circuits [@Ustinov2003; @Buzdin2005; @Feofanov2010]. However, in standard metallic SFS JJs, the $I{_{\text{c}}} R{_{\text{N}}}$ product is of the order of a few microvolts or less [@Buzdin2005; @Robinson2010; @Khaire2010], $I{_{\text{c}}}$ and $R{_{\text{N}}}$ being, respectively, the critical current and the normal state resistance. All these JJs are overdamped and thus characterized by high quasiparticle dissipation [@Buzdin2005; @Kato2007; @Massarotti2018]. This has hampered the use of ferromagnetic JJs as active switching elements in different classical and quantum circuits, since for such applications it is important to have a rather high $I{_{\text{c}}} R{_{\text{N}}}$ product and low damping [@Kato2007]. Low-dissipative ferromagnetic junctions use an additional insulating layer between one of the superconducting electrodes and the ferromagnetic barrier (SIFS JJs) [@Weides2006; @Bannykh2009; @Wild2010; @Larkin2012] or a ferromagnetic insulator barrier ($\text{SI}{_{\text{f}}}\text{S}$ JJs) [@Terzioglu1998; @Ioffe1999; @Kawabata2006; @Kawabata2010; @Vasenko2011] and may present key advantages for some applications, thus increasing the overall impact of JJs based on ferromagnetic barriers [@Bannykh2009; @Wild2010; @Larkin2012; @20; @21]. Heterostructures incorporating ferromagnetic insulator tunnel barriers have been theoretically proposed as quantum devices such as quiet ferromagnetic flux-qubits, based on anomalous $0$-$\pi$ transitions [@Ioffe1999; @Kawabata2006; @Kawabata2010], and as classical devices for digital electronics [@Terzioglu1998] and efficient electron refrigeration [@Kawabata2013refr]. Among the ferromagnetic insulators, has been used in superconducting spin valves [@Zhu2016], switchable JJs based on the interfacial exchange field [@Cascales2019], and in spin-filter // JJs, which represent the first $\text{SI}{_{\text{f}}}\text{S}$ JJs. Some of their properties have been studied in Refs. [@Senapati2011; @28; @Pal2014; @Massarotti2015; @Caruso2019]. The first evidence of macroscopic quantum tunneling (MQT) in ferromagnetic JJs is an indication that spin-filter JJs can be used as active quantum devices [@Massarotti2015]. These JJs are characterized by a thickness-dependent spin polarization because of the splitting in the insulator band structure induced by its magnetic exchange energy [@Senapati2011]. This property, together with the nontrivial magnetic structure of the barrier, causes an incomplete $0$-$\pi$ transition for the spin-filter efficiency ($P$) above $90\%$. Such an incomplete $0$-$\pi$ transition could be related to the presence of spin-triplet correlations, with implications for the $0$-$\pi$ technology [@Caruso2019]. \ This work aims at providing a self-consistent determination of the electrodynamic parameters in highly spin-polarized // junctions \[Fig. \[Fig:1\] (a)\]. For conventional JJs in the underdamped regime and with large $I{_{\text{c}}}$, measurements of Fiske steps have been successfully used to derive the capacitance $C$, while the amplitude of the hysteresis in the $I-V$ curve allows us to estimate the quality factor $Q$ within the resistively and capacitively shunted junction (RCSJ) model [@Barone1982; @Likharev1986]. However, when the junctions fall in the moderately damped regime or are characterized by low values of $I{_{\text{c}}}$ (or critical current density $J{_{\text{c}}}=I{_{\text{c}}}/A$, where $A$ is the cross section), it is more complicated to isolate the effective capacitance and the intrinsic dissipation sources of the junction from contributions due to the environment and the external circuit. Thus, more sophisticated methods are required for the analysis of the dissipation [@Barone1982; @Devoret1984; @Martinis1987; @Martinis1989; @Kautz1990]. We use the conventional tunnel junction microscopic (TJM) model to obtain a self-consistent estimation of $C$, $Q$, and the resistance associated with the quasiparticle dynamics $R{_{\text{sg}}}$, which are essential to define the electrodynamic properties of devices with $I{_{\text{c}}}$ down to few nanoamperes. The merit of this approach is a comparative analysis of tunnel-ferromagnetic JJs with barrier thickness spanning from $2.5$ to $\SI{4.0}{\nm}$. This allows us to explore substantially quite different transport regimes. If one wants to place the junction in a circuit or to couple it to a cavity [@Devoret2013; @Krantz2019], knowledge of the electrodynamic parameters and how they scale with the barrier thickness is fundamental. Therefore, this study provides a pathway to the engineering of tunnel-ferromagnetic JJs for specific applications. Methods {#Methods} ======= Dissipation in a JJ is frequency dependent and the quality factor is given by $Q(\omega)= \omega{_{\text{p}}} R(\omega) C$, where $\omega{_{\text{p}}}=(2eI{_{\text{c}}}/(\hbar C) )^{1/2}$ is the plasma frequency [@Barone1982; @Likharev1986]. In terms of the phase dynamics in the tilted washboard potential [@Barone1982], the phase particle in the supercurrent branch oscillates in one well of the potential at the plasma frequency $\omega{_{\text{p}}}$, while the voltage state involves steady motion of the phase particle ($\omega\sim 0$) [@Kautz1990]\[Fig. \[Fig:1\] (b)\]. High-frequency ($\omega\sim\omega{_{\text{P}}}$) dissipation at the switching from the superconducting to the resistive state (see the brown double arrow in Fig. \[Fig:1\] (b)) is determined by the high-frequency damping $Q_1$ and is mainly affected by the environment, i.e., the circuit in which the junction is embedded [@Devoret1984; @Martinis1987; @Martinis1989; @Kautz1990]. Low-frequency dissipation in the subgap branch of the $I-V$ curves ($\omega\sim0$) (see the green dashed arrow in Fig. \[Fig:1\] (b)) and the corresponding low-frequency damping $Q_0$ are affected by the intrinsic tunnel resistance, which is set by the subgap resistance $R{_{\text{sg}}}$ [@53; @54; @56; @Martinis1989; @Kautz1990] as $$\label{Eq:3} Q_0=\omega{_{\text{p}}} C R{_{\text{sg}}}.$$ The TJM model provides a complete microscopic description of a JJ, using the tunneling-Hamiltonian formalism [@Barone1982; @Likharev1986], and it is commonly employed for modeling superconducting quantum-interference devices (SQUIDs) and rapid-single-flux-quantum (RSFQ) logic gates and circuits [@43; @Odintsov1987; @44; @45]. This model can describe the subgap branch and the low-frequency electrodynamics of any JJ that shows tunneling conduction, without taking into account the exact expression for the current-phase relation (CPR), which could be nontrivial in the case of unconventional JJs such as the spin-filter JJs analyzed in this work [@Pal2014]. Therefore, it provides a powerful tool to investigate and determine $Q_0$ in junctions far from the underdamped regime and it enables us to isolate the dissipative components coming from the environment. It is particularly relevant since quasiparticle tunneling is a figure of merit in all classical and quantum circuits and has been, in general, a limit for standard SFS JJs. Measurements down to $\SI{300}{\milli\K}$ of the $I-V$ characteristics are performed by using an evaporation cryostat, while measurements down to $\SI{20}{\milli\K}$ are performed in a wet dilution refrigerator. Customized RC, copper powder filters, and room-temperature electromagnetic interference (EMI) filters guarantee high precision and resolution in the microvolt and nanoampere range. More details on the measurement setup can be found in Refs. [@41; @Massarotti2015], while information regarding the fabrication processes is given in Refs. [@Senapati2011; @Blamire2012; @Pal2014]. We measure the $I-V$ curves of junctions with different thickness $t$ at $\SI{20}{\milli\K}$, $\SI{300}{\milli\K}$, and $\SI{4.2}{\K}$ by current biasing the samples with a triangular waveform at $\SI{11.123}{\hertz}$ and by measuring the voltage across the junction. We extract $I{_{\text{c}}}$ at a voltage value far from the noise detected in the supercurrent branch. The normal resistance $R{_{\text{N}}}$ is calculated with a linear fit above $V{_{\text{g}}}=(\Delta_1+\Delta_2)/e=\SI{3.50}{\milli\V}$, with $\Delta_1$ and $\Delta_2$ being the gap energies of the two superconducting electrodes. TJM simulations are calculated by using <span style="font-variant:small-caps;">pscan2</span> [@url], a <span style="font-variant:small-caps;">python</span> module optimized to simulate SFQ logic-based superconducting circuits that typically work at $\SI{4.2}{\K}$. One of the subroutines of this software allows to simulate the $I-V$ characteristic of a JJ in electronic circuits with different degrees of complexity ${}^{1}$(The <span style="font-variant:small-caps;">pscan2</span> subroutine calculates time-averaged voltages $V$ across the device as a function of a bias current $I$). $I{_{\text{c}}}$, the Stewart-McCumber parameter $\beta=Q_0^2$, the gap voltage $V{_{\text{g}}}$, the ratio $I{_{\text{c}}}R{_{\text{N}}}/V{_{\text{g}}}$, and the ratio $R{_{\text{N}}}/R{_{\text{sg}}}$, $R{_{\text{sg}}}$ being the resistance of the subgap branch, are the software parameters that govern the shape of the $I-V$ curves. $I{_{\text{c}}}$ and $V{_{\text{g}}}$ measured directly from the $I-V$ curves in our experimental setup are affected by errors of $1\%$ and $2\%$, respectively, while $R{_{\text{N}}}$ is obtained by fitting the ohmic region of the $I-V$ curves and is affected by an error of $3\%$. Since these values can be obtained with high precision, they can be set as fixed parameters, as well as the ratio $I{_{\text{c}}}R{_{\text{N}}}/V{_{\text{g}}}$. $\beta$ and the ratio $R{_{\text{N}}}/R{_{\text{sg}}}$ are the fitting parameters. The Stewart-McCumber parameter modifies the amplitude of the hysteresis in the $I-V$ curve, without affecting the subgap region \[Fig. \[Fig:5\] (a)\]. The ratio $R{_{\text{N}}}/R{_{\text{sg}}}$, instead, modifies both the subgap shape and the hysteresis amplitude \[Fig. \[Fig:5\] (b)\]. In our simulations, we reproduce the current biasing of a JJ with a current generator in series with the filtered lines of our experimental setup (approximately $\SI{200}{\ohm}$). For each spin-filter JJ with a certain thickness $t$, we choose the best-fit parameters $Q_0$ and $R{_{\text{sg}}}$ in such a way that the deviations from the experimental curves are minimal. The errors on $Q_0$ and $R{_{\text{sg}}}$ represent the range of values that provide a significant overlap between the experimental $I-V$ characteristics and the simulated curves within the TJM model and are of $6\%$ and $10\%$ respectively. ![$I-V$ curves in normalized units simulated by using <span style="font-variant:small-caps;">pscan2</span>, by fixing $V{_{\text{g}}}=1.4$, $I{_{\text{c}}}R{_{\text{N}}}=1.0$: in a) we fixed $R{_{\text{N}}}/R{_{\text{sg}}}=0.1$ and we changed $\beta$; in b) we fixed $\beta=10$ and we changed $R{_{\text{N}}}/R{_{\text{sg}}}$.[]{data-label="Fig:5"}](TJM){width="8.6cm"} The thicknesses in the junctions analyzed in this work range from $2.5$ to $\SI{4.0}{\nm}$, while $P$ ranges from $88\%$ to $98\%$, respectively (Tab. \[Tab:1\]), falling in the highly spin-polarized regime. In the special case of spin-polarized systems, $R{_{\text{N}}}$ has to be redefined as the combination of the two resistances associated with the presence of different tunnel conductances for spin-up and spin-down electrons, because of the spin-filtering effect (see the Appendix). The subgap shape in the $I-V$ curves is linked to the quasiparticle dynamics in the junction. The quasiparticle current in a spin-polarized system has been expressed theoretically and analytically in the case of symmetric spin-filter JJs by taking into account the magnetic nature of the tunnel barrier and the spin-filtering effect [@Bergeret2012]. Simple calculations allow us to verify that the quasiparticle current in these devices has the same expression both in the case of conventional tunnel JJs, i.e., for $P=0$ and a magnetic exchange field in the tunnel barrier $h=0$, and in the ideal and extreme situation of perfect spin polarization ($P=100\%$) (see the Appendix). The conventional TJM model does not take into account the magnetic exchange field of the $\text{I}{_{\text{f}}}$ barrier in spin-filter junctions, which can be important in the intermediate regime between these two extreme cases. The systematic fitting of the $I-V$ curves at $\SI{4.2}{\K}$ as a function of the barrier thickness confirms that the shape of the $I-V$ curves is mostly determined by the standard parameters of the junction ($C$, $R{_{\text{sg}}}$, $Q_0$). Further consistency is given by the $I-V$ fitting through the frequency dependent RCSJ model for the junction with the highest $P$, as shown in section \[Results\]. Below $\SI{4.2}{\K}$, // JJs with $P$ up to $98\%$ show an incipient $0$-$\pi$ transition in the $I{_{\text{c}}}(T)$ curves, which can be understood in terms of spin-triplet correlations arising because of the presence of both the spin-filtering effect and a nonuniform magnetic activity in the $\text{I}{_{\text{f}}}$ barrier [@Caruso2019]. Therefore, deviations between the experimental curves and simulations at $\SI{300}{\milli\K}$ can be due to the magnetic nature of the barrier, which the TJM model does not take into account. However, the estimated fitting parameters give an upper bound to $Q_0$ and $R{_{\text{sg}}}$, and a term of comparison for possible applications of spin filter JJs at very low temperatures, as discussed in section \[Discussions\]. Results {#Results} ======= As one can observe in Fig. \[Fig:4\], the critical current density $J{_{\text{c}}}(t)$, with cross section $A=\SI{49}{\micro\m^2}$, and the $R{_{\text{N}}}A(t)$ curves at $\SI{300}{\milli\K}$ obey to a typical tunnel behavior, thus confirming the insulating nature of the ballistic barrier [@Caruso2019]. $R{_{\text{N}}}A(t)$ exhibits the characteristic exponential thickness dependence: $$R{_{\text{N}}}A(t)=\frac{2tA}{3\sqrt{4m{_{\text{e}}}\bar{E}}}\left(h/e\right)^2e^{\frac{2t}{\hbar}\sqrt{4m{_{\text{e}}} \bar{E}}}, \label{Eq:1}$$ where $m{_{\text{e}}}$ is the free electron mass, $e$ is the electron charge and $\bar{E}$ is the mean energy-barrier height seen by the charge carriers [@51]. ![In black, $R{_{\text{N}}}A (t)$ product measured at $\SI{300}{\milli\K}$ (black circles) as a function of the barrier thickness along with a fit using Eq. \[Eq:1\] (dashed curve). In red, critical current density $J{_{\text{c}}}(t)$ measured at $\SI{300}{\milli\K}$ as a function of the thickness $t$ (diamonds) along with an exponential fit (full line). The error bars are of the order of $1\%$ on measured values for $I{_{\text{c}}}$, and of $3\%$ for $R{_{\text{N}}}$.[]{data-label="Fig:4"}](Ic){width="8.6cm"} $t\;(\si{\nm})$ $P\;(\%)$ $I{_{\text{c}}}R{_{\text{N}}}@\SI{4.2}{\K}\;(\si{\micro\volt})$ $I{_{\text{c}}}R{_{\text{N}}}@\SI{300}{\milli\K}\;(\si{\micro\volt})$ $I{_{\text{c}}}R{_{\text{N}}}@\SI{20}{\milli\K}\;(\si{\micro\volt})$ ----------------- ----------- ----------------------------------------------------------------- ----------------------------------------------------------------------- ---------------------------------------------------------------------- $2.5$ $88$ $156\pm3$ $179\pm4$ - $3.0$ $93$ $24.2\pm0.5$ $38.3\pm0.8$ $44.0\pm0.9$ $3.5$ $96$ $9.9\pm0.2$ $19.0\pm0.4$ - $4.0$ $98$ $2.8\pm0.1$ $5.1\pm0.2$ $6.1\pm0.2$ \[Tab:3\] In Tab. \[Tab:3\], we report $I{_{\text{c}}}R{_{\text{N}}}$ at $\SI{4.2}{\K}$, at $\SI{300}{\milli\K}$, and at $\SI{20}{\milli\K}$, measured from the $I-V$ curves. The characteristic voltage $I{_{\text{c}}} R{_{\text{N}}}$ decreases by increasing the barrier thickness, as well as the corresponding Josephson frequency $\omega{_{\text{c}}}=I{_{\text{c}}}R{_{\text{N}}}2e/\hbar$. At $\SI{4.2}{\K}$, it ranges from $\SI{80}{\giga\hertz}$ for the thinnest junction to $\SI{1}{\giga\hertz}$ for the thickest one. At lower temperatures, we measure higher values of the $I{_{\text{c}}}R{_{\text{N}}}$ product. These values are higher than those usually achieved for SFS JJs and comparable to those of some SIFS heterostructures [@Bergeret2005; @Buzdin2005; @Weides2006; @Bannykh2009; @Khaire2010; @Larkin2012]. For barrier thicknesses lower than $\SI{2.5}{\nm}$, the characteristic voltage is as high as a few millivolts [@Senapati2011; @Massarotti2015; @Caruso2019]. $t\;(\si{\nm})$ $R{_{\text{sg}}}@\SI{4.2}{\K}\;(\si{\ohm})$ $R{_{\text{sg}}}@\SI{300}{\milli\K}\;(\si{\ohm})$ $Q_0@\SI{4.2}{\K}$ $Q_0@\SI{300}{\milli\K}$ $C$ ($\si{\pico\farad}$) ----------------- --------------------------------------------- --------------------------------------------------- -------------------- -------------------------- -------------------------- $2.5$ $59$ $93$ $16$ $48$ $1.6\pm0.3$ $3.0$ $82$ $350$ $7.3 $ $35$ $1.1\pm0.2$ $3.5$ $440$ $1700$ $6.6$ $32$ $0.26\pm0.05$ $4.0$ $3000$ $13000$ $2.6$ $26$ $0.018\pm0.003$ \[Tab:1\] ![Measured $I-V$ curves at $\SI{4.2}{\K}$ (black points) and TJM model simulation by using <span style="font-variant:small-caps;">pscan2</span>\ software (red curve) for high spin-filter JJs with thicknesses $t$: a) $\SI{2.5}{\nm}$, b) $\SI{3.0}{\nm}$, c) $\SI{3.5}{\nm}$, d) $\SI{4.0}{\nm}$. Quality factor $Q_0$ and subgap resistance $R{_{\text{sg}}}$ estimated from the simulations are collected in Tab. \[Tab:1\]. The blue squares in d) represent the frequency dependent RCSJ model fit curve, obtained for $Q_0=2.8$ and $Q_1=0.13$.[]{data-label="Fig:2"}](fit){width="8.6cm"} In Fig. \[Fig:2\], we present the $I-V$ curves measured at $\SI{4.2}{\K}$ (black points) and TJM simulations (red straight lines) obtained by using <span style="font-variant:small-caps;">pscan2</span>. We collect in Tab. \[Tab:1\] the fitting parameters $R{_{\text{sg}}}$ and $Q_0$. The thicker the barrier is, the higher is the subgap resistance [@53; @54; @56]. The low-frequency quality factor $Q_0$ decreases with the thickness. This is due to both the decrease of $I{_{\text{c}}}$ and of $C$ of the barrier with the thickness [@Barone1982]. ![In black: capacitance values of spin-filter JJs as a function of the barrier thickness $t$ (black circles), along with parallel-plate capacitance $C(t)$ fit (dashed curve). In red: specific capacitance $C{_{\text{s}}}$ of the analyzed junctions as a function of $R{_{\text{N}}}A$ (red diamonds) along with a tunnel barrier model fit (straight line, see Eq. \[Eq:2\]). The error bars on $C$ and $C{_{\text{s}}}$ are calculated using the propagation of the errors on $R{_{\text{sg}}}$, $Q_0$ and $I{_{\text{c}}}$.[]{data-label="Fig:3"}](Ct){width="8.6cm"} The decrease in the $Q_0$ factor for increasing $t$ indicates a smooth transition from an underdamped regime ($Q_0\sim10$) to a moderately damped regime with phase diffusion (PD) ($Q_0\sim 1$) [@Kautz1990; @Massarotti2012; @Stornaiuolo2013]. The presence of the PD regime is confirmed by the finite slope in the supercurrent branch for the junction with a $\SI{4.0}{\nm}$-thick barrier [@Kautz1990], which <span style="font-variant:small-caps;">pscan2</span> simulations cannot reproduce, since they do not take PD processes into account. Monte Carlo simulations can reproduce the finite slope in the supercurrent branch, taking into account multiple escape and retrapping processes in the phase dynamics, which are particularly relevant for low values of the $Q_1$ factor and $E{_{\text{J}}}$ comparable with the thermal energy $k{_{\text{B}}}T$, as in the case of the spin-filter junction with a $\SI{4.0}{\nm}$-thick barrier [@Kautz1990]. In Fig. \[Fig:2\] (d), a Monte Carlo fit according to the frequency dependent RCSJ model is shown (blue square points), with high-frequency $Q_1 = 0.13$ and low-frequency $Q_0 = 2.8$. This is consistent with the outcomes based on the TJM model. The environment plays an important role in determining the value of $Q_1$. The ratio between the low- and high-frequency quality factors $Q_1/Q_0$ equals the ratio between the resistance of the environment $R{_{\text{env}}}$ and the subgap resistance, $R{_{\text{env}}}/R{_{\text{sg}}}$, since $Q_0$ is written in terms of the quasiparticle dissipation (Eq. \[Eq:3\]), while $Q_1$ can be expressed in terms of the environment resistance $R{_{\text{env}}}$ [@Martinis1987; @Martinis1989; @Kautz1990]. For the junction with a barrier thickness of $\SI{4.0}{\nm}$, $R{_{\text{env}}}$ is approximately $\SI{150}{\ohm}$, which is of the same order of magnitude of the resistance of the lines in our experimental setup. Our analysis allows us to estimate the capacitance $C$ of the barrier and its dependence on the barrier thickness, using Eq. \[Eq:3\] (see Tab. \[Tab:1\]). The value of $C$ for the thinnest barrier is consistent with a previous estimation based on SCDs measurements [@Massarotti2015]. In Fig. \[Fig:3\], we plot the junction capacitance $C$ as a function of the barrier thickness $t$ (black circle points) and the fitting function for the capacitance in a parallel-plate capacitor $C=\epsilon_0\epsilon{_{\text{r}}} A/t$ (black dashed line), where $\epsilon_0=\SI{8.85}{\pico\farad/m}$ is the vacuum dielectric permittivity and $\epsilon{_{\text{r}}}$ is the relative permittivity, which acts as a fitting parameter. The estimated $\epsilon{_{\text{r}}}=(20\pm8)$ is consistent with the permittivity $\epsilon{_{\text{r}}}=26.5$ obtained with spectroscopic measurements on isolated thin films [@63], providing an additional validation of the fitting procedure. The $R{_{\text{N}}}A$ product as a function of the specific capacitance $C{_{\text{s}}}=C/A$ (red diamonds in Fig. \[Fig:3\]), follows the expected behavior for tunnel JJs [@Kawakami2003]. The red line in Fig. \[Fig:3\] is the function $$R{_{\text{N}}}A(C{_{\text{s}}})=\frac{2A\epsilon_0\epsilon{_{\text{r}}}}{3C{_{\text{s}}}\sqrt{4m{_{\text{e}}}\bar{E}}}\left(h/e\right)^2e^{\frac{2\epsilon_0\epsilon{_{\text{r}}}}{\hbar C{_{\text{s}}}}\sqrt{4m{_{\text{e}}} \bar{E}}}, \label{Eq:2}$$ which is obtained by replacing $t$ in equation \[Eq:1\] with its dependence on the specific capacitance $C{_{\text{s}}}$, $t=\epsilon_0\epsilon{_{\text{r}}}/C{_{\text{s}}}$. In Fig. \[Fig:7\] (a) we show the $I-V$ characteristic measured at $\SI{300}{\milli\K}$ (black points) and TJM simulations (red straight lines) for the junction with a $\SI{4.0}{\nm}$-thick barrier, which corresponds to the highest spin-filtering efficiency analyzed in this work. $Q_0$ and $R{_{\text{sg}}}$ for all the devices are collected in Tab. \[Tab:1\]. The best-fit curve at $\SI{300}{\milli\K}$ is characterized by a smaller $R{_{\text{sg}}}$ compared to the experimental one. We can attribute this deviation to the unconventional magnetic activity discussed in Ref. [@Caruso2019], which is at a maximum in the case of most spin-polarized JJs, where the magnetic nature of the barriers manifests in a steep increase of $I{_{\text{c}}}(T)$ below $\SI{2}{\K}$ \[Fig. \[Fig:7\] (b)\] [@Caruso2019]. The conventional TJM model does not take the magnetic activity in the $I{_{\text{f}}}$ barrier into account, nor the spin-dependent tunneling mechanism and the unconventional thermal behavior of $I{_{\text{c}}}$, thus giving a systematic underestimation of $R{_{\text{sg}}}$, as shown in Fig. \[Fig:7\] (a). However, despite the presence of these deviations, $R{_{\text{sg}}}$ estimated for all the junctions increases when decreasing the temperature $T$ due to the tunnel nature of the conduction mechanisms in the system [@54] and $Q_0$ increases because of the increase of $R{_{\text{sg}}}$, as expected. ![In a): measured $I-V$ characteristic at $\SI{300}{\milli\K}$ (black points) and TJM model simulation by using <span style="font-variant:small-caps;">pscan2</span> software (red straight line) for the spin-filter JJ with $t=\SI{4.0}{\nm}$. Quality factor $Q_0$ and subgap resistance $R{_{\text{sg}}}$ estimated from the simulations are collected in Tab. \[Tab:1\]. In b): incipient $0$-$\pi$ transition in the $I{_{\text{c}}}(T)$ for the spin-filter JJ with $t=\SI{4.0}{\nm}$, as reported in Ref. [@Caruso2019].[]{data-label="Fig:7"}](TJM300mK){width="8.6cm"} In Fig. \[Fig:9\], we finally present a comparison between the normalized $I-V$ curves measured at $\SI{4.2}{\K}$ and $\SI{300}{\milli\K}$ and the $I-V$ characteristics at $\SI{20}{\milli\K}$ for two of the junctions with the highest $P$: in a) $t=\SI{3.0}{\nm}$ and $P=93\%$ and b) $t=\SI{4.0}{\nm}$ and $P=98\%$. The current is normalized to $I{_{\text{c}}}$, while the voltage is normalized to the switching value $V{_{\text{s}}}$ to compare the subgap branches of the $I-V$ curves. $I{_{\text{c}}}$ at $\SI{20}{\milli\K}$ are $\SI{4.75}{\micro\ampere}$ for a) and $\SI{29}{\nano\ampere}$ for b). The critical current at $\SI{300}{\milli\K}$ is $\SI{4.64}{\micro\ampere}$ for a) and $\SI{26.7}{\nano\ampere}$ for b). The amplitude of the hysteresis in the $I-V$ curves increases when going toward lower temperatures, pointing to an increase of $Q_0$ and also as a consequence of $R{_{\text{sg}}}$. ![Normalized $I-V$ curves at $\SI{20}{\milli\K}$ (black points), $\SI{300}{\milli\K}$ (red squares) and $\SI{4.2}{\K}$ (blue diamonds) for spin-filter JJs with thicknesses $t$: a) $\SI{3.0}{\nm}$ and b) $\SI{4.0}{\nm}$. The current is normalized to the critical current $I{_{\text{c}}}$, while the voltage is normalized to the switching value $V{_{\text{s}}}$.[]{data-label="Fig:9"}](comparison_low){width="8.6cm"} Discussions and concluding remarks {#Discussions} ================================== The use of the TJM model on parent compounds allows to achieve a consistent and robust picture of tunnel-ferromagnetic JJs with a quantitative insight on key electrodynamic parameters, such as $Q_0$, $R{_{\text{sg}}}$ and $C$. The estimated $Q_0$ values at $\SI{4.2}{\K}$ are up to two orders of magnitude higher compared to those of standard SFS heterostructures that typically operate in the overdamped regime like SNS JJs, with $\beta$ ranging from $10^{-3}$ to $10^{-1}$ [@Bulaevskii1977; @Barone1982; @Likharev1986]. $Q_0$ values are of the same order of magnitude of conventional SIS junctions commonly used to drive and for the read-out of components in quantum and classical circuits [@Castellano2006; @Shcherbakova2015]. Moreover, the $Q_0$ values increase up to one order of magnitude for the $\SI{4.0}{\nm}$ thick barrier, when lowering $T$ to $\SI{300}{\milli\K}$. This sets a lower limit that can only increase at lower temperatures (see Fig. \[Fig:9\]), and suggests possible implementation of spin-filter JJs in low-dissipative $\pi$-qubits. $\pi$-superconducting RF-SQUIDS with ferromagnetic-insulating barriers were only theoretically suggested as quiet qubits efficiently decoupled from the fluctuations of an external magnetic field [@Kawabata2006; @Kawabata2010]. A spin-filter JJ with $t=\SI{3.0}{\nm}$ and an $A\sim\SI{50}{\micro\m^2}$ has an estimated charging energy $E{_{\text{c}}}\sim\SI{900}{\micro\K}$, and a Josephson energy at $\SI{20}{\milli\K}$ $E{_{\text{J}}}\sim\SI{100}{\K}$, which means $E{_{\text{J}}}/E{_{\text{c}}}\sim10^5$, suitable for a flux-qubit [@Devoret2013; @Krantz2019]. In the frame of the $0$-$\pi$ technology, spin filter JJs analyzed in this work can be implemented also as complementary $\pi$-junctions for phase-bias of conventional flux-qubit (passive elements) in which the high values of the subgap resistance $R{_{\text{sg}}}$ could increase the dephasing time of the overall circuit [@Kato2007]. The dephasing time is proportional to $E{_{\text{J}}}^2R{_{\text{sg}}}$ of the $\pi$-junction [@Kato2007]. The subgap resistance in this work ranges from tens of ohms to some kilo-ohms at $\SI{4.2}{\K}$, but when decreasing $T$, $R{_{\text{sg}}}$ increases from a factor $2$ to $5$ increasing $t$ at $\SI{300}{\milli\K}$. The dephasing time of a circuit with a spin-filter JJ with a $t=\SI{3.0}{\nm}$ thick barrier can be comparable with that of circuits with SIFS $\pi$-junctions [@Weides2006], and can increase of at least a factor $100$ compared to circuits with standard metallic $\pi$ shifters [@Feofanov2010; @Shcherbakova2015]. In standard metallic SFS JJs typical resistances are at most $\sim\SI{1}{\ohm}$, while $R{_{\text{sg}}}$ for the junction with a $\SI{3.0}{\nm}$ thick barrier at dilution temperature is at least $\SI{350}{\ohm}$. The subgap resistance is crucial for the engineering of transmon qubits. As suggested in Ref. [@Serniak2018], in these circuits quasiparticle tunneling can affect the relaxation and coherence times [@Serniak2018]. The values obtained in this work can be promising even for potential application of tunnel-ferromagnetic JJs in transmon qubits. The order of magnitude of the ratio $E{_{\text{J}}}/E{_{\text{c}}}$ for the investigated junctions scales with the thickness from $10^6$ to $10$. Adapting the area of the devices to conventional dimensions in transmon qubits ($A\sim \SI{1}{\micro\m^2}$), lower values of $E{_{\text{J}}}/E{_{\text{c}}}$ can be achieved, falling in the typical range of transmon qubit [@11; @Devoret2013; @Krantz2019]. As an example, reducing the cross section to $A\sim\SI{1}{\micro\m^2}$, $E{_{\text{J}}}$ of the spin-filter JJ with $t=\SI{3.5}{\nm}$ becomes $\sim\SI{280}{\milli\K}$, while $E{_{\text{c}}}$ becomes $\sim\SI{180}{\milli\K}$, so that $E{_{\text{J}}}/E{_{\text{c}}}\sim2$. Moreover, reducing the junction area by a factor $\sim 50$, $R{_{\text{sg}}}$ should increase up to values of the order of $50-\SI{100}{\kilo\ohm}$, thus further reducing quasiparticle noise. The same arguments are valid for the junction with $t=\SI{4.0}{\nm}$ barrier, which is characterized by a subgap resistance $\sim 10$ times higher. In conclusion, this work represents the first electrodynamic characterization of spin-filter JJs, and a fundamental step to use these devices as active elements in superconducting circuits. Our comparative and self-consistent approach allows to obtain the scaling-law as a function of the barrier thickness of fundamental electrodynamic parameters, such as $C{_{\text{s}}} (t)$, $R{_{\text{sg}}}(t)$ and $Q_0(t)$, providing the possibility to engineer spin-filter JJs as a function of the junction area in order to meet specific circuit requirements. Even if the ferromagnetic JJs analyzed in this work are not ideal SIS JJs, we succeeded in the determination of these fundamental electrodynamic parameters at $\SI{4.2}{\K}$ by using a conventional TJM model, and we provided a lower bound for $R{_{\text{sg}}}$ and $Q_0$ at $\SI{300}{\milli\K}$. The underestimation of $R{_{\text{sg}}}(t)$ observed at $\SI{300}{\milli\K}$ is due to the absence of the spin-filtering effect and of the magnetic activity of the barrier in the TJM model. Further studies are needed to implement a microscopic modelization of peculiar properties of the $\text{I}{_{\text{f}}}$ barrier, such as spin-selective tunneling mechanisms and triplet correlations. The same approach can be successfully extended to different types of tunnel junctions other than conventional SIS JJs, for instance, multilayered SIFS JJs, and can provide the possibility to engineer special circuits other than conventional flux and transmon qubits, in which ferromagnetic-tunnel junctions can be tuned by external microwaves and are capacitively coupled to standard superconducting circuits [@20; @21]. The authors thank G. Campagnano, S. Poletto, A. Miano and A. Kirichenko for fruitful discussions. H.G.A., R.C., D.M. and F.T. also thank NANOCOHYBRI project (COST Action CA 16218). The spin-filtering effect {#Appendix2} ========================= Josephson junctions with barriers show a spin-filtering effect due to the simultaneous presence of tunnel conduction mechanisms and a magnetic exchange field $h$ in the ferromagnetic phase of the barrier. When the becomes ferromagnetic ($T{_{\text{Curie}}}\sim\SI{40}{\K}$), the presence of exchange interactions leads to a spin selectivity of the tunneling processes: spin up (down) will see different barrier heights $E_{\uparrow(\downarrow)}=E_0\mp h/2$, with $E_0$ energy barrier height in the paramagnetic phase of the . Carriers that relive a higher barrier will be filtered out, thus giving a net spin-polarized current [@Senapati2011]. ![Measured $R(T)$ curves (black points) and semiconducting model fit (red straight line) above the Curie temperature $T{_{\text{Curie}}}$ of the device (dashed black line) for the junction with $t=\SI{4.0}{\nm}$. The dash dotted blue line indicates the temperature $\SI{15}{\K}$ at which we calculate the $P$.[]{data-label="Fig:6"}](fit_R){width="8.6cm"} An experimental measure of the spin-filtering efficiency is obtained from the $R(T)$ curves, since it can be defined as $$P=\left\|\frac{\sigma_{\uparrow}-\sigma_{\downarrow}}{\sigma_{\uparrow}+\sigma_{\downarrow}}\right\|,$$ where $\sigma_{\uparrow(\downarrow)}$ is the tunnel conductance through the barriers seen by spin up (down) carriers. In the limit of small magnetic exchange fields, the spin-filtering efficiency reduces to $$P\sim\tanh\left(\coth^{-1}\left(\frac{R^{}}{R}\right)\right),$$ where $R$ and $R^{}$ correspond to the measured resistance and the resistance in the absence of magnetic exchange field, respectively [@Senapati2011]. ![Comparison between the normalized subgap branch $J{_{\text{qp}}}=I{_{\text{qp}}}R{_{\text{N}}}/V$ for a non-magnetic tunnel JJ (red straight line) and a perfect spin-filter tunnel JJ (black points). The expression of the quasiparticle current used for the simulations can be found in [@Bergeret2012]. The parameters used to reproduce the curves are: $P=0$, $h=0$, $\Delta=1$, $\eta=0.01\Delta$ (damping factor) and $T=\SI{4.2}{\K}$ for the non-magnetic junction and $P=1$, $h=0.4\Delta$, $\Delta=1$, $\eta=0.01\Delta$ for the spin-filter JJ. The angles $\alpha$ and $\beta$ between $h$ and the magnetization induced in the superconducting electrodes are $\alpha=\beta=\pi/2$.[]{data-label="Fig:8"}](BERGERET2){width="8.6cm"} In Fig. \[Fig:6\] we show the $R(T)$ curve (black points) for the spin-filter JJ with a barrier thickness of $\SI{4.0}{\nm}$. The red straight line is the semiconducting fit performed in the paramagnetic phase of the barrier. This curve allows to obtain the resistance in the absence of magnetic exchange field $R^{*}$. Below $T{_{\text{Curie}}}$ (dashed black line), we can observe a decrease in the resistance because one spin channel is favored in the conduction. The spin-selective tunneling processes affect the normal resistance $R{_{\text{N}}}$ too, which is defined as $R{_{\text{N}}}=1/(4\pi (eN(0))^2(\sigma_{\uparrow}+\sigma_{\downarrow}))$, with $N(0)$ density of state at the Fermi level [@Bergeret2012]. The magnetic exchange field $h$ in the barrier induces a magnetization in the superconducting electrodes of spin-filter JJs [@Bergeret2012]. The angles between the magnetization in the electrodes and the magnetic exchange field will be denoted as $\alpha$ and $\beta$ [@Bergeret2012]. Tunneling of spin-polarized carriers appears only if the angles $\alpha$ and $\beta$ between $h$ and the magnetization in the left and right superconducting electrodes, respectively, are different from $0$ and $\pi$ \[Fig. \[Fig:8\]\] [@Bergeret2012]. The quasiparticle current in non-magnetic devices ($P=0$, $h=0$), and in magnetic JJs with total spin polarization ($P=1$, $h<\Delta$, being $\Delta$ the superconducting gap of the electrodes) and maximum non-collinearity between $h$ and the magnetization in the superconducting electrodes ($\alpha=\beta=\pi/2$) has the same analytic expression, and the $I-V$ curves are comparable \[Fig. \[Fig:8\]\]. We verify this statement using the expression for the quasiparticle current in spin-filter JJs proposed in [@Bergeret2012]. This result justifies the use of a conventional TJM model, in which there is no explicit introduction of a magnetic exchange field in the barrier, when fitting the $I-V$ curves in the ideal case of perfect spin polarization.	{ "pile_set_name": "ArXiv" }
--- abstract: 'The Gottesman-Kitaev-Preskill (GKP) encoding of a qubit within an oscillator provides a number of advantages when used in a fault-tolerant architecture for quantum computing, most notably that Gaussian operations suffice to implement all single- and two-qubit Clifford gates. The main drawback of the encoding is that the logical states themselves are challenging to produce. Here we present a method for generating optical GKP-encoded qubits by coupling an atomic ensemble to a squeezed state of light. Particular outcomes of a subsequent spin measurement of the ensemble herald successful generation of the resource state in the optical mode. We analyze the method in terms of the resources required (total spin and amount of squeezing) and the probability of success. We propose a physical implementation using a Faraday-based quantum non-demolition interaction.' author: - 'Keith R. Motes' - 'Ben Q. Baragiola' - Alexei Gilchrist - 'Nicolas C. Menicucci' bibliography: - 'bibliography.bib' - 'MenicucciPapersRefs.bib' title: Encoding qubits into oscillators with atomic ensembles and squeezed light --- Introduction {#sec:intro} ============ Noise is ubiquitous in physical devices, and quantum devices—including quantum computers [@bib:NielsenChuang00]—are no exception. Fortunately, the threshold theorem states that, under some reasonable assumptions about the type of noise, as long as the level of that noise is below a nonzero minimum value called the fault-tolerance threshold, quantum error correction can be used to reduce them to arbitrarily low levels (see Ref. [@Gottesman:2009ug] for a review). Quantum error correction relies on three key components: (1) redundantly encoded quantum information, (2) a method to detect and correct errors, and (3) a way to prepare, manipulate, and measure the quantum information in its encoded form. In 2001, Gottesman, Kitaev, and Preskill (GKP) proposed a method to encode a qubit into a harmonic oscillator in a way that protects against small displacement errors [@Gottesman2001]. In fact, this protection is fully general since any error acting on the oscillator can be expanded as a superposition of displacements [@Gottesman2001]. The main advantage of this scheme is that all logical Clifford operations—the most important and most common qubit gates [@bib:NielsenChuang00]—are implemented through Gaussian operations at the physical level. This is an advantage for optical implementations because Gaussian operations are much easier to implement than non-Gaussian ones [@Braunstein2005a]. On the other hand, the logical states are ideally an infinite comb of $\delta$ functions (i.e., a Dirac comb), which are impossible to produce because such states are unphysical (requiring infinite energy). As such, approximate states must be used instead. A comb of narrow Gaussian spikes whose heights are distributed according to a wide Gaussian envelope serves as the usual approximate wave function for the logical qubits, carrying with it an associated “embedded error” [@Gottesman2001] because it does not lie fully within the ideal logical subspace. As the spikes get ever narrower and the envelope of heights ever wider, these approximate states tend to their ideal counterparts. Even in their approximate form, GKP states are notoriously difficult to make in the laboratory, despite numerous proposals to do so [@Gottesman2001; @Vasconcelos:2010gb; @bib:Travaglione02; @Pirandola:2004jo; @Pirandola:2006gh; @Pirandola:2006bf; @Terhal2016]. Recent years have seen renewed interest in conquering these difficulties and producing GKP-encoded qubits precisely because of the ability to implement qubit-level Clifford gates as Gaussian operations. Specifically, they dovetail well with continuous-variable (CV) measurement-based quantum computing, in which all Gaussian operations can be implemented simply with homodyne detection on a pre-made CV cluster state [@Menicucci2006]. This architecture is particularly promising because optical CV cluster states can be made on an exceptionally large scale [@Yokoyama:2013jp; @Chen:2014jx; @Yoshikawa2016] with minimal experimental equipment [@Menicucci2011a; @Wang:2014im; @Alexander2016]. The price one pays for this unprecedented scalability is ubiquitous noise due to finite squeezing [@Alexander:2014ew]. Nevertheless, this noise can be handled—and fault-tolerant quantum computing achieved [@Menicucci:2014cx]—by processing GKP-encoded qubits [@Gottesman2001] and applying standard quantum error correction techniques [@Gottesman:2009ug]. In this work, we propose a method to generate optical GKP states by coupling an atomic ensemble, prepared in a spin coherent state, to a squeezed state of light. In Sec. \[sec:GKPanatomy\], we briefly review GKP states and their error-correction properties, and we identify the target states of our proposal. In Secs. \[sec:lightensemble\] and \[sec:stategen\], we describe the light-matter interaction and show that the resulting optical states match the target states. In Sec. \[sec:requirements\], we analyze the method in terms of its parameters and success probability as a function of output-state quality. In Sec. \[sec:PhysicalImplementation\], we discuss the experimental requirements to implement the method using a Faraday interaction with an atomic ensemble. Section \[sec:conclusion\] concludes. Ideal and approximate GKP states {#sec:GKPanatomy} ================================ Ideal GKP states and correcting errors -------------------------------------- The logical subspace of the GKP-encoded qubit is spanned by the logical states $\{{\left\lvert0\right\rangle}_{\text{L,ideal}}, {\left\lvert1\right\rangle}_{\text{L,ideal}}\}$, where $$\begin{aligned} \label{eq:idealGKP} {\left\lvertj\right\rangle}_{\text{L,ideal}}\propto \sum_{s = -\infty}^\infty \qket {(2s+j) g} \propto \sum_{s = -\infty}^\infty (-1)^{js} \pket {s\pi/g},\end{aligned}$$ where $\qket u$ and $\pket u$ are, respectively, position and momentum eigenstates (identified by the subscript on the ket), which satisfy ${\hat{q}} \qket u = u \qket u$ and ${\hat{p}} \pket u = u \pket u$. The wave function for ${\left\lvertj\right\rangle}_{\text{L,ideal}}$ is a Dirac comb in position with spacing $2g$ and offset $jg$ or, equivalently, a superposition of two Dirac combs in momentum, each with spacing $2\pi/g$, offset from each other by $\pi/g$, and with a relative phase of $(-1)^j$. Note that we have omitted the normalization because these ideal logical states are not normalizable but can nevertheless be approximated arbitrarily well by physical states. (We discuss approximate states below.) GKP states are protected against shift errors in position and/or momentum [@Gottesman2001]. The peaks in the position-space wave function of the logical ${\left\lvert0\right\rangle}_{\text{L,ideal}}$ and ${\left\lvert1\right\rangle}_{\text{L,ideal}}$ states are interleaved such that the closest distance between peaks of one state and those of the other is $g$. If any logical GKP state ${\left\lvert\psi\right\rangle} = \alpha {\left\lvert0\right\rangle}_{\text{L,ideal}}+ \beta {\left\lvert1\right\rangle}_{\text{L,ideal}}$ is shifted by a small amount $\delta q$ in position, then this shift error can be corrected by measuring ${\hat{q}} \mod g$ (by coupling to an ancilla prepared in ${\left\lvert0\right\rangle}_{\text{L,ideal}}$) and then shifting the state back into the code space [@Gottesman2001]. When ${\left\lvert{\delta q}\right\rvert} < g/2$, this correction succeeds perfectly. But when $g/2 < {\left\lvert{\delta q}\right\rvert} < g$, the attempt at correction results in a Pauli-$X$ error on the logical state (since it is more likely that the given measurement of ${\hat{q}} \mod g$ would have resulted from a shift by $\delta q - g$ rather than by $\delta q$). Similar properties hold for momentum shifts $\delta p$, which are correctable when ${\left\lvert{\delta p}\right\rvert} < \pi/2g$ and result in a Pauli-$Z$ error for larger shifts [@Gottesman2001]. Since displacements in position and momentum (Weyl operators) form a complete basis for expanding any operator, the ability to correct shift errors is tantamount to protection against any (small) error [@Gottesman2001], including common ones like amplitude damping (photon loss). The syndrome measurement procedure prescribed in Ref. [@Gottesman2001] projects the more general error into definite shifts in position and momentum, which are then (approximately) corrected. This ability to project more general errors into a basis of correctable ones through syndrome measurements is the critical feature of any scheme of active quantum error correction [@Gottesman:2009ug]. Approximate GKP states ---------------------- Both in the original work of GKP [@Gottesman2001] and here, when we talk of a “GKP state” we mean a physical state approximating a state in the GKP logical subspace. Given an ideal GKP state ${\left\lvert\psi\right\rangle}$, we will limit ourselves to considering approximate states of the form $$\begin{aligned} \label{eq:psieta} {\left\lvert\psi\right\rangle}_{\xi}\propto {\hat{{\xi}}} {\left\lvert\psi\right\rangle} = \int du\, dv\, {\xi}(u, v) e^{-iu{\hat{p}} + iv{\hat{q}}} {\left\lvert\psi\right\rangle}, $$ where ${\hat{{\xi}}}$ is a member of the affine oscillator semigroup [@Howe1988] (we will come back to this), and ${\xi}(u,v)$ is the Weyl-operator expansion of ${\hat{{\xi}}}$. We assume that ${\hat{{\xi}}}$ is close to the identity so that ${\left\lvert\psi\right\rangle}_{\xi}$ is a good approximation to ${\left\lvert\psi\right\rangle}$. Without ambiguity, we will refer to both ${\hat{{\xi}}}$ and ${\xi}(u,v)$ as the embedded error in the approximate GKP state. This terminology is consistent with Ref. [@Gottesman2001] and makes sense because we can think of ${\left\lvert\psi\right\rangle}_{\xi}$ as an ideal GKP state ${\left\lvert\psi\right\rangle}$ followed immediately by a trace-decreasing CP map (an error), described by the single Kraus operator ${\hat{\xi}}$. As a GKP state is used for computation, its embedded error will get larger, but this can be partially corrected if one has access to ancillary systems prepared in a (still approximate) logical basis state [@Gottesman2001]. It is easiest to see this by noting that ${\hat{{\xi}}}$, like any operator, can be written as a superposition of shifts in position and momentum, Eq. , which gets projected to just a single shift during the syndrome measurement and then approximately corrected, as discussed above and in Ref. [@Gottesman2001]. This is why it is important to have a supply of high-quality approximations to GKP basis states: a fresh ancilla is needed each time one wants to do GKP error correction. We consider slightly more general embedded errors than GKP themselves did because of the semigroup we have allowed ${\hat{{\xi}}}$ to inhabit. The mathematical details of this semigroup (which can be found in §18 of [@Howe1988]) are rather complicated and beyond the scope of this work, so instead we will merely highlight important features and the way our states go beyond those considered by GKP. In their original work [@Gottesman2001], the most general embedded error considered by GKP is $$\begin{aligned} \label{eq:etaGKP} {\xi}(u,v) \propto \exp \left( -\frac {u^2} {2\Delta^2} - \frac {v^2} {2\kappa^2} \right),\end{aligned}$$ which, for $\Delta \ll 1$ and $\kappa \ll 1$ (and appropriate normalization), corresponds to $$\begin{aligned} \label{eq:etaopGKP} {\hat{{\xi}}} & \approx e^{-\Delta^2 {\hat{p}}^2/2} e^{-\kappa^2 {\hat{q}}^2/2} \approx e^{-\kappa^2 {\hat{q}}^2/2} e^{-\Delta^2 {\hat{p}}^2/2} .\end{aligned}$$ An operator of the form $e^{-\sigma^2 {\hat{q}}^2/2}$ acts on a position-space wave function $\psi(q)$ as multiplication by a Gaussian envelope with zero mean and variance $1/\sigma^2$ \[i.e., $\psi(q) \mapsto e^{-\sigma^2 q^2/2} \psi(q)$\]. The equivalent action of this operator on a momentum-space wave function $\tilde \psi(p)$ is to convolve it with a zero-mean Gaussian with variance $\sigma^2$ \[i.e., $\tilde \psi(p) \mapsto (2 \pi \sigma ^2)^{-1/2} \int d\tau\, {\tilde \psi(p - \tau)} e^{-\tau^2/2\sigma^2}$\], resulting in a blurred version of the original wave function. The operator $e^{-\sigma^2 {\hat{p}}^2/2}$ behaves exactly the same, except with the roles of position and momentum exchanged. When the parameter $\sigma \ll 1$, these two operators approximately commute, and therefore we can regard ${\hat{{\xi}}}$ in Eq. as applying a large Gaussian envelope in position with variance $1/\kappa^2$ and replacing each spike in position with a narrow Gaussian with variance $\Delta^2$, without caring too much about the order in which these happen. This is the type of approximate state originally considered by GKP [@Gottesman2001]. Operators of the form $e^{-\sigma^2 {\hat{q}}^2/2}$ or $e^{-\sigma^2 {\hat{p}}^2/2}$ are members of the (ordinary) oscillator semigroup [@Howe1988]. This semigroup has a nontrivial multiplication law, which is why we have restricted discussion to the simpler special case $\sigma \ll 1$. This restriction nevertheless allows us to represent all possible combinations of large, zero-mean, Gaussian envelopes applied in position and/or momentum. So why is this not enough? In other words, why must we consider any other states beyond just these? There are two reasons. The first is that the main application we have in mind for these states—fault-tolerant CV measurement-based quantum computing [@Menicucci:2014cx]—has ubiquitous displacements that must be corrected [@Menicucci2006], which, combined with intrinsic error at each step due to finite squeezing [@Alexander:2014ew; @Menicucci2006], means that even if we start with this type of state, we will very quickly end up with one outside of this class. The second reason is that the generation procedure we describe below sometimes produces states that are outside of this class (see Sec. \[sec:stategen\]). Specifically, we extend the embedded errors to include those that arise when ${\hat{{\xi}}}$ is a member of the affine oscillator semigroup [@Howe1988]—which includes the ordinary one as a sub-semigroup. This allows us to represent all possible combinations of large, arbitrary-mean, Gaussian envelopes applied in position and/or momentum. We want this because the ideal GKP codespace is periodic in position and in momentum, so there is no fundamental reason to prefer zero-mean envelopes to more general ones, and these more general states arise naturally anyway, as discussed above. Importantly, note that applying these envelopes does not shift the locations of the spikes when applied to an ideal GKP state; they merely enlarge the class of embedded errors that can be discussed. Logical states, target states, resource states, and specifying their quality {#sec:GKPtarget} ---------------------------------------------------------------------------- Having discussed what it means for a state to be an approximate GKP state and how we have expanded this definition beyond that discussed originally by GKP [@Gottesman2001], we now identify the particular parameters of the target approximate GKP states we wish to produce. The most important motivating factor in this decision is the main application we have in mind: fault-tolerant CV measurement-based quantum computing [@Menicucci:2014cx]. In this scenario, the actual errors that accumulate in the GKP states will be roughly symmetric in position and momentum over the course of a computation. Therefore, we want an encoding that is unbiased with respect to position and momentum. This involves two requirements. First, we require $$\begin{aligned} g = \sqrt \pi\end{aligned}$$ so that the spacing between spikes in the grid comprising the logical subspace is equal for both position and momentum—see Eq. , and choose $g = \pi/g$. Second, we want a logical state whose embedded error is roughly symmetric in position and momentum: $$\begin{aligned} \label{eq:emberreq} \sigma \coloneqq \Delta = \kappa .\end{aligned}$$ For the rest of this article we focus on approximate GKP states ${\left\lvertj\right\rangle}_{\text{L}}= {\hat{{\xi}}} {\left\lvertj\right\rangle}_{\text{L,ideal}}$, which have embedded errors of the form $$\begin{aligned} \label{eq:emberrspec} {\hat{{\xi}}} & \approx e^{-\sigma^2 {\hat{p}}^2/2} e^{-\sigma^2 ({\hat{q}}-q_0)^2/2} \nonumber \\ & \approx e^{-\sigma^2 ({\hat{q}}-q_0)^2/2} e^{-\sigma^2 {\hat{p}}^2/2} ,\end{aligned}$$ for small $\sigma$. The momentum-space envelope always has a mean of zero, but that of the position-space envelope may be nonzero (see Sec. \[sec:stategen\]). The possibility of a nonzero mean in the position-space envelope is the key difference between the target states we study here and those analyzed in depth by GKP. This difference is largely inconsequential, however, because the resulting states are all decent approximations to ideal states when the applied envelopes are large. Henceforth, we refer to ${\left\lvertj\right\rangle}_{\text{L}}$ (and superpositions thereof) as logical states, keeping in mind that they include the embedded error as described above. For reasons that will become clear in Sec. \[sec:stategen\], we focus on creating Pauli-$X$-basis logical states ${\left\lvert\pm\right\rangle}_{\text{L}}= {\frac {1} {\sqrt 2} \bigl({\left\lvert0\right\rangle}_{\text{L}}\pm {\left\lvert1\right\rangle}_{\text{L}}\bigr)}$, which we refer to as target states. ![image](psitargetgrid_ALT){width=".9\textwidth"} We will present the target states in terms of their wave functions. We denote the position- and momentum-space wave functions of a state ${\left\lvert\psi\right\rangle}$, respectively, by $$\begin{aligned} \psi(u) \coloneqq \inprodsubsub u \psi q {} \qquad \text{and} \qquad \tilde\psi(u) \coloneqq \inprodsubsub u \psi p {} ,\end{aligned}$$ where the notation for position and momentum eigenstates is defined below Eq. . These are related by the Fourier transform $$\begin{aligned} \label{eq:} \tilde \psi (p) & = \frac {1} {\sqrt {2\pi}} \int dq\, e^{-i q p} \psi (q).\end{aligned}$$ Our target states have the position-space wave function $$\begin{aligned} \label{eq:targetq} \psi_{\pm} (q) & {{\mathpalette{\mathrel{\vcenter{ \offinterlineskip\halign{\hfil$##$\cr \relax\propto\cr\noalign{\kern2pt}\relax\sim\cr\noalign{\kern-2pt}}}}}}e^{-\sigma^2 (q-q_0)^2/2} \sum_{s \in \integers} (\pm 1)^s \exp\left[-\frac {(q-s\sqrt \pi)^2} {2\sigma^2} \right] ,\end{aligned}$$ with ${{\mathpalette{\mathrel{\vcenter{ \offinterlineskip\halign{\hfil$##$\cr \relax\propto\cr\noalign{\kern2pt}\relax\sim\cr\noalign{\kern-2pt}}}}}}$ indicating both approximation (valid for $\sigma \ll 1$) and proportionality up to a constant. We also require ${\left\lvert{q_0}\right\rvert} \le \sqrt \pi/2$ because if ${\left\lvert{q_0}\right\rvert}$ were larger than this, we could reduce it to the indicated range by shifting the state in position by an integer multiple of $\sqrt \pi$. The equivalent momentum-space wave functions are (approximately) $$\begin{aligned} \label{eq:targetp} \tilde \psi_{\pm} (p) &{{\mathpalette{\mathrel{\vcenter{ \offinterlineskip\halign{\hfil$##$\cr \relax\propto\cr\noalign{\kern2pt}\relax\sim\cr\noalign{\kern-2pt}}}}}}e^{-\sigma^2 p^2/2} \sum_{\mathclap{s \in {\genfrac{\lbrace}{\rbrace}{0pt}{}{2\integers\phantom{+1}}{2\integers+1}}}} e^{-i q_0 (p-s\sqrt \pi)} \exp\left[- \frac{(p-s\sqrt \pi)^2}{2 \sigma ^2} \right] , $$ where the upper and lower sets (evens and odds) correspond to $\pm$, respectively. Several examples of these wave functions are shown in Fig. \[fig:SymmResourceState\]. Instead of directly calculating the Fourier transform of Eq. —which, with $\sigma \ll 1$, would ever so slightly shift the spike locations—we have elected to use the second form of ${\hat{{\xi}}}$ from Eq. and apply it directly to the ideal momentum-space representation of ${\left\lvert\pm\right\rangle}_{\text{L,ideal}}$. This results in a simpler form of the approximate wave function and is valid for our purposes because our target state is one with $\sigma \ll 1$. Our goal is to produce high-quality physical approximations to ideal GKP states for use in CV measurement-based quantum computing. To be useful for this purpose, such a state—assumed pure for the present discussion—must have a wave function that closely approximates one of the target states ${\left\lvert\pm\right\rangle}_{\text{L}}$ described above. We henceforth refer to any such state as a resource state. Since the logical states, target states, and resource states under consideration all have symmetric embedded error \[Eq. \], we can measure the quality of any of these states by the single parameter, $\sigma^2$, which represents the variance of an individual spike in the superposition. We can express this parameter in dB, and when we do so, we call this the squeezing of the state. It is a measure of overall quality of the state because it is the only parameter required to specify the embedded error (up to shifts in the envelope). And it has operational value because of its connection to the amount of optical squeezing that would be required to produce just one of the spikes in the superposition (see Sec. \[sec:lightensemble\]). Specifically, $$\begin{aligned} \label{eq:squeezingofstate} \text{(squeezing of the state in dB)} \coloneqq -10 \log_{10} \sigma^2.\end{aligned}$$ Inverting this relation specifies the embedded error \[through Eq. \] and has the following visual interpretation with respect to the state’s wave function, as shown in Fig. \[fig:SymmResourceState\]: $$\begin{aligned} \label{eq:sigmafromsqueezing} \text{(envelope width)}^{-1} = \text{(spike width)} = \sigma = 10^{-\text{(\#dB)}/20},\end{aligned}$$ where (\#dB) is the squeezing of the state in dB. Light-ensemble interaction {#sec:lightensemble} ========================== In this section we describe how resource states can be conditionally prepared by entangling an atomic ensemble and an optical mode followed by projective measurement of the collective atomic spin. Where appropriate, the spin state of the atomic ensemble and the optical state of the field mode are labelled with $A$ and $O$, respectively. This section follows the progress of the joint spin-optical state through the circuit in Fig. \[fig:CircuitModel\], which begins with an unentangled state ${\left\lvert\Psi_a\right\rangle}_{AO}={\left\lvertJ,m_x\!=\!J\right\rangle}_A \otimes {\left\lvert0\right\rangle}_O$ at (a) and ends with a conditional optical state prepared upon a measurement of the spin at (d). A succinct derivation of the protocol for general initial angular-momentum and optical states is given in Appendix \[sec:GKPKraus\]. Initially the atomic ensemble with total collective spin $J$ is prepared in a spin-coherent state polarized along the $x$-axis, which we will denote as ${\left\lvertJ, m_x\!=\!J\right\rangle}_A $. Expressed in terms of eigenstates along the $z$-axis, ${\left\lvertJ,m_z\right\rangle}_A$, the initial spin state is $$\begin{aligned} {\left\lvertJ,m_x\!=\!J\right\rangle}_A &= {\hat{R}}_y(\pi/2) {\left\lvertJ,m_z\!=\!J\right\rangle}_A \nonumber \\ &= \sum_{\mathclap{m=-J}}^J d_{m,J}^{(J)}(\pi/2) {\left\lvertJ,m\right\rangle}_A,\end{aligned}$$ where the coefficients ${d_{m,m'}^{(J)}(\beta) \coloneqq {\left\langleJ,m\right\rvert} {\hat{R}}_y(\beta) {\left\lvertJ,m'\right\rangle}}$, known as Wigner’s (small) $d$-matrix elements [@Varshalovich1988], are matrix elements of the operator ${\hat{R}}_y(\beta)=\exp(-i \beta {\hat{J}}_y)$, which rotates the spin state around the $y$-axis by an angle $\beta$. For our purposes, $J$ is fixed and $\beta=\pi/2$; henceforth, we drop unnecessary notation and define $$\begin{aligned} \label{eq:dmmd} d_{m,m'} &\coloneqq d_{m,m'}^{(J)}\left(\frac \pi 2 \right) = \frac{1}{2^J} \sum_{k} (-1)^{k-m'+m} \nonumber \\ &\quad \times \frac{\sqrt{(J+m')!(J-m')!(J+m)!(J-m)!}}{(J+m'-k)!k!(J-k-m)!(k-m'+m)!},\end{aligned}$$ where the sum is only over integers $k$ such that no factorial in the denominator is of a negative number. We have used the explicit form for $d_{m,m'}^{(J)}(\beta)$ that appears, $e.g.$, in Sakurai [@bib:sakurai2011modern] Sec. 3.8, but this expression can in fact be summed analytically and written in terms of Jacobi polynomials [@Varshalovich1988]. We include these details in Appendix \[app:analyticform\] and use it to derive two special cases used in the main text. ![ Circuit that creates a resource state. The optical mode is prepared in the vacuum state ${\left\lvert0\right\rangle}_O$, and an atomic ensemble, with total collective spin $J$, is prepared in a spin-coherent state polarized along $x$, ${\left\lvertJ, m_x=J\right\rangle}_A$. The optical mode is first squeezed via the standard squeezing operator ${\hat{S}}({r})$ and is then coupled to the collective spin via a controlled position shift ${\hat{D}}_c(g)$. The spin is then projectively measured yielding outcome $x$, conditionally preparing the optical mode in state ${\left\lvert\psi_d^x\right\rangle}_O$. When certain conditions are met (see Sec. \[sec:stategen\]), this state may serve as a resource state, as defined at the end of Sec. \[sec:GKPtarget\].\[fig:CircuitModel\]](PreparationCircuit_ALT){width=".9\columnwidth"} One of these special cases, however, is simple enough to derive directly from the explicit sum, Eq. (\[eq:dmmd\]): When $m'=\pm J$, only a single term in the sum has non-negative factorials in the denominator, and the matrix element simplifies to $$\begin{aligned} \label{eq:GKPmatrixElements} d_{m,\pm J} = \frac{(\pm 1)^{J+m}}{2^J}\begin{pmatrix} 2J \\ J-m\end{pmatrix}^{1/2}.\end{aligned}$$ (See Appendix \[app:analyticform\] for an alternative derivation.) Specifically, the factorials in the denominator are non-negative only when $k=J-m$ for $m'=+J$; and when $k=0$ for $m'=-J$. The values of $m'=\pm J$ are significant as these yield desirable resource states, as we will see in Sec. \[sec:stategen\]. The initial optical state at (a) is the bosonic vacuum ${\left\lvert0\right\rangle}$ with creation and annihilation operators ${\hat{a}}$ and ${\hat{a}}^\dagger$ obeying the usual commutation relations $[{\hat{a}}, {\hat{a}}^\dagger]=1$. We also define the associated position- and momentum-quadrature operators, $$\begin{aligned} {\hat{q}} \coloneqq \frac {1} {\sqrt 2} ({\hat{a}} + {\hat{a}}^\dag) \quad \text{and} \quad {\hat{p}} \coloneqq \frac {-i} {\sqrt 2} ({\hat{a}} - {\hat{a}}^\dag),\end{aligned}$$ respectively, so that $[{\hat{q}}, {\hat{p}}] = i$ with $\hbar = 1$, giving a numerical vacuum variance of $\langle {\hat{q}}^2 \rangle = \langle {\hat{p}}^2 \rangle = \frac 1 2$. To discuss this circuit, we will need the standard definitions [@bib:kok2010] for the squeezing operator $$\begin{aligned} \label{eq:BenBlows} {\hat{S}}({r}) &\coloneqq \exp \left[\frac{1}{2}({r}^{\hat{a}}^2-{r}{\hat{a}}^{\dagger 2})\right], \intertext{displacement operator} \label{eq:ILikeToMoveItMoveIt} \qquad {\hat{D}}(\alpha) &\coloneqq \exp (\alpha {\hat{a}}^\dag - \alpha^ {\hat{a}}), \intertext{and displaced squeezed vacuum state} {\left\lvert\alpha,{r}\right\rangle} &\coloneqq {\hat{D}}(\alpha) {\hat{S}}({r}) {\left\lvert0\right\rangle}.\end{aligned}$$ For brevity, when $\alpha=0$ we will simply write ${\left\lvert{r}\right\rangle} \coloneqq {\left\lvert0,{r}\right\rangle}$ for a squeezed vacuum state. Note that when ${{r}>0}$, ${\left\lvert{r}\right\rangle}$ is squeezed in ${\hat{q}}$ and anti-squeezed in ${\hat{p}}$, and vice versa for ${{r}<0}$. Using this notation, the joint state at (b) in Fig. \[fig:CircuitModel\] is $$\begin{aligned} \label{eq:stateatb} {\left\lvert\Psi_b \right\rangle}_{AO}= \sum_{\mathclap{m=-J}}^J d_{m,J}{\left\lvertJ,m\right\rangle}_A \otimes {\left\lvert{r}\right\rangle}_O.\end{aligned}$$ The atomic ensemble and optical field are entangled with a controlled interaction that applies a displacement to the field proportional to the collective spin along the $z$-direction: $$\begin{aligned} \label{eq:BigD} {\hat{D}}_c(g) &\coloneqq \exp (-i g {\hat{p}}{\hat{J}}_z) = {\hat{D}}({\slashed{g}}{\hat{J}}_z) = {\hat{R}}_z(g{\hat{p}}),\end{aligned}$$ where we have defined $$\begin{aligned} {\slashed{g}}&\coloneqq \frac {g} {\sqrt 2}\end{aligned}$$ for brevity of notation and ease of converting between the Glauber [@Glauber1963] and Weyl-Heisenberg displacement operators. As shown in Eq. , the interaction ${\hat{D}}_c(g)$ can be thought of in two equivalent ways: (a) ${{\hat{D}}_c(g) = {\hat{D}}({\slashed{g}}{\hat{J}}_z)}$, which is a position displacement of the optical mode controlled on ${\hat{J}}_z$ of the ensemble, and (b) ${{\hat{D}}_c(g) = {\hat{R}}_z(g{\hat{p}})}$, which is a $z$-rotation of the collective spin controlled on ${\hat{p}}$ of the mode. The real parameter $g$ represents the strength of the coupling in both of these interpretations. (We use the symbol $g$ for this coupling strength because it will directly determine the spike spacing $g$ of the resource state, which is discussed in Sec. \[sec:GKPanatomy\].) ${\hat{D}}_c(g)$ has the following effect on the components of the superposition in Eq. : $$\begin{aligned} \label{eq:displacedsqueezed} {\hat{D}}_c(g) {\left\lvertJ,m\right\rangle}_A\otimes {\left\lvert{r}\right\rangle}_O = {\left\lvertJ,m\right\rangle}_A \otimes {\left\lvert{\slashed{g}}m,{r}\right\rangle}_O. \end{aligned}$$ After the controlled displacement given by Eq. (\[eq:BigD\]), the joint state at (c) in Fig. \[fig:CircuitModel\] is $$\begin{aligned} \label{eq:stateatc} {\left\lvert\Psi_c\right\rangle}_{AO} = \sum_{\mathclap{m=-J}}^J d_{m,J}{\left\lvertJ,m\right\rangle}_A \otimes {\left\lvert {\slashed{g}}m, {r}\right\rangle}_O.\end{aligned}$$ Finally, the collective spin is projectively measured along the $x$-direction as shown in (d) in Fig. \[fig:CircuitModel\]. We express ${\left\lvert\Psi_c\right\rangle}_{AO}$ in the basis of $x$-eigenstates: $$\begin{aligned} \label{eq:psic} {\left\lvert\Psi_c\right\rangle}_{AO} &= \sum_{\mathclap{m,m'=-J}}^J d_{m,J}d_{m,m'}{\left\lvertJ, m_x=m'\right\rangle}_A \otimes {\left\lvert {\slashed{g}}m, {r}\right\rangle}_O,\end{aligned}$$ where we used the fact that the matrix elements in Eq. (\[eq:dmmd\]) satisfy ${\left\langleJ,m'\right\rvert}{\hat{R}}_y (-\pi/2){\left\lvertJ,m\right\rangle} = d_{m,m'}$. Then, measurement trivially collapses the $m'$ summation in Eq. (\[eq:psic\]) to a single term. That is, given measurement outcome $x$, the optical state is projected to $$\begin{aligned} \label{eq:projectedOpticalState} {\left\lvert\psi^x_{d}\right\rangle}_O = \frac{1}{\sqrt{\mathcal{P}(x)}}\sum_{m=-J}^J d_{m,J}d_{m,x}{\left\lvert{\slashed{g}}m, {r}\right\rangle}_O.\end{aligned}$$ The state is normalized by the probability of obtaining $x$, $$\begin{aligned} \label{eq:GKPmeasProb} \mathcal{P}(x) = \sum_{\mathclap{m,m'}}d_{m,J}d_{m,x}d_{m',J}d_{m',x}e^{-\frac{1}{4}g^2e^{2{r}}(m-m')^2},\end{aligned}$$ as shown in Appendix \[sec:GKPKraus\], noting that the squeezing parameter ${r}$ is taken to be real. The conditional optical state can be written as a wave function in the position and momentum quadratures, respectively, as \[eq:psicondJ\] $$\begin{aligned} \label{eq:psicondJpos} \psi_d^x(q) = & \frac{e^{{r}/2}}{\mathcal{P}(x)^{1/2}\pi^{1/4}} \sum_{\mathclap{m=-J}}^J d_{m,J}d_{m,x} \exp\left[-\frac{(q-gm)^2}{2 e^{-2{r}}}\right], \\ \label{eq:psicondJmom} \tilde \psi_d^x(p) = & \frac{e^{-{r}/2}}{\mathcal{P}(x)^{1/2}\pi^{1/4}} \sum_{\mathclap{m=-J}}^J d_{m,J}d_{m,x}\exp\left[-igmp-\frac{p^2}{2e^{2{r}}}\right].\end{aligned}$$ For total spin ${J = 4}$ and ${J=\frac 9 2}$, the position- and momentum-space wave functions are plotted in Fig. \[fig:resourceState\]. Each measurement outcome $x$ prepares a conditional position-space wave function, Eq. (\[eq:psicondJpos\]), which is a superposition of displaced squeezed states separated by $g$ with amplitudes governed by the product of matrix elements ${d_{m,J}d_{m,x}}$. The momentum-space wave function, Eq. (\[eq:psicondJmom\]), is less straightforward, and we defer interpreting it until the next section wherein we restrict to ${x = \pm J}$. Resource state generation {#sec:stategen} ========================= The conditional optical states in Eq. (\[eq:projectedOpticalState\]) are useful as resource states (as defined in Sec. \[sec:GKPtarget\]) when three conditions are met: 1. The state’s position-space wave function consists of a periodic comb of spikes with either uniform or alternating phase (corresponding to ${\left\lvert\pm\right\rangle}_{\text{L}}$, respectively). 2. The spikes are separated in position by $\sqrt \pi$. 3. The embedded error of the state is symmetric—i.e., approximately equal in position and momentum. The first condition is met for spin-measurement outcomes $x=\pm J$. Tuning the coupling strength to $g=\sqrt{\pi}$ meets the second condition. The third condition is met by enforcing a constraint relating the optical squeezing ${r}$ and total spin $J$. Below, we formalize these requirements and reveal the correspondence of the resource with the target GKP states ${\left\lvert\pm\right\rangle}_{\text{L}}$. The wave functions for the resource states appear in the top and bottom rows of both sides of Fig. \[fig:resourceState\]. ![image](psiJxintgrid_ALT){width="\columnwidth"}![image](psiJxhalfintgrid_ALT){width="\columnwidth"} We begin with the first condition. Before enforcing any restrictions on $g$, the wave functions for the conditional optical states ${\left\lvert\psi_d^{\pm J}\right\rangle}_O$ are \[eq:psiJ\] $$\begin{aligned} \label{eq:psiqJ} \psi_d^{\pm J}(q) & = \frac{e^{{r}/2}}{\mathcal{P}(\pm J)^{1/2}\pi^{1/4}} 2^{-2J} \sum_{m=-J}^J \binom{2J}{J-m} (\pm 1)^{J+m} \nonumber \\ & \qquad\times \exp\left[-\frac{(q-gm)^2}{2 e^{-2{r}}}\right], \\ \label{eq:psipJ} \tilde \psi_d^{\pm J}(p) & = \frac{e^{-{r}/2}}{\mathcal{P}(\pm J)^{1/2}\pi^{1/4}} \nonumber \\ & \qquad\times \exp\left(-\frac{p^2}{2e^{2{r}}}\right) {\genfrac{\lbrace}{\rbrace}{0pt}{}{\cos^{2J}(g p/2)}{e^{iJ\pi}\sin^{2J}(g p/2)}} ,\end{aligned}$$ where we have used the $d_{m,\pm J}$ matrix elements in Eq. (\[eq:GKPmatrixElements\]), and the top and bottom choices within the braces correspond to $\pm$, respectively. The position-space wave function $\psi_d^{\pm J}(q)$ is a superposition of $2J+1$ Gaussians, each with variance $$\begin{aligned} \label{eq:GKPvarQ} \sigma_{q}^2=e^{-2{r}},\end{aligned}$$ separated by $g$, and distributed according to a binomial envelope arising from the binomial coefficients in Eq. (\[eq:psiqJ\]). The binomial coefficients may be approximated by a Gaussian using (see Appendix \[app:analyticform\]) $$\begin{aligned} \label{eq:binomGauss} 2^{-2J} \binom{2J}{J-m} &\approx \frac{1} {\sqrt{\pi J}} \exp \left(-\frac{m^2}{J}\right) ,\end{aligned}$$ which has variance $J/2$. Since the spikes are located at $q=gm$ (instead of $q=m$), the binomial coefficients provide an approximately Gaussian envelope with variance $$\begin{aligned} \label{eq:GKPvarEnvQ} \sigma_{q,\text{env}}^2= \frac {g^2J} {2}.\end{aligned}$$ Note, however, that the measured variance of each spike is $\frac 1 2 \sigma_q^2$, and that of the envelope is $\frac 1 2 \sigma_{q,\text{env}}^2$, due to the need to square the wave function to get measured probabilities. Similarly, the momentum-space wave function $\tilde \psi_d^{\pm J}(p)$ can be described as a product of a Gaussian envelope with variance $$\begin{aligned} \label{eq:GKPvarEnvP} \sigma_{p,\text{env}}^2=e^{2{r}}\end{aligned}$$ and a comb of approximately Gaussian peaks generated by $\cos^{2J}(g p/2)$ or $\sin^{2J}(g p/2)$ and hence separated by $2\pi/g$. The variance of the individual peaks is given by $$\begin{aligned} \label{eq:GKPvarP} \sigma_{p}^2 &= \frac{2 \bigl[J^2\zeta(2,J)-1 \bigr]}{g^2J^2} \approx \sigma_{q,\text{env}}^{-2} +O \left( \frac {1} {J^2} \right), \end{aligned}$$ where $\zeta(s,a)$ is the Hurwitz zeta function, as shown in Appendix \[app:cosvariance\]. As above, the measured variances of the spikes and envelope in momentum are $\frac 1 2 \sigma_p^2$ and $\frac 1 2 \sigma_{p,\text{env}}^2$, respectively. Under these approximations the wave functions are \[eq:psiJapprox\] $$\begin{aligned} \label{eq:psiqJapprox} \psi_d^{\pm J}(q) &{{\mathpalette{\mathrel{\vcenter{ \offinterlineskip\halign{\hfil$##$\cr \relax\propto\cr\noalign{\kern2pt}\relax\sim\cr\noalign{\kern-2pt}}}}}}e^{-q^2/2\sigma_{q,\text{env}}^2} \sum_{\mathclap{m=-J}}^J (\pm 1)^{J+m} \exp\left[-\frac{(q-gm)^2}{2 \sigma_q^2}\right], \\ \label{eq:psipJapprox} \tilde \psi_d^{\pm J}(p) &{{\mathpalette{\mathrel{\vcenter{ \offinterlineskip\halign{\hfil$##$\cr \relax\propto\cr\noalign{\kern2pt}\relax\sim\cr\noalign{\kern-2pt}}}}}}e^{-p^2/2\sigma_{p,\text{env}}^2} \sum_{\mathclap{s \in {\genfrac{\lbrace}{\rbrace}{0pt}{}{2\integers\phantom{+1}}{2\integers+1}}}} e^{i s J \pi} \exp \left[-\frac{(p-s \pi / g)^2}{2 \sigma_p^2} \right] ,\end{aligned}$$ The position representation in Eq. results from modifying the approximation in Eq. by the replacement $m \mapsto q/g$ (since this is where the original function has nontrivial support). The momentum representation, Eq. , results from approximating $\cos^{2J} (gp/2)$ and $\sin^{2J} (gp/2)$ by an infinite series of appropriately sized Gaussian spikes centered on $p = s \pi/g$ (where $s$ is even for $+$ and odd for $-$), with each multiplied by the original function value at the center of the spike. For $g = \sqrt{\pi}$, the conditional optical states ($x=\pm J$) meet the second condition—i.e., they have equally spaced spikes in position and momentum and thus resemble the target states, Eqs. and . However, there is an important difference between the conditional optical states when they arise from spin measurements of integer $J$ versus half-integer $J$. Refer to the left side of Fig. \[fig:resourceState\]. For any integer $J$ (case $J=4$ shown), the outcome ${x=+J}$ produces a resource state close to the target state ${\left\lvert+\right\rangle}_{\text{L}}= \frac {1} {\sqrt{2}}({\left\lvert0\right\rangle}_{\text{L}}+{\left\lvert1\right\rangle}_{\text{L}})$ because it is an equally spaced comb with uniform phase on every spike. Similarly, the outcome ${x=-J}$ produces a resource state close to the target state ${\left\lvert-\right\rangle}_{\text{L}}= \frac {1} {\sqrt{2}}({\left\lvert0\right\rangle}_{\text{L}}-{\left\lvert1\right\rangle}_{\text{L}})$ since the spikes have alternating sign within the superposition. Now refer to the right side of that figure. For any half-integer $J$ (case $J=\frac 9 2$ shown), the outcomes ${x=\pm J}$, respectively, produce a resource state close to a slightly shifted version of the target states ${\left\lvert\pm\right\rangle}_{\text{L}}$. This is because the logical subspace, by definition [@Gottesman2001], has a spike centered at 0. This corresponds to the $m=0$ branch of the superposition in Eq. , which does not exist for half-integer $J$. These states, therefore, must be displaced into the codespace by applying ${\hat{D}}({\slashed{g}}/2) = \exp(-i g{\hat{p}} / 2)$ before they can be considered (approximate) logical GKP states. Finally, the third condition is met when the embedded error is balanced in both quadratures ($\sigma_q^2 = \sigma_p^2$). Combining this requirement with the necessary coupling $g=\sqrt{\pi}$ and using Eqs. , , and gives the following relationship between squeezing ${r}$ and total spin $J$: $$\begin{aligned} \label{eq:xisymmetric} J = \frac{2}{\pi} e^{2 {r}} \qquad \Longleftrightarrow \qquad {r}= \frac 1 2 \log \left( \frac {\pi J}{2} \right).\end{aligned}$$ This relationship symmetrizes the spike and envelope variances: $$\begin{aligned} \label{eq:symmvars} \sigma_{(q,p)}^2 = e^{-2{r}} = \frac {2} {\pi J} = \sigma^{-2}_{(q,p),\text{env}}.\end{aligned}$$ Indeed, it is straightforward to verify that plugging these parameters into Eqs. and displacing the state if required (for half-integer $J$) produces wave functions of the form of Eqs. and , respectively, with $q_0 = \sqrt \pi/ 2$ for half-integer $J$. Applying Eq. , we can see that the squeezing of the state is equal to the optical squeezing (measured in dB). Specifically, \[eq:ressqueezing\] $$\begin{aligned} & \text{(squeezing of the state in dB)} \nonumber \\ & \qquad = -10 \log_{10}e^{-2 {r}} \approx 8.686\,{r}\\ & \qquad = -10 \log_{10} \frac {2} {\pi J} \approx 1.9612 + 10 \log _{10} J .\end{aligned}$$ We can invert the latter equation to obtain the required total spin $J$ as a function of the resource state’s squeezing (in dB): $$\begin{aligned} \label{eq:Jfromsqueezing} J=\frac{2}{\pi} 10^{\text{(\#dB)}/10}.\end{aligned}$$ The maroon curve (a) in Figure \[fig:squeezing\] plots this relation for various levels of squeezing in the resource state. Total angular momentum ${J\approx 63.5}$, which corresponds to 20 dB of squeezing, is sufficient for fault-tolerant measurement-based quantum computing with continuous-variable cluster states [@Menicucci:2014cx], although more modest levels (e.g., ${J\approx 20}$; 15 dB of squeezing) may still be useful in certain circumstances [@Menicucci:2014cx]. In summary, the resulting resource states produced by this method are $$\begin{aligned} \label{eq:shiftedlogical} {\left\lvert { \psi_{\pm}^{\rm res} }\right\rangle} \coloneqq \begin{cases} {\hat{D}}\left( {\slashed{g}}/2 \right) {\left\lvert\psi_d^{\pm J}\right\rangle} & \text{for half-integer~$J$,} \\ \phantom{{\hat{D}}\left( {\slashed{g}}/2 \right)} {\left\lvert\psi_d^{\pm J}\right\rangle} & \text{for integer~$J$.} \end{cases} $$ As these have a central spike at $q=0$ in their position-space wave function, equal spacing of spikes in position and momentum, and equal embedded error, they approximate well the target states ${\left\lvert\pm\right\rangle}_{\text{L}}$, whose wave functions are shown in Fig. \[fig:SymmResourceState\] (with $q_0 = \sqrt \pi/2$ for half-integer $J$). Further, when the optical squeezing ${r}$ and total spin $J$ are both increased while maintaining the functional relationship between them in Eq. , the resource states become better approximations to the target states, which themselves become better approximations to ideal GKP states. Outcome x=0 for integer J {#sec:zerooutcome} ------------------------- When $J$ is an integer, there is another way to produce an approximate GKP state: the conditional state when $x=0$. We choose not to focus too much on this option because (a) the embedded error of the resulting state is different from that of the $x=\pm J$ outcomes, and (b) it is not available for half-integer $J$. Still, we detail this case here for completeness. The fact that $x=0$ makes a useful state can be seen directly in Fig. \[fig:resourceState\] by noting that this outcome produces a momentum-space wave function $\tilde \psi_d^0(p)$ that seems to resemble $\psi_d^J(q)$, the position-space wave function for $x=J$. This would mean that it approximates the logical state ${\left\lvert0\right\rangle}_{\text{L}}$ (since $p$ and $q$ are related by a Fourier transform, which represents the logical Hadamard gate [@Gottesman2001]). Let us formalize this resemblance. ![(a) Required total spin $J$ \[Eq. \] and (b) success probability $\mathcal P_s$ \[Eq. \] for generating a resource state—i.e., an approximate GKP state satisfying the three conditions identified at the beginning of Sec. \[sec:stategen\]—with a given amount of squeezing. The resource state’s squeezing level is defined in Eq. and evaluated in Eqs. . []{data-label="fig:squeezing"}](JfromdB_ALT){width="1.0\columnwidth"} Using the results from Appendix \[app:analyticform\], we can immediately write the output wave functions for $x=0$: \[eq:x0psiJ\] $$\begin{aligned} \label{eq:x0psiqJ} \psi_d^0(q) & = \frac{e^{{r}/2}}{\mathcal{P}(0)^{1/2}\pi^{1/4}} 2^{-2J} {\binom{2 J}{J}}^{1/2} \nonumber \\ &\qquad \times \sum_{\mathclap{m=-J}}^J {\operatorname{Re}}(i^{J+m}) \binom{J}{\frac{J-m}{2}} \exp\left[-\frac{(q-gm)^2}{2 e^{-2{r}}}\right] , \\ \label{eq:x0psipJ} \tilde \psi_d^0(p) & = \frac{e^{-{r}/2}}{\mathcal{P}(0)^{1/2}\pi^{1/4}} i^J 2^{-J} {\binom{2 J}{J}}^{1/2} \nonumber \\ &\qquad \times \exp\left( -\frac{p^2}{2e^{2{r}}} \right) \sin^J (g p) .\end{aligned}$$ Analogously to how we approximated Eqs. as Eqs. , we can approximate these as \[eq:x0psiJapprox\] $$\begin{aligned} \label{eq:x0psiqJapprox} \psi_d^0(q) &{{\mathpalette{\mathrel{\vcenter{ \offinterlineskip\halign{\hfil$##$\cr \relax\propto\cr\noalign{\kern2pt}\relax\sim\cr\noalign{\kern-2pt}}}}}}e^{-q^2/4\sigma_{q,\text{env}}^2} \sum_{\mathclap{m=-J}}^J {\operatorname{Re}}(i^{J+m}) \exp\left[-\frac{(q-gm)^2}{2 \sigma_q^2}\right], \\ \label{eq:x0psipJapprox} \tilde \psi_d^0(p) &{{\mathpalette{\mathrel{\vcenter{ \offinterlineskip\halign{\hfil$##$\cr \relax\propto\cr\noalign{\kern2pt}\relax\sim\cr\noalign{\kern-2pt}}}}}}e^{-p^2/2\sigma_{p,\text{env}}^2} \nonumber \\ &\quad \times \sum_{\mathclap{n \in \integers}} (-1)^{J n} \exp \left\{-\frac{\bigl[p-(n+\frac 1 2) \pi / g \bigr]^2}{\sigma_p^2} \right\} ,\end{aligned}$$ where the definitions of each of the $\sigma$’s is unmodified from above. It is evident that applying a momentum shift ${\hat{D}}(i \pi / 2 g \sqrt 2) = \exp(i \pi {\hat{q}} / 2 g)$ will make $\psi_d^0 {{\mathpalette{\mathrel{\vcenter{ \offinterlineskip\halign{\hfil$##$\cr \relax\propto\cr\noalign{\kern2pt}\relax\sim\cr\noalign{\kern-2pt}}}}}}\tilde \psi_+$ and $\tilde \psi_d^0 {{\mathpalette{\mathrel{\vcenter{ \offinterlineskip\halign{\hfil$##$\cr \relax\propto\cr\noalign{\kern2pt}\relax\sim\cr\noalign{\kern-2pt}}}}}}\psi_+$, where the latter are defined in Eqs. and , respectively. As discussed above, this means that the $x=0$ output state (after the shift) approximates the logical state ${\left\lvert0\right\rangle}_{\text{L}}$. The most important difference between the wave functions in Eqs. and those in Eqs. is the embedded error. Specifically, the momentum-space spikes in Eq. are half the variance of the position-space spikes in Eq. . Equivalently, the position-space envelope in Eq. is twice the variance of the momentum-space envelope in Eq. . Allowing for the outcome $x=0$ in certain cases may prove beneficial for some applications, but for the reasons mentioned at the top of this subsection, we will not consider it further. Instead, we will focus on the results that are valid for all $J$—i.e., the states resulting from the outcomes $x=\pm J$. Success probability {#sec:requirements} =================== Recall that the procedure for generating resource states described in Sec. \[sec:lightensemble\] is probabilistic, with the probability of spin-measurement outcome $x$ given by Eq. , repeated here for reference: $$\begin{aligned} \label{eq:GKPmeasProb2} \mathcal{P}(x) = \sum_{\mathclap{m,m'}}d_{m,J}d_{m,x}d_{m',J}d_{m',x}e^{-\frac{1}{4}g^2e^{2{r}}(m-m')^2}.\end{aligned}$$ Valid outcomes for producing useful resource states are $x=\pm J$, which becomes \[using Eq. \] $$\begin{aligned} \mathcal{P}(\pm J) & = 2^{-4J} \sum_{\mathclap{m,m'}} (\pm 1)^{m+m'} \binom{2 J}{J-m} \binom{2 J}{J-m'} \nonumber \\ &\qquad \times e^{-\frac{1}{4}g^2e^{2{r}}(m-m')^2} .\end{aligned}$$ Furthermore, for a symmetrically encoded state with balanced embedded error (see above), $g^2 e^{2{r}} = \pi^2 J/2$, which means that, unless $J=0$, we really only need to consider the terms where ${\left\lvert{m-m'}\right\rvert} \leq 1$ (the others being made negligible by the Gaussian). Therefore, for reasonably large $J$, we may approximate $$\begin{aligned} e^{-\frac{1}{8}\pi^2 J(m-m')^2} &\approx \delta_{m,m'} + e^{-\frac{1}{8}\pi^2 J} (\delta_{m+1,m'} + \delta_{m,m'+1})\end{aligned}$$ in the double sum (where $\delta$ is the Kronecker delta), which leads to $$\begin{aligned} \mathcal{P}(\pm J) \approx \mathcal P_0(J) \pm 2 e^{-\frac{1}{8}\pi^2 J} \mathcal P_1(J),\end{aligned}$$ where $$\begin{aligned} \mathcal P_0(J) &\coloneqq 2^{-4J} \sum_{k=0}^{2J} \begin{pmatrix}2J \\ k\end{pmatrix}^2 = 2^{-4J} \begin{pmatrix}4J \\ 2J\end{pmatrix}, \\ \mathcal P_1(J) &\coloneqq 2^{-4J} \sum_{k=0}^{2J-1} \begin{pmatrix}2J \\ k\end{pmatrix} \begin{pmatrix}2J \\ k+1\end{pmatrix} = 2^{-4J} \begin{pmatrix}4J \\ 2J + 1\end{pmatrix}.\end{aligned}$$ To obtain this expression, we shifted the (dummy) summation indices by $J$ and used the Chu-Vandermonde identity to perform the sums. Since either outcome $x = \pm J$ yields a useful state, the total success probability $\mathcal P_s(J)$ is the sum of both cases. This means $\mathcal P_1(J)$ is irrelevant, giving (for $J \ge \tfrac 1 2$) $$\begin{aligned} \label{eq:PsJapprox} \mathcal P_s(J) &\approx 2 \times 2^{-4J} \begin{pmatrix}4J \\ 2J\end{pmatrix} . $$ Furthermore, we verified numerically that the following asymptotic approximation to Eq. (for large $J$) is correct to within 1.3% of the exact result \[Eq. \] for all values of $J \ge \tfrac 1 2$: $$\begin{aligned} \label{eq:PsuccessLargeJ} \mathcal P_s(J) &\approx \sqrt {\frac {2} {\pi J}} \left( 1 - \frac {1} {16 J} \right).\end{aligned}$$ Using Eq. , we can write this in terms of the squeezing of the state (measured in dB): $$\begin{aligned} \label{eq:PsuccessfromdB} \mathcal P_s & \approx 10^{-\text{(\#dB)}/20}-\frac{\pi}{32} 10^{-3 \text{(\#dB)}/20}.\end{aligned}$$ This is shown as the blue curve (b) in Fig. \[fig:squeezing\]. ![The probability $\mathcal{P}(x)$ \[Eq. \] of a particular spin-measurement outcome $x$. The plot uses $J=50$ and a symmetric encoding with balanced embedded error (see the beginning of Sec. \[sec:requirements\]). Of the $2J+1$ possible outcomes, two of them—namely, $x=\pm J$, which also just happen to be the most probable—yield useful resource states. (In the case of integer $J$, the $x=0$ case also produces a useful state, albeit with slightly different properties; see Sec. \[sec:zerooutcome\].) []{data-label="fig:hightails"}](HighTails_ALT){width="\columnwidth"} Critical here is that the success probability scales as $\mathcal P_s \sim J^{-1/2}$ for large $J$ (i.e., for high-quality resource states), whereas in a similar scheme proposed in Ref. [@bib:Travaglione02] using a single spin-$\frac{1}{2}$ atom with repeated interactions and measurements, the success probability is approximately $$\begin{aligned} \label{eq:iteratedsucc} \mathcal P_{s,\text{iter}}(J) \sim 2^{-2J+1}.\end{aligned}$$ This is exponential in $J = {{N}}/2$, where ${{N}}$ is the number of iterations of interaction with the atom. We obtained this formula from noting that the first measurement of this iterative method will succeed on either outcome, but on subsequent iterations, only one of the two outcomes will lead to a useful state [@bib:Travaglione02], and each outcome is approximately unbiased. (Contrary to what is claimed in that reference, these other outcomes are not even useful in the error-correction gadget [@Gottesman2001] because they do not occupy the logical subspace.) The advantage of our protocol can be explained by noting that the measurement is collective, with outcomes arising solely within the spin-$J$ subspace of ${{{N}}= 2J}$ particles. Thus, the total number of measurement outcomes is $2J+1$, and on top of that, the desired outcomes ($\pm J$) are the most likely ones—see Fig. \[fig:hightails\]. In contrast, consider the procedure proposed in Ref. [@bib:Travaglione02], whereby the individual spin-$\frac{1}{2}$ atom is measured iteratively a total of ${{N}}=2J$ times. The measurement samples the full $2^{2J}$-dimensional Hilbert space, with the overwhelming majority of outcomes failing to produce the desired resource state. Because of the exponential scaling, such an iterative method has essentially no hope of producing a resource state with squeezing much higher than about 10 dB (for which $\mathcal P_{s,\text{iter}} \approx 3\times 10^{-4}$), whereas 10 dB of squeezing in our proposal ($J \approx 6.5$) has a much more reasonable success probability $\mathcal P_s \approx 31\%$. Physical Implementation {#sec:PhysicalImplementation} ======================= The required coupling between an optical mode and a spin system, Eq. , arises in the linearized limit of the dispersive Faraday (or Kerr-type) interaction [@bib:Hammerer2010]. The polarization of light is transformed due to circular birefringence in an ensemble of polarizable quantum scatterers [@bib:deutsch10]. For strong enough coupling, the interaction produces the symmetric encoding in Sec. \[sec:requirements\] and provides the foundation for a quantum non-demolition measurement of the spin required to produce approximate GKP states in the optical mode. Faraday-based QND interaction {#sec:GKPinteraction} ----------------------------- We consider an optical field coupled to the collective spin formed by an ensemble of polarizable neutral atoms. Consider ${{N}}$ such atoms, each with effective spin ${\hat{j}}_z = \frac{1}{2} \big({\left\lvert{\uparrow}\middle\rangle\middle\langle{\uparrow}\right\rvert} - {\left\lvert{\downarrow}\middle\rangle\middle\langle{\downarrow}\right\rvert} \big)$ defined on metastable ground states $\{\uparrow, \downarrow \}$. The atoms couple to a common mode of light possessing two orthogonal linear polarizations, horizontal ($H$) and vertical ($V$), with respective annihilation operators ${\hat{a}}_H$ and ${\hat{a}}_V$. For an off-resonant field, the atoms and light become entangled via the dispersive Faraday interaction, ${\hat{U}} = e^{-i\chi {\hat{S}}_3 {\hat{J}}_z}$, which describes a coupling of the collective atomic spin, ${\hat{J}}_z = \sum_{n=1}^{{N}}{\hat{j}}_z^{(n)}$, to the 3-component of the field’s Stokes vector [@bib:deutsch10], $$\begin{aligned} \label{Eq::Stokes} {\hat{S}}_3 &= \frac{1}{2i} \big( {\hat{a}}^\dagger_H {\hat{a}}_V - {\hat{a}}^\dagger_V {\hat{a}}_H \big). \end{aligned}$$ The Faraday interaction generates a rotation of the Stokes vector around the 3-axis proportional to the atomic spin projection along ${\hat{J}}_z$ with interaction strength characterized by the dimensionless, single-photon rotation angle $\chi$. The controlled displacement required for the resource-state generation, Eq. (\[eq:BigD\]), arises by preparing the $H$-mode of light in a coherent state with ${{N_L}}$ photons. Making the linearization ${\hat{a}}_H \rightarrow \sqrt{{{N_L}}}$, the Stokes operator in Eq. becomes ${\hat{S}}_3 \approx \sqrt{ {{N_L}}/2 }\, {\hat{p}},$ where ${\hat{p}} = -i({\hat{a}}_V - {\hat{a}}_V^\dagger )/\sqrt{2}$ is the momentum quadrature of the $V$-polarized mode [@bib:Hammerer2010]. This linearized Faraday interaction generates the requisite controlled translations of $V$-mode photons, Eq. (\[eq:BigD\]), with effective coupling strength $$\begin{aligned} \label{Eq::Effectiveg} g = \chi \sqrt{ \frac {{{N_L}}} {2} }. \end{aligned}$$ To begin the protocol, the state ${\left\lvert\psi_b\right\rangle}_{AO}$ in Eq. (\[eq:stateatb\]) is prepared. Optical pumping intializes the atomic ensemble in a spin-coherent state along $x$, where $z$ is defined by the propagation direction of the light, and squeezed vacuum is separately prepared by pumping an optically nonlinear medium [@bib:Walls2008]. Then, the Faraday interaction entangles the spins and light, producing the state ${\left\lvert\psi_c\right\rangle}_{AO}$ in Eq. (\[eq:stateatc\]) when $g = \sqrt{\pi}$. This is achieved for interaction time $t = 2 \pi/\chi^2 \dot{N}_L$ for an $H$-polarized classical pump with photon flux $\dot{N}_L$. Projective spin measurement {#sec:GKPmeasurement} --------------------------- Once the optical field and atomic ensemble have become entangled, the collective spin state is projectively measured. A single atomic spin may be measured by driving a cycling transition and detecting the resulting fluorescence [@bib:Wineland11]. Concatenating with unitary transformations to cycle through the measurement basis, a projective measurement can be realized. With access to universal two-qubit control, this can be extended to many spins [@bib:Lloyd2001]. However, since our resource-state protocol benefits from large collective spins composed of many constituent atoms, such a procedure is prohibitively time- and resource-consuming. We focus instead on a quantum nondemolition (QND) measurement of the collective spin. The collective spin is coupled via the same Faraday interaction to a second optical field that serves as a meter. The meter experiences a spin-dependent polarization rotation that is measured via homodyne polarimetry [@bib:deutsch10]. When the spin-meter coupling is strong, relative rotations from different projective $m$-values are distinguishable over the inherent shot noise in the meter, and the collective spin measurement becomes projective. This is indeed the same strong-coupling requirement for the GKP-state peaks to be sufficiently separated. To implement the measurement, the collective spin is first rotated into the $x$-basis with a $\pi/2$-pulse. The meter, propagating again along $z$, is initialized with ${{N_L}}$ photons in the $H$-mode. The quantum-mechanical $V$-mode is prepared in position-squeezed vacuum ${\left\lvert{r}\right\rangle}_M$ (M stands for ‘meter’), which has shot-noise variance $\sigma^2_M = e^{-2{r}}$ for $q$-measurements. Polarimetry in the diagonal polarization basis implements a homodyne measurement of the position quadrature for $V$-mode photons, with the $H$-mode serving as the local oscillator [@bib:baragiola14]. The degree to which the measurement is projective is determined by the distinguishability of the meter states, ${\left\lvert{\slashed{g}}m, {r}\right\rangle}_M = e^{ -i g m {\hat{p}} } {\left\lvert{r}\right\rangle}_M$, given by Eq. (\[eq:Overlap\]), $$\begin{aligned} \label{Eq::Resolution} \big\| {\langle{ {\slashed{g}}m, {r}}\|{{\slashed{g}}m', {r}}\rangle} \big\|^2 & = \exp \left[ - \left( \frac {g} {2 \sigma_M} \right)^2(m - m')^2 \right]. \end{aligned}$$ In the limit that $(m - m')^2 \gg 4 \sigma_M^2/g^2 $, the meter states become orthogonal. Thus, distinguishing neighboring eigenstates of ${\hat{J}}_z$ requires $$\begin{aligned} \frac {4 \sigma_M^2} {g^2} \ll 1,\end{aligned}$$ with limits set by the characteristic coupling strength $g$ and the squeezed shot noise in the polarimeter, $\sigma_M^2$. Note that in contrast to the procedure for producing the GKP target states in Sec. \[sec:requirements\], the spin-meter coupling here is not constrained by a specific value of $g$ for a given input squeezing ${r}$ since the goal is only to produce distinguishable meter states. Practical limitations on ${{N_L}}$ in both the preparation and the measurement stage arise for two related reasons. First, the Faraday interaction is an effective description of the light-matter coupling when the quantum emitters remain far below saturation [@bib:Hammerer2010]. Second, increased ${{N_L}}$ precipitates more spontaneous photon scattering that spoils the QND interaction and measurement. Indeed, this has restricted QND spin squeezing to the Gaussian regime, far from a projective measurement. Overcoming the effects of decoherence necessitates that the coupling to the collective optical mode is large relative to all other modes. This is characterized by the optical density per atom, $\eta \coloneqq \sigma_0/A_m$, the ratio of the resonant atomic scattering cross section $\sigma_0$ to the transverse mode area $A_m$. While typical optical densities per atom in free space, $\eta \sim 10^{-5}$ [@bib:baragiola14], are far too weak for our purposes, those in engineered photonic environments such as photonic crystal waveguides can be much larger: $\eta \sim 1$ [@bib:goban14]. Operating near a band edge, “slow light" can further enhance the interaction by several orders of magnitude [@bib:goban14; @bib:Hood2016]. A detailed study of optical pumping for atoms very near and strongly coupled to a waveguide is beyond the scope of our work; nevertheless, an estimate of the required coupling can be found from a free-space model for alkali atoms. Here, the Faraday rotation angle per photon per unit angular momentum given above is $\chi = \eta \Gamma / 2 \Delta $, for detuning $\Delta$ and spontaneous emission rate $\Gamma$ [@bib:deutsch10]. To realize the coupling strength in Eq. required for symmetric code states and an approximately projective measurement while limiting the number of free-space scattered photons, we find that for ${{N_L}}= 10^4$ and $\Delta = 500 \Gamma$ the required optical density per atom is $\eta \sim 25$, within the reach of near-term technology. It may be possible to augment the atom-light coupling with an optical cavity [@bib:mcconnell15] and to suppress the deleterious effects of photon scattering by judiciously selecting the effective spin within each atom [@bib:Bohnet14]. Alternative fruitful avenues have opened in other physical architectures, where demonstrated strong coupling of “artificial atoms" to photonic environments could provide the necessary interaction strength [@bib:devoret13; @bib:arnold15]. In such systems, Purcell enhancement of the total coupling rate has the potential to reduce the collective spin’s susceptibility to other sources of noise. Conclusion {#sec:conclusion} ========== The fact that photons do not interact with each other directly—only indirectly via material interactions—both helps and hinders optical approaches to quantum computing. On the one hand, this non-interaction means that room-temperature experiments are possible with low noise from the environment. On the other hand, some matter-based mechanism is required to get the photons to interact—even if this is just interaction with a detector followed by postselection [@Knill2001]. The Gottesman-Kitaev-Preskill (GKP) encoding of qubits within light modes [@Gottesman2001] offers easier Clifford operations and built-in robustness to small errors at the price of resource states that are challenging to produce. This is analogous to how measurement-based quantum computing [@Raussendorf2001; @Menicucci2006] replaces the need for on-demand, controlled interactions with the generation of a resource state—in this case, a cluster state [@Briegel2001; @Zhang2006]—after which only (much easier) adaptive local measurements are required. In both cases, one trades away the requirement of many difficult interactions during the computation for the up-front single challenge of creating the required resource states, along with more modest requirements at run time. Despite the challenges associated with their generation, there is every reason to take GKP states seriously as a viable encoding for quantum information. In addition to their potential for use with optical CV cluster states [@Menicucci:2014cx], which is what we have focussed on here, subsequent results have shown that circuit QED may also be a viable platform for implementing fault-tolerant CV measurement-based quantum computing. A recent proposal exists to produce GKP qubits in circuit QED [@Terhal2016], and scalable CV cluster states can be made in circuit QED, as well [@Grimsmo2016]. Furthermore, GKP states have recently been shown to provide additional error-correction benefits when combined with a surface-code architecture [@Terhal2015], which could be used in both optics and circuit QED. N.C.M. thanks Gerard Milburn for the initial suggestion to consider interactions with atomic ensembles. B.Q.B. thanks Gavin Brennen and Ivan Deutsch for helpful discussions. K.R.M., B.Q.B., and A.G. acknowledge the Australian Research Council Centre of Excellence for Engineered Quantum Systems (Project number CE110001013). N.C.M. acknowledges the Australian Research Council Centre of Excellence for Quantum Computation and Communication Technology (Project number CE170100012). N.C.M. was additionally supported by the Australian Research Council under grant number DE120102204 and by the U.S. Defense Advanced Research Projects Agency (DARPA) Quiness program under Grant No. W31P4Q-15-1-0004. Kraus-operator formalism {#sec:GKPKraus} ======================== The theory of generalized quantum measurements provides a convenient formalism (see, for example, Ref. [@bib:NielsenChuang00]) to describe the state-preparation protocol described above. The conditional optical state is expressed using a set of Kraus operators corresponding to the outcomes of a spin measurement. We begin with a product state between a spin of total angular momentum $J$ and an optical field, ${\left\lvert\psi_b\right\rangle} = {\left\lvert\phi\right\rangle}_A \otimes {\left\lvert\psi\right\rangle}_O$, where the spin and field states are arbitrary. For measurement outcome $x$, the normalized, conditional field state is given by $$\begin{aligned} \label{eq:CondState} {\left\lvert\psi_d^x\right\rangle}_O = \frac{{\hat{A}}_{x} {\left\lvert\psi\right\rangle}_O}{ \sqrt{ \mathcal{P}(x) } }, \end{aligned}$$ where ${\hat{A}}_{x}$ is a Kraus operator and $\mathcal{P}(x)$ is the probability of outcome $x$. Expressing the spin state in the $z$-basis, ${\left\lvert\phi\right\rangle}_A = \sum_m c_m {\left\lvertm\right\rangle}$, the Kraus operator associated with the controlled-displacement interaction in Eq. (\[eq:BigD\]) is $$\begin{aligned} {\hat{A}}_{x} &\coloneqq {\left\langlex\right\rvert} {\hat{D}}_c(g) {\left\lvert\phi\right\rangle}_A = \sum_{m=-J}^J c_m d_{m,x} \, e^{ -i g m {\hat{p}} } . \label{eq:KrausOp} \end{aligned}$$ This describes the conditional operation that implements a set of displacements on the field proportional to the initial spin distribution, $c_m$, and outcome $x$. The probability of outcome $x$ is obtained by tracing over the initial field state, $$\begin{aligned} \label{eq:ProbKraus} \mathcal{P}(x) & = \mbox{Tr} \big[ {\hat{A}}^{\dagger}_{x} {\hat{A}}_{x} {\left\lvert{\psi}\middle\rangle\middle\langle{\psi}\right\rvert} \big] \\ & = \sum_{m, m'= -J}^J c_m c_{m'}^* d_{m,x} d_{m', x} {\left\langle\psi\right\rvert} e^{-i g(m-m'){\hat{p}}}{\left\lvert\psi\right\rangle}.\nonumber \end{aligned}$$ For a more general input state of light, ${\hat{\rho}}_O$, which may be mixed due to losses and errors in the squeezing procedure, the conditional field state is given by $$\begin{aligned} \label{Eq:CondRho} {\hat{\rho}}^x_{d} = \frac{{\hat{A}}_{x} {\hat{\rho}}_O {\hat{A}}^\dagger_{x}}{ \mathcal{P}(x) } \end{aligned}$$ with probability $\mathcal{P}(x) = \mbox{Tr}[ {\hat{A}}_{x}^\dagger {\hat{A}}_{x} {\hat{\rho}}_O ]$. Measurement probability {#measurement-probability .unnumbered} ----------------------- In Sec. \[sec:lightensemble\] we consider the case of a position-squeezed input field, Eq. (\[eq:BenBlows\]), and an initial spin-coherent state corresponding to $c_m = d_{m,J}$ in Eq. (\[eq:KrausOp\]). The Kraus operator description, Eq. (\[eq:CondState\]), then gives the conditional field state, Eq. (\[eq:projectedOpticalState\]). The probability of outcome $x$ follows directly from Eq. (\[eq:ProbKraus\]), $$\begin{aligned} \mathcal{P}(x) = \sum_{m,m'} d_{m,J}d_{m,x}d_{m',J}d_{m',x}{\left\langle{ {\slashed{g}}m , {r}}\middle\vert{ {\slashed{g}}m' , {r}}\right\rangle}.\end{aligned}$$ To evaluate the expression, we note that for real $\alpha$ and ${r}$, a displaced squeezed state can be written as $$\begin{aligned} {\left\lvert\alpha,{r}\right\rangle} = {\hat{D}}(\alpha){\hat{S}}({r}){\left\lvert0\right\rangle} = {\hat{S}}({r}){\hat{D}}\big(\alpha e^{{r}} \big){\left\lvert0\right\rangle}.\end{aligned}$$ Then, the overlap between states of different displacements is calculated simply, $$\begin{aligned} {\langle{\alpha,{r}}\|{\beta,{r}}\rangle} &= \exp\left[ -{\mbox{$\frac{1}{2}$}}e^{2{r}}(\alpha-\beta)^2\right], \label{eq:Overlap}\end{aligned}$$ and the probability becomes $$\begin{aligned} \label{eq:Pxappendix} \mathcal{P}(x) = \sum_{m,m'}d_{m,J}d_{m,x}d_{m',J}d_{m',x}e^{-\frac{1}{4}g^2e^{2{r}}(m-m')^2}.\end{aligned}$$ Analytic form of Wigner’s (small) d-matrix elements {#app:analyticform} =================================================== The matrix elements $d^{(J)}_{m,m'}(\beta)$ of Wigner’s (small) $d$-matrix can be summed analytically [@Varshalovich1988] using Jacobi polynomials $P_n^{(a,b)} (z)$: $$\begin{aligned} d_{m,m'}^{(J)}(\beta) & = \sum_k (-1)^{k-m'+m} \frac{\sqrt{(J+m')!(J-m')!(J+m)!(J-m)!}}{(J+m'-k)!k!(J-k-m)!(k-m'+m)!} \left( \cos \frac {\beta} {2} \right)^{2j-2k+m'-m} \left( \sin \frac {\beta} {2} \right)^{2k-m'+m} \\ & = {\varsigma}_{m,m'} \left[ \frac {s! (s+\mu+\nu)!} {(s+\mu)! (s+\nu)!} \right]^{\tfrac 1 2} \left( \sin \frac {\beta} {2} \right)^\mu \left( \cos \frac {\beta} {2} \right)^\nu P_s^{(\mu,\nu)} (\cos \beta) ,\end{aligned}$$ where $$\begin{aligned} \mu &= {\left\lvert{m-m'}\right\rvert} & \nu &= {\left\lvert{m+m'}\right\rvert} & s &= J - \frac 1 2 (\mu + \nu), & {\varsigma}_{m,m'} = \begin{cases} 1 & m' \ge m, \\ (-1)^{m'-m} & m' < m. \end{cases}\end{aligned}$$ When we evaluate this for $\beta = \pi/2$, this simplifies to $$\begin{aligned} d_{m,m'} \coloneqq d_{m,m'}^{(J)}\left( \frac \pi 2 \right) & = {\varsigma}_{m,m'} \left[ \frac {(2s+\mu+\nu)!} {(s+\mu)! (s+\nu)!} \frac {s! (s+\mu+\nu)!} {(2s+\mu+\nu)!} \right]^{\tfrac 1 2} 2^{s-J} P_s^{(\mu,\nu)} (0) \\ & = {\varsigma}_{m,m'} {\binom{2 J}{s+\mu }}^{\tfrac 1 2}{\binom{2 J}{s}}^{-\tfrac 1 2} 2^{s-J} P_s^{(\mu,\nu)} (0) .\end{aligned}$$ This is an equivalent analytic expression for the sum in Eq. . The two special cases of this we need for analysis are $m'=\pm J$ and $m'=0$. The first ($m'=\pm J$) is rather trivial since $s=0$, for which the Jacobi polynomial $P_0^{(a,b)}(z)=1$, leaving $$\begin{aligned} d_{m,\pm J} & = \frac {(\pm 1)^{J+m}} {2^J} {\binom{2 J}{J-m}}^{\tfrac 1 2} ,\end{aligned}$$ which agrees with Eq. . The second case ($m'=0$) takes a bit more work to evaluate. In this case, $\mu = \nu = {\left\lvert{m}\right\rvert}$ and $s = J - {\left\lvert{m}\right\rvert}$, which gives $$\begin{aligned} \label{eq:dm0part1} d_{m,0} & = {\varsigma}_{m,0} {\binom{2 J}{J}}^{\tfrac 1 2}{\binom{2 J}{J-{\left\lvert{m}\right\rvert}}}^{-\tfrac 1 2} 2^{-{\left\lvert{m}\right\rvert}} P_{J-{\left\lvert{m}\right\rvert}}^{({\left\lvert{m}\right\rvert},{\left\lvert{m}\right\rvert})} (0) .\end{aligned}$$ The Jacobi polynomial in this case can be simplified by first expressing it in terms of an ordinary hypergeometric function using the relation [@Bateman1953vol2] (§10.8) $$\begin{aligned} P^{(a,b)}_n(z) = 2^{-n} (z-1)^n \binom{b+n}{n} \,_2F_1 \left(-n,-a-n;b+1;\frac {z+1} {z-1} \right) .\end{aligned}$$ When $b = a$ and $z=0$, this simplifies to $$\begin{aligned} P^{(a,a)}_n(0) & = (-2)^{-n} \binom{a+n}{n} \,_2F_1(-n,-a-n;a+1; -1) \nonumber \\ & \overset{\text {(a)}}{=} (-2)^{-n} \binom{a+n}{n} \frac {\Gamma (1+a) \Gamma (1-\frac{n}{2})} {\Gamma (1-n) \Gamma (1+a+\frac{n}{2})} \nonumber \\ & = (-2)^{-n} \binom{a+n}{\frac{n}{2}} \frac {\Gamma (1+\frac{n}{2}) \Gamma (1-\frac{n}{2})} {\Gamma (1+n) \Gamma (1-n)} \nonumber \\ & \overset{\text {(b)}}{=} (-2)^{-n} \binom{a+n}{\frac{n}{2}} \cos \frac {n \pi} {2} .\end{aligned}$$ In (a) we have used Kummer’s Theorem [@Bailey1935] (§2.3) to reduce the hypergeometric function in this particular instance to an expression involving only gamma functions, and in (b) we employed Euler’s reflection lemma, which can be written $\Gamma(1+z) \Gamma(1-z) = \pi z \csc \pi z$. Plugging into Eq. , we get $$\begin{aligned} d_{m,0} & = {\varsigma}_{m,0} {\binom{2 J}{J}}^{\tfrac 1 2}{\binom{2 J}{J-{\left\lvert{m}\right\rvert}}}^{-\tfrac 1 2} 2^{-{\left\lvert{m}\right\rvert}} (-2)^{-(J-{\left\lvert{m}\right\rvert})} \binom{J}{\frac{J-{\left\lvert{m}\right\rvert}}{2}} \cos \frac {(J-{\left\lvert{m}\right\rvert})\pi} {2} \nonumber \\ & = {\operatorname{Re}}(i^{J+m}) 2^{-J} {\binom{2 J}{J}}^{\tfrac 1 2}{\binom{2 J}{J-m}}^{-\tfrac 1 2} \binom{J}{\frac{J-m}{2}} .\end{aligned}$$ Note that we have used complex notation in the final answer instead of a cosine (for brevity) and that no absolute values remain. We can now calculate the products we need for the main text: $$\begin{aligned} \label{eq:dmJdmpmJ} d_{m,J}d_{m,\pm J} & = \frac {(\pm 1)^{J+m}} {2^{2J}} {\binom{2 J}{J-m}} , \\ d_{m,J}d_{m,0} & = \frac {{\operatorname{Re}}(i^{J+m})} {2^{2J}} {\binom{2 J}{J}}^{\tfrac 1 2} \binom{J}{\frac{J-m}{2}} .\end{aligned}$$ Furthermore, using the Gaussian approximation [@Spencer2014] (§5.4) $$\begin{aligned} \binom{n}{k} & \approx 2^{n} \sqrt{\frac{2} {\pi n}} \exp \left(-\frac{(n-2k)^2}{2n}\right) ,\end{aligned}$$ we can approximate these products as $$\begin{aligned} d_{m,J}d_{m,\pm J} & \approx \frac{(\pm 1)^{J+m}} {\sqrt{\pi J}} \exp \left(-\frac{m^2}{J}\right) , \\ d_{m,J}d_{m,0} & \approx {\operatorname{Re}}(i^{J+m}) \frac {\sqrt 2} {(\pi J)^{3/4}} \exp \left(-\frac {m^2}{2J} \right) ,\end{aligned}$$ valid for $m = o(J^{2/3})$—i.e., everywhere except at the far tails of the distribution (where it is very small anyway and thus irrelevant, assuming a reasonably sized $J \gg \frac 1 2$). These are used in deriving the approximations in Eqs. and , respectively. Variance of peaks in momentum representation {#app:cosvariance} ============================================ The momentum representation of the resource states, Eq. (\[eq:psipJ\]), is a Gaussian envelope multiplying a comb generated by $\cos^{2J}(pg/2)$. We want to approximate each peak in the comb by a Gaussian by matching the peak’s variance. Treating a single peak as a probability distribution we have $$\begin{aligned} P(p) = \frac{g\Gamma(J+1)}{2\sqrt{\pi}\Gamma(J+ \frac 1 2)}\cos^{2J}(pg/2)\end{aligned}$$ where the prefactors ensure the normalisation $$\begin{aligned} \int_{-\pi/g}^{\pi/g}P(p)dp = 1.\end{aligned}$$ The variance is then calculated in the usual way: $$\begin{aligned} \sigma_p^2 = \langle p^2 \rangle - \langle p \rangle^2 = \int_{-\pi/g}^{\pi/g}P(p)p^2dp.\end{aligned}$$ Performing the integral we find $$\begin{aligned} \sigma_p^2 = \frac{2(J^2\zeta(2,J)-1)}{g^2J^2},\end{aligned}$$ where $\zeta(s,J)$ is the Hurwitz zeta function $$\begin{aligned} \zeta(s,J) = \sum_{k=0}^\infty \frac{1}{(k+J)^{s}}.\end{aligned}$$ For large $J$, $\zeta(2,J)\approx \frac{1}{J}+\frac{1}{2J^2}+O(\frac{1}{J^3})$ (see for example [@bib:paris04]). Hence, $$\begin{aligned} \sigma_p^2 \approx \frac{2}{g^2 J} + O\left( \frac {1} {J^2} \right) = \sigma_{q,\text{env}}^{-2} + O\left( \frac {1} {J^2} \right).\end{aligned}$$	{ "pile_set_name": "ArXiv" }
--- abstract: 'In this paper we consider the (weighted) spectral measure $\mu_n$ of a $n\times n$ random matrix, distributed according to a classical Gaussian, Laguerre or Jacobi ensemble, and show a moderate deviation principle for the standardised signed measure $\sqrt{n/a_n}(\mu_n -\sigma)$. The centering measure $\sigma$ is the weak limit of the empirical eigenvalue distribution and the rate function is given in terms of the $L^2$-norm of the density with respect to $\sigma$. The proof involves the tridiagonal representations of the ensembles which provide us with a sequence of independent random variables and a link to orthogonal polynomials.' author: - Jan Nagel bibliography: - 'detnag.bib' title: Moderate deviations for spectral measures of random matrix ensembles --- Keywords: spectral measure, random matrix theory, moderate deviations Introduction ============ In the theory of random matrices the three central distributions on the set of self-adjoint matrices are the Gaussian, the Laguerre and the Jacobi ensemble. For these distributions, a lot is known about the asymptotic properties of the empirical eigenvalue distribution $$\begin{aligned} \hat\mu_n = \frac{1}{n} \sum_{i=1}^n \delta_{\lambda_i} ,\end{aligned}$$ when $\lambda_1,\dots ,\lambda_n$ are the eigenvalues of the matrix under consideration and $\delta_x$ denotes the Dirac measure in $x$. The random measure $\hat\mu_n$ converges almost surely weakly to a limit distribution $\sigma$ with compact support. For the Gaussian ensemble, this is the famous semicircle law ([@wigner1958]). In the Laguerre case and Jacobi case, the limit distributions are the Marchenko-Pastur law and the semicircle law (see [@marpas1967] and [@col2005]). An overview and more recent results can be found in the monographs of [@mehta2004], [@forrester2010] and [@agz2010]. Recently, a number of paper deal with the weighted version of the empirical eigenvalue distribution, the spectral measure defined as $$\begin{aligned} \label{wspectralmeasure} \mu_n = \sum_{i=1}^n w_i \delta_{\lambda_i} ,\end{aligned}$$ with weights $w_i = \|\langle u_i,e_1\rangle \|^2$ given by the squared modulus of the top entries of the eigenvectors $u_1,\dots ,u_n$, see the paper of [@gamloz2004], [@gamrou2009b; @gamrou2009a]. This weighted measure encodes some information about the eigenvectors and appears in the general spectral theorem. For the general formulation suppose $A$ is a linear, continuous and self-adjoint operator on a Hilbert space $\mathcal{H}$ with a cyclic vector $h \in \mathcal{H}$ (i.e., the span of $h, Ah, A^2h\dots $ is $\mathcal{H}$). The spectral theorem yields the existence of a spectral measure $\mu$ on some measureable space $(X, \mathcal{X})$, a unitary isometry $U: L^2(\mu) \rightarrow \mathcal{H}$ and a bounded measureable function $\phi$ on $X$ such that $$\begin{aligned} \left( U^* A U f \right) (x) = \phi (x) f(x)\end{aligned}$$ for each $f \in L^2(\mu)$ (see [@dunfschw1963]). This statement is sometimes abbreviated to “every Hermitian operator is unitarily equivalent to a multiplication”. The isometry $U$ maps a polynomial $p(x)$ to $p(A) h$, which implies that the moments of $\mu$ satisfy $$\begin{aligned} \int x^k \mu(dx) = \langle 1,x^k \rangle_{L^2} = \langle h,A^k h \rangle \end{aligned}$$ for all $k\geq 1$. If $\mathcal{H}$ is $n$-dimensional and $h=e_1$, then this gives the moments of . By the functional calculus, $\int fd\mu_n = f(A)_{1,1}$ for suitable test functions $f$, so the behaviour of $\mu_n$ is related to matrix elements of functions of random matrices, which were studied by [@lytpast2009; @lytpast2009b] and [@pizzo2012]. We are concerned with moderate deviation properties of the spectral measure. To be self-contained, we recall the definition of large and moderate deviations. Let $U$ be a topological Hausdorff space, let $\mathcal{I}: U \rightarrow [0,\infty]$ be a lower semicontinuous function and $a_n$ a sequence of positive real numbers with $a_n\to \infty$. We say that a sequence $(P_{n})_n$ of probability measures on $(U,\mathcal{B}(U))$ satisfies a large deviation principle (LDP) with speed $a_n$ and rate function $\mathcal{I}$ if: - For all closed sets $F \subset U$: $$\begin{aligned} \limsup_{n\rightarrow\infty} \frac{1}{a_n} \log P_{n}(F)\leq -\inf_{x\in F}\mathcal{I}(x)\end{aligned}$$ - For all open sets $O \subset U$: $$\begin{aligned} \liminf_{n\rightarrow\infty} \frac{1}{a_n} \log P_{n}(O)\geq -\inf_{x\in O}\mathcal{I}(x)\end{aligned}$$ The rate function $\mathcal{I}$ is good if its level sets $\{x\in U \|\ \mathcal{I}(x)\leq a\}$ are compact for all $a\geq 0$. We say that a sequence of $U$-valued random variables satisfies an LDP if their distributions satisfy an LDP. If $a_n \to \infty$, but $a_n/n \to 0$, we say that $(P_{n})_n$ satisfies a moderate deviation principle (MDP). Moderate deviations are concerned with the convergence of measure at a scaling between the law of large numbers and the central limit theorem. On this scale, the rate function is identical for all possible speeds. The empirical eigenvalue distribution $\hat\mu_n$ satisfies a large deviation principle with speed $n^2$ and rate function related to Voiculescu’s entropy, see [@bengui1997], [@hiaipetz1998] and [@hiaipetz2006]. The spectral measure $\mu_n$ has the same weak limit $\sigma$ as the empirical distribution, the large deviation properties however, are different: the random weights reduce the speed to $n$. In the case of the Gaussian ensemble, the rate function is explicit and involves the Kullback-Leibler distance and a sum over atoms outside the limit support ([@gamrou2009b]). In several cases, moderate deviations of empirical eigenvalue distributions were proven. [@demguizei2003] considered the perturbation of a Gaussian matrix and [@doereich2010; @doereich2011] showed moderate deviations for Wigner matrices generalising the Gaussian ensemble. Our main result is Theorem \[haupt\], a moderate deviation principle for the rescaled measure $\sqrt{n/a_n}(\mu_n -\sigma)$, when $\mu_n$ is the random spectral measure of one of the three classical ensembles. The proof of Theorem \[haupt\] consists of two main steps. The first one is a moderate deviation principle for the moments of the rescaled measure. The key part is a transformation to new random variables which are independent and related to the polynomials orthogonal with respect to the spectral measure. In the second step we apply the contraction principle to lift the result into the space of measures. Due to this method, the MDP will be in the topology induced by the convergence of moments. This paper is structures as follows: in Section 2 we introduce the classical ensembles and corresponding spectral measures and we state our results. The transformation in the context of orthogonal polynomials is formulated in Section 3. Finally, Section 4 contains the proofs. Random matrix ensembles, results ================================ The central distribution of Hermitian matrices and a generalisation of the scalar normal distribution is the Gaussian ensemble. Following the threefold way of [@dyson1962b] it comes in a real, complex and quaternion version and is motivated by several physical observations (see [@mehta2004]). In the real case the Gaussian orthogonal ensemble (GOE) is the distribution of a $n\times n$ matrix with centered Gaussian entries such that the random variables on and above the diagonal are independent, the diagonal entries have variance $\frac{2}{n}$ and the off-diagonal entries have variance $\frac{1}{n}$. For the complex Gaussian unitary ensemble (GUE), the off diagonal entries are complex Gaussian distributed with variance $\frac{1}{2n}$, the real diagonal entries have variance $\frac{1}{n}$. The quaternion matrices can be constructed in a similar way. In all three cases, the density of the eigenvalues can be written as $$\begin{aligned} \label{ewge} f_G(\lambda) = c_G \| \Delta (\lambda )\|^{\beta} \prod_{i=1}^n e^{-\frac{\beta n}{4}\lambda_i^2},\end{aligned}$$ with $\| \Delta (\lambda )\|=\prod_{i< j} \|\lambda_i-\lambda_j\|$ the Vandermonde determinant, $\beta$ the real dimension of the entries and $c_G$ a normalisation constant. The Laguerre ensemble (or Wishart ensemble) appears as the distribution of covariance estimators in multivariate statistics ([@wishart1928]). It can be constructed as the square $XX^$ of a real, complex or quaternion Gaussian matrix $X$, the eigenvalue density is given by $$\begin{aligned} \label{ewle} f_L(\lambda) = c_L^\gamma \| \Delta (\lambda )\|^{\beta} \prod_{i=1}^n \lambda_i^{\gamma} e^{-\frac{\beta n}{2} \lambda_i} \mathbbm{1}_{\{ \lambda_i>0\} },\end{aligned}$$ where $\gamma>-1 $ is a parameter determined by the size of $X$. The third distribution is the Jacobi ensemble with applications in multivariate analysis of variance (MANOVA, see [@muirhead1982]) and eigenvalue density $$\begin{aligned} \label{ewje} f_J(\lambda) = c_J^{\gamma,\delta} \| \Delta (\lambda )\|^{\beta} \prod_{i=1}^n \lambda_i^{\gamma} (1-\lambda_i)^{\delta} \mathbbm{1}_{\{ 0<\lambda_i<1 \} } .\end{aligned}$$ and with parameters $\gamma ,\delta >-1$. The corresponding matrix model can be written as the ratio of two Laguerre matrices. It is a common feature of the three ensembles that the distribution is invariant under orthogonal (or unitary or symplectic) conjugations. As a consequence, the matrix of eigenvalues is Haar distributed on the corresponding compact group and independent of the eigenvalues, see [@dawid1977]. Consequently, the first unit vector is almost surely cyclic and the weight vector containing the square moduli of the top entries of the eigenvectors $$\begin{aligned} w = (w_1,\dots ,w_n) = \left( \|\langle u_1,e_1\rangle \|^2,\dots ,\|\langle u_n,e_1\rangle \|^2 \right) \end{aligned}$$ follows a Dirichlet distribution Dir($\frac{\beta}{2}$) with density $$\begin{aligned} \label{dirdensity} \frac{\Gamma (\tfrac{n\beta}{2} )}{\Gamma (\tfrac{\beta}{2})^n} w_1^{\beta/2-1} \dots w_n^{\beta/2-1} \mathbbm{1}_{\{ w_1+\dots + w_n=1, w_i> 0 \} } \end{aligned}$$ ([@johnson1976]). The distribution of the eigenvalues and the Dirichlet distribution of the weights $w$ are well-defined not only for $\beta \in \{1,2,4\}$ but for all values $\beta >0$ and we can consider random spectral measures of the three ensembles in the general case. To be precise, we say that $$\begin{aligned} \mu_n = \sum_{i=1}^n w_i \delta_{\lambda_i}\end{aligned}$$ is a spectral measure of a classical ensemble if the distribution of the support points is given by , or independent of the weights, which have density . The eigenvalue distributions defined here are normalised such that the empirical eigenvalue distribution $\hat\mu_n$ has a weak limit almost surely. Moreover, they satisfy a large deviation principle with speed $n^2$ and good rate functions related to Voiculescu’s entropy. For the Gaussian ensemble, the equilibrium measure and thus the limit of $\hat\mu_n$ is Wigner’s famous semicircle law $$\begin{aligned} \label{semicircle} \rho(dx) = \frac{1}{2\pi} \sqrt{4-x^2} dx .\end{aligned}$$ In the Laguerre case, the limit is the Marchenko-Pastur distribution $$\begin{aligned} \label{marchpast} \eta(dx) = \frac{\sqrt{x(4-x)}}{2\pi x} \mathbbm{1}_{ \{ 0<x<4 \} } dx\end{aligned}$$ and the eigenvalue distribution of the Jacobi ensemble converges to the arcsine law $$\begin{aligned} \label{arcsine} \nu (dx) = \frac{1}{\pi \sqrt{x(1-x)}} \mathbbm{1}_{ \{ 0<x<1 \} } dx .\end{aligned}$$ The same convergence holds for the weighted versions, the spectral measures $\mu_n$. More precisely, suppose that the expected moments of $\hat\mu_n$ converge, then bounding the moments of the Dirichlet weights, we can show $$\begin{aligned} \left\| m_k(\mu_n) - m_k (\hat{\mu}_n)\right\| = O_P\left( \frac{1}{\sqrt{n}} \right)\end{aligned}$$ for all $k$ as $n\to \infty$, where $$\begin{aligned} m_k(\mu) = \int x^k \mu(dx)\end{aligned}$$ denotes the $k$-th moment of a measure $\mu$. Interestingly, proving convergence of expected moments is precisely the way a number of limit theorems in the framework of free probability are proven, where a comprehensive overview is given by [@nicspe2006]. In order to present the results in a unified way and to avoid repetitions, we denote by $\mu_n$ the spectral measure of one of the three ensembles and the limit law by $\sigma$. We regard the standardised measure $\mu_n-\sigma$ as a random element of the set $$\begin{aligned} \mathcal{M}_0 = \left\{ \mu \| \mu \mbox{ finite signed measure on } \mathbb{R} \mbox{ with compact support, } \mu(\mathbb{R})=0 \right\} .\end{aligned}$$ We endow the space $\mathcal{M}_0$ with the topology of convergence of moments, that is, $\mu_n \to \mu$ if for all moments $m_k(\mu_n) \to m_k(\mu)$ holds. The family of monomials $x^k$ is separating on $\mathcal{M}_0$, in the sense that for $\mu \neq \nu$ there is a $k$ such that $\int x^kd\mu \neq \int x^k d\nu$. A metric for this topology is then given by $$\begin{aligned} \label{metric} d_\mathbf{m} (\mu,\nu) =\sum_{k=0}^\infty 2^{-k} \frac{\left\| \int x^k d(\mu-\nu) \right\|}{1+\left\| \int x^k d(\mu-\nu) \right\|} .\end{aligned}$$ Note that on $\mathcal{M}_0$, the topology of convergence of moments is neither stronger nor weaker than the topology of weak convergence. The $\sigma$-algebra is then the usual Borel algebra. The following theorem is our main result. \[haupt\] Let $\mu_n$ be a random spectral measure of a classical ensemble with weak limit $\sigma$. Then $(\sqrt{n/a_n}(\mu_n - \sigma))_n$ satisfies the moderate deviation principle in $\mathcal{M}_0$ with speed $a_n$ and good rate function $$\begin{aligned} \mathcal{I}(\mu) = \frac{\beta}{4} \int \left( \frac{\partial \mu}{\partial \sigma }\right)^2 d\sigma ,\end{aligned}$$ where $\frac{\partial \mu}{\partial \sigma }$ is the signed density with respect to $\sigma$. If $\mu$ is not absolutely continuous with respect to $\sigma$, we have $\mathcal{I}(\mu)=\infty$. For the first part of the proof, we consider the behaviour of moments of $\mu_n$. Motivated from the theory of random moment sequences, [@detnag2012] obtained central limit theorems for the moment vectors of the classical ensembles. Let $m^{(k)}(\mu) = (m_1(\mu) ,\dots ,m_k(\mu))^T$ denote the vector of the first $k$ moments. Then the convergence in distribution $$\begin{aligned} \label{clt} \sqrt{\frac{\beta n}{2}} (m^{(k)}(\mu_n) - m^{(k)}(\sigma)) \xrightarrow[n \rightarrow \infty ]{\mathcal{D}} \mathcal{N}_k(0,D_kD_k^T),\end{aligned}$$ holds, where $D_k$ is a lower triangular matrix depending on the ensemble under consideration. For the Gaussian ensemble, $$\begin{aligned} \label{covgau} (D_k^{(G)})_{i,j} = \binom{i}{\frac{i-j}{2}} - \binom{i}{\frac{i-j}{2} - 1}.\end{aligned}$$ for $i\geq j$ and $i+j$ even and $(D_k^{(G)})_{i,j}=0$ otherwise. As usual we let $\binom{i}{-1}=0$. The asymptotic covariance for the Laguerre ensemble satisfies $$\begin{aligned} \label{covlag} (D_k^{(L)})_{i,j} = \binom{2i}{i-j} - \binom{2i}{i-j-1} \qquad j \le i \end{aligned}$$ and in the Jacobi case we have $$\begin{aligned} \label{covjac} (D_k^{(J)})_{i,j} = \frac{2^{-2i+1}}{\sqrt{2}} \binom{2i}{i-j} \qquad j \le i .\end{aligned}$$ From the central limit theorem, we are able to prove a moderate deviation principle on the level of moments. \[step1\] The sequence of moment vectors $\sqrt{n/a_n}(m^{(k)}(\mu_n) - m^{(k)}(\sigma))$ satisfies a moderate deviation principle with speed $a_n$ and good rate function $$\begin{aligned} \mathcal{I}_m(m)= \frac{\beta}{4} \|\|D_k^{-1} m\|\|^2 ,\end{aligned}$$ where $D_k$ is the matrix in . In order to apply the contraction principle, we need an interpretation of $D_k^{-1}$ in terms of $\sigma$. As it turns out, the covariance matrix in has a very interesting structure. \[step2\] Let $\mu_n$ be the spectral measure of a classical ensemble with weak limit $\sigma$, then the covariance matrix in has the representation $$\begin{aligned} (D_kD_k^T)_{i,j} = m_{i+j}(\sigma) - m_i(\sigma) m_j(\sigma) .\end{aligned}$$ Orthogonal polynomials, tridiagonal representations =================================================== Let $\mu$ be a measure on the real line with compact support, and let $p_0(x),p_1(x),\dots$ be the monic polynomials orthogonal with respect to $\mu$, i.e., $$\begin{aligned} \int p_k(x)p_l(x) \mu(dx) = c_k \delta_{k,l}\end{aligned}$$ and the leading coefficient of $p_k(x)$ is 1. The polynomials satisfy a three-term-recursion $$\begin{aligned} \label{polrek} x p_k(x) = p_{k+1}(x) + b_{k+1} p_k(x) + a_k p_{k-1}(x) \quad \mbox{ for } 1\leq k \leq n-1,\end{aligned}$$ with $a_k >0$ and initial conditions $p_0(x) = 1,p_1(x) = x-b_1$. Indeed, Favard’s Theorem (see [@szego1939]) says that any sequence of polynomials satisfying with $a_k>0$ is orthogonal with respect to some measure $\mu$. Moreover, the measure $\mu$ is supported on $[0,\infty)$ if and only if there is a sequence of variables $z_k$ such that the decompositions $$\begin{aligned} a_k =& z_{2k-1} z_{2k} \label{zerlegungan},\\ b_k =& z_{2k-2} + z_{2k-1} \label{zerlegungbn}, \end{aligned}$$ hold with $z_k >0$ for $k \geq 1$ and $z_0=0$ (see [@chihara1978]). We will call the quantities $z_1,z_2,\dots $ recursion variables. If the support of the measure is contained in the compact interval $[0,1]$, the recursion variables form a chain sequence $$\begin{aligned} \label{zerlegungzn} z_k = (1-p_{k-1})p_k\end{aligned}$$ with $p_k \in (0,1)$ for $k=1,\dots ,2n-1$ and $p_0 = 0$ ([@wall1948]). The $p_k$ are called canonical moments, where for a comprehensive overview on canonical moments we refer to the monograph of [@dettstud1997].\ Note that for any $k$, there is a one-to-one relation between the first $k$ moments of the measure $\mu$ and the first $k$ recursion coefficients or, if they exists, recursion variables or canonical moments. This can be seen by looking at the first orthogonal polynomials: fixing $2k-1$ moments is equivalent to fixing the first $k$ polynomials while the first $2k$ moments uniquely determine the first $k$ polynomials and additionally the norm of $p_k(x)$.\ To be precise, the degree up to which the polynomials can be defined depends on the number of support points of $\mu$. If $a_1\dots a_k>0$, then orthogonal polynomials up to degree $k-1$ exist. In the cases we consider, all relevant polynomials can be defined.\ The limit eigenvalue distributions of the ensembles all have remarkably simple representations in the new variables: The recursion coefficients of the semicircle distribution satisfy $a_k=1, b_k=0$ for all $k$, the resulting polynomials are the Chebychev-polynomials of the second kind (rescaled on $[-2,2]$). For the Marchenko-Pastur law, the recursion variables $z_k$ are all equal to 1 and the canonical moments of the arcsine law are all given by $\tfrac{1}{2}$ ([@skibinsky1969]). Now suppose that $\mu_n$ is the spectral measure of the tridiagonal matrix $$\begin{aligned} \label{tridiag} T_n = \begin{pmatrix} b_1 & \sqrt{a_1} & & \\ \sqrt{a_1} & b_2 & \ddots & \\ & \ddots & \ddots & \sqrt{a_{n-1}} \\ & & \sqrt{a_{n-1}} & b_n \end{pmatrix}\end{aligned}$$ with a positive subdiagonal, then the polynomials defined by the recursion coefficients $a_k,b_k$ are orthogonal with respect to $\mu_n$ (see [@deift2000]). In the language of the spectral theorem, $T_n$ represents the multiplication by the identity in the basis of orthogonal polynomials.\ For the three classical ensembles, there exist tridiagonal reductions of the full matrix models such that for $\beta \in \{1,2,4\}$, the spectral measure of $T_n$ has the same distribution as the spectral measure of the full matrices. Moreover, the tridiagonal matrix yields the right distribution of eigenvalues and weights of $\mu_n$ for any value $\beta>0$. For the first two ensembles, the tridiagonal representation was obtained by [@dumede2002]. In the Gaussian case, $$\begin{aligned} a_k \sim \operatorname{Gamma}(\tfrac{\beta}{2}(n-k),\tfrac{2}{\beta n}), \qquad b_k \sim \mathcal{N}(0,\tfrac{2}{\beta n}) \end{aligned}$$ and all recursion coefficients are independent. For the Laguerre ensemble with parameter $\gamma$, the spectral measure is supported on $[0,\infty)$ and the recursion variables satisfy $$\begin{aligned} z_{2k-1} \sim \operatorname{Gamma}(\tfrac{\beta}{2}(n-k)+\gamma +1,\tfrac{2}{\beta n}), \qquad z_{2k} \sim \operatorname{Gamma}(\tfrac{\beta}{2}(n-k),\tfrac{2}{\beta n})\end{aligned}$$ and $z_1,\dots z_{2n-1}$ are independent. [@kilnen2004] constructed a reduction of the Jacobi ensemble. In this case, the canonical moments $p_1,\dots ,p_{2k-1}$ are independent and $$\begin{aligned} p_k \sim \begin{cases} \operatorname{Beta}\left( \tfrac{2n-k}{4} \beta , \tfrac{2n-k-2}{4} \beta +\gamma+\delta+2 \right)\quad & k \mbox{ even,} \\ \operatorname{Beta}\left( \tfrac{2n-k-1}{4} \beta +\gamma+1, \tfrac{2n-k-1}{4} \beta +\delta+1 \right)\quad & k \mbox{ odd.} \end{cases} \end{aligned}$$ In any of the three cases, this provides us with a bijective mapping $$\begin{aligned} \psi_k: m^{(k)} \longmapsto r^{(k)}\end{aligned}$$ where $r^{(k)}$ is the vector of first $k$ independent variables (recursion coefficient, recursion variables or canonical moments, respectively). $\psi_k$ is continuous and if $\mu_n$ is a spectral measure, the new variables $r^{(k)}(\mu_n)$ satisfy $r^{(k)}(\mu_n) \to r^{(k)}(\sigma)$ almost surely. Moreover, we have (see [@detnag2012]) $$\begin{aligned} \frac{\partial \psi^{-1}_k}{\partial r^{(k)}} (r^{(k)}(\sigma)) = D_k\end{aligned}$$ with the matrix $D_k$ as in . Proofs ====== \[mdps\] Let $\alpha,\beta,\gamma$ be positive constants, and $a_n$ be a sequence with $a_n\to \infty$ and $a_n/n\to 0$. - The sequence $(\sqrt{\tfrac{n}{a_n}}X_n)_n$, where $X_n\sim \mathcal{N}(0,\tfrac{\alpha}{n})$ satisfies a MDP with speed $a_n$ and good rate function $I_1(x) = \tfrac{1}{2\alpha} x^2$. - The sequence $(\sqrt{\tfrac{n}{a_n}}(Y_n-1))_n$ with $Y_n \sim \operatorname{Gamma}(\alpha n+\beta,\tfrac{1}{\alpha n})$ satisfies a MDP with speed $a_n$ and good rate function $I_2(x) = \tfrac{\alpha}{2} x^2$. - The sequence $(\sqrt{\tfrac{n}{a_n}}(Z_n-\tfrac{1}{2}))_n$ with $Z_n \sim \operatorname{Beta}(\alpha n+\beta,\alpha n+\gamma)$ satisfies a MDP with speed $a_n$ and good rate function $I_3(x) = 4\alpha x^2$. The first and second moderate deviations are well-known and follow from Theorem 3.7.1 in the book of [@demzei1998]. For the third sequence, we have the equality in distribution $$\begin{aligned} Z_n = \frac{W_1}{W_1+W_2} =:\phi(W_1,W_2)\end{aligned}$$ with $W_1,W_2$ independent, $W_1 \sim \operatorname{Gamma}(\alpha n+\beta,\tfrac{1}{\alpha n}),\ W_2 \sim \operatorname{Gamma}(\alpha n+\gamma,\tfrac{1}{\alpha n})$. From $(ii)$ we know that the vector $$\begin{aligned} \sqrt{n/a_n}(W_1-1,W_2-1)^T\end{aligned}$$ satisfies a MDP with speed $a_n$ and good rate function $$\begin{aligned} I(x,y) = \tfrac{\alpha}{2}(x^2+y^2) .\end{aligned}$$ Then the $\delta$-method for moderate deviations ([@gaozhao2011]) implies a MDP for $\sqrt{\tfrac{n}{a_n}}(Z_n-\tfrac{1}{2}) = \sqrt{\tfrac{n}{a_n}}(\phi(W_1,W_2)-\phi(1,1))$ with good rate function $$\begin{aligned} I_3(z) &= \inf \left\{ I(x,y)\ :\ \phi'(1,1)\cdot (x,y) = z\right\}\\ &= \inf \left\{ \tfrac{\alpha}{2}(x^2+y^2) :\ \tfrac{1}{4}x - \tfrac{1}{4}y = z\right\}\\ &= \tfrac{\alpha}{2} ((2z)^2+(-2z)^2) = 4{\alpha} x^2\end{aligned}$$ Proof of Theorem \[step1\]:\ Let $\mu_n$ be the spectral measure of the Jacobi ensemble with parameters $\gamma, \delta$. Then by the tridiagonal reduction of the Jacobi ensemble in Section 3, the canonical moments $p_1,\dots ,p_{2n-1}$ are independent and $$\begin{aligned} p_k \sim \operatorname{Beta}(\tfrac{\beta}{2} n+c_k,\tfrac{\beta}{2} n+d_k) \end{aligned}$$ with constants $c_k,d_k$ not depending on $n$. By Proposition \[mdps\], each normalised canonical moment $\sqrt{\tfrac{n}{a_n}}(p_k-\tfrac{1}{2})$ satisfies the MDP with speed $a_n$ and good rate function $I_3(x) = 2\beta x^2$. By the independence, the vector $$\begin{aligned} \sqrt{\frac{n}{a_n}}(p_1-\tfrac{1}{2},\dots ,p_k-\tfrac{1}{2})^T\end{aligned}$$ satisfies the MDP with speed $a_n$ and rate $I_k(x) = 2\beta \|\|x\|\|^2$, $x\in \mathbb{R}^k$. Recall that $\psi^{-1}_k$ maps the first $k$ canonical moments to the first $k$ moments. The canonical moments of the arcsine law $\nu$ are all identical to $\tfrac{1}{2}$, so $\psi^{-1}_k(p_1,\dots ,p_k) = m^{(k)}(\mu_n)$ and $\psi^{-1}_k(\tfrac{1}{2},\dots ,\tfrac{1}{2}) = m^{(k)}(\nu)$. The $\delta$-method yields a MDP for the normalised moment vector $\sqrt{\tfrac{n}{a_n}}(m^{(k)}(\mu_n) - m^{(k)}(\nu))$ with good rate function $$\begin{aligned} I_m(m) =& \inf \left\{ I_k(x)\ :\ \psi_k'(\tfrac{1}{2},\dots ,\tfrac{1}{2})\cdot x = m \right\} \\ =& 2 \beta \left\| \left\| \left( \psi_k'(\tfrac{1}{2},\dots ,\tfrac{1}{2})\right)^{-1} m \right\| \right\|^2 .\end{aligned}$$ The derivative of $\psi$ at $(\tfrac{1}{2},\dots ,\tfrac{1}{2})^T$ was first calculated by [@chakemstu1993] and is given by $2\sqrt{2} A$ with the $k\times k$ matrix $A$ given by . This proves the theorem for the Jacobi ensemble. For the other two ensembles, the arguments are the same: first, Proposition \[mdps\] implies a MDP for the independent auxiliary variables (the recursion coefficients or the recursion variables), then the $\delta$-method tranfers this MDP to the moments. The derivative appearing in the rate function is given by the matrices and , as calculated in the paper of [@detnag2012]. Proof of Lemma \[step2\]:\ For the Jacobi ensemble, this is Lemma 2.1 in the paper of [@chakemstu1993]. In the Laguerre and Gaussian case, we use a combinatoric interpretation in terms of the generalised Catalan numbers $$\begin{aligned} d_{i,j}= \binom{i+j}{i} - \binom{i+j}{i-1} .\end{aligned}$$ for $i\leq j$, see [@finu1976]. We remark that the Gaussian case could also be proven using the CLT-result of [@lytpast2009], however since it follows directly from our arguments in the Laguerre case, we will follow this route. The generalised Catalan number $d_{i,j}$ counts the number of paths in the triangular lattice $\big( (i,j)\big)_{j\geq i}$ starting at $(i,j)$ and ending in $(0,0)$ and in each vertex are only allowed to move up or to the left (see Figure 1). For $i=j$ this is a usual Catalan path and we obtain the Catalan number $d_{i,i}$, which is the $i$-th moment of the Marchenko-Pastur distribution $\eta$ ([@nicspe2006]). \[path\] (0,0)(8,7) (0,6)(7,6) (1,5)(7,5) (2,4)(7,4) (3,3)(7,3) (4,2)(7,2) (5,1)(7,1) (7,6)(7.5,6) (7,5)(7.5,5) (7,4)(7.5,4) (7,3)(7.5,3) (7,2)(7.5,2) (7,1)(7.5,1) (1,6)(1,5) (2,6)(2,4) (3,6)(3,3) (4,6)(4,2) (5,6)(5,1) (6,6)(6,1) (7,6)(7,1) (6,1)(6,0.5) (7,1)(7,0.5) (7,1)(7,2)(6,2)(6,4)(4,4)(4,5)(1,5)(1,6)(0,6) (0,6)(7,1) (0.4,6.5)[$(0,0)$]{} (7.8,1.4)[$(5,7)$]{} Recall that the matrix $D_k^{(L)}$ defining the covariance of the moments in the Laguerre case has entries $$\begin{aligned} (D_k^{(L)})_{i,j} = \binom{2i}{i-j} - \binom{2i}{i-j-1} = d_{i-j,i+j} \end{aligned}$$ for $j\ge i$ and 0 else. The asymptotic covariance of the $i$-th and the $j$-th moment is then given by $$\begin{aligned} \left( D_k^{(L)}(D_k^{(L)})^T\right)_{i,j} =& \sum_{r=1}^{\min \{ i,j\} } d_{i-r,i+r} d_{j-r,j+r} \\ =& -d_{i,i} d_{j,j} + \sum_{r=0}^{\min \{ i,j\} } d_{i-r,i+r} d_{j-r,j+r}.\end{aligned}$$ Without loss of generality assume $i\leq j$. Consider a generalised Catalan path to $(0,0)$ in the triangular lattice starting in $(j-r,j+r)$. This path moves $j-r$ times up and $j+r$ times to the left. Replacing each move upwards by a move to the left and vice versa and reversing the order of movements, we obtain a unique path starting in $(i+j,i+j)$ and ending in $(i-r,i+r)$ (moving $j+r$ times up and $j-r$ times to the left). Therefore, $$\begin{aligned} & \sum_{r=0}^i d_{i-r,i+r} d_{j-r,j+r} \\ =& \sum_{r=0}^i \left( \# \left\{ \mbox{Paths from } (i-r,i+r) \mbox{ to } (0,0) \right\} \right) \left( \# \left\{ \mbox{Paths from } (i+j,i+j) \mbox{ to } (i-r,i+r) \right\} \right) .\end{aligned}$$ This last sum gives precisely the number of Catalan paths from $(i+j,i+j)$ to $(0,0)$ since each of those paths crosses the diagonal $\{(i-r,i+r)\}_{0\leq r\leq i}$ exactly once (see Figure 2). In conclusion, $$\begin{aligned} \left( D_k^{(L)}(D_k^{(L)})^T\right)_{i,j} = d_{i+j,i+j} -d_{i,i} d_{j,j} = m_{i+j}(\eta) - m_i(\eta) m_j(\eta) .\end{aligned}$$ \[path2\] (0,0)(17,13) (2,13)(15,13) (3,13)(3,12) (3,12)(15,12) (4,13)(4,11) (4,11)(15,11) (5,13)(5,10) (5,10)(15,10) (6,13)(6,9) (6,9)(15,9) (7,13)(7,8) (7,8)(15,8) (8,13)(8,7) (8,7)(15,7) (9,13)(9,6) (9,6)(15,6) (10,13)(10,5) (10,5)(15,5) (11,13)(11,4) (11,4)(15,4) (12,13)(12,3) (12,3)(15,3) (13,13)(13,2) (13,2)(15,2) (14,13)(14,1) (14,1)(15,1) (15,13)(15,1) (15,1)(15,0.5) (7,8)(12,13) (2,13)(10,11)(12,9)(14,1)(7,8) (2,13)(4,13)(4,12)(6,12)(6,11)(9,11)(9,9)(12,9) (2,13)(7,13)(7,12)(10,12)(10,11) (10,11)(10,8)(12,8)(12,5)(13,5)(13,3)(14,3)(14,1) (13,0.4)[$(i+j,i+j)$]{} (6.5,7.4)[$(i,i)$]{} (2,12.4)[$(0,0)$]{} Now consider the remaining Gaussian ensemble, where the covariances are given by $$\begin{aligned} \left( D_k^{(G)}(D_k^{(G)})^T\right)_{i,j} = \sum_{r=1}^{\min \{i,j\}} d_{\frac{i-r}{2},\frac{i+r}{2}} \mathbbm{1}_{\{ i+r \mbox{ even}\} } \cdot d_{\frac{j-r}{2},\frac{j+r}{2}} \mathbbm{1}_{\{ j+r \mbox{ even}\} } .\end{aligned}$$ The even moments of the semicircle distribution $\rho$ are given by $m_{2k}(\rho) = d_{k,k}$ and the odd moments vanish. Suppose first that $i$ and $j$ are even, so we obtain from the calculations in the Laguerre case: $$\begin{aligned} \left( D_k^{(G)}(D_k^{(G)})^T\right)_{i,j} =& - d_{\frac{i}{2},\frac{i}{2}} d_{\frac{j}{2},\frac{j}{2}} + \sum_{r=0}^{\min \{i,j\}} d_{\frac{i-r}{2},\frac{i+r}{2}} d_{\frac{j-r}{2},\frac{j+r}{2}} \mathbbm{1}_{\{ r \mbox{ even}\} } \\ =& - d_{\frac{i}{2},\frac{i}{2}} d_{\frac{j}{2},\frac{j}{2}} + \sum_{r=0}^{\min \{i,j\}/2} d_{\frac{i}{2}-r,\frac{i}{2}+r} d_{\frac{j}{2}-r,\frac{j}{2}+r} \\ =& -d_{\frac{i}{2},\frac{i}{2}} d_{\frac{j}{2},\frac{j}{2}} + d_{\frac{i+j}{2},\frac{i+j}{2}} \\ =& m_{i+j}(\rho) - m_i(\rho) m_j(\rho) .\end{aligned}$$ If one of the indices $i,j$ is even and the other one is odd, all summands vanish and in this case also $m_i(\rho) m_j(\rho) - m_{i+j}(\rho)=0$. For the last case, assume $i,j$ are both odd, then $$\begin{aligned} \left( D_k^{(G)}(D_k^{(G)})^T\right)_{i,j} =& \sum_{r=0}^{\min \{i,j\}} d_{\frac{i-r}{2},\frac{i+r}{2}} d_{\frac{j-r}{2},\frac{j+r}{2}} \mathbbm{1}_{\{ r \mbox{ odd}\} } \\ =& \sum_{r=0}^{\min \{i-1,j-1\}/2} d_{\frac{i-1}{2}-r,\frac{i+1}{2}+r} d_{\frac{j-1}{2}-r,\frac{j+1}{2}+r} \\ =& \sum_{r=0}^{\min \{i^,j^\}} d_{i^-r,i^+1+r} d_{j^-r,j^+1+r}\end{aligned}$$ with $i^= \frac{i-1}{2}$ and $j^=\frac{j-1}{2}$. Now assume that $i \leq j$ and apply the path transformation described above, where a generalised Catalan path starting in $(j^-r,j^+1+r)$ is uniquely identified with a path starting in $(i^+j^+1,i^+j^+1)$ and ending in $(i^-r,i^+1+r)$. Therefore, $$\begin{aligned} & \sum_{r=0}^{\min \{i^,j^\}} d_{i^-r,i^+1+r} d_{j^-r,j^+1+r} \\ =& \sum_{r=0}^{i^} \left( \# \left\{ \mbox{Paths from } (i^-r,i^+1+r) \mbox{ to } (0,0) \right\} \right) \\ & \qquad \cdot \left( \# \left\{ \mbox{Paths from } (i^+j^+1,i^+j^+1) \mbox{ to } (i^-r,i^+1+r) \right\} \right) \\ =& d_{i^+j^+1,i^+j^+1}\end{aligned}$$ The last identity follows since each path crosses the diagonal $\{ (i^-r,i^+1+r) \}_{r\geq0}$ exactly once. We conclude, $$\begin{aligned} \left( D_k^{(G)}(D_k^{(G)})^T\right)_{i,j} = d_{\frac{i+j}{2},\frac{i+j}{2}} = m_{i+j}(\rho) = m_{i+j}(\rho) - m_i(\rho) m_j(\rho) .\end{aligned}$$ This proves the third case of the Lemma. Proof of Theorem \[haupt\]:\ Let $\mathbb{M}_0$ be the set of all moment sequences $m= m(\mu)= (m_1(\mu),m_2(\mu),\dots)$ of signed measures $\mu \in \mathcal{M}_0$ with compact support, endowed with the topology of pointwise convergence and the Borel $\sigma$-algebra. Theorem \[step1\] yields a MDP for the projection $m^{(k)}$ of $m$ onto the first $k$ coordinates. Denote by $D$ the infinite dimensional triangular matrix with upper left blocks given by $D_k$ as in Theorem \[step1\]. By the Dawson-Gärtner Theorem (see [@demzei1998]) the random sequence $$\begin{aligned} \sqrt{a_n^{-1}n}(m(\mu_n)-m(\sigma)) = m\big( \sqrt{a_n^{-1}n}(\mu_n - \sigma)\big) \in \mathbb{M}_0\end{aligned}$$ satisfies a moderate deviation principle with speed $a_n$ and good rate function $$\begin{aligned} \mathcal{I}_m(m) = \sup_{k\in \mathbb{N}} \mathcal{I}_k(m^{(k)}) = \frac{\beta}{4} \|\|D^{-1} m\|\|^2 .\end{aligned}$$ Note that since $D$ is a lower triangular matrix, $D^{-1} m$ is well-defined (with possible entries $\infty$). By the continuity of the mapping $m(\mu) \mapsto \mu$ from $\mathbb{M}_0$ to $\mathcal{M}_0$, the contraction principle yields a MDP for $\sqrt{a_n^{-1}n}(\mu_n - \sigma)$ with rate function $$\begin{aligned} \mathcal{I}(\mu) = \frac{\beta}{4} \|\|D^{-1} m(\mu)\|\|^2 .\end{aligned}$$ To prove the integral representation of the rate function, we first apply an idea presented in the paper of [@chakemstu1993]. Define $X_i: x\mapsto x^i-m_i(\sigma)$ for $i\in \mathbb{N}$ as a random variable on $(\mathbb{R},\mathcal{B}(\mathbb{R}),\sigma)$, then $$\begin{aligned} \mathbb{E}[X_i] = 0,\qquad \operatorname{Cov}(X_i,X_j) = m_{i+j}(\sigma) -m_i(\sigma)m_j(\sigma) .\end{aligned}$$ That is, Lemma \[step2\] implies that the expectation and covariance of $X = (X_1,\dots ,X_k)^T$ coincides with the asymptotic expectation and covariance of $\sqrt{\frac{\beta n}{2}} (m^{(k)}(\mu_n) - m^{(k)}(\sigma))$. But then the entries of $D_k^{-1}X$ are uncorrelated with variance 1, which means that $D_k^{-1}X$ has as entries the first $k$ polynomials $(p_1(x),\dots ,p_k(x))$ orthonormal with respect to $\sigma$. Therefore, the entries in the $i$-th row of $D_k^{-1}$ are given by the coefficients of the $i$-th orthonormal polynomial, except for the constant term (since $X_i$ is normalised). We can conclude that $$\begin{aligned} D^{-1} m(\mu) & = \left( \int p_1(x)-p_1(0) \mu(dx), \int p_2(x)-p_2(0) \mu(dx), \dots \right)^T \\ & = \left( \int p_1(x) \mu(dx), \int p_2(x) \mu(dx), \dots \right)^T\end{aligned}$$ where the last identity follows from $\mu(\mathbb{R})=0$. The rate function is then $$\begin{aligned} \mathcal{I}(\mu) = \frac{\beta}{4} \sum_{k=1}^\infty \left( \int p_k(x) \mu(dx) \right)^2 \end{aligned}$$ and the assertion of the theorem follows from Lemma \[rate\]. \[rate\] Let $\sigma$ be a probability measure with infinite and compact support and let $p_0,p_1,\dots$ denote the polynomials orthonormal with respect to $\sigma$. Then for any $\mu \in \mathcal{M}_0$, we have the equality of rate functions, $$\begin{aligned} \sum_{k=0}^\infty \left( \int p_k(x) \mu(dx) \right)^2 = \int \left(\tfrac{\partial \mu}{\partial \sigma} (x)\right)^2 \sigma (dx)\end{aligned}$$ and both sides equal $\infty$ if $\mu$ is not absolutely continuous with respect to $\sigma$. First, note that if $d\mu = h\cdot d\sigma$ is absolutely continuous with respect to $\sigma $ with signed density $h$, Lemma \[rate\] follows from Parseval’s identity. We have to show that the left hand side equals $\infty$ as soon as $\mu$ is not absolutely continuous with respect to $\sigma$. Suppose $\sigma(A)=0$ for a measurable set $A$, but $\mu(A)\neq 0$. We can choose polynomials $h_n$ of degree $n$ such that $\int x^k h_n(x)\sigma(dx)=\int x^k \mu(dx)$ for all $k\leq n$. Indeed, the infinite support of $\sigma$ implies that the Hankel-matrices $(m_{i+j}(\sigma))_{0\leq i,j\leq n}$ are positive definite ([@shotam1963]) and such polynomials exist. Then the measures $d\mu_n = h_n\cdot d\sigma $ have the first $n$ moments equal to the first $n$ moments of $\mu$ and therefore converge to $\mu$ in the moment topology. The support of $\mu_n$ is given by the support of $\sigma$ and is in particular uniformly bounded. Since $\int p_kd\mu$ depends only on the first $k$ moments of $\mu$, $$\begin{aligned} \sum_{k=0}^\infty \left( \int p_k(x) \mu(dx) \right)^2 = \lim_{n\to \infty} \sum_{k=0}^\infty \left( \int p_k(x) \mu_n(dx) \right)^2 = \lim_{n\to \infty} \int h_n(x)^2 \sigma (dx)\end{aligned}$$ Assume that there exists a subsequence of measures $\mu_{n_k}$ along which the integrals on the right hand side are uniformly bounded, i.e. $\sup_k \int h_{n_k}(x)^2 \sigma (dx)<\infty$. Then the total variation $$\begin{aligned} \operatorname{TV} (\mu_{n_k}) = \sup_{f: \|f\|\leq 1} \int f d\mu_{n_k} \leq \sup_{f: \|f\|\leq 1} \left( \int (fh_{n_k})^2 d\sigma \right)^{1/2} \leq \left( \int h_{n_k}^2 d\sigma \right)^{1/2}\end{aligned}$$ is uniformly bounded as well. Let $f$ be a continuous function and for $\varepsilon>0$, let $p_f$ a polynomial approximating $f$ in the sup-norm on the compact support of $\sigma$ and $\mu$, $$\begin{aligned} \sup_{x\in \operatorname{supp}(\sigma)\cup \operatorname{supp}(\mu)} \|f(x)-p_f(x)\| <\varepsilon ,\end{aligned}$$ then we can bound $$\begin{aligned} \left\| \int fd\mu_{n_k} - \int fd\mu \right\| \leq \left\| \int p_fd\mu_{n_k} - \int p_fd\mu \right\| + \varepsilon (\operatorname{TV}(\mu_{n_k}) + \operatorname{TV}(\mu) ) .\end{aligned}$$ Since all moments converge, the right hand side can be made arbitrarily small as $n\to \infty$ and we get $\mu_{n_k} \xrightarrow[n \rightarrow \infty ]{} \mu$ not only in the moment topology but also weakly. Moreover, the bounded total variation and Helly’s selection theorem yield (possibly taking another subsequence) the weak convergence of the positive parts $\mu_{n_k}^+$ to the positive part $\mu^+$, when $\mu=\mu^+-\mu^-$ is the Hahn-Jordan decomposition of a measure $\mu$ into the nonnegative measures $\mu^+,\mu^-$. Without loss of generality, assume that the set $A$ is a subset of the support of $\mu^+$. By the (outer) regularity of $\sigma$, for any $\varepsilon>0$ there is an open set $A_\varepsilon$ containing $A$ such that $\sigma(A_\varepsilon)<\varepsilon$. The Cauchy-Schwarz inequality implies $$\begin{aligned} \int h_{n_k}^2 d\sigma \geq \sigma(A_\varepsilon)^{-1} \left( \int_{A_\varepsilon} \|h_{n_k}\| d\sigma \right)^2 \geq \varepsilon^{-1} \mu_{n_k}^+(A_\varepsilon)^2 \end{aligned}$$ Since $A_\varepsilon$ is open, we have $\liminf_{k\to \infty} \mu_{n_k}^+(A_\varepsilon)\geq \mu^+(A_\varepsilon)>0$ by the Portmanteau-Theorem and the weak convergence of $\mu_{n_k}^+$. Letting $\varepsilon \to 0$, this contradicts the boundedness of $\int h_{n_k}(x)^2 \sigma (dx)$ and we get $\lim_{n\to \infty} \int h_n^2 d\sigma =\infty$. Acknowledgments. The author would like to thank Fabrice Gamboa and Alain Rouault for several helpful discussions.	{ "pile_set_name": "ArXiv" }
--- author: - \| Toshiya Hachisuka\ The Unviersity of Tokyo bibliography: - 'egbibsample.bib' ---	{ "pile_set_name": "ArXiv" }
--- abstract: 'We propose a type II upgrade of type I see-saw models leading to new classes of models where partially degenerate neutrinos are as natural as hierarchical ones. The additional type II contribution to the neutrino mass matrix, which determines the neutrino mass scale, is forced to be proportional to the unit matrix by a SO(3) flavour symmetry. The type I see-saw part of the neutrino mass matrix, which controls the mass squared differences and mixing angles, may be governed by sequential right-handed neutrino dominance and a natural alignment for the SO(3)-breaking vacuum. We focus on classes of models with bi-large mixing originating from the neutrino mass matrix although we also briefly discuss other classes of models where large mixing stems from the charged lepton mass matrix. We study renormalization group corrections to the neutrino mass squared differences and mixings and find that the low energy values do not depend sensitively on the high energy values for partially degenerate neutrinos with a mass scale up to about 0.15 eV. Our scenario predicts the effective mass for neutrinoless double beta decay to be approximately equal to the neutrino mass scale and therefore neutrinoless double beta decay will be observable if the neutrino mass spectrum is partially degenerate. We also find that all observable CP phases as well as $\theta_{13}$ become small as the neutrino mass scale increases.' --- SHEP/0403 [ From Hierarchical to Partially Degenerate Neutrinos\ via Type II Upgrade of Type I See-Saw Models ]{}\ S. Antusch[^1], S. F. King[^2]\ [Department of Physics and Astronomy, University of Southampton,\ Southampton, SO17 1BJ, U.K. ]{} Introduction ============ The observation of flavour conversions of neutrinos together with their interpretation by neutrino oscillations has brought crucial new information about fermion masses and mixings. Neutrinos are massive, with very small measured mass squared differences, and contrary to the quark sector, there is large flavour mixing among the leptons. The present 3$\sigma$ ranges for the parameters are $\theta_{12} \in [27.6^\circ,36.3^\circ]$, $\Delta m^2_\mathrm{sol}:= m_2^2 - m_1^2 \in [5.4\cdot 10^{-5}\:\mbox{eV}^2,9.5\cdot 10^{-5}\:\mbox{eV}^2]$ , $\theta_{23} \in [32.2^\circ,51.4^\circ]$, $\|\Delta m^2_\mathrm{atm}\|:= \|m_3^2 - m_1^2\| \in [1.4\cdot 10^{-3}\:\mbox{eV}^2,3.7\cdot 10^{-3}\:\mbox{eV}^2]$ and for $\theta_{13}$ the current upper bound which mainly stems from the CHOOZ data [@Apollonio:1999ae] is $\theta_{13} \lesssim 15^\circ$. The values have been taken from the global analysis [@Maltoni:2003da] which includes the the Super-Kamiokande atmospheric data [@Toshito:2001dk], the KamLAND results [@Eguchi:2002dm] and recent SNO salt results [@Ahmed:2003kj]. One of the most interesting missing piece of information is the neutrino mass scale. At present, the most stringent bounds are $m_i < 0.23$ eV from WMAP [@Spergel:2003cb] and $\Braket{m_\nu}\lesssim 0.35\,\mathrm{eV}$, with some uncertainty due to nuclear matrix elements, from neutrino-less double beta ($00\nu\beta$) decay experiments [@Klapdor-Kleingrothaus:2000sn; @Aalseth:2002rf]. The latter search for an effective mass defined by $\Braket{m_\nu}= \left\| \sum_i (U_\mathrm{MNS})_{1i}^2 \, m_i \right\| $ and are exclusively sensitive to Majorana masses. Future experiments which are under consideration at present might increase the sensitivity to $\Braket{m_\nu}$ by more than an order of magnitude. The neutrino mass spectrum for the mass range accessible to this next round of $0\nu\beta\beta $ decay searches shows at least a partial degeneracy. It is therefore interesting to investigate theoretical scenarios which could account for such neutrino mass schemes. The most promising scenarios for giving masses to neutrinos use a version of the see-saw mechanism [@Yanagida:1980; @Glashow:1979vf; @Gell-Mann:1980vs; @Mohapatra:1980ia], which provides a convincing explanation for their smallness. Models for strongly degenerate neutrinos have been considered e.g. in [@Caldwell:1993kn; @Bamert:1994vc; @Lee:1994qx; @Ioannisian:1994nx; @Joshipura:1994jy; @Joshipura:1995ax; @Ghosal:1997vs; @Carone:1998bb; @Ma:1998db; @Ray:1998eq; @Lazarides:1998jt; @Wetterich:1998vh; @Wu:1998if; @Ghosal:1999jb; @Barbieri:1999km; @Wu:1999mu; @Wu:1999yz; @Ma:2001dn; @Babu:2002dz; @Patgiri:2003ah; @Mohapatra:2003tw]. In most of them, the degeneracy is achieved by discrete symmetries. The non-Abelian flavour symmetry group $\SO(3)$ in connection with degenerate neutrinos has e.g. been considered in [@Wu:1998if; @Carone:1998bb; @Ma:1998db; @Wetterich:1998vh; @Barbieri:1999km]. It has turned out that using the type I see-saw mechanism in order to explain the smallness of neutrino masses, it seems to be difficult, if not impossible, to obtain a nearly degenerate neutrino mass spectrum in a natural way. In order to find natural explanations for partially degenerate neutrino masses, it is therefore promising to consider the type II see-saw mechanism (see e.g. [@Lazarides:1980nt; @Mohapatra:1981yp; @Wetterich:1981bx; @Ma:1998dx]), as has been argued for example in [@Caldwell:1993kn; @Joshipura:1995ax]. In this work, we propose a type II upgrade of type I see-saw models leading to new classes of models where partially degenerate neutrinos are as natural as hierarchical ones. This is achieved by a SO(3) flavour symmetry, which forces the additional type II mass term to be proportional to the unit matrix in leading order. The addition of a type II unit matrix contribution to the type I neutrino mass matrix with a particular phase structure turns out to enable the neutrino mass scale to be increased almost arbitrarily, while leaving the mixing angles approximately unchanged. In this approach the type I see-saw part of the neutrino mass matrix, which controls the mass squared differences and mixing angles, may be governed by sequential right-handed neutrino dominance and a natural alignment for the SO(3)-breaking vacuum. We focus on classes of models with bi-large mixing originating from the neutrino mass matrix although we also briefly discuss other classes of models where large mixing stems from the charged lepton mass matrix. We also study renormalization group corrections to the mass squared differences and mixings and find that the low energy values do not depend sensitively on the high energy values for partially degenerate neutrinos with a mass scale up to about 0.15 eV. This framework predicts the effective mass for neutrinoless double beta decay to be approximately equal to the neutrino mass scale and therefore neutrinoless double beta decay will be observable if the neutrino mass spectrum is partially degenerate. We also find that all observable CP phases become small as the neutrino mass scale increases. The layout of the paper is as follows: After giving a motivation in section 2, we outline how naturally small neutrino masses could emerge from type I and type II see-saw mechanisms in section 3. In section 4 we discuss the type II see-saw scenario with spontaneously broken SO(3) flavour symmetry and analyze the consequences for the ingredients of the type II see-saw formula. In section 5 we consider a real alignment mechanism for the SO(3)-breaking vacuum expectation values (vevs) and show how it leads to classes of models where the observed bi-large neutrino mixing can naturally be obtained. In section 6 we focus on models with sequential right-handed neutrino dominance [@King:1999mb; @King:2002nf] for the type I part of the neutrino mass matrix. We extract the mixing angles, masses and CP phases analytically and discuss the predictions for these parameters from our scenario. In section 7 we study renormalization group corrections to the neutrino mass squared differences and mixings. In section 8 we analyze the predictions for the effective mass for neutrinoless double beta decay. Section 9 contains a discussion and our conclusions. Motivation ========== For a given mass of the lightest neutrino and e.g. a normal scheme for the neutrino masses $m_1 < m_2 < m_3$, the remaining two masses can be calculated from the requirement that the experimentally measured mass squared differences are produced. Figure \[fig:NuMasses\] shows the neutrino mass eigenvalues as a function of the mass of the lightest neutrino. For a lightest neutrino heavier than about $0.02$ eV, the mass eigenvalues of the neutrinos are of the same order. We will refer to the neutrinos as partially degenerate, if the mass of the lightest neutrino is roughly in the range $[0.02,0.15]$ eV, where $m_1$ and $m_2$ are nearly degenerate. This is below what is usually called quasi-degenerate where the masses of all neutrinos are approximately degenerate. This mass range is particularly interesting, since it is not disfavored by unnaturally large radiative corrections and it might be accessible to future $0\nu\beta\beta $ decay searches. [$\vcenter{\hbox{\includegraphics[scale=0.9]{NeutrinoMasses.eps}}}$]{} With three left-handed neutrinos contained in the lepton doublets and three right-handed neutrinos which are singlets under $G_{321}:= \SU (3)_\mathrm{C}\times \SU (2)_\mathrm{L} \times \U (1)_\mathrm{Y}$, the general neutrino mass matrix is given by $$\begin{aligned} \mathscr{L}_{M_\nu} &=& - \frac{1}{2}\, \left( \begin{array}{c} \overline{\nu_{\mathrm{L}}^{f}} \\[1mm] \overline{\nu^{\ChargeC i}_\mathrm{R}} \end{array} \right)^T \left( \begin{array}{cc} (m^{\mathrm{II}}_{\mathrm{LL}})_{fg} \vphantom{\nu_{\mathrm{L}}^{\ChargeC g}}& (m_{\mathrm{LR}})_{fj}\\[1mm] (m_{\mathrm{LR}}^T)_{ig}\vphantom{\nu_{\mathrm{L}}^{\ChargeC g}}& (M_{\mathrm{RR}})_{ij} \end{array} \right) \, \left( \begin{array}{c} \nu_{\mathrm{L}}^{\ChargeC g} \\[1mm] {\nu^{ j}_\mathrm{R}} \end{array} \right)+\text{h.c.}\; .\end{aligned}$$ Under the assumption that the mass eigenvalues $M_{\mathrm{R}i}$ of $M_{\mathrm{RR}}$ are very large compared to the components of $m^{\mathrm{II}}_{\mathrm{LL}}$ and $m_{\mathrm{LR}}$, the mass matrix can approximately be diagonalized yielding $$\begin{aligned} \mathscr{L}_{M_\nu} &\approx& - \frac{1}{2}\, \left( \begin{array}{c} \overline{\nu{}'{}_{\mathrm{L}}^{f}} \\[1mm] \overline{\nu{}'{}^{\ChargeC i}_\mathrm{R}} \end{array} \right)^T \left( \begin{array}{cc} (m^{\nu}_{\mathrm{LL}})_{fg} \vphantom{\nu_{\mathrm{L}}^{\ChargeC g}}& 0\\[1mm] 0\vphantom{\nu_{\mathrm{L}}^{\ChargeC g}}& (M_{\mathrm{RR}})_{ij} \end{array} \right) \, \left( \begin{array}{c} \nu{}'{}_{\mathrm{L}}^{\ChargeC g} \\[1mm] {\nu{}'{}^{ j}_\mathrm{R}} \end{array} \right)+\text{h.c.}\; ,\end{aligned}$$ where, neglecting $\mathscr{O}(M_{\mathrm{R}i}^{-1})$-terms, $\nu{}{'}_{\mathrm{L}}^f\approx\nu_{\mathrm{L}}^f$ and $\nu{}{'}{}^{\ChargeC i}_\mathrm{R}\approx \nu^{\ChargeC i}_\mathrm{R}$. The Majorana mass matrix $m^\nu_{\mathrm{LL}}$ for the light left-handed neutrinos is given by $$\begin{aligned} \label{eq:TypIIMassMatrix} m^\nu_{\mathrm{LL}} \approx m^{\mathrm{II}}_{\mathrm{LL}} + m^{\mathrm{I}}_{\mathrm{LL}}\end{aligned}$$ with $$\begin{aligned} m^{\mathrm{I}}_{\mathrm{LL}} := - m_{\mathrm{LR}} \,M^{-1}_{\mathrm{RR}}\,m^T_{\mathrm{LR}}\; .\end{aligned}$$ The suppression by $M_{\mathrm{RR}}^{-1}$ provides a natural explanation for the smallness of neutrino masses from $m^{\mathrm{I}}_{\mathrm{LL}}$. This is referred to as the type I see-saw mechanism [@Yanagida:1980; @Glashow:1979vf; @Gell-Mann:1980vs; @Mohapatra:1980ia]. As we will sketch in section \[sec:SeeSaw\], the direct mass term $m^{\mathrm{II}}_{\mathrm{LL}}$ can also provide a naturally small contribution to the light neutrino masses if it stems e.g. from a see-saw suppressed induced vev. We will refer to the general case, where both possibilities are allowed, as the II see-saw mechanism [@Lazarides:1980nt; @Mohapatra:1981yp; @Wetterich:1981bx; @Ma:1998dx]. In type I see-saw models, it seems to be difficult, if not impossible, to obtain a partially degenerate or quasi-degenerate neutrino mass spectrum in a natural way, whereas hierarchical masses seem to be natural. For a review on neutrino mass models, see e.g. [@King:2003jb]. The direct mass term in type II models on the other hand has the potential to provide a natural way for generating neutrino masses with a partial degeneracy. Imagine for example that by symmetry, the direct mass term is forced to be proportional to the unit matrix in flavour space, $m^{\mathrm{II}}_{\mathrm{LL}} = m^{\mathrm{II}} \,\mathbbm{1}$. We will realize such a direct mass term, which gives a common mass to all the neutrinos, in section \[sec:SO(3)\] via SO(3) flavour symmetry. The type II formula of equation (\[eq:TypIIMassMatrix\]) is then realized by $$\begin{aligned} m^\nu_{\mathrm{LL}} \approx m^{\mathrm{II}}\, \left(\begin{array}{ccc} 1&0&0\\ 0&1&0\\ 0&0&1 \end{array}\right) + \left(\begin{array}{ccc} (m^{\mathrm{I}}_{\mathrm{LL}})_{11} &(m^{\mathrm{I}}_{\mathrm{LL}})_{12} & (m^{\mathrm{I}}_{\mathrm{LL}})_{13}\\ (m^{\mathrm{I}}_{\mathrm{LL}})_{21} &(m^{\mathrm{I}}_{\mathrm{LL}})_{22} &(m^{\mathrm{I}}_{\mathrm{LL}})_{23}\\ (m^{\mathrm{I}}_{\mathrm{LL}})_{31} &(m^{\mathrm{I}}_{\mathrm{LL}})_{32} &(m^{\mathrm{I}}_{\mathrm{LL}})_{33} \end{array}\right)\end{aligned}$$ and the direct mass term $m^{\mathrm{II}}\, \mathbbm{1}$ naturally allows for partially degenerate neutrinos. The neutrino mass splittings and mixing angles in this scenario are mainly controlled by the type I see-saw contribution $m_{\mathrm{LL}}^{\mathrm{I}} $. To analyze the effect of the type II contribution $m_{\mathrm{LL}}^{\mathrm{II}}$, we consider the diagonalization of $m_{\mathrm{LL}}^{\mathrm{I}}$ by a unitary transformation $(m_{\mathrm{LL}}^{\mathrm{I}})_{\mathrm{diag}} = V m_{\mathrm{LL}}^{\mathrm{I}} V^T $. If we assume for the moment that the type I see-saw mass matrix $m_{\mathrm{LL}}^{\mathrm{I}}$ is real, which implies that $V$ is an orthogonal matrix, we obtain $$\begin{aligned} \label{eq:TypIISeeSawFormulaUnitMatrix} (m^\nu_{\mathrm{LL}})_{\mathrm{diag}} \;= \; m^{\mathrm{II}}\, V V^T + V m_{\mathrm{LL}}^{\mathrm{I}} V^T \;=\; m^{\mathrm{II}}\, \mathbbm{1} + (m_{\mathrm{LL}}^{\mathrm{I}})_{\mathrm{diag}} \; . \end{aligned}$$ The additional direct mass term leaves the predictions for the mixings from the type I see-saw contribution unchanged in this case. This allows to transform many type I see-saw models for hierarchical neutrino masses into type II see-saw models for partially degenerate or quasi-degenerate neutrino masses while maintaining the predictions for the mixing angles. Obviously, in the general complex case, it is no longer that simple since for a unitary matrix $V V^T \not= \mathbbm{1}$ and the phases will have impact on the predictions for the mixings. We will return to this issue in section \[sec:Typw2SeqRhdNuDom\] where we will see that e.g. in some classes of models with sequential right-handed neutrino dominance (RHND) [@King:1999mb; @King:2002nf] for the type I contribution to the neutrino mass matrix, the known techniques and mechanisms for explaining the bi-large lepton mixings can be directly applied also in the presence of CP phases. Type I and Type II See-Saw Mechanisms {#sec:SeeSaw} ===================================== We now outline how naturally small neutrino masses could emerge from minimal realizations of type I and type II see-saw mechanisms. Let us consider a minimal extension of the MSSM in order to allow for neutrino masses. We discuss the supersymmetric case here since the modifications in oder to obtain the non-supersymmetric version are straightforward. Dirac masses for the neutrinos can be achieved by adding chiral superfields ${\hat{\nu}}^{\ChargeC \Nf}$ $(\Nf\in \{1,2,3\})$ in the representation $(\boldsymbol{1},\boldsymbol{1},0)$ of $G_{321}$ to the particle content, which contain the right-handed neutrinos $\nu^{\Nf}_\mathrm{R}$ as fermionic components. If not protected by symmetry, these singlets are expected to obtain large masses $M_{\mathrm{R}i}$, associated with the scale of lepton number breaking. Lepton masses now arise from the superpotential $$\begin{aligned} \label{eq:SuperpotentialInTheMSSM} \mathcal{W}_{\ell} & = & - \,(Y_e)_{gf}({\hat{L}}^{g} \cdot \,{\hat{H}}_d )\,{\hat{e}}^{\chargec f} +(Y_\nu)_{f \Ng}({\hat{L}}^{f}\cdot {\hat{H}}_u)\, {\hat{\nu}}^{\chargec \Ng} + \frac{1}{2} {\hat{\nu}}^{\chargec \Nf} (M_\mathrm{RR})_{\Nf \Ng} {\hat{\nu}}^{\chargec \Ng} \; , \end{aligned}$$ where the dot indicates the $\SU (2)_\mathrm{L}$ invariant product, i.e. $({\hat{L}}^{\Ng}\cdot {\hat{H}}_u) := {\hat{L}}_a^{\Ng}(i\tau_2)^{ab} ({\hat{H}}_u)_b$ with $\tau_A$ $(A\in \{1,2,3\})$ being the Pauli matrices. The superfields ${\hat{H}}_u$ and ${\hat{H}}_d$ contain the Higgs fields which also give masses to the up-type and down-type quarks respectively. ${\hat{L}}$ contains the lepton doublets and ${\hat{e}}^{\chargec}$ the charged leptons. This yields naturally small neutrino masses by the type I see-saw relation $$\begin{aligned} \label{eq:TypISeeSawFormula} m^{\mathrm{I}}_\mathrm{LL} = - \, v^2_u\, Y_\nu \, M_\mathrm{RR}^{-1}\, Y_\nu^T \; .\end{aligned}$$ In many cases however, the left-handed neutrinos may also obtain a naturally small direct mass term. This happens for example if $\SU (2)_\mathrm{L}$-triplet Higgs superfields ${\hat{\Delta}}$ and $\bar{{\hat{\Delta}}}$ are added to the particle spectrum which have weak hypercharge $q_\mathrm{Y}=+1$ and $q_\mathrm{Y}=-1$, respectively. The representations of the chiral superfields involved in this minimal setup are given in table \[tab:QuantumNumbersOfNuMSSM\]. $\vphantom{\sqrt{\big\|}}$Field ${\hat{H}}_d$ ${\hat{H}}_u$ ${\hat{\Delta}}$ $\bar{{\hat{\Delta}}}$ ${\hat{L}}^{f}$ ${{\hat{e}}}^{\ChargeC f}$ ${\hat{\nu}}^{\ChargeC f}$ ------------------------------------------------------ ------------------ ------------------ ------------------ ------------------------ ------------------ ---------------------------- ---------------------------- $\vphantom{\sqrt{\big\|}^C}\mathrm{SU}(3)_\mathrm{C}$ $\boldsymbol{1}$ $\boldsymbol{1}$ $\boldsymbol{1}$ $\boldsymbol{1}$ $\boldsymbol{1}$ $\boldsymbol{1}$ $\boldsymbol{1}$ $\vphantom{\sqrt{\big\|}}\mathrm{SU}(2)_\mathrm{L}$ $\boldsymbol{2}$ $\boldsymbol{2}$ $\boldsymbol{3}$ $\boldsymbol{3}$ $\boldsymbol{2}$ $\boldsymbol{1}$ $\boldsymbol{1}$ $\vphantom{\sqrt{\big\|}} q_\mathrm{Y}$ $-\tfrac{1}{2}$ $+\tfrac{1}{2}$ $1$ $-1$ $-\tfrac{1}{2}$ $+1$ $0$ : \[tab:QuantumNumbersOfNuMSSM\] Representations under $G_{321}$ of the chiral superfields involved in the generation of lepton masses in the MSSM extended by three singlet superfields which contain the right-handed neutrinos and by two $\mathrm{SU}(2)_\mathrm{L}$-triplets. Only the superfield ${\hat{\Delta}}$ contributes to the generation of fermion masses via the coupling $$\begin{aligned} \label{eq:SuperpotentialDelta} \mathcal{W}_\Delta & = & \frac{1}{2} \,(Y^\dagger_\Delta)_{fg} \,{\hat{L}}^T{}^f\,i\tau_2\, {\hat{\Delta}} \,{\hat{L}}^g\; \end{aligned}$$ to the lepton doublets. We have written the $\mathrm{SU}(2)_\mathrm{L}$-triplets as traceless $2\times2$-matrices $$\begin{aligned} {\hat{\Delta}}= \left(\begin{array}{cc} {\hat{\Delta}}^{+}& {\hat{\Delta}}^{++} \\ {\hat{\Delta}}^{0} &-{\hat{\Delta}}^{+} \end{array}\right) \:\mbox{and}\;\; \bar{{\hat{\Delta}}}= \left(\begin{array}{cc} \bar{{\hat{\Delta}}}^{+}& \bar{{\hat{\Delta}}}^{++} \\ \bar{{\hat{\Delta}}}^{0} &-\bar{{\hat{\Delta}}}^{+} \end{array}\right) . \end{aligned}$$ A vev $v_\Delta=\<{\Delta}^0\>$ of the neutral component of the scalar field contained in ${\hat{\Delta}}$ gives a direct mass for the left-handed neutrinos. We choose $v_\Delta$ to be real and positive by a proper phase choice for ${\hat{\Delta}}$. In order to estimate the natural size of $v_\Delta$, we consider the Higgs potential from the part $$\begin{aligned} \mathcal{W}_{H} = M_\Delta \operatorname{Tr}({\hat{\Delta}} \,\bar{{\hat{\Delta}}}) + \lambda_u \, {\hat{H}}^T_u\,i\tau_2\, \bar{{\hat{\Delta}}}\,{\hat{H}}_u +\lambda_d\, {\hat{H}}^T_d\,i\tau_2\, {\hat{\Delta}}\,{\hat{H}}_d + \mu \,({\hat{H}}_d\cdot{\hat{H}}_u) \end{aligned}$$ of the superpotential. Using the expansion of the superfields $ {\hat{H}}_u = H_u + \sqrt{2}\, \theta \Tilde{H}_u + \theta\theta\,\,F_{H_u} $ and $ {\hat{\Delta}} = \Delta + \sqrt{2}\, \theta \Tilde{\Delta} + \theta\theta\,\,F_{\Delta} $, we see that the scalar potential from $\|F_{\bar{\Delta}}\|^2$ contains the terms $$\begin{aligned} \mathcal{V} = M_\Delta \lambda_u\, H^T_u \,i\tau_2\, \Delta^* H_u + M^2_\Delta \operatorname{Tr}(\Delta^* \Delta) \;+\;\text{h.c.}\; .\end{aligned}$$ After EW symmetry breaking, this results in a tadpole which forces $\Delta$ to get an induced vev of the order $$\begin{aligned} v_\Delta \approx \frac{\lambda_u v_u^2}{M_\Delta}\; .\end{aligned}$$ Analogously, the neutral component of the superfield $\bar{{\hat{\Delta}}}$ obtains a small vev as well, however it is not relevant here since $\bar{{\hat{\Delta}}}$ does not couple to the fermions. If $M_\Delta$ is large, say of the order of one of the eigenvalues of $M_\mathrm{RR}$, this leads to another naturally small contribution $$\begin{aligned} m^{\mathrm{II}}_{\mathrm{LL}} = Y_{\Delta} v_\Delta\end{aligned}$$ to the neutrino mass matrix, which is now given by the type II see-saw formula $$\begin{aligned} \label{eq:TypIISeeSawFormula} m^\nu_\mathrm{LL} = m^{\mathrm{II}}_{\mathrm{LL}} + m^{\mathrm{I}}_{\mathrm{LL}} = Y_{\Delta} v_\Delta - v^2_u \, Y_\nu \, M_\mathrm{RR}^{-1}\, Y_\nu^T \; . \end{aligned}$$ The contributions to neutrino masses in the considered minimal type II scenario are illustrated in figure \[fig:TypeIIDiagrams\]. Note that we do not require left-right symmetry throughout this paper. Additional triplets like the ones used here to provide a minimal example for a type II see-saw mechanism typically appear in models with a left-right symmetric particle content like minimal left-right symmetric models [@Pati:1974yy; @Mohapatra:1975gc; @Senjanovic:1975rk], Pati-Salam models [@Pati:1973uk] or $\SO (10)$-GUTs [@GeorgiSO10; @Fritzsch:1975nn]. However, a naturally small direct vev can also originate from other sources like e.g. from higher-dimensional operators and most of the results of this paper can also be applied to this case. Type II See-Saw with $\SO(3)$ Flavour Symmetry {#sec:SO(3)} ============================================== We now show how a contribution to the neutrino mass matrix proportional to the unit matrix can be achieved by a spontaneously broken $\SO(3)$ flavour symmetry. We will analyze the consequences of such a framework for the masses and mixings in the lepton sector, where we consider the case that $\SO(3)$ acts on the lepton doublets. This has impact on the ingredients of the type II see-saw formula for the neutrino mass matrix of equation (\[eq:TypIISeeSawFormula\]) as well as for the mass matrix of the charged leptons. Structure of the Ingredients of the Type II See-Saw --------------------------------------------------- In order to break $\SO(3)$ spontaneously, we introduce additional heavy $G_{321}$-singlet superfields $\SingletH_{I}$ ($I \in \{1,2,3\}$) which are flavour triplets and acquire vevs $\<\SingletHiB_{f}\>$. In order to associate each flavon field $\SingletH_{I}$ with one of the right-handed neutrino superfields ${\hat{\nu}}^{\ChargeC i}$, we introduce three discrete symmetries ${Z}^{I}_2$. The representations of the fields involved in the generation of the neutrino mass matrix are given in table \[tab:repsSO(3)\]. $\vphantom{\sqrt{\big\|}}$Field ${\hat{L}}$ ${\hat{\Delta}}$ ${\hat{H}}_u$ ${\hat{\nu}}^{\ChargeC 1}$ ${\hat{\nu}}^{\ChargeC 2}$ ${\hat{\nu}}^{\ChargeC 3}$ $\SingletH_1$ $\SingletH_2$ $\SingletH_3$ ----------------------------------- ------------------ ------------------ ------------------ ---------------------------- ---------------------------- ---------------------------- ------------------ ------------------ ------------------ $\vphantom{\sqrt{\big\|}} \SO(3)$ $\boldsymbol{3}$ $\boldsymbol{1}$ $\boldsymbol{1}$ $\boldsymbol{1}$ $\boldsymbol{1}$ $\boldsymbol{1}$ $\boldsymbol{3}$ $\boldsymbol{3}$ $\boldsymbol{3}$ $\vphantom{\sqrt{\big\|}}{Z}^1_2 $ $+$ $+$ $+$ $-$ $+$ $+$ $-$ $+$ $+$ $\vphantom{\sqrt{\big\|}}{Z}^2_2 $ $+$ $+$ $+$ $+$ $-$ $+$ $+$ $-$ $+$ $\vphantom{\sqrt{\big\|}}{Z}^3_2 $ $+$ $+$ $+$ $+$ $+$ $-$ $+$ $+$ $-$ : \[tab:repsSO(3)\] Representations of $\SO(3) \times \:({Z}_2)^3$ involved in the generation of neutrino masses. The renormalizable term for neutrino Yukawa couplings is forbidden by the $\SO(3)$ flavour symmetry. Dirac masses for the neutrinos can however be generated by the higher-dimensional operators $$\begin{aligned} \label{eq:effectiveYnu} \mathcal{W}_{Y_\nu} &=& a_1 \,({\hat{L}}^f\cdot {\hat{H}}_u)\, {\hat{\nu}}^{\chargec 1} \, \frac{\<\SingletHaB_{f}\> }{\MSingletHa}+ a_2\,( {\hat{L}}^f\cdot{\hat{H}}_u)\, {\hat{\nu}}^{\chargec 2} \, \frac{\<\SingletHbB_{f}\> }{\MSingletHb}\nonumber \\ &&+\, a_3\,( {\hat{L}}^f \cdot{\hat{H}}_u)\, {\hat{\nu}}^{\chargec 3} \, \frac{\<\SingletHcB_{f}\> }{\MSingletHc} +\; \dots\; .\end{aligned}$$ $\MSingletHi$ correspond to the masses of some Froggatt-Nielsen fields [@Froggatt:1980sz] which are integrated out and produce the effective operators. In leading order the neutrino Yukawa matrix is thus given by $$\begin{aligned} \label{eq:LOYnu} \label{eq:Ynu} Y^0_\nu = \left(\begin{array}{ccc} \displaystyle a_1 \frac{\<\SingletHaB_{1}\>}{\MSingletHa}&\displaystyle a_2 \frac{\<\SingletHbB_{1}\>}{\MSingletHb} &\displaystyle a_3 \frac{\<\SingletHcB_{1}\>}{\MSingletHc} \\[4mm] \displaystyle a_1 \frac{\<\SingletHaB_{2}\>}{\MSingletHa}&\displaystyle a_2 \frac{\<\SingletHbB_{2}\>}{\MSingletHb} &\displaystyle a_3 \frac{\<\SingletHcB_{2}\>}{\MSingletHc}\\[4mm] \displaystyle a_1 \frac{\<\SingletHaB_{3}\> }{\MSingletHa}&\displaystyle a_2 \frac{\<\SingletHbB_{3}\>}{\MSingletHb} &\displaystyle a_3 \frac{\<\SingletHcB_{3}\>}{\MSingletHc} \end{array} \right) .\vspace{1mm}\end{aligned}$$ The Majorana mass term for the right-handed neutrinos is restricted by the discrete symmetries and for our choice of $Z_2$ symmetries given by $$\begin{aligned} \label{eq:MR0withSO(3)} \mathcal{W}_{M_\mathrm{RR}} & =& \frac{1}{2} {\hat{\nu}}^{\chargec 1} M_{\mathrm{R}1}{\hat{\nu}}^{\chargec 1} + \frac{1}{2} {\hat{\nu}}^{\chargec 2} M_{\mathrm{R}2}{\hat{\nu}}^{\chargec 2} + \frac{1}{2} {\hat{\nu}}^{\chargec 3} M_{\mathrm{R}3}{\hat{\nu}}^{\chargec 3}\nonumber \\ &&+ \frac{1}{2} \sum_{i,j,I,J=1}^3{\hat{\nu}}^{\chargec i} (M'_\mathrm{RR})_{ij}{\hat{\nu}}^{\chargec j}\, \,\delta_{iI}\,\delta_{jJ}\,\sum_{f=1}^3 \frac{\<(\SingletHB_I)_{f}\> \<(\SingletHB_J)_{f}\>}{M_{N' I} M_{N' J}} + \;\dots\; .\end{aligned}$$ At leading order, the mass matrix is thus forced to a diagonal structure $$\begin{aligned} \label{eq:LOMR} M_\mathrm{RR}^0 = \begin{pmatrix} M_{\mathrm{R}1} &0&0\\ 0&M_{\mathrm{R}2}&0\\ 0&0&M_{\mathrm{R}3} \end{pmatrix} .\end{aligned}$$ In the presence of the $\SO(3)$ symmetry, the couplings of the lepton doublets to the triplet Higgs superfield ${\hat{\Delta}}$ are given by $$\begin{aligned} \label{eq:YDeltaSO(3)} \mathcal{W}_{\Delta} = y_\Delta\, {\hat{L}}^T{}^f \, i\tau_2\,{\hat{\Delta}} \,{\hat{L}}^f +\sum_{f,g=1}^3 {\hat{L}}^T{}^f \,i\tau_2\, {\hat{\Delta}} \,{\hat{L}}^g \, \sum_{I=1}^3 b_I \frac{\<\SingletHiB_{f}\> \<\SingletHiB_{g}\>}{M_{L I}^2} + \;\dots\; .\end{aligned}$$ At leading order, this results in a contribution $Y_\Delta v_\Delta$ to the neutrino mass matrix proportional to the unit matrix in flavour space, $$\begin{aligned} \label{eq:LOmLL} Y_{\Delta}^0 = y_{\Delta} \begin{pmatrix} 1&0&0\\ 0&1&0\\ 0&0&1 \end{pmatrix} .\end{aligned}$$ Assuming $M_{NI} = M_{N'I} = M_{LI}$, the next-to-leading order operators are of the order $(\<\SingletHiB_{f}\> / \MSingletHi)^2 \approx(Y_\nu^0)^2_{If}$. If the neutrino Yukawa couplings are very small, the next-to-leading order operators are strongly suppressed and can approximately be neglected. However, if the effective field theory expansion parameters $\<\SingletHiB_{f}\> / \MSingletHi$ are relatively large, we should stress that a careful analysis of higher-dimensional operators of the superpotential and also of the Kähler potential (see e.g. [@King:2003xq]) has to be performed. Let us consider for example the effect of a neutrino Yukawa coupling $(Y_\nu^0)_{33} \approx \<\SingletHcB_{3}\> / M_{N 3} \approx 0.1$. The next-to-leading order operators of equation (\[eq:YDeltaSO(3)\]) induce a contribution to the atmospheric mass squared difference $\Delta m^2_\mathrm{atm}:= m_3^2 - m_1^2$ for a given scale of the direct mass term $m^{\mathrm{II}} \mathbbm{1} = y_\Delta v_\Delta \mathbbm{1}$. If we take for example $m^{\mathrm{II}}\approx 0.1$ eV and set $b_3= y_\Delta$ and $a_3=1$ for simplicity, this would induce a contribution to $\Delta m^2_\mathrm{atm}$ of about $2 \cdot 10^{-4}$ eV$^2$, which is still about an order of magnitude below the observed experimental value and thus provides only a rather small correction. Since we will not assume particularly large neutrino Yukawa couplings in this work, we will ignore the next-to-leading order operators in the following. The Mass Matrix of the Charged Leptons {#sec:Ye} -------------------------------------- Before we turn to classes of models which illustrate the mechanism, we discuss the generation of the mass matrix for the charged leptons. The most unrestricted case can be achieved by introducing three new additional flavour-triplet Higgs superfields $\SingletHei_{f}$ and three additional ${Z}'_2$ symmetries, as specified in table \[tab:repsSO(3)ChargedLeptons\]. $\vphantom{\sqrt{\big\|}}$Field ${\hat{e}}^{\ChargeC 1}$ ${\hat{e}}^{\ChargeC 2}$ ${\hat{e}}^{\ChargeC 3}$ $\SingletHe_{1}$ $\SingletHe_{2}$ $\SingletHe_{3}$ ------------------------------------ -------------------------- -------------------------- -------------------------- ------------------ ------------------ ------------------ $\vphantom{\sqrt{\big\|}} \SO(3)$ $\boldsymbol{1}$ $\boldsymbol{1}$ $\boldsymbol{1}$ $\boldsymbol{3}$ $\boldsymbol{3}$ $\boldsymbol{3}$ $\vphantom{\sqrt{\big\|}}{Z}'^1_2 $ $-$ $+$ $+$ $-$ $+$ $+$ $\vphantom{\sqrt{\big\|}}{Z}'^2_2 $ $+$ $-$ $+$ $+$ $-$ $+$ $\vphantom{\sqrt{\big\|}}{Z}'^3_2 $ $+$ $+$ $-$ $+$ $+$ $-$ : \[tab:repsSO(3)ChargedLeptons\] Representations of the charged leptons and new $G_{321}$-singlets under the horizontal symmetries $\SO(3) \times \:({Z}'_2)^3$, leading to the most unrestricted scenario. The fields are singlets under the symmetries ${Z}^1_2,{Z}^2_2$ and ${Z}^3_2$. This results in a general charged lepton Yukawa matrix generated by higher-dimensional operators, which can be hierarchical from a hierarchy of the vevs $\<\SingletHeaB_{f}\>,\<\SingletHebB_{f}\>$ and $\<\SingletHecB_{f}\>$ or from the masses of the Froggatt-Nielsen fields. The superpotential operators for the Yukawa interactions are $$\begin{aligned} \label{eq:effectiveYe} \mathcal{W}_{Y_e} &=& - \,a'_1 \, ({\hat{L}}^f\cdot{\hat{H}}_d)\, {\hat{e}}^{\chargec 1} \, \frac{\<\SingletHeaB_{f}\> }{\MSingletHea} - a'_2\, ({\hat{L}}^f\cdot{\hat{H}}_d) \, {\hat{e}}^{\chargec 2} \, \frac{\<\SingletHebB_{f}\> }{\MSingletHeb}\nonumber \\ &&-\, a'_3\, ({\hat{L}}^f\cdot{\hat{H}}_d)\, {\hat{e}}^{\chargec 3} \, \frac{\<\SingletHecB_{f}\> }{\MSingletHec} - \;\dots \; .\end{aligned}$$ The ${\cal O}(1)$-coefficients $ a_1,a_2,a_3 $ and $ a'_1,a'_2,a'_3 $ stem from the realization of the effective operators and are in principle calculable within an underlying full theory. The leading order Yukawa matrix for the charged leptons is given by $$\begin{aligned} \label{eq:Ye} Y^0_e = \left(\begin{array}{ccc} \displaystyle a'_1 \frac{\<\SingletHeaB_{1}\>}{\MSingletHea} &\displaystyle a'_2 \frac{\<\SingletHebB_{1}\>}{\MSingletHeb} &\displaystyle a'_3 \frac{\<\SingletHecB_{1}\>}{\MSingletHec} \\[4mm] \displaystyle a'_1 \frac{\<\SingletHeaB_{2}\>}{\MSingletHea} &\displaystyle a'_2 \frac{\<\SingletHebB_{2}\>}{\MSingletHeb} &\displaystyle a'_3 \frac{\<\SingletHecB_{2}\>}{\MSingletHec} \\[4mm] \displaystyle a'_1 \frac{\<\SingletHeaB_{3}\> }{\MSingletHea} &\displaystyle a'_2 \frac{\<\SingletHebB_{3}\>}{\MSingletHeb} &\displaystyle a'_3 \frac{\<\SingletHecB_{3}\>}{\MSingletHec} \end{array} \right) .\vspace{1mm}\end{aligned}$$ Different choices of discrete symmetries and flavon fields involved in the generation of the charged lepton mass matrix can give more predictive scenarios. Obviously, the $\SO(3)$ symmetry might leave the quark sector completely unaffected. On the other hand, the framework discussed so far can also be extended to the quark sector, yielding rather unrestricted quark masses with a natural hierarchy among the columns of the quark Yukawa matrices. For the present however, and in the rest of the paper, we restrict ourselves to the lepton sector. SO(3) Vacuum Alignment {#sec:VacuumAlignment} ====================== To stay as minimal as possible, we consider in the following the case that the charged leptons couple to the the same flavon fields $\SingletH_I$ as the neutrinos. This corresponds to replacing ${Z'}_2^{I} \rightarrow Z_2^{I}$ and $\SingletHei_f \rightarrow \SingletHi_f$ in section \[sec:Ye\]. The Yukawa matrix of the charged leptons is then related to the neutrino Yukawa matrix. The $\SO(3)$ flavour symmetry in the lepton sector is spontaneously broken by the vevs $\<\SingletHaB_f\>$. Let us consider first the general case with complex vevs. A simultaneous SO(3) rotation of the three vevs allows to eliminate three real degrees of freedom. We will shortly consider a scenario where the SO(3) vacuum is aligned such that the vevs are real. Complex phases in the Yukawa matrices $Y_\nu$ and $Y_e$ then stem entirely from the coefficients $ a_1,a_2,a_3 $ and $ a'_1,a'_2,a'_3 $. It turns out that with this vacuum alignment, which allows for three texture zeros in the Yukawa matrices, the type II scenario can realize the observed masses and mixings in the lepton sector in a particularly natural way. Real Alignment for the SO(3) Vacuum {#sec:RealVacuumAlignment} ----------------------------------- A possibility to achieve real vevs is to introduce three driving superfields ${\hat{A}},{\hat{B}}$ and ${\hat{C}} $ and to assume a superpotential of the form $$\begin{aligned} \mathcal{W} = {\hat{A}} (\SingletH_1^2 - \Lambda_1^2) + {\hat{B}} (\SingletH_2^2 - \Lambda_2^2) + {\hat{C}} (\SingletH_3^2 - \Lambda_3^2) \end{aligned}$$ with positive soft mass squareds $$\begin{aligned} m^2_{\theta_1} >0 \; , \quad m^2_{\theta_2} >0 \quad \mbox{and} \quad m^2_{\theta_3} >0 \end{aligned}$$ for the scalar components $\theta_1,\theta_2$ and $\theta_3$ of the flavon superfields $\SingletH_1,\SingletH_2$ and $\SingletH_3$ (see e.g. [@Barbieri:1999km]). $\Lambda_1,\Lambda_2$ and $\Lambda_3$ could stem from vevs of some $\SO(3)$-singlets. They can be chosen positive and real by a proper phase choice for these fields. Let us explicitly consider the vacuum alignment for $\theta_1$. Minimization of $\|F_{A}\|^2$ yields $$\begin{aligned} \Lambda_1^2=\sum_f (\mathrm{Re}\<\SingletHaB_{f}\>)^2 - \sum_i (\mathrm{Im}\<\SingletHaB_{f}\>)^2 + 2 i\sum_f \mathrm{Re}\<\SingletHaB_{f}\>\mathrm{Im}\<\SingletHaB_{f}\>\: .\end{aligned}$$ With our choice $\Lambda_1 \in \mathbbm{R}^+$, this leads to the two conditions $$\begin{aligned} 0 &=& \sum_f \mathrm{Re}\<\SingletHaB_{f}\>\mathrm{Im}\<\SingletHaB_{f}\> \; ,\\ \label{eqn:MinCondFA_1} \Lambda_1^2&=&\sum_f (\mathrm{Re}\<\SingletHaB_{f}\>)^2 - \sum_f (\mathrm{Im}\<\SingletHaB_{f}\>)^2\; . \end{aligned}$$ The soft mass term $\mathcal{V}_{s} = m^2_{\theta_1} \theta_1^* \theta_1$ deforms the driving potential and its minimization under the conditions from $\|F_{A}\|^2$ results in $\mathrm{Im}\<\SingletHaB_{f}\>=0$ for all components $f \in \{1,2,3\}$. This can be seen by plugging equation (\[eqn:MinCondFA\_1\]) into $\mathcal{V}_{s}$, which yields $$\begin{aligned} \mathcal{V}_{s} &=& m^2_{\theta_1}\,\left[ \sum_f (\mathrm{Re}\<\SingletHaB_{f}\>)^2 + \sum_f (\mathrm{Im}\<\SingletHaB_{f}\>)^2\right] \nonumber \\ &=& m^2_{\theta_1}\,\left[ \Lambda_1^2 + 2 \sum_f (\mathrm{Im}\<\SingletHaB_{f}\>)^2 \right] .\end{aligned}$$ Thus, the small positive mass squared forces the vev $\<\theta_1\>$ to be real. Treating $\<\theta_2\>$ and $\<\theta_3\>$ analogously, we obtain a real alignment for the vevs of all scalar components of the flavon superfields, $$\begin{aligned} \mathrm{Im}\<\SingletHB_{1}\>=0 \; , \quad \mathrm{Im}\<\SingletHB_{2}\>=0 \; , \quad \mathrm{Im}\<\SingletHB_{3}\>=0 \; . \end{aligned}$$ The moduli of the vev vectors are given by $$\begin{aligned} \|\<\SingletHB_{1}\>\|^2 \approx \Lambda_1^2 \; , \quad \|\<\SingletHB_{2}\>\|^2 \approx \Lambda_2^2 \; , \quad \|\<\SingletHB_{3}\>\|^2 \approx \Lambda_3^2 \; , \end{aligned}$$ which could lead to a hierarchy among the vevs via a hierarchy among $\Lambda_1$, $\Lambda_2$ and $\Lambda_3$. We can now use the $\SO(3)$ freedom to set two components of one of the vev vectors, say $\<\SingletHiB_{f}\>$, and one corresponding component of a second vev vector, say $\<\SingletHjB_{f}\>$, to zero. Defining the real expansion parameters \[eq:DefEpsilon\]$$\begin{aligned} &&\varepsilon_I:=\|a_I\|\frac{\Lambda_I}{\MSingletHi}\; , \;\; \varepsilon_J:=\|a_J\|\frac{\Lambda_J}{\MSingletHj}\; , \;\;\varepsilon_K:=\|a_K\|\frac{\Lambda_K}{\MSingletHk}\; , \\ &&\varepsilon'_I:=\|a'_I\|\frac{\Lambda_I}{\MSingletHei}\; , \;\; \varepsilon'_J:=\|a'_J\|\frac{\Lambda_J}{\MSingletHej}\; ,\;\; \varepsilon'_K:=\|a'_K\|\frac{\Lambda_K}{\MSingletHek}\; , \;\;\end{aligned}$$ without loss of generality we can write \[eq:VacAlignmentColumns\]$$\begin{aligned} a_I \frac{\<\SingletHB_I\>}{\MSingletHi} \!\!\!&:=&\!\!\! \left(\!\begin{array}{c} 0 \\ 0 \\ h \,e^{i \delta_I} \varepsilon_I\end{array}\!\!\right) \!, \; a_J\frac{\<\SingletHB_J\>}{\MSingletHj} := \! \left(\!\begin{array}{c} 0 \\ e\,e^{i \delta_J}\varepsilon_J \\ f \,e^{i\delta_J}\varepsilon_J \end{array}\!\!\right)\! , \; a_K \frac{\<\SingletHB_K\>}{\MSingletHk} := \! \left(\!\begin{array}{c} a \,e^{i \delta_K}\varepsilon_K \\ b\,e^{i\delta_K}\varepsilon_K \\ c \,e^{i \delta_K}\varepsilon_K \end{array}\!\!\right)\! , \\ a'_I \frac{\<\SingletHB_I\>}{\MSingletHei} \!\!\!&:=&\!\!\! \left(\!\begin{array}{c} 0 \\ 0 \\ h \,e^{i \delta'_I}\varepsilon'_I \end{array}\!\!\right) \!, \; a'_J \frac{\<\SingletHB_J\>}{\MSingletHej} := \left(\!\begin{array}{c} 0 \\ e\,e^{i \delta'_J}\varepsilon'_J \\ f \,e^{i\delta'_J}\varepsilon'_J \end{array}\!\!\right) \!, \; a'_K \frac{\<\SingletHB_K\>}{\MSingletHek} := \! \left(\!\begin{array}{c} a \,e^{i \delta'_K}\varepsilon'_K \\ b\,e^{i\delta'_K}\varepsilon'_K \\ c\,e^{i \delta'_K}\varepsilon'_K \end{array}\!\!\right)\!,\end{aligned}$$ with real coefficients $\{a,b,c,e,f,h\}$ satisfying $$\begin{aligned} &&h \approx 1 \; , \;\; e^2 + f^2 \approx 1 \; , \;\; a^2 + b^2 + c^2 \approx 1 \; .\end{aligned}$$ We choose the convention that the labels $\{1,2,3\}$ are assigned to $I,J$ and $K$ such that $$\begin{aligned} \varepsilon'_3 > \varepsilon'_2 > \varepsilon'_1 \end{aligned}$$ holds in the charged lepton sector. This also defines the labels of $\varepsilon_1,\varepsilon_2$ and $\varepsilon_3$ and of the heavy mass eigenvalues $M_{\mathrm{R}1},M_{\mathrm{R}2}$ and $M_{\mathrm{R}3}$ of the right-handed neutrinos. In the following, by a global phase transformation of all the leptons, we arrange for the direct mass term for the neutrinos proportional to the unit matrix to be real and positive. In addition, we absorb possible phases of $M_{\mathrm{R}1},M_{\mathrm{R}2}$ and $M_{\mathrm{R}3}$ in the columns of $Y_\nu$. The phases in the Yukawa matrices then stem from the coefficients $a_1,a_2,a_3$ and $a'_1,a'_2,a'_3$ which are in general complex, $$\begin{aligned} &&\da := \mbox{Arg} \,(a_1) \; ,\; \; \db := \mbox{Arg} \,(a_2) \; ,\; \;\dc := \mbox{Arg} \,(a_3) \; ,\; \; \\ &&\daP := \mbox{Arg} \,(a'_1) \; ,\; \; \dbP := \mbox{Arg} \,(a'_2) \; ,\; \;\dcP := \mbox{Arg} \,(a'_3) \; .\end{aligned}$$ Textures for the Yukawa Matrices and Type II Scenarios ------------------------------------------------------ The parameterization of the the SO(3)-breaking vacuum specified in equation (\[eq:VacAlignmentColumns\]) uses the basis where the Yukawa matrices $Y_\nu$ and $Y_e$ each have three zero entries. Motivation for this choice of basis would come from a full theory beyond the framework presented here. Table \[tab:TypeIIScenarios\] shows the textures for $Y_\nu$ and $Y_e$ for different proportions of $\varepsilon'_I, \varepsilon'_J$ and $\varepsilon'_K$ using the ordering convention $\varepsilon'_3 > \varepsilon'_2 > \varepsilon'_1$. Additional characteristic features of the Yukawa matrices are that each column has a common complex phase and that, without additional symmetries, the non-zero components of each column have a common typical order of magnitude defined by the expansion parameters $\varepsilon_1,\varepsilon_2,\varepsilon_3$ and $\varepsilon'_1,\varepsilon'_2,\varepsilon'_3$. The hierarchy among the masses of the charged leptons can naturally be realized via $\varepsilon'_3\gg\varepsilon'_2\gg\varepsilon'_1$, which we will assume in the following. $$\begin{aligned} \begin{array}{\|c\|c\|c\|} \hline \mbox{Model}&Y^0_\nu&Y^0_e\\ \hline \mbox{A1} & \left(\begin{array}{ccc} a\,e^{i \da}\epsa&0&0\\ b\,e^{i \da}\epsa&e\,e^{i \db}\epsb&0\\ c\,e^{i \da}\epsa&f\,e^{i \db}\epsb&h\,e^{i \dc}\epsc \end{array} \right) & \left(\begin{array}{ccc} a\,e^{i \daP}\epsaP&0&0\\ b\,e^{i \daP}\epsaP&e\,e^{i \dbP}\epsbP&0\\ c\,e^{i \daP}\epsaP&f\,e^{i \dbP}\epsbP&h\,e^{i \dcP}\epscP \end{array} \right)\\ \hline \mbox{A2} & \left(\begin{array}{ccc} 0&a\,e^{i \db}\epsb&0\\ e\,e^{i \da}\epsa&b\,e^{i \db}\epsb&0\\ f\,e^{i \da}\epsa&c\,e^{i \db}\epsb&h\,e^{i \dc}\epsc \end{array} \right) & \left(\begin{array}{ccc} 0&a\,e^{i \dbP}\epsbP&0\\ e\,e^{i \daP}\epsaP&b\,e^{i \dbP}\epsbP&0\\ f\,e^{i \daP}\epsaP&c\,e^{i \dbP}\epsbP&h\,e^{i \dcP}\epscP \end{array} \right)\\ \hline \mbox{B1} & \left(\begin{array}{ccc} a\,e^{i \da}\epsa&0&0\\ b\,e^{i \da}\epsa&0&e\,e^{i \dc}\epsc\\ c\,e^{i \da}\epsa&h\,e^{i \db}\epsb&f\,e^{i \dc}\epsc \end{array} \right) & \left(\begin{array}{ccc} a\,e^{i \daP}\epsaP&0&0\\ b\,e^{i \daP}\epsaP&0&e\,e^{i \dcP}\epscP\\ c\,e^{i \daP}\epsaP&h\,e^{i \dbP}\epsbP&f\,e^{i \dcP}\epscP \end{array} \right)\\ \hline \mbox{B2} & \left(\begin{array}{ccc} 0&a\,e^{i \db}\epsb&0\\ 0&b\,e^{i \db}\epsb&e\,e^{i \dc}\epsc\\ h\,e^{i \da}\epsa&c\,e^{i \db}\epsb&f\,e^{i \dc}\epsc \end{array} \right) & \left(\begin{array}{ccc} 0&a\,e^{i \dbP}\epsbP&0\\ 0&b\,e^{i \dbP}\epsbP&e\,e^{i \dcP}\epscP\\ h\,e^{i \daP}\epsaP&c\,e^{i \dbP}\epsbP&f\,e^{i \dcP}\epscP \end{array} \right)\\ \hline \mbox{C1} & \left(\begin{array}{ccc} 0&0&a\,e^{i \delta_3}\,\varepsilon_3 \\ 0&e\,e^{i \delta_2}\,\varepsilon_2&b\,e^{i \delta_3}\,\varepsilon_3 \\ h\,e^{i \delta_1}\,\varepsilon_1&f\,e^{i \delta_2}\,\varepsilon_2&c\,e^{i \delta_3}\,\varepsilon_3 \end{array} \right) & \left(\begin{array}{ccc} 0&0&a\,e^{i \delta'_3}\,\varepsilon'_3 \\ 0&e\,e^{i \delta'_2}\,\varepsilon'_2&b\,e^{i \delta'_3}\,\varepsilon'_3 \\ h\,e^{i \delta'_1}\,\varepsilon'_1&f\,e^{i \delta'_2}\,\varepsilon'_2&c\,e^{i\delta'_3}\,\varepsilon'_3 \end{array} \right) \\ \hline \mbox{C2} & \left(\begin{array}{ccc} 0&0&a\,e^{i \delta_3}\,\varepsilon_3 \\ e\,e^{i \delta_1}\,\varepsilon_1&0&b\,e^{i \delta_3}\,\varepsilon_3 \\ f\,e^{i \delta_1}\,\varepsilon_1&h\,e^{i \delta_2}\,\varepsilon_2&c\,e^{i \delta_3}\,\varepsilon_3 \end{array} \right) & \left(\begin{array}{ccc} 0&0&a\,e^{i \delta'_3}\,\varepsilon'_3 \\ e\,e^{i \delta'_1}\,\varepsilon'_1&0&b\,e^{i \delta'_3}\,\varepsilon'_3 \\ f\,e^{i \delta'_1}\,\varepsilon'_1&h\,e^{i \delta'_2}\,\varepsilon'_2&c\,e^{i\delta'_3}\,\varepsilon'_3 \end{array} \right) \\ \hline \end{array}\end{aligned}$$ ### Models Type A: Large Mixing $\theta_{23}$ from the Neutrino Mass Matrix In the models of type A, the mixing $\theta^e_{23}$ from the charged lepton mass matrix $M_e$ is approximately zero. The almost maximal total lepton mixing $\theta_{23}$ thus has to be generated in the neutrino mass matrix. This can for example be achieved if the dominant contribution to the type I part of the neutrino mass matrix stems from the right-handed neutrino $\nu_{\mathrm{R}}^{ 2}$ for model A1 or $\nu_{\mathrm{R}}^{ 1}$ for model A2. As in the single right handed neutrino dominance case [@King:1998jw; @King:1999cm] for a pure type I neutrino mass matrix, the condition for a nearly maximal $\theta_{23}$ is $\|e\| \approx \|f\|$. In addition, the zero in the first element of the column containing the coefficients $e$ and $f$ in general avoids the generation of a large $\theta^\nu_{13}$ from the neutrino mass matrix. Model A1 has the additional feature that $\theta^e_{12}$ and $\theta^e_{13}$ are very small, whereas in model type A2 a charged lepton mixing $\tan (\theta^e_{12}) = a/b$ is generated which induces a contribution to the total lepton mixings $\theta_{12}$ and $\theta_{13}$. In order not to violate the CHOOZ bound on $\theta_{13}$, $\theta^e_{12}$ has to be rather small which means that $a$ has to be somewhat smaller than $b$. Consequently, the large mixing $\theta_{12} \approx 32^\circ$ should be generated mainly in the neutrino sector. With sequential right-handed neutrino dominance [@King:1999mb; @King:2002nf] for $m^\mathrm{I}_{\mathrm{LL}}$, this can easily be realized. We will study model A1 in detail in section \[sec:Typw2SeqRhdNuDom\]. ### Models Type B: Large Mixing $\theta_{23}$ from the Charged Leptons The mixing $\theta^e_{23}$ from the charge lepton mass matrix is given by $\tan (\theta^e_{23}) \approx e/f$. $e \approx f$ can explain the nearly maximal total lepton mixing $\theta_{23}$ in a lopsided mass model, provided that the contribution $\theta^\nu_{23}$ is small, which can e.g. be achieved if $\nu_{\mathrm{R}}^{ 2}$ is the dominant right-handed neutrino for model B1 and $\nu_{\mathrm{R}}^{ 1}$ for model B2. Since as for the model A2, a large $\theta^e_{12}$ in model B2 would induce a large contribution to $\theta_{13}$, the large solar mixing should be generated by the neutrino mass matrix. ### Models Type C: Large Angles for all the Charged Lepton Mixings For models of type C, it is possible to generate bi-large lepton mixing entirely from $Y_e$. Small mixing from the neutrino mass matrix would correspond to $\nu_{\mathrm{R}}^{ 1}$ being the dominant right-handed neutrino and $\nu_{\mathrm{R}}^{ 2}$ being subdominant for model C1 or to $\nu_{\mathrm{R}}^{ 1}$ being dominant and $\nu_{\mathrm{R}}^{ 2}$ being subdominant for model C2. For the model C2, the large mixings are given by $\tan (\theta_{12})=a/b$ and $\tan (\theta_{23})=\sqrt{a^2+b^2}/c$ (see also [@Babu:2001cv]). While $\theta_{13}\approx 0$ for model C2, it has to be taken care that $\theta_{13}$ is below the experimental bounds in model C1. A Type II Scenario with Sequential RHND {#sec:Typw2SeqRhdNuDom} ======================================== As an example for a type II model where the bi-large lepton mixing stems from the neutrino mass matrix, we now consider explicitly the model A1 of table \[tab:TypeIIScenarios\]. We make the additional assumption of sequential righthanded neutrino dominance (RHND) [@King:1999mb; @King:2002nf], leading to a hierarchical type I part $m_\nu^{\mathrm{I}}$ of the neutrino mass matrix. The vev structure for this case allows to define (see equation (\[eq:VacAlignmentColumns\])) $$\begin{aligned} a_3 \frac{\<\SingletHB_3\>}{\MSingletHc} \!\!\!&:=&\!\!\! \!\left(\begin{array}{c} 0 \\ 0 \\ h \,e^{i \delta_3} \varepsilon_3\end{array}\right) \!, \;\; \! a_2\frac{\<\SingletHB_2\>}{\MSingletHb} := \! \left(\begin{array}{c} 0 \\ e\,e^{i \delta_2}\varepsilon_2 \\ f \,e^{i\delta_2}\varepsilon_2 \end{array}\right)\! , \;\;\! a_1 \frac{\<\SingletHB_1\>}{\MSingletHa} := \! \left(\begin{array}{c} a \,e^{i \delta_1}\varepsilon_1 \\ b\,e^{i\delta_1}\varepsilon_1 \\ c \,e^{i \delta_1}\varepsilon_1 \end{array}\right)\! , \\ a'_3 \frac{\<\SingletHB_3\>}{\MSingletHec} \!\!\!&:=&\!\!\! \!\left(\begin{array}{c} 0 \\ 0 \\ h \,e^{i \delta'_3}\varepsilon'_3 \end{array}\right) \!, \;\;\! a'_2 \frac{\<\SingletHB_2\>}{\MSingletHeb} := \left(\begin{array}{c} 0 \\ e\,e^{i \delta'_2}\varepsilon'_2 \\ f \,e^{i\delta'_2}\varepsilon'_2 \end{array}\right) \!, \;\;\! a'_1 \frac{\<\SingletHB_1\>}{\MSingletHea} := \! \left(\begin{array}{c} a \,e^{i \delta'_1}\varepsilon'_1 \\ b\,e^{i\delta'_1}\varepsilon'_1 \\ c\,e^{i \delta'_1}\varepsilon'_1 \end{array}\right)\!,\end{aligned}$$ leading to the Yukawa matrices $$\begin{aligned} \label{eqn:YukawaMatrices_A1} Y^0_\nu = \left(\begin{array}{ccc} a\,e^{i \da}\epsa&0&0\\ b\,e^{i \da}\epsa&e\,e^{i \db}\epsb&0\\ c\,e^{i \da}\epsa&f\,e^{i \db}\epsb&h\,e^{i \dc}\epsc \end{array} \right) \!, \;\; Y^0_e = \left(\begin{array}{ccc} a\,e^{i \daP}\epsaP&0&0\\ b\,e^{i \daP}\epsaP&e\,e^{i \dbP}\epsbP&0\\ c\,e^{i \daP}\epsaP&f\,e^{i \dbP}\epsbP&h\,e^{i \dcP}\epscP \end{array} \right) \!.\end{aligned}$$ In order to match notation with reference [@King:1999mb; @King:2002nf], we define $$\begin{aligned} M^0_\mathrm{RR} =: \left(\begin{array}{ccc} \!X&0&0\!\\ \!0&Y&0\!\\ \!0&0&X'\! \end{array} \right)\!,\end{aligned}$$ denoting the mass of the dominant right-handed neutrino by $Y$ and the mass of the subdominant one by $X$. The sequential RHND condition we impose is then $$\begin{aligned} \label{eq:SequSubDominace} \left\|\frac{\epsb^2}{Y}\right\| \gg \left\|\frac{\epsa^2}{X}\right\|\gg \left\|\frac{\epsc^2}{X'}\right\| \; .\end{aligned}$$ The leading order type II neutrino mass matrix is then given by $$\begin{aligned} \label{eq:TypeIIMassMatrix_A1} m_{\mathrm{LL}}^\nu &\!\!\!=\!\!\!& m^\mathrm{II} \!\left(\begin{array}{ccc} \!1&\!0\!&0\!\\ \!0&\!1\!&0\!\\ \!0&\!0\!&1\! \end{array} \right)\! - \frac{e^{i2\db} \epsb^2 v_u^2}{Y} \! \left(\begin{array}{ccc} \!0&0\!&0\!\\ \!0&e^2\!&ef\!\\ \!0&ef\!&f^2\! \end{array} \right)\! - \frac{e^{i2\da} \epsa^2 v_u^2}{X}\! \left(\begin{array}{ccc} \vphantom{f^2}\!a^2&ab&ac\!\vphantom{f^2}\\ \vphantom{f^2}\!ab&b^2&bc\!\vphantom{f^2}\\ \vphantom{f^2}\!ac&bc&c^2\!\vphantom{f^2} \end{array} \right)\! . \end{aligned}$$ We now analyze the lepton masses, mixings and CP phases which are generated by $m^\nu_{\mathrm{LL}}$ and by the lepton mass matrix $M_e=v_d Y_e$ analytically. Analytic Results for Neutrino Masses, Lepton Mixings and CP Phases ------------------------------------------------------------------ The mixing matrix in the lepton sector, the MNS matrix $U_{\mathrm{MNS}}$, is defined by the charged electroweak current $\overline{{e}_\mathrm{L}}^f \gamma^\mu U_{MNS} {\nu}^f_\mathrm{L}$ in the mass basis. Defining the diagonalization matrices $U_{e_\mathrm{L}},U_{e_\mathrm{R}}$ and $U_{\nu_\mathrm{L}}$ by $$\begin{aligned} U_{e_\mathrm{L}} \, M_e \,U^\dagger_{e_\mathrm{R}} = \left(\begin{array}{ccc} \!m_e&0&0\!\\ \!0&m_\mu&0\!\\ \!0&0&m_\tau\! \end{array} \right)\! , \quad U_{\nu_\mathrm{L}} \,m^\nu_{\mathrm{LL}}\,U^T_{\nu_\mathrm{L}} = \left(\begin{array}{ccc} \!m_1&0&0\!\\ \!0&m_2&0\!\\ \!0&0&m_3\! \end{array} \right)\! ,\end{aligned}$$ the MNS matrix is given by $$\begin{aligned} U_{\mathrm{MNS}} = U_{e_\mathrm{L}} U^\dagger_{\nu_\mathrm{L}}\; .\end{aligned}$$ We use the parameterization $ U_{\mathrm{MNS}} = R_{23} U_{13} R_{12} P_0 $ with $R_{23}, U_{13}, R_{12}$ and $P_0$ being defined as $$\begin{aligned} R_{12}:= \left(\begin{array}{ccc} c_{12} & s_{12} & 0\\ -s_{12}&c_{12} & 0\\ 0&0&1\end{array}\right) , \:& \quad U_{13}:=\left(\begin{array}{ccc} c_{13} & 0 & s_{13}e^{-i\delta}\\ 0&1 & 0\\ - s_{13}e^{i\delta}&0&c_{13}\end{array}\right) , \nonumber \\ R_{23}:=\left(\begin{array}{ccc} 1 & 0 & 0\\ 0&c_{23} & s_{23}\\ 0&-s_{23}&c_{23} \end{array}\right) , \:& \quad P_0:= \begin{pmatrix} 1&0&0\\0&e^{i\beta_2}&0\\0&0&e^{i\beta_3} \end{pmatrix}. \end{aligned}$$ The masses of the charged leptons are given by $m_\tau = h \,\epscP\, v_d,m_\mu = e \,\epsbP\, v_d$ and $m_e = a \,\epsaP\, v_d$. Unless $h,e$ or $a$ are zero, we can easily choose $\epscP,\epsbP$ and $\epsaP$ such that the right charged lepton masses are produced. In addition we note that the mixings $\theta^e_{12},\theta^e_{13}$ and $\theta^e_{23}$, which stem from $U_{e_\mathrm{L}}$ and could contribute to the MNS matrix, are very small. Furthermore, in leading order each column of $M_e$ has a common complex phase, which can be absorbed by $U_{e_\mathrm{R}}$. Therefore, the charged leptons do not influence the leptonic CP phases in this approximation. Using the analytical methods for diagonalizing neutrino mass matrices with small $\theta_{13}$ derived in [@King:2002nf], from $m^\nu_{\mathrm{LL}} = m_{\mathrm{LL}}^{\mathrm{II}} + m_{\mathrm{LL}}^{\mathrm{I}}$ we find for the mixing angles $$\begin{aligned} \label{AnalyticResultForT23} \tan (\theta_{23}) &\approx& \frac{\|e\|}{\|f\|} \; , \\ \label{AnalyticResultForT13} \tan (2 \theta_{13}) &\approx& \frac{2 \,\|a\|\,\epsa^2\,v_u^2}{X} \, \frac{\|\sin (\theta_{23}) \|b\| + \cos (\theta_{23}) \|c\| \,\mbox{sign}\,(b\, c\, e\, f)\|}{ \|2\, m^{\mathrm{II}}\, \sin (\widetilde \delta) + m_3^{\mathrm{I}} e^{i(2\db + 3\pi/2 -\widetilde \delta)}\| } \; ,\\ \label{AnalyticResultForT12}\tan (\theta_{12}) &\approx& \frac{\|a\|}{\|\cos (\theta_{23}) \|b\| - \sin (\theta_{23}) \|c\|\,\mbox{sign}\,(b\, c\, e\, f)\| } \; ,\end{aligned}$$ where $m_i^{\mathrm{I}}$ ($i\in \{1,2,3\}$) are the mass eigenvalues of the hierarchical $m_{\mathrm{LL}}^{\mathrm{I}}$ given by $$\begin{aligned} m_1^{\mathrm{I}} &=& {\cal O} \left(\frac{\epsc^2 v_u^2}{X'}\right) \;\approx\;0 \; , \vphantom{\frac{\|e\|}{\|f\|}}\\ \label{eq:m2I_A1} m_2^{\mathrm{I}} &\approx& \frac{(\|a\|^2 + \|\cos (\theta_{23}) \|b\| - \sin (\theta_{23}) \|c\|\,\mbox{sign}\,(b\, c\, e\, f)\|^2) \,\epsa^2\,v_u^2}{X} \;\approx\;\frac{\|a\|^2\,\epsa^2\,v_u^2 }{\sin^2 (\theta_{12}) X}\;,\\ m_3^{\mathrm{I}} &\approx& \frac{(\|e\| \sin (\theta_{23})+\|f\|\cos (\theta_{23}))^2\,\epsb^2\,v_u^2 }{Y}\;, \end{aligned}$$ and with $\widetilde \delta$ defined by $$\begin{aligned} \label{eq:DiracCP_A1_1} \tan (\widetilde \delta) := \frac{m_3^{\mathrm{I}} \, \sin (2 \db - 2 \da)}{ m_3^{\mathrm{I}} \,\cos (2 \db - 2 \da) - 2 \, m^{\mathrm{II}} \, \cos (2 \da)}\, .\end{aligned}$$ Given $\tan (\widetilde \delta)$, $\widetilde \delta$ has to be chosen such that $$\begin{aligned} \frac{\sin (\theta_{23}) \|b\| + \cos (\theta_{23}) \|c\| \,\mbox{sign}\,(b\, c\, e\, f)}{ \mbox{sign}\,(a\, b) \, [ 2\, m^{\mathrm{II}}\,e^{-i (2\da+ 3\pi/2)} \, \sin (\widetilde \delta) + m_3^{\mathrm{I}} e^{-i(2\da -2\db+\widetilde \delta)} ] } \ge 0 \;.\end{aligned}$$ This does not effect $\theta_{13}$, which we have defined to be $\ge 0$, however it is relevant for extracting the Dirac CP phase $\delta$, given by $$\begin{aligned} \label{eq:DiracCP_A1_2} \delta &\approx& \left\{ \begin{array}{cl} \widetilde \delta & \mbox{for $P\ge0$\,,}\\ \widetilde \delta + \pi & \mbox{for $P<0$\,,} \end{array} \right.\end{aligned}$$ with $P$ being defined by $$\begin{aligned} P := \frac{ \cos (\theta_{23}) \|b\| - \sin (\theta_{23}) \|c\|\,\mbox{sign}\,(b\, c\, e\, f) }{ \mbox{sign}\,(a\, b) \, [ (\cos (\theta_{23}) \|b\| - \sin (\theta_{23}) \|c\|\,\mbox{sign}\,(b\, c\, e\, f))^2 - \|a\|^2 ] }\; .\end{aligned}$$ The mass eigenvalues of the complete type II neutrino mass matrix are given by \[eq:ComplMassEigenvOfMnuTypeII\]$$\begin{aligned} m_1 &\approx& \|m^{\mathrm{II}}\|\;,\\ m_2 &\approx& \| m^{\mathrm{II}} - m_2^{\mathrm{I}} \,e^{i 2 \da}\|\;,\\ m_3 &\approx& \| m^{\mathrm{II}} - m_3^{\mathrm{I}} \,e^{i 2 \db}\|\;, \end{aligned}$$ and, for $m^{\mathrm{II}} \not= 0$, the Majorana phases $\beta_2$ and $\beta_3$ can be extracted by \[eq:MajPhases\_A1\]$$\begin{aligned} \beta_2 &\approx& \frac{1}{2}\mbox{arg}\,( m^{\mathrm{II}} - m_2^{\mathrm{I}} \,e^{i 2 \da})\;,\\ \beta_3 &\approx& \frac{1}{2}\mbox{arg}\,( m^{\mathrm{II}} - m_3^{\mathrm{I}} \,e^{i 2 \db})\;.\end{aligned}$$ A graphical illustration of the correlation between the masses and phases is given in figure \[fig:MassesinComplexPlane\]. $\vcenter{\hbox{\includegraphics[scale=1]{ComplexMasses_m3.eps}}}$ Discussion ---------- We find that for the mixings, only $\theta_{13}$ is affected by the direct mass term proportional to the unit matrix, whereas $\theta_{12}$ and $\theta_{23}$ are independent of $m^\mathrm{II}$. This is due to the fact that $\theta_{23}$ only depends on the dominant contribution to the type I part of the neutrino mass matrix and $\theta_{12}$ depends only on the subdominant type I part to leading order in $\theta_{13}$. On the other hand, $\theta_{13}$ depends on both, the dominant and the subdominant part. Looking at equation (\[AnalyticResultForT13\]) for $\theta_{13}$, it naively seems as $\theta_{13}$ goes to zero as $m^\mathrm{II}$ increases. However, by using the result for $\widetilde \delta$, the denominator of equation (\[AnalyticResultForT13\]) can be rewritten as $$\begin{aligned} \|2\, m^{\mathrm{II}}\, \sin (\widetilde \delta) + m_3^{\mathrm{I}} e^{i(2\db + 3\pi/2 -\widetilde \delta)}\| = \left\| m_3^{\mathrm{I}}\, \frac{\cos (2 \db - \widetilde \delta)}{\cos (2 \da)}\right\| ,\end{aligned}$$ which does no longer depend on $m^\mathrm{II}$ explicitly. Note that the term $\cos (2 \db - \widetilde \delta)/\cos (2 \da)$ goes to $1$ for small $m^\mathrm{II}$ and to $\cos (2 \db)/\cos (2 \da)$ for large $m^\mathrm{II}$, because $ \widetilde \delta$ goes to $0$ as discussed later and is shown in figure \[fig:DiracCPphase\_A1\].[^3] Besides the additional explicit dependence on the parameters $a,b$ and $c$, the mixing angle $\theta_{13}$ is only suppressed by a factor of $m_2^\mathrm{I}/m_3^\mathrm{I}$. However, as we discuss later and is shown in figure \[fig:m2Iandm3I\], $m_2^\mathrm{I}$ gets smaller for increasing $m^\mathrm{II}$ much faster that $m_3^\mathrm{I}$. Consequently, $\theta_{13}$ in fact decreases with increasing $m^\mathrm{II}$, however it does not go to zero for large $m^\mathrm{II}$. The dependence of $\theta_{13}$ on the type II mass scale is illustrated in figure \[fig:Theta13vsmII\]. As the type II contribution $m^{\mathrm{II}}_\mathrm{LL}$ gets larger and the neutrino mass scale increases, the mass splittings have to get smaller (see e.g. figure \[fig:NuMasses\]) in order to match the experimentally observed mass squared differences $\Delta m^2_\mathrm{sol}:= m_2^2 - m_1^2$ and $\|\Delta m^2_\mathrm{atm}\|:= \|m_3^2 - m_1^2\|$. Depending on $m^\mathrm{II}$ and the complex phases $\da$ and $\db$, this determines the eigenvalues $m_2^\mathrm{I}$ and $m_3^\mathrm{I}$ of the type I part of the neutrino mass matrix (see figure \[fig:MassesinComplexPlane\]). Note that with a hierarchical type I part $m^{\mathrm{I}}_{\mathrm{LL}}$, a normal mass ordering as well as an inverse ordering can be achieved. The latter is only possible if $(m^\mathrm{II} \cos (2 \db))^2 > \|\Delta m^2_\mathrm{atm}\|$. The dependence of $m_2^\mathrm{I}$ and $m_3^\mathrm{I}$ on the type II mass scale $m^{\mathrm{II}}$ is shown in figure \[fig:m2Iandm3I\]. For producing the small mass squared differences for a larger $m^\mathrm{II}$ in a natural way, i.e. without cancellations of two relatively large terms $m^{\mathrm{I}}_{\mathrm{LL}}$ and $m^{\mathrm{II}}_{\mathrm{LL}}$, $m_2^\mathrm{I}$ and $m_3^\mathrm{I}$ have to be smaller. From equation (\[eq:ComplMassEigenvOfMnuTypeII\]) and the definitions of $m_2^{\mathrm{I}}$ and $m_3^{\mathrm{I}}$ we find that the masses $Y$ and $X$ of the right-handed neutrinos, which control these mass splittings, can be chosen appropriately in a natural way. As we have already noted, the ratio $m_2^\mathrm{I}/m_3^\mathrm{I}$ for a partially degenerate mass spectrum is smaller than for hierarchical neutrino masses. This means that in the type II scenario with partially degenerate neutrino masses, sequential righthanded neutrino dominance is even more natural than in the pure type I see-saw case. As can be seen from equation (\[eq:MajPhases\_A1\]), the $m^{\mathrm{II}}$-dependence of $m_2^{\mathrm{I}}$ and $m_3^{\mathrm{I}}$ implies that the Majorana phases $\beta_1$ and $\beta_2$ get smaller for larger $m^{\mathrm{II}}$ (see figure \[fig:MajPhases\]). From equations (\[eq:DiracCP\_A1\_1\]) and (\[eq:DiracCP\_A1\_2\]), we conclude that the Dirac CP phase $\delta$ generically gets smaller for larger $m^{\mathrm{II}}$ as well. Quantitatively, the dependence of $\delta$ on the type II mass scale is shown in figure \[fig:DiracCPphase\_A1\] for some sample choices of $\da$ and $\db$. [${\ensuremath{\vcenter{\hbox{\includegraphics[scale=1]{Theta13vsmII_2.eps}}}}}$]{} \ \ [$\vcenter{\hbox{\includegraphics[scale=1]{DiracCPphase_A1.eps}}}$]{} Renormalization Group Corrections {#sec:RGRunning} ================================= In order to compare the predictions of see-saw scenarios with the experimental data obtained at low energy, the renormalization group (RG) running of the effective neutrino mass matrix has to be taken into account. It is known that for partially degenerate neutrino masses, RG corrections to the neutrino mixing angles can be significant. The corrections to the masses and mass squared differences are relevant even for strongly hierarchical neutrino masses and they can be enhanced or suppressed for a partially degenerate neutrino mass spectrum depending on the CP phases. Let us assume for example that the considered type II scenarios are embedded into a unified model where all additional degrees of freedom except for the ones contained in the right-handed neutrino superfields ${\hat{\nu}}_{\mathrm{R}}^{ \Nf}$ are integrated out above some energy scale $M_\mathrm{U}$. At $M_\mathrm{U}$, which could be the scale of gauge coupling unification, the model parameters are defined. In this case, the effective neutrino mass matrix has to be run from $M_\mathrm{U}$ to low energy using the $\beta$-functions for the various energy ranges above and between the see-saw scales [@Antusch:2002rr; @Antusch:2002ek] and below the mass scale of the lightest right-handed neutrino [@Chankowski:1993tx; @Babu:1993qv; @Antusch:2001ck; @Antusch:2001vn; @Antusch:2002ek]. For an accurate computation of the RG corrections for the neutrino mass parameters, the coupled system of differential equations has to be solved successively for the various effective theories [@King:2000hk; @Antusch:2002rr]. In addition to the parameters of the MNS matrix, the RG running between $M_\mathrm{U}$ and the electroweak scale $M_{\mathrm{EW}}$ then depends on the additional degrees of freedom corresponding to the lepton Yukawa couplings and the masses of the right-handed neutrinos. For a small difference of the Majorana CP phases corresponding to the mass eigenvalues $m_1$ and $m_2$, which is generically the case in the presented type II framework, radiative corrections to the lepton mixings have a characteristic property, as has been pointed out in [@Antusch:2002hy; @Antusch:2002fr]. The running of the solar mixing $\theta_{12}$ is generically larger than the RG corrections to the other mixings, if the atmospheric mixing is large already at high energy. We will see that in particular $\theta_{12}$ can be significantly lower at high energy, depending on $m^{\mathrm{II}}$ and $\tan \beta$. We will now estimate the generic size of the RG corrections for the mixing angles by making additional simplifications and assumptions. At first, let us consider the case that the heaviest right-handed neutrino corresponds to the column of the neutrino Yukawa matrix with the largest entries and that it has a mass larger than $M_\mathrm{U}\approx 2\cdot 10^{16}$. If we then assume that the other entries of the neutrino Yukawa matrix are very small compared to $y_\tau$, we can approximately neglect the running due to the neutrino Yukawa matrix and, for a rough estimate, simply use the RGEs for the neutrino mass operator below the see-saw scales. In order to see which quantities control the size of the RG effects, it is useful to consider the RGEs for the parameters of the MNS matrix [@Chankowski:1999xc; @Casas:1999tg; @Antusch:2003kp]. For example, the running of the mixing angles in leading order in the small quantity $\theta_{13}$ in the MSSM is given by [@Antusch:2003kp] $$\begin{aligned} \label{eq:AnalyticApproxT12} \frac{\mathrm{d}}{\mathrm{d} t}\, \theta_{12} & \!=\! & -\frac{y_\tau^2}{32\pi^2} \, \sin 2\theta_{12} \, s_{23}^2\, \frac{ \| {m_1} + {m_2}\, e^{\I 2 \beta_2}\|^2 }{\Delta m^2_\mathrm{sol} } + \mathscr{O}(\theta_{13}) \;, \label{eq:Theta12Dot}\end{aligned}$$ $$\begin{aligned} \label{eq:AnalyticApproxT13} \frac{\mathrm{d}}{\mathrm{d} t}\,\theta_{13} & \!=\! & \frac{y_\tau^2}{32\pi^2} \, \sin 2\theta_{12} \, \sin 2\theta_{23} \, \frac{m_3}{\Delta m^2_\mathrm{atm} \left( 1+\zeta \right)} \times \nonumber\\ && \quad \times \left[ m_1 \cos(2\beta_3-\delta) - \left( 1+\zeta \right) m_2 \, \cos(2\beta_3-2\beta_2-\delta) - \zeta m_3 \, \cos\delta \right] \nonumber \\ && + \mathscr{O}(\theta_{13}) \;, \label{eq:Theta13Dot}\end{aligned}$$ $$\begin{aligned} \label{eq:AnalyticApproxT23} \frac{\mathrm{d}}{\mathrm{d} t}\, \theta_{23} & \!=\! & -\frac{y_\tau^2}{32\pi^2} \, \sin 2\theta_{23} \, \frac{1}{\Delta m^2_\mathrm{atm}} \left[ c_{12}^2 \, \|m_2\, e^{\I 2 (\beta_3-\beta_2)} + m_3\|^2 + s_{12}^2 \, \frac{\|m_1\, e^{\I 2 \beta_3} + m_3\|^2}{1+\zeta} \right] \nonumber\\ & & {} + \mathscr{O}(\theta_{13}) \;, \label{eq:Theta23Dot}\vspace{1mm}\end{aligned}$$ where, with the renormalization scale $\mu$, $t$ is defined by $t:=\ln (\mu/\mu_0)$ and where we have used the abbreviation $\zeta := {\Delta m^2_\mathrm{sol}}/{\Delta m^2_\mathrm{atm}}$. Compared to the SM, $y_\tau$ in the MSSM is given by $y_\tau=y_\tau^\mathrm{SM} \sqrt{ 1+\tan^2\beta }$, which yields an enhancement of the RG effects for large $\tan \beta$. In addition, the running is generically enhanced due to factors of the type $m^2_i/\Delta m^2_\mathrm{sol}$ or $m^2_i/\Delta m^2_\mathrm{atm}$ if the neutrino mass scale is larger than the mass differences. For partially degenerate neutrinos, this is in particular the case for $m^2_i/\Delta m^2_\mathrm{sol}$, which enhances the running of $\theta_{12}$. For an estimate of the RG corrections, we have solved the evolution of the parameters numerically from low energy to high energy under the additional assumptions that $\theta_{13}\|_{M_\mathrm{EW}} = 0^\circ$ and $\beta_3\|_{M_\mathrm{EW}} = \beta_2\|_{M_\mathrm{EW}} = 0^\circ$. Below the SUSY breaking scale, which we have taken to be $1$ TeV, we assume the theory to be effectively the Standard Model. The values for $\theta_{12}\|_{M_\mathrm{U}}$ and $\theta_{23}\|_{M_\mathrm{U}}$ required in order to produce the best fit value $\theta_{12}\|_{M_\mathrm{EW}}\approx 32^\circ$ and $\theta_{23}\|_{M_\mathrm{EW}}\approx 45^\circ$ are shown in figure \[fig:RGCorrectionsMixingAngles\]. The RG correction factors for $\Delta m^2_\mathrm{sol}$ and $\Delta m^2_\mathrm{atm}$ with respect to the low energy values $\Delta m^2_\mathrm{sol} \approx 6.9 \cdot 10^{-5}$ eV$^2$ and $\Delta m^2_\mathrm{atm} \approx 2.6 \cdot 10^{-3}$ eV$^2$ are shown in figure \[fig:RGCorrectionsMassSquaredDiffs\] for the case of a normal mass ordering, i.e. $\Delta m^2_\mathrm{atm}>0$. It is known that instability under radiative corrections can require unnatural fine-tuning of the high energy parameters in order to produce the experimentally observed mass squared differences and mixings at low energy. Indeed, this is the case for nearly degenerate neutrino masses close to the experimental upper bounds and in addition large $\tan \beta$, such that the slopes $\frac{\mathrm{d}}{\mathrm{d} t}\theta_{12}$ and $\frac{\mathrm{d}}{\mathrm{d} t}\Delta m^2_\mathrm{sol}$ become very large. As we can see from figures \[fig:RGCorrectionsMixingAngles\] and \[fig:RGCorrectionsMassSquaredDiffs\], for the considered ranges of $\tan \beta \in [5,50]$ and $m^\mathrm{II}\in [0.03,1.5]$ eV, the RG corrections are well behaved and do not require any fine-tuning of the high energy parameters. Depending on $\tan\beta$, the RG effects for a given neutrino mass scale can either lead to significantly changed values for the mixings and mass squared difference at high energy or cause only rather small corrections. We conclude that though the requirement of naturalness restricts the neutrino mass scale for a given $\tan \beta$ via stability arguments in our scenarios, this does not lead to any problems for partially degenerate neutrinos. Nevertheless, the running has to be included in a careful analysis and it can easily be estimated using the figures \[fig:RGCorrectionsMixingAngles\] and \[fig:RGCorrectionsMassSquaredDiffs\]. \ \ Neutrinoless Double Beta Decay ============================== We now discuss the implications of the type II see-saw scenarios with spontaneously broken SO(3) flavour symmetry for $00\nu\beta$ decay. In order to be explicit, we focus on the models of type A1 and B1 with sequential right-handed neutrino dominance and vacuum alignment as introduced in section \[sec:Typw2SeqRhdNuDom\]. The effective mass $\Braket{m_\nu}$ for $00\nu\beta$ decay is then given by $$\begin{aligned} \label{eq:0nbbDecayMass_A1} \Braket{m_\nu} \;=\; \left\|\, \sum_i (U_\mathrm{MNS})_{1i}^2 \, m_i \,\right\| \;=\; \left\| (m^\nu_\mathrm{LL})_{11} \right\| \;\approx\; \left\|\,m^{\mathrm{II}} - \sin^2 (\theta_{12}) \, m_2^{\mathrm{I}} \, e^{2 i \da} \,\right\| ,\end{aligned}$$ as can be seen from equations (\[eq:TypeIIMassMatrix\_A1\]) and (\[eq:m2I\_A1\]). A discussion of the RG corrections to $\Braket{m_\nu}$ can be found in [@Antusch:2003kp]. The dependence of $\Braket{m_\nu}$ on the direct mass term $m_\nu^{\mathrm{II}}=m^{\mathrm{II}}\mathbbm{1}$ and on the complex phase $\da$ is illustrated in figure \[fig:EffMass0nbbDecay\]. For the models A1 and B1, we obtain $< m_\nu > \approx m_\nu^{\mathrm{II}}$ already for a lightest neutrino mass of about $0.02$ eV. For all the models with a real vacuum alignment, the general statement holds that when the direct mass term $m^{\mathrm{II}}\mathbbm{1}$ becomes dominant over the type I part of the mass matrix, the effective mass $< m_\nu >$ is approximately given by $m^{\mathrm{II}}$. The cancellations which can occur in the presence of a large difference of the Majorana phases associated with $m_1$ and $m_2$ are then absent and thus there are good prospects are for detecting $\Braket{m_\nu}$ in these classes of type II see-saw models. \ Discussion and Conclusions ========================== We have proposed a type II upgrade of type I see-saw models leading to new classes of models where partially degenerate neutrinos are as natural as hierarchical ones. A spontaneously broken SO(3) flavour symmetry forces the direct mass term for the neutrinos to be proportional to the unit matrix at leading order. This allows in principle to boost the mass of the lightest neutrino to any desired value leading from hierarchical to nearly degenerate mass spectra. Naturalness of models with nearly degenerate neutrino masses consists of two issues: First there is tree-level naturalness, which means that all parameters and in particular the small mass squared differences and bi-large mixing are produced by the model in a natural way. We have shown that our framework is natural in this respect and could produce any desired level of degeneracy at tree-level. Second, there is naturalness with respect to RG corrections. The latter should not require fine-tuning of the high-energy model parameters in order to produce the low-energy experimental values. RG corrections are un-suppressed in our scenario and thus, depending on $\tan \beta$, the naturalness requirement restricts the neutrino mass scale. Instead of nearly degenerate neutrinos, we have therefore considered partially degenerate neutrinos with a mass scale up to about $0.15$ eV, where the running does not lead to any naturalness problems. This mass range is particularly interesting since it might be accessible to experiments on neutrinoless double beta decay. For breaking the SO(3) flavour symmetry, we have considered a minimal set of flavon fields and a real alignment mechanism for their SO(3)-breaking vevs. The real vacuum alignment implies that there exists a basis where the Yukawa matrices have three texture zeros. This leads to classes of type II see-saw models with either small mixing from the charged leptons (type A), almost maximal atmospheric mixing from the charged leptons (type B) or models where in principle all mixing can be produced via the charged lepton mass matrix (type C). Characteristic features of them are that each column of the Yukawa matrices has a common complex phase and that, without additional symmetries, the non-zero components of each column have a common typical order of magnitude. In addition, for partially degenerate neutrino masses where the type II contribution $m^{\mathrm{II}}$ dominates over the type I part, the Majorana phases associated with the neutrino mass eigenvalues are predicted to be small and the effective mass for neutrinoless double beta decay is approximately given by $m^{\mathrm{II}}$. Future experiments on neutrinoless double beta decay can thus in principle rule out or confirm partially degenerate neutrino masses in this framework. Bi-large neutrino mixing and the two small mass squared differences can naturally be realized with sequential right-handed neutrino dominance [@King:1999mb; @King:2002nf] for the type I contribution to the neutrino mass matrix. One of the right-handed neutrinos and the corresponding column of the neutrino Yukawa matrix gives the dominant contribution to the type I mass matrix and is responsible for the large atmospheric mixing $\theta_{23}$ and the mass squared difference $\Delta m^2_\mathrm{atm}:= m_3^2 - m_1^2 $. The subdominant right-handed neutrino and the corresponding column then generate the smaller mass squared difference $\Delta m^2_\mathrm{sol}:= m_2^2 - m_1^2 $ and can account for a large solar neutrino mixing $\theta_{12}$. Due to the small difference of the Majorana phases corresponding to $m_1$ and $m_2$, the RG effects for $\theta_{12}$ in see-saw models are generically larger than for the other lepton mixing angles if $\theta_{23}$ is large already at high energy [@Antusch:2002hy; @Antusch:2002fr]. Depending on $\tan \beta$ and the neutrino mass scale, they can be sizable or lead only to rather small corrections. For partially degenerate neutrinos with a mass scale up to about $0.15$ eV, the running in general does not cause problems with naturalness, but it allows for a wider range of possible mass and mixing patterns at high energy. A careful analysis has to include RG effects and we have provided figures where estimates for the corrections can easily be read off. In the classes of type II see-saw models with sequential right-handed neutrino dominance for the type I contribution to the neutrino mass matrix and real vacuum alignment, the solar and the atmospheric neutrino mixings $\theta_{12}$ and $\theta_{23}$ are independent of the type II mass scale $m^{\mathrm{II}}$ and of the complex phases of the neutrino Yukawa matrix. This provides a natural way to upgrade these types of models continously from hierarchical neutrino mass spectra to partially degenerate ones, while maintaining the predictions for the two large lepton mixings. The mixing angle $\theta_{13}$ is generically small and furthermore decreases with increasing neutrino mass scale. In addition, we find that our scenario predicts that all observable CP phases, i.e. the Dirac CP phase $\delta$ relevant for neutrino oscillations and the Majorana CP phases $\beta_2$ and $\beta_3$, become small as the neutrino mass scale increases. This implies in particular that the effective mass for neutrinoless double beta decay is approximately equal to the neutrino mass scale and therefore neutrinoless double beta decay will be observable if the neutrino mass spectrum is partially degenerate. In our framework a partially degenerate neutrino mass spectrum is [a priori]{} as natural as a hierarchical spectrum. If neutrinos are partially degenerate, neutrinoless double beta decay has the potential to measure the neutrino mass scale. Acknowledgements {#acknowledgements .unnumbered} ================ We acknowledge support from the PPARC grant PPA/G/O/2002/00468. [10]{} CHOOZ, M. Apollonio et al., Limits on neutrino oscillations from the [CHOOZ]{} experiment, Phys. Lett. B466 (1999), 415–430, `hep-ex/9907037`. M. Maltoni, T. Schwetz, M. A. Tortola, and J. W. F. Valle, Status of three-neutrino oscillations after the SNO-salt data, (2003), `hep-ph/0309130`. Super[K]{}amiokande, T. Toshito, Super-[K]{}amiokande atmospheric neutrino results, (2001), `hep-ex/0105023`. KamLAND, K. Eguchi et al., First results from KamLAND: Evidence for reactor anti-neutrino disappearance, Phys. Rev. Lett. 90 (2003), 021802, `hep-ex/0212021`. SNO, S. N. Ahmed et al., Measurement of the total active B-8 solar neutrino flux at the Sudbury Neutrino Observatory with enhanced neutral current sensitivity, (2003), `nucl-ex/0309004`. D. N. Spergel et al., First year [W]{}ilkinson [M]{}icrowave [A]{}nisotropy [P]{}robe [(WMAP)]{} observations: [D]{}etermination of cosmological parameters, (2003), `astro-ph/0302209`. H. V. Klapdor-Kleingrothaus et al., Latest results from the [H]{}eidelberg-[M]{}oscow double-beta-decay experiment, Eur. Phys. J. A12 (2001), 147–154, `hep-ph/0103062`. 16EX Collaboration, C. E. Aalseth et al., The [IGEX Ge-76]{} neutrinoless double-beta decay experiment: [P]{}rospects for next generation experiments, Phys. Rev. D65 (2002), 092007, `hep-ex/0202026`. T. Yanagida, in Proceedings of the Workshop on the Unified Theory and the Baryon Number in the Universe (O. Sawada and A. Sugamoto, eds.), KEK, Tsukuba, Japan, 1979, p. 95. S. L. Glashow, The future of elementary particle physics, in Proceedings of the 1979 Carg[è]{}se Summer Institute on Quarks and Leptons (M. L[é]{}vy, J.-L. Basdevant, D. Speiser, J. Weyers, R. Gastmans, and M. Jacob, eds.), Plenum Press, New York, 1980, pp. 687–713. M. Gell-Mann, P. Ramond, and R. Slansky, Complex spinors and unified theories, in Supergravity (P. van Nieuwenhuizen and D. Z. Freedman, eds.), North Holland, Amsterdam, 1979, p. 315. R. N. Mohapatra and G. Senjanovi[ć]{}, Neutrino mass and spontaneous parity violation, Phys. Rev. Lett. 44 (1980), 912. D. O. Caldwell and R. N. Mohapatra, Neutrino mass explanations of solar and atmospheric neutrino deficits and hot dark matter, Phys. Rev. D48 (1993), 3259–3263. P. Bamert and C. P. Burgess, Naturally degenerate neutrinos, Phys. Lett. B329 (1994), 289–294, `hep-ph/9402229`. D.-G. Lee and R. N. Mohapatra, An SO(10) $\times$ $S_4$ scenario for naturally degenerate neutrinos, Phys. Lett. B329 (1994), 463–468, `hep-ph/9403201`. A. Ioannisian and J. W. F. Valle, SO(10) grand unification model for degenerate neutrino masses, Phys. Lett. B332 (1994), 93–99, `hep-ph/9402333`. A. S. Joshipura, Almost degenerate neutrinos, Z. Phys. C64 (1994), 31–35. A. S. Joshipura, Degenerate neutrinos in left-right symmetric theory, Phys. Rev. D51 (1995), 1321–1325, `hep-ph/9404354`. A. Ghosal, Almost degenerate neutrino mass in an $SU(2)_{QL}\! \times SU(2)_{LL}\! \times \! U(1)_Y$ model, Phys. Lett. B398 (1997), 315–320. C. D. Carone and M. Sher, Supersymmetric model of quasi-degenerate neutrinos, Phys. Lett. B420 (1998), 83–90, `hep-ph/9711259`. E. Ma, Splitting of three nearly mass-degenerate neutrinos, Phys. Lett. B456 (1999), 48–53, `hep-ph/9812344`. A. K. Ray and S. Sarkar, Almost degenerate neutrinos in a left-right symmetric model with discrete symmetries, Phys. Rev. D58 (1998), 055010. G. Lazarides, Degenerate neutrinos and supersymmetric inflation, Phys. Lett. B [452]{} (1999) 227, `hep-ph/9812454`. C. Wetterich, Natural maximal $\nu_\mu$-$\nu_\tau$ mixing, Phys. Lett. B451 (1999), 397–405, `hep-ph/9812426`. Y.-L. Wu, SO(3) gauge symmetry and neutrino-lepton flavor physics, Phys. Rev. D60 (1999), 073010, `hep-ph/9810491`. A. Ghosal, Bi-maximal neutrino mixing with SO(3) flavour symmetry, `hep-ph/9905470`. R. Barbieri, L. J. Hall, G. L. Kane, and G. G. Ross, Nearly degenerate neutrinos and broken flavour symmetry, (1999), `hep-ph/9901228`. Y.-L. Wu, Two/three-flavor oscillation and MSW/just-so solution of neutrinos in the SO(3) gauge model, Int. J. Mod. Phys. A14 (1999), 4313–4330, `hep-ph/9901320`. Y.-L. Wu, Spontaneous breaking of flavor symmetry and naturalness of nearly degenerate neutrino masses and bi-maximal mixing, Sci. China A43 (2000), 988–995, `hep-ph/9906435`. E. Ma and G. Rajasekaran, Softly broken $A_4$ symmetry for nearly degenerate neutrino masses, Phys. Rev. D64 (2001), 113012, `hep-ph/0106291`. K. S. Babu, E. Ma, and J. W. F. Valle, Underlying $A_4$ symmetry for the neutrino mass matrix and the quark mixing matrix, Phys. Lett. B552 (2003), 207–213, `hep-ph/0206292`. M. Patgiri and N. N. Singh, Right-handed Majorana neutrino mass matrices for generating bimaximal mixings in degenerate and inverted models of neutrinos, Int. J. Mod. Phys. A18 (2003), 743–754, `hep-ph/0301254`. R. N. Mohapatra, M. K. Parida, and G. Rajasekaran, High scale mixing unification and large neutrino mixing angles, (2003), `hep-ph/0301234`. G. Lazarides, Q. Shafi and C. Wetterich, Proton lifetime and fermion masses in an SO(10) model, Nucl. Phys. B [181]{} (1981) 287. R. N. Mohapatra and G. Senjanovi[ć]{}, Neutrino masses and mixings in gauge models with spontaneous parity violation, Phys. Rev. D23 (1981), 165. C. Wetterich, Neutrino masses and the scale of [$B-L$]{} violation, Nucl. Phys. B187 (1981), 343. E. Ma and U. Sarkar, Neutrino masses and leptogenesis with heavy [H]{}iggs triplets, Phys. Rev. Lett. 80 (1998), 5716–5719, `hep-ph/9802445`. S. F. King, Neutrino mass models, Rept. Prog. Phys. 67 (2004), 107–158, `hep-ph/0310204`. S. F. King, Large mixing angle MSW and atmospheric neutrinos from single right-handed neutrino dominance and U(1) family symmetry, Nucl. Phys. B576 (2000), 85–105, `hep-ph/9912492`. S. F. King, Constructing the large mixing angle MNS matrix in see-saw models with right-handed neutrino dominance, JHEP 09 (2002), 011, `hep-ph/0204360`. J. C. Pati and A. Salam, Lepton number as the fourth color, Phys. Rev. D10 (1974), 275–289. R. N. Mohapatra and J. C. Pati, A ’natural’ left-right symmetry, Phys. Rev. D11 (1975), 2558. G. Senjanovi[ć]{} and R. N. Mohapatra, Exact left-right symmetry and spontaneous violation of parity, Phys. Rev. D12 (1975), 1502. J. C. Pati and A. Salam, Unified lepton - hadron symmetry and a gauge theory of the basic interactions, Phys. Rev. D8 (1973), 1240. H. Georgi, Particles and fields, (edited by Carlson, C. E.), A.I.P., 1975, p. 575. H. Fritzsch and P. Minkowski, Unified interactions of leptons and hadrons, Ann. Phys. 93 (1975), 193–266. C. D. Froggatt and H. B. Nielsen, Statistical analysis of quark and lepton masses, Nucl. Phys. B164 (1980), 114. S. F. King and I. N. R. Peddie, Canonical normalisation and Yukawa matrices, (2003), `hep-ph/0312237`. S. F. King, Atmospheric and solar neutrinos with a heavy singlet, Phys. Lett. B439 (1998), 350–356, `hep-ph/9806440`. S. F. King, Atmospheric and solar neutrinos from single right-handed neutrino dominance and U(1) family symmetry, Nucl. Phys. B562 (1999), 57–77, `hep-ph/9904210`. K. S. Babu and S. M. Barr, Bimaximal neutrino mixings from lopsided mass matrices, Phys. Lett. B525 (2002), 289–296, `hep-ph/0111215`. S. Antusch, J. Kersten, M. Lindner, and M. Ratz, Neutrino mass matrix running for non-degenerate see-saw scales, Phys. Lett. B538 (2002), 87–95, `hep-ph/0203233`. S. Antusch and M. Ratz, Supergraph techniques and two-loop beta-functions for renormalizable and non-renormalizable operators, JHEP 07 (2002), 059, `hep-ph/0203027`. P. H. Chankowski and Z. Pluciennik, Renormalization group equations for seesaw neutrino masses, Phys. Lett. B316 (1993), 312–317, `hep-ph/9306333`. K. S. Babu, C. N. Leung, and J. Pantaleone, Renormalization of the neutrino mass operator, Phys. Lett. B319 (1993), 191–198, `hep-ph/9309223`. S. Antusch, M. Drees, J. Kersten, M. Lindner, and M. Ratz, Neutrino mass operator renormalization revisited, Phys. Lett. B519 (2001), 238–242, `hep-ph/0108005`. S. Antusch, M. Drees, J. Kersten, M. Lindner, and M. Ratz, Neutrino mass operator renormalization in [T]{}wo [H]{}iggs [D]{}oublet [M]{}odels and the [MSSM]{}, [hep-ph/0110366]{} (2001). S. F. King and N. N. Singh, Renormalisation group analysis of single right-handed neutrino dominance, Nucl. Phys. B591 (2000), 3–25, `hep-ph/0006229`. S. Antusch, J. Kersten, M. Lindner, and M. Ratz, The [LMA]{} solution from bimaximal lepton mixing at the GUT scale by renormalization group running, Phys. Lett. B544 (2002), 1–10, `hep-ph/0206078`. S. Antusch and M. Ratz, Radiative generation of the [LMA]{} solution from small solar neutrino mixing at the [GUT]{} scale, JHEP 11 (2002), 010, `hep-ph/0208136`. P. H. Chankowski, W. Krolikowski, and S. Pokorski, Fixed points in the evolution of neutrino mixings, Phys. Lett. B473 (2000), 109, `hep-ph/9910231`. J. A. Casas, J. R. Espinosa, A. Ibarra, and I. Navarro, General [RG]{} equations for physical neutrino parameters and their phenomenological implications, Nucl. Phys. B573 (2000), 652, `hep-ph/9910420`. S. Antusch, J. Kersten, M. Lindner, and M. Ratz, Running neutrino masses, mixings and CP phases: Analytical results and phenomenological consequences, Nucl. Phys. B674 (2003), 401–433, `hep-ph/0305273`. [^1]: E-mail: `[email protected]` [^2]: E-mail: `[email protected]` [^3]: Note that there are special choices of the complex phases where mass eigenvalues of the type I part of the neutrino mass matrix do not become smaller for larger $m^\mathrm{II}$. This is e.g. the case for $\db=45^\circ$, as can be seen from figure \[fig:MassesinComplexPlane\]. In this case $\widetilde \delta$ does not go to zero as $m^\mathrm{II}$ increases. For large $m^\mathrm{II}$, we regard this special choices for the phases as unnatural, since two large quantities from $m^{\mathrm{II}}_{\mathrm{LL}}$ and $m^{\mathrm{II}}_{\mathrm{LL}}$ have to conspire in order to produce a small mass squared difference.	{ "pile_set_name": "ArXiv" }
--- abstract: 'Yang-Mills theories supplemented by an additional coordinate constraint, which is solved and substituted in the original Lagrangian, provide examples of the so called Nambu models, in the case where such constraints arise from spontaneous Lorentz symmetry breaking. Some explicit calculations have shown that, after additional conditions are imposed, Nambu models are capable of reproducing the original gauge theories, thus making Lorentz violation unobservable and allowing the interpretation of the corresponding massless gauge bosons as the Goldstone bosons arising from the spontaneous symmetry breaking. A natural question posed by this approach in the realm of gauge theories is to determine under which conditions the recovery of an arbitrary gauge theory from the corresponding Nambu model, defined by a general constraint over the coordinates, becomes possible. We refer to these theories as extended Nambu models (ENM) and emphasize the fact that the [defining coordinate]{} constraint is not treated as a standard gauge fixing term. At this level, the mechanism for generating the constraint is irrelevant and the case of spontaneous Lorentz symmetry breaking is taken only as a motivation, which naturally bring this problem under consideration. Using a non-perturbative Hamiltonian analysis we prove that the ENM yields the original gauge theory after we demand current conservation for all time, together with the imposition of the Gauss laws constraints as initial conditions upon the dynamics of the ENM. The Nambu models yielding electrodynamics, Yang Mills theories and linearized gravity are particular examples of our general approach.' address: - \| CENTRA, Departamento de Física, Universidade do Algarve, 8005-139 Faro, Portugal.\ [email protected] - \| Instituto de Ciencias Nucleares, Universidad Nacional Autónoma de México, 04510 México, Distrito Federal, México\ [email protected] author: - 'C. A. Escobar' - 'L. F. Urrutia' title: 'Extended Nambu Models: their relation to gauge theories' --- Introduction ============ Gauge theories, symmetry principles and spontaneous symmetry breaking have been successfully put together in building the electroweak sector of the Standard Model [@Electrodebil; @Electrodebil2]. Another well known example including these concepts is the construction of pion interactions in the nonlinear sigma model, which is characterized by spontaneous chiral symmetry breaking [@chiral; @chiral2; @chiral3]. Here, pions are interpreted as the massless Goldstone bosons (GBs) generated by such breaking and the construction of their interactions arises from the most general Lagrangian respecting the remaining unbroken symmetries. It was precisely the understanding of pions as GBs what motivated the possibility of looking at fundamental massless particles, like photons and gravitons for example, as the GBs arising from some spontaneous symmetry breaking. Since these particles are of tensorial nature, a tensor-valued vacuum expectation value (VEV) would be required. Such a non-zero VEV produces fixed directions on the spacetime, thus leading to spontaneous Lorentz symmetry breaking (SLSB). The main goal of this proposal is to provide a dynamical setting for the gauge principle. One of the first realizations of this idea is the abelian Nambu model (ANM), which was proposed in Ref. as a description of electrodynamics arising from SLSB. The ANM is defined by the standard Maxwell Lagrangian plus a constraint over the vector potential $$\mathcal{L}(A_{\mu })=-\frac{1}{4}F_{\mu \nu }F^{\mu \nu }-A_{\mu }J^{\mu },\quad \quad \quad A_{\mu }A^{\mu }=\mathfrak{n}^{2}M^{2}, \label{NANMabeliano}$$ where $\mathfrak{n}^{\mu }$ is a properly oriented constant vector in the Lorentz space, while $M$ is the proposed scale associated with the SLSB. The origin of such a constraint can be understood as the consequence of a non-zero vacuum expectation value of the vector potential: $\langle A_{\mu }\rangle = \mathfrak{n}_{\mu }M$ produced by the SLSB. In a way similar to the non linear sigma model, the constraint is to be solved and substituted in the Lagrangian ${\mathcal L}$, thus drastically modifying the properties of the original gauge theory. In fact, the resulting model defined by Eq. (\[NANMabeliano\]) has more degrees of freedom (DOF), current conservation is not fulfilled and it does not exhibit gauge invariance. The aim in Ref. was to make explicit the conditions under which the ANM turns out to be equivalent to standard QED, instead of yielding a physical violation of the Lorentz symmetry. The idea that gauge particles (photons and gravitons, for example) might arise as the GBs of a theory with SLSB has been widely studied and goes back a long way[@RUSOS; @Bjorken; @Guralnik; @Guralnik1]. In this approach, the masslessness of these gauge particles can then be understood in terms of the Goldstone theorem [@TeoGoldstone; @TeoGoldstone2], instead of gauge invariance requirements. Nambu models have been further considered in relation to electrodynamics [@Azatov-Chkareuli; @Urru-Mont] and generalized to the Yang Mills [@JLCH1; @JLCH2; @JLCH3; @NANM_ES_UR] and the gravitational [@JLCH4] cases. Perturbative calculations in Nambu models show that, to the order considered and under some appropriate initial conditions, all SLSB contributions to physical processes cancel out, yielding an equivalence with the original gauge theory [@Nambu-Progr; @Azatov-Chkareuli; @JLCH1; @JLCH2; @JLCH3; @JLCH4]. Additional previous works in gravitation can be found in Refs. , respectively. In this way, Nambu models can be understood as a gauge theory plus some constraint upon the coordinates, which is to be solved and substituted in the corresponding Lagrangian, thus producing a quite different model as a first step. However, since the final goal is to recover the original gauge theory, this motivates the problem of determining under which conditions this is accomplished. This non-standard procedure is illustrated for a Yang-Mills theory in Ref. . In this paper we deal with this question for what we call an extended Nambu model (ENM) and emphasize the fact that the constraint is not treated as a standard gauge fixing term. The ENM is defined by an arbitrary gauge invariant Lagrangian density, which we call the mother gauge theory (MGT), supplemented by a constraint among the coordinates analogous to that appearing in Eq. (\[NANMabeliano\]). From the general point of view we adopt here, the way in which the constraint is generated is irrelevant for our purposes and the case of SLSB is taken only as a motivation, which naturally bring this problem under consideration. The MGT is defined by a Lagrangian density $\mathcal{L}$ from which only first class constraints (FCC) arise, being the generators of a non trivial gauge symmetry. The paper is organized as follows. In Section \[gaugetheory\] we define the general MGT which we are going to deal with. Then we employ the Dirac method [@Dirac_method] to perform the Hamiltonian analysis leading to presence of FCC. We close this section by writing the Hamiltonian and the canonical algebra that define the dynamics of the MGT, which will be used as benchmark to establish the equivalence between this MGT and the corresponding ENM. In Section \[NGmodel\] we present the ENM, which is defined by the same Lagrangian density $\mathcal{L}$ introduced in Section \[gaugetheory\], plus one constraint $F$ among the coordinates of the MGT. This constraint is solved for two generic cases and subsequently substituted in the MGT Lagrangian. Such an ENM leads to a theory without gauge invariance, which is manifest in the appearance of second class constraints only. Under the conditions imposed for $\mathcal{L}$ in Section \[gaugetheory\], we subsequently perform the Hamiltonian analysis of the ENM. Finally, identifying a suitable transformation between the canonical variables of the MGT and those of the ENM, we show that the Hamiltonian describing the MGT and the Hamiltonian corresponding to the ENM have the same functional form. Also, such transformation allows us to prove that the canonical algebra of the ENM induces the canonical algebra of the MGT. In this way, after suitable conditions are imposed in order to recover gauge invariance, the ENM is shown to be equivalent to the original MGT. In Section \[examples\], we present some examples of MGT where the conditions required in Section \[gaugetheory\] are fulfilled and, as a consequence, the equivalence with any associated ENM, defined by the constraint $F$, is established. The \[CURCONS1\] includes the calculation of the generalized current conservation equation in our MGT, which arises as a consequence of the gauge invariance of the Hamiltonian action. \[Mechanics1a\] contains a detailed discussion of the method employed in the case of a simple mechanical model. In \[Apendice1\] and \[Apendice2\], we explicitly show that the transformations between the canonical variables of the MGT and those of the ENM yield the canonical algebra of the former starting from the canonical algebra of the latter. To this end we also need to calculate the algebra of the second class constraints of the ENM, which is relevant to establish the relation between the canonical algebras of both theories. Mother Gauge Theories {#gaugetheory} ===================== In this section we state the conditions that define the class of MGT that we are going to deal with. We will show that these conditions allow for the construction of an ENM, which ultimately turns out to be equivalent to the original MGT, after some conditions are imposed. We will prove this by construction. The MGT is defined by the Lagrangian density $$\mathcal{L}=\mathcal{L}(\theta_{l},\lambda _{A},\dot{\theta}_{l}),\quad \quad \quad \quad l=1,2,...,n.\quad \quad \quad A=1,2,...,K. \label{Lagrangian}$$ where both $\theta_{l}$ and $\lambda _{A}$ are independent field coordinates depending upon the spacetime labels $t, {\mathbf x}=\{x_a; a=1,2,3\}$. To maintain the notation simple, in the following we do not write the spacetime dependence of the fields, unless some confusion arises. Here $ \mathcal{L}$ is independent of the velocities $\dot{\lambda}_{A}$. To consider the coupling with external currents, we can split the Lagrangian density as $$\mathcal{L}= \mathcal{L}_0(\theta_{l},\lambda _{A},\dot{\theta}_{l})-(\theta_lJ_l+\lambda_A J_A).$$ The term $-(\theta_lJ_l+\lambda_A J_A)$ is a standard way to couple an external current, which looks like [$-{\rm Tr}(\mathbf{A}_\mu \mathbf{J}^\mu$)]{} in the covariant form of the Yang Mills case. We are going to deal with theories where a generalized form of current conservation is satisfied and we will require to establish how this property is expressed in our [case]{}. Notice that the functional form of $\mathcal{L}$ is completely arbitrary. The coordinates $\lambda _{A}$ play a similar rôle to the standard Lagrange multipliers in a gauge theory and ensure that primary constrains appear. As usual, the canonical momenta are given by $$\Pi _{l}^{\theta}=\frac{\partial \mathcal{L}}{\partial \dot{\theta}_{l}},\quad \quad \quad \quad \Pi _{A}^{\lambda }=\frac{\partial \mathcal{L}}{\partial \dot{ \lambda}_{A}}, \label{moments}$$ with the [equal time]{} non-zero canonical Poisson brackets (PBs) algebra $$\{\theta_{l}(\mathbf{x}),\Pi _{k}^{\theta}(\mathbf{y})\}=\delta _{lk} \delta^3(\mathbf{x}-\mathbf{y})\,,\quad \quad \{\lambda _{A}(\mathbf{x}),\Pi _{B}^{\lambda}(\mathbf{y})\}=\delta _{AB} \delta^3(\mathbf{x}-\mathbf{y}). \label{commutationfree}$$ In general we will dispense of the coordinates dependence of the PB’s and we will write $\{\lambda _{A},\Pi _{B}\}=\delta _{AB}$, for example. This encodes the implicit assumption that the corresponding subindexes also carry the space coordinate dependence. [In order to have a familiar perspective of the procedure, we will keep in mind the [four dimensional]{} Yang Mills case as a guiding example. Here we identify [$\theta_{l}\rightarrow A_{a}^{\alpha}$, $\lambda _{A}\rightarrow A_{0}^{\alpha}$, $\Pi_l^\theta \rightarrow \Pi^\alpha_a=-E^\alpha_a$, $\Pi^\lambda_A \rightarrow \Pi^\alpha_0$, $J_l \rightarrow J^\alpha_a$ and $J_A \rightarrow J_0^\alpha$ where $\alpha$ and $ (0,a=1,2,3)$ ]{}denote the indices in the gauge group and in the Lorentz space, respectively.]{} Given that the velocities $\dot{\lambda}_{A}$ are not present in $\mathcal{L} $, the theory has primary constraints $\Phi _{A}^{1}=\Pi _{A}^{\lambda }=0$ and we take the standard Dirac method to perform the Hamiltonian analysis, starting from the canonical Hamiltonian density $$\mathcal{H}_{c}=\dot{\theta}_{l}\Pi _{l}^{\theta}-\mathcal{L}.$$ As a matter of notation, for any label $Z$, we will always write $H_Z$ for the Hamiltonian corresponding to the Hamiltonian density ${\mathcal H}_Z(y)$, i. e. $H_Z=\int d^3y {\mathcal H}_Z(y) $. From now on, we restrict ourselves to MGTs satisfying the next three conditions: 1. The canonical Hamiltonian density is linear in the coordinates $ \lambda _{A}$ and can be written as follows $$\mathcal{H}_{c}=\mathcal{H}_{F}(\theta_{i},\Pi _{i}^{\theta})+\lambda _{A}(G^0_{A}(\theta_{i},\Pi _{i}^{\theta})+ J_A) +\theta_l J_l. \label{hamiltonianocanonic}$$ [The corresponding Hamiltonian is $H_c$]{}. We refer to the quantities [$$G_{A}(\theta_{i},\Pi _{i}^{\theta})\equiv G^0_{A}(\theta_{i},\Pi _{i}^{\theta}) +J_A$$ ]{} as Gauss functions. They will turn out to be FCC in the MGT (the corresponding Gauss laws), but not in the ENM. Let us emphasize that $J_A$ and $J_l$, being external currents, do not play any rôle in the calculation of PBs. 2. The only primary constraints are $$\Phi _{A}^{1}=\Pi _{A}^{\lambda }=0, \label{condition2}$$ i.e., the momenta canonically conjugated to the coordinates $\lambda_{A}$. 3. The conditions $$\begin{aligned} &&\{G^0_{A}(\mathbf{x}),\mathcal{H}_{F}(\mathbf{y}\}=C_{AB}G^0_{B}(\mathbf{x})\delta^3(\mathbf{x}-\mathbf{y}),\quad \quad \nonumber \\ &&\{G^0_{A}(\mathbf{x}),G^0_{B}(\mathbf{y})\}=C_{ABC}G^0_{C}(\mathbf{x}) \delta^3(\mathbf{x}-\mathbf{y}), \label{condition3}\end{aligned}$$ hold at equal times. For example, in the Yang Mills case we have [$G_{A}(\theta_{i},\Pi _{i}^{\theta})\rightarrow -(D_aE_a-J_0)^\alpha$]{} with [$G^0_A\rightarrow -(D_aE_a)^\alpha$]{}. [Here $D_a$ denotes the corresponding covariant derivative.]{} We shall prove that the conditions (\[hamiltonianocanonic\]), (\[condition2\]) and (\[condition3\]) are enough to ensure that we are dealing with a theory having only first class constraints. We also demand that the constraints $G_A$ generate non-trivial gauge symmetry transformations, which must be verified in each particular case. In this way, such transformations will lead to the gauge invariance of the action. We assume a positive answer in the following. At the same time, a full Hamiltonian analysis, respecting such conditions, will be done in order to determine: the number of degrees of freedom, the identification of the constraints, the extended Hamiltonian density and the canonical algebra. Following the Dirac method, the extended Hamiltonian density is $$\mathcal{H}_{E}=\mathcal{H}_{c}+\beta _{A}\Pi _{A}^{\lambda }\;=\mathcal{H} _{F}+\lambda _{A}G_{A}+\theta_lJ_l+\beta _{A}\Pi _{A}^{\lambda },$$ where $\beta _{A}$ are arbitrary functions. [The corresponding Hamiltonian is $H_E$]{}. The time evolution condition of the primary constraints $\Phi _{A}^{1}=\Pi _{A}^{\lambda }$ yields $$\dot{\Phi}_{A}^{1}(\mathbf{x})=\{\Pi _{A}^{\lambda }(\mathbf{x}), H_E \}=\int d^{3}y\,\{\Pi _{A}^{\lambda }( \mathbf{x}),\lambda _{B}(\mathbf{y})\}G_{B}(\mathbf{y})=-G_{A}(\mathbf{x}),$$ i.e, to the secondary constraints $\Phi _{A}^{2}=G_{A}$. Let us remark that the constraints turn out to be the $G_{A}$’s instead of the $G^0_{A}$’s, in such a way that in some cases it is useful to express the results in the right hand side of Eq.(\[condition3\]) in terms of the former. Given that $ G_{A}=G_{A}(\theta_{i},\Pi _{i}^{\theta})$, we have $\{\lambda _{B},G_{A}\}=0$ and $ \{\Pi _{B}^{\lambda },G_{A}\}=0$. Considering Eq. (\[condition3\]), it follows that the time evolution condition of the secondary constraints $\dot{ \Phi}_{A}^{2}$ is $$\dot{\Phi}_{A}^{2}=\dot{G}_A=\{G_{A},{H}_{E}\}\approx \dot{J}_A+ C_{AB}G^0_{B} +{C}_{ABC}\lambda_B G^0_{C}+\int d^3y \, \{G^0_{A},\theta_l(\mathbf y)\}J_l(\mathbf{y}), \label{gaussbracket}$$ Introducing the constraints $G_A \approx 0$ in the above equation we can write $\dot{G}_A \approx (DJ)_A$, where $$(DJ)_A \equiv \dot{J}_A-C_{AB}J_B- C_{ABC}\lambda_BJ_C+\int d^3y \, \{G^0_{A},\theta_l(\mathbf y)\}J_l(\mathbf{y}). \label{DEFDJA}$$ In \[CURCONS1\] we show that $$(DJ)_A=0, \label{currentcons}$$ as a consequence of the invariance of the Hamiltonian action under the gauge transformations generated by the first class constraints $G_A$, which correspond to the generalized Gauss laws. In this way $(DJ)_A=0$ is the generalized version of current conservation in our model. Once the explicit form of the generators $G_A$ is given, one can recover the standard expressions in terms of the covariant derivatives, for example. A detailed calculation of the right hand side of Eq.(\[DEFDJA\]) in the case of a Yang-Mills theory can be found in the Appendix A of Ref. . Under the condition (\[currentcons\]), the quantities $\dot{\Phi}_{A}^{2}$ are weakly equal zero and there are no more constraints in the theory. We remark that the PBs among the quantities $(\Pi _{A}^{\lambda },G_{B})$ are zero or weakly zero, leading to the appearance of $2K$ first class constraints. The presence of these quantities implies that we are dealing with a proper gauge theory. We have $n+K$ variables in coordinate space and $ 2K$ first class constraints, which implies that the number of degrees of freedom (DOF) is $$\#\text{DOF}=\frac{1}{2}\bigg(2(n+K)-2(2K)\bigg)=n-K. \label{dofgauge}$$ For a Yang-Mills theory, when the number of generators of the gauge group is $N$, we have $n=3N$ and $K=N$, yielding the correct number of $2N$ DOF. In general, the presence of first class constraints implies unphysical degrees of freedom, which might be conveniently removed to obtain the reduced phase space. To this end we have to fix the gauge by imposing as many suitable gauge constraints as the number of first class constraints that we want to eliminate. These gauge constraints have to be admissible and should convert the set of gauge conditions plus the set of first class constraints to be eliminated into a set of second class constraints. Subsequently, each constraint in this set is set strongly equal to zero, after the introduction of the corresponding Dirac brackets (DBs). In our case we choose to eliminate the variables $\lambda _{A}$ and $\Pi _{B}^{\lambda}$. The gauge is partially fixed by adding to Eq. [(\[condition2\])]{} the constraints [$$\Phi _{A}^{3}=\lambda _{A}-\Theta _{A}\approx 0,$$ ]{} where $ \Theta _{A}$ are arbitrary functions to be consistently determined after the remaining first class constraints $G_{A}$ are fixed. The constraints $\Phi _{A}^{3}$ and $\Phi _{B}^{1}=$ $\Pi _{B}^{\lambda }$ become in fact second class, i.e. the matrix $Q_{AB}=\{\Phi _{A}^{1},\Phi _{B}^{3}\}$ can be inverted. Fixing strongly $\Phi _{A}^{1}=0$ and $\Phi _{B}^{3}$ $=0\;$we obtain, after introducing the Dirac brackets, the partially reduced Hamiltonian density $$\mathcal{H}_{E}=\mathcal{H}_{F}+\Theta _{A}G_{A}+\theta_lJ_l. \label{hamiltonianofinal}$$ In order to compute the corresponding Dirac brackets $\{\mathcal{A},\mathcal{ B}\}_{D}$, we require the $2N\times 2N$ matrix constructed with the PBs of the constraints $\Phi_A^1$ and $\Phi_B^3$ $$M=\left( \begin{array}{cc} 0 & Q \\ -Q^{T} & R \end{array} \right) , \label{MATRIX}$$ where $Q=[Q_{AB}]$, $R=[R_{AB}]$ with $Q_{AB}=\{\Phi _{A}^{1},\Phi _{B}^{3}\} $, $R_{AB}=\{\Phi _{A}^{3},\Phi _{B}^{3}\}$. Here we have made use of the PBs $\{\Phi _{A}^{1},\Phi _{B}^{1}\}=0=\{\Pi _{A}^{\lambda },\Pi _{B}^{\lambda }\}$. The inverse matrix is given by $$M^{-1}=\left( \begin{array}{cc} (Q^{T})^{-1}RQ^{-1} & -(Q^{T})^{-1} \\ Q^{-1} & 0 \end{array} \right) . \label{inversa}$$ Let us clarify another point in the notation. The matrix elements in Eq. (\[MATRIX\]), for example, also carry space-coordinate labels which are suppressed. In other words, $M_{AB}$ really stands for the equal-time object $M_{AB}(t, {\mathbf x};t,{\mathbf y})$. This is relevant for matrix multiplication where the product $M_{AB}=P_{AC}Q_{CB}$ corresponds to $M_{AB}(t,{\mathbf x}; t,{\mathbf y})=\int d^3 z \, P_{AC}(t,{\mathbf x};t, {\mathbf z})Q_{CB}(t, {\mathbf z}; t,{\mathbf y})$. The matrix elements are evaluated at fixed time $t$, so that the notation can be further simplified to $M_{AB}(x,y)=\int d^3z \, P_{AC}(x,z)Q_{CB}(z,y)$. The equal-time Dirac brackets are defined as $$\{\mathcal{A}(\mathbf{x}),\mathcal{B}(\mathbf{y})\}_{D}=\{\mathcal{A}( \mathbf{x}),\mathcal{B}(\mathbf{y})\}-\int \,d^{3}u\,d^{3}v\{\mathcal{A}( \mathbf{x}),\chi _{i}(\mathbf{u})\}(M^{-1})^{ij}\{\chi _{j}(\mathbf{v}), \mathcal{B}(\mathbf{y})\},$$ where $\chi _{j}$ denote any of the constraints $\Phi _{A}^{1}$ and $\Phi _{B}^{3}$. Using $M^{-1}$ given in (\[inversa\]), together with the fact that $\{\Phi _{A}^{1},\theta_{j}\}=\{\Phi _{A}^{1},\Pi _{j}^{\theta}\}=0$ we obtain the canonical algebra for the remaining variables $\theta_{i}$ and $\Pi _{j}^{\theta}$ $$\{\theta_{i},\theta_{j}\}_{D}=0,\quad \quad \quad \{\Pi^\theta _{i},\Pi _{j}^{\theta}\}_{D}=0,\quad \quad \quad \{\theta_{i},\Pi _{j}^{\theta}\}_{D}=\delta _{ij}. \label{algebrafinal}$$ We could further fix the constraints $G_{A}$, however at this point the Hamiltonian (\[hamiltonianofinal\]) together with the canonical algebra (\[algebrafinal\]) are sufficient to determine the dynamics of the MGT. The Extended Nambu Model {#NGmodel} ======================== In this section we define the extended Nambu model (ENM) associated with the MGT previously introduced. Our main goal is to establish the equivalence between both models by finding which additional conditions have to be imposed upon the ENM in order that its Hamiltonian and canonical algebra reduce to (\[hamiltonianofinal\]) and (\[algebrafinal\]), respectively. We will prove that the aforementioned equivalence can be reached provided that the conditions (\[hamiltonianocanonic\]), (\[condition2\]) and (\[condition3\]) are fulfilled. The ENM is given by the same Lagrangian (\[Lagrangian\]) defining the MGT $$\mathcal{L}=\mathcal{L}(\theta_{l},\lambda _{A},\dot{\theta}_{l}),\quad \quad \quad l=1,...,n.\quad \quad \quad A=1,...,K. \label{lag2}$$ plus one constraint $$F(\theta_{l},\lambda _{A})=0. \label{constriccion}$$ In the standard abelian Nambu model, $\mathcal{L}=\mathcal{L}(A_{\mu })= \mathcal{L}(A_{i},A_{0},\dot{A}_{i})$ corresponds to the Maxwell Lagrangian density, $\theta_{i}\rightarrow A_{i}$, $\lambda \rightarrow A_{0}$ and $F(A_{\mu })=A_{\mu }A^{\mu }-n^{2}M^{2}$. Here, we consider a wider class of Lagrangian densities $\mathcal{L}$ and constraints $F$ to define the ENM. As in the [case of the]{} MGT, [where external currents are considered]{}, the Lagrange density and the quantities $G_A$ will be split into $\mathcal{L}=\mathcal{L}_0-(\lambda_AJ_A+\theta_lJ_l)$ and $G_A=G_A^0+J_A$, respectively. For the moment, we deal with the case where $F(\theta_{i},\lambda _{A})=0$ includes all the variables $\lambda _{A}$. The function $F=F(\theta_l,\lambda_A)$ is not completely arbitrary but must satisfy some general conditions to be determined in the \[Apendice1\] and [the]{} \[Apendice2\], depending how the constraint F is solved. In comparison with the MGT, the introduction of the additional constraint [in Eq. (\[constriccion\])]{}, which is not [handled]{} as a gauge fixing [condition]{}, drastically modifies the structure and the dynamics of the [ENM]{}. In this way, it is not completely straightforward how an equivalence between the ENM and the MGT can be established. To mention just an example, significant differences between the ANM and the standard electrodynamics are: (i) the ANM has only second class constraints (there is no gauge invariance), (ii) the number of DOF in the ANM is three, while standard electrodynamics has only two DOF, (iii) the equations of motion do not match and (iv) current conservation does not follow from the equations of motion in the ANM. The general procedure through which we analyze the ENM is by solving explicitly the constraint (\[constriccion\]) for one variable and substituting this solution into the Lagrangian density (\[lag2\]). There are several ways to solve Eq. (\[constriccion\]) and we present the two generic cases, which reduce to solve either for one coordinate $\theta_l$ or for one coordinate $\lambda_A$. Both cases yield the same conditions for the equivalence we intend to establish. The general strategy is: (1) After solving the constraint (\[constriccion\]) we identify the canonical variables of the ENM, together with its canonical algebra. (2) Since both theories arise from a common Lagrangian it is possible to write the canonical variables of the MGT in terms of those of the ENM. (3) Through these substitutions we find that: (i) the canonical algebra of the MGT can be derived from the canonical algebra of the ENM and (ii) the Hamiltonian density of the ENM reduces to the Hamiltonian density of the MGT. (4) Nevertheless, at this stage the Gauss functions $G_A$ in the ENM are not constraints. In this way we need to impose additional conditions in order to recover gauge invariance. (5) These are realized by demanding the Gauss functions to be zero at some initial time and by recognizing that the dynamics of the ENM is consistent with this requirement, so that they become zero for all time. In this way, they can be added to the Hamiltonian density of the ENM as first class constraints, thus fully recovering the gauge invariance of the MGT. As we mentioned above, since the ENM is not gauge invariant, the conservation of an external current is not guaranteed [a priori]{}, therefore it has to be imposed as an additional condition. We will show that in some cases current conservation follows from the imposition of the Gauss constraints as initial conditions. Solving the coordinate $\theta_1$ {#x1} --------------------------------- In this case, we solve the constraint (\[constriccion\]) as $$\theta_{1}=f(\theta_{\bar{l}},\lambda _{A}),\quad \quad \quad \quad \bar{l}=2,...,n, \label{constriction2}$$ which yields $$\quad \quad \quad \dot{\theta}_{1}=\;\frac{\partial f}{\partial \theta_{\bar{l}}}\dot{\theta }_{\bar{l}}+\frac{\partial f}{\partial \lambda _{A}}\dot{\lambda}_{A},=f_{\theta_{ \bar{l}}}\dot{\theta}_{\bar{l}}+f_{\lambda _{A}}\dot{\lambda}_{A}, \label{hamiltonianconstric}$$ where the time derivative is denoted by an overdot. Substituting the relations (\[constriction2\]) and (\[hamiltonianconstric\]) directly in the Lagrangian density (\[lag2\]), we obtain $$\bar{\mathcal{L}}(\theta_{\bar{l}},\lambda _{A},\dot{\theta}_{\bar{l}},\dot{\lambda} _{A})=\mathcal{L}(\theta_{1}(\theta_{\bar{l}},\lambda _{A}),\theta_{\bar{l}},\lambda _{A}, \dot{\theta}_{1}(\theta_{\bar{l}},\lambda _{A},\dot{\theta}_{\bar{l}},\dot{\lambda}_{A}),\; \dot{\theta}_{\bar{l}}),$$ ### Hamiltonian and canonical algebra In this section we [make]{} explicit the relation among the canonical coordinates of the ENM and those of the MGT. Using [such relation]{} we show that: (i) the canonical algebra of the MGT is derived from the canonical algebra of the ENM and (ii) the form of the canonical Hamiltonian density of the MGT, given in the previous section, is obtained from the canonical Hamiltonian density of the ENM. In this case the independent coordinates of the ENM are $\theta_{\bar{l}}$ and $ \lambda _{A}$. We note that, the quantities $\dot{\lambda}_{A}$ appear only as the result of imposing the constraint (\[constriction2\]) by means of the substitution of the velocity $\dot{\theta}_{1}$. Therefore, the momenta associated to the $\lambda _{A}$ variables are not zero in the ENM. Note that in the ENM, $\theta_{1}$ and $\dot{\theta}_{1}$ are just labels to specify a particular combination of the coordinates $\theta_{\bar{l}}$ and $\lambda _{A}$ and velocities $\dot{\theta}_{\bar{l}}$ and $\dot{\lambda}_{A} $. After the substitutions (\[constriction2\]) and (\[hamiltonianconstric\]), the Lagrange densities $\bar{ \mathcal{L}}$ and $\mathcal{L}$ have different functional form; however, let us emphasize that the labels $\theta_{1}$ and $\dot{\theta}_{1}$ in the ENM allow us to write, for example, $$\frac{\partial \bar{\mathcal{L}}}{\partial \lambda _{A}}=\frac{\partial \mathcal{L}}{\partial \theta_{1}}\frac{\partial \theta_{1}}{\partial \lambda _{A}}+ \frac{\partial \mathcal{L}}{\partial \lambda _{A}},$$ making use of the chain rule, with $\mathcal{L}=\mathcal{L}(\theta_{l},\lambda _{A},\dot{\theta}_{l})$. Relations of this kind will prove very useful to compare the Hamiltonian structure of the ENM with that of the MGT. The canonical momenta for the ENM are given by $$\begin{aligned} \bar{\Pi}_{\bar{l}}^{\theta} &\equiv &\frac{\partial \bar{\mathcal{L}}}{\partial \dot{\theta}_{\bar{l}}}=\left[ \frac{\partial \mathcal{L}}{\partial \dot{\theta}_{\bar{ l}}}\right] _{\theta_{1}=f}+\left[ \frac{\partial \mathcal{L}}{\partial \dot{\theta} _{1}}\right] _{\theta_{1}=f}\frac{\partial \dot{\theta}_{1}}{\partial \dot{\theta}_{\bar{l}} }=\left[ \frac{\partial \mathcal{L}}{\partial \dot{\theta}_{\bar{l}}}\right] _{\theta_{1}=f}+\left[ \frac{\partial \mathcal{L}}{\partial \dot{\theta}_{1}}\right] _{\theta_{1}=f}f_{\theta_{\bar{l}}}\,,\quad \label{momenta2} \\ \quad \bar{\Pi}_{A}^{\lambda } &\equiv &\frac{\partial \bar{\mathcal{L}}}{ \partial \dot{\lambda}_{A}}=\left[ \frac{\partial \mathcal{L}}{\partial \dot{ \theta}_{1}}\right] _{\theta_{1}=f}\frac{\partial \dot{\theta}_{1}}{\partial \dot{\lambda} _{A}}=\left[ \frac{\partial \mathcal{L}}{\partial \dot{\theta}_{1}}\right] _{\theta_{1}=f}f_{\lambda _{A}}, \label{momenta3}\end{aligned}$$ where we have obtained $\partial \dot{\theta}_{1}/\partial \dot{\theta}_{\bar{l}}$ and $\partial \dot{\theta}_{1}/\partial \dot{\lambda}_{A}$ from (\[hamiltonianconstric\]). The notation is $f_y=\partial f/\partial y$. We assume that from Eqs. (\[momenta2\]) we can invert the velocities $\dot{\theta} _{\bar{l}}$ in terms of the momenta $\bar{\Pi}_{\bar{l}}^{\theta}$, so these equations do not define constraints of the theory. On the [other hand]{}, after eliminating $\left[ \partial \mathcal{L}/\partial \dot{\theta}_{1}\right] _{\theta_{1}=f}$ in favor of $\bar{\Pi}_{1}^{\lambda }$, the Eqs. (\[momenta3\]) provide $(K-1)$ primary constraints which we choose as $$\phi _{\bar{A}}^{1}=\bar{\Pi}_{\bar{A}}^{\lambda }-\frac{\bar{\Pi} _{1}^{\lambda }}{f_{\lambda _{1}}}f_{\lambda _{\bar{A}}}\approx 0,\;\;\;\;\; \bar{A}=2,3,...,K.$$ Let us recall that $$\frac{\partial \mathcal{L}}{\partial \dot{\theta}_{1}}\equiv \Pi _{1}^{\theta},\;\;\;\;\;\;\frac{\partial \mathcal{L}}{\partial \dot{\theta}_{\bar{l}}} \equiv \Pi _{\bar{l}}^{\theta}, \label{GTMOM}$$ define the corresponding canonically conjugated momenta in the MGT [arising]{} from the Lagrangian density $\mathcal{L}=\mathcal{L}(\theta_{l},\lambda _{A},\dot{\theta}_{l})$. In this way, when substituting $\theta_{1}=f(\theta_{\bar{l}},\lambda _{A})$ in the expressions for $\frac{\partial \mathcal{L}}{\partial \dot{\theta}_{1}}$ and $ \frac{\partial \mathcal{L}}{\partial \dot{\theta}_{\bar{l}}}$ of Eqs. (\[momenta2\]), $\Pi _{1}^{\theta}\;$and $\Pi _{\bar{l}}^{\theta}\;$become just labels used to rewrite a particular combination of coordinates and velocities in the ENM. However, when we reinstate the notation in terms of $\theta_{1}$ and $ \dot{\theta}_{1}$, i.e. when going back to $\mathcal{L}=\mathcal{L}(\theta_{l},\lambda _{A},\dot{\theta}_{l}),$ they recover their definition as the canonically conjugated momenta corresponding to the MGT. In other words, the relations (\[momenta2\]) allow us to relate the canonically conjugated momenta of the ENM (labeled as $\bar{\Pi}$) with those of the MGT (labeled as $\Pi $) in the following way. $$\bar{\Pi}_{\bar{l}}^{\theta}=\Pi _{\bar{l}}^{\theta}+\Pi _{1}^{\theta}f_{\theta_{\bar{l} }}\,,\quad \quad \quad \quad \bar{\Pi}_{A}^{\lambda }=\Pi _{1}^{\theta}f_{\lambda _{A}}. \label{momentaNG}$$ In the ENM the coordinates and canonical momenta ($\theta_{\bar{l}},\lambda _{A},\bar{\Pi}_{ \bar{l}}^{\theta},\bar{\Pi}_{A}^{\lambda }$) satisfy the non-zero canonical PBs $$\{\theta_{\bar{\imath}},\bar{\Pi}_{\bar{j}}^{\theta}\}=\delta _{{\bar{\imath}}{\bar{j}} },\quad \quad \quad \{\lambda _{A},\bar{\Pi}_{B}^{\lambda }\}=\delta _{AB}. \label{commutation_constric}$$ From (\[momentaNG\]) we can express the labels $\Pi _{1}^{\theta}$ and $\Pi _{ \bar{l}}^{\theta}$ in terms of the canonical momenta $\bar{\Pi}_{\bar{l}}^{\theta}$ and $\bar{\Pi}_{l}^{\lambda }$ of the ENM, as $$\Pi _{\bar{l}}^{\theta}=\bar{\Pi}_{\bar{l}}^{\theta}-\bar{\Pi}_{1}^{\lambda }\frac{ f_{\theta_{\bar{l}}}}{f_{\lambda _{1}}}\,,\quad \quad \quad \quad \Pi _{1}^{\theta}= \frac{\bar{\Pi}_{1}^{\lambda }}{f_{\lambda _{1}}}. \label{momentaTTNM}$$ In the \[Apendice1\], we show that when considering $\Pi _{1}^{\theta}$ and $\Pi _{\bar{l}}^{\theta}$ as labels in terms of the canonical variables of the ENM $(\theta_{\bar{l}},\bar{\Pi}_{\bar{l}}^{\theta},\lambda _{A},\bar{ \Pi}_{A}^{\lambda })$, we obtain the following PB algebra $$\{\theta_{i},\theta_{j}\}=0,\quad \quad \quad \{\Pi _{i}^{\theta},\Pi _{j}^{\theta}\}=0,\quad \quad \quad \{\theta_{i},\Pi _{j}^{\theta}\}=\delta _{ij},$$ by using the canonical algebra (\[commutation\_constric\]) of the ENM. The calculation of the Hamiltonian density $\mathcal{H}_{c}^{ENM}$ of the ENM gives $$\begin{aligned} \mathcal{H}_{c}^{ENM} &=&\dot{\lambda}_{A}\bar{\Pi}_{A}^{\lambda }+\dot{\theta}_{ \bar{l}}\bar{\Pi}_{\bar{l}}^{\theta}-\bar{\mathcal{L}}, \notag \\ &=&\dot{\lambda}_{A}\Pi _{1}^{\theta}f_{\lambda _{A}}+\dot{\theta}_{\bar{l}}(\Pi _{ \bar{l}}^{\theta}+\Pi _{1}^{\theta}f_{\theta_{\bar{l}}})-\bar{\mathcal{L}}, \notag \\ &=&\Pi _{1}^{\theta}(f_{\lambda _{A}}\dot{\lambda}_{A})+\dot{\theta}_{\bar{l}}(\Pi _{ \bar{l}}^{\theta}+\Pi _{1}^{\theta}f_{\theta_{\bar{l}}})-\bar{\mathcal{L}}, \notag \\ &=&\Pi _{1}^{\theta}(\dot{\theta}_{1}-f_{\theta_{\bar{l}}}\dot{\theta}_{\bar{l}})+\dot{\theta}_{\bar{l }}(\Pi _{\bar{l}}^{\theta}+\Pi _{1}^{\theta}f_{\theta_{\bar{l}}})-\bar{\mathcal{L}}, \notag \\ &=&\dot{\theta}_{1}\Pi _{1}^{\theta}+\dot{\theta}_{\bar{l}}\Pi _{\bar{l}}^{\theta}-\bar{\mathcal{ L}}, \notag \\ &=&\dot{\theta}_{k}\Pi _{k}^{\theta}-\bar{\mathcal{L}},\quad \quad \quad \quad \quad \quad \quad \quad \quad \quad (k=1,...,n), \label{HCENM}\end{aligned}$$ where we have substituted $f_{\lambda _{A}}\dot{\lambda}_{A}$ from Eq. (\[hamiltonianconstric\]) and $(\bar{\Pi}_{\bar{l}}^{\theta},\bar{\Pi}_{A}^{\lambda })$ from Eq. (\[momentaNG\]). Undoing the substitution (\[constriction2\]), that is to say, inserting back the original variables $\theta_{1}$ and $\dot{\theta}_{1}$ in Eq. (\[HCENM\]), we realize that $\bar{\mathcal{L}}$ reduces to $\mathcal{L}$ and that $\Pi _{k}^{\theta}$ are the corresponding canonically conjugated momenta of the MGT, according to Eqs. (\[GTMOM\]). In this way, the Hamiltonian density (\[HCENM\]) has the same form that [${\mathcal H}_{c}$]{} given in Eq. (\[hamiltonianocanonic\]) for the MGT. Next, we consider the extended Hamiltonian density which is given by $$\begin{aligned} \mathcal{H}_{E}^{ENM} &=&\mathcal{H}_{c}^{ENM}+\mu _{\bar{A}}\phi _{\bar{A} }^{1}, \notag \\ &=&\mathcal{H}_{F}+\lambda _{A}G_{A}+\theta_lJ_l+\mu _{\bar{A}}\left( \bar{\Pi}_{\bar{A} }^{\lambda }-\frac{\bar{\Pi}_{1}^{\lambda }}{f_{\lambda _{1}}}f_{\lambda _{ \bar{A}}}\right) ,\end{aligned}$$ where we have used that $\mathcal{H}_{c}^{ENM}$ has the same form that $ \mathcal{H}_{c}$ in (\[hamiltonianocanonic\]) and $\mu _{\bar{A}}$ are arbitrary functions. In the \[Apendice1\], we show that $\{\phi _{ \bar{A}}^{1},\theta_{l}\}=0$, $\{\phi _{\bar{A}}^{1},\Pi _{l}^{\theta}\}=0$ and $ \{\phi _{\bar{A}}^{1},\phi _{\bar{B}}^{1}\}=0$. Using the previous results, the time evolution condition of the primary constraints leads to $$\dot{\phi}_{\bar{A}}^{1}=\{\phi _{\bar{A}}^{1},{H} _{E}^{ENM}\}={\int d^3y \{\phi _{\bar{A}}^{1},\lambda _{\bar{B}}(y)\}G_{\bar{B}}(y)}=\left( G_{ \bar{A}}-G_{1}\frac{f_{\lambda _{\bar{A}}}}{f_{\lambda _{1}}}\right) ,$$ where we identify the secondary constraints $$\phi _{\bar{A}}^{2}=f_{\lambda _{\bar{A}}}-\frac{G_{\bar{A}}}{G_{1}} f_{\lambda _{1}}\approx 0. \label{constrsecun}$$ Next, we calculate the time evolution of $\phi _{\bar{A}}^{2}$, with the result $$\begin{aligned} \dot{\phi}_{\bar{A}}^{2}=\{\phi _{\bar{A}}^{2},{{H}}_{E}^{ENM}\} =U_{\bar{A}}-\mu _{\bar{B}}T_{\bar{B}\bar{A}}\approx 0, \end{aligned}$$ with $U_{\bar{A}}=\{\phi _{\bar{A}}^{2},{H}_{c}^{ENM}\}$ and $T= \left[ T_{\bar{A}\bar{B}}\right] =\left[ \{\phi _{\bar{A}}^{1},\phi _{\bar{B} }^{2}\}\right] $. Unless confusion arises, here and in the following we make use of the convention introduced after Eq. (\[inversa\]), whereby $\mu _{\bar{B}}T_{\bar{B}\bar{A}}$ is given by $\int d^3 y \, \mu_{\bar{B}}(y)T_{\bar{B}\bar{A}}(y,x)$, for example. In the \[Apendice1\], we show that it is possible to choose particular functions $f(\theta_{\bar{l}},\lambda _{A})$ such that, the matrix $T_{ \bar{A}\bar{B}}$ is invertible, which means that $\phi _{\bar{A}}^{1}$ and $ \phi _{\bar{B}}^{2}$ are second class constraints. The arbitrary functions $ \mu _{\bar{B}}$ are fixed as $$\mu _{\bar{A}}=(T^{-1})_{\bar{B}\bar{A}}U_{\bar{B}}.$$ The Dirac method stops and the ENM only has the following $2(K-1)\;$second class constraints $$\phi _{\bar{A}}^{1}=\bar{\Pi}_{\bar{A}}^{\lambda }-\frac{\bar{\Pi} _{1}^{\lambda }}{f_{\lambda _{1}}}f_{\lambda _{\bar{A}}},\quad \quad \quad \phi _{\bar{A}}^{2}=f_{\lambda _{\bar{A}}}-f_{\lambda _{1}}\frac{G_{\bar{A}} }{G_{1}}. \label{constspacelike}$$ Thus, the number of DOF of the ENM is $$\#\text{DOF}=\frac{1}{2}[2(n-1+K)-2(K-1)]=n. \label{dofx1a}$$ As previously emphasized, the ENM is not a gauge invariant theory, it has only second class constraints, and the number of degrees of freedom is not the same as in the MGT. From Eq. (\[dofgauge\]), we can observe that the ENM has $K$ degrees of freedom more than the MGT, so that, if we want to establish an equivalence, we will have to impose $K$ additional constraints to the ENM. The next step is to set strongly equal zero the second class constraints (\[constspacelike\]). To this end we introduce the corresponding DBs and further calculate them among the remaining variables. We require the matrix constructed with the PBs of the constraints $$M=\left[ \begin{array}{cc} \left[ R_{\bar{A}\bar{B}}\right] & \left[ T_{\bar{A}\bar{B}}\right] \\ -\left[ T_{\bar{B}\bar{A}}\right] & \;\left[ S_{\bar{A}\bar{B}}\right] \end{array} \right] =\left[ \begin{array}{cc} R & T \\ -T^{T} & \;S \end{array} \right] , \label{SL_PP_MATRIX}$$ where $$R_{\bar{A}\bar{B}}=\left\{ \phi _{\bar{A}}^{1},\;\;\phi _{\bar{B} }^{1}\right\} ,\;\;\;T_{\bar{A}\bar{B}}=\left\{ \phi _{\bar{A}}^{1},\;\;\phi _{\bar{B}}^{2}\right\} ,\;\;\;\;S_{\bar{A}\bar{B}}=\left\{ \phi _{\bar{A} }^{2},\;\;\phi _{\bar{B}}^{2}\right\} \;. \label{SL_PP_ENTRIES}$$ In the \[Apendice1\] we show that $R_{\bar{A}\bar{B}}=0$. The inverse matrix becomes $$\;\;M^{-1}=\left[ \begin{array}{cc} T^{-1}ST^{-1} & -T^{-1} \\ T^{-1} & 0 \end{array} \right] .$$ The DBs are $$\begin{aligned} \{\mathcal{A},\mathcal{B}\}_{D} &=&\{\mathcal{A},\mathcal{B}\}-\{\mathcal{A} ,\phi _{\bar{A}}^{1}\}(T^{-1}ST^{-1})_{\bar{A}\bar{B}}\{\phi _{\bar{B}}^{1}, \mathcal{B}\} \notag \label{bracketDspace1} \\ &&+\{\mathcal{A},\phi _{\bar{A}}^{1}\}(T^{-1})_{\bar{A}\bar{B}}\{\phi _{\bar{ B}}^{2},\mathcal{B}\}-\{\mathcal{A},\phi _{\bar{A}}^{2}\}(T^{-1})_{\bar{A} \bar{B}}\{\phi _{\bar{B}}^{1},\mathcal{B}\},\end{aligned}$$ which again leads to the result $$\{\mathcal{A}(x),\mathcal{B}(y)\}_{D}=\{\mathcal{A}(x),\mathcal{B}(y)\}, \label{DBEPBSL}$$ for the coordinates $\theta_{j}$ and momenta $\Pi _{j}^{\theta}\;$written in terms of the canonical variables $\theta_{\bar{l}},\;\bar{\Pi}_{\bar{l}}^{\theta}\;$of the ENM. The above result arises from the fact that each of the additional PBs in (\[bracketDspace1\]) includes a contribution containing $\phi _{\bar{A}}^{1}$, which has zero PB with $\theta_{j}$ and $\Pi _{j}^{\theta}$, according to the results in \[Apendice1\]. [Using]{} the transformations [given in ]{} (\[constriction2\]) and (\[momentaTTNM\]) we can rewrite all quantities of the ENM in terms of the labels $\theta_{j}$ and $\Pi _{j}^{\theta}$. [Moreover, since]{} we know the canonical algebra that [the latter]{} satisfy, at some stage it becomes more convenient to employ such labels to describe the ENM, instead of its own canonical variables. The above analysis shows that we recover the canonical algebra (\[algebrafinal\]) of the MGT. Once the Lagrange multipliers $\mu _{\bar{A}}$ have been fixed and the constraints $\phi _{\bar{A}}^{1}\;$and $\phi _{\bar{A}}^{2}\;$have been imposed strongly, the Hamiltonian [density]{} of the ENM is given by $$\mathcal{H}_{E}^{ENM}=\mathcal{H}_{F}+\lambda _{A}G_{A}+\theta_lJ_l. \label{hamiltonianofinal2}$$ ### Conservation of the quantities $G_{A}$ Up to this stage we have only verified that, with the appropriate change of variables (\[constriction2\]) and (\[momentaTTNM\]), the canonical algebra of the ENM induces the canonical algebra of the related MGT. Under the same transformations, the canonical Hamiltonian of the ENM adopts the same form as that of the MGT. Now we explore the conditions under which the full MGT emerges from the ENM. At this point, two fundamental differences arise: (a) the quantities $\lambda _{A}$ in the extended Hamiltonian density (\[hamiltonianofinal2\]) are functions of the coordinates and momenta of the MGT, i. e. $\lambda_A=\lambda_A(\theta_l, \Pi^\theta_k)$ instead of been arbitrary functions, as it is required in the Hamiltonian density of the MGT, where they should correspond to the $\Theta _{A}$’s appearing in Eq. (\[hamiltonianofinal\]). Such relations arise after imposing strongly the second class constraints (\[constspacelike\]) of the ENM. (b) The Gauss functions $G_{A}$ in Eq. (\[hamiltonianofinal2\]) are not constraints in the ENM, while in the MGT they should be realized as first class constraints, being the generators of the gauge symmetry. To deal with these issues, we study the time evolution of the quantities $ G_{A}$ under the dynamics of the ENM $$\begin{aligned} \dot{G}_{A}&=&\{G_{A},{H}_{E}^{ENM}\}={{\dot J}_A}+\{G_{A},{H}_{F}\}+\int d^3y \lambda _{B}(y)\{G_{A},G_{B}(y)\} \notag \\ &&+\int d^3y ( \{G_A,\theta_l(y)\}J_l(y)+ \{G_{A},\lambda_{B}(y)\} G_{D}(y)). \label{GMDOT}\end{aligned}$$ Using the the equivalence between the canonical algebras, which we have already proved, we can employ the results arising from the MGT to evaluate the first two brackets in Eq. (\[GMDOT\]). From Eq. (\[condition3\]) we obtain $$\begin{aligned} \dot{G}_{A}=\dot{J}_A+{C}_{AB}G^0_{B}&+& C_{ABC}\lambda_B G_C^0 + \int d^3y (\{G_A,\theta_l(y)\}J_l(y)+\{G_{A},\lambda _{B}(y)\}G_{B}(y)), \nonumber \\ \dot{G}_{A} &=& {C}_{AB}G_{B}+ C_{ABC}\lambda_B G_C + \int d^3y \{G_{A},\lambda _{B}(y)\}G_{B}(y)\nonumber \\ &+& \dot{J}_A-{C}_{AB}J_{B}- C_{ABC}\lambda_B J_C + \int d^3y \{G^0_{A},\theta_l(y)\}J_l(y), \nonumber \\ \dot{G}_{A} &=& {C}_{AB}G_{B}+ C_{ABC}\lambda_B G_C + \int d^3y\{G_{A},\lambda _{B}(y)\}G_{B}(y)+(DJ)_A, \notag \\ \label{DOTGAENM1}\end{aligned}$$ where we have used the relation $G_A^0=G_A-J_A$. Let us recall that $(DJ)_A$ was defined in Eq. (\[DEFDJA\]). At this stage $\dot{G}_{A} \neq 0$ for two reasons: (1) the Gauss functions $G_A$ are not constraints and (2) the generalized current conservation $(DJ)_A=0$ is not valid due to the lack of gauge invariance in the ENM. Then, in order to have $\dot{G}_{A}=0$ it is enough to impose the following conditions upon the ENM: (i) $(DJ)_A=0$ for all times and (ii) $G_{A}=0$ as an initial condition, at $t=0$ for example, which also leads to ${\dot G}_A=0$ at $t=0$. In this way, Eq. (\[DOTGAENM1\]) yields $G_{A}=0$ for all time. Under the above conditions, we can recover the MGT by adding the quantities $G_{A}$ as Hamiltonian first class constraints, $G_{A}\approx 0$, with arbitrary functions $N_{A}$. This requires to add $N_{A}G_{A}$ to the Hamiltonian (\[hamiltonianofinal2\]) and to redefine $\lambda _{A}+N_{A}=\Theta _{A}$, which leads to $$\mathcal{H}_{E}=\mathcal{H}_{F}+\Theta _{A}G_{A}+\theta_lJ_l, \label{HDENM1}$$ where $\Theta _{A}$ are now arbitrary functions. In other words, we have regained the Hamiltonian (\[hamiltonianofinal\]) together with the canonical algebra (\[algebrafinal\]) of the MGT. Summarizing, the equivalence between the MGT and the ENM model can be established only after imposing the generalized current conservation $(DJ)_A=0$ for all times, together with $G_{A}=0$ as initial conditions in the dynamics of the ENM. As we have previously shown, the number of DOF of the ENM is $n$, but when the $K$ relations $G_{A}=0$ are imposed as first class constraints into the Hamiltonian density, the remaining theory has only $n-K$ DOF, which is the same number of DOF of the MGT. Solving the coordinate $\protect\lambda _{1}$ --------------------------------------------- A similar analysis can be performed for this case. We solve the constraint (\[constriccion\]) as $$\lambda _{1}=g(\theta_{i},\lambda _{\bar{B}})\quad \quad \quad \bar{B}=2,3,...,K. \label{const2}$$ in such a way that the canonical variables for the ENM are now $\theta_{i},\; \bar{\Pi}_{i}^{\theta},\;\lambda _{\bar{B}},\;\bar{\Pi}_{\bar{B}}^{\lambda }.\;$ The velocities $\dot{\lambda}_{A}$ do not appear in the Lagrangian (\[Lagrangian\]), and since the constraint (\[const2\]) does not introduce additional velocity dependent terms, we have $$\frac{\partial \bar{\mathcal{L}}}{\partial \dot{\theta}_{i}}=\bar{\Pi} _{i}^{\theta}=\Pi _{i}^{\theta},\quad \quad \quad \frac{\partial \bar{\mathcal{L}}}{ \partial \dot{\lambda}_{\bar{B}}}=\Pi _{\bar{B }}^{\lambda }=0, \label{constr123a}$$ where $\Pi _{i}^{\theta}$ and $\Pi _{\bar{B}}^{\lambda }$ are given by (\[moments\]) satisfying the same PBs indicated in (\[commutationfree\]). ### Hamiltonian and canonical algebra of the ENM In this case, $\lambda _{1}$ and $\Pi _{1}^{\lambda }$ are not independent canonical variables of the ENM, in fact $\lambda_1$ is a function of the remaining variables $\lambda _{1}=g(\theta_{i},\lambda _{\bar{B}})$ and $\Pi _{1}^{\lambda }$ is not included in the ENM. The calculation of the canonical Hamiltonian density proceeds just as in the previous Section \[gaugetheory\] and we find $$\begin{aligned} \mathcal{H}_{c}^{ENM} &=&\mathcal{H}_{F}+\lambda _{1}G_{1}+\lambda _{\bar{B} }G_{\bar{B}}+\theta_lJ_l, \notag \\ &=&\mathcal{H}_{F}\;+g(\theta_{m},\lambda _{\bar{B}})G_{1}+\lambda _{\bar{B}}G_{ \bar{B}}+\theta_lJ_l.\end{aligned}$$ The theory contains $(K-1)$ primary constraints $\phi _{\bar{B}}^{1}=\bar{\Pi }_{\bar{B}}^{\lambda }\approx 0$, $\bar{B}=2,3,...,K$. The extended Hamiltonian is given by $$\begin{aligned} \mathcal{H}_{E}^{ENM} &=&\mathcal{H}_{c}^{ENM}+\mu _{\bar{B}}\bar{\Pi}_{\bar{ B}}^{\lambda }, \\ &=&\mathcal{H}_{F}+g(\theta_{m},\lambda _{\bar{B}})G_{1}+\lambda _{\bar{B}}G_{ \bar{B}}+\theta_lJ_l+\mu _{\bar{B}}\bar{\Pi}_{\bar{B}}^{\lambda }, \notag\end{aligned}$$ where $\mu _{\bar{B}}$ are arbitrary functions. The time evolution condition for the primary constraints gives $$\begin{aligned} \dot{\phi}_{\bar{B}}^{1} =\{\phi _{\bar{B}}^{1},{H}_{E}^{ENM}\} =\{\Pi _{\bar{B}}^{\lambda },{H}_{E}^{ENM}\} =-(g_{\lambda _{\bar{B}}}G_{1}+G_{\bar{B}}), \end{aligned}$$ yielding $(K-1)$ secondary constraints, which we write as $$\phi _{\bar{B}}^{2}=g_{\lambda _{\bar{B}}}+\frac{G_{\bar{B}}}{G_{1}}\approx 0.$$ From the time evolution condition for the secondary constraints, we obtain $$\begin{aligned} \dot{\phi}_{\bar{B}}^{2}&=&\{\phi _{\bar{B}}^{2},{H} _{E}^{ENM}\}=\{\phi _{\bar{B}}^{2},{H}_{c}^{ENM}\} + \nonumber \\ && + {\int d^3y \,\mu _{\bar{A}}(y)\{\phi _{\bar{B }}^{2},\phi _{\bar{A}}^{1}(y)\}}=W_{\bar{B}}-\mu _{\bar{A}}X_{\bar{B}\bar{A}}\approx 0,\end{aligned}$$ where $W_{\bar{B}}=\{\phi _{\bar{B}}^{2},\mathcal{H}_{c}^{ENM}\}$ and $X_{ \bar{A}\bar{B}}=\{\phi _{\bar{A}}^{1},\phi _{\bar{B}}^{2}\}$. Once again, in the \[Apendice2\] we show that there exist particular functions $g(\theta_{m},\lambda _{\bar{B}})$ such that the matrix $X_{\bar{A}\bar{ B}}$ is invertible, therefore, the Lagrange multipliers $\mu _{\bar{B}}$ are fixed as $$\mu _{\bar{B}}=(X^{-1})_{\bar{B}\bar{A}}W_{\bar{A}}.$$ The existence of $(X^{-1})_{\bar{B}\bar{A}}$ guarantees that $\phi _{\bar{B} }^{1}$ and $\phi _{\bar{B}}^{2}$ are second class constraints and that the Dirac method stops. The number of DOF is $$\#\text{DOF}=\frac{1}{2}[2(n+K-1)-2(K-1)]=n. \label{dof1ab}$$ As usual, we set strongly equal zero the constraints $\phi _{\bar{B}}^{1}$ and $\phi _{\bar{B}}^{2}$ to subsequently introduce the corresponding DBs. We require the matrix constructed with the PBs of the constraints $$M=\left[ \begin{array}{cc} \left[ R_{\bar{A}\bar{B}}\right] & \left[ X_{\bar{A}\bar{B}}\right] \\ -\left[ X_{\bar{B}\bar{A}}\right] & \;\left[ S_{\bar{A}\bar{B}}\right] \end{array} \right] =\left[ \begin{array}{cc} R & X \\ -X^{T} & \;S \end{array} \right] ,$$ where $$R_{\bar{A}\bar{B}}=\left\{ \phi _{\bar{A}}^{1},\;\;\phi _{\bar{B} }^{1}\right\} ,\;\;\;X_{\bar{A}\bar{B}}=\left\{ \phi _{\bar{A}}^{1},\;\;\phi _{\bar{B}}^{2}\right\} ,\;\;\;\;S_{\bar{A}\bar{B}}=\left\{ \phi _{\bar{A} }^{2},\;\;\phi _{\bar{B}}^{2}\right\} \,.\;$$ Again, in the \[Apendice2\] we show that $R_{\bar{A}\bar{B} }=0$. The inverse matrix $M^{-1}$ is given by $$\;\;M^{-1}=\left[ \begin{array}{cc} X^{-1}SX^{-1} & -X^{-1} \\ X^{-1} & 0 \end{array} \right] .$$ The DBs are $$\begin{aligned} \{\mathcal{A},\mathcal{B}\}_{D} &=&\{\mathcal{A},\mathcal{B}\}-\{\mathcal{A} ,\phi _{\bar{A}}^{1}\}(X^{-1}SX^{-1})_{\bar{A}\bar{B}}\{\phi _{\bar{B}}^{1}, \mathcal{B}\} \notag \label{bracketDspace} \\ &&+\{\mathcal{A},\phi _{\bar{A}}^{1}\}(X^{-1})_{\bar{A}\bar{B}}\{\phi _{\bar{ B}}^{2},\mathcal{B}\}-\{\mathcal{A},\phi _{\bar{A}}^{2}\}(X^{-1})_{\bar{A} \bar{B}}\{\phi _{\bar{B}}^{1},\mathcal{B}\},\end{aligned}$$ which again leads to the result $$\{\mathcal{A}(\mathbf{x}),\mathcal{B}(\mathbf{y})\}_{D}=\{\mathcal{A}( \mathbf{x}),\mathcal{B}(\mathbf{y})\},$$ for the final variables $\theta_{j}$ and $\Pi _{j}^{\theta}$. The above result arises from the fact that each one of the additional PBs in (\[bracketDspace\]) include a contribution with $\phi _{\bar{A}}^{1}$, which has zero PB with $\theta_{j}$ and $\Pi _{j}^{\theta}$, according to the \[Apendice2\]. In other words, we recover the canonical algebra (\[algebrafinal\]), together the final extended Hamiltonian density $$\mathcal{H}_{E}^{ENM}=\mathcal{H}_{F}+\lambda _{A}G_{A}+\theta_lJ_l. \label{hamiltonianofinal3}$$ ### Conservation of the quantities $G_{A}$ As in the case of Subsection \[x1\], we have to deal with the issue that neither the $ G_{A}$’s are constraints, nor the $\lambda _{A}$’s are arbitrary functions in the extended Hamiltonian density (\[hamiltonianofinal3\]) of the ENM. However, [the transformation of the Gauss functions $G_{A}$ into constraints]{} proceeds in the same way as in the previous case. In fact, [ after employing the equivalence between the canonical algebras of the MGT and the ENM in the calculation]{}, the dynamics of the ENM yields [ again Eq.(\[DOTGAENM1\])]{}. Once more, in order to recover the MGT, it is enough to impose the generalized current conservation $(DJ)_A=0$ for all times and $G_{A}=0$ as initial conditions, since the dynamics of the ENM guarantees that $G_{A}(t)=0$ for all time. Therefore, we can include the quantities $\ G_{A}$ as constraints in the extended Hamiltonian (\[hamiltonianofinal3\]), through arbitrary functions $N_{A}$, by adding $N_{A}G_{A}$ and redefining $\lambda _{A}+N_{A}=\Theta _{A}$. We obtain $$\mathcal{H}_{E}=\mathcal{H}_{F}+\Theta _{A}G_{A}+\theta_lJ_l, \label{HDENM2}$$ where now $\Theta _{A}\;$are arbitrary functions. In this way we recover the extended Hamiltonian density (\[hamiltonianofinal\]) together the canonical algebra (\[algebrafinal\]) of the MGT. The conditions for the equivalence between the MGT$\;$and the ENM have been established, yielding the same results as in the previous case of Subsection \[x1\]. We close this Subsection with a comment related to the rôle of the gauge fixing in the MGT when achieving its equivalence with the ENM. Let us recall that such equivalence was obtained basically in two steps: (i) we proved that the canonical algebra of the ENM yields the canonical algebra Eq. (\[algebrafinal\]) of the MGT and (ii) after the imposition of suitable conditions (Gauss laws and current conservation) in the ENM we were able to show that the corresponding Hamiltonian densities (\[HDENM1\]) and (\[HDENM2\]) have exactly the same form as the Hamiltonian density (\[algebrafinal\]) of the MGT. The only explicit gauge fixing in the MGT was to set strongly the first class constraints $\Phi^1_A=\Pi^\lambda_A \approx 0$ by adding the condition $\Phi^3_A=\lambda_A-\Theta_A$, in order to eliminate the variables $\lambda_A$ and $\Pi^\lambda_A$. The canonical algebra (\[algebrafinal\]), together with the Hamiltonian density (\[hamiltonianofinal\]) of the MGT were obtained at this stage. Anyway, the above gauge fixing is completely general because the arbitrary functions $\Theta_A$ remained unspecified and can only be determined once a further gauge fixing of the remaining first class constraints $G_A$ is performed. This has not been done and both the canonical algebra and the Hamiltonian densities of the MGT and the ENM coincide while keeping $G_A$ as first class constraints. In other words, the equivalence has been proved for an arbitrary gauge fixing in the MGT, after the variables $\lambda_A$ and $\Pi^\lambda_A$ were decoupled. Under the general conditions stated in the \[Apendice1\] and the \[Apendice2\] such equivalence is completely independent of the specific constraint $F(\theta_l, \lambda_A)=0$ which defines the ENM. Examples ======== In this section, we present some examples where the above ideas can be applied. In each case we outline a brief description of the corresponding MGT, showing that the conditions (\[hamiltonianocanonic\]), (\[condition2\]) and (\[condition3\]) in Section \[gaugetheory\] hold. These requirements are enough to establish an equivalence with an arbitrary ENM, which would be defined by the same Lagrangian of the MGT$\;$ plus a constraint $F(\theta_{l},\lambda _{A})=0$. This constraint leads to a new theory which is different of the MGT. However, after imposing suitable conditions we can recover the original MGT as explained in the Section \[NGmodel\]. A Mechanical Model ------------------ As a first example, we consider a mechanical model [@Dayi] defined by the Lagrangian $$\mathcal{L}=\frac{1}{2}(\dot{x}^{2}+\dot{y}^{2})-\lambda (x\dot{y}-y\dot{x})+ \frac{1}{2}\lambda ^{2}(x^{2}+y^{2})-V(x^{2}+y^{2}), \label{MECMOD}$$ which is invariant under the following gauge transformations, with parameter $\epsilon(t)$, $$\delta\, x=\epsilon(t) \,y, \qquad \delta\, y=-\epsilon(t)\,x, \qquad \delta \, \lambda = {\dot \epsilon}(t),$$ and describes time dependent rotations around the $z$ axis. It can be easily proved that the canonical Hamiltonian is given by $$\begin{aligned} \mathcal{H}_{c} &=&\frac{\Pi _{x}^{2}}{2}+\frac{\Pi _{y}^{2}}{2} +V(x^{2}+y^{2})+\lambda (x\Pi _{y}-y\Pi _{x}), \\ &=&\mathcal{H}_{F}(x,y,\Pi _{x},\Pi _{y})+\lambda (x\Pi _{y}-y\Pi _{x}), \notag\end{aligned}$$ where $$\mathcal{H}_{F}=\frac{\Pi _{x}^{2}}{2}+\frac{\Pi _{y}^{2}}{2} +V(x^{2}+y^{2}),\quad \quad \quad \Pi _{a}=\frac{\partial \mathcal{L}}{ \partial \dot{a}},\quad \quad \quad a=x,y.$$ The model has only the primary constraint $\Phi _{1}=\Pi _{\lambda }=0$, which leads to the secondary constraint $\Phi _{2}=x\Pi _{y}-y\Pi _{x}=0$. We note that, $\Phi _{2}=\Phi _{2}(x,y,\Pi _{x},\Pi _{y})$, i.e. it is $ \lambda $ independent. Following Section \[gaugetheory\] we identify $ \lambda \rightarrow \lambda $ and $G_{A\;\;}\rightarrow \Phi _{2}=x\Pi _{y}-y\Pi _{x}$. The conditions $\{\Phi _{2},\mathcal{H}_{F}\}=0$ and $\{\Phi _{2},\Phi _{2}\}=0$ hold, therefore, there are no [additional]{} constraints in the theory. The conditions (\[hamiltonianocanonic\]), (\[condition2\]) and (\[condition3\]) demanded in Section \[gaugetheory\] are fulfilled. The explicit equivalence between the MGT and the corresponding ENM arising from this mechanical model will be proved in the \[Mechanics1a\]. This is a detailed example which shows how does our procedure works. Yang-Mills Theory ----------------- This case was studied in Ref. , where the non abelian Nambu’s model was defined by the constraint $A_{\mu}^a A^{a\mu}=n^2M^2$, where $\mu$ and $a$ are indices in the Lorentz and group spaces, respectively, $n^2=n_\beta n^\beta=0,\pm1$ and $M^2$ is a constant. The space-like case presented in [@NANM_ES_UR] is a particular example what we have proved in this work. A brief review of the Yang-Mills case starts with the standard Lagrangian density $$\mathcal{L}(A_\mu^a)=-\frac{1}{4}F_{\mu\nu}^aF^{a\mu\nu}-A_\mu^aJ^{a\mu}, \label{LYM}$$ which produces the Hamiltonian density $$\begin{aligned} \mathcal{H}_{c} &=&\frac{1}{2}(\mathbb{E}^{2}+\mathbb{B}^{2})-A_{0}^{a\;\;}( \partial _{i}E_{i}-J_{0})^{a\;\;}+A_{i}^{a}J^{ai},\notag \\ &=&\mathcal{H}_{F}(A_{i},E_{i})-A_{0}^{a\;\;}(\partial _{i}E_{i}-J_{0})^{a} +A_{i}^{a}J^{ai}, \end{aligned}$$ where $$\mathcal{H}_{F}=\frac{1}{2}(\mathbb{E}^{2}+\mathbb{B}^{2}).$$ The non trivial gauge transformations in the Lagrangian (\[LYM\]) are $\delta A_\mu^a=D_\mu\Lambda^a(x)$, where $D_\mu$ is the covariant derivative. The primary constraints are $\Pi _{0}^{a\;\;}=0$, which lead to the secondary constraints $G_{a}=(\partial _{i}E_{i}-J_{0})_{a\;\;}=G_{a}(A_{i}^{b},E_{i}^{b})$. Following Section \[gaugetheory\], we identify $A_{0}^{a\;\;}\rightarrow \lambda _{A\;\;}$ and $G_{a\;\;}\rightarrow G_{A}$. It is well known that $\{G_{a}, \mathcal{H}_{F}\}=0$ and $\{G_{a},G_{b}\}\approx 0$, therefore, there are no more constraints. The conditions (\[hamiltonianocanonic\]), (\[condition2\]) and (\[condition3\]) are satisfied. As we previously mentioned, in some cases current conservation follows from the imposition of Gauss constraints as initial conditions, as it happens in the non-Abelian Nambu model, for example. We clarify this issue in the case discussed in Ref. , defined by $$\mathcal{L}(A_\mu)=-\frac{1}{4}F_{\mu\nu}^aF^{a\mu\nu}-A_\mu^aJ^{a\mu}, \quad\quad\quad A_{\mu}^a A^{a\mu}=n^2M^2.$$ The equations of motion for the space-like case are given by $$\mathcal{E}^{ia}-\mathcal{E}^{31}\frac{A_i^a}{A_3^1}=0, \quad\quad i \neq 3, \quad a\neq 1. \label{NANMeq1}$$ $$\mathcal{E}^{0a}-\mathcal{E}^{31}\frac{A_0^a}{A_3^1}=0, \label{NANMeq2}$$ with the notation $$\mathcal{E}^{\nu a}=(D_\mu F^{\mu\nu}-J^\nu)^a.$$ From the dynamics of the non-Abelian Nambu model the time evolution of the Gauss functions $\Omega^a=\mathcal{E}^{0a}=(D_i E_i- J_0)^a$ is $$\dot{\Omega}^a=-g C^{abc}A_0^b \Omega^c - D_\mu J^{\mu a} + D_k \bigg(\bigg(\frac{A_k^a}{A_0^1}\bigg)\Omega^1\bigg), \label{NANMGauss}$$ Imposing the Gauss constraints as initial conditions ($\Omega^a(t=t_0)=0$) upon Eqs. (\[NANMeq1\])-(\[NANMeq2\]), the standard Yang-Mills equations of motion ($\mathcal{E}^{\nu a}=0$) are recovered and they are valid at $t=t_0$. As a consequence of the antisymmetry of Maxwell tensor, the relation $$0=D_\nu \mathcal{E}^{\nu a}= D_\nu (D_\mu F^{\mu\nu}-J^\nu)^a= -D_\nu J^{\nu a}=0, \quad \quad (\textrm{at } t=t_0),$$ holds, yielding current conservation at $t=t_0$. Using the above in Eq. (\[NANMGauss\]), we obtain $\dot{\Omega}^a(t=t_0)=0$. Since the relations $\Omega^a(t=t_0)=0$ and $\dot{\Omega}^a(t=t_0)=0$ are fulfilled, we obtain $$\begin{aligned} \nonumber \Omega^a(t=t_0+\delta t_1)&=&\Omega^a(t=t_0)+\dot{\Omega}^a(t=t_0)\delta t_1+\cdots, \\ &=&0.\end{aligned}$$ Using that $\Omega^a(t=t_0+\delta t_1)=0$, we observe that the Yang Mills equations are now valid at $t=t_0+\delta t_1$, which again, due to antisymmetry of the Maxwell tensor, imply current conservation at $t=t_0+\delta t_1$ and $\dot{\Omega}^a(t=t_0+\delta t_1)=0$ via Eq. (\[NANMGauss\]). The relations $\Omega^a(t=t_0+\delta t_1)=0$ and $\dot{\Omega}^a(t=t_0+\delta t_1)=0$ imply $\Omega^a(t=t_0+\delta t_1+\delta t_2)=0$. Iterating the previous process, always applying the dynamical equation (\[NANMGauss\]) of the ENM, it follows that current conservation is valid for all time. Summarizing, the above analysis shows that in the case of the non-Abelian Nambu model, the specific gauge structure of the theory allows us to impose only the Gauss constraints as initial conditions, which necessarily yield current conservation for all time, in order to obtain the corresponding MGT. Linearized Gravity ------------------ The main example in this paper corresponds to present the equivalence between the linearized Einstein gravity and one ENM. We start from the Fierz-Pauli Lagrange density [@LFierz; @ADMDirac; @Lingrav2; @Lingrav3; @Lingrav4] $$\mathcal{L}=\frac{1}{2}[\partial_\lambda h_{\mu\nu}\partial^\mu h^{\lambda\nu}-\partial_\lambda h \partial_\mu h^{\mu\lambda}]+\frac{1}{4}[\partial_\mu h \partial^\mu h-\partial_\lambda h_{\mu\nu}\partial^\lambda h^{\mu\nu}],$$ where $h_{\mu\nu}$ is a second rank symmetric tensor, $h=\eta^{\mu\nu}h_{\mu\nu}$ with $\eta_{\mu\nu}=diag(-,+,+,+)$ being the metric tensor of the flat space. The non trivial gauge transformations in the above Lagrangian are $\delta h_{\mu\nu}=\partial_\mu \xi_\nu(x)+\partial_\nu \xi_\mu(x)$. Considering the action $\mathcal{S}=\int \,d^4x\, \mathcal{L}$, and after integration by parts, the Fierz-Pauli Lagrangian density can be rewritten as follows [@Lingrav2], $$\begin{aligned} \mathcal{L}&=&-\frac{1}{4}\dot{h}_{ii}\dot{h}_{jj}-\frac{1}{2}\partial_k h_{ii}\partial_kh_{00}+\frac{1}{4}\partial_ih_{jj}\partial_ih_{kk}+\frac{1}{4}\dot{h}_{ij}\dot{h}_{ij} \nonumber \\ && +\frac{1}{2}\partial_ih_{0j}\partial_ih_{0j}-\frac{1}{4}\partial_ih_{jk}\partial_i h_{jk} +\dot{h}_{ii}\partial_j h_{0j}-\frac{1}{2}\partial_i h_{kk}\partial_jh_{ij} \nonumber \\ && +\frac{1}{2}\partial_i h_{00}\partial_j h_{ij} -\dot{h}_{ik}\partial_ih_{0k}-\frac{1}{2}\partial_ih_{0i}\partial_jh_{0j}+\frac{1}{2}\partial_ih_{jk}\partial_jh_{ik}.\end{aligned}$$ where the latin indices stand for the pure space coordinates ($i,j=1,2,3$) and $\dot{f}\equiv\partial_0 f$. Introducing the momenta conjugate to $h_{\mu\nu}$ ($\mu,\nu=0,1,2,3$) $$p_{\mu\nu}=\frac{\delta \mathcal{L}}{\delta{\dot{h}_{\mu\nu}}},$$ one can directly write the primary constraints $$p_{0\nu}=0. \label{graviconstr}$$ The non-zero momenta are given by $$p_{ij}=-\frac{1}{2}\delta_{ij}\dot{h}_{kk}+\frac{1}{2}\dot{h}_{ij}+\delta_{ij}\partial_k h_{k0}-\frac{1}{2}(\partial_ih_{0j}+\partial_jh_{0i}).$$ The canonical Hamiltonian density is [@Lingrav2] $$\begin{aligned} \label{hamiltgravit} \mathcal{H}_c&=&p^{ij}\dot{h}_{ij}-\mathcal{L}, \nonumber \\ &=&p_{ij}p_{ij}-\frac{1}{2}p_{kk}p_{ll}+\frac{1}{2}(\partial_ih_{kk}\partial_jh_{ij}-\partial_ih_{jk}\partial_jh_{ik})+\frac{1}{4}(\partial_ih_{jk}\partial_ih_{jk}-\partial_ih_{jj}\partial_ih_{kk}) \nonumber \\ &&-\frac{1}{2}h_{00}(\partial_i\partial_ih_{kk}-\partial_i\partial_jh_{ij})-2h_{0j}\partial_ip_{ij}, \nonumber \\ &=& \mathcal{H}_F(h_{ij},p_{ij})-\frac{1}{2}h_{00}(\partial_i\partial_ih_{kk}-\partial_i\partial_jh_{ij})-2h_{0j}\partial_ip_{ij},\end{aligned}$$ where $$\mathcal{H}_F(h_{ij},p_{ij})=p_{ij}p_{ij}-\frac{1}{2}p_{kk}p_{ll}+\frac{1}{2}(\partial_ih_{kk}\partial_jh_{ij}-\partial_ih_{jk}\partial_jh_{ik})+\frac{1}{4}(\partial_ih_{jk}\partial_ih_{jk}-\partial_ih_{jj}\partial_ih_{kk})\,.$$ The time evolution of the primary constraints $\Omega^\mu=p^{0\mu}=0$ gives the secondary constraints $\Omega^4=\partial_i\partial_ih_{kk}-\partial_i\partial_jh_{ij}$ and $\Omega^{4+i}=\partial_jp_{ji}$. No tertiary constraints appear and the Dirac’s method closes. It can be proved that all constraints $\Omega$’s have vanishing Poisson brackets among themselves [@ADMDirac; @Lingrav2; @Lingrav3], so all of them are of first class. Also, the conditions $\{\Omega^4,\mathcal{H}_F\}\approx 0$ and $\{\Omega^{4+i},\mathcal{H}_F\}\approx 0$ are fulfilled [@ADMDirac; @Lingrav2; @Lingrav3]. The properties stated in this paragraph, together with (\[graviconstr\]) and (\[hamiltgravit\]) fulfill the basic requirements (\[hamiltonianocanonic\]), (\[condition2\]) and (\[condition3\]). Following Section \[gaugetheory\], we identify $\lambda_m\rightarrow h_{00}, \, h_{0j}$ and $G_m\rightarrow \partial_i\partial_ih_{kk}-\partial_i\partial_jh_{ij}, \, \partial_jp_{ji}$. Summary and conclusions ======================= Spontaneous Lorentz symmetry breaking (SLSB) has attracted considerable attention in recent years [@Colladay-Kostelecky; @Colladay-Kostelecky2], both from the experimental and theoretical points of view. One of the rewards in considering SLSB is the possibility of giving a dynamical setting to the gauge principle by considering photons, gravitons and non-Abelian gauge fields as the massless Nambu-Goldstone bosons arising from such a breaking [@Nambu-Progr; @Azatov-Chkareuli; @Urru-Mont; @JLCH1; @JLCH2; @JLCH3; @NANM_ES_UR; @JLCH4; @NANM-Abel; @emergent_gauge; @emergent_gauge2]. This approach can be codified under the name of different Nambu models. Usually these models arise as the low energy limit of spontaneous Lorentz symmetry breaking (SLSB) in the so called Bumblebee models [@BUMBLEBEE4; @BUMBLEBEE42; @BUMBLEBEE43; @BUMBLEBEE44; @BUMBLEBEE45; @bumblebee1; @HERNASKI; @Constri]. In practice Nambu models turn out to be described by the Lagrangian density of a gauge theory, the mother gauge theory (MGT), plus a non-linear constraint, arising from the non-zero vacuum expectation value of the corresponding gauge fields due to the SLSB. This constraint is not considered as an standard gauge fixing, but is explicitly solved and substituted in the gauge Lagrangian thus destroying gauge invariance. Some explicit calculations, like for example in Refs. , have shown that, under some conditions, the violation of the Lorentz symmetry and the lack of gauge invariance, introduced by the non-linear constraint, become unobservable in such a way that the appearing GBs can be interpreted as the gauge particles of the original MGT. A natural question posed by this approach in the realm of gauge theories, is to determine under which conditions the recovery of an arbitrary MGT theory from the corresponding Nambu model, defined by a general constraint over the coordinates, becomes possible. We refer to these theories as extended Nambu models (ENM), to differentiate them from the case where the constraint is treated as a standard gauge fixing term. At this level, the mechanism for generating the constraint is irrelevant and the case of SLSB is taken only as a motivation, which naturally bring this problem under consideration. The equivalence between gauge theories and ENMs is not straightforward and one has to consider the following issues: (i) because of the additional constraint, ENM models are not gauge invariant, (ii) the number of degrees of freedom in the ENM is larger than that of the MGT, and (iii) the equations of motion of the two theories do not match. In this way, it becomes clear that additional requirements have to be imposed upon the ENM in order to recover the original MGT. The strategy we follow is a generalization of the non-perturbative Hamiltonian analysis developed for the abelian Nambu model in Ref. and for the non-abelian Nambu model in Ref. . In Section \[gaugetheory\], we define the class of MGT under consideration by the requirements given in Eqs. (\[hamiltonianocanonic\]), (\[condition2\]) and (\[condition3\]), starting from a Lagrangian $ L(\theta_{l},\lambda _{A},\;\dot{\theta}_{l})$, where $l=1,2,...,n$ and $ A=1,2,...,K$. We also perform the Dirac method to calculate the Hamiltonian (\[hamiltonianofinal\]) and the canonical algebra (\[algebrafinal\]) of the MGT. Subsequently, the corresponding ENM is introduced in Section \[NGmodel\], by having the same Lagrangian as the MGT plus a constraint $F(\theta_{l},\lambda _{A})=0$ among the coordinates of the MGT. We solve the nonlinear constraint (\[constriccion\]) in two generic different ways, and we identify the canonical coordinates and momenta of the ENM, together with their canonical algebra and extended Hamiltonian. At this level, only second class constraints arise, reflecting the fact that the ENM is a theory without gauge invariance. Since both theories arise from the same Lagrangian, it is possible to rewrite the standard canonical variables of the MGT in terms of the canonical variables of the ENM. In this way, using the canonical algebra of the ENM, we show that the canonical algebra of the MGT is recovered. \[Apendice1\] and \[Apendice2\] include the calculation of the required PBs that prove the previous statement. The second class constraints in the ENM are further imposed strongly, by introducing the corresponding Dirac brackets (DBs), in order to eliminate the canonical variables $\lambda _{A},\;\bar{\Pi}_{B}^{\lambda }$. The DBs of the remaining variables are also calculated yielding no modifications with respect to the original PBs. The final extended Hamiltonian for the ENM, rewritten in terms of the canonical variables of the MGT, has the same form as the Hamiltonian corresponding to the MGT, except that the Gauss functions $G_{A}$ do not appear as first class constraints, as it should be in the MGT. This is because their coefficients in the extended Hamiltonian are not arbitrary functions, but specific functions of the coordinates of the ENM. In order to remedy this issue we calculate the time evolution of the functions $G_{A}$, according to the ENM dynamics, and find that demanding generalized current conservation [$(DJ)_A=0$]{} for all time, together $G_{A}=0$ as initial conditions yields $G_{A}(t)=0 $ for all time. This allows us to include the quantities $G_{A}$ as first class constraints in the extended Hamiltonian through arbitrary functions $N_{A}$ adding the term $N_{A}G_{A}$ to the ENM Hamiltonian. In this way, the Hamiltonian describing the ENM becomes the same as the Hamiltonian of the MGT. It is important to recall that the generalized current conservation condition $(DJ)_A=0$ follows from the gauge invariance generated by the Gauss law constraints $G_A\approx 0$ in the Hamiltonian action of the MGT. We prove this statement in the \[CURCONS1\]. Summarizing, the correct statement is that gauge invariance is indeed dynamically recovered in the ENM provided we impose current conservation for all times and the Gauss laws only as initial conditions. Let us recall that the canonical algebra of the MGT has been already recovered from that of the ENM. In this way, the complete equivalence of the MGT with the ENM plus those suitable conditions is obtained. We emphasize that such equivalence has been proved for an arbitrary gauge fixing in the MGT and that it is completely unrelated to the specific constraint which defines the ENM. Section \[examples\] presents some particular cases of MGTs in which our general result can be directly applied to construct arbitrary ENMs, from where the MGT can be ultimately recovered after imposing the appropriate conditions previously stated in our general analysis. Electrodynamics, Yang Mills theories and linearized gravity are typical examples described previously in the literature, where the constraint $ F(\theta_{l},\lambda _{A})=0$ arises from spontaneous Lorentz symmetry breaking. Since on one hand the standard Nambu models make explicit reference to spontaneous Lorentz symmetry breaking and on the other hand they constitute particular cases of the ENM we have considered; our results confirm and clarify the statement that, under the imposition of current conservation for all time together with the Gauss constraints as initial conditions, Lorentz invariance violation in such Nambu models is physically unobservable. Acknowledgements {#acknowledgements .unnumbered} ================ L. F. U. has been partially supported by the project CONACyT \# 237503. C. A. E. and L. F. U. acknowledge support from the project UNAM (Dirección General de Asuntos del Personal Académico) \# IN104815. C. A. E. is supported by the CONACyT Postdoctoral Grant No. 234745. The generalized current conservation {#CURCONS1} ==================================== [We generalize to the Hamiltonian formulation the general idea that current conservation is a consequence of the invariance of the action under the gauge transformations, which in this case are generated by the Gauss constraints. The simplest example of this relation arises in the Lagrangian formulation of electrodynamics with the action]{} $$S=\int d^4x \, \left(-\frac{1}{4} F^{\mu\nu} F_{\mu\nu}-J^\mu A_\mu\right).$$ [Demanding $\delta S=0$ under the gauge transformations $\delta A_\mu= \partial_\mu \alpha$, with parameter $\alpha$, yields]{} $$0=\delta S= -\int d^4x\, \alpha \,(\partial_\mu J^\mu), \quad \longrightarrow \quad \partial_\mu J^\mu=0,$$ after integrating by parts. In our case we start from the MGT described by the extended Hamiltonian density (\[hamiltonianofinal\]) $${\cal H}_E= {\cal H}_F+ \Theta_A (G^0_A+ J_A) + \theta_l J_l,$$ where we have fixed the constraints $\lambda_A= \Theta_A, \, \Pi^\lambda_A=0$ such that the corresponding canonical variables $\theta_k, \, \Pi^\theta_l$ satisfy the standard Dirac brackets (\[algebrafinal\]). Here $\Theta_A$ play the rôle of Lagrange multipliers. The Hamiltonian action can be written in a compact way as $$S=\int dt\left[ \Pi _{k}^{\theta }\;\dot{\theta}_{k}-\left( {H}_{F}+\Theta_{A}(G_{A}^{0}+J_{A}\right) +\theta _{l}J_{l})\right], \label{HAMACTION}$$ with the convention that the contraction of the indices $k$ and $A$ include an integral over the respective coordinates. In other words we denote $$P_kQ_k=\int d^3x\, P_k(t,{\mathbf x})Q_k(t,{\mathbf x}), \quad R_AS_A=\int d^3x\, R_A(t,{\mathbf x})S_A(t,{\mathbf x}).$$ Also, $H_F$ is the Hamiltonian corresponding to the Hamiltonian density ${\mathcal H}_F$ and we identify $G_A=G_A^0 +J_A$ as the remaining first class constraints. Let us recall that $J_{A},\, J_{l}$ are external currents independent of the canonical variables. The following calculation is an extension of the discussion related to the gauge invariance of the action in Chapter 3 of Ref. and we heavily rely upon the results included there. Let us consider the change of the action under the gauge transformations generated by the Gauss constraints $G_A$. The general transformation of any function of the canonical variables is $$\delta F=\;\alpha _{B}\left\{ F,G_{B}\right\} =\alpha _{B}\left\{ F,G_{B}^{0}\right\}, \label{CHANGE}$$ where $\alpha_A(x)$ are the gauge parameters. In particular, one can show [@HT] $$\delta \left( \Pi _{k}^{\theta }\;\dot{\theta}_{k}\right) =\frac{d}{dt}\left[ \alpha _{A}\left( \frac{\partial G_{A}^{0}}{\partial \Pi _{k}^{\theta }}\Pi _{k}^{\theta }-G_{A}^{0}\right) \right] +\frac{\partial \alpha _{A}}{\partial t}G_{A}^{0}.$$ Equation (\[CHANGE\]) also leads to $$\begin{aligned} &&\delta H_{F}=\alpha_{A}\left\{ H_{F},G_{A}^{0}\right\} =-\alpha _{A}C_{AB}G_{B}^{0}, \nonumber \\ &&\delta \theta _{l}=\;\alpha _{B}\left\{ \theta _{l},G_{B}\right\} =\alpha _{A}\left\{ \theta _{l},G_{A}^{0}\right\}, \nonumber \\ &&\delta G_{A}^{0}=\;\alpha _{B}\left\{ G^0_{A},G_{B}\right\} =\alpha _{B}\left\{ G_{A}^{0},G_{B}^{0}\right\} =C_{ABC}\alpha _{B}G_{C}^{0}.\end{aligned}$$ The next step is to establish the transformation law for the Lagrange multipliers. According to Ref. this is given by $$\delta \Theta_{A}=\frac{\partial \alpha ^{A}}{\partial t}+\alpha _{B} \Theta_{C}C_{BCA}+\alpha _{B}C_{BA}.$$ In this way we obtain $$\begin{aligned} \delta S &=&\int dt \left[ \frac{\partial \alpha _{A}}{\partial t}G_{A}^{0} + \alpha_{A}C_{AB}G_{B}^{0}- \alpha_{B}C_{BA}G_{A}^{0} \right. \nonumber \\ && \left. -\frac{\partial \alpha _{A}}{\partial t}G_{A}^{0} - [C_{BAC}+C_{ABC}] \Theta_{A}\alpha _{B}G_{C}^{0}\right] \nonumber \\ && - \int dt\left[ \left( \frac{\partial \alpha _{A}}{\partial t}+\Theta_{C}\alpha _{B}C_{BCA}+\alpha _{B}C_{BA}\right) J_{A}+\delta \theta _{l}J_{l}\right], \label{FINALCONS}\end{aligned}$$ where we have separated the term dependent on the currents in the third line of the above equation. It is interesting to observe that the cancellation in the second line of Eq. (\[FINALCONS\]) depends only on the antisymmetry of $C_{ABC}$ on the first two index, as required by their definition. This cancellation is the statement of the gauge invariance of the MGT in the absence of external currents. Imposing $\delta S=0$ in the general case we are left with $$\frac{\partial J_{A}(x)}{\partial t}-C_{ABC}\lambda_{B}J_{C}(x)-C_{AB}J_{B}(x)+ \int d^3y\left\{ \, G_{A}^{0}(x),\theta_{l}(y)\right\} J_{l}(y)=0, \label{FINCURCONS1}$$ which is our generalized statement of current conservation. In the above equation, which reproduces Eqs. (\[DEFDJA\]) and (\[currentcons\]), we have used $\Theta_A=\lambda_A$ and we have restated the coordinate dependence. We find it very remarkable that the above statement of current conservation can be obtained without making any reference to the detailed structure of the constraints $G_A$. A Mechanical Model: detailed calculation {#Mechanics1a} ======================================== In this Appendix we explicitly show how the equivalence between the MGT and the ENM is carried out for the mechanical model defined in Eq.(\[MECMOD\]) of Section \[examples\], which is a theory with first class constraints and non trivial gauge transformations that defines the mother gauge theory (MGT). The Dirac algorithm is performed and the canonical algebra, together with the Hamiltonian describing the dynamics of this MGT are presented. The next step is to build the ENM and to prove the equivalence with the MGT, showing that both the Hamiltonian and the algebra of this ENM correspond to those of the MGT, after suitable conditions are imposed to the ENM. The Gauge theory ---------------- The MGT is defined by the Lagrangian $${L}=\frac{1}{2}(\dot{x}^{2}+\dot{y}^{2})-\lambda (x\dot{y}-y\dot{x})+ \frac{1}{2}\lambda ^{2}(x^{2}+y^{2})-V(x^{2}+y^{2}). \label{MGTL}$$ It can be easily proved that, the canonical Hamiltonian is given by $$\begin{aligned} \nonumber {H}_{c} &=& \dot{x}\Pi_x+\dot{y}\Pi_y-{{L}}, \\ \nonumber &=&\frac{\Pi _{x}^{2}}{2}+\frac{\Pi _{y}^{2}}{2} +V(x^{2}+y^{2})+\lambda (x\Pi _{y}-y\Pi _{x}), \\ &=&{H}_{F}(x,y,\Pi _{x},\Pi _{y})+\lambda (x\Pi _{y}-y\Pi _{x}), \label{HamilMGT1a}\end{aligned}$$ where $$\Pi_\lambda=\frac{\partial \mathcal{L}}{\partial\dot{\lambda}}=0,\quad\quad \Pi _{x}=\frac{\partial \mathcal{L}}{ \partial \dot{x}}=\dot{x}+\lambda y,\quad \quad \Pi _{y}=\frac{\partial \mathcal{L}}{ \partial \dot{y}}=\dot{y}-\lambda x,$$ $${H}_{F}=\frac{\Pi _{x}^{2}}{2}+\frac{\Pi _{y}^{2}}{2} +V(x^{2}+y^{2}). \label{canmomegaug}$$ The canonical algebra is given by the non-zero Poisson brackets $$\{x,\Pi_x\}=\{y,\Pi_y\}=\{\lambda,\Pi_\lambda\}=1.$$ The model only has the primary constraint $\Phi _{1}=\Pi _{\lambda }\approx0$, which leads to the secondary constraint $\Phi _{2}=x\Pi _{y}-y\Pi _{x}\approx0$. They are first class constraints, which implies that the model has one DOF. This agrees with the counting in Eq. (\[dofgauge\]). Now we construct the extended Hamiltonian $${H}_{E} = {H}_{c}+ \beta\,\Pi_\lambda,$$ where $\beta$ is a Lagrange multiplier. In order to eliminate $\lambda$ and $\Pi_\lambda$ we add the gauge condition $\Phi_3=\lambda-\Theta\approx0$, where $\Theta$ is an arbitrary function to be consistently determined after the remaining first class constraint $\Phi_2$ is fixed. The constraints $\Phi_1$ and $\Phi_3$ become second class and we can introduce the Dirac’s brackets to describe the dynamics. The matrix constructed with $\Phi_1$ and $\Phi_3$ is given by $$M_{ij} = \left( \begin{array}{cc} 0 & -1 \\ 1 & 0 \end{array} \right),$$ with the inverse $$(M^{-1})_{ij} = \left( \begin{array}{cc} 0 & 1 \\ -1 & 0 \end{array} \right),$$ Using the above matrix, we calculate the Dirac’s brackets for the remaining variables with the result $$\{x,x\}_D=\{y,y\}_D=0,\,\, \{\Pi_x,\Pi_x\}_D=\{\Pi_y,\Pi_y\}_D=0,\,\, \{x,\Pi_y\}_D=\{y,\Pi_x\}_D=0, \label{algex1}$$ $$\{x,\Pi_x\}_D=\{y,\Pi_y\}_D=1. \label{algex2}$$ Fixing strongly the constraints $\Phi_1=\Pi_\lambda=0$ and $\Phi_3=\lambda-\Theta=0$, together with the introduction of the Dirac’s brackets, the Hamiltonian becomes $$\begin{aligned} \nonumber {H}_{E} &=&\frac{\Pi _{x}^{2}}{2}+\frac{\Pi _{y}^{2}}{2} +V(x^{2}+y^{2})+\Theta (x\Pi _{y}-y\Pi _{x}), \\ \label{HamilEx2} &=&{H}_{F}(x,y,\Pi _{x},\Pi _{y})+\Theta (x\Pi _{y}-y\Pi _{x}).\end{aligned}$$ At this point, the dynamics of this theory is determined by the extended Hamiltonian ${H}_{E}$ given in Eq. (\[HamilEx2\]), together with the algebra (\[algex1\]-\[algex2\]). In the next Subsection we are going to define an extended Nambu model using the Lagrangian (\[MGTL\]), plus a relation between the coordinates. The goal will be to recover the extended Hamiltonian ${H}_{E}$ and the algebra (\[algex1\]-\[algex2\]) using the dynamics of the ENM. The extended Nambu model {#the-extended-nambu-model} ------------------------ According to Section \[NGmodel\], we can construct a Nambu model using the above Lagrangian plus a suitable relation among the coordinates $(x,y,\lambda)$. We illustrate the two generic cases: a) solving the coordinate $x$ and b) solving the coordinate $\lambda$. We choose a particular function in order to show explicitly the equivalence. ### Solving for the coordinate $x$ In this case we define the extended Nambu model by the Lagrangian $${L}=\frac{1}{2}(\dot{x}^{2}+\dot{y}^{2})-\lambda (x\dot{y}-y\dot{x})+ \frac{1}{2}\lambda ^{2}(x^{2}+y^{2})-V(x^{2}+y^{2}), \label{LagranC1}$$ plus the relation $$x=x(y,\lambda)=\frac{1}{2}(y^2+\lambda^2), \label{ConsC1}$$ which implies $$\dot{x}=y\dot{y}+\lambda\dot{\lambda}. \label{Constric1a}$$ When Eq. (\[Constric1a\]) is substituted into the Lagrangian (\[LagranC1\]), we obtain the Lagrangian $$\tilde{ L}(y, \lambda)={ L}(x=x(y, \lambda), y, \lambda)$$ for the ENM, where we also substitute ${\dot x}=y {\dot y} + \lambda {\dot \lambda}$. The canonical momenta can be obtained making use of the chain rule and they are given by $$\tilde{\Pi}_y= \frac{\partial \tilde{ L}}{\partial \dot{y}}=\dot{y}-\lambda x +(\dot{x}+\lambda y)y, \qquad \tilde{\Pi}_\lambda= \frac{\partial \tilde{ L}}{\partial \dot{\lambda}} =(\dot{x}+\lambda y )\lambda. \label{canmometa1}$$ We emphasize that $x$ and $\dot{x}$ in the above expressions are just labels for the combinations (\[ConsC1\]) and (\[Constric1a\]), respectively. In other words, $x$ is not a coordinate of the ENM. It can be shown that the relations (\[canmometa1\]) can be solved for the velocities $\dot{y}$ and $\dot{\lambda}$ in terms of the momenta, such that there are no constraints in the theory; therefore, the number of DOF is two, which agrees with Eq. (\[dofx1a\]). The canonical algebra between the coordinates $y$, $\lambda$ and their momenta is given by the non-zero Poisson brackets $$\{y,\tilde{\Pi}_y\}=1,\quad\quad\quad\{\lambda,\tilde{\Pi}_\lambda\}=1. \label{algxc1}$$ We note that we can rewrite the canonical momenta (\[canmometa1\]) in terms of the canonical momenta of the gauge theory (\[canmomegaug\]) as $$\tilde{\Pi}_y=\Pi_y+\Pi_x y,\quad\quad\quad \tilde{\Pi}_\lambda=\Pi_x \lambda, \label{moment1ab}$$ where the definitions $$\Pi_x=\dot{x}+\lambda y,\quad\quad \Pi_y=\dot{y}-\lambda x,$$ are, again, just labels to specify a particular combination of the variables $(y,\lambda)$, their velocities and their momenta. Solving the expressions in Eq. (\[moment1ab\]) in favor of $\Pi_x$ and $\Pi_y$ we have $$\Pi_x=\frac{\tilde{\Pi}_\lambda}{\lambda},\quad\quad\quad \Pi_y= \tilde{\Pi}_y-\frac{\tilde{\Pi}_\lambda}{\lambda} y.$$ [Next we calculate]{} the Poisson brackets among the following quantities: $x=x(y,\lambda), \,\, y, \,\, \Pi_x=\Pi_x(y,\lambda,\tilde{\Pi}_y,\tilde{\Pi}_\lambda)$ and $ \Pi_y=\Pi_y(y,\lambda,\tilde{\Pi}_y,\tilde{\Pi}_\lambda)\}$, using the canonical algebra of the ENM in Eq. (\[algxc1\]). We find that the only non-zero Poisson brackets are $$\{x,\Pi_x\}=1,\quad\quad\quad\{y,\Pi_y\}=1. \label{CANAENM}$$ The canonical Hamiltonian of the ENM can be computed as $$\begin{aligned} {H}_c^{ENM}&=& \dot{y}\tilde{\Pi}_y+\dot{\lambda}\tilde{\Pi}_\lambda- \tilde{ L}, \nonumber \\ &=& \dot{y}(\Pi_y+\Pi_x y)+\dot{\lambda} \Pi_x \lambda -\tilde{ L}, \nonumber \\ &=& \dot{y}\Pi_y+(y \dot{y}+\lambda\dot{\lambda})\Pi_x-\tilde{ L}, \nonumber \\ &=& \dot{y}\Pi_y+\dot{x}\Pi_x-{L}.\end{aligned}$$ In the last line of the above equation we have undone the substitution (\[ConsC1\]) in $\tilde{L}$ and we recover ${L}(x,y,\lambda)$. After substituting the velocities $\dot{x}$ and $\dot{y}$ in terms of the momenta $\Pi_x$ and $\Pi_y$, according to Eq.(\[canmomegaug\]), we can write $${H}_c^{ENM}=\frac{\Pi _{x}^{2}}{2}+\frac{\Pi _{y}^{2}}{2} +V(x^{2}+y^{2})+\lambda (x\Pi _{y}-y\Pi _{x}). \label{HamilC11}$$ When going back to the reduced MGT variables $x$ and $y$ we recall that the relation (\[ConsC1\]) yields $\lambda=\lambda(x,y)=\sqrt{2x- y^2}$ in our case. At this point we have recovered in Eq.(\[CANAENM\]) the canonical algebra of the MGT given in Eqs (\[algex1\]-\[algex2\]). Also, the canonical Hamiltonian of the ENM in Eq.(\[HamilC11\]) has the same form of the extended Hamiltonian of the MGT in Eq. (\[HamilEx2\]). The only difference arises from the fact that in this case the quantity $G\equiv(x\Pi _{y}-y\Pi _{x})$ is not a constraint, which is reflected in that the factor $\lambda$ is not an arbitrary function. To deal with this issue we study the time evolution of the quantity $G$ in the ENM. The dynamics of the ENM leads to $$\dot{G}=\{ (x\Pi _{y}-y\Pi _{x}),{H}_c^{ENM}\}=\{(x\Pi _{y}-y\Pi _{x}),\lambda(x,y)\}G=\frac{y}{\lambda}(1+x)\, G.$$ The evaluation of the previous bracket is performed in an easier way by using the algebra of the MGT, which we have previously proved that can be derived from the ENM algebra. The above equation shows that if we demand that $G(t=0)=0$, we obtain that $\dot{G}=0$ as well at t=0. This proves that the relation $G(t)=0$ will hold for all time. Therefore, we can include the quantity $G\equiv(x\Pi _{y}-y\Pi _{x})$ as a constraint in the Hamiltonian (\[HamilC11\]), through an arbitrary function $N$, by adding $NG$ and redefining $\lambda+N=\Theta$. We obtain $${H}_{c} ={H}_{F}(x,y,\Pi _{x},\Pi _{y})+\Theta (x\Pi _{y}-y\Pi _{x}),$$ where now $\Theta$ is an arbitrary function. In this way we recover the extended Hamiltonian (\[HamilEx2\]) together the algebra (\[algex1\]-\[algex2\]) of the MGT, thus proving the equivalence between the ENM and the MGT for this case, once $G=0$ has been imposed as an initial condition. ### Solving for the coordinate $\lambda$ Now we define the extended Nambu model by the Lagrangian $${L}=\frac{1}{2}(\dot{x}^{2}+\dot{y}^{2})-\lambda (x\dot{y}-y\dot{x})+ \frac{1}{2}\lambda ^{2}(x^{2}+y^{2})-V(x^{2}+y^{2}), \label{LagranC2}$$ plus the relation $$\lambda=\lambda(x,y)=x^2+y^2. \label{ConsC2}$$ Since the relation (\[ConsC2\]) does not modify the velocities of the Lagrangian (\[LagranC2\]), because $\dot{\lambda}$ is not present there, the canonical momenta $\Pi_x$ and $\Pi_y$ coincide with those of the MGT theory and satisfy the canonical algebra given by the non-zero Poisson brackets $$\{x,\Pi_x\}=\{y,\Pi_y\}=1.$$ As anticipated by Eqs. (\[constr123a\]) and (\[dof1ab\]) there are no constraints and the number of DOF in this model is two. The canonical Hamiltonian is obtained in the same manner as in the MGT, leading to $$\begin{aligned} \nonumber {H}_{c}^{ENM} &=&\frac{\Pi _{x}^{2}}{2}+\frac{\Pi _{y}^{2}}{2} +V(x^{2}+y^{2})+\lambda(x,y) (x\Pi _{y}-y\Pi _{x}), \\ \label{HamilC2} &=&{H}_{F}(x,y,\Pi _{x},\Pi _{y})+\lambda(x,y) (x\Pi _{y}-y\Pi _{x}), \end{aligned}$$ where $\lambda(x,y)$ is not an arbitrary function, but it is given by Eq. (\[ConsC2\]). In this model, the quantity $G\equiv(x\Pi _{y}-y\Pi _{x})$ is not a constraint; however, it is a conserved quantity, which follows from the bracket $$\dot{G}=\{ (x\Pi _{y}-y\Pi _{x}),{H}_{c}\}=0.$$ This show that if we demand $G(t=0)=0$, then $G(t)=0$ for all times. Therefore, we can include the quantity $G\equiv(x\Pi _{y}-y\Pi _{x})$ as a constraint in the Hamiltonian (\[HamilC2\]), through an arbitrary function $N$, by adding $NG$ and redefining $\lambda(x,y)+N=\Theta$. We obtain $${H}_{c} ={H}_{F}(x,y,\Pi _{x},\Pi _{y})+\Theta (x\Pi _{y}-y\Pi _{x}),$$ where now $\Theta$ is an arbitrary function. Therefore, we recover the Hamiltonian (\[HamilEx2\]) together the algebra (\[algex1\]-\[algex2\]) of the MGT. The algebra resulting when solving for the coordinate $\theta_{1}$ in the ENM constraint {#Apendice1} ======================================================================================== In this Appendix we show that the algebra for the canonical variables $\theta_{i}$ and $\Pi _{j}^{\theta}$ in the MGT, given in Eq. (\[algebrafinal\]), is recovered from the algebra of canonical variables corresponding to the ENM given in Subsection \[x1\]. The algebra of the second class constraints of the ENM is also calculated. Here the canonical variables of the ENM ($\theta_{\bar{l}},\lambda_{A},\bar{\Pi} _{\bar{l}}^{\theta},\bar{\Pi}_{B}^{\lambda }$) have the non-zero PBs $$\{\theta_{\bar{\imath}},\bar{\Pi}_{\bar{j}}^{\theta}\}=\delta _{{\bar{\imath}}{\bar{j}}, }\quad \quad \quad \quad \{\lambda _{A},\bar{\Pi}_{B}^{\lambda }\}=\delta _{AB}.\quad \quad \quad (\bar{\imath},\bar{j}=2,3,...,n). \label{commutation_constric2}$$ The transformation from canonical variables of the ENM to those of the MGT is given by $$\theta_{1}=f(\theta_{\bar{l}},\lambda _{A}),\quad \theta_{\bar{l}}=\theta_{ \bar{l}},\quad \Pi _{1}^{\theta}=\frac{\bar{\Pi}_{1}^{\lambda }}{ f_{\lambda _{1}}},\quad \Pi _{\bar{l}}^{\theta}=\bar{\Pi}_{\bar{l} }^{\theta}-\frac{\bar{\Pi}_{1}^{\lambda }}{f_{\lambda _{1}}}f_{\theta_{\bar{l}}},\quad (\bar{l}=2,3,...,n). \label{transformation2}$$ Our goal is to calculated the algebra among the variables $\theta_{i},\Pi _{j}$ of the MGT in terms of the canonical algebra (\[commutation\_constric2\]) of the ENM. The $\theta_i-\theta_j$ sector ------------------------------ $$\{\theta_{1},\theta_{1}\}=\{f(\theta_{\bar{\imath}},\lambda _{A}),f(\theta_{\bar{j}},\lambda _{A})\}=0.$$ $$\{\theta_{1},\theta_{l}\}=\{f(\theta_{\bar{n}},\lambda _{A}),\theta_{\bar{l}}\}=0.$$ $$\{\theta_{\bar{l}},\theta_{\bar{m}}\}=0.$$ The $\Pi^\theta_i-\Pi^\theta_j$ sector -------------------------------------- $$\{\Pi _{1}^{\theta},\Pi _{1}^{\theta}\}=\bigg\{\frac{\bar{\Pi}_{1}^{\lambda }}{ f_{\lambda _{1}}},\frac{\bar{\Pi}_{1}^{\lambda }}{f_{\lambda _{1}}}\bigg\}=0.$$ $$\begin{aligned} \{\Pi _{1}^{\theta},\Pi _{\bar{l}}^{\theta}\} &=&\bigg\{\frac{\bar{\Pi}_{1}^{\lambda } }{f_{\lambda _{1}}},\bar{\Pi}_{\bar{l}}^{\theta}-\frac{\bar{\Pi}_{1}^{\lambda }}{ f_{\lambda _{1}}}f_{\theta_{\bar{l}}}\bigg\}=\bigg\{\frac{\bar{\Pi}_{1}^{\lambda } }{f_{\lambda _{1}}},\bar{\Pi}_{\bar{l}}^{\theta}\bigg\}-\frac{\bar{\Pi} _{1}^{\lambda }}{f_{\lambda _{1}}}\bigg\{\frac{\bar{\Pi}_{1}^{\lambda }}{ f_{\lambda _{1}}},f_{\theta_{\bar{l}}}\bigg\}, \notag \\ &=&\bar{\Pi}_{1}^{\lambda }\bigg\{\frac{1}{f_{\lambda _{1}}},\bar{\Pi}_{\bar{ l}}^{\theta}\bigg\}-\frac{\bar{\Pi}_{1}^{\lambda }}{f_{\lambda _{1}}f_{\lambda _{1}}}\bigg\{\bar{\Pi}_{1}^{\lambda },f_{\theta_{\bar{l}}}\bigg\}, \notag \\ &=&-\frac{\bar{\Pi}_{1}^{\lambda }}{f_{\lambda _{1}}^{2}}f_{{\lambda _{1}}\theta_{ \bar{l}}}+\frac{\bar{\Pi}_{1}^{\lambda }}{f_{\lambda _{1}}^{2}}f_{\theta_{\bar{l} }\lambda _{1}}=0.\end{aligned}$$ $$\begin{aligned} \{\Pi _{\bar{l}}^{\theta},\Pi _{\bar{m}}^{\theta}\} &=&\bigg\{\bar{\Pi}_{\bar{l}}^{\theta}- \frac{\bar{\Pi}_{1}^{\lambda }}{f_{\lambda _{1}}}f_{\theta_{\bar{l}}},\bar{\Pi}_{ \bar{m}}^{\theta}-\frac{\bar{\Pi}_{1}^{\lambda }}{f_{\lambda _{1}}}f_{\theta_{\bar{m}}} \bigg\}, \notag \\ &=&-\bigg\{\bar{\Pi}_{\bar{l}}^{\theta},\frac{f_{\theta_{\bar{m}}}}{f_{\lambda _{1}}} \bigg\}\bar{\Pi}_{1}^{\lambda }+\bigg\{\bar{\Pi}_{\bar{m}}^{\theta},\frac{f_{\theta_{ \bar{l}}}}{f_{\lambda _{1}}}\bigg\}\bar{\Pi}_{1}^{\lambda } \notag \\ &&+\bigg\{\bar{\Pi}_{1}^{\lambda },\frac{f_{\theta_{\bar{m}}}}{f_{\lambda _{1}}} \bigg\}\bar{\Pi}_{1}^{\lambda }\frac{f_{\theta_{\bar{l}}}}{f_{\lambda _{1}}}- \bigg\{\bar{\Pi}_{1}^{\lambda },\frac{f_{\theta_{\bar{l}}}}{f_{\lambda _{1}}} \bigg\}\bar{\Pi}_{1}^{\lambda }\frac{f_{\theta_{\bar{m}}}}{f_{\lambda _{1}}}, \notag \\ &=&\bigg(\frac{f_{\theta_{\bar{m}}\theta_{\bar{l}}}}{f_{\lambda _{1}}}-\frac{f_{\theta_{ \bar{m}}}f_{\lambda _{1}\theta_{\bar{l}}}}{(f_{\lambda _{1}})^{2}}\bigg)\bar{\Pi} _{1}^{\lambda }-\bigg(\frac{f_{\theta_{\bar{l}}\theta_{\bar{m}}}}{f_{\lambda _{1}}}- \frac{f_{\theta_{\bar{l}}}f_{\lambda _{1}\theta_{\bar{m}}}}{(f_{\lambda _{1}})^{2}} \bigg)\bar{\Pi}_{1}^{\lambda } \notag \\ &&-\bigg(\frac{f_{\theta_{\bar{m}}\lambda _{1}}}{f_{\lambda _{1}}}-\frac{f_{\theta_{ \bar{m}}}f_{\lambda _{1}\lambda _{1}}}{(f_{\lambda _{1}})^{2}}\bigg)\bar{\Pi} _{1}^{\lambda }\frac{f_{\theta_{\bar{l}}}}{f_{\lambda _{1}}}+\bigg(\frac{f_{\theta_{ \bar{l}}\lambda _{1}}}{f_{\lambda _{1}}}-\frac{f_{\theta_{\bar{l}}}f_{\lambda _{1}\lambda _{1}}}{(f_{\lambda _{1}})^{2}}\bigg)\bar{\Pi}_{1}^{\lambda } \frac{f_{\theta_{\bar{m}}}}{f_{\lambda _{1}}}, \notag \\ &=&-\bigg(\frac{f_{\theta_{\bar{m}}}f_{\lambda _{1}\theta_{\bar{l}}}}{(f_{\lambda _{1}})^{2}}\bigg)\bar{\Pi}_{1}^{\lambda }+\bigg(\frac{f_{\theta_{\bar{l} }}f_{\lambda _{1}\theta_{\bar{m}}}}{(f_{\lambda _{1}})^{2}}\bigg)\bar{\Pi} _{1}^{\lambda } \notag \\ &&-\bigg(\frac{f_{\theta_{\bar{m}}\lambda _{1}}}{f_{\lambda _{1}}}\bigg)\bar{\Pi} _{1}^{\lambda }\frac{f_{\theta_{\bar{l}}}}{f_{\lambda _{1}}}+\bigg(\frac{f_{\theta_{ \bar{l}}\lambda _{1}}}{f_{\lambda _{1}}}\bigg)\bar{\Pi}_{1}^{\lambda }\frac{ f_{\theta_{\bar{m}}}}{f_{\lambda _{1}}}=0.\end{aligned}$$ The $\theta_i-\Pi^\theta_j$ sector ---------------------------------- $$\{\theta_{1},\Pi _{1}^{\theta}\}=\bigg\{f(\theta_{\bar{l}},\lambda _{A}),\frac{\bar{\Pi} _{1}^{\lambda }}{f_{\lambda _{1}}}\bigg\}=\frac{f_{\lambda _{1}}}{f_{\lambda _{1}}}=1.$$ $$\begin{aligned} \{\theta_{1},\Pi _{\bar{l}}^{\theta}\} &=&\{\theta_{1},\bar{\Pi}_{\bar{l}}^{\theta}-\Pi _{1}^{\theta}f_{\theta_{\bar{l}}}\}, \notag \\ &=&\{\theta_{1},\bar{\Pi}_{\bar{l}}^{\theta}\}-\{\theta_{1},\Pi _{1}^{\theta}f_{\theta_{\bar{l}}}\}, \notag \\ &=&\{f(\theta_{\bar{l}},\lambda _{A}),\bar{\Pi}_{\bar{l}}^{\theta}\}-\{f(\theta_{\bar{l} },\lambda _{A}),\Pi _{1}^{\theta}\}f_{\theta_{\bar{l}}}, \notag \\ &=&f_{\theta_{\bar{l}}}-f_{\theta_{\bar{l}}}=0. \end{aligned}$$ $$\{\theta_{\bar{l}},\Pi _{1}^{\theta}\}=\{\theta_{\bar{l}},\frac{\bar{\Pi}_{1}^{\lambda }}{ f_{\lambda _{1}}}\}=0.$$ $$\{\theta_{\bar{l}},\Pi _{\bar{m}}^{\theta}\}=\{\theta_{\bar{l}},\bar{\Pi}_{\bar{m}}^{\theta}-\Pi _{1}^{\theta}f_{\theta_{\bar{m}}}\}=\{\theta_{\bar{l}},\bar{\Pi}_{\bar{m}}^{\theta}\}=\delta _{ \bar{l}\bar{m}}.$$ The previous calculations show that the algebra (\[commutationfree\]) of the standard MGT theory is recovered from the algebra (\[commutation\_constric2\]) of the ENM. The transformation between both models is given by (\[transformation2\]). The $\protect\phi _{\bar{A}}^{1}-(\theta_{j},\Pi _{j}^{\theta})$ sector ----------------------------------------------------------------------- The primary constraints are given by $$\phi _{\bar{A}}^{1}=\bar{\Pi}_{\bar{A}}^{\lambda }-\frac{\bar{\Pi} _{1}^{\lambda }}{f_{\lambda _{1}}}f_{\lambda _{\bar{A}}}.$$ Therefore $$\begin{aligned} \{\phi _{\bar{A}}^{1},\theta_{1}\} &=&\{\phi _{\bar{A}}^{1},f(\theta_{\bar{l}},\lambda _{A})\}, \notag \\ &=&\{\bar{\Pi}_{\bar{A}}^{\lambda },f(\theta_{\bar{l}},\lambda _{A})\}-\frac{ f_{\lambda _{\bar{A}}}}{f_{\lambda _{1}}}\bigg\{\bar{\Pi}_{1}^{\lambda },\;f(\theta_{\bar{l}},\lambda _{A})\bigg\}, \notag \\ &=&-f_{\lambda _{\bar{A}}}+\frac{f_{\lambda _{\bar{A}}}}{f_{\lambda _{1}}} f_{\lambda _{1}}=-f_{\lambda _{\bar{A}}}+f_{\lambda _{\bar{A}}}=0.\end{aligned}$$ $$\{\phi _{\bar{A}}^{1},\theta_{\bar{l}}\}=\{\bar{\Pi}_{\bar{A}}^{\lambda }-\frac{ \bar{\Pi}_{1}^{\lambda }}{f_{\lambda _{1}}}f_{\lambda _{\bar{A}}},\theta_{\bar{l} }\}=0.$$ $$\begin{aligned} \{\phi _{\bar{A}}^{1},\Pi _{1}^{\theta}\} &=&\bigg\{\phi _{\bar{A}}^{1},\frac{ \bar{\Pi}_{1}^{\lambda }}{f_{\lambda _{1}}}\bigg\}, \notag \\ &=&\bigg\{\bar{\Pi}_{\bar{A}}^{\lambda },\frac{\bar{\Pi}_{1}^{\lambda }}{ f_{\lambda _{1}}}\bigg\}-\bigg\{\frac{\bar{\Pi}_{1}^{\lambda }}{f_{\lambda _{1}}}f_{\lambda _{\bar{A}}},\frac{\bar{\Pi}_{1}^{\lambda }}{f_{\lambda _{1}} }\bigg\}, \notag \\ &=&\bigg\{\bar{\Pi}_{\bar{A}}^{\lambda },\frac{1}{f_{\lambda _{1}}}\bigg\} \bar{\Pi}_{1}^{\lambda }-\frac{\bar{\Pi}_{1}^{\lambda }}{f_{\lambda _{1}}} \bigg\{f_{\lambda _{\bar{A}}},\bar{\Pi}_{1}^{\lambda }\bigg\}\frac{1}{ f_{\lambda _{1}}}, \notag \\ &=&\frac{\bar{\Pi}_{1}^{\lambda }}{(f_{\lambda _{1}})^{2}}f_{\lambda _{1}\lambda _{\bar{A}}}-\frac{\bar{\Pi}_{1}^{\lambda }}{(f_{\lambda _{1}})^{2}}f_{\lambda _{\bar{A}}\lambda _{1}}=0.\end{aligned}$$ $$\begin{aligned} \{\phi _{\bar{A}}^{1},\Pi _{\bar{m}}^{\theta}\} &=&\{\phi _{\bar{A}}^{1},\bar{\Pi} _{\bar{m}}^{\theta}-\Pi _{1}^{\theta}f_{\theta_{\bar{m}}}\}, \notag \\ &=&\{\phi _{\bar{A}}^{1},\bar{\Pi}_{\bar{m}}^{\theta}\}-\{\phi _{\bar{A} }^{1},f_{\theta_{\bar{m}}}\}\Pi _{1}^{\theta}, \notag \\ &=&\bigg\{\bar{\Pi}_{\bar{A}}^{\lambda }-\frac{\bar{\Pi}_{1}^{\lambda }}{ f_{\lambda _{1}}}f_{\lambda _{\bar{A}}},\bar{\Pi}_{\bar{m}}^{\theta}\bigg\}- \bigg\{\bar{\Pi}_{\bar{A}}^{\lambda }-\frac{\bar{\Pi}_{1}^{\lambda }}{ f_{\lambda _{1}}}f_{\lambda _{\bar{A}}},f_{\theta_{\bar{m}}}\bigg\}\Pi _{1}^{\theta}, \notag \\ &=&-\bar{\Pi}_{1}^{\lambda }\bigg\{\frac{f_{\lambda _{\bar{A}}}}{f_{\lambda _{1}}},\bar{\Pi}_{\bar{m}}^{\theta}\bigg\}- \bigg\{\bar{\Pi}_{\bar{A}}^{\lambda },f_{\theta_{\bar{m}}}\bigg\}\Pi _{1}^{\theta}+\frac{f_{\lambda_{\bar{A}}}}{ f_{\lambda _{1}}}\bigg\{\bar{\Pi}_{1}^{\lambda },f_{\theta_{\bar{m}}}\bigg\}\Pi _{1}^{\theta}, \notag \\ &=&-\bar{\Pi}_{1}^{\lambda }\bigg(\frac{f_{\lambda _{\bar{A}}\theta_{\bar{m}}}}{ f_{\lambda _{1}}}- \frac{f_{\lambda _{\bar{A}}}f_{\lambda _{1}\theta_{\bar{m}}}}{ (f_{\lambda _{1}})^{2}}\bigg)+f_{\theta_{\bar{m}}\lambda _{\bar{A}}}\Pi _{1}^{\theta}- \frac{f_{\lambda _{\bar{A}}}}{f_{\lambda _{1}}}f_{\theta_{\bar{m}}\lambda _{1}}\Pi _{1}^{\theta}, \notag \\ &=&-\Pi _{1}^{\theta}f_{\lambda _{\bar{A}}\theta_{\bar{m}}}+\Pi _{1}^{\theta}\frac{ f_{\lambda _{\bar{A}}}f_{\lambda _{1}\theta_{\bar{m}}}}{f_{\lambda _{1}}}+f_{\theta_{ \bar{m}}\lambda _{\bar{A}}}\Pi _{1}^{\theta}-\frac{f_{\lambda _{\bar{A}}}}{ f_{\lambda _{1}}}f_{\theta_{\bar{m}}\lambda _{1}}\Pi _{1}^{\theta}=0.\notag \\\end{aligned}$$ The $\protect\phi _{\bar{A}}^{1}-\protect\phi _{\bar{B}}^{1}$ sector -------------------------------------------------------------------- $$\begin{aligned} \{\phi _{\bar{A}}^{1},\phi _{\bar{B}}^{1}\} &=&\bigg\{\bar{\Pi}_{\bar{A} }^{\lambda }-\frac{\bar{\Pi}_{1}^{\lambda }}{f_{\lambda _{1}}}f_{\lambda _{ \bar{A}}},\bar{\Pi}_{\bar{B}}^{\lambda }-\frac{\bar{\Pi}_{1}^{\lambda }}{ f_{\lambda _{1}}}f_{\lambda _{\bar{B}}}\bigg\}, \notag \\ &=&-\bigg\{\bar{\Pi}_{\bar{A}}^{\lambda },\frac{f_{\lambda _{\bar{B} }}}{f_{\lambda _{1}}}\bigg\}\bar{\Pi}_{1}^{\lambda }-\bar{\Pi} _{1}^{\lambda }\bigg\{\frac{f_{\lambda _{\bar{A}}}}{f_{\lambda _{1}}},\bar{ \Pi}_{\bar{B}}^{\lambda }\bigg\}+\bigg\{\frac{\bar{\Pi}_{1}^{\lambda }}{f_{\lambda _{1}}}f_{\lambda _{\bar{A}}},\frac{\bar{\Pi}_{1}^{\lambda }}{ f_{\lambda _{1}}}f_{\lambda _{\bar{B}}}\bigg\}, \notag \\ &=&\frac{f_{\lambda _{\bar{B}}\lambda _{\bar{A}}}}{f_{\lambda _{1}}} \bar{\Pi}_{1}^{\lambda }-\frac{f_{\lambda _{\bar{B}}}f_{\lambda _{1}\lambda _{\bar{A}}}}{(f_{\lambda _{1}})^{2}}\bar{\Pi}_{1}^{\lambda }- \frac{f_{\lambda _{\bar{A}}\lambda _{\bar{B}}}}{f_{\lambda _{1}}} \bar{\Pi}_{1}^{\lambda }+\frac{f_{\lambda _{\bar{A}}}f_{\lambda _{1}\lambda _{\bar{B}}}}{(f_{\lambda _{1}})^{2}}\bar{\Pi}_{1}^{\lambda } \notag \\ &&+\bar{\Pi}_{1}^{\lambda }\bigg\{\frac{f_{\lambda _{\bar{A}}}}{f_{\lambda _{1}}},\bar{\Pi}_{1}^{\lambda }\bigg\}\frac{f_{\lambda _{\bar{B}}}}{ f_{\lambda _{1}}}+\frac{f_{\lambda _{\bar{A}}}}{f_{\lambda _{1}}}\bigg\{\bar{ \Pi}_{1}^{\lambda },\frac{f_{\lambda _{\bar{B}\;\;\;\;}}}{f_{\lambda _{1}}} \bigg\}\bar{\Pi}_{1}^{\lambda }, \notag \\ &=&-\frac{f_{\lambda _{\bar{B}}}f_{\lambda _{1}\lambda _{\bar{A}}}}{ (f_{\lambda _{1}})^{2}}\bar{\Pi}_{1}^{\lambda }+\frac{f_{\lambda _{\bar{A} }}f_{\lambda _{1}\lambda _{\bar{B}\;\;\;\;}}}{(f_{\lambda _{1}})^{2}}\bar{\Pi }_{1}^{\lambda } \notag \\ &&+\frac{\bar{\Pi}_{1}^{\lambda }}{f_{\lambda _{1}}}\bigg(\frac{f_{\lambda _{ \bar{B}}}f_{\lambda _{\bar{A}}\lambda _{1}}}{f_{\lambda _{1}}}-\frac{ f_{\lambda _{\bar{B}}}f_{\lambda _{\bar{A}}}f_{\lambda _{1}\lambda _{1}}}{(f_{\lambda _{1}})^{2}}\bigg)-\frac{\bar{\Pi}_{1}^{\lambda }}{ f_{\lambda _{1}}}\bigg(\frac{f_{\lambda _{\bar{A}}}f_{\lambda _{\bar{B} }\lambda _{1}}}{f_{\lambda _{1}}}-\frac{f_{\lambda _{\bar{A} }}f_{\lambda _{\bar{B}}}f_{\lambda _{1}\lambda _{1}}}{(f_{\lambda _{1}})^{2}}\bigg), \notag \\ &=&-\frac{f_{\lambda _{\bar{B}}}f_{\lambda _{1}\lambda _{\bar{A}}}}{ (f_{\lambda _{1}})^{2}}\bar{\Pi}_{1}^{\lambda }+\frac{f_{\lambda _{\bar{A} }}f_{\lambda _{1}\lambda _{\bar{B}}}}{(f_{\lambda _{1}})^{2}}\bar{\Pi }_{1}^{\lambda } \notag \\ &&+\frac{\bar{\Pi}_{1}^{\lambda }}{f_{\lambda _{1}}}\bigg(\frac{f_{\lambda _{ \bar{B}}}f_{\lambda _{\bar{A}}\lambda _{1}}}{f_{\lambda _{1}}}-\frac{ f_{\lambda _{\bar{A}}}f_{\lambda _{\bar{B}}\lambda _{1}}}{f_{\lambda _{1}}}\bigg)=0.\end{aligned}$$ The $\protect\phi _{\bar{A}}^{1}-\protect\phi _{\bar{B} }^{2}$ sector ------------------------------------------------------- $$\phi _{\bar{A}}^{1}=\bar{\Pi}_{\bar{A}}^{\lambda }-\frac{\bar{\Pi} _{1}^{\lambda }}{f_{\lambda _{1}}}f_{\lambda _{\bar{A}}},\quad \quad \quad \phi _{\bar{B}}^{2}=f_{\lambda _{\bar{B}}}-\frac{f_{\lambda _{1}}}{G_{1}}G_{\bar{B}},$$ $$\{\phi _{\bar{A}}^{1},\phi _{\bar{B}}^{2}\}=\{\phi _{\bar{A} }^{1},f_{\lambda _{\bar{B}}}\}-\frac{G_{\bar{B}}}{G_{1}} \{\phi _{\bar{A}}^{1},f_{\lambda _{1}}\},$$ where we have used that $\{\phi _{\bar{A}}^{1},G_{m}(\theta_{l},\Pi _{m}^{\theta})\}=0$, because we have proved that $\{\phi _{\bar{A}}^{1},\theta_{l}\}=\{\phi _{\bar{A} }^{1},\Pi _{m}^{\theta}\}=0$. Therefore $$\begin{aligned} \label{matrizx1} \{\phi _{\bar{A}}^{1},\phi _{\bar{B}}^{2}\} &=&\{\bar{\Pi}_{\bar{A} }^{\lambda },f_{\lambda _{\bar{B}}}\}-\frac{f_{\lambda _{\bar{A}}}}{ f_{\lambda _{1}}}\bigg\{\bar{\Pi}_{1}^{\lambda },f_{\lambda _{\bar{B} }}\bigg\}-\frac{G_{\bar{B}}}{G_{1}}\bigg(\{\bar{\Pi}_{\bar{A} }^{\lambda },\;f_{\lambda _{1}}\}-\frac{f_{\lambda _{\bar{A}}}}{f_{\lambda _{1}}}\bigg\{\bar{\Pi}_{1}^{\lambda },\;f_{\lambda _{1}}\bigg\}\bigg), \notag \\ &=&-f_{\lambda _{\bar{B}}\lambda _{\bar{A}}}+\frac{f_{\lambda _{\bar{A}}}}{ f_{\lambda _{1}}}f_{\lambda _{\bar{B}}\lambda _{1}}-\frac{G_{\bar{B}} }{G_{1}}\bigg(f_{\lambda _{1}\lambda _{\bar{A}}}-\frac{f_{\lambda _{\bar{A}}} }{f_{\lambda _{1}}}f_{\lambda _{1}\lambda _{1}}\bigg), \notag \\ &=&-f_{\lambda _{\bar{B}}\lambda _{\bar{A}}}+\frac{f_{\lambda _{\bar{A}}}}{ f_{\lambda _{1}}}f_{\lambda _{\bar{B}}\lambda _{1}}-\frac{f_{\lambda _{\bar{B }\;\;\;\;}}}{f_{\lambda _{1}}}\bigg(f_{\lambda _{1}\lambda _{\bar{A}}}-\frac{ f_{\lambda _{\bar{A}}}}{f_{\lambda _{1}}}f_{\lambda _{1}\lambda _{1}}\bigg),\end{aligned}$$ where in the last line we have employed the constraint $\phi _{\bar{B}}^{2}=0$ to obtain $$\frac{G_{\bar{B}}}{G_{1}}=\frac{f_{\lambda _{\bar{B}}}}{ f_{\lambda _{1}}}.$$ The invertibility of the matrix $T_{\bar{A}\bar{B}}=\{\phi _{\bar{A} }^{1},\phi _{\bar{B}}^{2}\}$ depends of the function $f=f(\theta_{\bar{l} },\lambda _{m})$. A direct calculation shows that if we take $$f=f(\Psi ),\quad \quad \quad \quad \Psi =\sum_{i}\frac{\tilde{X}_{i}^{2}}{2}, \label{particularNM}$$ where $\tilde{X}_{i}$ denotes the variables $\theta_{\bar{l}}$ and $\lambda _{A}$, the matrix $T_{\bar{A}\bar{B}}=\{\phi _{\bar{A}}^{1},\phi _{\bar{B}}^{2}\}$ becomes $$T_{\bar{A}\bar{B}}=\{\phi _{\bar{A}}^{1},\phi _{\bar{B}}^{2}\}=f^{\prime }\times \bigg(\delta _{\bar{A}\bar{B}}+\frac{\lambda _{\bar{B}\;}\lambda _{ \bar{A}}}{\lambda _{1}^{2}}\bigg),\;\;\;\;f^{\prime }=\frac{df}{d\Psi },$$ which is invertible with $$(T^{-1})_{\bar{A}\bar{B}}=\frac{1}{f^{\prime }}\times \bigg(\delta _{ \bar{A}\bar{B}}-\frac{\lambda _{\bar{B}}\lambda _{\bar{A}}}{\lambda _{c}^{2}} \bigg),\quad \quad \quad c=1,2,...,K.$$ The function (\[particularNM\]) is a generalization of the constraint $A_\mu A^\mu=n^2M^2$ used to define the abelian Nambu model (\[NANMabeliano\]), which in this case would be written as $$A_1=\sqrt{A_0^2-n^2M^2-A_jA_j},\quad\quad\quad j=1,2,$$ where we have made the identifications $$\theta_1=\theta_1(\theta_{\bar{l}},\lambda_A)=f(\theta_{\bar{l}},\lambda_A) \, \rightarrow \, A_1=A_1(A_j,A_0)=\sqrt{A_0^2-n^2M^2-A_jA_j},\, j=1,2.$$ In this case, the constraint $F=F(\theta_l,\lambda_A)=0 \,$ is subjected to the following conditions: (i) It is possible to solve for $\theta_1$ and (ii) the function $\theta_1=f(\theta_{\bar{l}},\lambda_A)$ is such that the matrix defined by Eq. (\[matrizx1\]) is invertible. The algebra resulting when solving for the coordinate $\protect\lambda _{1}$ in the ENM constraint {#Apendice2} ================================================================================================== In this situation the calculation is direct since the canonical variables of the ENM and those of the MGT coincide. The $\protect\phi^1_{\bar{m}}-(\theta_i,\Pi^\theta_i)$ sector ------------------------------------------------------------- In this case the primary constraints are $\phi _{\bar{b}}^{1}=\Pi _{\bar{b}}^{\lambda }$ and the canonical algebra of the ENM gives $$\{\phi _{\bar{b}}^{1},\theta_{i}\}=\{\phi _{\bar{b}}^{1},\Pi _{i}^{\theta}\}=0.$$ The $\protect\phi^1_{\bar{m}}-\protect\phi^1_{\bar{m}}$ sector -------------------------------------------------------------- $$\{\phi^1_{\bar{a}},\phi^1_{\bar{b}}\}=\{\Pi^\lambda_{\bar{a}},\Pi^\lambda_{ \bar{b}}\}=0.$$ The $\protect\phi^1_{\bar{m}}-\protect\phi^2_{\bar{m}}$ sector -------------------------------------------------------------- $$\phi _{\bar{b}}^{1}=\Pi _{\bar{b}}^{\lambda },\quad \quad \quad \phi _{\bar{b }}^{2}=g_{\lambda _{\bar{b}}}+\frac{G_{\bar{b}}}{G_{1}}.$$ $$X_{\bar{a}\bar{b}}=\{\phi _{\bar{a}}^{1},\phi _{\bar{b}}^{2}\}=\bigg\{\Pi _{ \bar{a}}^{\lambda },g_{\lambda _{\bar{b}}}+\frac{G_{\bar{b}}}{G_{1}}\bigg\} =g_{\lambda _{\bar{b}}\lambda _{\bar{a}}}. \label{matrix22}$$ Again, we can take as an example $$g=g(\Psi ),\quad \quad \quad \quad \Psi =\sum_{i}\frac{\tilde{X}_{i}^{2}}{2}, \label{particularNM2}$$ where $\tilde{X}_{i}$ denotes the variables $\theta_{l}$ and $\lambda _{\bar{m}}$. The matrix $X_{\bar{l}\bar{m}}=\{\phi _{\bar{l}}^{1},\phi _{\bar{m}}^{2}\}$ becomes $$X_{\bar{l}\bar{m}}=\{\phi _{\bar{l}}^{1},\phi _{\bar{m}}^{2}\}=g^{\prime }\times \bigg(\delta _{\bar{l}\bar{m}}+\frac{\lambda _{\bar{m}}\lambda _{ \bar{l}}}{g^{\prime }}g^{\prime \prime }\bigg),\quad\quad\quad g^\prime= \frac{dg}{d\Psi},$$ which is invertible $$(X^{-1})_{\bar{l}\bar{m}}=\frac{1}{g^{\prime }}\times \bigg(\delta _{\bar{l} \bar{m}}-\frac{g^{\prime \prime }}{(g^{\prime }+\lambda _{\bar{c}}\lambda _{ \bar{c}}g^{\prime \prime })}\lambda _{\bar{m}}\lambda _{\bar{l}}\bigg).$$ The function (\[particularNM2\]) is a generalization of the constraint $A_\mu A^\mu=n^2M^2$ used to define the abelian Nambu model (\[NANMabeliano\]), which in this case would be written as $$A_0=\sqrt{n^2M^2+A_iA_i},\quad i=1,2,3,$$ after making the identification $$\lambda_1=\lambda_1(\theta_l,\lambda_{\bar{A}})=g(\theta_l,\lambda_{\bar{A}})\rightarrow A_0=A_0(A_i)=\sqrt{n^2M^2+A_iA_i},\quad\quad\quad i=1,2,3.$$ In this case, the requirements upon the constraint $F=F(\theta_l,\lambda_A)=0$ are: (i) It is possible to solve for $\lambda_1$ and (ii) the function $\lambda_1=g(\theta_l,\lambda_{\bar{A}})$ is such that the matrix defined by Eq. (\[matrix22\]) is invertible. [00]{} S.L. Glashow, [Nucl. Phys.]{} 22, 579-588 (1961). S. Weinberg, [Phys. Rev. Lett.]{} 19, 1264-1266 (1967). Y. Nambu, [Phys. Rev. Lett.]{} 4, 380 (1960). Y. Nambu and G. Jona-Lasinio, [Phys. Rev.]{} 122, 345 (1961). Y. Nambu and G. Jona-Lasinio [Phys. Rev.]{} 124, 246 (1961). Y. Nambu, [Progr. Theor. Phys. Suppl. Extra]{} 190 (1968). E. A. Ivanov and V. I. Ogievetsky, [Lett. Math. Phys.]{} 1, 309 (1976). J.D. Bjorken, [Ann. Phys. (N.Y.)]{} 24, 174 (1963). G. S. Guralnik, [Phys. Rev.]{} 136, B1404 (1964). G.S. Guralnik, [Phys. Rev. Letter]{} 13, 295 (1964). J. Goldstone, [Nuovo Cim.]{} 19 (1961) 154. J. Goldstone, A. Salam and S. Weinberg, [Phys. Rev.]{} 127 (1962) 965. A.T. Azatov and J.L. Chkareuli, [Phys. Rev. D]{} 73, 065026 (2006). O.J. Franca, R. Montemayor and L.F. Urrutia, [Phys. Rev. D]{} 85, 085008 (2012). J. L. Chkareuli and Z. R. Kepuladze, [Phys. Lett. B]{} 644, 212 (2007). J. L. Chkareuli and J. G. Jejelava, [Phys. Lett. B]{} 659, 754 (2008). J. L. Chkareuli, C. D. Froggatt and H. B. Nielsen, [Nucl. Phys. B]{} 821, 65 (2009). C. A. Escobar and L. F. Urrutia, [Phys. Rev. D]{} 92 , 025013 (2015). J. L. Chkareuli, J. G. Jejelava and G. Tatishvili, [Phys. Lett. B]{} 696, 124 (2011). C. A. Escobar and L. F. Urrutia, [Phys. Rev. D]{} 92, 025042 (2015). J.L. Chkareuli, C.D. Froggatt, H.B. Nielsen, [Phys. Rev. Lett.]{} 87, 091601 (2001). J.L. Chkareuli, C.D. Froggatt, H.B. Nielsen, [Nucl. Phys. B]{} 821, 65-73 (2009) . P. A. M. Dirac, [Lectures in Quantum mechanics]{} (Academic Press, New York, 1965). S. M. Carroll, H. Tam and I. K. Wehus, [Phys. Rev. D]{} 80, 025020 (2009). P. Kraus and E. T. Tomboulis, [Phys. Rev. D]{} 66, 045015 (2002). R. Bluhm and V.A. Kostelecký, [Phys. Rev. D]{} 71, 065008 (2005). R. Potting and V.A. Kostelecký, [Phys. Rev. D]{} 79, 065018 (2009). $\ddot{O}$. F. Dayi, [Phys. Lett. B]{} 228, 3 (1989). V. A. Kosteleck[ý]{} and S. Samuel, [Phys. Rev. D]{} [40]{}, 1886 (1989). V. A. Kosteleck[ý]{} and S. Samuel, [Phys. Rev. D]{} [39]{}, 683 (1989). V. A. Kosteleck[ý]{} and S. Samuel, [Phys. Rev. Lett.]{} [63]{}, 224 (1989). V. A. Kosteleck[ý]{} and R. Lehnert, [Phys. Rev. D]{} [63]{}, 065008 (2001). V. A. Kostelecký, [Phys. Rev. D]{} [69]{}, 105009 (2004). Q. G. Bailey and V. A. Kostelecký, [Phys. Rev. D]{} 74, 045001 (2006). C. A. Hernaski, [Phys. Rev. D]{} 90, 124036 (2014). R. Bluhm, N.L. Gagne, R. Potting and A. Vrublevskis, [Phys. Rev. D]{} 77, 125007 (2008). D. Colladay and V.A. Kosteleck[ý]{}, [Phys. Rev. D]{} 55, 6760 (1997). D. Colladay and V.A. Kosteleck[ý]{}, [Phys. Rev. D]{} 58, 116002 (1998). M. Fierz and W. Pauli, [Proc. Roy. Soc. Lond.]{} A173 (1939) 211. P. Mukherjee and A. Saha, [Int. J. Mod. Phys. A]{} 24 :4305-4315 (2009). A.F. Ferrari, J.R. Nascimento, E. Passos, A.Yu. Petrov, A. J. da Silva, [Phys. Lett. B]{} 652, 174 (2007). K. Hinterbichler, [Rev. Mod. Phys.]{} 84, 671-710 (2012). M. Henneaux and C. Teitelboim, [Phys. Rev. D]{} 71, 024018 (2005).	{ "pile_set_name": "ArXiv" }
--- abstract: 'Feature selection methods are widely used in order to solve the ’curse of dimensionality’ problem. Many proposed feature selection frameworks, treat all data points equally; neglecting their different representation power and importance. In this paper, we propose an unsupervised hypergraph feature selection method via a novel point-weighting framework and low-rank representation that captures the importance of different data points. We introduce a novel soft hypergraph with low complexity to model data. Then, we formulate the feature selection as an optimization problem to preserve local relationships and also the global structure of data. Our approach for global structure preservation helps the framework overcome the problem of unavailability of data labels in unsupervised learning. The proposed feature selection method treats with different data points based on their importance in defining the data structure and representation power. Moreover, since the robustness of feature selection methods against noise and outlier is of great importance, we adopt low-rank representation in our model. Also, we provide an efficient algorithm to solve the proposed optimization problem. The computational cost of the proposed algorithm is lower than many state-of-the-art methods which is of high importance in feature selection tasks. We conducted comprehensive experiments with various evaluation methods on different benchmark data sets. These experiments indicate significant improvement, compared with state-of-the-art feature selection methods.' address: 'Amirkabir University of Technology, Department of Computer Engineering and Information Technology' author: - 'Ammar Gilani[^1]' - 'Maryam Amirmazlaghani[^2]' bibliography: - 'ref.bib' title: 'Unsupervised Hypergraph Feature Selection via a Novel Point-Weighting Framework and Low-Rank Representation' --- feature selection, hypergraph embedding, low-rank representation, soft hypergraph, joint learning Introduction ============ Nowadays confronting high dimensional data is expected in practical applications. The complexity of many algorithms in various fields like machine learning and pattern recognition highly depends on the dimensionality of data. Two types of dimensionality reduction algorithms can be used to solve the ’curse of dimensionality’ problem: feature extraction and feature selection. Feature extraction methods try to define new features based on primary features and redescribe data based on them. In opposite, feature selection techniques aim to select the most descriptive subset of features. Unlike feature extraction techniques, feature selection algorithms preserve the primary representation of the data [@RFSTB]. Feature selection methods are categorized into three groups: supervised, semi-supervised, and unsupervised. Supervised and semi-supervised methods like generalized fisher score (GFS) [@GFS], hypergraph based information-theoretic feature selection (MII\_HG) [@MII_HG], and hypergraph regularized Lasso (HLasso) [@HLasso] need labeled data to perform feature selection. Although these methods can be effective, collecting labels for data can be time-consuming or even impractical in many cases. Unlike supervised and semi-supervised methods, unsupervised methods such as feature selection via joint embedding learning and sparse regression (JELSR) [@JELSR] and self-representation hypergraph low-rank feature selection (SHLFS) [@SHLFS] need no labeled data for feature selection. Hence unsupervised methods are used more widely in many applications.\ In unsupervised feature selection methods, providing an efficient representation of data is of great importance. Because the structure of data is preserved based on the constructed representation which without having an effective representation, preservation of the defined structure can be ineffective. In the last decade, various algorithms such as JELSR [@JELSR], laplacian score (LapScore) [@LapScore], similarity preserving feature selection (SPFS) [@SPFS], and multi-cluster feature selection (MCFS) [@MCFS] used graphs to model data. In graphs each edge can only be connected to at most another edge; So, graph-based methods can be adopted to model only dual relationships among data points. On the other hand, in real applications, different data points can be related to each other as a group and as a result, there exist multiple relationships among them which should be modeled and in the next steps, preserved. To overcome this problem, in the last few years, some algorithms have been using hypergraphs instead to model interconnection among data points; For instance, JHLSR [@JHLSR], MII\_HG [@MII_HG], and SHLFS [@SHLFS]. In hypergraphs, hyperedges are used instead of conventional edges. Hyperedges can connect any number of vertices. Although hypergraphs improve ordinary graphs in many ways, they have an important limitation: there are only two states for a vertex and a hyperedge: a hyperedge completely consists of that vertex or not, but the vertex may need to have partial participation in that hyperedge; Because importance of other vertices in that hyperedge can be more or less critical. So, a generalized form of hypergraphs is needed in which amount of attendance of a vertex in a hyperedge can adopt continuous values instead of only two states.\ Feature selection methods usually encounter the problem of inaccurate measurements and outlier in practical applications. So, it is of great importance for a feature selection method to overcome this problem. Proposing a model without handling noise and outlier can lead to focus on wrong data structure and as a result, inefficient feature selection. One way to overcome this problem is using low-rank representation in hypergraph-based feature selection methods proposed in [@HLR_FS; @SHLFS]. In other words, hypergraph low-rank feature selection (HLR\_FS) [@HLR_FS] and self-representation hypergraph low-rank feature selection (SHLFS) [@SHLFS] use low-rank representation in order to deal with noise and outlier.\ One common disadvantage of the mentioned methods is behaving different data points equally; While different data points can vary much in importance and role in defining the data structure.\ Another important point in feature selection methods is computational cost. As it was clarified, it is awaited to meet high dimensional data in the task of feature selection. Therefore, the computational complexity of an algorithm is of high importance. An algorithm can be effective on small-sized data, but it may be infeasible to adopt it on a high dimensional practical data set.\ In this paper, to overcome the weaknesses of the previously proposed methods, we introduce a novel low-rank feature selection method using soft hypergraphs which captures the importance of different data points. In the proposed point-weighting framework, we treat different data points based on their representation power and role in defining the data structure. We propose using soft hypergraphs in order to resolve the limitation of ordinary hypergraphs. Unlike ordinary hypergraphs, in soft hypergraphs, different vertices can have different participation in a hyperedge which leads to more accurate modeling. Moreover, the proposed approach uses cluster centroids to build the hypergraph which results in reducing the computational cost, and also better handling of noise, outlier, and redundancy. The reason for this can be described as follows: Usually, in real situations, some data points can have similar properties. So in order to preserve data structure, there is no need to use all data points. Points with analogous properties can be grouped as a cluster and a representative data point can be used instead of the cluster to represent the properties of that group. In this way, computational complexity decreases and the effect of noise and outlier reduces; Because the small differences in properties of some data points can be only the result of noise; Moreover, redundant information of almost similar data points will be discarded. The best choice for this representative point is the centroid of that cluster; Since it has a good representation power. We use the constructed hypergraph to preserve local relationships between centroids.\ To preserve the complete structure of data, the global structure needs to be preserved along with local relationships between centroids. For this purpose, we try to maintain the global structure of data by adding a term which also helps resolve the problem of data labels unavailability in unsupervised learning.\ So, in brief, in this paper, by weighting points, we focus on data points with high representation power; Especially cluster centroids. We also adopt low-rank representation and propose a novel term for global structure preservation. We have tested our proposed framework using various evaluation methods on different benchmark datasets which indicates significant improvement, compared with state-of-the-art feature selection methods. The innovative aspects of the proposed method can be summarized as follows: i) Introducing a new soft hypergraph to model data for unsupervised feature selection. ii) Proposing a point-weighting framework. iii) Adopting low-rank representation in the point-weighting framework. iv) Introducing a new approach to preserve the global structure of data. v) Providing an efficient algorithm to solve the proposed optimization problem.\ The notations used in this paper and their definitions are summarized in table \[table:notations\]. Here the organization of the rest of the paper is explained. In Section \[hypergraph-based feature selection\], we review concepts of hypergraph-based feature selection. In Section \[proposed method for feature selection\], we explain our proposed framework for feature selection. In section \[optimization\], we provide an efficient algorithm to solve the proposed optimization problem. In Section \[computational cost and convergence study\], we calculate the computational cost of our proposed algorithm. Section \[experiments\] exhibits our experimental results. Finally, in Section \[conclusion\], we conclude our paper. [ \|P[3cm]{}\|l\| ]{}\ $m_i^j$ & the element in ith row and jth column of matrix M (a typical matrix)\ tr(M) & the trace of M\ $\|\|M\|\|_F$ & the Frobenius norm of M\ $M^T$ & the transpose of M\ $M^{-1}$ & the inverse of M\ X $\in$ $R^{n\times d}$ & the data matrix; Each row is a data point\ n & the number of data points\ m & the number of centroids, which is also the number of vertices and hyperedges\ C $\in$ $R^{m\times d}$ & centroids matrix; Each row is a centroid\ $\zeta_i$ & the cluster number of the ith data point\ H $\in$ $R^{m\times m}$ & the incidence matrix of the soft hypergraph\ d(v) & the degree of vertex v\ $D_v$ $\in$ $R_+^{m\times m}$ & the diagonal matrix of vertices degrees\ $\delta(e)$ & the degree of hyperedge e\ $D_e$ $\in$ $R_+^{m\times m}$ & the diagonal matrix of hyperedges degrees\ w(e) & the weight function which assigns each hyperedge like e, a non-negative weight\ W $\in$ $R_+^{m\times m}$ & the diagonal matrix of hyperedges weights\ diag(W) $\in$ $R_+^m$ & the vector of hyperedges weights (the main diagonal of W)\ T $\in$ $R^{d\times k}$ & the linear transformation matrix used to embed data\ $\Delta $ & the hypergraph laplacian\ $I_n$ $\in$ $R^{n\times n}$ & the identity matrix\ k & the dimensionality of embedding\ d & the initial number of features\ Hypergraph-based feature selection {#hypergraph-based feature selection} ================================== An important step in feature selection algorithms is modeling data. Graphs can be used in order to model local relationships between data points. But they can model only dual relationships. To overcome this problem, in recent studies, hypergraphs has been used instead [@SHLFS; @JHLSR; @HLasso]. In this section, first, we explain the fundamentals of hypergraphs. Afterward, principals of soft hypergraphs are explained. Then, we review hypergraph-based feature selection methods quickly. Hypergraph fundamentals {#hypergraph fundamentals} ----------------------- An ordinary graph can be defined as the triplet: G = (V, E, w); Where V is the set of vertices, E is the set of edges, and w is a function which assigns each edge a weight. An edge consists of at most two vertices; So it can only model dual relationships among vertices. As it was explained earlier, in Section \[introduction\], in many applications, there exist multiple relationships among data points. So a conventional graph is not appropriate to thoroughly model relationships among them. To tackle this limitation of conventional graphs, hypergraphs can be used instead. Hypergraphs can be considered as the generalized form of simple graphs. They can be shown using the following notation: $G_H$ = (V, E, w). Similar to graphs, V and w are the set of vertices and the weight function respectively. Although instead of simple edges, hyperedges are adopted in hypergraphs to show interconnection among vertices; They are denoted by E. Hyperedges are the extended mode of traditional edges and can include multiple vertices; As a result, they can model multiple relationships amid vertices. In order to determine the relationship between different vertices and hyperedges, the incidence matrix H is defined as follows: $$\label{eq:H} h_v^{e}= \begin{cases} 1, & \text{if}\ vertex~v \in hyperedge~e \\ 0, & \text{elsewhere} \end{cases}$$ Where $h_v^{e}$ is the element in vth row and eth column of matrix H. Akin to conventional graphs, degree is defined for vertices and hyperedges. Degree of a vertex is the total weights of hyperedges which it is comprised in, i.e. $$\label{eq:dv} d(v)=\sum_{e~=~1}^{\|E\|} w(e)h_v^{e}$$ Where $\|E\|$ is the number of hyperedges. Degree of a hyperedge is the total amount of contributions of different vertices in it which in conventional hypergraphs is equal to the number of vertices it embraces, i.e. $$\label{eq:delta_e} \delta(e)=\sum_{v~=~1}^{\|V\|} h_v^{e}$$ Where $\|V\|$ is the number of vertices, $h_{v}^{e}$ is the element in vth row and eth column of matrix H. Soft hypergraph --------------- Even though hypergraphs cover the main limitation of ordinary graphs, they have the problem of equal participation of different vertices in their hyperedges. In hyperedges, all vertices are treated equally, but different vertices have different importance and contributions in hyperedges. In addition, there is no need to define the incidence matrix binary, While it can be continuous; At least in our case. This type of hypergraph is called soft hypergraph [@IVFH]. An example of soft hypergraph is provided in figure \[figure:hypergraph example\]; $E=\{e_1, e_2, e_3\}$ is the set of hyperedges and $V=\{v_1, v_2, v_3, v_4, v_5\}$ is the set of vertices. Hyperedge $e_3$ models multiple relationship between vertices $v_3$, $v_4$, and $v_5$. Since the incidence matrix is not binary, these vertices could have different participation in that hyperedge, i.e. 0.18, 0.59, and 0.42 respectively. [0.325]{} ![An example of soft hypergraph.[]{data-label="figure:hypergraph example"}](images/hypergraph_example.png "fig:") [0.325]{} $$H=~ \begin{blockarray}{cccc} e_1 & e_2 & e_3\\ \begin{block}{(ccc)c} 0 & 0.38 & 0 & v_1\\ 1 & 0.15 & 0 & v_2\\ 0.73 & 0 & 0.18 & v_3\\ 0 & 0 & 0.59 & v_4\\ 0 & 0 & 0.42 & v_5\\ \end{block} \end{blockarray}$$ A review on hypergraph-based feature selection methods {#A review on hypergraph-based feature selection methods} ------------------------------------------------------ In this section, we review some state-of-the-art feature selection methods which use hypergraphs to perform data modeling. We list related algorithms and provide a brief explanation about them.\ JHLSR [@JHLSR]: This framework adopts sparse representation [@RFRSR] to construct its hypergraph. Sparse representation for a data point can be described as the minimal subset of linearly dependent vertices. JHLSR needs one hyperedge for each data point; Moreover, Because of not using soft hypergraph, they have to repeat the process of finding the sparse representation of each data point several times with different sparsification parameters. As a result, the constructed hypergraph becomes very large which increases the computational complexity of the proposed model. As it was explained earlier, encountering noise and outlier is inevitable, but JHLSR does not provide any solution to handle this problem.\ SHLFS [@SHLFS]: Some supervised feature selection methods such as [@HLR_FS; @ASFS] try to generate the response vector using the linear combination of features and in the last step, report features with the most contribution in generating the response vector, as output. SHLFS modifies this method to be applicable in unsupervised learning and tries to generate each feature, using the linear combination of other features. SHLFS also adopts low-rank representation to handle noise and outlier.\ HLasso [@HLasso]: In order to construct the hypergraph, HLasso connects every data point to its k nearest neighbors which results in a large hypergraph which imposes much computational cost; Moreover, since it adopts ordinary hypergraphs, instead of soft hypergraphs, the nearest neighbors of each data point are treated equally, while they can have different distances from that point. After modeling data, HLasso uses hypergraph laplacian to preserve the distances of each point from its nearest neighbors. HLasso does not handle noise and outlier.\ HLR\_FS [@HLR_FS]: This method constructs an ordinary hypergraph by connecting data points with the same label. HLR\_FS also uses a low-rank constraint to make the model robust against noise and outlier.\ The above methods overlook the importance of treating different data points based on their representation power and role in defining the data structure. Moreover, none of them adopt soft hypergraphs in order to model data. Proposed method for feature selection {#proposed method for feature selection} ===================================== In this paper, we propose an unsupervised hypergraph feature selection method via a novel point-weighting framework and low-rank representation (referred to as HPWL). Since each centroid is the indicator of its cluster and its corresponding points, it has good representational power. So at the first step, we should find the centroids. In this way, any general clustering method can be adopted. We use the k-means algorithm in our experiments. the k-means algorithm is as follows: Input: X, m.\ Output: $\mu$, $\zeta$.\ Step 1: Initialization:\ Initialize cluster centers $\mu_1$, $\mu_2$, ..., $\mu_m$ randomly;\ Step 2: Updating clusters and their centers:\ repeat\ $\zeta_i = \operatorname{\arg\!\min}_j{\|\|x_i - \mu_j\|\|^2}$, for i = 1, 2, ..., n\ $\mu_j = \frac{\sum_{i=1}^m 1\{\zeta_i=j\}x_i}{\sum_{i=1}^m 1\{\zeta_i=j\}}$, for j = 1, 2, ..., m\ until* convergence;\ Where X is the data matrix, $x_i$ is the ith data point, m is the number of clusters, n is the number of data points, $\mu_i$ is the center of the ith cluster, and $\zeta_i$ indicates the cluster number of the ith data point. Using k-means, we determine m clusters. Since in our proposed method each centroid needs to be one of the data points, we consider our ith centroid, i.e. $c_i$ the nearest data point to $\mu_i$; For i = 1, 2, ..., m. We use obtained centroids for hypergraph construction. Treating all data points equally instead of focusing on centroids, results in useless computational complexity and redundant information; Moreover, it makes the algorithm sensitive to noise and outlier which leads to an ineffective selection of features. Soft hypergraph construction ---------------------------- We construct our soft hypergraph by focusing on the defined centroids. By using soft hypergraph, we can model and then preserve multiple relationships among different centroids precisely. Each centroid is considered as a vertex in the hypergraph. To generate hyperedges, we connect every centroid to its $l$ nearest centroids. To construct the H matrix, suppose that we are generating the hyperedge corresponding to the centroid i, i.e. $h^i$. Its jth element is computed as follows: $$\label{eq:e} h_j^i= \begin{cases} a_j^i, & \text{if}\ c_j \in \text{\{$l$ nearest centroids of } c_i\}\\ 0, & \text{elsewhere} \end{cases}$$ Where $h_j^i$ is the element in jth row and ith column of matrix H, $a_j^i = {exp}{\Big(-\frac{\|\|c_i - c_j\|\|_{2}^{2}}{\sigma^{2}}\Big)}$; $\sigma$ is average Euclidean distance among centroids and $c_i$, $c_j$ are two centroids. $H~=~[h^1,~h^2,...~,~h^m]$ where $h^i$ is the ith column of H, i.e. the ith hyperedge which is the hyperedge corresponding to the ith centroid. Our algorithm is capable of differentiating the importance of different centroids. We use equal values to initialize the weights of hyperedges. To select the most informative hyperedges and assign a higher weight to them, hyperedges weights should be updated. So the number of hyperedges directly affects the computational complexity. Hence a large number of hyperedges as used in [@JHLSR] imposes too much computational complexity. In the proposed method, the number of vertices and the number of hyperedges are equal to the number of centroids which is much lower than the number of data points. So it does not generate too many hyperedges and its computational cost is much lower than other hypergraph-based methods such as [@JHLSR; @EMHG; @HLasso; @HSIR]. Data embedding {#data embedding} -------------- To perform the task of feature selection, we use a transformation matrix (T) to project data into a low dimensional space. If $x_i$, $x_j$ are close to each other, we want $x_i T$, $x_j T$ to be close to each other too. After finding the optimal T, we use $\ell_{2}$ norm to find primary features which have the most role in setting up this transformation matrix.\ As it was explained earlier in Section \[introduction\], real-world data sets generally are corrupted and include noise and outlier; As a result, the rank of matrices is high in practical applications [@SRRDR; @MCNE]; Moreover, it has been demonstrated that high-dimensional data sets usually have low-dimensional representations, i.e. they can be redescribed in a low-dimensional space [@DGAHNP]. In order to remedy the aforementioned corruption and robust selection of features, we can use low-rank representation. Hence, we suppose that the transformation matrix is low-rank and we consider the following low-rank constraint for it: Since features are finally ranked according to matrix T, in order not to use uninformative and redundant features in data embedding, matrix T also needs to be sparse; Moreover, some elements of T may have small values which can be only the effect of errors. Trying to make matrix T sparse, also helps to set the values of these elements to zero. Local relationships between centroids {#local relationships between centroids} ------------------------------------- Importance of preservation of local neighborhood relationships among data points has been well discussed in recent works [@LapScore; @JELSR]. In order to maintain data local relationships, we focus on the centroids and then we try to preserve local neighborhood relationships among them. It has been demonstrated that using Gaussian kernel as similarity function can model local neighborhoods [@SpecCl]. So, we try to preserve the Gaussian similarity between different centroids. In other words, we try to maintain the distances between different centroids. To achieve this goal, we need to minimize the following objective function: $$\label{eq:local1} \Psi^{(1)}(W,~T)~=~\frac{1}{2}{\mathlarger{\sum}_{c_{i},c_{j}\in V}}\alpha_{i,j}\times\|\|(c_iT)^T~-~(c_jT)^T\|\|_2^2$$ where $c_iT$ and $c_jT$ are transformed centroids, $\|\|.\|\|_2$ is $\ell_2$ norm of a vector, and $\alpha_{i,j}$ is calculated using the following formula: $$\label{eq:local2} \alpha_{i,j}~=~\mathlarger{\sum}_{e \in E}~\frac{w(e)h(c_i,~e)h(c_j,~e)}{\delta(e)d(c_i)}$$ Where $c_i$ is the ith centroid, $d(c_i)$ is its degree in hypergraph, w(e) is the weight of hyperedge e, and $\delta(e)$ is its degree. $\alpha_{i,j}$ determines the importance of closing $c_iT$ and $c_jT$ which is zero in case centroids i and j are not connected to each other; Because h($c_i$, e)h($c_j$, e) = 0. In the other case, namely h($c_i$, e)h($c_j$, e) $>$ 0, the higher value of this factor means these two centroids are closer which leads to a larger $\alpha_{i,j}$; Moreover, $\frac{w(e)}{\delta(e)d(c_i)}$ determines the normalized weight of that hyperedge for $c_i$. So the larger $\alpha_{i,j}$ is, the higher the importance of closing $c_iT$ and $c_jT$ is. As it was explained, $\alpha_{i,j}$ also indicates importance of closing $c_i$ and $c_j$. So minimizing $\Psi^{(1)}$ leads to preservation of distances among different centroids. We have: $$\label{eq:local3} \Psi^{(1)}(W,~T)~=~\frac{1}{2}{\mathlarger{\sum}_{c_{i},c_{j}\in V}}\alpha_{i,j}\times\|\|(c_iT)^T~-~(c_jT)^T\|\|_2^2~=~tr(T^TC^T\Delta CT)$$ Where C is matrix of centroids and $\Delta$ is the unnormalized hypergraph laplacian matrix defined by [@LHC]: $$\label{eq:local4} \Delta ~=~I_{\|V\|}~-~D_v^{-1}HWD_e^{-1}H^T$$ $D_v$, $D_e$, and W are diagonal matrices of vertices degrees, hyperedges degrees, and hyperedges weights respectively and $I_{\|V\|}$ is a $\|V\| \times \|V\|$ identity matrix. It is worth noting that generally, minimizing $tr\big(F(C)^T\Delta F(C)\big)$, where F(C) is a function (in our case CT), leads to closing $f(c_i)$ and $f(c_j)$ where $f(c_i)=F(C)_i$, and $c_i$ and $c_j$ are in at least one common hyperedge; In addition, the amount of effort to close $f(c_i)$ and $f(c_j)$ is determined by contributions of $c_i$ and $c_j$ in their common hyperedge and the normalized hyperedge weight. This analysis clarifies the importance of using soft hypergraphs where different vertices are eligible to have different participation in a hyperedge because elements of the incidence matrix can adopt continuous values.\ Considering the low-rank constraint for transformation matrix, we have: $$\label{eq:local5} \begin{aligned} \min_{W,~T}\Psi^{(1)}(W,~T)~=~tr(T^TC^T\Delta CT)\\ \text{s.t. rank(T) $\leq$ r} \end{aligned}$$ Where C is the matrix of centroids and considering low-rank constraint is helpful in handling noise and outlier. We also add $\ell2,1$ norm of matrix T as a regularizer to try to make it sparse and as a result, not to use uninformative features in data embedding and reducing the effect of errors: $$\label{eq:W1} \Theta^{(1)}(T)~=~\|\|T\|\|_{2, 1}$$ Although $\|\|T\|\|_{2, 1}$ is convex, its derivative does not exist when at least one of its rows is zero. Hence we adopt the definition in [@JELSR]: $$\label{eq:W2} \Theta(T)~=~tr(T^TBT)$$ Where B is a $d\times d$ diagonal matrix which its ith element is calculated as follows: $$\label{eq:B} b_i^i~=~\frac{1}{2\|\|t_i\|\|_2}$$ Where $t_i$ is the ith row of T. Adding this term, changes the optimization problem to: $$\label{eq:local5_2} \begin{aligned} \min_{W,~T}\Psi^{(1)}(W,~T)~+~\rho\Theta(T)=~tr(T^TC^T\Delta CT)~+~\rho\times tr(T^TBT)\\ \text{s.t. rank(T) $\leq$ r} \end{aligned}$$ Where $\rho$ is a regularization parameter. In the next step, in order to control model complexity and setting the weight of uninformative or redundant hyperedges to zero, two constraints are added to the optimization problem (\[eq:local5\_2\]) which results: $$\label{eq:local6} \begin{aligned} \min_{W,~T}\Psi(W,~T)~+~\rho\Theta(T)~=~tr(T^TC^T\Delta CT)~+~\rho\times tr(T^TBT)~+~\kappa\|\|diag(W)\|\|_2^2\\ s.t.~\sum_{i~=~1}^mW_i=1,~W_i\geq0~for~i=0,1,...,m,~\text{rank(T) $\leq$ r} \end{aligned}$$ Where diag(W) denotes the main diagonal of W, i.e. the vector of hyperedges weights, $W_i = w_i^i$ which is the weight of the ith hyperedge, $\kappa$ is a regularization parameter, and $\|\|.\|\|_2$ is $\ell_{2}$ norm of a vector. Although the low-rank constraint added in Section \[data embedding\] helps our proposed framework become robust to noise and outlier, it makes the optimization problem non-convex and NP-hard [@CSCTR]. An appropriate way to apply this constraint is to produce T by multiplying two low-rank matrices, i.e. T = PQ where P $\in$ $R^{d\times r}$ and Q $\in$ $R^{r\times k}$: $$\label{eq:local7} \begin{aligned} \min_{W,~P,~Q}~[\Psi(W,~P,~Q)+\rho\Theta(P,~Q)]~=~tr(Q^TP^TC^T\Delta CPQ)~+~\rho\times tr(Q^TP^TBPQ)~+\\ \kappa\|\|diag(W)\|\|_2^2~~~s.t.~\sum_{i~=~1}^mW_i=1,~W_i\geq0~for~i=0,1,...,m \end{aligned}$$ Global structure ---------------- As it was discussed, our proposed framework preserves local relationships between centroids by maintaining the distances among them. But in order to conserve the complete structure of data, the global structure of data should be preserved along with local relationships. Since data labels have a significant role in defining the structure of data, many supervised methods such as [@mRMR; @PattC; @MII_HG; @GFS] focus on them to preserve data structure. But in unsupervised learning, data labels are not available. This motivates us to introduce an alternative way of preserving global structure in our unsupervised framework. It has been demonstrated that data points within the same class have high linear dependency and correlation [@RFRSR]. So, trying to preserve the correlation between data points, can be used as an alternative to focusing on data labels for unsupervised learning. Hence, in the proposed framework, we try to maintain the correlation between data points for global structure preservation; Especially data points with more representation power and role in defining the data structure. Since hypergraph construction and using it have high computational cost, we proposed using cluster centroids in Section \[local relationships between centroids\]. However, using only centroids leads to inevitable loss of some information. In order to use all the information with minor computational cost, we use all data points in the equation (\[eq:global1\]). In this way, our goal can be minimizing the following objective function: $$\label{eq:global1} \Upsilon^{(1)}(T)~=~\|\|(XT)(XT)^T~-~XX^T\|\|_F^2$$ Where XT is the transformed data and $\|\|.\|\|_F$ is the Frobenius norm of a matrix. In order to normalize the above correlation matrices, we divide each matrix by the length of its generating vectors, i.e. k and d respectively: $$\label{eq:global2} \Upsilon^{(2)}(T)~=~\|\|\frac{XT(XT)^T}{k}~-~\frac{XX^T}{d}\|\|_F^2$$ Which can be reformulated as follows: $$\label{eq:global3} \Upsilon^{(3)}(T)~=~\|\|XT(XT)^T~-~\frac{k}{d}~\times~XX^T\|\|_F^2$$ Where k is the dimensionality of embedding and d is the initial number of features. As in local relationships preservation, more representative data points should be emphasized on, especially the centroids. As a data point gets further from its corresponding centroid, not only its importance and representation power reduces, but it gets more likely an outlier. So it is not of high importance to preserve correlation for these data points. For this purpose, we define diagonal matrix D which its ith element is calculated as follows: $$\label{eq:D} d_i^i= \begin{cases} 1, & \text{if}\ x_i \text{ is one of the centroids}\\ affinity(x_i, c_{\zeta_i}), & \text{elsewhere} \end{cases}$$ Where $c_{\zeta_i}$ is the centroid of the cluster which $x_i$ belongs to. So, $affinity(x_i, c_{\zeta_i})$ is the Gaussian similarity between the data point i and its corresponding centroid. Note that $affinity(x_i, c_{\zeta_i})~\leq~1$. The further a data point gets from its corresponding centroid, the smaller its corresponding value in D will be. To focus on data points with better representation power and neglect outliers we change the equation (\[eq:global3\]) to: $$\label{eq:global23} \Upsilon^{(4)}(T)~=~\|\|D^\frac{1}{2}[XT(XT)^T~-~\frac{k}{d}~\times~XX^T]D^\frac{1}{2}\|\|_F^2$$ We are actually multiplying the changes of correlations between different data points by their importance, prior to calculating the Frobenius norm which is equivalent to emphasizing on more representative points, especially the centroids. The problem is that the defined term is not convex with respect to T. So we need to modify its representation: $$\label{eq:global4} \begin{aligned} \Upsilon^{(4)}(T)~=~\|\|D^\frac{1}{2}[XT(XT)^T~-~\frac{k}{d}~\times~XX^T]D^\frac{1}{2}\|\|_F^2\\ =~\|\|D^\frac{1}{2}XT(XT)^TD^\frac{1}{2}~-~\frac{k}{d}~\times~D^\frac{1}{2}XX^TD^\frac{1}{2}\|\|_F^2\\ =~\|\|D^\frac{1}{2}XTT^TX^TD^\frac{1}{2}~-~\frac{k}{d}~\times~D^\frac{1}{2}XX^TD^\frac{1}{2}\|\|_F^2 \end{aligned}$$ And in the last step, to make it convex, we use an approximation provided in [@SPFS] and we have: $$\label{eq:global5} \Upsilon(T)~=~\|\|D^\frac{1}{2}XT~-~Z_k\|\|_F^2$$ Where $Z_k = \Gamma_k\Xi_k^{1/2}$, $\Xi_k$ is the diagonal matrix of k eigenvalues of $[\frac{k}{d}~\times~D^\frac{1}{2}XX^TD^\frac{1}{2}]$ sorted descendingly, and $\Gamma_k$ is the matrix of their corresponding eigenvectors put together column-wise. So, minimizing $\Upsilon(T)$ results in preserving the correlation between data points. Final optimization problem -------------------------- To find appropriate solutions for W and T, we should minimize $\Psi(W, T)$, $\Theta(T)$, and $\Upsilon(T)$ simultaneously, and T needs to be replaced with PQ. In this way, our optimization problem becomes as follows: $$\label{eq:main_final} \begin{aligned} \min_{W,~P,~Q}~[\Psi(W,~P,~Q)+\tau\Upsilon(P,~Q)+\rho\Theta(P,~Q)]~=~tr(Q^TP^TC^T\Delta CPQ)~+~\tau\|\|D^\frac{1}{2}XPQ~-~Z_k\|\|_F^2\\ +~\kappa\|\|diag(W)\|\|_2^2~+~\rho\times tr(Q^TP^TBPQ)~~~s.t.~\sum_{i~=~1}^mW_i=1,~W_i\geq0~for~i=0,1,...,m \end{aligned}$$ The optimization problem (\[eq:main\_final\]) is the final optimization problem we aim to solve. Optimization ============ Our proposed objective function consists of three factors: W, P, and Q. Even though our optimization problem is generally non-convex and complicated, if we fix two factors and consider the other factor a variable, the problem becomes much simpler. So, we use coordinate descent algorithm to solve it which is summarized in algorithm \[algorithm:main\]. Each outer iteration includes three steps as follows: 1. Fix Q, W, optimize P. 2. Fix P, W, find the optimal value of Q. 3. Solve for W by fixing P and Q. Input: X and parameters $\tau$, $\rho$, $\kappa$, and r.\ Output: features scores.\ Step 1: Hypergraph construction.\ 1.1 cluster X using k-means algorithm;\ 1.2 generate matrix H based on the equation (\[eq:e\]);\ Step 2: Updating P, Q:\ repeat\ 2.1 Update P, Q using the equations (\[eq:opt\_p3\]), (\[eq:opt\_q3\]) respectively;\ 2.2 Update B according to the equation (\[eq:B\]);\ until convergence;\ Step 3: Updating W:\ Update W using the coordinate descent algorithm explained in Section \[updating w\];\ Step 4: Updating $\Delta $:\ Update $\Delta $ based on the equation (\[eq:local4\]);\ Step 5: Checking convergence and calculating features scores if necessary:\ If P, Q have converged, calculate features scores using $\ell2$ norm of T rows, otherwise go to Step 2;\ Updating P ---------- By considering Q and W as constants, we need to solve the following optimization problem: $$\label{eq:opt_p1} \begin{aligned} \min_P~tr(Q^TP^TC^T\Delta CPQ)~+~\tau\|\|D^\frac{1}{2}XPQ~-~Z_k\|\|_F^2\\ +~\rho\times tr(Q^TP^TBPQ) \end{aligned}$$ Since the objective function in (\[eq:opt\_p1\]) is convex with respect to P, it suffices to take its derivative with respect to P and set it to zero: $$\label{eq:opt_p2} \begin{aligned} C^T(\Delta + \Delta^T)CPQQ^T~+~2\tau X^TD^\frac{1}{2}(D^\frac{1}{2}XPQ-Z_k)Q^T+~2\rho BPQQ^T~=~0 \end{aligned}$$ Which results: $$\label{eq:opt_p3} \begin{aligned} P~=~\tau[C^T\Delta^\prime C~+~\tau X^TDX~+~\rho B]^{-1}X^TD^\frac{1}{2}Z_kQ^{-1} \end{aligned}$$ Where $\Delta^\prime = (\Delta + \Delta^T) / 2$. Updating Q ---------- Similarly, we need to solve the following optimization problem to update Q: $$\label{eq:opt_q1} \begin{aligned} \min_Q~tr(Q^TP^TC^T\Delta CPQ)~+~\tau\|\|D^\frac{1}{2}XPQ~-~Z_k\|\|_F^2\\ +~\rho\times tr(Q^TP^TBPQ) \end{aligned}$$ Setting its derivative with respect to Q, to zero results: $$\label{eq:opt_q2} \begin{aligned} P^TC^T(\Delta + \Delta^T)CPQ~+~2\tau P^TX^TD^\frac{1}{2}(D^\frac{1}{2}XPQ-Z_k)+~2\rho P^TBPQ~=~0 \end{aligned}$$ Which leads to: $$\label{eq:opt_q3} \begin{aligned} Q~=~\tau P^{-1}[C^T\Delta^\prime C~+~\tau X^TDX~+~\rho B]^{-1}X^TD^\frac{1}{2}Z_k \end{aligned}$$ Where $\Delta^\prime = (\Delta + \Delta^T) / 2$. Updating W {#updating w} ---------- Since P and Q are fixed, in order to update W, we need to solve the following optimization problem: $$\label{eq:opt_w1} \begin{aligned} \min_W~tr(Q^TP^TC^T\Delta CPQ)~+~\kappa\|\|diag(W)\|\|_2^2~~~s.t.~\sum_{i~=~1}^mW_i=1,~W_i\geq0~for~i=0,1,...,m \end{aligned}$$ Based on the equation (\[eq:local4\]), $\Delta ~=~I_{\|V\|}~-~D_v^{-1}HWD_e^{-1}H^T$. The optimization problem (\[eq:opt\_w1\]) can be simplified to: $$\label{eq:opt_w2} \begin{aligned} \min_W~tr(Q^TP^TC^TD_v^{-1}HWD_e^{-1}H^TCPQ)~+~\kappa\|\|diag(W)\|\|_2^2\\ s.t.~\sum_{i~=~1}^mW_i=1,~W_i\geq0~for~i=0,1,...,m \end{aligned}$$ Let $R=Q^TP^TC^TD_v^{-1}H$, $S=D_e^{-1}H^TCPQ$. for simplification. We have: $$\label{eq:opt_w3} \begin{aligned} \min_W~tr(RWS)~+~\kappa\|\|diag(W)\|\|_2^2\\ s.t.~\sum_{i~=~1}^mW_i=1,~W_i\geq0~for~i=0,1,...,m \end{aligned}$$ We have the following equalities: $$\label{eq:opt_w4} \begin{aligned} tr(RWS)~+~\kappa\|\|diag(W)\|\|_2^2~=~tr\Bigg(\begin{bmatrix}r^1~r^2~\hdots~r^m\end{bmatrix}\begin{bmatrix}W_1&&\\ &\ddots&\\&&W_m\end{bmatrix}\begin{bmatrix}s_1\\ s_2\\ \vdots\\ s_m\end{bmatrix}\Bigg)\\+~\kappa\|\|diag(W)\|\|_2^2~=~tr\bigg(\sum_{i=1}^m W_i r^i s_i\bigg)+~\kappa\|\|diag(W)\|\|_2^2~=~\sum_{i=1}^m\bigg[\sum_{j=1}^k r_j^i s_i^j W_i\bigg]\\ +~\kappa\|\|diag(W)\|\|_2^2~=~\sum_{i=1}^m\bigg[\Omega_iW_i~+~\kappa W_i^2\bigg] \end{aligned}$$ Where $\Omega_i=\sum_{j=1}^k r_j^i s_i^j$, $s_i$ is the ith row of S, and $r^i$ is the ith column of R. So the optimization problem (\[eq:opt\_w3\]) can be simplified to: $$\label{eq:opt_w4.5} \begin{aligned} \min_W~\sum_{i=1}^m\bigg[\Omega_iW_i~+~\kappa W_i^2\bigg]\\ s.t.~\sum_{i~=~1}^mW_i=1,~W_i\geq0~for~i=0,1,...,m \end{aligned}$$ Where $\Omega_i=\sum_{j=1}^k r_j^i s_i^j$. We use the coordinate descent algorithm to solve the optimization problem (\[eq:opt\_w4.5\]); Because it has the potential to provide a sparse solution, i.e. setting weights of uninformative hyperedges to zero. In each iteration of the coordinate descent algorithm, we select two consecutive hyperedges, like $W_i$ and $W_{i+1}$, to update their weights. Since $\sum_{i~=~1}^mW_i=1$, $(W_i + W_{i+1})$ will not change after each inner iteration. So we have: $$\label{eq:opt_w4.6} \begin{aligned} W_i^, W_{i+1}^~=~\operatorname{\arg\!\min}_{W_i, W_{i+1}}{\bigg[\Omega_iW_i~+~\Omega_{i+1}W_{i+1}~+~\kappa (W_i^2+W_{i+1}^2)\bigg]}\\ s.t.~W_i, W_{i+1} \geq 0,~~W_i + W_{i+1} = c \end{aligned}$$ Where $W_i^$, $W_{i+1}^$ are the optimal values of $W_i$, $W_{i+1}$ respectively. Replacing $W_{i+1}$ with $(c-W_i)$ results in: $$\label{eq:opt_w4.7} \begin{aligned} W_i^~=~\operatorname{\arg\!\min}_{W_i}{\bigg[\Omega_iW_i~+~\Omega_{i+1}(c-W_i)~+~\kappa (W_i^2+(c-W_i)^2)\bigg]}\\ s.t.~0 \leq W_i \leq c \end{aligned}$$ Which can be simplified to: $$\label{eq:opt_w4.8} \begin{aligned} W_i^~=~\operatorname{\arg\!\min}_{W_i}{\bigg[2\kappa W_i^2 + (\Omega_i-\Omega_{i+1}-2\kappa c)W_i\bigg]}\\ s.t.~0 \leq W_i \leq c \end{aligned}$$ Since the objective function in (\[eq:opt\_w4.8\]) is convex, we have the following equations: $$\label{eq:opt_w5} \begin{cases} W_i^ = 0,~W_{i+1}^* = c & if~2\kappa c \leq (\Omega_i - \Omega_{i + 1})\\ W_{i+1}^* = 0,~W_i^* = c & if~2\kappa c \leq (\Omega_{i + 1} - \Omega_i)\\ W_i^* = \frac{-\Omega_i+\Omega_{i+1}+2\kappa c}{4\kappa},~W_{i+1}^=c - W_i^ & \text{elsewhere} \end{cases}$$ Where $W_i^$, $W_{i+1}^$ are the updated values of $W_i$, $W_{i+1}$ respectively, $\Omega_i=\sum_{j=1}^k r_j^i s_i^j$, and $c = (W_i + W_{i+1})$. Using this method, the value of the objective function in (\[eq:main\_final\]) reduces after each inner iteration of updating W. As it can be seen in the equation (\[eq:opt\_w5\]), the provided solution by the coordinate descent algorithm is potentially sparse. After finding the optimal values of P and Q, matrix T is constructed using T=PQ. Each row of T corresponds to a feature and score of each feature is calculated based on $\ell2$ norm of its corresponding row in T. Finally, top-ranked features are selected. Computational cost and convergence study {#computational cost and convergence study} ======================================== Computational cost {#computational cost} ------------------ The first step of algorithm \[algorithm:main\] is hypergraph construction. To construct the hypergraph, data points should be clustered using the k-means method. its computational cost is $\mathcal{O}(t_0mn)$; Where $t_0\ll n$, $m\ll n$, and n are the number of iterations, clusters, and data points respectively. In step 2, matrices P and Q need to be updated. In Section \[local relationships between centroids\], because only centroids participated in the hypergraph, the incidence matrix H was defined $\in R^{m \times m}$. But, if needed, H can be defined $\in R^{n \times m}$ by adding zero valued rows for non-centroid data points, without any change in the hypergraph. Updating P, Q requires calculating the inverse of a $d\times d$ matrix where d is the number of features. Since in feature selection problems, usually the number of data points is much fewer than the number of features, based on [@l21RFS], by defining matrix H $\in R^{n \times m}$, the equations (\[eq:opt\_p3\]), (\[eq:opt\_q3\]) can be reformulated as follows: $$\label{eq:opt_p_f} P~=~\frac{\tau}{\rho}B^{-1}X^T[\frac{1}{\rho}(\Delta^\prime +\tau D)XB^{-1}X^T+I_n]^{-1}D^\frac{1}{2}Z_kQ^{-1}$$ $$\label{eq:opt_q_f} Q~=~\frac{\tau}{\rho} P^{-1}B^{-1}X^T[\frac{1}{\rho}(\Delta^\prime +\tau D)XBX^T + I_n]^{-1}D^\frac{1}{2}Z_k$$ By this approach the inverses of $n\times n$ matrices can be calculated instead which yields the time complexity of $\mathcal{O}(min(d,n)^3)$ for the second step. The third step consists of m inner iterations and computational cost of each iteration is k; Where k is the dimensionality of embedding. As a result, the time complexity of this step is $\mathcal{O}(mk)$. Since in equation (\[eq:local4\]), after each outer iteration, the values of H and $D_e$ does not change, there is no need to compute $\Delta $ from the ground up. So, the computational complexity of this step is $m^2\log m$. Since $m\ll n$, in almost all cases its lower than step 2. So The computational cost of each step of algorithm \[algorithm:main\] is $max\{\mathcal{O}(min(d,n)^3), \mathcal{O}(mk), \mathcal{O}(t_0mn)\}$, $m,t_0\ll n$; Which is less than several state-of-the-art algorithms such as [@JHLSR; @LapScore; @JELSR; @MCFS] in most cases. Convergence study {#convergence study} ----------------- In Section \[updating w\], it was proved that the value of the objective function decreases after each inner iteration. The proposed objective function has lower bound of zero; Moreover, we have the following proposition: Proposition 1. By using coordinate descent method, value of the objective function in (\[eq:main\_final\]) reduces after each iteration of step 2 of algorithm \[algorithm:main\]. Proof: let $$\label{eq:convergence1} P^{(i)}, Q^{(i)} =\operatorname*{\arg\!\min}_{P, Q}{\Psi(P, Q, W) + \Upsilon(P, Q) + \rho \times tr(Q^TP^TB^{(i-1)}PQ)}$$ Where $P^{(i)}, Q^{(i)}$ are the values of P, Q after the ith iteration and $B^{(i-1)}$ is the value of B after iteration (i-1). As a result: $$\label{eq:convergence2} \begin{aligned} \Psi(P^{(i)}, Q^{(i)}, W) + \Upsilon(P^{(i)}, Q^{(i)}) + \rho \times tr((Q^{(i)})^T(P^{(i)})^TB^{(i-1)}P^{(i)}Q^{(i)}) \leq\\ \Psi(P^{(i-1)}, Q^{(i-1)}, W) + \Upsilon(P^{(i-1)}, Q^{(i-1)}) + \rho \times tr((Q^{(i-1)})^T(P^{(i-1)})^TB^{(i-1)}P^{(i-1)}Q^{(i-1)}) \end{aligned}$$ Since based on the definition in (\[eq:B\]), $b_i^i~=~\frac{1}{2\|\|t_i\|\|_2}$, equation (\[eq:convergence2\]) can be rewritten as follows: $$\label{eq:convergence3} \begin{aligned} \Psi(P^{(i)}, Q^{(i)}, W) + \Upsilon(P^{(i)}, Q^{(i)}) + \rho \times \sum_{j=1}^d\frac{\|\|t_j^{(i)}\|\|_2^2}{\|\|t_j^{(i-1)}\|\|_2} \leq\\ \Psi(P^{(i-1)}, Q^{(i-1)}, W) + \Upsilon(P^{(i-1)}, Q^{(i-1)}) + \rho \times \sum_{j=1}^d\frac{\|\|t_j^{(i-1)}\|\|_2^2}{\|\|t_j^{(i-1)}\|\|_2} \end{aligned}$$ Based on a lemma provided in [@l21RFS], we have: $$\label{eq:convergence4} \begin{aligned} \sum_{j=1}^d \|\|t_j^i\|\|_2 - \sum_{j=1}^d \frac{\|\|t_j^i\|\|_2^2}{\|\|t_j^{(i-1)}\|\|_2} \leq \sum_{j=1}^d \|\|t_j^{(i-1)}\|\|_2 - \sum_{j=1}^d \frac{\|\|t_j^{(i-1)}\|\|_2^2}{\|\|t_j^{(i-1)}\|\|_2} \end{aligned}$$ According to the equations (\[eq:convergence3\]), (\[eq:convergence4\]), the following result holds: $$\label{eq:convergence5} \begin{aligned} \Psi(P^{(i)}, Q^{(i)}, W) + \Upsilon(P^{(i)}, Q^{(i)}) + \rho \times \sum_{j=1}^d \|\|t_j^{(i)}\|\|_2^2 \leq\\ \Psi(P^{(i-1)}, Q^{(i-1)}, W) + \Upsilon(P^{(i-1)}, Q^{(i-1)}) + \rho \times \sum_{j=1}^d \|\|t_j^{(i-1)}\|\|_2^2 \end{aligned}$$ Which is equivalent to: $$\label{eq:convergence_final} \begin{aligned} \Psi(P^{(i)}, Q^{(i)}, W) + \Upsilon(P^{(i)}, Q^{(i)}) + \rho \times \Theta(P^{(i)}, Q^{(i)}) \leq\\ \Psi(P^{(i-1)}, Q^{(i-1)}, W) + \Upsilon(P^{(i-1)}, Q^{(i-1)}) + \rho \times \Theta(P^{(i-1)}, Q^{(i-1)}) \end{aligned}$$ This completes the proof. Algorithm \[algorithm:main\] generally converges in two iterations; See figures \[figure:convergence\_T\], \[figure:convergence\_obj\]. This number of iterations is fewer than several other feature selection methods such as [@JHLSR; @JELSR]. Comparison with other methods {#comparison with other methods} ----------------------------- The computational cost of some state-of-the-art feature selection methods which mentioned their computational complexity, along with the number of outer iterations which they need to converge, is provided in table \[table:computational\_cost\]. As it can be seen, focusing on centroids helps our proposed method need a fewer number of iterations to converge and have lower computational complexity in most cases. [ \|P[1.5cm]{}\|P[10cm]{}\|P[4cm]{}\| ]{} method & computational cost & number of iterations\ JHLSR & $max\{\mathcal{O}(min(d,n)^3),~\mathcal{O}(mk),~\mathcal{O}(r^3+n^2)\}$, $m \gg n$ & 4\ LapScore & $\mathcal{O}(n^2d)$ & -\ JELSR & $max\{\mathcal{O}(n^2d), \mathcal{O}(n^3), \mathcal{O}(knd)\}$ & 10\ MCFS & $max\{\mathcal{O}(n^2d), \mathcal{O}(n^3), \mathcal{O}(knd)\}$ & -\ proposed & $max\{\mathcal{O}(min(d,n)^3), \mathcal{O}(mk), \mathcal{O}(t_0mn)\}$, $m,t_0\ll n$ & 2\ Parameter determination {#parameter determination} ----------------------- Our proposed objective function includes three parameters that need to be optimized manually: $\tau$, $\kappa$, and $\rho$. $\kappa$ and $\rho$ are regularization parameters which adjust sparsity of W and T. Parameter $\tau$ is used to set the importance of global structure preservation versus maintaining local neighborhood relationships among different centroids. We tuned these parameters empirically by grid search. In our experiments, we used $m = \left \lfloor{\frac{n}{10}}\right \rfloor$ for the number of centroids. Experiments =========== Evaluation methods {#evaluation methods} ------------------ In order to exhibit the effectiveness of our proposed method, we tested it using classification methods. Since our proposed method does not collaborate with any specific classifier, it can be tested using any classification method. We evaluated the proposed method using four general classification methods: K-Nearest Neighbours (KNN), Support Vector Machine with linear kernel (L-SVM), Support Vector Machine with radial basis kernel (RB-SVM), and Naïve Bayes (NB). These classifiers have been used to evaluate feature selection methods widely in recent studies [@mRMR; @RSR; @JELSR; @MCFS; @trace-ratio; @ELMS]. Data sets {#data sets} --------- Six publicly available benchmark data sets were used in our experiments: Gene-Expression [@CGAP], Smoke-Cancer [@AEGE], Various-Cancers [@RSCTC], Burkitt-Lymphoma [@RSCTC], Mouse-Type [@RSCTC], and Hepatitis-C [@RSCTC]. The number of classes in these datasets ranges from 2 to 10. These datasets also have a different number of samples: from 187 to 801. Moreover, these data sets are from different areas. The details of these data sets are provided in table \[table:data sets\]. [ \|P[4cm]{}\|P[2cm]{}\|P[2cm]{}\|P[1.5cm]{}\| ]{} Data set & Samples & Features & Classes\ Gene-Expression & 801 & 20,531 & 2\ Smoke-Cancer & 187 & 19,993 & 2\ Various-Cancers & 383 & 54,676 & 10\ Burkitt-Lymphoma & 220 & 22,284 & 3\ Mouse-Type & 214 & 45,102 & 7\ Hepatitis-C & 283 & 54,622 & 3\ Comparison methods {#comparison methods} ------------------ We compared the accuracy of our proposed method with seven state-of-the-art feature selection frameworks. These methods treat all data points equally and do not focus on more representative points, Also the importance of handling noise, redundancy and outlier is underestimated in them. Here we provide a brief explanation of each one of them: JHLSR [@JHLSR] Uses sparse representation for hypergraph construction. FScore [@PattC] This supervised framework scores the features one by one. This method tries to select the features which have the same value for data points within the same class, and different values for data points from different classes. l21RFS [@l21RFS] Tries to preserve similarity between data points and measures regression loss by $\ell_{2, 1}$ norm. This framework is supervised with class labels. TraceRatio [@trace-ratio] The main idea of this framework is similar to FScore. But it evaluates features jointly. We used the supervised form of this method in our experiments. LapScore [@LapScore] Selects features which can best preserve local structure of data. JELSR [@JELSR] Adopts a graph to model data structure. This method performs feature selection by jointly learning embedding and sparse regression. MCFS [@MCFS] First carries out manifold learning and then spectral regression. From the above methods, FScore, l21RFS, and TraceRatio are supervised with class labels and other methods are unsupervised. The differences between the unsupervised methods are provided in table \[table:differences\]. Experimental settings {#Experimental settings} --------------------- Since the optimal number of features is unknown, we performed comprehensive experiments with 10, 20, 30, ..., 200 number of features on each benchmark dataset. As in [@JHLSR; @JELSR] we randomly select 50% of data as training data and the other 50% as test data. We repeated this procedure five times and reported average and standard deviation of accuracies. Training data is given to feature selection methods as input. When the selection of features is conducted, reported features of test data are selected and along with corresponding labels are passed to a classification algorithm. Finally, classification accuracy is reported. Experiments Results {#experiments results} ------------------- The result of the exhaustive experiments conducted on various data sets prove efficiency of our proposed method. Although our proposed unsupervised framework, has lower computational complexity in most cases, it outperforms state-of-the-art supervised and unsupervised feature selection methods. Average and standard deviation of accuracies of each method on every data set are reported in tables \[table:results\_knn\], \[table:results\_lsvm\], \[table:results\_rbsvm\], \[table:results\_nb\]. The highest accuracies are boldfaced. We enhance average classification accuracy by 3.79% (JHLSR), 16.88% (FScore), 7.56% (l21RFS), 7.3% (TraceRatio), 8.38% (LapScore), 8.12% (JELSR), and 7.38% (MCFS) by KNN classification method. Using L-SVM classifier, we improve average classification accuracy by 3.47% (JHLSR), 16.75% (FScore), 8.98% (l21RFS), 8.12% (TraceRatio), 9.13% (LapScore), 8.98% (JELSR), and 8.22% (MCFS). By classifying using RB-SVM, we improve average classification accuracy by 3.93% (JHLSR), 18.86% (FScore), 8.1% (l21RFS), 8.15% (TraceRatio), 9.02% (LapScore), 8.93% (JELSR), and 8.29% (MCFS). Using NB classification method, we enhance average classification accuracy by 4.75% (JHLSR), 14.19% (FScore), 7.33% (l21RFS), 7.46% (TraceRatio), 9.37% (LapScore), 8.57% (JELSR), and 7.46% (MCFS). Accuracies of different methods for different number of features is also compared in figures \[figure:results\_knn\], \[figure:results\_lsvm\], \[figure:results\_rbsvm\], \[figure:results\_nb\]. [0.325]{} ![KNN classification accuracy vs. number of selected features on benchmark data sets.[]{data-label="figure:results_knn"}](images/GE_knn_accs.png "fig:") [0.325]{} ![KNN classification accuracy vs. number of selected features on benchmark data sets.[]{data-label="figure:results_knn"}](images/SC_knn_accs.png "fig:") [0.325]{} ![KNN classification accuracy vs. number of selected features on benchmark data sets.[]{data-label="figure:results_knn"}](images/VC_knn_accs.png "fig:") [0.325]{} ![KNN classification accuracy vs. number of selected features on benchmark data sets.[]{data-label="figure:results_knn"}](images/BL_knn_accs.png "fig:") [0.325]{} ![KNN classification accuracy vs. number of selected features on benchmark data sets.[]{data-label="figure:results_knn"}](images/MT_knn_accs.png "fig:") [0.325]{} ![KNN classification accuracy vs. number of selected features on benchmark data sets.[]{data-label="figure:results_knn"}](images/HC_knn_accs.png "fig:") [0.325]{} ![Various-Cancers](images/GE_l_svm_accs.png "fig:") [0.325]{} ![Various-Cancers](images/SC_l_svm_accs.png "fig:") [0.325]{} ![Various-Cancers](images/VC_l_svm_accs.png "fig:") [0.325]{} ![L-SVM classification accuracy vs. number of selected features on benchmark data sets.[]{data-label="figure:results_lsvm"}](images/BL_l_svm_accs.png "fig:") [0.325]{} ![L-SVM classification accuracy vs. number of selected features on benchmark data sets.[]{data-label="figure:results_lsvm"}](images/MT_l_svm_accs.png "fig:") [0.325]{} ![L-SVM classification accuracy vs. number of selected features on benchmark data sets.[]{data-label="figure:results_lsvm"}](images/HC_l_svm_accs.png "fig:") [0.325]{} ![Various-Cancers](images/GE_rb_svm_accs.png "fig:") [0.325]{} ![Various-Cancers](images/SC_rb_svm_accs.png "fig:") [0.325]{} ![Various-Cancers](images/VC_rb_svm_accs.png "fig:") [0.325]{} ![RB-SVM classification accuracy vs. number of selected features on benchmark data sets.[]{data-label="figure:results_rbsvm"}](images/BL_rb_svm_accs.png "fig:") [0.325]{} ![RB-SVM classification accuracy vs. number of selected features on benchmark data sets.[]{data-label="figure:results_rbsvm"}](images/MT_rb_svm_accs.png "fig:") [0.325]{} ![RB-SVM classification accuracy vs. number of selected features on benchmark data sets.[]{data-label="figure:results_rbsvm"}](images/HC_rb_svm_accs.png "fig:") [0.325]{} ![NB classification accuracy vs. number of selected features on benchmark data sets.[]{data-label="figure:results_nb"}](images/GE_nb_accs.png "fig:") [0.325]{} ![NB classification accuracy vs. number of selected features on benchmark data sets.[]{data-label="figure:results_nb"}](images/SC_nb_accs.png "fig:") [0.325]{} ![NB classification accuracy vs. number of selected features on benchmark data sets.[]{data-label="figure:results_nb"}](images/VC_nb_accs.png "fig:") [0.325]{} ![NB classification accuracy vs. number of selected features on benchmark data sets.[]{data-label="figure:results_nb"}](images/BL_nb_accs.png "fig:") [0.325]{} ![NB classification accuracy vs. number of selected features on benchmark data sets.[]{data-label="figure:results_nb"}](images/MT_nb_accs.png "fig:") [0.325]{} ![NB classification accuracy vs. number of selected features on benchmark data sets.[]{data-label="figure:results_nb"}](images/HC_nb_accs.png "fig:") [0.325]{} ![Various-Cancers](images/GE_T.png "fig:") [0.325]{} ![Various-Cancers](images/SC_T.png "fig:") [0.325]{} ![Various-Cancers](images/VC_T.png "fig:") [0.325]{} ![Convergence of matrix T on different data sets.[]{data-label="figure:convergence_T"}](images/BL_T.png "fig:") [0.325]{} ![Convergence of matrix T on different data sets.[]{data-label="figure:convergence_T"}](images/MT_T.png "fig:") [0.325]{} ![Convergence of matrix T on different data sets.[]{data-label="figure:convergence_T"}](images/HC_T.png "fig:") Convergence Analysis {#convergence analysis} -------------------- Since features are finally selected based on matrix T, we analyze changes of matrix T, along with the value of the objective function through iterations in order to study the convergence of algorithm \[algorithm:main\]. Let $$\label{eq:err} err(i) = \|\|T^{(i)} - T^{(i-1)}\|\|_F^2~/~(d \times k)$$ Where $T^{(i)}$ is the value of T after outer iteration number i and $(d \times k)$ is the number of elements of T. err(i) calculates the normalized change of T after an iteration. Generally, both T and the objective function converge to a global minimum after two iterations. See figures \[figure:convergence\_T\], \[figure:convergence\_obj\]. [0.325]{} ![Various-Cancers](images/GE_obj.png "fig:") [0.325]{} ![Various-Cancers](images/SC_obj.png "fig:") [0.325]{} ![Various-Cancers](images/VC_obj.png "fig:") [0.325]{} ![Convergence of the proposed objective function on different data sets.[]{data-label="figure:convergence_obj"}](images/BL_obj.png "fig:") [0.325]{} ![Convergence of the proposed objective function on different data sets.[]{data-label="figure:convergence_obj"}](images/MT_obj.png "fig:") [0.325]{} ![Convergence of the proposed objective function on different data sets.[]{data-label="figure:convergence_obj"}](images/HC_obj.png "fig:") Effect of weighting points {#effect of weighting points} -------------------------- To show the importance of weighting points of the proposed framework we conducted some experiments. We implemented the ordinary version of our point-weighting framework. In this new framework, all data points are treated equally. For this purpose, we replaced matrix D with the identity matrix. We also changed the soft hypergraph construction method. The new hypergraph was constructed by connecting each data point to its $l$ nearest neighbors. This soft hypergraph is widely used in recent studies [@HSIR; @HLasso]. This hypergraph contains n vertices and n hyperedges; While the hypergraph of our proposed method contains only m vertices and m hyperedges; Where $m \ll n$. Although our point-weighting framework has much lower computational complexity than the ordinary version, conducted experiments support its effectiveness by achieving higher classification accuracy on different benchmark datasets using different evaluation methods, in most cases. See figures \[figure:non-point-weighting-knn\], \[figure:non-point-weighting-lsvm\], \[figure:non-point-weighting-rbsvm\], and \[figure:non-point-weighting-nb\]. These experiments exhibit the importance of behaving different data points based on their representation power and role in defining the data structure. [0.325]{} ![KNN classification accuracies of different methods on benchmark data sets.[]{data-label="figure:non-point-weighting-knn"}](images/GE_knn_non_point_weighting.png "fig:") [0.325]{} ![KNN classification accuracies of different methods on benchmark data sets.[]{data-label="figure:non-point-weighting-knn"}](images/SC_knn_non_point_weighting.png "fig:") [0.325]{} ![KNN classification accuracies of different methods on benchmark data sets.[]{data-label="figure:non-point-weighting-knn"}](images/MT_knn_non_point_weighting.png "fig:") [0.325]{} ![L-SVM classification accuracies of different methods on benchmark data sets.[]{data-label="figure:non-point-weighting-lsvm"}](images/GE_l_svm_non_point_weighting.png "fig:") [0.325]{} ![L-SVM classification accuracies of different methods on benchmark data sets.[]{data-label="figure:non-point-weighting-lsvm"}](images/SC_l_svm_non_point_weighting.png "fig:") [0.325]{} ![L-SVM classification accuracies of different methods on benchmark data sets.[]{data-label="figure:non-point-weighting-lsvm"}](images/MT_l_svm_non_point_weighting.png "fig:") [0.325]{} ![RB-SVM classification accuracies of different methods on benchmark data sets.[]{data-label="figure:non-point-weighting-rbsvm"}](images/GE_rb_svm_non_point_weighting.png "fig:") [0.325]{} ![RB-SVM classification accuracies of different methods on benchmark data sets.[]{data-label="figure:non-point-weighting-rbsvm"}](images/SC_rb_svm_non_point_weighting.png "fig:") [0.325]{} ![RB-SVM classification accuracies of different methods on benchmark data sets.[]{data-label="figure:non-point-weighting-rbsvm"}](images/MT_rb_svm_non_point_weighting.png "fig:") [0.325]{} ![NB classification accuracies of different methods on benchmark data sets.[]{data-label="figure:non-point-weighting-nb"}](images/GE_nb_non_point_weighting.png "fig:") [0.325]{} ![NB classification accuracies of different methods on benchmark data sets.[]{data-label="figure:non-point-weighting-nb"}](images/SC_nb_non_point_weighting.png "fig:") [0.325]{} ![NB classification accuracies of different methods on benchmark data sets.[]{data-label="figure:non-point-weighting-nb"}](images/MT_nb_non_point_weighting.png "fig:") Effect of soft hypergraph {#effect of soft hypergraph} ------------------------- In soft hypergraphs, different vertices can have different participation in a hyperedge. This participation can be based on their importance and role in that hyperedge. This helps the framework to model data structure more precisely. We carried out some experiments to show the superiority of soft hypergraphs to ordinary hypergraphs. To do so, we converted the soft hypergraph of our framework to an ordinary hypergraph and tested it on some benchmark data sets using diverse evaluation methods. As it can be seen in figures \[figure:ordinary-hypergraph-knn\], \[figure:ordinary-hypergraph-lsvm\], \[figure:ordinary-hypergraph-rbsvm\], and \[figure:ordinary-hypergraph-nb\], using soft hypergraph results more classification accuracy in most cases. [0.325]{} ![KNN classification accuracies of different methods on benchmark data sets.[]{data-label="figure:ordinary-hypergraph-knn"}](images/GE_knn_ordinary_hypergraph.png "fig:") [0.325]{} ![KNN classification accuracies of different methods on benchmark data sets.[]{data-label="figure:ordinary-hypergraph-knn"}](images/SC_knn_ordinary_hypergraph.png "fig:") [0.325]{} ![KNN classification accuracies of different methods on benchmark data sets.[]{data-label="figure:ordinary-hypergraph-knn"}](images/MT_knn_ordinary_hypergraph.png "fig:") [0.325]{} ![L-SVM classification accuracies of different methods on benchmark data sets.[]{data-label="figure:ordinary-hypergraph-lsvm"}](images/GE_l_svm_ordinary_hypergraph.png "fig:") [0.325]{} ![L-SVM classification accuracies of different methods on benchmark data sets.[]{data-label="figure:ordinary-hypergraph-lsvm"}](images/SC_l_svm_ordinary_hypergraph.png "fig:") [0.325]{} ![L-SVM classification accuracies of different methods on benchmark data sets.[]{data-label="figure:ordinary-hypergraph-lsvm"}](images/MT_l_svm_ordinary_hypergraph.png "fig:") [0.325]{} ![RB-SVM classification accuracies of different methods on benchmark data sets.[]{data-label="figure:ordinary-hypergraph-rbsvm"}](images/GE_rb_svm_ordinary_hypergraph.png "fig:") [0.325]{} ![RB-SVM classification accuracies of different methods on benchmark data sets.[]{data-label="figure:ordinary-hypergraph-rbsvm"}](images/SC_rb_svm_ordinary_hypergraph.png "fig:") [0.325]{} ![RB-SVM classification accuracies of different methods on benchmark data sets.[]{data-label="figure:ordinary-hypergraph-rbsvm"}](images/MT_rb_svm_ordinary_hypergraph.png "fig:") [0.325]{} ![NB classification accuracies of different methods on benchmark data sets.[]{data-label="figure:ordinary-hypergraph-nb"}](images/GE_nb_ordinary_hypergraph.png "fig:") [0.325]{} ![NB classification accuracies of different methods on benchmark data sets.[]{data-label="figure:ordinary-hypergraph-nb"}](images/SC_nb_ordinary_hypergraph.png "fig:") [0.325]{} ![NB classification accuracies of different methods on benchmark data sets.[]{data-label="figure:ordinary-hypergraph-nb"}](images/MT_nb_ordinary_hypergraph.png "fig:") [0.325]{} ![KNN classification accuracies of different methods on benchmark data sets.[]{data-label="figure:global-str-knn"}](images/GE_knn_global_str.png "fig:") [0.325]{} ![KNN classification accuracies of different methods on benchmark data sets.[]{data-label="figure:global-str-knn"}](images/SC_knn_global_str.png "fig:") [0.325]{} ![KNN classification accuracies of different methods on benchmark data sets.[]{data-label="figure:global-str-knn"}](images/MT_knn_global_str.png "fig:") Effect of global structure preservation {#Effect of global similarity preservation} --------------------------------------- In the proposed framework, we preserve the global structure of data by maintaining the correlation between data points. Since in unsupervised frameworks data labels are unavailable and data points from the same class generally have high linear dependency and correlation [@RFRSR], this approach performs as an alternative to employment of data labels in supervised learning for structure preservation; Moreover, it helps to repair the inevitable loss of information caused by our low computational hypergraph which is constructed only based on cluster centroids. In order to test the importance of this term, we performed some experiments on different datasets using various evaluation methods. Although adding global structure preservation term has minor effect in computational cost of the proposed algorithm, as it can be seen in figures \[figure:global-str-knn\], \[figure:global-str-lsvm\], \[figure:global-str-rbsvm\], and \[figure:global-str-nb\], it leads to much more effective selection of features in most cases. [0.325]{} ![L-SVM classification accuracies of different methods on benchmark data sets.[]{data-label="figure:global-str-lsvm"}](images/GE_l_svm_global_str.png "fig:") [0.325]{} ![L-SVM classification accuracies of different methods on benchmark data sets.[]{data-label="figure:global-str-lsvm"}](images/SC_l_svm_global_str.png "fig:") [0.325]{} ![L-SVM classification accuracies of different methods on benchmark data sets.[]{data-label="figure:global-str-lsvm"}](images/MT_l_svm_global_str.png "fig:") [0.325]{} ![RB-SVM classification accuracies of different methods on benchmark data sets.[]{data-label="figure:global-str-rbsvm"}](images/GE_rb_svm_global_str.png "fig:") [0.325]{} ![RB-SVM classification accuracies of different methods on benchmark data sets.[]{data-label="figure:global-str-rbsvm"}](images/SC_rb_svm_global_str.png "fig:") [0.325]{} ![RB-SVM classification accuracies of different methods on benchmark data sets.[]{data-label="figure:global-str-rbsvm"}](images/MT_rb_svm_global_str.png "fig:") [0.325]{} ![NB classification accuracies of different methods on benchmark data sets.[]{data-label="figure:global-str-nb"}](images/GE_nb_global_str.png "fig:") [0.325]{} ![NB classification accuracies of different methods on benchmark data sets.[]{data-label="figure:global-str-nb"}](images/SC_nb_global_str.png "fig:") [0.325]{} ![NB classification accuracies of different methods on benchmark data sets.[]{data-label="figure:global-str-nb"}](images/MT_nb_global_str.png "fig:") Conclusion ========== In this paper, a novel point-weighting framework for hypergraph feature selection was proposed which adopts low-rank representation to handle noise and outlier, and captures the importance of different data points. Focusing on centroids helps the proposed method to have lower computational complexity with respect to many state-of-the-art feature selection methods, in most cases. We also behave different data points based on their representation power and role in defining the data structure. We conducted experiments to show the effectiveness of the innovative ideas of this paper. Exhaustive experiments exhibit the effectiveness of our proposed method in comparison with other state-of-the-art feature selection methods. [^1]: [email protected] [^2]: [email protected]	{ "pile_set_name": "ArXiv" }
--- abstract: \| For an efficient implementation of Buchberger’s Algorithm, it is essential to avoid the treatment of as many unnecessary critical pairs or obstructions as possible. In the case of the commutative polynomial ring, this is achieved by the Gebauer-Möller criteria. Here we present an adaptation of the Gebauer-Möller criteria for non-commutative polynomial rings, i.e. for free associative algebras over fields. The essential idea is to detect unnecessary obstructions using other obstructions with or without overlap. Experiments show that the new criteria are able to detect almost all unnecessary obstructions during the execution of Buchberger’s procedure. [Keywords:]{} Gröbner basis, free associative algebra, obstruction, Buchberger procedure [AMS classification:]{} 16-08, 20-04, 13P10 author: - \| Martin Kreuzer[^1], Xingqiang Xiu[^2]\ Fakultät für Informatik und Mathematik\ Universität Passau, D-94030 Passau, Germany title: 'Non-Commutative Gebauer-Möller Criteria' --- Introduction {#sec1} ============ Ever since B. Buchberger’s thesis [@Bu65], Gröbner bases have become a fundamental tool for computations in commutative algebra and algebraic geometry. The most time-consuming part in Buchberger’s Algorithm is the computation of the normal remainder of an S-polynomial corresponding to a critical pair. Therefore a significant amount of energy has been spent on reducing the number of critical pairs which have to be treated. After the discovery of various criteria for discarding critical pairs ahead of time by B. Buchberger and H.M. Möller (see [@Bu79], [@Bu85] and [@Mo85]), this subject found an initial resolution via the Gebauer-Möller installation presented in [@GM88] which offers a good compromise between efficiency and the success rate for detecting unnecessary critical pairs. A very different picture presents itself for Gröbner basis computations for two-sided ideals in non-commutative polynomial rings. The basic Gröbner basis theory in this case was described by G.H. Bergman (see [@Be78]), T. Mora (see [@Mo86] and [@Mo94]) and others, and obstructions, the non-commutative analogue of critical pairs, were studied in [@Mo94]. However, since only a few authors endeavoured to implement efficient versions of Buchberger’s Procedure for the non-commutative polynomial ring (i.e. the free associative algebra), the subject of minimizing the number of obstructions which have to be treated has received comparatively little attention, and merely a few rules were developed. For instance, the package [Plural]{} of the computer algebra system [Singular]{} implements a version of the product and the chain criterion, but not the multiply criterion or the leading word criterion. On the other hand, the system [Magma]{} appears to be based on a variant of the F4 Algorithm which does not use criteria for unnecessary obstructions. For an overview on rules which have been developed see for instance [@Co07]. In this paper, we present generalizations of the Gebauer-Möller criteria for non-commutative polynomials. They cover not only the known cases of useless obstructions discussed in [@Mo94], Lemma 5.11 and [@Co07], but form a complete analogue of the results in the commutative case. One of the key ingredients we use for this purpose is the consideration of obstructions without overlaps. We detect useless obstructions, i.e. obstructions that can be represented by other obstructions, using not only obstructions with overlaps but using also those without overlaps. We show that the consideration of obstructions without overlaps does not increase unnecessary computations, since a Gröbner representation is inherent in the S-polynomial of every obstruction without overlaps. Consequently, we reduce the number of obstructions efficiently and obtain a non-commutative version of the Gebauer-Möller criteria. This paper is organised as follows. In Section \[sec2\] we recall the basic theory of Gröbner bases for two-sided ideals in non-commutative polynomial rings. In particular, we introduce and study obstructions (see Definitions \[sec2def4\] and \[sec2def9\], and Lemmas \[sec2lem6\] and \[sec2lem8\]), present the Buchberger Criterion (see Proposition \[sec2pro10\]), and formulate the Buchberger Procedure (see Theorem \[sec2the11\]). The non-commutative analogues of the Gebauer-Möller criteria are developed in Section \[sec3\]. They are based on a careful study of the set of newly constructed obstructions which are produced during the execution of Buchberger’s Procedure. As a result, we are able to formulate the Non-Commutative Multiply Criterion (see Proposition \[ncMCrit\]), the Non-Commutative Leading Word Criterion (see Proposition \[ncLWCrit\]) and the Non-Commutative Backward Criterion (see Proposition \[ncBKCrit\]). When we combine these criteria, the result is a new Improved Buchberger Procedure \[sec3the14\]. The second author has implemented a version of the Buchberger Procedure for non-commutative polynomial rings in a package for the computer algebra system [ApCoCoA]{} which includes the non-commutative Gebauer-Möller criteria developed here (see [@Ap10]). In the last section, we present experimental results about the efficiency of the criteria for some cases of moderately difficult Gröbner basis computations. Unless mentioned otherwise, we adhere to the definitions and terminology given in [@KR00] and [@KR05]. Gröbner Bases in $K\langle X\rangle$ {#sec2} ==================================== In the following we let $X=\{x_1,\dots,x_n\}$ be a finite set of indeterminates (or a finite alphabet), and $\langle X\rangle$ the monoid of all words (or terms) $x_{i_1}\cdots x_{i_l}$ where the multiplication is concatenation of words. The empty word will be denoted by $\lambda$. Furthermore, let $K$ be a field, and let $$K\langle X\rangle=\{c_1w_1+\cdots+c_sw_s\ \|\ s\in\mathbb{N},c_i\in K\setminus\{0\},w_i\in\langle X\rangle\}$$ be the non-commutative polynomial ring generated by $X$ over $K$ (or the free associative $K$-algebra generated by $X$). We introduce basic notions of Gröbner basis theory in this setting. \[sec2def1\] A word ordering on $\langle X\rangle$ is a well-ordering ${\sigma}$ which is compatible with multiplication, i.e. $w_1\geq_{\sigma} w_2$ implies $w_3w_1w_4\geq_{\sigma} w_3w_2w_4$ for all words $w_1,w_2,w_3,w_4\in\langle X\rangle$. In the commutative case, a word ordering is usually called a term ordering or monomial ordering. For instance, the length-lexicographic ordering $\LLex$ is a word ordering. It first compares the length of two words and then breaks ties using the non-commutative lexicographic ordering with respect to $x_1>_{\LLex}\cdots>_{\LLex}x_n$. Note that the non-commutative lexicographic ordering by itself is not a word ordering, since it is neither a well-ordering nor compatible with multiplication. \[sec2def2\] Let ${\sigma}$ be a word ordering on $\langle X\rangle$. - Given a polynomial $f\in K\langle X\rangle\setminus\{0\}$, there exists a unique representation $f=c_1w_1+\cdots+c_sw_s$ with $c_1,\dots,c_s\in K\setminus\{0\}$ and $w_1,\dots,w_s\in\langle X\rangle$ such that $w_1>_{\sigma}\cdots>_{\sigma} w_s$. The word $\lw_{\sigma}(f)=w_1$ is called the leading word of $f$ with respect to ${\sigma}$. The element $\lc_{\sigma}(f)=c_1$ is called the leading coefficient. We let $\lm_{\sigma}(f)=c_1w_1$ and call it the leading monomial of $f$. - Let $I\subseteq K\langle X\rangle$ be a two-sided ideal. The set $\lw_{\sigma}\{I\}=\{\lw_{\sigma}(f)\ \|\ f\in I\setminus\{0\}\}\subseteq\langle X\rangle$ is called the leading word set of $I$. The two-sided ideal $\lw_{\sigma}(I)=\langle\lw_{\sigma}(f)\ \|\ f\in I\setminus\{0\}\rangle\subseteq K\langle X\rangle$ is called the leading word ideal of $I$. - A subset $G$ of a two-sided ideal $I\subseteq K\langle X\rangle$ is called a ${\sigma}$-Gröbner basis of $I$ if the set of the leading words $\lw_{\sigma}\{G\}=\{\lw_{\sigma}(f)\ \|\ f\in G\setminus\{0\}\}$ generates the leading word ideal $\lw_{\sigma}(I)$. In the following we focus on computations of Gröbner bases for two-sided ideals in $K\langle X\rangle$. For readers who want to know further properties and applications of non-commutative Gröbner bases, we refer to [@Mo94] and [@Xiu12]. Throughout this paper we assume that ${\sigma}$ is a word ordering on $\langle X\rangle$. The next algorithm is a central part of all Gröbner basis computations. \[sec2the3\] Let $f\in K\langle X\rangle$, $s\geq 1$, and $G=\{g_{1},\dots,g_{s}\}\subseteq K\langle X\rangle\setminus\{0\}$. Consider the following sequence of instructions. - Let $k_{1}=\cdots=k_{s}=0, p=0$, and $v=f$. - Find the smallest index $i\in\{1,\dots,s\}$ such that $\lw_{\sigma}(v)=w\lw_{\sigma}(g_{i})w'$ for some words $w,w'\in\langle X\rangle$. If such an $i$ exists, increase $k_{i}$ by $1$, set $c_{ik_{i}}=\frac{\lc_{\sigma}(v)}{\lc_{\sigma}(g_{i})}, w_{ik_{i}}=w, w'_{ik_{i}}=w'$, and replace $v$ by $v-c_{ik_{i}}w_{ik_{i}}g_{i}w'_{ik_{i}}$. - Repeat step [(D2)]{} until there is no more $i\in\{1,\dots,s\}$ such that $\lw_{\sigma}(v)$ is a multiple of $\lw_{\sigma}(g_{i})$. If now $v\neq 0$, then replace $p$ by $p+\lm_{\sigma}(v)$ and $v$ by $v-\lm_{\sigma}(v)$, continue with step [(D2)]{}. - Return the tuples $(c_{11},w_{11},w'_{11}),\dots,(c_{sk_{s}},w_{sk_{s}},w'_{sk_{s}})$ and $p$. This is an algorithm which returns tuples $(c_{11},w_{11},w'_{11}),\dots,(c_{sk_{s}},w_{sk_{s}},w'_{sk_{s}})$ and a polynomial $p\in K\langle X\rangle$ such that the following conditions are satisfied. - We have $f=\sum^{s}_{i=1}\sum^{k_{i}}_{j=1}c_{ij}w_{ij}g_{i}w'_{ij}+p$. - No element of ${\rm Supp}(p)$ is contained in $\langle\lw_{\sigma}(g_{1}),\dots,\lw_{\sigma}(g_{s})\rangle$. - For all $i\in\{1,\dots,s\}$ and all $j\in\{1,\dots,k_{i}\}$, we have $\lw_{\sigma}(w_{ij}g_{i}w'_{ij})\leq_{\sigma}\lw_{\sigma}(f)$. If $p\neq 0$, we have $\lw_{\sigma}(p)\leq_{\sigma}\lw_{\sigma}(f)$. - For all $i\in\{1,\dots, s\}$ and all $j\in\{1,\dots,k_{i}\}$, we have $\lw_{\sigma}(w_{ij}g_{i}w'_{ij})\notin\langle\lw_{\sigma}(g_{1}),\dots,\lw_{\sigma}(g_{i-1})\rangle$. Note that the resulting tuples $(c_{11},w_{11},w'_{11}),\dots,(c_{sk_{s}},w_{sk_{s}},w'_{sk_{s}})$ and polynomial $p$ satisfying conditions (a)-(d) are not unique. This is due to the fact that in step (D2) of the Division Algorithm there might exist more that one pair $(w,w')$ satisfying $\lw_{\sigma}(v)=w\lw_{\sigma}(g_{i})w'$ (see [@Xiu12], Example 3.2.2). A polynomial $p\in K\langle X\rangle$ obtained in Theorem \[sec2the3\] is called a normal remainder of $f$ with respect to $G$ and is denoted by ${\rm NR}_{\sigma,G}(f)$. For $s\geq 1$, we let $F_s=(K\langle X\rangle\otimes_{K}K\langle X\rangle)^s$ be the free two-sided $K\langle X\rangle$-module of rank $s$ with the canonical basis $\{e_{1},\dots,e_{s}\}$, where $e_{i}=(0,\dots,0,1\otimes1,$ $0,\dots,0)$ with $1\otimes1$ occurring in the $i^{\rm th}$ position for $i=1,\dots,s$, and we let $\mathbb{T}(F_{s})$ be the set of terms in $F_{s}$, i.e. $\mathbb{T}(F_{s})=\{we_{i}w'\ \|\ i\in\{1,\dots,s\}, w,w'\in\langle X\rangle\}$. \[sec2def4\] Let $G=\{g_1,\dots,g_s\}\subseteq K\langle X\rangle\setminus\{0\}$ with $s\geq 1$, and let $i,j\in\{1,\dots,s\}$ such that $i\leq j$. - If there exist some words $w_{i},w'_{i},w_{j},w'_{j}\in\langle X\rangle$ such that $w_{i}\lw_{\sigma}(g_{i})w'_{i}=w_{j}\lw_{\sigma}(g_{j})w'_{j}$, then we call the element $${\rm o}_{i,j}(w_{i},w'_{i};w_{j},w'_{j})=\frac{1}{\lc_{\sigma}(g_{i})}w_{i}e_{i}w'_{i}-\frac{1}{\lc_{\sigma}(g_{j})}w_{j}e_{j}w'_{j}\in F_{s}\setminus\{0\}$$ an obstruction of $g_i$ and $g_j$. If $i=j$, it is called a self obstruction of $g_{i}$. We will denote the set of all obstructions of $g_{i}$ and $g_{j}$ by ${\rm Obs}(i,j)$. - Let ${\rm o}_{i,j}(w_{i},w'_{i};w_{j},w'_{j})\in {\rm Obs}(i,j)$ be an obstruction of $g_{i}$ and $g_{j}$. The polynomial $$S_{i,j}(w_{i},w'_{i};w_{j},w'_{j})=\frac{1}{\lc_{\sigma}(g_{i})}w_{i}g_{i}w'_{i}-\frac{1}{\lc_{\sigma}(g_{j})}w_{j}g_{j}w'_{j}\in K\langle X\rangle$$ is called the S-polynomial of ${\rm o}_{i,j}(w_{i},w'_{i};w_{j},w'_{j})$. Using these definitions, we can characterize Gröbner bases in the following way. \[sec2pro5\] Let $G=\{g_1,\dots,g_s\}\subseteq K\langle X\rangle\setminus\{0\}$ be a set of polynomials which generate a two-sided ideal $I=\langle G\rangle\subseteq K\langle X\rangle$. Then the following conditions are equivalent. - The set $G$ is a ${\sigma}$-Gröbner basis of $I$. - For every obstruction ${\rm o}_{i,j}(w_{i},w'_{i};w_{j},w'_{j})$ in the set $\bigcup_{1\leq i\leq j\leq s}{\rm Obs}(i,j)$, its S-polynomial $S_{i,j}(w_{i},w'_{i};w_{j},w'_{j})$ has a representation $$S_{i,j}(w_{i},w'_{i};w_{j},w'_{j})=\sum^{\mu}_{k=1}c_{k}w_{k}g_{i_{k}}w'_{k}$$ with $c_{k}\in K, w_{k},w'_{k}\in\langle X\rangle$, and $g_{i_{k}}\in G$ for all $k\in\{1,\dots,\mu\}$ such that $\lw_{\sigma}(w_{j}g_{j}w'_{j})>_{\sigma}\lw_{\sigma}(w_{k}g_{i_{k}}w'_{k})$ if $c_k\neq 0$ for some $k\in\{1,\dots,\mu\}$. See [@Xiu12], Proposition 4.1.2. A presentation of $S_{i,j}(w_{i},w'_{i};w_{j},w'_{j})$ as in Proposition \[sec2pro5\].b is called a (weak) Gröbner representation of $S_{i,j}(w_{i},w'_{i};w_{j},w'_{j})$ in terms of $G$. Observe that there are infinitely many obstructions in each set ${\rm Obs}(i,j)$, due to the following two types of trivial obstructions. - If ${\rm o}_{i,j}(w_{i},w'_{i};w_{j},w'_{j})\in {\rm Obs}(i,j)$, then, for all $w,w'\in\langle X\rangle$, we have ${\rm o}_{i,j}(ww_{i},w'_{i}w';ww_{j},w'_{j}w')\in{\rm Obs}(i,j)$. - For all $w\in\langle X\rangle$, we have ${\rm o}_{i,j}(\lw_{\sigma}(g_{j})w,1;1,w\lw_{\sigma}(g_{i})), {\rm o}_{i,j}(1, w\lw_{\sigma}(g_{j});$ $\lw_{\sigma}(g_{i})w,1)\in{\rm Obs}(i,j)$. Before going on, let us get rid of these two types of trivial obstructions. The following lemma handles trivial obstructions of type (T1). \[sec2lem6\] If the S-polynomial of ${\rm o}_{i,j}(w_{i},w'_{i};w_{j},w'_{j})\in {\rm Obs}(i,j)$ has a Gröbner representation in terms of $G$, then, for all $w,w'\in\langle X\rangle$, the S-polynomial of ${\rm o}_{i,j}(ww_{i},w'_{i}w';ww_{j},w'_{j}w')$ also has a Gröbner representation in terms of $G$. Without loss of generality, we assume that $S_{i,j}(w_{i},w'_{i};w_{j},w'_{j})$ is non-zero. We write $S_{i,j}(w_{i},w'_{i};w_{j},w'_{j})=\sum^{\mu}_{k=1}c_{k}w_{k}g_{i_{k}}w'_{k}$, where $c_{k}\in K\setminus\{0\}$, $w_{k},w'_{k}\in\langle X\rangle$, and $g_{i_{k}}\in G$ such that $\lw_{\sigma}(w_{j}g_{j}w'_{j})>_{\sigma}\lw_{\sigma}(w_{k}g_{i_{k}}w'_{k})$ for all $k\in\{1,\dots,\mu\}$. For all $w,w'\in\langle X\rangle$, it is clear that $S_{i,j}(ww_{i},w'_{i}w';ww_{j},w'_{j}w')$ $=\sum^{\mu}_{k=1}c_{k}ww_{k}g_{i_{k}}w'_{k}w'$. Since the word ordering $\sigma$ is compatible with multiplication, we have $w\lw_{\sigma}(w_{j}g_{j}w'_{j})w'>_{\sigma}w\lw_{\sigma}(w_{k}g_{i_{k}}w'_{k})w'$ for all $k\in\{1,\dots,\mu\}$. Hence we have $\lw_{\sigma}(ww_{j}g_{j}w'_{j}w')>_{\sigma}\lw_{\sigma}(ww_{k}g_{i_{k}}w'_{k}w')$ for all $k\in\{1,\dots,\mu\}$ and $S_{i,j}(ww_{i},w'_{i}w';ww_{j},w'_{j}w')=\sum^{\mu}_{k=1}c_{k}ww_{k}g_{i_{k}}w'_{k}w'$ is a Gröbner representation in terms of $G$. To deal with trivial obstructions of type (T2), we introduce some terminology as follows. \[sec2def7\] Let $G=\{g_1,\dots,g_s\}\subseteq K\langle X\rangle\setminus\{0\}$ with $s\geq 1$. - Let $w_{1},w_{2}\in\langle X\rangle$ be two words. If there exist some words $w,w',w''\in\langle X\rangle$ and $w\neq 1$ such that $w_{1}=w'w$ and $w_{2}=ww''$, or $w_{1}=ww'$ and $w_{2}=w''w$, or $w_{1}=w$ and $w_{2}=w'ww''$, or $w_{1}=w'ww''$ and $w_{2}=w$, then we say $w_{1}$ and $w_{2}$ have an overlap at $w$. Otherwise, we say that $w_{1}$ and $w_{2}$ have no overlap. - Let ${\rm o}_{i,j}(w_{i},w'_{i};w_{j},w'_{j})\in {\rm Obs}(i,j)$ be an obstruction. If $\lw_{\sigma}(g_{i})$ and $\lw_{\sigma}(g_{j})$ have an overlap at $w\in\langle X\rangle\setminus\{1\}$ and if $w$ is a subword of $w_{i}\lw_{\sigma}(g_{i})w'_{i}$, then we say that ${\rm o}_{i,j}(w_{i},w'_{i};w_{j},w'_{j})$ has an overlap at $w$. Otherwise, we say that ${\rm o}_{i,j}(w_{i},w'_{i};w_{j},w'_{j})$ has no overlap. Thus, as shown in (T2), there are infinitely many obstructions without overlaps in each ${\rm Obs}(i,j)$. The following lemma gets rid of these trivial obstructions. \[sec2lem8\] If ${\rm o}_{i,j}(w_{i},w'_{i};w_{j},w'_{j})\in {\rm Obs}(i,j)$ has no overlap, then $S_{i,j}(w_{i},w'_{i};$ $w_{j},w'_{j})$ has a Gröbner representation in terms of $G$. See [@Mo94], Lemma 5.4. Observe that Lemma \[sec2lem8\] is indeed a non-commutative version of the product criterion (or criterion 2) of Buchberger (cf. [@Bu85]). \[sec2def9\] Let $G=\{g_1,\dots,g_s\}\subseteq K\langle X\rangle\setminus\{0\}$ with $s\geq 1$. - Let $i,j\in\{1,\dots,s\}$ and $i<j$. An obstruction in ${\rm Obs}(i,j)$ is called non-trivial if it has an overlap and is of the form ${\rm o}_{i,j}(w_{i},1;1,w'_{j})$, or ${\rm o}_{i,j}(1,w'_{i};$ $w_{j},1)$, or ${\rm o}_{i,j}(w_{i},w'_{i};1,1)$, or ${\rm o}_{i,j}(1,1;w_{j},w'_{j})$ with $w_{i},w'_{i},w_{j},w'_{j}\in\langle X\rangle$. - Let $i\in\{1,\dots,s\}$. A self obstruction in ${\rm Obs}(i,i)$ is called non-trivial if it has an overlap and is of the form ${\rm o}_{i,i}(1,w'_{i};w_{i},1)$ with $w_{i},w'_{i}\in\langle X\rangle\setminus\{1\}$. - Let $i,j\in\{1,\dots,s\}$ and $i\leq j$. The set of all non-trivial obstructions of $g_{i}$ and $g_{j}$ will be denoted by ${\rm NTObs}(i,j)$. In the literature, a non-trivial obstruction of the form ${\rm o}_{i,j}(w_{i},1;1,w'_{j})$ is called a left obstruction, a non-trivial obstruction of the form ${\rm o}_{i,j}(1,w'_{i};w_{j},1)$ is called a right obstruction, and a non-trivial obstruction of the form ${\rm o}_{i,j}(w_{i},w'_{i};$ $1,1)$ or ${\rm o}_{i,j}(1,1;w_{j},w'_{j})$ is called a center obstruction. We picture four types of obstructions as follows. [\[3]{}[C]{}]{} &\ &\ &&\ & &\ \ && [\[3]{}[C]{}]{} &\ &\ &&\ \ & &\ && At this point we can refine the characterization of Gröbner bases given in Proposition \[sec2pro5\] in the following way. \[sec2pro10\] Let $G=\{g_1,\dots,g_s\}\subseteq K\langle X\rangle$ be a set of non-zero polynomials which generate a two-sided ideal $I=\langle G\rangle\subseteq K\langle X\rangle$. Then the set $G$ is a ${\sigma}$-Gröbner basis of $I$ if and only if, for each non-trivial obstruction ${\rm o}_{i,j}(w_{i},w'_{i};w_{j},w'_{j})\in\bigcup_{1\leq i\leq j\leq s}{\rm NTObs}(i,j)$, its S-polynomial $S_{i,j}(w_{i},w'_{i};w_{j},w'_{j})$ has a Gröbner representation in terms of $G$. This follows directly from Proposition \[sec2pro5\] and Lemmas \[sec2lem6\] and \[sec2lem8\]. In view of Lemma \[sec2lem8\], it suffices to consider each obstruction with overlap, which is either a non-trivial obstruction or a multiple of a non-trivial obstruction. Further, Lemma \[sec2lem6\] treats a multiple of a non-trivial obstruction via the corresponding non-trivial obstruction. Therefore, it is sufficient to consider only non-trivial obstructions. The Buchberger Criterion enables us to formulate the following procedure for computing Gröbner bases of two-sided ideals. Note that, in the procedure, by a fair strategy we mean a selection strategy which ensures that every obstruction is selected eventually. Since these Gröbner bases need not be finite, we have to content ourselves with an enumerating procedure. \[sec2the11\] Let $s\geq 1$, and let $G=\{g_1,\dots,g_s\}\subseteq K\langle X\rangle$ be a set of non-zero polynomials which generate a two-sided ideal $I=\langle G\rangle\subseteq K\langle X\rangle$. Consider the following sequence of instructions. - Let $B=\bigcup_{1\leq i\leq j\leq s}{\rm NTObs}(i,j)$. - If $B=\emptyset$, return the result ${G}$. Otherwise, select an obstruction ${\rm o}_{i,j}(w_{i},w'_{i};$ $w_{j},w'_{j})\in B$ using a fair strategy and delete it from $B$. - Compute the S-polynomial $S=S_{i,j}(w_{i},w'_{i};w_{j},w'_{j})$ and its normal remainder $S'={\rm NR}_{\sigma,G}(S)$. If $S'=0$, continue with step [(B2)]{}. - Increase $s$ by one, append $g_{s}=S'$ to the set ${G}$, and append the set of obstructions $\bigcup_{1\leq i\leq s}{\rm NTObs}(i,s)$ to the set $B$. Then continue with step [(B2)]{}. This is a procedure that enumerates a ${\sigma}$-Gröbner basis ${G}$ of $I$. If $I$ has a finite ${\sigma}$-Gröbner basis, the procedure stops after finitely many steps and the resulting set ${G}$ is a finite ${\sigma}$-Gröbner basis of $I$. Note that this is a straightforward generalization of the commutative version of Buchberger’s algorithm to the non-commutative case. We refer to [@Mo94] for the original form of this procedure and to [@Xiu12], Theorem 4.1.14 for a detailed proof. Non-Commutative Gebauer-Möller Criteria {#sec3} ======================================= In this section we present non-commutative Gebauer-Möller criteria. They check whether an obstruction can be represented by “smaller” obstructions. If so, we declare such obstructions to be unnecessary. Before going into details, we define a certain well-ordering $\tau$ on $\mathbb{T}(F_{s})=\{we_{i}w'\ \|\ i\in\{1,\dots,s\}, w,w'\in\langle X\rangle\}$ and use it to order obstructions. In the following, let $s\geq1$, and let $G=\{g_1,\dots,g_{s}\}\subseteq K\langle X\rangle\setminus\{0\}$ be a set of non-commutative polynomials. \[sec3def1\] Let us define a relation $\tau$ on $\mathbb{T}(F_{s})$ as follows. For two terms $w_{1}e_{i}w'_{1},w_{2}e_{j}w'_{2}\in\mathbb{T}(F_{s})$, we let $w_{1}e_{i}w'_{1}\geq_{\tau}w_{2}e_{j}w'_{2}$ if - $w_{1}\lw_{\sigma}(g_{i})w'_{1}>_{\sigma}w_{2}\lw_{\sigma}(g_{j})w'_{2}$, or - $w_{1}\lw_{\sigma}(g_{i})w'_{1}=w_{2}\lw_{\sigma}(g_{j})w'_{2}$ and $i>j$, or - $w_{1}\lw_{\sigma}(g_{i})w'_{1}=w_{2}\lw_{\sigma}(g_{j})w'_{2}$ and $i=j$ and $w_{1}\geq_{\sigma}w_{2}$. One can check that $\tau$ is a well-ordering and is compatible with scalar multiplication. The relation $\tau$ is called the module term ordering induced by $({\sigma},{G})$ on $\mathbb{T}(F_{s})$. By definition, for every obstruction ${\rm o}_{i,j}(w_{i},w'_{i};w_{j},w'_{j})\in\bigcup_{1\leq i\leq j\leq s}{\rm Obs}(i,j)$, we have $w_{i}e_{i}w'_{i}<_{\tau}w_{j}e_{j}w'_{j}$. We extend the ordering $\tau$ to the set of obstructions $\bigcup_{1\leq i\leq j\leq s}{\rm Obs}(i,j)$ by committing the following slight abuse of notation. \[sec3def2\] Let $\tau$ be the module term ordering induced by $({\sigma},{G})$ on $\mathbb{T}(F_{s})$. Let ${\rm o}_{i,j}(w_{i},w'_{i};w_{j},w'_{j}), {\rm o}_{k,l}(w_{k},w'_{k};w_{l},w'_{l})$ be two obstructions in the set $\bigcup_{1\leq i\leq j\leq s}{\rm Obs}(i,j)$. If we have $w_{j}e_{j}w'_{j}>_{\tau}w_{l}e_{l}w'_{l}$, or if we have $w_{j}e_{j}w'_{j}=w_{l}e_{l}w'_{l}$ and $w_{i}e_{i}w'_{i}\geq_{\tau}w_{k}e_{k}w'_{k}$, then we let ${\rm o}_{i,j}(w_{i},w'_{i};w_{j},w'_{j})\geq_{\tau}{\rm o}_{k,l}(w_{k},w'_{k};$ $w_{l},w'_{l})$. The ordering $\tau$ is called the ordering induced by $({\sigma},{G})$ on the set of obstructions. One can verify that $\tau$ is also a well-ordering on $\bigcup_{1\leq i\leq j\leq s}{\rm Obs}(i,j)$ and compatible with scalar multiplication. Now we are ready to generalize the commutative Gebauer-Möller criteria (see [@CKR04] and [@GM88]) to the non-commutative case. Recall that, in step (B4) of the Buchberger Procedure, when a new generator $g_{s}$ is added, we immediately construct new obstructions $\bigcup_{1\leq i\leq s}{\rm NTObs}(i,s)$. We want to detect unnecessary obstructions in the set $\bigcup_{1\leq i\leq s}{\rm NTObs}(i,s)$ of newly constructed obstructions as well as in the set $\bigcup_{1\leq i\leq j\leq s-1}{\rm NTObs}(i,j)$ of previously constructed obstructions. We achieve this goal via the following three steps. Firstly, we detect unnecessary obstructions in the set $\bigcup_{1\leq i\leq s}{\rm NTObs}(i,s)$ with the aid of other obstructions also in this set. This step is called a head reduction step in [@CKR04]. Secondly, we detect unnecessary obstructions in the set $\bigcup_{1\leq i\leq s}{\rm NTObs}(i,s)$ with the aid of obstructions in the set $\bigcup_{1\leq i\leq j\leq s-1}{\rm NTObs}(i,j)$. This step is called a tail reduction step in [@CKR04]. Thirdly, we detect unnecessary obstructions in the set $\bigcup_{1\leq i\leq j\leq s-1}{\rm NTObs}(i,j)$ with the aid of the new generator $g_{s}$. Indeed, the first step corresponds to the commutative Gebauer-Möller criteria $M$ and $F$, and the last step corresponds to criterion $B_k$ (c.f. [@GM88], Subsection 3.4). The following lemma helps us to implement the first step, that is, to detect unnecessary obstructions in the set $\bigcup_{1\leq i\leq s}{\rm NTObs}(i,s)$ of newly constructed obstructions via other obstructions in this set. \[sec3lem3\] Let ${\rm o}_{i,s}(w_{i},w'_{i};u_{s},u'_{s})$ and ${\rm o}_{j,s}(w_{j},w'_{j};v_{s},v'_{s})$ be two distinct non-trivial obstructions in $\bigcup_{1\leq i\leq s}{\rm NTObs}(i,s)$ with two words $w,w'\in\langle X\rangle$ satisfying $u_{s}=$ $wv_{s}$ and $u'_{s}=v'_{s}w'$. - If $i<j$ and $ww'\neq1$, then we have $${\rm o}_{i,s}(w_{i},w'_{i};u_{s},u'_{s})=w{\rm o}_{j,s}(w_{j},w'_{j};v_{s},v'_{s})w'+{\rm o}_{i,j}(w_{i},w'_{i};ww_{j},w'_{j}w')$$ with ${\rm o}_{i,s}(w_{i},w'_{i};u_{s},u'_{s})>_{\tau}{\rm o}_{j,s}(w_{j},w'_{j};v_{s},v'_{s})$ and ${\rm o}_{i,s}(w_{i},w'_{i};u_{s},u'_{s})>_{\tau}{\rm o}_{i,j}(w_{i},w'_{i};ww_{j},w'_{j}w')$. Further, if the S-polynomials $S_{j,s}(w_{j},w'_{j};v_{s},v'_{s})$ and $S_{i,j}(w_{i},w'_{i};ww_{j},w'_{j}w')$ have Gröbner representations in terms of $G$, then so does $S_{i,s}(w_{i},w'_{i};u_{s},u'_{s})$. - If $i>j$, then we have $${\rm o}_{i,s}(w_{i},w'_{i};u_{s},u'_{s})=w{\rm o}_{j,s}(w_{j},w'_{j};v_{s},v'_{s})w'-{\rm o}_{j,i}(ww_{j},w'_{j}w';w_{i},w'_{i})$$ with ${\rm o}_{i,s}(w_{i},w'_{i};u_{s},u'_{s})>_{\tau}{\rm o}_{j,s}(w_{j},w'_{j};v_{s},v'_{s})$ and ${\rm o}_{i,s}(w_{i},w'_{i};u_{s},u'_{s})>_{\tau}{\rm o}_{j,i}(ww_{j},w'_{j}w';w_{i},w'_{i})$. Further, if the S-polynomials $S_{j,s}(w_{j},w'_{j};v_{s},v'_{s})$ and $S_{j,i}(ww_{j},w'_{j}w';w_{i},w'_{i})$ have Gröbner representations in terms of $G$, then so does $S_{i,s}(w_{i},w'_{i};u_{s},u'_{s})$. - If $i=j$ and $ww'\neq 1$ or if $i=j$ and $ww'=1$ and $w_i>_{\sigma}w_j$, then we have $${\rm o}_{i,s}(w_{i},w'_{i};u_{s},u'_{s})=w{\rm o}_{j,s}(w_{j},w'_{j};v_{s},v'_{s})w'+{\rm o}_{i,j}(w_{i},w'_{i};ww_{j},w'_{j}w')$$ with ${\rm o}_{i,s}(w_{i},w'_{i};u_{s},u'_{s})>_{\tau}{\rm o}_{j,s}(w_{j},w'_{j};v_{s},v'_{s})$ and ${\rm o}_{i,s}(w_{i},w'_{i};u_{s},u'_{s})>_{\tau}{\rm o}_{i,j}(w_{i},w'_{i};ww_{j},w'_{j}w')$. Further, if the S-polynomials $S_{j,s}(w_{j},w'_{j};v_{s},v'_{s})$ and $S_{i,j}(w_{i},w'_{i};ww_{j},w'_{j}w')$ have Gröbner representations in terms of $G$, then so does $_{i,s}(w_{i},w'_{i};u_{s},u'_{s})$. We prove case (a). Cases (b) and (c) can be proved similarly. The equation in case (a) follows from Definition \[sec2def4\].a and from the conditions $u_{s}=wv_{s}, u'_{s}=v'_{s}w'$ and $i<j$. Because of $ww'>1$, we have $u_{s}\lw(g_{s})u'_{s}=wv_{s}\lw(g_{s})v'_{s}w'>_{\sigma}v_{s}\lw(g_{s})v'_{s}$. Consequently, we get $u_{s}e_{s}u'_{s}>_{\tau}v_{s}e_{s}v'_{s}$ and ${\rm o}_{i,s}(w_{i},w'_{i};u_{s},u'_{s})>_{\tau}{\rm o}_{j,s}(w_{j},w'_{j};v_{s},v'_{s})$. From $u_{s}\lw(g_{s})u'_{s}$ $=w_{i}\lw(g_{i})w'_{i}=ww_{j}\lw(g_{j})w'_{j}w'$ and $s>j$, we get the inequalities $u_{s}e_{s}u'_{s}>_{\tau}ww_{j}e_{j}w'_{j}w'$ and ${\rm o}_{i,s}(w_{i},w'_{i};u_{s},u'_{s})>_{\tau}{\rm o}_{i,j}(w_{i},w'_{i};ww_{j},w'_{j}w')$. Next we show that, if $S_{j,s}(w_{j},w'_{j};v_{s},v'_{s})$ and $S_{i,j}(w_{i},w'_{i};ww_{j},w'_{j}w')$ have Gröbner representations in terms of $G$, then so does $S_{i,s}(w_{i},w'_{i};u_{s},u'_{s})$. Clearly we have $$S_{i,s}(w_{i},w'_{i};u_{s},u'_{s})=wS_{j,s}(w_{j},w'_{j};v_{s},v'_{s})w'+S_{i,j}(w_{i},w'_{i};ww_{j},w'_{j}w').$$ Without loss of generality, we assume that $S_{i,s}(w_{i},w'_{i};u_{s},u'_{s}), S_{j,s}(w_{j},w'_{j};v_{s},v'_{s})$ and $S_{i,j}(w_{i},w'_{i};ww_{j},w'_{j}w')$ are non-zero. Since there is a Gröbner representation for $S_{j,s}(w_{j},w'_{j};v_{s},v'_{s})$, we have $$S_{j,s}(w_{j},w'_{j};v_{s},v'_{s})=\sum^\mu_{k=1}a_kw_kg_{i_k}w'_k$$ with $a_k\in K\setminus\{0\},\ w_k,w'_k\in\langle X\rangle,\ g_{i_k}\in G$ for all $k\in\{1,\dots,\mu\}$, such that $\lw_{\sigma}(v_{s}g_{s}v'_{s})>_{\sigma}\lw_{\sigma}(a_kw_kg_{i_k}w'_k)$. Similarly, for $S_{i,j}(w_{i},w'_{i};ww_{j},w'_{j}w')$ we have $$S_{i,j}(w_{i},w'_{i};ww_{j},w'_{j}w')=\sum^\nu_{l=1}b_lw_lg_{i_l}w'_l$$ with $b_l\in K\setminus\{0\},\ w_l,w'_l\in\langle X\rangle,\ g_{i_l}\in G$ for all $l\in\{1,\dots,\nu\}$, such that $\lw_{\sigma}(ww_{j}g_{j}w'_{j}w')>_{\sigma}\lw_{\sigma}(b_lw_lg_{i_l}w'_l)$. Therefore we have $$\begin{aligned} S_{i,s}(w_{i},w'_{i};u_{s},u'_{s})&=&w(\sum^\mu_{k=1}a_kw_kg_{i_k}w'_k)w'+\sum^\nu_{l=1}b_lw_lg_{i_l}w'_l\\ &=&\sum^\mu_{k=1}a_kww_kg_{i_k}w'_kw'+\sum^\nu_{l=1}b_lw_lg_{i_l}w'_l.\end{aligned}$$ As $u_{s}\lw_{\sigma}(g_{s})u'_{s}=wv_{s}\lw_{\sigma}(g_{s})v'_{s}w'$, we have $\lw_{\sigma}(u_{s}g_{s}u'_{s})=\lw_{\sigma}(wv_{s}g_{s}v'_{s}w')$ $>_{\sigma}\lw_{\sigma}(ww_kg_{i_k}w'_kw')$ for all $k\in\{1,\dots,\mu\}$. By Definition \[sec2def4\], we have $\lw_{\sigma}(u_{s}g_{s}u'_{s})=\lw_{\sigma}(w_{i}g_iw'_{i})=\lw_{\sigma}(ww_{j}g_jw'_{j}w')>_{\sigma}\lw_{\sigma}(b_lw_lg_{i_l}w'_l)$ for all $l\in\{1,\dots,\nu\}$. Therefore $$S_{i,s}(w_{i},w'_{i};u_{s},u'_{s})=\sum^\mu_{k=1}a_kww_kg_{i_k}w'_kw'+\sum^\nu_{l=1}b_lw_lg_{i_l}w'_l$$ is a Gröbner representation of $S_{i,s}(w_{i},w'_{i};u_{s},u'_{s})$. The following example shows that the obstruction ${\rm o}_{i,j}(w_{i},w'_{i};ww_{j},w'_{j}w')$ in case (a) of Lemma \[sec3lem3\] can be a non-trivial obstruction, i.e. a multiple of a non-trivial obstruction or an obstruction without overlap. Similar phenomena occur in cases (b) and (c) of Lemma \[sec3lem3\], as well as in Lemmas \[sec3lem7\] and \[sec3lem10\]. \[sec3exa4\] Consider polynomials $G=\{g_{1},g_{2},g_{3}\}$ in the non-commutative polynomial ring $K\langle x,y\rangle$. - Assume that $\lm_{\sigma}(g_{1})=y^3, \lm_{\sigma}(g_{2})=x^2y^{2}$ and $\lm_{\sigma}(g_{3})=xyx^{2}y$. Then we have ${\rm o}_{1,3}(xyx^{2},1;1,y^{2}), {\rm o}_{2,3}(xy,1;1,y)\in\bigcup_{1\leq i\leq 3}{\rm NTObs}(i,3)$, and $${\rm o}_{1,3}(xyx^{2},1;1,y^{2})={\rm o}_{2,3}(xy,1;1,y)y+{\rm o}_{1,2}(xyx^2,1;xy,y).$$ Observe that ${\rm o}_{1,2}(xyx^2,xy;y)=xy{\rm o}_{1,2}(x^2,1;1,y)$ is a multiple of the non-trivial obstruction ${\rm o}_{1,2}(x^2,1;1,y)$. - Now assume that $\lm_{\sigma}(g_{1})=(xy)^{2}, \lm_{\sigma}(g_{2})=y$ and $\lm_{\sigma}(g_{3})=xyx^{2}y$. Then we have ${\rm o}_{1,3}(xyx,1;1,xy), {\rm o}_{2,3}(x,x^{2}y;1,1)\in\bigcup_{1\leq i\leq 3}{\rm NTObs}(i,3)$, and $${\rm o}_{1,3}(xyx,1;1,xy)={\rm o}_{2,3}(x,x^{2}y;1,1)xy+{\rm o}_{1,2}(xyx,1;x,x^{2}yxy).$$ One can check that ${\rm o}_{1,2}(xyx,1;x,x^{2}yxy)$ is an obstruction without overlap. In the following, we present the non-commutative multiply criterion and the leading word criterion. They are non-commutative analogues of the Gebauer-Möller criteria M and F, respectively. \[ncMCrit\] Suppose that ${\rm o}_{i,s}(w_{i},w'_{i};u_{s},u'_{s})$ and ${\rm o}_{j,s}(w_{j},w'_{j};v_{s},v'_{s})$ are two distinct non-trivial obstructions in $\bigcup_{1\leq i\leq s}{\rm NTObs}(i,s)$ such that there exist two words $w,w'$ in $\langle X\rangle$ satisfying $u_{s}=wv_{s}$ and $u'_{s}=v'_{s}w'$. Then we can remove the obstruction ${\rm o}_{i,s}(w_{i},w'_{i};u_{s},u'_{s})$ from $\bigcup_{1\leq i\leq s}{\rm NTObs}(i,s)$ in the execution of the Buchberger Procedure if $ww'\neq 1$. By the previous lemma, the obstruction ${\rm o}_{i,s}(w_{i},w'_{i};u_{s},u'_{s})$ can be represented as $${\rm o}_{i,s}(w_{i},w'_{i};u_{s},u'_{s})=w{\rm o}_{j,s}(w_{j},w'_{j};v_{s},v'_{s})w'+a{\rm o}_{k,l}(w_{k},w'_{k};w_{l},w'_{l})$$ with $a\in\{1,-1\}$ and $k=\min\{i,j\},l=\max\{i,j\}$. To prove that ${\rm o}_{i,s}(w_{i},w'_{i};$ $u_{s},u'_{s})$ is strictly larger than ${\rm o}_{j,s}(w_{j},w'_{j};v_{s},v'_{s})$ and ${\rm o}_{k,l}(w_{k},w'_{k};w_{l},w'_{l})$, we consider two cases. If $i>j$, then the result follows from Lemma \[sec3lem3\].b; if $i\leq j$, then the result follows from Lemma \[sec3lem3\].a and \[sec3lem3\].c and the condition $ww'\neq 1$. Moreover, the S-polynomial $S_{i,s}(w_{i},w'_{i};u_{s},u'_{s})$ has a Gröbner representation in terms of $G$ if $S_{j,s}(w_{j},w'_{j};v_{s},v'_{s})$ and $S_{k,l}(w_{k},w'_{k};w_{l},w'_{l})$ have Gröbner representations in terms of $G$. Theorem \[sec2the11\] ensures that $S_{j,s}(w_{j},w'_{j};v_{s},v'_{s})$ has a Gröbner representation in terms of $G$. Note that the obstruction ${\rm o}_{k,l}(w_{k},w'_{k};w_{l},w'_{l})$ can be either a multiple of a non-trivial obstruction or an obstruction without overlap (for instance, see Example \[sec3exa4\]). If ${\rm o}_{k,l}(w_{k},w'_{k};w_{l},w'_{l})$ is a multiple of a non-trivial obstruction, then Lemma \[sec2lem6\] and Theorem \[sec2the11\] guarantee that $S_{k,l}(w_{k},w'_{k};w_{l},w'_{l})$ has a Gröbner representation in terms of $G$. If ${\rm o}_{k,l}(w_{k},w'_{k};w_{l},w'_{l})$ is an obstruction without overlap, then, by Lemma \[sec2lem8\], its S-polynomial has a Gröbner representation in terms of $G$. Now the conclusion follows from Proposition \[sec2pro10\] and Theorem \[sec2the11\]. \[ncLWCrit\] Suppose that ${\rm o}_{i,s}(w_{i},w'_{i};u_{s},u'_{s})$ and ${\rm o}_{j,s}(w_{j},w'_{j};v_{s},v'_{s})$ are two distinct non-trivial obstructions in $\bigcup_{1\leq i\leq s}{\rm NTObs}(i,s)$ such that there exist two words $w,w'$ in $\langle X\rangle$ satisfying $u_{s}=wv_{s}$ and $u'_{s}=v'_{s}w'$. Then ${\rm o}_{i,s}(w_{i},w'_{i};u_{s},u'_{s})$ can be removed from $\bigcup_{1\leq i\leq s}{\rm NTObs}(i,s)$ in the execution of the Buchberger Procedure if one of the following conditions is satisfied. - $i>j$. - $i=j$ and $ww'=1$ and $w_{i}>_{\sigma}w_{j}$. Observe that condition (a) corresponds to Lemma \[sec3lem3\].b, while condition (b) corresponds to Lemma \[sec3lem3\].c. We represent ${\rm o}_{i,s}(w_{i},w'_{i};u_{s},u'_{s})$ as $${\rm o}_{i,s}(w_{i},w'_{i};u_{s},u'_{s})=w{\rm o}_{j,s}(w_{j},w'_{j};v_{s},v'_{s})w'-{\rm o}_{j,i}(ww_{j},w'_{j}w';w_{i},w'_{i}).$$ By Lemma \[sec3lem3\].b and \[sec3lem3\].c, we have ${\rm o}_{i,s}(w_{i},w'_{i};u_{s},u'_{s})$ is strictly larger than ${\rm o}_{j,s}(w_{j},w'_{j};v_{s},v'_{s})$ and ${\rm o}_{j,i}(ww_{j},w'_{j}w';w_{i},w'_{i})$. Moreover, if the S-polynomials $S_{j,s}(w_{j},w'_{j};v_{s},v'_{s})$ and $S_{j,i}(ww_{j},w'_{j}w';w_{i},w'_{i})$ have Gröbner representations in terms of $G$, then so does $S_{i,s}(w_{i},w'_{i};u_{s},u'_{s})$. Theorem \[sec2the11\] ensures $S_{j,s}(w_{j},w'_{j};$ $v_{s},v'_{s})$ has a Gröbner representation in terms of $G$. Note that the obstruction ${\rm o}_{j,i}(ww_{j},w'_{j}w';w_{i},w'_{i})$ can be either a multiple of a non-trivial obstruction or an obstruction without overlap (for instance, see Example \[sec3exa4\]). If ${\rm o}_{j,i}(ww_{j},w'_{j}w';w_{i},w'_{i})$ is a multiple of a non-trivial obstruction, then Lemma \[sec2lem6\] and Theorem \[sec2the11\] guarantee that $S_{j,i}(ww_{j},w'_{j}w';w_{i},w'_{i})$ has a Gröbner representation in terms of $G$. If ${\rm o}_{j,i}(ww_{j},w'_{j}w';w_{i},w'_{i})$ is an obstruction without overlap, then, by Lemma \[sec2lem8\], its S-polynomial has a Gröbner representation in terms of $G$. Now the conclusion follows from Proposition \[sec2pro10\] and Theorem \[sec2the11\]. Next we work on detecting unnecessary obstructions in $\bigcup_{1\leq i\leq s}{\rm NTObs}(i,s)$ via obstructions in the set $\bigcup_{1\leq i\leq j\leq s-1}{\rm NTObs}(i,j)$ of previously constructed obstructions. \[sec3lem7\] Let ${\rm o}_{j,s}(u_{j},u'_{j};w_{s},w'_{s})$ and ${\rm o}_{i,j}(w_{i},w'_{i};v_{j},v'_{j})$ be non-trivial obstructions in $\bigcup_{1\leq i\leq s}{\rm NTObs}(i,s)$ and $\bigcup_{1\leq i\leq j\leq s-1}{\rm NTObs}(i,j)$, respectively. If there exist two words $w,w'\in\langle X\rangle$ such that $u_{j}=wv_{j}$ and $u'_{j}=v'_{j}w'$, then we have $${\rm o}_{j,s}(u_{j},u'_{j};w_{s},w'_{s})=-w{\rm o}_{i,j}(w_{i},w'_{i};v_{j},v'_{j})w'+ {\rm o}_{i,s}(ww_{i},w'_{i}w';w_{s},w'_{s})$$ where the inequalities ${\rm o}_{j,s}(u_{j},u'_{j};w_{s},w'_{s}) >_{\tau}{\rm o}_{i,j}(w_{i},w'_{i};v_{j},v'_{j})$ and ${\rm o}_{j,s}(u_{j},u'_{j};$ $w_{s},w'_{s}) >_{\tau}{\rm o}_{i,s}(ww_{i},w'_{i}w';w_{s},w'_{s})$ hold. Further, if the S-polynomials $S_{i,j}(w_{i},w'_{i};$ $v_{j},v'_{j})$ and $S_{i,s}(ww_{i},w'_{i}w';w_{s},w'_{s})$ have Gröbner representations in terms of $G$, then so does $S_{j,s}(u_{j},u'_{j};w_{s},w'_{s})$. The claimed equality follows from Definition \[sec2def4\].a and from the conditions $u_{j}=wv_{j}$ and $u'_{j}=v'_{j}w'$. We have ${\rm o}_{j,s}(u_{j},u'_{j}; w_{s},w'_{s})>_{\tau}{\rm o}_{i,j}(w_{i},w'_{i};v_{j},v'_{j})$ for $w_{s}e_{s}w'_{s}>_{\tau}u_{j}e_{j}u'_{j}= wv_{j}e_{j}v'_{j}w\geq_{\tau}v_{j}e_{j}v'_{j}$. From the inequality $u_{j}e_{j}u'_{j}= wv_{j}e_{j}v'_{j}w>_{\tau}ww_{i}e_{i}w'_{i}w'$, it follows that ${\rm o}_{j,s}(u_{j},u'_{j};w_{s},w'_{s}) >_{\tau}{\rm o}_{i,s}(ww_{i},w'_{i}w';$ $w_{s},w'_{s})$. Again, we can prove the second part by following the same argument as in the proof of Lemma \[sec3lem3\].a. Note that the obstruction ${\rm o}_{i,s}(ww_{i},w'_{i}w';w_{s},w'_{s})$ in Lemma \[sec3lem7\] can be either a multiple of a non-trivial obstruction or an obstruction without overlap. However, it suffices for us to consider only the latter case, since the former case has been considered in Proposition \[ncMCrit\], and, more precisely, in Lemma \[sec3lem3\].b. \[ncTailRed\] Suppose that ${\rm o}_{j,s}(u_{j},u'_{j};w_{s},w'_{s})$ and ${\rm o}_{i,j}(w_{i},w'_{i};v_{j},v'_{j})$ are non-trivial obstructions in $\bigcup_{1\leq i\leq s}{\rm NTObs}(i,s)$ and $\bigcup_{1\leq i\leq j\leq s-1}{\rm NTObs}(i,j)$, respectively, such that there exist two words $w,w'\in\langle X\rangle$ satisfying $u_{j}=wv_{j}$ and $u'_{j}=v'_{j}w'$. If the word $ww_{i}$ is a multiple of $w_{s}\lw_{\sigma}(g_{s})$, or if the word $w'_{i}w'$ is a multiple of $\lw_{\sigma}(g_{s})w'_{s}$, then ${\rm o}_{j,s}(u_{j},u'_{j};w_{s},w'_{s})$ can be removed from $\bigcup_{1\leq i\leq s}{\rm NTObs}(i,s)$ in the execution of the Buchberger Procedure. By Lemma \[sec3lem7\], the obstruction ${\rm o}_{j,s}(u_{j},u'_{j};w_{s},w'_{s})$ can be represented as $${\rm o}_{j,s}(u_{j},u'_{j};w_{s},w'_{s})= -w{\rm o}_{i,j}(w_{i},w'_{i};v_{j},v'_{j})w'+{\rm o}_{i,s}(ww_{i},w'_{i}w';w_{s},w'_{s})$$ where the inequalitites ${\rm o}_{j,s}(u_{j},u'_{j};w_{s},w'_{s})>_\tau{\rm o}_{i,j}(w_{i},w'_{i};v_{j},v'_{j})$ and ${\rm o}_{j,s}(u_{j},u'_{j};$ $w_{s},w'_{s})>_\tau{\rm o}_{i,s}(ww_{i},w'_{i}w';w_{s},w'_{s})$ hold. Further, if the S-polynomials $S_{i,j}(w_{i},w'_{i};$ $v_{j},v'_{j})$ and $S_{i,s}(ww_{i},w'_{i}w';w_{s},w'_{s})$ have Gröbner representations in terms of $G$, then so does $S_{j,s}(u_{j},u'_{j};w_{s},w'_{s})$. Theorem \[sec2the11\] ensures that $S_{i,j}(w_{i},w'_{i};v_{j},v'_{j})$ has a Gröbner representation in terms of $G$. Note that $ww_{i}$ is a multiple of $w_{s}\lw_{\sigma}(g_{s})$ or $w'_{i}w'$ is a multiple of $\lw_{\sigma}(g_{s})w'_{s}$. This implies that ${\rm o}_{i,s}(ww_{i},$ $w'_{i}w';w_{s},w'_{s})$ has no overlap. By Lemma \[sec2lem8\], $S_{i,s}(ww_{i},w'_{i}w';w_{s},w'_{s})$ has a Gröbner representation in terms of $G$. Now the conclusion follows from Proposition \[sec2pro10\] and Theorem \[sec2the11\]. \[sec3rem9\] Our experiments in the final section show that, after applying the previous two criteria, the Non-Commutative Tail Reduction is unlikely to apply in the Buchberger Procedure. This may be due to the fact that frequently the Non-Commutative Multiply Criterion and the Non-Commutative Leading Word Criterion have already detected all unnecessary obstructions in the set $\bigcup_{1\leq i\leq s}{\rm NTObs}(i,s)$ of newly constructed obstructions. So far we have detected unnecessary obstructions in the set $\bigcup_{1\leq i\leq s}{\rm O}(i,s)$ of newly constructed obstructions. Intuitively, we are also able to detect unnecessary obstructions in the set $\bigcup_{1\leq i\leq j\leq s-1}{\rm Obs}(i,j)$ of previously constructed obstructions. Thus, in the last step, we detect unnecessary obstructions in this set by using the new generator $g_{s}$. \[sec3lem10\] Let ${\rm o}_{i,j}(w_{i},w'_{i};w_{j},w'_{j})\in\bigcup_{1\leq i\leq j\leq s-1}{\rm NTObs}(i,j)$ be a non-trivial obstruction. If there are two words $w,w'\in\langle X\rangle$ satisfying $w_{j}\lw_{\sigma}(g_{j})w'_{j}=w\lw_{\sigma}(g_{s})w'$, then we can represent ${\rm o}_{i,j}(w_{i},w'_{i};w_{j},w'_{j})$ as $${\rm o}_{i,j}(w_{i},w'_{i};w_{j},w'_{j})={\rm o}_{i,s}(w_{i},w'_{i};w,w')-{\rm o}_{j,s}(w_{j},w'_{j};w,w').$$ Moreover, if $S_{i,s}(w_{i},w'_{i};w,w')$ and $S_{j,s}(w_{j},w'_{j};w,w')$ have Gröbner representations in terms of $G$, then so does $S_{i,j}(w_{i},w'_{i};w_{j},w'_{j})$. The claimed equality follows from Definition \[sec2def4\].a and the condition $w_{j}\lw_{\sigma}(g_{j})w'_{j}$ $=w\lw_{\sigma}(g_{s})w'$. The proof of the second part is analogous to the proof of the second part of Lemma \[sec3lem3\].a. The following example shows that the obstruction ${\rm o}_{i,s}(w_{i},w'_{i};w,w')$ in the equation of Lemma \[sec3lem10\] can be either an obstruction without overlap or a multiple of a non-trivial obstruction. In the case that ${\rm o}_{i,s}(w_{i},w'_{i};w,w')$ is a multiple of a non-trivial obstruction, say ${\rm o}_{i,s}(\tilde{w}_{i},\tilde{w}'_{i};\tilde{w},\tilde{w}')$, the example shows that it is not necessary to have ${\rm o}_{i,s}(w_{i},w'_{i};w,w')>_{\tau}{\rm o}_{i,s}(\tilde{w}_{i},\tilde{w}'_{i};\tilde{w},\tilde{w}')$ (compared to Lemmas \[sec3lem3\] and \[sec3lem7\]). The same also holds for the obstruction ${\rm o}_{j,s}(w_{j},w'_{j};w,w')$ in the equation of Lemma \[sec3lem10\]. \[sec3exa11\] Consider polynomials $G=\{g_{1},g_{2},g_{3}\}$ in the non-commutative polynomial ring $K\langle x,y\rangle$ with $\lm_{\sigma}(g_{1})=x^3yx, \lm_{\sigma}(g_{2})=x^2$ and $\lm_{\sigma}(g_{3})=x$. We have ${\rm o}_{1,2}(1,1;x,yx)\in\bigcup_{1\leq i\leq j\leq 2}{\rm NTObs}(i,j)$ and $x\lw_{\sigma}(g_{2})yx=x^3yx=x^3y\lw_{\sigma}(g_{3})$ and $${\rm o}_{1,2}(1,1;x,yx)={\rm o}_{1,3}(1,1;x^3y,1)-{\rm o}_{2,3}(x,yx;x^3y,1).$$ One can check that ${\rm o}_{1,3}(1,1;x^3y,1)$ is a non-trivial obstruction in ${\rm NTObs}(1,3)$ and ${\rm o}_{1,2}(1,1;x,yx)<_{\tau}{\rm o}_{1,3}(1,1;x^3y,1)$. Moreover, ${\rm o}_{2,3}(x,yx;x^3y,1)$ is an obstruction without overlap. The following is a non-commutative analogue of the Gebauer-Möller criterian $B_k$, which is also known as the chain criterion (or criterion 1) of Buchberger (cf. [@Bu85]). \[ncBKCrit\] Suppose that ${\rm o}_{i,j}(w_{i},w'_{i};w_{j},w'_{j})\in\bigcup_{1\leq i\leq j\leq s-1}{\rm NTObs}(i,j)$ is a non-trivial obstruction. Then in the execution of the Buchberger Procedure ${\rm o}_{i,j}(w_{i},w'_{i};w_{j},w'_{j})$ can be removed from $\bigcup_{1\leq i\leq j\leq s-1}{\rm NTObs}(i,j)$ if the following three conditions are satisfied. - There are two words $w,w'\in\langle X\rangle$ such that $w_{j}\lw_{\sigma}(g_{j})w'_{j}=w\lw_{\sigma}(g_{s})w'$. - The obstruction ${\rm o}_{i,s}(w_{i},w'_{i};w,w')$ is either an obstruction without overlap or a multiple of a non-trivial obstruction in $\bigcup_{1\leq i\leq s}{\rm NTObs}(i,s)$. - The obstruction ${\rm o}_{j,s}(w_{j},w'_{j};w,w')$ is either an obstruction without overlap or a multiple of a non-trivial obstruction in $\bigcup_{1\leq i\leq s}{\rm NTObs}(i,s)$. By Lemma \[sec3lem10\], we can represent ${\rm o}_{i,j}(w_{i},w'_{i};w_{j},w'_{j})$ as $${\rm o}_{i,j}(w_{i},w'_{i};w_{j},w'_{j})={\rm o}_{i,s}(w_{i},w'_{i};w,w')-{\rm o}_{j,s}(w_{j},w'_{j};w,w').$$ Moreover, $S_{i,j}(w_{i},w'_{i};w_{j},w'_{j})$ has a Gröbner representations in terms of $G$ if $S_{i,s}(w_{i},w'_{i};w,w')$ and $S_{j,s}(w_{j},w'_{j};w,w')$ have Gröbner representations in terms of $G$. If ${\rm o}_{i,s}(w_{i},w'_{i};w,w')$ is an obstruction without overlap, then, by Lemma \[sec2lem8\], its S-polynomial has a Gröbner representations in terms of $G$. If it is a multiple of a non-trivial obstruction in $\bigcup_{1\leq i\leq s}{\rm NTObs}(i,s)$, then Lemma \[sec2lem8\] and Theorem \[sec2the11\] ensure that $S_{i,s}(w_{i},w'_{i};w,w')$ has a Gröbner representations in terms of $G$. By the same argument, one can show that $S_{j,s}(w_{j},w'_{j};w,w')$ has a Gröbner representations in terms of $G$. Now the conclusion follows from Proposition \[sec2pro10\] and Theorem \[sec2the11\]. We would like to mention that the Non-Commutative Backward Criterion given in Proposition \[ncBKCrit\] covers in particular all useless obstructions presented by T. Mora in [@Mo94], Lemma 5.11. \[sec3rem13\] In order to apply Propositions \[ncMCrit\], \[ncLWCrit\], \[ncTailRed\] and \[ncBKCrit\] to remove unnecessary obstructions during the execution of the Buchberger Procedure, it is crucial to make sure that the S-polynomials of those removed obstructions have Gröbner representations. - Propositions \[ncMCrit\], \[ncLWCrit\] and \[ncTailRed\] remove unnecessary non-trivial obstructions, say ${\rm o}_{i,s}(w_{i},w'_{i};w_{s},w'_{s})$, from the set $\bigcup_{1\leq i\leq s}{\rm NTObs}(i,s)$ of newly constructed obstructions. The Gröbner representation of the S-polynomial $S_{i,s}(w_{i},w'_{i};w_{s},w'_{s})$ depends on the Gröbner representations of the S-polynomials of two smaller obstructions in the set $\bigcup_{1\leq i\leq j\leq s-1}{\rm Obs}(i,j)$ and the set $\bigcup_{1\leq i\leq s}{\rm Obs}(i,s)$. - Proposition \[ncBKCrit\] removes unnecessary obstructions, say ${\rm o}_{i,j}(w_{i},w'_{i};w_{j},w'_{j})$, from the set $\bigcup_{1\leq i\leq j\leq s-1}{\rm Obs}(i,j)$ of previously constructed obstructions. The Gröbner representation of $S_{i,j}(w_{i},w'_{i};w_{j},w'_{j})$ depends on the Gröbner representations of the S-polynomials of two obstructions, say ${\rm o}_{k,s}(w_{k},w'_{k};$ $u_{s},u'_{s})$ and ${\rm o}_{l,s}(w_{l},w'_{l};v_{s},v'_{s})$, in $\bigcup_{1\leq i\leq s}{\rm Obs}(i,s)$, which are not necessarily smaller than ${\rm o}_{i,j}(w_{i},w'_{i};w_{j},w'_{j})$. If ${\rm o}_{k,s}(w_{k},w'_{k};u_{s},u'_{s})$ is a multiple of a non-trivial obstruction, say ${\rm o}_{k,s}(\tilde{w}_{k},\tilde{w}'_{k};\tilde{u}_{s},\tilde{u}'_{s})$, in $\bigcup_{1\leq i\leq s}{\rm NTObs}(i,s)$, then, before removing ${\rm o}_{i,j}(w_{i},w'_{i};w_{j},w'_{j})$, it is important to ensure that ${\rm o}_{k,s}(\tilde{w}_{k},\tilde{w}'_{k};\tilde{u}_{s},\tilde{u}'_{s})$ is in $\bigcup_{1\leq i\leq s}{\rm NTObs}(i,s)$. The same check should be applied to ${\rm o}_{l,s}(w_{l},w'_{l};v_{s},v'_{s})$. Observe that Propositions \[ncMCrit\], \[ncLWCrit\] and \[ncBKCrit\] are actually generalizations of the well-known Gebauer-Möller criteria (see [@CKR04] and [@GM88]) in commutative polynomial rings. More precisely, Propositions \[ncMCrit\], \[ncLWCrit\] and \[ncBKCrit\] correspond to criterion $M$, criterion $F$ and criterion $B_k$, respectively (c.f. [@GM88], Subsection 3.4). Using the Gebauer-Möller criteria, we can improve the Buchberger Procedure as follows. \[sec3the14\] In the setting of Theorem \[sec2the11\], we replace step [(B4)]{} by the following sequence of instructions. - Increase $s$ by one. Append $g_{s}=S'$ to the set ${G}$, and form the set of non-trivial obstructions ${\rm NTObs}(s)=\bigcup_{1\leq i\leq s}{\rm NTObs}(i,s)$. - Remove from ${\rm NTObs}(s)$ all obstructions ${\rm o}_{i,s}(w_{i},w'_{i};u_{s},u'_{s})$ such that there exists an obstruction ${\rm o}_{j,s}(w_{j},w'_{j};v_{s},v'_{s})\in{\rm NTObs}(s)$ with the properties that there exist two words $w,w'\in\langle X\rangle$ satisfying $u_{s}=wv_{s}$, $u'_{s}=v'_{s}w'$ and $ww'\neq1$. - Remove from ${\rm NTObs}(s)$ all obstructions ${\rm o}_{i,s}(w_{i},w'_{i};u_{s},u'_{s})$ such that there exists an obstruction ${\rm o}_{j,s}(w_{j},w'_{j};v_{s},v'_{s})\in{\rm NTObs}(s)$ with the properties that there exist two words $w,w'\in\langle X\rangle$ satisfying $u_{s}=wv_{s}$, $u'_{s}=v'_{s}w'$, and such that $i>j$, or $i=j$ and $ww'=1$ and $w_{i}>_{\sigma}w_{j}$. - Remove from ${\rm NTObs}(s)$ all obstructions ${\rm o}_{j,s}(u_{j},u'_{j};w_{s},w'_{s})$ such that there exists an obstruction ${\rm o}_{i,j}(w_{i},w'_{i};v_{j},v'_{j})\in B$ with the properties that there exist two words $w,w'\in\langle X\rangle$ satisfying $u_{j}=wv_{j}, u'_{j}=v'_{j}w'$, and such that ${\rm o}_{i,s}(ww_{i},w'_{i}w';w_{s},w'_{s})$ has no overlap. - Remove from $B$ all obstructions ${\rm o}_{i,j}(w_{i},w'_{i};w_{j},w'_{j})$ such that there exist two words $w,w'\in\langle X\rangle$ satisfying $w\lw_{\sigma}(g_{s})w'=w_{j}\lw_{\sigma}(g_{j})w'_{j}$, and such that the following conditions are satisfied. - ${\rm o}_{i,s}(w_{i},w'_{i};w,w')$ is either an obstruction without overlap or a multiple of a non-trivial obstruction in ${\rm NTObs}(s)$. - ${\rm o}_{j,s}(w_{j},w'_{j};w,w')$ is either an obstruction without overlap or a multiple of a non-trivial obstruction in ${\rm NTObs}(s)$. - Replace $B$ by $B\cup {\rm NTObs}(s)$ and continue with step [(B2)]{}. Then the resulting set of instructions is a procedure that enumerates a ${\sigma}$-Gröbner basis ${G}$ of $I$. If $I$ has a finite ${\sigma}$-Gröber basis, it stops after finitely many steps and the resulting set ${G}$ is a finite ${\sigma}$-Gröbner basis of $I$. This follows from Theorem \[sec2the11\] and Propositions \[ncMCrit\], \[ncLWCrit\], \[ncTailRed\] and \[ncBKCrit\]. Experiments and Conclusions {#sec4} =========================== In this section we want to present some experimental data which illustrate the performance of the Gebauer-Möller criteria presented in Propositions \[ncMCrit\], \[ncTailRed\] and \[ncBKCrit\]. The computations are based on an implementation (using C++) in an experimental version of the ApCoCoA library (see [@Ap10]) by the second author. \[sec4exa1\] Consider the non-commutative polynomial ring $\mathbb{Q}\langle a,b\rangle$ equipped with the word ordering $\LLex$ on $\langle a,b\rangle$ such that $a>_{\LLex}b$. We take the list of finite generalized triangle groups from [@RS02], Theorem 2.12 and construct a list of ideals in $\mathbb{Q}\langle a,b\rangle$. For $k=1,\dots,13$ let $I_{k}=\langle G_{k}\rangle\subseteq\mathbb{Q}\langle a,b\rangle$ be the ideal generated by the following set of polynomials $G_{k}\subseteq\mathbb{Q}\langle a,b\rangle$. $$\begin{aligned} G_{1}&=&\{a^2-1,b^3-1,(ababab^2ab^2)^2-1\},\\ G_{2}&=&\{a^2-1,b^3-1,(ababab^2)^3-1\},\\ G_{3}&=&\{a^3-1,b^3-1,(abab^2)^2-1\},\\ G_{4}&=&\{a^3-1,b^3-1,(aba^2b^2)^2-1\},\\ G_{5}&=&\{a^2-1,b^5-1,(abab^2)^2-1\},\\ G_{6}&=&\{a^2-1,b^5-1,(ababab^4)^2-1\}, \\ G_{7}&=&\{a^2-1,b^5-1,(abab^2ab^4)^2-1\},\\ G_{8}&=&\{a^2-1,b^4-1,(ababab^3)^2-1\},\\ G_{9}&=&\{a^2-1,b^3-1,(abab^2)^2-1\}, \\ G_{10}&=&\{a^2-1,b^3-1,(ababab^2)^2-1\},\\ G_{11}&=&\{a^2-1,b^3-1,(abababab^2)^2-1\},\\ G_{12}&=&\{a^2-1,b^3-1,(ababab^2abab^2)^2-1\},\\ G_{13}&=&\{a^2-1,b^3-1,(ababababab^2ab^2)^2-1\}.\end{aligned}$$ The following table lists some numbers of polynomials and obstructions treated by the Improved Buchberger Procedure given in Theorem \[sec3the14\]. $k$ $\#(Gb)$ $\!\!\#(RGb)\!\!$ $\#(Tot)$ $\#(Sel)$ $\#(M)$ $\#(F)$ $\#(B_k)$ $\rho$ ----- ---------- ------------------- ----------- ----------- --------- --------- ----------- -------- -- 1 62 35 7032 248 6512 48 224 0.0353 2 133 96 31700 533 30571 70 526 0.0168 3 50 40 2828 197 2489 11 131 0.0697 4 64 28 4702 253 4185 46 218 0.0538 5 35 21 1580 115 1348 24 93 0.0728 6 199 164 51175 882 49126 26 1141 0.0172 7 200 164 51864 886 49818 17 1143 0.0170 8 53 37 3756 192 3357 19 188 0.0511 9 11 5 150 31 98 8 13 0.2067 10 22 15 741 74 605 18 44 0.0999 11 30 21 1573 116 1324 50 83 0.0737 12 97 70 16841 365 15989 97 390 0.0217 13 220 194 87673 1021 85136 153 1363 0.0116 Here we used the following abbreviations. - $\#(Gb)$ is the number of elements of the Gröbner basis returned by the procedure. - $\#(RGb)$ is the cardinality of the reduced Gröbner basis of the corresponding ideal. - $\#(Tot)$ is the total number of non-trivial obstructions constructed during the Buchberger Procedure. - $\#(Sel)$ is the number of actually selected and analysed non-trivial obstructions. - $\#(M)$ is the number of unnecessary non-trivial obstructions detected by the Non-Commutative Multiply Criterion given in Proposition \[ncMCrit\]. - $\#(F)$ is the number of unnecessary non-trivial obstructions detected by the Non-Commutative Leading Word Criterion given in Proposition \[ncLWCrit\]. - $\#(B_k)$ is the number of unnecessary non-trivial obstructions detected by the Non-Commutative Backward Criterion given in Proposition \[ncBKCrit\]. - $\rho=\#(Sel)/\#(Tot)$. Note that $\#(RGb)$ is an invariant of the ideal which only depends on chosen word ordering. Other numbers in the table rely also on the selection strategy. In our experiments we used the normal strategy which first chooses the obstruction whose S-polynomial has the lowest degree and then breaks ties by choosing the obstruction whose S-polynomial has the smallest leading word with respect to the word ordering. In these experiments, the Non-Commutative Tail Reduction given in Proposition \[ncTailRed\] detected no of unnecessary non-trivial obstruction (see Remark \[sec3rem9\]). Apparently, the Non-Commutative Multiply Criterion and the Non-Commutative Leading Word Criterion had already detected all unnecessary obstructions in the set $\bigcup_{1\leq i\leq s}{\rm NTObs}(i,s)$ of newly constructed obstructions. The low ratios $\rho$ in the table indicate that the non-commutative Gebauer-Möller criteria we obtained can detect most unnecessary obstructions during the procedure. \[sec4exa2\] The following ideals [braid3]{} and [braid4]{} in the non-commutative polynomial ring $\mathbb{Q}\langle x_{1},x_{2},x_{3}\rangle$ are taken from [@SL09], Section 5. More precisely, [braid3]{} is the ideal generated by the set $\{-x_{2}x_{3}x_{1}+x_{3}x_{1}x_{3},$ $x_{2}x_{1}x_{2}-x_{3}x_{2}x_{3}$, $x_{1}x_{2}x_{1}-x_{3}x_{1}x_{2}, x_{1}^3+x_{1}x_{2}x_{3}+x_{2}^3+x_{3}^3\}$, and [braid4]{} is the ideal generated by the set $\{-x_{2}x_{3}x_{1}+x_{3}x_{1}x_{3}, x_{2}x_{1}x_{2}-x_{3}x_{2}x_{3}, x_{1}x_{2}x_{3}-x_{3}x_{1}x_{2}, x_{1}^3+x_{1}x_{2}x_{3}+x_{2}^3+x_{3}^3\}$. These ideals are generated by sets of homogeneous generators. The following table lists the results of the computations of Gröbner bases truncated at degree $11$ with respect to $\LLex$ on $\langle x_{1},x_{2},x_{3}\rangle$ such that $x_{1}>_{\LLex} x_{2}>_{\LLex}x_{3}$, via the Improved Buchberger Procedure. $\#(Gb)$ $\#(Tot)$ $\#(Sel)$ $\rho$ ----------- ---------- ----------- ----------- -------- braid3-11 726 289642 1663 0.0057 braid4-11 416 93252 1150 0.0123 The meaning of the symbols is the same as in Example \[sec4exa1\]. In this experiment we also used the normal strategy. Moreover, since we compute truncated Gröbner bases, we discard those obstructions whose S-polynomial have degrees larger than the degree of truncation. Thus the ratios $\rho$ in the table are lower than the ratios in the table of Example \[sec4exa1\]. Again, the non-commutative Gebauer-Möller criteria detect most unnecessary obstructions during the procedure. The experimental data in Examples \[sec4exa1\] and \[sec4exa2\] show that the generalizations of the Gebauer-Möller criteria presented in Propositions \[ncMCrit\], \[ncLWCrit\] and \[ncBKCrit\] can successfully detect a large number of unnecessary obstructions. In fact, they apparently detect almost all unnecessary obstructions during the Buchberger Procedure. Acknowledgements {#acknowledgements .unnumbered} ---------------- The second author is grateful to the Chinese Scholarship Council (CSC) for providing partial financial support. Both authors thank G. Studzinski for valuable discussions about non-commutative Gröbner bases. And both authors appreciate anonymous referees for careful reading and useful suggestions. [00]{} ApCoCoA team, ApCoCoA: Applied Computations in Commutative Algebra, available at http://www.apcocoa.org. G.H. Bergman, The diamond lemma for ring theory, Adv. Math. 29 (1978), 178-218. B. Buchberger, Ein Algorithmus zum Auffinden der Basiselemente des Restklassenrings nach einem nulldimensionalen Polynomideal, Dissertation, Universität Inssbruck, Austria, 1965. B. Buchberger, A criterion for detecting unnecessary reductions in the construction of Groebner bases, in: Proceedings of the International Symposiumon on Symbolic and Algebraic Computation (EUROSAM ’79), Edward W. Ng (Ed.), Springer-Verlag, London, UK, 3-21, 1979. B. Buchberger, Gröbner bases: an algorithmic method in polynomial ideal theory, in: Multidimensional Systems Theory-Progress, Directions and Open Problems in Multidimensional Systems, N.K. Bose (Ed.), Reidel Publishing Company, Dodrecht-Boston-Lancaster, 184-232, 1985. M. Caboara, M. Kreuzer and L. Robbiano, Efficiently computing minimal sets of critical pairs, Journal of Symbolic Computation 38 (2004), 1169-1190. A.M. Cohen, Non-commutative polynomial computations, 2007, available at www.win.tue.nl/$\sim$amc/pub/gbnpaangepast.pdf. R. Gebauer and H.M. Möller, On an installation of Buchberger’s algorithm, Journal Symbolic Computation 6 (1988), 275-286. M. Kreuzer and L. Robbiano, Computational Commutative Algebra 1, Springer, Heidelberg, 2000. M. Kreuzer and L. Robbiano, Computational Commutative Algebra 2, Springer, Heidelberg, 2005. H.M. Möller, A reduction strategy for the Taylor resolution, Proc. EUROCAL 85, Springer L.N. in Comp. Sci. 162, 526-534. T. Mora, Gröbner bases for non-commutative polynomial rings, in: Proceedings of the 3rd International Conference on Algebraic Algorithms and Error-Correcting Codes (AAECC-3), Jacques Calmet (Ed.), Springer-Verlag, London, UK, 1986, 353-362. T. Mora, An introduction to commutative and non-commutative Gröbner Bases, Journal of Theoretical Computer Science 134 (1994), 131-173. G. Rosenberger and M. Scheer, Classification of the finite generalized tetrahedron groups, Contemporary Mathematics 296 (2002), 207- 229. R. Scala and V. Levandovskyy, Letterplace ideals and non-commutative Gröbner bases, Journal of Symbolic Computation 44 (2004), 1374-1393. X. Xiu, Non-Commutative Gröbner Bases and Applications, Dissertation, Universität Passau, Germany, 2012. [^1]: [[email protected]]{} (Corresponding author) [^2]: [[email protected]]{}	{ "pile_set_name": "ArXiv" }
[What does a convolutional neural network recognize in the moon?]{}\ \ [Daigo Shoji$^{1}$]{}\ \ 1. Earth-Life Science Institute, Tokyo Institute of Technology, 2-12-1 Ookayama, Meguro-ku, Tokyo ([email protected])\ Abstract {#abstract .unnumbered} ======== Many people see a human face or animals in the pattern of the maria on the moon. Although the pattern corresponds to the actual variation in composition of the lunar surface, the culture and environment of each society influence the recognition of these objects (i.e., symbols) as specific entities. In contrast, a convolutional neural network (CNN) recognizes objects from characteristic shapes in a training data set. Using CNN, this study evaluates the probabilities of the pattern of lunar maria categorized into the shape of a crab, a lion and a hare. If Mare Frigoris (a dark band on the moon) is included in the lunar image, the lion is recognized. However, in an image without Mare Frigoris, the hare has the highest probability of recognition. Thus, the recognition of objects similar to the lunar pattern depends on which part of the lunar maria is taken into account. In human recognition, before we find similarities between the lunar maria and objects such as animals, we may be persuaded in advance to see a particular image from our culture and environment and then adjust the lunar pattern to the shape of the imagined object. Introduction ============ Because the periods of rotation and revolution are equal in the synchronous rotation of the moon, everyone on Earth sees almost the same lunar surface. The surface of the moon has color variation caused by the differences in composition. The large dark areas on the moon are called maria, which are composed of basalts erupted from the interior of the moon \[1, 2\]. People recognize familiar objects from the pattern of the lunar maria, although the objects differ according to one’s society and environment. In Europe, the lunar pattern is considered to be a human face \[2\]. On the other hand, the lunar pattern is considered to be a hare in Asian countries. In addition to the human face and the hare, a crab and a lion have also been reported as animals in the moon (Fig. \[fig1\]) \[3\]. Although these objects in the moon are imagined due to the similarity between the lunar pattern and the shape of an animal, the culture and the environment in each society also influence the kind of object recognized in the moon. For example, in one of the old Buddhist tales, which are called “Jataka”, a hare is exalted to the moon \[4\]. Thus, the hare in the moon believed in Asian countries is related to Buddhism. Recently, deep learning has achieved great success for recognizing and categorizing the objects in images. For example, deep learning of the CNN has been applied to the recognition of hand-written digits \[5\], human faces \[6\], and the detection of craters on Mars \[7\]. The CNN recognizes objects by learning the shape of training images. Thus, unlike human recognition, the CNN does not identify an object via culture and environment. In this work, shape recognition of the lunar pattern for the crab, lion and hare is performed by the CNN. First, the CNN learns the silhouette images, which consist of outlines of the three animals. By using silhouette images, even a simple CNN can categorize images with high accuracy and reduced computation time. Moreover, images of real animals are not required because the outline of the maria pattern is important. Evaluation using the real animal images is beyond the scope of this work. After training of the CNN and then testing the shapes of the lunar maria, the probabilities of the lunar pattern categorized to each animal are evaluated. It should be noted that the CNN does not learn the shape of the lunar pattern from the training of images. That is, the CNN recognizes the pattern of the lunar maria only by the shapes of the selected animals. Training of CNN =============== The structure of the CNN is shown in Fig. \[fig2\]. The structure is composed of two convolutional layers and two affine layers. Training and testing of images are conducted with the Keras library \[8\]. The kernel size of the two convolutional layers is set at 3$\times$3, and the pooling layer decreases the size of the featured maps by 50%. The optimizer is Adam, and loss functions and probabilities are calculated by cross-entropy and soft-max functions, respectively \[8\]. The numbers of collected silhouette images of the crab, lion and hare were 100, 100 and 105, respectively (I bought silhouette images from Deposit Photos (https://jp.depositphotos.com/), and downloaded free images from Silhouette AC (https://www.silhouette-ac.com/index.html), Illust AC (https://www.ac-illust.com), Clipart Library (http://clipart-library.com), Silhouette Design (http://kage-design.com/)). Example images are shown in Fig. \[fig3\]. Of these 305 images, 240 images (80 images in each class) were used for training of the CNNs and 65 images were used as the test data set. The size of each image was 50$\times$50 pixels and the intensity of the images was normalized between 0 and 1 by dividing each intensity by 255. To increase the accuracy even with the small number of images, k-fold cross validation was used. A training data set with 240 images was randomly divided into four groups (60 images per group), and four CNNs with the same structure were trained by using three groups (180 images) as training data sets and one group (60 images) as a validation data set (Fig. \[fig3\]). In each training of CNN, the group of validation data was changed. Each training was conducted up to 40 epochs, and the number of steps at one epoch to update the weights of the CNNs was set at 20 \[8\]. Within 40 epochs, weights resulting in the minimum loss function of the validation data set were conserved. Because the number of training images was not sufficient, the batch size was set at 180 and the training images were increased by rotating between -180$^{\circ}$ and +180$^{\circ}$ and by shifting up to 10% of the width and height of the images \[8\]. After trainings of the CNNs, the accuracy of the test data set was calculated from the mean probability by using the conserved weights of the four CNNs (Fig. \[fig4\]). An image of the lunar maria is shown in Fig. \[fig4\]. The lunar maria are represented in black color and other effects such as the round shape of the moon itself and the color variations in the maria are removed because the purpose of this work was to evaluate the probability that the CNN can recognize and categorize the lunar maria pattern into the three animals. Results ======= For up to 40 epochs, each CNN recognized the validation images with 92%-98% accuracy. By using the trained weights and the mean probabilities of the four networks, the CNNs classified 61 of 65 test images correctly, and thus the accuracy was approximately 93.8%. The probabilities of a lunar image categorized into the three animals are shown in Fig. \[fig5\]. Each image of the lunar maria was tested four times by rotating the image four times by 90 degrees. At every rotation angle, the probability of categorization of the lion was the highest, and the CNN recognized that the lunar pattern was not the shape of the crab. In the case of the lion, the dark band on the moon corresponding to Mare Frigoris \[1\] is regarded as the tail of the lion (Fig. \[fig1\]). The shapes of the crab and the hare do not contain the band of Mare Frigoris. Fig. \[fig6\] shows the probabilities of the lunar image without Mare Frigoris. The probabilities to the hare become highest when the dark band of Mare Frigoris is not included in the lunar pattern. Most silhouette images of the lion contain the lion’s tail (Fig. \[fig3\]). Thus, when the band is removed, the CNN concludes that the lunar image is the shape of the hare rather than that of the lion. In addition to the increased probabilities of the hare, the probabilities of categorizing the crab decreased drastically to almost zero (Fig. \[fig6\]). The images of the crab used in this work contain legs as the primary shape of the crab (Fig. \[fig3\]). In the case of the lunar pattern with Mare Frigoris, the CNN might recognize a crab by its claw legs, and so the probability decreases for the lunar image without Mare Frigoris. Thus, the reason why the probabilities of categorizing the crab is the lowest in both lunar images is that the CNN learned the claw legs as the characteristic shape of the crab during training and the lunar images did not have sufficient areas related to the these legs. Conclusion ========== The probability of recognizing and categorizing the lunar pattern into three animals (crab, lion, hare) depends on which part of the lunar maria is included in the image. Of the three animals, the pattern of the lunar maria is recognized as the shape of the lion if the dark band of Mare Frigoris is included in the lunar pattern. On the other hand, the probability of recognizing the hare becomes the highest for the pattern without the band, and the probability of recognizing the crab is almost zero because the CNN recognize the area of Mare Frigoris as part of the crab’s leg. The CNN classifies the lunar pattern from the characteristic shapes of the training images. Thus, if the lunar pattern is different from the shapes of the training images, the probability of recognition becomes low, as in the case of the crab in this work. This process is different from recognition by humans. Although people as well as CNN recognize the similarity of shapes, we can also relate different objects by their character (this may be known as “hierophany”, shown in the work of Mircea Eliade). This character cannot be derived from the shape itself. For example, the moon, the crab and the hare have been symbols of immortality, rebirth and fertility in many areas of the world \[9\], and thus the moon pattern is recognized as such animals via these symbolic characters. These characters are induced from the behaviors of the animals and the moon rather than from their shapes. Myths also influence the relationship between different objects, as shown by the “Jataka” tales in Buddhism \[4\]. In areas where lions do not live, it is difficult for people to connect a lion to the moon. Thus, the environment of a society is also important. If the culture and the environment of each society influence our recognition of the lunar pattern, we may decide what we see in the moon in advance, and then we may find the similarities of shapes by adjusting the lunar pattern (e.g., by removing the area of Mare Frigoris for the hare’s shape or seeing the crab’s shape without legs), which is opposite to the recognition mechanism of the CNN. However, we cannot connect objects that are completely different from the shape of the lunar pattern. Thus, feedback may exist between the similarity of shape and the influence of culture for the recognition of symbols (culture induces people to find similarity of some objects, and if they can find rough similarity, additional cultures such as myth are generated in the society, which strengthens the similarity unconsciously in people’s mind). Acknowledgements {#acknowledgements .unnumbered} ================ This work was supported by a JSPS Research Fellowship. I bought silhouette images from deposit photos (https://jp.depositphotos.com/), and downloaded free images from Silhouette AC (https://www.silhouette-ac.com/index.html), Illust AC (https://www.ac-illust.com), Clipart Library (http://clipart-library.com), Silhouette Desighn (http://kage-design.com/). Hiesinger, H., Head, J. W., Wolf, U., Jaumann, R., Neukum, G. 2003. Ages and stratigraphy of mare basalts in Oceanus Procellarum, Mare Nubium, Mare Cognitum, and Mare Insularum. J. Geophys. Res. 108, 5065. Aharonson, O., Goldreich, P., Sari, R. 2012. Why do we see the man in the Moon? Icarus 219, 241-243. “JAXA Space Education Center” https://edu.jaxa.jp/campaign/moon2014/ in Japanese. McByde, J. M. 1911. Brer rabbit in the fork-tales of the Negro and other races. The Sewanee Review 19, 185-206. Simard, P. Y., Steinkraus, D., Platt, J. C. 2003. Best practices for convolutional neural networks applied to visual document analysis. ICDAR 3, 958-962. Lawrence, S., Giles, C. L., Tsoi, A. C., Back, A. D. 1997. Face recognition: A convolutional neural-network approach. IEEE transactions on neural networks 8, 98-113. Palafox, L. F., Hamilton, C. W., Scheidt, S. P., Alvarez, A. M., 2017. Automated detection of geological landforms on Mars using convolutional neural networks. Comput. Geosci. 101, 48-56. “Keras documentation” https://keras.io/. Nevsky, N. 1971. Moon and immortality. The Toyo-Bunko. in Japanese. ![Surface of the moon and the animals seen on the lunar surface. The illustrations are drawn based on the webpage of JAXA \[3\].[]{data-label="fig1"}](fig1.pdf){width="10cm"} ![Schematic view of the convolutional neural network. The network has two convolutional layers with 32 and 64 kernels, respectively. The size of a kernel is 3$\times$3 and the width of the stride is set at 1. After the max pooling layer and two affine layers, the probabilities are evaluated. Two dropout layers are included in front of the affine layers.[]{data-label="fig2"}](fig2.pdf){width="17cm"} ![Examples of silhouette images of the animals and the schematic view of the CNN training. Training images are randomly divided into four groups (G1, G2, G3 and G4). Each CNN uses three groups as the training data sets and one group as the validation data set. The group of the validation data set is different for each CNN. Weights of CNNs are conserved when the loss function of the validation data set is the minimum value.[]{data-label="fig3"}](fig3.pdf){width="17cm"} ![Schematic view of the evaluation of probabilities of the lunar pattern. For the lunar images, the outline of the lunar maria is considered. Each CNN with the trained weights calculates the probabilities of the lunar image categorized as one of the three animals. The mean value of each probability is regarded as the final probability. []{data-label="fig4"}](fig4.pdf){width="15cm"} ![Probabilities of the lunar pattern with Mare Frigoris for the three animals. Images are tested from four angles. Probabilities are rounded to two significant digits.[]{data-label="fig5"}](fig5.pdf){width="10cm"} ![Probabilities of the lunar pattern without Mare Frigoris for the three animals. Images are tested from four angles. Probabilities are rounded to two significant digits.[]{data-label="fig6"}](fig6.pdf){width="10cm"}	{ "pile_set_name": "ArXiv" }
--- abstract: 'Quantum Field Theory (QFT) is the basis of some of the most fundamental theories in modern physics, but it is not an easy subject to learn. In the present article we intend to pave the way from quantum mechanics to QFT for students at early graduate or advanced undergraduate level. More specifically, we propose a new way of visualizing the wave function $\Psi$ of a linear chain of interacting quantum harmonic oscillators, which can be seen as a model for a simple one-dimensional bosonic quantum field. The main idea is to draw randomly chosen classical states of the chain superimposed upon each other and use a [color]{} scale to represent the value of $\Psi$ at the corresponding coordinates of the quantized system. Our goal is to establish a better intuitive understanding of the mathematical objects underlying quantum field theories and solid state physics.' author: - Helmut Linde title: A New Way of Visualizing Quantum Fields --- This is an author-created, un-copyedited version of an article accepted for publication/published in European Journal of Physics. IOP Publishing Ltd is not responsible for any errors or omissions in this version of the manuscript or any version derived from it. The Version of Record is available online at https://doi.org/10.1088/1361-6404/aaa032. Introduction ============ Quantum Field Theories (QFT) are celebrated for being the framework of the Standard Model and for making predictions which coincide with experimental data to extreme accuracy. Yet QFT is perceived as difficult to learn by many and this seems to be at least partially due to the lack of visualization options which would help to develop an intuition for the theory. In quantum mechanics (or ‘first quantization’), for example, a localized particle is associated with a $\delta$-function and a uniformly moving particle with a plane wave, both of which are relatively easy to depict in a graph. Many textbooks show the energy eigenstates of hydrogen atoms, harmonic oscillators or quantum wells as plotted wave functions. A survey of visualization methods in quantum mechanics can be found in Thaller’s books [@Thaller]$^,$[@Thaller2]. In QFT, on the other hand, visualizations of the fundamental concepts seem to be scarce. One prominent exception are the Feynman diagrams, which are an important tool in studying processes of interacting particles. Yet these diagrams are firmly connected to the specific computational method of perturbation theory and they are not meant as a graphical representation of the quantum field itself. Another approach is to compute and plot statistical quantities, like the particle density or the two point correlation function of a many body system. But a lot of information is lost when moving from the full description of its state to these aggregated quantities. Therefore they tend to be more useful to understand macroscopic properties of the system rather than the underlying details. The difficulties in visualizing a quantum field obviously stem from the fact that such a system has too many degrees of freedom: Even if the field is simplified to a discrete lattice with $N$ atoms oscillating in only one space dimension each (like, e.g., in the ‘quantized mattress’ model in Zee’s book[@Zee]), the quantum state is given by a wave function $\Psi: {\mathbb{R}}^N \rightarrow {\mathbb{C}}$, which cannot be plotted in a straight-forward way for any $N > 2$. Such lattices of coupled quantum oscillators are also studied in several other areas of physics and especially in solid state physics they are applied to model phonons in a crystal. Johnson and Gutierrez[@JohnsonGutierrez] visualize phonon states of a one-dimensional quantum lattice by projecting the probability density of the system to each of the one-dimensional spaces in which the atoms oscillate. In the present article we propose a new way of visualizing the state of a bosonic quantum field and apply it to the example of the harmonic chain, which is essentially a discretized, one-dimensional model for a boson field. The target audience for this material are mainly graduate students who completed the courses on quantum mechanics and who are now about to advance towards quantum field theory. We assume that the reader is familiar especially with the treatment of the quantum harmonic oscillator and base changes in Hilbert space via Fourier transformation. The article is organized as follows: In Section \[SecMonteCarlo\] we present a Monte Carlo method for visualizing quantum wave functions in (first quantization) quantum mechanics. Then we introduce a quantum harmonic chain in Section \[SecLinHarChain\] and we compute its dynamics by diagonalizing its Hamiltonian via a Fourier transformation. In Section \[ConnectingDots\] we explain how the harmonic chain can be interpreted in the pictures of solid state physics (where each atom in the chain is a particle) and in quantum field theory (where a particle is an excitation of the chain). Setion \[SecVisualizationOfChain\] contains our main result - a new visualization method for the state of a quantum harmonic chain and the application to several interesting special cases. We conclude with a discussion of our findings and an outlook to future research in Section \[SecDiscussion\]. Monte Carlo plot of wave functions {#SecMonteCarlo} ================================== De Aquino et. al. have proposed the following method to visualize the wave function $\Psi: {\mathbb{R}}^2 \rightarrow {\mathbb{C}}$ of a single quantum particle in two dimensions[@DeAquinoEtAl]: Associate ${\mathbb{R}}^2$ with the two orthogonal axes of a scatter plot and then draw many dots into this plane, with the location of each dot being randomly chosen according to the probability density function $\|\Psi\|^2$. The resulting charts show a high density of dots where $\|\Psi\|^2$ is relatively large and only very few dots where $\Psi$ is close to zero. In our visualization method we follow a similar Monte Carlo approach, but we impose the following rules: 1. Choose the locations of the dots randomly according to a uniform probability distribution in a rectangular window within ${\mathbb{R}}^2$. 2. Represent the value of $\Psi$ at each dot via its color. For energy eigenstates, which can always be written as real-valued functions, such a visualization is possible on a black and white scale. For complex-valued functions a color spectrum can be used. To start with a simple example from ordinary quantum mechanics, consider a two-dimensional harmonic quantum oscillator with the Hamiltonian $$\hat H = \frac{\hat{p_1}^2}{2m}+\frac{\hat{p_2}^2}{2m} + \frac{1}{2} \kappa (\hat{q_1}^2 + \hat{q_2}^2),$$ where $q_l$ with $l \in \{1;2\}$ are the position coordinates of the particle with mass $m$, $\hat{p_l} = -i \partial/\partial q_l$ are the momentum operators, and $\kappa$ controls the strength of the attractive potential. Here and in the rest of the article we use units of measure such that $\hbar = 1$. It is well known[@Liboff] that the Hamiltonian is separable and the (non-normalized) energy eigenstates of the system can be written as $$\Psi_{\nu_1,\nu_2} = \prod_{l=1,2}{H_{\nu_l}(\sqrt{m\omega}\,q_l)e^{-m\omega q_l^2/2}},$$ where $\omega^2 = \kappa/m$, $H_\nu$ is the $\nu$-th order Hermite polynomial and $\nu_{1,2} \in {\mathbb{N}}_0$ are the two quantum numbers which enumerate the energy eigenstates of the system. As an example, Fig. \[2DHarmonic\] shows a Monte Carlo plot of the energy state $\Psi_{2,1}$ created according to the two rules above. [Blue (red)]{} dots represent negative (positive) values of $\Psi_{2,1}$ and dots where $\Psi_{2,1} \approx 0$ blend in with the chart’s [white]{} background. We have omitted a quantitative color scale since the normalization of the wave function is not important. It is easy to recognize the general shape of $\Psi_{2,1}$ and its nodal lines. Obviously such a Monte Carlo plot is only of limited relevance in two dimensions, since we could have simply colored the whole chart area according to the value of $\Psi_{2,1}$ instead of just picking a few points, or we could have drawn the wave function in a 3D plot. We will see later that such a Monte Carlo visualization actually becomes quite useful when moving beyond two dimensions. [![Monte Carlo representation of the two-dimensional harmonic oscillator energy eigenstate $\Psi_{2,1}$. [Red (blue)]{} dots correspond to positive (negative) values of the wave function. Horizontal and vertical nodal lines of the function can be recognized as areas without any visible dots. As expected, the state $\Psi_{2,1}$ has two nodal lines intersecting with the $q_1$ axis and one intersecting with the $q_2$ axis.[]{data-label="2DHarmonic"}](2DHarmonic-COL.eps "fig:"){width="180mm"}]{} Quantized Linear Harmonic Chain {#SecLinHarChain} =============================== Consider a linear chain of $N$ coupled harmonic oscillators whose positions are defined by the coordinates $q_n$ with $n = 1,...,N$, respectively. In classical mechanics this system’s Hamiltonian is $$\label{EqClassicalH} H = \sum_{n=1}^N \left(\frac{p_n^2}{2m} + \frac{1}{2} \kappa q_n^2 + \frac{1}{2} \gamma (q_n - q_{n+1})^2\right),$$ where $p_n = m \dot q_n$. $\kappa > 0$ determines the strength of binding between the oscillating masses and their respective neutral positions, while $\gamma > 0$ controls the coupling between neighbours. For reasons of simplicity we impose periodic boundary conditions, i.e. $q_0$ is the same as $q_N$ and $q_{N+1}$ means just $q_1$. This system is the one-dimensional version of the ‘mattress’ that Zee[@Zee] uses to introduce quantum field theory, or it can be seen as a simple model for phonons traveling through a crystal lattice. We shall now follow the standard procedure: We will decompose the system’s dynamics into normal modes which can be excited and evolve in time independently of one another. Mathematically speaking, this means to diagonalize the Hamiltonian $H$ via a Fourier transformation, revealing the system’s equivalence to $N$ decoupled harmonic oscillators. One option is to expand $\vec q$ in real-valued sine and cosine modes, like in the article of Johnson and Gutierrez[@JohnsonGutierrez]. Alternatively, a complex-valued decomposition into modes of $e^{ikn}$ can be applied like in the book of Greiner[@Greiner] on p.18. We shall also follow the latter approach and we will exhibit the technical steps in detail in order to make this article as self-contained as possible. We will deviate slightly from Greiner’s path since we also need to find (among other things) the explicit coordinate transformation between the harmonic chain and the decoupled oscillators (eq. \[EqQtildeq\] below). We write the Hamiltonian (\[EqClassicalH\]) as $$H = \frac{\|\vec{p}\|^2}{2m} + \frac{1}{2} {\vec{q}}^T D \vec{q},$$ where $D$ is an $N \times N$ matrix taking care of the terms with $\kappa$ and $\gamma$ in (\[EqClassicalH\]): $$D = \begin{pmatrix} \kappa+2\gamma & -\gamma & 0 & 0 & \dots & 0 & -\gamma\\ -\gamma & \kappa+2\gamma & -\gamma & 0 & \dots &0 & 0\\ 0 & -\gamma & \kappa+2\gamma & -\gamma & \dots &0 & 0\\ \dots & \dots & \dots & \dots & \dots &\dots & \dots \end{pmatrix}$$ The matrix $D$ is a circulant, which means that each row vector is rotated one element to the right compared to the preceding row vector. In this sense $D$ is generated from the $N$-vector $(\kappa + 2\gamma, -\gamma, 0, 0, ..., 0 , -\gamma)$ which forms the top line of the matrix. Now remember that any circulant matrix can be diagonalized via a discrete Fourier transform[@Davis] which, in turn, means to expand the position coordinates $q_n$ with respect to the orthonormal basis $\{\vec{f^{(k)}}\}$ with $$\label{EqONB} f^{(k)}_n = \frac{1}{\sqrt N} e^{-\frac{2\pi i}{N}k n} \quad \textrm{for} \quad k = -\frac{(N-1)}{2}, \dots, \frac{(N-1)}{2}.$$ The vectors $\vec{f^{(k)}}$ are also known as the ‘normal modes’ of the system. We have made the assumption that $N$ is an odd number in order to save us a cumbersome treatment of special cases. Since $N$ is considered to be a large number in all practically relevant cases, this is no real restriction. We call the position coordinates in the new basis $Q_k$ and they are related to the old ones by $$\label{EqFourierDecomp} q_n = \sum_k Q_k f^{(k)}_n = \frac{1}{\sqrt N} \sum_k Q_k e^{-\frac{2\pi i}{N}k n}.$$ Here and in the rest of the article a sum over $k$ always means that $k$ runs from $-(N-1)/2$ to $+(N-1)/2$ unless explicitly stated otherwise. Note that the $Q_k$ are complex numbers. In order to ensure that the coordinates $q_n$ remain real-valued, we have to impose the constraint $$\label{EqConstraint} Q_k = \overline{Q_{-k}}.$$ An analogous transformation maps the momentum $\vec p$ to the momentum $\vec P$ with $P_k = m \dot Q_k$. In the new coordinates the Hamiltonian takes the form $$\label{EqHnew} H = \sum_k \left(\frac{\|P_k\|^2}{2m} + \frac{1}{2} \omega_k \|Q_k\|^2 \right).$$ Here we have used the fact that our transformation was essentially a discrete Fourier transformation and it diagonalized the circulant matrix $D$. This implies that the eigenvectors of $D$ are the base vectors $\vec{f^{(k)}}$. We have chosen to call the respective eigenvalues $\omega_k$. Since $D$ is real symmetric, the $\omega_k$ are real. We have $$\omega_k \vec{f^{(k)}} = D \vec{f^{(k)}} = D \overline{\vec{f^{(-k)}}} = \overline{D \vec{f^{(-k)}}} = \overline{\omega_{-k} \vec{f^{(-k)}}} = \omega_{-k} \vec{f^{(k)}}$$ and thus $\omega_k=\omega_{-k}$. In order to get rid of the complex coordinates before doing a canonical quantization, we perform a second coordinate transformation, splitting up each $Q_k$ into two real-valued coordinates: $$\label{EqQtilde} Q_k = \frac 12 (1+i) \tilde Q_k + \frac 12 (1-i) \tilde Q_{-k} \quad\textrm{with } \tilde Q_k \in {\mathbb{R}}.$$ The constraint (\[EqConstraint\]) is automatically observed then. By putting (\[EqQtilde\]) into (\[EqFourierDecomp\]) we get the transformation rule between the $q_n$ and the $\tilde Q_k$: $$\label{EqQtildeq} q_n = \sum_k \tilde Q_k \frac{1}{2} \left((1+i) f^{(k)}_n+(1-i)f^{(-k)}_n\right) =: \sum_k \tilde Q_k \tilde f^{(k)}_n$$ We note that the $\{\tilde f^{(k)}_n\}$, in which we have developed $\vec q$, form another orthonormal basis in ${\mathbb{R}}^N$. This can be verified by direct computation, using the fact that the $\{f^{(k)}_n\}$ are also orthonormal. The system’s Hamiltonian (\[EqHnew\]) looks very similar in the new coordinates:\ $$\label{EqHnew2} H = \sum_k \left(\frac{\tilde P_k^2}{2m} + \frac{1}{2} \omega_k \tilde Q_k^2\right),$$ where $\tilde P_k = m \partial/\partial_t \tilde Q_k$. $H$ is now separated into $N$ one-dimensional harmonic oscillators with the eigenfrequencies $\omega_k$. Next, we perform a canonical quantization on $H$, promoting the location and momentum coordinates to Hilbert space operators. As a result we get a set of $N$ decoupled quantum oscillators. Solving the harmonic oscillator is one of the most common introductory problems in quantum mechanics and it is presented in many textbooks[@Liboff]. It can be achieved either by solving the oscillator’s differential (Schroedinger) equation, or in an algebraic way based on commutation relations. Following the latter approach, we introduce the usual creation/annihilation operators $a^\dagger_k$ and $a_k$, so that the quantum version of (\[EqHnew2\]) becomes $$\label{EqHnew3} \hat H = \sum_k \omega_k \left(a^\dagger_k a_k + \frac 12 \right).$$ The excitations of the harmonic chain created by the $a_k^\dagger$ are also called ‘particles’ since this is how they often manifest themselves in experiments. Connecting the dots {#ConnectingDots} =================== By now we have established all the machinery needed to advance to the main ideas of this article, the Monte Carlo visualization of bosonic quantum fields. But let’s first step back and review how our work so far is related to first and second quantization. This section is meant to remind ourselves of the ‘big picture’ and provide some orientation to novices in the field. It can be omitted by the advanced reader. We had started with one quantum particle which was free to move in two dimensions $q_1$ and $q_2$ subject to an harmonic potential. As known from introductory quantum mechanics courses, the particle’s wave function (or, more precisely, the square of its absolute value) tells us the probability density of finding the particle at a given point in space when performing a measurement. We have visualized such a wave function for the specific eigenstate $\Psi_{2,1}$ in Fig. \[2DHarmonic\]. The coordinates $q_1$ and $q_2$ are linked to location operators $\hat q_1$ and $\hat q_2$ which in turn describe a quantum measurement of the particle’s position in the respective coordinate. Then we have shown how to solve the dynamics of the harmonic oscillator chain, which can serve for at least two physical models: On the one hand, we can think of each point in the chain as being an actual physical particle like the atoms in a crystal lattice. In this case $q_n$ is still the spacial coordinate of each mass point relative to its equilibrium position, just like in the two dimensional case before, and $n$ simply enumerates the atoms of the chain. We are still in the ‘first quantization world’ and if we were to measure the exact positions of the $N$ particles at the same time, the wave function of the system would tell us the probability of finding the chain is a certain configuration. The $\hat q_n$ remain simple location operators which describe a position measurement of the $n$th particle. For example, if the system’s normalized state is $\Psi$ and we measure the position of the third particle, then the expectation value of the result is $\langle \Psi, \hat q_3 \Psi \rangle$. The normal modes of oscillation in the chain, which we have identified in Section \[SecLinHarChain\], are called phonons - an important concept in solid state physics. On the other hand, the chain can serve as a model for a bosonic quantum field, which belongs to the realm of ‘second quantization’. In this case $n$ becomes the discrete version of a spacial coordinate and $q_n$ is the field strength at the point $n$. The wave function of the system would give us the probability to encounter the field in a given classical state if it were possible to measure the field strength at all points in space at the same time. The operators $\hat q_n$ are now called ‘field operators’ and they correspond to a measurement of the field strength in the spacial point $n$. In quantum field theory one would typically proceed to the continuum limit, replacing $n$ by $\vec x \in {\mathbb{R}}^3$ and $\hat q_n$ by some $\hat\phi(\vec x)$, for example. So all we have is a field, but where are particles - the bosons - which we wanted to describe? It turns out that what we call ‘particles’ are simply excitation modes of the quantum field. Such excitations can be limited to a region in space and they can travel through the field very similarly to how particles in first quantization can be more or less localized and move in space. But, in contrast to first quantization, excitations can also become stronger or weaker - which means that particles can be created or destroyed. In the following section we shall see some examples of such excitations, how they can be visualized, and how they correspond to physical particles. Visualization of the Quantum Harmonic Chain {#SecVisualizationOfChain} =========================================== Visualization Method -------------------- We have noted in Section \[SecLinHarChain\] that the Hamiltonian of the quantum harmonic chain of length $N$ is equivalent to the $N$-dimensional quantum harmonic oscillator (\[EqHnew2\]) up to a coordinate transformation. The latter has the convenient property of being separable, so that its energy eigenstates are simply products of the familiar one-dimensional harmonic oscillator states: $$\Psi_{\{\nu_k\}} = \prod_{k}{H_{\nu_k}(\sqrt{m\omega_k}\,\tilde Q_k)e^{-m\omega_k\tilde Q_k^2/2}},$$ where $H_\nu$ is the $\nu$-th order Hermite polynomial and $\nu_k \in {\mathbb{N}}_0$ for $k = -(N-1)/2,\dots,(N-1)/2$ are the quantum numbers. Any wave function $\Psi$ which is given in terms of the coordinates $\tilde Q_k$ can be transformed back to the original coordinates $q_k$ via (\[EqQtildeq\]). Once $\Psi(\vec q)$ is known, we can apply the Monte Carlo visualization which we have introduced in section \[SecMonteCarlo\]. Compared to our simple example in Fig. \[2DHarmonic\], the main difference is that we now have $N$ coordinates instead of only two. Therefore we have to adapt our visualization prescription as follows: 1. Choose points $\vec q$ randomly according to a uniform probability distribution in a rectangular window within ${\mathbb{R}}^N$. 2. Each point $\vec q$ is visualized as a polyline in a parallel axes plot and the value of $\Psi(\vec q)$ is represented by the color of the polyline. The visualization in a parallel axes diagram has the advantage that each point $\vec q$ is represented by a polyline which can be intuitively associated with the corresponding state of a classical (i.e., non-quantum) oscillator chain. In order to plot such visualizations we have developed a small program in the numerically oriented programming language Scilab[@Scilab]. The source code is available on request from the author of this article. Ground State ------------ As a first example of the results, Fig. \[GroundState\] represents the ground state $\Psi_0 = \Psi_{\{\nu_k=0\,\forall k\}}$ of a quantum chain consisting of $N = 15$ oscillators. In the language of QFT, this state is called the ‘vacuum’ since no excitations (=particles) are present. The wave function of the ground state is unique only up to multiplication with a complex number and in this case it has been chosen such that $\Psi_0$ is real and positive. Each line in the plot represents one possible configuration of the 15 oscillators’ positions. The [color]{} of each line corresponds to the value of the wave function $\Psi_0$ for the respective configuration. The color scale is chosen such that the value $\Psi_0 = 0$ blends in with the [white]{} background and the [more colorful]{} the line, the higher the value of $\Psi_0$. Each polyline is plotted on top of the lines with lower $\|\Psi_0\|$, so that the most important lines (i.e. those with high $\|\Psi_0\|$) are more clearly visible in the chart. Those lines represent the most probable configurations of the chain in the sense of a quantum mechanical measurement and they tend to be relatively close to the $\vec q = 0$ line, which is the classical equilibrium state of the system. Particles at rest ----------------- As further examples we consider excitations of the chain which can be interpreted as one or two particles at rest, respectively. The one-particle state $\Psi = a_0^\dagger \Psi_0$ is visualized in Fig. \[OnePartRest\] and the two-particle state $\Psi = (a_0^\dagger)^2 \Psi_0$ in Fig. \[TwoPartRest\]. As before, the color scale is chosen such that configurations with $\Psi = 0$ blend in invisibly with the [white]{} background, and the lines [in deepest red/blue ]{}correspond to the largest (positive) / smallest (negative) values of $\Psi$. Similar to the ground state, the configurations with the highest absolute value of $\Psi$ tend to be those with little fluctuation along the $n$-axis. But in the one-particle case we note that the polylines with mostly positive (negative) $q_n$ tend to correspond to positive (negative) values of $\Psi$. In the two-particle case the polylines close to the $\vec q = 0$ line tend to correspond to negative values of $\Psi$ while those farther away from that line are rather associated with positive values of $\Psi$. This pattern obviously resembles the positive and negative parts of the energy eigenfunctions of a one-dimensional harmonic oscillator. The classical analog to Fig. \[OnePartRest\] and Fig. \[TwoPartRest\] is a chain whose points masses oscillate synchronously and a stronger excitation of the chain as a whole corresponds to a higher number of particles in the quantum field. Particle with Non-Zero Wave Number ---------------------------------- Now we turn to excitations of the chain with non-zero wave number. For example, the state $\Psi = a_1^\dagger \Psi_0$ (which we can also write as $\Psi_{\{\nu_k\}}$ where $\nu_k$ is equal to $1$ for $k=1$ and zero else) is represented in Fig. \[OnePartKOne\] for $N=15$. Its classical analog is a standing wave in the oscillator chain. The location of the nodes near $n = 6$ and $n = 13$ can be explained by transforming back to the $q_n$ coordinates: In our state $a_1^\dagger \Psi_0$ the oscillating system is excited along the $\tilde Q_1$ coordinate, since this is how we defined the creation operators in (\[EqHnew3\]). Using (\[EqQtildeq\]) we see that an oscillation along $\tilde Q_1$ transforms back to an oscillation along the vector $\vec q$ with $$q_n = \frac{1}{2}\left((1+i) f_n^{(1)} + (1-i) f_n^{(-1)}\right)= \frac{1}{\sqrt{N}}\left(\cos\frac{2\pi n}{N} + \sin \frac{2\pi n}{N}\right),$$ which is zero for $n = 3N/8 \approx 5.6$ or for $n = 7N/8 \approx 13.1$. A graph of the state $a_{-1}^\dagger \Psi_0$ would look very similar to Fig. \[OnePartKOne\] with the exception that the standing wave is shifted by an offset of $N/4$ along the $n$-axis compared to the state $a_1^\dagger \Psi_0$. In the particle language of excitations $a_1^\dagger \Psi_0$ and $a_{-1}^\dagger \Psi_0$ would both be one-particle states, which is clear from how it is constructed via one creation operator. We can also construct the quantum analog to a progressive wave in the chain, for example the state $(a_1^\dagger + i a_{-1}^\dagger) \Psi_0$. But the resulting wave function would not be real-valued anymore. A complex-valued $\Psi$ can be visualized with our method by using different colors, but for the present article we limit ourselves to real-valued $\Psi$ functions which can be visualized [with two colors only]{}. Localized Particle ------------------ Our next example is a state with one spatially localized particle. Recall that we had expressed $\vec q$ in the orthonormal basis $\{\tilde f^{(k)}_n\}$ in (\[EqQtildeq\]) and note that the function $\delta_{n,n_0}$ (which is equal to one for $n=n_0$ and else zero) can be expanded in this basis as $$\label{EqDelta} \delta_{n,n_0} = \sum_k \langle \tilde f^{(k)}_\cdot ,\delta_{\cdot,n_0} \rangle \tilde f^{(k)}_n = \sum_k \tilde f^{(k)}_{n_0} \tilde f^{(k)}_n.$$ Remember that our operators $a^\dagger_k$ create a particle ‘in the state’ $\tilde f^{(k)}_n$. Thus, in analogy to (\[EqDelta\]) we construct new creation operators $$\label{EqCreatorsLocalized} b^\dagger_n := \sum_k \tilde f^{(k)}_n a^\dagger_{k}$$ and we expect $b^\dagger_n$ to create a particle localized in space at the position $n$. Fig. \[OnePartLocalized\] is a visualization of the state $$\label{EqPsiLoc} \Psi = b^\dagger_{5} \Psi_0.$$ It is apparent that the sign of $\Psi(\vec q)$ depends only on the sign of the single coordinate $q_5$. This is exactly what one should expect from a quantum oscillator which has $N = 11$ (in the case of Fig. \[OnePartLocalized\]) degrees of freedom, but which is excited only along the fifth coordinate. Yet the effect of the coupling between the point masses is also visible: At least at the positions $n=4$ and $n=6$ there is a notable correlation of the polylines’ [color]{} with those at $n=5$ in the sense that [red (blue)]{} lines dominate where $q > 0$ $(q<0)$. It looks like the excitation of the quantum field is ‘blurred’ around that point. Where does this blur come from, given that $\Psi$ in (\[EqPsiLoc\]) is constructed in close analogy to the $\delta$ function (\[EqDelta\]) with its sharp peak? The deeper reason is that $a^\dagger_{k} \Psi_0$ is not only associated with a wave $\tilde f^{(k)}_n$, which stems from the Fourier transformation and which gives rise to the analogy between (\[EqDelta\]) and (\[EqPsiLoc\]), but also with an oscillation amplitude in the direction of the $\tilde Q_k$ coordinate. This amplitude depends on the strength of the harmonic potential in the $\tilde Q_k$ direction, which in turn depends (via the diagonalization of the system’s Hamiltonian) on the coupling constant $\gamma$ between the mass points. Only in the trivial case $\gamma = 0$ the potential is radially symmetric in the space spanned by the $q_n$ or the $\tilde Q_k$ coordinates, which implies that the oscillation amplitudes of excited states don’t depend on $n$ or $k$, so that the analogy between (\[EqDelta\]) and (\[EqPsiLoc\]) perfectly holds. In this case the ‘blurring’ of the particle’s position in Fig. \[OnePartLocalized\] vanishes, which is also clear from the fact that for $\gamma = 0$ there is not even the concept of two ‘neighboring’ point masses built into the physical system. Two Localized Particles ----------------------- In our final example we consider two localized particles, which are described by the state $$\label{EqPsi2Loc} \Psi = b^\dagger_{n_1} b^\dagger_{n_2} \Psi_0.$$ This case is particularly interesting because we can visualize how two bosons move gradually from two different states into one and the same. The Figures \[TwoPartLocalized-large-dist\], \[TwoPartLocalized-small-dist\] and \[TwoPartLocalized-no-dist\] contain Monte-Carlo visualizations for different values of the pair $(n_1,n_2)$, namely for $(3,8)$, $(5,6)$ and $(5,5)$, respectively. The two particles in Fig \[TwoPartLocalized-large-dist\] are relatively far apart and each of the two localized excitations resemble the localized one-particle state from Fig. \[OnePartLocalized\], even though the phase structure of the $\Psi$ function (i.e. [red]{} lines vs. [blue]{} lines) is much more complicated. In Fig. \[TwoPartLocalized-no-dist\], on the other hands, two particles are located at the same point in space. Note the similarity of this plot with the two-particle state shown in Fig. \[TwoPartRest\]: Both exhibit the pattern of an harmonic oscillator’s second excited state (a function with two changes of sign), but in this case the excitation is limited to a small neighborhood of the point $n=5$. Fig. \[TwoPartLocalized-small-dist\] shows how the transition between the two extreme cases (two particles far apart vs. localized in the same point) comes about. Discussion of Results {#SecDiscussion} ===================== We have presented a new way of visualizing states of a quantum field and we conclude this article with a discussion of advantages and limitations of the new method, as well as an outlook for future investigation. We hope that our visualizations can help students of QFT and solid state physics to develop a better intuitive understanding of some of their fundamental concepts. Compared to the method presented by Johnson and Gutierrez [@JohnsonGutierrez] we feel that our graphs are a slightly more direct representation of the quantum state, since they indicate the value of $\Psi$ for (a randomly chosen set of) points in the state space directly instead of showing a projection of $\Psi$ to the $q_n$ coordinates. For example, the spatial ‘blurriness’ of a localized particle is clearly visible in Fig. \[OnePartLocalized\] but not in the analogous Fig. 23 of Johnson’s and Gutierrez’ paper. An inherent limitation of our method is that the graphs are not exactly reproducible. Due to the random nature of the Monte-Carlo approach two graphs generated with the exact same parameters will look similar, but not exactly the same. Also, the graphs are quite sensitive to parameters chosen in the Monte-Carlo algorithm. For example, if the number of random points chosen is too small, then none of these points might be in a region where $\|\Psi\|$ is large and the resulting graph will look quite noisy and without any clear structure. Whether that number is ’small’ also depends on the number $N$ of point masses. The volume from which interesting points can be chosen (e.g. the set of all $\vec q \in {\mathbb{R}}^N$ with $\|q_j\| < 1$ for all $j = 1,\dots,N$) grows exponentially with $N$. Consequently, the necessary number of random points and the computation time of the algorithm also grow exponentially with $N$. Finally, for too ‘complex’ states (e.g. those with more than just a few particles of different wave numbers) the graphs become too busy to be readable. In principle, our method could be extended to two-dimensional fields, where the polylines from the plots in the present paper would be replaced by overlapping surfaces in a 3D plot. In order to avoid a very busy and confusing graph, the number of surfaces drawn would have to be limited strictly to just a few with the very highest values of $\|\Psi\|$. An interesting subject for future work will be the extension of the method presented here to more complex phenomena in quantum fields, for example by considering interactions between fields or by making the shift from bosons to fermions. Acknowledgments =============== The author is grateful to the anonymous referees for thorough reviews and several helpful suggestions which have made this article better. [99]{} P.J. Davis: “Circulant Matrices” (Chelsea 1994) V.M. de Aquino, V.C. Aguilera-Navarro, M. Goto, H. Iwamoto: “Monte Carlo image representation”, Am. J. Phys., 69(7) (2001) R. Greiner: “Field Quantization” (Springer-Verlag Berlin Heidelberg 1996) S.C. Johnson, T.D. Gutierrez: “Visualizing the phonon wave function”, Am. J. Phys., 70(3) (2002) R.L. Liboff: “Introductory Quantum Mechanics” (Addison-Wesley Publishing Company, 1980) M. Ligare, R. Oliveri: “The calculated photon: Visualization of a quantum field”, Am. J. Phys., 70(1) (2002) B. Thaller: “Visual Quantum Mechanics: Selected Topics with Computer-Generated Animations of Quantum-Mechanical Phenomena” (Springer/TELOS, Berlin, 2000) B. Thaller: “Advanced Visual Quantum Mechanics” (Springer New York 2004) A. Zee: “Quantum Theory in a Nutshell” (Princeton University Press, 2010) Scilab Enterprises (2012). Scilab: Free and Open Source software for numerical computation (OS, Version 6.0.0) \[Software\]. Available from: http://www.scilab.org	{ "pile_set_name": "ArXiv" }
--- author: - Isaac Vikram Chenchiah - Anja Schlömerkemper bibliography: - 'References.bib' - 'Papers.bib' - 'Manuscripts.bib' title: 'Non-laminate Microstructures in Monoclinic-I Martensite' --- Introduction ============ Shape-memory alloys are materials that undergo a diffusionless solid-to-solid phase transformation due to change of temperature. They are capable of large macroscopic deformations and recover their original shape upon heating. While such materials usually form a cubic lattice (austenite) above a critical temperature they develop microstructures at lower temperatures (martensite). In this article we are interested in the cubic-to-monoclinic-I phase transformation, which occurs, eg., in NiTi, which is industrially one of the most important shape-memory alloys. In this case, there are twelve transformation strains, see Section \[sec:strains\] for details. Of interest is the set of all strains that can be recovered upon heating. In the variational approach to martensite [@Ball:1987] this set is modelled by the quasiconvex hull (Definition \[def:semi-convex\]) of the transformation strains. Unfortunately the quasiconvex hull of a set is difficult to calculate. Bhattacharya and Kohn [@Bhattacharya:1997p99] consider various phase transformations (cubic to tetragonal, cubic to trigonal, cubic to orthorhombic and cubic to monoclinic) in the context of geometrically-linear elasticity and observe that except for cubic-to-monoclinic martensite the symmetrised quasiconvex hull coincides with the convex hull. For cubic-to-monoclinic martensite they show that the symmetrised quasiconvex hull is strictly smaller than the convex hull and present an inner bound for it. We aim to find a better approximation of the symmetrised quasiconvex hull, again in the context of geometrically-linear elasticity. To this end, we are interested in the symmetrised lamination convex hull and the symmetrised rank-one convex hull of the transformation strains; these give inner bounds on the symmetrised quasiconvex hull, see Section \[sec:semi-convexity\]. Our analysis shows that there are points in the symmetrised rank-one convex hull of monoclinic-I martensite which are attained by non-laminate microstructures. This suggests (see below) that the symmetrised rank-one convex hull is strictly larger than the lamination convex hull of the twelve transformation strains. Since the symmetrised rank-one convex hull is a subset of the symmetrised quasiconvex hull (Remark \[rem:semi-convex\]), the strains attained by these non-laminate microstructures belong to the set of recoverable strains. Next we give more details on the strategy we use, our results and the organisation of the paper. Finally we will fix some notation. #### Strategy. Our strategy is to exploit firstly, the algebraic structure of symmetrised rank-one convex cones, secondly, the faceting structure of the convex hull of a finite set, and thirdly, the interaction between the two. That rank-one convex cones are varieties has been exploited to develop algorithms to calculate semi-convex hulls [@Kreiner:2003; @Kreiner:2004]. The relevance of convex polytopes to semi-convex hulls (compare Lemma \[lem:pairwise-compatible\] with Theorem \[thm:edge-compatible\]) has been noticed in [@Zhang:2006; @Tang:2008] but has not, to our knowledge, been exploited to determine semi-convex hulls. This paper represents a first attempt in this direction. The central idea is the following: Given a finite set whose symmetrised rank-one convex hull we wish to compute we proceed as follows. First we compute the symmetrised rank-one convex hulls of all its one-dimensional facets, i.e., edges in the language of convex polytopes; this is trivial, cf. below. We use this, together with knowledge of the structure of symmetrised rank-one convex cones on two-dimensional affine subspaces (Section \[sec:2D-cone\]), to determine the symmetrised rank-one convex hulls of all its two-dimensional facets. We then repeat this for higher dimensions. When the finite set we are interested in is the set of transformation strains of a material capable of a phase transformation from austenite to martensite, the set lies in the five-dimensional affine plane of strains with constant trace and thus the process above terminates when the symmetrised rank-one convex hull of the five-dimensional facet of the set (which is the convex hull of the set) is computed. In this bootstrapping strategy the steps become progressively more difficult as the dimension increases. Indeed while we can completely implement the two-dimensional step (Section \[sec:2D-cone\] and [@Chenchiah-Schloemerkemper-PRSL]) and have partial results for the three-dimensional step [@Chenchiah-Schloemerkemper-JMPS] we have a reason to believe that the steps for dimension four and five are considerably more difficult than that for steps two and three: unlike in lower dimensions, the symmetrised rank-one convex cone in the higher dimensions is an algebraic surface of a polynomial which is necessarily irreduccible [@Chenchiah-Schloemerkemper-PRSL]. #### Results. We have three main results: First we show that there are two kinds of monoclinic-I martensites which differ qualitatively with regard to the polytope-structure of their convex hulls. It follows that their semi-convex hulls are qualitatively different as well. Curiously all known monoclinic-I martensites belong to one of these kinds, which we name monoclinic-Ia (the other being monoclinic-Ib). The question as to whether $T_3$s (which are non-trivial symmetrised rank-one convex hulls of 3-tuples of pairwise incompatible strains, which are attained by non-laminate microstructures, see Section \[sec:T3\]) can be formed from the twelve variants of Monoclinic-II martensite is raised in [@Bhattacharya:1994p843 p863]. There it is shown that this is possible when a certain lattice parameter is sufficiently small. Here we prove a stronger result: In Section \[sec:T3\] we present a simple test for $T_3$s (Lemma \[lem:T3\]) that shows that in fact $T_3$s can form for all (non-zero) values of the lattice parameter, and that the same is true for Monoclinic-I martensite as well. This is our second main result. Our third result is a consequence of this: We show that for Monoclinic-Ia martensite, the symmetrised rank-one convex hull of the twelve transformation strains contains a five-dimensional continuum of points which are attained by non-laminate microstructures. Moreover the intersection of this continuum with the boundary of the convex hull is four-dimensional. This suggests that the symmetrised rank-one convex hull of the transformation strains is strictly larger than the symmetrised lamination convex hull. This would then imply that the symmetrised quasiconvex envelope of the energy density of this material is different from the symmetrised lamination convex envelope. It is well known that lamination convex envelopes can differ from rank-one convex envelopes, cf. [@Dacorogna:2007 Sect. 4], and (in dimensions larger than two) that rank-one convex envelopes can differ from quasiconvex envelopes [@Sverak:1992]. However, in the context of materials science, all quasiconvex and rank-one convex envelopes that have been evaluated so far have in fact coincided with lamination convex envelopes; Monoclinic-Ia martensite is the first material for which we now have a strong indication that they differ. #### Organisation of the paper. In Section \[sec:sroc\] we refresh the reader’s memory of some basic facts and results about strain compatibility, convex sets (in particular, convex polytopes) and semi-convex functions and sets. In Section \[sec:2D-cone\] we study the structure of symmetrised rank-one convex cones on two-dimensional affine subspaces of ${\ensuremath{{\mathcal S}^{3 \times 3}_{\text{\ensuremath{c}}} }}$ (Lemma \[lem:2D-compatible\_directions\]), see below for notation. The results presented here enable the computation of the symmetrised rank-one convex hull of any finite set in two-dimensional affine subspaces of ${\ensuremath{{\mathcal S}^{3 \times 3}_{\text{\ensuremath{c}}} }}$ and the characterisation of those compact sets in these spaces that possess non-trivial symmetrised rank-one convex hulls; we present some results in Section \[sec:T3\] but postpone a more extensive discussion to [@Chenchiah-Schloemerkemper-PRSL]. We extend these results to higher dimensions in [@Chenchiah-Schloemerkemper-JMPS] but Lemma \[lem:T3-3D\] and Section \[sec:nonlaminates\] provide a glimpse of the utility of the results in Section \[sec:T3\] even in higher dimensions. We then turn from abstract results to the specific class of materials of interest to us, monoclinic-I martensite. After some preliminary observations in Section \[sec:strains\] on the compatibility and symmetry relations between the twelve transformation strains of materials in this class we determine, in Section \[sec:polytope\], the facets of the convex hull of the twelve transformation strains of monoclinic-I martensite (Observations \[obs:e<d\], \[obs:e>d\] and \[obs:e=d\]). This leads to the discovery that there are in fact two kinds of monoclinic-I martensitic materials. With this foundation behind us, in Section \[sec:nonlaminates\] we investigate the (theoretical) possibility of non-laminate zero-energy microstructures occurring in these materials. We construct an open set in the symmetrised rank-one convex hull of the transformation strains for which $T_3$-microstructures are optimal (Construction \[cons:5D-set-T3s\]). We then deduce that in monoclinic-Ia martensite this set intersects the boundary of the convex hull, and thus the boundary of the symmetrised rank-one convex hull. (In Sections \[sec:polytope\] and \[sec:nonlaminates\] we use Mathematica to simplify computations but these computations are non-numerical. The Mathematica code that we have used together with explanatory notes can be found in the electronic supplementary material accompanying this article [@Chenchiah-Schloemerkemper1ESM].) We conclude with Section \[sec:conclusions\] with some questions raised by the preceding two sections. One of these is whether monoclinic-Ib martensite might have a larger set of recoverable strains (i.e., quasiconvex hull) than monoclinic-Ia martensite (modulo appropriate normalisation of the lattice parameters). This naturally also leads to the question as to whether a material that lies at the boundary of monoclinic-Ia and monoclinc-Ib martensite might demonstrate the best behaviour of all. #### Notation. In the geometrically linear theory of elasticity, the strains (pointwise) belong to the space of real symmetric $3\times 3$ matrices denoted by ${\ensuremath{{\mathcal S}^{3 \times 3}_{\text{\ensuremath{}}} }}$. We phrase some of our results for real symmetric $d\times d$ matrices and then use the symbol ${\ensuremath{{\mathcal S}^{d \times d}_{\text{\ensuremath{}}} }}$. The space of real symmetric $d \times d$ matrices whose trace is an arbitrary (but fixed) constant $c$ is denoted by ${\ensuremath{{\mathcal S}^{d \times d}_{\text{\ensuremath{c}}} }}$. We introduce an inner product $\langle \cdot, \cdot \rangle$ on ${\ensuremath{{\mathcal S}^{d \times d}_{\text{\ensuremath{}}} }}$ by ${\ensuremath{{\mathcal S}^{d \times d}_{\text{\ensuremath{}}} }} \ni A,B \mapsto \langle A, B \rangle = {\ensuremath{\operatorname{Tr}}}(AB)$; the norm induced by this inner product is $\\| \cdot \\|$. For $e_1,e_2 \in {\ensuremath{{\mathcal S}^{d \times d}_{\text{\ensuremath{}}} }}$ we set $$\begin{aligned} (e_1,e_2) &:= \{ \lambda e_1 + (1-\lambda) e_2\ \|\ \lambda \in (0,1) \}, \\ [e_1,e_2] &:= \{ \lambda e_1 + (1-\lambda) e_2\ \|\ \lambda \in [0,1] \}.\end{aligned}$$ By a direction in ${\ensuremath{{\mathcal S}^{d \times d}_{\text{\ensuremath{}}} }}$ we mean a one-dimensional affine subspace of ${\ensuremath{{\mathcal S}^{d \times d}_{\text{\ensuremath{}}} }}$. We denote the affine span of $S \subset {\ensuremath{{\mathcal S}^{d \times d}_{\text{\ensuremath{}}} }}$ by ${\ensuremath{\operatorname{aff}}}{\ensuremath{\operatorname{span}}}(S)$, the relative boundary of $S \subset {\ensuremath{{\mathcal S}^{d \times d}_{\text{\ensuremath{}}} }}$ by ${\ensuremath{\operatorname{rel}}}\partial S$ and its relative interior by ${\ensuremath{\operatorname{rel}}}{\ensuremath{\operatorname{int}}}{S}$. Preliminaries {#sec:sroc} ============= Strain compatibility {#sec:compatibility} -------------------- \[def:compatibility\] Let $e_1, e_2 \in {\ensuremath{{\mathcal S}^{3 \times 3}_{\text{\ensuremath{}}} }}$ and $S_1, S_2 \subset {\ensuremath{{\mathcal S}^{3 \times 3}_{\text{\ensuremath{}}} }}$. 1. $e_1$ is compatible with $e_2$ (or $e_1$ and $e_2$ are compatible), $e_1 {\ensuremath{\interleave}}e_2$, if there exist $a,b \in {\mathbb{R}}^3$ such that $e_1 - e_2 = \frac12 (a \otimes b + b \otimes a)$, where $\otimes$ denotes the dyadic product. 2. $e_1$ is compatible with 0 (or compatible for short) if $e_1 {\ensuremath{\interleave}}0$. 3. $e_1$ is incompatible with $e_2$ (or $e_1$ and $e_2$ are incompatible), $e_1 {\ensuremath{\cancel{\interleave}}}e_2$, if $e_1$ and $e_2$ are not compatible. 4. $e_1$ is incompatible with 0 (or incompatible for short) if $e_1 {\ensuremath{\cancel{\interleave}}}0$. 5. $S_1$ is totally compatible with $S_2$ (or $S_1$ and $S_2$ are totally compatible), $S_1 {\ensuremath{\interleave}}S_2$, if for all $e_1 \in S_1$ and for all $e_2 \in S_2$, $e_1 {\ensuremath{\interleave}}e_2$. \[it:local1\] 6. $S_1$ is compatible if $S_1 {\ensuremath{\interleave}}S_1$, i.e., if for all $e_1,e_2 \in S_1$, $e_1 {\ensuremath{\interleave}}e_2$. \[it:local9\] 7. $S_1$ is totally incompatible with $S_2$ (or $S_1$ and $S_2$ are totally incompatible), $S_1 {\ensuremath{\cancel{\interleave}}}S_2$, if for all $e_1 \in S_1$ and for all $e_2 \in S_2$, $e_1 {\ensuremath{\cancel{\interleave}}}e_2$. \[it:local2\] We observe that $[e_1,e_2]$ is compatible if and only if $e_1 {\ensuremath{\interleave}}e_2$. An alternative term for compatibility is “symmetrised rank-one connectedness”. In the figures compatible lines are represented by solid lines and incompatible lines by dashed lines, cf. eg. Figure \[fig:T3\]. The following lemma follows immediately from [@Kohn:1991p193 Lemma 4.1]. \[lem:SymTr3-compatibility\] Let $e_1, e_2 \in {\ensuremath{{\mathcal S}^{3 \times 3}_{\text{\ensuremath{c}}} }}$. Then $e_1 {\ensuremath{\interleave}}e_2$ iff $\det(e_1 - e_2) = 0$. Let $S \subset {\ensuremath{{\mathcal S}^{3 \times 3}_{\text{\ensuremath{c}}} }}$ and $x \in S$. The compatible cone in $S$ at $x$ is the set $$\Lambda_{S,x} := \{ y \in S\ \|\ y {\ensuremath{\interleave}}x \}.$$ When $0 \in S$ we set $\Lambda_S := \Lambda_{S,0}$. Note that the compatible cone $\Lambda_{S,x}$ is an affine cone with vertex $x$ in the linear algebraic sense, i.e., it is closed under multiplication by positive reals. However, the compatible cone is not geometrically a cone. Convex sets {#sec:convexity} ----------- We recall some elementary definitions and results from convex analysis. For more details see, eg. [@Rockafellar:1996]. Let $E$ be a subset of a vector space. We denote the convex hull of $E$ by ${\ensuremath{\mathcal{C}}}(E)$. Let $S\subset {\mathbb{R}}^d$ be convex. Then $S' \subseteq S$ is an [extreme subset]{} of $S$ if $S'$ is convex and satisfies: If $x,y\in S$ and $\exists \lambda \in (0,1)$ such that $\lambda x + (1-\lambda)y\in S'$, then $x,y\in S'$. Of special interest to us are convex polytopes which are convex hulls of finite sets. Definition \[def:polytope\] and Remark \[rem:polytope\] suffice for us. For an introduction to convex polytopes we refer the reader to, eg. [@Barvinok:2002; @Brondsted:1982; @Ewald:1996; @Grunbaum:2003; @Gruber:2007; @Ziegler:1994]. \[def:polytope\] The vertices of a convex polytope are its extreme points, its edges are its one-dimensional extreme subsets and its facets are its extreme subsets with co-dimension one. Let $E$ be a finite set. We denote the set of $n$-dimensional extreme subsets of ${\ensuremath{\mathcal{C}}}(E)$ by ${\ensuremath{\mathcal{F}}}_n(E)$. Thus the set of vertices of ${\ensuremath{\mathcal{C}}}(E)$ is ${\ensuremath{\mathcal{F}}}_0(E)$, the set of its edges is ${\ensuremath{\mathcal{F}}}_1(E)$, and the set of facets is ${\ensuremath{\mathcal{F}}}_{\dim({\ensuremath{\mathcal{C}}}(E))-1}(E)$ and ${\ensuremath{\mathcal{F}}}_{\dim{{\ensuremath{\mathcal{C}}}(E)}}(E) = \{ {\ensuremath{\mathcal{C}}}(E) \}$. \[rem:polytope\] For $S \in {\ensuremath{\mathcal{F}}}_{n}(E)$, $n=1,\dots, \dim({\ensuremath{\mathcal{C}}}(E))$, the relative boundary ${\ensuremath{\operatorname{rel}}}\partial S$ is a union of elements of ${\ensuremath{\mathcal{F}}}_{n-1}(E)$. Semi-convex functions and sets {#sec:semi-convexity} ------------------------------ Next we recall some elementary definitions and results about semi-convex functions and sets. For more details see, eg. [@Muller:1999; @Dacorogna:2007]. Since we work in the context of $S^{d \times d}$ (as opposed to $R^{d \times d}$) we have appended the qualifier “symmetrised” to the various notions of semi-convexity. The qualifier “symmetric” is also used in the literature. The term “symmetrised rank-one convex” (Defintion \[def:semi-convex\] below) might be misleading in that the matrices involved need not be of rank one but rather are the symmetric parts of rank-one matrices. An alternative name could be “wave-cone convex”, because the symmetric tensor products form the wave cone of a second-order linear differential operator whose kernel are the symmetrised gradients [@Tartar:1979; @Murat:1981]. 1. $f \colon S^{d \times d} \to {\mathbb{R}}$ is symmetrised rank-one convex if for all $\lambda \in [0,1]$ $$f(\lambda e_1 + (1-\lambda) e_2) \leqslant \lambda f(e_1) + (1-\lambda) f(e_2) \quad \forall e_1, e_2 \in S^{d \times d} \text{ with } e_1 {\ensuremath{\interleave}}e_2.$$ 2. A locally-bounded Borel function $f \colon S^{d \times d} \to {\mathbb{R}}$ is symmetrised quasiconvex if for an open and bounded set $U\subset {\mathbb{R}}^d$ with $\|\partial U\|=0$ one has $$f(e) \leqslant \frac{1}{\|U\|} \int_U f(e + D_s\varphi) \,dx \quad \forall \varphi \in W^{1,\infty}_0(U,{\mathbb{R}}^d),$$ whenever the integral on the right hand side exists, where $D_s \varphi$ is the symmetrised gradient of $\varphi$. \[def:semi-convex\] Let $E$ be a compact set in $S^{d \times d}$. 1. The symmetrised quasiconvex hull of $E$ is defined as $${\ensuremath{\mathcal{Q}}}(E) := \{ e \in S^{d \times d} \, \| \, f(e) \leqslant \sup_{e' \in E} f(e') \quad \forall f \colon S^{d \times d} \to {\mathbb{R}}\text{ quasiconvex} \}.$$ 2. The symmetrised rank-one convex hull of $E$ is defined as $$\begin{gathered} {\ensuremath{\mathcal{R}}}(E) := \{ e\in S^{d \times d} \, \| \, f(e) \leqslant \sup_{e' \in E} f(e') \\ \forall f \colon S^{d \times d} \to {\mathbb{R}}\text{ symmetrised rank-one convex} \}.\end{gathered}$$ 3. The symmetrised lamination convex hull of $E$, ${\ensuremath{\mathcal{L}}}(E)$, is the smallest set $\widetilde{E} \supseteq E$ such that $e_1, e_2 \in \widetilde{E}$ and $e_1 {\ensuremath{\interleave}}e_2$ implies that $[e_1, e_2] \subset \widetilde{E}$. We note that symmetrised lamination convex hulls can be constructed as follows: For $n \in {\mathbb{N}}_0$, let $$\begin{aligned} {\ensuremath{\mathcal{L}}}_0(E) &:= E, \\ {\ensuremath{\mathcal{L}}}_{n+1}(E) &:= \left\{ e\ \|\ \exists e_1, e_2 \in {\ensuremath{\mathcal{L}}}_n(E),\ e_1 {\ensuremath{\interleave}}e_2,\ e \in [e_1,e_2] \right\}.\end{aligned}$$ Then, $${\ensuremath{\mathcal{L}}}(E) = \bigcup_{n \in {\mathbb{N}}_0} {\ensuremath{\mathcal{L}}}_n(E).$$ Symmetrised rank-one convex hulls also have an alternate characterisation (cf. eg. [@Kreiner:2003 Lemma 2.5]): $$\begin{gathered} \label{eq:sroch} {\ensuremath{\mathcal{R}}}(E) := \{ e\in S^{d \times d} \, \| \, f(e) = 0 \forall f \colon S^{d \times d} \to [0,\infty) \\ \text{ symmetrised rank-one convex with } f(E)=\{0\} \}.\end{gathered}$$ We shall repeatedly use Remark \[rem:semi-convex\] and Lemma \[lem:pairwise-compatible\] without explicitly citing them: \[rem:semi-convex\] ${\ensuremath{\mathcal{L}}}(E) \subseteq {\ensuremath{\mathcal{R}}}(E) \subseteq {\ensuremath{\mathcal{Q}}}(E) \subseteq {\ensuremath{\mathcal{C}}}(E)$. \[lem:pairwise-compatible\] Let $E \subset {\ensuremath{{\mathcal S}^{d \times d}_{\text{\ensuremath{}}} }}$ be finite and compatible. Then ${\ensuremath{\mathcal{L}}}(E) = {\ensuremath{\mathcal{C}}}(E)$. In particular, $$\begin{aligned} {2} &\forall x,y \in {\ensuremath{{\mathcal S}^{3 \times 3}_{\text{\ensuremath{c}}} }}, \quad &{\ensuremath{\mathcal{R}}}(\{x,y\}) &= \begin{cases} [x,y] & \text{if } x {\ensuremath{\interleave}}y, \\ \{x,y\} & \text{else}. \end{cases} \label{eq:1D}\end{aligned}$$ Lemma \[lem:pairwise-compatible\] however is too weak for our purposes; we present a sharp version [@Zhang:2006 p.38][@Tang:2008 p.1266]: \[thm:edge-compatible\] Let $E \subset {\ensuremath{{\mathcal S}^{d \times d}_{\text{\ensuremath{}}} }}$ be finite. Then ${\ensuremath{\mathcal{L}}}(E) = {\ensuremath{\mathcal{C}}}(E)$ if and only if every edge of ${\ensuremath{\mathcal{C}}}(E)$ is compatible, i.e., if and only if $$\forall e_1, e_2 \in E, \qquad \ e_1 {\ensuremath{\cancel{\interleave}}}e_2 \implies [e_1,e_2] \notin {\ensuremath{\mathcal{F}}}_1(E).$$ Necessity. Let $[e_1,e_2] \in {\ensuremath{\mathcal{F}}}_1(E)$ with $e_1 {\ensuremath{\cancel{\interleave}}}e_2$. We show that $(e_1,e_2) \not\subset {\ensuremath{\mathcal{L}}}(E)$: Let $e \in [e_1,e_2]$ and $e \in {\ensuremath{\mathcal{L}}}(E)$. Then $\exists n \in {\mathbb{N}}$ finite, $f_1, f_2 \in {\ensuremath{\mathcal{L}}}_n(E)$ such that $e \in [f_1,f_2]$. However by extremality of $[e_1,e_2]$, $f_1,f_2 \in [e_1,e_2]$. Applying the same argument to $f_1$ and $f_2$ we conclude that in fact $e \in {\ensuremath{\mathcal{L}}}(\{e_1,e_2\})$ but from , ${\ensuremath{\mathcal{L}}}(\{e_1,e_2\}) = \{e_1,e_2\}$. Thus $e \in [e_1,e_2]$ and $e \in {\ensuremath{\mathcal{L}}}(E) \implies e \in \{e_1,e_2\}$. Sufficiency. Let every edge of ${\ensuremath{\mathcal{C}}}(E)$ be compatible. We show by induction that $$S \in {\ensuremath{\mathcal{F}}}_n(E) \implies S \subseteq {\ensuremath{\mathcal{L}}}_n(E). \tag{$P_n$}$$ for $n=0,1,\dots, \dim({\ensuremath{\mathcal{C}}}(E))$. Since ($P_{\dim({\ensuremath{\mathcal{C}}}(E))}$) is the statement that ${\ensuremath{\mathcal{C}}}(E) \subseteq {\ensuremath{\mathcal{L}}}(E)$, the result follows from Remark \[rem:semi-convex\]. ($P_0$) is trivially true since $e \in {\ensuremath{\mathcal{F}}}_0$ implies $e \in E$ and thus $e \in {\ensuremath{\mathcal{L}}}_0(E)$; ($P_1$) is true since every edge of ${\ensuremath{\mathcal{C}}}(E)$ is compatible by assumption. Now let ($P_n$) be true for some fixed $n=1,\dots, \dim({\ensuremath{\mathcal{C}}}(E))-1$. We show that ($P_{n+1}$) is true: Let $S \in {\ensuremath{\mathcal{F}}}_{n+1}(E)$ and $e \in S$. If $e \in {\ensuremath{\operatorname{rel}}}\partial S$, then $e$ is contained in an element of ${\ensuremath{\mathcal{F}}}_{n}(E)$ (Remark \[rem:polytope\]) and thus, by the inductive hypothesis, $e \in {\ensuremath{\mathcal{L}}}_n(E) \subset {\ensuremath{\mathcal{L}}}_{n+1}(E)$. We consider the case $e \in {\ensuremath{\operatorname{rel}}}{\ensuremath{\operatorname{int}}}S$: Pick $v \in {\ensuremath{\mathcal{F}}}_1(E)\cap {\ensuremath{\operatorname{rel}}}\partial S$. We view $v$ as a vector in ${\ensuremath{{\mathcal S}^{d \times d}_{\text{\ensuremath{}}} }}$. Since $S$ is bounded, the inclusion $$e + \lambda v \in {\ensuremath{\operatorname{rel}}}\partial S, \quad \lambda \in {\mathbb{R}},$$ has two solutions $\lambda_- < 0$ and $\lambda_+ > 0$. Note that $(e + \lambda_+ v)-(e + \lambda_- v)$ is parallel to $v$, which is compatible (by assumption). Thus $e + \lambda_\pm v \in {\ensuremath{\mathcal{L}}}_n(E)$ (by ($P_n$)), $e + \lambda_- v {\ensuremath{\interleave}}e + \lambda_+ v$ and $$e = \frac{\lambda_+}{\lambda_+ - \lambda_-} (e + \lambda_- v) + \frac{-\lambda_-}{\lambda_+ - \lambda_-} (e + \lambda_+ v).$$ It follows that $e \in {\ensuremath{\mathcal{L}}}_{n+1}(E)$ and thus $S \subseteq {\ensuremath{\mathcal{L}}}_{n+1}(E)$. The compatible cone in two-dimensional affine subspaces of ${\ensuremath{{\mathcal S}^{3 \times 3}_{\text{\ensuremath{c}}} }}$ {#sec:2D-cone} ============================================================================================================================== Our first task is to characterise, both geometrically (Lemma \[lem:2D-compatible\_directions\]) and algebraically (Remark \[rem:2D-compatible\_directions\]), the symmetrised rank-one convex cone in two-dimensional affine subspaces of ${\ensuremath{{\mathcal S}^{3 \times 3}_{\text{\ensuremath{c}}} }}$. \[lem:2D-compatible\_directions\] Let $S$ be a two-dimensional subspace of ${\ensuremath{{\mathcal S}^{3 \times 3}_{\text{\ensuremath{c}}} }}$. Then either 1. $S$ contains precisely one, two or three compatible directions, or 2. $S$ is compatible. Let $\{e_1,e_2\}$ be a basis for $S$. By Lemma \[lem:SymTr3-compatibility\], an arbitrary non-zero element of $S$, $xe_1 + ye_2$, $(x,y) \neq 0$, is compatible iff $$\det(xe_1 + ye_2) = 0.$$ This is equivalent to $$\label{eq:local6} x^3 \det(e_1) + x^2 y \langle {\ensuremath{\operatorname{cof}}}(e_1), e_2 \rangle + x y^2 \langle e_1, {\ensuremath{\operatorname{cof}}}(e_2) \rangle + y^3 \det(e_2) = 0,$$ where ${\ensuremath{\operatorname{cof}}}(e)$ is the cofactor of $e$: $$\begin{gathered} {\ensuremath{\operatorname{cof}}}\begin{pmatrix} e_{11} & e_{12} & e_{13} \\ e_{21} & e_{22} & e_{23} \\ e_{31} & e_{32} & e_{33} \end{pmatrix} \\ = \begin{pmatrix} e_{22} e_{33} - e_{32} e_{23} & -\left( e_{21} e_{33} - e_{31} e_{23} \right) & e_{21} e_{32} - e_{31} e_{22} \\ -\left( e_{12} e_{33} - e_{32} e_{13} \right) & e_{11} e_{33} - e_{31} e_{13} & -\left( e_{11} e_{32} - e_{31} e_{12} \right) \\ e_{12} e_{23} - e_{22} e_{13} & -\left( e_{11} e_{23} - e_{21} e_{13} \right) & e_{11} e_{22} - e_{21} e_{12} \end{pmatrix}^T.\end{gathered}$$ If the polynomial in is the zero-polynomial (i.e., $\det(e_1)=0, \langle {\ensuremath{\operatorname{cof}}}(e_1), e_2 \rangle=0$ etc.) then $S$ is compatible. Otherwise: If either $\det(e_1) = 0$ or $\det(e_2) = 0$ then has one or two solutions $(x,y) \neq 0$ and thus $S$ contains one or two compatible directions. Assume on the contrary that both $\det(e_1)$ and $\det(e_2) \neq 0$. Since $(x,y) \neq 0$, dividing by either $x^3$ or $y^3$ we obtain a polynomial (in either $\frac{x}{y}$ or $\frac{y}{x}$) which, being cubic, has one, two or three distinct real roots. Thus $S$ contains one, two or three compatible directions. Let $S$ be a subspace of ${\ensuremath{{\mathcal S}^{3 \times 3}_{\text{\ensuremath{c}}} }}$ with $\dim(S) > 1$. Then $S$ contains at least one compatible direction. The following examples show that each of the possibilities referred to in Lemma \[lem:2D-compatible\_directions\] can occur: The determinant of these matrices is proportional to $x(x^2+y^2)$ so the compatible direction is $x=0$: $$\left\{ \left. \begin{pmatrix} x & 0 & 0 \\ 0 & x & y \\ 0 & y & -2x \end{pmatrix} \right\| x,y \in {\mathbb{R}}\right\}.$$ The determinant of these matrices is proportional to $x^3$ so the compatible direction is $x=0$: $$\left\{ \left. \begin{pmatrix} y & x & x \\ x & -y & x \\ x & x & 0 \end{pmatrix} \right\| x,y \in {\mathbb{R}}\right\}.$$ The determinant of these matrices is proportional to $xy^2$ so the two compatible directions are $x=0$ and $y=0$: $$\left\{ \left. \begin{pmatrix} x & 0 & 0 \\ 0 & -x & y \\ 0 & y & 0 \end{pmatrix} \right\| x,y \in {\mathbb{R}}\right\}.$$ The determinant of these matrices is proportional to $xy(x+y)$ so the three compatible directions are $x=0$, $y=0$ and $x=-y$: $$\left\{ \left. \begin{pmatrix} x & 0 & 0 \\ 0 & y & 0 \\ 0 & 0 & -x-y \end{pmatrix} \right\| x,y \in {\mathbb{R}}\right\}.$$ The determinant of these matrices is proportional to $x(x^2-y^2)$ so the three compatible directions are $x=0$, $x=y$ and $x=-y$: $$\left\{ \left. \begin{pmatrix} -2x & 0 & 0 \\ 0 & x & y \\ 0 & y & x \end{pmatrix} \right\| x,y \in {\mathbb{R}}\right\}.$$ $$\begin{gathered} \left\{ \left. \begin{pmatrix} 0 & 0 & 0 \\ 0 & x & y \\ 0 & y & -x \end{pmatrix} \right\| x,y \in {\mathbb{R}}\right\}, \quad \left\{ \left. \begin{pmatrix} 0 & x & y \\ x & 0 & 0 \\ y & 0 & 0 \end{pmatrix} \right\| x,y \in {\mathbb{R}}\right\}.\end{gathered}$$ We note that Lemma \[lem:2D-compatible\_directions\] follows also from the following characterisation of real homogeneous cubic polynomials in two variables: \[rem:2D-compatible\_directions\] A homogeneous cubic polynomial on ${\mathbb{R}}^2$ can, by an appropriate choice of basis, be written in precisely one of the five following forms cf. eg. [@Gurevich:1964 Sec.23, p.263-6 ],[@Weinberg:1988p655] or [@Olver:1999 p.28]: 1. $x(x^2+y^2)$, 2. $x^3$, 3. $xy^2$, 4. $xy(x+y)$ or, equivalently, $x(x^2-y^2)$, 5. $0$. The examples above illustrate these possibilities. Lemma \[lem:2D-compatible\_directions\] shows that from the perspective of symmetrised rank-one convexity there are four kinds of two-dimensional affine subspaces of ${\ensuremath{{\mathcal S}^{3 \times 3}_{\text{\ensuremath{c}}} }}$, namely those for which the compatible cone is 1. a line, \[it:2D-1-line\] 2. the union of two (distinct) lines, \[it:2D-2-lines\] 3. the union of three (distinct) lines, and \[it:2D-3-lines\] 4. the subspace itself. \[it:2D-plane\] The next step is to investigate compatible hulls in these subspaces. Cases and are the simplest: In Case compatible hulls are obtained by convexifying in the compatible direction, in Case compatible hulls are identical to convex hulls. Cases and are reminiscent of separate convexity in ${\mathbb{R}}^2$ [@Tartar:1993p191; @Kreiner:2003]: Symmetrised rank-one convexity is geometrically identical to separate convexity in ${\mathbb{R}}^2$ in Case , and is similar to separate convexity in ${\mathbb{R}}^2$ in Case . As an aside, the inclusion-minimal configurations [@Szekelyhidi:2005p253] in these situations are either $T_3s$ or $T_4s$ as might be expected. Only $T_3$s are relevant to our immediate purposes so we discuss them in Section \[sec:T3\] and leave the rest for [@Chenchiah-Schloemerkemper-PRSL]. $T_3$s {#sec:T3} ====== $T3$s occur when there are precisely three directions in the compatible cone. Our definition is equivalent/identical to earlier definitions in the literature such as [@Bhattacharya:1994p843]\[p. 855, (3.9) on p. 862 and Fig. 3.1b on p. 856\]. \[def:T3-1\] Three points $e_1, e_2, e_3 \in {\ensuremath{{\mathcal S}^{3 \times 3}_{\text{\ensuremath{c}}} }}$ form a $T_3$ if 1. They are pairwise incompatible, and \[it:local3\] 2. There exist $e_{1,1} \in (e_2,e_3), e_{2,2} \in (e_3,e_1), e_{3,3} \in (e_1,e_2)$ such that $e_{i,i} {\ensuremath{\interleave}}e_i$, $i=1,2,3$. \[it:local4\] A schematic representation of a $T_3$ is shown in Figure \[fig:T3\]. \[rem:T3\] Note that ${\ensuremath{\operatorname{span}}}\{e_1 - e_{1,1}\}$, ${\ensuremath{\operatorname{span}}}\{e_2 - e_{2,2}\}$ and ${\ensuremath{\operatorname{span}}}\{e_3 - e_{3,3}\}$ are distinct compatible directions. By Lemma \[lem:2D-compatible\_directions\] there are no others. It follows that $e_{1,1}$, $e_{2,2}$ and $e_{3,3}$ are unique. \[it:local5\] schematicT3.pdf\_t Next we show that it is easy to check whether three points form a $T_3$: \[lem:T3\] Three points $e_1, e_2, e_3 \in {\ensuremath{{\mathcal S}^{3 \times 3}_{\text{\ensuremath{0}}} }}$ form a $T_3$ if and only if $$\label{eq:local7} {\ensuremath{\operatorname{sign}}}\det(e_1 - e_2) = {\ensuremath{\operatorname{sign}}}\det(e_2 - e_3) = {\ensuremath{\operatorname{sign}}}\det(e_3 - e_1) \neq 0.$$ Let $e_1, e_2, e_3 \in {\ensuremath{{\mathcal S}^{3 \times 3}_{\text{\ensuremath{0}}} }}$ satisfy . From Lemma \[lem:SymTr3-compatibility\] it follows immediately that Definition \[def:T3-1\] is satisfied. Consider the polynomial $[0,1] \ni \lambda \mapsto \det( (\lambda e_1 + (1-\lambda) e_2 ) - e_3 )$. By assumption, this polynomial has opposite signs at $0$ and $1$. In conjunction with Lemma \[lem:SymTr3-compatibility\] this implies: \[eq:T3-scaffold\] $$\begin{aligned} {2} \label{eq:T3-scaffolda} \exists \lambda_{12} \in (0,1), \quad &e_{3,3} &:= (\lambda_{12} e_1 + (1-\lambda_{12}) e_2 ) &{\ensuremath{\interleave}}e_3.\end{aligned}$$ Similarly, $$\begin{aligned} {2} \label{eq:T3-scaffoldb} \exists \lambda_{23} \in (0,1), \quad &e_{1,1} &:= (\lambda_{23} e_2 + (1-\lambda_{23}) e_3 ) &{\ensuremath{\interleave}}e_1, \\ \exists \lambda_{31} \in (0,1), \quad &e_{2,2} &:= (\lambda_{31} e_3 + (1-\lambda_{31}) e_1 ) &{\ensuremath{\interleave}}e_2. \label{eq:T3-scaffoldc}\end{aligned}$$ This shows that Definition \[def:T3-1\] is satisfied. Thus $\{e_1, e_2, e_3\}$ forms a $T_3$. Conversely, assume that $\{e_1, e_2, e_3\} \subset {\ensuremath{{\mathcal S}^{3 \times 3}_{\text{\ensuremath{0}}} }}$ form a $T_3$. Then, from Definition \[def:T3-1\], ${\ensuremath{\operatorname{sign}}}\det(e_1 - e_2), {\ensuremath{\operatorname{sign}}}\det(e_2 - e_3), {\ensuremath{\operatorname{sign}}}\det(e_3 - e_1) \neq 0$. Suppose it is not the case that ${\ensuremath{\operatorname{sign}}}\det(e_1 - e_2) = {\ensuremath{\operatorname{sign}}}\det(e_2 - e_3) = {\ensuremath{\operatorname{sign}}}\det(e_3 - e_1)$. Then, relabelling the points if necessary, we have $$\label{eq:local8} \det(e_1 - e_3), \det(e_2 - e_3) > 0.$$ Now consider the cubic polynomial ${\mathbb{R}}\ni \lambda \mapsto \det( (\lambda e_1 + (1-\lambda) e_2 ) - e_3 )$. By assumption there are three compatible directions passing through $e_3$, which are parallel to $[e_1,e_{1,1}]$, $[e_2,e_{2,2}]$ and $[e_3,e_{3,3}]$; see Figure \[fig:T3\]. Thus the cubic polynomical has three distinct real roots, one in each of $(-\infty,0)$, $(0,1)$ and $(1,\infty)$. Thus each root is simple; it follows that the polynomial changes sign around each root. On the other hand from the polynomial is positive at $0$ and $1$. It follows that it has one further root in $(0,1)$, which is a contradiction. We conclude that ${\ensuremath{\operatorname{sign}}}\det(e_1 - e_2) = {\ensuremath{\operatorname{sign}}}\det(e_2 - e_3) = {\ensuremath{\operatorname{sign}}}\det(e_3 - e_1)$. Let $e_1, e_2, e_3 \in {\ensuremath{{\mathcal S}^{3 \times 3}_{\text{\ensuremath{c}}} }}$ form a $T_3$ as in Definition \[def:T3-1\]. Then $e_1, e_2, e_3$ are referred to as the vertices of this $T_3$. When $e_1, e_2, e_3 \in {\ensuremath{{\mathcal S}^{3 \times 3}_{\text{\ensuremath{c}}} }}$ form a $T_3$ then their symmetrised rank-one convex hull ${\ensuremath{\mathcal{R}}}(\{e_1, e_2, e_3\})$ is (also) called a $T_3$. The context will make clear whether “$T_3$” refers to the set of three strains $\{e_1, e_2, e_3\}$ or to their symmetrised rank-one convex hull ${\ensuremath{\mathcal{R}}}(\{e_1, e_2, e_3\})$. It is known (cf., e.g., [@Bhattacharya:1994p843 §3]) that $$\label{eq:T3-hull} {\ensuremath{\mathcal{R}}}(\{e_1, e_2, e_3\}) \supseteq [e_1,e_{1,2}] \cup [e_2,e_{2,3}] \cup [e_3,e_{3,1}] \cup {\ensuremath{\mathcal{C}}}(\{e_{1,2}, e_{2,3}, e_{3,1}\}),$$ where $e_{1,2}$, $e_{2,3}$, $e_{3,1}$ are the nodes of the $T_3$ (Definition \[def:T3-nodes\] below), see Figure \[fig:T3-hull\]. However for the convenience of the reader we provide a proof of this in Proposition \[prop:T3-hull\] below. rkT3.pdf\_t \[def:T3-nodes\] Let $e_1, e_2, e_3 \in {\ensuremath{{\mathcal S}^{3 \times 3}_{\text{\ensuremath{c}}} }}$ form a $T_3$. Let $e_{1,1}$, $e_{2,2}$, $e_{3,3}$ be as defined in –. We define the nodes of the $T_3$ to be the three points $$e_{i,j} := [e_i,e_{i,i}] \cap [e_j,e_{j,j}], \quad i \neq j,\ i,j=1,2,3$$ (see Figure \[fig:T3\]). From Definition \[def:T3-1\] the nodes of a $T_3$ are pair-wise compatible and $e_{i,j}$ is compatible with $e_i$ and $e_j$. Later we will encounter symmetric $T_3$s and similar $T_3$s as defined below: \[def:T3-symmetry\] Let $e_1, e_2, e_3 \in {\ensuremath{{\mathcal S}^{3 \times 3}_{\text{\ensuremath{c}}} }}$ form a $T_3$. Let $\lambda_{12}$, $\lambda_{23}$, $\lambda_{31}$ be as defined in –. Then the $T_3$ is symmetric if $\lambda_{12} = \lambda_{23} = \lambda_{31}$. The following remark is elementary but we explicitly state it since it arises frequently in applications (cf. Section \[sec:nonlaminates\]): \[rem:T3-symmetry\] Let $R \in SO(3)$ such that $R^3$ is the identity. Let $e_1 \in {\ensuremath{{\mathcal S}^{3 \times 3}_{\text{\ensuremath{c}}} }}$, $e_2 := R^T e_1 R$ and $e_3 := R^T e_2 R$. Then $\det(e_1-e_2) = \det(e_2-e_3) = \det(e_3-e_1)$ and if this is non-zero then $e_1$, $e_2$, $e_3$ form a symmetric $T_3$. \[def:T3-similarity\] Let $e_1, e_2, e_3 \in {\ensuremath{{\mathcal S}^{3 \times 3}_{\text{\ensuremath{c}}} }}$ and $e'_1, e'_2, e'_3 \in {\ensuremath{{\mathcal S}^{3 \times 3}_{\text{\ensuremath{c}}} }}$ form $T_3$s. Let $\lambda_{ij}$, $\lambda'_{ij}$, $i \neq j$, $i,j=1,2,3$ be defined as in . 1. We say that these $T_3$s are similar if, for some permutation $\sigma \colon \{1,2,3\} \to \{1,2,3\}$, $\lambda'_{12} = \lambda_{\sigma(1)\sigma(2)}$, $\lambda'_{23} = \lambda_{\sigma(2)\sigma(3)}$ and $\lambda'_{31} = \lambda_{\sigma(3)\sigma(1)}$. \[it:local11\] 2. Corresponding points in these $T_3$s are points with the same barycentric coordinates: For $\mu_1, \mu_2, \mu_3 \in [0,1]$, $\sum_{i=1}^3 \mu_i = 1$, $\mu_1 e_1 + \mu_2 e_2 + \mu_3 e_3$ and $\mu_1 e'_{\sigma(1)} + \mu_2 e'_{\sigma(2)} + \mu_3 e'_{\sigma(3)}$ are corresponding points where $\sigma$ is the permutation in item above. We now prove : \[prop:T3-hull\] Let $e_1, e_2, e_3 \in {\ensuremath{{\mathcal S}^{3 \times 3}_{\text{\ensuremath{c}}} }}$ form a $T_3$. Then $${\ensuremath{\mathcal{R}}}(\{e_1, e_2, e_3\}) \supseteq [e_1,e_{1,2}] \cup [e_2,e_{2,3}] \cup [e_3,e_{3,1}] \cup {\ensuremath{\mathcal{C}}}(\{e_{1,2}, e_{2,3}, e_{3,1}\}). \tag{\ref{eq:T3-hull}}$$ We use the same strategy as [@Kreiner:2003 Proposition 2.7]. Let $e_{1,2}$, $e_{2,3}$, $e_{3,1}$ be related to $e_1$, $e_2$, $e_3$ as in the left-hand side of Figures \[fig:T3\] and \[fig:T3-hull\]; the proof in the other case is similar. Let $f$ be non-negative, symmetrised rank-one convex and vanish on $\{ e_1, e_2, e_3 \}$. Since $e_1 {\ensuremath{\interleave}}e_{1,2}$, it follows from the convexity of $f$ on $[e_1,e_{1,2}]$ that $f(e_{1,2}) \geqslant f(e_{3,1})$. Similarly, since $e_2 {\ensuremath{\interleave}}e_{2,3}$, it follows that $f(e_{2,3}) \geqslant f(e_{1,2})$. Finally from $e_3 {\ensuremath{\interleave}}e_{3,1}$ it follows that $f(e_{3,1}) \geqslant f(e_{2,3})$. In other words, $f(e_{2,3}) \geqslant f(e_{1,2}) \geqslant f(e_{3,1}) \geqslant f(e_{2,3})$. We conclude that $f(e_{2,3}) = f(e_{1,2}) = f(e_{3,1})= f(e_{2,3})$. But since $f(e_1)=0$, convexity of $f$ on $[e_1,e_{1,2}]$ shows that in fact $f(e_{1,2}) = f(e_{2,3}) = f(e_{3,1}) = 0$. From we conclude that $$[e_1,e_{1,2}] \cup [e_2,e_{2,3}] \cup [e_3,e_{3,1}] \subset {\ensuremath{\mathcal{R}}}(\{e_1,e_2,e_3\}).$$ The last step is to notice that since $e_{1,2}$, $e_{2,3}$, $e_{3,1}$ are pair-wise compatible, from Lemma \[lem:pairwise-compatible\], $${\ensuremath{\mathcal{C}}}(\{e_{1,2}, e_{2,3}, e_{3,1}\}) \subset {\ensuremath{\mathcal{R}}}(\{e_1,e_2,e_3\}),$$ which completes the proof. We end this section by giving an example of the utility of two-dimensional results in higher dimensions. \[lem:T3-3D\] Let $e_1, e_2, e_3 \in {\ensuremath{{\mathcal S}^{3 \times 3}_{\text{\ensuremath{c}}} }}$ form a $T_3$. Let $e_0 \in {\ensuremath{{\mathcal S}^{3 \times 3}_{\text{\ensuremath{c}}} }}$ such that $e_i {\ensuremath{\interleave}}e_0$ for $i=1,2,3$. For $\lambda \in {\mathbb{R}}$ and $i=1,2,3$ let $$\begin{aligned} e_i^\lambda &:= \lambda e_0 + (1-\lambda)e_i.\end{aligned}$$ Then $e_1^\lambda, e_2^\lambda, e_3^\lambda$ also form a $T_3$ and $$\label{eq:local12} \bigcup_{\lambda \in [0,1]} {\ensuremath{\mathcal{R}}}(\{e_1^\lambda, e_2^\lambda, e_3^\lambda\}) \subseteq {\ensuremath{\mathcal{R}}}(\{e_0,e_1,e_2,e_3\})$$ which, when $e_0 \notin {\ensuremath{\operatorname{aff}}}{\ensuremath{\operatorname{span}}}\{ e_1, e_2, e_3 \}$, is a three-dimensional continuum of $T_3$s. 3dT3.pdf\_t Let $\lambda \in (0,1)$ and $i=1,2,3$. Since $e_0 {\ensuremath{\interleave}}e_i$ it follows that $e_i^\lambda \in {\ensuremath{\mathcal{L}}}(\{e_0,e_1,e_2,e_3\})$. Since ${\ensuremath{\operatorname{sign}}}\det(e_i^\lambda - e_j^\lambda) = {\ensuremath{\operatorname{sign}}}\det (e_i - e_j)$, $i,j=1,2,3$, it follows that $e_1^\lambda, e_2^\lambda, e_3^\lambda$ form a $T_3$. The result follows. The transformation strains of Monoclinic-I martensite {#sec:strains} ===================================================== In this and the next section we prepare to apply the results of the preceding sections to monoclinic-I martensite by exploring first the symmetry and the geometry of this material. We denote the transformation strains of the twelve variants of cubic-to-monoclinic-I martensite, listed in Table \[tab:strains\], by $$\begin{aligned} e^{(i)} \in {\ensuremath{\mathcal{E}}}&:= \{ e^{(i)} \ \|\ i \in {\ensuremath{\mathcal{I}}}\}, \\ {\ensuremath{\mathcal{I}}}&:= \{1,2, \dots,12\} \end{aligned}$$ or simply by $i \in {\ensuremath{\mathcal{I}}}$ if the meaning is clear from the context. The transformation strains involve four lattice parameters $\alpha$, $\beta$, ${\varepsilon}$ and $\delta$. These have been chosen such that ${\varepsilon}> 0$ and $\delta > 0$. Typical lattice parameters are listed in Table \[tab:parameters\]. Note that ${\ensuremath{\operatorname{Tr}}}e = 2\alpha+\beta$ for $e \in {\ensuremath{\mathcal{E}}}$; thus ${\ensuremath{\mathcal{E}}}\subset {\ensuremath{{\mathcal S}^{3 \times 3}_{\text{\ensuremath{2\alpha+\beta}}} }}$. [l\|l\|\|l\|l\|\|l\|l\|\|l\|l]{} i & $e^{(i)}$ & i & $e^{(i)}$ & i & $e^{(i)}$ & i & $e^{(i)}$\ \ 1 & $\begin{pmatrix} \alpha & \delta & {\varepsilon}\\ \delta & \alpha & {\varepsilon}\\ {\varepsilon}& {\varepsilon}&\beta \end{pmatrix}$ & 2 & $\begin{pmatrix} \alpha & \delta & -{\varepsilon}\\ \delta & \alpha & -{\varepsilon}\\ -{\varepsilon}& -{\varepsilon}&\beta \end{pmatrix}$ & 3 & $\begin{pmatrix} \alpha & -\delta & -{\varepsilon}\\ -\delta & \alpha & {\varepsilon}\\ -{\varepsilon}& {\varepsilon}&\beta \end{pmatrix} $ & 4 & $\begin{pmatrix} \alpha & -\delta & {\varepsilon}\\ -\delta & \alpha & -{\varepsilon}\\ {\varepsilon}& -{\varepsilon}&\beta \end{pmatrix}$\ &&&&&&\ 5 & $\begin{pmatrix} \alpha & {\varepsilon}& \delta \\ {\varepsilon}& \beta & {\varepsilon}\\ \delta & {\varepsilon}&\alpha \end{pmatrix}$ & 6 & $\begin{pmatrix}\alpha & -{\varepsilon}& \delta \\ -{\varepsilon}& \beta & -{\varepsilon}\\ \delta & -{\varepsilon}&\alpha \end{pmatrix}$ & 7 & $\begin{pmatrix} \alpha & -{\varepsilon}& -\delta \\ -{\varepsilon}& \beta & {\varepsilon}\\ -\delta & {\varepsilon}&\alpha \end{pmatrix}$ & 8 & $\begin{pmatrix}\alpha & {\varepsilon}& -\delta \\ {\varepsilon}& \beta & -{\varepsilon}\\ -\delta & -{\varepsilon}&\alpha \end{pmatrix}$\ &&&&&&\ 9 & $\begin{pmatrix} \beta & {\varepsilon}& {\varepsilon}\\ {\varepsilon}& \alpha & \delta \\ {\varepsilon}& \delta &\alpha \end{pmatrix}$ & 10 & $\begin{pmatrix} \beta & -{\varepsilon}& -{\varepsilon}\\ -{\varepsilon}& \alpha & \delta \\ -{\varepsilon}& \delta &\alpha \end{pmatrix}$ & 11 & $\begin{pmatrix} \beta & -{\varepsilon}& {\varepsilon}\\ -{\varepsilon}& \alpha & -\delta \\ {\varepsilon}& -\delta &\alpha \end{pmatrix} $ & 12 & $\begin{pmatrix} \beta & {\varepsilon}& -{\varepsilon}\\ {\varepsilon}& \alpha & -\delta \\ -{\varepsilon}& -\delta &\alpha \end{pmatrix}$ $\alpha$ $\beta$ $\delta$ ${\varepsilon}$ Reference -------- ---------- --------- ---------- ----------------- ---------------------------------------------- NiTi 0.0243 -0.0437 0.0580 0.0427 [@Otsuka:1971; @Knowles:1981; @Hane:1999-07] CuZr 0.0348 0.0229 0.1067 0.0929 [@Seo:1998-02a; @Seo:1998-02b] TiNiCu 0.0232 -0.0410 0.0532 0.0395 [@Nam:1990] : Typical lattice parameters for monoclinic-I martensite, cf. eg. [@Shu:1998p5457 Table 2, p. 5459] and [@Bhattacharya:2003 p. 55 and 184].[]{data-label="tab:parameters"} #### Compatibility. Compatibility and incompatibility between the transformation strains is of critical importance to us. A simple calculation [@Chenchiah-Schloemerkemper1ESM Compatibility.nb] shows that for $e,f \in {\ensuremath{\mathcal{E}}}$, $$\label{eq:determinant} \det(e-f) \in \left\{ 0, \pm 4 {\varepsilon}\left( (\alpha - \beta) \delta + {\varepsilon}^2 - \delta^2 \right) \right\},$$ which gives the compatibility/incompatibility of the strains by Lemma \[lem:SymTr3-compatibility\]. Note that if the material parameters happen to be such that $(\alpha-\beta) \delta + {\varepsilon}^2 - \delta^2 = 0$ then all strains in ${\ensuremath{\mathcal{E}}}$ are pairwise compatible. Then by Lemma \[lem:pairwise-compatible\], ${\ensuremath{\mathcal{L}}}({\ensuremath{\mathcal{E}}}) = {\ensuremath{\mathcal{C}}}({\ensuremath{\mathcal{E}}})$. (In this case the material is able to form many more twins than usual [@Pitteri:1998].) Pairs of compatible and incompatible transformation strains in ${\ensuremath{\mathcal{E}}}$ are listed in Table \[tab:compatibility\]. Here and henceforth (including in Mathematica calculations) we assume that $(\alpha-\beta) \delta + {\varepsilon}^2 - \delta^2 \neq 0$ and, more generally, that the lattice parameters are generic. We also assume that $\alpha \neq \beta$, the (mathematical) reason for this will become clear in Section \[sec:polytope\]. There we will also see that the case ${\varepsilon}= \delta$ is special so we shall specifically consider this possibility. i $\det(e^{(\cdot)} - e^{(i)}) = 0$ $\det(e^{(\cdot)} - e^{(i)}) \gtrless 0$ $\det(e^{(\cdot)} - e^{(i)}) \lessgtr 0$ ---- ----------------------------------- ------------------------------------------ ------------------------------------------ 1 2, 3, 4, 5, 7, 9, 11 8, 12 6, 10 2 1, 3, 4, 6, 8, 10, 12 5, 9 7, 11 3 1, 2, 4, 5, 7, 10, 12 6, 11 8, 9 4 1, 2, 3, 6, 8, 9, 11 7, 10 5, 12 5 1, 3, 6, 7, 8, 9, 12 4, 11 2, 10 6 2, 4, 5, 7, 8, 10, 11 1, 9 3, 12 7 1, 3, 5, 6, 8, 10, 11 2, 12 4, 9 8 2, 4, 5, 6, 7, 9, 12 3, 10 1, 11 9 1, 4, 5, 8, 10, 11, 12 3, 7 2, 6 10 2, 3, 6, 7, 9, 11, 12 1, 5 4, 8 11 1, 4, 6, 7, 9, 10, 12 2, 8 3, 5 12 2, 3, 5, 8, 9, 10, 11 4, 6 1, 7 : Compatible and incompatible transformation strains [@Chenchiah-Schloemerkemper1ESM Compatibility.nb]. The signs in the second and third columns depend on the material parameters; the sign in the third column is opposite to the one in the second.[]{data-label="tab:compatibility"} #### Distances. Also of importance is the distance between the transformation strains [@Chenchiah-Schloemerkemper1ESM Distances.nb]: \[obs:distance\]Every pair of incompatible transformation strains is equidistant: For $e,f \in {\ensuremath{\mathcal{E}}}$ with $e {\ensuremath{\cancel{\interleave}}}f$, \[eq:distance\] $$\label{eq:distance-incompatible} \\| e - f \\|^2 = 2 (\alpha - \beta)^2 + 4 \delta^2 + 12 \epsilon^2.$$ Remark \[rem:incompatible-distances\] sheds more light on this. However for pairs of compatible transformation strains the situation is more complex: For $e,f \in {\ensuremath{\mathcal{E}}}$ with $e {\ensuremath{\interleave}}f$ and $e \neq f$, $$\label{eq:distance-compatible-1} \\| e - f \\|^2 \in \left\{ 16 {\varepsilon}^2, 8 (\delta^2 + {\varepsilon}^2), 2 (\alpha - \beta)^2 + 4 (\delta - {\varepsilon})^2, 2 (\alpha - \beta)^2 + 4 (\delta + {\varepsilon})^2 \right\}.$$ Table \[tab:distances\] presents the full picture. (See also Remark \[rem:compatible-distances\].) We exclude (until Section \[sec:conclusions\]) the special case ${\varepsilon}= \delta$ for which, for $e,f \in {\ensuremath{\mathcal{E}}}$ with $e \neq f$, $$\\| e - f \\|^2 \in \left\{ 16 {\varepsilon}^2, 2 (\alpha - \beta)^2, 2 (\alpha - \beta)^2 + 16 {\varepsilon}^2 \right\}.$$ Note that in this case, we cannot anymore distinguish between compatible and incompatible strains on the basis of the distance between them. ---- --------------------------------- ------------------------------------ --------------------------------------------------------- --------------------------------------------------------- i $\\| e^{(\cdot)} - e^{(i)} \\|^2$ $\\| e^{(\cdot)} - e^{(i)} \\|^2$ $\\| e^{(\cdot)} - e^{(i)} \\|^2$ $\\| e^{(\cdot)} - e^{(i)} \\|^2$ $= 16 {\varepsilon}^2$ $= 8 (\delta^2 + {\varepsilon}^2)$ $= 2 (\alpha - \beta)^2 + 4 (\delta - {\varepsilon})^2$ $= 2 (\alpha - \beta)^2 + 4 (\delta + {\varepsilon})^2$ 1 2 3, 4 5, 9 7, 11 2 1 3, 4 8, 12 6, 10 3 4 1, 2 7, 10 5, 12 4 3 1, 2 6, 11 8, 9 5 6 7, 8 1, 9 3, 12 6 5 7, 8 4, 11 2, 10 7 8 5, 6 3, 10 1, 11 8 7 5, 6 2, 12 4, 9 9 10 11, 12 1, 5 4, 8 10 9 11, 12 3, 7 2, 6 11 12 9, 10 4, 6 1, 7 12 11 9, 10 2, 8 3, 5 ---- --------------------------------- ------------------------------------ --------------------------------------------------------- --------------------------------------------------------- : Distances between compatible transformation strains [@Chenchiah-Schloemerkemper1ESM Distances.nb].[]{data-label="tab:distances"} #### Symmetry. Now that we have knowledge of the compatibilities and the distances between the transformation strains we are ready to analyse the symmetry between them: \[def:symmetrygroups\] 1. A map $\tau \colon {\ensuremath{\mathcal{E}}}\to {\ensuremath{\mathcal{E}}}$ is a symmetry of ${\ensuremath{\mathcal{E}}}$ if it preserves distance and compatibility in ${\ensuremath{\mathcal{E}}}$. That is, $\forall e,f \in {\ensuremath{\mathcal{E}}}$, $$\begin{aligned} \\| e - f \\| &= \\| \tau e - \tau f \\|, \\ \det (e - f) &= \pm \det( \tau e - \tau f ).\end{aligned}$$ A symmetry group of ${\ensuremath{\mathcal{E}}}$ is a group of symmetries of ${\ensuremath{\mathcal{E}}}$. 2. Let $n \in {\ensuremath{\mathcal{I}}}\setminus \{1\}$ and ${\mathcal E}_n$ be a set of subsets of ${\ensuremath{\mathcal{E}}}$, all with cardinality $n$. That is, $${\mathcal E}_n \subset \left\{ S \subset {\ensuremath{\mathcal{E}}}\ \|\ \# S = n \right\}.$$ A map $\tau \colon {\mathcal E}_n \to {\mathcal E}_n$ is a symmetry of ${\mathcal E}_n$ if it preserves distance and compatibility in ${\mathcal E}_n$. That is, $\forall S \in {\mathcal E}_n$, $\forall e,f \in S$, $$\begin{aligned} \\| e - f \\| &= \\| \tau e - \tau f \\|, \\ \det (e - f) &= \pm \det( \tau e - \tau f ).\end{aligned}$$ A symmetry group of ${\mathcal E}_n$ is a group of symmetries of ${\mathcal E}_n$. There are four sets that are of interest to us here. These are: (i) ${\ensuremath{\mathcal{E}}}$ itself, (ii) ${\ensuremath{\mathcal{E}}}^2_{{\ensuremath{\interleave}}}$, the set of pairs of compatible transformation strains, (iii) ${\ensuremath{\mathcal{E}}}^2_{{\ensuremath{\cancel{\interleave}}}}$, the set of pairs of incompatible transformation strains, and (iv) ${\ensuremath{\mathcal{E}}}^3_{{\ensuremath{\cancel{\interleave}}}}$, the set of three-tuples of incompatible transformation strains: $$\begin{aligned} {\ensuremath{\mathcal{E}}}^2_{{\ensuremath{\interleave}}} &:= \left\{ \{e,f\}\ \|\ e,f \in {\ensuremath{\mathcal{E}}},\ e \neq f,\ e {\ensuremath{\interleave}}f \right\}, \\ {\ensuremath{\mathcal{E}}}^2_{{\ensuremath{\cancel{\interleave}}}} &:= \left\{ \{e,f\}\ \|\ e,f \in {\ensuremath{\mathcal{E}}},\ e {\ensuremath{\cancel{\interleave}}}f \right\}, \\ {\ensuremath{\mathcal{E}}}^3_{{\ensuremath{\cancel{\interleave}}}} &:= \left\{ \{e,f,g\}\ \|\ e,f,g \in {\ensuremath{\mathcal{E}}},\ e {\ensuremath{\cancel{\interleave}}}f {\ensuremath{\cancel{\interleave}}}g {\ensuremath{\cancel{\interleave}}}e \right\}. \end{aligned}$$ We characterise the symmetry of ${\ensuremath{\mathcal{E}}}$ and ${\ensuremath{\mathcal{E}}}^2_{{\ensuremath{\interleave}}}$ in Lemma \[lem:symmetry1\] and those of ${\ensuremath{\mathcal{E}}}^2_{{\ensuremath{\cancel{\interleave}}}}$ and ${\ensuremath{\mathcal{E}}}^3_{{\ensuremath{\cancel{\interleave}}}}$ in Lemma \[lem:symmetry2\] below. In order to do so we begin with some observations. For further investigation of these sets see [@Chenchiah-Schloemerkemper-JMPS]. \[obs:oriented\_cube\] Since the transformation strains in ${\ensuremath{\mathcal{E}}}$ are obtained through a phase transformation from a cubic crystal, $S_4$, the group of rotational symmetries of a cube, is a symmetry group of ${\ensuremath{\mathcal{E}}}$. This group and its action on ${\ensuremath{\mathcal{E}}}$ can be generated as follows: Let $R_1$, $R_2$ and $R_3$ be anticlockwise rotations of $\frac{\pi}{2}$ about the coordinate axes: $$R_1 = \begin{pmatrix} 1 & 0 & 0 \\ 0 & 0 & -1 \\ 0 & 1 & 0 \end{pmatrix}, \qquad R_2 = \begin{pmatrix} 0 & 0 & 1 \\ 0 & 1 & 0 \\ -1 & 0 & 0 \end{pmatrix}, \qquad R_3 = \begin{pmatrix} 0 & -1 & 0 \\ 1 & 0 & 0 \\ 0 & 0 & 1 \end{pmatrix}.$$ For $i=1,2,3$, let $r_i$ be the map $$\label{eq:coordinate-rotations} {\ensuremath{{\mathcal S}^{3 \times 3}_{\text{\ensuremath{}}} }} \ni e \mapsto r_i e := R_i e R_i^T.$$ It is immediate that these are distance and determinant (and thus, compatibility) preserving: $\forall i=1,2,3$, $\forall e,f \in {\ensuremath{\mathcal{E}}}$, \[eq:rotations\] $$\begin{aligned} \\| e - f \\| &= \\| r_i e - r_i f \\|, \\ \det (e - f) &= \det( r_i e - r_i f ).\end{aligned}$$ Then $\{r_1,r_2,r_3\}$ generates $S_4$. (In fact any two of $r_1,r_2,r_3$ generate $S_4$ but it is convenient to retain all three.) The action of $S_4$ on ${\ensuremath{\mathcal{E}}}$ is listed, eg. in [@Hane:1999-07 Table 1, p. 2607] but their numbering of the transformation strains is different from ours. $S_4$ is isomorphic to a group of permutations on ${\ensuremath{\mathcal{I}}}$. We denote this group too by $S_4$; the images of $r_1$, $r_2$, $r_3$ under this isomorphism are also denoted by $r_1$, $r_2$, $r_3$. Table \[tab:generators\] lists the action of $r_1$, $r_2$, $r_3$ on ${\ensuremath{\mathcal{I}}}$ [@Chenchiah-Schloemerkemper1ESM Symmetry.nb]. $i$ 1 2 3 4 5 6 7 8 9 10 11 12 --------- ---- ---- --- ---- ---- --- ---- ---- ---- ---- ---- ---- $r_1 i$ 6 5 8 7 4 3 2 1 11 12 10 9 $r_2 i$ 12 11 9 10 8 7 5 6 2 1 3 4 $r_3 i$ 3 4 2 1 10 9 12 11 7 8 5 6 : The action of $r_1$, $r_2$, $r_3$ on ${\ensuremath{\mathcal{I}}}$.[]{data-label="tab:generators"} \[rem:cube\] Since $S_4$ is the group of rotational symmetries of a cube, it is natural and convenient to identify ${\ensuremath{\mathcal{E}}}$ (and ${\ensuremath{\mathcal{I}}}$) with the edges of a cube as shown in Figure \[fig:oriented-cube\]. (We could also have identified them with the diagonals of the faces.) Then $r_1$, $r_2$, $r_3$ are anticlockwise rotations of $\frac{\pi}{2}$ along an axis perpendicular to the face of the cube and passing through its centre. From Table \[tab:compatibility\] we note that the four edges with which an edge is incompatible are precisely the four edges with which it shares a vertex. Thus, 1. \[it:corners-of-faces\] The 24 elements of ${\ensuremath{\mathcal{E}}}^2_{{\ensuremath{\cancel{\interleave}}}}$ can be identified with the 24 corners of the faces of a cube, and 2. \[it:corners\]The 8 elements of ${\ensuremath{\mathcal{E}}}^3_{{\ensuremath{\cancel{\interleave}}}}$ can be identified with the 8 corners of a cube. cube\_one.pdf\_t cube\_three.pdf\_t Our next task is to show that ${\ensuremath{\mathcal{E}}}$ does not have the reflection symmetries of a cube. We will accomplish this in Observation \[obs:inversion-distance\] below. \[def:inversion\] Let $r_0$ be the permutation on ${\ensuremath{\mathcal{I}}}$ which interchanges $1$ and $2$, $3$ and $4$, $5$ and $6$, $7$ and $8$, $9$ and $10$, and $11$ and $12$. As can be seen from Figure \[fig:oriented-cube\], $r_0$ is an inversion (reflection through the centre) of the cube. This immediately shows both that $r_0 \notin S_4$ and that $\{ r_0, r_1, r_2, r_3 \}$ generates $S_4 \times C_2$, the group of symmetries of a cube (including reflections). \[rem:inversion\] The permutation $r_0$ corresponds to replacing ${\varepsilon}$ with $-{\varepsilon}$ in the transformation strains (cf. Table \[tab:strains\]). Unlike $r_1$, $r_2$ and $r_3$, $r_0$ cannot be identified with a linear operator on ${\ensuremath{{\mathcal S}^{3 \times 3}_{\text{\ensuremath{2\alpha+\beta}}} }}$. Assume, on the contrary, that there exists ${\ensuremath{\mathcal{L}}}\colon {\ensuremath{{\mathcal S}^{3 \times 3}_{\text{\ensuremath{2\alpha+\beta}}} }} \to {\ensuremath{{\mathcal S}^{3 \times 3}_{\text{\ensuremath{2\alpha+\beta}}} }}$, linear, such that $\forall i \in {\ensuremath{\mathcal{I}}}$, $\mathcal{L} e^{(i)} = e^{(r_0 i)}$. Then, $$\begin{aligned} {2} {\ensuremath{\mathcal{L}}}\left( e^{(1)} + e^{(2)} + e^{(3)} + e^{(4)} \right) &= e^{(1)} + e^{(2)} + e^{(3)} + e^{(4)} \\ &\implies {\ensuremath{\mathcal{L}}}\begin{pmatrix} \alpha & 0 & 0 \\ 0 & \alpha & 0 \\ 0 & 0 & \beta \end{pmatrix} = \begin{pmatrix} \alpha & 0 & 0 \\ 0 & \alpha & 0 \\ 0 & 0 & \beta \end{pmatrix}, \\ {\ensuremath{\mathcal{L}}}\left( e^{(5)} + e^{(6)} + e^{(7)} + e^{(8)} \right) &= e^{(5)} + e^{(6)} + e^{(7)} + e^{(8)} \\ &\implies {\ensuremath{\mathcal{L}}}\begin{pmatrix} \alpha & 0 & 0 \\ 0 & \beta & 0 \\ 0 & 0 & \alpha \end{pmatrix} = \begin{pmatrix} \alpha & 0 & 0 \\ 0 & \beta & 0 \\ 0 & 0 & \alpha \end{pmatrix}, \\ {\ensuremath{\mathcal{L}}}\left( e^{(9)} + e^{(10)} + e^{(11)} + e^{(12)} \right) &= e^{(9)} + e^{(10)} + e^{(11)} + e^{(12)} \\ &\implies {\ensuremath{\mathcal{L}}}\begin{pmatrix} \beta & 0 & 0 \\ 0 & \alpha & 0 \\ 0 & 0 & \alpha \end{pmatrix} = \begin{pmatrix} \beta & 0 & 0 \\ 0 & \alpha & 0 \\ 0 & 0 & \alpha \end{pmatrix}.\end{aligned}$$ Thus (i) ${\ensuremath{\mathcal{L}}}$ is an identity on the diagonal components of ${\ensuremath{{\mathcal S}^{3 \times 3}_{\text{\ensuremath{2\alpha+\beta}}} }}$. Moreover $$\begin{aligned} {4} {\ensuremath{\mathcal{L}}}\left( e^{(1)} + e^{(3)} \right) &= e^{(2)} + e^{(4)} &\implies {\ensuremath{\mathcal{L}}}\begin{pmatrix} \alpha & 0 & 0 \\ 0 & \alpha & {\varepsilon}\\ 0 & {\varepsilon}& \beta \end{pmatrix} &= \begin{pmatrix} \alpha & 0 & 0 \\ 0 & \alpha & -{\varepsilon}\\ 0 & -{\varepsilon}& \beta \end{pmatrix}, \\ {\ensuremath{\mathcal{L}}}\left( e^{(5)} + e^{(8)} \right) &= e^{(6)} + e^{(7)} &\implies {\ensuremath{\mathcal{L}}}\begin{pmatrix} \alpha & {\varepsilon}& 0 \\ {\varepsilon}& \beta & 0 \\ 0 & 0 & \alpha \end{pmatrix} &= \begin{pmatrix} \alpha & -{\varepsilon}& 0 \\ -{\varepsilon}& \beta & 0 \\ 0 & 0 & \alpha \end{pmatrix}, \\ {\ensuremath{\mathcal{L}}}\left( e^{(9)} + e^{(11)} \right) &= e^{(10)} + e^{(12)} &\implies {\ensuremath{\mathcal{L}}}\begin{pmatrix} \beta & 0 & {\varepsilon}\\ 0 & \alpha & 0 \\ {\varepsilon}& 0 & \alpha \end{pmatrix} &= \begin{pmatrix} \beta & 0 & -{\varepsilon}\\ 0 & \alpha & 0 \\ -{\varepsilon}& 0 & \alpha \end{pmatrix}.\end{aligned}$$ Thus (ii) ${\ensuremath{\mathcal{L}}}$ is a negative of the identity on the off-diagonal components of ${\ensuremath{{\mathcal S}^{3 \times 3}_{\text{\ensuremath{2\alpha+\beta}}} }}$. It is easy to check that (i) and (ii) are contradictory, eg. ${\ensuremath{\mathcal{L}}}e^{(1)} = e^{(2)}$. Now we address the question of whether $r_0$ is a symmetry of ${\ensuremath{\mathcal{E}}}$: \[obs:inversion-compatibility\] As can be verified [@Chenchiah-Schloemerkemper1ESM Symmetry.nb] from Table \[tab:compatibility\] and , $r_0$ is compatibility preserving: $\forall e,f \in {\ensuremath{\mathcal{E}}}$, $$\det (e - f) = - \det( r_0 e - r_0 f ).$$ \[obs:inversion-distance\] From Observation \[obs:distance\] and Remark \[rem:inversion\] it follows that $r_0$ is distance-preserving on pairs of incompatible transformation strains. However it is distance-preserving only on some pairs of compatible transformation strains: For $e,f \in {\ensuremath{\mathcal{E}}}$, $$\label{eq:inversion-distance} \\| e - f \\|^2 = \\| r_0 e - r_0 f \\|^2 \qquad \iff \qquad \\| e - f \\|^2 \neq 2 (\alpha - \beta)^2 + 4 (\delta \pm {\varepsilon})^2.$$ This is easy to verify in view of the following: Let $e,f \in {\ensuremath{\mathcal{E}}}$ with $e {\ensuremath{\interleave}}f$. Then, $$\label{eq:distance-compatible-2} \\| e - f \\|^2 = \begin{cases} 16 {\varepsilon}^2 &\text{ if } e = r_0 f, \\ 8 \delta^2 + 8 {\varepsilon}^2 &\text{ if } \exists i \in \{1,2,3\},\ e = r_i f, \\ 2 (\alpha - \beta)^2 + 4 (\delta \pm {\varepsilon})^2 &\text{ else}. \end{cases}$$ (The middle case corresponds to parallel edges of the cube that share a face.) Thus $r_0$ is not a symmetry of ${\ensuremath{\mathcal{E}}}$ nor is it a symmetry of ${\ensuremath{\mathcal{E}}}^2_{{\ensuremath{\interleave}}}$. In the light of Observation \[obs:oriented\_cube\] we conclude that ${\ensuremath{\mathcal{E}}}$ and ${\ensuremath{\mathcal{E}}}^2_{{\ensuremath{\interleave}}}$ have the rotation symmetries of a cube but not its reflection symmetries: \[lem:symmetry1\] $S_4$ is a symmetry group of ${\ensuremath{\mathcal{E}}}$ and ${\ensuremath{\mathcal{E}}}^2_{{\ensuremath{\interleave}}}$, whereas $S_4 \times C_2$ is not a symmetry group of ${\ensuremath{\mathcal{E}}}$ or ${\ensuremath{\mathcal{E}}}^2_{{\ensuremath{\interleave}}}$. \[rem:compatible-distances\] From we deduce that the orbits of $S_4$ partition ${\ensuremath{\mathcal{E}}}^2_{{\ensuremath{\interleave}}}$ into four equivalence classes. (Equation relates this to the edges of a cube.) Observations \[obs:distance\], \[obs:oriented\_cube\] and \[obs:inversion-compatibility\] show that ${\ensuremath{\mathcal{E}}}^2_{{\ensuremath{\cancel{\interleave}}}}$ and ${\ensuremath{\mathcal{E}}}^3_{{\ensuremath{\cancel{\interleave}}}}$ have the symmetries of a cube: \[lem:symmetry2\] $S_4 \times C_2$ is a symmetry group of ${\ensuremath{\mathcal{E}}}^2_{{\ensuremath{\cancel{\interleave}}}}$ and of ${\ensuremath{\mathcal{E}}}^3_{{\ensuremath{\cancel{\interleave}}}}$. \[rem:incompatible-distances\] From Remark \[rem:cube\] it is clear that $S_4 \times C_2$ (or indeed $S_4$) acts transitively on ${\ensuremath{\mathcal{E}}}^2_{{\ensuremath{\cancel{\interleave}}}}$, i.e., the orbit of any element of ${\ensuremath{\mathcal{E}}}^2_{{\ensuremath{\cancel{\interleave}}}}$ under $S_4 \times C_2$ equals ${\ensuremath{\mathcal{E}}}^2_{{\ensuremath{\cancel{\interleave}}}}$. This explains Observation \[obs:distance\]. Likewise, from Remark \[rem:cube\] it is clear that $S_4 \times C_2$ (or indeed $S_4$) acts transitively on ${\ensuremath{\mathcal{E}}}^3_{{\ensuremath{\cancel{\interleave}}}}$. The convex polytope formed by the transformation strains {#sec:polytope} ======================================================== In this section we study ${\ensuremath{\mathcal{C}}}({\ensuremath{\mathcal{E}}})$, the convex hull of ${\ensuremath{\mathcal{E}}}$, which is a five-dimensional polytope [@Chenchiah-Schloemerkemper1ESM Dimension.nb], when, as we have assumed, $\alpha \neq \beta$ (otherwise it is a three-dimensional polytope). We are interested in the facets of ${\ensuremath{\mathcal{C}}}({\ensuremath{\mathcal{E}}})$. However the vertices and edges of ${\ensuremath{\mathcal{C}}}({\ensuremath{\mathcal{E}}})$ are also of interest and we begin with them. For convenience we set $\Lambda = \{ \lambda \in [0,1]^{12},\ \sum_{i=1}^{12} \lambda_i = 1 \}$. The following linear functionals will be helpful in studying ${\ensuremath{\mathcal{C}}}({\ensuremath{\mathcal{E}}})$: The linear functionals $H_i \colon {\ensuremath{{\mathcal S}^{3 \times 3}_{\text{\ensuremath{}}} }} \to {\mathbb{R}}$, $i=0,1,2,3$, are defined by $$\begin{aligned} {2} &\phantom{==}H_0\, e &&= -e_{12} - e_{23} - e_{31}, \\ H_1\, e &= H_0\, r_1 e &&= -e_{12} + e_{23} + e_{31}, \\ H_2\, e &= H_0\, r_2 e &&= e_{12} - e_{23} + e_{31}, \\ H_3\, e &= H_0\, r_3 e &&= e_{12} + e_{23} - e_{31}; \end{aligned}$$ and $H_{ij} \colon {\ensuremath{{\mathcal S}^{3 \times 3}_{\text{\ensuremath{}}} }} \to {\mathbb{R}}$, $i,j=1,2,3$, by $$H_{ij}\, e = e_{ij},$$ where $e_{ij}$ denotes the $(i,j)$-component of the matrix $e$. For the convenience of the reader we summarise the images of the transformation strains under the functionals $H_i$, $i=0,1,2,3$, in Table \[tab:H\]. We also list the extremisers of these functionals; note that the extremisers when ${\varepsilon}= \delta$ are the union of the extremisers when ${\varepsilon}< \delta$ and when ${\varepsilon}> \delta$. i $H_0\, e^{(i)}$ $H_1\, e^{(i)}$ $H_2\, e^{(i)}$ $H_3\, e^{(i)}$ ---- ----------------------------- ----------------------------- ----------------------------- ----------------------------- 1 $-\delta - 2 {\varepsilon}$ $-\delta + 2 {\varepsilon}$ $\delta$ $\delta$ 2 $-\delta + 2 {\varepsilon}$ $-\delta - 2 {\varepsilon}$ $\delta$ $\delta$ 3 $\delta$ $\delta$ $-\delta - 2 {\varepsilon}$ $-\delta + 2 {\varepsilon}$ 4 $\delta$ $\delta$ $-\delta + 2 {\varepsilon}$ $-\delta - 2 {\varepsilon}$ 5 $-\delta - 2 {\varepsilon}$ $\delta$ $\delta$ $-\delta + 2 {\varepsilon}$ 6 $-\delta + 2 {\varepsilon}$ $\delta$ $\delta$ $-\delta - 2 {\varepsilon}$ 7 $\delta$ $-\delta + 2 {\varepsilon}$ $-\delta - 2 {\varepsilon}$ $\delta$ 8 $\delta$ $-\delta - 2 {\varepsilon}$ $-\delta + 2 {\varepsilon}$ $\delta$ 9 $-\delta - 2 {\varepsilon}$ $\delta$ $-\delta + 2 {\varepsilon}$ $\delta$ 10 $-\delta + 2 {\varepsilon}$ $\delta$ $-\delta - 2 {\varepsilon}$ $\delta$ 11 $\delta$ $-\delta + 2 {\varepsilon}$ $\delta$ $-\delta - 2 {\varepsilon}$ 12 $\delta$ $-\delta - 2 {\varepsilon}$ $\delta$ $-\delta + 2 {\varepsilon}$ : The images of the transformation strains under the linear functionals $H_i$, $i=0,1,2,3$, along with their extremisers [@Chenchiah-Schloemerkemper1ESM Linear\_Functionals.nb].[]{data-label="tab:H"} $i \mapsto H_0\, e^{(i)}$ $i \mapsto H_1\, e^{(i)}$ $i \mapsto H_2\, e^{(i)}$ $i \mapsto H_3\, e^{(i)}$ ------------ --------------------------- --------------------------- --------------------------- --------------------------- minimisers 1,5,9 2,8,12 3,7,10 4,6,11 : The images of the transformation strains under the linear functionals $H_i$, $i=0,1,2,3$, along with their extremisers [@Chenchiah-Schloemerkemper1ESM Linear\_Functionals.nb].[]{data-label="tab:H"} Maximisers $i \mapsto H_0\, e^{(i)}$ $i \mapsto H_1\, e^{(i)}$ $i \mapsto H_2\, e^{(i)}$ $i \mapsto H_3\, e^{(i)}$ ------------------------------ ------------------------------------------------- -------------------------------------------------- -------------------------------------------------- -------------------------------------------------- When ${\varepsilon}< \delta$ 3,4,7,8,11,12 3,4,5,6,9,10 1,2,5,6,11,12 1,2,7,8,9,10 When ${\varepsilon}= \delta$ ${\ensuremath{\mathcal{I}}}\setminus \{1,5,9\}$ ${\ensuremath{\mathcal{I}}}\setminus \{2,8,12\}$ ${\ensuremath{\mathcal{I}}}\setminus \{3,7,10\}$ ${\ensuremath{\mathcal{I}}}\setminus \{4,6,11\}$ When ${\varepsilon}> \delta$ 2,6,10 1,7,11 4,8,9 3,5,12 : The images of the transformation strains under the linear functionals $H_i$, $i=0,1,2,3$, along with their extremisers [@Chenchiah-Schloemerkemper1ESM Linear\_Functionals.nb].[]{data-label="tab:H"} #### Vertices. The vertices of ${\ensuremath{\mathcal{C}}}({\ensuremath{\mathcal{E}}})$ are the transformation strains: \[lem:verticesofC\] The set of vertices of ${\ensuremath{\mathcal{C}}}({\ensuremath{\mathcal{E}}})$ is ${\ensuremath{\mathcal{E}}}$. We show this explicitly for $e^{(1)}$, the proof for the other vertices follows from symmetry. Let $\lambda \in \Lambda$ such that $e^{(1)} = \sum_{i=1}^{12} \lambda_i e^{(i)}$. Then, $$\beta = H_{33} e^{(1)} = \sum_{i=1}^{12} \lambda_i H_{33} e^{(i)} = \sum_{i=1}^{4} \lambda_i \beta + \sum_{i=5}^{12} \lambda_i \alpha,$$ from which we conclude that $\lambda_i = 0$ for $i=5,\dots,12$ and thus $e^{(1)} = \sum_{i=1}^{4} \lambda_i e^{(i)}$. Now, using $H_0$ we obtain $$-\delta - 2\epsilon = H_0 e^{(1)} = \sum_{i=1}^{4} \lambda_i H_0 e^{(i)}.$$ Since ${\varepsilon}, \delta > 0$ it is easy to see (cf. Table \[tab:H\]) that $\lambda_1 = 1$ and $\lambda_i = 0$, $i=2,3,4$. We conclude that $e^{(1)}$ is a vertex. In the interest of brevity in future proofs of extremality of subsets we will only name the relevant family of four-dimensional hyperplanes, eg. for the above lemma we would say that this follows from $H_{33}$ and $H_0$. #### Edges. Contrary to what we are used to in two and three dimensions (see Remark \[rem:local1\] below), the convex hull of every pair of vertices is an edge of ${\ensuremath{\mathcal{C}}}({\ensuremath{\mathcal{E}}})$: \[lem:edges\] The set of edges of ${\ensuremath{\mathcal{C}}}({\ensuremath{\mathcal{E}}})$ is $\{ [e,f]\ \|\ e,f \in {\ensuremath{\mathcal{E}}}\}$. Henceforth, a compatible edge is the convex hull of a pair of compatible vertices, and an incompatible edge is the convex hull of a pair of incompatible vertices. We prove Lemma \[lem:edges\] here except for incompatible edges when ${\varepsilon}> \delta$. In Remark \[rem:edges\] we present a Mathematica-aided proof which is valid for ${\varepsilon}\neq \delta$. By symmetry it suffices to prove that the eleven edges $[e^{(1)},e^{(i)}]$, $i \in {\ensuremath{\mathcal{I}}}\setminus \{1\}$, are extremal. This is easy to verify for the compatible edges: For example, $H_{33}$ and $H_{12}$ show that $[e^{(1)},e^{(2)}]$ is extremal. The proof for the other compatible edges (with $e^{(1)}$ as a vertex) is similar. We now turn to the incompatible edges, for example $[e^{(1)},e^{(6)}]$. Let $\mu \in [0,1]$ and $\lambda \in \Lambda$ such that $$\mu e^{(1)} + (1-\mu) e^{(6)} = \sum_{i=1}^{12} \lambda_i e^{(i)}.$$ Consider first the case ${\varepsilon}< \delta$. Then $H_{11}$ and $H_2$ show that in fact $$\mu e^{(1)} + (1-\mu) e^{(6)} = \sum_{i \in \{1,2,5,6\}} \lambda_i e^{(i)}.$$ However $\dim {\ensuremath{\operatorname{aff}}}{\ensuremath{\operatorname{span}}}\{ e^{(1)},e^{(2)},e^{(5)},e^{(6)} \} = 3$ [@Chenchiah-Schloemerkemper1ESM Dimension\_Calculations.nb] and thus\ ${\ensuremath{\mathcal{C}}}(\{e^{(1)},e^{(2)},e^{(5)},e^{(6)}\})$ is a three-dimensional tetrahedron. It follows that $\lambda_5 = \lambda_2 = 0$ and thus $[e^{(1)},e^{(6)}]$ is extremal. Next consider the case ${\varepsilon}= \delta$. Then $H_{11}$, $H_{13}$, $H_1$ and $H_2$ show that in fact $$\mu e^{(1)} + (1-\mu) e^{(6)} = \sum_{i \in \{1,5,6\}} \lambda_i e^{(i)}.$$ However $\dim {\ensuremath{\operatorname{aff}}}{\ensuremath{\operatorname{span}}}\{ e^{(1)},e^{(5)},e^{(6)} \} = 2$ [@Chenchiah-Schloemerkemper1ESM Dimension\_Calculations.nb] and thus\ ${\ensuremath{\mathcal{C}}}(\{e^{(1)},e^{(5)},e^{(6)}\})$ is a triangle. It follows that $\lambda_5 = 0$ and thus $[e^{(1)},e^{(6)}]$ is extremal. The extremality of the other incompatible edges (with $e^{(1)}$ as a vertex) follows from symmetry. \[rem:local1\] In dimensions less than four the only polytopes for which the convex hull of every pair of vertices is an edge are the $n$-tetrahedra (line segments, triangles and tetrahedra in dimensions one, two and three, respectively). However for every $n>3$ and every $d>n$ there exists an n-dimensional convex polytope with d vertices for which the convex hull of every pair of vertices is extremal. See, eg. [@Brondsted:1982 Section 13], [@Ewald:1996 Section 3] or [@Ziegler:1994 Corollary 0.8]. #### The facets of ${\ensuremath{\mathcal{C}}}({\ensuremath{\mathcal{E}}})$. The algorithm we use to determine the facets of ${\ensuremath{\mathcal{C}}}({\ensuremath{\mathcal{E}}})$ is as follows [@Chenchiah-Schloemerkemper1ESM Faceting.nb]. It assumes that the affine span of the set is five-dimensional, that the cardinality of the set is small and thus that computational efficiency is not a consideration. \[alg:facets\] 1. First we form a set of all four-dimensional tetrahedra with vertices in ${\ensuremath{\mathcal{E}}}$ as follows: 1. Pick all five-tuples from ${\ensuremath{\mathcal{E}}}$: $$\{ S \subset {\ensuremath{\mathcal{E}}}\ \|\ \#S = 5 \}.$$ 2. Discard those five-tuples whose affine span is not four dimensional: $${\ensuremath{\mathcal{G}}}_1 := \{ S \subset {\ensuremath{\mathcal{E}}}\ \|\ \#S = 5, \dim {\ensuremath{\operatorname{aff}}}{\ensuremath{\operatorname{span}}}(S) = 4 \}.$$ 2. Of these tetrahedra we discard those whose convex hull is not contained in $\partial {\ensuremath{\mathcal{C}}}({\ensuremath{\mathcal{E}}})$. We do this as follows: 1. Let ${\ensuremath{\mathcal{G}}}_2 = {\ensuremath{\mathcal{G}}}_1$. 2. \[it:repeatS1\] Pick $S \in {\ensuremath{\mathcal{G}}}_2$. 3. Translate the origin to some $e \in S$. (The next two steps are carried out in this co-ordinate system.) 4. Compute a normal $n \in {\ensuremath{{\mathcal S}^{3 \times 3}_{\text{\ensuremath{2\alpha+\beta}}} }}$ to ${\ensuremath{\operatorname{aff}}}{\ensuremath{\operatorname{span}}}(S)$. 5. \[it:repeatE1\] If $\langle n, e \rangle$ has the same sign for all $e \in {\ensuremath{\mathcal{E}}}\setminus S$ then ${\ensuremath{\mathcal{C}}}(S) \subset \partial {\ensuremath{\mathcal{C}}}({\ensuremath{\mathcal{E}}})$. Otherwise remove $S$ from $G_2$ 6. Repeat steps to till all tetrahedra in ${\ensuremath{\mathcal{G}}}_2$ have been tested. We now obtain $${\ensuremath{\mathcal{G}}}_2 = \{ S \subset {\ensuremath{\mathcal{E}}}\ \|\ \#S = 5, \dim {\ensuremath{\operatorname{aff}}}{\ensuremath{\operatorname{span}}}(S) = 4, {\ensuremath{\mathcal{C}}}(S) \subset \partial {\ensuremath{\mathcal{C}}}({\ensuremath{\mathcal{E}}}) \}.$$ This is the set of all four-dimensional tetrahedra (with vertices in ${\ensuremath{\mathcal{E}}}$) whose union is $\partial {\ensuremath{\mathcal{C}}}({\ensuremath{\mathcal{E}}})$. 3. The final step is to form the facets of ${\ensuremath{\mathcal{C}}}({\ensuremath{\mathcal{E}}})$ by judiciously taking unions of sets in ${\ensuremath{\mathcal{G}}}_2$ as follows: 1. Let ${\ensuremath{\mathcal{G}}}_3 = {\ensuremath{\mathcal{G}}}_2$. 2. \[it:repeatS2\] Pick $S_1,S_2 \in {\ensuremath{\mathcal{G}}}_3$. 3. \[it:repeatE2\] If $\dim {\ensuremath{\operatorname{aff}}}{\ensuremath{\operatorname{span}}}(S_1 \cup S_2) = 4$ then $S_1$ and $S_2$ are parts of the same facet. In ${\ensuremath{\mathcal{G}}}_3$, replace $S_1$ and $S_2$ by $S_1 \cup S_2$. 4. Repeat steps and until every union of sets in $G_3$ increases the dimension, i.e., until it is true that $$\label{eq:local1} \forall S_1, S_2 \in {\ensuremath{\mathcal{G}}}_3, \qquad \dim {\ensuremath{\operatorname{aff}}}{\ensuremath{\operatorname{span}}}(S_1 \cup S_2) =4 \implies S_1 = S_2.$$ This is the set of all the facets of ${\ensuremath{\mathcal{C}}}({\ensuremath{\mathcal{E}}})$. Note that it is a set of $n$-tuples where $n \geqslant 5$. The results of a Mathematica implementation of Algorithm \[alg:facets\] are summarised in Observations \[obs:e<d\], \[obs:e=d\] and \[obs:e>d\] below. These reveal that the facet structure depends on whether ${\varepsilon}< \delta$, ${\varepsilon}= \delta$ or ${\varepsilon}> \delta$. All three possibilities are realisable in that there exist cubic-to-monoclinic-I phase transformations corresponding to each. (See, e.g., [@Bhattacharya:2003 Fig. 4.3 and (4.11) on p. 52–53] for the relationship between ${\varepsilon}$, $\delta$ and the unit cells of the cubic and monoclinic lattices.) However, curiously ${\varepsilon}< \delta$ for all the monoclinic-I materials of which we are aware, cf. Table \[tab:parameters\]; we return to this point in Section \[sec:conclusions\]. In the observations below, each group of facets is the orbit under $S_4$ of any facet in it [@Chenchiah-Schloemerkemper1ESM Facet\_Symmetry.nb]. Within each group the facets are listed in lexical order. Facets that occur for both ${\varepsilon}< \delta$ and ${\varepsilon}> \delta$ are shown in bold face. \[obs:e<d\] When ${\varepsilon}< \delta$, the 25 four-dimensional facets of ${\ensuremath{\mathcal{C}}}({\ensuremath{\mathcal{E}}})$ consist of the convex hulls of 1. 12 facets with 5 vertices each: \[it:local12\] $$\begin{aligned} {4} & \{1, 2, 3, 7, 10\}, \ && \{1, 2, 4, 6, 11\}, \ && \{1, 3, 4, 5, 9\}, \ && \{1, 5, 7, 8, 9\}, \\ & \{1, 5, 9, 11, 12\}, \ && \{2, 3, 4, 8, 12\}, \ && \{2, 5, 6, 8, 12\}, \ && \{2, 8, 9, 10, 12\},\\ & \{3, 5, 6, 7, 10\}, \ && \{3, 7, 10, 11, 12\}, \ && \{4, 6, 7, 8, 11\}, \ && \{4, 6, 9, 10, 11\}; \end{aligned}$$ 2. 4 pairs of $T_3$s (see Section \[sec:nonlaminates\]), each facet is invariant under $r_0$: $$\label{eq:T3-pair} \{1, 2, 5, 6, 11, 12\}, \ \{1, 2, 7, 8, 9, 10\}, \ \{3, 4, 5, 6, 9, 10\}, \ \{3, 4, 7, 8, 11, 12\}$$ 3. 6 pairs of pairwise compatible three-tuples: \[it:local13\] $$\begin{aligned} {3} & \mathbf{\{1, 2, 5, 8, 9, 12\}}, \ && \mathbf{\{1, 3, 5, 7, 9, 10\}}, \ && \mathbf{\{1, 4, 5, 6, 9, 11\}}, \\ & \mathbf{\{2, 3, 7, 8, 10, 12\}},\ && \mathbf{\{2, 4, 6, 8, 11, 12\}}, \ && \mathbf{\{3, 4, 6, 7, 10, 11\}}; \end{aligned}$$ 4. 3 facets with 8 vertices each; each facet is invariant under $r_0$: $$\label{eq:8vertex-facets} \mathbf{\{1, 2, 3, 4, 5, 6, 7, 8\}}, \ \mathbf{\{1, 2, 3, 4, 9, 10, 11, 12\}}, \ \mathbf{\{5, 6, 7, 8, 9, 10, 11, 12\}}.$$ \[obs:e=d\] When ${\varepsilon}= \delta$, the 7 four-dimensional facets of ${\ensuremath{\mathcal{C}}}({\ensuremath{\mathcal{E}}})$ consist of the convex hulls of 1. 4 facets with 9 vertices each: $$\begin{aligned} {2} & \{1, 2, 3, 5, 7, 8, 9, 10, 12\}, \ && \{1, 2, 4, 5, 6, 8, 9, 11, 12\}, \notag \\ & \{1, 3, 4, 5, 6, 7, 9, 10, 11\}, \ && \{2, 3, 4, 6, 7, 8, 10, 11, 12\}; \label{eq:9vertex-facets} \end{aligned}$$ 2. 3 facets with 8 vertices each; each facet is invariant under $r_0$: $$\mathbf{\{1, 2, 3, 4, 5, 6, 7, 8\}}, \ \mathbf{\{1, 2, 3, 4, 9, 10, 11, 12\}}, \ \mathbf{\{5, 6, 7, 8, 9, 10, 11, 12\}}.\tag{\ref{eq:8vertex-facets}}$$ \[obs:e>d\] When ${\varepsilon}> \delta$, the 33 four-dimensional facets of ${\ensuremath{\mathcal{C}}}({\ensuremath{\mathcal{E}}})$ consist of the convex hulls of 1. 12 facets which are the images under $r_0$ of the five-vertex facets that occur when ${\varepsilon}< \delta$: $$\begin{aligned} {4} & \{1, 2, 3, 5, 12\}, \ && \{1, 2, 4, 8, 9\}, \ && \{1, 3, 4, 7, 11\}, \ && \{1, 5, 6, 7, 11\}, \\ & \{1, 7, 9, 10, 11\}, \ && \{2, 3, 4, 6, 10\}, \ && \{2, 6, 7, 8, 10\}, \ && \{2, 6, 10, 11, 12\}, \\ & \{3, 5, 7, 8, 12\}, \ &&\{3, 5, 9, 10, 12\}, \ && \{4, 5, 6, 8, 9\}, \ && \{4, 8, 9, 11, 12\}; \end{aligned}$$ 2. 12 other five vertex facets (together these are invariant under $r_0$): $$\begin{aligned} {4} & \{1, 3, 5, 9, 12\}, \ && \{1, 3, 7, 10, 11\}, \ && \{1, 4, 5, 8, 9\}, \ && \{1, 4, 6, 7, 11\}, \\ & \{1, 5, 7, 9, 11\}, \ && \{2, 3, 5, 8, 12\}, \ && \{2, 3, 6, 7, 10\}, \ && \{2, 4, 6, 10, 11\}, \\ & \{2, 4, 8, 9, 12\}, \ && \{2, 6, 8, 10, 12\}, \ && \{3, 5, 7, 10, 12\}, \ && \{4, 6, 8, 9, 11\};\end{aligned}$$ 3. 6 pairs of pairwise compatible three-tuples: $$\begin{aligned} {3} & \mathbf{\{1, 2, 5, 8, 9, 12\}}, \ && \mathbf{\{1, 3, 5, 7, 9, 10\}}, \ && \mathbf{\{1, 4, 5, 6, 9, 11\}}, \\ & \mathbf{\{2, 3, 7, 8, 10, 12\}},\ && \mathbf{\{2, 4, 6, 8, 11, 12\}}, \ && \mathbf{\{3, 4, 6, 7, 10, 11\}}; \end{aligned}$$ 4. 3 facets with 8 vertices each; each facet is invariant under $r_0$: $$\mathbf{\{1, 2, 3, 4, 5, 6, 7, 8\}}, \ \mathbf{\{1, 2, 3, 4, 9, 10, 11, 12\}}, \ \mathbf{\{5, 6, 7, 8, 9, 10, 11, 12\}}. \tag{\ref{eq:8vertex-facets}}$$ The extremality of facets that are invariant under $r_0$ can be verified as in the proof of Lemma \[lem:verticesofC\]: $H_{11}, H_{22}, H_{33}$ show that the facets in are extremal. $H_i$, $i=0,1,2,3$, show that pairs of $T_3$s in are extremal since ${\varepsilon}< \delta$ implies $-\delta-2{\varepsilon}< -\delta < \delta-2{\varepsilon}< \delta < \delta+2{\varepsilon}$. In addition $H_i$, $i=0,1,2,3$, also show that the facets in are extremal, cf. Table \[tab:H\]. For the remaining facets extremality can be verified through a computation of normals [@Chenchiah-Schloemerkemper1ESM Facet\_Normal.nb]. \[rem:edges\] Observations \[obs:e<d\] and \[obs:e>d\] lead to a proof of Lemma \[lem:edges\] when ${\varepsilon}\neq \delta$: Observe that every edge is shared by at least four facets. (This is particularly easy to check for the incompatible edges: Each incompatible edge is contained in precisely one facet from each group.) It follows that every edge is extremal (cf. eg. [@Barvinok:2002 Chapter 6] or the other references listed in Section \[sec:convexity\]). We remark that this phenomenon of polytope facet structure depending on lattice parameters is not possible for the other martensites (i.e., cubic-to-tetragonal, cubic-to-trigonal and cubic-to-orthorhombic) because they form $n$-tetrahedra (for $n=2,3,5$, respectively) and thus their facet structure is fixed. Non-laminate microstructures in ${\ensuremath{\mathcal{R}}}({\ensuremath{\mathcal{E}}})$ {#sec:nonlaminates} ======================================================================================== In this section we use the results of the preceding sections to derive our central results about non-laminate microstructures in monoclinic-I martensite. These include $T_3$ microstructures formed by the strains in ${\ensuremath{\mathcal{E}}}$ (Section \[sec:T3s-level1\]) and $T_3$ microstructures formed by the nodes of these $T_3$s (Section \[sec:T3s-level2\]). (Level-1) $T_3$s and related microstructures {#sec:T3s-level1} -------------------------------------------- As mentioned earlier there are precisely eight 3-tuples of pairwise incompatible vertices (cf. Table \[tab:compatibility\]): $$\begin{gathered} {\ensuremath{\mathcal{E}}}^3_{{\ensuremath{\cancel{\interleave}}}} = \left\{ \{ e^{(1)}, e^{(6)}, e^{(12)} \},\ \{ e^{(1)}, e^{(8)}, e^{(10)} \},\ \{ e^{(2)}, e^{(5)}, e^{(11)} \},\ \{ e^{(2)}, e^{(7)}, e^{(9)} \}, \right. \\ \left. \{ e^{(3)}, e^{(6)}, e^{(9)} \},\ \{ e^{(3)}, e^{(8)}, e^{(11)} \},\ \{ e^{(4)}, e^{(5)}, e^{(10)} \},\ \{ e^{(4)}, e^{(7)}, e^{(12)} \} \right\}. $$ Since $${\ensuremath{\operatorname{sign}}}\det( e^{(1)} - e^{(6)} ) = {\ensuremath{\operatorname{sign}}}\det( e^{(6)} - e^{(12)} ) = {\ensuremath{\operatorname{sign}}}\det( e^{(12)} - e^{(1)} ) \neq 0$$ and likewise for the other 3-tuples, cf. Table \[tab:compatibility\], we obtain by Lemma \[lem:T3\] that each of these 3-tuples forms a $T_3$. (See also Remark \[rem:incompatible-distances\].) Let $v_{i,j,k} = \{ e^{(i)}, e^{(j)}, e^{(k)} \} \in {\ensuremath{\mathcal{E}}}^3_{{\ensuremath{\cancel{\interleave}}}}$. We set $$\begin{aligned} \tau_{i,j,k} &= {\ensuremath{\mathcal{R}}}(v_{i,j,k}), \\ {\ensuremath{\mathcal{T}}}&:= \{ \tau_{i,j,k}\ \|\ v_{i,j,k} \in {\ensuremath{\mathcal{E}}}^3_{{\ensuremath{\cancel{\interleave}}}} \}.\end{aligned}$$ For $\tau \in {\ensuremath{\mathcal{T}}}$ and $r \in S_4 \times C_2$ by $r \tau$ we mean the $T_3$ formed by the image under $r$ of the vertices of $\tau$. (The existence of such $T_3$s follows from $S_4 \times C_2$ being a symmetry group of ${\ensuremath{\mathcal{E}}}^3_{{\ensuremath{\cancel{\interleave}}}}$ as was shown in Lemma \[lem:symmetry2\]). The symmetry relations between the $T_3$s is illustrated in Figure \[fig:T3-symmety\]. As we shall see (Example \[eg:toblerone\] below) each $\tau \in {\ensuremath{\mathcal{T}}}$ is specially related to $r_0 \tau$; we refer to it as the dual of $\tau$. symmetryT3s\_one.pdf\_t symmetryT3s\_three.pdf\_t Before we proceed further we note that each of these eight $T_3$s is symmetric (Definition \[def:T3-symmetry\]) and has distinct nodes (Definition \[def:T3-nodes\]). Moreover all eight $T_3$s are similar (Definition \[def:T3-similarity\]): \[lem:T3-equivalence\] All $T_3$s in ${\ensuremath{\mathcal{T}}}$ are similar, and each is symmetric and has distinct nodes. The symmetry of each $T_3$ and the similarity of all $T_3$s in ${\ensuremath{\mathcal{T}}}$ follow from Remark \[rem:incompatible-distances\]. Now consider the nodes of $\tau_{1,8,10}$. Suppose, on the contrary, that $e_{1,8} = e_{8,10} = e_{10,1}$. (Here we use the notation of Section \[sec:T3\].) Then, by symmetry, the nodes coincide with the barycentre of the $T_3$ which is $\frac{1}{3} \left( e^{(1)} + e^{(8)} + e^{(10)} \right)$. However, an elementary calculation shows that this barycentre is incompatible with $e_1$, $e_8$ and $e_{10}$ which is a contradiction (cf. Definition \[def:T3-1\]). Thus $\tau_{1,8,10}$ has distinct nodes and by symmetry this is true for all eight $T_3$s. From Proposition \[prop:T3-hull\] and symmetry it follows that the barycentre of each $v \in {\ensuremath{\mathcal{E}}}^3_{{\ensuremath{\cancel{\interleave}}}}$ is contained in ${\ensuremath{\mathcal{R}}}(v)$. From Observations \[obs:e<d\] and \[obs:e=d\], when ${\varepsilon}\leqslant \delta$, each $T_3$ in ${\ensuremath{\mathcal{T}}}$ along with its dual is contained in a facet of ${\ensuremath{\mathcal{C}}}({\ensuremath{\mathcal{E}}})$. In particular for each $\tau \in {\ensuremath{\mathcal{T}}}$, the symmetrised rank-one convex hull of $\tau \cup r_0 \tau$ is contained in a facet of ${\ensuremath{\mathcal{C}}}({\ensuremath{\mathcal{E}}})$ and is thus a part of the boundary of the symmetrised rank-one convex hull of ${\ensuremath{\mathcal{E}}}$: $$\label{eq:T3-pairs-are-on-boundary} {\ensuremath{\mathcal{R}}}(\tau \cup r_0 \tau) \subset \partial \mathcal{R}({\ensuremath{\mathcal{E}}}).$$ Example \[eg:toblerone\] below reveals that ${\ensuremath{\mathcal{R}}}(\tau \cup r_0 \tau)$ is four dimensional and that each point in it can be attained by laminates of $T_3$ microstructures. In other words ${\ensuremath{\mathcal{R}}}(\tau \cup r_0 \tau)$ contains a four-dimensional set of $T_3$s. Before we show this we first construct a three-dimensional set of $T_3$s which lies in $\partial \mathcal{R}({\ensuremath{\mathcal{E}}})$ when ${\varepsilon}\leqslant \delta$. \[eg:3D-set-T3s\] Let $\tau \in {\ensuremath{\mathcal{T}}}$. From Table \[tab:compatibility\] the elements of ${\ensuremath{\mathcal{E}}}$ with which all the vertices of $\tau$ are compatible are precisely the vertices of $r_0 \tau$. Following the construction introduced in Lemma \[lem:T3-3D\], we construct a three-dimensional set of $T_3$s from $\tau$ and (any) one vertex from $r_0 \tau$. There are three such continua of $T_3$s, one for each vertex of $r_o \tau$. Analogously there are three such sets constructed from $r_0 \tau$ and the vertices of $\tau$. Note that this example shows that $T_3$s exist not only in ${\ensuremath{\mathcal{E}}}$ but also in ${\ensuremath{\mathcal{L}}}_1({\ensuremath{\mathcal{E}}}) \setminus {\ensuremath{\mathcal{E}}}$ since the vertices of almost all of the $T_3$s so constructed are themselves attained by a lamination of strains in ${\ensuremath{\mathcal{E}}}$. However in the next example the construction of $T_3$s precedes the construction of laminates: \[eg:toblerone\] From each $\tau \in {\ensuremath{\mathcal{T}}}$ and its dual we construct a four-dimensional set of laminates of nodes of $T_3$s. We do this explicitly for $\tau_{1,8,10}$ and its dual $\tau_{2,7,9}$; the construction for the other pairs is similar. First we note that each point in one of these $T_3$s is compatible with the corresponding point (i.e., the point with the same barycentric coordinates) in its dual. In fact this is true even after a cyclic permutation of the vertices: $\forall x,y,z \in {\mathbb{R}}$, $$x e^{(1)} + y e^{(8)} + z e^{(10)} \, {\ensuremath{\interleave}}\, x e^{(2)} + y e^{(9)} + z e^{(7)}, \, x e^{(7)} + y e^{(2)} + z e^{(9)}, \, x e^{(9)} + y e^{(7)} + z e^{(2)}.$$ This is immediate from the calculation [@Chenchiah-Schloemerkemper1ESM Pair\_of\_Level-1\_T3s.nb]: $$\begin{aligned} &\det \left( ( x e^{(1)} + y e^{(8)} + z e^{(10)} ) - ( x e^{(2)} + y e^{(9)} + z e^{(7)} ) \right) \\ &= \det \left( ( x e^{(1)} + y e^{(8)} + z e^{(10)} ) - ( x e^{(7)} + y e^{(2)} + z e^{(9)} ) \right) \\ &= \det \left( ( x e^{(1)} + y e^{(8)} + z e^{(10)} ) - ( x e^{(9)} + y e^{(7)} + z e^{(2)} ) \right) \\ &= 0.\end{aligned}$$ Thus, in particular, each node of $\tau_{1,8,10}$ is compatible with every node of $\tau_{2,7,9}$ (and vice versa). Since the nodes of a $T_3$ are pair-wise compatible it follows that these six nodes (i.e. in the notation of Section \[sec:T3\], $e_{1,8}$, $e_{8,10}$, $e_{10,1}$, $e_{2,9}$, $e_{9,7}$ and $e_{7,2}$) are pair-wise compatible (see Figure \[fig:toblerone\], note that the figure is schematic; in fact $\tau$ and $r_0 \tau$ share the same barycenter). We conclude that: $${\ensuremath{\mathcal{C}}}(\{ e_{1,8}, e_{8,10}, e_{10,1}, e_{2,9}, e_{9,7}, e_{7,2} \}) \subset {\ensuremath{\mathcal{R}}}(\{ e_1, e_2, e_7, e_8, e_9, e_{10} \}).$$ Each point in this convex hull is attained by a lamination of the nodes of $\tau_{1,8,10}$ and $\tau_{2,7,9}$. A simple Mathematica verification [@Chenchiah-Schloemerkemper1ESM Dimension.nb] shows that this convex hull is four-dimensional. (Thus the maximum depth of lamination required is also four, cf. proof of Theorem \[thm:edge-compatible\].) Note that when ${\varepsilon}\leqslant \delta$, from , $$\label{eq:toberone-is-on-boundary} {\ensuremath{\mathcal{C}}}(\{ e_{1,8}, e_{8,10}, e_{10,1}, e_{2,9}, e_{9,7}, e_{7,2} \}) \subset \partial \mathcal{R}({\ensuremath{\mathcal{E}}}).$$ Level-2 $T_3$s and related microstructures {#sec:T3s-level2} ------------------------------------------ Next we construct new $T_3$s from the nodes of the $T_3$s in ${\ensuremath{\mathcal{T}}}$. We refer to the former $T_3$s as Level-1 $T_3$s and to the new $T_3$s as Level-2 $T_3$s. Level-2 $T_3$s allow us to construct a five-dimensional set of $T_3$s, see Construction \[cons:5D-set-T3s\] below. \[cons:level2-T3s\] Let $\tau \in {\ensuremath{\mathcal{T}}}$ and let $\tau_1,\tau_2,\tau_3 \in {\ensuremath{\mathcal{T}}}$ be chosen such that, in Figure \[fig:T3-symmety\], the line joining $\tau$ and $\tau_i$ is an edge of the cube for $i=1,2,3$ (thus $\tau_i = r_j^{\pm1} \tau$ for some $j=1,2,3$). Note that the set $\{\tau_1,\tau_2,\tau_3\}$ is invariant under any element of $S_4$ that leaves $\{\tau, r_0 \tau\}$ invariant (these are rotations of $\frac{2\pi}{3}$ through the major diagonal formed by $\tau$ and $r_0 \tau$). Let $r$ be an element of this group (i.e., one of two such rotations). Now let $n_1$ be a node of $\tau_1$. Then $n_1$, $r n_1$, $r^2 n_1$ form a symmetric $T_3$ with distinct nodes. Similarly for $n_2$, $r n_2$, $r^2 n_2$ and $n_3$, $r n_3$, $r^2 n_3$. This is illustrated in Figure \[fig:level2-T3s\]. We show that $n_1$, $r n_1$, $r^2 n_1$ form a $T_3$ explicitly for $\tau = \tau_{3,8,11}$, $\tau_1 = \tau_{1,8,10}$, $\tau_2 = \tau_{2,5,11}$ and $\tau_3 = \tau_{3,6,9}$; the result for the other 3-tuples follows by symmetry. Since $n_1 \in \tau_{1,8,10}$, it has the barycentric representation $x e^{(1)} + y e^{(8)} + z e^{(10)}$ for some $x,y,z \in [0,1]$ with $x+y+z=1$. Then, as can be easily checked (Table \[tab:generators\]), $$\{r n_1, r^2 n_1 \} = \{x e^{(5)} + y e^{(11)} + z e^{(2)}, x e^{(9)} + y e^{(3)} + z e^{(6)} \}.$$ (Note the order of the vertices.) It can be verified [@Chenchiah-Schloemerkemper1ESM Level-2\_T3s.nb] that $$\begin{aligned} &\det \left( ( x e^{(1)} + y e^{(8)} + z e^{(10)} ) - ( x e^{(5)} + y e^{(11)} + z e^{(2)} ) \right) \\ &= \det \left( ( x e^{(5)} + y e^{(11)} + z e^{(2)} ) - ( x e^{(9)} + y e^{(3)} + z e^{(6)} ) \right) \\ &= \det \left( ( x e^{(9)} + y e^{(3)} + z e^{(6)} ) - ( x e^{(1)} + y e^{(8)} + z e^{(10)} ) \right) \\ &\neq 0.\end{aligned}$$ Thus, by Lemma \[lem:T3\], Remark \[rem:T3-symmetry\] and , $n_1$, $r n_1$, $r^2 n_1$ form a symmetric $T_3$. (From Remark \[rem:T3-symmetry\] it would have sufficed to check that one of the determinants above is non-zero.) To show that the nodes of this $T_3$ are distinct it suffices to check that the barycentre of the $T_3$ is incompatible with (one of) its nodes (cf. proof of Lemma \[lem:T3-equivalence\]). From [@Chenchiah-Schloemerkemper1ESM Level-2\_T3s.nb]: $$\begin{gathered} x e^{(1)} + y e^{(8)} + z e^{(10)},\ x e^{(5)} + y e^{(11)} + z e^{(2)},\ x e^{(9)} + y e^{(3)} + z e^{(6)} \\ {\ensuremath{\cancel{\interleave}}}\frac{1}{3} \left( ( x e^{(1)} + y e^{(8)} + z e^{(10)} ) + ( x e^{(5)} + y e^{(11)} + z e^{(2)} ) + ( x e^{(9)} + y e^{(3)} + z e^{(6)} ) \right).\end{gathered}$$ Which completes the proof. Since there are eight choices of $\tau$ and three choices of $n_1$ for each choice of $\tau$, by this construction we obtain 24 $T_3$s. Finally Example \[eg:toblerone\] and Construction \[cons:level2-T3s\] can be combined to construct a five-dimensional set of $T_3$s whose vertices are themselves laminates of nodes of (level-1) $T_3$s. \[cons:5D-set-T3s\] Let $\tau \in {\ensuremath{\mathcal{T}}}$ and let $\tau_1, \tau_2, \tau_3 \in {\ensuremath{\mathcal{T}}}$ and $r \in S_4$ be as in Construction \[cons:level2-T3s\] above. Let $n_{1,i}$ and $n'_{1,i}$, $i=1,2,3$ be the nodes of $\tau_1$ and its dual respectively. Pick $\mu_i, \mu'_i \in [0,1]$ such that $\sum_{i=1}^3 \mu_i + \sum_{i=1}^3 \mu'_i = 1$ and let $$p_1 = \sum_{i=1}^3 \mu_i n_{1,i}+ \sum_{i=1}^3 \mu'_i n'_{1,i}.$$ Note that $p_1$ is an element of the four-dimensional set constructed in Example \[eg:toblerone\] and thus can be attained by a lamination of $n_{1,i}, n'_{1,i}$, $i=1,2,3$. We assert that $p_1$, $r p_1$, $r^2 p_1$ form a symmetric $T_3$. The union of the $T_3$s as $p_1$ varies yields a five-dimensional set. When ${\varepsilon}\leqslant \delta$ from , this set intersects the boundary of the symmetrised rank-one convex hull of ${\ensuremath{\mathcal{E}}}$. A Mathematica calculation [@Chenchiah-Schloemerkemper1ESM Level-2\_T3s.nb] shows this explicitly for $\tau_1 = \tau_{1,8,10}$, $\tau_2 = \tau_{2,5,11}$ and $\tau_3 = \tau_{3,6,9}$ (i.e., $\tau = \tau_{3,8,11}$), the construction for the other 3-tuples is similar. To see that the set of $T_3$s constructed here is five-dimensional, it suffices to note that $p_1$ is picked from a four-dimensional set which, not being closed under $r$, does not contain the $T_3$s formed by $p_1$, $r p_1$, $r^2 p_1$. It follows that the union of the $T_3$s (as $p_1$ varies) constructed here is a five-dimensional set. It is natural at this point to ask whether the nodes of level-2 $T_3$s form level-3 $T_3$ and, more generally, whether the nodes of level-$n$ $T_3$s from level-$(n+1)$ $T_3$s. We postpone these questions to [@Chenchiah-Schloemerkemper-JMPS] and instead conclude by considering some implications of the results presented here. Conclusions {#sec:conclusions} =========== Mathematical comments --------------------- While we have, in the later half of this paper, focused on monoclinic-I martensite, it is clear that our general strategy can, in principle, be applied to any finite set in ${\ensuremath{{\mathcal S}^{3 \times 3}_{\text{\ensuremath{c}}} }}$; indeed in [@Chenchiah-Schloemerkemper-JMPS] we apply it also to monoclinic-II martensite. Here we briefly comment on the two main components of our strategy, namely an understanding of the algebraic structure of symmetrised rank-one convex cones, and an understanding of the polytope structure of the given set. #### The algebraic structure of symmetrised rank-one convex cones. While we have a complete understanding of symmetrised rank-one convexity in two-dimensions (Sections \[sec:2D-cone\] and \[sec:T3\], and [@Chenchiah-Schloemerkemper-PRSL]), the algebraic structure of symmetrised rank-one convex cones is not yet sufficiently well understood in higher dimensions. A key missing ingredient is a characterisation, in terms of canonical forms, of real cubic polynomials in several variables. This problem in invariant theory seems unsolved for three and more variables. The fruitfulness of our approach in two-dimensions suggests that it might be valuable to more fully explore the algebraic aspects of symmetrised rank-one convexity. When ${\varepsilon}< \delta$ we demonstrate the existence of $T_3$s that attain points on the boundary of ${\ensuremath{\mathcal{C}}}({\ensuremath{\mathcal{E}}})$, more precisely, that attain points on the four-dimensional facets of ${\ensuremath{\mathcal{C}}}({\ensuremath{\mathcal{E}}})$. Though (as a Mathematica calculation [@Chenchiah-Schloemerkemper1ESM Facet-T3Pair.nb] shows the $T_3$s do not belong to the three-dimensional facets (of the four-dimensional facets) of ${\ensuremath{\mathcal{C}}}({\ensuremath{\mathcal{E}}})$, we suspect that the symmetrised rank-one convex hull of monoclinic-I martensite is strictly larger than the lamination hull. If so, the question arises as to how much larger it is. In terms of dimensions a perturbation argument shows that ${\ensuremath{\mathcal{R}}}({\ensuremath{\mathcal{E}}}) \setminus {\ensuremath{\mathcal{L}}}({\ensuremath{\mathcal{E}}})$ would be at least two-dimensional. In fact we suspect that it is full (i.e., five) dimensional. #### Convex polytopes. Since the convex hull of any finite set is a convex polytope, it is natural that an attempt to determine the semi-convex hull of a finite sets takes advantage of the structure of the convex polytopes they generate. This seems not to have been considered in the literature except for Theorem \[thm:edge-compatible\]. Lemma \[lem:edges\] and Remark \[rem:local1\] demonstrate the counter-intuitive behaviour of high-dimensional polytopes and thus the usefulness of knowledge of the theory of convex polytopes. Moreover, when, as in the example of monoclinc-I martensite considered here, the faceting structure of the polytope depends on the material parameters, it might be expected that qualitative features of the semi-convex hulls and envelopes depend on the material parameters as well. If so, this heightens the possible utility of these polytopes for evaluating semi-convex hulls and envelopes. Implications for mechanics -------------------------- #### Two kinds of monoclinic-I martensite. Curiously, all cubic-to-monoclinic-I materials that we are aware of are monoclinic-Ia martensites (i.e., those for which ${\varepsilon}<\delta$), see Table \[tab:parameters\]. It is natural to ask whether monoclinic-I martensite recovers more strains (modulo appropriate normalisation of the lattice parameters) as ${\varepsilon}-\delta$ approaches zero (with ${\varepsilon}=\delta$ being the ideal), and whether monoclinic-Ib martensites (i.e., those for which ${\varepsilon}>\delta$) would demonstrate greater shape memory effect. #### Monoclinic-II martensite. We have reason to believe that there are multiple kinds of monoclinic-II martensite as well. We hope to settle this question in [@Chenchiah-Schloemerkemper-JMPS]. As can be easily verified (eg. [@Bhattacharya:2003]), the compatibility relations between the twelve transformation strains of monoclinic-II martensite are identical to those between the twelve transformation strains of monoclinic-I martensite. Then, with the help of Lemma \[lem:T3\], exactly as for monoclinic-I martensite, eight $T_3$s can be formed from these strains. This positively answers the question raised in [@Bhattacharya:1994p843 p863] as to whether $T_3$s can be formed from the twelve transformation strains of monoclinic-II martensite. Indeed Lemma \[lem:T3\] presents an elementary test by which this question can be answered for any three-tuple of strains that have the same trace. #### Special parameters. Throughout this paper, including in the (non-numerical) Mathematica computations, we have assumed the lattice parameters to be generic except that we considered the case ${\varepsilon}=\delta$. We did identify one special case, $(\alpha-\beta) \delta + {\varepsilon}^2 - \delta^2 = 0$, in which all twelve strains of monoclinic-I martensite are pair-wise compatible. The question arises as to whether a material recovers more strains as $(\alpha-\beta) \delta + {\varepsilon}^2 - \delta^2$ approaches zero. We suggest another parameter of importance, $\lambda := \lambda_{12} = \lambda_{23} = \lambda_{31}$ (Definition \[def:T3-symmetry\]) which however appears related to $(\alpha-\beta) \delta + {\varepsilon}^2 - \delta^2$ for the three materials considered in Table \[tab:special-parameters\]. Material $\lambda$ $(\alpha-\beta) \delta + {\varepsilon}^2 - \delta^2$ ---------- ----------- ------------------------------------------------------ NiTi $0.6830$ $0.0024$ TiNiCu $0.6683$ $0.0021$ CuZr $0.0396$ $-0.0015$ : $\lambda$ and $(\alpha-\beta) \delta + {\varepsilon}^2 - \delta^2$ for the $T_3$s in ${\ensuremath{\mathcal{T}}}$ for NiTi, CuZr and TiNiCu.[]{data-label="tab:special-parameters"} Our reasoning is that as $\lambda$ becomes close to either $0$ or $1$, the nodes of a $T_3$ become closer to its vertices and the energetic penalty for a $T_3$ microstructure being approximated by a finite-rank laminate becomes smaller. Indeed, instead of constructing a finite-rank laminate from the three nodes of a $T_3$, it suffices to move only one of the vertices to a node. As Table \[tab:special-parameters\] shows $\lambda = 0.0396$ for CuZr. We hypothesise that for this material the symmetrised lamination convex hull is very close to the convex hull and thus the convex hull is a very good approximation to all its semi-convex hulls. The same would apply to the (semi-)convex envelopes of the corresponding energy density as well. A similar reasoning might explain the remarkable closeness between the symmetrised lamination convex hull and the convex hull for CuAlNi (a monoclinic-II martensite) observed in [@Govindjee:2007]. #### Microstructure corresponding to $T_3$s. For monoclinic-I martensite with ${\varepsilon}< \delta$ we have shown that there exists a five-dimensional set of $T_3$s which reaches the boundary of ${\ensuremath{\mathcal{C}}}({\ensuremath{\mathcal{E}}})$. This raises the question as to whether the microstructures experimentally observed for strains in this set are laminate approximations to $T_3$ microstructures. We wonder too if experimental observations of such microstructures would provide insight into mechanisms governing microstructure formation (dynamics) and the role of surface energy. #### Acknowlegements. This research was supported by the Royal Society (International Joint Project “Symmetrised rank-one convex hull of monoclinic martensite”). It was initiated while the authors were at the Max Planck Institute for Mathematics in the Sciences, Leipzig and continued while AS was at the Department of Mathematics, University of Erlangen-Nuremberg, the Hausdorff Center for Mathematics and the Institute for Applied Mathematics, University of Bonn. IVC thanks these institutes and the Institute for Mathematics, University of Würzburg for hospitality during visits. IVC was introduced to the problem of determining the semi-convex hulls of monoclinic-I martensite when Sanjay Govindjee and Valery Smyshlyaev visited Kaushik Bhattacharya at Caltech. We thank Kaushik Bhattacharya, Richard James, Robert Kohn and Stefan Müller for helpful discussions. We thank the referees for helpful comments, for suggesting the name “wave convex”, and for bringing [@Tartar:1979; @Murat:1981] to our attention.	{ "pile_set_name": "ArXiv" }
--- abstract: 'We look at the limit distributions of sums of deterministic chaotic variables in unimodal maps and find a remarkable renormalization group (RG) structure associated to the operation of increment of summands and rescaling. In this structure - where the only relevant variable is the difference in control parameter from its value at the transition to chaos - the trivial fixed point is the Gaussian distribution and a novel nontrivial fixed point is a multifractal distribution that emulates the Feigenbaum attractor, and is universal in the sense of the latter. The crossover between the two fixed points is explained and the flow toward the trivial fixed point is seen to be comparable to the chaotic band merging sequence. We discuss the nature of the Central Limit Theorem for deterministic variables.' address: - '$^1$Santa Fe Institute, 1399 Hyde Park Road, Santa Fe, New Mexico 87501, USA' - '$^2$Centro Atómico Bariloche, Instituto Balseiro and CONICET, 8400 Bariloche, Argentina' - '$^3$ Center for Advanced Studies in Ecology and Biodiversity, Facultad de Ciencias Biológicas, Pontificia Universidad Católica de Chile, Casilla 114-D, Santiago CP 6513677, Chile' - '$^4$Instituto de Física, Universidad Nacional Autónoma de México, Apartado Postal 20-364, México 01000 DF, Mexico' author: - 'Miguel Angel Fuentes$^{1,2,3}$ and Alberto Robledo$^{4}$' title: Renormalization group structure for sums of variables generated by incipiently chaotic maps --- Introduction ============ As it is well documented [@kaminska1], increasingly larger sums of iterates of chaotic mappings give rise to a Gaussian stationary distribution in the same way independent random variables do according to the ordinary central limit theorem [@vankampen1; @khinchin1]. This deep-seated property, remarkably shared by deterministic and random systems composed of essentially uncorrelated variables, naturally raises questions about the existence, and if so, uniqueness or diversity, of limit distributions for systems made up of deterministic correlated variables. Related to these issues recent [@tsallis1]-[@grassberger1] numerical explorations of time averages of iterates at the period-doubling transition to chaos [@schuster1; @Hu] have been presented. Since the trajectories linked to this critical attractor are nonergodic and nonmixing the question of whether there is such stationary distribution for sums of iterates at the transition to chaos holds added interest as it may provide new angles to appraise the statistical mechanical analogy that is found in chaotic dynamics [robledo1]{}. The dynamics toward and at the Feigenbaum attractor is now known in much detail [@robledo1; @robledo2], therefore, it appears feasible to analyze also the properties of sums of iterate positions for this classic nonlinear system with the same kind of analytic reasoning and numerical thoroughness. Here we present the results for sums of chronological positions of trajectories associated to quadratic unimodal maps. We consider the case of the sum of positions of trajectories inside the Feigenbaum attractor as well as those within the chaotic $2^{K}$-band attractors obtained when the control parameter is shifted to values larger than that at the transition to chaos. Time and ensemble averages differ at the transition to chaos and here we chose to study the time average of a single trajectory initiated inside the attractor since all such trajectories, as explained below, are simply related. Clearly, time and ensemble averages are equivalent for chaotic attractors. From the information obtained we draw conclusions on the properties of the stationary distributions for these sums of variables. Our results, that reveal a multifractal stationary distribution that mirrors the features of the Feigenbaum attractor, can be easily extended to other critical attractor universality classes and other routes to chaos. About the relevance of our findings to physical systems, it is interesting to note, as one example, the parallels that have been found to exist between the dynamics at the noise-perturbed onset of chaos in unimodal maps and the dynamics of glass formation [@Brobledo]. In this connection chaotic band merging plays a central role in the relaxation properties of time correlation functions, while the multifractal attractor and multiband attractors in its neighborhood display the characteristic aging scaling property of glass formers [@Brobledo]. The overall picture we obtain is effectively described within the framework of the renormalization group (RG) approach for systems with scale invariant states or attractors. Firstly, the RG transformation for the distribution of a sum of variables is naturally given by the change due to the increment of summands followed by a suitable restoring operation. Second, the limit distributions can be identified as fixed points reached according to whether the acting relevant variables are set to zero or not. Lastly, the universality class of the non-trivial fixed-point distribution can be assessed in terms of the existing set of irrelevant variables. As it is well known [@schuster1; @Hu] a few decades ago the RG approach was successfully applied to the period-doubling route to chaos displayed by unimodal maps. In that case the RG transformation is functional composition and rescaling of the mapping and its effect re-enacts the growth of the period doubling cascade. In our case the RG transformation is the increment of terms and adjustment of the sum of positions and its effect is instead to go over again the merging of bands in the chaotic region. Specifically, we consider the Feigenbaum map $g(x)$, obtained from the fixed point equation $g(x)=\alpha g(g(x/\alpha ))$ with $g(0)=1$ and $g^{\prime }(0)=0$, and where $\alpha =-2.50290...$ is one of Feigenbaum’s universal constants [@schuster1; @Hu]. (For expediency we shall from now on denote the absolute value $\left\vert \alpha \right\vert $ by $\alpha $). Numerically, the properties of $g(x)$ can be conveniently obtained from the logistic map $f_{\mu }(x)=1-\mu x^{2},\;-1\leq x\leq 1$, with $\mu =$ $\mu _{\infty }=1.401155189092..$. The dynamics associated to the Feigenbaum map is determined by its multifractal attractor. For a recent detailed description of these properties see [@robledo1; @robledo2]. For values of $\mu >$ $\mu _{\infty }$ we employ a well-known scaling relation supported by numerical results. Initially we present properties of the sum of the absolute values $\left\vert x_{t}\right\vert $ of positions $x_{t}=f_{\mu _{\infty }}(x_{t-1}),\;t=1,2,3,...$, as a function of total time $N$ visited by the trajectory with initial position $x_{0}=0$, and obtain a patterned linear growth with $N$. We analyze this intricate fluctuating pattern, confined within a band of finite width, by eliminating the overall linear increment and find that the resulting stationary arrangement exhibits features inherited from the multifractal structure of the attractor. We derive an analytical expression for the sum that corroborates the numerical results and provide an understanding of its properties. Next, we consider the straight sum of $x_{t}$, where the signs taken by positions lessen the growth of its value as $N$ increases and the results are consistently similar to those for the sum of $\left\vert x_{t}\right\vert $, i.e. linear growth of a fixed-width band within which the sum displays a fluctuating arrangement. Numerical and analytical details for the sum of $x_{t}$ are presented. Then, we show numerical results for the sum of iterated positions obtained when the control parameter is shifted into the region of chaotic bands. In all of these cases the distributions evolve after a characteristic crossover towards a Gaussian form. Finally, we rationalize our findings in terms of an RG framework in which the action of the Central Limit Theorem plays a fundamental role and provide details of the crossover from multiband distributions to the gaussian distribution. We discuss our results. Sums of positions at the chaos threshold ======================================== Sum of absolute values $\left\vert x_{t}\right\vert $ ----------------------------------------------------- The starting point of our study is the evaluation of $$y_{\mu }(N)\equiv \sum\limits_{t=1}^{N}\left\vert x_{t}\right\vert , \label{sumabs1}$$with $\mu =\mu _{\infty }$ and with $x_{0}=0$. Fig. 1A shows the result, where it can be observed that the values recorded, besides a repeating fluctuating pattern within a narrow band, increase linearly on the whole. The measured slope of the linear growth is $c=0.56245...$ Fig. 1B shows an enlargement of the band, where some detail of the complex pattern of values of $y_{\mu _{\infty }}(N)$ is observed. A stationary view of the mentioned pattern is shown in Fig. 1C, where we plot$$y_{\mu _{\infty }}^{\prime }(N)\equiv \sum\limits_{t=1}^{N}\left( \left\vert x_{t}\right\vert -c\right) , \label{sumabs2}$$in logarithmic scales. There, we observe that the values of $y_{\mu _{\infty }}^{\prime }(N)$ fall within horizontal bands interspersed by gaps, revealing a fractal or multifractal set layout. The top (zeroth) band contains $y_{\mu _{\infty }}^{\prime }$ for all the odd values of $N$, the 1st band next to the top band contains $y_{\mu _{\infty }}^{\prime }$ for the even values of $N$ of the form $N=2+4m$, $m=0,1,2,...$ The 2nd band next to the top band contains $y_{\mu _{\infty }}^{\prime }(N)$ for $N=2^{2}+2^{3}m$, $m=0,1,2,...$, and so on. In general, the $k$-th band next to the top band contains $y_{\mu _{\infty }}^{\prime }(2^{k}+2^{k+1}m)$, $m=0,1,2,...$ Another important feature in this figure is that the $y_{\mu _{\infty }}^{\prime }(N)$ for subsequences of $N$ each of the form $N=(2l+1)2^{k}$, $k=0,1,2,...$, with $l$ fixed at a given value of $l=0,1,2,...$, appear aligned with a uniform slope $s=-1.323...$ The parallel lines formed by these subsequences imply the power law $y_{\mu _{\infty }}^{\prime }(N)\sim N^{s}$ for $N$ belonging to such a subsequence. ![A) Sum $y_{\mu _{\infty }}(N)$ of absolute values of visited points $x_{t}$, $t=0,...,N$, of the Feigenbaum’s attractor with initial condition $x_{0}=0$. B) A closer look of the path of the sum (see dotted circle in A), for values of $N$ around 67410. C) Centered sum $y_{\mu _{\infty }}^{\prime }(N)$ in logarithmic scales. See text.[]{data-label="f1"}](figure1.pdf){width="12cm"} ![Absolute value of trajectory positions $x_{t}$, $t=0,...$, for the logistic map $f_{\mu }(x)$ at $\mu _{\infty }$, with initial condition $x_{0}=0$, in logarithmic scale as a function of the logarithm of the time $t$, also shown by the numbers close to the points.[]{data-label="f2"}](figure2.pdf){width="10cm"} It is known [@robledo1; @robledo3] that these two characteristics of $y_{\mu _{\infty }}^{\prime }(N)$ are also present in the layout of the absolute value of the individual positions $\left\vert x_{t}\right\vert $, $t=1,2,3,...$ of the trajectory initiated at $x_{0}=0$; and this layout corresponds to the multifractal geometric configuration of the points of the Feigenbaum’s attractor, see Fig. 2. In this case, the horizontal bands of positions separated by equally-sized gaps are related to the set of period-doubling ‘diameters’ [@schuster1; @Hu] employed for the construction of the multifractal [@robledo2]. The identical slope shown in the logarithmic scales by all the position subsequences $\left\vert x_{t}\right\vert $, $t=(2l+1)2^{k}$, $k=0,1,2,...$, each formed by a fixed value of $l=0,1,2,...$, implies the power law $\left\vert x_{t}\right\vert \sim t^{s}$, $s=-\ln \alpha /\ln 2=-1.3236...$, as the $\left\vert x_{t}\right\vert $ can be expressed as $\left\vert x_{t}\right\vert \simeq \left\vert x_{2l+1}\right\vert \ \alpha ^{-k},t=(2l+1)2^{k}$, $k=0,1,2,...$, or, equivalently, $\left\vert x_{t}\right\vert \sim t^{s}$. Notice that the index $k$ also labels the order of the bands from top to bottom. The power law behavior involving the universal constant $\alpha $ of the subsequence positions reflect the approach of points in the attractor toward its most sparse region at $x=0$ from its most compact region, as the positions at odd times $\left\vert x_{2l+1}\right\vert =x_{2l+1}$, those in the top band, correspond to the densest region of the set. Having uncovered the through manifestation of the multifractal structure of the attractor into the sum $y_{\mu _{\infty }}^{\prime }(N)$ we proceed to derive this property and corroborate the numerical evidence. Consider Eq. (\[sumabs1\]) with $N=2^{k}$, $k=0,1,2,...$, the special case $l=0$ in the discussion above. Then the numbers of terms $\left\vert x_{t}\right\vert $ per band in $y_{\mu _{\infty }}(2^{k})$ are: $2^{k-1}$ in the top band ($j=0 $), $2^{k-2}$ in the next band ($j=1$),..., $2^{0}$ in the ($k-1$)-th band, plus an additional position in the $k$-th band. If we introduce the average of the positions on the top band$$\left\langle a\right\rangle \equiv 2^{-(k-1)}\sum_{j=0}^{2^{k-1}}x_{2j+1}, \label{ave1}$$the sum $y_{\mu _{\infty }}(2^{k})$ can be written as$$y_{\mu _{\infty }}(2^{k})=\left\langle a\right\rangle 2^{k-1}\sum_{j=0}^{k-2}(2\alpha )^{-j}+\alpha ^{-(k-1)}+\alpha ^{-k}. \label{sumabs3}$$Doing the geometric sum above and expressing the result as $y_{\mu _{\infty }}(2^{k})=c2^{k}+d\alpha ^{-k}$, we have$$c=\frac{\left\langle a\right\rangle \ \alpha }{2\alpha -1},~~~ d=\left( 1-\frac{\left\langle a\right\rangle \ 2\alpha }{2\alpha -1}\right) \ \alpha +1\ . \label{slopeandshift1}$$Evaluation of Eq. (\[ave1\]) yields to $\left\langle a\right\rangle =0.8999...$, and from this we obtain $c=0.56227...$ and $d=0.68826...$ We therefore find that the value of the slope $c$ in Fig. 1A is properly reproduced by our calculation. Also, since $\ln \left[ y_{\mu _{\infty }}(2^{k})-c2^{k}\right] =\ln d-k\ln \alpha $, or, equivalently, $\ln y_{\mu _{\infty }}^{\prime }(N)=\ln d-N\ln \alpha /\ln 2$,$\;N=2^{k}$,$\;k=0,1,2,... $, we corroborate that the value of the slope $s$ in the inset of Fig. 1C is indeed given by $s=-\ln \alpha /\ln 2=1.3236...$ (We have made use of the identity $\alpha ^{-k}=N^{-\ln \alpha /\ln 2}$,$\;N=2^{k}$,$\;k=0,1,2,...$). Sum of values of $x_{t}$ ------------------------ ![A) Sum $z_{\mu _{\infty }}(N)$ of values of visited points $x_{t}$, $t=0,...,N$, of the Feigenbaum’s attractor with initial condition $x_{0}=0$. B) A closer look of the path of the sum (see dotted circle in A), for values of $N$ around 59000. C and D) Centered sum $z_{\mu _{\infty }}^{\prime }(N)$ in logarithmic scales. See text.[]{data-label="f3"}](figure3.pdf){width="11cm"} When considering the signs taken by positions $x_{t}$ we note that their sum, $$z_{\mu }(N) \equiv \sum _{t=0}^{N}x_{t}, \label{sumnorm1}$$ when $N=2^{k}$, $\mu =$ $\mu _{\infty }$ and $x_{0}=0$, can be immediately obtained from the above derivation for $y_{\mu _{\infty }}(2^{k})$ simply by replacing $\alpha ^{-j}$ by $(-1)^{j}\alpha ^{-j}$, as the $x_{t}$ of different signs of the trajectory starting at $x_{0}=0$ fall into separate alternating bands (described above and shown in Fig. 2). In short, $x_{t}\simeq (-1)^{j}x_{2l+1}\ \alpha ^{-j}$, $t=(2l+1)2^{k}$, $k=0,1,2,...$ Writing $z_{\mu _{\infty }}(2^{k})$ as $z_{\mu _{\infty }}(2^{k})=e2^{k}+f(-1)^{k}\alpha ^{-k}$, we have$$e=\frac{\left\langle a\right\rangle \alpha }{2\alpha +1},~~~ f=\left( 1-\frac{\left\langle a\right\rangle 2\alpha }{2\alpha +1}\right) (-\alpha )+1. \label{slopeandshift2}$$Use of $\left\langle a\right\rangle =0.8999...$, leads to $e=0.37503...$ and $f=0.37443...$ Since $\ln \left( z_{\mu _{\infty }}(2^{k})-e2^{k}\right) =\ln f-k\ln \alpha $, or $\ln z_{\mu _{\infty }}^{\prime }(N)=\ln f-N\ln \alpha /\ln 2,\;N=2^{k},\;k=0,1,2,...$, where $z_{\mu _{\infty }}^{\prime }(2^{k})\equiv $ $z_{\mu _{\infty }}(2^{k})-e2^{k}$, we obtain for the value of the slope $s^{\prime }$ associated to the plot of $z_{\mu _{\infty }}^{\prime }(N)\sim N^{s^{\prime }}$ in logarithmic scales the number $s^{\prime }=-\ln (\alpha ^{2}-1)\ln \alpha /\ln 2=-2.1984...$(where the factor $\ln (\alpha ^{2}-1)$ takes into account the fact that consecutive values of the same sign in $z_{\mu _{\infty }}^{\prime }(2^{k})$ have $k=2m$ or $2m+1$, $m=0,1,2,...$) Our numerical evaluations for $z_{\mu _{\infty }}(N)$ and $z_{\mu _{\infty }}^{\prime }(N)$, shown in Fig. 3, reproduce the values given above for the slopes $e$ and $s^{\prime }$. Therefore, our numerical and analytical results are in agreement also in this case. The relationship between the sum $y_{\mu _{\infty }}(N)$ with initial condition $x_{0}=0$ and all other sums $Y_{\mu _{\infty }}(M)$ of consecutive positions with any initial condition $x_{0}$ inside the attractor can be obtained by inspection of Fig. 2. The sum $Y_{\mu _{\infty }}(M)$ differs only from $y_{\mu _{\infty }}(N)$ in the initial and final consecutive terms, $\sum_{t=1}^{t=t_{0}}\left\vert x_{t}\right\vert $ and $\sum_{t=N}^{t=M}\left\vert x_{t}\right\vert $, respectively, where $t_{0}$ is the time (shown in Fig. 2) at which the position $x_{0}$ is visited by the trajectory initiated at the origin $x=0$. When $t_{0}$, $N$, and $M$ are all powers of $2$ the differences between the sums become simpler and expressable in terms of $y_{\mu _{\infty }}(N)$. When $N\rightarrow \infty $ and $M\rightarrow \infty $ the difference between them is only a finite term $y_{\mu _{\infty }}(t_{0})$. Similar properties hold for the equivalent sums $z_{\mu _{\infty }}(N)$ and $Z_{\mu _{\infty }}(M)$ that take into account the signs of positions $x_{t}$. ![Distributions for the sums of positions $x_{t}$, $t=0,...,N$, of a single trajectory with initial condition $x_{0}=0$ within the $2^{3}$-band attractor at $\Delta \mu =0.0028448109$. The number of summands $N$ are indicated in each panel. See text.[]{data-label="f4"}](figure4.pdf){width="12cm"} Sums of positions for chaotic bands =================================== We turn now to study the sum of positions of trajectories when $\Delta \mu \equiv \mu -$ $\mu _{\infty }>0$. We recall that in this case the attractors are made up of $2^{K}$, $K=1,2,3,...$, bands and that their trajectories consist of an interband periodic motion of period $2^{K}$ and an intraband chaotic motion. We evaluated numerically the sums $z_{\mu }(N)$ for a single trajectory with initial condition $x_{0}=0$ for different values of $\Delta \mu $. The sum $z_{\mu }^{\prime }(N)$ was then obtained similarly to Eq. (\[sumabs2\]) by substracting the average $\left\langle z_{\mu }(N)\right\rangle _{x_{0}}$ and rescaling with a factor $N^{-1/2}$. The panels in Fig. 4 show the evolution of the distributions for increasing number of summands $N$ for a value of $\Delta \mu $ (chosen for visual clarity) when the attractor consists of $2^{3}$ chaotic bands. Initially the distributions are multimodal with disconnected domains, but as $N$ increases we observe merging of bands and development of a single-domain bell-shaped distribution that as $N\longrightarrow \infty $ converges in all cases to a Gaussian distribution. As a check of the ergodic property of chaotic band attractors we have also evaluated the distributions of sums of positions starting with an ensemble of uniformly-distributed initial positions $x_{0}$ and obtained results equivalent in all respects to those shown in Fig. 4. Faster convergence to the Gaussian distribution is achieved in this latter case. These numerical results can be understood as follows. We recollect [@schuster1; @Hu] that the relationship between the number of bands, $2^{K}$, $K\gg 1$, of a chaotic attractor and the control parameter distance $\Delta \mu $ at which it is located is given by $2^{K}\sim \Delta \mu ^{-\kappa }$, $\kappa =\ln 2/\ln \delta _{F}$, where $\delta _{F}=$ $0.46692...$ is the universal constant that measures both the rate of convergence of the values of $\mu $ at period doublings or at band splittings to $\mu _{\infty }$. For $\Delta \mu $ small and fixed, the sum of sequential positions of the trajectory initiated at $x_{0}=0$, Eq. (\[sumabs1\]), exhibits two different growth regimes as the total time $N$ increases. To specify them we introduce the difference in value $\delta x_{t}\equiv x_{t}(\mu )-\overline{x}_{t}(\mu )$ between the position at time $t$ and the average position within the band $\overline{x}_{t}(\mu )$ occupied at time $t$. Clearly, when $K\gg 1$ the average positions $\overline{x}_{t}(\mu )$ approximate the multifractal positions $x_{t}(\mu _{\infty })$ for $t\leq 2^{K}$. In the first regime, when $N\ll 2^{K}$, the properties of the sum $z_{\mu }(N)$ do not differ qualitatively from those of $z_{\mu _{\infty }}(N)$. This is because the fine structure of the Feigenbaum attractor is not suppressed by the fluctuations $\delta x_{t}$, as these contribute to the sum individually during the first cycle of the interband periodic motion. The discrete multi-scale nature of the distribution for $\mu _{\infty }$ is preserved when the interband motion governs the sum $z_{\mu }(N)$. The distributions for $z_{\mu }^{\prime }(N)$ and $z_{\mu _{\infty }}^{\prime }(N)$ are indistinguishable. In the second regime, when $N\gg 2^{K}$, the situation is opposite, after many interband cycles the fluctuations $\delta x_{t}$ add up in the sum and progressively wipe up the fine structure of the Feigenbaum attractor, leading to merging of bands and to the dominance of the fluctuating intraband motion. Ultimately, as $N\longrightarrow \infty $ the evolution of the distribution is similar to the action of the Central Limit Theorem and leads to a Gaussian stationary end result. It is also evident that as $\Delta \mu $ increases the first regime is shortened at the expense of the second, whereas when $\Delta \mu \longrightarrow 0$ the converse is the case. Therefore there exists an unambiguous $\Delta \mu $-dependent crossover behavior between the two radically different types of stationary distributions. This crossover is set out when the $\delta x_{t}$ fluctuations begin removing the band structure in $z_{\mu }^{\prime }(N)$ and ends when these fluctuations have broadened and merged all the chaotic bands and $z_{\mu }^{\prime }(N)$ forms a single continuous interval. When $\mu =\mu _{\infty }$ this process never takes place. An RG approach for sums of positions ==================================== We are in a position now to put together the numerical and analytical information presented above into the general framework of the RG approach. As known, this method was designed to characterize families of systems containing amongst their many individual states (or in this case attractors) a few exceptional ones with scale invariant properties and common to all systems in the family. We recall [@fisher1] that in the language of a minimal RG scheme there are two fixed points, each of which can be reached by the repeated application of a suitable transformation of the system’s main defining property. One of the fixed points, is termed trivial and is reached via the RG transformation for almost all initial settings. i.e. for all systems in the family when at least one of a small set of variables, named relevant variables, is nonzero. To reach the other fixed point, termed nontrivial, it is necessary that the relevant variables are all set to zero, and this implies a severely restricted set of initial settings that ensure such critical RG paths. The nontrivial fixed point embodies the scale invariant properties of the exceptional state that occurs in each system in the family and defines a universality class, while the differences amongst the individual systems are distinguished through a large set of so-called irrelevant variables. The variables in the latter set gradually vanish as the RG transformation is applied to a system that evolves toward the nontrivial fixed point. Further, when any system in the family is given a nonzero but sufficiently small value to (one or more of) the relevant variables, the RG transformation converts behavior similar to that of the nontrivial fixed point into that resembling the trivial fixed point through a well-defined crossover phenomenon. RG transformation and fixed points ---------------------------------- The recognition of the RG framework in the properties of the sums of positions of trajectories in unimodal maps and their associated distributions is straightforward. It can be concluded right away that in this problem (as defined here) there is only one relevant variable, the control parameter difference $\Delta \mu $. There is an infinite number of irrelevant variables, those that specify the differences between all possible unimodal maps (with quadratic maximum) and the Feigenbaum map $g(x)$. The RG transformation consists of the increment of one or more summands in the sum (\[sumabs1\]) followed by centering like in Eq. (\[sumabs2\]). The effect of the transformation in the distribution of the sum is then recorded . For sums of independent variables the transformation is equivalent to the convolution of distributions [@note1]. Notice that the transformation involves no scaling for the sums of positions at $\mu _{\infty }$. Examination of either Fig. 1 or Fig. 3 indicates that the values of the sums $y_{\mu _{\infty }}^{\prime }(N)$ and $z_{\mu _{\infty }}^{\prime }(N)$ are contained within a band of fixed width for all $N$ and therefore there should not be any scaling for such sums. The Feigenbaum attractor is not chaotic and its positions $x_{t}$ are not random variables. On the contrary the values of the sums $y_{\mu }^{\prime }(N)$ and $z_{\mu _{\infty }}^{\prime }(N)$ with $\mu >\mu _{\infty }$ spread as $N$ increases and scaling with a factor $N^{-1/2}$ maintains their values contained for $N\gg 1$. In this case the $x_{t}$ behave as independent random variables. There are two fixed-point distributions, the trivial continuum-space Gaussian distribution and the nontrivial discrete-space multifractal distribution (as observed in Figs. 1C, 3C and 3D). As explained above, there is a distinct crossover link between the two fixed-point distributions. Our results correspond to the dynamics inside the attractors, however, if the interest lies in considering only the stationary distribution of sums that do not contain the transient behavior of trajectories in their way to the attractor [@tsallis1]-[@tsallis3] our results are expected to give the correct answers for this case (see the Summary and discussion). Crossover via band merging --------------------------- In Fig. 5 we show a sector of the chaotic bands for the logistic map $f_{\mu }(x)$ where we indicate the widths of these bands at the control parameter values $\widehat{\mu }_{K}$ when they split each into two new bands. Interestingly, if we assume that for a given value of $K$ the widths of comparable lengths have equal lengths then these widths can be obtained from the widths of shortest and longest lengths via a simple scale factor consisting of an inverse power of $\alpha $. See Fig. 5. This introduces some degeneracy in the widths that propagates across the band splitting structure. Specifically, the widths scale now with increasing $K$ according to a binomial combination of the scaling of those that converge to the most crowded and most sparse regions of the multifractal attractor at $\mu _{\infty }$. As seen in Fig. 5 the widths form a Pascal triangle across the band splitting cascade. The total length $L_{K}$ of such chaotic $2^{K}$-band attractor can be immediately evaluated to yield$$L_{K}=\left( \alpha ^{-1}+\alpha ^{-2}\right) ^{K}. \label{lengthK}$$(See Ref. [@robledo2] for a similar evaluation relating to the supercycle diameters occurring across the bifurcation tree for $\mu <\mu _{\infty }$). ![Sector of the band splitting cascade for the logistic map $f_{\mu }(x) $ that shows the formation of a Pascal triangle of band widths (blue lines) at splitting according to the scaling approximation explained in the text, where $\alpha \simeq 2.5091$ is the absolute value of Feigenbaum’s universal constant.[]{data-label="f5"}](figure5.pdf){width="12cm"} Now, the sum $z_{\mu }(N)$ can be split into two terms, $z_{\mu }(N)=\overline{z}_{\mu }(N)+\delta z_{\mu }(N)$,$$\overline{z}_{\mu }(N)=\sum\limits_{t=0}^{N}\overline{x}_{t}\quad,~~~ \quad \delta z_{\mu }(N)=\sum\limits_{t=0}^{N}\delta x_{t}. \label{sumsz}$$Similarly to what we have seen for the sums $y_{\mu _{\infty }}(N)$ and $z_{\mu _{\infty }}(N)$, the first term $\overline{z}_{\mu }(N)$ is made of a narrow band of fixed width that shifts altogether linearly with $N$ to larger values while the band consists of a pattern of period $2^{K}$. Clearly $\overline{z}_{\mu }(N)$ does not participate in the band merging process, it is the second term $\delta z_{\mu }(N)$ that fluctuates and accomplishes band merging. Considering that all the correlated motion of $x_{t}$ has been taken over by $\overline{z}_{\mu }(N)$ the fluctuations of $\delta z_{\mu }(N)$ correspond to those of independent variables and band enlargement is measured by the mean square root displacement $$\left\langle \left[ \delta z_{\mu }(N)\right] ^{2}\right\rangle ^{1/2}\sim N^{1/2}. \label{msqrt1}$$We can estimate the number of summands $N_{K}$ necessary to achieve the merging of $2^{K}$ bands into a single one by matching the two lengths $\left\langle \left[ \delta z_{\mu }(N_{K})\right] ^{2}\right\rangle ^{1/2}$ and $1-L_{K}$. From Eqs. (\[lengthK\]) and (\[msqrt1\]) we obtain$$N_{K}\sim \left( \alpha ^{-1}+\alpha ^{-2}\right) ^{K}\left[ 1-\left( \alpha ^{-1}+\alpha ^{-2}\right) \right] , \label{match1}$$and considering $N_{K}$ to be of the form $N_{K}=2^{n_{K}}$ with $K\gg 1$ we obtain $$n_{K}\sim 2K\ln \left[ \left( \alpha ^{-1}+\alpha ^{-2}\right) /2\right] . \label{nKabout2K}$$The crossover estimate in Eq. (\[nKabout2K\]) for multiband into single-band distributions coincides in order of magnitude with the number of summands found necessary in Refs. [@tsallis2] and [@tsallis3] for numerical observations of long-tailed distributions resembling the so-called $q$-Gaussians [@tsallis2]. Summary and discussion ====================== In summary, we have found that the stationary distribution of the sum of iterate positions within the Feigenbaum attractor has a multifractal structure stamped by that of the initial multifractal set, while that involving sums of positions within the attractors composed of $2^{K}$ chaotic bands is the Gaussian distribution. We considered only the properties of a single sequence. At this transition the dynamics is nonergodic and nonmixing, therefore the two natural options, (i) time average with fixed initial condition and (ii) ensemble average of initial conditions at a large fixed time, are non- equivalent. A third option is to perform both ensemble and time averages. We chose option (i) because all sums with fixed initial positions within the attractor are simply related to each other via a deterministic term (as explained in Section II). Also, option (i) allowed us to obtain analytical results in closed form. For the ergodic chaotic band attractors the choice of a single initial condition leads to the same result as the use of an ensemble. In Refs. [@tsallis1]-[@tsallis3] sums of subsequent values of trajectories with uniformly distributed initial values across $-1\leq x\leq 1 $ were computed, and their properties were studied after discarding long transients. In Ref. [@tsallis3] transients were discarded of lengths ranging from 2048 to 65536 and their results are reported to be insensitive to the transient length. A cursory inspection of Ref. [@robledo2] (with regards to Figs. 14 to 18 and the text associated to them) makes it evident that after iteration times smaller than the smaller transient discarded in Ref. [@tsallis3] the trajectories considered have for all purposes fallen into the attractor. Thus, these sums are exceptionally well reproduced by sums of subsequent values $x_{t}$ with initial values inside the attractor. Next, when it is taken into account that sums belonging to every such initial value are simply related to that with $x_{0}=0$, it follows that the choice of sums we considered here capture the limiting properties of the sums studied in [@tsallis1]-[@tsallis3]. The numerical results that approximate $q$-Gaussians and/or Lévy distributions in [@tsallis1]-[@grassberger1] may be understood by observing that the power laws in the $\Delta \mu=0$ multifractal distribution (see Fig. 3) are preserved at the tails of the distributions throughout the band-merging crossover. Only in the limit $N \rightarrow \infty$ the distribution tails become truly exponential. We have also shown that the entire problem can be couched in the language of the RG formalism [@note1] in a way that makes clear the identification of the existing stationary distributions and the manner in which they are reached. These basic features suggest a degree of universality, and therefore limited to the critical attractor under consideration, in the properties of sums of deterministic variables at the transitions to chaos. Namely, the sums of positions of memory-retaining trajectories evolving under a vanishing Lyapunov exponent appear to preserve the particular features of the multifractal critical attractor under examination. Thus we expect that varying the degree of nonlinearity of a unimodal map would affect the scaling properties of time averages of trajectory positions at the period doubling transition to chaos, or alternatively, that the consideration of a different route to chaos, such as the quasiperiodic route, would lead to different scaling properties of comparable time averages. For instance, the known dependence of the universal constant $\alpha $ on the degree of nonlinearity $\varsigma $ of a unimodal map would show as a $\varsigma $-dependent exponents $s$ and $s^{\prime }$ that control the scale invariant property of the sums of trajectory positions with $x_{0}=0$ (shown in Figs. 1C, 3C and 3D). It is worthwhile expanding here our previous comment about the applicability of our method to other maps. For instance, the stationary distributions associated to the prototypical circle map [@schuster1; @Hu] could be determined similarly, thus extending our study to the route to chaos via quasi-periodicity. The dynamics at the classic golden mean critical attractor of the circle map exhibits counterparts with the Feigenbaum attractor concerning all the basic features that we have made use of here [@hernandez]. This can be corroborated via comparison of Fig. 2 with Fig. 3 of [@hernandez] and related text therein. We have contributed to clarify the nature and the circumstances under which a stationary distribution with universal properties (in the RG sense) may arise from sums of deterministic variables at the transition between regular and chaotic behavior, such as those studied here for variables evolving at zero Lyapunov exponent. In the absence of the fluctuating element present in chaotic attractors the distribution of the sums of these variables remains defined on a discrete multifractal set and is kept away from becoming a known (Gaussian or otherwise) continuum-space limit distribution for random variables. Acknowledgements. We appreciate partial financial support by DGAPA-UNAM and CONACYT (Mexican agencies). AR is grateful for hospitality received at the SFI. [99]{} Mackey M C and Tyran-Kaminska M 2006 Phys. Rep. [422]{} 167. van Kampen N G 1981 Stochastic Processes in Physics and Chemistry (North-Holland, Amsterdam). Khinchin A Y 1949 Mathematical Foundations of Statistical Mechanics (Dover, New York). Tirnakli U, Beck C and Tsallis C 2007 Phys. Rev. E [75]{} 040106(R). Tirnakli U, Tsallis C and Beck C 2009 Phys. Rev. E [79]{} 056209. Tirnakli U, Tsallis C and Beck C (arXiv:0906.1262) \[cond-mat.stat-mech\]. Grassberger P 2009 Phys. Rev. E [79]{} 057201. Schuster H G 1988 Deterministic Chaos. An Introduction (2nd Revised Edition, VCH Publishers, Weinheim). Hu B 1982 Phys. Rep. [91]{} 233. Mayoral E and Robledo A 2005 Phys. Rev. E [72]{} 026209. Robledo A and Moyano L G 2008 Phys. Rev. E [77]{} 036213 1-14. Baldovin F and Robledo A 2005 Phys. Rev. E [72]{} 066213. Robledo A 2006 Physica A [370]{} 449. Fisher M E 1998 Rev. Mod. Phys. [70]{} 653. A helpful note about the differences between the original RG transformation for unimodal maps [@schuster1; @Hu] and that presented here may be useful for the casual reader. (i) Feigenbaum’s RG transformation consists of functional composition (and fixed scaling with the universal constant $\alpha $), while ours is summation of variables (with no scaling at the transition and scaling with the inverse square root of total number of terms beyond the transition). (ii) The fixed points in Feigenbaum’s RG are maps, while the fixed points in ours are distributions. (iii) Feigenbaum’s RG generates the period doubling cascade, while ours produces band merging. (iv) The purpose of Feigenbaum’s RG is to characterize the critical attractor, while ours is to characterize stationary distributions. Hernández-Saldaña H and Robledo A 2006 Physica A [370]{} 286.	{ "pile_set_name": "ArXiv" }
--- abstract: 'Learning disentanglement aims at finding a low dimensional representation which consists of multiple explanatory and generative factors of the observational data. The framework of variational autoencoder (VAE) is commonly used to disentangle independent factors from observations. However, in real scenarios, factors with semantics are not necessarily independent. Instead, there might be an underlying causal structure which renders these factors dependent. We thus propose a new VAE based framework named CausalVAE, which includes a Causal Layer to transform independent exogenous factors into causal endogenous ones that correspond to causally related concepts in data. We further analyze the model identifiabitily, showing that the proposed model learned from observations recovers the true one up to a certain degree. Experiments are conducted on various datasets, including synthetic and real word benchmark CelebA. Results show that the causal representations learned by CausalVAE are semantically interpretable, and their causal relationship as a Directed Acyclic Graph (DAG) is identified with good accuracy. Furthermore, we demonstrate that the proposed CausalVAE model is able to generate counterfactual data through “do-operation" to the causal factors.' author: - \| Mengyue Yang$^{1}$, Furui Liu$^{1}$, Zhitang Chen$^{1}$, Xinwei Shen$^{2}$, Jianye Hao$^{1}$, Jun Wang$^{3}$\ $^{1}$ Noah’s Ark Lab, Huawei, Shenzhen, China\ $^{2}$ The Hong Kong University of Science and Technology, Hong Kong, China\ $^{3}$ University College London, London, United Kingdom\ `yangmengyue2,liufurui2,chenzhitang2,[email protected]`\ `[email protected]`\ `[email protected]` bibliography: - 'references.bib' title: 'CausalVAE: Disentangled Representation Learning via Neural Structural Causal Models ' --- Introduction ============ Disentangled representation learning is of great importance in various applications such as computer vision, speech and natural language processing, and recommender systems [@hsu2017unsupervised; @ma2019learning; @hsieh2018learning]. The reason is that it might help enhance the performance of models, i.e. improving the generalizability, robustness against adversarial attacks as well as the explanability, by learning data’s latent disentangled representation. One of the most common frameworks for disentangled representation learning is Variational Autoencoders (VAE), a deep generative model trained to disentangle the underlying explanatory factors. Disentanglement via VAE can be achieved by a regularization term of the Kullback-Leibler (KL) divergence between the posterior of the latent factors and a standard Multivariate Gaussian prior, which enforces the learned latent factors to be as independent as possible. It is expected to recover the latent variables if the observation in real world is generated by countable independent factors. To further enhance the independence, various extensions of VAE consider minimizing the mutual information among latent factors. For example, @higgins2017beta and @burgess2018understanding increased the weight of the KL divergence term to enforce independence. @kim2018disentangling [@chen2018isolating] further encourage the independence by reducing total correlation among factors. Most existing works of disentangled representation learning make a common assumption that the real world observations are generated by countable independent factors. Nevertheless we argue that in many real world applications, latent factors with semantics of interest are causally related and thus we need a new framework that supports causal disentanglement. -0.2in Consider a toy example of a swinging pendulum in Fig. \[true\_pendulum\]. The position of the illumination source and the angle of the pendulum are causes of the position and the length of the shadow. Through causal disentangled representation learning, we aim at learning representations that correspond to the above four concepts. Obviously, these concepts are not independent and existing methods may fail to extract those factors. Furthermore, causal disentanglement allow us to manipulate the causal system to generate counterfactual data. For example, we can manipulate the latent code of shadow to create new pictures without shadow even there are pendulum and light. This corresponds to the “do-operation” [@pearl2009causality] in causality, where the system operates under the condition that certain variables are controlled by external forces. A deep generative model that supports “do-operation” is of tremendous value as it allows us to ask “what-if" questions when making decisions. In this paper, we propose a VAE-based causal disentangled representation learning framework by introducing a novel Structural Causal Model layer, which allows us to recover the latent factors with semantics and structured via a causal DAG. The input signal passes through an encoder to obtain independent exogenous factors and then a Causal Layer to generate causal representation which is taken by the decoder to reconstruct the original input. We call the whole process Causal Disentangled Representation Learning. Unlike unsupervised disentangled representation learning of which the feasibility is questionable [@locatello2018challenging], additional information is required as weak supervision signals to achieve causal representation learning. By “weak supervision", we emphasize that in our work, the causal structure of the latent factors is automatically learned, instead of being given as a prior in [@causalgan]. To train our model, we propose a new loss function which includes the VAE evidence lower bound loss and an acyclicity constraint imposed on the learned causal graph to guarantee its “DAGness". In addition, we analyze the identifiablilty of the proposed model, showing that the learned parameters of the disentangled model recover the true one up to certain degree. The contribution of our paper is three-fold. (1) We propose a new framework named CausalVAE that supports causal disentanglement and “do-operation"; (2) Theoretical justification on model identifiability is provided; (3) We conduct comprehensive experiments with synthetic and real world face images to demonstrate that the learned factors are with causal semantics and can be intervened to generate counterfactual images that do not appear in training data. Related Works ============= In this section, we review state-of-the-art disentangled representation learning methods, including some recent advances on combining causality and disentangled representation learning. We also present preliminaries of causal structure learning from pure observations which is a key ingredient of our proposed CausalVAE framework. Disentangled Representation Learning: Conventional disentangled representation learning methods learn mutually independent latent factors by an encoder-decoder framework. In this process, a standard normal distribution is used as a prior of the latent code. A variational posterior $q(\mathbf{z}\|\mathbf{x})$ is then used to approximate the unknown true posterior $p(\mathbf{z}\|\mathbf{x})$. This framework was further extended by adding new independence regularization terms to the original loss function, leading to various algorithms. $\beta$-VAE [@higgins2017beta] proposes an adaptation framework which adjusts the weight of KL term to balance between independence of disentangled factors and the reconstruction performance. While factor VAE [@chen2018isolating] proposes a new framework which focuses solely on the independence of factors. Ladder VAE [@laddervae] on the other hand, leverages the structure of ladder neural network to train a structured VAE for hierarchical disentanglement. ![Model structure of CausalVAE. The encoder takes observation $\mathbf{x}$ as inputs to generate independent exogenous variable $\bm{\epsilon}$, whose prior distribution is assumed to be standard Multivariate Gaussian. Then it is transformed by the Causal Layer into causal representations $\mathbf{z}$ (Eq. \[sem\]) with a conditional prior distribution $p(\mathbf{z}\|\mathbf{u})$. A Mask Layer is then applied to $\mathbf{z}$ to resemble the SCM in Eq. \[maskedlayer\]. After that, $\mathbf{z}$ is taken as the input of the decoder to reconstruct the observation $\mathbf{x}$. []{data-label="cvae_structure"}](dataflow_new.pdf){width="1\columnwidth"} -0.4in Nevertheless the aforementioned unsupervised disentangled representation learning algorithms do not perform well in some situations where there is complex causal relationship among factors. Furthermore, they are challenged for lacking inductive bias and thus the model identifiability cannot be guaranteed [@locatello2018challenging]. The identifiability problem of VAE is defined as follows: if the parameters $\tilde{\bm{\theta}}$ learned from data lead to a marginal distribution equal to the true one parameterized by $\bm{\theta}$, i.e., $p_{\tilde{\bm{\theta}}}(\mathbf{x}) = p_{\bm{\theta}}(\mathbf{x}) $, then the joint distributions also match, i.e. $p_{\tilde{\bm{\theta}}}(\mathbf{x},\mathbf{z}) = p_{\bm{\theta}}(\mathbf{x},\mathbf{z}) $. Therefore, the rotation invariance of prior $p(\mathbf{z})$ (standard Multivariate Gaussian distribution) will lead the unindentifiable of $p(\mathbf{z})$. @nonlinearica prove that there is infinite number of distinct models entailing the same joint distributions, which means that the underlying generative model is not identifiable through unsupervised learning. On the contrary, by leveraging a few labels, one is able to recover the true model [@mathieu2018disentangling; @locatello2018challenging]. @kulkarni2015deep and @locatello2019disentangling use additional labels to reduce the model ambiguity. @nonlinearica gives an identifiability of VAE with additional inputs, by leveraging the theory of nonlinear Independent Component Analysis (nonlinear ICA) [@brakel2017learning]. Causal Discovery & Causal Disentangled Representation Learning: We refer to causal representation as ones structured by a causal graph. Discovering the causal graph from pure observations has attracted large amounts of attention in the past decades [@hoyer2009nonlinear; @zhang2012identifiability; @shimizu2006linear]. @pearl2009causality introduced a Probabilistic Graphical Models (PGMs) based language to describe causality among variables. @shimizu2006linear proposed an effective method called LiNGAM to learn the causal graph and they prove the model identifiability under the linearity and non-Gaussianity assumption. @zheng2018dags proposed NOTEARs with a fully differentiable DAG constraint for causal structure learning, which drastically reduces a very complicated combinatorial optimization problem to a continuous optimization problem. @zhu2019causal proposed a flexible and efficient Reinforcement Learning (RL) based method to search over a DAG space for a best graph with a highest score. Recently, the community has raised interest of combining causality and disentangled representation. @suter2018robustly used causality to explain disentangled latent representations. @causalgan proposed a method called CausalGAN which supports “do-operation” on images but it requires the causal graph given as a prior. Instead of assuming independent latent factors, @besserve2018counterfactuals allows dependent latent factors. However, in their work, the dependence is induced by some latent confounders, instead of a causal graph among latent factors investigated in this paper. @scholkopf2019causality stressed the importance and necessity of causal disentangled representation learning but it still remains conceptual. To the best of our knowledge, our work is the first one that successfully implements the idea of causal disentanglement. Causal Disentanglement in Variational Autoencoder ================================================= We start with the definition of causal representation, and then propose a new framework to achieve causal disentanglement by leveraging additional inputs, e.g. labels of concepts. Firstly, we give an overview of our proposed CausalVAE model structure in Fig. \[cvae\_structure\]. A Causal Layer, which essentially describes a Structural Causal Model (SCM) [@shimizu2006linear], is introduced to a conventional VAE network. The Causal Layer transforms the independent exogenous factors to causal endogenous factors corresponding to causally related concepts of interest. A mask mechanism [@ng2019masked] is then used to propagate the effect of parental variables to their children, mimicking the assignment operation of SCMs. Such a Causal Layer is the key to supporting intervention or “do-operation" to the system. Transforming Independent Exogenous Factors into Causal Representations ---------------------------------------------------------------------- Our model is within the framework of VAE-based disentanglement. In addition to the encoder and the decoder structures, we introduce a Structural Causal Model (SCM) layer to learn causal representations. To formalize causal representation, we consider $n$ concepts of interest in data. The concepts in observations are causally structured by a Directed Acyclic Graph (DAG) with an adjacency matrix $\mathbf{A}$. Though a general nonlinear SCM is preferred, for simplicity, in this work, the Causal Layer exactly implements a Linear SCM as described in Eq. \[sem\] (shown in Fig. \[cvae\_structure\] ), $$\mathbf{z} = \mathbf{A}^T\mathbf{z}+\bm{\epsilon} = (I-\mathbf{A}^T)^{-1}\bm{\epsilon},\\ \bm{\epsilon}\sim \mathcal{N}(\mathbf{0},\mathbf{I}), \label{sem}$$ where $\mathbf{A}$ is the parameters to be learnt in this layer. $\bm{\epsilon}$ are independent Gaussian exogenous factors and $\textbf{z}\in \mathbb{R}^{n}$ is structured causal representation of $n$ concepts that is generated by a DAG and thus $\mathbf{A}$ can be permuted into a strictly upper triangular matrix. Unsupervised learning of the model might be infeasible due to the identifiability issue as discussed in [@locatello2018challenging]. To address this problem, similar to iVAE [@nonlinearica], we adopt additional information $\mathbf{u}$ associated with the true causal concepts as supervising signals. In our work, we use the labels of the concepts. The additional information $\mathbf{u}$ is utilized in two ways. Firstly, we propose a conditional prior $p(\mathbf{z\|u})$ to regularize the learned posterior of $\mathbf{z}$. This guarantees that the learned model belongs to an identifiable family. Secondly, we also leverage $\mathbf{u}$ to learn the causal structure $\mathbf{A}$. Besides learning the causal representations, we further enable the model to support intervention to the causal system to generate counterfactual data which does not exist in the training data. Structural Causal Model Layer {#mask_layer} ----------------------------- Once the causal representations $\textbf{z}$ is obtained, it passes through a Mask Layer [@ng2019masked] to reconstruct itself. Note that this step resembles a SCM which depicts how children are generated by their corresponding parental variables. We will show why such a layer is necessary to achieve intervention. Let $z_i$ be the $i$th variable in the vector $\mathbf{z}$. The adjacency matrix associated with the causal graph is $\mathbf{A} = [\mathbf{A}_1\|\dots\|\mathbf{A}_n]$ where $\mathbf{A}_i \in \mathbb{R}^n$ is the weight vector such that $A_{ji}$ encodes the causal strength from $z_j$ to $z_i$. We have a set of mild nonlinear and invertible functions $ [g_1, g_2,\dots,g_n]$ that map parental variables to the child variable. Then we write $$\begin{aligned} \label{maskedlayer} z_i = g_i(\mathbf{A}_i \circ \mathbf{z}; \bm{\eta}_i) + \epsilon_i,\end{aligned}$$ where $\circ$ is the element-wise multiplication and $\bm{\eta}_i$ is the parameter of $g_i(\cdot)$ (as shown in Fig. \[cvae\_structure\] ). Note that according to Eq. \[sem\], we can simply write $z_i = \mathbf{A}_i^T \mathbf{z}+ \epsilon_i$. However, we find that adding a mild nonlinear function $g_i$ results in more stable performances. To show how this masking works, consider a variable $z_i$ and $\mathbf{A}_i \circ \mathbf{z}$ equals a vector that only contains its parental information as it masks out all $z_i$’s non-parent variables. By minimizing the reconstruction error, the adjacency matrix $\mathbf{A}$ and the parameter $\bm{\eta}_i$ of the mild nonlinear function $g_i$ are trained. This layer makes intervention or “do-operation” possible. Intervention [@pearl2009causality] in causality refers to modifying a certain part of a system by external forces and one is interested in the outcome of such manipulation. To intervene $z_i$, we set $z_i$ on the RHS of Eq. \[maskedlayer\] (corresponding to the $i-$th node of $\mathbf{z}$ in the first layer in Fig. \[cvae\_structure\]) to a fixed value, and then its effect is delivered to all its children as well as itself on the LHS of Eq. \[maskedlayer\] (corresponding to some nodes of $\mathbf{z}$ in the second layer). Note that intervening the cause will change the effect, whereas intervening the effect, on the other hand, does not change the cause because information can only flow into the next layer from the previous one in our model, which is aligned with the definition of causal effects. A Probabilistic Generarive Model for CausalVAE {#prob_analysis} ---------------------------------------------- We give a probabilistic formulation of the proposed generative model (shown in Fig. \[cvae\_structure\] ). Denote by $\mathbf{x}\in \mathbb{R}^{d}$ the observed variables and $\mathbf{u}\in \mathbb{R}^{n}$ the additional information. $u_i$ is the label of the $i$-th concept of interest in data. Let $\bm{\epsilon} \in \mathbb{R}^n$ be the latent exogenous independent variables and $\mathbf{z}\in \mathbb{R}^{n}$ be the latent endogenous variables with semantics where $\mathbf{z} = \mathbf{A}^{T}\mathbf{z} + \bm{\epsilon} = (\mathbf{I} - \mathbf{A}^T)^{-1}\bm{\epsilon}$. For simplicity, we denote $\mathbf{C}=(\mathbf{I} - \mathbf{A}^T)^{-1}$. We treat both $\mathbf{z}$ and $\bm{\epsilon}$ as latent variables. Consider the following conditional generative model parameterized by $\bm{\theta} = (\mathbf{f}, \mathbf{h},\mathbf{C}, \mathbf{T}, \bm{\lambda})$: $$\begin{aligned} p_{\bm{\theta}}(\mathbf{x,z,\bm{\epsilon}\|u}) &= p_{\bm{\theta}}(\mathbf{x\|z,\bm{\epsilon},u})p_{\bm{\theta}}(\mathbf{\bm{\epsilon},z\|u}). \label{generator} % &= p_{\bm{\theta}}(\mathbf{x}\|\mathbf{C}\bm{\epsilon})p_{\bm{\theta}}(\mathbf{C}\bm{\epsilon}\|\mathbf{u}), \\ % &=p_{\bm{\theta}}(\mathbf{x},\mathbf{C}\bm{\epsilon}\|\mathbf{u}).\end{aligned}$$ Let $\mathbf{f}(\mathbf{z})$ denote the decoder which is assumed to be an invertible function and $\mathbf{h}(\mathbf{x}, \mathbf{u})$ denotes the encoder. We define the generative and inference models as follows: $$\label{Eq: conditional} p_{\bm{\theta}}(\mathbf{x\|z,\bm{\epsilon},u})=p_{\bm{\theta}}(\mathbf{x}\|\mathbf{z}) \equiv p_{\bm{\xi}}(\mathbf{x} - \mathbf{f}(\mathbf{z})),~~~ q_{\bm{\phi}}(\mathbf{z},\bm{\epsilon}\|\mathbf{x},\mathbf{u}) \equiv q(\mathbf{z}\|\bm{\epsilon})q_{\bm{\zeta}}(\bm{\epsilon} - \mathbf{h}(\mathbf{x},\mathbf{u})),$$ which is obtained by assuming the following decoding and encoding processes: $$\label{xi} \mathbf{x}=\mathbf{f}(\mathbf{z})+\bm{\xi},~~~\bm{\epsilon}=\mathbf{h(x,u)}+\bm{\zeta},$$ where $\bm{\xi}$ and $\bm{\zeta}$ are the vectors of independent noise with probability densities $p_{\bm{\xi}}$ and $q_{\bm{\zeta}}$. When $\bm{\xi}$ and $\bm{\zeta}$ are infinitesimal, the encoder and decoder can be regarded as deterministic ones. We define the joint prior $p_{\bm{\theta}}(\bm{\epsilon},\mathbf{z}\|\mathbf{u})$ for latent variables $\mathbf{z}$ and $\bm{\epsilon}$ as $$\label{eq:joint_prior} p_{\bm{\theta}}(\bm{\epsilon},\mathbf{z}\|\mathbf{u}) = p_{\bm{\epsilon}}(\bm{\epsilon})p_{\bm{\theta}}(\mathbf{z\|u}),$$ where $p_{\bm{\epsilon}}(\bm{\epsilon})=\mathcal{N}(\mathbf{0},\mathbf{I})$ and the prior of latent endogenous variables $p_{\bm{\theta}}(\mathbf{z\|u})$ is a factorized Gaussian distribution conditioning on the additional observation $\mathbf{u}$, i.e. $$\begin{aligned} p_{\bm{\theta}}(\mathbf{z}\|\mathbf{u}) = \mathcal{N}(\lambda_1(u_i),\lambda_2^2(u_i)), \label{prior}\end{aligned}$$ where $\lambda_1$ and $\lambda_2$ are an arbitrary functions. In this paper, we let $\lambda_1(\mathbf{u}) = \mathbf{u}$ and $\lambda_2(\bm{u})\equiv 1$. Learning Strategy {#learning_strategy} ================= In this section, we discuss how to train the CausalVAE model in order to learn the causal representation as well as the causal graph simultaneously. Evidence Lower Bound of CausalVAE --------------------------------- We apply variational Bayes to learn a tractable distribution $q_\phi(\bm{\epsilon},\mathbf{z}\|\mathbf{x},\mathbf{u})$ to approximate the true posterior $p_{\bm{\theta}}(\bm{\epsilon},\mathbf{z}\|\mathbf{x},\mathbf{u})$. Given data set $\mathcal{X}$ with the empirical data distribution $q_\mathcal{X}(\mathbf{x},\mathbf{u})$, the parameters $\bm{\theta}$ and $\bm{\phi}$ are learned by optimizing the following evidence lower bound (ELBO): $$\begin{aligned} \resizebox{0.9\hsize}{!}{ $\mathbb{E}_{q_\mathcal{X}}[\log p_{\bm{\theta}}(\mathbf{x}\|\mathbf{u})] \ge \text{ELBO} = \mathbb{E}_{q_\mathcal{X}}[\mathbb{E}_{\bm{\epsilon},\mathbf{z}\sim q_{\bm{\phi}}}[\log p_{\bm{\theta}}(\mathbf{x}\|\mathbf{z},\bm{\epsilon},\mathbf{u})]-\mathcal{D}(q_{\bm{\phi}} (\bm{\epsilon},\mathbf{z}\|\mathbf{x},\mathbf{u})\|\|p_{\bm{\theta}}(\bm{\epsilon},\mathbf{z}\|\mathbf{u}))]$}, \label{elbo}\end{aligned}$$ where $\mathcal{D}(\cdot\\|\cdot)$ denotes KL divergence. Eq. \[elbo\] is intractable in general. However, thanks to the one-to-one correspondence between $\bm{\epsilon}$ and $\mathbf{z}$, we simplify the variational posterior as follows: $$\begin{aligned} \small q_{\bm{\phi}} (\bm{\epsilon},\mathbf{z}\|\mathbf{x},\mathbf{u}) &= q_{\bm{\phi}} (\bm{\epsilon}\|\mathbf{x},\mathbf{u})\delta(\bm{z}=\mathbf{C}\bm{\epsilon})=q_{\bm{\phi}} (\mathbf{z}\|\mathbf{x},\mathbf{u})\delta(\bm{\epsilon}=\mathbf{C}^{-1}\bm{z}), \label{inferencemodel}\end{aligned}$$ where $\delta(\cdot)$ is the Dirac delta function. According to the model assumptions introduced in Section \[prob\_analysis\], i.e., generation process (Eq. \[Eq: conditional\]) and prior (Eq. \[eq:joint\_prior\]), we attain a neat form of ELBO loss as follows: Proposition 1. ELBO defined in Eq. \[elbo\] can be written as: $$\begin{aligned} \label{eq:elbo_decom} \small \text{ELBO} &= \mathbb{E}_{q_\mathcal{X}}[\mathbb{E}_{q_\phi(\mathbf{z}\|\mathbf{x},\mathbf{u})}[\log p_{\bm{\theta}}(\mathbf{x}\|\mathbf{z})] - \mathcal{D}(q_\phi(\bm{\epsilon}\|\mathbf{x},\mathbf{u})\|\|p_{\bm{\epsilon}}(\bm{\epsilon}))-\mathcal{D}(q_\phi(\mathbf{z}\|\mathbf{x},\mathbf{u}) \|\| p_{\bm{\theta}}(\mathbf{z}\|\mathbf{u}))].\end{aligned}$$ Details of the proof are given in the Appendix. With this form, we can easily implement a loss function to train the CausalVAE model. Learning the Causal Structure of Latent Codes --------------------------------------------- In addition to the encoder and decoder, our CausalVAE model involves a Causal Layer with a DAG structure to be learned. Note that both $\mathbf{z}$ and $\mathbf{A}$ are unknown, to ease the training task, we leverage the additional labels $\mathbf{u}$ to construct the following constraint: $$\small l_u=\mathbb{E}_{q_{\mathcal{X}}}\Vert \mathbf{u} - \sigma(\mathbf{A}^T \mathbf{u})\Vert_2^2\leq \kappa_1, \label{constraint_A}$$ where $\sigma$ is a logistic function as our lables are binary and $\kappa_1$ is the small positive constant value. This follows the idea that $\mathbf{A}$ should also describe the causal relations among labels well. Similarly we apply the same constraint to the learned latent code $\mathbf{z}$ as follows: $$\label{constraint_m} l_m= \mathbb{E}_{\mathbf{z}\sim q_{\bm{\phi}}}\sum_{i=1}^n \Vert z_i - g_i(\mathbf{A}_i \circ\mathbf{z}; \bm{\eta}_i) \Vert^2 \leq \kappa_2,$$ where $\kappa_2$ is the small positive constant value. Lastly, the causal adjacency matrix $\mathbf{A}$ is constrained to be a DAG. Instead of using traditional DAG constraint that is combinatorial, we adopt a continuous differentiable constraint function [@zheng2018dags; @DBLP:journals/corr/abs-1906-04477; @DBLP:journals/corr/abs-1911-07420; @yu2019dag] . The function attains 0 if and only if the adjacency matrix $\mathbf{A}$ corresponds to a DAG [@yu2019dag], i.e. $$\begin{aligned} H(\mathbf{A}) \equiv tr((\mathbf{I}+\mathbf{A}\circ \mathbf{A})^n) - n = 0. \label{dag}\end{aligned}$$ The training procedure of our CausalVAE model reduces to the following constrained optimization: $$\begin{aligned} \text{maximize} ~~~\text{ELBO}, ~~~ \text{s.t.}~~~(\ref{constraint_A})(\ref{constraint_m})(\ref{dag}). \nonumber\end{aligned}$$ By lagrangian multiplier method, we have the new loss function $$\begin{aligned} \mathcal{L} = -\text{ELBO} + \alpha H(\mathbf{A})+ \beta l_u + \gamma l_m, \label{causal_elbo} \end{aligned}$$ where $\alpha,\beta, \gamma$ denote regularization hyperparameters. Identifiability Analysis {#identifiability_analysis} ======================== In this section, we present the identifiability of our proposed model. We adopt the $\sim$-identifiability [@nonlinearica] as follows: Definition 1. Let $\sim$ be the binary relation on $\Theta$ defined as follows: $$\begin{split} (\mathbf{f},\mathbf{h},\mathbf{C}, \mathbf{T},\bm{\lambda})\sim (\tilde{\mathbf{f}},\tilde{\mathbf{h}}, \tilde{\mathbf{C}}, \tilde{\mathbf{T}},\tilde{\bm{\lambda}}) &\Leftrightarrow \exists \mathbf{B_1},\mathbf{B_2},\mathbf{b_1},\mathbf{b_2} \vert\\ \mathbf{T}(\mathbf{h}(\mathbf{x},\mathbf{u})) = \mathbf{B_1}\tilde{\mathbf{T}}(\tilde{\mathbf{h}}(\mathbf{x},\mathbf{u}))+\mathbf{b_1},\mathbf{T}(\mathbf{f^{-1}}(\mathbf{x})) &= \mathbf{B_2}\tilde{\mathbf{T}}( {\mathbf{ \tilde{f}^{-1}}}(\mathbf{x}))+\mathbf{b_2},\forall \mathbf{x}\in \mathcal{X}, \end{split}$$ where $\mathbf{C}=(\mathbf{I} - \mathbf{A}^T)^{-1}$. If $\mathbf{B_1}$ is an invertible matrix and $\mathbf{B_2}$ is an invertible diagonal matrix with diagonal elements associated to $u_i$. We say that the model parameter is $\sim$-identifiable. Following [@nonlinearica], we obtain the identifiability of our causal generative model as follows. Theorem 1. Assume that the data we observed are generated according Eq. \[generator\]-\[Eq: conditional\] and the following assumptions hold, 1. The set $\{\mathbf{}x\in\mathcal{X}\|\phi_\xi(\mathbf{x})=0\}$ has measure zero, where $\phi_\xi$ is the characteristic function of the density $p_\xi$ defined in Eq. \[xi\]. 2. The decoder function $\mathbf{f}$ is differentiable and the Jacobian matrix of $\mathbf{f}$ is of full rank [^1]. 3. The sufficient statistics $T_{i,s}({z}_i)\not=0$ almost everywhere for all $1\le i \le n$ and $1\le s \le2$, where $T_{i,s}({z}_i)$ is the $s$th statistic of variable $z_i$. 4. The additional observations $u_i\not=0$. Then the parameters $(\mathbf{f}, \mathbf{h}, \mathbf{C}, \mathbf{T}, \bm{\lambda})$ are $\sim$-identifiable. The identifiability of the model under supervision of additional information is obtained thanks to the conditional prior $p_{\bm{\theta}}(\mathbf{z}\|\mathbf{u})$. The conditional prior guarantees that sufficient statistics of $p_{\bm{\theta}}(\mathbf{z}\|\mathbf{u})$ are related to the value of $\mathbf{u}$. A complete proof of Theorem 1 is available in Appendix. Experiments =========== In this section, we conduct experiments using both synthetic dataset and real human face image dataset and we compare our CausalVAE model against existing state of the art methods on disentangled representation learning. We focus on examing whether a certain algorithm is able to learn interpretable representations and whether outcomes of intervention on learned latent code is consistent to our understanding of the causal system. 0.2in ![The results of Intervention experiments on the pendulum dataset. Each row shows the result of controlling the <span style="font-variant:small-caps;">pendulum angle</span>, <span style="font-variant:small-caps;">light angle</span>, <span style="font-variant:small-caps;">shadow length</span>, and <span style="font-variant:small-caps;">shadow location</span> respectively. The bottom row is the original input image.[]{data-label="pendulum_res"}](cvae_pendulum_result/pendulum_results.png){width="\columnwidth"} -0.1in -0.2in ![Results of CausalVAE model on CelebA. The controlled factors are <span style="font-variant:small-caps;">gender</span>, <span style="font-variant:small-caps;">smile, eyes open</span> and <span style="font-variant:small-caps;">mouth open</span> respectively.[]{data-label="do_smile"}](cvae_doface_results/face_results.png){width="1\columnwidth"} -0.4in -0.1in ![The learning process of causal matrix $\mathbf{A}$. The concepts include: <span style="font-variant:small-caps;">Gender</span>, <span style="font-variant:small-caps;">Smile</span>, <span style="font-variant:small-caps;">Eyes Open</span>, <span style="font-variant:small-caps;">Mouth Open</span> (top-to-bottom and left-to-right order); (c) converged $\mathbf{A}$, (d) ground truth .[]{data-label="learn_A"}](result_graph/smile/smile_graph.png){width="0.6\columnwidth"} -0.1in -0.3in Dataset, Baselines & Metrics ---------------------------- Dataset: We conduct experiments on a synthetic dataset and a benchmark face dataset CelebA. The synthetic one is named Pendulum which includes images of causally related objects. Each image contains 3 entities (<span style="font-variant:small-caps;">pendulum</span>, <span style="font-variant:small-caps;">light</span>, <span style="font-variant:small-caps;">shadow</span>), and 4 concepts ((<span style="font-variant:small-caps;">pendulum angle</span>, <span style="font-variant:small-caps;">light angle</span>) $\to$ (<span style="font-variant:small-caps;">shadow location</span>, <span style="font-variant:small-caps;">shadow length</span>)). We also use a real world dataset CelebA[^2], a widely used dataset in the computer vision community. In this dataset, there are in total 200k human face images with labels on different concepts, and we focus on 4 causally related concepts (<span style="font-variant:small-caps;">gender</span>, <span style="font-variant:small-caps;">smile</span>, <span style="font-variant:small-caps;">eyes open</span>,<span style="font-variant:small-caps;">mouth open</span>), where <span style="font-variant:small-caps;">gender</span> and <span style="font-variant:small-caps;">smile</span> cause <span style="font-variant:small-caps;">eyes open</span>, and <span style="font-variant:small-caps;">smile</span> causes <span style="font-variant:small-caps;">mouth open</span>. More experimental results on other concepts are provided in the Appendix. Baselines: We compare our method with some state of the arts. They are categorized into supervised and unsupervised methods. CausalVAE-unsup, LadderVAE [@laddervae] and $\beta$-VAE [@higgins2017beta] are unsupervised methods. CausalVAE-unsup is a reduced version of our model whose structure is the same as CausalVAE except that the Mask Layer and the supervision conditional prior $p(\mathbf{z}\|\mathbf{u})$ are removed. Supervised methods include disentangled representation learning method DC-IGN [@kulkarni2015deep] and causal generative model CausalGAN [@causalgan]. As CausalGAN does not focus on representation learning, we only compare our CausalVAE with CausalGAN on intervention experiment (results given in Appendix). For these methods, the prior conditioning on the labels are given, and the dimensionality of the latent representation is the same as CausalVAE. Metrics: We use Maximal Information Coefficient (MIC) and Total Information Coefficient (TIC) [@kinney2014equitability] as our evaluation metrics. Both of them indicate the degree of information relevance between the learned representation and the ground truth labels of concepts. Intervention experiments ------------------------ Intervention experiments aim at testing if a certain dimension of the latent representation has interpretable semantics. The value of a latent code is manipulated by “do-operation” as introduced in previous sections, and we observe how the generated image appears. Intervention is conducted by the following steps: 1) a generative model is trained; 2) an arbitrary image from the training set is fed to the encoder to generate a latent code $\mathbf{z}$. 3) we manipulate the value of $z_i$ corresponding to a concept of interest. For CausalVAE, as Fig. \[cvae\_structure\] shows, we need to manipulate both the input and output nodes of the SCM layer. Note that the effect of manipulation to a parental node will be propagated to its children; 4) The intervened latent code $\tilde{\mathbf{z}}$ passes through the decoder to generate a new image. In the experiments, all images in the dataset are used to train our proposed model CausalVAE and other baselines. Parameters $(\alpha,\beta,\gamma)=(1,1,1)$ for all experiments unless specified. We first conduct intervention experiments on the Pendulum dataset, with 4 latent concepts and results are given in Fig. \[pendulum\_res\]. We intervene a certain concept by setting the corresponding latent code value to 0. We expect that the pattern of the manipulated concept will be fixed across all images under the same intervention. For example, when we intervene the pendulum <span style="font-variant:small-caps;">angle</span> as shown in the first line of Fig. \[pendulum\_res\] (a), the <span style="font-variant:small-caps;">angle</span> of pendulum of different images are almost the same. Meanwhile, we also observe that the <span style="font-variant:small-caps;">Shadow location</span> and <span style="font-variant:small-caps;">Shadow length</span> change in a correct way that aligns with the physics law. Note that this is also related to the concept of modularity, meaning that intervening a certain part of the generative system usually does not affect the other parts of the system. Similar phenomenon is observed in other intervention experiments, demonstrating that our model correctly implement the underlying causal system. The results of DC-IGN, a supervised method without considering the causal structure, are given in Fig. \[pendulum\_res\] (b). There exists a problem that manipulating the latent codes of effects sometimes has no influence to the whole image. This is probably because they do not explicitly consider causal disentanglement. Fig. \[do\_smile\] demonstrates the good result of CausalVAE on real world banchmark dataset CelebA, with subfigures showing the experiments on intervening concepts <span style="font-variant:small-caps;">gender, smile, eyes open</span> and <span style="font-variant:small-caps;">mouth open</span> respectively. We observe that when we intervene the cause concept <span style="font-variant:small-caps;">smile</span>, the status of <span style="font-variant:small-caps;">mouth open</span> also changes. In contrast, intervening effect concept <span style="font-variant:small-caps;">mouth open</span> does not cause the cause concept <span style="font-variant:small-caps;">smile</span> to change. Table 1 records the mutual information (MIC/TIC) between the learned representation and the ground truth concept labels of all compared methods. Our model achieves best alignment with the concept labels, justifying the effectiveness of our proposed method. On the contrary, factors learned by those compared methods have low correlation with the ground truth labels, indicating that those factors are at least not corresponding to the causal concepts of interest. In addition, we show in Fig. \[learn\_A\] the learned adjacency matrix $\mathbf{A}$. As the training epoch increases, we see that the graph learned by our model quickly converges to the true one, which shows that our method is able to correctly learn the causal relationship among the factors. Conclusion ========== In this paper, we investigate an important task of learning disentangled representations of causally related concepts in data, and propose a new framework called CausalVAE which includes a SCM layer to model the causal generation mechanism of data. We prove that the proposed model is fully identifiability given additional supervision signal. Experimental results with synthetic and real data show that CausalVAE successfully learns representations of causally related concepts and allows intervention to generate counterfactual outputs as expected according to our understanding of the causal system. To the best of our knowledge, our work is the first one that successfully implement causal disentanglement and is expected to bring new insights into the domain of disentangled representation learning. Proof of Proposition 1 ====================== Write the KL term in ELBO defined in Eq. 8 in the main text as $$\begin{split} \mathcal{D}[q_{\phi}(\bm{\epsilon},\mathbf{z}\|\mathbf{x},\mathbf{u})\Vert p_{\theta}(\bm{\epsilon},\mathbf{z}\|\mathbf{u})] =& \iint q_{\phi}(\bm{\epsilon},\mathbf{z}\|\mathbf{x},\mathbf{u})\log \frac{q_{\phi}(\bm{\epsilon},\mathbf{z}\|\mathbf{x},\mathbf{u})}{p_{\bm{\epsilon}}(\bm{\epsilon})p_{\theta}(\mathbf{z}\|\mathbf{u})}d\bm{\epsilon}d\mathbf{z}\\ =&\iint q_{\phi}(\bm{\epsilon},\mathbf{z}\|\mathbf{x},\mathbf{u})\log \frac{q_{\phi}(\bm{\epsilon},\mathbf{z}\|\mathbf{x},\mathbf{u})}{p_{\bm{\epsilon}}(\bm{\epsilon})}d\bm{\epsilon}d\mathbf{z}\nonumber\\ &+ \iint q_{\phi}(\bm{\epsilon},\mathbf{z}\|\mathbf{x},\mathbf{u})\log \frac{q_{\phi}(\bm{\epsilon},\mathbf{z}\|\mathbf{x},\mathbf{u})}{p_{\theta}(\mathbf{z}\|\mathbf{u})}d\bm{\epsilon}d\mathbf{z}. \end{split}$$ Based on Eq. 9 in the main text, we have $$\begin{split} &\iint q_{\phi}(\bm{\epsilon},\mathbf{z}\|\mathbf{x},\mathbf{u})\log \frac{q_{\phi}(\bm{\epsilon},\mathbf{z}\|\mathbf{x},\mathbf{u})}{p_{\bm{\epsilon}}(\bm{\epsilon})}d\bm{\epsilon}d\mathbf{z}\\ =&\int q_{\phi}(\bm{\epsilon}\|\mathbf{x},\mathbf{u})\log \frac{q_{\phi}(\bm{\epsilon}\|\mathbf{x},\mathbf{u})}{p_{\bm{\epsilon}}(\bm{\epsilon})}\int\delta(\mathbf{z}=\mathbf{C}\bm{\epsilon})d\mathbf{z}d\bm{\epsilon}\\ &+ \int q_{\phi}(\bm{\epsilon}\|\mathbf{x},\mathbf{u})\int \delta(\mathbf{z}=\mathbf{C}\bm{\epsilon})\log \delta(\mathbf{z}=\mathbf{C}\bm{\epsilon}))d\mathbf{z} d\bm{\epsilon}\\ =&\mathcal{D}[q_{\phi}(\bm{\epsilon}\|\mathbf{x},\mathbf{u})\Vert p_{\bm{\epsilon}}(\bm{\epsilon})]+0\\ =&\mathcal{D}[q_{\phi}(\bm{\epsilon}\|\mathbf{x},\mathbf{u})\Vert p_{\bm{\epsilon}}(\bm{\epsilon})], \end{split}$$ and $$\begin{split} &\iint q_{\phi}(\bm{\epsilon},\mathbf{z}\|\mathbf{x},\mathbf{u})\log \frac{q_{\phi}(\bm{\epsilon},\mathbf{z}\|\mathbf{x},\mathbf{u})}{p_{\theta}(\mathbf{z}\|\mathbf{u})}d\bm{\epsilon}d\mathbf{z}\\ =&\int q_{\phi}(\bm{z}\|\mathbf{x},\mathbf{u})\log \frac{q_{\phi}(\bm{z}\|\mathbf{x},\mathbf{u})}{p_{\theta}(\mathbf{z}\|\mathbf{u})}\int\delta(\bm{\epsilon}=\mathbf{C}^{-1}\bm{z})d\bm{\epsilon}d\mathbf{z}\\ &+ \int q_{\phi}(\bm{z}\|\mathbf{x},\mathbf{u})\int \delta(\bm{\epsilon}=\mathbf{C}\bm{z})\log \delta(\bm{\epsilon}=\mathbf{C}^{-1}\bm{z}) d\bm{\epsilon} d\mathbf{z}\\ =&\mathcal{D}[q_{\phi}(\bm{z}\|\mathbf{x},\mathbf{u})\Vert p_{\theta}(\mathbf{z}\|\mathbf{u})]+0\\ =&\mathcal{D}[q_{\phi}(\bm{z}\|\mathbf{x},\mathbf{u})\Vert p_{\theta}(\mathbf{z}\|\mathbf{u})]. \end{split}$$ Adding up the above two terms leads to the desired form of Proposition 1. Identifiability =============== Proof of Theorem 1 ------------------ The general logic of the proofing follows [@nonlinearica], but we focus on both encoder and decoder. In our setting, we has joint latent variables $\bm{\epsilon}, \mathbf{z}$, and we prove identidfiabilty of both of them. Another different setting from iVAE is that we consider a slighter transformation matrix, since our additional observations $\mathbf{u}$ of each concepts align to each causal representations $\mathbf{z}$. Sketch of proof: We analyze the identifiability of $\bm{\epsilon}$ starting with $p_{\bm{{\bm{{\bm{\theta}}}}}}\mathbf{(\mathbf{x}\|u)}=p_{\widetilde{{\bm{\theta}}}}(\mathbf{x}\|\mathbf{u})$. Then we define a new invertible matrix $\mathbf{L}$ which contains additional observation $u_i$ in causal system, and use it to prove that the learned $\mathbf{\tilde{T}}$ is the transformation of $\mathbf{T}$. [Step 2:]{} We take the inference model into consideration and analyze the identifiablity of the inference model by relating the inference model to the generative model. Details: At the begining of proof, we consider a simple condition that the dimension of observation data $d$ equals to the dimension of latent variables $n$. The distribution has two sufficient statistics, the mean and variance of $\mathbf{z}$, which are denoted by sufficient statistics $\mathbf{T}(\mathbf{z}) = (\bm{\mu}(\mathbf{z}), \bm{\sigma}(\mathbf{z}))=(T_{1,1}(z_1),\dots,T_{n,2}(z_n))$. We use these notations for model identifiability analysis in Section 5. $$\begin{aligned} &p_{\bm{\theta}}\mathbf{(\mathbf{x}\|u)}=p_{\widetilde{{\bm{\theta}}}}(\mathbf{x}\|\mathbf{u}), \nonumber \\ \Rightarrow &\iint_{\mathbf{z},\bm{\epsilon}} p_{\bm{\theta}}(\mathbf{x}\|\mathbf{z},\bm{\epsilon})p_{{\bm{\theta}}}(\mathbf{z},\bm{\epsilon}\|\mathbf{u})d\mathbf{z}d\bm{\epsilon} = \iint_{\mathbf{z},\bm{\epsilon}} p_{\widetilde{{\bm{\theta}}}}(\mathbf{x}\|\mathbf{z},\bm{\epsilon})p_{\widetilde{{\bm{\theta}}}}(\mathbf{z},\bm{\epsilon}\|\mathbf{u})d\mathbf{z}d\bm{\epsilon},\nonumber \\ \Rightarrow &\int_{\mathbf{z}} p_{\bm{\theta}}(\mathbf{x}\|\mathbf{z})p_{{\bm{\theta}}}(\mathbf{z}\|\mathbf{u})d\mathbf{z} = \iint_{\mathbf{z}} p_{\widetilde{{\bm{\theta}}}}(\mathbf{x}\|\mathbf{z})p_{\widetilde{{\bm{\theta}}}}(\mathbf{z}\|\mathbf{u})d\mathbf{z},\nonumber \\ \Rightarrow &\int_{\mathbf{x}'} p_{{\bm{\theta}}}(\mathbf{x}\|\mathbf{f^{-1}}(\mathbf{x}'))p_{{\bm{\theta}}}(\mathbf{f^{-1}}(\mathbf{x}')\|\mathbf{u})\|\det(J_{\mathbf{f^{-1}}}(\mathbf{x}'))\|d\mathbf{x}'\nonumber\\ &= \int_{\mathbf{x}'} p_{{\bm{\theta}}}(\mathbf{x}\|\mathbf{\tilde{f}^{-1}}(\mathbf{x}'))p_{\widetilde{{\bm{\theta}}}}(\mathbf{\tilde{f}^{-1}}(\mathbf{x}')\|\mathbf{u})\|\det(J_{\mathbf{\tilde{f}^{-1}}}(\mathbf{x}'))\|d\mathbf{x}'. \label{eq1}\end{aligned}$$ In determining function $\mathbf{f}$, there exist a Gaussian distribution $p_{\bm{\xi}}(\bm{\xi})$ which has infinitesimal variance. Then, the $p_{{\bm{\theta}}}(\mathbf{x}\|\mathbf{f^{-1}}(\mathbf{x}'))$ can be written as $p_{\bm{\xi}}(\mathbf{x}-\mathbf{x}')$. As the assumption (1) holds, this term is vanished. Then in our method, there exists the following equation: $$\begin{aligned} p_{{\bm{\theta}}}(\mathbf{f^{-1}}(\mathbf{x}')\|\mathbf{u})\|\det(J_{\mathbf{f^{-1}}}(\mathbf{x}'))\|&= p_{\widetilde{{\bm{\theta}}}}(\mathbf{\tilde{f}^{-1}}(\mathbf{x}')\|\mathbf{u})\|\det(J_{\widetilde{\mathbf{f}}^{-1}}(\mathbf{x}'))\|,\nonumber\\ \Rightarrow\widetilde{p}_{\bm{\theta}}(\mathbf{x}) &= \widetilde{p}_{\widetilde{{\bm{\theta}}}} (\mathbf{x}).\end{aligned}$$ Adopting the definition of multivariate Gaussian distribution, we define $$\bm{\lambda}_s(\mathbf{u})= \left[ \begin{array}{ccc} \lambda_1^s(u_1) & & \\ & \ddots & \\ & & \lambda_n^s(u_n) \\ \end{array}\right].$$ There exists the following equations: $$\begin{aligned} &\log\|\det(J_{\mathbf{f}^{-1}}(\mathbf{x}))\| - \log \mathbf{Q}(\mathbf{f}^{-1}(\mathbf{x}))+\log\mathbf{Z(u)}+\sum_{s=1}^2 \mathbf{T}_s (\mathbf{f}^{-1}(\mathbf{x}))\bm{\lambda}_s(\mathbf{u}),\nonumber\\ = &\log\|\det(J_{\widetilde{\mathbf{h}}}(\mathbf{x}))\| -\log \widetilde{\mathbf{Q}}(\tilde{\mathbf{f}}^{-1}(\mathbf{x}))+\log\mathbf{\tilde{Z}(u)}+\sum_{s=1}^2\widetilde{\mathbf{T}}_{s}(\tilde{\mathbf{f}}^{-1}(\mathbf{x}))\widetilde{\bm{\lambda}}_s(\mathbf{u}),\label{original}\end{aligned}$$ where $\mathbf{Q}$ denotes the base measure. In Gaussian distribution, it is $\bm{\sigma}(\mathbf{z})$. In learning process, $\mathbf{\widetilde{A}}$ is restricted as DAG. Thus, the $\mathbf{\widetilde{C}}$ exists which is full rank matrix. The item which is not related to $u$ in Eq. \[original\] are cancelled out [@sorrenson2020disentanglement]. $$\begin{aligned} &\sum_{s=1}^2{\mathbf{T}}_{s}(\mathbf{f}^{-1}(\mathbf{x}))\bm{\lambda}_s(\mathbf{u})= \sum_{s=1}^2\widetilde{\mathbf{T}}_{s}(\tilde{\mathbf{f}}^{-1}(\mathbf{x}))\widetilde{\bm{\lambda}}_s(\mathbf{u}) + \mathbf{b}, \label{eqsimple2}\end{aligned}$$ where $\mathbf{b}$ is a vector related to $\mathbf{u}$. In our model, there exist a deterministic relationship $\mathbf{C}$ between $\bm{\epsilon}$ and $\mathbf{z}$ where $\mathbf{C} = (\mathbf{I}-\mathbf{A}^T)^{-1}$. Thus we could get equivalent of Eq. \[eqsimple2\] as follows, $$\begin{aligned} &\sum_{s=1}^2{\mathbf{T}}_{s}(\mathbf{C}\mathbf{h}(\mathbf{x}))\bm{\lambda}_s(\mathbf{u})= \sum_{s=1}^2\widetilde{\mathbf{T}}_{s}(\widetilde{\mathbf{C}}\widetilde{\mathbf{h}}(\mathbf{x}))\widetilde{\bm{\lambda}}_s(\mathbf{u}) + \mathbf{b}‘, \label{eqsimple3}\end{aligned}$$ where $s$ denote the index of sufficient statistics of Gaussian distributions, indexing the mean (1) and the variance (2). By assuming that the additional observation $u_i$ is different, it is guaranteed that coefficients of the observations for different concepts are distinct. Thus, there exists an invertible matrix corresponding to additional information $\mathbf{u}$: $$\begin{aligned} \mathbf{L}= \left[ \begin{array}{ccc} \bm{\lambda}_1(\mathbf{u}) & \\ & \bm{\lambda}_2(\mathbf{u}) \end{array}\right].\label{matl}\end{aligned}$$ Since the assumption that $u_i\not=0$ holds, $\mathbf{L}$ is $2n \times 2n$ invertible and full rank diagonal matrix. Then, function of $\bm{\lambda}$ in Eq. \[eqsimple2\] and Eq. \[eqsimple3\] are replcaed by Eq. \[matl\], we could get: $$\begin{aligned} &\mathbf{LT}(\mathbf{f}^{-1}(\mathbf{x})) = \widetilde{\mathbf{L}}\widetilde{\mathbf{T}}(\widetilde{\mathbf{f}}^{-1}(\mathbf{x}))+ \mathbf{b},\\ &\mathbf{T}(\mathbf{f}^{-1}(\mathbf{x})) = \mathbf{B_2}\widetilde{\mathbf{T}} (\widetilde{\mathbf{f}}^{-1}(\mathbf{x})) + \mathbf{b_2},\label{iz}\end{aligned}$$ where $$\mathbf{B_2}= \left[ \begin{array}{ccc} {\lambda}_{1,1}(u_1)^{-1}\widetilde{\lambda}_{1,1}(u_1) & & \\ & \ddots & \\ & & {\lambda}_{n,2}(u_n)\widetilde{\lambda}_{n,2}(u_n) \\ \end{array}\right].$$ We replace $\mathbf{f}^{-1}$ with $\mathbf{Ch}$ and we could get the equations as below: $$\begin{aligned} \mathbf{B_3}\mathbf{L}\mathbf{T}(\mathbf{Ch}(\mathbf{x})) = \widetilde{\mathbf{L}}{\widetilde{\mathbf{T}}}(\tilde{\mathbf{C}}\tilde{\mathbf{h}}(\mathbf{x})) \Rightarrow \mathbf{T}(\mathbf{h}(\mathbf{x})) = \mathbf{B_1}{\widetilde{\mathbf{T}}} (\widetilde{\mathbf{h}}(\mathbf{x})) + \mathbf{b_1},\label{ie}\end{aligned}$$ where $\mathbf{B_3}$ is invertible matrix which corresponds to $\mathbf{C}$ and $\mathbf{B_1=\mathbf{L^{-1}B_3^{-1}\widetilde{L}}}$. The definition of $\widetilde{\mathbf{L}}$ on learning model migrates the definition of $\mathbf{L}$ on ground truth. Then we adopt the definitions following [@nonlinearica]. According to the Lemma 3 in [@nonlinearica], we are able to pick out a pair $(\epsilon_i,\epsilon_i^2)$ such that, $(\textbf{T}'_i(z_i),\textbf{T}'_i(z_i^2))$ are linearly independent. Then concat the two points into a vector, and denote the Jacobian matrix $\mathbf{Q} = [J_\mathbf{T}(\bm{\epsilon}),J_\mathbf{T}(\bm{\epsilon}^2)]$, and define $\mathbf{\tilde{Q}}$ on $\mathbf{\tilde{T}}(\mathbf{\tilde{h}} \circ \mathbf{Cf}(\bm{\epsilon}))$ in the same manner. By differentiating Eq. \[ie\], we get $$\begin{aligned} \mathbf{Q} = \mathbf{B_1} \mathbf{\tilde{Q}}.\label{invertibleb}\end{aligned}$$ Since the assumptiom (2) that Jacobian of $\mathbf{f}^{-1}$ is full rank holds, it can prove that both $\mathbf{Q}$ and $\mathbf{\tilde{Q}}$ are invertible matrix. Thus from Eq. \[invertibleb\], $\mathbf{B_1}$ is invertible matrix. Using the same way as shown in Eq. \[invertibleb\], it can prove that $\mathbf{B_2}$ is invertible matrix. Eq. \[iz\] and Eq. \[ie\] both hold. Combining the two results supports the identifiability result in CausalVAE. Extension of Definition 1 ------------------------- In most of scenarios, latent variable is a low dimensional representation of the observation, since we are not interested in all the information in observations. Therefore, we usually have $d>n$. We called it the reduction of dimension. We add auxiliary term as $\mathbf{\lambda(x) = \{\lambda(u),\lambda'\}}$ In our model, Only $n$ components of the latent variable are modulated, and its density has the form: $$\begin{aligned} p_{\bm{\theta}}(\mathbf{z}\|\mathbf{u}) = \frac{\mathbf{Q}(\mathbf{z})}{\mathbf{Z(u)}}\exp{{\sum_i^n\mathbf{T_i}(z_i)\lambda_i(u_i)}}\label{lowdimensional}\end{aligned}$$ and the term $e^{\sum_{n+1}^d\mathbf{T}(z_i)\lambda_i}$ is simply absorbed into $\mathbf{Q(z)}$. When we evaluate Eq. \[original\] by new definition (Eq. \[lowdimensional\]), the dimension of $p(\mathbf{z\|u})$ is $n$, because the remaining part is cancelled out. Assume that $p_{\bm{\theta}}\mathbf{(\mathbf{x}\|u)}$ equal to $p_{\widetilde{{\bm{\theta}}}}(\mathbf{x}\|\mathbf{u})$. For all the observational pairs $(\mathbf{x},\mathbf{u})$, let $J_h$ denote the Jacobian matrix of the encoder function. Following the definition in Theorem 2 in i VAE [@nonlinearica], $\mathbf{B}$ will be indexed by 4 indicates $(i,l, a,b)$, where $1<i<d$ and $1<l<s$ refer to the rows and $1<a<d$ and $1<b<s$ refer to the columns. We define a following equation: $$\begin{aligned} \mathbf{v} = \mathbf{\tilde{C}\circ \tilde{h}\circ f(z)}.\end{aligned}$$ The goal is to show that $v_i(\mathbf{z})$ is a function of only one $z_j$. We denote by $v_i^r := \frac{\partial v_i}{\partial z_r}$ and $v_i^{rt} := \frac{\partial^2 v_i}{\partial z_r \partial z_t}$. By differentiating Eq. \[iz\] with respect to $z_s$, we could get: $$\begin{aligned} T'_{i,l}(z_i) = \sum_{a =1}^d\sum_{b = 1}^s B_{2,(i,l,a,b)} \tilde{T}'_{a,b}(v_a(\mathbf{z}))v_a^r(\mathbf{z}).\end{aligned}$$ Lemma 1 (from Lemma 9 in @khemakhem2020ice): Consider a distribution that follows a strongly exponential family. Its sufficient statistic $\mathbf{\tilde{T}}$ is differentiable almost surely. Then $\tilde{T}'_i \not= 0$ almost everywhere on $\mathbb{R}$ for all $1 \le i \le s$. For $r>n$, $T'_{i,l}(z_i) = 0$, according to Lemma 1, $\tilde{T}'_{a,b}(v_a(\mathbf{z})) \not= 0$, since $\mathbf{B}_2$ is an invertible matrix, we can conclude that $v_a^r(\mathbf{z})=0$ for all $a<n$ and $r>n$. Therefore, we can conclude that each of the first $n$ components of $\mathbf{v}$ is only a function of one different $z_j$. Thus, when $d>n$, we could get the same conclusion as Theorem 1. Implementation Details ====================== We use one NVIDIA Tesla P40 GPU as our training and inference device. For the implementation of CausalVAE and other baselines, we extend $\mathbf{z}$ to matrix $\mathbf{z}\in\mathbb{R}^{n\times k}$ where $n$ is the number of concepts and $k$ is the latent dimension of each $\mathbf{z}_i$. The corresponding prior or conditional prior distributions of CausalVAE and other baselines are also adjusted (this means that we extend the multivariate Gaussian to the matrix Gaussian). The subdimensions $k $ for each synthetic (pendulum, water) experiments are set to be 4, and 32 for CelebA experiments. The implementation of continuous DAG constraint $H(\mathbf{A})$ follows the code of [@yu2019dag] [^3]. Data Preprocessing ------------------ ### Sythetic Simulator 0.2in ![Generate Policy of Pendulum Simulator[]{data-label="pendulum_simulator"}](pendulum_simulator.pdf){width="0.2\columnwidth"} -0.2in Fig. \[pendulum\_simulator\] shows our policy of generating synthetic Pendulum data. The picture includes a pendulum. The angles of pendulum and the light are changing overtime, and projection laws are used to generate the shadows. Given the light <span style="font-variant:small-caps;">position</span> and pendulum <span style="font-variant:small-caps;">angle</span>, we get the angles $\varphi_1$ and $\varphi_2$. Then the system can calculate the shadow <span style="font-variant:small-caps;">position</span> and <span style="font-variant:small-caps;">length</span> using triangular functions. The causal graph of concepts is shown in Fig. \[causal\_example\] (a). In Pendulum generator, the image size is set to be $96 \times 96$ with 4 channels. We generate about 7k images (6k for training and 1k for inference), $\varphi_1$ and $\varphi_2$ are ranged in around $[-\frac{\pi}{4},-\frac{\pi}{4}]$. ### Data Preprocess of CelebA CelebA dataset contains 20K human face images. We preprocess the original dataset by following two steps: \(1) We divided the whole dataset into training dataset $85\%$ and test dataset $15\%$. \(2) We only focus on facial features and resize the picture to be squared ($128\times128$ with 3 channels). Intervention Experiments ------------------------ ### Synthetic In synthetic experiments, we train the model on synthetic data for 80 epochs, and use this model to generate latent code of representations. The hyperparameters of baselines are defined as default. For CausalVAE, we set the $\alpha = 0.3$ and $(\beta, \gamma) = (1,1)$. We use $\mathcal{N}(\mathbf{u,\|u\|})$ as the condition prior $p_{\bm{\theta}}(\mathbf{z\|u})$. In the implementation of CausalVAE, $\|\mathbf{z}_{\text{mean}}\|$ is used as the variance of condition prior. The details of the neural networks are shown in Table \[do\_table1\]. ### CelebA We also present the DO-experiments of CausalVAE and CausalGAN. In the training of the models, we use face labels (<span style="font-variant:small-caps;">age</span>, <span style="font-variant:small-caps;">gender</span> and <span style="font-variant:small-caps;">beard</span>). For CausalVAE, we set the $\alpha = 0.3$ and $(\beta, \gamma) = (1,1)$. We use $\mathcal{N}(\mathbf{u,I})$ as the condition prior $p_{\bm{\theta}}(\mathbf{z\|u})$. For all the baseline, default hyperparameters and one common encoder and decoder structure are employed. For CausalGAN, we use the publicly available code[^4]. For all the VAE-based methods, mean and variance of the distribution of the latent variable are learned during training, and the latent code $z$ are sampled from Conditional Gaussian Distribution $p_{\bm{\theta}}(\mathbf{z\|u})$. In all experiments, we rescale the variance of learned representation $\mathbf{z}$ by multiplying a factor 0.1 to the original one. Training epoches for the model is set to be 80, and our proposed CausalVAE has a pretrain step to learn causal graph $\mathbf{A}$, which takes 10 epochs. The details of the neural networks are shown in Table \[do\_table2\]. The Pretrain Step for Causal Graph Learning ------------------------------------------- In our model, we need to learn the latent representation $\mathbf{z}$ and causal graph $\mathbf{A}$ simultaneously, whose optimal solution is not easy to find. Thus we adopt a pretrain stage to learn the causal graph $\mathbf{A}$ in the Mask Layer. We adopt the augmented Lagrangian to learn $\mathbf{A}$ in CausalVAE from the labels $\mathbf{u}$ in Mask Layer first. During the pretrain process, we truncate the gradient of other part of model and solve the optimization problem in Eq. \[pretrain\_eq\] to learn $\mathbf{A}$. The augmentation approach is widely used in causal discovery method, like NOTEARS [@shimizu2006linear], DAG-GNN [@yu2019dag]. The pretrain is a stage that learns the graph by optimizing the following objective functions: $$\begin{aligned} \text{minimize} & ~~~ l_u=\mathbb{E}_{q_D}\Vert \mathbf{u} -\mathbf{A}^T \mathbf{u}\Vert_2^2 \nonumber \\ \text{subject to} & ~~~ H(\mathbf{A})=0\end{aligned}$$ Then, we define an augmented Lagrangian: $$\begin{aligned} \label{pretrain_eq} l_{pre} = l_u + \lambda H(\mathbf{A}) + \frac{c}{2} H^2(\mathbf{A})\end{aligned}$$ where $\lambda$ is the Lagrangian multiplier and $c$ is the penalty. The following policy is used to update the $\lambda$ and $c$: $$\begin{aligned} \lambda_{s+1} = \lambda_{s} +c_{s} H(\mathbf{A}_s)\end{aligned}$$ $$c_{s+1}=\left\{ \begin{aligned} c_{s} & = \eta c_{s}, & ~~~ if \|H(\mathbf{A}_s)\| > \gamma\|H(\mathbf{A}_{s-1})\| \\ c_{s} & = c_{s}, & ~~~ otherwise \end{aligned} \right.$$ where $s$ is the iteration. In our experiments, we set $\eta = 10$ and $\gamma = \frac{1}{4}$. \[tab:freq1\] encoder decoder ------------------------------- ---------------------------------------------- 4\96\96$\times$900 fc. 1ELU concepts$\times$( 4$\times$ 300 fc. 1ELU ) 900$\times$300 fc. 1ELU concepts$\times$ (300$\times$300 fc. 1ELU) 300$\times$2\concepts\k fc. concepts$\times$(300$\times$ 1024 fc. 1ELU) - concepts$\times$(1024$\times$ 4\96\96 fc.) \[tab:freq1\] encoder decoder --------------------------------------------- ------------------------------------------------------- - (1$\times$1 conv. 128 1LReLU(0.2), stride 1) 4$\times$4 conv. 32 1LReLU (0.2), stride 2 (4$\times$4 convtranspose. 64 1LReLU (0.2), stride 1) 4$\times$4 conv. 64 1LReLU (0.2), stride 2 (4$\times$4 convtranspose. 64 1LReLU (0.2), stride 2) 4$\times$4 conv. 64 1LReLU(0.2), stride 2 (4$\times$4 convtranspose. 32 1LReLU (0.2), stride 2) 4$\times$4 conv. 64 1LReLU (0.2), stride 2 (4$\times$4 convtranspose. 32 1LReLU (0.2), stride 2) 4$\times$4 conv. 256 1LReLU (0.2), stride 2 (4$\times$4 convtranspose. 32 1LReLU (0.2), stride 2) 1$\times$1 conv. 3, stride 1 (4$\times$4 convtranspose. 3 , stride 2) Additional Experimental Results =============================== In this section, we show more experimental results. Fig. \[causal\_example\] shows the causal graph among concepts in different dataset respectively. We here show results including experiments analyzing the properties of learned representation, intervening results and the learning process of the causal graph. 0.2in ![Causal graphs of three dataset. (a) shows the causal graph in pendulum dataset. The concepts are <span style="font-variant:small-caps;">pendulum</span> <span style="font-variant:small-caps;">angle</span>, light <span style="font-variant:small-caps;">position</span>, <span style="font-variant:small-caps;">shadow position</span> and <span style="font-variant:small-caps;">shadow length</span>. (b) shows the causal graph in CelebA, on concepts <span style="font-variant:small-caps;">age</span>, <span style="font-variant:small-caps;">gender</span> and <span style="font-variant:small-caps;">beard</span> and <span style="font-variant:small-caps;">bald</span>. (c) shows the causal graph in CelebA, on concepts <span style="font-variant:small-caps;">gender</span>, <span style="font-variant:small-caps;">smile</span>, <span style="font-variant:small-caps;">eyes open</span> and <span style="font-variant:small-caps;">mouth open</span>. []{data-label="causal_example"}](True_graph/true_graph.pdf){width="0.8\columnwidth"} -0.2in The Property of Learned Representation -------------------------------------- We test our method and baselines on both synthetic data and benchmark human face data. In the previous section, we already show the relationships between the learned representation $\tilde{\mathbf{z}}$ and the target representation $\mathbf{z}$ (related by a linear transformation formed as a diagonal matrix). In this section, we visualize it by scatter plot. One of the important aspect of the generative model is that whether the learned representation aligns to the conditional prior we set. Our conditional prior is generated by the true label of each concept. The results show that the learned representations align to the expected representations. In figures, points are sampled from the joint distribution, and each color corresponds to one dimension. The additional observations (labels) of Pendulum dataset and those of CelebA dataset are different. In Pendulum, the labels are values within a fixed range The labels in CelebA dataset are discrete (in $\{-1,1\}$). Thus the scatter plots are different. The results show that the performance of our proposed method is better than all the baselines, including the supervised method and unsupervised method. 0.2in ![The figure shows the alignment of ground truth $p(\mathbf{z}\|\mathbf{u})$ and the learned latent factors $q(\mathbf{z}\|\mathbf{x,u})$ on pendulum experiments. Although DC-IGN is also the supervised method, our proposed CausalVAE shows a better performance. []{data-label="align_pendulum"}](result_align/pendulum/align_df_80.pdf){width="\columnwidth"} -0.2in 0.2in ![The figure shows the alignment of ground truth $p(\mathbf{z}\|\mathbf{u})$ and the learned latent factors $q(\mathbf{z}\|\mathbf{x,u})$ on CelebA for the concepts Group 1. The ground truth is a discrete distribution over $\{-1,1\}$, and the color of the points indicates different dimensions. The factors learned by CausalVAE show the best alignment among all.[]{data-label="align_beard"}](result_align/face/align_face_80.pdf){width="\columnwidth"} -0.2in ![The figure shows the alignment between ground truth $p(\mathbf{z}\|\mathbf{u})$ and the learned latent factors $q(\mathbf{z}\|\mathbf{x,u})$ on CelebA for 5 methods (CausalVAE, DC-IGN, $\beta$-VAE, CausalVAE-unsup, LadderVAE from left to right). The ground truth is a distribution with mean taken from $\{-1,1\}$, and the color of the points indicates different dimensions. The factors learned by CausalVAE show the best alignment among all. The concepts include: 1 <span style="font-variant:small-caps;">gender</span>; 2 <span style="font-variant:small-caps;">smile</span>; 3 <span style="font-variant:small-caps;">eyes open</span>; 4 <span style="font-variant:small-caps;">mouth open</span>.[]{data-label="align_smile"}](result_align/face/align_face_smile80.pdf){width="\columnwidth"} -0.2in The Learned Graph ----------------- We demonstrate the learning process of causal graph in this section. Fig. \[beard\_graph\] shows the graph learned process of CelebA Group 1. In this process, we initialize all the entries in $\mathbf{A}$ as 0.5. After 5 epochs, the graph converges. We observe an almost correct graph in this group of concepts. Intervention Results -------------------- 0.2in ![Intervention method[]{data-label="decodernet"}](DO.pdf){width="0.3\columnwidth"} -0.2in The intervention operations are as: - For the learned model, we first put an random observed image $\mathbf{x}$ into the encoder. In this process we could get $\bm{\epsilon}$ and $\mathbf{z}$. - Then for i-th concept, we fix the value of $z_i$ and $g_i({\mathbf{A_i\circ z}})$ as constants. - Finally, we put the new $\mathbf{z}$ into the decoder and get $\mathbf{x'}$. The Fig. \[face\_res\_1\] demonstrates the result of CausalVAE on real world banchmark dataset CelebA Group 1, with subfigures (a) (b) (c) (d) showing the intervention experiments on concepts of <span style="font-variant:small-caps;">age</span>, <span style="font-variant:small-caps;">gender</span>, <span style="font-variant:small-caps;">bald</span> and <span style="font-variant:small-caps;">beard</span> respectively. The interventions perform well that when we intervened the cause concept <span style="font-variant:small-caps;">gender</span>, the <span style="font-variant:small-caps;">beard</span> changes correspondingly. Similarly, when the cause concept <span style="font-variant:small-caps;">age</span> in intervened, its child concept <span style="font-variant:small-caps;">bald</span> also changes. In contrast, intervening effect concept <span style="font-variant:small-caps;">beard</span> does not influence the causal concepts <span style="font-variant:small-caps;">gender</span> and other unrelated concepts in Fig. \[face\_res\_1\] (d). Fig. \[ganbeard\] demonstrates the results of CausalGAN, with subfigures (a) (b) (c) (d) showing the intervention experiments on concepts CelebA Group 1. We observe that when we intervene <span style="font-variant:small-caps;">gender</span>, the <span style="font-variant:small-caps;">beard</span> are changed. But when we intervene <span style="font-variant:small-caps;">beard</span>, concept <span style="font-variant:small-caps;">gender</span> is also changed in third line as shown by Fig. \[ganbeard\] (d). In general, the ’do-intervention’ of CausalGAN performs worse than CausalVAE. The Fig. \[face\_res\_2\] demonstrates the result of CausalVAE on real world banchmark dataset CelebA Group 2, with subfigures (a) (b) (c) (d) showing the intervention experiments on concepts of <span style="font-variant:small-caps;">gender</span>, <span style="font-variant:small-caps;">smile</span>, <span style="font-variant:small-caps;">mouth open</span> and <span style="font-variant:small-caps;">eyes open</span> respectively. The interventions perform well that when we intervened the cause concept <span style="font-variant:small-caps;">gender</span>, not only the appearance of <span style="font-variant:small-caps;">gender</span> but the eyes changed. When we intervened the cause concept <span style="font-variant:small-caps;">smile</span>, not only the appearance of <span style="font-variant:small-caps;">smile</span> but the <span style="font-variant:small-caps;">mouth open</span>. In contrast, intervening effect concept <span style="font-variant:small-caps;">mouth open</span> does not influence the causal concepts <span style="font-variant:small-caps;">smile</span> in Fig. \[face\_res\_2\] (d). Fig. \[gansmile\] demonstrates the results of CausalGAN, with subfigures (a) (b) (c) (d) showing the intervention experiments on concepts CelebA Group 2. We find that when we control <span style="font-variant:small-caps;">smile</span>, the mouth is changed, as shown in the second line of Fig. \[gansmile\] (b). But we find sometimes the control of <span style="font-variant:small-caps;">smile</span> influence other unrelated concepts like <span style="font-variant:small-caps;">gender</span> (shown in first line of Fig. \[gansmile\] (b)). In this concepts group, CausalGAN also shows relatively unstable intervention experiments compared to that of ours. [^1]: (rank equals to its smaller dimension) [^2]: http://mmlab.ie.cuhk.edu.hk/projects/CelebA.html [^3]: https://github.com/fishmoon1234/DAG-GNN [^4]: https://github.com/mkocaoglu/CausalGAN	{ "pile_set_name": "ArXiv" }
--- abstract: 'Assimilation of data into a fire-spread model is formulated as an optimization problem. The level set equation, which relates the fire arrival time and the rate of spread, is allowed to be satisfied only approximately, and we minimize a norm of the residual. Previous methods based on modification of the fire arrival time either used an additive correction to the fire arrival time, or made a position correction. Unlike additive fire arrival time corrections, the new method respects the dependence of the fire rate of spread on diurnal changes of fuel moisture and on weather changes, and, unlike position corrections, it respects the dependence of the fire spread on fuels and terrain as well. The method is used to interpolate the fire arrival time between two perimeters by imposing the fire arrival time at the perimeters as constraints.' author: - Angel Farguell Caus - James Haley - \| Adam K. Kochanski\ Ana Cortés Fit[é]{} - Jan Mandel bibliography: - '../../references/geo.bib' - '../../references/other.bib' - '../../references/bddc.bib' - '../../references/slides.bib' title: Assimilation of fire perimeters and satellite detections by minimization of the residual in a fire spread model --- Fire spread model ================= The state of the fire spread model is the fire arrival time $T\left( x,y\right) $ at locations $\left( x,y\right) $ in a rectangular simulation domain $\Omega\subset\mathbb{R}^{2}$. The isoline $T\left( x,y\right) =c$ is then the fire perimeter at time $c$. The normal vector to the isoline is $\nabla T/\left\Vert \nabla T\right\Vert $. The rate of spread in the normal direction and the fire arrival time at a location on the isoline then satisfy the eikonal equation$$\left\Vert \nabla T\right\Vert =\frac{1}{R}.\label{eq:eikonal}$$ We assume that $R$ depends on location (because of different fuel, fuel moisture, and terrain) and time (because of wind and fuel moisture changing with time). Rothermel’s model [@Rothermel-1972-MMP-x] for 1D fire spread postulates$$R=R_{0}(1+\phi_{w}+\phi_{s}),\label{eq:roth}$$ where $R_{0}$ is the omnidirectional rate of spread, $\phi_{w}$, the wind factor, is a function of wind in the spread direction, and $\phi_{s}$, the slope factor, is a function of the terrain slope. The 1D model was adapted to the spread over 2D landscape by postulating that the wind factor and the slope factor are functions of the components of the wind vector and the terrain gradient in the normal direction. Thus,$$R=R\left( x,y,T\left( x,y\right) ,\nabla T\left( x,y\right) \right) .\label{eq:R}$$ The fire spread model is coupled to an atmospheric model. The fire emits sensible and latent heat fluxes, which change the state of the atmosphere, and the changing atmospheric conditions in turn impact the fire (Fig. \[fig:diagram\]). Wind affects the fire directly by the wind factor, and temperature, relative humidity and rain affect the fire through changing fuel moisture. The fire model is implemented on a rectangular mesh by finite differences. For numerical reasons, the gradient in the eikonal equation (\[eq:eikonal\]) needs to be implemented by an upwinding-type method [@Osher-2003-LSM], which avoids instabilities caused by breaking causality in fire propagation: for the computation of $\nabla T$ at a location $\left( x,y\right) $, only the values from the directions that the fire is coming from should be used, so the methods switch between one-sided differences depending on how the solution evolves. Sophisticated methods of upwinding type, such as ENO or flux-limiters [@Rehm-2009-FPU], aim to use more accurate central differences and switch to more stable one-sided upwind differences only as needed. Unfortunately, the switching causes the numerical gradient of $T$ at a mesh node become a nondifferentiable function of the values of $T$ at that point and its neighbors. In addition, we have added a penalty term to prevent the creation of local minima. It was observed in [@Mandel-2009-DAW] that if, in the level set method, a local minimum appears on the boundary, its value keeps decreasing out of control; we have later found out that this can in fact happen anywhere in the presence of spatially highly variable rate of spread, and we have observed a similar effect here during the minimization process. Fitting the fire spread model to data ===================================== Minimal residual formulation {#sec:eikonal} ---------------------------- Consider the situation when the two observed fire perimeters $\Gamma_{1}$ and $\Gamma_{2}$ at times $T_{1}<T_{2}$ are known, and we are interested in the fire progression between the two perimeters. Aside from immediate uses (visualization without jumps, post-fire analysis), such interpolation is useful to start the fire simulation from the larger perimeter $\Gamma_{2}$ at time $T_{2}$ by a spin-up of the atmospheric model by the heat fluxes from the interpolated fire arrival time between the fire perimeters; the coupled model can then start from perimeter $\Gamma_{2}$ at time $T_{2}$ in a consistent state between the fire and the atmosphere. Interpolation between an ignition point and a perimeter can be handled the same way, with the perimeter $\Gamma_{1}$ consisting of just a single point. In this situation, we solve the eikonal equation (\[eq:eikonal\]) only approximately, $$\left\Vert \nabla T\right\Vert \approx\frac{1}{R} \label{eq:approx}$$ imposing the given fire perimeters as constraints,$$T=T_{1}\text{ at }\Gamma_{1}\text{,}\quad T=T_{2}\text{ at }\Gamma_{2}. \label{eq:perimeter-conditions}$$ We formalize (\[eq:approx\]) as the minimization problem $$J(T)=\left( \int_{\Omega}\left\vert f(\left\Vert \nabla T\right\Vert _{2}^{2},R^{2})\right\vert ^{p}\right) ^{1/p}\rightarrow\min_{T}\text{ subject to (\ref{eq:perimeter-conditions}),} \label{eq:constrained-min}$$ where $f\left( x,y\right) $ is a function such that $f\left( x,y\right) =0$ if and only if $xy=1$, and $\Omega$ is the simulation domain. We mostly use the function $f\left( x,y\right) =1-xy$ but other functions, such as $f\left( x,y\right) =x-1/y$ have advantages in some situations. There are no boundary conditions imposed on the boundary of $\Omega$. Discretization and the constraint matrix {#sec:H} ---------------------------------------- The fire simulation domain is discretized by a logically rectangular grid (aligned approximately with longitude and latitude) and perimeters are given as shape files, i.e., collections of points on the perimeter. We express (\[eq:perimeter-conditions\]) in the form $$HT=g, \label{eq:perimeter-cond-discrete}$$ where $H$ is a sparse matrix. Since the points in the shape files do not need to lie on the grid, the rows of $H$ are the coefficients of an interpolation from the grid to the points in the shape files, which define the perimeters. We find the coefficients from barycentric interpolation. The rectangles of the grid are split into two triangles each, and, for each triangle, we compute the barycentric coordinates of the points in the shapefile, i.e., the coefficients of the unique linear combination of the vertices of the triangle that equals to the point in the shape file. If all 3 barycentric coordinates are in $\left[ 0,1\right] $, we conclude that the point is contained in the triangle, the barycentric coordinates are the sought interpolation coefficients, and they form one row of $H$. For efficiency, most points in the shapefile are excluded up front, based on a comparison of their coordinates with the vertices of the triangle, which is implemented by a fast binary search. When there is more than one point of the shapefile in any triangle, we condense them into a single constraint, obtained by adding the relevant rows of $H$. This way, we avoid over constraining the fire arrival time near the perimeter, which should be avoided for the same reason as limiting the number of constraints in mixed finite elements to avoid locking, cf., e.g., [@Brezzi-1991-MHF]. Numerical minimization of the residual {#sec:nummin} -------------------------------------- To solve (\[eq:constrained-min\]) numerically, we use a multiscale descent method similar to multigrid, combining line searches in the direction of changes of the value of $T$ at a single point, and linear combinations of point values as in [@McCormick-1983-UMS]. We use bilinear coarse grid functions with the coarse mesh step growing by a factor of 2. See Fig. \[fig:multigrid\](b) for an example of a coarse grid function with distance between nodes 16 mesh steps on the original, finest level. We start from an initial approximate solution that satisfies the constraint $HT=g$ exactly, and project all search directions on the subspace $Hu=0$, so that the constraint remains satisfied throughout the iterations. To find a reasonable initial approximation to the fire arrival time, we solve the quadratic minimization problem$$I\left( T\right) =\frac{1}{2}\int\limits_{\Omega}\left\Vert \left( -\triangle\right) ^{\alpha/2}T\right\Vert ^{2}dxdy\rightarrow\min_{T}\text{ subject to (\ref{eq:perimeter-conditions}) and }\frac{\partial T}{\partial\nu }=0, \label{eq:laplace}$$ where $\nu$ is the normal direction, $\triangle=\frac{\partial^{2}}{\partial x^{2}}+\frac{\partial^{2}}{\partial y^{2}}$ is the Laplace operator, and $\alpha>1$ is generally non-integer. The reason for choosing $\alpha>1$ is that $\sqrt{I\left( T\right) }$ is the Sobolev $W^{\alpha,2}\left( \Omega\right) $ seminorm and in 2D, the space $W^{\alpha,2}\left( \Omega\right) $ is embedded in continuous functions if and only if $\alpha>1$. Consequently, $I\left( T\right) $ is not a bound on the value $T\left( x,y\right) $ at any particular point, only averages over some area can be controlled. Numerically, when $\alpha=1$, minimizing $I\left( T\right) $ with a point constraint, such as an ignition point, results in $T$ taking the shape of a sharp funnel at that point (Fig. \[fig:initial\]), which becomes thinner as the mesh is refined. That would be definitely undesirable. The discrete form of (\[eq:laplace\]) is $$\frac{1}{2}\left\langle ST,T\right\rangle -\left\langle f,T\right\rangle \rightarrow\min_{T}\text{ subject to }HT=g,\label{eq:constr-discrete}$$ where $S=A^{\alpha}$ with $(-A)$ a discretization of the Laplace operator with Neumann boundary conditions. To solve (\[eq:constr-discrete\]), we first find a feasible solution $u_{0}=H^{\prime}\left( HH^{\prime}\right) ^{-1}g$, so that $Hu_{0}=g$, substitute $T=u_{0}+v$ to get $$\frac{1}{2}\left\langle S\left( u_{0}+v\right) ,u_{0}+v\right\rangle -\left\langle f,u_{0}+v\right\rangle \rightarrow\min_{T}\text{ subject to }Hv=0,$$ and augmenting the cost fuction, we get that (\[eq:constr-discrete\]) is equivalent to$$\frac{1}{2}\left\langle SPv,Pv\right\rangle +\frac{\rho}{2}\left\langle \left( I-P\right) v,v\right\rangle -\left\langle f_{0},v\right\rangle \rightarrow\min_{T}\text{ subject to }Hv=0\text{,}\label{eq:constr-homog}$$ where $f_{0}=f-Su_{0}$, $P=I-H^{\prime}\left( H^{\prime}H\right) ^{-1}H$ is the orthogonal projection on the nullspace of $H$, and $\rho>0$ is an arbitrary regularization parameter. We solve the minimization problem (\[eq:constr-homog\]) approximately by preconditioned conjugate gradients for the equivalent symmetric positive definite linear system$$P\left( SPv-f_{0}\right) +\rho\left( I-P\right) v=0.\label{eq:linear}$$ Since $S$ is discretization of the Neumann problem, the preconditioner requires some care. Define $Z$ as the vector that generates the nullspace of $S$, which consists of the discrete representation of constant functions, and $P_{Z}=I-Z^{\prime}\left( Z^{\prime}Z\right) ^{-1}Z$ the orthogonal projection on its complement. We use the preconditioner $$M:r\mapsto PP_{Z}S^{+}P_{Z}Pr,$$ where $S^{+}$ is the inverse of $S$ on the complement of its nullspace, and recover the solution by $T=u_{0}+Pv$. The method only requires access to matrix-vector multiplications by $S$ and $S^{+}$, which are readily implemented by cosine FFT. We only need to solve (\[eq:linear\]) to low accuracy to get a reasonable starting point for the nonlinear iterations, but the satisfaction of the constraint $HT=g$ to rounding precision is important. Conclusions =========== We have presented a new method for fitting data by an approximate solution of a fire spread model. The method was illustrated on an idealized example. Application to a real problem are forthcoming. Acknowledgments =============== This research was partially supported by grants NSF ICER-1664175 and NASA NNX13AH59G, and MINECO-Spain under contract TIN2014-53234-C2-1-R. High-performance computing support at CHPC at the University of Utah and Cheyenne (doi:10.5065/D6RX99HX) at NCAR CISL, sponsored by the NSF, are gratefully acknowledged. \[sect:bib\]	{ "pile_set_name": "ArXiv" }
--- abstract: 'In the central regions of active galaxies, dense molecular medium are exposed to various types of radiation and energy injections, such as UV, X-ray, cosmic ray, and shock dissipation. With the rapid progress of chemical models and implementation of new-generation mm/submm interferometry, we are now able to use molecules as powerful diagnostics of the physical and chemical processes in galaxies. Here we give a brief overview on the recent ALMA results to demonstrate how molecules can reveal underlying physical and chemical processes in galaxies. First, new detections of Galactic molecular absorption systems with elevated HCO/H$^{13}$CO$^+$ column density ratios are reported, indicating that these molecular media are irradiated by intense UV fields. Second, we discuss the spatial distributions of various types of shock tracers including HNCO, CH$_3$OH and SiO in NGC 253 and NGC 1068. Lastly, we provide an overview of proposed diagnostic methods of nuclear energy sources using ALMA, with an emphasis on the synergy with sensitive mid-infrared spectroscopy, which will be implemented by JWST and SPICA to disentangle the complex nature of heavily obscured galaxies across the cosmic time.' --- ALMA detections of new Galactic molecular absorption systems: HCO as an indicator of intense UV fields ====================================================================================================== Molecular medium in front of bright background continuum sources can often be observed as absorption lines, allowing us to make sensitive studies of physical and chemical properties of the interstellar medium (ISM) in various environments, including quiescent and translucent molecular media in the Milky Way, dense molecular gas in the vicinity of active galactic nuclei (AGNs), and the ISM in the early epoch of the universe. In fact, because the absorption line depths are independent of the distance to the absorption system, they offer a very powerful technique to uncover cold molecular gas in the distant universe, which is often difficult to observe in emission unless it is intensely heated by either star-formation or active nuclei. Since the first cosmological detection of such molecular absorption system toward PKS 1413+135 at $z=0.247$ ([@Wiklind1994 Wiklind & Combes 1994]), it has been demonstrated that molecular absorption systems are extremely useful for studying chemical properties of quiescent ISM (e.g., [@Muller2014 Muller 2014]), for determining the redshift of a foreground object in a galaxy-galaxy lens system ([@Wiklind1996 Wiklind & Combes 1996]), and for constraining fundamental physical quantities and constants such as cosmic background temperature (e.g., [@Henkel2009 Henkel 2009], [@Muller2011 Muller 2011]) and proton-electron mass ratios (e.g., [@Kanekar2015 Kanekar 2015]) across cosmic time. Despite of their usefulness, however, only four molecular absorption systems beyond $z>0.1$ are currently known at millimeter wavelengths ([@Combes2008 Combes 2008]; [@Curran2011 Curran 2011]). Another important case is a molecular absorption system in which absorption occurs near an AGN, i.e., presumably in the putative dense obscuring tori around AGNs. For instance, Centaurus A is known to exhibit strong absorption line features seen in both molecules (e.g., CO and HCO$^+$) and neutral hydrogen (e.g., [@Espada2010 Espada 2010]; see also Ott in this volume). However, the number of such systems, which are believed to originate from such $\sim$pc scale tori, is very small again. NGC 4261 is another example showing nuclear absorption lines in both CO and HI ([@Jaffe1994 Jaffe & McNamara 1994]) but recent sensitive search for mm-wave absorption features using the Nobeyama Millimeter Array (NMA) and the Plateau de Bure Interferometer (PdBI) yields no significant detections of either CO(1-0) or CO(2-1), giving very small optical depths for these transitions ($\tau_{\rm CO(2-1)} <0.098$, [@Okuda2013 Okuda 2013]). Such molecular absorption lines are also observed in the diffuse medium within the Galaxy against the background extragalactic continuum sources. This also gives an important way to unveil “dark” molecular medium, i.e., invisible in emission because they are often in the equilibrium with the background radiation. Currently, the number of such extragalactic continuum sources with known Galactic molecular absorption lines is limited to $\sim$30. Recently, the first detections of HCO, the formyl radical have been reported toward two extragalactic continuum sources ([@Liszt2014 Liszt 2014]), opening a new window for studying the impact of UV radiation on these diffuse media because the coexistence of H$_2$ and C$^+$, i.e., the presence of photodissociation regions (PDRs), is required for efficient HCO production ([@Gerin2009 Gerin 2009]). Again we need to expand the sample of Galactic absorption systems with such rare species to obtain a comprehensive picture of the diffuse molecular media within the Galaxy, but a survey of new molecular absorption systems is always very time consuming and entails a significant risk of non-detection. [@Ando2015 Ando (2015)] propose using the ALMA archive to search for new molecular absorption systems in mm/submm wavelengths, because the ALMA archive already offers a vast number of sensitive mm/submm spectra of extragalactic continuum sources, i.e., bandpass and/or visibility calibrators, which are essentially radio-loud quasars including BL-LAC objects. Because of the enormous improvement in sensitivity provided by ALMA, these calibrators in the ALMA archive are potentially ideal [free lunch]{} in the quest for molecular absorption systems from the Galaxy to the high-$z$ universe without any new investment of observing time! They analyzed 36 calibrator sources in the ALMA archive to create 3D cubes and search for any absorption features. This resulted in four new detections of absorption line features, and three of which are found to be newly detected Galactic absorption systems. Another one is NRAO530, which is a known absorption system, but strong HCO absorption lines have been detected toward NRAO530 for the first time (Figures \[fig1\] and \[fig2\]). With these detections, this doubles the number of Galactic molecular absorption systems exhibiting HCO absorption features, and it was shown that these Galactic diffuse molecular media are in an extreme condition similar to the Horsehead Nebula, a prototypical PDR (Figure \[fig3\]). ![ALMA band-3 spectrum toward the continuum source J1717-337 showing detections of hyperfine components of C$_2$H in absorption. At least four different velocity components are visible despite of the short integration time (of just a few minutes), demonstrating the superb sensitivity of ALMA. Figure modified from [@Ando2015 Ando (2015)]. []{data-label="fig1"}](Kohno-fig1.eps){width="50.00000%"} ![HCO radical and H$^{13}$CO$^+$ absorption lines toward J1717-337 and NRAO530. Figure modified from [@Ando2015 Ando (2015)]. []{data-label="fig2"}](Kohno-fig2.eps){width="50.00000%"} ![Column density ratios of HCO to H$^{13}$CO$^+$ as a function of the column density of H$^{13}$CO$^+$, including two new detections of HCO absorption (J1717-337 and NRAO530; [@Ando2015 Ando 2015]). Molecular absorption systems toward J1717-337, NRAO530, 3C111, and BL-LAC show elevated HCO/H$^{13}$CO$^+$ column density ratios similar to the value in the Horsehead Nebula ([@Gerin2009 Gerin 2009]), indicating that the diffuse media in the Galaxy seen toward these radio sources are exposed in the intense UV field. Figure modified from [@Ando2015 Ando (2015)]. []{data-label="fig3"}](Kohno-fig3.eps){width="50.00000%"} Shock tracers in the central regions of active galaxies: HNCO, CH$_3$OH, and SiO in NGC 253 and NGC 1068 ======================================================================================================== SiO is considered to be a tracer of shocks, although it could also be related to the presence of X-ray radiation (e.g., [@Garcia-Burillo2000 García-Burillo 2000]). SiO abundance is enhanced by ejection of significant Si from sputtered silicate grain cores in high velocity shocks (e.g., [@Martin-Pintado1992 Martín-Pintado 1992]). In contrast, HNCO forms in the grain mantle, and it can be sublimated in weaker shock events; therefore the HNCO/SiO ratio can be an indicator of shock strength (see also Kelly et al. in this volume). HNCO is also known to be fragile, being easily photodissociated by UV radiation. It is then proposed that the HNCO/CS ratio can give an indicator of starburst evolutionary stage balancing between shock and photodissociation ([@Martin2009 Martín 2009]). Methanol, CH$_3$OH, is formed through grain-surface reactions. A series of hydrogenations starting with CO in cold ($\sim$10 K) dust ([@Watanabe2002 Watanabe & Kouchi 2002]) forms CH$_3$OH efficiently, which then sublimes into the gas phase as the result of shock waves ([@Viti2011 Viti 2011]) or warms up dust grains in hot cores ([@Garrod2008 Garrod 2008]). It is also suggested that methanol will be easily dissociated by cosmic-ray and UV radiation (e.g., [@Aladro2013 Aladro 2013]). Recent ALMA observations of the central few-100-pc regions of NGC 253 depict the spatial variation of these shock tracers; SiO(2-1) shows a series of clumps with a significant peak at the very center of NGC 253, where an intense H40$\alpha$ recombination line is observed. Meanwhile, HNCO has no significant peak at the very center of the galaxy, suggesting that HNCO is photodissociated by intense UV radiation ([@Meier2015 Meier 2015]), although the overall extent of both emissions is quite similar. In contrast, in the central $\sim$kpc region of NGC 1068, there are striking differences between tracers of mild and strong shocks. [@Takano2014 Takano (2014)] presented the first methanol image of NGC 1068, showing clumps across the starburst ring or arms, as seen in the case of CO and its isotopologs, suggesting that mild shocks are ubiquitous in the arms hosting bursts of massive star formation. Meanwhile, PdBI observations reveal that SiO is concentrated near the circumnuclear disk (CND), without any significant emission across the starbursting spiral arms ([@Garcia-Burillo2010 García-Burillo 2010]; see also Kelly, Viti & Garcia-Burillo in this volume for new SiO(3-2) and HNCO(6-5) data taken with PdBI). New ALMA cycle 2 images of CH$_3$OH and HNCO with improved sensitivity are displayed in Figure \[fig4\] (Tosaki in prep.). We find a striking similarity between CH$_3$OH and HNCO, confirming that mild shocks are ubiquitous in the starbursting spiral arms. It is also evident that these two mild shock tracers are clearly detected in the CND, although the modeling of physical and chemical properties of the CND is not straightforward ([@Viti2014 Viti 2014]; [@Nakajima2015 Nakajima 2015]). A new SiO multi-transition study obtaining four transitions of SiO with ALMA in the central region of NGC 1068 is in progress (Taniguchi in prep.), and a preliminary $\sim0''.5$ resolution ALMA image shows that SiO(6-5) is concentrated at the western knot of the CND and no significant emission can be seen in any other regions including the eastern knot, molecular ridges along the inner bar seen in $^{12}$CO (e.g., [@Schinnerer2000 Schinnerer 2000]; [@Garcia-Burillo2014 García-Burillo 2014]) and spiral arms. This would suggest that the shock properties are drastically different between the Western knot and other regions. Further quantitative analysis of SiO fractional abundance distribution employing rotation diagrams of SiO, CS, and $^{13}$CO are on-going (Taniguchi in prep.) to understand the origin of such sharp and anisotropic variation of SiO properties. ![Distribution of CH$_3$OH (left) and HNCO (middle), tracers of mild shock, in the central $\sim$1 kpc region of NGC 1068. The CH$_3$OH/$^{13}$CO(1-0) intensity ratios, indicated as the sizes of circles superposed on the $^{13}$CO(1-0) integrated intensity image, are displayed in the right panel. Figures adopted from Tosaki (in prep.). []{data-label="fig4"}](Kohno-fig4.eps){width="70.00000%"} Unveiling the nature of obscured galaxies ========================================= With the enormous improvement of the sensitivity and spatial resolution in mm/submm interferometry accomplished by ALMA, it is now strongly expected that ALMA will provide unique diagnostic probes of energy sources in heavily obscured galaxies by penetrating deep dust condensations via mm/submm imaging and/or spectroscopy. One of the promising methods is to perform high spatial resolution measurements of the dust continuum to constrain the IR luminosity density and luminosity-to-mass (L/M) ratio in the heart of obscured galaxies. Theoretically, an extreme starburst, which will expel the ISM via radiation pressure onto dust and quench star-formation there, is estimated to have a maximum luminosity surface density of $\sim10^{13}$ $L_\odot$ kpc$^{-2}$ and L/M ratio of $\sim$1,000 $L_\odot/M_\odot$ ([@Thompson2005 Thompson 2005]; [@Sakamoto2008 Sakamoto 2008]), so spatially resolved measurements of these values would be crucial to test whether such galaxies are powered by a non-stellar source. In fact, PdBI and SMA observations of double nuclei in Arp 220 suggest that the western nucleus of Arp 220 can be heated by an AGN rather than a starburst ([@Downes2007 Downes & Eckart 2007]; [@Sakamoto2008 Sakamoto 2008]). Recent ALMA band-9 observations of the 430$\mu$m dust continuum at a resolution of $0''.36 \times 0''.20$ suggest that the IR luminosity density and the L/M ratio reach $2\times10^{14}$ $L_\odot$ kpc$^{-2}$ and 540 $L_\odot/M_\odot$, respectively, in the western nucleus of Arp 220, indicating the presence of a Compton-thick, heavily obscured AGN at least in the western nucleus of Arp 220 ([@Wilson2014 Wilson 2014]). Another promising, physically clean indicator would be (the ratio of) recombination lines of H, He, and He$^+$ because they have significantly different ionization potentials (13.6 to 54.4 eV). Specifically, [@Scoville2013 Scoville & Murchikova (2013)] demonstrate that the H26$\alpha$ (353.623 GHz) to HeII42$\alpha$ (342.894 GHz) luminosity ratio will show a factor of 50 difference between an AGN and a starburst. These two recombination lines can be simultaneously observed with ALMA band-7 with a reasonable integration time for local IR-luminous galaxies, and they will provide an ambiguous basis to calibrate other proposed diagnostic methods. A significant enhancement of HCN emission with respect to other molecules such as CO and HCO$^+$ at the centers of local AGNs was first indicated by [@Jackson1993 Jackson (1993)] and [@Tacconi1994 Tacconi (1994)] based on the NMA and PdBI observations of HCN(1-0) in NGC 1068. Since then, mounting supporting evidence for such elevated HCN emission at the centers of local Seyfert galaxies has been accumulated (e.g., [@Kohno2001 Kohno 2001]; [@Imanishi2006 Imanishi 2006]; [@Krips2008 Krips 2008]), giving an empirical AGN diagnostic. There are on-going debates on the origin of such elevated HCN, however. X-ray induced chemistry is one of the attractive scenarios because the CND of NGC 1068 is suggested to be a giant X-ray dominated region (XDR; [@Maloney1996 Maloney 1996]) based on detections of HOC$^+$ and other molecules that are preferentially enhanced in the XDR ([@Usero2004 Usero 2004]), although chemical models often do not reproduce such HCN enhancement in the XDR (e.g., [@Meijerink2007 Meijerink 2007]). More recently, spatially resolved ALMA observations of HCN(4-3) and HCO$^+$(4-3) at the CND of NGC 1068 at a $\sim$50 pc resolution reveal that the HCN(4-3)/HCO$^+$(4-3) line ratios are indeed highly elevated at the CND but not significantly at the very position of the active nucleus ([@Garcia-Burillo2014 García-Burillo 2014]; [@Viti2014 Viti 2014]). This indicates that HCN enhancement is not simply caused by X-ray irradiation. In fact, recent HCN(4-3)/HCO$^+$(4-3) measurements of the high-luminosity type-1 AGN NGC 7469 ([@Izumi2015a Izumi 2015a]) with $L_{\rm X} = 2 \times 10^{43}$ erg s$^{-1}$ suggest that the degree of HCN enhancement is not very different from that in NGC 1097 ([@Izumi2013 Izumi 2013]), a much lower luminosity type-1 AGN ($L_{\rm X} =7\times10^{40}$ erg s$^{-1}$), also adding further evidence against the XDR origin of enhanced HCN in AGNs (see also [@Martin2015 Martín 2015]). Mechanical heating, which is caused by the dissipation of shock-driven energy ([@Loenen2008 Loenen 2008]) and high-temperature-induced chemistry ([@Harada2010 Harada 2010]; [@Harada2013 Harada 2013]; [@Izumi2013 Izumi 2013]) are other possible mechanisms that could explain such HCN enhancement. Elevated HCN/HCO$^+$ ratios in massive (and presumably shocked) molecular outflows driven by the powerful AGN in Mrk 231 would be consistent with this view ([@Aalto2012 Aalto 2012]). A possible HCN maser amplification is also proposed ([@Matsushita2015 Matsushita 2015]) in the case of the HCN enhancement in M51 ([@Kohno1996 Kohno 1996]). Figure \[fig5\] displays the latest compilation of measurements of HCN(4-3)/HCO$^+$(4-3) and HCN(4-3)/CS(7-6) line ratios currently available, presented by [@Izumi2015b Izumi (2015b)]. This plot demonstrates that the HCN enhancement is still visible even in the $\sim$50 pc resolution ALMA measurements by [@Garcia-Burillo2014 García-Burillo (2014)] if we compare those values with star-forming and starburst galaxies. Although the origin of HCN enhancement is yet controversial (see also [@Viti; @2015 Viti 2015, this volume]), more (spatially resolved) observational work with ALMA will improve our understanding of such physical and chemical processes and result in identification of useful and reliable tracers of power sources in dusty galaxies. ![HCN(4-3)/HCO$^+$(4-3) and HCN(4-3)/CS(7-6) line ratios of local AGNs, starburst galaxies, and IR-luminous galaxies with buried AGNs. All these lines can be simultaneously observed using ALMA Band-7. In NGC 1068, these line ratios significantly depend on the size of the observing beam; a measurement with APEX ($\sim$1 kpc scale beam; [@Zhang2014 Zhang 2014]) is shown as an open red circle, whereas spatially resolved ALMA measurements ($\sim$50 pc; [@Garcia-Burillo2014 García-Burillo 2014]) are indicated by red filled circles. It was found that the HCN enhancement at the nucleus of NGC 1068 is not very significant compared with those in the peaks of the CND at $\sim$50 pc resolution, but it is still distinguishable from purely star-forming galaxies, suggesting that these ratios can give an empirical diagnostic of the dominant power sources within the observing beam. Note that NGC 1365 and NGC 4945 (two red open circles), with ratios similar to starbursts, are measured with a single dish telescope (and therefore dominated by circumnuclear starburst). Figure adopted from Izumi (2015b). []{data-label="fig5"}](Kohno-fig5.eps){width="65.00000%"} Regarding the nature of HCN enhancement, it is also important to consider the role of intense mid-IR radiation because HCN can be pumped by 14$\mu$m radiation, indicating the indispensable synergy between mm/submm and mid-IR facilities such as JWST and SPICA. Vibrationally excited HCN has been detected in IR-luminous galaxies such as NGC 4418, Mrk 231, and Arp 220 (e.g., [@Sakamoto2010 Sakamoto 2010]; [@Costagliola2015 Costagliola 2015]; [@Aalto2015a Aalto 2015a,b]). Mid-IR pumping seems to play a significant role in the enhanced high-$J$ HCN emission in the $z=3.91$ quasar APM 08279+5255 (e.g., [@Riechers2010 Riechers 2010]). Figure \[fig6\] presents a collection of submm (taken with ALMA) and mid-IR spectra (AKARI and Spitzer) toward the local ultraluminous IR galaxy IRAS 20551-4250, exhibiting vibrationally excited HCN emission ([@Imanishi2013 Imanishi & Nakanishi 2013]). Because vibrationally excited HCN is preferentially found at heavily obscured nuclei with highly elevated mid-IR luminosity (e.g., [@Aalto2015b Aalto 2015b]), deep silicate absorption features at 9.7 $\mu$m and 18 $\mu$m (indicating $A_{\rm V}>$50 - 100 mag) are often seen in the mid-IR spectra (although their connection is not yet established). Furthermore, PAH band features (the brightest one being at 7.7 $\mu$m) are also useful to constrain the relative contribution of AGN to the total IR luminosity in such dusty nuclei (e.g., [@Menendez-Delmestre2009 Menéndez-Delmestre 2009] and references therein; see also [@Privon2015 Privon (2015)] for an example of joint analysis of mm-wave HCN(1-0) and mid-IR properties.). JWST will be a powerful tool for studying such mid-IR radiation for local to low-$z$ galaxies, whereas SPICA, equipped with SMI (for 12 - 36 $\mu$m)[^1] and SAFARI (for 35 - 210 $\mu$m), will be crucial for higher redshift galaxies to catch these redshifted major mid-IR features. ![Mid-IR and submm spectra of IRAS 20551-4250, a heavily obscured ULIRG at $z=0.043$. Vibrationally-excited HCN emission has been detected (right). Figures taken from [@Imanishi2010 Imanishi (2010)], [@Sani2012 Sani & Nardini (2012)], and [@Imanishi2013 Imanishi & Nakanishi (2013)]. []{data-label="fig6"}](Kohno-fig6.eps){width="\textwidth"} 2012, A&A, 537, A44 2015, A&A, 574, A85 2015, A&A, in press (arXiv:1504.06824) 2013, A&A, 549, A39 2015, PASJ, in press (arXiv:1510.05004) , 2015, in Conditions and Impact of Star Formation From Lab to Space (EAS Publications Series), in press. 2008, Ap&SS, 313, 321 2015, A&A, 582, 91 2011, MNRAS, 416, 2143 2007, A&A, 468, L57 2010, ApJ, 720, 666 2010, A&A, 519, A2 2014, A&A, 567, A125 2008, ApJ, 682, 283 2009, A&A, 494, 977 2010, ApJ, 721, 1570 2013, ApJ, 765, 108 2009, A&A, 500, 725 2006, AJ, 131, 2888 2010, ApJ, 721, 1233 2013, AJ, 146, 91 2013, PASJ, 65, 100 2015a, ApJ, 811, 39 2015b, ApJ, in press (arXiv:1512.03438) 1993, ApJ (Letters), 418, L13 2005, ApJ, 631, 976 2015, MNRAS, 448, 104 1996, ApJ (Letters), 461, L29 2001, in: J.H. Knapen, J.E. Beckman, I. Shlosman, & T.J. Mahoney (eds.), The Central Kiloparsec of Starbursts and AGN: The La Palma Connection (San Francisco: Astronomical Society of the Pacific), p.672 2014, A&A, 564, A64 2008, A&A, 488, L5 1998, A&A, 337, 246 1996, ApJ, 466, 561 2009, ApJ, 694, 610 2015, A&A, 573, 116 1992, A&A, 254, 315 2015, ApJ, 799, 26 2015, ApJ, 801, 63 2007, A&A, 461, 793 2009, ApJ, 699, 667 2015, PASJ 67, 8 2013, ApJ, 768, 19 2015, ApJ, 814, 39 2010, ApJ, 725, 1032 2008, ApJ (Letters), 684, 957 2010, ApJ (Letters), 725, L228 2012, Advances in Astronomy, vol. 2012, id. 783451 2000, ApJ, 533, 85 2013, ApJ, 779, 75 1994, ApJ (Letters), 426, L77 2014, PASJ, 66, 75 2005, ApJ, 630, 167 2004, A&A, 419, 897 2011, ApJ (Letters), 740, L3 2014, A&A, 570, 28 2002, ApJ (Letters), 571, L173 1994, A&A, 286, L9 1996, Nature, 379, 139 2014, ApJ (Letters), 789, L36 2014, ApJ (Letters), 784, L31 [^1]: SMI contains 3 spectrometer modules with different resolutions, including the High Resolution Spectrometer, HRS, allowing us to obtain $R$ = 25,000 spectra at 12 - 17 $\mu$m. Note that JWST-MIRI does not have such high spectral resolution spectroscopic capability. See [@Sibthorpe2015 Sibthorpe 2015] for details of SMI and the newly defined SPICA mission, a candidate to ESA’s 5th medium class mission.	{ "pile_set_name": "ArXiv" }
--- abstract: 'Motivated by recent results in random matrix theory we will study the distributions arising from products of complex Gaussian random matrices and truncations of Haar distributed unitary matrices. We introduce an appropriately general class of measures and characterize them by their moments essentially given by specific Jacobi polynomials with varying parameters. Solving this moment problem requires a study of the Riemann surfaces associated to a class of algebraic equations. The connection to random matrix theory is then established using methods from free probability.' author: - 'Wolfgang Gawronski[^1], Thorsten Neuschel, Dries Stivigny [^2]' title: Jacobi polynomial moments and products of random matrices --- [Keywords: Moment problem; Jacobi polynomials; Raney distributions; Random matrices; Distribution of eigenvalues; Free probability theory; Free multiplicative convolution]{} [Mathematics Subject Classification 2010: 30E05 ; 15B52 , 30F10 , 46L54]{} Introduction ============ Products of random matrices are subject to research for many years now. It dates back to the 1960’s when Furstenberg and Kesten [@Furstenberg_Kesten] studied products of random matrices letting the number of factors grow to infinity while keeping the dimension fixed. This work was very influential and had applications to, for example, Schödinger operator theory [@Bougerol_Lacroix]. A more recent development is the study of the distribution of the eigenvalues and (squared) singular values of products consisting of a fixed number of factors as the dimensions grow to infinity. Different approaches have been found, e.g. free probability theory, to obtain the so-called limiting global eigenvalue or (squared) singular value distribution [@Alexeev_Gotze; @Burda_Janik_Waclaw; @Burda_1; @Burda_extended; @Dupic; @Gotze_Tikhomirov; @ORourke_Soshnikov; @Penson_Z]. In particular, the product of independent standard complex Gaussian matrices (these matrices are called Ginibre random matrices) has attracted interest with applications in, e.g., so-called multiple-input and multiple-output (MIMO) communication networks [@Akemann_Kieburg_Wei] (see [@Tulino_Verdu] for a more general introduction). In this context, it is also of interest to consider products involving Ginibre matrices and other random matrices. In [@Forrester] this was done for the product of Ginibre matrices and inverses of Ginibre matrices and in [@Kuijlaars_Stivigny] for the product of Ginibre matrices and truncations of unitary matrices (for applications, see, e.g., the introduction of [@Akemann_Nagao]). Let $r,s \in \mathbb{N} = \{0, 1, 2, \ldots\}$ with $s < r$ and let $T_1, \ldots, T_s$ be $s$ independent truncations of Haar distributed unitary matrices (such a matrix $T_j$ can be considered to be the upper left block of a Haar distributed unitary matrix). Moreover, let $G_{s+1}, \ldots, G_r$ be $r-s$ independent Ginibre random matrices. The motivation of this paper is to characterize the limiting distributions of the squared singular values of the product $$Y_{r,s} := G_r \ldots G_{s+1} T_s \ldots T_1.$$ This is equivalent with studying the limiting eigenvalue distribution of the Wishart-type matrix $Y_{r,s}^{\ast}Y_{r,s}$. The case $s = 0$, where we only have Ginibre random matrices, has been studied in [@Penson_Z; @Neuschel]. The limiting distribution was shown to be characterized by its moments $$FC_r(n) := \frac{1}{rn+1} {rn+n \choose n}, \qquad n \in \mathbb{N},$$ for fixed $r$. These numbers are called Fuss-Catalan numbers of order $r$ and historically arose in the context of combinatorial problems [@Knuth]. We will denote the corresponding distributions by $FC_r$. In case $s=1$, it turns out that the limiting distribution of the squared singular values of $Y_{r,1}$ coincides with a specific Raney distribution (Theorem \[thm: connection\_rmt\], Remark \[remark: lim\_distr\_case\_s\_1\] and also [@MNPZ; @Neuschel_Stivigny]). These distributions are a generalization of the Fuss-Catalan distributions and are defined by their moments, the so-called Raney numbers $$R_{\alpha, \beta}(n) := \frac{\beta}{n\alpha + \beta} {n\alpha + \beta \choose n}, \qquad n \in \mathbb{N},$$ for given $\alpha, \beta \in \mathbb{N}$ such that $\alpha > 1$ and $0 \leq \beta \leq \alpha$. These numbers have a combinatorial interpretation as well, see [@Forrester_Liu] for an overview. We will denote the corresponding distributions by $R_{\alpha, \beta}$. One can easily see that $R_{r+1, 1}(n) = FC_r(n)$ and so in the cases $s=0$ and $s=1$ the above mentioned limiting distributions of squared singular values are contained in the class of Raney distributions. However, for $s > 1$ the limiting distribution turns out not to belong to this class anymore. The main goal of this paper is to introduce and characterize an appropriately general class of measures that contains all these limiting distributions. In the language of free probability theory, this means we want to characterize a class of measures containing all multiplicative free convolutions (see Section 3.1) of the form $$FC_{r-s} \boxtimes R_{1, \frac{1}{2}}^{\boxtimes s}.$$ With this in mind, in Section \[sec: jac\_moments\] we introduce a sequence, depending on $a > 0$ and $r,s \in \mathbb{N}$ such that $s < r$, of positive numbers $$\label{eq: def_mu_0} J_{r,s,a}(0) := a$$ and for $n \in \mathbb{N}$ and $n \geq 1$ $$\label{eq: def_mu_n} J_{r,s,a}(n) := \frac{a}{n} \left(\frac{a^r}{(1+a)^s}\right)^n P_{n-1}^{(\alpha_{n-1}, \beta_{n-1})}\left(\frac{1-a}{1+a}\right),$$ where $P_n^{(\alpha_n, \beta_n)}(x)$ are the Jacobi polynomials with varying parameters $\alpha_n = rn + r + 1$ and $\beta_n = -(r+1-s)n - (r+2-s)$ as defined in [@Szego_Orth_Pol]. We will prove in Section 2 that these numbers indeed form a (Hausdorff) moment sequence of a compactly supported measure $J_{r,s,a}$. More precisely, we prove the following theorem. Let $r,s \in \mathbb{N}$ such that $s < r$ and let $a$ be a positive real number. Then there exists a unique measure $J_{r,s,a}$ on $[0, x^{\ast}]$ with total mass $a$ such that the moments are given by the numbers and . The right endpoint $x^{\ast}$ of the support of $J_{r,s,a}$ is defined below in . The proof of this result heavily relies on a study of the Riemann surface associated to the algebraic equation $$w^{r+1} - x(w-a)(w+1)^s = 0$$ which is done in Proposition \[prop: study\_alg\_eq\]. In Section \[sec: appl\_rmt\] we establish the connection with random matrix theory and we prove in Theorem \[thm: connection\_rmt\] that for $s < r$ we have $$J_{r,s,1} = FC_{r-s} \boxtimes R_{1, \frac{1}{2}}^{\boxtimes s}.$$ In particular, we can identify $J_{r,s,1}$ in the case $s = 0$ with the Fuss-Catalan distribution $FC_r$ and in the case $s = 1$ with the Raney distribution $R_{\frac{r+1}{2}, \frac{1}{2}}$. Finally, we want to emphasize the remarkable fact that the combination of Theorem \[thm: measure\_mu\] and Theorem \[thm: connection\_rmt\] establishes a further connection between random matrix theory and the theory of classical orthogonal polynomials. Jacobi polynomial moments {#sec: jac_moments} ========================= We start by showing that $J_{r,s,a}(n)$ is a positive-valued sequence. \[prop: positivity\_mu\_n\] Let $a > 0$ be a positive real number and $r,s$ positive integers such that $s \leq r$. Then $J_{r,s,a}(n) > 0$ for all $n \in \mathbb{N}$ where $J_{r,s,a}(n)$ is given by . To prove this, we need the following lemma. \[lemma: mu\_n\_id\_derivative\] Let $a > 0$ be a positive real number and $r,s$ positive integers such that $s \leq r$. Then $$\label{eq: mu_n_id_derivative} J_{r,s,a}(n) = \frac{1}{n!}\frac{d^{n-1}}{dz^{n-1}} \left.\left(\frac{z^{n(r+1)}}{(1+z)^{ns}}\right)\right\|_{z = a}$$ for all $n \in \mathbb{N}$ and $n \geq 1$. This follows from Leibniz’ rule. Indeed, we know that $$\begin{aligned} \frac{d^{n-1}}{dz^{n-1}} \left(\frac{z^{n(r+1)}}{(1+z)^{ns}}\right) &= \sum_{k = 0}^{n-1} {n-1 \choose k} \frac{d^{n-1-k}}{dz^{n-1-k}}\left(z^{n(r+1)}\right) \frac{d^k}{dz^k}\left((1+z)^{-ns}\right) \\ &= \sum_{k = 0}^{n-1} {n-1 \choose k} (nr + k + 2)_{n-1-k} z^{nr +k + 1} (-ns - k +1)_k (1+z)^{-ns-k} \\ &= \frac{z^{nr + 1}}{(1+z)^{ns}} P_{n-1}\left(\frac{z}{1+z}\right)\end{aligned}$$ where $(a)_k$ denotes the Pochhammer symbol and $$P_n(x) := \sum_{k = 0}^n {n \choose k} ((n+1)r + k + 2)_{n-k} (-(n+1)s - k +1)_k x^k.$$ Using the representation (see, e.g., [@Szego_Orth_Pol], p.62) $$P_n^{(\alpha_n, \beta_n)}(z)=\frac{1}{n!} \sum_{k=0}^n\binom{n}{k} (n+\alpha+\beta+1)_k (\alpha+k+1)_{n-k} \left(\frac{z-1}{2}\right)^k,$$ it is now straightforward to check that $$P_n(x) = n! P_n^{(\alpha_n, \beta_n)}(1-2x)$$ with $\alpha_n = rn + r + 1$ and $\beta_n = -(r+1 - s)n - (r+2 - s)$. This is clearly true for $n = 0$, so let $n \in \mathbb{N}$ and $n \geq 1$. We claim that $$\frac{d^{k}}{dz^{k}} \left.\left(\frac{z^{n(r+1)}}{(1+z)^{ns}}\right)\right\|_{z = x} > 0, \quad k = 0, \ldots, n, \quad x > 0$$ from which the statement then immediately follows by using Lemma \[lemma: mu\_n\_id\_derivative\]. A simple argument using Leibniz’ rule shows that it suffices to prove this claim for $r = s$. We start with the identity $$\frac{1}{(1+z)^{nr}} = \frac{1}{\Gamma(rn)} \int_{0}^{\infty} e^{-(1+z)t} t^{rn - 1} dt, \quad z>0.$$ Hence, after the substitution $y = tz$, we obtain $$\begin{aligned} \frac{d^{k}}{dz^{k}} \left(\frac{z^{n(r+1)}}{(1+z)^{nr}}\right) &= \frac{d^k}{dz^k} \left(\frac{z^n}{\Gamma(rn)} \int_0^{\infty} e^{-y(1 + \frac{1}{z})} y^{rn-1} dy\right) \\ &= \frac{1}{\Gamma(rn)} \int_0^{\infty} \frac{d^{k}}{dz^{k}} \left(z^n e^{-\frac{y}{z}}\right) e^{-y} y^{rn-1} dy \\ &= \frac{1}{\Gamma(rn)} \int_0^{\infty} \frac{d^{k}}{du^{k}} \left.\left(u^n e^{-\frac{1}{u}}\right)\right\|_{u = \frac{z}{y}} y^{n-k} e^{-y} y^{rn-1} dy.\end{aligned}$$ Using [@Szego_Orth_Pol Ex. 73 p. 388] and the sum representation for Laguerre polynomials (see e.g. [@Szego_Orth_Pol Formula 5.1.6]) $$L_k^{\alpha}(x) = \sum_{j = 0}^k (-1)^j {k + \alpha \choose k-j} \frac{x^j}{j!},$$ this can be rewritten as $$\frac{k!}{\Gamma(rn)} \int_0^{\infty} e^{-y(1 + \frac{1}{z})} y^{rn-1} z^{n-k} \left(\sum_{j = 0}^k {k-n-1 \choose k-j} \frac{(-1)^{k-j}}{j!} \left(\frac{y}{z}\right)^j\right) dy.$$ Since the sign of ${k-n-1 \choose k-j}$ is given by $(-1)^{k-j}$ the claim follows. We are now ready to state our main result of this section. First, we define $$\label{eq: w_ast} w^{\ast} := \frac{a(r + 1 - s) - r + \sqrt{(a(r + 1 - s) - r)^2 + 4a(r+1)(r-s)}}{2(r-s)}$$ and $$\label{eq: x_ast} x^{\ast} := \frac{r+1}{s+1} \frac{(w^{\ast})^r}{(w^{\ast} + 1)^{s-1}\left(w^{\ast} - \frac{as - 1}{s+1}\right)}.$$ These quantities are derived in Proposition \[prop: study\_alg\_eq\]. \[thm: measure\_mu\] Let $r,s \in \mathbb{N}$ such that $s < r$ and let $a$ be a positive real number. Then there exists a unique measure $J_{r,s,a}$ on $[0, x^{\ast}]$ with total mass $a$ such that the moments are given by the numbers and . First, notice that $x^{\ast} > 0$ for all $a > 0$ and $s < r$. Indeed, one can easily check that $w^{\ast} > a$ and thus, since $\frac{as-1}{s+1} < a$, we have that $x^{\ast} > 0$. Consider now the algebraic equation $$\label{eq: algebraic_eq_w} w^{r+1} - x(w - a)(w+1)^s = 0.$$ As we will show in Proposition \[prop: study\_alg\_eq\] this algebraic equation has a unique solution $w(x)$ which is analytic at infinity such that $w(x) \to a$ as $x \to \infty$. Furthermore, this solution has an analytic continuation to $\mathbb{C}\setminus[0, x^{\ast}]$. Since $w(x)$ is a solution of we know that $$w(x) = a + \frac{1}{x} \frac{w(x)^{r+1}}{(w(x)+1)^s}$$ and now applying the Lagrange-Bürmann theorem gives us that $$w(x) = a + \sum_{n = 1}^{\infty} \frac{1}{n!} \frac{d^{n-1}}{dz^{n-1}} \left.\left(\frac{z^{n(r+1)}}{(1+z)^{ns}}\right)\right\|_{z = a} x^{-n}$$ in a neighbourhood of infinity. Because of Lemma \[lemma: mu\_n\_id\_derivative\] this can be rewritten as $$\label{eq: w_moments_series} w(x) = \sum_{n = 0}^{\infty} J_{r,s,a}(n) x^{-n}.$$ We define now $$\label{eq: def_density} \rho(x) := \frac{1}{2\pi i} \left\lbrace\frac{w_{-}(x)}{x} - \frac{w_+(x)}{x}\right\rbrace, \qquad x \in (0, x^{\ast})$$ where $w_-(x)$, resp. $w_+(x)$, denotes the limiting value of $w(z)$ as $z$ approaches $x$ with $\text{Im}(z) < 0$, resp. $\text{Im}(z) > 0$. First of all, we notice that $w_-(x) = \overline{w_+(x)}$ so that $\rho(x)$ is a real-valued function. Furthermore, we claim that $\rho(x)$ is an integrable, everywhere positive function such that $$\int_0^{x^{\ast}} x^n \rho(x) dx = J_{r,s,a}(n) \qquad \text{for all } n \in \mathbb{N}$$ and thus is the density of a (unique) measure on $[0, x^{\ast}]$ with the numbers $J_{r,s,a}(n)$ as moments. By standard arguments and using $w(0) = 0$ one can see that $$\int_0^{x^{\ast}} x^n \rho(x) dx = \oint_{K} z^n \frac{w(z)}{z} dz = \oint_{K} z^{n-1} w(z) dz$$ with $K$ a positively oriented, closed contour encircling the cut $[0, x^{\ast}]$. Since $z^{n-1} w(z)$ has no singularities in $\mathbb{C}\setminus [0, x^{\ast}]$, we can can compute the residue at infinity and use to obtain that $$\oint_{K} z^{n-1} w(z) dz = J_{r,s,a}(n)$$ and hence $$\int_0^{x^{\ast}} x^n \rho(x) dx = J_{r,s,a}(n).$$ Finally, to prove that $\rho(x) > 0$ for all $x \in (0, x^{\ast})$, it now suffices to show that $w_-(x) \neq w_+(x)$ if $x \in (0, x^{\ast})$. Indeed, if $w_-(x) \neq w_+(x)$, then $\rho(x) \neq 0$ and thus, because of the continuity, $\rho(x)$ is either everywhere positive or everywhere negative. Since $$\int_0^{x^{\ast}} \rho(x) dx = J_{r,s,a}(0) = a$$ we obtain that $\rho(x)$ must be positive-valued. So let $x \in (0, x^{\ast})$ and let $K_x$ be a positively oriented circle around the origin with radius $x$ starting at the point $x$. Then we have $$1 = \frac{1}{2\pi i} \oint_{K_x} \frac{1}{z} dz = \frac{1}{2\pi i} \oint_{K_x} \frac{1}{f(w(z))} dz$$ where we used the fact that $$z = \frac{w(z)^{r+1}}{(w(z) - a)(w(z)+1)^s} =: f(w(z)).$$ Making $u = w(z)$ the new variable of integration and observing that $1 = f'(w(z)) w'(z)$ and $$f'(w) = \left(\frac{r+1}{w} - \frac{1}{w-a} - \frac{s}{w+1}\right)f(w)$$ we get that $$1 = \frac{1}{2\pi i} \oint_{w(K_x)} \left(\frac{r+1}{u} - \frac{1}{u-a} - \frac{s}{u+1}\right) du.$$ Here $w(K_x)$ is a contour starting at $w_+(x)$ and ending at $w_-(x)$. Assume now that $w_-(x) = w_+(x)$, i.e. $w_-(x) = w_+(x)$ is a real number. Because cannot have nonnegative solutions for $w$ if $x \in (0, x^{\ast})$ we see that $w(K_x)$ will be a closed contour starting in a point on the negative axis, going to the complex plane, crossing the real axis exactly one more time between the origin and $a$ and returning to its starting point. The orientation can be positive or negative, and the contour can encircle $-1$ but not $a$. Thus we get the following possibilities $$\frac{1}{2\pi i} \oint_{w(K_x)} \left(\frac{r+1}{u} - \frac{1}{u-a} - \frac{s}{u+1}\right) du = \begin{cases}r+1 \\ r+1-s \\ -(r+1) \\ -(r+1-s)\end{cases} \neq 1.$$ We see that in all cases we get a contradiction and thus $w_-(x) \neq w_+(x)$. Hence we can conclude that indeed defines a density. We end this section with the following proposition, which was needed in our proof of Theorem \[thm: measure\_mu\]. \[prop: study\_alg\_eq\] Let $r, s$ be positive integers such that $s < r$ and let $a > 0$. Then the equation $$\label{eq: algebraic_eq_w_2} w^{r+1} - x(w - a)(w+1)^s = 0$$ defines an algebraic function $w(x)$ which has an analytic branch at infinity with $w(x) \to a$, as $x \to \infty$. Moreover, this branch admits an analytic continuation to $\mathbb{C} \setminus [0, x^{\ast}]$ where $x^{\ast}$ is given by . The existence of a solution which is analytic at infinity can be seen by rewriting as $$w = a + \frac{1}{x} \frac{w^{r+1}}{(w+1)^s},$$ which permits one to apply Lagrange-Bürmann’s theorem. We confine this solution to the first sheet of the Riemann surface associated to and we denote it by $w_1(x)$. It remains to prove that this solution has an analytic continuation to $\mathbb{C} \setminus [0, x^{\ast}]$. To this end, we want to find all the branch points and thus we have to solve the following system of equations in the variables $x$ and $w$ $$\label{eq: solve_system} \begin{cases}w^{r+1} - x(w-a)(w+1)^s = 0 \\(r+1)w^r - x(w+1)^{s-1}((s+1)w - (as-1)) = 0\end{cases}.$$ One can now immediately see that $x = 0$ is a branch point with $w = 0$. This is a multiple branch point connecting all the $r+1$ sheets of the associated Riemann surface. Moreover, in the case $a = \frac{1}{s}$, the second equation of gives us that $w = 0$ and thus $x = 0$. So from now on we assume that $a \neq \frac{1}{s}$. Then the second equation can be rewritten as $$x = \frac{r+1}{s+1} \frac{w^r}{(w+1)^{s-1} \left(w - \frac{as-1}{s+1}\right)}.$$ Substituting this in the first equation gives us $$w^{r+1} - \frac{r+1}{s+1} \frac{w^r}{(w+1)^{s-1} \left(w - \frac{as-1}{s+1}\right)} (w-a)(w+1)^s = 0.$$ Assuming that $w \neq 0$, this can be simplified to $$w((s+1)w - (as-1)) - (r+1)(w-a)(w+1) = 0.$$ One can now easily check that the two solutions of this quadratic equation are given by $w = w^{\ast}$ where $w^{\star}$ is defined in and by $w = \tilde{w}$ with $$\label{eq: w_tilde} \tilde{w} := \frac{a(r + 1 - s) - r - \sqrt{(a(r + 1 - s) - r)^2 + 4a(r+1)(r-s)}}{2(r-s)}.$$ Hence, we can conclude that the only possible branch points are at the real points $x = 0$, $x = x^{\ast}$, $x = \tilde{x}$ with $$\label{eq: x_tilde} \tilde{x} := \frac{r+1}{s+1} \frac{(\tilde{w})^r}{(\tilde{w} + 1)^{s-1}\left(\tilde{w} - \frac{as - 1}{s+1}\right)}$$ and at infinity. To conclude that on the first sheet $w(x)$ only has branch points at $x = 0$ and at $x = x^{\ast}$, it now suffices to show that there is no branch point at $x = \tilde{x}$ on the first sheet. Indeed, due to the analyticity of $w_1(x)$ there cannot be a branch point at infinity on this sheet. From equation it can be observed that $w_1(x)$ admits an analytic continuation starting at infinity travelling along the negative real axis up to the origin. In the same manner, $w_1(x)$ can be analytically continued starting at infinity travelling along the positive real axis up to $x = x^{\ast}$. Moreover, we have $w_1(x) > 0$ on $\mathbb{R}\setminus [0, x^{\ast}]$. As all branch points are real, we can conclude that $w_1(x)$ admits an analytic continuation onto $\mathbb{C}\setminus [0, x^{\ast}]$. Taking into account that $\tilde{x} \notin [0, x^{\ast}]$ and $w(\tilde{x}) = \tilde{w} < 0$, this shows that there can be no further branch point on the first sheet which completes the proof. Using the observations made in the proof of Proposition \[prop: study\_alg\_eq\] we can now describe the geometry of the Riemann surface associated to the algebraic equation . We illustrate this in Figure \[fig: riemann\_surface\] for the case $r = 5, s=3$. (0,0)–(1,1)–(4,1)–(3,0)–cycle; (8,0)–(9,1)–(12,1)–(11,0)–cycle; (2, 0.5)–(2.8,0.5); (8.5,0.5)–(10,0.5); (10,0.5)–(10,-3.5); (8.5,0.5)–(8.5,-3.5); (2,0.5)–(2,-1.5); (8.5,0.5)–(5.5,-1.5); (2.8,0.5)–(2.8,-1.5); (10,0.5)–(7,-1.5); (2, -1.5)–(2.8,-1.5); (5.5,-1.5)–(7,-1.5); (0,-2)–(1,-1)–(4,-1)–(3,-2)–cycle; (5,-2)–(6,-1)–(9,-1)–(8,-2)–cycle; (0.5,-1.5)–(2,-1.5); (7,-1.5)–(8,-1.5); (0.5,-1.5)–(0.5,-3.5); (7,-1.5)–(2,-3.5); (2,-1.5)–(2,-3.5); (8,-1.5)–(3,-3.5); (0.5,-3.5)–(2,-3.5); (2,-3.5)–(3,-3.5); (8.5,-3.5)–(5.5,-1.5); (10,-3.5)–(7,-1.5); (8.5,-3.5)–(10,-3.5); (0,-4)–(1,-3)–(4,-3)–(3,-4)–cycle; (8,-4)–(9,-3)–(12,-3)–(11,-4)–cycle; at (2.8,0.5) [$x^{\ast}$]{}; (2.8,0.5) circle (2pt); at (2,0.5) [$0$]{}; (2,0.5) circle (2pt); at (2.8,-1.5) [$x^{\ast}$]{}; (2.8,-1.5) circle (2pt); (2,-1.5) circle (2pt); (2,-3.5) circle (2pt); at (3,-3.5) [$\tilde{x}$]{}; (3,-3.5) circle (2pt); at (8,-1.5) [$\tilde{x}$]{}; (8,-1.5) circle (2pt); (7,-1.5) circle (2pt); (10,0.5) circle (2pt); (10,-3.5) circle (2pt); Application to random matrix theory {#sec: appl_rmt} =================================== In this section we will show how the measures obtained in Theorem \[thm: measure\_mu\] arise naturally in random matrix theory and free probability. We start with a small introduction in free probability theory which we need to state our second theorem. For more details, we refer the reader to [@Voiculescu_DN], [@Anderson_GZ], [@Speicher] or [@Akemann_Handbook]. Free probability and random matrices ------------------------------------ Given a (compactly supported) probability measure $\mu$ on $\mathbb{R}$ such that $\int_{\mathbb{R}}x d\mu(x) \neq 0$ we define its $S$-transform as follows. Let $G_{\mu}$ denote the Stieltjes transform of the measure $\mu$, i.e. $$\label{eq: def_stieltjes} G_{\mu}(z) := \int_{\mathbb{R}} \frac{1}{z-x} d\mu(x), \qquad z \in \mathbb{C}\setminus\text{supp}(\mu)$$ and define $$\label{eq: def_psi_mu} \psi_{\mu}(z) := \frac{1}{z} G_{\mu}\left(\frac{1}{z}\right) - 1.$$ Let $\chi_{\mu}$ be the unique function, analytic in a neighbourhood of zero, satisfying $$\label{eq: def_chi_mu} \chi_{\mu}(\psi_{\mu}(z)) = z.$$ Then the $S$-transform, denoted by $S_{\mu}$, is defined as $$\label{eq: def_s_tr} S_{\mu}(z) := \frac{z+1}{z} \chi_{\mu}(z).$$ Given two (compactly supported) probability measures $\mu$ and $\nu$ with non-vanishing first moments, the free multiplicative convolution, denoted by $\mu \boxtimes \nu$, is the unique (compactly supported) probability measure that satisfies the identity $$S_{\mu \boxtimes \nu}(z) = S_{\mu}(z) S_{\nu}(z).$$ Notice that this identity shows us that the free multiplicative convolution is commutative, i.e. $\mu \boxtimes \nu = \nu \boxtimes \mu$. The $S$-transform is an important tool in free probability theory to compute the distribution of, for instance, the product of free random variables. By $\mu_A$ we denote the empirical eigenvalue distribution of an $n \times n$ random matrix $A$, i.e. $$\mu_A = \frac{1}{n} \sum_{i = 1}^n \delta_{\lambda_i(A)}$$ with $\lambda_i(A)$ the $n$ random eigenvalues of $A$. With this in mind, we now have the following result ([@Couillet_Debbah], Theorem 4.7, p. 82, see also [@Voiculescu] and [@Akemann_Handbook]). \[thm: distr\_prod\_matrices\] Let $\{A_n\}$ and $\{B_n\}$ be two sequences of random matrices of size $n \times n$ such that $A_n > 0$, i.e. all eigenvalues are positive (with probability 1), and such that $A_n$ and $B_n$ are asymptotically free almost surely for all $n$. Moreover, suppose that there exist two compactly supported probability measures $\mu_1$ and $\mu_2$ such that $$\mu_{A_n} \stackrel{w}{\longrightarrow} \mu_1 \qquad a.s. \qquad \text{and} \qquad \mu_{B_n} \stackrel{w}{\longrightarrow} \mu_2 \qquad a.s.,$$ as $n \to \infty$, and where $\mu_{A_n}$ resp. $\mu_{B_n}$ denote the empirical eigenvalue distribution of $A_n$ resp. $B_n$. Then $$\mu_{A_nB_n} \stackrel{w}{\longrightarrow} \mu_1 \boxtimes \mu_2 \qquad a.s.,$$ as $n \to \infty$. Here, by $$\mu_{A_n} \stackrel{w}{\longrightarrow} \mu_1 \qquad a.s.,$$ as $n \to \infty$, we mean that $$\int_{\mathbb{R}} f(t) d\mu_1(t) = \lim_{n \to \infty} \frac{1}{n} \sum_{i = 1}^n f(\lambda_i(A_n))$$ holds with probability $1$ for each bounded, continuous function $f(t)$. We say that $\mu_{A_n}$ converges weakly, almost surely to $\mu_1$. Basically, Theorem \[thm: distr\_prod\_matrices\] holds for all sequences $\{A_n\}$ and $\{B_n\}$ for which $A_n$ and $B_n$ are independent and for which at least the distribution of $A_n$ or $B_n$ is invariant under left and right multiplication by Haar distributed unitary matrices [@Couillet_Debbah; @Akemann_Handbook; @Voiculescu]. Product of Ginibre and truncated unitary matrices ------------------------------------------------- Let $U$ be a Haar distributed unitary random matrix of size $l \times l$ and let $T$ be the $m \times n$ upper left block of $U$ such that $l \geq m+n$. We call $T$ a truncated unitary (random) matrix of size $m \times n$ and the distribution is proportional to (see, e.g., [@Fyodorov_Sommers Eq. (69)]) $$\det(I - T^{\ast}T)^{l - m-n} \chi_{T^{\ast}T \leq I}(T) dT$$ where $$\chi_{T^{\ast}T \leq I}(T) := \begin{cases}1 & \mbox{ if } I-T^{\ast}T \mbox{ is positive-definite} \\ 0 & \mbox{else}\end{cases}.$$ A complex Ginibre matrix $G$ of size $m \times n$ has independent entries whose real and imaginary parts are independent and have a standard normal distribution with fixed variance. The probability distribution of $G$ is proportional to $$e^{- \operatorname{Tr}G^{\ast} G} dG.$$ We now take $s$ independent truncated unitary matrices $T_j$ of size $(n + \nu_j) \times (n + \nu_{j-1})$, with $\nu_j \geq 0$ and $\nu_0 = 0$, coming from an $l_j \times l_j$ unitary matrix. Furthermore, we take $r-s$ independent Ginibre random matrices $G_j$ of size $(n + \nu_j) \times (n + \nu_{j-1})$ for $j = s+1, \ldots, r$ and we define the product of independent matrices $$\label{eq: def_prod_yrs} Y_{r,s} := G_r \ldots G_{s+1}T_s \ldots T_1.$$ We then have the following theorem: \[thm: connection\_rmt\] Let $T_j$ and $G_j$ be as described above. Furthermore, suppose that we have that $l_j - 2n\geq 0$ and $\nu_i$ remain fixed for all $j = 1, \ldots, s$ and all $i = 0, \ldots, r$, as $n \to \infty$. Then we have $$\mu_{Y_n} \stackrel{w}{\longrightarrow} J_{r,s,1} \qquad a.s.,$$ as $n \to \infty$, and where $J_{r,s,1}$ is described in Theorem \[thm: measure\_mu\] and $Y_n$ is defined as the rescaled Wishart-type product $$Y_n := \frac{1}{n^{r-s}}Y_{r,s}^{\ast}Y_{r,s}.$$ We start with the observation that $Y_n$ has the same non-zero eigenvalues (counted with multiplicity) as $$Z_n := \frac{1}{n^{r-s}} (G_r \ldots G_{s+1})^{\ast}(G_r \ldots G_{s+1})(T_s \ldots T_1)(T_s \ldots T_1)^{\ast}$$ Moreover, the difference in the number of eigenvalues equal to zero is $\nu_{s}$ and thus the limiting eigenvalue distributions of $Y_n$ and $Z_n$ have to be equal, if they exist. It is known that the empirical eigenvalue distribution of $\frac{1}{n^{r-s}}(G_r \ldots G_{s+1})^{\ast}(G_r \ldots G_{s+1})$ converges weakly, almost surely to the Fuss-Catalan distribution $FC_{r-s}$, as $n \to \infty$ (see, e.g., [@Penson_Z]), which can be written as the Raney distribution $R_{r-s+1,1}$. Moreover, the distribution of each $G_j$ is invariant under left and right multiplication of Haar distributed unitary matrices and thus the same holds true for $$\frac{1}{n^{r-s}}(G_r \ldots G_{s+1})^{\ast}(G_r \ldots G_{s+1}).$$ In order to apply Theorem \[thm: distr\_prod\_matrices\] we now have to determine the limiting eigenvalue distribution of $\tilde{T}_n := (T_s \ldots T_1)(T_s \ldots T_1)^{\ast}$. By the same arguments as before, we know that $(T_s \ldots T_1)(T_s \ldots T_1)^{\ast}$ has the same non-zero eigenvalues (counted with multiplicity) as $(T_{s-1} \ldots T_1)(T_{s-1} \ldots T_1)^{\ast}T_s^{\ast}T_s$ and the difference in the number of eigenvalues equal to zero is $\|\nu_s - \nu_{s-1}\|$. It is known that the empirical eigenvalue distribution of $T_s^{\ast}T_s$ converges weakly, almost surely to the arcsine measure on $(0, 1)$ if $l_s - 2n$ is fixed, as $n \to \infty$. By comparing the moments, one can see that this is the Raney distribution $R_{1, \frac{1}{2}}$. Moreover, one can check that the distribution of $T_s$ is also invariant under left and right multiplication of Haar distributed unitary matrices and so is the distribution of $T_s^{\ast}T_s$. As before, to apply Theorem \[thm: distr\_prod\_matrices\] we now have to determine the limiting eigenvalue distribution of $(T_{s-1} \ldots T_1)(T_{s-1} \ldots T_1)^{\ast}$. Repeating this argument $s-1$ times we can conclude that $$\mu_{\tilde{T}_n} \stackrel{w}{\longrightarrow} R_{1, \frac{1}{2}}^{\boxtimes s} \qquad a.s.,$$ as $n \to \infty$, and with $\tilde{T}_n := (T_s \ldots T_1)(T_s \ldots T_1)^{\ast}$. An application of Theorem \[thm: distr\_prod\_matrices\] gives us that $$\mu_{Z_n} \stackrel{w}{\longrightarrow} \kappa \qquad a.s.,$$ as $n \to \infty$, where $$\label{eq: def_kappa} \kappa := R_{r-s+1, 1} \boxtimes R_{1, \frac{1}{2}}^{\boxtimes s}.$$ It remains to show that $\kappa = J_{r,s,1}$. Using the result from Mlotkowski [@Mlotkowski Proposition 4.3], we know that $$S_{R_{r-s+1}, 1}(z) = \frac{1}{(1+z)^{r-s}}, \qquad S_{R_{1, \frac{1}{2}}}(z) = \frac{z+2}{z+1}$$ and hence $$S_{\kappa}(z) = \frac{(z+2)^s}{(z+1)^r}.$$ Using this can be rewritten as $$\chi_{\kappa}(z) = \frac{z(z+2)^s}{(z+1)^{r+1}}.$$ Thus, if we replace $z$ by $\psi_{\kappa}(z)$ and use we obtain $$z = \frac{\psi_{\kappa}(z) (\psi_{\kappa}(z) + 2)^s}{(\psi_{\kappa}(z) + 1)^{r+1}}.$$ Finally applying identity and replacing $z$ by $1/z$ we arrive at $$\frac{1}{z} = \frac{(zG_{\kappa}(z) - 1)(z G_{\kappa}(z) + 1)^s}{(z G_{\kappa}z)^{r+1}}$$ and from this we can conclude that $w(x) = x G_{\kappa}(x)$ satisfies the algebraic equation $$w(x)^{r+1} - x (w(x) - 1)(w(x) + 1)^s = 0.$$ This is equation with $a = 1$ and thus, because of Theorem \[thm: measure\_mu\], we obtain $$\label{eq: kappa_equal_mu1} \kappa = J_{r,s,1}.$$ \[remark: lim\_distr\_case\_s\_1\] Theorem \[thm: connection\_rmt\] in combination with equations and gives us that $$J_{r,0, 1} = FC_r, \qquad J_{r,1,1} = R_{\frac{r+1}{2}, \frac{1}{2}}.$$ This can be seen immediately by using $$R_{r, 1} \boxtimes R_{1, \frac{1}{2}} = R_{\frac{r+1}{2}, \frac{1}{2}},$$ which is a special case of the identity stated in [@Mlotkowski Proposition 4.3]. This means in particular that $$J_{r, 0, 1}(n) = FC_r(n), \qquad J_{r, 1, 1}(n) = R_{\frac{r+1}{2}, \frac{1}{2}}(n)$$ for every $n \in \mathbb{N}$. It is interesting to remark that for these distributions explicit and elementary forms of the densities can be found by the method of parametrization (see, e.g., [@Forrester_Liu; @MNPZ; @Neuschel; @Neuschel_Stivigny]). The statements of Theorem \[thm: measure\_mu\] and Theorem \[thm: connection\_rmt\] are restricted to the case $s < r$ because of several technical issues that arise in the case $r = s$. However, it is natural and interesting to ask whether our results can be extended to this case. Acknowledgments {#acknowledgments .unnumbered} =============== We thank Prof. Arno Kuijlaars for many valuable discussions. The last two authors are supported by KU Leuven Research Grant OT/12/073 and the Belgian Interuniversity Attraction Pole P07/18. [99]{} G. Akemann, Z. Burda, M. Kieburg and T. Nagao, Universal microscopic correlation functions for products of truncated unitary matrices, preprint arXiv: 1310.6395. G. Akemann, M. Kieburg, and L. Wei, Singular value correlation functions for products of [W]{}ishart random matrices, J. Phys. A 46 (2013), 275205, 22 pp. N. Alexeev, F. G[ö]{}tze and A. Tikhomirov, Asymptotic distribution of singular values of powers of random matrices, Lith. Math. J. 50 (2010), no. 2, 121–132. G.W. Anderson, A. Guionnet and O. Zeitouni, An introduction to random matrices, Cambridge Studies in Advanced Mathematics, Vol. 118, Cambridge University Press, Cambridge, 2010. P. Bougerol and J. Lacroix, Products of random matrices with applications to [S]{}chrödinger operators, Progress in Probability and Statistics, 8 (1985), xii+283. Z. Burda, R.A. Janik and B. Waclaw, Spectrum of the product of independent random [G]{}aussian matrices, Phys. Rev. E (3) 81 (2010), no. 4, 041132. Z. Burda, A. Jarosz, G. Livan, M.A. Nowak, and A. Swiech, Eigenvalues and singular values of products of rectangular [G]{}aussian random matrices, Phys. Rev. E (3) 82 (2010), no. 6, 061114. Z. Burda, A. Jarosz, G. Livan, M.A. Nowak, and A. Swiech, Eigenvalues and singular values of products of rectangular [G]{}aussian random matrices — the extended version, Acta Phys. Polon. B 42 (2011), no. 5, 939–985. R. Couillet and M. Debbah, Random matrix methods for wireless communications, Cambridge University Press, Cambridge, 2011. T. Dupic and I. Isaac Pérez Castillo, Spectral density of products of Wishart dilute random matrices. Part I: the dense case, preprint arXiv:1401.7802. P.J. Forrester, Eigenvalue statistics for product complex Wishart matrices, preprint arXiv: 1401.2572. P.J. Forrester and D.-Z. Liu, Raney distributions and random matrix theory, preprint arXiv: 1404:5759v1. H. Furstenberg and H. Kesten, Products of random matrices, Ann. Math. Statist. 31 (1960), 457–469. Y.V. Fyodorov and H.-J. Sommers, Random matrices close to [H]{}ermitian or unitary: overview of methods and results, J. Phys. A 36 (2003), no. 12, 3303–3347. F. Götze and A. Tikhomirov, On the asymptotic spectrum of products of independent random matrices, preprint arXiv:1012.2710. R.L. Graham, D.E. Knuth and O. Patashnik, Concrete mathematics, Addison-Wesley Publishing Company, Reading, MA, 2nd edition, 1994. A. Kuijlaars, D. Stivigny, Singular values of products of random matrices, Random Matrices: Theory and Applications (2014) DOI 10.1142/S2010326314500117. W. Mlotkowski, Fuss-Catalan numbers in noncommutative probability, Documenta Mathematica 15 (2010), 939–955. W. Mlotkowski, M.A. Nowak, K.A. Penson and K. Zycskowski, Spectral density of generalized Wishart matrices and free multiplicative convolution, preprint arXiv: 1407.1282. T. Neuschel, Plancherel-Rotach formulae for average characteristic polynomials of products of Ginibre random matrices and the Fuss-Catalan distribution, Random Matrices: Theory and Appl. 03 (2014), no. 01, 1450003. T. Neuschel and D. Stivigny Asymptotics for characteristic polynomials of Wishart type products of complex Gaussian and truncated unitary random matrices, preprint arXiv:1407.2755. S. O’Rourke and A. Soshnikov, Products of independent non-[H]{}ermitian random matrices, Electron. J. Probab. 16 (2011), no. 81, 2219–2245. K.A. Penson and K. [Ż]{}yczkowski, Product of [G]{}inibre matrices: [F]{}uss-[C]{}atalan and [R]{}aney distributions, Phys. Rev. E 83 (2011), 061118, 9 pp. R. Speicher, Free probability and random matrices, preprint arXiv: 1404.3393. R. Speicher, Free Probability Theory, Chapter 22 of The Oxford handbook of random matrix theory, Oxford Univ. Press, Oxford, 2011. G. Szeg[ő]{}, Orthogonal polynomials, American Mathematical Society, Colloquium Publications, Vol. XXIII, American Mathematical Society, Providence, R.I., 1975. A.M. Tulino and S. Verdú, Random Matrix Theory and Wireless Communications, Commun. Inf. Theory 1 (2004), no. 1, 1–182. D.V. Voiculescu, Limit laws for random matrices and free products, Invent. Math. 104 (1991), no. 1, 201–220. D.V. Voiculescu, K.J. Dykema and A. Nica, Free random variables, CRM Monograph Series, Vol. 1, American Mathematical Society, Providence, RI, 1992. [^1]: Department of Mathematics, University of Trier, 54286 Trier, Germany. E-mail: [email protected] [^2]: Department of Mathematics, KU Leuven, Celestijnenlaan 200B box 2400, BE-3001 Leuven, Belgium. E-mail: [email protected], [email protected]	{ "pile_set_name": "ArXiv" }
--- abstract: 'A known drawback of ‘decentralised’ contact tracing architectures is that users who have been in contact with an infected person are able to precisely identify the relevant contact, and thereby perhaps identify the infected person. In [@chan2020pact], the PACT team discuss a simple DH-based protocol to mitigate this problem, but dismiss it because it is vulnerable to a malicious user who may deviate from the specified behaviour. This note presents a modified protocol which achieves robustness against a fully malicious user, and establishes some simple security properties.' author: - \| David Mestel[^1]\ University of Luxembourg\ bibliography: - 'mybib.bib' title: Robust ambiguity for contact tracing --- Introduction ============ In the design of contact tracing apps, the choice between ‘centralised’ and ‘decentralised’ architectures has received great public attention. The latter, which has been adopted by the majority of countries, by the Google-Apple API and by the DP3T consortium [@troncoso2020decentralized], has many privacy advantages, but one disadvantage is that each individual user is able to determine which of the many tokens they have collected came from an infected user, and consequently (by recalling the precise time and strength of the contact), may be able to identify the infected individual among their contacts [@vaudenay2020centralized]. This may compromise the privacy of the infected person, and violates the principle of manual contact tracing that a person should be told only that they have been in contact with an infected person, and not the person’s identity [@cdc2020contact]. One possible solution to this problem, discussed in [@chan2020pact], is for infected users to send to the system the tokens they have collected rather than those they have distributed. These can then be rerandomised before being broadcast to all users, so that users are able to recognise a rerandomised token as being derived from one they have broadcast, but not specifically which one. The simple protocol described in [@chan2020pact] is as follows (working in a multiplicative group ${\mathcal{G}}$ of prime order with generator $g$): 1. Each user $u$ generates a keypair $(s_u,g^{s_u})$ with random $s_u$. 2. Each broadcast token is of the form $t_i=(g^{\alpha_i},g^{\alpha_is_u})$, where each $\alpha_i$ is chosen uniformly at random. 3. On testing positive, a user who has received the token $t=(x,y)$ uploads $(x^\beta,y^\beta)$ for fresh random $\beta$. 4. To determine whether they are at risk, user $u$ checks whether a token of the form $(x,y)$ with $y=x^{s_u}$ is present on the server. Conditional on the Decisional Diffie-Hellman (DDH) assumption on ${\mathcal{G}}$, the tokens generated by each user are pseudorandom, and furthermore a user who learns they are at risk cannot tell which of their tokens (generated using $s_u$) was reported, because of the exponentiation by fresh random $r$. However, as the authors note, this protocol is fatally flawed in the presence of malicious users. Such a user can simply use a fresh $s_u$ for every token they generate, and this cannot be detected or prevented (one could imagine requiring every token to be accompanied by a zero-knowledge proof that it was generated using one of the keys in a public list of public keys, but this would not be remotely practical). In this note we show how to modify the protocol so as to be robust against fully malicious users, at the cost of requiring the server to send a ‘personalised’ (but non-secret) set of rerandomised tokens to each user. The essential idea is to extend the rerandomisation step such that the messages corresponding to malformed tokens are flat random. Protocol description ==================== We describe the protocol in three phases: registration; broadcasting, where a user has contacts with others and transmits tokens; and infection, after the user has tested positive. Throughout, ${\mathcal{G}}$ is assumed to be a multiplicative abelian group of prime order $p$, with generator $g$. 1. Registration phase: sample $s\leftarrow {\mathcal{U}}([0,p-1])$ and send $g^s$ (non-anonymously) to the server, which adds it to the list of public keys. Since registration is not required to be anonymous, the server can ensure that each individual is only able to register a single key. 2. Broadcast phase: sample $\alpha \leftarrow {\mathcal{U}}([1,p-1])$, and broadcast the token $${\textit{Tok}}(s,\alpha) {\vcentcolon=}(g^\alpha,g^{s\alpha}),$$ replacing $\alpha$ with a fresh random value after a suitable period. 3. Infection phase: a user who tests positive sends the server the list of tokens it has received. The server verifies that each token $t=(x,y)$ has $x\neq g$ (discarding those that fail). At the end of each day, for each user $u$, say with public key $g^{s_u}$, for each token $t=(x,y)$ in its list the server samples $\beta,\gamma \leftarrow {\mathcal{U}}([0,p-1])$ and sends to $u$ $${\textit{Shuff}}((x,y),\beta,\gamma,g^{s_u}) {\vcentcolon=}(x^\beta g^\gamma,y^\beta (g^{s_u})^\gamma).$$ On receiving $t'=(z,w)$, user $u$ checks whether $w=z^{s_u}$, and if so knows that they are infected. Note that ambiguity (but not other privacy properties) is dependent on the honesty of the server; similarly, in manual contact tracing ambiguity is dependent on the discretion of the tracer. To establish correctness, observe that if $t'$ came from a token broadcast by $u$, then we have $$\begin{aligned} t'&={\textit{Shuff}}\left({\textit{Tok}}(s,\alpha),\beta,\gamma,g^{s_u}\right)\\ &=\left((g^\alpha)^\beta g^\gamma,(g^{s_u\alpha})^\beta (g^{s_u})^\gamma\right)\\ &=\left(g^{\alpha\beta+\gamma},g^{s_u(\alpha\beta+\gamma)}\right),\end{aligned}$$ as required (for some $\alpha,\beta,\gamma$). The communication cost of this protocol is equivalent to that of just sending each user a list of all the tokens from infected users, with $O(1)$ computational cost per message passed from server to user (a single exponentiation by the user, and four exponentiations by the server). Security properties =================== In this section we establish three key security properties of the protocol. First, conditional on the DDH assumption on ${\mathcal{G}}$, the tokens broadcast by a user with randomly chosen key are computationally indistinguishable from independent random group elements, even with knowledge of the public key (Theorem \[thm:indist\]), and so no privacy is lost by uninfected users. Second, for each user $u$ the output of ${\textit{Shuff}}$ (as a probability distribution on ${\mathcal{G}}\times{\mathcal{G}}$ with random $\beta,\gamma$) is equal on all tokens honestly generated by $u$ (Theorem \[thm:honest\]), and so an honest-but-curious $u$ will be unable to determine which of their broadcast tokens corresponded to contact with an infected person. Third, for any $u$ the output of ${\textit{Shuff}}$ on any input other than a token honestly generated by $u$ is uniformly random (Theorem \[thm:dishonest\]), and so a malicious user is not able to defeat ambiguity by broadcasting malformed tokens. \[thm:indist\] Let $k$ be a positive integer, and $(S, A_1,\ldots,A_k)\sim {\mathcal{U}}([0,p-1]\times[0,p-1]^k)$. If ${\mathcal{G}}$ satisfies the DDH assumption then $$(g^S,{\textit{Tok}}(S,A_1),\ldots,{\textit{Tok}}(S,A_k)) {\stackrel{c}{\equiv}}{\mathcal{U}}({\mathcal{G}}\times ({\mathcal{G}}\times {\mathcal{G}})^k).$$ Let ${\mathcal{A}}$ be a PPT algorithm distinguishing the two distributions, and let $(g_1,g_2,g_3)$ be a DDH challenge (so either $(g_1,g_2,g_3) = (g^x,g^y,g^{xy})$ or $(g^x,g^y,g^z)$ for random $x,y,z$). Run ${\mathcal{A}}$ on $(g_1,(g_2^{s_1},g_3^{s_1}),(g_2^{s_2},g_3^{s_2}),\ldots,(g_2^{s_k},g_3^{s_k}))$ for random $s_1,s_2,\ldots,s_k\in [0,p-1]$. \[thm:honest\] Let $(X,Y)\sim {\mathcal{U}}([0,p-1]^2)$ and $Z\sim {\mathcal{U}}({\mathcal{G}})$. Then for all $s\in [0,p-1], \alpha\in [1,p-1],$ we have $${\textit{Shuff}}\left({\textit{Tok}}(s,\alpha),X,Y,g^s\right) \equiv (Z,Z^s).$$ We have ${\textit{Shuff}}({\textit{Tok}}(s,\alpha),\beta,\gamma,g^s) = (g^{\alpha\beta+\gamma},(g^{\alpha\beta + \gamma})^s)$. Since $\gamma$ is uniformly distributed, so is $\alpha\beta+\gamma$ and hence so is $g^{\alpha\beta+\gamma}$, as required. \[thm:dishonest\] Let $(X,Y)\sim {\mathcal{U}}([0,p-1]^2)$. Then for all $s\in [0,p-1], t\in ({\mathcal{G}}\setminus \{e\})\times {\mathcal{G}},$ either $t = {\textit{Tok}}(s,\alpha)$ for some $\alpha\in [1,p-1]$ or we have $${\textit{Shuff}}\left(t,X,Y,g^s\right) \equiv {\mathcal{U}}({\mathcal{G}}\times{\mathcal{G}}).$$ Without loss of generality $t=(g^\alpha,g^{s'\alpha})$ for some $s'$ and $\alpha\neq 0$. If $t\neq {\textit{Tok}}(s,\alpha)$ then $s'\neq s$. Then $$\begin{aligned} {\textit{Shuff}}(t,\beta,\gamma,g^s) &= \left((g^\alpha)^\beta g^\gamma,(g^{s'\alpha})^\beta (g^s)^\gamma\right)\\ &=\left(g^{\alpha\beta + \gamma},g^{s'\alpha\beta + s\gamma}\right)\\ &=\left(g^{\alpha\beta + \gamma},g^{s'(\alpha\beta + \gamma) + (s-s')\gamma}\right).\end{aligned}$$ Since $\alpha \neq 0$ and $\beta,\gamma$ are independently uniformly distributed ([IUD]{}), we have that $\alpha\beta$ and $\gamma$ are [IUD]{}, and hence so are $\alpha\beta + \gamma$ and $\gamma$. Hence since $s-s'\neq 0$ we have that $\alpha\beta + \gamma$ and $s'(\alpha\beta+\gamma)+(s-s')\gamma$ are IUD and hence so are $g^{\alpha\beta + \gamma}$ and $g^{s'(\alpha\beta + \gamma) + (s-s')\gamma}$, as required. Discussion ========== Related work {#related-work .unnumbered} ------------ The other approach for achieving ambiguity of which the author is aware is to use a Private Set Intersection Cardinality (PSI-CA) protocol to allow users to determine whether the set of tokens they have collected intersects with the set of tokens held by the server from infected users, without learning which tokens are in the intersection. This was proposed independently in [@trieu2020epione] and in [@contrail2020]. The security analysis in [@trieu2020epione] is expressly limited to the semi-honest setting, although it is suggested that one could guard against a dishonest user by requiring them to provide zero-knowledge proofs of correct behaviour, no doubt with significant performance consequences. The protocol in [@contrail2020] is similarly clearly flawed in the presence of a fully malicious user (specifically, at step 2 of the protocol, Alice may use different values of $\alpha$ for different $x_i$ and thereby reidentify elements despite Bob’s permutation). Moreover, no proofs are provided for the claimed security properties, and it seems that even in the semi-honest setting the claim that the server obtains no information about the contacts of undiagnosed users may be incorrect (for example, if the authorities can send to a suspect two tokens $x$ and $x'$ such that $x'=x^2$ then they will be able to identify the suspect as Alice when she performs the protocol). Open questions {#open-questions .unnumbered} -------------- The trick for this protocol was to ensure correct behaviour not by cumbersome zero-knowledge proofs but by rerandomising in such a way that a malformed token just results in the malefactor seeing random noise. The most important question for future work is whether a similar trick can be applied to obtain a lightweight DH-based protocol for PSI-CA which is robust against a fully malicious adversary. This would be extremely desirable because it could easily be added to DP3T-style systems with no changes to the system structure or to the technically-constrained Bluetooth Low Energy tokens. A second question is whether it is possible for the server, rather than sending all the rerandomised tokens to each user, to instead combine them in some way such that the user can tell whether they included at least one of the special form $(x,x^{s_u})$. This would be desirable for both performance and privacy reasons, since it would prevent users from learning how many of the tokens sent in by infected individuals were theirs. If the question was whether they were all of the special form then this would be trivial: just multiply together all the tokens componentwise. Unfortunately we have been unable to find a similar solution for the ‘disjunctive’ task. [^1]: This work was supported by FNR under project SmartExit.	{ "pile_set_name": "ArXiv" }
--- abstract: 'The scanning tunneling microscope is arguably the most versatile instrument for probing the [local]{} density of states of material surfaces, molecules, and devices. Despite its versatility, it has a limited range of accessible energies, whereas other spectroscopic techniques typically have a limited spatial resolution. Tunable atomic systems, though, can mimic, for instance, materials and electronics and probe them in ways not easily achievable by traditional techniques, especially for transport phenomena. Here, we fuse atomic and tunneling techniques to demonstrate a method that provides both spatial and energy resolution, as well as expands the accessible energy range to that prevalent in many-body systems. In this hybrid approach, the current supplies a simple, yet quantitative operational definition of a local density of states for both interacting and non-interacting systems as the rate at which particles can be siphoned from the system of interest by a narrow energy band of non-interacting states. Ultra-cold atomic lattices are a natural platform for implementing this concept to visualize the energy and spatial dependence of the atom density in interacting, inhomogeneous lattices, including ones with nontrivial topologies.' author: - Daniel Gruss - 'Chih-Chun Chien' - Julio Barreiro - Massimiliano Di Ventra - Michael Zwolak title: 'An energy-resolved atomic scanning probe' --- Introduction ============ The density of states is ubiquitous in classical and quantum physics, as it quantifies the energy distribution of available states. The local density of states (LDOS), in particular, gives the available states at position $\mathbf{r}$ and frequency $\omega$ according to [@Hedin70] $$D(\mathbf{r},\omega)=\sum_{n}\|\langle\mathbf{r}\|\phi_n \rangle\|^2 \delta(\omega-\omega_n).$$ Here, $\|\phi_n\rangle$ is the $n$-th eigenfunction of the full Hamiltonian with eigenvalue $\omega_n$. Spectroscopic means allow for the measurement of the density of states over the entirety of a system’s energy spectrum, but do not generally provide spatial resolution. Scanning probes, though, measure the LDOS at the position of the probe tip, but do not have access to the whole spectrum. In the standard operation of a scanning tunneling microscope (STM) to probe the LDOS $D(\mu)$ at frequency $\mu$, the tip distance is held constant while the sample bias $-V$ is changed, leading to a steady-state current [@hansma1987scanning; @chen1993introduction; @stroscio1993methods; @wiesendanger1994scanning] $$I \propto \int_{-eV/\hbar}^0 D(\omega) d\omega . \label{eq:STM1}$$ The LDOS is found from the differential conductance, $$D(\mu) \propto \left. \frac{d I}{dV} \right\|_{eV/\hbar=\mu} . \label{eq:STM2}$$ This expression neglects the voltage dependence of the electronic transmission from tip to sample, among other factors. Interpretational issues notwithstanding, this limits the accurate extraction of the native electronic density of states to the linear response regime, as not all changes in the current are due to changes in the LDOS. In particular, when the system has strong many-body interactions, such as a poorly screened electronic impurity, large applied biases will disturb the natural local state by disrupting the nearby electron density. ![The many-body LDOS of an interacting system ${\mathcal{S}}$ is measured by putting it “in contact” with a non-interacting, narrow band probe ${\mathcal{P}}$ via the focused laser beam ${\mathcal{C}}$. (a) The cold-atom system can be realized by optical tweezers or optical lattice potentials, as shown schematically, with alkaline-earth atoms [@stellmer2014]. The lattice laser beams’ intensities and dimensions determine the coupling rates and trapping frequencies. The system (probe) lattice traps only atoms in their ground (metastable excited) electronic state [@prl-101-170504]. Coupling between the probe and system lattices is driven by a laser beam (i) tuned to the transition between the two electronic states of the atoms, and (ii) focused on the sites where the atoms “flow” between ${\mathcal{S}}$ and ${\mathcal{P}}$. The probing band has a frequency offset $\mu$ and a small hopping frequency $\omega_{\mathcal{P}}$. (b) A representation of the setup in frequency space shows the narrower range of frequencies and offset of the probe. \[fig:expdiagram\]](figure1){width="\linewidth"} Here, we bring the tunability of cold-atom lattices to bear on this problem. These systems provide a means to simulate condensed matter [@bloch2008many; @fertig2005strongly; @strohmaier2007interaction; @Lamacraft12; @chien2015quantum], including transport phenomena [@ott2004collisionally; @gunter2006bose; @brantut2012conduction; @krinner2015observation], while simultaneously yielding opportunities to go beyond solid-state scenarios. In this vein, rather than an applied bias, we propose to use a narrow band probe ${\mathcal{P}}$ that scans in energy (see Fig. \[fig:expdiagram\]) to interrogate the many-body system ${\mathcal{S}}$. The current into (out of) an empty (full) reservoir of bandwidth $4\omega_{\mathcal{P}}$ offset to a frequency $\mu$ is proportional to the LDOS, i.e., the fraction of particles at frequency $\mu$, $$\label{eq:propcurr} I \approx - 2e \omega_{\mathcal{P}}\int_{\mu - 2\omega_{\mathcal{P}}}^{\mu + 2\omega_{\mathcal{P}}} D(\omega) d\omega \approx -8e\omega_{\mathcal{P}}^2 D(\mu)\,,$$ so long as $\omega_{\mathcal{P}}$ is small relative to variations of $D(\omega)$. The occupied (unoccupied) local density of states is given by using an empty (full) probe ${\mathcal{P}}$. This setup requires tunability of ${\mathcal{P}}$: Its chemical potential, occupation, bandwidth/hopping, and “contact” magnitude/location with ${\mathcal{S}}$ need to be adjustable without compromising its non-interacting behavior. Ultra-cold atoms in artificial lattice potentials are naturally suited to implement this setup, as we will show \[see Fig. \[fig:expdiagram\](a) for a schematic\]. In this context, the method is applicable to a wider class of systems and simulations than spatially resolved radio-frequency spectroscopy, a powerful, yet invasive, tomographic method better suited to homogeneous systems with appropriate atomic hyperfine electronic states and photodissociative mechanisms [@prl-90-230404; @prl-99-090403]. ![The current $I$ versus time $t$ for the (a) occupied and (b) unoccupied ${\mathcal{S}}$ states using an initially empty and filled ${\mathcal{P}}$, respectively, offset to frequency $\mu$. Here, $\omega_{\mathcal{S}}= 1$ ms$^{-1}$ (a typical cold-atom tunneling frequency), $\omega_{\mathcal{P}}=0.1$ ms$^{-1}$, and the total lattice length is $N=64$. (c) The LDOS for ${\mathcal{S}}$ determined from Eq. using the average current in the region $t \in {\mathcal{T}}=[2$ ms, $32$ ms$]$ \[demarcated by dashed lines in (a) and (b)\], which minimizes transient and edge effects. The error bars (shaded regions) indicate the standard deviation of the current in the region ${\mathcal{T}}$ combined with the broadening error 2$\omega_{\mathcal{P}}$. The Fermi level, $\omega_F$, is found from the point were the occupied and unoccupied states cross over. (d) Integrated error of the LDOS versus total lattice size and (e) averaging time \[for $N=32$ (blue solid line) and $N=512$ (green dashed line)\]. The baseline error \[when $N\to\infty$, the dotted line in (d) and (e)\] is set by the non-zero $\omega_{\mathcal{P}}$, which broadens the actual LDOS \[see Eq. \]. For short lattices and times, the error in the LDOS is already small, only a few percent. Thus, even modest cold-atom systems or numerical many-body calculations can effectively reconstruct the LDOS. \[fig:avecurrnonint\]](figure2){width="\linewidth"} ![(a) The current versus time for the occupied (left) and unoccupied (right) states for interaction strengths $U_i = 0$ ms$^{-1}$, $2$ ms$^{-1}$, $4$ ms$^{-1}$ (from top to bottom), and total size $N=32$. (b) The resulting LDOS using the same procedure as in Fig. \[fig:avecurrnonint\] and with the shaded regions indicating the standard deviation of the current combined with the probe broadening. Note that a steady-state current still forms as the system ceases to be a true insulator once it is connected to the non-interacting probe. As $U$ increases, a gap opens up between the occupied and unoccupied bands, in addition to causing a pronounced broadening. The dashed lines on the figure indicate the LDOS found using a mean-field approximation (see Appendix \[sec:mfapprox\]). \[fig:avecurrintmott\]](figure3a "fig:"){width="\linewidth"} ![(a) The current versus time for the occupied (left) and unoccupied (right) states for interaction strengths $U_i = 0$ ms$^{-1}$, $2$ ms$^{-1}$, $4$ ms$^{-1}$ (from top to bottom), and total size $N=32$. (b) The resulting LDOS using the same procedure as in Fig. \[fig:avecurrnonint\] and with the shaded regions indicating the standard deviation of the current combined with the probe broadening. Note that a steady-state current still forms as the system ceases to be a true insulator once it is connected to the non-interacting probe. As $U$ increases, a gap opens up between the occupied and unoccupied bands, in addition to causing a pronounced broadening. The dashed lines on the figure indicate the LDOS found using a mean-field approximation (see Appendix \[sec:mfapprox\]). \[fig:avecurrintmott\]](figure3b "fig:"){width="\linewidth"} Results ======= General Approach ---------------- The Hamiltonian, $H$, of both the system and probe is $$\label{eq:tothamil} H_{\mathcal{S}}+ H_{\mathcal{C}}- \hbar \omega_{\mathcal{P}}\sum_{i \in {\mathcal{P}}} \left( {c_{i}^{\dagger}} {c_{i+1}^{{\phantom\dagger}}} + \text{h.c.} \right) + \hbar \mu \sum_{i \in {\mathcal{P}}} {c_{i}^{\dagger}} {c_{i}^{{\phantom\dagger}}},$$ where $\omega_{\mathcal{P}}$ and $\mu$ are the hopping and relative onsite frequencies, respectively (i.e., the trapping frequencies for a cold-atom system) and ${c_{i}^{\dagger}}$ (${c_{i}^{{\phantom\dagger}}}$) are the creation (annihilation) operators of site $i$. The many-body Hamiltonian of ${\mathcal{S}}$ is $H_{\mathcal{S}}$ , and the contact Hamiltonian between ${\mathcal{S}}$ and ${\mathcal{P}}$ is $H_{\mathcal{C}}$, which we also take to have hopping frequency $\omega_{\mathcal{P}}$, thus giving a weak tunneling into a narrow band. Although not necessary, for simplicity, the probe is taken to be one dimensional. Optical lattices are ideal for implementing the scanning probe concept above. Independent manipulation of the lattices for the ground and excited (metastable) electronic configurations of atoms allows for the construction of the system and probe [@prl-101-170504]. After ${\mathcal{S}}$ and ${\mathcal{P}}$ are loaded with ultracold atoms in different electronic configurations and in their independent lattice ground states, laser beam $\theta t$-pulses of duration $t$ and phase $\delta$ focused on sites $s$ and $p$ will activate the contact between ${\mathcal{S}}$ and ${\mathcal{P}}$: $H_{\mathcal{C}}=\hbar\omega_{\mathcal{P}}a^\dagger_p a_s + \text{h.c.}$ with $\omega_{\mathcal{P}}=i\theta e^{i\delta}$. The composite system then evolves, giving rise to the current $I$, which yields the LDOS at frequency $\mu$ according to Eq. . The measurement is repeated for each $\mu$ of interest. Furthermore, the system could be probed at multiple contact points, repeatedly, and using several probes (see Appendix \[sec:realize\]). Any experimental realization of this method requires tunability of the hopping frequencies (the probe bandwidth) and the onsite frequencies (the probe offset), among other parameters. The trap depth and spacing controlling the hopping can be tuned via the magnitude of the trapping potential. The onsite frequency is related to the ground state of each well, which can be adjusted with the transverse size of the lattice beam (see Appendix \[sec:realize\]). In ${\mathcal{S}}$, Bose-Fermi mixtures can be loaded or the optical potential engineered beyond a lattice to allow for the simulation of different types of many-body systems, such as spatially inhomogeneous or varying strength interactions. Further control, including interactions via optical or orbital Feshbach resonances [@prl-115-135301; @prl-115-265302; @prl-115-265301], can be achieved by assembling the two non-overlapping lattices atom by atom [@s-354-1021; @s-354-1024]. Non-interacting LDOS -------------------- We illustrate this method by examining a many-body system with Hamiltonian $$H_{\mathcal{S}}= -\hbar \sum_{i \in {\mathcal{S}}, \sigma} \omega_{i} \left( {c_{i,\sigma}^{\dagger}} {c_{i+1,\sigma}^{{\phantom\dagger}}} + \text{h.c.} \right) + \hbar \sum_{i \in {\mathcal{S}}} U_i n_{i, \uparrow} n_{i, \downarrow} ,$$ where $\omega_{i}$ are tunneling frequencies in the system (sometimes taken to be uniform, i.e., $\omega_i = \omega_{\mathcal{S}}$), $n_{i,\sigma}={c_{i,\sigma}^{\dagger}} {c_{i,\sigma}^{{\phantom\dagger}}}$ is the number operator, and $U_i$ is the interaction frequency. The two (or more) components may refer to the spins of electrons or internal states of ultracold atoms. When inter-spin interactions are present, $H_{\mathcal{P}}$ is expanded in a similar manner to include spin. We first examine a spin-polarized, non-interacting system ${\mathcal{S}}$: $U_i=0$ and $\omega_{i}=\omega_{\mathcal{S}}$ for all $i \in {\mathcal{S}}$, which can be both solved exactly for the current [@Zwolak04-1; @chien2012bosonic; @chien2014landauer] and the LDOS (see Appendix \[sec:nonintldos\]). Figure \[fig:avecurrnonint\](c) shows the reconstruction of the LDOS using Eq. with a finite lattice and time average, giving quantitative agreement with the exact LDOS. The dominant source of error is the finite probe bandwidth $\omega_{\mathcal{P}}$, not the length of the lattice or the time of the average \[Figs. \[fig:avecurrnonint\](d) and (e)\]. This demonstrates that cold-atom systems or many-body simulations are well suited to implement this method, even though they are limited to finite lengths and times. As the total lattice length, $N \to \infty$, the averaging time, ${\left\| {\mathcal{T}}\right\|}\to \infty$, and the probe bandwidth, $\omega_{\mathcal{P}}\to 0$ (in this order), the exact LDOS would be recovered. Many-body LDOS -------------- We now apply the same approach to an interacting system with a constant $U_i = U$. Figure \[fig:avecurrintmott\] shows the LDOS of a Mott-insulator like state. As the interaction strength increases, the band splits and a gap forms between the occupied and unoccupied bands, as is typical for a Mott insulator. However, the Fermi level, $\omega_F \approx U/2$, and the occupied band is shifted to higher frequency. The observations also agree with predictions from Green’s function calculations in the hopping-only and interaction-only limits (see Appendix \[sec:mfapprox\]). ![(a) Schematic of a dimerized lattice system ${\mathcal{S}}$ —the SSH model with interactions—with alternating hopping coefficients $\omega_{1}$ and $\omega_{2}$ being interrogated by the narrow band probe ${\mathcal{P}}$. (b,c) LDOS of the interacting SSH model with $\omega_{1} = 1$ ms$^{-1}$, $\omega_{2} = 1.5$ ms$^{-1}$, and $U = 2$ ms$^{-1}$ as a function of lattice site and chemical potential. When the lattice terminates on a weak bond, edge modes appear and live in the gap between the electronic bands. The LDOS is for (b) a half-filled system and (c) and three-quarters-filled system, both of which show the presence of edge modes. In the former case, the edge modes are half-occupied in the ground state and the interaction term is essentially equivalent to an onsite energy shift. In the latter, the larger overall filling causes a higher occupation in these same modes and broadens them into the upper band. The energy resolved atomic scanning probe allows a clear visualization of the sublattice localization of the edge modes. The Fermi level is found empirically from where the full and empty bands overlap (averaged over all the sites). Its error can be large when the LDOS is very small over some range of energies (other ways to find the Fermi level, such as pinning it at the midpoint between features, may yield better estimates). \[fig:toposurf\]](figure4){width="\linewidth"} Moreover, the energy-resolved local density of states elucidates the role of interactions on the physical response. For instance, a filling-dependent, conducting-to-nonconducting transition occurs as a function of $U$ for interacting-induced transport [@chien2013interaction]: An inhomogeneous quench in $U$, where the interaction strength is taken from $0$ to a finite value for half the lattice, drives particles from that half of the lattice to the other (non-interacting) half so long as $U$ is not too strong and the filling not too large. Figure \[fig:avecurrintmott\](b) demonstrates that it is this shift of the occupied bands to higher frequency that aligns occupied states in the interacting side to open states in the non-interacting side, allowing particles to flow. As $U$ is increased further, eventually only the tail of the occupied band is aligned with open states, thus giving a decreasing—but nonzero—current solely due to many-body interactions. The mean-field solution, however, predicts the current should go exactly to zero. When the spatial dependence of the density of states is of interest, the probe can be used in a regime analogous to a “scanning-mode”: The probe can be coupled to the system at any lattice site, as shown in Fig. \[fig:toposurf\](a). As an example, we examine the Su-Schrieffer-Heeger (SSH) model of electrons hopping in polyacetylene [@SSH79; @np-9-795] in the presence of many-body interactions. The SSH model has alternating electronic hopping coefficients that dimerize the lattice. Here, $\omega_{i}$ is $\omega_1 (\omega_2)$ when $i \in {\mathcal{S}}$ is even (odd). In the non-interacting case, it has a topological invariant—the winding number. When this number is 1, edge states will be present at the boundary. When the lattice is half-filled, the presence of a homogeneous interaction $U$ does not remove the edge states. These can be seen in Figs. \[fig:toposurf\](b) and (c) as the sharp, localized peak of the LDOS at the boundary, and the interaction acts by simply shifting the LDOS. When the filling is increased to three-quarters, the LDOS peak at the boundary is broadened into the upper band. Moreover, the splitting that occurs for half-filling is washed out as the Fermi frequency increases beyond the bandgap. Conclusion ========== We conclude by noting that the density of states is a central concept in our description of matter. We have provided an operational definition of the LDOS for many-body systems, applicable in and out of equilibrium (e.g., exciting ${\mathcal{S}}$ via a quench or some other process), to fermionic or bosonic systems, etc. The core principle is that for a current to flow—whether in a steady state or not—into an empty, narrow probe band, there must be occupied states at that energy (similarly for a full narrow band and unoccupied states). In contrast to other methods (such as the single-site method from Ref. ), this approach uses the restriction of the probe to access the long-time properties of the total current and does not require time-dependent variation once the particles are in motion. The measurement is resolved in [both]{} energy [and]{} space, as well in other characteristics (e.g., spin-resolved). We demonstrated that a cold-atom setup will allow for the measurement of this operational definition: The many-body LDOS can be extracted with minimal disturbance to the system. While cold-atom systems allow for tunability, issues still can arise regarding, e.g., the orbital character of the local states and higher energy excitations. Unlike solid-state systems, however, the effect of these issues can be separated or even corrected in this setup. A related approach would be to use tunable cold-atom systems to more controllably—i.e., with less disturbance to the native state—implement Eqs. and . The energy-resolved scanning probe, however, acts not to mimic solid-state systems, but rather to implement the ideal, tunneling-based probe, one that minimizes the total current flowing and other disturbances. This will complement the quantum-gas microscope [@Bakr09], which resolves the spatial location of atoms. The flourishing of quantum simulations, from emulating condensed matter [@bloch_quantum_2012; @cirac_goals_2012] to the physics of the early universe [@hung2013cosmology; @kasamatsu2013atomic], demonstrates the need to probe—both experimentally and numerically—the undisturbed density of states with spatial and energy resolution up to many-body scales. In this vein, the energy-resolved atomic scanning probe will illuminate the nature of excitations and symmetry breaking in everything from the mundane to the exotic. Acknowledgments =============== Daniel Gruss acknowledges support under the Cooperative Research Agreement between the University of Maryland and the National Institute of Standards and Technology Center for Nanoscale Science and Technology, Award 70NANB14H209, through the University of Maryland. Massimiliano Di Ventra acknowledges support from the DOE Grant No. DE-FG02-05ER46204. Numerical considerations and error quantification ================================================= In order to extract the LDOS from real-time measurements on finite lattices, the current must be averaged over a finite time, i.e., the current will be in a quasi-steady state [@di2008electrical]. The estimate of the LDOS is thus $$D(\mu) \propto \frac{1}{{\left\| {\mathcal{T}}\right\|}} \int_{\mathcal{T}}I(t) dt .$$ As ${\left\| {\mathcal{T}}\right\|},N \to \infty$, this will converge to the true steady state. As we demonstrate, an accurate LDOS is already apparent for small lattices and short times, and thus it requires only modest resources (it is not experimentally or numerically taxing). The numerical calculations are as follows: For the non-interacting system, we integrate the equations of motion to find the current [@Zwolak04-1; @chien2012bosonic; @chien2014landauer]. The transient current when the probe “comes in contact” with the system is damped on the characteristic tunneling time and the recurrence time is proportional to the lattice size, which dictates both the lower and upper limits to the time region ${\mathcal{T}}$ given a finite lattice length. We use ${\mathcal{T}}=[ 2$ ms, $N/2$ ms$]$, where $N$ is the lattice length. We can also define an error for the non-interacting case, $$\text{Percentage Error} = 100 \; \% \cdot \frac{\sqrt{\int [D(\mu) - {D_{\text{ex}}}(\mu)]^2 \; d\mu}}{\sqrt{\int {D_{\text{ex}}}(\mu)^2 \; d\mu}} ,$$ where ${D_{\text{ex}}}(\mu)$ is the exact LDOS. The uncertainty due to probe broadening is due to contributions to the current from anywhere within the probe bandwidth of $4\omega_{\mathcal{P}}$, which results in an error, $\sigma_+(\mu)$, of $$D(\mu) - \max \left[ D(\mu), D(\mu-2\omega_{\mathcal{P}}), D(\mu+2\omega_{\mathcal{P}}) \right]$$ for the positive $\mu$ side and, $\sigma_-(\mu)$, $$D(\mu) - \min \left[ D(\mu), D(\mu-2\omega_{\mathcal{P}}), D(\mu+2\omega_{\mathcal{P}}) \right]$$ for the negative side. We combine this with the standard deviation, $\sigma_\text{stdev}$, from the time-dependent current, giving a total error of $$\label{eq:toterror} \sigma^\pm_\text{tot} = \sqrt{\sigma_\pm^2 + \sigma_\text{stdev}^2} .$$ For interacting systems, we perform time-dependent, density matrix renormalization group calculations [@vidal2003efficient; @vidal2004efficient] within the ITensor tensor product library [@itensorsite]. In all simulations, we decrease the time step until the calculation converges with respect to energy and we allow the matrix product bond dimension to increase without bound. The energy cutoff is $10^{-9} \omega_{\mathcal{S}}$. The averaging is in the region ${\mathcal{T}}=[ 2$ ms, $N/2$ ms$]$, as with the non-interacting case. Non-interacting LDOS {#sec:nonintldos} ==================== The non-interacting lattice is exactly solvable for both the current and the LDOS. For simplicity, we take the probing lattice to be a paradigmatic, one-dimensional lattice with homogeneous hopping frequency $\omega_{{\mathcal{P}}}$. Since we are interested in a narrow band reservoir, we take the weak hopping $\omega_{{\mathcal{P}}}$ to also be the coupling between ${\mathcal{S}}$ and ${\mathcal{P}}$. For the current, we need the retarded Green’s function for this semi-infinite lattice with a frequency offset $\mu$ (i.e., an onsite potential shift), which is [@zwolak2002dna; @chien2014landauer] $$\label{eq:nonintgr} g^r_{{\mathcal{P}}}(\omega) = \frac{1}{2 \omega_{{\mathcal{P}}}^2} \left[ (\omega-\mu) - \im \sqrt{4 \omega_{{\mathcal{P}}}^2 - (\omega-\mu)^2 } \right] .$$ Using this expression, the current for the infinite lattice, $N \to \infty$, is given by the Landauer formula [@di2008electrical], $$I = -\frac{e}{2 \pi} \int_{-\infty}^{\infty} d\omega \left[ f_{\mathcal{S}}(\omega) - f_{\mathcal{P}}(\omega) \right] T(\omega),$$ where $f_{{\mathcal{S}}({\mathcal{P}})}$ are the initial particle distributions (Fermi-Dirac distributions or completely filled/empty) in ${\mathcal{S}}({\mathcal{P}})$ and $T(\omega)$ is the transmission coefficient $$T(\omega)=\frac{4 {\operatorname{Re}}\left[ \sqrt{4 \omega_{\mathcal{P}}^2 - (\omega - \mu)^2} \right] {\operatorname{Re}}\left[ \sqrt{4 \omega_{\mathcal{S}}^2 - \omega^2} \right]} {\left\| \mu + \im \sqrt{4 \omega_{\mathcal{P}}^2 - (\omega-\mu)^2 } + \im \sqrt{4 \omega_{\mathcal{S}}^2 - \omega^2 } \right\|^2} .$$ Implementing the setup with $f_{\mathcal{P}}(\omega) \in \{0, 1\}$ gives the reconstructed LDOS as simply a sum of the occupied and unoccupied states, and also directly yields the Fermi level, $\omega_F$. Above, the system of interest ${\mathcal{S}}$ is half filled and in its zero temperature ground state. Here, we show, for simplicity, how the LDOS of a fully filled non-interacting system can be mapped out. The current in this case is $I=\frac{-e}{2\pi}\int_{\mu-2\omega_{\mathcal{P}}}^{\mu+2\omega_{\mathcal{P}}}T(\omega)d\omega\approx -(2/\pi)e\omega_{\mathcal{P}}T(\mu)$ when $\omega_{\mathcal{P}}\ll \omega_{\mathcal{S}}$. Thus, $$I \approx -\frac{4}{\pi}e \left(\frac{\omega_{\mathcal{P}}}{\omega_{\mathcal{S}}}\right)^2 \sqrt{4 \omega_{\mathcal{S}}^2 - \mu^2}$$ for $\| \mu \| \lesssim \omega_{\mathcal{S}}$. Then Eq. gives the exact LDOS, ${D_{\text{ex}}}(\mu)=\sqrt{4 \omega_{\mathcal{S}}^2 - \mu^2}/(2\pi\omega_{\mathcal{S}}^2)$, for a non-interacting ${\mathcal{S}}$. The remaining terms in the full expression broaden the reconstructed LDOS by approximately the probe bandwidth, $4\omega_{\mathcal{P}}$, which thus has to be small enough to discriminate features in the LDOS of ${\mathcal{S}}$ and, when using a cold-atom setup, the bandwidth should be large enough to get an appreciable current. For systems in the thermodynamic limit, the LDOS is related to the real-space Green’s function by [@RickayzenBook] $$\label{eq:LDOSGreen} D(\mathbf{r},\omega)=-\frac{1}{\pi} {\operatorname{Im}}\left[ g^r(\mathbf{r},\mathbf{r},\omega) \right].$$ These two equations give the same expression for a noninteracting semi-infinite lattice. The energy dispersion is $\omega_k=-2\omega_{\mathcal{S}}\cos[k\pi/(N+1)]$ and $\langle r=0\|\phi_k\rangle=\sqrt{2/(N+1)}\sin[k\pi/(N+1)]$ with $k=0,1,\cdots, N$. Thus, $(L=N+1)$ $$\begin{aligned} D(r=0,\omega) &= \int dk \frac{2}{L}\sin^2\left(\frac{k\pi}{L}\right)\frac{L\delta(k-k_{\mathcal{S}})}{2\pi\omega_{\mathcal{S}}\sin(\frac{k\pi}{L})}\notag \\ &=\frac{1}{2\pi\omega_{\mathcal{S}}^2}\sqrt{4\omega_{\mathcal{S}}^2-\omega^2}.\end{aligned}$$ Here, $k_{\mathcal{S}}$ is the value when $\omega-\omega_{\mathcal{S}}=0$ is satisfied. The Green’s function is similar to Eq. . Explicitly, $g^r_{{\mathcal{S}}}(\omega) = \left( \omega - \im \sqrt{4 \omega_{{\mathcal{S}}}^2 - \omega^2 } \right) / (2 \omega_{{\mathcal{S}}}^2)$, so $D(r=0,\omega)=-(1/\pi)\mbox{Im}(g^r)=\sqrt{4\omega_{\mathcal{S}}^2-\omega^2}/(2\pi\omega_{{\mathcal{S}}}^2)$. These both agree with that obtained from Eq. and the Landauer formula. Figure \[fig:noninttopo\] shows the topological system of dimerized lattices from Fig. \[fig:toposurf\](b) without an applied interaction term, which yields to non-perturbed state of the SSH system. ![The topological system from Fig. \[fig:toposurf\] with interaction strength $U=0$ ms$^{-1}$ and at half filling. In this case, the even-sites can be exactly half-occupied, removing the shift on the Fermi frequency and the splitting of the modes when half-filled. If the filling is increased or decreased, the structure of the LDOS remains the same. \[fig:noninttopo\]](figure5){width="\linewidth"} ![Scanning-mode applied to an inhomogeneous system. (a) The time-dependent current shows the formation of a steady-state flow into ${\mathcal{P}}$ when attached to the center of ${\mathcal{S}}$. (b) ${\mathcal{S}}$ has an inhomogeneous interaction profile: a linear decrease from $U=4$ ms$^{-1}$ on one end to $U=0$ ms$^{-1}$ on the other. (c) The LDOS for occupied and unoccupied states as a function of frequency offset $\mu$ and contact lattice position $i$. The weakly-interacting region of the system allows more particles to occupy lower frequency states, while the more strongly-interacting side forces the open states well beyond the Fermi level. In addition, a superimposed even-odd effect is visible, which is due to the finite size of the lattice, creating oscillations from the boundary. \[fig:avecurrintinhomo\]](figure6a "fig:"){width="\linewidth"} ![Scanning-mode applied to an inhomogeneous system. (a) The time-dependent current shows the formation of a steady-state flow into ${\mathcal{P}}$ when attached to the center of ${\mathcal{S}}$. (b) ${\mathcal{S}}$ has an inhomogeneous interaction profile: a linear decrease from $U=4$ ms$^{-1}$ on one end to $U=0$ ms$^{-1}$ on the other. (c) The LDOS for occupied and unoccupied states as a function of frequency offset $\mu$ and contact lattice position $i$. The weakly-interacting region of the system allows more particles to occupy lower frequency states, while the more strongly-interacting side forces the open states well beyond the Fermi level. In addition, a superimposed even-odd effect is visible, which is due to the finite size of the lattice, creating oscillations from the boundary. \[fig:avecurrintinhomo\]](figure6bc "fig:"){width="\linewidth"} ![The integral of the measured LDOS from Fig. \[fig:avecurrintinhomo\] (blue x’s) below the Fermi level gives the total particle occupation, which matches that found from a direct measurement of $n$, the occupation number, from the numerical calculation of the ground state (green line). The formation of a steady state current then provides not only a measure of the occupied and unoccupied states but also yields the real space occupation. The error bars represent the propagated error from the broadening and standard deviation as in the previous figures. \[fig:occupation\]](figure7){width="\linewidth"} Mean-field approximation {#sec:mfapprox} ======================== The Green’s function of the infinite, uniform Hubbard model in the mean-field approximation is [@RickayzenBook] $$\label{eq:meanfieldpoles} g_{inf}(k,\sigma,\omega)=\frac{\omega+\mu_{\mathcal{S}}-U(1-n)}{(\omega+\mu_{\mathcal{S}}-\omega_k)(\omega+\mu_{\mathcal{S}}-U)-Un\omega_k}.$$ Here, we focus on the 1D case with $\omega_k=-2\omega_{\mathcal{S}}\cos(k \pi / L)$ and chemical potential $\mu_{\mathcal{S}}$ representing an onsite energy shift. $\sigma=\uparrow,\downarrow$ denotes the chosen spin species, and $n$ is the filling factor. The poles of the Green’s function lead to the energy dispersion, which has the form $(\omega-\omega^{-}_k)(\omega-\omega^{+}_{k})$ with $\omega^{\pm}_{k}$ giving the two bands: $$\omega^{\pm}_{k} = \frac{U}{2} - \mu_{\mathcal{S}}+ \frac{\omega_k}{2} \pm \sqrt{\frac{U^2}{4} + \left( n - \frac{1}{2} \right) U \omega_k + \frac{\omega_k^2}{4}}.$$ The retarded Green’s function in real space is $g_{inf}^r(r,\sigma,\omega)=(1/N_l)\sum_{k} e^{ikr} g_{inf}(k,\sigma,\omega+i0^{+})$, where the summation is within the first Brillouin zone and $N_l$ is the lattice size. Since the Green’s function is for a uniform, translational invariant system, we may choose $r=0$. The LDOS, however, is from the Green’s function of a half-infinite lattice. An infinite lattice can be thought of as an assembly of two half-infinite lattices connected by an additional central site, so we use a derivation similar to Ref. [@chien2014landauer] and obtain $g_{inf}^r (r=0,\sigma,\omega)=[\omega+\mu_{\mathcal{S}}-Un-2\omega_{\mathcal{S}}^2g^r (r=0,\sigma,\omega)]^{-1}$. Inverting the relation, we find $g^r (r=0,\sigma,\omega)$ for a half-infinite chain, and the LDOS can be obtained from Eq. . In the isolated-site limit (i.e., the “atomic limit”) where $(\omega_{\mathcal{S}}/U)\rightarrow 0$, the Green’s function is exactly solvable and on a selected site it is $g_{\mathcal{S}}(\sigma, \omega)=\left[\frac{1-(n/2)}{\omega+\mu_{\mathcal{S}}}+\frac{n/2}{\omega+\mu_{\mathcal{S}}-U} \right]$. In this limit there is no distinction between an infinite lattice and a half-infinite lattice because there is no tunneling between sites. By an analytic continuation $\omega\rightarrow \omega-i0^{+}$, we obtain $g_{\mathcal{S}}^r$. Then, $D(\sigma,\omega)=\{[1-(n/2)]\delta(\omega+\mu_{\mathcal{S}})+(n/2)\delta(\omega+\mu_{\mathcal{S}}-U)\}$. Thus, there are two peaks at $-\mu_{\mathcal{S}}$ and $-\mu_{\mathcal{S}}+U$. Away from the isolated-site limit, the two peaks broaden into two bands separated by $U$, and this agrees with our observation shown in Fig. \[fig:avecurrintmott\]. LDOS of an inhomogeneous interaction ==================================== In previous sections, we focus on systems that have a uniform interaction term $U$ applied to the entirety of ${\mathcal{S}}$. As an additional example, we examine a system in scanning-mode with spatially inhomogeneous interactions—e.g., a linear decrease in the interaction strength $U$, Fig. \[fig:avecurrintinhomo\](b)—can determine both how the particle density shifts in space and energy. Figure \[fig:avecurrintinhomo\](c) shows the occupied and unoccupied LDOS of this inhomogeneous lattice as a function of position. The spatial decrease of the interactions forces particles to the region with small interactions, where at the very end the lattice has an LDOS similar to a non-interacting system. Just near the non-interacting boundary, however, a large peak in occupied density of states forms, i.e., states pinned well below the Fermi level. On the interacting side, the number of particles is small with an LDOS just below the Fermi level. A superimposed even-odd effect is visible, which is due to finite lattice effects, creating oscillations away from the boundaries. Up to the calculated error, there is a direct correspondence between the particle occupation and the integrated occupied LDOS, as with the typical non-interacting LDOS. Figure \[fig:occupation\] shows the comparison between the particle occupation from the simulation and the LDOS integrated over the full energy range from the application of the scanning-mode configuration to the inhomogeneous, interacting system. We note that the atomic scanning probe approach to [numerically]{} computing the local density of states gives a straightforward alternative to other numerical methods [@schoenauer_observation_2016]. Experimental details {#sec:realize} ==================== Among several embodiments of the energy-resolved atomic scanning probe with ultracold atoms, such as one using nano-patterned magnetic traps, we envision a highly-controllable one based on engineered optical potentials with alkaline-earths, specifically, fermionic strontium atoms, $^{87}$Sr [@stellmer2014]. In the setup introduced in the text, we consider the electronic ground state $^1S_0$ and the metastable state $^3P_0$ coupled by a clock laser [@s-341-632]. The system (probe) lattice is tuned to the magic-zero wavelength of the metastable (ground) state [@prl-101-170504; @pra-92-040501]. Therefore, atoms of ${\mathcal{S}}$ in the ground state do not see the lattice of ${\mathcal{P}}$, and, vice versa, atoms of ${\mathcal{P}}$ in the metastable state do not see the lattice of ${\mathcal{S}}$. This guarantees independent control of both the system and probe trapping potentials and interactions. Multiple energy-resolved atomic scanning probes could be potentially implemented, and each probing performed multiple times in an scheme technically demanding, but not fundamentally limited. For example, a 1D lattice system ${\mathcal{S}}$ could be probed by multiple ${\mathcal{P}}_i$ aligned perpendicular to the system. For fermionic strontium atoms, the number of atoms in each probe ${\mathcal{P}}_i$ could be measured via fluorescence in a cycling transition as follows [@prl-101-170504]. First, carry out a Raman transfer of the atoms in ${\mathcal{P}}_i$ from the state 5s5p$^3P_0$ to 5s5p$^3P_2$ via 5s6s$^3S_1$. Second, implement pulsed Raman sideband imaging on the cycling transition 5s5p$^3P_2$-5s6d$^3D_3$ as in Ref. [@prl-114-213002] but using the Zeeman manifold [@prl-110-070403]. Third, bring atoms back to the original probe state 5s5p$^3P_0$ with a Raman transfer as in the first step. At this point, the probe ${\mathcal{P}}_i$ could be brought again into contact with the system. [46]{}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 [**, ()]{} @noop [, ()]{} @noop []{} (, ) @noop [ ()]{} @noop []{} (, ) in [](\doibase 10.1142/9789814590174_0001) (, ) pp. @noop [**, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} in @noop []{}, (, , ) @noop [**, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} [, ()](\doibase 10.1103/PhysRevLett.90.230404) [, ()](\doibase 10.1103/physrevlett.99.090403) [, ()](\doibase 10.1103/PhysRevLett.115.135301) [, ()](\doibase 10.1103/PhysRevLett.115.265302) [, ()](\doibase 10.1103/PhysRevLett.115.265301) [, ()](\doibase 10.1126/science.aah3778) [, ()](\doibase 10.1126/science.aah3752) @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} [, ()](\doibase 10.1038/nphys2790) @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop [, ()]{} @noop []{} (, , ) @noop [**, ()]{} @noop [, ()]{} @noop [“,” ]{}, @noop [, ()]{} @noop []{} (, ) @noop [ ()]{} [**, ()](\doibase 10.1126/science.1236929) [, ()](\doibase 10.1103/physreva.92.040501) [, ()](\doibase 10.1103/physrevlett.114.213002) [**, ()](\doibase 10.1103/PhysRevLett.110.070403)	{ "pile_set_name": "ArXiv" }
--- abstract: 'We characterize the observational properties of the convectively driven vortex flows recently discovered on the quiet Sun, using magnetograms, Dopplergrams and images obtained with the 1-m balloon-borne [Sunrise]{} telescope . By visual inspection of time series, we find some $3.1\times 10^{-3}$ vortices Mm$^{-2}$min$^{-1}$ , which is a factor of $\sim$1.7 larger than previous estimates. The mean duration of the individual events turns out to be 7.9min, with a standard deviation of 3.2min. In addition, we find several events appearing at the same locations along the duration of the time series (31.6min). Such recurrent vortices show up in the proper motion flow field map averaged over the time series. The typical vertical vorticities are $\lesssim 6\times$10$^{-3}\,$sec$^{-1}$, which corresponds to a period of rotation of some 35min. The vortices show a preferred counterclockwise sense of rotation, which we conjecture may have to do with the preferred vorticity impinged by the solar differential rotation.' author: - 'J. A. Bonet, I. Márquez, J. Sánchez Almeida, J. Palacios, V. Martínez Pillet, S. K. Solanki, J. C. del Toro Iniesta, V. Domingo, T. Berkefeld, W. Schmidt, A. Gandorfer, P. Barthol, and M. Knölker' title: 'SUNRISE/IMaX observations of convectively driven vortex flows in the Sun' --- Introduction {#intro} ============ Solar surface convection is driven by localized downdrafts that collect the cold plasma returning to the solar interior after releasing internal energy [e.g. @spr90; @stei98; @ras98]. Angular momentum conservation forces the plasma to spin up as it approaches the sinkhole, and vortices are formed at the downdrafts. Such convectively driven vortices were theoretically predicted and sought for long [e.g., @spr90; @bal98], but their observational discovery is fairly recent . In addition to supporting numerical models of solar surface convection, the photospheric vortices may be of importance as heating sources for the outer solar atmosphere. The downdrafts not only advect vorticity but also magnetic fields, which are intensified to kG field strengths in and around them. Buoyancy and the vertical geometry of the downdraft tend to align the magnetic field lines with the vertical, so that the spinning motions at photospheric levels can be propagated upward using the field lines as guides [e.g., @cho93; @zir93; @bal98]. Waves thus excited transport photospheric energy that can be deposited in higher layers of the atmosphere. Moreover, downdrafts often trap structures of mixed polarity, so that the swirling motions wind up opposite polarity field lines, facilitating magnetic reconnection and the ensuing energy release. These convectively driven vortex flows are a recently discovered phenomenon poorly characterized from an observational point of view. So far, we only know that the vortices are quite common, have no preferred sense of rotation at the solar equator, and last (at least) minutes [@bon08]. They are also visible in the chromosphere, where they seem to be associated with significant blueshifts [@wed09]. Most of them are small-scale [$\lesssim 0.5$Mm @bon08], but some have a much larger radius of influence [up to 20Mm, @att09; @bra88]. Lifetimes can be longer than 20min, and several observables (such as circular polarization and G-band intensity) simultaneously indicate the presence of vortical motions [@bal10]. In terms of global properties rather than individual eddies, the vertical vorticity inferred from proper motions seems to be higher in downflow regions, suggesting excess vorticity in intergranular lanes [@wan95b; @pot05]. @nise03 searched for evidence of vorticity in the motions of isolated G-band bright points. [Sunrise]{} is a 1-m balloon-borne solar telescope [@bar10] which, together with its Imaging Magnetograph eXperiment [IMaX, @mar10], provides time series with state-of-the-art high spatial resolution images and magnetograms. The dataset is ideal for a systematic characterization of the poorly known physical properties of the vortices. Thus this Letter presents a comprehensive observational characterization of the vortices in the photosphere. The actual data set and the procedure to detect vortices are described in § \[description\]. The main results are summarized in § \[results\]. Based on such results, we compare the observed properties with the predictions of the numerical simulations of solar surface convection (§ \[discussion\]). Observation and analysis procedure {#description} ================================== The data were gathered with IMaX near the solar disk center on June 9, 2009 (UT 01:31-02:02; although the exact location is not known) during the first science flight of [Sunrise]{} [@bar10; @sol1]. The IMaX magnetograph uses a LiNbO$_3$ etalon operating in double pass, liquid crystal variable retarders as the polarization modulator, and a beam splitter as the polarization analyzer. We use here data recorded in the so-called V5-6 observing mode [see @mar10] where images of the four Stokes parameters were taken at five wavelengths along the profile of the magnetic-sensitive line Fe [i]{}$\lambda$5250.2 Å ($\pm 80, \pm 40~$mÅ from line center, plus continuum at +227mÅ). After the science observing run a calibration set consisting of 30 in-focus and out-of-focus image-pairs was recorded for post-facto retrieval of the point spread function (PSF) using phase diversity [@gon82; @pax96]. The science images were reconstructed by deconvolution using a modified Wiener filter and the calibrated PSF of the optical system. IMaX provides 85mÅ spectral resolution and between 015 and 018 angular resolution in the reconstructed images. Dopplergrams and magnetograms are derived from the Stokes parameters by using the approach described in [@mar10]. All in all, the reduction procedure renders time series of images, magnetograms, and Dopplergrams with a cadence of 33sec, a spatial sampling of 0055, and an effective field-of-view (FOV) of 45$\times$45. As inferred from the standard deviation of the polarization signals at the continuum wavelength, the circular polarization noise is $5\times 10^{-4}$ in units of the continuum intensity. The observing material analyzed here consists of a time series lasting 31.6 min. Movies were generated after rigid alignment and $p$-mode subsonic filtering [@tit89]. @bon08 found the small-scale vortex flows by visual feature tracking of magnetic G-band bright points (BPs). The same downdrafts producing vortices also advect and concentrate magnetic flux (see § \[intro\]), which often appears as BPs when the field strength is in the kG regime [see, e.g., @san04a and references therein]. As Bonet and colleagues acknowledge, the technique is rather limited since whirlpools without BPs are expected, and they escape from detection. Taking advantage of the combined high spatial resolution and high polarimetric sensitivity of IMaX/[Sunrise]{} data, we tried to detect and study vortices in longitudinal magnetograms, which are sensitive not only to kGs fields, but to plasmas with the full range of field strengths. In addition, we broaden the study using continuum intensity, line minimum intensity, line-of-sight (LOS) velocity, and line width (the last three parameters obtained from a Gaussian fit to the five sampled wavelengths). Individual vortices are identified and characterized through the following steps: \(1) The detection is based on a visual inspection of longitudinal-magnetogram movies. Playing back and forth these movies, we identify those locations and time intervals where structures seem to rotate. (2) Once a vortex candidate is thus located, it is isolated in a 55$\times$55 sub-field, where the corresponding sub-fields in the other four physical parameters are visually inspected for swirling motions (see Fig. \[fig1\]). (3) Horizontal velocity maps of the event in all the five parameters are created from proper motions. The horizontal motions are measured in these reduced sub-fields employing the local correlation tracking (LCT) algorithm of @nov88, as implemented by @mol94, and with a Gaussian tracking window of about 04 FWHM. In order to help the algorithm, the original images are interpolated in time and space so as to have a pixel of 0028 and a cadence of 11sec. The horizontal velocities obtained by comparing successive images are time averaged the duration of the event. Examples of such velocity maps are shown in Fig. \[fig2\]. The size of the tracking window was chosen as a trade off to be large enough for the LCT algorithm to have structures to track, yet small enough to minimize the presence of several structures with different velocities. If the velocity maps do not show a regular closed shape in at least two physical parameters, then we discard the vortex candidate and start from step 1. \(4) We compute the vertical vorticity, the divergence, and the curvature corresponding to the LCT horizontal velocities. Given the velocity [U]{}, the vertical vorticity $(\nabla\times {\bf U})_z$ can be interpreted in terms of the local angular velocity since a plasma in pure rotational motion has $(\nabla\times {\bf U})_z= 2 w$, with $w$ the angular velocity. Similarly, the curvature of such motion is $\kappa = {{1}\over{2}}\,(\nabla\times {\bf U})_z\,\|{\bf U}\|^{-1}$, with $\kappa^{-1}$ the radius of curvature. Examples of vorticity and curvature maps are shown in Figs. \[fig3\], \[fig4\]. \(5) Using the LCT horizontal velocities, we track the evolution of passively advected tracers ([corks]{}) spread out all over the sub-field [@yi92]. As time goes by, the corks end up in the sinkhole, revealing the position of the sink according to different physical parameters (see the white points in Fig. \[fig3\]). If the sinkhole positions inferred from the different physical parameters do not agree within 04, then the vortex is discarded and we return to step 1. We find the curvature maps to be an efficient complementary tool for vortex center detection. In most cases these maps show up the sinkholes as conspicuous point-like features (see Fig. \[fig4\]) with positions that agree well with the centers determined as the final destination of the corks in the corresponding movie (c.f. Figs. \[fig3\], \[fig4\]). The above list outlines the general procedure, but we do not disregard casual detections, e.g., when a second vortex was observed in any of the subfields corresponding to another vortex. Thus we find by chance some vortices which do not show up in the magnetogram signals. In addition to the LCT velocity maps of the individual vortices, we also computed the flow field for the full FOV during the full duration of the sequence. Isolated point-like features in the corresponding curvature map suggest the presence of vortices, and their existence motivates further inspection of magnetograms for swirling motions. As one can image from this cumbersome procedure, the FOV has been unequally searched. We focused on those regions were the magnetograph signals were largest, so that the [effective]{} FOV of our research is only 285 $\times$ 285. This area is used in the estimates below unless otherwise stated. Two animations, one showing the event in Figs. \[fig1\]-\[fig4\] and another with a different one, are given in the on-line material in the electronic edition of the Journal. ![image](f1.ps){width="90.00000%"} ![image](f2.ps){width="90.00000%"} The 1 -Mm radius circles centered in the sinkhole are included for reference. \[fig2\] ![image](f3.ps){width="90.00000%"} ![image](f4.ps){width="90.00000%"} Results ======= Following the procedure outlined above, we detected 42 vortices with proper velocity maps. They imply a space-time-density of $d\simeq 3.1\times 10^{-3}$ vortices Mm$^{-2}$min$^{-1}$. In addition, 31 structures showing vortical motion were discarded because they did not fullfill our strict selection criteria. The selected events are used here to characterize the observational properties of vortices. We have assigned a duration to each vortex, computed as the time interval in which the vortex motions are clearest. These durations span from 5min to 20min with a mean of $\tau$ $\simeq$ 7.9min and a standard deviation of 3.2min. The interpretation of these intervals as life-times is not devoid of uncertainty, though. Often we shorten the interval to assure a most pure swirling motion. In addition, some vortices appear in the mean flow field corresponding to the full time series, indicating that they probably last longer than the time span of the series itself. These long lasting vortices often involve a complex behavior: several short-lived vortices appear and disappear in the same location, giving rise to recurrent vortices. The position of their vortex centers may be static or drift with time. The recurrent vortices may or may not keep the same sense of rotation. In the latter case, however, we cannot consider the vortices to be strictly recurrent. We even find cases where the presence of a vortex is hinted as a clear pointlike feature in the curvature image, but we failed to identify any vortex at that position during the sequence. The largest surprise of our analysis is the finding of a preferred sense of rotation: 27 counterclockwise vs. 15 clockwise. This is a big difference with respect to @bon08, where the two senses of rotation were observed equally. We analyze this issue in § \[discussion\]. ![Probability density function (PDF) of vertical vorticities obtained from LCT proper motions. The 5min time averaged distributions for the full FOV are shown as a dotted line, and a dashed line, depending on whether they were derived from the continuum intensity or the magnetogram, respectively. The dotted-dashed line corresponds to the vorticity in a region 0.5 Mm wide around a number of well defined vortices (with the vorticity signed so that all vortices have positive vorticity). The solid line also represents a local histogram considering a region of 2 Mm. The excess of vorticity at some 0.005 sec$^{-1}$ is produced by the vortices. []{data-label="histograms"}](f5.ps){width="50.00000%"} Figure \[histograms\] shows histograms of vertical vorticities to characterize our measurements. The solid line corresponds to the vorticity in a region 2 Mm wide around a number of well defined vortices. The histogram considering only 0.5 Mm is shown as the dotted-dashed line. These histograms reveal the signature of the vortices, which turn out to have vorticities up to 0.006 sec$^{-1}$, corresponding to a period of rotation of some 35min. These are vorticities inferred from the magnetograms, which are systematically smaller than those obtained from the other observables. The difference can be pinned down to the proper motion velocity field, which tends to zero outside the large magnetic concentrations, where the polarization signals are low. The proper motions inferred from the other parameters extend throughout (Fig. \[fig2\]), so that the histograms of vorticities have less contribution at low values and show extended tails (cf. the dashed line and the dotted line in Fig. \[histograms\], which represent the histograms of vorticities for the full FOV inferred from the magnetogram and the continuum intensity, respectively). If a vortex has its axis tilted with respect to the [LOS]{}, it should produce a characteristic Doppler signal similar to the rotation curve of a galaxy, with a close pair redshift-blueshift centered at the sinkhole. The expected signals are of the order of a few hundred ms$^{-1}$ for moderate-high inclinations (30), which are at the limit of our observation. We unsuccessfully seek for such signals in the Doppler maps, meaning that the vortex motions are not highly tilted with respect to the horizontal plane. Moreover, we note the discovery of horizontal vortex flows near the edges of granules reported by @ste1 using the same [Sunrise]{}/IMaX data and that are likely to be of a different nature to those analyzed here. Discussion ========== Small-scale vortex flows in quiet Sun are detected using five different observational parameters: magnetogram, continuum intensity, line core intensity, LOS velocity, and line width. The fact that in most cases the detection is consistent in three or more of these observables (showing different morphology) reinforces the reliability of the events found. The number density of vortices is $\sim$1.7 times larger than that found by @bon08, even though we have been far more strict here. The increase can be ascribed to the use of a larger variety of physical parameters to detect the swirls. Most of the vortices are shortlived events observed during less than 10 min, but some of them last longer than the full time series, with recurrent vortices appearing in roughly the same place. Vortices have a typical vorticity smaller than 6$\times$10$^{-3}\,$sec$^{-1}$, which corresponds to a period of rotation of some 35min. For reference, the large maelstrom found by @bra88 had a vorticity ten times smaller, with an associated period of some 6 hours. The measured vorticities are generally much smaller than those predicted by the numerical simulations of magneto convection [@stei98 Stein 2010 private communication]. We think that the bulk of this difference can be attributed to the limited spatial-temporal resolution of the observations. We are unable to identify vortices smaller than the tracking window, and/or lasting less than 8-10 frames, which sets an upper limit to the vorticity of some 0.04 sec$^{-1}$. Simulations indicate that the highest vorticities occur at the smallest resolvable scales and, thus, the predicted distribution critically depends on the numerical resolution of the simulation [see Fig. 31 in @stei98]. The curvature maps (i.e., maps of the inverse radius of curvature) show the presence of vortices much better than the vorticity maps (see Figs. \[fig3\], \[fig4\]). The vorticity is sensitive to the flow speeds, which are large outside vortices, creating spurious vorticity signals. The curvature, however, only enhances areas where the swirling motions occur at small scales, independently of the velocity. The sinkholes show up conspicuously as local extremes in the curvature maps, and this new property should be employed when devising automatic algorithms to detect vortices. We find a preferred sense of rotation for the vortices (27 counterclockwise vs 15 clockwise). If the two senses of rotation were equally-probable then our observation would be highly unlikely (the probability is 4.4% assuming a binomial distribution). Nevertheless, the statistics is not large enough to provide a firm conclusion. The role of Coriolis forces on setting up this difference can also be discarded since the vortex motions involve time scales much too short to be affected by the solar rotation. The preferred sense of rotation may have to do with the solar differential rotation. The plasma poleward from the sink tends to lag behind, whereas the plasma equatorward from the sink moves forward. Such a difference impinges a preferred counterclockwise sense of rotation in the northern hemisphere, and a clockwise sense in the southern hemisphere. Back-of-the-envelope estimates indicate that the effect produces the right order of magnitude vorticity.[^1] If this conjecture turns out to be correct, it naturally explains the difference with respect to @bon08, whose observations correspond to the solar equator where there is no preferred sense of rotation. In our case the dominant counterclockwise rotation is consistent with an observed FOV in the northern hemisphere. Thanks are due to R. Stein for discussions on the comparison with numerical simulations, and to C. Pastor for her support during the data interpretation. The German contribution to SUNRISE is funded by the Bundesministerium für Wirtschaft und Technologie through Deutsches Zentrum für Luft- und Raumfahrt e.V. (DLR), Grant No. 50 OU 0401, and by the Innovationsfond of the President of the Max Planck Society (MPG). The Spanish contribution has been funded by the Spanish MICINN under projects ESP2006-13030-C06, AYA2009-14105-C06 (including European FEDER funds), AYA2007-66502, AYA2007-63881 and by the EC (SOLAIRE Network – MTRN-CT-2006-035484). The HAO contribution was partly funded through NASA grant number NNX08AH38G. This work has been partly supported by the WCU grant (No R31-10016) funded by the Korean Ministry of Education, Science & Technology [Facilities:]{} [23]{} natexlab\#1[\#1]{} , R., [Innes]{}, D. E., & [Potts]{}, H. E. 2009, , 493, L13 , L., [Vargas Domínguez]{}, S., [Palacios]{}, J., [Cabello]{}, I., & [Domingo]{}, V. 2010, , 513, L6 , P., [Gandorfer]{}, A., [Solanki]{}, S. K. [et al.]{} 2010, , in press (arXiv:1009.2689) , J. A., [M[á]{}rquez]{}, I., [S[á]{}nchez Almeida]{}, J., [Cabello]{}, I., & [Domingo]{}, V. 2008, , 687, L131 , P. N., [Scharmer]{}, G. B., [Ferguson]{}, S., [et al.]{} 1988, , 335, 238 , A. R., [Auffret]{}, H., & [Priest]{}, E. R. 1993, , 143, 49, revise , R. A. 1982, Optical Engineering, 21, 829 , P. R., [Yurchyshyn]{}, V., [Cao]{}, W., [et al.]{} 2010, , 714, L31 , V., [del Toro Iniesta]{}, J. C., [Álvarez-Herrero]{}, A., [et al.]{} 2010, , in press (arXiv:1009.1095) , R. & [Yi]{}, Z. 1994, [Internal Rep.]{} 31,Institute of Theoretical Astrophysics, University of Oslo, Oslo , P., [van Ballegooijen]{}, A. A., [de Wijn]{}, A. G., & [Sütterlin]{}, P. 2003, , 587, 458 , L. J. & [Simon]{}, G. W. 1988, , 333, 427 , R. G., [Seldin]{}, J. H., [Löfdahl]{}, M. G., [Scharmer]{}, G. B., & [Keller]{}, C. U. 1996, , 466, 1087 , W. & [Brandt]{}, P. N. 2005, Hvar Observatory Bulletin, 29, 61 , M. P. 1998, Journal of Fluid Mechanics, 369, 125 , J., [M[' a]{}rquez]{}, I., [Bonet]{}, J. A., [Dom[í]{}nguez Cerde[\~ n]{}a]{}, I., & [Muller]{}, R. 2004, , 609, L91 , S. K., [Barthol]{}, P., [Danilovic]{}, S., [Feller]{}, A. [et al.]{} 2010, , this issue , H. C., [Nordlund]{}, [Å]{}., & [Title]{}, A. M. 1990, , 28, 263 , R. F. I. & [Nordlund]{}, [Å]{}. 1998, , 499, 914 , O., [Franz]{}, M., [Bello González]{}, N., [Nutto]{}, Ch. [et al.]{} 2010, , this issue , A. M., [Tarbell]{}, T. D., [Topka]{}, K. P., [et al.]{} 1989, , 336, 475 , A. A., [Nisenson]{}, P., [Noyes]{}, R. W., [et al.]{} 1998, , 509, 435 , Y., [Noyes]{}, R. W., [Tarbell]{}, T. D., & [Title]{}, A. M. 1995, , 447, 419 , S. & [Rouppe van der Voort]{}, L. 2009, , 507, L9 , Z. 1992, PhD thesis, University of Oslo, Oslo , J. B. 1993, , 147, 47 [^1]: \[foot\]Assume plasma separated by 40Mm (i.e., the size of a supergranule) and converging to a sinkhole in the middle. Then the difference of differential rotations at latitude 30 deg produces a difference of velocities of some 35 msec$^{-1}$. If the circulation is approximately conserved during the convergence, a vortex 0.5 Mm wide would have a vorticity of 0.01 sec$^{-1}$.	{ "pile_set_name": "ArXiv" }
--- abstract: 'We study the bosonic analog of Andreev reflection at a normal-superfluid interface where the superfluid is a boson condensate. We model the normal region as a zone where nonlinear effects can be neglected. Against the background of a decaying condensate, we identify a novel contribution to the current of reflected atoms. The group velocity of this Andreev reflected component differs from that of the normally reflected one. For a three-dimensional planar or two-dimensional linear interface Andreev reflection is neither specular nor conjugate.' author: - 'I. Zapata' - 'F. Sols' title: Andreev reflection in bosonic condensates --- Andreev reflection occurs at a normal-superconductor interface when an incident electron/hole on the normal side is reflected as a hole/electron [@Andreev64]. It is a current-carrying process which, within a mean-field description, conserves the quasiparticle but not the charge current. The Andreev reflection of an incident hole is equivalent to the emission of a singlet electron pair from the superconductor into the normal metal [@2e-picture]. It has been observed in superconductors [@scr-exp] and superfluid $^{3}$He [@Enrico93], and has been proposed for fermionic quantum gases [@Schaeybroeck07]. In these Fermi superfluids the conversion between particles and holes reflects the fermion-pair structure of the superfluid phase. Andreev reflection determines the transport properties of fermionic superfluid-normal interfaces. It seems natural to investigate its bosonic analog. Andreev-like processes have been proposed for cold atom realizations of Bose-Hubard models [@Daley08] and Luttinger liquids [@Tokuno07]. These two works investigate the generation of reflected dips in the particle density following the incidence of a density bump on an interface. Here we investigate the possible analog of Andreev reflection at the interface between a bosonic superfluid and a normal gas. We seek an analog closer to the fermionic case by exploring the conversion between different types of Bogoliubov quasiparticles, which for bosons are defined through the transformation $\psi=e^{i\phi}\sum_{n}(u_{n}\gamma_{n}-v_{n}^{\ast}\gamma _{n}^{+})$, where $\psi$ is the field fluctuation operator, $\phi$ is the condensate phase, and $\gamma_{n}$ destroys quasiparticle $n$ [@Fetter72]. In bosonic fields the $u$ and $v$ components play asymmetric roles; for instance, under usual circumstances $v=0$ in a normal gas. In this context, we define Andreev reflection in a normal-superfluid (NS) interface as the scattering event which generates a quasiparticle with $v\neq0$ on the N side following the incidence of a conventional ($v=0$) quasiparticle from the same side. The hole-like quasiparticle (which however contributes positively to the density) travels at a speed different from the normally reflected one, which should be experimentally observable. A natural way of forming a NS interface in boson gases is to create a potential step that forces a difference in the atom densities on each side of the step. At a given temperature, the densities may be such that the gas is normal on the left and superfluid on the right. However, since in a normal boson gas the chemical potential stays below zero kinetic energy, transport near equilibrium is possible only if the condensate is confined. The $v$ component represents propagation below the chemical potential, so it can only be evanescent on the N side within that scenario [@Wynveen00; @Poulsen03]. As Andreev reflection from a confined condensate is ruled out, one is led to consider the alternative case of a decaying condensate. We define the normal (or non-superfluid) region as that in which the condensate flows faster than the local speed of sound; in our case, the outgoing coherent beam. For simplicity, we assume the interaction coupling strength to be zero on the N side, which does not change the essential physics. In general, transport through an interface separating subsonic from supersonic flow poses a new paradigm in superfluid transport [@Leboeuf01]. Like for fermions, particle current is not conserved within a (non-selfconsistent) mean-field description of Andreev processes [@SolsA94; @Sanchez97]. In both cases the condensate is responsible for the loss or gain of particles. However, there are several important differences with respect to the fermion case. The eigenvalue problem whose solutions are the Bogoliubov quasiparticles is non-Hermitian. This results in a peculiar normalization condition \[$\nu\equiv\int(\|u\|^{2}-\|v\|^{2})dx=1$\] for the quasiparticle wave functions [@Pethick02]. We will see that the conservation of $\nu$ yields unconventional relations between the scattering amplitudes. A counterintuitive but straightforward consequence is that, unlike for fermions, both particle-like and hole-like excitations carry an increase in the density (with respect to the Bogoliubov vacuum), since $\rho\sim \|u\|^{2}+\|v\|^{2}$. The existence of these differences suggests that bosonic quasiparticles define a rich novel class of quantum transport problems. The need for a confined condensate to generate Andreev reflection can be better appreciated from a study of the Bogoliubov – de Gennes (BdG) equations for bosons [@Dalfovo99; @Leggett01]: \[ptb\] [NDispersionRelation.eps]{} $$\begin{array} [c]{c}i\partial_{t}\dbinom{u}{v}=\left( \begin{array} [c]{cc}\mathcal{L}(x) & -n_{0}(x)g(x)\\ n_{0}(x)g(x) & -\mathcal{L}^{\ast}(x) \end{array} \right) \dbinom{u}{v}\\ \\ \mathcal{L}(x)\equiv-\frac{1}{2}D_{x}^{2}+V(x)+2n_{0}(x)g(x)-\mu \end{array} \label{eqnBdG}$$ Here, $\hbar=m=1$ units have been used, $n_{0}(x)$ is the condensate density, $g(x)$ is the (piecewise constant) coupling strength, $D_{x}\equiv\partial _{x}-i\phi^{\prime}$ with $\phi(x)$ the condensate phase, and $\mu$ is the condensate chemical potential. The scattering channels are defined by the plane wave solutions ($u,v\sim e^{i(kx-\varepsilon t)}$) in the flat region $V(x)=V_{0}$. The dispersion relations in the non-superfluid region ($g(x)n_{0}(x)=0$) are shown in Fig. \[grPlotNDispersionRelation\]. In the confined case ($\delta\mu\equiv\mu-V_{0}<0$), the left Fig. \[grPlotNDispersionRelation\] shows that, away from the interface, asymptotic propagation with $\varepsilon>0$ is only possible for the $u$ (particle-like) component, while the $v$ (hole-like) component can only be evanescent. A decaying ($\delta\mu>0$) condensate exhibits a richer scenario. The right Fig. \[grPlotNDispersionRelation\] shows a crossing of the $u$- and $v$-branches. This permits elastic conversion from a particle into a hole, in formal analogy with fermionic Andreev reflection. For a given energy within the allowed crossing range $\varepsilon\in(-\delta\mu,\delta\mu)$, there are four propagating waves. The right (left) going wave numbers are $k_{N}^{+}$ and $-k_{N}^{-}$ ($-k_{N}^{+}$ and $k_{N}^{-}$), where $k_{N}^{\pm}\equiv\sqrt{2(\delta\mu \pm\varepsilon)}$. The group velocities are those of a quadratic dispersion relation, $w_{N}^{\pm}=k_{N}^{\pm}$. Thus, for an incident particle from the normal side with velocity $w_{i}=\sqrt{2(\delta\mu+\varepsilon)}$, there are two outgoing channels with velocity $w_{f}$ satisfying $w_{f}=w_{i}$ (normal reflection) or $w_{i}^{2}+w_{f}^{2}=4~\delta\mu$ (Andreev reflection). The prediction of this novel form of propagation on the normal side is the main result of this paper. On the superfluid side, excitations travel with wavevector $k_{S}^{+}\equiv\sqrt{2}[\sqrt{\varepsilon^{2}+\mu^{2}}-\mu]^{1/2}$. Here, $\mu=gn_{+}$, with $n_{\pm}\equiv\lim_{x\rightarrow\pm\infty}n_{0}(x)$. At low energies, the speed $w_{S}=d\varepsilon/dk_{S}^{+}$ approaches the speed of sound $c=gn_{+}$. An Andreev transmission process occurs for energies $\varepsilon \in(0,\delta\mu)$ when a phonon incident from S is transmitted into N as a hole-like ($u=0$) quasiparticle. The other transmission channel is particle-like ($v=0$) and travels in N at a different velocity ($w_{N}^{+}$ vs. $w_{N}^{-}$). Remarkably, the outgoing hole-like scattering channels which characterize an Andreev process may exist despite its anomalous normalization $\nu<0$. We note, however, that the conservation of $\nu$ applies to the total scattering state: a particular channel may have $\nu<0$ whereas the total state retains the standard normalization $\nu>0$. \[ptb\] [PotentialPlot.eps]{} Bogoliubov quasiparticles are generally defined against the background of a particular condensate wave function. We consider a one-atom potential of the type shown in Fig. \[grPotentialPlot\], which provides a model for a leaking condensate. The condensate wave function $\Psi$ satisfies the time-dependent Gross-Pitaevskii equation $$\begin{aligned} i\partial_{t}\Psi & =-\tfrac{1}{2}\partial_{xx}\Psi+V(x)\Psi+g(x)\|\Psi \|^{2}\Psi~,\label{eqnGrossPitaevskii}\\ V(x) & =Z\,\delta(x)+V_{0}\,\theta(-x)~. \label{eqnPotentialProfile}$$ In the asymptotic regions it admits a solution of the type $\Psi (x\rightarrow\pm\infty)\sim e^{i(qx-\mu t)}$. In the previous discussion on scattering channels, we have assumed $q=0$ for simplicity, which —it can be shown— does not change the main conclusions. A stationary solution to (\[eqnGrossPitaevskii\]) and (\[eqnPotentialProfile\]) can be found analytically [@details]. An important question is whether or not that solution is stable. For $Z/\sqrt {\mu}\gg1$, the exact stationary solution resembles closely the wave function $\Psi(x)=\sqrt{n_{-}}\theta(-x)+\sqrt{n_{+}}\theta(x)$. For this approximate, step-like condensate, we have proved that the solutions to the BdG equations (\[eqnBdG\]) do not have complex eigenvalues. In other words, in the thick barrier limit the exact stationary solution resembles a solution which has been proven to be stable. In general, the condensed atom density on the N side is uniform and equal to $n_{-}=\|j\|/\sqrt{2\delta\mu}$, with atom velocity $w=-\sqrt{2\delta\mu}$, as determined by simple energy considerations. In this thick barrier limit, the resulting current is $j\simeq-\mu^{2}\sqrt{2\delta \mu}/gZ^{2}$. This solution applies to an interface separating a supersonic (normal) from a subsonic (superfluid) region [@Leboeuf01]. Next we write the general form of the scattering states within the approximation of a flowless condensate. For an atom incident from the N side with $\|\varepsilon\|<\delta\mu$ we have $$\psi_{N}\equiv\left\{ \begin{array} [c]{ll}\binom{1}{0}\left[ \chi_{N}^{+}+r_{n}\left( \chi_{N}^{+}\right) ^{\ast }\right] +\binom{0}{1}r_{a}\chi_{N}^{-}, & x\rightarrow-\infty\\ \binom{u_{0}}{v_{0}}t_{p}\chi_{S}+\binom{-v_{0}}{u_{0}}t_{e}~e^{-k_{S}^{-}~x}, & x\rightarrow\infty \end{array} \right. \label{eqnScatteringFromTheLeft}$$ where $\chi_{N}^{\pm}\equiv e^{ik_{N}^{\pm}x}/\sqrt{w_{N}^{\pm}}$, $\chi _{S}\equiv e^{i\mathrm{sgn}(\varepsilon)k_{S}^{+}x}/\sqrt{w_{S}}$ and $k_{S}^{-}\equiv\sqrt{2}[\sqrt{\varepsilon^{2}+\mu^{2}}+\mu]^{1/2}$ and $u_{0},v_{0}$ are such that $u_{0}^{2}+v_{0}^{2}=\sqrt{\varepsilon^{2}+\mu^{2}}/\|\varepsilon\|$ and $u_{0}/v_{0}=\sqrt{1+\varepsilon^{2}/\mu^{2}}+\varepsilon/\mu$. For an incoming phonon from the S side with energy $0<\varepsilon<\delta\mu$, we have $$\psi_{S}\equiv\left\{ \begin{array} [c]{ll}\binom{1}{0}t_{n}\left( \chi_{N}^{+}\right) ^{\ast}+\binom{0}{1}t_{a}\chi_{N}^{-}, & x\rightarrow-\infty,\\ \binom{u_{0}}{v_{0}}\left[ \chi_{S}^{\ast}+r_{p}\chi_{S}\right] +\binom{-v_{0}}{u_{0}}r_{e}~e^{-k_{S}^{-}~x}, & x\rightarrow\infty, \end{array} \right. \label{eqnScatteringFromTheRight}$$ The subindeces stand for normal ($a$), Andreev ($n$), phonon ($p$) and evanescent ($e$), while $r$ and $t$ represent reflection and transmission coefficients. Equations (\[eqnScatteringFromTheLeft\]) and (\[eqnScatteringFromTheRight\]) do not include the case $\varepsilon >\delta\mu$, which shows no Andreev processes, or that of an incoming channel with $\nu<0$, which does not seem realizable unless the beam comes from another Bose condensate. Conservation laws may be obtained from the generalized Wronskian $$W_{ij}\equiv u_{i}D_{x}^{\ast}u_{j}^{\ast}-u_{j}^{\ast}D_{x}u_{i}+v_{i}D_{x}v_{j}^{\ast}-v_{j}^{\ast}D_{x}^{\ast}v_{i}~, \label{eqnWronskian}$$ which is constant for any two stationary solutions of Eq. (\[eqnBdG\]) with same energy $\varepsilon$. When applied to $\psi_{N}$ and $\psi_{S}$, the constancy of (\[eqnWronskian\]) generates the following relations between the scattering amplitudes: $$\begin{array} [c]{c}\|r_{n}\|^{2}-\|r_{a}\|^{2}+\mathrm{sgn}(\varepsilon)~\|t_{p}\|^{2}=\|t_{n}\|^{2}-\|t_{a}\|^{2}+\|r_{p}\|^{2}=1\\ r_{n}~t_{n}^{\ast}+r_{p}^{\ast}~t_{p}-r_{a}~t_{a}^{\ast}=0 \end{array} \label{eqnWronskianEqualities}$$ The negative signs here can be traced back to the non-Hermitian character of the effective Hamiltonian in the bosonic BdG equations (\[eqnBdG\]). This represents a major difference with respect to the fermionic Bogoliubov problem and is a source of difficulties in the interpretation of the particle-hole transformation for bosons. If we assume a step-like profile for the condensate density, the scattering amplitudes can be found analytically. Here we note some important trends in the low transparency limit ($Z/\sqrt{\mu}\gg1$): The Andreev reflection amplitude ${\|{{r_{a}}}\|}$ scales like $Z^{-2}$, which indicates that, like for superconductors [@Blonder82], an Andreev reflection process involves the transmission of two atoms. This reflects the pairing correlations which exist in the quantum depletion cloud of the condensate and which are expressed through the structure of the Bogoliubov transformation and the BdG equations [@Leggett01]. Analogously, the scaling $\|t_{a}\|,\|t_{p}\|\sim Z^{-1}$ suggests that for large $Z$ the corresponding scattering processes involve the transmission of one atom. More information on conservation laws can be obtained from the continuity equation $$\partial_{t}\|u\|^{2}+\partial_{x}j_{u}=\partial_{t}\|v\|^{2}+\partial _{x}j_{v^{\ast}}=2~n_{0}(x)g(x)\operatorname{Im}(uv^{\ast})~, \label{eqnContinuity}$$ where $j_{s}\equiv\operatorname{Im}(s^{\ast}D_{x}s)$, which tells us simultaneously about the lack of conservation of the atom mass current (density $\|u\|^{2}+\|v\|^{2}$, current $j_{u}+j_{v^{\ast}}$) and the conservation of the quasiparticle current (density $\|u\|^{2}-\|v\|^{2}$, current $j_{u}-j_{v^{\ast}}$), the latter being also expressed by the constancy of the Wronskian (\[eqnWronskian\]). Both mass density and current are defined as their total values in the presence of the excitation of wave function $\binom{u}{v}$ minus the corresponding values of the Bogoliubov vacuum, which includes the zero-temperature depletion cloud [@Fetter72] and thus a fluctuating atom flux. We note that, while the $u$ component carries the current in the usual way, the $v$ does it in the opposite one. Another important feature is that, unlike for its fermionic counterpart [@Blonder82], the mass current has the same sign as the group velocity in all channels (the mass density is always positive). As in the case of superconductivity [@Sanchez97; @Blonder82], we interpret the non-conservation of atom number as the exchange of atoms with the Bogoliubov vacuum. The net balance of extracted atoms per unit time can be calculated by substracting, from the current at a point deep in the superfluid region, the current at any point in the normal region, $\Delta I\equiv j(x_{+})-j(x_{-})$. We obtain $$\Delta I=\|r_{n}\|^{2}+\|r_{a}\|^{2}+\frac{\|t_{p}\|^{2}}{u_{0}^{2}+v_{0}^{2}}-1\simeq\frac{-2~\mathrm{sgn}(\varepsilon)~k_{N}^{+}{}~k_{S}^{+}}{1+\varepsilon^{2}/\mu^{2}}\frac{1}{Z^{2}}~. \label{eqnCurrentBalance}$$ For the second equality, which gives the current to leading order in $Z^{-1}$, we have used (\[eqnWronskianEqualities\]) and the exact results for the scattering amplitudes [@details]. Thus, for an atom incident from the N side, the scattering process which follows is such that atoms are lost to the Bogoliubov vacuum if $\varepsilon>0$, while they are extracted from that vacuum if $\varepsilon<0$. From the continuity equation (\[eqnContinuity\]) it is clear that the transfer of atoms between the condensate and the quasiparticle field takes place exclusively on the S side, since $g(x)=0$ in N. \[ptb\] [Andreev3D.eps]{} The generalization of the above discussion to the case of a three-dimensional planar or two-dimensional linear interface yields qualitative differences with respect to the superconducting case. Invoking the conservation of energy and parallel momentum, we find that for $\varepsilon\in(-\delta\mu,\delta\mu)$ the beam angles (see Fig. \[grAndreev3D\]) satisfy the relation$$(1-\varepsilon/\delta\mu)\tan^{2}\alpha_{i}-(1+\varepsilon/\delta\mu)\tan ^{2}\alpha_{a}=2~\varepsilon/\delta\mu~. \label{angles}$$ This result leads to two conclusions: (a) If $0<\varepsilon<\delta\mu$, no Andreev reflection occurs for $\tan\alpha_{i}<\sqrt{\varepsilon/2(\delta \mu-\varepsilon)}$ and one has $\alpha_{a}<\alpha_{i}$. (b) If $-\delta \mu<\varepsilon<0$ Andreev reflection is possible, but only satisfying $\alpha_{a}>\alpha_{i}$ and $\tan\alpha_{a}\geq\sqrt{-\varepsilon/2(\delta \mu+\varepsilon)}$. The observation of bosonic Andreev reflection poses a challenge. A tunable coupling constant can be achieved through the handling of Feshbach resonances [@Fattori08] or, in the 1D limit, by increasing the width of the confining channel. Such long channels can form in the vicinity of a planar semiconductor chip. A potential barrier at the interface can be introduced with a properly focused, blue-detuned laser. The crucial measurement of escaping velocities may rely on precise Bragg spectroscopy or on the interference between incoming and outgoing beams. The 2D/3D case presents greater potential flexibility because anomalous reflection can occur in a range of different directions. Confinement by a linear interface in 2D could be achieved with a laser beam. For instance, $^{87}$Rb atoms may be confined in a waveguide with transverse trapping frequency of 500 Hz and experience a delta-barrier of $Z=5\sqrt{\mu}$ if they are exposed to a narrow Gaussian laser of 1 $\mu$m waist, blue-detuned by 10 nm to the 5$^{2}$S$_{1/2}$-5$^{2}$P$_{1/2}$ transition [@Metcalf99] (780.24 nm, linewidth 12$\pi$ Hz, saturation intensity 1.64 mW cm$^{-2}$). With a step barrier such that $\delta\mu=0.5\mu $, bulk superfluid linear densities in the range 0.2–20 $\mu$m$^{-1}$ place the system in the low-density mean-field region, where the present estimate applies. For that density range we obtain 1D coherence-length 7–0.7 $\mu$m, laser intensity 80–800 mW cm$^{-2}$, and 10–1000 $\mu$m s$^{-1}$ for the Andreev velocity scale ($\sim2\sqrt{\delta\mu}$), close to the superfluid sound speed of 70–700 $\mu$m s$^{-1}$. For these parameters the emitted condensate linear density is 0.04 times the bulk S density, while the corresponding condensate escape velocity is the same as the speed of sound in S, the equality being due to our choice of $\delta\mu$. We acknowledge valuable discussions with J. Brand, N. Davidson, E. Demler, S. Foelling, A. J. Leggett, C. Lobo, H. Michinel, E. Muller, and G. Shlyapnikov. This work has been supported by the Ramón Areces Foundation and by MEC (Spain) through Grant FIS2007-65723. The authors thank the hospitality and support of the Institut Henri Poincare - Centre Emile Borel, where part of this work was done. [99]{} A. F. Andreev, Sov. Phys. JETP 19, 1228 (1964). P. Samuelsson et al., Phys. Rev. Lett. 91, 157002 (2003); New J. Phys. 7, 176 (2005); E. Prada and F. Sols, Eur. Phys. J. B 40, 379 (2004). S. I. Bozhko et al., JETP Lett. 36, 153 (1982); P. A. M. Benistant et al., Phys. Rev. Lett. 51, 817 (1983); G. E. Blonder and M. Tinkham, Phys. Rev. B 27, 112 (1983). M. P. Enrico et al., Phys. Rev. Lett. 70, 1846 (1993). B. Van Schaeybroeck and A. Lazarides, Phys. Rev. Lett. 98, 170402 (2007). A. J. Daley et al., Phys. Rev. Lett. 100, 110404 (2008). A. Tokuno et al., Phys. Rev. Lett. 100, 140402 (2008). A. L. Fetter, Ann. Phys. (New York) 70, 67 (1972). A. Wynveen et al., Phys. Rev. A 62, 023602 (2000). U. V. Poulsen and K. Mølmer, Phys. Rev. A 67, 013610 (2003). P. Leboeuf and N. Pavloff, Phys. Rev. A 64, 033602 (2001). F. Sols and J. Ferrer, Phys. Rev. B 49, 15913 (1994). J. Sánchez-Cañizares and F. Sols, Phys. Rev. B 55, 531 (1997). C. J. Pethick and H. Smith, Bose-Einstein Condensation in Dilute Gases (Cambridge University Press, Cambridge, 2002). F. Dalfovo et al., Rev. Mod. Phys. 71, 463 (1999). A. J. Leggett, Rev. Mod. Phys. 73, 307 (2001). Technical details will be given elsewhere. G. E. Blonder et al., Phys. Rev. B 25, 4515 (1982). M. Fattori et al., Phys. Rev. Lett. 100, 080405 (2008); M. Gustavsson et al., ibid. 100, 080404 (2008). H. J. Metcalf and P. van der Straten, Laser Cooling and Trapping (Springer-Verlag, New York, 1999).	{ "pile_set_name": "ArXiv" }
--- abstract: \| We show that 1. [Rothberger bounded subgroups of $\sigma$-compact groups are characterized by Ramseyan partition relations. (Corollary 4)]{} 2. [For each uncountable cardinal $\kappa$ there is a ${\sf T}_0$ topological group of cardinality $\kappa$ such that ONE has a winning strategy in the point-open game on the group and the group is not a subspace of any $\sigma$-compact space. (Theorem \[rothbwonbytwo\])]{} 3. [For each uncountable cardinal $\kappa$ there is a ${\sf T}_0$ topological group of cardinality $\kappa$ such that ONE has a winning strategy in the point-open game on the group and the group is $\sigma$-compact. (Corollary \[rothbdedtorothbomega\])]{} address: \| Department of Mathematics\ Boise State University\ Boise, Idaho 83725 author: - Marion Scheepers title: Rothberger bounded groups and Ramsey theory --- Overview ======== In topological groups boundedness properties are counterparts for covering properties in general topological spaces: Guran’s notion of $\aleph_0$-boundedness is a counterpart of the Lindelöf covering property -[@Guran], while Okunev’s notion of o-boundedness (named Menger-boundedness by Kočinac, who introduced this notion independently) and Tkachenko’s corresponding property of strict o-boundedness are counterparts of $\sigma$-compactness -[@CH]. In this paper we consider a boundedness property which approximates Borel’s metric notion of strong measure zero. This boundedness property was introduced in unpublished work by Galvin, was later independently introduced by Kočinac under the name of Rothberger boundedness, and initially investigated in [@coc11]. In [@coc11] it was shown that a subgroup (or subset) of a metrizable topological group is Rothberger bounded if, and only if, it is strong measure zero in all left invariant metrics of the group. In [@coc11] we also extended some of the characterizations of strong measure zero from [@smzpow] to Rothberger boundedness, but we did not have techniques to also extend the Ramsey-theoretic characterization to this context. Now, in Corollary 4 of this paper, we obtain the Ramsey-theoretic characterization. The point-open game introduced in [@Galvin] by Galvin is closely related to the notion of strong measure zero. Galvin proved that for Lindelöf spaces in which each point is an intersection of countably many open sets player ONE of the point-open game has a winning strategy if, and only if, the space is countable. Apparently few, if any, uncountable examples of Hausdorff spaces where ONE has a winning strategy in the point-open game, have been pointed out in the literature. Another objective of this paper is to show that classical work of Comfort [@Comfort] and also classical work of Corson [@Corson] provide a wide range of topological groups that are such examples. Our analysis of these examples strengthen some results of Hernandez [@CH] on the theory of $\aleph_0$-bounded groups. Some terminology and notation ============================= For collections $\mathcal{A}$ and $\mathcal{B}$ the symbol ${{\sf S}_1}(\mathcal{A},\mathcal{B})$ denotes the statement that > For each sequence $(A_n:n<\omega)$ of elements of $\mathcal{A}$ there is a sequence $(b_n:n<\omega)$ such that for each $n$ $b_n\in A_n$, and $\{b_n:n<\omega\}\in\mathcal{B}$. Let $Y$ be a topological space. Then ${\mathcal{O}}$ denotes the collection of all open covers of $Y$. If the selection principle ${{\sf S}_1}({\mathcal{O}},{\mathcal{O}})$ holds for $Y$ we say $Y$ is a Rothberger space (or has the Rothberger property). This property is named after F. Rothberger who introduced it in his study of Borel’s notion of strong measure zero - [@rothberger38]. A subspace $X$ of the metric space $(Y,d)$ is a strong measure zero subspace if there is for each sequence $(\epsilon_n:n<\omega)$ of positive real numbers a partition $X=\cup_{n<\omega}X_n$ such that for each $n$ the $d$-diameter of $X_n$ is less than $\epsilon_n$. In [@smzpow] we characterized the strong measure zero subspaces of $\sigma$-compact metric spaces in terms of Ramseyan partition relations. For a topological group $(G,)$ and a neighborhood $U$ of its identity element we define: ${\mathcal{O}}(U)=\{xU:x\in G\}$. Then ${\mathcal{O}}(U)$ is an open cover of $G$. ${\mathcal{O}}_{nbd}$ denotes the family of all open covers of $G$ of the form ${\mathcal{O}}(U)$. A topological group is said to be Rothberger bounded if it has the property ${{\sf S}_1}({\mathcal{O}}_{nbd},{\mathcal{O}})$. For $X$ be a subset of the topological space $G$ let ${\mathcal{O}}_X$ denote the covers of $X$ by sets open in $G$. Then $X$ is said to be Rothberger bounded in $G$ if ${{\sf S}_1}({\mathcal{O}}_{nbd},{\mathcal{O}}_X)$ holds. In topological groups the property ${{\sf S}_1}({\mathcal{O}},{\mathcal{O}})$ is generally stronger than ${{\sf S}_1}({\mathcal{O}}_{nbd},{\mathcal{O}})$. The symbol ${{\sf G}_1}(\mathcal{A},\mathcal{B})$ denotes the following game of length $\omega$: Players ONE and TWO play an inning per $n<\omega$. In inning $n$ ONE first selects a member $O_n\in\mathcal{A}$, and then TWO responds by choosing a $T_n\in O_n$. A play $(O_0,\, T_0,\, \cdots,\, O_n,\, T_n,\,\cdots)$ is won by TWO if $\{T_n:n<\omega\}\in\mathcal{B}$; else, it is won by ONE. F. Galvin [@Galvin] introduced the game ${{\sf G}_1}({\mathcal{O}},{\mathcal{O}})$ and proved that it is related as follows to the well-known point-open game[^1]: ONE has a winning strategy in the point-open game if, and only if, TWO has a winning strategy in ${{\sf G}_1}({\mathcal{O}},{\mathcal{O}})$. TWO has a winning strategy in the point-open game if, and only if, ONE has a winning strategy in ${{\sf G}_1}({\mathcal{O}},{\mathcal{O}})$. A Ramseyan characterization of some Rothberger bounded groups. ============================================================== Let $\mathcal{A}$ and $\mathcal{B}$ be families of sets and let $n$ and $k$ be positive integers. The symbol $$\mathcal{A}\rightarrow(\mathcal{B})^n_k$$ denotes the statement that for each $A\in\mathcal{A}$ and for each function $f:\lbrack A\rbrack^n\rightarrow\{1,\,\cdots,\, k\}$ there is a $B\subseteq A$ and an $i\in\{1,\,\cdots,\,k\}$ such that $B\in\mathcal{B}$ and $f$ is constant of value $i$ on $\lbrack B\rbrack^n$. An open cover $\mathcal{U}$ of $Y$ is said to be an $\omega$-cover if $Y\not\in\mathcal{U}$, but for each finite set $F\subseteq Y$ there is a $U\in\mathcal{U}$ with $F\subseteq U$. $\Omega$ denotes the family of all open $\omega$-covers of $Y$. Theorem 9 of [@smzpow] gives the following characterization of strong measure zero subsets of $\sigma$-compact metric spaces: \[smzpowTh9\] For $X$ a subspace of a $\sigma$-compact metric space $Y$ the following are equivalent: 1. [$Y$ has the property ${{\sf S}_1}({\mathcal{O}},\mathcal{O}_X)$.]{} 2. [$X$ has strong measure zero (in all equivalent metrics on $Y$).]{} 3. [ONE has no winning strategy in the game ${{\sf G}_1}({\mathcal{O}},\mathcal{O}_X)$.]{} 4. [For each positive integer $k$, $\Omega\rightarrow({\mathcal{O}}_X)^2_k$.]{} Failure of Lebesgue’s Covering Lemma (which holds for compact metrizable spaces) was the main obstacle towards extending Theorem \[smzpowTh9\] beyond metrizable $\sigma$-compact spaces. We found two non-metric situations in which an appropriate generalization of Lebesgue’s Covering Lemma holds. Here is the first of the two[^2]: \[lebesguecov\] Let $(G,)$ be a ${\sf T}_0$ topological group, let $\mathcal{U}$ be an open cover of $G$ and let $K$ be a nonempty compact subset of $G$. Then there is a neighborhood $N$ of the identity of $G$ such that for each $x\in K$ there is a $U\in\mathcal{U}$ with $xN\subseteq U$. Let $\mathcal{U}$ be an open cover of $G$. For each $x\in K$ choose a $U(x)\in\mathcal{U}$ with $x\in U(x)$. Then choose for each $x$ a neighborhood $N_x$ of $G$’s identity such that $x{N_x}^2\subseteq U(x)$. Now $\{xN_x:x\in K\}$ is an open cover of $K$. Since $K$ is compact this cover has a finite subset that covers $K$, say $\{x_1N_{x_1},\cdots,x_kN_{x_k}\}$. Define $N$ by $$N=N_{x_1}\cap\cdots\cap N_{x_k}.$$ Then $N$ is as required. For consider any $k\in K$. Choose $i$ so that $k\in x_iN_{x_i}$. Now consider any $q\in kN$. Fix $a\in N$ with $q = ka$. Then we have $q\in x_iN_{x_i}a \subseteq x_iN^2_{x_i} \subseteq U(x_i)$. Since $q$ was an arbitrary element of $kN$ we find that $kN\subseteq U(x_i)$. $\Box$ Here is the version of Theorem \[smzpowTh9\] for $\sigma$-compact topological groups: \[rboundedinsigmacompact\] Let $(G,)$ be a $\sigma$-compact topological group and let $X$ be a subset of $G$. The following are equivalent: 1. [$(G,)$ satisfies ${{\sf S}_1}({\mathcal{O}},{\mathcal{O}}_X)$.]{} 2. [$(G,)$ satisfies ${{\sf S}_1}({\mathcal{O}}_{nbd},{\mathcal{O}}_X)$.]{} 3. [$X$ is a Rothberger bounded subset of $(G,)$.]{} 4. [ONE has no winning strategy in the game ${{\sf G}_1}({\mathcal{O}},{\mathcal{O}}_X)$.]{} 5. [For each positive integer $k$, $\Omega\rightarrow ({\mathcal{O}}_X)^2_k$]{} The implications $(2)\Rightarrow (3)$, $(4)\Rightarrow (5)$ and $(5)\Rightarrow(1)$ follow the ideas of the proof of Theorem 9 of [@smzpow], with some standard modifications. We prove $(3)\Rightarrow(4)$ here since its proof uses the perhaps new Lemma \[lebesguecov\]. [ ]{} Let $F$ be a strategy for player ONE in the game ${{\sf G}_1}({\mathcal{O}},{\mathcal{O}}_X)$ on $G$. Since $G$ is $\sigma$-compact, write $G=\bigcup_{n<\omega}G_n$ where for each $n$ we have $id_G\in G_n\subseteq G_{n+1}$ and $G_n$ is compact. For each $n$, $X_n = X\cap G_n$ is a Rothberger bounded subset of $G$. To defeat ONE’s strategy TWO will concentrate attention on specific $X_n$’s in specific innings. To this end, partition $\omega$ into infinitely many infinite subsets $S_n$. For innings numbered by members of $S_n$ TWO will focus on $X_n$. We now use ONE’s strategy $F$ to recursively define a sequence $(N_k:k<\omega)$ and an array $(\mathcal{U}(T_0,\cdots,T_k):k<\omega)$ where 1. [For each $k$, $N_k$ is a symmetric[^3] neighborhood of the identity of $G$;]{} 2. [With $n_0$ such that $0\in S_{n_0}$, $\mathcal{U}(\emptyset)$ is a finite subset of $F(\emptyset)$ (ONE’s first move) which covers $G_0$, and $N_0$ is a neighborhood of the identity of $G$ such that for each $x\in G_0$ there is a $V\in\mathcal{U}(\emptyset)$ with $x(N_0N_0)\subseteq V$.]{} 3. [For each $(T_0,\cdots,T_k)$ such that $T_0\in F(\emptyset)$, $T_1\in F(T_0)$, $\cdots$ and $T_k\in F(T_0,\cdots,T_{k-1})$ and for $n_{k+1}$ such that $k+1\in S_{n_{k+1}}$ we have $\mathcal{U}(T_0,\cdots,T_k)$ a finite subset of $F(T_0,\cdots,T_k)$ that covers $G_{n_{k+1}}$. Note that there are only finitely many such $(T_0,\cdots,T_k)$. $N_{k+1}$ is a neighborhood of the identity of $G$ such that for each such sequence $(T_0,\cdots,T_k)$ and for each $x\in G_{n_{k+1}}$ there is a $U\in\mathcal{U}(T_0,\cdots,T_{k})$ with $xN_{k+1}N_{k+1}\subseteq U$.]{} With this data available, construct a play against $F$ won by TWO as follows: Fix an $m<\omega$. Since $X_m$ is Rothberger bounded select for each $k\in S_m$ an $x_k\in G$ such that $(x_kN_k:k\in S_m)$ covers $X_m$. We may assume each $x_k$ is in $X_m$ - for suppose an $x_k$ is not in $X_m$. If $x_kN_k \cap X_k = \emptyset$, we may with impunity replace this $x_k$ by one from $X_k$. However, if $x_kN_k\cap X_k\neq\emptyset$, then let $y$ be an element of this intersection. We claim that $x_kN_k\cap X_k\subseteq yN_k^2$. For let $z\in x_kN_k\cap X_k$ be given. Choose $v$ in $N_k$ with $z = x_kv$. Since we also have $y\in x_kN_k$, choose $w\in N_k$ with $y=x_kw$. Now $x_k = yw^{-1} \in N_kN_k$ and so $z = y(w^{-1}v)\in yN_k^2$. Thus, we may choose for each $k\in S_m$ an $x_k\in X_k$ such that $(x_kU_k^2:k\in S_m)$ covers $X_k$. Finally, recursively choose a sequence $(T_k:k<\omega)$ as follows: Choose $T_0\in \mathcal{U}(\emptyset)$ with $x_0N_0^2 \subseteq T_0$. With $T_0,\, \cdots,\, T_m$ chosen, choose $T_{m+1}\in\mathcal{U}(T_0,\cdots,T_m)$ with $x_{m+1}N_{m+1}^2\subseteq T_{m+1}$. Then the sequence $$F(\emptyset),\, T_0,\, F(T_0),\, \cdots,\, T_k,\, F(T_0,\cdots,T_k),\, T_{k+1},\, \cdots$$ is an $F$-play lost by ONE. In the proof of (5) implies (1) we use that ${{\sf S}_1}({\mathcal{O}},{\mathcal{O}}_X)$ is equivalent to ${{\sf S}_1}(\Omega,{\mathcal{O}}_X)$. $\Box$ Note that $(4)$ also implies that ONE has no winning strategy in the game ${{\sf G}_1}({\mathcal{O}}_{nbd},{\mathcal{O}}_X)$ on $(G,)$, which in turn implies $(2)$. In particular we have the following Corollary: \[rothbbdedgp\] Let $(H,)$ be a subgroup of a $\sigma$-compact group $(G,)$. The following are equivalent: 1. [$(G,)$ satisfies ${{\sf S}_1}({\mathcal{O}},{\mathcal{O}}_H)$.]{} 2. [$(H,)$ is Rothberger bounded.]{} 3. [ONE has no winning strategy in the game ${{\sf G}_1}({\mathcal{O}},{\mathcal{O}}_H)$.]{} 4. [On $H$ ONE has no winning strategy in the game ${{\sf G}_1}({\mathcal{O}}_{nbd},{\mathcal{O}})$.]{} 5. [For each positive integer $k$, $(G,)$ satisfies $\Omega\rightarrow ({\mathcal{O}}_H)^2_k$]{} \(1) implies that $(G,)$ satisfies ${{\sf S}_1}({\mathcal{O}}_{nbd},{\mathcal{O}}_H)$. But then by Theorem 13 of [@coc11], $(H,)$ satisfies ${{\sf S}_1}({\mathcal{O}}_{nbd},{\mathcal{O}})$ in the relative topology, proving (2). Since (2) states that $H$ is a Rothberger bounded subset of $G$, Theorem \[rboundedinsigmacompact\] gives the implication from (2) to (3) and from (3) to (5). Also, (3) implies (4) which implies (2). The proof that (5) implies (1) uses the fact that ${{\sf S}_1}({\mathcal{O}},{\mathcal{O}}_H)$ is equivalent to ${{\sf S}_1}(\Omega,{\mathcal{O}}_H)$. $\Box$\ Corollary \[rothbbdedgp\] improves Theorem 22 of [@coc11] in that it does not require the group $(G,)$ to be metrizable. One might wonder how widely applicable Corollary \[rothbbdedgp\] really is. We shall address this in the next two sections by showing that: - [For each infinite cardinal number $\kappa$ there is a ${\sf T}_0$ Rothberger bounded group $(G,)$ of cardinality $\kappa$ which is not a subgroup of any $\sigma$-compact group, and TWO has a winning strategy in the game ${{\sf G}_1}({\mathcal{O}}_{nbd},{\mathcal{O}})$ on $(G,)$.]{} - [For each infinite cardinal number $\kappa$ there is a ${\sf T}_0$ $\sigma$-compact Rothberger bounded group of cardinality $\kappa$ for which TWO has a winning strategy in the game ${{\sf G}_1}({\mathcal{O}},{\mathcal{O}})$.]{} Rothberger groups not embedding into any $\sigma$-compact space. ================================================================ Not every ${\sf T}_0$ Rothberger bounded group is a subgroup of a $\sigma$-compact group, as illustrated by the following example of Comfort and Ross ([@CR], Example 3.2). We precede the example with a few general remarks about $P$-spaces. A topological space is said to be a $P$-space if each ${\sf G}_{\delta}$ set is open. Evidently, every subspace of a $P$-space is a $P$-space. Every countably infinite subspace of a ${\sf T}_2$ $P$-space is closed and discrete. It follows that a compact $P$-space is finite, and thus a $\sigma$-compact $P$-space is countable. Thus, no uncountable Lindelöf $P$-space is a subspace of a $\sigma$-compact ${\sf T}_2$-space. If a topological group $(G,)$ is a Lindelöf $P$-space then it is Rothberger bounded in a strong sense: Let $(U_n:n<\omega)$ be a sequence of neighborhoods for the identity. Then $U = \cap_{n<\omega}U_n$ is a neighborhood for the identity. Since the group is $\aleph_0$ bounded fix a sequence $(x_n:n<\omega)$ of elements of the group such that $x_nU,\, n<\omega$ covers the group. Then the sequence $(x_nU_n:n<\omega)$ witnesses that the group is Rothberger bounded. We now define the example: The underlying set of the group $G$ is $$G:=\{f\in\,^{\omega_1}2:\vert\{\alpha: f(\alpha)\neq 0\}\vert<\aleph_0\};$$ Endow $G$ with the [G]{}$_{\delta}$ topology. [@CR] shows that $(G,\oplus)$ is a Lindelöf $P$-group (and thus ${\sf T}_4$). Thus $(G,\oplus)$ is an uncountable Rothberger bounded group that is not contained in a $\sigma$-compact group. Theorem 2.3 of Comfort in [@Comfort] implies the following generalization of the above example: \[Comfortcountable\] Let $(G_i,_i)$, $i\in I$, be a family of countable topological groups. Endow the product $\prod_{i\in I}G_i$ with the countable box topology. Then the subgroup $$G:=\{f\in\prod_{i\in I}G_i: \vert\{j\in I:f(j)\neq {\sf id}_j\}\vert<\aleph_0\}$$ is a Lindelöf $P$-group. In particular, for each uncountable cardinal number $\kappa$ there is a ${\sf T}_0$ Lindelöf $P$ group of cardinality $\kappa$. One can prove an analogue of Theorem \[rboundedinsigmacompact\] also for Lindelöf $P$-groups, because: Galvin proved a result that implies that if a space is a Lindelöf P-space then it is a Rothberger space - see the Lemma in Section 2 of [@GN]. Thus, Lindelöf $P$ groups satisfy the stronger selection principle ${{\sf S}_1}({\mathcal{O}},{\mathcal{O}})$. But the following are equivalent: 1. [Topological space $X$ satisfies ${{\sf S}_1}({\mathcal{O}},{\mathcal{O}})$;]{} 2. [ONE has no winning strategy in the game ${{\sf G}_1}({\mathcal{O}},{\mathcal{O}})$ played on $X$;]{} 3. [For each positive integer $k$, $\Omega\rightarrow({\mathcal{O}})^2_k$ holds for $X$.]{} The equivalence of (1) and (2) is due to Pawlikowski [@JP], and the equivalence with (3) was proved in [@msrothbramsey]. Using ideas of [@CH] we now show: \[twowinsrothberger\] For each uncountable cardinal $\kappa$ there is a ${\sf T}_0$ Rothberger bounded group of cardinality $\kappa$ such that TWO has a winning strategy in ${{\sf G}_1}({\mathcal{O}}_{nbd},{\mathcal{O}})$. In the proof we will make use of the following elementary game of $\omega$ innings played on an infinite set $S$: In inning $n$ player ONE chooses a countable subset $W_n$ of $S$ and TWO responds by choosing a point $b_n\in W_n$. ONE must further obey the rule that for each $n$, $W_n\subseteq W_{n+1}$. A play $(W_0,\, b_0,\, \cdots,\, W_n,\, b_n,\,\cdots)$ is won by TWO if for each $x\in\bigcup_{n<\omega}W_n$ there are infinitely may $n$ with $b_n = x$. A standard argument shows that TWO has a winning perfect information strategy in this game. Call this game the “countable - one game". Let $\kappa$ be an uncountable cardinal number. Let $(G_{\alpha}:\alpha<\kappa)$ be a sequence of discrete countable groups and define $G$ to be the direct product $$\prod_{\alpha<\kappa}G_{\alpha}$$ endowed with the countable box topology. Then the subset $G^* = \{f\in G:\vert \{\alpha: f(\alpha)\neq {\sf id}_{\alpha}\}\vert<\aleph_0\}$ endowed with the relative topology is by Theorem \[Comfortcountable\] a Lindelöf ${\sf P}$-group. For a countable set $B\subset\kappa$, let $\Pi_B$ denote the projection of $G$ onto $\prod_{\alpha\in B}G_{\alpha}$. Then the set $$U_B = G^* \bigcap \Pi^{\leftarrow}_B\lbrack\{{\sf id}_B\}\rbrack$$ is a basic neighborhood of the identity element of $G^$. Also $\mathcal{D} = \{U_B:B\in\lbrack\kappa\rbrack^{\leq\aleph_0}\}$ is a neighborhood basis for the identity element of $G^$, and each $U_B$ is a subgroup of the group $G^$. Since $G^$ is a Lindelöf ${\sf P}$-group, each open cover of the form $\mathcal{O}(U_B)$ has a countable subcover, and this means that the subgroup $U_B$ has countably many distinct left cosets in $G^$. Now we show that TWO has a winning strategy in the game ${{\sf G}_1}({{\mathcal{O}}_{nbd}},{\mathcal{O}})$ played on $G^$. Since $\mathcal{D}$ is a neighborhood basis of the identity element of $G^$, we may assume that for each $n<\omega$ ONE’s $n$-th move is of the form ${\mathcal{O}}(U_{B_n})$, $B_n$ a countable subset of $\kappa$. And since TWO may replace ONE’s move ${\mathcal{O}}(U_B)$ with a move ${\mathcal{O}}(U_C)\subset {\mathcal{O}}(U_B)$ and respond to the replacement move instead, we may further assume that ONE’s moves are such that for each $n$, $B_n\subseteq B_{n+1}$, that is, $U_{B_{n+1}}\subseteq U_{B_n}$. Also, for each move ${\mathcal{O}}(U_{B_n})$ by ONE, TWO chooses a countable set $A_n\subseteq G^$ such that $\{x* U_{B_n}: x\in A_n\}$ is the set of distinct left cosets of $U_{B_n}$ in $G^$. Since TWO has perfect information and for each $n$ $U_{B_{n+1}}\subseteq U_{B_n}$, TWO may select the sets $A_n$ such that for each $n$ we have $A_n\subseteq A_{n+1}$. We may assume for each $n$ that for each $x\in A_n$, if $x(\alpha)\neq {\sf id}_{G_{\alpha}}$ then $\alpha\in B_n$. Now let $F$ be a winning perfect information strategy for TWO in the countable-1 game on $G^$. We define a strategy $\sigma$ for TWO in ${{\sf G}_1}({{\mathcal{O}}_{nbd}},{\mathcal{O}})$ as follows: Given ONE’s move ${\mathcal{O}}(U_{B_0})$ in ${{\sf G}_1}({{\mathcal{O}}_{nbd}},{\mathcal{O}})$ TWO first fixes $A_0$ as above, considered as a move of ONE of the countable-1 game. Then in that game TWO moves $x_0 = F(A_0)$. Then TWO responds in ${{\sf G}_1}({{\mathcal{O}}_{nbd}},{\mathcal{O}})$ with $\sigma({\mathcal{O}}(U_{B_0}) = x_0U_{B_0}$. In the next inning of ${{\sf G}_1}({{\mathcal{O}}_{nbd}},{\mathcal{O}})$ ONE moves ${\mathcal{O}}(U_{B_1})$. TWO first fixes the countble set $A_1$ as above and consider it as a move of ONE in the countable - 1 game on $G^$. In that game TWO moves $x_1 = F(A_0,A_1)$. Then TWO responds in ${{\sf G}_1}({{\mathcal{O}}_{nbd}},{\mathcal{O}})$ with $\sigma({\mathcal{O}}(U_{B_0}),{\mathcal{O}}(U_{B_1})) = x_1U_{B_1}$, and so on, as depicted in the following diagram. [c\|l]{}\ \ ONE & TWO\ ${\mathcal{O}}(U_{B_0})$ &\ &\ & $x_0U_{B_0}$\ ${\mathcal{O}}(U_{B_1})$ &\ &\ & $x_1U_{B_1} $\ $\vdots$ & $\vdots$\ [c\|l]{}\ \ ONE & TWO\ &\ $A_0$ & $x_0 = F(A_0)$\ &\ &\ $A_1$ & $x_1=F(A_0,A_1)$\ &\ $\vdots$ & $\vdots$\ We shall now see that $\sigma$ is a winning strategy for TWO in the game ${{\sf G}_1}({{\mathcal{O}}_{nbd}},{\mathcal{O}})$. Thus, consider a $\sigma$-play $${\mathcal{O}}(U_{B_0}),\, x_0U_{B_0},\,{\mathcal{O}}(U_{B_1}),\, x_1U_{B_1},\,\cdots,\, {\mathcal{O}}(U_{B_n}),\, x_nU_{B_n},\,\cdots$$ Let $x\in G^$ be given. We must show that $x\in\bigcup_{n<\omega}x_nU_{B_n}$. Put $B=\cup_{n<\omega}B_n$. Then evidently $U_B\subseteq U_{B_n}$ holds for each $n$. Also $L = xU_B$ is left coset of $U_B$ in $G^$. For each $n<\omega$ there is a $y\in A_n$ with $L\subseteq yU_{B_n}$. To see this, fix $n<\omega$. Since $U_B\subseteq U_{B_n}$ we have $xU_B\subseteq xU_{B_n}$. But $xU_{B_n}$ is a left coset of $U_{B_n}$ in $G^$, and so by the choice of $A_n$ there is a $y\in A_n$ with $xU_{B_n} = yU_{B_n}$. Thus, choose for each $n$ a $y_n\in A_n$ such that $L\subseteq y_nU_{B_n}$. For each $n<\omega$ we have $y_{n+1}U_{B_{n+1}}\subseteq y_nU_{B_n}$. For suppose on the contrary that $y_{n+1}U_{B_{n+1}}\not\subseteq y_nU_{B_n}$. Then we also have $y_{n+1}U_{B_n}\not\subseteq y_nU_{B_n}$, so that these are distinct left cosets of $U_{B_n}$ in $G^$, and so are disjoint. But this contradicts the fact that $\emptyset\neq L\subseteq y_{n+1}U_{B_{n+1}}\cap y_nU_{B_{n}}$. Towards proving the next claim first note that if $xU_{B_n}\subseteq yU_{B_n}$ then $(\forall \alpha\in B_n)(x(\alpha)=y(\alpha))$. For each $n<\omega$ we have $support(y_n)\subseteq support(y_{n+1})$. For by Claim 2 we find a $u\in U_{B_n}$ such that $y_{n+1} = y_nu$. Now $support(y_n)\cap support(u) = \emptyset$, and so $support(y_{n+1}) = support(y_n)\cup support(u)\supseteq support(y_n)$. For each $n<\omega$ we have $support(y_n)\subseteq support(x)$. Since $x\in L\subseteq y_nU_{B_n}$ it follows that $y_n^{-1}x\in U_{B_n}$, and so for each $\alpha\in B_n$ we have $x(\alpha) = y_n(\alpha)$. Since $support(y_n)\subseteq B_n$, Claim 4 follows. Since $x$ is in $G^$ it has finite support. Claims 3 and 4 imply that there is a $k$, from now on fixed, such that for all $n\geq k$ we have $support(y_n) = support(y_k)$. It follows that for all $n\ge k$, $y_n = y_k$. But then, for all $n\ge k$, $x\in y_kU_{B_n}$, which implies that $x\in y_kU_B$. But for infinitely many $n$ we have $x_n = y_k$,a nd so for such an $n$ larger than $k$, $x\in x_nU_{B_n}$. This completes the proof that $\sigma$ is a winning strategy for TWO. $\Box$ From this we now derive that TWO in fact has a winning strategy in the game ${{\sf G}_1}({\mathcal{O}},{\mathcal{O}})$, typically a harder game for player TWO. Towards this we need another generalization of the Lebesgue Covering Lemma, this time for Lindelöf P-groups: \[pgpLebesgue\] Let $(G,)$ be a Lindelöf P-group. Then there is for each open cover $\mathcal{U}$ of $G$ a neighborhood $N$ of the identity of $G$ such that for each $x\in G$ there is a $U\in\mathcal{U}$ such that $xN\subseteq U$. Let $\mathcal{U}$ be an open cover of $G$. For each $x\in G$ choose a neighborhood $U_x$ of $G$’s identity such that $U_x$ is an open subgroup of $G$, and there is a $U\in\mathcal{U}$ with $xU_x\subseteq U$. Since $G$ is Lindelöf we find $x_n,\, n<\omega$ such that $\mathcal{F} = \{x_nU_{x_n}:n<\omega\}$ is an open cover of $G$ and refines $\mathcal{U}$. Now since $G$ is a $P$-space, choose an open neighborhood $N$ of the identity such that $N$ is a subgroup of $G$, and $N\subseteq \bigcap_{n<\omega}U_{x_n}$. We claim that $N$ is as required. For consider any $x\in G$. Then $xN$ is a left coset of $N$ in $G$. We claim there is an $n$ with $xN\subseteq x_nU_{x_n}$. For if not, then for each $n$ we have $xN\not\subseteq x_nU_{x_n}$. But we have $xN\subseteq xU_{x_n}$, a left coset of $U_{x_n}$ in the group $G$. Thus $xU_{x_n}\neq x_nU_{x_n}$, and as $x_nU_{x_n}$ is also a left coset of $U_{x_n}$, we have $xN\cap x_nU_{x_n} = \emptyset$. But then the family $\{x_nU_{x_n}:n<\omega\}$ does not cover the subset $xN$ of $G$, contradicting the fact that $\mathcal{F}$ is a cover of $G$. $\Box$ \[rothbwonbytwo\] For each infinite cardinal $\kappa$ there is a ${\sf T}_0$ Lindelöf $P$-group of cardinality $\kappa$ such that TWO has a winning strategy in the game ${{\sf G}_1}({\mathcal{O}},{\mathcal{O}})$. Let $(G,)$ be a Lindelöf $P$-group for which TWO has a winning strategy in ${{\sf G}_1}({\mathcal{O}}_{nbd},{\mathcal{O}})$ (as for example in Proposition \[twowinsrothberger\]). Let $F$ be TWO’s winning strategy in that game. Define a strategy $\sigma$ for TWO in the game ${{\sf G}_1}({\mathcal{O}},{\mathcal{O}})$ as follows: When ONE plays the open cover $\mathcal{U}_1$, choose a neighborhood $N_1$ of the identity of $G$ as in Proposition \[pgpLebesgue\], and then let $\sigma(\mathcal{U}_1)$ be an element $U$ of $\mathcal{U}_1$ such that $F(N_1)N_1\subseteq U$. When ONE plays the next open cover $\mathcal{U}_2$ choose a neighborhoood $N_2$ of the identity of $G$ as in Proposition \[pgpLebesgue\], and then let $\sigma(\mathcal{U}_1,\mathcal{U}_2)$ be an element $U$ of $\mathcal{U}_2$ such that $F(N_1,N_2)N_2\subseteq U$, and so forth. Then $\sigma$ is a winning strategy for TWO. $\Box$ Call an open cover $\mathcal{U}$ of a topological space a $\gamma$-cover if $\mathcal{U}$ is infinite, and each infinite subset of $\mathcal{U}$ still covers the space. The symbol $\Gamma$ denotes the collection of open $\gamma$ covers of a space. In [@GN] Gerlits and Nagy introduced the notion of a $\gamma$-space: A topological space which satisfies the selection principle ${{\sf S}_1}(\Omega,\Gamma)$ is said to be a $\gamma$-space. It is evident that each $\gamma$-space is a Rothberger space. In Theorem 1 of [@GN] the authors prove For a ${\sf T}_{3\frac{1}{2}}$-space TWO has a winning strategy in ${{\sf G}_1}({\mathcal{O}},{\mathcal{O}})$ if, and only if, TWO has a winning strategy in ${{\sf G}_1}(\Omega,\Gamma)$. It is also evident that if TWO has a winning strategy in the game ${{\sf G}_1}(\Omega,\Gamma)$, then the underlying space is a $\gamma$-space. \[GNcorollary\] For each uncountable cardinal number $\kappa$ there is a Lindelöf $P$-group of cardinality $\kappa$ on which TWO has a winning strategy in the game ${{\sf G}_1}(\Omega,\Gamma)$. By a result of Gerlits and Nagy if TWO has a winning strategy in ${{\sf G}_1}({\mathcal{O}},{\mathcal{O}})$, then TWO has a winning strategy in ${{\sf G}_1}(\Omega,\Gamma)$. $\Box$ Large $\sigma$-compact Rothberger bounded ${\sf T}_0$ groups. ============================================================= In Proposition 4 of [@Corson], Corson proves essentially the following theorem[^4]: \[Corson\] Let $\{X_i:i\in I\}$ be a family of $\sigma$-compact topological groups and for each $i$ let $e_i$ be the identity element of $X_i$. Then the subgroup $$G:=\{f\in\prod_{i\in I}X_i:\vert\{j\in I:f(j)\neq e_j\}\vert< \aleph_0\}$$ is $\sigma$-compact. \[corsoncor\] For each infinite cardinal number $\kappa$ there is a ${\sf T}_0$ $\sigma$-compact Rothberger bounded group of cardinality $\kappa$. Let cardinal number $\kappa$ be given, and take $I$ to be $\kappa$. For each $i\in I$ take $X_i$ to be ${{\mathbb Z}}$, the additive group of integers. Now consider the group $G$ as in Corson’s Theorem. $G$ is in fact Rothberger bounded. To see this let for each $n$ a neighborhood $U_n$ of the identity element of $G$ be given. We may assume each $U_n$ is a basic open set, and thus that there is a finite set $F_n\subseteq I$ and for each $i\in F_n$ a neighborhood $N_{n,i}$ of $e_i$ such that $U_n$ is of the form $\{f\in G:(\forall i\in F_n)(f(i)\in N_{n,i})\}$. Now $C = \cup_{n<\omega}F_n$ is a countable subset of $I$ and $G_C = \{f\lceil_C:f \in G\}$ is evidently Rothberger bounded in $\prod_{i\in C}X_i$. For each $n$ choose a $g_n\in G_C$ such that $G_C\subseteq \cup_{n<\omega}g_nU_n\lceil_C$. For each $n$ choose $f_n\in G$ with $f_n\lceil_C = g_n$. Then it follows that $G\subseteq \cup_{n<\omega}f_nU_n$. Thus $G$ is a $\sigma$-compact Rothberger bounded group of cardinality $\kappa$. $\Box$ Using the method of proof of Proposition \[twowinsrothberger\], one proves \[twowinssigmacompactrothberger\] In the groups of Corollary \[corsoncor\] TWO has a winning strategy in the game ${{\sf G}_1}({{\mathcal{O}}_{nbd}},{\mathcal{O}})$. Note that for each finite subset $B$ of $\kappa$ the set $U_B =\{f\in G:(\forall \alpha\in B)(f(\alpha) = id_{G_{\alpha}}\}$ is in fact a subgroup of $G$, is a neighborhood of the identity element of $G$, and the set of such $U_B$ form a neighborhood basis of the identity element of $G$. Now apply the argument of Proposition \[twowinsrothberger\]. $\Box$ In fact the groups of Corollary \[corsoncor\] satisfy the stronger selection principle ${{\sf S}_1}({\mathcal{O}},{\mathcal{O}})$: One can prove more, namely \[twowinssigmacompact\] In the groups of Corollary \[corsoncor\] TWO has a winning strategy in the game ${{\sf G}_1}({\mathcal{O}},{\mathcal{O}})$ on $G$. Let $(G,)$ be such a group and $F$ be a winning strategy for TWO in the game ${{\sf G}_1}({{\mathcal{O}}_{nbd}},{\mathcal{O}})$ on this group. Since the group is $\sigma$-compact we write $G = \bigcup_{n=0}^{\infty}G_n$ where for each $n$ $G_n\subseteq G_{n+1}$ and $G_n$ is compact. For each open cover $\mathcal{U}$ of $G$ and for each $n$ choose a neighborhood $U(\mathcal{U},n)$ of the identity element of $G$ such that $U(\mathcal{U},n)$ is a subgroup of $G$ and for each $x\in G_n$ there is a $V\in\mathcal{U}$ such that $x* U(\mathcal{U},n) \subseteq V$. Let $V(x,\mathcal{U},n)$ be such a $V$. Define a strategy $\sigma$ for TWO of ${{\sf G}_1}({\mathcal{O}},{\mathcal{O}})$ as follows: When ONE plays the open cover $O_1$ in the first inning, TWO simulates a move for ONE in ${{\sf G}_1}({{\mathcal{O}}_{nbd}},{\mathcal{O}})$ as ${\mathcal{O}}(U(O_1,1))$, applies the winning strategy $F$ to this move to obtain $x_1U(O_1,1) = F({\mathcal{O}}(U(O_1,1)))$. Then if $G_1\cap x_1U(O_1,1)\neq \emptyset$ we find for an $x\in G_1$ that $xU(O_1,1) = x_1U(O_1,1)$, a left coset of $U(O_1,1)$ in $G$. Then TWO fixes such an $x$ and plays $$\sigma(O_1) = V(x,O_1,1) \in O_1.$$ If $G_1\cap x_1U(O_1,1) = \emptyset$ then TWO chooses an arbitrary element $x\in G_1$ and plays $$\sigma(O_1) = V(x,O_1,1) \in O_1.$$ When ONE next moves $O_2$, TWO simulates a move for ONE in ${{\sf G}_1}({{\mathcal{O}}_{nbd}},{\mathcal{O}})$ as follows: Define ${\mathcal{O}}= \{U\cap V:U\in O_1 \mbox{ and }V\in O_1\}\setminus\{\emptyset\}$ and let ONE’s move be ${\mathcal{O}}(U({\mathcal{O}},2))$. TWO’s response using $F$ is $x_2U({\mathcal{O}},2)=F({\mathcal{O}}(U(O_1,1)),{\mathcal{O}}(U({\mathcal{O}},2))$. Consider $x_2U({\mathcal{O}},2)\cap G_2$. If this is nonempty select any $x$ in this intersection. as before we have $xU({\mathcal{O}},2)=x_2U({\mathcal{O}},2)$, and now TWO responds with $$\sigma(O_1,O_2) = V(x,{\mathcal{O}},2)\in O_2.$$ If on the other hand the intersection is empty then TWO chooses any $x\in G_2$ and responds with $$\sigma(O_1,O_2) = V(x,{\mathcal{O}},2)\in O_2.$$ This procedure describes a strategy for TWO in the game ${{\sf G}_1}({\mathcal{O}},{\mathcal{O}})$ on $G$. To see that $\sigma$ is a winning strategy for TWO, consider any $\sigma$-play $$O_1,\, \sigma(O_1),\, O_2,\,\sigma(O_1,O_2),\,\cdots$$ Let an $x\in G$ be given. Fix the least $m$ with $x\in G_m$. By the definition of $\sigma$ we have an associated sequence $$U_n = U({\mathcal{O}}_n,n)$$ of subgroups of $G$ that are neighborhoods for the identity element where for each $n$ we have $U_{n+1}\subset U_n$, and ${\mathcal{O}}_1 = O_1$ while ${\mathcal{O}}_{n+1} = \{U\cap V:\, U\in {\mathcal{O}}_n \mbox{ and }V\in O_{n+1}\}\setminus\{\emptyset\}$, and elements $x_n$ of $G$ such that $$x_1U_1 = F({\mathcal{O}}(U_1)) \mbox{ and } x_{n+1}U_{n+1} = F({\mathcal{O}}(U_1),\cdots,{\mathcal{O}}(U_n)).$$ But then this is an $F$-play of ${{\sf G}_1}({{\mathcal{O}}_{nbd}},{\mathcal{O}})$ and thus won by TWO, meaning there are infinitely many $n$ with $x\in x_nU_n$. Thus, fix an $n>m$ with $x\in x_nU_n$. But then $G_n\cap x_nG_n\neq \emptyset$ and thus $\sigma(O_1,\cdots,O_n)\supseteq xU_n$, meaning $x\in \sigma(O_1,\cdots,O_n)$ $\Box$ \[sigmacompactGN\] For each infinite cardinal number $\kappa$ there is a $\sigma$-compact ${\sf T}_0$ topological group of cardinality $\kappa$ such that TWO has a winning strategy in the game ${{\sf G}_1}(\Omega,\Gamma)$. \[rothbdedtorothb\] Let $(G,)$ be a $\sigma$-compact ${\sf T}_0$ topological group with property ${{\sf S}_1}({{\mathcal{O}}_{nbd}},{\mathcal{O}})$. Then $G$ has the property ${{\sf S}_1}(\Omega,\Gamma)$. This follows directly from Corollary \[rothbbdedgp\] part (1) that $(G,)$ has property ${{\sf S}_1}({\mathcal{O}},{\mathcal{O}})$ since in the notation of that corollary $G=H$ and ${\mathcal{O}}_H={\mathcal{O}}$. Recall that a ${\sf T}_0$ topological group is ${\sf T}_{3\frac{1}{2}}$. Since a compact ${\sf T}_3$-space is Rothberger if, and only if, it is scattered ([@BCM] Proposition 34), if, and only if, it is a $\gamma$-space (Theorem 4 of [@GN] and its Corollary), and since the countable union of compact $\gamma$-spaces is a $\gamma$-space (The union of two compact $\gamma$-spaces is a compact Rothberger space and thus a compact $\gamma$ space. Now apply Jordan’s theorem [@Jordan] Corollary 14), these topological groups are in fact $\sigma$-compact (thus $\sigma$-scattered) $\gamma$-groups. $\Box$ Thus for any cardinal number $\kappa$, the topological group ${{\mathbb R}}^{\kappa}$ contains $\sigma$-compact $\gamma$ subgroups of cardinality $\kappa$. It follows for example that the elements with finite support of any power of the integers is a $\sigma$-compact Rothberger bounded topological group. \[rothbdedtorothbomega\] For each infinite cardinal number $\kappa$ there is a ${\sf T}_0$ topological group $(G,)$ of cardinality $\kappa$ which is a $\sigma$-compact Rothberger space in all finite powers. L. Babinkostova, Lj.D.R. Kočinac and M. Scheepers, Combinatorics of open covers (XI): Menger- and Rothberger bounded groups, [Topology and its Applications]{} 154 (2007), 1269-1280. M. Bonanzinga, F. Cammoroto and M. Matveev, Projective versions of selection principles, [Topology and its Applications]{} 157 (2010), 874- 893. W.W. Comfort, Compactness like properties for generalized weak topological sums, [Pacific Journal of Mathematics]{} 60:1 (1975), 31 - 37 W.W. Comfort and K.A. Ross, Pseudocompactness and uniform continuity in topological groups, [Pacific Journal of Mathematics]{} 16:3 (1966), 483 - 496. H.H. Corson, Normality in subsets of product spaces, [American Journal of Mathematics]{} 81 (1959), 784 - 796. F. Galvin, Indeterminacy of point-open games, [Bulletin de L’Acádemie Polonaise des Sciences]{} 26 (1978), 445 - 448. J. Gerlits, Zs. Nagy, Some properties of $C(X)$, I, [Topology and its Applications]{} 14 (1982), 151–161. I.I. Guran, On topological groups close to being Lindelöf, [Soviet Math. Dokl.]{} 23 (1981), 173 - 175. C. Hernandez, Topological groups close to being $\sigma$-compact, [Topology and its Applications]{} 102 (2000), 101 - 111. E. Hewitt and K.A. Ross, Abstract Harmonic Analysis I, in [Die Grundlehren der Mathematischen Wissehschaften]{} 115 [Springer-Verlag]{}, 1963. F. Jordan, There are no hereditary productive $\gamma$-spaces, [Topology and its Applications]{} 155 (2008), 1786 - 1791. J. Pawlikowski, Undetermined sets of point-open games, [Fundamenta Mathematicae]{} 144 (1994), 279 - 285. F. Rothberger, Eine Verschärfung der Eigenschaft C, [Fundamenta Mathematicae]{} 30 (1938), 50 - 55. M. Scheepers, Finite powers of strong measure zero sets, [The Journal of Symbolic Logic]{} 64:3 (1999), 1295 - 1306. M. Scheepers, Open covers and partition relations, [Proceedings of the American Mathematical Society]{} 127 (1999), 577 – 581. [^1]: Since we do not need this correspondence here, we refer readers to Galvin’s paper [@Galvin] for a definition of the point-open game. [^2]: Lemma \[lebesguecov\] must be well-known but I was not able to track down a reference for it. [^3]: i.e., $N_k^{-1}=N_k$. [^4]: Corson formulates the proposition for the case when the factor spaces $X_i$ are all the real line. But the argument gives the more general result of Theorem \[Corson\].	{ "pile_set_name": "ArXiv" }
--- abstract: 'This article is concerned with the constants that appear in Harish-Chandra’s character formula for stable discrete series of real reductive groups, although it does not require any knowledge about real reductive groups or discrete series. In Harish-Chandra’s work the only information we have about these constants is that they are uniquely determined by an inductive property. Later Goresky–Kottwitz–MacPherson and Herb gave different formulas for these constants; see [@GKM Theorem 3.1] and [@Herb-2S Theorem 4.2]. In this article we generalize these formulas to the case of arbitrary finite Coxeter groups (in this setting, discrete series no longer make sense), and give a direct proof that the two formulas agree. We actually prove a slightly more general identity that also implies the combinatorial identity underlying the discrete series character identities of [@Morel Proposition 3.3.1]. We also introduce a signed convolution of valuations on polyhedral cones in Euclidean space and show that the resulting function is a valuation. This gives a theoretical framework for the valuation appearing in [@GKM Appendix A]. In Appendix \[appendix\_2\_structures\] we extend the notion of $2$-structures (due to Herb) to pseudo-root systems.' author: - '[Richard EHRENBORG]{}, [Sophie MOREL]{} and [Margaret READDY]{}' bibliography: - 'TS2\_short.bib' date: '.' title: \| A generalization of combinatorial identities\ for stable discrete series constants --- =1 =8.5truein =11truein Introduction ============ Although this paper deals exclusively with the combinatorics of real hyperplane arrangements and Coxeter complexes, it has its origin in the representation theory of real reductive groups and its connections with the cohomology of locally symmetric spaces, and in particular of Shimura varieties. We start by explaining some of this background. This explanation can be safely skipped by the reader not interested in Shimura varieties. Let $G$ be an algebraic group over ${{\mathbb{Q}}}$. To simplify the exposition, we assume that $G$ is connected and semisimple. Let $K_\infty$ be a maximal compact subgroup of $G({{\mathbb{R}}})$ and $K$ be an open compact subgroup of $G(\mathbb{A}^\infty)$, where $\mathbb{A}^\infty=\widehat{{{\mathbb{Z}}}}\otimes_{{\mathbb{Z}}}{{\mathbb{Q}}}$ is the ring of finite adèles of ${{\mathbb{Q}}}$. We consider the double quotient $X_K=G({{\mathbb{Q}}})\backslash(G({{\mathbb{R}}})\times G(\mathbb{A}^\infty))/(K_\infty\times K)$. This is a real analytic variety for $K$ small enough, and the projective system $(X_K)_{K\subset G(\mathbb{A}^\infty)}$ has an action of $G(\mathbb{A}^\infty)$ by Hecke correspondences, which induces an action of the Hecke algebra at level $K$ on the cohomology of $X_K$, for any reasonable cohomology theory. We restrict our attention further to the case where the real Lie group $G({{\mathbb{R}}})$ has a discrete series. This is the so-called “equal rank case” because it occurs if and only if the groups $G({{\mathbb{R}}})$ and $K_\infty$ have the same rank. Then the $L^2$-cohomology $H^_{(2)}(X_K)$ is finite-dimensional, and Matsushima’s formula, proved in this generality by Borel and Casselman [@Borel_Casselman], gives a description of this cohomology and of its Hecke algebra action in terms of discrete automorphic representations of $G$ whose infinite component is a cohomological representation of $G({{\mathbb{R}}})$, and in particular, either a discrete series or a special type of non-tempered representation. Another cohomology of interest in this case is the intersection cohomology $IH^(\overline{X}_K)$ of the minimal Satake compactification $\overline{X}_K$ of $X_K$. In order to study this cohomology, Goresky, Harder and MacPherson introduced in [@GHM] a family of cohomology theories called “weighted cohomologies” and showed that the two middle weighted cohomologies agree with $IH^(\overline{X}_K)$ if $X_K$ has the structure of a complex algebraic variety. This result was later generalized by Saper in [@Saper]. All the cohomology theories that we discussed have actions of the Hecke algebra, and the isomorphism of the previous paragraph is equivariant for this action. Zucker conjectured that there should be a Hecke-equivariant isomorphism between $H^_{(2)}(X_K)$ and $IH(\overline{X}_K)$. This conjecture was proved by Looijenga [@Loo], Looijenga–Rapoport [@LooRap] and Saper–Stern [@SaSt] if $X_K$ has the structure of a complex algebraic variety and by Saper [@Saper] in general. In particular, by comparing the formulas for the action of a Hecke operator on weighted cohomology (this was calculated by Goresky and MacPherson using topological methods in [@GM]) and on $L^2$-cohomology (this was calculated by Arthur using the Arthur–Selberg trace formula in [@A-L2]), we obtain a formula for averaged discrete series characters of the group $G({{\mathbb{R}}})$. One of the goals of the paper [@GKM] of Goresky–Kottwitz–MacPherson was to prove this identity directly. If moreover the space $X_K$ is the set of complex points of a Shimura variety, then it descends to an algebraic variety over an explicit number field $E$ known as the reflex field, as does the minimal Satake compactification, and so the intersection cohomology has a natural action of the absolute Galois group $\operatorname{Gal}(\overline{E}/E)$. We can further complicate the calculation by trying to calculate the trace on $IH^(\overline{X}_K)$ of Hecke operators twisted by elements of the group $\operatorname{Gal}(\overline{E}/E)$, for example, powers of Frobenius maps. In the case where $X_K$ is a Siegel modular variety, this was done by the second author in [@Morel]. It requires a slightly different character identity for averaged discrete series characters of $G({{\mathbb{R}}})$, also involving discrete series characters of the endoscopic groups of $G$, and whose relationship with the Goresky–Kottwitz–MacPherson identity was not clear. We return to a discussion of the current article. In a previous article [@Ehrenborg_Morel_Readdy], we investigate the character identity of [@Morel]. In particular we relate it to the geometry of the Coxeter complex of the symmetric group and give a simpler and more natural proof than the brute force calculation in the appendix of [@Morel]. The goal of the present article is to generalize the approach and methods of [@Ehrenborg_Morel_Readdy] and to prove a combinatorial identity (Theorem \[thm\_main\]) that implies the character formulas of [@GKM] and of [@Morel] (see Section \[section\_applications\]). To obtain the character formula of [@GKM] from our results, we need to use Herb’s formula for averaged discrete series characters (see for example [@Herb] and [@Herb-2S]). We also generalize, in Corollary \[cor\_shelling\_order\] and Lemma \[lemma\_Coxeter\_A\], the geometric result of [@Ehrenborg_Morel_Readdy] (see Theorem 4.3 of that article). In fact, we prove an identity for all Coxeter systems with finite Coxeter groups, and not just the root systems that are generated by strongly orthogonal roots. The representation-theoretic interpretation of our identity in the general case is still unclear. We now describe in more detail the different sections of the article. In Section \[section\_hyperplane\_arrangements\] we review some background material about real hyperplane arrangements and introduce our main geometric construction, which we call the weighted complex. The weighted complex is the set of all the faces of a fixed hyperplane arrangement that are on the nonnegative side of an auxiliary hyperplane $H_\lambda$. It contains what is known as the bounded complex* in the theory of affine oriented matroids, and coincides with it if $H_\lambda$ is in general position. We also prove that, under a hypothesis about the dihedral angles between the hyperplanes of the arrangement (Condition \[(A)\] in Subsection \[section\_weighted\_complex\], which is always true in the Coxeter case), the weighted complex is shellable. In Section \[section\_definition\_sum\] we introduce one half of our main identity (see Definition \[def\_psi\]) and study its properties. In particular, we establish a recursion result in Proposition \[prop\_induction\_formula\_for\_psi\] which will be the heart of the proof of Theorem \[thm\_main\]. Our treatment is inspired by that of Sections 1-2 and Appendix A of [@GKM], but we make no assumptions on the Coxeter system. In fact, we prove the recursion result for arbitrary hyperplane arrangements. The main technical tool is Theorem \[theorem\_signed\_convolution\], which generalizes [@GKM Proposition A.4]. Section \[section\_main\_thm\] contains the statement and proof of our main theorem (Theorem \[thm\_main\]). We specialize to the case of Coxeter systems and introduce the second half of our main identity, generalizing a construction of Herb that we review in detail in Appendix \[appendix\_2\_structures\]. The proof that both halves of the identity are equal follows the same lines as the proofs of the character formulas in the Goresky–Kottwitz–MacPherson and Herb articles: First, using the recursive expressions for both sides (Propositions \[prop\_induction\_formula\_for\_psi\] and \[prop\_induction\_RHS\]), we reduce to the case where $\lambda$ is dominant so that $H_\lambda$ is as far as possible from the base chamber. Then we treat this case directly by a geometric argument based on the results of Section \[section\_hyperplane\_arrangements\]. In Section \[section\_applications\] we explain how Theorem \[thm\_main\] implies the identities of [@GKM Theorem 3.1] and of [@Ehrenborg_Morel_Readdy Theorem 6.4], and in Section \[section\_concluding\_remarks\] we include concluding remarks. We finish with two appendices. Each can be read independently from the rest of the article. The goal of our Appendix \[appendix\_valuations\] is to generalize [@GKM Proposition A.4], which is a key part of the induction in the proof of our main theorem. In Appendix A of their article [@GKM], Goresky–Kottwitz–MacPherson show that a certain function, which they call $\psi_C(x,\lambda)$, is a valuation (see Definition \[def\_valuation\]) on closed convex polyhedral cones, although they do not phrase it in these terms. We show that their function is a special case of a general construction that takes two valuations and produces a third one via a signed convolution. See Definition \[definition\_signed\_convolution\] for the precise definition of this signed convolution. In Appendix \[appendix\_2\_structures\] we review the theory of $2$-structures, due to Herb; see for example Herb’s review article [@Herb-2S]. We believe that this will be useful to the reader for a number of reasons. The proofs of the fundamental results of this theory are somewhat scattered in the literature and sometimes left as exercises. Furthermore, we needed to slightly adapt a number of results so that they continue to hold for Coxeter systems that do not necessarily arise from a (crystallographic) root system. Hyperplane arrangements {#section_hyperplane_arrangements} ======================= Background material {#background} ------------------- We fix a finite-dimensional ${{\mathbb{R}}}$-vector space $V$ with an inner product $(\cdot,\cdot)$. If $\alpha\in V$, we write $$\begin{aligned} H_\alpha & = \{x\in V : (\alpha,x)=0\}, & H^+_\alpha & = \{x\in V : (\alpha,x)> 0\}, & H^-_\alpha & = \{x\in V : (\alpha,x)< 0\}.\end{aligned}$$ We also denote by $s_\alpha$ the (orthogonal) reflection across the hyperplane $H_\alpha$. Let $(\alpha_e)_{e \in E}$ be a finite family of nonzero vectors in $V$. The corresponding (central) hyperplane arrangement is the family of hyperplanes ${\mathcal{H}}= (H_{\alpha_e})_{e \in E}$. Let $V_{0}$ be the intersection of all the hyperplanes, that is, $V_0=\bigcap_{e \in E} H_{\alpha_e}$. We say that the arrangement ${\mathcal{H}}$ is essential if $V_0 = \{0\}$, which means that the family $(\alpha_e)_{e \in E}$ spans $V$. We will mostly be interested in hyperplane arrangements that are essential and have no repeated hyperplanes, that is, where $\alpha_e\not\in{{\mathbb{R}}}\alpha_f$ for distinct $e, f \in E$. However, some of the arrangements arising in the inductive formula of Proposition \[prop\_induction\_formula\_for\_psi\] might not satisfy these conditions, so we do not want to impose this restriction earlier than necessary. Consider the map $s:V \longrightarrow \{+,-,0\}^E$ sending $x\in V$ to the family $(\operatorname{sign}((\alpha_e,x)))_{e \in E}$, where $\operatorname{sign}:{{\mathbb{R}}}\longrightarrow \{+,-,0\}$ is the map sending positive numbers to $+$, negative numbers to $-$ and zero to $0$. The image of the map $s:V \longrightarrow \{+,-,0\}^E$ is the set of covectors of an oriented matroid (see for example [@OM Definition 4.1.1]), which is the oriented matroid corresponding to the hyperplane arrangement. In fact, some of our results extend to general oriented matroids. In this article we have chosen to concentrate on hyperplane arrangements to keep the exposition more concrete. In particular, we do not assume that the reader knows what an oriented matroid is. We denote by ${\mathscr{L}}$ the set of nonempty subsets of $V$ of the form $C=s^{-1}(X)$, for a sign vector $X\in\{+,-,0\}^E$. The elements of ${\mathscr{L}}$ are called faces of the arrangement. The set ${\mathscr{L}}$ has a natural partial order given by $C \leq D$ if and only if $C \subseteq \overline{D}$. The relation $C \leq D$ is equivalent to the fact that for every $e\in E$ we have $s(C)_e=0$ or $s(C)_e=s(D)_e$. The set ${\mathscr{L}}$ with this partial order is called the face poset of the arrangement. Note that $V_{0}$ is the minimal element of ${\mathscr{L}}$. When we adjoin a maximal element $\widehat{1}$ to the poset ${\mathscr{L}}$, we obtain a lattice ${\mathscr{L}}\cup \{\widehat{1}\}$ known as the face lattice. Note that, under our convention, faces other than $V_0$ are not closed subsets of $V$: for every $C\in{\mathscr{L}}$, the closure $\overline{C}$ is a closed convex polyhedral cone in $V$, and it is an intersection of closed half-spaces $\overline{H^{\pm}_{\alpha_e}}$. The poset ${\mathscr{L}}$ is graded with the rank of a face $C\in{\mathscr{L}}$ given by $\rho(C) = \dim(C)-\dim(V_0)$, where we write $\dim(C)$ for $\dim(\operatorname{Span}(C))$. We denote by ${\mathscr{T}}$ the set of maximal faces of ${\mathscr{L}}$. These elements are often called chambers, regions or topes, and are the connected components of $V-\bigcup_{e\in E}H_{\alpha_e}$. If $T \in {\mathscr{T}}$ then $T$ is an open subset of $V$, and its closure is a closed convex polyhedral cone of dimension $\dim(V)$. If $X,Y\in\{+,-,0\}^E$, their composition $X\circ Y$ is the sign vector defined by $$(X\circ Y)_e = \begin{cases}X_e & \text{ if }X_e\not=0,\\ Y_e & \text{ otherwise.}\end{cases}$$ If $C,D\in{\mathscr{L}}$ then $s(C)\circ s(D)$ is also the image of a face of ${\mathscr{L}}$, and we denote this face by $C\circ D$. This is the unique face of ${\mathscr{L}}$ that contains all vectors of $V$ of the form $x+\varepsilon y$, with $x\in C$, $y\in D$ and $\varepsilon>0$ sufficiently small (relative to $x$ and $y$). Define the separation set of $C$ and $D$ to be the set $$S(C,D)=\{e \in E : s(C)_e=-s(D)_e\not=0\}.$$ This is the set of $e \in E$ such that $C$ and $D$ are on different sides of the hyperplane $H_{\alpha_e}$. Fix a chamber $B\in{\mathscr{T}}$. We can then define a partial order $\preceq_B$ on ${\mathscr{T}}$ by declaring that $T\preceq_B T'$ if and only if $S(B,T) \subseteq S(B,T')$. The resulting poset is called the chamber poset with base chamber $B$. We will denote it by ${\mathscr{T}}_B$. It is a poset with minimal element $B$ and maximal element $-B$. When all the hyperplanes are distinct, this poset is also graded with the poset rank function $\rho(T) = \|S(B,T)\|$; see [@OM Proposition 4.2.10]. We also consider the graph with vertex set ${\mathscr{T}}$, where two chambers $T,T'\in{\mathscr{T}}$ are connected by an edge if and only if $\overline{T}\cap\overline{T}'$ spans a hyperplane (necessarily one of the $H_{\alpha_e}$). In this situation, we say that this hyperplane is a wall of the chambers $T$ and $T'$. This graph is called the chamber graph. In the case when all the hyperplanes of the arrangement ${\mathcal{H}}$ are distinct, the distance between two chambers $T$ and $T'$ in this graph is $\|S(T,T')\|$; see [@OM Proposition 4.2.3]. Consider the sphere $\mathbb{S}$ of center $0$ and radius $1$ in $V/V_0$. The intersections $\overline{C}\cap\mathbb{S}$, for $C\in{\mathscr{L}}$, form a regular cell decomposition $\Sigma({\mathscr{L}})$ of $\mathbb{S}$, and we will identify ${\mathscr{L}}$ with the face poset of this regular cell decomposition. We will need the following notion and result. For instance, see [@OM Definition 4.7.14]. A pure $n$-dimensional polytopal complex $\Delta$ is [shellable]{} if it is $0$-dimensional (and hence a collection of a finite number of points), or if there is a linear order of the facets $F_{1}, F_{2}, \ldots, F_{k}$ of $\Delta$, called a shelling order, such that: - The boundary complex of $F_{1}$ is shellable. - For $1 < j \leq k$ the intersection of $\overline{F_{j}}$ with the union of the closures of the previous facets is nonempty and is the beginning of a shelling of the $(n-1)$-dimensional boundary complex of $F_{j}$, that is, $$\overline{F_{j}} \cap (\overline{F_{1}} \cup \overline{F_{2}} \cup \cdots \cup \overline{F_{j-1})} = \overline{G_{1}} \cup \overline{G_{2}} \cup \cdots \cup \overline{G_{r}} ,$$ where $G_{1}, G_{2}, \ldots, G_{r}, \ldots, G_{t}$ is a shelling order of $\partial F_{j}$ and $r\geq 1$. ([@OM Theorem 4.3.3].) Let $B$ be a chamber in ${\mathscr{T}}$. Then any linear extension of the chamber poset with base chamber $B$ is a shelling order on the facets of $\Sigma({\mathscr{L}})$. \[thm\_shelling\_order\] Finally, we recall the definition of the star of a face in ${\mathscr{L}}$. Let $C\in{\mathscr{L}}$. The star of $C$ in ${\mathscr{L}}$ is $\{D\in{\mathscr{L}}: C\leq D\}$. Geometrically it is the set of faces of ${\mathscr{L}}$ whose closure contains $C$. We will denote it by ${\mathscr{L}}_{\geq C}$. \[lemma\_star\] Let $C\in{\mathscr{L}}$ and let $E(C)=\{e\in E : C\subset H_{\alpha_e}\}$. Consider the hyperplane arrangement ${\mathcal{H}}(C)=(H_{\alpha_e})_{e\in E(C)}$ and let ${\mathscr{L}}_{{\mathcal{H}}(C)}$ be its face poset. Then the following three statements hold: - Each face $D$ of ${\mathscr{L}}$ is contained in a unique face $D'$ of ${\mathscr{L}}_{{\mathcal{H}}(C)}$, and the map $D \longmapsto D'$ induces an isomorphism of posets $\iota_C: {\mathscr{L}}_{\geq C} \stackrel{\sim}{\longrightarrow} {\mathscr{L}}_{{\mathcal{H}}(C)}$. In particular, it sends the chambers of ${\mathscr{T}}\cap{\mathscr{L}}_{\geq C}$ to the chambers of ${\mathscr{L}}_{{\mathcal{H}}(C)}$. - If $D_1,D_2\in{\mathscr{L}}_{\geq C}$ then the inclusion $S(D_1,D_2)\subset E(C)$ holds. In particular, we have the equality $S(D_1,D_2)=S(\iota_C(D_1),\iota_C(D_2))$, where the isomorphism $\iota_C$ is as in (i). - The isomorphism $\iota_C:{\mathscr{L}}_{\geq C} \stackrel{\sim}{\longrightarrow} {\mathscr{L}}_{{\mathcal{H}}(C)}$ preserves composition and dimension, that is, for all $D,D'\in{\mathscr{L}}_{\geq C}$, the identities $\iota_C(D \circ D') = \iota_C(D)\circ\iota_C(D')$ and $\dim(\iota_C(D)) = \dim(D)$ hold. In particular, ${\mathscr{L}}_{\geq C}$ is also isomorphic to the face poset of a regular cell decomposition of the unit sphere in $V/\bigcap_{e \in E(C)} H_{\alpha_e}$ that we denote by $\Sigma({\mathscr{L}}_{\geq C})$. Statement (i) is clear. We prove (ii). Let $D_1,D_2\in{\mathscr{L}}_{\geq C}$, and let $e\in S(D_1,D_2)$. Suppose for example that $s(D_1)_e=+$ and $s(D_2)_e=-$. (The other case is similar.) Then $\overline{D}_1\subset\overline{H_{\alpha_e}^+}$ and $\overline{D}_2\subset\overline{H_{\alpha_e}^-}$, so $C \subset \overline{D}_1\cap\overline{D}_2 \subset \overline{H_{\alpha_e}^+} \cap \overline{H_{\alpha_e}^-}= H_{\alpha_e}$, which implies that $e\in E(C)$. The first statement of (iii) follows easily from the definitions: the composition $D\circ D'$ is defined on the sign vectors of $D$ and $D'$, and the isomorphism $\iota_{C}$ just forgets the coordinates outside of $E(C)$ in these sign vectors. We prove the second statement of (iii). Let $D\in{\mathscr{L}}_{\geq C}$, and let $D'$ be the unique face of ${\mathscr{L}}_{{\mathcal{H}}(C)}$ containing $D$. We clearly have $\dim(D)\leq\dim(D')$. If $\dim(D')>\dim(D)$ then there exists $e\in E$ such that $D\subset H_{\alpha_e}$ and $D'\not\subset H_{\alpha_e}$. But $C\subset\overline{D}$, so this implies that $e\in E(C)$. As $D'$ is not included in $H_{\alpha_e}$, it must be contained in one of the open half-spaces $H^{\pm}_{\alpha_e}$, contradicting the fact that $D'$ contains $D$. \[rmk\_star\_convex\] Let $C\in{\mathscr{L}}$ and let $F'=\{e\in E : C\not\subset H_{\alpha_e}\}$. Then the set ${\mathscr{T}}\cap{\mathscr{L}}_{\geq C}$ is equal to $\{T\in{\mathscr{T}}: \forall e\in F',\ s(T)_e=s(C)_e\}$, so it is a $T$-convex subset of ${\mathscr{T}}$ in the sense of [@OM Definition 4.2.5]; see [@OM Proposition 4.2.6]. In other words, it contains every shortest path in the chamber graph between any two of its elements, so it is a lower order ideal in ${\mathscr{T}}_B$ for every choice of base chamber $B\in{\mathscr{T}}\cap{\mathscr{L}}_{\geq C}$. The weighted complex {#section_weighted_complex} -------------------- Let $\lambda\in V$. We consider the following subset of the face poset ${\mathscr{L}}$: $${\mathscr{L}}_{\lambda}=\{C\in{\mathscr{L}}: C \subseteq \overline{H_\lambda^+}\}.$$ More generally, if $C_0$ is a fixed face of ${\mathscr{L}}$, we also consider the intersection $${\mathscr{L}}_{\lambda,\geq C_0}={\mathscr{L}}_\lambda\cap{\mathscr{L}}_{\geq C_0} .$$ (See [@OM Section 4.5] for definitions.) If $\lambda\not=0$, then the hyperplane arrangement $\{H_\lambda\}\cup\{H_{\alpha_e} : e \in E\}$ defines an affine oriented matroid with distinguished hyperplane $H_\lambda$. If $H_\lambda$ is in general position relative to the $H_{\alpha_e}$, that is, if $\lambda$ is not in the span of any family $(\alpha_e)_{e \in F}$ for $\|F\| \leq \dim(V)-1$, then ${\mathscr{L}}_{\lambda}$ coincides with the bounded complex of this affine oriented matroid. In general, ${\mathscr{L}}_{\lambda}$ is larger. The basic properties of the subsets ${\mathscr{L}}_{\lambda}$ and ${\mathscr{L}}_{\lambda,\geq C_0}$ are given in the following proposition. The following two statements hold: - For a fixed face $C_0$ of ${\mathscr{L}}$ the set ${\mathscr{L}}_{\lambda,\geq C_0}$ is a lower order ideal in ${\mathscr{L}}_{\geq C_0}$. - Let $C\in{\mathscr{L}}_{\lambda}$. Then there exists $T\in{\mathscr{T}}\cap{\mathscr{L}}_{\lambda}$ such that $C\leq T$. \[prop\_weighted\_complex\_basic\] It suffices to prove (i) when $C_0$ is the minimal face of ${\mathscr{L}}$. Let $C,D\in{\mathscr{L}}$ such that $C \leq D$ and $D\in{\mathscr{L}}_{\lambda}$. The hypothesis implies that $C\subset\overline{D}$ and $D\subset\overline{H_\lambda^+}$. As $\overline{H_\lambda^+}$ is closed, this immediately gives $C\subset\overline{H_\lambda^+}$, hence $C\in{\mathscr{L}}_{\lambda}$. To show (ii) let $D_{1},D_{2}, \ldots, D_{p}$ be the chambers of ${\mathscr{T}}$ that are larger than $C$ with respect to the partial order $\leq$. If one of them is contained in $\overline{H_\lambda^+}$, then we are done. Otherwise, for every $i\in\{1,2,\ldots,p\}$, we can find a point $x_i\in D_i$ such that $(\lambda,x_i)<0$. Let $\mathcal{H}^{\prime}$ be the subarrangement of $\mathcal{H}$ where we remove all the hyperplanes of $\mathcal{H}$ that contain the cone $C$. In the arrangement $\mathcal{H}^{\prime}$ all the points $x_{i}$ are contained in the same chamber $C^{\prime}$. In particular, the convex hull $P$ of the points $x_{1}, x_{2}, \ldots, x_{p}$ is contained in $C^{\prime}$. The convex hull $P$ intersects the linear span of the cone $C$ in a point $x$. Since all the points $x_{i}$ are in the open half-space $H_\lambda^-$, so is the point $x$, that is, $(\lambda,x) < 0$. By inserting the hyperplanes of $\mathcal{H}$ that contain the cone $C$ back in the arrangement $\mathcal{H}^{\prime}$, we subdivide the region $C^{\prime}$ into regions $C_{1}, C_{2}, \ldots, C_{p}$. But the point $x$ belongs to the closure of each region $C_{i}$, thus the point $x$ belongs to the cone $C$. This is a contradiction since $C$ is contained in the half-space $\overline{H_{\lambda}^+}$, so $(\lambda,x) \geq 0$. The subcomplex of the cell decomposition $\Sigma({\mathscr{L}})$, respectively $\Sigma({\mathscr{L}}_{\geq C_0})$, whose face poset is the lower order ideal ${\mathscr{L}}_\lambda$, respectively ${\mathscr{L}}_{\lambda,\geq C_0}$, is called the weighted complex and denoted by $\Sigma({\mathscr{L}}_\lambda)$, respectively $\Sigma({\mathscr{L}}_{\lambda,\geq C_0})$. By Proposition \[prop\_weighted\_complex\_basic\] it is pure of the same dimension as $\Sigma({\mathscr{L}}_\lambda)$, respectively $\Sigma({\mathscr{L}}_{\lambda,\geq C_0})$. We now introduce a geometric condition on the hyperplane arrangement ${\mathcal{H}}$ that will imply the shellability of the weighted complex. Denote by [ (A)currentlabel\[(A)\]]{} the following condition on the family $(\alpha_e)_{e \in E}$ (or the corresponding arrangement): For every $T\in{\mathscr{T}}$ and for every $e \in E$ such that $S = \overline{T}\cap H_{\alpha_e}$ is of dimension $\dim(V)-1$, that is, $S$ is a facet of the convex cone $\overline{T}$, the following inclusions hold: $$\begin{aligned} T\subseteq\operatorname{relint}(S)+{{\mathbb{R}}}_{> 0}\alpha_e \text{ if } T\subseteq H^+_{\alpha_e}, \\ T\subseteq\operatorname{relint}(S)+{{\mathbb{R}}}_{< 0}\alpha_e \text{ if } T\subseteq H^-_{\alpha_e},\end{aligned}$$ where $\operatorname{relint}{S}$ is the relative interior of the cone $S$, that is, the interior of $S$ in $\operatorname{Span}(S)$. Geometrically, condition \[(A)\] means that if $T\in{\mathscr{T}}$ then the dihedral angle between any two adjacent facets (facets whose intersection is a face of dimension $\dim(V)-2$) of the convex polyhedral cone $\overline{T}$ is acute, that is, less than or equal to $\pi/2$. \[prop\_crossing\_wall\] Suppose that the arrangement ${\mathcal{H}}$ satisfies condition \[(A)\]. Let $T,T'\in{\mathscr{T}}$ and $e \in E$ such that $S(T,T')=\{e\}$, the inner product $(\alpha_e,\lambda)$ is nonnegative and the inclusion $T' \subset H^-_{\alpha_e}$ holds. Then $T'\in{\mathscr{L}}_{\lambda}$ implies that $T\in{\mathscr{L}}_\lambda$. The hypothesis implies that $\overline{T}\cap\overline{T}' = \overline{T}\cap H_{\alpha_e} = \overline{T}'\cap H_{\alpha_e}$. We denote this intersection by $S$. It is a facet of both $\overline{T}$ and $\overline{T}'$. By condition \[(A)\], we have $T \subset \operatorname{relint}{S}+{{\mathbb{R}}}_{> 0}\alpha_e$ and $T' \subset \operatorname{relint}{S}+{{\mathbb{R}}}_{< 0}\alpha_e$. In particular, if $x\in T$ then there exists $c>0$ such that $x - c \cdot \alpha_e \in T'$. Then we have $(x,\lambda)=(x- c \cdot \alpha_e,\lambda)+c \cdot (\alpha_e,\lambda)\geq 0$. This implies that $T\subset\overline{H_\lambda^+}$, that is, $T\in{\mathscr{L}}_{\lambda}$. Suppose that the arrangement ${\mathcal{H}}$ satisfies condition \[(A)\]. If $(\lambda,\alpha_e)\geq 0$ for every $e \in E$ and if there exists $B\in{\mathscr{T}}$ such that $B\subset H^+_{\alpha_e}$ for every $e \in E$, then ${\mathscr{T}}\cap {\mathscr{L}}_{\lambda}$ is a lower order ideal in ${\mathscr{T}}_B$. More generally, if $C\in{\mathscr{L}}$ and $E(C)=\{e\in E : C\subset H_{\alpha_e}\}$, if $(\lambda,\alpha_f)\geq 0$ for every $f\in E(C)$ and if there exists $B\in{\mathscr{T}}\cap{\mathscr{L}}_{\geq C}$ such that $B\subset H^+_{\alpha_f}$ for every $f\in E(C)$, then ${\mathscr{T}}\cap{\mathscr{L}}_{\lambda,\geq C}$ is a lower order ideal in ${\mathscr{T}}_B$. \[cor\_lower\_order\_ideal\] It suffices to prove the second statement. Let $T,T'$ be such that $S(B,T)\subset S(B,T')$ and $T' \in {\mathscr{L}}_{\lambda,\geq C}$. We want to show that $T\in{\mathscr{L}}_{\lambda,\geq C}$. As ${\mathscr{T}}_B$ is a graded poset and the intersection ${\mathscr{T}}\cap{\mathscr{L}}_{\geq C}$ is a lower order ideal in ${\mathscr{T}}_B$ (see Remark \[rmk\_star\_convex\]), we know that $T\in{\mathscr{T}}\cap{\mathscr{L}}_{\geq C}$, and it suffices to treat the case where $S(T',B)-S(T,B)$ is a singleton. Let $f$ be the single index of $S(B,T')-S(B,T)$. As $B,T'\in{\mathscr{T}}\cap{\mathscr{L}}_{\geq C}$, we have $f\in E(C)$ by Lemma \[lemma\_star\](ii), so $B\subset H^+_{\alpha_f}$. As $f \in S(B,T')-S(B,T)$, we have $T'\subset H^-_{\alpha_f}$ and $T\subset H^+_{\alpha_f}$. Also, as $f \in E(C)$, we have $(\lambda,\alpha_f) \geq 0$. So we may apply Proposition \[prop\_crossing\_wall\], and we obtain that $T\in{\mathscr{L}}_{\lambda}$. Suppose that the arrangement ${\mathcal{H}}$ satisfies condition \[(A)\]. Then the complex $\Sigma({\mathscr{L}}_\lambda)$ is shellable. Moreover, there exists a shelling order on its chambers which is an initial shelling of $\Sigma({\mathscr{L}})$. In particular, if $\lambda\not=0$ then $\Sigma({\mathscr{L}}_\lambda)$ is a shellable $PL$ ball of dimension $\dim(V/V_0)-1$. \[cor\_shelling\_order\] If $\lambda=0$, then ${\mathscr{L}}_\lambda={\mathscr{L}}$, so $\Sigma({\mathscr{L}}_\lambda)=\Sigma({\mathscr{L}})$, and the corollary is just Theorem \[thm\_shelling\_order\]. We now assume that $\lambda\not=0$. By Theorem \[thm\_shelling\_order\] and Corollary \[cor\_lower\_order\_ideal\], it suffices to find a family of signs $(\varepsilon_e) \in \{\pm 1\}^E$ such that: - for every $e \in E$, we have $(\lambda,\varepsilon_e\alpha_e)\geq 0$; - there exists a chamber $B\in{\mathscr{L}}_\lambda$ with $B\subset H^+_{\varepsilon_e\alpha_e}$ for every $e \in E$. Indeed, Corollary \[cor\_lower\_order\_ideal\] will then imply that ${\mathscr{T}}\cap{\mathscr{L}}_\lambda$ is a lower order ideal in ${\mathscr{T}}_B$, so it will be an initial segment for at least one linear extension of $\preceq_B$. Let $F = \{e\in E : (\lambda,\alpha_e)\not=0\}$. For every $e\in F$, we choose $\varepsilon_e \in \{\pm 1\}$ such that $(\lambda,\varepsilon_e\alpha_e)>0$. Let $x_0$ be a point in $V$ not on any hyperplane of ${\mathcal{H}}$, that is, $x_0\in V - \bigcup_{e \in E} H_{\alpha_e}$. Then for every $e \in F$, the inner product $(x_0 + c \cdot \lambda,\varepsilon_e\alpha_e) = (x_0,\varepsilon_e\alpha_e)+c \cdot (\lambda,\varepsilon_e\alpha_e) $ tends to $+\infty$ as $c$ tends to $+\infty$, so it is positive for $c$ large enough. Similarly, the inner product $(x_0+c \cdot \lambda,\lambda)=(x_0,\lambda)+c \cdot (\lambda,\lambda)$ is positive for $c$ large enough. On the other hand, if $e\in E - F$, then $(x_0+c \cdot \lambda,\alpha_e)=(x_0,\alpha_e)\not=0$ for every $c\in{{\mathbb{R}}}$. So, if $c\in{{\mathbb{R}}}$ is large enough, then $x=x_0+c \cdot \lambda\in V - \bigcup_{e \in E} H_{\alpha_e}$, and $x$ is in $H^+_{\varepsilon_e\alpha_e}$ for every $e\in F$ and in $H^+_\lambda$. In particular, there exists a chamber $B\in{\mathscr{T}}$ such that $x\in B$, and $B$ is included in $H^+_{\varepsilon_e\alpha_e}$ for every $e\in F$ and in $H^+_\lambda$. Now, if $e\in E - F$, we choose $\varepsilon_e\in\{\pm 1\}$ such that $B\subset H^+_{\varepsilon_e\alpha_e}$. As $(\lambda,\alpha_e)=0$, we clearly have $(\lambda,\varepsilon_e\alpha_e)\geq 0$. Coxeter arrangements {#section_Coxeter_arrangements} -------------------- Let $(W,S)$ be a Coxeter system, that is, $W$ is the group generated by the set $S$ and the relations between the generators are of the form $(st)^{m_{s,t}} = 1$ where $m_{s,s} = 1$ and $m_{s,t} \geq 2$ for $s \neq t$; see [@BB Section 1.1]. The corresponding Coxeter graph has vertex set $S$, and two generators $s$ and $t$ are connected with an edge if $m_{s,t} \geq 3$. If $m_{s,t} \geq 4$ it is customary to label the edge by the integer $m_{s,t}$. There are three natural partial orders on the elements of the Coxeter group $W$. First the [strong Bruhat order]{} is defined by the following cover relation: $z {\prec}w$ if there exists $s\in S$ and $u\in W$ such that $(usu^{-1}) z = w$ and $\ell(z)+1 = \ell(w)$ where $\ell$ is the length function on $W$; see for example [@BB Definition 2.1.1]. Next, we have the [right]{} (respectively [left]{}) [weak Bruhat order]{}, where the cover relation is $z {\prec}w$ if there exists $s\in S$ such that $z \cdot s= w$ (respectively $s \cdot z = w$) and $\ell(z) + 1 =\ell(w)$. The strong Bruhat order refines both the left and right weak Bruhat orders. Let $V=\bigoplus_{s\in S}{{\mathbb{R}}}e_s$, with the symmetric bilinear form $(\cdot,\cdot)$ defined by $$(e_s,e_t)=-\cos\left({\pi}/{m_{s,t}}\right).$$ In particular, $(e_s,e_s) = 1$. The canonical representation of $(W,S)$ is the representation of $W$ on $V$ given by $$\begin{aligned} \label{equation_canonical_representation} s(v) & = v - 2 \cdot (e_s,v) \cdot e_s,\end{aligned}$$ for every $s\in S$ and every $v\in V$. Note that this formula defines an orthogonal isomorphism of $V$ for the symmetric bilinear form $(\cdot,\cdot)$. We refer the reader to [@Bourbaki Chapitre V, § 4, 8, Théorème 2 p. 98] for the next result. \[thm\_W\_finite\] Equation defines a faithful representation of $W$ on $V$, and the form $(\cdot,\cdot)$ is positive definite if and only if $W$ is finite. From now on, we assume that $W$ is finite, and we write $\Phi=\{w(e_s) : w\in W,\ s\in S\}$ and $\Phi^+=\Phi\cap\sum_{s\in S}{{\mathbb{R}}}_{\geq 0} e_s$. The set $\Phi$ is a [pseudo-root system]{}, its subset $\Phi^+$ is the set of [positive pseudo-roots]{}, and the set $\Phi^- = -\Phi^+ = \Phi - \Phi^+$ is the set of [negative pseudo-roots]{}; see Definitions \[def\_pseudo\_root\_system\] and \[def\_positive\_pseudo\_roots\]. Then $\mathcal{H}=(H_\alpha)_{\alpha\in\Phi^+}$ is an essential hyperplane arrangement on $V$. The set of chambers ${\mathscr{T}}$ of this arrangement is in canonical bijection with $W$: the unit element $1\in W$ corresponds to the chamber $B=\bigcap_{\alpha\in\Phi^+}H_\alpha^+=\bigcap_{s\in S}H_{e_s}^+$, and an arbitrary element $w$ of $W$ corresponds to the chamber $w(B)$. More generally, a parabolic subgroup of $W$ is a subgroup $W_I$ generated by a subset $I$ of $S$, and the left cosets of parabolic subgroups of $W$ are called standard cosets. The Coxeter complex $\Sigma(W)$ of $W$ is the set of standard cosets of $W$ ordered by reverse inclusion. It is a simplicial complex, and we have an isomorphism of posets from $\Sigma(W)$ to the face poset ${\mathscr{L}}$ of ${\mathcal{H}}$ sending a standard coset $wW_I$ to the cone $\{x\in V : \forall s\in I,\ (x,w(e_s))=0\text{ and } \forall s\in S-I,\ (x,w(e_s))>0\}$. The fact that this is an isomorphism is proved in [@Bourbaki Chapitre V § 4 6 pp. 96–97], since the representation of $W$ on $V^\vee$ is isomorphic to its canonical representation on $V$ by Theorem \[thm\_W\_finite\]. The fact that $\Sigma(W)$ is a simplicial complex then follows from [@Bourbaki Chapitre V § 3 3 Proposition 7 p. 85]. The definitions of $B$ and of the isomorphism ${\mathscr{T}}\simeq W$ imply that, if $w,w'\in W$ and $T_w,T_{w'}\in{\mathscr{T}}$ are the corresponding chambers, then $$\begin{aligned} S(T_w,T_{w'}) & = \{\alpha\in\Phi^+ : w^{-1}(\alpha) \in \Phi^+ \text{ and } {w'}^{-1}(\alpha) \in \Phi^- \} \\ & \quad\cup \{\alpha\in\Phi^+ : w^{-1}(\alpha) \in \Phi^- \text{ and } {w'}^{-1}(\alpha) \in \Phi^+\} ,\end{aligned}$$ and in particular $$S(B,T_w)=\{\alpha\in\Phi^+ : w^{-1}(\alpha) \in \Phi^-\}.$$ By [@BB Propositions 3.1.3 and 4.4.6] this implies that the isomorphism ${\mathscr{T}}\simeq W$ sends the partial order $\preceq_B$ to the right weak Bruhat order on $W$. Let ${\mathcal{H}}=(H_{\alpha_e})_{e\in E}$ be a finite hyperplane arrangement on a finite-dimensional inner product space $V$, whose inner product is denoted by $(\cdot,\cdot)$. We say that $\mathcal{H}$ is a Coxeter arrangement if $\alpha_e\not\in{{\mathbb{R}}}\alpha_f$ for distinct $e,f\in E$ and if for every $e \in E$ the family of hyperplanes $\mathcal{H}$ is stable by the (orthogonal) reflection $s_{\alpha_e}$ across $H_{\alpha_e}$. \[thm\_Coxeter\_arrangements\] The hyperplane arrangement associated to a Coxeter system with finite Coxeter group is a Coxeter arrangement. Conversely, suppose that ${\mathcal{H}}$ is a Coxeter arrangement on an inner product space $V$, and that there exists a chamber $B$ of ${\mathcal{H}}$ that is on the positive side of each hyperplane in ${\mathcal{H}}$. Let $W$ be the subgroup of ${\mathbf{GL}}(V)$ generated by the set $\{s_{\alpha_e} : e \in E\}$, let $F$ be the set of $e \in E$ such that $\overline{B}\cap H_{\alpha_e}$ is a facet of $\overline{B}$ and let $S=\{s_{\alpha_f} : f \in F\}$. Then $(W,S)$ is a Coxeter system, the group $W$ is finite, and the hyperplane arrangement induced by ${\mathcal{H}}$ on $V/\bigcap_{e\in E}H_{\alpha_e}$ is isomorphic to the arrangement associated to the Coxeter system $(W,S)$. The first statement is an immediate consequence of the definition of the arrangement associated to a Coxeter system. The second and fourth statements follow from [@Bourbaki Chapitre V § 3 2 Théorème 1 p. 74]. The fact that $W$ is finite follows from [@Bourbaki Chapitre V § 3 7 Proposition 4 p. 80] and from the fact that the arrangement ${\mathcal{H}}$ is central. \[lemma\_Coxeter\_A\] Every Coxeter arrangement $\mathcal{H}$ satisfies Condition \[(A)\]. In a Coxeter arrangement, the dihedral angle between any two adjacent facets is $\pi/n$, with $n\geq 2$. In particular, Corollaries \[cor\_lower\_order\_ideal\] and \[cor\_shelling\_order\] apply to Coxeter arrangements. But we can actually prove a stronger result in this case. We fix a Coxeter arrangement ${\mathcal{H}}$ on an inner product space $V$, and we use the notation introduced above. We say that a vector $\lambda\in V$ is dominant if $(\lambda,\alpha)\geq 0$ for every $\alpha\in\Phi^+$. Suppose that $\lambda\in V$ is dominant. Denote by $B$ the chamber of ${\mathcal{H}}$ corresponding to $1\in W$. Let $z,w \in W$ such that $z \leq w$ in the strong Bruhat order of $W$. Then for every $x\in B$ the following inequality holds: $$(z^{-1}(\lambda),x) \geq (w^{-1}(\lambda),x) .$$ \[lemma\_Bruhat\_order\_implies\] We may assume that $w$ covers $z$, so that there exists $s\in S$ and $u\in W$ such that $w=(usu^{-1})z$. Let $\alpha$ be the unique pseudo-root of $\Phi^+$ such that $u(e_s)$ is a multiple (positive or negative) of $\alpha$. If $s_\alpha$ is the reflection across $H_\alpha$, we have $usu^{-1}=s_\alpha$, and so $w=s_\alpha z$ and $s_\alpha w = z$. Since the elements of $\Phi^+$ are unit vectors, $s_\alpha$ is given by the following formula: $s_{\alpha}(\mu) = \mu - 2\cdot (\mu,\alpha) \cdot \alpha$ for $\mu \in V$. Hence $$\begin{aligned} (s_{\alpha} w)^{-1}(\lambda) & = (w^{-1} s_{\alpha})(\lambda) = w^{-1}(\lambda) - 2\cdot (\lambda,\alpha) \cdot w^{-1}(\alpha), \\ \intertext{and so, if $x \in B$,} ((s_\alpha w)^{-1}\lambda,x) & = (w^{-1}(\lambda),x) - 2\cdot(\lambda,\alpha) \cdot (w^{-1}(\alpha),x) .\end{aligned}$$ As $\lambda$ is dominant, we have $(\lambda,\alpha) \geq 0$. By [@BB Equation (4.25)], we have the equivalence $w^{-1}(\alpha) \in \Phi^+ \Longleftrightarrow \ell(w^{-1}s_\alpha) > \ell(w^{-1})$, and [@BB Proposition 1.4.2(iv)] states that $\ell(v^{-1}) = \ell(v)$ for every $v\in W$. Using these two facts and the condition $\ell(s_{\alpha} w)<\ell(w)$, we see that $w^{-1}(\alpha) \in \Phi^-$. Thus $(w^{-1}(\alpha),x) < 0$ by definition of $B$. Hence the term $-2\cdot (\lambda,\alpha) \cdot (w^{-1}(\alpha),x)$ is nonnegative, that is, $(z^{-1}(\lambda),x)=((s_{\alpha} w)^{-1}(\lambda),x) \geq (w^{-1}(\lambda),x)$. \[prop\_strong\_Bruhat\] Let $(W,S)$ be a Coxeter system, and let ${\mathcal{H}}=(H_\alpha)_{\alpha\in\Phi^+}$ be the associated hyperplane arrangement on the space $V$ of the canonical representation of $(W,S)$. Let $\lambda\in V$ be a dominant vector. Then the set $W_\lambda$ of $w\in W$ such that the corresponding chamber of ${\mathcal{H}}$ is in ${\mathscr{T}}\cap{\mathscr{L}}_\lambda$ is a lower order ideal with respect to the strong Bruhat order on $W$. We denote by $B$ the chamber of ${\mathcal{H}}$ corresponding to $1\in W$. By definition of $W_\lambda$, an element $w$ of $W$ is in $W_\lambda$ if and only if for every $x\in B$ we have $(\lambda,w(x))=(w^{-1}(\lambda),x)\geq 0$. By Lemma \[lemma\_Bruhat\_order\_implies\], if $z,w \in W$ and $w$ is greater than $z$ in the strong Bruhat order then for every $x\in B$, we have $(z^{-1}(\lambda),x) \geq (w^{-1}(\lambda),x)$. If moreover $w\in W_\lambda$, this immediately implies that $z \in W_\lambda$. The weighted sum {#section_definition_sum} ================ Definition and first properties ------------------------------- We use the notation of Subsection \[section\_weighted\_complex\]. Given a chamber $B\in{\mathscr{T}}$, we define a function $f_B$ from the face poset ${\mathscr{L}}$ to the set of chambers ${\mathscr{T}}$ by $f_B(C)=C\circ B$. The next proposition gives some basic properties of the function $f_B$. \[prop\_f\_B\] The following two statements hold: - Fix a face $C\in{\mathscr{L}}$, and consider the poset isomorphism $\iota_C:{\mathscr{L}}_{\geq C} \stackrel{\sim}{\longrightarrow} {\mathscr{L}}_{{\mathcal{H}}(C)}$ of Lemma \[lemma\_star\]. If $B\in{\mathscr{T}}\cap{\mathscr{L}}_{\geq C}$ then for every $D\in{\mathscr{L}}_{\geq C}$ we have $\iota_C(f_B(D))=f_{\iota_C(B)}(\iota_C(D))$. - Suppose that ${\mathcal{H}}$ is a Coxeter arrangement with a chamber $B$ that is on the positive side of every hyperplane, and let $(W,S)$ be the associated Coxeter system. Identify ${\mathscr{L}}$ with the Coxeter complex $\Sigma(W)$ as in Section \[section\_Coxeter\_arrangements\]. If the face $C\in{\mathscr{L}}$ corresponds to a standard coset $c\subset W$, then the element $w\in W$ corresponding to the chamber $f_B(C)$ is the shortest element of $c$ which is also the minimal element of the coset $c$ in the right weak Bruhat order. In particular, part (ii) implies that, for the type $A$ Coxeter complex, the function $f_B$ defined here (for $B$ the chamber corresponding to $1\in W$) is equal to the function $f$ defined at the beginning of [@Ehrenborg_Morel_Readdy Section 4]. Note that the existence of a minimal element in every standard coset is proved in [@BB Proposition 2.4.4]. Statement (i) follows immediately from Lemma \[lemma\_star\]. We now prove (ii). By definition of the composition $\circ$, the chamber $f_B(C)=C\circ B$ is the element of ${\mathscr{T}}\cap{\mathscr{L}}_{\geq C}$ closest to $B$ in the chamber graph; in other words, it is the minimal element of ${\mathscr{T}}\cap{\mathscr{L}}_{\geq C}$ for the order $\preceq_B$; see Subsection \[background\]. As we know that $\preceq_B$ corresponds to the right weak Bruhat order on $W$ (see the discussion after Theorem \[thm\_W\_finite\]), and as the elements of $W$ corresponding to the chambers of ${\mathscr{T}}\cap{\mathscr{L}}_{\geq C}$ are the elements of the coset $c$, the result follows. The link between the function $f_B$ and the shellings of Theorem \[thm\_shelling\_order\] is established in the following proposition. For the type $A$ Coxeter complex, this result appeared implicitly in the proof of [@Ehrenborg_Morel_Readdy Proposition 4.1]. Let $B\in{\mathscr{T}}$, and let $B = T_1,T_2,\ldots,T_r = -B$ be a linear order of ${\mathscr{T}}$ refining the partial order $\preceq_B$. Then for every index $1 \leq i \leq r$ the fiber of $f_B$ over $T_i$ is given by $$\begin{aligned} f_B^{-1}(T_i) & = \{C\in{\mathscr{L}}: C\subset\overline{T}_i - (\overline{T}_1 \cup \overline{T}_2 {\cup \cdots \cup}\overline{T}_{i-1})\} \label{equation_fiber_of_f_B_1} \\ & = \{C\in{\mathscr{L}}: C\leq T_i \text{ and }C\not\leq T_j \text{ for }1\leq j<i\}. \label{equation_fiber_of_f_B_2}\end{aligned}$$ \[prop\_fibers\_of\_f\_B\] By Theorem \[thm\_shelling\_order\] the linear order $T_1,T_2,\ldots,T_r$ is a shelling order of the chambers of $\Sigma({\mathscr{L}})$. In particular, the shelling order defines a partition of the faces of $\Sigma({\mathscr{L}})$: $${\mathscr{L}}= \coprod_{i=1}^r \{C\in{\mathscr{L}}: C\leq T_i \text{ and }C\not\leq T_j \text{ for }1\leq j<i\} .$$ Proposition \[prop\_fibers\_of\_f\_B\] says that this partition is independent of the linear refinement $T_1,T_2,\ldots,T_r$ of $\preceq_B$, and that its blocks are the fibers of the map $f_B:{\mathscr{L}}\longrightarrow {\mathscr{T}}$. The equivalence between equalities and is an immediate consequence of the definition of the order $\leq$ on ${\mathscr{L}}$. Let us prove equality . Let $C\in f_B^{-1}(T_i)$, that is, $C\circ B=T_i$. In particular, we have $C\leq C\circ B=T_i$. Suppose that $T\in{\mathscr{T}}$ is another chamber such that $C\leq T$. Then for every $e\in S(B,T_i)$ we have $s(T_i)_e\not=s(B)_e$, but $s(C\circ B)_e=s(T_i)_e$, so $0 \neq s(C)_e=s(T_i)_e$. As $C\leq T$, this implies that $s(T)_e = s(C)_e = s(T_i)_e \neq s(B)_e$, hence that $e\in S(B,T)$. So we have proved that $S(B,T_i)\subseteq S(B,T)$, which means that $T_i\preceq_B T$. In particular, if $1 \leq j \leq i-1$ then $C\not\leq T_j$. Conversely, let $C\in{\mathscr{L}}$ be such that $C\leq T_i$ and $C\not\leq T_j$ for $1\leq j<i$, and let $T=C\circ B$. If $e\in S(B,T)$ then $0\not=s(C)_e=s(T)_e$. As $C\leq T_i$, this implies that $s(C)_e = s(T_i)_e$, so $s(T_i)_e = s(C)_e = s(T)_e \not= s(B)_e$, that is, $e\in S(B,T_i)$. So we have proved that $S(B,T) \subseteq S(B,T_i)$, which means that $T\preceq_B T_i$. Hence there exists an index $1 \leq i' \leq i$ such that $T=T_{i'}$. As $C\leq T$ and $C\not\leq T_j$ for $1\leq j<i$, we must have $i'=i$, that is, $f_B(C) = C\circ B = T = T_i$. We are interested in the following quantity. \[def\_psi\] Let ${\mathcal{H}}$ be a hyperplane arrangement in an inner product space $V$. Let $\lambda$ be a vector in $V$ and $B$ a chamber of ${\mathcal{H}}$, that is, $B \in{\mathscr{T}}$. Recall that ${\mathscr{L}}_\lambda$ is the collection of faces on the nonnegative side of the hyperplane $H_\lambda$. The [weighted sum]{} is defined to be $$\begin{aligned} \psi_{\mathcal{H}}(B,\lambda) & = \sum_{D\in{\mathscr{L}}_\lambda} (-1)^{\dim(D)} \cdot (-1)^{\|S(B,D\circ B)\|}. \label{equation_psi_H}\end{aligned}$$ More generally, if $C$ is a face of the arrangement ${\mathcal{H}}$, that is, $C\in{\mathscr{L}}$, and if $B$ is a chamber whose closure contains the face $C$, that is, $B\in{\mathscr{T}}\cap{\mathscr{L}}_{\geq C}$, we define the weighted sum to be $$\begin{aligned} \psi_{{\mathcal{H}}/C}(B,\lambda) & = \sum_{D\in{\mathscr{L}}_{\lambda,\geq C}} (-1)^{\dim(D)} \cdot (-1)^{\|S(B,D\circ B)\|}. \label{equation_psi_H_C}\end{aligned}$$ In our previous paper, we used the notation $S(\lambda)$ (see [@Ehrenborg_Morel_Readdy Equation (6.1)]) to denote what turns out to be a particular case of the sum in equation in the type $B$ Coxeter case; see equation in Subsection \[second\_application\] for the precise relation between the two. In this paper, we decided to follow the notation of [@GKM] in order to avoid overusing the letter $S$. Let $C\in{\mathscr{L}}$ and $B\in{\mathscr{T}}\cap{\mathscr{L}}_{\geq C}$. Suppose that ${\mathscr{T}}\cap{\mathscr{L}}_{\lambda,\geq C}$ is a lower order ideal in the chamber poset ${\mathscr{T}}_B$. Then the following identity holds: $$\psi_{{\mathcal{H}}/C}(B,\lambda) = \begin{cases} (-1)^{\dim(V)+\|E(C)\|} & \text{ if } {\mathscr{L}}_{\lambda,\geq C}={\mathscr{L}}_{\geq C}, \\ 0 & \text{ if } {\mathscr{L}}_{\lambda,\geq C} \subsetneq{\mathscr{L}}_{\geq C} . \end{cases}$$ \[prop\_base\_case\_for\_psi\] In particular, if $C$ is the minimal face $V_0$ in ${\mathscr{L}}$, then we have ${\mathscr{L}}_{\geq V_0}={\mathscr{L}}$. Hence the condition ${\mathscr{L}}_{\lambda,\geq V_0}={\mathscr{L}}_{\geq V_0}$ is equivalent to $\lambda=0$. If the hyperplane arrangement satisfies Condition \[(A)\], then we can use Corollary \[cor\_lower\_order\_ideal\] to verify the hypothesis of Proposition \[prop\_base\_case\_for\_psi\]. However, Proposition \[prop\_base\_case\_for\_psi\] holds in greater generality than in the case where the somewhat unnatural angle requirement of Condition \[(A)\] is satisfied. We use the notation of Lemma \[lemma\_star\]. In particular, we get a subarrangement ${\mathcal{H}}(C)$ of ${\mathcal{H}}$ whose hyperplanes are the hyperplanes of ${\mathcal{H}}$ containing $C$ and an isomorphism of posets $\iota_C:{\mathscr{L}}_{\geq C} \stackrel{\sim}{\longrightarrow} {\mathscr{L}}_{{\mathcal{H}}(C)}$. Let $B'=\iota_C(B)$. By part (ii) of Lemma \[lemma\_star\] this isomorphism induces an isomorphism of posets ${\mathscr{T}}_B\cap{\mathscr{L}}_{\geq C} \stackrel{\sim}{\longrightarrow} {\mathscr{T}}_{{\mathcal{H}}(C),B'}$, where ${\mathscr{T}}_{{\mathcal{H}}(C)}$ is the set of chambers of ${\mathcal{H}}(C)$. Choose a linear order $T_1,T_2,\ldots,T_r$ of ${\mathscr{T}}_{{\mathcal{H}}(C),B'}$ refining $\preceq_{B'}$. As $\iota_C({\mathscr{T}}\cap{\mathscr{L}}_{\lambda,\geq C})$ is a lower order ideal in ${\mathscr{T}}_{{\mathcal{H}}(C),B'}$ by assumption and by the previous paragraph, we may choose this linear order so that $\iota_C({\mathscr{T}}\cap {\mathscr{L}}_{\lambda,\geq C}) = \{T_1, T_2, \ldots, T_s\}$, for some $1 \leq s \leq r$. By Theorem \[thm\_shelling\_order\] the chosen linear order on ${\mathscr{T}}_{{\mathcal{H}}(C),B'}$ is a shelling order, and by Proposition \[prop\_fibers\_of\_f\_B\] the corresponding partition of ${\mathscr{L}}_{{\mathcal{H}}(C)}$ is $${\mathscr{L}}_{{\mathcal{H}}(C)}=\coprod_{i=1}^r f_{B'}^{-1}(T_i).$$ In particular, $\coprod_{i=1}^s f_{B'}^{-1}(T_i)$ is the set of faces of $\overline{T_1}\cup\overline{T_2}\cup \cdots\cup\overline{T_s}$, which is also equal to $\iota_C({\mathscr{L}}_{\lambda,\geq C})$ by Proposition \[prop\_weighted\_complex\_basic\]. As the function $D\longmapsto(-1)^{\|S(B',\iota_C(D)\circ B)'\|}$ is constant and equal to $(-1)^{\|S(B',T)\|}$ on the fiber $f_{B'}^{-1}(T)$ for every $T\in{\mathscr{T}}_{{\mathcal{H}}(C)}$ and as the isomorphism $\iota_C:{\mathscr{L}}_{\geq C} \stackrel{\sim}{\longrightarrow} {\mathscr{L}}_{{\mathcal{H}}(C)}$ preserves dimensions (by part (iii) of Lemma \[lemma\_star\]), we obtain $$\begin{aligned} \psi_{{\mathcal{H}}/C}(B,\lambda) & = \sum_{i=1}^s (-1)^{\|f_{B'}(T_i)\|} \cdot \sum_{D\in f_{B'}^{-1}(T_i)}(-1)^{\dim(D)}. \label{equation_psi_inner_sum}\end{aligned}$$ Let $a_{i}$ be the inner sum of , that is, $a_i = \sum_{D\in f_{B'}^{-1}(T_i)} (-1)^{\dim(D)}$ for $1 \leq i \leq r$. Our tool to evaluate $a_{i}$ is the [Euler characteristic with compact support]{}, denoted by $\chi_{c}$. This is defined for example in van den Dries’s book [@van_den_Dries Section 2 of Chapter 4], where it is simply called “Euler characteristic”. The connection with cohomology with compact support is explained in Schürmann’s book [@Schurmann Section 2.0.1, page 84]. The properties that will be most relevant to us are the facts that $\chi_c(A)=(-1)^d$ if $A$ is an open ball of dimension $d$, and that $\chi_{c}$ is a valuation, that is, $\chi_{c}(A) = \chi_{c}(A) + \chi_{c}(A\setminus A')$ for $A$ a nice enough topological space and $A'$ a nice enough subspace of $A$. Here “nice enough” means that $A$ is locally compact and has a stratification by locally closed subsets called cells which are all homeomorphic to open balls, and $A'$ is a union of some of the cells of $A$. First consider the case $i=1$. Then all the cones $D$ in $f_{B'}^{-1}(T_1)$ except for the minimal cone, which is $\operatorname{Span}(C)$, intersect the unit sphere $\mathbb{S}$ in $V$ in an open ball of dimension $\dim(D)-1$. In particular, we have $(-1)^{\dim(D)}=-\chi_c(D\cap\mathbb{S})$. As for $\operatorname{Span}(C)$, its intersection with $\mathbb{S}$ is a sphere of dimension $\dim(C)-1$. Note that if $C=\{0\}$ then this intersection is empty so it is a sphere of dimension $-1$. We have $(-1)^{\dim(C)}=1-\chi_c(\operatorname{Span}(C)\cap\mathbb{S})$. Hence we obtain $$\begin{aligned} a_1 & = 1- \sum_{D\in f_{B'}^{-1}(T_1)} \chi_c(D\cap\mathbb{S}) = 1 - \chi_c(\overline{T_1}\cap\mathbb{S}) .\end{aligned}$$ But $\overline{T_1} \cap \mathbb{S}$ is a closed ball of dimension $\dim(V)-1$, hence $\chi_c(\overline{T_1}\cap\mathbb{S})=1$ and so $a_{1} = 0$. Suppose now that $i\geq 2$. Then we have $$\begin{aligned} a_i & =-\sum_{D\in f_{B'}^{-1}(T_i)}\chi_c(D\cap\mathbb{S}) = -\chi_c(\overline{T_i}\cap\mathbb{S})+\chi_c\left(\overline{T_i}\cap \left(\overline{T_1}\cup\overline{T_2}\cup\cdots\cup\overline{T_{i-1}}\right) \cap\mathbb{S}\right).\end{aligned}$$ Again $\overline{T_i}\cap\mathbb{S}$ is a closed ball of dimension $\dim(V)-1$, hence $\chi_c(\overline{T_i}\cap\mathbb{S})=1$. On the other hand, $\overline{T_i}\cap(\bigcup_{j=1}^{i-1}\overline{T_j}) \cap\mathbb{S}$ is a partial shelling of the boundary of $\overline{T_i}\cap\mathbb{S}$. As $\Sigma({\mathscr{L}}_{{\mathcal{H}}(C)})$ is a sphere, $T_r$ is the only homology facet in the shelling given by the order $T_1,T_2,\ldots,T_r$. So if $i\leq r-1$ then $\overline{T_i}\cap(\bigcup_{j=1}^{i-1}\overline{T_j})\cap\mathbb{S}$ is not equal to $\partial\overline{T_i}\cap\mathbb{S}$, so it is a closed ball of dimension $\dim(V)-2$. In particular $\chi_c(\overline{T_i}\cap(\bigcup_{j=1}^{i-1}\overline{T_j})\cap\mathbb{S})=1$ and thus $a_i=0$. Finally, if $i=r$ then $\overline{T_r}\cap(\bigcup_{j=1}^{r-1}\overline{T_j})\cap\mathbb{S}$ is equal to the boundary of $\overline{T_r}\cap\mathbb{S}$, so $a_r=-\chi_c(T_r\cap\mathbb{S})=(-1)^{\dim(V)}$. As $\psi_{{\mathcal{H}}/C}(B,\lambda)=\sum_{i=1}^s (-1)^{\|S(B',T_i)\|}a_i$, this shows that $\psi_{{\mathcal{H}}/C}(B,\lambda)=0$ unless $s=r$. But $s=r$ if and only if ${\mathscr{L}}_{\geq C}={\mathscr{L}}_{\lambda,\geq C}$. In this case, we have $\psi_{{\mathcal{H}},C}(B,\lambda)=(-1)^{\|S(B',T_r)\|} \cdot a_r= (-1)^{\|S(B',T_r)\| + \dim(V)}$. Finally, observe that the last facet $T_{r}$ is given by $-B'$, so $S(B',-B')$ is the total index set of hyperplanes in ${\mathcal{H}}(C)$, that is, $E(C)$, and so $\psi_{{\mathcal{H}}/C}(B,\lambda)=(-1)^{\dim(V)+\|E(C)\|}$. We state the following lemma, which reduces the calculation of $\psi_{{\mathcal{H}}/C}(B,\lambda)$ to the case of an essential arrangement. Recall that $V_0$ is the intersection of all the hyperplanes of ${\mathcal{H}}$, that is, $V_0=\bigcap_{e\in E}H_{\alpha_e}$. Let $\pi$ denote the projection $V \longrightarrow V/V_0$. Let ${\mathcal{H}}/V_0$ be the hyperplane arrangement $(H_{\alpha_e}/V_0)_{e\in E}$ on V/$V_0$. Note that $\pi$ induces an isomorphism between the posets of faces of ${\mathcal{H}}$ and of ${\mathcal{H}}/V_0$. Let $C\in{\mathscr{L}}$, let $B$ be a chamber of ${\mathcal{H}}$ such that $B\in{\mathscr{L}}_{\geq C}$ and let $\lambda\in V$. Then the following identity holds: $$\psi_{{\mathcal{H}}/C}(B,\lambda) = \begin{cases} (-1)^{\dim(V_0)} \cdot \psi_{({\mathcal{H}}/V_0)/\pi(C)}(\pi(B),\pi(\lambda)) & \text{ if }\lambda\in V_0^\perp, \\ 0 & \text{ if } \lambda\not\in V_0^\perp . \end{cases}$$ \[lemma\_psi\_inessential\_arrangement\] Note that $\lambda\in V_0^\perp$ if and only if $V_0\subset H_\lambda$. If $\lambda\not\in V_0^\perp$ then the linear functional $(\lambda,\cdot)$ takes both positive and negative values on $V_0$. As $V_0\subset\overline{D}$ for every $D\in{\mathscr{L}}$, this linear functional also takes both positive and negative values on $D$, so $D\not\in{\mathscr{L}}_\lambda$. This shows that ${\mathscr{L}}_\lambda=\varnothing$ if $\lambda\not\in V_0^\perp$ and gives the second case. Now suppose that $\lambda\in V_0^\perp$. Then $V_0\subset H_\lambda$, and it is easy to see that $D\in{\mathscr{L}}_\lambda$, respectively $D\in{\mathscr{L}}_{\geq C}$, if and only if $\pi(D)\subset \overline{H^+_{\pi(\lambda)}}$, respectively $\pi(D)\geq\pi(C)$, and that $\dim(\pi(D))=\dim(D)-\dim(V_0)$. This yields the first case. Suppose that $V=V_1\times\cdots\times V_r$ with the $V_i$ mutually orthogonal subspaces of $V$ and that ${\mathcal{H}}$ also decomposes as a product ${\mathcal{H}}_1\times\cdots\times{\mathcal{H}}_r$. By this, we mean that there is a decomposition $E=E_1\sqcup\cdots\sqcup E_r$ such that, for $i = 1, \ldots, r$ and every $e\in E_i$, we have $\alpha_{e} \in V_i$. The arrangement ${\mathcal{H}}_i=(V_i\cap H_{\alpha_e})_{e\in E_i}$ is a hyperplane arrangement on the subspace $V_i$, and each hyperplane of ${\mathcal{H}}$ is of the form $H \times\prod_{j\not=i}V_j$, where $i = 1, \ldots, r$ and $H$ is one of the hyperplanes of ${\mathcal{H}}_i$. Let ${\mathscr{L}}_i$ be the face poset of ${\mathcal{H}}_i$ for $i = 1, \ldots, r$. Then the faces of ${\mathscr{L}}$ are exactly the products $C_1\times\cdots \times C_r$ where $C_i\in{\mathscr{L}}_i$, and the order on ${\mathscr{L}}$ is the product order. In particular, $C$ is a chamber in ${\mathscr{L}}$ if and only if all the $C_i$ are chambers in ${\mathscr{L}}_i$. \[lemma\_product\] Assume that the arrangement ${\mathcal{H}}$ factors as ${\mathcal{H}}_1\times\cdots\times{\mathcal{H}}_r$ as described in the two previous paragraphs. Let $C=C_1\times\cdots\times C_r$ be a face in ${\mathscr{L}}$, and let $B=B_1\times\cdots\times B_r$ be a chamber in ${\mathscr{L}}_{\geq C}$. Finally, let $\lambda\in V$. Then $$\psi_{{\mathcal{H}}/C}(B,\lambda)=\prod_{i=1}^r\psi_{{\mathcal{H}}_i/C_i}(B_i,\lambda_i),$$ where, for $i = 1, \ldots, r$, $\lambda_i$ is the orthogonal projection of $\lambda$ on $V_i$. The expression for $\psi_{{\mathcal{H}}/C}(B,\lambda)$ follows from the fact that ${\mathscr{L}}={\mathscr{L}}_1\times\cdots\times{\mathscr{L}}_r$ as posets once we prove the following statement: Let $D=D_1\times\cdots\times D_r\in{\mathscr{L}}$, with $D_i\in{\mathscr{L}}_i$. Then $D\in{\mathscr{L}}_\lambda$ if and only if $D_i\in {\mathscr{L}}_{i,\lambda_i}$ for $i = 1, \ldots, r$. We prove this last fact. Note that $\lambda=(\lambda_1,\ldots,\lambda_r)$ in $V_1\times\cdots\times V_r=V$ because the $V_i$ are pairwise orthogonal. If $D_i\in{\mathscr{L}}_{i,\lambda_i}$ for every $i = 1,\ldots,r$ then for every $x=(x_1,\ldots,x_r)\in V$ we have $(\lambda,x)=\sum_{i=1}^r(\lambda_i,x_i)\geq 0$. Conversely, suppose that $D\in{\mathscr{L}}_{\lambda}$. Let $x_j\in D_j$ for $j = 1, \ldots, r$. As all the $D_j$ are cones, for every $\varepsilon>0$, the element $\varepsilon x_j$ is in $D_j$. Fix $i\in\{1,\ldots,r\}$, and consider the element $x_\varepsilon= (\varepsilon x_1,\ldots,\varepsilon x_{i-1},x_i,\varepsilon x_{i+1},\ldots,\varepsilon x_r)$. Then $x_\varepsilon$ is in $D$. Thus we have the inequality $0 \leq (\lambda,x_\varepsilon) =(\lambda_i,x_i)+\varepsilon\sum_{j\not=i} (\lambda_j,x_j)$. Letting $\varepsilon$ tend to $0$, we obtain $(\lambda_i,x_i)\geq 0$. So $D_i\in {\mathscr{L}}_{i,\lambda_i}$. A recursive expression ---------------------- The goal of this subsection is to prove a recursive expression for $\psi_{{\mathcal{H}}/C}(B,\lambda)$ in terms of a lower-dimensional hyperplane arrangement. Let $V'\subset V$ be a linear subspace of codimension $2$. We define the intersection multiplicity of the arrangement ${\mathcal{H}}$ at $V'$ to be the cardinality $$\|\{e\in E : H_{\alpha_e} \supseteq V'\}\|.$$ Here is the main result of this subsection. Let $C$ be a face in ${\mathscr{L}}$, $B$ a chamber in ${\mathscr{T}}\cap {\mathscr{L}}_{\geq C}$ and $\lambda \in V$. Fix $e_0\in E$ such that $H_0=H_{\alpha_{e_0}}$ contains $C$ and is a wall of $B$, that is, the intersection $H_0\cap\overline{B}$ is the closure of a facet $A'$ of $B$. Let $B'\in{\mathscr{T}}$ be the chamber on the other side of $A'$ from $B$, that is, $B'=A'\circ(-B)$. Assume that $\alpha_e\not\in{{\mathbb{R}}}\alpha_{e_0}$ (which implies that $H_e\not=H_0$) for every $e\in E - \{e_0\}$. - The following identity holds: $$\psi_{{\mathcal{H}}/C}(B,\lambda) + \psi_{{\mathcal{H}}/C}(B',\lambda) = 2 \cdot \psi_{{\mathcal{H}}'_0/C}(A',\mu),$$ where ${\mathcal{H}}'_0$ is the hyperplane arrangement in $H_0$ defined by the family of orthogonal projections on $H_0$ of the $\alpha_e$, $e\in E - \{e_0\}$, and $\mu\in H_0$ is the orthogonal projection of $\lambda\in V$. Note that, by our hypothesis on $H_0$, the orthogonal projection of $\alpha_e$ on $H_0$ is nonzero for every $e\in E - \{e_0\}$, and that $C$ is also a face of ${\mathcal{H}}'_0$. - Let $E(C)=\{e\in E : C\subset H_{\alpha_e}\}$. Note that $e_0\in E(C)$ by assumption. Let $F_1$ be the set of $f\in E(C) - \{e_0\}$ such that the intersection multiplicity of ${\mathcal{H}}$ at $H_0\cap H_{\alpha_f}$ is even. Choose a subset $F$ of $F_1$ such that, for every $f_1\in F_1$, there is exactly one $f\in F$ with $H_0\cap H_{\alpha_{f_1}}=H_0\cap H_{\alpha_f}$. Let ${\mathcal{H}}_0$ be the hyperplane arrangement in $H_0$ defined by the family of orthogonal projections on $H_0$ of the $\alpha_f$, for $f\in F\sqcup(E-E(C))$. Since ${\mathcal{H}}_0$ is a subarrangement of ${\mathcal{H}}'_0$, there is a unique chamber of ${\mathcal{H}}_0$ containing $A'$, respectively $C$. Denote this chamber by $A$, respectively $C_0$. Then the following identity holds: $$\psi_{{\mathcal{H}}/C}(B,\lambda) + \psi_{{\mathcal{H}}/C}(B',\lambda) = 2 \cdot \psi_{{\mathcal{H}}_0/C_0}(A,\mu).$$ \[prop\_induction\_formula\_for\_psi\] at (-1,10) [(a)]{}; at (0,10.5) [${\mathcal{H}}$]{}; (0,10) – node\[very near end,above\] [$H_{0}$]{} (10,10); (1.5,11.5) – (4.5,8.5); (4.5,11.5) – (1.5,8.5); (5.5,11.5) – (8.5,8.5); (8.5,11.5) – (5.5,8.5); (7,11.5) – (7,8.5); at (-1,6.5) [(b)]{}; at (0,7) [${\mathcal{H}}'_{0}$]{}; (0,6.5) – (10,6.5); (3,6.5) circle (2pt); (3,6.66) circle (2pt); (7,6.5) circle (2pt); (7,6.66) circle (2pt); (7,6.34) circle (2pt); at (-1,4.5) [(c)]{}; at (0,5) [${\mathcal{H}}_{0}$]{}; (0,4.5) – (10,4.5); (7,4.5) circle (2pt); The proof of this proposition will occupy the rest of this subsection. The following definition will be useful. Let $C\in{\mathscr{L}}$ and $\lambda\in V$. If $D,D'\in{\mathscr{L}}_{\geq C}$, we define $\psi_{D/C}(D',\lambda)$ by the sum $$\psi_{D/C}(D',\lambda)=\sum_{\substack{C'\in{\mathscr{L}}_{\lambda,\geq C} \\ C'\circ D'\leq D}}(-1)^{\dim(C')}.$$ Suppose that $D'$ is a chamber. Then $C'\circ D'$ is a chamber for every $C'\in{\mathscr{L}}$, so $\psi_{D/C}(D',\lambda)=0$ unless $D$ is also a chamber. If $D$ is a chamber, we have $$\psi_{D/C}(D',\lambda)=\sum_{\substack{C'\in{\mathscr{L}}_{\lambda,\geq C} \\ C'\circ D'=D}}(-1)^{\dim(C')}.$$ \[rmk\_psi\_for\_chambers\] The functions $\psi_{D/C}(D',\lambda)$ are related to $\psi_{{\mathcal{H}}/C}(B,\lambda)$ by the following lemma. Let $C\in{\mathscr{L}}$ and $B\in{\mathscr{T}}\cap{\mathscr{L}}_{\geq C}$. Then for every $\lambda \in V$ the following identity holds: $$\psi_{{\mathcal{H}}/C}(B,\lambda) = \sum_{T\in{\mathscr{T}}\cap{\mathscr{L}}_{\geq C}}(-1)^{\|S(B,T)\|} \cdot \psi_{T/C}(B,\lambda).$$ \[lemma\_psi\_vs\_psi\] Indeed, if $D\in{\mathscr{L}}_{\geq C}$ then the chamber $D\circ B$ is also in ${\mathscr{L}}_{\geq C}$. Hence, using in Definition \[def\_psi\] and Remark \[rmk\_psi\_for\_chambers\], we obtain: $$\begin{aligned} \psi_{{\mathcal{H}}/C}(B,\lambda) & = \sum_{T\in{\mathscr{T}}\cap{\mathscr{L}}_{\geq C}}(-1)^{\|S(B,T)\|} \cdot \sum_{\substack{D\in{\mathscr{L}}_{\lambda,\geq C} \\ D\circ B=T}}(-1)^{\dim(D)} = \sum_{T\in{\mathscr{T}}\cap{\mathscr{L}}_{\geq C}} (-1)^{\|S(B,T)\|} \cdot \psi_{T/C}(B,\lambda). $$ Before Corollary \[cor\_psi\_C\_is\_a\_valuation\] of Appendix \[appendix\_valuations\], we define, for $K$ a closed convex polyhedral cone in $V$, a function $\psi_K:V\times V^\vee \longrightarrow {{\mathbb{R}}}$. For fixed $(x,\ell)\in V\times V^\vee$, the function $K \longmapsto\psi_K(x,\ell)$ is a valuation on the set of closed convex polyhedral cones in $V$ (see Definition \[def\_valuation\]). This function is related to the functions $\psi_{D/C}(D',\lambda)$ in the following way. Let $C\in{\mathscr{L}}$, let $D\in{\mathscr{L}}_{\geq C}$ and let $\lambda\in V$. Denote by $\ell\in V^\vee$ the linear functional $(\cdot,\lambda)$. Then for every $D'\in{\mathscr{L}}_{\geq C}$ the following identity holds: $$\psi_{D/C}(D',\lambda) = \psi_{\overline{D}}(x,\ell),$$ where $x$ is any point in $D'_1=(-C)\circ D'$. \[lemma\_on\_the\_definition\_of\_psi\] As before we write $E(C)=\{e\in E : C\subset H_{\alpha_e}\}$. Note that $s(D'_1)_e=s(D')_e$ for $e\in E(C)$, and $s(D'_1)=-s(C)_e\not=0$ for $e\in E-E(C)$. Also, by definition of ${\mathscr{L}}_{\geq C}$, if $e$ is any index of $E - E(C)$ and $C'\in{\mathscr{L}}_{\geq C}$ then $s(C)_e=s(C')_e\not=0$. We claim that, for every $C'\in{\mathscr{L}}$, we have $C'\circ D'_1\leq D$ if and only if $C'\in{\mathscr{L}}_{\geq C}$ and $C'\circ D'\leq D$. Indeed, suppose first that $C'\in{\mathscr{L}}_{\geq C}$ and $C'\circ D'\leq D$. Then for every $e\in E(C)$ we have $s(C'\circ D'_1)_e=s(C'\circ D')_e\leq s(D)_e$. Moreover, if $e\in E-E(C)$ then $s(C')_e=s(C)_e=s(D)_e\not=0$, so $s(C'\circ D'_1)_e=s(C')_e=s(D)_e$. This shows that $C'\circ D'_1\leq D$. Conversely, suppose that $C'$ is a face of ${\mathscr{L}}$ such that $C'\circ D'_1\leq D$. If $e\in E-E(C)$ then $0\not=s(C)_e=s(D)_e=-s(D'_1)_e$, thus $s(C')_e\not=0$, and so $s(C')_e=s(C'\circ D'_1)_e=s(D)_e$. This implies that $C'\in{\mathscr{L}}_{\geq C}$. Moreover, if $e\in E(C)$ we have $s(D'_1)_e=s(D')_e$, thus $s(C'\circ D')_e=s(C'\circ D'_1)_e\leq s(D)_e$. Hence we conclude that $C'\circ D'\leq D$. By the claim, we obtain $$\psi_{D/C}(D',\lambda) = \sum_{\substack{C'\in{\mathscr{L}}_{\lambda} \\ C'\circ D'_1\leq D}} (-1)^{\dim(C')} = \psi_D(D'_1,\lambda) .$$ We wish to show that this is equal to $\psi_{\overline{D}}(x,\ell)$, if $x\in D'_1$. As in Appendix \[appendix\_valuations\], we denote by $\mathcal{F}(\overline{D})$ the set of faces of the closed convex polyhedral cone $\overline{D}$. We have $\mathcal{F}(\overline{D})=\{\overline{C'}\ :\ C'\in{\mathscr{L}},\ C'\leq D\}$, and the set $\{C'\in{\mathscr{L}}: C'\circ D'_1\leq D\}$ is included in the set $\{C'\in{\mathscr{L}}: C'\leq D\}$. To prove the equality above, it suffices to show that the two following statements hold for $C'\in{\mathscr{L}}$ such that $C'\leq D$ (see Lemma \[lemma\_GKM\] for the definition of $\psi_x$ and $\psi_\ell$, and the beginning of Section \[subsection\_signed\_convolution\] for the notation $\overline{C'}^{\perp,\overline{D}}$): - The face $C'$ belongs to ${\mathscr{L}}_\lambda$ if and only $\psi_\ell(\overline{C'})=1$. - The inequality $C'\circ D'_1\leq D$ holds if and only if $\psi_x\left(\overline{C'}^{\perp,\overline{D}}\right)=1$. Statement (a) is just a direct translation of the definition of ${\mathscr{L}}_\lambda$. We prove (b). The assertion that $C'\circ D'_1\leq D$ is equivalent to the fact that, for every $y\in C'$ and every sufficiently small $\varepsilon>0$, we have $y+\varepsilon x\in\overline{D}$. On the other hand, the assertion that $\psi_x(\overline{C'}^{\perp,\overline{D}})=1$ is equivalent to the fact that $x\in(\overline{C'}^{\perp,\overline{D}})^=\overline{D}+\operatorname{Span}(C')$. (Note that $\operatorname{Span}(\overline{C'})=\operatorname{Span}(C')$ because $\operatorname{Span}(C')$ is a vector subspace of $V$, hence closed in $V$.) If $y+\varepsilon x\in\overline{D}$ for some $y\in C'$ and some $\varepsilon>0$, then we clearly have $x\in\overline{D}+\operatorname{Span}(C')$. Conversely, suppose that $x\in\overline{D}+\operatorname{Span}(C')$. Write $x=a+b$, with $a\in\overline{D}$ and $b\in\operatorname{Span}(C')$. Let $y\in C'$. If $\varepsilon>0$ is sufficiently small, then $y+\varepsilon b\in C'$, so $y+\varepsilon x\in\overline{D}+C'=\overline{D}$. We are now ready to prove the main result of this section. Note that $B'$ is also in ${\mathscr{L}}_{\geq C}$, because we chose $e_0$ such that $C \subseteq H_{\alpha_{e_0}}$. By Lemma \[lemma\_psi\_vs\_psi\], we have $$\psi_{{\mathcal{H}}/C}(B,\lambda) = \sum_{T\in{\mathscr{T}}\cap{\mathscr{L}}_{\geq C}} (-1)^{\|S(B,T)\|} \cdot \psi_{T/C}(B,\lambda).$$ Similarly, $$\psi_{{\mathcal{H}}/C}(B',\lambda)=\sum_{T\in{\mathscr{T}}\cap{\mathscr{L}}_{\geq C}} (-1)^{\|S(B',T)\|} \cdot \psi_{T/C}(B',\lambda)$$ also holds. Fix $T\in{\mathscr{T}}\cap{\mathscr{L}}_{\geq C}$. We wish to evaluate the sum $$a(T) = (-1)^{\|S(B,T)\|} \cdot \psi_{T/C}(B,\lambda) +(-1)^{\|S(B',T)\|} \cdot \psi_{T/C}(B',\lambda) .$$ Suppose first that $e_0\not\in S(B,T)$. Then $S(B',T)$ is the disjoint union $S(B,T)\sqcup\{e_0\}$. Let $D\in{\mathscr{L}}$. Then $D\circ B=D\circ B'$ if and only if $s(D)_{e_0}\not=0$. Moreover, if $s(D)_{e_0}=0$ then $D\circ B'$ cannot be equal to $T$, because they are on opposite sides of the hyperplane $H_0$. This implies that $$\begin{aligned} a(T) &= \sum_{\substack{D\in{\mathscr{L}}_{\lambda,\geq C} \\ D\circ B=T \\ D\circ B'=T}} \left((-1)^{\|S(B,T)\|}+(-1)^{\|S(B,T')\|}\right) \cdot (-1)^{\dim(D)} + \sum_{\substack{D\in{\mathscr{L}}_{\lambda,\geq C} \\ s(D)_{e_0}=0 \\D\circ B=T}} (-1)^{\|S(B,T)\|} \cdot (-1)^{\dim(D)} \\ & = \sum_{\substack{D\in{\mathscr{L}}_{\lambda,\geq C} \\ s(D)_{e_0}=0 \\ D\circ B=T}} (-1)^{\|S(B,T)\|} \cdot (-1)^{\dim(D)}.\end{aligned}$$ Let ${\mathscr{L}}_0'$ be the face poset of ${\mathcal{H}}'_0$. Then the following equalities hold: $${\mathscr{L}}'_0=\{D\in{\mathscr{L}}: D\subset H_0\}=\{D\in{\mathscr{L}}: s(D)_{e_0}=0\}.$$ In particular, the cone $C$ is also a face in ${\mathscr{L}}'_0$. Let $D\in{\mathscr{L}}$ such that $D\subset H_0$, and let us still denote by $D$ the corresponding face of ${\mathscr{L}}'_0$. Then $D\in{\mathscr{L}}'_{0,\mu}$ if and only if $D\in{\mathscr{L}}_\lambda$, and $D\in{\mathscr{L}}'_{0,\geq C}$ if and only if $D\in{\mathscr{L}}_{0,\geq C}$. Also, $D\circ A'$ is a chamber of ${\mathcal{H}}'_0$, and it is on the same side of $H_{\alpha_e}$ as $D\circ B$ for every $e\in E - \{e_0\}$, so we can only have $D\circ B=T$ if $H_0$ is a wall of $T$, that is, if $T$ has a facet $T'$ such that $T'\subset H_0$. In this case, we have $D\circ B=T$ if and only if $D\circ A'=T'$. We conclude that $a(T)=0$ unless $H_0$ is a wall of $T$. Moreover, if $H_0$ is a wall of $T$ and $T'$ is as before the unique facet of $T$ such that $T'\subset H_0$, then we have $T'\in{\mathscr{L}}'_{0,\geq C}$ and $S(A',T')=S(B,T)$. In this case we have $$\begin{aligned} a(T) & = \sum_{\substack{D\in{\mathscr{L}}'_{0,\mu,\geq C} \\ D\circ A'=T'}} (-1)^{\|S(A',T')\|} \cdot (-1)^{\dim(D)} = (-1)^{\|S(A',T')\|} \cdot \psi_{T'/C}(A',\mu). \label{equation_a_I}\end{aligned}$$ If $e_0\in S(B,T)$ then $e_0\not\in S(B',T)$. Exchanging the roles of $B$ and $B'$ in the previous paragraph, and noting that $A'$ is also the unique facet of $B'$ contained in $H_0$, we see that $a(T)=0$ unless $H_0$ is a wall of $T$. Moreover, if $H_0$ is a wall of $T$ and $T'$ is as before the unique facet of $T$ such that $T'\subset H_0$, then we have $T'\in{\mathscr{L}}'_{0,\geq C}$ and $S(A',T')=S(B',T)$, and thus the expression for $a(T)$ is given by the same formula as in equation . Finally, each chamber $T'$ of ${\mathcal{H}}'_0$ appears as a facet of exactly two chambers of ${\mathcal{H}}$, one on either side of $H_0$. Hence it contributes twice the right-hand side of equation . Summing over $T'$ in the set of chambers ${\mathscr{T}}'_0$ of ${\mathcal{H}}'_0$ and using Lemma \[lemma\_psi\_vs\_psi\] again, we obtain $$\begin{aligned} \psi_{{\mathcal{H}}/C}(B,\lambda)+\psi_{{\mathcal{H}}/C}(B',\lambda) & = \sum_{T\in{\mathscr{T}}\cap{\mathscr{L}}_{\geq C}} a(T) \\ & = 2 \cdot \sum_{T'\in{\mathscr{T}}'_0\cap{\mathscr{L}}'_{0,\geq C}} (-1)^{\|S(A',T')\|} \cdot \psi_{T'/C}(A',\mu)\\ & = 2 \cdot \psi_{{\mathcal{H}}'_0/C}(A',\mu). $$ By (i), it suffices to prove that $\psi_{{\mathcal{H}}_0'/C}(A',\mu)= \psi_{{\mathcal{H}}_0/C_0}(A,\mu)$. As in the proof of (i), we denote by ${\mathscr{L}}'_0$, respectively ${\mathscr{T}}_0'$, the set of faces, respectively chambers, of ${\mathcal{H}}_0'$. We also denote by ${\mathscr{L}}_0$, respectively ${\mathscr{T}}_0$, the set of faces, respectively chambers, of ${\mathcal{H}}_0$. As ${\mathcal{H}}_0$ is a subarrangement of ${\mathcal{H}}_0'$, each chamber $T'\in{\mathscr{T}}'_0$ is contained in a unique chamber $T\in{\mathscr{T}}_0$. So, using Lemma \[lemma\_psi\_vs\_psi\], we can write $$\psi_{{\mathcal{H}}'_0/C}(A',\mu) = \sum_{T\in{\mathscr{T}}_0}\ \sum_{\substack{T'\in{\mathscr{T}}'_0\cap{\mathscr{L}}'_{0,\geq C} \\ T'\subseteq T}} (-1)^{\|S(A',T')\|} \cdot \psi_{T'/C}(A',\mu).$$ Fix $T\in{\mathscr{T}}_0$. We claim that, for every $T'\in{\mathscr{T}}'_0$ such that $T'\subset T$, the cardinalities $\|S(A',T')\|$ and $\|S(A,T)\|$ have the same parity. Let us prove this claim. Let $f \in S(A,T)$. Then every $f' \in F_1$ such that $H_{\alpha_{f'}} \cap H_0 = H_{\alpha_{f}} \cap H_0$ belongs to the separation set $S(A',T')$. Furthermore, there are an odd number of such elements $f' \in F_1$. Hence the parity of $\|S(A,T)\|$ is the same as the parity of $\|S(A',T') \cap F_1\|$. Next consider the set $S(A',T') \cap F_0$ where $F_0$ is the complement $(E - \{e_0\}) - F_1$. Assume that $f \in S(A',T') \cap F_0$. Then there are an even number of $f' \in F_0$ such that $H_{\alpha_{f'}} \cap H_0 = H_{\alpha_{f}} \cap H_0$, and they all belong to the set $S(A',T')$. Hence the cardinality of $S(A',T') \cap F_0$ is even, yielding $\|S(A,T)\| \equiv \|S(A',T')\| \bmod 2$. We then have $$\begin{aligned} \psi_{{\mathcal{H}}'_0/C}(A',\mu) = \sum_{T\in{\mathscr{T}}_0} (-1)^{\|S(A,T)\|} \cdot \sum_{\substack{T'\in{\mathscr{T}}'_0\cap{\mathscr{L}}'_{0,\geq C} \\ T'\subseteq T}} \psi_{T'/C}(A',\mu).\end{aligned}$$ Next we note that if $T\in{\mathscr{T}}_0$ then $\{T'\in{\mathscr{T}}'_0\cap {\mathscr{L}}'_{0,\geq C} : T'\subset T\}$ is empty if $T\not\in{\mathscr{L}}_{0,\geq C_0}$. Indeed, if there exists $T'\in{\mathscr{T}}'_0\cap{\mathscr{L}}'_{0,\geq C}$ such that $T'\subset T$, then $\overline{T}\supset\overline{T'}\supset C$, so $\overline{T}\supset C_0$. Moreover, if $T\in{\mathscr{T}}_0\cap{\mathscr{L}}_{0,\geq C_0}$, then any $T'\in{\mathscr{T}}'_0$ such that $T'\subset T$ is in ${\mathscr{L}}'_{0,\geq C}$. Indeed, to go from ${\mathcal{H}}'_0$ to ${\mathcal{H}}_0$, we only removed hyperplanes that contain $C$. Thus if $T'\in{\mathscr{T}}'_0$ is such that $T'\subset T$ then for every $e\in E-E(C)$ we have $s(T')_e=s(T)_e=s(C_0)_e=s(C)_e$, which implies that $T'\geq C$. So we get that $$\psi_{{\mathcal{H}}'_0/C}(A',\mu)=\sum_{T\in{\mathscr{T}}_0\cap{\mathscr{L}}_{0,\geq C_0}} (-1)^{\|S(A,T)\|} \cdot \sum_{\substack{T'\in{\mathscr{T}}'_0 \\ T'\subseteq T}}\psi_{T'/C}(A',\mu).$$ It remains to show that for every $T\in{\mathscr{T}}_0\cap{\mathscr{L}}_{0,\geq C_0}$ the following holds: $$\psi_{T/C_0}(A,\mu)= \sum_{\substack{T'\in{\mathscr{T}}'_0 \\ T'\subseteq T}}\psi_{T'/C}(A',\mu).$$ Let $\ell\in H_0^\vee$ be the linear functional defined by $\ell=(\mu,\cdot)$, and let $x\in (-C)\circ A'$. By Lemma \[lemma\_on\_the\_definition\_of\_psi\] there exists a valuation $K \longmapsto\psi_K(x,\ell)$ on the set of closed convex polyhedral cones in $H_0$ such that - for every $D'\in{\mathscr{L}}'_{0,\geq C}$, we have $\psi_{D'/C}(A',\mu)=\psi_{\overline{D'}}(x,\ell)$; - for every $D\in{\mathscr{L}}_{0,\geq C_0}$, we have $\psi_{D/C_0}(A,\mu)=\psi_{\overline{D}} (x,\ell)$ (recall that $C\subset C_0$ and $A'\subset A$, so $x\in (-C_0)\circ A$). Let $T\in{\mathscr{T}}_0\cap{\mathscr{L}}_{0,\geq C_0}$. Then we have $$\overline{T} = \bigcup_{\substack{T'\in{\mathscr{T}}'_0 \\ T'\subseteq T}} \overline{T'}$$ and, if $T'_1,T'_2,\ldots,T'_i$ are pairwise distinct elements of ${\mathscr{T}}'_0$ that are included in $T$ and if $i\geq 2$, then $\overline{T'_1}\cap\overline{T'_2}\cap\cdots\cap \overline{T'_i}$ is the closure of a face of ${\mathscr{L}}'_0$ that is not a chamber. As $\psi_{D'/C}(A',\mu)=0$ for every $D'\in{\mathscr{L}}'_0$ that is not a chamber (see Remark \[rmk\_psi\_for\_chambers\]), the general properties of valuations (see [@Gr Theorem 1]) imply that $$\begin{aligned} \psi_{T/C_0}(A,\mu) & = \sum_{\substack{T'\in{\mathscr{T}}'_0 \\ T'\subseteq T}} \psi_{T'/C}(A',\mu),\end{aligned}$$ as desired. The case of Coxeter arrangements {#section_induction_Coxeter_case} -------------------------------- In this section we make the recursive expression of Proposition \[prop\_induction\_formula\_for\_psi\](ii) more explicit in the case of Coxeter arrangements. Suppose that ${\mathcal{H}}$ is a Coxeter arrangement, and fix a chamber $B$ of ${\mathcal{H}}$. We suppose that $B$ is on the positive side of every hyperplane of ${\mathcal{H}}$. As in Theorem \[thm\_Coxeter\_arrangements\], we obtain a Coxeter system $(W,S)$ with $W$ finite. Recall that $W$ is the subgroup of ${\mathbf{GL}}(V)$ generated by the reflections across the hyperplanes of ${\mathcal{H}}$, and that $S\subset W$ is the set of reflections across the hyperplanes that are walls of the base chamber $B$. We also consider the following action of $W$ on $E$. If $e\in E$ and $w\in W$ then by assumption $w(H_{\alpha_e})$ is a hyperplane of ${\mathcal{H}}$, so it is of the form $H_{\alpha_f}$ for a unique $f\in E$, and we set $w(e)=f$. This is clearly a left action. We have the following lemma. Let $\lambda\in V$, $C,D\in{\mathscr{L}}$ and $w\in W$. Then the following statements hold: - $w({\mathscr{L}}_\lambda)={\mathscr{L}}_{w(\lambda)}$; - $w(C\circ D)=w(C)\circ w(D)$; - $S(w(C),w(D))= w(S(C,D))$; - $\psi_C(D,\lambda)=\psi_{w(C)}(w(D),w(\lambda))$; - for every $B\in{\mathscr{T}}$, we have $\psi_{\mathcal{H}}(B,\lambda)=\psi_{\mathcal{H}}(w(B),w(\lambda))$. \[lemma\_W\_equivariance\] Statements (i), (ii) and (iii) are straightforward consequences of the fact that the action of $W$ on $V$ preserves the inner product $(\cdot,\cdot)$. Statement (iv) follows immediately from (i), (ii) and the definition of $\psi_C(D,\lambda)$, and (v) follows immediately from (iii) and (iv). More generally, if $w\in{\mathbf{GL}}(V)$ is an isometry, the same proof shows that we have $\psi_{\mathcal{H}}(B,\lambda)=\psi_{w({\mathcal{H}})}(w(B),w(\lambda))$. \[rmk\_W\_equivariance\] We now consider the situation of Proposition \[prop\_induction\_formula\_for\_psi\](ii), with $B$ the fixed base chamber of ${\mathscr{T}}$. For $e_0\in E$, the hyperplane $H_0=H_{\alpha_{e_0}}$ is a wall of $B$ if and only the reflection $s_0$ across this hyperplane is an element of $S$ (by definition of $S$). We fix such a $e_0$. We have the following essential fact. Let $e\in E - \{e_0\}$. Then the intersection multiplicity of ${\mathcal{H}}$ at $H_0\cap H_{\alpha_e}$ is even if and only if there exists $f\in E - \{e_0\}$ such that $H_0\cap H_{\alpha_f}=H_0\cap H_{\alpha_e}$ and that $(\alpha_{e_0},\alpha_f)=0$. Let $E'=\{f\in E - \{e_0\} : H_0\cap H_{\alpha_f}=H_0\cap H_{\alpha_e}\}$. Then $\|E'\|+1$ is the intersection multiplicity of ${\mathcal{H}}$ at the intersection $H_0 \cap H_{\alpha_e}$, and we have $s_0(E')=E'$. As $s_0^2=1$, the set $E'$ has odd cardinality if and only $s_0$ has a fixed point on $E'$. But $f\in E'$ is fixed by $s_0$ if and only if $s(H_{\alpha_f})=H_{\alpha_f}$, which is equivalent to the condition that $(\alpha_{e_0},\alpha_f)=0$. In particular, we can choose the set $F$ of Proposition \[prop\_induction\_formula\_for\_psi\](ii) to be $F=\{f\in E : (\alpha_{e_0},\alpha_f)=0\}$, and we obtain a hyperplane arrangement ${\mathcal{H}}_0=(H_0\cap H_{\alpha_f})_{f\in F}$ on $H_0$. For every $f\in F$, if $s_f\in W$ is the reflection across $H_{\alpha_f}$ then $s_f(H_0)=H_0$ and $s_f(F)=F$. In particular, $s_f$ preserves the arrangement ${\mathcal{H}}_0$. We conclude that ${\mathcal{H}}_0$ is a Coxeter arrangement on $H_0$. Finally, using Lemma \[lemma\_W\_equivariance\](v), the conclusion of Proposition \[prop\_induction\_formula\_for\_psi\](ii) becomes that, for every $\lambda\in V$, $$\psi_{\mathcal{H}}(B,\lambda)+\psi_{\mathcal{H}}(B,s(\lambda)) = 2 \cdot \psi_{{\mathcal{H}}_0}(A,\mu),$$ where $A$ and $\mu$ are defined in Proposition \[prop\_induction\_formula\_for\_psi\](ii). [Consider the $A_{3}$ arrangement in ${\mathbb{R}}^{4}$. This arrangement has $6$ positive roots: $\Phi^+ = \{e_{i} - e_{j} : 1 \leq i < j \leq 4\}$. Let $B = \{x : x_1 > x_2 > x_3 > x_4\}$ be the base chamber. Choose $\alpha_0$ to be the root $e_3 - e_4$, that is, the hyperplane $H_0$ is given by $\{x : x_3 = x_4\}$. The arrangement ${\mathcal{H}}'_0$ consists of the five hyperplanes $\{x \in H_0: x_1 = x_3\} = \{x \in H_0 : x_1 = x_4\}$, $\{x \in H_0: x_2 = x_3\} = \{x \in H_0 : x_2 = x_4\}$ and $\{x \in H_0: x_1 = x_2\}$. But since the only hyperplane orthogonal to $H_0$ is $\{x : x_1 = x_2\}$, the arrangement ${\mathcal{H}}_0$ only consists of the hyperplane $\{x \in H_0: x_1 = x_2\}$. Thus the arrangement ${\mathcal{H}}_0$ is the Coxeter arrangement of type $A_1$ embedded in a $3$-dimensional space. Furthermore, the chamber $A'$ in ${\mathcal{H}}'_0$ is given by $\{x : x_1 > x_2 > x_3 = x_4\}$, whereas the chamber $A$ in ${\mathcal{H}}_0$ is given by $\{x : x_1 > x_2\text{ and }x_3=x_4\}$. ]{} [Consider the $B_{3}$ arrangement in ${\mathbb{R}}^{3}$. This arrangement has $9$ positive roots: $\Phi^+ = \{e_{i} \pm e_{j} : 1 \leq i < j \leq 3\} \cup \{e_{i} : 1 \leq i \leq 3\}$. Let $B = \{x : x_1 > x_2 > x_3 > 0\}$ be the base chamber. Then the simple roots are $\{e_{1}-e_{2}, e_{2}-e_{3}, e_3\}$. Now pick $\alpha_0$ to be the root $e_3$, that is, the hyperplane $H_0$ is given by $\{x : x_3 = 0\}$. The arrangement ${\mathcal{H}}'_0$ consists of the eight hyperplanes $\{x \in H_0: x_1 = x_2\}$, $\{x \in H_0 : x_1 = -x_2\}$ (both of multiplicity $1$), and $\{x \in H_0: x_1 = 0\}$, $\{x \in H_0 : x_2 = 0\}$ (both of multiplicity $3$). Since all the multiplicities are odd, the two arrangements ${\mathcal{H}}_0$ and ${\mathcal{H}}'_0$ have the same set of hyperplanes. But all the hyperplanes of ${\mathcal{H}}_0$ have multiplicity $1$, thus the arrangement ${\mathcal{H}}_0$ is a Coxeter arrangement of type $B_2$. Furthermore, the two chambers $A$ and $A'$ are equal and are given by $\{x : x_1 > x_2 > x_3=0\}$. ]{} In the next section we consider a slightly more general situation. Suppose that $E=E^{(1)}\sqcup E^{(2)}$, with ${\mathcal{H}}^{(1)}=(H_{\alpha_e})_ {e\in E^{(1)}}$ a Coxeter arrangement whose Coxeter group $W$ stabilizes ${\mathcal{H}}^{(2)}=(H_{\alpha_e})_{e\in E^{(2)}}$, and that $C=(\bigcap_{e\in E^{(1)}}H_{\alpha_e})\cap(\bigcap_ {e\in E^{(2)}}H^+_{\alpha_e})$, so that $E(C)=E^{(1)}$. We apply Proposition \[prop\_induction\_formula\_for\_psi\] for some $e_0\in E^{(1)}$. As in the case of a Coxeter arrangement, we can choose the subset $F$ of $E^{(1)}$ to be $F=\{f\in E^{(1)} : (\alpha_{e_0},\alpha_f)=0\}$. The resulting hyperplane arrangement ${\mathcal{H}}_0$ on $H_0=H_{\alpha_0}$ decomposes again as ${\mathcal{H}}_0={\mathcal{H}}_0^{(1)}\sqcup{\mathcal{H}}_0^{(2)}$, where ${\mathcal{H}}_0^{(1)}=(H_0\cap H_{\alpha_f})_{f\in F}$ and ${\mathcal{H}}_0^{(2)}=(H_0\cap H_{\alpha_e})_{e\in E^{(2)}}$, and we see as before that ${\mathcal{H}}_0^{(1)}$ is a Coxeter arrangement on $H_0$. Moreover, the unique face $C_0$ of ${\mathcal{H}}_0$ that contains $C$ is clearly $C_0=(\bigcap_{e\in F}(H_0\cap H_{\alpha_e}))\cap(\bigcap_ {e\in E^{(2)}}(H_0\cap H^+_{\alpha_e}))$. If $\lambda\in V$ and if $B$ is a chamber whose closure contains $C$ and such that $H_0$ is a wall of $B$, we have $$\psi_{{\mathcal{H}}/C}(B,\lambda)+\psi_{{\mathcal{H}}/C}(B,s_{\alpha_{e_0}}(\lambda))= 2 \cdot \psi_{{\mathcal{H}}_0/C_0}(A,\mu),$$ where $A$ and $\mu$ are defined in Proposition \[prop\_induction\_formula\_for\_psi\](ii). \[ex\_parabolic\_arrangement\] [Let $(W,S)$ be a Coxeter system, let $V$ be the canonical representation of $W$, and let ${\mathcal{H}}=(H_\alpha)_{\alpha\in\Phi^+}$ be the associated hyperplane arrangement on $V$ as in Section \[section\_Coxeter\_arrangements\]. Let $I$ be a subset of $S$, set $\Phi^{(1)} = \Phi^+ \cap\bigl(\sum_{\alpha\in I}{{\mathbb{R}}}\alpha\bigr)$ and $\Phi^{(2)}=\Phi^+-\Phi^{(1)}$. Then ${\mathcal{H}}^{(1)}=(H_\alpha)_{\alpha\in \Phi^{(1)}}$ is a Coxeter arrangement with associated Coxeter system $(W_I,I)$, where $W_I$ is the subgroup of $W$ generated by $I$, and $W_I$ preserves the arrangement ${\mathcal{H}}^{(2)}=(H_\alpha)_{\alpha\in\Phi^{(2)}}$. If $C = (\bigcap_{\alpha\in \Phi^{(1)}} H_\alpha) \cap (\bigcap_{\alpha\in \Phi^{(2)}}H^+_\alpha)$ as before, then the chamber $B$ corresponding to $1\in W$ is in ${\mathscr{L}}_{\geq C}$. ]{} An identity for Coxeter arrangements {#section_main_thm} ==================================== We now specialize to the case at the end of Subsection \[section\_induction\_Coxeter\_case\]. That is, we assume that $E=E^{(1)}\sqcup E^{(2)}$, with ${\mathcal{H}}^{(1)}=(H_{\alpha_e})_ {e\in E^{(1)}}$ a Coxeter arrangement whose Coxeter group stabilizes ${\mathcal{H}}^{(2)}=(H_{\alpha_e})_{e\in E^{(2)}}$, and that $C= (\bigcap_{e\in E^{(1)}}H_{\alpha_e})\cap(\bigcap_ {e\in E^{(2)}}H^+_{\alpha_e})$. To simplify some of the notation, without loss of generality we may assume that the vectors $\alpha_e$, where $e\in E$, are all unit vectors. We assume that there exists a chamber $B$ of ${\mathcal{H}}$ that is on the positive side of every hyperplane of ${\mathcal{H}}$ (not just ${\mathcal{H}}^{(1)}$). This also defines a chamber of the arrangement ${\mathcal{H}}^{(1)}$, and we denote by $(W,S)$ the associated Coxeter system, as in Theorem \[thm\_Coxeter\_arrangements\]. The set $\Phi=\{\pm\alpha_e : e\in E^{(1)}\}$ is a normalized pseudo-root system (see Definition \[def\_pseudo\_root\_system\]), the subset $\Phi^+=\{\alpha_e : e\in E^{(1)}\}$ is a system of positive pseudo-roots in $\Phi$ (see Definition \[def\_positive\_pseudo\_roots\]), and $(W,S)$ is the corresponding Coxeter system. See Proposition \[prop\_pseudo\_root\_system\_vs\_Coxeter\_system\]. The main theorem ---------------- Recall the definition of $2$-structures from Subsection \[subsection\_2-structures\]. A $2$-structure* for $\Phi$ is a subset $\varphi \subseteq \Phi$ such that: - $\varphi$ is a pseudo-root system whose irreducible components are all of type $A_1$, $B_2$ or $I_2(2^k)$ with $k\geq 3$; - for every $w\in W$ such that $w(\varphi\cap\Phi^+)= \varphi\cap\Phi^+$, we have $\det(w)=1$. Recall that $\mathcal{T}(\Phi)$ is the set of $2$-structures of $\Phi$. By Proposition \[prop\_W\_acts\_transitively\_on\_T\_Phi\], the group $W$ acts transitively on $\mathcal{T}(\Phi)$. In Definition \[def\_sign\_2\_structure\] we define the sign $\epsilon(\varphi)=\epsilon(\varphi,\Phi^+)$ of any $2$-structure $\varphi\in\mathcal{T}(\Phi)$. If $\varphi\in\mathcal{T}(\Phi)$, we also write $\varphi^+=\varphi\cap\Phi^+$, and we denote by ${\mathcal{H}}_\varphi$ the hyperplane arrangement $(H_{\alpha})_{\alpha\in\varphi^+\sqcup E^{(2)}}$ and by $B_\varphi$, respectively $C_\varphi$, the unique chamber of ${\mathcal{H}}_\varphi$ containing $B$, respectively $C$. By the choice of $B$, the chamber $B_\varphi$ is also the unique chamber on the positive side of every hyperplane in ${\mathcal{H}}_\varphi$. The main theorem of this article is the following. Let ${\mathcal{H}}$, $C$ and $B$ be as above. For every $\lambda\in V$, we have $$\psi_{{\mathcal{H}}/C}(B,\lambda)=\sum_{\varphi\in\mathcal{T}(\Phi)} \epsilon(\varphi) \cdot \psi_{{\mathcal{H}}_\varphi/C_\varphi}(B_\varphi,\lambda).$$ \[thm\_main\] This theorem will be proved in Subsection \[section\_proof\_main\_theorem\]. We note the following corollary, which extends [@Herb-DSC Theorem 5.3] to the case of Coxeter systems. Note we do not obtain a new proof of Herb’s theorem, as Herb uses a similar induction argument to prove it. As we are using the definition of the sign of a $2$-structure from [@Herb-Pl], our formula looks a bit different from the one of [@Herb-DSC Theorem 5.3]. This is explained in Remark 5.1 of [@Herb-DSC], and we generalize the comparison between the two definitions of the sign in Corollary \[cor\_comparison\_signs\]. \[cor\_sum\_of\_signs\] The sum of the signs of all $2$-structures of a pseudo-root system is equal to $1$, that is, $$\sum_{\varphi\in\mathcal{T}(\Phi)}\epsilon(\varphi)=1.$$ Take $E^{(2)}=\varnothing$ and $\lambda=0$ in the identity of Theorem \[thm\_main\]. By Proposition \[prop\_base\_case\_for\_psi\], we have $\psi_{{\mathcal{H}}}(B,0)=(-1)^{\dim(V)+\|E\|}=(-1)^{\dim(V)+\|\Phi^+\|}$. By the same proposition, for every $\varphi\in\mathcal{T}(\Phi)$, we have $\psi_{{\mathcal{H}}_\varphi}(B_\varphi,0)=(-1)^{\dim(V)+\|\varphi^+\|}$. The equality then follows from the fact that $\|\Phi^+-\varphi^+\|$ is even for every 2-structure $\varphi$, which is proved in Lemma \[lemma\_r\_Phi\]. Let $\varphi\in\mathcal{T}(\Phi)$, and let $W(\varphi,\Phi^+)=\{w\in W:w(\varphi^+)\subset\Phi^+\}$ and $W_1(\varphi,\Phi^+)=\{w\in W:w(\varphi^+)\subset\varphi^+\}$. Then the sign $\epsilon(\varphi,\Phi^+)$ is given by $$\epsilon(\varphi,\Phi^+) =\frac{1}{\|W_1(\varphi,\Phi^+)\|} \sum_{w\in W(\varphi,\Phi^+)}\det(w).$$ \[cor\_comparison\_signs\] By Corollary \[cor\_T\_Phi\_as\_quotient\], we have a bijection $W(\varphi,\Phi^+)/W_1(\varphi,\Phi^+) \longrightarrow \mathcal{T}(\Phi)$, $w \longmapsto w(\varphi)$. So, by Corollary \[cor\_sum\_of\_signs\] and and Lemma \[lemma\_sign\_of\_w(varphi)\], we get $$\begin{aligned} 1 & = \frac{1}{\|W_1(\varphi,\Phi^+)\|}\sum_{w\in W(\varphi,\Phi^+)} \epsilon(w(\varphi),\Phi^+) = \epsilon(\varphi,\Phi^+) \frac{1}{\|W_1(\varphi,\Phi^+)\|}\sum_{w\in W(\varphi,\Phi^+)} \det(w). $$ Recursive formula for the right-hand side ----------------------------------------- In this subsection, we prove a recursive formula for the right-hand side of the identity in Theorem \[thm\_main\], mirroring the identity of Proposition \[prop\_induction\_formula\_for\_psi\](ii). Let $\alpha_0\in\Phi^+$ be a simple pseudo-root (see Definition \[def\_positive\_pseudo\_roots\]), and let $s_0=s_{\alpha_0}$ be the corresponding element of the set of simple reflections $S$. As in Subsection \[section\_induction\_Coxeter\_case\], we set $H_0=H_{\alpha_0}$ and $F = \{e\in E^{(1)} : (\alpha_e,\alpha_0)=0\}$, and we consider the arrangement ${\mathcal{H}}_0={\mathcal{H}}_0^{(1)}\sqcup{\mathcal{H}}_0^{(2)}$ on $H_0$, where ${\mathcal{H}}_0^{(1)}=(H_0\cap H_{\alpha_f})_{f\in F}$ and ${\mathcal{H}}_0^{(2)}=(H_0\cap H_{\alpha_e})_{e\in E^{(2)}}$. Then ${\mathcal{H}}_0^{(1)}$ is a Coxeter arrangement, with normalized pseudo-root system $\Phi_0=\{\pm\alpha_f : f\in F\}$ and system of positive pseudo-roots $\Phi_0^+=\{\alpha_f : f\in F\}$. Its Coxeter group preserves ${\mathcal{H}}_0^{(2)}$, because it preserves $H_0$ and is a subgroup of $W$. We denote by $A$ the unique chamber of ${\mathcal{H}}_0$ containing the interior (in $H_0$) of $\overline{B}\cap H_0$. This is also the unique chamber on the positive side of all the hyperplanes of ${\mathcal{H}}_0$. Finally, we fix $\lambda\in V$, and we denote by $\mu$ the orthogonal projection of $\lambda$ on $H_0$. The recursive formula is the following result: We have: $$\begin{aligned} \sum_{\varphi\in\mathcal{T}(\Phi)} \epsilon(\varphi) \cdot \psi_{{\mathcal{H}}_\varphi/C_\varphi} (B_\varphi,\lambda) & + \sum_{\varphi\in\mathcal{T}(\Phi)} \epsilon(\varphi) \cdot \psi_{{\mathcal{H}}_\varphi/C_\varphi} (B_\varphi,s_0(\lambda)) \nonumber \\ & = 2 \cdot \sum_{\varphi_0\in\mathcal{T}(\Phi_0)} \epsilon(\varphi_0) \cdot \psi_{{\mathcal{H}}_{0,\varphi_0}/C_{0,\varphi_0}} (A_{\varphi_0},\mu). \label{equation_recursive_RHS}\end{aligned}$$ \[prop\_induction\_RHS\] As $s_0^2=1$, the left-hand side of is equal to $$\sum_{\varphi\in\mathcal{T}(\Phi)} \epsilon(\varphi) \cdot \psi_{{\mathcal{H}}_\varphi/C_\varphi}(B_\varphi,\lambda) + \sum_{\varphi\in\mathcal{T}(\Phi)} \epsilon(s_0(\varphi)) \cdot \psi_{{\mathcal{H}}_{s_0(\varphi)}/C_{s_0(\varphi)}} (B_{s_0(\varphi)},s_0(\lambda)).$$ Let $\mathcal{T}'=\{\varphi\in\mathcal{T} : s_0(\varphi)\not=\varphi\}$ and $\mathcal{T}''=\{\varphi\in\mathcal{T} : s_0(\varphi)=\varphi\}$. Suppose that $\varphi\in\mathcal{T}'$. Then $\alpha_0\not\in\varphi$ holds, so $s_0(\varphi^+)\subset s_0(\Phi^+-\{\alpha_0\})\subset\Phi^+$ by [@BB Lemma 4.4.3]. Hence $\epsilon(s_0(\varphi))=-\epsilon(\varphi)$ by Lemma \[lemma\_sign\_of\_w(varphi)\]. Moreover, again using the fact that $s_0(\varphi^+)\subset\Phi^+$, we see that $s_0(\varphi)\cap\Phi^+=s_0(\varphi^+)$, so ${\mathcal{H}}_{s_0(\varphi)}=s_0({\mathcal{H}}_\varphi)$, $C_{s_0(\varphi)}=C_\varphi=s_0(C_\varphi)$ and $B_{s_0(\varphi)}=s_0(B_\varphi)$. By Remark \[rmk\_W\_equivariance\] we have $\psi_{{\mathcal{H}}_{s_0(\varphi)}/C_{s_0(\varphi)}}(B_{s_0(\varphi)},s_0(\lambda))= \psi_{{\mathcal{H}}_\varphi/C_\varphi}(B_\varphi,\lambda)$. This shows that $$\sum_{\varphi\in\mathcal{T}'} \epsilon(\varphi) \cdot \psi_{{\mathcal{H}}_\varphi/C_\varphi}(B_\varphi,\lambda)+ \sum_{\varphi\in\mathcal{T}'} \epsilon(s_0(\varphi)) \cdot \psi_{{\mathcal{H}}_{s_0(\varphi)}/C_{s_0(\varphi)}} (B_{s_0(\varphi)},s_0(\lambda))=0,$$ so the left hand-side of is equal to $$\begin{aligned} \sum_{\varphi\in\mathcal{T}''} \epsilon(\varphi) \cdot \psi_{{\mathcal{H}}_\varphi/C_\varphi}(B_\varphi,\lambda) & + \sum_{\varphi\in\mathcal{T}''} \epsilon(s_0(\varphi)) \cdot \psi_{{\mathcal{H}}_{s_0(\varphi)}/C_{s_0(\varphi)}}(B_{s_0(\varphi)},s_0(\lambda)) \\ & = \sum_{\varphi\in\mathcal{T}''} \epsilon(\varphi) \cdot (\psi_{{\mathcal{H}}_\varphi/C_\varphi}(B_\varphi,\lambda) + \psi_{{\mathcal{H}}_\varphi/C_\varphi}(B_\varphi,s_0(\lambda))).\end{aligned}$$ We wish to prove that the right-hand side of this last equality is equal to the right-hand side of . Note first that if $\varphi\in\mathcal{T}''$, we must have $\alpha_0\in\varphi$. Indeed, we know that $s_{\alpha_0}(\Phi^+-\{\alpha_0\})\subset\Phi^+$ (see [@BB Lemma 4.4.3]), so, if $\alpha_0\not\in\varphi$, then $s_0(\varphi^+) \subseteq \Phi^+\cap\varphi= \varphi^+$, contradicting condition (b) in the definition of a $2$-structure. By Proposition \[prop\_induction\_formula\_for\_psi\](ii) and Subsection \[section\_induction\_Coxeter\_case\], if $\varphi\in\mathcal{T}''$ then we have $$\psi_{{\mathcal{H}}_\varphi/C_\varphi}(B_\varphi,\lambda)+ \psi_{{\mathcal{H}}_\varphi/C_\varphi}(B_\varphi,s_0(\lambda)) = 2 \cdot \psi_{{\mathcal{H}}_{0,\varphi\cap\Phi_0}/C_{0,\varphi_0}}(A_{\varphi\cap\Phi_0},\mu),$$ where ${\mathcal{H}}_{0,\varphi\cap\Phi_0}$, $C_{0,\varphi_0}$ and $A_{\varphi\cap\Phi_0}$ are defined in the natural way, that is, as in the paragraph before the statement of the proposition, even though $\varphi\cap\Phi_0$ is not necessarily a $2$-structure for $\Phi_0$. Let $\mathcal{T}''_1$ be the set of $\varphi\in\mathcal{T}''$ such that $\varphi\cap\Phi_0$ is a $2$-structure for $\Phi_0$, and $\mathcal{T}''_2=\mathcal{T}''- \mathcal{T}''_1$. By statements (1) and (2) of Lemma \[lemma\_T\_Phi\_induction\_step\], we have $$\begin{aligned} \sum_{\varphi\in\mathcal{T}''_1} \epsilon(\varphi) \cdot \psi_{{\mathcal{H}}_{\varphi\cap\Phi_0}/C_{0,\varphi\cap\Phi_0}} (A_{\varphi\cap\Phi_0},\mu) & = \sum_{\varphi_0\in\mathcal{T}(\Phi_0)} \epsilon(\varphi_0) \cdot \psi_{{\mathcal{H}}_{0,\varphi_0}/C_{0,\varphi_0}}(A_{\varphi_0},\mu). \label{equation_almost_at_the_end_of_the_proof}\end{aligned}$$ Also, by statement (3) of Lemma \[lemma\_T\_Phi\_induction\_step\], there exists an involution $\iota$ of $\mathcal{T}''_2$ such that $\iota(\varphi)\cap\Phi_0=\varphi\cap\Phi_0$ and $\epsilon(\iota(\varphi))=-\epsilon(\varphi)$ for every $\varphi\in\mathcal{T}''_2$. This immediately implies that $$\begin{aligned} \sum_{\varphi\in\mathcal{T}''_2} \epsilon(\varphi) \cdot \psi_{{\mathcal{H}}_{\varphi\cap\Phi_0}/C_{0,\varphi\cap\Phi_0}}(A_{\varphi\cap\Phi_0},\mu) & =0 . \label{equation_the_end_of_the_proof}\end{aligned}$$ Putting equations and together, we get the desired result. Proof of the main theorem {#section_proof_main_theorem} ------------------------- We prove the result by induction on the rank of $\Phi$. Suppose that $\Phi$ is empty. Then ${\mathcal{H}}={\mathcal{H}}^{(2)}$, $C=B$ is a chamber of ${\mathcal{H}}$, and the only element of $\mathcal{T}(\Phi)$ is the empty pseudo-root system. So both sides of the identity of Theorem \[thm\_main\] are equal to $(-1)^{\dim(V)}$ if $B\in{\mathscr{L}}_\lambda$ and $0$ otherwise. Suppose now that $\Phi$ is not empty, and that we know the result for all pseudo-root systems of smaller rank. We first treat the case where ${\mathscr{L}}_{\lambda,\geq C}={\mathscr{L}}_{\geq C}$. This means that $\bigcup_{D\in{\mathscr{L}}_{\geq C}}D\subset\overline{H^+_\lambda}$. As $\psi_{{\mathcal{H}}/C}(B,\lambda)$ only depends on $\lambda$ through ${\mathscr{L}}_{\lambda,\geq C}$, we have $\psi_{{\mathcal{H}}/C}(B,\lambda)= \psi_{{\mathcal{H}}/C}(B,0)$. Similarly, for every $\varphi\in\mathcal{T}(\Phi)$, if ${\mathscr{L}}_\varphi$ is the face poset of ${\mathcal{H}}_\varphi$, then we have ${\mathscr{L}}_{\varphi,\lambda,\geq C_\varphi}={\mathscr{L}}_{\varphi, C_\varphi}$ because $\bigcup_{D\in{\mathscr{L}}_{\varphi,\geq C_\varphi}} D = \bigcup_{D\in{\mathscr{L}}_{\geq C}} D = \bigcap_{e\in E^{(2)}}H^+_{\alpha_e}$, hence $\psi_{{\mathcal{H}}_\varphi/C_\varphi}(B_\varphi,\lambda)= \psi_{{\mathcal{H}}_\varphi/C_\varphi}(B_\varphi,0)$. So we may assume that $\lambda=0$. Choose any simple pseudo-root $\alpha_0$, and let $s_0=s_{\alpha_0}$. As $s_0(0)=0$, Proposition \[prop\_induction\_formula\_for\_psi\](ii) and Subsection \[section\_induction\_Coxeter\_case\] imply that $\psi_{{\mathcal{H}}/C}(B,0)=\psi_{{\mathcal{H}}_0/C_0}(A,0)$, and Proposition \[prop\_induction\_RHS\] implies that $$\sum_{\varphi\in\mathcal{T}(\Phi)} \epsilon(\varphi) \cdot \psi_{{\mathcal{H}}_\varphi/C_\varphi}(B_\varphi,0) = \sum_{\varphi_0\in\mathcal{T}(\Phi_0)} \epsilon(\varphi_0) \cdot \psi_{{\mathcal{H}}_{0,\varphi_0}/C_{0,\varphi_0}} (A_{\varphi_0},0).$$ So the result follows from the induction hypothesis. Finally, suppose that ${\mathscr{L}}_{\lambda,\geq C}\not={\mathscr{L}}_{\geq C}$. Choose $w\in W$ such that $\lambda^{\prime}=w^{-1}(\lambda)$ satisfies $(\lambda^{\prime},\alpha)\geq 0$ for every $\alpha\in\Phi^+$. We do a second induction, now on the length of $w$. If $w$ has length $0$, that is, if $w=1$, then $B$ and $\lambda$ satisfy the assumptions of Corollary \[cor\_lower\_order\_ideal\], so by Proposition \[prop\_base\_case\_for\_psi\], we have $\psi_{{\mathcal{H}}/C}(B,\lambda)=0$. Let $\varphi\in\mathcal{T}(\Phi)$. Then $B_\varphi$ and $\lambda$ also satisfy the assumptions of Corollary \[cor\_lower\_order\_ideal\], so by Proposition \[prop\_base\_case\_for\_psi\] we have $\psi_{{\mathcal{H}}_\varphi/C_\varphi} (B_\varphi,\lambda)=0$. So we get the identity in this case. Suppose that $w$ has positive length, and choose a simple pseudo-root $\alpha_0$ such that $w=s_{\alpha_0}w'$ with $w'$ of length one less than $w$. By the induction hypothesis on $w$, we know that the identity holds for $s_{\alpha_0}(\lambda)=w'(\lambda')$. The result now follows from the recursive expressions for each side, that is, Proposition \[prop\_induction\_formula\_for\_psi\](ii) and Subsection \[section\_induction\_Coxeter\_case\] for the left-hand side, and Proposition \[prop\_induction\_RHS\] for the right-hand side, and from the induction hypothesis for $\Phi$. Applications {#section_applications} ============ Coxeter systems --------------- We now specialize Theorem \[thm\_main\] to the case where ${\mathcal{H}}={\mathcal{H}}^{(1)}$ is a Coxeter arrangement. In particular, $C$ is the minimal face of ${\mathscr{L}}$, so $\psi_{{\mathcal{H}}/C}(B,\lambda)=\psi_{\mathcal{H}}(B,\lambda)$ for every $\lambda\in V$. Let $\varphi\in\mathcal{T}(\Phi)$, and let $\varphi=\varphi_1 \sqcup\varphi_2\sqcup\cdots\sqcup\varphi_r$ be the decomposition of $\varphi$ into irreducible pseudo-root systems. Let $V_{i,\varphi}=\operatorname{Span}(\varphi_i)$ for $1 \leq i \leq r$. Then $V=V_{0,\varphi}\times V_{1,\varphi}\times\cdots\times V_{r,\varphi}$, where $V_{0,\varphi}=\varphi^\perp$. Note that the dimension of $V_{0,\varphi}$ is equal to $\dim(V)-\operatorname{rank}(\varphi)$, so it is independent of $\varphi$ by Proposition \[prop\_W\_acts\_transitively\_on\_T\_Phi\]. Let ${\mathcal{H}}_{i,\varphi}$ be the hyperplane arrangement given by $\varphi_i\cap\Phi^+$ on $V_{i,\varphi}$ where $1 \leq i \leq r$. For a fixed index $i$ let $B_{i,\varphi}$ be the chamber of the arrangement ${\mathcal{H}}_{i,\varphi}$ that is on the positive side of every hyperplane, and let $\lambda_{i,\varphi}$ be the orthogonal projection of $\lambda$ on $V_{i,\varphi}$. Combining Theorem \[thm\_main\] with Lemmas \[lemma\_psi\_inessential\_arrangement\] and \[lemma\_product\], we obtain: \[cor\_root\_systems\] For every $\lambda\in V$, we have $$\psi_{\mathcal{H}}(B,\lambda) = (-1)^{\dim(V)-R} \cdot \sum_{\substack{\varphi\in\mathcal{T}(\Phi)\\ \lambda\in\operatorname{Span}(\varphi)}} \epsilon(\varphi) \cdot \prod_{i=1}^r\psi_{{\mathcal{H}}_{i,\varphi}}(B_{i,\varphi}, \lambda_{i,\varphi}),$$ where $R$ is the rank of any $\varphi\in\mathcal{T}(\Phi)$. To finish the calculation of $\psi_{\mathcal{H}}(B,\lambda)$ in this case, we use the following proposition, whose proof is a straightforward calculation. In types $A_1$, $B_2 = I_2(4)$ and $I_2(2^k)$ for $k \geq 3$, the function $\psi$ is given by the following expressions: - Type $A_1$: Suppose that $V={{\mathbb{R}}}e_1$ and that $\Phi^+=\{e_1\}$. Then $\psi$ is given by $$\psi_{\mathcal{H}}(B,c e_1)=\begin{cases}0 & \text{ if } c>0, \\ 1 & \text{ if } c=0, \\ 2 & \text{ if } c<0.\end{cases}$$ - Type $I_2(2^k)$, where $k\geq 2$: Let $V={{\mathbb{R}}}e_1\oplus{{\mathbb{R}}}e_2$ with the usual inner product. For every $v\in V-\{0\}$, let $\theta(v)\in[0,2\pi)$ be the angle from $e_1$ to $v$. Suppose that $\Phi$ is the set of unit vectors that have an angle of $r\pi/2^k$ with $e_1$, where $r\in{{\mathbb{Z}}}$, and that $B$ is the set of nonzero vectors $v\in V$ such that $0<\theta(v)<\pi/2^k$. Then $\psi$ is given by $$\psi_{\mathcal{H}}(B,\lambda) = \begin{cases} 1 & \text{ if } \lambda=0, \\ 2 & \text{ if }\lambda\not=0\text{ and }\theta(\lambda)=r\pi/2^k\text{ with }2^{k-1}+1\leq r\leq 3\cdot 2^{k-1}, \\ 4 & \text{ if }\lambda\not=0\text{ and }r\pi/2^k<\theta(\lambda)<(r+1)\pi/2^k \\ & \text{ with } r \text{ odd and }2^{k-1}+1\leq r\leq 3\cdot 2^{k-1} - 1, \\ 0 & \text{ otherwise.} \end{cases}$$ in [0, 22.5, ..., 337.5]{} [(0,0) – ([3\cos()]{},[3\sin()]{});]{}; in [112.5, 135, ..., 270]{} [([3.2\cos()]{},[3.2\sin()]{}) node[$2$]{};]{}; in [123.75, 168.75, ..., 258.75]{} [([2.5\cos()]{},[2.5\sin()]{}) node[$4$]{};]{}; (0,0) – (2,0) node\[above\][$e_{1}$]{}; (0,0) – (0,2) node\[right\][$e_{2}$]{}; If $(W,S)$ arises from a root system $\Phi$ and $-1$ is an element of $W$ (or, equivalently, the root system is generated by strongly orthogonal roots), then [@GKM Theorem 3.1] and [@Herb-2S Theorem 4.2] give two different expressions for the coefficients appearing in the formula for the averaged discrete series characters of a real reductive group with root system $\Phi$. Corollary \[cor\_root\_systems\] asserts the equality of these two formulas. (Its proof is an induction similar to the ones used in the proofs of the two cited theorems.) In general, although there are no discrete series anymore, the formulas of Goresky–Kottwitz–MacPherson and Herb still make sense, and Corollary \[cor\_root\_systems\] says that they are still equal. Also, Corollary \[cor\_root\_systems\] implies that $\psi_{\mathcal{H}}(B,\lambda)=0$ if $\lambda$ is not in the span of any $2$-structure for $\Phi$, so it implies [@GKM Theorem 5.3]. It is not clear whether this is an easier proof than the one given in [@GKM]. The type [[A]{}]{} identity involving ordered set partitions {#second_application} -------------------------------------------------------------- We now show how to deduce [@Ehrenborg_Morel_Readdy Theorem 6.4] [^1] from Theorem \[thm\_main\]. We take $V={{\mathbb{R}}}^n$ with the usual inner product, and we denote by $(e_1,\ldots,e_n)$ the standard basis of $V$. We consider the hyperplane arrangement ${\mathcal{H}}$ of type $B_n$ on $V$, that is, ${\mathcal{H}}=(H_\alpha)_{\alpha\in\Phi_B^+}$, where $\Phi_B^+= \{e_i\pm e_j:1\leq i<j\leq n\}\cup\{e_1,\ldots,e_n\}$. We write $\Phi_B^+=\Phi^{(1)}\sqcup\Phi^{(2)}$, where $\Phi^{(1)}=\{e_i-e_j:1\leq i<j\leq n\}$, and we denote by ${\mathcal{H}}={\mathcal{H}}^{(1)}\sqcup{\mathcal{H}}^{(2)}$ the corresponding decomposition of ${\mathcal{H}}$. The arrangement ${\mathcal{H}}^{(1)}$ is a Coxeter arrangement of type $A_{n-1}$, and we denote by $\Phi=\Phi^{(1)}\cup(-\Phi^{(1)})$ the associated root system. Let $C$ be the intersection $(\bigcap_{\alpha\in\Phi^{(1)}}H_\alpha)\cap(\bigcap_{\alpha\in \Phi^{(2)}}H^+_\alpha)$. Then $C$ is the open ray ${{\mathbb{R}}}_{>0}\cdot(e_1+e_2+\cdots+e_n)$. Recall that ${\mathscr{L}}$ is the face poset of ${\mathcal{H}}$. We will now give a description of ${\mathscr{L}}$ in terms of signed ordered partitions. See also [@ER Section 5] for this description. A signed block is a nonempty subset ${\widetilde{B}}$ of $\{\pm 1,\ldots,\pm n\}$ such that, for every $i\in\{1, \ldots,n\}$, at most one of $\pm i$ is in ${\widetilde{B}}$. We then denote by $B$ the subset of $\{1,\ldots,n\}$ defined by $B=\{\|i\| : i\in{\widetilde{B}}\}$. A signed ordered partition of a subset $I$ of $\{1,\ldots,n\}$ is a list $({\widetilde{B}}_1,\ldots,{\widetilde{B}}_r)$ of signed blocks such that $(B_1,\ldots,B_r)$ is an ordered partition of $I$. We consider the poset ${\Pi^{\text{ord},B}_{n}}$ whose elements are pairs $\pi=({\widetilde{\pi}},Z)$, where $Z \subseteq \{1,\ldots,n\}$ and ${\widetilde{\pi}}$ is a signed ordered partition of $\{1,\ldots,n\}-Z$, and the cover relation is given by the following two rules: $$\begin{aligned} (({\widetilde{B}}_1,\ldots,{\widetilde{B}}_r),Z) & {\prec}(({\widetilde{B}}_1,\ldots,{\widetilde{B}}_{r-1}),B_r\cup Z) , \\ (({\widetilde{B}}_1,\ldots,{\widetilde{B}}_r),Z) & {\prec}(({\widetilde{B}}_1,\ldots,{\widetilde{B}}_{i-1}, {\widetilde{B}}_i\cup{\widetilde{B}}_{i+1},{\widetilde{B}}_{i+2},\ldots,{\widetilde{B}}_{r}),Z).\end{aligned}$$ The set $Z$ is usually called the [zero block]{} of $\pi$. Let $\pi=({\widetilde{\pi}},Z)$ be an element of ${\Pi^{\text{ord},B}_{n}}$, with ${\widetilde{\pi}}=({\widetilde{B}}_1,\ldots,{\widetilde{B}}_r)$. We define the cone $C_\pi$ to be the set of $(x_1,\ldots,x_n)\in V$ such that (with the convention that $x_{-i}=-x_i$ for $1 \leq i \leq n$): - if $Z=\{i_1,\ldots,i_m\}$, then the equalities $x_{i_1}= \cdots =x_{i_m}=0$ hold; - for every block ${\widetilde{B}}= \{i_1,\ldots,i_m\}$ in ${\widetilde{\pi}}$, the equalities and inequality $x_{i_1}= \cdots =x_{i_m}> 0$ hold; - for every two consecutive blocks ${\widetilde{B}}_{s}$ and ${\widetilde{B}}_{s+1}$ in ${\widetilde{\pi}}$ with $i \in {\widetilde{B}}_s$ and $j \in {\widetilde{B}}_{s+1}$, the inequality $\|x_i\| > \|x_j\|$ holds. It is easy to see that the map $\varphi:{\Pi^{\text{ord},B}_{n}}\longrightarrow {\mathscr{L}}$ sending $\pi$ to $C_\pi$ is a bijection, and that it induces an order-reversing isomorphism between the poset ${\Pi^{\text{ord},B}_{n}}$ and the face poset ${\mathscr{L}}$. The inverse image of the ray $C = {{\mathbb{R}}}_{>0}\cdot(e_1+e_2+\cdots+e_n)$ by this bijection is the element $\pi_0=((\{1,\ldots,n\}),\varnothing)$ of ${\Pi^{\text{ord},B}_{n}}$, so the elements of ${\mathscr{L}}_{\geq C}$ correspond exactly to the (unsigned) ordered partitions of $\{1,\ldots,n\}$. In other words, the bijection $\varphi$ induces an order-reversing isomorphism between the poset ${\Pi^{\text{ord}}_{n}}$ of ordered partitions of $\{1,\ldots,n\}$ defined in [@Ehrenborg_Morel_Readdy Section 2] and the poset ${\mathscr{L}}_{\geq C}$. Let $\lambda=(\lambda_1,\ldots,\lambda_n)\in{{\mathbb{R}}}^n$. For a signed block ${\widetilde{B}}$, we set $\lambda_{{\widetilde{B}}}=\sum_{i\in B} \lambda_i$, with the convention that $\lambda_{-i}=-\lambda_{i}$ for $1 \leq i \leq n$. Define the subset ${\Pi^{\text{ord},B}_{n}}(\lambda)$ of ${\Pi^{\text{ord},B}_{n}}$ by $${\Pi^{\text{ord},B}_{n}}(\lambda) = \left\{ (({\widetilde{B}}_{1}, {\widetilde{B}}_{2}, \ldots, {\widetilde{B}}_{r}),Z) \in {\Pi^{\text{ord},B}_{n}}\:\: : \:\: \sum_{i=1}^{s} \lambda_{{\widetilde{B}}_{i}} \geq 0 \text{ for } 1 \leq s \leq r \right\} .$$ Then an element $\pi$ of ${\Pi^{\text{ord},B}_{n}}$ is in ${\Pi^{\text{ord},B}_{n}}(\lambda)$ if and only if $C_\pi$ is in ${\mathscr{L}}_\lambda$. Moreover, the subset ${\mathscr{L}}_{\lambda,\geq C}$ corresponds to the set ${\Pi^{\text{ord}}_{n}}(\lambda)$ of ordered partitions $(B_1,\ldots,B_r)$ of $\{1,\ldots,n\}$ such that, for every $1 \leq s \leq r$, we have $\sum_{i=1}^s\lambda_{B_i}\geq 0$. This is almost the set $\mathcal{P}(\lambda)$ of [@Ehrenborg_Morel_Readdy Section 3]; the only difference is that the inequalities defining $\mathcal{P}(\lambda)$ are strict. So we can give the following identity relating these two sets: For every $\varepsilon\in{{\mathbb{R}}}$, let $\lambda_\varepsilon= (\lambda_1-\varepsilon,\ldots,\lambda_n-\varepsilon)$. Then, if $\varepsilon>0$ is sufficiently small, we have ${\Pi^{\text{ord}}_{n}}(\lambda_\varepsilon)=\mathcal{P}(\lambda)$. Let $B$ be the unique chamber of ${\mathscr{L}}$ that is on the positive side of every hyperplane, that is, $B=\{x_1>x_2>\cdots>x_n>0\}$. As we already observed, ${\mathscr{L}}_{\geq C}$ is isomorphic to the face poset of the arrangement ${\mathcal{H}}^{(1)}$, which is a Coxeter arrangement of type $A_{n-1}$, and the unique chamber of this arrangement containing $B$ corresponds to the identity element in the symmetric group $\mathfrak{S}_n$. It then follows from Proposition \[prop\_f\_B\] that the function $f_B:{\mathscr{L}}_{\geq C} \longrightarrow {\mathscr{T}}\cap{\mathscr{L}}_{\geq C}$ corresponds via $\varphi:{\Pi^{\text{ord}}_{n}}\stackrel{\sim}{\longrightarrow} {\mathscr{L}}_{\geq C}$ to the function $f:{\Pi^{\text{ord}}_{n}}\longrightarrow \mathfrak{S}_n$ of [@Ehrenborg_Morel_Readdy Section 4]. We obtain the equality: $$\psi_{{\mathcal{H}}/C}(B,\lambda) = \sum_{\pi\in{\Pi^{\text{ord}}_{n}}(\lambda)} (-1)^{\|\pi\|} \cdot (-1)^{f(\pi)},$$ where $\|\pi\|$ denotes the number of blocks of the ordered partition $\pi = (B_1, \ldots, B_r)$, in other words, $\|\pi\|=r$. Let $\overline{\lambda}$ denote the reverse of $\lambda$, that is, $\overline{\lambda}=(\lambda_n,\ldots,\lambda_1)$. If $\varepsilon\in{{\mathbb{R}}}$, we write $\overline{\lambda}_\varepsilon$ for $(\lambda_n-\varepsilon, \ldots,\lambda_1-\varepsilon)$. By [@Ehrenborg_Morel_Readdy Lemma 7.1], we have $$\psi_{{\mathcal{H}}/C}(B,\overline{\lambda}) = (-1)^{\binom{n}{2}} \cdot \sum_{\pi\in{\Pi^{\text{ord}}_{n}}(\lambda)}(-1)^{\|\pi\|} \cdot (-1)^{g(\pi)},$$ where $g:{\Pi^{\text{ord}}_{n}}\longrightarrow \mathfrak{S}_n$ is the function defined at the beginning of [@Ehrenborg_Morel_Readdy Section 6]. Finally, using the fact that ${\Pi^{\text{ord}}_{n}}(\lambda_\varepsilon)=\mathcal{P}(\lambda)$ for sufficiently small $\varepsilon>0$, then the sum $S(\lambda)$ of [@Ehrenborg_Morel_Readdy Section 6] is given by the expression: $$\label{eq_S_vs_psi} S(\lambda) = (-1)^{\binom{n}{2}} \cdot \psi_{{\mathcal{H}}/C}(B,\overline{\lambda}_\varepsilon)$$ for any sufficiently small $\varepsilon>0$. We now find an expression for the sum $T(\lambda)$ of [@Ehrenborg_Morel_Readdy Section 6] in terms of $2$-structures. As in [@Ehrenborg_Morel_Readdy], we denote by $M_n$ the set of maximal matchings on $\{1,2,\ldots,n\}$. Then we have a bijection $M_n \stackrel{\sim}{\longrightarrow} \mathcal{T}(\Phi)$ sending a matching $p=\{p_1,\ldots,p_m\}$, where $p_1=\{i_1<j_1\},\ldots,p_m=\{i_m<j_m\}$ are the edges of $p$, to the $2$-structure $\varphi_p=\{\pm(e_{i_1}-e_{j_1}),\ldots,\pm(e_{i_m}-e_{j_m})\}$. Moreover, we have $(-1)^p=\epsilon(\varphi_p)$. We can calculate $\psi_{{\mathcal{H}}_{\varphi_p}/C_{\varphi_p}}(B_{\varphi_p},\lambda)$ using Lemma \[lemma\_product\] for the decomposition $V=V_0\times V_1\times\cdots\times V_m$, where $V_k={{\mathbb{R}}}e_{i_k}+{{\mathbb{R}}}e_{j_k}$ for $1\leq k\leq m$, $V_0=\{0\}$ if $n$ is even, and $V_0={{\mathbb{R}}}e_i$ if $n$ is odd and $i$ is the unique unmatched element of $\{1,\ldots,n\}$. By Lemma \[lemma\_product\], we have $$\psi_{{\mathcal{H}}_{\varphi_p}/C_{\varphi_p}}(B_{\varphi_p},\lambda)= \prod_{k=1}^{m} d_2(\lambda_{i_k},\lambda_{j_k}) \cdot \begin{cases} 1 & \text{if $n$ is even,} \\ d_{1}(\lambda_{i}) & \text{if $n$ is odd,} \end{cases}$$ where: - The function $d_1:{{\mathbb{R}}}\longrightarrow {{\mathbb{R}}}$ is defined by $d_1(a)=\psi_{{\mathcal{H}}_1/C_1} (B_1,a)$, where ${\mathcal{H}}_1$ is the hyperplane arrangement $(H_e)$ on ${{\mathbb{R}}}e$ and $B_1=C_1={{\mathbb{R}}}_{>0}e$. - The function $d_2:{{\mathbb{R}}}^2 \longrightarrow {{\mathbb{R}}}$ is defined by $d_2(a,b)=\psi_{{\mathcal{H}}_2/C_2} (B_2,(a,b))$, where ${\mathcal{H}}_2$ is the hyperplane arrangement $(H_e,H_f, H_{e-f},H_{e+f})$ on ${{\mathbb{R}}}e\oplus{{\mathbb{R}}}f$, $C_2=\{\alpha e+\beta f : \alpha=\beta>0\}$ and $B_2=\{\alpha e+\beta f : \alpha>\beta>0\}$. In other words, the functions $d_1$ and $d_2$ are precisely the function $\psi_{{\mathcal{H}}/C}(B,\lambda)$ that we are trying to determine in the cases $n=1$ and $n=2$. A direct calculation yields: $$d_1(a)= \begin{cases} -1 & \text{ if } a\geq 0, \\ 0 & \text{ if } a<0, \end{cases} \:\:\:\: \text{ and } \:\:\:\: d_2(a,b)=\begin{cases} -1 & \text{ if } a,b \geq 0, \\ -2 & \text{ if } b \geq -a>0, \\ 0 & \text{ otherwise}. \end{cases}$$ Comparing this with the formula defining $c(p,\lambda)$ in [@Ehrenborg_Morel_Readdy Section 6], we see that, for all $a,b\in{{\mathbb{R}}}$, if $\varepsilon>0$ is sufficiently small relative to $a$ and $b$, we have $d_{1}(a-\varepsilon) = - c_{1}(a)$ and $d_{2}(a-\varepsilon,b-\varepsilon) = - c_{2}(b,a)$, and hence $$\begin{aligned} \psi_{{\mathcal{H}}_{\varphi_p}/C_{\varphi_p}}(B_{\varphi_p},\overline{\lambda}_\varepsilon) & = c(p,\lambda)\cdot \begin{cases}(-1)^{n/2} & \text{ if } n\text{ is even,} \\ (-1)^{(n+1)/2} & \text{ if }n\text{ is odd}\end{cases} \\ & = (-1)^{n} \cdot (-1)^{\binom{n}{2}} \cdot c(p,\lambda),\end{aligned}$$ if $\varepsilon>0$ is sufficiently small relative to the $\lambda_i$. Combining all these calculations, we see that if $\varepsilon>0$ is sufficiently small, then $$\sum_{\varphi\in\mathcal{T}(\Phi)} \epsilon(\varphi) \cdot \psi_{{\mathcal{H}}_\varphi/C_\varphi}(B_\varphi,\overline{\lambda}_ \varepsilon) = (-1)^{n} \cdot (-1)^{\binom{n}{2}} \cdot \sum_{p\in M_n}(-1)^p \cdot c(p,\lambda) = (-1)^{n} \cdot (-1)^{\binom{n}{2}} \cdot T(\lambda).$$ The identity $S(\lambda)=(-1)^n \cdot T(\lambda)$ in [@Ehrenborg_Morel_Readdy Theorem 6.4] now follows from Theorem \[thm\_main\], applied to $\overline{\lambda}_\varepsilon$ for $\varepsilon>0$ sufficiently small. Concluding remarks {#section_concluding_remarks} ================== As mentioned in the introduction, we are not aware of whether or not there is a representation-theoretic interpretation of the identity in Theorem \[thm\_main\]. In other words, what is the meaning of the constants $\psi_{{\mathcal{H}}/C}(B,\lambda)$ for different values of $\lambda$? In Appendix \[appendix\_valuations\] we construct a signed convolution of two valuations that yields a new valuation. We wonder if there are other convolutions of valuations that also yield valuations. For instance, can the signed convolution of Definition \[definition\_signed\_convolution\] be extended to valuations on (not necessarily polyhedral) cones in Euclidean space? The main results in this paper are for Coxeter arrangements. However, to obtain the recursion formula in Section \[section\_induction\_Coxeter\_case\] we first prove it for general hyperplane arrangements in Proposition \[prop\_induction\_formula\_for\_psi\]. Can the other side of the recursion also be viewed in a larger setting? That is, is there some analogue of $2$-structures for more general hyperplane arrangements? In Appendix \[appendix\_2\_structures\] the proof of Proposition \[prop\_W\_acts\_transitively\_on\_T\_Phi\] that the group $W$ acts transitively on the set of $2$-structures $\mathcal{T}$ consiste of verifying the result for all irreducible pseudo-root systems. Is there a general proof that does not use the classification of irreducible pseudo-root systems? Extending the construction of a valuation by Goresky, Kottwitz and MacPherson {#appendix_valuations} ============================================================================= We introduce a signed convolution of valuations on closed convex polyhedral cones in a finite-dimensional real vector space. As a special case we obtain in Corollary \[cor\_psi\_C\_is\_a\_valuation\] a valuation due to Goresky, Kottwitz and MacPherson; see [@GKM Proposition A.4]. Subsection \[subsection\_signed\_convolution\] of this appendix contains definitions and statements of results. The proofs are relegated to Subsection \[subsection\_signed\_convolution\_proofs\]. The signed convolution {#subsection_signed_convolution} ---------------------- Let $V$ be a finite-dimensional real vector space and $V^\vee$ be its dual. A closed convex polyhedral cone in $V$ is a subset of the form ${{\mathbb{R}}}_{\geq 0} v_{1} + {{\mathbb{R}}}_{\geq 0} v_{2} + \cdots + {{\mathbb{R}}}_{\geq 0} v_{k}$, where $v_{1}, v_{2}, \ldots, v_{k} \in V$ and $k\geq 0$. For a subset $X$ of the space $V$, define $X^{\perp} = \{ \alpha \in V^{\vee} : \forall x \in X, \: {\langle \alpha, x \rangle} = 0 \}$ and $X^{} = \{ \alpha \in V^{\vee} : \forall x \in X, \: {\langle \alpha, x \rangle} \geq 0 \}$. Note that $X^{\perp}$ is a subspace of $V^{\vee}$ and depends only on the linear span of $X$, and that $X^$ is a convex cone in $V^{\vee}$ and depends only on the closed convex polyhedral cone generated by $X$. For $F$ a face[^2] of a closed convex polyhedral cone $K$, define $F^{\perp,K} = F^{\perp} \cap K^{}$. The map $F \longmapsto F^{\perp,K}$ is an order-reversing bijection from the set of faces of $K$ to the set of faces of $K^$. This statement and other basic properties of closed convex polyhedral cones are proved in [@Fulton Section 1.2]. We denote by ${\mathcal{C}}(V)$ the set of closed convex polyhedral cones in $V$. A valuation on ${\mathcal{C}}(V)$ with values in an abelian group $A$ is a function $f:{\mathcal{C}}(V) \longrightarrow A$ such that $f(\varnothing)=0$ and that for any $K,K'\in{\mathcal{C}}(V)$ such that $K\cup K'\in{\mathcal{C}}(V)$, we have $$\begin{aligned} f(K\cup K')+f(K\cap K') & = f(K)+f(K').\end{aligned}$$ \[def\_valuation\] For $\lambda\in V^\vee$, we define the hyperplane $H_{\lambda}$ and the two open half-spaces $H^{+}_\lambda$ and $H^{-}_\lambda$ by $$H_\lambda=\{x\in V : {\langle \lambda, x \rangle}=0\}, \quad H^{+}_\lambda=\{x\in V : {\langle \lambda, x \rangle} > 0\} \quad \text{ and } \quad H^{-}_\lambda=\{x\in V : {\langle \lambda, x \rangle} < 0\}.$$ The closed half-spaces are given by $\overline{H_\lambda^+} = \{x\in V : {\langle \lambda, x \rangle} \geq 0\}$ and $\overline{H_\lambda^-} = \{x\in V : {\langle \lambda, x \rangle} \leq 0\}$. We have the following criterion for recognizing valuations on closed convex polyhedral cones. This is known as [Groemer’s first extension theorem]{} and is proved in [@Gr Theorem 2]. Let $A$ be an abelian group and $f:{\mathcal{C}}(V) \longrightarrow A$ be a function such that $f(\varnothing)=0$. Suppose that for every $K\in{\mathcal{C}}(V)$ and every $\mu \in V^\vee$ the following holds: $$\begin{aligned} f(K)+f(K\cap H_\mu) & = f\bigl(K\cap \overline{H_\mu^{+}}\bigr) + f\bigl(K \cap \overline{H_\mu^{-}}\bigr). \label{equation_Groemer}\end{aligned}$$ Then the function $f$ is a valuation. \[theorem\_Groemer\] Let $X$ be a subset of $V$ such that the complement $V - X$ is convex. Then the function $\phi_X: {\mathcal{C}}(V) \longrightarrow {{\mathbb{Z}}}$ defined by $$\phi_X(K) = \begin{cases} 1 & \text{ if } \emptyset \subsetneq K\subset X, \\ 0 & \text{ otherwise,} \end{cases}$$ is a valuation. In particular, if $\lambda\in V^\vee$ then the function $\psi_\lambda=\phi_{\overline{H_\lambda^+}}$ is a valuation. \[lemma\_GKM\] Let $K\in{\mathcal{C}}(V)$ be nonempty and let $\mu\in V^\vee$. Let $K_{0}=K\cap H_\mu$, $K_{+}=K\cap \overline{H_\mu^{+}}$ and $K_{-}=K\cap \overline{H_\mu^{-}}$. We must check Criterion in Theorem \[theorem\_Groemer\], that is, $\phi_X(K) + \phi_X(K_{0}) = \phi_X(K_{+}) + \phi_X(K_{-})$. If $K\subset X$ then $K_{0}$, $K_{+}$ and $K_{-}$ are also included in $X$, and the equality above is clear. If $K_{+}\subset X$ but $K_{-}\not\subset X$, then $K_{0}\subset X$ and $K\not\subset X$, so again the desired equality holds. The case where $K_{-}\subset X$ and $K_{+}\not\subset X$ is symmetric. Finally, suppose that $K_{+},K_{-}\not\subset X$. Then $K\not\subset X$, and so we must show that $K_{0}\not\subset X$. Take $x\in K_{+}-X$ and $y\in K_{-}-X$. Then the segment $[x,y]$ is contained in the convex set $V-X$. As this segment intersects $K_{0}$, this shows that $K_{0}\not\subset X$. Let $A$ be a ring, and let $f_{1}:{\mathcal{C}}(V) \longrightarrow A$ and $f_{2}:{\mathcal{C}}(V^\vee) \longrightarrow A$ be two functions. Define their [signed convolution]{} $f_{1} * f_{2}:{\mathcal{C}}(V) \longrightarrow A$ by $$\begin{aligned} (f_{1} * f_{2})(K) & = \sum_{F \in \mathcal{F}(K)} (-1)^{\dim(F)} \cdot f_{1}(F) \cdot f_{2}(F^{\perp,K}),\end{aligned}$$ where, for every $K\in\mathcal{C}(V)$, $\mathcal{F}(K)$ is the set of faces of the cone $K$. \[definition\_signed\_convolution\] The main result of this appendix is the following theorem whose proof is in Subsection \[subsection\_signed\_convolution\_proofs\]. The signed convolution of two valuations is a valuation. \[theorem\_signed\_convolution\] Given $x \in V$ and $\lambda \in V^{\vee}$, we have two valuations $\psi_\lambda:{\mathcal{C}}(V){\longrightarrow}{{\mathbb{Z}}}$ and $\psi_x:{\mathcal{C}}(V^\vee){\longrightarrow}{{\mathbb{Z}}}$ defined in Lemma \[lemma\_GKM\]. Let $K\longmapsto\psi_{K}(x,\lambda)$ be their signed convolution, that is, for $K\in{\mathcal{C}}(V)$, $$\begin{aligned} \psi_{K}(x,\lambda) & = (\psi_{\lambda} * \psi_{x})(K) .\end{aligned}$$ This function is defined in [@GKM Appendix A] (at the top of page 540). For every $x\in V$ and every $\lambda\in V^\vee$, the function $K \longmapsto\psi_K(x,\lambda)$ from ${\mathcal{C}}(V)$ to ${{\mathbb{R}}}$ is a valuation. \[cor\_psi\_C\_is\_a\_valuation\] Since any valuation satisfies the additivity property, we obtain the next corollary, which is [@GKM Proposition A.4]. \[Goresky–Kottwitz–MacPherson\] Let $K$ be a closed convex polyhedral cone. Suppose that its relative interior ${K^{\circ}}$ is the disjoint union of the relative interiors ${K_1^{\circ}}, {K_2^{\circ}}, \ldots, {K_r^{\circ}}$ of $r$ closed convex polyhedral cones $K_1, K_2, \ldots, K_r$. Then for every $x\in V$ and every $\lambda\in V^\vee$ $$\psi_K(x,\lambda) = \sum_{i=1}^r (-1)^{\dim(K) - \dim(K_i)} \cdot \psi_{K_i}(x,\lambda).$$ As a final note, valuations on ${\mathcal{C}}(V)$ can be extended to relatively open cones as well. Let $G$ be a collection of sets that is closed under finite intersections. Define $B(G)$ to be the Boolean algebra generated by $G$, that is, the smallest collection of sets that contains $G$ and is closed under finite unions, finite intersections and complements. Groemer’s Integral Theorem states that a valuation on $G$ can be extended to a valuation on the Boolean algebra $B(G)$; see [@Gr] and also [@Klain_Rota Chapter 2]. In the case where $G = {\mathcal{C}}(V)$, that is, the collection of closed convex polyhedral cones in $E$, the associated Boolean algebra $B({\mathcal{C}}(V))$ contains all cones that are obtained by intersecting closed and open half-spaces. Proofs {#subsection_signed_convolution_proofs} ------ Before proving Theorem \[theorem\_signed\_convolution\], we state and prove the following lemma. Let $K \subset V$ be a closed convex polyhedral cone, let $F$ be a face of the cone $K$ and let $\mu\in V^\vee$. We write $K_{0}=K\cap H_\mu$, $K_{+}=K\cap \overline{H^+_\mu}$ and $K_{-}=K\cap \overline{H^-_\mu}$. - Assume that $F\subset \overline{H_\mu^{+}}$ but $F\not\subset H_\mu$, that is, $F$ is a face of $K_{+} $ but not of $K_{0}$. Then the equality $F^{\perp}\cap K^=F^{\perp}\cap K_{+}^$ holds. - Assume that $F \cap H_\mu^+\not= \varnothing$ and $F \cap H_\mu^{-}\not= \varnothing$, in other words, the hyperplane $H_\mu$ cuts the face $F$ in two. Then the equality $F^{\perp,K} = F^{\perp}\cap K_{0}^{}$ holds. - In the situation of (b), let $F_{0}=F\cap H_\mu$. Then the equality $F_{0}^{\perp}\cap K^=F^{\perp,K}$ holds. \[lemma\_cut\_face\] We first prove (a). The inclusion $F^{\perp}\cap K^\subset F^{\perp}\cap K_{+}^$ clearly holds, so we just need to show the reverse inclusion. Note that $K_{+}^=K^+{{\mathbb{R}}}_{\geq 0}\mu$. Let $\lambda\in F^{\perp} \cap K_{+}^$, and write $\lambda=\lambda'+a\mu$, with $\lambda'\in K^$ and $a \geq 0$. Choose $x^{+} \in F \cap H_\mu^{+}$. Then we have $0 ={\langle \lambda, x^{+} \rangle} = {\langle \lambda', x^{+} \rangle} + a{\langle \mu, x^{+} \rangle}$, with ${\langle \lambda', x^{+} \rangle} \geq 0$ because $x^{+}\in F \subset K$ and ${\langle \mu, x^{+} \rangle} > 0$ because $x^{+} \in H_\mu^+$, so $a\leq 0$. Hence $a=0$, which implies that $\lambda\in K^$. We now prove (b). The relation $F^{\perp,K} = F^{\perp}\cap K^\subset F^{\perp}\cap K_{0}^$ clearly holds, so we just need to verify the reverse inclusion. Note that $K_{0}^=K^{}+{{\mathbb{R}}}\mu$. Let $\lambda \in F^{\perp} \cap K_{0}^$, and write $\lambda=\lambda'+a\mu$, with $\lambda'\in K^$ and $a \in {{\mathbb{R}}}$. Choose $x^{+} \in F\cap H_\mu^+$ and $x^{-} \in F\cap H_\mu^-$. Then we have $0 = {\langle \lambda, x^{+} \rangle} = {\langle \lambda', x^{+} \rangle} + a {\langle \mu, x^{+} \rangle}$, with ${\langle \lambda', x^{+} \rangle} \geq 0$ because $x^{+} \in F \subset K$, and ${\langle \mu, x^{+} \rangle} > 0$ because $x^{+} \in H_\mu^+$, so $a\leq 0$. On the other hand, we have $0 = {\langle \lambda, x^{-} \rangle} = {\langle \lambda', x^{-} \rangle} + a {\langle \mu, x^{-} \rangle}$, with ${\langle \lambda', x^{-} \rangle} \geq 0$ because $x^{-} \in F \subset K$, and ${\langle \mu, x^{-} \rangle} < 0$ because $x^{-} \in H_\mu^-$, so $a \geq 0$. This shows that $a=0$, hence that $\lambda\in K^$. Finally, we prove (c). We have $F^{\perp,K} \subset F_{0}^{\perp} \cap K^$, so we just need to show the reverse inclusion. Note that $F_{0}^{\perp}=F^{\perp}+{{\mathbb{R}}}\mu$. Let $\lambda\in F_{0}^{\perp}\cap K^$, and write $\lambda=\lambda'+a\mu$, with $\lambda'\in F^{\perp}$ and $a \in {{\mathbb{R}}}$. As in the proof of (b), choose $x^{+} \in F\cap H_\mu^+$ and $x^{-} \in F\cap H_\mu^-$. Then we have $0 \leq {\langle \lambda, x^{+} \rangle} = {\langle \lambda', x^{+} \rangle} + a {\langle \mu, x^{+} \rangle} = a {\langle \mu, x^{+} \rangle}$ and ${\langle \mu, x^{+} \rangle} > 0$, so $a\geq 0$. On the other hand, we have $0 \leq {\langle \lambda, x^{-} \rangle} = {\langle \lambda', x^{-} \rangle} + a {\langle \mu, x^{-} \rangle} = a {\langle \mu, x^{-} \rangle}$ and ${\langle \mu, x^{-} \rangle} < 0$, so $a \leq 0$. Hence we must have $a=0$, implying $\lambda \in F^{\perp}$. at (-0.5,3.7) [(a)]{}; (-1,0) – node\[very near end,above\] [$H_{\mu}$]{} (5.5,0); (0,-1) – (0,2) – (2,4) – (4,0) – (2,-2) – (0,-1); at (1.7,1.7) [$K_{+}$]{}; at (1.7,0.3) [$K_{0}$]{}; at (1.7,-1) [$K_{-}$]{}; at (-0.5,3.7) [(b)]{}; (-1,0) – (5.5,0); (0,-1) – (0,2) – (2,4) – (4,0) – (2,-2) – (0,-1); at (0.79,3.31) [$S_{(i)}$]{}; (0,2) – (2,4); at (4.4,0.3) [$S_{(iii)}$]{}; (4,0) circle (2pt); at (1.4,1.2) [$1$st term of $S_{(iv)}$]{}; (0,0) – (0,2); at (1.4,0.3) [$3$rd term of $S_{(iv)}$]{}; (0,0) circle (2pt); Let $g = f_1 * f_2$ be the signed convolution of the valuations $f_{1} : {\mathcal{C}}(V) \longrightarrow A$ and $f_{2} : {\mathcal{C}}(V^{\vee}) \longrightarrow A$. We check the criterion of Theorem \[theorem\_Groemer\]. Let $K\in{\mathcal{C}}(V)$ and let $\mu \in V^\vee$. We define as before three closed convex polyhedral cones $K_{+} = K \cap \overline{H_\mu^{+}}$, $K_{-} = K \cap \overline{H_\mu^{-}}$, $K_{0} = K \cap H_{\mu}$. Note that $K_{0}^=K_{+}^\cup K_{-}^$ and $K^=K_{+}^\cap K_{-}^$. The faces $F$ of the cone $K$ come in four disjoint categories. For each category, we consider the contribution to the sum defining $g(K)$. - $F$ is a face of $K_{+}$, but not of $K_{0}$, that is, ${F^{\circ}} \subset H_\mu^{+}$. Then by Lemma \[lemma\_cut\_face\](a) we have $F^{\perp}\cap K_{+}^* = F^{\perp}\cap K^$, that is, $F^{\perp,K}=F^{\perp,K_{+}}$. Hence the contribution is $$\begin{aligned} S_{(i)} & = \sum_{\substack{F \in \mathcal{F}(K) \cap \mathcal{F}(K_{+}) \\ F\not\in \mathcal{F}(K_{0})}} (-1)^{\dim(F)} \cdot f_{1}(F) \cdot f_{2}(F^{\perp,K}) \\ & = \sum_{F \in (\mathcal{F}(K_{+})\cap\mathcal{F}(K))-\mathcal{F}(K_{0})} (-1)^{\dim(F)} \cdot f_{1}(F) \cdot f_{2}(F^{\perp,K_{+}}) .\end{aligned}$$ - $F$ is a face of $K_{-}$, but not of $K_{0}$, that is, ${F^{\circ}} \subset H_\mu^{-}$. As in case (i), we have $F^{\perp,K}=F^{\perp,K_{-}}$, and the contribution is $$\begin{aligned} S_{(ii)} & = \sum_{\substack{F \in \mathcal{F}(K) \cap \mathcal{F}(K_{-}) \\ F\not\in \mathcal{F}(K_{0})}} (-1)^{\dim(F)} \cdot f_{1}(F) \cdot f_{2}(F^{\perp,K}) \\ & = \sum_{F \in (\mathcal{F}(K_{-})\cap\mathcal{F}(K))-\mathcal{F}(K_{0})} (-1)^{\dim(F)} \cdot f_{1}(F) \cdot f_{2}(F^{\perp,K_{-}}).\end{aligned}$$ - $F$ is a face of all three cones $K_{+}$, $K_{-}$ and $K_{0}$, that is, we have $F \subset H_{\mu}$. Here the contribution is $$\begin{aligned} S_{(iii)} & = \sum_{\substack{F \in \mathcal{F}(K) \\ F \subset H_{\mu}}} (-1)^{\dim(F)} \cdot f_{1}(F) \cdot f_{2}(F^{\perp,K}) \\ & = \sum_{F \in \mathcal{F}(K_{+})\cap\mathcal{F}(K_{-})\cap\mathcal{F}(K_{0})} (-1)^{\dim(F)} \cdot f_{1}(F) \cdot \bigl(f_{2}(F^{\perp,K_{+}}) + f_{2}(F^{\perp,K_{-}}) - f_{2}(F^{\perp,K_{0}})\bigr) ,\end{aligned}$$ since $f_{2}$ is a valuation and $F^{\perp,K_{+}} \cup F^{\perp,K_{-}} = F^{\perp,K_{0}}$ and $F^{\perp,K_{+}} \cap F^{\perp,K_{-}} = F^{\perp,K}$. - The face $F$ gets cut into three faces: $F_{+} = F\cap K_{+}$ in $K_{+}$, $F_{-} = F\cap K_{-}$ in $K_{-}$ and $F_{0} = F\cap K_{0}$ in $K_{0}$. Then we have $F^{\perp}=F_{+}^{\perp}=F_{-}^{\perp}$ because $F$, $F_{+}$ and $F_{-}$ have the same span. By Lemma \[lemma\_cut\_face\](b), we have $F^{\perp}\cap K_{0}^= F^{\perp}\cap K^$, and so $$F^{\perp,K}=F_{+}^{\perp,K_{+}}=F_{-}^{\perp,K_{-}}=F^{\perp}\cap K_{0}^.$$ By Lemma \[lemma\_cut\_face\](c), we also have $F_{0}^{\perp}\cap K^=F^{\perp,K}$. So the contribution is $$\begin{aligned} S_{(iv)} & = \sum_{\substack{F \in \mathcal{F}(K) \\ F \text{ being cut}}} (-1)^{\dim(F)} \cdot f_{1}(F) \cdot f_{2}(F^{\perp,K}) \\ & = \sum_{\substack{F \in \mathcal{F}(K) \\ F \text{ being cut}}} (-1)^{\dim(F)} \cdot (f_{1}(F_{+})+f_{1}(F_{-})-f_{1}(F_{0})) \cdot f_{2}(F^{\perp,K}) \\ & = \sum_{\substack{F \in \mathcal{F}(K) \\ F \text{ being cut}}} (-1)^{\dim(F)} \cdot (f_{1}(F_{+}) f_{2}(F_{+}^{\perp,K_{+}})+f_{1}(F_{-}) f_{2}(F_{-}^{\perp,K_{-}})-f_{1}(F_{0}) f_{2}(F^{\perp,K})) \intertext{Using the fact that $F^{\perp,K}=F_{0}^{\perp}\cap K^$, we get $f_{2}(F^{\perp,K}) = f_{2}(F_{0}^{\perp,K_{+}}) + f_{2}(F_{0}^{\perp,K_{-}}) - f_{2}(F_{0}^{\perp,K_{0}})$ and so} S_{(iv)} & = \sum_{\substack{F \in \mathcal{F}(K) \\ F \text{ being cut}}} (-1)^{\dim(F)} \cdot \biggl( f_{1}(F_{+}) f_{2}(F_{+}^{\perp,K_{+}}) + f_{1}(F_{-}) f_{2}(F_{-}^{\perp,K_{-}})\\ & \hspace{40mm} - f_{1}(F_{0}) f_{2}(F_{0}^{\perp,K_{+}}) - f_{1}(F_{0}) f_{2}(F_{0}^{\perp,K_{-}}) + f_{1}(F_{0}) f_{2}(F_{0}^{\perp,K_{0}})\biggr).\end{aligned}$$ Now expand $g(K)$ as $S_{(i)}+S_{(ii)}+S_{(iii)}+S_{(iv)}$. We use use the fact that $(-1)^{\dim(F)} = - (-1)^{\dim(F_{0})}$ in the third, fourth and fifth terms of $S_{(iv)}$. The contributions to $g(K_{+})$, respectively $g(K_{-})$, are given by the sum $S_{(i)}$, respectively $S_{(ii)}$, the first term in the sum $S_{(iii)}$, respectively the second term, and the first and third terms in the sum $S_{(iv)}$, respectively the second and fourth terms. See Figure \[figure\_one\] (b). Finally, the third term of the sum $S_{(iii)}$ and the fifth term of the sum $S_{(iv)}$ yield the sum for $-g(K_{0})$, which proves that $g(K)=g(K_{+})+g(K_{-})-g(K_{0})$. Review of 2-structures {#appendix_2_structures} ====================== The concept of $2$-structure for a root system was introduced by Herb to calculate discrete series characters on real reductive groups. See for example Section 5 of [@Herb-DSC] or Section 4 of the review article [@Herb-2S]. In this section we review Herb’s constructions and adapt them so that they work for an arbitrary Coxeter system having finite Coxeter group. We fix a finite-dimensional ${{\mathbb{R}}}$-vector space $V$ and an inner product $(\cdot,\cdot)$ on $V$. For every $v\in V-\{0\}$, we denote by $s_v$ the (orthogonal) reflection across the hyperplane $v^\perp$. Whenever we need to describe the irreducible root systems, we use the description given in the tables at the end of [@Bourbaki], except that we write $(e_1,\ldots,e_n)$ for the canonical basis of ${{\mathbb{R}}}^n$. When we need a system of positive roots in these root systems, we also use the ones given in these tables. This appendix is organized as follows. Subsections \[subsection\_B\_1\] and \[subsection\_2-structures\] contain the definitions and results respectively. Subsections \[subsection\_B\_3\] and \[section\_2\_structures\_irreducibles\] contain the technical proofs. The verification that the Coxeter group $W$ acts transitively on the set of $2$-structures takes place in the fourth subsection. Pseudo-root systems {#subsection_B_1} ------------------- A finite subset $\Phi$ of $V-\{0\}$ is called a pseudo-root system* if it satisfies the following conditions: - for every $\alpha\in\Phi$, we have $\Phi\cap{{\mathbb{R}}}\alpha=\{\pm\alpha\}$; - for every $\alpha,\beta\in\Phi$, the reflection $s_\alpha$ sends $\beta$ to a vector of the form $c\gamma$, with $c\in{{\mathbb{R}}}_{>0}$ and $\gamma\in\Phi$. If all the elements of $\Phi$ are unit vectors, we call $\Phi$ a normalized pseudo-root system. In that case, condition (b) become “$s_\alpha(\beta)\in\Phi$”. \[def\_pseudo\_root\_system\] We use this definition because it is convenient in the context of Coxeter systems. A root system (in the usual sense) is a pseudo-root system, which is not normalized in general. The converse is not true, even if we allow ourselves to replace the elements of $\Phi$ by scalar multiples, because of the existence of non-crystallographic Coxeter systems (see Proposition \[prop\_pseudo\_root\_system\_vs\_Coxeter\_system\]). Pseudo-root systems are called “root systems” in [@Hu-Cox Section 1.2] and [@BB Section 4.4]. We avoid this terminology because it is not compatible with the established definition of root systems in representation theory. If $\Phi$ is normalized or an actual root system then the group $W$ preserves $\Phi$, so the action of $W$ on $V$ restricts to an action of $W$ on $\Phi$. In general, we can still make $W$ act on $\Phi$ by declaring that if $w\in W$ and $\alpha\in\Phi$ then $w\cdot\alpha$ is the unique element $\beta$ of $\Phi$ such that $w(\alpha)\in{{\mathbb{R}}}_{>0}\beta$. This reduces to the previous action if $\Phi$ is normalized or an actual root system. Whenever we write an element of $W$ acting on an element of $\Phi$, this is the action that we mean. \[rmk\_action\_W\_Phi\] Let $\Phi\subset V$ be a pseudo-root system. A subset $\Delta$ of $\Phi$ is called a system of simple pseudo-roots if - The set $\Delta$ is a vector space basis for the linear span of $\Phi$. - For every $\alpha\in\Phi$, we can write $\alpha=\sum_{\beta\in\Delta} n_\beta\beta$, where the coefficients $n_\beta$ are in ${\mathbb{R}}$ and they are either all nonnegative or all nonpositive. The corresponding system of positive pseudo-roots is then $$\Phi^+ = \Phi \cap \biggl\{\sum_{\beta\in\Delta} n_\beta \beta\ : n_\beta\in{{\mathbb{R}}}_{\geq 0}\ \forall\beta\in\Delta\biggr\}.$$ We also write $\Phi^- = - \Phi^+$. \[def\_positive\_pseudo\_roots\] Let $\Phi\subset V$ be a pseudo-root system. We say that $\Phi$ is irreducible if there is no partition $\Phi=\Phi_1\sqcup\Phi_2$, with $\Phi_1$ and $\Phi_2$ nonempty pseudo-root systems such that $(\alpha_1,\alpha_2)=0$ for every $\alpha_1\in\Phi_1$ and every $\alpha_2\in\Phi_2$. The following two statements hold: - ([@Hu-Cox Section 1.9] and [@Hu-Cox Section 1.4].) Let $\Phi\subset V$ be a pseudo-root system and $\Delta\subset\Phi$ be a system of simple pseudo-roots. Let $W=W(\Phi)$ be the subgroup of ${\mathbf{GL}}(V)$ generated by the reflections $s_\alpha$ for $\alpha\in\Phi$, and let $S=\{s_\alpha : \alpha\in\Delta\}$. Then $(W,S)$ is a Coxeter system where $W$ is finite, and the Coxeter graph of $(W,S)$ is connected if and only if $\Phi$ is irreducible. Moreover, $W$ acts transitively on the set of systems of positive pseudo-roots if we use the action of Remark \[rmk\_action\_W\_Phi\]. - ([@Hu-Cox Section 5.4].) Conversely, let $(W,S)$ be a Coxeter system with $W$ finite, and let $\rho:W \longrightarrow {\mathbf{GL}}(V)$ be its canonical representation on $V=\bigoplus_{s\in S} {{\mathbb{R}}}e_s$ (see the beginning of Subsection \[section\_Coxeter\_arrangements\]). Then $\Phi=\{\rho(w)(e_s) : w\in W,\ s\in S\}$ is a normalized pseudo-root system and $\Delta=\{e_s : s\in S\}$ is a system of simple pseudo-roots in $\Phi$. \[prop\_pseudo\_root\_system\_vs\_Coxeter\_system\] Let $\Phi\subset V$ be an irreducible pseudo-root system. We say that $\Phi$ is of type $A_n$, respectively $B_n$, $D_n$, $E_6$, $E_7$, $E_8$, $F_4$, $H_3$, $H_4$, $I_2(m)$ with $m \geq 3$, if the corresponding Coxeter system is of that type. Here we use the classification of simple finite Coxeter systems proved in [@GB Chapter 5]. See Table 1 in [@BB Appendix A]. The Coxeter group of type $I_2(m)$ is the dihedral group of order $2m$. Note that types $I_2(3)$ and $A_2$ are isomorphic, types $I_2(4)$ and $B_2$ are isomorphic, and types $I_2(6)$ and $G_2$ are isomorphic. We did not include $I_2(2)$ in the list of irreducible types, because the corresponding Coxeter system is not irreducible, as it is isomorphic to $A_1\times A_1$. We will use the following lemma when introducing the sign associated to a $2$-structure in Proposition \[prop\_sign\_2\_structure\]. Recall that, if $r\geq 1$, then the lexicographic order on ${{\mathbb{R}}}^r$ is defined by $(x_1,\ldots,x_r) < (y_1,\ldots,y_r)$ if there exists $1 \leq i \leq r$ such that $x_i < y_i$ and that $x_j=y_j$ for $1\leq j\leq i-1$. It is a total order. Furthermore we say that a vector $x$ is [positive]{} if $x > (0,0, \ldots, 0)$. Let $\Phi\subset V$ be a pseudo-root system. Let $v_1, v_2, \ldots, v_r$ be linearly independent elements of $V$ such that no element of $\Phi$ is orthogonal to every $v_i$. Define $\Phi^+$ to be the set of $\alpha\in\Phi$ such that the element $((\alpha,v_1),(\alpha,v_2),\ldots, (\alpha,v_r))$ of ${{\mathbb{R}}}^r$ is positive with respect to the lexicographic order on ${{\mathbb{R}}}^r$. Then $\Phi^+$ is a system of positive pseudo-roots. \[lemma\_defining\_positive\_roots\] We complete $(v_1,\ldots,v_r)$ to a basis $(v_1,\ldots,v_n)$ of $V$, where $n$ is the dimension of $V$. If $v,w\in V$, we say that $v<w$ if $((v,v_1),\ldots,(v,v_n))<((w,v_1),\ldots,(w,v_n))$ in the lexicographic order on ${{\mathbb{R}}}^n$. This defines a total order on $V$ in the sense of [@Hu-Cox Section 1.3], and $\Phi^+$ is the corresponding positive system in $\Phi$. By the theorem in [@Hu-Cox Section 1.3], $\Phi^+$ is a system of positive pseudo-roots in the sense of Definition \[def\_positive\_pseudo\_roots\]. If $\theta=(v_1,\ldots,v_r)$ is a sequence of linearly independent elements of $V$ such that $\theta^\perp\cap\Phi=\varnothing$, we denote the system of positive pseudo-roots of Lemma \[lemma\_defining\_positive\_roots\] by $\Phi^+_\theta$. \[def\_Phi\_theta\] 2-structures {#subsection_2-structures} ------------ We define $2$-structures, generalizing a notion introduced by Herb for root systems; see for example the beginning of [@Herb-2S Section 4]. We also generalize some of the results of [@Herb-DSC Section 5] to Coxeter systems with finite Coxeter groups. We fix a pseudo-root system $\Phi$ in $V$ and a system of positive pseudo-roots $\Phi^+\subset\Phi$. We denote by $(W,S)$ the corresponding Coxeter system (see Proposition \[prop\_pseudo\_root\_system\_vs\_Coxeter\_system\]). A $2$-structure for $\Phi$ is a subset $\varphi$ of $\Phi$ satisfying the following properties: - The subset $\varphi$ is a disjoint union $\varphi=\varphi_1\sqcup\varphi_2\sqcup\cdots\sqcup\varphi_r$, where the $\varphi_i$ are pairwise orthogonal subsets of $\varphi$ and each of them is an irreducible pseudo-root system of type $A_1$, $B_2$ or $I_2(2^n)$, for $n\geq 3$. - Let $\varphi^+=\varphi\cap\Phi^+$. If $w\in W$ is such that $w(\varphi^+)=\varphi^+$ then $\det(w) = 1$. \[def\_2\_structure\] Although condition (b) involves the set of positive pseudo-roots $\varphi^+$ in $\varphi$, it does not actually depend on the choice of $\varphi^+$, because the Coxeter group of $\varphi$ acts transitively on sets of positive pseudo-roots in $\varphi$. If $\varphi \subseteq \Phi$ is a $2$-structure then there is no $\alpha\in\Phi$ that is orthogonal to every element of $\varphi$. Indeed, if such an $\alpha$ existed then the associated reflection $s_\alpha$ would fix every element of $\varphi$, and in particular send $\varphi^+$ to itself, which would contradict condition (b) of Definition \[def\_2\_structure\]. \[rmk\_no\_orthogonal\_root\] Let $\mathcal{T}(\Phi) \subseteq 2^{\Phi}$ be the set of all $2$-structures for the pseudo-root system $\Phi$. The following proposition is proved in Subsection \[section\_2\_structures\_irreducibles\], where we also show that each irreducible pseudo-root system contains a $2$-structure and give the type of this $2$-structure. This introduces no circularity in the arguments: the only results in this appendix that depend on Proposition \[prop\_W\_acts\_transitively\_on\_T\_Phi\] are Lemmas \[lemma\_T\_Phi\_induction\_step\] and \[lemma\_r\_Phi\], and these lemmas are not used in Subsections \[subsection\_B\_3\] and \[section\_2\_structures\_irreducibles\]. The group $W$ acts transitively on the collection of $2$-structures $\mathcal{T}(\Phi)$. \[prop\_W\_acts\_transitively\_on\_T\_Phi\] Let $\varphi\in\mathcal{T}(\Phi)$. We write $\varphi^+=\varphi\cap\Phi^+$ and $\varphi^-=\varphi\cap\Phi^-$, and we define $$\begin{aligned} W(\varphi,\Phi^+) & = \{w\in W:w(\varphi^+)\subset\Phi^+\} , \\ W_1(\varphi,\Phi^+) & = \{w\in W:w(\varphi^+)\subset\varphi^+\} = \{w\in W:w(\varphi^+) = \varphi^+\} .\end{aligned}$$ Note that $W_1(\varphi,\Phi^+)$ is a subgroup of $W$, and that the subset $W(\varphi,\Phi^+)$ of $W$ is stable by right translations by elements of $W_1(\varphi,\Phi^+)$. Let $\varphi\in\mathcal{T}(\Phi)$. Then the map $W \longrightarrow \mathcal{T}(\Phi)$, $w \longmapsto w(\varphi)$ induces a bijection $$W(\varphi,\Phi^+)/W_1(\varphi,\Phi^+) {\stackrel{\sim}{{\longrightarrow}}}\mathcal{T}(\Phi).$$ \[cor\_T\_Phi\_as\_quotient\] We denote by $f:W{\longrightarrow}\mathcal{T}(\Phi)$ the map defined by $f(w)=w(\varphi)$. If $u\in W_1(\varphi,\Phi^+)$, then $u(\varphi)=\varphi$, so $f(wu)=f(w)$ for every $w\in W$. So the map $f$ does induce a map from $W(\varphi,\Phi^+)/W_1(\varphi,\Phi^+)$ to $\mathcal{T}(\Phi)$, that we denote by $\overline{f}$. We show that $\overline{f}$ is surjective. Let $\varphi'\in \mathcal{T}(\Phi)$. By Proposition \[prop\_W\_acts\_transitively\_on\_T\_Phi\], there exists $w\in W$ such that $w(\varphi)=\varphi'$. By the theorem in [@Hu-Cox Section 1.3], the set $w^{-1}(\Phi^+)\cap\varphi$ is a system of positive pseudo-roots in $\varphi$, so, by Proposition \[prop\_pseudo\_root\_system\_vs\_Coxeter\_system\], there exists $v\in W(\varphi)$, where $W(\varphi)$ is the Coxeter group of $\varphi$, such that $v(\varphi^+)=w^{-1}(\Phi^+)\cap\varphi$. Then $wv(\varphi^+)=\Phi^+\cap w(\varphi)\subset\Phi^+$, so $wv\in W(\varphi,\Phi^+)$, and $wv(\varphi)=w(\varphi)=\varphi'$, that is, $f(wv)=\varphi'$. We show that $\overline{f}$ is injective. Let $w,w'\in W(\varphi,\Phi^+)$ such that $w(\varphi)=w'(\varphi)$. Then we have $w^{-1}w'(\varphi)=\varphi$, and, again by the theorem in [@Hu-Cox Section 1.3], the set $w^{-1}w'(\varphi^+)$ is a system of positive pseudo-roots in $\varphi$, so there exists $v\in W(\varphi)$ such that $v^{-1}w^{-1}w'(\varphi^+)=\varphi^+$. This means that we have $w'=wvu$ with $u\in W_1(\varphi,\Phi^+)$. So we will be done if we show that $v=1$. Note that $wv(\varphi^+)=wvu(\varphi^+)= w'(\varphi^+)\subset\Phi^+$. Suppose that $v\not=1$; then there exists $\alpha\in\varphi^+$ such that $v(\alpha)\in \varphi^-$, and then $wv(\alpha)=-w(-v(\alpha))\in \Phi^-$ (because $w\in W(\varphi,\Phi^+)$), contradicting the fact that $wv(\varphi^+) \subset\Phi^+$. So $v=1$. The following proposition, which follows immediately from Lemma 5.6 of [@Herb-DSC] for root systems, can be deduced from Lemmas \[lemma\_one\_structures\] and \[lemma\_choice\_of\_theta\] for irreducible root systems not of type $G_2$, and proved via a direct calculation for the remaining irreducible types. We will not need this result, so we do not go into details. Let $\mathcal{T}_{(a)}(\Phi)$ be the set of $\varphi \subseteq \Phi$ that satisfy condition (a) of Definition \[def\_2\_structure\]. Then $\mathcal{T}(\Phi)$ is exactly the set of elements of $\mathcal{T}_{(a)}(\Phi)$ that are maximal with respect to inclusion. \[prop\_maximality\_of\_2\_structures\] Let $\varphi \subseteq \Phi$ be a $2$-structure. Define an ordered subset $\theta$ of $\varphi$ as follows. Select a linear order of the irreducible components $\varphi_1, \varphi_2, \ldots, \varphi_r$ of $\varphi$. If $\varphi_{i}$ is a pseudo-root system of type $A_1$, let $\theta_{i}$ be the singleton $\varphi_{i} \cap \varphi^{+}$. If $\varphi_{i}$ is a pseudo-root system of type $B_{2}$ or $I_{2}(2^{k})$ for $k \geq 3$, pick two orthogonal elements $\alpha$ and $\alpha'$ from $\varphi_{i} \cap \varphi^{+}$ such that $\varphi_{i} \cap \varphi^{+}= \varphi^+_{i,(\alpha,\alpha')}$, that is, such that an element $\beta$ of $\varphi_i$ is in $\varphi^+$ if and only if either $(\beta,\alpha)>0$, or $(\beta,\alpha)=0$ and $(\beta,\alpha')>0$. Let $\theta_{i}$ be the sequence $(\alpha, \alpha')$. Finally let $\theta$ be the concatenation of the sequences $\theta_{1}, \theta_{2}, \ldots, \theta_{r}$. Let $\Phi_\theta^+$ be the system of positive pseudo-roots defined by the sequence $\theta$ as in Lemma \[lemma\_defining\_positive\_roots\], and let $w_\theta$ be the unique element of $W$ such that $w_\theta\cdot \Phi^+=\Phi_\theta^+$. Then the sign $\det(w_\theta)$ depends only on $\varphi$ and not on the choices made to form $\theta$. \[prop\_sign\_2\_structure\] in [0, 22.5, ..., 337.5]{} [([3\cos()]{},[3\sin()]{}) – ([-3\cos()]{},[-3\sin()]{});]{}; (-2.5,2.5) node[$\Phi^+$]{}; ([4\cos(-11.25)]{},[4\sin(-11.25)]{}) – ([-4\cos(-11.25)]{},[-4\sin(-11.25)]{}); (0,0) – (0,2) node\[right\][$\alpha$]{}; (0,0) – (2,0) node\[above\][$\alpha'$]{}; in [0, 22.5, ..., 337.5]{} [([3\cos()]{},[3\sin()]{}) – ([-3\cos()]{},[-3\sin()]{});]{}; (-2.5,2.5) node[$\Phi^+$]{}; ([4\cos(-11.25)]{},[4\sin(-11.25)]{}) – ([-4\cos(-11.25)]{},[-4\sin(-11.25)]{}); (0,0) – ([2\cos(3\22.5)]{},[2\sin(3\22.5]{}) node\[left\][$\alpha$]{}; (0,0) – ([2\cos(7\22.5)]{},[2\sin(7\22.5]{}) node\[above\][$\alpha'$]{}; Note that there are several choices when producing the ordered set $\theta$. First we have to select an order $\varphi_1, \varphi_2, \ldots, \varphi_r$. There are $r!$ ways to do this. Second, if $\varphi_i$ is of type $B_{2}$ or of type $I_{2}(2^{k})$, there are two possible choices for the pseudo-roots $\alpha$ and $\alpha'$; see Figure \[figure\_choice\_theta\]. These selections do not influence the sign of $w_\theta$, although they do of course affect the set $\Phi^+_\theta$. Let $\theta$ and $\theta'$ be the results of two possible sequences of choices. For an element $w$ in $W$, recall that its length $\ell(w)$ [^3] is also given by the cardinality of the intersection $w \cdot \Phi^{+} \cap \Phi^{-}$; see [@BB Proposition 4.4.4]. Note that $\Phi^+_{\theta} \cap \Phi^-_{\theta'} = w_{\theta} \cdot \Phi^+ \cap w_{\theta'} \cdot \Phi^- = w_{\theta'} \cdot (w_{\theta'}^{-1} w_{\theta} \cdot \Phi^+ \cap \Phi^-)$ which has cardinality $\ell(w_{\theta'}^{-1} w_{\theta})$. Hence to prove that the signs agree, that is, that $\det(w_\theta)=\det(w_{\theta'})$, it suffices to show that the set $\Phi^+_{\theta} \cap \Phi^-_{\theta'}$ has an even number of elements. We can reduce to the following two cases: - there exists $1 \leq i \leq r$ such that $\theta$ and $\theta'$ differ only by the choice of the two pseudo-roots in the factor $\varphi_i$; - there exists $1 \leq i \leq r-1$ such that $\theta_i=\theta'_{i+1}$, $\theta_{i+1}=\theta'_i$ and $\theta_j=\theta'_j$ if $j\not=i,i+1$. We begin by treating case (a). We write $\theta_i=(\alpha, \alpha')$ and $\theta'_i=(\beta, \beta')$. Let $\gamma \in \Phi^+_\theta \cap \Phi^-_{\theta'}$. Then $\gamma$ is orthogonal to $\varphi_1,\ldots,\varphi_{i-1}$, and it is not orthogonal to $\varphi_i$. Also, as the sets of positive pseudo-roots in $\varphi_i$ defined by $\theta_i$ and $\theta'_i$ are equal by assumption, we cannot have $\gamma\in\varphi_i$. Write $\gamma=c\alpha+c'\alpha'+\lambda$, with $\lambda\in\varphi_1^\perp\cap\cdots\cap\varphi_i^\perp$. By the previous sentence, we have $\lambda\not=0$. The vector $\iota(\gamma)=-(s_{\alpha}s_{\alpha'})(\gamma)=c\alpha+c'\alpha'-\lambda$ is also in $\Phi$. It is not equal to $\gamma$ because $\lambda\not=0$, and it is in $\Phi^+_\theta \cap \Phi^-_{\theta'}$ because $\gamma$ and $\iota(\gamma)$ have the same inner product with any element of the set $\{\alpha,\alpha',\beta,\beta'\}$. Note that we clearly have $\iota(\iota(\gamma))=\gamma$. We have constructed a fixed-point free involution $\iota$ on the set $\Phi^+_\theta \cap \Phi^-_{\theta'}$, which proves that this set has even cardinality. We treat case (b). Suppose first that $\varphi_i$ and $\varphi_{i+1}$ are both of type $A_1$, so we can write $\theta_i=(\alpha_i)$ and $\theta_{i+1}=(\alpha_{i+1})$. Let $\Phi'$ be the pseudo-root system $\Phi\cap({{\mathbb{R}}}\alpha_i+{{\mathbb{R}}}\alpha_{i+1})$. If $\Phi'$ is of type $I_2(m)$ with $m\geq 3$, then $m$ must be even because $\Phi'$ contains two orthogonal pseudo-roots. But then $\Phi'$ contains a multiple $\beta$ of $\alpha_i-\alpha_{i+1}$, and the reflection $s_\beta$ sends $\varphi^+$ to $\varphi^+$ because it fixes every element of $\varphi_j$ for $j=i,i+1$ and exchanges $\alpha_i$ and $\alpha_{i+1}$, contradicting the definition of a $2$-structure. Hence $\Phi'$ is of type $A_1\times A_1$, and then the fact that $\|\Phi^+_{\theta} \cap \Phi^-_{\theta'}\|$ is even follows from Lemma \[lemma\_change\_order\_theta\_sign\]. Suppose that $\varphi_i$ is of type $A_1$ and $\varphi_{i+1}$ is of type $I_2(2^m)$ with $m\geq 2$. Then we can write $\theta_i=(\alpha_i)$ and $\theta_{i+1}=(\alpha_{i+1},\alpha'_{i+1})$. Let $\Phi'$, respectively $\Phi''$, be the pseudo-root system $\Phi\cap({{\mathbb{R}}}\alpha_i+{{\mathbb{R}}}\alpha_{i+1})$, respectively $\Phi\cap({{\mathbb{R}}}\alpha_i+{{\mathbb{R}}}\alpha'_{i+1})$, and let $\theta''$ be the sequence that we obtain from $\theta$ by switching $\alpha_i$ and $\alpha_{i+1}$. As $\varphi_{i+1}$ is of type $I_2(2^m)$, it (and hence $\Phi$) contains a pseudo-root $\beta$ proportional to $\alpha_{i+1}-\alpha'_{i+1}$, and then $s_\beta(\Phi')=\Phi''$, so $\Phi'$ and $\Phi''$ are of the same type. By Lemma \[lemma\_change\_order\_theta\_sign\], the cardinalities of the sets $\Phi^+_\theta \cap \Phi^-_{\theta''}$ and $\Phi^+_{\theta''} \cap \Phi^-_{\theta'}$ have the same parity, and so $\|\Phi^+_{\theta} \cap \Phi^-_{\theta'}\|$ is even. The case where $\varphi_i$ is of rank $2$ and $\varphi_{i+1}$ of rank $1$ follows from the previous case by switching the roles of $\varphi_i$ and $\varphi_{i+1}$. Finally, suppose that both $\varphi_i$ and $\varphi_{i+1}$ are of rank $2$. Then we can write $\theta_i=(\alpha_i,\alpha_i')$ and $\theta_{i+1}=(\alpha_{i+1},\alpha'_{i+1})$. We move from $\theta$ to $\theta'$ by the following sequence of operations: - We switch $\alpha_i'$ and $\alpha_{i+1}$. By Lemma \[lemma\_change\_order\_theta\_sign\] and Remark \[remark\_change\_order\_theta\_sign\], this changes the sign of $w_\theta$ by $(-1)^{m_1/2 - 1}$, where the pseudo-root system $\Phi_1=\Phi\cap({{\mathbb{R}}}\alpha_i'+{{\mathbb{R}}}\alpha_{i+1})$ is of type $I_2(m_1)$. - We switch $\alpha_i'$ and $\alpha_{i+1}'$. By the same lemma and remark, this changes the sign by $(-1)^{m_2/2 - 1}$, where the pseudo-root system $\Phi_2 = \Phi\cap({{\mathbb{R}}}\alpha_i'+{{\mathbb{R}}}\alpha'_{i+1})$ is of type $I_2(m_2)$. - We switch $\alpha_{i}$ and $\alpha_{i+1}$. By the same lemma and remark, this changes the sign by $(-1)^{m_3/2 - 1}$, where the pseudo-root system $\Phi_3=\Phi\cap({{\mathbb{R}}}\alpha_{i}+{{\mathbb{R}}}\alpha_{i+1})$ is of type $I_2(m_3)$. - We switch $\alpha_{i}$ and $\alpha_{i+1}'$. By the same lemma and remark, this changes the sign by $(-1)^{m_4/2 - 1}$, where the pseudo-root system $\Phi_4=\Phi\cap({{\mathbb{R}}}\alpha_{i}+{{\mathbb{R}}}\alpha'_{i+1})$ is of type $I_2(m_4)$. The reflections $s_i=s_{\alpha_i-\alpha_i'}$ and $s_{i+1}=s_{\alpha_{i+1}-\alpha_{i+1}'}$ are both in $W$ because $\varphi_i$ contains a multiple of $\alpha_i-\alpha_i'$ and $\varphi_{i+1}$ contains a multiple of $\alpha_{i+1}-\alpha_{i+1}'$. Observe now that $s_i(\Phi_1)=\Phi_3$, $s_i(\Phi_2)=\Phi_4$, $s_{i+1}(\Phi_1)=\Phi_2$ and $s_{i+1}(\Phi_3)=\Phi_4$. Thus the four pseudo-root systems $\Phi_1$, $\Phi_2$, $\Phi_3$ and $\Phi_4$ are isomorphic and hence $m_{1} = m_{2} = m_{3} = m_{4}$. Hence performing operations (1) to (4) changes the sign by $((-1)^{m_1/2 - 1})^{4} = 1$, that is, $\det(w_\theta)=\det(w_{\theta'})$. Let $\varphi \subseteq \Phi$ be a $2$-structure, and let $w_\theta$ be as in Proposition \[prop\_sign\_2\_structure\]. Then the sign $(-1)^{r+r'}\det(w_\theta)$, where $r$ is the number of irreducible factors of $\Phi$ of type $A_{2n}$ with $n$ odd and $r'$ is the number of irreducible factors of $\Phi$ of type $I_2(2n'+1)$ with $n'\geq 3$ odd, is called the sign of $\varphi$ and denoted by $\epsilon(\varphi,\Phi^+)$, or by $\epsilon(\varphi)$ if the system of positive pseudo-roots $\Phi^+$ is understood. \[def\_sign\_2\_structure\] For a root system, this coincides with the definition of the sign of $\varphi$ from Herb’s paper [@Herb-Pl], and it differs from the definition in Section 5 of Herb’s paper [@Herb-DSC]; see Remark 5.1 of [@Herb-DSC] and Corollary \[cor\_comparison\_signs\]. Let $\varphi \subseteq \Phi$ be a $2$-structure, that is, $\varphi\in\mathcal{T}(\Phi)$. - For every $w\in W$, the identity $\epsilon(w(\varphi), w(\Phi^+))=\epsilon(\varphi,\Phi^+)$ holds. - Let $w\in W$ be such that $w(\varphi^+) \subseteq \Phi^+$. Then the identity $\epsilon(w(\varphi),\Phi^+) = \det(w) \cdot \epsilon(\varphi,\Phi^+)$ holds. \[lemma\_sign\_of\_w(varphi)\] Both identities follow easily from the definition of $\epsilon(\varphi,\Phi^+)$. Indeed, let $\theta$ be a subset of $\varphi$ chosen as in Proposition \[prop\_sign\_2\_structure\]. For every $w\in W$, $w(\varphi)$ is a $2$-structure for $\Phi$ and its subset $w(\theta)$ satisfies the same conditions for the system of positive pseudo-roots $w(\Phi^+)$, and also for the system of positive pseudo-roots $\Phi^+$ if $w(\varphi^+) \subset \Phi^+$. Also, we have $\Phi^+_{w\cdot \theta}=w\cdot\Phi^+_\theta$. This immediately yields (i) and (ii). Let $\alpha_0\in\Phi$ be a simple pseudo-root, let $s_0$ be the simple reflection defined by $\alpha_0$, let $\Phi_0 = \alpha_0^\perp\cap\Phi$ and $\Phi_0^+=\Phi_0\cap\Phi^+$. Let $\mathcal{T}''$ be the set of $\varphi\in\mathcal{T}(\Phi)$ such that $s_0(\varphi)=\varphi$; we also consider the subsets $\mathcal{T}''_1=\{\varphi\in\mathcal{T}'':\varphi\cap \Phi_0\in\mathcal{T}(\Phi_0)\}$ and $\mathcal{T}''_2= \mathcal{T}''-\mathcal{T}''_1$. Then the following statements hold: - The map $\mathcal{T}''_1{\longrightarrow}\mathcal{T}(\Phi_0)$, $\varphi\longmapsto\varphi\cap\Phi_0$ is bijective. - For every $\varphi\in\mathcal{T}''_1$, we have $\epsilon(\varphi,\Phi^+)=\epsilon(\varphi\cap\Phi_0,\Phi_0^+)$. - There exists an involution $\iota$ of $\mathcal{T}''_2$ such that, for every $\varphi\in\mathcal{T}''_2$, we have $\varphi\cap\Phi_0=\iota(\varphi)\cap\Phi_0$ and $\epsilon(\iota(\varphi), \Phi^+)=-\epsilon(\varphi,\Phi^+)$. \[lemma\_T\_Phi\_induction\_step\] Note that a $2$-structure $\varphi$ for $\Phi$ is in $\mathcal{T}''$ if and only if $\alpha_0\in\varphi$. Indeed, if $\alpha_0\in\varphi$, then $s_0$ is in the Coxeter group of $\varphi$, so $s_0(\varphi)=\varphi$; conversely, we have $s_{\alpha_0}(\Phi^+-\{\alpha_0\})\subset\Phi^+$ by [@BB Lemma 4.4.3], so, if $\varphi\in\mathcal{T}''$ and $\alpha_0\not\in\varphi$, then $s_0(\varphi^+) \subseteq \Phi^+\cap\varphi= \varphi^+$, contradicting condition (b) in the definition of a $2$-structure. Note also that the subset $\varphi\cap\Phi_0$ of $\Phi_0$ always satisfies condition (a) in the definition of a $2$-structure, but it does not always satisfy condition (b). We prove (1). We may assume that $\Phi$ is irreducible, and we will freely use the explicit description of $2$-structures given in Subsection \[section\_2\_structures\_irreducibles\]. If $2$-structures for $\Phi$ are all of type $A_1^s$ for some $s$, which happens in types $A_n$, $D_n$, $E_6$, $E_7$, $E_8$, $H_3$, $H_4$ and $I_2(m)$ for $m$ odd, then $\varphi\cap\Phi_0\in\mathcal{T}(\Phi_0)$ for every $\varphi\in\mathcal{T}(\Phi)$, that is, $\mathcal{T}_1''=\mathcal{T}''$, and we see in the explicit description of $2$-structures that the map of statement (1) is a bijection. It is easy to check that the same statement holds in type $I_2(m)$ for $m$ even. We now suppose that $\Phi$ is of type $B_n$ or $F_4$. (Recall that from the point of view of Coxeter systems types $B_n$ and $C_n$ are isomorphic.) For convenience, in this case, we take $\Phi$ to be the actual root system, with possibly non-normalized roots; this does not affect any of the definitions that we made before. To study the map of (1), we may assume that $\alpha_0 = e_n$ or $\alpha_0 = e_1-e_2$. Suppose first that $\alpha_0 = e_1-e_2$. Then $\Phi_0$ is reducible. Furthermore, it is of type $A_1\times B_{n-2}$ if $\Phi$ is of type $B_n$, and of type $A_1\times B_2$ if $\Phi$ is of type $F_4$, where the $A_1$ factor is $\{\pm(e_1+e_2)\}$. In both cases, it is easy to see that $\mathcal{T}''_1=\mathcal{T}''$ and that (1) holds. Suppose that $\alpha_0=e_n$. Then $\Phi_0$ is irreducible. Furthermore, it is of type $B_{n-1}$ if $\Phi$ is of type $B_n$, and of type $B_3$ if $\Phi$ is of type $F_4$. If $\Phi$ is of type $F_4$ or $B_n$ with $n$ even then again it is easy to see that $\mathcal{T}''_1=\mathcal{T}''$ and that (1) holds. Finally, suppose that $\Phi$ is of type $B_n$ with $n$ odd and that $\alpha_0=e_n$. If $\varphi\in\mathcal{T}''$ then we have $\varphi\in\mathcal{T}''_1$ if and only if $\{\pm e_n\}$ is an irreducible component of $\varphi$. The map sending $\varphi_0\in\mathcal{T}(\Phi_0)$ to $\varphi_0\sqcup\{\pm e_n\}$ is thus an inverse to the map of (1), so statement (1) holds. We now prove (3). We have seen in the proof of (1) that $\mathcal{T}''_2=\varnothing$ unless $\Phi$ is of type $B_n$ with $n$ odd and $\alpha_0$ is the short simple root. Assume that we are in this case, which means that $\alpha_0=e_n$. Let $\varphi\in\mathcal{T}''_2$. Then there exists $2 \leq i \leq n$ such that $\varphi_1=\{\pm e_n,\pm e_i,\pm e_n\pm e_i\}$ is an irreducible component of $\varphi$. Write $\varphi=\varphi_1\sqcup\varphi_2\sqcup\cdots\sqcup\varphi_r$, where the $\varphi_k$ are irreducible and $\varphi_2=\{\pm e_j\}$ is the unique rank $1$ component of $\varphi$. Set $\iota(\varphi)=\{\pm e_n,\pm e_j,\pm e_n\pm e_j\}\sqcup\{\pm e_i\}\sqcup\varphi_3 \sqcup\cdots\sqcup\varphi_r$. This map switches the roles of $e_i$ and $e_j$. Then $\iota(\varphi)$ is also in $\mathcal{T}''_2$, it is not equal to $\varphi$, we have $\iota(\varphi)\cap\Phi_0=\varphi\cap\Phi_0$ and $\iota(\iota(\varphi))=\varphi$. To finish the proof of (3), it suffices to show that $\epsilon(\iota(\varphi),\Phi^+)= -\epsilon(\varphi,\Phi^+)$ for every $\varphi\in{\mathscr{T}}''_2$. But this follows immediately from the definition of $\iota(\varphi)$ and from Lemma \[lemma\_change\_order\_theta\_sign\]. We finally prove (2). Let $\varphi\in\mathcal{T}''_1$. Choose an ordered subset $\theta=\{\alpha_1,\ldots,\alpha_r\}$ of $\varphi$ as in Proposition \[prop\_sign\_2\_structure\]. We may assume that $\alpha_0\in \theta$. If $\alpha_0$ is in an irreducible component of $\varphi$ of type $A_1$, we may assume that $\alpha_0=\alpha_1$. If $\alpha_0$ is in an irreducible component of $\varphi$ of rank $2$, then, as it is a simple pseudo-root, it cannot be the first element of $\theta$ coming from this rank $2$ factor of $\varphi$ (see Figure \[figure\_choice\_theta\] for an illustration in the case of $I_2(8)$, the general case is similar), so we may assume that $\alpha_0=\alpha_r$. Suppose first that $\alpha_0$ is in an irreducible component of $\varphi$ of rank $2$ and that $\alpha_0=\alpha_r$. By the description of $2$-structures in Subsection \[section\_2\_structures\_irreducibles\], this can only happen if $\Phi$ is of type $B_n$, $F_4$ or $I_2(m)$ with $m$ even. The set $\{\alpha_1, \ldots,\alpha_{r-1}\}$ is an ordered subset of $\varphi_0$ satisfying the conditions of Proposition \[prop\_sign\_2\_structure\], and $\Phi^+_{0,\theta_0}=\Phi^+_\theta\cap\Phi_0$. So the statement of (2) will follow if we can show that $X=(\Phi^+_{\theta}-\Phi^+_{0,\theta_0}) \cap \Phi^-$ has even cardinality. Let $s=s_{\alpha_r}$. We claim that $s(X)=X$ and that $s$ has no fixed points in $X$, which implies that $X$ has even cardinality because $s^2=1$. The fact that $s$ has no fixed point in $X$ follows from the facts that the fixed points of $s$ are the elements of $\alpha_r^\perp$, that $\Phi\cap\alpha_r^\perp=\Phi_0$ and that $X\cap\Phi_0= \varnothing$. As $\alpha_r\not\in\Phi^-$ and $-\alpha_r\not\in\Phi^+_\theta$, we have $X=(\Phi^--\{-\alpha_r\})\cap(\Phi^+_\theta-(\Phi^+_{0,\theta_0} \cup\{\alpha_r\}))$. As $\alpha_r$ is a simple pseudo-root, we have $s(\Phi^--\{-\alpha_r\})\subset\Phi^--\{-\alpha_r\}$ by [@BB Lemma 4.4.3]. So it suffices to prove that $s$ preserves $\Phi^+_\theta-(\Phi^+_{0,\theta_0}\cup\{\alpha_r\})$. If $\beta\in\Phi^+_\theta-\Phi^+_{0,\theta_0}$ is such that $\beta\not=\alpha_r$, then we cannot have $(\beta,\alpha_i)=0$ for every $i\in\{1,\ldots,r-1\}$; indeed, as $\Phi$ is of type $B_n$, $F_4$ or $I_2(m)$ with $m$ even, the family $(\alpha_1,\ldots,\alpha_r)$ is an orthonormal basis of $V$, so the only element of $\Phi^+_\theta$ that is orthogonal to $\alpha_1,\ldots,\alpha_{r-1}$ is $\alpha_r$. So $\Phi^+_\theta-(\Phi^+_{0,\theta_0}\cup\{\alpha_r\})$ is the set pseudo-roots $\beta\in\Phi$ such that $((\beta,\alpha_1),\ldots, (\beta,\alpha_{r-1}))>0$ (for the lexicographic order on ${{\mathbb{R}}}^{r-1}$) and that $(\beta,\alpha_r)\not=0$. This set is stable by $s$, because, for every $\beta\in V$, we have $(s(\beta),\alpha_i)=(\beta,s(\alpha_i))=(\beta,\alpha_i)$ if $1\leq i\leq r-1$ and $(s(\beta),\alpha_r)=(\beta,s(\alpha_r))= -(\beta,\alpha_r)$. Now we suppose that $\alpha_0$ is in an irreducible component of $\varphi$ of rank $1$ and that $\alpha_0=\alpha_1$. Then $\theta_0=\{\alpha_2, \ldots,\alpha_r\}$ is an ordered subset of $\varphi_0$ satisfying the conditions of Proposition \[prop\_sign\_2\_structure\], and $\Phi^+_{0,\theta_0}=\Phi_0\cap\Phi^+_\theta$, so $\Phi^+_{\theta}-\Phi^+_{0,\theta_0}=\{\beta\in\Phi : (\beta,\alpha_1)>0\}$. Statement (2) will follow if we can show that $$X=(\Phi^+_{\theta}-\Phi^+_{0,\theta_0}) \cap \Phi^-=\{\beta\in\Phi^-: (\beta,\alpha_1)>0\}$$ has even cardinality if $\Phi$ is not of type $A_{2n}$ or $I_2(2n'+1)$ with $n'$ odd, and odd cardinality otherwise. As $\varphi$ has an irreducible component of rank $1$, we cannot be in type $F_4$. We can check that $X$ has even cardinality by a computer calculation in the exceptional types $E$, $G$ and $H$. We now go through the remaining types one by one (in cases $A$, $B$ and $D$, we use the description of the roots from the tables at the end of [@Bourbaki], and not the normalized pseudo-root system): - : If $m$ is even, then $\varphi$ is a rank $2$ pseudo-root system; so $m$ must be odd, and then $\varphi_0$ is empty and $\epsilon(\varphi_0)=1$. There are exactly $m$ pseudo-roots $\beta$ such that $(\beta,\alpha_1)>0$, and $(m-1)/2$ of these are in $\Phi^-$. So $\epsilon(\varphi)=(-1)^{(m-1)/2}(-1)^{(m-1)/2}=1$, which is what we wanted. - : We write $\alpha_0=e_i-e_{i+1}$, with $1\leq i\leq n$. Then $$X=\{e_j-e_k: 1\leq k<j=i\mbox{ or }i+1=k<j\leq n+1\}$$ has cardinality $n-1$, that is, even if and only if $n$ is odd. - : As $\varphi$ has an irreducible component of rank $1$, the integer $n$ must be odd and $\alpha_0$ is the short simple root, that is, $\alpha_0=e_n$. Then $$X=\{-e_i+e_n:1\leq i\leq n\}$$ has cardinality $n-1$, which is even. - : If $\alpha_0=e_i-e_{i+1}$ with $1\leq i\leq n-1$, then $$\begin{aligned} X = & \{e_j-e_k: 1\leq k<j=i\mbox{ or }i+1=k<j\leq n\} \\ & \cup \{-(e_j+e_k):i+1=j<k\leq n\mbox{ or }i\not=j<k=i+1\}\end{aligned}$$ has cardinality $2n-4$. If $\alpha_0=e_{n-1}+e_n$, then $$X=\{e_j-e_k:n=j>k\not=n-1\mbox{ or }j=n-1>k\}$$ also has cardinality $2n-4$. Let $\varphi \subseteq \Phi$ be a $2$-structure. Then $\|\Phi^+-\varphi^+\|$ is an even integer. More precisely, if $\Phi$ is irreducible, we have $$\|\Phi^+-\varphi^+\|= \begin{cases} 2n\bmod 4 & \text{ if }\Phi\text{ is of type } A_{2n}, \\ 0\bmod 4 & \text{ if }\Phi\text{ is of type } A_{2n+1}, B, D, E, F_4, G_2, \text{or } H, \\ 2^r(m-1) & \text{ if } \Phi \text{ is of type } I_2(2^r m) \text{ with } m \text{ odd}. \end{cases}$$ \[lemma\_r\_Phi\] This follows from the explicit description of $2$-structures for the irreducible types in Subsection \[section\_2\_structures\_irreducibles\]. Orthogonal sets of pseudo-roots and 2-structures {#subsection_B_3} ------------------------------------------------ For a pseudo-root system $\Phi$ let ${\mathscr{O}}(\Phi)$ be the set of all finite sequences $(\alpha_1,\alpha_2,\ldots,\alpha_r)$ of elements of $\Phi$ which are pairwise orthogonal and such that their entries all have the same length, that is, the following two conditions hold: - $(\alpha_i,\alpha_j)=0$ for all $1 \leq i < j \leq r$; - $\\|\alpha_1\\|=\\|\alpha_2\\|=\cdots=\\|\alpha_r\\|$. Recall that in the root systems of types $A_n$, $D_n$, $E_6$, $E_7$ and $E_8$, all the roots have the same length. Hence in part (ii) of the following lemma there is no need for an extra condition in these types. Suppose that $\Phi$ is a root system. - Let $\theta=(\alpha_1,\ldots,\alpha_r)$ and $\theta'=(\beta_1,\ldots,\beta_s)$ be elements of ${\mathscr{O}}(\Phi)$, and suppose that $\theta^\perp\cap\Phi=(\theta')^\perp\cap\Phi=\varnothing$ and that the elements of $\theta$ and $\theta'$ have the same length. Then there exists $w\in W$ such that $\{\alpha_1,\ldots,\alpha_r\}= \{w(\beta_1),\ldots,w(\beta_s)\}$. In particular, $r=s$ holds. - Suppose that $\Phi$ is irreducible and not of type $G_2$. Let $\theta\in{\mathscr{O}}(\Phi)$. If $\Phi$ is of type $B_{n}$ and $\theta$ consists of long roots of the root system then assume that $n$ is even. Similarly, if $\Phi$ is of type $C_{n}$ and $\theta$ consists of short roots then again assume that $n$ is even. Then there exists $\theta'\in{\mathscr{O}}(\Phi)$ such that $\theta$ is an initial segment of $\theta'$ and $(\theta')^\perp\cap\Phi=\varnothing$. \[lemma\_one\_structures\] We start by proving (i). Let $\Phi^+_\theta$, respectively $\Phi^+_{\theta'}$, be the system of positive roots defined by $\theta$, respectively $\theta'$, as in Definition \[def\_Phi\_theta\]. As $W$ acts transitively on the set of systems of positive roots, there exists $w\in W$ such that $w(\Phi^+_{\theta'})=\Phi^+_{\theta}$. As $w(\Phi^+_{\theta'})= \Phi^+_{w(\beta_1),\ldots,w(\beta_s)}$, we may assume that $\Phi^+_\theta=\Phi^+_{\theta'}$. We then wish to prove that $\theta$ and $\theta'$ are equal up to reordering their entries. We proceed by induction on the length of $\theta$. If $\theta$ is empty then $\Phi$ is also empty because of the condition $\Phi\cap\theta^\perp=\varnothing$, so $\theta'$ is empty and we are done. Suppose that $r\geq 1$. Let $j_0$ be the smallest index $j$ such that $(\alpha_1,\beta_j)\not=0$. Since $\Phi\cap(\theta')^\perp=\varnothing$, this minimum exists. As $\beta_{j_0}\in\Phi^+_{\theta'}= \Phi^+_\theta$, we cannot have $(\alpha_1,\beta_{j_0})<0$, so $(\alpha_1,\beta_{j_0})>0$. As $\Phi$ is a root system and not just a pseudo-root system, the corollary after [@Bourbaki Chapitre VI, § 1, 3, Théorème 1] implies that the difference $\gamma=\alpha_1-\beta_{j_0}$ is an element of $\Phi\cup\{0\}$. Suppose that $\gamma\in\Phi^+_{\theta'}$. As $(\gamma,\beta_j)=0$ for $1\leq j<j_0$, we must then have $0 \leq (\gamma,\beta_{j_0}) = (\alpha_1,\beta_{j_0}) - (\beta_{j_0},\beta_{j_0})$. The hypothesis states that $\\|\alpha_1\\|=\\|\beta_{j_0}\\|$ and hence we deduce that $(\alpha_1,\beta_{j_0}) \geq \\|\beta_{j_0}\\|^2 = \\|\alpha_1\\| \cdot \\|\beta_{j_0}\\|$. This inequality implies that $\alpha_1=\beta_{j_0}$, contradicting the fact that $\gamma$ is nonzero. Suppose that $\gamma\in \Phi^-_\theta$. Then $0 \leq (\alpha_1,-\gamma) =(\alpha_1,\beta_{j_0}) - (\alpha_1,\alpha_1)$, so $(\alpha_1,\beta_{j_0})\geq\\|\alpha_1\\|^2$, and again this implies that $\alpha_1=\beta_{j_0}$ and contradicts the assumption. Hence we conclude that $\gamma=0$, that is, $\alpha_1=\beta_{j_0}$. Let $\Phi_0=\alpha_1^\perp\cap \Phi=\beta_{j_0}^\perp\cap\Phi$, $\theta_0=(\alpha_2,\ldots,\alpha_r)$ and $\theta'_0=(\beta_1,\ldots,\widehat{\beta_{j_0}},\ldots,\beta_s)$. Then $\Phi_0$ is a root system, $\theta_0$ and $\theta_0'$ are in ${\mathscr{O}}(\Phi_0)$, $\theta_0^\perp\cap\Phi_0=(\theta_0')^\perp\cap\Phi_0= \emptyset$, and $\Phi^+_{0,\theta_0}=\Phi^+_\theta\cap\Phi_0= \Phi^+_{\theta'}\cap\Phi_0=\Phi^+_{0,\theta'_0}$. We can apply the induction hypothesis to conclude that $\{\alpha_2,\ldots,\alpha_r\}= \{\beta_1,\ldots,\widehat{\beta_{j_0}},\ldots,\beta_s\}$, and this immediately implies that $\{\alpha_1,\ldots,\alpha_r\}= \{\beta_1,\ldots,\beta_s\}$. We use the classification of irreducible root systems to prove (ii). - Let $\Phi=A_n$. Then $\theta=(e_{i_1}-e_{j_1},\ldots,e_{i_s}-e_{j_s})$ with $i_1,j_1,\ldots,i_s,j_s\in\{1,\ldots,n+1\}$ all distinct. After applying an element of $W = \mathfrak{S}_{n+1}$ to $\theta$, we may assume that $(i_1,j_1,\ldots,i_s,j_s)=(1,2,\ldots,2s)$, and then we take $\theta'=(e_{1}-e_{2}, \ldots, e_{2m-1}-e_{2m})$, where $m=\lfloor\frac{n+1}{2}\rfloor$. - $\Phi=B_n$. Suppose that the elements of $\theta$ have the same length as the short roots in $\Phi$. Then $\theta=(\pm e_{i_1},\ldots,\pm e_{i_s})$, so after applying an element of $W$, we may assume that $\theta=(e_1,\ldots,e_s)$, and we can take $\theta' = (e_1,\ldots,e_n)$. Suppose that $n$ is even and that the elements $\theta$ have the same length as the long roots in $\Phi$. Then the elements of $\theta$ are of the form $\pm e_i\pm e_j$, with the following condition. If $\pm(e_i+e_j)$, respectively $\pm(e_i-e_j)$, appears, then $\mp(e_i+e_j)$, respectively $\mp(e_i-e_j)$, cannot appear and neither can any $\pm e_k\pm e_l$ with $\|\{i,j\}\cap\{k,l\}\| = 1$. After applying an element of $W$, we may assume that $\theta$ up to some reordering is given by the concatenation $(e_1+e_2,e_1-e_2,\ldots,e_{2s-1}+e_{2s},e_{2s-1}-e_{2s}) \circ (e_{2s+1}+e_{2s+2},\ldots,e_{2(s+t)-1}+e_{2(s+t)})$. Extend $\theta$ to a reordering $\theta'$ of $(e_1+e_2,e_1-e_2,\ldots,e_{2m-1}+e_{2m},e_{2m-1}-e_{2m})$, where $m=n/2$. - The cases of $C_n$ and $D_n$ are similar, except for some parity issues in the case of $D_n$. - The cases of $F_4$, $E_6$, $E_7$ and $E_8$ can be checked with a computer. (Data from the calculations can be found in Table \[table\_the\_number\_of\_orthogonal\_sets\].) $$\begin{array}{r \| r r r r r r r} k & E_{6} & E_{7} & E_{8} & F_{4} & F_{4} & H_{3} & H_{4} \\ & & & & \text{\small short} & \text{\small long} & & \\ \hline 1 & 72 & 126 & 240 & 24 & 24 & 30 & 120 \\ 2 & 1080 & 3780 & 15120 & 72 & 72 & 60 & 1800 \\ 3 & 4320 & 32760 & 302400 & 96 & 96 & 40 & 2400 \\ 4 & 2160 & 75600 & 1965600 & 48 & 48 & & 1200 \\ 5 & & 90720 & 3628800 & & & & \\ 6 & & 60480 & 3628800 & & & & \\ 7 & & 17280 & 2073600 & & & & \\ 8 & & & 518400 & & & & \\ \end{array}$$ \[table\_the\_number\_of\_orthogonal\_sets\] Let $\Phi$ be a normalized pseudo-root system, let $\theta=(\alpha_1, \ldots,\alpha_r)$ be a sequence of pairwise orthogonal elements of $\Phi$ such that $\theta^\perp\cap\Phi=\varnothing$, and let $\theta'$ be the sequence obtained from $\theta$ by exchanging $\alpha_i$ and $\alpha_{i+1}$. Consider the subroot system $\Phi'=\Phi\cap({{\mathbb{R}}}\alpha_i+{{\mathbb{R}}}\alpha_{i+1})$. Then $\Phi'$ is of type $A_1\times A_1$ or $I_2(m)$ with $m\geq 4$ even, and the parity of the cardinality of $\Phi^+_\theta \cap \Phi^-_{\theta'}$ is given by $$\begin{array}{rll} \|\Phi^+_\theta \cap \Phi^-_{\theta'}\| & \equiv 0 \ \bmod 2 & \text{ if }\Phi'=A_1\times A_1, \\ \|\Phi^+_\theta \cap \Phi^-_{\theta'}\| & \equiv m/2-1 \ \bmod 2 & \text{ if }\Phi'=I_2(m). \end{array}$$ \[lemma\_change\_order\_theta\_sign\] As $\Phi'$ is a pseudo-root system of rank $2$ (because it is contained in a $2$-dimensional vector space and contains the two linearly independent pseudo-roots $\alpha_i$ and $\alpha_{i+1}$), it is of type $A_1 \times A_1$ or $I_2(m)$ with $m\geq 3$. Moreover, $\Phi'$ contains two orthogonal pseudo-roots, so it cannot be of type $I_2(m)$ with $m$ odd. We now set $C=\Phi^+_\theta \cap \Phi^-_{\theta'}$ and calculate the parity of $\|C\|$. Let $\gamma\in C$. Then $\gamma$ is orthogonal to $\alpha_1,\ldots,\alpha_{i-1}$, so we can write $\gamma=c\alpha_i+d\alpha_{i+1}+\lambda$ with $\lambda\in\operatorname{Span}( \alpha_1,\ldots,\alpha_{i+1})^\perp$ and $c\alpha_i+d\alpha_{i+1}\not=0$. Set $\iota(\gamma)=- s_{\alpha_i}s_{\alpha_{i+1}}(\gamma)$. Then $\iota(\gamma)\in\Phi$ and $\iota(\gamma)=c\alpha_i+d \alpha_{i+1}-\lambda$, so $\iota(\gamma)\in C$. Also, we clearly have $\iota(\iota(\gamma))=\gamma$, and $\iota(\gamma)$ is equal to $\gamma$ if and only if $\lambda=0$, that is, if and only if $\gamma\in\Phi'$. We have defined an involution $\iota$ of $C$, and we conclude that $\|C\| \equiv \|C_0\| \bmod 2$, where $C_0 = \Phi'\cap C$ is the set of fixed points of $\iota$ in $C$. If $\Phi'$ is of type $A_1\times A_1$ then we easily see that $C_0$ is empty, so we are done. Suppose that $\Phi'$ is of type $I_2(m)$ with $m$ even. Let $\gamma=c\alpha_i+d\alpha_{i+1}\in\Phi'$, with $c,d\in{{\mathbb{R}}}$. Then $\gamma\in C$ if and only if $c>0$ and $d<0$. The set $C_0$ contains exactly one quarter of the elements of $\Phi'-\{\pm\alpha_i,\pm\alpha_{i+1}\}$, that is, $\|C_0\| = (2m-4)/4 = m/2-1$. If we view the root system $A_1 \times A_1$ as the dihedral pseudo-root system $I_2(2)$ then the conclusion of Lemma \[lemma\_change\_order\_theta\_sign\] is that $\|\Phi^+_\theta \cap \Phi^-_{\theta'}\| \equiv m/2-1 \bmod 2$ if $\Phi'=I_2(m)$ with $m$ even and $m \geq 2$. \[remark\_change\_order\_theta\_sign\] Suppose that $\Phi$ is an irreducible root system (not just a pseudo-root system) and not of type $G_2$. Let $\Phi^+$ be a system of positive roots of $\Phi$ and let $\varphi \subseteq \Phi$ be a $2$-structure. Define a subset $\theta$ of $\varphi$ as in Proposition \[prop\_sign\_2\_structure\]. Then there is a choice of the sequences $\theta_i$ for which $\theta$ is an element of ${\mathscr{O}}(\Phi)$. Moreover, if $\Phi$ is of type $B_n$ or $F_4$ we can choose $\theta$ to consist of short roots. Similarly, if $\Phi$ is of type $C_n$ or $F_4$ we can choose $\theta$ to consist of long roots. \[lemma\_choice\_of\_theta\] By Remark \[rmk\_no\_orthogonal\_root\] we have $\theta^\perp\cap\Phi=\varnothing$. We use the notation of Proposition \[prop\_sign\_2\_structure\]. If all the roots of $\Phi$ have the same length (which is the case for $A_n$, $D_n$, $E_6$, $E_7$ and $E_8$), then there is nothing to prove. Note also that if $\varphi_i$ is an actual root system of type $B_2$ (that is, with the correct root lengths), then the two possible choices for $\theta_i$ are the set of short positive roots and the set of long positive roots. Suppose that $\Phi$ is of type $B_n$. If $\varphi$ has no irreducible component of type $A_1$, then we choose the two short positive roots in each $\varphi_i$. Suppose that $\varphi$ has a factor of type $A_1$. We show that this factor cannot contain long roots. Suppose on the contrary that this occurs. Without loss of generality, we may assume that $\varphi_1=\{\pm (e_1+e_2)\}$. The rank $2$ factors of $\varphi$ cannot contain $e_1-e_2$, so they are all in $e_1^\perp\cap e_2^\perp$. All the rank $1$ factors that do not contain $e_1-e_2$ must also be in $e_1^\perp\cap e_2^\perp$. If $e_1-e_2$ were not in $\varphi$ then the reflection $s_{e_1-e_2}$ would act as the identity on all the elements on $\varphi$, which contradicts the definition of a $2$-structure. Hence $\{\pm(e_1-e_2)\}$ is another rank $1$ factor of $\varphi$. But then the reflection $s_{e_1}$ preserves $\varphi^+$, which is impossible. Hence all the $A_1$ factors of $\varphi$ contain only short roots, and we choose the $\theta_i$ in the $B_2$ factors to contain the two short positive roots. The case of $C_n$ is similar, with the roles of short and long roots uniformly exchanged. Finally suppose that $\Phi=F_4$. In this case we can similarly show that the $2$-structure $\varphi$ has type $B_{2}^{2}$, allowing us to pick either short or long roots in each factor. 2-structures in the irreducible types {#section_2_structures_irreducibles} ------------------------------------- In this subsection we prove Proposition \[prop\_W\_acts\_transitively\_on\_T\_Phi\], that is, the fact that the group $W$ acts transitively on the collection of $2$-structures $\mathcal{T}(\Phi)$. It is enough to prove this result for irreducible pseudo-root systems. We proceed by a case by case analysis. ### Types $A_n$, $D_n$, $E_6$, $E_7$ and $E_8$ {#types-a_n-d_n-e_6-e_7-and-e_8 .unnumbered} Suppose that $\Phi$ is a root system of type $A_n$, $D_n$ or $E_m$ with $m\in\{6,7,8\}$. As all the roots of $\Phi$ have the same length and as $\Phi$ contains no $B_2$ root system, the $2$-structures for $\Phi$ are exactly the maximal sets $\varphi=\{\pm\alpha_1,\ldots,\pm\alpha_r\}$ such that $(\alpha_1,\ldots,\alpha_r)\in{\mathscr{O}}(\Phi)$. By Lemma \[lemma\_one\_structures\](i), for any $(\alpha_1,\ldots,\alpha_r)$ and $(\beta_1,\ldots,\beta_s)$ on ${\mathscr{O}}(\Phi)$, there exists $w\in W$ such that $\{\alpha_1,\ldots,\alpha_r\}=\{w(\beta_1),\ldots,w(\beta_s)\}$. Hence the group $W$ acts transitively on $\mathcal{T}(\Phi)$. In particular, all the $2$-structures for $\Phi$ are isomorphic, so we can determine their type. See Table \[table\_2-structure\_in\_A\_D\_E\]. $$\begin{array}{c \| l l} \text{Type of root} & & \text{Type of} \\ \text{system } \Phi & \text{$2$-structures are isomorphic to} & \text{$2$-structure} \\ \hline A_n & \{\pm(e_1-e_2), \:\:\: \pm(e_3-e_4), \:\:\: \ldots, \:\:\: \pm(e_{2m-1}-e_{2m})\} & A_1^m \\ D_n & \{\pm e_1 \pm e_2, \:\:\: \pm e_3 \pm e_4, \:\:\: \ldots, \:\:\: \pm e_{2m-1} \pm e_{2m}\} & A_1^{2m} \\ E_6 & \{\pm e_1 \pm e_2, \:\:\: \pm e_3 \pm e_4\} & A_1^{4} \\ E_7 & \{\pm e_1 \pm e_2, \:\:\: \pm e_3 \pm e_4, \:\:\: \pm e_5 \pm e_6, \:\:\: \pm(e_7-e_8)\} & A_1^{7} \\ E_8 & \{\pm e_1 \pm e_2, \:\:\: \pm e_3 \pm e_4, \:\:\: \pm e_5 \pm e_6, \:\:\: \pm e_7 \pm e_8\} & A_1^{8} \end{array}$$ \[table\_2-structure\_in\_A\_D\_E\] ### Types $B_n$ and $C_n$ {#types-b_n-and-c_n .unnumbered} Suppose that $\Phi$ is a root system of type $B_n$. This will also give the type $C_n$ case, since $B_n$ and $C_n$ correspond to the same Coxeter system. We claim that $W$ acts transitively on $\mathcal{T}(\Phi)$. In particular, all the $2$-structures for $\Phi$ are isomorphic to $$\varphi_0=\begin{cases} \{\pm e_1, \pm e_2, \pm e_1\pm e_2\} \sqcup \cdots \sqcup \{\pm e_{2m-1},\pm e_{2m},\pm e_{2m-1}\pm e_{2m}\} \!\!\!\!& \text{if } n=2m, \\ \{\pm e_1, \pm e_2, \pm e_1\pm e_2\} \sqcup \cdots \sqcup \{\pm e_{2m-1},\pm e_{2m},\pm e_{2m-1}\pm e_{2m}\} \sqcup \{\pm e_{2m+1}\} \!\!\!\!& \text{if } n=2m+1, \end{cases}$$ so they are of type $B_2^{m}$ if $n=2m$ is even, and of type $B_2^{m}\times A_1$ if $n=2m+1$ is odd. We prove the claim by induction on $n$. The case $n=1$ is clear. Suppose that $n\geq 2$. Let $\varphi,\varphi'\in\Phi$. By Lemma \[lemma\_choice\_of\_theta\], we can choose sequences $\theta$ of $\varphi$ and $\theta'$ of $\varphi'$ as in Proposition \[prop\_sign\_2\_structure\] such that $\theta,\theta'\in{\mathscr{O}}(\Phi)$ and that these subsets contain only short roots. By Lemma \[lemma\_one\_structures\](i), we may assume that $\theta$ and $\theta'$ coincide up to the order of their elements. Denote by $\varphi=\varphi_1\sqcup\cdots\sqcup\varphi_s$ and $\varphi'=\varphi'_1\sqcup\cdots\sqcup\varphi'_t$ the decomposition into irreducible systems that gave rise to $\theta$ and $\theta'$. We can always change the order on the $\varphi_i$ and the $\varphi'_j$. Suppose that $\varphi_1$ is of rank $1$, so that $\varphi_1=\{\pm\alpha_1\}$. We may assume that $\alpha_1\in\varphi'_1$. If $\varphi'_1$ is of rank $1$ then $\varphi'_1=\varphi_1$. As $\Phi\cap\varphi_1^\perp$ is an irreducible root system of type $B_{n-1}$, the conclusion follows by the induction hypothesis. If $\varphi'_1$ is of rank $2$ then $\varphi_1'$ is a $B_2$ root system whose short positive roots are $\alpha_1$ and some $\alpha_2$, and we may assume that $\alpha_2\in\varphi_2$. In particular, $\beta=\alpha_1-\alpha_2\in\Phi$. If $\varphi_2=\{\pm\alpha_2\}$ then the reflection $s_\beta$ preserves $\varphi^+$, which is not possible. So $\varphi_2$ is of rank $2$ (in particular, $n\geq 3$), which means that it is a $B_2$ root system whose short roots are $\alpha_2$ and some $\alpha_3$. We may assume that $\alpha_3\in\varphi'_2$. In particular, $\alpha_2-\alpha_3\in\Phi$, so $\gamma=s_\beta(\alpha_2-\alpha_3)=\alpha_1-\alpha_3$ is also a root. The irreducible components of $s_\gamma(\varphi)$ are $\varphi'_1,\{\pm\alpha_3\},\varphi_3,\ldots,\varphi_s$. As $\Phi\cap(\varphi_1')^\perp$ is a root system of type $B_{n-2}$, the induction hypothesis implies that there is a $w\in W$ such that $w(\varphi')=s_\gamma(\varphi)$, which finishes the proof in this case. Suppose that $\varphi_1$ is of rank $2$, and call its other short positive root $\alpha_2$. We may assume that $\alpha_1\in\varphi'_1$. If $\varphi'_1$ is of rank $1$ then $\varphi'_1=\{\pm\alpha_1\}$, and we can repeat the reasoning of the previous paragraph with the roles of $\varphi$ and $\varphi'$ exchanged. If $\varphi'_1=\varphi_1$ then the conclusion follows from the induction hypothesis applied to the $B_{n-2}$ root system $\varphi_1^\perp\cap\Phi$. Finally, suppose that $\varphi'_1$ is of rank $2$ and $\varphi'_1\not=\varphi_1$. Let $\alpha_3$ be the other short positive root of $\varphi'_1$. As $\alpha_2$ and $\alpha_3$ are both short roots, $\beta=\alpha_2-\alpha_3\in\Phi$. Note that the irreducible components of $s_\beta(\varphi)$ are $\varphi'_1,s_\beta(\varphi_2),\ldots,s_\beta(\varphi_s)$, so again the induction hypothesis implies that there exists $w\in W$ such that $s_\beta(\varphi)=w(\varphi')$, and we are done. ### Type $F_4$ {#typef_4 .unnumbered} Suppose that $\Phi$ is a root system of type $F_4$. Then we can show that $W$ acts transitively on $\mathcal{T}(\Phi)$ exactly as in type $B_n$. In particular, any $2$-structure is isomorphic to $\varphi_0 = \{\pm e_1,\pm e_2,\pm e_1\pm e_2\} \sqcup \{\pm e_3,\pm e_4, \pm e_{3}\pm e_{4}\}$, so it is of type $B_2^2$. ### Dihedral types {#dihedral-types .unnumbered} Suppose that $\Phi$ is a pseudo-root system of type $I_2(m)$ with $m\geq 5$ (this includes the type $G_2$ root system). It is straightforward to see that $W$ acts transitively on $\mathcal{T}(\Phi)$. If $m$ is odd then all the $2$-structures for $\Phi$ are isomorphic to $\varphi_0=\{\pm e_1\}$, and in particular of type $A_1$. If $m$ is even then all the $2$-structures for $\Phi$ are of type $I_2(2^r)$, where $2^r$ is the largest power of $2$ dividing $m$. ### Types $H_3$ and $H_4$ {#types-h_3-and-h_4 .unnumbered} Suppose that $\Phi$ is of type $H_3$ or $H_4$. We use the description of the pseudo-root systems $H_3$ and $H_4$ given in [@GB Table 5.2] where they are called $I_3$ and $I_4$. In particular, we choose $\Phi$ to be normalized. We claim that $W$ acts transitively on $\mathcal{T}(\Phi)$, and so every $2$-structure for $\Phi$ is isomorphic to $$\begin{aligned} \varphi_0=\begin{cases} \{\pm e_1,\pm e_2,\pm e_3\} & \text{ if } \Phi=H_3,\\ \{\pm e_1,\pm e_2,\pm e_3,\pm e_4\} & \text{ if }\Phi=H_4, \end{cases} \label{equation_type_H_2-structure}\end{aligned}$$ and in particular it is of type $A_1^3$ if $\Phi=H_3$ and of type $A_1^4$ if $\Phi=H_4$. It is clear by the chosen description of $\Phi$ that all of the inner products of elements of $\Phi$ are in ${{\mathbb{Q}}}[\sqrt{5}]$, and in particular $1/\sqrt{2}$ never appears. So there are no pseudo-roots in $\Phi$ with an angle of ${\pi}/{4}$ between them, which implies that $\Phi$ does not contain any pseudo-root system of type $I_2(m)$ with $m$ a multiple of $4$, and so $2$-structures for $\Phi$ (if they exist) can only have irreducible components of type $A_1$. We check easily that the set $\varphi_0$ given in equation is a $2$-structure, so it remains to show that all the maximal sets of pairwise orthogonal pseudo-roots are conjugate under $W$ to $\zeta_0$, where $\zeta_0=\{e_1,e_2,e_3\}$ if $\Phi=H_3$ and $\zeta_0=\{e_1,e_2,e_3,e_4\}$ if $\Phi=H_4$. Any element of the stabilizer $W_0$ of $\zeta_0$ in $W$ must act on $\operatorname{Span}(\Phi)$ by a permutation of the coordinates, and it must be an even permutation to be in $W$. This implies that the cardinality of $W_0$ is $3$ for $\Phi=H_3$ and $12$ for $\Phi=H_4$. Using a computer, it is not hard to count all the maximal sets of pairwise orthogonal pseudo-roots in $H_3$ and $H_4$ (see Table \[table\_the\_number\_of\_orthogonal\_sets\]). We find that there are $40$ such sets for $H_3$ and $1200$ such sets for $H_4$. In both cases, this number is equal to $\|W\|/\|W_0\|$, so $W$ does act transitively on the set of maximal sets of pairwise orthogonal pseudo-roots, and hence also on ${\mathscr{T}}(\Phi)$. Acknowledgements {#acknowledgements .unnumbered} ================ This work was partially supported by the LABEX MILYON (ANR-10-LABX-0070) of Université de Lyon, within the program “Investissements d’Avenir” (ANR-11-IDEX-0007) operated by the French National Research Agency (ANR). More precisely, the authors would like to thank the École Normale Supérieure de Lyon (ÉNS de Lyon) for its hospitality and support to the second author during the academic year 2017–2018, and to the first and third author during one week visits. The first and third authors also thank Princeton University for hosting four one-week visits during the academic year 2018–2019, and the Institute for Advanced Study for hosting a research visit Summer 2019, and the second author thanks the University of Kentucky for its hospitality during a one-week visit in the Fall of 2019. This work was also partially supported by grants from the Simons Foundation (\#429370 to Richard Ehrenborg and \#422467 to Margaret Readdy). Finally, we used SageMath and Maple for innumerable root system computations. [University of Kentucky, Department of Mathematics, Lexington, KY 40506.]{} . [Princeton University, Department of Mathematics, Princeton, NJ 08540.]{} . [University of Kentucky, Department of Mathematics, Lexington, KY 40506.]{} . [^1]: This is a reformulation of [@Morel Proposition A.4]. [^2]: In this appendix, we take all faces to be closed faces, unlike in the rest of the article. [^3]: By definition, this is the minimal number of factors in an expression of $w$ as a product of reflections corresponding to simple pseudo-roots.	{ "pile_set_name": "ArXiv" }
--- abstract: \| Path planning plays an essential role in many areas of robotics. Various planning techniques have been presented, either focusing on learning a specific task from demonstrations or retrieving trajectories by optimizing for hand-crafted cost functions which are well defined a priori. In this work, we present an incremental adversarial learning-based framework that allows inferring implicit behaviour, i.e. the natural characteristic of a set of given trajectories. To achieve adversarial learning, a zero-sum game is constructed between a planning algorithm and an adversary - the discriminator. We employ the discriminator within an optimal motion planning algorithm, such that costs can be learned and optimized iteratively, improving the integration of implicit behavior. By combining a cost-based planning approach with trained intrinsic behaviour, this can be be integrated also with other constraints such as obstacles or general cost factors within a single planning framework. We demonstrate the proposed method on a dataset for collision avoidance, as well as for the generation of human-like trajectories from motion capture data. Our results show that incremental adversarial learning is able to generate paths that reflect the natural implicit behaviour of a dataset, with the ability to improve on performance using iterative learning and generation. author: - 'Salvatore Virga$^{1}$, Christian Rupprecht$^{1, }$, Nassir Navab$^{1,2}$ and Christoph Hennersperger$^{1,3}$[^1][^2] [^3] [^4]' bibliography: - 'IEEEabrv.bib' - 'references.bib' title: 'Incremental Adversarial Learning for Optimal Path Planning* ' --- ACKNOWLEDGMENT {#acknowledgment .unnumbered} ============== This work partially received funding by the project IOTMA supported by the central innovation program for SMEs (ZIM), the BayMED project 5G-MedServices funded by the Bavarian state, as well as from the European Union’s Horizon 2020 research and innovation program EDEN2020 under Grant Agreement No. 688279. The motion capture data used in this project was obtained from mocap.cs.cmu.edu which was created with funding from NSF EIA-0196217. We gratefully acknowledge the support of NVIDIA Corporation with the donation of the Titan X Pascal GPU used here. [^1]: $^{1}$Salvatore Virga, Christoph Hennersperger, Christian Ruprecht and Nassir Navab are with Computer Aided Medical Procedures, Technical University Munich, Boltzmannstr. 3, 85748 Garching bei Munich, Germany [[email protected]]{} [^2]: $^{2}$Nassir Navab is with Johns Hopkins University, Baltmore, USA [^3]: $^{3}$Christoph Hennersperger is with Trinity College Dublin, College Green, Dublin 2, Ireland [^4]: $^{*}$Christian Rupprecht is now at the University of Oxford, United Kingdom.	{ "pile_set_name": "ArXiv" }
--- abstract: 'In this paper, we are interested in symbiotic radio networks, in which an Internet-of-Things (IoT) network parasitizes in a primary network to achieve spectrum-, energy-, and infrastructure-efficient communications. Specifically, the BS serves multiple cellular users using time division multiple access (TDMA) and each IoT device is associated with one cellular user for information transmission. We focus on the user association problem, whose objective is to link each IoT device to an appropriate cellular user by maximizing the sum rate of all IoT devices. However, the difficulty in obtaining the full real-time channel information makes it difficult to design an optimal policy for this problem. To overcome this issue, we propose two deep reinforcement learning (DRL) algorithms, both use the historical information to infer the current information in order to make appropriate decisions. One algorithm, centralized DRL, makes decisions for all IoT devices at one time with global information. The other algorithm, distributed DRL, makes a decision only for one IoT device at one time using local information. Finally, simulation results show that the two DRL algorithms achieve comparable performance as the optimal user association policy which requires perfect real-time information, and the distributed DRL algorithm has the advantage of scalability.' author: - \| Qianqian Zhang, Ying-Chang Liang, Fellow, IEEE, and H. Vincent Poor, Fellow, IEEE\ [^1] title: Intelligent User Association for Symbiotic Radio Networks using Deep Reinforcement Learning --- Symbiotic radio networks (SRN), ambient backscatter communication (AmBC), user association, deep reinforcement learning. Introduction {#sec:intro} ============ The exponential growth in the number of Internet-of-Things (IoT) devices will lead to an enormous demand on wireless spectrum and network infrastructure [@andrews2014will; @Wang2017A; @zhang2016spectrum]. To support massive IoT connections, it is highly desirable to design spectrum-, energy-, and infrastructure-efficient communication technologies. Symbiotic radio networks (SRN) [@zhang2019backscatter; @long2019full; @guo2019resource], in which an IoT network parasitizes in a primary network, is envisioned as a promising technique to achieve this goal. In addition, when ambient backscatter communication (AmBC) [@liu2013ambient] is used for IoT transmission, the IoT devices in SRN transmit their messages to theirs destinations by reflecting the signals received from the primary transmitter without requiring active radio-frequency (RF) transmitter chain. That means, the data transmission of the IoT device uses the passive radio technology and does not require dedicated spectrum and infrastructure. As such, SRN has attracted increasing attention from both academia and industry recently [@wang2016ambient; @qian2017semi; @ZhangLiangGlobecom17; @yang2018modulation; @yang2018cooperative; @guo2019exploiting; @zhang2019constellation; @kang2018riding]. In AmBC-based SRN, the IoT network is an always beneficial party, and thus there are three types of symbiotic relationships based on the interaction between the two coexisting networks: parasitism, commensalism, and mutualism. Consider a simple SRN model, which consists of three nodes: an RF source, a backscatter IoT device, and a reader. The IoT device backscatters the ambient RF signal by changing its reflection coefficient, through which the information of the IoT device is transmitted to the reader. When the backscatter link is relatively strong as compared to the direct link, and the IoT and primary transmissions have the same baud rate, the IoT gains the transmission opportunity, but it causes severe interference to the primary transmission. Thus, the two networks form the parasitism relationship [@zhang2019backscatter; @kang2018riding]. When the backscatter link is very weak as compared to the direct link, the effect of backscatter link on the primary transmission is negligible. Thus, in this case, the two networks form the commensalism relationship. Due to the weak backscatter signal, in general, the IoT transmission is much slower than the primary transmission to enhance the transmission performance. Energy detector, which is simple and easy to accomplish, is used in [@liu2013ambient; @wang2016ambient; @qian2017semi; @ZhangLiangGlobecom17] to recover the IoT message. However, since the direct link signal is treated as interference in energy detector, the performance suffers from degradation. The performance of the IoT transmission can be improved through interference cancellation [@yang2018modulation; @guo2019exploiting] or cooperative receiver [@yang2018cooperative; @zhang2019constellation]. In fact, the backscatter link signal contains the RF source signal and the IoT transmission rate is typically much lower than the primary transmission rate. Thus, the backscatter link can be seen as an additional path of the primary transmission and the slowly changing reflection coefficient of the IoT device introduces time variation for the channel. This observation indicates that the existence of the IoT transmission can improve the performance of the primary system. To achieve it, we need the cooperation between the IoT transmission and the primary network. An example of the cooperation is that the primary receiver and the reader are integrated as a cooperative receiver, which decodes the messages not only from the RF source, but also from the IoT device. In [@yang2018cooperative], the signal detection problem is considered for this scenario and the results show that the existence of the backscatter link benefits the detection of the RF source message based on the joint decoding. Thus, the cooperative design can achieve a mutualism relationship between the primary and IoT transmissions. In this paper, we are interested in the user association problem for AmBC-based SRN. The base station (BS) in the primary network serves the cellular users through time division multiple access (TDMA), and each IoT device is associated with one cellular user for information transmission by reflecting the signals received from the BS, and each cellular user decodes the messages from the BS and the associated IoT devices using the successive interference cancelation (SIC) strategy. For user association problem in SRN, the BS determines which cellular user an IoT device should be associated with in order to maximize the sum rate of all IoT devices. In order to obtain the optimal user association strategy, the full real-time channel information is required. However, it is impractical for the BS to obtain all channel information since it involves a great amount of overhead. To overcome this challenge, we use deep reinforcement learning (DRL) approach to infer the real-time channel information by using the historical channel knowledge[^2] based on the channel correlation between different frames. We propose two DRL algorithms, referred to as centralized DRL algorithm and distributed DRL algorithm, to make proper decisions for the user association problem. The centralized DRL algorithm uses the historical global information as the current state to make decisions for all IoT devices at one time, while the distributed DRL algorithm uses the historical local information[^3] as the current state to make a decision for one IoT device at one time. Compared with the centralized DRL algorithm, the distributed DRL algorithm has the advantage of scalability, though at the cost of a slightly more information. In a nutshell, the main contributions of this paper are summarized as follows: - We formulate the user association problem in SRN, which is a challenging task especially for complicated environment. - We propose two DRL-based user association algorithms, namely, centralized DRL and distributed DRL, without the requirement of the full real-time channel information. - The two DRL algorithms use the historical channel information to infer the current information for decision making. - We show that the two proposed DRL algorithms can achieve a performance close to that of the optimal policy with perfect real-time channel information. - Finally, we show that the centralized DRL algorithm needs less information to converge while the distributed DRL algorithm is scalable. Related Works: Recently, DRL has been widely and successfully applied in wireless communication systems, see [@luong2018applicationsDeep] for an excellent overview. In particular, in [@anh2018deep], the authors study the time scheduling problem in RF-powered backscatter cognitive radio systems using DRL to maximize the total transmission rate. In [@chu2018reinforcement], a DRL-based access channel control problem is studied for the uplink wireless system with limited access channels. A DRL-based algorithm is proposed in [@zhang2018deep] to select proper modulation and coding scheme in cognitive heterogenous networks by learning the interference patten. DRL is adopted in [@he2017deep] to schedule users in order to enhance the sum rate in a caching network. In [@wang2018handover], DRL algorithm is used to reduce the handover rate under a constraint of the minimum sum rate. In [@yu2018deep], a distributed DRL multiple access algorithm is proposed to enhance the uplink sum rate in a multi-user wireless system. DRL approach is used in [@zhao2018deep] for user association and resource allocation in heterogeneous networks to maximize the overall network utility. Handoff policy in mmWave scenario is studied in [@sun2018smart] by taking into account the mmWave channel characteristics and the quality of service (QoS) requirements of users. Distributed dynamic power control problem is studied in [@nasir2018deep] for wireless networks using DRL algorithm. Organization: The rest of the paper is organized as follows. In Section II, the SRN model is established in detail. In Section III, we formulate the user association problem and analyze the optimal policy. Section IV present the two proposed DRL algorithms. Section V presents substantial simulation results for demonstrating the performance. Finally, the paper is concluded in Section VI. System Model {#sec:system model} ============ The system model for the SRN considered in this paper is shown in Fig. \[fig:system model\], in which an IoT network parasitizes in a primary network. In particular, the BS in the primary network serves $M$ cellular users through TDMA manner (see Fig. \[fig:frame\]), while $N$ IoT devices in the IoT network transmit their messages to the associated cellular users by reflecting the received signals from the BS. Specifically, as shown in Fig. \[fig:frame\], each IoT device only transmits information in one time slot corresponding to one associated cellular user. The cellular user decodes the signals from both the BS and the associated IoT devices using SIC strategy. In the following, we provide the channel model, the signal model, and the signal-to-interference-plus-noise ratio (SINR) model for the SRN. ![System model.[]{data-label="fig:system model"}](systemmodel.pdf){width=".99\columnwidth"} ![Frame structure of: (a) the primary network; (b) IoT device $n$.[]{data-label="fig:frame"}](frame.pdf){width=".99\columnwidth"} Channel Model ------------- Here, each channel in the SRN consists of two components: a large-scale fading component and a small-scale fading component. Denote by $h_m$ the channel coefficient from BS to User $m$ with $h_m = \sqrt{\lambda_{m}}\tilde{h}_m$, by $f_n$ the channel coefficient from BS to IoT Device $n$ with $f_n = \sqrt{\lambda_{n}}\tilde{f}_n$, and by $g_{m,n}$ the channel coefficient from IoT Device $n$ to User $m$ with $g_{m,n} = \sqrt{\lambda_{m,n}}\tilde{g}_{m,n}$, where $\lambda_{m}$, $\lambda_{n}$, and $\lambda_{m,n}$ represent the corresponding large-scale fading components, and $\tilde{h}_m$, $\tilde{f}_n$, and $\tilde{g}_{m,n}$ represent the corresponding small-scale fading components. The large-scale fading components remain unchanged for a fixed distance between the two corresponding nodes, while the small-scale fading components remain unchange in one frame, but vary in different frames. We use Jakes’ model to represent the variation of the small-scale fading component for each channel in frame $t$, which yields [@liang2017spectrum] $$\begin{aligned} \tilde{h}_m{(t)} &= \rho\tilde{h}_m{(t-1)} + e_{m}{(t)}, \label{eq:channelRelation1}\\ \tilde{f}_n{(t)} &= \rho\tilde{f}_n{(t-1)} + e_{n}{(t)},\label{eq:channelRelation2}\\ \tilde{g}_{m,n}{(t)} &= \rho\tilde{g}_{m,n}{(t-1)} + e_{m,n}{(t)},\label{eq:channelRelation3}\end{aligned}$$ for $m = 1, \cdots, M$ and $n = 1, \cdots, N$, where $\tilde{h}_m{(0)}\sim\mathcal {CN}(0,1)$, $\tilde{f}_n{(0)}\sim\mathcal {CN}(0,1)$, and $\tilde{g}_{m,n}{(0)}\sim\mathcal {CN}(0,1)$, and $e_{m}{(t)}$, $e_{n}{(t)}$, and $e_{m,n}{(t)}$ are the independent and identically distributed random variables for any frame $t$ with distribution $\mathcal {CN}(0,1-\rho^2)$, and $\mathcal{CN}(\mu, \sigma^2)$ denotes the complex Gaussian distribution with mean $\mu$ and variance $\sigma^2$. The variable $\rho$ represents the correlation of channels between different frames. Signal Model ------------ As shown in Fig. \[fig:system model\], the BS transmits message $x_m$ with unit power to User $m$ in one time slot during one frame, while the IoT Device $n$ backscatters the received BS signals with its own message $c_n$ to one associated cellular user. Suppose that the symbol period for each IoT device covers $K$ BS symbol periods [@liu2013ambient]. The received signals at User $m$ can be written as $$\label{eq:receivedSignal} y_m = \sqrt{p}h_mx_m+\sum_{n=1}^N a_{m,n}\sqrt{p}\alpha_n f_ng_{m,n}x_mc_n + u_m,$$ where $p$ is the transmitted power at the BS, $\alpha_n$ denotes the reflection coefficient of IoT Device $n$, $u_m$ is the complex Gaussian noise at User $m$ with $u_m\sim\mathcal {CN}(0,\sigma^2)$, and $a_{m,n}\in\{0,1\}$ is the user association indicator. If $a_{m,n} = 1$, IoT Device $n$ is associated with User $m$, i.e., IoT Device $n$ transmits information when the BS serves User $m$; otherwise $a_{m,n} = 0$. SINR model ---------- The cellular user adopts SIC strategy to decode the messages for its own and the associated IoT devices. Due to the double fading, the backscatter link is weaker than the direct link. Thus, the cellular user needs to decode its own message first. After that, the cellular user decodes the messages of the associated IoT devices. When there are multiple IoT devices are associated with the same cellular user, the cellular user first decodes the message of the strongest IoT device by treating other IoT devices’ signals as interference. According to this strategy, we first define ${h}_{m,n} \triangleq \|\alpha_n\|^2\|f_n\|^2\|g_{m,n}\|^2$, and use set $\tilde{I}_n = \{l\|h_{m,l}< h_{m,n}, l = 1,\cdots,N\}$ to indicate the identify numbers of the IoT devices that may interfere with IoT Device $n$. Then, the SINR of IoT Device $n$ at User $m$ is given by [@liang1999downlink] $$\begin{aligned} \label{eq:SINRdmn} \gamma_{m,n}=\frac{a_{m,n}Kp \|\alpha_n\|^2\|f_n\|^2\|g_{m,n}\|^2}{ \sum_{l\in \tilde{I}_n} a_{m,l} Kp{h}_{m,l} +\sigma^2}.\end{aligned}$$ Optimal User Association Policy =============================== In this section, we first formulate the user association problem for the SRN, which associates each IoT device with a suitable cellular user to maximize the sum rate of the IoT devices. Then we present the optimal policy for this formulated user association problem. Problem Formulation ------------------- In SRN, the IoT transmission relies on the primary cellular transmission. Hence, different association scheme yields different IoT transmission rate, due to the different channel gains. Specifically, based on , since each channel gain may vary in different frames, the IoT devices may need to be associated with different cellular users in different frames to achieve higher SINR, thereby higher transmission rate. Meanwhile, if there are multiple IoT devices associated with the same cellular user, there will exist interference which affects the IoT transmission rate. Thus, it is significantly important to design a suitable user association policy. In what follows, we will formulate the user association problem mathematically. The achievable rate, $R_{m,n}$, for IoT Device $n$ backscattering the signals to User $m$ is given by $$\begin{aligned} \label{eq:ratemn} R_{m,n}=\frac{1}{K}a_{m,n}\log_2(1+\gamma_{m,n}).\end{aligned}$$ The sum rate for all IoT devices in the SRN can be written as $\sum_{m = 1}^{M}\sum_{n=1}^{N}R_{m,n}$. Thus, the user association problem is expressed as $$\begin{aligned} \label{eq:ProblemFormulation} \mathbf{P1}: \;\; \max\limits_{\mathbb{A}} &\;\;\sum_{m = 1}^{M}\sum_{n=1}^{N}R_{m,n}\\ s.t.&\;\;\sum_{m=1}^{M}a_{m,n} = 1, \label{eq:constraint}\end{aligned}$$ where $\mathbb{A}$ represents the association index set composed by $a_{m,n}, m = 1,\cdots,M, n = 1,\cdots, N$, and means each IoT device only selects one time slot for information transmission in one frame, which is consistent with Fig.\[fig:frame\](b). Note that the user association policy can be performed either at the IoT devices or at the BS. In this paper, we consider the user association policy is performed at the BS since the BS has stronger computing capacity than the IoT devices. The Optimal Policy {#sec:policy} ------------------ To obtain the optimal index set $\mathbb{A}^$, it is clear that three steps are required: 1) list all possibility index set $\mathbb{A}$ satisfying ; 2) calculate the sum rate of all IoT devices for each possible index set $\mathbb{A}$; 3) select the optimal index set $\mathbb{A}^$ for maximizing $\sum_{m = 1}^{M}\sum_{n=1}^{N}R_{m,n}$. It is noted that to solve the optimization problem $\mathbf{P1}$, the complete real-time channel information is required to calculate the real-time SINR in . However, according to the frame structure in Fig. \[fig:frame\], each cellular user only receives the signals at its corresponding time slot, while each IoT device transmits its message only in one chosen time slot. Thus, the BS can only get the channel information from the IoT device to its associated cellular user. In other words, it is difficult for the BS to obtain the full real-time channel information from the IoT devices to all cellular users. Therefore, it is impractical for the BS to calculate the optimal index set $\mathbb{A}^$ and derive the optimal policy. Deep Reinforcement Learning Algorithms ====================================== In this section, we provide two DRL algorithms to solve the user association problem in SRN without requiring the full real-time channel information. One DRL algorithm, referred to as centralized DRL, uses the globally available information as the current states and obtains the user association decisions for all IoT devices at one time. The other DRL algorithm, referred to as distributed DRL, uses the locally available information as the current local states and obtains the user association decision only for one IoT device at one time. In the following, we will elaborate the basic principle, introduce the overview of DRL, and present the two DRL algorithms in detail. Basic Principle {#sec:basicPrinciple} --------------- As described in Section \[sec:policy\], the optimal policy of the user association problem requires the full real-time SINR in , which means that the full real-time channel information is needed. However, it is impractical for the BS to obtain the full real-time channel estimation since in one frame, the BS can only obtain the channel information between the IoT device and its associated cellular user instead of all channel information between all IoT devices and all cellular users. In fact, the channels between different frames are correlated due to the following two reasons: 1) for the channels in different frames, the large-scale fading component remains constant if the location is unchanged; 2) the small-scale fading component follows the first-order complex Gauss-Markov process based on , , and . Thus, if the BS can learn the correlation between the channels in different frames by exploring and exploiting the historical channel information, it is possible for BS to infer the current channel information and associate each IoT device with an appropriate cellular user to maximize the IoT sum transmission rate in each frame. DRL can effectively learn a hidden correlation by trial-and-error and design its optimal policy from the interaction with the environment [@mnih2015human]. Therefore, we can use DRL to learn the channel correction and design a proper user association policy to maximize the sum transmission rate. Overview of DRL --------------- In this section, we will present the overview of DRL technology. For that, we first elaborate the RL framework. ### RL Framework ![RL framework.[]{data-label="fig:RLframe"}](RLframework.pdf){width=".88\columnwidth"} In reinforcement learning process, the agent learns its own best policy through interacting with its environment over time [@kaelbling1996reinforcement; @sutton2018reinforcement]. Here, we first define the main elements of RL. Denote by $\mathcal{S}$ the set of all possible environment states $s$, by $\mathcal{A}$ the set of all possible actions $a$, by $r(s,a)$ the immediate reward when adopting action $a\in\mathcal{A}$ under the environment state $s \in \mathcal{S}$, and by $\pi$ the policy that the agent uses to map the current environment state to the pending action. As shown in Fig. \[fig:RLframe\], the agent first observes the current state $s\in \mathcal{S}$, and then takes action $a \in \mathcal{A} $ by the current policy $\pi$. After taking an action, the environment state changes from $s$ to $s'\in \mathcal{S}$, and the agent gets an immediate reward $r(s,a)$. According to the observed information, $s'$ and $r(s,a)$, the agent repeatedly adjusts its policy to approach to the optimal policy. The agent adjusts its policy to maximize the long-term reward. Notice that the maximization of the long-term reward is not equivalent to that of the immediate reward since for one state-action pair $(s,a)$ with a high immediate reward, its next state-action pair $(s',a')$ may suffer from a low immediate reward. Thus, the long-term reward includes not only the immediate reward but also the future reward, which can be expressed as $$\begin{aligned} \label{eq:long-termReward} \mathbf{\emph{Q}}(s,a)=r(s,a)+\gamma\sum_{s'\in \mathcal{S}}\sum_{a'\in\mathcal{A}}P_{s,s'}(a)\mathbf{\emph{Q}}(s',a'),\end{aligned}$$ where $\gamma\in [0,1]$ is a discount factor indicating the impact of the future reward and $P_{s,s'}(a)$ denotes the transition probability from the state $s$ to the state $s'$ when taking action $a$. In the RL process, the agent aims to take an optimal action and find the optimal policy $\pi^(s)$ under the current state $s$ by maximizing the long-term reward. Thus, based on , the optimal long-term reward $\mathbf{\emph{Q}}^(s,a)$ can be written as $$\begin{aligned} \label{eq:optimalLong-termReward} \mathbf{\emph{Q}}^(s,a)=r(s,a)+\gamma\sum_{s'\in \mathcal{S}}P_{s,s'}(a)\max_{a'\in\mathcal{A} }\mathbf{\emph{Q}}^(s',a').\end{aligned}$$ And the optimal policy $\pi^(s)$ is $$\begin{aligned} \label{eq:optimalPolicy} \pi^(s)=\arg \max_{a\in\mathcal{A} } \mathbf{\emph{Q}}^(s,a).\end{aligned}$$ Actually, the optimal policy in and the optimal long-term reward in are very difficult to be obtained directly since the transition probability $P_{s,s'}(a)$ is typically unknown for the agent especially with the complicated environment. The Q-learning algorithm is a well-known model-free RL algorithm to obtain the optimal policy, which does not require the transition probability $P_{s,s'}(a)$. Specifically, the Q-learning algorithm constructs a lookup $\|\mathcal{S}\|\times\|\mathcal{A}\|$ Q-table, in which $\mathbf{\emph{Q}}(s,a)$ as element indicates the long-term rewards of all possible state-action pairs. In addition, the agent takes actions through the $\epsilon$-greedy policy for each time step and obtains the corresponding experience $(s,a,r,s')$. After each experience $(s,a,r,s')$, the Q-learning algorithm updates the corresponding element in Q-table according to $$\label{eq:Qtable} Q(s,a)\leftarrow (1-\alpha)Q(s,a) +\alpha\left[r(s,a)+\gamma\max_{a'\in\mathcal{A} }Q(s',a')\right],$$ where $\alpha$ is the learning rate. Note that the Q-table is initialized randomly. The $\epsilon$-greedy policy implies that the agent takes a random action from the action space $\mathcal{A}$ with probability $\epsilon$, whereas executes the action $a^$ that makes the maximum value in the Q-table given a current state $s$, i.e., $a^ = \arg\max_{a\in\mathcal{A} }Q(s,a)$, with probability $1-\epsilon$ [@mnih2015human]. The $\epsilon$-greedy policy can avoid falling into the local optimum. The main reason is that the random action with a probability of $\epsilon$ can explore more possible action and experience the best action to update the Q-table. In fact, when the state space and the action space are small, the Q-learning algorithm can rapidly experience all possible state-action pairs to update the Q-table, thereby high performance. However, in practice, the size of the state and action spaces are typically large, especially with complicated environment. In this case, the performance of the Q-learning algorithm is degraded since it is difficult to experience all possible actions especially the best action and it is unacceptable to storage the large Q-table. To overcome the shortcoming of the Q-learning algorithm, DRL is introduced to find the optimal policy under the large state-action spaces. In the following, we will provide the DRL framework. ### DRL Framework In DRL, a deep neural network, referred to as deep Q-network (DQN), instead of the Q-table is implemented to estimate the long-term reward $\mathbf{\emph{Q}}(s,a)$, as shown in Fig. \[fig:DRLframe\]. The DQN can be expressed as $\mathbf{\emph{Q}}(s,a;\bm \theta)$, where $\bm \theta$ is the weights of the DQN. The input of the DQN is one of the environment states, i.e., $s\in \mathcal{S}$, and the output is the long-term reward $\mathbf{\emph{Q}}(s,a;\bm \theta)$ of each possible action $a$ in $\mathcal{A}$ for a given environment state $s$. In fact, for DRL, to achieve an approximate value $Q^(s,a)$, the agent needs to update the DQN weights $\bm \theta$, which is equivalent to the update of Q-table in RL. Similarly, the DRL uses each experience $(s,a,r,s')$ obtained by the $\epsilon$-greedy policy to train the DQN. The process of training DQN aims to minimize the loss function $\mathbb{L(\bm \theta)}$, which can be expressed as $$\label{eq:LossFunction} \mathbb{L(\bm \theta)} = \mathbb{E} \left[\|y_{tar}-\mathbf{\emph{Q}}(s,a;\bm \theta)\|^2\right],$$ where $y_{tar}$ is the target value, which is given by $$\label{eq:TargetValue} y_{tar} = r(s,a) + \gamma\max_{a'\in\mathcal{A} } \mathbf{\emph{Q}}(s',a';\bm \theta^-),$$ where $\bm \theta^-$ is the old weights of the DQN, which is updated once per $T_u$ steps. We call $\mathbf{\emph{Q}}(s,a;\bm \theta^-)$ the target Q-network, which updates its weights $\bm \theta^-$ frequently but slowly. The target Q-network can stabilize the learning algorithm by removing the correlations among the targets and the estimated Q-values. Note that in DQN, experience replay mechanism is also used to overcome the instability of the learning algorithm [@mnih2015human]. During the learning process, the agent not only uses the current experience $(s,a,r,s')$, but also uses the old experiences. In particular, the neural network is trained by randomly sampling a minibatches of $Z$ experiences from the replay memory $\mathbb{D}$. The replay memory $\mathbb{D}$ is used to store the experiences $(s,a,r,s')$ with a first-in-first-out principle. Once getting a new experience, the agent puts it into the replay memory $\mathbb{D}$. The size of this replay memory $\mathbb{D}$ is $N_E$. By using the experience replay mechanism, the experiences used for learning are more like independent and identically distributed, thereby reducing the correlations among the observations. Therefore, the experience replay mechanism increases the stability of the learning process. ![DRL framework.[]{data-label="fig:DRLframe"}](DRLframework.pdf){width=".88\columnwidth"} Centralized DRL-based User Association Algorithm {#sec:centrolized} ------------------------------------------------ In this subsection, we present the centralized DRL-based user association algorithm, in which the BS serves as the agent. In this algorithm, the BS makes the user association decisions for all IoT devices at one time for a given environment state. To begin with, we introduce the action space, the state space, and the immediate reward function for this algorithm. ### Actions Since the centralized DRL algorithm aims to associate each IoT device with a proper cellular user to maximize the sum rate, the action space needs to include all possible and available association schemes. Thus, the action space is given by $$\label{eq:C_action} \mathcal{A}_c = \{\{b_1,\cdots, b_N\}_1,\cdots, \{b_1,\cdots, b_N\}_{M^N}\},$$ where $b_n\in\{1, \cdots, M\}$ denotes the index of the cellular user associated with the IoT Device $n$. The number of possible actions is $M^N$, i.e., the size of this action space is $M^N$. We take an example to understand this action space. Assuming that there are $M = 2$ cellular users and $N = 2$ IoT devices, the action space $\mathcal{A}_c$ is $\mathcal{A}_c = \{\{1,1\},\{1,2\},\{2,1\},\{2,2\}\}$, which means there are $2^2 = 4$ possible actions. ### States {#sec:C_state} Since the DRL agent trains the DQN based on each experience $(s,a,r,s')$, it is important for the DRL agent to collect a proper and available state to provide useful knowledge for decision making. In Section \[sec:policy\], we have stated that the full real-time channel information is difficult to be obtained. However, the channels between different frames are correlated, which has been discussed in Section \[sec:basicPrinciple\]. As such, we can use the historical channel information as the state to optimize the policy. Here, we denote by $\mathcal{H}_{L} = \{h_{m,n}\}$ the historical channel information of all backscatter links. After each interaction with environment, $\mathcal{H}_{L}$ will be update. In particularly, at the end of frame $t$, User $m$ transmits the backscatter channels information $h_{m,n}(t)$ from IoT Device $n$ associated with it to the BS. Then the BS updates $\mathcal{H}_{L}(t)$ with the information $h_{m,n}(t)$ and considers the updated $\mathcal{H}_{L}(t)$ as the state for the frame $(t+1)$. To summarize, the state in frame $t$ is given by $$\label{eq:C_state} s_c(t) = \mathcal{H}_{L}(t-1).$$ Note that before feeding $s_c(t)$ into DQN, we first normalize it to guarantee the performance of the centralized DRL algorithm. ### Reward Function The goal of this centralized DRL algorithm is to maximize the sum rate of all IoT devices. Thus the immediate reward function $r_c(t)$ in frame $t$ shall be the sum rate of all IoT devices. i.e., $$\label{eq:C_reward} r_c(t) = \sum_{m=1}^{M}\sum_{n=1}^{N}R_{m,n}(t).$$ Note that after taking action $a$ by the observed state $s$, the BS will obtain the immediate reward by the feedback from the cellular users. Fig. \[fig:C\_structureDRL\] shows the structure of the proposed centralized DRL algorithm. In this algorithm, the agent delivers the decision $a_c(t)$ made according to $\epsilon$-greedy policy to the IoT devices. The IoT devices access the associated cellular user based on the decision from the BS. And the cellular users decode the signals of the associated IoT devices and feedback all useful and available information to BS for the calculation and the update of $r_c(t)$ and $s_c(t+1)$. Then the BS storages the experience $(s_c(t),a_c(t),r_c(t),s_c(t+1))$ into the replay memory $\mathbb{D}$, and randomly samples a minibatch of experiences in $\mathbb{D}$ to train the DQN. The DQN is used to make decision for the next frame according to $\epsilon$-greedy policy. In addition, the pseudocode of the proposed centralized DRL-based user association algorithm is shown in Algorithm \[alg:centralized\]. ![The structure of the proposed centralized DRL-based user association algorithm.[]{data-label="fig:C_structureDRL"}](CentralizedDRL.pdf){width=".88\columnwidth"} \[alg:centralized\] Initialize the weights $\bm \theta_c$ of the DQN randomly;\ Initialize the weights $\bm \theta^-_c$ of the target Q-network with $\bm \theta^-_c = \bm \theta_c$;\ Initialize the size of minibatch $Z$;\ Initialize the target Q-network replacement frequency $T_u$;\ The agent takes actions randomly and storages the corresponding experience $(s_c,a_c,r_c,s_c')$ into the replay memory $\mathbb{D}$ until there are $Z$ experiences.\ The agent selects an action $a_c(t)$ through $\epsilon$-greedy policy in frame $t, (t>Z)$;\ The agent calculates the immediate reward $r_c(t)$ after taking action $a_c(t)$ in frame $t$;\ The agent observes a new state $s_c(t+1)$ in frame $(t+1)$;\ The agent stores the new experience $(s_c(t),a_c(t), r_c(t), s_c(t+1))$ into the replay memory $\mathbb{D}$;\ The agent randomly samples a minibatch of $Z$ experiences $(s_c,a_c,r_c,s_c')$ from the replay memory $\mathbb{D}$ to train the DQN;\ The agent updates the DQN weights $\bm \theta_c$;\ The agent updates the target Q-network weights $\bm \theta^-_c$ once per $T_u$ steps with $\bm \theta^-_c = \bm \theta_c$.\ For the centralized DRL algorithm, when $N$ is large, the state-action space becomes very large. In this case, it is difficult for this algorithm to train the DQN successfully. In addition, if $N$ increases, this algorithm can not work since the state-action space changes, resulting in the inability to use the designed DQN. In other words, the centralized DRL algorithm is not a scalable algorithm. To overcome the above challenges, we propose another algorithm called distributed DRL-based user association algorithm. In the following, we will present this algorithm. Distributed DRL-based User Association Algorithm ------------------------------------------------ In this subsection, we provide a distributed DRL-based user association algorithm, in which the BS serves as the agent and uses a centralized training and distributed execution framework [@calabrese2016learning]. In this algorithm, the BS allocates $N$ computing units to make decision for $N$ IoT devices individually. In other words, Unit $n$ inputs the state of IoT Device $n$ and outputs the action of IoT Device $n$. Here, we first introduce the action space, the state space, and the immediate reward function. ### Actions In the distributed DRL algorithm, the computing unit makes decision only for one IoT device at one time with a given state of the corresponding IoT device. Thus, the action space is given by $$\label{eq:D_action} \mathcal{A}_d = \{1,2,\cdots, M\}.$$ ### States {#states} Since the units make the user association decision individually, it is difficult to control the decision of other unit in the distributed DRL algorithm. We notice that the optimal user association policy for this algorithm is not only related to the channel information but also related to the interference information. This means that the state requires not only the historical channel information, but also the interference information. The interference information includes two components: interferer information and interfered information. In what follows, for the distributed DRL algorithm, we describe the state $s_d^{n}(t)$ of IoT Device $n$ conditioned on associated with User $m$ at frame $t$, which is divided into three feature groups. - Local Information: According to , the agent needs to feed the channel information into the DQN to provide useful knowledge for learning the optimal policy. Since it is difficult to obtain the channel information between IoT Device $n$ and all cellular users, the agent uses the historical information to explore and infer the current channel information, which is similar to Section \[sec:C\_state\]. Thus, the state is designed to include the historical channel information $\mathcal{H}_L^n(t-1)$, where $\mathcal{H}_L^n(t-1)$ is the historical channel information between IoT Device $n$ and all cellular users updated in frame $(t-1)$. Meanwhile, the state at frame $t$ includes the action taking by IoT Device $n$ at frame $(t-1)$ to suggest the effect of historical action. In addition, since the agent trains the DQN using all environment experiences, in order to identify all IoT devices, the state is designed to include the identity number, i.e., $n$. - Interferer Information: The state is designed to include the interferer information to observe the interference from other IoT devices when decoding the message from IoT Device $n$. In particular, according to , if IoT Device $n$ is associated with User $m$ in frame $(t-1)$, User $m$ will feedback the interferer information when decoding the IoT Device $n$ message, $I_n(t-1)$, to the BS, where $I_n(t-1) = \sum_{l\in \tilde{I}_n(t-1)}p a_{m,l}^{d}(t-1){{h}}_{m,l}(t-1)$ and $a_{m,l}^{d}(t-1)\in \{0,1\}$ indicates whether the IoT Device $l$ is associated with User $m$ in frame $(t-1)$. - Interfered Information: Finally, the agent uses the feedback from User $m$ to sense the interference $O_n(t-1)$ from IoT Device $n$ to other IoT devices in frame $(t-1)$, where $O_n(t-1) = \sum_{l\in \tilde{O}_n(t-1)}p a_{m,l}^{d}(t-1){{h}}_{m,l}(t-1)$ and $\tilde{O}_n(t-1) = \{l\|h_{m,l}(t-1)> h_{m,n}(t-1), l = 1,\cdots,N\}$ is the identify number set of the IoT device that may be interfered by IoT Device $n$ at frame $(t-1)$. And the state is designed to include the interference information $O_n(t-1)$. To summarize, the state $s_d^{n}(t)$ of IoT Device $n$ at frame $t$ is given by $$\label{eq:D_state} s_d^{n}(t) = \{\mathcal{H}_L^n(t-1),a_d^n(t-1), n, I_n(t-1), O_n(t-1) \},$$ where $a_d^n(t-1)$ is the action of IoT Device $n$ at frame $(t-1)$. Notice that $s_d^{n}(t)$ is normalized to guarantee the performance of the distributed DRL algorithm. ### Reward Function The immediate reward function should evaluate the effect of the action taken on the goal of maximizing the sum rate. Here, the immediate reward includes not only the current transmission rate of IoT Device $n$, but also the interference with other IoT devices. The main reason is that if the decision process for each IoT device aims to maximize its own transmission rate, it is difficult to converge to an optimal policy for maximizing the sum rate. To quantify the effect of interference, similar to [@nasir2018deep], the agent first calculates the transmission rate, $R_{m,l}^{-n}(t)$, without the interference from IoT Device $n$ for IoT Device $l\in \tilde{O}_n(t)$, which can be expressed as $$\begin{aligned} \label{eq:D_withoutInterference} &R_{m,l}^{-n}(t)= \frac{1}{K}a_{m,l}^{d}(t)\times \nonumber \\ &~~~~~~\log_2\left(1+\frac{a_{m,l}^{d}(t)Kp \|\alpha_l\|^2\|f_l\|^2\|g_{m,l}\|^2}{ \sum_{i\in\tilde{I}_l(t) ,i\neq n} a_{m,i}^{d}(t) Kp{{h}}_{m,i} +\sigma^2}\right),\end{aligned}$$ Then, the agent computes the effect of IoT Device $n$ on the IoT devices in $\tilde{O}_n(t)$ by $$\label{eq:D_interferenceEffect} \beta_{l}^{-n}(t) = R_{m,l}^{-n}(t)- R_{m,l}(t).$$ Thus the immediate reward function can be written as $$\label{eq:D_reward} r_d^n(t) = R_{m,n}(t)- \sum_{l\in \tilde{O}_n(t)}\beta_{l}^{-n}(t).$$ The reward in consists of two components: its contribution to the sum rate and the penalty about interference to other IoT device. This reward function ensures that the agent considers not only the maximization of each IoT device rate, but also the effect on other IoT devices, thereby guaranteeing the optimal policy rapidly. The structure of the proposed distributed DRL algorithm is shown in Fig. \[fig:D\_structureDRL\]. The information delivery between cellular users, the IoT devices, and the BS is the same with the centralized DRL algorithm, which is discussed in Section \[sec:centrolized\]. The difference between these two algorithms is that the BS needs to allocate $N$ computing units to make decisions for $N$ IoT devices individually in the distributed DRL algorithm. In addition, in the distributed DRL algorithm, after training the DQN, the BS delivers the updated DQN weights $\bm \theta_d$ to each computing unit. Then, the $N$ computing units make decisions, individually, for the $N$ IoT devices according to the $\epsilon$-greedy policy. In addition, the pseudocode of the proposed distributed DRL-based user association algorithm is shown in Algorithm \[alg:distributed\]. Note that if the number of IoT devices $N$ changes, the BS just changes the number of computing units to execute the distributed DRL algorithm without redesigning the DQN. In other words, this distributed DRL algorithm has the advantage of scalability. In addition, here, we consider the units at the BS make decisions for the IoT devices due to the limited computing capability of the IoT devices. If the IoT devices have enough computing capacity, it is reasonable that each IoT device as the agent makes its own decisions based on the proposed distributed DRL algorithm. ![The structure of the proposed distributed DRL-based user association algorithm.[]{data-label="fig:D_structureDRL"}](DistributedDRL.pdf){width=".88\columnwidth"} \[alg:distributed\] Initialize the weights $\bm \theta_d$ of the DQN randomly;\ Initialize the weights $\bm \theta^-_d$ of the target Q-network with $\bm \theta^-_d = \bm \theta_d$;\ Initialize the size of minibatch $Z$;\ Initialize the target Q-network replacement frequency $T_u$;\ The agent takes actions randomly and storages the corresponding experience $(s^n_d,a^n_d,r^n_d,(s^n_d)')$ of each IoT device into the replay memory $\mathbb{D}$ until there are $Z$ experiences.\ Unit $n$ selects an action $a_d^n(t)$ through $\epsilon$-greedy policy in frame $t, (t>Z)$ for IoT Device $n$, $n = 1, \cdots, N$;\ Unit $n$ calculates the immediate reward $r_d^n(t)$ after taking action $a_d^n(t)$ in frame $t$ for IoT Device $n$, $n = 1, \cdots, N$;\ The agent observes a new state $s_d^n(t+1)$ in frame $(t+1)$ of IoT Device $n$, $n = 1, \cdots, N$;\ The agent stores all new experiences $(s_d^n(t),a_d^n(t), r_d^n(t), s_d^n(t+1))$, $n = 1,\cdots, N$ into the replay memory $\mathbb{D}$;\ The agent randomly samples a minibatch of $Z$ experiences from the replay memory $\mathbb{D}$ to train the DQN;\ The agent updates the DQN weights $\bm \theta_d$;\ The agent updates the target Q-network weights $\bm \theta^-_d$ once per $T_u$ steps with $\bm \theta^-_d = \bm \theta_d$;\ The agent delivers the updated DQN weights $\bm \theta_d$ to $N$ computing units.\ Performance Evaluation {#sec:simulations} ====================== In this section, simulation results are presented to evaluate the performance of the two proposed DRL-based user association algorithms. For comparison, we consider two benchmark algorithms: random policy and optimal policy. In the random policy, each IoT device is associated with a cellular user randomly. For the optimal policy, we assume that the BS knows full perfect real-time channel information and obtains the optimal policy by the method proposed in Section \[sec:policy\]. Since it is impractical for the BS to perfectly know the full real-time channel information, the performance of the optimal policy is just the theoretical upper bound. In the following, we will present the simulation setup and the performance of the two proposed DRL-based user association algorithms. Simulation Setup ---------------- To begin with, we consider the locations of the BS, the cellular users, and the IoT devices, are in a $100$ meters by $100$ meters region. The BS is located at the center of this region. And the IoT devices and the cellular users are placed randomly based on a uniform distribution within a distance of $10\thicksim 100$ meters from the BS. We set the transmit power of the BS to $p = 40$dBm and the background noise power to $\sigma^2 = -114$dBm. We consider a distance-dependent path loss model, which is $32.45+20\log_{10}(f)+20\log_{10}(d)-G_t-G_r$ (in dB), where $f$ is the carrier frequency in Mhz, $d$ is the distance in km, $G_t$ denotes the transmit antenna gain, and $G_r$ denotes the receive antenna gain. Here we set $f = 2.4$GHz, $G_t = G_r = 2.5$dB. We assume all IoT devices have the same reflection coefficient $\alpha_n = \alpha = 0.8$ for $n = 1,\cdots, N$. And the period ratio between the IoT device and the BS is set to $K = 50$. \[tab:central\] [c c]{} Parameters & Value\ C-DQN: number of hidden layers & $3$\ C-DQN: neuron network size & $256\times128\times64$\ D-DQN: number of hidden layers & $3$\ D-DQN: neuron network size & $128\times64\times32$\ Activation function & ReLU\ Optimizer & Adam\ Learning rate & $0.01$\ Mini-batch size ($Z$) & $64$\ Replay memory size ($N_E$) & $800$\ Target-DQN updating frequency ($T_u$) & $100$\ Next, we describe the design of the hyper-parameters for the two DRL algorithms. First, the two DRL algorithms are implemented using TensorFlow, and the parameters of the two DQNs corresponding to the two DRL algorithms are listed in Table I. Furthermore, we set the discount factor to $\gamma = 0.3$. In addition, the $\epsilon$-greedy policy is used to take actions. At first, we set $\epsilon(0) = 0.2$, which means a random action is chosen with a probability of $0.2$ to explore the experiences. Then, to move from a more explorative policy to a more exploitative policy, the probability $\epsilon$ follows $\epsilon(t+1) = \max\{\epsilon_{\min},(1-\lambda_\epsilon)\epsilon(t)\}$, where $\epsilon_{\min} = 0.005$ and $\lambda_\epsilon = 0.005$. Performance for the Proposed Algorithms --------------------------------------- Fig. \[fig:DRL1\] illustrates the average sum transmission rate of all IoT devices using different algorithms. In this figure, we consider a quasi-static channel scenario by setting $\rho = 0.99$, which means the channel changes slowly. Meanwhile, we consider the number of the cellular users is $M = 3$ and the number of the IoT devices is $N = 3$ in this figure. It can be seen that both the centralized DRL algorithm and the distributed DRL algorithm can almost achieve the optimal sum transmission rate gradually in a quasi-static scenario. This observation indicates that the two DRL algorithms can learn almost perfect knowledge and design almost optimal policy in a quasi-static scenario. Meanwhile the average sum rate of the proposed two DRL algorithms is around $0.45$ bits/frame/Hz, while the average sum rate of the random policy is around $0.25$ bits/frame/Hz. That indicates the average sum rate of the proposed two DRL algorithms is almost twice the average sum rate of random policy. [.99]{}\[fig:1a\] ![The average sum transmission rate comparison with $\rho = 0.99$. Each value is a moving average of the pervious $200$ frames.[]{data-label="fig:DRL1"}](c_dc_1.eps "fig:"){width="8.5cm"} \ [.99]{}\[fig:1b\] ![The average sum transmission rate comparison with $\rho = 0.99$. Each value is a moving average of the pervious $200$ frames.[]{data-label="fig:DRL1"}](d_dc_1.eps "fig:"){width="8.5cm"} Fig. \[fig:DRL5\] presents the average sum transmission rate for different algorithms in a relative dynamic channel scenario with $\rho = 0.5$. In this figure, we set $M = N = 3$. From this figure, we can see that the two DRL algorithms can approach the performance of the optimal policy. Compared with the quasi-static scenario, the two DRL algorithms in this more dynamic scenario have a little gaps with the optimal policy. The main reason is that when the channel changes rapidly, it is more difficult to infer the next channel state. ![The average sum transmission rate comparison with $\rho = 0.5$. Each value is a moving average of the pervious $200$ frames.[]{data-label="fig:DRL5"}](dc_5.eps){width=".99\columnwidth"} Fig. \[fig:DRL0\] shows the performance of the average sum transmission rate for different algorithms in a highly dynamic scenario with $\rho = 0$. In this scenario, the small-scale fading component changes rapidly without correlation between different frames. We set $M = N = 3$. From this figure, it is seen that there exist gaps between the two proposed DRL algorithms and the optimal policy. The main reason is that when $\rho = 0$, the small-scale fading component is difficult to be learnt from the historical channel information since the channel changes without correlation between different frames. However, the proposed algorithms can approach the optimal policy. This is because the agent can learn the large-scale fading information. These observations indicate that the proposed two algorithms are effective even in a highly dynamic scenario. ![The average sum transmission rate comparison with $\rho = 0$. Each value is a moving average of the pervious $200$ frames.[]{data-label="fig:DRL0"}](dc_0.eps){width=".99\columnwidth"} The Scalability of the Distributed DRL Algorithm ------------------------------------------------ In this subsection, we present the scalability of the proposed distributed DRL algorithm. When the number of IoT devices changes, the centralized DRL algorithm can not work effectively due to the change of the action space. Fig. \[fig:DDRL5\] shows the performance of the average sum rate of different algorithms when the number of IoT devices changes with $\rho = 0$ and $M = 3$. It is seen that regardless of whether the number of IoT devices increases or decreases, the distributed DRL algorithm can approach the optimal policy, and always be better than the random policy. This figure validates the scalability of the proposed distributed DRL algorithm when the environment changes in a highly dynamic way. [.9]{}\[fig:de\] ![The average sum transmission rate comparison with $\rho = 0$. Each value is a moving average of the pervious $200$ frames.[]{data-label="fig:DDRL5"}](decrease_0.eps "fig:"){width="8.5cm"} \ [.9]{}\[fig:in\] ![The average sum transmission rate comparison with $\rho = 0$. Each value is a moving average of the pervious $200$ frames.[]{data-label="fig:DDRL5"}](increase_0.eps "fig:"){width="8.5cm"} Fig. \[fig:DDRLi\] depicts the performance of the average sum rate for different algorithms with $\rho = 0.5$, $M = 8$, and $N = 8$. In this case, the size of action space for the centralized DRL algorithm is $8^8 = 1.68\times10^7$. Thus, it is impractical to use the centralized DRL algorithm to make decisions. In addition, the optimal policy needs to search $8^8 = 1.68\times10^7$ possible index sets to obtain the optimal decision. Thus, it is too complicated to obtain the performance of the optimal policy. Therefore, in Fig. \[fig:DDRLi\], we only show the performance of the proposed distributed DRL algorithm and the random policy. From this figure, we can see that the average sum transmission rate is about $1.1$ bits/frame/Hz for the proposed distributed DRL algorithm, while the rate for the random policy is about $0.6$ bits/frame/Hz. This observation indicates that the proposed distributed DRL algorithm is effective when the number of the IoT devices and the number of the cellular users are large. ![The average sum transmission rate comparison with $\rho = 0.5$. Each value is a moving average of the pervious $200$ frames.[]{data-label="fig:DDRLi"}](dc_8.eps){width=".99\columnwidth"} Conclusions {#sec:conclusions} =========== This paper has studied the user association problem in AmBC-based SRN using the DRL approaches. Since it is difficult to obtain the full real-time channel information, we use the historical information to infer the current information by the DRL approaches to make appropriate decisions. In particular, we propose two DRL algorithms, namely, centralized DRL and distributed DRL. The centralized DRL algorithm involves the globally available information as current state and outputs one action that involves decisions for all IoT device. While the distributed DRL algorithm uses the locally available information as current state and outputs decisions individually for each IoT device. Finally, simulation results have demonstrated that the two DRL algorithms can perform close to the optimal policy with perfect real-time information. In addition, the centralized DRL algorithm needs less information than the distributed DRL algorithm, while the distributed DRL algorithm has the advantage of scalability, which means it can still work effectively even when the number of the IoT devices changes. [10]{} \[1\][\#1]{} url@samestyle \[2\][\#2]{} \[2\][[l@\#1=l@\#1\#2]{}]{} J. G. Andrews, S. Buzzi, W. Choi, S. V. Hanly, A. Lozano, A. C. Soong, and J. C. Zhang, “What will [5G]{} be?” IEEE J. Sel. Areas Commun., vol. 32, no. 6, pp. 1065–1082, Jun. 2014. Y.-P. E. [Wang]{}, X. [Lin]{}, A. [Adhikary]{}, A. [Grovlen]{}, Y. [Sui]{}, Y. [Blankenship]{}, J. [Bergman]{}, and H. S. [Razaghi]{}, “A primer on [3GPP]{} narrowband [Internet of Things]{},” IEEE Commun. Mag., vol. 55, no. 3, pp. 117–123, March 2017. L. Zhang, Y.-C. Liang, and M. Xiao, “Spectrum sharing for [Internet of Things]{}: A survey,” arXiv preprint arXiv:1810.04408, 2018. Q. [Zhang]{}, L. [Zhang]{}, Y. [Liang]{}, and P. [Kam]{}, “Backscatter-[NOMA]{}: A symbiotic system of cellular and [Internet-of-Things]{} networks,” IEEE Access, vol. 7, pp. 20000–20013, 2019. R. [Long]{}, H. [Guo]{}, L. [Zhang]{}, and Y. [Liang]{}, “Full-duplex backscatter communications in symbiotic radio systems,” IEEE Access, vol. 7, pp. 21597–21608, 2019. H. [Guo]{}, Y. [Liang]{}, R. [Long]{}, S. [Xiao]{}, and Q. [Zhang]{}, “Resource allocation for symbiotic radio system with fading channels,” IEEE Access, vol. 7, pp. 34333–34347, 2019. V. Liu, A. Parks, V. Talla, S. Gollakota, D. Wetherall, and J. R. Smith, “Ambient backscatter: Wireless communication out of thin air,” Proc. ACM SIGCOMM, vol. 43, no. 4, pp. 39–50, Oct. 2013. G. Wang, F. Gao, R. Fan, and C. Tellambura, “Ambient backscatter communication systems: Detection and performance analysis,” IEEE Trans. Commun., vol. 64, no. 11, pp. 4836–4846, Nov. 2016. J. Qian, F. Gao, G. Wang, S. Jin, and H. Zhu, “Semi-coherent detection and performance analysis for ambient backscatter system,” IEEE Trans. Commun., vol. 65, no. 12, pp. 5266–5279, Dec. 2017. Q. Zhang and Y.-C. Liang, “Signal detection for ambient backscatter communications using unsupervised learning,” in IEEE GLOBECOM Workshop 2017, Singapore, Dec. 2017, pp. 1–6. G. Yang, Y.-C. Liang, R. Zhang, and Y. Pei, “Modulation in the air: Backscatter communication over ambient [OFDM]{} carrier,” IEEE Trans. Commun., vol. 66, no. 3, pp. 1219–1233, Mar. 2018. G. Yang, Q. Zhang, and Y.-C. Liang, “Cooperative ambient backscatter communications for green [Internet-of-Things]{},” IEEE Internet Things J., vol. 5, no. 2, pp. 1116–1130, Apr. 2018. H. Guo, Q. Zhang, S. Xiao, and Y.-C. Liang, “Exploiting multiple antennas for cognitive ambient backscatter communication,” IEEE Internet Things J., vol. 6, no. 1, pp. 765–775, 2019. Q. Zhang, H. Guo, Y.-C. Liang, and X. Yuan, “Constellation learning-based signal detection for ambient backscatter communication systems,” IEEE J. Sel. Areas Commun., vol. 37, no. 2, pp. 452–463, 2019. X. Kang, Y.-C. Liang, and J. Yang, “Riding on the primary: A new spectrum sharing paradigm for wireless-powered [IoT]{} devices,” IEEE Trans. on Wireless Commun., vol. 17, no. 9, pp. 6335–6347, Sep. 2018. N. C. Luong, D. T. Hoang, S. Gong, D. Niyato, P. Wang, Y.-C. Liang, and D. I. Kim, “Applications of deep reinforcement learning in communications and networking: A survey,” arXiv preprint arXiv:1810.07862, 2018. T. T. Anh, N. C. Luong, D. Niyato, Y.-C. Liang, and D. I. Kim, “Deep reinforcement learning for time scheduling in [RF]{}-powered backscatter cognitive radio networks,” arXiv preprint arXiv:1810.04520, 2018. M. Chu, H. Li, X. Liao, and S. Cui, “Reinforcement learning based multi-access control with energy harvesting,” in 2018 IEEE GLOBECOM.1em plus 0.5em minus 0.4emIEEE, 2018, pp. 1–6. L. Zhang, J. Tan, Y.-C. Liang, G. Feng, and D. Niyato, “Deep reinforcement learning based modulation and coding scheme selection in cognitive heterogeneous networks,” arXiv preprint arXiv:1811.02868, 2018. Y. He, Z. Zhang, F. R. Yu, N. Zhao, H. Yin, V. C. Leung, and Y. Zhang, “Deep-reinforcement-learning-based optimization for cache-enabled opportunistic interference alignment wireless networks,” IEEE Trans. Veh. Technol., vol. 66, no. 11, pp. 10433–10445, 2017. Z. Wang, L. Li, Y. Xu, H. Tian, and S. Cui, “Handover control in wireless systems via asynchronous multiuser deep reinforcement learning,” IEEE Internet Things J., vol. 5, no. 6, pp. 4296–4307, 2018. Y. Yu, T. Wang, and S. C. Liew, “Deep-reinforcement learning multiple access for heterogeneous wireless networks,” in 2018 IEEE ICC.1em plus 0.5em minus 0.4emIEEE, 2018, pp. 1–7. N. Zhao, Y.-C. Liang, D. Niyato, Y. Pei, and Y. Jiang, “Deep reinforcement learning for user association and resource allocation in heterogeneous networks,” in 2018 IEEE GLOBECOM.1em plus 0.5em minus 0.4emIEEE, 2018, pp. 1–6. Y. Sun, G. Feng, S. Qin, Y.-C. Liang, and T.-S. P. Yum, “The [SMART]{} handoff policy for millimeter wave heterogeneous cellular networks,” IEEE Trans. Mobile Comput., vol. 17, no. 6, pp. 1456–1468, 2018. Y. S. Nasir and D. Guo, “Deep reinforcement learning for distributed dynamic power allocation in wireless networks,” arXiv preprint arXiv:1808.00490, 2018. L. [Liang]{}, J. [Kim]{}, S. C. [Jha]{}, K. [Sivanesan]{}, and G. Y. [Li]{}, “Spectrum and power allocation for vehicular communications with delayed csi feedback,” IEEE Wireless Commun. Lett., vol. 6, no. 4, pp. 458–461, Aug 2017. Y.-C. [Liang]{}, P. S. F. [Chin]{}, and K. J. R. [Liu]{}, “Downlink beamforming for [DS-CDMA]{} mobile radio with multimedia services,” IEEE Trans. Commun., vol. 49, no. 7, pp. 1288–1298, Jul. 2001. V. Mnih, K. Kavukcuoglu, D. Silver, A. A. Rusu, J. Veness, M. G. Bellemare, A. Graves, M. Riedmiller, A. K. Fidjeland, G. Ostrovski et al., “Human-level control through deep reinforcement learning,” Nature, vol. 518, no. 7540, p. 529, 2015. L. P. Kaelbling, M. L. Littman, and A. W. Moore, “Reinforcement learning: A survey,” J. Artificial Intell. Research, vol. 4, pp. 237–285, 1996. R. S. Sutton and A. G. Barto, Reinforcement learning: An introduction.1em plus 0.5em minus 0.4emMIT press, 2018. F. D. [Calabrese]{}, L. [Wang]{}, E. [Ghadimi]{}, G. [Peters]{}, L. [Hanzo]{}, and P. [Soldati]{}, “Learning radio resource management in rans: Framework, opportunities, and challenges,” IEEE Commun. Mag., vol. 56, no. 9, pp. 138–145, Sep. 2018. [^1]: This work is supported by National Natural Science Foundation of China under Grants 61631005, U1801261, and 61571100. (Corresponding author: Ying-Chang Liang.*) Q. Zhang is with the National Key Laboratory of Science and Technology on Communications, and the Center for Intelligent Networking and Communications (CINC), University of Electronic Science and Technology of China (UESTC), Chengdu 611731, China (e-mail: qqzhang [email protected]). Y.-C. Liang is with the Center for Intelligent Networking and Communications (CINC), University of Electronic Science and Technology of China (UESTC), Chengdu 611731, China (e-mail: [email protected]). H. V. Poor is with the Department of Electrical Engineering, Princeton University, Princeton, NJ 08544 USA (e-mail: [email protected]). [^2]: When decoding messages, the cellular user needs to estimate the channel information, from which process, the historical channel information is obtained. [^3]: The local information represents the available information at one IoT device.	{ "pile_set_name": "ArXiv" }
--- abstract: 'The $Q^2$ evolution of polarised parton distributions at small $x$ is studied. Various analytic approximations are critically discussed. We compare the full evolution with that obtained from the leading-pole approximation to the splitting functions, and show that the validity of this approximation depends critically on the $x \to 0$ behaviour of the starting distributions. A new analytic solution which is valid at small $x$ is obtained, and its domain of applicability is discussed.' --- 20.5cm DTP/95/62\ July 1995\ 1.cm [*Analytic Approaches to the Evolution of Polarised\ Parton Distributions at Small [x]{}* ]{} 1.cm [T. Gehrmann and W.J. Stirling]{} .4cm [Departments of Physics and Mathematical Sciences, University of Durham\ Durham DH1 3LE, England ]{}\ Introduction ============ The first moment of the polarised structure function $g_1(x,Q^2)$, the Ellis-Jaffe sum rule [@ej], determines the overall spin content of the nucleon. Measurements of the Ellis-Jaffe sum rule [@exp; @slac] involve an extrapolation of $g_1$ for $x\rightarrow 0$, which is usually performed by fitting a Regge-motivated form $$g_1(x) = C x^{\alpha} \label{one}$$ to the experimental data points with the lowest $x$-values. This procedure can be problematic, since these data points are usually taken at a relatively low scale of $Q^2\sim 1 \; \mbox{GeV}^2$, whereas the overall sum rule is evaluated at the average $Q^2$ of the experiment, which is typically between $3$ and $10\;\mbox{GeV}^2$. In the recent past, various authors have attempted to calculate the asymptotic behaviour of $g_1(x)$ (see for example Ref. [@CR] for a review of the various approaches). At scales of low momentum transfer ($Q^2 \approx 1 \; \mbox{GeV}^2$), a non–perturbative calculation [@bass] of the flavour singlet contribution to $g_1$ shows good agreement with $g_1^p$ at small $x$, but it should be noted that the normalisation of this non–perturbative contribution is highly sensitive to the only approximately known value of the vacuum quark condensate. The experimental discrepancy between $g_1^p$ and $g_1^d$ in the small-$x$ region seems to contradict the above result. As the singlet distribution is identical for both targets, this discrepancy indicates a sizeable valence-quark contribution in this region. With increasing $Q^2$, perturbative corrections become more and more important. These corrections affect both the valence and the singlet contributions ($\Delta\Sigma = \sum_q(\Delta q + \Delta \bar q)$) to $g_1$ and give rise to an evolution of the corresponding parton densities [@ap] $$\begin{aligned} \frac{\partial}{\partial \ln Q^2} \Delta q_{val} (x,Q^2) & = & \int_x^1 \frac{\d y}{y} \Delta P_{qq} (y) \;\Delta q_{val} (x/y,Q^2) \nonumber\\ \label{evo} \frac{\partial}{\partial \ln Q^2} \left(\begin{array}{c} \Delta \Sigma \\ \Delta G \end{array} \right) (x,Q^2) & = & \frac{\alpha_s(Q^2)}{2\pi} \int_x^1 \frac{\d y}{y} \left(\begin{array}{cc} \Delta P_{qq} & \Delta P_{qg} \\ \Delta P_{gq} & \Delta P_{gg}\end{array} \right) (y)\; \left(\begin{array}{c} \Delta \Sigma \\ \Delta G \end{array} \right)(x/y,Q^2),\nonumber\\\end{aligned}$$ without determining the densities themselves. These enter the above equation in the form of the initial distributions $\Delta q_{val}(x,Q_0^2)$, $\Delta \Sigma(x,Q_0^2)$ and $\Delta G(x,Q_0^2)$, which form the boundary conditions for the solution. In experimental measurements, these perturbative corrections are incorporated by rescaling the value of $g_1$ to the average $Q^2$ of the experiment. This rescaling procedure relies on the assumption that the asymmetry $g_1(x) /F_1(x)$ satisfies exact Bjorken scaling, i.e. that the $Q^2$-dependence of $g_1$ coincides with that of $F_1$. Although this assumption is consistent with the present data (which cover only a small range of $Q^2$ values at fixed $x$), there is no theoretical justification for it. In particular, examination of the polarised and unpolarised splitting functions [@ap] shows that $g_1(x) /F_1(x)$ should indeed show only a very weak $Q^2$ dependence in the large-$x$ region, where both structure functions are dominated by the valence quark content, as $\Delta P_{qq}(x)$ and $P_{qq}(x)$ are identical. In contrast to this, the splitting functions in the singlet sector, which dominates the small-$x$ behaviour of $F_1$, are different. The unpolarised $P_{gq}(x)$ and $P_{gg}(x)$ have a soft gluon singularity at $x=0$, which causes the steep rise of $F_1$ in the small-$x$ region. As this singularity is absent in the polarised splitting functions (soft gluon emission does not change the spin of the parent parton), one would expect the ratio $\mid \!g_1(x) /F_1(x) \!\mid$ to decrease with increasing $Q^2$. With the exact splitting functions $$\begin{aligned} \Delta P_{qq}^{(f)} (x) & = & \frac{4}{3} \; \left[ 2\; \frac{1}{(1-x)_{+}} - 1 -x + \frac{3}{2} \delta(1-x) \right] \nonumber \\ \Delta P_{qg}^{(f)} (x) & = & 2n_f \; \frac{1}{2} \; \left[ 2x-1 \right] \nonumber\\ \Delta P_{gq}^{(f)} (x) & = & \frac{4}{3} \; \left[ 2-x \right] \nonumber\\ \label{evoful} \Delta P_{gg}^{(f)} (x) & = & 3 \; \left[ 2\; \frac{1}{(1-x)_{+}} + 2 -4x + \frac{11}{6} \delta(1-x)\right] -\;\frac{n_f}{3} \delta(1-x)\end{aligned}$$ it is not possible to find an analytic solution to (\[evo\]) with realistic boundary conditions for the whole range of $x$. By restricting themselves to small values of $x$ (although it is not [a priori]{} clear which values of $x$ can be regarded as small), various authors have attempted to determine the asymptotic behaviour of $g_1$ in the limit $Q^2\to \infty$. One possible approach [@CR] is to assume that all the $Q^2$ dependence is dominated by the evolution of the gluon, i.e. by $\Delta P_{gg}(x)$. This method gives successful predictions for the unpolarised structure functions, due to the $1/x$ pole in the unpolarised $P_{gg}$. As this pole is not present in $\Delta P_{gg}$, the validity of this approach needs to be examined more carefully. Another possible approach [@bf] to the asymptotic small-$x$ behaviour is to transform (\[evoful\]) into moment space and to expand around the rightmost singularity at $N=0$: $$\langle \Delta P\rangle_N = {A\over N} + B + O(N) \Rightarrow \Delta P(x) \approx A + B \delta(1-x) .$$ This procedure yields the following approximate splitting functions[^1]: $$\begin{aligned} \Delta P_{qq}^{(l)} (x) & = & \frac{4}{3} \; \left[ 1 +\frac{1}{2} \delta (1-x) \right] \nonumber \\ \Delta P_{qg}^{(l)} (x) & = & 2n_f \; \frac{1}{2} \; \left[ -1 + 2 \delta (1-x) \right] \nonumber\\ \Delta P_{gq}^{(l)} (x) & = & \frac{4}{3} \; \left[ 2-\delta (1-x) \right] \nonumber\\ \label{bfevo} \Delta P_{gg}^{(l)} (x) & = & 3 \; \left[ 4 - \frac{13}{6} \delta(1-x)\right] -\;\frac{n_f}{3} \delta(1-x)\end{aligned}$$ With these simplified splitting functions, one can analytically solve (\[evo\]) for asymptotic values of $Q^2$ with realistic boundary conditions in the small-$x$ region. This approach is based on the fact that the behaviour of the parton distributions at small $x$ is governed by the region around $N=0$ in moment space. This property can be understood from the $N$-singularity structure of the initial distributions: a logarithmic ($\sim 1/x$) singularity coincides with a pole at $N=0$ in the moment transform, a power-like singularity of the form $x^{\alpha}$ transforms into $\Gamma(\alpha+N)$, which has a singularity at $N = -\alpha$ (see Fig. 1). It is important to notice, however, that the expansion around the $N=0$ pole in moment space agrees with the full splitting function only within a circle of unit radius (Fig. 1). Outside this circle, the series might still be convergent, but its value will be different from that given by the full splitting function. This especially affects the reliability of this approach for low values of $\alpha$. In the extreme case $\alpha$ could approach $-1$ giving rise to a pole close to the boundary of the circle of convergence. In this letter, we examine the validity of analytical approaches to the small-$x$ behaviour of $g_1$. Section 2 contains a study of the evolution matrix on the right-hand side of (\[evo\]). Its properties in the case of power-like ($x^{\alpha}$) boundary conditions are discussed, using the full and the leading-pole expanded splitting functions. By examining the sensitivity of the evolution matrix to the form of the parton distributions in the large-$x$ region, we are able to assess when $x$ can be regarded as small. In Section 3 we present an analytic solution of (\[evo\]), which becomes exact for $x\to 0$. Finally, Section 4 contains some phenomenological implications and conclusions. Study of the evolution matrix ============================= Several qualitative features of the polarised parton densities can already be determined by inserting simple test distributions in the right-hand side of Eq. (\[evo\]). The resulting elements of the evolution matrix determine the local change of the parton densities with increasing $\ln (Q^2)$. Furthermore, for $Q^2/Q_0^2$ not too large one can approximate the solution of the Altarelli-Parisi equations by $$\begin{aligned} \left(\begin{array}{c} \Delta \Sigma \\ \Delta G \end{array} \right) (x,Q^2) & = & \left(\begin{array}{c} \Delta \Sigma \\ \Delta G \end{array} \right) (x,Q_0^2) \nonumber\\ & + & \frac{\alpha_s(Q_0^2)}{2\pi} \int_x^1 \frac{\d y}{y} \left(\begin{array}{cc} \Delta P_{qq} & \Delta P_{qg} \\ \Delta P_{gq} & \Delta P_{gg}\end{array} \right) (y) \left(\begin{array}{c} \Delta \Sigma \\ \Delta G \end{array} \right)(x/y,Q_0^2) \; \ln \left(\frac{Q^2}{Q_0^2}\right).\end{aligned}$$ A realistic choice of test distribution is $$t(x) = x^{\alpha} (1-x)^{\beta}\qquad \mbox{with} \quad (-1<\alpha<0,\;\beta>0),$$ which is similar to the analytic forms of the parton densities at $Q_0^2$ used in recent fits to the polarised structure function data [@bf; @gs]. The exponent $\alpha$ determines the behaviour of the distribution in the small-$x$ regime, whereas the large-$x$ behaviour is controlled by $\beta$. Variations of $\beta$ should therefore not affect any predictions of the small-$x$ behaviour of the parton distributions. This property can be used to define the range of validity of these predictions, i.e. to indicate if $x$ can be regarded as small or not. The elements of the evolution matrix $$A_{ij} = \int_x^1 \frac{\d y}{y}\; \Delta P_{ij} (y) \; t\left(\frac{x}{y} \right)$$ can be computed analytically. The necessary integrals are given in Appendix A. Using the full splitting functions (\[evoful\]), we find[^2] $$\begin{aligned} A_{qq}^{(f)} (x) & = & \frac{4}{3} \left[ 2 A_1(x) -A_2(x) -A_3(x) + \frac{3}{2} A_4(x) \right] \nonumber \\ A_{qg}^{(f)} (x) & = & 3 \left[-A_2(x) + 2 A_3(x) \right] \nonumber \\ A_{gq}^{(f)} (x) & = & \frac{4}{3} \left[ 2 A_2(x) -A_3(x) \right] \nonumber \\ A_{gg}^{(f)} (x) & = & 3 \left[ 2 A_1(x) + 2 A_2(x) - 4 A_3(x) + \frac{3}{2} A_4(x) \right],\end{aligned}$$ while the leading-pole expanded [@bf] splitting functions of (\[bfevo\]) yield $$\begin{aligned} A_{qq}^{(l)} (x) & = & \frac{4}{3} \left[ A_2(x) + \frac{1}{2} A_4(x) \right] \nonumber\\ A_{qg}^{(l)} (x) & = & 3 \left[ -A_2(x) + 2 A_4(x) \right] \nonumber\\ A_{gq}^{(l)} (x) & = & \frac{4}{3} \left[ 2 A_2(x) -A_4(x) \right] \nonumber\\ A_{gg}^{(l)} (x) & = & 3 \left[ 4 A_2(x) - \frac{5}{2} A_4(x) \right].\end{aligned}$$ A closer inspection of the $A_{ij}$ shows that all of them diverge like $x^{\alpha}$ as $x\to 0$. The behaviour in the limit $x \to 0$ can therefore be written as $$\label{limit} \lim_{x \to 0} A_{ij} (x) = a_{ij} x^{\alpha}.$$ Provided that both the initial quark singlet and the initial gluon distributions have power-like boundary conditions in the limit $x\to 0$, these most singular terms will dominate the right-hand side of (\[evo\]). The replacement of the $A_{ij}^{(f)}$ by the above expressions (\[limit\]) in (\[evo\]) should therefore enable us to find an analytic solution for $\Delta \Sigma(x,Q^2)$ and $\Delta G(x,Q^2)$, which becomes exact for $x\to 0$. This exercise will be performed in the following section. The $a_{ij}$ coefficients for the full and the leading-pole expanded splitting functions are [not]{} identical: $$\begin{array}{ll} \displaystyle a_{qq}^{(f)} = \frac{4}{3} \;\left[2(-\psi(-\alpha) - \gamma_E) + \frac{1-2\alpha}{\alpha(1-\alpha)}+ \frac{3}{2} \right] \qquad & \displaystyle a_{qq}^{(l)} = \frac{4}{3} \; \frac{-2+\alpha}{2\alpha} \\ \vspace{3mm} \displaystyle a_{qg}^{(f)} = 3 \; \frac{1+\alpha}{\alpha(1-\alpha)} \qquad & \displaystyle a_{qg}^{(l)} = 3 \; \frac{1+2\alpha}{\alpha} \\ \vspace{3mm} \displaystyle a_{gq}^{(f)} = \frac{4}{3} \;\frac{-2+\alpha}{\alpha(1-\alpha)} \qquad & \displaystyle a_{gq}^{(l)} = \frac{4}{3} \; \frac{-2-\alpha}{\alpha} \\ \vspace{3mm} \displaystyle a_{gg}^{(f)} = 3 \; \left[ 2(-\psi(-\alpha) - \gamma_E) - \frac{2+2\alpha}{\alpha(1-\alpha)} + \frac{3}{2} \right] \qquad & \displaystyle a_{gg}^{(l)} = 3 \; \frac{-8-5\alpha}{2\alpha} \end{array}$$ Here $\psi(x)$ is the usual psi (digamma) function [@AbSteg]. Figure 2 shows the $A_{ij}^{(f)}$ for $\alpha=-0.25,-0.6$ and $\beta=4,9$, together with the approximate forms $A_{ij}^{(l)}$ and the limits $a_{ij}^{(f)}x^{\alpha}$. This figure displays the following important features of the evolution matrix in the small-$x$ region: - Although the test distributions $x^{\alpha}(1-x)^4$ and $x^{\alpha}(1-x)^9$ differ by less than 5% for $x\le 0.01$, the corresponding $A_{ij}^{(f)}$ differ by up to a factor of 2 in the same range. This clearly demonstrates that even at $x=0.01$ and below the evolution is sensitive to the behaviour of the parton distributions in the large-$x$ region. The insensitivity of the $A_{ij}^{(f)}$ to variations of $\beta$ can furthermore be used to define whether $x$ can be regarded as small. For example, by requiring $A_{ij}^{(f)}$ to vary by less than 30% for all combinations in $i$ and $j$ and both values of $\alpha$, we find that only $x\le 0.001$ can be regarded as small, and the more conservative bound of less than 10% deviation yields $x \le 0.0001$. It should therefore be clear that the mere knowledge of $g_1$ at the lowest $x$ values accessible with fixed-target experiments is insufficient to predict the asymptotic behaviour of $g_1$ in the small-$x$ limit, as the behaviour of the parton distributions at these values of $x$ is still closely correlated with the distributions in the large-$x$ region. - The convergence of the $A_{ij}^{(f)}$ towards $a_{ij}^{(f)}x^{\alpha}$ improves for smaller values of $\alpha$. This behaviour just reflects the fact that $A_{ij}^{(f)}$ contains, in addition to this leading term, less singular terms proportional to $\ln(x)$. In general, these lower $\mid \! A_{ij}^{(f)} \! \mid$. If $t(x)$ is less singular that $x^{-1/e}$, the logarithmic terms are larger than the power-like terms for $$x > x_0 (\alpha) = \left( \frac{\omega(\alpha)}{\alpha} \right)^{\frac{1}{\alpha}}$$ where $\omega(\alpha)$ is the branch of Lambert’s $\omega$-function which satisfies $\omega(-1/e)=-1$. As $x_0$ decreases very quickly with $\alpha$ ($x_0 \approx 10^{-15}$ for $\alpha=-0.1$), the replacement $A_{ij}^{(f)}(x) \to a_{ij} x^{\alpha}$, although formally still correct, loses its meaning for values of $\alpha$ close to 0 in any physically relevant region. - While the $A_{ij}^{(l)}$ resemble the $A_{ij}^{(f)}$ for values of $\alpha$ close to 0, they disagree for smaller $\alpha$. This feature becomes most striking for the $A_{qg}$ (see Fig. 2). The full splitting functions [@ap] predict that a positive gluon polarisation in the small-$x$ region will always generate a negative contribution to the sea polarisation. In contrast, the leading-pole expanded splitting functions of [@bf] predict a [positive]{} sea polarisation, if the gluon polarisation $\Delta G(x)$ is more singular than $x^{-0.5}$. This behaviour can be inferred from the $\alpha$ dependence of the $a_{ij}$ displayed in Fig. 3. The good agreement for higher values of $\alpha$ is due to the fact that all leading contributions in $\ln (x)$ are contained in the $N=0$ pole and hence are well approximated by the $A_{ij}^{(l)}$. As elaborated above, these contributions remain important for a finite range in $x > x_0 >0$. The asymptotic predictions of [@bf] will therefore still approximate the full evolution, provided they are restricted to this finite range. - The magnitude of $A_{gg}$ is larger by a factor 3 than the magnitude of all the other terms, but $A_{gg}$ is not more singular than any other contribution. Therefore, the small-$x$ estimate of Ref. [@CR] is quantitative at best, and should be expected to yield a less accurate prediction than the corresponding estimate of the unpolarised distributions. [\|c\|c\|c\|c\|]{} ------------------------------------------------------------------------ & $N\le -2$ & $N=-1$ & $N=0$\ ------------------------------------------------------------------------ $3/4 \;\Delta P_{qq}$ & 2 & 1 & 1\ ------------------------------------------------------------------------ $1/3 \;\Delta P_{qg}$ & 0 & 2 & -1\ ------------------------------------------------------------------------ $3/4 \;\Delta P_{gq}$ & 0 & -1 & 2\ ------------------------------------------------------------------------ $1/3 \;\Delta P_{gg}$ & 2 & -2 & 4\ - The agreement between leading pole expanded and full splitting functions is better for the $A_{gq}$ and $A_{gg}$ than it is for $A_{qq}$ and $A_{qg}$. This feature can be understood from the relative magnitude of the residues in the corresponding splitting functions (Table \[tab:res\]): the $N=0$ residue is dominant only in the $P_{gq}$ and $P_{gg}$ splitting functions, the other two splitting functions contain residues for $N<0$, which are twice as big as the $N=0$ residue. It should be clear from the above that the leading-pole expansion of Ref. [@bf] gives a reliable approximation to the evolution matrix in the small-$x$ region, provided that the initial distributions are significantly less singular than $x^{1/e}$. For more singular distributions, this approach results in a manifestly different evolution matrix and hence will yield a different small-$x$ behaviour of the polarised parton distributions. Solution of the Altarelli-Parisi equations in the limit $x\to 0$ ================================================================ Provided both polarised singlet quark and gluon densities have power-like boundary conditions in the small-$x$ region, $$\Delta \Sigma (x,Q_0^2) \sim x^{\alpha_q} , \quad \Delta G (x,Q_0^2) \sim x^{\alpha_G} \qquad \mbox{with } -1 < \alpha_q,\alpha_G <0 ,$$ one can find a solution of the Altarelli-Parisi equations which becomes exact in the limit $x\rightarrow 0$ and has the form $$\begin{aligned} \Delta q_{val} (x,Q^2) & = & R_v(Q^2,Q_0^2)\; x^{\alpha_v}\nonumber\\ \Delta \Sigma (x,Q^2) &=& R_{qq}(Q^2,Q_0^2) x^{\alpha_q} + R_{qg}(Q^2,Q_0^2) x^{\alpha_G}, \nonumber\\ \Delta G (x,Q^2) &=& R_{gq}(Q^2,Q_0^2) x^{\alpha_q} + R_{gg}(Q^2,Q_0^2) x^{\alpha_G}.\end{aligned}$$ A detailed derivation of this solution and the explicit forms of the $R$-functions is given in Appendix B. The above bounds on $\alpha$ cover the whole theoretically allowed range: as the first moments of the distributions have to be finite, we find $\alpha>-1$. Furthermore, inspection of the singularity structure of the evolution equations (Fig. 1) shows that any initial distribution, which is finite in the small-$x$ region, will develop a logarithmic divergence due to the $N=0$ singularity of the splitting functions. The case of finite or logarithmic boundary conditions can be treated correctly with the leading-pole approximation – its asymptotics are discussed in [@bf]. In a previous analysis [@gs] of the experimental data on polarised structure functions we have found $\alpha_q=\alpha_v\simeq -0.55$. The experimental data used in this analysis were insufficient to determine $\alpha_G$, and therefore it was fixed to be $0$. In contrast to this, more recent measurements at lower $Q^2$ [@slac] favour $\alpha_G<0$. As we have neglected all contributions of order $\ln (x)$ in the above solution, we expect it to be reliable only for $x < x_0(\mbox{max}(\alpha_q, \alpha_G))$. In order to compare this approach with the leading pole expansion of Ref. [@bf] and the numerical solution of (\[evo\]) with the full splitting functions, we have evaluated the distributions for $Q_0^2=4\;\mbox{GeV}^2$ and $Q^2=100\;\mbox{GeV}^2$, using $n_f=3, \Lambda^{QCD}=200 \;\mbox{MeV}$ and the following initial distributions: $$\begin{aligned} \Delta \Sigma(x,Q_0^2) & = & N_q x^{\alpha_q} (1-x)^\beta \nonumber \\ \Delta G(x,Q_0^2) & = & N_G x^{\alpha_G} (1-x)^\beta \\ \Delta q_{val}(x,Q_0^2) & = & N_{val} x^{\alpha_v} (1-x)^\beta .\nonumber\end{aligned}$$ To illustrate the validity of the various approximations, we adopt the following parameter values: $\alpha_q,\alpha_G,\alpha_v = -0.6, -0.25$, $\beta=4,9$, and for simplicity we take $N_q = N_g = N_v = 1$. Figures 4 (a), (b) and (c) show the behaviour of the gluon, singlet quark and valence quark distributions respectively, at small $x$ and $Q^2 = 4,\; 100$ GeV$^2$. The initial distributions $x^{\alpha} (1-x)^4$ are indicated as solid lines. Starting with the gluon distribution (Fig. 4(a)), we see that for $ x < 10^{-2}$, the leading-pole approximation to the splitting funcitons (dotted lines) gives excellent agreement with the full evolution (dashed line), especially for values of $\alpha_q, \alpha_G$ close to 0. This is consistent with the agreement between the corresponding $A_{gg}$ functions shown in Fig. 2 and can be understood from the $N=0$ dominance in the $\Delta P_{gq}$ and $\Delta P_{gg}$ splitting functions. In contrast, the $x^\alpha$ approximation (short-dashed line) significantly overestimates the evolution in the $x$ range shown, espcially for $\alpha_q, \alpha_G$ close to 0. Convergence of this approach can only be observed at even smaller values of $x$. Note, however, the sensitivity to the large-$x$ behaviour. While both the dotted and the dashed lines are computed with $\beta = 4$, the dot-dashed curve corresponds to full splitting function evolution for $\beta = 9$, i.e. a softer large-$x$ distribution. Evidently there is a significant sensitivity to the behaviour at large $x$ even for $x$ values as small as $O(10^{-3})$. This casts doubt on the idea of using data on the evolution of the small-$x$ structure functions alone to determine the gluon distribution. For the singlet quark distribution (Fig. 4(b)) the situation is rather different. Here the leading-pole approximation [overestimates]{} the evolution at small $x$. This is readily understood from the behaviour of the corresponding $A_{qq}$ and $A_{qg}$ functions in Fig. 2, both of which are systematically more positive for the leading-pole splitting functions. In fact we see that for $\alpha_q = -0.25$ and $\alpha_G = -0.6$, the full evolution gives a negative singlet distribution at small $x$, whereas the leading pole splitting functions give a positive distribution. Notice also that the evolution is less sensitive to the large-$x$ behaviour (compare the dashed and dot-dashed curves which correspond to $\beta = 4,9$ respectively) than for the gluon distribution. For $\alpha_q = \alpha_G = -0.6$, the $x^\alpha$ approximation is quite reasonable, and certainly better than the leading-pole approximation. However the opposite is true when both $\alpha_q, \alpha_G$ are close to $0$. Finally, Fig. 4(c) compares the valence quark evolution in the various approximations. This depends only on $\Delta P_{qq}$, and so the behaviour here is a direct reflection of the corresponding $A_{qq}$ shown in Fig. 2. In particular, for $\alpha_q = -0.6$ the $x^\alpha$ approximation is very good, while the leading-pole approximation overestimates the evolution at all $x$ values shown. For less singular small-$x$ behaviour ($\alpha_q = -0.25$), however, both approximations reproduce the full evolution, the leading-pole approximation showing slightly better convergence for $x> O(10^{-4})$. In practice, the normalisations of the singlet quark and gluon distribution, $N_q$ and $N_G$, will not be the same. As the evolution of the gluon densitity is dominated by the gluon-to-gluon splitting, it will be almost unaffected by changes of $N_q$. Only if $N_q$ is one or more orders of magnitude larger than $N_G$, will the impact of quark-to-gluon splitting become visible. More drastic effects of a change in the relative normalisation can be expected for the quark singlet distribution, as contributions from quark-to-quark and gluon-to-quark splitting have the same magnitude but opposite signs (cf. Fig. 3). Therefore, a relative increase of $N_G$ yields a faster evolution of the quark distribution to negative values. The convergence properties of the different analytic approaches are almost unaffected by changes in the normalisation. Only for $N_G \gg N_q$ do we find that convergence of the $x^{\alpha}$ approximation to the singlet distribution sets in for smaller values of $x$. This simply reflects an increased impact of the gluon-to-gluon splitting. Conclusions =========== In this letter we have studied the feasibility of two different analytic approaches to the evolution of polarised parton densities at small $x$, finding that none of these approaches is able to give reliable predictions for the whole theoretically allowed range of boundary conditions in the small-$x$ region. In the leading-pole expansion [@bf; @ber], the full splitting functions $\Delta P_{ij}$ are replaced by the leading terms of their Laurent series around $N=0$. As this approach correctly reproduces all terms proprotional to $\ln x$ generated in the evolution, it is found to be in good agreement with the full evolution if the initial quark and gluon distributions are less singular than $x^{-1/e}$. For more singular boundary conditions, only the gluon distribution is reproduced correctly, in particular the quark distribution is overestimated. Keeping only terms with powerlike singularities in the evolution equation, we were able to derive an exact solution of this equation in the limit $x \to 0$. As we have neglected all logarithmic terms in this approach, its convergence is best for boundary conditions of quark and gluon distributions more singular than $x^{-1/e}$. For less singular boundary conditions, this approach still converges towards the full solution, but its predictions are far away from the full solution for any realistic experimental value of $x$. We have also shown that the evolution of the polarised gluon distribution is sensitive to the shape of this distribution in the large-$x$ region. This observation raises doubts on the possibility of determining the gluon polarisation from the evolution of $g_1$ in the small-$x$ region. It furthermore demonstrates the need for complementary measurements of $\Delta G(x)$ (e.g. from $J/\Psi$-production or direct-$\gamma$ measurements). We have seen that the effects of the evolution on the quark distributions in the small-$x$ region are rather small, as the quark-to-quark and the gluon-to-quark splitting contribute with opposite signs. The gluon distribution is indeed rising with increasing $Q^2$, but only contributes to $g_1$ at order $\alpha_s(Q^2)$. Bearing in mind that $\Delta G$ contributes with a negative coefficient function to $g_1$, one expects that $g_1$ will become negative at small $x$ for asymptotic values of $Q^2$, due to the gluonic contribution and the negative sea polarisation generated from $g \to q \bar q$ splitting. In general, the effects of the evolution on the polarised parton densities will be more moderate than the effects on the unpolarised densities. The assumption of approximate scaling for $g_1(x)/F_1(x)$ in the small-$x$ region is therefore rather doubtful. It seems more realistic to assume approximate scaling for $g_1(x)$ for the range of fixed-target experiments, due to the partial cancellation of quark and gluon evolution as explained above. Acknowledgements {#acknowledgements .unnumbered} ================ Financial support from the UK PPARC (WJS), and from the Gottlieb Daimler- und Karl Benz-Stiftung and the Studienstiftung des deutschen Volkes (TG) is gratefully acknowledged. This work was supported in part by the EU Programme “Human Capital and Mobility”, Network “Physics at High Energy Colliders”, contract CHRX-CT93-0357 (DG 12 COMA). Convolution integrals of the test distribution ============================================== For the test distribution $$t(x) = x^{\alpha} (1-x)^{\beta}\qquad \mbox{with} \quad (-1<\alpha<0,\;\beta>0),$$ the convolution integrals $$A_{ij} = \int_x^1 \frac{\d y}{y}\; \Delta P_{ij} (y) \; t\left(\frac{x}{y} \right)$$ on the right-hand side of (\[evo\]) can be expressed in an analytic form. From the explicit forms of the splitting functions given in (\[evoful\]) and (\[bfevo\]), one sees that the required integrals are $$\begin{aligned} A_1 (x) & = & \int_x^1 \frac {\d y}{y}\; \frac{1}{(1-y)_{+}}\; \left( \frac{x}{y} \right)^{\alpha} \; \left( 1- \frac{x}{y} \right)^{\beta} \nonumber\\ & = & x^{\alpha} (1-x)^{\beta} \big[ \ln (1-x) + \frac{\alpha + \beta +1}{ \beta + 1} (1-x) \;_{3}F_{2} (2+\beta+\alpha,1,1;2,2+\beta;(1-x))\ \nonumber \\ & & -\psi(\beta+1) -\gamma_E\big] \nonumber\\ A_2 (x) & = & \int_x^1 \frac {\d y}{y}\; \left( \frac{x}{y} \right)^{\alpha} \; \left( 1- \frac{x}{y} \right)^{\beta} \nonumber\\ & = & (1-x)^{\beta+1} \frac{1}{\beta+1} \;_{2}F_{1} (1-\alpha,1+\beta;2+\beta;(1-x)) \nonumber\\ A_3 (x) & = & \int_x^1 \frac {\d y}{y}\; y \left( \frac{x}{y} \right)^{\alpha} \; \left( 1- \frac{x}{y} \right)^{\beta} \nonumber\\ & = & x (1-x)^{\beta+1} \left[\frac{1}{1-\alpha} x^{\alpha-1} - \frac{\alpha+\beta}{(1-\alpha)(\beta+1)} \;_{2}F_{1} (1-\alpha,1+\beta;2+\beta;(1-x))\right] \nonumber\\ A_4 (x) & = & \int_x^1 \frac {\d y}{y}\; \delta (1-y) \left( \frac{x}{y} \right)^{\alpha} \; \left( 1- \frac{x}{y} \right)^{\beta} \nonumber\\ & = & x^{\alpha} (1-x)^{\beta}.\end{aligned}$$ All these functions diverge like $x^{\alpha}$ as $x\to 0$, and the leading singular behaviour at small $x$ is found to be $$\begin{aligned} A_1(x) & \smx & x^{\alpha} \left[ -\psi(-\alpha) - \gamma_E \right] \nonumber\\ A_2(x) & \smx & \frac{1}{-\alpha} x^{\alpha} \nonumber\\ A_3(x) & \smx & \frac{1}{1-\alpha} x^{\alpha} \nonumber\\ A_4(x) & \smx & x^{\alpha}.\end{aligned}$$ Analytic solution of the Altarelli-Parisi equations for $x\to 0$ ================================================================= A solution of the Altarelli–Parisi evolution equations (\[evo\]) can never be more singular at $x=0$ than the starting distributions. It follows that the most singular parts of the valence quark, singlet quark and gluon distributions can be obtained by inserting the following ansatz $$\begin{aligned} \Delta q_{val} (x,Q^2) & = & R_v(Q^2,Q_0^2)\; x^{\alpha_v}\nonumber\\ \Delta \Sigma (x,Q^2) & = & R_{qq}(Q^2,Q_0^2) \;x^{\alpha_q} + R_{qg}(Q^2,Q_0^2) \; x^{\alpha_G}, \nonumber\\ \Delta G (x,Q^2) & = & R_{gq}(Q^2,Q_0^2) \;x^{\alpha_q} + R_{gg}(Q^2,Q_0^2)\; x^{\alpha_G}\end{aligned}$$ into (\[evo\]). Keeping only terms proportional to $x^{\alpha_v}$, $x^{\alpha_q}$ and $x^{\alpha_G}$ on the right-hand side, we obtain the following evolution equations for the $R$ coefficients ($\beta_0 = 11 - 2/3 n_f$) : $$\begin{aligned} \frac{\partial}{\partial \ln \alpha_s} R_v (Q^2,Q_0^2) & = & - \; \frac{2}{\beta_0} a_{qq}(\alpha_v) R_v (Q^2,Q_0^2) \nonumber \\ \frac{\partial}{\partial \ln \alpha_s} \left(\begin{array}{c} R_{qq} \\ R_{gq} \end{array} \right) (Q^2,Q_0^2) & = & - \; \frac{2}{\beta_0} \left(\begin{array}{cc} a_{qq} (\alpha_q) & a_{qg} (\alpha_q) \\ a_{gq} (\alpha_q)& a_{gg} (\alpha_q) \end{array} \right)\; \left(\begin{array}{c} R_{qq} \\ R_{gq} \end{array} \right)(Q^2,Q_0^2)\nonumber \\ \frac{\partial}{\partial \ln \alpha_s} \left(\begin{array}{c} R_{qg} \\ R_{gg} \end{array} \right) (Q^2,Q_0^2) & = & - \; \frac{2}{\beta_0} \left(\begin{array}{cc} a_{qq} (\alpha_G) & a_{qg} (\alpha_G) \\ a_{gq} (\alpha_G)& a_{gg} (\alpha_G) \end{array} \right)\; \left(\begin{array}{c} R_{qg} \\ R_{gg}\end{array} \right)(Q^2,Q_0^2).\end{aligned}$$ As we are interested in the asymptotic solution for the [full]{} splitting functions, all $a_{ij}$ in the above are $a_{ij}^{(f)}$. Introducing $$\begin{aligned} s & = & \ln \left( \frac{\displaystyle \ln (Q^2/\Lambda^2)}{\displaystyle \ln (Q_0^2/\Lambda^2)}\right) \nonumber\\ \omega_{\pm} (\alpha) & = & \frac{1}{2} \left( a_{qq} (\alpha) + a_{gg} (\alpha) \pm \sqrt{ \left( a_{qq} (\alpha) -a_{gg} (\alpha) \right)^2 + 4 a_{gq} (\alpha) a_{qg} (\alpha)}\right),\end{aligned}$$ the general solution of these equations reads $$\begin{aligned} R_v(Q^2,Q_0^2) & = & N_v \;\exp \left\{ \frac{2}{\beta_0} a_{qq} (\alpha_v) s \right\} \nonumber\\ R_{qq}(Q^2,Q_0^2) & = & R_{qq+} (Q_0^2) \;\exp \left\{ \frac{2}{\beta_0} \omega_{+} (\alpha_q) s \right\} + R_{qq-} (Q_0^2) \;\exp \left\{ \frac{2}{\beta_0} \omega_{-} (\alpha_q) s \right\} \nonumber\\ R_{gq}(Q^2,Q_0^2) & = & R_{gq+} (Q_0^2) \;\exp \left\{ \frac{2}{\beta_0} \omega_{+} (\alpha_q) s \right\} + R_{gq-} (Q_0^2) \;\exp \left\{ \frac{2}{\beta_0} \omega_{-} (\alpha_q) s \right\} \nonumber\\ R_{qg}(Q^2,Q_0^2) & = & R_{qg+} (Q_0^2) \;\exp \left\{ \frac{2}{\beta_0} \omega_{+} (\alpha_G) s \right\} + R_{qg-} (Q_0^2) \;\exp \left\{ \frac{2}{\beta_0} \omega_{-} (\alpha_G) s \right\} \nonumber\\ R_{gg}(Q^2,Q_0^2) & = & R_{gg+} (Q_0^2) \;\exp \left\{ \frac{2}{\beta_0} \omega_{+} (\alpha_G) s \right\} + R_{gg-} (Q_0^2) \;\exp \left\{ \frac{2}{\beta_0} \omega_{-} (\alpha_G) s \right\},\end{aligned}$$ where the $R_{ij\pm} (Q_0^2)$ are determined by the boundary conditions at $Q_0^2$. As we assume that the initial distributions for the quark singlet and the gluon have the form $$\Delta \Sigma (x,Q_0^2) = N_q \; x^{\alpha_q} , \quad \Delta G (x,Q_0^2) = N_G \;x^{\alpha_G},$$ these constants are determined to be $$\begin{array}{ll} \displaystyle R_{qq+} (Q_0^2) = \frac{\displaystyle \omega_{+} (\alpha_q) - a_{gg}(\alpha_q)}{\displaystyle \omega_{+} (\alpha_q) - \omega_{-} (\alpha_q)} N_{q}, & R_{qq-} (Q_0^2) = - \; \frac{\displaystyle \omega_{-} (\alpha_q) - a_{gg}(\alpha_q)} {\displaystyle \omega_{+} (\alpha_q) - \omega_{-} (\alpha_q)} N_{q},\\ \vspace{3mm} R_{gq+} (Q_0^2) = \frac{\displaystyle a_{gq} (\alpha_q)} {\displaystyle \omega_{+} (\alpha_q) - \omega_{-} (\alpha_q)} N_{q}, & R_{gq-} (Q_0^2) = - \; \frac{\displaystyle a_{gq} (\alpha_q)} {\displaystyle \omega_{+} (\alpha_q) - \omega_{-} (\alpha_q)} N_{q},\\ \vspace{3mm} R_{qg+} (Q_0^2) = \frac{\displaystyle a_{qg} (\alpha_G)} {\displaystyle \omega_{+} (\alpha_G) - \omega_{-} (\alpha_G)} N_{g}, & R_{qg-} (Q_0^2) = - \; \frac{\displaystyle a_{qg} (\alpha_G)} {\displaystyle \omega_{+} (\alpha_G) - \omega_{-} (\alpha_G)} N_{g},\\ \vspace{3mm} R_{gg+} (Q_0^2) = \frac{\displaystyle \omega_{+} (\alpha_G) - a_{qq}(\alpha_G)}{\displaystyle \omega_{+} (\alpha_G) - \omega_{-} (\alpha_G)} N_{g}, & R_{gg-} (Q_0^2) = - \; \frac{\displaystyle \omega_{-} (\alpha_G) - a_{qq}(\alpha_G)} {\displaystyle \omega_{+} (\alpha_G) - \omega_{-} (\alpha_G)} N_{g}. \end{array}$$ [10]{} J. Ellis and R.L. Jaffe, Phys. Rev. [D9]{} (1974) 1444, erratum [D10]{} (1974) 1669. SLAC-Yale collaboration: M.J. Alguard et al, Phys. Rev. Lett. [37]{} (1976) 1261; G. Baum et al., Phys. Rev. Lett. [45]{} (1980) 2000; [51]{} (1983) 1135.\ EMC collaboration: J. Ashman et al., Nucl. Phys. [B328]{} (1989) 1.\ SLAC-E142 collaboration: D.L. Anthony et al., Phys. Rev. Lett. [71]{} (1993) 959.\ SMC collaboration: B. Adeva et al., Phys. Lett. [B302]{} (1993) 553, D. Adams et al., Phys. Lett. [B329]{} (1994) 399. SLAC-E143 collaboration: K. Abe et al, Phys. Rev. Lett. [74]{} (1995) 346; [75]{} (1995) 25. F.E. Close and R.G. Roberts, Phys. Lett. [B336]{} (1994) 257. S.D. Bass and P.V. Landshoff, Phys. Lett. [B336]{} (1994) 537. G. Altarelli and G. Parisi, Nucl. Phys. [B126]{} (1977) 298. R.D. Ball, S. Forte and G. Ridolfi, Nucl. Phys. [B444]{} (1995) 287. A. Berera, Phys. Lett. [B293]{} (1992) 445. T. Gehrmann and W.J. Stirling, Z. Phys. [C65]{} (1995) 461. See for example M. Abramowitz and I.A. Stegun, [Handbook of Mathematical Functions]{}, National Bureau of Standards Applied Mathematics Series 55 (1964). Figure Captions {#figure-captions .unnumbered} --------------- Figure 1 : Singularity structure of the evolution equations in the complex $N$-moment plane. Dots ($\bullet$) denote the poles of the splitting functions, and the cross ($\times$) indicates the small-$x$ singularity of the initial distribution. The leading-pole expansion only converges to the splitting function in the unit circle around the origin. Figure 2 : Elements of the splitting matrix for the test distribution $x^{\alpha}(1-x)^{\beta}$. Solid line: full splitting functions for $\beta=4$, long-dashed line: same for $\beta=9$, short-dashed line: most singular $x^{\alpha}$ contribution, dotted line: leading-pole expanded splitting functions for $\beta=4$, dot-dashed line: same for $\beta=9$. For better visibility, all elements are multiplied by $x$. Figure 3 : Coefficients of the most singular pieces in the splitting matrix for the full (left) and the leading-pole expanded (right) splitting functions. Solid line: $a_{qq}$, long-dashed line: $a_{qg}$, short-dashed line: $a_{gq}$, dotted line: $a_{gg}$. Figure 4 : Evolution of test distributions for gluons ($x\Delta G(x,Q^2)$),(a), singlet quarks ($x\Delta \Sigma (x,Q^2)$),(b) and valence quarks ($x\Delta q_{val} (x,Q^2)$),(c) as described in the text. Solid line: starting distribution at $4 \; \mbox{GeV}^2$, long-dashed and dot-dashed line: evolved distributions at $100 \; \mbox{GeV}^2$ for different large-$x$ behaviour at $Q_0^2$, short dashed line: result of $x^{\alpha}$ approximaton, dotted line: result of leading-pole approximation. [^1]: Similar splitting functions containing only the residue at $N=0$ were studied in , giving qualitatively comparable results to [^2]: For simplicity, we take $n_f=3$ throughout this study	{ "pile_set_name": "ArXiv" }
--- abstract: 'We have computed wind models with time-dependent dust formation and grain-size dependent opacities, where (1) the problem is simplified by assuming a fixed dust-grain size, and where (2) the radiation pressure efficiency is approximated using grain sizes based on various means of the actual grain size distribution. It is shown that in critical cases, the effect of grain sizes can be significant. For well-developed winds, however, the effects on the mass-loss rate and the wind speed are small.' author: - 'Lars Mattsson, and Susanne Höfner' title: 'Relaxing the Small Particle Approximation for Dust-grain opacities in Carbon-star Wind Models' --- [Introduction.]{} In a new grid of C-star wind models Mattsson et al. (2010, A&A, 509, 14) used the so-called [small-particle limit]{} (SPL) of Mie theory, i.e., grains are assumed to be small compared to the photon wavelengths when calculating their optical properties. In that limit the wavelength- and grain size-dependence of the opacity can be greatly simplified. Fig. 1 shows the distribution of actual grain sizes found in a representative subsample of models taken from the model grid (presented in Mattsson et al. 2010, A&A, 509, 14). The maximum grain size $a_{\rm d}$ (the “dividing radius” between the SPL and non-SPL regemes) where deviations in opacities from the small particle limit are less than 10% is marked by the vertical dashed line. The vast majority of the grid models did not meet this criterion (see Fig. 1). [Models.]{} We have re-computated a selection of models adopting different relaxations of the small-particle approximation (SPA) and picked models such that two categories emerge: those with strong well-developed winds and those with slow critical winds (Mattsson & Höfner 2010, A&A, submitted). We use an “optimised” (OPT) constant grain radius $a_{\rm gr} = 3.55\cdot 10^{-5}$ cm for all grains (the peak in $Q_{\rm rp}$, see Fig.1). We also consider mean grain radii derived from one of the moments $K_i$ ($i = 1,2,3$) of the actual grain-size distribution as effective grain sizes used in the raditive transfer calculations. [Results and Conclusions.]{} In the critical-wind cases, the effect of grain sizes can be significant. Mass-loss rates may increase by a factor of two, or more, and wind speeds by as much as an order of magnitude (see Fig. 2, open symbols). Furthermore, the corresponding models with grain-size dependent opacities that have resultant winds tend to have much lower degrees of dust condensation, compared to their SPA counterparts (again, see Fig. 2). Consequently, the “dust-loss rates” are much lower in these new models. In contrast, for well-developed dust-driven winds (Fig. 2, filled symbols), where the dust formation has saturated, the effect of grain sizes on mass-loss rate, wind speed and dust-to-gas ratio is almost negligible. This work was partly supported by the Swedish Research Council (Vetenskapsrådet). ![Left: Histogram of grain sizes for wind-forming models with $M_\star=1M_\odot$, $\Delta u_{\rm p} = 4$ km s$^{-1}$ taken from Mattsson et al. (2010, A&A, 509, 14). The vertical dashed line marks the dividing radius $a_{\rm d}$. Right: The radiative pressure efficiency factor $Q_{\rm rp}$ and its components $Q_{\rm ext}$, $Q_{\rm abs}$ and $Q_{\rm sca}$, relative to the corresponding SPL values used in the SPA models, as functions of grain radius at $\lambda=1\mu$m. ](mattsson_f1a.eps "fig:"){width="35mm"} ![Left: Histogram of grain sizes for wind-forming models with $M_\star=1M_\odot$, $\Delta u_{\rm p} = 4$ km s$^{-1}$ taken from Mattsson et al. (2010, A&A, 509, 14). The vertical dashed line marks the dividing radius $a_{\rm d}$. Right: The radiative pressure efficiency factor $Q_{\rm rp}$ and its components $Q_{\rm ext}$, $Q_{\rm abs}$ and $Q_{\rm sca}$, relative to the corresponding SPL values used in the SPA models, as functions of grain radius at $\lambda=1\mu$m. ](mattsson_f1b.eps "fig:"){width="35mm"} ![From left to right: Mass-loss rates, wind speeds, mean degrees of dust condensation and dust-to-gas ratios for “$Q_{\rm rp}$-optimised” and K1, K2, K3 models versus the corresponding SPA models. The dashed lines show the case of equal values. \[agr\_spa\_Kopt\] ](mattsson_f2a "fig:"){width="28mm"} ![From left to right: Mass-loss rates, wind speeds, mean degrees of dust condensation and dust-to-gas ratios for “$Q_{\rm rp}$-optimised” and K1, K2, K3 models versus the corresponding SPA models. The dashed lines show the case of equal values. \[agr\_spa\_Kopt\] ](mattsson_f2b "fig:"){width="28mm"} ![From left to right: Mass-loss rates, wind speeds, mean degrees of dust condensation and dust-to-gas ratios for “$Q_{\rm rp}$-optimised” and K1, K2, K3 models versus the corresponding SPA models. The dashed lines show the case of equal values. \[agr\_spa\_Kopt\] ](mattsson_f2c "fig:"){width="28mm"} ![From left to right: Mass-loss rates, wind speeds, mean degrees of dust condensation and dust-to-gas ratios for “$Q_{\rm rp}$-optimised” and K1, K2, K3 models versus the corresponding SPA models. The dashed lines show the case of equal values. \[agr\_spa\_Kopt\] ](mattsson_f2d "fig:"){width="28mm"}\ ![From left to right: Mass-loss rates, wind speeds, mean degrees of dust condensation and dust-to-gas ratios for “$Q_{\rm rp}$-optimised” and K1, K2, K3 models versus the corresponding SPA models. The dashed lines show the case of equal values. \[agr\_spa\_Kopt\] ](mattsson_f2e "fig:"){width="28mm"} ![From left to right: Mass-loss rates, wind speeds, mean degrees of dust condensation and dust-to-gas ratios for “$Q_{\rm rp}$-optimised” and K1, K2, K3 models versus the corresponding SPA models. The dashed lines show the case of equal values. \[agr\_spa\_Kopt\] ](mattsson_f2f "fig:"){width="28mm"} ![From left to right: Mass-loss rates, wind speeds, mean degrees of dust condensation and dust-to-gas ratios for “$Q_{\rm rp}$-optimised” and K1, K2, K3 models versus the corresponding SPA models. The dashed lines show the case of equal values. \[agr\_spa\_Kopt\] ](mattsson_f2g "fig:"){width="28mm"} ![From left to right: Mass-loss rates, wind speeds, mean degrees of dust condensation and dust-to-gas ratios for “$Q_{\rm rp}$-optimised” and K1, K2, K3 models versus the corresponding SPA models. The dashed lines show the case of equal values. \[agr\_spa\_Kopt\] ](mattsson_f2h "fig:"){width="28mm"}\ ![From left to right: Mass-loss rates, wind speeds, mean degrees of dust condensation and dust-to-gas ratios for “$Q_{\rm rp}$-optimised” and K1, K2, K3 models versus the corresponding SPA models. The dashed lines show the case of equal values. \[agr\_spa\_Kopt\] ](mattsson_f2i "fig:"){width="28mm"} ![From left to right: Mass-loss rates, wind speeds, mean degrees of dust condensation and dust-to-gas ratios for “$Q_{\rm rp}$-optimised” and K1, K2, K3 models versus the corresponding SPA models. The dashed lines show the case of equal values. \[agr\_spa\_Kopt\] ](mattsson_f2j "fig:"){width="28mm"} ![From left to right: Mass-loss rates, wind speeds, mean degrees of dust condensation and dust-to-gas ratios for “$Q_{\rm rp}$-optimised” and K1, K2, K3 models versus the corresponding SPA models. The dashed lines show the case of equal values. \[agr\_spa\_Kopt\] ](mattsson_f2k "fig:"){width="28mm"} ![From left to right: Mass-loss rates, wind speeds, mean degrees of dust condensation and dust-to-gas ratios for “$Q_{\rm rp}$-optimised” and K1, K2, K3 models versus the corresponding SPA models. The dashed lines show the case of equal values. \[agr\_spa\_Kopt\] ](mattsson_f2l "fig:"){width="28mm"}\ ![From left to right: Mass-loss rates, wind speeds, mean degrees of dust condensation and dust-to-gas ratios for “$Q_{\rm rp}$-optimised” and K1, K2, K3 models versus the corresponding SPA models. The dashed lines show the case of equal values. \[agr\_spa\_Kopt\] ](mattsson_f2m "fig:"){width="28mm"} ![From left to right: Mass-loss rates, wind speeds, mean degrees of dust condensation and dust-to-gas ratios for “$Q_{\rm rp}$-optimised” and K1, K2, K3 models versus the corresponding SPA models. The dashed lines show the case of equal values. \[agr\_spa\_Kopt\] ](mattsson_f2n "fig:"){width="28mm"} ![From left to right: Mass-loss rates, wind speeds, mean degrees of dust condensation and dust-to-gas ratios for “$Q_{\rm rp}$-optimised” and K1, K2, K3 models versus the corresponding SPA models. The dashed lines show the case of equal values. \[agr\_spa\_Kopt\] ](mattsson_f2o "fig:"){width="28mm"} ![From left to right: Mass-loss rates, wind speeds, mean degrees of dust condensation and dust-to-gas ratios for “$Q_{\rm rp}$-optimised” and K1, K2, K3 models versus the corresponding SPA models. The dashed lines show the case of equal values. \[agr\_spa\_Kopt\] ](mattsson_f2p "fig:"){width="28mm"}	{ "pile_set_name": "ArXiv" }
--- abstract: 'Charged particle production in deep-inelastic $ep$ scattering is measured with the H1 detector at HERA. The kinematic range of the analysis covers low photon virtualities, , and small values of Bjorken-$x$, $10^{-4} < x < 10^{-2}$. The analysis is performed in the hadronic centre-of-mass system. The charged particle densities are measured as a function of pseudorapidity ($\eta^$) and transverse momentum ($p_T^$) in the range $0<\eta^<5$ and $0<p_T^ < 10$ GeV differentially in $x$ and $Q^2$. The data are compared to predictions from different Monte Carlo generators implementing various options for hadronisation and parton evolutions.' --- \#1\#2\#3\#4[[\#1]{} [\#2]{} (\#3) \#4]{} [DESY 13-012 ISSN 0418-9833]{}\ [January 2013]{}\ [Measurement of Charged Particle Spectra in Deep-Inelastic $ep$ Scattering at HERA]{} H1 Collaboration Submitted to C. Alexa$^{5}$, V. Andreev$^{25}$, A. Baghdasaryan$^{37}$, S. Baghdasaryan$^{37}$, W. Bartel$^{11}$, K. Begzsuren$^{34}$, A. Belousov$^{25}$, P. Belov$^{11}$, V. Boudry$^{28}$, I. Bozovic-Jelisavcic$^{2}$, G. Brandt$^{49}$, M. Brinkmann$^{11}$, V. Brisson$^{27}$, D. Britzger$^{11}$, A. Buniatyan$^{14}$, A. Bylinkin$^{24,46}$, L. Bystritskaya$^{24}$, A.J. Campbell$^{11}$, K.B. Cantun Avila$^{22}$, F. Ceccopieri$^{4}$, K. Cerny$^{31}$, V. Chekelian$^{26}$, J.G. Contreras$^{22}$, J. Cvach$^{30}$, J.B. Dainton$^{18}$, K. Daum$^{36,41}$, E.A. De Wolf$^{4}$, C. Diaconu$^{21}$, M. Dobre$^{5}$, V. Dodonov$^{13}$, A. Dossanov$^{12,26}$, G. Eckerlin$^{11}$, S. Egli$^{35}$, E. Elsen$^{11}$, L. Favart$^{4}$, A. Fedotov$^{24}$, R. Felst$^{11}$, J. Feltesse$^{10}$, J. Ferencei$^{16}$, D.-J. Fischer$^{11}$, M. Fleischer$^{11}$, A. Fomenko$^{25}$, E. Gabathuler$^{18}$, J. Gayler$^{11}$, S. Ghazaryan$^{11}$, A. Glazov$^{11}$, L. Goerlich$^{7}$, N. Gogitidze$^{25}$, M. Gouzevitch$^{11,42}$, C. Grab$^{39}$, A. Grebenyuk$^{11}$, T. Greenshaw$^{18}$, G. Grindhammer$^{26}$, S. Habib$^{11}$, D. Haidt$^{11}$, R.C.W. Henderson$^{17}$, E. Hennekemper$^{15}$, M. Herbst$^{15}$, G. Herrera$^{23}$, M. Hildebrandt$^{35}$, K.H. Hiller$^{38}$, J. Hladky$^{30}$, D. Hoffmann$^{21}$, R. Horisberger$^{35}$, T. Hreus$^{4}$, F. Huber$^{14}$, M. Jacquet$^{27}$, X. Janssen$^{4}$, L. Jönsson$^{20}$, H. Jung$^{11,4}$, M. Kapichine$^{9}$, C. Kiesling$^{26}$, M. Klein$^{18}$, C. Kleinwort$^{11}$, R. Kogler$^{12}$, P. Kostka$^{38}$, M. Krämer$^{11}$, J. Kretzschmar$^{18}$, K. Krüger$^{11}$, M.P.J. Landon$^{19}$, W. Lange$^{38}$, P. Laycock$^{18}$, A. Lebedev$^{25}$, S. Levonian$^{11}$, K. Lipka$^{11,45}$, B. List$^{11}$, J. List$^{11}$, B. Lobodzinski$^{11}$, R. Lopez-Fernandez$^{23}$, V. Lubimov$^{24, \dagger}$, E. Malinovski$^{25}$, H.-U. Martyn$^{1}$, S.J. Maxfield$^{18}$, A. Mehta$^{18}$, A.B. Meyer$^{11}$, H. Meyer$^{36}$, J. Meyer$^{11}$, S. Mikocki$^{7}$, I. Milcewicz-Mika$^{7}$, A. Morozov$^{9}$, J.V. Morris$^{6}$, K. Müller$^{40}$, Th. Naumann$^{38}$, P.R. Newman$^{3}$, C. Niebuhr$^{11}$, D. Nikitin$^{9}$, G. Nowak$^{7}$, K. Nowak$^{12,45}$, J.E. Olsson$^{11}$, D. Ozerov$^{11}$, P. Pahl$^{11}$, V. Palichik$^{9}$, M. Pandurovic$^{2}$, C. Pascaud$^{27}$, G.D. Patel$^{18}$, E. Perez$^{10,43}$, A. Petrukhin$^{11}$, I. Picuric$^{29}$, H. Pirumov$^{14}$, D. Pitzl$^{11}$, R. Plačakytė$^{11,45}$, B. Pokorny$^{31}$, R. Polifka$^{31,47}$, V. Radescu$^{11,45}$, N. Raicevic$^{29}$, T. Ravdandorj$^{34}$, P. Reimer$^{30}$, E. Rizvi$^{19}$, P. Robmann$^{40}$, R. Roosen$^{4}$, A. Rostovtsev$^{24}$, M. Rotaru$^{5}$, J.E. Ruiz Tabasco$^{22}$, S. Rusakov$^{25}$, D. Šálek$^{31}$, D.P.C. Sankey$^{6}$, M. Sauter$^{14}$, E. Sauvan$^{21,48}$, S. Schmitt$^{11}$, L. Schoeffel$^{10}$, A. Schöning$^{14}$, H.-C. Schultz-Coulon$^{15}$, F. Sefkow$^{11}$, S. Shushkevich$^{11}$, Y. Soloviev$^{11,25}$, P. Sopicki$^{7}$, D. South$^{11}$, V. Spaskov$^{9}$, A. Specka$^{28}$, Z. Staykova$^{4}$, M. Steder$^{11}$, B. Stella$^{32}$, G. Stoicea$^{5}$, U. Straumann$^{40}$, T. Sykora$^{4,31}$, P.D. Thompson$^{3}$, D. Traynor$^{19}$, P. Truöl$^{40}$, I. Tsakov$^{33}$, B. Tseepeldorj$^{34,44}$, J. Turnau$^{7}$, A. Valkárová$^{31}$, C. Vallée$^{21}$, P. Van Mechelen$^{4}$, Y. Vazdik$^{25}$, D. Wegener$^{8}$, E. Wünsch$^{11}$, J. Žáček$^{31}$, J. Zálešák$^{30}$, Z. Zhang$^{27}$, R. Žlebčík$^{31}$, H. Zohrabyan$^{37}$, and F. Zomer$^{27}$ Introduction ============ Deep-inelastic scattering (DIS) processes at the $ep$ collider HERA can access small values of Bjorken-$x$ at low four momentum transfers squared $Q^2$ of a few GeV$^2$. In the region of low $x$, characterised by high densities of gluons and sea quarks in the proton, the parton interaction with the virtual photon may originate from a cascade of partons emitted prior to the interaction as illustrated in . In perturbative Quantum Chromodynamics (QCD) such multi-parton emissions are described only within certain approximations valid in restricted phase space regions. At sufficiently large $Q^2$ and not too small $x$ the Dokshitzer-Gribov-Lipatov-Altarelli-Parisi (DGLAP) [@DGLAP] evolution equation is expected to be a good approximation. The DGLAP equation corresponds to a strong ordering of the transverse momenta of the propagator partons, $k_{T,i}$ , with respect to the proton direction, which implies strong ordering of the transverse momenta of the emitted partons, $p_{T,i} \ll p_{T,i+1}$, in the parton cascade from the proton towards the virtual photon. At small $x$ the DGLAP approximation is expected to become inadequate and the Balitsky-Fadin-Kuraev-Lipatov (BFKL) [@BFKL] scheme may be more appropriate, which has no ordering in $k_{T}$ of the partons along the ladder. The Ciafaloni-Catani-Fiorani-Marchesini (CCFM) evolution [@CCFM] aims to unify the DGLAP and BFKL approaches. It introduces angular ordering of gluon emissions to implement coherence effects. At small $x$ the CCFM evolution equation is almost equivalent to the BFKL approach, while it reproduces the DGLAP equations for sufficiently large $x$ and $Q^2$. Measurements of the proton structure function $F_2 (x, Q^2)$ [@F2] are well described by the Next-to-Leading-Order (NLO) or Next-to-Next-to-Leading Order (NNLO) DGLAP evolution [@F2-theory-MRST; @F2-theory-CT; @F2-theory-NNPDF; @F2-theory-ABM], suggesting that this observable may be too inclusive to exhibit signals for BFKL dynamics. Deviations from the $k_T$ ordering at HERA are observed in jet production [@dijet; @forw-incl-jet-1], transverse energy flow [@tr-energy-flow-1; @tr-energy-flow-2], forward jet production [@forw-jet-h1-zeus1; @forw-jet-h1-zeus2; @forw-jet-h1-zeus3] and measurements of forward $\pi^0$ production [@fwd-pi]. Studies of the transverse momentum spectrum of charged particles have been proposed in [@Kuhlen-pt] as a more direct probe of the underlying parton dynamics. In that paper it has been shown with the help of QCD models that the high-$p_T$ tail is sensitive to parton radiation, while the contribution from hadronisation is small. Previous measurements of the transverse momentum and pseudorapidity dependence of particle densities performed by the H1 collaboration [@pt-old] were limited in statistical precision. This paper presents a study of charged particle production in $ep$ collisions for . The analysis is performed in the hadronic centre-of-mass frame, i.e. in the virtual photon-proton rest frame. The charged particle densities as a function of pseudorapidity and transverse momentum are measured differentially in $x$ and $Q^2$. The data set used for the analysis was collected with the H1 detector in 2006 when positrons and protons collided with energies of $27.6$ GeV and $920$ GeV, respectively, corresponding to a centre-of-mass energy of $\sqrt{s}=319$ GeV. The integrated luminosity of the data set is $88.6$ pb$^{-1}$, which is about seventy times larger than what was available for the previous H1 measurement [@pt-old]. This allows for a more detailed study of the dynamical features of parton evolution in the proton at small $x$. QCD models {#sec-QCDmodels} ========== Parton cascade and hadronisation processes leading to charged particle production in $ep$ collisions are modeled using different Monte Carlo (MC) programs. Brief descriptions of the MC event generators considered in this analysis are given below. - The [Rapgap]{} [@RAPGAP] MC generator matches first order QCD matrix elements to DGLAP based leading logarithm approximations for parton showers with strongly ordered transverse momenta of subsequently emitted partons. The factorisation and renormalisation scales are set to $\mu_f=\mu_r=\sqrt{Q^2+\hat{p_T}^2}$, where $\hat{p}_T$ is the transverse momentum of the outgoing hard parton from the matrix element in the centre-of-mass frame of the hard subsystem. - The [Djangoh]{} [@DJANGOH] MC generator uses the Colour Dipole Model (CDM) as implemented in [ARIADNE]{} [@ARIADNE], which models first order QCD processes and creates dipoles between coloured partons. Gluon emission is treated as radiation from these dipoles, and new dipoles are formed from the emitted gluons from which further radiation is possible. The radiation pattern of the dipoles includes interference effects, thus modelling gluon coherence. The transverse momenta of the emitted partons are not ordered, producing a configuration similar to the BFKL treatment of parton evolution [@CDM-BFKL]. - The [Cascade]{} [@CASCADE] MC generator uses off-shell leading order QCD matrix elements, supplemented with parton emissions based on the CCFM evolution equation. The equation requires an unintegrated gluon density (see [@CASCADE]), which takes the transverse momenta of the propagators into account. In contrast to the DGLAP evolution equation, the CCFM equation only contains gluon splitting $g \to g g$. - The Herwig++ [@herwig] MC program with the POWHEG (POsitive Weight Hardest Emission) option [@POWHEG-DIS], combines the full matrix element including virtual corrections at $\mathcal{O}(\alpha_s)$ with a DGLAP-like parton shower simulation. The Herwig++ MC program uses the Coherent Parton Branching algorithm [@CoherentPB1; @CoherentPB2] which is based on colour coherence to suppress branchings outside an angular-ordered region of phase space. Here, final state radiation is angular ordered and initial state radiation is ordered in $E\cdot \theta \approx p_T$, where $E$ and $\theta$ are the energy and polar angle of the radiated parton, respectively. - Photoproduction background is generated with the [Phojet]{} [@Phojet] program, which uses a two-component dual parton model [@Phojet1] including diffractive processes and vector meson production. The [Rapgap]{} and [Djangoh]{} predictions are calculated using the CTEQ6L(LO) [@cteq6lo] set of Parton Distribution Functions (PDF), while [Cascade]{} uses the default A0 unintegrated gluon density set [@A0]. The predictions of Herwig++ were obtained with the default PDF MRST 02 NLO [@MRST]. To simulate hadronisation the Lund string fragmentation model [@Lund] is used, as implemented in JETSET [@JETSETandPYTHIA] for [Djangoh]{} and [Pythia]{} [@PYTHIA] for both [Rapgap]{} and [Cascade]{}. The parameters of the Lund string fragmentation model used here are tuned to describe $e^+e^-$ results [@ALEPH]. The tuning was performed by the ALEPH collaboration using hadronic $Z$ decay data and the [Pythia]{}6.1 simulation with Bose-Einstein correlations turned on. In addition, the tune obtained by the Professor tool [@profftune] using LEP data is also tested. Herwig++ incorporates the cluster model [@clustermodel1] of hadronisation, in which colour-singlet clusters of partons form after the perturbative phase and then decay into the observed hadrons. [Djangoh]{} and [Rapgap]{} are also used together with the H1 detector simulation in order to determine the acceptance and efficiency and to estimate the systematic uncertainties associated with the measurement. The programs are interfaced to [Heracles]{} [@HERACLES] to simulate the QED-radiative effects. The generated events are passed through a detailed simulation of the H1 detector response based on the [Geant]{} simulation program [@GEANT] and are processed using the same reconstruction and analysis program chain as for data. For the determination of the detector effects both the [Rapgap]{} and [Djangoh]{} predictions are studied. Both models describe all relevant control distributions reasonably well [@myThesis]. To improve the determination of the detector corrections the transverse momentum and pseudorapidity of charged particles as well as inelasticity $y$, defined as $y=Q^2/(s\cdot x)$, are reweighted to the data [@myThesis]. The reweighting is applied to the generator quantities. Experimental method =================== H1 detector ----------- A full description of the H1 detector can be found elsewhere [@H1det1; @H1det2; @SPACAL] and only the components most relevant for this analysis are briefly mentioned here. The coordinate system of H1 is defined such that the positive $z$ axis is pointing in the direction of the proton beam (forward direction) and the nominal interaction point is located at $z = 0$. The polar angle $\theta$ is then defined with respect to this axis. The pseudorapidity is defined to be $\eta = - \ln (\tan(\theta/2))$. Charged particles are measured within the central tracking detector (CTD) in the polar angle range , which is also used to reconstruct the interaction vertex. The CTD comprises two large cylindrical jet chambers (CJCs), and the silicon vertex detector [@CST34; @CST35]. The CTD is operated inside a $1.16$ T solenoidal magnetic field. The CJCs are separated by a cylindrical drift chamber which improves the $z$ coordinate reconstruction. A cylindrical multiwire proportional chamber [@CIP], which is mainly used in the trigger, is situated inside the inner CJC. The trajectories of charged particles are measured with a transverse momentum resolution of $\sigma (p_T)/p_T \approx 0.2\% p_T/$GeV$\oplus 1.5\%$. The forward tracking detector (FTD) [@FTDJINST_FromKarin] measures the tracks of charged particles at polar angles . In the region of angular overlap, FTD and short CTD track segments are used to reconstruct combined tracks, extending the detector acceptance for well-reconstructed tracks. Both the CTD and the combined tracks are linked to hits in the vertex detectors: the central silicon tracker (CST) [@CST34; @CST35], the backward silicon tracker (BST) [@BST37] and the forward silicon tracker (FST) [@FST36]. These detectors provide precise spatial coordinate measurements and therefore significantly improve the primary vertex spatial resolution. The CST consists of two layers of double-sided silicon strip detectors surrounding the beam pipe covering an angular range of for tracks passing through both layers. The BST consists of six double wheels of strip detectors measuring the transverse coordinates of charged particles. The FST design is similar to the BST and consists of five double wheels of single-sided strip detectors. The lead-scintillating fibre calorimeter (SpaCal) [@SPACAL] covering the region , has electromagnetic and hadronic sections. The calorimeter is used to measure the scattered positron and the backward hadronic energy flow. The energy resolution for positrons in the electromagnetic section is $\sigma (E)/E \approx 7.1\% / \sqrt{E/\text{GeV}} \oplus 1\%$, as determined in test beam measurements [@test-beam-Spacal]. The SpaCal provides energy and time-of-flight information used for triggering purposes. A backward proportional chamber (BPC) in front of the SpaCal is used to improve the angular measurement of the scattered lepton. The liquid argon (LAr) calorimeter [@Lar] covers the range and is used in this analysis in the reconstruction of the hadronic final state. It has an energy resolution of $\sigma (E)/E \approx 50\% / \sqrt{E/\text{GeV}} \oplus 2\%$ for hadronic showers, as obtained from test beam measurements [@test-beam-Lar]. Event reconstruction -------------------- The DIS kinematics is reconstructed based on the measurement of the scattered electron and the hadronic final state (HFS) particles. In the so-called $e\Sigma$-method [@eSigma] the kinematic variables $Q^2$, $y$ and $x$ are given by: $$Q^2 = 4E_eE_e'\cos^2 \left( \frac{\theta_e}{2} \right), \quad y = 2E_e \frac{\Sigma}{\left[\Sigma+E_e'(1-\cos\theta_e)\right]^2}\quad {\rm and}\quad x=\frac{Q^2}{s\cdot y},$$ where $s$ is the square of the centre-of-mass energy, $E_e'$ and $\theta_e$ the energy and polar angle of the scattered lepton, respectively, $E_e$ being the energy of incoming lepton and $\Sigma = \sum_i (E_i-p_{z,i})$ where the sum runs over all hadronic final state (HFS) particles $i$. This method provides an optimum in resolution of the kinematic variables and shows only little sensitivity to QED radiative effects. The HFS particles are reconstructed using an energy flow algorithm [@hadroo2]. This algorithm combines charged particle tracks and calorimetric energy clusters, taking into account their respective resolution and geometric overlap, into hadronic objects, while avoiding double counting of energy. Data selection -------------- DIS events were recorded using triggers based on electromagnetic energy deposits in the SpaCal calorimeter. The trigger efficiency is determined using independently triggered data. For DIS events the trigger inefficiency is negligible in the kinematic region of the analysis. The scattered lepton, defined by the most energetic SpaCal cluster, is required to have an energy $E'_e$ larger than $12$ GeV. The kinematical phase space is defined by $5 < Q^2 <$ and $0.05 < y < 0.6$, corresponding to the geometric acceptance of the SpaCal. The upper cut on $y$ reduces background from photoproduction. In addition, $x$ is required to be in the range of $0.0001 < x < 0.01$. Additional selections are made to reduce QED radiation effects and to suppress background events. The $z$ coordinate of the event vertex is required to be within $35$ cm of the nominal interaction point. Events with high energy initial state photon radiation are rejected by requiring $35< \sum_i (E_i-p_{z,i})<75$ GeV. Here, the sum extends over all HFS particles and the scattered electron. This cut further suppresses photoproduction background events to a level of about $0.5 \%$. The tracks used in the analysis are measured in the CTD alone (central tracks) or result from combinations of CTD and FTD information (combined tracks). Central tracks are required to have transverse momenta in the laboratory frame $p_{T} > 150$ MeV. The momentum of a combined track is required to be larger than $0.5$ GeV to ensure that the track has enough momentum to cross the endwall of the CJC. Both central and combined tracks are required to originate from the primary event vertex and to be in the pseudorapidity range $-2< \eta < 2.5$ measured in the laboratory frame. Using only tracks assigned to the event vertex, the contributions from in-flight decays of $K_S^{0}$, $\Lambda $ and from photon conversions and from other secondary decays are reduced. Further track quality cuts [@myThesis] are applied to ensure a high purity of the track reconstruction. Definition of experimental observables -------------------------------------- The results of this analysis are presented in the hadronic centre-of-mass frame (HCM), to minimise the effect of the transverse boost from the virtual photon. The transformation to the HCM frame is reconstructed with the knowledge of the kinematic variables $Q^2$ and $y$ [@myThesis]. The transverse momentum and pseudorapidity of charged particles in the HCM frame are labelled as $p_T^$ and $\eta^$. Since in this frame the positive $z^$ axis is defined by the direction of the virtual photon, HFS particles with $\eta^ > 0$ belong to the current hemisphere and particles with $\eta^* < 0$ originate from the target (proton remnant) hemisphere. Charged particle densities as a function of transverse momentum and pseudorapidity are defined as $(1/N)({\rm d}n/{\rm d}p_T^)$ and $(1/N)({\rm d}n/\rm{d}\eta^)$, respectively. Here, ${\rm d}n$ is the total number of charged particles with transverse momentum (pseudorapidity) in the ${\rm d}p_T^$ (${\rm d}\eta^$) bin and $N$ denotes the number of selected DIS events. For distributions measured differentially in $x$ and $Q^2$, ${\rm d}n$ and $N$ are the numbers for the respective $(x,Q^2)$ bin. Hadronisation is expected to be more relevant at small transverse momenta, while the hard parton radiation is expected to contribute more significantly at high $p_T^$ ($p_T^ > 1$ GeV) [@Kuhlen-pt]. To distinguish hadronisations effects from parton evolution signatures, the charged particle density is measured as a function of $\eta^$ for and for . The $p_T^$ dependence of the charged particle densities is studied in two different pseudorapidity intervals, $0 < \eta^* < 1.5$ and $1.5 < \eta^* < 5$, referred to as the “central region” and “current region”, respectively, as illustrated in . Such division approximately defines the regions where the sensitivity to the hard scatter is largest (current region), and where the parton shower models can be tested (central region). The target region, $\eta^{}<0$, is not accessible in this analysis. ![ The two pseudorapidity regions analysed in this paper. The region $0 < \eta^ < 1.5$ and $1.5 < \eta^* < 5$, are denoted as “central” and “current” regions, respectively.[]{data-label="fig:eta-feydiag"}](d13-012f2.eps){height="4.5cm" width="7.cm"} Data corrections ================ The data are corrected to the number of stable charged particles including charged hyperons, with proper lifetime $c\tau>10$ mm, in the phase space given in . Correction factors are calculated for each analysis bin from the ratio of the number of generated stable charged particles to the number of reconstructed tracks. The bin widths are chosen such that a purity[^1] of more than $75 \%$ is ensured in all bins. The correction takes into account detector effects like limited resolution and losses near the phase space boundaries, as well as a small residual contamination from weak decays of neutral particles (e.g. $K^0$ and $\Lambda$). [l\|l]{}\ Four momentum transfers squared & $5 < Q^2 < 100$ GeV$^2$\ Inelasticity & $0.05 < y < 0.6$\ \ Pseudorapidity in the laboratory frame & $ -2< \eta < 2.5$\ Transverse momentum in the laboratory frame & $p_{T} > 150$ MeV\ Pseudorapidity in the HCM frame & $ 0 < \eta^* < 5$\ Transverse momentum in the HCM frame & $ 0 < p_T^* < 10$ GeV\ In addition to migrations between bins inside the measurement phase space, there are migrations from outside of the analysis phase space and there is background from photoproduction. These contributions are subtracted prior to applying the correction factors according to the procedure outlined in [@forw-jet-h1-zeus3]. The [Djangoh]{} MC was used to correct the data. The differences to the correction factors obtained from [Rapgap]{} are taken as systematic uncertainties. The correction factors strongly depend on $\eta^$ and to a lesser extent on $p_T^$. In the $1.5 < \eta^* < 5$ region they vary between $1$ and $1.8$ with the largest values seen at high $p_T^$ and large $\eta^$. In the $0 < \eta^* < 1.5$ region, the correction factors rise up to $2.6$ at high $p_T^$, due to the limited detector acceptance in this region. The two MC models predict very similar correction factors for most of the phase space region, but differences up to $5.5 \%$ are observed at small $\eta^$ and high $p_T^$. Systematic uncertainties ======================== The following sources of systematic uncertainties are considered for all measured quantities. - The systematic uncertainty on the SpaCal energy scale is $1 \%$ [@ElCalib], which results in a systematic uncertainty of typically $0.4 \%$ for the measured single differential distributions. - The SpaCal angular resolution of $1$ mrad leads to a systematic uncertainty of about $0.1 \%$ for the measured distributions. - The hadronic energy scale uncertainty is known to a precision of $2\%$ [@HFSCalib]. The scale uncertainty enters into the uncertainty of the phase space calculations, which depend on $E-P_z$ of the HFS, and also affects the boost to the HCM frame. The systematic effect on the present measurements is about $0.3\%$. - The systematic uncertainty arising from the model dependence of the data correction is taken as the difference of the correction factors calculated using [Rapgap]{} and [Djangoh]{} MC. The resulting uncertainty on the measurements varies between $0.2 \%$ and $5.5 \%$. - The systematic uncertainty associated with the track reconstruction (e.g. track reconstruction efficiency, vertex reconstruction efficiency, weak decays and nuclear interaction uncertainties) is estimated to be: - $1 \%$ for central tracks, determined from the analysis of curling tracks and from the analysis of secondary vertices of charged particles located in the material between the two CJCs and originating from interactions with the detector material. The nuclear interaction cross sections of pions and kaons is found to be smaller in the simulation than in data. After correcting for these deficits, the agreement in the track efficiency between data and MC is found to be better than $1\%$. - $10 \%$ for combined tracks [@myThesis; @FTDJINST_FromKarin]. This was checked using all selected central tracks, as well as by using pions from $K_S^0$ decays, as a function of transverse momentum and pseudorapidity. Consistent results are obtained from both samples showing agreement of data and MC within $10 \%$. The systematic uncertainty associated with the track reconstruction is applied as an independent uncorrelated uncertainty on every data point. The resulting uncertainty on the measurements varies between $1 \%$ and $5.4 \%$ and is $1.6 \%$ on average. An additional systematic uncertainty of $0.2\%$ is assigned due to the different $K_S^0$ contamination seen in data and MC for both central and combined tracks. The corresponding effect arising from $\Lambda$ contamination is expected to be negligible. - The systematic uncertainty on the remaining photoproduction background is estimated to be $30\%$. This results in an uncertainty on the measured densities up to $0.9 \%$ at small $x$ and $Q^2$. At large $x$ and $Q^2$ the contribution from photoproduction is small and its uncertainty is negligible. The systematic uncertainties shown in the figures and tables are calculated by adding all contributions in quadrature. The total systematic uncertainty for the single differential measurements is below $2.5\%$ for most analysis bins. Results ======= The measurements of the charged particle densities as a function of pseudorapidity and transverse momentum in the phase space summarised in are listed in tables \[tab:pt-curr\] to \[tab:pt-curr7\] and shown in figures \[fig:eta-PDF-RAPGAP\] to \[fig:pt-bins-curr\]. Charged particle densities as a function of pseudorapidity ---------------------------------------------------------- The charged particle densities as a function of $\eta^$ were measured separately for and for , as shown in . In the soft $p_T^$ region, the pseudorapidity distribution is almost flat in the $1.5 \lesssim \eta^ \lesssim 3$ range with about $1.7$ charged particles per unit of pseudorapidity. The distribution falls at small $\eta^{}$ due to the cut on pseudorapidity in the laboratory frame. In the hard $p_T^$ region the distribution becomes more peaked near $\eta^{}=2.5$, with a maximum of $0.23$ charged particles per unit of pseudorapidity. For $1 < p_T^ < 10$ GeV the density increases rather strongly up to $\eta^{} \approx 2.5$, a behaviour expected from the strong ordering of transverse momentum towards the hard scattering vertex. also shows the predictions of the DGLAP-like model [Rapgap]{} based on different PDF sets. In the soft $p_T^$ region all NLO PDFs (HERAPDF1.0 [@HERAPDF10], CTEQ6.6 [@CTEQ66M], GRV98NLO [@GRV98]) show similar results and predict less particles compared to calculations using the default LO PDF set CTEQ6L(LO). All predictions are close to the data. At large $p_T^$, differences between the NLO PDF sets are observed, with CTEQ6L(LO) being closest to the data, although the differences to the data are still larger than the differences between the various PDF predictions. Similar PDF uncertainties are observed when using the CDM model as implemented in [Djangoh]{}. To check the sensitivity to hadronisation effects, the [Rapgap]{} predictions obtained with three sets of fragmentation parameters are compared to the data in : parameters tuned by ALEPH [@ALEPH], by the Professor tuning tool [@profftune] and default [Pythia]{}6.424 fragmentation parameters. Significant differences between these three samples are seen in the soft $p_T^$ region, where the data are best described by the ALEPH tune. At large transverse momenta they give similar results but none of them describes the data. Predictions from models with different approaches for QCD radiation (see section 2) are shown in . The data are compared to the CDM model [Djangoh]{}, the DGLAP-based MC [Rapgap]{} and Herwig++ and the CCFM model [Cascade]{}. In the soft $p_T^$ region, [Djangoh]{} and [Rapgap]{} describe the data within the PDF uncertainties (figure \[fig:eta-PDF-RAPGAP\]). Herwig++, which uses the cluster fragmentation model, provides a reasonable description of the data in the central region. The effect of not using the POWHEG option in Herwig++ also has been investigated. Only small differences were observed which are not considered further in this paper. [Cascade]{} predicts too high multiplicities in most of the measured $\eta^$ range. In the region of $1< p_T^{} <10$ GeV the best description of the data is achieved by [Djangoh]{}. [Rapgap]{} strongly undershoots the data in the central region. Herwig++ predicts a spectrum which is even below the prediction of [Rapgap]{}. [Cascade]{} is significantly above the data in a wide range of $\eta^$. The charged particle densities as a function of $\eta^$ are shown in for in eight different intervals of $x$ and $Q^{2}$ . The data are compared to predictions of the [Djangoh]{}, [Rapgap]{}, Herwig++ and [Cascade]{} generators. [Djangoh]{} provides a good description of the data over the full kinematic range. In general the description of the data by [Rapgap]{} is somewhat worse, with overshooting the data by about $10\%$ at low $x$. Herwig++ predicts smaller charged particle densities than observed in data in most of the phase space with differences of the order of $10\%$ at the highest $Q^2$. [Cascade]{} is significantly above the data for $\eta^{} < 3$ in all $(x,Q^2)$ bins. In figure \[fig:eta-hard\] the charged particle densities as a function of $\eta^$ are shown in $(x,Q^{2})$ intervals for $1< p_T^ < 10$ GeV . The shape of the distributions changes with $x$ and $Q^{2}$ more strongly than what is observed for (). At small values of $x$ and $Q^{2}$ the measured distribution is less dependent on $\eta^{}$ compared to the region at high $x$ and $Q^{2}$. None of the models describes all aspects of the data. In general [Djangoh]{} is closest to the data. However it fails to describe the data at low and medium $x$ in the central pseudorapidity region, with downwards deviations of the order of $20\%$. The [Rapgap]{} prediction is below the data, with the strongest deviation observed at small $x$ and small $\eta^$. Herwig++ significantly undershoots the data. The prediction of [Cascade]{} agrees reasonably well with the measurement at low $x$ and $Q^2$, but overshoots the data significantly as $x$ or $Q^2$ increases. Charged particle densities as a function of transverse momentum {#chap-result} --------------------------------------------------------------- In the charged particle densities as a function of $p_T^$ are shown for two pseudorapidity intervals, $0 < \eta^ < 1.5$ (central) and $1.5 < \eta^* < 5$ (current). The shapes of the measured $p_T^$ distributions in the two pseudorapidity ranges are similar. The spectrum falls by more than four orders of magnitude from $p_T^ < 1$ GeV to $p_T^* \approx 8$ GeV. The measurements are compared to the predictions of the [Djangoh]{}, [Rapgap]{}, Herwig++ and [Cascade]{} generators. [Djangoh]{} provides in general a good description of the data, while only at high $p_{T}^$ in the current region deviations from the measurement are observed. The other models fail to describe the data, with the strongest deviations being observed in the central region. The ratio of RAPGAP to data shows a sharp drop at $p_T^\approx 1$ GeV. The $p_T^$ spectra predicted by Herwig++ are even softer than those predicted by [Rapgap]{}. [Cascade]{} in general produces higher particle densities than measured. In the charged particle densities as a function of $p_T^$ are shown for eight $(x,Q^{2})$ intervals for the central and the current region, respectively. In the central region the measurement shows a dependence on $x$, such that the number of soft particles is decreasing with increasing $x$ for fixed $Q^{2}$. In the current region this effect is less pronounced. The [Djangoh]{} model provides in general a good description of the data over the full kinematic range in both pseudorapidity regions, degrading at high $p_{T}^$ in the lowest $(x,Q^{2})$ bin. Significant deviations of the [Rapgap]{} predictions from data are observed in the central region at low $x $ and $Q^{2}$. The description becomes somewhat better at larger values of $x$ and $Q^{2}$. The same trend is observed for the current region, but the overall data description is better. Herwig++ fails to describe the measurements at high $p_{T}^$ in the whole phase space. At lowest $x$ and $Q^{2}$ the spectrum is much softer than the one obtained with [Rapgap]{}, while at high $x$ and $Q^{2}$ both predictions are similar. [Cascade]{} describes the data in the lowest $x$ and $Q^{2}$ bin at high $p_{T}^$ only. Conclusion ========== This paper presents a study of charged particle production in $ep$ collisions at low $Q^2$ measured with the H1 detector. The kinematic range of the analysis covers low photon virtualities, , and small values of $x$, $10^{-4} < x < 10^{-2}$. The analysis is performed in the hadronic centre-of-mass system. The charged particle densities as a function of pseudorapidity ($\eta^$) and transverse momentum ($p_T^$) are measured differentially in $x$ and $Q^2$. The charged particle densities as a function of pseudorapidity show different shapes, depending on the $p_T^$ range. For $0<p_T^* < 1$ GeV the density of particles is approximately constant for $1<\eta^<3.5$, while for the density increases with increasing $\eta^{}$ up to $\eta^* \approx 2.5$, a behaviour expected from the strong ordering of transverse momentum towards the hard scattering vertex. The charged particle densities as a function of transverse momentum show an $x$ dependence at small $\eta^$ ($0 < \eta^ < 1.5$), such that the number of soft particles is decreasing with increasing $x$, while in the $1.5 < \eta^* < 5$ range this effect is less visible. In order to relate the charged hadron spectra to the parton dynamics at small $x$, the data are compared to QCD models with different evolution approaches for simulating the parton cascade and with different hadronisation schemes. The data allow the validity of different models to be tested. At small $p_T^$, the data are reasonably well described by [Djangoh]{} (based on the Colour Dipole Model), as well as by [Rapgap]{} (based on the DGLAP shower evolution). At high $p_T^$ and at low $\eta^$, [Rapgap]{} severely undershoots the data. The differences are most pronounced at lowest $x$ and $Q^2$, and decrease with increasing $x$ and $Q^2$ values. Herwig++ which is also based on DGLAP but uses a cluster fragmentation model is significantly below the data over the full phase space. [Cascade]{} (based on CCFM) gives a reasonable description only at the lowest $x$ and $Q^2$, but overall predicts higher charged particle densities than observed in data. The Colour Dipole Model implemented in [Djangoh]{} is the best among the considered models and provides a reasonable description of the data. Acknowledgement =============== We are grateful to the HERA machine group whose outstanding efforts have made this experiment possible. We thank the engineers and technicians for their work in constructing and maintaining the H1 detector, our funding agencies for financial support, the DESY technical staff for continual assistance and the DESY directorate for support and for the hospitality which they extend to the non-DESY members of the collaboration. We would like to give credit to all partners contributing to the WLCG computing infrastructure for their support for the H1 Collaboration. [10]{} V. Gribov and L. Lipatov, Sov. J. Nucl. Phys. [15]{} (1972) 438 and 675;\ L. Lipatov, Sov. J. Nucl. Phys. [20]{} (1975) 94;\ G. Altarelli and G. Parisi, Nucl. Phys. [B126]{} (1977) 298;\ Y. Dokshitzer, Sov. Phys. JETP [46]{} (1977) 641. E. Kuraev, L. Lipatov and V. Fadin, Sov. Phys. JETP [44]{} (1976) 443;\ E. Kuraev, L. Lipatov and V. Fadin, Sov. Phys. JETP [45]{} (1977) 199;\ Y. Balitsky and L. Lipatov, Sov. J. Nucl. Phys. [28]{} (1978) 822. M. Ciafaloni, Nucl. Phys. [B296]{} (1988) 49;\ M. Catani, F. Fiorani and G. Marchesini, Phys. Lett. [B234]{} (1990) 339;\ S. Catani, F. Fiorani and G. Marchesini, Nucl. Phys. [B336]{} (1990) 18;\ G. Marchesini, Nucl. Phys. [B445]{} (1995) 49. F.D. Aaron [et al.]{} \[H1 and ZEUS Collaborations\], JHEP [1001]{} (2010) 109 \[arXiv:0911.0884\];\ F.D. Aaron [et al.]{} \[H1 Collaboration\], JHEP [1209]{} (2012) 061 \[arXiv:1206.7007\]. R.S. Thorne, Phys. Rev. [D73]{} (2006) 054019 \[hep-ph/0601245\];\ A.D. Martin [et al.]{}, Eur. Phys. J. [C63]{} (2009) 189 \[arXiv:0901.0002\];\ R.S. Thorne \[arXiv:1201.6180\]. H.L. Lai [et al.]{}, Phys. Rev. [D82]{}, (2010) 074024 \[arXiv:1007.2241\];\ P. Nadolsky [et al.]{}, contribution to the Proceedings of the XX Workshop on Deep Inelastic Scattering and Related Subjects, Bonn, Germany, 26-30 March, 2012 \[arXiv:1206.3321\]. S. Forte [et al.]{}, Nucl. Phys. [B834]{}, (2010) 116 \[arXiv:1001.2312\];\ R.D. Ball [et al.]{} \[NNPDF Collaboration\], Nucl. Phys. [B849]{} (2011) 296 \[arXiv:1101.1300\];\ R.D. Ball [et al.]{} \[NNPDF Collaboration\], Nucl. Phys. [B855]{} (2012) 153 \[arXiv:1107.2652\]. S. Alekhin and S. Moch, Phys. Lett. [B699]{} (2011) 345 \[arXiv:1011.5790\];\ S. Alekhin [et al.]{}, Phys. Rev. [D81]{} (2010) 014032 \[arXiv:0908.2766\];\ S. Alekhin and S. Moch \[arXiv:1107.0469\]. A. Aktas [et al.]{} \[H1 Collaboration\], Eur. Phys. J. [C33]{} (2004) 477 \[hep-ex/0310019\]. C. Adloff [et al.]{} \[H1 Collaboration\], Phys. Lett. [B415]{} (1997) 418 . C. Adloff [et al.]{} \[H1 Collaboration\], Eur. Phys. J. [C12]{} (2000) 595 \[hep-ex/9907027\]. S. Aid [et al.]{} \[H1 Collaboration\], Eur. Phys. [B356]{} (1995) 118 . A. Aktas [et al.]{} \[H1 Collaboration\], Eur. Phys. J. [C46]{} (2006) 27 \[hep-ex/0508055\]. S. Chekanov [et al.]{} \[ZEUS Collaboration\], Eur. Phys. J. [C52]{} (2007) 515 \[arXiv:0707.3093\]. F.D. Aaron [et al.]{} \[H1 Collaboration\], Eur. Phys. J. [C72]{} (2012) 1910 \[arXiv:1111.4227\]. A. Aktas [et al.]{} \[H1 Collaboration\], Eur. Phys. J. [C36]{} (2004) 441 \[hep-ex/0404009\]. M. Kuhlen, Phys. Lett. [B382]{} (1996) 441 \[hep-ph/9606246\]. C. Adloff [et al.]{} \[H1 Collaboration\], Nucl. Phys [B485]{} (1997) 3 \[hep-ex/9610006\]. H. Jung, [Rapgap]{} 3.1, Comput. Phys. Commun [86]{} (1995) 147. K. Charchula, G. A. Schuler and H. Spiesberger, 1.4, Comput. Phys. Commun. [81]{} (1994) 381. L. Lönnblad, 4.10, Comput. Phys. Commun. [71]{} (1992) 15. L. Lönnblad, Z.Phys. [C65]{} (1995) 285;\ A.H. Mueller, Nucl. Phys. [B415]{} (1994) 373. H. Jung, Comput. Phys. Commun. [143]{} (2002) 100;\ H. Jung [et al.]{}, [Cascade]{} 2.2.0, Eur. Phys. J. [C70]{} (2010) 1237. S. Gieseke [et al.]{}, Herwig++ 2.5, \[arXiv:1102.1672\]. L. D’Errico and P. Richardson, Eur. Phys. J. [C72]{} (2011) 2042 \[arXiv:1106.2983\]. G. Marchesini, Nucl. Phys. [B445]{} (1995) 49 . S. Catani, B.R. Webber and G. Marchesini, Nucl. Phys. [B349]{} (1991) 635. R. Engel, Z. Phys. [C66]{} (1995) 203;\ R. Engel and J. Ranft, [Phojet]{} 1.6, Phys. Rev. [D54]{} (1996) 4244 \[hep-ph/9509373\]. A. Capella [et al.]{}, Phys. Rept. [236]{} (1994) 227. J. Pumplin [et al.]{}, JHEP [0207]{} (2002) 012 \[hep-ph/0201195\]. H. Jung, in “Proceedings of the XII International Workshop on Deep Inelastic Scattering (DIS2004)”, eds. D. Bruncko, J. Ferencei, P. Stríženec, Štrbské Pleso, Slovakia (2004), 299 \[hep-ph/0411287\]. A.D. Martin [et al.]{}, Phys. Lett. [B531]{} (2002) 216 \[hep-ph/0201127\]. B. Andersson [et al.]{}, Phys. Rept. [97]{} (1983) 31. T. Sjöstrand [et al.]{}, Comput. Phys. Commun. [82]{} (1994) 174 \[hep-ph/9508391\]. T. Sjöstrand [et al.]{}, Comput. Phys. Commun. [135]{} (2001) 238 \[hep-ph/0010017\]. S. Schael [et al.]{} \[ALEPH Collaboration\], Phys. Lett. [B606]{} (2005) 265. A. Buckley [et al.]{}, Eur. Phys. J. [C65]{} (2010) 331 \[arXiv:0907.2973\]. B.R. Webber, Nucl. Phys. [B238]{} (1984) 492;\ G. Marchesini and B.R. Webber, Nucl. Phys. [B310]{} (1988) 461. A. Kwiatkowski, H. Spiesberger, and H.J. Möhring, [Heracles]{}1.0, Comput. Phys. Commun. [69]{} (1992) 155. R. Brun [et al.]{}, 3, Technical Report CERN-DD/EE-84-1(1987). A. Grebenyuk, [ “Transverse Momentum of Charged Particles in low-$Q^2$ DIS at HERA"]{}, PhD thesis, Hamburg University, 2012, (available at\ <http://www-h1.desy.de/publications/theses_list.html>). I. Abt [et al.]{} \[H1 collaboration\], Nucl. Instrum. Meth. [A386]{} (1997) 310. I. Abt [et al.]{} \[H1 collaboration\], Nucl. Instrum. Meth. [A386]{} (1997) 348. R. D. Appuhn [et al.]{} \[H1 SpaCal Group\], Nucl. Instrum. Meth. [A386]{} (1997) 397. D. Pitzl [et al.]{}, Nucl. Instrum. Meth. [A454]{} (2000) 334 \[hep-ex/0002044\]. B. List, Nucl. Instrum. Meth. [A501]{} (2001) 49. J. Becker [et al.]{}, Nucl. Instrum. Meth. [A586]{} (2008) 190. P.J. Laycock [et al.]{}, JINST 7 (2012) T8003 \[arXiv:1206.4068\]. J. Kretzschmar, [A Precision Measurement of the Proton Structure Function $F_2$ with the H1 Experiment]{}, PhD thesis, Humboldt University, Berlin, 2008, (available at\ <http://www-h1.desy.de/publications/theses_list.html>). I. Glushkov, [“D\* Meson Production in Deep Inelastic Electron-Proton Scatter- ing with the Forward and Backward Silicon Trackers of the H1 Experiment at HERA”]{}, PhD thesis, Humboldt University, Berlin, 2007, (available at\ <http://www-h1.desy.de/publications/theses_list.html>). T. Nicholls [et al.]{} \[H1 SpaCal Group\], Nucl. Instrum. Meth. [A374]{} (1996) 149. B. Andrieu [et al.]{} \[H1 Calorimeter Group\], Nucl. Instrum. Meth. [A336]{} (1993) 460. B. Andrieu [et al.]{} \[H1 Calorimeter Group\], Nucl. Instrum. Meth. [A336]{} (1993) 499. U. Bassler and G. Bernardi, Nucl. Instrum. Meth. [A426]{} (1999) 583 \[hep-ex/9801017\]. M. Peez, [ “Search for Deviations from the Standard Model in High Transverse Energy Processes at the Electron Proton Collider HERA”]{} (In French), PhD thesis, Lyon University, 2003, (available at\ <http://www-h1.desy.de/publications/theses_list.html>);\ B. Portheault, [ “First Measurement of Charged and Neutral Current Cross Sections with the Polarised Positron Beam at HERA II and QCD-Electroweak Analyses”]{} (In French), PhD thesis, Paris XI ORSAY University, 2005, (available at\ <http://www-h1.desy.de/publications/theses_list.html>);\ S. Hellwig, [ “Untersuchung der $D^\ast-\pi_{\rm slow}$ Double Tagging Methode in Charmanalysen”]{}, Dipl. thesis, Hamburg University, 2004, (available at\ <http://www-h1.desy.de/publications/theses_list.html>). F.D. Aaron [et al.]{} \[H1 Collaboration\], Eur.Phys. J. [C63]{} (2009) 625 \[arXiv:0904.0929\]. D. Salek, [ “Measurement of the Longitudinal Proton Structure Function in Diffraction at the H1 Experiment and Prospects for Diffraction at LHC”]{}, PhD thesis, Charles Univ. Prague, DESY-THESIS-2011-013 (available at <http://www-h1.desy.de/publications/theses_list.html>). F.D. Aaron [et al.]{} \[H1 and ZEUS Collaboration\], JHEP [1001]{} (2010) 109 \[arXiv:0911.0884\]. P.M. Nadolsky [et al.]{}, Phys. Rev. [D78]{} (2008) 013004 \[arXiv:0802.0007\]. M. Glück, E. Reya and A. Vogt, Eur. Phys. J. [C5]{} (1998) 461 \[hep-ph/9806404\]. 5.8pt [\|c\|ccc\|]{}\ & $1\!/\!N \cdot {\rm d}n/{\rm d}\eta^\,$ & stat. & sys.\ & & &\ $0.0 - 1.0$ & $1.019$ & $0.06$ & $6.2$\ $1.0 - 1.6 $ & $1.577$ & $0.03$ & $3.4$\ $1.6 - 2.1$ & $1.717$ & $0.03 $ & $2.7$\ $2.1 - 2.6$ & $1.754$ & $0.03$ & $2.1$\ $2.6 - 3.1$ & $1.706$ & $0.03$ & $1.4$\ $3.1 - 3.7$ & $1.467$ & $0.04$ & $1.4$\ $3.7 - 5.0 $ & $0.691$ & $0.08 $ & $1.7$\ 5.8pt [\|c\|ccc\|]{}\ & $1\!/\!N \cdot {\rm d}n/\rm{d}\eta^\,$ & stat. & sys.\ & & &\ $0.0 - 0.5$ & $0.0807$ & $0.26$ & $6.9$\ $0.5 - 1.0$ & $0.1448$ &$ 0.14$ & $4.6$\ $1.0 - 1.5$ & $0.1835$ & $0.11$ & $2.4$\ $1.5 - 2.0$ & $0.2066$ & $0.10$ & $1.4$\ $2.0 - 2.5$ & $0.2255$ & $0.09$ & $1.8$\ $2.5 - 3.0$ & $0.2251$ & $0.09$ & $1.9$\ $3.0 - 3.7$ & $0.1668$ & $0.13$ & $2.3$\ $3.7 - 5.0$ & $0.0329$ & $0.42$ & $4.5$\ 5.8pt [ccc]{} [\|c\|c\|ccc\|]{}\ & & $1\!/\!N \cdot {\rm d}n/\rm{d}\eta^\,$ & stat. & sys.\ & & & &\ & $0.0 - 1.0$ & $1.398$ & $0.14$ & $5.8$\ & $1.0 - 1.6$ & $1.621$ & $0.09$ & $3.5$\ & $1.6 - 2.1$ & $1.727$ & $0.08$ & $2.8$\ & $2.1 - 2.6$ & $1.760$ & $0.08$ & $3.5$\ & $2.6 - 3.1$ & $1.749$ & $0.08$ & $3.5$\ & $3.1 - 3.7$ & $1.650$ & $0.10$ & $2.6$\ & $3.7 - 5.0$ & $0.683$ & $0.23$ & $2.1$\ & $0.0 - 1.5$ & $1.241$ & $0.22$ & $4.4$\ & $1.5 - 2.3$ & $1.682$ & $0.12$ & $2.7$\ & $2.3 - 2.8$ & $1.732$ & $0.10$ & $ 3.2$\ & $2.8 - 3.3$ & $1.671$ & $0.10$ & $2.1$\ & $3.3 - 3.9$ & $1.347$ & $0.12$ & $2.0$\ & $3.9 - 5.0$ & $0.652$ & $0.24$ & $2.1$\ &$ 0.5 - 2.0$ & $1.288$ & $0.21$ & $4.4$\ & $2.0 - 2.9$ & $1.613$ & $0.13$ & $1.8$\ & $2.9 - 3.7$ & $1.272$ & $0.15$ & $2.0$\ & $3.7 - 5.0$ & $0.554$ & $0.28$ &$ 2.5$\ & $0.0 - 1.0$ & $1.464$ & $0.15$ & $5.9$\ & $1.0 - 1.6$ & $1.721$ & $0.10$ & $2.5$\ & $1.6 - 2.1$ & $1.820$ & $0.09$ & $2.6$\ & $2.1 - 2.6$ & $1.865$ & $0.09$ & $2.8$\ & $2.6 - 3.1$ & $1.857$ & $0.09$ & $2.1$\ &$ 3.1 - 3.7$ & $1.680$ & $0.10$ & $1.6$\ & $3.7 - 5.0$ & $0.784$ & $0.23$ & $1.7$\ -- -------------- ------------------------------------------ -------- -------- $1\!/\!N \cdot {\rm d}n/{\rm d}\eta^\,$ stat. sys. $0.0 - 1.5$ $1.268$ $0.22$ $5.2$ $1.5 - 2.3$ $1.790$ $0.12$ $2.6$ $2.3 - 2.8$ $1.819$ $0.10$ $2.1$ $2.8 - 3.3$ $1.685$ $0.10$ $1.7$ $3.3 - 3.9$ $1.312$ $0.12$ $1.9$ $3.9 - 5.0$ $0.700$ $0.23$ $2.0$ $0.5 - 2.0$ $ 1.34$ $0.20$ $8.6$ $2.0 - 2.9$ $1.650$ $0.13$ $1.9$ $2.9 - 3.7$ $1.263$ $0.14$ $2.3$ $3.7 - 5.0$ $ 0.550$ $0.28$ $3.4$ $0.0 - 1.0$ $1.461$ $0.15$ $5.8$ $1.0 - 1.6$ $1.820$ $0.09$ $2.7$ $1.6 - 2.1$ $1.928$ $0.08$ $2.4$ $2.1 - 2.6$ $1.951$ $0.08$ $2.4$ $ 2.6 - 3.1$ $1.883$ $0.08$ $2.2$ $3.1 - 3.7$ $1.601$ $0.10$ $1.7$ $3.7 - 5.0$ $0.883$ $0.21$ $2.1$ $ 0.0 - 1.5$ $1.077$ $0.20$ $5.4$ $1.5 - 2.2$ $1.783$ $0.09$ $2.6$ $2.2 - 2.9$ $1.714$ $0.09$ $1.8$ $2.9 - 3.7$ $1.445$ $0.11$ $1.9$ $3.7 - 5.0$ $0.634$ $0.22$ $2.0 $ -- -------------- ------------------------------------------ -------- -------- 5.8pt [ccc]{} [\|c\|c\|ccc\|]{}\ & & $1\!/\!N \cdot {\rm d}n/\rm{d}\eta^\,$ & stat. & sys.\ & & & &\ & $0.0 - 0.5$ & $0.1365$ &$ 0.39$ & $5.7$\ & $0.5 - 1.0$ & $0.1551$ & $0.29$ & $3.5$\ & $1.0 - 1.5$ & $0.1679$ & $0.27$ & $2.3$\ & $1.5 - 2.0$ & $0.1818$ & $0.26$ & $3.5$\ & $2.0 - 2.5$ & $0.1961$ & $0.26$ & $2.4$\ & $2.5 - 3.0$ & $0.2084$ &$ 0.24$ & $2.4$\ & $3.0 - 3.7$ &$ 0.1984$ & $0.31$ & $2.3$\ & $3.7 - 5.0$ & $0.0501$ & $1.02$ & $3.6$\ & $0.0 - 1.0$ & $0.1115$ & $0.70$ & $ 7.3$\ & $1.0 - 1.5$ & $0.1575$ & $0.33$ & $3.4$\ & $1.5 - 2.0$ & $0.1758$ & $0.31$ & $ 3.3$\ & $2.0 - 2.5$ & $0.1903$ & $0.30$ & $5.2$\ & $2.5 - 3.0$ & $0.2037$ & $0.29$ & $1.9$\ & $3.0 - 3.7$ & $0.1626$ & $0.38$ & $2.6$\ & $3.7 - 5.0$ & $0.0311$ & $1.33$ & $8.0$\ & $0.5 - 1.5$ & $0.1244$ & $0.66$ & $ 4.5$\ & $1.5 - 2.0$ &$ 0.1675$ & $0.32$ & $4.7$\ & $2.0 - 2.5$ & $0.1811$ &$ 0.31$ & $ 3.0$\ & $2.5 - 3.0$ &$ 0.1686$ & $0.32$ & $2.3$\ & $3.0 - 3.7$ & $0.0978$ & $0.51$ & $3.0$\ & $3.7 - 5.0$ & $0.01071$ & $2.21$ & $4.7$\ & $0.0 - 0.5$ & $0.1506$ & $0.42$ & $6.1$\ & $0.5 - 1.0$ & $0.1764$ & $0.30$ & $2.1$\ & $1.0 - 1.5$ & $0.1959$ & $0.28$ & $2.7$\ & $1.5 - 2.0$ & $0.2180$ & $0.26$ & $1.7$\ & $2.0 - 2.5$ & $0.2444$ & $0.25$ & $1.9$\ & $2.5 - 3.0$ & $0.2646$ & $0.23$ & $2.7$\ & $3.0 - 3.7$ & $0.2334$ & $0.31$ & $1.4$\ & $3.7 - 5.0$ & $0.0552$ & $1.03$ & $2.7$\ -- -------------- ----------------------------------------- -------- -------- $1\!/\!N \cdot {\rm d}n/\rm{d}\eta^\,$ stat. sys. $0.0 - 1.0$ $0.123$ $0.69$ $8.3$ $1.0 - 1.5$ $0.1910$ $0.30$ $3.4$ $1.5 - 2.0$ $ 0.2145$ $0.27$ $3.1$ $2.0 - 2.5$ $0.2409$ $0.26$ $ 3.6$ $2.5 - 3.0$ $0.2415$ $0.26$ $2.6$ $3.0 - 3.7$ $0.1750$ $0.37$ $2.4$ $3.7 - 5.0$ $0.0315$ $1.34$ $3.7$ $0.5 - 1.5$ $0.1496$ $0.58$ $6.6$ $1.5 - 2.0$ $0.2086$ $0.27$ $3.8$ $2.0 - 2.5$ $0.2246$ $0.26$ $4.5$ $2.5 - 3.0$ $0.1997$ $0.28$ $3.8$ $3.0 - 3.7$ $0.1051$ $0.48$ $3.8$ $ 3.7 - 5.0$ $0.01028$ $2.38$ $ 7.7$ $0.0 - 0.5$ $0.1581$ $0.42$ $6.1$ $0.5 - 1.0$ $0.2105$ $0.27$ $2.3$ $ 1.0 - 1.5$ $0.2479$ $0.23$ $ 2.5$ $1.5 - 2.0$ $0.2820$ $0.22$ $2.1$ $2.0 - 2.5$ $0.3188$ $0.20$ $2.3$ $2.5 - 3.0$ $0.3386$ $0.20$ $1.4$ $ 3.0 - 3.7$ $0.2601$ $0.28$ $1.5$ $3.7 - 5.0$ $0.0602$ $0.95$ $1.8$ $0.0 - 1.0$ $0.118$ $0.63$ $9.0$ $1.0 - 1.5$ $0.251$ $0.22$ $4.5$ $1.5 - 2.0$ $0.2966$ $0.19$ $3.4$ $2.0 - 2.5$ $0.3167$ $0.19$ $2.1$ $2.5 - 3.0$ $0.2799$ $0.20$ $2.1$ $3.0 - 3.7$ $0.1587$ $0.34$ $2.4$ $3.7 - 5.0$ $0.02112$ $1.49$ $4.6$ -- -------------- ----------------------------------------- -------- -------- 5.8pt [\|c\|lcc\|]{}\ & $1\!/\!N \cdot {\rm d}n/{\rm d}p_T^\,$ & stat. & sys.\ & & &\ $ 0.2 - 0.4$ & $\ 3.952$ & $0.01$ & $2.0$\ $ 0.4 - 0.6$ & $\ 2.431$ & $0.02$ & $1.6$\ $ 0.6 - 1.0$ & $\ 0.954$ & $0.04$ & $1.8$\ $1.0 - 2.0$ & $\ 0.1686$ & $ 0.15$ & $2.5$\ $2.0 - 4.0$ & $\ 1.549\cdot 10^{-2}$ & $0.70$ & $2.0$\ $4.0 - 10.0$ & $\ 7.15\cdot 10^{-4}$ & $5.4$1 & $1.9$\ 5.8pt [\|c\|lcc\|]{}\ & $1\!/\!N \cdot {\rm d}n/{\rm d}p_T^\,$ & stat. & sys.\ & & &\ $0.0 - 0.3$ & $\ 5.24$ & $0.01$ & $1.8$\ $0.3 - 0.6$ & $\ 6.10$ & $0.01$ & $1.7$\ $0.6 - 1.0$ & $\ 2.234$ & $0.02$ & $1.8$\ $1.0 - 1.5$ & $\ 0.6193$ & $0.05$ & $1.5$\ $1.5 - 2.1$ & $\ 0.1849$ & $0.10$ & $1.5$\ $2.1 - 3.0$ & $\ 5.23\cdot 10 ^{-2}$ & $0.23$ & $2.0$\ $3.0 - 4.0$ & $\ 1.381\cdot 10 ^{-2}$ & $0.47$ & $2.0$\ $4.0 - 5.0$ & $\ 4.14\cdot 10 ^{-3}$ & $0.84$ & $2.4$\ $5.0 - 6.3$ & $\ 1.402\cdot 10 ^{-3}$ & $1.67$ & $2.8$\ $6.3 - 7.9$ & $\ 3.98\cdot 10 ^{-4}$ & $3.47$ & $2.5$\ $7.9 - 10.0$ & $\ 1.061\cdot 10 ^{-4}$ & $7.60$ & $3.2$\ 5.8pt [\|c\|c\|lcc\|]{}\ & & $1\!/\!N \cdot {\rm d}n/{\rm d}p_T^\,$ & stat. & sys.\ & & & &\ & $0.2 - 0.4$ & $\ 4.76$ & $0.03$ & $3.2$\ & $0.4 - 0.6$ & $\ 2.92$ & $0.04$ & $2.9$\ & $0.6 - 1.0$ &$ \ 1.144$ & $0.10$ & $3.7$\ & $1.0 - 2.0$ & $\ 0.1955$ & $0.39$ & $3.0$\ & $2.0 - 4.0$ & $\ 1.489\cdot10^{-2}$ & $1.89$ & $3.2$\ & $4.0 - 10.0$ & $\ 5.69\cdot 10^{-4}$ & $15.36$ & $6.0$\ & $0.2 - 0.4$ & $\ 3.99$ & $0.05$ & $2.5$\ & $0.4 - 0.6$ & $\ 2.53$ & $0.06$ & $2.6$\ & $0.6 - 1.0$ & $\ 0.994$ & $0.14$ &$ 2.9$\ & $1.0 - 2.0$ & $\ 0.1611$ & $0.52$ & $3.3$\ & $2.0 - 4.0$ &$ \ 1.286\cdot 10 ^{-2}$ & $2.58$ &$ 2.7$\ & $4.0 - 10.0$ & $\ 5.1\cdot 10 ^{-4}$ & $21.40$ & $4.9$\ & $0.2 - 0.6$ & $\ 2.097$ & $0.11$ & $3.0$\ & $0.6 - 1.0$ & $\ 0.659$ & $0.19$ & $3.4$\ & $1.0 - 2.0$ & $\ 0.1113$ & $0.74$ & $2.8 $\ & $2.0 - 4.0$ & $\ 8.82\cdot 10^{-3}$ & $3.40$ & $3.1$\ & $4.0 - 10.0$ & $\ 3.3\cdot 10^{-4}$ & $34.47$ & $5.5$\ -- -------------- ----------------------------------------- ---------- --------- $1\!/\!N \cdot {\rm d}n/{\rm d}p_T^\,$ stat. sys. $0.2 - 0.4$ $\ 4.97$ $0.04$ $2.5$ $0.4 - 0.6$ $\ 3.060$ $0.05$ $2.7$ $0.6 - 1.0$ $\ 1.229$ $0.11$ $2.6$ $1.0 - 2.0$ $\ 0.2155$ $0.40 $ $3.5$ $2.0 - 4.0$ $\ 1.960\cdot 10^{-2}$ $1.87$ $3.1 $ $4.0 - 10.0$ $\ 9.2\cdot 10^{-4}$ $14.13$ $3.3$ $0.2 - 0.4$ $\ 4.05$ $0.05$ $3.0$ $0.4 - 0.6$ $\ 2.593$ $0.06$ $2.9$ $ 0.6 - 1.0$ $\ 1.033$ $0.13$ $3.5$ $ 1.0 - 2.0$ $\ 0.1811$ $0.50$ $4.2$ $2.0 - 4.0$ $\ 1.623 \cdot 10 ^{-2}$ $2.34$ $5.8$ $4.0 - 10.0$ $\ 7.5 \cdot 10 ^{-4}$ $18.91$ $3.2$ $0.2 - 0.6$ $ \ 2.124$ $0.10$ $4.5$ $0.6 - 1.0$ $ \ 0.692$ $0.18$ $ 4.1 $ $ 1.0 - 2.0$ $ \ 0.129$ $ 0.66$ $4.8$ $2.0 - 4.0$ $ \ 1.119 \cdot 10 ^{-2}$ $3.12$ $3.0$ $4.0 - 10.0$ $\ 4.42 \cdot 10 ^{-4}$ $28.92$ $6.0$ $0.2 - 0.4$ $\ 5.00$ $0.04$ $2.5$ $0.4 - 0.6$ $\ 3.156$ $0.04$ $ 2.4$ $0.6 - 1.0$ $\ 1.296$ $0.97$ $2.8$ $1.0 - 2.0$ $\ 0.2474$ $ 0.36$ $3.7$ $2.0 - 4.0$ $\ 2.579 \cdot 10 ^{-2}$ $1.57$ $ 4.7$ $4.0 - 10.0$ $\ 1.40 \cdot 10 ^{-3}$ $11.28$ $2.8$ $0.2 - 0.4$ $\ 3.356$ $0.05$ $2.8$ $0.4 - 0.6$ $\ 2.254$ $0.05$ $2.8$ $0.6 - 1.0$ $\ 0.945$ $0.12$ $3.6$ $1.0 - 2.0$ $\ 0.1905$ $0.42$ $5.4$ $2.0 - 4.0$ $\ 2.185 \cdot 10 ^{-2}$ $1.74$ $ 6.9 $ $4.0 - 10.0$ $\ 1.21 \cdot 10 ^{-3}$ $1 3.18$ $3.5$ -- -------------- ----------------------------------------- ---------- --------- 5.8pt [\|c\|c\|lcc\|]{}\ & & $1\!/\!N \cdot {\rm d}n/{\rm d}p_T^\,$ & stat. & sys.\ & & & &\ & $0.0 - 0.3$ &$\ 5.60 $&$ 0.04 $&$ 2.5$\ & $0.3 - 0.6$ &$\ 6.83 $&$ 0.04 $&$ 2.8$\ & $0.6 - 1.0$ &$ \ 2.290 $&$ 0.07 $&$ 3.6$\ & $1.0 - 1.5$ &$ \ 0.639 $&$ 0.14 $&$ 3.3$\ & $1.5 - 2.1$ &$ \ 0.1897 $&$ 0.30 $&$ 2.7$\ & $2.1 - 3.0$ &$ \ 5.12 \cdot 10 ^{-2} $&$ 0.69 $&$ 4.0$\ & $3.0 - 4.0$ &$ \ 1.378 \cdot 10 ^{-2} $&$ 1.40 $&$ 4.8 $\ & $4.0 - 5.0$ &$ \ 3.97 \cdot 10 ^{-3} $&$ 2.46 $&$ 8.6$\ & $5.0 - 6.3$ &$ \ 1.40 \cdot 10 ^{-3} $&$ 4.68 $&$ 9.7$\ &$ 6.3 - 7.9$ &$ \ 4.34 \cdot 10 ^{-4} $&$ 9.33 $&$ 6.2$\ & $7.9 - 10.0$ &$ \ 9.3 \cdot 10 ^{-5} $&$ 20.80 $&$ 13.2$\ & $0.0 - 0.3$ &$ \ 5.57 $&$ 0.04 $&$ 2.8 $\ & $0.3 - 0.6$ &$ \ 6.64 $&$ 0.04 $&$ 2.6 $\ & $0.6 - 1.0$ &$ \ 2.207 $&$ 0.08 $&$ 2.2 $\ & $1.0 - 1.5$ &$ \ 0.587 $&$ 0.17 $&$ 1.7 $\ & $1.5 - 2.1$ &$ \ 0.161 $&$ 0.36 $&$ 1.8 $\ & $2.1 - 3.0$ &$ \ 4.12 \cdot 10 ^{-2} $&$ 0.85 $&$ 3.2 $\ & $3.0 - 4.0$ &$ \ 1.002 \cdot 10 ^{-2} $&$ 1.83 $&$ 5.7 $\ & $4.0 - 5.0$ &$ \ 2.63 \cdot 10 ^{-3} $&$ 3.46 $&$ 6.7 $\ & $5.0 - 6.3$ &$ \ 7.69 \cdot 10 ^{-4} $&$ 7.08 $&$ 5.7 $\ &$ 6.3 - 7.9$ &$ \ 2.46 \cdot 10 ^{-4} $&$ 13.98 $&$ 10.4 $\ & $7.9 - 10.0$ &$ \ 6.65 \cdot 10 ^{-5} $&$ 32.90 $&$ 18.6 $\ & $0.0 - 0.3$ &$ \ 4.96 $&$ 0.05 $&$ 2.2$\ &$ 0.3 - 0.6$ &$ \ 5.74 $&$ 0.04 $&$ 2.1$\ & $0.6 - 1.0$ &$ \ 1.863 $&$ 0.08 $&$ 3.4$\ & $1.0 - 1.5$ &$ \ 0.472 $&$ 0.19 $&$ 2.4$\ & $1.5 - 2.1$ &$ \ 0.1214 $&$ 0.40 $&$ 2.5$\ & $2.1 - 3.0$ &$ \ 2.85 \cdot 10 ^{-2} $&$ 1.02 $&$ 5.5$\ & $3.0 - 4.0$ &$ \ 6.14 \cdot 10 ^{-3} $&$ 2.40 $&$ 6.1$\ & $4.0 - 5.0$ &$ \ 1.55 \cdot 10 ^{-3} $&$ 4.94 $&$ 5.4$\ & $5.0 - 6.3$ &$ \ 4.401 \cdot 10 ^{-4} $&$ 10.49 $&$ 12.8$\ & $6.3 - 7.9$ &$ \ 1.11 \cdot 10 ^{-4} $&$ 24.17 $&$ 13.6$\ -- --------------- ----------------------------------------- ---------- --------- $1\!/\!N \cdot {\rm d}n/{\rm d}p_T^\,$ stat. sys. $ 0.0 - 0.3$ $\ 5.975$ $0.04$ $1.5$ $0.3 - 0.6$ $\ 7.16$ $0.04$ $2.0$ $0.6 - 1.0$ $\ 2.517$ $ 0.07$ $1.7$ $1.0 - 1.5$ $\ 0.742$ $0.15$ $2.4$ $1.5 - 2.1$ $\ 0.2317$ $0.29$ $2.3$ $2.1 - 3.0$ $\ 6.82 \cdot 10 ^{-2}$ $0.64$ $3.4$ $3.0 - 4.0$ $ \ 1.848 \cdot 10 ^{-2}$ $1.30$ $3.6$ $4.0 - 5.0$ $ \ 5.83 \cdot 10 ^{-3}$ $2.27$ $ 5.2$ $5.0 - 6.3$ $\ 2.00 \cdot 10 ^{-3}$ $4.45$ $2.5$ $ 6.3 - 7.9$ $\ 5.88 \cdot 10 ^{-4}$ $8.91$ $2.9 $ $7.9 - 10.0$ $\ 1.942 \cdot 10 ^{-4}$ $ 17.77$ $5.3$ $ 0.0 - 0.3$ $\ 5.824$ $0.05$ $ 1.5 $ $ 0.3 - 0.6$ $\ 6.77$ $0.04$ $1.7 $ $ 0.6 - 1.0$ $\ 2.332$ $ 0.08$ $2.2$ $ 1.0 - 1.5$ $\ 0.657$ $ 0.16$ $1.8 $ $ 1.5 - 2.1$ $ \ 0.1948$ $ 0.31$ $ 2.3 $ $ 2.1 - 3.0$ $ \ 0.053 8$ $ 0.74$ $3.0 $ $ 3.0 - 4.0$ $\ 1.297 \cdot 10 ^{-2}$ $1.56$ $ 5.9 $ $ 4.0 - 5.0$ $\ 3.94 \cdot 10 ^{-3}$ $2.90$ $4.5 $ $ 5.0 - 6.3$ $\ 1.33 \cdot 10 ^{-3}$ $5.80$ $ 6.9 $ $ 6.3 - 7.9$ $\ 3.40 \cdot 10 ^{-4}$ $12.91$ $ 6.3 $ $ 7.9 - 10.0$ $\ 7.7 \cdot 10 ^{-5}$ $30.96$ $8.6 $ $ 0.0 - 0.3$ $\ 5.12$ $ 0.05$ $2.2$ $ 0.3 - 0.6$ $\ 5.65$ $0.04$ $ 2.4 $ $ 0.6 - 1.0$ $\ 1.946$ $ 0.08$ $ 2.0$ $ 1.0 - 1.5$ $\ 0.538 $ $0.17$ $2.2 $ $ 1.5 - 2.1$ $\ 0.1481$ $0.35$ $ 3.9 $ $ 2.1 - 3.0$ $\ 3.89 \cdot 10 ^{-2}$ $0.85 $ $3.6$ $ 3.0 - 4.0$ $\ 8.643 \cdot 10 ^{-3}$ $ 1.94$ $3.4$ $ 4.0 - 5.0$ $\ 2.10 \cdot 10 ^{-3}$ $3.93 $ $4.7$ $ 5.0 - 6.3$ $\ 6.43 \cdot 10 ^{-4}$ $8.90$ $ 12.9$ $ 6.3 - 10.0$ $\ 7.7 \cdot 10 ^{-5}$ $46.70$ $10.6$ -- --------------- ----------------------------------------- ---------- --------- -- -------------- ----------------------------------------- ---------- --------- $1\!/\!N \cdot {\rm d}n/{\rm d}p_T^\,$ stat. sys. $0.0 - 0.3$ $\ 6.25 $ $0.04$ $1.6 $ $0.3 - 0.6$ $\ 7.32 $ $0.04$ $1.7 $ $0.6 - 1.0$ $\ 2.673$ $0.07$ $ 2.1$ $1.0 - 1.5$ $\ 0.851 $ $0.13$ $2.5$ $1.5 - 2.1$ $\ 0.2898$ $0.24$ $2.4 $ $2.1 - 3.0$ $\ 9.21 \cdot 10 ^{-2}$ $0.52$ $2.9$ $3.0 - 4.0$ $ \ 2.790 \cdot 10 ^{-2}$ $1.01$ $2.2$ $4.0 - 5.0$ $\ 9.30 \cdot 10 ^{-3}$ $ 1.75$ $3.0$ $5.0 - 6.3$ $\ 3.28 \cdot 10 ^{-3}$ $3.37$ $3.4$ $6.3 - 7.9 $ $\ 9.64 \cdot 10 ^{-4}$ $6.79$ $4.2$ $7.9 - 10.0$ $\ 3.06 \cdot 10 ^{-4}$ $14.41$ $3.9$ $0.0 - 0.3$ $ \ 5.652 $ $0.04$ $1.5 $ $0.3 - 0.6$ $\ 6.052 $ $ 0.04 $ $1.6 $ $0.6 - 1.0$ $ \ 2.244 $ $ 0.06 $ $1.7 $ $1.0 - 1.5$ $ \ 0.697 $ $ 0.13$ $1.8 $ $1.5 - 2.1$ $ \ 0.2323$ $ 0.24$ $1.9$ $ 2.1 - 3.0$ $ \ 7.12 \cdot 10 ^{-2}$ $0.52 $ $2.1 $ $3.0 - 4.0$ $ \ 2.025 \cdot 10 ^{-2}$ $1.05$ $1.9 $ $4.0 - 5.0$ $ \ 6.32 \cdot 10 ^{-3}$ $1.92$ $3.9 $ $5.0 - 6.3$ $ \ 2.20 \cdot 10 ^{-3}$ $3.87$ $4.0 $ $6.3 - 7.9 $ $ \ 5.91 \cdot 10 ^{-4}$ $8.42 $ $6.0$ $7.9 - 10.0$ $ \ 1.38 \cdot 10 ^{-4}$ $19.37 $ $20.1$ -- -------------- ----------------------------------------- ---------- --------- (18.,8.5)(0.0,0.0) (-0.0,0.)[![Charged particle density as a function of $\eta^$ for (a) and for (b) compared to [Rapgap]{} predictions with different proton PDFs. The predictions are obtained using the ALEPH tune.[]{data-label="fig:eta-PDF-RAPGAP"}](d13-012f3a.eps "fig:"){height="8.8cm" width="8.45cm"}]{} (8.0,0.)[![Charged particle density as a function of $\eta^$ for (a) and for (b) compared to [Rapgap]{} predictions with different proton PDFs. The predictions are obtained using the ALEPH tune.[]{data-label="fig:eta-PDF-RAPGAP"}](d13-012f3b.eps "fig:"){height="8.8cm" width="8.45cm"}]{} (0.6,8.2)[[ ** ]{}]{} (8.7,8.2)[[ ** ]{}]{} (18.,8.5)(0.0,0.0) (-0.0,0.)[![Charged particle density as a function of $\eta^$ for (a) $p_T^< 1$ GeV for (b) compared to [Rapgap]{} predictions for three different sets of fragmentation parameters. The predictions are obtained using CTEQ6L(LO) PDF.[]{data-label="fig:eta-aleph-prof-def"}](d13-012f4a.eps "fig:"){height="8.8cm" width="8.45cm"}]{} (8.0,0.)[![Charged particle density as a function of $\eta^$ for (a) $p_T^< 1$ GeV for (b) compared to [Rapgap]{} predictions for three different sets of fragmentation parameters. The predictions are obtained using CTEQ6L(LO) PDF.[]{data-label="fig:eta-aleph-prof-def"}](d13-012f4b.eps "fig:"){height="8.8cm" width="8.45cm"}]{} (0.6,8.2)[[ ** ]{}]{} (8.7,8.2)[[ ** ]{}]{} (18.,9.)(0.0,0.0) (-0.0,0.)[![Charged particle density as a function of $\eta^$ for (a) $p_T^< 1$ GeV for (b) compared to [Djangoh]{}, [Rapgap]{}, Herwig++ and [Cascade]{} Monte Carlo predictions.[]{data-label="fig:eta-PS"}](d13-012f5a.eps "fig:"){height="8.8cm" width="8.45cm"}]{} (8.0,0.)[![Charged particle density as a function of $\eta^$ for (a) $p_T^< 1$ GeV for (b) compared to [Djangoh]{}, [Rapgap]{}, Herwig++ and [Cascade]{} Monte Carlo predictions.[]{data-label="fig:eta-PS"}](d13-012f5b.eps "fig:"){height="8.8cm" width="8.45cm"}]{} (0.6,8.2)[[ ** ]{}]{} (8.7,8.2)[[ ** ]{}]{} th (17.,18.)(0.0,0.0) ![Charged particle density as a function of $\eta^$ for $p_T^< 1$ GeV for eight intervals of $Q^{2}$ and $x$ compared to [Djangoh]{}, [Rapgap]{}, Herwig++ and [Cascade]{} Monte Carlo predictions.[]{data-label="fig:eta-soft"}](d13-012f6.eps "fig:"){height="18.cm" width="17.cm"} (17.,18.)(0.0,0.0) ![Charged particle density as a function of $\eta^$ for $1< p_T^< 10$ GeV for eight intervals of $Q^{2}$ and $x$ compared to [Djangoh]{}, [Rapgap]{}, Herwig++ and [Cascade]{} Monte Carlo predictions.[]{data-label="fig:eta-hard"}](d13-012f7.eps "fig:"){height="18.cm" width="17.cm"} (18.,10.5)(0.0,0.0) (-0.3,0.)[![Charged particle density as a function of $p_T^$ in the ranges (a) $0 <\eta^< 1.5$ and (b) $1.5 <\eta^< 5$ compared to [Djangoh]{}, [Rapgap]{}, Herwig++ and [Cascade]{} Monte Carlo predictions. The ratios of MC predictions to the measurements are shown on the bottom of the figure.[]{data-label="fig:pt-mcdata-cen-RDC"}](d13-012f8a.eps "fig:"){height="10.4cm" width="9.4cm"}]{} (7.8,0.)[![Charged particle density as a function of $p_T^$ in the ranges (a) $0 <\eta^< 1.5$ and (b) $1.5 <\eta^< 5$ compared to [Djangoh]{}, [Rapgap]{}, Herwig++ and [Cascade]{} Monte Carlo predictions. The ratios of MC predictions to the measurements are shown on the bottom of the figure.[]{data-label="fig:pt-mcdata-cen-RDC"}](d13-012f8b.eps "fig:"){height="10.4cm" width="9.4cm"}]{} (0.6,9.2)[[ ** ]{}]{} (8.7,9.2)[[ ** ]{}]{} (17.,18.)(0.0,0.0) ![Charged particle density as a function of $p_T^$ in the range $0 <\eta^< 1.5$ for eight intervals of $Q^{2}$ and $x$ compared to [Djangoh]{}, [Rapgap]{}, Herwig++ and [Cascade]{} Monte Carlo predictions.[]{data-label="fig:pt-bins-cen"}](d13-012f9.eps "fig:"){height="18.cm" width="17.cm"} (17.,18.)(0.0,0.0) ![Charged particle density as a function of $p_T^$ in the range $1.5 <\eta^< 5$ for eight intervals of $Q^{2}$ and $x$ compared to [Djangoh]{}, [Rapgap]{}, Herwig++ and [Cascade]{} Monte Carlo predictions.[]{data-label="fig:pt-bins-curr"}](d13-012f10.eps "fig:"){height="18.cm" width="17.cm"} [^1]: The purity is defined as the ratio of the number of charged particles generated and reconstructed in a given bin to the total number of charged particles in the phase space of the analysis which are reconstructed in this bin.	{ "pile_set_name": "ArXiv" }
--- abstract: 'The global functional brain network (graph) is more suitable for characterizing brain states than local analysis of the connectivity of brain regions. Therefore, graph-theoretic approaches are the natural methods to study the brain. However, conventional graph theoretical analyses are limited due to the lack of formal statistical methods for estimation and inference for random graphs. For example, the concept of correlation between two vectors of graphs is yet not defined. The aim of this article to introduce a notion of correlation between graphs. In order to develop a framework to infer correlation between graphs, we assume that they are generated by mathematical models and that the parameters of the models are our random variables. Then, we define that two vectors of graphs are independent whether their parameters are independent. The problem is that, in real world, the model is rarely known, and consequently, the parameters cannot be estimated. By analyzing the graph spectrum, we showed that the spectral radius is highly associated with the parameters of the graph model. Based on it, we constructed a framework for correlation inference between graphs and illustrate our approach in a functional magnetic resonance imaging data composed of 814 subjects comprising 529 controls and 285 individuals diagnosed with autism spectrum disorder (ASD). Results show that correlations between default-mode and control, default-mode and somatomotor, and default-mode and visual sub-networks are higher ($p<0.05$) in ASD than in controls.' author: - \| André Fujita, Daniel Yasumasa Takahashi,\ Joana Bisol Balardin, and João Ricardo Sato bibliography: - 'main.bib' title: Correlation between graphs with an application to brain networks analysis --- Introduction {#sec:intro} ============ Cutting-edge brain mapping techniques such as functional magnetic resonance imaging (fMRI) generate huge amounts of datasets that allows the construction of whole brain functional networks. Attempts to analyze and quantitatively characterize the structural properties of these networks are based on techniques of an emergence new field, namely complex network analysis ([@Boccaletti]; [@Newman]; [@Strogatz]). Complex network analysis originated from mathematics, more specifically in graph theory, and aims to characterize the whole brain networks with a few number of measures. In this approach, a brain network is represented by a graph, in which its vertices represent the brain regions of interest (ROI), and edges represent the functional associations between ROIs (e.g. functional connectivity). Various graph-theoretic metrics can be used to investigate the mechanisms underlying the functional brain networks. Some examples are measures of functional integration, network motifs, centrality, and network resilience ([@Rubinov]). The analysis of the structural properties of the graphs allow us to visualize and understand the overall connectivity pattern of ROIs and also to quantitatively characterize its organization. These approaches became more popular over the last decade after it has provided an essential framework to elucidate the relationship between brain structure and function, and also to have proven by an increasing number of studies to give insights regarding the potential mechanisms involved in aging ([@Perry2015]), sex differences ([@Ingalhalikar2014]), various brain disorders ([@Stam2014]), and structural reconfiguration of the brain in response to external task modulation ([@Sherwin2015]). Although applications of methods developed in graph theory have been successful in the analysis of brain networks as aforementioned, there is still a gap between these graph-based computational approaches and Statistics. For example, to the best of our knowledge, the concept of correlation between graphs is unknown. The concept of correlation between graphs may aid the understanding of how brain sub-networks interact and also to identify differences in those interactions between controls and patients (subjects diagnosed with a disorder) that may be useful for the development of novel procedures for diagnosis and prognosis. Graphs are difficult to be manipulated from a statistical viewpoint because they are not numbers, but objects composed of one set of vertices and one set of edges. By observing the graphs depicted in Figure \[figure:illustration\], it is very difficult to identify correlation between them by only analyzing their structures. Thus, to construct a framework to infer correlation, one natural idea would be to imagine that a graph is generated by a mathematical model with a set of parameters. The parameters are the random variables. Intuitively, two vectors of graphs are correlated whether the parameters (random variables) of the graph model are correlated (Figure \[figure:illustration\]). However, given two vectors of graphs, the model that generates them is rarely known, and consequently, the parameters cannot be estimated. Thus it is necessary to identify a feature of the graph that is highly associated with the parameters of the graph. In order to identify the feature that contains the information of the parameter, we investigated the spectral properties of random graphs (set of eigenvalues of its adjacency matrix). It is known that some structural properties, such as the number of walks, diameter, and cliques can be described by the spectrum of the graph ([@Mieghem]). Here, we propose to estimate the correlation between graphs by using the spectral radius (largest eigenvalue) of the graphs. Our results show that the spectral radius is highly associated with the parameters that generate the graph, and thus, it can be a good feature to calculate correlation between two graphs. ![Two vectors of perfectly correlated random graphs [A]{} and [B]{}, each one of size four. Graphs [A]{} and [B]{} are caricatural representations of two distinct brain sub-networks (e.g. somatomotor and default-mode) of four subjects. The identification of correlation by directly analyzing the structure of graphs [A]{} and [B]{} is very difficult. Notice that although they are generated by the same model (in this illustration, the graphs were generated by an Erdös-Rényi random graph model) and parameters, they are structurally different. Thus, one solution to identify correlation between graphs consists in identifying correlation between the parameters of the random graph models. \[figure:illustration\]](./Figure1.pdf){width="4in"} We illustrate the usefulness of our method by analyzing a large fMRI dataset (ABIDE - The Autism Brain Imaging Data Exchange - Consortium website - <http://fcon_1000.projects.nitrc.org/indi/abide/>) composed of 814 participants comprising 529 controls and 285 individuals with ASD. Description of the method {#sec:meth} ========================= Graph ----- A graph is a pair of sets $G=(V,E)$, where $V$ is a set of $n$ vertices ($v_{1}, v_{2}, \ldots, v_{n}$) and $E$ is a set of $m$ edges that connect two vertices of $V$. Any undirected graph $G$ with $n$ vertices can be represented by its adjacency matrix ${\bf A}^G$ with $n \times n$ elements ${\bf A}_{ij}^G$ ($i,j=1, \ldots, n$), whose value is ${\bf A}_{ij}^G = {\bf A}_{ji}^G = 1$ if vertices $v_i$ and $v_j$ are connected, and 0 otherwise. The spectrum of graph $G$ is the set of eigenvalues of its adjacency matrix ${\bf A}^G$. Thus, an undirected graph with $n$ vertices has $n$ real eigenvalues $\lambda_1 \ge \lambda_2 \ge \ldots \ge \lambda_n$. Correlation between graphs {#sec:correlation} -------------------------- Two random variables are statistically independent whether knowledge about one of them does not aid in the prediction of the other. On the other hand, if they are not independent, then the values of one of the variables can be predicted by information provided about the other. In this study, we consider that the parameters of the graph models are random variables. Thus, we assume we have $k$ independent graphs randomly generated by the same graph model, but each one with a distinct set of parameters (sampled from a probabilistic distribution). Let $k$ and $\Theta$ be the number of graphs and their parameters, respectively. Then, let ${\underset{\widetilde{}}{\theta}}^1 = \{ \theta_{1}^{1}, \theta_{2}^{1}, \ldots, \theta_{k}^{1} \}$ and ${\underset{\widetilde{}}{\theta}}^2 = \{ \theta_{1}^{2}, \theta_{2}^{2}, \ldots, \theta_{k}^{2} \}$ be two samples of random variables $\Theta^1$ and $\Theta^2$, respectively, ${{\underset{\widetilde{}}{G}}}^1({\underset{\widetilde{}}{\theta}}^1) = \{ G_{1}^{1}(\theta_{1}^{1}), G_{2}^{1}(\theta_{2}^{1}), \ldots, G_{k}^{1}(\theta_{k}^{1})\}$ and ${{\underset{\widetilde{}}{G}}}^2({\underset{\widetilde{}}{\theta}}^2) = \{ G_{1}^{2}(\theta_{1}^{2}), G_{2}^{2}(\theta_{2}^{2}), \ldots, G_{k}^{2}(\theta_{k}^{2})\}$ be two samples of random graphs constructed by using ${\underset{\widetilde{}}{\theta}}^1$ and ${\underset{\widetilde{}}{\theta}}^2$, respectively. To illustrate this concept, suppose that ${{\underset{\widetilde{}}{G}}}^1$ and ${{\underset{\widetilde{}}{G}}}^2$ are two vectors of Erdös-Rényi random graphs ([@Erdos]). An Erdös-Rényi random graph ([@Erdos]) has $n$ labeled vertices in which each pair of vertices is connected by an edge with a given probability $p$. In this case, the probability $p$ is the parameter of graph $G$. Thus, the two vectors of Erdös-Rényi random graphs can be described as ${{\underset{\widetilde{}}{G}}}^1({\underset{\widetilde{}}{p}}^1)=\{ G_{1}^{1}(p_{1}^{1}), G_{2}^{1}(p_{2}^{1}), \ldots, G_{k}^{1}(p_{k}^{1})\}$ and ${{\underset{\widetilde{}}{G}}}^2({\underset{\widetilde{}}{p}}^2) = \{ G_{1}^{2}(p_{1}^{2}), G_{2}^{2}(p_{2}^{2}), \ldots, G_{k}^{2}(p_{k}^{2})\}$. We say that random graphs ${{\underset{\widetilde{}}{G}}}^1({\underset{\widetilde{}}{\theta}}^1)$ and ${{\underset{\widetilde{}}{G}}}^2({\underset{\widetilde{}}{\theta}}^2)$ are independent if the vectors of parameters ${{\underset{\widetilde{}}{\theta}}}^1$ and ${{\underset{\widetilde{}}{\theta}}}^2$ are independent. In our example for an Erdös-Rényi random graph, we say that graphs ${{\underset{\widetilde{}}{G}}}^1({\underset{\widetilde{}}{p}}^1)$ and ${{\underset{\widetilde{}}{G}}}^2({\underset{\widetilde{}}{p}}^2)$ are correlated if the vectors of probabilities ${{\underset{\widetilde{}}{p}}}^1=\{p_{1}^{1}, \ldots, p_{k}^{1}\}$ and ${{\underset{\widetilde{}}{p}}}^2=\{p_{1}^{2}, \ldots, p_{k}^{2}\}$ are correlated. Formally, two random variables $\Theta^{1}$ and $\Theta^{2}$ with probability density functions $f(\Theta^1)$ and $f(\Theta^2)$ are independent if and only if the combined random variable $(\Theta^1, \Theta^2)$ has a joint probability density function $f(\Theta^1,\Theta^2)=f(\Theta^1) \times f(\Theta^2)$. We say that two random variables $\Theta^1$ and $\Theta^2$ are dependent if they are not independent. The test of independence between $G^1$ and $G^2$ is described as a hypothesis test as follows:\ $\text{H}_0$: $\Theta^1$ and $\Theta^2$ are independent (null hypothesis)\ $\text{H}_1$: $\Theta^1$ and $\Theta^2$ are not independent (alternative hypothesis) One simple manner to identify correlation between $\Theta^1$ and $\Theta^2$ consists in, if the graph model is known, to estimate the parameters of the graphs, and then test the probabilistic dependence between them. However, the graph model is rarely known for real world graphs. Thus, the problem consists in detecting dependence only from the observation of random graphs (and not the parameters). In other words, it is necessary to identify a feature of the graph that is highly associated with the parameters of the graph. From spectral graph theory, the largest eigenvalue ($\lambda_1$) of a graph $G$ is known as its spectral radius or index (for simplicity, we will denote the largest eigenvalue $\lambda_1$ just as $\lambda$). For several random graphs, it is known that the spectral radius is a function of the parameters of the graph. For example, for the Erdös-Rényi random graph, let $n$ and $p$ be the number of vertices and the probability that two vertices are connected by an edge, respectively. Then, the spectral radius of an Erdös-Rényi random graph is $np$. Thus, we propose to use the spectral radius to identify correlation between graphs. Let ${{\underset{\widetilde{}}{G}}}^1 = \{ G_{1}^{1}, G_{2}^{1}, \ldots, G_{k}^{1},\}$ and ${{\underset{\widetilde{}}{G}}}^2 = \{ G_{1}^{2}, G_{2}^{2}, \ldots, G_{k}^{2},\}$ be two samples of random graphs and ${\underset{\widetilde{}}{\lambda}}^1 = \{ \lambda_{1}^{1}, \lambda_{2}^{1}, \ldots, \lambda_{k}^{1} \}$ and ${\underset{\widetilde{}}{\lambda}}^2 = \{ \lambda_{1}^{2}, \lambda_{2}^{2}, \ldots, \lambda_{k}^{2} \}$ be the spectral radii associated with ${{\underset{\widetilde{}}{G}}}^1$ and ${{\underset{\widetilde{}}{G}}}^2$, respectively. Thus, to identify correlation between graphs, one may test the independence between ${\underset{\widetilde{}}{\lambda}}^1$ and ${\underset{\widetilde{}}{\lambda}}^2$. ### Identification of the correlation between graphs {#section:spearman_correlation} Once defined the feature to be used to identify the correlation between graphs, then it is necessary to estimate the correlation itself. We propose the use of the Spearman’s rank correlation ($\rho$) because: (i) its implementation is simple; (ii) it is robust to outliers; and (iii) it does not require assumptions of linearity in the relationship between variables (it can identify monotonic nonlinear associations), nor the variables should be measured at interval scales, as it can be used for ordinal variables ([@spearman1904general]). Let $\hat{\rho}$ be the sample Spearman’s rank correlation coefficient. To estimate $\hat{\rho}$, first convert the raw values of $\lambda^1_i$ and $\lambda^2_i$ ($i=1, \ldots, k$) to ranks, and calculate the differences $d_i$ between the ranks of $\lambda^1_i$ and $\lambda^2_i$. Then, calculate the Spearman’s rank correlation coefficient ($\hat{\rho}$) as: = 1- where $\lim_{k \rightarrow \infty}\mathbb{E}[\hat{\rho}({\underset{\widetilde{}}{\lambda}}^1, {\underset{\widetilde{}}{\lambda}}^2)] := \rho(\lambda^1, \lambda^2) = 12 \mathbb{E}[F_1(\lambda^1)F_2(\lambda^2)] - 3$. Observe that if $\lambda^1$ and $\lambda^2$ are independent, then $\rho(\lambda^1, \lambda^2) = 0$. The main idea of this article is that in several cases $12 \mathbb{E}[F_1(\lambda^1)F_2(\lambda^2)] - 3 = 0$ if and only if $12 \mathbb{E}[F_1(\Theta^1)F_2(\Theta^2)] - 3 = 0$ therefore we can use $\hat{\rho}({\underset{\widetilde{}}{\lambda}}^1, {\underset{\widetilde{}}{\lambda}}^2)$ to estimate $12 \mathbb{E}[F_1(\Theta^1)F_2(\Theta^2)] - 3$. We can prove the validity of this idea in a simple case. Denote by $F$ the joint probability distribution for $(\Theta^1, \Theta^2)$ and the marginals for $\Theta^1$ and $\Theta^2$ by $F_1$ and $F_2$, respectively. Let $F$ be differentiable on both coordinates. Given i.i.d. copies $(\Theta^1_i, \Theta^2_i)_{i = 1,\ldots, k}$ of $(\Theta^1, \Theta^2)$, let $(G^1_i(\Theta^1_i))_{i = 1,\ldots, k}$ and $(G^2_i(\Theta^2_i))_{i = 1,\ldots, k}$ be independent ER random graphs of size $n$. Then, for any positive $\epsilon, \delta$ there exist an integer $k_0$ such that for all $k > k_0$ and $n > n_0(k)$ we have with probability larger than $1-\delta$ that $$\left\|\hat{\rho}({\underset{\widetilde{}}{\lambda}}^1, {\underset{\widetilde{}}{\lambda}}^2) - 12 \mathbb{E}[F_1(\Theta^1)F_2(\Theta^2)] +3\right\| \leq \epsilon$$ Let $\mathbb{P}$ be the joint probability measure for the sequences $(\Theta^1_i)_{i \geq 1}$, $(\Theta^2_i)_{i \geq 1}$, $(G^1_i(\Theta^1_i))_{i \geq 1}$, and $(G^2_i(\Theta^2_i))_{i \geq 1}$. To prove the proposition, it is enough to show that for suitable choices of $k$ and $n$, we have $$\label{eq:boundeig} \mathbb{P}\left(\left\|\hat{\rho}({\underset{\widetilde{}}{\lambda}}^1, {\underset{\widetilde{}}{\lambda}}^2) - \hat{\rho}({\underset{\widetilde{}}{\Theta}}^1, {\underset{\widetilde{}}{\Theta}}^2) \right\| > \epsilon/2\right)< \delta/2$$ and $$\label{eq:boundrho} \mathbb{P}\left(\left\|\hat{\rho}({\underset{\widetilde{}}{\Theta}}^1, {\underset{\widetilde{}}{\Theta}}^2) - 12 \mathbb{E}[F_1(\Theta^1)F_2(\Theta^2)] +3\right\| > \epsilon/2 \right) < \delta/2.$$ It is a classical result (see for example [@borkowf2002computing]) that $\hat{\rho}((\Theta^1_i)_{i = 1,\ldots, k}, (\Theta^2_i)_{i = 1,\ldots, k})$ converges in probability to $12 \mathbb{E}[F_1(\Theta^1)F_2(\Theta^2)] -3$, therefore, for sufficiently large $k$, we have that holds. Now, it remains to prove that there is $n_0(k)$ such that for all $n > n_0(k)$ inequality holds. Let $\hat{r}_i^1$ and $\hat{r}_i^2$ for $i =1, \ldots, k$ be the ranks of the spectral radii of the graphs $G^1_i(\Theta^1_i)$ and $G^2_i(\Theta^2_i)$, respectively. Also, let $r_i^1$ and $r_i^2$ for $i =1, \ldots, k$ be the ranks of the $\Theta^1_i$ and $\Theta^2_i$, respectively. From the definition of Spearman correlation, it is clear that if $\hat{r}_i^l = r_i^l$ for $i=1,\ldots, k$ and $l =1,2$, we have that $$\hat{\rho}({\underset{\widetilde{}}{\lambda}}^1, {\underset{\widetilde{}}{\lambda}}^2) = \hat{\rho}({\underset{\widetilde{}}{\Theta}}^1, {\underset{\widetilde{}}{\Theta}}^2).$$ Therefore, we have that $$\begin{aligned} &\mathbb{P}\left(\left\|\hat{\rho}({\underset{\widetilde{}}{\lambda}}^1, {\underset{\widetilde{}}{\lambda}}^2) - \hat{\rho}({\underset{\widetilde{}}{\Theta}}^1, {\underset{\widetilde{}}{\Theta}}^2) \right\| > \epsilon/2\right)\\ & \leq \mathbb{P}\left(\hat{\rho}({\underset{\widetilde{}}{\lambda}}^1, {\underset{\widetilde{}}{\lambda}}^2) \neq \hat{\rho}({\underset{\widetilde{}}{\Theta}}^1, {\underset{\widetilde{}}{\Theta}}^2)\right)\\ &\leq \mathbb{P} \left(\hat{r}^l_i \neq r^l_i \;\; \text{for some} \;\; i=1,\ldots, k\;\; \text{and}\;\; l =1,2 \right)\\ & \leq \sum_{l =1}^2\sum_{i=1}^k\mathbb{P}\left( \hat{r}^l_i \neq r^l_i \right)\end{aligned}$$ To prove , using the above inequalities, it is enough to show for $l = 1,2$ and $i = 1, \ldots, k$ that $$\mathbb{P}\left( \hat{r}^l_i \neq r^l_i \right) < \frac{\delta}{4k}.$$ Let $\gamma = \min\{\| \Theta^l_i - \Theta^l_j \|: i, j = 1, \ldots, k \;\; \text{and}\; l = 1,2\}$. We have that $$\mathbb{P}\left( \hat{r}^l_i \neq r^l_i \right) \leq \mathbb{P}\left(\left\|\frac{\lambda_i^l}{n} - \Theta^l_i \right\| > \gamma/2 \right).$$ For an increasing sequence of ER random graph with parameter $p$ and their respective spectral radii $\lambda(n)$, it is well known that $\lambda(n)/n$ converges in probability to $p$ ([@ding2010spectral]), therefore, for large enough $n_0$ we have that for all $n > n_0$ $$\mathbb{P}\left(\left\|\frac{\lambda_i^l}{n} - \Theta^l_i \right\| > \gamma/2 \right) < \frac{\delta}{4k}.$$ This concludes the proof of the proposition. Spearman’s correlation coefficient ($\rho$) assumes values between -1 and 1, where $\rho$ is $+1$ in the case of a perfect monotonically increasing relationship and $-1$ in the case of a perfect monotonically decreasing relationship. In the case of imperfect monotonically dependence, $-1 < \rho < +1$, and in the case of monotonically independent random variables, $\rho=0$. Thus, the hypothesis test to identify Spearman’s correlation between two vectors of graphs can be defined as: $\text{H}_0$: $\rho=0$ (null hypothesis)\ $\text{H}_1$: $\rho\ne0$ (alternative hypothesis) The Spearman’s rank correlation coefficient under the null hypothesis can be asymptotically approximated by a Student’s t-distribution with $k-2$ degrees of freedom as ([@spearman1904general]): t= Now, suppose we have two conditions A and B and consequently, we have four graphs ${{\underset{\widetilde{}}{G}}}^1_{\text{A}}$ and ${{\underset{\widetilde{}}{G}}}^2_{\text{A}}$ in condition A, and ${{\underset{\widetilde{}}{G}}}^1_{\text{B}}$ and ${{\underset{\widetilde{}}{G}}}^2_{\text{B}}$ in condition B. We are interested in testing whether the correlation between graphs ${{\underset{\widetilde{}}{G}}}^1_{\text{A}}$ and ${{\underset{\widetilde{}}{G}}}^2_{\text{A}}$ are equal to the correlation between ${{\underset{\widetilde{}}{G}}}^1_{\text{B}}$ and ${{\underset{\widetilde{}}{G}}}^2_{\text{B}}$. In other words, we would like to test $\text{H}_0: \rho_{\text{A}} = \rho_{\text{B}}$\ $\text{H}_1: \rho_{\text{A}} \ne \rho_{\text{B}}$. This test can be performed by using the procedure developed by [@Fisher]. Let $k_\text{A}$ and $k_\text{B}$ be the number of graphs in conditions A and B, respectively. First, transform each of the two correlation coefficients as $\hat{\rho}'_{\text{A}} = \frac{1}{2}\text{log}(\frac{1+\hat{\rho}_{\text{A}}}{1-\hat{\rho}_{\text{A}}})$ and $\hat{\rho}'_{\text{B}} = \frac{1}{2}\text{log}(\frac{1+\hat{\rho}_{\text{B}}}{1-\hat{\rho}_{\text{B}}} )$. Then, calculate the test statistic as $z=\frac{\hat{\rho}'_{\text{A}} - \hat{\rho}'_{\text{B}}}{\sqrt{ \frac{1}{k_\text{A} - 3} + \frac{1}{k_\text{B}-3}}}$. Finally, compute the $p$-value for the $z$-statistic. In the present study, we used the `R` function `cor.test` with parameter `method='spearman'` (package `stats`) to compute the Spearman’s correlation test. Simulation studies {#section:simulations} ================== We carried out Monte Carlo simulations in five different random graphs to illustrate the performance of the proposed framework. Among several classes of random graphs, we describe the Erdös-Rényi random graph ([@Erdos]), random geometric graph ([@Penrose]), random regular graph ([@Meringer]), Barabási-Albert random graph ([@Barabasi]), and Watts-Strogatz random graph ([@Watts]), due to their importance to model real world events or their well known theoretical results. Random graph models ------------------- ### Erdös-Rényi random graph {#section:random} Erdös-Rényi random graphs ([@Erdos]) are one of the most studied random graphs. Erdös and Rényi defined a random graph as $n$ labeled vertices in which each pair of vertices $(v_i,v_j)$ is connected by an edge with a given probability $p$. The spectral radius of an Erdös-Rényi random graph is $np$ ([@Furedi]). The `R` function used to generate an Erdös-Rényi random graph is `erdos.renyi.game` (package `igraph`). The `igraph` package can be downloaded from the `R` website (http://www.r-project.org). ### Random geometric graph A random geometric graph (RGG) is a spatial network. An undirected graph is constructed by randomly placing $n$ vertices in some topological space $\text{R}^d$ (e.g. a unit square - $d=2$) according to a specified probability distribution (e.g. uniform distribution) and connecting two vertices by an edge if their distance (according to some metric, e.g., Euclidian norm) is smaller than a certain neighborhood radius $r$. Hence, random geometric graphs have a spatial element absent in other random graphs. The spectral radius of a random geometric graph converges almost surely to $r^{d}$ ([@Bordenave]). The `R` function used to generate a geometric random graph is `grg.game` (package `igraph`). ### Random regular graph A random regular graph is a graph where each vertex has the same number of adjacent vertices; i.e. every vertex has the same degree. A random regular graph with vertices of degree $deg$ is called a random $deg$-regular graph or random regular graph of degree $deg$ ([@Meringer]). Random regular graphs of degree at most 2 are well known: a 0-regular graph consists of disconnected vertices; a 1-regular graph consists of disconnected edges; a 2-regular graph consists of disconnected cycles and infinite chains; a 3-regular graph is known as a cubic graph. The spectral radius of a random $deg$-regular graph is $deg$ ([@Alon]). The `R` function used to generate a regular random graph is `k.regular.game` (package `igraph`). ### Barabási-Albert random graph {#section:SF-graph} Barabási-Albert random graphs proposed by [@Barabasi] have a power-law degree distribution due to vertices preferential attachment (the more connected a vertex is, the more likely it is to receive new edges). [@Barabasi] proposed the following construction: start with a small number of ($n_0$) vertices and at every time-step, add a new vertex with $m_1$ ($m_1 \le n_0$) edges that connect the new vertex to $m_1$ different vertices already present in the system. When choosing the vertices to which the new vertex connects, assume that the probability that a new vertex will be connected to vertex $v_i$ is proportional to the degree of vertex $v_i$ and the scaling exponent $p_s$ ($P(v_i) \sim degree(v_i)^{p_s}$, where $degree(v_i)$ is the number of adjacent edges of vertex $v_i$ in the current time step) which indicates the order of the proportionality ($p_s=1$ linear; $p_s=2$ quadratic and so on). Let $k_0$ be the smallest degree, the spectral radius of the Barabási-Albert random graph is of the order of $k_{0}^{1/2}n^{1/2(p_s-1)}$ ([@Dorogovtsev]). The `R` function used to generate a Barabási-Albert random graph is `barabasi.game` (package `igraph`). ### Watts-Strogatz random graph {#section:WS-graph} Watts-Strogatz random graph ([@Watts]) is a random graph that interpolates between a regular lattice and an Erdös-Rényi random graph. This random graph present small-world properties (short average path lengths, i.e., most vertices are not neighbors of one another but can be reached from every other vertex by a small number of steps) and higher clustering coefficient (the number of triangles in the graph) than Erdös-Rényi random graphs. The algorithm to construct a Watts-Strogatz random graph is as follows: [Input]{}: Let $n$, $nei$, and $p_w$ be the number of vertices, the number of neighbors (mean degree), and the rewiring probability, respectively. 1. construct a ring lattice with $n$ vertices, in which every vertex is connected to its first $nei$ neighbors ($\frac{nei}{2}$ on either side); 2. choose a vertex and the edge that connects it to its nearest neighbor in a clockwise sense. With probability $p_w$, reconnect this edge to a vertex chosen uniformly at random over the entire ring. This process is repeated by moving clockwise around the ring, considering each vertex in turn until one lap is completed. Next, the edges that connect vertices to their second-nearest neighbors clockwise are considered. As in the previous step, each edge is randomly rewired with probability $p_w$; continue this process, circulating around the ring and proceeding outward to more distant neighbors after each lap, until each edge in the original lattice has been considered once. [Output]{}: the Watts-Strogatz random graph To the best of our knowledge, the spectral radius of a Watts-Strogatz random graph is not analytically defined, but there are empirical evidences that it is a function of $p_w$ and $nei$ ([@Mieghem]). The `R` function used to generate a Watts-Strogatz random graph is `watts.strogatz.game` (package `igraph`). Simulation description {#simulation} ---------------------- We designed three simulations to evaluate: (i) whether the Spearman’s correlation between the spectral radii indeed retrieves the association between the parameters of the graph; (ii) the control of the rate of type I error and power of the method based on the spectral radius; and (iii) the performance of the Fisher’s test on the spectral radius. ### Simulation 1 {#simulation1} In order to verify whether the Spearman’s correlation coefficient between the spectral radii in fact retrieves the association between the parameters of the graph, we compared the correlation estimated directly from the parameters with the correlation obtained by analyzing the spectral radius. The design of the experiment is as follows: we set the graph model as the Erdös-Rényi random graph, the number of graphs as $k=50$ and the size of the graph as $n=100$. The parameter $p$ of the Erdös-Rényi random graphs were generated from a bivariate normal distribution with mean zero and covariance matrix $\Sigma=\begin{pmatrix} 1 & s\\ s & 1\end{pmatrix}$ with $s = -1.0, -0.9, \ldots,0, \ldots, 0.9, 1.0$. Spearman’s correlation coefficients are estimated by using the spectral radii. For each value of $s$, we repeated this procedure 30 times and compared the estimated correlation coefficient with the correlation ($s$) in fact used to generate the graphs. ### Simulation 2 {#simulation2} In order to evaluate the control of the rate of false positives under the null hypothesis and also its statistical power to identify correlation between two vectors of graphs, we constructed the following simulation study. The parameters of the graphs are generated from a bivariate normal distribution with covariance matrix $\Sigma$ ($(\Theta^1, \Theta^2) \sim N(0, \Sigma)$). In order to evaluate the control of the rate of false positives under the null hypothesis (no correlation), we set $\Sigma=\begin{pmatrix} 1 & 0\\0 & 1\end{pmatrix}$. To evaluate the power of the test, we set $\Sigma=\begin{pmatrix} 1 & 0.5\\0.5 & 1\end{pmatrix}$. Notice that the vectors of parameters ($({\underset{\widetilde{}}{\theta}}^1, {\underset{\widetilde{}}{\theta}}^2)$) must be linearly normalized in the interval $[0, 1]$. The set-up of the parameters of the graphs is as follows: - Erdös-Rényi random graph: $p := \theta$ - Random geometric graph: $r := \theta$ and $d=2$ - Random regular graph: $deg$ := $\text{integer part of } 10 \times \theta$ - Barabási-Albert random graph: $k_0=3$ and $p_s := \text{integer part of } 10\times \theta$ - Watts-Strogatz random graph: $nei=3$, and $p_w := \theta$ The number of graphs varied in $k = 20, 40, 60, 80, 100$. The size of the graphs was set to $n=50$. The vectors of spectral radii (${\underset{\widetilde{}}{\lambda}}^1$ and ${\underset{\widetilde{}}{\lambda}}^2$) were computed by using the adjacency matrix of each graph; and the Spearman’s correlation test applied on both the parameters and the spectral radii. This process was repeated 1,000 times for each number of graphs $k$ and pair of graph models. In order to evaluate and compare the power of the test between applying the correlation test on the parameters of the graph or on the spectral radii, we constructed receiver-operating characteristic (ROC) curves. The ROC curve is a bi-dimensional plot with the one minus the specificity (number of true negatives/(number of true negatives+number of false positives)) on the $x$-axis and the sensitivity (number of true positives/(number of true positives+number of false negatives)) on the $y$-axis. A curve above and further the diagonal means high power while a curve close to the diagonal means random decisions. In our case, the nominal $p$-value is on the $x$-axis and the proportion of rejected null hypothesis (the proportion of associations identified between two random variables), on the $y$-axis. ROC curves were plot (i) to verify the control of the rate of false positives; (ii) to evaluate the power of the test; and (iii) to compare the performance of the correlation estimated by using the spectral radius and the original parameter of the graph. ### Simulation 3 {#simulation3} In order to evaluate the performance of Fisher’s test on both the control of the rate of false positives under the null hypothesis and also its statistical power between two conditions A and B, we constructed ROC curves for the following experimental set-up. The graphs were constructed by using the Erdös-Rényi random graph model with the parameter $p$ generated by bivariate normal distributions with means zero and covariance matrices $\Sigma_{\text{A}}$ and $\Sigma_{\text{B}}$ for conditions A and B, respectively. In order to evaluate the control of the rate of false positives under the null hypothesis (same correlation between conditions A and B), we set $\Sigma_{\text{A}}=\Sigma_{\text{B}}=\begin{pmatrix} 1 & 0.5\\0.5 & 1\end{pmatrix}$. To evaluate the power of the test in identifying differences in correlation between conditions A and B, we set $\Sigma_{\text{A}}=\begin{pmatrix} 1 & 0.3\\0.3 & 1\end{pmatrix}$ and $\Sigma_{\text{B}}=\begin{pmatrix} 1 & 0.6\\0.6 & 1\end{pmatrix}$. The number of graphs varied in $k = 20, 40, 60, 80, 100$. The size of the graphs was set to $n=50$. This process was repeated 1,000 times for each number of graphs $k$. Results and analysis of the simulations --------------------------------------- One first natural question is, instead of using the spectral radius, may one use another feature of the graph, such as the number of edges or measures of network centrality? In order to verify whether the spectral radius is indeed better than those measures, we simulated the five random graph models described in section \[section:random\] and compared the performance between the spectral radius against other six measures, namely transitivity centrality, betweenness centrality, closeness centrality, eigenvector centrality, degree centrality, and assortativity. The experimental set-up is as follows. The parameters of the graphs were generated by uniform distributions: Erdös-Rényi random graph $p \sim U(0,1)$, random geometric graph $r \sim U(0,1)$, random regular graph $k \sim \text{integer part of }U(1,10)$, Barabási-Albert random graph $p_s \sim \text{integer part of }U(1,4)$, Watts-Strogatz random graph $p_w \sim U(0,1)$. The number of graphs is set to $k=30$. The number of vertices of the graph varied in $n=25, 50, 75, 100$. Then, we calculated the Spearman’s correlation coefficient between the original parameter used to generate the graph and the feature (the spectral radius, transitivity centrality, betweenness centrality, closeness centrality, eigenvector centrality, degree centrality, and assortativity). This process was repeated 100 times for each $n$ and graph model. The correlation calculated here measures how much information regarding the parameter is represented in the feature. In other words, it measures how well the feature describes the parameter. ![Simulation study to select the most suited feature of the graph to be used in the identification of correlation between vectors of graphs. The $x$-axis and $y$-axis represent the size of the graph $n$ (number of vertices) and the average Spearman’s correlation coefficient between the original parameter and the feature. The error bars represent the 95% confidence interval. $\lambda$: spectral radius; TC: transitivity centrality; BC: betweenness centrality; CC: closeness centrality; EVC: eigenvector centrality; DC: degree centrality; AS: assortativity. Notice that for Erdös-Rényi, geometric, and $k$-regular graphs, spectral radius, transitivity, and closeness centralities are the best choices. For Albert-Barabási random graph model, the spectral radius and the closeness centrality are the best features. However, for Watts-Strogatz random graph model, the spectral radius is the most correlated to the parameter ($p_w$) of the graph. \[figure:best\_feature\]](./Figure2.pdf){width="4in"} Figure \[figure:best\_feature\] illustrates the average Spearman’s correlation coefficient and the 95% confidence interval between the actual parameter used to generate the graphs and the features. For Erdös-Rényi random graph, the spectral radius, transitivity, and closeness centralities presented the highest associations with the parameter $p$. For geometric, $k$-regular, and Barabási-Albert random graph models, the spectral radius and the closeness centrality are the features that better represent the actual parameter. For Watts-Strogatz random graph model, the spectral radius is the most correlated to the parameter of the graph. By combining all these results, we conclude that the spectral radius is the one that contains the highest information regarding the parameters and consequently is the most suited feature to be used to identify correlation between graphs. The quite high performance of closeness centrality can be explained by the fact that for Erdös-Rényi, geometric, $k$-regular, and Barabási-Albert random graph models, the parameters of the graph are associated with the number of edges. Notice that highly connected graphs tends to present higher closeness centrality. On the other hand, the parameter $p_w$ of the Watts-Strogatz random graph model represents the rewiring probability of the edges. In other words, what vary along Watts-Strogatz random graphs is their structure (connectivity), and not the number of edges. Thus, the performance of closeness centrality becomes poor and not adequate to identify correlation when the structure of the graph is modified without altering the number of edges. Some features such as the degree centrality and assortativity for the $k$-regular random graph model, and the transitivity centrality for the Barabási-Albert random graph model could not be calculated. It happened because a $k$-regular graph presents the same degree for all vertices; therefore, all the vertices present the same degree centrality. For the Barabási-Albert random graph model, a few vertices present very high degree while the majority of vertices present low degree. Figure \[figure:boxplot\] represents the correlation coefficient obtained by carrying out simulation 1 (section \[simulation1\]). The $x$-axis indicates the real correlation used to generate the parameters while the boxplots on the $y$-axis indicate the correlation estimated by using the Spearman’s correlation on the spectral radius. Notice that the correlation estimated by applying on the spectral radius ($y$-axis) is indeed monotonic (and also linear) in relation to the actual correlation between the parameters of the graph ($x$-axis). In other words, the higher the correlation between the parameters, the higher is the Spearman’s correlation coefficient. Therefore, the spectral radius is in fact a good feature to identify correlation between graphs. ![Boxplots to investigate the monotonicity of the correlation between spectral radii and the real correlation between the parameters used to generate the graphs. The $x$-axis indicates the real correlation used to generate the parameters while the $y$-axis indicates the correlation estimated by the Spearman’s correlation using the spectral radius. Notice that the estimated correlation is monotonic (and linear) in relation to the actual correlation between the parameters of the graphs. \[figure:boxplot\]](./Figure3.pdf){width="4in"} Figures \[figure:roc-H0\] and \[figure:roc-H1\] describe the ROC curves for the correlation between different classes of graph models (Erdös-Rényi, geometric, regular, Barabási-Albert, and Watts-Strogatz) under the null and alternative hypotheses, respectively, in 1,000 repetitions. For further details regarding the design of this experiment, refer to simulation 2 described in section \[simulation2\]. The panels on the upper triangle represent the ROCs curves obtained by using the spectral radius. The panel on the lower triangle represents the ROC curve (reference ROC curve) obtained by using the original parameters of the graph. By analyzing the ROC curves, it is possible to notice at least that: (i) the ROC curves under the null hypothesis are in the diagonal (Figure \[figure:roc-H0\]), i.e., the statistical test is effectively controlling the rate of false positives (the proportion of rejected null hypothesis is as expected by the $p$-value threshold); (ii) the power of the test increases as the number of graphs ($k$) increases (Figure \[figure:roc-H1\]); and (iii) ROC curves obtained by applying Spearman’s correlation in the spectral radius (ROC curves in the upper triangle of Figure \[figure:roc-H0\] and \[figure:roc-H1\]) are similar to the one obtained by applying on the original parameters of the graphs (ROC curves in the lower triangle (reference ROC curve) of Figure \[figure:roc-H0\] and \[figure:roc-H1\]). These simulation studies show that, in fact, it is possible to retrieve the underlying correlation between the parameters of the graphs by analyzing the spectral radius of their adjacency matrices, at least, for these five random graph models. ![On the upper triangle, ROC curves constructed based on the correlations estimated from spectral radii. On the lower triangle, the reference ROC curve constructed based on the parameter of the graph. The $x$-axis represents the $p$-value’s threshold and the $y$-axis represents the proportion of rejected null hypothesis in 1,000 repetitions. The different types of line (solid and dashed) represent the number of graphs ($k=20, 40, 60, 80, 100$) used in each repetition. Notice that all lines are in the diagonal, i.e., the statistical test is indeed controlling the rate of false positives as expected. \[figure:roc-H0\]](./Figure4.pdf){width="4in"} ![On the upper triangle, ROC curves constructed based on the correlations estimated from spectral radii. On the lower triangle, the reference ROC curve constructed based on the parameters of the graph. The $x$-axis represents the $p$-value’s threshold and the $y$-axis represents the proportion of rejected null hypothesis in 1,000 repetitions. The different types of line (solid and dashed) represent the number of graphs ($k=20, 40, 60, 80, 100$) used in each repetition. Notice that the greater the number of graphs ($k$), the higher is the power of the test. \[figure:roc-H1\]](./Figure5.pdf){width="4in"} Figure \[figure:fisher-roc\] panels (A) and (B) describe the ROC curves for simulation 3 (simulation to evaluate the Fisher’s test. For further details, refer to section \[simulation3\]) under the null and alternative hypotheses, respectively. By analyzing Figure \[figure:fisher-roc\]A, it is possible to notice that the test in fact controls the rate of false positives. By analyzing Figure \[figure:fisher-roc\]B, it is possible to see that the power of the test increases proportionally to the number of graphs. In summary, the Fisher’s test is indeed identifying distinct correlations between two conditions. ![Comparison of two conditions by using the Fisher’s test. The $x$-axis represents the $p$-value’s threshold and the $y$-axis represents the proportion of rejected null hypothesis in 1,000 repetitions. The different types of line (solid and dashed) represent the number of graphs ($k=20, 40, 60, 80, 100$) used in each repetition. (A) Under the null hypothesis, i.e., there is no difference in correlation between two conditions. The lines in the diagonal show that the statistical test is indeed controlling the rate of false positives. (B) The correlation between two conditions is in fact different. Notice that the greater the number of graphs ($k$), the higher is the power of the test. \[figure:fisher-roc\]](./Figure-ROC-Fisher.pdf){width="4in"} Application to Autism Spectrum Disorder dataset =============================================== Autism spectrum disorder (ASD) is a neurodevelopmental disorder usually diagnosed in early childhood. ASD etiology is complex and not completely understood ([@Ecker2015]), involving several risk factors, such as genetic, environmental, psychological, and neurobiological ([@Hallmayer; @Betancur]). It is usually diagnosed by a multidisciplinary group composed of physicians and psychologists that, through clinical interviews and tests, identify a combination of unusual behavioral characteristics and try to assess deficits in social communication, social reciprocity, and repetitive and stereotyped behaviors and interests ([@Wing]). These symptoms frequently manifest during the child’s first three years and are accompanied by developmental differences in brain anatomy, functioning, and functional brain connectivity. Current studies suggest that ASD is a disorder of brain systems ([@Wass2011; @Stevenson2012; @Just2012; @Frith2003]) and that anatomical abnormalities are subtle but widespread over the brain ([@Ecker2013]). Thus, one straightforward approach to enhance our comprehension of neural substrates of this disorder is to investigate differences in brain connectivity when compared to controls. In this context, most studies focus on finding differences between region-to-region functional connectivity or in vertex centrality measures. Due to the lack of a suitable methodological framework, investigations in how the structural organization in one brain sub-network is associated with the organization of another sub-network is scarce. Moreover, the description of these “correlations” among sub-networks in clinical populations remains unexplored. In the current study, we establish a novel framework to define correlation between graph structures and illustrate the usefulness of this method by enhancing our comprehension on the neurobiology of ASD. Dataset description ------------------- A large resting state fMRI dataset initially composed of 908 individuals comprising controls and subjects diagnosed with ASD was downloaded from the ABIDE Consortium website (<http://fcon_1000.projects.nitrc.org/indi/abide/>). The ABIDE dataset is fully anonymized in compliance with the HIPAA Privacy Rules and the 1000 Functional Connectomes Project/INDI protocols. Protected health information is not included in this dataset. Further details can be obtained from the ABIDE Consortium website. The pre-processing of the imaging data was performed using the Athena pipeline downloaded from (<http://www.nitrc.org/plugins/mwiki/index.php/neurobureau:AthenaPipeline>). The 351 regions of interest (ROIs) considered as the vertices of the brain network were defined by the CC400 atlas ([@Craddock]). A total of 35 ROIs including the ventricles were identified by using the MNI atlas and removed, resulting 316 ROIs for the construction of brain networks. The average time series within the ROIs were considered as to be the region representatives. Subject’s head movement during magnetic resonance scanning was treated by using the “scrubbing” procedure described by [@Power]. Individuals with a number of adequate scans less than 100 after the “scrubbing” were discarded, which resulted in 814 subjects for subsequent analyses. Thus, the dataset used in this study was composed of 529 controls (430 males, mean age $\pm$ standard deviation, $17.47 \pm 7.81$ years) and 285 ASD (255 males, $17.53 \pm 7.13$ years). Brain functional networks ------------------------- The schema of the entire fMRI data analysis can be seen in Figure \[figure:schema\]. A brain functional network can be modeled as a graph, i.e., a pair of sets $G=(V,E)$, in which $V$ is the set of regions of interest - ROIs (vertices), and $E$ is a set of functional connectivity (edges) connecting the ROIs. In the current study, the functional connectivity between two ROIs was obtained by calculating the Spearman’s correlation coefficient between ROIs $i$ and $j$ ($i,j = 1, \ldots, 316$) for each individual $q=1, \ldots, 814$. Thus, a brain functional network $G^q$ with 316 ROIs can be represented by its adjacency matrix ${\bf A}^q$ with $316 \times 316$ elements ${\bf A}^q_{ij}$ containing the connectivity (Spearman’s correlation coefficient) between the ROIs $i$ and $j$ ($i,j=1,\ldots, 316$; $q=1, \ldots, 814$). ![General pipeline schema of the fMRI data analysis. Raw fMRI data is pre-processed by using the Athena pipeline. Head movement effects are removed by using the “scrubbing” procedure. Functional networks were constructed by estimating the Spearman’s correlation among ROIs and site effects removed by a generalized linear model. The number of functional brain sub-networks (clusters) were estimated by the slope statistic and the sub-networks were obtained by applying the spectral clustering algorithm. Finally, the correlation among sub-networks were estimated by the Spearman’s correlation applied on the spectral radii. \[figure:schema\]](./Figure-schema.pdf){width="4in"} Site effects were modeled with a generalized linear model (GLM), i.e., with the site as a categorical predictor variable and the correlation coefficient as the response variable. The residuals of the model were used for subsequent analyses as the connectivity filtered by the site effect. $P$-values for each Spearman’s correlation coefficient between ROIs $i$ and $j$ were calculated and corrected for the false discovery rate (FDR) ([@Benjamini]). To estimate the number of sub-networks and also to identify the sub-networks themselves in a data-driven manner, we applied the slope criterion ([@Fujita]) and the spectral clustering algorithm ([@Ng]), respectively, on the average connectivity matrix (the average of the $z$-values associated with the $p$-values) taking into account the entire dataset. The application of the spectral clustering resulted in five (estimated by the slope statistic) well defined sub-networks namely somatomotor, visual, default-mode, cerebellar, and control, depicted in Figure \[figure:cluster\]. ![Brain functional sub-networks. The ROIs were clustered by the spectral clustering algorithm. The number of sub-networks was estimated as five by the slope statistic. Each sub-network is represented by a different color, namely somatomotor (green), visual (blue), default-mode (purple), cerebellar (red), and control (orange). R: right; L: left. \[figure:cluster\]](./Figure6.pdf){width="4in"} To obtain the adjacency matrix that represents the brain functional sub-network of each individual, we set ${\bf A}^{q}_{ij} = 1$ if the $p$-value corrected for the FDR is less than 0.05, and ${\bf A}^{q}_{ij} = 0$, otherwise. Notice that the Spearman’s correlation test is not used as a statistical test to identify correlation between two ROIs but only as an objective criterion to construct the adjacency matrix of the graph. Then, we calculated the spectral radius for each sub-network of each subject. Thus, we obtained five vectors (one for each sub-network) of size 529 and other five vectors of size 285, for controls and ASD, respectively. The 10 correlations among all the five sub-networks were estimated by using the spectral radius only for controls (Figure \[figure:net\]). Figure \[figure:net\] shows the statistically significant ($p<0.05$ after FDR correction for multiple tests) correlations between brain sub-networks. Interestingly, all sub-networks are positively correlated among them. The thickness of the edge represents the strength of the correlation, i.e., the thicker the edge, the higher is the correlation between sub-networks. ![Correlation between control brain sub-networks. The thickness of the edge represents the correlation coefficient, i.e., the thicker the edge, the higher is the absolute value of the correlation between sub-networks. Only statistically significant correlations at a significance threshold of 0.05 (after FDR correction for multiple tests) are shown. All identified correlations are positive. \[figure:net\]](./Figure7.pdf){width="4in"} Then, to identify correlations among sub-networks that are different between controls and ASD, we carried out the statistical test developed by Fisher ([@Fisher]) (section \[section:spearman\_correlation\]). Figure \[figure:diff-net\] illustrates the results obtained by comparing the correlations between controls versus ASD in the “scrubbed” data. The “scrubbing” procedure ([@Power]) is necessary to remove head movement effects that may cause spurious results. Only correlations that are statistically different between controls and ASD are represented by edges. At the edges there are the scatterplots with the linear regression lines that fit the data (blue and red represent the spectral radii of controls and ASD, respectively) and the Spearman’s correlation coefficient for controls and ASD. Interestingly, default-mode and control, default-mode and somatomotor, and default-mode and visual systems showed statistically significant higher inter-correlation in ASD when compared to controls ($p<0.05$ after FDR correction for multiple tests). ![Correlation between brain sub-networks in the “scrubbed” data. At the edges there are represented the scatterplots with the linear regression lines that fit the data and the Spearman’s correlation coefficient for controls and ASD. Blue and red dots represent the eigenvalues for controls and ASD, respectively. Default-mode and control, default-mode and somatomotor, and default-mode and visual systems showed higher inter-correlation in ASD than in controls ($p<0.05$ after FDR correction for multiple tests) \[figure:diff-net\]](./figure-net.pdf){width="4in"} Currently, ASD is characterized as a disconnection syndrome ([@geschwind2007autism]) that affects information processing at both the local and global levels ([@hernandez2014neural]). This hypothesis has been mainly derived from studies suggesting that the wide heterogeneity of autistic symptoms and traits are highly unlikely to be due to impairments in a single system, or brain region, but instead emerge from disruptions in multiple neurocognitive systems (for a review see [@Ecker2015]). The framework for correlation inference between graphs proposed in our study therefore represents a useful and suitable method for examining brain connectivity alterations in ASD related to abnormalities in the relationships between domain-specific sub-networks in the brain. Our findings indicate a different pattern of interactions between the default-mode and several sub-networks associated with sensorymotor, visual and executive processing in ASD. Abnormalities in the connectivity between nodes of the default-mode network (DMN) has been widely investigated in ASD ([@assaf2010abnormal; @kennedy2008functional; @weng2010alterations]) giving its associations with social cognition ([@buckner2008brain]). There are also functional connectivity studies reporting ASD-related differences in motor and visual networks ([@mcgrath2012atypical; @nebel2014precentral]). Moreover, task-based connectivity studies have also reported differences in individuals with ASD in fronto-parietal nodes of the control network in response to cognitive control tasks ([@kana2007inhibitory]). While these previous studies have contributed to the characterization of the disconnection model of ASD, evidence of commonly interacting abnormalities in several distinct neurocognitive systems remain scarce. The present results therefore go beyond previous research by providing evidence for altered interactions between sub-networks in ASD ranging from sensory and motor processing to higher-order cognitive functions. The mechanisms underlying this pattern of widespread connectivity abnormalities are likely to be related to the complex nature of atypical trajectory of brain maturation in ASD, which is probably mediated by many genetic and environmental factors and their interactions ([@abrahams2010connecting]). As a result, the impact of ASD on brain anatomy, functioning and connectivity is expected to be multidimensional and observable at multiple neural systems ([@Ecker2015]). For instance, ASD prediction accuracy using a support vector machine (SVM) analytic approach ([@ecker2010describing]) was shown to be improved by the use of a combined set of different morphometric features of the cortical surface rather than a specific anatomical characteristic; moreover a spatially distributed pattern of regions instead of an isolated brain region contributed with maximal classification weights to the prediction model. It is therefore likely that the distributed patterns of functional connectivity differences reported here also reflect some of these systems-level features of ASD pathology. This result reinforces the theory that ASD is related to abnormal neurodevelopmental processes (Frith, 2003), which is spread across the whole brain. By using the proposed methodological framework of graph correlations, our main contribution was to demonstrate that these latent processes result in correlated topological organization between multiple brain systems. To the best of our knowledge, this is the first description of this feature in ASD. Final remarks ============= In this study, we used the Spearman’s rank correlation as a measure to identify dependence between graphs due to its simplicity. However, instead of Spearman’s rank correlation, other methods that identify a broader type of dependence in data can be used, such as distance correlation ([@Szekely2007]), mutual information ([@shannon2015mathematical]), Hoeffding’s D measure ([@hoeffding1948non]), and the measure proposed by Heller-Heller-Gorfine ([@Heller]). Here, we focused on the development of a method to identify correlation for undirected graphs. For directed graphs, little is known about their spectrum. Notice that the adjacency matrix of an undirected graph is not symmetric, and consequently, the eigenvalues are not real numbers. Thus, in order to develop a framework to identify correlation between undirected graphs, it is necessary to better understand how are their eigenvalues (or eventually use another approach not based on their spectrum). To the best of our knowledge, there is no general analytical equation that describes the spectral radius as a function of the parameters for all kinds of random graphs models. However, for the Erdös-Rényi random graph model we prove that the proposed approach is consistent and not biased. For the other four random graph models used in this study, we showed by simulations that the spectral radius is indeed associated with the parameters of the graph and it is also a good feature to infer correlation between graphs. This approach based on the spectral radius seems to be promising and we hope it may open opportunities to develop other formal statistical methods in graphs.	{ "pile_set_name": "ArXiv" }
--- abstract: 'In Dynamic Ensemble Selection (DES) techniques, only the most competent classifiers are selected to classify a given query sample. Hence, the key issue in DES is how to estimate the competence of each classifier in a pool to select the most competent ones. In order to deal with this issue, we proposed a novel dynamic ensemble selection framework using meta-learning, called META-DES. The framework is divided into three steps. In the first step, the pool of classifiers is generated from the training data. In the second phase the meta-features are computed using the training data and used to train a meta-classifier that is able to predict whether or not a base classifier from the pool is competent enough to classify an input instance. In this paper, we propose improvements to the training and generalization phase of the META-DES framework. In the training phase, we evaluate four different algorithms for the training of the meta-classifier. For the generalization phase, three combination approaches are evaluated: Dynamic selection, where only the classifiers that attain a certain competence level are selected; Dynamic weighting, where the meta-classifier estimates the competence of each classifier in the pool, and the outputs of all classifiers in the pool are weighted based on their level of competence; and a hybrid approach, in which first an ensemble with the most competent classifiers is selected, after which the weights of the selected classifiers are estimated in order to be used in a weighted majority voting scheme. Experiments are carried out on 30 classification datasets. Experimental results demonstrate that the changes proposed in this paper significantly improve the recognition accuracy of the system in several datasets.' author: - - bibliography: - 'report.bib' title: 'META-DES.H: a dynamic ensemble selection technique using meta-learning and a dynamic weighting approach' --- Ensemble of classifiers; dynamic ensemble selection; dynamic weighting; classifier competence; meta-Learning. Introduction ============ Multiple Classifier Systems (MCS) aim to combine classifiers in order to increase the recognition accuracy in pattern recognition systems [@kittler; @kuncheva]. MCS are composed of three phases [@Alceu2014]: (1) Generation, (2) Selection, and (3) Integration. In the first phase, a pool of classifiers is generated. In the second phase, a single classifier or a subset having the best classifiers of the pool is(are) selected. We refer to the subset of classifiers as the Ensemble of Classifiers (EoC). In the last phase, integration, the predictions of the selected classifiers are combined to obtain the final decision [@kittler]. Recent works in MCS have shown that dynamic ensemble selection (DES) techniques achieve higher classification accuracy when compared to static ones [@Alceu2014; @CruzPR; @knora]. This is specially true for ill-defined problems, i.e., for problems where the size of the training data is small and there are not enough data available to train the classifiers [@paulo2; @logid]. The key issue in DES is to define a criterion to measure the level of competence of a base classifier. Most DES techniques [@knora; @lca; @mcb; @ijcnn2011] use estimates of the classifiers’ local accuracy in small regions of the feature space surrounding the query instance, called the region of competence, as a search criterion to estimate the competence level of the base classifier. However, in our previous work [@ijcnn2011], we demonstrated that the use of local accuracy estimates alone is insufficient to provide higher classification performance. To tackle this issue, in [@CruzPR] we proposed a novel DES framework, called META-DES, in which multiple criteria regarding the behavior of a base classifier are used to compute its level of competence. The framework is based on two environments: the classification environment, in which the input features are mapped into a set of class labels, and the meta-classification environment, where different properties from the classification environment, such as the classifier accuracy in a local region of the feature space, are extracted from the training data and encoded as meta-features. With the arrival of new test data, the meta-features are extracted using the test data as reference, and used as input to the meta-classifier. The meta-classifier decides whether the base classifier is competent enough to classify the test sample. The framework is divided into three steps: (1) Overproduction, where the pool of classifiers is generated; (2) Meta-training, where the meta-features are extracted, using the training data, and used as inputs to train a meta-classifier that works as a classifier selector, and (3) the Generalization phase, in which the meta-features are extracted from each query sample and used as input to the meta-classifier to perform the ensemble selection. In this paper, we propose two improvements to the META-DES framework. First, we modify the training routine of the meta-classifier. The modification made is motivated by the fact that there is a strong correlation between the performance of the meta-classifier for the selection of “competent” classifiers, i.e., classifiers that predict the correct label for a given query sample and the classification accuracy of the DES system [@icpr2014]. Hence, we believe that the proposed META-DES framework can obtain higher classification performance by focusing only on improving the performance of the system at the meta-classification level. This is an interesting feature of the proposed system especially when dealing with ill-defined problems due to critical dataset sizes [@CruzPR]. Four different classifier models are considered for the meta-classifier: MLP Neural Network, Support Vector Machines with Gaussian Kernel (SVM), s and Naive Bayes [@delgado14a]. Secondly, we propose three combination schemes for the generalization phase of the framework: Dynamic selection, Dynamic weighting and Hybrid. In the dynamic selection approach, only the classifiers that attain a certain level of competence are used to classify a given query sample. In the dynamic weighting approach, the meta-classifier is used to estimate the weights of all base classifiers in the pool. Then, their decisions are aggregated using a weighted majority voting scheme [@kuncheva]. Thus, classifiers that attain a higher level of competence, for the classification of the given query sample, have a greater impact on the final decision. In the hybrid approach, only the classifiers that attain a certain level of competence are selected. Then, the meta-classifier is used to compute the weights of the selected base classifiers to be used in a weighted majority voting scheme. The hybrid approach is based on the observation that the selected base classifiers might be associated with different levels of competence. It is feasible that classifiers that attained a higher level of competence should have more influence for the classification of the given test sample. The proposed framework differs from mixture of expert techniques [@NguyenAM06; @mixture], since our system is based on the mechanism used for the selection of dynamic ensembles [@Alceu2014; @CruzPR] rather than static ones [@mixture]. In addition, mixture of experts techniques are dedicated to the use of neural networks as base classifier, while, in the proposed framework, any classification algorithm can be used. We evaluate the generalization performance of the system over 30 classification problems derived from different data repositories. Furthermore, the recognition performance of the system is compared against eight state-of-the-art dynamic selection techniques according to a new survey on this topic [@Alceu2014]. Experimental results demonstrate that the choice of the meta-classifier has a significant impact on the classification accuracy of the overall system. The modifications proposed in this work significantly improve the performance of the framework when compared to state-of-the-art dynamic selection techniques. This paper is organized as follows: The META-DES framework is introduced in Section \[sec:proposed\]. Experimental results are given in Section \[sec:experiments\]. Finally the conclusion is presented in the last section. The META-DES Framework {#sec:proposed} ====================== The META-DES framework is based on the assumption that the dynamic ensemble selection problem can be considered as a meta-problem. This meta-problem uses different criteria regarding the behavior of a base classifier $c_{i}$, in order to decide whether it is competent enough to classify a given test sample $\mathbf{x}_{j}$. The meta-problem is defined as follows [@CruzPR]: - The meta-classes of this meta-problem are either “competent” (1) or “incompetent” (0) to classify $\mathbf{x}_{j}$. - Each set of meta-features $f_{i}$ corresponds to a different criterion for measuring the level of competence of a base classifier. - The meta-features are encoded into a meta-features vector $v_{i,j}$. - A meta-classifier $\lambda$ is trained based on the meta-features $v_{i,j}$ to predict whether or not $c_{i}$ will achieve the correct prediction for $\mathbf{x}_{j}$, i.e., if it is competent enough to classify $\mathbf{x}_{j}$ A general overview of the META-DES framework is depicted in Figure \[fig:overview\]. It is divided into three phases: Overproduction, Meta-training and Generalization. Overproduction -------------- In this step, the pool of classifiers $C = \{c_{1}, \ldots, c_{M}\}$, where $M$ is the pool size, is generated using the training dataset $\mathcal{T}$. The Bagging technique [@bagging] is used in this work in order to build a diverse pool of classifiers. Meta-Training ------------- In this phase, the meta-features are computed and used to train the meta-classifier $\lambda$. As shown in Figure \[fig:overview\], the meta-training stage consists of three steps: sample selection, meta-features extraction process and meta-training. A different dataset $\mathcal{T}_{\lambda}$ is used in this phase to prevent overfitting. ### Sample selection We decided to focus the training of $\lambda$ on cases in which the extent of consensus of the pool is low. This decision was based on the observations made in [@docs; @paulo2] the main issues in dynamic ensemble selection occur when classifying testing instances where the degree of consensus among the pool of classifiers is low, i.e., when the number of votes from the winning class is close to or even equal to the number of votes from the second class. We employ a sample selection mechanism based on a threshold $h_{C}$, called the consensus threshold. For each $\mathbf{x}_{j,train_{\lambda}} \in \mathcal{T}_{\lambda}$, the degree of consensus of the pool, denoted by $H \left ( \mathbf{x}_{j,train_{\lambda}}, C \right )$, is computed. If $H \left ( \mathbf{x}_{j,train_{\lambda}}, C \right )$ falls below the threshold $h_{C}$, $\mathbf{x}_{j,train_{\lambda}}$ is passed down to the meta-features extraction process. ### Meta-feature extraction {#sec:metafeatures} The first step in extracting the meta-features involves computing the region of competence of $\mathbf{x}_{j,train_{\lambda}}$, denoted by $\theta_{j} = \left \{ \mathbf{x}_{1}, \ldots, \mathbf{x}_{K} \right \}$. The region of competence is defined in the $\mathcal{T_{\lambda}}$ set using the K-Nearest Neighbor algorithm. Then, $\mathbf{x}_{j,train_{\lambda}}$ is transformed into an output profile, $\tilde{\mathbf{x}}_{j,train_{\lambda}}$. The output profile of the instance $\mathbf{x}_{j,train_{\lambda}}$ is denoted by $\tilde{\mathbf{x}}_{j,train_{\lambda}} = \left\lbrace \tilde{\mathbf{x}}_{j,train_{\lambda},1}, \tilde{\mathbf{x}}_{j,train_{\lambda},2}, \ldots, \tilde{\mathbf{x}}_{j,train_{\lambda},M} \right\rbrace $, where each $\tilde{\mathbf{x}}_{j,train_{\lambda},i}$ is the decision yielded by the base classifier $c_{i}$ for the sample $\mathbf{x}_{j,train_{\lambda}}$ [@paulo2]. The similarity between $\tilde{\mathbf{x}}_{j,train_{\lambda}}$ and the output profiles of the instances in $\mathcal{T}_{\lambda}$ is obtained through the Euclidean distance. The most similar output profiles are selected to form the set $\phi_{j} = \left \{ \tilde{\mathbf{x}}_{1}, \ldots, \tilde{\mathbf{x}}_{K_{p}} \right \}$, where each output profile $\tilde{\mathbf{x}}_{k}$ is associated with a label $w_{l,k}$. Next, for each base classifier $c_{i} \in C$, five sets of meta-features are calculated: - *- Neighbors’ hard classification:* First, a vector with $K$ elements is created. For each sample $\mathbf{x}_{k}$, belonging to the region of competence $\theta_{j}$, if $c_{i}$ correctly classifies $\mathbf{x}_{k}$, the $k$-th position of the vector is set to 1, otherwise it is 0. Thus, $K$ meta-features are computed. - *- Posterior Probability:* First, a vector with $K$ elements is created. Then, for each sample $\mathbf{x}_{k}$, belonging to the region of competence $\theta_{j}$, the posterior probability of $c_{i}$, $P(w_{l}\mid \mathbf{x}_{k})$ is computed and inserted into the $k$-th position of the vector. Consequently, $K$ meta-features are computed. - *- Overall Local Accuracy:* The accuracy of $c_{i}$ over the whole region of competence $\theta_{j}$ is computed and encoded as $f_{3}$. - *- Outputs’ profile classification:* First, a vector with $K_{p}$ elements is generated. Then, for each member $\tilde{\mathbf{x}}_{k}$ belonging to the set of output profiles $\phi_{j}$, if the label produced by $c_{i}$ for $\mathbf{x}_{k}$ is equal to the label $w_{l,k}$ of $\tilde{\mathbf{x}}_{k}$, the $k$-th position of the vector is set to 1, otherwise it is set to 0. A total of $K_{p}$ meta-features are extracted using output profiles. - *- Classifier’s confidence:* The perpendicular distance between the reference sample $\mathbf{x}_{j}$ and the decision boundary of the base classifier $c_{i}$ is calculated and encoded as $f_{5}$. A vector $v_{i,j} = \left\lbrace f_{1} \cup f_{2} \cup f_{3} \cup f_{4} \cup f_{5} \right\rbrace$ is obtained at the end of the process. If $c_{i}$ correctly classifies $\mathbf{x}_{j}$, the class attribute of $v_{i,j}$, $\alpha_{i,j} = 1$ (i.e., $v_{i,j}$ belongs to the meta-class “competent”), otherwise $\alpha_{i,j} = 0$. $v_{i,j}$ is stored in the meta-features dataset $\mathcal{T}_{\lambda}^{}$ that is used to train the meta-classifier $\lambda$. ### Training The last step of the meta-training phase is the training of $\lambda$. The dataset $\mathcal{T}_{\lambda}^{}$ is divided on the basis of 75% for training and 25% for validation. In this paper, we evaluate four classifier models for the meta-classifier: MLP Neural Network, Support Vector Machines with Gaussian Kernel (SVM), Random Forests and Naive Bayes. These classifiers were selected based on a recent study [@delgado14a] that ranked the best classification models in a comparison considering a total of 179 classifiers and 121 datasets. All classifiers were implemented using the Matlab PRTOOLS toolbox [@PRTools]. The parameters of each classifier were set as follows: 1. MLP Neural Network: The validation data was used to select the number of nodes in the hidden layer. We used a configuration with 10 neurons in the hidden layer since there were no improvement in results with more than 10 neurons. The training process for $\lambda$ was performed using the Levenberg-Marquadt algorithm. The training process was stopped if its performance on the validation set decreased or failed to improve for five consecutive epochs. 2. SVM: A radial basis SVM with a Gaussian Kernel was used. For each dataset, a grid search was performed in order to set the values of the regularization parameter $c$ and the Kernel spread parameter $\gamma$. 3. Random Forest: A total of $200$ decision trees were used. The depth of each tree was fixed at $5$. 4. Naive Bayes: A simple Naive Bayes classifier using a normal distribution to model numeric features. No parameters are required for this model. Generalization {#sec:generalization} -------------- Given the query sample $\mathbf{x}_{j,test}$, the region of competence $\theta_{j}$ is computed using the samples from the dynamic selection dataset $D_{SEL}$. Following that, the output profiles $\tilde{\mathbf{x}}_{j,test}$ of the test sample, $\mathbf{x}_{j,test}$, are calculated. The set with $K_{p}$ similar output profiles $\phi_{j}$, of the query sample $\mathbf{x}_{j,test}$, is obtained through the Euclidean distance applied over the output profiles of the dynamic selection dataset, $\tilde{D}_{SEL}$. Next, for each classifier $c_{i}$ belonging to the pool of classifiers $C$, the meta-features extraction process is called, returning the meta-features vector $v_{i,j}$. Then, $v_{i,j}$ is used as input to the meta-classifier $\lambda$. The support obtained by the meta-classifier for the “competent” meta-class, denoted by $\delta_{i,j}$, is computed as the level of competence of the base classifier $c_{i}$ for the classification of the test sample $\mathbf{x}_{j,test}$. Three combination approaches are considered: - META-DES.S: In this approach, the base classifiers that achieve a level of competence $\delta_{i,j} > \Upsilon$ are considered competent, and are selected to compose the ensemble $C'$. In this paper, we set $\Upsilon = 0.5$ (i.e., the base classifier is selected if the support for the “competent” meta-class is higher than the support for the “incompetent” meta-class). The final decision is obtained using the majority vote rule [@kittler]. Tie-breaking is handled by choosing the class with the highest a posteriori probability. - META-DES.W: Every classifier in the pool $C$ is used to predict the label of $\mathbf{x}_{j,test}$. The level of competence $\delta_{i,j}$ estimated by the meta-classifier $\lambda$ is used as the weight of each base classifier. The final decision is obtained using a weighted majority vote combination scheme [@kuncheva]. Thus, the decisions obtained by the base classifiers with a higher level of competence $\delta_{i,j}$ have a greater influence on the final decision. - META-DES.H: In this approach, first the base classifiers that achieve a level of competence $\delta_{i,j} > \Upsilon = 0.5$ are considered competent and are selected to compose the ensemble $C'$. Next, the level of competence $\delta_{i,j}$ estimated by the meta-classifier $\lambda$, for the classifiers in the ensemble $C'$, are used as its weights. Thus, the decisions obtained by the base classifiers with the highest level of competence $\delta_{i,j}$ have a greater influence in the final decision. A weighting majority voting scheme is used to predict the label $w_{l}$ of $\mathbf{x}_{j,test}$. Experiments {#sec:experiments} =========== Datasets -------- A total of 30 datasets are used in the comparative experiments, with sixteen taken from the UCI machine learning repository [@Lichman2013], four from the STATLOG project [@King95statlog], four from the Knowledge Extraction based on Evolutionary Learning (KEEL) repository [@FdezFLDG11], four from the Ludmila Kuncheva Collection of real medical data [@lkc], and two artificial datasets generated with the Matlab PRTOOLS toolbox [@PRTools]. The key features of each dataset are shown in Table \[table:datasets\]. Experimental Protocol --------------------- For the sake of simplicity, the same experimental protocol used in previous publications [@CruzPR; @icpr2014] was used. The experiments were carried out using 20 replications. For each replication, the datasets were randomly divided on the basis of 50% for training, 25% for the dynamic selection dataset (D$_{SEL}$), and 25% for the test set ($\mathcal{G}$). The divisions were performed while maintaining the prior probabilities of each class. For the proposed META-DES, 50% of the training data was used in the meta-training process $\mathcal{T}_{\lambda}$ and 50% for the generation of the pool of classifiers ($\mathcal{T}$). For the two-class classification problems, the pool of classifiers was composed of 100 Perceptrons generated using the bagging technique [@bagging]. For the multi-class problems, the pool of classifiers was composed of 100 multi-class Perceptrons. The use of Perceptron as base classifier is based on the observations that the use of weak classifiers can show more differences between the DES schemes [@knora], thus making it a better option for comparing different DES techniques. Furthermore, as reported by Leo Breiman, the bagging technique achieves better results when weak and unstable base classifiers are used [@bagging]. The values of the hyper-parameters $K$, $K_{p}$ and $h_{c}$ were set at 7, 5 and 70%, respectively. They were selected empirically based on previous publications [@ijcnn2011; @icpr2014; @CruzPR]. Comparison of different classification models as the Meta-Classifier {#sec:diffClassifiers} -------------------------------------------------------------------- In this experiment, we analyze the impact of the classifier model used for the meta-problem (i.e., for the selection of competent classifiers). The objective of this experiment is to verify whether we can improve the classification performance of the META-DES system, previously defined using an MLP neural network as the meta-classifier. The following classifier models are considered: Multi-Layer Perceptron (MLP) Neural Networks as in [@CruzPR], Support Vector Machines with Gaussian Kernel (SVM), Random Forests and Naive Bayes. Table \[table:diffClassifiers\] shows a comparison of the performance of the meta-classifier $\lambda$ and the recognition accuracy obtained by the META-DES system using each classification model. The best results are highlighted in bold. For each dataset, we compared the results obtained by the meta-classifier $\lambda$ and by the META-DES framework using the MLP network [@CruzPR], against the best result obtained by any of the other classifier models (SVM, Random Forest and Naive Bayes). The comparison was performed using the Kruskal-Wallis non-parametric statistical test, with a 95% confidence interval. Results that are significantly better are marked with a $\bullet$. We can observe that when the meta-classifier achieves a recognition performance that is statistically superior for a single dataset, such as, Banana, Faults and WDGV1, for instance, the META-DES is also likely to achieve superior accuracy for the same classification problem. Figure \[fig:barClassifierModel\] shows the number of datasets that each classifier model achieved the highest accuracy. The Naive Bayes classifier is ranked first, achieving the best results for 14 datasets, followed by the MLP Neural Network with 8. SVM and Random Forests achieved the best results for 4 datasets each. The strong performance of the Naive Bayes may be explained by the fact that the majority of the meta-features are binary, and this classifier model handles well binary input features different than MLP Networks. In addition, it might indicate that the proposed sets of meta-features are possibly independent [@Cruz2014ANNPR]. This is an interesting finding since the Naive Bayes model is much faster both in the training and testing stages when compared to an MLP Neural Network or an SVM classifier. Furthermore, in order to verify whether the difference in classification results obtained over the 30 datasets is statistically significant, we performed a Wilcoxon non-parametric signed rank test with 95% confidence for a pairwise comparison between the results obtained using an MLP Neural Network against the best result obtained using a different classifier for the meta-classifier. The Wilcoxon signed rank test was used since it was suggested in [@Demsar:2006] as a robust method for comparing the classification results of two algorithms over several datasets. The results of the Wilcoxon statistical test are shown in the last row of Table \[table:diffClassifiers\]. Techniques that achieve performance equivalent to the MLP network are marked with “\~”; those that achieve statistically superior performance are marked with a “+”, and those with inferior performance are marked with a “-”. When comparing the performance of the four meta-classifiers, the results achieved using Random Forests and Naive Bayes as the meta-classifier $\lambda$ are significantly superior. Hence, we can conclude that significant gains in classification accuracy can be achieved by choosing a more suitable classifier model for the meta-classifier $\lambda$. Although the choice of the best meta-classifier may vary according to the classification problem (Table \[table:diffClassifiers\]), the results of the META-DES using Naive Bayes as the meta-classifier achieves results that are statistically superior when compared to the MLP neural network over the 30 datasets studied in this work. Comparison Between Combination Approaches: Dynamic Selection, Dynamic Weighting and Hybrid {#sec:fusion} ------------------------------------------------------------------------------------------ In this section, we compare the three combination approaches presented in Section \[sec:generalization\]: Dynamic Selection, Dynamic weighting, and the Hybrid approach. For the sake of simplicity, we present only the results obtained using the Naive Bayes as the meta-classifier $\lambda$ since it achieved the highest classification accuracy in the previous experiments (Table \[table:diffClassifiers\]). The results achieved using the Naive Bayes as meta-classifier for the three combination approaches are shown in Table \[table:classificationApproaches\]. In order to select the best combination approach, we compare the average ranks of each approach computed using the Friedman test, which is a non-parametric equivalent of the repeated measures ANOVA used to compare several algorithms over multiple datasets [@Friedman1937Use; @Demsar:2006]. The Friedman test ranks each algorithm, with the best performing one getting rank 1, the second best rank 2, and so forth for each dataset separately. The average rank is then computed, considering all datasets. Thus, the best algorithm is the one with the lowest average rank. The approaches that use the proposed weighting scheme (Dynamic weighting and Hybrid) outperformed the Dynamic selection approach in accuracy. This can be explained by the fact the outputs given by the Naive Bayes classifier can be directly interpreted as the likelihood that the base classifier belongs to the “competent” meta-class. Thus, the supports provided by the meta-classifier can directly be used as the weights of each classifier for a weighted majority voting scheme. This is different from other classification models, such as Random Forests where their class supports cannot be directly interpreted as such. Hence, the meta-classifier can also be used for the fusion (integration) of the classifiers in the ensemble, rather than only for ensemble selection. Since the Hybrid combination approach presents the highest recognition accuracy when the 30 datasets are considered (lowest average rank) this combination approach is selected for the comparison against other state-of-the-art DES techniques. Comparison with the state-of-the-art DES techniques --------------------------------------------------- In this section, we compare the recognition rates obtained by the proposed META-DES.H against eight state-of-the-art dynamic selection techniques in the DES literature: the KNORA-ELIMINATE [@knora], KNORA-UNION [@knora], DES-FA [@ijcnn2011], Local Classifier Accuracy (LCA) [@lca], Overall Local Accuracy (OLA) [@lca], Modified Local Accuracy (MLA) [@Smits_2002], Multiple Classifier Behaviour (MCB) [@mcb] and K-Nearests Output Profiles (KNOP) [@paulo2]. For all techniques, we use the same pool of classifiers defined in the previous section (Section \[sec:diffClassifiers\]) in order to have a fair comparison. The size of the region of competence (neighborhood size), $K$ is set to $7$ since it achieved the best result in previous experiments [@Alceu2014; @ijcnn2011]. The comparative results are shown in Table \[table:ResultsBayes\]. Due to size constraints, we only show the results using Naive Bayes as the meta-classifier since it achieved the highest recognition accuracy in the previous experiment. For each dataset, a Kruskal-Wallis statistical test with 95% confidence was conducted to know if the classification improvement is statistically significant. Results that are statistically better are marked with a $\bullet$. The results of the proposed technique obtained the highest accuracy in 20 out of 30 datasets. In addition, the accuracy of the proposed system was statistically superior in 15 out of 30 datasets. The original META-DES framework [@CruzPR], without the improvements proposed in this paper, achieved results that are statistically superior in 10 out of the 30 datasets when compared with the state-of-the-art DES techniques. Furthermore, we also consider the Wilcoxon test with 95% confidence, for a pairwise comparison between the classification performances of the proposed system against the performance of the state-of-the-art DES techniques over multiple datasets. The results of the Wilcoxon test are shown in the last row of the table. The performance of the proposed META-DES.H system is statistically better when all 30 datasets are considered. Hence, the experimental results demonstrate that the changes proposed in this paper lead to a significant gains in performance when compared to other DES algorithms. Conclusion ========== In this paper, we proposed two modifications to the novel META-DES framework. First, we compared different classifier models, such as the MLP Neural Network, Support Vector Machines with Gaussian Kernel (SVM), Random Forests and Naive Bayes for the meta-classifier. Next, we evaluated three combination approaches to the framework: Dynamic selection, Dynamic weighting and Hybrid. In the Dynamic selection approach, only the classifiers that attain a certain level of competence are used to classify a given query sample. In the dynamic weighting approach, all base classifiers in the pool are considered to give the final decision, with the meta-classifier estimating the weight of each base classifier. In the hybrid approach, only the classifiers that attain a certain level of competence are initially selected, after which their decisions are aggregated in a weighted majority voting scheme. Thus, the base classifiers attaining higher levels of competence have a greater impact on the final decision. Experiments were conducted using 30 classification datasets derived from five different data repositories (UCI, KEEL, STATLOG, LKC and ELENA). First, we observed a significant improvement in accuracy using different classifier models for the meta-problem. The performance of the META-DES trained using a Naive Bayes for the meta-classifier achieves results that are statistically better compared to those achieved using an MLP Neural Network, according to the Wilcoxon Signed Rank test with 95% confidence. This finding confirms the initial hypothesis that the overall performance of the system improves when the recognition accuracy of the meta-classifier improves. As the META-DES framework considers the dynamic selection problem as a meta-classification problem, we can improve the recognition accuracy by focusing only on improving the classification performance in the meta-problem. This finding is especially useful for ill-defined problems since there is not enough data to properly train the base classifiers. Techniques such as stacked generalization for the generation of more meta-feature vectors in the data generation process as well as the use of feature selection techniques to achieve a more representative set of meta-features can be considered to improve the recognition performance at the meta-classification level. In addition, we demonstrate that the framework can also be used to compute the weights of the base classifiers. We found that the Naive Bayes classifier achieved the best result when the dynamic weighting (META-DES.W) or hybrid (META-DES.H) approach is used. This can be explained by the fact that the supports given by this classifier can be seen as the likelihood that the base classifier belongs to the “competent” meta-class. Thus, the classifiers that are more likely to be “competent” have greater influence on the classification of any given test sample. When compared to eight state-of-the-art techniques found in the dynamic ensemble selection literature, the proposed META-DES.H using a Naive Bayes classifier for the meta-classifier presented classification accuracy that is statistically better in 15 out of the 30 classification datasets. The original META-DES framework [@CruzPR] achieved results that are statistically better in 10 out of the 30 datasets when compared with the state-of-the-art DES techniques. Hence, the changes to the META-DES framework proposed in this paper lead to a significant gain in performance when compared against other DES algorithms. Acknowledgement {#acknowledgement .unnumbered} =============== This work was supported by the Natural Sciences and Engineering Research Council of Canada (NSERC), the École de technologie supérieure (ÉTS Montréal), CNPq (Conselho Nacional de Desenvolvimento Científico e Tecnológico) and FACEPE (Fundação de Amparo à Ciência e Tecnologia de Pernambuco).	{ "pile_set_name": "ArXiv" }
--- abstract: 'Linear Parameter-Varying (LPV) systems with piecewise differentiable parameters is a class of LPV systems for which no proper analysis conditions have been obtained so far. To fill this gap, we propose an approach based on the theory of hybrid systems. The underlying idea is to reformulate the considered LPV system as an equivalent hybrid system that will incorporate, through a suitable state augmentation, information on both the dynamics of the state of the system and the considered class of parameter trajectories. Then, using a result pertaining on the stability of hybrid systems, two stability conditions are established and shown to naturally generalize and unify the well-known quadratic and robust stability criteria together. The obtained conditions being infinite-dimensional, a relaxation approach based on sum of squares programming is used in order to obtain tractable finite-dimensional conditions. The approach is finally illustrated on two examples from the literature.' author: - 'Corentin Briat and Mustafa Khammash[^1]' title: Stability analysis of LPV systems with piecewise differentiable parameters --- Introduction ============ Linear Parameter-Varying (LPV) systems [@Toth:10; @Mohammadpour:12; @Briat:book1] are an important class of linear systems that can be used to model linear systems that intrinsically depend on parameters [@Wu:95] or to approximate nonlinear systems in the objective of designing gain-scheduled controllers [@Shamma:88phd; @Shamma:92]; see [@Briat:book1] for a classification attempt. They have been, since then, successfully applied to a wide variety of real-world systems such as automotive suspensions systems [@Poussot:08b; @Poussot:10], robotics [@KajiwaraAG:99a], aperiodic sampled-data systems [@Robert:10], aircrafts [@Gilbert:10; @Seiler:12; @Pfifer:15], etc. The field has also been enriched with very broad theoretical results and numerical tools [@Packard:94a; @Apkarian:95; @Apkarian:95a; @Apkarian:98a; @Wu:01; @Wu:06b; @Scherer:01; @Bokor:05a; @Scherer:12; @Scherer:15; @Briat:book1; @Briat:15d; @Balas:15; @Peni:16; @Wang:16]. The objective of the paper is to extend the stability results in [@Briat:15d] for LPV systems with piecewise constant parameters to the case of piecewise differentiable parameters. Such parameter trajectories may arise whenever an (impulsive) LPV system is used to approximate a nonlinear impulsive system or in linear systems where parameters naturally have such a behavior. Finally, they can also be used to approximate parameter trajectories that exhibit intermittent very fast, yet differentiable, variations. Despite being similar to the case of piecewise constant parameters, the fact that the parameters are time-varying between discontinuities leads to additional difficulties which prevent from straightforwardly extending the approach developed in [@Briat:15d]. To overcome this, we propose a different, more general, approach based on hybrid systems theory [@Goebel:12]. Firstly, we equivalently reformulate the considered LPV system into a hybrid system that captures in its formulation both the dynamics of the state of the system and the considered class of parameter trajectories. Recent results from hybrid systems theory [@Goebel:12] combined with the use of quadratic Lyapunov functions are then applied in order to derive sufficient stability conditions for the considered hybrid systems and, hence, for the associated LPV system and class of parameter trajectories. We then prove that the obtained stability conditions generalize and unify the well-known quadratic stability and robust stability conditions, which can be recovered as extremal/particular cases. The obtained infinite-dimensional stability conditions are reminiscent of those obtained in [@Briat:13d; @Briat:14f; @Briat:15i], a fact that strongly confirms the connection between the approaches. To make these conditions computationally verifiable, we propose a relaxation approach based on sum of squares [@Putinar:93; @Parrilo:00; @sostools3] and semidefinite programming [@Sturm:01a]. The approach is finally validated through two examples. Outline. The structure of the paper is as follows: in Section \[sec:preliminary\] preliminary definitions and results are given. Section \[sec:stab\] develops the main results of the paper. Examples are finally given in Section \[sec:examples\]. Notations. The set of nonnegative integers is denoted by $\mathbb{N}_0$. The set of symmetric matrices of dimension $n$ is denoted by $\mathbb{S}^n$ while the cone of positive (semi)definite matrices of dimension $n$ is denoted by ($\mathbb{S}^n_{\succeq0}$) $\mathbb{S}_{\succ0}^n$. For some $A,B\in\mathbb{S}^n$, the notation that $A\succ(\succeq)B$ means that $A-B$ is positive (semi)definite. The maximum and the minimum eigenvalue of a symmetric matrix $A$ are denoted by $\lambda_{max}(A)$ and $\lambda_{min}(A)$, respectively. A function $\alpha:\mathbb{R}_{\ge0}\mapsto\mathbb{R}_{\ge0}$ is of class $\mathcal{K}_\infty$ (i.e. $\alpha\in\mathcal{K}_\infty$), if it is continuous, zero at zero, increasing and unbounded. A function $\beta:\mathbb{R}_{\ge0}\mapsto\mathbb{R}_{\ge0}$ is positive definite if $\beta(s)>0$ for all $s>0$ and $\beta(0)=0$. Given a vector $v\in\mathbb{R}^d$ and a closed set $\mathcal{A}\subset\mathbb{R}^d$, the distance of $v$ to the set $\mathcal{A}$ is denoted by $\|v\|_\mathcal{A}$ and is defined by $\|v\|_\mathcal{A}:=\inf_{y\in\mathcal{A}}\|v-y\|$. For any differentiable function $f(x,y)$, the partial derivatives with respect to the first and second argument evaluated at $(x,y)=(x^,y^)$ are denoted by $\partial_x f(x^,y^)$ and $\partial_y f(x^,y^)$, respectively. Preliminaries {#sec:preliminary} ============= LPV systems ----------- LPV systems are dynamical systems that can be described as $$\label{eq:mainsyst} \begin{array}{rcl} \dot{x}(t)&=&A(\rho(t))x(t)\\ x(0)&=&x_0 \end{array}$$ where $x,x_0\in\mathbb{R}^n$ are the state of the system and the initial condition, respectively. The matrix-valued function $A(\cdot)\in\mathbb{R}^{n\times n}$ is assumed to be bounded and continuous. The parameter vector trajectory $\rho:\mathbb{R}_{\ge0}\to\mathcal{P}\subset\mathbb{R}^N$, $\mathcal{P}$ compact and connected, is assumed to be piecewise differentiable with derivative in $\mathcal{D}\subset\mathbb{R}^N$, where $\mathcal{D}$ is also compact and connected. When the parameters are independent of each other, then we can assume that $\mathcal{P}$ is a box and that the following decompositions hold $\mathcal{P}=:\mathcal{P}_1\times\ldots\times\mathcal{P}_N$ where $\mathcal{P}_i:=[\underline{\rho}_i,\ \bar{\rho}_i]$, $\underline{\rho}_i\le\bar{\rho}_i$ and $\mathcal{D}=:\mathcal{D}_1\times\ldots\times\mathcal{D}_N$ where $\mathcal{D}_i:=[\underline{\nu}_i,\ \bar{\nu}_i]$, $\underline{\nu}_i\le\bar{\nu}_i$. We also define the set of vertices of $\mathcal{D}$ as $\mathcal{D}^v$; i.e. $\mathcal{D}^v:=\{\underline{\nu}_1,\ \bar{\nu}_1\}\times\ldots\{\underline{\nu}_N,\ \bar{\nu}_N\}$. Hybrid systems -------------- Let us consider here the following hybrid system $$\label{eq:hybrid} \begin{array}{rcl} \dot{\chi}(t)&\in&F(\chi(t))\ \textnormal{if }\chi(t)\in C\\ \chi(t^+)&\in&G(\chi(t))\ \textnormal{if }\chi(t)\in D \end{array}$$ where $\chi(t)\in\mathbb{R}^d$, $C\subset\mathbb{R}^d$ is open, $D\subset\mathbb{R}^d$ is compact and $C\subset G(D)$. The flow map and the jump map are the set-valued maps $F:C\rightrightarrows\mathbb{R}^n$ and $G:D\rightrightarrows C$, respectively. Note that the trajectories of the above system are left-continuous and the right-handed limit is given denoted by $\chi(t^+)=\lim_{s\downarrow t}\chi(s)$. It is often convenient to define a solution (or hybrid arc) $\phi$ to the above system over a hybrid time domain $\operatorname{dom}\phi\subset\mathbb{R}_{\ge0}\times\mathbb{Z}_{\ge0}$ where the first component denotes the usual continuous-time while the second one counts the number of jumps[^2]. We also assume for simplicity that the solutions are complete (i.e. $\operatorname{dom}\phi$ is unbounded). We then have the following stability result: \[th:goebel\] Let $\mathcal{A}\subset\mathbb{R}^d$ be closed. Assume that there exist a function $V:\bar{C}\cup D\mapsto\mathbb{R}$ that is continuously differentiable on an open set containing $\bar C$ (i.e. the closure of $C$), functions $\alpha_1,\alpha_2\in\mathcal{K}_\infty$ and a continuous positive definite function $\alpha_3$ such that (a) $\alpha_1(\|\chi\|_\mathcal{A})\le V(x)\le\alpha_2(\|\chi\|_\mathcal{A})$ for all $\chi\in\bar{C}\cup D$; (b) $\langle\nabla V(\chi),f\rangle\le-\alpha_3(\|\chi\|_\mathcal{A})$ for all $\chi\in C$ and $f\in F(\chi)$; (c) $V(g)-V(\chi)\le0$ for all $\chi\in D$ and $g\in G(\chi)$. Assume further that for each $r>0$, there exists a $\gamma_r\in\mathcal{K}_\infty$ and an $N_r\ge0$ such that for every solution $\phi$ to the system , we have that $\|\phi(0,0)\|_{\mathcal{A}}\in(0,r]$, $(t,j)\in\operatorname{dom}\phi$, $t+j\ge T$ imply $j\ge \gamma_r(T)-N_r$, then $\mathcal{A}$ is uniformly globally asymptotically stable for the system . The above stability result only requires the flow part of the system to be stabilizing while the jump part is only required to be non-expansive. This result is fully adapted to the analysis of LPV systems with piecewise differentiable parameters since discontinuous changes in the values of the parameters can not have any stabilizing effect – quite the opposite. In this regard, the asymptotic stability of the system can only be ensured by the flow part of the system where parameters are smoothly varying. Main results {#sec:stab} ============ The objective of this section is to present the main results on the paper: the stability problem in the case of periodic parameter discontinuities is addressed in Section \[sec:cstDT\] and extended to the aperiodic case in Section \[sec:minDT\]. The results are connected to existing ones in Section \[sec:QR\]. Finally, some computational discussions are provided in Section \[sec:computational\]. Stability under constant dwell-time {#sec:cstDT} ----------------------------------- In this section, we will consider the family of piecewise differentiable parameter trajectories given by $$\mathscr{P}_{\hspace{-2pt}\scriptscriptstyle{\bar{T}}}:=\left\{\begin{array}{c} \rho:\mathbb{R}_{\ge0}\to\mathcal{P}\left\|\begin{array}{c} \dot{\rho}(t)\in\mathcal{Q}(\rho(t)),t\in(t_k,t_{k+1})\\ T_k:=t_{k+1}-t_k=\bar T, k\in\mathbb{N}_{0} \end{array}\right.\end{array}\right\}$$ where $\bar T>0$, $t_0=0$, and $\mathcal{Q}(\rho)=\mathcal{Q}_1(\rho)\times\ldots\times\mathcal{Q}_N(\rho)$ with $$\mathcal{Q}_i(\rho):=\left\{\begin{array}{ccl} \mathcal{D}_i&&\textnormal{if }\rho_i\in(\underline{\rho}_i,\bar{\rho}_i),\\ \mathcal{D}_i\cap\mathbb{R}_{\ge0}&&\textnormal{if }\rho_i=\underline{\rho}_i,\\ \mathcal{D}_i\cap\mathbb{R}_{\le0}&&\textnormal{if }\rho_i=\bar{\rho}_i. \end{array}\right.$$ In other words, the trajectories contained in this family can only exhibit jumps at the times $t_k=k\bar{T}$, $k>0$ (we assume that no discontinuity occurs at $t_0=0$) and hence the distance between two potential successive discontinuities – the so-called dwell-time* – is given by $T_k:=t_{k+1}-t_k=\bar T$ and is constant, whence the name constant dwell-time. The associated hybrid system is given by $$\label{eq:mainsystH_RDT} \begin{array}{l} \left\{\left.\begin{array}{rcl} \dot{x}(t)&=&A(\rho(t))x(t)\\ \dot{\rho}(t)&\in&\mathcal{Q}(\rho(t))\\ \dot{\tau}(t)&=&1 \end{array}\right\| \textnormal{if }(x(t),\rho(t),\tau(t))\in C\right\}\\ \left\{\left.\begin{array}{rcl} x(t^+)&=&x(t)\\ \rho(t^+)&\in&\mathcal{P}\\ \tau(t^+)&=&0 \end{array}\right\|\textnormal{if }(x(t),\rho(t),\tau(t))\in D\right\} \end{array}$$ where $$ C := \mathbb{R}^n\times\mathcal{P}\times[0,\bar T)\ \textnormal{and}\ D :=\mathbb{R}^n\times\mathcal{P}\times\{\bar{T}\}. $$ The initial condition for this system is chosen such that $(x(0),\rho(0),\tau(0))\in\mathbb{R}^n\times\mathcal{P}\times\{0\}$. Note, moreover, that $C\cap D=\emptyset$, $\bar{C}\cap D=D$ and $G(D)\subset C$. Hence, starting from the above initial condition, the solution of the hybrid system is complete (i.e. it is defined for all $t\ge0$) and it is confined in $C\cup D$. The above hybrid system interestingly incorporates both the dynamics of the state of the system and the considered class of parameter trajectories. In addition to that, the state is also augmented to contain a clock that will measure the time elapsed since the last jump in the parameter trajectories. Indeed, starting from the chosen initial condition, the system will smoothly flow until the clock $\tau$ reaches the value $\bar T$ upon which the jump map is activated. The jump map changes the value of the parameter (thereby introducing a discontinuity in the parameter trajectories) and resets the clock, which places back the system in flow-mode. In this regard, we can easily observe that this formulation incorporates all the necessary information about the system and the parameter trajectories in order to provide accurate stability results. Such a result based on the use of a quadratic Lyapunov function is given below: \[th:cstDT\] Let $\bar T\in\mathbb{R}_{>0}$ be given and assume that there exist a bounded continuously differentiable matrix-valued function $S:[0, \bar{T}]\times\mathcal{P}\to\mathbb{S}^n_{\succ0}$ and a scalar $\operatorname{\varepsilon}>0$ such that the conditions $$\label{eq:cst:1} \partial_\tau S(\tau,\theta)+\partial_\rho S(\tau,\theta)\mu+\operatorname{Sym}[S(\tau,\theta)A(\theta)]+\operatorname{\varepsilon}I_n\preceq0$$ and $$\label{eq:cst:2} S(0,\theta)-S(\bar{T},\eta)\preceq0$$ hold for all $\theta,\eta\in\mathcal{P}$, all $\mu\in\mathcal{D}^v$ and all $\tau\in[0, \bar{T}]$. Then the LPV system with parameter trajectories in $\mathscr{P}_{\hspace{-2pt}\scriptscriptstyle{\bar{T}}}$ is asymptotically stable.[$\vartriangle$]{} Let $\mathcal{A}=\{0\}\times\mathcal{P}\times[0, \bar{T}]$ and note that the LPV system with parameter trajectories in $\mathscr{P}_{\hspace{-2pt}\scriptscriptstyle{ T}}$ is asymptotically stable if and only if $\mathcal{A}$ is asymptotically stable for the system . We prove now that for the choice of the Lyapunov function $V(x,\tau,\rho):=x^TS(\tau,\rho)x$ with $S(\tau,\rho)\succ0$ for all $\tau\in[0,T_{max}]$ and all $\rho\in\mathcal{P}$, the feasibility of the conditions of Theorem \[th:cstDT\] implies the feasibility of those in Theorem \[th:goebel\]. Since $S(\tau,\rho)\succ0$ for all $\tau\in[0,T_{max}]$ and all $\rho\in\mathcal{P}$, then $V(x,\tau,\rho)=0$ if and only if $x=0$ and, hence, the conditions of statement (a) of Theorem \[th:goebel\] are verified with the functions $$\begin{array}{rcl} \alpha_1(\|(x,\tau,\rho)\|_\mathcal{A})&=&\min_{(\tau,\rho)\in[0,\bar{T}]\times\mathcal{P}}\lambda_{min}(S(\tau,\rho))\|\|x\|\|_2^2\\ \alpha_2(\|(x,\tau,\rho)\|_\mathcal{A})&=&\max_{(\tau,\rho)\in[0,\bar{T}]\times\mathcal{P}}\lambda_{max}(S(\tau,\rho))\|\|x\|\|_2^2 \end{array}$$ which are both $\mathcal{K}_\infty$ functions. To prove that the feasibility of the condition implies that of statement (b) of Theorem \[th:goebel\], let $\Psi_{0}(\tau,\rho,\mu)$ be the matrix on the left-hand side of when $\operatorname{\varepsilon}=0$. Using the linearity in $\mu$, it is immediate to get that $\Psi_0(\tau,\rho,\mu)+\operatorname{\varepsilon}I_n\preceq0$ for all $(\tau,\theta,\mu)\in[0,\bar{T}]\times\mathcal{P}\times\mathcal{D}^v$ if and only if $\Psi_0(\tau,\rho,\mu)+\operatorname{\varepsilon}I_n\preceq0$ for all $(\tau,\theta,\mu)\in[0,\bar{T}]\times\mathcal{P}\times\mathcal{D}$. Hence, is equivalent to saying that $$\langle \nabla V(x,\tau,\rho),f\rangle=x^T\Psi_0(\tau,\rho,\mu)x\le-\operatorname{\varepsilon}\|\|x\|\|_2^2$$ for all $(x,\tau,\theta,\mu)\in\mathbb{R}^n\times[0,\bar{T}]\times\mathcal{P}\times\mathcal{D}$. Therefore, we have that the feasibility of the condition implies that of statement (b) of Theorem \[th:goebel\] with $\alpha_3(\|(x,\tau,\rho)\|_\mathcal{A})=\operatorname{\varepsilon}\|\|x\|\|^2_2$, which is positive definite. Finally, the condition of statement (c) reads $$V(g)-V(x)=x^T(S(0,\theta)-S(\bar{T},\eta))x\le0$$ and must hold for all $(x,\theta,\eta)\in\mathbb{R}^n\times\mathcal{P}\times\mathcal{P}$. This is equivalent to the condition . Finally, we need to check the time-domain condition. First, note that $$\operatorname{dom}\phi=\bigcup_{j=0}^\infty([j\bar T,(j+1)\bar T],j)$$ which, together with $t+j\ge \tilde T$ for some $\tilde T>0$, implies that $t+1+t/\bar T\ge \tilde T$ since $j\le 1+t/\bar T$ for all $(t,j)\in\operatorname{dom}\phi$. Hence, $t\ge(1+\bar{T}^{-1})^{-1}(\tilde T-1)$ and, as a result, the set $\mathcal{A}$ is asymptotically stable for the system and the result follows. Stability under minimum dwell-time {#sec:minDT} ---------------------------------- Let us consider now the family of piecewise differentiable parameter trajectories given by $$\mathscr{P}_{\hspace{-1mm}{\scriptscriptstyle\geqslant\bar{T}}}:=\left\{\begin{array}{c} \rho:\mathbb{R}_{\ge0}\to\mathcal{P}\left\|\begin{array}{c} \dot{\rho}(t)\in\mathcal{Q}(\rho(t)), t\in[t_k,t_{k+1})\\ T_k:=t_{k+1}-t_k\ge\bar T,\ k\in\mathbb{N}_{0} \end{array}\right.\end{array}\right\}$$ where $\bar T>0$, $t_0=0$ (we again assume that no discontinuity can occur at $t_0$). The corresponding hybrid system is given by $$\label{eq:mainsystH_MDT} \begin{array}{l} \left\{\left.\begin{array}{rcl} \dot{x}(t)&=&A(\rho(t))x(t)\\ \dot{\rho}(t)&\in&\mathcal{Q}(\rho(t))\\ \dot{\tau}(t)&=&1\\ \dot{T}(t)&=&0\\ \end{array}\right\| \textnormal{if }(x(t),\rho(t),\tau(t),T(t))\in C\right\}\\ \left\{\left.\begin{array}{rcl} x(t^+)&=&x(t)\\ \rho(t^+)&\in&\mathcal{P}\\ \tau(t^+)&=&0\\ T(t^+)&\in&[\bar{T},\infty) \end{array}\right\|\textnormal{if }(x(t),\rho(t),\tau(t),T(t))\in D\right\} \end{array}$$ where $$\begin{array}{rcl} C&=& \mathbb{R}^n\times\mathcal{P}\times E_<,\\ D&=&\mathbb{R}^n\times\mathcal{P}\times E_=\\ E_\square&=&\{\varphi\in\mathbb{R}_{\ge0}\times[\bar T,\infty):\varphi_1\square\varphi_2\},\ \square\in\{<,=\}. \end{array}$$ The initial condition for this system is chosen such that $(x(0),\rho(0),\tau(0),T(0))\in\mathbb{R}^n\times\mathcal{P}\times\{0\}\times [\bar T,\infty)$. This system contains an additional state compared to the system considered in Section \[sec:cstDT\] in order to avoid the use of a time-dependent jump set (a purely technical requirement that can be relaxed). Using the current formulation, the current dwell-time is drawn each time the system jumps and the jumping condition is satisfied when $\tau(t)=T(t)$. This leads to the following result: \[th:minDT\] Let $\bar T\in\mathbb{R}_{>0}$ be given and assume that there exist a bounded continuously differentiable matrix-valued function $S:[0, \bar{T}]\times\mathcal{P}\to\mathbb{S}^n_{\succ0}$ and a scalar $\operatorname{\varepsilon}>0$ such that the conditions $$\label{eq:minDT:1} \partial_\rho S(\bar{T},\theta)\mu+\operatorname{Sym}[S(\bar{T},\theta)A(\theta)]+\operatorname{\varepsilon}I\preceq0$$ $$\label{eq:minDT:2} \partial_\tau S(\tau,\theta)+\partial_\rho S(\tau,\theta)\mu+\operatorname{Sym}[S(\tau,\theta)A(\theta)]+\operatorname{\varepsilon}I\preceq0$$ and $$\label{eq:minDT:3} S(0,\theta)-S(\bar{T},\eta)\preceq0$$ hold for all $\theta,\eta\in\mathcal{P}$, $\mu\in\mathcal{D}^v $ and all $\tau\in[0, \bar{T}]$. Then, the LPV system with parameter trajectories in $\mathscr{P}_{\hspace{-1mm}{\scriptscriptstyle\geqslant \bar{T}}}$ is asymptotically stable. Assume that the full trajectory of $T(t)$ is known and such that $T(t)\ge\bar T$ for all $t\ge0$. Note that this is possible since $T(t)$ is independent of the rest of the state of the hybrid system . Then, there exists a $T_{max}<\infty$ such that $\bar{T}\le T(t)\le T_{max}$ for all $t\ge0$ and define the set $\mathcal{A}=\{0\}\times\mathcal{P}\times((E_<\cup E_=)\cap[0,T_{max}]^2)$. Let us consider here the following Lyapunov function $$V(x,\tau,\rho)=\left\{\begin{array}{lcl} x^TS(\tau,\rho)x&&\textnormal{if }\tau\le\bar{T},\\ x^TS(\bar{T},\rho)x&&\textnormal{if }\tau>\bar{T}. \end{array}\right.$$ The rest of the proof is analogous to that of Theorem \[th:cstDT\] and it thus omitted. Connection with quadratic and robust stability {#sec:QR} ---------------------------------------------- Interestingly, it can be shown that the minimum dwell-time result stated in Theorem \[th:minDT\] naturally generalizes and unifies the quadratic and robust stability conditions through the concept of minimum dwell-time. This is further explained in the results below: When $\bar T\to0$ and $t_k\to\infty$ as $k\to\infty$, then the conditions of Theorem \[th:minDT\] are equivalent to saying that there exists a matrix $P\in\mathbb{S}_{\succ0}^n$ such that $$\label{eq:quadstab} A(\theta)^TP+PA(\theta)\prec0$$ for all $\theta\in\mathcal{P}$. First note that since $\bar T>0$ and $t_k\to\infty$ as $k\to\infty$, then the solution of the hybrid system is complete. Let $\bar T=\epsilon>0$, then we have that $$\begin{array}{rcl} S(0,\theta)-S(\epsilon,\theta)&=&-\epsilon \partial_\tau S(0,\theta)+o(\epsilon),\\ S(0,\theta)-S(\epsilon,\eta)&=&(S(0,\theta)-S(0,\eta))-\epsilon \partial_\tau S(0,\eta)+o(\epsilon),\\ S(0,\eta)-S(\epsilon,\theta)&=&(S(0,\eta)-S(0,\theta))-\epsilon \partial_\tau S(0,\theta)+o(\epsilon) \end{array}$$ where it is assumed that $\eta\ne\theta$ and where $o(\epsilon)$ is the little-o notation. From , we get that the first equation implies that $\partial_\tau S(0,\theta)\succeq0$. Therefore, for the second expression to be negative semidefinite for any arbitrarily small $\epsilon>0$, we need that $S(0,\theta)-S(0,\eta)$ be negative semidefinite. However, this contradicts the last one and hence we need that $S(0,\eta)=S(0,\theta)$; i.e. $S$ is independent of $\rho$. Finally, since $\epsilon>0$ is arbitrarily small, then both - can be satisfied with a matrix-valued function $S$ that is independent of $\tau$. Hence, we need that $S(\tau,\theta):=P\succ0$ and substituting it in - yield the quadratic stability condition . When $\bar T\to\infty$, then the conditions of Theorem \[th:minDT\] are equivalent to saying that there exists a differentiable matrix-valued function $P:\mathcal{P}\mapsto\mathcal{S}_{\succ0}^n$ such that $$\label{eq:robstab} \partial_\rho P(\theta)\mu+A(\theta)^TP(\theta)+P(\theta)A(\theta)\prec0$$ for all $\theta\in\mathcal{P}$ and all $\mu\in\mathcal{D}^v $. Clearly, when $\bar T\to\infty$, then there are no jumps anymore and hence we can remove the condition as it never occurs. Consequently, we can choose $S(\tau,\theta)=P(\theta)$ and the conditions - immediately reduce to . Computational considerations {#sec:computational} ---------------------------- The conditions formulated in Theorem \[th:cstDT\] and Theorem \[th:minDT\] are infinite-dimensional semidefinite programs that are intractable per se. To make them tractable, we propose to consider an approach based on sum of squares programming [@Parrilo:00] that will result in a approximate finite-dimensional semidefinite program that can be solved using standard semidefinite programming solvers such as SeDuMi [@Sturm:01a]. The relaxation procedure can be performed with the help of the package SOSTOOLS [@sostools3] This conversion to an SOS program is described below. Assuming that $\mathcal{P}$ is a compact semialgebraic set, then it can defined as $$\mathcal{P}=\left\{\theta\in\mathbb{R}^N: g_i(\theta)\ge0, i=1,\ldots,M \right\}$$ where the functions $g_i:\mathbb{R}^N\mapsto\mathbb{R}$, $i=1,\ldots,M $ are polynomial. We further have that $$[0, \bar{T}]=\left\{\tau\in\mathbb{R}:\ f(\tau):=\tau( \bar{T}-\tau)\ge0\right\}.$$ In what follows, we say that a symmetric matrix-valued function $\Theta(\cdot)$ is a sum of squares matrix (SOS matrix) if there exists a matrix $\Xi(\cdot)$ such that $\Theta(\cdot)=\Xi(\cdot)^{T}\Xi(\cdot)$. The following result provides the sum of squares formulation of Theorem \[th:cstDT\]: \[prog:periodic\] Let $\operatorname{\varepsilon}, \bar{T}>0$ be given and assume that the sum of squares program is feasible. Then, the conditions of Theorem \[th:cstDT\] hold with the computed polynomial matrix $S(\tau,\theta)$ and the system is asymptotically stable for all $\rho\in\mathscr{P}_{\hspace{-2pt}\scriptscriptstyle{\bar T}}$. The conditions of Theorem \[th:minDT\] can be checked by simply adding to the SOS program of Proposition \[prog:periodic\] the constraints - $ \Upsilon_3^i(\tau,\theta,\mu)$, $i=1,\ldots,M$, are SOS matrices for all $\mu\in\mathcal{D}^v $ - $ -\partial_\rho S(\bar{T},\theta)\mu-\operatorname{Sym}[S( T,\theta)A(\theta)]-\sum_{i=1}^{M }\Upsilon_3^i(\tau,\theta,\mu)g_i(\theta)-\operatorname{\varepsilon}I_n$ is an SOS matrix for all $\mu\in\mathcal{D}^v $ where $\Upsilon_3^i:\mathcal{D}^v \times [0,\bar{T}]\times\mathcal{P}\mapsto\mathbb{S}^n$, $i=1,\ldots,M$, are additional polynomial variables. When the parameter set $\mathcal{P}$ is also defined by equality constraints $h_i(\theta)=0$, $i=1,\ldots, M'$, these constraints can be simply added in the sum of squares programs in the same way as the inequality constraints, but with the particularity that the corresponding multiplier matrices be simply symmetric instead of being symmetric SOS. Examples {#sec:examples} ======== We consider now two examples. The first one is a 2-dimensional toy example considered in [@Xie:97] whereas the second one is a 4-dimensional system considered in [@Wu:95] and inspired from an automatic flight control design problem. Example 1 --------- Let us consider the system with the matrix [@Xie:97; @Briat:15d] $$\label{eq:ex1} A(\rho)=\begin{bmatrix} 0 & &1\\ -2-\rho &\ &-1 \end{bmatrix}$$ where the time-varying parameter $\rho(t)$ takes values in $\mathcal{P}=[0,\bar{\rho}]$, $\bar{\rho}>0$. It is known [@Xie:97] that this system is quadratically stable if and only if $\bar{\rho}\le3.828$ but it is was later proven in the context of piecewise constant parameters [@Briat:15d] that this bound can be improved provided that discontinuities do not occur too often. We now apply the conditions of Theorem \[th:cstDT\] and Theorem \[th:minDT\] in order to characterize the impact of parameter variations between discontinuities. To this aim, we consider that $\|\dot{\rho}(t)\|\le\nu$ with $\nu\ge0$ and that $\bar{\rho}\in\{0,0.1,\ldots,10\}$. For each value for the upper-bound $\bar{\rho}$ in that set, we solve for the conditions Theorem \[th:cstDT\] and Theorem \[th:minDT\] to get estimates (i.e. upper-bounds) for the minimum stability-preserving constant and minimum dwell-time. We use here $\operatorname{\varepsilon}_1=\operatorname{\varepsilon}_2=\operatorname{\varepsilon}_3=0.01$ and polynomials of degree 4 in the sum of squares programs. Note that we have, in this case, $M=1$, $M'=0$ and $g_1(\theta)=\theta(\bar{\rho}-\theta)$. The complexity of the approach can be evaluated here through the number of primal/dual variables of the semidefinite program which is 2209/273 in the constant dwell-time case and 2409/315 in the minimum dwell-time case. Time-wise, the average preprocessing/solving time is given by 4.83/0.89 sec in the constant dwell-time case and 6.04/1.25 sec in the minimum dwell-time case. The results are depicted in Fig. \[fig:1\] and Fig. \[fig:2\] where we can see that the obtained minimum values for the dwell-times increase with the rate of variation $\nu$ of the parameter, which is an indicator of the fact that increasing the rate of variation of the parameter destabilizes the system and, consequently, the dwell-time needs to increase in order to preserve the overall stability of the system. For this example, it is interesting to note that the constant and minimum dwell-time curves seem to converge to each other when $\nu$ increases which could indicate that when the parameter is fast-varying between jumps then the aperiodicity of jumps do not affect much the stability of the system. ![Evolution of the computed minimum upper-bound on the constant dwell-time using Theorem \[th:cstDT\] for the system - using an SOS approach with polynomials of degree 4.[]{data-label="fig:1"}](./Data_cstDT_nu.pdf){width="65.00000%"} ![Evolution of the computed minimum upper-bound on the minimum stability-preserving minimum dwell-time using Theorem \[th:minDT\] for the system - with $\|\dot{\rho}\|\le\nu$ using an SOS approach with polynomials of degree 4.[]{data-label="fig:2"}](./Data_minDT_nu.pdf){width="65.00000%"} Example 2 --------- Let us consider now the system with the matrix considered in [@Wu:95 p. 55]: $$\label{eq:ex2} A(\rho)=\begin{bmatrix} 3/4 &\ & 2 &\ & \rho_1 &\ & \rho_2\\ 0 &\ & 1/2 &\ & -\rho_2 &\ & \rho_1\\ -3\upsilon\rho_1/4 &\ & \upsilon\left(\rho_2-2\rho_1\right) &\ & -\upsilon &\ & 0\\ -3\upsilon\rho_2/4 &\ & \upsilon\left(\rho_1-2\rho_2\right) &\ & 0 &\ & -\upsilon \end{bmatrix}$$ where $\upsilon=15/4$ and $\rho\in\mathcal{P}=\{z\in\mathbb{R}^2:\|\|z\|\|_2=1\}$. It has been shown in [@Wu:95] that this system is not quadratically stable but was proven to be stable under minimum dwell-time equal to 1.7605 when the parameter trajectories are piecewise constant [@Briat:15d]. We propose now to quantify the effects of smooth parameter variations between discontinuities. Note, however, that the set $\mathcal{P}$ is not a box as considered along the paper but the next calculations show that this is not a problem since a proper set for the values of the derivative of the parameters can be defined. To this aim, let us define the parametrization $\rho_1(t)=\cos(\beta(t))$ and $\rho_2(t)=\sin(\beta(t))$ where $\beta(t)$ is piecewise differentiable. Differentiating these equalities yields $\dot{\rho}_1(t)=-\dot{\beta}(t)\rho_2(t)$ and $\dot{\rho}_2(t)=\dot{\beta}(t)\rho_1(t)$ where $\dot{\beta}(t)\in[-\nu,\nu]$, $\nu\ge0$, at all times where $\beta(t)$ is differentiable. In this regard, we can consider $\dot{\beta}$ as an additional parameter that enters linearly in the stability conditions and hence the conditions can be checked at the vertices of the interval, that is, for all $\dot{\beta}\in\{-\nu,\nu\}$. Note that, in this case, we have $M=0$, $M'=1$ and $h_1(\rho)=\rho_1^2+\rho_2^2-1$. $\nu=0$ $\nu=0.1$ $\nu=0.3$ $\nu=0.5$ $\nu=0.8$ $\nu=0.9$ primal/dual vars. time (sec) ------- --------- ----------- ----------- ----------- ----------- ----------- ------------------- ------------ $d=2$ 2.7282 2.9494 3.5578 4.6317 11.6859 26.1883 9820/1850 20/27 $d=4$ 1.7605 1.8881 2.2561 2.9466 6.4539 43300/4620 212/935 We now consider the conditions of Theorem \[th:minDT\] and we get the results gathered in Table \[tab\] where we can see that, as expected, when $\nu$ increases then the minimum dwell-time has to increase to preserve stability. Using polynomials of higher degree allows to improve the numerical results at the expense of an increase of the computational complexity. As a final comment, it seems important to point out the failure of the semidefinite solver due to too important numerical errors when $d=4$ and $\nu=0.9$. Conclusion ========== Reformulating LPV systems into hybrid systems enabled the derivation of tractable conditions for establishing the stability of LPV systems with piecewise differentiable parameters. These results extend those obtained in [@Briat:15d] for piecewise constant parameters through the use of a different, yet connected, approach. It is shown that the obtained stability conditions generalize and unify the well-known quadratic and robust stability conditions in a single formulation. Possible extensions include controller/filter/observer design and performance analysis. [10]{} R. T[óth]{}. . Springer, Germany, 2010. J. Mohammadpour and C. W. Scherer, editors. . Springer, New York, USA, 2012. C. Briat. , volume 3 of [Advances on Delays and Dynamics]{}. Springer-Verlag, Heidelberg, Germany, 2015. F. Wu. . PhD thesis, University of California Berkeley, 1995. J. S. Shamma. . PhD thesis, Laboratory for Information and decision systems - Massachusetts Institute of Technology, 1988. J. S. Shamma and M. Athans. Gain scheduling: potential hazards and possible remedies. , 12(3):101–107, 1992. C. Poussot-Vassal, O. Sename, L. Dugard, P. Gáspár, Z. Szabó, and J. Bokor. New semi-active suspension control strategy through [LPV]{} technique. , 2008. S. M. Savaresi, C. [Poussot-Vassal]{}, C. Spelta, O. Sename, and L. Dugard. . Butterworth Heinemann, 2010. H. Kajiwara, P. Apkarian, and P. Gahinet. techniques for control of an inverted pendulum. , 19:44–54, 1999. D. Robert, O. Sename, and D. Simon. An $\mathcal{H}_\infty$ [LPV]{} design for sampling varying controllers: Experimentation with a [T]{}-inverted pendulum. , 18(3):741–749, 2010. W. Gilbert, D. Henrion, J. Bernussou, and D. Boyer. Polynomial [LPV]{} synthesis applied to turbofan engines. , 18(9):1077–1083, 2010. P. Seiler, G. J. Balas, and A. Packard. Linear parameter-varying control for the [X]{}-53 active aeroelastic wing. In J. Mohammadpour and C. W. Scherer, editors, [Control of Linear Parameter Varying Systems with Applications]{}, pages 483–512. Springer Science, 2012. H. Pfifer, C. P. Moreno, J. Theis, A. Kotikapuldi, A. Gupta, B. Takarics, and P. Seiler. Linear parameter varying techniques applied to aeroservoelastic aircraft: In memory of [G]{}ary [B]{}alas. In [1st IFAC Workshop on Linear Parameter Varying Systems]{}, 2015. A. Packard. Gain scheduling via [L]{}inear [F]{}ractional [T]{}ransformations. , 22:79–92, 1994. P. Apkarian, P. Gahinet, and G. Becker. Self-scheduled control of linear parameter varying systems: A design example. , 31(9):1251–1261, 1995. P. Apkarian and P. Gahinet. A convex characterization of gain-scheduled $\mathcal{H}_\infty$ controllers. , 5:853–864, 1995. P. Apkarian and R. J. Adams. Advanced gain-scheduling techniques for uncertain systems. , 6:21–32, 1998. F. Wu. A generalized [LPV]{} system analysis and control synthesis framework. , 74:745–759, 2001. F. Wu and K. Dong. Gain-scheduling control of [LFT]{} systems using parameter-dependent [L]{}yapunov functions. , 42:39–50, 2006. C. W. Scherer. control and full-block multipliers. , 37:361–375, 2001. J. Bokor. Linear parameter varying systems : a geometric theory and applications. , 2005. C. W. Scherer and I. E. K[ö]{}se. . , 57(9):2219–2234, 2012. C. W. Scherer. Gain-scheduling control with dynamic multipliers by convex optimization. , 53(3):1224–1249, 2015. C. Briat. Stability analysis and control of [LPV]{} systems with piecewise constant parameters. , 82:10–17, 2015. G. Balas, A. Hjartarson, A. Packard, and P. Seiler. . Software and user’s manual, 2015. T. Peni and P. J. Seiler. Computation of lower bounds for the induced $\mathcal{{L}}_2$ norm of [LPV]{} systems. , 26(4):646–661, 2016. S. Wang, H. Pfifer, and P. Seiler. Robust synthesis for linear parameter varying systems using integral quadratic constraints. , 68:111–118, 2016. R. Goebel, R. G. Sanfelice, and A. R. Teel. . Princeton University Press, 2012. C. Briat. Convex conditions for robust stability analysis and stabilization of linear aperiodic impulsive and sampled-data systems under dwell-time constraints. , 49(11):3449–3457, 2013. C. Briat. Convex conditions for robust stabilization of uncertain switched systems with guaranteed minimum and mode-dependent dwell-time. , 78:63–72, 2015. C. Briat. Stability analysis and stabilization of stochastic linear impulsive, switched and sampled-data systems under dwell-time constraints. , 74:279–287, 2016. M. Putinar. Positive polynomials on compact semi-algebraic sets. , 42(3):969–984, 1993. P. Parrilo. . PhD thesis, California Institute of Technology, Pasadena, California, 2000. A. Papachristodoulou, J. Anderson, G. Valmorbida, S. Prajna, P. Seiler, and P. A. Parrilo. , 2013. J. F. Sturm. Using [SEDUMI]{} $1. 02$, a [M]{}atlab [T]{}oolbox for [O]{}ptimization [O]{}ver [S]{}ymmetric [C]{}ones. , 11(12):625–653, 2001. L. Xie, S. Shishkin, and M. Fu. Piecewise [L]{}yapunov functions for robust stability of linear time-varying systems. , 31(3):165–171, 1997. [^1]: Corentin Briat and Mustafa Khammash are with the Department of Biosystems Science and Engineering, ETH-Zürich, Switzerland; email: [email protected], [email protected], [email protected]; url: https://www.bsse.ethz.ch/ctsb/, http://www.briat.info. [^2]: See [@Goebel:12] for more details about the solutions of .	{ "pile_set_name": "ArXiv" }
--- abstract: 'We consider the smoothed maximum likelihood estimator and the smoothed Grenander-type estimator for a monotone baseline hazard rate $\lambda_0$ in the Cox model. We analyze their asymptotic behavior and show that they are asymptotically normal at rate $n^{m/(2m+1)}$, when $\lambda_0$ is $m\geq 2$ times continuously differentiable, and that both estimators are asymptotically equivalent. Finally, we present numerical results on pointwise confidence intervals that illustrate the comparable behavior of the two methods.' address: \| Delft Institute of Applied Mathematics\ Delft University of Technology\ van Mourik Broekmanweg 6, 2628 XE Delft, Netherlands\ \ author: - - title: Smoothed isotonic estimators of a monotone baseline hazard in the Cox model --- ,	{ "pile_set_name": "ArXiv" }
--- abstract: 'The [Kepler]{} mission has provided a treasure trove of eclipsing binaries (EBs), observed at extremely high photometric precision, nearly continuously for several years. We are carrying out a survey of $\sim$100 of these EBs to derive dynamical masses and radii with precisions of 3% or better. We use multiplexed near-infrared H band spectroscopy from the Sloan Digital Sky Survey-III and -IV APOGEE instrument and optical spectroscopy from the Hobby Eberly telescope High-Resolution Spectrograph to derive double-lined spectroscopic orbits and dynamical mass-ratios ([$\mathit{q}$]{}) for the EB sample, two of which we showcase in this paper. This orbital information is combined with [Kepler]{} photometry to derive orbital inclination, dynamical masses of the system component, radii and temperatures. These measurements are directly applicable for benchmarking stellar models that are integrating the next generation of improvements, such as the magnetic suppression of convection efficiency, updated opacity tables, and fine-tuned equations of state. We selected our EB sample to include systems with low-mass ($M\la0.8\,{\ensuremath{\mathit{\,M_\odot}}}$) primary or secondary components, as well as many EBs expected to populate the relatively sparse parameter space below $\sim0.5\,M_{\odot}$. In this paper, we describe our EB sample and the analysis techniques we are utilizing, and also present masses and radii for two systems that inhabit particularly underpopulated regions of mass-radius-period space: KIC 2445134 and KIC 3003991. Our joint spectrosopic and photometric analysis of KIC 2445134 (${\ensuremath{\mathit{q}}}{}=0.411\pm0.001$) yields masses and radii of ${\ensuremath{\mathit{M}_A}}{}=1.29\pm0.03{\ensuremath{\mathit{\,M_\odot}}}$, ${\ensuremath{\mathit{M}_B}}{}=0.53\pm0.01{\ensuremath{\mathit{\,M_\odot}}}$, ${\ensuremath{\mathit{R}_A}}{}=1.42\pm0.01{\ensuremath{\mathit{\,R_\odot}}}$, ${\ensuremath{\mathit{R}_B}}{}=0.510\pm0.004{\ensuremath{\mathit{\,R_\odot}}}$, and a temperature ratio of $T_B/T_A=0.635\pm0.001$; our analysis of KIC 3003991 (${\ensuremath{\mathit{q}}}{}=0.298\pm0.006$) yields ${\ensuremath{\mathit{M}_A}}{}=0.74\pm0.04{\ensuremath{\mathit{\,M_\odot}}}$, ${\ensuremath{\mathit{M}_B}}{}=0.222\pm0.007{\ensuremath{\mathit{\,M_\odot}}}$, ${\ensuremath{\mathit{R}_A}}{}=0.84\pm0.01{\ensuremath{\mathit{\,R_\odot}}}$, ${\ensuremath{\mathit{R}_B}}{}=0.250\pm0.004{\ensuremath{\mathit{\,R_\odot}}}$, and a temperature ratio of $T_B/T_A=0.662\pm0.001$.' author: - Suvrath Mahadevan - 'Chad F. Bender' - Kelly Hambleton - 'Scott W. Fleming' - Rohit Deshpande - Kyle Conroy - Gal Matijevič - Leslie Hebb - Arpita Roy - Hasan Ak - Blaž Leban - Andrej Prša title: 'The SDSS-HET Survey of [Kepler]{} Eclipsing Binaries. Description of the Survey and First Results.' --- Introduction ============ Eclipsing binaries (EBs) have served as benchmarks for stellar astrophysics for hundreds of years [e.g., @1783RSPT...73..474G]. Advances in radial velocity (RV) measurement precision over the past several decades have steadily improved the precision of masses derived from EBs. Compilations of EBs show that the total number, diversity, and measurement precisions of detached, main-sequence EBs have steadily increased: 72 stars with masses and radii ($M$, $R$) measured to $<$ 15% in , 88 stars with $M$, $R$ $<$ 2% in , 188 stars with $M$, $R$ $<$ 3% in , and 198 stars with $M$, $R$ $<$2% in DEBCat[^1], an updated version of . Despite this steady growth in sample size, low-mass stars (defined here as $M < 0.8\,M_{\odot}$) remain a relatively small fraction of the overall sample. For example, DEBCat contains only 54 stars with $M < 0.8\,M_{\odot}$, while the sample has only 10 such stars, none of which have orbital periods longer than five days. More recent results are starting to expand the low-mass sample [e.g., @Schwamb2013; @Gomez2014; @Zhou2015; @Dittmann2017; @Lubin2017; @Casewell2018], but a significant increase in sample size is still lacking, predominantly due to the resource expense associated with obtaining high-precision RVs (the method used to derive most of the EB dynamical masses in the current compilations), and because the poor flux ratio of a low-mass K/M dwarf orbiting a larger primary makes it difficult to detect in the optical. Observations in the near-infrared (NIR) result in the tangible benefit of improving the flux contrast for EBs with a small mass-ratio (${\ensuremath{\mathit{q}}}={\ensuremath{\mathit{M}_B}}/{\ensuremath{\mathit{M}_A}}$). For example, the flux ratio of an M5 dwarf, with $T_{\rm{eff}}\sim3300$K to a G2 dwarf, with $T_{\rm{eff}}\sim5800$K, is $\sim10$ times more favorable in the $H$ band compared to the $V$ band. This allows for the extraction of measured RVs from fainter secondary stars, and therefore extends the lower limit on [$\mathit{q}$]{} for which masses of dwarf stars can be derived. While optical spectra tend to lose sensitivity at $q \sim 0.5$, NIR spectra can push down to $q \sim 0.1$ [@Prato2002; @2003ApJ...599.1344M; @2008ApJ...689..416B]. Measurements of select bright binaries have achieved mass ratio precisions of 0.02-0.42% (in $M\sin{i}$) [@2010ApJ...719.1293K], but precisions of 1-3% are still sufficient to distinguish between model parameters [@2010ApJ...718..502M]. The high photometric precision of [Kepler]{} makes it possible to determine the masses of some EBs purely photometrically [@2011Sci...331..562C; @2012ApJ...746..185F] via Doppler boosting and ellipsoidal effects [@2011MNRAS.415.3921F]. While the technique is promising, @2011Sci...331..562C achieved only a 10% mass precision using the [Kepler]{} photometry alone (i.e., without including any spectroscopic RVs), while @2012ApJ...746..185F found that only 5 of their 7 binaries had RV semi-amplitudes in agreement with those predicted from their photometric analysis. In a large follow-up program, @talor2015 observed 281 targets, confirming 70 binary systems while finding many of their false positives were due to pulsating red giants. The use of spectroscopic RVs therefore remains an essential and reliable method for obtaining mass precisions at the 1-3% level, motivating the continual use of spectra to measure the stellar masses of EBs. Stellar models rely on the precision measurements that EBs afford for calibrating the physical parameters used in their calculations (leading to well-characterized EBs often being referred to as stellar “benchmarks”). Stars with $M > 0.8\,M_{\odot}$ generally agree with theoretical models to within observational uncertainties. Lower mass stars, however, are often observed to have radii that are 5-15% larger than model predictions . There is observational and theoretical evidence to suggest that magnetic fields could be the cause , due to interaction with the partially convective outer atmospheres and/or generation of cool starspots at polar latitudes. Unidentified opacity sources have also been suggested [@2006ApJ...644..475B], although a lack of metallicity measurements for many stars in the current EB sample has prevented an in-depth examination of that possible correlation. Models that use a different equation of state than previous works agree with observations of one low-mass [Kepler]{} EB [KOI-126; @2011ApJ...740L..25F], opening another regime of parameter space to explore. In this paper we introduce the SDSS-HET Survey of Kepler Eclipsing Binaries, which is combining [Kepler]{} photometry with ground-based spectroscopy to precisely measure orbital parameters, dynamical masses and radii, and temperature ratios for a sample of 109 EB candidates selected from the @Kirk2016 catalog, and listed in Table \[ebsampletable\]. We utilize both optical and NIR spectroscopy to solve these EBs as double-lined spectroscopic binaries (SB2s), and combine these results with [Kepler]{} photometry to derive masses and radii with precisions of better than 3% for most of the sample, and as good as 1% for a subset. Here we present two objects in our sample. Our complete sample is restricted to EBs that are classified as fully detached and have $H<13$. Orbital periods range from a few days to more than one hundred days. The total sample size is comparable in number to the $M<2{\ensuremath{\mathit{\,M_\odot}}}$ members in and DEBCat (110 and 285 EBs, respectively), but we have included many EBs with low-mass primary or secondary components in order to substantially increase the population of well measured low-mass stars. One of the reasons the compilation from is so useful to the astronomical community is that their study homogenously re-computed orbital and stellar parameters from the compiled list of literature EBs. Our survey will intrinsically possess this quality, as our analysis applies a singular set of tools to a homogeneous data set. In §2 we describe the facilities and data products we are using, and in §3 we discuss the analysis techniques applied to the spectroscopy and photometry, as well as probabilistic analysis that we use to derive realistic parameter uncertainties. In §4 we present two low-mass EBs from our sample, and provide our derived orbits, masses, and radii for these systems. In §5 we discuss these systems in the context of the pre-existing population of precisely measured low-mass EBs, and describe our plans for analyzing the remainder of our [Kepler]{} EB sample. As part of this program, we constructed a semi-automated reduction pipeline for the HET High-Resolution Spectrograph, which we describe in Appendix A. Facilities and Datasets\[facilities\] ===================================== The [Kepler]{} Mission\[kepler\] ---------------------------------- [Kepler]{} is a space-borne, 0.95-m, high-precision photometer equipped with a broadband filter covering 420 nm – 865 nm. From 2009 – 2013, [Kepler]{} monitored a single field located $13.5^{\circ}$ above the galactic plane in the direction of Cygnus, with a mission to detect transiting, habitable-zone exoplanets [@2010Sci...327..977B]. Its ability to conduct photometry with high precision [$\sim80$ ppm for $K_p = 12$; @2010ApJ...713L..92C] and to observe with a nearly continuous cadence facilitated numerous ancillary stellar astrophysics programs, including the detection of EBs over a range of orbital periods that are challenging to observe from the ground. The majority of [Kepler]{} EBs are faint [only $\sim12$% of detached systems have $K_p < 12$ in @Kirk2016], and followup spectroscopy of such systems requires a combination of large telescopes and long exposure times. The Kepler Eclipsing Binary Catalog[^2] [@2005ApJ...628..426P; @2011AJ....142..160S; @Matijevic2012; @Conroy2014; @Kirk2016; @Abdul2016] has used [Kepler]{} photometry to identify thousands of EBs, and has derived extremely precise orbital periods, which greatly simplifies the process of turning individual spectroscopic measurements into an SB2 orbit. The SDSS-III APOGEE Spectrometer\[apogee\] ------------------------------------------ The Apache Point Observatory Galactic Evolution Experiment (APOGEE, [@Majewski2017]) is a fiber-fed, multi-object, near-infrared spectrometer that uses a volume phase holographic grating and a linear array of three Hawaii-2RG detectors to record spectra from 1.51[$\mathrm{\,\mu{}m}$]{} to 1.68[$\mathrm{\,\mu{}m}$]{} with a spectral resolution of $\lambda/\Delta\lambda\sim22,500$ [@2010SPIE.7735E..46W]. The instrument is located at Apache Point Observatory on the 2.5m SDSS telescope [@2006AJ....131.2332G; @York2000], and was commissioned in the spring and summer of 2011 for a three year survey aimed at a galactic evolution experiment [@2008AN....329.1018A; @2011AJ....142...72E]. A small fraction of the survey time ($\sim5\%$) was devoted to ancillary science programs [@2013AJ....146...81Z], which includes the [Kepler]{} EB program described in this paper. The spectrograph is stabilized in a vacuum-sealed cryostat cooled via liquid nitrogen, which minimizes thermal variations and yields a typical radial velocity precision of $100-200{\ensuremath{\mathrm{\,m\,s^{-1}}}}$ on our EB sample. APOGEE can simultaneously observe 300 targets over the telescope’s 3$^\circ$ diameter field-of-view, which is also fortuitously the approximate size of each [Kepler]{} module. This multiplexing capability allows us to efficiently observe many EBs with a single integration, also referred to as a visit. Each visit is typically comprised of eight consecutive eight-minute exposures, which are later combined by the APOGEE data reduction pipeline [@Nidever2015] to produce a single visit spectrum with a total integration time of slightly more than one hour. All analysis described in this paper utilizes APOGEE data at the visit level of processing. Our ancillary APOGEE program included two fields overlapping [Kepler]{} modules centered on the open clusters NGC 6791 (3 visits) and NGC 6819 (6 visits). We observed a total of 42 and 67 detached EBs in each field, respectively. All spectra for this program were obtained in 2011, during the first year of survey operations, and have been publicly released as part of SDSS Data Release 10 [@2013arXiv1307.7735A]. Prior to analyzing the APOGEE spectra, we perform additional post-processing beyond that provided by the APOGEE pipeline. We use a low order polynomial to remove continuum and normalize each spectrum. Residuals caused by imperfect correction of telluric absorption and sky emission are present in most pipeline reduced APOGEE spectra; we manually correct these by interpolating over neighboring pixels. The pipeline version used for the DR10 release flagged wavelengths with suspected bad pixels by setting their flux to zero. This flagging complicates our RV measurements because a cross-correlation analysis interprets such pixels as having strong, discrete signal that is not present in correlation templates, which reduces the overall amplitude of a real correlation signal. We interpolate over these regions to minimize their impact. The HET High-Resolution Spectrograph\[hrs\] ------------------------------------------- The small number of visits obtained by APOGEE in each of our fields are, by themselves, inadequate for deriving stellar mass at the 3% level for most of our EBs, but the infrared bandwidth provides essential leverage on the low-mass companions in high-contrast systems. Consequently, for 55 of our EBs (Table \[ebsampletable\], column 7) we supplemented the APOGEE spectroscopy with optical spectroscopy from the Hobby-Eberly Telescope to achive higher precision in spectroscopic orbital determination for the primary star. The High-Resolution Spectrograph [@1998SPIE.3355..387T hereafter HRS] is a visible light fiber-fed, cross-dispersed, echelle spectrometer located on the 9.2m Hobby-Eberly Telescope [@1998SPIE.3352...34R hereafter HET]. The HET was designed to carry out narrow-field spectroscopy of faint objects, and so is well suited for targets like our [Kepler]{} EB sample. The spectrograph is fiber-fed and housed in an isolated enclosure in the HET spectrograph room, which provides a moderately stable environment capable of achieving long-term radial velocity stability of $25{\ensuremath{\mathrm{\,m\,s^{-1}}}}$ on bright, low-mass, main-sequence stars using standard ThAr emission lamps for wavelength calibration [@2012ApJ...751L..31B]. The HET operates under a queue-based observing scheme [@2007PASP..119..556S] that allows us to request observations of a target to occur within a narrow time window. This capability allowed us to efficiently observe 55 members of our EB sample from 2011 – 2013, while targeting each system at specific orbital phases. A similar set of observations could not easily be obtained at a classically scheduled facility, making the HET a unique resource for studying binary stars. We use the HRS with a 2” fiber in a configuration that provides a spectral resolution of $\lambda / \Delta\lambda=30,000$ over a bandwidth from 4076Å– 7838Å, except a small gap at 5936Å where light falls between the two HRS CCDs, distributed in 73 spectral orders. This configuration provides a large number of these features from which we derive precise radial velocities. The HET queue-based observing mode allows for the spectrograph configuration to change multiple times throughout a single night as targeting requests from different observers are carried out. Standard sequences of calibration frames, including biases, flats, and ThAr wavelength references, are typically obtained at the end of the observing session. The echelle and cross-disperser positions are not precisely repeatable at the sub-pixel level, so the standard HET operations introduce a discontinuity between the target and the ThAr observations, which can result in RV shifts of several hundred m s$^{-1}$. To avoid this, we obtain additional ThAr frames immediately before or after each target observation, without altering the instrument configuration. We do not obtain extra flats because the telescope overhead would be severe and our experience has shown that the misalignment inherent in the standard queue procedures is usually small enough to prevent fringing in reduced images. To efficiently and uniformly reduce the large HRS data set generated by our [Kepler]{} EB program, we created a semi-automated data handling pipeline for the HRS. This pipeline is written in the Interactive Data language (IDL) and carries out image processing, spectral extraction, and wavelength calibration tasks. Appendix A gives a detailed description of the pipeline. As with the APOGEE spectra, we apply several post-processing steps to our extracted HRS spectra prior to analysis. Each spectral order is continuum normalized. Strong telluric contamination is mostly restricted to isolated wavelength regions at optical wavelengths. Rather than attempt a telluric correction across our HRS bandpass, we choose to retain only those regions of the spectrum with telluric contamination of 0.5% or smaller, modulus a continuity function that preserves large unbroken chunks of spectrum. In practice, we retain eight spectral windows: 4390–5025Å, 5100–5410Å, 5475–5680Å, 5770–5855Å, 6020–6260Å, 6365–6430Å, 6620–6850Å, and 7450–7580Å. Finally, sky emission lines are removed by interpolating over neighboring pixels. Analysis Techniques\[analysis\] =============================== Measurement of Radial Velocities \[rvanalysis\] ----------------------------------------------- All spectroscopy of unresolved binary stars contain the blended light of both the primary and secondary components. The S/N and the wavelength-dependent contrast ratio of an individual spectrum dictate whether that spectrum can be solved as a double-lined spectroscopic binary (SB2) or only as a single-lined spectroscopic binary (SB1). Our APOGEE spectroscopy (§\[apogee\]) was designed to exploit the NIR contrast advantage of EBs containing a pair of main-sequence stars with a small [$\mathit{q}$]{}, thereby observing the systems as SB2s. Because our target sample contains a wide range of binary types, including EBs with small [$\mathit{q}$]{}, EBs with equal mass components, and EBs with one or both components evolving or evolved, the suitability of optical or NIR spectroscopy for solving each EB as an SB2 is determined on a case by case basis. We analyze our processed APOGEE and HET spectra identically, using a combination of one-dimensional and two-dimensional cross-correlation algorithms to measure RVs for the primary and secondary components of each EB. For the two-dimensional case, we have implemented the TODCOR algorithm [@1994ApJ...420..806Z] as an interactive IDL program, SXCORR, which allows the user great flexibility in quickly optimizing the correlation templates while examining multiple epochs of spectra. SXCORR simultaneously cross-correlates two template spectra against a target spectrum containing the blended light from a binary to disentangle the component RVs. Our SXCORR implementation of TODCOR includes the maximum-likelihood extension described by @2003MNRAS.342.1291Z, modified slightly to allow segments of spectrum with variable lengths. SXCORR automatically resamples both target and template spectra into log-lambda wavelength space [@1979AJ.....84.1511T], as needed depending on the input spectra. Previous investigations have extensively described our procedures for using TODCOR techniques to measure the RVs of an SB2 [e.g. @2014arXiv1402.0846L; @2012ApJ...751L..31B; @2008ApJ...689..416B], and we refer the interested reader to the discriptions therein. Our one-dimensional correlation analysis is the trivial simplification of the two-dimensional case. Concurrently with our HRS EB observations, we observed an extensive spectral template library of known single dwarf stars using an identical HRS configuration. This library ranges in spectral type from early F through mid M. Additionally, we have supplemented this library with synthetic templates generated from the `PHOENIX`-based BT-Settl model grid [@2011ASPC..448...91A]. These synthetic models cover a much wider range of $T_{eff}$, $\mathrm{[M/H]}$, and $\log{g}$ than our observed library, although some demonstrate substantial discrepancies with observed spectra at high-resolution [@2014ApJ...782...61T] that manifest in our RV analysis as reduced correlation power. To generate a template from the BT-Settl library, we convolve the raw synthetic spectrum to the proper resolution (22,500 for APOGEE; 30,000 for HRS), and resample to 3-pixels per resolution element. We additionally apply a suite of rotational broadening kernels generated from a four parameter non-linear limb-darkening model and the appropriate stellar parameters. We do not have an observed template library for the H-band APOGEE spectra, so all APOGEE RVs are measured using BT-Settl templates. Binary Star Modeling -------------------- To model the [Kepler]{} light curve with and radial velocity data simultaneously, we used 1.0 binary star modeling software [@2005ApJ...628..426P], which is based on the Wilson-Devinney code [@1971ApJ...166..605W hereafter WD]. We fit the data within a Bayesian framework, by utilizing , a implementation of the affine invariant Markov chain Monte Carlo (MCMC) ensemble sampler proposed by @Goodman2010 and implemented by @2013PASP..125..306F. We model the noise, instrumental variations and stellar variations caused by spots by employing , a Gaussian process library [@DFM2017]. Prior to modeling, we minimally prepared the data by flux-normalizing and stitching the [Kepler]{} quarters and removing any obvious spurious points. The entire process is streamlined into the pipeline that is capable of processing the data autonomously. We describe the details below. ### Uncertainty Determination for the Data As the data uncertainties are commonly underestimated, we determined the uncertainties by identifying the standard deviation of sections of the light curve. For each object we identified 10 sections that contained slowly varying instrumental noise and spots, and Gaussian noise. We determined the noise of the 10 sections individually and subsequently used the average as our uncertainty value for all light curve points. To ensure we did not underestimate the noise in the light curve data, we included a Gaussian noise term in our fitting procedure, which is discussed in more detail in §\[sec:gp\]. ### The PHOEBE Model {#sec:ph_model} The modeling software combines the complete treatment of the Roche potential with the detailed treatment of surface and horizon effects such as limb darkening, reflection and gravity brightening to derive an accurate model of the binary parameters. The current implementation uses the WD method of summing over the discrete trapezoidal surface elements, which cover the distorted stellar surfaces, to determine an accurate representation of the total observed flux and, consequently, a complete set of stellar and orbital parameters. incorporates all the functionality of the WD code, but also provides an intuitive graphical user interface alongside many other improvements, including updated filters, PHOENIX model atmospheres [@Husser2013] and bindings that enable interfacing between and . To decrease the computational cost of using with , for each iteration we created the light curve model containing 2000 data points in phase space. We then unfolded this model light curve into time space prior to adding the noise model and determining the log likelihood. This allowed the light curve to be computed in a relatively short amount of time, but fit the model of the light curve combined with the instrumental and stellar trends to the complete data set. The light curve and radial velocity data were fit simultaneously. Within our models, we fit the following parameters: inclination, $i$; eccentricity, $e$; argument of periastron, the angle from the ascending node to periastron, measured in the direction of motion, $\omega$; the primary and secondary potentials, proportional to the inverse radius, $\Omega_1$ and $\Omega_2$, respectively; third light $l3$; gamma velocity, the motion of the center of mass of the binary, $\gamma$; mass ratio, $q$; and semi-major axis, $sma$. We further set the albedos (reflective properties of the stars) and gravity darkening exponents (which relate to the change in temperature of the stars due to their obliquity), to the theoretical values of $A$=0.6 [@Rucinski1969b] and $\beta$=0.32 [@Lucy1967] for stars with convective envelopes ($T_\mathrm{eff} < 7000$K) and $A$=1.0 [@Rucinski1969a] and $\beta$=1.0 [@vonZeipel1924] for stars with radiative envelopes ($T_\mathrm{eff} > 7000$K). As @Diaz-Cordoves1992 showed that the square-root limb darkening model is preferable for objects that radiate towards the IR, we applied the square-root limb darkening law to our models and updated the limb-darkening coefficients after each iteration. Limb darkening is a thorny issue, with its implications discussed in detail in @prsa2016; it affects eclipse ingress and egress and is degenerate with stellar radii. We use limb darkening coefficients computed from the PHOENIX model atmospheres for the range of temperatures, surface gravities and chemical abundances applicable to our systems. We enforced consistency by interpolating the limb darkening coefficients for any explored combination of atmosphere parameters. While there are inherent limitations of using the square-root limb darkening model, the compounded systematic effects that arise from its use (compared to other limb darkening models) are below 0.5%. ### Markov Chain Monte Carlo Derived Uncertainties\[mcmcanalysis\] Objective comparison between masses and radii measured observationally and those predicted by theory requires that realistic uncertainties be derived for the measurements. We use a Markov Chain Monte Carlo (MCMC) sampler to sample the posterior probability distribution function. At each iteration we compute the posterior log-probability distribution function: $$\log{P(\theta\|D)} = \log{P(F\|\theta)} + \log{P(RV\|\theta)} + \log{P(\theta)} + C,$$ where $D$ denotes the data, $F$ are the light curve measurements, $RV$ are the radial velocity measurements, $\theta$ is the parameter vector that contains the fitted parameters (specified in §\[sec:ph\_model\]) and $C$ is an arbitrary constant. We incorporate into our analysis to sample the probability distribution function within a Bayesian framework. A significant advantage of this is that the results are presented as posterior probability distribution functions, which indicate parameter correlations and provide more robust uncertainty estimates. explores the parameter space using a set of walkers, in our case 142. These chains begin with a uniform prior probability distribution for each parameter. At every iteration, each chain assesses its likelihood with respect to that of another chain and then elects whether to move towards that chain. The new parameters are based on the covariance matrix of the two chains. If the move increases the posterior likelihood then it is accepted; if the move decreases it then it may be accepted with a certain probability. During the initial burn-in time, Markov chains are converging towards their maximum likelihood position. The statistics of a large number of iterations provide converged posterior distributions for the model parameters. We create a model of all available data (light curves and radial velocities) for each binary star. To assess convergence, we use auto-correlation timescales. The auto-correlation timescale is used to estimate the number of iterations required to generate an independent sample, i.e. the number of iterations required for the chain to “forget” where it started. We require a minimum of 30 auto-correlation time-scales to achieve convergence. Eclipse depths are related to the temperature ratio. To obtain an accurate uncertainty for the temperature ratios, and thus secondary component temperatures, we marginalized over the primary and secondary effective temperatures, and report the posterior probability of the temperature ratio. From this, we calculate the secondary temperature and its uncertainty. The minimal rotational radial velocity that can be derived from APOGEE spectra is limited by the resolving power, allowing us to measure $v_\mathrm{rot} \gtrsim 10$km/s. We do not detect any broadening for any component in the studied systems, implying that all measured rotational velocities are below the $10$km/s detection threshold. ### [Gaussian Process Regression]{} {#sec:gp} A Gaussian process (GP) is used to model noise, both instrumental and astrophysical; it is defined as a collection of random variables for which any finite number have a Gaussian distribution with a specified covariance structure. When using GPs, we use the data to condition the GP prior so that the GPs are normally distributed with respect to the data. A significant advantage of GP modeling is that it handles correlations in the data that are poorly understood by specifying only the high-level properties of a covariance kernel. We elected to use GPs to address the issues associated with stellar variations due to spots, correlated noise and instrumental systematics that are present in the light curves for both objects. We further incorporated a white noise kernel to assess our computed uncertainties. Prior to the application of Gaussian processes, we create a combined binary star light and radial velocity curve model using and . We then apply Gaussian processes to the light curve model, which removes the need to associate the systematics in the light curve with an explicit functional form. That way, we are able to model a wide range of systematics with a small number of tunable parameters. The kernel, or covariance function, describes the similarity between two adjacent data points. For our kernel, we elected to use a term that approximates a Matern 3/2 function, which has a slowly and a rapidly varying component: $$\label{eqn:matern} k(\tau) = \sigma^2 \left[ (1+1/\epsilon)e^{-(1-\epsilon)\sqrt{3}\tau/\rho} (1+1/\epsilon)e^{-(1+\epsilon)\sqrt{3}\tau/\rho}\right]$$ Parameter $\epsilon$ controls the quality of the approximation since, in the limit as ${\epsilon \to 0}$, this becomes the Matern-3/2 function. For our computations, we set $\epsilon$ to 0.01. The Matern 3/2 function was selected as it is capable of modeling the noise, stellar variations and instrumental systematics in the light curve using only two parameters, $\sigma$ and $\rho$. We additionally incorporated a white noise term into our model. The purpose of this term is to ensure that our uncertainties are well estimated. Poorly estimated uncertainties lead to higher values for the jitter term (high in relation to the associated noise level) and the addition of white noise of the form: $$\label{eqn:noise} k(\tau_{n,m}) = \sigma^2\delta_{n,m},$$ where $\sigma$ is a tunable parameter. Masses and Radii for two Example [Kepler]{} EBs\[solvedebs\] ============================================================== Here we describe the results of KIC 2445134 and KIC 3003991 to demonstrate the pipeline introduced in §\[analysis\]. These objects were selected as they are total eclipsers, have small [$\mathit{q}$]{}, and the data are adequate to derive precise masses and radii. KIC 2445134 ----------- KIC 2445134 is comprised of an F-type dwarf primary and an M-type secondary in an 8.4 day period, and has a flux ratio of $\sim0.01$ in the [Kepler]{} bandpass. It has a [Kepler]{} magnitude of 13.55 and an H-band magnitude of 12.40. We obtained six observations of KIC 2445134 with the HRS, which we solved as an SB1 using a mid-F type HRS template, and three observations with APOGEE that we solved as an SB2 using the BT-Settl templates. These measurements constrained the spectroscopic orbital parameters and allowed us to re-examine our HRS cross-correlation functions and detect the faint companion in two epochs using a mid-M star HRS template and with correlation power slightly above the noise. Table \[rv2445134\] lists the date and corresponding barycentric Julian Date for each of our observations, the RVs we measure, and the associated spectrograph. Figure \[fig:fluxratio\] shows the BT-Settl models corresponding to the primary and secondary components, along with the flux ratio for a synthetic EB constructed from these models. We carried out the MCMC analyses described in §\[mcmcanalysis\] on the RVs in Table \[rv2445134\] and the [Kepler]{} photometry, and derived the orbital and physical parameters listed in Table \[orbpar2445134\]. Figures \[fig:model2445lc\] and\[fig:model2445rv\] depict the light and radial velocity curve models, respectively. Due to the large mass ratio, temperature ratio and the orientation of the orbit, the light curve of KIC2445134 contains Doppler boosting. We incorporated Doppler boosting into our model using the framework of @Bloemen2011 and present the Doppler boosting parameters for the primary and secondary components, $B_A$ and $B_B$, respectively, in Table \[orbpar2445134\]. Our measurement of the mass ratio, ${\ensuremath{\mathit{q}}}=0.411\pm0.001$, has a precision of better than 1%. Our derived masses (${\ensuremath{\mathit{M}_A}}=1.29\pm0.03\,{\ensuremath{\mathit{\,M_\odot}}}$, ${\ensuremath{\mathit{M}_B}}=0.53\pm0.01\,{\ensuremath{\mathit{\,M_\odot}}}$) and radii (${\ensuremath{\mathit{R}_A}}=1.42\pm0.01\,{\ensuremath{\mathit{\,R_\odot}}}$, ${\ensuremath{\mathit{R}_A}}=0.510\pm0.004\,{\ensuremath{\mathit{\,R_\odot}}}$) have measurement precisions of 1% or better, except for the primary mass which is constrained to 2%. To assess the dependency of the parameter uncertainties on our light-curve per-point uncertainties, we ran our software in the same manner as outlined above, but with the uncertainties multiplied by 0.5 and 2 for KIC2445134. The outcome was that all fundamental parameters remained within the one sigma uncertainties quoted and the uncertainties were unchanged with the exception of the secondary radius uncertainty, which changed from 0.004${\ensuremath{\mathit{\,R_\odot}}}$ to 0.005${\ensuremath{\mathit{\,R_\odot}}}$ (still providing a 1% uncertainty) for the case of the increased per-point uncertainties. KIC 3003991 ----------- KIC 3003991 represents the faint end of our EB sample, with [Kepler]{} mag = 13.9, and is comprised of a late G-star and a mid M-star in a 7.2 day orbit with a [Kepler]{} flux ratio of $\sim$0.005. We obtained six observations with HRS and six observations with APOGEE. Once again, we were able to easily detect the low mass companion in the H-band spectra, with $\alpha\sim0.10$, and used the measurement of [$\mathit{K}_B$]{} from the APOGEE spectra to re-analyze the HRS spectra and recover the companion RV for many of the HRS epochs. The relative faintness of this EB resulted in a poorer S/N than KIC 2445134 for both the APOGEE and HRS spectroscopy, which manifests as larger uncertainties on primary and secondary RVs. Table \[rv3003991\] lists the date and corresponding barycentric Julian Date for each of our observations, the RVs we measure, and the associated spectrograph. Figure \[fig:model3003lc\] and Figure \[fig:model3003rv\] depict the light curve model and phase folded spectroscopic orbit for KIC 3003991, corresponding to the orbital and physical parameters in Table \[orbpar3003991\]. Figure \[fig:3003post\] shows the posterior distributions and Figure \[fig:3003postblob\] shows distributions of quantities calculated from our model posteriors (where the calculations were performed after each iteration). We measure a mass ratio of ${\ensuremath{\mathit{q}}}=0.289\pm0.006$, and derive masses of ${\ensuremath{\mathit{M}_A}}=0.74\pm0.04\,{\ensuremath{\mathit{\,M_\odot}}}$ and ${\ensuremath{\mathit{M}_B}}=0.222\pm0.007\,{\ensuremath{\mathit{\,M_\odot}}}$, corresponding to precisions of 5% and 3%, respectively. The uncertainty on both is driven largely by the uncertainty on the semi-amplitude, which is difficult to reduce further without a significant additional spectroscopic investment. As the precision we derive for [$\mathit{M}_B$]{} complies with our program objective of 3%, it will provide useful constraints on stellar models due to the paucity of precisely measured masses in the $M\la0.25\,{\ensuremath{\mathit{\,M_\odot}}}$ regime. Additionally, the radii we measure for this EB, ${\ensuremath{\mathit{R}_A}}=0.84\pm0.01\,{\ensuremath{\mathit{\,R_\odot}}}$ and ${\ensuremath{\mathit{R}_B}}=0.250\pm0.004\,{\ensuremath{\mathit{\,R_\odot}}}$, have precisions of 1% and 2%, respectively. False Positives in our Sample ============================= We selected our EB sample (Table \[ebsampletable\]) in 2011, while the analysis pipeline being used by the [Kepler]{} team to identify EBs and substellar companions was still being refined. In a few cases, the eclipse signal detected in [Kepler]{} photometry results from a background EB that is spatially unresolved from the primary KIC target at the [Kepler]{} plate scale [@Abdul2016]. Many of these “false-positives” have subsequently been identified by the [Kepler]{} team using advanced light-curve analysis of the photometry at the pixel level, and have been flagged in the Kepler Eclipsing Binary Catalog. We included 14 stars now labeled as false-positives for which we have APOGEE or HRS spectra to look for RV variability. In all cases, these targets show no variability within the RV measurement precisions listed in §\[facilities\]. These targets, which we have dynamically verified as false positives, are identified in the last column of Table \[ebsampletable\]. Figure \[fig:falsepos\] illustrates the attained RV precision of a star (top panel) that does not have a stellar component with a period within our baseline, amounting to better than $\sim$100m/s. The magnitude and RV precision of this star is representative of stars observed from this program with the HET. The RV measurement precision of a false positive observed with APOGEE is also shown (bottom panel), and is representative of the typical RV uncertainty for our APOGEE targets. Discussion & Future Prospects ============================= The two EBs presented in §\[solvedebs\] illustrate the contributions made by our ongoing work to the direct and precise measurement of masses and radii for low-mass stars. Many systems in our EB sample have primary components with mass $M > 0.8\,M_{\odot}$, a regime where models of stellar mass and radii are well matched to observational constraints. Many have small mass ratios ($q<0.6$), and so have low-mass secondaries, specifically targeting the regime with disagreement between models and existing observations. The high mass primary allows traditional spectroscopic analysis tools to obtain metallicity for the spectra since the secondary star flux is low enough that the effect of spectral contamination is minimal. As discussed in @2012ApJ...760L...9T the stellar metallicity is an important constraint on the models, as well as being necessary to derive ages. This project aims to derive masses and radii for low mass stars using a combination of precision photometry from [Kepler]{} coupled with radial velocities from stable fiber-fed optical and near-infrared high resolution spectrographs. Figure \[fig:debcatcomp1\] depicts the mass-radius relationship for the stars in this work (filled blue and black circles), compared with those from DEBCat (open circles). Figure \[fig:debcatcomp2\] depicts the same but with \[Fe/H\] = -1 isochrones (as opposed to Solar metlicity). The primary and secondary component of KIC2445134 both show good agreement (within 1$\sigma$) with the 2Gyr Solar metalicity isochrones. The primary component of KIC3003991 shows good agreement (within 1$\sigma$) with the 10Gyr isochrones, however, the secondary component is 3$\sigma$ away. This kind of disagreement is not unusual in the literature for low mass stars . Our program with APOGEE will continue to fill in parameter space by obtaining double-line spectroscopic RVs for high-flux-contrast systems. KIC 2445134 and KIC 3003991 would have been especially challenging without the benefit of NIR spectroscopy. For these systems we succeeded in obtaining the 3% program goal for the uncertainties on the masses and radii (with the exception of the primary mass of KIC3003991), making these valuable benchmark systems. Additional results emerging from this program will further populate the mass-radius plane for low mass dwarfs, study systems with equal-mass components, and provide high precision checks on parameters derived from asteroseismology for a few, rare giants in EB systems with detected stellar oscillations. We acknowledge support from two NASA ADAP grants (NNX13AF32G and 16-ADAP16-0201), and NSF grant AST 1517592 for this project. This work was partially supported by funding from the Center for Exoplanets and Habitable Worlds. The Center for Exoplanets and Habitable Worlds is supported by the Pennsylvania State University, the Eberly College of Science, and the Pennsylvania Space Grant Consortium. We acknowledge support from NSF grant AST 1006676 and AST 1126413 in our pursuit of precision radial velocities in the NIR. This research has made use of the SIMBAD database, operated at CDS, Strasbourg, France. This publication makes use of data products from the Two Micron All Sky Survey, which is a joint project of the University of Massachusetts and the Infrared Processing and Analysis Center/California Institute of Technology, funded by the National Aeronautics and Space Administration and the National Science Foundation. Finally, we acknowledge the thorough review of the anonymous reviewer whose comments saved us from a small embarrassment. This work was based on observations with the SDSS 2.5-meter telescope. Funding for SDSS-III has been provided by the Alfred P. Sloan Foundation, the Participating Institutions, the National Science Foundation, and the U.S. Department of Energy of Science. The SDSS-III web site is http://www.sdss3.org/. SDSS-III is managed by the Astrophysical Research Consortium for the Participating Institutions of the SDSS-III Collaboration including the University of Arizona, the Brazilian Participation Group, Brookhaven National Laboratory, University of Cambridge, Carnegie Mellon University, University of Florida, the French Participation Group, the German Participation Group, Harvard University, the Instituto de Astrofísica de Canarias, the Michigan State/Notre Dame/JINA Participation Group, Johns Hopkins University, Lawrence Berkeley National Laboratory, Max Planck Institute for Astrophysics, Max Planck Institute for Extraterrestrial Physics, New Mexico State University, New York University, Ohio State University, Pennsylvania State University, University of Portsmouth, Princeton University, the Spanish Participation Group, University of Tokyo, University of Utah, Vanderbilt University, University of Virginia, University of Washington, and Yale University. Data presented herein were also obtained at the Hobby Eberly Telescope (HET), a joint project of the University of Texas at Austin, the Pennsylvania State University, Stanford University, Ludwig-Maximilians- Universitaät Munchen, and Georg-August-Universität Göttingen. The HET is named in honor of its principal benefactors, William P. Hobby and Robert E. Eberly. Appendix A - Detailed Description of Our HET HRS Pipeline {#appendixa .unnumbered} ========================================================= Our EB project makes use of an internal data reduction pipeline for the HRS on HET. For future reference with other papers in this series, and for others using this instrument on the HET, we provide a detailed description of our pipeline, which is composed of a series of independent subroutines that can be combined into nightly processing scripts; we refer to on-sky observations of science targets generically as target frames, and individual calibrations either generically as calibrations or by their specific function. As an aside, the pipeline can easily process HRS data taken in any of the possible HRS configurations, and so, as a service to other observers who may be interested in utilizing its functionality, the description we provide here is intended to be general, rather than specific to the HRS configuration listed above. A raw HRS FITS image contains three components: the main FITS header describes the instrument configuration and target exposure, while the first and second FITS extensions contain the red and blue image data, respectively. Instrument keywords in the main header can reliably reconstruct the HRS configuration used for any given observation, without relying on external log sheets or records. Keywords describing the target, such as OBSTYPE and OBJECT, are manually specified by the HET resident astronomer carrying out the observation, and are not always used consistently. To begin reducing a night’s observations, we copy the raw target and calibration images into a single working directory. The pipeline automatically classifies the various calibration and target frames, and carries out basic image processing on them. If the working directory contains images taken with different HRS configurations, each configuration will be parsed and handled automatically. First, `HRS_SPLIT` extracts the red and blue images from the multi-extension FITS file, saves them as individual zero-extension FITS images with a suffix of ’\_r’ or ’\_b’, and assigns the data contained in the original header. This step ensures that the original FITS files are not modified in any way by the pipeline. We then perform an overscan correction (`HRS_OVERSCAN`) by triming each image to the data region and subtracting a pedestal value determined by computing the median of the overscan region. Higher order corrections that could be derived from the overscan are instead corrected with the bias frames. Different HRS configurations utilize different on-chip CCD binning schemes, and `HRS_BIAS` automatically groups and combines bias frames based on their binning, producing master bias frames for each configuration used during a night. A similar process is then performed on all available flat fields (`HRS_FLAT`), although the number of possible configurations is much larger for flats than for bias frames. `HRS_FLAT` properly separates multiple sets of flats from the same configuration that were obtained with different exposure times, which is occasionally necessary when very high signal-to-noise flats are required. In addition, non-standard flat fields taken through the HRS target fibers illuminated by the Medium Resolution Spectrograph [@2003SPIE.4841.1036R] flat lamp or twilight sky are properly recognized. Each master flat field is automatically bias corrected and saved with a unique name identifying the light source, instrument configuration, and exposure time. Similar corrections are applied to individual target frames. These image processing steps can generally be carried out without any human input or supervision, and are conveniently grouped in a script, `HRS_IPROC`, which requires no input parameters or keywords. Extracting one dimensional spectra from a two dimensional image requires locating and tracing each echelle order in the target image. The HRS dispersion runs vertically across the detector (our y coordinate), and we make the simplifying (but essentially correct) assumption that the HRS slit is aligned perfectly with detector pixel rows (our x coordinate). Each order is traced independently, and each trace begins by identifying one x & y pair that intercepts the fiber profile. At that position we extract a subarray with an x dimension of 1.5 times the physical width of the fiber in pixels, which fully encloses the illuminated pixels without overlapping into the adjacent orders, and a y dimension of 4 pixels, which averages over noise and increases the speed of the tracing procedure. The illumination profile of a fiber resembles a top hat, with some variation across the illuminated pixels due to imperfect radial scrambling in the fiber. This functional form is not well represented by the Gaussian profile commonly used for slit fed spectrographs. Instead, to identify the x coordinate of the beam center we compute the derivative of the profile and fit that with a combination of two Gaussians of equal width, separated by the slit width in pixels. This parameterization reliably recovers the beam center position at the selected y coordinate. The algorithm then walks up and down the beam, computing x for each binned y. We fit these measured beam positions with a low order polynomial to derive the beam position at each dispersion position. The tracing algorithm is implemented in two separate subroutines: `HRS_MTRACE` and `HRS_ATRACE`. `HRS_MTRACE` allows the user to interactively identify beams, and annotate each with the echelle order and the type of spectrum (absorption, such as a target or flat field; or emission, such as a sky or ThAr). This information is encapsulated in a template file that fully describes the HRS configuration. `HRS_ATRACE` is fully automated, and uses previously derived templates to process routine sets of images, retracing all beams to compensate for any misalignment between the template and the target frame due to imprecision in repositioning the HRS components. Preserving the relative flux information in our science target frames requires that the master flat frames be normalized to remove both the echelle blaze and the fiber illumination profile. `HRS_FLATNORM` derives this normalization, using a procedure conceptually similar to optimal spectral extraction. Such schemes have been described extensively in the past [e.g. @1986PASP...98..609H; @1990PASP..102..183M; @2004PASP..116..362C], and we have adapted them for use with a fiber-fed echelle spectrograph. Each master flat is traced, and then each beam is rectified. The rectified beam is used to derive a dispersion dependent spatial profile for the beam, following the procedure outlined by @2004PASP..116..362C and figures therein. This profile is the desired function for normalizing the flat fields. Reversing the image rectification on the beam profile yields the normalization image that can be applied to the original master flat frame. To extract the individual target spectra, `HRS_OPTEXT` divides the normalized flat field into the two dimensional target image, and then applies an optimal extraction algorithm similar to that used for normalizing the flat fields. If multiple exposures of the same target were obtained, the user has the option to extract each image individually, or coadd them in two dimensions and extract the composite spectrum. Optimal fiber profiles are automatically derived for each beam, and provide the proper weighting functions for optimal extraction. A primary benefit often associated with optimal extraction algorithms is their ability to optimize the S/N of the extracted spectrum by minimizing the contribution of detector noise. While this is useful for some of our fainter [Kepler]{} EB targets, more useful is the algorithm’s ability to automatically identify bad pixels in the two dimensional images caused by detector defects or cosmic-rays. These pixels are automatically excluded and the weighting function at each affected dispersion element is adjusted accordingly. `HRS_OPTEXT` returns optimally extracted spectra, sum extracted spectra for comparison purposes, and wavelength dependent variances that provide realistic measures of the spectrum S/N. These data products, as well as descriptions of each echelle order and beam, are encapsulated in a FITS table for each target image. Our wavelength calibration pipeline relies on a construct similar to that behind the tracing algorithms: time intensive manual line identification defines templates for each HRS configuration, and these templates are used in an automated re-identification procedure for routine data processing. The templates are generated using the IDL based WAVECAL package to associate the wavelengths of known ThAr spectral lines with their y pixel location for each spectral order. Wavelengths are taken from the linelist of @2007MNRAS.378..221M, and typically number $\sim10-20$ per order. The set of lines comprising each order is fit with a low order polynomial (typically 4th order) to derive the template dispersion solution for that order. To calibrate a target spectrum, we automatically measure the ThAr image taken consecutively to the target observation, using the template as a guide. The dispersion solution for each order is automatically solved for, using a sigma-clipping rejection to detect poorly fit lines, and the resulting wavelength solution is applied to the corresponding target spectrum. Although the HRS is not a pressure or temperature stabilized instrument, this careful use of ThAr calibrations consecutive with target observations can yield radial velocity stability of $25\,\mathrm{m s^{-1}}$ or better for observations taken on different nights, weeks, or even months [e.g. @2012ApJ...751L..31B]. Abolfathi, B., Aguado, D. S., Aguilar, G., et al. 2018, , 235, 42 Abdul-Masih, M., Pr[š]{}a, A., Conroy, K., et al. 2016, , 151, 101 Ahn, C. P., Alexandroff, R., Allende Prieto, C., et al. 2013, arXiv:1307.7735 Allard, F., Homeier, D., & Freytag, B. 2011, 16th Cambridge Workshop on Cool Stars, Stellar Systems, and the Sun, 448, 91 Allende Prieto, C., Majewski, S. R., Schiavon, R., et al. 2008, Astronomische Nachrichten, 329, 1018 Andersen, J. 1991, , 3, 91 Andrews, J. J., Anguiano, B., Chanam[é]{}, J., et al. 2019, , 871, 42 Bender, C. F., & Simon, M. 2008, , 689, 416 Bender, C. F., Mahadevan, S., Deshpande, R., et al. 2012, , 751, L31 Berger, D. H., Gies, D. R., McAlister, H. A., et al. 2006, , 644, 475 Bloemen, S., Marsh, T. R., [Ø]{}stensen, R. H., et al. 2011, , 410, 1787 Borucki, W. J., Koch, D., Basri, G., et al. 2010, Science, 327, 977 Caldwell, D. A., Kolodziejczak, J. J., Van Cleve, J. E., et al. 2010, , 713, L92 Carter, J. A., Fabrycky, D. C., Ragozzine, D., et al. 2011, Science, 331, 562 Casewell, S. L., Raynard, L., Watson, C. A., et al. 2018, , 481, 1897 Castelli, F., & Kurucz, R. L. 2004, arXiv:astro-ph/0405087 Chabrier, G., Gallardo, J., & Baraffe, I. 2007, , 472, L17 Choi, J., Dotter, A., Conroy, C., et al. 2016, , 823, 102 Claret, A., Hauschildt, P. H., & Witte, S. 2012, , 546, A14 Conroy, K. E., Pr[š]{}a, A., Stassun, K. G., et al. 2014, , 126, 914 Cruz, P., Diaz, M., Birkby, J., et al. 2018, , 476, 5253 Cushing, M. C., Vacca, W. D., & Rayner, J. T. 2004, , 116, 362 Diaz-Cordoves, J., & Gimenez, A. 1992, , 259, 227 Dittmann, J. A., Irwin, J. M., Charbonneau, D., et al. 2017, , 836, 124 Dotter, A. 2016, , 222, 8 Doyle, L. R., Carter, J. A., Fabrycky, D. C., et al. 2011, Science, 333, 1602 Eisenstein, D. J., Weinberg, D. H., Agol, E., et al. 2011, , 142, 72 El-Badry, K., Rix, H.-W., Ting, Y.-S., et al. 2018, , 473, 5043 Faigler, S., & Mazeh, T. 2011, , 415, 3921 Faigler, S., Mazeh, T., Quinn, S. N., Latham, D. W., & Tal-Or, L. 2012, , 746, 185 Feiden, G. A., Chaboyer, B., & Dotter, A. 2011, , 740, L25 Feiden, G. A., & Chaboyer, B. 2013, , 779, 183 Foreman-Mackey, D., Hogg, D. W., Lang, D., & Goodman, J. 2013, , 125, 306 Foreman-Mackey, D., Agol, E., Ambikasaran, S., & Angus, R. 2017, Astrophysics Source Code Library, ascl:1709.008 Garc[í]{}a P[é]{}rez, A. E., Allende Prieto, C., Holtzman, J. A., et al. 2016, , 151, 144 G[ó]{}mez Maqueo Chew, Y., Morales, J. C., Faedi, F., et al. 2014, , 572, A50 Goodman, J., & Weare, J. 2010, Communications in Applied Mathematics and Computational Science, Vol. 5, No. 1, p. 65-80, 2010, 5, 65 Goodricke, J. 1783, Royal Society of London Philosophical Transactions Series I, 73, 474 Gray, D. F. 1992, Camb. Astrophys. Ser., Vol. 20,, Gunn, J. E., Siegmund, W. A., Mannery, E. J., et al. 2006, , 131, 2332 Higl, J., & Weiss, A. 2017, , 608, A62 Horne, K. 1986, , 98, 609 Husser, T.-O., Wende-von Berg, S., Dreizler, S., et al. 2013, , 553, A6 Irwin, J. M., Quinn, S. N., Berta, Z. K., et al. 2011, , 742, 123 Kesseli, A. Y., Kirkpatrick, J. D., Fajardo-Acosta, S. B., et al. 2018, arXiv:1810.07702 Kirk, B., Conroy, K., Pr[š]{}a, A., et al. 2016, , 151, 68 Konacki, M., Muterspaugh, M. W., Kulkarni, S. R., & He[ł]{}miniak, K. G. 2010, , 719, 1293 Kraus, A. L., Tucker, R. A., Thompson, M. I., Craine, E. R., & Hillenbrand, L. A. 2011, , 728, 48 Lockwood, A. C., Johnson, J. A., Bender, C. F., et al. 2014, arXiv:1402.0846 L[ó]{}pez-Morales, M. 2007, , 660, 732 Lubin, J. B., Rodriguez, J. E., Zhou, G., et al. 2017, , 844, 134 Lucy, L. B. 1967, , 65, 89 MacDonald, J., & Mullan, D. J. 2017, , 850, 58 Majewski, S. R., Schiavon, R. P., Frinchaboy, P. M., et al. 2017, , 154, 94 Markwardt, C. B. 2009, Astronomical Data Analysis Software and Systems XVIII, 411, 251 Matijevi[č]{}, G., Pr[š]{}a, A., Orosz, J. A., et al. 2012, , 143, 123 Mazeh, T., Simon, M., Prato, L., Markus, B., & Zucker, S. 2003, , 599, 1344 Mighell, K. J., & Plavchan, P. 2013, , 145, 148 Morales, J. C., Ribas, I., & Jordi, C. 2008, , 478, 507 Morales, J. C., Gallardo, J., Ribas, I., et al. 2010, , 718, 502 Mukai, K. 1990, , 102, 183 Murphy, M. T., Tzanavaris, P., Webb, J. K., & Lovis, C. 2007, , 378, 221 Nidever, D. L., Holtzman, J. A., Allende Prieto, C., et al. 2015, , 150, 173 Piskunov, N. E., & Valenti, J. A. 2002, , 385, 1095 Planck, M. 1900, V. Deut. Phys. Ges. 2, 202 Planck, M. 1900, V. Deut. Phys. Ges. 2, 237 Popper, D. M. 1980, , 18, 115 Prato, L., Simon, M., Mazeh, T., et al. 2002, , 569, 863 Pr[š]{}a, A., Batalha, N., Slawson, R. W., et al. 2011, , 141, 83 Pr[š]{}a, A., & Harmanec, P. 2012, IAU Symposium, 282, 339 Pr[š]{}a, A., & Zwitter, T. 2005, , 628, 426 , A., [Conroy]{}, K. E., [Horvat]{}, M., et al. 2016, , 227, 29 Ramsey, L. W., Adams, M. T., Barnes, T. G., et al. 1998, , 3352, 34 Ramsey, L. W., Engel, L. G., Sessions, N., et al. 2003, , 4841, 1036 Ruci[ń]{}ski, S. M. 1969, , 19, 125 Ruci[ń]{}ski, S. M. 1969, , 19, 245 Shetrone, M., Cornell, M. E., Fowler, J. R., et al. 2007, , 119, 556 Slawson, R. W., Pr[š]{}a, A., Welsh, W. F., et al. 2011, , 142, 160 Schwamb, M. E., Orosz, J. A., Carter, J. A., et al. 2013, , 768, 127 Tal-Or, L., Faigler, S., & Mazeh, T. 2015, , 580, A21 Terrien, R. C., Fleming, S. W., Mahadevan, S., et al. 2012, , 760, L9 Terrien, R. C., Mahadevan, S., Deshpande, R., et al. 2014, , 782, 61 Tonry, J., & Davis, M. 1979, , 84, 1511 Torres, G., Andersen, J., & Gim[é]{}nez, A. 2010, , 18, 67 Tull, R. G. 1998, , 3355, 387 von Zeipel, H. 1924, , 84, 702 Wilson, J. C., Hearty, F., Skrutskie, M. F., et al. 2010, , 7735, Wilson, R. E., & Devinney, E. J. 1971, , 166, 605 Wilson, R. E. 1979, , 234, 1054 Wilson, R. E. 1993, New Frontiers in Binary Star Research, 38, 91 Wilson, R. E., & Wyithe, S. B. 2003, GAIA Spectroscopy: Science and Technology, 298, 313 Wilson, R. E. 2007, Astronomical and Astrophysical Transactions, 26, 3 Wilson, R. F., Teske, J., Majewski, S. R., et al. 2018, , 155, 68 York, D. G., Adelman, J., Anderson, J. E., Jr., et al. 2000, , 120, 1579 Zasowski, G., Johnson, J. A., Frinchaboy, P. M., et al. 2013, , 146, 81 Zhou, G., Bayliss, D., Hartman, J. D., et al. 2015, , 451, 2263 Zucker, S. 2003, , 342, 1291 Zucker, S., & Mazeh, T. 1994, , 420, 806 [ccrrc]{} 2011 Aug 27 & 55800.804144 & 9.614 $\pm$ 0.0450 & & HRS\ 2011 Sep 06 & 55810.756457 & 48.652 $\pm$ 0.056 & $-$45.01 $\pm$ 3.24 & HRS\ 2011 Sep 11 & 55815.724269 & $-$15.022 $\pm$ 0.053 & & HRS\ 2011 Sep 24 & 55828.729442 & 57.916 $\pm$ 0.058 & $-$68.20 $\pm$ 3.66 & HRS\ 2011 Oct 21 & 55855.631824 & 28.320 $\pm$ 0.050 & & HRS\ 2011 Oct 29 & 55863.602364 & 39.702 $\pm$ 0.056 & & HRS\ 2011 Sep 07 & 55811.613072 & 58.55 $\pm$ 0.290 & $-$67.79 $\pm$ 1.37 & APOGEE\ 2011 Oct 06 & 55840.593338 & $-$12.33 $\pm$ 0.22 & 101.92 $\pm$ 0.89 & APOGEE\ 2011 Oct 17 & 55851.578523 & 17.48 $\pm$ 0.200 & 28.34 $\pm$ 0.81 & APOGEE\ [ccrrc]{} 2011 Oct 08 & 55842.687885 & $-$144.119 $\pm$ 0.072 & $-$50.68 $\pm$ 8.35 & HRS\ 2012 Mar 21 & 56007.990683 & $-$120.157 $\pm$ 0.065 & & HRS\ 2012 Apr 24 & 56041.897439 & $-$101.114 $\pm$ 0.064 & $-$201.51 $\pm$ 7.01 & HRS\ 2012 May 28 & 56075.798304 & $-$143.297 $\pm$ 0.061 & $-$50.92 $\pm$ 8.21 & HRS\ 2012 Jun 08 & 56086.797501 & $-$103.329 $\pm$t 0.064 & $-$185.18 $\pm$ 12.08 & HRS\ 2012 Jun 10 & 56088.766870 & $-$140.972 $\pm$ 0.06 & $-$54.93 $\pm$ 5.63 & HRS\ 2011 Sep 09 & 55813.6973400 & $-$144.110 $\pm$ 0.290 & $-$39.99 $\pm$ 4.58 & APOGEE\ 2011 Sep 19 & 55823.7244600 & $-$115.150 $\pm$ 0.200 & $-$151.93 $\pm$ 4.65 & APOGEE\ 2011 Oct 06 & 55840.6597900 & $-$106.210 $\pm$ 0.310 & $-$165.85 $\pm$ 6.30 & APOGEE\ 2011 Oct 15 & 55849.5774200 & $-$139.530 $\pm$ 0.200 & $-$69.54 $\pm$ 3.22 & APOGEE\ 2011 Oct 17 & 55851.6479100 & $-$136.690 $\pm$ 0.170 & $-$72.76 $\pm$ 3.51 & APOGEE\ 2011 Nov 01 & 55866.5693400 & $-$128.360 $\pm$ 0.170 & $-$96.77 $\pm$ 4.51 & APOGEE\ [lc]{}\ $P$ (days) & $8.41201\pm 0.00077$\ $T_{transit}$ & $2454972.647749$\ $i$ (deg) & $88.032\pm0.001$\ [$\mathit{e}$]{}& $0.00555\pm0.00001$\ [$\mathit{\omega}$]{} (rad) & $4.799\pm0.001$\ $\Omega_A$ & $15.433\pm0.002$\ $\Omega_B$ & $18.4\pm0.1$\ $L$3 (%) & $4.0\pm0.1$\ $T_{peri.}$ & $2455826.85\pm0.25$\ [$\mathit{K}_A$]{} ([$\mathrm{\,km\,s^{-1}}$]{}) & $37.2\pm0.3$\ [$\mathit{K}_B$]{} ([$\mathrm{\,km\,s^{-1}}$]{}) & $90\pm1$\ [$\mathit{\gamma}$]{} ([$\mathrm{\,km\,s^{-1}}$]{}) & $21.6\pm0.3$\ [$\mathit{q}$]{} & $0.411\pm0.001$\ $sma$ ([$\mathit{\,R_\odot}$]{})& $21.3\pm0.2$\ \ $B_A$ & $3.873\pm0.002$\ $B_B$ & $6.098\pm0.006$\ \ [$\mathit{M}_A$]{} ([$\mathit{\,M_\odot}$]{}) & $1.29\pm0.03$\ [$\mathit{M}_B$]{} ([$\mathit{\,M_\odot}$]{}) & $0.53\pm0.01$\ [$\mathit{R}_A$]{} ([$\mathit{\,R_\odot}$]{}) & $1.42\pm0.01$\ [$\mathit{R}_B$]{} ([$\mathit{\,R_\odot}$]{}) & $0.510\pm0.004$\ [$\mathit{T}_B$]{}/[$\mathit{T}_A$]{} & $0.635\pm0.001$\ $Teff B$ (K) & $3976\pm170$\ \ $\log$($\phi_\mathrm{matern}$) & $ -0.91\pm0.05$\ $\log$($\rho_\mathrm{matern}$) & $ 5.09\pm0.07$\ $\log$($\phi_\mathrm{jitter}$) & $ -9.992\pm0.004$\ [lc]{}\ $P$ (days) & $7.24478\pm 0.00062$\ $T_{transit}$ & $2454964.859085$\ $i$ (deg) & $88.178\pm0.008$\ [$\mathit{e}$]{}& $0.00030\pm0.00003$\ [$\mathit{\omega}$]{} (rad) & $4.09\pm0.03$\ $\Omega_A$ & $18.76\pm0.02$\ $\Omega_B$ & $19.9\pm0.3$\ $L$3 (%) & $3.5\pm0.2$\ $T_{peri.}$ & $2455953.65\pm0.18$\ [$\mathit{K}_A$]{} ([$\mathrm{\,km\,s^{-1}}$]{}) & $24.97\pm0.05$\ [$\mathit{K}_B$]{} ([$\mathrm{\,km\,s^{-1}}$]{}) & $83\pm1$\ [$\mathit{\gamma}$]{} ([$\mathrm{\,km\,s^{-1}}$]{}) & $-122.55\pm0.03$\ [$\mathit{q}$]{} & $0.298\pm0.006$\ $sma$ ([$\mathit{\,R_\odot}$]{})& $15.6\pm0.2$\ \ \ [$\mathit{M}_A$]{} ([$\mathit{\,M_\odot}$]{}) & $0.74\pm0.04$\ [$\mathit{M}_B$]{} ([$\mathit{\,M_\odot}$]{}) & $0.222\pm0.007$\ [$\mathit{R}_A$]{} ([$\mathit{\,R_\odot}$]{}) & $0.84\pm0.01$\ [$\mathit{R}_B$]{} ([$\mathit{\,R_\odot}$]{}) & $0.250\pm0.004$\ [$\mathit{T}_B$]{}/[$\mathit{T}_A$]{} & $0.662\pm0.001$\ $Teff B$ (K) & $3536\pm140$\ \ $\log$($\phi_{matern}$) &$-0.3\pm0.1$\ $\log$($\rho_{matern}$) &$4.3\pm0.1$\ $\log$($\phi_{jitter}$) &$-10.996\pm0.004$\ ![BT-Settl models corresponding to the components in KIC 2445134 (solid blue and red lines), and the flux ratio (dashed line). The [Kepler]{}, HRS, and APOGEE bandpasses are indicated for reference. The model spectra have been degraded to a spectral resolution of R=2,000, while the flux ratio is shown for R=100. The contrast is an order of magnitude more favorable for detecting the secondary in the H-band than in the optical.[]{data-label="fig:fluxratio"}](fig_alpha_v2.pdf){width="5.2in"} ![ light curve (black) and ten models including GPs (blue) for KIC2445134, where the different panels emphasize different regions of the light curve.[]{data-label="fig:model2445lc"}](2445_lc_all.pdf){width="5.2in"} ![Phase folded RVs for KIC2445134 derived from the HET spectra, squares and circles, and the APOGEE spectra, triangels and stars, for the primary and secondary components, respectively. The (upper panel) depicts the radial velocities and ten models from the final iteration (where each model is depicted by a red or blue line). The (middle panel) and (lower panel) depict the residuals from the average model fit to the primary and secondary radial velocity data, respectively. The residuals are measured in $\kms$.[]{data-label="fig:model2445rv"}](2445_rvs_FINAL.pdf){width="5.2in"} ![ Kepler light curve (black) and ten models including GPs (blue) for KIC3003991, where the different panels emphasize different regions of the light curve.[]{data-label="fig:model3003lc"}](3003_model.pdf){width="5.2in"} ![Phase folded RVs for KIC3003991. The layout is identical to Figure\[fig:model2445rv\]. The thickness of the lines depict the spread of the ten models caused by the model uncertainty. This is particularly significant for the secondary component (red lines).[]{data-label="fig:model3003rv"}](3003_rvs_FINAL.pdf){width="5.2in"} ![Posterior distributions for the KIC3003991 parameters derived from [Kepler]{} photometry, and APOGEE and HRS radial velocities. Boxes in the upper right contain the standard correlation coefficient, with 1 corresponding to perfect correlation, -1 perfect anti-correlation, and 0 no correlation. Final parameter values are listed along the top.[]{data-label="fig:3003post"}](3003_posteriors.pdf){width="7.in"} ![Parameter distributions for the calculated parameters of KIC3003991. Layout is as Figure\[fig:3003post\][]{data-label="fig:3003postblob"}](3003_posteriors_blob.pdf){width="6.5in"} ![Radius versus mass, plotted for the EBs in the DEBCat Eclipsing Binary Catalog (open circles), along with the two EBs we describe in §\[solvedebs\], where KIC2445134 is depicted with solid blue circles and KIC3003991 is depicted with solid black circles. For comparison, we show ( Isochrones and Stellar Tracks) stellar isochrones (@Dotter2016 [@Choi2016]) for solar metallicity stars. The inserts depict magnified regions for the secondary components of KIC2445134 (left insert) and KIC3003991 (right insert). The paucity of precision measurements below $\sim0.8\,{\ensuremath{\mathit{\,M_\odot}}}$, and particularly below $\sim0.25\,{\ensuremath{\mathit{\,M_\odot}}}$, is evident.[]{data-label="fig:debcatcomp1"}](apogee_iso_00.pdf){width="5.2in" height="4.8in"} ![The same as Figure\[fig:debcatcomp1\], but with \[Fe/H\]=-1. In this case, the primary component of KIC3003991 now agrees with the 10Gyr isochrone, however, the secondary component does not. This is inline with a known disparity between theoretical stellar models and observations in the low mass region, and further highlights the importance of modeling low mass binary components .[]{data-label="fig:debcatcomp2"}](apogee_iso_-1.pdf){width="5.2in" height="4.8in"} ![RV measurements of KIC 3248033 identified as a false-positive based on pixel level analysis of the photometry and of a red-giant KIC 2438070 with no stellar eclipsing companion. Our RV measurements dynamically verify that the eclipse signal in the light-curve data results from an unresolved background EB which is highly diluted by the foreground KIC target.[]{data-label="fig:falsepos"}](False_positives_v2.pdf){width="5.2in"} [^1]: <http://www.astro.keele.ac.uk/jkt/debcat/> [^2]: http://keplerebs.villanova.edu/	{ "pile_set_name": "ArXiv" }
--- abstract: 'The geometric phases of the cyclic states of a generalized harmonic oscillator with nonadiabatic time-periodic parameters are discussed in the framework of squeezed state. A class of cyclic states are expressed as a superposition of an infinte number of squeezed states. Then, their geometric phases are obtained explicitly and found to be $-(n+1/2)$ times the classical nonadiabatic Hannay angle. It is shown that the analysis based on squeezed state approach provide a clear picture of the geometric meaning of the quantal phase.' address: \| $^1$ Department of Physics and Centre for Nonlinear Studies, Hong Kong Baptist University, Hong Kong, China\ $^2$ Institute of Applied Physics and Computational Mathematics, P.O.Box.8009, 100088 Beijing, China\ $^3$ Department of Physics, University of Houston, Houston, TX 77204 author: - 'Jie Liu $^{1,2}$, Bambi Hu $^{1,3}$, and Baowen Li$^1$' title: ' Nonadiabatic Geometric Phase and Hannay Angle: A Squeezed State Approach ' --- [2]{} Berry phase[@Berry84], which reveals the gauge structure associated with a phase shift in adiabatic processes in quantum mechanics, has attracted great theoretical interests and has been repeatedly corroborated by experiments (see e.g.[@SW89]). This quantum adiabatic phase has a classical analog -Hannay angle [@HB85]. The relaxation of the adiabatic approximation is an important step [@AA87; @BH88]. Aharanov and Anandan [@AA87] studied the phase associated with a cyclic evolution in quantum mechanics (which occurs when a state returns to its initial condition), and shown that the phase is a geometric property of the curve in the projective Hilbert space which is naturally associated with the motion. The significance of Aharanov and Anandan’s generalization are twofold. On the one hand, the cyclic evolution of a physical system is of most interest in physics both experimentally and theoretically. On the other hand, the universal existence of the cyclic evolution is guaranteed for any quantum system. This can be easily recognized by considering the eigenvectors of the unitary evolution operator for a quantum system. An explicit example is a time-periodic Hamiltonian system where the Floquet theorem applies. The eigenfunctions of the Floquet operator, which are so-called the Bloch wave functions in the condensed matter physics, are obviously cyclic solutions and of great interest in physics. Unlike the adiabatic case, however, in the nonadiabatic case, calculating the eigenvectors and extracting the nonadiabatic geometric phase from the quasi-energy term for a time-dependent Hamiltonan is far from trival, except for such a special example as the spin particle in a magnetic field. Recent works of Ge and Child[@GC97] made a step further in this direction. They found a [special]{} cyclic state of Gaussian wave packet’s form for a generalized harmonic oscillator. The nonadiabatic geometric phase is explicitly calculated and found to be one half of the classical nonadiabatic Hannay angle. In this letter, we would like to study the nonadiabatic geometric phase of the [general]{} cyclic evolutions of the generalized harmonic oscillator. To this end, an alternative way - squeezed state approach will be used. In particular, we shall construct a class of quantum states based on a superposition of an infinite number of squeezed states. We find that the condition for them to be cyclic evolutions is nothing but a quantization rule without Maslov-Morse correction. The nonadiabatic geometric phases are obtained [explicitly]{}, and found to be related to the classical Hannay angle by a factor $n+1/2$. Furthermore, the quantum phase can be interpreted as a sum of the area difference on the expectation value plane through a cannonical transformation and the area on the quantum fluctuation plane swept out by a periodic orbit. This interpretation gives a unified picture of the geometric meaning of the quantal phase for the adiabatic and nonadiabatic case. Squeezed state approach has been successfully applied in many branches of physics such as quantum optics, high energy physics and condensed matter physics. Recent years have witnessed a growing application of squeezed state to study the chaotic dynamical systems[@PS94; @ZF95; @LS97; @HLLZ98]. In this letter, we shall employ this approach to discuss geometric phase and Hannay angle for a generalized harmonic oscillator. An apparent reason for this choice is that this system admits the squeezed state as an exact solution. The squeezed state approach[@JK79; @ZFG90; @TF91] starts from the time-dependent variational principle (TDVP) formulation, $$\delta \int dt\langle\Phi,t\|i\hbar\frac{\partial}{\partial t} - \hat H \|\Phi,t\rangle = 0.$$ Variation w.r.t $\langle\Phi,t\|$ and $\|\Phi,t\rangle$ gives rise to the Schrödinger equation and its complex conjugate, respectively. The squeezed state is chosen as the trial wave function, which is defined by the ordinary harmonic oscillator displacement operator acting on a squeezed vacuum state $\|0\rangle$: $$\|\Psi\rangle = \exp\left(\alpha \hat a^{+} - \alpha^{} \hat a\right)\|\phi\rangle,$$ $$\|\phi\rangle = \exp\left(\frac 1 2 (\beta{\hat a^{+2}} - \beta {\hat a}^2)\right)\|0\rangle.$$ ${\hat a}^{+}$ and ${\hat a}$ are boson creation and annihilation operator which satisfy the canonical commutation relation: $[\hat a,{\hat a}^+] =1$. From the TDVP, we can obtain the dynamical equations for the expectation values $(q,p)$ and the quantum fluctuations $ \Delta p^2 \equiv \langle\Psi,t\|(\hat p-p)^2\|\Psi,t \rangle=\hbar(\frac{1}{4G}+4\Pi^2G), \Delta q^2 \equiv \langle\Psi,t\|(\hat q-q)^2\|\Psi,t\rangle=\hbar G, $ $$\begin{aligned} \dot q =\frac{\partial H_{eff}}{\partial p},\qquad \dot p = -\frac{\partial H_{eff}}{\partial q},\nonumber\\ \hbar \dot G = \frac{\partial H_{eff}}{\partial \Pi},\qquad \hbar \dot \Pi = -\frac{\partial H_{eff}}{\partial G},\end{aligned}$$ where the dot denotes the time derivative. The effective Hamiltonian $H_{eff}$ is defined on the extended space $(q,p,G,\Pi)$, taking the form $H_{eff} = \langle\Psi,t\|\hat H\|\Psi,t\rangle$. The time-dependent variational principle leaves an ambiguity of a time-dependent phase $\lambda (t)$, which can be fixed with the aid of the Schrödinger equation, $$\dot \lambda(t) = \langle\Psi,t\|i\frac{\partial}{\partial t}\| \Psi,t\rangle -\frac{1}{\hbar}\langle\Psi,t\|\hat H\|\Psi,t\rangle.$$ This phase is well defined for general [nonadiabatic]{} and [noncyclic]{} evolution of a squeezed state. It represents a phase change of the squeezed state during a time-evolution. Obviously, the phase consists of two parts. The meaning of the second part is obvious: a measure of the time of evolution. It is the [dynamical]{} phase and can be rewritten as, $$\lambda_D(t) = -\frac{1}{\hbar}\int_0^t H_{eff} dt.$$ The first part can be viewed as a difference of the [total]{} phase and the [dynamical]{} phase. We call it [geometric]{} phase since it just is the Aharanov-Anandan’s phase for the case of cyclic evolution. From the expression of the squeezed state, the geometric phase is equal to $$\lambda_G(t) =\int_0^t \left(\frac{1}{2\hbar}(p\dot q - q\dot p) -\dot\Pi G\right) dt.$$ It is clear that the evolution of expectation values $(q,p)$ as well as the evolution of the quantum fluctuations $(G,\Pi)$ contribute to the geometric phase. The contribution from the former one is explicitly $\hbar$ dependent, while the contribution from quantum fluctuation is $\hbar$ independent. For the case of cyclic evolution of squeezed state the quantal phase is equal to a sum of the projective areas on the coordinates plane $(q,p)$ and fluctuation plane $(G,\Pi)$ swept out by a periodic orbit of the effective Hamiltonian. The Hamiltonian of the generalized harmonic oscillator takes the form, $$\hat H(q,p,t) = \frac 1 2 \left(a(t)\hat q^2 + b(t)\hat p^2 + c(t)(\hat q \hat p + \hat p \hat q)\right), \label{GHO}$$ where real parameter ${a(t),b(t),c(t)}$ are time-periodic functions with common period $T$. Our discussions are restricted to the elliptic case, namely, $a(t)b(t) > c^2(t)$. Applying the squeezed state to this system, from Eq. (\[GHO\]) one obtains an effective Hamiltonian in the extended phase space $(q,p;G,\Pi)$, $$H_{eff}(q,p;G,\Pi;t) = H_{cl}(q,p,t) + \hbar H_{fl}(G,\Pi,t),$$ where $$H_{cl} = \frac 1 2 \left(a(t) q^2 + b(t) p^2 + 2 c(t) q p\right),$$ describes the motion of the expectation values; $$H_{fl} = \frac 1 2 \left(a(t)G + b(t)(\frac{1}{4G} +4\Pi^2G) + 4c(t)G\Pi\right),$$ depicts the evolution of the quantum fluctuations. Starting from this effective Hamiltonian, it is easy to analyse the dynamical properties. The motions of both degree of freedom are decoupled. In the fluctuation plane $(G,\Pi)$, whole motions are restricted on the invariant tori except for a unique T-periodic solution denoted by $(G_p(t),\Pi_p(t))$. The Hamiltonian $H_{cl}$ which describes the motion of the expectation values $(q,p)$ is identical to the classical version of the system (\[GHO\]). The $(q=0,p=0)$ is obviously a fixed point. Other motions are quasi-periodic trajectories confined on the tori. Through a canonical transformation, $q=q(\bar I,\bar\phi,t), p=p(\bar I,\bar\phi,t)$, the Hamiltonian $H_{cl}(q,p,t)$ can be transformed to a new Hamiltonian $\bar H(\bar I,t)$ which does not contain angle variable $\bar\phi$. Its solution is described by $\bar I =\bar I_0; \bar\phi(t) = \bar\phi_0 + \int_0^t \frac{\partial \bar H(\bar I_0, t)} {\partial \bar I_0 } dt .$ For this canonical transformation is explicitly time dependent, the new Hamiltonian $\bar H$ differs from the old one $H_{cl}$ both in value and in functional form. Thus, we introduce a function $A$ to measure the difference, $$A(\bar \phi, \bar I,t) = \bar H(\bar I,t) - H_{cl}\left((\phi(\bar\phi,\bar I,t), I (\bar \phi,\bar I,t),t\right).$$ Therefore the classical non-adiabatic Hannay angle is $$\Theta_H =\langle \int_0^T \frac{\partial A}{\partial \bar I}dt \rangle_ {\bar \phi_0}, \label{Hannay}$$ where the bracket denotes averaging around the invariant torus, $\langle \cdots\rangle=\frac{1}{2\pi}\int_0^{2\pi}\cdots d\bar\phi_0$. Now we turn to the quantum system (\[GHO\]). Since it is a time-periodic Hamiltonian system, the Floquet theory applies. A unitary time evolution operator refering to one period T, the so-called Floquet operator $\hat U(T)$ is worthy of consideration. We can construct a state as a superposition of infinite number of squeezed states $$\|S_1\rangle = c\int_0^{2\pi} e^{\frac{i}{\hbar} \bar I_0 \bar\phi_0}\| \bar I_0,\bar\phi_0;G_0,\Pi_0\rangle d\bar\phi_0 ,$$ where $\|\bar I_0,\bar\phi_0;G_0,\Pi_0\rangle$ represents a squeezed state centered at $ q(\bar I_0,\bar \phi_0,t=0) ,p(\bar I_0,\bar \phi_0,t=0)$ with fluctuations $G_0,\Pi_0$; The $G_0$ and $\Pi_0$ are chosen on the unique periodic orbit $(G_0=G_p(t=0), \Pi_0=\Pi_p(t=0))$ ; $c$ is a normalization constant. Consider the situation that $\hat U(mT)$ (or $\hat U^m(T)$) acts on the state $\|S_1\rangle$, $$\hat U(mT)\|S_1\rangle = c\int_0^{2\pi}e^{\frac{i}{\hbar}\bar I_0 \bar \phi_0} e^{i\lambda} \|\bar I_0,\bar\phi_0+\bar \phi^m ;G_0,\Pi_0\rangle d\bar \phi_0,$$ where $\bar \phi^m=\int_0^{mT}\frac{\partial \bar H(\bar I_0,t)}{\partial \bar I_0}dt$, and $ \lambda = \lambda_D(mT) + \lambda_G(mT). $ The dynamical part is $ \lambda_D(mT)=-\frac{1}{\hbar}\int_0^{mT} H_{eff}dt,\quad $ and the geometric part $ \lambda_G(mT) = \frac{1}{\hbar}\int_0^{mT}\frac {1}{2} (p\dot q -q\dot p)dt -\int_0^{mT}\dot\Pi_p G_p dt. $ They can be expressed as, $$\lambda_D(mT)= \langle\lambda_D(mT)\rangle_{\bar \phi_0} + \{\lambda_D(mT) \}(\bar\phi_0), \label{pch1}$$ $$\lambda_G(mT)=\langle\lambda_G(mT)\rangle_{\bar \phi_0} + \{\lambda_G(mT) \}(\bar\phi_0). \label{pch2}$$ respectively. Where the symbols $\langle \cdots\rangle_{\bar\phi_0}$ denotes the average over the $\bar \phi_0$ as in Eq.(\[Hannay\]); $\{\cdots\}(\bar\phi_0)$ represent the terms relating to $\bar\phi_0$. Then, $$\langle\lambda_G(mT)\rangle_{\bar\phi_0} = \frac{m}{\hbar}\langle\int_0^T \left(\frac 1 2 (p\dot q -q\dot p)\right)dt\rangle_{\bar\phi_0} -m\oint G_p d\Pi_p .$$ Making variables transformation $\bar\phi_0' = \bar\phi_0 + \bar \phi^m$, we have $$\hat U(mT)\|S_1\rangle = c e^{i\lambda_m^1} \int_{\bar \phi^m}^{2\pi+\bar \phi^m} e^{\frac{i}{\hbar}\bar I_0\bar\phi_0'}$$ $$e^{i\{\lambda_D(mT)\}(\bar\phi_0')+i\{\lambda_G(mT)\}(\bar\phi_0')}\| \bar I_0,\bar \phi_0';G_0,\Pi_0\rangle d\bar\phi_0' , \label{ums1}$$ where $ \lambda_m^1 = m (\lambda_G^R+\lambda_D^R). $ The geometrical part and the dynamical part take the forms as follows, $$\lambda_G^R = \frac{1}{\hbar} \left(\langle\int_0^T \frac 1 2 (p\dot q -q\dot p)dt\rangle_{\phi_0} -\bar I_0\int_0^T\frac{\partial \bar H}{\partial \bar I_0} dt\right)$$ $$-\oint G_p d\Pi_p. \label{Fgphase}$$ $$\lambda_D^R= -\frac{1}{\hbar} \langle\int_0^T H_{eff} dt\rangle_{\phi_0}.$$ The integral in Eq.(\[ums1\]) can be written as $ \int_{0}^{2\pi}\cdots + \int_{2\pi }^{2\pi+\bar \phi^m}\cdots - \int_{0 }^{\bar \phi^m} \cdots. $ The last two terms will cancell each other if and only if $ e^{\frac{i}{\hbar} \bar I_0 2\pi}=1$, which gives rise to $$\bar I_0 = n\hbar . \label{quru}$$ This is nothing but the quantization rule without Maslov-Morse correction. The motion of the expectation values $(q,p)$ confined on the invariant torus $\bar I_0$ is quasi-periodic. The ergodicity of the motion guarantees that temporal average is equivalent to the spatial average supposing that the time is long enough. Then, from the ergodicity principle, we can choose an integer $r$, which is large enough so that the phase change (see(\[pch1\]) and (\[pch2\])) during the time interval $rT$ does not relate to $\bar\phi_0$. Then, we construct a state $\|S_r\rangle$ like[@LHL98], $$\|S_r\rangle = \|S_1\rangle + \cdots + e^{-i\lambda_m^1}\hat U(mT)\|S_1\rangle + \cdots$$ $$+e^{-i\lambda_{r-1}^1}\hat U((r-1)T)\|S_1\rangle. \label{stater}$$ and under the condition (\[quru\]) we can prove that[@LHL98], $$\hat U(T)\|S_r\rangle = e^{i(\lambda_D^R +\lambda_G^R)}\|S_r\rangle.$$ In fact, the above relation indicates that the state $\|S_r\rangle$ is an eigenstate of the Floquet operator, $n$ is the state number. $\lambda_D^R$ and $\lambda_G^R$ is the dynamical and geometric phase relating to the cyclic states , respectively. To see the meaning of the geometric phase $\lambda_G^R$ expressed by Eq.(\[Fgphase\]), let us consider following differential 2-form which is preserved under the canonical transformation, i.e. $ dp\wedge dq -dH\wedge dt = d\bar I\wedge d\bar \phi -d\bar H\wedge dt. $ We rewrite this into another form, $$dp\wedge dq - d\bar I\wedge d\bar \phi = -d(\bar H -H_{cl})\wedge dt. \label{2form}$$ Let us first make an integration of the above equation for one period ($T$) and then average over the variable $\bar\phi_0$ . Keeping in mind that the area meaning of the differential 2-form, One will find immediately that the term bracketed in the expression of the geometric phase Eq.(\[Fgphase\]) corresponds to the left hand side of the above equation, whereas the right hand side will equal to $n\hbar$ times the classical Hanny’s angle (see Eq. (\[Hannay\])). The $\frac 1 2$ relation between the last term in Eq.(\[Fgphase\]) and the classical angle is given by Ge and Child[@GC97] and verified by our explicit perturbative results in follows. Then, we can reach a simple relation between the geometric phase and non-adiabatic Hannay angle, $$\lambda_G^R=-(n+\frac 1 2)\Theta_H . \label{Gphagle}$$ Now we take a specific choice of the periodic parameters as an example to demonstrate the above approach and verify our findings. Set that $a(t)=1+\epsilon \cos(\omega t), b(t)=1-\epsilon \cos(\omega t) , c(t)=\epsilon\sin(\omega t)$. Our discussions are restricted to the elliptic case, namely, $a(t)b(t) > c^2(t)$, i.e. $\epsilon < 1$. The perturbation method will be employed in the following discussions. Our solutions of power series are accurate to second order. Now, we rewrite the classical Hamiltonian in terms of the action-angle variables, i.e. $q=\sqrt{2I}\sin\phi,\quad p=\sqrt{2I}\cos\phi$, $$H_{cl} = H_0 (I) + \epsilon H_1(I,\phi), \label{Hcl}$$ where $H_0 = I, H_1 = -I\cos(\omega t + 2\phi)$. It is convenient to employ the Lie transformation[@LL83] method to make a canonical transformation, so that the new Hamiltonian $\bar H(\bar I)$ contains the action variable only, $$\bar H(\bar I) = \bar I -\frac{\bar I}{\omega+2} \epsilon^2. \label{HI}$$ The generating functions are $w_1=I\sin (\omega t + 2 \phi)/(\omega +2)$ and $w_2 = 0$, respectively. The relation between the old variables and the new variables is given by $(\phi, I)= {\cal T}^{-1} (\bar \phi, \bar I)$, where the transformation operator ${\cal T}^{-1} = 1 + \epsilon {\cal L}_1 + \epsilon^2 ({\cal L}_2/2+{\cal L}_1^2/2). {\cal L}_n$ is Lie operator defined by ${\cal L}_n = [w_n,~]. [~,~]$ represents a Poission Bracket. With the help of Eqs. (\[Hannay\]) and (\[Hcl\],\[HI\]) we arrive at the expression of the classical angle analytically, $$\Theta_H = \frac{2\pi\epsilon^2}{(\omega +2)^2}. \label{Hannay2}$$ Obviously, this classical non-adiabatic Hannay’s angle is independent of the action. T-periodic solution $(G_p(t),\Pi_p(t))$ of the Hamiltonian $H_{fl}$, can be derived by using the power-series expansion, $$G_p(t)=\frac {1} {2} -\frac{\cos(\omega t)}{\omega+2}\epsilon,\qquad \Pi_p(t)=-\frac{\sin(\omega t)}{\omega+2}\epsilon. \label{GPperiod}$$ Notice the fact that an arbitrary $\omega$ can be approached by a series of rational number like $q/p$, we can repeat the above process by constructing a state as in Eq.(\[stater\]), where the $r=q$[@LHL98]. Finally, we obtain the analytic expression of the geometric phase, $$\lambda_G^R=-(\frac{\bar I_0}{\hbar}+\frac 1 2) \frac{2\pi\epsilon^2}{(\omega+2)^2}.$$ Considering the quantization rule (${\bar I_0}=n\hbar$) and the explicit expression of Hannay’s angle (\[Hannay2\]), the above equation coincise with the relation (\[Gphagle\]). An interesting example is given by the case $n=0$, i.e. the ground state of the Floquet states. It corresponds to a cyclic squeezed state with period $T$, whose expectation values keep fixed at the zero point, while its fluctuations change periodically (see Eq.(\[GPperiod\])). The geometric phase of this cyclic state resulting only from the periodic evolution of the fluctuations’ part, is equal to one half of the classical Hannay angle. This is just what obtained by Ge and Child[@GC97]. In summary, the squeezed state approach is used to study the nonadiabatic geometric phase relating to the cyclic evolutions of a generalized harmonic oscillator. The quantum phases are obtained explicitly and found to be $-(n+1/2)$ times the Hannay angle. The quantum phase can be interpreted as a sum of the area difference on the expectation value plane through the cannonical transformation and the area on the quantum fluctuation plane swept out by a periodic orbit. The explanation given here provides a unified picture of the geometric meaning of the quantal phase for the adiabatic case as well as the nonadiabatic case. In the adiabatic limit, our $n+1/2$ relation is identical to the elegant formula of Berry[@HB85]. However, the semiclassical approximation has not been envoked.\ We would like to thank Profs. Shi-Gang Chen, Lei-Han Tang, and Wei-Mou Zheng for helpful discussions and comments. This work was supported in part by the grants from the Hong Kong Research Grants Council (RGC) and the Hong Kong Baptist University Faculty Research Grants (FRG). [99]{} M. V. Berry, Proc. Roy. Soc. London, A [392]{}, 45 (1984). , ed. A. Shapere and F. Wilczek. (1989), World Scientific. J. H. Hannay, J. Phys. A [18]{}, 221 (1985); M. V. Berry, J. Phys. A [18]{}, 15 (1985). Y. Aharonov and J. Anandan, Phys. Rev. Lett. [58]{}, 1593 (1987). M. V. Berry and J. H. Hannay, J. Phys. A [21]{}, L325 (1988). Y. C. Ge and M. S. Child, Phys. Rev. Lett. [78]{}, 2507 (1997). A. K. Pattanayak and W. C. Schieve, Phys. Rev. Lett. [72]{}, 2855 (1994) W. M. Zhang and D. H. Feng, Phys. Rep. [252]{}, 1 (1995), and the references therein. W. V. Liu and W. C. Schieve, Phys. Rev. Lett. [78]{}, 3278 (1997) B. Hu, B. Li, J. Liu and J. L. Zhou, Phys. Rev. E [58]{}, 1743 (1998). R. Jackiw and A. Kerman, Phys. Lett. A [71]{}, 158 (1979). W. M. Zhang, D. H. Feng and R. Gilmore, Rev. Mod. Phys. [62]{}, 867 (1990). Y. Tsui and Y. Fujiwara, Prog. Theor. Phys. [86]{}, 443 (1991). J. Liu, B. Hu, and B. Li, Report No. HKBU-CNS-9807, quant-th/9806076. A. J. Lichtenberg and M. A. Lieberman, [Regular and Stochastic Motion*]{}, P123, Springer-Verlag, (1983).	{ "pile_set_name": "ArXiv" }
[**]{}\ [*]{}\ Vincenzo Branchina[^1]\[one\] Department of Physics, University of Catania and\ INFN, Sezione di Catania, Via Santa Sofia 64, I-95123, Catania, Italy Marco Di Liberto[^2]\[two\] Scuola Superiore di Catania, Via S. Nullo 5/i, Catania, Italy Ivano Lodato[^3]\[three\] Scuola Superiore di Catania, Via S. Nullo 5/i, Catania, Italy and\ INFN, Sezione di Catania, Via Santa Sofia 64, I-95123, Catania, Italy [Abstract]{}\ The fluctuation-dissipation theorem (FDT) is very general and applies to a broad variety of different physical phenomena in condensed matter physics. With the help of the FDT and following the famous work of Caldeira and Leggett, we show that, whenever linear response theory applies, any generic bosonic or fermionic system at finite temperature $T$ can be mapped onto a fictitious system of free harmonic oscillators. To the best of our knowledge, this is the first time that such a mapping is explicitly worked out. This finding provides further theoretical support to the phenomenological harmonic oscillator models commonly used in condensed matter. Moreover, our result helps in clarifying an interpretation issue related to the presence and physical origin of the Bose-Einstein factor in the FDT. Introduction ============ The idea of modeling physical systems as a collection of harmonic oscillators has a long history and dates back to even before the birth of quantum mechanics. One of the best known example is Planck’s work on black body radiation at the edge of the classical (beginning of quantum) era. More recently, this model has been of great importance in connection with the study of dissipation in quantum mechanics[@seni; @ford; @mohri]. In their famous paper devoted to the study of tunneling in dissipative systems[@caleg2], Caldeira and Legget observed that “any physical system which is weakly perturbed around its equilibrium state can be adequately represented (at T=0 at least) by regarding that system as equivalent to a set of simple harmonic oscillators”. They supplemented this statement with an explicit computation, where they showed that given a quantum system at $T=0$, in the lowest order approximation, its dynamics can be reproduced with the help of a properly constructed system of harmonic oscillators. They also stressed that the study of the $T \neq 0$ case (not considered in their paper) needed separate discussion. Caldeira and Legget then suggested the form of the total lagrangian of a physical system in interaction with a certain “environment” (the “heat bath”) and worked out the consequences of this assumption in connection with the tunneling problem[@caleg2; @caleg1]. Prompted by this pioneering work, harmonic oscillator models are nowadays extensively used and there is little doubt that, from a phenomenological point of view, they reproduce quite well the physics of the systems under investigation. As a specific example, we can consider an electrical circuit, where the resistance is modeled with a collection of harmonically oscillating electrical dipoles. From a theoretical point of view, however, it would be more satisfactory if we could prove that, for any generic system at finite temperature $T \neq 0$, it is possible to find an equivalent system of harmonic oscillators such that the statistical (thermodynamical) properties of the real physical system are properly reproduced by the system of oscillators. This would be an extension of the Caldeira Legget result and would provide further theoretical support to the commonly used phenomenological models. The main scope of this work is to present a new and very general result which provides the above mentioned extension of the Caldeira-Legget one. By working within the framework of the FDT[@cawe], we show that, whenever linear response theory applies, any generic bosonic and/or fermionic system at finite temperature can be mapped onto the Fock space of a fictitious system of free harmonic oscillators at the same temperature. As a byproduct of our analysis, we shall see that our finding should help in clarifying an interpretation issue concerning the Bose-Einstein (BE) distribution factor which appears in the FDT. Actually, an often raised question concerns the physical meaning and/or origin of the BE factor which appears in the relation between the power spectrum of the fluctuating quantity and the corresponding generalized susceptibility. Sometimes this term is interpreted as due to an harmonic oscillator composition of the physical system under investigation. Such an interpretation, however, is not supported by the derivation of the theorem itself (see for instance[@kubo; @jetz2; @noi]). Moreover, the FDT applies to any generic bosonic or fermionic system (irrespectively of its statistics). Far from being an academic question, the resolution of this interpretation issue is of very practical importance in many different contexts[@koch; @koch1; @bema1; @bema2]. From a real understanding of the origin of this term often depends the correct physical interpretation of theoretical and experimental results[@jetz2; @noi; @jetz1; @doran; @maha; @taylor]. As we shall see, our results suggest that this term does not originate from underlying physical oscillator degrees of freedom of the system but is rather a general property related to the approximation (linear response) involved in the derivation of the FDT. According to this theorem, whenever linear response theory is applicable, given a generic system which interacts with an external field $f(t)$ through the interaction term $\hat V = - f(t)\,\hat A$, where $\hat A$ is an observable of the system, the mean square of the Fourier transform $\hat A(\omega)$ of $\hat A(t)$ is related to the imaginary part $\chi_{_A}''(\omega)$ of the corresponding (Fourier transformed) generalized susceptibility by the relation[@cawe] : \[fdt1\] \^2()= \_[\_A]{}\^[”]{}() [coth]{}( ) =2 \_[\_A]{}\^[”]{}() (1 2 + ), where $\beta=1/k T$, with $T$ the temperature of the system and $k$ the Boltzmann constant. For instance, in the case of a resistively shunted Josephson junction[@koch], when applied to the power spectrum $S_I(\omega)$ of the noise current (fluctuation) in the resistive shunt (dissipation), the theorem takes the form ($R$ is the shunt resistance)[@koch1]: \[spectr\] S\_I() = (+ ) . The power spectrum $S_I(\omega)$ has been measured[@koch] and good agreement between the experimental results and Eq.(\[spectr\]) has been found. The above $\frac{\hbar\omega}{2}$ term is sometimes presented[@koch; @koch1; @gardizo; @kogan] as due to zero point energies and the experimental results[@koch] as a measurement of them. In fact, the term in parenthesis in Eq.(\[spectr\]) coincides with the mean energy of an harmonic oscillator of frequency $\omega$ in a thermal bath. The same holds true for the general case of Eq.(\[fdt1\]), where the similar term is the mean energy of an harmonic oscillator in $\hbar\omega$ units, i.e.the BE distribution function. In the following we shall see that our result (the mapping) strongly suggests that the agreement between the experimental results[@koch] and Eq.(\[spectr\]) cannot be considered as a signature of measurement of zero point energies. The rest of the paper is organized as follows. In section 2 we briefly review the derivation of the FDT and establish some relations useful for the following. In section 3 we establish our new result, the [mapping]{}, i.e. we show that for any given bosonic or fermionic system at finite temperature $T \neq 0$ we can always find a fictitious system of harmonic oscillators in such a manner that the physical quantities which appear in the FDT can be obtained from this equivalent system of oscillators. Section 4 is for our comments and conclusions. In particular, in this last section we present our comments on the interpretation issue related to the BE term in the FDT. The fluctuation-dissipation theorem =================================== Let us begin by briefly reviewing the derivation of the FDT. Consider a macroscopic system with unperturbed Hamiltonian $\hat{H}_0$ under the influence of the perturbation \[inter\] = - f(t)(t), where $\hat{A}(t)$ is an observable (a bosonic operator) of the system and $f(t)$ an external generalized force[^4]. Let $\|E_n\rangle$ be the $\hat{H}_0$ eigenstates (with eigenvalues $E_n$) and $\langle E_n\|\hat{A}(t)\|E_n \rangle =0$. Within the framework of linear response theory, the quantum-statistical average $\langle\hat{A}(t)\rangle_f$ of the observable $\hat{A}(t)$ in the presence of $\hat{V}$ is given by \[resp2\] (t)\_f = \_[-]{}\^t d t’ \_[\_[A]{}]{}(t-t’) f(t’) where $\chi_{_{A}}(t - t')$ is the generalized susceptibility, \[chi\] \_[\_[A]{}]{}(t - t’)=(t-t’) = -G\_R(t - t’) , with $\langle ... \rangle = \sum_{n} \varrho_n \langle E_n\| ... \|E_n \rangle$, $\varrho_n= e^{-\beta E_n}/Z$ , $Z=\sum_n e^{- \beta E_n}$, $G_R(t-t')$ being the retarded Green’s function and $\hat{A}(t)=e^{i\hat{H_0}t/\hbar}\hat{A}e^{-i\hat{H_0}t/\hbar}$. Defining the correlators (from now on $t^{'}=0$): \[correla\] G\_[>]{}(t)= (t)(0) [and]{} G\_[<]{}(t)= (0)(t) , so that $ G_R(t)=-i\theta(t)(G_{>}(t) - G_{<}(t))$, and the corresponding Fourier transforms, $ G_{>}(\omega)$ and $ G_{<}(\omega)$ respectively, it is a matter of few lines to show that: \[aux\] G\_[>]{}()=-G\_R() ; G\_[<]{}()=e\^[-]{}G\_[>]{}() . Finally, by noting that \[oo2\] \^2()= 12 (G\_[>]{}() + G\_[<]{}()) and that the Fourier transform of $\chi_{_{A}}(t)$ is $\chi_{_{A}}(\omega) = \chi_{_{A}}^{\,'}(\omega) + i \chi_{_{A}}^{\,''}(\omega) = -\frac{1}{\hbar} G_{R}(\omega)$ we get: \[fddt\] \^2()= \_[\_[A]{}]{}\^[”]{}() = \_[\_[A]{}]{}\^[”]{}() [coth]{}( ) =2\_[\_[A]{}]{}\^[”]{}() (1 2 + ) which is Eq.(\[fdt1\]), the celebrated FDT. As observed by Kubo et al.[@kubo] (and shown in the derivation sketched above), the BE factor is simply due to a peculiar combination of Boltzmann factors in Eq.(\[fddt\]) and there is no reference to physical harmonic oscillators of the system whatsoever. Despite such an authoritative remark, some people insist in interpreting the $\left(\frac1 2 + \frac{1}{e^{\beta\hbar\omega}-1}\right)$ term as related to harmonic oscillator degrees of freedom of the physical system. In the case of the measured[@koch] power spectrum of Eq.(\[spectr\]), some authors[@bema1; @bema2] interpret this term as due to the electromagnetic field in the resistive shunt and therefore the first term in parenthesis of Eq.(\[spectr\]) as originating from zero point energies of this electromagnetic field. Such an interpretation, however, is not supported by any physical derivation and has been strongly criticized in[@jetz2; @jetz1; @doran; @maha]. Very recently, starting from the results of the present work, we have also carefully investigated this issue[@noi], providing arguments which strongly support previous criticisms (see[@noi] for details). Let us go back now to our analysis. For our purposes, it is useful to show that from Eqs.(\[chi\]) and (\[oo2\]) we can easily derive the following expressions for $\chi_{_A}^{\,''}(\omega)$ and $\langle \hat{A}^2(\omega)\rangle$: \[chi2\] \_[\_A]{}”()=\_[i, j]{}\_i \|A\_[i j]{}\|\^2 , and \[o2\] \^2()=\_[i, j]{}\_i \|A\_[ij]{}\|\^2 , where $A_{ij}=\langle E_i\|\hat A \| E_j\rangle $ . In fact, by inserting in Eq.(\[chi\]) the expressions:$\theta(t-t')=-\int_{-\infty}^{+\infty}\frac{d\omega} {2\pi i}\frac{e^{-i\omega(t-t')}}{\omega+i\eta}$ , $I=\sum_i \|E_i\rangle\langle E_i\|$ and $\hat{A}(t)=e^{i\hat{H_0}t/\hbar}\hat{A}e^{-i\hat{H_0}t/\hbar}$ we get: $$\chi_{_A}(t-t')=-\frac1 \hbar \int \frac{d\omega}{2\pi}\frac{e^{-i\omega(t-t')}}{\omega+i\eta} \sum_{i, j} \varrho_i \|A_{ij}\|^2 \left(e^{i(E_i-E_j)(t-t')/\hbar}- e^{-i(E_i-E_j)(t-t')/\hbar} \right)$$ Then, making use of $\lim_{\eta\rightarrow 0}\frac{1}{\omega+i\eta}= \mathcal{P} \left(\frac1 \omega\right) - i\pi\delta(\omega)$, Eq.(\[chi2\]) follows immediately. As for Eq.(\[o2\]), from Eq.(\[correla\]) for $G_>(t)$, we have: \[uno\] G\_>(t) = \_[i,j]{}\_i E\_i\| e\^[i H t]{}A e\^[-i H t]{}\|E\_jE\_j\|A\|E\_i=\ =\_[i,j]{}\_i e\^[-i (E\_j - E\_i)t]{}\|A\_[ij]{}\|\^2. Working out the similar expression for $G_<(t)$, for the correlation function $G(t)$ we get: $$\label{due} G(t)=\frac1 2 (G_>(t) + G_<(t))=\frac1 2 \sum_{i , j}\rho_i \|A_{ij}\|^2 (e^{-\frac i \hbar(E_j - E_i)t}+e^{-\frac i \hbar(E_i - E_j)t})\, ,$$ so that the Fourier transform $\tilde G(\omega)$ is: \[ft\] G() =\_[i , j]{}\_i\|A\_[ij]{}\|\^2. As $\langle\hat{A}^2\rangle = G(0)$, $\tilde G(\omega)$ is the spectral density $\langle \hat{A}^2(\omega)\rangle$ of $\langle\hat{A}^2\rangle$ ($\langle \hat{A}^2\rangle=\int_{-\infty}^{+\infty}\langle \hat{A}^2(\omega) \rangle\frac{d\omega}{2\pi}$). Then, making use of Eq.(\[ft\]), we finally get Eq.(\[o2\]). For our purposes, it is also useful to write Eq.(\[o2\]) in a different manner. After some straightforward manipulations, Eq.(\[o2\]) can be written as: \^2()&=& \_[j > i]{}(\_i - \_j) \|A\_[ij]{}\|\^2 [coth]{}( ) \[o5\]\ &=& [coth]{}()\_[j > i]{} (\_i - \_j) \|A\_[ij]{}\|\^2 ,\[o6\] where we have introduced the notation: $\omega_{ji}=\frac{E_j-E_i}{\hbar}$. By following similar steps, Eq.(\[chi2\]) can also be written as: \[chii3\] ”()=\_[j > i]{}(\_i - \_j) \|A\_[i j]{}\|\^2 . Clearly, comparing Eq.(\[o6\]) with Eq.(\[chii3\]), we find, as we should, the FDT theorem. Now, starting from Eqs.(\[o5\]) and (\[chii3\]) and taking inspiration from the seminal work of Caldeira and Leggett[@caleg2], we shall be able to establish a formal mapping between the real system considered so far and a system of fictitious harmonic oscillators. A similar mapping, restricted however to the $T=0$ case, was considered in[@caleg2], where it was also noted that the $T\neq 0$ case needs separate discussion. The mapping that we are going to construct in the present work deals with the $T\neq 0$ general case. Mapping Boson and Fermion systems onto harmonic oscillators =========================================================== To prepare the basis for the construction of this mapping, let us consider first a real system ${\cal S}_{osc}$ of harmonic oscillators (each of which is labeled below by the double index $\{ji\}$ for reasons that will become clear in the following) whose free Hamiltonian is: \[armonico\] H\_[osc]{} = \_[j > i]{}( + q\_[ji]{}\^[2]{}), where $\omega_{ji}$ are the proper frequencies of the individual harmonic oscillators and $M_{ji}$ their masses. Let $\| n_{j i}\rangle$ ($n_{ji}=0, 1,2,...$) be the occupation number states of the $\{ji\}$ oscillator out of which the Fock space of ${\cal S}_{osc}$ is built up. Let us consider also ${\cal S}_{osc}$ in interaction with an external system through the one-particle operator: \[armint\] V\_[osc]{} = - f(t) A\_[osc]{}, with \[onepart\] [A]{}\_[osc]{} = \_[j > i]{} (\_[j i]{} [q]{}\_[ji]{} ). Obviously, the FDT applied to ${\cal S}_{osc}$ gives $\langle {\hat A}_{osc}^2(\omega)\rangle = \hbar \chi_{osc}^{\,''}(\omega) \,{\rm coth}\left( \frac{\beta\hbar\omega}{2}\right)$, but this is not what matters to us. What is important for our purposes is that, differently from any other generic system, for ${\cal S}_{osc}$ we can exactly compute $\langle {\hat A}_{osc}^2(\omega) \rangle$ and $\chi_{osc}^{\,''}(\omega)$ from Eqs.(\[chi2\]) and (\[o2\]) because we can explicitly compute the matrix elements of ${\hat A}_{osc}$. In fact, if we apply Eqs.(\[uno\]) and (\[ft\]) to ${\cal S}_{osc}$ and replace the double index notation ($\omega_{ji}$ ; $n_{ji}$ ; $M_{ji}$ ; etc.) with the more convenient (for the time being) and self explanatory one index notation ($\omega_{1}$, $\omega_{2}$, ... ; $n_{1}$, $n_{2}$, ... ; $M_{1}$, $M_{2}$, ... ; etc.), for $\langle \hat{A}_{osc}^2(\omega)\rangle$ we have: \[osc\] \_[osc]{}\^2()&=&\_[n\_1,n\_2,..]{} \_[m\_1,m\_2,..]{}(\_[n\_1]{} \_[n\_2]{}) \|n\_1,n\_2,..\|\_[osc]{} \| m\_1,m\_2,..\|\^2\ &&, where $l_k=n_k-m_k$, $\varrho_{n_k} = e^{-\beta (n_k + 1/2)\hbar\omega_k}/Z_k$, $Z_k = \sum_{n_k} e^{-\beta (n_k + 1/2)\hbar\omega_k}$ (note also that in this one index notation $\hat{A}_{osc}$ is written as ${\hat A}_{osc} = \sum_{k} \left(\alpha_{k} \,{\hat q}_{k} \right)$). Now, as n\_k\| [q]{}\_[k]{} \| m\_k= ( n\_k + 1\| m\_k + n\_k - 1\| m\_k ) , we immediately get: \[osc2\] \_[osc]{}\^2()&=&\_[n\_1,n\_2,..]{} \_[m\_1,m\_2,..]{}(\_[n\_1]{} \_[n\_2]{}) \ &&\^2\ && . Let us concentrate our attention to the square in the second line of Eq.(\[osc2\]). Due to the presence of the Kronecker deltas, all the crossed terms in this square, i.e. all the terms with different values of the index $k$, vanish. In other words, the square of the sum is equal to the sum of the squares: \^2\ = \_k (\_k ( \_[m\_k, n\_k+1]{} + \_[m\_k, n\_k-1]{}) \_[hk]{} \_[m\_h, n\_h]{})\^2 For the same reason, the same holds true for each value of the index $k$, i.e.: &&(\_k ( \_[m\_k, n\_k+1]{} + \_[m\_k, n\_k-1]{}) \_[hk]{} \_[m\_h, n\_h]{})\^2\ &=&\_k\^2 ( (n\_k + 1) \_[m\_k, n\_k+1]{} + n\_k \_[m\_k, n\_k-1]{}) \_[hk]{} \_[m\_h, n\_h]{}\[osscc\]. Therefore, as $l_k=n_k-m_k $, for $\langle \hat{A}_{osc}^2(\omega)\rangle$ we get: \[osc3\] \_[osc]{}\^2()=\_[n\_1,n\_2,..]{} (\_[n\_1]{} \_[n\_2]{}) \_k \_k\^2 (2n\_k + 1) ((-\_k) +(+\_k) ). Finally, as $ \sum_{n_k}\, \varrho_{n_k}=1 $, the above expression becomes: \_[osc]{}\^2()&=&\_k \_k\^2 ((-\_k) +(+\_k)) \_[n\_k]{} \_[n\_k]{}(2n\_k + 1) \[si\]\ &=&\_k \_k\^2 [coth]{}() ((-\_k) +(+\_k))\[osc4\]. Going back to the original double index notation: \_[osc]{}\^2()&=&\_[j>i]{} \_[ji]{}\^2 [coth]{}() ((-\_[ji]{}) +(+\_[ji]{}))\[osc5\]\ &=& [coth]{}()\_[j>i]{} \_[ji]{}\^2 ((-\_[ji]{}) - (+\_[ji]{}))\[oscc5\]. We have just seen that given a real system ${\cal S}_{osc}$ of harmonic oscillators and the one particle operator $\hat{A}_{osc}$ of Eq.(\[onepart\]), for such an operator is possible to evaluate explicitly $\langle \hat{A}_{osc}^2(\omega)\rangle$. We find that each of the individual harmonic oscillators gives rise to a term ${\rm coth}\left(\frac{\beta \hbar\omega_{ji}}{2}\right)$ which in turn comes from the term $\sum_{n_{ji}}\, \varrho_{n_{ji}}\,(2\,n_{ji} + 1)$ of Eq.(\[si\]). Let us now consider $\chi_{osc}^{''} (\omega)$, which (see Eqs.(\[chi2\]) and (\[osc2\])) is nothing but: \[cchi2\] \^[”]{}\_[osc]{} ()&=& \_[n\_1,n\_2,..]{} \_[m\_1,m\_2,..]{}(\_[n\_1]{} \_[n\_2]{}) \ &&\^2\ && . Apart from the factor $1/\hbar$, Eq.(\[cchi2\]) differs from Eq.(\[osc2\]) because it contains the difference (rather than the sum) of delta functions in the last line. If we proceed for $\chi_{osc}^{''} (\omega)$ as we have just done for $\langle \hat{A}_{osc}^2(\omega)\rangle$, we immediately note that the only difference with the previous computation is due to this minus sign. In fact, its presence causes that rather than the combination $(2\,n_k + 1)$ of Eq.(\[osc3\]), which comes from the sum $(n_k + 1) + n_k$ of Eq.(\[osscc\]), we get the combination $(n_k + 1) - n_k = 1$. Therefore, for $\chi_{osc}^{''} (\omega)$ we do not get the sum $\sum_{n_k}\, \varrho_{n_k}\,(2\,n_k + 1)= {\rm coth}\left(\frac{\beta \hbar\omega_k}{2}\right)$ of Eq.(\[si\]), but rather $\sum_{n_k}\, \varrho_{n_k}\, = 1$. Then: \[chi3\] \_[osc]{}\^[”]{} ()= \_[j>i]{} \_[ji]{}\^2 ((-\_[ji]{}) - (+\_[ji]{})). Naturally, comparing Eq.(\[oscc5\]) with Eq.(\[chi3\]) we see that for ${\cal S}_{osc}$ the FDT holds true, as it should. However, what is important for our purposes is to note that for this system we have been able to compute separately $\langle \hat{A}_{osc}^2(\omega)\rangle$ and $\chi_{osc}^{''} (\omega)$ and found that the ${\rm coth}\left(\frac{\beta \hbar\omega}{2}\right)$ factor of the FDT originates from the individual contributions ${\rm coth}\left(\frac{\beta \hbar\omega_{ji}}{2}\right)$ of each of the harmonic oscillators of ${\cal S}_{osc}$. We are now in the position to build up our mapping. Let us consider the original system ${\cal S}$, described by the unperturbed Hamiltonian $\hat H_0$, in interaction with an external field $f(t)$ through the interaction term $\hat V = - f (t)\,\hat A$ (see Eq.(\[inter\])), and construct a fictitious system of harmonic oscillators ${\cal S}_{osc}$, described by the free Hamiltonian ${\hat H}_{osc}$ of Eq.(\[armonico\]), in interaction with the same external field $f(t)$ through the interaction term ${\hat V}_{osc}$ of Eq.(\[armint\]), with $\hat A_{osc}$ given by Eq.(\[onepart\]), where for $\alpha_{j i}$ we choose: \[alfa\] \_[j i]{} = ()\^[12]{} (\_i - \_j)\^[12]{} \|A\_[ij]{}\| and for the proper frequencies $\omega_{ji}$ of the oscillators: \[omega\] \_[ji]{}= (E\_j-E\_i)/> 0, with $E_i$ the eigenvalues of the Hamiltonian ${\hat H}_0$ of the real system. Comparing Eq.(\[oscc5\]) with Eq.(\[o6\]) and Eq.(\[chi3\]) with Eq.(\[chii3\]), it is immediate to see that with the above choices of $\alpha_{j i}$ and $\omega_{ji}$ we have: \[cen1\] \^2()= \_[osc]{}\^2() [and]{} \_[\_A]{}\^[”]{} () = \_[osc]{}\^[”]{} (). Eqs.(\[alfa\]) and (\[omega\]) are the central results of our analysis. Actually, these are the equations which allow to establish our mapping. In fact, with such a choice of the $\alpha$’s and the $\omega$’s, we are able to map the real system ${\cal S}$ onto a fictitious system of harmonic oscillators ${\cal S}_{osc}$, [S]{}\_[osc]{}, in such a manner that $\chi_{_A}^{''} (\omega)$ and $\langle \hat{A}^2(\omega)\rangle$ of the real system are equivalently obtained by computing the corresponding quantities of the fictitious one (Eqs.(\[cen1\])). This is the desired result. What we have just shown is that any generic boson or fermion system at finite temperature $T$ is equivalent to a system of harmonic oscillators at the same temperature. From a theoretical point of view, the relevance of such a result should be immediately clear. As we have already observed, in fact, harmonic oscillator models are quite common in modeling generic physical systems. Now, the typical physical situation we have to deal with is that of a system (bosonic or fermionic) at finite temperature $T \neq 0$. In this respect, our mapping fills up the gap mentioned by Caldeira and Legget (see Appendix C of[@caleg2]) by extending the $T=0$ mapping put forward by them to the general finite temperature case, thus providing further theoretical support to the use of these models. In the following section, we would like to add some more comments on the above results. In particular, we are going to consider the previously mentioned interpretation issue concerning the presence and origin of the BE distribution factor in the FDT. Comments and conclusions ======================== First of all, we point out that, in order to construct the above mapping, the key ingredient we used is the hypothesis that linear response theory is applicable, which is the main hypothesis under which the FDT is established. When this is not the case, Eq.(\[resp2\]) cannot be derived and we do not arrive to Eqs.(\[chi2\]) and (\[o2\]), which are crucial to build up our mapping. Note now that, by considering the “equivalent” harmonic oscillators system ${\cal S}_{osc}$ rather than the real one, we are somehow allowed to regard the BE distribution factor ${\rm coth}\left(\frac{\beta\hbar\omega}{2}\right)$ of the FDT in Eq.(\[fdt1\]) as originating from the individual contributions ${\rm coth}\left(\frac{\beta\hbar\omega_{ji}}{2}\right)$ of each of the oscillators of the fictitious system (see above, Eqs.(\[osc5\]), (\[oscc5\]) and (\[chi3\])). In this sense, such a mapping allows for an oscillator interpretation of the BE term in the FDT. At the same time, however, our result shows that this BE factor does not describe the physics of the system, i.e.it does not encode any real, physical, harmonic oscillator degrees of freedom of the system (see also the considerations below). In this respect, it is worth to point out that what we have implemented is not a canonical transformation, i.e.it is not a transformation which allows to describe the system in terms of new degrees of freedom (such as normal modes), but a formal mapping, a mathematical construct, which can be established, we repeat ourselves, only within the framework of linear response theory. In our opinion, then, our finding provides an answer to the questions of the “physical meaning” or “physical origin” of the BE term in the FDT or, stated differently, to the question of whether this BE distribution factor possibly describes the physical nature of the system or not[@taylor]. In fact, from the derivation of the FDT, we know that the BE factor derives from a peculiar combination of Boltzmann factors (see[@kubo] and Eq.(\[fddt\]) above). At the same time, we have shown that, regardless the bosonic or fermionic nature of the (real) system ${\cal S}$, it is always possible to establish a mapping which relates ${\cal S}$ to a system of harmonic oscillators ${\cal S}_{osc}$ so that this BE factor can be regarded as “originating” from the individual oscillators of the “equivalent” system ${\cal S}_{osc}$. Therefore, it is not the physical nature of the system which is encoded in this BE term but rather a fundamental quantum property of any bosonic and/or fermionic system: [whenever linear response theory is applicable, any generic system is, at least with respect to the FDT, equivalent (in the sense defined above) to a system of quantum harmonic oscillators]{}. Before ending this section, it is probably worth to spend few words on some examples of realistic systems where our mapping is at work. In this respect, we would like to note that there are several applications in the literature where fermionic systems, after bosonization, are actually described by a system in interaction with a bosonic bath. This is, for instance, the case of the anisotropic Kondo model, which is shown to be equivalent to a spin-boson model (a two level system in interaction with a bosonic bath). The same is also true for a quantum dot interacting with external leads. In the above examples, the system is described with the help of a spin-boson Hamiltonian, thus providing concrete realizations of the mapping discussed in this work. What amounts to the same thing, they are worked out examples where the Caldeira-Legget model is explicitly derived. In this respect, in fact, it is important to note that any application of our mapping is concretely substantiated in the Caldeira-Leggett model. At the same time, we stress again that the present work is focused on the question of deeply understanding what is really behind the fact that this modelization is so successful in covering the essential features of dissipative systems. We believe that we achieve this goal by performing a thorough analysis of the fluctuation-dissipation theorem. In summary, we have found that when linear response theory applies, any generic system can be mapped onto a fictitious system of harmonic oscillators so that that the mean square $\langle \hat{A}^2(\omega)\rangle$ of the fluctuating observable and the corresponding imaginary part of the generalized susceptibility $\chi_{_A}^{''} (\omega)$ of the real system are given by the corresponding quantities of the fictitious one. Moreover, we have seen that such a mapping allows to consider the BE distribution factor which appears in the FDT as originating from the individual harmonic oscillators of the fictitious equivalent system. This strongly suggests that it is only in this sense that this BE factor can be interpreted in terms of harmonic oscillators and that no other physical meaning can be superimposed to it. We believe that our mapping has a broader range of applicability than the worked case of the FDT discussed in this paper. Work is in progress in this direction. [[Acknowledgements]{}]{} We thank Luigi Amico, Marcello Baldo, Pino Falci, J.C. Taylor and Dario Zappalà for many useful discussions. [99]{} I.R.Senitzky, Phys.Rev.[[119]{}]{}, 670 (1960). G.W.Ford, M.Kac, P. Mazur, J.Math.Phys.[[6]{}]{}, 504 (1965). K.Mohring, U.Smilansky, Nucl.Phys.[[A338]{}]{}, 227 (1980). A.O.Caldeira, A.J.Leggett, Ann.Phys.(N.Y.) [[149]{}]{}, 374 (1983). A.O.Caldeira, A.J.Leggett, Phys.Rev.Lett.[[46]{}]{}, 211 [1981]{}. H.B.Callen, T.A.Welton, Phys.Rev.[[83]{}]{}, 34 (1951). R.Kubo, M.Toda, N.Hashitsume, [Statistical Physics II]{}, Springer-Verlag, Berlin (1985). P.Jetzer, N.Straumann, Phys.Lett.[[B639]{}]{} 57 (2006). V.Branchina, M.Di Liberto, I.Lodato, J. Cosmol. Astropart. Phys. 08(2009)011. R.H.Koch, D.J.Van Harlingen, J.Clarke, Phys.Rev.[[B26]{}]{}, 74 (1982). R.H.Koch, D.J.Van Harlingen, J.Clarke, Phys.Rev.Lett.[[45]{}]{}, 2132 (1980). C.Beck, M.C.Mackey, Phys.Lett.[[B605]{}]{}, 295 (2005). C.Beck, M.C.Mackey, Physica [[A379]{}]{}, 101 (2007). P.Jetzer, N.Straumann, Phys.Lett. [[B606]{}]{} 77 (2005). M.Doran, J.Jaeckel, J. Cosmol. Astropart. Phys. 08(2006)010. G.Mahajan, S.Sarkar, T.Padmanahban, Phys.Lett.[[B641]{}]{}, 6 (2006). J.C.Taylor, J.Phys.: Condens.Matter [[19]{}]{}, 106223 (2007). C.W.Gardiner, P.Zoller, [Quantum Noise]{} (Springer, Berlin, 2000). Sh.Kogan, [Electronic Noise and Fluctuation in Solids*]{} (Cambridge University Press, Cambridge, 1996). [^1]: [email protected] [^2]: [email protected] [^3]: [email protected] [^4]: More generally, we could consider a local observable and a local generalized force, in which case we would have $\hat{V} = -\int d^3\,\vec r \hat{A}(\vec{r})f(\vec{r},t)$, and successively define a local susceptibility $\chi(\vec{r},t;\vec{r'},t')$ (see Eq.(\[chi\]) below). As this would add nothing to our argument, we shall restrict ourselves to $\vec r$-independent quantities. The extension to include local operators is immediate.	{ "pile_set_name": "ArXiv" }
--- abstract: 'Phase structure of rapidly quenched solid helium samples is studied by the NMR technique. The pulse NMR method is used for measurements of spin-lattice $T_1$ and spin-spin $T_2$ relaxation times and spin diffusion coefficient $D$ for all coexisting phases. It was found that quenched samples are two-phase systems consisting of the hcp matrix and some inclusions which are characterized by $D$ and $T_2$ values close to those in liquid phase. Such liquid-like inclusions undergo a spontaneous transition to a new state with anomalously short $T_2$ times. It is found that inclusions observed in both the states disappear on careful annealing near the melting curve. It is assumed that the liquid-like inclusions transform into a new state — a glass or a crystal with a large number of dislocations. These disordered inclusions may be responsible for the anomalous phenomena observed in supersolid region.' author: - 'A.P. Birchenko,' - 'N.P. Mikhin,' - 'E.Ya. Rudavskii,' - 'Ye.O. Vekhov' bibliography: - 'D:/vekhov/Thesis/My\_Diss/bib\_disser\_vekhov.bib' title: NMR Study of Disordered Inclusions in the Quenched Solid Helium --- B.Verkin Institute for Low Temperature Physics and Engineering of the National Academy of Sciences of Ukraine, 47 Lenin Ave., Kharkov 61103, Ukraine\ yegor\[email protected] Introduction ============ The search for the supersolid state in solid helium has recently led to the detection of an anomalous behavior first in torsional experiments [@Kim.2004] and then in investigations of elastic properties [@Day.2007], specific heat [@Lin.2007], and mass transfer [@Ray.2010]. Although the effects observed are still waiting for a consistent explanation, the anomalies are most often credited to some type of disorder that can developed in a solid. For example, these maybe a system of dislocations, grain boundaries, liquid inclusions or a disordered (glassy) phase. To understand the phenomenon of supersolid, it is fundamentally important to clear up the conditions creating a disorder in solid helium and the properties of the appearing disordered phase. Much research has been carried out along this line employing various theoretical and experimental methods. The Monte-Carlo simulations [@Boninsegni.2006; @Clark.2006; @Pollet.2007] support the view that the supersolid state is impossible in a perfect hcp crystal of $^4$He. However, the signs of a disordered glassy phase were detected in torsion experiments investigating the relaxation dynamics [@Aoki.2008; @Hunt.2009]. A large contribution of a disordered (glassy) phase to the pressure of $^4$He crystals grown on fast cooling [@Grigorev2.2007] and samples deformed $in\,situ$ [@Degtyarev.2010] was registered by precise barometry. The contribution of a glassy phase to the properties of solid helium was calculated in Refs. [@Andreev.2007.ru; @Balatsky.2007]. Interesting information was derived from visual observation of the disorder in solid $^4$He [@Sasaki.2008]: the samples grown by the blocking capillary method were polycrystals with grains of micrometer size and the liquid phase may well exist at the grain boundaries. Much information about the phase composition of a sample can be obtained by the method of nuclear magnetic resonance (NMR). For example, the investigations of spin diffusion in a region of a bcc-hcp transition [@Mikhin.2001] revealed an additional diffusive process in a sample in which both crystallographic phases coexisted. The process is characterized by high diffusion coefficients and indicated that liquid-like inclusions can form in the sample in the course of the bcc-hcp transition. Later the presence of such inclusions in pure $^4$He and a dilute $^3$He-$^4$He solid mixtures was confirmed by precise pressure measurements [@Mikhin.2007]. The subsequent NMR investigations of the hcp phase [@Vekhov.2010] showed that regions with high diffusion coefficients appeared readily in fast-grown single-phase samples. It was found that in addition to the high diffusion coefficient, the inclusion had such spin-spin relaxation times which are inherent in the liquid phase [@Birchenko.2011.LT26]. The detected liquid-like inclusions disappeared after thorough annealing. This study, which is a continuation of the previous NMR experiments [@Vekhov.2010; @Birchenko.2011.LT26], is concentrated on a detailed investigation of the revealed non-equilibrium phase. It is focused on identification of coexisting phases rather than the study of kinetics of the occurring processes. Since NMR measurements in helium are usually made on $^3$He nuclei having a non-zero magnetic moment, a $^4$He crystal for this investigation must contain a certain amount of the $^3$He impurity so that technique used could be efficient. It is especially important for measurements in the region of supersolid effects (below $\AC 300$ mK) because the anomalies in question are sensitive even to low concentration of the $^3$He impurity. That is why the NMR experiments performed at very low temperatures [@Toda.2010; @Kim.2010; @Huan.2011] could furnish important information about the state of $^3$He in the supersolid region. These NMR measurements were made at high temperatures (above $\AC 1.3$ K) to pursue another objective: to use $^3$He atoms as probes for identifying and investigating the metastable (disordered) phase in solid helium. Experimental Technique ====================== The experimental cell used for NMR measurements in solid helium is shown schematically in Fig. \[fig\_cell\]. The cell cavity (1) in a form of a cylinder 16 mm long and 8 mm in diameter was filled with helium. NMR coil (2) was wound in the inner surface of the cavity (sample). Straty-Adams pressure gauges (3) of $\pm 5-10$ mbar resolution was fixed at the cylinder ends. They were separated from the cylinder cavity with specially designed fluoroplastic displacers (4) to reduce edge effects on the NMR signal. The filling capillary (5) of the cell was made of stainless steel, its inner diameter being 0.1 mm. It was thermally connected to a 1K-pot to apply the blocking capillary method of growing solid helium samples. A resistance thermometer was mounted on the copper body (6) of the cell to measure the sample temperature (it had an accuracy of $\pm 5$ mK and a sensitivity of $\pm 1$ mK). The thermostabilizing system (heater and thermometer) attached in the upper part of a copper cooling finger (7) ensured temperature stabilization within $\pm 1$ mK. The systems of thermometry, thermostabilization, barometry, and NMR measurements were completely automated. The measurements were made in the temperature range $1.3-2.0$ K. The hcp crystals under investigation corresponded to the pressures $34-40$ bar ($V_m=20.3-20.0$ cm$^3$/mol), i.e. were grown from the normal liquid above the upper triple point (bcc-hcp-He I). The crystals were obtained by fast cooling along the melting curve at a rate of $\AC 2-6$ mK/s, which produced a large number of defects in the sample. Under such conditions the sample of solid helium contains also a large number of liquid-like inclusions that might be captured under crystallization (see Ref. [@Birchenko.2011.LT26]). The times of nuclear magnetic relaxation and the spin diffusion coefficients were measured in the NMR experiments. The measurements were made at the frequency $f_0=9.15$ MHz using sequences of probe pulses (the Carr-Purcell ($CP$) method) $90^o-\tau-180^o$ [@Carr.1954], where $\tau$ is the time interval between the pulses. The time of spin-spin relaxation $T_2$ and the spin diffusion coefficient $D$ were estimated by measuring the dependence of the relative echo-signal amplitude $h/h_0$ on the time interval $\tau$ between the probe pulses. In the $CP$ method, this dependence is [@Carr.1954]: $$\frac{h}{h_0}=\sum_{i}\alpha_i\left[1-\exp \left( \frac{\Delta t}{(T_1)_i}\right)\right]\exp\left[-\frac{2\tau}{\left(T_2\right)_i}-\frac{2}{3}\gamma^2\tau^3G^2D_i\right] , \label{DCarr}$$ where $\Delta t$ is the time interval between the pulse sequences, $G$ is the magnetic field gradient, $(T_1)_i$ is the time of spin-lattice relaxation in the i-th phase, $(T_2)_i$ is the time of spin-spin relaxation in the i-th phase, $D_i$ and $\alpha_i$ are the diffusion coefficient and the relative volume content of the i-th phase, respectively; $\gamma$ is the gyromagnetic ratio. $T_1$, $T_2$, and $D$ can be found by varying the correspondent parameters in Eq. (\[DCarr\]). The time of spin-lattice relaxation $T_1$ can be obtained from the dependence $h/h_0(\Delta t)$ on varying $\Delta t$, the other parameters being constant. The spin-spin relaxation time $T_2$ can be obtained from the dependence $h/h_0(\tau)$ which is measured under the lowest magnetic field gradient ($G$ is close to zero) on varying $\tau$ and holding $\Delta t$ constant. Finally, the diffusion coefficient $D$ is derived from the dependence $h/h_0(\tau)$ under a non-zero magnetic field gradient with a invariable $\Delta t$. If the system contains several subsystems (coexisting phases), Eq. (\[DCarr\]) gives a sum of the corresponding exponents and sought-for kinetic coefficients are obtainable from the analysis of the dependence obtained. The data obtained are reliable, which was proven by check NMR experiments in the single-phase regions of the mixture. They agree with available data for mixtures $1.94\%$ $^3$He for relaxation times [@Miyoshi.1970; @Guyer.1971], $0.75\%$ and $2.17\%$ $^3$He for diffusion coefficients [@Grigorev.1973; @Grigorev.1974] taking into account the difference in the $^3$He concentrations. Anomalous behavior of spin-spin relaxation time {#T2} =============================================== The analysis of $T_1$-, $T_2$-, and $D$-values shows that the parameter $T_2$ (spin-spin relaxation time) and $D$ (diffusion coefficient) [@Vekhov.2010; @Birchenko.2011.LT26] can provide much information necessary for identifying the phase structure of the sample. $T_2$ can tell much about the changes of the intensity of the spin motion. Such information was used to investigate the properties of solid helium in [@Allen.1982; @Mikhin.2000]. According to the Bloembergen-Purcell-Pound (BPP) model [@Bloembergen.1948], the rate of spin-spin relaxation decreases as the intensity of the relative motion of nuclei with a non-zero magnetic moment ($^3$He nuclei) increases. This correlation between $T_2$ and $D$ was quite fully considered in [@Cowan.1997.book]. On the other hand, when there is no additional strong magnetic effects (e.g. ferro- and diamagnetic impurities [@Korb.2010]) $T_2$, unlike $D$, is independent of the shape and the size of fine-grained objects. It is therefore appropriate to discuss the results obtained beginning with the time $T_2$. The starting experimental dependences $h/h_0(\tau)$ taken on the same crystal in several states are illustrated in Fig. \[fig\_T2\_hcp\]. The as-grown hcp crystal obtained on cooling along the melting curve at the rate of $\AC 5$ mK/s is shown in Fig. \[fig\_T2\_hcp\]a. The dependence $h(\tau)$ has two slopes corresponding to two processes — fast relaxation with the time $T_2=36\pm 21$ ms and slow relaxation with the time $T_2=362\pm 62$ ms. $T_2$ for the fast process measured at $T=1.7$ K is in good agreement with the spin-spin relaxation time for a bulk liquid at $P\AC 25$ bar; in the case of the slow process this time corresponds to the hcp phase. These data for the single-phase states of the investigating system were obtained in special calibration experiments [@Birchenko.2011.LT26]. They support our previous conclusion [@Vekhov.2010; @Birchenko.2011.LT26] that metastable liquid-like inclusions are formed readily in fast grown crystals of solid helium. ![(Color online) The dependence of the spin-echo amplitude on the time interval between probe pulses ($G=0$) for different states of sample: a) as-grown fast-cooled sample, $T=1.7$ K, $P=35$ bar ($\Delta t=150$ s); b) the same sample after isothermal 3 hour exposure ($\Delta t=100$ s); c) the same sample after annealing near the melting curve and subsequent cooling to $T=1.7$ K ($\Delta t=100$ s). Dashed lines — extrapolation of the dependence h/h0 for each phase.[]{data-label="fig_T2_hcp"}](fig_T2_hcp_liq+){width="1\linewidth"} ![(Color online) The dependence of the spin-echo amplitude on the time interval between probe pulses ($G=0$) for different states of sample: a) as-grown fast-cooled sample, $T=1.7$ K, $P=35$ bar ($\Delta t=150$ s); b) the same sample after isothermal 3 hour exposure ($\Delta t=100$ s); c) the same sample after annealing near the melting curve and subsequent cooling to $T=1.7$ K ($\Delta t=100$ s). Dashed lines — extrapolation of the dependence h/h0 for each phase.[]{data-label="fig_T2_hcp"}](fig_T2_hcp_glass+){width="1\linewidth"} ![(Color online) The dependence of the spin-echo amplitude on the time interval between probe pulses ($G=0$) for different states of sample: a) as-grown fast-cooled sample, $T=1.7$ K, $P=35$ bar ($\Delta t=150$ s); b) the same sample after isothermal 3 hour exposure ($\Delta t=100$ s); c) the same sample after annealing near the melting curve and subsequent cooling to $T=1.7$ K ($\Delta t=100$ s). Dashed lines — extrapolation of the dependence h/h0 for each phase.[]{data-label="fig_T2_hcp"}](fig_T2_hcp+){width="1\linewidth"} An unexpected effect was observed in the course of NMR measurements when a fast grown sample with metastable liquid-like inclusions was investigated for about three hours at $T=1.7$ K. In this case the obtained dependence $h/h_0(\tau)$ is also a superposition of two exponents (Fig. \[fig\_T2\_hcp\]b), where the time $T_2$ for the hcp phase is now $443\pm 51$ ms (within the total experimental error it can be taken as practically unchanged). The time $T_2$ for the fast process connected with the formation of liquid-like inclusions decreased by an order of magnitude and was $3.2\pm 3.0$ ms. In all the cases this change in $T_2$ was rather fast: it occurred within one measurement run, i.e. no longer than $\Delta t\AC 100-300$ s. It is natural to assume that the liquid-like inclusions undergo some evolution or a phase transition in the hcp matrix. The dependences processed by Eq. (\[DCarr\]) (Figs. \[fig\_T2\_hcp\]a and \[fig\_T2\_hcp\]b) show that the weighting factor $\alpha_i$ for non-equilibrium inclusions (see Fig. \[fig\_T2\_hcp\]b) is about three times higher than in the case of Fig. \[fig\_T2\_hcp\]a (estimation of $\alpha_i$ is detailed in Section \[discussion\]). Note that the dependence $h(\tau)$ taken after thorough ($\AC 2$ hours) annealing of the sample near the melting curve ($T=1.95$ K) and its subsequent cooling to $T=1.7$ K can be described by one exponent with $T_2=350\pm 19$ ms, which is characteristic of the hcp phase at this temperature (see Fig. \[fig\_T2\_hcp\]c). This means that after annealing the non-equilibrium inclusions disappeared almost completely and the sample became a single-phase hcp crystal. Similar results were also obtained at other temperatures (see Fig. \[fig\_T2allinclusions\]). Fig. \[fig\_T2allinclusions\] carries in addition the experimental data for the single-phase states of the system the liquid state and the hcp phase (see Ref. [@Birchenko.2011.LT26]). In the liquid state $T_2$ is practically temperature-independent (within the scatter of experimental data) in the region far from the degeneracy temperature. In the hcp phase $T_2$ decreases with lowering temperature and at $T\AC 1.5$ K the $T_2$-values of the crystal and the liquid almost coincide. The reason for this accidental coincidence is as follows: in the investigated temperature region the rate of spin-spin relaxation in the crystal is mainly determined by the concentration and the mobility of vacancies which intensify the motion of $^3$He impurity atoms. According to the BPP theory [@Bloembergen.1948], the influence of the magnetic fields of the neighboring nuclear spins is averaged more efficiently and the spin-spin relaxation slows down. The Arrhenius equation suggests that $T_2$ decreases as the temperature of the crystal lowers. As a result, the contribution of the liquid-like inclusions is practically inseparable at $T<1.7$ K. The $T_2$-values measured after the transition of the liquid-like inclusions to a new state are shown in Fig. \[fig\_T2allinclusions\]. They are over the order of magnitude lower than $T_2$ of liquid-like inclusions. Note that the transition of liquid-like inclusions to a new state was observed in most of the investigated crystals with the delay time varying from half an hour to several hours. The process was spontaneous with no distinct correlation with temperature. Discussion ========== The spin-spin relaxation time $T_2$ has been analyzed quite thoroughly (see the previous section). Unfortunately, the results obtained do not allow an unambiguous analysis of the spin-lattice relaxation time $T_1$. This is because $T_1$ is measured under operation of two mechanisms of relaxation in the liquid phase: the true relaxation is connected with recovery of equilibrium magnetization over the sample volume; a wall relaxation is determined by non-controllable interaction between $^3$He and the magnetic impurities at the cell walls. Although the experiment revealed some change in $T_1$ caused by the transition of the sample from the state with liquid-like inclusions to the new phase, an accurate quantitative analysis of the phase composition was problematic. Further experiments are therefore planned to investigate the behavior of the spin-lattice relaxation time in this system. The behavior of the diffusion coefficient $D$ in samples grown on fast cooling is analyzed using the $\tau$-dependences of the spin echo-signal amplitude. These dependences are a superposition of two exponents [@Vekhov.2010; @Birchenko.2011.LT26] — a slowly attenuating exponent and a fast attenuating one which correspond to the hcp phase and the liquid-like inclusions, respectively. Note that the $D$-values measured for liquid-like inclusions are proportional to the rate of crystal growth [@Birchenko.2011.LT26]. This suggest that the size of the metastable liquid drops captured in the process of crystallization is the smaller for the slower the crystallization rate. This may be due to a limited diffusion, like that usually observed when the size of liquid-like inclusions is comparable to the diffusion length $\sqrt{D\tau}$ (e.g., see Ref. [@Wayne.1966]). In this case the coefficient $D$ is dependent on the size of the inclusions, $\tau$ being invariable. The situation changed as the liquid-like inclusions transformed into a new state, with anomalously short times $T_2$. The echo-signal amplitudes has a wide scatter in data, which prohibited an accurate separation of the contribution made by the fast-diffusion process to the echo attenuation against the background of the hcp phase diffusion. However, the obtained dependence $h/h_0(\tau)$ could be described in terms of the slow-diffusion process typical of the hcp phase. This means that all of the liquid-like inclusions changed into the new state and the fast diffusion process did not observed. Yet, in some experiments the signs of the fast diffusion process were seen against the background of a wide data scatter, which suggest that some liquid-like inclusions did not transform into the new state. We may thus conclude that the diffusion coefficient of the new state is either smaller than $D$ of the hcp phase or it is of the same order of magnitude. According to Eq. (\[DCarr\]), correct estimation of the relative contribution of $\alpha_i$ to the amplitudes of the echo-signal of the coexisting phases is possible if the time intervals $\Delta t$ between the sequences of probe pulses in NMR measurement are several times longer than the maximum time of spin-lattice relaxation in all phases, i.e. when magnetization is recovered completely. Only under this condition does the echo signal reach its maximum value. Assuming equal $^3$He concentrations in the phases, we can obtain the weighing factor $\alpha_i$ of the i-th phase.[^1] Otherwise, magnetization of some phases can be recovered incomplete during NMR measurement because of the difference in the rates of spin-lattice relaxation in the phases, and this leads to a distortion of $\alpha_i$-values. This requirement is not always satisfied in real experiments because metastable inclusions evolve even during one run of $D$ or $T_2$ measurement. The obtained $\alpha_i$-values are therefore rather approximate. The estimation performed for annealed hcp samples using $\Delta t$ and $T_1$ shows that $\alpha_i$ of metastable inclusions can be as high as $\AC 20-30\%$. The detected new phase is a long-living metastable state (see above) which disappears only after thorough annealing of the sample. Its properties differ from those of the hcp phase and liquid. We believe that the most probable reason for the decrease in $T_2$ of these inclusions is a transition of the liquid-like inclusions having higher diffusion coefficients to a new state with low $D$. The decrease in $D$ of the liquid entails a reduction of the relaxation time $T_2$ (see Section \[T2\]). This situation is similar to solidification of a liquid without the crystal order, i.e. the transition to an amorphous or glass-like state. Previously [@Menges.1991] the feasibility of formation of the amorphous phase in solid rare gases was found from heat capacity measurements on samples grown by condensation on a cold substrate. Unfortunately we are unaware of data on the relaxation time $T_2$ and the diffusion coefficient $D$ in amorphous helium, which makes the identification of the structure of the obtained phase somewhat uncertain. This assumption was however supported by precise measurements of pressure in solid $^4$He samples grown on fast cooling [@Grigorev2.2007]. Apart from the phonon influence, an additional contribution to the pressure $P_g$ proportional to $T^2$ was observed, which occurs typically in a disordered (glassy) phase. Recently, an amorphous helium phase has also been registered in neutron diffraction experiments [@Bossy.2010]. As follows from the structural factor analysis, liquid helium confined in a porous medium holds its liquid state to the pressure $P\AC 38$ bar at $T\AC 0.4$ K and then an amorphous phase forms when liquid helium solidifies. Crystallization of liquid-like inclusions generating a large number of dislocations can be another factor responsible for the lower $T_2$-values. It is known that dislocations in crystals act as centers attracting impurities. In our case this attraction enriches the regions around dislocations in $^3$He atoms. The increasing local $^3$He concentration reduces the time of spin-spin relaxation in this region as $T_2 = 7.5\cdot 10^{-5}/x$ (s), where $x$ is the relative $^3$He concentration [@Mikhin.2000]. This dependence describes satisfactory other authors results [@Miyoshi.1970; @Mikhin.2000; @Greenberg.1972] and closely agrees with the calculation according to the Torrey theory [@Torrey.1953]. There is one more scenario of a significant decrease in the time of nuclear magnetic relaxation [@Korb.2010]. Within a limited geometry the relaxation processes accelerate significantly in the presence of magnetic impurities on the walls. In this experiment a limited geometry could be created by crystalline dendrites growing into the volume of the liquid-like inclusions [@Franck.1986], and the $^3$He atoms possessing nuclear paramagnetism could act as magnetic impurities. However, according to the Curie Law, the influence of such impurities is essential at rather low temperatures [@Mikhin.2004; @Mikhin.2005.QFS2005]. It was therefore hardly probable in this experiment. We also considered the possibility of the formation of bcc inclusions (in the $P-T$ phase diagram the bcc phase is quite close to the solid helium region under investigation). However, $T_2$ in the bcc phase is about $\AC 0.1-1$ s [@Polturak.1995], which is one or two orders of magnitude higher than the times of new state. Note that in the course of investigation on as-grown rapidly quenched samples both the pressure gauges in the cell exhibited a long-duration (up to 20 h) monotonic growth of pressure at a constant temperature. The rate of the pressure growth increased with the temperature, which suggests the thermally activated origin of the effect. The effect is attributed to the pressure gradient in the rapidly quenched samples. The pressure did not increase in samples grown carefully and slowly during $\AC 0.5-1$ h. The narrow areas in the displacers (see Fig. \[fig\_cell\]) might considerably slow down the process of pressure equalizing. Therefore, we were unable to estimate pressure change due to annealing the metastable inclusions even having two pressure gauges. Conclusion ========== The series of experiments in this study and Ref. [@Birchenko.2011.LT26] have shown that a disordered metastable long living phase is readily formed in fast-grown helium crystals. The new phase coexists with the equilibrium crystalline phase. The measured diffusion coefficient and the spin-spin relaxation time of this phase correspond at first to the values typical for the liquid phase. The size of the liquid-like inclusions is the larger if the growth rate of the crystal is higher. The liquid-like inclusions do not form in crystals grown on cooling along the melting curve at comparatively low rates. The liquid-like inclusion disappear after thorough annealing near the melting temperature. They form at the stage of the crystal growth, and their size and quantity can be controlled by varying the growth rate of the crystal. A new effect — a spontaneous transition of liquid-like inclusions to another state has been detected. The state has an anomalously short time of spin-spin relaxation unusual in both a crystal and a liquid. It is assumed that the new state is either an amorphous (glassy) phase or a crystal with a large number of dislocations. To identify the disordered phase definitely, further experimental investigations are necessary, in particular by structural methods. In the context of the supersolid problem it is important to allow for the influence of the disordered phase on the properties of helium crystals. It is likely that the metastable disordered inclusions with very short times of spin-spin relaxation may be responsible for the anomalous phenomena in the region of supersolid. The authors are indebted to V.A. Maidanov, V.D. Natsik, A.I. Prokhvatilov, and D.A. Tayurskiy for useful discussions while preparing this paper. The study was supported by the Science and Technology Center in Ukraine, Project \#5211. [^1]: The difference in the $^3$He concentration between crystalline and liquid phases is essential at temperatures below those used in this experiment.	{ "pile_set_name": "ArXiv" }
--- abstract: 'We derive the dependence of the fraction of passive central galaxies on the mass of their supermassive black holes for a sample of over 400,000 SDSS galaxies at z $<$ 0.2. Our large sample contains galaxies in a wide range of environments, with stellar masses $8 < {\rm log}(M_{}/M_{\odot}) < 12$, spanning the entire morphological spectrum from pure disks to spheroids. We derive estimates for the black hole masses from measured central velocity dispersions and bulge masses, using a variety of published empirical relationships. We find a very strong dependence of the passive fraction on black hole mass, which is largely unaffected by the details of the black hole mass estimate. Moreover, the passive fraction relationship with black hole mass remains strong and tight even at fixed values of galaxy stellar mass ($M_{}$), dark matter halo mass ($M_{\rm halo}$), and bulge-to-total stellar mass ratio ($B/T$). Whereas, the passive fraction dependence on $M_{}$, $M_{\rm halo}$ and $B/T$ is weak at fixed $M_{BH}$. These observations show that, for central galaxies, $M_{BH}$ is the strongest correlator with the passive fraction, consistent with quenching from AGN feedback.' author: - \| Asa F. L. Bluck$^{1,}$, Sara L. Ellison$^{1}$, David R. Patton$^{2}$, Luc Simard$^{3}$, J. Trevor Mendel$^{4}$, Hossein Teimoorinia$^{1}$, Jorge Moreno$^{5,6}$, and Else Starkenburg$^{7,8}$\ $^1$ Department of Physics and Astronomy, University of Victoria, 3800 Finnerty Road, Victoria, British Columbia, V8P 1A1, Canada\ $^2$ Department of Physics and Astronomy, Trent University, 1600 West Bank Drive, Peterborough, Ontario, K9J 7B8, Canada\ $^3$ National Research Council of Canada, Herzberg Institute of Astrophysics, 5071 West Saanich Road, Victoria, British Columbia, V9E 2E7, Canada\ $^4$ Max-Planck-Institut für Extraterrestrische Physik (MPE), Giessenbachstrasse, D-85748 Garching, Germany\ $^5$ Department of Physics and Astronomy, California State Polytechnic University Pomona, Pomona, CA 91768, USA\ $^6$ TAPIR, Mailcode 350-17, California Institute of Technology, Pasadena, CA 91125, USA\ $^7$ Schwarzschild Fellow, Leibniz-Institut für Astrophysik Potsdam (AIP), An der Sternwarte 16, 14482 Potsdam, Germany\ $^8$ CIfAR Global Scholar\ $$ Email: [email protected] title: 'Why do galaxies stop forming stars? I. The passive fraction - black hole mass relation for central galaxies' --- Galaxies: formation, evolution, AGN, morphology; black holes; star formation Introduction ============ Understanding why galaxies stop forming stars is an important unresolved question in galaxy evolution. Only $\sim$10% of baryons reside within galaxies (e.g. Fukugita & Peebles 2004, Shull et al. 2012), yet since galaxies lie at nodes in the cosmic web corresponding to local minima in the gravitational potential well, naively one would expect far more baryons to collate in galaxies, ultimately forming more stars. Theoretical models offer a wide range of solutions to this problem, relying on the physics of gas, stars, and black hole accretion disks as so called ‘baryonic feedback’. However, observational studies are required to test these models and provide evidence for their range and applicability. Since the discovery that most galaxies contain a supermassive black hole (e.g. Kormendy & Richstone 1995), the energy released from forming these objects has become a popular mechanism for regulating gas flows and star formation in simulations, particularly for more massive galaxies (e.g. Croton et al. 2006, Bower et al. 2006, 2008, Somerville et al. 2008, Guo et al. 2011, Vogelsberger et al. 2014a,b, Schaye et al. 2015). In fact, substantial feedback from accretion around supermassive black holes is required in cosmological semi-analytic models, semi-empirical models, and hydrodynamical simulations to achieve the steep slope of the high-mass-end of the galaxy stellar mass function (e.g. Henriques et al. 2014, Schaye et al. 2015, Vogelsberger et al. 2014a,b). Observationally, predominantly only very indirect means for linking the star formation in large populations of galaxies to their AGN have been found, e.g. the strong dependence of the passive fraction on galaxy stellar mass (e.g. Baldry et al. 2006, Peng et al. 2010, 2012) and galaxy structure (e.g. Driver et al. 2006, Cameron et al. 2009, Cameron & Driver 2009, Bluck et al. 2014). These galaxy properties both correlate (weakly) with black hole mass and, hence, the total energy released in forming the black hole (e.g. Silk & Rees 1998, Fabian 1999, Bluck et al. 2011, 2014). More direct measurements of AGN driven winds in galaxies and radio bubbles in galaxy haloes have provided evidence for the mechanisms of AGN feedback, but typically only for a very small number of galaxies (e.g. McNamara et al. 2000, Nulsen et al. 2005, McNamara et al. 2007, Dunn et al. 2010, Fabian 2012, Cicone et al. 2013). More recent work has linked the passive fraction of large populations of galaxies to the central density within 1 kpc (Cheung et al. 2012, Fang et al. 2013, Woo et al. 2014) and to the mass of the galactic bulge (Bluck et al. 2014, Lang et al. 2014). Both of these galaxy properties are expected to correlate strongly with the mass of the central black hole (e.g. Haring & Rix 2004, McConnell & Ma 2013) and hence may provide qualitative support for AGN feedback driven quenching. We find in Bluck et al. (2014) that ‘bulge mass is king’ in the sense that bulge mass is a tighter and steeper correlator to the passive fraction for centrals than any other variable considered (including stellar mass, halo mass, disk mass, local density, and galaxy structure). We argue for an explanation of this observation through AGN feedback, relying on the tight relationships between the bulge and central black hole (e.g. Haring & Rix 2004). The primary motivation for this Letter is to expand on that work by testing the role of AGN feedback in quenching central galaxies more directly. In this Letter we estimate black hole masses from their bulge masses and central velocity dispersions, using a wide range of published empirical relationships with dynamical measurements of black hole masses (e.g. McConnell & Ma 2013). From this, we derive the relationship between passive fraction and estimated black hole mass. Our goal is to test whether AGN feedback is a viable route for quenching galaxies, and test whether or not it is dominant for a large sample of central galaxies. Throughout this Letter we assume a $\Lambda$CDM cosmology with H$_{0}$ = 70 km s$^{-1}$, $\Omega_{m}$ = 0.3, $\Omega_{\Lambda}$ = 0.7, and adopt AB magnitude units. Data Overview & Passive Fraction ================================ We use the Sloan Digital Sky Survey Data Release 7 (SDSS DR7, Abazajian et al. 2009) spectroscopic sample as our data source. From this we collate a sample of 538046 galaxies (423480 centrals and 114566 satellites) with $8 < {\rm log}(M_{}/M_{\odot}) < 12$ at z $<$ 0.2. In this Letter (the first paper in the series) we concentrate on the central galaxies. The star formation rates (SFR) for these galaxies are calculated from emission lines for non-AGN star forming galaxies and from the strength of the 4000 Å break for non-emission line galaxies and AGN (Brinchmann et al. 2004). A fibre correction is applied based on galaxy colour and magnitude outside the aperture. The stellar masses for the galaxies, and their component disks and spheroids, are derived in Mendel et al. (2014), based on SED fitting to a dual Sérsic fit of the $ugriz$ wavebands (Simard et al. 2011). An n = 4 bulge and n = 1 disk model is used, and we test the reliability of this approach in Bluck et al. (2014) via model data. We define the galaxy structure (or morphology) to be $B/T = M_{\rm bulge} / M_{}$, where $M_{}$ is the total stellar mass of the galaxy. The velocity dispersions are derived from the widths of absorption lines taken from Bernardi et al. (2003) with an updated method implemented as in Bernardi et al. (2007) to the later data releases. Velocity dispersions from absorption lines with a S/N $<$ 3.5 are discarded from our sample, and those with $\sigma$ $<$ 70 km/s are deselected for some analyses, due to the instrumental resolution of the SDSS spectra. We also restrict our final sample to galaxies with an error on the velocity dispersion of $\sigma_{\rm err}$ $<$ 50 km/s. We recover $\sim$ 80 % of our parent sample which pass these data quality cuts. The halo masses used in this Letter are derived from an abundance matching technique applied to the total stellar mass of the group or cluster (from Yang et al. 2007, 2008, 2009). Testing of the group finding algorithm on model galaxies from the Millennium Simulation (Springel et al. 2005) showed that over 90 % of galaxies are correctly assigned to groups at $M_{\rm halo} > 10^{12} M_{\odot}$. Using these group catalogues, centrals are defined as the most massive galaxy in the group, with satellites being any other galaxy within the group. We follow the prescription for defining the passive fraction in Bluck et al. (2014). We define a galaxy to be passive if it is forming stars at a rate a factor of ten times lower than (emission line, non-AGN) star forming galaxies matched at the same stellar mass and redshift. This cleanly divides the two peaks of the bimodal distribution in SFR at all mass ranges. The passive fraction is the ratio of passive-to-total galaxies in each binning. We correct for the flux limit of the SDSS by weighting each galaxy in the passive fraction by the inverse of the volume over which its $ugriz$ magnitudes would pass all of the selection criteria. The errors on the passive fraction are computed in this work via the jack-knife technique. See Bluck et al. (2014) §2 & 3 for full details on these data and techniques. Black Hole Masses ================= Today there are less than 100 reliable dynamical measurements of central black hole masses in existence (see McConnell & Ma 2013 for a compilation). This is far too few to investigate the dependence of the passive fraction on black hole mass directly, due in part to the statistical nature of the passive fraction and the need for a wide range in galaxy properties. Fortunately, there exist tight relationships between dynamical measurements of central black hole masses and the properties of their host galaxies, particularly the bulge or spheroid in which the black hole resides (e.g. Magorrian et al. 1998; Ferrarese & Merritt 2000; Gebhardt et al. 2000; Haring & Rix 2004). On a case by case basis, using these relationships to estimate a given galaxy’s black hole mass would lead to a high level of uncertainty, but for very large samples the average values should yield reliable measurements, in the regime where the relationships are tested. In this work, we use a number of the galaxy - black hole relationships from the literature to estimate black hole masses. Specifically, we use central velocity dispersion ($\sigma_{c}$), stellar mass of the bulge or spheroid ($M_{\rm bulge}$), and combinations of these two. We also test separating the sample into early- and late-type systems and using separate scaling laws for each (as suggested in McConnell et al. 2011). The formulae used to estimate the black hole masses, and the references for these, are all presented in Table 1. For the velocity dispersions, we first make an aperture correction, so that all measurements are made at the same effective aperture. We use the formula in Jorgensen et al. (1995), specifically calculating: $$\sigma_{c} = \big(\frac{R_{c}}{R_{\rm app}}\big)^{-0.04} \sigma_{\rm app}$$ where the central radius, $R_{c}$, is chosen to be in line with the measurements made in the literature, with typical values of $R_{e}$ or $R_{e}/8$ used. $R_{e}$ is the bulge (or spheroid) effective radius, calculated in our sample through Sérsic index fitting (see Simard et al. 2011 for full details). We note that the aperture correction only affects the final mass estimate by typically $<$10%. For the bulge masses no such correction is necessary, and thus, this provides a natural test to the reliability of the method. Since our goal is to present the passive fraction dependence on black hole mass for all galaxies, we do not restrict the sample to pure spheroids. Thus, there is a substantial population of disks present in our sample. This poses no significant difficulty for measuring the black hole masses via bulge mass (but see Bluck et al. 2014 appendices for a discussion on their reliability). However, for the velocity dispersions, contamination in their measurement from rotating disks could be an issue. We test this in the next section by looking at early- and late-type galaxies separately, and comparing estimates of $M_{BH}$ from velocity dispersion methods to bulge mass methods. We additionally test the impact of the choice of black hole mass calibration by comparing a number of different techniques, noting a high level of consistency between them (see next section). Source log($M_{BH}/M_{\odot}$) = Morphology Dispersion (dex) $N_{\rm Gal}$ Fig. 1 Colour ---------------------------- ------------------------------------------------------------------------------------------------------------ ------------ ------------------ --------------- --------------- Hopkins et al. (2007) 0.54 $\times$ log($M_{\rm bulge} / 10^{11} M_{\odot}$) + 2.18 $\times$ log($\sigma_{c}$ / 200 km/s) + 8.24 No Cut 0.22 38 Blue Ferrarese & Merritt (2000) 4.80 $\times$ log($\sigma_{c}$) - 2.90 No Cut 0.31 12 Green Haring & Rix (2004) 1.12 $\times$ log($M_{\rm bulge} / 10^{11} M_{\odot}$) + 8.2 No Cut 0.33 30 Cyan McConnell & Ma (2013) 5.64 $\times$ log($\sigma_{c}$ / 200 km/s) + 8.32 No Cut 0.38 72 Orange McConnell & Ma (2013) 5.20 $\times$ log($\sigma_{c}$ / 200 km/s) + 8.39 ETGs 0.34 53 Red$^{}$ McConnell & Ma (2013) 5.06 $\times$ log($\sigma_{c}$ / 200 km/s) + 8.07 LTGs 0.46 19 Red$^{}$ $^{}$ for the ETG/ LTG samples, galaxies defined as early types ($B/T >$ 0.5) are fit with the ETG formula and those defined as late types ($B/T <$ 0.5) are fit with the LTG formula, they are then recombined into the same sample, and shown in red in Fig. 1. $N_{\rm Gal}$ is the number of galaxies used in the dynamical fit. Results ======= The $f_{\rm passive} - M_{BH}$ Relationship for Central Galaxies ---------------------------------------------------------------- In this section we compute, for the first time, the relationship between the fraction of passive galaxies and the mass of their central supermassive black holes, for a population of over 400,000 central galaxies (see Fig. 1). The left panel shows the $f_{\rm passive} - M_{BH}$ relationship for all central galaxies, which are defined as the most massive group members in the catalogues of Yang et al. (2009). We show this relationship via a number of different black hole mass calibrations, summarized in Table 1. These include relationships from various authors based on central velocity dispersion, bulge mass, and a combination of the two. For the full morphological sample, there is remarkably good agreement between all of the methods used for estimating black hole masses. The $f_{\rm passive} - M_{BH}$ relationship is very steep for central galaxies, with galaxies with low black hole masses being predominantly star forming and galaxies with high black hole masses being predominantly passive. The cross-over mass, where 50% of galaxies are passive occurs at $M_{BH} \sim 10^{7.5} M_{\odot}$. Values of black hole mass $< 10^{6} M_{\odot}$ are highly uncertain due to the $M_{BH} - \sigma$ , $M_{\rm bulge}$ relations not extending into this domain. Additionally, velocity dispersions $<$ 70 km/s are less reliable than for the rest of the sample (e.g. Bernardi et al. 2007), and this corresponds to approximately $M_{BH}$ $< 10^{6} M_{\odot}$ as well. For these reasons, we advise readers to treat the low $M_{BH}$ values more as upper limits (this threshold is indicated with a dashed vertical line in all relevant plots). In the middle and right panels of Fig. 1, we show the passive fraction - black hole mass relationship for centrals split into early-types ($B/T >$ 0.5) and late-types ($B/T <$ 0.5) respectively (computed in Mendel et al. 2014). For the early-type sample, all of the black hole mass estimates lead to very similar passive fraction relationships, as with the full sample. However, for the late-type galaxies, there is some disagreement between methods which use velocity dispersion alone and those which incorporate bulge mass. Specifically, the passive fraction is lower for velocity dispersion estimates of black hole mass than with bulge mass estimates. This is most probably explained by the velocity dispersions being contaminated by disk rotation into the plane of the sky, and by measurements of the absorption line widths being contaminated by multiple emission lines obscuring the true continua. Both of these effects can lead to an over-estimate of the black hole mass, and hence lead to lower passive fractions for the measured black hole masses from velocity dispersion methods compared to bulge mass methods. For bulge mass only methods contamination of the kinematic measurement in disks is clearly not an issue, but interestingly, in the case where we use $\sigma_{c}$ and $M_{\rm bulge}$ together (Hopkins et al. 2007), the contamination is also negligible (compare blue and cyan lines in Fig. 1, right panel). Thus, we posit that to estimate black hole masses for the full morphological sample (i.e. including disk galaxies) a measure of the mass of the bulge is [essential]{} to include. For the remainder of the analyses in this Letter we focus on the black hole mass estimates from bulge mass and velocity dispersion together (from Hopkins et al. 2007, see top row of Table 1), where kinematic contamination is not a significant issue. This technique draws on multiple complementary measurements of the central regions of galaxies, from spectroscopic kinematics and photometric profile and SED fitting. We note that all of our conclusions remain unchanged if we restrict to early-type galaxies only and use velocity dispersion methods to compute our black hole masses. The advantage of incorporating the bulge mass is to leave the sample complete. In an upcoming paper in the series (Paper II, Bluck et al. in prep.) we show the passive fraction relationship with black hole mass for satellites. We find that, whereas centrals with low black hole masses are essentially all star forming, satellites with low black hole masses are more frequently passive. This hints at the need for additional quenching mechanisms to AGN feedback for satellites, but not necessarily for centrals. Is the $f_{\rm passive} - M_{BH}$ Relationship Fundamental for Central Galaxies? -------------------------------------------------------------------------------- The steep dependence of the passive fraction on supermassive black hole mass for centrals (Fig. 1, left panel) is consistent with the quenching of central galaxies being driven by AGN feedback. However, given that there are observationally well known, and theoretically expected, correlations between the black hole mass (derived from $M_{\rm bulge}$ and/or $\sigma_{c}$) and other galaxy properties (e.g. $M_{\rm halo}$, $M_{}$, $B/T$), it is not immediately clear whether the $f_{\rm passive} - M_{BH}$ relationship is fundamental or not (i.e. whether it is derivative of some other, more fundamental, parameter). To test the possibility that another galaxy property is more fundamental to passivity than black hole mass, we look at the $f_{\rm passive} - M_{BH}$ relationship at a fixed values of the following galaxies properties: $M_{\rm halo}$, $M_{}$, and $B/T$ (presented in Fig. 2, top row). We find that the $f_{\rm passive} - M_{BH}$ relationship remains tight and steep even at a fixed halo mass, stellar mass and galaxy structure ($B/T$). Out of these three variables, we find the greatest variation at a fixed $M_{BH}$ with $B/T$, suggesting that the structure of a galaxy is an important secondary consideration for whether a galaxy will be forming stars or not. In Fig. 2 (bottom row) we present the $f_{\rm passive} - M_{\rm halo}, M_{}$, and $B/T$ relationships, each at fixed $M_{BH}$. These are all remarkably flat compared to the $f_{\rm passive} - M_{BH}$ relationship, suggesting that it must be more fundamental than these alternatives. Particularly, the $f_{\rm passive} - M_{\rm halo}$ relationship, at a fixed $M_{BH}$, is almost entirely flat, indicating essentially no dependence of galaxy quenching on the mass of the halo, once the correlation with black hole mass is accounted for. This has a profound implication for models of galaxy quenching. Methods that derive the energy to do work on the gas in a galaxy from the dark matter halo cannot be driving the quenching of centrals (e.g. halo mass quenching, Dekel & Birnboim 2006, Dekel et al. 2009, Woo et al. 2013). However, at high $M_{BH}$ values, there is some hint of a subtle positive correlation of the passive fraction with $M_{\rm halo}$, which may be explained if AGN feedback is more efficient in higher mass haloes (e.g. Dekel et al. 2014). The $f_{\rm passive} - M_{}$ relationship at fixed $M_{BH}$ is either flat, or in certain $M_{BH}$ ranges, [negative]{}. This striking result clearly indicates that stellar mass is not the primary driver to passivity in central galaxies (as advocated in Baldry et al. 2006, Peng et al. 2010, 2012). This result further suggests that the quenching of central galaxies cannot be driven by stellar or supernova feedback, which would both correlate with $M_{}$ (as integrated star formation rate) more strongly than with $M_{BH}$. The negative trend is most probably explained by the fact that increasing $M_{}$ at fixed $M_{BH}$ results in an increase in $M_{\rm disk}$, and hence a decrease in $B/T$. This follows directly from the $M_{BH} - M_{\rm bulge}$ relationship, and we have seen already that $B/T$ is an important [secondary]{} correlator to the passive fraction. The $f_{\rm passive} - M_{BH}$ relationship is also steeper and tighter than the $f_{\rm passive} - B/T$ relationship (which was suggested to be primary in Driver et al. 2006, Cameron et al. 2009). This implies that major galaxy mergers (which can destroy in situ disks, e.g., Cole et al. 2000) are not necessary for quenching central galaxies, but substantial central black holes are. However, see Hopkins et al. (2013) for evidence that disks can regrow in the merger process complicating this issue. Additionally, morphological quenching (e.g. Martig et al. 2009) cannot be the primary driver to the quenching of star formation in centrals, since black hole mass correlates much stronger with the passive fraction than galaxy structure. See Bluck et al. (2014) §5 for a more thorough discussion on the role of various processes in quenching central galaxies. We consider the relative role of the supermassive black hole in quenching central and satellites galaxies compared to many other galaxy properties (including bulge mass, disk mass, and local density) in an upcoming work in this series (Paper III, Teimoorinia et al., in prep.). Black hole mass remains the most important single correlator to the passive fraction in this larger analysis as well. Conclusions =========== In this Letter, we present for the first time the relationship between the fraction of passive galaxies and the mass of their supermassive black holes, for a sample of over 400,000 central galaxies at z $<$ 0.2 (Fig. 1). We find that the $f_{\rm passive} - M_{BH}$ relationship is very steep, with the vast majority of galaxies with low black hole masses being star forming and the vast majority of galaxies with high black hole masses being passive. We find a cross-over mass (where 50% of galaxies are passive) at $M_{BH} \sim 10^{7.5} M_{\odot}$. The steep dependence of the passive fraction on black hole mass is consistent with AGN feedback being the dominant mechanism for quenching central galaxies. We test splitting our sample into early- and late-types, finding generally very good accord between all of the black hole mass estimates, except for late-type galaxies. Here we find that incorporating bulge mass into the black hole mass estimate is essential because velocity dispersion alone methods are vulnerable to contamination from disk rotation. Finally, we find that black hole mass is a more fundamental correlator to the passive fraction than total stellar mass, group halo mass, or galaxy structure (Fig. 2). This is highly suggestive of the quenching of central galaxies being driven by AGN feedback. However, it is not inconceivable that the mechanism which grows the central region of galaxies (and presumably their supermassive black holes as well) is directly responsible for the quenching of central galaxies. Yet this would have to occur in such a way as to leave the central density as the primary correlator to the passive fraction, with $M_{\rm halo}$, $M_{}$, and $B/T$ being substantially less important. We conclude that this is less likely than AGN feedback being responsible. [99]{} Abazajian K. et al., 2009, ApJS, 182, 543 \ Baldry I. K., Balogh M. L., Bower R. G., Glazebrook K., Nichol R. C., Bamford S. P., Budavari T., 2006, MNRAS, 373, 469\ Bernardi M. et al., 2003, AJ, 125, 1817\ Bernardi M., Hyde J. B., Sheth R. K., Miller C. J., Nichol R. C., 2007, AJ, 133, 1741\ Bluck A. F. L., Conselice C. J., Almaini O., Laird E., Nandra K. & Gruetzbauch R., 2011, MNRAS, 410, 1174\ Bluck A. F. L., Mendel J. T., Ellison S. L., Moreno J., Simard L., Patton D. R., Starkenburg E., 2014, MNRAS, 441, 599\ Bower R. G. et al., 2006, MNRAS, 370, 645\ Bower R. G., McCarthy I. G., Benson, A. J., 2008, MNRAS, 390, 1399\ Brinchmann J., Charlot S., White S. D. M., Tremonti C., Kauffmann G., Heckman T., Brinkmann J., 2004, MNRAS, 351, 1151 \ Cameron E., Driver S. P., Graham A. W., Liske J., 2009, ApJ, 699, 105\ Cameron E., Driver S. P., 2009, A&A, 493, 489\ Cheung E. et al., 2012, ApJ, 760, 131\ Cicone C. et al., 2013, A&A in press, arXiv:1311.2595\ Cole S., Lacey C. G., Baugh C. M., Frenk C. S., 2000, MNRAS, 319, 168\ Croton D. J. et al., 2006, MNRAS, 365, 11 \ Dekel A., Birnboim Y., 2006, MNRAS, 368, 2\ Dekel A. et al., 2009, Nature, 457, 451\ Dekel A., Burkert A., 2014, MNRAS, 438, 1870\ Driver S. P. et al., 2006, MNRAS, 368, 414\ Dunn J. P. et al., 2010, ApJ, 709, 611 \ Fabian A. C., 1999, MNRAS, 303L, 59\ Fabian A. C., 2012, ARA&A, 50, 455\ Fang J. J., Faber S. M., Koo D. C., Dekel A., 2013, ApJ, 776, 63\ Ferrarese L., Merritt D., 2000, ApJ, 539L, 9\ Fukugita M., Peebles P. J. E., 2004, ApJ, 616, 643 \ Gebhardt K. et al., 2000, ApJ, 539, 13\ Guo Q. et al., 2011, MNRAS, 413, 101 \ Haring N., Rix H., 2004, ApJ, 604L, 89\ Henriques B. et al., 2014, MNRAS submitted, arXiv:1410.0365\ Hopkins P. F., Bundy K., Hernquist L., Ellis R. S., 2007, ApJ, 659, 976\ Hopkins P. F., Cox T. J., Hernquist L., Narayanan D., Hayward C. C., Murray N., 2013, MNRAS, 430, 1901 \ Jorgensen I., Franx M., Kjaergaard P., 1995, MNRAS, 276, 1341 \ Kormendy J., Richstone D., 1995, ARA&A, 33, 581 \ Lang P. et al., 2014, ApJ, 788, 11 \ Magorrian J. et al., 1998, AJ, 115, 2285\ Martig M., Bournaud F., Teyssier R., Dekel A., 2009, ApJ, 707, 250\ McConnell N. J., Ma C. P., 2013, ApJ, 764, 184\ McConnell N. J. et al. 2011, Nature, 480, 215\ McNamara B. R. et al., 2000, ApJ, 534L, 135\ McNamara B. R., Nulsen P. E. J., 2007, ARA&A, 45, 117\ Mendel J. T., Simard L., Palmer M., Ellison S. L., Patton D. R., 2014, ApJS, 210, 3 \ Nulsen P. E. J., McNamara B. R., Wise M. W., David L. P., 2005, ApJ, 628, 629 \ Peng Y. et al., 2010, ApJ, 721, 193\ Peng Y., Lilly S. J., Renzini A., Carollo M., 2012, ApJ, 757, 4 \ Schaye J. et al., 2015, MNRAS, 446, 521\ Shull J. M., Smith B. D., Danforth C. W., 2012, ApJ, 759, 23\ Silk J., Rees M. J., 1998, A&A, 331, 1\ Simard L., Mendel J. T., Patton D. R., Ellison S. L., McConnachie A. W., 2011, ApJ, 196, 11\ Somerville R. S., Hopkins P. F., Cox T. J., Robertson B. E., Hernquist L., 2008, MNRAS, 391, 481\ Springel V. et al., 2005, Nature, 435, 629 \ Vogelsberger M. et al., 2014a, Nature, 509, 177\ Vogelsberger M. et al., 2014b, 2014, MNRAS, 444, 1518 \ Woo J. et al., 2013, MNRAS, 428, 3306\ Woo J., Dekel A., Faber S. M., Koo D. C., 2014, MNRAS submitted, arXiv:1406.5372 \ Yang X., Mo H. J., van den Bosch F. C., Pasquali A. L., Cheng B. M., 2007, ApJ, 671, 153\ Yang X., Mo H. J., van den Bosch F. C., 2008, ApJ, 676, 248\ Yang X., Mo H. J., van den Bosch F. C., 2009, ApJ, 695, 900	{ "pile_set_name": "ArXiv" }
--- abstract: 'We study the 1D dynamics of dark-dark solitons in the miscible regime of two density-coupled Bose-Einstein condensates having repulsive interparticle interactions within each condensate ($g>0$). By using an adiabatic perturbation theory in the parameter $g_{12}/{g}$, we show that, contrary to the case of two solitons in scalar condensates, the interactions between solitons are attractive when the interparticle interactions between condensates are repulsive $g_{12}>0$. As a result, the relative motion of dark solitons with equal chemical potential $\mu$ is well approximated by harmonic oscillations of angular frequency $w_r=(\mu/\hbar)\sqrt{({8}/{15}){g_{12}}/{g}}$. We also show that in finite systems, the resonance of this anomalous excitation mode with the spin density mode of lowest energy gives rise to alternating dynamical instability and stability fringes as a function of the perturbative parameter. In the presence of harmonic trapping (with angular frequency $\Omega$) the solitons are driven by the superposition of two harmonic motions at a frequency given by $w^2=(\Omega/\sqrt{2})^2+w_r^2$. When $g_{12}<0$, these two oscillators compete to give rise to an overall effective potential that can be either single well or double well through a pitchfork bifurcation. All our theoretical results are compared with numerical solutions of the Gross-Pitaevskii equation for the dynamics and the Bogoliubov equations for the linear stability. A good agreement is found between them.' author: - 'I. Morera' - 'A. Muñoz Mateo' - 'A. Polls' - 'B. Juliá-Díaz' title: 'Dark-dark-soliton dynamics in two density-coupled Bose-Einstein condensates' --- Introduction ============ Over the last two decades, Bose-Einstein condensation (BEC) has enabled the study of numerous physical concepts [@Pitaevskii2003; @Pethick2008]. Among them, an interesting scenario is the connection between the non-linear waves and atomic systems, that leads to the so-called matter-wave solitons [@Kevrekidis2008; @Kevrekidis2015]. These structures emerge from the balance between linear dispersion and interactions, which are accounted for from non-linear terms in the equations of motion. Specifically, mean field descriptions of BECs, as provided by the Gross-Pitaevskii equation (GP), incorporate a non-linear term proportional to the interatomic interaction strength $g$, which is measured by the s-wave scattering length $a$. Depending on the sign and magnitude of the latter and the dimensionality of the system, a large number of structures can be found: dark [@Frantzeskakis2010] and bright [@Strecker2003] solitons, vortices [@Fetter2001], etc. Dark solitons in BECs are localized non-linear excitations that present a notch in the condensate density and a phase step across its center. They have been observed in numerous experiments using different techniques [@Burger1999; @Becker2008; @Weller2008; @Lamporesi2013; @Anderson1999] and this has inspired a large number of theoretical works (see [@Frantzeskakis2010] and references therein). Specifically, the problem of dark solitons in BECs confined by parabolic external traps has been extensively studied [@William1997; @Muryshev1999; @Busch2000; @Pelinovsky2005]. Another interesting aspect of BECs is the study of multi-component systems, which can be described by a set of coupled GP equations. These systems were soon realized in ultracold-gas experiments by coupling two condensates made of either different atomic species [@Cornell1998] or different hyperfine states of the same atomic species [@Myatt1997]. This fact inspired the study of matter-wave solitons in these settings, and new families of solitonic structures have been found: dark-dark [@Ohberg2001; @Hoefer2011; @Yan2012], dark-bright [@Middelkamp2011; @Yan2011; @Achilleos2011], dark-antidark [@Danaila2016], etc. Particular attention has been given to the so-called Manakov limit in 1D settings, where the intra- and inter-condensate particle interactions match [@Yan2012]. The aim of this work is to study the dynamics of dark-dark solitons in a one-dimensional setting of two density-coupled condensates out of the Manakov limit, for varying coupling between condensates. Specifically, we consider a mean-field description of BECs composed by two hyperfine states of the same alkali species, and inspect both the case without any axial trap, termed untrapped, and the case with a harmonic trap in the axial direction. Within the Hamiltonian approach of the perturbation theory for solitons [@Uzunov1993; @Kivshar1994; @Frantzeskakis2002], we obtain analytical expressions for the adiabatic evolution of the dark-dark soliton. For the untrapped case we find that when the interaction between components is repulsive the dark-dark soliton can be seen as a bound state of two dark solitons performing a relative harmonic motion. For attractive interactions such a bound state can not exist because the dark solitons repel each other. An equivalent study is also performed for the confined case, where the competition between soliton interactions and harmonic trapping is found to produce sizable changes on the dynamics, leading for example to bound states even with attractive interactions. Our analytical results are tested with different numerical techniques: direct simulations of the Gross-Pitaevskii equations, Fourier analysis to extract characteristic frequencies of the motion, and numerical solution of the Bogoliubov equations for the linear excitations of stationary states. In what follows, we first introduce the theoretical model in section II, and next we show our numerical results in section III. To sum up, we present our conclusions in section IV. Theoretical model ================= The dynamics of a BEC at zero temperature can be accurately described within the mean field approach in terms of a wave-function $\psi(\mathbf{r},t)$. In a one dimensional setting, the wave functions of two trapped, density-coupled BECs are governed by corresponding Gross-Pitaevskii equations $$\begin{aligned} \begin{aligned} &i\hbar \frac{\partial}{\partial t} \psi_1 = \left(\frac{-\hbar^2}{2m}\frac{\partial^2}{\partial z^2}+\frac{m\Omega_z^2 z^2}{2}+ g \|\psi_{1}\|^2 + g_{12} \|\psi_{2}\|^2 \right)\psi_{1} \\ &i\hbar \frac{\partial}{\partial t} \psi_2 = \left(\frac{-\hbar^2}{2m}\frac{\partial^2}{\partial z^2}+\frac{m\Omega_z^2 z^2}{2}+ g \|\psi_{2}\|^2 + g_{12} \|\psi_{1}\|^2 \right)\psi_{2}\, , \end{aligned} \label{eq:ebmr_dim}\end{aligned}$$ where $\Omega_z$ is the angular frequency of the axial harmonic trapping, $g=2\hbar \Omega_\perp a$ is the reduced 1D strength of the interaction between particles of the same condensate, which is proportional to the energy of the transverse trapping $\hbar \Omega_\perp$ and to the scattering length $a$, and $g_{12}=2\hbar \Omega_\perp a_{12}$ is the strength of the density coupling between condensates, proportional to the scattering length between particles of different condensates $a_{12}$. A stationary soliton solution to Eq. (\[eq:ebmr\_dim\]) $\psi_i(z,t)$, with $i=1,2$, can be written as the product of a time-independent (and real function) background component $\phi_i$ and the soliton excitation $v_i$, so that $$\begin{aligned} \psi_i(z,t)=\phi_i(z) v_i(z,t) e^{-i \mu_i t/\hbar}, \label{DS}\end{aligned}$$ where $\mu_i$ is the chemical potential. It is important to remark that both $\psi_i$ and $\phi_i$ satisfy the coupled GP equations for the same chemical potential: $$\begin{aligned} &\mu_i \phi_i = \left(-\frac{1}{2}\frac{\partial^2}{\partial z^2}+\frac{1}{2}\Omega^2 z^2+ g \phi_{i}^2 + g_{12} \phi_{j}^2 \right)\phi_{i} \label{eq:ebmr20} \\ &i \frac{\partial}{\partial t} \psi_i = \left(-\frac{1}{2}\frac{\partial^2}{\partial z^2}+\frac{1}{2}\Omega^2 z^2+ g \|\psi_{i}\|^2 + g_{12} \|\psi_{j}\|^2 \right)\psi_{i} \,, \label{eq:ebmr2}\end{aligned}$$ where $\Omega=\Omega_z/\Omega_\perp$ is the trap aspect ratio, and we have written the equations in dimensionless form by using transverse trap units, $a_\perp=\sqrt{\hbar/m\Omega_\perp}$ as unit length and $t_\perp=\Omega_\perp^{-1}$ as unit time, and for the sake of a simple notation have kept the same letters for the dimensionless couplings ${g}$ and ${g}_{12}$. So, from Eqs. (\[eq:ebmr20\])-(\[eq:ebmr2\]) we can obtain corresponding equations for the soliton wave functions $v_i(z,t)$ $$\begin{aligned} &\left(i \frac{\partial}{\partial t} + \frac{1}{2}\frac{\partial^2 }{\partial z^2}- g\phi^2_{i} \left(\|v_{i}\|^2-1\right) - g_{12}\phi^2_j\left(\|v_{j}\|^2-1\right)\right)v_i= \nonumber\\ &\hspace{.5cm} =-\frac{\partial v_i}{\partial z}\,\frac{\partial \ln \phi_i}{\partial z}\,, \label{eq:v}\end{aligned}$$ where the external trap does not appear explicitly, although its information is encoded in the background wavefunctions $\phi_i$. Untrapped case -------------- Here we set $\Omega=0$, hence the ground state solutions to Eq. (\[eq:ebmr20\]) are the constant density $\|\phi_i\|^2=n_i$ states $$n_i=\frac{g\mu_i-g_{12}\mu_j}{g^2-g^2_{12}}. \label{eq:n_i}$$ The equations of motion for soliton states Eq. (\[eq:v\]) can be written as $$\begin{aligned} \left(i \frac{\partial }{\partial t} + \frac{1}{2}\frac{\partial^2 }{\partial z^2} - g\,n_i\, \left(\|v_{i}\|^2-1\right) - g_{12}\,n_j\, \left(\|v_{j}\|^2-1\right)\right)v_i=0. \label{eq:v2}\end{aligned}$$ As we said before, we are going to deal with the coupling between condensates in a perturbative way. Then, assuming that $g_{12}<<g$, the background densities Eq. (\[eq:n\_i\]) become $n_i\approx \mu_i/g-(g_{12}/g)\,\mu_j/g$, and neglecting terms proportional to $g_{12}^2$ in Eq. (\[eq:v2\]) we get $$\begin{aligned} \begin{aligned} &i \frac{\partial v_i}{\partial t} + \frac{1}{2}\frac{\partial^2 v_i}{\partial z^2}- \mu_i \left(\|v_{i}\|^2-1\right)v_i = \\ &=- \frac{g_{12}}{g}\mu_j \left(\|v_{i}\|^2-\|v_{j}\|^2\right)v_i\equiv \frac{g_{12}}{g}\mu_j P(v_i,v_j). \end{aligned} \label{eq:v3}\end{aligned}$$ On the left hand side of this equality we have the well known non-linear Schrödinger equation for the wave function $v_i$, whereas on the right hand side we get a perturbative term in the parameter $g_{12}/g$. In order to obtain analytical expressions for the interactions between solitons, from now on we will focus on the case of equal backgrounds $\phi_1=\phi_2=\phi$, and then $\mu_1=\mu_2=\mu=(g+g_{12})n$. So that, by dividing the whole equation by $\mu$, we set new units for space $\hbar/\sqrt{m\mu}$ and time $\hbar/\mu$. In these units we introduce the general dark soliton solution $$\begin{aligned} \begin{aligned} &v_i(z,t)=\cos\varphi_i \tanh \zeta_i + i \sin\varphi_i, \end{aligned} \label{soliton}\end{aligned}$$ where $\zeta_i=\cos\varphi_i(z-t\sin\varphi_i )$, and $\varphi_i$ parametrizes the soliton darkness, that is $\|v_i\|^2=(1-\cos^2\varphi_i \, \mbox{sech}^2 \, \zeta_i)$, and the soliton velocity $\dot z_0=\sin \varphi_i$, that is the time derivative of the soliton position $z_0$, taken at minimum density. In the absence of perturbation, $P(v_i,v_j)=0$, the solution Eq. (\[soliton\]) describes overlapping solitons moving at constant velocity, and hence $\ddot z_0=\dot\varphi_i=0$. However, the effect of the perturbation $P(v_i,v_j)$ leads to the adiabatic time evolution of the soliton parameters, $(\varphi, z_0)\rightarrow (\varphi(t), z_0(t))$. This evolution can be described by the equation [@Kivshar1994] $$\begin{aligned} \dot\varphi_i=\frac{g_{12}/g}{2 \cos^2 \varphi_i \sin \varphi_i} \Re\left[\int_{-\infty}^{\infty} P(v_i,v_j)\frac{\partial v_i^}{\partial t} dz\right]. \label{pert}\end{aligned}$$ Assuming solitons with high density depletions $\varphi_i<<1$ and separated by a relative distance $2z_0$, such that $\zeta_1=\cos \varphi_1 \left(z-z_0\right)$ and $\zeta_2=\cos \varphi_2 \left(z+z_0\right)$, Eq. (\[pert\]) gives $$\begin{aligned} \dot\varphi_1=-\frac{g_{12}}{g}f(z_0)\equiv \, F_{\rm eff}^{(i)}(z_0), \label{eq:phi}\end{aligned}$$ where $f(z_0)=8\sinh(2z_0)e^{6z_0}(3-3e^{8z_0}+(1+4e^{4z_0}+e^{8z_0})4z_0)/(e^{4z_0} -1)^5$. As can be seen, for $z_0=0$, that is for overlapped solitons, $\dot \varphi_1=0$ and the solitons evolve without relative motion. Since, in this limit ($\varphi<<1$) $\dot z_0\approx \varphi$, Eq. (\[eq:phi\]) is also an equation for $\ddot z_0$, with the right hand side playing the role of a classical force $F_{\rm eff}^{(i)}(z_0)$ derived from an effective potential $U_{\rm eff}^{(i)}(z_0)$ : $$\begin{aligned} \ddot z_0=-\frac{d U_{\rm eff}^{(i)}(z_0)}{dz_0}. \label{eq:rel}\end{aligned}$$ This potential accounts for the interactions between the solitons and is the source of their relative motion. After integration of the force Eq. (\[eq:phi\]) over $z_0$, we get $$\begin{aligned} U_{\rm eff}^{(i)}(z_0)=\frac{g_{12}}{g}\,\frac{2(1+e^{4z_0})(1-2z_0)-4} { e^{-4z_0}\,(e^{4z_0}-1)^3}. \label{eq:int_pot}\end{aligned}$$ As it has been anticipated, it presents a minimum at $z_0=0$, allowing for the existence of a bound state made of overlapped solitons, and tends exponentially to zero for finite values of the inter-soliton distance $2z_0$ (see Fig. \[fig:pert\]). For small separation $z_0\rightarrow 0$, the Taylor expansion reads $$\begin{aligned} U_{\rm eff}^{(i)}(z_0)\sim \frac{g_{12}}{g}\left(-\frac{1}{6}+\frac{4}{15} z^2_0 \right) +O[z^4_0]\,, $$ ![Top panel: Effective force experienced by a dark soliton forming a dark-dark soliton system for $g_{12}/g=0.01$ in the absence of trapping. Bottom panel: Soliton interaction potential Eq. (\[eq:int\_pot\]) from which the above effective force is derived. The solitons oscillate in a relative motion around the minimum at $z_0=0$.[]{data-label="fig:pert"}](Pot_g12_pos_a_001b){width="\columnwidth"} and the equation of relative motion Eq. (\[eq:rel\]) simplifies to that of a harmonic oscillator, $\ddot z_0+w_r^2 \,z_0=0$, with angular frequency (for generic chemical potential $\mu$) $$w_r=\mu\sqrt{\frac{8}{15}\frac{g_{12}}{g}}. \label{eq:freq_rel}$$ This is the frequency of small oscillations of the two solitons around their center of mass, moving at constant velocity $v_{CM}=(\sin \varphi_1+\sin \varphi_2)/2$. It is also interesting to note that this model predicts an instability (imaginary frequency) when $g_{12}<0$, due to the fact that the effective potential $U_{\rm eff}^{(i)}$ presents a maximum for attractive interparticle interactions between particles of different condensates, and as a consequence, it prevents the existence of a bound solitonic state in this case. Trapped case ------------ In the presence of axial harmonic trapping $V(z)=\frac{1}{2}\Omega^2 z^2$, we rely on the Thomas-Fermi (TF) approximation to study the strong (interparticle) interacting regime ($\mu_i>>\Omega$) [@Pitaevskii2003]. There the inhomogeneous ground state densities, $n_i(z)=\|\phi_i\|^2$, are given by $$\begin{aligned} n_i(z)=\frac{g\mu_i(z)-g_{12}\mu_j(z)}{g^2-g^2_{12}}, \label{eq:TF}\end{aligned}$$ where $\mu_i(z)=\mu_i-V(z)$ are the local chemical potentials. In the limit of $g_{12}<<g$, we get $$\begin{aligned} g\,n_i(z)={\mu_i(z)}-\frac{g_{12}}{g}{\mu_j(z)},\end{aligned}$$ and, as a result, the equations of motion for soliton solutions Eq. (\[eq:v\]) become $$\begin{aligned} &i \frac{\partial v_i}{\partial t} + \frac{1}{2}\frac{\partial^2 v_i}{\partial z^2}- \mu_i(z) \left(\|v_{i}\|^2-1\right)v_i = \nonumber \\ &=- \frac{g_{12}}{g}\mu_j(z) \left(\|v_{i}\|^2-\|v_{j}\|^2\right)v_i -\frac{\partial v_i}{\partial z}\,\frac{d\ln \phi_i}{dz} .\end{aligned}$$ The local chemical potentials and the last term in the right hand side of this equation are the main differences with respect to the untrapped case Eq. (\[eq:v3\]). In spite of these differences, and due to the fact that the background densities change slowly inside the TF regime ($dV/dz\rightarrow 0$), an analogue perturbative approach can still be followed, and the governing equation for the dark soliton $v_i$ is $$\begin{aligned} i \frac{\partial v_i}{\partial t} + \frac{1}{2}\frac{\partial^2 v_i}{\partial z^2}-\mu_i(z) v_i \left(\|v_{i}\|^2-1\right)=P(v_i,v_j;V(z)), \label{eq:v5}\end{aligned}$$ where the perturbation $P(v_i,v_j;V)$ is given by $$\begin{aligned} &P(v_i,v_j;V)\approx - \frac{g_{12}}{g} \mu_j(z) \left(\|v_{i}\|^2-\|v_{j}\|^2\right)v_i+ \nonumber\\ &\hspace{.5cm} +\frac{1}{2\mu(z)}\frac{dV}{dz}\frac{\partial v_i}{\partial z},\end{aligned}$$ which, along with the term associated to the interaction between dark solitons, contains a new term introduced by the external trap [@Frantzeskakis2010]. Again, we focus on the symmetric case $\mu_1=\mu_2=\mu$, where the local chemical potential reads $\mu(z)=\mu-V(z)=(g+g_{12})\,n(z)$. In the particular case of motion around the center of the harmonic potential $\mu(z)\rightarrow \mu$, and by following the same procedure as in the untrapped case (with $\hbar/\sqrt{m\mu}$ and $\hbar/\mu$ as space and time units respectively) we can obtain a particle-like evolution for the soliton position $z_0$ ![(Color online) Forces (top panel) and total, single-well potential (bottom panel) experienced by each dark soliton forming a dark-dark soliton system with $\mu=1$ and $g_{12}>0$ in the presence of a harmonic trap with $\Omega=0.1$. The dashed (red) line represents the force caused by the trap, whereas the continuous lines represent the effective force due to the interaction with the other dark soliton.[]{data-label="fig:potg12posa001"}](Pot_g12_pos_a_001){width="\linewidth"} $$\begin{aligned} \ddot{z}_0=-\frac{\Omega^2}{2 \mu^2} z_0- F_{\rm eff}^{(i)} , \label{eq:freq_trap}\end{aligned}$$ where the right hand side is the total effective force acting on the soliton $F_{\rm eff}=-dU_{\rm eff}/d\,z_0$, and embraces the superposition of the action of the external trap and the soliton interaction. For small-amplitude oscillations $z_0\approx 0$, we get $$\begin{aligned} \ddot{z}_0=-\left(\frac{\Omega^2}{2\mu^2}+\frac{8}{15}\frac{g_{12}}{g} \right)z_0+\frac{64}{63}\frac{g_{12}}{g}z_0^3+O(z_0^5), \label{eq:cl2}\end{aligned}$$ where the balance between the two harmonic force terms (inside the parenthesis) leads to different scenarios depending on the sign of the coupling interaction $g_{12}$. ![(Color online) Same as Fig. \[fig:potg12posa001\] for $g_{12}<0$. Three cases are depicted for different $g_{12}/g$ values, below (blue), at (orange) and above (green) the threshold $g_c$ indicating the change in the shape of the total effective potential from single well $g_{12}/g< c_c$ to double well $g_{12}/g>g_c$ (see text). []{data-label="fig:potg12nega001"}](Pot_g12_neg_a_001){width="\columnwidth"} ### Repulsive coupling: ${g_{12}>0}$ Figure \[fig:potg12posa001\] shows the effective force and effective potential defined in Eq. (\[eq:freq\_trap\]) for varying values of $g_{12}/g$ at a given trapping $\Omega=0.1$. As can be seen, the different parameters do not produce qualitative changes in the potential, which presents a single minimum capable to bound the coupled solitons. As in the untrapped case, our theory predicts that the dark solitons will perform small-amplitude oscillations around $z_0=0$ with angular frequency (in transverse oscillator units) $$w=\left(\frac{\Omega^2}{2}+\frac{8}{15}\frac{g_{12}}{g} \mu^2\right)^\frac { 1 } { 2 }. \label{eq:fre}$$ As we will see later, our numerical results demonstrate that the minimum supports a bound state that exists and is stable for small $g_{12}/g$. However, for increasing values of this parameter both stable and unstable cases can be found depending on the particular valued of the chemical potential of the system. ### Attractive coupling: ${g_{12}<0}$ The phenomenology is richer for attractive interactions between particles of different condensates. In contrast to the untrapped case, the presence of the harmonic potential allows for the generation of local minima in the total effective potential that can support stationary states made of two solitons. As shown in Fig. \[fig:potg12nega001\], by increasing the parameter $\|g_{12}/g\|$, the total effective potential modifies from a single well (an also a unique fixed point $z_0^=0$ in the equation of motion) to a double well potential (with three fixed points). Specifically, the system shows a pitchfork bifurcation at $g_c=\|g_{12}/g\|={15}(\Omega/\mu)^2/16$, hence for $\|g_{12}/g\|>g_c$ the fixed point at $z_0^=0$ loses its stability and two new off-center, stable fixed points appear. From Eq. (\[eq:cl2\]) we get their position at $$z_{0_\pm}^{} = \pm \frac{1}{\sqrt{\mu}}\sqrt{\frac{21}{40}- \frac{63}{128} \left\|\frac{g}{g_{12}}\right\| \left(\frac{\Omega}{\mu}\right)^2 }. \label{eq:fixed_double}$$ ![(Color online) Stationary state made of two solitons situated at the fixed points $z_{0_\pm}^$ of the effective potential (see text) for $\mu=10\,\Omega$ and $g_{12}/g=-0.1$.[]{data-label="fig:separated"}](separated_solitons){width="\columnwidth"} As can be noted, the separation between minima increases with $\mu/\, \Omega$ for given $g_{12}/g$, and saturates at a distance of $2z_0^=1.44$. Fig. \[fig:separated\] shows an example of a stationary state with two separated solitons occupying the two minima of the effective potential at $\mu=10 \, \Omega$ and $g_{12}/g=-0.1<g_c$. For small distances $\epsilon(t)$ around the fixed points $z_{0_\pm}^$, from the substitution of $z_0(t)=z_0^+\epsilon(t)$ in Eq. (\[eq:cl2\]), the solitons oscillate according to $$\ddot{\epsilon}=-\left( w_1(z_{0_\pm}^)^2 +w^2 \right) \epsilon \equiv -w_{\rm eff}^2 \epsilon , \label{eq:epsilon}$$ up to linear terms in the perturbation $\epsilon(t)$, where $w$ is the angular frequency given by , and $ w_1(z_0)^2=-2 f(z_0) \left( \mbox{sech}(z_0)- {5}/{\sinh(2z_0)} + 3\right)-$ $\left( \left(1+4e^{4z_0} +e^{ 8z_0}\right)+ \left(-12e^{8z_0}+\left(16e^{4z_0}+8e^{8z_0}\right) z_0 \right) \right)$ $ 64e^{6z_0} / \left(e^{4z_0}-1\right)^5 $. ![(Color online) Bogoliubov modes in 1D rings of different size (untrapped case). The upper panel depicts the real part of the frequencies and the lower panel the imaginary parts. The solid and dashed lines correspond to Eqs. (\[eq:freq\_rel\]) and (\[eq:spin\]), respectively, the crossing of which indicates the appearance of instabilities.[]{data-label="fig:bog_notrap"}](modes_mu05_Z10.pdf){width="\columnwidth"} Numerical results ================= In what follows, in order to test our analytical predictions on the dark-dark-soliton dynamics, we first numerically solve the GP equation to obtain these stationary states for varying chemical potentials and interactions strengths. Afterwards, the soliton stability is monitored both in the nonlinear regime (by simulating the real time evolution with the GP equation) and by linear analysis around the equilibrium states. The latter is performed by solving the Bogoliubov equations for the linear excitations $[u(z,t),v(z,t)]$ around the dark-dark soliton states $\psi(z,t)=\exp(-i \mu t)[\psi(z) +\sum_\omega(u \,e^{-i\omega t}+ v^ e^{i\omega t})]$. These equations are: $$\begin{aligned} B \left(\begin{array}{c} u_1 \\ v_1 \\ u_2 \\ v_2 \end{array}\right) = \omega \left(\begin{array}{c} u_1 \\ v_1 \\ u_2 \\ v_2 \end{array}\right), \label{eq:BdG}\end{aligned}$$ where: $$\begin{aligned} B= \left(\begin{array}{cccc} h_1 & g\psi_1^2 & g_{12}\psi_ 1 \psi_2^* & g_{12}\psi_1 \psi_2 \\ -g\psi_1^{2} & -h_1 & -g_{12}\psi_1^ \psi_2^* & -g_{12}\psi_1^* \psi_2 \\ g_{12}\psi_1^* \psi_2 & g_{12}\psi_1 \psi_2 & h_2 & g\psi_2^2 \\ -g_{12}\psi_1^* \psi_2^* & - g_{12}\psi_1 \psi_2^* &- g\psi_2^{2} & -h_2 \end{array}\right).\end{aligned}$$ Here $h_i=-\frac{1}{2} \frac{\partial^2}{\partial z^2} + \frac{1}{2}\Omega^2z^2 + 2g \|\psi_i\|^2 + g_{12} \|\psi_{j\neq i}\|^2 - \mu_i$. They can be seen as an eigenvalue problem with a non-trivial solution given by ${\rm det}\|B- \omega I\|=0$. Apart from the perturbative case $g_{12}<<g$, we also explore numerically the stability of overlapped solitons in the whole miscible regime $g_{12}<g$, and show that the finite size of the system determines the stability properties. Untrapped case -------------- ![(Color online) Spatio-temporal evolution of the density of two untrapped dark-dark solitons with $\mu=1$, $g_{12}=0.015$ and $g=1$ obtained from the numerical solution of the GP Eq. . The top and down panel correspond to component 1 and 2 respectively. The continuous (colour) lines represent the evolution of the dark solitons positions given by Eq. (\[eq:freq\_rel\]).[]{data-label="fig:osci003"}](untrap_osci_0015){width="\columnwidth"} ![Same as Fig. \[fig:osci003\] but with $g_{12}=0.1$. []{data-label="fig:osci01"}](untrap_osci_01){width="\columnwidth"} ![(Color online) Soliton relative motion frequency as a function of the interparticle strength ratio in the absence of external trap. The continuous (blue) line represents our theoretical prediction Eq. (\[eq:freq\_rel\]) and the (orange) dots are values extracted by means of Fourier analysis from the numerical solution of the time dependent GP equation.[]{data-label="fig:fourier"}](Fourier2){width="\columnwidth"} The excitation frequencies of the stationary states can be readily extracted by solving the Bogoliubov equations Eq. (\[eq:BdG\]). In the Appendix we analytically show that there are two zero modes of excitation at $g_{12}/g=0$, and $1$, and hence that the system of overlapped dark solitons is expected to be unstable in between. However such analysis assume an infinite system, and the situation is quite different in systems of finite size, where ranges of dynamical stability can be found. Fig. \[fig:bog\_notrap\] shows two examples of this phenomenon, where the Bogoliubov modes are computed for overlapped solitons with the same chemical potential $\mu=0.5$ in 1D rings of different sizes $L=20,\; 40$. The instability emerge from a Hopf bifurcation, which occurs due to the collision of two excitation modes (see the discussion about this collision in the next section): one mode associated to the oscillations around the minimum of the effective potential of soliton interactions, given by expression (\[eq:freq\_rel\]) and represented in Fig. \[fig:bog\_notrap\] by the solid red curve, and the background spin density mode of lowest energy (represented by the dashed curve for the longer ring), which is given by the analytical expression (in full units) [@Abad2013]: $$\hbar\omega=\sqrt{\frac{\hbar^2 k^2}{2m} \left( \frac{\hbar^2 k^2}{2m} + 2(g-g_{12})n\right)}. \label{eq:spin}$$ The small disagreement between the crossing of these analytical curves and the beginning of instabilities in the numerical results arises from the curvature of the modes near the bifurcation point, and decreases for longer rings. This first instability triggers new collisions between modes and, as a consequence, more instability regions. The longer the ring the higher the number of instability regions in the system, approaching the prediction for the infinite case. It is worth remarking that the bound state mode predicted by Eq. (\[eq:freq\_rel\]), in excellent agreement with the numerics, does not change with the size of the ring. To analyze the dynamics of dark-dark solitons in the bound state allowed by the repulsive interparticle interactions $g_{12}>0$, we excite the relative motion of the solitons by imposing the initial ansatz $$\begin{aligned} \begin{aligned} &\psi_1=\sqrt{n}\tanh\left(z-z_0\right) \\ &\psi_2=\sqrt{n}\tanh\left(z+z_0\right), \end{aligned}\end{aligned}$$ and we fix $\mu=1$. Figures \[fig:osci003\]–\[fig:osci01\] show the comparison between the subsequent motion of the solitons from the numerical solution of GP Eq. , and the analytical prediction by Eq. (\[eq:freq\_rel\]) fitted to $y(t)=z_0\cos{wt}$. As can be seen, it provides a reasonable good estimate for small $g_{12}$ (Fig. \[fig:osci003\]) but fails for larger $g_{12}$ (Fig. \[fig:osci01\]) or also long times. In order to obtain a more quantitative comparison we have run different real time evolutions for varying $g_{12}$ and equal chemical potential $\mu=1$. By tracking the position of the solitons (at minimum density), we have computed their characteristic frequency from a Fourier analysis in time. The numerical results are presented in Fig. \[fig:fourier\], and show a very good agreement with our analytical prediction for small values of $g_{12}/g$. Trapped case ------------ ### Attractive interaction between condensates: ${g_{12}<0}$ ![(Color online) Frequency of the dark soliton relative motion in a harmonic trap with $\Omega=0.1$. The real (top panel) and imaginary (bottom) parts of this frequency are shown according to our analytical prediction Eq. (continuous orange line) and the numerical solution of the excitation spectrum from Eq. for the out-of-phase anomalous mode (dash-dotted lines).[]{data-label="fig:freqg12nega00175"}](Freq_g12_neg_a_00175.pdf){width="\columnwidth"} As anticipated, in this case the configuration of overlapped solitons at $z_0=0$ is unstable for $\|g_{12}/g\|>g_c$. This instability can also be detected by the appearance of an imaginary frequency in the excitation spectrum. Fig. \[fig:freqg12nega00175\] shows our results for the linear excitations of such a stationary state with $\mu=10\, \Omega$ from the solution of the Bogoliubov equations (dash-dotted lines). The frequency of the out-of-phase anomalous mode (see a discussion of this mode in next section) takes real values for $\|g_{12}/g\|<g_c$ and pure imaginary for $\|g_{12}/g\|>g_c$. This critical point $g_c$ is associated to the change of the total effective potential, from single-well to double-well. Along with the change of the total effective potential two new stable fixed points appear in the system. In Fig. \[fig:eqpos2\] we compare these points (red dots), obtained by extracting the mean position of the soliton oscillations around the equilibrium positions, with our analytical approach Eq. (orange line).The latter fails for increasing values of $\|g/g_{12}\|$. However the direct numerical computation of Eq. (blue line) provides a very good agreement for the regime of interest, and shows that the distance between equilibrium points increases with $\|g_{12}/g\|$ instead of being saturated at $2 z_0^=1.44$. ![(Color online) Fixed points of the effective double well potential given by Eq. (continuous blue line) as a function of $g_{12}/g$. The continuous (orange) line represents the analytical result and the (red) dots are the mean position of the soliton oscillations extracted from the numerical solution of GP equation. []{data-label="fig:eqpos2"}](eq_pos2){width="1.\columnwidth"} ![Spatio-temporal evolution of a dark-dark-soliton density obtained from the numerical solution of the GP Eq. for $g_{12}/g=-0.046$ and harmonic trapping $\Omega=0.1$. Each panel corresponds to a different condensate. The dashed (red) lines represent the position of the equilibrium points given by . []{data-label="fig:doublewell0046"}](insta_outcenter_0046){width="\columnwidth"} We have numerically solved the GP Eq. (\[eq:ebmr\_dim\]) for the real-time evolution of dark-dark solitons with values of $g_{12}$ below and above the bifurcation point $g_c$ for the change of stability. First, we have computed a case (see Fig. \[fig:doublewell0046\]) with overlapped solitons situated at $z_0=0$ and $g_{12}/g=-0.046<g_c$, hence unstable according to the linear prediction. The initial stationary state has been perturbed with white-noise of $1\%$ amplitude. As expected the system is unstable, and eventually the solitons separate by moving towards the minima of the effective double well potential Eq. (\[eq:fixed\_double\]). In this case, each dark soliton has enough energy to pass through the energy barrier created at the trap center $z_0=0$, so that collisions between them are observed to cause a shift in their trajectories. If instead the two dark solitons are initially situated at different locations, close to the positions of the fixed points Eq. (\[eq:fixed\_double\]) (see Fig. \[fig:osci\_outcenter\_00125\]), the solitons oscillate symmetrically around such points. Their time evolution can be fitted by $x(t)=z_0^+\left(-z_0^+z_0 \right)\cos(w_{\rm eff}t)$, where $w_{\rm eff}$ is given by , and provides a good approximation to the real time evolution obtained from the GP equation. ![Same as Fig. \[fig:doublewell0046\] with $g_{12}/g=-0.0125$. In this spatio-temporal evolution each dark soliton is situated off center ($z_0=\pm0.3$, near the equilibrium points given by ). The continuous line is fitted according to our analytical results (see text).[]{data-label="fig:osci_outcenter_00125"}](osci_outcenter_00125198){width="\columnwidth"} ### Repulsive interaction between condensates: ${g_{12}>0}$ ![(Color online) Numerical results from the Bogoliubov spectrum of a dark-dark soliton with $\mu=10 \,\Omega$ in a trap. The real (top panel) and imaginary (bottom panel) parts of the excitation frequencies are shown against the interaction strength ratio $g_{12}/g$. The continuous (orange) line represents the analytical values from Eq. .[]{data-label="fig:spectrumrealimag"}](spectrum_trap01_mu1.pdf){width="1.0\columnwidth"} Figure \[fig:spectrumrealimag\] represents the excitation spectrum of overlapped dark-dark solitons at $z_0^=0$, in the range $g_{12}/g\in[0,1]$, for $\mu=10\, \Omega$, within the Thomas Fermi regime of the axial harmonic oscillator. At $g_{12}=0$, corresponding to uncoupled solitons, the linear modes have double degeneracy, and the lowest energy excitations are the anomalous mode, with frequency $\omega_0=\Omega/\sqrt{2}$, and the hydro-dynamical excitations $\omega_n=\sqrt{{n(n+1)}/{2}}\,\Omega$, with $n=1,2,...$ [@Stringari1996; @Busch2000; @Kevrekidis2015; @Frantzeskakis2010]. The degeneracy is broken for non null $g_{12}$ and gives rise to two branches of in-phase and out-of-phase modes. In the hydro-dynamical case, the in-phase modes are associated to excitations in the total density of the background, and remain constant for varying $g_{12}$. On the other hand, the out-of-phase or spin modes account for variations in the difference of the background densities [@Abad2013], and decrease their energy for increasing values of $g_{12}$. ![Real time evolution of three dark-dark solitons with varying coupling out of and inside the first instability region of Fig. \[fig:spectrumrealimag\], showing corresponding stable and unstable dynamics. []{data-label="fig:instab_trap"}](abc_evolution.pdf "fig:"){width="\columnwidth"}\ ![Real time evolution of three dark-dark solitons with varying coupling out of and inside the first instability region of Fig. \[fig:spectrumrealimag\], showing corresponding stable and unstable dynamics. []{data-label="fig:instab_trap"}](imaginary_freq_mu_1.pdf "fig:"){width="0.9\columnwidth"} The anomalous modes are characterized by the negative value of the quantity $norm \times energy$ [@MacKay1987], and in scalar condensates their frequency coincide with the oscillation frequency of the solitons in the trap [@Busch2000]. There are also two different anomalous modes for non null $g_{12}$, an in-phase one with constant energy, and an out-of-phase mode whose energy increases with $g_{12}$. The in-phase anomalous mode is associated with the small amplitude, abreast oscillations of the solitons in the trap, hence it is the same as in decoupled condensates: $\omega_0=\Omega/\sqrt{2}$. However, the out-of-phase anomalous mode is associated with the relative motion of the solitons. This mode, given by the analytical expression Eq. , is depicted (orange line) in the upper panel of Fig. \[fig:spectrumrealimag\], in good agreement with the numerical results for small values of $g_{12}/g$. The lower panel of Fig. \[fig:spectrumrealimag\] present our numerical results for the imaginary part of the spectrum. As previously commented in the untrapped case, these instabilities are characterized by the collision of the out-of-phase anomalous mode with the spin mode associated with the background. These collisions produce Hamiltonian-Hopf bifurcations where a complex frequency quartet appears in the excitation spectrum. It is interesting to see that the out-of-phase anomalous mode only collides with odd spin modes. Regions of stability and instability alternates up to a value of the coupling with $g_{12}/g$ close to 1, but interestingly inside the immiscible regime. The higher the chemical potential, the closer is this value to $g_{12}/g=1$, according with the Bogoliubov analysis for the untrapped case. Such a value can be well approximated within the Thomas-Fermi regime by Eq. (\[eq:spin\]) evaluated at maximum density $n=\mu/(g+g_{12})$. Again the crossing of this analytical frequency with the function for small oscillations Eq. (\[eq:fre\]) provides a good estimate for the beginning of the first instability region in the range $g_{12}/g\in[0,1]$. It is worth noticing that such instability is not captured by Eq. alone, due to the fact that for its derivation the soliton motion was assumed to be decoupled from the background. Figure \[fig:instab\_trap\] shows examples of stable and unstable dark-dark solitons near the first instability region of Fig. \[fig:spectrumrealimag\]. The evolution of the nonlinear systems develops according the linear stability analysis and demonstrate the existence of dynamically stable coupled solitons, cases (a) and (c), that could be experimentally realized. Conclusions =========== The dynamics of dark-dark soliton states in two density coupled BECs has been studied within GP theory. By performing a perturbation analysis in the parameter $g_{12}/g$ for solitons with equal chemical potential, we have derived analytical expressions describing their relative motion both in harmonic traps and untrapped systems. Contrary to the case of solitons in scalar condensates, our theoretical model predicts that the interaction between dark solitons excited in different condensates is attractive (repulsive) for repulsive (attractive) interparticle interactions $g_{12}$. In harmonically trapped systems, the scenario is specially interesting for negative $g_{12}$, where the effective potential felt by the solitons modifies its shape as a function of $g_{12}$ (through a pitchfork bifurcation) from a single-well potential to a double-well potential, then allowing for stationary states made of solitons located at different positions. The theoretical analytical predictions have been shown to be in good agreement with the numerical solutions of the Gross Pitaevskii equation for the real time evolution, and with the Bogoliubov equations for the linear excitations of the dark-dark solitons. In particular, we have demonstrated that the resonance of two out-of-phase modes, the anomalous one giving the frequency of the relative motion between solitons, and the lowest energy mode associated to the spin density excitation of the background, give rise to instabilities (Hopf bifurcations) that produce the decay of dark-dark solitons. This fact translate in finite systems, either harmonically trapped condensates or ring geometries, into alternating regions of dynamical stability and instability. The existence of dynamically stable dark-dark solitons open up the way for their experimental realization. The current availability of Feshbach resonances for tuning both interaction parameters $g$ and $g_{12}$ allows to choose a stable fringe in the spectrum. Also in this regard, as a natural extension of this work, it would be interesting to explore the stability of equivalent states (soliton-soliton or vortex-vortex states) in multidimensional systems. The authors acknowledge financial support by grants 2014SGR-401 from Generalitat de Catalunya and FIS2014-54672-P from the MINECO (Spain). [99]{} L. P. Pitaevskii and S. Stringari, [Bose-Einstein Condensation]{}, (Oxford University Press, Oxford, 2003). C. J. Pethick and H. Smith, [Bose-Einstein Condensation in Dilute Gases]{}, (Cambridge University Press, Cambridge, 2008). P. G. Kevrekidis, D. J. Frantzeskakis, and R. Carretero- González, [The Defocusing Nonlinear Schrödinger Equation: From Dark Solitons to Vortices and Vortex Rings]{}, (SIAM, Philadelphia, 2015). P. G. Kevrekidis, D. J. Frantzeskakis, and R. Carretero González (eds.), [Emergent Nonlinear Phenomena in Bose-Einstein Condensates. Theory and Experiment]{}, (Springer-Verlag, Berlin, 2008). D. J. Frantzeskakis, J. Phys. A [43]{}, 213001 (2010). K. E. Strecker, G. B. Partridge, A. G. Truscott, and R. G. Hulet, New J. Phys. [5]{}, 731 (2003). A. L. Fetter and A. A. Svidzinsky, J. Phys.: Cond. Matt. [13]{}, R135 (2001). S. Burger, K. Bongs, S. Dettmer, W. Ertmer, K. Sengstock, A. Sanpera, G. V. Shlyapnikov, and M. Lewenstein, Phys. Rev. Lett. [83]{}, 5198 (1999). C. Becker, S. Stellmer, P. Soltan-Panahi, S. Dörscher, M. Baumert, E.-M. Richter, J. Kronjäger, K. Bongs, and K. Sengstock, Nat. Phys. [4]{}, 496 (2008). A. Weller, J. P. Ronzheimer, C. Gross, J. Esteve, M. K. Oberthaler, G. Theocharis, and P. G. Kevrekidis, Phys. Rev. Lett. [101]{}, 130401 (2008). G. Lamporesi, S. Donadello, S. Serafini, F. Dalfovo, and G. Ferrari, Nat. Phys. [9]{}, 656 (2013). B. P. Anderson, P. C. Haljan, C. A. Regal, D. L. Feder, L. A. Collins, C. W. Clark, and E. A. Cornell, Phys. Rev. Lett. [86]{}, 2926 (1999). William P. Reinhardt and Charles W. Clark, J. Phys. B: At. Mol. Opt. Phys. [30]{}, L785 (1997). A. E. Muryshev, H. B. van Linden van den Heuvell, and G. V. Shlyapnikov, Phys. Rev. A [60]{}, R2665(R) (1999). Th. Busch and J. R. Anglin, Phys. Rev. Lett. [84]{}, 2298 (2000). D. E. Pelinovsky, D. J. Frantzeskakis, and P. G. Kevrekidis, Phys. Rev. E [72]{}, 016615 (2005). D. S. Hall, M. R. Matthews, J. R. Ensher, C. E. Wieman, and E. A. Cornell, Phys. Rev. Lett. [81]{}, 1539 (1998). C. J. Myatt, E. A. Burt, R. W. Ghrist, E. A. Cornell, and C. E. Wieman, Phys. Rev. Lett. [78]{}, 586 (1997). P. Öhberg and L. Santos, Phys. Rev. Lett. [86]{}, 2918 (2001). D. Yan, J. J. Chang, C. Hamner, M. Hoefer, P. G. Kevrekidis, P. Engels, V. Achilleos, D. J. Frantzeskakis, and J. Cuevas, J. Phys. B: At. Mol. Opt. Phys. [45]{}, 115301 (2012). M. A. Hoefer, J. J. Chang, C. Hamner, and P. Engels, Phys. Rev. A [84]{}, 041605(R) (2011). S. Middelkamp, J. J. Chang, C. Hamner, R. Carretero-González, P. G. Kevrekidis, V. Achilleos, D. J. Frantzeskakis, P. Schmelcher, and P. Engels, Phys. Lett. A [375]{}, 642 (2011). D. Yan, J. J. Chang, C. Hamner, P. G. Kevrekidis, P. Engels, V. Achilleos, D. J. Frantzeskakis, R. Carretero-González, and P. Schmelcher, Phys. Rev. A [84]{}, 053630 (2011). V. Achilleos, P. G. Kevrekidis, V. M. Rothos, and D. J. Frantzeskakis, Phys. Rev. A [84]{}, 053626 (2011). I. Danaila, M. A. Khamehchi, V. Gokhroo, P. Engels, and P. G. Kevrekidis, Phys. Rev. A [94]{}, 053617 (2016). A. M. Kamchatnov and V. S. Shchesnovich, Phys. Rev. A [70]{}, 023604 (2004). I. M. Uzunov and V. S. Gerdjikov, Phys. Rev. A [47]{}, 1582 (1993). Yu. S. Kivshar and X. Yang, Phys. Rev. E [49]{}, 1657 (1994). D. J. Frantzeskakis, G. Theocharis, F. K. Diakonos, P. Schmelcher, and Yu. S. Kivshar, Phys. Rev. A [66]{}, 053608 (2002). S. Stringari, Phys. Rev. Lett. [77]{}, 2360 (1996). M. Abad and A. Recati, Eur. Phys. J. D [67]{}, 148 (2013). R. S. MacKay and J. D. Meiss, [Hamiltonian Dynamical Systems]{}, (Hilger, Bristol, 1987). N. Rosen and Philip M. Morse, Philip M, Phys. Rev. [42]{}, 210, (1932). S. S. Shamailov and J. Brand, arXiv:1709.00403 (2017). Appendix: Bogoliubov equations for overlapping dark solitons without external trap {#appendix-bogoliubov-equations-for-overlapping-dark-solitons-without-external-trap .unnumbered} ================================================================================== In this case, the ground state of Eqs. (\[eq:ebmr\_dim\]) is the constant density solution $\psi_1(x)=\psi_2(x)=\sqrt{n}$, such that $\mu=(g+g_{12})n$. The dark soliton solutions healing to the ground state with density $n$ are given by $$\psi_{1,2}(x)=\psi_d(x)e^{-i\mu\,t/\hbar}=\sqrt{\frac{\mu}{g+g_{12}}}\, \tanh\left(\frac{x}{\xi}\right) e^{-i\mu\,t/\hbar}\,, \label{eq:dark}$$ where $\xi=\sqrt{\hbar^2/m \mu}$ is the soliton healing length. As in the scalar case, it can be seen that the soliton state depends on both, the chemical potential $\mu$ and the (sum of the) interaction strength $g+g_{12}$. We check the stability of Eq. (\[eq:dark\]) by solving the Bogoliubov equations $$\begin{aligned} H_0 \, u_1+ \psi_d^2 \left[(2g+g_{12})u_1+g v_1 +g_{12}(u_2+v_2) \right] = \omega \, u_1 \nonumber\\ -H_0 \, v_1 - \psi_d^2 \left[(2g+g_{12})v_1+g u_1 +g_{12}(u_2+v_2) \right] = \omega \, v_1 \nonumber\\ H_0 \, u_2+ \psi_d^2 \left[(2g+g_{12})u_2+g v_2 +g_{12}(u_1+v_1) \right] = \omega \, u_2 \nonumber\\ -H_0 \, v_2 - \psi_d^2 \left[(2g+g_{12})v_2+g u_2 +g_{12}(u_1+v_1) \right] = \omega \, v_2 , \label{eq:bog0}\end{aligned}$$ where $H_0 = -(\hbar^2/2m)\partial_{xx} -\mu$, and $[u_{1}(x), v_{1}(x), u_{2}(x) ,v_{2}(x)]$ are the linear modes with energy $\omega$. By adding and substracting the the two first previous equations, on the one hand, and the two last, on the other hand, we obtain new equations for the linear combinations $f_{j\pm}=u_{j}(x) \pm v_{j}(x)$ $$\begin{aligned} \left(\frac{-\hbar^2}{2m}\partial_{xx}+ \, (g+g_{12}) \, \psi_d^2-\mu\right) f_{1-}= \omega f_{1+} \nonumber\\ \left(\frac{-\hbar^2}{2m}\partial_{xx}+ \, (g+g_{12}) \, \psi_d^2-\mu\right) f_{2-}= \omega f_{2+}, \label{eq:bog1-} \\ \left( \frac{-\hbar^2}{2m}\partial_{xx} + (3g+g_{12}) \psi_d^2 \, -\mu \right) f_{1+} +2g_{12}\psi_d^2 f_{2+}=\omega f_{1-} \, \nonumber\\ \left( \frac{-\hbar^2}{2m}\partial_{xx} + (3g+g_{12}) \psi_d^2 \, -\mu \right) f_{2+} +2g_{12}\psi_d^2 f_{1+}=\omega f_{2-} \,. \label{eq:bog1+}\end{aligned}$$ The first two equations Eqs. (\[eq:bog1-\]) are already decoupled for the modes $1$ and $2$, and (for $\omega=0$) contains zero energy excitations associated to the $U(1)$ symmetry presented within each condensate, so that a global phase can be arbitrarily picked in the soliton solutions Eq. (\[eq:dark\]). However, Eqs. (\[eq:bog1+\]) are still coupling the modes $1$ and $2$. In order to decouple them, we make the symmetric $f_{s\pm}=f_{1\pm}(x) + f_{2\pm}(x)$ and antisymmetric $f_{a\pm}=f_{1\pm}(x) - f_{2\pm}(x)$ linear combinarions to get $$\begin{aligned} \left(\frac{-\hbar^2}{2m}\partial_{xx}+ \, (g+g_{12}) \, \psi_d^2-\mu\right) f_{s-}= \omega f_{s+} \nonumber\\ \left( \frac{-\hbar^2}{2m}\partial_{xx} + 3(g+g_{12}) \psi_d^2 \, -\mu \right) f_{s+} =\omega f_{s-} \,, \label{eq:bog_s}\end{aligned}$$ and $$\begin{aligned} \left(\frac{-\hbar^2}{2m}\partial_{xx}+ \, (g+g_{12}) \, \psi_d^2-\mu\right) f_{a-}= \omega f_{a+} \, \nonumber\\ \left( \frac{-\hbar^2}{2m}\partial_{xx} + (3g-g_{12}) \psi_d^2 \, -\mu \right) f_{a+} =\omega f_{a-} \,. \label{eq:bog_a}\end{aligned}$$ Equations (\[eq:bog\_s\]) for the symmetric combinations $f_{s\pm}=f_{1\pm}(x) + f_{2\pm}(x)$ are equivalent to the Bog equations of a scalar condensate in a dark soliton state for an interaction strength $g+g_{12}$. As a result they do not present unstable modes, and contain two Goldstone modes (with $\omega=0$) reflecting the mentioned $U(1)$ symmetry and the translational invariance of the system. On the other hand, the Eqs. (\[eq:bog\_a\]) for the antisymmetric combinations $f_{a\pm}=f_{1\pm}(x) - f_{2\pm}(x)$ are relevant for the unstable modes having complex frequencies $\omega$. These modes can first appear at $\omega=0$ for particular values of the system parameters $\{\mu, g, g_{12}\}$ (for a given particle species of mass $m$), indicating a bifurcation of the soliton solutions giving place to a new stationary state. However, as we have seen in the text, the instabilities can also appear from a couple of nonzero real frequencies which become complex, indicating a Hopf bifurcation with a subsequent oscillatory dynamics. Below, we analyze the first case associated to the existence of zero modes. At $\omega=0$, since the first equation (\[eq:bog\_s\]) provides the commented Goldstone modes, we only have to look for solutions to $$\begin{aligned} \left( \frac{-\hbar^2}{2m}\partial_{xx} + \mu \frac{3g-g_{12}}{g+g_{12}}\, \tanh^2\left(\frac{x}{\xi} \, \right)-\mu \right)f_{a+} =0\,,\end{aligned}$$ that in units of the healing length gives $$\begin{aligned} -\frac{1}{2}\partial_{xx} f_{a+}- (1+\delta g) \, \mbox{sech}^2(x) f_{a+} = -\delta g \, f_{a+}, \label{eq:zero}\end{aligned}$$ where $\delta g=2(g-g_{12})/({g+g_{12}})=2(g-g_{12})n/\mu$ is the relevant parameter for the emergence of instabilities. Eq. (\[eq:zero\]) is a Schrödinger equation for the asymmetric wavefunction $f_{a+}$ in the potential well $\mbox{sech}^2(x)$ with depth $(1+\delta g)$. This well has bound states with energy $-\delta g$ whenever [@Rosen; @Sophie2017] $$\sqrt{2 \,\delta g+\frac{9}{4}}-\sqrt{2\,\delta g}-\frac{1}{2}=n \, ,$$ with $n=0,1,2,\dots$ and $n \leq \sqrt{2\delta g+9/4}-1/2$. The latter condition ensures that $g_{12}<g$ and saturates with $g=g_{12}$ at the starting point of the inmiscible regime.	{ "pile_set_name": "ArXiv" }
--- author: - 'Yangyang Zhang, Dachun Yang[^1], Wen Yuan and Songbai Wang' title: 'Real-Variable Characterizations of Orlicz-Slice Hardy Spaces ' --- [In this article, the authors first introduce a class of Orlicz-slice spaces which generalize the slice spaces recently studied by P. Auscher et al. Based on these Orlicz-slice spaces, the authors introduce a new kind of Hardy type spaces, the Orlicz-slice Hardy spaces, via the radial maximal functions. This new scale of Orlicz-slice Hardy spaces contains the variant of the Orlicz-Hardy space of A. Bonami and J. Feuto as well as the Hardy-amalgam space of Z. V. de P. Ablé and J. Feuto as special cases. Their characterizations via the atom, the molecule, various maximal functions, the Poisson integral and the Littlewood-Paley functions are also obtained. As an application of these characterizations, the authors establish their finite atomic characterizations, which further induce a description of their dual spaces and a criterion on the boundedness of sublinear operators from these Orlicz-slice Hardy spaces into a quasi-Banach space. Then, applying this criterion, the authors obtain the boundedness of $\delta$-type Calderón-Zygmund operators on these Orlicz-slice Hardy spaces. All these results are new even for slice Hardy spaces and, moreover, for Hardy-amalgam spaces, the Littlewood-Paley function characterizations, the dual spaces and the boundedness of $\delta$-type Calderón-Zygmund operators on these Hardy-type spaces are also new. ]{} Introduction\[s1\] ================== The Hardy spaces $H^p(\rn)$, with $p\in (0, 1]$, are known to be one of the most important working spaces on $\rn$ in harmonic analysis and partial differential equations, which play key roles in many branches of analysis; see, for example, [@fs72; @fs82; @sto89; @em; @m94]. In particular, $H^p(\rn)$, with $p\in (0, 1]$, are good substitutes of Lebesgue spaces $L^p(\rn)$ when studying the boundedness of Calderón-Zygmund operators. In recent decades, in order to meet the requirements arising in the study of the boundedness of operators, partial differential equations and some other fields, various variants of Hardy spaces have been introduced and developed, such as weak Hardy spaces (see, for example, [@frs74; @fs87]), Hardy-Lorentz spaces (see, for example, [@at07; @abr17; @lyy16; @lyy17; @lyy18]) and Orlicz-Hardy spaces (see, for example, [@l66; @l67; @sto76; @yll]). Recently, in [@af1], as a generalization of the classical Hardy space and the Lorentz-Hardy space, Ablé and Feuto introduced the Hardy type space $\mathcal{H}^{(p,q)}(\rn)$ with $p,\,q\in(0,\infty)$ based on the N. Weiner amalgam spaces $(L^p,\ell^q)(\rn)$ and obtained an atomic decomposition of these Hardy-amalgam spaces when $q\in(0,\fz)$ and $p\in(0,\min\{1,q\})$. In [@bf], Bonami and Feuto introduced the Hardy type spaces $H_^\Phi(\rn)$ and $h_^\Phi(\rn)$ with respect to the amalgam space $(L^\Phi,\ell^1)(\rn)$, where $\Phi(t):=\frac{t}{\log(e+t)}$ for any $t\in[0,\infty)$ is an Orlicz function, and applied these Hardy-type spaces to study the linear decomposition of the product of the Hardy space $H^1(\rn)$ and its dual space $\BMO(\rn)$ as well as the local Hardy space $h^1(\rn)$ and its dual space $\bmo(\rn)$. Moreover, very recently, Cao et al. [@cky1] applied $h_^\Phi(\rn)$ to study the bilinear decomposition of the product of the local Hardy space $h^1(\rn)$ and its dual space $\bmo(\rn)$. Recall that both the Hardy type spaces $H_^\Phi(\rn)$ and $h_^\Phi(\rn)$ were defined in [@bf] via the (local) radial maximal functions, while $h_^\Phi(\rn)$ in [@cky1] was defined via the local grand maximal function. Moreover, no other real-variable characterizations of both the Hardy type spaces $H_^\Phi(\rn)$ and $h_^\Phi(\rn)$ are known so far. On the other hand, recently, to study the classification of weak solutions in the natural classes for the boundary value problems of a $t$-independent elliptic system in the upper plane, Auscher and Mourgoglou [@am] introduced the slice spaces $E^q_t(\rn)$. In [@ap], Auscher and Prisuelos-Arribas further introduced a more general slice space $(E^q_r)_t(\rn)$ and applied it to study the action of operators, such as the Hardy-Littlewood maximal operator, Calderón-Zygmund operators and Riesz potentials, on tent spaces. More precisely, recall that the tent space $T^q_r(\rr_+^{n+1})$, with $q,\,r\in(0,\fz)$, consists of all measurable functions $F$ on $\rr_+^{n+1}:=\rn\times(0,\fz)$ such that $$\\|F\\|_{T^q_r(\rr_+^{n+1})}:=\lf\\|\lf[\int_0^\fz\int_{B(\cdot,t)}\|F(y,t)\|^r \,\frac{dy\,dt}{t^{n+1}}\r]^{1/r}\r\\|_{L^q(\rn)}<\fz,$$ here and hereafter, for any $(x,t)\in\rr_+^{n+1}$, $B(x,t):=\{y\in\rn:\ \|y-x\|<t\}$. It is known (see [@cms]) that $T^q_r(\rr_+^{n+1})$ can be represented as $\sum_{i=1}^\fz\lz_i A_i$ with $\{\lz_i\}_{i\in\nn}\in \ell^q$ and $\{A_i\}_{i\in\nn}$ being $T^q_r(\rr_+^{n+1})$-atoms, that is, for any $i\in\nn$, there exists a ball $B_i\subset \rn$ such that $$\supp (A_i)\subset \widehat{B}_i:=\{(x,t)\in\rr_+^{n+1}:\ d(x, \rn\setminus B_i)\ge t\}$$ and $\iint_{\widehat{B}_i}\|A_i(x,t)\|^r\,\frac{dx\,dt}{t}\le \|B_i\|^{1-\frac{r}{q}}$, where $d(x, \rn\setminus B_i):=\inf\{\|x-y\|:\ y\in\rn\setminus B_i\}$. As a subspace of $T^q_r(\rr_+^{n+1})$, Auscher and Prisuelos-Arribas [@ap] introduced the space $\mathfrak{T}^q_r(\rr_+^{n+1})$ consisting of all functions $F\in T^q_r(\rr_+^{n+1})$ which can be represented as $\sum_{i=1}^\fz\lz_i A_i$ with $\{\lz_i\}_{i\in\nn}\in \ell^q$ and $T^q_r(\rr_+^{n+1})$-atoms $\{A_i\}_{i\in\mathbb{N}}$ satisfying the additional moment condition $\int_{\rr_+^{n+1}}A_i(x,t)\,dx=0$ for almost every $t\in (0,\fz)$ and any $i\in\mathbb{N}$. In [@ap], Auscher and Prisuelos-Arribas studied the behaviors of the Hardy-Littlewood maximal operator, Calderón-Zygmund operators and Riesz potentials on $T^q_r(\rr_+^{n+1})$ and $\mathfrak{T}^q_r(\rr_+^{n+1})$. As Auscher and Prisuelos-Arribas mentioned in [@ap], “it would be interesting to explore further these spaces (interpolation, etc) and their applications". One key tool used in [@ap] is the slice space which is defined via slicing the classical tent space norm at a fixed height. Recall that, for any $q,\ r,\ t\in(0,\fz)$, the slice space $(E^q_r)_t(\rn)$ in [@ap] is defined as the space of all locally $r$-integrable functions $f$ on $\rn$ such that $$\label{slice} \\|f\\|_{(E^q_r)_t(\rn)}:=\lf\{\int_\rn\left[t^{-n}\int_{B(x,t)}\|f(y)\|^r\,dy\right]^{q/r} \,dx\right\}^{1/q}<\fz.$$ In particular, $E^q_t(\rn):=(E^q_2)_t(\rn)$ was introduced in [@am]. A subspace $(\mathfrak{C}_r^q)_t(\rn)$ of $(E^q_r)_t(\rn)$ was also introduced in [@ap] in a way similar to $\mathfrak{T}^q_r(\rr_+^{n+1})$ (see also Definition \[auscher\] below). These slice spaces $(E^q_r)_t$ and $(\mathfrak{C}_r^q)_t$ were proved in [@ap] to be the retracts of the tent spaces $T^q_r(\rr_+^{n+1})$ and $\mathfrak{T}^q_r(\rr_+^{n+1})$, respectively. They are also special cases of the Wiener-amalgam spaces (see [@f]) which were first introduced by N. Wiener and further developed in time-frequency analysis and sampling theory. Properties of slice spaces such as the duality, the atomic decomposition and the interpolation were also clarified in [@am; @ap]. Observe that the Hardy type space $(\mathfrak{C}_r^q)_t(\rn)$ \[and also $\mathfrak{T}^q_r(\rr_+^{n+1})$\] was introduced in [@ap] via atoms and no other real-variable characterizations of these Hardy type spaces are known so far. Let $\Phi$ be an Orlicz function on $[0,\fz)$ and $q,\ t\in (0,\fz)$. Motivated by the aforementioned works, in this article, we first introduce a class of Orlicz-slice spaces, $(E_\Phi^q)_t(\rn)$, which generalize the slice spaces \[in this case, $\Phi(\tau):=\tau^r$ for any $\tau\in [0,\fz)$ with $r\in(0,\infty)$\] recently defined and studied by Auscher and Mourgoglou [@am] (the case $r=2$) as well as by Auscher and Prisuelos-Arribas [@ap]. Based on these Orlicz-slice spaces, we then introduce a new kind of Hardy-type spaces, the Orlicz-slice Hardy spaces $(HE_\Phi^q)_t(\rn)$, via the radial maximal functions. This new scale of Orlicz-slice Hardy spaces contains the variant of the Orlicz-Hardy space, $H_^\Phi(\rn)$ \[in this case, $q=t=1$\], of Bonami and Feuto [@bf] as well as the Hardy-amalgam space \[in this case, $t=1$ and $\Phi(\tau):=\tau^p$ for any $\tau\in [0,\fz)$ with $p\in (0,\fz)$\] of Ablé and Feuto [@af1] as special cases. Their characterizations via the atom, the molecule, various maximal functions, the Poisson integral and the Littlewood-Paley functions are also obtained. As an application of these characterizations, we then establish finite atomic characterizations of Orlicz-slice Hardy spaces, which further induce a description of their dual spaces and a criterion on the boundedness of sublinear operators from these Orlicz-slice Hardy spaces into a quasi-Banach space. Then, applying this criterion, we obtain the boundedness of $\delta$-type Calderón-Zygmund operators on these Orlicz-slice Hardy spaces. Moreover, the relations between the Orlicz-slice space and the Orlicz-slice Hardy space, or between the Hardy-type space $(\mathfrak{C}_r^q)_t(\rn)$, with $t\in (0,\fz)$, $r\in (1,\fz)$ and $q\in (\frac n{n+1}, 1]$, from [@ap] and $(HE_\Phi^q)_t(\rn)$ in the case when $\Phi(\tau):=\tau^s$ for any $\tau\in [0,\fz)$ with $s\in (\frac n{n+1},q]$ are also clarified. All these results of this article are new even for slice Hardy spaces and, moreover, for Hardy-amalgam spaces, the molecular characterization, the Littlewood-Paley function characterizations, the dual spaces and the boundedness of $\delta$-type Calderón-Zygmund operators on these Hardy-type spaces are also new. Thus, the results obtained in this article essentially complement and generalize the real-variable theories of the Hardy-amalgam space in [@af1] as well as the Hardy-type space $H_^\Phi(\rn)$ in [@bf]. To be more precise, in Section \[s2\] of this article, we introduce the notion of Orlicz-slice spaces $(E_\Phi^q)_t(\rn)$ and then present some basic properties of $(E_\Phi^q)_t(\rn)$, such as their equivalence relation with the Orlicz-amalgam spaces (see Proposition \[th2\] below), the Fefferman-Stein vector-valued inequality for the Hardy-Littlewood maximal operator on $(E_\Phi^q)_t(\rn)$ (see Theorem \[main\] below), the boundedness of the Hardy-Littlewood maximal operator on $(E_\Phi^q)_t(\rn)$ (see Corollary \[main2\] below), and the dual spaces of $(E_\Phi^q)_t(\rn)$ (see Theorem \[dual\] below). The boundedness of the Hardy-Littlewood maximal operator on $(E_\Phi^q)_t(\rn)$ is a key tool in this article. Recall that the boundedness of the Hardy-Littlewood maximal operator on the amalgam space $(L^p,\ell^q)(\rn)$ with $p,\,q\in(1,\infty)$ was obtained in [@chh]. However, the approach used in [@chh] for $(L^p,\ell^q)(\rn)$ is no longer feasible for $(E_\Phi^q)_t(\rn)$ because the quasi-norm $\\|\cdot\\|_{L^\Phi(\rn)}$ cannot be represented as an integral and hence cannot apply the weighted boundedness of the Hardy-Littlewood maximal operator. To overcome this obstacle, we employ a different method, namely, we first establish a generalization of [@ap Lemma 4.1] via replacing the maximal function and $L^r(\rn)$ norm therein, respectively, by the vector-valued maximal function and $L^\Phi(\rn)$ norm here (see Lemma \[mm\] below), which plays a key role in establishing the Fefferman-Stein vector-valued inequality for the Hardy-Littlewood maximal operator on $(E_\Phi^q)_t(\rn)$ (see the proof of Theorem \[main\]); from Theorem \[main\], we immediately induce the desired boundedness of the Hardy-Littlewood maximal operator on $(E_\Phi^q)_t(\rn)$. We also point out that the proof of Theorem \[dual\] strongly depends on Proposition \[th2\] and the well-known dual spaces of Orlicz-amalgam spaces. Moreover, in Lemma \[ballproof\] below, we further prove that the Orlicz-slice spaces $(E_\Phi^q)_t(\rn)$ are ball quasi-Banach function spaces considered in [@ykds] and hence all results from [@ykds] are applicable to $(E_\Phi^q)_t(\rn)$. In Section \[s3\], based on the Orlicz-slice spaces $(E_\Phi^q)_t(\rn)$, we first introduce the Orlicz-slice Hardy spaces, $(HE_\Phi^q)_t(\rn)$, which are defined via the radial maximal functions (see Definition \[dh\] below) and then present some fundamental properties of these Orlicz-slice Hardy spaces $(HE_\Phi^q)_t(\rn)$ including characterizations via the grand and the non-tangential maximal functions (see Theorem \[mdj\] below), the poisson integral (see Theorem \[poisson\] below), the atom (see Theorem \[atom ch\] below), the molecule (see Theorem \[molecular\] below), the Littlewood-Paley functions (see Theorems \[lusin\], \[gfunction\] and \[glamda\] below) and the finite atomic decomposition (see Theorem \[finite\] below). We also clarify the relations between $(E_\Phi^q)_t(\rn)$ and $(HE_\Phi^q)_t(\rn)$ in Theorem \[dayu1\] below. The proofs of all main results in Section \[s3\] are given in Section \[s4\]. Recall that a real-variable theory of Hardy spaces related to ball quasi-Banach function spaces was recently developed in [@ykds]. The results obtained in [@ykds] are of so wide generality that, in Section \[s4\], we can directly apply them to obtain the atomic and the molecular characterizations as well as those characterizations via various maximal functions, the Poisson integral and the Lusin area function of the Orlicz-slice Hardy space $(HE_\Phi^q)_t(\rn)$ as well as the relation between $(HE_\Phi^q)_t(\rn)$ and $(E_\Phi^q)_t(\rn)$. Then, using the atomic characterization, we further establish the Littlewood-Paley $g$-function and $g_\lz^\ast$-function characterizations of $(HE_\Phi^q)_t(\rn)$ and also the finite atomic characterization. We point out that, in [@af1], Ablé and Feuto introduced the Hardy-amalgam space $\mathcal{H}^{(p,q)}(\rn)$ and obtained the non-tangential maximal function characterization, the Poisson integral characterization and the atomic decomposition as well as the finite atomic decomposition of this space. To the best of our knowledge, this might be the first article to deal with the real-variable theory of Hardy spaces based on amalgam spaces. Comparing with [@af1], the approach used in this article via the general theory of [@ykds] for the corresponding characterizations of $(HE_\Phi^q)_t(\rn)$ is much simpler. Also, comparing with the atomic characterization of $\mathcal{H}^{(p,q)}(\rn)$ obtained in [@af1], the atomic characterization of $(HE_\Phi^q)_t(\rn)$ obtained in this article holds true on a wider range even when $(HE_\Phi^q)_t(\rn)$ is reduced to $\mathcal{H}^{(p,q)}(\rn)$ \[in this case, $t=1$ and $\Phi(\tau):=\tau^p$ for any $\tau\in [0,\fz)$ with $p\in (0,\fz)$\], which improves the related result in [@af1]. In Section \[s5\], as an application of both the atomic characterization (Theorem \[atom ch\]) and the finite atomic characterization (Theorem \[finite\]) of $(HE_\Phi^q)_t(\rn)$ obtained in Section \[s3\], we prove that the dual spaces of $(HE_\Phi^q)_t(\rn)$ can be described as certain Campanato spaces related to the Orlicz-slice spaces (see Theorem \[dual2\] below). The last section, Section \[s6\], is devoted to some further applications of the characterizations obtained in Section \[s3\]. We first establish a criterion on the boundedness of sublinear operators from $(HE_\Phi^q)_t(\rn)$ into a quasi-Banach space (see Theorem \[suanzi1\] and Corollary \[suanzi\] below), which are further used to obtain the boundedness of the $\delta$-type Calderón-Zygmund operators on $(HE_\Phi^q)_t(\rn)$ (see Theorems \[cz\] and \[cz2\]). Moreover, in Proposition \[au\] below, we clarify the relation between $(\mathfrak{C}_r^q)_t(\rn)$, with $t\in (0,\fz)$, $r\in (1,\fz)$ and $q\in (\frac n{n+1}, 1]$, from [@ap] and $(HE_\Phi^q)_t(\rn)$ in the case when $\Phi(\tau):=\tau^s$ for any $\tau\in [0,\fz)$ with $s\in (0,q]$. Observe that a real-variable theory of local Hardy spaces based on the Orlicz-slice spaces can also be developed. However, to limit the length of this article, we will consider this local version in another article. Finally, we make some convention on notation. For any $x\in\rn$ and $r\in(0,\infty)$, let $B(x,r):=\{y\in\rn:\|x-y\|<r\}$ and $\bar{B}(x,r)$ be its closure in $\rn$. For any $r\in(0,\infty)$, $f\in L^1_\loc(\rn)$ and $x\in\rn$, let $$\dashint_{B(x,\,r)}f(y)\,dy:=\frac1{\|B(x,r)\|}\int_{B(x,r)}f(y)\,dy,$$ here and hereafter, $L^1_\loc(\rn)$ denotes the space of all locally integrable functions. For any set $E$, we use $\chi_{E}$ to denote its characteristic function and $\# E$ its cardinality. We also use $\vec{0}_n$ to denote the origin of $\rn$. Let $\mathcal{S}(\rn)$ denote the collection of all Schwartz functions on $\rn$, equipped with the classical well-known topology, and $\mathcal{S}'(\rn)$ its topological dual, namely, the collection of all bounded linear functionals on $\mathcal{S}(\rn)$ equipped with the weak-$\ast$ topology. Let $\mathbb{N}:=\{1,\,2,...\}$ and $\mathbb{Z}_+:=\mathbb{N}\bigcup\{0\}$. Denote by the symbol $\mathcal{Q}$ the set of all cubes having their edges parallel to the coordinate axes. Also, for any $x\in\rn$ and $l\in(0,\fz)$, $Q(x,l)$ denotes the cube with the center $x$ and the side-length $l$. Furthermore, for any cube $Q\in\mathcal{Q}$ and $j\in\mathbb{Z}_+$, let $S_j(Q):=(2^{j+1}Q)\setminus(2^jQ)$ with $j\in\mathbb{N}$ and $S_0(Q):=2Q$. For any $\varphi\in\mathcal{S}(\rn)$ and $t\in(0,\infty)$, let $\varphi_t(\cdot):=t^{-n}\varphi(t^{-1}\cdot)$. For any $s\in\mathbb{R}$, we denote by $\lfloor s\rfloor$ the largest integer not greater than $s$. For any $p\in[0,1]$, let $p'$ be its conjugate index, that is, $p'$ satisfies $1/p+1/p'=1$. We always use $C$ to denote a positive constant, which is independent of the main parameter, but it may vary from line to line. Moreover, we use $C_{(\gamma,\ \beta,\ \ldots)}$ to denote a positive constant depending on the indicated parameters $\gamma,\ \beta,\ \ldots$. If, for any real functions $f$ and $g$, $f\leq Cg$, we then write $f\lesssim g$ and, if $f\lesssim g\lesssim f$, we then write $f\sim g$. Orlicz-slice spaces\[s2\] ========================= In this section, we introduce the slice spaces related to Orlicz functions and present some of their basic properties such as the boundedness of maximal operators, which are used in the later sections. We begin with the notions of both Orlicz functions and Orlicz spaces (see, for example, [@mmz]). \[d1.1\] A function $\Phi:\ [0,\infty)\ \rightarrow\ [0,\infty)$ is called an Orlicz function if it is non-decreasing and satisfies $\Phi(0)= 0$, $\Phi(t)>0$ whenever $t\in(0,\infty)$ and $\lim_{t\rightarrow\infty}\Phi(t)=\infty$. An Orlicz function $\Phi$ is said to be of lower (resp., upper) type $p$ with $p\in(-\infty,\infty)$ if there exists a positive constant $C_{(p)}$, depending on $p$, such that, for any $t\in[0,\infty)$ and $s\in(0,1)$ \[resp., $s\in [1,\infty)$\], $$\Phi(st)\le C_{(p)}s^p \Phi(t).$$ A function $\Phi:\ [0,\infty)\ \rightarrow\ [0,\infty)$ is said to be of positive lower (resp., upper) type if it is of lower (resp., upper) type $p$ for some $p\in(0,\infty)$. \[d1.2\] Let $\Phi$ be an Orlicz function with positive lower type $p_{\Phi}^-$ and positive upper type $p_{\Phi}^+$. The Orlicz space $L^\Phi(\rn)$ is defined to be the set of all measurable functions $f$ such that $$\\|f\\|_{L^\Phi(\rn)}:=\inf\lf\{\lambda\in(0,\infty):\ \int_{\rn}\Phi\lf(\frac{\|f(x)\|}{\lambda}\r)\,dx\le1\r\}<\infty.$$ We now give some basic properties of Orlicz functions. \[lem1\] Let $\Phi$ be an Orlicz function with positive upper type $p_{\Phi}^+$. Then there exists a positive constant $C$ such that $$\Phi(t_1+t_2)\le C\lf[\Phi(t_1)+\Phi(t_2)\r],\quad\forall\ t_1,\ t_2\in[0,\infty).$$ Obviously we only need to consider the case when $t_1+t_2>0$. If $p_{\Phi}^+ \in(0,1]$, then, for any $i\in\{1,2\}$, $$\frac{t_i}{t_1+t_2}\Phi(t_1+t_2)\lesssim\Phi(t_i)$$ and hence $$t_i\Phi(t_1+t_2)\lesssim\Phi(t_i)(t_1+t_2),$$ which, via taking the summation on $i$ on both side, further implies the desired conclusion. If $p_{\Phi}^+ \in(1,\infty)$, then let $\widetilde{\Phi}(t):=\Phi(t^{1/p_{\Phi}^+})$ for any $t\in[0,\infty)$. It is easy to check that $\widetilde{\Phi}$ is an Orlicz function of upper type $1$ and hence, by the proved conclusion, we have $$\Phi(t_1+t_2)=\widetilde{\Phi}\lf([t_1+t_2]^{p_{\Phi}^+}\r)\lesssim\widetilde{\Phi}\lf(t_1^{p_{\Phi}^+}\r) +\widetilde{\Phi}\lf(t_2^{p_{\Phi}^+}\r) \sim\Phi(t_1)+\Phi(t_2).$$ This finishes the proof of Lemma \[lem1\]. When $\Phi$ is an Orlicz function with positive upper type $p_{\Phi}^+$, from Lemma \[lem1\], it is easy to deduce that $\\|\cdot\\|_{L^\Phi(\rn)}$ is a quasi-norm. The following lemma is well known. \[lem2\] Let $\Phi$ be an Orlicz function with positive lower type $p_{\Phi}^-$ and positive upper type $p_{\Phi}^+$ and $$\widetilde{\Phi}(t):=\int_0^t\frac{\Phi(s)}{s}\,ds,\quad \forall\,t\in(0,\infty).$$ Then $\widetilde{\Phi}$ is also an Orlicz function, which is equivalent to $\Phi$ and $\widetilde{\Phi}$ is continuous and strictly increasing. \[re\] Observe that all the results stated in this article are invariant under the change of equivalent Orlicz functions. Moreover, equivalent Orlicz functions share the same positive upper and the same lower type numbers. In what follows, by Lemma \[lem2\], without loss of generality, we may always assume that an Orlicz function $\Phi$ is continuous and strictly increasing. \[lem3\] Let $\Phi$ be an Orlicz function with positive lower type $p_{\Phi}^-$. If the inequality that $$\int_{\rn}\Phi\lf(\frac{\|f(x)\|}{\lambda}\r)\,dx\leqslant\widetilde{C}\quad for\ some\ \lambda\in(0,\infty) \ and\ positive\ constant\ \widetilde{C}$$ holds true, then there exists a positive constant $C$, depending on $\widetilde{C}$ and $p_{\Phi}^-$, such that $\\|f\\|_{L^\Phi(\rn)}\le C\lambda$. The proof is simple and we can take $C:=(1+\widetilde{C}C_{(p_{\Phi}^-)})^{1/p_{\Phi}^-}$ with $\widetilde{C}$ as in the assumption of Lemma \[lem3\]. This finishes the proof of Lemma \[lem3\]. Now we introduce the Orlicz-slice space and the Orlicz-amalgam space. The former is a generalization of the slice spaces introduced in [@am; @ap], and the latter is a generalization of the classical amalgam space $(L^p,\ell^q)$ defined by N. Wiener in 1926, in the formulation of his generalized harmonic analysis. \[d2\] Let $t,\ q\in(0,\infty)$ and $\Phi$ be an Orlicz function with positive lower type $p_{\Phi}^-$ and positive upper type $p_{\Phi}^+$. The Orlicz-slice space $(E_\Phi^q)_t(\rn)$ is defined to be the set of all measurable functions $f$ such that $$\\|f\\|_{(E_\Phi^q)_t(\rn)} :=\lf\{\int_{\rn}\lf[\frac{\\|f\chi_{B(x,t)}\\|_{L^\Phi(\rn)}} {\\|\chi_{B(x,t)}\\|_{L^\Phi(\rn)}}\r]^q\,dx\r\}^{\frac{1}{q}}<\infty.$$ \[d1.3\] Let $t,\ q\in(0,\infty)$ and $\Phi$ be an Orlicz function with positive lower type $p_{\Phi}^-$ and positive upper type $p_{\Phi}^+$. The Orlicz-amalgam space $\ell^q(L^\Phi_t)(\rn)$ is defined to be the set of all measurable functions $f$ such that $$\\|f\\|_{\ell^q(L^\Phi_t)(\rn)}:=\lf[\sum_{k\in \mathbb{Z}^n}\lf\\|f\chi_{Q_{tk}}\r\\|^q_{L^\Phi(\rn)}\r]^{\frac{1}{q}}<\infty,$$ where $Q_{tk}:=t[k+[0,1)^n]$ for any $t\in(0,\infty)$ and $k\in\mathbb{Z}^n$. \[re2.10\] - Both the Orlicz-slice space and the Orlicz-amalgam space fall into the scale of Wiener-amalgam spaces introduced by Feichtinger [@f]. By Lemmas \[lem1\], \[lem3\] and [@f Theorem 1], we know that both the Orlicz-slice space and the Orlicz-amalgam space are quasi-Banach spaces. - If $t=1$ and $\Phi(\tau):=\tau^p$ for any $\tau\in [0,\fz)$ with $p\in (0,\fz)$, then $(E_\Phi^q)_t(\rn)$ coincides with the Weiner amalgam spaces $(L^p,\ell^q)(\rn)$ in [@af1]. By [@af1 Proposition 2.1], we have $(L^p,\ell^q)(\rn)\subset L^p(\rn)\bigcap L^q(\rn)$ when $p\in(0, q)$ and $L^p(\rn)\bigcup L^q(\rn)\subset(L^p,\ell^q)(\rn)$ when $ q\in(0, p)$, here and hereafter, for any $r\in(0,\infty]$, the symbol $L^r(\rn)$ denotes the set of all measurable functions $f$ such that $$\\|f\\|_{L^r(\rn)}:=\lf\{\int_{\rn}\|f(x)\|^r\,dx\r\}^{1/r}<\infty$$ with the usual modification made when $r=\infty$. - If $\Phi(\tau):=\tau^r$ for any $\tau\in[0,\fz)$ with $r\in(0,\fz)$, then $(E_\Phi^q)_t(\rn)$ and $(E_r^q)_t(\rn)$ from [@am; @ap] coincide with equivalent quasi-norms. The following proposition clarifies the relation between $(E_r^q)_t(\rn)$, with $t,\ q,\ r\in(0,\infty)$, and $L^q(\rn)$, whose proof is a slight modification of the proof of [@af1 Proposition 2.1]. \[ggg\] Let $t,\ q,\ r\in(0,\infty)$. - If $r\in(0, q]$, then $L^r(\rn)\cup L^q(\rn)\subset (E_r^q)_t(\rn)$; precisely, for any $f\in L^r(\rn)\cup L^q(\rn)$, then $f\in (E_r^q)_t(\rn)$ and $\\|f\\|_{(E_r^q)_t(\rn)}\leq\min\{\\|f\\|_{L^r(\rn)}, \\|f\\|_{L^q(\rn)}\}$; - If $q\in(0, r]$, then $(E_r^q)_t(\rn)\subset L^q(\rn)$; precisely, for any $f\in(E_r^q)_t(\rn)$, then $f\in L^q(\rn)$ and $\\|f\\|_{L^q(\rn)}\leq\\|f\\|_{(E_r^q)_t(\rn)}$. - $(E_q^q)_t(\rn)$ and $L^q(\rn)$ coincide with the same quasi-norms. Observe that (iii) is an immediate corollary of (i) and (ii). Thus, to complete the proof of this proposition, we only need to show (i) and (ii). We first show (i). In this case, for any $f\in L^q(\rn)$, using the Hölder inequality and the Fubini theorem, we obtain $$\begin{aligned} \\|f\\|_{(E_r^q)_t(\rn)}&\le \lf[\int_{\rn}\lf\{\frac1{\|B(x,t)\|}\lf[\int_{B(x,t)}\|f(y)\|^q\,dy\r]^\frac rq\|B(x,t)\|^\frac1{(q/r)'}\r\}^\frac qr\,dx\r]^\frac1q\\ &=\lf\{\int_{\rn}\frac1{\|B(x,t)\|}\int_{B(x,t)}\|f(y)\|^q\,dy\,dx\r\}^\frac1q=\\|f\\|_{L^q(\rn)}.\end{aligned}$$ Also, for any $f\in L^r(\rn)$, applying the Minkowski inequality, we conclude that $$\\|f\\|_{(E_r^q)_t(\rn)}\le\lf[\int_{\rn}\lf\{\int_{\rn}\lf[t^{-n}\chi_{B(x,t)}(y)\|f(y)\|^r\r]^\frac qrdx\r\}^\frac rqdy\r]^\frac1r=\\|f\\|_{L^r(\rn)}.$$ Thus, $L^r(\rn)\bigcup L^q(\rn)\subset(E_r^q)_t(\rn)$, that is, (i) holds true. Now, we prove (ii). In this case, by the Hölder inequality, we have $$\begin{aligned} \\|f\\|_{L^q(\rn)}&=\lf\{\int_{\rn}\frac1{\|B(x,t)\|}\int_{B(x,t)}\|f(y)\|^q\,dy\,dx\r\}^\frac1q\\ &\leq\lf\{\int_{\rn}\frac1{\|B(x,t)\|}\lf[\int_{B(x,t)}\|f(y)\|^r\,dy\r]^{\frac{q}{r}} \lf[\int_{B(x,t)}dy\r]^{\frac{1}{(r/q)'}}\,dx\r\}^\frac1q =\\|f\\|_{(E_r^q)_t(\rn)}.\end{aligned}$$ From this, we deduce $(E_r^q)_t(\rn)\subset L^q(\rn)$, which completes the proof of (ii) and hence of Proposition \[ggg\]. Observing that, for any $x\in\rn$ and $t\in(0,\infty)$, $$\label{qiu} \\|\chi_{B(x,t)}\\|_{L^\Phi(\rn)}=\lf[\Phi^{-1}\lf(\frac{1}{\|B(x,t)\|}\r)\r]^{-1} =\lf[\Phi^{-1}\lf(\frac{1}{\varepsilon_nt^n}\r)\r]^{-1}=:\widetilde{C}_{(\Phi,t)}$$ is independent of $x,$ where $\varepsilon_n$ denotes the volume of the unit ball in $\rn$ and $\Phi^{-1}$ the inverse function of $\Phi$, we have the next proposition, which shows that, for any $t\in(0,\infty)$, the Orlicz-slice space $(E_\Phi^q)_t(\rn)$ is equivalent to the Orlicz-amalgam space $\ell^q(L^\Phi_t)(\rn)$. \[th2\] Let $t,\ q\in(0,\infty)$ and $\Phi$ be an Orlicz function with positive lower type $p_{\Phi}^-$ and positive upper type $p_{\Phi}^+$. Then $(E_\Phi^q)_t(\rn)$ and $\ell^q(L^\Phi_t)(\rn)$ coincide and, for any $f\in(E_\Phi^q)_t(\rn)$, $$\lf[t^{n}\sum_{k\in \mathbb{Z}^n}\lf\\|f\chi_{Q_{ kt}}\r\\|^q_{L^\Phi(\rn)}\r]^{\frac{1}{q}} \sim\lf\{\int_{\rn}\lf[\lf\\|f\chi_{B(x,t)}\r\\|_{L^\Phi(\rn)}\r]^q\,dx\r\}^{\frac{1}{q}},$$ where the equivalent positive constants are independent of $f$ and $t$. We first show that $$\label{dj} \lf[t^{n}\sum_{k\in \mathbb{Z}^n}\lf\\|f\chi_{Q_{tk}}\r\\|^q_{L^\Phi(\rn)}\r]^{\frac{1}{q}} \sim\lf\{\int_{\rn}\lf[\lf\\|f\chi_{B(x,2\sqrt{n}t)}\r\\|_{L^\Phi(\rn)}\r]^q\,dx\r\}^{\frac{1}{q}}.$$ Indeed, it is easy to see that $$\begin{aligned} \lf[t^{n}\sum_{k\in \mathbb{Z}^n}\lf\\|f\chi_{Q_{tk}}\r\\|^q_{L^\Phi(\rn)}\r]^{\frac{1}{q}} =\lf\\|\sum_{k\in \mathbb{Z}^n}\lf\\|f\chi_{Q_{tk}}\r\\|_{L^\Phi(\rn)}\chi_{Q_{tk}}\r\\|_{L^q(\rn)}.\end{aligned}$$ For any $x\in Q_{tk}$, from $Q_{tk}\subset B(x,2\sqrt{n}t)$, it follows that $$\lf\\|f\chi_{Q_{tk}}\r\\|_{L^\Phi(\rn)}\chi_{Q_{tk}}(x)\le\lf\\|f\chi_{B(x,2\sqrt{n}t)}\r\\|_{L^\Phi(\rn)}.$$ Thus, combining the above two formulas, we conclude that $$\label{yyy} \lf\\|\sum_{k\in \mathbb{Z}^n}\lf\\|f\chi_{Q_{tk}}\r\\|_{L^\Phi(\rn)}\chi_{Q_{tk}}\r\\|_{L^q(\rn)} \le\lf\\|\lf\\|f\chi_{B(\cdot,2\sqrt{n}t)}\r\\|_{L^\Phi(\rn)}\r\\|_{L^q(\rn)}.$$ To prove the opposite inequality, for any given $x\in \rn$, we let $$M_x:=\lf\{k\in\mathbb{Z}^n:\ Q_{tk} \cap B(x,2\sqrt{n}t)\neq\emptyset\r\}.$$ Then the cardinality $\#M_x\lesssim1$ and, if $k\in M_x$, then $x\in B(tk,4\sqrt{n}t)$, which further implies that $$\begin{aligned} \lf\\|f\chi_{B(x,2\sqrt{n}t)}\r\\|_{L^\Phi(\rn)}&=\lf\\|\sum_{k\in \mathbb{Z}^n}f \chi_{B(x,2\sqrt{n}t)}\chi_{Q_{tk}}\r\\|_{L^\Phi(\rn)} \lesssim\sum_{k\in M_x}\lf\\|f\chi_{B(x,2\sqrt{n}t)}\chi_{Q_{tk}}\r\\|_{L^\Phi(\rn)}\\ &\lesssim\sum_{k\in \mathbb{Z}^n}\lf\\|f\chi_{Q_{tk}}\r\\|_{L^\Phi(\rn)}\chi_{B(tk,4\sqrt{n}t)}(x).\end{aligned}$$ Thus, we have $$\label{yy} \lf\\|\lf\\|f\chi_{B(\cdot,2\sqrt{n}t)}\r\\|_{L^\Phi(\rn)}\r\\|_{L^q(\rn)}\lesssim\lf\\|\sum_{k\in \mathbb{Z}^n}\lf\\|f\chi_{Q_{tk}}\r\\|_{L^\Phi(\rn)}\chi_{B(tk,4\sqrt{n}t)}\r\\|_{L^q(\rn)}.$$ It is easy to see that there exist $N\in\mathbb{N}$ and $\{k_1,\ \ldots,\ k_N\}\subset\mathbb{Z}^n$, independent of $t$, such that $N\lesssim1$ and $B(\vec{0}_n,4\sqrt{n}t)$ $\subseteq\bigcup_{m=1}^N Q_{tk_m}$ and hence $$\sum_{k\in \mathbb{Z}^n}\lf\\|f\chi_{Q_{tk}}\r\\|_{L^\Phi(\rn)}\chi_{B(tk,4\sqrt{n}t)} \le\sum_{m=1}^N\sum_{k\in \mathbb{Z}^n}\lf\\|f\chi_{Q_{tk}}\r\\|_{L^\Phi(\rn)}\chi_{Q_{t(k_m+k)}}.$$ By this, the translation invariance of the Lebesgue measure and $N\lesssim1$, we further obtain $$\begin{aligned} \lf\\|\sum_{k\in \mathbb{Z}^n}\lf\\|f\chi_{Q_{tk}}\r\\|_{L^\Phi(\rn)}\chi_{B(tk,4\sqrt{n}t)}\r\\|_{L^q(\rn)} &\leqslant\lf\\|\sum_{m=1}^N\sum_{k\in \mathbb{Z}^n}\lf\\|f\chi_{Q_{tk}}\r\\|_{L^\Phi(\rn)}\chi_{Q_{t(k_m+k)}}\r\\|_{L^q(\rn)}\\ &\lesssim\lf[t^{n}\sum_{k\in \mathbb{Z}^n}\lf\\|f\chi_{Q_{tk}}\r\\|^q_{L^\Phi(\rn)}\r]^{\frac{1}{q}},\end{aligned}$$ which, together with (\[yy\]), implies that the opposite inequality of (\[yyy\]) holds true. Thus, (\[dj\]) holds true. Now, to complete the proof of Proposition \[th2\], we only need to show that $$\lf\\|\lf\\|f\chi_{B(\cdot,2\sqrt{n}t)}\r\\|_{L^\Phi(\rn)}\r\\|_{L^q(\rn)} \backsim\lf\\|\lf\\|f\chi_{B(\cdot,t)}\r\\|_{L^\Phi(\rn)}\r\\|_{L^q(\rn)}.$$ Since both $\overline{B}(\vec{0}_n,4\sqrt{n}t)$ and $\overline{B}(\vec{0}_n,t)$ are compact subsets of $\rn$ with nonempty interiors, it follows that there exist $M\in\mathbb{N}$ and $\{x_1,\ \ldots,\ x_M\}\subset\rn$, independent of $t$, such that $M\lesssim1$ and $B(\vec{0}_n,4\sqrt{n}t)\subseteq\bigcup_{m=1}^MB(x_m,t).$ Thus, for any $x\in\rn$, we have $$\begin{aligned} \lf\\|f\chi_{B(x,2\sqrt{n}t)}\r\\|_{L^\Phi(\rn)}&=\lf\\|f\sum_{m=1}^M\chi_{B(x+x_m,t)}\r\\|_{L^\Phi(\rn)} \lesssim\sum_{m=1}^M\lf\\|f\chi_{B(x+x_m,t)}\r\\|_{L^\Phi(\rn)}.\end{aligned}$$ By this, the translation invariance of the Lebesgue measure and $M\lesssim1$, we further obtain $$\lf\\|\lf\\|f\chi_{B(\cdot,2\sqrt{n}t)}\r\\|_{L^\Phi(\rn)}\r\\|_{L^q(\rn)} \lesssim\sum_{m=1}^M\lf\\|\lf\\|f\chi_{B(x+x_m,t)}\r\\|_{L^\Phi(\rn)}\r\\|_{L^q(\rn)} \lesssim\lf\\|\lf\\|f\chi_{B(\cdot,t)}\r\\|_{L^\Phi(\rn)}\r\\|_{L^q(\rn)}.$$ The reverse inequality obviously holds true. This finishes the proof of Proposition \[th2\]. Recall that the centered Hardy-Littlewood maximal operator $\mathcal{M}$ is defined by setting, for any locally integrable function $f$ and $x\in\rn$, $$\mathcal{M}(f)(x):=\sup_{r\in(0,\infty)}\dashint_{B(x,r)}\|f(x)\|\,dy,$$ and the uncentered Hardy-Littlewood maximal operator $\mathcal{M}_u$ is defined by setting, for any locally integrable function $f$ and $x\in\rn$, $$\mathcal{M}_u(f)(x):=\sup_{x\in B} \dashint_B \|f(y)\|\,dy,$$ where the supremum is taken over all balls $B$ of $\rn$ containing $x$. Borrowing some ideas from the proof of [@ap Lemma 4.1], we have the following very useful technical lemma, which plays a vital role in the proof of Theorem \[main\] below. \[mm\] Let $t\in(0,\infty),\ r\in(1,\infty)$ and $\Phi$ be an Orlicz function with positive lower type $p_{\Phi}^-\in(1,\infty)$ and positive upper type $p_{\Phi}^+$. Then, for any sequence $\{f_j\}_{j\in\mathbb{Z}}$ of locally integrable functions and $x\in\rn$, it holds true that $$\begin{aligned} \lf\\|\lf\{\sum _{j\in\mathbb{Z}}\lf[\mathcal{M}(f_j)\r]^r\r\}^{\frac{1}{r}}\chi_{B(x,t)}\r\\|_{L^\Phi(\rn)} &\le C\lf\\|\lf\{\sum_{j\in\mathbb{Z}}\lf\|f_j\r\|^r\r\}^{\frac{1}{r}}\chi_{B(x,2t)}\r\\|_{L^\Phi(\rn)} \\&\quad+C\\|\chi_{B(x,t)}\\|_{L^\Phi(\rn)} \lf\{\sum_{j\in\mathbb{Z}}\lf[\mathcal{M}_u\lf(\dashint_{B(\cdot,t)}\lf\|f_j(z)\r\|\,dz\r)(x)\r]^r\r\}^{\frac{1}{r}},\end{aligned}$$ where the positive constant $C$ is independent of $\{f_j\}_{j\in\mathbb{Z}}$, $x\in\rn$ and $t\in(0,\infty)$. Let $\{f_j\}_{j\in\mathbb{Z}}$ be a sequence of locally integrable functions. Fix $x\in\rn$. Then we have $$\begin{aligned} \lf\\|\lf\{\sum _{j\in\mathbb{Z}}\lf[\mathcal{M}(f_j)\r]^r\r\}^{\frac{1}{r}}\chi_{B(x,t)}\r\\|_{L^\Phi(\rn)} &\lesssim\lf\\|\lf\{\sum _{j\in\mathbb{Z}}\lf[\sup_{s\in(0, t]}\dashint_{B(\cdot,s)}\lf\|f_j(z)\r\|\,dz\r]^r\r\}^{\frac{1}{r}}\chi_{B(x,t)}\r\\|_{L^\Phi(\rn)}\\ &\quad+\lf\\|\lf\{\sum _{j\in\mathbb{Z}}\lf[\sup_{s\in(t,\infty)}\dashint_{B(\cdot,s)}\lf\|f_j(z)\r\|\,dz\r]^r \r\}^{\frac{1}{r}}\chi_{B(x,t)}\r\\|_{L^\Phi(\rn)}\\ &=:\mathrm{I}+\mathrm{II}.\end{aligned}$$ Since $B(y,s)\subset B(x,2t)$ whenever $s\in(0,t]$ and $y\in B(x,t)$, it follows that $$\begin{aligned} \mathrm{I}&\sim\lf\\|\chi_{B(x,t)}\lf\{\sum _{j\in\mathbb{Z}}\lf[\sup_{s\in(0, t]}\dashint_{B(\cdot,s)}\lf\|f_j(z)\r\|\chi_{B(x,2t)}(z)\,dz\r]^r\r\}^{\frac{1}{r}}\r\\|_{L^\Phi(\rn)}\\ &\lesssim\lf\\|\lf\{\sum _{j\in\mathbb{Z}}\lf[\mathcal{M}\lf(f_j\chi_{B(x,2t)}\r)\r]^r\r\}^{\frac{1}{r}}\r\\|_{L^\Phi(\rn)} \lesssim\lf\\|\lf\{\sum _{j\in\mathbb{Z}}\lf\|f_j\r\|^r\r\}^{\frac{1}{r}}\chi_{B(x,2t)}\r\\|_{L^\Phi(\rn)},\end{aligned}$$ where, in the last inequality, we used the Orlicz Fefferman-Stein vector-valued inequality (see, for example, [@kk] or [@yll Theorem 2.1.4]). As for $\mathrm{II}$, observe that, for any $\xi,\ z\in\rn$, $\xi\in B(z,t)$ if and only if $z\in B(\xi,t)$ and, moreover, if $z\in B(y,s)$, $\xi\in B(z,t)$ and $s\in(t,\infty)$, then $\xi\in B(y,2s)$. Besides, observe that $y\in B(x,t)$ and $s\in(t,\infty)$ imply that $x\in B(y,2s).$ From these observations and the Fubini theorem, we deduce that $$\begin{aligned} \mathrm{II}&\sim\lf\\|\chi_{B(x,t)}\lf\{\sum_{j\in\mathbb{Z}}\lf[\sup_{s\in(t,\infty)} \dashint_{B(\cdot,s)}\dashint_{B(z,t)}\lf\|f_j(z)\r\| \,d\xi \,dz\r]^r\r\}^{\frac{1}{r}}\r\\|_{L^\Phi(\rn)}\\ &\lesssim\lf\\|\chi_{B(x,t)}\lf\{\sum_{j\in\mathbb{Z}}\lf[\sup_{s\in(t,\infty)} \dashint_{B(\cdot,2s)}\dashint_{B(\xi,t)}\lf\|f_j(z)\r\|\,dz\,d\xi\r]^r\r\}^{\frac{1}{r}}\r\\|_{L^\Phi(\rn)}\\ &\lesssim\lf\\|\chi_{B(x,t)}\lf\{\sum_{j\in\mathbb{Z}} \lf[\mathcal{M}_u\lf(\dashint_{B(\cdot,t)}\lf\|f_j(z)\r\|\,dz\r)(x)\r]^r\r\}^{\frac{1}{r}}\r\\|_{L^\Phi(\rn)}\\ &\lesssim\lf\\|\chi_{B(x,t)}\r\\|_{L^\Phi(\rn)}\lf\{\sum_{j\in\mathbb{Z}} \lf[\mathcal{M}_u\lf(\dashint_{B(\cdot,t)}\lf\|f_j(z)\r\|\,dz\r)(x)\r]^r\r\}^{\frac{1}{r}}.\end{aligned}$$ Gathering the estimates for $\mathrm{I}$ and $\mathrm{II}$, we then obtain the desired conclusion, which completes the proof of Lemma \[mm\]. \[d1.4\] A convex function $\Phi:\ [0,\infty)\ \rightarrow\ [0,\infty)$ is called a Young function if $\Phi$ is non-decreasing, $\Phi(0)= 0$ and $\lim_{t\rightarrow\infty}\Phi(t)=\infty$. For any Young function $\Phi$, its complementary function $\Psi:\ [0,\infty)\ \rightarrow\ [0,\infty)$ is defined by setting, for any $y\in[0,\infty)$ $$\Psi(y):=\sup\lf\{xy-\Phi(x):\ x\in[0,\infty)\r\}.$$ \[ddd\] A Young function $\Phi:\ [0,\infty)\ \rightarrow\ [0,\infty)$ is called an N-function if $\Phi(0)= 0$, $\Phi(t)>0$ for any $t\in(0,\infty)$, $\lim_{t\rightarrow\infty}\frac{\Phi(t)}{t}=\infty$ and $\lim_{t\rightarrow0^+}\frac{\Phi(t)}{t}=0$, here and hereafter, $t\rightarrow0^+$ means $t\in(0,\infty)$ and $t\rightarrow0$. \[odj\] Let $\Phi$ be an Orlicz function with lower type $p_{\Phi}^-\in(1,\infty)$ and positive upper type $p_{\Phi}^+$. Then there exists an N-function $\widetilde{\Phi}$, which is equivalent to $\Phi$. Consider the function $$\widetilde{\Phi}(t):=\begin{cases}\displaystyle\int_0^t \sup_{\tau\in(0,s)}\frac{\Phi(\tau)}{\tau}\,ds,\quad\forall\,t \in (0,\infty],\\ \ 0,\quad t=0. \end{cases}$$ Then it is easy to prove that $\widetilde{\Phi}$ is convex on $[0,\infty)$. By the assumption that $p_{\Phi}^-\in[1,\infty)$, we know that, for any $t\in(0,\infty)$, $$\widetilde{\Phi}(t)\le t\sup_{\tau\in(0,t)}\frac{\Phi(\tau)}{\tau}\le C_{(p_\Phi^-)}t\sup_{\tau\in(0,t)}\lf(\frac{\tau}{t}\r)^{p_\Phi^-}\frac{\Phi(t)}{\tau}\le C_{(p_\Phi^-)}\Phi(t).$$ On the other hand, for any $t\in(0,\infty)$, we have $$\Phi(t)\le C_{(p_\Phi^+)}2^{p_\Phi^+}\Phi(t/2)\le C_{(p_\Phi^+)}2^{p_\Phi^+}\int_{t/2}^{t}\sup_{\tau\in(0,s)}\frac{\Phi(\tau)}{\tau}ds \le C_{(p_\Phi^+)}2^{p_\Phi^+}\widetilde{\Phi}(t).$$ Thus, we obtain $\Phi\sim\widetilde{\Phi}$. Moreover, it is easy to prove that $\widetilde{\Phi}$ is an N-function, which completes the proof of Lemma \[odj\]. \[nfunction\] 1. Observe that all the results stated in this article are invariant under the change of equivalent Orlicz functions. In what follows, by Lemma \[odj\] and its proof, without loss of generality, we may always assume that an Orlicz function $\Phi$ of lower type $p_\Phi^-\in(1,\infty)$ is also an N-function and an Orlicz function $\Phi$ of lower type $p_\Phi^-=1$ is also a Young function. 2. Let $q\in[1,\infty)$ and $\Phi$ be a Young function with lower type $p_{\Phi}^-\in[1,\infty)$ and positive upper type $p_{\Phi}^+$. We know that $L^\Phi(\rn)$ is a Banach space (see [@mmz p.67, Theorem 10]). Then it is easy to prove that $(E_\Phi^q)_t(\rn)$ is also a Banach space. The following two lemmas come from [@mmz p.13, Proposition 1(ii); p.58, Proposition 1], respectively. \[un\] Let $\Phi$ be an N-function and $\Psi$ its complementary function. Then $\Phi$ and $\Psi$ are strictly increasing and hence their inverses $\Phi^{-1}$ and $\Psi^{-1}$ are uniquely defined and, for any $t\in(0,\infty)$, $$t<\Phi^{-1}(t)\Psi^{-1}(t)<2t.$$ \[hold\] Let $\Phi$ be a Young function and $\Psi$ its complementary function. If $f\in L^\Phi(\rn)$ and $g\in L^\Psi(\rn)$, then $$\int_{\rn}\lf\|f(x)g(x)\r\|dx\le2\\|f\\|_{L^\Phi(\rn)}\\|g\\|_{L^\Psi(\rn)}.$$ The following Fefferman-Stein type inequality for Orlicz-slice spaces extends the well-known Fefferman-Stein vector-valued maximal inequality [@fs Theorem 1(1)] , which plays an important role in the succeeding sections. \[main\] Let $t\in(0,\infty)$, $q,\ r\in(1,\infty)$ and $\Phi$ be an Orlicz function with lower type $p_{\Phi}^-\in(1,\infty)$ and positive upper type $p_{\Phi}^+$. Then there exists a positive constant $C$ such that, for any $\{f_j\}_{j\in\mathbb{Z}}\subset (E_\Phi^q)_t(\rn)$, $$\left\\|\left\{\sum _{j\in\mathbb{Z}}\left[\mathcal{M}(f_j)\right]^r\right\} ^{\frac{1}{r}}\right\\|_{(E_\Phi^q)_t(\rn)} \le C\left\\|\left\{\sum _{j\in\mathbb{Z}}\|f_j\|^r\right\}^{\frac{1}{r}}\right\\|_{(E_\Phi^q)_t(\rn)},$$ where $C$ is independent of $\{f_j\}_{j\in\mathbb{Z}}$ and $t$. For any $\{f_j\}_{j\in\mathbb{Z}}\subset (E_\Phi^q)_t(\rn)$, applying Lemma \[mm\], we have $$\begin{aligned} &\int_{\rn} \lf\{\frac{\\|[\sum _{j\in\mathbb{Z}}[\mathcal{M}(f_j)]^r]^{\frac{1}{r}}\chi_{B(x,t)}\\|_{L^\Phi(\rn)}} {\\|\chi_{B(x,t)}\\|_{L^\Phi(\rn)}}\r\}^q\,dx\\ &\quad\lesssim \int_{\rn}\lf\{\frac{\\|\{\sum_{j\in\mathbb{Z}}\|f_j\|^r\}^{\frac{1}{r}} \chi_{B(x,2t)}\\|_{L^\Phi(\rn)}}{\\|\chi_{B(x,t)}\\|_{L^\Phi(\rn)}}\r\}^q\, dx +\int_{\rn}\lf\{\sum_{j\in\mathbb{Z}}\lf[\mathcal{M}_u\lf(\dashint_{B(\cdot,t)}\|f_j(z)\|dz\r)(x)\r]^r\r\}^{\frac{q}{r}}dx \\&\quad=:\mathrm{I}+\mathrm{II}.\end{aligned}$$ Since both $\overline{B}(\vec{0}_n,2t)$ and $\overline{B}(\vec{0}_n,t)$ are compact subsets of $\rn$ with nonempty interiors, it follows that there exist $N\in\mathbb{N}$ and $\{x_1,\ \ldots,\ x_N\}\subset\rn$, independent of $t$, such that $N\lesssim1$ and $B(\vec{0}_n,2t)\subseteq\bigcup_{m=1}^NB(x_m,t)$. Thus, by this, (\[qiu\]) and the translation invariance of the Lebesgue measure, we conclude that $$\begin{aligned} \mathrm{I}&\sim\lf[\widetilde{C}_{(\Phi,t)}\r]^q\int_{\rn}\lf\\|\lf\{\sum_{j\in\mathbb{Z}}\lf\|f_j\r\|^r \r\}^{\frac{1}{r}}\chi_{B(x,2t)}\r\\|_{L^\Phi(\rn)}^q\, dx\\ &\lesssim\lf[\widetilde{C}_{(\Phi,t)}\r]^q\int_{\rn}\lf\\| \lf\{\sum_{j\in\mathbb{Z}}\lf\|f_j\r\|^r\r\}^{\frac{1}{r}}\sum_{m=1}^N\chi_{B(x+x_m,t)}\r\\|_{L^\Phi(\rn)}^q\, dx\\ &\lesssim\lf[\widetilde{C}_{(\Phi,t)}\r]^q\sum_{m=1}^N\int_{\rn} \lf\\|\lf\{\sum_{j\in\mathbb{Z}}\lf\|f_j\r\|^r\r\}^{\frac{1}{r}}\chi_{B(x+x_m,t)}\r\\|_{L^\Phi(\rn)}^q\, dx\\ &\lesssim\lf[\widetilde{C}_{(\Phi,t)}\r]^q\int_{\rn}\lf\\|\lf\{\sum_{j\in\mathbb{Z}} \lf\|f_j\r\|^r\r\}^{\frac{1}{r}}\chi_{B(x,t)}\r\\|_{L^\Phi(\rn)}^q\, dx\sim \lf\\|\lf\{\sum _{j\in\mathbb{Z}}\lf\|f_j\r\|^r\r\}^{\frac{1}{r}}\r\\|_{(E_\Phi^q)_t(\rn)}^q,\end{aligned}$$ where $\widetilde{C}_{(\Phi,t)}$ is as in . As for $\mathrm{II}$, by the Fefferman-Stein vector-valued inequality in $L^q(\rn)$ (see [@fs]), we have $$\mathrm{II} \lesssim\int_{\rn}\lf\{\sum_{j\in\mathbb{Z}}\lf[\dashint_{B(x,t)}\|f_j(z)\|\,dz\r]^r\r\}^{\frac{q}{r}}dx.$$ Let $r':=\frac{r}{r-1}.$ Then there exists $\{b_j\}_{j\in\mathbb{Z}}\in \ell^{r'}$, with $\\|\{b_j\}_{j\in\mathbb{Z}}\\|_{\ell^{r'}}=1,$ such that $$\int_{\rn}\lf\{\sum_{j\in\mathbb{Z}}\lf[\dashint_{B(x,t)}\|f_j(z)\|\,dz\r]^r\r\}^{\frac{q}{r}}dx =\int_{\rn}\lf[\sum_{j\in\mathbb{Z}}b_j\dashint_{B(x,t)}\|f_j(z)\|\,dz\r]^{q}\,dx.$$ Using Lemma \[hold\] and the Hölder inequality, we further conclude that $$\begin{aligned} &\int_{\rn}\lf[\sum_{j\in\mathbb{Z}}b_j\dashint_{B(x,t)}\|f_j(z)\|\,dz\r]^{q}\,dx\\ &\quad\lesssim\int_{\rn}\lf\{\dashint_{B(x,t)}\lf[\sum_{j\in\mathbb{Z}}\|f_j(z)\|^r\r]^{\frac{1}{r}} \lf(\sum_{j\in\mathbb{Z}}b_j^{r'}\r)^{\frac{1}{r'}}\,dz\r\}^{q}\,dx\\ &\quad\lesssim\int_{\rn}\lf\{\lf\\|\lf[\sum_{j\in\mathbb{Z}}\lf\|f_j\r\|^r\r]^{\frac{1}{r}} \chi_{B(x,t)}\r\\|_{L^\Phi(\rn)}\frac{\\|\chi_{B(x,t)}\\|_{L^\Psi(\rn)}}{\|B(x,t)\|}\r\}^q\, dx.\end{aligned}$$ Applying Lemma \[un\], we obtain $$\begin{aligned} \frac{\\|\chi_{B(x,t)}\\|_{L^\Psi(\rn)}}{\|B(x,t)\|} &=\frac{1}{\|B(x,t)\|\Psi^{-1}(\frac{1}{\|B(x,t)\|})} =\frac{\Phi^{-1}(\frac{1}{\|B(x,t)\|})}{\|B(x,t)\|\Phi^{-1}(\frac{1}{\|B(x,t)\|})\Psi^{-1}(\frac{1}{\|B(x,t)\|})}\\ &<\Phi^{-1}\lf(\frac{1}{\|B(x,t)\|}\r)=\frac{1}{\\|\chi_{B(x,t)}\\|_{L^\Phi(\rn)}}.\end{aligned}$$ Thus, $$\mathrm{II}\lesssim\lf\\|\lf\{\sum _{j\in\mathbb{Z}}\|f_j\|^r\r\}^{\frac{1}{r}}\r\\|_{(E_\Phi^q)_t(\rn)}^q,$$ which completes the proof of Theorem \[main\]. Let $t\in(0,\infty),$ $q\in(1,\infty)$ and $\Phi(\tau):=\tau^q$ for any $\tau\in[0,\infty).$ Then, by Remark \[re2.10\](ii) and Proposition \[ggg\](iii), we know that $(E_\Phi^q)_t(\rn)=L^q(\rn)$ and, in this case, Theorem \[main\] is just the well-known Fefferman-Stein vector-valued maximal inequality [@fs Theorem 1(1)]. As an immediate consequence of Theorem \[main\], we have the following boundedness of Hardy-Littlewood maximal operators on Orlicz-slice spaces. \[main2\] Let $t\in(0,\infty)$, $q\in(1,\infty)$ and $\Phi$ be an Orlicz function with positive lower type $p_{\Phi}^-\in(1,\infty)$ and positive upper type $p_{\Phi}^+$. Then the central Hardy-Littlewood maximal function $\mathcal{M}$ is bounded on the Orlicz-slice space $(E_\Phi^q)_t(\rn)$ with the operator norm independent of $t$. Let $t\in(0,\infty)$ and $q,\ r\in(1,\infty).$ Recall that Auscher and Prisuelos-Arribas [@ap Proposition 8.3(a)] obtained the boundedness of the Hardy-Littlewood maximal operator $\mathcal{M}$ on the space $(E_r^q)_t(\rn).$ It is easy to see that, if $\Phi(\tau):=\tau^r$ for any $\tau\in[0,\infty)$, then $(E_\Phi^q)_t(\rn)=(E_r^q)_t(\rn)$ and, in this case, Proposition \[main2\] is just [@ap Proposition 8.3(a)]. Thus, Proposition \[main2\] essentially generalizes [@ap Proposition 8.3(a)]. Let $q\in(1,\infty)$ and $\{E_k\}_{k\in\mathbb{N}}$ be a sequence of Banach spaces. The amalgam space $\ell^q(\{E_k\}_{k\in\mathbb{N}})$ is defined to be set of all sequences $x: = \{x_k\}_{k\in\mathbb{N}}$ such that $$\\|x\\|_{\ell^q(\{E_k\}_{k\in\mathbb{N}})}:=\lf[\sum_{k=1}^\infty\\|x_k\\|_{E_k}^q\r]^{\frac{1}{q}}<\infty.$$ The following lemma comes from [@kg p.359]. \[dou\] Let $q\in(1,\infty)$. Then the space $\ell^q(\{E_k\}_{k\in\mathbb{N}})$ is a Banach space and its dual space is $\ell^{q'}(\{(E_k)^\}_{k\in\mathbb{N}})$, where $\frac{1}{q}+\frac{1}{q'}=1$ and $(E_k)^$ denotes the dual space of $E_k$. \[dual\] Let $t\in(0,\infty)$, $q\in(1,\infty)$ and $\Phi$ be an Orlicz function with lower type $p_{\Phi}^-\in(1,\infty)$ and positive upper type $p_{\Phi}^+$. Let $\Psi$ be the complementary function of $\Phi$. Then the dual space of $(E_\Phi^q)_t(\rn)$ is isomorphic and homeomorphic to $(E_\Psi^{q'})_t(\rn)$. Let $t\in(0,\infty)$. Using Proposition \[th2\], we obtain $$(E_\Phi^q)_t(\rn)=\ell^q(\{L^\Phi(Q_{tk})\}_{k\in\mathbb{Z}^n})$$ and, for any $f\in(E_\Phi^q)_t(\rn)$, $$\\|f\\|_{(E_\Phi^q)_t(\rn)}\sim\frac{t^{\frac{n}{q}}}{[\Phi^{-1}(\frac{1}{\varepsilon_nt^n})]^{-1}} \\|f\\|_{\ell^q(\{L^\Phi(Q_{tk})\}_{k\in\mathbb{Z}^n})}.$$ Then it is easy to prove that $((E_\Phi^q)_t(\rn))^=(\ell^q(\{L^\Phi(Q_{tk})\}_{k\in\mathbb{Z}^n}))^$ and, for any $f\in((E_\Phi^q)_t(\rn))^$, $$\label{11} \frac{t^{\frac{n}{q}}}{[\Phi^{-1}(\frac{1}{\varepsilon_nt^n})]^{-1}} \\|f\\|_{((E_\Phi^q)_t(\rn))^}\sim\\|f\\|_{(\ell^q(\{L^\Phi(Q_{tk})\}_{k\in\mathbb{Z}^n}))^},$$ where the equivalent positive constants are independent of $t$ and $f$. Applying [@mmz p.110, Theorem 7] or [@af Theorem 8.19], we know that $(L^\Phi(Q_{tk}))^$ is isomorphic and homeomorphic to $L^\Psi(Q_{tk})$, which, together with Lemma \[dou\], implies that $(\ell^q(\{L^\Phi(Q_{tk})\}_{k\in\mathbb{Z}^n}))^* =\ell^{q'}(\{L^\Psi(Q_{tk})\}_{k\in\mathbb{Z}^n})$ and, for any $f\in(\ell^q(\{L^\Phi(Q_{tk})\}_{k\in\mathbb{Z}^n}))^$, $$\label{22} \\|f\\|_{(\ell^q(\{L^\Phi(Q_{tk})\}_{k\in\mathbb{Z}^n}))^} \sim\\|f\\|_{\ell^{q'}(\{L^\Psi(Q_{tk})\}_{k\in\mathbb{Z}^n})},$$ where the equivalent positive constants are independent of $f$ and $t$. Using Proposition \[th2\] again, we find that, for any $f\in \ell^{q'}(\{L^\Psi(Q_{tk})\}_{k\in\mathbb{Z}^n})$, $$\\|f\\|_{\ell^{q'}(\{L^\Psi(Q_{tk})\}_{k\in\mathbb{Z}^n})}\sim\frac{[\Psi^{-1} (\frac{1}{\varepsilon_nt^n})]^{-1}}{t^{\frac{n}{q'}}} \\|f\\|_{(E_\Psi^{q'})_t(\rn)},$$ which, combined with (\[11\]) and (\[22\]), implies that $((E_\Phi^q)_t(\rn))^=(E_\Psi^{q'})_t(\rn)$ and $$\frac{t^n}{[\Phi^{-1}(\frac{1}{\varepsilon_nt^n})]^{-1}[\Psi^{-1} (\frac{1}{\varepsilon_nt^n})]^{-1}}\\|f\\|_{((E_\Phi^q)_t)(\rn)^} \sim\\|f\\|_{(E_\Psi^{q'})_t(\rn)},$$ where all equivalent positive constants are independent of $f$ and $t$. By Lemma \[un\], we have $$\frac{t^n}{[\Phi^{-1}(\frac{1}{\varepsilon_nt^n})]^{-1}[\Psi^{-1} (\frac{1}{\varepsilon_nt^n})]^{-1}} \sim 1$$ with the equivalent positive constants independent of $f$, which further implies the desired conclusion and hence completes the proof of Theorem \[dual\]. Next, we recall the notion of ball quasi-Banach function spaces defined in [@ykds Definition 2.1]. In what follows, the symbol $\mathbb{M}(\rn)$ denotes the set of all measurable functions on $\rn$. \[ball\] A quasi-Banach space $X\subset\mathbb{M}(\rn)$ is called a ball quasi-Banach function space on $\rn$ if it satisfies 1. $\\|f\\|_{X}=0$ implies that $f=0$ almost everywhere; 2. $\|g\|\le\|f\|$ almost everywhere implies that $\\|g\\|_{X}\le\\|f\\|_{X}$; 3. $0\le f_m\uparrow f$ almost everywhere on $\rn$ implies that $\\|f_m\\|_{X}\uparrow\\|f\\|_{X}$; 4. $B\in\mathbb{B}$ implies that $\chi_B\in X$, where $$\mathbb{B}:=\{B(x,r):\ x\in\rn\ \ \mbox{and}\ \ r\in(0,\infty)\}.$$ Recall that Sawano et al. [@ykds] developed a real-variable theory of Hardy spaces associated with ball quasi-Banach function spaces. Next we show that the Orlicz-slice spaces are ball quasi-Banach function spaces, which further implies that the Orlicz-slice Hardy space is a special case of the Hardy type space considered in [@ykds]. \[ballproof\] Let $t,\ q\in(0,\infty)$ and $\Phi$ be an Orlicz function with positive lower type $p_{\Phi}^-$ and positive upper type $p_{\Phi}^+$. Then $(E_\Phi^q)_t(\rn)$ is a ball quasi-Banach function space. By Remark \[re\], without loss of generality, we may assume that $\Phi$ is continuous and strictly increasing. Then, from the definition of $\\|\cdot\\|_{(E_\Phi^q)_t(\rn)}$, it is easy to deduce that $(E_\Phi^q)_t(\rn)$ satisfies (i), (ii) and (iv) of Definition \[ball\]. We now prove that $(E_\Phi^q)_t(\rn)$ satisfies Definition \[ball\](iii). To this end, let $\{f_m\}_{m\in\mathbb{N}}\subset(E_\Phi^q)_t(\rn)$ and $f\in(E_\Phi^q)_t(\rn)$ satisfy $0\le f_m\uparrow f$ almost everywhere on $\rn$. For any fixed $x\in\rn$ and $t\in(0,\infty)$, let $A_{(x,t)}\in(0,\\|f\chi_{B(x,t)}\\|_{L^\Phi(\rn)})$. Then, by the definition of $\\|f\chi_{B(x,t)}\\|_{L^\Phi(\rn)}$, we have $$\int_{B(x,t)}\Phi\lf(\frac{\|f(y)\|}{A_{(x,t)}}\r)\,dy>1,$$ which, together with the monotone convergence theorem, implies that there exists $N\in\mathbb{N}$ such that $$\int_{B(x,t)}\Phi\lf(\frac{\|f_K(y)\|}{A_{(x,t)}}\r)\,dy>1,\quad \forall\,K\ge N.$$ Thus, when $K\ge N$, $\\|f_K\chi_{B(x,t)}\\|_{L^\Phi(\rn)}>A_{(x,t)}$, which, together with the arbitrariness of $A_{(x,t)}\in(0,\\|f\chi_{B(x,t)}\\|_{L^\Phi(\rn)})$, implies that, for any $x\in\rn$, $\lim_{m\rightarrow\infty}\\|f_m\chi_{B(x,t)}\\|_{L^\Phi(\rn)}=\\|f\chi_{B(x,t)}\\|_{L^\Phi(\rn)}$. Then, by the monotone convergence theorem in $L^q(\rn)$, we obtain $$\lim_{m\rightarrow\infty}\\|f_m\\|_{(E_\Phi^q)_t(\rn)}=\\|f\\|_{(E_\Phi^q)_t(\rn)}.$$ This finishes the proof of Lemma \[ballproof\]. \[lian\] A ball quasi-Banach function space $X$ is said to have an absolutely continuous quasi-norm if $\\|\chi_{E_j}\\|_{X}\downarrow0$ as $j\rightarrow\infty$ whenever $\{E_j\}_{j=1}^\infty$ is a sequence of measurable sets in $\rn$ satisfying that $E_j\supset E_{j+1}$ for any $j\in\mathbb{N}$ and $\cap_{j=1}^\infty E_j=\emptyset$. \[convex\] Let $X$ be a ball quasi-Banach function space and $p\in(0,\infty)$. 1. The $p$-convexification $X^p$ of $X$ is defined by setting $X^p:=\{f\in\mathbb{M}(\rn):\ \|f\|^p\in X\}$ equipped with the quasi-norm $\\|f\\|_{X^p}:=\\|\|f\|^p\\|_{X^p}^{\frac{1}{p}}$ for any $f\in X^p$. 2. The space $X$ is said to be p-convex if there exists a positive constant $C$ such that, for any $\{f_j\}_{j\in\mathbb{N}}\subset X^{\frac{1}{p}}$, $$\lf\\|\sum_{j=1}^{\infty}\lf\|f_j\r\|\r\\|_{X^{\frac{1}{p}}}\le C\sum_{j=1}^{\infty}\lf\\|f_j\r\\|_{X^{\frac{1}{p}}}.$$ In particular, when $C=1$, $X$ is said to be strictly p-convex. \[littlewood\] Let $t$, $q\in(0,\infty)$ and $\Phi$ be an Orlicz function with positive lower type $p_{\Phi}^-$ and positive upper type $p_{\Phi}^+$. Let $r\in(0,\min\{p_\Phi^-,q\})$. Then $\mathcal{M}$ is bounded on $[(E_\Phi^q)_t(\rn)]^{\frac{1}{r}}$ with the operator norm independent of $t$, where $[(E_\Phi^q)_t(\rn)]^{\frac{1}{r}}$ is the $\frac{1}{r}$-convexification of $(E_\Phi^q)_t(\rn)$. For any $\tau\in(0,\infty)$, let $\Phi_r(\tau):=\Phi(\sqrt[r]{\tau})$. Then $\Phi_r$ is of upper type $\frac{p_{\Phi}^+}{r}$ and of lower type $\frac{p_{\Phi}^-}{r}$, and $\frac{p_{\Phi}^-}{r}\in(1,\infty)$. This implies that, for any $t\in(0,\infty)$, $f\in[(E_\Phi^q)_t(\rn)]^{\frac{1}{r}}$ and $x\in\rn$, $$\lf\\|\|f\|^{\frac{1}{r}}\chi_{B(x,t)}\r\\|_{L^\Phi(\rn)}=\lf\\|\|f\|\chi_{B(x,t)}\r\\|_{L^{\Phi_r}(\rn)}^{\frac{1}{r}}$$ and $$\lf\\|\chi_{B(x,t)}\r\\|_{L^\Phi(\rn)}=\lf\\|\lf[\chi_{B(x,t)}\r]^{\frac{1}{r}}\r\\|_{L^\Phi(\rn)} =\lf\\|\chi_{B(x,t)}\r\\|_{L^{\Phi_r}(\rn)}^{\frac{1}{r}}.$$ Combining this and Definition \[convex\], we obtain $$\begin{aligned} \lf\\|\mathcal{M}(f)\r\\|_{[(E_\Phi^q)_t(\rn)]^{\frac{1}{r}}} &=\lf\{\int_{\rn}\lf[\frac{\\|\|\mathcal{M}(f)\|^{\frac{1}{r}}\chi_{B(x,t)}\\| _{L^\Phi(\rn)}}{\\|\chi_{B(x,t)}\\|_{L^\Phi(\rn)}}\r]^q\,dx\r\}^{\frac{r}{q}}\\ &=\lf\{\int_{\rn}\lf[\frac{\\|\|\mathcal{M}(f)\|\chi_{B(x,t)}\\| _{L^{\Phi_r}(\rn)}}{\\|\chi_{B(x,t)}\\|_{L^{\Phi_r}(\rn)}}\r]^{\frac{q}{r}}\,dx\r\}^{\frac{r}{q}} =\lf\\|\mathcal{M}(f)\r\\|_{(E_{\Phi_r}^{q/r})_t(\rn)}.\end{aligned}$$ Since $\frac{p_{\Phi}^-}{r}\in(1,\infty)$ and $\frac{q}{r}\in(1,\infty)$, from Proposition \[main2\], it follows that $$\lf\\|\mathcal{M}(f)\r\\|_{(E_{\Phi_r}^{q/r})_t(\rn)} \lesssim\lf\\|f\r\\|_{(E_{\Phi_r}^{q/r})_t(\rn)} \sim\lf\\|f\r\\|_{[(E_\Phi^q)_t(\rn)]^{\frac{1}{r}}}.$$ This finishes the proof of Lemma \[littlewood\]. Orlicz-slice Hardy spaces \[s3\] ================================ In this section, we introduce the Orlicz-slice Hardy spaces, which are defined via the radial maximal functions. We then present a series of real-variable characterizations of these Orlicz-slice Hardy spaces, including characterizations via grand and non-tangential maximal functions, poisson integrals, atoms and finite atoms, and Littlewood-Paley functions. A Lebesgue-Hardy type coincidence relation is also established between Orlicz-slice spaces and Orlicz-slice Hardy spaces. Let us begin with the following notion of the radial maximal function. \[31\] Let $\varphi\in\mathcal{S}(\rn)$ and $f\in\mathcal{S}'(\rn)$. The radial maximal function $M(f,\varphi)$ is defined by setting $$M(f,\varphi)(x):=\sup_{s\in(0,\infty)}\lf\|(\varphi_s \ast f)(x)\r\|,\quad \forall\,x\in\rn.$$ \[dh\] Let $t$, $q\in(0,\infty) $ and $\Phi$ be an Orlicz function with positive lower type $p_{\Phi}^-$ and positive upper type $p_{\Phi}^+$. Then the Orlicz-slice Hardy space $(HE_\Phi^q)_t(\rn)$ is defined by setting $$(HE_\Phi^q)_t(\rn):=\lf\{f\in\mathcal{S}'(\rn):\ \\|f\\|_{(HE_\Phi^q)_t(\rn)}:= \\|M(f,\varphi)\\|_{(E_\Phi^q)_t(\rn)}<\infty \r\},$$ where $\varphi\in\mathcal{S}(\rn)$ satisfies $ \int_{\rn}\varphi(x)\,dx\neq0. $ In particular, when $\Phi(s):=s^r$ for any $s\in[0,\fz)$ with $r\in(0,\fz)$, the Hardy-type space $(HE_r^q)_t(\rn):=(HE_{\Phi}^q)_t(\rn)$ is called the slice Hardy space. \[ReHpq\] 1. If $t=1$ and $\Phi(\tau):=\tau^p$ for any $\tau\in [0,\fz)$ with $p\in (0,\fz)$, then $(HE_\Phi^q)_t(\rn)$ coincides with the Hardy-amalgam space $\mathcal{H}^{p,q}(\rn)$ in [@af1]. 2. If $t=q=1$, then $(HE_\Phi^q)_t(\rn)$ coincides with the variant of the Orlicz-Hardy space $H_^\Phi(\rn)$ of Bonami and Feuto [@bf]. Characterizations in terms of various maximal functions\[s3.1\] --------------------------------------------------------------- We now present some maximal function characterizations of $(HE_\Phi^q)_t(\rn)$, whose proofs are given in Section \[s4\]. Define, for any $N\in\mathbb{N}$ and $\varphi\in\mathcal{S}(\rn)$, $$p_N(\varphi):=\sum_{\alpha\in\mathbb{Z}^n_+,\|\alpha\|\le N} \sup_{x\in\rn}(1+\|x\|)^{N+n}\|\partial^\alpha\varphi(x)\|,$$ and let $\mathcal{F}_N(\rn):=\{\varphi\in\mathcal{S}(\rn):\ \ p_N(\varphi)\le1\}$. Also recall that $\mathbb{R}_+^{n+1}:=\rn\times(0,\infty)$. \[jdhs\] Let $\varphi\in\mathcal{S}(\rn)$, $N\in\mathbb{N}$, $a,\ b\in(0,\infty)$ and $f\in\mathcal{S}'(\rn)$. 1. The grand maximal function $M_N(f)$* is defined by setting, for any $x\in\rn$, $$M_N(f)(x):=\sup\lf\{\|\varphi_s \ast f(y)\|:\ s\in(0,\infty),\ \|x-y\|<s,\ \varphi\in\mathcal{F}_N(\rn)\r\};$$ 2. The grand radial maximal function $M_N^0(f)$ is defined by setting, for any $x\in\rn$, $$M_N^0(f)(x):=\sup\lf\{\|\varphi_s \ast f(x)\|:\ s\in(0,\infty),\ \varphi\in\mathcal{F}_N(\rn)\r\};$$ 3. The non-tangential maximal function $M_a^(f,\varphi)$, with aperture $a\in(0,\infty)$, is defined by setting, for any $x\in\rn$, $$M_a^(f,\varphi)(x):=\sup_{s\in(0,\infty)}\lf\{\sup_{y\in\rn,\|y-x\|<as}\lf\|(\varphi_s \ast f)(y)\r\|\r\};$$ 4. The maximal function $M_b^{}(f,\varphi)$ of Peetre type is defined by setting, for any $x\in\rn$, $$M_b^{*}(f,\varphi)(x):=\sup_{(y,s)\in\mathbb{R}_+^{n+1}}\frac{\|(\varphi_s \ast f)(x-y)\|}{(1+s^{-1}\|y\|)^b};$$ 5. The grand maximal function $M(f,\varphi)$ of Peetre type* is defined by setting, for any $x\in\rn$, $$M_{b,\ N}^{*}(f)(x):=\sup_{\psi\in\mathcal{F}_N(\rn)}\lf\{\sup_{(y,s)\in\mathbb{R}_+^{n+1}}\frac{\|(\psi_s \ast f)(x-y)\|}{(1+s^{-1}\|y\|)^b}\r\}.$$ It is easy to see that, for any $N\in\mathbb{Z}_+$, there exists a positive constant $C_{(N)}$, depending on $N$, such that, for any $f\in\mathcal{S}'(\rn)$ and $x\in\rn$, $$\label{zhujida} M_N^0(f)(x)\le M_N(f)(x)\le C_{(N)}M_N^0(f)(x);$$ see [@b Proposition 3.10]. Via the above maximal functions, we can characterize $(HE_\Phi^q)_t(\rn)$ as follows. \[mdj\] Let $t,\ a,\ b,\ q\in(0,\infty)$. Let $\Phi$ be an Orlicz function with positive lower type $p_{\Phi}^-$ and positive upper type $p_{\Phi}^+$. Let $\varphi\in\mathcal{S}(\rn)$ satisfy $\int_{\rn}\varphi(x)\,dx\neq0.$ 1. Let $N\ge\lfloor b+1\rfloor$ be an integer. Then, for any $f\in\mathcal{S}'(\rn)$, it holds true that $$\lf\\|M(f,\varphi)\r\\|_{(E_\Phi^q)_t(\rn)}\lesssim\lf\\|M_a^(f,\varphi)\r\\|_{(E_\Phi^q)_t(\rn)} \lesssim\lf\\|M_b^{}(f,\varphi)\r\\|_{(E_\Phi^q)_t(\rn)},$$ $$\lf\\|M(f,\varphi)\r\\|_{(E_\Phi^q)_t(\rn)}\lesssim\lf\\|M_N(f)\r\\|_{(E_\Phi^q)_t(\rn)} \lesssim\lf\\|M_{\lfloor b+2 \rfloor}(f)\r\\|_{(E_\Phi^q)_t(\rn)}\lesssim\lf\\|M_b^{}(f,\varphi)\r\\|_{(E_\Phi^q)_t(\rn)}$$ and $$\lf\\|M_b^{}(f,\varphi)\r\\|_{(E_\Phi^q)_t(\rn)}\sim\lf\\|M_{b,\ N}^{}(f)\r\\|_{(E_\Phi^q)_t(\rn)},$$ where the implicit positive constants are independent of $f$ and $t$. 2. Assume $b\in(\frac{n}{\min\{p_\Phi^-,q\}},\infty)$. Then, for any $f\in\mathcal{S}'(\rn)$, $$\lf\\|M_{b,\ N}^{*}(f)\r\\|_{(E_\Phi^q)_t(\rn)}\lesssim\lf\\|M(f,\varphi)\r\\|_{(E_\Phi^q)_t(\rn)},$$ where the implicit positive constant is independent of $f$ and $t$. In particular, when $N\ge\lfloor b+1\rfloor$, if one of the following quantities $$\lf\\|M(f,\varphi)\r\\|_{(E_\Phi^q)_t(\rn)},\ \lf\\|M_a^(f,\varphi)\r\\|_{(E_\Phi^q)_t(\rn)},\ \lf\\|M_N(f)\r\\|_{(E_\Phi^q)_t(\rn)},$$ $$\lf\\|M_b^{}(f,\varphi)\r\\|_{(E_\Phi^q)_t(\rn)}\quad and \quad \lf\\|M_{b,\ N}^{}(f)\r\\|_{(E_\Phi^q)_t(\rn)}$$ is finite, then the others are also finite and mutually equivalent with the implicit positive constants independent of $f$ and $t$. Characterization in terms of Poisson integrals\[s3.2\] ------------------------------------------------------ In this section, we characterize $(HE_\Phi^q)_t(\rn)$ by means of the Poisson integral. Recall that $f\in\mathcal{S}'(\rn)$ is said to be a bounded tempered distribution if, for any $\varphi\in\mathcal{S}(\rn)$, $\varphi \ast f \in L^{\infty}(\rn)$. Moreover, for any bounded tempered distribution $f$, the Poisson semigroup of $f$ is defined by setting, for any $s\in(0,\infty)$, $$P_sf:=e^{-s\sqrt{-\triangle}}f:=\mathcal{F}^{-1}(e^{-s\|\cdot\|}\mathcal{F}f)$$ (see, for example, [@em p.89] for the details), where $\mathcal{F}$ denotes the Fourier transform. Recall that $\mathcal{F}f$ is defined by setting, for any $\varphi\in\mathcal{S}(\rn)$, $\la\mathcal{F}f,\varphi\ra:=\la f,\mathcal{F}\varphi\ra$, where, for any $\xi:=(\xi_1,...,\xi_n)\in\rn$, $$\mathcal{F}\varphi(\xi):=(2\pi)^{-n/2}\int_{\rn}\varphi(x)e^{-ix\xi}\,dx$$ with $x\xi=\sum_{i=1}^{n}x_i\xi_i$ for any $x:=(x_1,...,x_n)\in\rn$; also, $\mathcal{F}^{-1}$ denotes the inverse Fourier transform which is defined by setting, for any $f\in\mathcal{S}(\rn)$ \[or $\mathcal{S}'(\rn)$\] and $\xi\in\rn$, $\mathcal{F}^{-1}f(\xi):=\mathcal{F}f(-\xi)$. Then we have the following characterization of $(HE_\Phi^q)_t(\rn)$. \[poisson\] Let $t$, $q\in(0,\infty) $ and $\Phi$ be an Orlicz function with positive lower type $p_{\Phi}^-$ and positive upper type $p_{\Phi}^+$. Assume that $f\in\mathcal{S}'(\rn)$. Then $f\in (HE_\Phi^q)_t(\rn)$ if and only if $f$ is a bounded tempered distribution and $\sup_{s\in(0,\infty)}\|P_s\ast f\|\in (E_\Phi^q)_t(\rn)$. \[remax\] Let $t,\ q\in(0,\infty)$ and $\Phi(\tau):=\tau^q$ for any $\tau\in[0,\infty).$ Then, by Remark \[re2.10\](ii) and Proposition \[ggg\](iii), we know that $(E_\Phi^q)_t(\rn)=L^q(\rn)$ and, in this case, $(HE_\Phi^q)_t(\rn)=H^q(\rn),$ where $H^q(\rn)$ denotes the classical Hardy space, and Theorems \[mdj\] and \[poisson\] coincide with the well-known results on $H^q(\rn)$ (see, for example, [@l] or [@g p.60, Theorem 1]). Relations between $(E_\Phi^q)_t(\rn)$ and $(HE_\Phi^q)_t(\rn)$\[s3.3\] ---------------------------------------------------------------------- In this section, we discuss the relation between the spaces $(E_\Phi^q)_t(\rn)$ and $(HE_\Phi^q)_t(\rn)$. More precisely, we generalize the classical result that $H^p(\rn)=L^p(\rn)$ with $p\in(1,\infty)$ as follows. \[dayu1\] Let $t\in(0,\infty)$, $q\in(1,\infty) $ and $\Phi$ be an Orlicz function with lower type $p_{\Phi}^-\in(1,\infty)$ and positive upper type $p_{\Phi}^+$. 1. It holds true that $(E_\Phi^q)_t(\rn)\ \hookrightarrow\ \mathcal{S}'(\rn)$. 2. If $f\in (E_\Phi^q)_t(\rn)$, then $f\in (HE_\Phi^q)_t(\rn)$. 3. If $f\in {(HE_\Phi^q)_t}(\rn)$, then there exists a locally integrable function $g\in (E_\Phi^q)_t(\rn)$ such that $g$ represents $f$, which means that $f=g$ in $\mathcal{S}'(\rn)$ and $\\|f\\|_{(HE_\Phi^q)_t(\rn)}=\\|g\\|_{(HE_\Phi^q)_t(\rn)}$. Atomic and molecular characterizations\[s3.4\] ---------------------------------------------- In this section, we present the atomic and the molecular characterizations of $(HE_\Phi^q)_t(\rn)$. In what follows, for any $L\in\mathbb{Z}_+$, the symbol $\mathcal{P}_L(\rn)$ denotes the set of all polynomials on $\rn$ of degree not greater than $L$. For any $a\in L^1(\rn)$ satisfying $$\int_{\rn}(1+\|x\|)^L\|a(x)\|\,dx<\infty,$$ we write $a\bot\mathcal{P}_L(\rn)$ if $$\int_{\rn}a(x)x^\alpha\,dx=0$$ for any $\alpha\in\mathbb{Z}_+^n$ with $\|\alpha\|\le L$. \[deatom\] Let $t,\ q\in(0,\infty)$, $r\in[1,\infty]$ and $d\in\mathbb{Z}_+$. Let $\Phi$ be an Orlicz function with positive lower type $p_{\Phi}^-$ and positive upper type $p_{\Phi}^+$. The function $a$ is called an $((E_\Phi^q)_t(\rn),\ r,\ d)$-atom if there exists a cube $Q\in\mathcal{Q}$ such that $\supp(a)\subset Q$, $$\\|a\\|_{L^r(\rn)}\le\frac{\|Q\|^{\frac{1}{r }}}{\\|\chi_Q\\|_{(E_\Phi^q)_t(\rn)}}$$ and $a\bot\mathcal{P}_d(\rn)$. \[atomic hardy\] Let $t,\ q\in(0,\infty)$ and $\Phi$ be an Orlicz function with positive lower type $p_{\Phi}^-$ and positive upper type $p_{\Phi}^+$. Let $r\in(\max\{1,q,\ p_{\Phi}^+\},\infty]$, $s \in(0,\min\{p^-_{\Phi},q,1\})$ and $d\in\mathbb{Z}_+$ satisfying $d\ge\lfloor n(\frac{1}{s}-1)\rfloor$. The atomic Orlicz-slice Hardy space $(HE_\Phi^q)_t^{r,d}(\rn)$ is defined to be the set of all $f \in\mathcal{S}'(\rn)$ satisfying that there exist a sequence $\{a_j\}_{j=1}^\infty$ of $((E_\Phi^q)_t(\rn),\ r,\ d)$-atoms supported, respectively, on the cubes $\{Q_j\}_{j=1}^\infty\subset\mathcal{Q}$ and a sequence $\{\lambda_j\}_{j=1}^\infty\subset[0,\infty)$ such that $$\label{33} f=\sum_{j=1}^{\infty}\lambda_j a_j\ \ \mathrm{in}\ \mathcal{S}'(\rn)$$ and $$\\|f\\|_{(HE_\Phi^q)_t^{r,d}(\rn)}:=\inf \lf\\|\lf\{\sum_{j=1}^{\infty} \lf[\frac{\lambda_j} {\\|\chi_{Q_j}\\|_{(E_\Phi^q)_t(\rn)}} \r]^s\chi_{Q_j} \r\} ^{\frac{1}{s}}\r\\|_{(E_\Phi^q)_t(\rn)}<\infty,$$ where the infimum is taken over all decompositions of $f$ as above. We have the following atomic characterization of $(HE_\Phi^q)_t(\rn)$. \[atom ch\] Let all assumptions be as in Definition \[atomic hardy\]. Then $(HE_\Phi^q)_t(\rn)=(HE_\Phi^q)_t^{r,d}(\rn)$ with equivalent quasi-norms. \[demole\] Let $t,\ q\in(0,\infty)$ and $\Phi$ be an Orlicz function with positive lower type $p_{\Phi}^-$ and positive upper type $p_{\Phi}^+$. Let $r\in[1,\infty]$, $d\in\mathbb{Z}_+$ and $\tau\in(0,\infty)$. A measurable function $m$ on $\rn$ is called an $((E_\Phi^q)_t(\rn),\ r,\ d,\ \tau)$-molecule centered at a cube $Q\in\mathcal{Q}$ if, for any $j\in\mathbb{Z}_+$, $$\lf\\|\chi_{S_j(Q)}m\r\\|_{L^r(\rn)}\le2^{-\tau j}\frac{\|Q\|^{\frac{1}{r}}}{\\|\chi_Q\\|_{(E_\Phi^q)_t(\rn)}}$$ and $a\bot\mathcal{P}_d(\rn)$. In analogy, one defines an $((E_\Phi^q)_t(\rn),\ r,\ d,\ \tau)$-molecule centered at a ball $B$. \[molecular\] Let $t,\ q\in(0,\infty)$ and $\Phi$ be an Orlicz function with positive lower type $p_{\Phi}^-$ and positive upper type $p_{\Phi}^+$. Let $r\in(\max\{1,q,\ p_{\Phi}^+\},\infty]$ and $s \in(0,\min\{p^-_{\Phi},q,1\})$. Assume that $d\in\mathbb{Z}_+$ satisfies $d\ge\lfloor n(\frac{1}{s}-1)\rfloor$ and $\tau\in(0,\infty)$ satisfies $\tau>n(\frac{1}{s}-\frac{1}{r}).$ Then $f\in (HE_\Phi^q)_t(\rn)$ if and only if there exist a sequence $\{m_j\}_{j=1}^\infty$ of $((E_\Phi^q)_t(\rn),\ r,\ d,\ \tau)$-molecules centered, respectively, at the cubes $\{Q_j\}_{j=1}^\infty\subset\mathcal{Q}$ and $\{\lambda_j\}_{j=1}^\infty\subset[0,\infty)$ satisfying $$\lf\\|\lf\{\sum_{j=1}^{\infty} \lf[\frac{\lambda_j}{\\|\chi_{Q_j}\\|_{(E_\Phi^q)_t(\rn)}}\r]^s\chi_{Q_j} \r\} ^{\frac{1}{s}}\r\\|_{(E_\Phi^q)_t(\rn)}<\infty$$ such that $$f=\sum_{j=1}^{\infty}\lambda_j m_j\ \ in\ \mathcal{S}'(\rn).$$ Moreover, $$\\|f\\|_{(HE_\Phi^q)_t(\rn)}\sim\inf\lf\\|\lf\{\sum_{j=1}^{\infty} \lf[\frac{\lambda_j} {\\|\chi_{Q_j}\\|_{(E_\Phi^q)_t(\rn)}} \r]^s\chi_{Q_j} \r\} ^{\frac{1}{s}}\r\\|_{(E_\Phi^q)_t(\rn)},$$ where the infimum is taken over all decompositions of $f$ as above and the equivalent positive constants are independent of $f$ and $t$. \[reatom\] - Let $t,\ q$ and $\Phi$ be as in Remark \[remax\]. In this case, we have $(E_\Phi^q)_t(\rn)=L^q(\rn)$ and $(HE_\Phi^q)_t(\rn)=H^q(\rn)$ and, for any $\tau\in(0,\infty),\ r\in[1,\infty]$ and $d\in\zz_+,$ any $((E_\Phi^q)_t(\rn),r,d)$-atom from Definition \[deatom\] and any $((E_\Phi^q)_t(\rn),\ r,\ d,\ \tau)$-molecule from Definition \[demole\] just become, respectively, a well-known classical atom (see, for example, [@l Definition 1.1] or [@em p.112]) and a well-known classical molecule (see, for example, [@hyz Definition 1.2] with $X=\rn$). - Let $t,\ q$ and $\Phi$ be as in Remark \[remax\]. In this case, when $r\in[1,\infty]\cap(q,\fz]$ and $s=q$, then Theorem \[atom ch\] coincides with the classical atomic characterization of $H^q(\rn)$ (see, for example, [@l p.34, Theorem 3.1] and [@em p.107, Theorem 2]) and Theorem \[molecular\] with the classical molecular characterization of $H^q(\rn)$ (see, for example, [@hyz Theorem 2.2] with $X=\rn$). However, it is still unclear whether or not both Theorems \[atom ch\] and \[molecular\] still hold true when $r=\max\{1,q,\ p_{\Phi}^+\}$ and $s=\min\{p^-_{\Phi},q,1\}$. Observe that the atomic and the molecular characterizations obtained, respectively, in Theorems \[atom ch\] and \[molecular\] are more close, in spirit, to the atomic characterization ([@ns Theorem 4.6]) and the molecular characterization ([@ns Theorem 5.2]) of variable Hardy spaces, respectively. As a corollary of the above theorems, we have the following conclusion. \[shoulian\] Let all the assumptions be as in Definition \[atomic hardy\]. Then 1. $(HE_\Phi^q)_t(\rn)\cap L^\infty(\rn)$ is dense in $(HE_\Phi^q)_t(\rn)$. 2. The summations in converge in $(HE_\Phi^q)_t(\rn)$. Characterizations in terms of Littlewood-Paley functions\[s3.5\] ---------------------------------------------------------------- In this section, we establish various Littlewood-Paley function characterizations of $(HE_\Phi^q)_t(\rn)$. \[cone\] For any $x\in\rn$, let $\Gamma(x):=\{(y,s)\in\mathbb{R}_+^{n+1}:\ \|x-y\|<s\}$, which is called the cone of aperture $1$ with vertex $x\in\rn$. For any $\tau\in(0,\infty)$, $f\in\mathcal{S}'(\rn)$ and $\varphi\in\mathcal{S}(\rn)$, let $$\varphi(\tau D)(f):=\mathcal{F}^{-1}[\varphi(\tau\cdot)\mathcal{F}f].$$ Recall that a distribution $f\in\mathcal{S}'(\rn)$ is said to be vanish weakly at infinity if $\lim_{t\downarrow0}\varphi(tD)(f)=0$ in $\mathcal{S}'(\rn)$ for any $\varphi\in\mathcal{S}(\rn)$. Let $\varphi\in\mathcal{S}(\rn)$ be such that $$\chi_{B(\vec{0}_n,4) \setminus B(\vec{0}_n,2)}\le\varphi\le\chi_{B(\vec{0}_n,8) \setminus B(\vec{0}_n,1)}.$$ For any $f\in\mathcal{S}'(\rn)$, the Littlewood-Paley $g$-function $g(f)$, the Lusin area function $S(f)$ and the Littlewood-Paley $g_\lambda^$-function* $g_\lambda^(f)$, with $\lambda\in(1,\infty)$, of $f$ are defined, respectively, by setting, for any $x\in\rn$, $$g(f)(x):=\lf\{\int_0^{\infty}\|\varphi(\tau D)(f)(x)\|^{2}\,\frac{d\tau}{\tau}\r\}^{ \frac{1}{2}},$$ $$S(f)(x):=\lf\{\int_{\Gamma(x)}\|\varphi(\tau D)(f)(y)\|^{2}\,\frac{dy\,d\tau}{\tau^{n+1}}\r\}^{ \frac{1}{2}}$$ and $$g_\lambda^(f)(x):=\lf\{\int_0^{\infty}\int_{\rn}\lf(\frac{\tau}{\tau+\|x-y\|}\r)^{\lambda n}\|\varphi(\tau D)(f)(y)\|^{2}\,\frac{dy\,d\tau}{\tau^{n+1}}\r\}^{ \frac{1}{2}}.$$ Using these functions, we have the following characterizations. \[lusin\](Lusin area function characterization) Let $t,\ q\in(0,\infty)$ and $\Phi$ be an Orlicz function with positive lower type $p_{\Phi}^-$ and positive upper type $p_{\Phi}^+$. Then $f\in (HE_\Phi^q)_t(\rn)$ if and only if $f$ vanishes weakly at infinity and $$\lf\\|S(f)\r\\|_{(E_\Phi^q)_t(\rn)}<\infty.$$ Moreover, $$\\|f\\|_{(HE_\Phi^q)_t(\rn)}\sim\lf\\|S(f)\r\\|_{(E_\Phi^q)_t(\rn)},$$ where the equivalent positive constants are independent of $f$ and $t$. \[gfunction\](Littlewood-Paley g-function characterization) Let $t,\ q\in(0,\infty)$ and $\Phi$ be an Orlicz function with positive lower type $p_{\Phi}^-$ and positive upper type $p_{\Phi}^+$. Then $f\in (HE_\Phi^q)_t(\rn)$ if and only if $f$ vanishes weakly at infinity and $$\lf\\|g(f)\r\\|_{(E_\Phi^q)_t(\rn)}<\infty.$$ Moreover, $$\\|f\\|_{(HE_\Phi^q)_t(\rn)}\sim \lf\\|g(f)\r\\|_{(E_\Phi^q)_t(\rn)},$$ where the equivalent positive constants are independent of $f$ and $t$. \[glamda\](Littlewood-Paley $g_\lambda^$-function characterization) Let $t,\ q\in(0,\infty)$ and $\Phi$ be an Orlicz function with positive lower type $p_{\Phi}^-$ and positive upper type $p_{\Phi}^+$. Let $\lambda\in (1+\frac{2}{\min\{p_{\Phi}^-,\ q\}},\infty)$. Then $f\in (HE_\Phi^q)_t(\rn)$ if and only if $f$ vanishes weakly at infinity and $$\lf\\|g_\lambda^(f)\r\\|_{(E_\Phi^q)_t(\rn)}<\infty.$$ Moreover, $$\\|f\\|_{(HE_\Phi^q)_t(\rn)}\sim\lf\\|g_\lambda^(f)\r\\|_{(E_\Phi^q)_t(\rn)},$$ where the equivalent positive constants are independent of $f$ and $t$. Let $t,\ q$ and $\Phi$ be as in Remark \[remax\]. In this case, we have $p^{-}_\Phi=q$, $(E_\Phi^q)_t(\rn)=L^q(\rn)$ and $(HE_\Phi^q)_t(\rn)=H^q(\rn)$ and the best known range of $\lambda$ in Theorem \[glamda\] is $\lambda\in(2/q,\infty)$ (see, for example, [@fs82 Corollary 7.4]). However, it is still unclear whether or not Theorem \[glamda\] still holds true when $\lambda\in(\frac{2}{\min\{p^{-}_\Phi,q\}},1+\frac{2}{\min\{p^{-}_\Phi,q\}}]$. Finite atomic characterizations\[s3.6\] --------------------------------------- In this section, we establish a finite atomic decomposition theorem on $(HE_\Phi^q)_t(\rn)$. \[definite\] Let $t,\ q\in(0,\infty)$ and $\Phi$ be an Orlicz function with positive lower type $p_{\Phi}^-$ and positive upper type $p_{\Phi}^+$. Let $r\in(\max\{1,q,\ p_{\Phi}^+\},\infty]$, $s \in(0,\min\{p^-_{\Phi},q,1\})$ and $d\in\mathbb{Z}_+$ satisfying $d\ge\lfloor n(\frac{1}{s}-1)\rfloor$. The finite atomic Orlicz-slice Hardy space* $(HE_\Phi^{q,r,d})_t^{\fin}(\rn)$ is define to be the set of all finite linear combinations of $((E_\Phi^q)_t(\rn),\ r,\ d)$-atoms. The quasi-norm $\\|\cdot\\|_{(HE_\Phi^{q,r,d})_t^{\fin}(\rn)}$ in $(HE_\Phi^{q,r,d})_t^{\fin}(\rn)$ is defined by setting, for any $f\in(HE_\Phi^{q,r,d})_t^{\fin}(\rn)$ $$\begin{aligned} &\\|f\\|_{(HE_\Phi^{q,r,d})_t^{\fin}(\rn)}\\ &\quad :=\inf\lf\{\lf\\|\lf\{\sum_{j=1}^{m} \lf[\frac{\lambda_j} {\\|\chi_{Q_j}\\|_{(E_\Phi^q)_t(\rn)}} \r]^s\chi_{Q_j} \r\} ^{\frac{1}{s}}\r\\|_{(E_\Phi^q)_t(\rn)}:\ m\in\mathbb{N},\ f=\sum_{j=1}^{m}\lambda_j a_j,\ \{\lambda_j\}_{j=1}^m\subset[0,\infty)\r\},\end{aligned}$$ where the infimum is taken over all finite linear combinations of $f$ via $((E_\Phi^q)_t(\rn),\ r,\ d)$-atoms $\{a_j\}_{j=1}^m$ supported, respectively, on cubes $\{Q_j\}_{j=1}^m$. Then we have the following conclusion. In what follows, the symbol $\mathcal{C}{(\rn)}$ is defined to be the set of all continuous complex-valued functions on $\rn.$ \[finite\] Let $t,\ q\in(0,\infty)$ and $\Phi$ be an Orlicz function with positive lower type $p_{\Phi}^-$ and positive upper type $p_{\Phi}^+$. Let $r\in(\max\{1,q,\ p_{\Phi}^+\},\infty]$, $s\in(0,\min\{p^-_{\Phi},q,1\})$ and $d\in\mathbb{Z}_+$ satisfying $d\ge\lfloor n(\frac{1}{s}-1)\rfloor$. 1. If $r\in(\max\{1,q,\ p_{\Phi}^+\},\infty)$, then $\\|\cdot\\|_{(HE_\Phi^q)_t(\rn)}$ and $\\|\cdot\\|_{(HE_\Phi^{q,r,d})_t^{\fin}(\rn)}$ are equivalent on the space $(HE_\Phi^{q,r,d})_t^{\fin}(\rn)$ with the equivalent positive constant independent of $t$. 2. If $r=\fz$, then $\\|\cdot\\|_{(HE_\Phi^q)_t(\rn)}$ and $\\|\cdot\\|_{(HE_\Phi^{q,\infty,d})_t^{\fin}(\rn)}$ are equivalent on $(HE_\Phi^{q,\infty,d})_t^{\fin}(\rn)\cap\mathcal{C}{(\rn)}$ with the equivalent positive constant independent of $t.$ Let $t,\ q,\ r,\ d$ and $\Phi$ be as in Remark \[reatom\](ii). In this case, when $r\in[1,\infty]\cap(q,\fz]$ and $s=q$, Theorem \[finite\] coincides with the classical finite atomic decomposition theorem of $H^q(\rn)$ (see, for example, [@msv Theorem 3.1, Remark 3.3] and [@gly Theorem 5.6] with $X=\rn$). However, it is still unclear whether or not Theorem \[finite\] still holds true when $r=\max\{1,q,\ p_{\Phi}^+\}$ and $s=\min\{p^-_{\Phi},q,1\}$. Further remarks\[s3.7\] ----------------------- Amalgam spaces were first introduced by N. Wiener in 1926. In general, for any $t,\ p,\ q\in(0,\infty)$, the amalgam space $\ell^q(L^p_t)(\rn)=(L^p_t,\ell^q)(\rn)$ is defined by setting $$\ell^q(L^p_t)(\rn):=\lf\{f\ \mbox{measurable}:\ \\|f\\|_{\ell^q(L^p_t)(\rn)}:=\lf[\sum_{k\in \mathbb{Z}^n}\lf\\|f\chi_{Q_{tk}}\r\\|^q_{L^p(\rn)}\r]^{\frac{1}{q}}<\infty\r\}.$$ It is easy to see that the amalgam space $\ell^q(L^p_t)(\rn)$ is a special case of the Orlicz-amalgam space $\ell^q(L^\Phi_t))(\rn)$ in Definition \[d1.3\]. In [@af1], Ablé and Feuto introduced the Hardy type space $\mathcal{H}^{(p,q)}(\rn)$ with $p,\ q\in(0,\infty)$ based on the amalgam space $\ell^q(L^p_1)(\rn)$ and obtained their atomic characterization when $q\in(0,\fz)$ and $p\in(0,\min\{1,q\})$. The atomic characterization obtained in Section \[s3.4\] of this article essentially generalizes [@af1 Theorem 4.4]. In [@bf], Bonami and Feuto introduced the Hardy type spaces $H_^\Phi(\rn)$ with respect to the amalgam space $(L^\Phi,\ell^1)(\rn)=\ell^1(L^\Phi_1)(\rn)$ with $\Phi(t):=\frac{t}{\log(e+t)}$ for any $t\in[0,\infty)$, and applied these spaces to study the linear decomposition of the product of the Hardy space $H^1(\rn)$ and its dual space $\BMO(\rn)$; see also [@cky1]. Since $\ell^1(L^\Phi_1)(\rn)$ is a special case of the Orlicz-amalgam spaces introduced in Definition \[d1.3\], from Proposition \[th2\], we deduce that the space $H_^\Phi(\rn)$ is also a special case of the Orlicz-slice Hardy spaces $(HE_\Phi^{q})_t(\rn)$ considered in this article. Proofs of main results from Section \[s3\]\[s4\] ================================================ In this section, we give the proofs of the results presented in Section \[s3\]. Since Orlicz-slice spaces are ball quasi-Banach function spaces (see Lemma \[ballproof\]), some of these results can be deduced directly from [@ykds], in which a real-variable theory of Hardy spaces related to ball quasi-Banach function spaces was developed. However, some properties and characterizations of $(HE_\Phi^q)_t(\rn)$, such as Littlewood-Paley function and finite atomic characterizations, need independent and detailed proofs. We begin with the proof of Theorem \[mdj\]. By Lemmas \[ballproof\] and \[littlewood\], we know that $(E_\Phi^q)_t(\rn)$ is a ball quasi-Banach function space and, for any $r\in(0,\min\{p_\Phi^-,q\})$, $\mathcal{M}$ is bounded on $[(E_\Phi^q)_t(\rn)]^{\frac{1}{r}}$, where $[(E_\Phi^q)_t(\rn)]^{\frac{1}{r}}$ is the $\frac{1}{r}$-convexification of $(E_\Phi^q)_t(\rn)$ as in Definition \[convex\](i). Thus, all the assumptions of [@ykds Theorem 3.1] are satisfied, which further implies that all the conclusions of Theorem \[mdj\] hold true. This finishes the proof of Theorem \[mdj\]. We point out that, by a carefully checking on the proof of [@ykds Theorem 3.1], we find that $\lfz b+2\rfz$ in [@ykds Theorem 3.1] should be $\lfz b+1\rfz$. By Lemma \[littlewood\], we know that, for any $r\in(\frac{1}{\min\{p_\Phi^-,q\}},\infty)$, $\mathcal{M}$ is bounded on $[(E_\Phi^q)_t(\rn)]^{r}$. Moreover, by , for any $t\in(0,\infty)$ and $z\in\rn$, we have $$\begin{aligned} \lf\\|\chi_{B(z,1)}\r\\|_{(E_\Phi^q)_t(\rn)}&=\lf\{\int_{\rn}\lf[\frac{\\|\chi_{B(z,1)}\chi_{B(x,t)}\\| _{L^\Phi(\rn)}}{\\|\chi_{B(x,t)}\\|_{L^\Phi(\rn)}}\r]^q\,dx\r\}^{\frac{1}{q}} \ge\lf\{\int_{B(z,\frac{1}{2})}\lf[\frac{\\|\chi_{B(z,1)}\chi_{B(x,t)}\\| _{L^\Phi(\rn)}}{\\|\chi_{B(x,t)}\\|_{L^\Phi(\rn)}}\r]^q\,dx\r\}^{\frac{1}{q}}\\ &\ge\lf\{\int_{B(z,\frac{1}{2})}\lf[\frac{\\|\chi_{B(x,\min\{t,1/2\})}\\| _{L^\Phi(\rn)}}{\\|\chi_{B(x,t)}\\|_{L^\Phi(\rn)}}\r]^q\,dx\r\}^{\frac{1}{q}} \sim 1,\end{aligned}$$ which further implies that $\inf_{z\in\rn}\lf\\|\chi_{B(z,1)}\r\\|_{(E_\Phi^q)_t(\rn)}\gs 1$. Thus, all assumptions of [@ykds Theorem 3.3] are satisfied, from which we deduce all the desired conclusions of Theorem \[poisson\]. This finishes the proof of Theorem \[poisson\]. For any $\theta\in(0,\infty)$, the powered Hardy-Littlewood maximal operator $\mathcal{M}^{(\theta)}$ is defined by setting, for any $f\in L^1_{\loc}(\rn)$ and $x\in\rn$, $$\mathcal{M}^{(\theta)}(f)(x):=\lf\{\mathcal{M}\lf(\|f\|^{\theta}\r)(x) \r\}^{\frac{1}{\theta}}.$$ \[sconvex\] Let $t$, $q\in(0,\infty)$ and $\Phi$ be an Orlicz function with positive lower type $p_{\Phi}^-$ and positive upper type $p_{\Phi}^+$. Let $s\in (0,\min\{p^-_{\Phi},q\}]$ . Then $(E_\Phi^q)_t(\rn)$ is a strictly $s$-convex ball quasi-Banach function space as in Definition \[convex\](ii). By Lemma \[ballproof\], we already know that $(E_\Phi^q)_t(\rn)$ is a ball quasi-Banach function space. Now, we show that $(E_\Phi^q)_t(\rn)$ is strictly $s$-convex. To this end, let $s\in(0,\min\{p_{\Phi}^-,q\}]$ and, for any $\tau\in[0,\infty)$, let $\Phi_s(\tau):=\Phi(\sqrt[s]{\tau})$. Then $\Phi_s$ is of upper type $\frac{p_{\Phi}^+}{s}$ and of lower type $\frac{p_{\Phi}^-}{s}$, and $\frac{p_{\Phi}^-}{s}\in[1,\infty)$. Thus, for any $t\in(0,\infty)$, $f\in[(E_\Phi^q)_t(\rn)]^{\frac{1}{s}}$ and $x\in\rn$, we have $$\lf\\|\|f\|^{\frac{1}{s}}\chi_{B(x,t)}\r\\|_{L^\Phi(\rn)}=\lf\\|f\chi_{B(x,t)}\r\\|_{L^{\Phi_s}(\rn)}^{\frac{1}{s}} \quad\mathrm{and}\quad \lf\\|\chi_{B(x,t)}\r\\|_{L^\Phi(\rn)} =\lf\\|\chi_{B(x,t)}\r\\|_{L^{\Phi_s}(\rn)}^{\frac{1}{s}}.$$ By this and Definition \[convex\](i), we know that, for any $t\in(0,\infty)$ and $\{f_j\}_{j=1}^{\infty}\subset[(E_\Phi^q)_t(\rn)]^{\frac{1}{s}}$, $$\begin{aligned} \label{eq} \lf\\|\sum_{j=1}^{\infty}\|f_j\|\r\\|_{[(E_\Phi^q)_t(\rn)]^{\frac{1}{s}}} &=\lf\{\int_{\rn}\lf[\frac{\\|(\sum_{j=1}^{\infty}\|f_j\|)^{\frac{1}{s}}\chi_{B(x,t)}\\| _{L^\Phi(\rn)}}{\\|\chi_{B(x,t)}\\|_{L^\Phi(\rn)}}\r]^q\,dx\r\}^{\frac{s}{q}}\\ \noz &=\lf\{\int_{\rn}\lf[\frac{\\|\sum_{j=1}^{\infty}\|f_j\|\chi_{B(x,t)}\\| _{L^{\Phi_s}(\rn)}}{\\|\chi_{B(x,t)}\\|_{L^{\Phi_s}(\rn)}}\r]^{\frac{q}{s}}\,dx\r\}^{\frac{s}{q}} =\lf\\|\sum_{j=1}^{\infty}\|f_j\|\r\\|_{(E_{\Phi_s}^{q/s})_t(\rn)}.\end{aligned}$$ Since $\frac{p_{\Phi}^-}{s}\in[1,\infty)$ and $\frac{q}{s}\in[1,\infty)$, from Remark \[nfunction\](i), we deduce that $(E_{\Phi_s}^{q/s})_t(\rn)$ is a Banach space, which, together with , further implies that $$\lf\\|\sum_{j=1}^{\infty}\|f_j\|\r\\|_{[(E_\Phi^q)_t(\rn)]^{\frac{1}{s}}} =\lf\\|\sum_{j=1}^{\infty}\|f_j\|\r\\|_{(E_{\Phi_s}^{q/s})_t(\rn)} \le\sum_{j=1}^{\infty}\lf\\|f_j\r\\|_{(E_{\Phi_s}^{q/s})_t(\rn)} = \sum_{j=1}^{\infty}\lf\\|f_j\r\\|_{[(E_\Phi^q)_t(\rn)]^{\frac{1}{s}}}.$$ Thus, $(E_\Phi^q)_t(\rn)$ is strictly $s$-convex, which completes the proof of Lemma \[sconvex\]. \[man28\] Let $t,\ q\in(0,\infty)$ and $\Phi$ be an Orlicz function with positive lower type $p_{\Phi}^-$ and positive upper type $p_{\Phi}^+$. Let $s\in (0,\min\{p^-_{\Phi},q\}]$ . Then, for any $\theta\in(0,s)$, there exists a positive constant $C_{(s,\theta)}$, depending on $\theta$ and $s$, but independent of $t$, such that, for any $\{f_j\}_{j=1}^{\infty}\subset\mathbb{M}(\rn)$, $$\label{28} \lf\\|\lf\{\sum_{j=1}^{\infty}\lf[\mathcal{M}^{(\theta)}\lf(f_j\r)\r]^s\r\}^{\frac{1}{s}}\r\\|_{(E_\Phi^q)_t(\rn)} \le C_{(s,\theta)} \lf\\|\lf\{\sum_{j=1}^{\infty}\lf\|f_j\r\|^s\r\}^{\frac{1}{s}}\r\\|_{(E_\Phi^q)_t(\rn)}.$$ Let $\theta$ and $s$ be as in the lemma. For any $\tau\in[0,\infty)$, let $\Phi_\theta(\tau):=\Phi(\sqrt[\theta]{\tau})$. Then $\Phi_\theta$ is of upper type $\frac{p_{\Phi}^+}{\theta}$ and of lower type $\frac{p_{\Phi}^-}{\theta}$, and $\frac{p_{\Phi}^-}{\theta}$, $\frac{q}{\theta}\in(1,\infty)$. Then, by Definition \[d2\], for any $f\in(E_\Phi^q)_t(\rn)$, we have $$\lf\\|\lf\{\sum _{j=1}^\infty\lf[\mathcal{M}^{(\theta)}\lf(f_j\r)\r]^s\r\}^{\frac{1}{s}}\r\\|_{(E_\Phi^q)_t(\rn)} =\lf\\|\lf\{\sum _{j=1}^\infty\lf[\mathcal{M}\lf(\|f_j\|^{\theta}\r)\r]^{\frac{s}{\theta}}\r\}^{\frac{\theta} {s}\frac{1}{\theta}}\r\\|_{(E_\Phi^q)_t(\rn)} =\lf\\|\lf\{\sum _{j=1}^\infty\lf[\mathcal{M}\lf(\|f_j\|^{\theta}\r)\r]^{\frac{s}{\theta}}\r\}^{\frac{\theta} {s}}\r\\|_{(E_{\Phi_\theta}^{q/\theta})_t(\rn)}^{\frac{1}{\theta}}.$$ From this and Theorem \[main\], it follows that $$\lf\\|\lf\{\sum _{j=1}^\infty\lf[\mathcal{M}^{(\theta)}(f_j)\r]^s\r\}^{\frac{1}{s}}\r\\|_{(E_\Phi^q)_t(\rn)} \lesssim\lf\\|\lf\{\sum _{j=1}^\infty\|f_j\|^s\r\}^{\frac{1}{s}}\r\\|_{(E_\Phi^q)_t(\rn)},$$ namely, holds true, which completes the proof of Lemma \[man28\]. \[man37\] Let $t,\ q\in(0,\infty)$ and $\Phi$ be an Orlicz function with positive lower type $p_{\Phi}^-$ and positive upper type $p_{\Phi}^+$. Let $r\in(\max\{q,\ p_{\Phi}^+\},\infty]$ and $s\in (0,\min\{p^-_{\Phi},q\})$. Then there exists a positive constant $C_{(s,r)}$, depending on $s$ and $r$, but independent of $t$, such that, for any $f\in\mathbb{M}(\rn)$, $$\label{jdjd} \lf\\|\mathcal{M}^{((r/s)')}(f) \r\\|_{([(E_\Phi^q)_t(\rn)]^{1/s})^}\leq C_{(s,r)}\lf\\|f \r\\|_{([(E_\Phi^q)_t(\rn)]^{1/s})^},$$ here and hereafter, $[(E_\Phi^q)_t(\rn)]^{1/s}$ is the $\frac1s$-convexification of $(E_\Phi^q)_t(\rn)$ as in Definition \[convex\](i) and $([(E_\Phi^q)_t(\rn)]^{1/s})^$ denotes its dual space. For any $\tau\in[0,\infty)$, let $\Phi_s(\tau):=\Phi(\sqrt[s]{\tau})$. Then $\Phi_s$ is of upper type $p_{\Phi}^+/s$ and of lower type $p_{\Phi}^-/s$, and $p_{\Phi}^-/s\in(1,\infty)$. As in the proof of Lemma \[littlewood\], we know that, for any $f\in\mathbb{M}(\rn)$, $$\lf\\|f\r\\|_{[(E_\Phi^q)_t(\rn)]^{1/s}}=\lf\\|f\r\\|_{(E_{\Phi_s}^{q/s})_t(\rn)}.$$ From this, Theorem \[dual\] and [@ykds Proposition 7.8], we deduce that $$\lf\\|\mathcal{M}^{((r/s)')}(f) \r\\|_{([(E_\Phi^q)_t(\rn)]^{1/s}(\rn))^}=\lf\\|\mathcal{M}^{((r/s)')}(f) \r\\|_{((E_{\Phi_s}^{q/s})_t(\rn))^}\sim\lf\\|\mathcal{M}^{((r/s)')}(f) \r\\|_{(E_{\Psi_s}^{(q/s)'})_t(\rn)},$$ where $\Psi_s$ is the complementary function to $\Phi_s$ and $\Psi_s$ is of upper type $(p_{\Phi}^-/s)'$ and lower type $(p_{\Phi}^+/s)'.$ Thus, we have $$\lf\\|\mathcal{M}^{((r/s)')}(f) \r\\|_{(E_{\Psi_s}^{(q/s)'})_t(\rn)}= \lf\\|\lf[\mathcal{M}\lf(\|f\|^{(r/s)'}\r)\r]^{\frac{1}{(r/s)'}} \r\\|_{(E_{\Psi_s}^{(q/s)'})_t(\rn)}.$$ Since $(r/s)'\in(0, \min\{(p_{\Phi}^+/s)',(q/s)'\})$, from Lemma \[littlewood\], it follows that $$\lf\\|\lf[\mathcal{M}\lf(\|f\|^{(r/s)'}\r)\r]^{\frac{1}{(r/s)'}} \r\\|_{(E_{\Psi_s}^{(q/s)'})_t(\rn)} \lesssim\lf\\|f \r\\|_{(E_{\Psi_s}^{(q/s)'})_t(\rn)},$$ which further implies (\[jdjd\]) and hence completes the proof of Lemma \[man37\]. \[aboso\] Let $t,\ q\in(0,\infty)$ and $\Phi$ be an Orlicz function with positive lower type $p_{\Phi}^-$ and positive upper type $p_{\Phi}^+$. Then $(E_\Phi^q)_t(\rn)$ has an absolutely continuous quasi-norm as in Definition \[lian\]. Let $\{E_j\}_{j=1}^\infty$ be a sequence of measurable sets that satisfy $E_j\supset E_{j+1}$ for any $j\in\mathbb{N}$ and $\cap_{j=1}^\infty E_j=\emptyset$. By the fact that $$\lf\\|\chi_{E_j}\chi_{B(x,t)}\r\\| _{L^\Phi(\rn)}\lesssim\max\lf\{\lf[\int_{B(x,t)}\Phi\lf(\chi_{E_j}(y)\r)\,dy\r]^{1/p_{\Phi}^+},\ \lf[\int_{B(x,t)}\Phi\lf(\chi_{E_j}(y)\r)\,dy\r]^{1/p_{\Phi}^-} \r\},$$ we have $\lim_{j\rightarrow\infty}\\|\chi_{E_j}\chi_{B(x,t)}\\|_{L^\Phi(\rn)}=0$, which further implies that $$\lim_{j\rightarrow\infty}\\|\chi_{E_j}\\|_{(E_\Phi^q)_t(\rn)} =\lim_{j\rightarrow\infty}\lf\{\int_{\rn}\lf[\frac{\\|\chi_{E_j}\chi_{B(x,t)}\\| _{L^\Phi(\rn)}}{\\|\chi_{B(x,t)}\\|_{L^\Phi(\rn)}}\r]^q\,dx\r\}^{\frac{1}{q}}=0.$$ This shows that $(E_\Phi^q)_t(\rn)$ has an absolutely continuous quasi-norm and hence finishes the proof of Lemma \[aboso\]. By Lemmas \[ballproof\], \[sconvex\], \[man28\] and \[man37\], we know that $(E_\Phi^q)_t(\rn)$ satisfies all the assumptions of [@ykds Theorems 3.6 and 3.7]. Thus, Theorem \[atom ch\] is a direct consequence of [@ykds Theorems 3.6 and 3.7]. From [@ykds Theorems 3.9 and 3.21], we further deduce Theorems \[molecular\] and \[lusin\]. Using Lemma \[aboso\] and [@ykds Corollary 3.11], we also obtain Proposition \[shoulian\]. This finishes the proofs of Theorems \[atom ch\], \[molecular\], \[lusin\] and Proposition \[shoulian\]. We first prove the necessity of Theorem \[gfunction\]. Let $f\in (HE_\Phi^q)_t(\rn)$. By Theorem \[lusin\], we know that $f$ vanishes weakly at infinity. Now we prove that $g(f)\in (E_\Phi^q)_t(\rn)$ and $\\|g(f)\\|_{(E_\Phi^q)_t(\rn)}\lesssim\\|f\\|_{(HE_\Phi^q)_t(\rn)}$. Let $s\in (0,\min\{p^-_{\Phi},q,1\})$. Then, by Theorem \[atom ch\] and the fact that $f\in (HE_\Phi^q)_t(\rn)$, we find that $$\label{450} f=\sum_{Q\in\mathcal{Q}}\lambda_Q a_Q\ \ \text{in}\ \ \mathcal{S}'(\rn),$$ where, for any $Q\in \mathcal{Q}$, $a_Q$ is an $((E_\Phi^q)_t(\rn),\ \infty,\ d)$-atom supported on $Q$, $d\in[\lfloor n(\frac{1}{s}-1)\rfloor,\infty)\bigcap\mathbb{Z}_+$ and $\{\lambda_Q\}_{Q\in\mathcal{Q}}\subset[0,\infty)$ satisfying $$\label{451} \lf\\|\lf\{\sum_{Q\in \mathcal{Q}} \lf[\frac{\lambda_Q} {\\|\chi_{Q}\\|_{(E_\Phi^q)_t(\rn)}} \r]^s\chi_{Q} \r\} ^{\frac{1}{s}}\r\\|_{(E_\Phi^q)_t(\rn)}\lesssim_s\\|f\\|_{(HE_\Phi^q)_t(\rn)}.$$ From (\[450\]), we deduce that, for any $x\in \rn$, $$\label{454} g(f)(x)\le\sum_{Q\in\mathcal{Q}}\lambda_Q g(a_Q)(x).$$ Let $r\in(\max\{1,\ q,\ p_{\Phi}^+\},\infty]$. Since $g$ is bounded on $L^r(\rn)$, it follows that, for any $Q\in\mathcal{Q}$, $$\lf\\|\chi_{2Q}g(a_Q)\r\\|_{L^r(\rn)}\lesssim\\|a_{Q}\\|_{L^r(\rn)} \lesssim\|Q\|^{1/r}\lf\\|\chi_{Q}\r\\|_{(E_\Phi^q)_t(\rn)}^{-1},$$ which, combined with [@ykds Theoreom 2.10], implies that $$\label{455} \lf\\|\sum_{Q\in\mathcal{Q}}\lambda_Q\chi_{2Q}g(a_Q)\r\\|_{(E_\Phi^q)_t(\rn)} \lesssim \lf\\|\lf\{\sum_{Q\in \mathcal{Q}} \lf[\frac{\lambda_Q} {\\|\chi_{Q}\\|_{(E_\Phi^q)_t(\rn)}} \r]^s\chi_{Q} \r\} ^{\frac{1}{s}}\r\\|_{(E_\Phi^q)_t(\rn)}.$$ Let $\theta\in(0,s)$. Repeating the proof of [@ns (4.4)] with $\\|\chi_Q\\|_{L^{p(\cdot)}}$ replaced by $\\|\chi_Q\\|_{(E_\Phi^q)_t(\rn)}$, we find that, for any $x\in\rn\setminus(2Q)$, $$g(a_Q)(x)\lesssim\lf(\frac{l_Q}{\|x-x_Q\|}\r)^{n+d+1}\lf\\|\chi_Q\r\\|_{(E_\Phi^q)_t(\rn)}^{-1} \lesssim\frac{1}{\\|\chi_Q\\|_{(E_\Phi^q)_t(\rn)}}\mathcal{M}^{(\theta)}(\chi_Q)(x),$$ where $l_Q$ and $x_Q$ denote the side-length and the center of $Q$, respectively. This, together with Lemma \[man28\], further implies that $$\lf\\|\sum_{Q\in\mathcal{Q}}\lambda_Q\chi_{\rn\setminus 2Q}g(a_Q)\r\\|_{(E_\Phi^q)_t(\rn)} \lesssim \lf\\|\lf\{\sum_{Q\in \mathcal{Q}} \lf[\frac{\lambda_Q} {\\|\chi_{Q}\\|_{(E_\Phi^q)_t(\rn)}} \r]^s\chi_{Q} \r\} ^{\frac{1}{s}}\r\\|_{(E_\Phi^q)_t(\rn)}.$$ From this, (\[454\]), (\[455\]) and (\[451\]), we deduce that $$\lf\\|g(f)\r\\|_{(E_\Phi^q)_t(\rn)}\le \lf\\|\sum_{Q\in\mathcal{Q}}\lambda_Q g(a_Q)\r\\|_{(E_\Phi^q)_t(\rn)} \lesssim\lf\\|\lf\{\sum_{Q\in \mathcal{Q}} \lf[\frac{\lambda_Q} {\\|\chi_{Q}\\|_{(E_\Phi^q)_t(\rn)}} \r]^s\chi_{Q} \r\} ^{\frac{1}{s}}\r\\|_{(E_\Phi^q)_t(\rn)}\lesssim\\|f\\|_{(HE_\Phi^q)_t(\rn)}.$$ Therefore, $g(f)\in (E_\Phi^q)_t(\rn)$ and $\\|g(f)\\|_{(E_\Phi^q)_t(\rn)}\lesssim\\|f\\|_{(HE_\Phi^q)_t(\rn)}$, which completes the proof of the necessity of Theorem \[gfunction\]. To complete the proof of Theorem \[gfunction\], it remains to show the sufficiency of Theorem \[gfunction\]. To this end, by Theorem \[lusin\], we only need to prove that, if $f\in\mathcal{S}'(\rn)$ vanishes weakly at infinity and $g(f)\in(E_\Phi^q)_t(\rn)$, then $$\label{349} \lf\\|S(f)\r\\|_{(E_\Phi^q)_t(\rn)}\lesssim\lf\\|g(f)\r\\|_{(E_\Phi^q)_t(\rn)}.$$ Let $\varphi\in\mathcal{S}(\rn)$ be such that $$\chi_{B(\vec{0}_n,4) \setminus B(\vec{0}_n,2)}\le\varphi\le\chi_{B(\vec{0}_n,8) \setminus B(\vec{0}_n,1)}.$$ Let $\psi\in\mathcal{S}(\rn)$ satisfy that $\mathcal{F}^{-1}(\psi)=\varphi$. Then it is easy to prove that $$\int_{\rn}\psi(x)x^\alpha\,dx=0,\ \ \ \ \ \forall \alpha\in\mathbb{Z}_+^n.$$ For any $a,\ \tau\in(0,\infty)$, $f\in\mathcal{S}'(\rn)$ and $x\in\rn$, let $$(\psi_\tau^f)_a(x):=\sup_{y\in\rn}\frac{\|\psi_\tauf(y)\|}{(1+\|x-y\|/\tau)^a}.$$ For any $l\in \mathbb{Z}$, denote $\psi_{2^{-l}}$ and $(\psi_{2^{-l}}^)_a$ simply by $\psi_{l}$ and $(\psi_{l}^)_a$, respectively. It is easy to see that, for any $x\in\rn$, $$\begin{aligned} S(f)(x) &=\lf\{\int_{\Gamma(x)}\|\varphi(\tau D)(f)(y)\|^{2}\,\frac{dy\,d\tau}{\tau^{n+1}}\r\}^{ \frac{1}{2}} \lesssim\lf\{\int_{0}^{\infty}\sup_{\{y\in\rn:\|y-x\|<\tau\}}\|\varphi(\tau D)(f)(y)\|^{2}\,\frac{dy\,d\tau}{\tau}\r\}^{ \frac{1}{2}}\\ &\lesssim\lf\{\int_{0}^{\infty}\lf[(\psi_\tau^f)_a(x)\r]^{2}\,\frac{d\tau}{\tau}\r\}^{ \frac{1}{2}}.\end{aligned}$$ Let, for any $x\in\rn$, $$P_a(f)(x):=\lf\{\int_{0}^{\infty}\lf[(\psi_\tau^f)_a(x)\r]^{2}\,\frac{d\tau}{\tau}\r\}^{ \frac{1}{2}}.$$ Thus, to show (\[349\]), it suffices to prove that, if $f\in\mathcal{S}'(\rn)$ and $g(f)\in(E_\Phi^q)_t(\rn)$, then $$\label{77} \lf\\|P_a(f)\r\\|_{(E_\Phi^q)_t(\rn)}\lesssim\lf\\|g(f)\r\\|_{(E_\Phi^q)_t(\rn)}.$$ Let $a\in(\frac{n}{\min\{p_{\Phi}^-,\ q\}},\infty)$. We choose $r\in(\frac{n}{a},\min\{p_{\Phi}^-,\ q\})$. Then, by [@lsuy Lemma 3.5], we find that, for any $l\in \mathbb{Z}$, $\tau\in[1,2]$, $N\in \mathbb{N}$, $a\in(0,N]$ and $x\in\rn$, $$\lf[(\psi_{2^{-l}\tau}^f)_a(x)\r]^{r} \lesssim\sum_{k=0}^{\infty}2^{-kNr}2^{(k+l)n}\int_{\rn}\frac{\|(\psi_{k+l})_\tauf(y)\|^r}{(1+2^l\|x-y\|)^{ar}}dy.$$ From the Minkowski inequality, it follows that $$\begin{aligned} \lf\{\int_{1}^{2}\lf[(\psi_{2^{-l}\tau}^\ast f)_a(x)\r]^{2}\,\frac{d\tau}{\tau}\r\}^{ \frac{r}{2}} &\lesssim\lf\{\int_{1}^{2}\lf[\sum_{k=0}^{\infty}2^{-kNr}2^{(k+l)n} \int_{\rn}\frac{\|(\psi_{k+l})_\tauf(y)\|^r}{(1+2^l\|x-y\|)^{ar}}dy\r]^{\frac{2}{r}}\,\frac{d\tau}{\tau}\r\}^{ \frac{r}{2}}\\ &\lesssim\sum_{k=0}^{\infty} 2^{-kNr}2^{(k+l)n}\int_{\rn} \frac{[\int_{1}^{2}\|(\psi_{k+l})_\tauf(y)\|^2\frac{d\tau}{\tau}]^{\frac{r}{2}}}{(1+2^l\|x-y\|)^{ar}}\,dy\\ &\lesssim\sum_{k=0}^{\infty} 2^{-kNr}2^{kn} \lf(g_l\ast\lf[\int_{1}^{2}\|(\psi_{k+l})_\tau\ast f(\cdot)\|^2\frac{d\tau}{\tau}\r]^{\frac{r}{2}}\r)(x)\\ &\lesssim\sum_{k=0}^{\infty} 2^{-k(Nr-n)} \mathcal{M}\lf(\lf[\int_{1}^{2}\|(\psi_{k+l})_\tau\ast f(\cdot)\|^2\frac{d\tau}{\tau}\r]^{\frac{r}{2}}\r)(x),\end{aligned}$$ where, for any $l\in\mathbb{Z}$ and $x\in\rn$, $$g_l(x):=\frac{2^{nl}}{(1+2^l\|x\|)^{ar}}\ \in L^1(\rn)\ \ \ \mbox{and}\ \ \ \\|g_l\\|_{L^1(\rn)}\lesssim1.$$ Then, by the Minkowski inequality, we find that $$\begin{aligned} \lf\\|P_a(f)\r\\|_{(E_\Phi^q)_t(\rn)} &=\lf\\|\lf\{\sum_{l=-\infty}^{\infty} \int_{2^{-l}}^{2^{-l+1}}\lf[(\psi_\tau^\ast f)_a(\cdot)\r]^{2}\,\frac{d\tau}{\tau}\r\}^{ \frac{1}{2}}\r\\|_{(E_\Phi^q)_t(\rn)}\\ &=\lf\\|\lf\{\sum_{l=-\infty}^{\infty} \int_{1}^{2}\lf[(\psi_{2^{-l}}^\ast f)_a(\cdot)\r]^{2}\,\frac{d\tau}{\tau}\r\}^{ \frac{1}{2}}\r\\|_{(E_\Phi^q)_t(\rn)}\\ &\lesssim\lf\\|\lf\{\sum_{l=-\infty}^{\infty}\lf[ \sum_{k=0}^{\infty} 2^{-k(Nr-n)} \mathcal{M}\lf(\lf[\int_{1}^{2}\|(\psi_{k+l})_\tau\ast f(\cdot)\|^2\frac{d\tau}{\tau}\r]^{\frac{r}{2}}\r)\r]^{\frac{2}{r}}\r\}^{ \frac{1}{2}}\r\\|_{(E_\Phi^q)_t(\rn)}\\ &\lesssim\lf\\|\lf\{\sum_{l=-\infty}^{\infty}\lf[ \mathcal{M}\lf(\lf[\int_{1}^{2}\|(\psi_{l})_\tau\ast f(\cdot)\|^2\frac{d\tau}{\tau}\r]^{\frac{r}{2}}\r)\r]^{\frac{2}{r}}\r\}^{ \frac{1}{2}}\r\\|_{(E_\Phi^q)_t(\rn)}.\end{aligned}$$ By the fact that $r\in(\frac{n}{a},\min\{p_{\Phi}^-,\ q\})$ and Theorem \[main\], we conclude that $$\begin{aligned} &\lf\\|\lf\{\sum_{l=-\infty}^{\infty}\lf[ \mathcal{M}\lf(\lf[\int_{1}^{2}\|(\psi_{l})_\tau\ast f(\cdot)\|^2\frac{d\tau}{\tau}\r]^{\frac{r}{2}}\r)\r]^{\frac{2}{r}}\r\}^{ \frac{1}{2}}\r\\|_{(E_\Phi^q)_t(\rn)}\\ &\quad\lesssim \lf\\|\lf\{\sum_{l=-\infty}^{\infty}\int_{1}^{2}\|(\psi_{l})_\tau\ast f(\cdot)\|^2\frac{d\tau}{\tau}\r\}^{ \frac{1}{2}}\r\\|_{(E_\Phi^q)_t(\rn)}\lesssim\lf\\|g(f)\r\\|_{(E_\Phi^q)_t(\rn)},\end{aligned}$$ which implies that (\[77\]) holds true. This finishes the proof of Theorem \[gfunction\]. To prove this theorem, we only need to show the necessity, since the sufficiency is easy because of Theorem \[lusin\] and the obvious fact that, for any $f\in\mathcal{S}'(\rn)$ and $x\in\rn$, $S(f)(x)\le g_\lambda^(f)(x)$. To show the necessity, for any $f\in (HE_\Phi^q)_t(\rn)$, by Theorem \[lusin\], we know that $f$ vanishes weakly at infinity. From the fact that $\lambda\in(1+\frac{2}{\min\{p_{\Phi}^-,\ q\}},\infty)$, we deduce that there exists $a\in(\frac{n}{\min\{p_{\Phi}^-,\ q\}},\infty)$ such that $\lambda\in(1+\frac{2a}{n},\infty)$ and, for any $x\in\rn$, $$\begin{aligned} g_\lambda^(f)(x)&=\lf\{\int_0^{\infty}\int_{\rn}\lf(\frac{\tau}{\tau+\|x-y\|}\r)^{\lambda n}\|\varphi(\tau D)(f)(y)\|^{2}\,\frac{dy\,d\tau}{\tau^{n+1}}\r\}^{\frac{1}{2}}\\ &\lesssim\lf\{\int_0^{\infty}\lf[(\psi_\tau^f)_a(x)\r]^2\int_{\rn}\lf(1+\frac{\|x-y\|}{\tau}\r)^{2a-\lambda n}\,\frac{dy\,d\tau}{\tau^{n+1}}\r\}^{\frac{1}{2}}\\ &\sim\lf\{\int_0^{\infty}\lf[(\psi_\tau^f)_a(x)\r]^2\,\frac{d\tau}{\tau}\r\}^{\frac{1}{2}}\sim P_a(f)(x),\end{aligned}$$ which, combined with (\[77\]) and Theorem \[gfunction\], implies that $$\lf\\|g_\lambda^(f)\r\\|_{(E_\Phi^q)_t(\rn)}\lesssim\\|f\\|_{(HE_\Phi^q)_t(\rn)}.$$ This finishes the proof of Theorem \[glamda\]. To show Theorem \[finite\], we need the following lemma. \[youjie\] Let $t,\ q\in(0,\infty)$ and $\Phi$ be an Orlicz function with positive lower type $p_{\Phi}^-$ and positive upper type $p_{\Phi}^+$. Let $N\in\mathbb{N}\cap(\lfloor\frac{n}{\min\{p_\Phi^-,q\}}+1\rfloor,\infty)$. Suppose $f\in(HE_\Phi^q)_t(\rn)$, $\\|f\\|_{(HE_\Phi^q)_t(\rn)}=1$ and $\supp(f)\subset B(\vec{0}_n,R)$ with $R\in(1,\infty)$. Then there exists a positive constant $C_{(N)}$, depending on $N$, but independent of $f$ and $t$, such that, for any $x\notin B(\vec{0}_{n},4R)$, $$\label{423} M_N(f)(x)\le C_{(N)}\lf\\|\chi_{B(\vec{0}_n,R)}\r\\|_{(E_\Phi^q)_t(\rn)}^{-1}.$$ For any $x\notin B(\vec{0}_n,4R)$, by (\[zhujida\]), we have $$M_N(f)(x)\le M_N^0(f)(x).$$ To prove (\[423\]), it suffices to show that, for any $\varphi\in\mathcal{F}_N(\rn)$, $\tau\in(0,\infty)$ and $x\notin B(\vec{0}_{n},4R)$, $$\|\varphi_\tau\ast f(x)\|\lesssim \lf\\|\chi_{B(\vec{0}_n,R)}\r\\|_{(E_\Phi^q)_t(\rn)}^{-1}.$$ Let $\theta\in\mathcal{S}(\rn)$ be such that $\supp(\theta)\subset B(\vec{0}_n,2)$, $0\le\theta\le1$ and $\theta\equiv1$ on $B(\vec{0}_n,1)$. We distinguish two cases with respect to the size of $\tau$. For any $\tau\in [R,\infty)$ and $x\notin B(\vec{0}_{n},4R)$, arguing as in the proof of [@cw Lemma 7.10], we have $$\label{424} \varphi_\tau\ast f(x)=\psi_R\ast f(\vec{0}_n)$$ and $c\psi\in\mathcal{F}_N(\rn)$ with $c:=C_{(N)}$, where, for any $\tau\in [R,\infty)$ and $z\in\rn$, $$\psi(z):=\lf(\frac{R}{\tau}\r)^n\Phi\lf(\frac{x}{\tau}+\frac{Rz}{\tau}\r)\theta(z).$$ Therefore, (\[424\]) ensures that, for any $x\notin B(\vec{0}_n,4R)$, $$\|\varphi_\tau\ast f(x)\|\lesssim M_N(f)(z),\quad \forall\ z\in B(\vec{0}_n,R),$$ which, together with $\\|f\\|_{(HE_\Phi^q)_t(\rn)}=1$, further implies that, for any $x\notin B(\vec{0}_n,4R)$, $$\begin{aligned} \label{111} \|\varphi_\tau\ast f(x)\|&\lesssim \inf_{z\in B(\vec{0}_n,R)}M_N(f)(z)\\\noz &\lesssim \frac{\\|\chi_{B(\vec{0}_n,R)}\inf_{z\in B(\vec{0}_n,R)}M_N(f)(z)\\|_{(E_\Phi^q)_t(\rn)}}{\\|\chi_{B(\vec{0}_n,R)}\\|_{(E_\Phi^q)_t(\rn)}}\\ \noz &\lesssim\lf\\|\chi_{B(\vec{0}_n,R)}\r\\|_{(E_\Phi^q)_t(\rn)}^{-1}\lf\\|M_N(f)\r\\|_{(E_\Phi^q)_t(\rn)} \lesssim\lf\\|\chi_{B(\vec{0}_n,R)}\r\\|_{(E_\Phi^q)_t(\rn)}^{-1}.\end{aligned}$$ For any $\tau\in(0, R)$ and $u\in B(\vec{0}_n,\frac{R}{2})$, following [@cw Lemma 7.10], we obtain $$\varphi_\tau\ast f(x)=\psi_\tau\ast f(u)$$ and $c\psi\in\mathcal{F}_N(\rn)$ with $c:=C_{(N)}$, where, for any $\tau\in(0, R)$ and $z\in\rn$, $$\psi(z):=\Phi\lf(\frac{x-u}{\tau}+z\r)\theta\lf(\frac{u}{R}-\frac{tz}{R}\r).$$ Thus, for any $x\notin B(\vec{0}_n,4R)$, we have $$\|\varphi_\tau\ast f(x)\|\lesssim M_N(f)(u),\quad \forall\ u\in B\lf(\vec{0}_n,\frac{R}{2}\r).$$ By proceeding as in (\[111\]), we further conclude that, for any $\tau\in(0, R)$ and $x\notin B(\vec{0}_n,4R)$, $$\|\varphi_\tau\ast f(x)\|\lesssim \lf\\|\chi_{B(\vec{0}_n,R)}\r\\|_{(E_\Phi^q)_t(\rn)}^{-1}.$$ This finishes the proof of Lemma \[youjie\]. Obviously, from Theorem \[atom ch\], we deduce that $$(HE_\Phi^{q,r,d})_t^{\fin}(\rn)\subset(HE_\Phi^q)_t(\rn)$$ and, for any $f\in(HE_\Phi^{q,r,d})_t^{\fin}(\rn)$, $$\\|f\\|_{(HE_\Phi^q)_t(\rn)}\lesssim\\|f\\|_{(HE_\Phi^{q,r,d})_t^{\fin}(\rn)}.$$ Thus, to complete the proof of Theorem \[finite\], we still need to show that, for any given $t,\ q,\ d$ as in Theorem \[finite\] and $r\in(\max\{1,\ q,\ p_{\Phi}^+\},\infty)$ and any $f\in(HE_\Phi^{q,r,d})_t^{\fin}(\rn)$, $$\\|f\\|_{(HE_\Phi^{q,r,d})_t^{\fin}(\rn)}\lesssim\\|f\\|_{(HE_\Phi^q)_t(\rn)},$$ and that a similar estimate also holds true for $r=\infty$ and any given $t,\ q,\ d$ as in Theorem \[finite\] and any $f\in(HE_\Phi^{q,\infty,d})_t^{\fin}(\rn)\cap\mathcal{C}{(\rn)}$. Assume that $r\in(\max\{1,\ q,\ p_{\Phi}^+\},\infty]$ and, by the homogeneity of both $\\|\cdot\\|_{(HE_\Phi^{q,r,d})_t^{\fin}(\rn)}$ and $\\|\cdot\\|_{(HE_\Phi)_t(\rn)}$, without loss of generality, we may also assume that $f\in(HE_\Phi^{q,r,d})_t^{\fin}(\rn)$ and $\\|f\\|_{(HE_\Phi^q)_t(\rn)}=1$. Since $f$ is a finite linear combination of $((E_\Phi^q)_t(\rn),\ r,\ d)$-atoms, it follows that there exists $R\in(1,\infty)$ such that $f$ is supported on $B(\vec{0}_n,R)$. Thus, if let $N$ be as in Lemma \[youjie\], then, by Lemma \[youjie\], there exists a positive constant $C_{(N)}$ such that, for any $x\notin B(\vec{0}_{n},4R)$, $$\label{78} M_N(f)(x)\le C_{(N)}\lf\\|\chi_{B(\vec{0}_n,R)}\r\\|_{(E_\Phi^q)_t(\rn)}^{-1}.$$ For each $j\in\mathbb{Z}$, let $\mathcal{O}_j:=\lf\{x\in\rn:\ M_N(f)(x)>2^j\r\}$. Denote by $j'$ the largest integer $j$ such that $$\label{130} 2^{j'}<C_{(N)}\lf\\|\chi_{B(\vec{0}_n,R)}\r\\|_{(E_\Phi^q)_t(\rn)}^{-1}.$$ Then, by (\[78\]), for any $j\in\{j'+1, j'+2, \ldots\}$, $$\label{oo} \mathcal{O}_j\subset B(\vec{0}_n,4R).$$ Since $f\in L^r(\rn)$, from the proof of [@ykds Proposition 4.3], it follows that there exist a sequence $\{(a_{j,k},Q_{j,k})\}_{j\in\mathbb{Z},k\in K_j}$ of pairs of $((E_\Phi^q)_t(\rn),\ \infty,\ d)$-atoms and their supports, and a sequence of scalars, $\{\lambda_{j,k}\}_{j\in\mathbb{Z},k\in K_j}\subset[0,\infty)$, such that $$\label{eqq} f=\sum_{j=-\infty}^{\infty}\sum_{k\in K_j}\lambda_{j,k}a_{j,k}$$ in both $\mathcal{S}'(\rn)$ and almost everywhere, where $\{K_j\}_{j\in\mathbb{Z}}$ is a set of indices and $\{Q_{j,k}\}_{j\in\mathbb{Z},k\in K_j}$ a family of closed cubes with disjoint interiors such that $\mathcal{O}_j=\cup_{k\in K_j}Q_{j,k}$ as in [@ykds Lemma 2.23]. Moreover, for some given $s\in (0,\min\{p_\Phi^-,q\})$, we have $$\label{331} \lf\\|\lf\{\sum_{j=-\infty}^{\infty}\sum_{k\in K_j} \lf[\frac{\lambda_{j,k}} {\\|\chi_{Q_{j,k}}\\|_{(E_\Phi^q)_t(\rn)}} \r]^s\chi_{Q_{j,k}} \r\} ^{\frac{1}{s}}\r\\|_{(E_\Phi^q)_t(\rn)}\lesssim\\|f\\|_{(HE_\Phi^q)_t(\rn)}.$$ Define $$\label{eq2} h:=\sum_{j=-\infty}^{j'}\sum_{k\in K_j}\lambda_{j,k}a_{j,k}\quad\mathrm{and} \quad l:=\sum_{j=j'+1}^{\infty}\sum_{k\in K_j}\lambda_{j,k}a_{j,k},$$ where the series converge in both $\mathcal{S}'(\rn)$ and almost everywhere. Clearly $f=h+l$ and, by (\[oo\]), $\supp(l)\subset\cup_{j>j'}\mathcal{O}_j\subset B(\vec{0}_n,4R)$. Therefore, $h=l=0$ on $\rn\setminus B(\vec{0}_n,4R)$ and hence $\supp(h)\subset B(\vec{0}_n,4R)$. Moreover, by the proof of [@ykds Proposition 4.3], we know that there exists a positive constant $C_0$ such that $\\|\lambda_{j,k}a_{j,k}\\|_{L^{\infty}(\rn)}\le C_02^j$. Since $f\in L^r(\rn)$ and $r\in(1,\infty]$, from the boundedness on $L^r(\rn)$ of the Hardy-Littlewood maximal operator, it follows that $M_N(f)\in L^r(\rn)$. Then we have $$\begin{aligned} \label{ll} \\|l\\|_{L^r(\rn)}&\leq\lf\\|\sum_{j=j'+1}^{\infty}\sum_{k\in K_j}\|\lambda_{j,k}a_{j,k}\|\r\\|_{L^r(\rn)}\lesssim\lf\\|\sum_{j=j'+1}^{\infty} \sum_{k\in K_j}2^j\chi_{Q_{j,k}}\r\\|_{L^r(\rn)}\\ \noz &\lesssim\lf\\|\sum_{j=j'+1}^{\infty}2^j\chi_{\mathcal{O}_j}\r\\|_{L^r(\rn)}\lesssim\\|M_N(f)\\|_{L^r(\rn)}.\end{aligned}$$ Thus, $l\in L^r(\rn)$ and so $h=f-l\in L^r(\rn)$. It follows from (\[ll\]) and the Hölder inequality that, for any $\|\beta\|\leq d$, $$\begin{aligned} \int_{\rn}\sum_{j=j'+1}^{\infty}\sum_{k\in K_j}\lf\|x^\beta\r\|\|\lambda_{j,k}a_{j,k}(x)\|\,dx &\le\lf\\|\sum_{j=j'+1}^{\infty}\sum_{k\in K_j}\|\lambda_{j,k}a_{j,k}\|\r\\|_{L^r(\rn)} \lf\{\int_{B(\vec{0}_n,4R)}\lf\|x^\beta\r\|^{r'}\,dx\r\}^{\frac1r'}\\ &\lesssim_R\\|M_N(f)\\|_{L^r(\rn)}<\infty.\end{aligned}$$ This, combined with the vanishing moments of $a_{j,k}$, implies that $l$ has vanishing moments up to $d$ and hence so does $h$ by $h=f-l$. In order to estimate the size of $g$ in $B(\vec{0}_n,4R)$, recall that $$\label{129} \lf\\|\lambda_{j,k}a_{j,k}\r\\|_{L^{\infty}(\rn)}\lesssim2^j,\ \supp(a_{j,k})\subset Q_{j,k} \quad\mathrm{and}\quad\sum_{k\in K_j}\chi_{Q_{j,k}}\lesssim1.$$ It is easy to show that $$\label{131} \lf\\|\chi_{B(\vec{0}_n,R)}\r\\|_{(E_\Phi^q)_t(\rn)}\sim\lf\\|\chi_{Q(\vec{0}_n,8R)}\r\\|_{(E_\Phi^q)_t(\rn)}.$$ Indeed, it is easy to see that there exist $M\in\mathbb{N}$ and $\{x_1, \ldots, x_M\}\subset\rn$, independent of $t$ and $f$, such that $M\lesssim1$ and $Q(\vec{0}_n,8R)\subseteq\bigcup_{m=1}^MB(x_m,R)$, which further implies that $$\label{555} \lf\\|\chi_{Q(\vec{0}_n,8R)}\r\\|_{(E_\Phi^q)_t(\rn)}\lesssim \lf\\|\sum_{m=1}^M\chi_{B(x_m,R)}\r\\|_{(E_\Phi^q)_t(\rn)}\lesssim \sum_{m=1}^M\lf\\|\chi_{B(x_m,R)}\r\\|_{(E_\Phi^q)_t(\rn)}.$$ Observing that, for any $t\in(0,\infty)$, $m\in\mathbb{N}$ and $x\in\rn $, $$\lf\\|\chi_{B(x_m,R)}\chi_{B(x,t)}\r\\|_{L^\Phi(\rn)} =\lf\\|\chi_{B(\vec{0}_n,R)}\chi_{B(x-x_m,t)}\r\\|_{L^\Phi(\rn)},$$ by this and with $\widetilde{C}_{(\Phi,t)}$ as therein, we have $$\begin{aligned} \lf\\|\chi_{B(x_m,R)}\r\\|_{(E_\Phi^q)_t(\rn)} &=\lf\{\int_{\rn}\lf[\frac{\\|\chi_{B(\vec{0}_n,R)}\chi_{B(x-x_m,t)}\\|_{L^\Phi(\rn)}} {\\|\chi_{B(x,t)}\\|_{L^\Phi(\rn)}}\r]^q\,dx\r\}^{\frac{1}{q}}\\ &=\frac{1}{\widetilde{C}_{(\Phi,t)}}\lf\{\int_{\rn}\lf\\|\chi_{B(\vec{0}_n,R)}\chi_{B(x-x_m,t)}\r\\|_{L^\Phi(\rn)} ^q\,dx\r\}^{\frac{1}{q}}\\ &=\frac{1}{\widetilde{C}_{(\Phi,t)}}\lf\{\int_{\rn}\lf\\|\chi_{B(\vec{0}_n,R)}\chi_{B(x,t)}\r\\|_{L^\Phi(\rn)} ^q\,dx\r\}^{\frac{1}{q}} =\lf\\|\chi_{B(\vec{0}_n,R)}\r\\|_{(E_\Phi^q)_t(\rn)},\end{aligned}$$ which, together with $\eqref{555}$, implies that $\\|\chi_{Q(\vec{0}_n,8R)}\\|_{(E_\Phi^q)_t(\rn)}\lesssim\\|\chi_{B(\vec{0}_n,R)}\\|_{(E_\Phi^q)_t(\rn)}.$ The converse inequality holds true obviously. Thus, we obtain . Combining (\[130\]), (\[129\]) and (\[131\]), we conclude that $$\\|h\\|_{L^{\infty}(\rn)}\le\sum_{j\le j'}\lf\\|\sum_{k\in K_j}\|\lambda_{j,k}a_{j,k}\|\r\\|_{L^{\infty}(\rn)} \lesssim\sum_{j\le j'}2^j\lesssim2^{j'}\lesssim\lf\\|\chi_{B(\vec{0}_n,R)}\r\\|_{(E_\Phi^q)_t(\rn)}^{-1} \le\widetilde{C}\lf\\|\chi_{Q(\vec{0}_n,8R))}\r\\|_{(E_\Phi^q)_t(\rn)}^{-1},$$ where $\widetilde{C}$ is a positive constant independent of $f$ and $t$. From this and the fact that $h$ has vanishing moments up to $d$, it follows that $\widetilde{C}^{-1}h$ is an $((E_\Phi^q)_t(\rn),\ \infty,\ d)$-atom. Now, to complete the proof of Theorem \[finite\](i), we assume that $r\in(\max\{1,\ q,\ p_{\Phi}^+\},\infty)$. We rewrite $l$ as a finite linear combination of $((E_\Phi^q)_t(\rn),\ r,\ d)$-atoms. For any $i\in\mathbb{N}$, let $$F_i:=\lf\{(j,k)\in\mathbb{Z}\times\mathbb{Z}_+:\ j\in\lf\{j'+1,j'+2, \ldots\r\},\ k\in K_j,\ \|j\|+k\le i\r\},$$ and $l_i:=\sum_{(j,k)\in F_i}\lambda_{j,k}a_{j,k}$. Since the series $l=\sum_{j=j'+1}^{\infty}\sum_{k\in K_j}\lambda_{j,k}a_{j,k}$ converges in $L^r(\rn)$, it follows that there exists a positive integer $i_0$, which may depend on $t$ and $f$, such that $$\\|l-l_{i_0}\\|_{L^r(\rn)}\le\frac{\|Q(\vec{0}_n,8R)\|^{\frac{1}{r }}}{\\|\chi_{Q(\vec{0}_n,8R)}\\|_{(E_\Phi^q)_t(\rn)}}.$$ Thus, $l-l_{i_0}$ is an $((E_\Phi^q)_t(\rn),\ r,\ d)$-atom, because $\supp(l-l_{i_0})\subset B(\vec{0}_n,4R)\subset Q(\vec{0}_n,8R)$ and, for any $\|\beta\|\le d$, $\int_{\rn}(l-l_{i_0})(x)x^\beta dx=0$. Therefore, $$f=h+l=\widetilde{C}\widetilde{C}^{-1}h+(l-l_{i_0})+l_{i_0}$$ is a finite decomposition of $f$ in terms of $((E_\Phi^q)_t(\rn),\ r,\ d)$-atoms. Moreover, by (\[331\]), we have $$\begin{aligned} \\|f\\|_{(HE_\Phi^{q,r,s})_t^{\fin}(\rn)} &\le\lf\\|\lf\{\lf[\frac{\widetilde{C}} {\\|\chi_{Q(\vec{0}_n,8R)}\\|_{(E_\Phi^q)_t(\rn)}} \r]^s\chi_{Q(\vec{0}_n,8R)}+\lf[\frac{1} {\\|\chi_{Q(\vec{0}_n,8R)}\\|_{(E_\Phi^q)_t(\rn)}} \r]^s\chi_{Q(\vec{0}_n,8R)}\r.\r.\\ &\quad\lf.\lf.+\sum_{(j,k)\in F_{i_0}}\lf[\frac{\lambda_{j,k}} {\\|\chi_{Q_{j,k}}\\|_{(E_\Phi^q)_t(\rn)}} \r]^s\chi_{Q_{j,k}} \r\} ^{\frac{1}{s}}\r\\|_{(E_\Phi^q)_t(\rn)}\\ &\lesssim1+\lf\\|\lf\{\sum_{(j,k)\in F_{i_0}}\lf[\frac{\lambda_{j,k}} {\\|\chi_{Q_{j,k}}\\|_{(E_\Phi^q)_t(\rn)}} \r]^s\chi_{Q_{j,k}} \r\} ^{\frac{1}{s}}\r\\|_{(E_\Phi^q)_t(\rn)}\\ &\lesssim1+\lf\\|\lf\{\sum_{j=-\infty}^{\infty}\sum_{k\in K_j} \lf[\frac{\lambda_{j,k}} {\\|\chi_{Q_{j,k}}\\|_{(E_\Phi^q)_t(\rn)}} \r]^s\chi_{Q_{j,k}} \r\} ^{\frac{1}{s}}\r\\|_{(E_\Phi^q)_t(\rn)}\\ &\lesssim1+\\|f\\|_{(HE_\Phi^q)_t(\rn)}\lesssim1.\end{aligned}$$ Thus, $\\|f\\|_{(HE_\Phi^{q,r,d})_t^{\fin}(\rn)}\lesssim1$. This finishes the proof of Theorem \[finite\](i). To prove Theorem \[finite\](ii), we assume that $f\in(HE_\Phi^{q,\infty,d})_t^{\fin}(\rn)\cap\mathcal{C}(\rn)$ and $\\|f\\|_{(HE_\Phi^q)_t(\rn)}=1$. Since $f$ has a compact support, it follows that $f$ is uniformly continuous. Then, by this, the proof of [@ykds Proposition 4.3] and the argument presented in [@em pp.108-109], we know that each $((E_\Phi^q)_t(\rn),\ \infty,\ d)$-atom $a_{j,k}$ in is continuous. Since $f$ is bounded, from the boundedness of $M_N(f)$ on $L^{\infty}(\rn)$, it follows that there exists a positive integer $j''>j'$ such that $\mathcal{O}_j=\emptyset$ for any $j\in\{j''+1, j''+2, \ldots\}$. Consequently, in this case, $l$ in becomes $$l=\sum_{j=j'+1}^{j''}\sum_{k\in K_j}\lambda_{j,k}a_{j,k}.$$ Let $\epsilon\in(0,\infty)$. Since $f$ is uniformly continuous, it follows that there exists $\delta\in(0,\infty)$ such that, if $\|x-y\|<\delta$, then $\|f(x)-f(y)\|<\epsilon$. Write $l=l_1^\epsilon+l_2^\epsilon$ with $l_1^\epsilon:=\sum_{(j,k)\in F_1}\lambda_{j,k}a_{j,k}$ and $l_2^\epsilon:=\sum_{(j,k)\in F_2}\lambda_{j,k}a_{j,k}$, where $$F_1:=\lf\{(j,k)\in\mathbb{Z}\times\mathbb{Z}_+:\ j\in\lf\{j'+1, \ldots,j''\r\},\ k\in K_j, \ {\rm diam}(Q_{j,k})\ge\delta\r\}$$ and $$F_2:=\lf\{(j,k)\in\mathbb{Z}\times\mathbb{Z}_+:\ j\in\lf\{j'+1, \ldots, j''\r\},\ k\in K_j,\ {\rm diam}(Q_{j,k})<\delta\r\}.$$ Observe that $l_1^\epsilon$ is a finite summation. Since the atoms are continuous, we know that $l_1^\epsilon$ is also a continuous function. Furthermore, using this fact and repeating the proof of [@bly Theorem 6.2], we conclude that $$\\|l_2^\epsilon\\|_{L^\infty(\rn)}\lesssim(j''-j')\epsilon.$$ This means that one can write $l$ as the sum of one continuous term and one which is uniformly arbitrarily small. Thus, $l$ is continuous and so is $h=f-l$. To find a finite atomic decomposition of $f$, we use again the splitting $l=l_1^\epsilon+l_2^\epsilon$. It is clear that, for any $\epsilon\in(0,\infty)$, $l_1^\epsilon$ is a finite combination of continuous $((E_\Phi^q)_t(\rn),\ \infty,\ d)$-atoms. Also, since both $l$ and $l_1^\epsilon$ are continuous and have vanishing moments up to order $d$, it follows that $l_2^\epsilon=l-l_1^\epsilon$ is also continuous and has vanishing moments up to order $d$. Moreover, $\supp(l_2^\epsilon)\subset B(\vec{0}_n,4R)\subset Q(\vec{0}_n,8R)$ and $ \\|l_2^\epsilon\\|_{L^\infty(\rn)}\lesssim(j''-j')\epsilon$. So we can choose $\epsilon$ small enough such that $l_2^\epsilon$ becomes an arbitrarily small multiple of a continuous $((E_\Phi^q)_t(\rn),\ \infty,\ d)$-atom. Therefore, $f=h+l_1^\epsilon+l_2^\epsilon$ is a finite linear continuous atomic combination. Then, by an argument similar to the proof of (i), we obtain $\\|f\\|_{(HE_\Phi^{q,\infty,d})_t^{\fin}(\rn)}\lesssim1$. This finishes the proof of (ii) and hence of Theorem \[finite\]. Dual spaces of Orlicz-slice Hardy spaces {#s5} ======================================== In this section, we provide a description of the dual space of the Orlicz-slice Hardy space $(HE_\Phi^q)_t(\rn)$, with $\max\{p_{\Phi}^+,\ q\}\in(0,1]$, in terms of Campanato spaces. This description is a consequence of both their atomic characterization from Theorem \[atom ch\] and their finite atomic characterization from Theorem \[finite\] as well as some basic tools from functional analysis. A function $\Phi:\ [0,\infty)\rightarrow\mathbb{R}$ is said to be concave if, for any $t,\ s\in[0,\infty)$ and $\lambda\in[0,1]$, $$\lambda\Phi(t)+(1-\lambda)\Phi(s)\leq \Phi(\lambda t+(1-\lambda)s).$$ \[djdj\] Let $\Phi$ be an Orlicz function with positive lower type $p_{\Phi}^-$ and positive upper type $p_{\Phi}^+$ satisfying $p_{\Phi}^+\in(0,1].$ Then there exists a concave function $\widetilde{\Phi}$ with the same types as $\Phi$, which is equivalent to $\Phi$. Consider the function $$\widetilde{\Phi}(t):=\begin{cases} \displaystyle\int_0^t \inf_{\tau\in(0,s)}\frac{\Phi(\tau)}{\tau}\,ds,\quad &t \in (0,\infty],\\ \ 0,\quad &t=0. \end{cases}$$ Then it is easy to prove that $\widetilde{\Phi}$ is concave on $[0,\infty)$. By the assumption that $p_{\Phi}^+\in(0,1]$, we know that, for any $t\in[0,\infty)$, $$\widetilde{\Phi}(t)\geq t\inf_{\tau\in(0,t)}\frac{\Phi(\tau)}{\tau}\gtrsim t\inf_{\tau\in(0,t)}\lf(\frac{\tau}{t}\r)^{p_\Phi^+}\frac{\Phi(t)}{\tau}\sim \Phi(t).$$ On the other hand, for any $t\in[0,\infty)$, we have $$\widetilde{\Phi}(t)=\int_{0}^{t}\inf_{\tau\in(0,s)}\frac{\Phi(\tau)}{\tau}\,ds \lesssim\frac{\Phi(t)}{t^{p_\Phi^-}}\int_{0}^{t}\inf_{\tau\in(0,s)}\frac{1}{\tau^{1-p_\Phi^-}}\,ds \sim\frac{\Phi(t)}{t^{p_\Phi^-}}\int_{0}^{t}\frac{1}{s^{1-p_\Phi^-}}\,ds\sim\Phi(t).$$ Thus, we obtain $\Phi\sim\widetilde{\Phi}$. Moreover, it is easy to prove that $\widetilde{\Phi}$ is an Orlicz function with positive lower type $p_{\Phi}^-$ and positive upper type $p_{\Phi}^+$ satisfying $p_{\Phi}^+\in(0,1]$, which completes the proof of Lemma \[djdj\]. \[concave\] Observe that all the results of this article are invariant under the change of equivalent Orlicz functions. By this and Lemma \[djdj\], without loss of generality, in this section, we may always assume that an Orlicz function with positive lower type $p_{\Phi}^-$ and positive upper type $p_{\Phi}^+$ satisfying $p_{\Phi}^+\in(0,1]$ is also concave. \[quasi\] Let $t\in(0,\infty)$, $q\in(0,1]$ and $\Phi$ be an Orlicz function with positive lower type $p_{\Phi}^-$ and positive upper type $p_{\Phi}^+\in(0,1]$. Then there exists a nonnegative constant $C$ such that, for any sequence $\{f_j\}_{j\in\nn}\subset (E_\Phi^{q})_t(\rn)$ of nonnegative functions such that $\sum_{j\in\mathbb{N}}f_j$ converges in $(E_\Phi^{q})_t(\rn),$ $$\lf\\|\sum_{j\in\mathbb{N}}f_j\r\\|_{(E_\Phi^{q})_t(\rn)} \ge C\sum_{j\in\mathbb{N}}\lf\\|f_j\r\\|_{(E_\Phi^{q})_t(\rn)}.$$ By Lemma \[djdj\], we know that there exists a concave function $\widetilde{\Phi}$ with same types of $\Phi$, which is equivalent to $\Phi$. Thus, for any $f\in(E_\Phi^{q})_t(\rn)$, $$\lf\\|f\r\\|_{(E_\Phi^{q})_t(\rn)}\sim\lf\\|f\r\\|_{(E_{\widetilde{\Phi}}^{q})_t(\rn)}$$ and, to prove this lemma, by the Levi theorem, we only need to show that, for any nonnegative $f_1,\ f_2\in(E_\Phi^{q})_t(\rn)$, $$\lf\\|f_1+f_2\r\\|_{(E_{\widetilde{\Phi}}^{q})_t(\rn)}\geq\lf\\|f_1\r\\|_{(E_{\widetilde{\Phi}}^{q})_t(\rn)}+ \lf\\|f_2\r\\|_{(E_{\widetilde{\Phi}}^{q})_t(\rn)}.$$ Fix $x\in\rn$ and let $a_1,\ a_2\in\rr$ satisfy $a_1 \in(0,\\|f_1\\|_{L^{\wz\Phi}(B(x,t))})$ and $a_2\in (0,\\|f_2\\|_{L^{\wz\Phi}(B(x,t))})$. Since $\widetilde{\Phi}$ is concave, it follows that $$\begin{aligned} \int_{B(x,t)}\widetilde{\Phi}\lf(\frac{f_1+f_2}{a_1+a_2}\r)\,dx &=\int_{B(x,t)}\widetilde{\Phi}\lf(\frac{f_1}{a_1}\frac{a_1}{a_1+a_2}+\frac{f_2}{a_2}\frac{a_2}{a_1+a_2}\r)\,dx\\ &\geq\frac{a_1}{a_1+a_2}\int_{B(x,t)}\widetilde{\Phi}\lf(\frac{f_1}{a_1}\r)\,dx +\frac{a_2}{a_1+a_2}\int_{B(x,t)}\widetilde{\Phi}\lf(\frac{f_2}{a_2}\r)\,dx\\ &>\frac{a_1}{a_1+a_2}+\frac{a_2}{a_1+a_2}=1.\end{aligned}$$ Thus, $$\lf\\|f_1+f_2\r\\|_{L^{\wz\Phi}(B(x,t))}\geq a_1+a_2,$$ which further implies that $$\lf\\|f_1+f_2\r\\|_{L^{\wz\Phi}(B(x,t))}\geq \lf\\|f_1\r\\|_{L^{\wz\Phi}(B(x,t))}+\lf\\|f_2\r\\|_{L^{\wz\Phi}(B(x,t))}.$$ From this and the definition of $(E_{\widetilde\Phi}^{q})_t(\rn)$, it easily follows that $$\begin{aligned} \lf\\|f_1+f_2\r\\|_{(E_{\widetilde{\Phi}}^{q})_t(\rn)}\geq\lf\\|f_1\r\\|_{(E_{\widetilde{\Phi}}^{q})_t(\rn)}+ \lf\\|f_2\r\\|_{(E_{\widetilde{\Phi}}^{q})_t(\rn)},\end{aligned}$$ which completes the proof of Lemma \[quasi\]. \[atom\] Let $t\in(0,\infty)$, $q\in(0,1]$ and $\Phi$ be an Orlicz function with positive lower type $p_{\Phi}^-$ and positive upper type $p_{\Phi}^+\in(0,1]$. Let $r\in(1,\infty]$, $s\in (0,\min\{p^-_{\Phi},q\})$ and $d\in\mathbb{Z}_+$ satisfying $d\ge\lfloor n(\frac{1}{s}-1)\rfloor$. Suppose $L$ is a continuous linear functional on $(HE_\Phi^q)_t(\rn)=(HE_\Phi^q)_t^{r,d}(\rn)$ . Then $$\begin{aligned} \\|L\\|_{((HE_\Phi^q)_t^{r,d}(\rn))^} :=&\sup\lf\{\|Lf\|:\ \\|f\\|_{(HE_\Phi^q)_t^{r,d}(\rn)}\leq1\r\}\\ \sim&\sup\lf\{\|La\|:\ a\ is\ an\ ((E_\Phi^q)_t(\rn),\ r,\ d)\text{-atom}\r\}\end{aligned}$$ with the equivalent positive constant independent of $L$ and $t$. Observing that any $((E_\Phi^q)_t(\rn),\ r,\ d)$-atom $a$ satisfies $\\|a\\|_{(HE_\Phi^q)_t^{r,d}(\rn)}\leq1$, to prove this lemma, we only need to show $$\label{85} \sup\lf\{\|Lf\|:\ \\|f\\|_{(HE_\Phi^q)_t^{r,d}(\rn)}\leq1\r\} \lesssim\sup\lf\{\|La\|:\ a\ \text{is a}\ ((E_\Phi^q)_t(\rn),\ r,\ d)\text{-atom}\r\}.$$ Take any $f\in(HE_\Phi^q)_t(\rn)$ and $\\|f\\|_{(HE_\Phi^q)_t^{r,d}(\rn)}\leq1$, which is reasonable by Theorem \[atom ch\]. By Definition \[atomic hardy\], we know that, for any $\epsilon\in(0,\infty)$, there exist a sequence of $((E_\Phi^q)_t(\rn),\ r,\ d)$-atoms and a sequence $\{\lambda_j\}_{j=1}^\infty\subset[0,\infty)$ such that $f=\sum_{j=1}^{\infty}\lambda_j a_j$ in $\mathcal{S}'(\rn)$ and $$\lf\\|\lf\{\sum_{j=1}^{\infty} \lf[\frac{\lambda_j} {\\|\chi_{Q_j}\\|_{(E_\Phi^q)_t(\rn)}} \r]^s\chi_{Q_j} \r\} ^{\frac{1}{s}}\r\\|_{(E_\Phi^q)_t(\rn)}\leq1+\epsilon.$$ Combining this and Lemma \[quasi\], we have $$\begin{aligned} \label{52} \sum_{j=1}^{\infty}\|\lambda_j\| &\lesssim\lf\\|\sum_{j=1}^{\infty} \frac{\lambda_j} {\\|\chi_{Q_j}\\|_{(E_\Phi^q)_t(\rn)}} \chi_{Q_j} \r\\|_{(E_\Phi^q)_t(\rn)}\\ \noz &\lesssim\lf\\|\lf\{\sum_{j=1}^{\infty} \lf[\frac{\lambda_j} {\\|\chi_{Q_j}\\|_{(E_\Phi^q)_t(\rn)}} \r]^s\chi_{Q_j} \r\} ^{\frac{1}{s}}\r\\|_{(E_\Phi^q)_t(\rn)} \lesssim1+\epsilon.\end{aligned}$$ Observe that, by Proposition \[shoulian\], we know that $f=\sum_{j=1}^{\infty}\lambda_j a_j$ holds true in $(HE_\Phi^q)_t(\rn)$. From this and , it follows that $$\|Lf\|\leq\sum_{j=1}^{\infty}\|\lambda_j\|\|La_j\|\lesssim(1+\epsilon) \sup\lf\{\|La\|:\ a\ \mathrm{is\ an}\ ((E_\Phi^q)_t(\rn),\ r,\ d)\text{-atom}\r\}.$$ Letting $\epsilon\rightarrow 0^+$, we then obtain , which completes the proof of Lemma \[atom\]. \[decamp\] Let $t\in(0,\infty)$, $q\in(0,1]$ and $\Phi$ be an Orlicz function with positive lower type $p_{\Phi}^-$ and positive upper type $p_{\Phi}^+\in(0,1]$. Let $r\in[1,\infty)$, $s\in(0,\min\{p^-_{\Phi},q\})$ and $d\in\mathbb{Z}_+$ satisfying $d\ge\lfloor n(\frac{1}{s}-1)\rfloor$. The Campanato space* $\mathcal{L}_{\Phi,t}^{q,r,d}(\rn)$ is defined to be the space of all locally $L^r(\rn)$ functions $g$ such that $$\\|g\\|_{\mathcal{L}_{\Phi,t}^{q,r,d}(\rn)}:=\sup_{B\subset\rn}\inf_{P\in\mathcal{P}_d(\rn)} \frac{\|B\|}{\\|\chi_{B}\\|_{(E_\Phi^q)_t(\rn)}}\lf[\frac{1}{\|B\|}\int_{B}\|g(x)-P(x)\|^r\,dx\r]^{\frac{1}{r}}<\infty,$$ where the first supremum is taken over all the balls $B\subset\rn$ and $\mathcal{P}_d(\rn)$ denotes the space of all polynomials on $\rn$ with order not greater than $d$. As usual, by a little abuse of notation, we identify $f\in\mathcal{L}_{\Phi,t}^{q,r,d}(\rn)$ with an equivalent class $f+\mathcal{P}_d(\rn)$. \[dual2\] Let $t\in(0,\infty)$, $q\in(0,1]$ and $\Phi$ be an Orlicz function with positive lower type $p_{\Phi}^-$ and positive upper type $p_{\Phi}^+\in(0,1]$. Let $r\in(1,\infty]$, $s \in(0,\min\{p^-_{\Phi},q\})$ and $d\in\mathbb{Z}_+$ satisfying $d\ge\lfloor n(\frac{1}{s}-1)\rfloor$. Then the dual space of $(HE_\Phi^q)_t(\rn)$, denoted by $((HE_\Phi^q)_t(\rn))^$, is $\mathcal{L}_{\Phi,t}^{q,r',d}(\rn)$ in the following sense: 1. Any $g\in\mathcal{L}_{\Phi,t}^{q,r',d}(\rn)$ induces a linear functional given by $$\label{linf} L_g:\ f\mapsto\ L_g(f):=\int_{\rn}f(x)g(x)dx,$$ which is initially defined on $(HE_\Phi^{q,r,d})_t^{\fin}(\rn)$ and has a bounded extension to $(HE_\Phi^q)_t(\rn)$. 2. Conversely, any continuous linear functional on $(HE_\Phi^q)_t(\rn)$ is of the form for a unique $g\in\mathcal{L}_{\Phi,t}^{q,r',d}(\rn)$. Moreover, in any case, $\\|g\\|_{\mathcal{L}_{\Phi,t}^{q,r',d}(\rn)}$ is equivalent to $\\|L_g\\|_{((HE_\Phi^q)_t(\rn))^}$ with the equivalent positive constants independent of $t$, here and hereafter, $\\|\cdot\\|_{((HE_\Phi^q)_t(\rn))^}$ denotes the norm of $((HE_\Phi^q)_t(\rn))^$. Let $t\in(0,\infty)$. - Let $q\in(0,1],\ r\in[1,\infty)$ and $\Phi(\tau):=\tau^q$ for any $\tau\in[0,\infty)$. In this case, via some simple computations, we know that, for any ball $B\subset\rn$, $\\|\chi_B\\|_{(E_\Phi^q)_t(\rn)}=\|B\|^\frac1q$. Thus, in this case, $\mathcal{L}_{\Phi,t}^{q,r,d}(\rn)$ coincides with the classical Campanato space $L_{\frac1q-1,r,d}(\rn)$ which was introduced by Campanato [@c]. - Let $q\in(0,1)$ and $\Phi(\tau):=\tau^q$ for any $\tau\in[0,\infty)$. In this case, we have $p^{-}_\Phi=q$, $(E_\Phi^q)_t(\rn)=L^q(\rn)$ and $(HE_\Phi^q)_t(\rn)=H^q(\rn)$, and the best known range of $r$ in Theorem \[dual2\] is $[1,\infty]$ (see, for example, [@l Theorem 4.1]). However, it is still unclear whether or not Theorem \[glamda\] still holds true when $r=1$ and $\max\{p_{\Phi}^+,\ q\}\in(0,1)$. \[bbbb\] Let $t,\ q\in(0,\infty)$ and $\Phi$ be an Orlicz function with positive lower type $p_{\Phi}^-$ and positive upper type $p_{\Phi}^+$. Let $x_0\in\rn$ and $r\in(0,\infty)$. Then $Q(x_0,\frac{2r}{\sqrt{n}})\subset B(x_0,r)\subset Q(x_0,2r)$ and there exists a positive constant $C$, independent of $t$, $x_0$ and $r$, such that $$\lf\\|\chi_{Q(x_0,\frac{2r}{\sqrt{n}})}\r\\|_{(E_\Phi^q)_t(\rn)}\leq \lf\\|\chi_{B(x_0,r)}\r\\|_{(E_\Phi^q)_t(\rn)}\leq\lf\\|\chi_{Q(x_0,2r)}\r\\|_{(E_\Phi^q)_t(\rn)} \leq C\lf\\|\chi_{Q(x_0,\frac{2r}{\sqrt{n}})}\r\\|_{(E_\Phi^q)_t(\rn)}.$$ Obviously, for any $x_0\in\rn$ and $r\in(0,\infty)$, we have $$Q\lf(x_0,\frac{2r}{\sqrt{n}}\r)\subset B\lf(x_0,r\r)\subset Q\lf(x_0,2r\r)$$ and $$\lf\\|\chi_{Q(x_0,\frac{2r}{\sqrt{n}})}\r\\|_{(E_\Phi^q)_t(\rn)}\leq \lf\\|\chi_{B(x_0,r)}\r\\|_{(E_\Phi^q)_t(\rn)}\leq\lf\\|\chi_{Q(x_0,2r)}\r\\|_{(E_\Phi^q)_t(\rn)}.$$ Thus, to complete the proof of Lemma \[bbbb\], we only need to show that, for any $x_0\in\rn$ and $r\in(0,\infty)$, $$\lf\\|\chi_{Q(x_0,2r)}\r\\|_{(E_\Phi^q)_t(\rn)}\lesssim\lf\\|\chi_{Q(x_0,\frac{2r}{\sqrt{n}})}\r\\|_{(E_\Phi^q)_t(\rn)}.$$ Assume that $r\in(0,\infty)$ and, without loss of generality, we may assume that $x_0=\vec{0}_n$. Then it is easy to see that there exist $M\in\mathbb{N}$ and $\{x_1, \ldots, x_M\}\subset\rn$, independent of $t$ and $r$, such that $M\lesssim1$ and $Q(\vec{0}_n,2r)\subseteq\bigcup_{m=1}^MQ(x_m,\frac{2r}{\sqrt{n}})$, which implies that $$\label{1111} \lf\\|\chi_{Q(\vec{0}_n,2r)}\r\\|_{(E_\Phi^q)_t(\rn)}\leq \lf\\|\sum_{m=1}^M\chi_{Q(x_m,\frac{2r}{\sqrt{n}})}\r\\|_{(E_\Phi^q)_t(\rn)}\lesssim \sum_{m=1}^M\lf\\|\chi_{Q(x_m,\frac{2r}{\sqrt{n}})}\r\\|_{(E_\Phi^q)_t(\rn)}.$$ Observing that, for any $m\in\mathbb{N}$, $t\in(0,\infty)$ and $x\in\rn$, $$\lf\\|\chi_{Q(x_m,\frac{2r}{\sqrt{n}})}\chi_{B(x,t)}\r\\|_{L^\Phi(\rn)} =\lf\\|\chi_{B(\vec{0}_n,\frac{2r}{\sqrt{n}})}\chi_{B(x-x_m,t)}\r\\|_{L^\Phi(\rn)},$$ by this and with $\widetilde{C}_{(\Phi,t)}$ as therein, we have $$\begin{aligned} \lf\\|\chi_{Q(x_m,\frac{2r}{\sqrt{n}})}\r\\|_{(E_\Phi^q)_t(\rn)} &=\lf\{\int_{\rn}\lf[\frac{\\|\chi_{Q(\vec{0}_n,\frac{2r}{\sqrt{n}})}\chi_{B(x-x_m,t)}\\|_{L^\Phi(\rn)}} {\\|\chi_{B(x,t)}\\|_{L^\Phi(\rn)}}\r]^q\,dx\r\}^{\frac{1}{q}}\\ &=\frac{1}{\widetilde{C}_{(\Phi,t)}}\lf\{\int_{\rn}\lf\\| \chi_{Q(\vec{0}_n,\frac{2r}{\sqrt{n}})}\chi_{B(x-x_m,t)}\r\\|_{L^\Phi(\rn)} ^q\,dx\r\}^{\frac{1}{q}}\\ &=\frac{1}{\widetilde{C}_{(\Phi,t)}}\lf\{\int_{\rn}\lf\\| \chi_{Q(\vec{0}_n,\frac{2r}{\sqrt{n}})}\chi_{B(x,t)}\r\\|_{L^\Phi(\rn)} ^q\,dx\r\}^{\frac{1}{q}} =\lf\\|\chi_{Q(\vec{0}_n,\frac{2r}{\sqrt{n}})}\r\\|_{(E_\Phi^q)_t(\rn)},\end{aligned}$$ which, combined with $\eqref{1111}$, implies that $\\|\chi_{Q(\vec{0}_n,2r)}\\|_{(E_\Phi^q)_t(\rn)}\lesssim\\|\chi_{Q(\vec{0}_n,\frac{2r}{\sqrt{n}})} \\|_{(E_\Phi^q)_t(\rn)}.$ Thus, for any $x_0\in\rn$ and $r\in(0,\infty)$, $$\lf\\|\chi_{Q(x_0,2r)}\r\\|_{(E_\Phi^q)_t(\rn)}\lesssim\lf\\|\chi_{Q(x_0,\frac{2r}{\sqrt{n}})}\r\\|_{(E_\Phi^q)_t(\rn)}.$$ This finishes the proof Lemma \[bbbb\]. We first show (i). By Theorem \[atom ch\], to prove $\mathcal{L}_{\Phi,t}^{q,r',d}(\rn)\subset((HE_\Phi^q)_t(\rn))^$, it suffices to show $$\mathcal{L}_{\Phi,t}^{q,r',d}(\rn)\subset((HE_\Phi^q)_t^{r,d}(\rn))^.$$ Let $g\in\mathcal{L}_{\Phi,t}^{q,r',d}(\rn)$ and $a$ be an $((E_\Phi^q)_t(\rn),\ r,\ d)$-atom supported on a cube $Q\subset\rn$. Let the ball $B\subset\rn$ such that $Q\subset B$ and $\|Q\|\sim\|B\|$. Then, by the moment and the size conditions of $a$, together with the Hölder inequality and Lemma \[bbbb\], we know that $$\begin{aligned} \|L_g(a)\|:&=\lf\|\int_{\rn}a(x)g(x)\,dx\r\| =\inf_{P\in \cp_d(\rn)}\lf\|\int_{\rn}a(x)\lf[g(x)-P(x)\r]\,dx\r\|\\ &\leq\\|a\\|_{L^r(\rn)}\inf_{P\in \cp_d(\rn)}\lf[\int_{\rn}\|g(x)-P(x)\|^{r'}\,dx\r]^{\frac{1}{r'}}\\ &\leq\frac{\|Q\|^{\frac{1}{r}}}{\\|\chi_{Q}\\|_{(E_\Phi^q)_t(\rn)}} \inf_{P\in \cp_d(\rn)}\lf[\int_{\rn}\|g(x)-P(x)\|^{r'}\,dx\r]^{\frac{1}{r'}}\\ &\sim\frac{\|B\|^{\frac{1}{r}}}{\\|\chi_{B}\\|_{(E_\Phi^q)_t(\rn)}} \inf_{P\in \cp_d(\rn)}\lf[\int_{\rn}\|g(x)-P(x)\|^{r'}\,dx\r]^{\frac{1}{r'}} \lesssim\\|g\\|_{\mathcal{L}_{\Phi,t}^{q,r',d}(\rn)}.\end{aligned}$$ Moreover, for any $f\in(HE_\Phi^{q,r,d})_t^{\fin}(\rn)$, by Definition \[definite\], we know that there exist a sequence $\{a_j\}_{j=1}^m$ of $((E_\Phi^q)_t(\rn),\ r,\ d)$-atoms supported, respectively, on the cubes $\{Q_j\}_{j=1}^m$ and a sequence $\{\lambda_j\}_{j=1}^m\subset[0,\infty)$ such that $$\lf\\|\lf\{\sum_{j=1}^{m} \lf[\frac{\lambda_j} {\\|\chi_{Q_j}\\|_{(E_\Phi^q)_t(\rn)}} \r]^s\chi_{Q_j} \r\} ^{\frac{1}{s}}\r\\|_{(E_\Phi^q)_t(\rn)} \lesssim\\|f\\|_{(HE_\Phi^{q,r,d})_t^{\fin}(\rn)}.$$ From this and Lemma \[quasi\], it follows that $$\begin{aligned} \|L_g(f)\|:&=\lf\|\int_{\rn}f(x)g(x)\,dx\r\| \leq\sum_{j=1}^{m}\lambda_j\lf\|\int_{\rn}a_j(x)g(x)\,dx\r\|\\ &\lesssim\sum_{j=1}^{m}\lambda_j\\|g\\|_{\mathcal{L}_{\Phi,t}^{q,r',d}(\rn)} \lesssim\lf\\|\sum_{j=1}^{m} \frac{\lambda_j} {\\|\chi_{Q_j}\\|_{(E_\Phi^q)_t(\rn)}} \chi_{Q_j} \r\\|_{(E_\Phi^q)_t(\rn)}\\|g\\|_{\mathcal{L}_{\Phi,t}^{q,r',d}(\rn)}\\ &\lesssim\lf\\|\lf\{\sum_{j=1}^{m} \lf[\frac{\lambda_j} {\\|\chi_{Q_j}\\|_{(E_\Phi^q)_t(\rn)}} \r]^s\chi_{Q_j} \r\} ^{\frac{1}{s}}\r\\|_{(E_\Phi^q)_t(\rn)} \\|g\\|_{\mathcal{L}_{\Phi,t}^{q,r',d}(\rn)}\\ &\lesssim\\|f\\|_{(HE_\Phi^{q,r,d})_t^{\fin}(\rn)} \\|g\\|_{\mathcal{L}_{\Phi,t}^{q,r',d}(\rn)}.\end{aligned}$$ By this and the fact that $(HE_\Phi^{q,r,d})_t^{\fin}(\rn)$ is dense in $(HE_\Phi^q)_t(\rn)$ as well as Theorem \[finite\], we obtain Theorem \[dual2\](i). As for (ii), for any ball $B\subset\rn$, let $\pi_B :\ L^1(B)\rightarrow \mathcal{P}_d(\rn)$ be the natural projection such that, for any $f\in L^1(B)$ and $Q\in\mathcal{P}_d(\rn)$, $$\int_{B}\pi_B(f)(x)Q(x)\,dx=\int_{B}f(x)Q(x)\,dx.$$ It is well known (see, for example, [@b p.51,(8.9)] or [@l p.54,Lemma 4.1]) that $$\label{5.3} \sup_{x\in B}\|\pi_Bf(x)\|\lesssim \frac{1}{\|B\|}\int_{B}\|f(x)\|\,dx.$$ For any $r\in(1,\infty]$ and ball $B\subset\rn$, we define the closed subspace $L^r_0(B)$ of $L^r(B)$ by setting $$L^r_0(B): = \lf\{f\in L^r(B):\ \pi_Bf = 0\r\}.$$ Notice that $L^r(B)$ is the subspace of $L^r(\rn)$ consisting of all measurable functions vanishing outside $B$. Thus, if $f\in L^r_0(B)$, then $\frac{\|Q\|^{\frac{1}{r }}}{\\|\chi_Q\\|_{(E_\Phi^q)_t(\rn)}}\\|f\\|_{L^r(\rn)}^{-1}f$ is an $((E_\Phi^q)_t(\rn),\ r,\ d)$-atom, where $Q$ is a cube, $Q\supset B$ and the side length of $Q$ equals to 2 times radius of $B$. Suppose now $L \in((HE_\Phi^q)_t(\rn))^* = ((HE_\Phi^q)_t^{r,d}(\rn))^$. By Lemma \[atom\], we know that, for any $f\in L^r_0(B)$, $$\label{810} \|L(f)\| \lesssim\frac{\\|\chi_Q\\|_{(E_\Phi^q)_t(\rn)}}{\|Q\|^{\frac{1}{r }}} \\|L\\|_{((HE_\Phi^q)_t(\rn))^}\\|f\\|_{L^r(\rn)}.$$ Therefore, $L$ provides a bounded linear functional on $L^r_0(B)$. Thus, by the Hahn-Banach theorem (see, for example, [@y p.106, Theorem 1]), we know that there exists a linear functional $L_B$, which extends $L$ to the whole space $L^r(B)$ without increasing its norm. If $r\in(1,\infty)$, by the duality $(L^r(B))^* = L^{r'}(B)$, we find that there exists $h_B \in L^{r'}(B)\subset L^{1}(B) $ such that, for any $f\in L^r_0(B)$, $$L(f)=L_B(f)=\int_{B}f(x)h_B(x)\,dx.$$ For the case $r =\infty$, let $\widetilde{r}\in(1,\infty)$. By Theorems \[atom ch\], we know that $L \in((HE_\Phi^q)_t^{\infty,d}(\rn))^$ implies $L\in((HE_\Phi^q)_t^{\widetilde{r},d}(\rn))^$ without changing the norm of $L$. Thus, there exists $h_B \in L^{\widetilde{r}'}(B)\subset L^{1}(B)$ such that, for any $f\in L^\infty_0(B)$, $L(f)=\int_{B}f(x)h_B(x)\,dx$. Altogether, we find that, for any $r\in(1,\infty]$, there exists $h_B \in L^{r'}(B)$ such that, for any $f\in L^r_0(B)$, $L(f)=\int_{B}f(x)h_B(x)\,dx.$ Next we show that such $h_B$ is unique modulo $\cp_d(\rn)$. Indeed, assume that $h_B'$ is another element of $L^{r'}(B)$ such that $L(f)=\int_{B}f(x)h_B'(x)\,dx$ for any $f\in L^r_0(B)$. Then, for any $f\in L^\infty(B)$, we have $f-\pi_B(f)\in L^\infty_0 (B)$ and $$\begin{aligned} 0&=\int_B[f(x)-\pi_B(f)(x)][h_B(x)-h_B'(x)]\,dx\\ &=\int_Bf(x)[h_B(x)-h_B'(x)]\,dx-\int_B\pi_B(f)(x)\pi_B(h_B - h_B')(x)\,dx\\ &=\int_Bf(x)[h_B(x)-h_B'(x)]\,dx-\int_Bf(x)\pi_B(h_B-h_B')(x)\,dx\\ &=\int_B f(x)[h_B(x)-h_B'(x)-\pi_B(h_B - h_B')(x)]\,dx.\end{aligned}$$ The arbitrariness of $f$ implies that $h_B(x)-h_B'(x) = \pi_B(h_B - h_B')(x)$ for almost every $x\in B$. Therefore, after changing values of $h_B$ (or $h_B'$) on a set of measure zero, we have $h_B - h_B'\in \cp_d(\rn)$. Thus, the function $h_B$ is unique up to a polynomial of degree at most $d$ regardless of the exponent $r\in(1,\infty]$. For any $\rho\in\mathbb{N}$, let $g_\rho$ be the unique element of $L^{r'}(B(\vec 0_n,\rho))$ such that $L(f) =\int_{B(\vec 0_n,\rho)}f(x)g_\rho(x)\,dx$ for any $f\in L^r_0(B(\vec 0_n,\rho))$. The preceding arguments show that $g_\rho\|_{B(\vec 0_n,\ell)} = g_\ell$ for any $\ell\in\{1, \ldots, \rho\}$. Therefore, we can define a locally $L^{r'}(\rn)$ function $g$ by setting $g(x):= g_\rho(x)$ whenever $x\in B(\vec{0}_n,\rho)$. If $f$ is a finite linear combination of $((E_\Phi^q)_t(\rn),\ r,\ d)$-atoms, then $L(f) =\int_{\rn}f(x)g(x)\,dx$. By , for any ball $B\subset\rn$, the norm of $g$ as a linear functional on $L^r_0(B)$ satisfies $$\label{811} \\|g\\|_{(L^r_0(B))^} \leq\frac{\\|\chi_Q\\|_{(E_\Phi^q)_t(\rn)}}{\|Q\|^{\frac{1}{r }}} \\|L\\|_{((HE_\Phi^q)_t(\rn))^}.$$ It is known (see [@b p.52, (8.12)]) that $$\label{812} \\|g\\|_{(L^r_0(B))^}=\inf_{P\in \cp_d(\rn)}\\|g-P\\|_{L^{r'}(B)}.$$ Combining , and Lemma \[bbbb\], we have $$\\|g\\|_{\mathcal{L}_{\Phi,t}^{q,r',d}(\rn)}\leq\sup_{B\subset\rn} \frac{\|B\|^{\frac{1}{r}}}{\\|\chi_{B}\\|_{(E_\Phi^q)_t(\rn)}}\\|g\\|_{(L^r_0(B))^} \leq\\|L\\|_{((HE_\Phi^q)_t^{r,d}(\rn))^}.$$ This finishes the proof of (ii) and hence of Theorem \[dual2\]. Applications\[s6\] ================== In this section, we first establish a criterion on the boundedness of sublinear operators from $(HE_\Phi^q)_t(\rn)$ into a quasi-Banach space as an application of the finite atomic characterizations of $(HE_\Phi^q)_t(\rn)$ from Theorem \[finite\]. Then we clarify the relationship between the atomic space $(\mathfrak{C}_r^q)_t$ introduced in [@ap] and the Orlicz-slice Hardy space $(HE_\Phi^q)_t(\rn)$. As an application of the above boundedness criterion, we obtain the boundedness of $\delta$-type Calderón-Zygmund operators on $(HE_\Phi^q)_t(\rn)$. Boundedness of sublinear operators\[s5.1\] ------------------------------------------ The main purpose of this section is to establish a criterion on the boundedness of sublinear operators from $(HE_\Phi^q)_t(\rn)$ into a quasi-Banach space. Recall that a complete vector space is called a quasi-Banach space* $\mathcal{B}$ if its quasi-norm $\\|\cdot\\|_{\mathcal{B}}$ satisfies - $\\|f\\|_{\mathcal{B}}=0$ if and only if $f$ is the zero element of $\mathcal{B}$; - there exists a positive constant $C\in[1,\infty)$ such that, for any $f,\ g\in\mathcal{B}$, $$\\|f + g\\|_{\mathcal{B}} \leq C(\\|f \\|_{\mathcal{B}} + \\|g\\|_{\mathcal{B}}).$$ Obviously, when $C=1$, a quasi-Banach space $\mathcal{B}$ is just a Banach space. Next we recall the notion of $\gamma$-quasi-Banach spaces (see, for example, [@k], [@yll], [@yz] and [@yz2]). Let $\gamma\in(0,1]$. A quasi-Banach space $\mathcal{B}_\gamma$ with the quasi-norm $\\|\cdot\\|_{\mathcal{B}_\gamma}$ called a $\gamma$-quasi-Banach space if there exists a positive constant $\kappa\in[1,\infty)$ such that, for any $m\in\mathbb{N}$ and $\{f_j\}_{j=1}^{m}\subset\mathcal{B}_\gamma$, $$\lf\\|\sum_{j=1}^{m}f_j\r\\|_{\mathcal{B}_\gamma}^\gamma \le\kappa\sum_{j=1}^{m}\lf\\|f_j\r\\|^\gamma_{\mathcal{B}_\gamma}.$$ For any given $\gamma$-quasi-Banach space $\mathcal{B}_\gamma$, with $\gamma\in(0,1]$, and a linear space $\mathcal{V}$, an operator $T$ from $\mathcal{V}$ to $\mathcal{B}_\gamma$ is said to be $\mathcal{B}_\gamma$-sublinear if there exists a positive constant $\kappa\in[1,\infty)$ such that 1. for any $f,\ g\in\mathcal{V}$, $\\|T(f)-T(g)\\|_{\mathcal{B}_\gamma}\le\kappa\\|T(f-g)\\|_{\mathcal{B}_\gamma}$; 2. for any $m\in\mathbb{N}$, $\{f_j\}_{j=1}^{m}\subset\mathcal{V}$ and $\{\lambda_j\}_{j=1}^{m}\subset\mathbb{C}$, $$\lf\\|T\lf(\sum_{j=1}^{m}\lambda_jf_j\r)\r\\|_{\mathcal{B}_\gamma}^\gamma\le \kappa\sum_{j=1}^{m}\|\lambda_j\|^{\gamma}\\|T(f_j)\\|_{\mathcal{B}_\gamma}^\gamma.$$ \[suanzi1\] Let $t,\ q\in(0,\infty)$, $\gamma\in(0,1]$, $\Phi$ be an Orlicz function with positive lower type $p_{\Phi}^-$ and positive upper type $p_{\Phi}^+$ and $\mathcal{B}_\gamma$ a $\gamma$-quasi-Banach space. Let $r\in(\max\{1,\ q,\ p_{\Phi}^+\},\infty]$, $s\in (0,\min\{p^-_{\Phi},q,1\})$ and $d\in\mathbb{Z}_+$ satisfying $d\ge\lfloor n(\frac{1}{s}-1)\rfloor$. If either of the following two statements holds true: 1. $r\in(\max\{1,\ q,\ p_{\Phi}^+\},\infty)$ and $T:\ (HE_\Phi^{q,r,d})_t^{\fin}(\rn)\rightarrow\mathcal{B}_\gamma$ is a $\mathcal{B}_\gamma$-sublinear operator satisfying that there exists a positive constant $C_1$ such that, for any $f\in(HE_\Phi^{q,r,d})_t^{\fin}(\rn)$ $$\label{61} \\|T(f)\\|_{\mathcal{B}_\gamma}\le C_1\\|f\\|_{(HE_\Phi^{q,r,d})_t^{\fin}(\rn)};$$ 2. $T:\ (HE_\Phi^{q,\infty,d})_t^{\fin}(\rn)\cap\mathcal{C}(\rn)\rightarrow\mathcal{B}_\gamma$ is a $\mathcal{B}_\gamma$-sublinear operator satisfying that there exists a positive constant $C_2$ such that, for any $f\in(HE_\Phi^{q,\infty,d})_t^{\fin}(\rn)\cap\mathcal{C}(\rn)$ $$\\|T(f)\\|_{\mathcal{B}_\gamma}\le C_2\\|f\\|_{(HE_\Phi^{q,\infty,d})_t^{\fin}(\rn)},$$ then $T$ uniquely extends to a bounded $\mathcal{B}_\gamma$-sublinear operator from $(HE_\Phi^q)_t(\rn)$ into $\mathcal{B}_\gamma$. Moreover, there exists a positive constant $\widetilde{C}$ such that, for any $f\in(HE_\Phi^q)_t(\rn)$, $$\\|T(f)\\|_{\mathcal{B}_\gamma}\le\widetilde{C}\\|f\\|_{(HE_\Phi^q)_t(\rn)}.$$ By Theorem \[suanzi1\], we easily obtain the following corollary, which is a variant of Meda et al. [@msv Corollary 3.4] and Grafakos et al. [@gly Theorem 5.9] as well as Ky [@k Theorem 3.5] (see also [@yll Theorem 1.6.9]), the details being omitted. \[suanzi\] Let $t\in(0,\infty)$, $q\in(0,1]$, $\gamma\in(0,1]$, $\Phi$ be an Orlicz function with positive lower type $p_{\Phi}^-$ and positive upper type $p_{\Phi}^+\in(0,1]$ and $\mathcal{B}_\gamma$ a $\gamma$-quasi-Banach space. Let $r\in(1,\infty]$, $s\in (0,\min\{p^-_{\Phi},q\})$ and $d\in\mathbb{Z}_+$ satisfying $d\ge\lfloor n(\frac{1}{s}-1)\rfloor$. If either of the following two statements holds true: 1. $r\in(1,\infty)$ and $T:\ (HE_\Phi^{q,r,d})_t^{\fin}(\rn)\rightarrow\mathcal{B}_\gamma$ is a $\mathcal{B}_\gamma$-sublinear operator satisfying $$A:=\sup\lf\{\\|Ta\\|_{\mathcal{B}_\gamma}:\ a\ is\ a\ ((E_\Phi^q)_t(\rn),\ r,\ d)\text{-atom}\r\}<\infty;$$ 2. $T:\ (HE_\Phi^{q,\infty,d})_t^{\fin}(\rn)\bigcap\mathcal{C}(\rn)\rightarrow\mathcal{B}_\gamma$ is a $\mathcal{B}_\gamma$-sublinear operator satisfying $$A:=\sup\lf\{\\|Ta\\|_{\mathcal{B}_\gamma}:\ a\ is\ a\ continuous\ ((E_\Phi^q)_t(\rn),\ \infty,\ d)\text{-atom}\r\}<\infty,$$ then $T$ uniquely extends to a bounded $\mathcal{B}_\gamma$-sublinear operator from $(HE_\Phi^q)_t(\rn)$ into $\mathcal{B}_\gamma$. Moreover, there exists a positive constant $\widetilde{C}$ such that, for any $f\in(HE_\Phi^q)_t(\rn)$, $$\\|T(f)\\|_{\mathcal{B}_\gamma}\le\widetilde{C}\\|f\\|_{(HE_\Phi^q)_t(\rn)}.$$ \[remax2\] Let $t\in(0,\infty)$, $q\in(0,1]$ and $\Phi(\tau):=\tau^q$ for any $\tau\in[0,\infty)$. In this case, we have $(E_\Phi^q)_t(\rn)=L^q(\rn)$ and $(HE_\Phi^q)_t(\rn)=H^q(\rn)$ and Theorem \[suanzi\] coincides with the well-known criterion on the boundedness of sublinear operators from $H^q(\rn)$ into a quasi-Banach space except the case $r=1$ (see, for example, [@k Theorem 3.5], [@yll Theorem 1.6.9] and [@gly Theorem 5.9] with $X:=\rn$). Moreover, when $q=1$, Theorem \[suanzi\] is just [@msv Corollary 3.4]. We now prove Theorem \[suanzi1\]. Suppose that assumption (i) holds true and $f\in(HE_\Phi^q)_t(\rn)$. Then, by the density of $(HE_\Phi^{q,r,d})_t^{\fin}(\rn)$ in $(HE_\Phi^q)_t(\rn)$, we know that there exists a Cauchy sequence $\{f_k\}_{k\in\mathbb{N}}\subset(HE_\Phi^{q,r,d})_t^{\fin}(\rn)$ such that $$\lim_{k\rightarrow\infty}\\|f_k-f\\|_{(HE_\Phi^q)_t(\rn)}=0.$$ By this, (\[61\]) and Theorem \[finite\](i), we conclude that, for any $k,\ l\in\mathbb{N}$, as $k,\ l\rightarrow\infty$, $$\lf\\|T(f_k)-T(f_l)\r\\|_{\mathcal{B}_\gamma}\lesssim\\|T(f_k-f_l)\\|_{\mathcal{B}_\gamma} \lesssim\\|f_k-f_l\\|_{(HE_\Phi^{q,r,d})_t^{\fin}(\rn)}\sim\\|f_k-f_l\\|_{(HE_\Phi^q)_t(\rn)}\rightarrow0,$$ which implies that $\{T(f_k)\}_{k\in\mathbb{N}}$ is a Cauchy sequence in $\mathcal{B}_\gamma$. Therefore, by the completeness of $\mathcal{B}_\gamma$, we know that there exists some $g\in\mathcal{B}_\gamma$ such that $g=\lim_{k\rightarrow\infty}T(f_k)$ in $\mathcal{B}_\gamma$. Then let $T(f):=g$. From this, (\[61\]) and Theorem \[finite\](i) again, it is easy to deduce that $T(f)$ is well defined and, for any $f\in(HE_\Phi^q)_t(\rn)$, $$\begin{aligned} \\|T(f)\\|_{\mathcal{B}_\gamma}^\gamma&\lesssim \limsup_{k\rightarrow\infty}\lf[\\|T(f)-T(f_k)\\|_{\mathcal{B}_\gamma}^\gamma+\\|T(f_k)\\|_{\mathcal{B}_\gamma}^\gamma\r] \lesssim \limsup_{k\rightarrow\infty}\\|T(f_k)\\|_{\mathcal{B}_\gamma}^\gamma\\ &\lesssim\limsup_{k\rightarrow\infty}\\|f_k\\|_{(HE_\Phi^{q,r,d})_t^{\fin}(\rn)}^\gamma \sim\lim_{k\rightarrow\infty}\\|f_k\\|_{(HE_\Phi^q)_t(\rn)}^\gamma\sim\\|f\\|_{(HE_\Phi^q)_t(\rn)}^\gamma,\end{aligned}$$ which completes the proof of (i). Suppose that the assumption (ii) holds true. Similarly to the proof of (i), using Theorem \[finite\](ii), we also conclude that, for any $f\in(HE_\Phi^{q,\infty,d})_t^{\fin}(\rn)\cap\mathcal{C}{(\rn)}$, $\\|T(f)\\|_{\mathcal{B}_\gamma}\lesssim\\|f\\|_{(HE_\Phi^q)_t(\rn)}$. To extend $T$ to the whole $(HE_\Phi^q)_t(\rn)$, we only need to prove that $(HE_\Phi^{q,\infty,d})_t^{\fin}(\rn)\cap\mathcal{C}{(\rn)}$ is dense in $(HE_\Phi^q)_t(\rn)$. Observing that $(HE_\Phi^{q,\infty,d})_t^{\fin}(\rn)$ is dense in $(HE_\Phi^q)_t(\rn)$, to show this, it suffices to prove that $(HE_\Phi^{q,\infty,d})_t^{\fin}(\rn)\cap\mathcal{C}{(\rn)}$ is dense in $(HE_\Phi^{q,\infty,d})_t^{\fin}(\rn)$ in terms of the quasi-norm $\\|\cdot\\|_{(HE_\Phi^q)_t(\rn)}$. Actually, we show that $(HE_\Phi^{q,\infty,d})_t^{\fin}(\rn)\cap\mathcal{C}^{\infty}{(\rn)}$ is dense in $(HE_\Phi^{q,\infty,d})_t^{\fin}(\rn)$. To see this, let $f\in(HE_\Phi^{q,\infty,d})_t^{\fin}(\rn)$. Since $f$ is a finite linear combination of functions with bounded supports, it follows that there exists $R\in(0,\infty)$ such that $\supp(f)\subset B(\vec{0}_n,R)$. Take $\varphi\in\mathcal{S}(\rn)$ such that $\supp(\varphi)\subset B(\vec{0}_n,1)$ and $\int_\rn\varphi(x)\,dx=1$. It is easy to see that $\supp(\varphi_\tau\ast f)\subset B(\vec{0}_n,2R)$ for any $\tau\in(0,R)$ and $\varphi_\tau\ast f$ has vanishing moments up to order $d$, where $\varphi_\tau(x):=\tau^{-n}\varphi(\tau^{-1}x)$ for any $x\in\rn$. Thus, $\varphi_\tau\ast f\in (HE_\Phi^{q,\infty,d})_t^{\fin}(\rn)\cap\mathcal{C}^{\infty}{(\rn)}$. Likewise, $\supp( f-\varphi_\tau\ast f)\subset B(\vec{0}_n,2R)$ for any $\tau\in(0,R)$ and $\varphi_\tau\ast f$ has vanishing moments up to order $d$. Moreover, taking any $\delta\in(1,\infty)$, we have $$\lf\\|f-\varphi_\tau\ast f\r\\|_{L^\delta(\rn)}\rightarrow0\quad\mathrm{as}\quad\tau\rightarrow0.$$ Thus, $f-\varphi_\tau\ast f=c_\tau a_\tau$ for some $((E_\Phi^q)_t(\rn),\ \delta,\ d)$-atom $a_\tau$, and some constant $c_\tau$ which satisfies that $c_\tau\rightarrow0$ as $\tau\rightarrow0$. Thus, $\\|f-\varphi_\tau\ast f\\|_{(HE_\Phi^q)_t(\rn)}\rightarrow0$ as $\tau\rightarrow0$. This finishes the proof of Theorem \[suanzi1\]. Boundedness of Calderón-Zygmund operators\[s5.2\] ------------------------------------------------- In [@ap], Auscher and Prisuelos-Arribas obtained the boundedness on slice spaces $(E_r^q)_t(\rn)$ of operators such as the Hardy-Littlewood maximal operator, Calderón-Zygmund operators etc. Based on $(E_r^q)_t(\rn)$, Auscher and Prisuelos-Arribas in [@ap] also introduced a Hardy-type space $(\mathfrak{C}_r^q)_t(\rn)$ and proved the boundedness of Calderón-Zygmund operators on it. In this section, we first obtain the relationship between the space $(\mathfrak{C}_r^q)_t(\rn)$ and the Orlicz-slice Hardy space $(HE_\Phi^q)_t(\rn)$. Then, using the criterion for the boundedness of sublinear operators obtained in Theorem \[suanzi\], we establish the boundedness of $\delta$-type Calderón-Zygmund operators from $(HE_\Phi^q)_t(\rn)$ to $(HE_\Phi^q)_t(\rn)$ \[or to $(E_\Phi^q)_t(\rn)$\] with $\delta\in(0,1]$ and $\min\{p_{\Phi}^+,q\}\in(\frac{n}{n+\delta},1]$, respectively. \[auscher\] Let $t\in(0,\infty)$, $r\in(1,\infty)$ and $q\in(\frac{n}{n+1},1]$. A function $a\in(E_r^q)_t(\rn)$ is called an $(E_r^q)_t(\rn)$-atom if it is supported on a ball $B$ of radius $\tau\in[t,\infty)$ and satisfies $$\\|a\\|_{L^r(\rn)}\le\|B\|^{\frac{1}{r}-\frac{1}{q}}.$$ The Hardy-type space $(\mathfrak{C}_r^q)_t(\rn)$ is then defined to be the set of all measurable functions $f\in(E_r^q)_t(\rn)$ such that there exist a sequence of numbers, $\{\lambda_j\}_{j=1}^\infty\in \ell^q$, and a sequence of $(E_r^q)_t(\rn)$-atoms, $\{a_j\}_{j=1}^\infty$, supported, respectively, on the balls $\{B_j\}_{j=1}^\infty$, with $\int_{\rn}a_j(x)dx=0$ for any $j\in\mathbb{N}$ so that $f=\sum_{j=1}^\infty\lambda_ja_j$ with convergence in $(E_r^q)_t(\rn)$. \[au\] Let $t\in(0,\infty)$, $r\in(1,\infty)$, $q\in(\frac{n}{n+1},1]$ and $s\in(0,q]$. Then $$(\mathfrak{C}_r^q)_t(\rn)\subset H^q(\rn)\subset H^q(\rn)\cup H^s(\rn)\subset(HE_s^q)_t(\rn),$$ where the slice Hardy space $(HE_s^q)_t(\rn)$ is as in Definition \[dh\]. By the definition of the space $(\mathfrak{C}_r^q)_t(\rn)$, it is obvious that the Hardy-type space $(\mathfrak{C}_r^q)_t(\rn)$ is the subspace of the classical real Hardy space $H^q(\rn)$. Furthermore, by Proposition \[ggg\](i), we know that $(\mathfrak{C}_r^q)_t(\rn)\subset H^q(\rn)\subset H^q(\rn)\cup H^s(\rn)\subset(HE_s^q)_t(\rn)$, which completes the proof of Proposition \[au\]. Let $t\in(0,\infty)$, $r\in(1,\infty)$, $q\in(\frac{1}{2},1]$, $s\in(0,q]$ and $n=1$. When $q\in(\frac{1}{2},1)$, the difference $\delta_1-\delta_{-1}$ of Dirac masses lies in $H^q(\mathbb{R})$ but not in $(\mathfrak{C}_r^q)_t(\mathbb{R})$, because $\delta_1-\delta_{-1}$ is only a distribution, not a function (see also, for example, [@em p.129]). This shows that $(\mathfrak{C}_r^q)_t(\mathbb{R})\subsetneqq(HE_s^q)_t(\mathbb{R})$. When $q=1$ and $s\in(0,1),$ let $d=\lfloor s^{-1}-1\rfloor,$ the $d$-order derivative $(\delta_1-\delta_{-1})^{(d)}$ of $\delta_1-\delta_{-1}$ lies in $H^s(\rr)$, but not in $(\mathfrak{C}_r^1)_t(\mathbb{R})$; thus, in this case, we also have $(\mathfrak{C}_r^1)_t(\mathbb{R})\subsetneqq(HE_s^1)_t(\mathbb{R})$. \[con\] Let $\delta\in(0,1]$, a convolutional $\delta$-type Calderón-Zygmund operator $\mathcal{T}$ is a linear operator, which is bounded on $L^2(\rn)$ with kernel $k\in\mathcal{S}'(\rn)$ coinciding with a locally integrable function on $\rn\setminus\{\vec{0}_n\}$ and satisfying that there exists a positive constant $C$, independent of $f$, $x$ and $y$, such that, for any $x,\ y\in\rn$ with $\|x\|\ge\|2y\|$, $$\|k(x-y)-k(x)\|\le C\frac{\|y\|^\delta}{\|x\|^{n+\delta}}$$ and, for any $f\in L^2(\rn)$, $\mathcal{T}(f):=k\ast f$. \[noncon\] Let $\delta\in(0,1]$. A non-convolutional $\delta$-type Calderón-Zygmund operator is a linear operator which is bounded on $L^2(\rn)$ and satisfies that, for any $f\in L^2(\rn)$ with compact support and $x\notin\supp(f)$, $$\mathcal{T}(f)(x):=\int_{\rn}K(x,y)f(y)\,dy,$$ where $K$ denotes a measurable function on $(\rn\times\rn)\setminus\{(x,x):\ x\in\rn\}$ satisfying that there exists a positive constant $C$ such that, for any $x,\ y,\ z\in\rn$, $$\lf\|K(x,y)-K(x,z)\r\|\le C\frac{\|y-z\|^\delta}{\|x-y\|^{n+\delta}}\quad \mathrm{when}\ \|x-y\|>2\|y-z\|.$$ \[czsuan\] Let $t,\ q\in(0,\infty)$ and $\Phi$ be an Orlicz function with positive lower type $p_{\Phi}^-$ and positive upper type $p_{\Phi}^+$. Assume that $r\in(\max\{1,\ q,\ p_{\Phi}^+\},\infty]$, $s\in(0,\min\{p^-_{\Phi},q,1\})$. Let $\{\lambda_k\}_{k\in\mathbb{N}}\subset[0,\infty)$ and $\{Q_k\}_{k\in\mathbb{N}}$ be a sequence of cubes. Then, for any sequence $\{a_k\}_{k\in\mathbb{N}}\subset L^r(\rn)$ such that, for any $k\in\mathbb{N}$, $\supp(a_k)\subset Q_k$, $$\lf\\|a_k\r\\|_{L^r(\rn)}\le\frac{\|Q_k\|^{\frac{1}{r}}}{\\|\chi_{Q_k}\\|_{(E_\Phi^{q})_t(\rn)}}$$ and $$\lf\\|\lf\{\sum_{k\in\mathbb{N}}\lf[\frac{\lambda_k}{\\|\chi_{Q_k}\\|_{(E_\Phi^{q})_t(\rn)}}\r]^s \chi_{Q_k}\r\}^{\frac{1}{s}}\r\\|_{(E_\Phi^{q})_t(\rn)}<\infty,$$ it holds true that $$\lf\\|\sum_{k\in\mathbb{N}}\lambda_ka_k\r\\|_{(E_\Phi^{q})_t(\rn)} \lesssim\lf\\|\lf\{\sum_{k\in\mathbb{N}} \lf[\frac{\lambda_k}{\\|\chi_{Q_k}\\|_{(E_\Phi^{q})_t(\rn)}}\r]^s\chi_{Q_k}\r\}^{\frac{1}{s}}\r\\|_{(E_\Phi^{q})_t(\rn)},$$ where the implicit positive constant is independent of $\{\lambda_k\}_{k\in\mathbb{N}}$, $\{a_k\}_{k\in\mathbb{N}}$ and $t$. By Lemmas \[ballproof\], \[sconvex\], \[man28\] and \[man37\], we know that $(E_\Phi^q)_t(\rn)$ satisfies all the assumptions of [@ykds Theorem 2.10]. As a simple corollary of [@ykds Theorem 2.10], we immediately obtain the desired conclusion of Lemma \[czsuan\], which completes the proof of Lemma \[czsuan\]. Via borrowing some ideas from the proof of Yan et al. [@yyyz Theorem 7.4] and applying the criterion established in Theorem \[suanzi1\], we obtain the boundedness of convolutional $\delta$-type Calderón-Zygmund operators from $(HE_\Phi^{q})_t(\rn)$ to itself or to $(E_\Phi^{q})_t(\rn)$ (see Theorem \[cz\] below), which extends the corresponding results of Fefferman and Stein [@fs72 Theorem 12] to the present setting. \[cz\] Let $t,\ q\in(0,\infty)$, $\delta\in(0,1]$ and $\Phi$ be an Orlicz function with positive lower type $p_{\Phi}^-$ satisfying $\min\{p_{\Phi}^-,\ q\}\in(\frac{n}{n+\delta},1]$ and positive upper type $p_{\Phi}^+$. 1. If $\mathcal{T}$ is a convolutional $\delta$-type Calderón-Zygmund operator as in Definition \[con\], then there exists a positive constant $C$ such that, for any $f\in (HE_\Phi^q)_t(\rn)$, $$\\|\mathcal{T}(f)\\|_{(E_\Phi^{q})_t(\rn)}\le C\\|f\\|_{(HE_\Phi^{q})_t(\rn)}.$$ 2. If $\mathcal{T}$ is a convolutional $\delta$-type Calderón-Zygmund operator as in Definition \[con\], then there exists a positive constant $C$ such that, for any $f\in (HE_\Phi^q)_t(\rn)$, $$\\|\mathcal{T}(f)\\|_{(HE_\Phi^{q})_t(\rn)}\le C\\|f\\|_{(HE_\Phi^{q})_t(\rn)},$$ where the positive constant $C$ is independent of $f$ and $t$. By similarity, we only prove (ii). Let $\mathcal{T}$ be a convolutional $\delta$-type Calderón-Zygmund operator as in Definition \[con\]. Let $r\in(\max\{1,\ q,\ p_{\Phi}^+\},\infty)$ and $f\in(HE_\Phi^{q,r,d})_t^{\fin}(\rn)$. Then, without loss of generality, we may assume that $\\|f\\|_{(HE_\Phi^{q})_t(\rn)}=1$. Thus, to prove (ii), by Theorem \[suanzi1\](i), we only need to show that $$\label{h2} \\|\mathcal{T}(f)\\|_{(HE_\Phi^{q})_t(\rn)}\lesssim1.$$ Noticing that $f\in(HE_\Phi^{q})_t(\rn)\bigcap L^r(\rn)$, by the proof of [@ykds Theorem 3.7], we know that there exist a sequence of $\{\lambda_j\}_{j=1}^{\infty}\subset[0,\infty)$ and $\{a_j\}_{j=1}^{\infty}$ of $((E_\Phi^q)_t(\rn),\ r,\ 0)$-atoms supported, respectively, on the cubes $\{Q_j\}_{j=1}^{\infty}:=\{Q(x_j,r_j)\}_{j=1}^{\infty}\subset\mathcal{Q}$ such that $f=\sum_{j=1}^{\infty}\lambda_j a_j$ converges in $L^r(\rn)$ and $$\lf\\|\lf\{\sum_{j=1}^{\infty} \lf[\frac{\lambda_j} {\\|\chi_{Q_j}\\|_{(E_\Phi^q)_t(\rn)}} \r]^s\chi_{Q_j} \r\} ^{\frac{1}{s}}\r\\|_{(E_\Phi^q)_t(\rn)}\lesssim\\|f\\|_{(HE_\Phi^q)_t(\rn)}\lesssim1.$$ From the fact that $\mathcal{T}$ is bounded on $L^r(\rn)$ (see, for example, [@d Theorem 5.1]), we deduce that $$\mathcal{T}(f)=\sum_{j=1}^{\infty}\lambda_j \mathcal{T}(a_j)$$ converges in $L^r(\rn)$. Using this and Theorem \[mdj\], we have $$\begin{aligned} \lf\\|\mathcal{T}(f)\r\\|_{(HE_\Phi^{q})_t(\rn)}&\sim\lf\\|M\lf(\mathcal{T}(f),\varphi\r)\r\\|_{(E_\Phi^q)_t(\rn)}\\ &\lesssim\lf\\|\sum_{j=1}^{\infty}\lambda_jM\lf(\mathcal{T}(a_j),\varphi\r) \chi_{4\sqrt{n}Q_j}\r\\|_{(E_\Phi^q)_t(\rn)}+ \lf\\|\sum_{j=1}^{\infty}\lambda_jM\lf(\mathcal{T}(a_j),\varphi\r)\chi_{\rn\setminus 4\sqrt{n}Q_j}\r\\|_{(E_\Phi^q)_t(\rn)}\\ &=:\mathrm{I}+\mathrm{II},\end{aligned}$$ where $\varphi\in\mathcal{S}(\rn)$ satisfies $\int_{\rn}\varphi(x)\,dx\neq0$ and $M(\mathcal{T}(f),\varphi)$ is as in Definition \[31\]. For $\mathrm{I}$, by the boundedness of $\mathcal{T}$ on $L^r(\rn)$ and the fact that $M(\mathcal{T}(a),\varphi)\lesssim\mathcal{M}(\mathcal{T}(a))$, we conclude that, for any $j\in\mathbb{N}$, $$\lf\\|M\lf(\mathcal{T}(a_j),\varphi\r)\chi_{4\sqrt{n}Q_j}\r\\|_{L^r(\rn)} \lesssim\lf\\|\mathcal{M}\lf(\mathcal{T}(a_j)\r)\r\\|_{L^r(\rn)}\lesssim\lf\\|\mathcal{T}(a_j)\r\\|_{L^r(\rn)} \lesssim\lf\\|a_j\r\\|_{L^r(\rn)}\lesssim\frac{\|Q_j\|^{\frac{1}{r}}}{\\|\chi_{Q_j}\\|_{(E_\Phi^q)_t(\rn)}},$$ which, combined with Lemma \[czsuan\], implies that $$\mathrm{I} \lesssim\lf\\|\lf\{\sum_{j=1}^{\infty}\lf[\lambda_jM(\mathcal{T}(a_j),\varphi) \chi_{4\sqrt{n}Q_j}\r]^s\r\}^{\frac{1}{s}}\r\\|_{(E_\Phi^q)_t(\rn)} \lesssim\lf\\|\lf\{\sum_{j=1}^{\infty} \lf[\frac{\lambda_j} {\\|\chi_{Q_j}\\|_{(E_\Phi^q)_t(\rn)}} \r]^s\chi_{Q_j} \r\} ^{\frac{1}{s}}\r\\|_{(E_\Phi^q)_t(\rn)} \lesssim1.$$ This is a desired estimate. As for $\mathrm{II}$, for any $\tau\in(0,\infty)$, let $k^{(\tau)}:=k\ast\varphi_\tau$ with $\varphi_\tau(\cdot):=\frac{1}{\tau^n}\varphi(\frac{\cdot}{\tau})$. By the proof of [@yyyz Theorem 7.4], we find that $k^{(\tau)}$ satisfies the same conditions as $k$. Now, by the vanishing moment condition of $a_j$ and the Hölder inequality, we know that, for any $x\notin 4\sqrt{n}Q_j$, $$\begin{aligned} \lf\|M\lf(\mathcal{T}(a_j),\varphi\r)(x)\r\| &=\sup_{\tau\in(0,\infty)}\lf\|\varphi_\tau\ast (k \ast a_j)(x)\r\| =\sup_{\tau\in(0,\infty)}\lf\|k^{(\tau)} \ast a_j(x)\r\|\\ &=\sup_{\tau\in(0,\infty)}\lf\|\int_{\rn} \lf[k^{(\tau)}(x-y)-k^{(\tau)}(x-x_j) \r]a_j(y)\,dy\r\|\\ &\lesssim\int_{\rn} \frac{\|y-x_j\|^\delta}{\|x-x_j\|^{n+\delta}}\lf\|a_j(y)\r\|\,dy \lesssim\frac{r_j^\delta}{\|x-x_j\|^{n+\delta}}\\|a_j\\|_{L^r(\rn)}\lf\|Q_j\r\|^{\frac{1}{r'}}\\ &\lesssim\frac{r_j^{n+\delta}}{\|x-x_j\|^{n+\delta}}\frac{1}{\\|\chi_{Q_j}\\|_{(E_\Phi^{q})_t(\rn)}} \lesssim\lf[\mathcal{M}\lf(\chi_{Q_j}\r)(x) \r]^{\frac{n+\delta}{n}}\frac{1}{\\|\chi_{Q_j}\\|_{(E_\Phi^{q})_t(\rn)}},\end{aligned}$$ which implies that, for any $x\notin 4\sqrt{n}Q_j$, $$\lf\|M\lf(\mathcal{T}(a_j),\varphi\r)(x)\r\|\chi_{\rn\setminus 4\sqrt{n}Q_j}(x) \lesssim\lf[\mathcal{M}\lf(\chi_{Q_j}\r)(x)\r]^{\frac{n+\delta}{n}}\frac{1}{\\|\chi_{Q_j}\\|_{(E_\Phi^{q})_t(\rn)}}.$$ Therefore, we have $$\begin{aligned} \mathrm{II} \lesssim\lf\\|\sum_{j=1}^{\infty}\frac{\lambda_j}{\\|\chi_{Q_j}\\|_{(E_\Phi^{q})_t(\rn)}} \lf[\mathcal{M}\lf(\chi_{Q_j}\r)\r]^{\frac{n+\delta}{n}}\r\\|_{(E_\Phi^q)_t(\rn)}.\end{aligned}$$ Let $u:=\frac{n}{n+\delta}$ and $\Phi_u(\tau):=\Phi(\sqrt[u]{\tau})$. Since $\min\{p_{\Phi}^-,\ q\}\in(\frac{n}{n+\delta},1]$, it follows that $\Phi_u$ is of upper type $\frac{p_{\Phi}^+}{u}$ and of lower type $\frac{p_{\Phi}^-}{u}$, and $\frac{p_{\Phi}^-}{u}$, $\frac{q}{u}\in(1,\infty)$. By this and Theorem \[main\], we further conclude that $$\begin{aligned} \label{gg} \mathrm{II} &\lesssim\lf\\| \lf\{ \sum_{j=1}^{\infty} \frac{\lambda_j}{\\|\chi_{Q_j}\\|_{(E_\Phi^{q})_t(\rn)}} \lf[\mathcal{M}(\chi_{Q_j})\r]^{\frac{1}{u}} \r\}^{u} \r\\|_{(E_{\Phi_u}^{q/u})_t(\rn)}^{\frac{1}{u}} \lesssim\lf\\| \lf\{ \sum_{j=1}^{\infty} \frac{\lambda_j\chi_{Q_j}}{\\|\chi_{Q_j}\\|_{(E_\Phi^{q})_t(\rn)}} \r\}^{u} \r\\|_{(E_{\Phi_u}^{q/u})_t(\rn)}^{\frac{1}{u}}\\ \noz &\lesssim\lf\\|\lf\{\sum_{j=1}^{\infty} \lf[\frac{\lambda_j} {\\|\chi_{Q_j}\\|_{(E_\Phi^q)_t(\rn)}} \r]^s\chi_{Q_j} \r\} ^{\frac{1}{s}}\r\\|_{(E_\Phi^q)_t(\rn)} \lesssim1.\end{aligned}$$ Combining the estimates for $\mathrm{I}$ and $\mathrm{II}$, we obtain , which completes the proof of (ii) and hence of Theorem \[cz\]. We recall the notion of $\beta$-order Calderón-Zygmund operators as follows (see, for example, [@yyyz]). For any given $\beta\in(0,\infty)\setminus\mathbb{N}$, a linear operator $\mathcal{T}$ is called a $\beta$-order Calderón-Zygmund operator if $\mathcal{T}$ is bounded on $L^2(\rn)$ and its kernel $$k:\ (\rn\times\rn)\setminus\lf\{(x,x):\ x\in\rn\r\}\rightarrow\mathbb{C}$$ satisfies that there exists a positive constant $C$ such that, for any $\alpha\in\mathbb{Z}_+^n$ with $\|\alpha\|\le\lfloor\beta\rfloor$ and $x,\ y,\ z\in\rn$ with $\|x-y\|>2\|y-z\|$, $$\label{71} \lf\|\partial_x^\alpha k(x,y)-\partial_x^\alpha k(x,z)\r\|\le C\frac{\|y-z\|^{\beta-\lfloor\beta\rfloor}}{\|x-y\|^{n+\beta}}$$ and, for any $f\in L^2(\rn)$ having compact support and $x\notin \supp f$, $$\mathcal{T}(f)(x)=\int_{\supp f}k(x,y)f(y)\,dy.$$ Here and hereafter, for any $\alpha:=(\alpha_1, \ldots, \alpha_n)\in\zz_+^n$, $\partial_x^{\alpha}:=(\frac{\partial}{\partial x_1}) ^{\alpha_1}\cdots(\frac{\partial}{\partial x_n})^{\alpha_n}$. Next, we establish the boundedness of the $\beta$-order Calderón-Zygmund operator $\mathcal{T}$ from the Orlicz-slice Hardy space $(HE_\Phi^{q})_t(\rn)$ to itself (see Theorem \[cz2\]) or to $(E_\Phi^{q})_t(\rn)$ (see Theorem \[cz3\]). Recall that, for any $l\in\mathbb{N}$, an operator $\mathcal{T}$ is said to have the vanishing moment condition up to order $l$ if, for any $a\in L^2(\rn)$ with compact support and satisfying that, for any $\gamma\in\zz_+^n$ with $\|\gamma\|\leq l$, $\int_{\rn}x^\gamma a(x)\,dx=0$, it holds true that $\int_{\rn}x^{\gamma}\mathcal{T}(a)(x)\,dx=0$. \[cz2\] Let $t\in(0,\infty)$, $q\in(0,2)$, $\beta\in(0,\infty)\setminus\mathbb{N}$ and $\Phi$ be an Orlicz function with positive lower type $p_{\Phi}^-$ satisfying $\min\{p_{\Phi}^-,\ q\}\in(\frac{n}{n+\beta},\frac{n}{n+\lfloor\beta\rfloor}]$ and positive upper type $p_{\Phi}^+\in(0,2)$. Let $\mathcal{T}$ be a $\beta$-order Calderón-Zygmund operator and have the vanishing moment condition up to order $\lfloor\beta\rfloor$. Then $\mathcal{T}$ has a unique extension on $(HE_\Phi^{q})_t(\rn)$ and, for any $f\in(HE_\Phi^{q})_t(\rn)$, $$\lf\\|\mathcal{T}(f)\r\\|_{(HE_\Phi^{q})_t(\rn)}\le C\\|f\\|_{(HE_\Phi^{q})_t(\rn)},$$ where $C$ is positive constant independent of $f$ and $t$. Let $\{\lambda_j\}_{j=1}^{\infty}$ and $\{a_j\}_{j=1}^{\infty}$ be the same as in the proof of Theorem \[cz\]. By an argument similar to that used in the proof of Theorem \[cz\], we know that, to prove Theorem \[cz2\], it suffices to show that $$\label{yuan} \lf\\|\sum_{j=1}^{\infty}\lambda_jM\lf(\mathcal{T}(a_j),\varphi\r)\r\\|_{(E_\Phi^q)_t(\rn)}\lesssim1,$$ where $M(\mathcal{T}(a_j),\varphi)$ is as in Definition \[31\]. To this end, it is easy to see that $$\begin{aligned} \lf\\|\sum_{j=1}^{\infty}\lambda_jM\lf(\mathcal{T}(a_j),\varphi\r)\r\\|_{(E_\Phi^q)_t(\rn)} &\lesssim\lf\\|\sum_{j=1}^{\infty}\lambda_jM\lf(\mathcal{T}(a_j),\varphi\r) \chi_{4\sqrt{n}Q_j}\r\\|_{(E_\Phi^q)_t(\rn)}\\ &\quad+\lf\\|\sum_{j=1}^{\infty}\lambda_jM\lf(\mathcal{T}(a_j),\varphi\r) \chi_{\rn\setminus 4\sqrt{n}Q_j}\r\\|_{(E_\Phi^q)_t(\rn)} =:\mathrm{I}+\mathrm{II},\end{aligned}$$ where, for any $j\in\mathbb{N}$, $Q_j:=Q(x_j,r_j)$ is the same as in the proof of Theorem \[cz\]. For $\mathrm{I}$, by an argument similar to that used in the proof of Theorem \[cz\], we conclude that $\mathrm{I}\lesssim1$. Next, we deal with $\mathrm{II}$. To this end, from the vanishing moment condition of $\mathcal{T}$ and the fact that $\lfloor\beta\rfloor\le n(\frac{1}{\min\{q,p_{\Phi}^-\}}-1)$ implies $\lfloor\beta\rfloor\le d$, it follows that, for any $j\in\mathbb{N}$, $\tau\in(0,\infty)$ and $x\notin4Q_j$, $$\begin{aligned} \label{79} \lf\|\varphi_\tau\ast\mathcal{T}(a_j)(x)\r\| &=\frac{1}{\tau^n}\lf\|\int_{\rn} \varphi\lf(\frac{x-y}{\tau}\r)\mathcal{T}(a_j)(y)\,dy\r\|\\\noz &\le\frac{1}{\tau^n}\int_{\rn}\lf\|\varphi\lf(\frac{x-y}{\tau}\r)- \sum_{\|\alpha\|\le\lfloor\beta\rfloor}\frac{\partial^\alpha\varphi (\frac{x-x_j}{\tau})}{\alpha!}\lf(\frac{y-x_j}{\tau}\r)^{\alpha}\r\| \lf\|\mathcal{T}(a_j)(y)\r\|\,dy\\\noz &=\frac{1}{\tau^n}\lf(\int_{\|y-x_j\|<2r_j}+\int_{2r_j\le\|y-x_j\|< \frac{\|x-x_j\|}{2}}+\int_{\|y-x_j\|\ge\frac{\|x-x_j\|}{2}}\r)\\\noz &\quad\times\lf\|\varphi\lf(\frac{x-y}{\tau}\r)- \sum_{\|\alpha\|\le\lfloor\beta\rfloor}\frac{\partial^\alpha \varphi(\frac{x-x_j}{\tau})}{\alpha!}\lf(\frac{y-x_j}{\tau}\r)^{\alpha}\r\| \lf\|\mathcal{T}(a_j)(y)\r\|\,dy\\\noz &=:\mathrm{II}_1+\mathrm{II}_2+\mathrm{II}_3,\end{aligned}$$ where $\varphi\in\mathcal{S}(\rn)$ satisfying $\int_{\rn}\varphi(x)\,dx\neq0$. For $\mathrm{II}_1$, by the Taylor remainder theorem, the Hölder inequality and the fact that $\mathcal{T}$ is bounded on $L^2(\rn)$, similarly to the estimation of [@yyyz (7.9)], we find that, for any $\tau\in(0,\infty)$ and $x\notin 4Q_j$, $$\begin{aligned} \mathrm{II}_1&\lesssim\frac{1}{\tau^n}\int_{\|y-x_j\|<2r_j} \frac{\tau^{n+\lfloor\beta\rfloor+1}}{\|x-x_j\|^{n+\lfloor\beta\rfloor+1}} \frac{\|y-x_j\|^{\lfloor\beta\rfloor+1}}{\tau^{\lfloor\beta\rfloor+1}}\lf\|\mathcal{T}(a_j)(y)\r\|\,dy\\ &\lesssim\frac{r_j^{\lfloor\beta\rfloor+1}}{\|x-x_j\|^{n+\lfloor\beta\rfloor+1}}\lf\\| \mathcal{T}a_j\r\\|_{L^2(\rn)}\lf\|Q_j\r\|^{\frac{1}{2}} \lesssim\frac{r_j^{n+\lfloor\beta\rfloor+1}}{\|x-x_j\|^{n+\lfloor\beta\rfloor+1}} \frac{1}{\\|\chi_{Q_j}\\|_{(E_\Phi^{q})_t(\rn)}}.\end{aligned}$$ For $\mathrm{II}_2$, by the Taylor remainder theorem, the vanishing moments of $a_j$, the fact that $\lfloor\beta\rfloor\le n(\frac{1}{\min\{q,p_{\Phi}^-\}}-1)\le d$, (\[71\]) and the Hölder inequality, similarly to the estimation of [@yyyz (7.10)], we conclude that, for any $\tau\in(0,\infty)$ and $x\notin4Q_j$, $$\begin{aligned} \mathrm{II}_2 &\lesssim\frac{1}{\|x-x_j\|^{n+\lfloor\beta\rfloor+1}}\int_{2r_j\le\|y-x_j\|<\frac{\|x-x_j\|}{2}} \|y-x_j\|^{\lfloor\beta\rfloor+1}\int_{Q_j} \|a_j(z)\| \frac{\|z-x_j\|^\beta}{\|y-x_j\|^{n+\beta}}\,dz\,dy\\ &\lesssim\frac{r_j^{\beta}}{\|x-x_j\|^{n+\lfloor\beta\rfloor+1}} \int_{2r_j\le\|y-x_j\|<\frac{\|x-x_j\|}{2}} \frac{1}{\|y-x_j\|^{n+\beta-\lfloor\beta\rfloor-1}}\,dy\\|a_j\\|_{L^2(\rn)}\|Q_j\|^{\frac{1}{2}}\\ &\lesssim\frac{r_j^{n+\beta}}{\|x-x_j\|^{n+\beta}}\frac{1}{\\|\chi_{Q_j}\\|_{(E_\Phi^{q})_t(\rn)}}.\end{aligned}$$ For $\mathrm{II}_3$, by the vanishing moments of $a_j$, the fact that $\lfloor\beta\rfloor\le n(\frac{1}{\min\{q,p_{\Phi}^-\}}-1)\le d$, (\[71\]) and the Hölder inequality, similarly to the estimation of [@yyyz (7.11)], we know that, for any $\tau\in(0,\infty)$ and $x\notin4Q_j$, $$\begin{aligned} \mathrm{II}_3 &\lesssim\int_{\|y-x_j\|\ge\frac{\|x-x_j\|}{2}} \|\varphi_\tau(x-y)\|\int_{Q_j} \|a_j(z)\|\frac{\|z-x_j\|^\beta}{\|y-x_j\|^{n+\beta}}\,dz\,dy\\ &\quad +\int_{\|y-x_j\|\ge\frac{\|x-x_j\|}{2}} \lf\|\frac{1}{\tau^n}\sum_{\|\alpha\|\le\lfloor\beta\rfloor} \frac{\partial^\alpha\varphi(\frac{x-x_j}{\tau})}{\alpha!} \lf(\frac{y-x_j}{\tau}\r)^{\alpha}\r\| \int_{Q_j} \|a_j(z)\|\frac{\|z-x_j\|^\beta}{\|y-x_j\|^{n+\beta}}\,dz\,dy\\ &\lesssim\frac{\|r_j\|^\beta}{\|x-x_j\|^{n+\beta}}\\|a_j\\|_{L^2(\rn)}\|Q_j\|^{\frac{1}{2}} \int_{\|y-x_j\|\ge\frac{\|x-x_j\|}{2}}\|\varphi_\tau(x-y)\|\,dy\\ &\quad +\sum_{\|\alpha\|\le\lfloor\beta\rfloor} \|r_j\|^\beta \\|a_j\\|_{L^2(\rn)} \|Q\|^{\frac{1}{2}} \int_{\|y-x_j\|\ge\frac{\|x-x_j\|}{2}} \frac{1}{\tau^n}\frac{\tau^{n+\alpha}}{\|x-x_j\|^{n+\alpha}} \frac{\|y-x_j\|^{\alpha}}{\tau^\alpha}\frac{1}{\|y-x_j\|^{n+\beta}}\,dy\\ &\lesssim \frac{1}{\\|\chi_{Q_j}\\|_{(E_\Phi^{q})_t(\rn)}}\frac{r_j^{n+\beta}}{\|x-x_j\|^{n+\beta}}.\end{aligned}$$ Combining (\[79\]) and the estimates of $\mathrm{II}_1$, $\mathrm{II}_2$ and $\mathrm{II}_3$, we conclude that, for any $x\notin4Q_j$, $$\begin{aligned} M(\mathcal{T}(a_j),\varphi)(x) &=\sup_{\tau\in(0,\infty)}\|\varphi_\tau\ast \mathcal{T}(a_j)(x)\|\\ &\lesssim\frac{r_j^{n+\beta}}{\|x-x_j\|^{n+\beta}} \frac{1}{\\|\chi_{Q_j}\\|_{(E_\Phi^{q})_t(\rn)}}\lesssim\lf[\mathcal{M}(\chi_{Q_j})(x)\r]^{\frac{n+\delta}{n}} \frac{1}{\\|\chi_{Q_j}\\|_{(E_\Phi^{q})_t(\rn)}},\end{aligned}$$ which further implies that, for any $x\in\rn$, $$M(\mathcal{T}(a_j),\varphi)(x)\chi_{4Q_j}(x)\lesssim\lf[\mathcal{M}(\chi_{Q_j})(x)\r]^{\frac{n+\delta}{n}} \frac{1}{\\|\chi_{Q_j}\\|_{(E_\Phi^{q})_t(\rn)}}.$$ Then, by an argument similar to that used in the proof of Theorem \[cz\], we know that holds true, which completes the proof of Theorem \[cz2\]. \[cz3\] Let $t\in(0,\infty)$, $q\in(0,2)$, $\beta\in(0,\infty)\setminus\mathbb{N}$ and $\Phi$ be an Orlicz function with positive lower type $p_{\Phi}^-$ satisfying $\min\{p_{\Phi}^-,\ q\}\in(\frac{n}{n+\beta},\frac{n}{n+\lfloor\beta\rfloor}]$ and positive upper type $p_{\Phi}^+\in(0,2)$. Let $\mathcal{T}$ be a $\beta$-order Calderón-Zygmund operator. Then $\mathcal{T}$ has a unique extension from $(HE_\Phi^{q})_t(\rn)$ to $(E_\Phi^{q})_t(\rn)$ and, for any $f\in(HE_\Phi^{q})_t(\rn)$, $$\\|\mathcal{T}(f)\\|_{(E_\Phi^{q})_t(\rn)}\le C\\|f\\|_{(HE_\Phi^{q})_t(\rn)},$$ where $C$ is positive constant independent of $f$ and $t$. Let $\{\lambda_j\}_{j=1}^{\infty}$ and $\{a_j\}_{j=1}^{\infty}$ be the same as in the proof of Theorem \[cz\]. By an argument similar to that used in the proof of Theorem \[cz\], we know that, to prove Theorem \[cz3\], it suffices to show that $$\label{yuan2} \lf\\|\sum_{j=1}^{\infty}\lambda_j\mathcal{T}(a_j)\r\\|_{(E_\Phi^q)_t(\rn)}\lesssim1.$$ To this end, it is easy to see that $$\begin{aligned} \lf\\|\sum_{j=1}^{\infty}\lambda_j\mathcal{T}(a_j)\r\\|_{(E_\Phi^q)_t(\rn)} &\lesssim\lf\\|\sum_{j=1}^{\infty}\lambda_j\mathcal{T}(a_j) \chi_{4\sqrt{n}Q_j}\r\\|_{(E_\Phi^q)_t(\rn)}\\ &\quad+\lf\\|\sum_{j=1}^{\infty}\lambda_j\mathcal{T}(a_j) \chi_{\rn\setminus 4\sqrt{n}Q_j}\r\\|_{(E_\Phi^q)_t(\rn)} =:\mathrm{I}+\mathrm{II},\end{aligned}$$ where, for any $j\in\mathbb{N}$, $Q_j:=Q(x_j,r_j)$ is the same as in the proof of Theorem \[cz\]. For $\mathrm{I}$, by the boundedness of $\mathcal{T}$ on $L^2(\rn)$, we conclude that, for any $j\in\mathbb{N}$, $$\\|\mathcal{T}(a_j)\\|_{L^2(\rn)} \lesssim\\|a_j\\|_{L^2(\rn)}\lesssim\frac{\|Q_j\|^{\frac{1}{2}}}{\\|\chi_{Q_j}\\|_{(E_\Phi^q)_t(\rn)}},$$ which, together with Lemma \[czsuan\], implies that $$\mathrm{I} \lesssim\lf\\|\lf\{\sum_{j=1}^{\infty}\lf[\lambda_j\mathcal{T}(a_j) \chi_{4\sqrt{n}Q_j}\r]^s\r\}^{\frac{1}{s}}\r\\|_{(E_\Phi^q)_t(\rn)} \lesssim\lf\\|\lf\{\sum_{j=1}^{\infty} \lf[\frac{\lambda_j} {\\|\chi_{Q_j}\\|_{(E_\Phi^q)_t(\rn)}} \r]^s\chi_{Q_j} \r\} ^{\frac{1}{s}}\r\\|_{(E_\Phi^q)_t(\rn)} \lesssim1.$$ This is a desired estimate. Next, we deal with $\mathrm{II}$. To this end, from the Taylor remainder theorem, vanishing moments of $a_j$, the fact that $\lfloor\beta\rfloor\le n(\frac{1}{\min\{q,p_{\Phi}^-\}}-1)$ implies $\lfloor\beta\rfloor\le d$ and the Hölder inequality, it follows that, for any $j\in\mathbb{N}$ and $z\in Q_j$, there exists $\xi(z)\in Q_j$ such that, for any $x\notin4\sqrt{n}Q_j$, $$\begin{aligned} \|\mathcal{T}(a_j)(x)\| &\leq\int_{Q_j}\|a(z)\|\|k(x,z)\|\,dz\\ &=\int_{Q_j}\|a(z)\|\lf\|k(x,z)-\sum_{\|\alpha\|<\lfloor\beta\rfloor}\frac{\partial_x^\alpha k(x,x_j) }{\alpha!}(z-x_j)^\alpha\r\|\,dz\\ &\sim\int_{Q_j}\|a(z)\|\lf\|\sum_{\|\alpha\|=\lfloor\beta\rfloor}\frac{\partial_x^\alpha k(x,x_j)-\partial_x^\alpha k(x,\xi(z)) }{\alpha!}(z-x_j)^\alpha\r\|\,dz\\ &\lesssim\int_{Q_j}\|a(z)\|\frac{r_j^\beta}{\|x-x_j\|^{n+\beta}}\,dz \lesssim\frac{r_j^\beta}{\|x-x_j\|^{n+\beta}}\\|a_j\\|_{L^2(\rn)}\|Q_j\|^{\frac{1}{2}}\\ &\lesssim\frac{r_j^{n+\beta}}{\|x-x_j\|^{n+\beta}}\frac{1}{\\|\chi_{Q_j}\\|_{(E_\Phi^{q})_t(\rn)}} \lesssim\lf[\mathcal{M}(\chi_{Q_j})(x)\r]^{\frac{n+\delta}{n}} \frac{1}{\\|\chi_{Q_j}\\|_{(E_\Phi^{q})_t(\rn)}}.\end{aligned}$$ Then, by an argument similar to that used in the proof of Theorem \[cz\], we know that holds true, which completes the proof of Theorem \[cz3\]. - Let $t$, $\Phi$ be as in Theorem \[cz2\]. Notice that, when $\beta:=\delta\in(0,1)$, the operators $\mathcal{T}$ in Theorems \[cz2\] and \[cz3\] is just a non-convolutional $\delta$-type Calderón-Zygmund operator. Thus, the operators in Theorems \[cz2\] and \[cz3\] include the non-convolutional $\delta$-type Claderón-Zygmund operators as special cases. Observe that, differently from Theorem \[cz\], in Theorems \[cz2\] and \[cz3\], we have a restriction on the ranges of $q$ and $p_\Phi^+$, namely, $q,\ p_\Phi^+\in (0, 2)$, which is caused by the fact that the $\beta$-order Calderón-Zygmund operator is only known bounded on $L^r(\rn)$ for any $r\in (1,2]$ (see, for example, [@d Theorem 5.10]). Thus, by [@d Theorem 5.10] again, if we further assume that the kernel $k$ of $\mathcal{T}$ satisfies (5.11) of [@d Theorem 5.10], we can then remove this restriction. - Let $t\in(0,\infty,),\ r\in(1,\infty),\ \delta\in(0,1]$ and $q\in(\frac{n}{n+\delta},1].$ Recall that Auscher and Prisuelos-Arribas [@ap Proposition 8.4] proved that the non-convolutional $\delta$-type Claderón-Zygmund operators are bounded from $(\mathfrak C_r^q)_t(\rn)$ to $(E_r^q)_t(\rn)$ and from $(\mathfrak C_r^q)_t(\rn)$ to $(\mathfrak C_r^q)_t(\rn)$. In Theorems \[cz2\] and \[cz3\], if let $\beta:=\delta\in(0,1],\ s\in(\frac{n}{n+\delta},q]$ and $\Phi(\tau):=\tau^s$ for any $\tau\in[0,\infty)$, then we know that the non-convolutional $\delta$-type Claderón-Zygmund operators are bounded from $(HE_s^q)_t(\rn)$ to $(E_s^q)_t(\rn)$ and from $(HE_s^q)_t(\rn)$ to $(HE_s^q)_t(\rn).$ By Propositions \[au\] and \[ggg\], we know that $(\mathfrak C_r^q)_t(\rn)\subsetneqq (HE_s^q)_t(\rn)$ and $(E_r^q)_t(\rn)\subsetneqq (E_s^q)_t(\rn)$ and hence [@ap Proposition 8.4] and Theorems \[cz2\] and \[cz3\] in this article can not cover each other. - When $t,\ q\in(0,\infty)$ and $\Phi(\tau):=\tau^q$ for any $\tau\in[0,\infty)$, $(HE_\Phi^q)_t(\rn)$ and $(E_\Phi^q)_t(\rn)$ respectively become the classical Hardy space $H^q(\rn)$ and Lebesgue space $L^q(\rn)$. In this case, we know that, if $\delta\in (0,1]$ and $q\in(\frac{n}{n+\delta}, 1]$, then Theorems \[cz2\] and \[cz3\] and (ii) of this remark give the boundedness of the classical $\delta$-type Claderón-Zygmund operator from $H^q(\rn)$ to $L^q(\rn)$ and from $H^q(\rn)$ to itself, which is well known (see, for example, [@am1 Theorem 1.1], [@em p.115, Theorem 4], [@l p.109, Theorem 4.1 and p.119, Theorem 4.5]). [10]{} Z. V. de P. Ablé and J. Feuto, Atomic decomposition of Hardy-amalgam spaces, J. Math. Anal. Appl. 455 (2017), 1899-1936. W. Abu-Shammala and A. Torchinsky, The Hardy-Lorentz spaces $H^{p,q}(\rn)$, Studia Math. 182 (2007), 283-294. R. A. Adams and J. J. F. Fournier, Sobolev Spaces, Second edition, Pure and Applied Mathematics (Amsterdam) 140, Elsevier/Academic Press, Amsterdam, 2003. V. Almeida, J. J. Betancor and L. Rodríguez-Mesa, Anisotropic Hardy-Lorentz spaces with variable exponents, Canad. J. Math. 69 (2017), 1219-1273. J. Alvarez and M. Milman, $H^p$ continuity properties of Calderón-Zygmund-type operators, J. Math. Anal. Appl. 118 (1986), 63-79. P. Auscher and M. Mourgoglou, Representation and uniqueness for boundary value elliptic problems via first order systems, Revista Mat. Iberoam. (to appear) or arXiv: 1404.2687v2. P. Auscher and C. Prisuelos-Arribas, Tent space boundedness via extrapolation, Math. Z. 286 (2017), 1575-1604. A. Bonami and J. Feuto, Products of functions in Hardy and Lipschitz or BMO spaces (English summary), in: Recent Developments in Real and Harmonic Analysis, 57-71, Appl. Numer. Harmon. Anal., Birkhäuser Boston, Inc., Boston, MA, 2010. M. Bownik, Anisotropic Hardy Spaces and Wavelets, Mem. Amer. Math. Soc. 164 (2003), no. 781, vi+122 pp. M. Bownik, D. Yang and Y. Zhou, Weighted anisotropic Hardy spaces and their applications in boundedness of sublinear operators, Indiana Univ. Math. J. 57 (2008), 3065-3100. S. Campanato, Proprietá di una familia di spazi funzionali, Ann. Scuola Norm. Super. Pisa 18 (1964), 137-160. J. Cao, L. D. Ky and D. Yang, Bilinear decompositions of products of local Hardy and Lipschitz or BMO spaces through wavelets, Commun. Contemp. Math. 20 (2018), no. 3, 1750025, 30 pp. C. Carton-Lebrun, H. P. Heinig and S. C. Hofmann, Integral operators on weighted amalgams, Studia Math. 109 (1994), 133-157. R. R. Coifman, Y. Meyer and E. M. Stein, Some new function spaces and their applications to harmonic analysis, J. Funct. Anal. 62 (1985), 304-335. D. Cruz-Uribe and L.-A. D. Wang, Variable Hardy spaces, Indiana Univ. Math. J. 63 (2014), 447-493. J. Duoandikoetxea, Fourier Analysis, Graduate Studies in Mathematics 29, American Mathematical Society, Providence, RI, 2001. C. Fefferman, N. M. Rivière and Y. Sagher, Interpolation between $H^p$ spaces: the real method, Trans. Amer. Math. Soc. 191 (1974), 75-81. C. Fefferman and E. M. Stein, Some maximal inequalities, Amer. J. Math. 93 (1971), 107-115. C. Fefferman and E. M. Stein, $H^p$ spaces of several variables, Acta Math. 129 (1972), 137-193. R. Fefferman and F. Soria, The spaces weak $H^1$, Studia Math, 85 (1987), 1-16. H. G. Feichtinger, Banach convolution algebras of Wiener type, in: Functions, Series, Operators, Vol. I, II (Budapest, 1980), 509-524, Colloq. Math. Soc. János Bolyai, 35, North-Holland, Amsterdam, 1983. G. B. Folland and E. M. Stein, Hardy Spaces on Homogeneous Groups, Mathematical Notes 28, Princeton University Press, Princeton, N.J., University of Tokyo Press, Tokyo, 1982. L. Grafakos, Modern Fourier Analysis, Third edition, Graduate Texts in Mathematics 250, Springer, New York, 2014. L. Grafakos, L. Liu and D. Yang, Maximal function characterizations of Hardy spaces on RD-spaces and their applications, Sci. China Ser. A 51 (2008), 2253-2284. G. Hu, D. Yang and Y. Zhou, Boundedness of singular integrals in Hardy spaces on spaces of homogeneous type, Taiwanese J. Math. 13 (2009), 91-135. V. Kokilashvili and M. Krbec, Weighted Inequalities in Lorentz and Orlicz Spaces, World Scientific Publishing Co., Inc., River Edge, NJ, 1991. L. D. Ky, New Hardy spaces of Musielak-Orlicz type and boundedness of sublinear operators, Integral Equations Operator Theory 78 (2014), 115-150. G. Köthe, Topological Vector Spaces. I, Translated from the German by D. J. H. Garling, Die Grundlehren der mathematischen Wissenschaften, Band 159, Springer-Verlag, New York Inc., New York, 1969. R. Leśniewicz, On Hardy-Orlicz spaces. I, Bull. Acad. Polon. Sci. Sér. Sci. Math. Astronom. Phys. 14 (1966), 145-150. R. Leśniewicz, On Hardy-Orlicz spaces. II, Bull. Acad. Polon. Sci. Sér. Sci. Math. Astronom. Phys. 15 (1967), 277-281. Y. Liang, Y. Sawano, T. Ullrich, D. Yang and W. Yuan, New characterizations of Besov-Triebel-Lizorkin- Hausdorff spaces including coorbits and wavelets, J. Fourier Anal. Appl. 18 (2012), 1067-1111. J. Liu, D. Yang and W. Yuan, Anisotropic Hardy-Lorentz spaces and their applications, Sci. China Math. 59 (2016), 1669-1720. J. Liu, D. Yang and W. Yuan, Anisotropic variable Hardy-Lorentz spaces and their real interpolation, J. Math. Anal. Appl. 456 (2017), 356-393. J. Liu, D. Yang and W. Yuan, Littlewood-Paley characterizations of anisotropic Hardy-Lorentz spaces, Acta Math. Sci. Ser. B Engl. Ed. 38 (2018), 1-33. S. Z. Lu, Four Lectures on Real $H^p$ spaces, World Scientific Publishing Co., Inc., River Edge, NJ, 1995. S. Meda, P. Sjögren and M. Vallarino, On the $H^1$-$L^1$ boundedness of operators, Proc. Amer. Math. Soc. 136 (2008), 2921-2931. S. Müller, Hardy space methods for nonlinear partial differential equations, Equadiff 8 (Bratislava, 1993), Tatra Mt. Math. Publ. 4 (1994), 159-168. E. Nakai and Y. Sawano, Hardy spaces with variable expoments and generalized Campanato spaces, J. Funct. Anal. 262 (2012), 3665-3748. M. M. Rao and Z. D. Ren, Theory of Orlicz Spaces, Monographs and Textbooks in Pure and Applied Mathematics 146, Marcel Dekker, Inc., New York, 1991. Y. Sawano, K. Ho, D. Yang and S. Yang, Hardy spaces for ball quasi-Banach function spaces, Dissertationes Math. (Rozprawy Mat.) 525 (2017), 1-102. E. M. Stein, Harmonic Analysis: Real-Variable Methods, Orthogonality, and Oscillatory Integrals, Princeton Mathematical Series 43, Monographs in Harmonic Analysis III, Princeton University Press, Princeton, NJ, 1993. J. O. Strömberg, Bounded mean oscillation with Orlicz norms and duality of Hardy spaces, Bull. Amer. Math. Soc. 82 (1976), 953-955. J. O. Strömberg and A. Torchinsky, Weighted Hardy Spaces, Lecture Notes in Mathematics 1381, Springer-Verlag, Berlin, 1989. D. Yang, Y. Liang and L. D. Ky, Real-Variable Theory of Musielak-Orlicz Hardy Spaces, Lecture Notes in Mathematics 2182, Springer, Cham, 2017, xiii+466 pp. X. Yan, D. Yang, W. Yuan and C. Zhuo, Variable weak Hardy spaces and their applications, J. Funct. Anal. 271 (2016), 2822-2887. D. Yang and Y. Zhou, A boundedness criterion via atoms for linear operators in Hardy spaces, Constr. Approx. 29 (2009), 207-218. D. Yang and Y. Zhou, Boundedness of sublinear operators in Hardy spaces on RD-spaces via atoms, J. Math. Anal. Appl. 339 (2008), 622-635. K. Yosida, Functional Analysis, Reprint of the sixth (1980) edition, Classics in Mathematics, Springer-Verlag, Berlin, 1995. Yangyang Zhang, Dachun Yang (Corresponding author) and Wen Yuan Laboratory of Mathematics and Complex Systems (Ministry of Education of China), School of Mathematical Sciences, Beijing Normal University, Beijing 100875, People’s Republic of China [E-mails]{}: `[email protected]` (Y. Zhang) `[email protected]` (D. Yang) `[email protected]` (W. Yuan) Songbai Wang College of Mathematics and Statistics, Hubei Normal University, Huangshi 435002, People’s Republic of China [E-mail]{}: `[email protected]` [^1]: Corresponding author / March 25, 2018 / newest version.	{ "pile_set_name": "ArXiv" }
--- abstract: 'The Pauli exclusion principle plays an important role in many-body fermion systems preventing them from collapsing by repulsion. For example, the Pauli principle causes a repulsive potential at short distances between two $\alpha$ particles. On the other hand, the existence of nuclear rainbows demonstrates that the inter-nuclear potential is sufficiently attractive in the internal region to cause refraction. The two concepts of repulsion and attraction are seemingly irreconcilable. Contrary to traditional understanding, it is shown that the Pauli principle causes a [universal structural Pauli attraction]{} between nuclei rather than a [structural repulsive core]{}. Through systematic studies of $\alpha$+$\alpha$, $\alpha$+$^{16}$O, $\alpha$+$^{40}$Ca and $^{16}$O+$^{16}$O systems, it is shown that the emergence of cluster structures near the threshold energy at low energies and nuclear rainbows at high energies is a direct consequence of the Pauli principle.' author: - 'S. Ohkubo' title: 'Luneburg-lens-like universal structural Pauli attraction in nucleus-nucleus interactions: origin of emergence of cluster structures and nuclear rainbows ' --- Why do cluster structures appear near the threshold energy, while molecular resonances occur at higher excitation energies, and nuclear rainbows at even high energies? The threshold rule, molecular resonance theory and nuclear rainbow theory have been proposed and extensively studied for more than fifty years. Until now, these independent theories - each describing successfully different facets of nuclear structure - had not been thought to be closely connected in the level of the fundamental principle. I will show that they share a common [raison d’être]{}: They are all a direct consequence of the universal Pauli attraction. The Pauli exclusion plays an important role in nuclei. The shell structure, in which nucleons behave like independent particles in a mean field potential and persist throughout the periodic table, is a consequence of the Pauli principle and the short-range character of the nuclear force [@Bohr1969]. The Pauli principle also plays an important role between nuclei at small distances where they overlap. For the typical $\alpha$+$\alpha$ system, microscopic Resonationg Group Method (RGM) studies have revealed [@Tamagaki1962; @Tamagaki1965; @Tamagaki1968; @Hiura1972] that the inter-nuclear interaction for $S$ and $D$ waves has a [repulsive]{} core at short distances and angular-momentum ($L$)-dependent [shallow]{} attraction in the outer region. The repulsive core explains the experimental phase shifts in $\alpha$+$\alpha$ scattering and the well-developed $\alpha$ cluster structure of $^8$Be well [@Tamagaki1965; @Tamagaki1968; @Hiura1972; @Ali1966]. The repulsive core was found to be a potential representation of the damped inner oscillations of the inter-cluster wave functions, with the energy-independent node at around 2 fm caused by the Pauli principle [@Tamagaki1968; @Hiura1972]. This is known as the [structural repulsive core]{} [@Otsuki1965]. For heavy ion systems such as $^{16}$O+$^{16}$O, the existence of a repulsive core at short distances has also been shown in microscopic model calculations [@Tohsaki1975; @Ando1978; @Buck1977]. Although the so-called standard optical potential model with a Woods-Saxon form factor witnessed tremendous success in the studies of light-ion and heavy-ion scattering and reactions [@Hodgson1978], it could not describe the Backward Angle Anomaly (BAA) or Anomalous Large Angle Scattering (ALAS) in $\alpha$+$^{16}$O and $\alpha$+$^{40}$Ca scattering [@Brink1985]. This was shown to be resolved using a [non-standard]{} optical model with a [deep]{} potential without a repulsive core, for $\alpha$+$^{16}$O in Refs.[@Ohkubo1977; @Michel1983] and for $\alpha$+$^{40}$Ca in Refs.[@Michel1977; @Delbar1978]. Furthermore the clear observation of the Airy minimum of the nuclear rainbow in $^{16}$O+$^{16}$O scattering at $E_L$=350 MeV [@Stiliaris1989] showed that the potential is [deep]{} in the internal region [@Khoa2007]. The deep potentials in the internal region from the ALAS and rainbow are inconsistent with the repulsive core picture concluded from the microscopic studies. On the other hand, the deep potentials are found to be similar to a double folding model potential derived from an effective two-body force. One might thus naively understand that the deep potentials, hence the ALAS and nuclear rainbow phenomena, may be a consequence of the strong attractive nature of the nuclear forces. However, in contrast to traditional understanding, I will show that the deep potentials, and therefore also the emergence of a nuclear rainbow and nuclear clustering, are a direct consequence of the Pauli principle. The purpose of this paper is to show that the Pauli principle between nuclei causes a strong Luneburg lens like [universal structural Pauli attraction]{} in the internal region, which is in contrast to the traditional understanding that the Pauli principle causes a repulsive core at short distances. The Luneburg lens like universal Pauli attraction allows the emergence of the simultaneous existence of cluster structures near the threshold energy in the low excitation energy region and a nuclear rainbow in the high energy region. ![\[fig.1\] [(Color online) The experimental angular distribution (points) [@Delbar1978] in $\alpha$+$^{40}$Ca rainbow scattering at $E_L$=61 MeV and the calculated one (red thick solid line), which is decomposed into the farside (red long dashed line) and nearside (red medium dashed line) contributions. The moduli of $S$-matrices, $\|S_L\|$, in the inset (red filled circles) are connected by lines to guide the eye. The angular distribution (brown thin solid line) and its farside component (brown long dashed line) calculated by switching off the imaginary potential (W=0) are also displayed. The cut-off calculations of $L=0-11$ partial waves with and without $W$ are shown by the blue short dashed lines. ]{} ](fig1-40Ca61MeV.eps){width="8.6cm"} In composite particle scattering absorption is mostly strong, which makes it difficult to determine the potential up to the internal region without ambiguity. However there are some exceptions where absorption is weak or incomplete and the nuclear rainbow and ALAS, in which scattering waves penetrate deep into the internal region, are observed in elastic scattering. $\alpha$+$^{16}$O [@Ohkubo1977; @Michel1983; @Michel1998] and $\alpha$+$^{40}$Ca [@Michel1977; @Delbar1978; @Michel1986; @Ohkubo1988; @Michel1998] scattering are such typical examples, for which a global potential, which works over a wide range of energies, has been determined. system $N_0$ $V_0$ (MeV) $N=0$ $N=2$ $N=4$ $N=6$ $N=8$ $N=10$ $N=12$ $N=14$ $N=16$ $N=18$ $N=20$ $N=22$ -------------------- ------- ------------- ------- ------- ------- ------- ------- -------- -------- -------- -------- -------- -------- -------- $\alpha$+$\alpha$ 4 119 0.99 0.99 $\alpha$+$^{16}$O 8 134 0.99 0.98 0.99 0.98 $\alpha$+$^{40}$Ca 12 151 1.00 1.00 1.00 1.00 0.99 0.93 $^{16}$O+$^{16}$O 24 321 0.98 0.93 0.91 0.89 0.89 0.91 0.93 0.96 0.96 0.90 0.74 0.43 \[Table1\] In Fig. 1 the angular distribution in $\alpha$+$^{40}$Ca scattering at $E_L$=61 MeV, calculated using a global potential with a Woods-Saxon squared form factor (thick solid line), which works well over a wide range of energies $E_L$=24 - 166 MeV [@Delbar1978], is compared with the experimental data. The decomposition of the calculated cross sections into the farside and nearside components shows that the minimum at around $\theta=80^\circ$ is caused by farside refractive scattering and is the first order Airy minimum $A1$ of the nuclear rainbow. The global potential can be uniquely determined by reproducing the Airy structure of the nuclear rainbow. The calculations in which the imaginary potential is switched off ($W=0$) show that the minimum at around $\theta=40^\circ$ is a remnant of the Airy minimum $A2$, and the broad bump in the experimental angular distribution in the $\theta=40-80^\circ$ region is a remnant of the $A2$ Airy maximum. The moduli of the $S$-matrices, $\|S_L\|$, which for $L$=0-11 is of the order of $10^{-2}$ (inset), shows that absorption is relatively weak. This makes the observation of a nuclear rainbow possible. The cut-off calculations for the smaller $L$ values, $L$=0-11, (short dashed line) show that these partial waves contribute to the correct description of the Airy structure, which is also confirmed in the same calculations with $W=0$ (short dashed line). The global potential also reproduces well the ALAS [@Delbar1978], the $\alpha$+$^{40}$Ca fusion oscillations in the lower energy region $E_L$=10-27 MeV [@Michel1986B] and the $\alpha$ cluster structure in $^{44}$Ti including the energy levels, $B(E2)$ values and $\alpha$ spectroscopic factors [@Michel1986; @Michel1998; @Yamaya1990; @Yamaya1998]. The semi-microscopic double folding potentials derived from the effective two-body Hasegawa-Nagata-Yamamoto (HNY) force [@Hasegawa1971] and the density-dependent M3Y (DDM3Y) force [@Kobos1982] are similar to the global potential and describe the $\alpha$+$^{40}$Ca system over a wide range of energies as well [@Ohkubo1988; @Atzrott1996]. ![\[fig.2\] [(Color online) The global nuclear potential (black solid lines) and the corresponding Luneburg lens potential (black circles) for the $\alpha$+$^{40}$Ca system. The potential including the Coulomb potentials and corresponding Luneburg lens potential are indicated by long dashed lines (pink) and squares (pink), respectively. The calculated eigenstates for $L=0$ with $N<N_0=12$ together with the $N=N_0$ band head $0^+$ of the $\alpha$+$^{40}$Ca cluster state in $^{44}$Ti are indicated by horizontal solid lines. The eigenenergies of the Luneburg potential including the Coulomb potential are indicated by horizontal dashed lines. ]{} ](fig2.eps){width="5.5cm"} In Fig. 2 the global potential (solid lines) used in Ref.[@Michel1986], D180 with the potential strength -180 MeV, and the potential including the Coulomb potential (long dashed lines) are displayed. The internal region of these potentials resemble the truncated harmonic oscillator (HO) potential well, which is called a Luneburg lens potential, as displayed by the filled circles (black) and squares (pink), respectively. The depth -$V_0$ and the truncation radius $R$, at which the HO potential is zero, are $V_0$=167 MeV and $R$=5.3 fm for the nuclear potential only and $V_0$=146 MeV and $R$=5.2 fm for the combined nuclear and Coulomb potential. In Fig. 2 the states with $N<N_0=12$ are Pauli forbidden, where $N\equiv 2n+L$ with $n$ being the number of the nodes in the wave functions. The $N=12$ state corresponds to the ground state with the $\alpha$+$^{40}$Ca cluster structure in $^{44}$Ti. The eigenenergies of the truncated HO potential (horizontal dashed lines) correspond well to those of the global potential. In Table I the overlap of the calculated wave functions with $N<N_0$ in the global potentials with the redundant Pauli forbidden HO wave functions is almost complete. This means that the redundant Pauli forbidden states of the RGM equations are almost completely embedded in the global potential. The situation is almost the same for the other $L<N_0 -2n$. ![\[fig.3\] [(Color online) Same as Fig. 1 for panel (a) and Fig. 2 for panel (b) but for the $\alpha$+$^{16}$O system with $N_0=8$. In the inset of panel (a) $\|S_L\|$ of the internal waves are additionally indicated by unfilled circles. In panel (b) the triangles (red) are the double folding potentials derived from the HNY force with -538 MeV for the triplet even state in the intermediate range [@Hasegawa1971]. The $N=8$ (solid horizontal line) shows the $\alpha$+$^{16}$O cluster ground state in $^{20}$Ne. ]{} ](fig3.eps){width="7.5cm"} This is true for other systems. In Fig. 3(a) the experimental angular distribution in $\alpha$+$^{16}$O rainbow scattering at $E_L$=49.5 MeV is compared with the one calculated using the phenomenological global potential, which was determined from the systematic analysis of the ALAS and nuclear rainbow scattering [@Michel1983]. The Airy structure with the Airy minimum $A1$ at around $\theta=75^\circ$ followed by the broad Airy maximum $A1$ is well reproduced by the global potential. The Airy structure is brought about by the farside refractive scattering. Furthermore, using the technique in Refs.[@Albinski82; @Michel2001], the Airy structure is found to be caused by the interference between the two sub-amplitudes, i.e., the farside-subcomponent of the internal-waves, which penetrate the potential barrier at the surface into the internal region, and the farside-subcomponent of the barrier waves, which are reflected at the barrier. In fact, in the inset of Fig. 3(a) the moduli of the $S$-matrix of the internal waves calculated using the technique in Ref.[@Albinski82] are significantly large. If one cuts off the contributions of the partial waves for $L=$0-7 (blue medium-dashed lines), the Airy minimum is destroyed in disagreement with the experiment. This means the waves with smaller $L$ values contribute to the correct reproduction of the Airy structure, i.e., to constraining the shape and depth of the potential in the internal region. In Fig. 3(b) the global potential for the $\alpha$+$^{16}$O system with the energy-dependent parameter $\alpha$=3.02 in Ref.[@Michel1983] is shown. This potential, which reproduces well the observed $\alpha$ cluster structure in $^{20}$Ne, the energy levels, $B(E2)$ values and $\alpha$ widths [@Michel1983; @Michel1998], resembles the semi-microscopic double folding potentials derived from the HNY force (triangles in Fig. 3) [@Michel1983; @Ohkubo1977] and the DDM3Y force [@Abele1993; @Hirabayashi2013] well. The internal region of the global potential (solid line) is well simulated by the Luneburg lens truncated HO potential with $V_0$=131 MeV and $R$=3.75 fm (filled circles) and, when the Coulomb potential is added (long dashed line), with $V_0$=144 MeV and $R$=3.9 fm (filled squares). The $N=8$ state corresponds to the ground state with the $\alpha$+$^{16}$O cluster structure in $^{20}$Ne. Table I shows that the overlap of the wavefunctions of the states with $N<$8 and the redundant Pauli forbidden HO wave functions is almost complete. ![\[fig.4\] [(Color online) Same as Fig. 2 but for the $\alpha$+$\alpha$ system with $N_0=4$. The solid horizontal line shows the $N=4$ ground state with $\alpha$+$\alpha$ cluster structure in $^{8}$Be. ]{} ](fig4.eps){width="5.5cm"} ![(Color online) Same as Fig. 2 but for the $^{16}$O+$^{16}$O system with $N_0=24$. The solid horizontal line shows the $N=24$ $0^+$ state with $^{16}$O+$^{16}$O cluster structure in $^{32}$S. \[fig5\] ](fig5.eps){width="5.5cm"} In Fig. 4 the global potential for $\alpha$+$\alpha$ [@Buck1977], which reproduces the experimental phase shifts of elastic scattering over a wide range of energies, is displayed. The internal region of the potential without (with) the Coulomb potential is well simulated by the truncated HO potential with $V_0$=120 MeV and $R$=2.5 fm (116 MeV and 2.4 fm). In Fig. 5 the global potential for $^{16}$O+$^{16}$O in Ref.[@Ohkubo2002] is displayed. The global potential reproduces the rainbow scattering [@Stiliaris1989; @Khoa2000; @Nicoli1999; @Michel2001; @Khoa2007] and molecular structure with the $^{16}$O+$^{16}$O cluster structure in a unified way [@Ohkubo2002]. The internal region of the potentials resemble the Luneburg lens truncated HO potentials well, with $V_0$=266 MeV and $R$=4.7 fm when the Coulomb potential is included and with 310 MeV and 4.8 fm for the nuclear potential only. For the $^{16}$O+$^{16}$O system it is noted that the region $r<2$ fm has some ambiguity and the slightly different folding potential in this region can reproduce the $^{16}$O+$^{16}$O scattering equally well [@Nicoli1999]. In Table I the overlap of the states for $N < N_0$ with the redundant Pauli forbidden HO wave functions is large except for $N=22$ near the threshold energy. Thus the physical wave functions with $N \geq N_0$ generated by the global potentials are found to be orthogonal to the redundant Pauli forbidden HO wave functions in the RGM. This orthogonality is closely related to the shape and depth of the potential in the internal region, i.e., the Luneburg lens like truncated HO potential. I will now show theoretically that the global potentials have a Luneburg lens like universal Pauli attraction in the internal region. The RGM equation for the antisymmetrized wave function for two clusters that are assumed to have HO shell model wave functions with the size parameter $\nu$ and spin 0 is given by $$\begin{aligned} (T_r+V_D(r)-E)\chi_{\it L}({\it r}) + \int K(r,r^\prime)dr^\prime \chi_{\it L}({\it r}^\prime) & = & 0 \label{eq:eq4}\end{aligned}$$ where $\chi_{\it L}({\it r})$, $T_r$, $V_D(r)$, $E$ and $K(r,r^\prime)$ are the relative wave function, kinetic energy operator of the relative motion, direct (double folding) potential, relative energy and exchange kernel, respectively. Since one knows that the local potential works very well, if one approximate $K(r,r^\prime) = V_{P}(r) \delta(r-r^\prime)$, Eq. (\[eq:eq4\]) becomes a local potential equation $$\begin{aligned} \{T_r + V_D(r) + V_{P}(r)-E\}\chi_{\it L}({\it r}) & = & 0, \label{eq:eq7}\end{aligned}$$ with the local potential $V(r)\equiv V_D(r) + V_{P}(r)$. One can impose the eigenfunctions $\chi_{\it L}^{(n)}({\it r})$ with $n< (N_0 -L)/2$ to satisfy $$\begin{aligned} \{T_r + V_{HO}(r)-(2n+L+3/2)\hbar \omega\} \chi_{\it L}^{(n)}({\it r}) & = & 0, \label{eq:eq8}\end{aligned}$$ where $V_{HO}(r)$ is the HO potential with a depth $- V_0$ at $r=0$ and the size parameter $\nu$. This guarantees that the physical wave functions of Eq. (\[eq:eq7\]) with $N\ge N_0$ are orthogonal to the redundant Pauli forbidden states. This is satisfied when $V(r)=V_{HO}(r) $ in the internal region $r < R$, where $R$ is the size of the Luneburg lens, which is a HO potential truncated at $r=R$ as given below. Thus in order that the wave functions of the physical states satisfy the Pauli principle, the local potential should resemble a deep HO potential in the internal region, i.e., a Luneburg lens potential. When the $V_D(r)$ itself resembles a deep HO as seen for the HNY force (triangles) in Fig. 3(b), the $V_{P}(r)$ is small [@Ohkubo1977]. On the other hand, when $V_D(r)$ is repulsive (in the case of, for example, Brink-Boeker force B1), the $V_{P}(r)$ must be deep so that the $V(r)$ resembles a Luneburg lens potential. Thus the Pauli principle plays the role generating a $ V_{P}(r)$ so that the $V(r)$ resembles a Luneburg lens like HO potential in the internal region. A Luneburg lens with a radius $R$ is a lens that refracts all the parallel incident trajectories to the focus $R_f$ ($< R$). For such a lens the refractive index $n$ is given by $$n^2(r \leq R) = ({R_f^2-r^2+R^2})/{R_f^2}, \quad n(r > R) = 1.$$ The potential having this property [@Michel2002] is $$V(r \leq R) = V_0 \left( {r^2}/{R^2}-1 \right),\quad V(r > R) = 0,$$ where $V_0 = E (R/R_f)^2$ is the depth at $r=0$. This is a HO potential truncated at $r=R$. The outer region of the nuclear potential has a diffuse surface, and so deviates from the ideal Luneburg lens. This causes astigmatism to occur, which is nothing but the emergence of a nuclear rainbow. Thus the emergence of the nuclear rainbow is due the properties of both the Luneburg lens like potential in the internal region and the diffuse attraction in the outer region. The values of the strength of the potential in Table I are consistent with those evaluated from the constraint of the Pauli principle at $E$=0, i.e., $V_0 = (N_0+3/2)\hbar \omega$, which are 121, 125, 157 and 305 MeV for $\alpha$+$\alpha$, $\alpha$+$^{16}$O, $\alpha$+$^{40}$Ca and $^{16}$O+$^{16}$O, respectively. Thus the deep nature of the potential is a direct consequence of the Pauli principle. This explains why a rainbow occurs in nuclear scattering in the potentials that generate cluster states at lower energies, near the threshold energy. A deep double folding potential derived from a density-dependent effective two-body force, such as the DDM3Y force [@Kobos1982], resembles the Luneburg lens like potential and has been successfully used in scattering and structure studies [@Khoa2007; @Atzrott1996; @Abele1993; @Hirabayashi2013; @Khoa2000; @Nicoli1999]. According to the present study, for which the depth and shape in the internal region are determined to be constrained by the Pauli principle, it seems that the Pauli principle manifests itself through the density-dependence as well as the exchange terms [@Gupta1984; @Chaudhuri1985; @Khoa1988]. From the structure viewpoint, shell model wave functions in the HO potential, which have almost complete overlaps with the Pauli forbidden states embedded in the local potential as shown in Table I, are equivalent to the cluster representation of Wildermuth and Kanellopoulos [@Wildermuth1958; @Wildermuth1966; @Bayman1958] and can be represented by the $SU(3)$ model [@Elliott1958]. This wave function with a Gaussian tail is damped at the surface. On the other hand, in the present local potential cluster model, which has an attractive potential with an exponential tail at the surface similar to a Woods-Saxon potential, the inner oscillations of the wave function are damped [@Tamagaki1965; @Tamagaki1968; @Hiura1972] due to the orthogonality to the Pauli forbidden states with $N<N_0$ embedded in the Luneburg lens like potential. This brings about the enhancement of the amplitude of the wave function at the surface, i.e., emergence and development of cluster structure. Thus the Pauli principle plays the dual role of causing (i) the [shell model]{} potential with a deep HO shape as the structural Pauli attraction and (ii) the [cluster structure]{} with the damped inner oscillations and enhanced surface amplitude by the orthogonality to the embedded Pauli forbidden states due to (i). When two nuclei with a typical shell model structure such as the double magic nucleus $\alpha$ particle, $^{16}$O, $^{40}$Ca and $^{208}$Pb come closer, the universal Pauli attraction inevitably makes possible the emergence of a cluster structure slightly above the highest Pauli-forbidden state because of the diffuse surface, i.e., near the threshold energy of the compound system. This is the reason why the $\alpha$ cluster structure typically appears for the double magic core + $\alpha$ systems like $^8$Be, $^{20}$Ne, $^{44}$Ti and $^{212}$Po. The shell model structure of the internal constituent nuclei (dynamical nature) and the existence of redundant Pauli forbidden states due to the Pauli principle in the wave functions of the relative motion (kinematical nature) are closely interrelated for the emergence of both the cluster structures in the compound system and nuclear rainbows. The Pauli principle does not only provide the [raison d’être]{} for the shell structure of nuclei but also for the emergence of the cluster structure near the threshold. This will not be limited to closed nuclei and two nuclear systems as long as redundant Pauli forbidden states exist. Also, the nucleon-nucleon potential may have a strong Pauli attraction due to the Pauli principle rather than the repulsive core [@Otsuki1964; @Otsuki1965; @Aoki2012], the effects of which could be seen in few body systems. To summarize, the existence of a Luneburg lens like universal structural Pauli attraction in the internal region of nucleus-nucleus interaction has been shown. This is different from the traditional view that a structural repulsive core exists at short distances. It is found that the depth and the shape of the potential in the internal region is constrained to a Luneburg lens like truncated harmonic oscillator potential by the Pauli principle. The present work reinforces the empirical threshold rule, which had intuitively been understood to be due to the saturation property of the nuclear force. The emergence of a cluster state near the threshold energy can now be seen as a consequence of both the Pauli principle in the internal region and diffuse attraction in the outer region. The emergence of cluster structures and rainbows are unified as a consequence of a Luneburg lens like universal structural Pauli attraction in the internal region (kinematical) and diffuse attraction in the outer region (dynamical). The author thanks the Yukawa Institute for Theoretical Physics, Kyoto University for the hospitality extended during a stay in February 2016 where this work was completed. [aa]{} A. Bohr and B. R. Mottelson: [Nuclear Structure]{}, Vol. I, (W. A. Benjamin, New York, 1969). I. Shimodaya, R. Tamagaki, and H. Tanaka, Prog. Theor. Phys. [27]{}, 793 (1962); [25]{}, 853 (1961). R. Tamagaki and H. Tanaka, Prog. Theor. Phys. [34]{}, 191 (1965). R. Tamagaki, Prog. Theor. Phys. Suppl. [E68]{}, 242 (1968) and references therein. J. Hiura and R. Tamagaki, Suppl. Prog. Theor. Phys. [52]{}, 25 (1972). S. Ali and A. R. Bodmer, Nucl. Phys. [A80]{}, 99 (1966). S. Otsuki, R. Tamagaki and M. Yasuno, in [Commemoration issue for the 30th of the Meson theory by Dr. H. Yukawa]{} Prog. Theor. Phys. Suppl. [E65]{}, 578 (1965). A. Tohsaki, F. Tanabe, and R. Tamagaki, Prog. Theor. Phys. [53]{}, 1022 (1975). T. Ando, K. Ikeda and A. Tohsaki-Suzuki, Prog. Theor. Phys. [59]{}, 306 (1978). B. Buck, H. Friedrich, and C. Wheatley, Nucl. Phys. [A275]{}, 246 (1977). P. E. Hodgson, [Nuclear Reactions and Nuclear Structure]{} (Clarendon Press, Oxford 1971); P. E. Hodgson, [Nuclear Heavy Ion Reactions]{} ( Clarendon Press, Oxford, 1978). D. M. Brink, [[Semi-classical Methods for Nucleus-Nucleus Scattering]{}]{} (Cambridge University Press, Cambridge, 1985). S. Ohkubo, Y. Kondo, and S. Nagata, Prog. Theor. Phys. [57]{}, 82 (1977). F. Michel [et al.]{}, Phys. Rev. C [28]{}, 1904 (1983). F. Michel and P. Vanderpoorten, Phys. Rev. C [16]{}, 142 (1977). Th. Delbar [et al.]{}, Phys. Rev. C [18]{}, 1237 (1978). E. Stiliaris [et al.]{}, Phys. Lett. [B 223]{}, 291 (1989). D. T. Khoa, W. von Oertzen, H. G. Bohlen, and S. Ohkubo, J. Phys. [G 34]{}, R111 (2007) and references therein. F. Michel, S. Ohkubo, and G. Reidemeister, Prog. Theor. Phys. Suppl. [132]{}, 7 (1998); S. Ohkubo, T. Yamaya, and P. E. Hodgson, Nuclear clusters, in [Nucleon-Hadron Many-Body Systems]{}, (edited by H. Ejiri and H. Toki) (Oxford University Press, Oxford, UK,1999), p. 150 and references therein. F. Michel, G. Reidemeister, and S. Ohkubo, Phys. Rev. Lett. [57]{}, 1215 (1986); Phys. Rev. C [37]{}, 292 (1988). S. Ohkubo, Phys. Rev. C [38]{}, 2377 (1988). F. Michel, G. Reidemeister, and S. Ohkubo, Phys. Rev. [34]{}, 1248 (1986). T. Yamaya, S. Oh-ami, M. Fujiwara, T. Itahashi, K. Katori, M. Tosaki, S. Kato, S. Hatori, and S. Ohkubo, Phys. Rev. C [42]{}, 1935 (1990). T. Yamaya, K. Katori, M. Fujiwara, S, Kato, and S. Ohkubo, Prog. Theor. Phys. Suppl. [132]{}, 73 (1998). A. Hasegawa and S. Nagata, Prog. Theor. Phys. [45]{}, 1786 (1971); Y. Yamamoto, Prog. Theor. Phys. [52]{}, 471 (1974). A. M. Kobos [et al.]{}, Nucl. Phys. [A384]{}, 65 (1982); A. M. Kobos [et al.]{}, Nucl. Phys. [A425]{}, 205 (1984). U. Atzrott, P. Mohr, H. Abele, C. Hillenmayer, and G. Staudt, Phys. Rev. C [53]{}, 1336 (1996). J. Albiński and F. Michel, Phys. Rev. C [25]{}, 213 (1982). F. Michel, G. Reidemeister, and S. Ohkubo, Phys. Rev. [63]{}, 034620 (2001); F. Michel, F. Brau, G. Reidemeister, and S. Ohkubo, Phys. Rev. Lett. [85]{}, 1824 (2000). H. Abele and G. Staudt, Phys. Rev. C [47]{}, 742 (1993). Y. Hirabayashi and S. Ohkubo, Phys. Rev. C [88]{}, 014314 (2013) and references therein. S. Ohkubo and K. Yamashita, Phys. Rev. C [66]{}, 021301(R) (2002). D. T. Khoa, W. von Oertzen, H. G. Bohlen, and F. Nuoffer, Nucl. Phys. [A672]{}, 387 (2000). M. P. Nicoli [et al.]{}, Phys. Rev. C [60]{}, 064608 (1999). F. Michel, G. Reidemeister, and S. Ohkubo, Phys. Rev. Lett. [89]{}, 152701 (2002). D. T. Khoa, Nucl. Phys. [A484]{}, 386 (1988). S. K. Gupta and B. Sinha, Phys. Rev. C [30]{}, 1093 (1985). A. K. Chaudhuri, D. N. Basu, and B. Sinha, Nucl. Phys. [A439]{}, 415 (1985); A. K. Chaudhuri and B. Sinha, Nucl. Phys. [A455]{}, 169 (1986). S. Otsuki, R. Tamagaki, and M. Wada, Prog. Theor. Phys. [32]{}, 220 (1964); S. Machida and M. Namiki; Prog. Theor. Phys. [33]{}, 125 (1965). K. Wildermuth and Th. Kanellopoulos, Nucl. Phys. [7]{}, 150 (1958), Nucl. Phys. [9]{}, 449 (1958/59). K. Wildermuth and W. McClure, Springer Tracts in Modern Physics, Vol. 41, [Cluster Representation of Nuclei]{} (Springer Verlag, Berlin, 1966). B. F. Bayman and A. Bohr, Nucl. Phys. [9]{}, 596 (1958/59). J. P. Elliott, Proc. Roy. Soc. London [A 245]{}, 562 (1958). S. Aoki, J. Balog, and P. Weisz, Prog. Theor. Phys. [128]{}, 1269 (2012).	{ "pile_set_name": "ArXiv" }
--- author: - Hideki Okawa - Josh Kunkle - Elliot Lipeles bibliography: - 'zhinv\_whitepaper.bib' title: \| Prospects on the search for invisible Higgs decays in the $ZH$ channel at the LHC and HL-LHC:\ A Snowmass White Paper ---	{ "pile_set_name": "ArXiv" }
--- abstract: \| This is the first of a series of papers describing a numerical implementation of the conformally rescaled Einstein equation, an implementation designed to calculate asymptotically flat spacetimes, especially spacetimes containing black holes.\ In the present paper we derive the new first order time evolution equations to be used in the scheme. These time evolution equations can either be written in symmetric hyperbolic or in flux-conservative form. Since the conformally rescaled Einstein equation, also called the conformal field equations, formally allow us to place the grid boundaries outside the physical spacetime, we can modify the equations near the grid boundaries and get a consistent and stable discretisation. Even if we calculate spacetimes containing black holes, there is no need for introducing artifical boundaries in the physical spacetime, which then would complicate, influence, or even exclude the computation of certain spacetime regions. author: - \| [Peter Hübner]{}\ ([email protected])\ Max-Planck-Institut für Gravitationsphysik\ Albert-Einstein-Institut\ Schlaatzweg 1\ D-14473 Potsdam\ FRG title: How to Avoid Artificial Boundaries in the Numerical Calculation of Black Hole Spacetimes --- \ [short title: Calculating Black Hole Spacetimes]{}\ [PACS numbers: 0420G, 0420H, 0430]{}\	{ "pile_set_name": "ArXiv" }
--- abstract: \| games are very useful in studying separation and equivalence results in logic. The standard finite game characterizes equivalence in first order logic. The standard game in infinitary logic characterizes equivalence in $L_{\infty\omega}$. The logic $\LL$ is the extension of first order logic with countable conjunctions and disjunctions. There was no game for $\LL$ in the literature. In this paper we develop an Game for $\LL$. This game is based on a game for propositional and first order logic introduced by Hella and Väänänen. Unlike the standard games which are modeled solely after the behavior of quantifiers, this new game also takes into account the behavior of connectives in logic. We prove the adequacy theorem for this game. We also apply the new game to prove complexity results about infinite binary strings. author: - \| Jouko Väänänen[^1]\ Department of Mathematics and Statistics\ University of Helsinki\ and\ Institute for Logic, Language and Computation,\ University of Amsterdam\ and\ Tong Wang[^2]\ Institute for Logic, Language and Computation,\ University of Amsterdam bibliography: - 'EF.bib' title: 'An Ehrenfeucht-Fraïssé Game for $L_{\omega_1\omega}$' --- Introduction ============ One major limitation of the expressive power of first order logic is that many familiar mathematical concepts (e.g. finiteness, connectedness of graphs, and well-foundedness) are not definable. These concepts involve, explicitly or implicitly, the notion of infinity which is difficult to handle in first order logic, partly because first order logic only has [finitary]{} formulas. One way to transcend this boundary is to consider stronger logics where [infinitary]{} formulas are also admissible. Infinitary logics made their first appearance in print with the papers of Scott and Tarski [@ScottTarski58] and Tarski [@Tarski58]. Among all infinitary logics, the logic $\LL$ turns out to be an especially interesting example. The logic $\LL$ adds to first order logic countable[^3] conjunctions and disjunctions of formulas. It is in a sense the smallest infinitary logic. It strikes a nice balance—$\LL$ is expressive enough to define many familiar mathematical concepts which are not definable in first order logic (including the three aforementioned), while it still enjoys model-theoretic properties such as having the Completeness Theorem [@Karp64] and the Interpolation Theorem [@Lopez-Escobar65].** Throughout this paper we assume $L$ to be a relational vocabulary. The terms of $L$ are defined as usual. The set of $\LL$-formulas of $L$ is the smallest set closed under the following operations: 1. If $t_1$ and $t_2$ are terms, then $t_1=t_2$ is a formula. 2. If $R$ is an $n$-place relational symbol and $t_1,\ldots, t_n$ are terms, then $Rt_1\ldots t_n$ is a formula. 3. If $\phi$ is a formula, then so is $\neg\phi$. 4. If $\phi$ and $\psi$ are formulas, then so is $\phi\vee\psi$. 5. If $\phi$ and $\psi$ are formulas, then so is $\phi\wedge\psi$. 6. If $I$ is a countable set and for every $i\in I$, $\phi_i$ is a formula, then so is $\bigvee_{i\in I}\phi_i$. 7. If $I$ is a countable set and for every $i\in I$, $\phi_i$ is a formula, then so is $\bigwedge_{i\in I}\phi_i$. 8. If $\phi$ is a formula and $x_n$ a variable, then $\exists x_n \phi$ is a formula. 9. If $\phi$ is a formula and $x_n$ a variable, then $\forall x_n \phi$ is a formula. Rather surprisingly, some questions in the basic model theory of $\LL$ have been left unanswered. In this paper we address one such question, namely the question of separation in $\LL$. Let us fix some notation. We use $A$, $B$ to denote $L$-structures and use $\A$, $\B$ to denote classes of $L$-structures. We say that a formula $\phi$ [separates]{} a pair of structures $A$ and $B$ if $A\models \phi$ and $B\not\models \phi$. We say that a formula $\phi$ [separates]{} two classes of structures $\mA$ and $\mB$, denoted by $(\mA, \mB)\models \phi$, if for all structures $A\in\mA$ we have $A\models \phi$ and for all structures $B\in \mB$ we have $B\not\models \phi$. Let $A\equiv B$ denote that $A$ and $B$ are elementarily equivalent. If $\alpha$ is an ordinal, let $A\equiv^\alpha_{\infty\omega} B$ denote that $A$ and $B$ satisfy the same $L_{\infty\omega}$ [^4] sentences up to quantifier rank $\alpha$. Finally we denote by $A\equiv_{\infty\omega} B$ that $A$ and $B$ are equivalent in $L_{\infty\omega}$. Note that the notions of separation and equivalence are opposite sides of the same coin: two structures can be separated by a formula in a certain language if and only if they are not equivalent in this language. The central question that we try to answer in this paper is the following:** Question 1: : Given two structures $A$ and $B$, is there a formula $\phi$ in $\LL$ that separates them? More generally, given two classes of structures $\mA$ and $\mB$, is there a formula $\phi$ in $\LL$ that separates them? A very useful machinery in logic to study equivalence and separation results is the game, also known as the back-and-forth game. The game was first formulated for first order logic by Ehrenfeucht [@Ehrenfeucht57; @Ehrenfeucht6061], based on the ideas of Fraïssé [@Fraisse55]. Let us denote this game by EF$_n$. For $n$ a natural number, the game EF$_n$ characterizes equivalence in first order logic up to formulas of quantifier rank $n$. Karp [@Karp64] developed the idea into a characterization of equivalence in infinitary logic using back-and-forth sequences, which was later formulated as a game in [@Benda69]. See also [@Vaananen11 Chapter 7]. For $\alpha$ an ordinal, the game EFD$_\alpha$ characterizes equivalence in $L_{\infty\omega}$ up to formulas of quantifier rank $\alpha$. The logic $\LL$ is intermediate between first order logic and $L_{\infty\omega}$. It is only natural to ask: is there an game that, standing somehow in between the standard games EF$_n$ and EFD$_\alpha$, characterizes equivalence in $\LL$? The question of finding such a game has been open since 1970s. In this paper we introduce an game for $\LL$. The text is divided as follows. In Section 2 we analyze the reason why the standard game EFD$_\alpha$ fails to characterize equivalence in $\LL$. The idea is roughly that the standard game only models the behavior of quantifiers in logic, and while this is enough for studying separation in $L_{\infty\omega}$, it is not enough for $\LL$ since the latter is defined by explicit reference to the [countability]{} of conjunctions and disjunctions. An game for the latter needs to model also the behavior of connectives. Fortunately there already exists an game for first order logic that is in this spirit, namely the game EFB$_n$ introduced by Hella and Väänänen [@HellaVaananen]. We present this game in Section 3. In Section 4 we extend this game to cover $\LL$-formulas and we prove the adequacy theorem for the new game. In Section 5 we consider other alternatives for an game for $\LL$. In Section 6 we test our new tool on the complexity of information coded by infinite binary strings. The Standard Game ================= Let us first recapitulate the standard game in infinitary logic. Let $L$ be a relational vocabulary and $A$, $B$ be $L$-structures. A partial mapping $\pi: A\rightarrow B$ is a [partial isomorphism]{} if it is an isomorphism between $A\upharpoonright{\rm{dom}(\pi)}$ and $B\upharpoonright{\rm{ran}(\pi)}$. Let $L$ be a relational vocabulary and $A$, $B$ be $L$-structures. Let $\alpha$ be an ordinal. The game EFD$_\alpha(A, B)$ has two players. The number $\alpha$ is called the [rank]{} of the game. The positions in the game are of the form $(\bar{a}, \bar{b}, \gamma)$, where $\bar{a}$ and $\bar{b}$ are tuples in $A$ and $B$ respectively and $\gamma$ is an ordinal. The game begins from the position $(\emptyset, \emptyset, \alpha)$. Suppose the game is in position $(\bar{a}, \bar{b}, \gamma)$, $\gamma>0$, there are the following possibilities for the next move in the game. 1. Player [I]{} picks an element $c\in A$. He also picks an ordinal $\beta<\gamma$. Player [II]{} responds by picking an element $d\in B$. The game continues from the position $(\bar{a}c, \bar{b}d, \beta)$.** 2. Player [I]{} picks an element $d\in B$. He also picks an ordinal $\beta<\gamma$. Player [II]{} responds by picking an element $c\in A$. The game continues from the position $(\bar{a}c, \bar{b}d, \beta)$. The game terminates in a position $(\bar{a}, \bar{b}, \gamma)$ if $\gamma=0$. Player [II]{} wins if $\bar{a}\mapsto \bar{b}$ is a partial isomorphism from $A$ to $B$. Otherwise player [I]{} wins. The following result is essentially due to Karp [@Karp64], although it was originally formulated in terms of back-and-forth sequences. Let $L$ be a relational vocabulary, $A$ and $B$ be $L$-structures and $\alpha$ be an ordinal. Then the following are equivalent: 1. $A\equiv^\alpha_{\infty\omega} B$. 2. Player [II]{} has a winning strategy in $_\alpha(A, B)$. The game EFD$_\alpha$ provides important information about our Question 1. Make the following simple observation. \[p2\] For any formula $\phi\in \LL$, the quantifier rank of $\phi$ is a countable ordinal. \[c3\] If player [II]{} has a winning strategy in $\mathrm{EFD}_{\omega_1}(A, B)$, then there is no formula in $\LL$ that separates $A$ and $B$.** The standard games EF$_n$ and EFD$_\alpha$ characterize a hierarchy of equivalence relations. They provide in a sense upper bounds and lower bounds for Question 1. The situation can be illustrated by Figure \[f1\]. ![Hierarchy of equivalence relations[]{data-label="f1"}](Figure.pdf){width="6cm" height="5cm"} If two structures $A$ and $B$ are not even elementarily equivalent, then obviously there is an $\LL$-formula separating the two because this formulas is already in first order logic. On the other hand, if $A$ and $B$ are $L_{\infty\omega}$-equivalent, then there is no hope of finding an $\LL$-formula separating the two structures. Corollary \[c3\] shows that we can improve the upper bound: if two structures are equivalent in $L_{\infty\omega}$ up to formulas of quantifier rank $\omega_1$, then they cannot be separated by any formula in $\LL$. Therefore, from the point of view of $\LL$ all the excitement concentrates in the interval between $A\equiv B$ and $A\equiv^{\omega_1}_{\infty\omega}B$. However the game EFD$_\alpha$ is not able to tell us much about what happens in this interval. The reason is that although a small number of connectives implies a small quantifier rank (as in Proposition \[p2\]), there is no implication in the reverse direction. We may very well have a separation formula of very low quantifier rank, yet it contains a large number of conjunctions and disjunctions, and the standard game cannot detect the existence of such a formula. In particular, the standard game does not answer the question whether there is a quantifier-free formula in $\LL$ separating two structures, and still, the elimination of quantifiers in favor of connectives is the oldest application of infinitary logic.[^5] The failure of the standard game in characterizing equivalence in $\LL$ can be understood in the following way. An game, meaning a set of rules and winning conditions, corresponds to a way of measuring the complexity of formulas. If a game on the one hand, and a complexity measure on the other hand, form a good match, we would have the adequacy theorem for this game. In the case of the game EFD$_\alpha$, the complexity of a formula is measured by its quantifier rank. The two types of moves in the game EFD$_\alpha$ model the nature of existential and universal quantifiers respectively. The logic $\LL$, however, is defined by reference to the countability of conjunctions and disjunctions, and quantifier rank simply does not capture this information. Therefore there is a sense in saying that the game EFD$_\alpha$ does not solve the question of separation in $\LL$ because it measures the complexity of $\LL$-formulas in an unsuitable way. In this light, we need two things in order to develop an game for $\LL$: (a) a new game and (b) a new measure of the complexity of infinitary formulas. These two components have to be compatible so that we can have the adequacy theorem. Moreover, both of them should take into account not only the behavior of quantifiers, but also of connectives. A Game for First Order Logic ============================ What is a suitable complexity measure for $\LL$-formulas then? We draw inspiration from the game EFB$_n$, proposed by Hella and Väänänen in [@HellaVaananen] as a refinement of the standard game for first order logic. This game resembles the game for independence friendly logic, presented in [@Vaananen02]. An innovative feature of the game EFB$_n$ is that it measures the complexity of a formula by making use of our intuition about the length of a formula. Let $L$ be a relational vocabulary as usual. Henceforth we shall assume all formulas to be in negation normal form. \[d5\] The [size]{}, denoted by $\mathrm{s}(\phi)$, of a formula $\phi$ in first order logic is defined inductively as follows: $$\begin{aligned} \mathrm{s}(\phi)&=1 \textrm{ if } \phi \textrm{ is an atomic or negated atomic formula},\\ \mathrm{s}(\phi \wedge \psi)&= \mathrm{s}(\phi)+\mathrm{s}(\psi),\\ \mathrm{s}(\phi\vee \psi)&=\mathrm{s}(\phi)+\mathrm{s}(\psi),\\ \mathrm{s}(\exists x \phi)&=\mathrm{s}(\phi)+1,\\ \mathrm{s}(\forall x \phi)&= \mathrm{s}(\phi)+1. \end{aligned}$$ The size of a first order formula is a positive integer. It is always bigger or equal to the quantifier rank. To define the game EFB$_n$ we need some notation. We use $x_j, j\in {{{\Bbb}N}}$ to denote variables. A variable assignment for a structure $A$ is a finite partial mapping $\alpha: {{{\Bbb}N}}\rightarrow A$. The finite domain of $\alpha$ is denoted by dom$(\alpha)$. The game EFB$_n$ is defined on pairs of classes of structures rather than on pairs of structures. We consider classes $\mA$ of structures $(A, \alpha)$, where $A$ is a model and $\alpha$ is a variable assignment on $A$. We assume that whenever $(A, \alpha), (B, \beta)\in \mA$, then $A$ and $B$ have the same vocabulary, and $\alpha$ and $\beta$ have the same domain, which we denote by dom$(\mA)$. If $\alpha$ is an assignment on $A$, $a\in A$ and $j\in {{{\Bbb}N}}$, then $\alpha(a/j)$ is the assignment that maps $j$ to $a$ and agrees with $\alpha$ otherwise. If $F$ is a choice function on $\mA$, namely that $F$ is a function defined on $\mA$ such that $F(A, \alpha)\in A$ for all $(A, \alpha)\in \mA$, then $\mA(F/j)$ is defined as $\{(A, \alpha(F(A, \alpha)/j)\|(A, \alpha)\in \mA\}$. Finally, $\mA(\star/j)=\{(A, \alpha(a/j))\|(A, \alpha)\in \mA, a\in A)\}$. Let $L$ be a relational vocabulary and $\mA_0$, $\mB_0$ classes of $L$-structures. Let $n$ be a positive integer. The game $_n(\mA_0, \mB_0)$ has two players. The number $n$ is called the [rank]{} of the game. The positions in the game are of the form $(\mA, \mB, m)$ where $\mA, \mB$ are classes of $L$-structures and $m\in {{{\Bbb}N}}$. The game begins from position $(\mA_0, \mB_0, n)$. Suppose the game is in position $(\mA, \mB, m)$. There are the following possibilities for the next move in the game. Left splitting move: : Player [I]{} first represents $\mA$ as a union $\mA_1\cup \mA_2$. He also chooses positive numbers $m_1$ and $m_2$ such that $m_1+m_2=m$. Now the game continues from the position $(\mA_1, \mB, m_1)$ or from the position $(\mA_2, \mB, m_2)$, and player [II]{} can choose which.** Right splitting move: : Player [II]{} first represents $\mB$ as a union $\mB_1\cup \mB_2$. He also chooses positive numbers $m_1$ and $m_2$ such that $m_1+m_2=m$. Now the game continues from the position $(\mA, \mB_1, m_1)$ or from the position $(\mA, \mB_2, m_2)$, and player [II]{} can choose which. Left supplementing move: : Player [I]{} picks an element from each structure $(A, \alpha)\in \mA$. More precisely, [I]{} chooses a natural number $j$ and a choice function $F$ for $\mA$. Then the game continues from the position$(\mA(F/j), \mB(\star/j), m-1)$. Right supplementing move: : Player [I]{} picks an element from each structure $(B, \beta)\in \mB$. More precisely, [I]{} chooses a natural number $j$ and a choice function $F$ for $\mB$. Then the game continues from the position$(\mA(\star/j), \mB(F/j), m-1)$. The game ends in a position $(\mA, \mB, m)$ and player [I]{} wins if there is an atomic or a negated atomic formula $\phi$ such that $(\mA, \mB)\models \phi$. Player [II]{} wins the game if they reach a position $(\mA, \mB, m)$ such that $m=1$ and [I]{} does not win in this position.**** The game EFB$_n$ characterizes separation in first order logic up to formulas of size $n$: \[10\] Let $L$ be a relational vocabulary, $\mA$ and $\mB$ be classes of $L$-structures, and let $n$ be a positive integer. Then the following are equivalent: 1. Player [I]{} has a winning strategy in the game $_n(\mA, \mB)$.** 2. There is a first order $L$-formula of size $\leq n$ such that $(\mA, \mB)\models \phi$. The proof of this theorem is postponed to the next section, where we prove the adequacy theorem for the generalized game. For now the interested reader is referred to [@HellaVaananen]. Extending the Game to $L_{\omega_1\omega}$ ========================================== The game EFB$_n$ sets up a good model in our quest. Our next step is to extend this game to $\LL$. Again, there are two things that we need to do: a) modifying the rules of the game, and b) extending the definition of size to cover $\LL$-formulas. We shall start from the latter. At a first glance, the most straightforward way of extending Definition \[d5\] to infinitary formulas is to leave the clauses for atomic formulas, quantifiers, finite conjunctions and disjunctions unchanged, and define the size of a countable conjunction or disjunction as the supremum of the size of its proper subformulas. However this approach leads to undesirable consequences. Here we encounter the interesting phenomenon that the arithmetic of infinite ordinals behave differently from the finite case. For one thing, addition of infinite ordinals is not commutative. Consider for example a formula $\phi$ with $\mathrm{s}(\phi)=\omega$ and a formula $\psi$ with $\mathrm{s}(\psi)=1$. We would have that $\mathrm{s}(\phi\wedge\psi)=\omega+1$, which is not equal to $\mathrm{s}(\psi\wedge\phi)=1+\omega=\omega$. This is counterintuitive. Moreover, in this example the formula $\psi\wedge \phi$ has the same size as its proper subformula $\phi$, a fact that is fatal to the induction proof of the adequacy theorem. We need a new way of adding up ordinals. The operation that we turn to is the [natural sum]{}. The natural sum is defined using the Cantor normal form of ordinals. Let $\gamma_1$ and $\gamma_2$ be ordinals. One can represent $\gamma_1$ and $\gamma_2$ uniquely in the form $$\gamma_1=\omega^{\alpha_1}\cdot k_1+\ldots+\omega^{\alpha_n}\cdot k_n$$ $$\gamma_2=\omega^{\alpha_1}\cdot j_1+\ldots+\omega^{\alpha_n}\cdot j_n$$ where $\alpha_1>\ldots>\alpha_n$ is a sequence of ordinals, $k_1,\ldots, k_n$ and $j_1,\ldots, j_n$ are natural numbers satisfying $k_i+j_i>0$ for all $i$. Define the [natural sum]{} of $\gamma_1$ and $\gamma_2$, denoted by $\gamma_1\#\gamma_2$, of $\gamma_1$ and $\gamma_2$ as $$\gamma_1\#\gamma_2=\omega^{\alpha_1}\cdot(k_1+j_1)+\ldots+\omega^{\alpha_n}\cdot(k_n+j_n).$$ The natural sum is commutative and associative. We also introduce the natural sum for countable sequences of ordinals. Let $\{\gamma_i\|i\in \omega\}$ be a sequence of ordinals. Let $S_n$ denote the natural sum of the first $n$ items in the sequence: $$S_n=\gamma_0\#\ldots\#\gamma_{n-1}.$$ Define the [infinite natural sum]{}, denoted by $\S \gamma_i$, of the sequence $\{\gamma_i\|i\in \omega\}$ as $$\S\gamma_i=\sup_{n\in\omega} S_n.$$ We are now ready to define the notion of size for $\LL$-formulas. Again assume all formulas to be in negation normal form. \[d8\] The [size]{}, denoted by $\mathrm{s}(\phi)$, of a formula $\phi$ in $\LL$ is defined inductively as follows: $$\begin{aligned} \mathrm{s}(\phi)&=1 \textrm{ if } \phi \textrm{ is an atomic or negated atomic formula},\\ \mathrm{s}(\phi \wedge \psi)&= \mathrm{s}(\phi)\#\mathrm{s}(\psi),\\ \mathrm{s}(\phi\vee \psi)&=\mathrm{s}(\phi)\#\mathrm{s}(\psi),\\ \mathrm{s}(\exists x \phi)&=\mathrm{s}(\phi)+1,\\ \mathrm{s}(\forall x \phi)&= \mathrm{s}(\phi)+1,\\ \mathrm{s}(\bigwedge_{i\in \omega} \phi_i)&=\S \mathrm{s}(\phi_i),\\ \mathrm{s}(\bigvee_{i\in \omega} \phi_i)&=\S \mathrm{s}(\phi_i). \end{aligned}$$ If $I$ is any countable set and for each $i\in I$, $\phi_i$ is a formula, we define $\mathrm{s}(\bigwedge_{i\in I} \phi_i)$ to be $\mathrm{s}(\bigwedge_{i\in\omega}\phi_{p(i)})$, where $p$ is a bijection from $\omega$ to $I$, and similarly for disjunction. In the next proposition we show that this definition does not depend on the choice of $p$. The size of an $\LL$-formula is always a countable ordinal. The following are straightforward properties of size. \[p5\] 1. If $\phi$ is a first order formula, then Definition \[d5\] and Definition \[d8\] assign it the same size. 2. If $\phi$ is a proper subformula of $\psi$, then $\mathrm{s}(\phi)<\mathrm{s}(\psi)$. 3. The size of a conjunction is invariant under permutations of the conjuncts. More precisely, if $\phi$ and $\psi$ are formulas then $\mathrm{s}(\phi\wedge\psi)=\mathrm{s}(\psi\wedge\phi)$. If $\phi_i, i\in \omega$ is a sequence of formulas and $p$ is a permutation of $\omega$, then $\mathrm{s}(\bigwedge_{i\in\omega}\phi_i)=\mathrm{s}(\bigwedge_{i\in\omega}\phi_{p(i)})$. The same applies to disjunctions. 1\. When restricted to finite ordinals, the natural sum coincides with the usual ordinal sum. 2\. Every formula has a non-zero size. If $\gamma_1$ and $\gamma_2$ are non-zero ordinals, we always have $\gamma_1<\gamma_1\#\gamma_2$ and $\gamma_2<\gamma_1\#\gamma_2$. Hence the size of a conjunction is always greater than both of the conjuncts. The same applies to disjunctions. The quantifier case is obvious. 3\. The natural sum of ordinals is commutative. For the infinitary case, it suffices to observe that every finite sum $S_n$ of the sequence $\{\mathrm{s}(\phi_i)\| i\in\omega\}$ is subsumed by a finite sum of the sequence $\{\mathrm{s}(\phi_{p(i)})\| i\in\omega\}$. Now we move on to the other half of our plan, namely adjusting the rules of the game EFB$_n$ for $\LL$. We add, in addition to the four types of moves in EFB$_n$, two new types of moves: the infinite left splitting move and the infinite right splitting move. These two types of moves are supposed to model infinite conjunctions and disjunctions. Let $L$ be a relational vocabulary and $\mA_0$, $\mB_0$ classes of $L$-structures. Let $\alpha$ be a countable ordinal. The game $_\alpha(\mA_0, \mB_0)$ has two players. The number $\alpha$ is called the [rank]{} of the game. The positions in the game are of the form $(\mA, \mB, \gamma)$ where $\mA, \mB$ are classes of $L$-structures and $\gamma$ is an ordinal. The game begins from position $(\mA_0, \mB_0, \alpha)$. Suppose the game is in position $(\mA, \mB, \gamma)$. There are the following possibilities for the next move in the game. Finite left splitting move: : Player [I]{} first represents $\mA$ as a union $\mA_1\cup \mA_2$. He also chooses non-zero ordinals $\gamma_1$ and $\gamma_2$ such that $\gamma_1\#\gamma_2\leq \gamma$. Now the game continues from the position $(\mA_1, \mB, \gamma_1)$ or from the position $(\mA_2, \mB, \gamma_2)$, and player [II]{} can choose which.** Finite right splitting move: : Player [I]{} first represents $\mB$ as a union $\mB_1\cup \mB_2$. He also chooses non-zero ordinals $\gamma_1$ and $\gamma_2$ such that $\gamma_1\#\gamma_2\leq \gamma$. Now the game continues from the position $(\mA, \mB_1, \gamma_1)$ or from the position $(\mA, \mB_2, \gamma_2)$, and player [II]{} can choose which. Infinite left splitting move: : Player [I]{} first represents $\mA$ as a countable union $\bigcup_{i\in\omega} \mA_i$. He also chooses a sequence of non-zero ordinals $\{\gamma_i\|i\in \omega\}$ such that $\S\gamma_i\leq \gamma$. Now the game continues from the position $(\mA_i, \mB, \gamma_i)$ for some $i\in \omega$, and player [II]{} can choose which. Infinite right splitting move: : Player [I]{} first represents $\mB$ as a countable union $\bigcup_{i\in\omega} \mB_i$. He also chooses a sequence of non-zero ordinals $\{\gamma_i\|i\in \omega\}$ such that $\S\gamma_i\leq \gamma$. Now the game continues from the position $(\mA, \mB_i, \gamma_i)$ for some $i\in \omega$, and player [II]{} can choose which. Left supplementing move: : Player [I]{} picks an element from each structure $(A, \alpha)\in \mA$. More precisely, [I]{} chooses a natural number $j$ and a choice function $F$ for $\mA$. He also chooses an ordinal $\delta<\gamma$. Then the game continues from the position $(\mA(F/j), \mB(\star/j), \delta)$. Right supplementing move: : Player [I]{} picks an element from each structure $(B, \beta)\in \mB$. More precisely, [I]{} chooses a natural number $j$ and a choice function $F$ for $\mB_m$. He also chooses an ordinal $\delta<\gamma$. Then the game continues from the position $(\mA(\star/j), \mB(F/j), \delta)$. The game ends in a position $(\mA, \mB, \gamma)$ and player [I]{} wins if there is an atomic or a negated atomic formula $\phi$ such that $(\mA, \mB)\models \phi$. Player [II]{} wins the game if they reach a position $(\mA, \mB, \gamma)$ such that $\gamma=1$ and [I]{} does not win in this position.**** The game EFB$_\alpha$ characterizes separation in $\LL$ up to size $\alpha$. The following theorem is the central result in this paper. \[t7\] Let $L$ be a relational vocabulary,[^6] $\mA$ and $\mB$ classes of $L$-structures, and let $\alpha$ be a countable ordinal. Then the following are equivalent: 1. Player [I]{} has a winning strategy in the game $_\alpha(\mA, \mB)$. 2. There is an $L$-formula $\phi$ in $\LL$ of size $\leq \alpha$ such that $(\mA, \mB)\models \phi$. We prove this theorem by induction on the rank $\alpha$. When $\alpha=1$ the proposition is obvious. Suppose the proposition is true for all ordinals $\alpha<\gamma$. Now consider the case $\alpha=\gamma$. (1$\Longrightarrow$ 2) Suppose [I]{} has a winning strategy in the game EFB$_\gamma(\mA, \mB)$. We look at the first move in this strategy. Depending on which type of move it is, there are the following possibilities. [Case 1.]{} Player [I]{} first makes a finite left splitting move. He represents $\mA$ as a union $\mA_1\cup \mA_2$. He chooses non-zero ordinals $\gamma_1$ and $\gamma_2$ such that $\gamma_1\#\gamma_2\leq \gamma$. Now since this is a winning strategy, both $(\mA_1, \mB, \gamma_1)$ and $(\mA_2, \mB, \gamma_2)$ are winning positions for [I]{}. Note that both $\gamma_1$ and $\gamma_2$ are strictly smaller than $\gamma$. By the induction hypothesis, there are $L$-formulas $\psi$ and $\theta$ in $\LL$ such that $\mathrm{s}(\psi)\leq \gamma_1$, $\mathrm{s}(\theta)\leq \gamma_2$, $(\mA_1, \mB)\models \psi$ and $(\mA_2, \mB)\models \theta$. Let $\phi$ be the formula $\psi\vee \theta$. Then we have that $\mathrm{s}(\phi)=\mathrm{s}(\psi)\#\mathrm{s}(\theta)\leq\gamma_1\#\gamma_2\leq \gamma$, and that $(\mA, \mB)\models \phi$.**** [Case 2.]{} Player [I]{} first makes a finite right splitting move. He represents $\mB$ as a union $\mB_1\cup \mB_2$. He chooses non-zero ordinals $\gamma_1$ and $\gamma_2$ such that $\gamma_1\#\gamma_2\leq\gamma$. Now since this is a winning strategy, both $(\mA, \mB_1, \gamma_1)$ and $(\mA, \mB_2, \gamma_2)$ are winning positions for [I]{}. Note that both $\gamma_1$ and $\gamma_2$ are strictly smaller than $\gamma$. By the induction hypothesis, there are formulas $\psi$ and $\theta$ such that $\mathrm{s}(\psi)\leq \gamma_1$, $\mathrm{s}(\theta)\leq \gamma_2$, $(\mA, \mB_1)\models \psi$ and $(\mA, \mB_2)\models \theta$. Let $\phi$ be the formula $\psi\wedge \theta$. Then we have that $\mathrm{s}(\phi)= \mathrm{s}(\psi)\#\mathrm{s}(\theta)\leq \gamma_1\#\gamma_2\leq\gamma$, and that $(\mA, \mB)\models \phi$.**** [Case 3.]{} Player [I]{} first makes an infinite left splitting move. He represents $\mA$ as a union $\bigcup_{i\in\omega} \mA_i$. He chooses ordinals $\{\gamma_i\|i\in \omega\}$ such that $\S\gamma_i\leq \gamma$. Note that every $\gamma_i$ is strictly smaller than $\gamma$. Now since this is a winning strategy, for every $i\in \omega$ $(\mA_i, \mB, \gamma_i)$ is a winning position for him. By the induction hypothesis, there are $L$-formulas $\phi_i$ in $\LL$ such that $\mathrm{s}(\phi_i)\leq \gamma_i$ and $(\mA_i, \mB)\models \phi_i$ for all $i\in \omega$. Let $\phi$ be the formula $\bigvee_{i\in \omega} \phi_i$. It is an $\LL$-formula. We have that $\mathrm{s}(\phi)=\S \mathrm{s}(\phi_i)\leq\S \gamma_i\leq \gamma$, and that $(\mA, \mB)\models \phi$. [Case 4.]{} Player [I]{} first makes an infinite right splitting move. He represents $\mB$ as a union $\bigcup_{i\in\omega} \mB_i$. He chooses ordinals $\{\gamma_i\|i\in \omega\}$ such that $\S\gamma_i\leq \gamma$. Note that every $\gamma_i$ is strictly smaller than $\gamma$. Now since this is a winning strategy, for every $i\in \omega$ $(\mA, \mB_i, \gamma_i)$ is a winning position for him. By the induction hypothesis, there are $L$-formulas $\phi_i$ in $\LL$ such that $\mathrm{s}(\phi_i)\leq \gamma_i$ and $(\mA, \mB_i)\models \phi_i$ for all $i\in \omega$. Let $\phi$ be the formula $\bigwedge_{i\in \omega} \phi_i$. It is an $\LL$-formula. We have that $\mathrm{s}(\phi)=\S \mathrm{s}(\phi_i)\leq\S \gamma_i\leq \gamma$, and that $(\mA, \mB)\models \phi$. [Case 5.]{} Player [I]{} first makes a left supplementing move. He chooses a natural number $j$ and a choice function $F$ for $\mA$. He also chooses an ordinal $\delta<\gamma$. The game continues from $(\mA(F/j), \mB(\star/j), \delta)$. Since this is a winning strategy for [I]{}, the position $(\mA(F/j), \mB(\star/j), \delta)$ is a winning position for him too. By the induction hypothesis, there is a formula $\psi$ such that $\mathrm{s}(\psi)\leq \delta$, and $(\mA(F/j), \mB(\star/j))\models \psi$. Let $\phi$ be the formula $\exists x_j\psi$. Then we have that $\mathrm{s}(\phi)=\mathrm{s}(\psi)+1\leq \delta+1\leq \gamma$, and that $(\mA, \mB)\models \phi$.**** [Case 6.]{} Player [I]{} first makes a right supplementing move. He chooses a natural number $j$ and a choice function $F$ for $\mB$. He also chooses an ordinal $\delta<\gamma$. The game continues from $(\mA(\star/j), \mB(F/j), \delta)$. Since this is a winning strategy for [I]{}, the position $(\mA(\star/j), \mB(F/j), \delta)$ is a winning position for him too. By the induction hypothesis, there is a formula $\psi$ such that $\mathrm{s}(\psi)\leq \delta$, and $(\mA(\star/j), \mB(F/j))\models \psi$. Let $\phi$ be the formula $\forall x_j\psi$. Then we have that $\mathrm{s}(\phi)=\mathrm{s}(\psi)+1\leq \delta+1\leq \gamma$, and that $(\mA, \mB)\models \phi$.**** Now for the converse direction (2 $\Longrightarrow$ 1). Suppose there is an $\LL$-formula $\phi$ of size $\leq \gamma$ such that $(\mA, \mB)\models \phi$. Depending on the shape of $\phi$, there are the following possibilities. [Case 1.]{} The formulas $\phi$ is an atomic formula. By definition player [I]{} wins the game EFB$_1(\mA, \mB)$. [Case 2.]{} The formula $\phi$ is $\psi\vee\theta$. Let $\mA_1$ be the class of structures $(A, \alpha)\in \mA$ such that $(A, \alpha)\models \psi$, $\mA_2$ be the class of structures $(A, \alpha)\in \mA$ such that $(A, \alpha)\models \theta$. Since $(A, \alpha)\models \phi$ for all $(A, \alpha)\in \mA$, we have $\mA=\mA_1\cup \mA_2$. Moreover since $(B, \beta)\not\models \phi$ for all $(B, \beta)\in \mB$, we have $(\mA_1, \mB)\models \psi$ and $(\mA_2, \mB)\models \theta$. Finally, as $\mathrm{s}(\phi)\leq \gamma$, there are ordinals $\gamma_1$ and $\gamma_2$ such that $\mathrm{s}(\psi)\leq \gamma_1$, $\mathrm{s}(\theta)\leq \gamma_2$, and $\gamma_1\#\gamma_2\leq \gamma$. By the induction hypothesis, [I]{} has a winning strategy in both $(\mA_1, \mB, \gamma_1)$ and $(\mA_2, \mB, \gamma_2)$. Therefore he has a winning strategy in $(\mA, \mB, \gamma)$ by first making a finite left splitting move, and then follow the winning strategy in $(\mA_1, \mB, \gamma_1)$ or $(\mA_2, \mB, \gamma_2)$. [Case 3.]{} The formula $\phi$ is $\psi\wedge\theta$. This case is dual to case 2 and we omit the details. [Case 4.]{} The formula $\phi$ is $\bigvee_{i\in\omega}\phi_i$. For each $i$, let $\mA_i$ be the class of structures $(A, \alpha)\in \mA$ such that $(A, \alpha)\models \phi_i$. Since $(A, \alpha)\models \phi$ for all $(A, \alpha)\in \mA$, we have $\mA=\bigcup_{i\in\omega}\mA_i$. Moreover since $(B, \beta)\not\models\phi$ for all $(B, \beta)\in \mB$, we have $(\mA_i, \mB)\models \phi_i$ for every $i$. Finally, as $\mathrm{s}(\phi)\leq \gamma$, there are ordinals $\{\gamma_i\|i\in \omega\}$ such that $\mathrm{s}(\phi_i)\leq \gamma_i$ and $\S\gamma_i\leq \gamma$. By the induction hypothesis, [I]{} has a winning strategy in $(\mA_i, \mB, \gamma_i)$ for every $i$. Therefore he has a winning strategy in $(\mA, \mB, \gamma)$ by first making an infinite left splitting move, and then follow the winning strategy in some $(\mA_i, \mB, \gamma_i)$. [Case 5.]{} The formula $\phi$ is $\bigwedge_{i\in\omega}\phi_i$. This case is dual to case 4 and we omit the details. [Case 6.]{} The formula $\phi$ is $\exists x_j \psi$. Since $(A, \alpha)\models \phi$ for all $(A, \alpha)\in \mA$, there is a choice function $F$ for $\mA$ such that $(A, \alpha(F(A, \alpha)/j))\models \psi$. Thus $(A, \alpha^)\models \psi$ for all $(A, \alpha^)\in \mA(F/j)$. On the other hand, we have that $(\mB, \beta^)\not\models \psi$ for all $(\mB, \beta^)\in \mB(\star/j)$. Therefore $(\mA(F/j), \mB(\star/j))\models \psi$. By the induction hypothesis, player [I]{} has a winning strategy in $(\mA(F/j), \mB(\star/j), \mathrm{s}(\psi))$. Note that $\mathrm{s}(\phi)=\mathrm{s}(\psi)+1$. Hence [I]{} has a winning strategy in $(\mA, \mB, \gamma)$ by first making a left supplementing move, chooses the ordinal $\mathrm{s}(\psi)<\mathrm{s}(\phi)\leq\gamma$ and then follow the winning strategy in $(\mA(F/j), \mB(\star/j), \mathrm{s}(\psi))$.**** [Case 7.]{} The formula $\phi$ is $\forall x_j \psi$. This case is dual to case 6 and we omit the details. \[c8\] Given two classes of $L$-structures $\mA$ and $\mA'$, let $\mA\subseteq \mA'$ denote that every structure in $\mA$ is contained in $\mA'$. Let $\mA, \mA', \mB, \mB'$ be classes of $L$-structures and $\alpha$ be a countable ordinal. Suppose $\mA'\subseteq\mA$ and $\mB'\subseteq\mB$. If player [I]{} has a winning strategy in $_\alpha(\mA, \mB)$, then he also has a winning strategy in $_\alpha(\mA', \mB')$. Suppose [I]{} has a winning strategy in EFB$_\alpha(\mA, \mB)$. By Theorem \[t7\], there is a formula $\phi$ with $\mathrm{s}(\phi)\leq \alpha$ such that $(\mA, \mB)\models \phi$. Since $\mA'\subseteq \mA$ and $\mB'\subseteq \mB$, we also have that $(\mA', \mB')\models \phi$. By Theorem \[t7\] again, [I]{} has a winning strategy in EFB$_\alpha(\mA', \mB')$. The six types of moves in the game EFB$_\alpha$ model the six formula-forming operations in $\LL$ respectively. The finite left splitting move corresponds to finite disjunction, the finite right splitting move to finite conjunction, the infinite left splitting move to infinite disjunction, the infinite right splitting move to infinite conjunction, the left supplementing move to existential quantification, and the right supplementing move to universal quantification. This feature enables us to consider various fragments of the game EFB$_\alpha$ which characterize separation in different fragments of $\LL$. If we forbid infinite splitting moves, what we get is essentially the old game EFB$_n$. If we only allow splitting moves in the game, the result is a game for the propositional fragment of $\LL$. \[c9\] Let $^P_\alpha(\mA, \mB)$ be the restriction of the game $_\alpha(\mA, \mB)$ where player [I]{} can only make finite and infinite splitting moves, but no supplementing moves. Then the following are equivalent. 1. Player [I]{} has a winning strategy in the game $^P_\alpha(\mA, \mB)$.** 2. There is a quantifier-free $\LL$-formula $\phi$ of size $\leq \alpha$ such that $(\mA, \mB)\models \phi$. Alternative Conceptions of Size =============================== Let us take some time to reflect on our result. We have shown that measuring the complexity of formulas by size is [a]{} way to define an game for $\LL$. However, our viewpoint naturally suggests the possibility of a plurality of complexity measures, and hence a plurality of games for $\LL$. Can we find other complexity measures which give rise to an game for $\LL$? Furthermore, to which extent are these complexity measures [reasonable]{}?** We shall start by looking at the bare minimum, namely which complexity measures could give rise to an game. It turns out that if we merely want an adequacy theorem to hold, we can be quite relaxed about the choice of complexity measures. Let $q:\mathrm{Ord}\rightarrow \mathrm{Ord}$, $r:\mathrm{Ord}^2\rightarrow \mathrm{Ord}$ and $t:\mathrm{Ord}^\omega\rightarrow \mathrm{Ord}$ be ordinal functions. Consider an abstract complexity measure $c$ defined by $q$, $r$ and $t$: $$\label{e1} \begin{aligned} c(\phi)&=1 \textrm{ if } \phi \textrm{ is an atomic or negated atomic formula},\\ c(\phi \wedge \psi)&= r((c(\phi),c(\psi)),\\ c(\phi\vee \psi)&=r((c(\phi),c(\psi)),\\ c(\exists x \phi)&=q(c(\phi)),\\ c(\forall x \phi)&= q(c(\phi)),\\ c(\bigwedge_{i\in \omega} \phi_i)&=t( \{c(\phi_i)\|i\in\omega\}),\\ c(\bigvee_{i\in \omega} \phi_i)&=t(\{c(\phi_i)\|i\in\omega\}). \end{aligned}$$ We say that $c$ is [nice]{} if $q$, $r$ and $t$ satisfy the following conditions: 1. If $\gamma>0$, then $q(\gamma)>\gamma$. 2. If $\gamma_1,\gamma_2>0$, then $r(\gamma_1, \gamma_2)>\gamma_1$ and $r(\gamma_1, \gamma_2)>\gamma_2$. 3. If $\{\gamma_i\|i\in\omega\}$ is a sequence of non-zero ordinals, then $t(\{\gamma_i\|i\in\omega\})>\gamma_i$ for all $i\in\omega$. Niceness guarantees that the complexity of a formula is always strictly larger than the complexity of its proper subformulas. Granted this, we can generalize the game EFB$_\alpha$ to any nice complexity measure by performing minor cosmetic surgery. Let EFB$^c_\alpha$ denote the game for $\LL$ under the complexity measure $c$. The starting position and the winning condition in the game EFB$^c_\alpha$ remain the same as in EFB$_\alpha$. The way that the rank of the game is counted is modified. We only give the ‘left’ moves in the game EFB$^c_\alpha$ here; the ‘right’ moves are dual. Finite left splitting move: : Player [I]{} first represents $\mA$ as a union $\mA_1\cup \mA_2$. He also chooses non-zero ordinals $\gamma_1$ and $\gamma_2$ such that $r(\gamma_1,\gamma_2)\leq \gamma$. Now the game continues from the position $(\mA_1, \mB, \gamma_1)$ or from the position $(\mA_2, \mB, \gamma_2)$, and player [II]{} can choose which.** Infinite left splitting move: : Player [I]{} first represents $\mA$ as a countable union $\bigcup_{i\in\omega} \mA_i$. He also chooses a sequence of non-zero ordinals $\{\gamma_i\|i\in \omega\}$ such that $t(\{\gamma_i\|i\in\omega\})\leq \gamma$. Now the game continues from the position $(\mA_i, \mB, \gamma_i)$ for some $i\in \omega$, and player [II]{} can choose which. Left supplementing move: : Player [I]{} picks an element from each structure $(A, \alpha)\in \mA$. More precisely, [I]{} chooses a natural number $j$ and a choice function $F$ for $\mA$. He also chooses an ordinal $\delta$ such that $q(\delta)\leq\gamma$. Then the game continues from the position $(\mA(F/j), \mB(\star/j), \delta)$. It is an easy observation that the proof of the adequacy theorem goes through if we substitute size for any nice complexity measure. We have the following generalized adequacy theorem. Let $c$ be a nice complexity measure. Let $L$ be a relational vocabulary, $\mA$ and $\mB$ classes of $L$-structures, and let $\alpha$ be an ordinal. Then the following are equivalent: 1. Player [I]{} has a winning strategy in the game $^c_\alpha(\mA, \mB)$.** 2. There is an $L$-formula $\phi$ in $\LL$ with $c(\phi)\leq\alpha$ such that $(\mA, \mB)\models \phi$. In particular, size is a nice complexity measure: take $q$ to be the successor function, $r$ the natural sum, and $t$ the infinite natural sum. Another interesting nice complexity measure $c_1$ can be obtained by taking $q$ to be the successor function, $r(\gamma_1, \gamma_2)=\sup(\gamma_1+1, \gamma_2+1)$ and $t(\{\gamma_i\|i\in\omega\})=\sup_{i\in\omega}(\gamma_i+1)$. Note that if we take $q$ to be the successor function and $r, t$ to be the supremum, then the complexity measure is just quantifier rank—but this is not a nice complexity measure. To determine which of the nice complexity measures are the more natural and reasonable ones is a subtler question. As a general rule of thumb, we would like a complexity measure to carry some meaning, so that we have an idea of what exactly it is measuring. We also want a complexity measure to be invariant under certain logical equivalences, so that it is compatible with our intuition about ‘complexity’. Between size and $c_1$, size has the advantage of having a clear intuition behind it. This is especially so in the propositional case: the size of a propositional formula is just the number of propositional symbols in this formula. The quantifier and infinitary cases are natural extensions of this basic intuition. For $c_1$ the intuition is not so clear. As for invariance under logical equivalences, both size and $c_1$ are invariant under permutations of conjuncts and disjuncts. They part ways, however, when we try to count the complexity of repetitions of formulas. Consider atomic formulas $\phi_i, i\in\omega$. Let $\Theta$ be the infinite conjunction $\bigwedge_{i\in\omega}\phi_0$. Let $\Psi$ be the infinite conjunction $\bigwedge_{i\in\omega}\phi_i$. Size assigns $\Theta$ and $\Psi$ the complexity $\omega$. The measure $c_1$ assigns $\Theta$ and $\Psi$ the complexity $2$. It seems that $c_1$ gives the more reasonable complexity for $\Theta$, since $\Theta$ is no more than a repetition of the same formula $\phi_0$. On the other hand size seems to give the more reasonable complexity for $\Psi$, while $c_1$ does not reflect the fact that $\Psi$ contains much more information than any finite conjunction. However as long as complexity measures are defined purely syntactically, every complexity measure would assign the same complexity to $\Theta$ and $\Psi$, so there is no complexity measure that assigns the ‘right’ complexity to $\Theta$ and $\Psi$ at the same time. This suggests that there is no point in trying to find [the]{} reasonable complexity measure. We may indeed adopt a pluralistic view: we shall choose which is the appropriate complexity measure depending on what information we intend to model. An Application: the Complexity of Infinite Binary String Properties =================================================================== In this section we apply the newly defined game EFB$_\alpha$ to solve one instance of Question 1. We focus on information code by infinite binary strings. An infinitary binary string $t$ is a function from $\omega$ to $2$. We can think of a string $t$ simply as a sequence of $0$’s and $1$’s. A property $\P$ of infinite binary strings is a subset $\P\subset 2^\omega$. Infinite binary strings can be studied using the propositional fragment of $\LL$. Let $L$ be the propositional language with countably many propositional symbols $p_i$, $i\in \omega$. For an infinitary binary string $t$ and a propositional symbol $p_i$, we say that $$t\models p_i\ \textrm{if}\ t(i)=1,$$ and $$t\models \neg p_i\ \textrm{if} \ t(i)=0.$$ Intuitively, $p_i$ has the intended meaning that ‘the i’th position of $t$ is $1$’, and $\neg p_i$ has the intended meaning that ‘the i’th position of $t$ is $0$’. We pose the question: Question 2: : Given a string property $\P$, what is the minimal size of a propositional $\LL$-formula defining $\P$? Note that an infinite binary string $t$ codes a real number. Hence this is essentially a question about the definability of sets of real numbers. There is a straightforward way to define a string property $\P$. An infinite binary string $t$ can be characterized by the following sentence in $\LL$: $$\theta_t=\bigwedge_{i\in\omega}p_i^t,$$ where $$p_i^t=\left\{ \begin{array}{ll} p_i & \textrm{if $t(i)=1$,}\\ \neg p_i & \textrm{if $t(i)=0$.}\end{array}\right.$$ A property $\P$ of infinite binary strings can be defined by the following sentence: $$\Phi_\P=\bigvee_{t\in\P} \theta_t.$$ In general $\Phi_\P$ is an $L_{\infty\omega}$-sentence. When the cardinality of $\P$ is countable, this sentence is in $\LL$. In this case the sentence $\Phi_\P$ has size $$\mathrm{s}(\Phi_\P)=\sum_{t\in\P}^\#\mathrm{s}(\theta_t)=\sum_{i\in\omega}^\#\omega=\omega^2.$$ It is easy to see that if the complement of $\P$ in $2^\omega$ is countable, $\P$ can also be defined by an $\LL$-sentence of size $\omega^2$. The question is whether the sentence $\Phi_\P$ is minimal. Regarding each infinite binary string as a ‘model’, we can play the game $^P_\alpha$ on classes of infinite binary strings. This game solves Question 2 for an interesting class of string properties. A property $\P$ of infinite binary strings is said to be [dense]{} if it is dense in the space[^7] $2^\omega$ with the product topology. \[t10\] Let $\P_1$ and $\P_2$ be dense properties of infinite binary strings. Player [II]{} has a winning strategy in $_\alpha^P(\P_1, \P_2)$ for all $\alpha<\omega^2$. Together with Corollary \[c9\], this leads to the following theorem. If $\P_1$ and $\P_2$ are dense properties of infinitary binary strings and $\phi$ is a propositional $\LL$-formula separating $\P_1$ and $\P_2$, then the size of $\phi$ is at least $\omega^2$. \[c11\] If $\P$ is a string property such that both $\P$ and its complement are dense, then $\P$ cannot be defined by a propositional $\LL$-formula of size less than $\omega^2$. If a string property $\P$ is countable, then by Cantor diagonal argument its complement is dense. For a countable dense property $\P$ the standard formula $\Phi_\P$ is a minimal formula defining $\P$. In the remaining of this section we give a proof of Theorem \[t10\]. Let us first fix some notation. Let $2^{<\omega}$ denote the set of partial functions $\omega \rightarrow 2$ with a finite domain. We call a function $g\in 2^{<\omega}$ a [finite segment]{}. For an infinite binary string $h\in 2^\omega$ and a finite segment $g\in 2^{<\omega}$, we say that $h$ [agrees with]{} $g$, or that $h$ [extends]{} $g$, if $g\subset h$. A string property $\P$ is dense if and only if for any finite segment $g\in 2^{<\omega}$, there exists $h\in\P$ that extends $g$. In the game EFB$^P_\alpha$ the players are not allowed to play supplementing moves, but can only play splitting moves. When player [I]{} plays an infinite splitting move, he splits a class of structures $\mA$ (or $\mB$) into countably many pieces $\mA_i, i\in\omega$ (or $\mB_i, i\in \omega$). We say that an infinite splitting move is a [proper infinite splitting move]{} if there are infinitely many pieces $\mA_i$ (or $\mB_i$) that are non-empty. If player [I]{} splits $\mA$ (or $\mB$) into infinitely many pieces among which only finitely many are non-empty, we call it a [degenerate infinite splitting move]{}. Strange as it may seem, there is nothing in the rules that forbids *[I]{} playing in this way. A degenerate infinite splitting move is closer in nature to a finite series of finite splitting moves. Henceforth we refer to degenerate infinite splitting move and finite series of finite splitting moves as [generalized finite splitting moves]{}.******* Let $\alpha$ be an ordinal smaller than $\omega^2$. We describe the following strategy for player [II]{} in the game $_\alpha^P(\P_1, \P_2)$. [Player II’s strategy]{}: Let $(\mA, \mB, \gamma)$ denote a game position. In each round, player [II]{} makes sure that $$\label{e9} \begin{split} \exists f\in 2^{<\omega} \Bigg(&\forall h\in \mA (h\supset f) \wedge \\ &\forall h \in \mB (h\supset f)\wedge \\ &\forall g\in 2^{<\omega}\bigg( g\supset f\rightarrow \big(\exists h\in \mA (h\supset g) \wedge \exists h' \in \mB (h'\supset g)\big)\bigg)\Bigg). \end{split}$$ She maintains this condition until she sees an opportunity to finish off the game directly. Condition (\[e9\]) says that at any stage of the game, we can always find a finite segment $f$ such that all strings in $\mA$ and $\mB$ agree with $f$. Moreover, for any finite extension $g$ of $f$, there are string $h$ and $h'$ in $\mA$ and $\mB$ respectively that agree with $g$. Therefore if we imagine we ‘cut off’ the finite segment $f$, the remaining configuration in the game is almost identical to the starting position. We will specify later what counts as an opportunity for [II]{} to win the game directly. Theorem \[t10\] is established through a series of lemmas. Condition (\[e9\]) is true in the starting position $(\P_1, \P_2, \alpha).$ Let $f$ be the empty function. Then condition (\[e9\]) is just the density condition. \[l14\] If condition (\[e9\]) is true in position $(\mA, \mB, \gamma)$, then no atomic formula or its negation separates $\mA$ and $\mB$. There is an atomic formula or its negation separating $\mA$ and $\mB$ if and only if there exists $g\in 2^{<\omega}$ such that $\|\textrm{dom}(g)\|=1$, $g\subset h$ for all $h\in \mA$, and $g\not\subset h'$ for all $h'\in \mB$. We claim that condition (\[e9\]) guarantees that this situation does not occur. There are the following three possibilities for $g$, let us consider them one by one. Firstly, if $g\subset f$, then by condition (\[e9\]) all strings in $\mA$ and $\mB$ agree with $g$. Secondly, if $g$ disagrees with $f$, namely if dom($g$)$\subset$dom($f$) yet $g\not\subset f$, then by condition (\[e9\]) again all strings in $\mA$ and $\mB$ disagree with $g$. Finally, suppose $\textrm{dom}(g)\cap \textrm{dom}(f)=\emptyset$. The finite segment $g\cup f$ is an extension of $f$. Therefore by condition (\[e9\]) there are $h\in \mA$ and $h'\in \mB$ that agree with $g\cup f$ respectively. In particular they agree with $g$. \[l15\] If player [I]{} makes a generalized finite splitting move, player [II]{} can maintain condition (\[e9\]) to the next round. Suppose the game is in position $(\mA, \mB, \gamma)$ and condition (\[e9\]) holds. Suppose player [I]{} makes a generalized finite left splitting move. He represents $\mA$ as a union $\mA_1\cup\ldots\cup \mA_n$ of $n$ pieces. He also picks ordinals $\gamma_1,\ldots, \gamma_n$ such that $\gamma_1\#\ldots\#\gamma_n\leq\gamma$. We claim that there must exist a finite segment $f'\supset f$ and a piece $\mA_k$ such that $$\label{e10} \forall g\in 2^{<\omega}\big( g\supset f'\rightarrow \exists h\in \mA_k (h\supset g)\big).$$ Suppose this is not the case for $\mA_1,\ldots, \mA_{n-1}$. Then we have $$\label{e11} \forall f'\supset f \exists g\in 2^{<\omega}\bigg(g\supset f' \wedge \forall h\in 2^\omega (h\supset g\rightarrow h\notin \mA_i) \bigg)$$ for $i=1,\ldots, n-1$. Now let $f'=f$, apply condition (\[e11\]) to $\mA_1$: $$\exists g_1\in 2^{<\omega}\bigg(g\supset f \wedge \forall h\in 2^\omega\big( h\supset g_1\rightarrow h\notin \mA_1\big)\bigg).$$ This means that $$\exists g_1\in 2^{<\omega}\bigg(g\supset f \wedge \forall h\in 2^\omega \big( (h\supset g_1\wedge h\in \mA )\rightarrow h\in \mA_2\cup\ldots\cup \mA_n\big)\bigg).$$ Now let $f'=g_1$, apply condition (\[e11\]) to $\mA_2$: $$\exists g_2\in 2^{<\omega}\bigg(g_2\supset g_1\wedge \forall h\in 2^\omega \bigg( (h\supset g_2 \wedge h\in \mA)\rightarrow h\in \mA_3\cup\ldots\cup \mA_n\big)\bigg).$$ Iterate this process, we get: $$\exists g_{n-1}\in 2^{<\omega}\bigg(g_{n-1}\supset g_{n-2} \forall h\in 2^\omega\bigg( (h\supset g_{n-1}\wedge h\in \mA)\rightarrow h\in \mA_n\big)\bigg).$$ This means that condition (\[e10\]) is true for $\mA_n$ when we take $g_{n-1}$ for $f'$. Therefore the previous claim is true. Now let player [II]{} pick this piece $\mA_k$ and the corresponding ordinal $\gamma_k$. What remains to be done is a cosmetic surgery on $\mA_k$ and $\mB$. We throw away the strings that do not agree with $f'$. Let $$\mA'=\{h\in\mA_k\|h\supset f'\}$$ and let $$\mB'=\{h'\in\mB\|h'\supset f'\}.$$ By Corollary \[c8\], if player [II]{} has a winning strategy in this new position, she also has one in the old position. It is straightforward to check that condition (\[e9\]) holds for $\mA'$ and $\mB'$. Let the game continue from the position $(\mA', \mB', \gamma_k)$. The case where [I]{} plays a generalized finite right splitting move is similar. \[l16\] Suppose condition (\[e9\]) is true in position $(\mA, \mB, \gamma)$ with $\gamma<\omega^2$. If player [I]{} makes a proper infinite splitting move, then player [II]{} has a winning strategy in the rest of the game. Suppose the game is in position $(\mA, \mB, \gamma)$ and condition (\[e9\]) is true. Suppose player [I]{} makes a proper infinite splitting move. He represents $\mA$ as a union $\bigcup_{i\in \omega} \mA_i$, among which there are infinitely many non-empty pieces. He also chooses ordinals $\gamma_i, i\in \omega$ such that $\S\gamma_i\leq\gamma< \omega^2$. Note that there can only be finitely many $i\in\omega$ such that $\gamma_i$ is an infinite ordinal, for otherwise the natural sum $\S\gamma_i$ would be greater or equal to $\omega^2$, which is a contradiction. On the other hand there are infinitely many $i\in \omega$ such that $\mA_i$ is non-empty. Therefore there must exist $k\in \omega$ such that $\mA_k$ is non-empty and $\gamma_k$ is finite. If this situation occurs, we ask player [II]{} to jump out of the strategy of maintaining condition (\[e9\]) and go on to win the game directly. Let player [II]{} pick this piece $\mA_k$ and $\gamma_k$. The game continues from the position $(\mA_k, \mB, \gamma_k)$. Apply Corollary \[c8\] and suppose $\mA_k$ consists of a single string $h_A$. Note that since $\gamma_k$ is a finite number, in the rest of the game [I]{} can only make finite right splitting moves.**** Suppose [I]{} makes a finite right splitting move. He represents $\mB$ as a union $\mB_1\cup\ldots\cup\mB_n$ of $n$ pieces. Consider the witnessing finite segment $f$ for condition (\[e9\]). Let $k=\sup(\mathrm{dom}(f))$. For any natural number $m$, let $h_m$ be the string in $\mB$ such that $h_m$ agrees with $h_A$ on the first $k+m$ elements of $\omega$. By condition (\[e9\]) this string always exists. In other words, we have $$\forall m\in \omega \ \exists h_m\in \mB \big(h_m\upharpoonright (k+m)=h_A\upharpoonright (k+m)\big).$$ Among the $n$ pieces $\mB_1,\ldots, \mB_n$, there must be one piece $\mB_l$ that contains $h_m$ for infinitely many $m\in \omega$. Let player [II]{} pick this piece $\mB_l$. She uses the same strategy if [I]{} makes another finite right splitting move. She makes sure that at any stage of the remaining game, the piece $\mB_r$ at hand always contain $h_m$ for infinitely many $m\in \omega$. This guarantees that no atomic formula separates $\mA_k$ and $\mB$. Player [II]{} wins when the game terminates after finitely many rounds.**** Now we can see that the strategy prescribed above is indeed a winning strategy for player [II]{}. Player [II]{} hangs on in the game by maintaining condition (\[e9\]). If her opponent plays a proper infinite splitting move at some point, she goes on to win the rest of the game by appealing to Lemma \[l16\]. Otherwise she continues with condition (\[e9\]). This is possible because of Lemma \[l15\]. This strategy keeps her away from losing by Lemma \[l14\]. Player [II]{} will eventually prevail after finitely many rounds. Thus we have finished the proof of Theorem \[t10\].**** The following are some examples of dense properties of infinite binary strings. By Corollary \[c11\] none of these properties can be defined by a propositional $\LL$-formula of size less than $\omega^2$. 1. $\P_1=\{f\in 2^\omega\|f^{-1}(1) \textrm{ is finite}\}$ ‘$f$ has finitely many ones’. 2. $\P_2=\{f\in 2^\omega\|\|f^{-1}(1)\| \textrm{ is an odd number}\}$ ‘$f$ has an odd number of ones’. 3. $\P_3=\textrm{`$f$ codes a rational number'}$. 4. $\P_4=\textrm{`$f$ codes a decidable set'}$. This result is a first step in the application of EFB$_\alpha$. The game suggests many natural questions about the expressive power of $\LL$, for example: [Open Problems:]{}** 1. Can we use the game to prove higher complexity results for propositional $\LL$? 2. Can we find applications of the full game instead of the propositional fragment? 3. Can we prove complexity results for other nice complexity measures? These remain questions for further study. Finally, we point out that new problems arise when we move from $\LL$ to $L_{\omega_2\omega}$ and even higher infinitary languages. Suppose we want to define the notion of size for $L_{\omega_2\omega}$-formulas in such a way that it is coherent with our definition of size for $\LL$-formulas. To illustrate the problem, consider the uncountable conjunction $\Phi=\bigwedge_{i\in\omega_1}\phi_i$, where $\phi_i$ is an atomic formula for all $i\in\omega_1$. What is the right size of $\Phi$? There are the following three desiderata: 1. If $\theta$ is an $\LL$-formula, then the size of $\theta$ is the same as it is given in Definition \[d8\]. 2. The size of $\Phi$ is the supremum of the size of its proper subformulas, namely $\mathrm{s}(\Phi)=\sup_{\alpha<\omega_1}\mathrm{s}(\bigwedge_{i\in\alpha}\phi_i)$. 3. The size of $\Phi$ is at least $\omega_1$. It turns out that these three conditions cannot be satisfied simultaneously. Suppose there is a complexity measure $\mathrm{s}$ satisfying conditions 1 and 2. Consider any infinite ordinal $\alpha<\omega_1$. By condition 1, we have that $\mathrm{s}(\bigwedge_{i\in\alpha}\phi_i)=\mathrm{s}(\bigwedge_{i\in\omega}\phi_{p(i)})=\omega$, where $p$ is a bijection from $\omega$ to $\alpha$. This together with condition 2 renders the size of $\Phi$ to be $\omega$, contrary to condition 3. [^1]: Research partially supported by grant 40734 of the Academy of Finland. [^2]: Research supported by grant 4048 of the European Science Foundation project ‘New Frontiers of Infinity: Mathematical, Philosophical and Computational Prospects’. [^3]: In this paper we use ‘countable’ to mean ‘countably infinite’. [^4]: The logic $L_{\infty\omega}$ is the extension of first order logic with arbitrary conjunctions and disjunctions. [^5]: According to Barwise [@Barwise81], Charles Peirce thought of quantifiers as infinite conjunctions or disjunctions, and this was picked up by Löwenheim, Wittgenstein and others, and used in proof theory by Novikoff already in the 1940s. [^6]: Note that $L$ is not necessarily countable. [^7]: Called the Cantor space.	{ "pile_set_name": "ArXiv" }
--- abstract: 'In this paper we consider a class of optimization problems with a strongly convex objective function and the feasible set given by an intersection of a simple convex set with a set given by a number of linear equality and inequality constraints. A number of optimization problems in applications can be stated in this form, examples being the entropy-linear programming, the ridge regression, the elastic net, the regularized optimal transport, etc. We extend the Fast Gradient Method applied to the dual problem in order to make it primal-dual so that it allows not only to solve the dual problem, but also to construct nearly optimal and nearly feasible solution of the primal problem. We also prove a theorem about the convergence rate for the proposed algorithm in terms of the objective function and the linear constraints infeasibility.' author: - Alexey Chernov - Pavel Dvurechensky - Alexender Gasnikov title: 'Fast Primal-Dual Gradient Method for Strongly Convex Minimization Problems with Linear Constraints' --- Introduction {#introduction .unnumbered} ============ In this paper we consider a constrained convex optimization problem of the following form $$(P_1) \quad \quad \min_{x\in Q \subseteq E} \left\{ f(x) : A_1x =b_1, A_2x \leq b_2\right\}, \notag $$ where $E$ is a finite-dimensional real vector space, $Q $ is a simple closed convex set, $A_1$, $A_2$ are given linear operators from $E$ to some finite-dimensional real vector spaces $H_1$ and $H_2$ respectively, $b_1 \in H_1$, $b_2 \in H_2$ are given, $f(x)$ is a $\nu$-strongly convex function on $Q$ with respect to some chosen norm $\\|\cdot\\|_E$ on $E$. The last means that for any $x,y \in Q$ $f(y) \geq f(x) + {\langle}\nabla f(x) , y-x {\rangle}+ \frac{\nu}{2}\\|x-y\\|^2_E$, where $\nabla f(x)$ is any subgradient of $f(x)$ at $x$ and hence is an element of the dual space $E^$. Also we denote the value of a linear function $g \in E^$ at $x\in E$ by ${\langle}g, x {\rangle}$. Problem $(P_1)$ captures a broad set of optimization problems arising in applications. The first example is the classical entropy-linear programming (ELP) problem [@Fang_1997_Entropy] which arises in many applications such as econometrics [@Golan_1996_Maximum], modeling in science and engineering [@Kapur_1989_Maximum], especially in the modeling of traffic flows [@Gasnikov_2013_Introduction] and the IP traffic matrix estimation [@Rahman_2006_IP; @Zhang_2005_Estimating]. Other examples are the ridge regression problem [@Hastie_2009_Elements] and the elastic net approach [@Zou_2005_Regularization] which are used in machine learning. Finally, the problem class $(P_1)$ covers problems of regularized optimal transport (ROT) [@Cuturi_2013_Sinkhorn] and regularized optimal partial transport (ROPT) [@Benamou_2015_Iterative], which recently have become popular in application to the image analysis. The classical balancing algorithms such as [@Cuturi_2013_Sinkhorn; @Bregman_1967_Relaxation; @Bregman_1967_Proof] are very efficient for solving ROT problems or special types of ELP problem, but they can deal only with linear equality constraints of special type and their rate of convergence estimates are rather impractical [@Franklin_1989_Scaling]. In [@Benamou_2015_Iterative] the authors provide a generalization but only for the ROPT problems which are a particular case of Problem $(P_1)$ with linear inequalities constraints of a special type and no convergence rate estimates are provided. Unfortunately the existing balancing-type algorithms for the ROT and ROPT problems become very unstable when the regularization parameter is chosen very small, which is the case when one needs to calculate a good approximation to the solution of the optimal transport (OT) or the optimal partial transport (OPT) problem. In practice the typical dimensions of the spaces $E,H_1,H_2$ range from thousands to millions, which makes it natural to use a first-order method to solve Problem $(P_1)$. A common approach to solve such large-scale Problem $(P_1)$ is to make the transition to the Lagrange dual problem and solve it by some first-order method. Unfortunately, existing methods which elaborate this idea have at least two drawbacks. Firstly, the convergence analysis of the Fast Gradient Method (FGM) [@Nesterov_2005_Smooth] can not be directly applied since it is based on the assumption of boundedness of the feasible set in both the primal and the dual problem, which does not hold for the Lagrange dual problem. A possible way to overcome this obstacle is to assume that the solution of the dual problem is bounded and add some additional constraints to the Lagrange dual problem in order to make the dual feasible set bounded. But in practice the bound for the solution of the dual problem is usually not known. In [@Gasnikov_2015_Entropy] the authors use this approach with additional constraints and propose a restart technique to define the unknown bound for the optimal dual variable value. Unfortunately, the authors consider only the classical ELP problem with only the equality constraints and it is not clear whether their technique can be applied for Problem $(P_1)$ with inequality constraints. Secondly, it is important to estimate the rate of convergence not only in terms of the error in the solution of the Lagrange dual problem at it is done in [@Polyak_2013_Dual; @Necoara_2008_Applications] but also in terms of the error in the solution of the primal problem [^1] $\|f(x_k)-Opt[P_1]\|$ and the linear constraints infeasibility $\\|A_1x_k-b_1\\|_{H_1}$, $\\|(A_2x_k-b_2)_+\\|_{H_2}$, where vector $v_+$ denotes the vector with components $[v_+]_i=(v_i)_+=\max\{v_i,0\}$, $x_k$ is the output of the algorithm on the $k$-th iteration, $Opt[P_1]$ denotes the optimal function value for Problem $(P_1)$. Alternative approaches [@Goldstein_2012_Fast; @Shefi_2014_Rate] based on the idea of the method of multipliers and the quasi-Newton methods such as L-BFGS also do not allow to obtain the convergence rate for the approximate primal solution and the linear constraints infeasibility. Our contributions in this work are the following. We extend the Fast Gradient Method [@Nesterov_2005_Smooth; @Devolder_2014_First] applied to the dual problem in order to make it primal-dual so that it allows not only to solve the dual problem, but also to construct nearly optimal and nearly feasible solution to the primal problem $(P_1)$. We also equip our method with a stopping criterion which allows an online control of the quality of the approximate primal-dual solution. Unlike [@Cuturi_2013_Sinkhorn; @Benamou_2015_Iterative; @Polyak_2013_Dual; @Necoara_2008_Applications; @Goldstein_2012_Fast; @Shefi_2014_Rate; @Gasnikov_2015_Entropy] we provide the estimates for the rate of convergence in terms of the error in the solution of the primal problem $\|f(x_k)-Opt[P_1]\|$ and the linear constraints infeasibility $\\|A_1x_k-b_1\\|_{H_1}$, $\\|(A_2x_k-b_2)_+\\|_{H_2}$. In the contrast to the estimates in [@Nesterov_2005_Smooth], our estimates do not rely on the assumption that the feasible set of the dual problem is bounded. At the same time our approach is applicable for the wider class of problems defined by $(P_1)$ than approaches in [@Cuturi_2013_Sinkhorn; @Gasnikov_2015_Entropy]. In the computational experiments we show that our approach allows to solve ROT problems more efficiently than the algorithms of [@Cuturi_2013_Sinkhorn; @Benamou_2015_Iterative; @Gasnikov_2015_Entropy] when the regularization parameter is small. Preliminaries {#S:prel} ============= Notation -------- For any finite-dimensional real vector space $E$ we denote by $E^$ its dual. We denote the value of a linear function $g \in E^$ at $x\in E$ by ${\langle}g, x {\rangle}$. Let $\\|\cdot\\|_E$ denote some norm on $E$ and $\\|\cdot\\|_{E,}$ denote the norm on $E^$ which is dual to $\\|\cdot\\|_E$ $$\\|g\\|_{E,} = \max_{\\|x\\|_E \leq 1} {\langle}g, x {\rangle}.$$ In the special case when $E$ is a Euclidean space we denote the standard Euclidean norm by $\\|\cdot\\|_2$. Note that in this case the dual norm is also Euclidean. By $\partial f(x)$ we denote the subdifferential of the function $f(x)$ at a point $x$. Let $E_1, E_2$ be two finite-dimensional real vector spaces. For a linear operator $A:E_1 \to E_2$ we define its norm as follows $$\\|A\\|_{E_1 \to E_2} = \max_{x \in E_1,u \in E_2^} \{{\langle}u, A x {\rangle}: \\|x\\|_{E_1} = 1, \\|u\\|_{E_2,} = 1 \}.$$ For a linear operator $A:E_1 \to E_2$ we define the adjoint operator $A^T: E_2^ \to E_1^$ in the following way $${\langle}u, A x {\rangle}= {\langle}A^T u, x {\rangle}, \quad \forall u \in E_2^, \quad x \in E_1.$$ We say that a function $f: E \to \R$ has a $L$-Lipschitz-continuous gradient if it is differentiable and its gradient satisfies Lipschitz condition $$\\|\nabla f(x) - \nabla f(y) \\|_{E,} \leq L \\|x-y\\|_E.$$ We characterize the quality of an approximate solution to Problem $(P_1)$ by three quantities $\e_f,\e_{eq}, \e_{in} > 0$ and say that a point $\hat{x}$ is an $(\e_f,\e_{eq}, \e_{in})$-solution to Problem $(P_1)$ if the following inequalities hold $$\|f(\hat{x}) - Opt[P_1]\| \leq \e_f , \quad \\|A_1\hat{x}-b_1\\|_2 \leq \e_{eq}, \quad \\|(A_2\hat{x}-b_2)_+\\|_2 \leq \e_{in}. \label{eq:sol_def}$$ Here $Opt[P_1]$ denotes the optimal function value for Problem $(P_1)$ and the vector $v_+$ denotes the vector with components $[v_+]_i=(v_i)_+=\max\{v_i,0\}$. Also for any $t \in R$ we denote by $\lceil t \rceil$ the smallest integer greater than or equal to $t$. Dual Problem ------------ Let us denote $\Lambda =\{\lambda = (\lambda^{(1)},\lambda^{(2)})^T \in H_1^ \times H_2^: \lambda^{(2)} \geq 0\}$. The Lagrange dual problem to Problem $(P_1)$ is $$(D_1) \quad \quad \max_{\lambda \in \Lambda} \left\{ - {\langle}\lambda^{(1)}, b_1 {\rangle}- {\langle}\lambda^{(2)}, b_2 {\rangle}+ \min_{x\in Q} \left( f(x) + {\langle}A_1^T \lambda^{(1)} + A_2^T \lambda^{(2)} ,x {\rangle}\right) \right\}. \notag $$ We rewrite Problem $(D_1)$ in the equivalent form of a minimization problem. $$\begin{aligned} & (P_2) \quad \min_{\lambda \in \Lambda} \left\{ {\langle}\lambda^{(1)}, b_1 {\rangle}+ {\langle}\lambda^{(2)}, b_2 {\rangle}+ \max_{x\in Q} \left( -f(x) - {\langle}A_1^T \lambda^{(1)} + A_2^T \lambda^{(2)} ,x {\rangle}\right) \right\}. \notag $$ We denote $$\vp(\lambda) = \vp(\lambda^{(1)}, \lambda^{(2)}) = {\langle}\lambda^{(1)}, b_1 {\rangle}+ {\langle}\lambda^{(2)}, b_2 {\rangle}+ \max_{x\in Q} \left( -f(x) - {\langle}A_1^T \lambda^{(1)} + A_2^T \lambda^{(2)} ,x {\rangle}\right) \label{eq:vp_def}$$ Note that the gradient of the function $\vp(\lambda)$ is equal to (see e.g. [@Nesterov_2005_Smooth]) $$\nabla \vp(\lambda) = \left( \begin{aligned} b_1 - A_1 x (\lambda)\\ b_2 - A_2 x (\lambda)\\ \end{aligned} \right), \label{eq:nvp}$$ where $x (\lambda)$ is the unique solution of the problem $$\max_{x\in Q} \left( -f(x) - {\langle}A_1^T \lambda^{(1)} + A_2^T \lambda^{(2)} ,x {\rangle}\right). \label{eq:inner}$$ Note that this gradient is Lipschitz-continuous (see e.g. [@Nesterov_2005_Smooth]) with constant $$L = \frac{1}{\nu}\left(\\|A_1\\|_{E \to H_1}^2+ \\|A_2\\|_{E \to H_2}^2\right).$$ It is obvious that $$Opt[D_1]=-Opt[P_2]. \label{eq:D_1_P_2_sol}$$ Here by $Opt[D_1]$, $Opt[P_2]$ we denote the optimal function value in Problem $(D_1)$ and Problem $(P_2)$ respectively. Finally, the following inequality follows from the weak duality $$Opt[P_1] \geq Opt[D_1]. \label{eq:wD}$$ Main Assumptions {#S:main_assum} ---------------- We make the following two main assumptions 1. The problem is simple in the sense that for any $x \in Q$ it has a closed form solution or can be solved very fast up to the machine precision. 2. The dual problem $(D_1)$ has a solution $\lambda^=(\lambda^{(1)},\lambda^{(2)})^T$ and there exist some $R_1, R_2 >0$ such that $$\\|\lambda^{(1)}\\|_{2} \leq R_1 < +\infty, \quad \\|\lambda^{(2)}\\|_{2} \leq R_2 < +\infty. \label{eq:l_bound}$$ Examples of Problem $(P_1)$ --------------------------- In this subsection we describe several particular problems which can be written in the form of Problem $(P_1)$. Entropy-linear programming problem [@Fang_1997_Entropy]. $$\min_{x \in S_n(1)} \left\{ \sum_{i=1}^n x_i \ln \left(x_i/\xi_i\right) : Ax =b \right\} \notag $$ for some given $\xi \in \R^n_{++} = \{x \in \R^n: x_i > 0 , i=1,...,n \}$. Here $S_n(1) = \{x \in \R^n : \sum_{i=1}^n x_i = 1, x_i \geq 0 , i=1,...,n \}$. Regularized optimal transport problem [@Cuturi_2013_Sinkhorn]. $$\min_{X \in \R^{p \times p}_+} \left\{ \gamma \sum_{i,j=1}^p x_{ij} \ln x_{ij} + \sum_{i,j=1}^p c_{ij} x_{ij}: Xe=a_1, X^Te=a_2 \right\}, \label{eq:ROT}$$ where $e \in \R^p$ is the vector of all ones, $a_1,a_2 \in S_p(1)$, $c_{ij} \geq 0, i,j=1,...,p$ are given, $\gamma > 0$ is the regularization parameter, $X^T$ is the transpose matrix of $X$, $x_{ij}$ is the element of the matrix $X$ in the $i$th row and the $j$th column. Regularized optimal partial transport problem [@Benamou_2015_Iterative]. $$\min_{X \in \R^{p \times p}_+} \left\{ \gamma \sum_{i,j=1}^p x_{ij} \ln x_{ij} + \sum_{i,j=1}^p c_{ij} x_{ij}: Xe \leq a_1, X^Te \leq a_2, e^T X e = m \right\}, \notag $$ where $a_1,a_2 \in \R^p_+$, $c_{ij} \geq 0, i,j=1,...,p$, $m >0$ are given, $\gamma > 0$ is the regularization parameter and the inequalities should be understood component-wise. Algorithm and Theoretical Analysis {#S:2} ================================== We extend the Fast Gradient Method [@Nesterov_2005_Smooth; @Devolder_2014_First] in order to make it primal-dual so that it allows not only to solve the dual problem $(P_2)$, but also to construct a nearly optimal and nearly feasible solution to the primal problem $(P_1)$. We also equip it with a stopping criterion which allows an online control of the quality of the approximate primal-dual solution. Let $\{\alpha_i\}_{i \geq 0}$ be a sequence of coefficients satisfying $$\begin{aligned} & \alpha_0 \in (0,1], \notag \\ & \alpha_k^2 \leq \sum_{i=0}^k{\alpha_i} , \quad \forall k \geq 1. \notag$$ We define also $C_k = \sum_{i=0}^k{\alpha_i}$ and $\tau_{i}=\frac{\alpha_{i+1}}{C_{i+1}}$. Usual choice is $\alpha_i=\frac{i+1}{2}$, $i\geq 0$. In this case $C_k=\frac{(k+1)(k+2)}{4}$. Also we define the Euclidean norm on $H_1^* \times H_2^$ in a natural way $$\\|\lambda\\|_2^2=\\|\lambda^{(1)}\\|_2^2 + \\|\lambda^{(2)}\\|_2^2$$ for any $\lambda=(\lambda^{(1)},\lambda^{(2)})^T \in H_1^ \times H_2^$. Unfortunately we can not directly use the convergence results of [@Nesterov_2005_Smooth; @Devolder_2014_First] for the reason that the feasible set $\Lambda$ in the dual problem $(D_1)$ is unbounded and the constructed sequence $\hat{x}_k$ may possibly not satisfy the equality and inequality constraints. Choose $\lambda_0 = (\lambda^{(1)}_0,\lambda^{(2)}_0)^T = 0$. Set $k$ = 0. Let the assumptions listed in the subsection \[S:main\_assum\] hold and $\alpha_i=\frac{i+1}{2}$, $i\geq 0$ in Algorithm \[alg:FGM\_gen\]. Then Algorithm \[alg:FGM\_gen\] will stop after not more than $$N_{stop} = \max \left\{ \left\lceil \sqrt{\frac{8L(R_1^2+R_2^2)}{\tilde{\e}_f}} \right\rceil, \left\lceil \sqrt{\frac{8L(R_1^2+R_2^2)}{R_1\tilde{\e}_{eq}}} \right\rceil, \left\lceil \sqrt{\frac{8L(R_1^2+R_2^2)}{R_2\tilde{\e}_{in}}} \right\rceil \right\} -1$$ iterations. Moreover after not more than $$N = \max \left\{ \left\lceil \sqrt{\frac{16L(R_1^2+R_2^2)}{\e_f}} \right\rceil, \left\lceil \sqrt{\frac{8L(R_1^2+R_2^2)}{R_1\e_{eq}}} \right\rceil, \left\lceil \sqrt{\frac{8L(R_1^2+R_2^2)}{R_2\e_{in}}} \right\rceil \right\} -1$$ iterations of Algorithm \[alg:FGM\_gen\] the point $\hat{x}_N$ will be an approximate solution to Problem $(P_1)$ in the sense of . \[Th:FGM\_compl\] From the complexity analysis of the FGM [@Nesterov_2005_Smooth; @Devolder_2014_First] one has $$C_k \vp(\eta_k) \leq \min_{\lambda \in \Lambda} \left\{ \sum_{i=0}^k{\alpha_i \left( \vp(\lambda_i) + {\langle}\nabla \vp(\lambda_i), \lambda-\lambda_i {\rangle}\right) } + \frac{L}{2} \\|\lambda\\|_2^2 \right\} \label{eq:FGM_compl}$$ Let us introduce a set $\Lambda_R =\{\lambda = (\lambda^{(1)},\lambda^{(2)})^T: \lambda^{(2)} \geq 0, \\|\lambda^{(1)}\\|_2 \leq 2R_1, \\|\lambda^{(2)}\\|_2 \leq 2R_2 \}$ where $R_1$, $R_2$ are given in . Then from we obtain $$\begin{aligned} & C_k \vp(\eta_k) \leq \min_{\lambda \in \Lambda} \left\{ \sum_{i=0}^k{\alpha_i \left( \vp(\lambda_i) + {\langle}\nabla \vp(\lambda_i), \lambda-\lambda_i {\rangle}\right) } + \frac{L}{2} \\|\lambda\\|_2^2 \right\} \leq \notag \\ & \leq \min_{\lambda \in \Lambda_R} \left\{ \sum_{i=0}^k{\alpha_i \left( \vp(\lambda_i) + {\langle}\nabla \vp(\lambda_i), \lambda-\lambda_i {\rangle}\right) } + \frac{L}{2} \\|\lambda\\|_2^2 \right\} \leq \notag \\ & \leq \min_{\lambda \in \Lambda_R} \left\{ \sum_{i=0}^k{\alpha_i \left( \vp(\lambda_i) + {\langle}\nabla \vp(\lambda_i), \lambda-\lambda_i {\rangle}\right) } \right\} + 2L(R_1^2+R_2^2). \label{eq:proof_st_1}\end{aligned}$$ On the other hand from the definition of $\vp(\lambda)$ we have $$\begin{aligned} & \vp(\lambda_i) = \vp(\lambda_i^{(1)},\lambda_i^{(2)}) = {\langle}\lambda_i^{(1)}, b_1 {\rangle}+ {\langle}\lambda_i^{(2)}, b_2 {\rangle}+ \notag \\ & + \max_{x\in Q} \left( -f(x) - {\langle}A_1^T \lambda_i^{(1)} + A_2^T \lambda_i^{(2)} ,x {\rangle}\right) = \notag \\ & = {\langle}\lambda_i^{(1)}, b_1 {\rangle}+ {\langle}\lambda_i^{(2)}, b_2 {\rangle}- f(x(\lambda_i)) - {\langle}A_1^T \lambda_i^{(1)} + A_2^T \lambda_i^{(2)} ,x(\lambda_i) {\rangle}. \notag\end{aligned}$$ Combining this equality with we obtain $$\begin{aligned} & \vp(\lambda_i) - {\langle}\nabla \vp (\lambda_i), \lambda_i {\rangle}= \vp(\lambda_i^{(1)},\lambda_i^{(2)}) - {\langle}\nabla \vp(\lambda_i^{(1)},\lambda_i^{(2)}), (\lambda_i^{(1)},\lambda_i^{(2)})^T {\rangle}= \notag \\ & = {\langle}\lambda_i^{(1)}, b_1 {\rangle}+ {\langle}\lambda_i^{(2)}, b_2 {\rangle}- f(x(\lambda_i)) - {\langle}A_1^T \lambda_i^{(1)} + A_2^T \lambda_i^{(2)} ,x(\lambda_i) {\rangle}-\notag \\ & - {\langle}b_1-A_1 x(\lambda_i),\lambda_i^{(1)} {\rangle}- {\langle}b_2-A_2 x(\lambda_i),\lambda_i^{(2)} {\rangle}= - f(x(\lambda_i)). \notag\end{aligned}$$ Summing these inequalities from $i=0$ to $i=k$ with the weights $\{\alpha_i\}_{i=1,...k}$ we get, using the convexity of $f(\cdot)$ $$\begin{aligned} & \sum_{i=0}^k{\alpha_i \left( \vp(\lambda_i) + {\langle}\nabla \vp(\lambda_i), \lambda-\lambda_i {\rangle}\right) } = \notag \\ & = -\sum_{i=0}^k \alpha_i f(x(\lambda_i)) + \sum_{i=0}^k \alpha_i {\langle}(b_1-A_1 x(\lambda_i),b_2-A_2 x(\lambda_i))^T, (\lambda^{(1)},\lambda^{(2)})^T {\rangle}\leq \notag \\ & \leq -C_kf(\hat{x}_k) + C_k {\langle}(b_1-A_1 \hat{x}_k,b_2-A_2 \hat{x}_k)^T, (\lambda^{(1)},\lambda^{(2)})^T {\rangle}. \notag\end{aligned}$$ Substituting this inequality to we obtain $$\begin{aligned} & C_k \vp(\eta_k) \leq -C_kf(\hat{x}_k) + \notag \\ & + C_k \min_{\lambda \in \Lambda_R} \left\{ {\langle}(b_1-A_1 \hat{x}_k,b_2-A_2 \hat{x}_k)^T, (\lambda^{(1)},\lambda^{(2)})^T {\rangle}\right\} + 2L(R_1^2+R_2^2) . \notag\end{aligned}$$ Finally, since $$\begin{aligned} &\max_{\lambda \in \Lambda_R} \left\{ {\langle}(-b_1+A_1 \hat{x}_k,-b_2+A_2 \hat{x}_k)^T, (\lambda^{(1)},\lambda^{(2)})^T {\rangle}\right\} = \notag \\ & =2 R_1 \\|A_1 \hat{x}_k - b_1 \\|_2 + 2R_2 \\| (A_2 \hat{x}_k - b_2)_+ \\|_2, \notag\end{aligned}$$ we obtain $$\vp(\eta_k) + f(\hat{x}_k) +2 R_1 \\|A_1 \hat{x}_k - b_1 \\|_2 + 2R_2 \\| (A_2 \hat{x}_k - b_2)_+ \\|_2 \leq \frac{2L(R_1^2+R_2^2)}{C_k}. \label{eq:vpmfxh}$$ Since $\lambda^=(\lambda^{(1)}, \lambda^{(2)})^T$ is an optimal solution of Problem $(D_1)$ we have for any $x \in Q$ $$Opt[P_1]\leq f(x) + {\langle}\lambda^{(1)}, A_1 x -b_1 {\rangle}+ {\langle}\lambda^{(2)}, A_2 x -b_2 {\rangle}.$$ Using the assumption and that $\lambda^{*(2)} \geq 0$ we get $$f(\hat{x}_k) \geq Opt[P_1]- R_1 \\|A_1 \hat{x}_k - b_1 \\|_2- R_2 \\|(A_2 \hat{x}_k - b_2)_+ \\|_2. \label{eq:fxhat_est}$$ Hence $$\begin{aligned} & \vp(\eta_k) + f(\hat{x}_k) = \vp(\eta_k) - Opt[P_2]+Opt[P_2] + Opt[P_1] - Opt[P_1] + f(\hat{x}_k) \stackrel{\eqref{eq:D_1_P_2_sol}}{=} \notag \\ & = \vp(\eta_k) - Opt[P_2]-Opt[D_1]+Opt[P_1] - Opt[P_1] + f(\hat{x}_k) \stackrel{\eqref{eq:wD}}{\geq} \notag \\ & \geq - Opt[P_1] + f(\hat{x}_k) \stackrel{\eqref{eq:fxhat_est}}{\geq} - R_1 \\|A_1 \hat{x}_k - b_1 \\|_2- R_2 \\|(A_2 \hat{x}_k - b_2)_+ \\|_2. \label{eq:aux1}\end{aligned}$$ This and give $$R_1 \\|A_1 \hat{x}_k - b_1 \\|_2 + R_2 \\| (A_2 \hat{x}_k - b_2)_+ \\|_2 \leq \frac{2L(R_1^2+R_2^2)}{C_k}. \label{eq:R_norm_est}$$ Hence we obtain $$\vp(\eta_k) + f(\hat{x}_k) \stackrel{\eqref{eq:aux1},\eqref{eq:R_norm_est}}{\geq} - \frac{2L(R_1^2+R_2^2)}{C_k}. \label{eq:vppfxhatgeq}$$ On the other hand we have $$\vp(\eta_k) + f(\hat{x}_k) \stackrel{\eqref{eq:vpmfxh}}{\leq} \frac{2L(R_1^2+R_2^2)}{C_k}. \label{eq:vppfxhatleq}$$ Combining , , we conclude $$\begin{aligned} &\\|A_1 \hat{x}_k - b_1 \\|_2 \leq \frac{2L(R_1^2+R_2^2)}{C_kR_1}, \notag \\ &\\| (A_2 \hat{x}_k - b_2)_+ \\|_2 \leq \frac{2L(R_1^2+R_2^2)}{C_kR_2 }, \notag \\ &\|\vp(\eta_k) + f(\hat{x}_k)\| \leq \frac{2L(R_1^2+R_2^2)}{C_k}. \label{eq:untileq}\end{aligned}$$ As we know for the chosen sequence $\alpha_i = \frac{i+1}{2}, i \geq 0$ it holds that $C_k=\frac{(k+1)(k+2)}{4} \geq \frac{(k+1)^2}{4}$. Then in accordance to after given in the theorem statement number $N_{stop}$ of the iterations of Algorithm \[alg:FGM\_gen\] the stopping criterion will fulfill and Algorithm \[alg:FGM\_gen\] will stop. Now let us prove the second statement of the theorem. We have $$\begin{aligned} & \vp(\eta_k) + Opt[P_1] = \vp(\eta_k) - Opt[P_2]+Opt[P_2]+Opt[P_1] \stackrel{\eqref{eq:D_1_P_2_sol}}{=} \notag \\ & = \vp(\eta_k) - Opt[P_2]-Opt[D_1]+Opt[P_1] \stackrel{\eqref{eq:wD}}{\geq} 0. \notag $$ Hence $$f(\hat{x}_k)- Opt[P_1] \leq f(\hat{x}_k) + \vp(\eta_k) \label{eq:fxhatmopt}$$ On the other hand $$f(\hat{x}_k) - Opt[P_1] \stackrel{\eqref{eq:fxhat_est}}{\geq} - R_1 \\|A_1 \hat{x}_k - b_1 \\|_2- R_2 \\|(A_2 \hat{x}_k - b_2)_+ \\|_2. \label{eq:fxhat_est1}$$ Note that since the point $\hat{x}_k$ may not satisfy the equality and inequality constraints one can not guarantee that $f(\hat{x}_k) - Opt[P_1] \geq 0$. From , we can see that if we set $\tilde{\e}_f = \e_f$, $\tilde{\e}_{eq} = \min\{\frac{\e_f}{2R_1},\e_{eq}\}$, $\tilde{\e}_{in} = \min\{\frac{\e_f}{2R_2},\e_{in}\}$ and run Algorithm \[alg:FGM\_gen\] for $N$ iterations, where $N$ is given in the theorem statement, we obtain that fulfills and $\hat{x}_N$ is an approximate solution to Problem $(P_1)$ in the sense of . Note that other authors [@Cuturi_2013_Sinkhorn; @Benamou_2015_Iterative; @Polyak_2013_Dual; @Necoara_2008_Applications; @Goldstein_2012_Fast; @Shefi_2014_Rate; @Gasnikov_2015_Entropy] do not provide the complexity analysis for their algorithms when the accuracy of the solution is given by . Preliminary Numerical Experiments {#S:3} ================================= To compare our algorithm with the existing algorithms we choose the problem of regularized optimal transport [@Cuturi_2013_Sinkhorn] which is a special case of Problem $(P_1)$. The first reason is that despite insufficient theoretical analysis the existing balancing type methods for solving this type of problems are known to be very efficient in practice [@Cuturi_2013_Sinkhorn] and provide a kind of benchmark for any new method. The second reason is that ROT problem have recently become very popular in application to image analysis based on Wasserstein spaces geometry [@Cuturi_2013_Sinkhorn; @Benamou_2015_Iterative]. Our numerical experiments were carried out on a PC with CPU Intel Core i5 (2.5Hgz), 2Gb of RAM using Matlab 2012 (8.0). We compare proposed in this article Algorithm \[alg:FGM\_gen\] (below we refer to it as FGM) with the following algorithms - Applied to the dual problem $(D_1)$ Conjugate Gradient Method in the Fletcher–Reeves form [@Fletcher_1964_Function] with the stepsize chosen by one-dimensional minimization. We refer to this algorithm as CGM. - The algorithm proposed in [@Gasnikov_2015_Entropy] and based on the idea of Tikhonov’s regularization of the dual problem $(D_1)$. In this approach the regularized dual problem is solved by the Fast Gradient Method [@Nesterov_2005_Smooth]. We will refer to this algorithm as REG; - Balancing method [@Bregman_1967_Proof; @Cuturi_2013_Sinkhorn] which is a special type of a fixed-point-iteration method for the system of the optimality conditions for the ROT problem. It is referred below as BAL. The key parameters of the ROT problem in the experiments are as follows - $n := dim(E) = p^2$ – problem dimension, varies from $2^4$ to $9^4$; - $m_1 := dim(H_1)= 2\sqrt{n}$ and $m_2= dim(H_2) = 0$ – dimensions of the vectors $b_1$ and $b_2$ respectively; - $c_{ij}$, $i,j = 1,p$ are chosen as squared Euclidean pairwise distance between the points in a $\sqrt{p} \times \sqrt{p}$ grid originated by a 2D image [@Cuturi_2013_Sinkhorn; @Benamou_2015_Iterative]; - $a_1$ and $a_2$ are random vectors in $S_{m_1}(1)$ and $b_1 = (a_1,a_2)^T$; - the regularization parameter $\gamma$ varies from 0.001 to 1; - the desired accuracy of the approximate solution in is defined by its relative counterpart $\varepsilon_f^{rel}$ and $\varepsilon_g^{rel}$ as follows $$\varepsilon_f = \varepsilon_f^{rel} \cdot f(x(\lambda_0)) \quad \varepsilon_{eq} = \varepsilon_g^{rel} \cdot \\|A_1x(\lambda_0) - b_1\\|_2,$$ where $\lambda_0$ is the starting point of the algorithm. Note that $\e_{in} = 0$ since no inequality constraints are present in the ROT problem. Figure \[fig:exp\_res\_cmp\_scale\] shows the number of iterations for the FGM, BAL and CGM methods depending on the inverse of the regularization parameter $\gamma$. The results for the REG are not plotted since this algorithm required one order of magnitude more iterations than the other methods. In these experiment we chose $n = 2401$ and $\varepsilon_f^{rel} =\varepsilon_g^{rel} = 0.01$. One can see that the complexity of the FGM (i.e. proposed Algorithm \[alg:FGM\_gen\]) depends nearly linearly on the value of $1/\gamma$ and this complexity is smaller than that of the other methods when $\gamma$ is small. ![Complexity of FGM, BAL and CGM as $\gamma$ varies[]{data-label="fig:exp_res_cmp_scale"}](1scale_dep.jpg){width="\textwidth"} Figure \[fig:exp\_res\_cmp\_acc\] shows the the number of iterations for the FGM, BAL and CGM methods depending on the relative error $\varepsilon^{rel}$. The results for the REG are not plotted since this algorithm required one order of magnitude more iterations than the other methods. In these experiment we chose $n = 2401$, $\gamma = 0.1$ and $\varepsilon_f^{rel} =\varepsilon_g^{rel} = \varepsilon^{rel}$. One can see that in half of the cases the FGM (i.e. proposed Algorithm \[alg:FGM\_gen\]) performs better or equally to the other methods. ![Complexity of FGM, BAL and CGM as the desired relative accuracy varies[]{data-label="fig:exp_res_cmp_acc"}](2err_iter.jpg){width="\textwidth"} Conclusion {#conclusion .unnumbered} ========== This paper proposes a new primal-dual approach to solve a general class of problems stated as Problem $(P_1)$. Unlike the existing methods, we managed to provide the convergence rate for the proposed algorithm in terms of the primal problem error $\|f(\hat{x}_k - Opt[P_1]\|$ and the linear constraints infeasibility $\\|A_1\hat{x}_k-b_1\\|_{2}$, $\\|(A_2\hat{x}_k-b_2)_+\\|_{2}$. Our numerical experiments show that our algorithm performs better than existing methods for problems of regularized optimal transport which are a special instance of Problem $(P_1)$ for which there exist efficient algorithms. [50]{} Fang, S.-C., Rajasekera J., Tsao H.-S.: Entropy optimization and mathematical programming. Kluwer’s International Series (1997). Golan, A., Judge, G., Miller, D.: Maximum entropy econometrics: Robust estimation with limited data. Chichester, Wiley (1996). Kapur, J.: Maximum – entropy models in science and engineering. John Wiley & Sons, Inc. (1989). Gasnikov, A. et.al.: Introduction to mathematical modelling of traffic flows. Moscow, MCCME, (2013) (in russian) Rahman, M. M., Saha, S., Chengan, U. and Alfa, A. S.: IP Traffic Matrix Estimation Methods: Comparisons and Improvements// 2006 IEEE International Conference on Communications, Istanbul, p. 90–96 (2006) Zhang Y., Roughan M., Lund C., Donoho D.: Estimating Point-to-Point and Point-to-Multipoint Traffic Matrices: An Information-Theoretic Approach // IEEE/ACM Transactions of Networking, 13(5), p. 947–960 (2005) Hastie, T., Tibshirani, R., Friedman, R.: The Elements of statistical learning: Data mining, Inference and Prediction. Springer (2009) Zou, H., Hastie, T.: Regularization and variable selection via the elastic net // Journal of the Royal Statistical Society: Series B (Statistical Methodology) 67(2), p. 301–320 (2005) Cuturi, M.: Sinkhorn Distances: Lightspeed Computation of Optimal Transport // Advances in Neural Information Processing Systems, p. 2292–2300 (2013) Benamou, J.-D. , Carlier, G., Cuturi, M., Nenna, L., Peyre, G.: Iterative Bregman Projections for Regularized Transportation Problems //SIAM J. Sci. Comput., 37(2), p. A1111–A1138 (2015) Bregman, L.: The relaxation method of finding the common point of convex sets and its application to the solution of problems in convex programming // USSR computational mathematics and mathematical physics, 7(3), p. 200–217 (1967) Bregman, L.: Proof of the convergence of Sheleikhovskii’s method for a problem with transportation constraints // Zh. Vychisl. Mat. Mat. Fiz., 7(1), p. 147–156 (1967) Franklin, J. and Lorenz, J.: On the scaling of multidimensional matrices. // Linear Algebra and its applications, 114, 717–735 (1989) Nesterov, Yu.: Smooth minimization of non-smooth functions. // Mathematical Programming, Vol. 103, no. 1, p. 127–152 (2005) Gasnikov, A., Gasnikova, E., Nesterov, Yu., Chernov, A.: About effective numerical methods to solve entropy linear programming problem // Computational Mathematics and Mathematical Physics, , Vol. 56, no. 4, p. 514–524 (2016) <http://arxiv.org/abs/1410.7719> Polyak, R. A., Costa, J., Neyshabouri, J.: Dual fast projected gradient method for quadratic programming // Optimization Letters, 7(4), p.631–645 (2013) Necoara, I., Suykens, J.A.K.: Applications of a smoothing technique to decomposition in convex optimization // IEEE Trans. Automatic control, 53(11), p. 2674–2679 (2008). Goldstein, T., O’Donoghue, B., Setzer, S.: Fast Alternating Direction Optimization Methods. // Tech. Report, Department of Mathematics, University of California, Los Angeles, USA, May (2012) Shefi, R., Teboulle,M.: Rate of Convergence Analysis of Decomposition Methods Based on the Proximal Method of Multipliers for Convex Minimization // SIAM J. Optim., 24(1), p. 269–297 (2014) Devolder, O., Glineur, F., Nesterov, Yu.: First-order Methods of Smooth Convex Optimization with Inexact Oracle. Mathematical Programming 146(1–2), p. 37–75 (2014) 152, p. 381–404 (2015) Fletcher, R., Reeves, C. M.: Function minimization by conjugate gradients //Comput. J., 7, p. 149–154 (1964) [^1]: The absolute value here is crucial since $x_k$ may not satisfy linear constraints and hence $f(x_k)-Opt[P_1]$ could be negative.	{ "pile_set_name": "ArXiv" }
--- abstract: \| We present high resolution (R = 60,000) measurements of the NaI D1 $\&$ D2 (5890Å) and CaII K (3933Å) interstellar absorption line profiles recorded towards several post-AGB stars located within the M13 and M15 globular clusters, supplemented with a lower resolution spectrum of the CaII K-line observed in absorption towards an Ofpe/WN9 star in the central region of the M33 galaxy. The normalized interstellar absorption profiles have been fit with cloud component velocities, doppler widths and column densities in order to investigate the kinematics and physical conditions of the neutral and partially ionized gas observed along each sight-line. Our CaII observations towards M13 have revealed 4 absorption components that can be identified with galactic Intermediate Velocity Clouds (IVCs) spanning the -50 $>$ V$_{lsr}$ $>$ -80 km s$^{-1}$ range. The NaI/CaII ratio for these IVC’s is $<$ 0.3, which characterizes the gas as being warm (T $\sim$ 10$^{3}$K) and partially ionized. Similar observations towards two stars within M15 have revealed absorption due to a galactic IVC at V$_{lsr}$ $\sim$ +65 km s$^{-1}$. This IVC is revealed to have considerable velocity structure, requiring at least 3 cloud components to fit the observed NaI and CaII profiles. CaII K-line observations of a sight-line towards the center of the M33 galaxy have revealed at least 10 cloud components. A cloud at V$_{lsr}$ $\sim$ -130 km s$^{-1}$ is either an IVC associated with the M33 galaxy occurring at +45 km s$^{-1}$ with respect to the M33 local standard of rest, or it is a newly discovered HVC associated with our own Galaxy. In addition, 4 clouds have been discovered in the -165 $>$ V$_{lsr}$ $>$ -205 km s$^{-1}$ range. Three of these clouds are identified with the disk gas of M33, whereas a component at -203 km s$^{-1}$ could be IVC gas in the surrounding halo of M33. author: - 'Barry Y. Welsh, Jonathan Wheatley and Rosine Lallement' title: 'High resolution NaI and CaII absorption observations towards M13, M15 and M33.' --- Introduction ============ It has long been known that the Galaxy is surrounded by ionized and neutral gas clouds in the form of a hot and ionized galactic halo [@spitz56], in addition to the numerous neutral intermediate-velocity (IV) and high-velocity (HV) gas clouds [@wak97]. These two types of cloud are normally defined by their observed radial velocities, such that IVCs have velocities of $\mid$ V$_{LSR}$ $\mid$ 30 - 90 km s$^{-1}$ and HVCs have velocities of $\mid$ V$_{LSR}$ $\mid$ $>$ 90 km s$^{-1}$. Understanding the physical and chemical characteristics of both types of gas cloud is fundamental in defining our current ideas on how matter is exchanged between the Galaxy and the surrounding intergalactic medium. Although the origins of both IVCs and HVCs are still hotly debated [@spit08], it seems likely that they are linked to either expelled gas from a supernova-driven galactic fountain [@breg80] or they represent accumulations of tidally-stripped or condensed and cooled gas [@put03; @mall04]. Irrespective of their true origin, it is generally agreed that IVCs and HVCs are providing infalling gas, which is theorized to be required in order to sustain both the star formation process and the presence of an interstellar medium in our Galaxy [@van62]. If IVCs and HVCs are a major source of this gaseous infall, then knowledge of the mass and chemical composition of each cloud are key input parameters for any model of galactic evolution and star formation [@alib01]. The chemical composition and physical state of gas in HVCs and IVCs are best determined through absorption measurements at ultraviolet wavelengths [@wak01; @richt01; @coll07]. These data suggest that the majority of neutral HVCs have subsolar metallicities, implying that some fraction of the gas is of a primordial extragalactic origin. However recent observations of a new type of highly ionized HVCs that are not associated with neutral HI emission, have revealed supersolar metallicity values which suggests a galactic origin for the gas in these particular HVCs [@fox06; @zech08]. Similarly, the gas associated with IVCs also appears to possess near solar abundances [@richt01], except that the majority of IVCs are relatively nearby objects of distance $<$ 2kpc compared with the more distant HVCs of distance $>$ 5 kpc [@olan08]. However, we note that all these assumptions are based on relatively few observations of HVCs and IVCs, with the data often being of a low S/N ratio because of the relative faintness of the background sources used in the absorption observations. However, with the advent of high resolution spectrographs coupled to very large aperture ground-based telescopes, it is now possible to routinely investigate the interstellar gas residing beyond our own Galaxy through classical absorption spectroscopy techniques that involve distant (d $>$ 5 kpc) background stellar sources. In particular, the well-known interstellar lines of NaI-D (5890Å) and CaII-K (3933Å) have been extremely useful in revealing the presence of, and determining accurate distances to, IVCs and HVCs [@smok04; @smok07; @kenn98]. In particular, interstellar absorption observations of the sight-lines towards hot post-AGB stars within globular clusters of known distance has proven highly fruitful. Their sight-lines can provide (a) distance limits to absorption arising within foreground IVCs and HVCs as seen in their visible NaI and CaII interstellar absorption spectra [@lehn99] and (b) metal abundance determinations of foreground IVC and HVC gas, through interstellar absorption spectra of numerous ultraviolet resonance lines [@zech08]. In this Paper we present high resolution (R $\sim$ 60,000) observations of the NaI D-lines and the CaII K-line seen in absorption towards two post-AGB stars in the globular cluster NGC 7078 (M15) and toward one post-AGB star in NGC 6205 (M13). We also present a lower resolution CaII K-line spectrum of the hot star UIT-236 which resides in a planetary nebula in the central regions of the M33 galaxy. Using these data we discuss the velocity structure of the interstellar absorption observed along each sight-line and attempt to associate these cloud components with a galactic, extended halo or extra-galactic origin. These data will also supplement a forthcoming abundance analysis of these 4 sight-lines using ultraviolet absorption observations (with a resolving power of $\sim$ 20,000) using the Cosmic Origins Spectrograph due to be flown in mid-2009 on Servicing Mission 5 to the Hubble Space Telescope. Observations and Data Reduction =============================== We present absorption observations of the interstellar NaI D-line doublet at $\sim$ 5890Å and the CaII K-line at 3393Å recorded towards hot post-AGB stars in the M13 (NGC 6205) and M15 (NGC 7078) globular clusters, in addition to the Of/WN9 star UIT-236 in the M33 galaxy. The 4 target stars, together with associated astronomical information and appropriate references, are listed in Table 1. These data were obtained during service observing runs performed during 2008 March to July with the High Dispersion Spectrograph [@nog02] on the 8.2m Subaru telescope of the National Astronomical Observatory of Japan in Hawaii. The spectrograph was configured with a cross-disperser angle of 5.52$^{\circ}$ to receive light passing through a 0.6 arc sec wide x 5 arc sec long slit, with the spectra being recorded on two 4k x 2k pixel EEV CCD’s. Spectra were recorded separately for the NaI lines (using the StdYb instrumental set-up) and the CaII line (using the StdBc instrumental set-up). Unfortunately this set-up configuration precluded measurement of the CaII H-line at 3368Å, which appeared in an echelle order very close to the edge of the detector. The raw data were reduced in a similar manner to that described in Sfeir et al. [@sfeir99], which includes cosmic ray removal, CCD bias subtraction, flat-fielding and inter-order background subtraction. The spectra were wavelength calibrated using Th-Ar calibration lamp spectra, which resulted in a wavelength accuracy of $\sim$ 2 km s$^{-1}$. The spectral resolution of the M13 and M15 spectra was determined to be 5 km s$^{-1}$, whereas the resolution of the spectrum recorded towards M33 (which was recorded with 2 pixel binning) was 9 km s$^{-1}$. The telluric water vapor lines that contaminate the NaI absorption spectra were removed using a synthetic transmission spectrum described in Lallement et al. [@lall93]. The spectra were all well-exposed with typical S/N ratios in excess of 20:1. Finally, the spectral data were converted into velocities in the local standard of rest (LSR) frame. The calibrated spectra were fit with a high order polynomial to produce local continua in order to establish normalized residual intensities for the absorption lines of interest. These normalized profiles were then fit with multiple absorption components (i.e. interstellar clouds) using a line-fitting program described in Sfeir et al. [@sfeir99]. This program assigns a 3-parameter model fit to the observed profiles using values for the interstellar gas cloud velocity, V, a Gaussian velocity dispersion, $\it b$ and a cloud component column density, $\it N$. For the case of the NaI D1 and D2 line-doublet, the best-fit was performed simultaneously on both line profiles. Fits were carried out using the $\it minimum$ number of absorption components, with the addition of extra components only being deemed necessary if the $\chi$$^{2}$ residual error between the observed and model data points decreased by more than 11.1 [@vall93]. In order to derive meaningful interstellar gas temperatures for these fits, the $\it b$-values were constrained to be $<$ 2.9 km s$^{-1}$ ($<$20,000K ) for the CaII components and to be $<$3.3 km s$^{-1}$ ($<$ 15,000K) for the NaI components. The resultant best-fit values of V$_{lsr}$, $\it b$ and $\it N$ for the NaI and CaII absorption lines are listed in Table 2, and in Figure 1 we show the observed spectra and their respective model fits. We have also included upper limit estimates for certain important IV components where no absorption feature was detected with significance. These values were derived from conservative estimates of the strength of a potential absorption feature appearing at a level $>$ 2.5-$\sigma$ above the rms value of the local continuum. The corresponding upper limit on the column density of such features was determined under the assumption of a linear curve of growth and is reported in Table 2. Discussion ========== The sight-line to Barnard 29 in the M13 globular cluster -------------------------------------------------------- The 7.2kpc sight-line to Barnard 29 in the M13 globular cluster ($\it l$ = 59$^{\circ}$, $\it b$ = +41$^{\circ}$) passes through the galactic Hercules shell (GS 57+41) at a distance of $\sim$ 150pc [@lil92], as well as an Intermediate Velocity Cloud (IVC) first observed in 21cm HI emission [@kerr72]. The $\it I.U.E$ observations of Barnard 29 by de Boer and Savage [@deBoer83] revealed IV gas at a velocity centered at $\sim$ 80 km s$^{-1}$ in the MgII, CII and OI absorption lines. Due to the similar velocities of the UV and the HI data it was argued that this absorption arises at an unknown distance within the galactic halo. More recent Leiden-Dwingeloo HI observations have revealed the IV cloud to possess a central velocity of -64.8 km s$^{-1}$ [@smok06], and the HI maps of Kuntz $\&$ Danly [@kuntz96] show that M13 lies just off the boundary that defines the spatial extent of the IV Arch, a feature that extends over a large portion of the northern galactic hemisphere. Previous NaI D-lineand CaII K-line observations of sight-lines towards several stars within M13 (but not Barnard 29) have revealed IV absorption over the -65 to -72 km s$^{-1}$ range [@shaw96]. Observations of the CaII K-line in the sight-line towards Barnard 29 by Bates et al. [@bates95] have shown a blend of at least three IV absorption components spanning the -40 to -90 km s$^{-1}$ range, whereas the more recent measurements of Smoker et al. [@smok06] have revealed a similarly broad absorption feature that was fit with only one component centered at V$_{lsr}$ = -48.4 km s$^{-1}$. Both sets of data are consistent with our present CaII-K observations that reveal 4 components at V$_{lsr}$ = -50.1, -59.6, -70.1 and -78.6 km s$^{-1}$. Our NaI D-line spectra show no measurable absorption at velocities in the -40 to -90 km s$^{-1}$ range (to a detection level of $<$ 5 mÅ for the D2 line), in contrast with other (lower S/N) absorption observations of sight-lines towards 3 angular close stars in M13 [@shaw96]. Our integrated value of CaII column density for all 4 IV components is log N(CaII) = 11.64 cm$^{-2}$, which can be compared with the similar values of log N(CaII) = 11.65 and 11.92 cm$^{-2}$ measured respectively by Smoker et al [@smok06] and Bates et al. [@bates95]. Additionally, we note in Figure 3 of Smoker et al. [@smok06] there appears to be absorption features in both of the D2 and D1 line spectra of Barnard 29 in the -20 to -80 km s$^{-1}$ range. Unfortunately the velocities of these features are not common to both NaI lines, suggesting that they may be due to incomplete removal of telluric water vapor lines in this wavelength region. If this is the case, then both of our NaI D-line profiles argue against any measurable IV absorption formed over the same velocity range as the CaII-K line. There is a large ($\sim$ factor of 100) variation in the abundance of Na for high- and intermediate-velocity gas compared with that measured for low-velocity gas [@wak01], which may explain why, although NaI IV components have been detected along several other sight-lines to M13 [@shaw96], we have not been able to detect similar IV NaI towards Barnard 29. The NaI/CaII ratio for all four of the presently detected IV components is $<$ 0.3, which is a value found widely for interstellar clouds that can be characterized as being warm (T $\sim$ 10$^{3}$K), of low density (n$_{H}$ $<$ 1 cm$^{-3}$ and partially ionized [@hobbs75]. It has also been shown that the NaI/CaII ratio does not depend on velocity for sight-lines through IVCs and HVCs, since the ratio depends mostly on N(HI) which has lower values at higher velocities [@wak01]. The physical conditions of this IV gas can be compared with the main absorbing cloud at V$_{lsr}$ $\sim$ +10 km s$^{-1}$, whose higher NaI/CaII ratio of 3.0 is typical for clouds that are present in colder, more neutral and denser interstellar regions. As discussed in the Introduction measurement of the distance to IVCs and HVCs is an important parameter. We are able to place a lower limit to the distance of the IV absorption of d = 265pc, based on CaII absorption measurements towards the star HD 156633 ($\it l$ = 56.4$^{\circ}$, $\it b$ = +33.1$^{\circ}$) which does not reveal any measurable absorption components for V$_{lsr}$ $<$ - 30 km s$^{-1}$ [@welsh09]). The upper limit of d = 7.2kpc is obtained from the distance to the M13 globular cluster. Although we did not detect any IV absorption in the NaI lines, we note that this is also consistent with the NaI measurements towards the foreground star HD 147113 ($\it l$ = 61$^{\circ}$, $\it b$ = +46$^{\circ}$) of distance $\sim$600pc [@lil92]. The sight-lines to K648 and Zng1 in the M15 globular cluster ------------------------------------------------------------ Both of the 10.3 kpc long sight-lines towards the M15 globular cluster ($\it l$= 65$^{\circ}$, $\it b$=-27.3$^{\circ}$) pass through a previously known IVC with V$_{lsr}$ $\sim$ +70 km s$^{-1}$ [@cohen79; @kerr72]. More recent, higher resolution optical and HI observations of this IV gas (often called the ‘g1’ cloud) have resulted in a distance estimation of between 1.8 to 3.8 kpc to this IVC [@wakker08; @smoker02a; @smoker02b]. The absorbing gas within this cloud complex has been shown to exhibit significant small-scale structure with NaI column densities varying by up to a factor 16 across the foreground cloud [@meyer99]. The star Zng-1 has been previously observed at a spectral resolution of $\sim$ 7.5 km $^{-1}$ in both the NaI and CaII absorption lines [@smoker02a; @smoker02b]. These data revealed appreciable low velocity absorption covering the -30 to +20 km s$^{-1}$ range, which can be associated with gas in the galactic disk (and perhaps lower halo). An absorption feature centered at V$_{lsr}$ = +64 km s$^{-1}$ was observed in both the NaI and CaII lines, which matched the velocity of the HI emission from the foreground IVC gas recorded towards M15. An extra absorption component at V$_{lsr}$ = +53 km s$^{-1}$ was also tentatively suggested by these visible observations. Under the assumption that a value of the HI column density to Zng-1 is N(HI) = 5 x 10$^{19}$ cm$^{-2}$, the authors derived values of N(CaII) / N(HI) $\sim$ 5 x 10$^{-8}$ and N(NaI) / N(HI) $\sim$ 1.3 x 10$^{-8}$ for the entire IV gas cloud. The star K648, which is thought to be the central star of a planetary nebula lying within M15, has been observed in the ultraviolet and has been shown to possess a significant stellar wind with the nebula expanding at $\sim$ 15 km s$^{-1}$ with respect to the ambient medium [@bianch01] . No previous visible interstellar absorption observations exist for this star. Our present higher resolution NaI and CaII measurements towards both Zng-1 and K648 (which are separated by only 1.25 arc minute on the sky) show very similar patterns of visible absorption (see Figure 1). These profiles are dominated by low velocity ( -30 to +20 km s$^{-1}$) absorption originating in the galactic disk/halo, in addition to an IVC absorption centered at V$_{lsr}$ $\sim$ +65 km s$^{-1}$. The low-velocity galactic components generally have NaI/CaII ratios $>$ 1.0, consistent with the sight-lines sampling a cold and neutral diffuse interstellar medium. In contrast, the NaI/CaII ratios for all of the components associated with the IVC have values $<$ 1.0, suggestive of a warm, partially ionized and low density interstellar cloud medium. Our data also clearly show that the IVC gas sampled towards both stars has considerable velocity structure, requiring at least 3 cloud components to fit the observed NaI and CaII absorption profiles. This structure of the IVC, in which several clouds with similar velocities are present (i.e. +55 $<$ V$_{lsr}$ $<$ +75 km s$^{1}$), is consistent with the theoretical scenario in which possible cloud-to-cloud collisions could result in triggering the star formation process within such complexes [@christ97]. The sight-line towards the star UIT-236 in the M33 galaxy --------------------------------------------------------- M33, the Triangulum spiral galaxy (NGC 598), is a member of the Local Group lying at a distance of $\sim$ 820 kpc in the direction ($\it l$= 133.6$^{\circ}$, $\it b$=-31.3$^{\circ}$) and is about one quarter the size of both the Milky Way and Andromeda galaxies. It has a systemic velocity of V$_{lsr}$ $\sim$ -180 km s$^{-1}$ and is viewed near face on. The metallicity of the galaxy is sub-solar, with a radial abundance gradient that differs for different elements [@rosol08]. The interstellar sight-line towards the Of-type star UIT-236, which is part of the NGC 588 giant HII region, is close to the nucleus of the galaxy and low resolution ultraviolet observations have revealed significant mass-loss (through P-Cygni wind profiles) from this star [@bianch04]. Low resolution far ultra-violet absorption observations of UIT-236 recorded with $\it FUSE$ have revealed the presence of both hot 300,000$^{\circ}$K OVI and lower temperature CII gas at V$_{lsr}$ $\sim$ 180 km s$^{-1}$ that can be definitely associated with the host galaxy’s ISM [@wak03; @hutch04]. Observations of interstellar molecular H$_{2}$ and several low ionization ions towards the core of NGC 588 (i.e. UIT-236) with $\it FUSE$ have revealed a common absorption at V$_{lsr}$ $\sim$ -140 km s$^{-1}$ [@bluhm03]. This is a very similar absorption velocity to that of the ‘weak’ component seen in 21cm HI emission towards M33 by Rogstad et al. [@rog76]. Our present CaII-K absorption observations towards UIT-236 shown in Figure 1 reveal five major absorption systems centered at V$_{lsr}$ $\sim$ -205, -175, -130, -35 and 0 km s$^{-1}$. Unfortunately we have no accompanying interstellar NaI observations towards this star. However, we can associate the two sets of components observed at V$_{lsr}$ = 0 and -35 km s$^{-1}$ with absorption due to foreground galactic (and possibly inner galactic halo) interstellar gas. We note that the component at V$_{lsr}$ $\sim$ -35 km s$^{-1}$ is consistent with the velocity of gas associated with the outer region of the IVC HI gas which has been termed the ‘PP Arch’ [@wak01]. This stream of IV gas has an implied distance of 1.0 to 2.7 kpc with velocities in the -30 to -60 km s$^{-1}$ range [@wakker08]. If we assume an integrated HI column density of N(HI) $\sim$ 5 x 10$^{19}$ cm$^{-2}$ for the IV gas (-25 to -45 km s$^{-1}$) in this direction, then we obtain a N(CaII)/N(HI) ratio of $\sim$ 6 x 10$^{-8}$ for the IV cloud complex. Wakker $\&$ Mathis [@wakk00] discovered an empirical relationship between observed values of N(CaII) and N(HI) along many sight-lines, such that for the PP Arch IVC they predict a N(CaII)/N(HI) ratio that is slightly larger than our observed value. This supports the current notion that IVC’s seem deficient in ionized calcium compared with HVCs. The cloud component observed at V$_{lsr}$ $\sim$ -130 km s$^{-1}$ poses the question as to whether it is a high negative velocity HVC belonging to the Milky Way galaxy, or whether it is an IVC associated with the M33 galaxy that occurs at $\sim$ +45 km s$^{-1}$ with respect to the M33 local standard of rest? The observed velocity of this absorption component is very similar to that of both UV molecular H$_{2}$ and several low ionization lines measured towards NGC 588 with $\it FUSE$ [@bluhm03]. This is also very similar velocity to that of the ‘weak’ component seen in 21cm HI emission by Rogstad et al. [@rog76], who theorized that this absorption could be due to gas residing above and below the plane of the M33 galaxy. In addition the all-sky map of galactic HVC’s shows only highly negative (V $<$ -300 km s$^{-1}$) velocity gas present in the general direction of M33 [@wakker08]. We have not been able to confirm the presence of such HV gas with our present interstellar observations to a level of N(CaII) $<$ 10$^{11}$ cm$^{-2}$. Until future, more extensive, UV absorption observations of the V$_{lsr}$ $\sim$ -130 km s$^{-1}$ component are performed, we favor this component’s association with an IVC of the M33 galaxy. The interstellar CaII gas clouds observed in absorption with velocities in the range -175 $>$ V$_{lsr}$ $>$ - 200 km s$^{-1}$ can be directly associated with gas of the M33 galaxy, since such velocities are close to the systemic velocity of the host galaxy. As noted previously, the FUV absorption lines of interstellar molecular hydrogen, CII (1036Å), FeII (1144Å), ArI (1048Å) and OVI (1032Å) have also been detected with similar velocities towards M33 [@bluhm03; @hutch04]. The cloud component at V$_{lsr}$ $\sim$ -173 km s$^{-1}$ is saturated and is therefore most probably associated with gas in the disk of M33, which we are viewing almost face-on. The HI emission maps of M33 by Rogstad et al. [@rog76] show the presence of the main velocity components (-150 to - 200 km s$^{-1}$) occurring in the region within $\pm$5 arc min of the galaxy center. The component with V$_{lsr}$ = -203 km s$^{-1}$ could be an IVC of M33 with a relative velocity of $\sim$ -30 km s$^{-1}$ wrt the disk gas. We note that Grossi et al. [@gross08] have recently found several HI clouds surrounding M33 that fall into this velocity range. The current view of M33 is that it is a satellite of the larger M31 galaxy and the complex of surrounding HI (and HII) clouds may be either debris flowing into M33 from the IGM or from a previous interaction with M31. Gaining gas phase abundances of this component (from forthcoming UV absorption observations) would be extremely informative with respect to a better determination of the origin of this cloud. Conclusion ========== We have presented high resolution (R = 60,000) absorption measurements of the interstellar lines of NaI D1 $\&$ D2 (5890Å) and CaII-K (3933Å) recorded towards post AGB stars in the M13 and M15 globular clusters, supplemented with a lower resolution spectrum of the CaII-K line seen towards an Ofpe/WN9 star in the central region of the M33 galaxy. The absorption line-profiles have been fit with cloud component velocities and column densities such that the kinematics and physical conditions of the neutral and partially ionized gas components can be investigated. Four CaII-K absorption components that can be identified with galactic IVC’s spanning the -50 to -80 km s$^{-1}$ range have been detected towards the M13 globular cluster. The associated NaI/CaII column density ratio for this IVC gas is $<$ 0.3, which suggests that the gas is warm (T $\sim$ 10$^{3}$K) and partially ionized. The observations of two stars within M15 have revealed absorption due to a galactic IVC at V$_{lsr}$ $\sim$+65 km s$^{-1}$, which requires at least 3 separate cloud components (closely spaced in velocity) to adequately fit the NaI and CaII profiles. CaII K-line observations of the sight-line to the M33 galaxy have revealed at least 10 gas cloud components. A cloud at V$_{lsr}$ $\sim$ -130 km s$^{-1}$ is either an IVC associated with the M33 galaxy (occurring at +45 km s$^{-1}$ with respect to the M33 local standard of rest), or it is an newly discovered HVC associated with our own Galaxy. Also, 4 gas clouds have been detected in the -165 to -205 km s$^{-1}$ range, of which three are probably associated with the disk gas of M33. The cloud at V$_{lsr}$ = -203 km s$^{-1}$ could be an IVC residing in the surrounding halo of M33. Finally, we note that these high resolution visible data will be extremely useful in providing sight-line velocity templates for the forthcoming lower resolution UV absorption studies of these same stars to be carried out with the newly installed Cosmic Origins Spectrograph on the Hubble Space Telescope in mid-2009. We particularly acknowledge the dedicated team of engineers, technicians, and research staff who recorded these data with the Subaru Telescope, which is operated by the National Astronomical Observatory of Japan. This publication makes use of data products from the SIMBAD database, operated at CDS, Strasbourg, France. BYW acknowledges funding for this research through the NSF award AST-0507244. Alibes, A., Labay, J. and Canal, R., 2001, , 370, 1103 Bates, B., Shaw, C., Kemp, S. et al., 1995, , 444, 672 Bianchi, L., Bohlin, R., Catanzaro, G. et al., 2001, , 122, 1538 Bianchi, L., Bohlin, R. and Massey, P., 2004, , 610, 228 Bluhm, H., de Boer, K., Marggraf, O. et al., 2003, , 398, 983 Bregman, N., 1980, , 236, 577 Christodoulou, D., Tohline, J. and Keenan, F., 1997, , 468, 810 Cohen, J., 1979, , 231, 751 Collins, J., Shull, M. and Giroux, M., 2007, , 657, 271 de Boer, K. and Savage, B., 1983, , 265, 210 Dixon, van D. and Hurwitz, M., 1998, , 500, L29 Fox, A., Savage, B. and Wakker, B., 2006, , 165, 229 Grossi, M., Giovanardi, C., Corbelli, E. et al., 2008, , 487, 161 Hobbs, L., 1975, , 202, 628 Hutchings, J. and Butler, K., 2004, , 128, 2234 Kennedy, D., Bates, B. and Kemp, S., 1998, , 336, 315 Kerr, F. and Knapp, G., 1972, , 77, 354 Kuntz, K. and Danly, L., 1996, , 457, 703 Lallement, R., Bertin, P, Chassefiere, E. and Scott, N., 1993, , 271, 734 Lallement, R., Welsh, B.Y., Vergely, J.L. et al., 2003, 411, 447 Lehner, N., Rolleston, W., Ryans, R. et al., 1999, , 134, 257 Lilienthal, D., Hirth, W., Mebold, U. and de Boer, K., 1992, , 255, 323 Maller, A. and Bullock, J., 2004, , 355, 694 Meyer, D. and Lauroesch, J., 1999, , 520, L103 Mooney, C., Rolleston, W., Keenan, F. et al., 2004, , 419, 1123 Noguchi, K., Aoki, W., Kawanomoto, S. et al., 2002, PASJ, 54, 855 Olano, C., 2008, , 485, 457 Putman, M., Staveley-Smith, L., Freeman, K. et al., 2003, , 586, 170 Richter, P., Sembach, K., Wakker, B. et al., 2001, , 559, 318 Rogstad, D., Wright, M. and Lockhart, I., 1976, , 204, 703 Rosolowsky, E. and Joshua, S., 2008, , 675, 1213 Sfeir, D.M., Lallement, R., Crifo, F. and Welsh, B.Y., 1999, , 346, 785 Shaw, C., Bates, B., Kemp, S. et al., 1996, , 473, 849 Smoker, J., Keenan, F., Lehner, N. and Trundle, C., 2002a, , 387, 1057 Smoker, J., Haffner, L., Keenan, F. et al., 2002b, , 337, 385 Smoker, J., Lynn, B., Rolleston, W. et al., , 352, 1279 Smoker, J., Lynn, B., Christian, D. and Keenan, F., 2006, , 370, 151 Smoker, J., Hunter, I., Kalberla, P. et al., 2007, , 378, 947 Spitoni, E., Recchi, S. and Matteucci, F., 2008, , 484, 743 Spitzer, L.J., 1956, , 124, 20 Vallerga, J., Vedder, P., Craig, N. and Welsh, B.Y., 1993, , 411, 729 van den Bergh, S., 1962, , 67, 486 Wakker, B., 2001, , 136, 463 Wakker, B. $\&$ Mathis, J., 2000, , 544, L107 Wakker, B.P. $\&$ van Woerden, H., 1997, ARA$\&$A, 35, 217 Wakker, B., et al., 2003, , 146,1 Wakker, B.P., York, D., Wilhelm, R. et al., 2008, , 672, 298 Welsh, B. and Lallement, R., , (in prep) Zech, W., Lehner, N., Howk, C. et al., 2008, , 679, 460	{ "pile_set_name": "ArXiv" }
--- author: - 'P. F. L. Maxted' - 'R. J. Hutcheon' bibliography: - 'mybib.bib' date: Dates to be inserted title: 'Discovery and characterisation of long-period eclipsing binary stars from Kepler K2 campaigns 1, 2 and 3[^1],[^2]' --- [The Kepler K2 mission now makes it possible to find and study a wider variety of eclipsing binary stars than has been possible to-date, particularly long-period systems with narrow eclipses.]{} [Our aim is to characterise eclipsing binary stars observed by the Kepler K2 mission with orbital periods longer than $P\approx 5.5$ days. ]{} [ The `ellc` binary star model has been used to determine the geometry of eclipsing binary systems in Kepler K2 campaigns 1, 2 and 3. The nature of the stars in each binary is estimated by comparison to stellar evolution tracks in the effective temperature – mean stellar density plane. ]{} [43 eclipsing binary systems have been identified and 40 of these are characterised in some detail. The majority of these systems are found to be late-type dwarf and sub-giant stars with masses in the range 0.6–1.4 solar masses. We identify two eclipsing binaries containing red giant stars, including one bright system with total eclipses that is ideal for detailed follow-up observations. The bright B3V-type star HD 142883 is found to be an eclipsing binary in a triple star system. We observe a series of frequencies at large multiples of the orbital frequency in BW Aqr that we tentatively identify as tidally induced pulsations in this well-studied eccentric binary system. We find that the faint eclipsing binary EPIC 201160323 shows rapid apsidal motion. Rotational modulation signals are observed in 13 eclipsing systems, the majority of which are found to rotate non-synchronously with their orbits. ]{} [The K2 mission is a rich source of data that can be used to find long period eclipsing binary stars. These data combined with follow-up observations can be used to precisely measure the masses and radii of stars for which such fundamental data are currently lacking, e.g., sub-giant stars and slowly-rotating low-mass stars.]{} Introduction ============ Apart from the Sun and a few nearby stars, detached (i.e., non-interacting) eclipsing binaries (DEBS) provide the only means to measure accurate, model-independent masses and radii for normal stars. Using high-quality multi-wavelength photometry and high-resolution spectroscopy, masses and radii for stars in DEBS can be measured to $\pm 0.5$% or better . Spectral disentangling techniques also make it possible to determine the effective temperature (T$_{\rm eff}$) and surface composition of both stars in the binary from the analysis of their spectra [@2010ASPC..435..207P]. As a result, DEBS provide the most stringent test available for the accuracy of stellar evolution models for many different types of star . Empirical relations between mass, density, T$_{\rm eff}$ and metallicity based on DEBS can be used to estimate model-independent masses and radii for low-mass companions in SB1 eclipsing binaries, e.g., transiting hot-Jupiter systems [@2011MNRAS.417.2166S] or brown dwarf or very low mass stars in eclipsing binaries with solar-type stars . DEBS are also useful as distance indicators because their absolute magnitudes can be accurately estimated from the radii of the stars combined with a calibration of the stars’ surface brightness against colour or T$_{\rm eff}$ [@2017ApJ...837....7G]. DEBS have been used to investigate the systematic errors in parallax measurements for the Gaia DR1 data release [@2016ApJ...831L...6S], and to accurately measure the distance to the Magellanic Clouds [@2013Natur.495...76P; @2014ApJ...780...59G]. The Kepler K2 mission is providing very high quality photometry for thousands of moderately bright stars in selected regions of the sky (“campaign fields”) near the ecliptic plane [@2014PASP..126..398H]. Each campaign field is observed almost continuously for up to 80 days, making it possible to discover and characterise eclipsing binaries with orbital periods of weeks that are very hard to study using light curves obtained from ground-based instruments. Extracting high quality photometry from the K2 images is challenging because the spacecraft is being operated using only 2 reaction wheels. This operating mode has made it possible to extend the mission lifetime, but does result in the pointing of the spacecraft being less stable than during the original Kepler mission. Nevertheless, there is now a variety of algorithms available to correct for the instrumental noise caused by this pointing drift that make it possible to recover photometric performance better than 100ppm per 6-hours at 12th magnitude, close to the performance of the original Kepler mission . These algorithms are generally optimised for the detection of the periodic shallow eclipses in the light curves of transiting exoplanets. Eclipsing binary stars have been found both as a by-product of these searches for transiting exoplanets and by searches for variable stars of all types in the Kepler K2 data. To-date, the characterisation of these eclipsing binaries has not been very detailed, being limited to estimates of the period plus, in some cases, some basic characterisation of the eclipse properties, e.g., depth and width. At the time of writing, there are approximately 200 DEBS that have masses and radii measured to a precision of 2% or better [@2015ASPC..496..164S]. This sample is dominated by short-period systems ($P \loa 10$d) in which the components of the binary system are forced to co-rotate with the orbit. This makes it difficult to study phenomena such as interior mixing processes that can have subtle effects on the evolution of normal stars, but which may be disrupted by rapid rotation, particularly for sub-giant and giant stars. We have conducted our own search of the Kepler K2 data from campaigns 1, 2 and 3 and characterised the stars in these binaries in some detail using modelling of the Kepler K2 light curve plus existing optical and infrared photometry. Our study is motivated by the opportunity to study in detail stars of a type for which little fundamental accurate data are currently available. We have concentrated on bright stars with well-defined eclipses and long orbital periods that are ideal for detailed characterisation using high-resolution spectroscopy, but also discuss some other DEBS of interest that we have found in our survey. The results are presented here for the benefit of those who can share the task of characterising these binary systems and as a useful indicator of the number and properties of long-period eclipsing binaries that will be found in future large-scale photometric surveys. Analysis ======== Note that where we refer to the primary and secondary stars in the following description (star 1 and 2, respectively) these labels refer to the star eclipsed during the deeper and shallower eclipses in the K2 light curve, respectively, irrespective of the stars’ effective temperatures, masses, radii, etc. Target selection ---------------- Targets were identified by visual inspection of the detrended light curves generated by the [k2sff]{} algorithm [@2014PASP..126..948V]. We downloaded the light curve data from the Mikulski Archive for Space Telescopes[^3] (MAST) and used a simple script to plot the data for each system while making a note of any stars showing eclipse-like features in the light curve at least 5% deep and with orbital periods $P\ga 5.5$ days. We excluded stars from our list with a strong ellipsoidal effect in the light curve, i.e., a quasi-sinusoidal variation in flux with two maxima per orbital cycle due to the gravitational distortion of the stars in a close binary system. We also excluded systems fainter than Kepler magnitude ${\rm Kp} \approx 13$ unless they seemed particularly interesting based on an initial appraisal of the light curve or other information available. These points of interest are noted in section \[notes\_sec\]. The list of stars selected for further analysis is shown in Table \[lcinfo\] together with some basic characteristics of the light curves. The rotation periods $P_{\rm rot}$ listed in this table were determined as part of the detrending process described in section \[detrend\_sec\]. Aperture photometry and detrending \[detrend\_sec\] --------------------------------------------------- We downloaded the target pixel files for each target from MAST and used these data to produce light curves using synthetic aperture photometry. We first calculated the median value for every pixel in the data cube. The pixels in the lowest 10-percentile of this median image were then used to calculate the background level in the individual images. We used the target aperture specified in the target pixel file where available, otherwise we used a circular aperture centered on the flux-weighted centroid of the median image with a radius selected by-eye to encompass most of the flux in the star – typically 4–8 pixels. We also calculated the flux-weighted centroid within the target aperture for each image. The light curves produced by this method clearly show instrumental noise due to the varying position of the star on the detector. We used the [k2sc]{} algorithm [@2016MNRAS.459.2408A] to remove this instrumental noise. This algorithm uses Gaussian processes to decompose the light curve into a trend associated with the position of the star on the detector plus a trend with time that represents the intrinsic variability of the star. We first detrend the data using a squared-exponential kernel to describe the covariance properties of the trend with time. This kernel is suitable for smooth, aperiodic variations so we mask the eclipses for this calculation. We then use a Lomb-Scargle periodogram [@1992nrfa.book.....P] to characterise any periodic or quasi-periodic variability in the detrended light curve between the eclipses. This variability can be due to modulation of the light curve by star spots on one or both stars, or due to pulsations. The periods that we judged to be significant detected by this process are noted in Table \[lcinfo\] and are listed in order of power from strongest to weakest. For the stars whose period is noted in Table \[lcinfo\] we repeated the detrending using a quasi-periodic kernel for the time trend, again with the eclipses masked. In both cases (squared-exponential and quasi-periodic kernels) the trend with position determined from the data between the eclipses was used to interpolate a correction to the data during the eclipses. f These light curves are shown in Figs. \[lcplot1\], \[lcplot2a\], \[lcplot2b\] and \[lcplot3\]. ![image](Figs/lcplot1.png) ![image](Figs/lcplot2a.png) ![image](Figs/lcplot2b.png) ![image](Figs/lcplot3.png) WASP archive photometry ----------------------- The WASP project has obtained over 580 billion photometric observations for more than 30 million bright stars during a survey that has discovered more than 150 transiting exoplanets since observations started in May 2004 [@2006PASP..118.1407P]. WASP photometry is available for many of the systems in Table \[lcinfo\], but is of much lower quality than the K2 photometry. Nevertheless, WASP photometry has enabled us to determine or refine the orbital period for long-period binaries where only two or three eclipses have been observed by the Kepler K2 mission. ![image](Figs/201161715fit-eps-converted-to.pdf){width="49.00000%"} ![image](Figs/201246763fit-eps-converted-to.pdf){width="49.00000%"} ![image](Figs/201253025fit-eps-converted-to.pdf){width="49.00000%"} ![image](Figs/201379113fit-eps-converted-to.pdf){width="49.00000%"} ![image](Figs/201382417fit-eps-converted-to.pdf){width="49.00000%"} ![image](Figs/201408204fit-eps-converted-to.pdf){width="49.00000%"} ![image](Figs/201488365fit-eps-converted-to.pdf){width="49.00000%"} ![image](Figs/201576812fit-eps-converted-to.pdf){width="49.00000%"} ![image](Figs/201648133fit-eps-converted-to.pdf){width="49.00000%"} ![image](Figs/201665500fit-eps-converted-to.pdf){width="49.00000%"} ![image](Figs/201705526fit-eps-converted-to.pdf){width="49.00000%"} ![image](Figs/201723461fit-eps-converted-to.pdf){width="49.00000%"} ![image](Figs/202674012fit-eps-converted-to.pdf){width="49.00000%"} ![image](Figs/EPIC202674012_rvfit-eps-converted-to.pdf){width="49.00000%"} ![image](Figs/202843085fit-eps-converted-to.pdf){width="49.00000%"} ![image](Figs/203361171fit-eps-converted-to.pdf){width="49.00000%"} ![image](Figs/203371239fit-eps-converted-to.pdf){width="49.00000%"} ![image](Figs/203543668fit-eps-converted-to.pdf){width="49.00000%"} ![image](Figs/203610780fit-eps-converted-to.pdf){width="49.00000%"} ![image](Figs/203636784fit-eps-converted-to.pdf){width="49.00000%"} ![image](Figs/203728604fit-eps-converted-to.pdf){width="49.00000%"} ![image](Figs/204407880fit-eps-converted-to.pdf){width="49.00000%"} ![image](Figs/204748201fit-eps-converted-to.pdf){width="49.00000%"} ![image](Figs/204822807fit-eps-converted-to.pdf){width="49.00000%"} ![image](Figs/204760247fit-eps-converted-to.pdf){width="49.00000%"} ![image](Figs/204760247rvfit-eps-converted-to.pdf){width="49.00000%"} ![image](Figs/204870619fit-eps-converted-to.pdf){width="49.00000%"} ![image](Figs/205170307fit-eps-converted-to.pdf){width="49.00000%"} ![image](Figs/205020466fit-eps-converted-to.pdf){width="49.00000%"} ![image](Figs/205020466rvfit-eps-converted-to.pdf){width="49.00000%"} ![image](Figs/205546169fit-eps-converted-to.pdf){width="49.00000%"} ![image](Figs/205703649fit-eps-converted-to.pdf){width="49.00000%"} ![image](Figs/205919993fit-eps-converted-to.pdf){width="49.00000%"} ![image](Figs/205919993rvfit-eps-converted-to.pdf){width="49.00000%"} ![image](Figs/205982900fit-eps-converted-to.pdf){width="49.00000%"} ![image](Figs/206066862fit-eps-converted-to.pdf){width="49.00000%"} ![image](Figs/206066909fit-eps-converted-to.pdf){width="49.00000%"} ![image](Figs/206084435fit-eps-converted-to.pdf){width="49.00000%"} ![image](Figs/206109641fit-eps-converted-to.pdf){width="49.00000%"} ![image](Figs/206212261fit-eps-converted-to.pdf){width="49.00000%"} ![image](Figs/206241558fit-eps-converted-to.pdf){width="49.00000%"} ![image](Figs/206288770fit-eps-converted-to.pdf){width="49.00000%"} ![image](Figs/206433263fit-eps-converted-to.pdf){width="49.00000%"} The two WASP instruments are located at the Observatorio del Roque de los Muchachos, La Palma and at Sutherland Observatory, South Africa. Both instruments carry an array of eight wide-field cameras, each with a 2048 $\times$ 2048 pixel CCD detector. The majority of the survey has been conducted using 200-mm, f/1.8 lenses combined with a filter that defines a bandpass covering the wavelengths 400-700 nm [@2006PASP..118.1407P]. From July 2012 the WASP-South instrument has used 85-mm, f/1.2 lenses with SDSS r$^{\prime}$ filters [@2014CoSka..43..500S]. A dedicated pipeline is used to perform aperture photometry on the images at the position of catalogued stars within the images. The data are then processed by a detrending algorithm that has been developed from the SysRem algorithm of @2005MNRAS.356.1466T, as described by @2006MNRAS.373..799C. Light curve modeling \[ellc\] ----------------------------- We used version 1.6.1 of the `ellc` light curve model to determine the geometry and other parameters for each binary system. Note that the definition of the “third light” parameter used in this version of `ellc` to account for light from other stars in the photometric aperture is different to the one described in . In the new version, third light is described by the parameter $\ell_3$. This parameter is used to calculate the flux $ {\cal F}_3 = \ell_3 \left( {\cal F}_{\rm N,1} + {\cal F}_{\rm N,2}\right)$, where ${\cal F}_{\rm N,1}$ is the flux from star 1 emitted towards star 2 and vice versa. This value of ${\cal F}_3$ is then used in the calculation of the observed flux ${\cal F}_{i}$ at time $t_i$ as before, i.e., $${\cal F}_{i} = \frac{{\cal F}_{i,1} + {\cal F}_{i,2} + {\cal F}_3}{ {\cal F}_{\rm N,1} + {\cal F}_{\rm N,2}+ {\cal F}_3},$$ where ${\cal F}_{i,1}$ is the flux emitted by star 1 towards the observer at time $t_i$ and similarly for ${\cal F}_{i,2}$. A complete list of changes in `ellc` version 1.6.1 is provided in the file [CHANGELOG.rst]{} provided with the package distribution.[^4] The details of the analysis are not the same for every binary system because some binary systems have peculiarities that required special treatment. Here we outline the main features of the analysis applied to the majority of the systems analysed. Additional details and differences from this general approach for individual systems are described in Section \[notes\_sec\]. The free parameters in the model for each binary were: the sum of radii of the stars in units of the semi-major axis – $r_{\rm sum} = (R_1+R_2)/a$, the ratio of the radii – $k = R_2/R_1$; the surface brightness ratio in the Kepler band – $S_{\rm Kp}$; the orbital inclination, $i$; the time of primary eclipse – $T_0$; the orbital period – $P$; $f_s = \sqrt{e}\sin(\omega)$ and $f_c = \sqrt{e}\cos(\omega)$, where $e$ is the orbital eccentricity and $\omega$ is the longitude of periastron; and “third light” – $\ell_3$. We use $f_s$ and $f_c$ as parameters because a uniform prior probability distribution for these parameters corresponds to a uniform prior probability distribution for $e$. We use a quadratic limb-darkening law for both stars with priors on the coefficients calculated using [ldtk]{} [@2015MNRAS.453.3821P] based on the spherical model atmospheres by . To calculate these priors we assume $\log g =4.3 \pm 0.3$ and \[Fe/H\]$ = 0.0 \pm 0.2$ for all stars and effective temperature estimates from a preliminary analysis very similar to those derived in below in Section \[TeffSection\]. The standard error estimates on the coefficients inherited from the assumed errors on T$_{\rm eff}$, $\log g$ and \[Fe/H\] are likely to be underestimates of the true uncertainties since they do not account for systematic errors in the models and other issues with estimating limb darkening coefficients from models [@2011MNRAS.418.1165H]. To allow for this additional uncertainty we add 0.05 in quadrature to the standard error estimates for both coefficients. This estimate of the systematic error in the coefficients dominates the error budget for the limb darkening so we did not consider it necessary to re-calculate these coefficients for the slightly different values of $T_{\rm eff}$ derived in Section \[TeffSection\] cf. our preliminary solution. Rather than sampling the limb darkening coefficients $u_1$ and $u_2$ directly, we use the parameters $q_1 = (u_1 + u_2)^2$ and $q_2 = 0.5u_1/(u_1 + u_2)$ since this makes it easier to uniformly sample the allowed parameter space [@2013MNRAS.435.2152K]. Unless otherwise noted, we used spheres to model the shape of these well-detached stars so gravity darkening was ignored. There is little or no information about the geometry of the binary system in the observations between the eclipses. For the light curve modeling of most stars we used only observations over a range 1.5 times the full eclipse width centered on each eclipse. This had the advantage of speeding up the calculation. We used numerical integration of the eclipse model to account for the exposure time of 1765s for data obtained near or during an eclipse. It is notoriously difficult to include star spots in the model for an eclipsing binary star because the number of free parameters required is large and the constraints on these parameters from the light curve are generally weak and highly degenerate. We did not attempt to model star spots for any of the binary systems here since the amplitude of the star spot modulation is generally quite small ($\la 0.5\%$) so the resulting systematic error in the parameters derived will, in general, not be large enough to alter our conclusions regarding the nature of the binary. Instead, we simply divide-out the time trend due to star spot modulation established from the Gaussian process fit to the out-of-eclipse data. We used [emcee]{} [@2013PASP..125..306F], a [python]{} implementation of an affine invariant Markov chain Monte Carlo (MCMC) ensemble sampler, to calculate the posterior probability distribution of the model parameters. We used an ensemble with at least twice the number samples per chain step (“walkers”) as there were model parameters and 5,000 or 10,000 steps in the chain used for the results quoted below. The convergence of the chain was judged by visual inspection of the parameters and the likelihood as a function of step number. In cases where we suspected the chain had not sampled the posterior probability distribution accurately we calculated a new Markov Chain starting from the best-fit parameters in the previous chain and using an increased number of chain steps and/or an increased number of walkers with a large spread of initial parameter values to ensure convergence. The standard error per observation was either assumed to be constant for all the data, or assumed to be constant within each of two blocks of data where there is a gap in the observations. These values of the standard error were included as free parameters in the MCMC analysis by including the necessary term in the calculation of the likelihood for each chain step. Unless otherwise stated, we only use data within a range of 1.5 times the eclipse width (as listed in Table \[lcinfo\]) centred on each eclipse in this analysis. This ensures that these standard error estimates (and, hence, the error estimates on the model parameters) are determined by the scatter in the residuals through the eclipse, rather than the much lower scatter in the residuals between the eclipses. From preliminary fits to the complete light curves we found that the out-of-eclipse level is always very close to the value 1 with a very small error and is not correlated with the other parameters so we fix this parameter at 1 for the analysis presented here. The aim of this analysis is to characterise each binary system in order to identify systems of interest for further study and for comparison to binary population models. The parameters we have derived are reliable enough for this purpose but further work is needed to determine the accuracy of these parameters. The K2 data clearly have the potential to produce very precise parameters for some binary systems, but we have not attempted to characterise the level of systematic error in these parameters for all the systems studied. We advise that a careful study of these issues should be done before the parameters of individual binary systems are used to test stellar evolution models. Effective temperature estimates\[TeffSection\] ---------------------------------------------- We have used empirical colour – effective temperature and colour – surface brightness relations to estimate the effective temperatures of the individual stars in the binary and triple systems we have studied. We extracted photometry for each target from the following catalogues – B$_{\rm T}$ and V$_{\rm T}$ magnitudes from the Tycho-2 catalogue ; B, V, g$^{\prime}$, r$^{\prime}$ and i$^{\prime}$ magnitudes from data release 9 of the AAVSO Photometric All Sky Survey [APASS9, @2016yCat.2336....0H]; J, H and K$_{\rm s}$ magnitudes from the Two-micron All Sky Survey [2MASS, @2006AJ....131.1163S]; i$^{\prime}$, J and K magnitudes from the Deep Near-infrared Southern Sky Survey [DENIS, @2005yCat.2263....0T]. Not all stars have data in all these catalogues. Photometry from the Sloan digital sky survey (SDSS) can be unreliable for these bright stars because they saturate the detectors, but we have used g$^{\prime}$-, r$^{\prime}$- and i$^{\prime}$-band “psfMag” magnitudes from data release 9 of the SDSS [@2012ApJS..203...21A] in some cases, as noted in Table \[TeffTable\]. Magnitudes from the APASS9 catalogue that are given with a standard error estimate of 0.00 were not included in our analysis. Our model for the observed photometry then has the following free parameters that are determined by a least-squares fit to the observed apparent magnitudes and other data for each system – g$^{\prime}_{0, i}$, the apparent g$^{\prime}$-band magnitudes for stars $i=1$, $i=2$ and (for triple systems) $i=3$, corrected for extinction; $T_{\rm eff,i}$ the effective temperatures for each star in the binary or triple system; E$({\rm B}-{\rm V})$, the reddening to the system; $\sigma_{\rm ext}$ the additional systematic error added in quadrature to each synthetic magnitude to account for systematic errors in the conversion to observed magnitudes. For each trial combination of these parameters we use the empirical colour – effective temperature relations by @2013ApJ...771...40B to predict the apparent magnitudes for each star in each of the observed bands. We used the same transformation between the Johnson and 2MASS photometric systems as Boyajian et al.. We used the Cousins I$_{\rm C}$ band as an approximation to the DENIS Gunn i$^{\prime}$ band and the 2MASS K$_{\rm s}$ band as an approximation to the DENIS K band [see Fig. 4 @2005ARA+A..43..293B]. We used interpolation in Table 3 of @2000PASP..112..961B to transform the Johnson B, V magnitudes to Tycho-2 B$_{\rm T}$ and V$_{\rm T}$ magnitudes. We assume that the extinction in the V band is $3.1\times {\rm E}({\rm B}-{\rm V})$. Extinction in the SDSS and 2MASS bands is calculated using A$_{\rm r} = 2.770\times {\rm E}({\rm B}-{\rm V})$ from and extinction coefficients relative to the r$^{\prime}$ band from @2014MNRAS.440.3430D. We use the transformation from Sloan g$^{\prime}$, r$^{\prime}$ and i$^{\prime}$ magnitudes by @2011AJ....142..112B to estimate Kepler Kp magnitudes for each star in the system. This enables us to include the flux ratio $\ell_{\rm Kp}$ as a constraint in the analysis of the published photometry. Another useful constraint is the surface brightness ratio in the Kepler band, $S_{\rm Kp}$, which we account for by using the empirical relation between the V-band surface brightness $S_{\rm V}$ and $(\rm B - \rm K)$ from @2017ApJ...837....7G. The comparison between the predicted and observed values is done in terms of the surface brightness parameter $$S_i = m_{i,0} + 5\log\phi,$$ where $i$ denotes a particular band (V or Kp), $\phi$ is the angular diameter in milli-arcseconds, and $m_{i,0}$ is the de-reddened apparent magnitude in a given band, so that $S_{\rm Kp} = S_{\rm V} + ({\rm Kp}-{\rm V})$. We used [emcee]{} [@2013PASP..125..306F] to sample the posterior probability distribution for our model parameters. We used the reddening maps by @2011ApJ...737..103S to estimate the total line-of-sight extinction to each target, ${\rm E}({\rm B}-{\rm V})_{\rm map}$. This value is used to impose the following (unnormalized) prior on $\Delta = {\rm E}({\rm B}-{\rm V}) - {\rm E}({\rm B}-{\rm V})_{\rm map}$: $$P(\Delta) = \left\{ \begin{array}{ll} 1 & \Delta \le 0 \\ \exp(-0.5(\Delta/0.034)^2) & \Delta > 0 \\ \end{array} \right.$$ The constant 0.034 is taken from @2014MNRAS.437.1681M and is based on a comparison of ${\rm E}({\rm B}-{\rm V})_{\rm map}$ to ${\rm E}({\rm B}-{\rm V})$ from Strömgren photometry for 150 A-type stars. A least-squares optimisation algorithm was used to find an initial set of parameters for the chain and the Markov chains were calculated using 64 walkers and 256 steps following a burn-in run of 128 steps. An example of the output from the program used to implement our method is shown in Fig. \[fitmag\]. Read 17 lines from EPIC201161715.phot Calculating least-squares solution... g_1 = 15.50 T_eff,1 = 4931 K g_2 = 17.25 T_eff,2 = 5264 K E(B-V) = 0.00 chi-squared = 3.44 Ndf = 12 sigma_ext = 0.05 Starting emcee chain of 384 steps with 64 walkers Median acceptance fraction = 0.487 Best log(likelihood) = 42.52 in walker 12 at step 155 Parameter median values, standard deviations and best-fit values. g_1 = 15.362 +/- 0.102 [ 15.460 ] T_eff,1 = 5031 +/- 84 K [ 4958 ] g_2 = 17.148 +/- 0.207 [ 17.231 ] T_eff,2 = 5366 +/- 120 K [ 5282 ] E(B-V) = 0.034 +/- 0.025 [ 0.011 ] sig_ext = 0.027 +/- 0.016 [ 0.019 ] chi-squared = 9.90 type band value_obs error source value_fit value_A value_B z ---- ---- --------- ------ -------- --------- -------- -------- ---- mag B 15.759 0.208 APASS9 15.7578 15.9591 17.6869 0.01 mag H 12.769 0.021 2MASS 12.8287 12.9722 15.0962 2.11 mag Ic 13.916 0.03 DENIS 13.9310 14.0947 16.0660 0.42 mag J 13.227 0.07 DENIS 13.2451 13.3976 15.4518 0.25 mag J 13.283 0.027 2MASS 13.2451 13.3976 15.4518 1.15 rat K_p 0.18 0.03 paper-v3 0.1771 14.8825 16.7617 0.10 sb2 K_p 1.4 0.1 paper-v3 1.3832 5.3211 4.9689 0.17 mag Ks 12.602 0.14 DENIS 12.6763 12.8175 14.9610 0.53 mag Ks 12.725 0.03 2MASS 12.6763 12.8175 14.9610 1.37 mag V 14.881 0.024 APASS9 14.8917 15.0734 16.9224 0.35 mag g 15.2956 0.019 SDSS 15.3090 15.5032 17.2736 0.50 mag g 15.259 0.11 APASS9 15.3090 15.5032 17.2736 0.45 mag i 14.413 0.05 APASS9 14.3643 14.5308 16.4828 0.91 mag i 14.3639 0.0192 SDSS 14.3643 14.5308 16.4828 0.02 mag r 14.581 0.103 APASS9 14.5961 14.7691 16.6755 0.14 mag r 14.6083 0.0133 SDSS 14.5961 14.7691 16.6755 0.53 Nobs = 17 Nmag = 14 Ndf = 11 BIC = -68.04 Completed analysis of EPIC201161715.phot There will be some systematic error in the T$_{\rm eff}$ estimates for stars in eclipsing binaries cooler than 4900K because we have extrapolated the empirical $S_{\rm V}$ – $(\rm B - \rm K)$ relation in this regime. The empirical colour – temperature relations we have used are valid over the approximate range T$_{\rm eff} = 3450$K to 8600K. Our results may be biased in systems where one of the stars has an effective temperature near either of these limits because we exclude trial solutions with any T$_{\rm eff, i}$ value outside this range. Between these limits we use uniform priors on the values of T$_{\rm eff, i}$. We also use uniform priors for g$^{\prime}_{0, 1}$ and g$^{\prime}_{0, 2}$. In systems where there is evidence of third light from the light curve analysis and the star appears unresolved in sky survey images we compare solutions with a uniform prior on g$^{\prime}_{0, 3}$ and with a constraint on g$^{\prime}$$_{0, 3}$ assuming that the third light is due to a main-sequence star at the same distance as the eclipsing binary star. We use the stellar model from the Dartmouth stellar evolution database [@2008ApJS..178...89D] for solar composition to define the limits of the main sequence in the T$_{\rm eff}$ – $M_{\rm g^{\prime}}$ plane, where $M_{\rm g^{\prime}}$ is the absolute magnitude of star $i$ in the g$^{\prime}$ band. For each trial solution we use interpolation between these model isochrones to define limits to g$^{\prime}_{0, 3}$ assuming that the fainter star in the eclipsing binary is a main-sequence star, i.e., we reject solutions where the combination of T$_{\rm eff, 3}$, T$_{\rm eff, B}$, g$^{\prime}_{0, 3}$ and g$^{\prime}_{0, B}$ cannot be reproduced by two stars between the zero-age main sequence and terminal-age main-sequence in the T$_{\rm eff}$ – $M_{{\rm g}^{\prime}}$ plane, where $B = 1$ or 2 is the index for the star in the eclipsing binary that is fainter in the g$^{\prime}$-band. Systems where we adopted solutions including this constraint are noted with a $\star$ symbol in Table \[TeffTable\], together with the median and standard deviation of the model parameters derived using [emcee]{}. Our method requires an estimate of the apparent g$^{\prime}$ magnitude. In cases where no such estimate is available from APASS9 we either use the SDSS g$^{\prime}$ magnitude or infer a value from the Tycho-2 B$_{\rm T}$ and V$_{\rm T}$ magnitudes using equation (6a) from @2011AJ....142..112B. In either case, we assign an nominal standard error of 0.5 magnitudes to this estimate. We also found for some stars that the magnitudes from the DENIS and 2MASS surveys were significantly different. In general, we used the 2MASS magnitudes in these cases and excluded the DENIS photometry from the fit – these cases are noted in Table \[TeffTable\]. Results ======= The parameters derived from our analysis of the K2 light curves for each target are given in Tables \[ellcpar1\] and \[ellcpar2\]. The best fits to the K2 light curves are shown in Figs. \[lcfit1\], \[lcfit2\], \[lcfit3\] and \[lcfit4\]. The effective temperature estimates for the components of each system are given in Table \[TeffTable\]. In Table \[TrhoTable\] we give an estimate of the mean stellar density ($\rho_{\star}$) for the two stars in each eclipsing binary calculated using the following expression derived from Kepler’s third law. $$\mbox{$\rho_{\star}$} = \frac{3\mbox{M$_{\star}$}}{4\pi R_{\star}^3} = \frac{3\pi}{GP^2(1+q)} \left(\frac{a}{\mbox{R$_{\star}$}}\right)^3.$$ Here, $P$ and $a$ are the period and semi-major axis of the Keplerian orbit, and $q = M_\mathrm{c}/\mbox{$M_{\star}$}$ is the mass ratio for a companion with mass $M_\mathrm{c}$ to a star with mass $M_{\star}$ and radius $R_{\star}$. The value of $q$ was estimated by calculating the position of the stars in Fig. \[TeffDensity\] for various values of $q$ and then choosing the value which is consistent with the approximate masses inferred from the stellar evolution tracks shown in this figure. ![image](Figs/TeffDensity-eps-converted-to.pdf){width="99.00000%"} Notes on individual objects \[notes\_sec\] ------------------------------------------ ### EPIC 201160323 This faint star shows rapid apsidal motion. The period measured from the times of primary and secondary eclipse in the K2 light curve are $P_{\rm pri} =22.272$d and $P_{\rm sec} = 22.300$d, respectively. We therefore included the rate of change of the longitude of periastron in the light curve model as a free parameter and hence obtained the value $\frac{d\omega}{dt} = -0.10^{\circ} \pm 0.01^{\circ}$ per anomalistic period. This corresponds to an apsidal motion period of approximately 220 years if this rate is assumed to be constant. ![image](Figs/EPIC201160323_lcfit-eps-converted-to.pdf){width="99.00000%"} Our best-fit model is shown in Fig. \[201160323\_lcfit\], where the drift in eclipse times relative to a single linear ephemeris calculated with the average period can be clearly seen. There are no nearby stars listed in the GAIA DR1 catalogue that might explain the large value for the third light parameter derived from the light curve analysis ($\ell_3 = 0.43 \pm 0.05$). This suggests that EPIC 201160323 is a triple or multiple star system in which the gravitational interaction between the eclipsing binary and an additional body or bodies is causing the rapid change in the orientation of its orbit. This star is listed in the K2 Variable Catalogue as an eclipsing binary with a period of 22.299969d, which matches closely our estimate of $P_{\rm sec}$. Note that the orbital period values given in Tables \[lcinfo\] and \[ellcpar1\] are the anomalistic period. We did not attempt to estimate the effective temperatures of the stars in this system from the published photometry because there are no reliable photometric measurements at optical wavelengths – EPIC 201160323 is too faint to appear in either the APASS9 or Tycho-2 catalogues. ### EPIC 201161715 Star 1 is much larger than star 2 but the stars have similar effective temperatures so we assume that Star 1 is a sub-giant or red giant and $q =M_2/M_1 < 1$ (since the more massive star will have evolved off the main sequence first). For any reasonable choice of $q<1$ we find that star 1 is a red giant with a mass $M_1\approx 1.4\,M_{\sun}$. The evolution tracks for different masses have similar values of T$_{\rm eff}$ on the red giant branch so this mass is quite uncertain if we consider the properties of star 1 only. However, star 2 appears near the main-sequence turn off point so must have a mass $\ga 0.8\,M_{\sun}$. Both stars are in relatively short-lived evolutionary phases and the main-sequence life time decreases rapidly with increasing mass, so the mass ratio cannot be very different from 1. We conclude that $q\approx 0.8$ such that $M_1\approx 1.2\,M_{\sun}$ and $M_2\approx 0.95\,M_{\sun}$. From Fig. \[TeffDensity\] it can be seen that if $q\approx 0.8$ then this binary contains a star near the main-sequence turn-off point (MSTO) and a star at the base of the red giant branch, similar to the well-known systems AI Phe and TZ For . This makes this system an attractive target for calibrating stellar models. This star is listed in the K2 Variable Catalogue as an eclipsing binary with a period of 59.889024d, which agrees well with our period estimate. ### EPIC 201246763 The K2 light curve of this star shows one primary eclipse and two secondary eclipses. The position of the star on the detector during the second of the secondary eclipses is not well sampled by the other observations of this star so the detrending corrections applied to some of the data in this eclipse are extrapolated from the out-of-eclipse data. There are distinct differences between the shape and depth of this eclipse between the first and second observation of this feature in the K2 light curve. This makes it difficult to determine a precise value for the orbital period using the K2 data alone. Fortunately, the observations of this star from the WASP photometric archive have good coverage of both eclipses of this star that can be used to measure the orbital period to good precision. We used a least-squares fit with the [jktebop]{}[^5] model to 664 observations around primary and secondary eclipse from the WASP photometric archive to measure the orbital period of the binary. The WASP data cover the minima of two primary eclipses and one secondary eclipse plus a few observations of the ingress or egress to an eclipse. The first eclipse in the WASP data occurs on JD 2454881. We included the time of mid-eclipse from a preliminary fit to the K2 light curve as a constraint in this fit. The geometric parameters of the binary system were fixed at values from the same preliminary fit to the K2 light curve. The orbital period value we obtained is $43.68281 \pm 0.00003 $ days. We imposed this value as a prior on the orbital period for our final analysis of the K2 light curve using [emcee]{}. From Fig. \[TeffDensity\] it can be seen that this binary contains two main-sequence stars with masses $M_1\approx 1.0\,M_{\sun}$ and $M_2\approx 1.2\,M_{\sun}$. These mass estimates are quite robust because both stars lie near the evolution tracks with these masses for any reasonable choice of the mass ratio, $q$. ### EPIC 201253025 The aperture used to calculate the light curve is contaminated by another star approximately 4.7 arcseconds to the west of the main target and 1.6 magnitudes fainter in the G band according to the Gaia DR1 data release . We found that we could not get a good fit to the entire data set using one set of parameters, partly because the level of contamination from the nearby star is not constant. To deal with this problem we analysed separately the two parts of the light curve either side of the gap in the data at BJD 2456849. The results for two subsets of data are both given in Tables \[ellcpar1\] and \[ellcpar2\]. This approach does improve the fit to the two parts of the light curve, but residuals of about 0.5% are still apparent for some eclipses, presumably as a result of star spots on one or both stars that are also the cause of the quasi-periodic variations in flux between the eclipses. Despite these problems there is very good agreement between the geometric parameters derived from the two parts of the light curve. We set $\ell_3 = 0$ for our analysis of the published photometry to estimate T$_{\rm eff}$ because we assume that the value of $\ell_3$ in Table \[ellcpar2\] is due to the star 4.7 arcseconds to the west of the main target. This star is listed in the K2 Variable Catalogue as an eclipsing binary with a period of 6.785544d, which agrees well with our period estimate. High resolution imaging by @2016AJ....151..159S did not detect any companions to this star, with the quoted upper limit to the relative brightness at I-band being 2.02 magnitudes at 0.25arcsec. EPIC 201253025 contains a pair of quite similar stars so we assume $q\approx 1$, in which case the stars are towards the end of their main sequence lifetimes with masses $\approx 1.2M_{\sun}$ (Fig. \[TeffDensity\]). The rotation periods detected in the K2 light curve show that the stars in this binary system rotate non-synchronously, with one star rotating slightly faster than predicted for synchronous rotation and one slightly slower. The values of $R_{\star}/a \approx 0.07$ for these stars put them near the boundary between synchronous and non-synchronous rotation for stars with convective envelopes . This makes EPIC 201253025 an interesting test case for theories of the tidal interactions between low mass stars. ### EPIC 201379113 The secondary eclipse is very shallow (1.5%) and partial so it is not possible to determine a reliable value of $\ell_3$ from the K2 light curve alone. In addition, the observed flux between the eclipses varies by up to 0.4% on time scales of 10 days or more. There may be a rotation modulation signal with a period of about 22days in these flux variations, but we are not confident of this detection. We divided out these slow flux variations so the observed secondary depth varies systematically by a few parts per thousand. To derive the parameters in Tables \[ellcpar1\] and \[ellcpar2\] we fixed the value $\ell_3 = 0$. Even with this restriction, the additional noise in the eclipse depths results in quite large errors on the light curve model parameters for this binary. The precision of these parameters can certainly be improved using constraints on the luminosity ratio and third-light contribution from spectroscopy. This star is listed in the K2 Variable Catalogue as an eclipsing binary with a period of 21.186043d, which is slightly shorter than the period that we find from our analysis. From the location of these stars in the T$_{\rm eff}$ – $\rho_{\star}$ plane we estimate that they are dwarf stars with masses $M_1\approx 0.8M_{\sun}$ and $M_2\approx 0.6M_{\sun}$. This conclusion is not affected by the assumed value for the mass ratio for any reasonable estimate of $q\approx 0.7$. ### EPIC 201382417 The light curve between the eclipses shows a quasi-periodic variation that gradually increases from being barely detectable at the start of the K2 observing sequence up to an amplitude of 0.4%. We have divided out this trend rather than trying to fit a model to this variation. As a result, the depth of the secondary eclipse relative to this “corrected” out-of-eclipse level varies from about 1.5% at the start of the observing sequence to 1.2% in the second half of the data set. The best-fit solution to this corrected light curve has a secondary eclipse depth of 1.27%. The parameters in Tables \[ellcpar1\] and \[ellcpar2\] are very precise but there are certainly systematic errors in these values as a result of the detrending process, i.e., these parameters are much less accurate than implied that the quoted precision. To obtain a more accurate solution it will be necessary to identify and characterise the source or sources of the variation between the eclipses, i.e., to determine whether it is due to spot modulation on a third star that dominates the flux from this system (assuming our estimate of $\ell_3$ is accurate), or from the primary star in the eclipsing binary system, or a combination of both. Although the error bars quoted in Tables \[ellcpar1\] and \[ellcpar2\] are underestimates of the current accuracy in these parameters they do give a useful estimate of the accuracy that may be possible with a more complete model for this system. This star is listed in the K2 Variable Catalogue as an eclipsing binary with a period of 5.1976386d, which agrees well with our estimate of the orbital period. From the location of these stars in the T$_{\rm eff}$ – $\rho_{\star}$ plane (Fig. \[TeffDensity\]) we estimate that they are dwarf stars with masses $M_1\approx 1.2M_{\sun}$ and $M_2\approx 0.7M_{\sun}$. Both stars lie near the evolution tracks for these masses for any reasonable choice of $q=M_2/M_1$. ### EPIC 201408204 The stars in this binary system have very similar effective temperatures and radii so we assume $q\approx 1$. High resolution imaging by @2016AJ....151..159S did not detect any companions to this star, with the quoted upper limit to the relative brightness at I-band being 2.84 magnitudes at 0.25arcsec. This star is listed in the K2 Variable Catalogue as an eclipsing binary with a period of 8.482149d, which agrees well with our estimate of the orbital period. The rotation periods detected in the K2 light curve suggest that this pair of main-sequence stars with masses $M\approx 1 M_{\sun}$ (Fig. \[TeffDensity\]) rotate non-synchronously with the orbit although one of the rotation periods is close to the orbital period. However, the orbital eccentricity of this binary is quite large ($e\approx 0.2$) so in this case it makes more sense to compare the observed rotation periods to the “pseudo-synchronisation” rotation period determined by matching the angular velocity of the star to the orbital angular velocity at periastron . The corresponding ratio of the orbital and pseudo-synchronisation rotation periods is $(1+e)^2/(1-e^2)^{3/2} = 1.54$, suggesting that neither of the stars rotates pseudo-synchronously. This is another useful system for testing models of tidal dissipation in solar-type stars. ### EPIC 201488365 = [@lrr]{} & &\ Mass \[$M_{\sun}$\] & 1.32 $\pm$ 0.01 & 1.29 $\pm$ 0.01\ Radius \[$R_{\sun}$\] & 1.634 $\pm$ 0.005 & 1.498 $\pm$ 0.006\ $\log g$ \[cms$^{-2}$\] & 4.132 $\pm$ 0.002 & 4.197 $\pm$ 0.003\ T$_{\rm eff}$ \[K\] & 6430 $\pm$ 155 & 6420 $\pm$ 155\ $\log(L/L_{\sun})$ & 0.61 $\pm$ 0.04 & 0.54 $\pm$ 0.04\ M$_{\rm V}$ & 3.24 $\pm$ 0.16 & 3.44 $\pm$ 0.16\ Orbital period \[d\] &\ Mass ratio &\ Distance \[pc\] &\ @2010MNRAS.402.2424R have published spectroscopic orbits for both components of FM Leo together with an analysis of the light curves available to them at that time. We have used the semi-amplitudes $K_1$ and $K_2$ from Ratajczak et al. together with the parameters from our analysis of the Kepler K2 light curves with [jktabsdim]{}[^6] to derive the absolute parameters for FM Leo given in Table \[tab:fmleo\_abspar\]. The masses derived ($1.29M_{\sun}$ and $1.32M_{\sun}$) are in reasonable agreement with the estimate $M\approx 1.25M_{\sun}$ implied from the position of the stars in the T$_{\rm eff}$ – $\rho_{\star}$ plane (Fig. \[TeffDensity\]). The precision of the radius measurements is improved by an order of magnitude compared to what was possible with the data available to Ratajczak et al.. FM Leo could be a very useful system for testing stellar models if more precise estimates for the metallicity and effective temperature of the stars become available. The scatter in the residuals through the eclipses is approximately a factor of 2 larger than the residuals between the eclipses so it is likely that there is additional systematic error in the parameters derived from the K2 light curve comparable to the quoted error bars. This star is listed in the K2 Variable Catalogue as an eclipsing binary with a period of 3.364700d, which is approximately half of the correct orbital period. ### EPIC 201576812 = @2011AJ....142...50F present a detailed analysis of the WASP light curve and high-resolution spectroscopy of this eclipsing binary. They did not detect the secondary star in their spectroscopy and so to estimate the masses and radii of the stars they adopted the value $M_1 = 0.92 \pm 0.1M_{\sun}$ for the primary star mass based on the values T$_{\rm eff} = 5483$–5957K and \[Fe/H\] $= -0.28$ from the analysis of its spectrum compared to stellar evolution models. This star is listed in the K2 Variable Catalogue as an eclipsing binary with a period of 5.728410d, which agrees well with our period estimate. High resolution imaging by @2016AJ....151..159S did not detect any companions to this star, with the quoted upper limit to the relative brightness at I-band being 2.21 magnitudes at 0.25arcsec. As there is no evidence for third light in the spectrum of this star and there are no bright companions within the photometric aperture we have used, we set $\ell_3 = 0$ in our analysis of the K2 light curve. The geometric light curve parameters we obtain are not quite consistent with those of @2011AJ....142...50F at the 1-$\sigma$ level. This level of disagreement is not surprising given that the light curve of this star shows a shallow partial secondary eclipse plus rotational spot modulation visible between the eclipses with an amplitude $\approx 1\%$. ### EPIC 201648133 The K2 light curve of this star shows two primary eclipses and two secondary eclipses, with a gap in the data at the time of a primary eclipse near the middle of the observing sequence. A least-squares fit of a simple light curve model to the WASP photometry provides three times of primary eclipse as follows: HJD 2454852.4432(6), 2454922.4931(4), 2455237.7094(4), where figures in parentheses denote the standard error in the final digit of these values. From a fit to these times of mid-eclipse plus one further time of mid-eclipse from a preliminary fit to the K2 light curve we obtain $P=35.02402(1)$d. We imposed this value of the period with its standard error as a prior for our analysis of the K2 light curve. This star is listed in the K2 Variable Catalogue as an eclipsing binary, but no period estimate is given. High resolution imaging by @2016AJ....151..159S did not detect any companions to this star, with the quoted upper limit to the relative brightness at I-band being 3.18 magnitudes at 0.25arcsec. The location of these stars in the T$_{\rm eff}$ – $\rho_{\star}$ plane (Fig. \[TeffDensity\]) is consistent with the assumptions that they are dwarf stars with masses $M_1\approx 1.1M_{\sun}$ and $M_2\approx 0.85M_{\sun}$ for any reasonable estimate of the mass ratio, $q$. ### EPIC 201665500 This star is included in our study because we initially assumed the orbital period is approximately 6.1 days and that there are two similar eclipses in the light curve. In fact, the orbital period is half this value and there is a very shallow secondary eclipse visible in the K2 light curve. The primary eclipse in this light curve is a transit of a solar-type star by a low mass star. The secondary eclipse is very shallow compared to the star spot modulation visible between the eclipses (few parts per thousand cf. peak-to-peak amplitude $\approx 1.5$%) so there is considerable scatter in this secondary eclipse depth caused by dividing out the star modulation. As the secondary eclipse is not well defined we decided to fix the third-light value at $\ell_3 = 0$. High resolution imaging by @2016AJ....151..159S did not detect any companions to this star, with the quoted upper limit to the relative brightness at I-band being 2.37 magnitudes at 0.25arcsec. This star is listed in the K2 Variable Catalogue as an eclipsing binary with a period of 3.053723d, which agrees well with our estimate of the orbital period. Star 1 has T$_{\rm eff}\approx 6300$K while star 2 is very cool and much smaller than star 1 so we assume that this system consists of a solar-type star and a K- or M-dwarf companion. In this case $q\ll 1$ so the position of the stars in Fig. \[TeffDensity\] does not depend strongly on the assumed value of $q$. From the location of these stars in the T$_{\rm eff}$ – $\rho_{\star}$ plane (Fig. \[TeffDensity\]) we estimate that they are dwarfs stars with masses $M_1\approx 1.2M_{\sun}$ and $M_2\approx 0.5M_{\sun}$. ### EPIC 201705526 = The orbital period shown in Table \[ellcpar1\] was measured from 85,962 WASP photometric measurements obtained over 1148 days using the [hunter]{} algorithm [@2006MNRAS.373..799C]. This value is in fair agreement with the period of 18.120439d given in the K2 Variable Catalogue . include this star in their table of planetary candidates. This appears to be based on the depth and width of the secondary eclipse in the K2 light curve. We speculate that their outlier rejection algorithm may have removed the narrow primary eclipse data from the K2 light curve resulting in the misclassification of this eclipsing binary as a transiting planet candidate. A good fit to the K2 light curve is also possible for solutions with a surface brightness ratio $S_{\rm Kp}\approx$ 7 and $R_2/R_1 \approx 0.9$ but this leads to estimates of the mean stellar densities and effective temperatures that are not plausible. In contrast, the location of these stars in the T$_{\rm eff}$ – $\rho_{\star}$ plane for the parameters we have adopted (Fig. \[TeffDensity\]) suggests that they are dwarf stars with masses $M_1\approx 1.3M_{\sun}$ and $M_2\approx 0.7M_{\sun}$. Both stars appear near or below the zero-age main sequence for solar-metallicity models of stars with these masses for any reasonable choice of the mass ratio, $q$. ### EPIC 201723461 We decided to fix the third-light parameter at the value $\ell_3 = 0$ since the eclipses in this light curve are partial and the secondary eclipse is quite shallow. Even with this assumption, the ratio of the radii is only weakly constrained by the light curve. This star is listed in the K2 Variable Catalogue as an eclipsing binary with a period of 22.713572d, which agrees well with our estimate of the orbital period. Although the plotted position of the cooler star is less dense than the hotter star in Fig. \[TeffDensity\], the uncertainty in the radius ratio is large enough to accommodate solutions where these stars have mean densities as expected for dwarf stars with masses $M\approx 0.7M_{\sun}$. Changing the mass ratio from our assumed value of $q=1$ does not alter this conclusion. [@lrr]{} & &\ Mass \[$M_{\sun}$\] & 1.48 $\pm$ 0.16 & 1.16 $\pm$ 0.13\ Radius \[$R_{\sun}$\] & 2.18 $\pm$ 0.07 & 1.22 $\pm$ 0.04\ $\log g$ \[cms$^{-2}$\] & 3.93 $\pm$ 0.02 & 4.33 $\pm$ 0.02\ T$_{\rm eff}$ \[K\] & 6250 $\pm$ 285 & 6150 $\pm$ 285\ $\log(L/L_{\sun})$ & 0.82 $\pm$ 0.08 & 0.29 $\pm$ 0.09\ M$_{\rm V}$ \[mag\] & 2.7 $\pm$ 0.2 & 4.1 $\pm$ 0.2\ Orbital period \[d\] &\ Mass ratio &\ Distance \[pc\] &\ ### EPIC 202674012 = We downloaded four spectra of this star observed with the FEROS spectrograph from the ESO science archive. We used cross correlation over the wavelength range 400–680nm against a numerical mask from an F0-type template star in iSpec to measure the radial velocities given in Table \[rvtable\]. The full widths at half minimum of the dips in the cross correlation function measured by a simultaneous fit of two Gaussian profiles are 23kms$^{-1}$ and 17kms$^{-1}$ for star 1 and star 2, respectively. The ratio of depths of these dips is 0.41, which is in reasonable agreement with the value of $\ell_{\rm Kp}$ given in Table \[ellcpar2\] if some allowance is made for the different wavelength range covered by these spectra cf. the Kepler band pass. We used [emcee]{} to find the best fit Keplerian orbit to these radial velocity measurements including Gaussian priors on the parameters $f_s$, $f_c$, $T_0$ and $P$ taken from the values shown in Table \[ellcpar1\]. We assumed a single value for the standard error on these radial velocity measurements and included this as a free parameter in the analysis by including the appropriate term in the likelihood function. The semi-amplitudes derived from this fit are $K_1 = 45.3 \pm 2.4$kms$^{-1}$ and $K_2 = 57.8 \pm2.6$kms$^{-1}$, and the standard error for the maximum-likelihood solution was 0.26 kms$^{-1}$. The absolute parameters of the stars derived from these values and the data in Tables \[ellcpar1\] and \[ellcpar2\] are given in Table \[202674012abspar\]. The spectral type is F3(V) [@1982MSS...C03....0H], which implies a mean value of T$_{\rm eff} \approx 6435$K [@2013ApJ...771...40B]. This agrees well with our estimates for T$_{\rm eff,1}$ and T$_{\rm eff,2}$ in Table \[TeffTable\]. There is also good agreement between the measured masses of the stars and their expected masses given their position in Fig. \[TeffDensity\] relative to stellar evolution tracks for solar composition. ### EPIC 202843085 The location of these stars in the T$_{\rm eff}$ – $\rho_{\star}$ plane (Fig. \[TeffDensity\]) suggests that they are a pair of dwarf stars with masses $M\approx 1.4M_{\sun}$ near the end of the main sequence. Star 2 is larger than star 1 so $q>1$ but it is very unlikely that both stars would appear near the MSTO if they have very different masses so we assume the value $q=1$ for purposes of plotting these stars in Fig. \[TeffDensity\]. ### EPIC 203361171 We used different apertures to calculate the light curve of this star for images obtained before and after a change in the spacecraft orientation near BJD 2456936.8. Both apertures include a star approximately 21 arcseconds to the south-west of the main target and 2.4 magnitudes fainter in the G band according to the Gaia DR1 data release . We did not include the value of $\ell_3$ given in Table \[ellcpar2\] in our analysis to estimate the effective temperatures of the stars because we assume that this value is dominated by the star 21 arcseconds to the south-west of the main target whose flux will not be included in the published catalogue photometry. The location of these stars in the T$_{\rm eff}$ – $\rho_{\star}$ plane (Fig. \[TeffDensity\]) suggests that they are a pair of dwarf stars both with masses $M\approx 1.2M_{\sun}$ and near the end of the main sequence. This is a similar case to EPIC 205703649 so we again assume $q=1$ for the purposes of plotting these stars in Fig. \[TeffDensity\]. ### EPIC 203371239 The light curve of this star between eclipses shows a very clear signal due to multi-periodic pulsations (Fig. \[203371239\_ft\]) with frequencies near 0.8 cycles/day and 0.4 cycles/day, and amplitudes of about 1%. These frequencies and amplitudes taken with the effective temperature estimates given in Table \[TeffTable\] suggest that one or both of the stars in this binary is a $\gamma$ Dor-type pulsator [@2011MNRAS.415.3531B]. The location of these stars in the T$_{\rm eff}$ – $\rho_{\star}$ plane (Fig. \[TeffDensity\]) suggests that they are a pair of dwarf stars with masses $M\approx 1.3M_{\sun}$ and $M\approx 1.2M_{\sun}$. This is a similar case to EPIC 202843085 and EPIC 202843085 so we again assume $q=1$ for the purposes of plotting these stars in Fig. \[TeffDensity\]. ![image](Figs/EPIC203371239_ft-eps-converted-to.pdf){width="99.00000%"} ### EPIC 203543668 The photometric aperture we used to construct the K2 light includes the flux from some nearby stars, but this is not enough to account for the value of $\ell_3$ we obtain from the fit to the light curve. The location of these stars in the T$_{\rm eff}$ – $\rho_{\star}$ plane (Fig. \[TeffDensity\]) suggests that the primary is a star similar to the Sun and the secondary is a dwarf star with a mass $M\approx 0.7M_{\sun}$. Both stars appear near the zero-age main sequence for solar-metallicity models of stars with these masses for any reasonable choice of the mass ratio, $q$. ### EPIC 203610780 The parameters we have derived for this binary system from the analysis of the K2 light curve are quite robust because the eclipses are total. Star 2 is much larger and cooler than star 1 so we can assume $q>1$, but the actual value of $q=1.2$ used to plot the stars in Fig. \[TeffDensity\] is quite uncertain. The position of the hotter star below the zero-age main sequence for solar-type stars suggests that this may be a low-metallicity system. This conclusion is not affected by the exact choice of $q$. The complicating factor for this interpretation is the large amount of third light in this system that leads to large uncertainties in the effective temperature estimates. ### EPIC 203636784 The rotation signal in the K2 light curve has an amplitude of about 1.5% at the start of the observing sequence that gradually decreases to an amplitude of about 0.5%. The rotation period is consistent with the assumption of pseudo-synchronous rotation. Star 1 has T$_{\rm eff}\approx 6000$K while star 2 is much cooler and smaller than star 1 so we assume that this system consists of a solar type star and a K- or M-dwarf companion. The position of the stars in Fig. \[TeffDensity\] does not depend strongly on the assumed value of $q$ provided that this value is significantly less than 1. The location of these stars in the T$_{\rm eff}$ – $\rho_{\star}$ plane (Fig. \[TeffDensity\]) suggests that the primary is a star near the main-sequence turn-off with a mass $M\approx 1.1M_{\sun}$ and the secondary is a dwarf star with a mass $M\approx 0.7M_{\sun}$. ### EPIC 203728604 The K2 light curve of this star between the eclipse shows a well defined periodic signal with a period of 2.306d and an amplitude of about 400ppm. The coherence of this signal suggests that this is a pulsation signal rather than rotational modulation due to star spots, perhaps due to $\gamma$ Dor-type pulsations in one of the stars. The location of these stars in the T$_{\rm eff}$ – $\rho_{\star}$ plane (Fig. \[TeffDensity\]) suggests that the primary is a star near the main-sequence turn-off with a mass $M\approx 1.5M_{\sun}$ and the secondary is a star similar to the Sun. This conclusion is not affected by the choice of mass ratio for any value $q\ga 0.8$. The mass ratio is almost certainly has a value $q>1$ since star 1 has a much larger radius than star 2. The periodogram of the data between the eclipses for this star is shown in Fig. \[203728604\_ft\]. ![image](Figs/EPIC203728604_ft-eps-converted-to.pdf){width="99.00000%"} ### EPIC 204407880 The WASP 200-mm data for this star include three nights covering the primary eclipse. To measure the times of mid-eclipse from these data we used a model with the geometric parameters fixed to the values determined from a preliminary fit to the K2 light curve. The three times of mid-eclipse and the surface brightness ratio in the WASP bandpass were free parameters in a fit to the WASP data using [emcee]{} to determine the optimum value of these parameters and their standard errors. The times of mid-eclipse derived using this method were BJD$_{\rm TDB}-2450000=$ 3893.334(2), 4271.385(1), 4649.435(2), where the values in parentheses denote the standard error in the final digit. From a linear fit to these times of mid-eclipse plus the value 2456917.71237(14) from a preliminary fit to the K2 light curve we find an orbital period of $34.36789 \pm 0.00002$d. This period was included as a prior in the analysis of the K2 light curve. ### EPIC 204576757 This star is listed as a planetary candidate system with a period of 23.277669 days by @2016ApJS..222...14V, although the estimated radius of the companion ($\sim 3 R_{\rm Jup}$) is rather large for a planetary-mass object. Three total eclipses due to the transit of the companion are visible in the K2 light curve but there is no clear secondary eclipse visible in these data. This may be because the companion contributes less than about 0.25% of the flux at optical wavelengths, or the orbital eccentricity may be large enough for there to be no secondary eclipse. Given this ambiguity over the configuration of this binary system we have not attempted any further analysis of the K2 light curve. ### EPIC 204748201 Although this is a binary with total eclipses, the secondary eclipse is very shallow so including third light contamination in the analysis results in parameters that have large uncertainties. We decided to fix the third light parameter at $\ell_3 = 0$ for this preliminary characterisation of this system. Star 1 has T$_{\rm eff}\approx 6100$K while star 2 is very cool and much smaller than star 1 so we assume that this system consists of a solar type star and a K-dwarf companion. In this case $q$ must be significantly less than 1, so the position of the stars in Fig. \[TeffDensity\] does not depend strongly on the assumed value of $q$. With these assumptions, the location of the stars in the T$_{\rm eff}$ – $\rho_{\star}$ plane (Fig. \[TeffDensity\]) suggests that they are dwarf stars near the zero-age main sequence with masses $M\approx 1.2M_{\sun}$ and $M\approx 0.5M_{\sun}$. ### EPIC 204760247 = This bright B3V star (V=5.84) is listed in SIMBAD as a Cepheid variable star – this is not correct. @1967ApJS...14..263H found this star to be a variable with a possible period 0.2872days based on 20 observations in each of the U and B bands but note that “because of the extremely small amplitude of the variation …this star must be considered a [tentative]{} $\beta$ Cephei variable.” noted that this star is a double-lined spectroscopic binary with the secondary component “much fainter”. In a later study they estimated a mass ratio for this binary of $0.38\pm0.03$. @1987ApJS...64..487L report 8 radial velocity measurements from which they claimed the first spectroscopic solution for this star with an orbital period near 10 days, but with a very large eccentricity that is not consistent with the data described below. @2002MNRAS.331...45K note that this is a variable star on the basis of the Hipparcos epoch photometry but were not able to classify the type variability. @2011MNRAS.416.2477W used observations from the STEREO mission to correctly identify the variability of this star as being due to eclipses with a period of 9.20days. This star is a member of the Upper Scorpius OB association . ![image](Figs/EPIC204760247_ft-eps-converted-to.pdf){width="99.00000%"} We conducted aperture photometry for this star including the extensive charge overspill region provided in the K2 target pixel file. This provided useful photometry for the interval BJD 2456894.5 to 2456935.5. There is a very clear pulsation signal in the data between the eclipses with an period of 0.908 days and an amplitude of 0.18% (Fig. \[204760247\_ft\]). There is also a periodic signal in these data with a period close to the orbital period and an amplitude of 0.1% that may be due to irradiation of the companion star by the B3V primary star. We have not included this effect in our model of the light curve. The surface brightness ratio from a preliminary light curve solution combined with an estimate for the primary star effective temperature $T_{\rm eff}=18,000$K based on its spectral type implies $T_{\rm eff}\approx 10,000$K for the secondary star. We used these $T_{\rm eff}$ estimates and the tabulation by to estimate the quadratic limb darkening coefficients $(a,b) = (0.11,0.24)$ and $(0.21, 0.29)$ for the primary and secondary, respectively. We assume standard errors of 0.05 on all these coefficients when imposing them as priors in the light curve analysis. We downloaded six spectra of this star observed with the FEROS spectrograph from the ESO science archive. We used cross correlation over the wavelength range 400–680nm against a numerical mask from an A0-type template star in iSpec to measure the radial velocities given in Table \[rvtable\]. The full widths at half minimum (FWHM) of the dips in the cross correlation function (CCF) measured by a simultaneous fit of two Gaussian profiles are 22kms$^{-1}$ and 28kms$^{-1}$ for star 1 and star 2, respectively. A third dip is visible in the CCF with a radial velocity of $-13$kms$^{-1}$ and FWHM of 15kms$^{-1}$ and a strength approximately half that of the peak for star 2. The mismatch between the spectral type of the primary star and the template in this case makes it difficult to interpret the strength of the dip in the CCF for this star – no template is available for spectral type earlier than A0 in the current version of iSpec. We used [emcee]{} to find the best fit Keplerian orbit to these radial velocity measurements assuming a circular orbit ($e=0$). We did not find a satisfactory fit to these data using the values of $T_0$ and $P$ taken from the values shown in Table \[ellcpar1\]. Instead we noted that there is a secondary eclipse visible in the Hipparcos epoch photometry for this star and used this to estimate an orbital period of 9.199724(4) days. With this orbital period imposed as a prior we find that the semi-amplitudes are $K_1 = 62.8 \pm 1.7$kms$^{-1}$ and $K_2 = 136.6 \pm1.4$kms$^{-1}$. The masses of the stars are $M_1 = 5.18\pm 0.15$M$_{\odot}$ and $M_2 = 2.38\pm 0.11$M$_{\odot}$ and their radii are $R_1 = 2.50 \pm 0.04$R$_{\odot}$ and $R_2 = 1.63\pm 0.04$R$_{\odot}$. ### EPIC 204822807 The argument that leads to the conclusion $q\approx 1$ for EPIC 201161715 also applies to this binary system, but in this case it is star 2 that is the cooler and larger star so we assume the value $q= 1.1$ to plot the position of the stars in Fig. \[TeffDensity\]. This system is a bright, totally eclipsing binary system that contains a star on the red giant branch and a star with a mass similar to the Sun near the main-sequence turn-off point, similar to the well-known systems AI Phe and TZ For . This makes this system an attractive target for calibrating stellar models. ### EPIC 204870619 This is a long-period binary in which a sub-giant star with a mass $\approx 1.2\,M_{\odot}$ produces total eclipses of a dwarf star with a mass $\approx 0.8\,M_{\odot}$. The position of the stars in the T$_{\rm eff}$–$\rho_{\star}$ plane lie near the stellar evolution tracks for these masses for any reasonable choice of $q$ so these masses should be quite accurate. It may be challenging to measure precise radial velocities for the dwarf star in this binary since it only contributes about 5% of the flux at optical wavelengths and the system is quite faint. Nevertheless, this is an attractive target for follow-up observations to test stellar models given that, in principle, the masses and radii for these stars can be determined to an accuracy much better than 1% and there do not appear to be any complications in the analysis due to star spots or pulsations. ### EPIC 205020466 We obtained 4 spectra of this star using the High Resolution Spectrograph (HRS) on the Southern African Large Telescope (SALT) [@2014SPIE.9147E..6TC]. We used the medium resolution mode to obtain spectra at a resolving power $R=43,000$ on the blue arm of the instrument and $R=40,000$ on the red arm. The exposure time of 577s on both arms resulted in a signal-to-noise per pixel of approximately 10 on the blue arm and 20 on the red arm. We used spectra reduced automatically using the pipeline developed by @2017ASPC..510..480K for our analysis. We used iSpec to measure the radial velocities for both stars from Gaussian profile fits to the cross-correlation function (CCF) calculated using a numerical mask based on the solar spectrum. The results of these fits are given in Table \[rvtable\]. We did not attempt any further analysis of the HRS spectra because the signal-to-noise is quite low and the reduced spectra are not corrected for the blaze function of the spectrograph. We used [emcee]{} to find the best fit Keplerian orbit to these radial velocity measurements including Gaussian priors on the parameters $f_s$, $f_c$, $T_0$ and $P$ taken from the values shown in Table \[ellcpar1\]. We assumed a single value for the standard error on these radial velocity measurements and included this as a free parameter in the analysis by including the appropriate term in the likelihood function. The semi-amplitudes derived from this fit are $K_1 = 60.5 \pm 1.3$kms$^{-1}$ and $K_2 = 75.5 \pm 1.3 $kms$^{-1}$ and the standard error per observation derived is 2.0kms$^{-1}$. The absolute parameters of the stars derived from these values and the data in Tables \[ellcpar1\] and \[ellcpar2\] are given in Table \[205020466abspar\]. The masses in this table agree well with the values that can be inferred from the location of the stars in the T$_{\rm eff}$ – $\rho_{\star}$ plane (Fig. \[TeffDensity\]) if the errors in T$_{\rm eff}$ are accounted for. [@lrr]{} & &\ Mass \[$M_{\sun}$\] & 1.05 $\pm$ 0.04 & 0.84 $\pm$ 0.04\ Radius \[$R_{\sun}$\] & 0.89 $\pm$ 0.015 & 0.86 $\pm$ 0.015\ $\log g$ \[cms$^{-2}$\] & 4.56 $\pm$ 0.01 & 4.49 $\pm$ 0.015\ T$_{\rm eff}$ \[K\] & 5300 $\pm$ 480 & 5070 $\pm$ 425\ $\log(L/L_{\sun})$ & $-0.25 \pm 0.16$ &$ -0.35 \pm 0.15 $\ M$_{\rm V}$ \[mag\] & 5.5 $\pm$ 0.5 & 5.9 $\pm$ 0.5\ Orbital period \[d\] &\ Mass ratio &\ Distance \[pc\] &\ ### EPIC 205170307 The analysis of this system is complicated by substantial 3rd light contamination ($\ell_3\approx 17$%) but the eclipses are total and well-defined so a robust determination of the system parameters is possible. The eclipsing pair are dwarf stars with masses $\approx 0.65\,M_{\odot}$ and $\approx 1.0\,M_{\odot}$ with the more massive star near the main-sequence turn-off point in Fig. \[TeffDensity\]. The position of the stars in the T$_{\rm eff}$–$\rho_{\star}$ plane lie near the stellar evolution tracks for these masses for any reasonable choice of the mass ratio, $q$. ### EPIC 205546169 Based on the parameters we have derived, the eclipsing pair in this system are both dwarf stars with masses $\approx 1.2\,M_{\odot}$ with one star near the zero-age main sequence and one near the end of the main sequence. This conclusion does not depend on the assumed mass ratio for any reasonable choice of $q$. We assume $q=1.1$ to plot these stars in Fig. \[TeffDensity\] since star 2 is apparently more evolved than star 1. This combination is difficult to reconcile with the very similar effective temperatures for the two stars. The secondary eclipse in this system is very shallow, the eclipses are partial and there may be third-light contamination $\ell_3 \approx 9$% so the parameters we have derived here may be subject to quite large systematic error. Spectroscopic observations to determine more robust estimates for $\ell_3$ and $\ell_{\rm Kp}$ will be very helpful for the analysis of this system. ### EPIC 205703649 The photometric aperture we used for this star is contaminated by nearby stars of comparable brightness to the target star. This is accounted for in the photometric fit by including $\ell_3$ as a free parameter and accounts for the large value of this parameter. We have assumed that this contaminating flux does not affect the catalogue photometry for the target star and so we set $\ell_3=0$ to derive the effective temperature estimates in Table \[TeffTable\]. With this assumption, both stars appear to be dwarf stars near the MSTO with masses $\approx 1.0\,M_{\odot}$. Star 2 is larger than star 1 so $q>1$ seems a reasonable choice. However, it is very unlikely that both stars would appear near the MSTO if they have different masses, therefore we assume the value $q=1$ for the purposes of plotting these stars in Fig. \[TeffDensity\]. ### EPIC 205919993 = LP 819-72 was identified as an eclipsing binary using data from the WASP project prior to the start of the K2 mission. The primary eclipse was never observed with WASP because the orbital period is so close to exactly 11 days. As such, it was misclassified as an eclipsing binary with a transiting low mass companion (“EBLM”) with an orbital period of 3.666 days. We submitted LP 819-72 to the guest observer program on K2 and also obtained 33 spectra of this system with the fibre-echelle spectrograph on the CTIO 1.5-m telescope operated by the SMARTS Consortium. The spectra were typically observed in groups of three with an exposure time of 900s plus an accompanying arc spectrum for wavelength calibration. We extracted a single order from these echelle spectra using the optimal extraction routines and wavelength calibration routines [pamela]{} and [molly]{} [@1989PASP..101.1032M]. The spectral order selected covers the wavelength range 660.9–647.7nm and the resolving power of the instrument is $R\approx 37,000$. The signal-to-noise ratio per pixel at the centre of the order is typically ${\rm S/N} \approx 20$. We used iSpec to measure the radial velocities for both stars in those spectra where the lines from the two stars are clearly resolved. The radial velocities were measured using Gaussian profile fits to the cross-correlation function (CCF) calculated using a numerical mask based on a K5 spectral-type template. The results of these fits are given in Table \[rvtable\]. The individual spectra observed on the night JD 2455477 were of lower signal-to-noise than other spectra so we co-added these spectra for analysis. We also co-added the group of three spectra with the highest S/N in order to look for additional dips in the CCF. No such dips were detected so we estimate that the contribution from any third star in the system does not exceed about 10% at these wavelengths, assuming that any such star is a slowly rotating star with a late-type spectrum. For simplicity in the analysis below we fixed the third light parameter $\ell_3=0$ for our analysis of the K2 light curve. We also imposed a prior on the flux ratio $\ell_{\rm Kp}$ from the ratio of the depths of dips in the CCF. The mean and standard error in the mean in this ratio is $1.28\pm0.03$ but we use a Gaussian prior on $\ell_{\rm Kp}$ with mean 1.28 and standard deviation 0.05 to allow for some uncertainty in converting the depth of the dip in the CCF to a flux in the Kepler bandpass. This information from the spectroscopy is extremely useful for the analysis of the K2 light curve because without these priors on $\ell_3$ and the flux ratio the best-fit solutions tend to imply a flux ratio for the stars that is inconsistent with the radius ratio. We used [emcee]{} to find the best fit Keplerian orbit to the radial velocity measurement in Table \[rvtable\] including Gaussian priors on the parameters $f_s$, $f_c$, $T_0$ and $P$ taken from the values shown in Table \[ellcpar1\]. We find that the semi-amplitudes are $K_1 = 54.1 \pm 0.6$kms$^{-1}$ and $K_2 = 49.2 \pm 0.6 $kms$^{-1}$. The absolute parameters of the stars derived from these values and the data in Tables \[ellcpar1\] and \[ellcpar2\] are given in Table \[205919993abspar\]. [@lrr]{} & &\ Mass \[$M_{\sun}$\] & 0.59 $\pm$ 0.02 & 0.65 $\pm$ 0.02\ Radius \[$R_{\sun}$\] & 0.60 $\pm$ 0.01 & 0.58 $\pm$ 0.01\ $\log g$ \[cms$^{-2}$\] & 4.65 $\pm$ 0.02 & 4.72 $\pm$ 0.02\ T$_{\rm eff}$ \[K\] & 4025 $\pm$ 60 & 4230 $\pm$ 60\ $\log(L/L_{\sun})$ & $-1.07 \pm 0.03$ &$ -1.01 \pm 0.03 $\ M$_{\rm V}$ \[mag\] & 8.38 $\pm$ 0.13 & 8.05 $\pm$ 0.12\ Orbital period \[d\] &\ Mass ratio &\ Distance \[pc\] &\ The distance to this system based on the parallax measurement from Gaia DR1 is $44.9\pm0.5$pc. The distance to this system based on the 2MASS K$_{\rm s}$-band magnitude transformed to Johnson K-band and the surface brightness – T$_{\rm eff}$ relation by is $42.1\pm 1.0$pc, which is a fair agreement with the Gaia DR1 estimate, particularly if the suspected systematic error of about 0.22mas in Gaia DR1 parallax values for stars near the ecliptic is taken into account [@2016arXiv160905390S]. It remains to be seen whether a more accurate estimate of $\ell_3$ will lead to better agreement between these distance estimates. This high proper motion star has a spectral type of K5V [@1986AJ.....92..139S]. This implies an effective temperature T$_{\rm eff} \approx 4436$K, which agrees reasonably well with the average of our estimates of T$_{\rm eff}$ for the two stars in Table \[TeffTable\]. ### EPIC 205982900 = A detailed study of this eclipsing binary has been presented by based on [uvby]{} light curves by and spectroscopic orbits for both components from . The times of primary and secondary eclipse show apsidal motion with a period of approximately 6000 years [@2014CoSka..43..419V]. The projected equatorial rotational velocities of the stars suggest that they both rotate at about half the rate expected assuming pseudo-synchronisation of the rotation with the orbital angular velocity at periastron. According to linear ephemeris by @2004AcA....54..207K the 2MASS observations of this star were obtained during primary eclipse so we did not include these data in the analysis to determine the T$_{\rm eff}$ estimates in Table \[TeffTable\]. These T$_{\rm eff}$ estimates are about 250K cooler than the values from based on the dereddened $(b-y)_0$ colour indices, which is significant at the 2-$\sigma$ level. This is a consequence of the larger reddening to this system derived by Clausen from Strömgren photometry compared to the value we obtained from broad-band photometry together with the prior on E($\rm B-\rm V$) from reddening maps. The position of these stars in the T$_{\rm eff}$ – $\rho_{\star}$ plane (Fig. \[TeffDensity\]) suggest that the higher effective temperature estimate is more reliable given that the masses of these stars are known to be $M\approx 1.4M_{\odot}$. The K2 light curve of BW Aqr shows an ellipsoidal effect with a semi-amplitude of about 0.1%. To account for this, we set the period of the quasi-periodic kernel used to account for the intrinsic variability of the star in the detrending process to half the orbital period. We then modeled this light curve including the ellipsoidal effect, i.e., we did not divide-out this trend. To account for the ellipsoidal effect and the resulting gravity darkening we used the Roche potential to calculate the shape of the ellipsoids used to model the two stars assuming a mass ratio $q=1$. The gravity darkening exponents were set to $y_1=0.267$ and $y_2=0.280$ for star 1 and 2, respectively, these values being interpolated from the tabulation by . The residuals from the our best-fit model light curve show structure at a level of about 500 ppm. The periodogram of these residuals shows peaks at various harmonics of the orbital period. We used light curves derived using the [everest]{} algorithm [@2016AJ....152..100L] to check that these features in the light curve are not a by-product of our data analysis method. We also checked the [everest]{} light curves for 3 other stars of similar brightness to BW Aqr and observed during the same K2 campaign and using the same channel of the Kepler instrument. The periodograms of the [everest]{} light curves for BW Aqr and one of these comparison stars are shown in Fig. \[BWAqr\_ft\]. We also analysed the [everest]{} light curve for EPIC 201576812 (TYC 272-458-1). For both BW Aqr and TYC 272-458-1 we excluded data obtained during eclipse from the calculation of the periodogram. The rotational modulation of TYC 272-458-1 combined with this masking of the eclipses results in closely separated peaks in the periodogram at harmonics of the orbital frequency up to approximately the 25$^{\rm th}$ harmonic. In contrast, the periodogram of BW Aqr shows a well defined sequence of single peaks up to the 50$^{\rm th}$ harmonic of the orbital frequency, at least. A likely interpretation of these frequencies is that they are tidally induced pulsations, similar to those seen in the non-eclipsing binary system KOI-54 [@2012MNRAS.420.3126F; @2012MNRAS.421..983B]. The strongest frequencies are at the 2$^{\rm nd}$ and 3$^{\rm rd}$ harmonic of the orbital frequency, which agrees well with the expectation for a binary system with an eccentricity e$\approx 0.18$ . The detection of tidally induced pulsations in BW Aqr open up the possibility of testing methods to derive stellar parameters using asteroseismology for non-eclipsing binary stars such as KOI-54, and to investigate the role of these pulsations in the circularisation and synchronisation of BW Aqr’s orbit. Such studies will be aided by the availability of K2 data with a temporal sampling of 58.8s (“SC” data) for this binary system. ![image](Figs/BWAqr_ft.png){width="99.00000%"} [@lrr]{} & &\ Mass \[$M_{\sun}$\] & 1.38 $\pm$ 0.02 & 1.48 $\pm$ 0.02\ Radius \[$R_{\sun}$\] & 1.732 $\pm$ 0.008 & 2.068 $\pm$ 0.009\ $\log g$ \[cms$^{-2}$\] & 4.100 $\pm$ 0.003 & 3.977 $\pm$ 0.003\ T$_{\rm eff}$ \[K\] & 6450 $\pm$ 100 & 6350 $\pm$ 100\ $\log(L/L_{\sun})$ & 0.67 $\pm$ 0.03 & 0.80 $\pm$ 0.03\ M$_{\rm V}$ \[mag\] & 3.07 $\pm$ 0.07 & 2.76 $\pm$ 0.07\ Orbital period \[d\] &\ Mass ratio &\ Eccentricity &\ Distance \[pc\] &\ ### EPIC 206066862 = We decided to fix the third light parameter at the value $\ell_3 = 0$ since the eclipses in this light curve are partial and the secondary eclipse is quite shallow. quote an orbital period P=11.08716751d for this binary, which is in good agreement with our estimate. These stars appear near the evolutionary tracks for masses of $0.8\,M_{\sun}$ and $1.2\,M_{\sun}$ for any reasonable choice of the mass ratio, so we assume $q\approx 0.7$. ### EPIC 206066909 The K2 light curve between the eclipses shows variability on timescales of a few days with an amplitude of about 0.1%. We were not able to identify a clear period for this variability and the interpretation of this signal is complicated by a substantial third-light contribution to the light curve in this multiple star system. The position of the stars in the T$_{\rm eff}$–$\rho_{\star}$ plane lie near the stellar evolution tracks for masses $0.7\,M_{\sun}$ and $1.3\,M_{\sun}$ for any reasonable choice of the mass ratio so we assume $q\approx 0.5$. ### EPIC 206075677 There are two partial secondary eclipses visible in the detrended K2 light curve with depths of 0.4% and 1.5% following division by the trends in the light curve due to star spot modulation. These trends show an amplitude of 1.5% and at least three well defined peaks in the periodogram at the periods listed in Table \[lcinfo\]. Given the complexity of this system and the lack of unambiguous information in the light curve, we decided not to attempt a model fit to this light curve. ### EPIC 206084435 There is only one primary eclipse visible in the K2 light curve for this very long-period binary ($P\approx 48$d) so the period is determined from the two shallow secondary eclipses. The primary star is a main-sequence star with a mass $M\approx 1.1M_{\odot}$ and the companion is a dwarf star with a mass $M\approx 0.65M_{\odot}$. The position of the stars in the T$_{\rm eff}$–$\rho_{\star}$ plane lie near the stellar evolution tracks for these masses for any reasonable choice of the mass ratio, $q$. ### EPIC 206109641 The K2 light curve of this star shows two eclipses with unequal depths separated by 7.68 days. Only one eclipse of each depth in visible within the 69.16 day span of the K2 observations. From the K2 data alone it is only possible to establish a lower limit to the orbital period of 41.77 days. Data from the WASP archive includes one night of 200-mm data covering the minimum of the secondary eclipse at HJD 2455141.285, another night of 200-mm data showing the start of the ingress to primary eclipse at HJD 2454632.6 and one night of 85-mm data covering the egress to the secondary eclipse at HJD 2456144.43. The only period consistent with these observations plus the lack of visible eclipses on other nights for which we have WASP data is $P\approx 62.6$d. We included the WASP photometry from the nights listed above plus data from a few additional nights to set the out-of-eclipse level in our analysis of the K2 light curve. The additional parameters required in the fit were the surface brightness ratio for the 200-mm WASP data, $S_{\rm 200}$, the surface brightness ratio for the 85-mm WASP data, $S_{\rm 85}$, the zero-points of the 85-mm and 200-mm data, and the standard deviation of the residuals for the 85-mm and 200-mm data sets, $\sigma_{\rm 200}$ and $\sigma_{\rm 85}$, respectively. Given the quality of the WASP data we decided to use the same limb darkening coefficients and values for third light for these data as for the K2 data rather than adding even more free parameters to the fit. The best-fit model light curves to the K2 and WASP data are shown in Fig. \[206109641\_lcfit\]. The best-fit values of the additional parameters were found to be $S_{\rm 200}=0.57 \pm0.01 $, $S_{\rm 85}=0.42\pm0.02$, $\sigma_{\rm 200}=0.016$ and $\sigma_{\rm 85}=0.077$. Both stars lie near the position of the Sun in the T$_{\rm eff}$–$\rho_{\star}$ diagram for any reasonable choice of mass ratio but the relative position of the stars is inconsistent with star 1 being larger and therefore more evolved unless $q\approx 0.9$. ![image](Figs/206109641_lcfit.png){width="99.00000%"} ### EPIC 206212261 The occultation of a dwarf star with a mass $M\approx 0.6M_{\odot}$ by a 1.0-$M_{\odot}$ star near the MSTO results in well-defined total eclipses in the K2 light curve of this system, so the parameters are determined to good precision despite the extreme flux ratio in this binary $\ell_{\rm Kp}\approx 1.7$%. This extreme flux ratio may make it challenging to obtain spectroscopic observations to determine the masses of these stars. Star 1 has T$_{\rm eff}\approx 5400$K while star 2 is cooler and much smaller than star 1 so we assume that this system consists of a solar type star and a K-dwarf companion. In this case $q$ must be significantly less than 1, so the position of the stars in Fig. \[TeffDensity\] does not depend strongly on the assumed value of $q$. ### EPIC 206241558 The K2 light curve of this star shows two eclipses with unequal depths separated by 12.32 days. Only one eclipse of each depth in visible within the 69.16 day span of the K2 observations. From the K2 data alone it is only possible to establish a lower limit to the orbital period of 53.1 days. Data from the WASP archive includes one night of 200-mm data covering the minimum of the primary eclipse at HJD 2455012.63 and another night of 200-mm data covering the egress to the secondary eclipse at HJD 2454685.4. The only period consistent with these observations plus the lack of visible eclipses on other nights for which we have WASP data is $P\approx 56.6$d. We included the WASP photometry from the nights listed above plus data from one additional night to set the out-of-eclipse level in our analysis of the K2 light curve. Given the quality of the WASP data we decided to use the same limb darkening coefficients and values for third light for these data as for the K2 data. We made several attempts to find a combination of starting parameters for the [emcee]{} analysis that lead to best-fit solutions with both $k < 1$ and $S_{\rm Kp} <1$ or vice versa, but always found that the Markov chains converged on solutions with $k < 1$ and $S_{\rm Kp} > 1$ or vice versa, depending on the initial parameters and whether or not we fixed $\ell_3 = 0$. We did not find any combination of parameters for which the two stars appear near the same isochrone in the effective temperature – mean stellar density plane so our estimate for the mass ratio $q\approx 1.0$ is very uncertain in this case. One possibility is that this star is the result of a merger between two stars in a triple system induced by Kozai-Lidov cycles in a triple star system [@2018arXiv180205854S]. The best-fit model light curves to the K2 and WASP data are shown in Fig. \[206241558\_lcfit\]. The surface brightness ratio in the 200-mm WASP data we found to be $S_{\rm 200}=1.5 \pm0.1 $ and the standard deviation for the residuals is $\sigma_{\rm 200}=0.023$. ![image](Figs/206241558_lcfit.png){width="99.00000%"} ### EPIC 206253908 The K2 light curve of this star shows only one eclipse, from which it is possible to establish a lower limit to the orbital period of 62.7 days. Eclipses are also seen in the WASP archive photometry at HJD 2454758.41 and 2455085.61. We estimate the period of this binary to be 65.45d. As there is no secondary eclipse visible in the K2 or WASP data we did not attempt any further analysis of this star. ### EPIC 206288770 The occultation of a dwarf star with a mass $M\approx 0.6\,M_{\odot}$ by a 1.2-$M_{\odot}$ main-sequence star results in well-defined total eclipses in the K2 light curve of this system, so the parameters are determined to good precision despite the extreme flux ratio in this binary $\ell_{\rm Kp}\approx 1.4$%. These mass estimates are quite robust because the position of the stars in the T$_{\rm eff}$–$\rho_{\star}$ plane lie near the stellar evolution tracks for these masses for any reasonable choice of $q$. The extreme flux ratio in this binary may make it challenging to obtain spectroscopic observations to determine the masses of these stars. ### EPIC 206433263 This eclipsing binary is a favourable target for detailed study of a star similar to the Sun. The eclipses of the star with a mass $M\approx 1.0\,M_{\odot}$ are total and the companion with a mass $M\approx 1.2\,M_{\odot}$ is near the MSTO. These mass estimates are quite robust because the position of the stars in Fig. \[TeffDensity\] lie near the stellar evolution tracks for these masses for any reasonable choice of $q$. The optical flux ratio is favourable for spectroscopic follow-up to determine accurate masses for both stars and the star is moderately bright. The mass of the star near the MSTO will put a tight constraint on the age of the system and there do not appear to be any complications due to star spots or third light contamination. Summary and conclusions ======================= We have used light curves from Kepler K2 to identify 42 long-period eclipsing binary systems ($P\ga 5.5$d) with narrow eclipses and little or no ellipsoidal effect, i.e., well-detached binaries with very weak tidal interaction between the stars. This includes systems with periods $P\ga 60$d for which the orbital period cannot be established from the K2 data alone (EPIC 206109641, EPIC 206241558 and EPIC 206253908). In these cases we have used data from the WASP project data archive to establish the orbital period of the binary system. For 40 targets we have determined the geometry of the binary system (fractional radii, inclination, eccentricity, etc.) from the analysis of the Kepler K2 light curve using the `ellc` eclipsing binary star model. For 38 of these systems we also estimate the effective temperature of the stars from an analysis of the observed apparent magnitudes and other data for the system. For these 38 systems we are able to estimate the mass and evolutionary state of the stars by comparing their mean stellar densities and effective temperatures to stellar models. They typically contain main-sequence or sub-giant stars with masses from 0.6$M_{\odot}$ to 1.4$M_{\odot}$, with sub-giant and giant companions being more common among the longer-period systems. We have used new radial velocity measurements to make preliminary estimates of the mass, radius and luminosity of the stars in 3 systems (EPIC 202674012, EPIC 205020466, EPIC 205919993). We have also re-calculated these absolute parameters for two systems with spectroscopic orbits that have been previously studied using light curves from ground-based instruments (EPIC 201576812 = FM Leo and EPIC 205982900 = BW Aqr). We have also estimated the masses and radii of the eclipsing stars in the early-type triple system EPIC 204760247 = HD 142883. We confirm the presence of variability between the eclipses in HD 142883 due to $\beta$ Cephei-type pulsations. Variability in the light curve between the eclipses due to $\gamma$ Doradus-type pulsations is seen in EPIC 203371239 and perhaps also EPIC 203728604. In BW Aqr we find variability due to pulsations which we suspect are induced by dissipation of tidal forces in this eccentric binary. Variability due to magnetic activity is seen in several systems and has been used to measure the rotation periods of one or both stars in 13 cases. Kepler K2 provides almost continuous observations for each campaign field for up to 80 days. This makes it possible to identify and characterise long-period eclipsing binary stars that are very hard to find and study using ground-based observations. The high quality of the photometry also makes it possible to identify features in the light curve such as low-amplitude rotational modulation due to star spots and pulsations that are also very hard to observe from the ground. The high quality of the photometry also enables the geometry of these eclipsing binaries to be measured to very high precision, particularly for those that show total eclipses. Additional work is needed to establish the accuracy of the parameters for eclipsing binary stars that can be derived using Kepler K2 photometry. The majority of the eclipsing binary systems we have identified in the Kepler K2 data with long orbital periods ($P \goa 10$d) contain dwarf or sub-giant stars. The high quality of the K2 photometry makes it possible to identify binaries where one star has a much lower mass than the other. These are useful systems for testing stellar evolution models because free-parameters in the models such as the mixing length parameter will affect the two stars in the binary in different ways. It can be challenging to obtain accurate mass estimates for the stars in such binaries because they often have extreme luminosity ratios as optical/infra-red wavelengths, but there are now several high-resolution spectrographs available on large telescopes that should make detailed characterisation for some of these systems feasible. We thank the anonymous referee for their careful reading of the manuscript and useful comments that have improved the quality of the paper. This paper includes data collected by the K2 mission. Funding for the K2 mission is provided by the NASA Science Mission directorate. This research has made use of the SIMBAD database, operated at CDS, Strasbourg, France. Funding for SDSS-III has been provided by the Alfred P. Sloan Foundation, the Participating Institutions, the National Science Foundation, and the U.S. Department of Energy Office of Science. The SDSS-III web site is http://www.sdss3.org/. SDSS-III is managed by the Astrophysical Research Consortium for the Participating Institutions of the SDSS-III Collaboration including the University of Arizona, the Brazilian Participation Group, Brookhaven National Laboratory, Carnegie Mellon University, University of Florida, the French Participation Group, the German Participation Group, Harvard University, the Instituto de Astrofisica de Canarias, the Michigan State/Notre Dame/JINA Participation Group, Johns Hopkins University, Lawrence Berkeley National Laboratory, Max Planck Institute for Astrophysics, Max Planck Institute for Extraterrestrial Physics, New Mexico State University, New York University, Ohio State University, Pennsylvania State University, University of Portsmouth, Princeton University, the Spanish Participation Group, University of Tokyo, University of Utah, Vanderbilt University, University of Virginia, University of Washington, and Yale University. Some of the observations reported in this paper were obtained with the Southern African Large Telescope (SALT) under program 2016-2-SCI-001 (PI: Maxted). Based on observations made with ESO Telescopes at the La Silla Paranal Observatory under programme IDs 70.D-0433(A), 091.C-0713(A), 089.D-0097(B) and 091.D-0145(B). We thank Dr Adrian Barker for sharing his expertise in tidally induced stellar pulsations with us. [^1]: Based on observations made with the Southern African Large Telescope (SALT) [^2]: Based on observations collected at the European Organisation for Astronomical Research in the Southern Hemisphere under ESO programmes 073.C-0337(A), 089.D-0097(B), 091.C-0713(A), 091.D-0145(B), 094.A-9029(R), 178.D-0361(B), 178.D-0361(F) [^3]: <http://archive.stsci.edu/missions/hlsp/k2sff> [^4]: <https://pypi.python.org/ellc> [^5]: [www.astro.keele.ac.uk/\~jkt/codes/jktebop.html](www.astro.keele.ac.uk/~jkt/codes/jktebop.html) [^6]: [www.astro.keele.ac.uk/jkt/codes/jktabsdim.html](www.astro.keele.ac.uk/jkt/codes/jktabsdim.html)	{ "pile_set_name": "ArXiv" }
--- abstract: \| The production rate of the X(5568) measured by D0 collaboration is quite large and difficult to be understood by various general hadronization mechanism. We propose the inclusive production formulation for the cross section, and predict the distributions and production rates of X(5568) at LHC energies, which are crucial information for the relevant measurements. We also suspect $D^{\pm}_s \pi^{\pm}$ state can be copiously produced and observed. Keywords: [multi-quark hadron, large production rate]{}\ author: - \| Yi Jin$^1$, Shi-Yuan Li$^2$ and Shu-Qing Li$^2$\ [$^1$ School of Physics and Technology, University of Jinan, Jinan 250022, P. R. China]{}\ [$^2$ School of Physics, Shandong University, Jinan 250100, P. R. China]{} title: 'New $B^0_s \pi^{\pm}$ and $D^{\pm}_s \pi^{\pm}$ states in high energy multi-production process' --- Recent years one witnesses the observation of exotic multi-quark hadrons [@bela; @belb; @cdfa; @belc; @lhcba]. The flavour number presented in the exotic hadron exceeding three is very crucial in clearly identifying the valence quark number. For a hadron with non zero quantum numbers as bottom or charm besides the isospin ($I_3$) and strangeness, one has no other choice than to introduce more than three quarks carrying them. Exotic multi-quark hadron with only ‘light’ (u, d, s) quarks might have produced in various processes, but it is not easy to confirm the valence quark number. [Light]{} Pentaquark is special, but yet of no confirmed evidence of its production. Another significant fact should also be paid great attention on is that the confirmed multi-quark states were all produced from heavier hadron decays. However, this situation may be changed: The D0 collaboration just found a new $B^0_s \pi^{\pm}$ state, X(5568) [@D0:2016mwd], with [four flavours]{} in this hadron, —if it is really a particle. This will be quite remarkable, as the first solid evidence of multi-quark state [directly produced in the multi-production process of high energy collision]{}, rather than from hadron decays. The bottom flavour here is very decisive. Because of the bottom flavour and the mass, it is hardly possible produced from decay of a heavier hadron (For the charm sector, however, one has to distinguish prompt ones from those from bottom decay). This observation of the X(5568) [@D0:2016mwd] surely turns on the significant new page in the study of the multi-quark states. Now one can study their [production mechanism]{} which is tightly related with the inner structure of the relevant hadrons [@long1; @former], [in multi-production processes]{} at high energy colliders like Tevatron and LHC. It is noticeable that in the high energy multi-production processes, the space-time evolution of the quark system and the produced hadron system are quite different from that of the decay processes. Besides the different partonic processes and kinematic phase space distributions, the final state interaction (if existing) is also different. So comparing the production of the ‘exotic hadron’ in the decay and multi-production processes can eliminate the misleading information possibly rendered by, e.g., phase space effects and final state interactions, etc. As a matter of fact, the measurement [@D0:2016mwd] has provided the important information of the production. The ratio $\rho$ of the yield of $X(5568)$ to the yield of the $B_s^0$ meson in two kinematic ranges, 10 $< p_T (B_s^0) <$ 15 GeV/$c$ and 15 $< p_T (B_s^0) <$ 30 GeV/$c$, is measured. The results for $\rho$ are (9.1 $\pm$ 2.6 $\pm$ 1.6)$\%$ and (8.2 $\pm$ 2.7 $\pm$ 1.6)$\%$, respectively, with an average of (8.6 $\pm$ 1.9 $\pm$ 1.4)$\%$. Here we assume $B^0_s \pi^{\pm}$ is the dominant decay mode of this new state. This large production rate itself first of all excludes the possibility of decay from heavier particles like $B_c$, which are difficult to produce. In this paper we investigate the production mechanism of this new $B^0_s \pi^{\pm}$ state, taking it as a particle, say, X(5568). We emphasize that its production rate is quite large, and difficult to be understood by various general hadronization models. So we propose the inclusive resonance production formulation to calculate its cross sections for various collision processes as well as energies, which provides useful information for the relevant detections, e.g., those at LHC. We demonstrate that the transverse and longitudinal momentum distributions, together with the special property of the decay channel are important for the set of the triggers and data selection requirements. When just the bottom quark is replaced by the charm quark in X(5568), assuming the structure hence the production mechanism for the corresponding four-quark charm hadron do not change, the production of the ‘new $D^{\pm}_s \pi^{\pm}$ state’ is studied. As mentioned above, the production mechanism is tightly related with the inner structure of the relevant hadron [@long1; @former]. Our analysis starts from assuming X(5568) as four quark state and directly produced together with other general mesons and baryons, employing various hadronization mechanism [@Jin:2010wg]. 1\) `String fragmentation model `. The bottom quark can only be produced by hard scattering between partons. In the hadronization process it links with another antiquark $\bar q $ by a string. Then this string fragments into several hadrons, among which are the bottom hadrons, e.g., $B^0_s$. It is well known that the relative probabilities of the creation of light flavours $u,d,s$ from vacuum are $1:1:\lambda$, with $\lambda \sim 0.3$ [@Agashe:2014kda; @Jin:2010wg]. These values can be extracted by looking into the production ratios, e.g., $B_s/B_0 \sim 0.3$ in various high energy multi-production processes. For the investigation here, we also need the production of diquark. Since we only want to estimate the production rate, not concentrating on correlations, we adopt the diquark mechanism rather than the popcorn mechanism for simplicity. And this will be consistent with the X(5568) structure when considered as diquark antidiquark pair. From the production ratios such as $\Lambda_b/B_s \sim 1/2$, we take it as $\lambda/2$. However, if the diquark has a strange flavour, e.g., $us$, a further factor $\lambda$ is introduced. Here is the relative probabilities: $$\begin{aligned} ~& u:d:s:ud:uu:dd:us:ds:ss \\ \nonumber = & 1:1:\lambda:\frac{\lambda}{2}:\frac{\lambda}{2}:\frac{\lambda}{2}:\frac{\lambda^2}{2}: \frac{\lambda^2}{2}:\frac{\lambda^3}{2}. \label{rostring}\end{aligned}$$ In string model, quarks and diquarks are created from vacuum in pair, so here ‘$u$’ means ‘$u \bar u$’, ‘$ud$’ means ‘$ud \overline{ud}$’, etc. These relative ratios are consistent with the production rates of $\Xi $, $\Sigma $, non-strange baryons, as well as their relative production ratios to the relevant mesons in a jet [@Agashe:2014kda]. On the other hand, the production of X(5568) needs a special string: The $bu$ or $bd$ diquark as one of the end of the string connecting another quark as the other end [@Jin:2013bra; @Jin:2014nva; @Jin:2015mla]. Needless to say, this kind of colour connection is suppressed. We set an ansatz that this special kind of string $bq -q$, relative to the usual $b-\bar q$ one, receives a suppression factor $\lambda/2$. This is reasonable since when the strings are ‘short’ enough and they just correspond to a single particle, the suppression factor just gives the correct relative production rate. Hence we estimate the production ratio $X(5568)/B^0_s$ in the string model as: $$P_X= \frac{2 \times \lambda/2}{D_1} \times \frac{\lambda^2/2}{D_2}.$$ Here $D_1=1+1+\lambda+\lambda/2+\lambda/2+\lambda^2/2$, and $D_2=1+1+\lambda+\lambda/2+\lambda/2+\lambda/2+\lambda^2/2+\lambda^2/2+\lambda^3/2$. $$P_{Bs}= \frac{1+1+\lambda}{D_1} \times \frac{\lambda}{D_2} ;$$ $$\rho_{string}= \frac{P_X}{P_{Bs}} \sim 2 \%.$$ This is much smaller than the average $8.6 \%$. Even, we emphasize that the decreasing nature with respect to the increasing of the transverse momentum indicated by the experiment (we will clarify that it must be) of the above relative ratio $\rho$ renders that the total production ratio is larger than $8.6 \%$. Furthermore, we ignore the fragmentation into several kinds of hadrons with similar flavour but different spins, which will even decrease the ratio. To get a higher production rate, one needs to tune the relative ratios in Eq. (\[rostring\]), i.e., to increase the relative ratio of the diquarks containing strange quark(s). However, similar problems as in other models [@lsy] to give correct production rates of the strange baryons will be raised. 2\) `cluster model and final state interaction `. In cluster model, we have to introduce a free parameter to describe the probability the cluster with mass large enough to decay into X(5568). The value of the parameter can be obtained by fitting the data. So it is not very meaningful to say it is large or small. By the help of the recent work [@fi], we find that if this peak is not a real particle but some kind of final state interaction effect, the observed ‘rate’ is also very small. The reason is that this kind of peak is quite sensitive to the cluster (refered as A in [@fi]) mass. But the probability for the mass of the cluster around the proper value is very small. So explaining the large ‘production rate’ needs fine tuning by the final state interaction models. 3\) `Combination model`. The key thing is how the light flavour produced and combined with the bottom quark, to become X(5568). The probability of creation of each flavour is as the above, only we do not need to introduce the diquark. So it is the $b$ quark combined with $\bar s u \bar d$ or $\bar s d \bar u$ for X(5568). If we assume that the total probability of any three light (anti)quarks combined with $ b$ as 1, then $$P_X=\frac{\lambda}{D} \frac{1}{D} \frac{1}{D} \times 2.$$ Here $D=2(1+1+\lambda)$. If we assume that $b$ combines with any other antiquark with probability 1, then $$P_{Bs}=\frac{\lambda}{(D/2)},$$ $$\rho^{max}_{combination}= \frac{P_X}{P_{Bs}} \sim 5 \%.$$ Here we write the superscript ‘$^{max}$’ to address the fact: We did not take into account that the above two cases, $b$ combining with an antiquark or $b$ combining with three quark cluster $q \bar q' \bar q''$ excludes each other, and there should be other cases, e.g., $b$ combining with $qq$ to form a baryon. All these possibilities add up to 1. This reflects the unitarity of the combination process [@long1; @lsy], i.e., one quark can combine with many other quarks or clusters, according to some law, but the total probability must be 1, since the quark is confined and has to go into a hadron. For the calculation of $\rho$ here, we can introduce a relative probability $\zeta = (b ~ combines ~ with~any ~q \bar q' \bar q'')/ (b~ combines ~with ~ any ~ \bar q)$. In general $\zeta$ is smaller than 1, according to the fact that four quark state is not copiously found in hadronization. Hence $\rho_{combination}=\rho^{max}_{combination} \times \zeta$, smaller than $5\%$. The above analysis is dependent on the SU(3) flavour symmetry breaking parameter $\lambda$, with different sensitivities for string model and combination model, respectively. In concrete, the result of the combination model is less sensitive to $\lambda$, our up limit varies from about $3 \%$ to $6 \%$ for $\lambda $ from 1 to 0. The result of the string model is much more sensitive, varies from 0 to about $16 \%$ for $\lambda $ from 0 to 1. Especially, if we want to get $8\%$ in the string model, we need to take $\lambda $ to be 0.65. However, $\lambda$ is a very steady value, and corresponds to the physics of producing a strange quark pair from the vacuum by tunneling effect in the hadronization process, independent from colliding particles (except heavy nuclei) and energies. It is taken as 0.3 in Pythia [@Pythia], and confirmed by experiments up to LHC energies [@Abelev:2012tca]. Though the above results can be considered consistent with data at the order of magnitude level, the tension is obvious. Since no other plausible parameters to tune, one has difficulty to raise to a larger production ratio (if confirmed by experiment) in these models. On the other hand, the combination model has provided a clue to improve the description on the large production ratio. If there is a special large $\zeta_{X}$ which is only applicable to this special X(5568) production, it can provide an enhancement factor. In the combination model, if it can take the value around 2, then we can get the experimental result. In the mean time, for the measurement [@D0:2016mwd], $B_s K$ and $B_d \pi$ are searched as cross check and no signal of new state found [@epaps]. It seems another evidence implying that the X(5568) is a very special structure. Its production may be quite unique, different from the other particles produced, though QCD is flavour blind. So this plays another support for us to deal it independently, ignoring the constrains from the unitarity as well as the general hadronization mechanism applied to the general hadrons. For other cases, the $\zeta$ is quite small, as the ansatz for the string model, a value of order of magnitude 10 per cent, so even the other combination can lead to a 4-quark state but no observable signal [@long1]. The above idea of special $\zeta_X$ can be more systematically realized in analogy of the inclusive resonance production framework, as that for the quarkonium. The key point is to describe its production in two steps, whatsoever taking it as a bound state of hadrons or 2-quark clusters, or even, two pieces of strings. And since this special production mechanism, the probability is not added with others to exhaust the unitarity constrain [@long1; @lsy; @Wang:2016vxa] mentioned above, besides not applied to other light flavours. As [@long1], we start from the amplitude $$%\label{startpoint} %\begin{array}{l l l} % A(P) = <X(p \bar p), X\|\hat{T}>\|J/\Psi> \frac{\sqrt{2}}{\sqrt{M}} A(P) = <H(A,B), X\|\hat{T}\|p \bar p > = \frac{1}{\sqrt{\mu}} \int \frac{d^3 k}{(2\pi)^3} \Phi(\vec k) {\bold M}(\vec k). %\end{array}$$ $\mu$ is the reduced mass. A and B are two clusters to be combined as the X(5568). Here we take the $p \bar p$ collision process as the example. The relative momentum $q$ of these two clusters can be considered as small in the rest frame of X(5568), we get $$\label{swave} A (P)= \frac{1}{\sqrt{\mu}} \Psi(0) \hat{O}(q=0), q=P_A-P_B.$$ Here $\Psi$ is the wave function. $\hat{O}$ is the amplitude of production of two free ingredient particles (with vanishing relative momentum and proper angular momentum state). We only consider the simplest S-wave case. For the cross section, we need the absolute value square, with the proper initial flux factor $1/F$ and phase space integral. From this formula, the X(5568) is produced in two steps: First, the production of the ingredient hadron/cluster pair; second, the combination of this pair to X(5568) with probability described by $\|\Psi(0)\|^2$. For the first step, it is the $$\begin{aligned} \label{int} && \frac{1}{F} \sum_{j \neq A, B} \hspace{-0.67 cm} \int \prod \frac{d^3 p_j}{(2 \pi)^3 2 E_j} \overline{ \|\hat{O} \|^2} (p_j, P_A+P_B=P_H, q=0) \nonumber \\ && \times (2 \pi)^4 \delta^{(4)}(P_{intial}-\sum_{j \neq A, B} p_j-P_H)\end{aligned}$$ to be calculated. Here the average is taken on various spin states. It is not possible to be calculated directly with some effective quantum field theory/model when the initial state is (anti) protons and A and B are hadrons or diquarks. However, $$\begin{aligned} \label{int2} && \frac{1}{N} \frac{d N}{d^3 P_H d^3 q} \\ && \propto \frac{1}{F} \sum_{j \neq A, B} \hspace{-0.67 cm} \int \prod \frac{d^3 p_j}{(2 \pi)^3 2 E_j} \overline{ \|\hat{O} \|^2} (p_j, P_A+P_B=P_H, q) \nonumber \\ && \times (2 \pi)^4 \delta^{(4)}(P_{intial}-\sum_{j \neq A, B} p_j-P_A-P_B) \nonumber\end{aligned}$$ can be calculated by an event generator such as Pythia [@Pythia] or equivalently SDQCM [@Jin:2010wg] for the case that hadrons/diaurks A and B on shell. It is the advantage that in the framework we employ, only the on shell case is considered, so that the numerical calculation with event generator is plausible. The quantity of Eq. (\[int2\]) describes the two hadrons/diquarks A and B correlation in the phase space. For the hadron case, by proper integral on components of $P_H$ and/or q, the resulting correlations can directly be compared with data and serve for tuning the parameters. Employing the event generator, one gets $$\label{ext} \frac{1}{N} \frac{d N}{d^3 P_H d^3 q}, \forall q,$$ then extrapolates to the special case $q=0$. Numerically, one can take an average around $q=0$ for the above quantity [@long1]. For $B_s \pi$, we take the simple average of two cases, hadron bound state (B, K) and diquark pair as mentoned above, then fit the X(5568) spectrum measured by D0 Collaboration to get the effective wave function at origin (Fig. 1(a)). ![Transverse momentum distributions at Tevatron. (a) The dashed line is that for X(5568), with the best fitting of the wave function to get the correct $\rho$ measured by D0 collaboration. The solid line is for $B_s$ as reference. (b) For the charm sector, dashed - X$_c(D_s ^\pm \pi^\pm)$; solid - $D_s$. []{data-label="bs"}](BDpt0b_all.eps) The X(5568) transverse spectrum is softer than that of $B_s$, as demonstrated here and indicated from the experiment. This is from the fact that we require the two clusters near to each other in phase space for combination. Realized in the above formulation, is the relative momentum vanishing. This is in contrary to the fragmentation spectrum, the more massive, the harder. The cross sections of X(5568) in other collision processes and energies are easy to be obtained, since the effective wave function at origin is process- and energy- independent. Here we show the pseudo-rapidity $\eta$ distributions for proton-proton collisions at $\sqrt{s}=8$ TeV as an example (Fig. 2(a)). For others we refer to [@long1]. The production rate begins to fall beyond $\eta =3$, as general B hadrons. The ‘rapidity plateau’ is much more narrow than those of light charged hadrons (mainly pions). These results are useful for various detectors. Based on our calculation, one can go further to estimate the kinematic distributions of the signal particles which are from the decay of X(5568) and directly detected. As an example, Fig. 2 (b) showes the $k_T-k$ (transverse momentum and total momentum) distribution for the signal pions from the decay process $X \to B_s + \pi$ in the LHCb detector ranges ($2< \eta <5$). The mass difference between X(5568) and $B_s+\pi$ is small, and the pion mass is small. These facts lead to that the produced pions are not energetic, e.g., only around $10 \%$ of the signal pions with $k_T > 0.5$ GeV/c (the requirement of the relevant measurement by LHCb Collaboration [@LHCb:2016ppf]). The formulation is also applicable for the cross section of X$_c(D^{\pm}_s\pi^{\pm}$) state production. Both charm and bottom are heavy, and can be calculated by perturbative QCD in the exactly same way once taking into account the different value of the mass. If $X_c$ exists, it is not difficult to be detected. $D^{\pm}_s$ can be detected from $D^{\pm}_s \to \phi \pi$ channel, by proper 3 charged particle tracks from the vertex displaced from the primary one. Then this reconstructed $D^{\pm}_s$ can be combined with a proper charged particle track considered as $\pi$ from the [primary vertex]{} to give the invariant mass distribution to look for the resonance. If $K^0_s$ is well measured, $D^{\pm}_s$ can also be reconstructed from the 2K channel and then combined with the $\pi$ from the primary vertex. This kind of pions can eliminate the possibility that the $X_c$ produced from the decay of bottom. Of course just by keeping or not this restriction, one can preliminarily investigate $X_c$ from multi-production or from weak decay. Here we would like to emphasize that, since the mass of $X_c$ are around half of X(5568), it has a larger boost factor $\gamma$ about two times of that of X(5568) for the same momentum. This means whether Tevatron or LHC, in both central and large rapidity regions, the signal pions are more energetic to be detectable. The transverse distribution of $X_c$ at Tevatron energy, comparing with that of $D_s$, can be seen from Fig. \[bs\] (b). Here we assume that replacing $b$ by $c$ quark will not change the value of the wave function at origin, since the reduced mass is insensitive to the heavy ingredient. The production ratio $\rho_c=X_c/D_s$ for the transverse momentum region 10 $< p_T (D_s) <$ 15 GeV/$c$ and 15 $< p_T (D_s) <$ 30 GeV/$c$, is $ 10.2 \%$ and $ 7.9 \%$, respectively, almost similar as those of the bottom sector. It is copious enough and the search from experiments is reasonable. --------- --------- [(a)]{} [(b)]{} --------- --------- We look forward for the new charm four quark partner. This will shed light to the understanding of this special new state, as well as deepen our understanding on the hadronization mechanism, besides its structure [@rw]. These productions of X(5568) or possible $X_c$ can also be realized in high energy heavy ion collisions, with a larger rate since there strangeness and/or diquark is enhanced. The large number of quarks in unit phase space volume in heavy ion collisions also indicates that one can get a larger value of $\hat{O}$. These lead to larger production ratio $X(5568)/B^0_s$ ($X_c/D_s$), hence larger $B^0_s/B^0$ from X(5568) decay (also $D_s/D$). This can be measured as the ‘anomalous strangeness enhancement’ in the B (D) meson sector. Furthermore, it is also interesting to combine X(5568) with other tracks to look for heavier hadron, e.g., $B_c$, as complementary study. If one finds an $X_c$, to study those from multi-production and from heavier hadron decay are also very helpful, as mentioned above. We thank Profs. T Gershon, Y. R. Liu, L.L. Ma and Z. G. Si for discussions. This work is supported in part by NSFC and NSF, Shandong Province. [150]{} S.-K. Choi, [et al.]{} (Belle Collaboration), Phys. Rev. Lett. [100]{}, 142001 (2008). R. Mizuk, [et al.]{} (Belle Collaboration), Phys. Rev. D [78]{}, 072004 (2008). T. Aaltonen, [et al.]{} (CDF Collaboration), Phys. Rev. Lett. [102]{}, 242002 (2009). A. Bondar, [et al.]{} (Belle Collaboration), Phys. Rev. Lett. [108]{}, 122001 (2012). R. Aaij, [et al.]{} (LHCb Collaboration), Phys. Rev. Lett. [115]{}, 072001 (2015). V. M. Abazov [et al.]{} \[D0 Collaboration\], \[arXiv:1602.07588 \[hep-ex\]\]. Yi Jin, Shi-Yuan Li, et al., ‘Exotic Hadron Bound State production in hadronic colliders’, in preparation. Shi-Yuan Li, ‘Calculation of Production and Deacy of Baryonium via Positronium (Quarkonium) Approach’, ICTP Visitor’s Work Report, 2005, unpublished. Y. Jin, Z. Si, Q. Xie and T. Yao, Sci. China Phys. Mech. Astron. [53]{}, 1998 (2010) , and refs. therein. B. Andersson, “The Lund model,” Camb. Monogr. Part. Phys. Nucl. Phys. Cosmol. [7]{} (1997) 1. K. A. Olive [et al.]{} \[Particle Data Group Collaboration\], Chin. Phys. C [38]{} (2014) 090001. Marchesini G, Webber B R. Simulation of QCD jets including soft gluon interference. Nucl Phys B, 1984, 238: 1-29; Webber B R. A QCD model for jet fragmentation including soft gluon interference. Nucl Phys B, 1984, 238: 492-528 Y. Jin, S. Y. Li, Z. G. Si, Z. J. Yang and T. Yao, Phys. Lett. B [727]{}, 468 (2013) . Y. Jin, S. Y. Li, Y. R. Liu, Z. G. Si and T. Yao, Phys. Rev. D [89]{}, no. 9, 094006 (2014) . Y. Jin, H. L. Li, S. Q. Li, S. Y. Li, Z. G. Si, T. Yao and X. F. Zhang, Phys. Rev. D [91]{}, no. 11, 114017 (2015) . X. H. Liu and G. Li, arXiv:1603.00708 \[hep-ph\]; T. J. Burns and E. S. Swanson, arXiv:1603.04366 \[hep-ph\]. W. Han, S. Y. Li, Y. H. Shang, F. L. Shao and T. Yao, Phys. Rev. C [80]{}, 035202 (2009) . R. Q. Wang, J. Song, K. J. Sun, L. W. Chen, G. Li and F. L. Shao, arXiv:1601.02835 \[hep-ph\]. See Supplemental material at [http://www-d0.fnal.gov/ Run2Physics/WWW/results/final/B/B16A/]{}. T. Sjöstrand [et al.]{}, Comp. Phys. Commun. [ [135]{}]{}, 238 (2001). B. Abelev [et al.]{} \[ALICE Collaboration\], Phys. Lett. B [718]{} (2012) 279 doi:10.1016/j.physletb.2012.10.049 \[arXiv:1208.1948 \[hep-ex\]\]. The LHCb Collaboration, LHCb-CONF-2016-004, CERN-LHCb-CONF-2016-004. S. S. Agaev, K. Azizi and H. Sundu, arXiv:1602.08642 \[hep-ph\]; Z. G. Wang, arXiv:1602.08711 \[hep-ph\]; W. Wang and R. Zhu, arXiv:1602.08806 \[hep-ph\]; W. Chen, H. X. Chen, X. Liu, T. G. Steele and S. L. Zhu, arXiv:1602.08916 \[hep-ph\]; C. M. Zanetti, M. Nielsen and K. P. Khemchandani, arXiv:1602.09041 \[hep-ph\]; S. S. Agaev, K. Azizi and H. Sundu, arXiv:1603.00290 \[hep-ph\]; Y. R. Liu, X. Liu and S. L. Zhu, arXiv:1603.01131 \[hep-ph\]; S. S. Agaev, K. Azizi and H. Sundu, arXiv:1603.01471 \[hep-ph\]; J. M. Dias, K. P. Khemchandani, A. M. Torres, M. Nielsen and C. M. Zanetti, arXiv:1603.02249 \[hep-ph\]; Z. G. Wang, arXiv:1603.02498 \[hep-ph\]; S. S. Agaev, K. Azizi and H. Sundu, arXiv:1603.02708 \[hep-ph\]; X. G. He and P. Ko, arXiv:1603.02915 \[hep-ph\]; F. Stancu, arXiv:1603.03322 \[hep-ph\]; L. Tang and C. F. Qiao, arXiv:1603.04761 \[hep-ph\]; F. K. Guo, U. G. Meißner and B. S. Zou, arXiv:1603.06316 \[hep-ph\]; Q. F. Lü and Y. B. Dong, arXiv:1603.06417 \[hep-ph\]; A. Esposito, A. Pilloni and A. D. Polosa, arXiv:1603.07667 \[hep-ph\]; M. Albaladejo, J. Nieves, E. Oset, Z. F. Sun and X. Liu, arXiv:1603.09230 \[hep-ph\]; A. Ali, L. Maiani, A. D. Polosa and V. Riquer, arXiv:1604.01731 \[hep-ph\]; R. Albuquerque, S. Narison, A. Rabemananjara and D. Rabetiarivony, arXiv:1604.05566 \[hep-ph\]; X. Chen and J. Ping, arXiv:1604.05651 \[hep-ph\]; K. Terasaki, arXiv:1604.06161 \[hep-ph\]; H. Y. Cheng, The Universe [3]{}, no. 3, 33 (2015); J. Wu, Y. R. Liu, K. Chen, X. Liu and S. L. Zhu, arXiv:1605.01134 \[hep-ph\]; W. Chen, H. X. Chen, X. Liu, T. G. Steele and S. L. Zhu, arXiv:1605.01647 \[hep-ph\]; S. S. Agaev, K. Azizi and H. Sundu, arXiv:1605.02496 \[hep-ph\]; A. Ali, arXiv:1605.05954 \[hep-ph\]; R. F. Lebed, arXiv:1605.07975 \[hep-ph\].	{ "pile_set_name": "ArXiv" }
--- abstract: \| For several years, MPI has been the de facto standard for writing parallel applications. One of the most popular MPI implementations is MPICH. Its successor, MPICH2, features a completely new design that provides more performance and flexibility. To ensure portability, it has a hierarchical structure based on which porting can be done at different levels. In this paper, we present our experiences designing and implementing MPICH2 over InfiniBand. Because of its high performance and open standard, InfiniBand is gaining popularity in the area of high-performance computing. Our study focuses on optimizing the performance of MPI-1 functions in MPICH2. One of our objectives is to exploit Remote Direct Memory Access (RDMA) in Infiniband to achieve high performance. We have based our design on the RDMA Channel interface provided by MPICH2, which encapsulates architecture-dependent communication functionalities into a very small set of functions. Starting with a basic design, we apply different optimizations and also propose a zero-copy-based design. We characterize the impact of our optimizations and designs using microbenchmarks. We have also performed an application-level evaluation using the NAS Parallel Benchmarks. Our optimized MPICH2 implementation achieves 7.6 $\mu$s latency and 857 MB/s bandwidth, which are close to the raw performance of the underlying InfiniBand layer. Our study shows that the RDMA Channel interface in MPICH2 provides a simple, yet powerful, abstraction that enables implementations with high performance by exploiting RDMA operations in InfiniBand. To the best of our knowledge, this is the first high-performance design and implementation of MPICH2 on InfiniBand using RDMA support. author: - 'Jiuxing Liu[^^]{}' - 'Weihang Jiang[^^]{}' - 'Pete Wyckoff[^^]{}' - 'Dhabaleswar K. Panda[^^]{}' - 'David Ashton[^\\^]{}' - 'Darius Buntinas[^\\^]{}' - 'William Gropp[^\\^]{}' - 'Brian Toonen[^\\^]{}' - \| \ [^^]{}Computer and Information Science, The Ohio State University, Columbus, OH 43210\ {liuj, jiangw, panda}@cis.ohio-state.edu - \| \ [^^]{}Ohio Supercomputer Center, 1224 Kinnear Road, Columbus, OH 43212\ [email protected] - \| \ [^\\^]{}Mathematics and Computer Science Division, Argonne National Laboratory, Argonne, IL 60439\ {ashton, buntinas, gropp, toonen}@mcs.anl.gov bibliography: - 'mydoc\_bib.bib' title: ' Design and Implementation of MPICH2 over InfiniBand with RDMA Support[^1] ' --- Introduction {#sec:intro} ============ During the past ten years, the research and industry communities have proposed and implemented user-level communication systems to address some of the problems associated with traditional networking protocols. The Virtual Interface Architecture (VIA) [@via_dunning98] was proposed to standardize these efforts. More recently, the InfiniBand Architecture [@IB-SPEC] has been introduced, which combines storage I/O with interprocess communication. In addition to send and receive operations, InfiniBand architecture supports Remote Direct Memory Access (RDMA). RDMA operations enable direct access to the address space of a remote process. These operations introduce new opportunities and challenges in designing communication protocols. In the area of high-performance computing, MPI [@mpi-ref] has been the de facto standard for writing parallel applications. After the original MPI standard (MPI-1), an enhanced standard (MPI-2) [@MPI2] was introduced, which includes features such as dynamic process management, one-sided communication, and I/O. MPICH [@MPICH] is one of the most popular MPI-1 implementations. Recently, work has begun to create MPICH2 [@MPICH2], which aims to support both MPI-1 and MPI-2 standards. It features a completely new design that provides better performance and flexibility than its predecessor. MPICH2 is also very portable and provides mechanisms that make it easy to retarget MPICH2 to new communication architectures such as InfiniBand. In this paper, we present our experiences in designing and implementing MPICH2 over InfiniBand using RDMA operations. Although MPICH2 supports both MPI-1 and MPI-2, our study focuses on optimizing the performance of MPI-1 functions. We have based our design on the RDMA Channel interface provided by MPICH2, which encapsulates architecture-dependent communication functionalities in a small set of functions. Despite its simple interface, we have shown that the RDMA Channel does not prevent one from achieving high performance. In our testbed, our MPICH2 implementation achieves 7.6 $\mu$s latency and 857 MB/s peak bandwidth, which are quite close to the raw performance of the InfiniBand platform. We have also evaluated our designs using the NAS Parallel Benchmarks [@NAS_BENCH]. Overall, we have demonstrated that the RDMA Channel interface is a simple, yet powerful, abstraction that makes it possible to design high-performance MPICH2 implementations with less development effort. In our design, communication between processes is done exclusively using RDMA operations. Our design starts with an emulation of a shared-memory-based implementation. Then we introduce various optimization techniques to improve its performance. To evaluate the impact of each optimization, we use latency and bandwidth microbenchmarks. We also propose a zero-copy design for large messages. Our results show that with [piggybacking]{} and [zero-copy]{} optimizations for large messages, our design achieves very good performance. The remainder of the paper is organized as follows. In Section 2, we provide an introduction to InfiniBand and its RDMA operations. In Section 3, we present an overview of MPICH2, its implementation structure, and the RDMA Channel interface. In Sections 4 and 5, we describe our designs and implementations. In Section 6, we compare our RDMA Channel-based design with another design based on a more complicated interface called CH3. In Section 7, we present an application level performance evaluation. In Section 8, we describe related work. In Section 9, we draw conclusions and briefly mention some future research directions. InfiniBand Overview {#sec:iba} =================== The InfiniBand Architecture (IBA) [@IB-SPEC] defines a switched network fabric for interconnecting processing nodes and I/O nodes. It provides a communication and management infrastructure for interprocessor communication and I/O. In an InfiniBand network, processing nodes and I/O nodes are connected to the fabric by channel adapters (CAs). Channel adapters usually have programmable DMA engines with protection features. There are two kinds of channel adapters: host channel adapter (HCA) and target channel adapter (TCA). HCAs sit on processing nodes. The InfiniBand communication stack consists of different layers. The interface presented by channel adapters to consumers belongs to the transport layer. A queue-based model is used in this interface. A queue pair in InfiniBand Architecture consists of two queues: a send queue and a receive queue. The send queue holds instructions to transmit data, and the receive queue holds instructions that describe where received data is to be placed. Communication operations are described in work queue requests (WQRs), or descriptors, and are submitted to the work queue. The completion of WQRs is reported through completion queues (CQs). Once a work queue element is finished, a completion queue entry is placed in the associated completion queue. Applications can check the completion queue to see whether any work queue request has been finished. InfiniBand also supports different classes of transport service. In this paper, we focus on the reliable connection (RC) service. RDMA Operations in InfiniBand Architecture ------------------------------------------ InfiniBand Architecture supports both channel and memory semantics. In channel semantics, send/receive operations are used for communication. To receive a message, the programmer posts a receive descriptor that describes where the message should be put at the receiver side. At the sender side, the programmer initiates the send operation by posting a send descriptor. In memory semantics, InfiniBand supports remote direct memory access (RDMA) operations, including RDMA write and RDMA read. RDMA operations are one sided and do not incur software overhead at the remote side. In these operations, the sender (initiator) starts RDMA by posting RDMA descriptors. A descriptor contains both the local data source addresses (multiple data segments can be specified at the source) and the remote data destination address. At the sender side, the completion of an RDMA operation can be reported through CQs. The operation is transparent to the software layer at the receiver side. Since RDMA operations enable a process to access the address space of another process directly, they have raised some security concerns. In InfiniBand architecture, a key-based mechanism is used. A memory buffer must first be registered before it can be used for communication. Among other things, the registration generates a remote key. This remote key must be presented when the sender initiates an RDMA operation to access the buffer. Compared with send/receive operations, RDMA operations have several advantages. First, RDMA operations themselves are generally faster than send/receive operations because they are simpler at the hardware level. Second, they do not involve managing and posting descriptors at the receiver side, which would incur additional overheads and reduce the communication performance. MPICH2 Overview {#sec:mpich2} =============== MPICH [@MPICH] is developed at Argonne National Laboratory. It is one of the most popular MPI implementations. The original MPICH provides support for the MPI-1 standard. As a successor of MPICH, MPICH2 [@MPICH2] aims to support not only the MPI-1 standard but also functionalities such as dynamic process management, one-sided communication, and MPI I/O, which are specified in the MPI-2 standard. However, MPICH2 is not merely MPICH with MPI-2 extensions. It is based on a completely new design, aiming to provide more performance, flexibility, and portability than the original MPICH. One of the notable features in the implementation of MPICH2 is that it can take advantage of RDMA operations if they are provided by the underlying interconnect. These operations can be used not only to support MPI-2 one-sided communication but also to implement normal MPI-1 communication. Although MPICH2 is still under development, beta versions are already available for developers. In the current version, all MPI-1 functions have been implemented. MPI-2 functions are not completely supported yet. In this paper, we mainly focus on the MPI-1 part of MPICH2. MPICH2 Implementation Structure ------------------------------- One of the objectives in MPICH2 design is portability. To facilitate porting MPICH2 from one platform to another, MPICH2 uses ADI3 (the third generation of the Abstract Device Interface) to provide a portability layer. ADI3 is a full-featured abstract device interface and has many functions, so it is not a trivial task to implement all of them. To reduce the porting effort, MPICH2 introduces the CH3 interface. CH3 is a layer that implements the ADI3 functions and provides an interface consisting of only a dozen functions. A “channel” implements the CH3 interface. Channels exist for different communication architectures such as TCP sockets and, SHMEM. Because only a dozen functions are associated with each channel interface, it is easier to implement a channel than the ADI3 device. To take advantage of architectures with globally shared memory or RDMA capabilities and to further reduce the porting overhead, MPICH2 introduces the RDMA Channel which implements the CH3 interface. The RDMA Channel interface only contains five functions. We will discuss the details of the RDMA Channel interface in the next subsection. The hierarchical structure of MPICH2 , as shown in Figure \[fig:mpich2\], gives much flexibility to implementors. The three interfaces (ADI3, CH3, and the RDMA Channel interface) provide different trade-offs between communication performance and ease of porting. ![MPICH2 Implementation Structure[]{data-label="fig:mpich2"}](./figs/mpich2.eps){width="\columnwidth"} MPICH2 RDMA Channel Interface ----------------------------- The MPICH2 RDMA Channel interface is designed specifically for architectures with globally shared memory or RDMA capabilities. It contains five functions, among which only two are central to communication. (Other functions deal with process management, initialization, and finalization.) These two functions are called [put (write)]{} and [get (read)]{}. Both [put]{} and [get]{} functions accept a connection structure and a list of buffers as parameters. They return to the caller the number of bytes that have been successfully put or gotten. If the number of bytes completed is less than the total length of buffers, the caller will retry the same [get]{} or [put]{} operation later. Figure \[fig:put\_get\] illustrates the semantics of [put]{} and [get]{}. Logically, a pipe is shared between the sender and the receiver. The [put]{} operation writes to the pipe, and the [get]{} operation reads from it. The data in the pipe is consumed in FIFO order. Both operations are nonblocking in the sense that they return immediately with the number of bytes completed, instead of waiting for the entire operation to finish. We note that [put]{} and [get]{} are different from RDMA write and RDMA read in InfiniBand. While RDMA operations in InfiniBand are one sided, [put]{} and [get]{} in the RDMA Channel interface are essentially two-sided operations. [Put]{} and [get]{} operations can be implemented on architectures with globally shared memory in a straightforward manner. Figure \[fig:shmem\] shows one example. In this implementation, a shared buffer (organized logically as a ring) is placed in shared memory, together with a head pointer and a tail pointer. The [put]{} operation copies the user buffer to the shared buffer and adjusts the head pointer. The [get]{} operation involves reading from the shared buffer and adjusting the tail pointer. In the case of buffer overflow or underflow (detected by comparing head and tail pointers), the operations return immediately, and the caller will retry them. ![Put and Get Operations[]{data-label="fig:put_get"}](./figs/put_get.eps){width="0.9\columnwidth"} ![Put and Get Implementation with Globally Shared Memory[]{data-label="fig:shmem"}](./figs/shmem.eps){width="0.8\columnwidth"} Working at the RDMA Channel interface level is better than writing a new CH3 or ADI3 implementation for many reasons: 1. Improvements done at this level can affect all shared-memory-like transports such as globally shared memory, RDMA over IP, Quadrics, and Myrinet. 2. Other protocols on InfiniBand need efficient processing, including one-sided communication in MPI-2, DSM systems, and parallel file systems. The RDMA Channel interface can potentially be used also for them. 3. Designing proper interfaces to similar systems improves performance and portability in general. In collaboration, the OSU and ANL teams are also currently working together to design an improved interface that can benefit communication systems in general. Designing and Optimizing MPICH2 over InfiniBand {#sec:design} =============================================== In this section, we present several different designs of MPICH2 over InfiniBand based on the RDMA Channel interface. We first start with a basic design that resembles the scheme described in Figure \[fig:shmem\]. Then we apply various optimization techniques to improve its performance. In this section, the designs are evaluated by using microbenchmarks such as latency and bandwidth. We show that by taking advantage of RDMA operations in InfiniBand, we can achieve not only low latency for small messages but also high bandwidth for large messages using the RDMA Channel interface. In Section 5, we present a zero-copy design. Experimental Testbed -------------------- Our experimental testbed consists of a cluster system with 8 SuperMicro SUPER P4DL6 nodes. Each node has dual Intel Xeon 2.40 GHz processors with a 512K L2 cache and a 400 MHz front side bus. The machines are connected by Mellanox InfiniHost MT23108 DualPort 4X HCA adapter through an InfiniScale MT43132 Eight 4x Port InfiniBand Switch. The HCA adapters work under the PCI-X 64-bit 133MHz interfaces. We used the Linux Red Hat 7.2 operating system with 2.4.7 kernel. The compilers we used were GNU GCC 2.96 and GNU FORTRAN 0.5.26. Basic Design ------------ In Figure \[fig:shmem\], we illustrated how the RDMA Channel interface can be implemented on shared-memory architectures. In a cluster connected by InfiniBand, however, there is no physically shared memory. In our basic design, we use RDMA write operations provided by InfiniBand to emulate this scheme. We put the shared-memory buffer in the receiver’s main memory. This memory is registered and exported to the sender. Therefore, it is accessible to the sender through RDMA operations. To avoid the relatively high cost of registering user buffers for sending every message, we also use a preregistered buffer at the sender that is the same size as the shared buffer at the receiver. User data is first copied into this buffer and then sent out. Head and tail pointers also need to be shared between the sender and the receiver. Since they are used frequently at both sides, we use replication to prevent polling through the network. For the tail pointer, a master copy is kept at the receiver, and a replica is kept at the sender. For the head pointer, a master copy is kept at the sender, and a replica is kept at the receiver. For each direction of every connection, the associated “shared” memory buffer, head and tail pointers are registered during initialization, and their addresses and remote keys are exchanged. At the sender, the [put]{} operation is implemented as follows: 1. Use local copies of head and tail pointers to decide how much empty space is available. 2. Copy user buffer to the preregistered buffer. 3. Use RDMA write operation to write the data to the buffer at the receiver side. 4. Adjust the head pointer based on the amount of data written. 5. Use another RDMA write to update the remote copy of head pointer. 6. Return the number of bytes written. At the receiver, the [get]{} operation is implemented in the following way: 1. Check local copies of head and tail pointers to see whether there is new data available. 2. Copy the data from the shared memory buffer to user buffer. 3. Adjust the tail pointer based on the amount of data that has been copied. 4. Use an RDMA write to update the remote copy of tail pointer. 5. Return the number of bytes successfully read. We note that copies of head and tail pointers are not always consistent. For example, after a sender adjusts its head pointer, it uses RDMA write to update the remote copy at the receiver. Therefore, the head pointer at the receiver is not up to date until the RDMA write finishes. However, this inconsistency does not affect the correctness of the scheme; it merely prevents the receiver from reading new data temporarily. Similarly, inconsistency of tail pointer may prevent the sender from writing to the shared buffer. But eventually the pointers will become up to date, and the sender or the receiver will be able to make progress. ### Performance of the Basic Design We use latency and bandwidth tests to evaluate the performance of our basic design. The latency test is conducted in a ping-pong fashion, and the results are derived from round-trip time. In the bandwidth test, a sender keeps sending back-to-back messages to the receiver until it has reached a predefined window size [W]{}. Then it waits for these messages to finish and send out another [W]{} messages. The results are derived from the total test time and the number of bytes sent. Figures \[fig:basic\_lat\] and \[fig:basic\_bw\] show the results. Our basic design achieves a latency of 18.6 $\mu$s for small messages and a bandwidth of 230 MB/s for large messages. (Unless stated otherwise, the unit MB in this paper is an abbreviation for $10{}^{6}$ bytes, not $2^{20}$ bytes.) These numbers are much worse than the raw performance numbers achievable by the underlying InfiniBand layer (5.9 $\mu$s latency and 870 MB/s bandwidth). ![MPI Latency for Basic Design[]{data-label="fig:basic_lat"}](./figs/basic_lat.eps){width="\columnwidth"} ![MPI Bandwidth for Basic Design[]{data-label="fig:basic_bw"}](./figs/basic_bw.eps){width="\columnwidth"} A careful look at the basic design reveals many inefficiencies. For example, a matching pair of send and receive operations in MPI require three RDMA write operations to take place: one for transfer of data, and two for updating head and tail pointers. These not only increase latency and host overhead but also generate unnecessary network traffic. For large messages, the basic scheme leads to two extra memory copies. The first one is from user buffer to the preregistered buffer at the sender side. The second one is from the shared buffer to user buffer at the receiver side. These memory copies consume resources such as memory bandwidth and CPU cycles. To make matters worse, in the basic design the memory copies and communication operations are serialized. For example, a sender first copies the whole message (or part of the message if it cannot fit itself in the empty space of the preregistered buffer). Then it initiates RDMA write to transfer the data. This serialization of copying and RDMA write greatly reduces the bandwidth for large messages. Optimization with Piggybacking Pointer Updates ---------------------------------------------- Our first optimization targeted to avoid separate head and tail pointer updates whenever possible. The technique we used is piggybacking, which combines pointer updates with data transfer. At the sender side, we combine data and the new value of head pointer into a single message. To help the receiver detect the arrival of the message, we attach the size with the message and put two flags at the beginning and the end of the message. The receiver detects arrival of the new message by polling on the flags. To avoid possible situations where the buffer content happens to have the same value as the flag, we divide the shared buffer into fixed-sized chunks. Each message uses a different chunk. In this way, the situations can be handled by using two polling flags, or “bottom fill.” Similar techniques have been used in [@MVAPICH_RDMA; @Rinku-ipdps03]. At the receiver side, instead of using RDMA write to update the remote tail pointer each time data has been read, we delay the updates until the free space in the shared buffer drops below a certain threshold. If messages are sent from the receiver to the sender, the pointer value is attached with the message, and no extra message is used to transfer pointer updates. If no messages are sent from the receiver to the sender, eventually we will explicitly send the updates by using an extra message. The sender updates its pointer after receiving this message. Even in this case, however, the traffic can be reduced because several consecutive updates of the tail pointer can be sent by using only one message. The use of piggybacking and delayed pointer updates can greatly improve the performance of small message. From Figure \[fig:piggyback\_lat\] we see that the latency is reduced from 18.6 $\mu$s to 7.4 $\mu$s. Figure \[fig:piggyback\_bw\] shows that the optimization also improves bandwidth for small messages. ![MPI Small-Message Latency with Piggybacking[]{data-label="fig:piggyback_lat"}](./figs/piggyback_lat.eps){width="\columnwidth"} ![MPI Small-Message Bandwidth with Piggybacking[]{data-label="fig:piggyback_bw"}](./figs/piggyback_bw.eps){width="\columnwidth"} Optimization with Pipelining of Large Messages ---------------------------------------------- As we have discussed, our basic design suffers from serialization of memory copies and RDMA writes. A better solution is to use pipelining to overlap memory copies with RDMA write operations. In our piggybacking optimization, we divide the shared-memory buffer into small chunks. When sending and receiving large messages, we need to use more than one such chunks. At the sender side, instead of starting RDMA writes after copying all the chunks, we initiate the RDMA transfer immediately after copying each chunk. In this way, the RDMA operation can be overlapped with the copying of the next chunk. Similarly, at the receiver side we start copying from the shared buffer to the user buffer immediately after a chunk is received. In this way, the receive RDMA operations can be overlapped with the copying. Figure \[fig:pipe\_bw\] compares the bandwidth of the pipelining scheme with the basic scheme. (Piggybacking is also used in the pipelining scheme.) We can see that pipelining combined with piggybacking has greatly improved MPI bandwidth. The peak bandwidth has been increased from 230 MB/s to over 500 MB/s. This result is still not satisfying, however, because InfiniBand is able to deliver bandwidth up to 870 MB/s. To investigate the performance bottleneck, we have conducted memory copy tests in our testbed. We have found that memory copy bandwidth is less than 800 MB/s for large messages. In our MPI bandwidth tests, with RDMA write operations and memory copies both using the memory bus, the bandwidth achievable at the application level is even less. Therefore, the memory bus clearly becomes a performance bottleneck for large messages because of the extra memory copies. In the pipelining optimization, it is important that we balance each stage of the pipeline so that we can get maximum throughput. One parameter we can change to balance pipeline stages is the chunk size, or how much data we copy each time for a large message. Figure \[fig:pipeline\_bufsize\] shows MPI bandwidth for different chunk sizes for the pipelining optimization. We observe that MPI does not give good performance when the chunk size is either too small (1K Bytes) or too large (32K Bytes). MPI performs comparably for chunk sizes of 2K to 16K Bytes. In all remaining tests, we have chosen a chunk size of 16K Bytes. ![MPI Bandwidth with Pipelining[]{data-label="fig:pipe_bw"}](./figs/pipe_bw.eps){width="\columnwidth"} ![MPI Bandwidth with Pipelining (Different Chunk Sizes)[]{data-label="fig:pipeline_bufsize"}](./figs/pipeline_bufsize.eps){width="\columnwidth"} Zero-Copy Design ================ As we have discussed in the preceding section, it is desirable to avoid memory copies for large messages. In this section, we describe a zero-copy design for large messages based on the RDMA Channel interface. In our new design, small messages are still transferred by using RDMA write, similar to the piggybacking scheme. For large messages, however, RDMA read, instead of RDMA write, is used for data transfer. The basic idea of our zero-copy design is to let the receiver “pull” the data directly from the sender using RDMA read. For each connection, shared buffers are still used for transferring small messages. However, the data for large messages is not transferred through the shared buffer. At the sender, when the [put]{} function is called, it checks the user buffer and decides whether to use zero-copy or not, based on the buffer size. If zero-copy is not used, the message is sent through the shared buffer as discussed before. If zero-copy is used, the function registers the user buffer, constructs a special packet that contains information about the user buffer such as address, size, and remote key, then sends the special packet by using RDMA write through the shared buffer. The [put]{} function returns a value of 0, at this stage, because no data has been transferred yet. Subsequent calls to [put]{} also return 0 until all of the data has been transferred, and the operation has completed. Once the operation has completed, [put]{} will return the number of bytes transferred. When the packet arrives at the other side and the [get]{} function is called, the receiver checks the shared buffer and processes all the packets in order. If a packet is a data packet, the data is copied to the user buffer. If it is a special packet, the user buffer is registered, and an RDMA read operation is issued to fetch the data from the remote side directly to the user buffer. After initiating the RDMA read, the [get]{} function returns with a value of 0, because the operation is still in progress. When the RDMA read is finished, calling the [get]{} function leads to an acknowledgment packet being sent to the sender. The [get]{} function then returns the number of bytes successfully transferred. When the acknowledgment packet is received at the sender side, the sender deregisters the user buffer, completing the operation, and the next call to the [put]{} function will return the number of bytes transferred. The zero-copy process is illustrated in Figure \[fig:zcopy\]. ![Zero-Copy Design[]{data-label="fig:zcopy"}](./figs/zcopy.eps){width="0.8\columnwidth"} In the current InfiniBand implementation, memory registration and deregistration are expensive operations. To reduce the number of registrations and deregistrations, we have implemented a registration cache [@PINDOWN-CACHE]. The basic idea is to delay the deregistration of user buffers and put them into a cache. If the same buffer is reused later, its registration information can be fetched directly from the cache instead of going through the expensive registration process. Deregistration happens only when there are too many registered user buffers. We note that the effectiveness of registration cache depends on buffer reuse patterns of applications. If applications rarely reuse buffers for communication, registration overhead cannot be avoided most of the time. Fortunately, our previous study with the NAS Parallel Benchmarks [@MVAPICH_SC] has demonstrated that buffer reuse rates are very high in these applications. We compare the bandwidth of the pipelining design and the zero-copy design in Figure \[fig:zcopy\_bw\]. We observe that zero-copy greatly improves the bandwidth for large messages. We achieve a peak bandwidth of 857 MB/s, which is quite close to the peak bandwidth at the InfiniBand level (870 MB/s). We also see that as a result of cache effect, bandwidth for large messages drops for the pipelining design. Because of the extra overhead in the implementation, the zero-copy design slightly increases the latency for small messages, which is now around 7.6 $\mu$s. ![MPI Bandwidth with Zero-Copy and Pipelining[]{data-label="fig:zcopy_bw"}](./figs/zcopy_bw.eps){width="\columnwidth"} Our zero-copy implementation uses RDMA read operations, which let the receiver “pull” data from the sender. An alternative is to use RDMA write operations and let the sender “push” data to the receiver. Before the sender can push the data, the receiver has to use special packets to advertise availability of new receive buffers. Therefore, this method can be very efficient if the [get]{} operations are called [before]{} the corresponding [put]{} operations. In the current MPICH2 implementation, however, the layers above the RDMA Channel interface are implemented in such a way that [get]{} is always called after [put]{} for large messages. Therefore, we have chosen an RDMA read-based implementation instead of RDMA write. Comparing CH3 and RDMA Channel Interface Designs ================================================ The RDMA Channel interface in MPICH2 provides a simple way to implement MPICH2 in many communication architectures. In the preceding section, we showed that this interface does not prevent one from achieving good performance. Nor does it prevent zero-copy implementation for large messages. Our results showed that with various optimizations, we can achieve a latency of 7.6 $\mu$s and a peak bandwidth of 857 MB/s. The CH3 interface is more complicated than the RDMA Channel interface. Therefore, porting it requires more effort. However, since CH3 provides more flexibility, it is possible to achieve better performance at this level. To study the impact of different interfaces on MPICH2 performance, we have also done a CH3-level implementation. This implementation uses RDMA write operations for transferring large messages, as shown in Figure \[fig:rdmawrite\]. Before transferring the message, a handshake happens between the sender and the receiver. User buffer at the receiver is registered and its information is sent to the sender through the handshake. The sender then uses RDMA write to transfer the data. A registration cache is also used in this implementation. Figures \[fig:rdmawrite\_lat\] and \[fig:rdmawrite\_bw\] compare this implementation with our RDMA Channel-based zero-copy design using latency and bandwidth microbenchmarks. We see that the two implementations perform comparably for small and large messages. However, the CH3-based design outperforms the RDMA Channel-based design for mid-sized messages (32K to 256K Bytes) in bandwidth. Figure \[fig:vapi\_readwrite\] shows the bandwidth of RDMA read and RDMA write at the InfiniBand VAPI level. (VAPI is the programming interface for our InfiniBand cards.) With the current VAPI implementation, RDMA write operations have a clear advantage over RDMA read for mid-sized messages. Therefore, the fact that CH3-based design outperforms RDMA Channel-based design for mid-sized messages is more the result of the raw performance difference between RDMA write and RDMA read than the designs themselves. ![CH3 Zero-Copy with RDMA Write[]{data-label="fig:rdmawrite"}](./figs/rdmawrite.eps){width="0.8\columnwidth"} ![MPI Latency for CH3 Design and RDMA Channel Interface Design[]{data-label="fig:rdmawrite_lat"}](./figs/rdmawrite_lat.eps){width="\columnwidth"} ![MPI Bandwidth for CH3 Design and RDMA Channel Interface Design[]{data-label="fig:rdmawrite_bw"}](./figs/rdmawrite_bw.eps){width="\columnwidth"} ![InfiniBand Bandwidth[]{data-label="fig:vapi_readwrite"}](./figs/vapi_readwrite.eps){width="\columnwidth"} Application-Level Evaluation {#sec:perf} ============================ In this section, we carry out an application-level evaluation of our MPICH2 designs using NAS Parallel Benchmarks [@NAS_BENCH]. We run class A benchmarks on 4 nodes and class B benchmarks on 8 nodes. Benchmarks SP and BT require a square number of nodes. Therefore, their results are only shown for 4 nodes. The results are shown in Figures \[fig:mpich2\_nas4\] and \[fig:mpich2\_nas8\]. We have evaluated three designs: RDMA Channel implementation with pipelining for large messages (Pipelining), RDMA Channel implementation with zero-copy for large messages (RDMA Channel), and CH3 implementation with zero-copy (CH3). Although the performance difference of these three designs is not much, we observe that the pipelining design performs the worst in all cases. The RDMA Channel-based zero-copy design performs very close to the the CH3-based zero-copy design. On average, the CH3-based design performs less than 1% better on both 4 nodes and 8 nodes. ![NAS Class A on 4 Nodes[]{data-label="fig:mpich2_nas4"}](./figs/mpich2_nas4.eps){width="\columnwidth"} ![NAS Class B on 8 Nodes[]{data-label="fig:mpich2_nas8"}](./figs/mpich2_nas8.eps){width="\columnwidth"} Related Work {#sec:related} ============ As the predecessor of MPICH2 and one of the most popular MPI implementations, MPICH supports a similar implementation structure as MPICH2. MPICH provides ADI2 (the second generation of Abstract Device Interface) and Channel interface. Various implementations exist based on these interfaces [@MPI-MAD; @MPIMyrinet; @MPIQuadrics; @MVICH]. Our MVAPICH implementation [@MVAPICH_RDMA], which exploits RDMA write in InfiniBand, is based on the ADI2 interface. Since MPICH2 is relatively new, there exists very little work describing its implementations on different architectures. In [@CH3], a CH3-level implementation based on TCP/IP is described. Work in [@MPICH2_DE] presents an implementation MPICH2 over InfiniBand, also using the CH3 interface. However, in our paper, our focus is on the RDMA Channel interface instead of the CH3 interface. MPICH2 is designed to support both MPI-1 and MPI-2 standards. There have been studies about supporting the MPI-2 standard, especially one-sided communication operations [@MPI2-SUN; @MPI2-NEC]. To date, we have concentrated on supporting MPI-1 functions in MPICH2. We plan to explore the support of MPI-2 functions in the future. Because of its high bandwidth and low latency, the InfiniBand Architecture has been used as the communication subsystem in a number of systems other than MPI, such as distributed shared-memory systems and parallel file systems [@IB-PVFS; @IB-DSM]. The RDMA Channel interface presents a stream-based abstraction somewhat similar to the traditional socket interface. There have been studies of how to implement user-level socket interface efficiently over high-speed interconnects such as Myrinet, VIA, and Gigabit Ethernet [@SOCKET-GM; @SOCKET-VIA; @SOCKET-EMP]. Recently, Socket Direct Protocol (SDP) [@SDP] has been proposed, which provides a socket interface over InfiniBand. The idea of our zero-copy scheme is similar to the Z-Copy scheme in SDP. However, there are also differences between the RDMA Channel interface and the traditional socket interface. For example, [put]{} and [get]{} functions in RDMA Channel interface are nonblocking, while functions in the traditional sockets are usually blocking. To support traditional socket interface, one has to make sure the same semantics are maintained. We do not have to deal with this issue for the RDMA Channel interface. Conclusions and Future Work {#sec:conclusion} =========================== In this paper, we present a study of using RDMA operations to implement MPICH2 over InfiniBand. Our work takes advantage of the RDMA Channel interface provided by MPICH2. The RDMA Channel interface provides a very small set of functions to encapsulate the underlying communication layer on which the whole MPICH2 implementation is built. Consisting of only five functions, the RDMA Channel interface is easy to implement for different communication architectures. However, the question arises whether this abstraction is powerful enough that one can still achieve good performance. Our study has shown that the RDMA Channel interface still provides the implementors much flexibility. With optimizations such as piggybacking, pipelining, and zero-copy, MPICH2 is able to deliver good performance to the application layer. For example, one of our designs achieves 7.6 $\mu$s latency and 857 MB/s peak bandwidth, which come quite close to the raw performance of InfiniBand. In our study, we characterize the impact of each optimization by using latency and bandwidth microbenchmarks. We have also conducted an application-level evaluation using the NAS Parallel Benchmarks. So far, our study has been restricted to a fairly small platform consisting of 8 nodes. In the future, we plan to use larger clusters to study various aspects of our designs regarding scalability. Another direction we are pursuing is to provide support for MPI-2 functionalities such as one-sided communication using RDMA and atomic operations in InfiniBand. We are also working on how to support efficient collective communication on top of InfiniBand. The submitted manuscript has been created in part by the University of Chicago as Operator of Argonne National Laboratory (“Argonne”) under Contract No. W-31-109-ENG-38 with the U.S. Department of Energy. The U.S. Government retains for itself, and others acting on its behalf, a paid-up, nonexclusive, irrevocable worldwide license in said article to reproduce, prepare derivative works, distribute copies to the public, and perform publicly and display publicly, by or on behalf of the Government. [^1]: This research is supported in part by Department of Energy’s Grant \#DE-FC02-01ER25506, a grant from Sandia National Laboratories, a grant from Intel Corporation, National Science Foundation’s grants \#CCR-0204429 and \#CCR-0311542, and by the Mathematical, Information, and Computational Sciences Division subprogram of the Office of Advanced Scientific Computing Research, Office of Science, U.S. Department of Energy, under Contract W-31-109-ENG-38.	{ "pile_set_name": "ArXiv" }
--- abstract: 'The space of Frobenius manifolds has a natural involutive symmetry on it: there exists a map $I$ which send a Frobenius manifold to another Frobenius manifold. Also, from a Frobenius manifold one may construct a so-called almost dual Frobenius manifold which satisfies almost all of the axioms of a Frobenius manifold. The action of $I$ on the almost dual manifolds is studied, and the action of $I$ on objects such as periods, twisted periods and flows is studied. A distinguished class of Frobenius manifolds sit at the fixed point of this involutive symmetry, and this is made manifest in certain modular properties of the various structures. In particular, up to a simple reciprocal transformation, for this class of modular Frobenius manifolds, the flows are invariant under the action of $I\,.$' address: \| Department of Mathematics\ University of Glasgow\ Glasgow G12 8QQ\ U.K. author: - 'Ewan K. Morrison, Ian A.B. Strachan' date: '3$^{\rm rd}$ June 2010' title: Modular Frobenius manifolds and their invariant flows --- Introduction ============ The main property of the Chazy equation $$\gamma'''(\tau)=6\gamma(\tau) \gamma''(\tau)-9{\gamma'(\tau)}^2 \,, \label{chazy}$$ is its modularity property, that is, its invariance under the action of the group $SL(2,\mathbb{Z})\,:$ $$\begin{aligned} \tau & \mapsto & \frac{a\tau + b}{c\tau + d}, \quad a\,,b\,,c\,,d \in \mathbb{Z}\,, ad-bc =1, \\ \gamma(\tau) & \mapsto & (c\tau + d)^2\gamma(\tau)+2c(c\tau + d).\end{aligned}$$ The aim of this paper is to study integrable equations of hydrodynamic type $$\frac{\partial t^i}{\partial T} = M^i_j({\bf t}) \frac{\partial t^j}{\partial X}\,, \qquad\qquad i\,,j=1\,,\ldots\,, N$$ which have a similar invariance under certain multivariable extensions of such modular transformations. The origins of such systems comes from the study of certain Frobenius manifolds [@dubrovin1]. Given an arbitrary Frobenius manifold one can associate to it natural families of hydrodynamic systems, and the systems that will be studied arise from this construction. Before discussing the symmetry properties of these integrable hydrodynamic equations we first examine the modularity properties of the underlying Frobenius manifold. The space of Frobenius manifolds possesses various symmetries and the main object of study in this paper is an involutive symmetry $I\,,$ also known as a type II-symmetry, which maps solutions $F$ of the underlying WDVV equations to a new solution $\widehat F$: $$I\,:\, F \longrightarrow {\widehat F}\,.$$ Up to some trivial redefinitions of variables by changes of sign, $I^2=id\,,$ the identity transformation [^1]. However, certain special Frobenius manifolds lie at the fixed point of this involution and these inherit special modularity properties: for convenience such manifolds will be referred to as modular Frobenius manifolds, and a characterization of these, in terms of spectral data, will be given. Examples of such modular Frobenius manifolds include: - Jacobi group orbit spaces [@B; @iabs2]; - the unfolding spaces of the three simple elliptic singularities [@iabs3] ${\widetilde E}_{6,7,8}\,.$ In the first part of the paper we combine the study of $I$ with the notion of an almost dual Frobenius manifold. Given a Frobenius manifold, and hence a solution $F$ of the underlying WDVV equations, Dubrovin [@dubrovin2] showed how one can construct a new solution $F^\star$ which satisfies most (but crucially not all) of the axioms of a Frobenius manifold. We denote this transformation by $\star:$ $$F \stackrel{\star}{\longrightarrow} F^\star$$ Combining this with the symmetry $I$ leads to the following picture $$\begin{array}{ccc} F&\overset{I}{\longrightarrow}& {\hat F}\\ \downarrow & & \downarrow \\ F^\star && {\hat F}^\star \end{array}$$ In section 4 we study the induced involutive symmetry $$F^\star \overset{I^\star}{\longrightarrow} {\hat F}^\star\,.$$ in terms of almost duality. It turns out that the action $I^\star$ is different for modular and non-modular Frobenius manifolds (the action of type I-symmetries were discussed in [@RS2]). The study of the action of the modular group of the associated hydrodynamic flows will be done in two stages: the action of the symmetry $I$ on the tensors $M^i_j({\bf t})$ will first be given for an arbitrary Frobenius manifold (similar results have also been recently obtained by Dingdian and Zhang [@zhang]) and then specialized to those tensors arising from modular Frobenius manifolds. This gives a new symmetry of the corresponding hydrodynamic flows. Flow equations in the dual picture can also be written down (this being essentially a Muira-type transformation) and the action of $I^\star$ on these flows can then be studied. One of the advantages of this dual picture is that is it easier to write down general examples of such flows. Thus large classes of explicit modular invariant flows may be written down explicitly. Preliminaries {#pre} ============= Frobenius Manifolds and the WDVV Equations ------------------------------------------ Frobenius manifolds were introduced as a way to give a geometric understanding to solutions of the Witten-Dijkraaf-Verlinde-Verlinde (WDVV) equations, $$\label{wdvv} \frac{\partial^3 F}{\partial t^{\alpha} \partial t^{\beta} \partial t^{\lambda}}\eta^{\lambda \mu} \frac{\partial^3 F}{\partial t^{\mu} \partial t^{\delta} \partial t^{\gamma}} = \frac{\partial^3 F}{\partial t^{\delta} \partial t^{\beta} \partial t^{\lambda}}\eta^{\lambda \mu} \frac{\partial^3 F}{\partial t^{\mu} \partial t^{\alpha} \partial t^{\gamma}}$$ for some quasihomogeneous function $F(t)$. Throughout this paper $\eta^{\alpha\beta}$ will be defined via $\eta^{\alpha\beta}\eta_{\beta\kappa}=\delta^{\alpha}_{\kappa}$ where $$\eta_{\alpha\beta} = \frac{\partial^3 F}{\partial t^1 \partial t^{\alpha} \partial t^{\beta}}$$ is constant and non-degenerate. We recall briefly how to establish the correspondence between Frobenius manifolds and solutions of WDVV. The triple $(A,\circ, \eta)$ is a Frobenius algebra if: - [$(A, \circ)$ is a commutative associative algebra over ${\mathbb C}$ with unity $e$;]{} - [The bilinear pairing $\eta$ and mutiplication $\circ$ satisfy the following Frobenius condition $$\eta( X\circ Y, Z ) = \eta( X, Y \circ Z )\,, \quad X, Y, Z \in A.$$]{} With this one may define a Frobenius manifold. Let ${\mathcal M}$ be a smooth manifold. ${\mathcal M}$ is called a Frobenius manifold if each tangent space $T_t{\mathcal M}$ is equipped with the structure of a Frobenius algebra varying smoothly with $t\in {\mathcal M}$, and further - [The invariant inner product $\eta$ defines a flat metric on ${\mathcal M}$. ]{} - [The unity vector field is covariantly constant with respect to the Levi-Civitá connection for $\eta$, $$\label{fm1} {}^\eta\nabla e =0.$$]{} - [Let $$c(X,Y,Z):=\eta( X\circ Y, Z)\,, \quad X, Y, Z \in T_t{\mathcal M},$$ then the $(0,4)$ tensor ${}^\eta\nabla_Wc(X,Y,Z)$ is totally symmetric. ]{} - [There exists a vector field $E\in\Gamma(T{\mathcal M})$ such that $\nabla\nabla E=0$ and $${\mathcal L}_E \eta = (2-d)\eta, \quad {\mathcal L}_E \circ = \circ, \quad {\mathcal L}_E e = -e.$$]{} $E$ is called the Euler vector field. Condition 1 implies there exist a choice of coordinates $(t^1, ..., t^N)$ such that the Gram matrix $\eta_{\alpha\beta}=\langle \partial_{\alpha}, \partial_{\beta} \rangle$ is constant. Furthermore, this may be done in such a way that $e=\partial_1$. In such a coordinate system, partial and covariant derivatives coincide, and condition 3 becomes $c_{\alpha\beta\gamma,\kappa} =c_{\alpha\beta\kappa,\gamma}$. Successive applications of the Poincaré lemma then implies local existence of a function $F(t)$ called the free energy of the Frobenius manifold such that $$c_{\alpha\beta\gamma} = \frac{\partial^3 F(t)}{\partial t^{\alpha}\partial t^{\beta} \partial t^{\gamma}}.$$ Since $\eta(X,Y) = \eta (e\circ X, Y) = c(e, X, Y)$, we have $$\eta_{\alpha\beta} = c_{1\alpha\beta}.$$ Defining $(\eta_{\alpha\beta})^{-1}=\eta^{\alpha\beta}$, the components of $\circ$ are given by $c^{\alpha}_{\beta\gamma}=\eta^{\alpha\varepsilon}c_{\varepsilon\beta\gamma}$. Associativity of $\circ$ is then equivalent to . Condition 4 leads to requiring $F$ be quasihomogeneous $${\mathcal L}_EF = E(F) = d_F \cdot F = (3-d)F \quad \mbox{modulo quadratic terms}.$$ We remark further that in the case where the grading operator $\nabla E$ of the Frobenius algebras is diagonalizable, the Euler vector field may be reduced to the form $$\label{euvect1} E(t) = \sum_{\alpha}(1-q_\alpha)t^{\alpha}\partial_{\alpha} + \sum_{\{\alpha:q_\alpha=1\}} r^\sigma \partial_\sigma.$$ We will, for the most part, restrict ourselves in this paper to those Frobenius manifolds with the property that $r^\sigma=0\,,\sigma=1\,,\ldots\,,N\,.$ This excludes those examples coming from quantum cohomology and the extended affine Weyl orbit spaces [@DZ]: the application of the inversion symmetry (defined below) even to the simplest prepotential $$F=\frac{1}{2} (t^1)^2 t^2 + e^{t^2}$$ leads to prepotentials of non-analytic functions (e.g. functions such as $e^{-1/x}$ will appear). The deformed connection {#gmeq} ----------------------- One may define on a Frobenius manifold a one parameter family of flat connections parameterized by $z \in {\mathbb P}^1$, $${}^\eta {\widetilde\nabla}_X Y = {}^\eta\nabla_X Y -z X \circ Y, \quad X, Y, Z \in \Gamma(T{\mathcal M}).$$ Here ${}^\eta\nabla$ is the Levi-Civita connection of the metric $\eta\,.$ This connection is flat identically in $z$ by virtue of the axioms of a Frobenius manifolds. Thus there exists a choice of coordinates[^2] $\mathfrak{t}_{\alpha}(t, \lambda)$ such that $${}^\eta{\widetilde\nabla} d\mathfrak{t}_{\alpha} = 0.$$ In the flat coordinates $(t^1, ... , t^N)$ of connection ${}^\eta\nabla$ this reads $$\label{gm1} \frac{\partial^2 \mathfrak{t}^{\alpha}}{\partial t^{\kappa}\partial t^{\varepsilon}} = z c^{\sigma}_{\kappa\varepsilon}\frac{\partial \mathfrak{t}^{\alpha}}{\partial t^{\sigma}}.$$ Taking in to account the fact that for $z=0$ the two coordinate systems $(t^1, ... t^N)$ and $(\mathfrak{t}^1, ... , \mathfrak{t}^N)$ coincide, one may construct power series solutions to : $$\mathfrak{t}^{\alpha}({\bf t},z) = \sum_{n=0}^{\infty} z^n h^{(n,\alpha)}({\bf t}).$$ This ansatz yields a recursion relation for the functions $h^{(n,\alpha)}$, $$\label{recur1} \frac{\partial^2 h^{(n,\alpha)}}{\partial t^{\kappa}\partial t^{\varepsilon}} = c^{\sigma}_{\kappa\varepsilon}\frac{\partial h^{(n-1,\alpha)}}{\partial t^{\sigma}}, \quad h^{(0,\alpha)} := t_{\alpha} = \eta_{\alpha\varepsilon}t^{\varepsilon}, \quad n\geq 0.$$ Thus starting from the Casimirs $t_\alpha$ one may construct the corresponding deformed flat coordinates recursively. From Frobenius Manifolds to Equations of Hydrodynamic Type ---------------------------------------------------------- Let ${\mathcal M}$ be a Frobenius manifold with local coordinates $(u^1, ... , u^N)$. We define on the the loop space of ${\mathcal M}$ a Poisson bracket of hydrodynamic type $$\label{hydrobrkt1} \{ H_{(n,\varepsilon)} , H_{(m,\nu)} \} = \int_{S^1} \frac{\delta H_{(n,\varepsilon)}}{\delta u^\alpha} \left( \eta^{\alpha\beta}\frac{\partial~}{\partial X} -\eta^{\alpha\sigma}\Gamma_{\sigma\kappa}^\beta \, t^{\kappa}_{,X} \right) \frac{\delta H_{(m,\nu)}}{\delta u^{\beta}} dX,$$ where the functional densities for $H_{(n,\varepsilon)}$ depend on $u$ and not its derivatives. Indeed, we define $$\label{conserv1} H_{(n,\varepsilon)}= \int_{S^1} h^{ (n,\varepsilon) }(u)dX.$$ The bracket may be used to define the so-called principal hierarchy of the Frobenius manifold $$\frac{\partial u^{\alpha}}{\partial T_{(n,\kappa)}} = \{ u^{\alpha} , H_{(n,\kappa)} \}.$$ In the flat coordinate system $(t^1, ... , t^N)$ this becomes $$\frac{\partial t^{\alpha}}{\partial T_{(n,\kappa)}} = \eta^{\alpha\beta}\frac{\partial~}{\partial X}\left( \frac{\partial h^{(n,\kappa)}}{\partial t^{\beta}} \right) = \eta^{\alpha\beta}\frac{\partial^2 h^{(n,\kappa)}}{\partial t^{\beta}\partial t^{\nu}}\frac{\partial t^{\nu}}{\partial X}.$$ One may use the relation to illuminate further how the properties of these equations depend precisely on the algebraic structure on the tangent spaces $T_t{\mathcal M}$ viz: $$\label{matrixform1} \frac{\partial t^{\alpha}}{\partial T_{(n,\kappa)}} = \eta^{\alpha\beta}c^{\sigma}_{\beta\lambda}\frac{\partial h^{(n-1,\kappa)}}{\partial t^{\sigma}} \frac{\partial t^{\lambda}}{\partial X}.$$ For future use we define $$M_{(n,\kappa)}({\bf t})^{\alpha}_{\hspace{0.1cm}{\lambda}}=\eta^{\alpha\beta}c^{\sigma}_{\beta\lambda}\frac{\partial h^{(n-1,\kappa)}}{\partial t^{\sigma}} \,.$$ and note these flow equations simplify when $\alpha=N\,,$ i.e. $$\frac{\partial t^N}{\partial T_{(n,\kappa)}} = \frac{\partial h^{(n-1,\kappa)}}{\partial X}\,. \label{PT}$$ Since $t_1=t^N$ is the Egorov potential of the metric $\eta\,,$ this shows that the potential is a conserved quantity of all the flows, in agreement with [@PT] in the more general setting of semi-Hamiltonian Egorov systems. Symmetries of the WDVV Equations -------------------------------- The space of solutions of the WDVV equation possesses certain natural symmetries. A symmetry of WDVV is a map $$F(t) \mapsto \hat{F}(\hat{t}), \quad t^{\alpha} \mapsto \hat{t}^{\alpha}, \quad \eta_{\alpha\beta} \mapsto \hat{\eta}_{\alpha\beta}$$ from one solution to another. In [@dubrovin1] two symmetries were defined: - Legendre-type transformations $S_\kappa$: This is defined by $$\begin{aligned} {\hat t}_\alpha & = & \partial_{t^\alpha} \partial_{t^\kappa} F(t)\,, \qquad ({\rm N.B.}\quad{\hat t}_\alpha=\eta_{\alpha\beta}t^\beta)\\ \frac{\partial^2{\hat F}}{\partial{\hat t}^\alpha \partial{\hat t}^\beta} & = & \frac{\partial^2{ F}}{\partial{ t}^\alpha \partial{ t}^\beta} \,,\\ {\hat\eta}_{\alpha\beta} & = & \eta_{\alpha\beta}\,.\end{aligned}$$ - Inversion Symmetry: This is defined by $$\hat{t}^1 = \frac{1}{2}\frac{t_{\sigma}t^{\sigma}}{t^N}, \quad \hat{t}^{\alpha} = \frac{t^{\alpha}}{t^N}, \quad\mbox{ (for }\alpha \neq 1, N), \quad \hat{t}^N = -\frac{1}{t^N},$$ $$\label{isdef} \hat{F}(\hat{t}) = (\hat{t}^N)^2 F \left( \frac{1}{2}\frac{\hat{t}_{\sigma}\hat{t}^{\sigma}}{\hat{t}^N}, -\frac{\hat{t}^{2}}{\hat{t}^N}, ... , -\frac{\hat{t}^{N-1}}{\hat{t}^N}, -\frac{1}{\hat{t}^N} \right) + \frac{1}{2}\hat{t}^1\hat{t}_{\sigma}\hat{t}^{\sigma},$$ $$\hat{\eta}_{\alpha\beta} = \eta_{\alpha\beta}.$$ In this paper we will concentrate on inversion symmetries (also called type II symmetries). For these, the structure constants of the inverted Frobenius manifold are related to those of the original by $$\label{ctrans1} \hat{c}_{\alpha\beta\gamma} = (t^N)^{-2}\frac{\partial t^{\lambda}}{\partial \hat{t}^{\alpha}}\frac{\partial t^{\mu}}{\partial \hat{t}^{\beta}}\frac{\partial t^{\nu}}{\partial \hat{t}^{\gamma}}c_{\lambda\mu\nu}.$$ Under the assumption that that $r_i=0$ for $i=1\,,\ldots\,,N$ the corresponding Euler vector field of the inverted Frobenius manifold $\hat{F}$ have the form $$\hat{E}(\hat{t}) = \sum_{\alpha}(1-\frac{\hat{d}}{2}-\hat{\mu}_{\alpha})\hat{t}^{\alpha}\hat{\partial}_{\alpha},$$ where $$\hat{d}=2-d, \quad \hat{\mu}_1 = \mu_N - 1, \quad \hat{\mu}_N = \mu_1 + 1, \quad \hat{\mu}_i = \mu_i, \quad \mbox{for }i \neq 1, N.$$ A Frobenius manifold is said to be modular if it lies at a fixed point of the inversion symmetry, $$\hat{d} = d\,,\qquad \hat{\mu}_i = \mu_i\,.$$ Comparison of the two Euler fields gives the following: A Frobenius manifold is modular if and only if $d=1$ and $r^N=0.$ Examples of such modular Frobenius manifolds were given in the introduction. The application of $I$ to such a modular manifold does not yield a new prepotential but rather a modular tranformation of itself, i.e. $$F(\hat{t}) = \left({\hat{t}}^N\right)^2 F\left(t({\hat{t}})\right) + \frac{1}{2} \hat{t}^1 {\hat{t}}_\sigma {\hat{t}}^\sigma\,. \label{invariantF}$$ \[basicexample\] The prepotential $$F=\frac{1}{2} u^2 \tau - \frac{1}{2} u ({\bf z},{\bf z}) + f({\bf z},\tau)$$ (where $({\bf x},{\bf y})$ is some inner product) satisfies under the action of $I$ if and only if $$f\left(\frac{{\bf z}}{\tau},-\frac{1}{\tau}\right) = \frac{1}{\tau^2} f({\bf z},\tau) - \frac{1}{4 \tau^3} ({\bf z},{\bf z})^2\,.$$ In one dimension the ansatz $f(z,\tau)=z^4 \gamma(\tau)$ reduced this condition to $$f\left(-\frac{1}{\tau}\right) = \frac{1}{\tau^2} f(\tau) - \frac{1}{4} \tau\,.$$ This is the transformation property of the solution to Chazy’s equation. Note that the WDVV equations still have to be satisfied - this argument just gives the modular transformation properties that the solution must satisfy. Remark. There exists another choice of coordinates on a semi-simple Frobenius manifold, namely one that simplifies its algebraic structure. In these coordinates that multiplication is trivial: $$\frac{\partial~}{\partial u^i} \circ \frac{\partial~}{\partial u^j} = \delta_{ij} \frac{\partial~}{\partial u^i}\,.$$ It turns out that these coordinates are the roots of $$\label{canondef} \det\left(g^{\alpha\beta}(t)-u\,\eta^{\alpha\beta}\right)=0\,,$$ where $g$ is the intersection form of the Frobenius manifold. The roots of the expression are invariant under the symmetry $I$, and so the canonical coordinates are preserved up to a re-ordering [@dubrovin1]. Inversion and Almost Duality ============================ Consider the vector field $E^{-1}$ defined by the condition $$E^{-1} \circ E = e\,.$$ This is defined on $M^{\star}= M\backslash \Sigma\,,$ where $\Sigma$ is the discriminant submanifold on which $E^{-1}$ is undefined. With this field one may define a new dual multiplication $\star: TM^\star \times TM^\star \rightarrow TM^\star$ by $$X \star Y = E^{-1} \circ X \circ Y\,, \qquad\qquad \forall\, X\,,Y \in TM^\star\,.$$ This new multiplication is clearly commutative and associative, with the Euler vector field being the unity field for the new multiplication. Furthermore, this new multiplication is compatible with the intersection form $g$ on the Frobenius manifold, i.e. $$g(X\star Y, Z) = g(X,Y\star Z)\,, \qquad\qquad \forall\, X\,,Y\,,Z \in TM^\star\,.$$ Here $g$ is defined by the equation $$g(X,Y)=\eta(X\circ Y, E^{-1})\,, \qquad\qquad \forall\, X\,,Y \in TM^\star$$ (and hence is well-defined on $M^\star\,$). Alternatively one may use the metric $\eta$ to extend the original multiplication to the cotangent bundle and define $$g^{-1}(x,y) = \iota_E(x\circ y) \,, \qquad\qquad \forall\, x\,,y \in T^\star M^\star\,.$$ The intersection form has the important property that it is flat, and hence there exists a distinguished coordinate system $\{ p^i\,,i=1\,,\ldots\,,N\}$ in which the components of the intersection form are constant. It may be shown that there exists a dual prepotential $F^\star$ such that its third derivatives give the structure functions $c^{\star}_{ijk}$ for the dual multiplication. More precisely [@dubrovin2]: Given a Frobenius manifold $M$, there exists a function $F^\star$ defined on $M^\star$ such that: $$\begin{aligned} c^{\star}_{ijk} & = & g\left( \frac{\partial~}{\partial p^i}\star \frac{\partial~}{\partial p^j}\,, \frac{\partial~}{\partial p^k} \right) \,,\\ & = &\frac{\partial^3 F^\star}{\partial p^i \partial p^j \partial p^k}\,.\end{aligned}$$ Moreover, the pair $(F^\star,g)$ satisfy the WDVV-equations in the flat coordinates $\{ p^i \}$ of the metric $g\,.$ Thus given a specific Frobenius manifold one may construct a dual solution to the WDVV-equations by constructing the flat-coordinates of the intersection form and using the above result to find the tensor $c^{\star}_{ijk}$ from which the dual prepotential may be constructed. Recall the following: A function $\mathrm{p}=\mathrm{p}(t;\lambda)$ is called a $\lambda$-period of the Frobenius manifolds if it satisfies the Gauss-Manin equations $$\left({}^g \nabla \,\,-\lambda \,\,{}^{\eta} \nabla \right) d\mathrm{p}=0\,.$$ The following fact were proved in [@dubrovin2]: - $p=\mathrm{p}(t;0)$ is a flat coordinate for the intersection form $g\,;$ - $\mathrm{p}(t;\lambda) = p(t^1-\lambda\,,t^2\,,\ldots\,,t^N)\,.$ To understand the action of $I$ in the dual picture it is first necessary to understand the relationship between the two $\lambda$ periods $\hat{\mathrm{p}}$ and $\mathrm{p}$, or just the relationship between the flat coordinates of $g$ and $\hat{g}\,.$ \[basiclemma\] Let $p$ be a flat coordinate for $g$ with the property $E(p)=\left(\frac{1-d}{2}\right)p\,.$ Then $${\hat p}=\frac{p}{t_1}$$ is a flat coordinate for ${\hat g}\,.$ Moreover $${\hat{\mathrm{p}}}=\frac{\mathrm{p}}{t_1}$$ Recall that a flat coordinate $p$ for the metric $g$ must satisfy the equation ${}^g\nabla dp=0\,.$ Thus one just has to calculate ${}^{\hat{g}}\nabla d{\hat{p}}.$ In canonical coordinates $$\eta=\sum_{i=1}^N H_i^2({\bf u}) du_i^2\,, \qquad g=\sum_{i=1}^N \frac{H_i^2({\bf u})}{u_i} du_i^2$$ with Egorov potential $t_1\,,$ i.e. $H_i^2=\partial_i t_1\,.$ In terms of the dual objects, ${\hat{u}}_i=u_i$ and $${\hat{H}}_i=\frac{H_i}{t_1}\,, \qquad \hat{t}_1=-\frac{1}{t_1}$$ and the rotation coefficients satisfy the relation ${\hat{\beta}}_{ij}=\beta_{ij} - \frac{H_i H_j}{t_1}\,.$ Hence the connections ${}^{\hat{g}}\nabla$ and ${}^{{g}}\nabla$ can be expressed in terms of the Darboux-Egorov data for $\eta$ and the result follows via straightforward calculations. We now move from a flat coordinate to flat coordinate systems. It is here that the differences between the cases $d\neq 1$ and $d=1$ becomes apparent. \[dneqone\] Let $\{p^i\,:i=1\,,\ldots\,,N\}$ be a flat coordinate system for $g\,.$ If $d\neq 1$ then $$\left\{{\hat{p}}^i=\frac{p^i}{t_1} \,:i=1\,,\ldots\,,N\right\}$$ is a flat coordinate system for ${\hat{g}}\,.$ Moreover the normalization conditions are preserved, and the Gram matrices coincide. In [@dubrovin2] it was shown that if $d\neq 1$ then $p$ may be normalized so $E(p)=\left(\frac{d-1}{2}\right)p\,,$ so the above Lemma shows that the ${\hat{p}}^i$ are flat. To show that they form a coordinate [system]{} one must calculate the Jacobian of the transformation. A simple calculation gives $$\frac{\partial({\hat{p}^1}\,,\ldots\,,{\hat{p}}^N)}{\partial(p^1\,,\ldots\,,p^N)} = - \frac{1}{t_1^{N+1}}$$ on using the result [@dubrovin2] (again for $d\neq 1$) that $t_1=\left( \frac{1-d}{2} \right) g_{ab}p^a p^b\,.$ The calculation of the dual normalization conditions and Gram matrices is also straightforward and yields the relation ${\hat{g}}_{ab}=g_{ab}\,.$ When $d=1$ certain special results hold. Suppose $d=1\,.$ Then: - ${}^g \nabla E=0\,;$ - $t_1$ is a flat coordinate for both $\eta$ and $g\,.$ The proof is by direct computation. For example, in flat coordinates $\{ t^i\,,i=1\,,\ldots\,,N\}$ for $\eta$ the Christoffel symbols for $g$ take the form [@dubrovin1; @dubrovin2] $${}^g \Gamma^{\alpha\beta}_\gamma = c^{\alpha\epsilon}_\gamma \left( \frac{1}{2} - \mathcal{V}\right)^\beta_\epsilon$$ and using this one finds that ${}^g\nabla_\alpha E^\beta=\left(\frac{1-d}{2}\right) \delta^\beta_\alpha\,.$ Alternatively, in canonical coordinates, $$\begin{aligned} {}^g \nabla_i E^i & = & 0 \,, \qquad i\neq j\,,\\ {}^g \nabla_i E^i & = & \frac{(1-d)}{2} \,,\qquad ({\rm no~sum})\,.\end{aligned}$$ The proof that $t_1$ is flat for $g$ (i.e. ${}^g\nabla dt_1=0$) is similar. Recall that $E$ plays the role of the unity vector field in the dual picture. Thus when $d=1$ the unity vector field is covariantly constant, so almost dual Frobenius manifolds at $d=1$ are even closer to Frobenius manifolds than for those with $d\neq1\,.$ We will return to this point later. Suppose $d=1$ and $r_N=0$ and let $\{p^i\,:i=1\,,\ldots\,,N\}$ be a flat coordinate system for $g\,.$ Let $$\hat{p}^1 = \frac{1}{2}\frac{p_{\sigma}p^{\sigma}}{p^N}, \quad \hat{p}^{\alpha} = \frac{p^{\alpha}}{p^N}, \quad\mbox{ (for }\alpha \neq 1, N), \quad \hat{p}^N = -\frac{1}{p^N},$$ where $p^N=t_1\,.$ Then $\{{\hat{p}}^i\,:i=1\,,\ldots\,,N\}$ are a flat coordinate system for ${\hat{g}}\,.$ From the above lemma $t_1$ is a flat coordinate for $g$ and hence we choose $p^N=t_1\,.$ With this $E(p^N)=0$ and since ${}^g\nabla E=0\,,$ the vector field $E$ must take the form $E=\sum_{i=1}^{N-1} c_i \frac{\partial~}{\partial p^i}$ for some constants $c_i\,.$ Using the freedom to redefine the $p^i$ for $i\neq N$ one may set $$E=\frac{\partial~}{\partial p^1}\,.$$ With this $g(E,E)=\eta(E^{-1} \circ E,E) = \eta(e,E)=r_N$ (again since $d=1$). Thus from [@dubrovin1] Lemma 1.1 one may redefine coordinates so $$g_{ab}=\delta_{a+b,N+1}$$ in the $\{p^i\}$-coordinates. Since $E(p^i)=0.p^i$ for $i=2\,,\ldots\,,N-1$ one may use Lemma \[basiclemma\] to show that ${\hat{p}}^i=p^i/p^N$ are flow coordinates. By direct calculation - for example, in canonical coordinates - one may show that ${\hat{p}}^1$ and ${\hat{p}}^N$ define above are also flat coordinates (again, one just has to show ${}^{\hat{g}} \nabla d{\hat p}=0\,$). It also follows immediately that ${\hat{g}}_{ab}=g_{ab}\,.$ It remains to compare the two induced prepotentials $F^\star$ and ${\hat{F^\star}}$ and the corresponding multiplications. Let $F$ define a Frobenius manifold and let $\hat{F}$ denote the induced manifold under the action of the symmetry $I\,.$ Let $F^\star$ and $\hat{F^\star}$ denote the corresponding almost dual structures. The $I^\star$, the induced symmetry act as: - [Case I: $d \neq 1\,:$]{} $$\begin{aligned} {\hat p}^i & = & \frac{p^i}{t_1} \,, \qquad i=1\,,\ldots\,, N\,,\\ {\hat{g}}_{ab} & = & g_{ab} \,, \\ {\hat{F^\star}}({\hat{\mathbf{p}}}) & = & \frac{F^\star(\mathbf{p}({\hat{\mathbf{p}}}))}{t_1^2}\end{aligned}$$ where $t_1=\frac{(1-d)}{2} g_{ab} p^a p^b\,.$ - [Case II: $d=1\,:$]{} $$\begin{aligned} {\hat p}^1 & = & \frac{1}{2}\frac{p_{\sigma}p^{\sigma}}{t_1}\,,\quad {\hat p}^i = \frac{p^i}{t_1} \,, \qquad i=2\,,\ldots\,, N-1\,,\quad {\hat p}^N = -\frac{1}{t_1}\,,\\ {\hat{g}}_{ab} & = & g_{ab} \,, \\ {\hat{F^\star}}(\hat{\mathbf{p}}) & = & (\hat{p}^N)^2 F \left( \mathbf{p}({\hat{\mathbf{p}}}) \right) + \frac{1}{2}\hat{p}^1\hat{p}_{\sigma}\hat{p}^{\sigma}\,,\end{aligned}$$ where $t_1=p^N\,.$ Note, in both cases $p^N$ is the Egorov potential for the metric $\eta\,.$ Recall [@dubrovin2] that for $d\neq 1$ the dual prepotential satisfies the homogeneity condition $$\sum_\alpha p^\alpha \frac{\partial F^\star}{\partial p^\alpha} = 2 F^\star + \frac{1}{1-d} g_{\alpha\beta} p^\alpha p^\beta\,.$$ Using this and the explicit coordinates given in Proposition \[dneqone\] one finds that $$\frac{\partial^3 {\hat{F^\star}}}{\partial{\hat{p}}^\alpha \partial{\hat{p}}^\beta \partial{\hat{p}}^\gamma} = t_1 \frac{\partial^3 {{F^\star}}}{\partial{{p}}^\alpha \partial{{p}}^\beta \partial{{p}}^\gamma} -\frac{2}{1-d} (g_{\alpha\beta} p_\gamma+g_{\alpha\gamma} p_\beta+g_{\beta\gamma} p_\alpha) + \frac{2}{t_1(1-d)} p_\alpha p_\beta p_\gamma\,.$$ From this it is straightforward to show that $\hat{F^\star}$ satisfies the WDVV equations in the $\{ p^i\}$-variables. If $d=1$ the proof is identical to the original inversion symmetry as presented in [@dubrovin1]. It remains to show how the deformed flat coordinates of both ${}^\eta {\widetilde{\nabla}}$ and ${}^g {\widetilde{\nabla}}$ behave under inversion symmetry. This will be done in the next section. We will end this section with some examples. Given an irreducible Coxeter group $W$ of rank $N\,,$ the Saito construction gives a Frobenius manifold structure on the orbit space $\mathbb{C}^N/W\,.$ The almost dual prepotential takes the form $$F^\star({\bf p}) = \frac{1}{4} \sum_{\alpha\in R_W} (\alpha,{\bf p}) \log (\alpha,{\bf p})^2$$ where $(,)$ is the metric $g\,.$ Application of the $I^\star$ transform (recall $d\neq 1$ for these examples) yields the solution $${\hat{F^\star}}({\hat{\bf p}}) = \frac{1}{4} \sum_{\alpha\in R_W} (\alpha,{\bf {\hat{p}}}) \log (\alpha,{\hat{{\bf p}}})^2 - \frac{h}{4} ({\hat{{\bf p}}},{\hat{{\bf p}}})\log({\hat{{\bf p}}},{\hat{{\bf p}}})$$ where $h$ is defined by the relation $\sum_{\alpha\in R_W} (\alpha,{\bf z}) = h ({\bf h},{\bf h})$ (and hence depends on the normalization of the roots $\alpha\in R_W\,).$ Thus the original solution is recovered but with the addition of a new radial term. Such solution have been constructed directly (i.e. without knowledge of its geometric origins) in [@LP]. The transformation property hold more generally than just for dual solutions, and hence may also be applied to the $\vee$-systems [@Veselov]. Given the Weyl group $A_N$ and $B_n$ with Lie algebra $\mathfrak{g}$ with Cartan subalgebra $\mathfrak{h}$ one may construct the so-called Jacobi group $J(\mathfrak{h})$ and orbit space $\Omega/J(\mathfrak{g})$ where $\Omega=\mathbb{C}\oplus {\mathfrak h} \oplus \mathbb{H}\,.$ This orbit space carries the structure of a Frobenius manifold and the dual prepotential takes the form $$\begin{aligned} F^\star(u\,,{\bf z}\,,\tau) & = & \frac{1}{2} \tau u^2 - \frac{1}{2} u ({\bf z},{\bf z})+\sum_{\alpha\in\mathfrak{U}} h(\alpha.{\bf z},\tau)\end{aligned}$$ Here the function $h$ is essentially the elliptic trilogarithm introduced by Beilinson and Levin [@BL; @Levin] and the set $\mathfrak{U}$ contains certain vectors - an elliptic generalization of classical root systems. The basic function $h$ satisfies the modularity property (c.f. example \[basicexample\]) $$h\left( \frac{z}{\tau} \right) = \frac{1}{\tau^2} h(z,\tau) - \frac{z^4}{4!\,\tau^3}$$ up to quadratic terms. The proof that this satisfies the inversion symmetry may be found in [@iabs2]. Inversion Symmetry and Principal Hierarchies {#mainresult} ============================================ A deformed flat coordinate $\mathfrak{t}$ must satisfy the equation ${}^\eta \widetilde{\nabla} d\mathfrak{t}=0$ and similarly a deformed flat coordinate $\mathfrak{p}\,,$ also known as twisted period, must satisfy the equation ${}^g \widetilde{\nabla} d\mathfrak{p}=0\,,$ where ${}^g {\widetilde\nabla}_X Y = {}^g \nabla_X Y - \nu X \star Y$ is the dual deformed connection. It is straightforward to prove, using canonical coordinates, that $$\hat{\mathfrak{t}}=\frac{\mathfrak{t}}{t_1} \qquad {\rm and~}\qquad {\hat{\mathfrak{p}}}=\frac{\mathfrak{p}}{t_1}$$ are corresponding dual deformed flat coordinates (and one follows from the other by the contour integral/Laplace transform methods in [@dubrovin2]). However, a more subtle form appears when the calculations are performed in flat coordinates - a shift appears in the labels. We begin by examining the relationship between $\mathfrak{t}$ and $\hat{\mathfrak{t}}$ since the coefficients will give the transformation between the Hamiltonian densities of the various flows. These results has also been proved independently in [@zhang]. \[propzero\] Under the inversion symmetry, the Hamiltonian densities of the principal hierarchy corresponding to $F$ are mapped to those of $\hat{F}$ according to the following rules $$\label{htrans1} h^{(n,\alpha)} (\mathbf{t}(\hat{\mathbf{t}})) = -\frac{1}{\hat{t}^N}\hat{h}^{(\tilde{n},\tilde{\alpha})}(\hat{\mathbf{t}}),$$ where $$\tilde{n} = \left\{ \begin{array}{ll} n+1, & \mbox{if } \alpha = N, \\ n, & \mbox{if }\alpha \neq 1, N, \\ n-1, &\mbox{if }\alpha = 1, \end{array}\right. \tilde{\alpha} = \left\{ \begin{array}{ll} 1, & \mbox{if } \alpha = N, \\ \alpha, & \mbox{if }\alpha \neq 1, N, \\ N, & \mbox{if }\alpha = 1. \end{array}\right.$$ This consists of inverting equation using and checking the solutions given in the proposition do in fact satisfy it. One may identify how the labels $\alpha$ are mapped using the fact that the unity field decreases the degree of the densities by one, $$e^{\tilde{n}}(h^{(\tilde{n},\tilde{\alpha})}) = t_{\tilde{\alpha}}.$$ Here $e^{\tilde{n}}$ denotes $\tilde{n}$ successive applications of the operator $e= \partial / \partial t^1$. This lifts in turn to the corresponding flows, leading to the following \[prop1\] Under the inversion symmetry, the flows on the loop space of the inverted Frobenius manifold are related to those of the original one via $$\label{flowtrans1} M_{(n,\alpha)}(\mathbf{t}(\hat{\mathbf{t}}))= -\hat{t}^N\hat{M}_{(\tilde{n},\tilde{\alpha})}(\hat{\mathbf{t}}) + \hat{h}^{(\tilde{n}-1,\tilde{\alpha})}(\hat{\mathbf{t}})\mathbf{1},$$ where $\tilde{n}$, $\tilde{\alpha}$ are as above. This is a calculation that follows from plugging the results of Proposition \[propzero\] into . Particularly, using equations , , and we can apply $I$ to to find $$\begin{gathered} \tag{2} \frac{\partial{\hat{t}}^{\rho}}{\partial T_{(n,\alpha)}} = -\underbrace{({\hat{t}}^N)^{-3} \frac{{\partial}{\hat{t}}^{\rho}}{{\partial}t^{\iota}} \hat{\eta}^{\iota\kappa} \hat{\eta}^{\mu\lambda} \frac{{\partial}{\hat{t}}^{\delta}}{{\partial}t^{\lambda}} \frac{{\partial}{\hat{t}}^{\gamma}}{{\partial}t^{\kappa}} \hat{c}_{\delta\gamma\xi}({\hat{t}})\frac{{\partial}{\hat{t}}^{{\varepsilon}}}{{\partial}t^{\mu}} \frac{{\partial}{\hat{h}}^{(\tilde{n}-1,\tilde{\alpha})}({\hat{t}})}{{\partial}t^{\mu}}\frac{\partial {\hat{t}}^{\xi}}{\partial X}}_{\mbox{term }1} \\ + \underbrace{({\hat{t}}^N)^{-2} \frac{{\partial}{\hat{t}}^{\rho}}{{\partial}t^{\iota}} \hat{\eta}^{\iota\kappa} \frac{{\partial}{\hat{t}}^{\gamma}}{{\partial}t^{\kappa}} \hat{\eta}_{\gamma\xi}{\hat{h}}^{(\tilde{n}-1,\tilde{\alpha})}({\hat{t}})\frac{\partial {\hat{t}}^{\xi}}{\partial X}}_{\mbox{term }2}.\end{gathered}$$ A straightforward but tedious calculation then gives simplification of the terms $$\mbox{term }1 = -{\hat{t}}^N \hat{\eta}^{\rho{\varepsilon}}\hat{c}^{\mu}_{{\varepsilon}\sigma}\frac{{\partial}{\hat{h}}^{(n-1,\alpha)}({\hat{t}})}{{\partial}{\hat{t}}^{\mu}}\frac{\partial {\hat{t}}^{\sigma}}{\partial X} \qquad \mbox{term }2= \delta^{\rho}_{\sigma}{\hat{h}}^{(n-1,\alpha)}({\hat{t}})\frac{\partial {\hat{t}}^{\sigma}}{\partial X}.$$ Proposition is proved. Naturally, one is led to consider the analogous problem in the almost dual picture. This gives the following Assume $d \neq 1$. Let $F^\star$ and $\hat{F}^\star$ be related by the almost dual inversion symmetry, $I^\star$. Then the successive approximations to the flat coordinates for the corresponding deformed connections $^g\tilde{\nabla}$ and $^{\hat{g}}\tilde{\nabla}$ are related via $$\label{ptrans1} l^{(n,\alpha)}(\mathbf{p}(\hat{\mathbf{p}})) = \frac{1}{\hat{t}_1}\hat{l}^{(n,\alpha)}(\hat{\mathbf{p}}),$$ where the approximations $l^{(n,\alpha)}$ are defined via $$\mathfrak{p}^{\alpha}(\mathbf{p},\nu) = \sum_{n=0}^{\infty} \nu^n l^{(n,\alpha)}(\mathbf{p})$$ This gives rise to the corresponding relationship between the almost dual flows $$\label{adflowtrans1} M_{(n,\alpha)}(\mathbf{p}(\hat{\mathbf{p}})) = \hat{t}_1\hat{M}_{(n,\alpha)}(\hat{\mathbf{p}}) - \frac{2}{1-d}\hat{l}^{(n-1,\alpha)}(\hat{\mathbf{p}})\mathbf{1}.$$ If $d=1$ the transformation is the same as in . This is carried out in an analogous manner to the above using $$\frac{\partial^3 {\hat{F^\star}}}{\partial{\hat{p}}^\alpha \partial{\hat{p}}^\beta \partial{\hat{p}}^\gamma} = t_1 \frac{\partial^3 {{F^\star}}}{\partial{{p}}^\alpha \partial{{p}}^\beta \partial{{p}}^\gamma} -\frac{2}{1-d} (g_{\alpha\beta}p_\gamma+g_{\alpha\gamma}p_{\beta}+g_{\beta\gamma}p_{\alpha}) + \frac{2}{t_1(1-d)} p_\alpha p_\beta p_\gamma\,.$$ One must also make use of certain homogeneity conditions, for example $$p^{\sigma}\frac{\partial^3F^\star}{\partial p^{\sigma} \partial p^{\varepsilon} \partial p^{\kappa}} = \frac{2g_{\kappa\varepsilon}}{1-d}.$$ At first sight the result is surprising: the flows associated to $F$ are not mapped under $I$ to the flows associated to $\hat{F}\,.$ However a simple reciprocal transformation $$\begin{aligned} d{\tilde{X}} & = & t^N dX + h^{({\tilde{n}}-1,{\tilde{\kappa}})} dT_{({\tilde{n}},{\tilde{\kappa}})}\,,\\ d{\tilde{T}_{({\tilde{n}},{\tilde{\kappa}})}} & = & dT_{({\tilde{n}},{\tilde{\kappa}})}\end{aligned}$$ (recall that the Egorov potential $t^N$ is conserved by (\[PT\])) transform the flow to that associated with ${\hat{F}}\,.$ Thus: Up to a simple reciprocal transformation, $I$ maps flows of $F$ to flows of $\hat{F}\,.$ Moreover, if the Frobenius manifold is modular, $I$ maps the flows to themselves, i.e. the flows are invariant. For semi-simple Frobenius manifolds these flow take diagonal form when written in terms of canonical coordinates (which are specific examples of Riemann invariants): $$\frac{\partial u^i}{\partial T_{(n,k)}} = \lambda^i_{(n,k)}({\bf u}) \frac{\partial u^i}{\partial X}\,.$$ Since the characteristic velocities are the eigenvalues of the tensor $M_{(n,k)}$ one can easily show that under inversion $${\hat \lambda}^i_{(n,k)} = t^N \lambda^i_{(n,k)} - h_{(n,k)}\,$$ and thus by applying a reciprocal transformation the system, when written in Riemann invariant form, is unchanged. This same result holds in the weaker setting of semi-Hamiltonian systems where the metric is Egorov [@PT]. A class of such examples may be found by restricting flows to certain natural submanifolds of a Frobenius manifold [@iabs1; @FV]. We end this section with an extended example which shows how the modularity properties of solutions to the Chazy equation results in, up to a reciprocal transformation, flows that are invariant under the modular transformations. Consider the Frobenius manifold defined by the free energy $$F = \frac{1}{2}t_1^2t_3+\frac{1}{2}t_1t_2^2 - \frac{t_2^4}{16}\gamma(t_3)\,; \quad E=t_1\frac{\partial}{\partial t_1} + \frac{1}{2}t_2\frac{\partial}{\partial t_2},$$ where $\gamma$ is some unknown 1-periodic function. For the duration of this example all coordinates will be written with lowered indices ($t_{\alpha} = t^{\alpha}$). In order for $F$ to satisfy WDVV, $\gamma$ must satisfy Chazy’s equation, $$\gamma'''(t_3)=6\gamma(t_3)\gamma''(t_3)-9(\gamma'(t_3))^2.$$ The main property of the Chazy equation is an $SL(2,{\mathbb Z})$ invariance: $$\begin{aligned} t_3 & \mapsto & \frac{at_3 + b}{ct_3 + d}, \quad ad-bc =1, \nonumber \\ \gamma(t_3) & \mapsto & (ct_3 + d)^2\gamma(t_3)+2c(ct_3 + d).\end{aligned}$$ This in turn allows us to apply the inversion symmetry. Since $d=1$ and $r^N=0$ this example is one of a class of examples of Frobenius manifolds that lie at fixed points of the inversion symmetry, and hence defines a modular Frobenius manifold. Consider the following solution to equation \[gm1\], $$h_{ (0,2) } = t_2.$$ Applying the inversion symmetry, we find $$h_{ (0,2) }(\hat{t}_1+\frac{1}{2}\frac{\hat{t}_2^2}{\hat{t}_3}, -\frac{\hat{t}_2}{\hat{t}_3}, -\frac{1}{\hat{t}_3}) = -\frac{\hat{t}_2}{\hat{t_3}} = -\frac{1}{\hat{t_3}}h_{(0,2)}(\hat{t}_1,\hat{t}_2,\hat{t}_3),$$ since $F$ lies at a fixed point of the inversion symmetry. Similarly, considering $$h_{ (1,1) } = t_1t_3+\frac{1}{2}t_2^2$$ one finds $$\begin{split} h_{ (1,1) } (\hat{t}_1+\frac{1}{2}\frac{\hat{t}_2^2}{\hat{t}_3}, -\frac{\hat{t}_2}{\hat{t}_3}, -\frac{1}{\hat{t}_3}) = \left( \hat{t}_1+\frac{1}{2}\frac{\hat{t}_2^2}{\hat{t}_3} \right)\left( -\frac{1}{\hat{t}_3} \right)+\frac{1}{2}\left( -\frac{\hat{t}_2}{\hat{t}_3} \right) = -\frac{1}{\hat{t}_3}\hat{t}_1 = -\frac{1}{\hat{t}_3}\hat{h}_{(0,3)}. \end{split}$$ The following depicts how the densities are mapped under $I$: (100,80)(-15,50) (0,50) (0,60) (79,60)[(1,0)[10]{}]{} (3,51)[(1,0)[45]{}]{} (3,50)[(1,-1)[47]{}]{} (50,3)[(-1,1)[47]{}]{} (50,50) (53,51)[(1,0)[45]{}]{} (53,50)[(1,-1)[47]{}]{} (100,3)[(-1,1)[47]{}]{} (100,50) (23,34) (73,34) (-15,25) (0,25) (-1,20) (5,18) (3,26)[(1,0)[45]{}]{} (50,25) (49,20) (55,18) (53,26)[(1,0)[45]{}]{} (100,25) (99,20) (105,18) (-15,0) (0,0) (-4,-6) (3,1)[(1,0)[45]{}]{} (50,0) (50,-6) (53,1)[(1,0)[45]{}]{} (100,0) (96,-6) The flow corresponding the the density $h_{(0,2)}$ is $$\frac{\partial}{\partial T_{(1,2)}}\left( \begin{array}{c} t_1 \\ t_2 \\ t_3 \end{array}\right) = \left( \begin{array}{ccc} 0 & \frac{-3}{4}t_2^4\gamma'(t_3) & -\frac{3}{4}t_2^3\gamma''(t_3) \\ 1 & -\frac{3}{2}\gamma(t_3)t_2 & -\frac{3}{4}t_2^2\gamma'(t_3) \\ 0 & 1 & 0 \end{array}\right) \frac{\partial}{\partial X}\left( \begin{array}{c} t_1 \\ t_2 \\ t_3 \end{array} \right),$$ Inverting this one finds $$\begin{aligned} \frac{\partial}{\partial T_{(1,2)}}\left( \begin{array}{c} {\hat{t}}_1 \\ {\hat{t}}_2 \\ {\hat{t}}_3 \end{array} \right) &=& \left( \begin{array}{ccc} {\hat{t}}_2 & \frac{3}{4}{\hat{t}}_2^2{\hat{t}}_3\gamma'({\hat{t}}_3) & \frac{1}{4}{\hat{t}}_2^3{\hat{t}}_3\gamma''({\hat{t}}_3) \\ -{\hat{t}}_3 & {\hat{t}}_2 + \frac{3}{2}\gamma({\hat{t}}_3){\hat{t}}_2{\hat{t}}_3 & \frac{3}{4}{\hat{t}}_2^2{\hat{t}}_3\gamma'({\hat{t}}_3) \\ 0 & -{\hat{t}}_3 & {\hat{t}}_2 \end{array} \right)\frac{\partial}{\partial X}\left( \begin{array}{c} {\hat{t}}_1 \\ {\hat{t}}_2 \\ {\hat{t}}_3 \end{array} \right) \nonumber \\ &=&\left( \left( \begin{array}{ccc} 0 & \frac{-3}{4}{\hat{t}}_2^4\gamma'({\hat{t}}_3) & -\frac{3}{4}{\hat{t}}_2^3\gamma''({\hat{t}}_3) \\ 1 & -\frac{3}{2}\gamma({\hat{t}}_3){\hat{t}}_2 & -\frac{3}{4}{\hat{t}}_2^2\gamma'({\hat{t}}_3) \\ 0 & 1 & 0 \end{array}\right) + {\hat{t}}_2\mathbf{1} \right) \frac{\partial}{\partial X}\left( \begin{array}{c} {\hat{t}}_1 \\ {\hat{t}}_2 \\ {\hat{t}}_3 \end{array} \right) \nonumber \\ &=& \left( -{\hat{t}}_3\hat{M}_{1,2}(\hat{\mathbf{t}}) + h_{(0,2)}(\hat{\mathbf{t}})\mathbf{1}\right) \frac{\partial}{\partial X}\left( \begin{array}{c} {\hat{t}}_1 \\ {\hat{t}}_2 \\ {\hat{t}}_3 \end{array} \right), \nonumber\end{aligned}$$ as predicted by . Similarly, one can see how the flows corresponding to $\alpha=1$ and $\alpha=3$ are related. The flow corresponding to $h_{(1,1)}$ is $$\frac{\partial}{\partial T_{(2,1)}} \left( \begin{array}{c} t_1 \\ t_2 \\ t_3 \end{array} \right) = \left( \begin{array}{ccc} t_1 & -\frac{3}{4}t_2^3\gamma'(t_3) - \frac{1}{4}t_2^3 t_3 \gamma''(t_3) & -\frac{1}{4}t_2^4\gamma''(t_3)-\frac{1}{16}t_2^4t_3\gamma'''(t_3) \\ t_2 & t_1 - \frac{3}{2}\gamma(t_3)t_2^2-\frac{3}{4}t_2^2t_3\gamma(t_3) & -\frac{3}{4}t_2^3\gamma'(t_3) - \frac{1}{4}t_2^3t_3\gamma''(t_3) \\ t_3 & t_2 & t_1 \end{array} \right)\frac{\partial}{\partial X}\left( \begin{array}{c} t_1 \\ t_2 \\ t_3 \end{array} \right).$$ This may be inverted to give $$\begin{aligned} \frac{\partial}{\partial T_{(2,1)}}\left( \begin{array}{c} \hat{t}_1 \\ \hat{t}_2 \\ \hat{t}_3 \end{array} \right) &=& \left( \begin{array}{ccc} \hat{t}_1 & \frac{1}{4}\hat{t}_2^3\hat{t}_3\gamma''(\hat{t}_3) & \frac{1}{16}\hat{t}_2^4\hat{t}_3\gamma'''(\hat{t}_3) \\ 0 & \hat{t}_1 + \frac{3}{4}\hat{t}_2^2\hat{t}_3\gamma'(\hat{t}_3) & \frac{1}{4}\hat{t}_2^3\hat{t}_3\gamma''(\hat{t}_3) \\ -\hat{t}_3 & 0 & \hat{t}_1 \end{array} \right) \frac{\partial}{\partial X} \left( \begin{array}{c} \hat{t}_1 \\ \hat{t}_2 \\ \hat{t}_3 \end{array} \right) \nonumber \\ &=&\left( -\hat{t}_3 \left( \begin{array}{ccc} 0 & -\frac{1}{4}\hat{t}_2^3\hat{t}_2\gamma''(\hat{t}_3) & -\frac{1}{16}\hat{t}_2^4\gamma'''(\hat{t}_3) \\ 0 & -\frac{3}{4}\hat{t}_2^2 \gamma'(\hat{t}_3) & -\frac{1}{4}\hat{t}_2^3\gamma''(\hat{t}_3) \\ 1 & 0 & 0 \end{array} \right) + \hat{t}_1\mathbf{1} \right) \frac{\partial}{\partial X} \left( \begin{array}{c} \hat{t}_1 \\ \hat{t}_2 \\ \hat{t}_3 \end{array} \right) \nonumber \\ &=& (-\hat{t}_3\hat{M}_{(1,3)}(\hat{t})+ \hat{h}_{(0,3)}(\hat{t})\mathbf{1})\frac{\partial}{\partial X}\left( \begin{array}{c} {\hat{t}}_1 \\ {\hat{t}}_2 \\ {\hat{t}}_3 \end{array} \right).\nonumber\end{aligned}$$ Note the use, in the above example, of the modular transformation properties of solutions of Chazy’s equation and their derivatives. Comments ======== For Frobenius manifolds with $d=1$ the Euler field - the unity field for the dual multiplication - is covariantly constant with respect to the Levi-Civita connection of the intersection form $g\,.$ Thus in this case one is even closer to the axioms of a Frobenius manifolds than are the almost duality axioms. The dual prepotential must take the form $${\hat{F}}({\bf p})={\rm cubic~} + f(p^2\,,\ldots\,,p^N)$$ where the precise form of the cubic terms depends on the original Frobenius manifold. If $\eta_{11}=0$ then $g(E,E)=r_N$ and hence depends on whether $r_N$ is zero or not, by Lemma 1.1 in [@dubrovin1]. This fact has already been observed in the explicit calculation of dual prepotentials of various Hurwitz space Frobenius manifolds [@RS1; @RS2], and had been built into trigonometric ansatz for solutions of the WDVV equations [@MH; @misha]. The dual prepotentials of the extended-affine Weyl group orbit spaces should also fall into this class [@DZ]. What is new in this case (for modular Frobenius manifolds) is the invariance property of the function $f$ under $I$ as described in Example \[basicexample\]. This is a quite severe restriction on the function $f\,$ and it would be interesting to find further examples of functions with this property. Similarly, it would be of interest to find further examples of elliptic $\vee$-systems. Some of these results in this paper may also be obtained by other means. For the simple elliptic singularities ${\tilde{E}}_{6,7,8}$ the modular property of the twisted periods may be obtained directly from the superpotential/unfolding of the underlying singularity. For example, for ${\tilde{E}}_6$ the superpotential is $$\begin{aligned} f_t({\bf z},{\bf s}) & = & z_1^3 + z_2^3 + z_3^3 + s_8 z_1 z_2 z_3 \\ &&+ s_7 z_1 z_2 + s_6 z_1 z_3 + s_5 z_2 z_3 + s_4 z_1 + s_3 z_2 + s_2 z_3 + s_1\,,\end{aligned}$$ where the relationship between the $s_i$ and the flat coordinates $t^i$ have been found explicitly [@VW]. The twisted periods are given by contour integrals [@E] $$\mathfrak{p}(t,\nu) = \sqrt{h(t^8)} \oint_\gamma f_t^{\nu-1} dx\wedge dy\wedge dz$$ (note the primitive factor not present in the corresponding formulae for simple singularities). It was noticed in [@VW] that $f_t$ is invariant under the action of $I\,.$ The simple transformation property of $h$ then gives the required relation $$\hat{\mathfrak{p}}=\frac{\mathfrak{p}}{t^N}\,.$$ Finally, one may calculate how objects such as the isomonodromic $\tau$-functions (denoted $\tau_I$ to avoid confusion with the variable $\tau$) transform under inversion [@me]: $${\hat{\tau}}_I = \frac{\tau_I}{\sqrt{t_1}}\,.$$ For modular Frobenius manifolds this translates in the following modularity property for $\tau_I\,,$ $${\hat{\tau}}_I \left( \frac{{\bf{z}}}{\tau},-\frac{1}{\tau} \right) = \frac{\tau_I({\bf z},\tau)}{\sqrt{\tau}}\,.$$ For the Chazy example $\tau_I^{-48} = (t^2)^{12} \Delta^3(t^3)\,,$ where $\Delta$ is the classical discriminant of the underlying elliptic curve, and one may verify the transformation property by using the well-known modular transformation properties of the discriminant [@KS]. The next step is to calculate higher-order corrections to the dispersive deformations of these flows (a problem also raised, with some conjectures, in [@zhang]), and in particular the higher-order corrections to flows originating from a modular Frobenius manifold. One would hope that the modularity properties would carry over to the full dispersive hierarchy in some form (recall that a reciprocal transformation was required for a precise statement of invariance even at zero order). At first order this is entirely computational, on using the formulae in [@DZ2]: the $G$-function is know explicitly [@me; @KS; @iabs3] for a wide range of modular Frobenius manifolds. An understanding of the modularity properties of the Universal Loop Equation [@DZ3] or of approaches based on singularity theory [@GM] for these modular Frobenius manifolds would be of great interest and is potentially a way to construct new examples of dispersive integrable hierarchies. Acknowledgements {#acknowledgements .unnumbered} ================ IABS would like to acknowledge financial support from the British Council (PMI2 Research Co-operation award) and EM would like to thank the EPSRC for financial support. We both would like to thank Misha Feigin for various useful conversation and remarks. [99]{} Beilinson, A. and Levin, A.,[The Elliptic Polylogarithm,]{} in [Motives]{} (ed. Jannsen, U., Kleiman, S,. Serre, J.-P.), Proc. Symp. Pure Math. vol [55]{}, Amer. Math. Soc., (1994), Part 2, 123-190. Dingdian, X. and Zhang, Y., [On symmetries of the WDVV Equations]{}, arXiv:1002:00341 Dubrovin, B., [On almost duality for Frobenius manifolds]{} in [Geometry, topology, and mathematical physics]{}, 75–132, Amer. Math. Soc. Transl. Ser. 2, 212, Amer. Math. Soc., Providence, RI, 2004. Eguchi, T, Yamada Y. and Yang, S-K., [Topological Field Theories and the Period Integrals]{}, Feigin, M.V., [Trigonometric solutions of WDVV equations and generalized Calogero-Moser-Sutherland systems]{}, SIGMA [5]{} (2009), 088, 10 pages. Feigin, M.V. and Veselov, A.P., [Logarithmic Frobenius structures and Coxeter discriminants]{}, Advances in Mathematics [212:1]{} (2007) 143-162. Feigin, M.V. and Veselov, A.P., [On the geometry of $\vee$-systems]{}, in [Geometry, Topology, and Mathematical Physics]{}, Amer. Math. Soc. Transl. Ser. 2, Vol. 224, Amer. Math. Soc., Providence, RI, 2008, 111-123. Lechtenfeld, O. and Polovnikov, K., [A new class of solutions to the WDVV equation]{}, Phys. Lett. [A 374]{}, (2010), 504-506. Levin, A., [Elliptic polylogarithms: an analytic theory]{} Compositio Math. 106 (1997), no. 3, 267–282. Martini, R., and Hoevenaars, L.K., [Trigonometric Solutions of the WDVV Equations from Root Systems]{}, Lett.Math.Phys. [65]{} (2003) 15-18. Pavlov, M.V. and Tsarev, S.P.,[Tri-Hamiltonian Structures of Egorov Systems of Hydrodynamic type]{}, Func. Anal. and its Applications [37:1]{} (2003) 32-45. Riley, A. and Strachan, I.A.B., [Duality for Jacobi Group Orbit Spaces and Elliptic Solutions of the WDVV Equations]{}, Lett. Math. Phys. [77:3]{} (2006) 221-234. Riley, A. and Strachan, I.A.B., [A note on the relationship between rational and trigonometric solutions of the WDVV equations,]{} J. Nonlinear Mathematical Physics 14(1) (2007) 82-94 Strachan, I.A.B., [Frobenius manifolds: natural submanifolds and induced bi-Hamiltonian structures]{}, Differential Geometry and its Applications, 20 (2004), 67-99. Strachan, I.A.B., [Weyl groups and elliptic solutions of the WDVV equations]{}, Advances in Mathematics [224:5]{}, (2010), 1801-1838. Strachan, I.A.B., [Simple Elliptic Singularities: a note on their $G$-function]{}, arXiv:1004.2140 Veselov, A.P., [[Deformations of the root systems and new solutions to generalized WDVV equations]{}, Phys. Lett [A 261]{} (1999) 297-302.]{} [^1]: For convenience we refer to a Frobenius manifold $F$ rather than to the Frobenius manifold whose prepotential is $F\,.$ [^2]: The notation $\mathfrak{t}$ will be used rather than the more standard $\tilde{t}$ to avoid, when we look action on $I$, symbols of the form $\hat{\tilde{t}}\,.$	{ "pile_set_name": "ArXiv" }
--- abstract: \| Theoretical design of global optimization algorithms can profitably utilize recent statistical mechanical treatments of potential energy surfaces (PES’s). Here we analyze a particular method to explain its success in locating global minima on surfaces with a multiple-funnel structure, where trapping in local minima with different morphologies is expected. We find that a key factor in overcoming trapping is the transformation applied to the PES which broadens the thermodynamic transitions. The global minimum then has a significant probability of occupation at temperatures where the free energy barriers between funnels are surmountable. [PACS numbers: 61.46.+w,02.60.Pn,36.40.Ei]{} address: - 'FOM Institute for Atomic and Molecular Physics, Kruislaan 407, 1098 SJ Amsterdam, The Netherlands' - 'University Chemical Laboratory, Lensfield Road, Cambridge CB2 1EW, UK' author: - 'Jonathan P. K. Doye' - 'David J. Wales' title: On the Thermodynamics of Global Optimization --- [2]{} Global optimization is a subject of intense current interest, since better optimization techniques can produce cost reductions and greater efficiency. It is therefore important to understand the key requirements for a successful global optimization method, rather than proceeding on purely empirical or intuitive grounds. One of the difficulties associated with global optimization is the exponential increase in the search space with system size. For example, the number of possible conformations of a typical protein is so large that it would take longer than the age of the universe to find the native state if the conformations were sampled randomly (Levinthal’s ‘paradox’[@Levinthal]). This problem can be more rigorously defined using computational complexity theory; finding the global minimum of a protein[@Ngo94] or a cluster[@Wille] is NP-hard. In fact, it is easy to design surfaces that result in efficient relaxation to the global minimum, despite very large configurational spaces[@Zwanzig92; @Zwanzig95; @JD96c]. Such surfaces are characterized by a single deep funnel[@BandK97] leading to the global minimum, and this feature may underlie the ability of a protein to fold to the native state[@Leopold; @Bryngel95]. Hence, global optimization methods should find surfaces with a single funnel relatively easy to tackle. The problem is significantly harder if there is more than one funnel, since there are then competing relaxation pathways. The extreme case would be when a funnel that does not lead to the global minimum is thermodynamically favoured down to low temperatures. On cooling the system would probably be trapped in the secondary funnel. This explains why naive simulated annealing will often fail. A number of atomic clusters bound by the Lennard-Jones (LJ) potential exhibit surfaces with just such a topography. In this paper we examine in detail one such case, an LJ cluster with 38 atoms, to show how a recently-developed global optimization method is able to overcome trapping. Our results indicate some necessary conditions the algorithm must satisfy if it is to succeed when applied to PES’s with multiple funnels. Most small LJ clusters have global minima based upon Mackay icosahedra[@Northby87]. However, for 38 atoms the global minimum is a face-centred cubic (fcc) truncated octahedron (Figure \[38barrier\]), and for 75–77 and 102–104 the global minima are based on Marks’ decahedra[@Marks84]. These minima were initially discovered by construction[@JD95c; @JD95d], and until our recent application of a ‘basin-hopping’ algorithm [@WalesD97] only the LJ$_{38}$ global minimum had been found by an unbiased global optimization method. 0.1cm \[38barrier\] -.5cm The multiple-funnel character of the LJ$_{38}$ PES is revealed in the energy profile of a pathway between the fcc global minimum and the lowest energy icosahedral minimum (Figure \[38barrier\])[@JD97a]. Maxima on this pathway correspond to true transition states (stationary points with a single negative Hessian eigenvalue), and the segments linking maxima and minima are approximate steepest descent paths. The two lowest energy minima are well separated in configuration space, and to cross the barrier between them the cluster must pass through high energy minima associated with the liquid-like state of the cluster. Transitions between the fcc and icosahedral regions of configuration space can only occur if the high energy ‘liquid-like’ minima have a significant probability of being occupied. At low temperatures the Boltzmann weights for these intermediate states are small leading to a large free energy barrier between the two regions. 0.1cm \[38un.therm\] -0.3cm Some of the equilibrium thermodynamic properties of LJ$_{38}$ are shown in Figure \[38un.therm\]. There is only a small energy difference between the global minimum and the lowest energy icosahedral minimum. However, the entropy associated with the icosahedral structures is larger and icosahedra become favoured at low temperature. The centre of this transition is at a temperature of 0.12$\,\epsilon k^{-1}$ ($\epsilon$ is the pair well depth of the LJ potential and $k$ is the Boltzmann constant); it gives rise to the small peak in the heat capacity (Figure \[38un.therm\]b). The finite-size analogue of the bulk melting transition occurs at 0.18$\,\epsilon k^{-1}$, producing the main peak in the heat capacity. These transitions are also reflected in the occupation probabilities for the different ‘phases’ (Figure \[38un.therm\]a). At low temperatures one observes the cluster trapped in either the fcc or icosahedral funnels, because of the large free energy barrier between them. On cooling from the liquid-like state, there is a thermodynamic driving force for entering the icosahedral region of configuration space. This effect is exacerbated by the topography of the PES. The structure of simple atomic liquids has significant polytetrahedral character[@NelsonS; @JD96b], and relaxation into the icosahedral funnel is more likely because icosahedra have more polytetrahedral character than the fcc structures. Hence, the icosahedral funnel is directly connected to the low energy liquid-like states, whereas the fcc funnel is only connected to the high energy liquid-like states. Therefore, it is extremely probable that the system will enter the icosahedral funnel on cooling and there become trapped. To observe the truncated octahedron we must simulate the cluster in the small temperature window where both the high energy liquid-like minima and the fcc structures have small, but significant, probabilities of being occupied. Indeed, we did observe the truncated octahedron in this temperature range, but only [once]{} in regular quenches from a 0.25$\,\mu$s simulation for Ar. The situation for the larger clusters with non-icosahedral global minima is even worse. The decahedral to icosahedral transition is sharper and lies even further below the melting transition than for LJ$_{38}$ (e.g. for LJ$_{75}$ it occurs at $T=0.09\,\epsilon k^{-1}$). The global optimization method that we analyze here belongs to the family of ‘hypersurface deformation’ methods[@StillW88]. In this approach the energy function is transformed, usually to a smoother surface with fewer minima. The lowest minimum of this new surface is then mapped back to the original surface, but there is no guarantee that the global minima on the two surfaces are related and often there are good reasons to think that they are not[@JD95c]. In contrast, the transformation that we apply is guaranteed to preserve the global minimum. The transformed energy $\tilde E$ is defined by: $$\tilde E({\bf X}) = min\left\{ E({\bf X}) \right\},$$ where ${\bf X}$ represents the vector of nuclear coordinates and $min$ signifies that an energy minimization is performed starting from ${\bf X}$. The above approach is very similar to Li and Scheraga’s Monte Carlo (MC) minimization[@Li87a]. In our recent application to LJ clusters it has outperformed all other methods in the literature, finding all the known lowest energy LJ clusters up to 110 atoms, including those with non-icosahedral global minima[@WalesD97]. The method we use to explore the $\tilde E$ surface is simply a canonical MC simulation at constant temperature, in this case 0.8$\,\epsilon k^{-1}$. Interestingly, the other methods that have been most successful for LJ clusters are genetic algorithms[@Deaven96; @Niesse96a] which use minimization to refine the new coordinates generated at each step, thus in effect performing a search on the same transformed surface, $\tilde E$. 0.1cm \[38trans.therm\] -0.3cm The topography of the transformed surface is that of a multi-dimensional staircase with each step corresponding to the basin of attraction surrounding a minimum. The transformation has a significant effect on the dynamics. Not only are transitions to a lower energy minimum barrierless, but they can also occur at any point along the boundary between basins of attraction, whereas on the untransformed surface transitions can occur only when the system passes along the transition state valley. As a result intrawell vibrational motion is removed and the system can hop directly between minima at each step. However, the increase in interbasin transitions will not necessarily aid the location of the global minimum on a multiple-funnel PES unless the transformation also changes the thermodynamics. For LJ$_{38}$ the probability of the system occupying the intermediate states between the funnels must be non-negligible under conditions where the icosahedral and fcc structures also have significant occupation probabilities. This is indeed what is observed for the $\tilde E$ surface (Figure \[38trans.therm\]). The transitions have been smeared out and there is now only one broad peak in the heat capacity. Most significantly, the probability that the system is in the basin of attraction of the global minimum decays much more slowly, and the temperature range for which both the high energy liquid-like minima and the global minimum have significant probabilities is large. MC simulations anywhere in this temperature range easily locate the global minimum from a random starting point. Furthermore, the $\tilde E$ transformation opens up new paths between the two lowest energy minima because atoms are now able to pass through oneanother. Such paths would obviously not be feasible on the untransformed PES. Consequently, even at $T=0.2\,\epsilon k^{-1}$ the cluster can escape from the icosahedral region of the transformed surface. We can understand the different thermodynamics for the two surfaces by writing the partition function as a sum over all the geometrically distinct minima on the PES. On the untransformed surface this gives for $p_s$, the probability that the cluster is in minimum $s$, $$p_s(\beta)={n_s\exp(-\beta E_s) \over \overline\nu_s^{3N-6}}\big/ \sum_s{n_s\exp(-\beta E_s) \over \overline\nu_s^{3N-6}},$$ where $\beta=1/kT$, $N$ is the number of atoms, $\overline\nu_s$ is the geometric mean vibrational frequency of minimum $s$ (calculated by diagonalizing the Hessian matrix), $n_s$ is the number of permutational isomers of minimum $s$ and is given by $n_s=2N!/h_s$, where $h_s$ is the order of the point group of minimum $s$. This equation is only approximate since it assumes that each well is harmonic. However, it does give a qualitatively correct picture of the thermodynamics[@Wales93a], and although anharmonic corrections can be found[@JD95a] they are rather cumbersome. For the transformed surface $$p_s(\beta)={n_s A_s\exp(-\beta E_s)\over \sum_s n_s A_s \exp(-\beta E_s)}.$$ where $A_s$ is the hyperarea on the PES for which minimization leads to minimum $s$. The $A_s$ values depend upon the available nuclear configuration space, which must be bounded to prevent evaporation. We have considered both placing the cluster in a container and restricting the configuration space to hyperspheres around each local minimum by setting the coordinates to those of the relevant minimum in the Markov chain. Similar results are obtained for appropriate choices of container and hypersphere radii; the $A_s$ values reported below and the results in reference [@WalesD97] were obtained by resetting the coordinates. Expressions (2) and (3) differ only in the vibrational frequency and $A_s$ terms. The fcc to icosahedral and the icosahedral to liquid transitions are caused by the greater number of minima associated with the higher temperature state, which leads to a larger entropy. On the untransformed surface this effect is reinforced by the decrease in the mean vibrational frequencies with increasing potential energy, $\overline\nu_{fcc}:\overline\nu_{icos}:\overline\nu_{liquid}=1:0.968:0.864$. Although these differences may seem small, when raised to the power $3N-6$ they increase the entropy of the higher energy states significantly, sharpening the transitions and lowering the temperature at which they occur. In contrast, $A_s$ decreases rapidly with increasing potential energy, $A_{fcc}:A_{icos}:A_{liquid}=1:0.0488:0.00122$. These values were obtained by inversion of the occupation probabilities obtained in MC runs of up to $10^6$ steps. The decrease in $A_s$ reduces the entropy of the higher energy states, causing the transitions to be broadened and the temperature at which they occur to increase. We can now explain in detail why the basin-hopping method is successful. First, the staircase transformation removes the barriers between minima on the PES without changing the identity of the global minimum, accelerating the dynamics. Second, it changes the thermodynamics of the surface, broadening the transitions so that even for a multiple-funnel surface such as that of LJ$_{38}$, the global minimum has a significant probability of occupation at temperatures where the free energy barrier for passage between the funnels is surmountable. It is this latter feature which is especially important in overcoming the difficulties associated with multiple funnels and represents a [new criterion for designing successful global optimization methods]{}. The broadened transition that results from the transformation also differs markedly from the optimum conditions for protein folding, if we assume that proteins have evolved single-funnel surfaces in order to fold efficiently. A steep funnel provides a large thermodynamic driving force for relaxation to the global minimum[@JD96c; @Bryngel95], and also leads to a sharp thermodynamic transition. However, in global optimizations one cannot make assumptions about the topography of the PES and on a multiple-funnel surface such features can exacerbate the difficulties associated with trapping in secondary funnels[@JD96c]. Moreover, in protein folding there is the extra requirement that the folded protein must remain localized in the native state, and this necessitates a sharp transition. There is no need for a global optimization method to mimic this property. Indeed, when applied to a PES with multiple funnels the optimum temperature for the basin-hopping approach is above the transition, where the system only spends a minority of its time in the global minimum. We have tested the basin-hopping method on a number of other systems, and its performance is equally impressive. For example, it succeeds in finding the global minima for clusters bound by short range Morse potentials, which have much rougher energy landscapes than LJ clusters[@JD95c]. We have also obtained results for water, metal and silicon clusters which are helping to interpret experimental results. D.J.W. is grateful to the Royal Society for financial support. The work of the FOM Institute is part of the scientific program of FOM and is supported by the Nederlandse Organisatie voor Wetenschappelijk Onderzoek (NWO). [10]{} C. Levinthal, in [Mössbauer Spectroscopy in Biological Systems, Proceedings of a Meeting Held at Allerton House, Monticello, Illinois]{}, edited by J. T. P. DeBrunner and E. Munck (University of Illinois Press, Illinois, 1969), pp. 22–24. J. T. Ngo, J. Marks, and M. Karplus, in [The Protein Folding Problem and Tertiary Structure Prediction]{}, edited by K. Merz and S. L. Grand (Birkhäuser, Boston, 1994), pp. 433–506. L. T. Wille and J. Vennik, J. Phys. A [18]{}, L419 (1985). R. Zwanzig, A. Szabo, and B. Bagchi, Proc. Natl. Acad. Sci. USA [89]{}, 20 (1992). R. Zwanzig, Proc. Natl. Acad. Sci. USA [92]{}, 9801 (1995). J. P. K. Doye and D. J. Wales, J. Chem. Phys. [105]{}, 8428 (1996). O. M. Becker and M. Karplus, J. Chem. Phys. [106]{}, 1495 (1997). P. E. Leopold, M. Montal, and J. N. Onuchic, Proc. Natl. Acad. Sci. USA [89]{}, 8271 (1992). J. D. Bryngelson, J. N. Onuchic, N. D. Socci, and P. G. Wolynes, Proteins: Structure, Function and Genetics [21]{}, 167 (1995). J. A. Northby, J. Chem. Phys. [87]{}, 6166 (1987). L. D. Marks, Phil. Mag. A [49]{}, 81 (1984). J. P. K. Doye, D. J. Wales, and R. S. Berry, J. Chem. Phys. [103]{}, 4234 (1995). J. P. K. Doye and D. J. Wales, Chem. Phys. Lett [247]{}, 339 (1995). D. J. Wales and J. P. K. Doye, J. Phys. Chem. A [101]{}, 5111 (1997). J. P. K. Doye and D. J. Wales, Z. Phys. D [40]{}, 194 (1997). J. P. K. Doye and D. J. Wales, J. Chem. Phys. [102]{}, 9659 (1995). D. R. Nelson and F. Spaepen, Solid State Phys. [42]{}, 1 (1989). J. P. K. Doye and D. J. Wales, J. Phys. B [29]{}, 4859 (1996). F. H. Stillinger and T. A. Weber, J. Stat. Phys. [52]{}, 1429 (1988). Z. Li and H. A. Scheraga, Proc. Natl. Acad. Sci. USA [84]{}, 6611 (1987). D. M. Deaven, N. Tit, J. R. Morris, and K. M. Ho, Chem. Phys. Lett [256]{}, 195 (1996). J. A. Niesse and H. R. Mayne, J. Chem. Phys. [105]{}, 4700 (1996). D. J. Wales, Mol. Phys. [78]{}, 151 (1993).	{ "pile_set_name": "ArXiv" }
--- author: - 'Arpan Chattopadhyay, Abhishek Sinha, Marceau Coupechoux, and Anurag Kumar \' bibliography: - 'IEEEabrv.bib' - 'arpan-techreport.bib' title: 'Deploy-As-You-Go Wireless Relay Placement: An Optimal Sequential Decision Approach using the Multi-Relay Channel Model[^1][^2][^3] ' --- Introduction {#Introduction} ============ Wireless interconnection of devices such as smart phones, or wireless sensors, to the wireline communication infrastructure is an important requirement. These are battery operated, resource constrained devices. Hence, due to the physical placement of these devices, or due to channel conditions, a direct one-hop link to the infrastructure “base-station” might not be feasible. In such situations, other nodes could serve as relays in order to realize a multi-hop path between the source device and the infrastructure. In the wireless sensor network context, the relays could be other wireless sensors or battery operated radio routers deployed specifically as relays. The relays are also resource constrained and a cost might be involved in placing them. Hence, there arises the problem of optimal relay placement. Such a problem involves the joint optimization of node placement and operation of the resulting network, where by “operation” we mean activities such as transmission scheduling, power allocation, and channel coding. ![A source and a sink connected by a multi-hop path comprising $N$ relay nodes along a line.[]{data-label="fig:general_line_network"}](general-line-network.pdf){height="1.2cm" width="8cm"} Our work in this paper is motivated by recent interest in problems of [impromptu (as-you-go)]{} deployment of wireless relay networks in various situations; for example, “first responders” in emergency situations, or quick deployment (and redeployment) of sensor networks in large terrains, such as forests (see [@souryal-etal07real-time-deployment-range-extension], [@howard-etal02incremental-self-deployment-algorithm-mobile-sensor-network], [@bao-lee07rapid-deployment-ad-hoc-backbone], [@sinha-etal12optimal-sequential-relay-placement-random-lattice-path], [@chattopadhyay-etal13measurement-based-impromptu-placement_wiopt]). In this paper, we are concerned with the situation in which a deployment agent walks [from the source node to the sink node]{}, along the line joining these two nodes, and places wireless relays (in an “as-you-go” manner) so as to create a source-to-sink multi-relay channel network with high data rate; see Figure \[fig:general\_line\_network\]. We first consider the scenario where the length $L$ of the line in Figure \[fig:general\_line\_network\] is known; the results of this case are used to formulate the as-you-go deployment in the case where $L$ is a priori unknown, but has exponential distribution with known mean $\overline{L}$. In order to capture the fundamental trade-offs involved in such problems, we consider an information theoretic model. For a placement of the relay nodes and allocation of transmission powers to these relays, we model the “quality” of communication between the source and the sink by the information theoretic achievable rate of the multi-relay channel (see [@cover-gamal79capacity-relay-channel], [@reznik-etal04degraded-gaussian-multirelay-channel] and [@xie-kumar04network-information-theory-scaling-law] for the single and multi-relay channel models). The relays are equipped with full-duplex radios[^4], and carry out decode-and-forward relaying. We consider scalar, memoryless, time-invariant, additive white Gaussian noise (AWGN) channels. We assume synchronous operation across all transmitters and receivers, and consider the exponential path-loss model for radio wave propagation. Related Work {#subsection:related_work} ------------ A formulation of the relay placement problem requires a model of the wireless network at the physical (PHY) and medium access control (MAC) layers. Most researchers have adopted the link scheduling and interference model, i.e., a scheduling algorithm determines radio resource allocation (channel and power) and interference is treated as noise (see [@georgiadis-etal06resource-allocation-cross-layer-control]); treating interference as noise leads to the model that simultaneous transmissions “collide” at receiving nodes, and transmission scheduling aims to avoid collisions. However, node placement for throughput maximization with this model is intractable because the optimal throughput is obtained by first solving for the optimum schedule assuming fixed node locations, followed by an optimization over those locations. Hence, with such a model, there appears to be little work on the problem of jointly optimizing the relay node placement and the transmission schedule. Reference [@firouzabadi-martins08optimal-node-placement] is one such work where the authors considered placing a set of nodes in an existing network such that a certain network utility is optimized subject to a set of linear constraints on link rates, under the link scheduling and interference model. They posed the problem as one of geometric programming assuming exponential path-loss, and proposed a distributed solution. The authors of [@zheng-etal12robust-relay-placement_arxiv] consider relay placement for utility maximization, assuming there are several source nodes, sink nodes and a few candidate locations for placing relays; they ignore interference because of highly directional antennas used in $60$ GHz mmWave networks, which may not always be valid. Relay placement for capacity enhancement has been studied in [@lu-etal11relay-placement-80216], but there interference is mitigated by scheduling transmissions over multiple channels. On the other hand, an information theoretic model for a wireless network often provides a closed-form expression for the channel capacity, or at least an achievable rate region. These results are asymptotic, and make idealized assumptions such as full-duplex radios, perfect interference cancellation, etc., but provide algebraic expressions that can be used to formulate tractable optimization problems which can provide useful insights. In the context of optimal relay placement, some researchers have already exploited this approach. Thakur et al., in [@thakur-etal10optimal-relay-location-power-allocation], report on the problem of placing a single relay node to maximize the capacity of a broadcast relay channel in a wideband regime. Lin et al., in [@lin-etal07relay-placement-80216], numerically solve the problem of a single relay node placement, under power-law path loss and individual power constraints at the source and the relay; however, our work is primarily focused on multi-relay placement, under the exponential path-loss model and a sum power constraint among the nodes. The linear deterministic channel model ([@avestimehr-etal11wireless-network-deterministic]) is used by Appuswamy et al. in [@appuswamy-etal10relay-placement-deterministic-line] to study the problem of placing two or more relay nodes along a line so as to maximize the end-to-end data rate. Our present paper is in a similar spirit; however, we use the achievable rate formulas for the $N$-relay channel (with decode and forward relays) to study the problem of placing relays on a line having length $L$, under a sum power constraint over the nodes. The most important difference of our paper with the literature reported above is that we address the problem of [sequential placement]{} of relay nodes along a line of an unknown random length. This paper extends our previous work in [@chattopadhyay-etal12optimal-capacity-relay-placement-line], which presents the analysis for the case of given $L$ and $N$; the study under given $L$ and $N$ is a precursor to the formulation of as-you-go deployment problem, since it motivates an additive cost structure that is essential for the formulation of the sequential deployment problem as a Markov decision process (MDP). The deploy-as-you-go problem has been addressed by previous researchers. For example, Howard et al., in [@howard-etal02incremental-self-deployment-algorithm-mobile-sensor-network], provide heuristic algorithms for incremental deployment of sensors in order to cover a deployment area. Souryal et al., in [@souryal-etal07real-time-deployment-range-extension], propose heuristic deployment algorithms for the problem of impromptu wireless network deployment, with an experimental study of indoor RF link quality variation. The authors of [@bao-lee07rapid-deployment-ad-hoc-backbone] propose a collaborative deployment method for multiple deployment agents, so that the contiguous coverage area of relays is maximized subject to a total number of relays constraint. However, until the work in [@mondal-etal12impromptu-deployment_NCC] and [@sinha-etal12optimal-sequential-relay-placement-random-lattice-path], there appears to have been no effort to rigorously formulate as-you-go deployment problem in order to derive optimal deployment algorithms. The authors of [@mondal-etal12impromptu-deployment_NCC] and [@sinha-etal12optimal-sequential-relay-placement-random-lattice-path] used MDP based formulations to address the problem of placing relay nodes sequentially along a line and along a random lattice path, respectively. The formulations in [@mondal-etal12impromptu-deployment_NCC] and [@sinha-etal12optimal-sequential-relay-placement-random-lattice-path] are based on the so-called “lone packet traffic model" under which, at any time instant, there can be no more than one packet traversing the network, thereby eliminating contention between wireless links. This work was later extended in [@chattopadhyay-etal13measurement-based-impromptu-placement_wiopt] to the scenario where the traffic is still lone packet, but a measurement-based approach is employed to account for the spatial variation of link qualities due to shadowing. In this paper, we consider as-you-go deployment along a line, but move away from the lone-packet traffic assumption by employing information theoretic achievable rate formulas (for full-duplex radios and decode-and-forward relaying). We assume exponential path-loss model (see [@franceschetti-etal04random-walk-model-wave-propagation] and Section \[subsection:motivation-for-exponential-path-loss\]). To the best of our knowledge, there is no prior work that considers as-you-go deployment under this physical layer model. Our Contribution ---------------- - [Optimal Offline Deployment:]{} Given the location of $N$ full-duplex relays to connect a source and a sink separated by a given distance $L$, and under the exponential path-loss model and a sum power constraint among the nodes, the optimal power split among the nodes and the achievable rate are expressed (Theorem \[theorem:multirelay\_capacity\]) in terms of the channel gains. We find expression for optimal relay location in the single relay placement problem (Theorem \[theorem:single\_relay\_total\_power\]). For the $N$ relay placement problem, numerical study shows that, the relay nodes are clustered near the source at low attenuation and are placed uniformly at high attenuation. Theorem \[theorem:large\_nodes\_uniform\] shows that, by placing large number of relays uniformly, we can achieve a rate arbitrarily close to the AWGN capacity. Only this part of our current paper was published in the conference version [@chattopadhyay-etal12optimal-capacity-relay-placement-line]. - [Optimal As-You-Go Deployment:]{} In Section \[sec:mdp\_total\_power\], we consider the problem of placing relay nodes in a deploy-as-you-go manner, so as to connect a source and a sink separated by an unknown distance, modeled as an exponentially distributed random variable $L$. Specifically, the problem is to start from a source, and walk along a line, placing relay nodes as we go, until the line ends, at which point the sink is placed. With a sum power constraint, the aim is to maximize a capacity limiting term derived from the deployment problem for known $L$, while constraining the expected number of relays. We “relax” the expected number of relays constraint via a Lagrange multiplier, and formulate the problem as a total cost MDP with uncountable state space and non-compact action sets. We prove the existence of an optimal policy and convergence of value iteration (Theorem \[theorem:convergence\_of\_value\_iteration\]); these results for uncountable state space and non-compact action space are not evident from standard literature. We study properties of the value function analytically. This is the first time that the as-you-go deployment problem is formulated to maximize the end-to-end data rate under the full-duplex multi-relay channel model. - [Numerical Results on As-You-Go Deployment:]{} In Section \[section:numerical\_work\_information\_theoretic\_model\], we study the policy structure numerically. We also demonstrate numerically that the proposed as-you-go algorithm achieves an end-to-end data rate sufficiently close to the maximum possible achievable data rate for offline placement. This is particularly important since there is no other benchmark in the literature, with which we can make a fair comparison of our policy. - The material in Section \[sec:mdp\_total\_power\] and Section \[section:numerical\_work\_information\_theoretic\_model\] were absent in the conference version [@chattopadhyay-etal12optimal-capacity-relay-placement-line]. Organization of the Paper ------------------------- In Section \[sec:system\_model\_and\_notation\], we describe our system model and notation. In Section \[sec:total\_power\_constraint\], we address the problem of relay placement on a line of known length. Section \[sec:mdp\_total\_power\] deals with the problem of as-you-go deployment along a line of unknown random length. Numerical work on as-you-go deployment has been presented in Section \[section:numerical\_work\_information\_theoretic\_model\]. Some discussions are provided in Section \[section:additional\_discussion\]. Conclusions are drawn in Section \[conclusion\]. System Model and Notation {#sec:system_model_and_notation} ========================= The Multi-Relay Channel ----------------------- The multi-relay channel was studied in [@xie-kumar04network-information-theory-scaling-law] and [@reznik-etal04degraded-gaussian-multirelay-channel] and is an extension of the single relay model presented in [@cover-gamal79capacity-relay-channel]. We consider a network deployed on a line with a source node and a sink node at the end of the line, and $N$ full-duplex relay nodes as shown in Figure \[fig:general\_line\_network\]. The relay nodes are numbered as $1, 2,\cdots,N$. The source and sink are indexed by $0$ and $N+1$, respectively. The distance of the $k$-th node [from the source]{} is denoted by $y_{k}:=r_{1}+r_{2}+\cdots+r_{k}$. Thus, $y_{N+1}=L$. As in [@xie-kumar04network-information-theory-scaling-law] and [@reznik-etal04degraded-gaussian-multirelay-channel], we consider the [scalar, time-invariant, memoryless, AWGN setting.]{} We use the model that a symbol transmitted by node $i$ is received at node $j$ after multiplication by the (positive, real valued) channel gain $h_{i,j}$ (an assumption often made in the literature, see e.g., [@xie-kumar04network-information-theory-scaling-law] and [@gupta-kumar03capacity-wireless-networks]). The power gain from Node $i$ to Node $j$ is denoted by $g_{i,j} = h_{i,j}^2$. We define $g_{i,i}=1$ and $h_{i,i}=1$. The Gaussian additive noise at any receiver is independent and identically distributed from symbol to symbol and has variance $\sigma^2$. An Inner Bound to the Capacity {#subsec:achievable_rate_multirelay_xie_and_kumar} ------------------------------ For the multi-relay channel, we denote the symbol transmitted by the $i$-th node at time $t$ ($t$ is discrete) by $X_{i}(t)$ for $i=0,1,\cdots,N$. $Z_{k}(t) \sim \mathcal{N}(0,\sigma^{2})$ is the additive white Gaussian noise at node $k$ and time $t$, and is assumed to be [independent and identically distributed across $k$ and $t$.]{} Thus, at symbol time $t$, node $k, 1 \leq k \leq N+1$ receives: $$Y_{k}(t)= \sum_{j\in \{0,1,\cdots,N\}, j \neq k} h_{j,k}X_{j}(t)+Z_{k}(t) \label{eqn:network_equation}$$ An inner bound to the capacity of this network, under any path-loss model, is given by (see [@xie-kumar04network-information-theory-scaling-law]): $$R=\min_{1 \leq k \leq N+1} C \bigg( \frac{1}{\sigma^{2}×} \sum_{j=1}^{k} ( \sum_{i=0}^{j-1} h_{i,k} \sqrt{P_{i,j}} )^{2} \bigg) \label{eqn:achievable_rate_multirelay}$$ where $C(x):=\frac{1}{2}\log_{2}(1+x)$, and node $i$ transmits to node $j$ at power $P_{i,j}$ (expressed in mW). In Appendix \[section:coding\_scheme\_description\], we provide a descriptive overview of the coding and decoding scheme proposed in [@xie-kumar04network-information-theory-scaling-law]. A sequence of messages are sent from the source to the sink; each message is encoded in a block of symbols and transmitted by using the relay nodes. The scheme involves [coherent transmission]{} by the source and relay nodes (this requires [symbol-level synchronization among the nodes]{}), and [successive interference cancellation]{} at the relay nodes and the sink. A node receives information about a message in two ways (i) by the message being directed to it cooperatively by all the previous nodes, and (ii) by overhearing previous transmissions of the message to the previous nodes. Thus node $k$ receives codes corresponding to a message $k$ times before it attempts to decode the message (a discussion on the practical feasibility of full-duplex decode-and-forward relaying scheme is provided in Section \[subsection:motivation-for-full-duplex-decode-forward\]). Note that, $C \bigg( \frac{1}{\sigma^{2}×} \sum_{j=1}^{k} ( \sum_{i=0}^{j-1} h_{i,k} \sqrt{P_{i,j}} )^{2} \bigg)$ in (\[eqn:achievable\_rate\_multirelay\]), for any $k$, denotes a possible rate that can be achieved by node $k$ from the transmissions from nodes $0,1,\cdots,k-1$. The smallest of these terms becomes the bottleneck, see (\[eqn:achievable\_rate\_multirelay\]). For the single relay channel, $N=1$. Thus, by (\[eqn:achievable\_rate\_multirelay\]), an achievable rate is given by (see also [@cover-gamal79capacity-relay-channel]): $$\begin{aligned} R=\min & \bigg\{ & C \left(\frac{g_{0,1}P_{0,1}}{\sigma^{2}×}\right), \nonumber\\ && C \left( \frac{g_{0,2}P_{0,1}+(h_{0,2}\sqrt{P_{0,2}}+h_{1,2}\sqrt{P_{1,2}})^{2}}{\sigma^{2}×} \right) \bigg\}\label{eqn:single_relay_genaral_capacity_formula}\end{aligned}$$ Here, the first term in the $\min\{\cdot,\cdot\}$ of (\[eqn:single\_relay\_genaral\_capacity\_formula\]) is the achievable rate at node $1$ (i.e., the relay node) due to the transmission from the source. The second term in the $\min\{\cdot,\cdot\}$ corresponds to the possible achievable rate at the sink node due to direct coherent transmission from the source and the relay and due to the overheard transmission from the source to the relay. The higher the channel attenuation, the less will be the contribution of farther nodes, “overheard” transmissions become less relevant, and coherent transmission reduces to a simple transmission from the previous relay. The system is then closer to simple store-and-forward relaying. The authors of [@xie-kumar04network-information-theory-scaling-law] have shown that any rate strictly less than $R$ is achievable through the coding and decoding scheme. This achievable rate formula can also be obtained from the capacity formula of a physically degraded multi-relay channel (see [@reznik-etal04degraded-gaussian-multirelay-channel]), since the capacity of the degraded relay channel is a lower bound to the actual channel capacity. [In this paper, we will seek to optimize $R$ in (\[eqn:achievable\_rate\_multirelay\]) over power allocations to the nodes and the node locations, keeping in mind that $R$ is a lower bound to the actual capacity. We denote the value of $R$ optimized over power allocation and relay locations by $R^$.]{} Path-Loss Model --------------- We model the power gain via the exponential path-loss model: the power gain at a distance $r$ is $e^{-\rho r}$ where $\rho > 0$. This is a simple model used for tractability (see [@firouzabadi-martins08optimal-node-placement], [@appuswamy-etal10relay-placement-deterministic-line]) and [@altman-etal11greec-cellular Section $2.3$] for prior work assuming exponential path-loss). However, for propagation scenarios involving randomly placed scatterers (as would be the case in a dense urban environment, or a forest, for example) analytical and experimental support has been provided for the exponential path-loss model (a discussion has been provided in Section \[subsection:motivation-for-exponential-path-loss\]). We also discuss in Section \[subsection:insights\_for\_power\_law\_from\_exponential\] how the insights obtained from the results for exponential path-loss can be used for power-law path-loss (power gain at a distance $r$ is $r^{-\eta}$, $\eta>0$). Deployment with other path-loss models is left in this paper as a possible future work. Under exponential path-loss, the channel gains and power gains in the line network become multiplicative, e.g., $h_{i,i+2}=h_{i,i+1}h_{i+1,i+2}$ and $g_{i,i+2}=g_{i,i+1}g_{i+1,i+2}$ for $i \in \{0,1,\cdots,N-1\}$. We discuss in Section \[subsection:shadowing-fading\] how shadowing and fading can be taken care of in our model, by providing a fade- margin in the power at each transmitter. Motivation for the Sum Power Constraint {#subsec:sum_power_constraint} --------------------------------------- In this paper we consider the sum power constraint $\sum_{i=0}^{N}P_{i}=P_{T}$ (in mW) over the source and the relays. This constraint has the following motivation. Let the fixed power expended in a relay (for reception and driving the electronic circuits) be denoted by $P_{\mathrm{rcv}}$ (expressed in mW), and the initial battery energy in each node be denoted by $E$ (in mJ unit). The information theoretic approach utilized in this paper requires that the nodes in the network are always on. Hence, the lifetime of node $i, 1 \leq i \leq N$, is $\tau_i=\frac{E}{P_i+P_{rcv}×}$, the lifetime of the source is $\tau_0=\frac{E}{P_0}$, and that of the sink is $\tau_{N+1}=\frac{E}{P_{rcv}}$. The rate of battery replacement at node $i$ is $\frac{1}{\tau_i}$. Hence, the rate at which we have to replace the batteries in the network is $\sum_{i=0}^{N+1}\frac{1}{\tau_i×}=\frac{1}{E×}(\sum_{i=0}^{N}P_i+(N+1)P_{rcv})$. The depletion rate $\frac{P_{rcv}}{E}$ is inevitable at any node, and it does not affect the achievable data rate. Hence, in order to reduce the battery replacement rate, we must reduce the sum transmit power in the entire network. Placement on a Line of Known Length {#sec:total_power_constraint} =================================== As a precursor to addressing the deploy-as-you-go problem over a line of unknown length, in this section we solve the problem of power constrained deployment of [a given number of relays on a line of known length]{}. We will often refer to this problem as [offline deployment problem]{}. The results of this section provide (i) first insights into the relay placements we obtain using the multi-relay channel model, (ii) a starting point for the formulation of as-you-go deployment problem, and (iii) a benchmark with which we can compare the performance of our as-you-go deployment algorithm. Optimal Power Allocation {#subsec:optimal_power_allocation} ------------------------ In this section, we consider the optimal placement of relay nodes on a line of given length, $L$, so as to to maximize $R$ (see (\[eqn:achievable\_rate\_multirelay\])), subject to a total power constraint on the source and relay nodes given by $\sum_{i=0}^{N}P_{i}=P_{T}$. We will first maximize $R$ in (\[eqn:achievable\_rate\_multirelay\]) over $P_{i,j}, 0 \leq i < j \leq (N+1)$ for any given placement of nodes (i.e., given $y_{1}, y_{2},\cdots,y_{N}$). This will provide an expression of achievable rate in terms of channel gains, which has to be maximized over $y_{1}, y_{2},\cdots,y_{N}$. Let $\gamma_{k}:=\sum_{i=0}^{k-1}P_{i,k}$ for $k \in \{1,2,\cdots,N+1\}$ (expressed in mW). Hence, the sum power constraint becomes $\sum_{k=1}^{N+1}\gamma_{k}=P_{T}$. \[theorem:multirelay\_capacity\] 1. Under the exponential path-loss model, for fixed location of relay nodes, the optimal power allocation that maximizes the achievable rate for the sum power constraint is given by: $$P_{i,j}= \begin{cases} \frac{g_{i,j}}{\sum_{l=0}^{j-1}g_{l,j}×}\gamma_{j}\,\, & \forall 0 \leq i <j \leq (N+1) \\ 0, \,\, &\text{if}\,\, j \leq i \end{cases}\label{eqn:power_gamma_relation}$$ where $$\begin{aligned} \gamma_{1}&=&\frac{P_{T}×}{1+g_{0,1}\sum_{k=2}^{N+1} \frac{(g_{0,k-1}-g_{0,k})}{g_{0,k}g_{0,k-1}\sum_{l=0}^{k-1}\frac{1}{g_{0,l}×}×} ×} \label{eqn:gamma_one}\\ \gamma_{j}&=&\frac{g_{0,1}\frac{(g_{0,j-1}-g_{0,j})}{g_{0,j}g_{0,j-1}\sum_{l=0}^{j-1}\frac{1}{g_{0,l}×}×}×}{1+g_{0,1}\sum_{k=2}^{N+1} \frac{(g_{0,k-1}-g_{0,k})}{g_{0,k}g_{0,k-1}\sum_{l=0}^{k-1}\frac{1}{g_{0,l}×}×} ×} P_{T} \,\,\,\, \forall \, j \geq 2 \label{eqn:gamma_k}\nonumber\end{aligned}$$ 2. The achievable rate optimized over the power allocation for a given placement of nodes is given by: $$R^{opt}_{P_T}(y_1,y_2,\cdots,y_N)=C \bigg( \frac{\frac{P_{T}}{\sigma^{2}×}} {\frac{1}{g_{0,1}×}+\sum_{k=2}^{N+1} \frac{(g_{0,k-1}-g_{0,k})}{g_{0,k}g_{0,k-1}\sum_{l=0}^{k-1}\frac{1}{g_{0,l}×}×} ×} \bigg)\label{eqn:capacity_multirelay}$$ The basic idea is to choose the power levels (i.e., $P_{i,j},\, 0 \leq i<j \leq N+1$) in (\[eqn:achievable\_rate\_multirelay\]) so that all the terms in the $\min\{\cdot\}$ in (\[eqn:achievable\_rate\_multirelay\]) become equal. We provide explicit expressions for $P_{i,j},\,0 \leq i<j \leq N+1$ and the achievable rate (optimized over power allocation) in terms of the power gains. See Appendix \[appendix:proof\_of\_multirelay\_channel\_capacity\_theorem\_after\_power\_allocation\] for the detailed proof. A result on the equality of certain terms under optimal power allocation has also been proved in [@reznik-etal04degraded-gaussian-multirelay-channel] for the coding scheme used in [@reznik-etal04degraded-gaussian-multirelay-channel]. But it was proved in the context of a degraded Gaussian multi-relay channel, and the proof depends on an inductive argument, whereas our proof utilizes LP (linear programming) duality. Recalling the exponential path-loss parameter $\rho$, and the source-sink distance $L$, let us define $\lambda := \rho L$, which can be treated as a measure of attenuation in the line. Let us now comment on the results of Theorem \[theorem:multirelay\_capacity\]: - In order to maximize $R^{opt}_{P_T}(y_1,y_2,\cdots,y_N)$, we need to place the relay nodes such that $\frac{1}{g_{0,1}}+\sum_{k=2}^{N+1}\frac{(g_{0,k-1}-g_{0,k})}{g_{0,k}g_{0,k-1}\sum_{l=0}^{k-1}\frac{1}{g_{0,l}×}×}$ is minimized. This quantity is viewed as the net attenuation the power $P_T$ faces. - When no relay is placed, the attenuation is $e^{\lambda}$. The ratio of attenuation with no relay and attenuation with relays is called the “relaying gain” $G(N,\lambda)$. $$G(N,\lambda):=\frac{e^{\lambda}}{\frac{1}{g_{0,1}×}+\sum_{k=2}^{N+1}\frac{(g_{0,k-1}-g_{0,k})}{g_{0,k}g_{0,k-1}\sum_{l=0}^{k-1}\frac{1}{g_{0,l}×}×}×} \label{eqn:gain_definition}$$ Rate is increasing with the number of relays, and is bounded above by $C(\frac{P_T}{\sigma^2})$. Hence, $G(N,\lambda) \in [1, e^{\lambda}]$. Also, note that $G(N,\lambda)$ does not depend on $P_T$. - By Theorem \[theorem:multirelay\_capacity\], we have $P_{k,j} \geq P_{i,j}$ for $i<k<j$. - Note that we have derived Theorem \[theorem:multirelay\_capacity\] using the fact that $g_{0,k}$ is nonincreasing in $k$. If there exists some $k \geq 1$ such that $g_{0,k}=g_{0,k+1}$, i.e, if $k$-th and $(k+1)$-st nodes are placed at the same position, then $\gamma_{k+1} = 0$, i.e., the nodes $i < k$ do not direct any power specifically to relay $k+1$. However, relay $k+1$ can decode the symbols received at relay $k$, and those transmitted by relay $k$. Then relay $(k+1)$ can transmit coherently with the nodes $l \leq k$ to improve effective received power in the nodes $j > k+1$. ![Single relay placement, total power constraint, exponential path-loss: $\frac{y_{1}^{}}{L×}$ and optimum $\frac{P_{0,1}}{P_{T}×}$ versus $\lambda$.[]{data-label="fig:single_relay_total_power"}](single-relay-total-power-position-power.pdf){height="3cm" width="8cm"} Optimal Placement of a Single Relay Node {#subsection:optimal_placement_single_relay_sum_power} ---------------------------------------- In the following theorem, we derive the optimal power allocation, node location and data rate when a single relay is placed. \[theorem:single\_relay\_total\_power\] For the single relay node placement problem with sum power constraint and exponential path-loss model, the normalized optimum relay location $\frac{y_{1}^{}}{L×}$, power allocation and optimized achievable rate are given as follows:[^5] 1. [For $\lambda \leq \log 3$]{}, $\frac{y_{1}^{}}{L×}=0$, $P_{0,1}=\frac{2P_{T}}{e^{\lambda}+1×}$, $P_{0,2}=P_{1,2}=\frac{e^{\lambda}-1}{e^{\lambda}+1×}\frac{P_{T}}{2×}$ and $R^{}=C \left(\frac{2P_{T}}{(e^{\lambda}+1)\sigma^{2}×}\right)$. 2. [For $\lambda \geq \log 3$]{}, $\frac{y_{1}^{}}{L×}=\frac{1}{\lambda×} \log \left(\sqrt{e^{\lambda}+1}-1\right)$, $P_{0,1}=\frac{P_{T}}{2×}$, $P_{0,2}=\frac{1}{\sqrt{e^{\lambda}+1}×}\frac{P_{T}}{2×}$, $P_{1,2}=\frac{\sqrt{e^{\lambda}+1}-1}{\sqrt{e^{\lambda}+1}×}\frac{P_{T}}{2×}$ and $R^{}=C \left( \frac{1}{\sqrt{e^{\lambda}+1}-1×}\frac{P_{T}}{2 \sigma^{2}×} \right)$ See Appendix \[appendix:proof\_of\_single\_relay\_sum\_power\_results\]. [Discussion:]{} It is easy to check that $R^{}$ obtained in Theorem \[theorem:single\_relay\_total\_power\] is strictly greater than the AWGN capacity $C \left(\frac{P_{T}}{\sigma^{2}×}e^{-\lambda}\right)$ for all $\lambda>0$. This happens because the source and relay transmit coherently to the sink. $R^{}$ becomes equal to the AWGN capacity only at $\lambda=0$. At $\lambda=0$, we do not use the relay since the sink can decode any message that the relay is able to decode. The variation of $\frac{y_{1}^{}}{L×}$ and $\frac{P_{0,1}}{P_{T}}$ with $\lambda$ has been shown in Figure \[fig:single\_relay\_total\_power\]. We observe that (from Figure \[fig:single\_relay\_total\_power\] and Theorem \[theorem:single\_relay\_total\_power\]) $\lim_{\lambda \rightarrow \infty} \frac{y_{1}^{}}{L}=\frac{1}{2×}$, $\lim_{\lambda \rightarrow \infty}P_{0,2}=0$ and $\lim_{\lambda \rightarrow 0}P_{0,1}=P_{T}$. For large values of $\lambda$, source and relay cooperation provides negligible benefit since source to sink attenuation is very high. So it is optimal to place the relay at a distance $\frac{L}{2×}$. The relay works as a repeater which forwards data received from the source to the sink. For small $\lambda$, the gain obtained from coherent transmission is dominant, and, in order to receive sufficient information (required for coherent transmission) from the source, the relay is placed near the source. Optimal Placement for a Multi-Relay Channel {#subsection:properties_of_gain} ------------------------------------------- As we discussed earlier, we need to place $N$ relay nodes such that $\frac{1}{g_{0,1}×}+\sum_{k=2}^{N+1} \frac{(g_{0,k-1}-g_{0,k})}{g_{0,k}g_{0,k-1}\sum_{l=0}^{k-1}\frac{1}{g_{0,l}×}×}$ is minimized. Here $g_{0,k}=e^{-\rho y_{k}}$. We have the constraint $0 \leq y_{1} \leq y_{2} \leq \cdots \leq y_{N} \leq y_{N+1}=L$. Now, writing $z_{k}=e^{\rho y_{k}}$, and defining $z_{0}:=1$, we arrive at the following problem: $$\begin{aligned} & & \min \bigg\{ z_{1}+\sum_{k=2}^{N+1} \frac{z_{k}-z_{k-1}}{\sum_{l=0}^{k-1} z_{l}}\bigg\}\nonumber\\ & s.t & \,\, 1 \leq z_{1} \leq \cdots \leq z_{N} \leq z_{N+1} = e^{\lambda} \label{eqn:multirelay_optimization}\end{aligned}$$ The objective function is convex in each of the variables $z_{1}, z_{2},\cdots, z_{N}$. The objective function is sum of linear fractionals, and the constraints are linear. [Remark:]{} From optimization problem (\[eqn:multirelay\_optimization\]) we observe that optimum $z_{1},z_{2},\cdots,z_{N}$ depend only on $\lambda:=\rho L$. Since $z_{k}=e^{\lambda \frac{y_{k}}{L}}$, the normalized optimal distance of relays from the source depend only on $\lambda$ and $N$. \[theorem:capacity\_increasing\_with\_N\] For fixed $\rho$, $L$ and $\sigma^{2}$, the optimized achievable rate $R^{}$ for a sum power constraint strictly increases with the number of relay nodes. See Appendix \[appendix:proof\_of\_capacity\_increases\_in\_N\]. \[theorem:G\_increasing\_in\_lambda\] For any fixed number of relays $N \geq 1$, $G(N,\lambda)$ is increasing in $\lambda$. See Appendix \[appendix:proof\_of\_G\_increasing\_in\_lambda\]. [A numerical study of multi-relay placement:]{} We discretize the interval $[0,L]$ and run a search program to find normalized optimal relay locations for different values of $\lambda$ and $N$. The results are summarized in Figure \[fig:multirelay-optimal-location\]. We observe that at low attenuation (small $\lambda$), relay nodes are clustered near the source node and are often at the source node, whereas at high attenuation (large $\lambda$) they are almost uniformly placed along the line. For large $\lambda$, the effect of long distance between any two adjacent nodes dominates the gain obtained by coherent relaying. Hence, it is beneficial to minimize the maximum distance between any two adjacent nodes and thus multihopping is a better strategy in this case. For small $\lambda$, the gain obtained by coherent transmission is dominant. In order to allow this, relays should be able to receive sufficient information from their previous nodes. Thus, they tend to be clustered near the source. In Figure \[fig:G\_vs\_N\] we plot the relaying gain $G(N,\lambda)$ in dB vs. the number of relays $N$, for various values of $\lambda$. As proved in Theorem \[theorem:capacity\_increasing\_with\_N\], we see that $G(N,\lambda)$ increases with $N$ for fixed $\lambda$. On the other hand, $G(N,\lambda)$ increases with $\lambda$ for fixed $N$, as proved in Theorem \[theorem:G\_increasing\_in\_lambda\]. ![$G (N,\lambda)$ vs $N$ for total power constraint.[]{data-label="fig:G_vs_N"}](optimal-placement-fixed-line.pdf){height="3cm" width="8cm"} Uniformly Placed Relays, Large $N$ ---------------------------------- When the relays are uniformly placed, the behaviour of $R^{opt}_{P_T}(y_1,\cdots,y_N)$ (called $R_{N}$ in the next theorem) for large number of relays is captured by the following: \[theorem:large\_nodes\_uniform\] For exponential path-loss and sum power constraint, if $N$ relay nodes are placed uniformly between the source and the sink, resulting in $R_{N}$ achievable rate, then $ \lim_{N \rightarrow \infty} R_N= C \left(\frac{P_{T}}{\sigma^{2}×}\right)$. See Appendix \[appendix:proof\_of\_large\_nodes\_uniform\]. [Remark:]{} From Theorem \[theorem:large\_nodes\_uniform\], it is clear that we can achieve a rate arbitrarily close to $C(\frac{P_{T}}{ \sigma ^{2}})$ (i.e., the effect of path-loss can completely be mitigated) by placing a large enough number of relay nodes. In this context, we would like to mention that the variation of broadcast capacity as a function of the number of nodes $N$ (located randomly inside a unit square) was studied in [@zheng06information-dissemination]; but the broadcast capacity in their paper increases with $N$ since they assume per-node power constraint. As-You-Go Deployment of Relays on a Line of Unknown Length {#sec:mdp_total_power} ========================================================== Having developed the problem of placing a given number of relays over a line of fixed, given length, we now turn to the deploy-as-you-go problem. An agent walks along a line, starting from the source and heading towards the sink which is at an unknown distance from the source location, deploying relays as he goes, so as to achieve a multi-relay network when he encounters the sink location (and places the sink there). We model the distance from the source to sink as an exponentially distributed random variable with mean $\overline{L}=\frac{1}{\beta}$.[^6] The deployment objective is to achieve a high data rate from the source to the sink, subject to a total power constraint and a constraint on the expected number of relays placed (note that, the number of relay nodes, $N$, is a random variable here, due to the randomness in $L$). Using the rate expression from Theorem \[theorem:multirelay\_capacity\], we formulate the problem as a total cost MDP. Such a deployment problem could be motivated by a situation where it is required to place a sensor (say, a video camera) to monitor an event or an object from a safe distance (e.g., the battlefront in urban combat, or a suspicious object that needs to be detonated, or a group of animals in a forest). In such a situation, the deployment agent, after placing the sensor, walk away from the scene of the event, along a forest trail, or a road, or a building corridor, placing relays as he walks, until a suitable safe sink location is found, in such a way that the number of relays is kept small while the end-to-end data rate is maximized. Formulation as an MDP {#subsection:mdp_formulation} --------------------- We now formulate the as-you-go deployment problem as an MDP. ### Deployment Policies In the as-you-go placement problem, the person carries a number of nodes and places them as he walks, under the control of a placement policy. A deployment policy $\pi$ is a sequence of mappings $\{\mu_1,\mu_2, \mu_3, \cdots \}$ from the state space to the action space; at the $k$-th decision instant (i.e., after placing the $(k-1)$-st relay), $\mu_k$ provides the distance at which the next relay should be placed (provided that the line does not end before that point), given the system state which is a function of the locations of previously placed nodes. Thus, the decisions are made based on the locations of the relays placed earlier. The first decision instant is the start of the line, and the subsequent decision instants are the placement points of the relays. Let $\Pi$ denote the set (possibly uncountable) of all deployment policies. Let $\mathbb{E}_{\pi}$ denote the expectation under policy $\pi$. ### The Unconstrained Problem We recall from (\[eqn:capacity\_multirelay\]) that for a fixed length $L$ of the line and a fixed $N$, $e^{\rho y_1}+\sum_{k=2}^{N+1}\frac{e^{\rho y_k}-e^{\rho y_{k-1}}}{1+e^{\rho y_1}+\cdots+e^{\rho y_{k-1}}×}$ has to be minimized in order to maximize $R^{opt}_{P_T}(y_1,y_2,\cdots,y_N)$. $e^{\rho y_1}+\sum_{k=2}^{N+1}\frac{e^{\rho y_k}-e^{\rho y_{k-1}}}{1+e^{\rho y_1}+\cdots+e^{\rho y_{k-1}}×}$ is basically a scaling factor which captures the effect of attenuation and relaying on the maximum possible SNR $\frac{P_T}{\sigma^2}$. Let $\xi>0$ be the cost of placing a relay. We are interested in solving the following problem: $$\begin{aligned} \inf_{\pi\in \Pi } \mathbb{E}_{\pi} \left( \left(e^{\rho y_1}+ \sum_{k=2}^{N+1}\frac{e^{\rho y_k}-e^{\rho y_{k-1}}}{1+e^{\rho y_1}+ \cdots+e^{\rho y_{k-1}}×} \right)+\xi N \right)\label{eqn:unconstrained_mdp}\end{aligned}$$ The “cost” function inside the outer parentheses in (\[eqn:unconstrained\_mdp\]) has two terms, one is the denominator of $G(N,\lambda)$ in (\[eqn:gain\_definition\]), and the other is a linear multiple of the number of relays. Thus, the cost function captures the tradeoff between the cost of placing relays (quantified as $\xi$ per relay), and the need to achieve high end-to-end data rate by making the denominator of $G(N,\lambda)$ small. Note that, due to the randomness in the length of the line, the $y_k, k \geq 1,$ and $N$ are all random variables.[^7] We will see in Theorem \[theorem:convergence\_of\_value\_iteration\] that an optimal policy always exists for this problem. ### The Constrained Problem Solving the problem in (\[eqn:unconstrained\_mdp\]) also helps in solving the following constrained problem: $$\begin{aligned} & & \inf_{\pi\in \Pi } \mathbb{E}_{\pi} \left(e^{\rho y_1}+ \sum_{k=2}^{N+1}\frac{e^{\rho y_k}-e^{\rho y_{k-1}}}{1+e^{\rho y_1}+\cdots+e^{\rho y_{k-1}}×}\right)\nonumber\\ &\textit{s.t., }&\mathbb{E}_{\pi}(N) \leq M \label{eqn:constrained_mdp}\end{aligned}$$ where $M>0$ is a constraint on the expected number of relays. [^8] The following standard result (see [@beutler-ross85optimal-policies-controlled-markov-chains-constraint Theorem $4.3$]) gives the optimal $\xi^$: \[lemma:choice-of-xi\] If there exists $\xi^>0$ and a policy $\pi_{\xi^}^ \in \Pi$ such that $\pi_{\xi^}^$ is an optimal policy for the unconstrained problem (\[eqn:unconstrained\_mdp\]) under $\xi^$ and $\mathbb{E}_{\pi_{\xi^}^}N=M$, then $\pi_{\xi^}^$ is also optimal for the constrained problem (\[eqn:constrained\_mdp\]). The motivation behind formulation (\[eqn:constrained\_mdp\]) is as follows. Suppose that one seeks to solve the following problem: $$\begin{aligned} && sup_{\pi } \mathbb{E}_{\pi} \log_2 \bigg(1+\frac{\frac{P_T}{\sigma^2×}}{e^{\rho y_1}+ \sum_{k=2}^{N+1}\frac{e^{\rho y_k}-e^{\rho y_{k-1}}}{1+e^{\rho y_1}+\cdots+e^{\rho y_{k-1}}×}}\bigg)\nonumber\\ &&\textit{s.t., } \mathbb{E}_{\pi}(N) \leq M \label{eqn:constrained_mdp_actual}\end{aligned}$$ Since $\log_2 (1+\frac{1}{x})$ is convex in $x$, we can argue by Jensen’s inequality that by solving (\[eqn:constrained\_mdp\]) we actually find a relay placement policy that maximizes a lower bound to the expected achievable data rate obtained from (\[eqn:constrained\_mdp\_actual\]). But formulation (\[eqn:constrained\_mdp\]) (and hence formulation (\[eqn:unconstrained\_mdp\]), by Lemma \[lemma:choice-of-xi\]) allows us to write the objective function as a summation of hop-costs; this motivates us to formulate the as-you-go deployment problem as an MDP, resulting in a substantial reduction in policy computation. However, in Section \[section:numerical\_work\_information\_theoretic\_model\], we will show numerically that solving (\[eqn:constrained\_mdp\]) is a reasonable approach to deal with the computational complexity of (\[eqn:constrained\_mdp\_actual\]); we will see that formulation (\[eqn:constrained\_mdp\]) allows us to achieve a reasonable performance. We now formulate the above “as-you-go" relay placement problem (\[eqn:unconstrained\_mdp\]) as a total cost Markov decision process. ### State Space, Action Space and Cost Structure Let us define $s_0:=1$, $s_k:=\frac{e^{\rho y_k}}{1+e^{\rho y_1}+\cdots+e^{\rho y_k}×}\, \forall \, k \geq 1$. Also, recall that $r_{k+1}=y_{k+1}-y_k$. Thus, we can rewrite (\[eqn:unconstrained\_mdp\]) as: $$\begin{aligned} \inf_{\pi\in \Pi } \mathbb{E}_{\pi} \left(1+\sum_{k=0}^{N}s_k(e^{\rho r_{k+1}}-1)+\xi N \right) \label{eqn:unconstrained_mdp_with_state}\end{aligned}$$ When the person starts walking from the source along the line, the state of the system is set to $s_0:=1$. At this instant the placement policy provides the location at which the first relay should be placed. The person walks towards the prescribed placement point. If the sink placement location is encountered before reaching this point, the sink is placed; if not, then the first relay is placed at the placement point. In general, the state after placing the $k$-th relay is denoted by $s_k$ (a function of the location of the nodes up to the $k$-th instant), for $k=1,2,\cdots$. At state $s_k$, the action is the distance $r_{k+1}$ where the next relay has to be placed (action $\infty$ means that no further relay will be placed). If the line ends before this distance, the sink node has to be placed at the end. [The randomness is coming from the random residual length of the line.]{} Let $l_k$ denote the residual length at the $k$-th instant. With this notation, the state of the system evolves as: $$\begin{aligned} s_{k+1}= \begin{cases} \frac{s_k e^{\rho r_{k+1}}}{1+s_k e^{\rho r_{k+1}}×}, \text{ if $l_k > r_{k+1}$} , \\ \mathbf{EOL}, \text{ \hspace{7mm} else}. \label{eqn:state-evolution} \end{cases}\end{aligned}$$ Here $\mathbf{EOL}$ denotes the end of the line, i.e., the termination state. The single stage cost (for problem (\[eqn:unconstrained\_mdp\_with\_state\])) for state $s$, action $a$ and residual length $l$, is: $$\begin{aligned} c(s,a,l)= \begin{cases} \xi + s(e^{\rho a}-1), \text{ if $l>a$},\\ s(e^{\rho l}-1), \text{\hspace{6.5mm} else}. \label{eqn:single-stage-cost} \end{cases}\end{aligned}$$ Also, $c(\mathbf{EOL},a,\cdot)=0$ for all $a$. From (\[eqn:state-evolution\]), it is clear that the next state $s_{k+1}$ depends on the current state $s_k$, the current action $r_{k+1}$ and the residual length of the line. Since the length of the line is exponentially distributed, from any placement point, the residual line length is exponentially distributed, and independent of the history of the process. The cost incurred at the $k$-th decision instant is given by (\[eqn:single-stage-cost\]), which depends on $s_k$, $r_{k+1}$ and $l_k$. Hence, our formulation in (\[eqn:unconstrained\_mdp\_with\_state\]) is an MDP with state space $\mathcal{S}:=(0,1]\cup \{\mathbf{EOL}\}$ and action space $\mathcal{A}\cup \{\infty \}$ where $\mathcal{A}:=[0,\infty)$. [Remark:]{} An optimal policy (if it exists) for the problem (\[eqn:unconstrained\_mdp\]) will be used to place relay nodes along a line whose length is a sample from an exponential distribution with mean $\frac{1}{\beta×}$. After the deployment is over, the power $P_T$ will be shared optimally among the source and the deployed relay nodes (according to Theorem \[theorem:multirelay\_capacity\]). Optimal Value Function {#subsection:analysis_of_mdp} ---------------------- Suppose $s_k=s$ for some $k \geq 0$. Then, the optimal value function (cost-to-go) at state $s$ is defined by: $$J_{\xi}(s)=\inf_{\pi \in \Pi} \mathbb{E} \left(\sum_{n=k}^{\infty} c(s_n, a_n, l_n)\|s_k=s \right)$$ If we decide to place the next relay at a distance $a<\infty$ and follow the optimal policy thereafter, the expected cost-to-go at a state $s \in (0,1]$ becomes: $$\int_{0}^{a}\beta e^{-\beta z}s(e^{\rho z}-1)dz + e^{-\beta a}\bigg(s(e^{\rho a}-1)+\xi+J_{\xi}\left(\frac{se^{\rho a}}{1+se^{\rho a}×}\right)\bigg)\label{eqn:cost-to-go}$$ The first term in (\[eqn:cost-to-go\]) corresponds to the case in which the line ends at a distance less than $a$ and we are forced to place the sink node. The second term corresponds to the case where the residual length of the line is greater than $a$ and a relay is placed at a distance $a$. Note that our MDP has an uncountable state space $\mathcal{S}=(0,1] \cup \{\mathbf{EOL}\}$ and a non-compact action space $\mathcal{A}=[0,\infty) \cup \{\infty \}$. Several technical issues arise in this kind of problems, such as the existence of optimal or $\epsilon$-optimal policies, measurability of the policies, etc. We, therefore, invoke the results provided by Schäl [@schal75conditions-optimality], which deal with such issues. Our problem is one of minimizing total, undiscounted, non-negative costs over an infinite horizon. Equivalently, in the context of [@schal75conditions-optimality], we have a problem of total reward maximization where the rewards are the negative of the costs. Thus, our problem specifically fits into the negative dynamic programming setting of [@schal75conditions-optimality] (i.e., the $\mathsf{N}$ case where single-stage rewards are non-positive). Now, the state $\mathbf{EOL}$ is absorbing. Also, no action is taken at this state and the cost at this state is $0$. Hence, we can think of this state as state $0$ in order to make our state space a Borel subset of the real line. \[thm:schal\_bellman\_eqn\] \[[@schal75conditions-optimality], Equation (3.6)\] The optimal value function $J_{\xi}(\cdot)$ satisfies the Bellman equation. [$\Box$]{} Thus, $J_\xi(\cdot)$ satisfies the following Bellman equation for each $s \in (0,1]$: $$\begin{aligned} J_{\xi}(s) &=& \min \bigg \{\inf_{a \geq 0} \bigg[\int_{0}^{a}\beta e^{-\beta z}s(e^{\rho z}-1)dz \nonumber\\ && + e^{-\beta a}\bigg(s(e^{\rho a}-1)+\xi+J_{\xi}\left(\frac{se^{\rho a}}{1+se^{\rho a}×}\right)\bigg)\bigg], \nonumber\\ && \int_{0}^{\infty}\beta e^{-\beta z}s(e^{\rho z}-1)dz \bigg \} \label{eqn:bellman_unbroken} \end{aligned}$$ where the second term inside $\min\{\cdot, \cdot \}$ is the cost of not placing any relay (i.e., $a=\infty$). We analyze the MDP for $\beta>\rho$ and $\beta \leq \rho$. ### Case I ($\beta>\rho$) We observe that the cost of not placing any relay (i.e., $a=\infty$) at state $s \in (0,1]$ is given by: $$\begin{aligned} \int_{0}^{\infty} \beta e^{-\beta z}s(e^{\rho z}-1)dz=\theta s\end{aligned}$$ where $\theta:=\frac{\rho}{\beta-\rho×}$ (using the fact that $\beta>\rho$). Since not placing a relay (i.e., $a = \infty$) is a possible action for every $s$, it follows that $J_{\xi}(s)\leq \theta s $. The cost in (\[eqn:cost-to-go\]), upon simplification, can be written as: $$\begin{aligned} \theta s + e^{-\beta a}\bigg(-\theta s e^{\rho a}+\xi+ J_{\xi}\left(\frac{se^{\rho a}}{1+se^{\rho a}×}\right)\bigg)\label{eqn:cost_to_go_beta_greater_than_rho}\end{aligned}$$ Since $J_{\xi}(s) \leq \theta$ for all $s \in (0,1]$, the expression in (\[eqn:cost\_to\_go\_beta\_greater\_than\_rho\]) is strictly less that $\theta s$ for large enough $a<\infty$. Hence, according to (\[eqn:bellman\_unbroken\]), it is not optimal to not place any relay and the Bellman equation (\[eqn:bellman\_unbroken\]) can be rewritten as: $$\begin{aligned} J_{\xi}(s)= \theta s + \inf_{a \geq 0} e^{-\beta a}\bigg(-\theta s e^{\rho a}+ \xi+J_{\xi}\left(\frac{se^{\rho a}}{1+se^{\rho a}×}\right)\bigg)\label{eqn:bellman_equation_simplified_in_a}\end{aligned}$$ ### Case II ($\beta \leq \rho$) Here the cost in (\[eqn:cost-to-go\]) is $\infty$ if we do not place a relay (i.e., if $a=\infty$). Let us consider a policy $\pi_1$ where we place the next relay at a fixed distance $0 <a <\infty$ from the current relay, irrespective of the current state. If the residual length of the line is $z$ at any state $s$, we will place less than $\frac{z}{a×}$ additional relays, and for each relay a cost less than $(\xi+(e^{\rho a}-1))$ is incurred (since $s \leq 1$). At the last step when we place the sink, a cost less than $(e^{\rho a}-1)$ is incurred. Thus, the value function of this policy is upper bounded by: $$\begin{aligned} && \int_{0}^{\infty} \beta e^{-\beta z} \frac{z}{a×}(\xi+(e^{\rho a}-1)) dz+(e^{\rho a}-1) \nonumber\\ &=& \frac{1}{\beta a ×}\left(\xi+(e^{\rho a}-1) \right)+(e^{\rho a}-1)\label{eqn:upper_bound_on_cost_beta_leq_rho}\end{aligned}$$ Hence, $J_{\xi}(s) \leq \frac{1}{\beta a ×}\left(\xi+(e^{\rho a}-1) \right)+(e^{\rho a}-1) < \infty$. Thus, by the same argument as in the case $\beta > \rho$, the minimizer in the Bellman equation lies in $[0,\infty)$, i.e., the optimal placement distance lies in $[0,\infty)$. Hence, (\[eqn:bellman\_unbroken\]) can be rewritten as: $$\begin{aligned} J_{\xi}(s) &=& \inf_{a \geq 0} \bigg\{\int_{0}^{a}\beta e^{-\beta z}s(e^{\rho z}-1)dz + \nonumber\\ && e^{-\beta a}\bigg(s(e^{\rho a}-1)+\xi+J_{\xi}\left(\frac{se^{\rho a}}{1+se^{\rho a}×}\right)\bigg)\bigg\} \label{eqn:bellman_beta_leq_rho} \end{aligned}$$ Upper Bound on the Optimal Value Function ----------------------------------------- \[prop:upper\_bound\_on\_cost\_beta\_geq\_rho\] If $\beta>\rho$, then $J_{\xi}(s) < \theta s$ for all $s \in (0,1]$. We know that $J_{\xi}(s) \leq \theta s \leq \theta$. Now, let us consider the Bellman equation (\[eqn:bellman\_equation\_simplified\_in\_a\]). It is easy to see that $(-\theta s e^{\rho a}+\xi+J_{\xi}(\frac{se^{\rho a}}{1+se^{\rho a}×}))$ is strictly negative for sufficiently large $a$. Hence, the R.H.S of (\[eqn:bellman\_equation\_simplified\_in\_a\]) is strictly less than $\theta s$. For $\beta>\rho$, $\lim_{s \rightarrow 0} J_{\xi}(s) \rightarrow 0$ for any $\xi>0$. Follows from Proposition \[prop:upper\_bound\_on\_cost\_beta\_geq\_rho\]. \[prop:upper\_bound\_on\_cost\] If $\beta>0$ and $\rho>0$ and $0<a<\infty$, then $J_{\xi}(s) <\frac{1}{\beta a ×}\left(\xi+(e^{\rho a}-1) \right)+(e^{\rho a}-1)$ for all $s \in (0,1]$. Follows from (\[eqn:upper\_bound\_on\_cost\_beta\_leq\_rho\]), since the analysis is valid even for $\beta>\rho$. Convergence of the Value Iteration ---------------------------------- The value iteration for our MDP is given by: $$\begin{aligned} J_{\xi}^{(k+1)}(s) &=& \inf_{a \geq 0} \bigg\{\int_{0}^{a}\beta e^{-\beta z}s(e^{\rho z}-1)dz + e^{-\beta a}\bigg(s(e^{\rho a}-1) \nonumber\\ &&+\xi+J_{\xi}^{(k)}\left(\frac{se^{\rho a}}{1+se^{\rho a}×}\right)\bigg)\bigg\}\label{eqn:value_iteration}\end{aligned}$$ Here $J_{\xi}^{(k)}(s)$ is the $k$-th iterate of the value iteration. Let us start with $J_{\xi}^{(0)}(s):=0$ for all $s \in (0,1]$. We set $J_{\xi}^{(k)}(\mathbf{EOL})=0$ for all $k \geq 0$. $J_{\xi}^{(k)}(s)$ is the optimal value function for a problem with the same single-stage cost and the same transition structure, but with the horizon length being $k$ (instead of infinite horizon as in our original problem) and $0$ terminal cost. Here, by horizon length $k$, we mean that there are $k$ number of relays available for deployment. Let $\Gamma_k(s)$ be the set of minimizers of (\[eqn:value\_iteration\]) at the $k$-th iteration at state $s$, if the infimum is achieved at some $a<\infty$. Let $\Gamma_{\infty}(s):=\{a \in \mathcal{A}:a$ be an accumulation point of some sequence $\{a_k\}$ where each $a_k \in \Gamma_{k}(s)\}$. Let $\Gamma^(s)$ be the set of minimizers in (\[eqn:bellman\_beta\_leq\_rho\]). In Appendix \[appendix:sequential\_placement\_total\_power\], we show that $\Gamma_k(s)$ for each $k \geq 1$, $\Gamma_{\infty}(s)$ and $\Gamma^(s)$ are nonempty. \[theorem:convergence\_of\_value\_iteration\] The value iteration given by (\[eqn:value\_iteration\]) has the following properties: 1. $J_{\xi}^{(k)}(s)\rightarrow J_{\xi}(s)$ for all $s \in (0,1]$, i.e., the value iteration converges to the optimal value function. 2. $\Gamma_{\infty}(s) \subset \Gamma^(s)$. 3. There is a stationary optimal policy $f^{\infty}=\{f,f,f,\cdots\}$ where $f:(0,1] \rightarrow \mathcal{A}$ and $f(s) \in \Gamma_{\infty}(s)$ for all $s \in (0,1]$. The proof is given in Appendix \[appendix:proof\_of\_value\_iteration\_convergence\]. It uses some results from [@schal75conditions-optimality], which have been discussed first. Next, we provide a general theorem (Theorem \[thm:value\_iteration\_general\]) on the convergence of value iteration, which has been used to prove Theorem \[theorem:convergence\_of\_value\_iteration\]. [Remark:]{} Since the action space is noncompact, it is not obvious from standard results whether the optimal policy exists. However, we are able to show that in our problem, for each state $s \in (0,1]$, the optimal action will lie in a compact set of the from $[0,a(s)]$, where $a(s)$ is continuous in $s$, and $a(s)$ could possibly go to $\infty$ as $s \rightarrow 0$. The results of [@schal75conditions-optimality] allow us to work with the scenario where for each state $s$, it is sufficient to focus only on a compact action space $[0,a(s)]$. Properties of the Value Function $J_{\xi}(s)$ --------------------------------------------- \[prop:increasing\_concave\_in\_s\] $J_{\xi}(s)$ is increasing and concave over $s \in (0,1]$. \[prop:increasing\_concave\_in\_lambda\] $J_{\xi}(s)$ is increasing and concave in $\xi$ for all $s \in (0,1]$. \[prop:continuity\_of\_cost\] $J_{\xi}(s)$ is continuous in $s$ over $(0,1]$ and continuous in $\xi$ over $(0,\infty)$. See Appendix \[appendix:proof\_of\_propositions\] for the proofs of these propositions. A Useful Normalization {#subsection:a_useful_normalization} ---------------------- Note that, $\beta L$ is exponentially distributed with mean $1$. Defining $\Lambda:=\frac{\rho}{\beta×}$ and $\tilde{z}_k:=\beta y_k$, $k=1,2,\cdots,(N+1)$, we can rewrite (\[eqn:unconstrained\_mdp\]) as follows: $$\begin{aligned} \inf_{\pi\in \Pi } \mathbb{E}_{\pi} \left(e^{\Lambda \tilde{z}_1}+ \sum_{k=2}^{N+1}\frac{e^{\Lambda \tilde{z}_k}-e^{\Lambda \tilde{z}_{k-1}}}{1+e^{\Lambda \tilde{z}_1}+ \cdots+e^{\Lambda \tilde{z}_{k-1}}×}+\xi N \right)\label{eqn:unconstrained_mdp_beta_one}\end{aligned}$$ Note that, $\Lambda$ plays the same role as $\lambda$ played in the known $L$ case (see Section \[subsection:optimal\_placement\_single\_relay\_sum\_power\]). Since $\frac{1}{\beta×}$ is the mean length of the line, $\Lambda$ can be considered as a measure of attenuation in the network. We can think of the new problem (\[eqn:unconstrained\_mdp\_beta\_one\]) in the same way as (\[eqn:unconstrained\_mdp\]), but with the length of the line being exponentially distributed with mean $1$ ($\beta'=1$) and the path-loss exponent being changed to $\rho'=\Lambda=\frac{\rho}{\beta×}$. The relay locations are also normalized ($\tilde{z}_k=\beta y_k$). One can solve the new problem (\[eqn:unconstrained\_mdp\_beta\_one\]) and obtain the optimal policy. Then the solution to (\[eqn:unconstrained\_mdp\]) can be obtained by multiplying each control distance (from the optimal policy of (\[eqn:unconstrained\_mdp\_beta\_one\])) with the constant $\frac{1}{\beta×}$. Hence, it suffices to work with $\beta=1$. A Numerical Study of As-You-Go Deployment {#section:numerical_work_information_theoretic_model} ========================================= ![$\beta=1$, $\Lambda:=\frac{\rho}{\beta}=2$; $a^$ vs. $s$.[]{data-label="fig:control_vs_state"}](control_vs_state.pdf){height="3cm" width="8cm"} ![$\beta=1$, $\Lambda:=\frac{\rho}{\beta×}=2$; $a^$ vs. $\xi$.[]{data-label="fig:control_vs_xi"}](control_vs_xi.pdf){height="3cm" width="8cm"} ![$\beta=1$, $\xi=0.01$; $a^$ vs. $\Lambda$.[]{data-label="fig:control_vs_rho"}](control_vs_rho.pdf){height="3cm" width="8cm"} Let us recall that the state of the system after placing the $k$-th relay is given by $s_k=\frac{e^{\Lambda \tilde{z}_k}}{\sum_{i=0}^{k}e^{\Lambda \tilde{z}_i}}$. The action is the normalized distance of the next relay to be placed from the current location. The single stage cost function for our total cost minimization problem is given by (\[eqn:single-stage-cost\]). In our numerical work, we discretized the state space $(0,1]$ into $100$ steps as $\{0.01,0.02,\cdots,0.99,1\}$, and discretized the action space into steps of size $0.001$, i.e., the action space becomes $\{0,0.001,0.002,\cdots\}$. Structure of the Optimal Policy {#subsection:policy_structure_numerical} ------------------------------- We performed numerical experiments to study the structure of the optimal policy obtained through value iteration for $\beta=1$ and some values of $\Lambda:=\frac{\rho}{\beta}$. The value iteration in these experiments converged and we obtained a stationary optimal policy, though Theorem \[theorem:convergence\_of\_value\_iteration\] does not guarantee the uniqueness of the stationary optimal policy. Figure \[fig:control\_vs\_state\] shows that the [normalized]{} optimal placement distance $a^$ is decreasing with the state $s \in (0,1]$. This can be understood as follows. The state $s$ (at a placement point) is small only if a sufficiently large number of relays have already been placed.[^9] Hence, if several relays have already been placed and $\sum_{i=0}^{k}e^{\Lambda \tilde{z}_i}$ is sufficiently large compared to $e^{\Lambda \tilde{z}_k}$ (i.e., $s_k$ is small), the $(k+1)$-st relay will be able to receive sufficient amount of power from the previous nodes, and hence does not need to be placed close to the $k$-th relay. A value of $s_k$ close to $1$ indicates that there is a large gap between relay $k$ and relay $k-1$, the power received at the next relay from the previous relays is small and hence the next relay must be placed closer to the previous one. On the other hand, $a^$ is increasing with $\xi$ (see Figure \[fig:control\_vs\_xi\]). Recall that $\xi$ is the price of placing a relay. This figure confirms the intuition that if the relay price is high, then the relays should be placed less frequently. -------------------- ----------------------------------------------------------- -------- $\mathbf{\Lambda}$ Normalised Optimal distances of the nodes No. of from the source relays $0.01$ 0, 0, 8.4180, 10.0000 3 $0.1$ 0, 0, 0, 0, 0, 0, 0, 0.2950, 0.5950, 0.9810, 1.3670, 33 1.7530, 2.1390, $\cdots,$ 9.0870, 9.4730, 9.8590, 10.0000 $5$ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.0020, 0.0080, 0.0140, 1677 0.0200, $\cdots$, 9.9860, 9.9920, 9.9980, 10.0000 -------------------- ----------------------------------------------------------- -------- : Sequential placement on a line of length $10$ for various $\Lambda$, using the corresponding optimal policies for $\xi=0.001$.[]{data-label="table:effect_of_rho_on_placement_1"} $\mathbf{\Lambda}$ Evolution of state in the process of sequential placement -------------------- ------------------------------------------------------------------------ $0.01$ 1, 0.5, 0.34, 0.27 $0.1$ 1, 0.5, 0.34, 0.26, 0.21, 0.18, 0.16, 0.14, 0.13, 0.12, 0.12, $\cdots$ $5$ 1, 0.5, 0.34, 0.26, 0.21, 0.18, 0.16, 0.14, 0.13, 0.12, 0.11, 0.1, 0.1, 0.1, $\cdots$ : Evolution of state in the process of sequential placement on a line of length $10$ for various values of $\Lambda$, using the corresponding optimal policies for $\xi=0.001$.[]{data-label="table:state-evolution_1"} -------------------- ----------------------------------------------------------- -------- $\mathbf{\Lambda}$ Normalised Optimal distances of the nodes No. of from the source relays $0.01$ 10 0 $0.1$ 5.3060, 10.0000 1 $5$ 0, 0.0050, 0.0510, 0.1220, 0.1930, 143 0.2640,$\cdots$, 9.9910, 10.0000 $8$ 0, 0.003, 0.019, 0.06, 0.101, $\cdots$, 9.982, 10 246 $20$ 0, 0.001, 0.003, 0.016, 0.031, 0.046, $\cdots$, 9.991, 10 669 -------------------- ----------------------------------------------------------- -------- : Sequential placement on a line of length $10$ for various $\Lambda$, using the corresponding optimal policies for $\xi=0.1$.[]{data-label="table:effect_of_rho_on_placement_3"} $\mathbf{\Lambda}$ Evolution of state in the process of sequential placement -------------------- ----------------------------------------------------------- $0.01$ 1 $0.1$ 1, 0.63 $5$ 1, 0.5, 0.34, 0.3, 0.3, $\cdots$ $8$ 1, 0.5, 0.34, 0.28, 0.28, $\cdots$ $20$ 1, 0.5, 0.34, 0.27, 0.26, 0.26, $\cdots$ : Evolution of state in the process of sequential placement on a line of length $10$ for various values of $\Lambda$, using the corresponding optimal policies for $\xi=0.1$.[]{data-label="table:state-evolution_3"} Figure \[fig:control\_vs\_rho\] shows that $a^$ is decreasing with $\Lambda$, for fixed values of $\xi$ and $s$. This happens because increased attenuation will require frequent placement of the relays. Relay Placement Patterns {#subsection:numerical_relay_plecement_pattern_as_you_go} ------------------------ The policy that we use corresponds to a line having exponentially distributed length with mean $1$, but it is applied to the scenario where the actual realization of the (normalised) length (see Section \[subsection:a\_useful\_normalization\]) of the line is $10$. Tables \[table:effect\_of\_rho\_on\_placement\_1\], \[table:effect\_of\_rho\_on\_placement\_3\], and \[table:effect\_of\_rho\_on\_placement\_4\] illustrate some examples of as-you-go placement of relay nodes along a line of normalised length $10$, using various values of $\Lambda$ and $\xi$. Tables \[table:state-evolution\_1\], \[table:state-evolution\_3\], and \[table:state-evolution\_4\] illustrate the corresponding evolution of state as the relays are placed in the examples in Tables \[table:effect\_of\_rho\_on\_placement\_1\], \[table:effect\_of\_rho\_on\_placement\_3\] and \[table:effect\_of\_rho\_on\_placement\_4\]. If the line actually ends at some point before (normalised) distance $10$, the process would end there with the corresponding placement of relays (as can be obtained from Tables \[table:effect\_of\_rho\_on\_placement\_1\], \[table:effect\_of\_rho\_on\_placement\_3\], and \[table:effect\_of\_rho\_on\_placement\_4\]) before the sink being placed at the end-point. Thus, for example, reading from Table \[table:effect\_of\_rho\_on\_placement\_3\] for $\xi=0.1$ and $\Lambda=5$, if the actual normalised length of the line is $0.99$, then one relay will be placed at $0$ (the source itself), followed by $15$ relays at normalised distances $0.005, 0.051, 0.122, 0.193, 0.264, \cdots, 0.974$ from the source, and finally the sink is placed at a normalised distance $0.99$, the end of the line. We observe that as $\Lambda$ increases, more relays need to be placed since the optimal control decreases with $\Lambda$ for each $s$ (see Figure \[fig:control\_vs\_rho\]). On the other hand, the number of relays decreases with increasing $\xi$ (the relay cost); this is in confirmation of the observations from Figure \[fig:control\_vs\_xi\]. Note that, initially one or more relays are placed at or near the source if $a^(s=1)$ is $0$ or small. But, after some relays have been placed, the relays are placed equally spaced apart. We see that this happens because, after a few relays have been placed, the state, $s$, does not change, hence, resulting in the relays being subsequently placed equally spaced apart. This phenomenon is evident in Table \[table:state-evolution\_1\], Table \[table:state-evolution\_3\], Table \[table:state-evolution\_4\], and Figure \[fig:state\_evolution\]. The state $s$ will remain unchanged after a relay placement if $s=\lceil{\frac{se^{\Lambda a^(s)}}{0.01(1+se^{\Lambda a^(s)})×}}\rceil \times 0.01$, since we have discretized the state space. After some relays are placed, the state becomes equal to a fixed point $s'$ of the function $\lceil{\frac{se^{\Lambda a^(s)}}{0.01(1+se^{\Lambda a^(s)})×}}\rceil \times 0.01$. Note that the deployment starts from $s_{0}:=1$, but for any value of $s_{0}$ (even with $s_{0}$ smaller than $s'$), we numerically observe the same phenomenon. Hence, $s'$ is an absorbing state. ------- ------------------------------------------------- -------- $\xi$ Normalised Optimal distances of the nodes No. of from the source relays $0.2$ 0 , 0.008, 0.03, 0.052, $\cdots$, 9.996, 10 456 $1$ 0.022, 0.069, 0.116, $\cdots$, 9.986, 10 213 $2$ 0.042, 0.103, 0.163, 0.223, $\cdots$, 9.943, 10 166 $10$ 0.099, 0.205, 0.311, $\cdots$, 9.957, 10 94 ------- ------------------------------------------------- -------- : Sequential placement on a line of length $10$ for of $\xi$, using the corresponding optimal policies for $\Lambda=20$.[]{data-label="table:effect_of_rho_on_placement_4"} $\xi$ Evolution of state in the process of sequential placement ------- ----------------------------------------------------------- $0.2$ 1, 0.5, 0.37, 0.37, $\cdots$ $1$ 1, 0.61, 0.61, $\cdots$ $2$ 1, 0.7, 0.71, 0.71, $\cdots$ $10$ 1, 0.88, 0.88, $\cdots$ : Evolution of state in the process of sequential placement on a line of length $10$ for various values of $\xi$, using the corresponding optimal policies for $\Lambda=20$.[]{data-label="table:state-evolution_4"} Numerical Examples for Practical Deployment {#subsection:numerical-example-practical-deployment} ------------------------------------------- In order to provide a more concrete illustration we adopt a path loss parameter from [@franceschetti-etal04random-walk-model-wave-propagation]. Figure $4$ of [@franceschetti-etal04random-walk-model-wave-propagation] shows that the attenuation in the received signal power in a dense urban environment is roughly $50$ dB when we move from $50$ m distance to $300$ m distance away from the transmitter. This yields a value of $\rho$ to be $0.04$ per meter for the exponential path-loss (see the discussion in Section \[subsection:motivation-for-exponential-path-loss\] on the motivation for choosing the exponential path-loss model in the light of the results from [@franceschetti-etal04random-walk-model-wave-propagation]). Then, $\frac{1}{\beta}=200$ m corresponds to $\Lambda=8$, and $\frac{1}{\beta}=500$ m corresponds to $\Lambda=20$. For $\Lambda=20$, normalised relay locations and state evolution $\{s_k\}_{k \geq 1}$ are available in Tables \[table:effect\_of\_rho\_on\_placement\_3\]-\[table:state-evolution\_4\], and, for $\Lambda=8$, normalised relay locations and state evolution $\{s_k\}_{k \geq 1}$ are available in Tables \[table:effect\_of\_rho\_on\_placement\_3\]-\[table:state-evolution\_3\]. Note that, under $\rho=0.04$ per meter and $\Lambda=20$, one unit normalised distance in the tables correspond to $500$ m distance in the dense urban environment (due to the normalization as in Section \[subsection:a\_useful\_normalization\]). For the sake of illustration, let us consider the sample deployment for $\xi=10$, $\Lambda=20$ (Table \[table:effect\_of\_rho\_on\_placement\_4\]). In this case, the first relay will be placed at a distance $0.099 \times 500=49.5$ m from the source, the second relay will be placed at a distance $0.205 \times 500=102.5$ m from the source, etc. Also, if we choose $\xi$ such that few relays will be placed on a typical line whose length is several hundreds of meters, then the relays will be placed almost uniformly on the line. But, for small $\xi$, more relays will be placed and some of them will be clustered near the source (see the deployment for $\Lambda=8$ and $\xi=0.1$ in Table \[table:effect\_of\_rho\_on\_placement\_3\]). ![Evolution of the state $s_k$ with $k$: initial state $s_0=1$.[]{data-label="fig:state_evolution"}](state_evolution.pdf){height="3cm" width="8cm"} ------- ----------- ------------ ------------- ------------- ------------ Average Mean Number of Maximum $\xi$ $\Lambda$ percentage number of cases where percentage difference relays used no relay difference was used 0.001 0.01 0.0068 2.0002 0 0.7698 0.001 0.1 0.3996 9.4849 0 6.8947 0.01 0.01 0 0 10000 0 0.01 0.1 0.3517 2.2723 0 4.6618 0.01 0.5 1.5661 7.7572 0 4.7789 0.1 0.01 0 0 10000 0 0.1 0.1 0.1259 0.0056 9944 25.9098 0.1 0.5 2.9869 1.8252 0 12.5907 0.1 2 4.7023 7.1530 0 9.0211 0.1 20 3.5472 27.9217 0 6.6223 0.1 8 4.0097 21.0671 0 7.8264 1 8 8.0286 7.8886 495 27.7362 1 20 5.2158 11.2342 402 26.0026 5 20 10.3341 7.1950 597 61.7460 ------- ----------- ------------ ------------- ------------- ------------ : Comparison of the performance (in terms of $H$; see text) of optimal sequential placement over a line of random length, with the optimal placement if the length was known. Results from 10000 samples of exponentially distributed line lengths.[]{data-label="table:comparison_optimal_mdp"} Comparison with Optimal Offline Deployment {#subsubsection:numerical_performance_as_you_go} ------------------------------------------ Since there is no prior work in the literature with which we can make a fair comparison of our as-you-go deployment policy for the full-duplex wireless multi-relay network, we compare the performance of our policy with optimal offline deployment. Thus, the numerical experiments reported in Table \[table:comparison\_optimal\_mdp\] are a result of asking the following question: how does the cost of as-you-go deployment over a line of exponentially distributed length compare with the cost of placing the same number of relays optimally over the line, once the length of the line has been revealed? For several combinations of $\Lambda$ and $\xi$, we generated $10000$ random numbers independently from an exponential distribution with parameter $\beta=1$. Each of these numbers was considered as a possible realization of the length of the line. Then we computed the placement of the relay nodes for each realization by optimal sequential placement policy, which gave us $H=\frac{1}{g_{0,1}×}+\sum_{k=2}^{N+1}\frac{(g_{0,k-1}-g_{0,k})}{g_{0,k}g_{0,k-1}\sum_{l=0}^{k-1}\frac{1}{g_{0,l}×}×}$, a quantity that we use to evaluate the quality of the relay placement. The significance of $H$ can be recalled from (\[eqn:capacity\_multirelay\]) where we found that the rate $C(1+\frac{P_T/\sigma^2}{H})$ can be achieved if total power $P_T$ is available to distribute among the source and the relays; i.e., $H$ can be interpreted as the net effective attenuation after power has been allocated optimally over the nodes. Also, for each realization, we computed $H$ for optimal relay placement, assuming that the length of the line is known before deployment and that the number of relays available is the [same]{} as the number of relays used by the corresponding sequential placement policy. For a given combination of $\Lambda$ and $\xi$, for the $k$-th realization of the length of the line, let us denote the two $H$ values by $H_{\mathsf{sequential}}^{(k)}$ and $H_{\mathsf{optimal}}^{(k)}$. Then the percentage [difference]{} for the $k$-th realization is: $$e_k:= \frac{\|H_{\mathsf{optimal}}^{(k)}-H_{\mathsf{sequential}}^{(k)}\|}{H_{\mathsf{optimal}}^{(k)}} \times 100 \label{eqn:error_or_difference_expression}$$ The average percentage difference in Table \[table:comparison\_optimal\_mdp\] is the quantity $\frac{\sum_{k=1}^{10000}e_k}{10000×}$. The maximum percentage difference is the quantity $\max_{k \in \{1,2,\cdots,10000\}}e_k$. Discussion of Table \[table:comparison\_optimal\_mdp\]:* 1. For small enough $\xi$, some relays will be placed at the source itself. For example, for $\Lambda=0.01$ and $\xi=0.001$, we will place two relays at the source (Table \[table:effect\_of\_rho\_on\_placement\_1\]). After placing the first relay, the next state will become $s=0.5$, and $a^(s=0.5)=0$. The state after placing the second relay becomes $s=0.34$, for which $a^(s=0.34)=8.41$ (see the placement in Table \[table:effect\_of\_rho\_on\_placement\_1\]). Now, the line having an exponentially distributed length with mean $1$ will end before $a^(s=0.34)=8.41$ distance with high probability, and the probability of placing the third relay will be very small. As a result, the mean number of relays will be $2.0002$. In case only $2$ relays are placed by the sequential deployment policy and we seek to place $2$ relays optimally for the same length of the line (with the length known), the optimal locations for both relays are close to the source location if the length of the line is small (i.e., if the attenuation $\lambda$ is small, recall the definition of $\lambda$ from Section \[subsection:optimal\_placement\_single\_relay\_sum\_power\]). If the line is long (which has a very small probability), the optimal placement will be significantly different from the sequential placement. Altogether, the difference (from (\[eqn:error\_or\_difference\_expression\])) will be small. 2. For $\Lambda=0.01$ and $\xi=0.1$, $a^(1)$ is so large that with high probability the line will end in a distance less than $a^(1)$ and no relay will be placed. 3. From (\[eqn:capacity\_multirelay\]) we know that for a given placement of relays on a line of given length $L$, the optimal power allocation yields an achievable rate $\log_2(1+\frac{P_T/\sigma^2}{H})$. At the end of as-you-go deployment the power is allocated optimally among the nodes deployed, and a rate $\log_2(1+\frac{P_T/\sigma^2}{H_{sequential}})$ can be achieved. If the same number of relays are optimally placed over the same line, with the same total power, then the inner bound is given by $\log_2(1+\frac{P_T/\sigma^2}{H_{optimal}})$. We seek to compare these two rates numerically. The maximum fractional difference in Table \[table:comparison\_optimal\_mdp\] is less than $\frac{2}{3}$, and substantially smaller than $\frac{2}{3}$ in most cases. Since, in (\[eqn:error\_or\_difference\_expression\]), $H_{sequential}^{(k)}$ is always greater than $H_{optimal}^{(k)}$, we have $H_{sequential}^{(k)} \leq \frac{5}{3} H_{optimal}^{(k)}$ for all $k \geq 1$ (i.e., for all realizations of $L$ in the simulation). Now, by the monotonicity of $\log_2(\cdot)$: $$\begin{aligned} && \frac{1}{2} \log_2 \bigg(1+\frac{P_T/\sigma^2}{H_{optimal}^{(k)}} \bigg)- \frac{1}{2} \log_2 \bigg(1+\frac{P_T/\sigma^2}{H_{sequential}^{(k)}} \bigg) \nonumber\\ & \leq & \frac{1}{2} \log_2 \bigg(1+\frac{P_T/\sigma^2}{H_{optimal}^{(k)}} \bigg)- \frac{1}{2} \log_2 \bigg(1+\frac{P_T/\sigma^2}{\frac{5}{3} H_{optimal}^{(k)}} \bigg) \nonumber\\\end{aligned}$$ Since $\log_2(\cdot)$ is a concave function, for any $x>y>0$, we have $\log_2 (1+x)-\log_2(1+y) \leq \log_2 (x) -\log_2 (y)$. Using this inequality, we can upper bound the difference in achievable rate from the previous equation by: $$\begin{aligned} \frac{1}{2} \log_2 \bigg(\frac{P_T/\sigma^2}{H_{optimal}^{(k)}} \bigg)- \frac{1}{2} \log_2 \bigg(\frac{P_T/\sigma^2}{\frac{5}{3} H_{optimal}^{(k)}} \bigg) = 0.3685\end{aligned}$$ This calculation implies that, for the large number of cases reported in Table \[table:comparison\_optimal\_mdp\], by using the approximation in (\[eqn:constrained\_mdp\]) and by using the corresponding optimal policy for as-you-go deployment, we lose at most $0.3685$ bits per channel use, compared to the case when the realization of the exponentially distributed source to sink distance is known apriori and when we use the same number of relays as used in the as-you-go deployment case. Note that, the statement of this claim holds [with high probability]{} since the maximum difference is taken over $10000$ sample deployments. Hence, it is reasonable to solve (\[eqn:constrained\_mdp\]) instead of (\[eqn:constrained\_mdp\_actual\]) which is intractable. Discussion {#section:additional_discussion} ========== Exponential Path-Loss Model {#subsection:motivation-for-exponential-path-loss} --------------------------- Exponential path-loss model has been used before in the context of relay placement (see [@firouzabadi-martins08optimal-node-placement], [@appuswamy-etal10relay-placement-deterministic-line]) and in the context of cellular networks (see [@altman-etal11greec-cellular Section $2.3$]). Analytical and experimental support for the exponential path-loss model have been provided by Franceschetti et al. ([@franceschetti-etal04random-walk-model-wave-propagation]). Franceschetti et al. used a random scattering model (applicable to an urban environment, or a forest environment) to show that the path-loss in such an environment is the product of an exponential function and a power function of the distance (see [@franceschetti-etal04random-walk-model-wave-propagation Equation $(14)$]). Figure $4$ of their paper, which is obtained from measurements made in an urban environment, shows that path-loss (in dB) varies linearly with distance beyond a distance of $40-50$ meters, which implies exponential path-loss for longer distance. These distances are practical for urban scenarios where the network is deployed over several hundreds of meters or several kilometers. Exponential path-loss was also proposed by Marano and Franceschetti for urban environment, and validated by theory and experiment (see [@marano-franceschetti05ray-propagation-random-lattice Figure $10$]).[^10] Incorporating Shadowing and Fading {#subsection:shadowing-fading} ---------------------------------- Shadowing (which is typically viewed as being static once a link is deployed) and time varying fading, can be incorporated in our setting by providing a fade-margin in the power at each transmitter. Thus, when expressed in dBm, the actual transmit power for any transmitter-receiver pair is the fade-margin plus the power used in the information theoretic capacity formulas; this fade-margin does not depend on the distance between the transmitter-receiver pair. Note that, this approach, though very conservative in nature, can remove the complexity in analysis arising out of fading in the network. Also note that, if the actual power gain between two nodes $r$ distance apart is $c_0e^{-\rho r}$ with $c_0 > 0$, then $c_0$ can be absorbed in the fade margin. Full-Duplex Decode-and-Forward Relaying {#subsection:motivation-for-full-duplex-decode-forward} --------------------------------------- Full-duplex radios might become a reality soon; see [@khandani13two-way-full-duplex-wireless], [@khandani10spatial-multiplexing-two-way-channel], [@choi-etal10single-channel-full-duplex], [@jain-etal11real-time-full-duplex] for recent efforts to realize them. Decode-and-forward relaying requires symbol-level synchronous operation across all nodes in the network. The requirement of globally coherent transmission and reception seems to be restrictive at the moment, but this problem will be solved with the advent of better clocks (with less drift) and efficient clock synchronization algorithms. Any research on impromptu deployment assuming imperfect synchronization, or half-duplex communication, or no interference cancellation, can use this paper as a benchmark for performance analysis. Insights on Power-Law Path-Loss {#subsection:insights_for_power_law_from_exponential} ------------------------------- In [@chattopadhyay-etal12optimal-capacity-relay-placement-line], we studied the problem of single-relay placement under a per-node power constraint at the source and the relay, for both exponential and power-law path-loss models. The variation of optimal relay location, as the amount of attenuation in the network varies, follow slightly different (but mostly similar) trends (see Figures $2$ and $3$ of [@chattopadhyay-etal12optimal-capacity-relay-placement-line]) because of the fact that power-law model allows unbounded power gain (unlike the exponential model) when the distance $r$ tends to $0$ ($\lim_{r \rightarrow 0} r^{-\eta}=\infty$). The findings are even more similar when we bound the power gain from above by some constant value in case of the power-law model (power gain is $\min\{r^{-\eta}, b^{-\eta} \}$ for some $b > 0$); see the similarity between Figures $2$ and $4$ in [@chattopadhyay-etal12optimal-capacity-relay-placement-line]. The results on the fixed node power case provide the insight that when the power gain is $r^{-\eta}$ or $\min\{r^{-\eta}, b^{-\eta} \}$; under the sum power constraint, the variation of the relay locations as a function of attenuation will follow a pattern similar to that in case of exponential path-loss. Conclusion ========== Motivated by the problem of as-you-go deployment of wireless relay networks, we first studied the problem of placing relay nodes along a line, in order to connect a sink at the end of the line to a source at the start of the line, so as to maximize the end-to-end achievable data rate. For the multi-relay channel with exponential path-loss and sum power constraint, we derived an expression for the achievable rate in terms of the power gains among all possible node pairs, and formulated an optimization problem in order to maximize the end-to-end data rate. Numerical work for the fixed source-sink distance suggests that at low attenuation the relays are mostly clustered close to the source in order to be able to cooperate among themselves, whereas at high attenuation they are uniformly placed and work as repeaters. Next, the deploy-as-you-go sequential placement problem was addressed; a sequential relay placement problem along a line having unknown random length was formulated as an MDP, the value function was characterized analytically, and the policy structure was investigated numerically. We found numerically that at the initial stage of the deployment process the inter-relay distances are smaller, and, as deployment progresses, the inter-relay distances increase gradually, and finally the relays start being placed at regular intervals. Our results are based on information theoretic achievable rate results. In order to utilize currently commercially available wireless devices, we have also been exploring non-information theoretic, packet forwarding models for optimal relay placement, with the aim of obtaining placement algorithms that can be easily reduced to practice (see [@chattopadhyay-etal13measurement-based-impromptu-placement_wiopt] for reference). The study of as-you-go deployment under the information theoretic model and under the packet forwarding model provides two complementary approaches for two different conditions in the physical layer and the MAC layer, and provides a more comprehensive development of the problem. [Arpan Chattopadhyay]{} obtained his B.E. in Electronics and Telecommunication Engineering from Jadavpur University, Kolkata, India in the year 2008, and M.E. and Ph.D in Telecommunication Engineering from Indian Institute of Science, Bangalore, India in the year 2010 and 2015, respectively. He is currently working in INRIA, Paris as a postdoctoral researcher. His research interests include networks and machine learning. [Abhishek Sinha]{} is currently a graduate student in the Laboratory for Information and Decision Systems (LIDS), at Massachusetts Institute of Technology, Cambridge, MA. Prior to joining MIT, he completed his Master’s studies in Telecommunication Engineering at the Indian Institute of Science, Bangalore, in the year 2012. His areas of interests include stochastic processes, information theory and network control. [Marceau Coupechoux]{} is an Associate Professor at Telecom ParisTech since 2005. He obtained his master from Telecom ParisTech in 1999 and from University of Stuttgart, Germany in 2000, and his Ph.D. from Institut Eurecom, Sophia-Antipolis, France, in 2004. From 2000 to 2005, he was with Alcatel-Lucent (Bell Labs former Research & Innovation and then in the Network Design department). In the Computer and Network Science department of Telecom ParisTech, he is working on cellular networks, wireless networks, ad hoc networks, cognitive networks, focusing mainly on layer 2 protocols, scheduling and resource management. From August 2011 to August 2012 he was a visiting scientist at IISc Bangalore. [Anurag Kumar]{} obtained his B.Tech. degree from the Indian Institute of Technology at Kanpur, and the PhD degree from Cornell University, both in Electrical Engineering. He was then with Bell Laboratories, Holmdel, N.J., for over 6 years. Since 1988 he has been on the faculty of the Indian Institute of Science (IISc), Bangalore, in the Department of Electrical Communication Engineering. He is currently also the Director of the Institute. From 1988 to 2003 he was the Coordinator at IISc of the Education and Research Network Project (ERNET), India’s first wide-area packet switching network. His area of research is communication networking, specifically, modeling, analysis, control and optimisation problems arising in communication networks and distributed systems. Recently his research has focused primarily on wireless networking. He is a Fellow of the IEEE, of the Indian National Science Academy (INSA), of the Indian Academy of Science (IASc), of the Indian National Academy of Engineering (INAE), and of The World Academy of Sciences (TWAS). He is a recepient of the Indian Institute of Science Alumni Award for Engineering Research for 2008. A Brief Description of the Coding Scheme of [@xie-kumar04network-information-theory-scaling-law] {#section:coding_scheme_description} ================================================================================================ Transmissions take place via block codes of $T$ symbols each. The transmission blocks at the source and the $N$ relays are synchronized. The coding and decoding scheme is such that a message generated at the source at the beginning of block $b, b \geq 1,$ is decoded by the sink at the end of block $b + N$, i.e., $N+1$ block durations after the message was generated (with probability tending to 1, as $T \to \infty)$. Thus, at the end of $B$ blocks, $B \geq N+1$, the sink is able to decode $B-N$ messages. It follows, by taking $B \to \infty$, that, if the code rate is $R$ bits per symbol, then an information rate of $R$ bits per symbol can be achieved from the source to the sink. As mentioned earlier, we index the source by $0$, the relays by $k, 1 \leq k \leq N$, and the sink by $N+1$. There are $(N+1)^2$ independent Gaussian random codebooks, each containing $2^{TR}$ codes, each code being of length $T$; these codebooks are available to all nodes. At the beginning of block $b$, the source generates a new message $w_b$, and, at this stage, we assume that each node $k, 1 \leq k \leq N+1,$ has a reliable estimate of all the messages $w_{b-j}, j \geq k$. In block $b$, the source uses a new codebook to encode $w_b$. In addition, relay $k, 1 \leq k \leq N,$ and [all]{} of its previous transmitters (indexed $0 \leq j \leq k-1$), use [another]{} codebook to encode $w_{b-k}$ (or their estimate of it). Thus, if the relays $1,2,\cdots,k$ have a perfect estimate of $w_{b-k}$ at the beginning of block $b$, they will transmit the same codeword for $w_{b-k}$. Therefore, in block $b$, the source and relays $1, 2, \cdots, k$ coherently transmit* the codeword for $w_{b-k}$. In this manner, in block $b$, transmitter $k, 0 \leq k \leq N,$ generates $N+1 - k$ codewords, corresponding to $w_{b-k}, w_{b-k-1}, \cdots, w_{b-N}$, which are transmitted with powers $P_{k,k+1}, P_{k,k+2}, \cdots, P_{k,N+1}$. In block $b$, node $k, 1 \leq k \leq N+1,$ receives a superposition of transmissions from all other nodes. Assuming that node $k$ knows all the powers, and all the channel gains, and recalling that it has a reliable estimate of all the messages $w_{b-j}, j \geq k$, it can subtract the interference from transmitters $k+1, k+2, \cdots, N$. At the end of block $b$, after subtracting the signals it knows, node $k$ is left with the $k$ received signals from nodes $0, 1, \cdots, (k-1)$ (received in blocks $b, b-1, \cdots, b-k+1$), which all carry an encoding of the message $w_{b-k+1}$. These $k$ signals are then jointly used to decode $w_{b-k+1},$ using joint typicality decoding. The codebooks are cycled through in a manner so that in any block all nodes encoding a message (or their estimate of it) use the same codebook, but different (thus, independent) codebooks are used for different messages. Under this encoding and decoding scheme, any rate strictly less than $R$ displayed in (\[eqn:achievable\_rate\_multirelay\]) is achievable. Proof of Theorem \[theorem:multirelay\_capacity\] {#appendix:proof_of_multirelay_channel_capacity_theorem_after_power_allocation} ================================================= We want to maximize $R$ given in (\[eqn:achievable\_rate\_multirelay\]) subject to the total power constraint, assuming fixed relay locations. Let us consider $C (\frac{1}{\sigma^{2}×} \sum_{j=1}^{k} ( \sum_{i=0}^{j-1} h_{i,k} \sqrt{P_{i,j}} )^{2})$, i.e., the $k$-th term in the argument of $\min \{\cdots\}$ in (\[eqn:achievable\_rate\_multirelay\]). By the monotonicity of $C(\cdot)$, it is sufficient to consider $\sum_{j=1}^{k} ( \sum_{i=0}^{j-1} h_{i,k} \sqrt{P_{i,j}} )^{2}$. Now since the channel gains are multiplicative, we have: $$\sum_{j=1}^{k} ( \sum_{i=0}^{j-1} h_{i,k} \sqrt{P_{i,j}} )^{2}=g_{0,k}\sum_{j=1}^{k} \bigg( \sum_{i=0}^{j-1}\frac{\sqrt{P_{i,j}}}{h_{0,i}×}\bigg)^{2}\nonumber$$ Thus our optimization problem becomes: $$\begin{aligned} & & \max \, \min_{k \in \{1,2,\cdots,N+1\}} g_{0,k} \sum_{j=1}^{k} \bigg( \sum_{i=0}^{j-1}\frac{\sqrt{P_{i,j}}}{h_{0,i}×}\bigg)^{2}\nonumber\\ & \textit{s.t} & \, \sum_{j=1}^{N+1}\gamma_{j} \leq P_{T} \,\, \textit{and} \,\sum_{i=0}^{j-1}P_{i,j}=\gamma_{j} \, \forall \, 1 \leq j \leq (N+1) \label{eqn:optimization_problem}\end{aligned}$$ Let us fix $\gamma_{1}, \gamma_{2},\cdots,\gamma_{N+1}$ such that their sum is equal to $P_{T}$. We observe that $P_{i,N+1}$ for $i \in \{0,1,\cdots,N\}$ appear in the objective function only once: for $k=N+1$ through the term $( \sum_{i=0}^{N}\frac{\sqrt{P_{i,N+1}}}{h_{0,i}×})^{2}$. Since we have fixed $\gamma_{N+1}$, we need to maximize this term over $P_{i,N+1},\, i \in \{0,1,\cdots,N\}$. So we have the following optimization problem: $$\begin{aligned} \max \sum_{i=0}^{N} \frac{\sqrt{P_{i,N+1}}}{h_{0,i}×} \,\,\,\,\, \textit{s.t} \,\,\,\,\, \sum_{i=0}^{N} P_{i,N+1}=\gamma_{N+1}\label{eqn:problem}\end{aligned}$$ By Cauchy-Schwartz inequality, the objective function in this optimization problem is upper bounded by (using the fact that $g_{0,i}=h_{0,i}^2$ $\forall i \in \{0,1,\cdots,N\}$): $$\sqrt{(\sum_{i=0}^{N} P_{i,N+1})(\sum_{i=0}^{N}\frac{1}{g_{0,i}×})}=\sqrt{\gamma_{N+1}\sum_{i=0}^{N}\frac{1}{g_{0,i}}}$$ The upper bound is achieved if there exists some $c>0$ such that $\frac{\sqrt{P_{i,N+1}}}{\frac{1}{h_{0,i}×}×}=c$ $\forall i \in \{0,1,\cdots,N\}$. So we have: $$P_{i,N+1}=\frac{c^2}{g_{0,i}×} \,\, \forall i \in \{0,1,\cdots,N\}\nonumber\\$$ Since $\sum_{i=0}^{N}P_{i,N+1}=\gamma_{N+1}$, we obtain $c^2=\frac{\gamma_{N+1}}{\sum_{l=0}^{N}\frac{1}{g_{0,l}×}×}$. Thus, $ P_{i,N+1}=\frac{\frac{1}{g_{0,i}×}}{\sum_{l=0}^{N}\frac{1}{g_{0,l}×}×}\gamma_{N+1}$. Here we have used the fact that $h_{0,0}=1$. Now $\{P_{i,N}: i=0,1,\cdots,(N-1)\}$ appear only through the sum $\sum_{i=0}^{N-1}\frac{\sqrt{P_{i,N}}}{h_{0,i}×}$, and it appears twice: for $k=N$ and $k=N+1$. We need to maximize this sum subject to the constraint $\sum_{i=0}^{N-1}P_{i,N}=\gamma_{N}$. This optimization can be solved in a similar way as before. Thus by repeatedly using this argument and solving optimization problems similar in nature to (\[eqn:problem\]), we obtain: $$P_{i,j}=\frac{\frac{1}{g_{0,i}×}}{\sum_{l=0}^{j-1}\frac{1}{g_{0,l}×}×}\gamma_{j} \,\, \forall 0 \leq i < j \leq (N+1)$$ Substituting for $P_{i,j},0 \leq i < j \leq (N+1)$ in (\[eqn:optimization\_problem\]), we obtain the following optimization problem: $$\begin{aligned} & & \max \min_{k \in \{1,2,\cdots,N+1\}} g_{0,k} \sum_{j=1}^{k}\bigg(\gamma_{j} \sum_{i=0}^{j-1}\frac{1}{g_{0,i}×} \bigg)\nonumber\\ & \textit{s.t.} & \,\, \sum_{j=1}^{N+1}\gamma_{j} \leq P_{T}\end{aligned}$$ Let us define $b_{k}:=g_{0,k}$ and $a_{j}:=\sum_{i=0}^{j-1}\frac{1}{g_{0,i}×}$. Observe that $b_{k}$ is decreasing and $a_{k}$ is increasing with $k$. Let us define: $$\tilde{s}_{k}(\gamma_{1},\gamma_{2},\cdots,\gamma_{N+1}) := b_{k} \sum_{j=1}^{k} a_{j} \gamma_{j}$$ With this notation, our optimization problem becomes: $$\begin{aligned} \max \min_{1 \leq k \leq N+1} \tilde{s}_{k}(\gamma_{1},\gamma_{2},\cdots,\gamma_{N+1}) \,\,\,\, \textit{s.t.} \,\, \sum_{j=1}^{N+1}\gamma_{j} \leq P_{T} \label{eqn:modified_optimization_problem}\end{aligned}$$ Under optimal allocation of $\gamma_{1}, \gamma_{2},\cdots,\gamma_{N+1}$ for the optimization problem (\[eqn:modified\_optimization\_problem\]), $\tilde{s}_{1}=\tilde{s}_{2}=\cdots=\tilde{s}_{N+1}$. [$\Box$]{} (\[eqn:modified\_optimization\_problem\]) can be rewritten as: $$\begin{aligned} && \max \zeta \nonumber\\ \textit{s.t} && \zeta \leq b_{k}\sum_{j=1}^{k}a_{j}\gamma_{j} \, \forall \, k \in \{1,2,\cdots,N+1\}, \nonumber\\ && \sum_{j=1}^{N+1}\gamma_{j} \leq P_{T}, \,\,\, \gamma_{j} \geq 0 \,\, \forall \, 1 \leq j \leq N+1\label{eqn:equality_problem_primal}\end{aligned}$$ The dual of this linear program is given by: $$\begin{aligned} &&\min P_{T}\theta \nonumber\\ \textit{s.t} && \sum_{k=1}^{N+1}\mu_{k}=1,\,\,\, \theta \geq 0,\nonumber\\ && a_{l}\sum_{k=l}^{N+1}b_{k}\mu_{k}+\nu_{l}=\theta \, \forall \, l \in \{1,2,\cdots,N+1\},\nonumber\\ && \mu_{l} \geq 0, \nu_{l} \geq 0 \, \forall \, l \in \{1,2,\cdots,N+1\} \label{eqn:equality_problem_dual}\end{aligned}$$ Now, let us consider a primal feasible solution $(\{\gamma_j^\}_{1 \leq j \leq N+1}, \zeta^)$ which satisfies: $$\begin{aligned} && b_{k}\sum_{j=1}^{k}a_{j}\gamma_{j}^=\zeta^ \, \forall \, k \in \{1,2,\cdots,N+1\}, \nonumber\\ && \sum_{j=1}^{N+1}\gamma_{j}^* = P_{T}\end{aligned}$$ Thus we have, $b_1 a_1 \gamma_1^=\zeta^$, i.e., $\gamma_1^=\frac{\zeta^}{b_1 a_1×}$. Again, $b_2(a_1 \gamma_1^+a_2 \gamma_2^)=\zeta^$, which implies $ \frac{b_2}{b_1×}\zeta^+b_2 a_2 \gamma_2^=\zeta^$. Thus we obtain $\gamma_2^=\frac{\zeta^}{a_2×}(\frac{1}{b_2×}-\frac{1}{b_1×})$. In general, we can write: $$\gamma_k^=\frac{\zeta^}{a_k×}\left(\frac{1}{b_k×}-\frac{1}{b_{k-1}×}\right) \, \forall k \in \{1,2,\cdots,N+1\}\nonumber\\$$ with $\frac{1}{b_{0}×}:=0$. Now, since $\sum_{k=1}^{N+1}\gamma_k^=P_T$, we obtain: $$\begin{aligned} \zeta^&=&\frac{P_{T}}{\sum_{k=1}^{N+1}\frac{1}{a_{k}×}(\frac{1}{b_k×}-\frac{1}{b_{k-1}×})×}\nonumber\\ \gamma_j^&=&\frac{ \frac{1}{a_j×} \left(\frac{1}{b_j×}-\frac{1}{b_{j-1}×}\right) } {\sum_{k=1}^{N+1}\frac{1}{a_{k}×}(\frac{1}{b_k×}-\frac{1}{b_{k-1}×})×}P_{T}, \, j \in \{1,2,\cdots,N+1\} \label{eqn:primal_optimal}\end{aligned}$$ It should be noted that since $b_{k}$ is nonincreasing in $k$, the primal variables above are nonnegative and satisfies feasibility conditions. Again, let us consider a dual feasible solution $(\{\mu_j^,\nu_j^\}_{1 \leq j \leq N+1}, \theta^)$ which satisfies: $$\begin{aligned} && \sum_{k=1}^{N+1}\mu_{k}^=1, \,\,\, \nu_l^=0 \, \forall \, l \in \{1,2,\cdots,N+1\}\nonumber\\ && a_{l}\sum_{k=l}^{N+1}b_{k}\mu_{k}^+\nu_{l}^=\theta^* \, \forall \, l \in \{1,2,\cdots,N+1\}\end{aligned}$$ Solving these equations, we obtain: $$\begin{aligned} \theta^&=&\frac{1}{\sum_{k=1}^{N+1}\frac{1}{b_k×}\left(\frac{1}{a_k×}-\frac{1}{a_{k+1}×}\right)×}\nonumber\\ \mu_j^&=&\frac{\frac{1}{b_j×}(\frac{1}{a_j×}-\frac{1}{a_{j+1}×})} {\sum_{k=1}^{N+1}\frac{1}{b_k×}\left(\frac{1}{a_k×}-\frac{1}{a_{k+1}×}\right)×},\, j \in \{1,2,\cdots,N+1\} \label{eqn:dual_optimal}\end{aligned}$$ where $\frac{1}{a_{N+2}×}:=0$. Since $a_k$ is increasing in $k$, all dual variables are feasible. It is easy to check that $\zeta^=P_T \theta^$, which means that there is no duality gap. Since the primal is a linear program, the solution $(\gamma_1^, \gamma_2^,\cdots,\gamma_{N+1}^, \zeta^)$ is primal optimal. Thus we have established the claim, since the primal optimal solution satisfies it. So let us obtain $\gamma_{1},\gamma_{2},\cdots,\gamma_{N+1}$ for which $\tilde{s}_{1}=\tilde{s}_{2}=\cdots=\tilde{s}_{N+1}$. Putting $\tilde{s}_{k}=\tilde{s}_{k-1}$, we obtain $b_{k} \sum_{j=1}^{k} a_{j} \gamma_{j}=b_{k-1} \sum_{j=1}^{k-1} a_{j} \gamma_{j}$. Thus, we obtain, $ \gamma_{k}=\frac{(b_{k-1}-b_{k})}{b_{k}×} \frac{1}{a_{k}×} \sum_{j=1}^{k-1}a_{j}\gamma_{j}$ Let $d_{k}:=\frac{(b_{k-1}-b_{k})}{b_{k}×} \frac{1}{a_{k}×}$. Hence, $ \gamma_{k}=d_{k} \sum_{j=1}^{k-1}a_{j}\gamma_{j}$. From this recursive equation, we have $\gamma_{2}=d_{2}a_{1}\gamma_{1}$, $\gamma_{3}=d_{3}(a_{1}\gamma_{1}+a_{2}\gamma_{2})=d_{3}a_{1}(1+a_{2}d_{2})\gamma_{1}$, and, in general for $k \geq 3$, $$\gamma_{k}=d_{k}a_{1}\Pi_{j=2}^{k-1}(1+a_{j}d_{j})\gamma_{1}\label{eqn:gamma_k_gamma_1}$$ Using the fact that $\gamma_{1}+\gamma_{2}+\cdots+\gamma_{N+1}=P_{T}$, we obtain: $$\gamma_{1}=\frac{P_{T}}{1+d_{2}a_{1}+ \sum_{k=3}^{N+1}d_{k}a_{1} \Pi_{j=2}^{k-1}(1+a_{j}d_{j}) ×} \label{eqn:gamma_1}$$ Thus if $\tilde{s}_{1}=\tilde{s}_{2}=\cdots=\tilde{s}_{N+1}$, there is a unique allocation $\gamma_{1},\gamma_{2},\cdots,\gamma_{N+1}$. So this must be the one maximizing $R$. Hence, optimum $\gamma_{1}$ is obtained by (\[eqn:gamma\_1\]). Then, substituting the values of $\{a_{k}:k=0,1,\cdots,N\}$ and $d_{k}:k=1,2,\cdots,N+1$ in (\[eqn:gamma\_k\_gamma\_1\]) and (\[eqn:gamma\_1\]), we obtain the values of $\gamma_{1},\gamma_{2},\cdots,\gamma_{N+1}$ as shown in Theorem \[theorem:multirelay\_capacity\]. Now under these optimal values of $\gamma_{1},\gamma_{2},\cdots,\gamma_{N+1}$, all terms in the argument of $\min \{\cdots\}$ in (\[eqn:achievable\_rate\_multirelay\]) are equal. So we can consider the first term alone. Thus we obtain the expression for $R$ optimized over power allocation among all the nodes for fixed relay locations as : $R_{P_T}^{opt}(y_1,y_2,\cdots,y_N)=C \left(\frac{g_{0,1}P_{0,1}}{\sigma^{2}×}\right)=C \left(\frac{g_{0,1}\gamma_{1}}{\sigma^{2}×}\right)$. Substituting the expression for $\gamma_{1}$ from (\[eqn:gamma\_one\]), we obtain the achievable rate formula (\[eqn:capacity\_multirelay\]).[$\Box$]{} $$\begin{aligned} \label{eqn:capacity_increasing_with_N} & & z_{1}^{}+\frac{z_{2}^{}-z_{1}^{}}{1+z_{1}^{}×}+\cdots+\frac{e^{\rho y}-z_{i}^{}}{1+z_{1}^{}+\cdots+z_{i}^{}×}+\frac{z_{i+1}^{}-e^{\rho y}}{1+z_{1}^{}+\cdots+z_{i}^{}+e^{\rho y}×}+\cdots.+\frac{e^{\rho L}-z_{N}^{}}{1+z_{1}^{}+\cdots+z_{i}^{}+e^{\rho y}+z_{i+1}^{}+\cdots+z_{N}^{}×}\nonumber\\ & & < z_{1}^{}+\frac{z_{2}^{}-z_{1}^{}}{1+z_{1}^{}×}+\cdots+\frac{e^{\rho L}-z_{N}^{}}{1+z_{1}^{}+\cdots+z_{i}^{}+z_{i+1}^{}+\cdots+z_{N}^{}×} \label{eqn:intermediate_eqn:capacity_increasing_in_N}\end{aligned}$$ ------------------------------------------------------------------------ Proof of Theorem \[theorem:single\_relay\_total\_power\] {#appendix:proof_of_single_relay_sum_power_results} ======================================================== Here we want to place the relay node at a distance $r_{1}$ from the source to minimize $\bigg\{\frac{1}{g_{0,1}×}+\frac{g_{0,1}-g_{0,2}}{g_{0,2}(1+g_{0,1})×}\bigg\}$ (see Equation (\[eqn:capacity\_multirelay\])). Hence, our optimization problem becomes : $$\min_{r_{1} \in [0,L]} \bigg\{e^{\rho r_{1}}+\frac{e^{-\rho r_{1}}-e^{-\rho L}}{e^{-\rho L}(1+e^{-\rho r_{1}})×}\bigg\}\nonumber\\$$ Writing $z_{1}=e^{\rho r_{1}}$, the problem becomes : $$\min_{z_{1} \in [1,e^{\rho L}]} \bigg\{z_{1}-1+\frac{e^{\rho L}+1}{z_{1}+1×}\bigg\}\nonumber\\$$ This is a convex optimization problem. Equating the derivative of the objective function to zero, we obtain $1-\frac{e^{\rho L}+1}{(z_{1}+1)^{2}×}=0$. Thus the derivative becomes zero at $z_{1}'=\sqrt{1+e^{\rho L}}-1>0$. Hence, the objective function is decreasing in $z_{1}$ for $z_{1} \leq z_{1}'$ and increasing in $z_{1} \geq z_{1}'$. So the minimizer is $z_{1}^{}=\max \{z_{1}',1 \}$. So the optimum distance of the relay node from the source is $y_{1}^{}=r_{1}^{}=\max \{0,r_{1}' \}$, where $r_{1}'=\frac{1}{\rho×} \log (\sqrt{1+e^{\rho L}}-1)$. Hence, $\frac{y_{1}^{}}{L×}=\max \{\frac{1}{\lambda×} \log \left(\sqrt{e^{\lambda}+1}-1 \right),0\}$. Now $r_{1}' \geq 0$ if and only if $\lambda \geq \log 3$. Hence, $\frac{y_{1}^{}}{L×}=0$ for $\lambda \leq \log 3$ and $\frac{y_{1}^{}}{L×}=\frac{1}{\lambda×} \log \left(\sqrt{e^{\lambda}+1}-1 \right)$ for $\lambda \geq \log 3$. [For $\lambda \leq \log 3$]{}, the relay is placed at the source. Then $g_{0,1}=1$ and $g_{0,2}=g_{1,2}=e^{-\lambda}$. Then $P_{0,1}=\gamma_{1}=\frac{2P_{T}}{e^{\lambda}+1×}$ (by Theorem $1$) and $R^{}=C \left(\frac{2P_{T}}{(e^{\lambda}+1)\sigma^{2}×}\right)$. Also $\gamma_{2}=\frac{e^{\lambda}-1}{e^{\lambda}+1×}P_{T}$. Hence, $P_{0,2}=P_{1,2}=\frac{e^{\lambda}-1}{e^{\lambda}+1×}\frac{P_{T}}{2×}$. [for $\lambda \geq \log 3$]{}, the relay is placed at $r_{1}'$. Substituting the value of $r_{1}'$ into Equation ($\ref{eqn:power_gamma_relation})$, we obtain $P_{0,1}=\gamma_{1}=\frac{P_{T}}{2×}$, $P_{0,2}=\frac{1}{\sqrt{e^{\lambda}+1}×}\frac{P_{T}}{2×}$, $P_{1,2}=\frac{\sqrt{e^{\lambda}+1}-1}{\sqrt{e^{\lambda}+1}×}\frac{P_{T}}{2×}$. So in this case $R^{}=C \left(\frac{g_{0,1}P_{0,1}}{\sigma^{2}×} \right)$. Since $P_{0,1}=\frac{P_{T}}{2×}$, we have $R^{}=C \left( \frac{1}{\sqrt{e^{\lambda}+1}-1×}\frac{P_{T}}{2 \sigma^{2}×} \right)$. [$\Box$]{} Proof of Theorem \[theorem:capacity\_increasing\_with\_N\] {#appendix:proof_of_capacity_increases_in_N} ========================================================== For the $N$-relay problem, let the minimizer in (\[eqn:multirelay\_optimization\]) be $z_{1}^{}, z_{2}^{},\cdots,z_{N}^{}$ and let $y_{k}^{}=\frac{1}{\rho×} \log z_{k}^{}$. Clearly, there exists $i \in \{0,1,\cdots,N\}$ such that $y_{i+1}^{}>y_{i}^{}$. Let us insert a new relay at a distance $y$ from the source such that $y_{i}^{}<y<y_{i+1}^{}$. Now we find that we can easily reach (\[eqn:capacity\_increasing\_with\_N\]) (see next page) just by simple comparison. For example, $$\begin{aligned} & & \frac{e^{\rho y}-z_{i}^{}}{1+z_{1}^{}+\cdots+z_{i}^{}×}+\frac{z_{i+1}^{}-e^{\rho y}}{1+z_{1}^{}+\cdots+z_{i}^{}+e^{\rho y}×} \\ & < & \frac{e^{\rho y}-z_{i}^{}}{1+z_{1}^{}+\cdots+z_{i}^{}×}+\frac{z_{i+1}^{}-e^{\rho y}}{1+z_{1}^{}+\cdots+z_{i}^{}×}\\ & =& \frac{z_{i+1}^{}-z_{i}^{}}{1+z_{1}^{}+\cdots+z_{i}^{}×}\end{aligned}$$ First $i$ terms in the summations of L.H.S (left hand side) and R.H.S (right hand side) of (\[eqn:capacity\_increasing\_with\_N\]) are identical. Also sum of the remaining terms in L.H.S is smaller than that of the R.H.S since there is an additional $e^{\rho y}$ in the denominator of each fraction for the L.H.S. Hence, we can justify (\[eqn:capacity\_increasing\_with\_N\]). Now R.H.S is precisely the optimum objective function for the $N$-relay placement problem (see (\[eqn:multirelay\_optimization\])). On the other hand, L.H.S is a particular value of the objective in (\[eqn:multirelay\_optimization\]), for $(N+1)$-relay placement problem. This clearly implies that by adding one additional relay we can strictly improve from $R^{}$ of the $N$ relay channel. Hence, $R^{}(N+1)>R^{}(N)$. [$\Box$]{} Proof of Theorem \[theorem:G\_increasing\_in\_lambda\] {#appendix:proof_of_G_increasing_in_lambda} ====================================================== Consider the optimization problem as shown in (\[eqn:multirelay\_optimization\]). Let us consider $\lambda_1$, $\lambda_2$, with $\lambda_1<\lambda_2$, the respective minimizers being $(z_{1}^{},\cdots,z_{N}^{})$ and $(z_{1}',\cdots,z_{N}')$. Clearly, $$\begin{aligned} G(N,\lambda_1)=\frac{e^{\lambda_1}}{ z_{1}^{}+\sum_{k=2}^{N+1} \frac{z_{k}^{}-z_{k-1}^{}}{\sum_{l=0}^{k-1} z_{l}^{}}×}\end{aligned}$$ with $z_{N+1}^{}=e^{\lambda_1}$ and $z_{0}^{}=1$. With $N \geq 1 $, note that $z_{1}^{}-\frac{z_{N}^{}}{1+z_{1}^{}+\cdots+z_{N}^{}×} \geq 0$, since $z_{1}^{} \geq 1$ and $\frac{z_{N}^{}}{1+z_{1}^{}+\cdots+z_{N}^{}×} \leq 1$. Hence, it is easy to see that $\frac{e^{\lambda}}{ z_{1}^{}-\frac{z_{N}^{}}{1+z_{1}^{}+\cdots+z_{N}^{}×}+\sum_{k=2}^{N} \frac{z_{k}^{}-z_{k-1}^{}}{\sum_{l=0}^{k-1} z_{l}^{}}+\frac{e^{\lambda}}{1+z_{1}^{}+\cdots+z_{N}^{}×}×}$ is increasing in $\lambda$ where $(z_{1}^{},\cdots,z_{N}^{})$ is the optimal solution of (\[eqn:multirelay\_optimization\]) with $\lambda=\lambda_1$. Hence, $$\begin{aligned} G(N,\lambda_1)&=&\frac{e^{\lambda_1}}{ z_{1}^{}+\sum_{k=2}^{N} \frac{z_{k}^{}-z_{k-1}^{}}{\sum_{l=0}^{k-1} z_{l}^{}}+\frac{e^{\lambda_1}-z_{N}^{}}{\sum_{l=0}^{N} z_{l}^{}}×}\nonumber\\ &\leq & \frac{e^{\lambda_2}}{ z_{1}^{}+\sum_{k=2}^{N} \frac{z_{k}^{}-z_{k-1}^{}}{\sum_{l=0}^{k-1} z_{l}^{}}+\frac{e^{\lambda_2}-z_{N}^{}}{1+z_{1}^{}+\cdots+z_{N}^{}×}×}\nonumber\\ &\leq & \frac{e^{\lambda_2}}{ z_{1}^{'}+\sum_{k=2}^{N} \frac{z_{k}^{'}-z_{k-1}^{'}}{\sum_{l=0}^{k-1} z_{l}^{'}}+\frac{e^{\lambda_2}-z_{N}^{'}}{1+z_{1}^{'}+\cdots+z_{N}^{'}×}×}\nonumber\\ &=& G(N,\lambda_2)\end{aligned}$$ The second inequality follows from the fact that $(z_{1}',\cdots,z_{N}')$ minimizes $z_{1}+\sum_{k=2}^{N+1} \frac{z_{k}-z_{k-1}}{\sum_{l=0}^{k-1} z_{l}}$ subject to the constraint $1 \leq z_1 \leq z_2 \leq \cdots \leq z_N \leq z_{N+1} =e^{\lambda_2}$. Hence, $G(N,\lambda)$ is increasing in $\lambda$ for fixed $N$. [$\Box$]{} Proof of Theorem \[theorem:large\_nodes\_uniform\] {#appendix:proof_of_large_nodes_uniform} ================================================== When $N$ relay nodes are uniformly placed along a line, we will have $y_{k}=\frac{kL}{N+1×}$. Then our formula for achievable rate $R_{P_T}^{opt}(y_1,y_2,\cdots,y_N)$ for sum power constraint becomes: $R_{N}=C(\frac{P_{T}}{\sigma^{2}×}\frac{1}{f(N)×})$ where $f(N)=a_N+\sum_{k=2}^{N+1}\frac{a_N^{k}-a_N^{k-1}}{1+a_N+\cdots+a_N^{k-1}×}$ with $a_N=e^{ \frac{\rho L}{N+1×}}=e^{\frac{\lambda}{N+1×}}$. Since $a_N>1$ for all $N<\infty$ and $\rho>0$, we have $f(N) > a_N$ for all $N \geq 1$ and hence, $\liminf_N f(N) \geq \lim_{N \rightarrow \infty} a_N =1$. Now, $$\begin{aligned} f(N)&=& a_N+\sum_{k=1}^{N}\frac{a_N^{k+1}-a_N^{k}}{1+a_N+\cdots+a_N^{k}×} \nonumber\\ &=& a_N+ (a_N-1)^{2}\sum_{k=1}^{N} \frac{a_N^{k}}{a_N^{k+1}-1×}\nonumber\\ &\leq& a_N+ (a_N-1)^{2}\sum_{k=1}^{N} \frac{a_N^{k}}{a_N^{k}-1×}\nonumber\\ &=& e^{\frac{\lambda}{N+1×}}+ (e^{\frac{\lambda}{N+1×}}-1)^{2}\sum_{k=1}^{N} \frac{e^{\frac{k \lambda}{N+1×}}}{e^{\frac{k \lambda}{N+1×}}-1×}\nonumber\\ &\leq& e^{\frac{\lambda}{N+1×}}+ (e^{\frac{\lambda}{N+1×}}-1)^{2}\sum_{k=1}^{N} \frac{e^{\frac{k \lambda}{N+1×}}}{\frac{k \lambda}{N+1×}}\nonumber\\ &=& e^{\frac{\lambda}{N+1×}}+ (e^{\frac{\lambda}{N+1×}}-1)^{2} \frac{(N+1)}{\lambda×} \sum_{k=1}^{N} \frac{e^{\frac{k \lambda}{N+1×}}}{k} \label{eqn:inequality_of_fN}\end{aligned}$$ where the first inequality follows from the fact that $a_N>1$ and the second inequality follows from the fact that $e^{\frac{k \lambda}{N+1×}} \geq 1+ \frac{k \lambda}{N+1×}$. Now, by Cauchy-Schwartz inequality, $$\begin{aligned} \sum_{k=1}^{N} \frac{e^{\frac{k \lambda}{N+1×}}}{k} \leq \sqrt{(\sum_{k=1}^{N}e^{\frac{2 k \lambda}{N+1×}}) (\sum_{k=1}^{N} \frac{1}{k^2×})}\end{aligned}$$ Since $\sum_{k=1}^{\infty}\frac{1}{k^2×}=\frac{\pi^2}{6×}$, we can write: $$\begin{aligned} \sum_{k=1}^{N} \frac{e^{\frac{k \lambda}{N+1×}}}{k} \leq \sqrt{(\sum_{k=1}^{N}e^{\frac{2 k \lambda}{N+1×}}) \frac{\pi^2}{6×}} \label{eqn:inequality_using_the_series}\end{aligned}$$ Hence, by (\[eqn:inequality\_using\_the\_series\]) and (\[eqn:inequality\_of\_fN\]), $$\begin{aligned} f(N) &\leq& e^{\frac{\lambda}{N+1×}}+ (e^{\frac{\lambda}{N+1×}}-1)^{2} \frac{(N+1)\pi}{\sqrt{6}\lambda×} \sqrt{\sum_{k=1}^{N}e^{\frac{2 k \lambda}{N+1×}}} \nonumber\\ &=& e^{\frac{\lambda}{N+1×}}+ (e^{\frac{\lambda}{N+1×}}-1)^{2} \frac{(N+1)\pi}{\sqrt{6}\lambda×} \sqrt{e^{\frac{2 \lambda}{N+1×}} \frac{(e^{\frac{2 N \lambda}{N+1×}}-1)}{(e^{\frac{2 \lambda}{N+1×}}-1)×}}\nonumber\\\end{aligned}$$ Now, since $e^{\frac{2 N \lambda}{N+1×}}-1 \leq e^{\frac{2 N \lambda}{N+1×}} \leq e^{2\lambda}$, we obtain: $$\begin{aligned} f(N) &\leq& e^{\frac{\lambda}{N+1×}}+ (e^{\frac{\lambda}{N+1×}}-1)^{2} \frac{(N+1)\pi}{\sqrt{6}\lambda×} e^{\lambda} e^{\frac{\lambda}{N+1×}} \sqrt{\frac{1}{(e^{\frac{2 \lambda}{N+1×}}-1)×}} \nonumber\\ &=& e^{\frac{\lambda}{N+1×}}+ (e^{\frac{\lambda}{N+1×}}-1)^{\frac{3}{2×}} \frac{(N+1)\pi}{\sqrt{6}\lambda×} e^{\lambda} e^{\frac{\lambda}{N+1×}} \sqrt{\frac{1}{(e^{\frac{ \lambda}{N+1×}}+1)×}} \nonumber\\\end{aligned}$$ Hence, $$\begin{aligned} && \limsup_N f(N) \nonumber\\ & \leq & 1+ \frac{\pi e^{\lambda}}{\sqrt{12}×} \lim_{N \rightarrow \infty} \frac{(N+1)}{\lambda} (e^{\frac{\lambda}{N+1×}}-1)^{\frac{3}{2×}} \nonumber\\\end{aligned}$$ Putting $q=\frac{\lambda}{N+1×}$, $$\begin{aligned} \limsup_N f(N) &\leq& 1+ \frac{\pi e^{\lambda}}{\sqrt{12}×} \lim_{q \rightarrow 0} \sqrt{q} \lim_{q \rightarrow 0} (\frac{e^q-1}{q×})^{\frac{3}{2×}} \nonumber\\ &=& 1 \nonumber\\\end{aligned}$$ Now, we have proved that $\limsup_N f(N) \leq 1 \leq \liminf_N f(N)$ and hence $\lim_{N \rightarrow \infty} f(N)=1$. Hence, $\lim_{N \rightarrow \infty} R_N = C(\frac{P_T}{\sigma^{2}})$ and the theorem is proved. [$\Box$]{} Proof of Theorem \[theorem:convergence\_of\_value\_iteration\] {#appendix:sequential_placement_total_power} ============================================================== \[appendix:proof\_of\_value\_iteration\_convergence\] As we have seen in Section \[sec:mdp\_total\_power\], our problem is a negative dynamic programming problem (i.e., the $\mathsf{N}$ case of [@schal75conditions-optimality], where single-stage rewards are non-positive). It is to be noted that Schäl [@schal75conditions-optimality] discusses two other kind of problems as well: the $\mathsf{P}$ case (single-stage rewards are positive) and the $\mathsf{D}$ case (the reward at stage $k$ is discounted by a factor $\alpha^k$, where $0<\alpha<1$). In this appendix, we first state a general-purpose theorem for the value iteration (Theorem \[thm:value\_iteration\_general\]), prove it by some results of [@schal75conditions-optimality], and then we use this theorem to prove Theorem \[theorem:convergence\_of\_value\_iteration\]. A General Result (Derived from [@schal75conditions-optimality]) {#appendix_subsection_schal-discussion} --------------------------------------------------------------- Consider an infinite horizon total cost MDP whose state space $\mathcal{S}$ is an interval in $\mathbb{R}$ and the action space $\mathcal{A}$ is $[0,\infty)$. Let the set of possible actions at state $s$ be denoted by $\mathcal{A}(s)$. Let the single-stage cost be $c(s,a,w) \geq 0$ where $s$, $a$ and $w$ are the state, the action and the disturbance, respectively. Let us denote the optimal expected cost-to-go at state $s$ by $V^(s)$. Let the state of the system evolve as $s_{k+1}=h(s_k,a_k,w_k)$, where $s_k$, $a_k$ and $w_k$ are the state, the action and the disturbance at the $k$-th instant, respectively. Let $s^{} \in \mathcal{S}$ be an absorbing state with $c(s^{},a, w)=0$ for all $a$, $w$. Let us consider the value iteration for all $s \in \mathcal{S}$, with $V^{(0)}(\cdot)= 0 $: $$\begin{aligned} V^{(k+1)}(s)&=&\inf_{a \in [0,\infty)} \mathbb{E}_{w} \bigg( c(s,a,w)+V^{(k)}(h(s,a,w)) \bigg), s \neq s^{}\nonumber\\ V^{(k+1)}(s^{})&=& 0\label{eqn:value_iteration_general}\end{aligned}$$ We provide some results and concepts from [@schal75conditions-optimality], which will be used later to prove Theorem \[theorem:convergence\_of\_value\_iteration\]. \[thm:schal\_convergence\_value\_iteration\] \[[Theorem 4.2 ([@schal75conditions-optimality])]{}\] $V^{(k)}(s)\rightarrow V^{(\infty)}(s)$ for all $s \in \mathcal{S}$, i.e., the value iteration (\[eqn:value\_iteration\_general\]) converges. [$\Box$]{} Let us recall that $\Gamma_k(s)$ is the set of minimizers of (\[eqn:value\_iteration\_general\]) at the $k$-th iteration at state $s$, if the infimum is achieved at some $a<\infty$. $\Gamma_{\infty}(s):=\{a \in \mathcal{A}:a$ is an accumulation point of some sequence $\{a_k\}$ where each $a_k \in \Gamma_{k}(s)\}$. $\Gamma^(s)$ is the set of minimizers in the Bellman Equation. Let $\mathcal{C}(\mathcal{A})$ be the set of nonempty compact subsets of $\mathcal{A}$. The Hausdorff metric $d$ on $\mathcal{C}(\mathcal{A})$ is defined as follows: $$d(C_1,C_2)=\max \{ \sup_{c \in C_1}\rho(c,C_2), \, \sup_{c \in C_2}\rho(c,C_1) \}$$ where $\rho(c,C)$ is the minimum distance between the point $c$ and the compact set $C$. \[prop:separable\_hausdorff\] \[Proposition 9.1([@schal75conditions-optimality])\] $(\mathcal{C}(\mathcal{A}),d)$ is a separable metric space. A mapping $\phi:\mathcal{S}\rightarrow \mathcal{C}(\mathcal{A})$ is called measurable if it is measurable with respect to the Borel $\sigma$-algebra of $(\mathcal{C}(\mathcal{A}),d)$. $\hat{\mathcal{F}}(\mathcal{S} \times \mathcal{A})$ is the set of all measurable functions $v:\mathcal{S} \times \mathcal{A} \rightarrow \mathbb{R}$ which are bounded below and where every such $v(\cdot)$ is the limit of a non-decreasing sequence of measurable, bounded functions $v_k:\mathcal{S} \times \mathcal{A} \rightarrow \mathbb{R}$. We will next present a condition followed by a theorem. The condition, if satisfied, implies the convergence of value iteration (\[eqn:value\_iteration\_general\]) to the optimal value function (according to the theorem). \[condition\_A\] \[[Derived from Condition A in [@schal75conditions-optimality]]{}\] 1. $\mathcal{A}(s)\in \mathcal{C}(\mathcal{A})$ for all $s \in \mathcal{S}$ and $\mathcal{A}:\mathcal{S}\rightarrow \mathcal{C}(\mathcal{A})$ is measurable. 2. $\mathbb{E}_{w}(c(s,a,w)+V^{(k)}(h(s,a,w))) $ is in $\hat{\mathcal{F}}(\mathcal{S}\times \mathcal{A})$ for all $k \geq 0$.[$\Box$]{} \[thm:schal\_main\_theorem\] \[[Theorem $13.3$, [@schal75conditions-optimality]]{}\] If $c(s,a,w) \geq 0$ for all $s,a,w$ and Condition \[condition\_A\] holds: 1. $V^{(\infty)}(s)=V^(s)$, $s \in \mathcal{S}$. 2. $\Gamma_{\infty}(s) \subset \Gamma^(s)$. 3. There is a stationary optimal policy $f^{\infty}$ where $f:\mathcal{S} \rightarrow \mathcal{A}$ and $f(s) \in \Gamma_{\infty}(s)$ for all $s \in \mathcal{S}$.[$\Box$]{} The next condition and theorem deal with the situation where the action space is noncompact. \[condition\_B\] \[[Condition B ([@schal75conditions-optimality])]{}\] There is a measurable mapping $\underline{\mathcal{A}}:\mathcal{S} \rightarrow \mathcal{C}(\mathcal{A})$ such that: 1. $\underline{{\mathcal{A}}}(s)\subset {\mathcal{A}}(s)$ for all $s \in \mathcal{S}$. 2. for all $k \geq 0$. [$\Box$]{} This condition requires that for each state $s$, there is a compact set $\underline{\mathcal{A}}(s)$ of actions such that no optimizer of the value iteration lies outside the set $\underline{\mathcal{A}}(s)$ at any stage $k \geq 0$. \[thm:schal\_compact\_action\] \[[Theorem $17.1$, [@schal75conditions-optimality]]{}\] If Condition \[condition\_B\] is satisfied and if the three statements in Theorem \[thm:schal\_main\_theorem\] are valid for the modified problem having admissible set of actions $\underline{\mathcal{A}}(s)$ for each state $s \in \mathcal{S}$, then those statements are valid for the original problem as well.[$\Box$]{} Now we will provide an important theorem which will be used to prove Theorem \[theorem:convergence\_of\_value\_iteration\]. \[thm:value\_iteration\_general\] If the value iteration (\[eqn:value\_iteration\_general\]) satisfies the following conditions: 1. For each $k$, $\mathbb{E}_{w}\bigg(c(s,a,w)+V^{(k)}(h(s,a,w))\bigg)$ is jointly continuous in $a$ and $s$ for $s \neq s^{}$. 2. The infimum in (\[eqn:value\_iteration\_general\]) is achieved in $[0,\infty)$ for all $s \neq s^{}$. 3. For each $s \in \mathcal{S}$, there exists $a(s)<\infty$ such that $a(s)$ is continuous in $s$ for $s \neq s^{}$, and no minimizer of (\[eqn:value\_iteration\_general\]) lies in $(a(s), \infty)$ for each $k \geq 0$. Then the following hold: 1. The value iteration converges, i.e., $V^{(k)}(s) \rightarrow V^{(\infty)}(s)$ for all $s \neq s^{}$. 2. $V^{(\infty)}(s)=V^(s)$ for all $s \neq s^{}$. 3. $\Gamma_{\infty}(s) \subset \Gamma^(s)$ for all $s \neq s^{}$. 4. There is a stationary optimal policy $f^{\infty}$ where $f:\mathcal{S} \setminus \{s^{}\} \rightarrow \mathcal{A}$ and $f(s) \in \Gamma_{\infty}(s) \,\,\forall \,\, s \neq s^{}$.[$\Box$]{} Proof of Theorem \[thm:value\_iteration\_general\]:* By Theorem \[thm:schal\_convergence\_value\_iteration\], the value iteration converges, i.e., $V^{(k)}(s)\rightarrow V^{(\infty)}(s)$. Moreover, $V^{(k)}(s)$ is the optimal cost for a $k$-stage problem with zero terminal cost, and the cost at each stage is positive. Hence, $V^{(k)}(s)$ increases in $k$ for every $s \in \mathcal{S}$. Thus, for all $s \in \mathcal{S}$, $V^{(k)}(s) \uparrow V^{(\infty)}(s)$. Now, Condition \[condition\_B\] and Theorem \[thm:schal\_compact\_action\] say that if no optimizer of the value iteration in each stage $k$ lies outside a compact subset $\underline{\mathcal{A}}(s)$ of $\mathcal{A}(s) \subset \mathcal{A}$, then we can deal with the modified problem having a new action space $\underline{\mathcal{A}}(s)$. If the value iteration converges to the optimal value in this modified problem, then it will converge to the optimal value in the original problem as well, provided that the mapping $\underline{\mathcal{A}}:\mathcal{S}\rightarrow \mathcal{C}(\mathcal{A})$ is measurable. Let us choose $\underline{\mathcal{A}}(s):=[0,a(s)]$, where $a(s)$ satisfies hypothesis (c) of Theorem \[thm:value\_iteration\_general\]. Since $a(s)$ is continuous at $s \neq s^{}$, for any $\epsilon>0$ we can find a $\delta_{s,\epsilon}>0$ such that $\|a(s)-a(s')\|<\epsilon$ whenever $\|s-s'\|<\delta_{s,\epsilon}$, $s \neq s^{}$, $s' \neq s^{}$. Now, when $\|a(s)-a(s')\|<\epsilon$, we have $d([0,a(s)],[0,a(s')])<\epsilon$. Hence, the mapping $\underline{\mathcal{A}}:\mathcal{S}\rightarrow \mathcal{C}(\mathcal{A})$ is continuous at all $s \neq s^{}$, and thereby measurable in this case. Hence, the value iteration (\[eqn:value\_iteration\_general\]) satisfies Condition \[condition\_B\]. Thus, the value iteration for $s \neq s^{}$ can be modified as: $$V^{(k+1)}(s)=\inf_{a \in [0,a(s)]} \mathbb{E}_{w} ( c(s,a,w)+V^{(k)}(h(s,a,w)) ) \label{eqn:modified_value_iteration_general}$$ Now, $\mathbb{E}_{w} ( c(s,a,w)+V^{(k)}(h(s,a,w)) )$ is continuous (can be discontinuous at $s=s^{}$, since this quantity is $0$ at $s=s^$) on $\mathcal{S} \times \mathcal{A}$ (by our hypothesis). Hence, $\mathbb{E}_{w} ( c(s,a,w)+V^{(k)}(h(s,a,w)) )$ is measurable on $\mathcal{S} \times \mathcal{A}$. Also, it is bounded below by $0$. Hence, it can be approximated by an increasing sequence of bounded measurable functions $\{v_{n,k}\}_{n \geq 1}$ given by $v_{n,k}(s,a)=\min \{\mathbb{E}_{w} ( c(s,a,w)+V^{(k)}(h(s,a,w)) ),\,n \}$. Hence, $\mathbb{E}_{w} ( c(s,a,w)+V^{(k)}(h(s,a,w)) )$ is in $\hat{\mathcal{F}}(\mathcal{S} \times \mathcal{A})$. Thus, Condition \[condition\_A\] is satisfied for the modified problem and therefore, by Theorem \[thm:schal\_main\_theorem\], the modified value iteration in (\[eqn:modified\_value\_iteration\_general\]) converges to the optimal value function. Now, by Theorem \[thm:schal\_compact\_action\], we can argue that the value iteration (\[eqn:value\_iteration\_general\]) converges to the optimal value function in the original problem and hence $V^{(\infty)}(s)=V^(s)$ for all $s \in \mathcal{S} \setminus s^$. Also, $\Gamma_{\infty}(s) \subset \Gamma^(s)$ for all $s \in \mathcal{S} \setminus s^$ and there exists a stationary optimal policy $f^{\infty}$ where $f(s) \in \Gamma_{\infty}(s)$ for all $s \in \mathcal{S} \setminus s^$ (by Theorem \[thm:schal\_main\_theorem\]). Proof of Theorem \[theorem:convergence\_of\_value\_iteration\] {#appendix_subsection-convergence-value-iteration-proof} -------------------------------------------------------------- This proof uses the results of Theorem \[thm:value\_iteration\_general\] provided in this appendix. Remember that the state $\mathbf{EOL}$ is absorbing and $c(\mathbf{EOL}, a, w)=0$ for all $a$, $w$. We can think of it as state $0$ so that our state space becomes $[0,1]$ which is a Borel set. We will see that the state $0$ plays the role of the state $s^$ as mentioned in Theorem \[thm:value\_iteration\_general\]. We need to check whether the conditions (a), (b), and (c) in Theorem \[thm:value\_iteration\_general\] are satisfied for the value iteration (\[eqn:value\_iteration\]). Of course, $J_{\xi}^{(0)}(s)=0$ is concave, increasing in $s \in (0,1]$. Suppose that $J_{\xi}^{(k)}(s)$ is concave, increasing in $s$ for some $k \geq 0$. Also, for any fixed $a \geq 0$, $\frac{se^{\rho a}}{1+se^{\rho a}×}$ is concave and increasing in $s$. Thus, by the composition rule for the composition of a concave increasing function $J_{\xi}^{(k)}(\cdot)$ and a concave increasing function $\frac{se^{\rho a}}{1+se^{\rho a}×}$, for any $a \geq 0$ the term $J_{\xi}^{(k)}\left(\frac{se^{\rho a}}{1+se^{\rho a}×}\right)$ is concave, increasing over $s \in (0,1]$. Hence, $\int_{0}^{a}\beta e^{-\beta z}s(e^{\rho z}-1)dz + e^{-\beta a}\bigg(s(e^{\rho a}-1)+\xi+J_{\xi}^{(k)}\left(\frac{se^{\rho a}}{1+se^{\rho a}×}\right)\bigg)$ (in (\[eqn:value\_iteration\])) is concave increasing over $s \in (0,1]$. Since the infimization over $a$ preserves concavity, we conclude that $J_{\xi}^{(k+1)}(s)$ is concave, increasing over $s \in (0,1]$. Hence, for each $k$, $J_{\xi}^{(k)}(s)$ is continuous in $s$ over $(0,1)$, since otherwise concavity w.r.t. $s$ will be violated. Now, we must have $J_{\xi}^{(k)}(1) \leq \lim_{s \uparrow 1} J_{\xi}^{(k)}(s)$, since otherwise the concavity of $J_{\xi}^{(k)}(s)$ will be violated. But since $J_{\xi}^{(k)}(s)$ is increasing in $s$, $J_{\xi}^{(k)}(1) \geq \lim_{s \uparrow 1} J_{\xi}^{(k)}(s)$. Hence, $J_{\xi}^{(k)}(1) = \lim_{s \uparrow 1} J_{\xi}^{(k)}(s)$. Thus, $J_{\xi}^{(k)}(s)$ is continuous in $s$ over $(0,1]$ for each $k$. Hence, $\int_{0}^{a}\beta e^{-\beta z}s(e^{\rho z}-1)dz + e^{-\beta a}(s(e^{\rho a}-1)+\xi+J_{\xi}^{(k)}(\frac{se^{\rho a}}{1+se^{\rho a}×}) )$ is continuous in $s,a$ for $s \neq 0$. Hence, condition (a) in Theorem \[thm:value\_iteration\_general\] is satisfied. Now, we will check condition (c) in Theorem \[thm:value\_iteration\_general\]. By Theorem \[thm:schal\_convergence\_value\_iteration\], the value iteration converges, i.e., $J_{\xi}^{(k)}(s)\rightarrow J_{\xi}^{(\infty)}(s)$. Also, $J_{\xi}^{(\infty)}(s)$ is concave, increasing in $s \in (0,1]$ and hence continuous. Moreover, $J_{\xi}^{(k)}(s)$ is the optimal cost for a $k$-stage problem with zero terminal cost, and the cost at each stage is positive. Hence, $J_{\xi}^{(k)}(s)$ increases in $k$ for every $s \in (0,1]$. Thus, for all $s \in (0,1]$, $J_{\xi}^{(k)}(s) \uparrow J_{\xi}^{(\infty)}(s)$. Again, $J_{\xi}^{(k)}(s)$ is the optimal cost for a $k$-stage problem with zero terminal cost. Hence, it is less than or equal to the optimal cost for the infinite horizon problem with the same transition law and cost structure. Hence, $J_{\xi}^{(k)}(s)\leq J_{\xi}(s)$ for all $k \geq 1$. Since $J_{\xi}^{(k)}(s)\uparrow J_{\xi}^{(\infty)}(s)$, we have $J_{\xi}^{(\infty)}(s) \leq J_{\xi}(s)$. Now, consider the following two cases: ### $\beta>\rho$ Let us define a function $\psi:(0,1]\rightarrow \mathbb{R}$ by $\psi(s)=\frac{J_{\xi}^{(\infty)}(s)+\theta s}{2×}$. By Proposition \[prop:upper\_bound\_on\_cost\_beta\_geq\_rho\], $J_{\xi}(s)<\theta s$ for all $s \in (0,1]$. Hence, $J_{\xi}^{(\infty)}(s)<\psi(s)<\theta s$ and $\psi(s)$ is continuous over $s \in (0,1]$. Since $\beta>\rho$ and $J_{\xi}^{(k)}(s) \in [0, \theta]$ for any $s$ in $(0,1]$, the expression $\theta s + e^{-\beta a}\bigg(-\theta s e^{\rho a}+\xi+J_{\xi}^{(k)}\left(\frac{se^{\rho a}}{1+se^{\rho a}×}\right)\bigg)$ obtained from the R.H.S of (\[eqn:value\_iteration\]) converges to $\theta s$ as $a \rightarrow \infty$. A lower bound to this expression is $\theta s + e^{-\beta a}(-\theta s e^{\rho a})$. With $\beta > \rho$, for each $s$, there exists $a(s)<\infty$ such that $\theta s + e^{-\beta a}(-\theta s e^{\rho a})> \psi(s)$ for all $a > a(s)$. But $\theta s + \inf_{a \geq 0} e^{-\beta a}\bigg(-\theta s e^{\rho a}+\xi+J_{\xi}^{(k)}\left(\frac{se^{\rho a}}{1+se^{\rho a}×}\right)\bigg)$ is equal to $J_{\xi}^{(k+1)}(s)<\psi(s)$. Hence, for any $s \in (0,1]$, the minimizers for (\[eqn:bellman\_equation\_simplified\_in\_a\]) always lie in the compact interval $[0,a(s)]$ for all $k \geq 1$. Since $\psi(s)$ is continuous in $s$, we can choose $a(s)$ as a continuous function of $s$ on $(0,1]$. ### $\beta \leq \rho$ Fix $A$, $0 <A <\infty$. Let $K:=\frac{1}{\beta A ×}\left(\xi+(e^{\rho A}-1) \right)+(e^{\rho A}-1)$. Then, by Proposition \[prop:upper\_bound\_on\_cost\], $J_{\xi}(s) \leq K$ for all $s \in (0,1]$. Now, we observe that the objective function (for minimization over $a$) in the R.H.S of (\[eqn:value\_iteration\]) is lower bounded by $\int_{0}^{a}\beta e^{-\beta z}s(e^{\rho z}-1)dz$, which is continuous in $s,a$ and goes to $\infty$ as $a \rightarrow \infty$ for each $s \in (0,1]$. Hence, for each $s \in (0,1]$, there exists $0<a(s)<\infty$ such that $\int_{0}^{a}\beta s e^{-\beta z}(e^{\rho z}-1)dz >2K$ for all $a>a(s)$ and $a(s)$ is continuous over $s \in (0,1]$. But $J_{\xi}^{(k+1)}(s) \leq J_{\xi}(s) \leq K$ for all $k$. Hence, the minimizers in (\[eqn:value\_iteration\]) always lie in $[0,a(s)]$ where $a(s)$ is independent of $k$ and continuous over $s \in (0,1]$. Let us set $a(0)=a(1)$.[^11] Then, the chosen function $a(s)$ is continuous over $s \in (0,1]$ and can be discontinuous only at $s=0$. Thus, condition (c) of Theorem \[thm:value\_iteration\_general\] has been verified for the value iteration (\[eqn:value\_iteration\]). Condition (b) of Theorem \[thm:value\_iteration\_general\] is obviously satisfied since a continuous function over a compact set always has a minimizer. [$\Box$]{} [Remark:]{} Observe that in our value iteration (\[eqn:value\_iteration\]) it is always sufficient to deal with compact action spaces, and the objective functions to be minimized at each stage of the value iteration are continuous in $s$, $a$. Hence, $\Gamma_{k}(s)$ is nonempty for each $s \in (0,1]$, $k \geq 0$. Also, since there exists $K>0$ such that $J_{\xi}^{(k)}(s) \leq K$ for all $k \geq 0$, $s \in (0,1]$, it is sufficient to restrict the action space in (\[eqn:value\_iteration\]) to a set $[0, a(s)]$ for any $s \in (0,1]$, $k \geq 0$. Hence, $\Gamma_k (s) \subset [0,a(s)]$ for all $s \in (0,1]$, $k \geq 0$. Now, for a fixed $s \in (0,1]$, any sequence $\{a_k\}_{k \geq 0}$ with $a_k \in \Gamma_k (s)$, in bounded. Hence, the sequence must have a limit point. Hence, $\Gamma_{\infty}(s)$ is nonempty for each $s \in (0,1]$. Since $\Gamma_{\infty}(s) \subset \Gamma^(s)$, $\Gamma^*(s)$ is nonempty for each $s \in (0,1]$. Proofs of Propositions \[prop:increasing\_concave\_in\_s\], \[prop:increasing\_concave\_in\_lambda\] and \[prop:continuity\_of\_cost\] {#appendix_subsection_policy-structure} ====================================================================================================================================== \[appendix:proof\_of\_propositions\] Proof of Proposition \[prop:increasing\_concave\_in\_s\] -------------------------------------------------------- Fix $\xi$. Consider the value iteration (\[eqn:value\_iteration\]). Let us start with $J_{\xi}^{(0)}(s):=0$ for all $s \in (0,1]$. Clearly, $J_{\xi}^{(1)}(s)$ is concave and increasing in $s$, since pointwise infimum of linear functions is concave. Now let us assume that $J_{\xi}^{(k)}(s)$ is concave and increasing in $s$. Then, by the composition rule, it is easy to show that $J_{\xi}^{(k)}(\frac{se^{\rho a}}{1+se^{\rho a}×})$ is concave and increasing in $s$ for any fixed $a\geq 0$. Hence, $J_{\xi}^{(k+1)}(s)$ is concave and increasing, since pointwise infimum of a set of concave and increasing functions is concave and increasing. By Theorem \[theorem:convergence\_of\_value\_iteration\], $J_{\xi}^{(k)}(s)\rightarrow J_{\xi}(s)$. Hence, $J_{\xi}(s)$ is concave and increasing in $s$.[$\Box$]{} Proof of Proposition \[prop:increasing\_concave\_in\_lambda\] ------------------------------------------------------------- Consider the value iteration (\[eqn:value\_iteration\]). Since $J_{\xi}^{(0)}(s):=0$ for all $s \in (0,1]$, $J_{\xi}^{(1)}(s)$ is obtained by taking infimum (over $a$) of a linear, increasing function of $\xi$. Hence, $J_{\xi}^{(1)}(s)$ is concave, increasing over $\xi \in (0, \infty)$. If we assume that $J_{\xi}^{(k)}(s)$ is concave and increasing in $\xi$, then $J_{\xi}^{(k)}(\frac{se^{\rho a}}{1+se^{\rho a}×})$ is also concave and increasing in $\xi$ for fixed $s$ and $a$. Thus, $J_{\xi}^{(k+1)}(s)$ is also concave and increasing in $\xi$. Now, $J_{\xi}^{(k)}(s) \rightarrow J_{\xi}(s)$ for all $s \in \mathcal{S}$, and $J_{\xi}^{(k)}(s)$ is concave, increasing in $\xi$ for all $k \geq 0$, $s \in \mathcal{S}$. Hence, $J_{\xi}(s)$ is concave and increasing in $\xi$. [$\Box$]{} Proof of Proposition \[prop:continuity\_of\_cost\] -------------------------------------------------- Clearly, $J_{\xi}(s)$ is continuous in $s$ over $(0,1)$, since otherwise concavity w.r.t. $s$ will be violated. Now, since $J_{\xi}(s)$ is concave in $s$ over $(0,1]$, we must have $J_{\xi}(1) \leq \lim_{s \uparrow 1}J_{\xi}(s)$. But since $J_{\xi}(s)$ is increasing in $s$, $J_{\xi}(1) \geq \lim_{s \uparrow 1}J_{\xi}(s)$. Hence, $J_{\xi}(1) = \lim_{s \uparrow 1}J_{\xi}(s)$. Thus, $J_{\xi}(s)$ is continuous in $s$ over $(0,1]$. Again, for a fixed $s \in (0,1]$, $J_{\xi}(s)$ is concave and increasing in $\xi$. Hence, $J_{\xi}(s)$ is continuous in $\xi$ over $\xi \in (0,c),\, \forall \, c>0$. Hence, $J_{\xi}(s)$ is continuous in $\xi$ over $(0,\infty)$. [$\Box$]{} [^1]: This work was supported by the Department of Science and Technology (DST), India, through the J.C. Bose Fellowship, by an Indo-Brazil cooperative project on “WIreless Networks and techniques with applications to SOcial Needs (WINSON)," and by a project funded by the Department of Electronics and Information Technology, India, and NSF, USA, titled “Wireless Sensor Networks for Protecting Wildlife and Humans in Forests.” [^2]: This paper is an extension of [@chattopadhyay-etal12optimal-capacity-relay-placement-line], and is also available in [@chattopadhyay-etal12optimal-capacity-relay-placement-line-arxiv-17April2013]. [^3]: Arpan Chattopadhyay and Anurag Kumar are with the Electrical Communication Engineering (ECE) Department, Indian Institute of Science (IISc), Bangalore-560012, India (e-mail: [email protected], [email protected]). Abhishek Sinha is with the Laboratory for Information and Decision Systems (LIDS), Massachusetts Institute of Technology, Cambridge, MA 02139 (e-mail: [email protected]). Marceau Coupechoux is with Telecom ParisTech and CNRS LTCI, Dept. Informatique et Réseaux, 23, avenue d’Italie, 75013 Paris, France (e-mail: [email protected]). This work was done during the period when he was a Visiting Scientist in the ECE Deparment, IISc. [^4]: Full-duplex radios are becoming practical; see [@khandani13two-way-full-duplex-wireless], [@khandani10spatial-multiplexing-two-way-channel], [@choi-etal10single-channel-full-duplex], [@jain-etal11real-time-full-duplex]. [^5]: $\log (\cdot)$ in this paper will mean the natural logarithm unless the base is specified. [^6]: A motivation for the use of the exponential distribution, given the prior knowledge of the mean length $\overline{L}$, is that it is the maximum entropy continuous probability density function with the given mean. By using the exponential distribution, we are leaving the length of the line as uncertain as we can, given the prior knowledge of its mean. [^7]: Recall Section \[subsec:sum\_power\_constraint\]. The battery depletion rate $\frac{P_{rcv}}{E}$ of a node due to the receive power alone can be absorbed into the relay cost $\xi$. [^8]: The constraint on the mean number of relays can be justified if we consider the relay deployment problem for multiple source-sink pairs over several lines of mean length $\overline{L}$, given a large pool of relays, and we are only interested in keeping small the total number of relays over all these deployments. [^9]: $\frac{e^{\Lambda \tilde{z}_k}}{\sum_{i=0}^{k}e^{\Lambda \tilde{z}_i}} \geq \frac{e^{\Lambda \tilde{z}_k}}{(k+1)e^{\Lambda \tilde{z}_k}}=\frac{1}{k+1}$; hence, if $s_k$ is small, $k$ must be large enough. [^10]: Marano and Franceschetti ([@marano-franceschetti05ray-propagation-random-lattice]) modeled a city as a random lattice, and the distance from the transmitter to the receiver is measured along the edges of the lattice instead of the Euclidean distance. Hence, this result renders the analysis in our paper valid even for deployment along the streets of a city with turns; deployment algorithm in that case will only consider the distances along the streets and not on the actual Euclidean distances. [^11]: Remember that at state $0$ (i.e., state $\mathbf{EOL}$), the single stage cost is $0$ irrespective of the action, and that this state is absorbing. Hence, any action at state $0$ can be optimal.	{ "pile_set_name": "ArXiv" }
--- abstract: 'We show that the Continuum Hypothesis is consistent with all regular spaces of hereditarily countable $\pi$-character being $C$-closed. This gives us a model of [ZFC]{} in which the Continuum Hypothesis holds and compact Hausdorff spaces of countable tightness are sequential.' address: - \| Department of Mathematics\ University of North Carolina, Charlotte\ Charlotte, NC 28223 - \| Department of Mathematics\ Ohio University\ Athens, OH 45701 author: - Alan Dow - Todd Eisworth title: '[C]{}[H]{} and the Moore-Mrowka Problem' --- Introduction ============ Our goal in this paper is to prove that the Continuum Hypothesis is consistent with the following statement: [$\circledast$]{} Regular spaces that are hereditarily of countable $\pi$–character are $C$-closed. The principle $\circledast$ may seem technical (and mysterious given the lack of definitions), but it turns out to be of interest for a couple of reasons: - It is unknown if $\circledast$ is a consequence of the Proper Forcing Axiom or even Martin’s Maximum, and our consistency proof seems to need that the Continuum Hypothesis holds in the final model. [.05in]{} - More importantly, in the presence of $2^{\aleph_0}<2^{\aleph_1}$, $\circledast$ is strong enough to imply that compact spaces of countable tightness are sequential, and so our model provides the final piece to the solution of the well-known “Moore-Mrowka + [CH]{} problem” in set-theoretic topology. [.05in]{} The Moore-Mrowka problem was long considered to be one of the major problems in set-theoretic topology. The reader may be interested in a brief review. The problem, asking if every countably tight compact space was sequential, was raised by Moore and Mrowka in 1964. In the mid-seventies two of the most famous examples in topology, Ostaszewski’s S-space [@Ost1976] and Fedorchuk’s S-space [@Fed1976], established that $\diamondsuit$ implied there were counterexamples. The Moore-Mrowka statement was proven to be consistent by Balogh [@Bal1989] by showing that it was a consequence of the proper forcing axiom. Nevertheless, since $\diamondsuit$ implies [CH]{}, Arhangelskii asked [@Arh1978] (Problem 26) if the Continuum was sufficient to produce a counterexample. This problem was again raised by Shakhmatov [@Shak1992](2.13) in the influential Recent Progress in General Topology. Before moving on, we provide the reader with some definitions, looking first at some important cardinal functions for topological spaces. Let $X$ be a topological space. 1. $z\in X$ is a point of countable tightness in $X$ ($t(z, X)=\aleph_0)$ if whenever $A\subseteq X$ and $z\in\operatorname{cl}_X(A)$, there is a countable $A_0\subseteq A$ such that $z\in\operatorname{cl}_X(A_0)$. [.05in]{} 2. $X$ is countably tight ($t(X)=\aleph_0$) if $t(z, X)=\aleph_0$ for every $z\in X$. [.05in]{} 3. $X$ has countable $\pi$-character ($\pi\chi(X)=\aleph_0$) if for any point $x\in X$ there is a countable collection $\{U_n:n\in\omega\}$ of non-empty open sets such that for any open neighborhood $U$ of $x$ there is an $n$ with $U_n\subseteq U$. Note that we are not requiring that $x$ is a member of $U_n$, so this is a weakening of first countability. [.05in]{} 4. $X$ is hereditarily of countable $\pi$-character ($h\pi\chi(X)=\aleph_0$) if $\pi\chi(Y)=\aleph_0$ whenever $Y\subseteq X$. [.05in]{} An elementary argument shows that $h\pi\chi(X)=\aleph_0\Longrightarrow t(X)=\aleph_0$, and the reverse implication is true if $X$ is compact (Hausdorff) by a deep theorem of Sapirovskii [@sapirovskii]. Let $X$ be a topological space. 1. $X$ is countably compact if every infinite subset has a point of accumulation, or equivalently, if every countable open cover of $X$ has a finite subcover. [.05in]{} 2. $X$ is $C$-closed if every countably compact subset of $X$ is closed in $X$. [.05in]{} 3. A subset $Y$ of $X$ is sequentially closed if whenever $\langle x_n:n<\omega\rangle$ is a convergent sequence of points from $Y$, the limit of the sequence is also in $Y$, that is, $Y$ is closed under the operation of taking limits of convergent sequences. [.05in]{} 4. $X$ is sequential if every sequentially closed subset of $X$ is closed. [.05in]{} Given the above definitions, we can now illustrate the power of our principle $\circledast$: \[chmm\] 1. Assume $2^{\aleph_0}<2^{\aleph_1} + \circledast$. Then compact spaces of countable tightness are sequential. [.05in]{} 2. $\circledast$ implies that if $X$ is a countably compact regular space satisfying $h\pi\chi(X)=\aleph_0$, then every non-isolated point in $X$ is the limit of a non-trivial convergent sequence. First, suppose $X$ is compact and of countable tightness. By Sapirovskii’s Theorem [@sapirovskii], we know $h\pi\chi(X)=\aleph_0$, and so by $\circledast$ we know that $X$ is $C$-closed. By a result of Ismail and Nyikos [@ismailnyikos], $2^{\aleph_0}<2^{\aleph_1}$ implies that a compact Hausdorff space is $C$-closed if and only if it sequential, and we are done. For the second, suppose $X$ is as described, and $x$ is a non-isolated point. Since $X$ is $C$-closed, we know that $X\setminus\{x\}$ is not countably compact. This means we can find a (countably infinite) set $$\{x_n:n\in\omega\}\subseteq X\setminus\{x\}$$ that is closed and discrete in $X\setminus\{x\}$. Since $X$ is countably compact, it follows that $\{x_n:n\in\omega\}$ must converge to $x$ in $X$. By (1) above, any model of [CH]{} + $\circledast$ gives us a model of [CH]{} in which compact spaces of countable tightness are sequential. At this point, we do not know if $\circledast$ can hold in a model where [CH]{} fails. The conclusion of (2) says that $X$ is [$C$-sequential]{}, a notion introduced and studied by Ran[č]{}in [@rancin]. Hajnal and Juh[á]{}sz show that [CH]{} implies the existence of a countably compact regular space of countable tightness with no non-trivial convergent sequences at all, so the use of hereditary $\pi$-character rather than tightness is critical. Bad triples =========== We call $\vec{X}=\langle X, Y, z\rangle$ a [relevant triple]{} if 1. $X$ and $Y$ are regular separable topological spaces [.05in]{} 2. the underlying set of $Y$ is $\omega_1$ [.05in]{} 3. $Y= X\cup\{z\}$ [.05in]{} 4. $X$ is countably compact, and [.05in]{} 5. $z$ is not isolated in $Y$. [.05in]{} A relevant triple $\langle X, Y, z\rangle$ is [bad]{} (and called a [bad triple]{}) if in addition 5. $h\pi\chi(X)=\aleph_0$, and [.05in]{} 6. $t(z, Y)=\aleph_0$. Note that the separability of $Y$ follows from the separability of $X$, so this assumption is superfluous. Also, the requirement that the underlying set of $Y$ is $\omega_1$ is included for technical reasons: if we require only that $\|Y\|=\aleph_1$, then clearly $Y$ is homeomorphic to some space with underlying set $\omega_1$. Bad triples are relevant to our project because of the following fact: \[prop2\] Assume [CH]{} holds. If $\circledast$ fails, then there is a bad triple. Assume $Z$ is a counterexample to $\circledast$, so $Z$ is regular, hereditarily of countable $\pi$-character, but not $C$-closed. Let $W$ be a countably compact non-closed subset of $Z$,and choose $z\in\operatorname{cl}_Z(W)\setminus W$. Since $h\pi\chi(Z)=\aleph_0$, we know $Z$ has countable tightness, and so there is a countable $W_0\subseteq W$ with $z\in\operatorname{cl}_Z(W_0)$. Since $\operatorname{cl}_Z(W_0)$ is also countably compact and not closed, we can without loss of generality assume that $W$ is separable. A separable regular space has weight at most $2^{\aleph_0}$. Since we are assuming [CH]{} that means $w(W)$ is at most $\aleph_1$. Certainly $w(W)$ is uncountable, as regular spaces with countable bases are metrizable and countably compact metrizable spaces are compact. Thus, $W$ has weight exactly $\aleph_1$. Since the Continuum Hypothesis holds, we can find $M$ such that - $M$ is an elementary submodel of $H(\chi)$ for some sufficiently large regular $\chi$, [.05in]{} - $M$ has cardinality $\aleph_1$, [.05in]{} - $M$ is closed under $\omega$-sequences, and [.05in]{} - $Z$, $W$, and $z$ are all elements of $M$. Now let $X= M\cap W$ (topologized as a subspace of $W$) and $Y= X\cup\{z\}$. We claim that $\langle X, Y, z\rangle$ is a bad triple. Elementarity arguments tell us that $X$ and $Y$ are regular, $X$ is hereditarily of countable $\pi$-character, $z$ is not isolated in $Y$, and $t(z, Y)=\aleph_0$. Note that $X$ (and hence $Y$) is separable, as $M$ will contain every member of a (countable) dense subset of $W$. The only remaining issue of substance is whether $X$ is countably compact, but this follows easily using the fact that $M$ is closed under $\omega$-sequences: any countably infinite subset $A$ of $X$ is a member of $M$, and by elementary $M$ will contain a limit point of $A$ (as $M$ knows that $W$ is countably compact) and therefore $A$ has a limit point in $X$. Adjoining a filter ================== This section is critical as it contains the main new idea necessary to produce our desired model. Suppose $\vec{X}=\langle X, Y, z\rangle$ is a bad triple. We define $$\mathbb{P}[\vec{X}]=\{A\subseteq X: \text{ $A$ is a separable closed subset of $X$ with } z\in\operatorname{cl}_Y(A)\}$$ and order $\mathbb{P}[\vec{X}]$ by setting $$A\leq B\Longleftrightarrow A\subseteq B.$$ if $A$ is a closed subset of $X$ and $z\in\operatorname{cl}_Y(A)$, then there is a $B\in\mathbb{P}[\vec{X}]$ such that $B\subseteq A$. With $\vec{X}=\langle X, Y, z\rangle$ and $\mathbb{P}=\mathbb{P}_{\vec{X}}$ as above, we have: 1. If $A$ is a closed subset of $X$ with $z\in\operatorname{cl}_Y(A)$, then there is a $B\in\mathbb{P}[\vec{X}]$ with $B\subseteq A$. [.05in]{} 2. $\mathbb{P}$ is a countably closed notion of forcing, and [.05in]{} 3. for each open neighborhood $U$ of $z$ in $Y$, $D^U:=\{A\in \mathbb{P}: A\subseteq U\}$ is dense in $\mathbb{P}$. The first statement is trivial as $t(z, Y)=\aleph_0$. For the second, suppose $\langle A_n:n<\omega\rangle$ is a decreasing sequence of conditions in $\mathbb{P}$, and let $A:=\bigcap\{A_n:n<\omega\}$. The set $A$ is clearly a closed subset of $X$, so in light of (1) it suffices to show $z\in\operatorname{cl}_Y(A)$. Suppose this fails, and let $U$ be an open neighborhood of $z$ in $Y$ for which $$\label{eqn3} A\cap U=\emptyset.$$ Since $Y$ is regular, we can find an open neighborhood $V$ of $z$ with $\operatorname{cl}_Y(V)\subseteq U$. Now let $W=\operatorname{cl}_Y(V)\cap X$. Since $W$ is a closed subset of $X$ it is countably compact. Furthermore, for each $n$ we know $A_n\cap W\neq\emptyset$ because $A_n$ meets $V$. Thus, $\langle A_n\cap V:n<\omega\rangle$ is a decreasing sequence of non-empty closed subsets of the countably compact space $W$. We conclude that $$\bigcap\{A_n\cap V:n<\omega\}=A\cap V\neq\emptyset,$$ which contradicts (\[eqn3\]). For (3), given $U$ and a condition $A\in\mathbb{P}$, we can choose an open neighborhood $V$ of $z$ with $\operatorname{cl}_Y(V)\subseteq U$, and the set $A\cap\operatorname{cl}_Y(V)$ contains a condition in $\mathbb{P}[\vec{X}]$ by (1). We will be using forcings like $\mathbb{P}[\vec{X}]$ in models where [CH]{} holds, and the next proposition shows why this cardinal arithmetic assumption is useful: \[usefulprop\] Assume the Continuum Hypothesis, and suppose $\vec{X}=\langle X, Y, z\rangle$ and $\mathbb{P}=\mathbb{P}[\vec{X}]$ are as above. Let $G$ be a generic subset of $\mathbb{P}$. In the extension $V[G]$, if we define $$\mathcal{F}:=\{A\subseteq X: \text{$A$ closed and $A\supseteq B$ for some $B\in G$ }\}$$ then 1. $\mathcal{F}$ is a maximal free filter of closed subsets of $X$, and [.05in]{} 2. for any open neighborhood $U$ of $z$ in $Y$, there is an $A\in\mathcal{F}$ with $A\subseteq U$. As far as (1) goes, the only issue is whether $\mathcal{F}$ is maximal in $V[G]$. It suffices to prove that whenever we have a condition $p\in\mathbb{P}$ and a $\mathbb{P}$-name $\dot A$ for which $$\label{eqn4} p{\Vdash}\dot A\text{ is a closed subset of $X$ that meets every member of $\dot G_{\mathbb{P}}$},$$ we can find an extension $q$ of $p$ in $\mathbb{P}$ such that $$q{\Vdash}q\subseteq \dot A.$$ Since we have assumed [CH]{}, we know that $Y$ has weight $\aleph_1$. The point $z$ cannot have a countable neighborhood base in $Y$ (as otherwise $X$ would not be countably compact), so it follows that $z$ has character $\aleph_1$ in $Y$. Let $\{U_\alpha:\alpha<\omega_1\}$ enumerate a neighborhood base for $z$ in $Y$. Note that (\[eqn4\]) shows us that $r{\Vdash}\dot A\cap r\neq\emptyset$ whenever $r\leq p$ in $\mathbb{P}$, and so we can construct a sequence $\langle p_\alpha:\alpha<\omega_1\rangle$ of conditions in $\mathbb{P}$ and sequence $\langle x_\alpha:\alpha<\omega_1\rangle$ of points in $X$ such that - $p_0=p$ [.05in]{} - $p_\alpha=\bigcap\{p_\beta:\beta<\alpha\}$ for $\alpha$ limit [.05in]{} - $p_{\alpha+1}\leq p_\alpha$ [.05in]{} - $p_{\alpha+1}{\Vdash}x_\alpha\in p_\alpha\cap U_\alpha\cap\dot A$. [.05in]{} Since $t(z, Y)=\aleph_0$, we can choose $\alpha<\omega_1$ least for which $$z\in\operatorname{cl}_Y\{x_\beta:\beta<\alpha\}.$$ Note that this implies $\alpha$ is a limit ordinal, and $$\beta<\alpha\Longrightarrow z\in\operatorname{cl}_Y\{x_\gamma:\beta<\gamma<\alpha\}.$$ Now define $$r:=\operatorname{cl}_X\{x_\beta:\beta<\alpha\},$$ clearly $r$ is a condition in $\mathbb{P}$. Moreover, given $\beta<\alpha$ we know $$\{x_\gamma:\beta<\gamma<\alpha\}\subseteq r\cap p_\beta,$$ and hence $r\cap p_\beta$ is a condition in $\mathbb{P}$ for each $\beta<\alpha$. Let us define $$q:=\bigcap_{\beta<\alpha}r\cap p_\beta.$$ We know that $q\in\mathbb{P}$ (see the proof that $\mathbb{P}$ is countably closed) and $q$ extends $p=p_0$ in $\mathbb{P}$. We have $q\leq p_{\beta+1}$ for each $\beta<\alpha$, and this means $$q{\Vdash}\{x_\beta:\beta<\alpha\}\subseteq\dot A.$$ By definition of $r$, we have $$q{\Vdash}r\subseteq\dot A$$ and since $q\subseteq r$, we conclude $$q{\Vdash}q\subseteq \dot A$$ as required. The following theorem summarizes our efforts in this section: \[summary1\] Suppose [CH]{} holds and $\vec{X} = \langle X, Y, z\rangle$ is a bad triple. Then 1. the notion of forcing $\mathbb{P}[\vec{X}]$ is countably closed and of cardinality $\aleph_1$, and [.05in]{} 2. if $G$ is a generic subset of $\mathbb{P}[\vec{X}]$, then in $V[G]$ there is an object $\mathcal{F}$ such that 1. $\mathcal{F}$ is a maximal free filter of of closed subsets of $X$, [.05in]{} 2. $\mathcal{F}$ has a base of separable sets, and [.05in]{} 3. for any open neighborhood $U$ of $z$ in $Y$, there is an $A\in\mathcal{F}$ with $A\subseteq U$. [.05in]{} Destroying a counterexample =========================== In this section we examine the problem of “killing” a given bad triple. We show that bad triples can be destroyed by a reasonable notion of forcing, as long as we have the Continuum Hypothesis available. Our plan is as follows: given a bad triple $\vec{X}=\langle X, Y, z\rangle$, we will first force with $\mathbb{P}[\vec{X}]$, and ask if $X$ is still hereditarily of countable $\pi$-character. If the answer is “no”, then $\vec{X}$ is no longer a bad triple and we are done. If the answer is “yes”, then we will do an additional forcing that achieves $t(z, Y)>\aleph_0$, and thus destroy the badness of $\vec{X}$ via a different route. Most of this section will concentrate on the second notion of forcing mentioned above. This forcing was isolated in previous work [@EI17], and we will review briefly its definition and salient properties. \[41\] We say that a pair $\langle X, \mathcal{F}\rangle$ [satisfies the requirements of [@EI17]]{} if - $X$ is a countably compact, non-compact regular space satisfying $h\pi\chi(X)=\aleph_0$, and [.05in]{} - $\mathcal{F}$ is a maximal free filter of closed subsets of $X$ with a separable base. [.05in]{} We will use such $X$ and $\mathcal{F}$ to define the notion of forcing, but before we can do this we need some more definitions lifted from earlier work: Suppose $X$ and $\mathcal{F}$ are as above 1. A subset $A$ of $X$ is [large]{} if $A$ meets every set in $\mathcal{F}$. A set that is not large is called [small]{}. [.05in]{} 2. A [promise]{} is a function whose domain is a large subset of $X$ such that for each $x\in\operatorname{dom}(f)$, $f(x)$ is an open neighborhood of $x$. [.05in]{} Given these definitions, we define the notion of forcing $\mathbb{P}[X,\mathcal{F}]$ as follows: A forcing condition $p$ is a triple $(\sigma_p, A_p,\Phi_p)$ where 1. $\sigma_p$ is a one–to–one function from some countable ordinal into $X$, and we define $[p]:=\operatorname{ran}(\sigma_p)$ 2. $A_p\in{\mathcal{F}}$ 3. $\Phi_p$ is a countable set of promises (see below) 4. $\operatorname{cl}_X([p])\cap A_p=\emptyset$ and a condition $q$ extends $p$ (written $q\leq p$) if 5. $\sigma_q{\supseteq}\sigma_p$ 6. $A_q{\subseteq}A_p$ 7. $\Phi_q{\supseteq}\Phi_p$ 8. $[q]\setminus[p]{\subseteq}A_p$ 9. if $f\in\Phi_p$, then the set $$Y(f, q, p):=\{x\in\operatorname{dom}f: [q]\setminus[p]{\subseteq}f(x)\}$$ is large, and $f{\upharpoonright}Y(f, q, p)\in\Phi_q$. Most of the paper [@EI17] is concerned with working out properties of the above notion of forcing. For example, the notion of forcing is proper, and in the generic extension every countable sequence of ground model elements is already in the ground model. We will be more systematic describing the properties of $\mathbb{P}[X,\mathcal{F}]$ later, but for our immediate purposes we need the fact that it adjoins a sequence $\langle x_\alpha:\alpha<\omega_1\rangle$ satisfying the following two conditions: $$\operatorname{cl}_X(\{x_\beta:\beta<\alpha\})\notin\mathcal{F}\text{ for every $\alpha<\omega_1$},$$ and $$\text{for every }A\in \mathcal{F}\text{ there is an $\alpha$ such that }\{x_\beta:\alpha\leq\beta<\omega_1\}\subseteq A.$$ The sequence just described is just $\bigcup\{\sigma_p:p\in G\}$, where $G$ is a generic subset of $\mathbb{P}[X,\mathcal{F}]$. The following lemma gives an additional property of $\mathbb{P}[X,\mathcal{F}]$ that will be needed for our argument: Suppose $X$ and $\mathcal{F}$ are as above, and $\mathcal{U}$ is an open cover of $X$. If $G$ is a generic subset of $\mathbb{P}[X,\mathcal{F}]$ then in $V[G]$ there is an $\alpha<\omega_1$ such that any countable subset of $\{x_\beta:\alpha\leq\beta<\omega_1\}$ is covered by an element of $\mathcal{U}$ Suppose $\mathcal{U}$ is an open cover of $X$; our plan is to show that the set of conditions forcing the existence of such an $\alpha$ is dense. To that end, let $p = (\sigma_p, A_p, \Phi_p)$ be a condition. For each $x\in X$, choose $U_x\in\mathcal{U}$ with $x\in U_x$, and let $f:X\rightarrow\mathcal{U}$ be the function mapping $x$ to $U_x$. Note that $f$ is a promise, and $$q:=(\sigma_p, A_p, \Phi_p\cup\{f\})$$ is an extension of $p$ in $\mathbb{P}[X,\mathcal{F}]$. Suppose now that $G$ is a generic subset of $\mathbb{P}$ containing $q$. We will work in the extension $V[G]$. First, we set $\alpha = \operatorname{dom}(p)$. It suffices to show that whenever $\alpha<\beta<\omega_1$, the set $\{x_\gamma:\alpha\leq\gamma<\beta\}$ is covered by an element of $\mathcal{U}$. Given such a $\beta$, we fix $r\in G$ with $\beta\subseteq\operatorname{dom}(r)$ (such an $r$ can be found by our discussion preceding the lemma). Since $G$ is a filter on $\mathbb{P}[X,\mathcal{F}]$, we may assume that $r$ extends $q$. Since $f\in\Phi_q$, it follows that $$[r]\setminus[q]\subseteq f(x)\text{ for a large set of }x.$$ But $f(x)$ is always an element of $\mathcal{U}$, and $$\{x_\gamma:\alpha\leq\gamma<\beta\}\subseteq [r]\setminus [p]=[r]\setminus [q],$$ so we have what we need. Armed with the above, we show how $\mathbb{P}[X,\mathcal{F}]$ can be used to destroy certain bad triples: \[usefullemma\] Suppose $\bar{X}=\langle X, Y, z\rangle$ is a bad triple. Further assume there is an $\mathcal{F}$ satisfying: - $\mathcal{F}$ is a maximal free filter of closed subsets of $X$, [.05in]{} - for any open neighborhood $U$ of $z$ in $Y$, there is an $A\in\mathcal{F}$ with $A\subseteq U$, and [.05in]{} - $\mathcal{F}$ has a base of separable sets. [.05in]{} Let $G$ be a generic subset of $\mathbb{P}[X,\mathcal{F}]$. Then $$V[G]\models \text{``}t(z, Y)>\aleph_0\text{"}.$$ We have already mentioned that in the generic extension there is a sequence $\langle x_\alpha:\alpha<\omega_1\rangle$ with the following properties: 1. for each $A\in\mathcal{F}$, there is an $\alpha$ such that $\{x_\beta:\alpha\leq\beta<\omega_1\}\subseteq U$, and [.05in]{} 2. if $\mathcal{U}$ is an open cover of $X$ in $V$, then there is an $\alpha$ such that any countable subset of $\{x_\beta:\alpha\leq\beta<\omega_1\}$ is covered by an element of $\mathcal{U}$. [.05in]{} Let $\mathcal{U}$ be (in $V$) an open cover of $X$ by sets $U$ with $z\notin\operatorname{cl}_Y(U)$ (this is possible by regularity of $Y$). In $V[G]$, we can find $\alpha<\omega_1$ such that any countable subset of $\{x_\beta:\alpha\leq\beta<\omega_1\}$ is covered by an element of $\mathcal{U}$. Now define $$W:=\{x_\beta:\alpha\leq\beta<\omega_1\}.$$ By (1) above, it follows that $z\in \operatorname{cl}_Y(W)$. By our choice of $\alpha$, $z$ is not in the closure of any countable subset of $W$. Thus, $W$ witnesses that $t(z, Y)>\aleph_0$, as needed. Now suppose [CH]{} holds, let $\vec{X}=\langle X, Y, z\rangle$ be a bad triple, and let $G$ be a generic subset of $\mathbb{P}[\vec{X}]$. In $V[G]$, we know there is a filter $\mathcal{F}$ as in the conclusion of Theorem \[summary1\]. If $\vec{X}$ is still a bad triple in $V[G]$, then the pair $\langle X, \mathcal{F}\rangle$ satisfies the assumptions of Lemma \[usefullemma\], and hence forcing with $\mathbb{P}[X,\mathcal{F}]$ over $V[G]$ will give us $t(z, Y)>\aleph_0$. In any case, we have demonstrated that a bad triple can be destroyed by a “reasonable” notion of forcing, as long as [CH]{} holds. The following two definitions give us some notation to keep things organized. \[type\] Let $\vec{X}$ be a relevant triple. - We say $\vec{X}$ is of Type 0 if $\vec{X}$ is not a bad triple. [.05in]{} - We say $\vec{X}$ is of Type 1 if it is a bad triple, but after forcing with $\mathbb{P}[\vec{X}]$ it is no longer bad. [.05in]{} - We say $\vec{X}$ is of Type 2 if it is of neither of the previous two types. Let $\vec{X}=\langle X, Y, z\rangle$ be a relevant triple. - We let $\mathbb{P}^0[\vec{X}]$ be the trivial (one-element) notion of forcing. [.05in]{} - $\mathbb{P}^1[\vec{X}]$ denotes the forcing $\mathbb{P}[\vec{X}]$. [.05in]{} - If $\vec{X}$ is of Type 2, we let $\mathbb{P}^2[\vec{X}]$ denotes the forcing $\mathbb{P}[\vec{X}]\mathbb{P}[X,\dot{\mathcal{F}}]$, as discussed in the paragraph prior to Definition \[type\]. [.05in]{} We end with the following summary of our work in this section: \[kill\] Suppose the Continuum Hypothesis holds. Then any relevant triple $\vec{X}$ is of Type n for some unique $n\in\{0, 1, 2\}$. Furthermore, if $\vec{X}$ is of Type $n$, then $$V^{\mathbb{P}^n[\vec{X}]}\models\text{``$\vec{X}$ is not a bad triple''.}$$ Putting it all together ======================= In this section we will build our model of [ZFC + CH + $\circledast$]{}. Our plan is standard: we assume [GCH]{} in the ground model and build a countable support iteration of length $\omega_2$ where at each stage we treat a relevant triple using Theorem \[kill\]. We start with an [ad hoc]{} definition: A notion of forcing $\mathbb{P}$ is [admissible]{} if $\mathbb{P}$ is trivial, or of the form $\mathbb{P}[\vec{X}]$ for some bad triple $\vec{X}$, or of the form $\mathbb{P}[X,\mathcal{F}]$ where $\|X\|=\aleph_1$ and $\langle X, \mathcal{F}\rangle$ is as in Definition \[41\]. A countable support iteration $\bar{\mathbb{P}}=\langle \text{ $\mathbb{P}_\alpha$, $\dot{\mathbb{Q}}_\beta:\alpha\leq\omega_2$, $\beta<\omega_2$}\rangle$ is [admissible]{} if for each $\alpha<\omega_2$, $${\Vdash}_{\alpha}\text{``$\dot Q_\alpha$ is admissible''}.$$ Our goal is to show that there is an admissible iteration $\bar{\mathbb{P}}$ such that $${\Vdash}_{\omega_2}\text{``{\sf CH } + $\circledast$''}.$$ Most of the results in this section will consist of pointers back to results already in the literature. The conclusions we draw are simply stated, but the results we need to obtain these conclusions often involve very technical concepts. Our plan is to treat much of the earlier work as a black box, as we do in the next proposition. Admissible iterations satisfy the assumptions of Theorem 4 in [@EI15]. Thus, the limit $\mathbb{P}_{\omega_2}$ is totally proper, and for any $\alpha<\omega_2$ the quotient forcing $\mathbb{P}_{\omega_2}/\dot{G}_\alpha$ is totally proper. Let $\bar{\mathbb{P}}$ be an admissible iteration. Theorem 4 of [@EI15] requires the iterands of $\bar{\mathbb{P}}$ to have three properties: they each must be totally proper, weakly $<\omega_1$-proper, and satisfy a complicated “iteration condition”. We do not need the details of these definitions here because sections 4, 5, and 6 of [@EI17] are devoted to showing that forcings of the form $\mathbb{P}[X,\mathcal{F}]$ satisfy these three conditions, and countably closed forcings satisfy them in a trivial way. The conclusion of Theorem 4 of [@EI15] is that $\mathbb{P}_{\omega_2}$ is totally proper, and the result regarding quotient forcing follows from Proposition 6.13 of the same paper. The quotient forcing portion of the preceding proposition is actually saying something quite simple: if $G$ is a generic subset of $\mathbb{P}_{\omega_2}$ and $G_\alpha$ is the canonical generic subset of $\mathbb{P}_\alpha$ obtained from $G$, then $V[G]$ is obtained from $V[G_\alpha]$ by forcing with a totally proper notion of forcing. \[cor1\] Let $\bar{\mathbb{P}}$ be an admissible iteration, and let $G$ be a generic in subset of $\mathbb{P}_{\omega_2}$. 1. If [CH]{} holds in $V$, it continues to hold in $V[G]$. [.05in]{} 2. Suppose $\alpha<\omega_2$ and $\vec{X}$ is a relevant triple in $V[G_\alpha]$. Then $\vec{X}$ is a relevant triple in $V[G]$. Furthermore, if $\vec{X}$ is not a bad triple in $V[G_\alpha]$, then it is not a bad triple in $V[G]$. The first statement follows since $\mathbb{P}_{\omega_2}$ is totally proper. For the second, let $\vec{X}=\langle X, Y, z\rangle$ be a relevant triple in $V[G_\alpha]$. Since $V[G]$ is obtained from $V[G_\alpha]$ through forcing with a totally proper notion of forcing, we know $X$ has the same countable subsets in both $V[G_\alpha]$ and $V[G]$, and hence $X$ remains countably compact in $V[G]$. The other aspects of relevance are preserved automatically. Suppose now that $\vec{X}$ is not a bad triple in $V[G_\alpha]$. This can happen by one of two ways: either $h\pi\chi(X)>\aleph_0$, or $t(z, Y)>\aleph_0$. Destroying either of these conditions would require adding a new countable sequence of elements of $V[G_\alpha]$ — either a countable sequence of basic open sets of $X$ in the first case, or a new countable of points of $X$ in the second. This cannot happen as $V[G]$ is obtained from $V[G_\alpha]$ via a totally proper notion of forcing. The above corollary tells us that if a bad triple is destroyed at some stage of an admissible iteration, then it stays dead. Clearly this will be an important ingredient in our construction, but we need to be able to argue that every relevant triple in the final extension has had its “badness” destroyed at some stage along the way. To do this, we need to show that the limit forcing $\mathbb{P}_{\omega_2}$ has the $\aleph_2$-chain condition. Again, this will follow after some appeals to the literature. In this case, we will need to take advantage of previous work on the so-called [$\aleph_2$-properness isomorphism condition]{} (abbreviated $\aleph_2$-p.i.c.). We will not give the (complicated) definition here, as all we need to know is encapsulated in the following three propositions: \[blackbox1\] Suppose $\bar{P}=\langle \mathbb{P}_\alpha,\dot{\mathbb{Q}}_\beta:\alpha\leq\omega_2, \beta<\omega_2\rangle$ is a countable support iteration such that $${\Vdash}_\alpha\text{``$\dot{\mathbb{Q}}_\alpha$ satisfies the $\aleph_2$-p.i.c.''}.$$ Then 1. $\mathbb{P}_\alpha$ satisfies the $\aleph_2$-p.i.c for each $\alpha<\omega_2$, and [.05in]{} 2. if $2^{\aleph_0}=\aleph_1$ then $\mathbb{P}_{\omega_2}$ satisfies the $\aleph_2$-c.c. This is a special case of Lemma 2.4 on page 410 of [@pif]. Abraham also gives a nice treatment of this in Section 5.4 of [@proper]. \[blackbox3\] Let $\mathbb{P}$ be an admissible notion of forcing. If [CH]{} holds, then $\mathbb{P}$ satisfies the $\aleph_2$-p.i.c. Lemma 2.5 on page 411 of [@pif] tells us that a proper notion of forcing of size $\aleph_1$ satisfies the $\aleph_2$-p.i.c. Clearly nothing more needs to be said for the trivial forcing, and under [CH]{} if $\vec{X}$ is a bad triple, then $\mathbb{P}[\vec{X}]$ is of cardinality $\aleph_1$ and hence satisfies the $\aleph_2$-p.i.c. as well. For $\mathbb{P}$ of the form $\mathbb{P}[X,\mathcal{F}]$ with $\|X\|=\aleph_1$, we get the $\aleph_2$-p.i.c. by way of Theorem 7.2 of [@EI17]. Now we can put the pieces together to obtain the following lemma: Suppose $\bar{\mathbb{P}}$ is an admissible iteration, and $2^{\aleph_0}=\aleph_1$ and $2^{\aleph_1}=\aleph_2$ in the ground model. Then - $\mathbb{P}_{\omega_2}$ satisfies the $\aleph_2$-chain condition, and [.05in]{} - ${\Vdash}_{\omega_2}\text{``}2^{\aleph_1}=\aleph_2\text{''}$. We know that our iteration satisfies the assumptions of Proposition \[blackbox1\], so for each $\alpha<\omega_2$, $\mathbb{P}_\alpha$ satisfies the $\aleph_2$-p.i.c., and since [CH]{} holds we know $\mathbb{P}_{\omega_2}$ has the $\aleph_2$-chain condition. We preserve $2^{\aleph_1}=\aleph_2$ for standard reasons, as we are iterating $\aleph_2$-c.c. posets, and (by induction) each iterand is forced to have cardinality at most $2^{\aleph_1}=\aleph_2$. We now have everything we need to construct our model. Assume $2^{\aleph_0}=\aleph_1$ and $2^{\aleph_1}=\aleph_2$. There is an admissible iteration $\bar{\mathbb{P}}$ so that $${\Vdash}_{\omega_2}\text{`` {\sf CH} } + \circledast \text{''}.$$ The iteration is built by induction on $\alpha<\omega_2$. At stage $\alpha$, some fixed bookkeeping procedure will hand us a $\mathbb{P}_\alpha$ name $\dot{\vec{X}}_\alpha$ for a relevant triple. Since [CH]{} holds in $V^{\mathbb{P}_\alpha}$, we know $${\Vdash}_\alpha\text{``$\dot{\vec{X}}_\alpha$ is of Type $n$ for some $n<3$''},$$ and we let $\dot{\mathbb{Q}}_\alpha$ be a $\mathbb{P}_\alpha$ name such that $${\Vdash}_\alpha\text{``$\dot{\mathbb{Q}}_\alpha$ is $\mathbb{P}^n[\dot{\vec{X}}_\alpha]$ where $\dot{\vec{X}}_\alpha$ is of Type $n$.''}$$ Standard arguments using the $\aleph_2$-chain condition allow us to arrange the bookkeeping so that every relevant triple in the final extension $V^{\mathbb{P}_{\omega_2}}$ is considered at some stage of the iteration. This is where we use our assumption that relevant triples consist of topological spaces with underlying set $\omega_1$, as this guarantees there are only $\aleph_2$ relevant triples that need to be considered. Now let $ {G}$ be a generic subset of $\mathbb{P}_{\omega_2}$, and for $\alpha<\omega_2$ let $G_\alpha$ be the generic subset of $\mathbb{P}_\alpha$ obtained from $ {G}$. We know that [CH]{} is true in $V[G]$ by virtue of Corollary \[cor1\], so we need only worry about $\circledast$. Since [CH]{} holds in $V[G]$, Proposition \[prop2\] tells us that it suffices to show that there are no bad triples in $V[G]$. If $\vec{X}$ is a relevant triple in $V[G]$, there is an $\alpha<\omega_2$ such that $\vec{X}\in V[G_\alpha]$, and $\vec{X}=\vec{X}_\alpha$. Given the forcing we do at stage $\alpha$, we know that $\vec{X}$ is not a bad triple in $V[G_{\alpha+1}]$, and we can then apply the second part of Corollary \[cor1\] to conclude that $\vec{X}$ is not a bad triple in $V[G]$. Therefore, $\circledast$ holds in $V[G]$. Questions ========= Finally, we collect here a few questions. A few of these are well-known and taken from Shakhmatov’s [@Shak1992], while others are specifically motivated by the construction presented here. Is it consistent that every countably compact, countably tight space is sequential? This is Problem 2.12 from [@Shak1992]. Dow [@dow2] has shown that a counterexample exists if $\mathfrak{b}=2^{\aleph_0}$. The proof of Proposition \[chmm\] required both $\circledast$ and $2^{\aleph_0}<2^{\aleph_1}$ in order to establish that compact spaces of countable tightness are sequential. It is not clear if $\circledast$ alone is sufficient for this: Does $\circledast$ imply that compact spaces of countable tightness are sequential? The preceding question is of course closely related to Ismail and Nyikos’s Problem ([@ismailnyikos], but see also Problem 3.2 in [@Shak1992]). If a compact space is $C$-closed, must it be sequential? Our proof of the consistency of $\circledast$ made heavy use of the fact that the Continuum Hypothesis held in the final model. It is not clear how to obtain the result without this assumption, and this leads us to the following question: Is $\circledast$ consistent with the failure of [CH]{}? Does it follow from [PFA]{}? Finally, we finish with a related question mentioned in [@dow2]: Is it consistent that every regular countably tight space is $C$-closed? [10]{} Uri Abraham. Proper forcing. In [Handbook of set theory. [V]{}ols. 1, 2, 3]{}, pages 333–394. Springer, Dordrecht, 2010. A. V. Arhangel$'$ski[ĭ]{}. The structure and classification of topological spaces and cardinal invariants. , 33(6(204)):29–84, 272, 1978. Zolt[á]{}n T. Balogh. On compact [H]{}ausdorff spaces of countable tightness. , 105(3):755–764, 1989. Alan Dow. A countably compact, countably tight, nonsequential space. In [General topology and applications ([M]{}iddletown, [CT]{}, 1988)]{}, volume 123 of [Lecture Notes in Pure and Appl. Math.]{}, pages 71–80. Dekker, New York, 1990. Todd Eisworth. Countable compactness, hereditary [$\pi$]{}-character, and the continuum hypothesis. , 153(18):3572–3597, 2006. Todd Eisworth and Peter Nyikos. First countable, countably compact spaces and the continuum hypothesis. , 357(11):4269–4299 (electronic), 2005. V. V. Fedor[č]{}uk. Completely closed mappings, and the compatibility of certain general topology theorems with the axioms of set theory. , 99 (141)(1):3–33, 135, 1976. Mohammad Ismail and Peter Nyikos. On spaces in which countably compact sets are closed, and hereditary properties. , 11(3):281–292, 1980. A. J. Ostaszewski. On countably compact, perfectly normal spaces. , 14(3):505–516, 1976. D. V. Ran[č]{}in. Tightness, sequentiality and closed coverings. , 232(5):1015–1018, 1977. B. [È]{}. [Š]{}apirovski[ĭ]{}. -character and [$\pi $]{}-weight in bicompacta. , 223(4):799–802, 1975. Dmitri B. Shakhmatov. Compact spaces and their generalizations. In [Recent progress in general topology ([P]{}rague, 1991)*]{}, pages 571–640. North-Holland, Amsterdam, 1992. Saharon Shelah. . Perspectives in Mathematical Logic. Springer-Verlag, Berlin, second edition, 1998.	{ "pile_set_name": "ArXiv" }
[Global unique solvability\ of the initial-boundary value problem for the equations of\ one-dimensional polytropic flows\ of viscous compressible multifluids]{} [Alexander Mamontov,Dmitriy Prokudin]{} [September 18, 2018]{} [Lavrentyev Institute of Hydrodynamics,\ Siberian Branch of the Russian Academy of Sciences\ pr. Lavrent’eva 15, Novosibirsk 630090, Russia]{} [Abstract]{} We consider the equations which describe polytropic one-dimensional flows of viscous compressible multifluids. We prove global existence and uniqueness of a solution to the initial-boundary value problem which corresponds to the flow in a bounded space domain. [Keywords:]{} multifluid, viscous compressible flow, initial-boundary value problem, polytropic flow, global existence, uniqueness Introduction ============ The paper is devoted to the analysis of the solvability of the equations of motion of multicomponent viscous compressible fluids (homogeneous mixtures of fluids, multifluids). Concerning the origin of the model and its physical interpretation, we refer the reader to [@mamprok.jfcs2]. An overview of the options for formulating the model and the known results can be found in [@mamprok.france] and [@mamprok.semi17]. Related multi-velocity models of multifluids are considered in [@mamprok.dorovski], [@mamprok.giovang], [@mamprok.nigm] and [@mamprok.raj]. As the first results on the well-posedness of the multidimensional equations of multifluids, we can refer to [@mamprok.frehsegm1], [@mamprok.frehsegm2] and [@mamprok.frehsew]. Solvability for related models is shown in [@mamprok.jms], [@mamprok.muchapz], [@mamprok.semiprok17] and [@mamprok.jfcsprok]. Weak solutions for multidimensional barotropic problems for the model considered in the paper are constructed in the steady version in [@mamprok.semi162] and [@mamprok.jaimprok] (polytropic case), and then in [@mamprok.smz1] and [@mamprok.smz2] (general case); in the unsteady version in [@mamprok.semi161] (polytropic case), and then in [@mamprok.izvran18] (general case). Similar results for the heat-conductive model are obtained in [@mamprok.izvran14]. For a number of reasons, including the purpose of constructing more regular solutions, one-dimensional formulations are of interest. It should be noted that the dimension with respect to the number of components (constituents) of a multifluid is not logically and technically related to the number of the spatial variables, and the interaction of the constituents via the viscous terms transforms the system of differential equations governing the motion of a multifluid into a system essentially different from the one-component system. Therefore, despite the developed theory of one-dimensional flows of a viscous gas (see [@mamprok.kazhshel] for example), the one-dimensional theory of multifluids remains relevant. The specificity of the paper is that we consider a variant of the model with an average velocity in the transport operator. Statement of the problem ======================== Let us consider the system of equations governing motions of multicomponent viscous compressible fluids without taking into account chemical reactions in the case of one spatial variable: $$\partial_{t}\rho_{i}+\partial_{x}(\rho_{i} v)=0,$$ $$\rho_{i}\left(\partial_{t}u_{i}+v\partial_{x}u_{i}\right)=\partial_{x}P_{i},\quad i=1,\ldots,N.$$ Here $N\geqslant 2$ is the number of components, $\rho_{i}$ is the density if the $i$-th constituent, $u_{i}$ is the velocity of the $i$-th component, $\displaystyle v=\frac{1}{N}\sum\limits_{i=1}^{N}u_{i}$ is the average velocity of the multifluid, and $P_{i}$ are the stresses. We accept generalized Newton’s hypothesis $$P_{i}=-p+\sum\limits_{j=1}^{N}\mu_{ij}\partial_{x}u_{j},$$ where $p$ is the pressure, and the viscosity coefficients $\{\mu_{ij}\}_{i, j=1}^{N}$ form the symmetric matrix $\textbf{M}$. Moreover, $\textbf{M}>0$, i. e. $(\textbf{M}\boldsymbol{\xi},\boldsymbol{\xi})\geqslant C_{0}(\textbf{M})\|\boldsymbol{\xi}\|^2$ for all $\boldsymbol{\xi}\in{\mathbb R}^N$ with a constant $C_{0}(\textbf{M})>0$. The written equations together with the constitutive equation $$p=K\rho^{\gamma},\quad \rho=\sum\limits_{i=1}^{N}\rho_{i},\quad K={\rm const}>0,\quad \gamma={\rm const}>1$$ form a closed system $$\begin{aligned} \label{newcontinuity.double}\partial_{t}\rho_{i}+\partial_{x}(\rho_{i} v)=0,\quad v=\frac{1}{N}\sum\limits_{i=1}^{N}u_{i},\end{aligned}$$ $$\begin{aligned} \label{newmomentum.double}\rho_{i}\left(\partial_{t}u_{i}+v\partial_{x}u_{i}\right)+K\partial_{x} \rho^{\gamma}= \sum\limits_{j=1}^N \mu_{ij}\partial_{xx}u_{j},\quad i=1,\ldots,N,\quad \rho=\sum\limits_{i=1}^{N}\rho_{i}.\end{aligned}$$ Let us consider this system in the rectangular $Q_{T}$ (here and below$Q_{t}=(0, 1)\times(0, t)$) with an arbitrary finite height $T$, $0<T<\infty$, and endow this system with the following initial and boundary conditions ($i=1,\ldots,N$): $$\begin{aligned} \label{nachusl.double}\rho_{i}\|_{t=0}=\rho_{0i}(x), \quad u_{i}\|_{t=0}=u_{0i}(x),\quad x\in [0, 1],\end{aligned}$$ $$\begin{aligned} \label{boundvelocity.double}u_{i}\|_{x=0}=u_{i}\|_{x=1}=0,\quad t\in [0, T].\end{aligned}$$ [Definition 1.]{} By a strong solution to the problem – we mean a collection of $2N$ functions $(\rho_{1},\ldots, \rho_{N}, u_{1},\ldots, u_{N})$ such that the equations and are satisfied a. e. in $Q_{T}$, the initial conditions are satisfied for a. a. $x\in (0, 1)$, the boundary conditions are satisfied for a. a. $ t\in (0, T)$, and the following inequalities and inclusions hold ($i=1,\ldots,N$) $$\rho_{i}>0,\quad \rho_{i}\in L_{\infty}\big(0, T; W^{1}_{2}(0, 1)\big), \quad \partial_{t}\rho_{i}\in L_{\infty}\big(0, T; L_{2}(0, 1)\big),$$ $$u_{i}\in L_{\infty}\big(0, T; W^{1}_{2}(0, 1)\big)\bigcap L_{2}\big(0, T; W^{2}_{2}(0, 1)\big),\quad \partial_{t}u_{i} \in L_{2}(Q_{T}).$$ Main result =========== Suppose that the initial data in satisfy the conditions $$\rho_{0i}\in W^{1}_{2}(0,1),\quad \rho_{0i}>0,\quad u_{0i}\in \overset{\circ}{W^1_2}(0, 1),\quad i=1,\ldots,N\quad (N\geqslant 2),$$ the symmetric viscosity matrix $\textbf{M}$ is positive definite, and the polytropic index $\gamma>1$ as well as the constants $0<K,T<\infty$ are given. Then there exists a unique strong solution to the problem – in the sense of Definition 1. Since the uniqueness and the corresponding local result are obtained in [@mamprok.local], the proof of Theorem 2 reduces to obtaining global a priori estimates, which is the main content of the paper. Lagrangian coordinates ====================== During the study of the problem –, the parallel use of the Lagrangian coordinates is convenient. Let us accept $\displaystyle y(x,t)=\int\limits_{0}^{x}\rho(s,t)\,ds$ and $t$ as new independent variables. Then the system , takes the form $$\begin{aligned} \label{1newcontinuity1lagr} \partial_{t}\rho_{i}+\rho\rho_{i}\partial_{y}v=0,\quad v=\frac{1}{N}\sum\limits_{i=1}^{N}u_{i},\end{aligned}$$ $$\begin{aligned} \label{1newmomentum1lagr} \frac{\rho_{i}}{\rho}\partial_{t}u_{i}+K\partial_{y} \rho^{\gamma}= \sum\limits_{j=1}^N \mu_{ij}\partial_{y}(\rho\partial_{y}u_{j}),\quad i=1,\ldots,N,\quad \rho=\sum\limits_{i=1}^{N}\rho_{i},\end{aligned}$$ the domain $Q_{T}$ is transformed into the rectangular $\Pi_{T}=(0, d)\times(0, T)$, where $\displaystyle d=\int\limits_{0}^{1}\rho_{0}\,dx>0$, $\displaystyle \rho_{0}=\sum\limits_{i=1}^{N}\rho_{0i}$, and the initial and boundary conditions take the form ($i=1,\ldots,N$) $$\begin{aligned} \label{1nachusl1lagr} \rho_{i}\|_{t=0}=\widetilde{\rho}_{0i}(y), \quad u_{i}\|_{t=0}=\widetilde{u}_{0i}(y),\quad y\in [0, d],\end{aligned}$$ $$\begin{aligned} \label{1boundvelocity1lagr} u_{i}\|_{y=0}=u_{i}\|_{y=d}=0,\quad t\in [0, T].\end{aligned}$$ Estimates of the concentrations =============================== Let us consider a hypothetical solution $(\rho_{1},\ldots, \rho_{N}, u_{1},\ldots, u_{N})$ to the problem – which possesses all necessary differential properties, and such that the densities $\rho_{i}$, $i=1,\ldots,N$, are positive and bounded (see Definition 1). First of all, we note that the summation of with respect to $i=1,\ldots,N$ gives $$\begin{aligned} \label{1602171} \partial_{t}\rho+\rho^{2}\partial_{y}v=0,\end{aligned}$$ and hence $$\partial_{t}\left(\frac{\rho_{i}}{\rho}\right)=0, \quad i=1,\ldots,N.$$ Hence, due to we get the equalities $$\begin{aligned} \label{1602173} \frac{\rho_{i}(y,t)}{\rho(y,t)}=\frac{\widetilde{\rho}_{0i}(y)}{\widetilde{\rho}_{0}(y)}\quad\text{as}\quad(y,t)\in[0, d]\times[0, T]\end{aligned}$$ for all $i=1,\ldots,N$, where $\displaystyle \widetilde{\rho}_{0}=\sum\limits_{i=1}^{N}\widetilde{\rho}_{0i}$. In the Eulerian coordinates $(x,t)$, the ratios $\displaystyle \frac{\rho_{i}}{\rho}$ satisfy the transport equations, and we only have the inequalities $$\begin{gathered} \label{2602171} \inf\limits_{[0,1]}\frac{\rho_{0i}(x)}{\rho_{0}(x)}\leqslant\frac{\rho_{i}(x,t)}{\rho(x,t)}\leqslant\sup\limits_{[0,1]}\frac{\rho_{0i}(x)}{\rho_{0}(x)}\leqslant 1\quad\text{as}\quad(x,t)\in[0, 1]\times[0, T],\\ i=1,\ldots,N.\end{gathered}$$ First a priori estimates ======================== Typically for the compressible Navier—Stokes theory, the energy inequality immediately entails the estimates for the kinetic energy, the rate of energy dissipation and the potential energy of the multifluid constituents, as we show below. [Lemma 3.]{} Under the assumptions of Theorem 2, there exists a positive constant[^1] $\displaystyle C_{1}\left(\left\{\inf\limits_{[0,1]}\frac{\rho_{0i}}{\rho_{0}}\right\}, \left\{\\|\sqrt{\rho_{0i}}u_{0i}\\|_{L_{2}(0,1)}\right\}, \\|\rho_{0}\\|_{L_{\gamma}(0,1)}, K, \textbf{M}, N, \gamma\right)$, such that the following estimate holds $$\label{lemma1} \sum\limits_{i=1}^{N}\left(\\|\sqrt{\rho}u_{i}\\|_{L_{\infty}\big(0, T;L_{2}(0, 1)\big)} +\\|\partial_{x}u_{i}\\|_{L_{2}(Q_{T})}\right)+\\|\rho\\|_{L_{\infty}\big(0, T;L_{\gamma}(0,1)\big)}\leqslant C_{1}.$$ [Proof.]{} Let us multiply the equations by $u_{i}$, integrate over $(0, 1)$ and sum with respect to $i=1,\ldots,N$. In view of , and $\textbf{M}>0$, the following relations hold $$\sum\limits_{i=1}^{N}\int\limits^{1}_{0}\Big(\rho_{i}\partial_{t}u_{i}+\rho_{i} v\partial_{x}u_{i}\Big)u_{i}\,dx= \frac{1}{2}\frac{d}{dt}\sum\limits_{i=1}^{N} \int\limits^{1}_{0}\rho_{i} u_{i}^{2}\, dx,$$ $$\sum\limits_{i=1}^{N}\int\limits^{1}_{0}u_{i}K\partial_{x}\rho^{\gamma}\,dx=-KN\int\limits^{1}_{0}\rho^{\gamma}\partial_{x}v\, dx= \frac{KN}{\gamma-1}\frac{d}{dt}\int\limits^{1}_{0}\rho^{\gamma}\, dx,$$ $$\begin{gathered} \label{lemma51.4} \sum\limits_{i,j=1}^N \mu_{ij}\int\limits\limits_{0}^{1}(\partial_{xx}u_{j})u_{i}\, dx=-\sum\limits_{i,j=1}^N \mu_{ij}\int\limits^{1}_{0}(\partial_{x}u_{i})(\partial_{x}u_{j})\,dx\leqslant\\ \leqslant -C_{0}(\textbf{M}) \sum\limits_{i=1}^N\int\limits\limits_{0}^{1}\|\partial_{x}u_{i}\|^{2}\,dx,\end{gathered}$$ and hence we get the inequality $$\label{lemma51.6} \frac{d}{dt}\sum\limits_{i=1}^{N}\int\limits^{1}_{0}\left(\frac{1}{2}\rho_{i} u_{i}^{2}+\frac{K}{\gamma-1}\rho^{\gamma}\right)\,dx+ C_{0}\sum\limits_{i=1}^N\int\limits\limits_{0}^{1}\|\partial_{x}u_{i}\|^{2}\,dx\leqslant 0.$$ We integrate over $(0, t)$, and using , we obtain the bound $$\sum\limits_{i=1}^{N}\int\limits^{1}_{0}\left(\frac{1}{2}\rho_{i} u_{i}^{2}+\frac{K}{\gamma-1}\rho^{\gamma}\right)\,dx+ C_{0}\sum\limits_{i=1}^N\int\limits\limits_{0}^{t}\int\limits\limits_{0}^{1}\|\partial_{x}u_{i}\|^{2}\,dxd\tau\leqslant$$ $$\leqslant\sum\limits_{i=1}^{N}\int\limits^{1}_{0}\left(\frac{1}{2}\rho_{0i}u_{0i}^{2}+\frac{K}{\gamma-1}\rho_{0}^{\gamma}\right)\,dx,$$ which, due to , implies the conclusion of Lemma 3. [Remark 4.]{} In the Lagrangian coordinates, the bound takes the form $$\begin{gathered} \label{lemma1lagr} \sum\limits_{i=1}^{N}\left(\left\\|u_{i}\right\\|_{L_{\infty}\big(0, T;L_{2}(0, d)\big)}+\\|\sqrt{\rho}\partial_{y}u_{i}\\|_{L_{2}(\Pi_{T})}\right)+ \\|\rho\\|_{L_{\infty}\big(0, T;L_{\gamma-1}(0,d)\big)}\leqslant\\ \leqslant C_{2}(C_{1},\gamma).\end{gathered}$$ [Remark 5.]{} In view of , we obviously get the following inequality from $$\label{lemma1lagr2602171} \sum\limits_{i=1}^{N}\left\\|u_{i}\right\\|_{L_{2}\big(0, T;L_{\infty}(0, 1)\big)}\leqslant C_{1}.$$ Bound of the density from above =============================== The crucial a priori estimates are the bounds of strict positiveness and boundedness of the densities of the multifluid constituents. First we prove the bound for the densities from above. [Lemma 6.]{} There exists a constant\ $C_{3}\left(C_{1}, \\|\rho_{0}\\|_{L_{\infty}(0,1)}, \{\\|\rho_{0i}u_{0i}\\|_{L_{1}(0, 1)}\}, K, \textbf{M}, N, T, d, \gamma\right)$ such that $$\label{supernew1602171} \rho(x,t)\leqslant C_{3}\quad\text{as}\quad(x,t)\in[0, 1]\times[0, T].$$ [Proof.]{} Let us rewrite the equations , using , in the form $$\begin{gathered} \label{new1602171} \frac{1}{N}\sum\limits_{j=1}^N \widetilde{\mu}_{ij}\left(\partial_{t}(\rho_{j}u_{j})+\partial_{x}(\rho_{j}v u_{j})\right)+\frac{K}{N}\left(\sum\limits_{j=1}^N \widetilde{\mu}_{ij}\right)\partial_{x} \rho^{\gamma}=\frac{1}{N}\partial_{xx}u_{i},\\ i=1,\ldots,N,\end{gathered}$$ where $\widetilde{\mu}_{ij}$ are the entries of the matrix $\widetilde{\textbf{M}}=\textbf{M}^{-1}>0$, and then sum with respect to $i=1,\ldots,N$, then we get $$\begin{aligned} \label{new1602172} \partial_{t}V=\partial_{x}\left(\partial_{x}v-\widetilde{K}\rho^{\gamma}-vV\right),\end{aligned}$$ where $\displaystyle V=\frac{1}{N}\sum\limits_{i,j=1}^N \widetilde{\mu}_{ij}\rho_{j}u_{j}$, $\displaystyle \widetilde{K}=\frac{K}{N}\sum\limits_{i, j=1}^{N}\widetilde{\mu}_{ij}>0$. We denote $$\label{lemma52.1} \alpha(x,t)=\int\limits_{0}^t\Big(\partial_{x}v-\widetilde{K}\rho^{\gamma}- v V\Big)\,d\tau+\int\limits_{0}^xV_{0}\,ds,$$ where $V_{0}(x)=V(x,0)$. In view of , we have $$\\|\partial_{x}\alpha\\|_{L_{\infty}\big(0, T;L_{1}(0, 1)\big)}=\\|V\\|_{L_{\infty}\big(0, T;L_{1}(0, 1)\big)}\leqslant C_{4}(C_{1}, \textbf{M}, d),$$ $$\sup\limits_{[0, T]}\left\|\int\limits^{1}_{0}\alpha\,dx\right\| \leqslant T\max\limits_{1\leqslant i,j\leqslant N}\|\widetilde{\mu}_{ij}\|\sum\limits_{i=1}^{N}\sup\limits_{[0, T]}\int\limits^{1}_{0}\rho u_{i}^{2}\, dx+\widetilde{K}T\sup\limits_{[0, T]}\int\limits^{1}_{0}\rho^{\gamma}\, dx+$$ $$+\max\limits_{1\leqslant i, j\leqslant N}\|\widetilde{\mu}_{ij}\| \sum\limits_{i=1}^{N}\int\limits_{0}^1\rho_{0i}\|u_{0i}\|\, dx\leqslant C_{5}\left(C_{1}, \{\\|\rho_{0i}u_{0i}\\|_{L_{1}(0, 1)}\},\textbf{M}, \widetilde{K}, T, \gamma\right),$$ and hence, using Poincaré’s inequality, we get[^2] $$\sup\limits_{[0, T]}\int\limits^{1}_{0}\|\alpha\|\, dx \leqslant \sup\limits_{[0, T]}\int\limits^{1}_{0}\|\partial_{x}\alpha \|\,dx+ \sup\limits_{[0, T]}\left\|\int\limits^{1}_{0}\alpha \,dx\right\|\leqslant C_{6}(C_{4}, C_{5}),$$ and we arrive at the boundedness of $\alpha$ in $L_{\infty}\big(0,T;W_{1}^{1}(0, 1)\big)$. Using this and the fact $W_{1}^{1}(0, 1)\hookrightarrow L_{\infty}(0,1)$, we obtain the estimate $$\\|\alpha\\|_{L_{\infty}(Q_{T})}\leqslant C_{7}\left(C_{4}, C_{6}\right).$$ Let us note that, in view of , and , the following relations hold $$d_{t}(\rho e^{\alpha})=-\widetilde{K}e^{\alpha}\rho^{\gamma+1}\leqslant 0,\quad \text{where}\quad d_{t}=\partial_{t}+v\partial_{x},$$ and hence $$\rho e^{\alpha}\leqslant\sup\limits_{[0,1]}{\rho_{0}}\exp\left(\int\limits_{0}^{1}\|V_{0}\|\,dx\right),$$ so that we arrive at the conclusion of Lemma 6. The bound for the derivative of the density =========================================== In order to obtain the bound for the densities from below, we first need to prove the boundedness of the first spatial derivative of the logarithm of the total density. Specifically, the following assertion holds. [Lemma 7.]{} There exists a constant $$\displaystyle C_{8}\left(C_{1}, C_{2}, C_{3}, \left\{\\|\widetilde{u}_{0i}\\|_{L_{2}(0,d)}\right\}, \left\{\left\\|\frac{\widetilde{\rho}_{0i}}{\widetilde{\rho}_{0}}\right\\|_{W^{1}_{2}(0,d)}\right\}, \\|\left(\ln{\widetilde{\rho}}_{0}\right)^{\prime}\\|_{L_{2}(0,d)}, \textbf{M}, N\right)$$ such that $$\begin{aligned} \label{eq01021710} \\|\partial_{y}\ln\rho\\|_{L_{\infty}\big(0, T; L_{2}(0, d)\big)}\leqslant C_{8}.\end{aligned}$$ [Proof.]{} Let us use the equations in the form , . We rewrite the equations as $$\begin{aligned} \label{newnewmomentum.double} \frac{1}{N}\sum\limits_{j=1}^{N}\widetilde{\mu}_{ij}\frac{\rho_{j}}{\rho}\partial_{t}u_{j}+\frac{K}{N}\left(\sum\limits_{j=1}^{N}\widetilde{\mu}_{ij}\right)\partial_{y} \rho^{\gamma}= \frac{1}{N}\partial_{y}(\rho\partial_{y}u_{i}),\quad i=1,\ldots,N,\end{aligned}$$ and then sum with respect to $i=1,\ldots,N$, then we get, using , the resulting relation $$\begin{aligned} \label{newnewmomentum1newnew} \frac{1}{N}\sum\limits_{i, j=1}^{N}\widetilde{\mu}_{ij}\frac{\widetilde{\rho}_{0j}}{\widetilde{\rho}_{0}}\partial_{t}u_{j}+\widetilde{K}\partial_{y}\rho^{\gamma}= \partial_{y}\left(\rho\partial_{y}v\right).\end{aligned}$$ We extract from that $$\begin{aligned} \label{eq0102173} \rho\partial_{y}v=-\partial_{t}\ln\rho\end{aligned}$$ and substitute this into , then we get $$\partial_{ty}\ln\rho+\widetilde{K}\partial_{y}\rho^{\gamma}=-\frac{1}{N}\sum\limits_{i,j=1}^{N}\widetilde{\mu}_{ij} \frac{\widetilde{\rho}_{0j}}{\widetilde{\rho}_{0}}\partial_{t}u_{j}.$$ We multiply this equality by $\displaystyle \partial_{y}\ln\rho=:w$ and integrate over $y\in (0,d)$, then we obtain $$\begin{aligned} \label{eq0102172} \frac{1}{2}\frac{d}{dt}\left(\int\limits_{0}^{d}w^{2}\, dy\right)+\widetilde{K}\gamma\int\limits_{0}^{d}\rho^{\gamma}w^{2}\,dy =-\frac{1}{N}\sum\limits_{i, j=1}^{N}\widetilde{\mu}_{ij}\int\limits_{0}^{d}\left(\frac{\widetilde{\rho}_{0j}}{\widetilde{\rho}_{0}}\partial_{t}u_{j}\right)w\,dy.\end{aligned}$$ Let us transform the right-hand side of via the integration by parts and using : $$\begin{aligned} \label{eq01021777}\begin{array}{c}\displaystyle -\frac{1}{N}\sum\limits_{i, j=1}^{N}\widetilde{\mu}_{ij}\int\limits_{0}^{d}\left(\frac{\widetilde{\rho}_{0j}}{\widetilde{\rho}_{0}}\partial_{t}u_{j}\right)w\,dy =-\frac{d}{dt}\left(\frac{1}{N}\sum\limits_{i, j=1}^{N}\widetilde{\mu}_{ij}\int\limits_{0}^{d} \frac{\widetilde{\rho}_{0j}}{\widetilde{\rho}_{0}}u_{j}w\,dy\right)+\\ \\ \displaystyle +\frac{1}{N}\sum\limits_{i, j=1}^{N}\widetilde{\mu}_{ij}\int\limits_{0}^{d} \rho u_{j}(\partial_{y}v)\left(\frac{\widetilde{\rho}_{0j}}{\widetilde{\rho}_{0}}\right)^{\prime}\,dy+ \frac{1}{N}\sum\limits_{i, j=1}^{N}\widetilde{\mu}_{ij}\int\limits_{0}^{d}\frac{\widetilde{\rho}_{0j}\rho}{\widetilde{\rho}_{0}}(\partial_{y}v)(\partial_{y}u_{j})\,dy. \end{array}\end{aligned}$$ Thus, after integration of with respect to $t$, taking into account and , we get $$\\|w\\|^{2}_{L_{2}(0, d)}+2\widetilde{K}\gamma\int\limits_{0}^{t}\int\limits_{0}^{d}\rho^{\gamma}w^{2}\,dyd\tau\leqslant$$ $$\leqslant\\|w_{0}\\|^{2}_{L_{2}(0, d)}-\frac{2}{N}\sum\limits_{i, j=1}^{N}\widetilde{\mu}_{ij}\int\limits_{0}^{d} \frac{\widetilde{\rho}_{0j}}{\widetilde{\rho}_{0}}u_{j}w\,dy+ \frac{2}{N}\sum\limits_{i, j=1}^{N}\widetilde{\mu}_{ij}\int\limits_{0}^{d} \frac{\widetilde{\rho}_{0j}}{\widetilde{\rho}_{0}}\widetilde{u}_{0j}w_{0}\,dy+$$ $$+\frac{2\sqrt{C_{3}}}{N}\sum\limits_{i, j=1}^{N}\|\widetilde{\mu}_{ij}\|\int\limits_{0}^{t}\left\\|\left(\frac{\widetilde{\rho}_{0j}}{\widetilde{\rho}_{0}}\right)^{\prime}\right\\|_{L_{2}(0, d)}\\|u_{j}\\|_{L_{\infty}(0, d)}\\|\sqrt{\rho}\partial_{y}v\\|_{L_{2}(0, d)}\,d\tau+$$ $$+\frac{2}{N}\sum\limits_{i, j=1}^{N}\|\widetilde{\mu}_{ij}\|\sup\limits_{[0,d]}\frac{\widetilde{\rho}_{0j}}{\widetilde{\rho}_{0}}\int\limits_{0}^{t}\\|\rho (\partial_{y}v)(\partial_{y}u_{j})\\|_{L_{1}(0, d)}\,d\tau,$$ where $w_{0}=(\ln{\widetilde{\rho}_{0}})^{\prime}$. Using the estimates and , we derive from this the inequality , concluding the proof of Lemma 7. The estimate of the density from below ====================================== In this section, we finish obtaining the crucial estimates of strict positiveness and boundedness of the densities via the following assertion. [Lemma 8.]{} There exists a constant $C_{9}(C_{8},d)$ such that $$\begin{aligned} \label{eq01021714} \rho(y,t)\geqslant C_{9}\quad\text{as}\quad(y,t)\in[0, d]\times[0, T].\end{aligned}$$ [Proof.]{} The continuity equation for $\rho$ immediately leads, for any$t\in[0,T]$, to the existence of a point $z(t)\in[0, d]$ such that $$\begin{aligned} \label{eq0102176} \rho(z(t),t)=d.\end{aligned}$$ Hence, we can use the representation $$\ln\rho(y,t)=\ln\rho(z(t),t)+\int\limits_{z(t)}^{y}\partial_{s}\ln\rho(s,t)\, ds,$$ from which, via Hölder’s inequality, and using and , we get $$\|\ln\rho(y,t)\|\leqslant \|\ln{d}\|+\sqrt{d}\\|w\\|_{L_{2}(0,d)}\leqslant C_{10}(C_{8},d).$$ This leads immediately to , and Lemma 8 is proved. [Remark 9.]{} The equalities and the estimates in Lemmas 6 and 8 imply that for all $i=1,\ldots,N$ we have $$\begin{aligned} \label{eq01021714111} C_{11}\leqslant\rho_{i}(y,t)\leqslant C_{3}\quad\text{as}\quad(y,t)\in[0, d]\times[0, T],\end{aligned}$$ where $\displaystyle C_{11}=C_{11}\left(C_{9},\left\{\inf\limits_{[0,d]}\frac{\widetilde{\rho}_{0i}(y)}{\widetilde{\rho}_{0}(y)}\right\}\right)$. Further bounds ============== Concluding the proof of Theorem 2, we obtain the estimates for the derivatives of the densities and velocities of the multifluid constituents. [Remark 10.]{} From the estimates in Lemmas 6 and 7 and the formula , it follows that $$\begin{aligned} \label{finalocenla} \left\\|\partial_{x}\rho_{i}\right\\|_{L_{\infty}(0, T; L_{2}(0,1))}\leqslant C_{12},\quad i=1,\ldots,N,\end{aligned}$$ where $\displaystyle C_{12}=C_{12}\left(C_{3}, C_{8}, \left\{\left\\|\left(\frac{\widetilde{\rho}_{0}}{\widetilde{\rho}_{0i}}\right)^{\prime}\right\\|_{L_{2}(0,d)}\right\}, \left\{\sup\limits_{[0,d]}\frac{\widetilde{\rho}_{0i}}{\widetilde{\rho}_{0}}\right\}\right)$. [Lemma 11.]{} There exists\ $C_{13}\left(C_{1}, C_{3}, C_{11}, C_{12}, \{\\|u_{0i}^{\prime}\\|_{L_{2}(0, 1)}\}, K, \textbf{M}, N, T, \gamma\right)$ such that $$\sum\limits_{i=1}^{N}\left(\\|\partial_{x}u_{i}\\|_{L_{\infty}(0, T; L_{2}(0,1))}+\\|\partial_{xx}u_{i}\\|_{L_{2}(Q_{T})}+\\|\partial_{t}u_{i}\\|_{L_{2}(Q_{T})}\right)\leqslant C_{13}.$$ [Proof.]{} We derive from that $$\begin{aligned} \label{eq01021715} \left\\|(\partial_{x}\rho)(t)\right\\|_{L_{2}(0,1)}\leqslant C_{14}(C_{12}, N)\quad \forall\, t\in[0,T].\end{aligned}$$ Using the idea of [@mamprok.basovshel], we square the momentum equations and sum the result with respect to $i=1,\ldots,N$, then we get $$\begin{gathered} \label{1303171}\sum\limits_{i=1}^{N}\rho_{i}(\partial_{t}u_{i})^{2}+\sum\limits_{i=1}^{N}\frac{1}{\rho_{i}} \left(\sum\limits_{j=1}^{N}\mu_{ij}\partial_{xx}u_{j}\right)^{2}-2\sum\limits_{i=1}^{N}(\partial_{t}u_{i})\left(\sum\limits_{j=1}^{N}\mu_{ij}\partial_{xx}u_{j}\right)=\\ =\sum\limits_{i=1}^{N}\rho_{i}\left(\frac{K\partial_{x}\rho^{\gamma}}{\rho_{i}}+v\partial_{x}u_{i}\right)^{2}.\end{gathered}$$ Let us introduce a function $\beta(t)$ via the relation $$\beta(t)=\sum\limits_{i, j=1}^{N}\mu_{ij}\int\limits_{0}^{1}(\partial_{x}u_{i})(\partial_{x}u_{j})\, dx+$$ $$+\sum\limits_{i=1}^{N}\int\limits_{0}^{t}\int\limits_{0}^{1}\left(\rho_{i}(\partial_{t}u_{i})^{2}+ \frac{1}{\rho_{i}}\left(\sum\limits_{j=1}^{N}\mu_{ij}\partial_{xx}u_{j}\right)^{2}\right)\, dxd\tau.$$ Then and the inequalities , , and give the estimate[^3] $$\beta^{\prime}(t)\leqslant C_{15}+C_{16}\left(\sum\limits_{j=1}^{N}\\|u_{j}\\|^{2}_{L_{\infty}(0,1)}\right)\left(\sum\limits_{i, j=1}^{N}\mu_{ij} \int\limits_{0}^{1}(\partial_{x}u_{i})(\partial_{x}u_{j})\, dx\right)\leqslant$$ $$\leqslant C_{15}+C_{16}\left(\sum\limits_{j=1}^{N}\\|u_{j}\\|^{2}_{L_{\infty}(0,1)}\right)\beta(t),$$ where $C_{15}=C_{15}(C_{3}, C_{11}, C_{14}, K, N, \gamma)$, $C_{16}=C_{16}(C_{3}, \textbf{M})$, from which, via Gronwall’s lemma (see also ), it follows that $$\beta(t)\leqslant C_{17}\left(C_{1}, C_{15}, C_{16}, \{\\|u_{0i}^{\prime}\\|_{L_{2}(0, 1)}\}, \textbf{M}, T\right),$$ and we arrive at the conclusion of Lemma 11. [Remark 12.]{} It follows immediately from the continuity equations and the estimates in Lemmas 6 and 11 and Remark 10, that $$\\|\partial_{t}\rho_{i}\\|_{L_{\infty}(0, T; L_{2}(0,1))}\leqslant C_{18}(C_{3}, C_{12}, C_{13}),\quad i=1,\ldots,N.$$ Proof of Theorem 2 ================== Basing on the global a priori estimates proved in Sections 5–10, we can continue the local solution (obtained in Theorem 2 from [@mamprok.local]) into the entire $Q_{T}$ (see, e. g., [@mamprok.akm], P. 40, or [@mamprok.weigant92], P. 20). The uniqueness of this solution is shown in [@mamprok.local]. Thus, Theorem 2 is proved. [99]{} S. N. Antontsev, A. V. Kazhikhov and V. N. Monakhov, [Boundary value problems in mechanics of nonhomogeneous fluids]{}, Studies in Mathematics and its Applications, [22]{}, North-Holland Publishing Co., Amsterdam, 1990. I. V. Basov and V. V. Shelukhin, [On equations of a nonlinear compressible fluid with a discontinuous constitutive law]{}, Siberian Math. J., [40]{}:3 (1999), 435–445. V. N. Dorovsky and Yu. V. Perepechko, [Theory of the partial melting]{}, Sov. Geology and Geophysics, 9 (1989), 56–64 (in Russian). J. Frehse, S. Goj and J. Malek, [On a Stokes-like system for mixtures of fluids]{}, SIAM J. Math. Anal., [36]{}:4 (2005), 1259–1281. J. Frehse, S. Goj and J. Malek, [A uniqueness result for a model for mixtures in the absence of external forces and interaction momentum]{}, Appl. Math., [50]{}:6 (2005), 527–541. J. Frehse and W. Weigant, [On quasi-stationary models of mixtures of compressible fluids]{}, Appl. Math., [53]{}:4 (2008), 319–345. V. Giovangigli, [Multicomponent flow modeling]{}, Birkhäuser, Boston, 1999. A. V. Kazhikhov and V. V. Shelukhin, [Unique global solution with respect to time of initial-boundary value problems for one-dimensional equations of a viscous gas]{}, Journal of Applied Mathematics and Mechanics, [41]{}:2 (1977), 273–282. A. E. Mamontov and D. A. Prokudin, [Viscous compressible multi-fluids: modeling and multi-D existence]{}, Methods and applications of analysis, [20]{}:2 (2013), 179–195. A. E. Mamontov and D. A. Prokudin, [Solubility of a stationary boundary-value problem for the equations of motion of a one-temperature mixture of viscous compressible heat–conducting fluids]{}, Izvestiya: Mathematics, [78]{}:3 (2014), 554–579. A. E. Mamontov and D. A. Prokudin, [Solvability of the regularized steady problem of the spatial motions of multicomponent viscous compressible fluids]{}, Siberian Math. J., [57]{}:6 (2016), 1044–1054. A. E. Mamontov and D. A. Prokudin, [Solubility of initial boundary value problem for the equations of polytropic motion of multicomponent viscous compressible fluids]{}, Siberian Electr. Math. Reports, [13]{} (2016), 541–583 (in Russian). A. E. Mamontov and D. A. Prokudin, [Solubility of steady boundary value problem for the equations of polytropic motion of multicomponent viscous compressible fluids]{}, Siberian Electr. Math. Reports, [13]{} (2016), 664–693 (in Russian). A. E. Mamontov and D. A. Prokudin, [Existence of weak solutions to the three-dimensional problem of steady barotropic motions of mixtures of viscous compressible fluids]{}, Siberian Math. J., [58]{}:1 (2017), 113–127. A. E. Mamontov and D. A. Prokudin, [Viscous compressible homogeneous multi-fluids with multiple velocities: barotropic existence theory]{}, Siberian Electr. Math. Reports, [14]{} (2017), 388–397. A. E. Mamontov and D. A. Prokudin, [Modeling viscous compressible barotropic multi-fluid flows]{}, J. of physics: conference series, 894(2017)012058. A. E. Mamontov and D. A. Prokudin, [Solubility of unsteady equations of multi-component viscous compressible fluids]{}, Izvestiya: Mathematics, [82]{}:1 (2018), 140–185. A. E. Mamontov and D. A. Prokudin, [Local solvability of initial-boundary value problem for one-dimensional equations of polytropic flows of viscous compressible multifluids]{}, J. of Math. Sciences, [231]{}:2 (2018), 227–242. A. E. Mamontov and D. A. Prokudin, [Unique solvability of initial-boundary value problem for one-dimensional equations of polytropic flows of multicomponent viscous compressible fluids]{}, Siberian Electr. Math. Reports, [15]{} (2018), 631–649. P. B. Mucha, M. Pokorny and E. Zatorska, [Heat-conducting, compressible mixtures with multicomponent diffusion: Construction of a weak solution]{}, SIAM J. Math. Anal., [47]{}:5 (2015), 3747–3797. R. I. Nigmatulin, [Dynamics of multiphase media, Vol. 1]{}, Hemisphere, N.Y., 1990. D. A. Prokudin, [Unique solvability of initial-boundary value problem for a model system of equations for the polytropic motion of a mixture of viscous compressible fluids]{}, Siberian Electr. Math. Reports, [14]{} (2017), 568–585 (in Russian). D. A. Prokudin, [Global solvability of the initial boundary value problem for a model system of one-dimensional equations of polytropic flows of viscous compressible fluid mixtures]{}, J. of Physics: Conference Series, 894 (2017) 012076, 6 P. D. A. Prokudin and M. V. Krayushkina, [Solvability of a stationary boundary value problem for a model system of the equations of barotropic motion of a mixture of compressible viscous fluids]{}, Journal of Applied and Industrial Mathematics, [10]{}:3 (2016), 417–428. K. L. Rajagopal and L. Tao, [Mechanics of mixtures, Series on Advances in Mathematics for Applied Sciences]{}, [35]{}, World Scientific, River Edge, NJ, 1995. W. Weigant, [Non-homogeneous boundary value problems for the Navier—Stokes equations of a viscous gas]{}. Ph. D. Thesis. Barnaul. 1992 (in Russian). [^1]: Hereinafter, $C$ with indices denotes positive constants which depend on the initial data, physical constants and $T$. [^2]: The bound $C_{6}$ depends on the size of the flow domain. [^3]: Here the symmetry of the matrix $\textbf{M}$ is used.	{ "pile_set_name": "ArXiv" }
--- abstract: \| We study the behavior of matrix string theory in the strong coupling region, where it is expected to reduce to discrete light-cone type IIA superstring. In the large $N$ limit, the reduction corresponds to the double-dimensional reduction from wrapped supermembranes on $R^{10} \times S^1$ to type IIA superstrings on $R^{10}$ in the light-cone gauge, which is shown classically, however it is not obvious quantum mechanically. We analyze the problem in matrix string theory by using the strong coupling ($1/g$) expansion. We find that the quantum corrections do not cancel out at $\mathcal{O}(1/g^2)$. Detailed calculations can be seen in Ref.[@UY]. author: - \| [Shozo Uehara]{}[^1] and [Satoshi Yamada]{}[^2]\ [Department of Physics, Nagoya University]{},\ [Nagoya 464-8602, Japan]{} title: 'On the quantum matrix string[^3] ' --- DPNU-02-34\ hep-th/0210261\ October 2002 Introduction ============ Supermembrane in eleven dimensions [@HLP; @BST] plays an important role to understand the fundamental degrees of freedom in M-theory. At the classical level, it was shown that the supermembrane is related to type IIA superstring in ten dimensions by the double-dimensional reduction [@DHIS]. The procedure is the following: (i) Consider the target space of $R^{10}\times S^1$. (ii) Set the compactified coordinate (with radius $L$) proportional to one of the spatial coordinates of the world volume, which we call $\rho$ coordinate. (iii) Simply ignore the infinite tower of the Kaluza-Klein (non-zero) modes. However, it is not obvious whether such a reduction is justified also in quantum theory. Sekino and Yoneya analyzed the double-dimensional reduction quantum mechanically with the light-cone supermembrane action [@SY]. They kept the Kaluza-Klein modes associated with the $\rho$ coordinate in the wrapped supermembrane theory on the target space $R^{10}\times S^1$ and integrated them out by using the perturbative expansion with respect to the radius $L$. Since the gauge coupling satisfies $g\sim 1/L$ in the wrapped supermembrane theory, the expansion can be regarded as the strong coupling expansion. They calculated the effective action for the zero modes along the $\rho$ direction to the one-loop order of $O(L^2)$ and found that the quantum corrections cancel out and the effective action agrees with the classical (free) action of type IIA superstring except at the points where the usual string interactions could occur. As is emphasized in their paper [@SY], however, the strong coupling expansion does not give a rigorous proof of the quantum double-dimensional reduction because the propagators are proportional to the two-dimensional $\delta$-function, $\delta^{(2)}(\xi) \equiv \delta(\tau)\delta(\sigma)$, which will cause the ultraviolet divergences of $\delta^{(2)}(0)$ type in loops. However it is very difficult to find a suitable regularization which respects symmetries, and hence the strong coupling expansion is not yet defined rigorously. In this sense, they gave a formal argument for the vanishing of the one-loop corrections of $O(L^2)$ by demonstrating that the coefficients of $\delta^{(2)}(0)$ coming from both bosonic and fermionic degrees of freedom cancel out. The purpose of our work is essentially to extend their analysis to the two-loop order of $O(L^2)$. However, the naive extension is not straightforward because at the two-loop level, even the coefficients of the $\delta^{(2)}(0)$ diverge due to the contribution of the infinite Kaluza-Klein towers. Thus, we need another regularization for the summation over the infinite tower of the Kaluza-Klein modes. We know the matrix regularization of the supermembrane on $R^{11}$ in the light-cone gauge [@dHN] and also that of the wrapped supermembrane on $R^{10} \times S^1$ in the light-cone gauge [@SY]. The former is called Matrix theory [@BFSS] and the latter is called matrix string theory [@Mot; @DVV] which will be a non-perturbative formulation of light-cone quantized type IIA superstring theory in the large $N$ limit. Furthermore, even at finite $N$, Matrix and matrix string theories are conjectured to be non-perturbative formulations of discrete light-cone quantized (DLCQ) M-theory and type IIA superstring theory, respectively [@Sus; @Sei; @Sen]. Thus, we consider matrix string theory and study whether the reduction from matrix strings to discrete light-cone type IIA superstrings is justified quantum mechanically.[^4] From wrapped supermembrane to matrix string =========================================== The supermembrane action on the target space $R^{11}$ [@dHN] in the light-cone gauge is given by $$\begin{aligned} S&=&LT\int d\tau \!\int_0^{2\pi}\!d\sigma d\rho \Biggl[ \frac{1}{2}(D_{\tau}X^i)^2 -\frac{1}{4L^2}\{ X^i,X^j\}^2{\nonumber\\}&&\hspace{20ex} +i\psi^T D_{\tau}\psi +\frac{i}{L} \psi^T\gamma^i\{ X^i,\psi\}\Biggr]\label{action1},\\ &&D_{\tau} ={\partial_\tau}-\frac{1}{L}\{A,~~\},\quad \{A,B\} = \partial_{\sigma}A\,\partial_{\rho}B - \partial_{\rho}A\, \partial_{\sigma}B ,\end{aligned}$$ where the indices $i,j$ run through $1,2,\cdots,9$, the spinor $\psi$ has sixteen real components[^5] and $T$ is the membrane tension. At this stage, $L$ is an arbitrary length parameter of no physical meaning. The action is invariant under the gauge transformation, $$\delta A={\partial_\tau}\Lambda + \frac{1}{L}\{\Lambda,A\},\quad \delta X^i= \frac{1}{L}\{\Lambda,X^i\},\quad \delta \psi= \frac{1}{L}\{\Lambda,\psi\}.$$ This gauge transformation generates the area-preserving diffeomorphism on the world volume. When the spatial surface of the supermembrane has a non-trivial topology, we have to impose further the global constraints. Now we consider the wrapped supermembrane theory on the target space $R^{10}\times S^1$ and discuss the correspondence with matrix string [@SY]. We take the $X^9$ direction as $S^1$ and identify the radius with the above parameter $L$, $$\begin{aligned} X^9=L\rho +Y.\label{L}\end{aligned}$$ Thus $L$ has the physical meaning of the radius of the $X^9$ direction which is regarded as the “eleventh” direction in M-theory. Substituting eq.(\[L\]) into eq.(\[action1\]), we obtain the light-cone gauge supermembrane action on $R^{10}\times S^1$, $$\begin{aligned} S&=&LT\int d\tau \int_0^{2\pi}d\sigma d\rho \left[ \frac{1}{2}F_{\tau\sigma}^2 +\frac{1}{2}(D_{\tau}X^k)^2 -\frac{1}{2}(D_{\sigma}X^k)^2 -\frac{1}{4L^2}\{ X^k,X^l\}^2 \right.{\nonumber\\}&&\hspace{20ex} \left.+i\psi^T D_{\tau}\psi -i\psi^T \gamma^9 D_{\sigma}\psi +\frac{i}{L}\psi^T\gamma^k\{ X^k,\psi\} \right], \label{action}\\ &&F_{\tau\sigma}={\partial_\tau}Y - \partial_{\sigma}A - \frac{1}{L}\{A,Y\},\quad D_{\sigma}=\partial_{\sigma}-\frac{1}{L}\{Y,~~\},\end{aligned}$$ where the indices $k,l$ run through $1,2,\cdots,8$. This is also an action of the gauge theory of the area-preserving diffeomorphism, where the gauge coupling $g\sim 1/L$. In Ref.[@SY], the area-preserving diffeomorphism in eq.(\[action\]) was regularized by the finite dimensional group $U(N)$ and it was shown that the matrix-regularized form of the action (\[action\]) agrees with that of matrix string theory, $$\begin{aligned} &&S=LT\int d\tau \int_0^{2\pi}d\theta \,tr\left[ \frac{1}{2}F_{\tau\theta}^2 +\frac{1}{2}(D_{\tau}X^k)^2 -\frac{1}{2}(D_{\theta}X^k)^2 +\frac{1}{4L^2}[ X^k,X^l]^2 \right.{\nonumber\\}&&\hspace{20ex} \left.+i\psi^T D_{\tau}\psi -i\psi^T \gamma^9 D_{\theta}\psi -\frac{1}{L}\psi^T\gamma^k[X^k,\psi] \right],\label{action2}\\ &&F_{\tau\theta}={\partial_\tau}Y - {\partial_\theta}A - \frac{i}{L}[A,Y],~ D_{\tau}={\partial_\tau}-\frac{i}{L}[A,~~],~ D_{\theta}={\partial_\theta}-\frac{i}{L}[Y,~~]\,,\end{aligned}$$ where each element of the matrices is a function of ($\tau,\theta$). The action (\[action2\]) is invariant under the $U(N)$ gauge transformations, $$\begin{aligned} &&\delta A={\partial_\tau}\Lambda + \frac{i}{L}[\Lambda,A],\quad \delta Y={\partial_\theta}\Lambda + \frac{i}{L}[\Lambda,Y],{\nonumber\\}&&\delta X^k= \frac{i}{L}[\Lambda,X^k],\quad \delta \psi= \frac{i}{L}[\Lambda,\psi].\label{gauge}\end{aligned}$$ The zero-modes along the $\rho$ direction in the wrapped supermembrane are mapped to the diagonal elements of matrix string while the Kaluza-Klein modes are mapped to the off-diagonal elements [@SY]. Note that in the matrix regularization of the wrapped supermembrane on $R^{10} \times S^1$, there are no obvious counterparts of the global constraints, because the (matrix-regularized) Gauss law constraint, which is derived from eq.(\[action2\]), cannot be manifestly interpreted as the integrability condition. In the classical double-dimensional reduction, the Kaluza-Klein modes of every field along the $\rho$ direction are set zero. And then in the $L\to0$ limit, the action (\[action\]) reduces to the light-cone type IIA superstring action. As for the matrix-regularized action (\[action2\]), the off-diagonal elements of every matrix are set zero in such a classical double-dimensional reduction. Then the action reduces to the DLCQ type IIA superstring action in the light-cone momentum $p^+=N/R$ sector. It is expected that the reductions are justified also in quantum theory, however, it is not so simple [@Rus; @SY]. In particular, the quantum double-dimensional reduction of the wrapped supermembrane was analyzed for the small radius $L$ in [@SY], which corresponds to the strong gauge coupling $g\sim 1/L$ in the wrapped supermembrane theory and also to the weak string coupling $g_s\sim L/\sqrt{\alpha'}$ in type IIA superstring theory. By using the perturbative expansion with respect to $L$, the Kaluza-Klein modes along the $\rho$ direction were integrated out to the one-loop order of $O(L^2)$ and it was found that the effective action for the zero modes agrees with the classical (free) action of the type IIA superstring except at the interaction points of perturbative strings. So far the result is consistent with the expectation that the wrapped supermembrane theory in the region of small radius $L$ agrees with the perturbative type IIA superstring theory. Then we analyze the quantum reduction of the matrix string (\[action2\]) to the diagonal elements for small radius $L$ to study whether the effective action for the diagonal matrix elements agrees with the classical (free) action of the DLCQ type IIA superstring. Strong coupling expansion in matrix string theory\[SCEinMST\] ============================================================= First every $N\times N$ hermite matrix in eq.(\[action2\]) is decomposed into the diagonal and off-diagonal parts, $$A\to a+A,\quad Y\to y+Y,\quad X^k\to x^k+X^k, \quad \psi\to \psi+\Psi,\label{decomp}$$ where $a,y,x^k$ and $\psi$ are the diagonal and $A,Y,X^k$ and $\Psi$ are the off-diagonal parts of the original matrices, respectively, which are plugged into (\[action2\]). The gauge transformations are also decomposed as $$\delta a={\partial_\tau}\lambda +\frac{i}{L}[\Lambda,A]_{{\rm diag}},\quad \delta A={\partial_\tau}\Lambda + \frac{i}{L}([\lambda,A]+[\Lambda,a] +[\Lambda,A]_{{\mbox{\scriptsize off-diag}}}),~\cdots$$ where $\lambda$ and $\Lambda$ are the diagonal and off-diagonal parts of the gauge parameter, respectively. At this stage, we impose a boundary condition in $\theta$-direction. Here, for simplicity, we choose such boundary conditions as to have the $N$ string bits having $p^+=1/R$ for the diagonal matrix elements, $$\phi_a(\theta +2\pi)=\phi_a(\theta).\quad (\mbox{for } \phi=a,y,x^k,\psi)$$ As for the off-diagonal matrix elements, we naturally impose $$\Phi_{ab}(\theta +2\pi)=\Phi_{ab}(\theta).\quad (\mbox{for } \Phi=A, Y, X^k, \Psi)$$ Next we fix the gauge. We choose the following gauge condition, $$a=y,\quad{\partial_\theta}Y-\frac{i}{L}[y,Y] -\frac{i}{L}[x^k,X^k]+\frac{i}{L}[a,A] -{\partial_\tau}A=0,\label{fixing}$$ and we proceed in the Landau gauge. Then we get $$\begin{aligned} T&=&\int {\mathcal{D}}y {\mathcal{D}}x^k {\mathcal{D}}\psi {\mathcal{D}}c {\mathcal{D}}\bar{c} {\mathcal{D}}A {\mathcal{D}}Y {\mathcal{D}}X^k {\mathcal{D}}\Psi {\mathcal{D}}C {\mathcal{D}}\bar{C} {\mathcal{D}}B{\nonumber\\}&&\hspace{2ex}\exp\Biggl[iLT \int d\tau \int_0^{2\pi} d\theta\, ({\mathcal{L}}^{string}+ {\mathcal{L}}^B_0 +L{\mathcal{L}}^B_1+L^2{\mathcal{L}}^B_2{\nonumber\\}&&\hspace{23ex}+{\mathcal{L}}^F_0 +L^{1/2}{\mathcal{L}}^F_{1/2} +L{\mathcal{L}}^F_1)\Biggr],\label{PI}\\ {\mathcal{L}}^{string}&=&{\mathop{\mathrm{Tr}}}\bigg[\frac{1}{2}({\partial_\tau}x^k)^2 - \frac{1}{2}({\partial_\theta}x^k)^2 +\frac{1}{2}\left\{({\partial_\tau}-{\partial_\theta})y\right\}^2{\nonumber\\}&&\hspace{2cm}+i\bar{c}({\partial_\tau}-{\partial_\theta})c + i\psi^T{\partial_\tau}\psi-i \psi^T\gamma^9{\partial_\theta}\psi \bigg],\\ {\mathcal{L}}^B_0&=&{\mathop{\mathrm{Tr}}}\bigg[-\frac{1}{2}\,[x^k,A]^2+\frac{1}{2}\,[x^k,Y]^2 +\frac{1}{2}\,[x^k,X^l]^2{\nonumber\\}&&\hspace{1ex}-\frac{1}{2}\left([y,Y]+[x^k,X^k]-[y,A]\right)^2{\nonumber\\}&&\hspace{1ex}-iB \left([y,Y]+[x^k,X^k]-[y,A]\right) +i[x^k,\bar{C}][x^k,C]\bigg],\label{freeb}\\ {\mathcal{L}}^F_0&=&{\mathop{\mathrm{Tr}}}\left[\Psi^T[y,\Psi]-\Psi^T\gamma^9[y,\Psi]- \Psi^T\gamma^k[x^k,\Psi]\,\right],\label{freef}\\ {\mathcal{L}}^F_{1/2}&=&{\mathop{\mathrm{Tr}}}\left[-2\Psi^T[\psi,A]+2\Psi^T\gamma^9[\psi,Y] +2\Psi^T\gamma^k[\psi,X^k]\,\right],\label{1f}\\ {\mathcal{L}}^B_1&=&{\mathop{\mathrm{Tr}}}\bigg[-i{\partial_\tau}Y [y,Y]+2i{\partial_\tau}Y [y,A] -i{\partial_\tau}A [y,Y]{\nonumber\\}&&\hspace{4ex}+ 2i{\partial_\theta}A [y,Y]- i{\partial_\theta}A[y,A]-i{\partial_\theta}Y[y,A]{\nonumber\\}&&\hspace{4ex}-i{\partial_\tau}X^k[y,X^k]+2i{\partial_\tau}X^k[x^k,A] -i{\partial_\tau}A[x^k,X^k]{\nonumber\\}&&\hspace{4ex}+i{\partial_\theta}X^k[y,X^k]-2i{\partial_\theta}X^k[x^k,Y] +i{\partial_\theta}Y[x^k,X^k]{\nonumber\\}&&\hspace{4ex}-[y,Y][A,Y]+[y,A][A,Y]-[y,X^k][A,X^k]{\nonumber\\}&&\hspace{4ex}+[x^k,A][A,X^k]+[y,X^k][Y,X^k]-[x^k,Y][Y,X^k]{\nonumber\\}&&\hspace{4ex}+[x^k,X^l][X^k,X^l]+B{\partial_\theta}Y-B{\partial_\tau}A -{\partial_\theta}\bar{C}[y,C]{\nonumber\\}&&\hspace{4ex}-[y,\bar{C}]{\partial_\theta}C +i[y,\bar{C}][Y,C] +i[x^k,\bar{C}][X^k,C]{\nonumber\\}&&\hspace{4ex}+ [y,\bar{C}]\,{\partial_\tau}C -i[y,\bar{C}][A,C] +{\partial_\tau}\bar{C}[y,C]\bigg],\label{1b}\\ {\mathcal{L}}^F_{1}&=&{\mathop{\mathrm{Tr}}}\Big[\,i\Psi^T{\partial_\tau}\Psi -i\Psi^T\gamma^9{\partial_\theta}\Psi{\nonumber\\}&&\hspace{4ex}+\Psi^T[A,\Psi]-\Psi^T\gamma^9[Y,\Psi] -\Psi^T\gamma^k[X^k,\Psi]\,\Big],\label{2f}\\ {\mathcal{L}}^B_2&=&{\mathop{\mathrm{Tr}}}\bigg[\,\frac{1}{2}({\partial_\tau}Y - {\partial_\theta}A)^2 +\frac{1}{2}({\partial_\tau}X^k)^2 -\frac{1}{2}({\partial_\theta}X^k)^2{\nonumber\\}&&\hspace{3ex}-i{\partial_\tau}Y[A,Y] +i{\partial_\theta}A[A,Y] -i{\partial_\tau}X^k[A,X^k]{\nonumber\\}&&\hspace{3ex} +i{\partial_\theta}X^k[Y,X^k] -\frac{1}{2}\,[A,Y]^2 -\frac{1}{2}\,[A,X^k]^2{\nonumber\\}&&\hspace{3ex}+ \frac{1}{2}\,[Y,X^k]^2+ \frac{1}{4}\,[X^k,X^l]^2 -i{\partial_\theta}\bar{C}{\partial_\theta}C -{\partial_\theta}\bar{C}[Y,C]{\nonumber\\}&&\hspace{3ex} +i{\partial_\tau}\bar{C}{\partial_\tau}C +{\partial_\tau}\bar{C}[A,C] +i[\bar{C},A]_{{\rm diag}}[C,A]_{{\rm diag}}{\nonumber\\}&&\hspace{3ex} -i[\bar{C},Y]_{{\rm diag}}[C,Y]_{{\rm diag}} -i[\bar{C},X^k]_{{\rm diag}}[C,X^k]_{{\rm diag}}\bigg],\label{2b}\end{aligned}$$ where ($c$,$\bar{c}$) are (ghost, anti-ghost) for the first condition of (\[fixing\]), while ($C$,$\bar{C}$,$B$) are (ghost, anti-ghost, $B$-field) for the second one, respectively, $a$ has been integrated out by using the Landau gauge condition for eq.(\[fixing\]) and some of the off-diagonal parts have been rescaled as [@SY] $$A\to LA, \quad Y\to LY,\quad X^k\to LX^k,\quad\Psi\to L^{1/2}\Psi, \quad C\to L^2C.$$ By using the above action, we perform the perturbative expansion with respect to $L$ and integrate only the off-diagonal matrix elements; $$\begin{aligned} &&T=\int {\mathcal{D}}y {\mathcal{D}}x^k {\mathcal{D}}\psi {\mathcal{D}}c {\mathcal{D}}\bar{c}\,\, \exp\left(iS_{eff}[y,\,x^k,\,\bar{c},\,c,\,\psi]\right),\\ &&S_{eff}[y,\,x^k,\,\bar{c},\,c,\,\psi]=\int d\tau \int_0^{2\pi} d\theta\, \left({\mathcal{L}}^{string} -i\ln Z[y,\,x^k,\,\psi]\right),\label{eff}\\ &&Z[y,\,x^k,\,\psi]=\int {\mathcal{D}}A {\mathcal{D}}Y {\mathcal{D}}X^k {\mathcal{D}}\Psi{\mathcal{D}}C {\mathcal{D}}\bar{C}{\mathcal{D}}B\,\exp \,(i\tilde{S})\,,\label{Z}\\ &&\tilde{S}= \int d\tau \int_0^{2\pi} d\theta\, \Big({\mathcal{L}}^B_0 +L{\mathcal{L}}^B_1 + L^2{\mathcal{L}}^B_2 +{\mathcal{L}}^F_0 +L^{\frac{1}{2}}{\mathcal{L}}^F_{1/2} +L{\mathcal{L}}^F_1\Big),\end{aligned}$$ where we have set $LT=1$ and $\xi=(\tau,\theta)$ for brevity. We regard (\[freeb\]) and (\[freef\]) as the free parts and (\[1f\])–(\[2b\]) the interactions. Then we read off the propagators from the free parts, $$\begin{aligned} {\langle \hat{X}_{ab}^K(\xi)\hat{X}_{ba}^L(\xi') \rangle} &=&-i\left(\delta^{KL}-\frac{(\hat{x}_a^K -\hat{x}_b^K)(\hat{x}_a^L-\hat{x}_b^L)}{(\hat{x}_a-\hat{x}_b)^2} \right)\frac{G(\xi,\xi')}{(\hat{x}_a-\hat{x}_b)^2}\,,\label{pro}\\ {\langle B_{ab}(\xi)Y_{ba}(\xi') \rangle} &=&{\langle B_{ab}(\xi)A_{ba}(\xi') \rangle} = \frac{y_a-y_b}{(x_a-x_b)^2}\,G(\xi,\xi'),\\ {\langle B_{ab}(\xi)X_{ba}^k(\xi') \rangle}&=& \frac{x_a^k-x_b^k}{(x_a-x_b)^2}\,G(\xi,\xi'),\\ {\langle \bar{C}_{ab}(\xi)C_{ba}(\xi') \rangle}&=& \frac{G(\xi,\xi')}{(x_a-x_b)^2},\\ {\langle \Psi_{ab}^{\alpha}(\xi)\Psi_{ba}^{\beta}(\xi') \rangle}&=& -\frac{i}{2} \frac{(y_a-y_b)(I+\gamma^9)_{\alpha\beta} +(x_a^k-x_b^k)\gamma^k_{\alpha\beta}}{(x_a-x_b)^2}G(\xi,\xi')\, \label{psipsi},\end{aligned}$$ where $G(\xi,\xi')\equiv \delta^{(2)}(\xi-\xi')$, $(x_a-x_b)^2\equiv(x_{a}^k-x_b^k)(x_{a}^k-x_b^k)$, the spinor indices $\alpha,\beta$ run through $1,2,\cdots,16$ and we have introduced the hatted variables $\hat{X}^K$ and $\hat{x}^K$ $(K=k,9,10\,\ (k = 1,2,\cdots,8))$, $$\hat{X}^k=X^k,\quad\hat{X}^9=Y,\quad\hat{X}^{10}=iA,\quad \hat{x}^k=x^k,\quad\hat{x}^9=y,\quad\hat{x}^{10}=iy,\label{hatX}$$ and $(\hat{x}_a-\hat{x}_b)^2\equiv(\hat{x}_a^K-\hat{x}_b^K) (\hat{x}_a^K-\hat{x}_b^K)=(x_a-x_b)^2$. Notice that $(x_{a}-x_b)^2$ (for $a\ne b$, $1\leq a,b\leq N$) must be non-zero in order that the perturbative expansion makes sense since the propagators are singular at $(x_a-x_b)^2=0$. We recall that in matrix string theory, the usual string interactions are described by the exchange of coincident diagonal matrix elements and hence the perturbative expansion does not make sense even for small radius $L$ at the interaction points. Thus, henceforth we ignore the interaction points and integrate out the off-diagonal matrix elements to get the effective action for the diagonal matrix elements, which is expected to agree with the classical (free) action of DLCQ type IIA superstring. The perturbative calculation is, however, formal as in Ref.[@SY]: The propagators (\[pro\])-(\[psipsi\]) are proportional to the $\delta$-function $G(\xi,\xi')=\delta^{(2)}(\xi-\xi')$ and the loops suffer from the ultraviolet divergences like $\delta^{(2)}(0)$. However, it is very difficult to find a suitable regularization which respects symmetries. If we adopt a certain regularization, e.g. cutoff regularization for large momenta, the regularized $\delta$-function $G_{(r)}(\xi,\xi')$ would not satisfy $f(\xi)G_{(r)}(\xi,\xi')=f(\xi')G_{(r)}(\xi,\xi')$, which causes an ambiguity of how we choose the arguments of $(x_a^k-x_b^k)$ and $(y_a-y_b)$, which appear in the propagators (\[pro\])-(\[psipsi\]). To avoid the ambiguity, henceforth we consider only the configurations of the diagonal matrix elements in which the differences of arbitrary two elements $(x_a^k-x_b^k)$, $(y_a-y_b)$ and $(\psi_a-\psi_b)$ are independent of $\xi$, although $x_a^k$, $x_b^k$, $y_a$, $y_b$, $\psi_a$ and $\psi_b$ themselves depend on $\xi$, in general. We have not yet found such a suitable regularization, however, we give a formal argument about the quantum corrections order by order in the strong-coupling expansion. \(i) $O(L^{0})$: The lowest order contribution in eq.(\[Z\]) is the one-loop determinant of the free action. Actually, the determinant is unity due to the coincidence between bosonic and fermionic degrees of freedoms. \(ii) $O(L^{1/2})$: The next contribution in eq.(\[Z\]) comes from $\tilde{S}_{1/2}^F=\int {\mathcal{L}}_{1/2}^F$. ${\langle i\tilde{S}_{1/2}^F \rangle}$ vanishes because there is no way to self-contract in $i\tilde{S}_{1/2}^F$. \(iii) $O(L^{1})$: The $O(L^{1})$ contributions in eq.(\[Z\]) come from $\tilde{S}_{1}^B$, $\tilde{S}_{1/2}^F$ and $\tilde{S}_{1}^F$. There are tree kinds of contributions, ${\langle i\tilde{S}_{1}^B \rangle}$, $(1/2!) {\langle i\tilde{S}_{1/2}^F \,i\tilde{S}_{1/2}^F \rangle}$ and ${\langle i\tilde{S}_{1}^F \rangle}$. The first one is given by $$\begin{aligned} {\langle i\tilde{S}_{1}^B \rangle} &=&iL\int d^2 \xi d^2 \xi' \sum_{a,b=1}^N\bigg\{i(y_a-y_b){\partial_\tau}{\langle Y_{ab}(\xi)Y_{ba}(\xi') \rangle}{\nonumber\\}&&\hspace{7ex}-i(y_a-y_b){\partial_\tau}{\langle Y_{ab}(\xi)A_{ba}(\xi') \rangle} -i(y_a-y_b){\partial_\theta}{\langle A_{ab}(\xi)Y_{ba}(\xi') \rangle}{\nonumber\\}&&\hspace{7ex}+i(y_a-y_b){\partial_\theta}{\langle A_{ab}(\xi)A_{ba}(\xi') \rangle} +i(y_a-y_b){\partial_\tau}{\langle X_{ab}^k(\xi)X_{ba}^k(\xi') \rangle}{\nonumber\\}&&\hspace{7ex}-i(x_a^k-x_b^k){\partial_\tau}{\langle X_{ab}^k(\xi)A_{ba}(\xi') \rangle} -i(y_a-y_b){\partial_\theta}{\langle X_{ab}^k(\xi)X_{ba}^k(\xi') \rangle}{\nonumber\\}&&\hspace{7ex}+i(x_a^k-x_b^k){\partial_\theta}{\langle X_{ab}^k(\xi)Y_{ba}(\xi') \rangle} -{\partial_\theta}{\langle B_{ab}(\xi)Y_{ba}(\xi') \rangle}{\nonumber\\}&&\hspace{7ex}+{\partial_\tau}{\langle B_{ab}(\xi)A_{ba}(\xi') \rangle} +2(y_a-y_b)\,{\partial_\theta}{\langle \bar{C}_{ab}(\xi)C_{ba}(\xi') \rangle}{\nonumber\\}&&\hspace{18ex}-2(y_a-y_b)\,{\partial_\tau}{\langle \bar{C}_{ab}(\xi)C_{ba}(\xi') \rangle}\bigg\} \,\delta^{(2)}(\xi-\xi').\end{aligned}$$ From eqs.(\[pro\])-(\[psipsi\]), we see that the quantity in the braces is antisymmetric in $a$ and $b$ and hence ${\langle i\tilde{S}_{1}^B \rangle}$ is zero. Similarly we find that both ${\langle i\tilde{S}_{1/2}^F\,i\tilde{S}_{1/2}^F \rangle}$ and ${\langle i\tilde{S}_{1}^F \rangle}$ are zero. Note that in order to show that the quantum correction of $O(L)$ is zero, we have never used the fact that $G(\xi,\xi')$ is a $\delta$-function. Hence it would hold even if $G(\xi,\xi')$ is some regularized $\delta$-function. \(iv) $O(L^{3/2})$: There are tree kinds of contributions, $$\frac{1}{2!}{\langle i\tilde{S}_{1}^B\,i\tilde{S}_{1/2}^F \rangle},\quad \frac{1}{3!}{\langle i\tilde{S}_{1/2}^F\,i\tilde{S}_{1/2}^F\, i\tilde{S}_{1/2}^F \rangle},\quad \frac{1}{2!}{\langle i\tilde{S}_{1}^F\,i\tilde{S}_{1/2}^F \rangle}.$$ Each of them is zero because there is no way of contraction, respectively. \(v) $O(L^{2})$: The contributions are $$\begin{aligned} &&{\langle i\tilde{S}_{2}^B \rangle},\quad \frac{1}{2!}{\langle i\tilde{S}_{1}^B i\tilde{S}_{1}^B \rangle},\quad \frac{1}{2!}{\langle i\tilde{S}_{1}^F \,i\tilde{S}_{1}^F \rangle},\quad \frac{1}{2!}{\langle i\tilde{S}_1^B i\tilde{S}_1^F \rangle},\\ &&\frac{1}{3!}{\langle i\tilde{S}_{1/2}^F i\tilde{S}_{1/2}^F i\tilde{S}_{1}^B \rangle},\quad \frac{1}{3!}{\langle i\tilde{S}_{1/2}^F i\tilde{S}_{1/2}^F i\tilde{S}_{1}^F \rangle},\quad \frac{1}{4!}{\langle i\tilde{S}_{1/2}^F i\tilde{S}_{1/2}^F i\tilde{S}_{1/2}^F i\tilde{S}_{1/2}^F \rangle}.\end{aligned}$$ The last three terms contain fermionic diagonal elements $\psi_a$ and they each vanish due to the anti-commutativity of the Grassmann variables $\psi_a$. Also it is easy to see that there is no contribution from $(1/2!){\langle i\tilde{S}_1^B i\tilde{S}_1^F \rangle}$. Then $(1/2!){\langle i\tilde{S}_{1}^B i\tilde{S}_{1}^B \rangle}$, ${\langle i\tilde{S}_{2}^B \rangle}$ and $(1/2!){\langle i\tilde{S}_{1}^F \,i\tilde{S}_{1}^F \rangle}$ are to be considered below. (v-1) One-loop: The one-loop contributions coming from $(1/2!){\langle i\tilde{S}_{1}^B i\tilde{S}_{1}^B \rangle}$, ${\langle i\tilde{S}_{2}^B \rangle}$ and $(1/2!){\langle i\tilde{S}_{1}^F \,i\tilde{S}_{1}^F \rangle}$ are referred to as $(1/2!){\langle i\tilde{S}_{1}^B i\tilde{S}_{1}^B \rangle}^{(1)}$, ${\langle i\tilde{S}_{2}^B \rangle}^{(1)}$ and $(1/2!){\langle i\tilde{S}_{1}^F \,i\tilde{S}_{1}^F \rangle}^{(1)}$, respectively. They are given by $$\begin{aligned} &&\frac{1}{2!}\,{\langle i\tilde{S}_{1}^B i\tilde{S}_{1}^B \rangle}^{(1)}=L^2\int d^2 \xi d^2 \xi' \sum_{a\ne b}\Biggl[-\frac{17(y_a-y_b)^2}{\{(x_a-x_b)^2\}^2} {\partial_\tau}{\partial_{\theta'}}G(\xi,\xi')\,G(\xi,\xi'){\nonumber\\}&&\hspace{6ex} +\left( \frac{1}{(x_a-x_b)^2}+\frac{17}{2}\,\frac{(y_a-y_b)^2} {\{(x_a-x_b)^2\}^2}\right) {\partial_\tau}{\partial_{\tau'}}G(\xi,\xi')\,G(\xi,\xi'){\nonumber\\}&&\hspace{6ex}-\left( \frac{1}{(x_a-x_b)^2}-\frac{17}{2}\, \frac{(y_a-y_b)^2}{\{(x_a-x_b)^2\}^2}\right) {\partial_\theta}{\partial_{\theta'}}G(\xi,\xi')\,G(\xi,\xi')\Biggr],\label{1-loop1}\\ &&{\langle i\tilde{S}_{2}^B \rangle}^{(1)}= L^2\int d^2 \xi d^2 \xi'\sum_{a\ne b}\Bigg[\frac{(y_a-y_b)^2}{\{(x_a-x_b)^2\}^2}\, {\partial_\tau}{\partial_{\theta'}}G(\xi,\xi')\,\delta^{(2)}(\xi-\xi'){\nonumber\\}&&\hspace{6ex}\left(\frac{3}{(x_a-x_b)^2} -\frac{1}{2}\frac{(y_a-y_b)^2}{\{(x_a-x_b)^2\}^2}\right) {\partial_\tau}{\partial_{\tau'}}G(\xi,\xi')\,\delta^{(2)}(\xi-\xi'){\nonumber\\}&&\hspace{6ex} -\left(\frac{3}{(x_a-x_b)^2} +\frac{1}{2}\frac{(y_a-y_b)^2}{\{(x_a-x_b)^2\}^2}\right) {\partial_\theta}{\partial_{\theta'}}G(\xi,\xi')\,\delta^{(2)}(\xi-\xi')\Bigg], \label{1-loop2}\\ &&\frac{1}{2!}{\langle i\tilde{S}_{1}^F i\tilde{S}_{1}^F \rangle}^{(1)} =L^2\int d^2 \xi d^2 \xi' \sum_{a \ne b}\Bigg[\frac{16(y_a-y_b)^2}{\{(x_a-x_b)^2\}^2}\, {\partial_\tau}{\partial_{\theta'}}G(\xi,\xi')\,G(\xi,\xi'){\nonumber\\}&&\hspace{6ex}+\left(\frac{-4}{(x_a-x_b)^2} -\frac{8(y_a-y_b)^2}{\{(x_a-x_b)^2\}^2}\right) {\partial_\tau}{\partial_{\tau'}}G(\xi,\xi')\,G(\xi,\xi'){\nonumber\\}&&\hspace{6ex}+\left(\frac{4}{(x_a-x_b)^2} -\frac{8(y_a-y_b)^2}{\{(x_a-x_b)^2\}^2}\right) {\partial_\theta}{\partial_{\theta'}}G(\xi,\xi')\,G(\xi,\xi')\Bigg]\label{1-loop3}.\end{aligned}$$ Note that we have never used the fact that $G(\xi,\xi')$ is the $\delta$-function in deriving eqs.(\[1-loop1\])-(\[1-loop3\]). Thus they are expected to be unaltered even if we adopt a certain regularization. At this stage we first use the fact that $G(\xi,\xi')$ is the $\delta$-function and it is shown that they are canceled, [^6] $$\frac{1}{2!}{\langle i\tilde{S}_{1}^B i\tilde{S}_{1}^B \rangle}^{(1)} +{\langle i\tilde{S}_{2}^B \rangle}^{(1)} + \frac{1}{2!}{\langle i\tilde{S}_{1}^F\,i\tilde{S}_{1}^F \rangle}^{(1)}=0.$$ (v-2) Two-loop: The two-loop contributions, which are not calculated in Ref.[@SY], are referred to as $(1/2!){\langle i\tilde{S}_{1}^B i\tilde{S}_{1}^B \rangle}^{(2)}$, ${\langle i\tilde{S}_{2}^B \rangle}^{(2)}$ and $(1/2!) {\langle i\tilde{S}_{1}^F \,i\tilde{S}_{1}^F \rangle}^{(2)}$, respectively. We obtain $$\begin{aligned} &&\frac{1}{2!}{\langle i\tilde{S}_{1}^B i\tilde{S}_{1}^B \rangle}^{(2)} =-i L^2\int d^2 \xi d^2 \xi' \sum_{a\ne b,\,b\ne c,\,c\ne a}\Biggl\{\frac{33}{2} \frac{1}{(x_a-x_b)^2(x_b-x_c)^2}{\nonumber\\}&&\hspace{22ex}-16\,\frac{\{(x_a^k-x_b^k)(x_c^k-x_a^k)\}^2} {(x_a-x_b)^2(x_b-x_c)^2\{(x_c-x_a)^2\}^2}{\nonumber\\}&&\hspace{22ex}-\frac{1}{2}\, \frac{\{(x_a^k-x_b^k)(x_b^k-x_c^k)\}^2}{\{(x_a-x_b)^2(x_b-x_c)^2\}^2} \Biggr\}\,(G(\xi,\xi'))^3,\label{1'}\\ &&{\langle i\tilde{S}_{2}^B \rangle}^{(2)}=iL^2\int d^2 \xi d^2\xi' \Biggl[\sum_{a\ne b,\, b\ne c,\,c\ne a} \left\{\frac{73}{2}\frac{1}{(x_a-x_b)^2(x_b-x_c)^2}\right.{\nonumber\\}&&\hspace{18ex}-\frac{1}{2}\frac{\{(x_a^k-x_b^k)(x_b^k-x_c^k)\}^2} {\{(x_a-x_b)^2(x_b-x_c)^2\}^2}\Biggr\} +54 \sum_{a\ne b} \frac{1}{\{(x_a-x_b)^2\}^2}\Biggr]{\nonumber\\}&&\hspace{20ex} \times (G(\xi,\xi'))^2\,\delta^{(2)}(\xi-\xi'),\label{2'}\\ &&\frac{1}{2!} {\langle i\tilde{S}_{1}^F \,i\tilde{S}_{1}^F \rangle}^{(2)} =-iL^2\int d^2 \xi d^2 \xi' \sum_{a\ne b,\,b\ne c,\,c\ne a} \bigg\{20\,\frac{1}{(x_a-x_b)^2(x_b-x_c)^2}{\nonumber\\}&&\hspace{15ex}+16\,\frac{\{(x_a^k-x_b^k)(x_c^k-x_a^k)\}^2} {(x_a-x_b)^2(x_b-x_c)^2\{(x_c-x_a)^2\}^2} \bigg\}(G(\xi,\xi'))^3. \label{3'}\end{aligned}$$ Note that in calculating eqs.(\[1’\])–(\[3’\]) we have never used the fact that $G(\xi,\xi')$ is the $\delta$-function and hence eqs.(\[1’\])–(\[3’\]) are expected to be unaltered even if we adopt a certain regularization. At this stage we first use the fact that $G(\xi,\xi')$ is the $\delta$-function and sum up eqs.(\[1’\])–(\[3’\]) to get $$\begin{aligned} &&\frac{1}{2!}{\langle i\tilde{S}_{1}^B i\tilde{S}_{1}^B \rangle}^{(2)}+{\langle i\tilde{S}_{2}^B \rangle}^{(2)} +\frac{1}{2!} {\langle i\tilde{S}_{1}^F \,i\tilde{S}_{1}^F \rangle}^{(2)}{\nonumber\\}&&\hspace{1ex}=iL^2\int d^2 \xi d^2\xi'\sum_{a\ne b} \frac{54}{\{(x_a-x_b)^2\}^2}\,(G(\xi,\xi'))^3.\label{eq:2Lsum}\end{aligned}$$ The two-loop quantum corrections at $O(L^2)$ do not cancel out! One comment is in order: The remaining term is exactly that of the second summation in eq.(\[2’\]). If we assume that the differences of the diagonal elements can be estimated as $(x_a^k-x_b^k)\sim O(N^{\alpha})$[^7] with some common constant $\alpha$ for large $N$, we will see that the terms canceled in eq.(\[eq:2Lsum\]) behave as $\sum_{a\ne b,\,b\ne c,\,c\ne a} (x_a^k-x_b^k)^{-2} (x_b^k-x_c^k)^{-2} \sim O(N^{3-4\alpha})$, while the remaining term behaves as $\sum_{a\ne b} (x_a^k-x_b^k)^{-4}\sim O(N^{2-4\alpha})$. In this sense, we could say that only the leading terms in the large $N$ are canceled in the two-loop quantum corrections to the classical string action at $O(L^2)$. Conclusion ========== We have studied in matrix string theory whether the reduction to the diagonal elements of the matrices is justified quantum mechanically. We have seen that the quantum corrections do not cancel out at $O(L^2)$. We should note, however, that no suitable regularization for the divergences of $\delta^{(2)}(0)$ type is found so far, and hence we have only studied a mechanism of cancellation of quantum corrections. In fact, we have found that at the two-loop level of $O(L^2)$, the sub-leading term in the large $N$ comes only from the bosonic degrees of freedom and cannot be canceled out. Even if we find a suitable regularization, such a structure seems to be unaltered and hence our result will be unchanged. Finally, we comment on the global constraints in the wrapped supermembrane theory. Such constraints should be taken into account in the calculations of the quantum double-dimensional reduction.[^8] In matrix string theory, however, there are no counterparts of such constraints. In particular, in the standard derivation of matrix string theory, they do not appear naturally. However we can show that the global constraint does not alter our result [@UY2]. Acknowledgments. One of the authors (SU) would like to thank the organizers of the “Third International Sakharov Conference on Physics” for the kind invitation. The work of SU is supported in part by the Grant-in-Aid for Scientific Research No.13135212. [99]{} S. Uehara and S. Yamada, JHEP [0209]{}, 019 (2002)\ \[arXiv:hep-th/0207209\]. J. Hughes, J. Liu and J. Polchinski, Phys. Lett. B [180]{}, 370 (1986). E. Bergshoeff, E. Sezgin and P. K. Townsend, Phys. Lett. B [189]{}, 75 (1987). M. J. Duff, P. S. Howe, T. Inami and K. S. Stelle, Phys. Lett. B [191]{}, 70 (1987). J. G. Russo, Nucl. Phys. B [492]{}, 205 (1997) \[arXiv:hep-th/9610018\]. Y. Sekino and T. Yoneya, Nucl. Phys. B [619]{}, 22 (2001) \[arXiv:hep-th/0108176\]. B. de Wit, J. Hoppe and H. Nicolai, Nucl. Phys. B [305]{}, 545 (1988). T. Banks, W. Fischler, S. H. Shenker and L. Susskind, Phys. Rev. D [55]{}, 5112 (1997) \[arXiv:hep-th/9610043\]. L. Motl, arXiv:hep-th/9701025. R. Dijkgraaf, E. Verlinde and H. Verlinde, Nucl. Phys. B [500]{}, 43 (1997) \[arXiv:hep-th/9703030\]. L. Susskind, arXiv:hep-th/9704080. N. Seiberg, Phys. Rev. Lett. [79]{}, 3577 (1997) \[arXiv:hep-th/9710009\]. A. Sen, Adv. Theor. Math. Phys. [2]{}, 51 (1998) \[arXiv:hep-th/9709220\]. T. Kugo and S. Uehara, Nucl. Phys. B [197]{}, 378 (1982). S. Uehara and S. Yamada, in preparation. [^1]: e-mail: [email protected] [^2]: e-mail: [email protected] [^3]: Presented at the 3rd International Sakharov Conference on Physics, Moscow, June 2002. [^4]: We use a convention of the light-cone coordinates such that $x^{\pm}=(x^{0}\pm x^{10})/\sqrt 2$. Furthermore, $x^-$ is compactified on $S^1$ with radius $R$ in DLCQ. [^5]: We use the real and symmetric gamma matrices $\gamma^i$, which satisfy $\{\gamma^i,\gamma^j\}=2\delta^{ij}$. [^6]: This result in matrix string theory is essentially the same as the one in the wrapped supermembrane theory [@SY]. In Ref.[@SY], however, the zero-mode gauge field $a$ is restricted to be zero by hand, while we have just fixed the gauge ($a=y$) and added the corresponding FP-ghost part following the standard procedure [@KU]. In this sense the configuration of $a$ is not restricted in our calculations. [^7]: According to the correspondence of a long string in matrix string theory with the wrapped supermembrane, $\alpha=-1$ for $\|a-b\|\ll N$ [@SY]. [^8]: The global constraints are not considered in the calculations [@SY].	{ "pile_set_name": "ArXiv" }
--- author: - - - - bibliography: - 'main.bib' title: Inferring Concise Specifications of APIs --- specification inference, postconditions	{ "pile_set_name": "ArXiv" }
--- abstract: \| This paper studies some analytical properties of weak solutions of 3D stochastic primitive equations with periodic boundary conditions. The martingale problem associated to this model is shown to have a family of solutions satisfying the Markov property, which is achieved by means of an abstract selection principle. The Markov property is crucial to extend the regularity of the transition semigroup from small times to arbitrary times. Thus, under a regular additive noise, every Markov solution is shown to have a property of continuous dependence on initial conditions, which follows from employing the weak-strong uniqueness principle and the Bismut-Elworthy-Li formula. Key words: Markov selection; $\mathcal{W}$-strong Feller; author: - \| Zhao Dong$^{1,}$, Rangrang Zhang$^{1,}$\ [$^1$ Institute of Applied Mathematics, Academy of Mathematics and Systems Science, Chinese Academy of Sciences,]{}\ [No. 55 Zhongguancun East Road, Haidian District, Beijing, 100190, P. R. China]{}\ ([[email protected]]{}, [[email protected]]{} ) title: 'Markov Selection and $\mathcal{W}$-strong Feller for 3D Stochastic Primitive Equations ' --- Introduction ============= The primitive equations (PEs) derived by Boussinesq approximation are a basic model in the study of large oceanic and atmospheric dynamics. These systems form the analytical core of the most advantaged general circulation models. For this reason and due to their challenging nonlinear and anisotropic structure, the PEs have recently received considerable attention from the mathematical community. The mathematical study of the PEs originated in a series of articles by J.L. Lions, R. Temam, and S. Wang in the early 1990s [@L-T-W-1; @L-T-W-2; @L-T-W-3; @L-T-W-4]. They set up the mathematical framework and showed the global existence of weak solutions. For the existence and uniqueness of strong solution, many works are concerned on it. For example, C. Hu, R. Temam and M. Ziane proved the global existence and uniqueness of strong solutions to the viscous primitive equations in thin domains for a large set of initial data whose size depends on the thickness of the domain in [@H-T-Z]. In [@GG-MN], F. $\rm Guill\acute{e}n-Gonz\acute{a}lez$, N. Masmoudi and M.A. Rodriguez-Bellido showed the local existence and uniqueness of strong solutions to the viscous primitive equations for any initial data. C. Cao and E.S. Titi developed a beautiful approach to dealing with the $L^6$-norm of the fluctuation $\tilde{v}$ of horizontal velocity and obtained the global well-posedness for the 3D viscous primitive equations in [@C-T-1]. For the uniqueness of weak solutions, in [@L-T], J. Li and E.S. Titi established some conditional uniqueness of weak solutions to the viscous primitive equations under periodic boundary conditions, and they proved the global existence and uniqueness of weak solutions with the initial data taken as small $L^{\infty}$ perturbations of functions in the space $X=\Big\{v\in (L^6(\mathcal{O}))^2\| \partial_z v\in (L^2(\mathcal{O}))^2\Big\}$. For the primitive equations in random case, many authors paid attention to it. In [@Guo], B. Guo and D. Huang obtained the existence of universal attractor of strong solution under the assumptions that the momentum equation is driven by an additive stochastic forcing and the thermodynamical equation is driven by a fixed heat source. A. Debussche, N. Glatt-Holtz, R. Temam and M. Ziane established the global well-posedness of strong solution, when the primitive equations are driven by multiplicative random noises in [@D-G-T-Z]. For the ergodicity, in [@RR], the authors obtained the existence of global weak solutions, and also obtained the exponential mixing property for the weak solutions which are limits of spectral Galerkin approximations of 3D stochastic primitive equations driven by regular multiplicative noise. For a special case that the stochastic primitive equations are in two space dimensions with small linear multiplicative noise, H. Gao and C. Sun obtained a Wentzell-Freidlin type large deviation principle for by weak convergence method in [@G-S], where they omit the spatial variable $y$ and only take $(x,z)$ into account. Furthermore, they established the Hausdorff dimension of the global attractor is finite in [@G-S-1]. When the primitive equations are driven by an infinite-dimensional additive fractional noise with Hilbert-space-valued, G. Zhou obtained the existence of random attractor in [@Z]. As we know, both in deterministic and stochastic case, the uniqueness of weak solutions is an important open problem, which results in many properties of weak solutions disappear. Thus, in order to have a deeper understanding of weak solutions and have some development on their uniqueness, it’s natural to explore more properties of them. This article presents a step in this direction. We establish that there exists an almost sure Markov family of the primitive equations forced by multiplicative noise. Furthermore, we obtain that every Markov solution has a property of continuous dependence on the initial conditions ($\mathcal{W}$-strong Feller) if the primitive equations are driven by a regular additive noise. In comparison with [@RR], we stress that the main improvement of our paper is that the $\mathcal{W}$-strong Feller is valid for all Markov solutions and not restricted to solutions which are limits of Galerkin approximations. Moreover, the conditions on the noise here is much weaker than those in [@RR]. When uniqueness of weak solutions is open, Markov property has no direct meaning but a natural question is the existence of a Markov selection. A sufficient condition for the existence of almost sure Markov selections was provided by B. Goldys, M. Röckner and X. Zhang in [@B-M-X], where they dealt with an abstract stochastic evolution equations. Here, we apply this sufficient condition to our equations (\[eq5-1\])-(\[eqq20\]) and obtain \[thm-2\] Under Hypothesis H0, there exists an almost sure Markov family $(P_x)_{x\in H}$ of (\[eq5-1\])-(\[eqq20\]). The definition of weak solution of (\[eq5-1\])-(\[eqq20\]) is in Sect 4.2. The important part of this paper is to investigate the continuity with respect to the initial conditions (strong Feller property) for the Markov family $(P_x)_{x\in H}$. To achieve this, $\mathcal{W}$-strong Feller is considered which is weaker than strong Feller in $H$ when $\mathcal{W}$ is a subspace of $H$. In the past two decades, there are several works concerned on $\mathcal{W}$-strong Feller for stochastic evolution equations. In particular, F. Flandoli and M. Romito established an abstract criterion Theorem 5.4 to obtain $\mathcal{W}$-strong Feller property for Markov selections of 3D Stochastic Navier-Stokes equations in [@F-M]. It says that if a Markov process coincides on a small positive random time with a strong Feller process, then it is strong Feller itself. The idea behind this is to use an approximation by a regularised problem, which has itself strong Feller solutions. For the concrete proof, two key points are needed: weak-strong uniqueness principle and the Bismut-Elworthy-Li formula. It’s worth mentioning that this technique is usually applied to handle locally Lipschitz nonlinearities in stochastic equations. To study the strong Feller property of our equations, we will follow the idea of Theorem 5.4 in [@F-M]. Firstly, we introduce an auxiliary Ornstein-Uhlenbeck process $Z$ which is a stationary ergodic solution to a stochastic Stokes equation, then, the approximation process $X^{(R)}_t$ is obtained. To achieve our goal, two steps are needed: the approximation process $X^{(R)}_t$ coincides with the original process on a small positive random time interval and $X^{(R)}_t$ is $\mathcal{W}$-strong Feller. The second step is more challenging, where we have to control the power of $Z$ to be less than 2 as we want to apply the Fernique’s theorem to $Z$. We overcome this difficulty by making use of $\Lambda \frac{\partial Z}{\partial z}$, which is a key idea in our proof, in that case, $Z\in C([0,T]; D(A))$ is needed, then the noise has to be chosen as $A^{-\frac{5}{4}-\varepsilon_0}$ with $\varepsilon_0>0$. Thus, the corresponding $\mathcal{W}$ is equal to $D(A^{\frac{3}{4}+\varepsilon_0})$. In comparison with 3D Navier-Stokes equations, the regularity of the noise here is required to be higher than $A^{-\frac{11}{12}-\varepsilon_0}$ with $\varepsilon_0>0$ for 3D Navier-Stokes equations because of strong nonlinear terms $(\int^z_{-1}\nabla_H \cdot vdz')\frac{\partial v}{\partial z}$ and $(\int^z_{-1}\nabla_H \cdot vdz')\frac{\partial T}{\partial z}$ in the primitive equations. Also, the advective structure of the primitive equations leads to a delicate asymmetry in the nonlinear terms, which requires a more refined calculation. \[thm-4\] Assume Hypothesis H1 holds. Let $(P_x)_{x\in H}$ be the Markov solution of (\[eq5-1\])-(\[eqq20\]) and $(P_t)_{t\geq 0}$ be the associated operators on $B_b(H)$ defined as (\[eq-30\]), then $(P_t)_{t\geq 0}$ is $\mathcal{W}$-strong Feller. The definition of $\mathcal{W}$-strong Feller is defined in Sect 5.1. 0.3cm This paper is organized as follows. In Sects. 2 and 3, we introduce the 3D stochastic primitive equations and make formulation of those equations. The abstract Markov selection principle and concrete proof are given in Sect.4. Finally, $\mathcal{W}$-strong Feller is proved in Sect. 5. Preliminaries ============= The 3D stochastic primitive equations of the large-scale ocean under a stochastic forcing, in a Cartesian system, are written as $$\begin{aligned} \label{eq-1} \frac{\partial v}{\partial t}+(v\cdot \nabla_H)v+\theta\frac{\partial v}{\partial z}+f{k}\times v +\nabla_H P - \Delta v &=&\sigma_1(v,T)\frac{dW_1}{dt},\\ \label{eq-2} \partial_{z}P+T&=&0,\\ \label{eq-3} \nabla_H\cdot v+\partial_{z}\theta&=&0,\\ \label{eq-4} \frac{\partial T}{\partial t}+(v\cdot\nabla_H)T+\theta\frac{\partial T}{\partial z}- \Delta T&=&\sigma_2(v,T)\frac{dW_2}{dt},\end{aligned}$$ where the horizontal velocity field $v=(v^{(1)},v^{(2)})$, the three-dimensional velocity field $(v^{(1)},v^{(2)},\theta)$, the temperature $T$ and the pressure $P$ are unknown functions. $f$ is the Coriolis parameter. ${k}$ is vertical unit vector. Set $\nabla_H=(\partial x,\partial y)$ to be the horizontal gradient operator and $\Delta=\partial^{2}_{x}+\partial^{2}_{y}+\partial^{2}_{z}$ to be the three dimensional Laplacian. $W_1$ and $W_2$ are two independent cylindrical Wiener processes on $H_1$ and $H_2$, respectively. $H_1$ and $H_2$ will be defined in Sect. 3. The spatial variable $(x,y,z)$ belongs to $\mathcal{M}:= \mathbb{T}^2\times (-1,0)$. For simplicity of the presentation, all the physical parameters (height, viscosity, size of periodic box) are set to 1. Refer to [@C-L-T], the boundary value conditions for (\[eq-1\])-(\[eq-4\]) are given by $$\begin{aligned} \label{eqq1} v, \ \theta\ and \ T\ are\ periodic\ in\ x\ and\ y,\\ \label{eqq2} (\partial_{z}v,\theta)\mid _{z=-1,0}=(0,0),\ T\mid_{z=-1}=1, \ T\mid_{z=0}=0.\\ \label{eqq3} (v,T)\mid_{t=0}=(v_0,T_0).\end{aligned}$$ Replacing $T$ and $P$ by $T+z$ and $P-\frac{z^2}{2}$, respectively, then (\[eq-1\])-(\[eq-4\]) with (\[eqq1\])-(\[eqq3\]) is equivalent to the following system $$\begin{aligned} \label{eqq4} \frac{\partial v}{\partial t}+(v\cdot \nabla_H)v+\theta\frac{\partial v}{\partial z}+f{k}\times v +\nabla_H P -\Delta v &=&\sigma_1(v,T+z)\frac{dW_1}{dt},\\ \label{eqq5} \partial_{z}P+T&=&0,\\ \label{eqq6} \nabla_H\cdot v+\partial_{z}\theta&=&0,\\ \label{eqq7} \frac{\partial T}{\partial t}+(v\cdot\nabla_H)T+\theta(\frac{\partial T}{\partial z}+1)-\Delta T&=&\sigma_2(v,T+z)\frac{dW_2}{dt},\end{aligned}$$ subject to the boundary and initial conditions $$\begin{aligned} \label{eqq8} v, \ \theta\ and \ T\ are\ periodic\ in\ x\ and\ y,\\ \label{eqq9} (\partial_{z}v,\theta)\mid _{z=-1,0}=(0,0),\ T\mid_{z=-1,z=0}=0,\\ \label{eqq10} (v,T)\mid_{t=0}=(v_0,T_0).\end{aligned}$$ Here, for simplicity, we still denote by $T_0$ the initial temperature in (\[eqq10\]), though it is now different from that in $(\ref{eqq3})$. Inherent symmetries in the equations show that the solution of the primitive equations on $\mathbb{T}^2\times (-1,0)$ with boundaries (\[eqq8\])-(\[eqq10\]) may be recovered by solving the equations with periodic boundary conditions in $x, y $ and $z$ variables on the extended domain $\mathbb{T}^2\times (-1,1):= \mathbb{T}^3$, and restricting to $z\in (-1,0)$. To see this, consider any solution of (\[eqq4\])-(\[eqq7\]) with boundaries (\[eqq8\])-(\[eqq10\]), we perform that $$\begin{aligned} v(x,y,z)&=&v(x,y,-z), \ \rm{for}\ (x,y,z)\in \mathbb{T}^2\times (0,1),\\ T(x,y,z)&=&-T(x,y,-z), \ \rm{for}\ (x,y,z)\in \mathbb{T}^2\times (0,1),\\ P(x,y,z)&=&P(x,y,-z), \ \rm{for}\ (x,y,z)\in \mathbb{T}^2\times (0,1),\\ \theta(x,y,z)&=&-\theta(x,y,-z),\ \rm{for}\ (x,y,z)\in \mathbb{T}^2\times (0,1).\end{aligned}$$ We also extend $\sigma_1$ in the even fashion and $\sigma_2$ in the odd fashion across $\mathbb{T}^2\times \{0\}$. Hence, we consider the primitive equations on the extended domain $\mathbb{T}^3=\mathbb{T}^2\times (-1,1)$, $$\begin{aligned} \label{eqq14} \frac{\partial v}{\partial t}+(v\cdot \nabla_H)v+\theta\frac{\partial v}{\partial z}+f{k}\times v +\nabla_H P -\Delta v &=&\phi({v},{T})\frac{dW_1}{dt},\\ \label{eqq15} \partial_{z}P+T&=&0,\\ \label{eqq16} \nabla_H\cdot v+\partial_{z}\theta&=&0,\\ \label{eqq17} \frac{\partial T}{\partial t}+(v\cdot\nabla_H)T+\theta(\frac{\partial T}{\partial z}+1)-\Delta T&=&\varphi({v},{T})\frac{dW_2}{dt},\end{aligned}$$ subject to the boundary and initial conditions $$\begin{aligned} \label{eqq11} v, \ \theta\ , P\ and \ T\ are\ periodic\ in\ x\ ,\ y,\ z,\\ \label{eqq12} v\ and\ P\ are\ even\ in \ z, \ \theta\ and\ T\ are\ odd\ in\ z,\\ \label{eqq13} (v,T)\mid_{t=0}=(v_0,T_0),\end{aligned}$$ where $$\phi(v,T)=\sigma_1(v,T+z)\quad \varphi({v},{T})=\sigma_2(v,T+z).$$ Because of the equivalent of the above two kinds of boundary and initial conditions, we consider, throughout this paper, the system (\[eqq14\])-(\[eqq13\]) defined on $\mathbb{T}^3$. Note that condition (\[eqq12\]) is a symmetry condition, which is preserved by system (\[eqq14\])-(\[eqq17\]), that is if a smooth solution to system (\[eqq14\])-(\[eqq17\]) exists and is unique, then it must satisfy the symmetry condition (\[eqq12\]), as long as it is initially satisfied. Note that the vertical velocity $\theta$ can be expressed in terms of the horizonal velocity $v$, through the incompressibility condition (\[eqq16\]) and the symmetry condition (\[eqq12\]), as $$\theta(t,x,y,z)=\Phi(v)(t,x,y,z)=-\int^{z}_{-1}\nabla_H\cdot v(t,x,y,z')dz',$$ moreover, $$\int^{1}_{-1}\nabla_H\cdot v dz=0.$$ Supposing that $p_{b}$ is a certain unknown function at $\Gamma_{b}:=\mathbb{T}^2\times\{-1\}$, and integrating (\[eqq15\]) from $-1$ to $z$, we have $$P(x,y,z,t)= p_{b}(x,y,t)-\int^{z}_{-1} T(x,y,z',t) dz'.$$ Now, (\[eqq14\])-(\[eqq13\]) can be rewritten as $$\begin{aligned} \label{eq5-1} \frac{\partial v}{\partial t}+(v\cdot \nabla_H)v+\Phi(v)\frac{\partial v}{\partial z}+f{k}\times v +\nabla_H p_{b}-\int^{z}_{-1}\nabla_H T dz' -\Delta v &=&\phi(v,T),\\ \label{eq-6-1} \frac{\partial T}{\partial t}+(v\cdot\nabla_H)T+\Phi(v)\frac{\partial T}{\partial z}+ \Phi(v) -\Delta T&=&\varphi(v,T),\\ \label{eq-7-1} \int^{1}_{-1}\nabla_H\cdot v dz&=&0.\end{aligned}$$ The boundary and initial conditions for (\[eq5-1\])-(\[eq-7-1\]) are given by $$\begin{aligned} \label{eqq18} v\ and \ T\ are\ periodic\ in\ x, \ y \ and \ z,\\ \label{eqq19} v\ and\ P\ are\ even\ in \ z, \ \theta\ and\ T\ are\ odd\ in\ z,\\ \label{eqq20} (v, T)\mid_{t=0}=(v_0,T_0).\end{aligned}$$ It is easy to know that Markov Selection and $\mathcal{W}$-strong Feller property for $(v, T)$ of (\[eq5-1\])-(\[eqq20\]) implies the same results of the original solution $(v,T)$ of the system (\[eqq14\])-(\[eqq13\]). In the following, we will focus on (\[eq5-1\])-(\[eqq20\]). Formulation of (\[eq5-1\])-(\[eqq20\]) ====================================== Functional Spaces ------------------- Let $\mathcal{L}(K_1;K_2)$ (resp. $\mathcal{L}_2(K_1;K_2)$) be the space of bounded (resp. Hilbert-Schmidt) linear operators from the Hilbert space $K_1$ to $K_2$, the norm is denoted by $\\|\cdot\\|_{\mathcal{L}(K_1;K_2)}(\\|\cdot\\|_{\mathcal{L}_2(K_1;K_2)})$. Denote by $\|\cdot\|_{L^p(\mathbb{T}^2)}$ the norm of $L^p(\mathbb{T}^2)$ and $\|\cdot\|_{p}$ the norm of $L^p(\mathbb{T}^3)$ for $p\in \mathbb{N}_{+}$. In particular, $\|\cdot\|$ and $(\cdot,\cdot)$ represent the norm and inner product of $L^2(\mathbb{T}^3)$. For the classical Sobolev space $W^{m,2}(\mathbb{T}^3)$, $m\in \mathbb{N}_+$, $$\notag \left\{ \begin{array}{ll} W^{m,2}(\mathbb{T}^3)=\Big\{U\in L^2(\mathbb{T}^3)\Big\| \partial_{\alpha}U\in L^2(\mathbb{T}^3)\ {\rm for} \ \|\alpha\|\leq m\Big\},& \\ \|U\|^2_{W^{m,2}(\mathbb{T}^3)}=\sum_{0\leq\|\alpha\|\leq m}\|\partial_{\alpha}U\|^2. & \end{array} \right.$$ It’s known that $(W^{m,2}(\mathbb{T}^3), \|\cdot\|_{W^{m,2}(\mathbb{T}^3)})$ is a Hilbert space. Define working spaces for equations (\[eq5-1\])-(\[eqq20\]). Let $$\begin{aligned} \notag &&\mathcal{V}_1:=\left\{v\in (C^{\infty}(\mathbb{T}^3))^2;\ \int^1_{-1}\nabla_H\cdot v dz=0, v \ is\ periodic\ in\ x, \ y \ and \ even\ in\ z, \int_{\mathbb{T}^3}vdxdydz=0\right\},\\ \notag &&\mathcal{V}_2:=\left\{T\in C^{\infty}(\mathbb{T}^3);\ \ T\ is\ periodic\ in\ x, \ y \ and \ odd\ in\ z, \int_{\mathbb{T}^3}Tdxdydz=0 \right\}, \end{aligned}$$ $V_1$= the closure of $ \mathcal{V}_1$ with respect to the norm $\|\cdot\|_{W^{1,2}(\mathbb{T}^3)}\times \|\cdot\|_{W^{1,2}(\mathbb{T}^3)}$,\ $V_2$= the closure of $ \mathcal{V}_2$ with respect to the norm $\|\cdot\|_{W^{1,2}(\mathbb{T}^3)}$,\ $H_1$= the closure of $ \mathcal{V}_1$ with respect to the norm $\|\cdot\|\times \|\cdot\|$,\ $H_2$= the closure of $ \mathcal{V}_2$ with respect to the norm $\|\cdot\|$,\ $$V=V_1\times V_2, \quad H=H_1\times H_2.$$ The inner products and norms on $V$, $H$ are given by $$\begin{aligned} (U,U_1)_{V}&=&(v,v_1)_{V_1}+(T,T_1)_{V_2},\\ (U,U_1)&=&(v,v_1)+(T,T_1)=(v^{(1)},v^{(1)}_1)+(v^{(2)},v^{(2)}_1)+(T,T_1),\\ (U,U)^{\frac{1}{2}}_{V}&=&(v,v)^{\frac{1}{2}}_{V_1}+(T,T)^{\frac{1}{2}}_{V_2}, \quad \\|U\\|_{V}=(U,U)^{\frac{1}{2}}_V.\end{aligned}$$ where $U=(v,T), U_1=(v_1,T_1), v=(v^{(1)}, v^{(2)})$ and $v_1=(v^{(1)}_1, v^{(2)}_1)$ . On the periodic domain $\mathbb{T}^3$, it’s known that $-\Delta$ is a self-adjoint compact operator, denote by $\{e_n\}_{n=1,2,\cdot\cdot\cdot}$ an eigenbasis and $\{\lambda_n\}_{n=1,2,\cdot\cdot\cdot}$ the corresponding increasing eigenvalue sequence of $-\Delta$. For $s\in \mathbb{R}^+$, define $$\\|f\\|^2_s=\sum^{\infty}_{k=1}\|\lambda_k\|^{s}\|(f,e_k)\|^2$$ and let $H^s(\mathbb{T}^3)$ denote the Sobolev space of all $f\in H$ for which $\\|f\\|_s$ is finite. It is easy to know that $\\|f\\|_0=\|f\|$ and $\\|f\\|_1=\|f\|_{W^{1,2}(\mathbb{T}^3)}$. For simplicity, denote $\\|\cdot\\|_1=\\|\cdot\\|$. For $s<0$, define $H^s(\mathbb{T}^3)$ to be the dual of $H^{-s}(\mathbb{T}^3)$. Set $\Lambda=(-\Delta)^{\frac{1}{2}}$, then $$\\|f\\|^2_s=\|\Lambda^s f\|^2, \quad \|\Lambda^s v\|^2=\|\Lambda^s v^{(1)}\|^2+\|\Lambda^s v^{(2)}\|^2, \quad \\|U\\|^2_s=\|\Lambda^s v\|^2+\|\Lambda^s T\|^2.$$ Functionals ----------- Define three bilinear forms $a:V\times V\rightarrow \mathbb{R}$, $a_1:V_1\times V_1\rightarrow \mathbb{R}$, $a_2:V_2\times V_2\rightarrow \mathbb{R}$, and their corresponding linear operators $A: V\rightarrow V^{'}$, $A_1: V_1\rightarrow V^{'}_1$, $A_2: V_2\rightarrow V^{'}_2$ by setting $$a(U,U_1):=(AU,U_1)= a_1(v,v_1)+ a_2(T,T_1)\\ ,$$ where $$\begin{aligned} \notag a_1(v,v_1):=(A_1v, v_1)=\int_{\mathbb{T}^3}\left(\nabla_H v\cdot \nabla_H v_1+\frac{\partial v}{\partial z}\cdot\frac{\partial v_1}{\partial z}\right)dxdydz,\end{aligned}$$ $$\begin{aligned} \notag a_2(T,T_1):=(A_2T, T_1)=\int_{\mathbb{T}^3}\left(\nabla_H T\cdot \nabla_H T_1+\frac{\partial T}{\partial z}\frac{\partial T_1}{\partial z}\right)dxdydz,\end{aligned}$$ for any $U=(v,T)$, $U_1=(v_1, T_1)\in V$. (i) : The forms $a$, $a_i\ (i=1,2)$ are coercive, continuous, and therefore, the operators $A:V\rightarrow V'$ and $A_i: V_i\rightarrow V'_i\ (i=1,2)$ are isomorphisms. Moreover, $$\begin{aligned} a(U,U_1)&\leq& C_1\\|U\\|_V\\|U_1\\|_V,\\ a(U,U)&\geq& C_2\\|U\\|^2_V,\end{aligned}$$ where $C_1$ and $C_2$ are two absolute constants (independent of the physically relevant constants $Re_i$, $Rt_i$, etc). (ii) : The isomorphism $A:V\rightarrow V'$ (respectively $A_i: V_i\rightarrow V'_i\ (i=1,2)$) can be extended to a self-adjoint unbounded linear operator on $H$ (respectively on $H_i$, i=1,2), with compact inverse $A^{-1}: H\rightarrow H$ (respectively $A^{-1}_i: H_i\rightarrow H_i \ (i=1,2)$). Now, we define three functionals $b: V\times V\times V\rightarrow \mathbb{R}$, $b_i: V_1\times V_i\times V_i\rightarrow \mathbb{R}\ (i=1,2)$ and the associated operators $B: V\times V\rightarrow V'$, $B_i: V_1\times V_i\rightarrow V'_i\ (i=1,2)$ by setting $$\begin{aligned} b(U,U_1,U_2)&:=&(B(U,U_1),U_2)= b_1(v,v_1,v_2)+ b_2(v, T_1, T_2) ,\\ b_1(v,v_1,v_2)&:=&(B_1(v,v_1), v_2)=\int_{\mathbb{T}^3}\left[(v\cdot \nabla_H)v_1+\Phi(v)\frac{\partial v_1}{\partial z}\right]\cdot v_2dxdydz,\\ b_2(v, T_1, T_2)&:=&(B_2(v,T_1), T_2)=\int_{\mathbb{T}^3}\left[(v\cdot \nabla_H)T_1+\Phi(v)\frac{\partial T_1}{\partial z}\right] T_2dxdydz,\end{aligned}$$ for any $U=(v, T)$, $U_i=(v_i,T_i)\in V$. Moreover, we define another functional $g: V\times V \rightarrow \mathbb{R}$ and the associated linear operator $G: V\rightarrow V'$ by $$\begin{aligned} \notag g(U,U_1)&:=&(G(U), U_1)\notag \\ &=&\int_{\mathbb{T}^3}\left[f(k\times v)\cdot v_1+(\nabla_H p_b-\int^z_{-1}\nabla_H Tdz')\cdot v_1+\Phi(v)\cdot T_1\right]dxdydz . \notag\end{aligned}$$ Finally, using the functionals defined above to obtain the following stochastic evolution equation $$\begin{aligned} \label{aa} \left\{ \begin{array}{ll} dU(t)+AU(t)dt+B(U(t),U(t))dt+G(U(t))dt=\Psi(U(t))dW(t), \\ U(0)=y, \end{array} \right.\end{aligned}$$ where $$\notag W=\left( \begin{array}{c} W_1\\ W_2 \\ \end{array} \right) ,\quad \Psi(U) =\left( \begin{array}{cc} \phi(v,T) & 0 \\ 0 & \varphi(v,T) \\ \end{array} \right).$$ Inequalities ------------- Firstly, we recall the integral version of Minkowshi inequality for the $L^p$ spaces, $p\geq 1$. Let $\mathcal{O}_1\subset \mathbb{R}^{m_1}$ and $\mathcal{O}_2\subset \mathbb{R}^{m_2}$ be two measurable sets, where $m_1$ and $m_2$ are two positive integers. Suppose that $f(\xi, \eta)$ is measurable over $\mathcal{O}_1\times \mathcal{O}_2$. Then $$\notag \left[\int_{\mathcal{O}_1}\left(\int_{\mathcal{O}_2}\|f(\xi,\eta)\|d\eta\right)^p d\xi\right]^{1/p}\leq \int_{\mathcal{O}_2}\left(\int_{\mathcal{O}_1}\|f(\xi, \eta)\|^p d\xi\right)^{1/p}d\eta.$$ ([@C-T-1])\[le-2\] If $v_1 \in H^1(\mathbb{T}^3),v_2 \in H^2(\mathbb{T}^3),v_3 \in H^1(\mathbb{T}^3)$, then (i) : $\|\int_{\mathbb{T}^3}(v_1 \cdot \nabla_H)v_2\cdot v_3 dxdydz\|\leq c\|\nabla_H v_2\|\|v_3\|_3\|v_1\|_6\leq c \|\nabla_H v_2\|\|v_3\|^{\frac{1}{2}}\|\nabla_H v_3\|^{\frac{1}{2}}\|\nabla_H v_1\|$, (ii) : $\|\int_{\mathbb{T}^3}\Phi(v_1)v_{2z}\cdot v_3 dxdydz\|\leq c\|\nabla_H v_1\|\|v_3\|^{\frac{1}{2}}\|\nabla_H v_3\|^{\frac{1}{2}}\|\partial_{z} v_2\|^{\frac{1}{2}}\|\nabla_H \partial_{z} v_2\|^{\frac{1}{2}}$. Markov Selection ================ In the following, we will introduce Markov selection for stochastic evolution equations using the same notations as [@B-M-X]. Preliminaries ------------- Let $(\mathbb{X}, \rho_{\mathbb{X}})$ be a polish space and set $\Omega := C([0,\infty); \mathbb{X})$. Denote by $\mathcal{B}$ the Borel $\sigma$-field of $\Omega$ and by $Pr(\Omega)$ the set of all probability measures on $(\Omega, \mathcal{B})$. Define the canonical process $\xi: \Omega\rightarrow \mathbb{X}$ as $$\xi_t(\omega)=\omega(t).$$ For fixed $t\geq 0$, let $\Omega^t := C([t,\infty); \mathbb{X})$ be the space of all continuous functions from $[t,\infty)$ to $\mathbb{X}$ with the metric $$\rho^t(x,y):=\sum^{\infty}_{m=\lfloor t\rfloor+1}\frac{1}{2^m}\left(\sup_{s\in[t,m]}\rho_{\mathbb{X}}\Big(x(s),y(s)\Big)\wedge 1\right)$$ where $\lfloor t\rfloor$ denotes the integer part of $t$. Then $(\Omega^t, \rho^t)$ is a Polish space. For $s\geq t$, define the $\sigma$-algebra $\mathcal{B}^t_s$ on $\Omega^t$ by $\mathcal{B}^t_s:=\sigma[\xi_r: t\leq r\leq s]$, and write $\mathcal{B}^t:=\bigcup_{s\geq t}\mathcal{B}^t_s$. Thus, we have a measurable space with filtration $(\Omega^t,\mathcal{B}^t, (\mathcal{B}^t_s)_{s\geq t})$. If $t=0$, we simply write $(\Omega, \mathcal{B},(\mathcal{B}_s)_{s\geq 0})$. Finally, define the map $\Phi_t: \Omega \rightarrow \Omega^t$ defined by $$\Phi_t(\omega)(s):= \omega(s-t), \quad s\geq t,$$ which establishes a measurable isomorphism between $(\Omega,\mathcal{B}, (\mathcal{B}_s)_{s\geq 0})$ and $(\Omega^t,\mathcal{B}^t, (\mathcal{B}^t_s)_{s\geq t})$. Given $P\in Pr(\Omega)$ and $t>0$, denote $\omega \mapsto P\|^{\omega}_{\mathcal{B}_t}: \Omega \rightarrow Pr(\Omega^t)$ a regular conditional probability distribution of $P$ on $\mathcal{B}_t$. Since $\Omega$ is a Polish space and every $\sigma$-field $\mathcal{B}_t$ is finitely generated, such a function exists and is unique, up to $P$-null sets. In particular, $$P\|^{\omega}_{\mathcal{B}_t}[\xi_t=\omega(t)]=1$$ for all $\omega\in \Omega$, and if $A\in \mathcal{B}_t$ and $B\in \mathcal{B}^t$, $$P(A\cap B)=\int_A P\|^{\omega}_{\mathcal{B}_t}(B)P(d\omega).$$ Refer to [@F-M], we introduce the following definitions. Given a family $(P_x)_{x\in H}$ of probability measures in $Pr(\Omega)$, the Markov property can be stated as $$P_x\|^{\omega}_{\mathcal{B}_t}=\Phi(t)P_{\omega(t)},\quad for\ P_x-a.e.\ \omega\in \Omega,$$ for each $x\in H$ and for all $t\geq 0$. The family $(P_x)_{x\in H}$ has the almost sure Markov property if for each $x\in H$, there is a set $\Gamma\subset (0,\infty)$ with null Lebesgue measure, such that $$P_x\|^{\omega}_{\mathcal{B}_t}=\Phi(t)P_{\omega(t)},\quad for\ P_x-a.e.\ \omega\in \Omega,$$ for all $t\notin \Gamma$. A General Criterion ------------------- Let $\mathbb{H}$ be a separable Hilbert space, with inner product $\langle\cdot,\cdot\rangle_{\mathbb{H}}$ and norm $\\|\cdot\\|_{\mathbb{H}}$. Let $\mathbb{X}, $ $\mathbb{U}$ be two separable and reflexive Banach spaces with norms $\\|\cdot\\|_{\mathbb{X}}$ and $\\|\cdot\\|_{\mathbb{U}}$, such that $$\mathbb{U}\subset \mathbb{H}\subset \mathbb{X}$$ continuously and densely. If we identify the dual of $\mathbb{H}$ with itself, then we get $$\mathbb{X}^\subset \mathbb{H}^\backsimeq \mathbb{H}\subset \mathbb{X}.$$ The dual pair between $\mathbb{X}$ and $\mathbb{X}^$ is denoted by $${}_{\mathbb{X}}\langle x,y\rangle_{\mathbb{X}^}, \ x\in \mathbb{X},\ y\in \mathbb{X}^.$$ We remark that if $x\in \mathbb{H}$, then $${}_{\mathbb{X}}\langle x,y\rangle_{\mathbb{X}^}= \langle x,y\rangle_{\mathbb{H}}.$$ Let $\mathcal{E}$ be a fixed countable dense subset of $\mathbb{X}^$ which will be chosen in each case and $(W(t))_{t\geq 0}$ be a cylindrical Brownian motion in another separable Hilbert space $(\mathbb{Y},\\|\cdot\\|_{\mathbb{Y}})$ with identity covariance. Consider the following evolution equation: $$\begin{aligned} \label{eq-12} dX(t)=\mathcal{A}(X(t))dt+\mathcal{R}(X(t))dW(t),\ t\geq 0, \ X(0)=x_0 \in \mathbb{H},\end{aligned}$$ where $\mathcal{A}: \mathbb{U}\rightarrow \mathbb{X}$ is $\mathcal{B}(\mathbb{U})/ \mathcal{B}(\mathbb{X})$-measurable and $\mathcal{R}: \mathbb{U}\rightarrow \mathcal{L}_2(\mathbb{Y}; \mathbb{H})$ is $\mathcal{B}(\mathbb{U})/ \mathcal{B}(\mathcal{L}_2(\mathbb{Y}; \mathbb{H}))$-measurable. [@B-M-X]\[dfn-1\] Let $x_0\in \mathbb{H}$. A probability measure $P\in Pr(\Omega)$ is called a martingale solution of (\[eq-12\]) with initial value $x_0$, if it satisfies (M1) : $P(X(0)=x_0)=1$ and for any $n\in \mathbb{N}_{+}$ $$P\Big\{X\in \Omega: \int^n_0\\|\mathcal{A}(X(s))\\|_{\mathbb{X}}ds+\int^n_0 \\|\mathcal{R}(X(s))\\|^2_{\mathcal{L}_2(\mathbb{Y}; \mathbb{H})}ds< +\infty\Big\}=1;$$ (M2) : for every $l\in \mathcal{E}$, the process $$M_l(t,X):=\langle X(t), l\rangle_{\mathbb{X}^}-\int^t_0\langle \mathcal{A}(X(s)), l\rangle_{\mathbb{X}^}ds$$ is a continuous square integrable $\mathcal{B}_t$-martingale with respect to $P$, whose quadratic variation process is given by $$\langle M_{l} \rangle (t,X):= \int^t_0\\|\mathcal{R}^(X(s))(l)\\|^2_{\mathbb{Y}}ds,$$ where the asterisk denote the adjoint operator of $\mathcal{R}(X(s))$; (M3) : for any $p\in \mathbb{N}$, there exists a continuous positive real function $t\mapsto C_{t,p}$ (only depending on $p$ and $\mathcal{A}, \mathcal{R}$), a lower semi-continuous functional $\mathcal{N}_p: \mathbb{U}\rightarrow [0,\infty]$, and a Lebesgue null set $\mathbb{T}_p\subset (0,\infty)$ such that for all $0\leq s\notin \mathbb{T}_p$ and all $t\geq s$ $$\begin{aligned} \mathbb{E}^P\Big( \sup_{r\in [s,t]}\\|X(r)\\|^{2p}_{\mathbb{H}}+\int^t_s\mathcal{N}_p(X(r))dr \Big\| \mathcal{B}_s\Big)\leq C_{t-s,p}\cdot (\\|X(s)\\|^{2p}_{\mathbb{H}}+1). \end{aligned}$$ The above definition of martingale solution is in the sense of Stroock and Varadhan’s martingale problem in [@S-V], which is weaker than that in [@RR]. In [@B-M-X], the authors give a sufficient conditions on $\mathcal{A}$ and $\mathcal{R}$ to obtain Markov family $\{P_{x_0}\}_{x_0\in \mathbb{H}}$ for (\[eq-12\]). For this purpose, they firstly introduced the following function class $\mathcal{U}^q$, $q\geq 1$ : A lower semi-continuous function $\mathcal{N}: \mathbb{U} \rightarrow [0,\infty]$ belong to $\mathcal{U}^q$ if $\mathcal{N}(y)=0$ implies $y=0$, and $$\mathcal{N}(cy)\leq c^q \mathcal{N}(y), \ \forall c\geq 0,\ y\in \mathbb{U}.$$ and $$\Big\{y\in \mathbb{U}: \mathcal{N}(y)\leq 1\Big\} \ is\ relatively\ compact\ in\ \mathbb{U}.$$ The assumptions on $\mathcal{A}$ and $\mathcal{R}$ are given as follows: (C1) : (Demi-Continuity) For any $x\in \mathbb{X}^$, if $y_n$ strong converges to $y$ in $\mathbb{U}$, then $$\lim_{n\rightarrow \infty}{}_{\mathbb{X}}\langle\mathcal{A}(y_n),x\rangle_{\mathbb{X}^}={}_{\mathbb{X}}\langle\mathcal{A}(y),x\rangle_{\mathbb{X}^},$$ and $$\lim_{n\rightarrow \infty}\\|\mathcal{R}^(y_n)(x)-\mathcal{R}^(y)(x)\\|_{\mathbb{Y}}=0.$$ (C2) : (Coercivity Condition) There exist $\lambda_1\geq 0$ and $\mathcal{N}_1\in \mathcal{U}^q$ for some $q\geq 2$ such that for all $x\in \mathbb{X}^$ $${}_{\mathbb{X}}\langle\mathcal{A}(x),x\rangle_{\mathbb{X}^}\leq -\mathcal{N}_1(x)+\lambda_1(1+\\|x\\|^2_{\mathbb{H}}).$$ (C3) : (Growth Condition) There exist $\lambda_2, \lambda_3, \lambda_4>0$ and $\gamma'\geq \gamma>1$ such that for all $x\in \mathbb{U}$ $$\begin{aligned} \\|\mathcal{A}(x)\\|^{\gamma}_{\mathbb{X}}&\leq& \lambda_2\mathcal{N}_1(x)+\lambda_3(1+\\|x\\|^{\gamma'}_{\mathbb{H}}),\\ \\|\mathcal{R}(x)\\|^2_{\mathcal{L}_2(\mathbb{Y};\mathbb{H})}&\leq& \lambda_4(1+\\|x\\|^2_{\mathbb{H}}), \end{aligned}$$ where $\mathcal{N}_1$ is as in $(C2)$. ([@B-M-X])\[thm-6\] Assume $(C1)-(C3)$ hold, for each $x_0\in\mathbb{H}$, there exists a martingale solution $P\in Pr(\Omega)$ starting from $x_0$ to (\[eq-12\]) in the sense of Definition \[dfn-1\]. Then, the main result is \[thm-1\] ( [@B-M-X]) Under $(C1)-(C3)$, there exists an almost sure Markov family $(P_x)_{x\in\mathbb{H}}$ to (\[eq-12\]). Proof of Theorem \[thm-2\] -------------------------- In this part, we will use Theorem \[thm-1\] to get an almost surely Markov family $\{P_{x}\}_{x\in H}$ for (\[aa\]). Firstly, define the operator $\mathcal{A}$ and $\mathcal{R}$ as follows: $$\begin{aligned} \mathcal{A}(y)&:=& Ay+B(y,y)+G(y),\\ \mathcal{R}(y)&:=& \Psi(y)\ \rm{for} \ y\in C^{\infty}(\mathbb{T}^3).\end{aligned}$$ Here, for our equation (\[aa\]), we choose $$\mathbb{U}=H^1(\mathbb{T}^3),\ \mathbb{Y}=\mathbb{H}=H,\ \mathbb{X}=(H^3(\mathbb{T}^3))^, \ \mathbb{X}^=H^3(\mathbb{T}^3),$$ then $\mathbb{X}$ is a Hilbert space and $\mathbb{X}^ \subset \mathbb{U}$ compactly. Moreover, the covariance operator $\Psi$ is assumed to satisfy Hypothesis H0 : \(i) $\Psi:$ $H^1(\mathbb{T}^3)\rightarrow \mathcal{L}_2(H;H)$ is a continuous and bounded Lipschitz mapping, i.e. $$\\|\Psi(y)\\|^2_{\mathcal{L}_2(H;H)}\leq \lambda_0\|y\|^2 +\rho \quad y\in H^1(\mathbb{T}^3),$$ for some constants $\lambda_0\geq0,\ \rho\geq0.$ \(ii) If $y,y_n\in H^1(\mathbb{T}^3)$, such that $y_n$ strongly converges to $y$ in $H^1(\mathbb{T}^3)$, then for any $x\in C^{\infty}(\mathbb{T}^3)$, $$\|\Psi(y_n)^(x)-\Psi(y)^(x)\|_H\rightarrow 0 \quad n\rightarrow \infty.$$ By Lemma \[lem-1\] below, $\mathcal{A}$ can be extended to an operator $\mathcal{A}: H^1(\mathbb{T}^3) \rightarrow \mathbb{X}$. For $y \notin H^1(\mathbb{T}^3)$, $\mathcal{A}(y):= \infty$. \[lem-1\] For any $y_1$, $y_2\in C^{\infty}(\mathbb{T}^3)$, $$\begin{aligned} \\|Ay_1-Ay_2\\|_{\mathbb{X}}&\leq& C_1\|y_1-y_2\|,\\ \\|B(y_1,y_1)-B(y_2,y_2)\\|_{\mathbb{X}}&\leq& C_2(\\|y_1\\|+\\|y_2\\|)\\|y_1-y_2\\|,\\ \\|G(y_1)-G(y_2)\\|_{\mathbb{X}}&\leq& C_3\|y_1-y_2\|.\end{aligned}$$ for constants $C_1, C_2, C_3$. In particular, the operator $\mathcal{A}: C^{\infty}(\mathbb{T}^3)\rightarrow \mathbb{X}$ extends to an operator $\mathcal{A}: H^1(\mathbb{T}^3)\rightarrow \mathbb{X}$ by continuity. Proof of Lemma \[lem-1\] We only prove the second assertion, the first and third estimates can be obtained by Hölder inequality. Refer to [@L-T-W-2], $$\\|B(y,y_1)\\|_{-3}\leq C\|y\|\\|y_1\\|.$$ Then, $$\begin{aligned} \\|B(y_1,y_1)-B(y_2,y_2)\\|_{\mathbb{X}} &=&\sup_{x\in C^{\infty}(\mathbb{T}^3): \\|x\\|_{3}\leq 1 }\|\langle B(y_1,y_1)-B(y_2,y_2),x\rangle\|\\ &=& \sup_{x\in C^{\infty}(\mathbb{T}^3): \\|x\\|_{3}\leq 1 }\|\langle B(y_1,y_1-y_2)-B(y_1-y_2,y_2),x\rangle\|\\ &\leq & C_2(\\|y_1\\|+\\|y_2\\|)\\|y_1-y_2\\|.\end{aligned}$$ $\hfill\blacksquare$ In order to use Theorem \[thm-1\], define the functional $\mathcal{N}_1$ on $\mathbb{U}$ as follows: $$\begin{aligned} \mathcal{N}_1(y):= \left\{ \begin{array}{ll} \\|y\\|^2, & if\ y\in H^1(\mathbb{T}^3), \\ +\infty, & otherwise. \end{array} \right.\end{aligned}$$ It is obvious that $\mathcal{N}_1\in \mathcal{U}^2$. Proof of Theorem \[thm-2\] By Theorem \[thm-1\], we only need to check (C1)-(C3) for $\mathcal{A}$ and $\mathcal{R}$. (I) : The demi-continuity condition (C1) holds by Lemma \[lem-1\] and Hypothesis H0. (II) : The coercivity condition (C2) follows since $$\langle B(y,y),y\rangle=0,$$ then, by Young inequality, we have $$\begin{aligned} {}_{H^{-3}(\mathbb{T}^3)}\langle \mathcal{A}(y),y\rangle_{H^3(\mathbb{T}^3)} &=&{}_{H^{-3}(\mathbb{T}^3)}\langle Ay+B(y,y)+G(y),y\rangle_{H^3(\mathbb{T}^3)}\\ &=&-\\|y\\|^2+C\|y\|\\|y\\|\\ &\leq& -\\|y\\|^2+ \frac{1}{2}\\|y\\|^2+C\|y\|^2\\ &\leq& -\frac{1}{2}\\|y\\|^2+\lambda_1(1+\|y\|^2). \end{aligned}$$ (III) : The growth condition (C3) is clear since by Lemma \[lem-1\], it gives $$\\|\mathcal{A}(y)\\|^2_{-3}\leq \lambda_1 \\|y\\|^2 +\lambda_2 (1+\|y\|^2),$$ and by Hypothesis H0, we have $$\\|\mathcal{R}(y)\\|^2_{\mathcal{L}_2(H;{H})}\leq \lambda_3(1+\|y\|^2).$$ $\hfill\blacksquare$ By Theorem \[thm-6\], for any $x_0\in H$, there exists a martingale solution $P_{x_0}\in Pr(\Omega)$ to (\[aa\]) in the sense of Definition \[dfn-1\]. Refer to [@B-M-X] and [@ROMITO], we know that $P_{x_0}$ is obtained by means of maximisation. $\mathcal{W}$-strong Feller =========================== In this section, we apply the abstract result ( Theorem 5.4 in [@F-M]) to obtain that every Markov selection in Sect. 4 has $\mathcal{W}$-strong Feller property. Preliminaries ------------- Firstly, we recall the following important lemma ([@Resnick], Lemma A.4): \[lem-2\] Suppose that $s>0$ and $p\in(1,\infty)$. If $f,g\in C^{\infty}(\mathbb{T}^3)$, then $$\|\Lambda^s(fg)\|_p\leq C(\|f\|_{p_1}\|\Lambda^sg\|_{p_2}+\|g\|_{p_3}\|\Lambda^sf\|_{p_4}),$$ with $p_i\in(1,\infty]$, $i=1,\cdot\cdot\cdot,4$ such that $$\frac{1}{p}=\frac{1}{p_1}+\frac{1}{p_2}=\frac{1}{p_3}+\frac{1}{p_4}.$$ $\|g\|_{p_3}\|\Lambda^sf\|_{p_4}$ can be equal to $\|f\|_{p_1}\|\Lambda^sg\|_{p_2}$ by choosing suitable parameters $p_3$ and $p_4$, in that case, we only write one of them. We will also use the following Sobolev inequality ([@Stein], Chapter V ): \[lem-3\] Suppose that $q>1$,$p\in[q,\infty)$ and $$\frac{1}{p}+\frac{\sigma}{3}=\frac{1}{q}.$$ If $\Lambda^{\sigma}f\in L^q(\mathbb{T}^3)$, then $f\in L^p(\mathbb{T}^3)$ and there is a constant $C\geq0$ independent of $f$ such that $$\|f\|_p\leq C\|\Lambda^{\sigma}f\|_q.$$ We shall use as well the following interpolation inequality ( [@Ju], (5.5)). \[lem-4\] For $f\in C^{\infty}(\mathbb{T}^3)$, we have $$\\|f\\|_s\leq C\\|f\\|^{\frac{s_2-s}{s_2-s_1}}_{s_1}\\|f\\|^{\frac{s-s_1}{s_2-s_1}}_{s_2},\quad s_1<s<s_2.$$ Refer to the appendix of [@C-T-2], we have the following lemma. For any $v,T $ and $\omega\in C^{\infty}(\mathbb{T}^3)$, $$\begin{aligned} &&\Big\|\int_{\mathbb{T}^3}\Phi(v)\frac{\partial T}{\partial z}\omega\Big\|\\ &\leq& \int_{\mathbb{T}^2}\Big(\int^1_{-1}\|\nabla_H v\|dz'\Big)\Big(\int^1_{-1}\|\frac{\partial T}{\partial z}\omega\|dz\Big)dxdy\\ &\leq &C\int_{\mathbb{T}^2}\Big(\int^1_{-1}\|\nabla_H v\|dz'\Big)\Big(\int^1_{-1}\|\frac{\partial T}{\partial z}\|^2dz\Big)^\frac{1}{2}\Big(\int^1_{-1}\|\omega\|^2dz\Big)^\frac{1}{2}dxdy\\ &\leq& C\|\omega\|\\|\nabla_H v\\|_{s_1}\\|T\\|_{s_2},\end{aligned}$$ where $s_1+s_2=1$. At last, we introduce the definition of $\mathcal{W}$-strong Feller. ($\mathcal{W}$-strong Feller) A given semigroup $(P_t)_{t\geq 0}$ on $B_b(H)$ is $\mathcal{W}$-strong Feller, if for any $t>0$ and $\psi\in B_b(H)$, $P_t\psi\in C_b(\mathcal{W})$. $\mathcal{W}$ Space and Hypothesis ----------------------------------- For any $\varepsilon_0>0$ and set $$\mathcal{W}=D(A^{\frac{3}{4}+\varepsilon_0}),\ \|U\|_{\mathcal{W}}=\|v\|_{\mathcal{W}}+\|T\|_{\mathcal{W}}=\|\Lambda^{\frac{3}{2}+2\varepsilon_0} v\|+\|\Lambda^{\frac{3}{2}+2\varepsilon_0} T\|.$$ In this section, we choose $$\Omega := C([0,\infty); H^{-\beta})$$ for some $\beta >3$ and $\mathcal{B}$ denote the Borel $\sigma$-algebra on $\Omega$. We assume the noise is additive, nondegenerate and regular. Concretely, Hypothesis H1 : There are an isomorphism $Q_0$ of $H$ and a number $\alpha_0 = \frac{1}{2}+\varepsilon_0$ such that $$\Psi=Q^{\frac{1}{2}}=A^{-\frac{3}{4}-\alpha_0}Q^{\frac{1}{2}}_0=A^{-\frac{5}{4}-\varepsilon_0}Q^{\frac{1}{2}}_0,$$ where the covariant $Q$: $H\rightarrow H$ is a symmetric non-negative trace-class operator on $H$. Firstly, we notice that Hypothesis H1 implies Hypothesis H0. Indeed, the operator $A^{-\frac{3}{4}-\varepsilon}$ is Hilbert-Schmidt in $H$, for every $\varepsilon >0$. Moreover, $A^{-\frac{3}{4}-\varepsilon}Q^{\frac{1}{2}}_0W(t)$ is a Brownian motion in H, for every $\varepsilon>0$ and every isomorphism $Q_0$ of H, where $W(t)$ is a cylindrical Wiener process on H. In conclusion, $A^{-\frac{3}{4}-\alpha_0}Q^{\frac{1}{2}}_0W(t)$ is a Brownian motion in $D(A^{\alpha})$ for every $\alpha_0>\alpha >0$. In the following, we will consider equations (\[aa\]) in the following abstract form: $$\begin{aligned} \label{bb} \left\{ \begin{array}{ll} dU(t)+AU(t)dt+B(U(t),U(t))dt+G(U(t))dt=Q^{\frac{1}{2}}dW(t), \\ U(0)=y. \end{array} \right.\end{aligned}$$ Under Hypothesis H1, in [@Guo], the authors have proved that for $y\in V$, there exists a unique strong solution $U=(v,T)$. However, for $y\in H$, the uniqueness of the weak solution is still open, hence, we have to deal with the selected Markov process. For $y\in H$, let $P_y$ denote the law of the corresponding solution $U(\cdot, y)$ to (\[bb\]). Since Hypothesis H1 implies Hypothesis H0, by Theorem \[thm-2\], the measures $P_y$, $y\in H$ form a Markov process. Let $(P_t)_{t\geq 0}$ be the associated transition semigroup on $B_b(H)$, defined as $$\begin{aligned} \label{eq-30} P_t(\varphi)(y):= \mathbb{E}[\varphi(U(t,y))] \quad \forall y\in H,\ \forall \varphi\in B_b(H).\end{aligned}$$ Proof of Theorem \[thm-4\] -------------------------- For the proof, we shall use Theorem 5.4 in [@F-M], which is an abstract result to prove the strong Feller property of Markov selection. In order to achieve this, we follow the idea of Theorem 5.11 in [@F-M] to construct $P^{(R)}_y$. We introduce an equation which differs from the original one by a cut-off only, so that with large probability they have the same trajectories on a small random time interval. Consider $$\begin{aligned} \label{eq-16} dU(t)+AU(t)dt+\chi_R(\|U\|^2_{\mathcal{W}})\Big[B(U(t),U(t))+G(U(t))\Big]dt=Q^{\frac{1}{2}}dW(t),\end{aligned}$$ where $\chi_R: \mathbb{R}\rightarrow [0,1]$ is of class $C^{\infty}$ such that $\chi_R(\|U\|)=1$ if $\|U\|\leq R$, and $\chi_R(\|U\|)=0$ if $\|U\|\geq R+1$ and its first derivative bounded by 1. \[thm-5\] (Weak-strong uniqueness) Suppose Hypothesis H1 holds. Then for every $y\in \mathcal{W}$, equation (\[eq-16\]) has a unique martingale solution $P^{(R)}_y$, with $$P^{(R)}_y[C([0,\infty);\mathcal{W})]=1.$$ Let $\tau_R:\Omega\rightarrow [0,\infty]$ be defined by $$\tau_R(\omega):= \inf\{t\geq 0: \|\omega(t)\|^2_{\mathcal{W}}\geq R\}$$ and $\tau_R(\omega):= \infty$ if this set is empty. If $y\in \mathcal{W}$ and $\|y\|^2_{\mathcal{W}}< R$, then $$\begin{aligned} \label{eq-39} \lim_{\varepsilon \rightarrow 0} P^{(R)}_{y+h}[\tau_R\geq \varepsilon]=1, \ uniformly \ in \ h\in \mathcal{W}, \ \|h\|_{\mathcal{W}}<1.\end{aligned}$$ Moreover, $$\begin{aligned} \label{eq-40} \mathbb{E}^{P^{(R)}_y}[\varphi(\omega_t)I_{[\tau_R \geq t]}]=\mathbb{E}^{P_y}[\varphi(\omega_t)I_{[\tau_R \geq t]}]\end{aligned}$$ for every $t\geq 0$ and $\varphi \in B_b(H)$. Proof of Theorem \[thm-5\] Let $Z$ be the solution to $$d Z(t)+A Z(t) dt = Q^{\frac{1}{2}}dW(t),$$ with the initial data $Z(0)=0$ and let $X^{(R)}_y$ be the solution to the auxilary problem $$\begin{aligned} \label{eq-17} \frac{dX^{(R)}(t) }{dt}+AX^{(R)}(t)+\chi_R(\|X^{(R)}+Z\|^2_{\mathcal{W}})\Big[B(X^{(R)}+Z,X^{(R)}+Z)+G(X^{(R)}+Z)\Big]=0,\end{aligned}$$ with $X^{(R)}(0)=y$. Moreover, define $U^{(R)}(t)=X^{(R)}(t)+Z(t)$, which is a weak solution to equation (\[eq-16\]). We denote its law on $\Omega$ by $P^{(R)}_y$. For the noise, by Hypothesis H1, the trajectories of the noise belong to $$\Omega^:= \bigcap_{\beta\in (0,\frac{1}{2}),\ \alpha\in [0,\frac{1}{2}+\varepsilon_0)}C^{\beta}([0,\infty);D(A^{\alpha}))$$ with probability one. Hence, the analyticity of the semigroup generated by $A$ implies that for each $\omega\in \Omega^$, $$\begin{aligned} \label{eq-22} Z(\omega)\in C([0,\infty);D(A^{1+\varepsilon_0-\varepsilon}))\subseteq C([0,\infty);\mathcal{W}), \end{aligned}$$ for every $\varepsilon\in(0,\frac{1}{4})$. Now, fix $\omega\in \Omega^$, we will prove that equation (\[eq-17\]) has a unique global weak solution in $C([0,\infty);\mathcal{W})$. Denoting by $X=(\kappa^{(R)},g^{(R)})$, $Z=(Z_1,Z_2)$, $\kappa^{(R)}=v^{(R)}-Z_1$, $g^{(R)}=T-Z_2$, then (\[eq-17\]) can be rewritten as $$\begin{aligned} \label{eq-18} \frac{\partial \kappa^{(R)}}{\partial t}&+&\Big((\kappa^{(R)}+ Z_1)\cdot \nabla_H\Big)(\kappa^{(R)}+ Z_1)+\Phi(\kappa^{(R)}+Z_1)\frac{\partial( \kappa^{(R)}+ Z_1)}{\partial z} +f{k}\times (\kappa^{(R)}+Z_1)\\ \notag &+&\nabla_H p_{b}-\int^{z}_{-1}\nabla_H (g^{(R)}+Z_2)dz'-\Delta\kappa^{(R)}=0,\end{aligned}$$ $$\begin{aligned} \label{eq-19} \frac{\partial g^{(R)}}{\partial t}+\Big((\kappa^{(R)}+ Z_1)\cdot \nabla_H\Big)(g^{(R)}+ Z_2)+\Phi(\kappa^{(R)}+Z_1)\frac{\partial (g^{(R)}+Z_2)}{\partial z}+\Phi(\kappa^{(R)}+Z_1)-\Delta g^{(R)} =0.\end{aligned}$$ (Existence of weak solution)* Multiplying (\[eq-18\]) by $-\Lambda^{3+4\varepsilon_0} \kappa^{(R)}$, integrating over $\mathbb{T}^3$, it follows that $$\begin{aligned} &&\frac{1}{2}\frac{d}{dt}\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\kappa^{(R)}\|^2+\|\Lambda^{\frac{5}{2}+2\varepsilon_0}\kappa^{(R)}\|^2\\ &=&\chi_R(\|U^{(R)}\|^2_{\mathcal{W}})\int_{\mathbb{T}^3}\Lambda^{3+4\varepsilon_0}\kappa^{(R)} \Big[\Big((\kappa^{(R)}+ Z_1)\cdot \nabla_H\Big)(\kappa^{(R)}+ Z_1)\Big]dxdydz\\ \notag && +\chi_R(\|U^{(R)}\|^2_{\mathcal{W}})\int_{\mathbb{T}^3}\Lambda^{3+4\varepsilon_0}\kappa^{(R)} \Big[\Phi(\kappa^{(R)}+Z_1)\frac{\partial( \kappa^{(R)}+ Z_1)}{\partial z}\Big]dxdydz\\ \notag &&+\chi_R(\|U^{(R)}\|^2_{\mathcal{W}})\int_{\mathbb{T}^3}\Lambda^{3+4\varepsilon_0}\kappa^{(R)} \Big[f{k}\times (\kappa^{(R)}+Z_1)+\nabla_H p_{b}-\int^{z}_{-1}\nabla_H (g^{(R)}+Z_2)dz'\Big] dxdydz\\ &:=& \chi_R(\|U^{(R)}\|^2_{\mathcal{W}})(I_1+I_2+I_3). \end{aligned}$$ For $I_1$, we have $$\begin{aligned} &&\Big\|\int_{\mathbb{T}^3}\Lambda^{3+4\varepsilon_0}\kappa^{(R)}[(\kappa^{(R)}\cdot \nabla_H)\kappa^{(R)}] dxdydz\Big\|\\ &=& \Big\|\int_{\mathbb{T}^3}\Lambda^{\frac{5}{2}+2\varepsilon_0}\kappa^{(R)}\Lambda^{\frac{1}{2}+2\varepsilon_0}[(\kappa^{(R)}\cdot \nabla_H)\kappa^{(R)}] dxdydz \Big\|\\ &\leq & C\|\Lambda^{\frac{5}{2}+2\varepsilon_0}\kappa^{(R)}\|\|\Lambda^{\frac{1}{2}+2\varepsilon_0+\sigma_1}\kappa^{(R)}\|\|\Lambda^{1+\sigma_2}\kappa^{(R)}\|+ C\|\Lambda^{\frac{5}{2}+2\varepsilon_0}\kappa^{(R)}\|\|\Lambda^{s_1}\kappa^{(R)}\|\|\Lambda^{\frac{3}{2}+2\varepsilon_0+s_2}\kappa^{(R)}\|\\ &\leq & C\|\Lambda^{\frac{5}{2}+2\varepsilon_0}\kappa^{(R)}\|\|\Lambda^{\frac{3}{2}}\kappa^{(R)}\|\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\kappa^{(R)}\|\\ &\leq & \varepsilon \| \Lambda^{\frac{5}{2}+2\varepsilon_0}\kappa^{(R)}\|^2+C\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\kappa^{(R)}\|^4,\end{aligned}$$ where the first equality follows from Lemma \[lem-2\]. In the first inequality, $\sigma_1+\sigma_2=\frac{3}{2}$, $s_1+s_2=\frac{3}{2}$, we choose $$\sigma_1=1-2\varepsilon_0, \quad \sigma_2=\frac{1}{2}+2\varepsilon_0,\quad s_1=\frac{3}{2}, \quad s_2=0.$$ The second inequality follows from Lemma \[lem-4\]. The Young inequality is used in the last inequality. By the same argument, we have $$\begin{aligned} &&\Big\|\int_{\mathbb{T}^3}\Lambda^{3+4\varepsilon_0}\kappa^{(R)}[(\kappa^{(R)}\cdot \nabla_H) Z_1)] dxdydz\Big\|\\ &=& \Big\|\int_{\mathbb{T}^3}\Lambda^{\frac{5}{2}+2\varepsilon_0}\kappa^{(R)}\Lambda^{\frac{1}{2}+2\varepsilon_0}[(\kappa^{(R)}\cdot \nabla_H) Z_1] dxdydz \Big\|\\ &\leq & C\|\Lambda^{\frac{5}{2}+2\varepsilon_0}\kappa^{(R)}\|\|\Lambda^{\frac{1}{2}+2\varepsilon_0+\sigma_1}\kappa^{(R)}\|\|\Lambda^{1+\sigma_2}Z_1\|+ C\|\Lambda^{\frac{5}{2}+2\varepsilon_0}\kappa^{(R)}\|\|\Lambda^{s_1}\kappa^{(R)}\|\|\Lambda^{\frac{3}{2}+2\varepsilon_0+s_2}Z_1\|\\ &\leq& C\|\Lambda^{\frac{5}{2}+2\varepsilon_0}\kappa^{(R)}\|\Big(\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\kappa^{(R)}\|\|\Lambda^{\frac{3}{2}}Z_1\| +\|\Lambda^{\frac{3}{2}+2\varepsilon_0}Z_1\|\|\Lambda^{\frac{3}{2}}\kappa^{(R)}\|\Big)\\ &\leq &\varepsilon \| \Lambda^{\frac{5}{2}+2\varepsilon_0}\kappa^{(R)}\|^2+C\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\kappa^{(R)}\|^2\|\Lambda^{\frac{3}{2}}Z_1\|^2 +C\|\Lambda^{\frac{3}{2}+2\varepsilon_0}Z_1\|^2\|\Lambda^{\frac{3}{2}}\kappa^{(R)}\|^2,\end{aligned}$$ where $\sigma_1+\sigma_2=\frac{3}{2}$, $s_1+s_2=\frac{3}{2}$, we choose $$\sigma_1=1, \quad \sigma_2=\frac{1}{2},\quad s_1=\frac{3}{2}, \quad s_2=0.$$ $$\begin{aligned} &&\Big\|\int_{\mathbb{T}^3}\Lambda^{3+4\varepsilon_0}\kappa^{(R)}[(Z_1\cdot \nabla_H) \kappa^{(R)})] dxdydz\Big\|\\ &=& \Big\|\int_{\mathbb{T}^3}\Lambda^{\frac{5}{2}+2\varepsilon_0}\kappa^{(R)}\Lambda^{\frac{1}{2}+2\varepsilon_0}[(Z_1\cdot \nabla_H)\kappa^{(R)}] dxdydz \Big\|\\ &\leq & C\|\Lambda^{\frac{5}{2}+2\varepsilon_0}\kappa^{(R)}\|\|\Lambda^{\frac{1}{2}+2\varepsilon_0+\sigma_1}Z_1\|\|\Lambda^{1+\sigma_2}\kappa^{(R)}\|+ C\|\Lambda^{\frac{5}{2}+2\varepsilon_0}\kappa^{(R)}\|\|\Lambda^{s_1}Z_1\|\|\Lambda^{\frac{3}{2}+2\varepsilon_0+s_2}\kappa^{(R)}\|\\ &\leq& C\|\Lambda^{\frac{5}{2}+2\varepsilon_0}\kappa^{(R)}\|\|\Lambda^{\frac{3}{2}}Z_1\|\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\kappa^{(R)}\|\\ &\leq &\varepsilon \| \Lambda^{\frac{5}{2}+2\varepsilon_0}\kappa^{(R)}\|^2+C\|\Lambda^{\frac{3}{2}}Z_1\|^2\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\kappa^{(R)}\|^2,\end{aligned}$$ where $\sigma_1+\sigma_2=\frac{3}{2}$, $s_1+s_2=\frac{3}{2}$, we choose $$\sigma_1=1-2\varepsilon_0, \quad \sigma_2=\frac{1}{2}+2\varepsilon_0,\quad s_1=\frac{3}{2}, \quad s_2=0.$$ $$\begin{aligned} &&\Big\|\int_{\mathbb{T}^3}\Lambda^{3+4\varepsilon_0}\kappa^{(R)}[(Z_1\cdot \nabla_H) Z_1)] dxdydz\Big\|\\ &=& \Big\|\int_{\mathbb{T}^3}\Lambda^{\frac{5}{2}+2\varepsilon_0}\kappa^{(R)}\Lambda^{\frac{1}{2}+2\varepsilon_0}[(Z_1\cdot \nabla_H)Z_1] dxdydz \Big\|\\ &\leq & C\|\Lambda^{\frac{5}{2}+2\varepsilon_0}\kappa^{(R)}\|\|\Lambda^{\frac{1}{2}+2\varepsilon_0+\sigma_1}Z_1\|\|\Lambda^{1+\sigma_2}Z_1\|+ C\|\Lambda^{\frac{5}{2}+2\varepsilon_0}\kappa^{(R)}\|\|\Lambda^{s_1}Z_1\|\|\Lambda^{\frac{3}{2}+2\varepsilon_0+s_2}Z_1\|\\ &\leq & C\|\Lambda^{\frac{5}{2}+2\varepsilon_0}\kappa^{(R)}\|\|\Lambda^{\frac{3}{2}+2\varepsilon_0}Z_1\|\|\Lambda^{\frac{3}{2}}Z_1\|\\ &\leq &\varepsilon \| \Lambda^{\frac{5}{2}+2\varepsilon_0}\kappa^{(R)}\|^2+C\|\Lambda^{\frac{3}{2}}Z_1\|^2\|\Lambda^{\frac{3}{2}+2\varepsilon_0}Z_1\|^2,\end{aligned}$$ where $\sigma_1+\sigma_2=\frac{3}{2}$, $s_1+s_2=\frac{3}{2}$, we choose $$\sigma_1=1, \quad \sigma_2=\frac{1}{2},\quad s_1=\frac{3}{2}, \quad s_2=0.$$ For $I_2$, we have $$\begin{aligned} &&\Big\|\int_{\mathbb{T}^3}\Lambda^{3+4\varepsilon_0}\kappa^{(R)} [\Phi(\kappa^{(R)})\frac{\partial \kappa^{(R)}}{\partial z}]dxdydz\Big\|\\ &=& \Big\|\int_{\mathbb{T}^3}\Lambda^{\frac{5}{2}+2\varepsilon_0}\kappa^{(R)}\Lambda^{\frac{1}{2}+2\varepsilon_0} [\Phi(\kappa^{(R)})\frac{\partial \kappa^{(R)}}{\partial z}]dxdydz\Big\|\\ &\leq & C\|\Lambda^{\frac{5}{2}+2\varepsilon_0}\kappa^{(R)}\|\|\Lambda^{\frac{3}{2}+2\varepsilon_0+\sigma_1}\kappa^{(R)}\|\|\Lambda^{1+\sigma_2} \kappa^{(R)}\|+C\|\Lambda^{\frac{5}{2}+2\varepsilon_0}\kappa^{(R)}\|\|\Lambda^{1+s_1}\kappa^{(R)}\|\|\Lambda^{\frac{3}{2}+2\varepsilon_0+s_2} \kappa^{(R)}\|\\ &\leq & C\|\Lambda^{\frac{5}{2}+2\varepsilon_0}\kappa^{(R)}\| \|\Lambda^{\frac{3}{2}+2\varepsilon_0}\kappa^{(R)}\|\|\Lambda^2\kappa^{(R)}\|\\ &\leq& C\|\Lambda^{\frac{5}{2}+2\varepsilon_0}\kappa^{(R)}\|^{\frac{3}{2}-2\varepsilon_0} \|\Lambda^{\frac{3}{2}+2\varepsilon_0}\kappa^{(R)}\|^{\frac{1}{2}+2\varepsilon_0}\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\kappa^{(R)}\|\\ &\leq & \varepsilon\|\Lambda^{\frac{5}{2}+2\varepsilon_0}\kappa^{(R)}\|^2+C\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\kappa^{(R)}\|^{2+\frac{4}{1+4\varepsilon_0}},\end{aligned}$$ where $\sigma_1+\sigma_2=1$, $s_1+s_2=1$, we choose $$\sigma_1=s_2=\frac{1}{2}-2\varepsilon_0, \quad \sigma_2=s_1=\frac{1}{2}+2\varepsilon_0.$$ $$\begin{aligned} &&\Big\|\int_{\mathbb{T}^3}\Lambda^{3+4\varepsilon_0}\kappa^{(R)} [\Phi(\kappa^{(R)})\frac{\partial Z_1}{\partial z}]dxdydz\Big\|\\ &=& \Big\|\int_{\mathbb{T}^3}\Lambda^{\frac{5}{2}+2\varepsilon_0}\kappa^{(R)}\Lambda^{\frac{1}{2}+2\varepsilon_0} [\Phi(\kappa^{(R)})\frac{\partial Z_1}{\partial z}]dxdydz\Big\|\\ &\leq & C\|\Lambda^{\frac{5}{2}+2\varepsilon_0}\kappa^{(R)}\|\|\Lambda^{\frac{3}{2}+2\varepsilon_0+\sigma_1}\kappa^{(R)}\|\|\Lambda^{1+\sigma_2} Z_1\|+C\|\Lambda^{\frac{5}{2}+2\varepsilon_0}\kappa^{(R)}\|\|\Lambda^{1+s_1}\kappa^{(R)}\|\|\Lambda^{\frac{3}{2}+2\varepsilon_0+s_2} Z_1\|\\ &\leq & \|\Lambda^{\frac{5}{2}+2\varepsilon_0}\kappa^{(R)}\|\|\Lambda^{\frac{3}{2}+2\varepsilon_0}Z_1\|\|\Lambda^{2}\kappa^{(R)}\|\\ &\leq& \|\Lambda^{\frac{5}{2}+2\varepsilon_0}\kappa^{(R)}\|^{\frac{3}{2}-2\varepsilon_0}\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\kappa^{(R)}\|^{\frac{1}{2}+2\varepsilon_0} \|\Lambda^{\frac{3}{2}+2\varepsilon_0}Z_1\|^2\\ &\leq& \varepsilon\|\Lambda^{\frac{5}{2}+2\varepsilon_0}\kappa^{(R)}\|^2+C\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\kappa^{(R)}\|^2\|\Lambda^{\frac{3}{2}+2\varepsilon_0}Z_1\|^{\frac{4}{1+4\varepsilon_0}},\end{aligned}$$ where $\sigma_1+\sigma_2=1$, $s_1+s_2=1$, we choose $$\sigma_1=\frac{1}{2}-2\varepsilon_0, \quad \sigma_2=\frac{1}{2}+2\varepsilon_0,\quad s_1=1,\quad s_2=0.$$ $$\begin{aligned} &&\Big\|\int_{\mathbb{T}^3}\Lambda^{3+4\varepsilon_0}\kappa^{(R)} [\Phi(Z_1)\frac{\partial \kappa^{(R)}}{\partial z}]dxdydz\Big\|\\ &=& \Big\|\int_{\mathbb{T}^3}\Lambda^{\frac{5}{2}+2\varepsilon_0}\kappa^{(R)}\Lambda^{\frac{1}{2}+2\varepsilon_0} [\Phi(Z_1)\frac{\partial \kappa^{(R)}}{\partial z}]dxdydz\Big\|\\ &\leq & C\|\Lambda^{\frac{5}{2}+2\varepsilon_0}\kappa^{(R)}\|\|\Lambda^{\frac{3}{2}+2\varepsilon_0+\sigma_1}Z_1\|\|\Lambda^{1+\sigma_2} \kappa^{(R)}\|+C\|\Lambda^{\frac{5}{2}+2\varepsilon_0}\kappa^{(R)}\|\|\Lambda^{1+s_1}Z_1\|\|\Lambda^{\frac{3}{2}+2\varepsilon_0+s_2} \kappa^{(R)}\|\\ &\leq & \|\Lambda^{\frac{5}{2}+2\varepsilon_0}\kappa^{(R)}\|\|\Lambda^{\frac{3}{2}+2\varepsilon_0}Z_1\|\|\Lambda^{2}\kappa^{(R)}\|\\ &\leq& \|\Lambda^{\frac{5}{2}+2\varepsilon_0}\kappa^{(R)}\|^{\frac{3}{2}-2\varepsilon_0}\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\kappa^{(R)}\|^{\frac{1}{2}+2\varepsilon_0} \|\Lambda^{\frac{3}{2}+2\varepsilon_0}Z_1\|^2\\ &\leq& \varepsilon\|\Lambda^{\frac{5}{2}+2\varepsilon_0}\kappa^{(R)}\|^2+C\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\kappa^{(R)}\|^2\|\Lambda^{\frac{3}{2}+2\varepsilon_0}Z_1\|^{\frac{4}{1+4\varepsilon_0}},\end{aligned}$$ where $\sigma_1+\sigma_2=1$, $s_1+s_2=1$, we choose $$\sigma_1=0, \quad \sigma_2=1,\quad s_1=\frac{1}{2}+2\varepsilon_0 ,\quad s_2=\frac{1}{2}-2\varepsilon_0.$$ $$\begin{aligned} &&\Big\|\int_{\mathbb{T}^3}\Lambda^{3+4\varepsilon_0}\kappa^{(R)} [\Phi(Z_1)\frac{\partial Z_1}{\partial z}]dxdydz\Big\|\\ &=& \Big\|\int_{\mathbb{T}^3}\Lambda^{\frac{5}{2}+2\varepsilon_0}\kappa^{(R)}\Lambda^{\frac{1}{2}+2\varepsilon_0} [\Phi(Z_1)\frac{\partial Z_1}{\partial z}]dxdydz\Big\|\\ &\leq& C\|\Lambda^{\frac{5}{2}+2\varepsilon_0}\kappa^{(R)}\|\|\Lambda^{\frac{3}{2}+2\varepsilon_0+\sigma_1}Z_1\|\|\Lambda^{1+\sigma_2} Z_1\|+C\|\Lambda^{\frac{5}{2}+2\varepsilon_0}\kappa^{(R)}\|\|\Lambda^{1+s_1}Z_1\|\|\Lambda^{\frac{3}{2}+2\varepsilon_0+s_2} Z_1\|\\ &\leq& \|\Lambda^{\frac{5}{2}+2\varepsilon_0}\kappa^{(R)}\|\|\Lambda^{\frac{3}{2}+2\varepsilon_0}Z_1\|\|\Lambda^{2} Z_1\| \\ &\leq & \varepsilon\|\Lambda^{2+2\varepsilon_0}\kappa^{(R)}\|^2+C\|\Lambda^{\frac{3}{2}+2\varepsilon_0}Z_1\|^2\|\Lambda^{2} Z_1\|^2,\end{aligned}$$ where $\sigma_1+\sigma_2=1$, $s_1+s_2=1$, we choose $$\sigma_1=0, \quad \sigma_2=1,\quad s_1=1 ,\quad s_2=0.$$ For $I_3$, $$\begin{aligned} &&\Big\|\int_{\mathbb{T}^3}\Lambda^{3+4\varepsilon_0}\kappa^{(R)} [f{k}\times (\kappa^{(R)}+Z_1)+\nabla_H p_{b}-\int^{z}_{-1}\nabla_H (g^{(R)}+Z_2)dz'] dxdydz\Big\|\\ &=&\Big\|\int_{\mathbb{T}^3}\Lambda^{\frac{5}{2}+2\varepsilon_0}\kappa^{(R)} \Lambda^{\frac{1}{2}+2\varepsilon_0}[f{k}\times (\kappa^{(R)}+Z_1)+\nabla_H p_{b}-\int^{z}_{-1}\nabla_H (g^{(R)}+Z_2)dz'] dxdydz\Big\|\\ &\leq & \varepsilon\|\Lambda^{\frac{5}{2}+2\varepsilon_0}\kappa^{(R)}\|^2+C\|\Lambda^{\frac{1}{2}+2\varepsilon_0}\kappa^{(R)}\|^2+C\|\Lambda^{\frac{1}{2}+2\varepsilon_0}Z_1\|^2 +C\|\Lambda^{\frac{3}{2}+2\varepsilon_0}g^{(R)}\|^2+C\|\Lambda^{\frac{3}{2}+2\varepsilon_0}Z_2\|^2,\end{aligned}$$ thus, $$\begin{aligned} \label{eq-20} &&\frac{1}{2}\frac{d}{dt}\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\kappa^{(R)}\|^2+\|\Lambda^{\frac{5}{2}+2\varepsilon_0}\kappa^{(R)}\|^2\\ \notag &\leq& \varepsilon\|\Lambda^{\frac{5}{2}+2\varepsilon_0}\kappa^{(R)}\|^2 +C(R+\|\Lambda^{\frac{3}{2}+2\varepsilon_0}Z_1\|^{2+\frac{4}{1+4\varepsilon_0}})+C(R+\|\Lambda^2 Z_1\|^2)\|\Lambda^{\frac{3}{2}+2\varepsilon_0}Z_1\|^2+\|\Lambda^{\frac{3}{2}+2\varepsilon_0}Z_2\|^2. \notag\end{aligned}$$ Multiplying (\[eq-19\]) by $-\Lambda^{3+4\varepsilon_0} g^{(R)}$, integrating over $\mathbb{T}^3$, it follows that $$\begin{aligned} &&\frac{1}{2}\frac{d}{dt}\|\Lambda^{\frac{3}{2}+2\varepsilon_0} g^{(R)}\|^2+\|\Lambda^{\frac{5}{2}+2\varepsilon_0} g^{(R)}\|^2\\ & =&\chi_R(\|U^{(R)}\|^2_{\mathcal{W}})\int_{\mathbb{T}^3}\Lambda^{3+4\varepsilon_0}g^{(R)}[\Big((\kappa^{(R)}+ Z_1)\cdot \nabla_H\Big)(g^{(R)}+ Z_2)]dxdydz\\ &+&\chi_R(\|U^{(R)}\|^2_{\mathcal{W}})\int_{\mathbb{T}^3}\Lambda^{3+4\varepsilon_0}g^{(R)}[\Phi(\kappa^{(R)}+Z_1)\frac{\partial( g^{(R)}+Z_2)}{\partial z}]dxdydz\\ &+&\chi_R(\|U^{(R)}\|^2_{\mathcal{W}})\int_{\mathbb{T}^3}\Lambda^{3+4\varepsilon_0}g^{(R)}\Phi(\kappa^{(R)}+Z_1)dxdydz\\ &:=& \chi_R(\|U^{(R)}\|^2_{\mathcal{W}})(I_4+I_5+I_6).\end{aligned}$$ For $I_4$, we have $$\begin{aligned} &&\Big\|\int_{\mathbb{T}^3}\Lambda^{3+4\varepsilon_0}g^{(R)}[(\kappa^{(R)}\cdot \nabla_H)g^{(R)}]dxdydz\Big\|\\ &=& \Big\|\int_{\mathbb{T}^3}\Lambda^{\frac{5}{2}+2\varepsilon_0}g^{(R)}\Lambda^{\frac{1}{2}+2\varepsilon_0}[(\kappa^{(R)}\cdot \nabla_H)g^{(R)}]dxdydz\Big\|\\ &\leq& C\|\Lambda^{\frac{5}{2}+2\varepsilon_0}g^{(R)}\|\|\Lambda^{\frac{1}{2}+2\varepsilon_0+\sigma_1}\kappa^{(R)}\|\|\Lambda^{1+\sigma_2}g^{(R)}\| +C\|\Lambda^{\frac{5}{2}+2\varepsilon_0}g^{(R)}\|\|\Lambda^{s_1}\kappa^{(R)}\|\|\Lambda^{\frac{3}{2}+2\varepsilon_0+s_2}g^{(R)}\|\\ &\leq& C\|\Lambda^{\frac{5}{2}+2\varepsilon_0}g^{(R)}\|\|\Lambda^{\frac{3}{2}+2\varepsilon_0}g^{(R)}\|\|\Lambda^{\frac{3}{2}}\kappa^{(R)}\|\\ &\leq& \varepsilon \|\Lambda^{\frac{5}{2}+2\varepsilon_0}g^{(R)}\|^2+C\|\Lambda^{\frac{3}{2}+2\varepsilon_0}g^{(R)}\|^2\|\Lambda^{\frac{3}{2}}\kappa^{(R)}\|^2,\end{aligned}$$ where $\sigma_1+\sigma_2=\frac{3}{2}$, $s_1+s_2=\frac{3}{2}$, we choose $$\sigma_1=1-2\varepsilon_0, \quad \sigma_2=\frac{1}{2}+2\varepsilon_0, \quad s_1=\frac{3}{2}, \quad s_2=0.$$ By the same argument, we have $$\begin{aligned} \Big\|\int_{\mathbb{T}^3}\Lambda^{3+4\varepsilon_0}g^{(R)}[(\kappa^{(R)}\cdot \nabla_H)Z_2]dxdydz\Big\| &\leq& \varepsilon \|\Lambda^{\frac{5}{2}+2\varepsilon_0}g^{(R)}\|^2+C\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\kappa^{(R)}\|^2\|\Lambda^{\frac{3}{2}}Z_2\|^2 +C\|\Lambda^{\frac{3}{2}+2\varepsilon_0}Z_2\|^2\|\Lambda^{\frac{3}{2}}\kappa^{(R)}\|^2,\\ \Big\|\int_{\mathbb{T}^3}\Lambda^{3+4\varepsilon_0}g^{(R)}[(Z_1\cdot \nabla_H)g^{(R)}]dxdydz\Big\| &\leq& \varepsilon \|\Lambda^{\frac{5}{2}+2\varepsilon_0}g^{(R)}\|^2+C\|\Lambda^{\frac{3}{2}+2\varepsilon_0}g^{(R)}\|^2\|\Lambda^{\frac{3}{2}}Z_1\|^2,\\ \Big\|\int_{\mathbb{T}^3}\Lambda^{3+4\varepsilon_0}g^{(R)}[(Z_1\cdot \nabla_H)Z_2]dxdydz\Big\| &\leq& \varepsilon \|\Lambda^{\frac{5}{2}+2\varepsilon_0}g^{(R)}\|^2+C\|\Lambda^{\frac{3}{2}+2\varepsilon_0}Z_2\|^2\|\Lambda^{\frac{3}{2}}Z_1\|^2.\end{aligned}$$ For $I_5$, similar to $I_4$, $$\begin{aligned} \Big\|\int_{\mathbb{T}^3}\Lambda^{3+4\varepsilon_0}g^{(R)}[\Phi(\kappa^{(R)})\frac{\partial g^{(R)}}{\partial z}]dxdydz\Big\| &\leq& \varepsilon \|\Lambda^{\frac{5}{2}+2\varepsilon_0}g^{(R)}\|^2+C\|\Lambda^{\frac{3}{2}+2\varepsilon_0}g^{(R)}\|^2\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\kappa^{(R)}\|^{\frac{4}{1+4\varepsilon_0}},\\ \Big\|\int_{\mathbb{T}^3}\Lambda^{3+4\varepsilon_0}g^{(R)}[\Phi(\kappa^{(R)})\frac{\partial Z_2}{\partial z}]dxdydz\Big\| &\leq& \varepsilon \|\Lambda^{\frac{5}{2}+2\varepsilon_0}g^{(R)}\|^2+C\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\kappa^{(R)}\|^2\|\Lambda^{2}Z_2\|^{2},\\ \Big\|\int_{\mathbb{T}^3}\Lambda^{3+4\varepsilon_0}g^{(R)}[\Phi(Z_1)\frac{\partial g^{(R)}}{\partial z}]dxdydz\Big\| &\leq& \varepsilon \|\Lambda^{\frac{5}{2}+2\varepsilon_0}g^{(R)}\|^2+C\|\Lambda^{\frac{3}{2}+2\varepsilon_0}g^{(R)}\|^2\|\Lambda^{\frac{3}{2}+2\varepsilon_0}Z_1\|^{\frac{4}{1+4\varepsilon_0}},\\ \Big\|\int_{\mathbb{T}^3}\Lambda^{3+4\varepsilon_0}g^{(R)}[\Phi(Z_1)\frac{\partial Z_2}{\partial z}]dxdydz\Big\| &\leq& \varepsilon \|\Lambda^{\frac{5}{2}+2\varepsilon_0}g^{(R)}\|^2+C\|\Lambda^{\frac{3}{2}+2\varepsilon_0}Z_1\|^2\|\Lambda^{2}Z_2\|^{2}.\end{aligned}$$ For $I_6$, we have $$\begin{aligned} &&\Big\|\int_{\mathbb{T}^3}\Lambda^{3+4\varepsilon_0}g^{(R)}\Phi(\kappa^{(R)+Z_1})dxdydz\Big\|\\ &\leq& C\|\Lambda^{\frac{5}{2}+2\varepsilon_0}g^{(R)}\|\|\Lambda^{\frac{3}{2}+2\varepsilon_0}(\kappa^{(R)}+Z_1)\|\\ &\leq& \varepsilon \|\Lambda^{\frac{5}{2}+2\varepsilon_0}g^{(R)}\|^2 +C\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\kappa^{(R)}\|^2+C\|\Lambda^{\frac{3}{2}+2\varepsilon_0}Z_1\|^{2}.\end{aligned}$$ Thus, $$\begin{aligned} \label{eq-21} &&\frac{1}{2}\frac{d}{dt}\|\Lambda^{\frac{3}{2}+2\varepsilon_0}g^{(R)}\|^2+\|\Lambda^{\frac{5}{2}+2\varepsilon_0}g^{(R)}\|^2 \\ \notag & \leq& \varepsilon\|\Lambda^{\frac{5}{2}+2\varepsilon_0}g^{(R)}\|^2 +C(R+\|\Lambda^{\frac{3}{2}+2\varepsilon_0}Z_1\|^{\frac{4}{1+4\varepsilon_0}}\|\Lambda^{\frac{3}{2}+2\varepsilon_0}Z_2\|^2 +\|\Lambda^{\frac{3}{2}+2\varepsilon_0}Z_1\|^{2}\|\Lambda^{2}Z_2\|^2). \notag\end{aligned}$$ Since $\|X^{(R)}\|^2_{\mathcal{W}}=\|\kappa^{(R)}\|^2_{\mathcal{W}}+\|g^{(R)}\|^2_{\mathcal{W}}$, combining (\[eq-20\]) and (\[eq-21\]), we have $$\begin{aligned} \label{eq-26} \frac{d\|X^{(R)}\|^2_{\mathcal{W}}}{dt}+\\|X^{(R)}\\|^2_{\frac{5}{2}+2\varepsilon_0}\leq C(R+\|\Lambda^{2}Z\|^{2+\frac{4}{1+4\varepsilon_0}}),\end{aligned}$$ by the property of $Z$ in (\[eq-22\]), (\[eq-17\]) has a weak solution $X^{(R)}$ in $L^{\infty}([0,T];\mathcal{W})\cap L^2([0,T]; D(A^{\frac{5}{4}+\varepsilon_0}))$. (Continuity of weak solution) Multiplying (\[eq-18\]) by $-\Lambda^{1+4\varepsilon_0} \frac{ d \kappa^{(R)}}{dt}$, integrating over $\mathbb{T}^3$, it follows that $$\begin{aligned} &&\frac{1}{2}\frac{d}{dt}\int_{\mathbb{T}^3}\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\kappa^{(R)}\|^2dxdydz+\| \Lambda^{\frac{1}{2}+2\varepsilon_0} \dot{\kappa}^{(R)}\|^2\\ &=&\chi_R(\|U^{(R)}\|^2_{\mathcal{W}})\int_{\mathbb{T}^3}\Big[\Big((\kappa^{(R)}+ Z_1)\cdot \nabla_H\Big)(\kappa^{(R)}+ Z_1)\Big] \Lambda^{1+4\varepsilon_0} \dot{\kappa}^{(R)}dxdydz\\ \notag &+&\chi_R(\|U^{(R)}\|^2_{\mathcal{W}})\int_{\mathbb{T}^3}\Big[\Phi(\kappa^{(R)}+Z_1)\frac{\partial( \kappa^{(R)}+ Z_1)}{\partial z}\Big] \Lambda^{1+4\varepsilon_0} \dot{\kappa}^{(R)}dxdydz\\ \notag &+&\chi_R(\|U^{(R)}\|^2_{\mathcal{W}})\int_{\mathbb{T}^3}\Big(f{k}\times (\kappa^{(R)}+Z_1)+\nabla_H p_{b}-\int^{z}_{-1}\nabla_H (g^{(R)}+Z_2)dz'\Big) \Lambda^{1+4\varepsilon_0} \dot{\kappa}^{(R)}dxdydz\\ &:=& \chi_R(\|U^{(R)}\|^2_{\mathcal{W}})(J_1+J_2+J_3), \end{aligned}$$ where $\dot{\kappa}^{(R)}$ denotes $\frac{d \kappa^{(R)}}{dt}$, and this symbol always denotes the deviation with respect to $t$. For $J_1$, By Hölder inequality and the Young inequality, we have $$\begin{aligned} &&\Big\|\int_{\mathbb{T}^3}[(\kappa^{(R)}\cdot \nabla_H)\kappa^{(R)}] \Lambda^{1+4\varepsilon_0} \dot{\kappa}^{(R)}dxdydz\Big\|\\ &\leq&C\|\Lambda^{\frac{1}{2}+2\varepsilon_0} \dot{\kappa}^{(R)}\|\|\Lambda^{\frac{1}{2}+2\varepsilon_0+\sigma_1} \kappa^{(R)}\|\|\Lambda^{1+\sigma_2}\kappa^{(R)}\|+C\|\Lambda^{\frac{1}{2}+2\varepsilon_0} \dot{\kappa}^{(R)}\|\|\Lambda^{s_1} \kappa^{(R)}\|\|\Lambda^{\frac{3}{2}+2\varepsilon_0+s_2}\kappa^{(R)}\|\\ &\leq&C\|\Lambda^{\frac{1}{2}+2\varepsilon_0} \dot{\kappa}^{(R)}\|\|\Lambda^{\frac{3}{2}+2\varepsilon_0} \kappa^{(R)}\|\|\Lambda^{\frac{3}{2}}\kappa^{(R)}\|\\ &\leq&\varepsilon \|\Lambda^{\frac{1}{2}+2\varepsilon_0} \dot{\kappa}^{(R)}\|^2+C\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\kappa^{(R)}\|^4,\end{aligned}$$ similarly, $$\begin{aligned} &&\Big\|\int_{\mathbb{T}^3}[(\kappa^{(R)}\cdot \nabla_H) Z_1] \Lambda^{1+4\varepsilon_0} \dot{\kappa}^{(R)}dxdydz\Big\|\\ &\leq&C\|\Lambda^{\frac{1}{2}+2\varepsilon_0} \dot{\kappa}^{(R)}\|\|\Lambda^{\frac{1}{2}+2\varepsilon_0+\sigma_1} \kappa^{(R)}\|\|\Lambda^{1+\sigma_2}Z_1\|+C\|\Lambda^{\frac{1}{2}+2\varepsilon_0} \dot{\kappa}^{(R)}\|\|\Lambda^{s_1} \kappa^{(R)}\|\|\Lambda^{\frac{3}{2}+2\varepsilon_0+s_2}Z_1\|\\ &\leq&C\|\Lambda^{\frac{1}{2}+2\varepsilon_0} \dot{\kappa}^{(R)}\|\Big(\|\Lambda^{\frac{3}{2}+2\varepsilon_0} \kappa^{(R)}\|\|\Lambda^{\frac{3}{2}}Z_1\|+\|\Lambda^{\frac{3}{2}+2\varepsilon_0} Z_1\|\|\Lambda^{\frac{3}{2}}\kappa^{(R)}\|\Big)\\ &\leq&\varepsilon \|\Lambda^{\frac{1}{2}+2\varepsilon_0} \dot{\kappa}^{(R)}\|^2+C\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\kappa^{(R)}\|^2\|\Lambda^{\frac{3}{2}+2\varepsilon_0}Z_1\|^2,\end{aligned}$$ $$\begin{aligned} &&\Big\|\int_{\mathbb{T}^3}[(Z_1\cdot \nabla_H)\kappa^{(R)}] \Lambda^{1+4\varepsilon_0} \dot{\kappa}^{(R)}dxdydz\Big\|\\ &\leq&C\|\Lambda^{\frac{1}{2}+2\varepsilon_0} \dot{\kappa}^{(R)}\|\|\Lambda^{\frac{1}{2}+2\varepsilon_0+\sigma_1} Z_1\|\|\Lambda^{1+\sigma_2}\kappa^{(R)}\|+C\|\Lambda^{\frac{1}{2}+2\varepsilon_0} \dot{\kappa}^{(R)}\|\|\Lambda^{s_1} Z_1\|\|\Lambda^{\frac{3}{2}+2\varepsilon_0+s_2}\kappa^{(R)}\|\\ &\leq&C\|\Lambda^{\frac{1}{2}+2\varepsilon_0} \dot{\kappa}^{(R)}\| \Big(\|\Lambda^{\frac{3}{2}+2\varepsilon_0}Z_1\|\|\Lambda^{\frac{3}{2}}\kappa^{(R)}\| +\|\Lambda^{\frac{3}{2}}Z_1\|\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\kappa^{(R)}\|\Big) \\ &\leq&\varepsilon \|\Lambda^{\frac{1}{2}+2\varepsilon_0} \dot{\kappa}^{(R)}\|^2+C\|\Lambda^{\frac{3}{2}+2\varepsilon_0}Z_1\|^2\|\Lambda^{\frac{3}{2}}\kappa^{(R)}\|^2 +C\|\Lambda^{\frac{3}{2}}Z_1\|^2\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\kappa^{(R)}\|^2,\end{aligned}$$ and $$\begin{aligned} &&\Big\|\int_{\mathbb{T}^3}[(Z_1\cdot \nabla_H)Z_1] \Lambda^{1+4\varepsilon_0} \dot{\kappa}^{(R)}dxdydz\Big\|\\ &\leq&C\|\Lambda^{\frac{1}{2}+2\varepsilon_0} \dot{\kappa}^{(R)}\|\|\Lambda^{\frac{1}{2}+2\varepsilon_0+\sigma_1} Z_1\|\|\Lambda^{1+\sigma_2}Z_1\|+C\|\Lambda^{\frac{1}{2}+2\varepsilon_0} \dot{\kappa}^{(R)}\|\|\Lambda^{s_1} Z_1\|\|\Lambda^{\frac{3}{2}+2\varepsilon_0+s_2}Z_1\|\\ &\leq& C\|\Lambda^{\frac{1}{2}+2\varepsilon_0} \dot{\kappa}^{(R)}\|\|\Lambda^{\frac{3}{2}+2\varepsilon_0}Z_1\|\|\Lambda^{\frac{3}{2}}Z_1\|\\ &\leq&\varepsilon \|\Lambda^{\frac{1}{2}+2\varepsilon_0} \dot{\kappa}^{(R)}\|^2+C\|\Lambda^{\frac{3}{2}+2\varepsilon_0}Z_1\|^2\|\Lambda^{\frac{3}{2}}Z_1\|^2.\end{aligned}$$ For $J_2$, we obtain $$\begin{aligned} &&\Big\|\int_{\mathbb{T}^3}[\Phi(\kappa^{(R)})\frac{\partial \kappa^{(R)}}{\partial z}] \Lambda^{1+4\varepsilon_0} \dot{\kappa}^{(R)}dxdydz\Big\|\\ &\leq&C\|\Lambda^{\frac{1}{2}+2\varepsilon_0} \dot{\kappa}^{(R)}\|\Big(\|\Lambda^{\frac{3}{2}+2\varepsilon_0+\sigma_1} \kappa^{(R)}\|\|\Lambda^{1+\sigma_2}\kappa^{(R)}\|+\|\Lambda^{1+s_1}\kappa^{(R)}\|\|\Lambda^{\frac{3}{2}+2\varepsilon_0+s_2} \kappa^{(R)}\|\Big)\\ &\leq&C\|\Lambda^{\frac{1}{2}+2\varepsilon_0} \dot{\kappa}^{(R)}\|\|\Lambda^{2} \kappa^{(R)}\|\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\kappa^{(R)}\|\\ &\leq& C\|\Lambda^{\frac{1}{2}+2\varepsilon_0} \dot{\kappa}^{(R)}\|\|\Lambda^{\frac{3}{2}+2\varepsilon_0} \kappa^{(R)}\|^{\frac{1}{2}+2\varepsilon_0}\|\Lambda^{\frac{5}{2}+2\varepsilon_0} \kappa^{(R)}\|^{\frac{1}{2}-2\varepsilon_0}\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\kappa^{(R)}\|\\ &\leq&\varepsilon \|\Lambda^{\frac{1}{2}+2\varepsilon_0} \dot{\kappa}^{(R)}\|^2+C\|\Lambda^{\frac{5}{2}+2\varepsilon_0}\kappa^{(R)}\|^2 +C\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\kappa^{(R)}\|^2 \|\Lambda^{\frac{3}{2}+2\varepsilon_0}\kappa^{(R)}\|^{\frac{4}{1+4\varepsilon_0}},\end{aligned}$$ similarly, $$\begin{aligned} &&\Big\|\int_{\mathbb{T}^3}[\Phi(\kappa^{(R)})\frac{\partial Z_1}{\partial z}] \Lambda^{1+4\varepsilon_0} \dot{\kappa}^{(R)}dxdydz\Big\|\\ &\leq&C\|\Lambda^{\frac{1}{2}+2\varepsilon_0} \dot{\kappa}^{(R)}\|\Big(\|\Lambda^{\frac{3}{2}+2\varepsilon_0+\sigma_1} \kappa^{(R)}\|\|\Lambda^{1+\sigma_2}Z_1\|+\|\Lambda^{1+s_1}\kappa^{(R)}\|\|\Lambda^{\frac{3}{2}+2\varepsilon_0+s_2} Z_1\|\Big)\\ &\leq&C\|\Lambda^{\frac{1}{2}+2\varepsilon_0} \dot{\kappa}^{(R)}\|\|\Lambda^{2} Z_1\|\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\kappa^{(R)}\|\\ &\leq&\varepsilon \|\Lambda^{\frac{1}{2}+2\varepsilon_0} \dot{\kappa}^{(R)}\|^2 +C\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\kappa^{(R)}\|^2\|\Lambda^{2}Z_1\|^2,\\ &&\Big\|\int_{\mathbb{T}^3}[\Phi(Z_1)\frac{\partial \kappa^{(R)}}{\partial z}] \Lambda^{1+4\varepsilon_0} \dot{\kappa}^{(R)}dxdydz\Big\|\\ &\leq&C\|\Lambda^{\frac{1}{2}+2\varepsilon_0} \dot{\kappa}^{(R)}\|\|\Lambda^{\frac{3}{2}+2\varepsilon_0} \kappa^{(R)}\|\|\Lambda^{2}Z_1\|\\ &\leq&\varepsilon \|\Lambda^{\frac{1}{2}+2\varepsilon_0} \dot{\kappa}^{(R)}\|^2 +C\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\kappa^{(R)}\|^2\|\Lambda^{2}Z_1\|^2,\\ &&\Big\|\int_{\mathbb{T}^3}[\Phi(Z_1)\frac{\partial Z_1}{\partial z}] \Lambda^{1+4\varepsilon_0} \dot{\kappa}^{(R)}dxdydz\Big\|\\ &\leq&C\|\Lambda^{\frac{1}{2}+2\varepsilon_0} \dot{\kappa}^{(R)}\|\|\Lambda^{\frac{3}{2}+2\varepsilon_0} Z_1\|\|\Lambda^{2}Z_1\|\\ &\leq&\varepsilon \|\Lambda^{\frac{1}{2}+2\varepsilon_0} \dot{\kappa}^{(R)}\|^2+C\|\Lambda^{\frac{3}{2} +2\varepsilon_0}Z_1\|^2\|\Lambda^{2}Z_1\|^2.\end{aligned}$$ For $J_3$, we have $$\begin{aligned} &&\Big\|\int_{\mathbb{T}^3}[f{k}\times (\kappa^{(R)}+Z_1)+\nabla_H p_{b}-\int^{z}_{-1}\nabla_H (g^{(R)}+Z_2)dz'] \Lambda^{1+4\varepsilon_0} \dot{\kappa}^{(R)}dxdydz\Big\|\\ &\leq & \varepsilon \|\Lambda^{\frac{1}{2}+2\varepsilon_0} \dot{\kappa}^{(R)}\|^2+C\|\Lambda^{\frac{1}{2}+2\varepsilon_0}(\kappa^{(R)}+Z_1)\|^2+C\|\Lambda^{\frac{3}{2}+2\varepsilon_0}(g^{(R)}+Z_2)\|^2.\end{aligned}$$ Multiplying (\[eq-19\]) by $-\Lambda^{1+4\varepsilon_0} \frac{d}{dt} g^{(R)}$, integrating over $\mathbb{T}^3$, it follows that $$\begin{aligned} &&\frac{1}{2}\frac{d}{dt}\int_{\mathbb{T}^3}\|\Lambda^{\frac{3}{2}+2\varepsilon_0} g^{(R)}\|^2dxdydz+\|\Lambda^{\frac{1}{2}+2\varepsilon_0}\dot{g}^{(R)}\|^2\\ &=&\chi_R(\|U^{(R)}\|^2_{\mathcal{W}})\int_{\mathbb{T}^3}\Big[\Big((\kappa^{(R)}+ Z_1)\cdot \nabla_H\Big)(g^{(R)}+ Z_2)\Big] \Lambda^{1+4\varepsilon_0}\dot{g}^{(R)}dxdydz,\\ &+&\chi_R(\|U^{(R)}\|^2_{\mathcal{W}})\int_{\mathbb{T}^3}\Big[\Phi(\kappa^{(R)}+Z_1)\frac{\partial( g^{(R)}+Z_2)}{\partial z}\Big]\Lambda^{1+4\varepsilon_0}\dot{g}^{(R)}dxdydz,\\ &+&\chi_R(\|U^{(R)}\|^2_{\mathcal{W}})\int_{\mathbb{T}^3}\Phi(\kappa^{(R)}+Z_1)\Lambda^{1+4\varepsilon_0}\dot{g}^{(R)}dxdydz,\\ &:=& \chi_R(\|U^{(R)}\|^2_{\mathcal{W}})(J_4+J_5+J_6), \end{aligned}$$ where $\dot{g}^{(R)}$ denotes $\frac{d g^{(R)}}{dt}$. For $J_4$, we have $$\begin{aligned} &&\Big\|\int_{\mathbb{T}^3}[(\kappa^{(R)}\cdot \nabla_H)g^{(R)}] \Lambda^{1+4\varepsilon_0}\dot{g}^{(R)}dxdydz\Big\|\\ &\leq& C\|\Lambda^{\frac{1}{2}+2\varepsilon_0} \dot{g}^{(R)}\|\Big(\|\Lambda^{\frac{1}{2}+2\varepsilon_0+\sigma_1}\kappa^{(R)}\|\|\Lambda^{1+\sigma_2}g^{(R)}\| +\|\Lambda^{s_1}\kappa^{(R)}\|\|\Lambda^{\frac{3}{2}+2\varepsilon_0+s_2}g^{(R)}\|\Big)\\ &\leq& C\|\Lambda^{\frac{1}{2}+2\varepsilon_0} \dot{g}^{(R)}\|\|\Lambda^{\frac{3}{2}+2\varepsilon_0} g^{(R)}\|\|\Lambda^{\frac{3}{2}}\kappa^{(R)}\|\\ &\leq & \varepsilon \|\Lambda^{\frac{1}{2}+2\varepsilon_0} \dot{g}^{(R)}\|^2+C\|\Lambda^{\frac{3}{2}+2\varepsilon_0} g^{(R)}\|^2\|\Lambda^{\frac{3}{2}}\kappa^{(R)}\|^2,\end{aligned}$$ similarly, $$\begin{aligned} \Big\|\int_{\mathbb{T}^3}[(\kappa^{(R)}\cdot \nabla_H) Z_2] \Lambda^{1+4\varepsilon_0}\dot{g}^{(R)}dxdydz\Big\| &\leq & \varepsilon \|\Lambda^{\frac{1}{2}+2\varepsilon_0} \dot{g}^{(R)}\|^2+C\|\Lambda^{\frac{3}{2}+2\varepsilon_0}Z_2\|^2\|\Lambda^{\frac{3}{2}}\kappa^{(R)}\|^2,\\ \Big\|\int_{\mathbb{T}^3}[(Z_1\cdot \nabla_H) g^{(R)}] \Lambda^{1+4\varepsilon_0}\dot{g}^{(R)}dxdydz\Big\| &\leq& \varepsilon \|\Lambda^{\frac{1}{2}+2\varepsilon_0} \dot{g}^{(R)}\|^2+C\|\Lambda^{\frac{3}{2}+2\varepsilon_0}g^{(R)}\|^2\|\Lambda^{\frac{3}{2}}Z_1\|^2,\\ \Big\|\int_{\mathbb{T}^3}[(Z_1\cdot \nabla_H) Z_2] \Lambda^{4\varepsilon_0}\dot{g}^{(R)}dxdydz\Big\| &\leq& \varepsilon \|\Lambda^{\frac{1}{2}+2\varepsilon_0} \dot{g}^{(R)}\|^2+C\|\Lambda^{\frac{3}{2}+2\varepsilon_0}Z_2\|^2\|\Lambda^{\frac{3}{2}}Z_1\|^2.\end{aligned}$$ For $J_5$, we have $$\begin{aligned} &&\Big\|\int_{\mathbb{T}^3}[\Phi(\kappa^{(R)})\frac{\partial g^{(R)}}{\partial z}] \Lambda^{1+4\varepsilon_0}\dot{g}^{(R)}dxdydz\Big\|\\ &\leq &\|\Lambda^{\frac{1}{2}+2\varepsilon_0} \dot{g}^{(R)}\|\|\Lambda^{\frac{3}{2}+2\varepsilon_0+\sigma_1} \kappa^{(R)}\|\|\Lambda^{1+\sigma_2} g^{(R)}\|^2+C\|\Lambda^{2\varepsilon_0} \dot{g}^{(R)}\|\|\Lambda^{1+s_1} \kappa^{(R)}\|\|\Lambda^{\frac{3}{2}+2\varepsilon_0+s_2} g^{(R)}\|^2\\ &\leq & C\|\Lambda^{\frac{1}{2}+2\varepsilon_0} \dot{g}^{(R)}\|\|\Lambda^{\frac{3}{2}+2\varepsilon_0} \kappa^{(R)}\|\|\Lambda^{2} g^{(R)}\|^2\\ &\leq & \varepsilon \|\Lambda^{\frac{1}{2}+2\varepsilon_0} \dot{g}^{(R)}\|^2+C\|\Lambda^{\frac{5}{2}+2\varepsilon_0} g^{(R)}\|^2+C\|\Lambda^{\frac{3}{2}+2\varepsilon_0}g^{(R)}\|^2 \|\Lambda^{\frac{3}{2}+2\varepsilon_0}\kappa^{(R)}\|^{\frac{4}{1+4\varepsilon_0}},\end{aligned}$$ similarly, we get $$\begin{aligned} \Big\|\int_{\mathbb{T}^3}[\Phi(\kappa^{(R)})\frac{\partial Z_2}{\partial z}] \Lambda^{1+4\varepsilon_0}\dot{g}^{(R)}dxdydz\Big\| &\leq &\varepsilon \|\Lambda^{\frac{1}{2}+2\varepsilon_0} \dot{g}^{(R)}\|^2+C\|\Lambda^{\frac{5}{2}+2\varepsilon_0} \kappa^{(R)}\|^2+C\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\kappa^{(R)}\|^2 \|\Lambda^{\frac{3}{2}+2\varepsilon_0}Z_2\|^{\frac{4}{1+4\varepsilon_0}},\\ \|\int_{\mathbb{T}^3}[\Phi(Z_1)\frac{\partial Z_2}{\partial z}] \Lambda^{1+4\varepsilon_0}\dot{g}^{(R)}dxdydz\| &\leq & \varepsilon \|\Lambda^{\frac{1}{2}+2\varepsilon_0} \dot{g}^{(R)}\|^2+C\|\Lambda^{\frac{5}{2}+2\varepsilon_0} g^{(R)}\|^2+C\|\Lambda^{\frac{3}{2}+2\varepsilon_0}g^{(R)}\|^2 \|\Lambda^{\frac{3}{2}+2\varepsilon_0}Z_1\|^{\frac{4}{1+4\varepsilon_0}},\\ \|\int_{\mathbb{T}^3}[\Phi(Z_1)\frac{\partial Z_2}{\partial z}] \Lambda^{1+4\varepsilon_0}\dot{g}^{(R)}dxdydz\| &\leq & \varepsilon \|\Lambda^{\frac{1}{2}+2\varepsilon_0} \dot{g}^{(R)}\|^2+C\|\Lambda^{2} Z_1\|^2\|\Lambda^{\frac{3}{2}+2\varepsilon_0}Z_2\|^2.\end{aligned}$$ For $J_6$, we have $$\begin{aligned} &&\Big\|\int_{\mathbb{T}^3}\Phi(\kappa^{(R)}+Z_1) \Lambda^{1+4\varepsilon_0}\dot{g}^{(R)}dxdydz\Big\|\\ &\leq & C\|\Lambda^{\frac{1}{2}+2\varepsilon_0} \dot{g}^{(R)}\|\|\Lambda^{\frac{3}{2}+2\varepsilon_0} (\kappa^{(R)}+Z_1)\|\\ &\leq & \varepsilon \|\Lambda^{\frac{1}{2}+2\varepsilon_0} \dot{g}^{(R)}\|^2+C\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\kappa^{(R)}\|^2 +C\|\Lambda^{\frac{3}{2}+2\varepsilon_0}Z_1\|^{2}.\end{aligned}$$ Combing the above estimations, we obtain $$\begin{aligned} \label{eq-25} &&\frac{d}{dt}\\|X^{(R)}\\|^2_{\frac{3}{2}+2\varepsilon_0}+\\|\dot{X}^{(R)}\\|^2_{\frac{1}{2}+2\varepsilon_0}\\ \notag &\leq& C\chi_R(\|U^{(R)}\|^2_{\mathcal{W}})\Big(C(R)+\|\Lambda^{\frac{5}{2}+2\varepsilon_0} \kappa^{(R)}\|^2+\|\Lambda^{\frac{5}{2}+2\varepsilon_0} g^{(R)}\|^2+\|\Lambda^{2} Z_1\|^{2+\frac{4}{1+4\varepsilon_0}}+\|\Lambda^{2} Z_2\|^{2+\frac{4}{1+4\varepsilon_0}}\Big).\end{aligned}$$ Integrating (\[eq-25\]) on $t$ from $0$ to $T$, as $\int^T_0\|\Lambda^{\frac{5}{2}+2\varepsilon_0} \kappa^{(R)}\|^2dt$ and $\int^T_0\|\Lambda^{\frac{5}{2}+2\varepsilon_0} g^{(R)}\|^2dt$ can be dominated by (\[eq-26\]), by the property of $Z$ in (\[eq-22\]), we get the time derivative $\frac{d{X}^{(R)}}{dt} \in L^2([0,T]; D(A^{\frac{1}{4}+\varepsilon_0}))$. Then by (\[eq-26\]) and [@T-R], we obtain $X^{(R)}\in C([0,T]; \mathcal{W})$. (Uniqueness of weak solution)Let $X_1=(\kappa^{(R)}_1, g^{(R)}_1)$, $X_2=(\kappa^{(R)}_2, g^{(R)}_2)$ be two solutions of (\[eq-17\]) in $C([0,T]; \mathcal{W})$ and set $$\begin{aligned} Y=X_1-X_2=(\kappa,g)=(\kappa^{(R)}_1-\kappa^{(R)}_2,g^{(R)}_1-g^{(R)}_2),\\ v_1=\kappa^{(R)}_1+Z_1, \quad T_1=g^{(R)}_1+Z_2,\quad v_2=\kappa^{(R)}_2+Z_1,\quad T_2=g^{(R)}_2+Z_2.\end{aligned}$$ Firstly, from (\[eq-17\]), we obtain $$\begin{aligned} \frac{dY}{dt}+AY&+&B(U_1,U_1)\Big(\chi_R(\|U_1\|^2_{\mathcal{W}})-\chi_R(\|U_2\|^2_{\mathcal{W}})\Big)+B(U_1,Y)\chi_R(\|U_2\|^2_{\mathcal{W}}) +B(Y,U_2)\chi_R(\|U_2\|^2_{\mathcal{W}})\\ &&+G(U_1)\Big(\chi_R(\|U_1\|^2_{\mathcal{W}})-\chi_R(\|U_2\|^2_{\mathcal{W}})\Big) +\tilde{G}(Y)\chi_R(\|U_2\|^2_{\mathcal{W}})=0,\end{aligned}$$ where $$\tilde{G}(Y)=\left( \begin{array}{c} fk\times \kappa-\int^{z}_{-1}\nabla_H g dz' \\ \Phi(\kappa) \\ \end{array} \right).$$ That is, $$\begin{aligned} \label{eq-27} \frac{d\kappa}{dt}&+&A_1\kappa+\Big((v_1\cdot\nabla_H)v_1+\Phi(v_1)\frac{\partial v_1}{\partial z}\Big)\Big(\chi_R(\|U_1\|^2_{\mathcal{W}})-\chi_R(\|U_2\|^2_{\mathcal{W}})\Big) \\ \notag &&+\Big((v_1\cdot\nabla_H)\kappa+\Phi(v_1)\frac{\partial \kappa}{\partial z}\Big)\chi_R(\|U_2\|^2_{\mathcal{W}})+\Big((\kappa\cdot\nabla_H)v_2+\Phi(\kappa)\frac{\partial v_2}{\partial z})\Big)\chi_R(\|U_2\|^2_{\mathcal{W}}) \\ \notag &&+(fk\times v_1+\nabla_H p_b-\int^z_{-1}\nabla_HT_1dz')\Big(\chi_R(\|U_1\|^2_{\mathcal{W}})-\chi_R(\|U_2\|^2_{\mathcal{W}})\Big) \\ \notag && +(fk\times \kappa-\int^z_{-1}\nabla_H gdz')\chi_R(\|U_2\|^2_{\mathcal{W}})=0,\end{aligned}$$ and $$\begin{aligned} \label{eq-28} \frac{dg}{dt}&+&A_2g+\Big((v_1\cdot\nabla_H)T_1+\Phi(v_1)\frac{\partial T_1}{\partial z}\Big)\Big(\chi_R(\|U_1\|^2_{\mathcal{W}})-\chi_R(\|U_2\|^2_{\mathcal{W}})\Big)\\ \notag &&+\Big((v_1\cdot\nabla_H)g+\Phi(v_1)\frac{\partial g}{\partial z}\Big)\chi_R(\|U_2\|^2_{\mathcal{W}}) +\Big((\kappa\cdot\nabla_H)T_2+\Phi(\kappa)\frac{\partial T_2}{\partial z}\Big)\chi_R(\|U_2\|^2_{\mathcal{W}})\\ \notag &&+\Phi(\kappa)\chi_R(\|U_2\|^2_{\mathcal{W}})=0.\end{aligned}$$ Multiplying (\[eq-27\]) by $-\Lambda^{1+4\varepsilon_0} \kappa$, then integrating over $\mathbb{T}^3$, we have $$\begin{aligned} &&\frac{1}{2}\frac{d}{dt}\int_{\mathbb{T}^3}\|\Lambda^{\frac{1}{2}+2\varepsilon_0} \kappa\|^2dxdydz+\|\Lambda^{\frac{3}{2}+2\varepsilon_0} \kappa\|^2\\ &=&\Big(\chi_R(\|U_1\|^2_{\mathcal{W}})-\chi_R(\|U_2\|^2_{\mathcal{W}})\Big)\int_{\mathbb{T}^3}[(v_1\cdot\nabla_H)v_1+\Phi(v_1)\frac{\partial v_1}{\partial z}]\Lambda^{1+4\varepsilon_0} \kappa dxdydz\\ \notag \quad \quad &&+ \chi_R(\|U_2\|^2_{\mathcal{W}})\int_{\mathbb{T}^3}[(v_1\cdot\nabla_H)\kappa+\Phi(v_1)\frac{\partial \kappa}{\partial z}]\Lambda^{1+4\varepsilon_0} \kappa dxdydz\\ \notag \quad \quad &&+\chi_R(\|U_2\|^2_{\mathcal{W}})\int_{\mathbb{T}^3}[(\kappa\cdot\nabla_H)v_2+\Phi(\kappa)\frac{\partial v_2}{\partial z}]\Lambda^{1+4\varepsilon_0} \kappa dxdydz\\ \notag \quad \quad &&+ \Big(\chi_R(\|U_1\|^2_{\mathcal{W}})-\chi_R(\|U_2\|^2_{\mathcal{W}})\Big)\int_{\mathbb{T}^3}(fk\times v_1+\nabla_H p_b-\int^z_{-1}\nabla_HT_1dz')\Lambda^{1+4\varepsilon_0} \kappa dxdydz\\ \notag \quad \quad &&+\chi_R(\|U_2\|^2_{\mathcal{W}})\int_{\mathbb{T}^3} \Big(fk\times \kappa -\int^z_{-1}\nabla_H gdz'\Big)\Lambda^{1+4\varepsilon_0} \kappa dxdydz\\ \notag &&:=I+II+III+IV+V.\end{aligned}$$ It’s easy to know that $$\|\chi_R(\|U_1\|^2_{\mathcal{W}})-\chi_R(\|U_2\|^2_{\mathcal{W}})\|\leq C(R)\|Y\|_{\mathcal{W}}[I_{[0,R+1]}(\|U_1\|^2_{\mathcal{W}})+I_{[0,R+1]}(\|U_2\|^2_{\mathcal{W}})].$$ For $I$, since $\int_{\mathbb{T}^3}((v_1\cdot\nabla_H)v_1\Lambda^{1+4\varepsilon_0} \kappa dxdydz$ is weaker than $\int_{\mathbb{T}^3}(\Phi(v_1)\frac{\partial v_1}{\partial z})\Lambda^{1+4\varepsilon_0} \kappa dxdydz$, we only need to estimate the term involved $\Phi$. For $\varepsilon_1\in(0,2\varepsilon_0)$, we have $$\begin{aligned} &&C(R)\|Y\|_{\mathcal{W}}\Big\|\int_{\mathbb{T}^3}[\Phi(v_1)\frac{\partial v_1}{\partial z}]\Lambda^{1+4\varepsilon_0} \kappa dxdydz\Big\|\\ &\leq& C(R)\|Y\|_{\mathcal{W}}\Big\|\int_{\mathbb{T}^3}\Lambda^{2\varepsilon_0-\frac{1}{2}+\varepsilon_1}[\Phi(v_1)\frac{\partial v_1}{\partial z}]\Lambda^{\frac{3}{2}+2\varepsilon_0-\varepsilon_1} \kappa dxdydz\Big\|\\ &\leq& C(R)\|Y\|_{\mathcal{W}} \Big(\|\Lambda^{2\varepsilon_0+\frac{1}{2}+\varepsilon_1+\sigma_1}v_1\|\|\Lambda^{1+\sigma_2}v_1\|\|\Lambda^{\frac{3}{2}+2\varepsilon_0-\varepsilon_1}\kappa\|+ \|\Lambda^{1+s_1}v_1\|\|\Lambda^{2\varepsilon_0+\frac{1}{2}+\varepsilon_1+s_2}v_1\|\|\Lambda^{\frac{3}{2}+2\varepsilon_0-\varepsilon_1}\kappa\| \Big)\\ &\leq& C(R)\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\kappa\|\|\Lambda^{1+\varepsilon_1}v_1\|\|\Lambda^{\frac{3}{2}+2\varepsilon_0}v_1\|\|\Lambda^{\frac{3}{2}+2\varepsilon_0-\varepsilon_1}\kappa\| +\|\Lambda^{\frac{3}{2}+2\varepsilon_0}g\|\|\Lambda^{1+\varepsilon_1}v_1\|\|\Lambda^{\frac{3}{2}+2\varepsilon_0}v_1\|\|\Lambda^{\frac{3}{2}+2\varepsilon_0-\varepsilon_1}\kappa\| \\ &\leq& C(R)\Big(\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\kappa\|+\|\Lambda^{\frac{3}{2}+2\varepsilon_0}g\|\Big)\|\Lambda^{\frac{1}{2}+2\varepsilon_0}\kappa\|^{\varepsilon_1} \|\Lambda^{\frac{3}{2}+2\varepsilon_0}\kappa\|^{1-\varepsilon_1}\|\Lambda^{\frac{3}{2}+2\varepsilon_0}v_1\|^2 \\ &\leq& \varepsilon\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\kappa\|^2+\varepsilon\|\Lambda^{\frac{3}{2}+2\varepsilon_0}g\|^2 +C(R)\|\Lambda^{\frac{3}{2}+2\varepsilon_0}v_1\|^{\frac{4}{\varepsilon_1}}\|\Lambda^{\frac{1}{2}+2\varepsilon_0}\kappa\|^2\\ &\leq& \varepsilon\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\kappa\|^2+\varepsilon\|\Lambda^{\frac{3}{2}+2\varepsilon_0}g\|^2 +C(R,\|U_1\|^2_{\mathcal{W}})\|\Lambda^{\frac{1}{2}+2\varepsilon_0}\kappa\|^2,\end{aligned}$$ where $\sigma_1+\sigma_2=1$, $s_1+s_2=1$, and we choose $$\sigma_1=s_2=\frac{1}{2}-2\varepsilon_0,\quad \sigma_2=s_1=\frac{1}{2}+2\varepsilon_0.$$ For $II$, $$\begin{aligned} &&\Big\|\int_{\mathbb{T}^3}[\Phi(v_1)\frac{\partial \kappa}{\partial z}]\Lambda^{1+4\varepsilon_0} \kappa dxdydz\Big\|\\ &\leq&\Big\|\int_{\mathbb{T}^3}\Lambda^{2\varepsilon_0-\frac{1}{2}} [\Phi(v_1)\frac{\partial \kappa}{\partial z}]\Lambda^{\frac{3}{2}+2\varepsilon_0} \kappa dxdydz\Big\|\\ &\leq&C\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\kappa\|\|\Lambda^{2\varepsilon_0+\frac{1}{2}+\sigma_1}v_1\|\|\Lambda^{1+\sigma_2}\kappa\| +C\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\kappa\|\|\Lambda^{1+s_1}v_1\|\|\Lambda^{2\varepsilon_0+\frac{1}{2}+s_2}\kappa\|\\ &\leq& C\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\kappa\|\|\Lambda^{\frac{3}{2}+2\varepsilon_0}v_1\|\|\Lambda \kappa\|\\ &\leq& C\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\kappa\| \|\Lambda^{\frac{3}{2}+\varepsilon_0}v_1\|\|\Lambda^{\frac{1}{2}+2\varepsilon_0}\kappa\|^{\frac{1}{2}+2\varepsilon_0}\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\kappa\|^{\frac{1}{2}-2\varepsilon_0}\\ &\leq& \varepsilon\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\kappa\|^2+C\|\Lambda^{\frac{3}{2}+2\varepsilon_0}v_1\|^{\frac{4}{1+4\varepsilon_0}}\|\Lambda^{\frac{1}{2}+2\varepsilon_0}\kappa\|^2\\ &\leq& \varepsilon\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\kappa\|^2+C(R,\|U_1\|^2_{\mathcal{W}})\|\Lambda^{\frac{1}{2}+2\varepsilon_0}\kappa\|^2,\end{aligned}$$ where $\sigma_1+\sigma_2=1$, $s_1+s_2=1$, and we choose $$\sigma_1=1,\quad \sigma_2=0,\quad s_1=\frac{1}{2}+2\varepsilon_0,\quad s_2=\frac{1}{2}-2\varepsilon_0.$$ For $III$, $$\begin{aligned} &&\Big\|\int_{\mathbb{T}^3}(\Phi(\kappa)\frac{\partial v_2}{\partial z})\Lambda^{1+4\varepsilon_0} \kappa dxdydz\Big\|\\ &\leq&\int_{\mathbb{T}^3}\Lambda^{2\varepsilon_0-\frac{1}{2}} [\Phi(\kappa)\frac{\partial v_2}{\partial z}]\Lambda^{\frac{3}{2}+2\varepsilon_0} \kappa dxdydz\\ &\leq&C\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\kappa\|\|\Lambda^{2\varepsilon_0+\frac{1}{2}+\sigma_1}\kappa\|\|\Lambda^{1+\sigma_2}v_2\| +C\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\kappa\|\|\Lambda^{1+s_1}\kappa\|\|\Lambda^{2\varepsilon_0+\frac{1}{2}+s_2}v_2\|\\ &\leq& C\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\kappa\|\|\Lambda \kappa\|\|\Lambda^{\frac{3}{2}+2\varepsilon_0}v_2\|\\ &\leq& \varepsilon\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\kappa\|^2+C(R,\|U_2\|^2_{\mathcal{W}})\|\Lambda^{\frac{1}{2}+2\varepsilon_0}\kappa\|^2,\end{aligned}$$ where $\sigma_1+\sigma_2=1$, $s_1+s_2=1$, and we choose $$\sigma_1=\frac{1}{2}-2\varepsilon_0,\quad \sigma_2=\frac{1}{2}+2\varepsilon_0,\quad s_1=0,\quad s_2=1.$$ For $IV$, $$\begin{aligned} &&\Big(\chi_R(\|U_1\|^2_{\mathcal{W}})-\chi_R(\|U_2\|^2_{\mathcal{W}})\Big)\Big\|\int_{\mathbb{T}^3}[fk\times v_1+\nabla_H p_b-\int^z_{-1}\nabla_HT_1dz']\Lambda^{1+4\varepsilon_0} \kappa dxdydz\Big\|\\ &\leq& C(R)[I_{[0,R+1]}(\|U_1\|^2_{\mathcal{W}})+I_{[0,R+1]}(\|U_2\|^2_{\mathcal{W}})]\|Y\|_{\mathcal{W}}\Big\|\int_{\mathbb{T}^3}\Lambda^{2\varepsilon_0-\frac{1}{2}+\varepsilon_1}[fk\times v_1\\ && +\nabla_H p_b-\int^z_{-1}\nabla_HT_1dz']\Lambda^{\frac{3}{2}+2\varepsilon_0-\varepsilon_1} \kappa dxdydz\Big\|\\ &\leq&C(R)[I_{[0,R+1]}(\|U_1\|^2_{\mathcal{W}})+I_{[0,R+1]}(\|U_2\|^2_{\mathcal{W}})]\|Y\|_{\mathcal{W}} \|\Lambda^{\frac{3}{2}+2\varepsilon_0-\varepsilon_1}\kappa\|\|\Lambda^{2\varepsilon_0-\frac{1}{2}+\varepsilon_1}v_1\|\\ && +C(R)[I_{[0,R+1]}(\|U_1\|^2_{\mathcal{W}})+I_{[0,R+1]}(\|U_2\|^2_{\mathcal{W}})]\|Y\|_{\mathcal{W}} \|\Lambda^{\frac{3}{2}+2\varepsilon_0-\varepsilon_1}\kappa\|\|\Lambda^{2\varepsilon_0+\frac{1}{2}+\varepsilon_1}T_1\|\\ &:=&C(R)[I_{[0,R+1]}(\|U_1\|^2_{\mathcal{W}})+I_{[0,R+1]}(\|U_2\|^2_{\mathcal{W}})](IV^{(1)}+IV^{(2)}),\end{aligned}$$ for $IV^{(1)}$, we have $$\begin{aligned} IV^{(1)}&=&\|Y\|_{\mathcal{W}} \|\Lambda^{\frac{3}{2}+2\varepsilon_0-\varepsilon_1}\kappa\|\|\Lambda^{2\varepsilon_0-\frac{1}{2}+\varepsilon_1}v_1\|\\ &\leq& C\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\kappa\|^{2-\varepsilon_1}\|\Lambda^{\frac{1}{2}+2\varepsilon_0}\kappa\|^{\varepsilon_1}\|\Lambda^{2\varepsilon_0-\frac{1}{2}+\varepsilon_1}v_1\| +\|\Lambda^{\frac{3}{2}+2\varepsilon_0}g\|\|\Lambda^{\frac{1}{2}+2\varepsilon_0}\kappa\|^{\varepsilon_1}\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\kappa\|^{1-\varepsilon_1}\|\Lambda^{2\varepsilon_0-\frac{1}{2}+\varepsilon_1}v_1\|\\ &\leq& \varepsilon\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\kappa\|^2+\varepsilon\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\kappa\|^2+C\|\Lambda^{\frac{1}{2}+2\varepsilon_0}\kappa\|^2\|\Lambda^{2\varepsilon_0-\frac{1}{2}+\varepsilon_1}v_1\|^{\frac{2}{\varepsilon_1}}\\ &\leq& \varepsilon\|\Lambda^{1+2\varepsilon_0}\kappa\|^2+\varepsilon\|\Lambda^{1+2\varepsilon_0}\kappa\|^2+ C(R,\|U_1\|^2_{\mathcal{W}})\|\Lambda^{\frac{1}{2}+2\varepsilon_0}\kappa\|^{2}\end{aligned}$$ $IV^{(2)}$ is similar to $IV^{(1)}$, we have $$\begin{aligned} IV^{(2)}&=&\|Y\|_{\mathcal{W}} \|\Lambda^{\frac{3}{2}+2\varepsilon_0-\varepsilon_1}\kappa\|\|\Lambda^{\frac{1}{2}+2\varepsilon_0+\varepsilon_1}T_1\|\\ &\leq& \varepsilon\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\kappa\|^2+\varepsilon\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\kappa\|^2+C\|\Lambda^{\frac{1}{2}+2\varepsilon_0}\kappa\|^2\|\Lambda^{2\varepsilon_0+\frac{1}{2}+\varepsilon_1}T_1\|^{\frac{2}{\varepsilon_1}}\\ &\leq& \varepsilon\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\kappa\|^2+\varepsilon\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\kappa\|^2+C(R,\|U_1\|^2_{\mathcal{W}})\|\Lambda^{\frac{1}{2}+2\varepsilon_0}\kappa\|^{2}.\end{aligned}$$ For $V$, we have $$\begin{aligned} &&\chi_{R}(\|U_2\|^2_{\mathcal{W}})\Big\|\int_{\mathbb{T}^3}\Big(fk\times \kappa-\int^z_{-1}\nabla_H g dz'\Big)\Lambda^{1+4\varepsilon_0} \kappa dxdydz\Big\|\\ &\leq& \chi_{R}(\|U_2\|^2_{\mathcal{W}})\int_{\mathbb{T}^3}\Lambda^{2\varepsilon_0-\frac{1}{2}}[fk\times \kappa-\int^z_{-1}\nabla_H g dz']\Lambda^{\frac{3}{2}+2\varepsilon_0} \kappa dxdydz\\ &\leq& \varepsilon\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\kappa\|^2+C\|\Lambda^{2\varepsilon_0-\frac{1}{2}}\kappa\|^2+C\|\Lambda^{\frac{1}{2}+2\varepsilon_0}g\|^2,\end{aligned}$$ thus, we have $$\begin{aligned} \label{eq-33} &&\frac{1}{2}\frac{d}{dt}\int_{\mathbb{T}^3}\|\Lambda^{\frac{1}{2}+2\varepsilon_0} \kappa\|^2dxdydz+\|\Lambda^{\frac{3}{2}+2\varepsilon_0} \kappa\|^2\\ \notag &\leq& \varepsilon\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\kappa\|^2 +C(R,\|U_1\|_{\mathcal{W}},\|U_2\|_{\mathcal{W}})(1+\|\Lambda^{\frac{1}{2}+2\varepsilon_0} \kappa\|^2+\|\Lambda^{\frac{1}{2}+2\varepsilon_0} g\|^2).\notag\end{aligned}$$ Multiplying (\[eq-28\]) by $-\Lambda^{1+4\varepsilon_0} g$, then integrating over $\mathbb{T}^3$, we have $$\begin{aligned} &&\frac{1}{2}\frac{d}{dt}\int_{\mathbb{T}^3}\|\Lambda^{\frac{1}{2}+2\varepsilon_0} g\|^2dxdydz+\|\Lambda^{\frac{3}{2}+2\varepsilon_0} g\|^2\\ &&= \Big(\chi_R(\|U_1\|^2_{\mathcal{W}})-\chi_R(\|U_2\|^2_{\mathcal{W}})\Big)\int_{\mathbb{T}^3} \Big[(v_1\cdot\nabla_H)T_1+\Phi(v_1)\frac{\partial T_1}{\partial z}\Big]\Lambda^{1+4\varepsilon_0} g dxdydz\\ &&+\chi_R(\|U_2\|^2_{\mathcal{W}})\int_{\mathbb{T}^3}\Big[(v_1\cdot\nabla_H)g+\Phi(v_1)\frac{\partial g}{\partial z}\Big]\Lambda^{1+4\varepsilon_0}g dxdydz\\ &&+\chi_R(\|U_2\|^2_{\mathcal{W}})\int_{\mathbb{T}^3}\Big[(\kappa\cdot\nabla_H)T_2+\Phi(\kappa)\frac{\partial T_2}{\partial z}\Big]\Lambda^{1+4\varepsilon_0} g dxdydz\\ &&+\chi_R(\|U_2\|^2_{\mathcal{W}})\int_{\mathbb{T}^3}\Phi(\kappa)\Lambda^{1+4\varepsilon_0} g dxdydz\\ &&:= VI+VII+VIII+IX. \end{aligned}$$ For $VI$, similar to $I$, we only need to estimate $$\begin{aligned} &&\|Y\|_{\mathcal{W}}\Big\|\int_{\mathbb{T}^3}\Big [\Phi(v_1)\frac{\partial T_1}{\partial z}\Big]\Lambda^{1+4\varepsilon_0} g dxdydz\Big\| \\ &=& \|Y\|_{\mathcal{W}}\Big\|\int_{\mathbb{T}^3} \Lambda^{2\varepsilon_0-\frac{1}{2}+\varepsilon_1}[\Phi(v_1)\frac{\partial T_1}{\partial z}]\Lambda^{\frac{3}{2}+2\varepsilon_0-\varepsilon_1} g dxdydz\Big\|\\ &\leq& \|Y\|_{\mathcal{W}}\|\Lambda^{\frac{3}{2}+2\varepsilon_0-\varepsilon_1} g\|\Big(\|\Lambda^{\frac{1}{2}+2\varepsilon_0+\varepsilon_1+\sigma_1} v_1\|\|\Lambda^{1+\sigma_2} T_1\|+ \|\Lambda^{1+s_1} v_1\|\|\Lambda^{\frac{1}{2}+2\varepsilon_0+\varepsilon_1+s_2}T_1\|\Big)\\ &\leq& \|Y\|_{\mathcal{W}}\|\Lambda^{\frac{3}{2}+2\varepsilon_0-\varepsilon_1} g\|\|\Lambda^{1+\varepsilon_1} v_1\|\|\Lambda^{\frac{3}{2}+2\varepsilon_0} T_1\|\\ &\leq& C\Big(\|\Lambda^{\frac{3}{2}+2\varepsilon_0} \kappa\|+\|\Lambda^{\frac{3}{2}+2\varepsilon_0} g\|\Big)\|\Lambda^{\frac{1}{2}+2\varepsilon_0} g\|^{\varepsilon_1}\|\Lambda^{\frac{3}{2}+2\varepsilon_0} g\|^{1-\varepsilon_1}\|\Lambda^{1+\varepsilon_1} v_1\|\|\Lambda^{\frac{3}{2}+2\varepsilon_0} T_1\|\\ &\leq& \varepsilon\|\Lambda^{\frac{3}{2}+2\varepsilon_0} g\|^2+\varepsilon\|\Lambda^{\frac{3}{2}+2\varepsilon_0} \kappa\|^2+C\|\Lambda^{1+2\varepsilon_0} v_1\|^{\frac{2}{\varepsilon_1}}\|\Lambda^{\frac{3}{2}+2\varepsilon_0} T_1\|^{\frac{2}{\varepsilon_1}}\|\Lambda^{\frac{1}{2}+2\varepsilon_0}g\|^2,\end{aligned}$$ where $\sigma_1+\sigma_2=1$, $s_1+s_2=1$, and we choose $$\sigma_1=\frac{1}{2}-2\varepsilon_0,\quad \sigma_2=\frac{1}{2}+2\varepsilon_0,\quad s_1=\varepsilon_1,\quad s_2=1-\varepsilon_1.$$ For $VII$, $$\begin{aligned} &&\Big\|\int_{\mathbb{T}^3}[\Phi(v_1)\frac{\partial g}{\partial z}]\Lambda^{1+4\varepsilon_0}g dxdydz\Big\|\\ &=&\Big\|\int_{\mathbb{T}^3}\Lambda^{2\varepsilon_0-\frac{1}{2}}[\Phi(v_1)\frac{\partial g}{\partial z}]\Lambda^{\frac{3}{2}+2\varepsilon_0}g dxdydz\Big\|\\ &\leq& C\|\Lambda^{\frac{3}{2}+2\varepsilon_0} g\|\Big(\|\Lambda^{2\varepsilon_0+\frac{1}{2}+\sigma_1} v_1\|\|\Lambda^{1+\sigma_2} g\|+\|\Lambda^{1+s_1} v_1\|\|\Lambda^{2\varepsilon_0+\frac{1}{2}+s_2} g\|\Big)\\ &\leq& C\|\Lambda^{\frac{3}{2}+2\varepsilon_0} g\|\|\Lambda^{1} g\|\|\Lambda^{\frac{3}{2}+2\varepsilon_0} v_1\|\\ &\leq&C\|\Lambda^{\frac{3}{2}+2\varepsilon_0} g\|\|\Lambda^{\frac{1}{2}+2\varepsilon_0} g\|^{\frac{1}{2}+2\varepsilon_0}\|\Lambda^{\frac{3}{2}+2\varepsilon_0} g\|^{\frac{1}{2}-2\varepsilon_0}\|\Lambda^{\frac{3}{2}+2\varepsilon_0} v_1\|\\ &\leq& \varepsilon\|\Lambda^{\frac{3}{2}+2\varepsilon_0} g\|^2+C\|\Lambda^{\frac{3}{2}+2\varepsilon_0} v_1\|^{\frac{4}{1+4\varepsilon_0}}\|\Lambda^{\frac{1}{2}+2\varepsilon_0}g\|^2,\end{aligned}$$ where $\sigma_1+\sigma_2=1$, $s_1+s_2=1$, and we choose $$\sigma_1=1,\quad \sigma_2=0, \quad s_1=\frac{1}{2}+2\varepsilon_0 ,\quad s_2=\frac{1}{2}-2\varepsilon_0.$$ For $VIII$, $$\begin{aligned} &&\Big\|\int_{\mathbb{T}^3}[\Phi(\kappa)\frac{\partial T_2}{\partial z}]\Lambda^{1+4\varepsilon_0} g dxdydz\Big\|\\ &=& \Big\|\int_{\mathbb{T}^3}\Lambda^{2\varepsilon_0-\frac{1}{2}}[\Phi(\kappa)\frac{\partial T_2}{\partial z}]\Lambda^{\frac{3}{2}+2\varepsilon_0} g dxdydz\Big\|\\ &\leq& C\|\Lambda^{\frac{3}{2}+2\varepsilon_0} g\|(\|\Lambda^{\frac{1}{2}+2\varepsilon_0+\sigma_1} \kappa\|\|\Lambda^{1+\sigma_2} T_2\|+\|\Lambda^{1+s_1} \kappa\|\|\Lambda^{\frac{1}{2}+2\varepsilon_0+s_2} T_2\|)\\ &\leq& C\|\Lambda^{\frac{3}{2}+2\varepsilon_0} g\|\|\Lambda^1 \kappa\|\|\Lambda^{\frac{3}{2}+2\varepsilon_0} T_2\|\\ &\leq& C\|\Lambda^{\frac{3}{2}+2\varepsilon_0} g\|\|\Lambda^{\frac{1}{2}+2\varepsilon_0} \kappa\|^{\frac{1}{2}+2\varepsilon_0}\|\Lambda^{\frac{3}{2}+2\varepsilon_0} \kappa\|^{\frac{1}{2}-2\varepsilon_0}\|\Lambda^{\frac{3}{2}+2\varepsilon_0} T_2\|\\ &\leq&\varepsilon\|\Lambda^{\frac{3}{2}+2\varepsilon_0} g\|^2+\varepsilon\|\Lambda^{\frac{3}{2}+2\varepsilon_0} \kappa\|^2+C\|\Lambda^{\frac{3}{2}+2\varepsilon_0} T_2\|^{\frac{4}{1+4\varepsilon_0}}\|\Lambda^{\frac{1}{2}+2\varepsilon_0}\kappa\|^2,\end{aligned}$$ where $\sigma_1+\sigma_2=1$, $s_1+s_2=1$, and we choose $$\sigma_1=\frac{1}{2}-2\varepsilon_0,\quad \sigma_2=\frac{1}{2}+2\varepsilon_0,\quad s_1=0,\quad s_2=1.$$ For $IX$, we have $$\begin{aligned} &&\Big\|\int_{\mathbb{T}^3}\Phi(\kappa)\Lambda^{1+4\varepsilon_0}g dxdydz\Big\|\\ &=&\Big\|\int_{\mathbb{T}^3}\Lambda^{2\varepsilon_0-\frac{1}{2}}\Phi(\kappa)\Lambda^{\frac{3}{2}+2\varepsilon_0}g dxdydz\Big\|\\ &\leq& C\|\Lambda^{\frac{3}{2}+2\varepsilon_0} g\|\|\Lambda^{\frac{1}{2}+2\varepsilon_0} \kappa\|\\ &\leq& \varepsilon\|\Lambda^{\frac{3}{2}+2\varepsilon_0} g\|^2+C\|\Lambda^{\frac{1}{2}+2\varepsilon_0} \kappa\|^2,\end{aligned}$$ thus, we obtain $$\begin{aligned} \label{eq-37} &&\frac{1}{2}\frac{d}{dt}\int_{\mathbb{T}^3}\|\Lambda^{\frac{1}{2}+2\varepsilon_0} g\|^2dxdydz+\|\Lambda^{\frac{3}{2}+2\varepsilon_0} g\|^2\\ \notag &\leq& \varepsilon\|\Lambda^{\frac{3}{2}+2\varepsilon_0} g\|^2+\varepsilon\|\Lambda^{\frac{3}{2}+2\varepsilon_0} \eta\|^2+ C(R,\|U_1\|_{\mathcal{W}},\|U_2\|_{\mathcal{W}})(1+\|\Lambda^{\frac{1}{2}+2\varepsilon_0}\kappa\|^2+\|\Lambda^{\frac{1}{2}+2\varepsilon_0}g\|^2). \notag\end{aligned}$$ Combining (\[eq-33\]) with (\[eq-37\]), we have $$\begin{aligned} \label{eq-38} \frac{d\\| Y\\|^2_{\frac{1}{2}+2\varepsilon_0}}{dt}+\\| Y\\|^2_{\frac{3}{2}+2\varepsilon_0} \leq C(R,\sup_{t\in [0,T]}\|U_1\|_{\mathcal{W}},\sup_{t\in [0,T]}\|U_2\|_{\mathcal{W}})(1+\\| Y\\|^2_{\frac{1}{2}+2\varepsilon_0}),\end{aligned}$$ by Gronwall inequality, (\[eq-38\]) yields that $\\|Y\\|_{\frac{1}{2}+2\varepsilon_0 }=0$, which implies $Y=0$. Up to now, we have proved that equation (\[eq-17\]) has a unique global weak solution in the space $C([0,T];\mathcal{W})$. Next, we prove (\[eq-39\]). In order to do so, it is sufficient to show that $P^{(R)}_y[\tau_R<\varepsilon]\leq C(\varepsilon, R)$ with $C(\varepsilon, R)\downarrow 0$ as $\varepsilon\downarrow 0$, for all $y\in \mathcal{W}$, with $\|y\|^2_{\mathcal{W}}\leq \frac{R}{8}$. So, fix $\varepsilon$ small enough, let $$\Theta_{\varepsilon, R}:= \sup_{t\in[0,\varepsilon]}\|AZ(t)\|$$ and assume that $\Theta^{6}_{\varepsilon, R}\leq \frac{R}{8}$. Moreover, setting $$\varphi(t):= \|X^{(R)}\|^2_{\mathcal{W}}+\Theta^{6}_{\varepsilon, R},$$ by (\[eq-26\]), we get $\dot{\varphi}\leq C(R)$. This implies, together with the bounds on $y$ and $\Theta_{\varepsilon, R}$, that $$\sup_{t\in[0,\varepsilon]}\|U^{(R)}(t)\|^2_{\mathcal{W}}\leq R$$ for $\varepsilon$ small enough. It follows that $\tau_R\geq \varepsilon$. Hence, $$P^{(R)}_y[\tau_R<\varepsilon]\leq P^{(R)}_y\Big[\sup_{t\in [0,\varepsilon]}\|A Z\|^{6}> \frac{R}{8}\Big],$$ letting $\varepsilon\downarrow 0$, we have $$P^{(R)}_y[\tau_R<\varepsilon]\rightarrow 0.$$ Since the probability above is independent of $y$, (\[eq-39\]) is proved. Finally, since $$U(t\wedge \tau_R(U^{(R)}))=U^{(R)}(t\wedge \tau_R(U^{(R)})) \quad \forall t,\ P-a.s.$$ and $U$ is $H$-valued weakly continuous, we obtain $\tau_R(U^{(R)})=\tau_R(U)$, thus (\[eq-40\]) is proved. $\hfill\blacksquare$ In order to apply Theorem 5.4 in [@F-M] to obtain Theorem \[thm-4\], we now only need to prove \[prp-1\] Assume Hypothesis H1 holds. For every $R>0$, the transition semigroup $(P^{(R)}_t)_{t\geq 0}$ associated to equation (\[eq-16\]) is $\mathcal{W}$-strong Feller. Proof of Proposition \[prp-1\] Let $(\Omega, \mathcal{B}, (\mathcal{B}_t)_{t\geq 0}, {P})$ be a filtered probability space, $(W_t)_{t\geq 0}$ a cylindrical Wiener process on $H$ and for every $y\in \mathcal{W}$, denote by $U^{(R)}_y$ the solution to (\[eq-16\]) with initial value $y\in \mathcal{W}$. By the Bismut-Elworthy-Li formula, $$D_z(P^{(R)}_t\psi)(y)=\frac{1}{t}E^{\mathbb{P}}\Big[\psi(U^{(R)}_y(t))\int^t_0(Q^{-\frac{1}{2}}D_zY^{(R)}_y(s),dW(s))\Big],$$ where $D_z(P^{(R)}_t\psi)$ denotes $(D(P^{(R)}_t\psi),y)$, for $y\in H$, $D_zY^{(R)}_y=DY^{(R)}_y\cdot z$, and $DY^{(R)}_y$ denotes the derivative of $U^{(R)}_y$ with respect to the initial value. Then for $\|\psi\|_{\infty}\leq 1$, by the B-D-G inequality, we have $$\begin{aligned} \label{eq-46} \|(P^{(R)}_t\psi)(y_0+h)-(P^{(R)}_t\psi)(y_0)\|\leq \frac{C}{t}\sup_{\eta\in[0,1]}E^{\mathbb{P}}\Big[(\int^t_0\|Q^{-\frac{1}{2}}D_hY^{(R)}_{y_0+\eta h}(s)\|ds)^{\frac{1}{2}}\Big].\end{aligned}$$ This proposition is proved once we prove that the right side of the above inequality converges to 0 as $\|h\|_{\mathcal{W}}\rightarrow 0$. For any $y\in \mathcal{W}, h\in H$, write $U=U^{(R)}_y, DY=D_h U=(D_h v, D_h T)=(\eta(t,y)\cdot h,\gamma(t,y)\cdot h)$ and denote $DY=\beta=(\eta,\gamma)$ for simplicity. Refer to (\[eq-16\]) and [@RR], we have $$\begin{aligned} \label{eq-41} &&\frac{\partial \beta}{\partial t}+\chi_R(\|U\|^2_{\mathcal{W}})\Big[(v\cdot \nabla_H)\beta+\Phi(v)\frac{\partial \beta}{\partial z}+(\eta\cdot \nabla_H)U+\Phi(\eta)\frac{\partial U}{\partial z}+\tilde{G}(\beta) \Big] +A\beta\\ \notag &&\quad \quad +2\chi'_R(\|U\|^2_{\mathcal{W}})(U,\beta)_{\mathcal{W}}\Big(B(U,U)+G(U)\Big)=0\end{aligned}$$ with initial value $\beta(0)=h$, and $$\tilde{G}(\beta)=\left( \begin{array}{c} fk\times\eta-\int^{z}_{-1}\nabla_H \gamma dz' \\ \Phi(\eta) \\ \end{array} \right).$$ We also can rewrite (\[eq-41\]) in the following form $$\begin{aligned} \label{eq-42} &&\frac{\partial \eta}{\partial t}+\chi_R(\|U\|^2_{\mathcal{W}})[(v\cdot \nabla_H)\eta+\Phi(v)\frac{\partial \eta}{\partial z}+(\eta\cdot \nabla_H)v+\Phi(\eta)\frac{\partial v}{\partial z}+fk\times \eta-\int^z_{-1}\nabla_H\gamma dz' ] \\ \notag &&\quad +2\chi'_R(\|U\|^2_{\mathcal{W}})(U,\beta)_{\mathcal{W}}[(v\cdot\nabla_H)v+\Phi(v)\frac{\partial v}{\partial z}+fk\times v+\nabla_H p_b-\int^z_{-1}\nabla_HTdz']-\Delta\eta=0,\end{aligned}$$ and $$\begin{aligned} \label{eq-43} &&\frac{\partial \gamma}{\partial t}+\chi_R(\|U\|^2_{\mathcal{W}})[(v\cdot \nabla_H)\gamma+\Phi(v)\frac{\partial \gamma}{\partial z}+(\eta\cdot \nabla_H)T+\Phi(\eta)\frac{\partial T}{\partial z}+\Phi(\eta) ] \\ \notag &&\quad +2\chi'_R(\|U\|^2_{\mathcal{W}})(U,\beta)_{\mathcal{W}}[(v\cdot\nabla_H)T+\Phi(v)\frac{\partial T}{\partial z}+\Phi(v)]-\Delta\gamma=0.\end{aligned}$$ Multiplying (\[eq-42\]) by $-\Lambda^{3+4\varepsilon_0} \eta$, then integrating over $\mathbb{T}^3$, we have $$\begin{aligned} &&\frac{1}{2}\frac{d\|\Lambda^{\frac{3}{2}+2\varepsilon_0} \eta\|^2}{dt}+\|\Lambda^{\frac{5}{2}+2\varepsilon_0} \eta\|^2\\ &=&\chi_R(\|U\|^2_{\mathcal{W}})\int_{\mathbb{T}^3}[(v\cdot \nabla_H)\eta+\Phi(v)\frac{\partial \eta}{\partial z}+(\eta\cdot \nabla_H)v+\Phi(\eta)\frac{\partial v}{\partial z}]\Lambda^{3+4\varepsilon_0}\eta dxdydz\\ && \quad +\chi_R(\|U\|^2_{\mathcal{W}})\int_{\mathbb{T}^3}[fk\times \eta-\int^z_{-1}\nabla_H\gamma dz' ]\Lambda^{3+4\varepsilon_0}\eta dxdydz\\ \notag &&\quad+2\int_{\mathbb{T}^3}\chi'_R(\|U\|^2_{\mathcal{W}})(U,\beta)_{\mathcal{W}}[(v\cdot\nabla_H)v+\Phi(v)\frac{\partial v}{\partial z}]\Lambda^{3+4\varepsilon_0} \eta dxdydz \\ \notag && \quad+2\int_{\mathbb{T}^3}\chi'_R(\|U\|^2_{\mathcal{W}})(U,\beta)_{\mathcal{W}}[fk\times v+\nabla_H p_b-\int^z_{-1}\nabla_HTdz']\Lambda^{3+4\varepsilon_0} \eta dxdydz\\ &:=& K_1+K_2+K_3+K_4.\end{aligned}$$ For $K_1$, we only need to estimate $$\begin{aligned} &&\Big\|\int_{\mathbb{T}^3}[\Phi(v)\frac{\partial \eta}{\partial z}+\Phi(\eta)\frac{\partial v}{\partial z}]\Lambda^{3+4\varepsilon_0}\eta dxdydz\Big\| \\ &=&\Big\|\int_{\mathbb{T}^3}\Lambda^{\frac{1}{2}+2\varepsilon_0}[\Phi(v)\frac{\partial \eta}{\partial z}+\Phi(\eta)\frac{\partial v}{\partial z}]\Lambda^{\frac{5}{2}+2\varepsilon_0}\eta dxdydz\Big\|\\ &\leq& \|\Lambda^{\frac{5}{2}+2\varepsilon_0}\eta \|\Big(\|\Lambda^{\frac{3}{2}+2\varepsilon_0+\sigma_1}\eta\|\|\Lambda^{1+\sigma_2}v\|+\|\Lambda^{1+s_1}\eta\|\|\Lambda^{\frac{3}{2}+2\varepsilon_0+s_2}v\|\Big)\\ &\leq& \|\Lambda^{\frac{5}{2}+2\varepsilon_0}\eta \|\|\Lambda^2\eta\|\|\Lambda^{\frac{3}{2}+2\varepsilon_0}v\|\\ &\leq& \|\Lambda^{\frac{5}{2}+2\varepsilon_0}\eta \|\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\eta\|^{\frac{1}{2}+2\varepsilon_0} \|\Lambda^{\frac{5}{2}+2\varepsilon_0}\eta\|^{\frac{1}{2}-2\varepsilon_0}\|\Lambda^{\frac{3}{2}+2\varepsilon_0}v\|\\ &\leq& \varepsilon\|\Lambda^{\frac{5}{2}+2\varepsilon_0}\eta \|^2+C\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\eta \|^2\|\Lambda^{\frac{3}{2}+2\varepsilon_0}v\|^{\frac{4}{1+4\varepsilon_0}},\end{aligned}$$ where $\sigma_1+\sigma_2=1$, $s_1+s_2=1$, and we choose $$\sigma_1=\frac{1}{2}-2\varepsilon_0,\quad \sigma_2=\frac{1}{2}+2\varepsilon_0,\quad s_1=1,\quad s_2=0.$$ For $K_2$, $$\begin{aligned} &&\Big\|\int_{\mathbb{T}^3}[fk\times \eta-\int^z_{-1}\nabla_H\gamma dz' ]\Lambda^{3+4\varepsilon_0}\eta dxdydz \Big\|\\ &\leq&\varepsilon\|\Lambda^{\frac{5}{2}+2\varepsilon_0}\eta \|^2+C\|\Lambda^{\frac{1}{2}+2\varepsilon_0}\gamma \|^2+C\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\eta\|^2.\end{aligned}$$ For $K_3$, $$\begin{aligned} &&\Big\|\int_{\mathbb{T}^3}(U,\beta)_{\mathcal{W}}[\Phi(v)\frac{\partial v}{\partial z}]\Lambda^{3+4\varepsilon_0} \eta dxdydz\Big\|\\ &\leq &(U,\beta)_{\mathcal{W}}\|\Lambda^{\frac{5}{2}+2\varepsilon_0}\eta \|\Big(\|\Lambda^{\frac{3}{2}+2\varepsilon_0+\sigma_1}v \|\|\Lambda^{\sigma_2}\frac{\partial v}{\partial z} \|+\|\Lambda^{1+s_1}v \|\|\Lambda^{\frac{1}{2}+2\varepsilon_0+s_2}\frac{\partial v}{\partial z} \|\Big)\\ &\leq& C(U,\beta)_{\mathcal{W}}\|\Lambda^{\frac{5}{2}+2\varepsilon_0}\eta \|\|\Lambda^{\frac{3}{2}+2\varepsilon_0}v \|\|\Lambda \frac{\partial v}{\partial z}\|\\ &\leq& C\|\Lambda^{\frac{3}{2}+2\varepsilon_0}v \|\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\eta \|\|\Lambda^{\frac{5}{2}+2\varepsilon_0}\eta \|\|\Lambda^{\frac{3}{2}+2\varepsilon_0}v \|\|\Lambda \frac{\partial v}{\partial z}\|+C\|\Lambda^{\frac{3}{2}+2\varepsilon_0}T \|\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\gamma \|\|\Lambda^{\frac{5}{2}+2\varepsilon_0}\eta \|\|\Lambda^{\frac{3}{2}+2\varepsilon_0}v \|\|\Lambda \frac{\partial v}{\partial z}\|\\ &\leq & \varepsilon\|\Lambda^{\frac{5}{2}+2\varepsilon_0}\eta \|^2+C\|\Lambda^{\frac{3}{2}+2\varepsilon_0}v \|^4\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\eta \|^2(\|\Lambda^{\frac{5}{2}+2\varepsilon_0}\kappa\|^2+\|\Lambda\frac{\partial Z_1}{\partial z}\|^2)\\ &&\quad +C\|\Lambda^{\frac{3}{2}+2\varepsilon_0}v \|^2\|\Lambda^{\frac{3}{2}+2\varepsilon_0}T \|^2\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\gamma \|^2(\|\Lambda^{\frac{5}{2}+2\varepsilon_0}\kappa\|^2+\|\Lambda\frac{\partial Z_1}{\partial z}\|^2),\end{aligned}$$ where $\sigma_1+\sigma_2=1$, $s_1+s_2=1$, and we choose $$\sigma_1=0,\quad \sigma_2=1,\quad s_1=\frac{1}{2}+2\varepsilon_0,\quad s_2=\frac{1}{2}-2\varepsilon_0.$$ For $K_4$, $$\begin{aligned} &&\Big\|\int_{\mathbb{T}^3}(U,\beta)_{\mathcal{W}}[\int^z_{-1}\nabla_HTdz']\Lambda^{3+4\varepsilon_0} \eta dxdydz\Big\|\\ &=& \Big\|\int_{\mathbb{T}^3}(U,\beta)_{\mathcal{W}}\Lambda^{\frac{1}{2}+2\varepsilon_0}[\int^z_{-1}\nabla_HTdz']\Lambda^{\frac{5}{2}+2\varepsilon_0} \eta dxdydz\Big\|\\ &\leq & (U,\beta)_{\mathcal{W}}\|\Lambda^{\frac{5}{2}+2\varepsilon_0}\eta \|\|\Lambda^{\frac{3}{2}+2\varepsilon_0}T\|\\ &\leq &C\|\Lambda^{\frac{3}{2}+2\varepsilon_0}v \|\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\eta \|\|\Lambda^{\frac{5}{2}+2\varepsilon_0}\eta \|\|\Lambda^{\frac{3}{2}+2\varepsilon_0}T\|+C\|\Lambda^{\frac{3}{2}+2\varepsilon_0}T \|\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\gamma \|\|\Lambda^{\frac{5}{2}+2\varepsilon_0}\eta \|\|\Lambda^{\frac{3}{2}+2\varepsilon_0}T\|\\ &\leq& \varepsilon\|\Lambda^{\frac{5}{2}+2\varepsilon_0}\eta \|^2+C\|\Lambda^{\frac{3}{2}+2\varepsilon_0}v \|^2\|\Lambda^{\frac{3}{2}+2\varepsilon_0}T \|^2\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\eta \|^2+C\|\Lambda^{\frac{3}{2}+2\varepsilon_0}T \|^4\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\gamma \|^2.\end{aligned}$$ Multiplying (\[eq-43\]) by $-\Lambda^{3+4\varepsilon_0} \gamma$, then integrating over $\mathbb{T}^3$, we have $$\begin{aligned} &&\frac{1}{2}\frac{d \|\Lambda^{\frac{3}{2}+2\varepsilon_0}\gamma\|^2}{d t}+\|\Lambda^{\frac{5}{2}+2\varepsilon_0} \gamma\|^2\\ &=&\chi_R(\|U\|^2_{\mathcal{W}})\int_{\mathbb{T}^3}\Big[(v\cdot \nabla_H)\gamma+\Phi(v)\frac{\partial \gamma}{\partial z}+(\eta\cdot \nabla_H)T+\Phi(\eta)\frac{\partial T}{\partial z} +\Phi(\eta)\Big]\Lambda^{3+4\varepsilon_0} \gamma dxdydz \\ &&\quad +2\int_{\mathbb{T}^3}\chi'_R(\|U\|^2_{\mathcal{W}})(U,\beta)_{\mathcal{W}}\int_{\mathbb{T}^3}\Big[(v\cdot\nabla_H)T+\Phi(v)\frac{\partial T}{\partial z}+\Phi(v)\Big]\Lambda^{3+4\varepsilon_0} \gamma dxdydz\\ &:=& K_5+K_6.\end{aligned}$$ For $K_5$, we only need to estimate the following term, $$\begin{aligned} &&\Big\|\int_{\mathbb{T}^3}[\Phi(v)\frac{\partial \gamma}{\partial z}+\Phi(\eta)\frac{\partial T}{\partial z} ]\Lambda^{3+4\varepsilon_0} \gamma dxdydz\Big\| \\ &=& \Big\|\int_{\mathbb{T}^3}\Lambda^{\frac{1}{2}+2\varepsilon_0} [\Phi(v)\frac{\partial \gamma}{\partial z}+\Phi(\eta)\frac{\partial T}{\partial z} ]\Lambda^{\frac{5}{2}+2\varepsilon_0} \gamma dxdydz\Big\|\\ &\leq& C\|\Lambda^{\frac{5}{2}+2\varepsilon_0} \gamma\|\Big(\|\Lambda^{\frac{3}{2}+2\varepsilon_0} v\|\|\Lambda^{2} \gamma\|+\|\Lambda^{\frac{3}{2}+2\varepsilon_0} T\|\|\Lambda^{2} \eta\|\Big)\\ &\leq& \varepsilon\|\Lambda^{\frac{5}{2}+2\varepsilon_0}\eta \|^2+\varepsilon\|\Lambda^{\frac{5}{2}+2\varepsilon_0}\gamma \|^2 +C\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\gamma \|^2\|\Lambda^{\frac{3}{2}+2\varepsilon_0}v \|^{\frac{4}{1+4\varepsilon_0}}+C\|\Lambda^{\frac{3}{2}+2\varepsilon_0}\eta \|^2\|\Lambda^{\frac{3}{2}+2\varepsilon_0}T \|^{\frac{4}{1+4\varepsilon_0}}.\end{aligned}$$ For $K_6$, $$\begin{aligned} &&\Big\|\int_{\mathbb{T}^3}\chi'_R(\|U\|^2_{\mathcal{W}})(U,\beta)_{\mathcal{W}}\int_{\mathbb{T}^3}[\Phi(v)\frac{\partial T}{\partial z}]\Lambda^{3+4\varepsilon_0} \gamma dxdydz\Big\|\\ &=& \Big\|\int_{\mathbb{T}^3}\chi'_R(\|U\|^2_{\mathcal{W}})(U,\beta)_{\mathcal{W}}\int_{\mathbb{T}^3}\Lambda^{\frac{1}{2}+2\varepsilon_0}[\Phi(v)\frac{\partial T}{\partial z}]\Lambda^{\frac{5}{2}+2\varepsilon_0} \gamma dxdydz\Big\|\\ &\leq& C(U,\beta)_{\mathcal{W}}\|\Lambda^{\frac{5}{2}+2\varepsilon_0} \gamma\|\Big(\|\Lambda^{\frac{3}{2}+2\varepsilon_0+\sigma_1}v\|\|\Lambda^{1+\sigma_2} T\|+\|\Lambda^{1+s_1}v\|\|\Lambda^{\frac{1}{2}+2\varepsilon_0+s_2} \frac{\partial T}{\partial z}\|\Big)\\ &\leq& C(U,\beta)_{\mathcal{W}}\|\Lambda^{\frac{5}{2}+2\varepsilon_0} \gamma\|\|\Lambda^{\frac{3}{2}+2\varepsilon_0} v\|\|\Lambda \frac{\partial T}{\partial z}\|\\ &\leq& C\|\Lambda^{\frac{3}{2}+2\varepsilon_0} v\|^2\|\Lambda^{\frac{3}{2}+2\varepsilon_0} \eta\|\|\Lambda \frac{\partial T}{\partial z}\|\|\Lambda^{\frac{5}{2}+2\varepsilon_0} \gamma\|+C\|\Lambda^{\frac{3}{2}+2\varepsilon_0} v\|\|\Lambda^{\frac{3}{2}+2\varepsilon_0} T\|\|\Lambda^{\frac{3}{2}+2\varepsilon_0} \gamma\|\|\Lambda \frac{\partial T}{\partial z}\|\|\Lambda^{\frac{5}{2}+2\varepsilon_0} \gamma\|\\ &\leq& C\|\Lambda^{\frac{3}{2}+2\varepsilon_0} v\|^2\|\Lambda^{\frac{3}{2}+2\varepsilon_0} \eta\|\|\Lambda^{\frac{5}{2}+2\varepsilon_0} g\|\|\Lambda^{\frac{5}{2}+2\varepsilon_0} \gamma\|+C\|\Lambda^{\frac{3}{2}+2\varepsilon_0} v\|\|\Lambda^{\frac{3}{2}+2\varepsilon_0} T\|\|\Lambda^{\frac{3}{2}+2\varepsilon_0} \gamma\|\|\Lambda^{\frac{5}{2}+2\varepsilon_0} g\|\|\Lambda^{\frac{5}{2}+2\varepsilon_0} \gamma\|\\ &&\quad +C\|\Lambda^{\frac{3}{2}+2\varepsilon_0} v\|^2\|\Lambda^{\frac{3}{2}+2\varepsilon_0} \eta\|\|\Lambda \frac{\partial Z_2}{\partial z}\|\|\Lambda^{\frac{5}{2}+2\varepsilon_0} \gamma\|+C\|\Lambda^{\frac{3}{2}+2\varepsilon_0} v\|\|\Lambda^{\frac{3}{2}+2\varepsilon_0} T\|\|\Lambda^{\frac{3}{2}+2\varepsilon_0} \gamma\|\|\Lambda \frac{\partial Z_2}{\partial z}\|\|\Lambda^{\frac{5}{2}+2\varepsilon_0} \gamma\|\\ &\leq& \varepsilon \|\Lambda^{\frac{5}{2}+2\varepsilon_0} \gamma\|^2+C\|\Lambda^{\frac{3}{2}+2\varepsilon_0} v\|^4\|\Lambda^{\frac{3}{2}+2\varepsilon_0} \eta\|^2(\|\Lambda^{\frac{5}{2}+2\varepsilon_0} g\|^2+\|\Lambda\frac{\partial Z_2}{\partial z}\|^2)\\ &&\quad+C\|\Lambda^{\frac{3}{2}+2\varepsilon_0} v\|^2\|\Lambda^{\frac{3}{2}+2\varepsilon_0} T\|^2\|\Lambda^{\frac{3}{2}+2\varepsilon_0} \gamma\|^2(\|\Lambda^{\frac{5}{2}+2\varepsilon_0} g\|^2+\|\Lambda\frac{\partial Z_2}{\partial z}\|^2).\end{aligned}$$ Combing all the above estimations, we have for $\|U\|^2_{\mathcal{W}}\leq R$, $$\begin{aligned} \frac{d\| \beta\|^2_{\mathcal{W}}}{dt}+\\|\beta\\|^2_{\frac{5}{2}+2\varepsilon_0} \leq C\Big(C(R)+\|\Lambda^{\frac{5}{2}+2\varepsilon_0} \kappa\|^2+\|\Lambda^{\frac{5}{2}+2\varepsilon_0} g\|^2+\|\Lambda\frac{\partial Z_1}{\partial z}\|^2+\|\Lambda\frac{\partial Z_2}{\partial z}\|^2\Big)\|\beta\|^2_{\mathcal{W}},\end{aligned}$$ by Gronwall’s inequality and (\[eq-26\]), we finally get $$\begin{aligned} &&\int^t_0\\|\beta(l)\\|^2_{\frac{5}{2}+2\varepsilon_0}dl\\ &\leq& C\| h\|^2_{\mathcal{W}}+\exp\left(C\int^t_0\Big(C(R)+\|\Lambda^{\frac{5}{2}+2\varepsilon_0} \kappa\|^2+\|\Lambda^{\frac{5}{2}+2\varepsilon_0} g\|^2+\|\Lambda\frac{\partial Z_1}{\partial z}\|^2+\|\Lambda\frac{\partial Z_2}{\partial z}\|^2\Big)dl\right)\| h\|^2_{\mathcal{W}}\\ &\leq& C\| h\|^2_{\mathcal{W}}+\exp\left(C\left(\| \Lambda^{\frac{3}{2}+2\varepsilon_0}y\|^2+\int^t_0\Big(C(R)+\|\Lambda\frac{\partial Z_1}{\partial z}\|^2+\|\Lambda\frac{\partial Z_2}{\partial z}\|^2\Big)dl\right)\right)\| h\|^2_{\mathcal{W}}.\end{aligned}$$ Since $Z_i$ is a Gaussian random variable in $C([0,\infty), D(\Lambda^{2+2\varepsilon_0-2\varepsilon}))$, for every $\varepsilon>0$, by Fernique’s theorem, we could choose $t_0$ small enough and obtain $$\mathbb{E}\int^{t_0}_0\\|\beta(l)\\|^2_{\frac{5}{2}+2\varepsilon_0}dl\leq C(t_0,R)\| h\|^2_{\mathcal{W}}.$$ From Hypothesis H1, it follows that $$Q^{-\frac{1}{2}}=Q^{-\frac{1}{2}}_0A^{\frac{5}{4}+\varepsilon_0},$$ thus, the assertion of (\[eq-46\]) holds for $t_0$. For general $t$, by the semigroup property, the assertion follows easily. $\hfill\blacksquare$ In order to apply Fernique’s theorem to Gaussian process $Z$, we have to make use of $\Lambda \frac{\partial Z}{\partial z}$ and control its power to be less than 2 during the estimate of $K_3$ and $K_6$. Proof of Theorem \[thm-4\] Theorem \[thm-5\] and Proposition \[prp-1\] imply the result by Theorem 5.4 in [@F-M]. 0.5cm [ This work was supported by National Natural Science Foundation of China (NSFC) (No. 11431014, No. 11371041, No. 11401557, No. 11271356)]{}, the Fundamental Research Funds for the Central Universities (No. 0010000048), Key Laboratory of Random Complex Structures and Data Science, Academy of Mathematics and Systems Science, Chinese Academy of Sciences(No. 2008DP173182), and the applied mathematical research for the important strategic demand of China in information science and related fields(No. 2011CB808000). [2]{} R.A. Adams : Sobolev Space. New York: Academic Press, 1975. C. Cao, J. Li, E. Titi: Local and global well-posedness of strong solutions to the 3D primitive equations with vertical eddy diffusivity. Arch. Rational Mech. Anal. 214 35-76 (2014). C. Cao, E.S. Titi: Global well-posedness and finite-dimensional global attractor for a 3D planetary geostrophic viscous model. Comm. Pure Appl. Math. 56, no. 2, 198-233 (2003). C. Cao, E.S. Titi: Global well-posedness of the three-dimensional viscous primitive equations of large-scale ocean and atmosphere dynamics. Ann. of Math. 166, 245-267 (2007). A. Debussche, N. Glatt-Holtz, R. Temam, M. Ziane: Global existence and regularity for the 3D stochastic primitive equations of the ocean and atmosphere with multiplicative white noise. Nonlinearity, 25, 2093-2118 (2012). Z. Dong, J. Zhai, R. Zhang: Exponential mixing for 3D stochastic primitive equations of the large scale ocean. Preprint. Available at arXiv: 1506.08514. B. Ewald, M. Petcu, P. Temam: Stochstic solutions of the two-dimensional primitive equations of the ocean and atmosphere with an additive noise, Anal. Appl., 5, 183-198 (2007). F. Flandoli, M. Romito : Markov selection for the 3D stochastic Navier-Stokes equations Probab. Theory Relat. Fields 140: 407-458 (2008). H. Gao, C. Sun: Hausdorff dimension of random attractor for stochastic Navier-Stokes-Voight equations and primitive equations. Dyn. Partial Differ. Equ. 7 , no. 4, 307-326 (2010). H. Gao, C. Sun: Well-posedness and large deviations for the stochastic primitive equations in two space dimensions. Commun. Math. Sci. Vol.10, No.2, 575-593 (2012). B. Goldys, M. Röckner, X. Zhang: Martingale solutions and Markov selections for stochastic partial differential equations. Stochastic Process. Appl. 119, no. 5, 1725-1764 (2009). F. $\rm Guill\acute{e}n-Gonz\acute{a}lez$, N. Masmoudi, M.A. Rodriguez-Bellido: Anisotropic estimates and strong solutions for the primitive equations. Diff. Int. Equ. 14, 1381-1408(2001). B. Guo, D. Huang: 3D stochastic primitive equations of the large-scale ocean: global well-posedness and attractors. Comm. Math. Phys. 286, no. 2, 697-723 (2009). C. Hu, R. Temam, M. Ziane: The primitive equations on the large scale ocean under the small depth hypothesis. Discrete Contin. Dynam. Systems 9, 97-131 (2003). N. Ju: Existence and uniqueness of the solution to the dissipative 2D quasi-geostrophic equations in the Sobolev space. Comm. Math. Phys. 251 365-376. MR2100059 (2004). J. Li, E.S. Titi: Existence and uniqueness of weak solutions to viscous primitive equations for certain class of discontinuous initial data. Arxiv:1512.00700 (2015). J.L. Lions, R. Temam, S. Wang : New formulations of the primitive equations of atmosphere and applications. Nonlinearity 5, 237-288 (1992). J.L. Lions, R. Temam, S. Wang : On the equations of the large scale ocean. Nonlinearity 5, 1007-1053 (1992). J.L. Lions, R. Temam, S. Wang : Models of the coupled atmosphere and ocean. Computational Mechanics Advance 1, 1-54 (1993). J.L. Lions, R. Temam, S. Wang : Mathematical theory for the coupled atmosphere-ocean models. J. Math. Pures Appl. 74, 105-163 (1995). S.G. Resnick: Dynamical problems in nin-linear advective partial differential equations. PH.D. thesis, Univ. of Chicago. MR2716577 (1995). M. Romito : Analysis of equilibrium states of markov solutions to the 3D Navier-Stokes equations driven by additive noise J. Stat. Phys. 131, no.3, 415-444 (2008). E. M. Stein: Singular Integrals and Differentiability Properties of Functions. Princeton Univ. Press, Princeton, N.J. MR0290095 (1970). D.W. Stroock, S.R.S. Varadhan: Multidimensional Diffusion Processes. Springer-Verlag, Berlin, 1979. R. Temam : Navier-Stokes Equations: Theory and Numeriacal Analysis, 3rd ed. Studies in Mathematical and Its Applications 2. North-Holland, Amsterdam (1984). G. Zhou: Global well-posedness and random attractor of the 3D viscous primitive equations driven by fractional noises. Preprint. Available at arXiv:1604.05376 (2016).	{ "pile_set_name": "ArXiv" }
--- abstract: 'In this paper, we are concerned with possibly degenerate stochastic partial differential equations (SPDEs). An $L^2$-theory is introduced, from which we derive the Hörmander theorem with an analytical approach. With the method of De Giorgi iteration, we obtain the maximum principle which states the $L^p$ ($p>0$) estimates for the time-space uniform norm of weak solutions.' author: - Jinniao Qiu title: ' $L^2$-Theory of Linear Degenerate SPDEs and $L^p$ ($p>0$) Estimates for the Uniform Norm of Weak Solutions ' --- [Mathematics Subject Classification (2010):]{} 60H15, 35R60, 35D30 [Keywords:]{} stochastic partial differential equation, $L^2$-theory, Hörmander theorem, maximum principle, De Giorgi iteration Introduction ============ Let $(\Omega,{{\mathscr{F}}},\{{{\mathscr{F}}}_t\}_{t \geq 0},{\mathbb{P}})$ be a complete filtered probability space, on which a $d_1$-dimensional Wiener process $W=(W_t)_{t\geq 0}$ is well defined. We consider SPDE of the form $$\label{SPDE} \left\{\begin{array}{l} \begin{split} du(t,x)=\,&\displaystyle \left[ \frac{1}{2}(L_k^2+M_k^2)u +{b}^jD_ju+cu+f+L_k'g^k+M_k'h^k \right](t,x)\, dt\\ &\displaystyle +\left[M_ku+\beta^k u+h^k\right](t,x)\, dW_{t}^{k}, \quad (t,x)\in Q:= [0,T]\times {\mathbb{R}}^d;\\ u(0,x)=\, &\underline{u}_0(x), \quad x\in{\mathbb{R}}^d. \end{split} \end{array}\right.$$ Here and throughout this paper, the summation over repeated indices is enforced unless stated otherwise, $T\in (0,\infty)$, $D=(D_1,\dots,D_d)$ is the gradient operator, and $L_k=\sigma^{jk}D_j$, $M_k=\theta^{jk}D_j$, $L_k'=D_j(\sigma^{jk}\cdot)$, $M_k'=D_j(\theta^{jk}\cdot)$, for $k=1,\dots,d_1$. SPDE is said to be degenerate when it fails to satisfy the super-parabolicity (SP): There exists $\lambda\in(0,\infty)$ such that $$\begin{aligned} \sigma^{ik}\sigma^{jk}(t,x)\xi^i\xi^j\geq \lambda \|\xi\|^2\,\,\,\,\,a.s.,\,\,\,\,\,\forall\, (t,x,\xi)\in [0,T]\times{\mathbb{R}}^d\times{\mathbb{R}}^d. \end{aligned}$$ We first investigate the solvability of linear, possibly degenerate SPDEs in $L^2$-spaces. An $L^2$-theory on linear degenerate SPDEs was initiated by Krylov and Rozovskii [@krylov-1977-Deg-L2], and it was developed recently by [@Deg-SPDE-3-2014; @Deg-SPDE-2014; @krylov2013hypoellipticity; @Deg-SPDE-2-2014]. Along this line, to get a solution of SPDE in space $L^2(\Omega;C([0,T];H^m))$ not only requires that $f+L_k'g^k+M_k'h^k$ is $H^m$-valued but also assumes that $h^k$ is $H^{m+1}$-valued, while in this work, $f,g$ and $h$ are allowed to be just $H^m$-valued, with $H ^{m}:=(1-\Delta)^{-\frac{m}{2}} L^{2}({\mathbb{R}}^d)$ for each $m\in{\mathbb{R}}$. Moreover, we get the estimate $L_ku\in L^2(\Omega\times[0,T];H^m) $, and at the price of assuming a Hörmander-type condition, we further have $u\in L^2(\Omega\times[0,T];H^{m+\eta})$ for some $\eta\in(0,1]$. For the proof, we apply the a priori estimates for solutions of the approximating super-parabolic SPDEs in line with the applications of pseudo-differential operator theory. As a byproduct, a Hörmander theorem for SPDE is derived from the established $L^2$-theory and an estimate on the Lie bracket (Lemma \[lem-lie-bracket\]). Most importantly, we prove the maximum principle for the weak solution of SPDE . More precisely, we obtain the $L^p$ ($p>0$) estimates for the time-space uniform norm of weak solutions, i.e., under suitable integrability assumptions on $\underline{u}_0,f,g$ and $h$, we have \[thm-MP-intr\] Let the Hörmander-type condition $({\mathcal H} )$ hold. For the weak solution $u$ of SPDE , we have for any $p\in (0,\infty)$ $$\begin{aligned} E\\|u^{\mp}\\|^p_{L^{\infty}(Q)} \leq C\,\Xi (\underline{u}_0^{\mp},f^{\mp},g,h),\end{aligned}$$ where $\Xi (\underline{u}_0^{\mp},f^{\mp},g,h)$ is expressed in terms of certain norms of $(\underline{u}_0,f^{\mp},g,h)$, and the constant $C$ depends on $d,p,T$ and the quantities related to the structure coefficients of SPDE . The novelty of our result is that it does not require the super-parabolic condition (SP), which, to the best of our knowledge, is always assumed in the existing literature on such kind of maximum principles for SPDEs. For the super-parabolic SPDEs, Krylov [@Kryl96; @Krylov_99] established the $L^p$-theory ($p\geq 2$), from which one can derive from the classical Sobolev embedding theorem the $L^p$ estimates of time-space uniform norm for the strong solutions that require more smoothness assumptions on the coefficients. For the weak solutions of super-parabolic SPDEs in bounded domains, the maximum principle was obtained by Denis, Matoussi and Stoica [@DenisMatoussiStoica2005] and further by [@D-bounddeness-SPDE-2013; @DenisMatoussi2009], but with $p\in [2,\infty)$. Their method relied on Moser’s iteration. Such method was also used by Denis, Matoussi and Zhang [@denis-2014-MP] to derive the maximum principle for weak solutions of super-parabolic SPDEs with obstacle. In comparison, we adopt a stochastic version of De Giorgi iteration scheme in this paper. We would also note that our method is inspired by the other two different versions of De Giorgi iteration used by Hsu, Wang and Wang [@hsu-2013-Stoch-DeGiorgi] to investigate the regularity of strong solutions for super-parabolic SPDEs and by Qiu and Tang [@QiuTangMPBSPDE11] to study the maximum principles of weak solutions for a class of backward SPDEs. The remainder of this paper is organized as follows. In Section 2, we set some notations and state our main result. The $L^2$-theory and the Hörmander theorem are proved in Section 3. In section 4, we prove the maximum principle. Preliminaries and the main results ================================== Let $L^2({\mathbb{R}}^d)$ ($L^2$ for short) be the usual Lebesgue integrable space with usual scalar product $\langle\cdot,\,\cdot\rangle$ and norm $\\|\cdot\\|$. For $n\in (-\infty,\infty)$, we denote by $H ^{n}$ the space of Bessel potentials, that is $H ^{n}:=(1-\Delta)^{-\frac{n}{2}} L^{2}$ with the norm $$\\|\phi\\|_{n}:=\\|(1-\Delta)^{\frac{n}{2}}\phi\\|, ~~ \phi\in H^{n}.$$ For each $l\in \mathbb{N}^+$ and domain $\Pi\subset {\mathbb{R}}^l$, denote by $C_c^{\infty}(\Pi)$ the space of infinitely differentiable functions with compact supports in $\Pi$. For convenience, we shall use $\langle \cdot,\,\cdot\rangle$ to denote the duality between $(H^n)^k$ and $(H^{-n})^k$ ($k\in{\mathbb{N}}^+,\,n\in{\mathbb{R}}$) as well as that between the Schwartz function space $\mathscr{D}$ and $C_c^{\infty}({\mathbb{R}}^d)$. Moreover, We always omit the superscript associated to the dimension when there is no confusion. For Banach space ($\mathbb{B}$, $\\|\cdot\\|_{\mathbb{B}}$) and $p\in[1,\infty]$, ${\mathcal{S}}^p (\mathbb{B})$ is the set of all the $\mathbb{B}$-valued, $({\mathscr{F}}_t)$-adapted and continuous processes $(X_{t})_{t\in [0,T]}$ such that $$\\|X\\|_{{\mathcal{S}}^p(\mathbb{B})}:= \left\\| \sup_{t\in [0,T]} \\|X_t\\|_{\mathbb{B}} \right\\|_{L^p(\Omega)}< \infty.$$ Denote by $\mathcal{L}^p(\mathbb{B})$ the totality of all the $\mathbb B$-valued, $({\mathscr{F}}_t)$-adapted processes $(X_{t})_{t\in [0,T]}$ such that $$\\|X\\|_{\mathcal{L}^p(\mathbb{B})}:= \left\\| \\|X_t\\|_{\mathbb{B}} \right\\|_{L^p(\Omega\times[0,T])}< \infty.$$ Obviously, both $({\mathcal{S}}^p(\mathbb{B}),\,\\|\cdot\\|_{{\mathcal{S}}^p(\mathbb{B})})$ and $(\mathcal{L}^p(\mathbb{B}),\\|\cdot\\|_{\mathcal{L}^p(\mathbb{B})})$ are Banach spaces. In addition, for $p\in(0,1)$, we denote by $L^p(\Omega;\mathbb B)$ the $\mathbb B$-valued ${\mathscr{F}}$-measurable functions $f$ such that $\\|f\\|_{\mathbb B}^p\in L^1(\Omega;{\mathbb{R}})$ with $\\|f\\|_{L^p(\Omega;\mathbb B)}:=\left\\|\\|f\\|_{\mathbb B}^p\right\\|_{L^1(\Omega;{\mathbb{R}})}^{1/p}$. By $C_b^{\infty}$, we denote the set of infinitely differentiable functions with bounded derivatives of any order. Denote by ${\mathcal{L}}^{\infty}(C_b^{\infty})$ the set of functions $h$ on $\Omega\times [0,T] \times {\mathbb{R}}^d$ such that $h(t,x)$ is infinitely differentiable with respect to $x$ and all the derivatives of any order belong to $ {\mathcal{L}}^{\infty}(L^\infty({\mathbb{R}}^d))$. Throughout this paper, we denote $I^n=(1-\Delta)^{\frac{n}{2}}$ for $n\in{\mathbb{R}}$. Then $I^n$ belongs to $\Psi_n$ that is the class of pseudo-differential operators of order $n$. By the pseudo-differential operator theory (see [@Hormander1983analysis] for instance), the $m$-th order differential operator belongs to $\Psi_m$ for $m\in{\mathbb{N}}^+$, the multiplication by elements of $C_b^{\infty}$ lies in $\Psi_0$, and for the reader’s convenience, two basic results are collected below. \[lem-pdo\] (i). If $J_1\in\Psi_{n_1}$ and $J_2\in\Psi_{n_2}$ with $n_1,n_2\in{\mathbb{R}}$, then $J_1J_2\in\Psi_{n_1+n_2}$ and the Lie bracket $[J_1,J_2]:=J_1J_2-J_2J_1\in\Psi_{n_1+n_2-1}$. (ii). For $m\in (0,\infty)$, let $\zeta$ belong to $C_b^{m}$ which is defined as usual. Then for any $n\in(-m,m)$ there exists constant $C$ such that $$\\|\zeta\phi\\|_n\leq C \\|\zeta\\|_{C^{m}}\\|\phi\\|_n,\quad \forall\,\phi\in H^n.$$ We introduce the definition for solution of SPDE . \[defn-solution\] A process $u$ is called a solution to SPDE if $u\in {\mathcal{S}}^2(H^m)$ for some $m\in{\mathbb{R}}$ and SPDE holds in the distributional sense, i.e., for any $\zeta\in C^{\infty}_c({\mathbb{R}})\otimes C_c^{\infty}({\mathbb{R}}^d)$ there holds almost surely $$\begin{split} &\langle \zeta(t),\,u(t)\rangle -\!\int_0^t\!\!\langle \partial_s \zeta(s),\, u(s) \rangle \, ds -\!\int_0^t \!\!\langle \zeta(s),\, (M_ku+\beta^ku+h^k)(s)\rangle \,dW_s^k \\ &=\langle \zeta(0),\, u_0\rangle + \int_0^t\!\! \bigg\langle \zeta,\, \frac{1}{2}(L_k^2+M_k^2)u +{b}^jD_ju+cu+f+L_k'g^k+M_k'h^k \bigg\rangle(s)\, ds , \quad \forall \ t\in[0,T]. \end{split}$$ In particular, if $u\in {\mathcal{S}}^2(L^2)$, it is said to be a weak solution. Set $${\mathbb{V}}_0=\{L_1,\dots,L_{d_1}\} \quad \text{and} \quad {\mathbb{V}}_{n+1}={\mathbb{V}}_n\cup\{[L_k,V]:\,V\in{\mathbb{V}}_n,\,k=1,\dots,d_1\}.$$ Denote by ${\mathbb{L}}_n$ the set of linear combinations of elements of ${\mathbb{V}}_n$ with coefficients of ${\mathcal{L}}^{\infty}(C^{\infty}_b)$. We introduce the following Hörmander-type condition. $({\mathcal H} )$ There exists $n_0\in{\mathbb{N}}_0$ such that $\{D_i:i=1,\dots,d\}\subset {\mathbb{L}}_{n_0}$. (Throughout this paper, $n_0$ is always chosen to be the smallest one.) It is easy to check that the super-parabolicity (SP) corresponds to the trivial case $n_0=0$. For a nontrivial example, we could consider the $2$-dimensional case with $d_1=d=2$: $L_1=D_1$ and $L_2=\cos \left((1+\alpha_t)x\right) D_2$ where $(\alpha_t)_{t\geq 0}$ can be any nonnegative bounded ${\mathscr{F}}_t$-adapted process. Then one has $D_2\notin {\mathbb{L}}_0$, though $\{D_1,D_2\}\subset {\mathbb{L}}_1$ since $[L_1,L_2]=-(1+\alpha_t)\sin \left((1+\alpha_t)x\right) D_2$. Hence, we have $n_0=1$. We also make the following assumptions.\ $({\mathcal A} 1)$ $\sigma^{ik},\theta^{ik},b^i,\beta, c\in{\mathcal{L}}^{\infty}(C_b^{\infty})$, for $i=1,\dots,d$, $k=1,\dots,d_1$;\ $({\mathcal A} 2)$ $c\geq 0$, $\underline{u}_0\in L^{\infty}(\Omega\times{\mathbb{R}}^d)\cap \cap_{q>0} L^q(\Omega,{\mathscr{F}}_0;L^2)$, $f,g^k,h^k\in {\mathcal{L}}^2(L^2)\cap \cap_{q>0} L^q(\Omega;L^2(Q))$, for $k=1,\dots,d_1$, and moreover, for some $ \bar p >d+2\eta$ $$(f,g,h)\in L^{\infty}(\Omega;L^{\frac{\bar p (d+2\eta)}{(\bar p + d + 2\eta)\eta}}(Q))\times L^{\infty}(\Omega;L^{\frac{\bar p }{\eta}}(Q)) \times \left( L^{\infty}(\Omega;L^{\frac{2\bar p (d+2\eta)}{(\bar p + d + 2\eta)\eta}}(Q)) \cap L^{\infty}(\Omega;L^{\frac{\bar p }{\eta}}(Q))\right),$$ where and in the following, we set $\eta=2^{-n_0}$. Throughout this paper, we denote $$\begin{aligned} \Lambda_{\bar p,\infty}^{\mp}&= \\|\underline{u}_0^{\mp}\\|_{L^{\infty}(\Omega\times{\mathbb{R}}^d)} +\operatorname{esssup}_{\omega\in\Omega}\\|f^{\mp}(\omega,\cdot,\cdot)\\|_{L^{\frac{\bar p (d+2\eta)}{(\bar p + d + 2\eta)\eta}}(Q)} +\operatorname{esssup}_{\omega\in\Omega}\\|(g,h)(\omega,\cdot,\cdot)\\|_{L^{\frac{\bar p }{\eta}}(Q)}\\ &\quad +\operatorname{esssup}_{\omega\in\Omega}\\|h(\omega,\cdot,\cdot)\\|_{L^{\frac{2\bar p (d+2\eta)}{(\bar p + d + 2\eta)\eta}}(Q)} ,\\ \Lambda_p^{\mp}&=\\|\underline{u}_0^{\mp}\\|_{L^p(\Omega;L^2)} + \\|(f^{\mp},g,h)\\|_{L^p(\Omega;L^2(Q))} ,\quad p\in(0,\infty).\end{aligned}$$* We now state our main results. \[thm-main\] Let assumption $({\mathcal{A}}1)$ hold. Given $f\in {\mathcal{L}}^2(H^m)$, $g,h\in {\mathcal{L}}^2((H^m)^{d_1})$ and $\underline{u}_0\in L^2(\Omega,{\mathscr{F}}_0;H^m)$ with some $m\in{\mathbb{R}}$, the following three assertions hold: \(i) SPDE admits a unique solution $u\in{\mathcal{S}}^2(H^m)$ with $L_ku\in{\mathcal{L}}^2(H^m)$, $k=1,\dots,d_1$, and $$\begin{aligned} &E\sup_{t\in[0,T]}\\|u(t)\\|_m^2+\sum_{k=1}^{d_1}E\int_{0}^T\\|L_ku(t)\\|_m^2\,dt \nonumber\\ &\leq C\left\{ E\\|\underline{u}_0\\|_m^2 + E\int_{0}^T \left(\\|f(s)\\|_m^2+\\|g(s)\\|_m^2+\\|h(s)\\|_m^2\right)\,ds\right\},\end{aligned}$$ with the constant $C$ depending on $T,m,\theta,\sigma,b,c$ and $\beta$. In particular, if condition $({\mathcal{H}})$ holds, we have further $$\begin{aligned} E\int_{0}^T\\|u(t)\\|_{m+\eta}^2\,dt \leq C\left\{ E\\|\underline{u}_0\\|_m^2 + E\int_{0}^T \left(\\|f(s)\\|_m^2+\\|g(s)\\|_m^2+\\|h(s)\\|_m^2\right)\,ds\right\},\end{aligned}$$ with $C$ depending on $T,m,n_0,\theta,\sigma,b,c$ and $\beta$. \(ii) Assume further $({\mathcal H} )$ and $f\in\cap_{n\in{\mathbb{R}}}{\mathcal{L}}^2(H^n)$, $g,h\in\cap_{n\in{\mathbb{R}}}{\mathcal{L}}^2((H^n)^{d_1})$. For any ${\varepsilon}\in(0,T)$, one has $u\in\cap_{n\in{\mathbb{R}}} L^2(\Omega;C([{\varepsilon},T];H^n))$, and for each $n\in{\mathbb{R}}$, $$\begin{aligned} &E\sup_{t\in[{\varepsilon},T]}\\| u(t)\\|_{n}^2 +E\int_{{\varepsilon}}^{T} \\| u(t)\\|_{n+\eta}^2\,dt \nonumber\\ &\leq C\left\{E\\|G\\|_m^2+ E\int_{0}^{T} \left(\\|f(s)\\|_{n}^2+ \\|g(s)\\|_{n}^2+ \\|h(s)\\|_{n}^2\right) ds\right\},\label{est-hmder}\end{aligned}$$ with the constant $C$ depending on ${\varepsilon},n,T,m,n_0,\sigma,\theta,\gamma,b$ and $c$. In particular, the random field $u(t,x)$ is almost surely infinitely differentiable with respect to $x$ on $(0,T]\times{\mathbb{R}}^d$ and each derivative is a continuous function on $(0,T]\times{\mathbb{R}}^d$. \(iii) Let assumption $({\mathcal{A}}2)$ and condition $({\mathcal{H}})$ hold. For the weak solution $u$ of SPDE , there exists $\theta_0\in(0,1]$ such that for any $p>0$, $$\begin{aligned} E\\|u^{\mp}\\|^p_{L^{\infty}(Q)} \leq C \left(\Lambda^{\mp}_{\bar p,\infty} + \Lambda^{\mp}_{\frac{p}{\theta_0}} \right)^p,\end{aligned}$$ with the constant $C$ depending on $d,p,n_0, T$ and the quantities related to the coefficients $\sigma,\theta,b,c$ and $\beta$. \[rmk-mainr\] Assertion (i) is a summary of Theorem \[thm-BSPDE\] and Corollary \[cor-grad-est\], in which an $L^2$-theory is presented for the linear, possibly degenerate SPDEs. Assertion (ii) is from Theorem \[thm-hormander\], which states a Hörmander-type theorem. The most important result of this paper is the maximum principle of assertion (iii), which corresponds to Theorem \[thm-Lp-inf\] below and states the $L^p$ ($p>0$) estimates for the time-space uniform norm of weak solutions for possibly degenerate SPDE in the whole space. $L^2$-theory and Hörmander theorem for SPDEs ============================================ $L^2$-theory of SPDEs {#sec-l2-thy} --------------------- We consider the following SPDE $$\label{SPDE-D} \left\{\begin{array}{l} \begin{split} du(t,x)=\,&\displaystyle \left[\delta\Delta u+ \frac{1}{2}(L_k^2+M_k^2)u +{b}^jD_ju+cu+f+L_k'g^k+M_k'h^k \right](t,x)\, dt\\ &\displaystyle +\left[M_ku+\beta^k u+h^k\right](t,x)\, dW_{t}^{k}, \quad (t,x)\in Q;\\ u(0,x)=\, &\underline{u}_0(x), \quad x\in{\mathbb{R}}^d, \end{split} \end{array}\right.$$ with $\delta \in [0,\infty)$. We first give an a priori estimate for the solution of SPDE . \[prop-apriori-estm\] Let assumption $({\mathcal{A}}1)$ hold. Assume $\underline{u}_0\in L^2(\Omega,{\mathscr{F}}_0;H^m)$ and $f,g^k,h^k\in {\mathcal{L}}^2(H^m)$ with $m\in {\mathbb{R}}$, for $k=1,\dots,d_1$. If $u\in {\mathcal{S}}^2(H^{m+1})\cap{\mathcal{L}}^2(H^{m+2})$ is a solution of SPDE , one has $$\begin{aligned} &E\sup_{t\in[0,T]}\\|u(t)\\|_m^2+E\int_{0}^T\left(\delta \\|Du(t)\\|_m^2+\sum_{k=1}^{d_1}\\|L_ku(t)\\|_m^2\right)\,dt \nonumber\\ &\leq C\left\{ E\\|\underline{u}_0\\|_m^2 + E\int_{0}^T \left(\\|f(s)\\|_m^2+\\|g(s)\\|_m^2+\\|h(s)\\|_m^2\right)\,ds\right\},\label{estim-aprioi-prop}\end{aligned}$$ with $C$ being independent of $\delta$. We have decompositions $L_k=L_k'+c_k$ and $M_k=M_k'+\alpha_k$ with $c_k=-(D_i\sigma^{ik})\cdot$ and $\alpha_k=-(D_i\theta^{ik})\cdot$, for $k=1,\dots,d_1$. Applying Itô formula for the square norm (see e.g. [@Krylov_Rozovskii81 Theorem 3.1]), one has almost surely for $t\in[0,T]$, $$\begin{aligned} &\\|I^mu(t)\\|^2+\int_0^t 2\delta\\|I^mDu(s)\\|^2\,ds -\int_0^t \!\!\!2\langle I^m u(s),\,I^m(Du\theta+\beta u+h)(s)\,dW_s\rangle\nonumber \\ &=\\|I^m\underline{u}_0\\|^2+ \int_0^t\left\langle I^mu,\, I^m\left((L_k^2+M_k^2)u+2M_k'h^k+2L_k'g^k \right)\right\rangle(s)\,ds \nonumber\\ &\quad +\int_0^t2\left\langle I^mu,\,I^m\left(b^jD_ju+cu+f\right)\right\rangle(s)\,ds +\int_0^t\\|I^m(Du\theta+\beta u+h)(s)\\|^2\,ds. \label{eq-prop-ito}\end{aligned}$$ First, basic calculations yield $$\begin{aligned} &\langle I^mu,\,I^m(L_k^2u+2L_k'g^k+2cu+2f)\rangle\nonumber\\ &=\langle I^mu,\,I^m(L'_k+c_k)L_ku\rangle+2\langle I^mu,\,I^m L'_kg^k\rangle +2\langle I^mu,\,I^m(cu+f)\rangle\nonumber\\ &=-\\| I^mL_ku\\|^2 +\langle [I^m,L_k]u,\,I^mL_ku\rangle+\langle I^mu,\,[I^m,L_k']L_ku+ I^mc_kL_ku\rangle \nonumber\\ &\,\,\, -2\langle I^mL_ku,\,I^mg^k\rangle+2\langle[I^m,L_k]u,\,I^m g^k\rangle +2\langle I^mu,\,[I^m,L_k']g^k\rangle +2\langle I^mu,\,I^m(cu+f)\rangle\nonumber\\ &\leq-(1-{\varepsilon})\\|I^mL_ku\\|^2+C_{{\varepsilon}}\left(\\|I^mu\\|^2+\\|I^mg^k\\|^2+\\|I^mf\\|^2\right),\quad{\varepsilon}\in(0,1),\label{eq-rela-u}\end{aligned}$$ and $$\begin{aligned} &\langle I^mu,\,I^m(M_k^2u+2M_k'h^k)\rangle\nonumber\\ &=-\\| I^mM_ku\\|^2 +\langle [I^m,M_k]u,\,I^mM_ku\rangle+\langle I^mu,\,[I^m,M_k']M_ku+ I^m\alpha_kM_ku\rangle \nonumber\\ &\,\,\, -2\langle I^mM_ku,\,I^mh^k\rangle+2\langle[I^m,M_k]u,\,I^m h^k\rangle +2\langle I^mu,\,[I^m,M_k']h^k\rangle \nonumber\\ &\leq -\\| I^mM_ku\\|^2 -2\langle I^mM_ku,\,I^mh^k\rangle +\langle [I^m,M_k]u,\,M_kI^mu\rangle+\langle I^mu,\,[I^m,M_k]M_ku+ \alpha_kM_kI^mu\rangle \nonumber\\ &\,\,\, +C \left(\\|I^mu\\|^2+\\|I^mh^k\\|^2\right)\nonumber\\ &\leq -\\| I^mM_ku\\|^2 -2\langle I^mM_ku,\,I^mh^k\rangle +\langle I^mu,\,[[I^m,M_k],M_k]u+ \alpha_kM_kI^mu\rangle \nonumber\\ &\,\,\, +C \left(\\|I^mu\\|^2+\\|I^mh^k\\|^2\right)\nonumber\\ &\leq -\\| I^mM_ku\\|^2 -2\langle I^mM_ku,\,I^mh^k\rangle + C \left(\\|I^mu\\|^2+\\|I^mh^k\\|^2\right),\label{eq-relat-uxi}\end{aligned}$$ where we have used the relation $$\begin{aligned} \langle I^mu,\,\alpha_kM_kI^mu\rangle=-\frac{1}{2}\langle I^mu,\,D_i(\alpha_k\theta^{ik})I^mu\rangle.\label{eq-relat-adu}\end{aligned}$$ Notice that for $i=1,\dots,d$, $k=1,\dots,d_1$, $$\begin{aligned} \\|I^m(h^k+\beta^k u+M_ku)\\|^2 &=\\|I^mh^k\\|^2+2\langle I^mh^k,\,I^mM_ku\rangle+\\|I^mM_ku\\|^2\nonumber\\ &\,\,\,\,+2\langle I^m(h^k+M_ku), I^m(\beta^k u)\rangle + \\|I^m(\beta^k u)\\|^2,\label{esti-diff-prop3-1}\\ \langle I^mu,\,I^m(b^iD_iu)\rangle &=-\frac{1}{2} \langle I^mu,\,D_ib^iI^mu +2 [b^iD_i,\,I^m]u\rangle,\nonumber\\ \langle I^mM_ku,\,I^m(\beta^ku) \rangle &\leq \langle M_kI^mu,\,\beta^kI^mu\rangle+C\\|I^m u\\|^2\nonumber\\ &= -\frac{1}{2} \langle I^mu, \,D_i(\beta^k\theta^{ik})I^mu\rangle +C\\|I^m u\\|^2.\nonumber\end{aligned}$$ Putting , and together, and taking expectations on both sides of , one gets by Gronwall inequality $$\begin{aligned} &\sup_{t\in[0,T]}E\\|u(t)\\|_m^2+E\int_{0}^T\left(\delta\\|Du(t)\\|_m^2+\sum_{k=1}^{d_1}\\|L_ku(t)\\|_m^2\right)dt \nonumber\\ &\leq C\left\{ E\\|\underline{u}_0\\|_m^2 + E\int_{0}^T \left(\\|f(s)\\|_m^2+\\|g(s)\\|_m^2+\\|h(s)\\|_m^2\right)\,ds\right\}. \label{eq-prf-0}\end{aligned}$$ On the other hand, one has for each $t\in[0,T)$, $$\begin{aligned} &E\sup_{\tau\in[0,t]}\bigg\|\int_0^{\tau} \!\!\!2\left\langle I^m u(s),\,I^m(h+\beta u+Du\theta)(s)\,dW_s\right\rangle\bigg\|\\ &\leq C \bigg(E\sum_{k=1}^{d_1}\int_0^t\left(\|\langle I^mu(s),\,I^m(h^k+\beta^ku)(s)\rangle\|^2 +\|\langle I^mu(s),\,(M_kI^m+[I^m,M_k])u(s)\rangle\|^2\right)ds\bigg)^{1/2}\\ &\leq C \bigg(E\int_0^t \left(\\|I^mu(s)\\|^2\\|I^mh(s)\\|^2+\\|I^mu(s)\\|^4\right)ds \bigg)^{1/2 } \\ &\leq {\varepsilon}E\sup_{s\in[0,t]}\\|I^mu(s)\\|^2 +C_{{\varepsilon}} E\int_0^t \left(\\|I^mh(s)\\|^2+\\|I^mu(s)\\|^2\right)ds,\quad {\varepsilon}\in(0,1).\end{aligned}$$ Together with , , and , the above estimate implies . The estimate plays important role in our $L^2$-theory for SPDEs, for which some unusual techniques are applied in the calculations of , and . Especially, we treat the term $2\langle I^mM_ku,\,I^mh^k\rangle$ as a unity and it allows us to weaken the assumptions on $h$ in the $L^2$-theory. An immediate consequence of Proposition \[prop-apriori-estm\] is the following \[cor-uniqn\] Let assumption $({\mathcal{A}}1)$ hold. Given $\underline{u}_0\in L^2(\Omega,{\mathscr{F}}_0;H^m)$ and $f,g^k,h^k\in {\mathcal{L}}^2(H^m)$ with $m\in {\mathbb{R}}$, for $k=1,\dots,d_1$, the solution of SPDE is unique. \[thm-BSPDE\] Let assumption $({\mathcal{A}}1)$ hold. Assume $\underline{u}_0\in L^2(\Omega,{\mathscr{F}}_0;H^m)$ and $f,g^k,h^k\in {\mathcal{L}}^2(H^m)$ with $m\in {\mathbb{R}}$, for $k=1,\dots,d_1$. SPDE with $\delta=0$ (equivalently, SPDE ) admits a unique solution $u\in {\mathcal{S}}^2(H^{m})$ with $L_ku\in{\mathcal{L}}^2(H^m)$, $k=1,\dots,d_1$, and $$\begin{aligned} &E\sup_{t\in[0,T]}\\|u(t)\\|_m^2+\sum_{k=1}^{d_1}E\int_{0}^T\\|L_ku(t)\\|_m^2\,dt \nonumber\\ & \leq C\left\{ E\\|\underline{u}_0\\|_m^2 + E\int_{0}^T \left(\\|f(s)\\|_m^2+\\|g(s)\\|_m^2+\\|h(s)\\|_m^2\right)\,ds\right\},\label{estim-thm}\end{aligned}$$ with $C$ depending on $T,m,\sigma,\theta,b,c$ and $\beta$. Choose $\{\delta_l\}_{l\in{\mathbb{N}}^+}\subset(0,1)$, $\{\underline u^n_0\}_{n\in{\mathbb{N}}^+}\subset L^2(\Omega,{\mathscr{F}}_0;H^{m+5})$ and $\{f_n,g_n^k,h_n^k\}_{n\in{\mathbb{N}}^+}\subset {\mathcal{L}}^2(H^{m+5})$, for $k=1,\dots,d_1$, such that $\delta_l$ converges down to $0$ and $$\lim_{n\rightarrow \infty} \,\\|\underline{u}_0^n-\underline{u}_0\\|_{L^2(\Omega;H^m)} +\\|\left(f_n-f,g_n-g,h_n-h \right)\\|_{{\mathcal{L}}^2(H^m)} \,=0.$$ By $L^p$-theory for SPDEs (see [@Kryl96] for instance), SPDE admits a unique solution $u_{l,n}\in {\mathcal{S}}^2(H^{m+5})\cap {\mathcal{L}}^2(H^{m+6}) $ associated with $(\delta_l,f_n,g_n,h_n,\underline{u}^n_0)$. Fixing $n$, one deduces from Proposition \[prop-apriori-estm\] that $\{(u_{l,n},L_ku_{l,n})\}_{l\in{\mathbb{N}}^+}$ is bounded in ${\mathcal{S}}^2(H^{m+4})\times {\mathcal{L}}^2(H^{m+4})$, $k=1,\dots,d_1$. Notice that $\delta_l\Delta u_{l,n}$ tends to zero in ${\mathcal{L}}^2(H^{m+2})$ as $l$ goes to infinity. Therefore, letting $l$ tend to infinity, we derive from Proposition \[prop-apriori-estm\] and Corollary \[cor-uniqn\] the unique solution $u_n$ for SPDE associated with $(f_n,g_n,h_n,\underline{u}_0^n)$ and $\delta=0$ such that $(u_n,L_ku_n)\in{\mathcal{S}}^{2}(H^{m+2})\times {\mathcal{L}}^2(H^{m+2})$, for $k=1,\dots,d_1$. Furthermore, letting $n$ go to infinity, again by Proposition \[prop-apriori-estm\] and Corollary \[cor-uniqn\], one obtains the unique solution $u$ and associated estimates. This completes the proof. \[rmk-ito-forml\] Consider the particular case $m=0$ in Theorem \[thm-BSPDE\]. In view of the approximations in the above proof, through similar calculations as in the proof of Proposition \[prop-apriori-estm\], we can get the following estimate $$\begin{aligned} &\\|u(t)\\|^{2} -\int_0^t \!\left\langle u(s),\,(-D_i\theta^{ik}u+2\beta^k u+2h^k)(s)\,dW^k_s\right\rangle\nonumber \\ &\leq\\|\underline u_0\\|^{2} -(1-{\varepsilon})\int_0^t\sum_{k=1}^{d_1}\\| L_ku(s)\\|^2\,ds +C_{{\varepsilon}}\int_0^t \\|u(s)\\|^{2}\,ds \nonumber\\ &\quad + \int_0^t \left(\\|h(s)\\|^2+ 2\left\langle u(s),\,(L_k'g^k+cu+f)(s)\right\rangle\right)\,ds\quad \text{a.s.},\,\,\forall {\varepsilon}\in(0,1). \label{est-l2-3}\end{aligned}$$ Assume further $c\geq 0$. Put $u_{\lambda}=(u-{\lambda})^+:=\max\{u-{\lambda},0\}$ for ${\lambda}\in [0,\infty)$. If we start from the Itô formula for the square norm of the positive part of solution (see [@QiuWei-RBSPDE-2013 Corollary 3.11]), in a similar way to the above estimate, we have $$\begin{aligned} &\\|u_{\lambda}(t)\\|^{2} -\int_0^t \!\!\!\langle u_{\lambda}(s),\,(-D_i\theta^{ik}u_{\lambda}+2\beta^k u_{\lambda}+2h^k)(s)\,dW^k_s\rangle\nonumber \\ &\leq\\|u_{\lambda}(0)\\|^{2} -(1-{\varepsilon})\int_0^t\sum_{k=1}^{d_1}\\| L_ku_{\lambda}(s)\\|^2\,ds +C_{{\varepsilon}}\int_0^t \left(\\|u_{\lambda}(s)\\|^{2}+\langle \|u_{\lambda}\|,\,\lambda 1_{\{u_{\lambda}>0\}}\rangle (s)\right)\,ds \nonumber\\ &\quad + \int_0^t \left(\\|h(s)1_{\{u_{\lambda}>0\}}\\|^2+ 2\left\langle u_{\lambda}(s),\,(L_k'g^k+f)(s)\right\rangle\right)\,ds\quad \text{a.s.},\,\,\forall {\varepsilon}\in(0,1). \label{est-l2-uk-1}\end{aligned}$$ where we note that $u\leq u_{\lambda} + {\lambda}1_{\{u_{\lambda}>0\}}$. Note that we do not assume the Hörmander-type condition $({\mathcal{H}})$ in Theorem \[thm-BSPDE\]. In fact, we can get more regularity properties of solutions of SPDE under condition $({\mathcal{H}})$, for which we first recall an estimate on the Lie bracket. \[lem-lie-bracket\]([@Qiu-2014-Hormander Lemma 4.1]). For $\{J,L\}\subset\cup_{l\geq 0} {\mathbb{V}}_l$, $m\in{\mathbb{R}}$ and ${\varepsilon}\in[0,1]$, there exists a positive constant $C$ such that almost surely for any $\phi\in H^{m}$ with $J\phi\in H^{m-1+{\varepsilon}}$ and $L\phi\in H^m$, it holds that $$\begin{aligned} \\|[J,L]\phi\\|_{m-1+\frac{{\varepsilon}}{2}} \leq C\left( \\|J\phi\\|_{m-1+{\varepsilon}}+\\|L\phi\\|_{m}+\\|\phi\\|_m \right).\end{aligned}$$ The above lemma basically generalizes [@krylov2014hmander-PDE Lemma 4.2] from the deterministic case for $m=0$ to the stochastic case for any $m\in{\mathbb{R}}$. Starting from estimate of Theorem \[thm-BSPDE\] and applying Lemma \[lem-lie-bracket\] iteratively to elements of $\mathbb{V}_0,\dots,\mathbb{V}_{n_0}$, we have \[cor-grad-est\] Assume the same hypothesis of Theorem \[thm-BSPDE\]. Let condition $({\mathcal{H}})$ hold. For the unique solution $u$ of SPDE , one has further $u\in{\mathcal{L}}^2(H^{m+\eta})$ with $$E\int_0^T\\|u(s)\\|_{m+\eta}^2\,ds \leq C\left\{ E\\|G\\|_m^2 + E\int_0^T\left( \\|f(s)\\|_m^2+\\|g(s)\\|_{m}^2+\\|h(s)\\|_m^2 \right)ds \right\},$$ where the constant $C$ depends on $T,m,n_0,\sigma,\theta,b,c$ and $\beta$. The estimate on solution of SPDE for the case $m=0$ in Corollary \[cor-grad-est\] plays an important role in Section \[sec:lp-inf\] for the maximum principle of weak solutions. Therefore, for the reader’s convenience, we provide the following proof of Lemma \[lem-lie-bracket\] from which Corollary \[cor-grad-est\] follows immediately. Assume first $\phi\in H^{m+1}$. Setting $A^n=I^{n-1}[J,L]$, we have $A^n\in\Psi_{n}$ almost surely for each $n\in{\mathbb{R}}$. As the adjoint operator of $J$ and $L$, $J^=-J+\tilde{c}$ and $L^=-L+\bar{c}$ with $\tilde{c},\bar{c} \in {\mathcal{L}}^{\infty}(C^{\infty}_b)$. By Lemma \[lem-pdo\], we have $$\begin{aligned} &\langle JL\phi,\,I^mA^{m-1+{\varepsilon}}\phi \rangle\\ &=\langle L\phi,\,(I^mJ^+[J^,I^m])A^{m-1+{\varepsilon}}\phi\rangle\\ &=\langle I^mL\phi,\,(A^{m-1+{\varepsilon}}J^+[J^,A^{m-1+{\varepsilon}}])\phi \rangle +\langle [I^m,J]L\phi,\,A^{m-1+{\varepsilon}}\phi\rangle\\ &\leq C\left( \\|L\phi\\|_m^2+ \\|J\phi\\|^2_{m-1+{\varepsilon}}+\\|\phi\\|_m^2 \right)\end{aligned}$$ and $$\begin{aligned} &\langle LJ\phi,\,I^mA^{m-1+{\varepsilon}}\phi \rangle\\ &=\langle J\phi,\,(I^{m-1+{\varepsilon}}L^+[L^,I^{m-1+{\varepsilon}}])A^m\phi\rangle\\ &=\langle I^{m-1+{\varepsilon}}J\phi,\,(A^mL^+[L^,A^m])\phi\rangle +\langle I^{m-1+{\varepsilon}}J\phi,\, I^{-(m-1+{\varepsilon})}[L^,I^{m-1+{\varepsilon}}]A^m\phi\rangle\\ &\leq C\left(\\|J\phi\\|_{m-1+{\varepsilon}}^2+\\|L\phi\\|_m^2+\\|\phi\\|_m^2 \right).\end{aligned}$$ Hence, $$\begin{aligned} \\|[J,L]\phi\\|_{m-1+\frac{{\varepsilon}}{2}} =\langle [J,L]\phi,\,I^mA^{m-1+{\varepsilon}}\phi \rangle^{\frac{1}{2}} \leq C\left(\\|J\phi\\|_{m-1+{\varepsilon}}+\\|L\phi\\|_m+\\|\phi\\|_m \right).\end{aligned}$$ Through standard density arguments, one verifies that the above estimate also holds for any $\phi\in H^{m}$ with $J\phi\in H^{m-1+{\varepsilon}}$ and $L\phi\in H^m$. Hörmander theorem for SPDEs --------------------------- Inspired by the filtering theory of partially observable diffusion processes, Krylov [@krylov2013hypoellipticity; @krylov2013-Hormder-SPDE] has just obtained the Hörmander theorem for SPDEs, which states the spacial smoothness of solutions. The method therein relies on the generalized Itô-Wentzell formula and associated results on deterministic PDEs. Next to the above established $L^2$-theory, we intend to derive the following Hörmander theorem for SPDE under the condition $({\mathcal{H}})$ with an analytical approach. \[thm-hormander\] Let assumptions $({\mathcal H} )$ and $({\mathcal{A}}1)$ hold. If $f\in\cap_{n\in{\mathbb{R}}}{\mathcal{L}}^2(H^n)$, $g,h\in\cap_{n\in{\mathbb{R}}}{\mathcal{L}}^2((H^n)^{d_1})$, and $\underline{u}_0\in L^2(\Omega;H^m)$ for some $m\in{\mathbb{R}}$, then for the unique solution $u$ of SPDE in Theorem \[thm-BSPDE\], one has for any ${\varepsilon}\in(0,T)$, $$u\in\cap_{n\in{\mathbb{R}}} L^2(\Omega;C([{\varepsilon},T];H^n)),$$ and for any $n\in{\mathbb{R}}$, $$\begin{aligned} &E\sup_{t\in[{\varepsilon},T]}\\| u(t)\\|_{n}^2 +E\int_{{\varepsilon}}^{T} \\| u(t)\\|_{n+\eta}^2\,dt \nonumber\\ &\leq C\left\{E\\|G\\|_m^2+ E\int_{0}^{T} \left(\\|f(s)\\|_{n}^2+ \\|g(s)\\|_{n}^2+ \\|h(s)\\|_{n}^2\right) ds\right\},\label{est-hmder}\end{aligned}$$ with the constant $C$ depending on ${\varepsilon},n,T,m,n_0,\sigma,\theta,\gamma,b$ and $c$. In particular, the random field $u(t,x)$ is almost surely infinitely differentiable with respect to $x$ on $(0,T]\times{\mathbb{R}}^d$ and each derivative is a continuous function on $(0,T]\times{\mathbb{R}}^d$. \[sec:proof-main-thm\] By Theorem \[thm-BSPDE\], SPDE admits a unique solution $u\in{\mathcal{S}}^2(H^m)$ and the random field $\bar{u}(t,x):=tu(t,x)$ is the unique solution of SPDE $$\label{SPDE-1} \left\{\begin{array}{l} \begin{split} d\bar{u}(t,x)=\,&\displaystyle \left[ \frac{1}{2}(L_k^2+M_k^2)\bar{u} +{b}^jD_j\bar{u}+c\bar{u}+u+t\left(f+L_k'g^k+M_k'h^k\right) \right](t,x)\, dt\\ &\displaystyle +\left[tM_k\bar{u}+t\beta^k \bar{u}+th^k\right](t,x)\, dW_{t}^{k}, \quad (t,x)\in Q;\\ \bar{u}(0,x)=\, &0, \quad x\in{\mathbb{R}}^d, \end{split} \end{array}\right.$$ with $$\begin{aligned} &E\sup_{t\in[0,T]}\\|\bar u(t)\\|_m^2+\sum_{k=1}^{d_1}E\int_{0}^T\\|L_k\bar u(t)\\|_m^2dt \nonumber\\ &\leq C\left(T^2+1\right)E\int_{0}^T \left(\\|f(s)\\|_m^2+\\|g(s)\\|_m^2+\\|h(s)\\|_m^2+\\|u(s)\\|_m^2\right)\,ds.\end{aligned}$$ Starting from the above estimate, we apply Lemma \[lem-lie-bracket\] iteratively to elements of ${\mathbb{V}}_0,\dots,{\mathbb{V}}_{n_0}$. Under condition $({\mathcal{H}})$, there arrives the estimate $$\begin{aligned} \label{grad-est} \int_0^T\\|D\bar{u}\\|^2_{m-1+\eta}ds\leq C\left(T^2+1\right)E\int_{0}^T \left(\\|f(s)\\|_m^2+\\|g(s)\\|_m^2+\\|h(s)\\|_m^2+\\|{u}(s)\\|_m^2\right)\,ds.\end{aligned}$$ Fix any ${\varepsilon}\in(0,T\wedge 1)$ and define ${\varepsilon}_l=\sum_{i=1}^l\frac{{\varepsilon}}{2^i}$ for $l\in{\mathbb{N}}^+$. By interpolation and Theorem \[thm-BSPDE\], we have $$\begin{aligned} &E\sup_{t\in[{\varepsilon}_1,T]}\\| u(t)\\|_m^2 +E\int_{{\varepsilon}_1}^{T}\\| u(t)\\|_{m+2^{-n_0}}^2 dt \nonumber\\ & \leq \frac{C2(T^2+1)}{{\varepsilon}}E\int_{0}^{T} \left(\\|f(s)\\|_m^2+\\|g(s)\\|_m^2+\\|h(s)\\|_m^2+\\|u(s)\\|_m^2\right)\,ds.\end{aligned}$$ Since $f\in\cap_{n\in{\mathbb{R}}}{\mathcal{L}}^2(H^n)$ and $g,h\in \cap_{n\in{\mathbb{R}}}{\mathcal{L}}^2((H^n)^{d_1})$, by iteration we obtain for any $j\in{\mathbb{N}}^+$, $$\begin{aligned} &E\sup_{t\in[{\varepsilon}_j,T]}\\| u(t)\\|_{m+(j-1)\eta}^2 +E\int_{{\varepsilon}_j}^{T} \\| u(t)\\|_{m+j\eta}^2\,dt \nonumber\\ &\leq \frac{C2^j(T^2+1)}{{\varepsilon}} E\int_{{\varepsilon}_{j-1}}^{T} \!\!\!\left(\\|f(s)\\|_{m+(j-1)\eta}^2+ \\|g(s)\\|_{m+(j-1)\eta}^2+ \\|h(s)\\|_{m+(j-1)\eta}^2+\\|u(s)\\|_{m+(j-1)\eta}\right)\! ds,\nonumber\end{aligned}$$ which together with estimate , implies by iteration that $$\begin{aligned} &E\sup_{t\in[{\varepsilon}_j,T]}\\| u(t)\\|_{m+(j-1)\eta}^2 +E\int_{{\varepsilon}_j}^{T} \\| u(t)\\|_{m+j\eta}^2\,dt \nonumber\\ &\leq C(j)\left\{E\\|G\\|_m^2+ E\int_{0}^{T} \left(\\|f(s)\\|_{m+(j-1)\eta}^2+ \\|g(s)\\|_{m+(j-1)\eta}^2+ \\|h(s)\\|_{m+(j-1)\eta}^2\right) ds\right\}.\nonumber\end{aligned}$$ Hence, for any ${\varepsilon}\in (0,T)$, one has $u\in\cap_{n\in{\mathbb{R}}} L^2(\Omega;C([{\varepsilon},T];H^n))$ and the estimate holds. In particular, by Sobolev embedding theorem, $u(t,x)$ is almost surely infinitely differentiable with respect to $x$ and each derivative is a continuous function on $(0,T]\times{\mathbb{R}}^d$. By Theorem \[thm-hormander\], we have the global spacial smoothness of the solution in time interval $(0,T]$. A similar result exists in Krylov’s recent work [@krylov2013hypoellipticity; @krylov2013-Hormder-SPDE], which states the local spacial smoothness of solution under a Hörmander-type condition of local type. However, the method therein relies on the generalized Itô-Wentzell formula and associated results on deterministic PDEs, while herein, we use an analytical approach on the basis of our $L^2$-theory and an estimate on the Lie bracket (Lemma \[lem-lie-bracket\]). In fact, our method has the potential to derive the associated local results, but we would not seek such a generality in the present paper. In addition, we would mention that, to the best of our knowledge, the hypoellipticity for SPDEs was first considered by Chaleyat-Maurel and Michel [@chaleyat1984hypoellipticity], where the coefficients depend on $(t,\omega)$ only through a substituted Wiener process. $L^p$ estimates for the uniform norm of solutions {#sec:lp-inf} ================================================= To the end, let assumptions $({\mathcal{A}}1)$, $({\mathcal{A}}2)$ and $({\mathcal{H}})$ hold. By Theorem \[thm-BSPDE\], SPDE has a unique weak* solutoin. In this section, we shall prove the $L^p$-estimates for the time-space uniform norm of the weak solution. \[thm-Lp-inf\] For the weak solution $u$ of SPDE , there exists $\theta_0\in (0,1]$ such that for any $p\in(0,\infty)$, $$\begin{aligned} E\\|u^{\mp}\\|^p_{L^{\infty}(Q)} \leq C \left(\Lambda^{\mp}_{\bar p,\infty} + \Lambda^{\mp}_{\frac{p}{\theta_0}} \right)^p,\end{aligned}$$ with the constant $C$ depending on $d,p,n_0, T$ and the quantities related to the coefficients $\sigma,\theta,b,c$ and $\beta$. An immediate consequence is the following comparison principle. Suppose that random field $u$ is the weak solution of SPDE . Let $\tilde u$ be the solution of SPDE with the initial value $\underline u_0$ and external force $f$ being replaced by $\tilde{\underline u}_0$ and $\tilde f$ respectively. Suppose further that $$f\leq \tilde f,\quad \mathbb{P}\otimes dt \otimes dx\text{-a.e. and }\underline{u}_0\leq \tilde{\underline u}_0,\quad \mathbb{P}\otimes dx\text{-a.e.}$$ Then, there holds $u\leq \tilde u$, $\mathbb{P}\otimes dt \otimes dx$-a.e. Before proving Theorem \[thm-Lp-inf\], we give the following embedding lemma that will be used frequently in what follows. \[lem-embedding\] For $\psi\in L^2(0,T;H^{\eta})\cap C([0,T];L^2)$, one has $\psi\in L^{\frac{2(d+2\eta)}{d}}(Q)$ and $$\begin{aligned} \\|\psi\\|_{L^{\frac{2(d+2\eta)}{d}}(Q)} \leq \\|\psi\\|^{\frac{d}{d+2\eta}}_{L^2(0,T;H^{\eta})} \\|\psi\\|^{\frac{2\eta}{d+2\eta}}_{C([0,T];L^2)} \label{eq-lem-embedding}\end{aligned}$$ with the positive constant $C$ depending on $d$ and $\eta$. By the fractional Gagliard-Nirenberg inequality (see [@Hajaiej-2012-frac-Gargld-Nirenberg Corollary 2.3] for instance), we have $$\\|\psi(s,\cdot)\\|_{L^q}^q \leq\ C~ \\|\psi(s,\cdot)\\|_{\eta}^{\alpha q}\\|\psi(s,\cdot)\\|^{q(1-\alpha)}, \quad \text{a.e.}\ s\in[0,T],$$ where $\alpha=d/(d+2\eta)$ and $q=2(d+2\eta)/d$. Integrating on $[0,T]$, we obtain $$\begin{split} \int_{Q}\|\psi(s,x)\|^qdxds \leq &\ C~ \\|\psi\\|_{\eta}^{2} \max_{s\in [0,T]}\\|\psi(s,\cdot)\\|^{(1-\alpha) q}. \end{split}$$ Therefore, $\psi\in L^{\frac{2(d+2\eta)}{d}}(Q)$ and there holds . For $\lambda>0$ and $z\in{\mathbb{N}}_0$, set $$\begin{aligned} u_z=(u-\lambda(1-2^{-z}))^+ \quad \textrm{and} \quad U_z=\sup_{t\in[0,T]}\\|u_z(t)\\|^2+\int_0^T\left( \\|u_z(t)\\|_{\eta}^2+\sum_{k=1}^{d_1}\\|L_ku_z(t)\\|^2 \right)dt.\end{aligned}$$ Obviously, for each $z\in{\mathbb{N}}^+$, one has $\|D_i u_{z-1}\|\geq \|D_i u_z\| $ for $i=1,\dots,d$, $$\begin{aligned} u_{z-1}\geq u_z ,\,\, u1_{\{u_z>0\}}=u_z+\lambda(1-2^{-z})1_{\{u_z>0\}} \quad \text{and}\quad 1_{\{u_z>0\}}\leq \left(\frac{2^zu_{z-1}}{\lambda}\right)^{q},\quad\forall \, q>0.\label{relatn-uz}\end{aligned}$$ As an immediate consequence of Lemma \[lem-embedding\], there follows $$\\|u_z\\|^2_{L^{\frac{2(d+2\eta)}{d}}(Q)} \leq C\, U_z,\quad a.s.$$ with the constant $C$ depending on $d$ and $\eta$. In view of Remark \[rmk-ito-forml\], the weak solution $u$ of SPDE satisfies $$\begin{aligned} &\\|u_z(t)\\|^{2} -\int_0^t \!\!\!\langle u_z(s),\,(-D_i\theta^{ik}u_z+2\beta^k u_z+2h^k)(s)\,dW^k_s\rangle\nonumber \\ &\leq\\|u_z(0)\\|^{2} -(1-{\varepsilon})\int_0^t\sum_{k=1}^{d_1}\\| L_ku_z(s)\\|^2\,ds +C_{{\varepsilon}}\int_0^t \left(\\|u_z(s)\\|^{2}+\left\langle \|u_z\|,\,\lambda (1-2^{-z})1_{\{u_z>0\}}\right\rangle (s)\right)\,ds \nonumber\\ &\quad + \int_0^t \left(\\|h(s)1_{\{u_z>0\}}\\|^2+ 2\left\langle u_z(s),\,(L_k'g^k+f)(s)\right\rangle\right)\,ds,\quad\text{a.s.,}\,\,\forall {\varepsilon}\in(0,1). \label{est-l2-uk-1}\end{aligned}$$ Taking ${\varepsilon}=1/2$, we have by Gronwall inequality $$\begin{aligned} &\sup_{s\in[0,t]}\\|u_z(s)\\|^{2}+\int_0^t\sum_{k=1}^{d_1}\\| L_ku_z(s)\\|^2\,ds \nonumber \\ &\leq C\bigg\{ \lambda (1-2^{-z})\int_0^t\langle \|u_z\|,\,1_{\{u_z>0\}}\rangle (s)\,ds +\sup_{\tau\in[0,t]}\int_0^{\tau} \!\!\left\langle u_z(s),\,(-D_i\theta^{ik}u_z+2\beta^k u_z+2h^k)(s)\,dW^k_s\right\rangle \nonumber\\ &\quad + \int_0^t \left(\\|h(s)1_{\{u_z>0\}}\\|^2+ 2\left\|\left\langle u_z(s),\,(L_k'g^k+f^+)(s)\right\rangle\right\|\right)\,ds +\\|u_z(0)\\|^{2} \bigg\},\quad\text{a.s.}\end{aligned}$$ Under condition $({\mathcal{H}})$, starting from the above estimate and applying Lemma \[lem-lie-bracket\] iteratively to elements of ${\mathbb{V}}_0,\dots,{\mathbb{V}}_{n_0}$, we get $$\begin{aligned} &\sup_{s\in[0,t]}\\|u_z(s)\\|^{2}+\int_0^t\left(\\|u_z(s)\\|^2_{\eta}+\sum_{k=1}^{d_1}\\| L_ku_z(s)\\|^2\right)\,ds \nonumber \\ &\leq C\bigg\{ \lambda (1-2^{-z})\int_0^t\langle \|u_z\|,\,1_{\{u_z>0\}}\rangle (s)\,ds +\sup_{\tau\in[0,t]}\int_0^{\tau} \!\!\left\langle u_z(s),\,(-D_i\theta^{ik}u_z+2\beta^k u_z+2h^k)(s)\,dW^k_s\right\rangle \nonumber\\ &\quad + \int_0^t \left(\\|h(s)1_{\{u_z>0\}}\\|^2+ 2\left\|\left\langle u_z(s),\,(L_k'g^k+f^+)(s)\right\rangle\right\|\right)\,ds +\\|u_z(0)\\|^{2} \bigg\} ,\quad\text{a.s.}\label{eq-ito-uz}\end{aligned}$$ Set $$M_z(t)=\int_0^t\left\langle u_z(s),\,(-D_i\theta^{ik}u_z+2\beta^k u_z+2h^k)(s)\,dW^k_s \right\rangle,\quad t\in [0,T].$$ The proof of Theorem \[thm-Lp-inf\] is started from the iteration inequality of the following lemma. \[lem-iteration\] Assume $\lambda \geq 2 \Lambda_{\bar p,\infty}^+ >1 $. For the solution of SPDE , there exists a positive constant $N$ such that for any $z\in{\mathbb{N}}^+$, $$\begin{aligned} U_z\leq \frac{N^z}{\lambda^{2\alpha_0}} \left( U_{z-1}\right)^{1+\alpha_0} + N\sup_{t\in[0,T]} M_z(t), \quad\text{a.s.}\label{itrn-schem}\end{aligned}$$ where $$0<\alpha_0:=\frac{(\bar{p}-2\eta)(d+2\eta)}{2\bar p d}-\frac{1}{2}.$$ We estimate each item involved in relation . Since $\bar{p}>d+2\eta$, basic calculations yield that $2<2+4\alpha_0<\frac{2(d+2\eta)}{d}$. Then, it holds that $$\begin{aligned} &\lambda (1-2^{-z})\int_0^T\langle \|u_z\|,\,1_{\{u_z>0\}}\rangle (s)\,ds \\ &\leq \lambda (1-2^{-z})\int_0^T\left\langle \|u_{z-1}\|,\,\left(\frac{2^zu_{z-1}}{\lambda}\right)^{1+2\alpha_0} \right\rangle (s)\,ds \\ &= \frac{(1-2^{-z})2^{(1+2\alpha_0)z}}{\lambda^{2\alpha_0}} \\|u_{z-1}\\|_{L^{2+2\alpha_0}(Q)}^{2+2\alpha_0} \\ &\leq \frac{(1-2^{-z})2^{(1+2\alpha_0)z}}{\lambda^{2\alpha_0}} \\|u_{z-1}\\|_{L^{\frac{2(d+2\eta)}{d}}(Q)}^{(2+2\alpha_0){\varepsilon}} \\|u_{z-1}\\|_{L^{2}(Q)}^{(2+2\alpha_0)(1-{\varepsilon})} \\ &\leq \frac{C(1-2^{-z})2^{(1+2\alpha_0)z}}{\lambda^{2\alpha_0}} \left(U_{k-1}\right)^{{1+\alpha_0}},\quad\text{a.s.}\end{aligned}$$ where by Lyapunov’s inequality, ${\varepsilon}\in(0,1)$ is chosen to satisfy $$\frac{1}{2+2\alpha_0}=\frac{d{\varepsilon}}{2(d+2\eta)}+\frac{1-{\varepsilon}}{2}.$$ Furthermore, we have $$\begin{aligned} &\int_0^T \langle u_z,\,f^+\rangle (s) \,ds\\ &\leq \\|u_z\\|_{L^{\frac{2(d+2\eta)}{d}}(Q)} \\|f^+\\|_{L^{\frac{\bar p (d+2\eta)}{(\bar p + d+2\eta)\eta}}(Q)} \left(\int_{Q}1_{\{u_z>0\}}\,dxds\right)^{\frac{1}{2}-\frac{\eta}{\bar p}}\\ &\leq \\|u_z\\|_{L^{\frac{2(d+2\eta)}{d}}(Q)} \\|f^+\\|_{L^{\frac{\bar p (d+2\eta)}{(\bar p + d+2\eta)\eta}}(Q)} \left(\int_{Q} \left\|\frac{2^zu_{z-1}}{\lambda}\right\|^{\frac{2(d+2\eta)}{d}} \,dxds\right)^{\frac{1}{2}-\frac{\eta}{\bar p}} \\ &\leq \left(\frac{2^z}{\lambda}\right)^{1+2\alpha_0} \\|f^+\\|_{L^{\frac{\bar p (d+2\eta)}{(\bar p + d+2\eta)\eta}}(Q)} \\|u_{z-1}\\|_{L^{\frac{2(d+2\eta)}{d}}(Q)}^{2+2\alpha_0} \\ &\leq C \left(\frac{2^z}{\lambda}\right)^{1+2\alpha_0}\\|f^+\\|_{L^{\frac{\bar p (d+2\eta)}{(\bar p + d+2\eta)\eta}}(Q)} (U_{z-1})^{1+\alpha_0}, \quad\text{a.s.}\end{aligned}$$ and $$\begin{aligned} &\int_0^T\|\langle u_z,\,L_k'g^k \rangle (s)\|\,ds \\ &=\int_0^T\|\langle L_ku_z,\,g^k \rangle (s)\|\,ds \\ &\leq \\|L_ku_z\\|_{L^2(Q)}\left(\int_Q g^21_{\{u_z>0\}}\,dxds\right)^{\frac{1}{2}}\\ & \leq \\|L_ku_z\\|_{L^2(Q)} \\|g\\|_{L^{\frac{\bar p}{\eta}}(Q)} \left(\int_Q 1_{\{u_z>0\}}\,dxds\right)^{\frac{1}{2}-\frac{\eta}{\bar p}}\\ &\leq \left(\frac{2^z}{\lambda}\right)^{1+2\alpha_0} \\|g\\|_{L^{\frac{\bar p}{\eta}}(Q)} \\|L_ku_{z-1}\\|_{L^2(Q)} \\|u_{z-1}\\|_{L^{\frac{2(d+2\eta)}{d}}(Q)}^{1+2\alpha_0} \\ &\leq C \left(\frac{2^z}{\lambda}\right)^{1+2\alpha_0}\\|g\\|_{L^{\frac{\bar p}{\eta}}(Q)} (U_{z-1})^{1+\alpha_0},\quad\text{a.s.}\end{aligned}$$ Let $q=\frac{\bar{p}(d+2\eta)}{\eta(\bar{p}+d+2\eta)}$ and $\tilde q = \frac{q}{q-1}$. There follows $\frac{2(d+2\eta)}{d\tilde{q}}=2+2\alpha_0$ and thus $$\begin{aligned} &\int_0^T \\|h(s) 1_{\{u_z>0\}} \\|^2ds\\ &\leq \\|h\\|_{L^{2q}(Q)}^{2}\left( \int_Q 1_{\{u_z>0\}}dxds \right)^{\frac{1}{\tilde q }}\\ &\leq \\|h\\|_{L^{2q}(Q)}^{2}\left( \int_Q \left( \frac{2^zu_{z-1}}{\lambda} \right)^{\frac{2(d+2\eta)}{d}}dxds \right)^{\frac{1}{\tilde q }}\\ &= \left(\frac{2^z}{\lambda}\right)^{2+2\alpha_0} \\|h\\|_{L^{2q}(Q)}^{2} \\| u_{z-1} \\|_{L^{\frac{2(d+2\eta)}{d}}(Q)} ^{2+2\alpha_0} \\ &\leq \frac{C2^{(2+2\alpha_0)z}}{\lambda^{2+2\alpha_0}} \\|h\\|_{L^{2q}(Q)}^{2} \left( U_{z-1} \right)^{1+\alpha_0},\quad\text{a.s.}\end{aligned}$$ Since $\lambda\geq 2 \Lambda^+_{\bar p,\infty}$, it follows that $u_z(0)\equiv 0$ for any $z\in{\mathbb{N}}^+$. Choosing $N$ to be big enough, we have by relation , $$\begin{aligned} U_z\leq \frac{N^z}{\lambda^{2\alpha_0}} \left( U_{z-1}\right)^{1+\alpha_0} + N\sup_{t\in[0,T]} M_z(t),\quad \text{a.s.}\end{aligned}$$ Next, let us deal with the martingale part $M_z(\cdot)$ in the iteration inequality . We shall prove that $M_z(\cdot)$ is comparable with $\left(U_{z-1}\right)^{1+\alpha_0}$. \[lem-est-prob\] Let $\lambda\geq \Lambda^+_{\bar p,\infty}$. There exists $N\in(1,\infty)$ such that for any $\kappa,\zeta\in(0,\infty)$, $${\mathbb{P}}\left(\left\{ \sup_{t\in[0,T]} M_z(t) \geq \kappa\zeta,\,\left(U_{z-1}\right)^{1+\alpha_0} \leq \zeta \right\}\right) \leq \exp\left\{-\frac{\kappa^2\lambda^{4\alpha_0}}{2N^z}\right\},\quad \forall\,z\in{\mathbb{N}}^+.$$ First, we have $$\begin{aligned} \langle M_z \rangle_T &=\sum_{k=1}^{d_1}\int_0^T \left\| \left\langle u_z,\,(-D_i\theta^{ik}u_z+2\beta^k u_z+2h^k)(s)\right\rangle \right\|^2ds\\ &\leq C\int_0^T \left( \\|u_z\\|^4 + \\|u_z\\|^2\\|h1_{\{u_z>0\}}\\|^2 \right)ds,\quad \text{a.s.}\end{aligned}$$ with the constant $C$ being independent of $z$. On the other hand, we have $$\begin{aligned} &\int_0^T \\|u_z\\|^2\\|h1_{\{u_z>0\}}\\|^2ds\\ &\leq \sup_{s\in[0,T]}\\|u_z(s)\\|^2\int_0^T \\|h1_{\{u_z>0\}}\\|^2ds\\ &\leq \sup_{s\in[0,T]}\\|u_z(s)\\|^2 \\|h\\|_{L^{\frac{\bar p}{\eta}}(Q)}^2 \left( \int_{Q} 1_{\{u_z>0\}} dxds\right)^{1-\frac{2\eta}{\bar p}} \\ &\leq \\|h\\|_{L^{\frac{\bar p}{\eta}}(Q)}^2 \sup_{s\in[0,T]}\\|u_z(s)\\|^2 \left(\int_{Q} \left\|\frac{2^zu_{z-1}}{\lambda}\right\|^{\frac{2(d+2\eta)}{d}} \,dxds\right)^{1-\frac{2\eta}{\bar p}} \\ &=\left(\frac{2^z}{\lambda}\right)^{2+4\alpha_0} \\|h\\|_{L^{\frac{\bar p}{\eta}}(Q)}^2 \sup_{s\in[0,T]}\\|u_z(s)\\|^2 \\|u_{z-1}\\|^{2+4\alpha_0}_{L^{\frac{2(d+2\eta)}{d}}(Q)} \\ & \leq C \left(\frac{2^z}{\lambda}\right)^{2+4\alpha_0} \\|h\\|_{L^{\frac{\bar p}{\eta}}(Q)}^2 \left( U_{z-1} \right)^{2+2\alpha_0},\quad \text{a.s.}\end{aligned}$$ and $$\begin{aligned} &\int_0^T \\|u_z(s)\\|^4ds \\ &\leq \sup_{s\in[0,T]}\\|u_z(s)\\|^2\int_0^T \\|u_z\\|^2ds\\ &\leq \sup_{s\in[0,T]}\\|u_{z}(s)\\|^2 \int_{Q}\left\|u_z\right\|^2 \left\|\frac{2^zu_{z-1}}{\lambda}\right\|^{4\alpha_0} \,dxds \\ &\leq \left(\frac{2^z}{\lambda}\right)^{4\alpha_0} \sup_{s\in[0,T]}\\|u_z(s)\\|^2 \\|u_{z-1}\\|^{2+4\alpha_0}_{2+4\alpha_0} \\ &\leq C \left(\frac{2^z}{\lambda}\right)^{4\alpha_0} \left( U_{z-1} \right)^{2+2\alpha_0},\quad\text{a.s.}\end{aligned}$$ Therefore, there exists $N\in (1,\infty)$ such that for any $z\in{\mathbb{N}}^+$, $$\begin{aligned} \langle M_z \rangle_T \leq C \left\{ \left(\frac{2^z}{\lambda}\right)^{4\alpha_0} +\left(\frac{2^z}{\lambda}\right)^{2+4\alpha_0} \\|h\\|_{L^{\frac{\bar p}{\eta}}(Q)}^2 \right\} \left( U_{z-1} \right)^{2+2\alpha_0} \leq \frac{N^z}{\lambda^{4\alpha_0}} \left( U_{z-1} \right)^{2+2\alpha_0} ,\quad \text{a.s.}\label{eq-martg}\end{aligned}$$ with the constant $C$ being independent of $z$. In view of relation , $\left(U_{z-1}\right)^{1+\alpha_0} \leq \zeta$ implies that $\langle M_z \rangle_T \leq \gamma := \frac{N^z\zeta^2}{\lambda^{4\alpha_0}}$. Note that there exists a Brownian motion $B$ such that $M_t=B_{\langle M \rangle_t}$. Hence, $$\begin{aligned} {\mathbb{P}}\left(\left\{ \sup_{t\in[0,T]} M_z(t) \geq \kappa\zeta,\,\left(U_{z-1}\right)^{1+\alpha_0}\leq \zeta \right\}\right) &\leq {\mathbb{P}}\left(\left\{ \sup_{t\in[0,T]} M_z(t) \geq \kappa\zeta,\,\langle M_z \rangle_T \leq \gamma \right\}\right) \\ &\leq {\mathbb{P}}\left(\left\{ \sup_{t\in[0,\gamma]} B_t \geq \kappa\zeta \right\}\right)\\ \text{(by the reflection principle) } &=2 {\mathbb{P}}\left(\left\{ B_{\gamma} \geq \kappa\zeta \right\}\right) \\ &\leq \exp\left\{-\frac{\kappa^2\zeta^2}{2\gamma}\right\}=\exp\left\{-\frac{\kappa^2\lambda^{4\alpha_0}}{2N^z}\right\},\end{aligned}$$ which completes the proof. Combining the iteration inequality and the estimate on martingale part $M_z(\cdot)$, we shall estimate the tail probability of $\\|u^{+}\\|_{L^{\infty}(Q)}$. \[lem-uplus\] There exist $\theta_0\in(0,1)$ and $\lambda_0\in (1,\infty)$ such that for any $\lambda\geq \lambda_0 $, $$\begin{aligned} {\mathbb{P}}\left(\left\{ \\|u^+\\|_{L^{\infty}(Q)} >\lambda, \, U_0 \leq \lambda^{2\theta_0} \right\}\right) \leq 2\exp\left\{-\lambda^{2\alpha_0} \right\}. \label{est-lem-uplus}\end{aligned}$$ For $z\in{\mathbb{N}}_0$, set $$A_z=\left\{ U_z\leq \frac{\lambda^{2\theta_0}}{\nu^z} \right\},$$ with the parameter $\nu>1$ waiting to be determined later. Observe that $$\begin{aligned} \left\{ \\|u^+\\|_{L^{\infty}(Q)} >\lambda, \, U_0 \leq \lambda^{2\theta_0} \right\} \subset \cup_{z\in{\mathbb{N}}_0} \left(A_z\right)^c \cap A_0 \subset \cup_{z\in {\mathbb{N}}^+} \left(A_z\right)^c\cap A_{z-1}\end{aligned}$$ which implies that $$\begin{aligned} {\mathbb{P}}\left(\left\{ \\|u^+\\|_{L^{\infty}(Q)} >\lambda, \, U_0 \leq \lambda^{2\theta_0} \right\}\right) \leq \sum_{z\in {\mathbb{N}}^+} {\mathbb{P}}\left(\left(A_z\right)^c\cap A_{z-1}\right). \label{eq-lem-prob}\end{aligned}$$ In view of Lemma \[lem-iteration\], the event in $\left(A_z\right)^c\cap A_{z-1}$ implies that $$\begin{aligned} \sup_{t\in[0,T]} M_z(t) &\geq \frac{\lambda^{2\theta_0}}{N\nu^z}-\frac{N^{z-1}}{\lambda^{2\alpha_0-2\theta_0(1+\alpha_0)}\nu^{(z-1)(1+{\alpha_0})}}\\ &=\frac{\lambda^{2\theta_0(1+\alpha_0)}}{\nu^{(z-1)(1+\alpha_0)}}\left[ \frac{\nu^{\alpha_0z-1-\alpha_0}}{N\lambda^{2\alpha_0\theta_0}} - \frac{N^{z-1}}{\lambda^{2\alpha_0}} \right].\end{aligned}$$ Put $$\zeta_z= \frac{\lambda^{2\theta_0(1+\alpha_0)}}{\nu^{(z-1)(1+\alpha_0)}} \quad\text{and}\quad \kappa_z= \frac{\nu^{\alpha_0z-1-\alpha_0}}{N\lambda^{2\alpha_0\theta_0}} - \frac{N^{z-1}}{\lambda^{2\alpha_0}},$$ and take $$\theta_0=\frac{1}{4} \quad \text{and}\quad \nu=(2N+1)^{\frac{1}{\alpha_0}}.$$ There exists $\lambda_0\in (1,\infty)$ such that for any $\lambda \geq \lambda_0$, one has $$\kappa_z\geq \frac{(2N+1)^{z}}{\lambda^{\alpha_0}},\quad \forall \,z\in{\mathbb{N}}^+.$$ By Lemma \[lem-est-prob\], it follows that for any $z\in{\mathbb{N}}^+$, $$\begin{aligned} {\mathbb{P}}\left(\left(A_z\right)^c\cap A_{z-1}\right) &\leq {\mathbb{P}}\left( \left\{ \sup_{t\in[0,T]} M_z(t) \geq \kappa_z\zeta_z,\,\left(U_{z-1}\right)^{1+\alpha_0} \leq \zeta_z \right\} \right)\\ &\leq \exp\left\{-\frac{\kappa_z^2\lambda^{4\alpha_0}}{2N^z}\right\} \leq \exp\left\{-\frac{(2N+1)^{2z}\lambda^{2\alpha_0}}{2N^z}\right\} \\ &\leq \exp\left\{ -2^z\lambda^{2\alpha_0} \right\} \leq \exp\left\{ -z\lambda^{2\alpha_0} \right\},\end{aligned}$$ which together with relation implies estimate . Finally, equipped with the above estimate on the tail probability, we are now at a position to prove the $L^p$-estimates for the time-space uniform norm of weak solutions. Taking $z=0$ in relation and applying Hölder inequality, we have for $0\leq \tau \leq T$, $$\begin{aligned} &\sup_{t\in[0,\tau]}\\|u^+(t)\\|^{2}+\int_0^{\tau}\left(\\|u^+(s)\\|^2_{\eta}+\sum_{k=1}^{d_1}\\| L_ku^+(s)\\|^2\right)\,ds \nonumber \\ &\leq C\bigg\{ \sup_{t\in[0,{\tau}]}\int_0^t \!\!\left\langle u^+(s),\,(-D_i\theta^{ik}u^++2\beta^k u^++2h^k)(s)\,dW^k_s\right\rangle \nonumber\\ &\quad + \int_0^{\tau} \left(\\|h(s)1_{\{u>0\}}\\|^2+ 2\left\|\left\langle u^+(s),\,(L_k'g^k+f^+)(s)\right\rangle\right\|\right)\,ds +\\|\underline{u}^+_0 \\|^{2} \bigg\}\nonumber\\ & \leq C\bigg\{ \sup_{t\in[0,{\tau}]}\int_0^t \!\!\left\langle u^+(s),\,(-D_i\theta^{ik}u^++2\beta^k u^++2h^k)(s)\,dW^k_s\right\rangle \nonumber\\ &\quad + \int_0^{\tau} \left(\\|h(s)1_{\{u>0\}}\\|^2+ \\|g(s)1_{\{u>0\}}\\|^2 +\\|f^+(s)1_{\{u>0\}}\\|^2\right)\,ds +\\|\underline{u}^+_0 \\|^{2} \bigg\}\nonumber\\ &+\frac{1}{2} \int_0^{\tau} \left( \\|u^+\\|^2+\sum_{k=1}^{d_1}\\|L_ku^+(s)\\|^2 \right)ds,\quad\text{a.s.,}\nonumber\end{aligned}$$ which implies that $$\begin{aligned} &\sup_{t\in[0,\tau]}\\|u^+(t)\\|^{2}+\int_0^{\tau}\left(\\|u^+(s)\\|^2_{\eta}+\sum_{k=1}^{d_1}\\| L_ku^+(s)\\|^2\right)\,ds \nonumber \\ &\leq C\left\{\sup_{t\in[0,{\tau}]} \tilde{M}_t + \\|(f^+,g,h)1_{\{u>0\}}\\|_{L^2([0,{\tau}]\times{\mathbb{R}}^d)}^2 + \\|\underline{u}^+_0 \\|^2 \right\},\quad\text{a.s.,}\label{eq-U0-l2}\end{aligned}$$ with $$\tilde M_t:=\int_0^t \!\!\left\langle u^+(s),\,(-D_i\theta^{ik}u^++2\beta^k u^++2h^k)(s)\,dW^k_s\right\rangle,\quad t\in[0,T].$$ Observe that for any $t\in[0,T]$ and $q> 0$, $$\begin{aligned} \langle \tilde M\rangle_t^{\frac{q}{2}} &=\left(\sum_{k=1}^{d_1}\int_0^t \left\| \left\langle u^+(s),\,(-D_i\theta^{ik}u^++2\beta^k u^++2h^k)(s)\right\rangle \right\|^2ds\right)^{\frac{q}{2}}\\ &\leq C\left(\int_0^t \left( \\|u^+\\|^4 + \\|u^+\\|^2\\|h1_{\{u>0\}}\\|^2 \right)(s)\,ds\right)^{\frac{q}{2}} \\ &\leq \left({\varepsilon}+C\tau^{\frac{q}{2}}\right) \sup_{s\in[0,t]}\\|u^+(s)\\|^{2q} + C_{{\varepsilon}} \left(\int_0^t\\|h1_{\{u>0\}}\\|^2 ds\right)^{q} .\end{aligned}$$ Take $${\varepsilon}=\frac{1}{4} \quad \text{and}\quad \tau=T \wedge \left(\frac{1}{4C}\right)^{\frac{2}{q}}.$$ By relation and the Burkholder-Davis-Gundy inequality, we have for $q>0$, $$\begin{aligned} &E\sup_{t\in[0,\tau]}\\|u^+(t)\\|^{2q}+E\left[\int_0^{\tau}\left(\\|u^+(s)\\|^2_{\eta}+\sum_{k=1}^{d_1}\\| L_ku^+(s)\\|^2\right)\,ds\right]^q \nonumber \\ &\leq \frac{1}{2} E\sup_{s\in[0,\tau]}\\|u^+(s)\\|^{2q} + CE \left[ \left( \\|(f^+,g,h)1_{\{u>0\}}\\|_{L^2([0,{\tau}]\times{\mathbb{R}}^d)}^2 + \\|\underline{u}^+_0 \\|^2 \right)^{q}\right].\end{aligned}$$ Starting from the interval $[0,\tau]$, within $\left\lceil\frac{T}{\tau} \right\rceil$ steps we arrive at $$\begin{aligned} E\left(U_0\right)^q \leq C E\left[\\|(f^+,g,h)1_{\{u>0\}}\\|_{L^2(Q)}^{2q} + \\|\underline{u}^+_0 \\|^{2q} \right].\end{aligned}$$ Taking $q=\frac{p}{2\theta_0}$ in the above inequality, we have by Proposition \[lem-uplus\], $$\begin{aligned} &E\\|u^+\\|^p_{L^{\infty}(Q)}\\ &=p\int_0^{\infty} {\mathbb{P}}\left( \left\{ \\|u^+\\|_{L^{\infty}(Q)} > \lambda \right\}\right)\lambda^{p-1}\,d\lambda\\ &\leq \lambda_0^p +\int_{\lambda_0}^{\infty} {\mathbb{P}}\left(\left\{ U_0 > \lambda^{2\theta_0} \right\}\right)\lambda^{p-1} \,d\lambda +\int_{\lambda_0}^{\infty} {\mathbb{P}}\left(\left\{ \\|u^+\\|_{L^{\infty}(Q)} > \lambda,\, U_0 \leq \lambda^{2\theta_0}\right\} \right)\lambda^{p-1} \,d\lambda \\ &\leq \lambda_0^p +\frac{1}{2\theta_0} E\left\|U_0\right\|^{\frac{p}{2\theta_0}} +\int_{\lambda_0}^{\infty} 2 \exp\left\{-\lambda^{2\alpha_0 }\right\} \lambda^{p-1} \,d\lambda \\ &<\infty.\end{aligned}$$ Hence, in view of Lemmas \[lem-iteration\] and \[lem-uplus\], we have by scaling $$\begin{aligned} E\\|u^+\\|^p_{L^{\infty}(Q)} \leq C \left(\Lambda^+_{\bar p,\infty } + \Lambda^+_{\frac{p}{\theta_0}} \right)^p,\end{aligned}$$ with the constant $C$ depending on $d,p,n_0, T$ and the quantities related to the coefficients $\sigma,\theta,b,c$ and $\beta$. The estimate on $u^-$ follows in a similar way. We complete the proof. By Theorem \[thm-Lp-inf\], we give the $L^p$ ($p>0$) estimates for the time-space uniform norm of weak solutions for possibly degenerate SPDE in the whole space. It seems to be new, even for the super-parabolic case (that is $n_0=0$ in $({\mathcal{H}})$), as the existing results on such kind of estimates for weak solutions of super-parabolic SPDEs are restricted in bounded domains (see [@D-bounddeness-SPDE-2013; @DenisMatoussi2009; @DenisMatoussiStoica2005]) with $p\in [2,\infty)$. In fact, our method of De Giorgi iteration in this section is applicable to the local maximum principle for weak solutions of SPDEs in either bounded or unbounded domains, by using the techniques of cut-off functions (see [@QiuTangMPBSPDE11] for instance). On the other hand, in Theorem \[thm-Lp-inf\] as well as in assertion (i) of Theorem \[thm-main\], we assume $({\mathcal{A}}1)$ which requires the spacial smoothness of coefficients $\sigma,\theta,b,c$ and $\beta$; in fact, such assumption is made for the sake of simplicity and it can be relaxed in a standard way due to the properties of multipliers in (ii) of Lemma \[lem-pdo\]. We would postpone such generalizations in domains with relaxed assumption $({\mathcal{A}}1)$ to a future work. [10]{} , [Hypoellipticity theorems and conditional laws]{}, Probab. Theory Relat. Fields, 65 (1984), pp. 573–597. , [A note on degenerate stochastic integro-differential equations]{}, arXiv:1406.5649, (2014). , [On the boundedness of solutions of [SPDEs]{}]{}, Stochastic Partial Differential Equations: Analysis and Computations, (2013), pp. 1–19. , [Maximum principle and comparison theorem for quasi-linear stochastic $\textrm{PDE}$¡¯s]{}, Electron. J. Probab., 14 (2009), pp. 500–530. , [$\textrm{L}^p$ estimates for the uniform norm of solutions of quasilinear $\textrm{SPDE}$’s]{}, Probab. Theory Relat. Fields, 133 (2005), pp. 437–463. , [Maximum principle for quasilinear stochastic [PDEs]{} with obstacle]{}, Electron. J. Probab, 19 (2014), pp. 1–32. , [On the solvability of degenerate stochastic partial differential equations in [Sobolev]{} spaces]{}, Stochastic Partial Differential Equations: Analysis and Computations, (2014), pp. 1–32. , [Fractional [Gagliardo–Nirenberg]{} and [Hardy]{} inequalities under [Lorentz]{} norms]{}, J. Math. Anal. Appl., 396 (2012), pp. 569–577. , [The Analysis of Linear Partial Differential Operators [III]{}]{}, vol. 257, Springer, 1983. , [Stochastic [De Giorgi]{} iteration and regularity of stochastic partial differential equation]{}, arXiv:1312.3311, (2013). , [On $\textrm{L}_p$-theory of stochastic partial differential equations]{}, J. Math. Anal., 27 (1996), pp. 313–340. height 2pt depth -1.6pt width 23pt, [*]{}, in Stochastic Partial Differential Equations: Six Perspectives, vol. 64 of Mathematic Surveys and Monographs, AMS, Providence, RI, 1999, pp. 185–242. , [Hörmander’s theorem for stochastic partial differential equations]{}, arXiv:1309.5543, (2013). height 2pt depth -1.6pt width 23pt, [Hypoellipticity for filtering problems of partially observable diffusion processes]{}, Probab. Theory Relat. Fields, (2013), pp. 1–32. height 2pt depth -1.6pt width 23pt, [Hörmander’s theorem for parabolic equations with coefficients measurable in the time variable]{}, SIAM J. Math. Anal., 46 (2014), pp. 854–870. , [On the cauchy problem for linear stochastic partial differential equations]{}, Izvestiya: Mathematics, 11 (1977), pp. 1267–1284. , [Stochastic evolution equations]{}, J. Sov. Math., 16 (1981), pp. 1233–1277. , [On degenerate linear stochastic evolutions equations driven by jump processes]{}, arXiv:1406.4541, (2014). , [Hörmander-type theorem for [Itô]{} processes and related backward [SPDEs]{}]{}, arXiv:1412.5481 \[math.AP\], (2014). , [Maximum principles for backward stochastic partial differential equations]{}, J. Funct. Anal., 262 (2012), pp. 2436–2480. , [On the quasi-linear reflected backward stochastic partial differential equations*]{}, J. Funct. Anal., 267 (2014), pp. 3598–3656.	{ "pile_set_name": "ArXiv" }
Introduction— We consider propagation of a blast wave due to strong explosion in the center of a power-law-density, $\rho \propto r^{-\alpha }$, spherically symmetric atmosphere. This problem is of interest for two reasons. An exploding star sends a blast wave into the stellar, and later into the circumstellar medium. The (circum) stellar medium has a varying but spherically symmetrical density, and power-law density profile would often be a reasonable approximation. Blast waves in a circumstellar medium are interesting astronomical sources (eg. [@Ch] and later papers by the same author). The strong explosion problem is also of methodological interest, being a canonical example of self-similarity in gas dynamics. Self-similar solutions, with shock radius $R\propto t^{\beta }$, of type I and of type II were known in this problem. [Type I:]{} For $\alpha <3$, the blast wave is decelerating, $\beta <1$ [@SeB]. The self-similar solution is of type I, meaning that the index $\beta$ is determined by the energy conservation law – from $R^3\rho (R)(dR/dt)^2 \propto {\rm const}$, one gets $\beta=2/(5-\alpha )$. This is a straightforward generalization of the Sedov-Taylor-von Neumann solution describing the explosion in a uniform atmosphere [@Se; @Ta; @vN; @LL]. [Type II:]{} For $\alpha >3.26$ [^1], the blast wave is accelerating, $\beta >1$ [@WS]. The self-similar solution is of type II, meaning that the index $\beta$ is determined by an eigenvalue problem – it must be chosen so as to remove a singularity of the system of ordinary differential equations describing the self-similar solution. Dimensional analysis is insufficient in this case, and $\beta$ should be determined numerically. [Type III:]{} For $3<\alpha <3.26$, the blast has a constant velocity, $\beta =1$ \[this paper\]. The self-similar solution is of a new type, type III, meaning that the index $\beta$ is determined by the requirement that the non-self-similar part of the blast wave serve as a piston for the self-similar part of the blast wave. Dimensional analysis does not work in this case, but one can calculate $\beta$ without solving an eigenvalue problem, it is just $\beta =1$. We consider only a spherically symmetrical model. In reality, deviations from spherical symmetry should exist [@Go], but are expected to be very small generically (we checked this numerically), should be especially small for $\alpha \approx 3$, and may be entirely absent for $\alpha > 3$. Gravity is neglected. We first calculate the new self-similar solution, and then describe numerical simulations which confirm that our self-similar solution is indeed an asymptotic of the strong explosion. Formulation of the problem At $t=0$ and $r\rightarrow \infty$, the atmosphere is spherically symmetrical, cold, motionless, and self-similar with $\rho \rightarrow r^{-\alpha }$. Also at $t=0$ but at small $r$ there is a non-zero energy density in arbitrary form. The problem is to calculate the resulting flow at $t\rightarrow \infty$. Self-similar solution One assumes that at large $t$, the flow is a self-similar blast wave (meaning a shock wave followed by a flow). In dimensionless units, without loss of generality, we will assume that the blast wave propagates into the density $$\rho_0(r)=r^{-\alpha }.$$ The shock radius is $$R=t^{\beta }.$$ The shock jump conditions give the density, velocity, and pressure at $r=R-0$: $\rho = 4\rho _0(R)$, $v=(3/4)\dot{R}$, $p=(3/4)\rho _0(R)\dot{R}^2$. Then the density, velocity, and pressure of the downstream gas are of the form $$\rho=4t^{-\alpha \beta}\rho (\xi ),$$ $$v=(3/4)\beta t^{\beta -1}v (\xi ),$$ $$p=(3/4)\beta ^2t^{2\beta -2-\alpha \beta}p (\xi ).$$ where $\rho (1)=v(1)=p(1)=1$. When (3)-(5) are substituted into the gas dynamics equations (continuity, momentum, and pressure equations), a system of ordinary differential equations for $\rho (\xi )$, $v (\xi )$, and $p (\xi )$ is obtained. With $'\equiv d/d\xi $, $$(4\xi -3v)\rho '/\rho =3(v'+2v/\xi )-4\alpha,$$ $$(4\xi -3v)v'=p'/\rho+4(1-1/\beta )v,$$ $$(4\xi -3v)p'/p =5(v'+2v/\xi )+4(2-2/\beta -\alpha ).$$ One can introduce new variables, reducing (6)-(8) to a system of two homogeneous equations [@SeB]. We put $$v (\xi )=V(\xi )\xi,~~~~~ p(\xi)=\rho (\xi)(C(\xi )\xi )^2.$$ Here $C$ is proportional to the sound speed, $V(1)=C(1)=1$, and, after re-defining $'\equiv \xi~d/d\xi =d/d\ln \xi $, we get $$V'={F(V,C)\over H(V,C)},$$ $$C'={G(V,C)\over (4-3V)H(V,C)},$$ where $$\begin{aligned} F=-\left( 15C^2-(4-3V)^2+4(1-\beta ^{-1})(4-3V)\right) V \nonumber\\ +4\left( (\alpha-2(1-\beta ^{-1})\right) C^2,\end{aligned}$$ $$H(V,C)=5C^2-(4-3V)^2,$$ $$G(V,C)=C\left( F+(6V-4\beta ^{-1})H\right) .$$ The system (10),(11) is integrated as follows: start at $\ln \xi =0$ with $V=C=1$, and integrate back to $\ln \xi =-\infty $. The integration can terminate at the singular point $H=0$, unless the singularity is removed by having $F=0$ at this point. [^2] [Type I, [@SeB]:]{} For $\alpha < 3$, energy conservation gives $\beta =2/(5-\alpha)$. The blast wave is decelerating, $\beta <1$. If also $\alpha > 2$, the solution terminates at the singular point $V=4/3$, $C=0$ at a finite value of $\ln \xi = \ln \xi _{\rm min}$, and an evacuated region forms at small $\xi$. Near the singular point, that is for small $x\equiv \ln \xi -\ln \xi _{\rm min}$, equations (10), (11) give $$V=4/3-(4/5)(\alpha -1)x, ~~~C\propto x^{1/2}.$$ Then (6) gives $$\rho \propto x^{-\nu }, ~~~\nu ={8\alpha -18\over 3\alpha -3}.$$ For $\alpha <3$, the density slope $\nu <1$. The density singularity is integrable, and the pressure at the singular point $p\propto \rho C^2$ is vanishing. [Type II, [@WS]:]{} For $\alpha > 3.26$, $\beta $ is determined numerically, by removing the singularity at $H=0$. The self-similar solution extends all the way to $\ln \xi =-\infty $. The blast wave is accelerating, $\beta >1$. [Type III, the new regime:]{} For $\alpha > 3$, the type I solution, $\beta =2/(5-\alpha)$ cannot be correct. This solution has infinite pressure at the termination point at finite $\xi$. Infinite pressure is not necessarily unphysical. We are considering just an asymptotic solution at infinite time. For $\alpha >3$ a finite fraction of the mass of the gas is not described by the self-similar solution, because it originates in the non-self-similar part of the unperturbed density (for a discussion see [@WS]). In terms of pure self-similar solution, this gas forms an infinitely heavy piston, which might be able to support the infinite pressure at the end of the self-similar flow. However the problem is that the piston moves at a constant velocity, and therefore it cannot keep up with the accelerated blast wave. We must search among different values of $\beta $. For $\alpha > 3$, because of the constant-velocity piston formed by the non-self-similar gas, the solution cannot be decelerating, that is $\beta \geq 1$. It turns out (by numerical integration of (10), (11)) that for $\alpha < 3.26$ and $\beta \geq 1$ the solution terminates at the singular point $V=4/3$, $C=0$ at a finite value of $\ln \xi = \ln \xi _{\rm min}$. Near the singular point, that is for small $x\equiv \ln \xi -\ln \xi _{\rm min}$, equations (10), (11) give for $\beta > 1$ $$V=4/3-(8/5)(2-\beta ^{-1})x, ~~~C\propto x^{1/2}.$$ Then (6) gives $$\rho \propto x^{-\nu }, ~~~\nu =1+{5(\alpha -3)\over 6(2-\beta ^{-1})}.$$ For $\alpha >3$, the density slope $\nu >1$. The pressure at the singular point $p\propto \rho C^2$ is diverging. But this is impossible for an accelerating solution, as discussed above. We are left with just one possibility, $\beta =1$. In this case, (10), (11) give near the singular point $$V=4/3-4(1-\alpha /5)x,$$ $$C\propto x^{\mu }, ~~~\mu ={\alpha \over 3(5-\alpha )}.$$ Then (6) gives $$\rho \propto x^{-2\mu }.$$ The pressure at the singular point $p\propto \rho C^2$ is finite. Since the blast wave moves at a constant speed, this end pressure can be provided by the piston of non-self-similar gas. Since $2\mu >1$, the mass is diverging, but this is OK, because in terms of an ideal self-similar solution the total mass of the exploding gas is infinite. Part of this mass forms an infinite-mass piston, another part forms an infinite-mass self-similar flow. The new solution exists only in a narrow interval of density slopes, $3<\alpha <3.26$. For $\alpha >3.26$, numerical integration of (10), (11) with $\beta =1$ does not terminate at the point $V=4/3$, $C=0$. The solution crosses the line $H=0$ in a different point, with $F\neq 0$. Therefore, our solution is no longer valid. The right solution is the type II solution of [@WS]. ![The $\alpha =2.85$ density case. Thick lines show the self-similar solution with $\beta = 2/(5-\alpha )=0.930$. Thin lines show the numerical solution. The horizontal axis gives $r/R_{\rm shock}$. Shown are $0.1\rho /\rho_{\rm shock}$, $1+v/v_{\rm max}$,$2+p/p_{\rm max}$. Resolution and run time: 10000 grid, 20 doublings of the grid. See main text for details.](gruzinovf1.ps){width=".4\textwidth"} ![Same as fig. 1, for $\alpha =3.15$, $\beta = 1$. ](gruzinovf2.ps){width=".4\textwidth"} Numerical simulation— There is no theorem guaranteeing that a self-similar solution is asymptotically reached starting from a generic initial condition. One has to assume this, or check numerically. We numerically simulated spherically symmetrical explosions in a power-law atmosphere. A spherically symmetric version of the code [@TP] was used. It was found that the second-order code becomes unstable near the termination part of the $\alpha \approx 3$ blast waves. We therefore had to use the first-order version of the code. It was found that for $\alpha \approx 3$ a very long run is needed to reach the theoretically predicted self-similar solution. This can be explained as follows. The energy conservation gives, approximately, $M(R)\dot{R} ^2=const$, where $M(R)$ is the enclosed mass $$M(R)=\int_1 ^Rdrr^{-\alpha}\propto R^{3-\alpha}-1.$$ Here 1 is the length of the numerical grid, and all the energy of the explosion is put into the first grid point at the initial time. If $\alpha$ is close to 3, one has to wait for a long time, when the shock moves to large $R$, before one of the terms in (22) becomes much larger than the other. Since our new solution exists only in narrow interval of $\alpha$, it is not expected to be very different from the Sedov-Taylor or Waxman-Shvarts solutions. To detect the small difference, we have to aim at a few percent accuracy for $\|\alpha -3\|\sim 0.15$. Putting $R^{\|3-\alpha \|}=30$ gives the greed of $\sim 10^{10}$ points. This is more than we can simulate directly. We used much smaller grids, resolutions of $10^4$ and $10^5$, but then the grid was doubled for up to twenty times, when the blast wave was reaching the end of the grid. We first made sure that this numerical method does reproduce the hollow Sedov-Taylor solutions ($2<\alpha <3$, $\beta =2/(5-\alpha)$). It does, as figure 1 shows. We also checked that the numerical method reproduces the Waxman-Shvarts solution (it does for $\alpha =3.40$ with 20 doublings of the grid). We then checked if the new type III solution ($3<\alpha <3.26$, $\beta =1$) is the true asymptotic of the explosion. It apparently is, as figure 2 shows. However, we have to say that our numerical proof is only tentative. It turns out that increased resolution makes the rear zone of the blast wave (numerically ?) unstable, fig. 3. The density and velocity profiles do not change with increased resolution and still agree with the predicted self-similar profiles. The fluctuating pressure is still centered around the predicted self-similar solution. But the lack of numerical convergence near the end of the blast wave (or a possible indication of a true weak instability?) must somewhat reduce our confidence that our self-similar solution is the true long-time asymptotic of the explosion. Conclusion A new, type III, self-similar solution of the strong explosion problem in the power-law atmosphere is given. This solution describes a constant speed ($\beta =1$) blast wave, and applies in a narrow interval of density slopes, $3<\alpha <3.26$. Numerical simulations confirm (tentatively) that the new solution is indeed a long-time asymptotic of the strong explosion. Acknowledgment I thank Drs. Goodman, Pen, Sari, Waxman for discussions. This work was supported by the David and Lucille Packard Foundation. ![$\alpha =3.15$, analytical solution is not shown, 125000 grid, 18 doublings of the grid.](gruzinovf3.ps){width=".4\textwidth"} R. A. Chevalier, Astrophys. J. [259]{}, L85 (1982) L. I. Sedov, Similarity and Dimensional Methods in Mechanics (New York, Academic, 1959) L. I. Sedov, Prikl. Mat. Mekh. [10]{}, 241 (1946) G. I. Taylor, Proc. R. Soc. London [A201]{}, 159 (1950) J. von Neumann, Los Alamos Sci. Lab. Tech. Series [7]{} (1947) L. D. Landau, E. M. Lifshitz, Fluid Mechanics (Oxford: Pergamon 1959) E. Waxman, D. Shvarts, Phys. Fluids [A5]{}, 1035 (1993) J. Goodman, Astrophys. J. [358]{}, 214 (1990) H. Trac, U.-L. Pen, astro-ph/0210611 [^1]: Here and in the rest of the paper only the astronomically interesting case of adiabatic index $\gamma =5/3$ is considered. Generalization to other values of $\gamma$ is straightforward. [^2]: The singularity $4-3V=0$ with $H\neq 0$ does not occur for the values of $\alpha$ that we consider.	{ "pile_set_name": "ArXiv" }
--- abstract: 'We discuss the role of intrinsic charm (IC) in the nucleon for forward production of $c$-quark (or $\bar c$-antiquark) in proton-proton collisions for low and high energies. The calculations are performed in collinear-factorization approach with on-shell partons, $k_T$-factorization approach with off-shell partons as well as in a hybrid approach using collinear charm distributions and unintegrated (transverse momentum dependent) gluon distributions. For the collinear-factorization approach we use matrix elements for both massless and massive charm quarks/antiquarks. The distributions in rapidity and transverse momentum of charm quark/antiquark are shown for a few different models of IC. Forward charm production is dominated by $gc$-fusion processes. The IC contribution dominates over the standard pQCD (extrinsic) $gg$-fusion mechanism of $c\bar c$-pair production at large rapidities or Feynman-$x_F$. We perform similar calculations within leading-order and next-to-leading order $k_T$-factorization approach. The $k_T$-factorization approach leads to much larger cross sections than the LO collinear approach. At high energies and large rapidities of $c$-quark or $\bar c$-antiquark one tests gluon distributions at extremely small $x$. The IC contribution has important consequences for high-energy neutrino production in the Ice-Cube experiment and can be, to some extent, tested at the LHC by the SHIP and FASER experiments by studies of the $\nu_{\tau}$ neutrino production.' author: - 'Rafa[ł]{} Maciu[ł]{}a' - 'Antoni Szczurek[^1]' title: 'Intrinsic charm in the nucleon and charm production at large rapidities in collinear, hybrid and -factorization approaches' --- Introduction ============ The text-book proton consists of $u u d$ valence quarks. This picture is by far too simplified. In fact there is strong evidence for internal strangeness and somewhat smaller for internal charm content of the nonperturbative proton. Different pictures of nonperturbative $c \bar c$ content were proposed in the past. A first example is relatively old BHPS model [@BHPS1980] which assumes $u u d c \bar c$ 5-parton Fock configurations (see also Refs. [@BHK1982; @VB1996]). Another picture proposed in the literature is a meson cloud model (MCM) [@NNNT1996; @MT1997; @SMT1999; @CDNN2001; @HLM2014], where the $p \to {\bar D}^0 \Lambda_c$ or $D \Sigma_c$ fluctuations of the proton are considered. While in the first model $c(x) = {\bar c}(x)$ in the MCM $c(x) \ne {\bar c}(x)$. The models do not allow to predict precisely the absolute probability for the $c$-quark or $\bar c$-antiquark content of the proton. Experimental data put only loose constraints on the charm content: $$\int_0^1 c(x) \; dx = \int_0^1 {\bar c}(x) \; dx < 0.01 \; . \label{charm_probability}$$ It is rather upper limit but this value depends somewhat on the model of charm content of a proton. In general, for sea-like models the probability can be slightly larger than for the BHPS one. In the sea-like case the charm is concentrated at lower values of $x$. Very recent lattice study of charm quark electromagnetic form factors suggested asymmetry of $c$-quark and $\bar c$-antiquark distributions [@latticeQCD]. Recently there is a renewed interest in the intrinsic charm (IC) which is related to experiments being performed at the LHC [@BBLS2015; @RKAG2016; @LLSB2016; @BBLLMST2018]. The intrinsic charm is often included in global parton analyses of world experimental data [@BKLSSV2015; @Ball:2014uwa; @Hou:2017khm]. The highly energetic neutrino experiments, such as IceCube, could put further constraints on the intrinsic charm [@ERS2008; @LB2016; @GGN2018]. Here, however, the IC contribution may compete with a concept of the subleading fragmentation [@Maciula:2017wov]. Similarly, future LHC high and low energy forward experiments like FASER and SHIP could also be very helpful in this context (see e.g. Ref. [@Bai:2020ukz] and Ref. [@Bai:2018xum], respectively). Also the LHCb experiment in its fixed-target mode could be sensitive to the contributions coming from intrinsic charm in a proton, especially in the case of open charm production [@Aaij:2018ogq], where some problems with a satisfactory theoretical description of the experimental data were reported (see also discussion in Ref. [@Maciula:2020cfy]). In this paper we concentrate therefore on forward production of charm quarks/antiquarks. There were already some studies performed with color glass condensate approach and compared to the dipole approach at forward directions [@GNU2010; @CGGN2017]. In our approach we will use instead collinear, hybrid and $k_T$-factorization approach. The latter two were not studied so far in the context of IC and forward production of charm. Models of intrinsic charm in a nucleon ====================================== In the five-quark Fock component $u u d c \bar c$ heavy quark/antiquark carries rather large fraction of the mother proton. In the BHPS model, after some approximations the probability to find $c$ or $\bar c$ (the same for both) can be expressed via a simple formula: $$\frac{dP}{dx} = c(x) = {\bar c}(x) = A x^2 \left( 6 x (1+x) ln(x) + (1-x)(1+10x+x^2) \right) \; .$$ The normalization constant $A$ depends on integrated probability for $c \bar c$ component and is 6 for 1 % probability. Please note that the quark mass is not explicit in this simplified formula. In the meson cloud models $c$ is in the baryon-like object and $\bar c$ in the meson-like object. Then the probabilistic distribution can be obtained as $$\begin{aligned} \frac{dP_c}{dx} &=& \int_x^1 \frac{dy}{y} f_B(y) f_{c/B}(x/y) \; , \\ \frac{dP_{\bar c}}{dx} &=& \int_x^1 \frac{dy}{y} f_M(y) f_{\bar c/M}(x/y) \; . \label{MCM}\end{aligned}$$ The $f_B$ and $f_M$ functions, the probability to find meson or baryon in proton, can be calculated from corresponding Lagrangians supplemented by a somewhat arbitrary and poorly known vertex form factors and can be found e.g. in Ref. [@HLM2014]. In general, such an approach leads to $c(x) \ne {\bar c}(x)$. In practice both models give rather similar distributions as will be shown in the following, so using one of them as an example is representative and sufficient. These are models of large-$x$ components of IC. In principle, the IC may have also small-$x$ component known under the name of sea-like, however, only simple ad hoc parametrizations were used in the literature. ![A dynamical process leading to sea-like IC. []{data-label="fig:sea-like"}](sea-like.eps){width="100.00000%"} There is another category of processes leading to sea-like IC (see Fig. \[fig:sea-like\] where an example of corresponding dynamical processes is shown). Using intrinsic glue in the nucleon (see e.g. Ref. [@EI1998]) one can generate intrinsic charm sea. The intrinsic gluon distribution fulfil by construction the relation: $$\int_0^1 \left( x u_v(x) + x d_v(x) + x g(x) \right) dx = 1 \; .$$ For massless charm the intrinsic charm can be calculated as the convolution with initial (intrinsic) glue $$c(x) = {\bar c}(x) = \alpha_s(4 m_c^2) / (2 \pi) \int_x^1 dy \left( \frac{1}{y} \right) P_{q g}\left( \frac{x}{y} \right) g(y) \; ,$$ where $g$ is the intrinsic gluon distribution. With the model from Ref. [@EI1998] the $c$ and $\bar c$ distributions are integrable, concentrated at $x \sim$ 0.1-0.2 and corresponding probability is 2.7 %. It should be less for massive quarks/antiquarks. In Fig. \[fig:xcharm\_IC\] we show $x$-distribution of the IC for the BHPS and for the sea-like model described above. ![Charm quark/antiquark distribution for the two different models of IC. The solid line represents the BHPS model while the dashed line is for sea-like glue as obtained in a way described above. In this calculation BHPS model with 1% probability was used for illustration. []{data-label="fig:xcharm_IC"}](xc_BHPS_vs_sea-like.eps){width="100.00000%"} In the GRV approach [@Gluck:1994uf] the charm contribution is calculated fully radiatively as the convolution of gluon distribution with appropriate mass-dependent splitting function: $$x c(x,Q^2) = \frac{\alpha_s(\mu'^2)}{2 \pi} \int_{a x}^1 dy \left( \frac{x}{y} \right) C_{g,2}^c \left(\frac{x}{y}, \ \frac{m_c^2}{Q^2} \right) g(y,\mu'^2) \; . \label{GRV_charm}$$ The explicit formula for $C_{g,2}^c$ and $a$, including mass of quarks/antiquarks, can be found in Ref. [@Gluck:1994uf]. In the following calculations we will use more modern gluon distributions. In this paper we concentrate on large-$x$ component and completely ignore the sea-like component(s). Charm can also be generated by evolution equations via $g \to c \bar c$ transition (splitting). Often it was included in the evolution as a massless parton with zero as initial condition at the starting scale $\mu^2 \sim m_c^2$. In a dedicated fits, the intrinsic charm distribution is used as initial condition for DGLAP evolved charm distributions (see e.g. Ref. [@PLT2007]). In the right panel Fig. \[fig:xc\] we show charm distribution in a proton without (dashed line) and with (solid line) the IC distribution taken as initial condition of the evolution. Cross section for associated charm production ============================================= The collinear approach ---------------------- In the present study we discuss production of the final states with one charm quark or charm antiquark. In the collinear approach [@Collins:1989gx] the final state charm must be associated with at least one additional gluon or (light) quark. Typical leading-order mechanisms for charm production initiated by charm quark in a initial state are shown in Fig. \[fig1\]. The diagrams correspond to the $gc \to gc$ (or $g \bar c \to g\bar c$) subprocesses that are expected to be dominant at high energies, however, the $q c \to q c$ and $\bar q c \to \bar q c$ (or $q \bar c \to q \bar c$ and $\bar q \bar c \to \bar q \bar c$) mechanisms with $q = u,d,s$ are also possible and will be taken into account in the following numerical calculations. ![ Typical leading-order (2 $\to$ 2) mechanisms of production of $c$ quarks or $\bar c$ antiquarks in the collinear parton model. []{data-label="fig1"}](2to2-IC-a.eps){width="100.00000%"} ![ Typical leading-order (2 $\to$ 2) mechanisms of production of $c$ quarks or $\bar c$ antiquarks in the collinear parton model. []{data-label="fig1"}](2to2-IC-b.eps){width="100.00000%"} ![ Typical leading-order (2 $\to$ 2) mechanisms of production of $c$ quarks or $\bar c$ antiquarks in the collinear parton model. []{data-label="fig1"}](2to2-IC-c.eps){width="100.00000%"} In the collinear approach the differential cross section for forward charm production within the $gc \to gc$ mechanism[^2] can be calculated as $$\begin{aligned} \frac{d \sigma}{d y_1 d y_2 d^2 p_t} = \frac{1}{16 \pi {\hat s}^2} \overline{\| {\cal M}_{g c \to g c} \|^2} x_{1} g(x_1, \mu^2) x_{2} c(x_2, \mu^2) \; ,\end{aligned}$$ where ${\cal M}_{g c \to g c}$ is the on-shell matrix element for $gc \to gc$ subprocesses and $g(x_1, \mu^2)$ and $c(x_2, \mu^2)$ are the collinear gluon and charm quark PDFs evaluated at longitudinal momentum fractions $x$ and factorization scale $\mu^{2}$. Including the mass of charm quark the on-shell matrix element takes the following form: $$\begin{aligned} \overline{\|{\cal M}_{gc \to gc}\|^2} &=& g_s^4 \left[ \left( - m_c^4 ( 3\hat{s}^2 + 14 \hat{s}\hat{u} + 3\hat{u}^2 ) + m_c^2 ( \hat{s}^3 + 7 \hat{s}^2\hat{u} + 7 \hat{s} \hat{u}^2 + \hat{u}^3) \right. \right. \nonumber \\ && \left. \left. + 6m_c^8-\hat{s}\hat{u} ( \hat{s}^2+\hat{u}^2 ) \right) \left( -18m_{c}^2 (\hat{s}+\hat{u}) +18m_c^4+9\hat{s}^2+9\hat{u}^2-\hat{t}^2 \right) \right] \nonumber \\ && / \left( 18\hat{t}^2 ( \hat{u}-m_c^2)^2 ( \hat{s}-m_c^2) \right)^2 ,\end{aligned}$$ where $g_s^2 = 4 \pi \alpha_{s}(\mu)$. In the massless limit $m_c \to 0$ one recovers the known textbook formula: $$\overline{\|{\cal M}_{gc \to gc}\|^2} = g_s^4 \left( -\frac{4}{9} \left( \frac{{\hat u}^2 + {\hat s}^2}{{\hat u}{\hat s}} \right) + \left( \frac{{\hat u}^2+{\hat s}^2}{{\hat t}^2} \right) \right) \; .$$ A role of the charm quark mass in the matrix element will be discussed when presenting numerical results. ![ Charm quark distributions in a proton as a function of longitudinal momentum fraction $x$. Here different models for initial intrinsic charm quark distributions are shown (left panel) and a comparison between charm quark distributions obtained with and without concept of intrinsic charm in the proton. []{data-label="fig:xc"}](fig1a-IC.eps){width="100.00000%"} ![ Charm quark distributions in a proton as a function of longitudinal momentum fraction $x$. Here different models for initial intrinsic charm quark distributions are shown (left panel) and a comparison between charm quark distributions obtained with and without concept of intrinsic charm in the proton. []{data-label="fig:xc"}](fig1b-IC.eps){width="100.00000%"} In the numerical calculations below the intrinsic charm PDFs are taken at the initial scale $m_{c} = 1.3$ GeV, so the perturbative charm contribution is intentionally not taken into account. We apply four different grids of the intrinsic charm distributions from the CT14nnloIC PDF [@Hou:2017khm] that correspond to the BHPS 1% and BHPS 3.5% as well as the sea-like LS (low-strength) and sea-like HS (high-strength) models for initial intrinsic charm distribution. The distributions are compared with each other in the left panel of Fig. \[fig:xc\]. In the right panel we present in addition the difference between the CT14nnloIC charm PDF obtained with and without intrinsic-charm concept. On the other hand the collinear gluon PDFs $g(x, \mu^2)$ are taken at the running factorization scale related to the averaged transverse momentum of the outgoing particles, i.e. $\mu = \sqrt{\frac{p_{t1}^2 + p_{t2}^2}{2} + m_c^2}$. The charm quark mass $m_{c} = 1.3$ GeV plays here a role of the minimal scale and ensures that we are not going beyond the fitted PDF grids where unconstrained extrapolation procedures are applied. We keep the charm quark mass here even when the massless matrix element and/or kinematics are used. As will be shown later, the numerical results strongly depend on how the longitudinal momentum fractions $x_1$ and $x_2$ (arguments of parton distributions) are calculated. In the massive scheme of the calculations the quantities are defined as follows: $$\begin{aligned} x_1 &=& \frac{p_{t1}}{\sqrt{s}} \exp(+y_1) + \frac{m_{t2}}{\sqrt{s}} \exp(+y_2) \; , \nonumber \\ x_2 &=& \frac{p_{t1}}{\sqrt{s}} \exp(-y_1) + \frac{m_{t2}}{\sqrt{s}} \exp(-y_2) \; . \end{aligned}$$ In this equations $p_{t1}$ is transverse momentum of the outgoing gluon (or light quark/antiquark) and the $m_{t2}$ is $c$ quark ($\bar c$ antiquark) transverse mass defined as $m_{t} = \sqrt{p_t^2+m_c^2}$. As will be discussed further it is crucial to include in kinematics the mass of the final charm, while the initial charm can be considered massless. In the following numerical studies all the calculations in the massless limit with massless matrix elements will be done within the kinematics corrected in the above manner. The effect of the correction will be also explicitly shown. Considering forward production of charm at the LHC energies one is exploring asymmetric kinematical regions where $x_1$ is very small (down to $10^{-5}$) and $x_2$ is rather large (about $10^{-1}$). Thus in this reaction small-$x$ gluon PDF and intrinsic large-$x$ charm content of the proton are probed simultaneously. As it is shown in Fig. \[fig:xcxg\] both distributions are not well constrained by the global experimental data. In the left panel we show the central fits for intrinsic charm distribution from the CT14nnloIC and the NNPDF30nloIC PDF sets [@Ball:2016neh] together with $1\sigma$ standard deviation. In the right panel we compare gluon PDF fits from different collaborations, including MMHT2014nlo [@Harland-Lang:2014zoa], JR14NLO08FF [@Jimenez-Delgado:2014twa] and CT14lo/nnlo sets. Clearly the current level of knowledge of both distributions is rather limited and the large uncertainties prevent definite conclusions. In principle, a study of far-forward production of charm may improve the situation by exploring unconstrained areas. ![ The intrinsic charm (left panel) and gluon (right panel) distributions in a proton as a function of longitudinal momentum fraction $x$. Here different sets of collinear PDFs are shown including uncertainties. []{data-label="fig:xcxg"}](fig2a-IC.eps){width="100.00000%"} ![ The intrinsic charm (left panel) and gluon (right panel) distributions in a proton as a function of longitudinal momentum fraction $x$. Here different sets of collinear PDFs are shown including uncertainties. []{data-label="fig:xcxg"}](fig2b-glue.eps){width="100.00000%"} In the present study we go beyond the leading-order mechanisms and include also higher-order processes that are expected to play important role. We take into account all $2\to 3$ and $2\to 4$ processes at tree-level that lead to a production of charm quark or antiquark and are driven by the $gc$ and $qc$ (or $\bar q c$) initial state interactions. Examples of the diagrams corresponding to the processes are shown in Fig. \[figHO\]. The relevant cross sections are calculated with the help of the <span style="font-variant:small-caps;">KaTie</span> Monte Carlo generator [@vanHameren:2016kkz]. ![ Examples of the $2\to3$ (left panel) and the $2\to4$ (right panel) mechanisms of production of $c$ quarks or $\bar c$ antiquarks in the collinear parton model. []{data-label="figHO"}](2to3-IC.eps){width="100.00000%"} ![ Examples of the $2\to3$ (left panel) and the $2\to4$ (right panel) mechanisms of production of $c$ quarks or $\bar c$ antiquarks in the collinear parton model. []{data-label="figHO"}](2to4-IC.eps){width="100.00000%"} Having massless partons (minijets) in the final states considered in the present work it is necessary to regularize the cross section that has a singularity in the $p_{t} \to 0$ limit. We follow here the known prescription adopted in <span style="font-variant:small-caps;">Pythia</span> where a special suppression factor is introduced at the cross section level [@Sjostrand:2014zea]: $$F(p_t) = \frac{p_t^2}{ p_{t0}^2 + p_t^2 } \; \label{Phytia_formfactor}$$ for each of the outgoing massless partons with transverse momentum $p_t$, where $p_{t0}$ is a free parameter of the form factor. The hybrid model ---------------- Within the asymmetric kinematic situation $x_1 \ll x_2$ described above the cross section for the processes under consideration can be also expressed in the so-called hybrid factorization model motivated by the works in Refs. [@Deak:2009xt; @Kutak:2012rf]. In this framework the small-$x$ gluon is taken to be off mass shell and the differential cross section e.g. for $pp \to g c X$ via $g^* c \to g c$ mechanism reads: $$\begin{aligned} d \sigma_{pp \to gc X} = \int d^ 2 k_{t} \int \frac{dx_1}{x_1} \int dx_2 \; {\cal F}_{g^{}}(x_1, k_{t}^{2}, \mu^2) \; c(x_2, \mu^2) \; d\hat{\sigma}_{g^{}c \to gc} \; ,\end{aligned}$$ where ${\cal F}_{g^{}}(x_1, k_{t}^{2}, \mu^2)$ is the unintegrated gluon distribution in one proton and $c(x_2, \mu^2)$ a collinear PDF in the second one. The $d\hat{\sigma}_{g^{}c \to gc}$ is the hard partonic cross section obtained from a gauge invariant tree-level off-shell amplitude. In the present paper we shall not discuss the validity of the hybrid model on the theoretical level and concentrate only on its phenomenological application in forward production. A derivation of the hybrid factorization from the dilute limit of the Color Glass Condensate approach can be found in Ref. [@Kotko:2015ura]. The gluon uPDF depends on gluon longitudinal momentum fraction $x$, transverse momentum squared $k_t^2$ of the gluons entering the hard process, and in general also on a (factorization) scale of the hard process $\mu^2$. In the numerical calculations we take different models of unintegrated parton densities from the literature: the JH-2013-set2 [@Hautmann:2013tba] model obtained from the CCFM evolution equations, the Kutak-Sapeta (KS) [@Kutak:2014wga] model being a solution of linear and non-linear BK evolution, the DGLAP-based PB-NLO-set1 [@Martinez:2018jxt] model from the parton-branching (PB) method and the Kimber-Martin-Ryskin (KMR) prescription [@Watt:2003mx]. All of the models, except the PB-NLO-set1, are constructed in the way that allows for resummation of extra hard emissions from the uPDFs. It means that in the hybrid model already at leading-order some part of radiative higher-order corrections can be effectively included via uPDFs. However, it is true only for those uPDF models in which extra emissions of soft and even hard partons are encoded, including $k_{t}^{2} > \mu^{2}$ configurations. Then, when calculating the charm production cross section via e.g. the $g^* c \to g c$ mechanism one could expect to effectively include contributions related to an additional extra partonic emission (i.e. $g^* c \to g g c$) which in some sense plays a role of the initial state parton shower. In Fig. \[fig-uPDFs\] we plot the gluon transverse momentum dependence of the different gluon uPDFs from the literature. At the small $x$-values and low scales the differences between the model are quite significant. ![ The ingoing gluon transverse momentum distributions from the different models of unintegrated gluon densities in a proton. []{data-label="fig-uPDFs"}](uPDFs_kT.eps){width="100.00000%"} There are ongoing intensive works on construction of the full NLO Monte Carlo generator for off-shell initial state partons that are expected to be finished in near future [@private-Hameren]. This framework seems to be necessary in phenomenological studies that are based on the PB uPDFs [@Maciula:2019izq]. The extra hard emissions from the DGLAP-based uPDFs are usually strongly suppressed which leaves a room for higher-order terms. Therefore, in this case one needs to include usual leading order subprocesses properly matched with a number of additional higher-order radiative corrections at the level of hard matrix elements. In the moment, it can be done only at tree-level. In consequence, the numerical calculations with the PB-NLO-set1 uPDFs, are done including in addition all $2\to 3$ and $2\to 4$ channels of partonic subprocesses that lead to a production of charm quark or antiquark and are driven by the $g^c$ and $q^c$ (or $\bar q^* c$) initial state interactions, similarly as in the collinear case. Here we follow a dedicated matching procedure to avoid double-counting as introduced in Ref. [@Maciula:2019izq], and further used in Refs. [@Lipatov:2019izq; @Maciula:2020cfy]. The $\bm{k_{T}}$-factorization ------------------------------ Another possible theoretical approach to perform the calculations for the processes considered here is the $k_{T}$-factorization [@kTfactorization]. This framework extends the hybrid model formalism and includes in addition effects related to off-shellness of the initial state charm quark. In principle, it allows to study intrinsic charm contribution to charm production via mechanisms where both incident partons are off mass shell. A topology of possible diagrams present in the $k_{T}$-factorization in the case of intrinsic charm studies is not the same as in the collinear case. Here one can follow two different ways of calculation and consider: - $g^* c^* \to gc$ (and/or $q^* c^* \to qc$, $\bar{q}^* c^* \to \bar qc$ ) mechanism, - $g^* c^* \to c$ mechanism. The second one is not present in other approaches and can be treated as leading-order. The first mechanism directly corresponds to the scheme of the calculations applied in the hybrid model and can be classified as higher-order. However, their mutual coincidence is not clear and strongly depends on the model of unintegrated PDFs used in the numerical calculations. ### The $2 \to 2$ partonic mechanism The $k_{T}$-factorization cross section for the $pp \to gc X$ reaction driven by the typical $2\to2$ mechanisms, e.g. like the $g^* c^* \to gc$, can be expressed as follows: $$\begin{aligned} \label{LO_kt-factorization} \frac{d \sigma(p p \to g c \, X)}{d y_1 d y_2 d^2p_{1,t} d^2p_{2,t}} &=& \int \frac{d^2 k_{1,t}}{\pi} \frac{d^2 k_{2,t}}{\pi} \frac{1}{16 \pi^2 (x_1 x_2 s)^2} \; \overline{ \| {\cal M}^{\mathrm{off-shell}}_{g^* c^* \to g c} \|^2} \\ && \times \; \delta^{2} \left( \vec{k}_{1,t} + \vec{k}_{2,t} - \vec{p}_{1,t} - \vec{p}_{2,t} \right) \; {\cal F}_g(x_1,k_{1,t}^2,\mu^2) \; {\cal F}_c(x_2,k_{2,t}^2,\mu^2) \; \nonumber , \end{aligned}$$ where ${\cal F}_g(x_1,k_{1,t}^2,\mu^2)$ and ${\cal F}_c(x_2,k_{2,t}^2,\mu^2)$ are the gluon and intrinsic charm quark uPDFs, respectively, for both colliding hadrons and ${\cal M}^{\mathrm{off-shell}}_{g^* c^* \to g c}$ is the off-shell matrix element for the hard subprocess. Here the Feynmann diagrams are the same as shown in Fig. \[fig1\]. The extra integration is over transverse momenta of the initial partons. Here, one keeps exact kinematics from the very beginning and additional hard dynamics coming from transverse momenta of incident partons. Explicit treatment of the transverse momenta makes the approach very efficient in studies of correlation observables. Considering forward production of charm one should not expect that the initial state (intrinsic) charm quark could have large transverse momenta. Rather small deviations from the collinear limit are more physically motivated here. Therefore, for the unintegrated charm distribution ${\cal F}_c(x,k_{t}^2,\mu^2)$ we will assume Gaussian distributions with rather small smearing parameter $\sigma_{0}$. The unintegrated $c$ ($\bar c$) distributions are constructed as: $${\cal F}_c(x,k_t^2) = \pi \; G(k_t^2) \cdot x c ( x,\mu^2 ), \label{UPDF_c}$$ where $$G(k_t^2) = \frac{1}{2 \pi \sigma_0^2} \exp\left( \frac{-k_t^2}{2 \sigma_0^2} \right)$$ is a standard two-dimensional Gaussian distribution and $\sigma_0$ is in principle a free parameter which governs the nonperturbative effects in the proton wave function. The factor $\pi$ is because of our normalization of unintegrated parton distributions: $$\int d k_t^2 {\cal F}_c(x,k_t^2) = x c(x) \; . \label{normalization_of_UPDFs}$$ The hard off-shell matrix element ${\cal M}^{\mathrm{off-shell}}_{g^* c^* \to g c}$ is known only in the massless limit. Within this limit the relevant calculations can be done in the <span style="font-variant:small-caps;">KaTie</span> Monte Carlo code, where the matrix element is computed numerically. Its analytic form can be obtained according to Parton-Reggeization-Approach (PRA) and was published in Ref. [@Nefedov:2013ywa]. For the higher-order tree-level diagrams with off-shell initial state partons and with extra partonic legs in the final state one can also use <span style="font-variant:small-caps;">KaTie</span>, which is very efficient in this type of calculations and which was very recently equipped with tools that allow in addition for generation of analytic form of matrix-elements for a given hard multileg processes [@vanHameren:2016kkz]. ### The $2\to 1$ partonic mechanism In the $k_T$-factorization framework the charm quark/antiquark can be created at one-order higher approach. A relevant formalism was used previously for production of forward pions in Ref. [@CS2006]. In Fig. \[fig:kt-factorization\_diagrams\] we show basic graphs for charm quark production within the $2 \to 1$ mechanisms. ![Two leading-order diagrams for charm quark (antiquark) production relevant for $k_t$-factorization approach. The extra explicit gluonic emissions suggest the use of unintegrated gluon distributions. []{data-label="fig:kt-factorization_diagrams"}](2to1-IC-a.eps){width="100.00000%"} ![Two leading-order diagrams for charm quark (antiquark) production relevant for $k_t$-factorization approach. The extra explicit gluonic emissions suggest the use of unintegrated gluon distributions. []{data-label="fig:kt-factorization_diagrams"}](2to1-IC-b.eps){width="100.00000%"} The emitted charm-quark (or antiquark) momentum-space distribution can be written as: $$\begin{aligned} \frac{d \sigma(p p \to c \, X)}{d y d^2 p_t}&& = \frac{16 N_c}{N_c^2 - 1} \cdot \frac{4}{9} \cdot \frac{1}{m_t^2} \times \; \nonumber \\ &&\int \alpha_s(\Omega^2) f_g(x_1,k_{1,t}^2,\mu^2) f_c(x_2,k_{2,t},\mu^2) \delta\left( \vec{k}_{1,t} + \vec{k}_{2,t} - \vec{p}_t \right) d^2 k_{1,t} d^2 k_{2,t} \; . \nonumber \\ \label{kt_factorization_at_LO}\end{aligned}$$ In the formula above $f$’s are unintegrated gluon or charm quark/antiquark distributions. For unintegrated gluon distributions we will take the ones used recently in the literature in the context of $\eta_c$ or $\chi_c$ production [@BPSS2019; @BPSS2020] where the kinematics is similar. For $\Omega^2$ we can take $\Omega^2 = \min(m_t,k_{1t}^2,k_{2t}^2)$ or just $\Omega^2 = m_t^2$. The longitudinal momentum fractions are calculated as $$\begin{aligned} x_1 &=& \frac{m_t}{\sqrt{s}} \exp(+y) \; , \nonumber \\ x_2 &=& \frac{m_t}{\sqrt{s}} \exp(-y) \; .\end{aligned}$$ Results ======= We divide the section with numerical results to four subsections. First three of them are devoted to numerical calculations obtained with the collinear-, hybrid- and the $k_T$-factorization approach, respectively. The last subsection contains explicit predictions for impact of intrinsic charm mechanism on forward production of charm in different experiments, including low energy LHC experiments like fixed-target LHCb and SHIP, as well as high energy FCC and LHC experiments, like proposed recently LHC-FASER. The collinear approach ---------------------- We start presentation of numerical predictions with the results for $pp \to gc X$ reaction driven by the $gc \to gc$ leading-order mechanism calculated in the collinear framework within massive matrix element and kinematics for the energy $\sqrt{s} = 7$ TeV. Here we take the gluon and the intrinsic charm distributions as encoded in the CT14nnloIC collinear PDFs. The three different lines in Fig. \[fig1\] correspond to a different choice of the $p_{t0}$ parameter used for the regularization of the cross section. We see that the predictions for charm quark transverse momentum (left panel) and rapidity (right panel) distributions are very sensitive to the choice of this parameter, especially, at small charm quark transverse momenta, which also affects the rapidity spectrum. In the numerical studies below $p_{t0}=1.0$ GeV will be taken as a default choice which leads to a central value of the uncertainty related to the choice of the parameter. ![ The charm quark transverse momentum (left) and rapidity (right) differential cross sections for $pp$-scattering at $\sqrt{s}=7$ TeV. The results correspond to the $g c \rightarrow g c$ mechanism calculated within the intrinsic charm concept in the collinear-approach with matrix element and kinematics for massive charm quark. Here three different values of the regularization parameter $p_{T0}$ are used. Details are specified in the figure. []{data-label="fig1"}](dsig_dpt_coll_pT0.eps){width="100.00000%"} ![ The charm quark transverse momentum (left) and rapidity (right) differential cross sections for $pp$-scattering at $\sqrt{s}=7$ TeV. The results correspond to the $g c \rightarrow g c$ mechanism calculated within the intrinsic charm concept in the collinear-approach with matrix element and kinematics for massive charm quark. Here three different values of the regularization parameter $p_{T0}$ are used. Details are specified in the figure. []{data-label="fig1"}](dsig_dy_coll_pT0.eps){width="100.00000%"} In Fig. \[fig2\] we present again collinear results for the leading-order $gc \to gc$ mechanism but here we applied four different sets of the intrinsic charm distribution in a proton at initial scale $\mu = 1.3$ GeV as incorporated in the CT14nnloIC PDFs. Again, we show the differential cross sections as a function of the charm quark transverse momentum (left panel) and rapidity (right panel). The solid, long-dashed, dotted and dash-dotted lines correspond to the BHPS 1%, BHPS 3.5%, sea-like LS and sea-like HS models, respectively. The sea-like models lead to a larger cross section than in the case of the BHPS model in the midrapidty region. On the other hand, a larger cross section in the forward direction is obtained within the BHPS models. Clearly, large uncertainties due to the intrinsic charm input are found. ![ The same as in Fig. \[fig1\] but here results obtained with the four different scenarios for intrinsic charm content in a proton are shown. Details are specified in the figure. []{data-label="fig2"}](dsig_dpt_coll_pT01_diffIC.eps){width="100.00000%"} ![ The same as in Fig. \[fig1\] but here results obtained with the four different scenarios for intrinsic charm content in a proton are shown. Details are specified in the figure. []{data-label="fig2"}](dsig_dy_coll_pT01_diffIC.eps){width="100.00000%"} The intrinsic charm component in the proton is not the only source of uncertainties related to the collinear PDFs. As it is shown in Fig. \[fig3\] the gluon PDF also leads to a significant uncertainties of the predictions. Here we show a comparison of the predictions obtained with the default CT14nnloIC (solid lines), the JR14NLO08VF (dotted lines) and the MMHT2014nlo (dashed lines) PDF sets. The gluon PDFs provided by different groups are probed here at small-$x$ and relatively small scales and lead to a quite different results, especially, at small transverse momenta of charm quark. ![ The same as in Fig. \[fig1\] but here results obtained with the three different collinear gluon PDFs are shown. Details are specified in the figure. []{data-label="fig3"}](dsig_dpt_coll_pT0_gluon.eps){width="100.00000%"} ![ The same as in Fig. \[fig1\] but here results obtained with the three different collinear gluon PDFs are shown. Details are specified in the figure. []{data-label="fig3"}](dsig_dy_coll_pT0_gluon.eps){width="100.00000%"} Now we wish to compare three different schemes for the collinear calculations of the $pp \to gc X$ reaction via the $gc \to gc$ leading-order partonic subprocess. In Fig. \[fig4\] we present theoretical distributions obtained within the matrix element with massive quarks (called massive ME for brevity) and kinematics including quark masses (solid lines, our default choice), within the massless matrix element and massless kinematics (dotted histograms), as well as within the massless matrix element and kinematics corrected for the charm quark mass (solid histograms). In each of the cases, we kept the same choice of the renormalization scale $\mu_{R}^{2} = p_{t0}^{2}+p_{t}^{2}+m_{c}^{2}$ and the factorization scale $\mu_{F}^{2} = p_{t}^{2}+m_{c}^{2}$. The charm quark transverse momentum distributions (left panel) are almost identical and some very small (almost invisible) discrepancies appear only at extremely small transverse momenta. The rapidity distributions (right panel) are found to be very sensitive to the charm quark mass effects. Neglecting the charm quark mass in the kinematics leads to a shift of its rapidity distribution to a far forward direction. Correction of the kinematics by inclusion of the outgoing particles mass in the calculation of $x$-values seems to approximately restore the full massive calculations. This step seems to be necessary in the case of massless calculations, otherwise shapes of the predicted rapidity distributions may not be correct. ![ The same as in Fig. \[fig1\] but here results of three different schemes of the collinear caclulations are compared. The solid lines correspond to the calculations with massive matrix element and kinematics, the dotted histograms show results for the calculations with massless matrix elements and kinematics, and the solid histograms represent calculations with massless matrix element and kinematics corrected for the charm quark mass. []{data-label="fig4"}](dsig_dpT_coll_kinematics.eps){width="100.00000%"} ![ The same as in Fig. \[fig1\] but here results of three different schemes of the collinear caclulations are compared. The solid lines correspond to the calculations with massive matrix element and kinematics, the dotted histograms show results for the calculations with massless matrix elements and kinematics, and the solid histograms represent calculations with massless matrix element and kinematics corrected for the charm quark mass. []{data-label="fig4"}](dsig_dy_coll_kinematics.eps){width="100.00000%"} Having discussed the dominant leading-order mechanism we wish to move beyond and consider importance of higher-order corrections for the charm quark forward production mechanisms with intrinsic charm in the initial state. In Fig. \[fig5\] we compare our collinear predictions for the leading-order $2\to2$ mechanisms, both $gc \to gc$ (dotted histograms) and $qc\to qc$ (short-dashed histograms) shown separately, and for the higher-order $2\to3$ (long-dashed histograms) and $2\to 4$ (dash-dotted histograms) mechanisms calculated at tree-level. A sum of the four different components denoted as $2\to 2+3+4$ is also shown but it does not follow any merging procedure here[^3]. For the higher-order contributions the partonic subprocesses with $gc$ and $qc$ initial states are added together. We report a huge contribution to the cross section coming from the higher-order mechanisms (more than order of magnitude). It clearly shows that the leading-order mechanisms are not enough in order to get reasonable predictions for the impact of intrinsic charm concept on forward charm quark production. Full NLO and even NNLO frameworks are required for precise studies of the subject within the collinear parton model. The situation in the case for other approaches, like the hybrid- and the $k_{T}$-factorization is quite different than in the collinear case what will be discussed in next two subsections. ![ The same as in Fig. \[fig1\] but here results of the $2 \to 2$ ($gc$ and $qc$ initial states), $2\to 3$ ($gc + qc$ initial states) and $2\to4$ ($gc + qc$ initial states) mechanisms are shown separately. The calculations are done with massless matrix element and kinematics corrected for the charm quark mass. Details are specified in the figure. []{data-label="fig5"}](dsig_dpT_coll_mechanisms.eps){width="100.00000%"} ![ The same as in Fig. \[fig1\] but here results of the $2 \to 2$ ($gc$ and $qc$ initial states), $2\to 3$ ($gc + qc$ initial states) and $2\to4$ ($gc + qc$ initial states) mechanisms are shown separately. The calculations are done with massless matrix element and kinematics corrected for the charm quark mass. Details are specified in the figure. []{data-label="fig5"}](dsig_dy_coll_mechanisms.eps){width="100.00000%"} The hybrid model ---------------- Now we wish to start presentation of our numerical results obtained in the hybrid model. Here the incident small-$x$ parton is assumed to be off-mass shell in contrast to the large-$x$ intrinsic charm which is kept on-shell. In Fig. \[fig6\] we show theoretical predictions for charm quark transverse momentum (left panel) and rapidity (right panel) distributions for forward charm production within the leading-order $g^c \to gc$ and the $q^c \to qc$ mechanisms. Here the KMR-CT14lo gluon and light quark/antiquark uPDFs are used. We observe that here much larger cross sections are obtained than in the analogous calculations done in the collinear framework (see two lowest histograms in Fig. \[fig5\]). Especially, in the hybrid model the gluonic component is much bigger than its collinear counterpart. Significant effects related to the off-shellness of the incident gluons are found. Considering far forward rapidities of the produced charm quark the transverse momenta (virtualities) of the incident small-$x$ gluons start to play a very important role and lead to a sizeable enhancement of the predicted cross section with respect to the leading-order collinear calculations. ![ The charm quark transverse momentum (left) and rapidity (right) differential cross sections for $pp$-scattering at $\sqrt{s}=7$ TeV. The results correspond to the $g^* c \rightarrow g c$ and $q^* c \rightarrow q c$ mechanisms calculated within the intrinsic charm concept in the hybrid model with off-shell initial state gluon and/or off-shell light-quark. Here the KMR-CT14lo unintegrated parton densities were used. []{data-label="fig6"}](dsig_dpT_hyb_kmr.eps){width="100.00000%"} ![ The charm quark transverse momentum (left) and rapidity (right) differential cross sections for $pp$-scattering at $\sqrt{s}=7$ TeV. The results correspond to the $g^* c \rightarrow g c$ and $q^* c \rightarrow q c$ mechanisms calculated within the intrinsic charm concept in the hybrid model with off-shell initial state gluon and/or off-shell light-quark. Here the KMR-CT14lo unintegrated parton densities were used. []{data-label="fig6"}](dsig_dy_hyb_kmr.eps){width="100.00000%"} Since in the hybrid model the leading-order quark component $q^* c \rightarrow q c$ is found to be negligible one can safely concentrate on the gluonic $g^* c \rightarrow g c$ channel only. In Fig. \[fig7\] we show the relevant predictions for different unintegrated gluon densities from the literature. We compare results obtained with the KMR-CT14lo (solid histograms), the CCFM JH-2013-set2 (dashed histograms) as well as the KS-linear (dotted histograms) and KS-nonlinear (dash-dotted histograms) gluon uPDFs. Different models lead to quite different results, however, they seem to be consistent with each other up to a factor of 5. Main differences appear at larger quark transverse momenta. At small transverse momenta predictions within the KMR-CT14lo, the JH-2013-set2 and the KS-linear uPDFs coincide. It translates also into the rapidity spectrum. Only the KS-nonlinear uPDF leads to a somewhat different behaviour of the cross section at small $p_{T}$’s. We observe that both the transverse momentum and rapidity distributions of charm quark are sensitive to the non-linear evolution effects that lead here to a sizeable damping of the predicted cross section. Thus, the forward production of charm within intrinsic charm concept might be a very good testing ground for studies of the non-linear term in the evolution of unintegrated gluon densities and may shed new light on phenomenon of parton saturation. ![ The same as in Fig. \[fig6\] but here results for four different unintegrated gluon densities in a proton are shown. Here only the $g^* c \rightarrow g c$ mechanism is taken into account. []{data-label="fig7"}](dsig_dpT_hyb_uGDFs.eps){width="100.00000%"} ![ The same as in Fig. \[fig6\] but here results for four different unintegrated gluon densities in a proton are shown. Here only the $g^* c \rightarrow g c$ mechanism is taken into account. []{data-label="fig7"}](dsig_dy_hyb_uGDFs.eps){width="100.00000%"} Above, we have used those gluon uPDF models that are assumed to allow for an effective resummation of extra real emissions (real higher-order terms). Therefore, their can be successfully used in phenomenological studies based even on leading-order matrix elements (see a discussion in Refs. [@Maciula:2019izq; @Maciula:2020cfy]). Here we wish to present results obtained within the scheme of the calculations where the higher-order corrections are not resummed in the uPDF but are taken into account via the hard-matrix elements. This procedure can be tested with the help of the DGLAP-based Parton-Branching uPDFs as was proposed in Ref. [@Maciula:2019izq] and further applied in Refs. [@Lipatov:2019izq; @Maciula:2020cfy]. In Fig. \[fig8\] we show predictions of the hybrid model for the $2 \to 2$, $2 \to 3$ and $2 \to 4$ mechanisms, as well as for their sum $2 \to 2+3+4$ obtained using a dedicated merging procedure. The results are calculated with the PB-NLO-set1 quark and gluon uPDFs. For the leading-order $2\to 2$ mechanisms we show $g^c$ and $q^c$ channels separately while for the higher-order components we plot sum of all possible gluonic and quark channels. As in the collinear case, the higher-order mechanisms are found to be very important also here. ![ The same as in Fig. \[fig6\] but here results for PB-NLO-set1 unintegrated parton densities obtained within the $2 \to 2+3+4$ scheme of the calculation. Here, the $2\to2$, $2\to3$, and $2\to4$ components as well as their sum $2\to 2+3+4$ obtained including merging procedure are shown separately. []{data-label="fig8"}](dsig_dpT_hyb_pb-dce.eps){width="100.00000%"} ![ The same as in Fig. \[fig6\] but here results for PB-NLO-set1 unintegrated parton densities obtained within the $2 \to 2+3+4$ scheme of the calculation. Here, the $2\to2$, $2\to3$, and $2\to4$ components as well as their sum $2\to 2+3+4$ obtained including merging procedure are shown separately. []{data-label="fig8"}](dsig_dy_hyb_pb-dce.eps){width="100.00000%"} For a better transparency in Fig. \[fig9\] we compare the hybrid model results obtained with the KMR-CT14lo (solid histograms) with the PB-NLO-set1 (dashed histograms) uPDFs, that correspond to the two different hybrid calculation schemes, together with the results obtained in the collinear approach (dotted histograms). Both types of the hybrid model calculations seem to lead to a very similar predictions. It seems to justify the proposed $2 \to 2+3+4$ hybrid calculation scheme with the PB uPDFs and with the applied merging in a qualitative way. On the other hand, the collinear $2 \to 2+3+4$ results seem to be larger by a factor of 2 than their hybrid model counterpart. However, this might be related to a lack of a relevant merging procedure in the collinear case. ![ The same as in Fig. \[fig6\] but here we compare results for the CT14nnlo collinear PDFs with $2 \to 2+3+4$ collinear model calculations, for the KMR-CT14lo uPDFs with the $2\to 2$ hybrid model calculations and for the PB-NLO-set1 uPDFs with $2 \to 2+3+4$ hybrid model calculations including merging. []{data-label="fig9"}](dsig_dpT_hyb_pb-dce-vs-kmr-vs-nlo.eps){width="100.00000%"} ![ The same as in Fig. \[fig6\] but here we compare results for the CT14nnlo collinear PDFs with $2 \to 2+3+4$ collinear model calculations, for the KMR-CT14lo uPDFs with the $2\to 2$ hybrid model calculations and for the PB-NLO-set1 uPDFs with $2 \to 2+3+4$ hybrid model calculations including merging. []{data-label="fig9"}](dsig_dy_hyb_pb-dce-vs-kmr-vs-nlo.eps){width="100.00000%"} The $\bm{k_{T}}$-factorization approach --------------------------------------- Now we wish to present results obtained within the $k_{T}$-factorization approach. So here we take into account also effects related to off-shellness of $c$ quark of the intrinsic charm in the proton. The transverse momentum dependent intrinsic charm uPDF is obtained by Gaussian smearing of the collinear PDF. Rather small smearing parameter is used that do not allow for a large transverse momenta of the intrinsic charm. It seems to be appropriate for the case of the forward production of charm. For the unintegrated gluon density the KMR-CT14lo model is used. In Fig. \[fig10\] we show predictions for the $g^c^ \to gc$ mechanism with both initial state partons being off-shell. Here three different values of the smearing parameter in the calculation of the intrinsic charm uPDF are used: $\sigma_{0} = 0.5$ GeV (solid histograms), $3.5$ GeV (dotted histograms) and $7.0$ GeV (dashed histograms). The larger $\sigma_{0}$ is taken the smaller cross section at small outgoing charm quark transverse momenta is obtained (left panel). The same is true for the rapidity spectrum in the forward region (right panel). ![ The charm quark transverse momentum (left) and rapidity (right) differential cross sections for $pp$-scattering at $\sqrt{s}=7$ TeV. The results correspond to the $g^* c^* \rightarrow g c$ mechanism calculated within the intrinsic charm concept in the $k_{T}$-factorization approach with both off-shell initial state partons. Here the KMR-CT14lo unintegrated gluon density and Gaussian $k_{t}$-distribution for off-shell charm quark were used. We show results for different values of the smearing parameter $\sigma$. []{data-label="fig10"}](dsig_dpT_kTfact-2to2-gauss.eps){width="100.00000%"} ![ The charm quark transverse momentum (left) and rapidity (right) differential cross sections for $pp$-scattering at $\sqrt{s}=7$ TeV. The results correspond to the $g^* c^* \rightarrow g c$ mechanism calculated within the intrinsic charm concept in the $k_{T}$-factorization approach with both off-shell initial state partons. Here the KMR-CT14lo unintegrated gluon density and Gaussian $k_{t}$-distribution for off-shell charm quark were used. We show results for different values of the smearing parameter $\sigma$. []{data-label="fig10"}](dsig_dy_kTfact-2to2-gauss.eps){width="100.00000%"} In Fig. \[fig11\] we illustrate mutual relations between the results obtained with the hybrid and the $k_T$-factorization frameworks. When the smearing parameter in the calculation of the intrinsic charm uPDF is small, e.g. $\sigma_{0} = 0.5$ GeV, the hybrid model $g^c \to gc$ results coincide with the $g^c^* \to gc$ results obtained within the full $k_T$-factorization approach. ![ The same as in Fig. \[fig10\] but here we compare results for the hybrid $g^c\to gc$ and the $k_{T}$-factorization $g^c^* \to gc$ calculations obtained with the KMR-CT14lo unintegrated gluon densities. The off-shell charm quark Gaussian $k_{t}$-distribution is obtained with the the smearing parameter $\sigma_0 = 0.5$ GeV. []{data-label="fig11"}](dsig_dpT_hyb-vs-kTfact.eps){width="100.00000%"} ![ The same as in Fig. \[fig10\] but here we compare results for the hybrid $g^c\to gc$ and the $k_{T}$-factorization $g^c^* \to gc$ calculations obtained with the KMR-CT14lo unintegrated gluon densities. The off-shell charm quark Gaussian $k_{t}$-distribution is obtained with the the smearing parameter $\sigma_0 = 0.5$ GeV. []{data-label="fig11"}](dsig_dy_hyb-vs-kTfact.eps){width="100.00000%"} Finally, we wish to present results of the $k_{T}$-factorization approach for the $g^c^ \to c$ mechanism. In Fig. \[fig12\] we compare the corresponding predictions obtained with the four different gluon uPDFs: the KMR-CT14lo (solid lines), the JH-2013-set2 (dotted lines), the KS-linear (dash-dotted lines) and the KS-nonlinear (dashed lines). Different models lead to quite different results. The discrepancies between the uPDF models obtained here seem to be larger than in the corresponding case of the $g^c \to gc$ calculations within the hybrid model. ![ The charm quark transverse momentum (left) and rapidity (right) differential cross sections for $pp$-scattering at $\sqrt{s}=7$ TeV. The results correspond to the $g^ c^* \rightarrow c$ mechanism calculated within the intrinsic charm concept in the $k_{T}$-factorization approach with both off-shell initial state partons. Here the Gaussian $k_{t}$-distribution for off-shell charm quark were used. We show results for different gluon uPDFs. []{data-label="fig12"}](dsig_dpT_kTfact_2to1.eps){width="100.00000%"} ![ The charm quark transverse momentum (left) and rapidity (right) differential cross sections for $pp$-scattering at $\sqrt{s}=7$ TeV. The results correspond to the $g^* c^* \rightarrow c$ mechanism calculated within the intrinsic charm concept in the $k_{T}$-factorization approach with both off-shell initial state partons. Here the Gaussian $k_{t}$-distribution for off-shell charm quark were used. We show results for different gluon uPDFs. []{data-label="fig12"}](dsig_dy_kTfact_2to1.eps){width="100.00000%"} Predictions for future experiments ---------------------------------- Before we go to predictions for different future and present experiments we wish to summarize the conclusions drawn in the previous subsection by a direct comparison of the results corresponding to the approaches discussed above. ![ The charm quark transverse momentum (left) and rapidity (right) differential cross sections for $pp$-scattering at $\sqrt{s}=7$ TeV. Here we compare predictions of the three different approaches used in the previous subsections: the $2 \to 2+3+4$ collinear, the hybrid $g^c \to gc$ and the $k_{T}$-factorization $g^c^* \to c$ calculations. []{data-label="fig13"}](dsig_dpT_summary.eps){width="100.00000%"} ![ The charm quark transverse momentum (left) and rapidity (right) differential cross sections for $pp$-scattering at $\sqrt{s}=7$ TeV. Here we compare predictions of the three different approaches used in the previous subsections: the $2 \to 2+3+4$ collinear, the hybrid $g^c \to gc$ and the $k_{T}$-factorization $g^c^* \to c$ calculations. []{data-label="fig13"}](dsig_dy_summary.eps){width="100.00000%"} In Fig. \[fig13\] we compare predictions of the three different approaches used in the previous subsections: the $2 \to 2+3+4$ collinear (dashed histograms), the hybrid $g^c \to gc$ (solid histograms) and the $k_{T}$-factorization $g^c^* \to c$ (solid lines) calculations. Different models lead to a very different results with more than one order of magnitude difference between the lowest and the highest predicted cross section. Huge cross section for $g c \to c$ or $c g \to c$ may be partly due to ignoring other emissions than $c$ or $\bar c$ in the evolution of $x_1$ and $x_2$. These large uncertainties of the predictions can be reduced only by a forward experiments at forward directions. Forward charm production data sets that will be dominated by the contribution from intrinsic charm are necessary to draw definite conclusions about the level of applicability of the different theoretical approaches. Therefore, now we wish to present results of the study of the impact of the intrinsic charm component on the forward charm particle production in already existing or future experiments at different energies. We start with predictions for the high energy experiments at the LHC and the FCC, at $\sqrt{s}=13$ and $50$ TeV, respectively (top and bottom panels in Fig. \[fig14\]). In the LHC case the considered kinematics correspond to the planned FASER experiment. Here we compare predictions of the $k_{T}$-factorization approach for the $g^g^ \to c\bar c$ mechanism which is known to give a very good description of the LHC open charm data [@Maciula:2019izq], and predictions of the $g^c \to gc$ mechanism (dashed) within the hybrid model. In both cases the charm production cross section starts to be dominated by the intrinsic charm component at very forward rapidities, i.e.* $y \geq 7$. In this far-forward region, the transverse momentum distribution of charm quark is also dominated by the contributions of the intrinsic charm. The predicted enhancement of the charm cross section could certainly be examined by the FASER experiment dedicated to a measurement of forward neutrinos originating from semileptonic decays of $D$ mesons. The actual predictions for neutrinos will be presented elsewhere. ![ Predictions of the impact of the intrinsic charm component in charm quark production in different experiments. Here we explore kinematics relevant for the FASER experiment at the LHC and an exemplary experiment at the FCC. []{data-label="fig14"}](dsig_dy_hyb_faser.eps){width="100.00000%"} ![ Predictions of the impact of the intrinsic charm component in charm quark production in different experiments. Here we explore kinematics relevant for the FASER experiment at the LHC and an exemplary experiment at the FCC. []{data-label="fig14"}](dsig_dpT_hyb_faser.eps){width="100.00000%"} \ ![ Predictions of the impact of the intrinsic charm component in charm quark production in different experiments. Here we explore kinematics relevant for the FASER experiment at the LHC and an exemplary experiment at the FCC. []{data-label="fig14"}](dsig_dy_hyb_fcc.eps){width="100.00000%"} ![ Predictions of the impact of the intrinsic charm component in charm quark production in different experiments. Here we explore kinematics relevant for the FASER experiment at the LHC and an exemplary experiment at the FCC. []{data-label="fig14"}](dsig_dpT_hyb_fcc.eps){width="100.00000%"} In addition, we also analysed a possibility of experimental study of the intrinsic charm concept at lower energies. In Fig. \[fig15\] we show predictions for the fixed-target LHC and the SHIP experiment, at $\sqrt{s}=86.6$ and $27.4$ GeV, respectively (top and bottom panels). We observe that also at relatively small energies the intrinsic charm contributions could be identified experimentally. It seems that the already existing data set on open charm meson production in the fixed-target LHC mode [@Aaij:2018ogq] needs to have the intrinsic charm component included in the theoretical description. Similarly, our results suggests that the predictions of the tau-neutrino flux that could be measured in the SHIP experiment should include effects related to a possible intrinsic charm content of the proton. ![ Predictions of the impact of the intrinsic charm component in charm production in different experiments. Here we explore kinematics of the fixed-target mode LHCb and the kinematics relevant for the SHIP experiment. []{data-label="fig15"}](dsig_dy_hyb_lhcb.eps){width="100.00000%"} ![ Predictions of the impact of the intrinsic charm component in charm production in different experiments. Here we explore kinematics of the fixed-target mode LHCb and the kinematics relevant for the SHIP experiment. []{data-label="fig15"}](dsig_dpT_hyb_lhcb.eps){width="100.00000%"} \ ![ Predictions of the impact of the intrinsic charm component in charm production in different experiments. Here we explore kinematics of the fixed-target mode LHCb and the kinematics relevant for the SHIP experiment. []{data-label="fig15"}](dsig_dy_hyb_ship.eps){width="100.00000%"} ![ Predictions of the impact of the intrinsic charm component in charm production in different experiments. Here we explore kinematics of the fixed-target mode LHCb and the kinematics relevant for the SHIP experiment. []{data-label="fig15"}](dsig_dpT_hyb_ship.eps){width="100.00000%"} Conclusions =========== In this paper we have discussed the effect of intrinsic charm in the proton on forward production of $c$ quark or $\bar c$ antiquark at different energies. Three different approaches: collinear, hybrid and $k_{T}$-factorization have been used with modern collinear and unintegrated parton distribution functions. The production mechanism of $c$-quarks and $\bar c$-antiquarks originating from intrinsic charm in the nucleon is concentrated in forward/backward directions, but details depend on collision energy. The absolute normalization strongly depends on the approach used. The leading-order (LO) collinear framework leads to the smallest cross section. The cross section becomes much bigger in the $k_{T}$-factorization or in the hybrid model which effectively include higher-order corrections. The next-to-leading (NLO) and even next-to-next-to-leading (NNLO) tree-level corrections are found to be very important here. Therefore, the $k_{T}$-factorization or the hybrid model will give stringent limits on the intrinsic charm which cannot be constrained at present from first principles. We have shown that in the collinear approach the LO calculations of the intrinsic charm component are insufficient. We have included the NLO and NNLO components at tree-level which were found to significantly contribute to the cross section. Working in the hybrid model or in the $k_{T}$-factorization approach we have shown that the effects related to the off-shellness of the incident partons (especially gluons) are large. In both cases higher-order corrections are effectively included already within the basic $gc \to gc$ mechanism. We have used different models for gluon unintegrated parton distribution functions (uPDFs) from the literature. We obtained different results for different gluon uPDFs. The forward charm production was recognized as a useful testing ground for the small-$x$ behaviour of the gluon uPDFs. We have shown in addition that the final results are also sensitive to the concept of gluon saturation in a proton. Unintegrated gluon densities derived from linear and non-linear evolution equations lead to a quite different results. We have performed also leading-order calculations within $k_T$-factorization approach where the basic process is either $g + c \to c$ or $c + g \to c$ as done for forward production of charm quarks. We have shown that the intrinsic charm component dominates over the standard pQCD (extrinsic) mechanism of $c\bar c$-pair production at forward (or far-forward) rapidities starting from low energy fixed-target experiment at $\sqrt{s}=27.4$ and $86.6$ GeV, through the LHC Run II nominal energy $\sqrt{s}=13$ TeV, and up to the energies relevant for the IceCube experiment ($\sqrt{s}=50$ TeV). The LHC experiments at low energies (fixed-target experiments) can provide valuable information already now. Future LHC experiments on $\nu_{\tau}$ neutrino production such as SHIP sand FASER are an interesting alternative in next few years. In the present study we intentionally limited to the production of charm quarks/antiquarks. The production of charmed mesons or baryons is currently uder discussion and a new fragmentation scheme was proposed [@szczurek2020] very recently. We leave the predictions for production of charmed hadrons and their semileptonic decays for a separate study. However, the consequences for high-energy neutrino production have been discussed shortly in the context of the IceCube experiment and experiments proposed at the LHC (SHIP and FASER). Acknowledgements {#acknowledgements .unnumbered} ================ We would like to thank Victor Goncalves for a discussion on IC. This study was partially supported by the Polish National Science Center grant UMO-2018/31/B/ST2/03537 and by the Center for Innovation and Transfer of Natural Sciences and Engineering Knowledge in Rzeszów. [100]{} S.J. Brodsky, P. Hoyer, C. Peterson and N. Sakai, Phys. Lett. [B93]{}, 451 (1980). V.D. Barger, F. Halzen and W.Y. Keung, Phys. Rev. [D25]{}, 112 (1982). R. Vogt and S.J. Brodsky, Nucl. Phys. [B438]{}, 261 (1995), [478]{}, 311 (1996). F.S. Navarra, M. Nielsen, C.A.A. Nunes and M. Teixeira, Phys. Rev. [D54]{}, 842 (1996). W. Melnitchouk and A.W. Thomas, Phys. Lett. [B414]{}, 134 (1997). F.M. Steffens, W. Melnitchouk and A.W. Thomas, Eur. Phys. J. [C11]{}, 673 (1999). F. Carvalho, F.O. Duraes, F.S. Navarra and M. Nielsen, Phys. Rev. Lett. [86]{} (2001) 5434. T.J. Hobbs, J.T. Londergan and W. Melnitchouk, Phys. Rev. [D89]{}, 074008 (2014). R.S. Sufian, T. Liu, A. Alexandru, S.J. Brodsky, G.F. de Teramond, H. G. Dosch, T. Draper, K.-F. Liu and Y.-B. Yang. arXiv:2003.01078 \[hep-lat\]. P.H. Beauchemin, V.A. Bednyakov, G.I. Lykasov, Y.Y. Stepanenko, Phys. Rev. [D92]{}, 0341014 (2015). S. Rostami, A. Khorramian, A. Aleedaneshvar and M. Goharpour, J. Phys. [G43]{}, 055001 (2016). A.V. Lipatov, G.I. Lykasov, Y.Y. Stepanenko and V.A. Bednyakov, Phys. Rev. [D94]{}, 053011 (2016). V.A. Bednyakov, S.J. Brodsky, A.V. Lipatov, G.I. Lykasov, M.A. Malyshev, J. Smiesko and S. Tokar, arXiv:1712.09096 \[hep-ph\]. S. Brodsky, A. Kusina, F. Lyonnet, I. Schienbein, H. Spiesberger and R. Vogt, Adv. High Energy Phys. 2015, 231547 (2015). R. D. Ball et al. \[NNPDF\], J. High Energy Phys. 04, 040 (2015). T. J. Hou, S. Dulat, J. Gao, M. Guzzi, J. Huston, P. Nadolsky, C. Schmidt, J. Winter, K. Xie and C. P. Yuan, J. High Energy Phys. 02, 059 (2018). R. Enberg, M.H. Reno and I. Sarcevic, Phys. Rev. [D78]{}, 043005 (2008). R. Laha and S. Brodsky, Phys. Rev. [D96]{}, 123002 (2017), arXiv:1607.08240. A.V. Giannini, V.P. Goncalves and F.S. Navarra, Phys. Rev. [D98]{}, 014012 (2018). R. Maciu[ł]{}a and A. Szczurek, Phys. Rev. D 97, no.7, 074001 (2018). W. Bai, M. Diwan, M. V. Garzelli, Y. S. Jeong and M. H. Reno, \[arXiv:2002.03012 \[hep-ph\]\]. W. Bai and M. H. Reno, J. High Energy Phys. 02, 077 (2019). R. Aaij et al. \[LHCb\], Phys. Rev. Lett. 122, no.13, 132002 (2019). R. Maciu[ł]{}a, \[arXiv:2003.05702 \[hep-ph\]\]. V.P. Goncalves, F.S. Navarra and T. Ulrich, Nucl. Phys. [A842]{} 59 (2010). F. Carvalho, A.V. Giannini, V.P. Goncalves and F.S. Navarra, Phys. Rev. [D96]{}, 094002 (2017). A. Edin and G. Ingelman, Phys. Lett. [B432]{}, 402 (1998). M. Gluck, E. Reya and A. Vogt, Z. Phys. C 67, 433-448 (1995). J. Pumplin, H.L. Lai and W.K. Tung, Phys. Rev. [D75]{}, 054029 (2007). J. C. Collins, D. E. Soper and G. F. Sterman, Adv. Ser. Direct. High Energy Phys. 5, 1-91 (1989). R. D. Ball et al. \[NNPDF\], Eur. Phys. J. C 76, no.11, 647 (2016). L. Harland-Lang, A. Martin, P. Motylinski and R. Thorne, Eur. Phys. J. C 75, no.5, 204 (2015). P. Jimenez-Delgado and E. Reya, Phys. Rev. D 89, no.7, 074049 (2014). A. van Hameren, Comput. Phys. Commun. 224, 371-380 (2018). T. Sjöstrand, S. Ask, J. R. Christiansen, R. Corke, N. Desai, P. Ilten, S. Mrenna, S. Prestel, C. O. Rasmussen and P. Z. Skands, Comput. Phys. Commun. 191, 159-177 (2015). M. Deak, F. Hautmann, H. Jung and K. Kutak, J. High Energy Phys. 09, 121 (2009). K. Kutak and S. Sapeta, Phys. Rev. D 86, 094043 (2012). P. Kotko, K. Kutak, C. Marquet, E. Petreska, S. Sapeta and A. van Hameren, J. High Energy Phys. 09, 106 (2015). F. Hautmann and H. Jung, Nucl. Phys. B [883]{}, 1 (2014). K. Kutak, Phys. Rev. D [91]{}, no. 3, 034021 (2015). A. Bermudez Martinez, P. Connor, H. Jung, A. Lelek, R. Žlebčík, F. Hautmann and V. Radescu, Phys. Rev. D [99]{}, no. 7, 074008 (2019). G. Watt, A. D. Martin and M. G. Ryskin, Eur. Phys. J. C [31]{}, 73 (2003). A. van Hameren, private communication R. Maciu[ł]{}a and A. Szczurek, Phys. Rev. D [100]{}, no. 5, 054001 (2019). A. V. Lipatov, M. A. Malyshev and H. Jung, Phys. Rev. D [101]{}, no. 3, 034022 (2020). R. Maciu[ł]{}a, arXiv:2003.05702 \[hep-ph\]. S. Catani, M. Ciafaloni and F. Hautmann, Phys. Lett. B242 (1990) 97; Nucl. Phys. B366 (1991) 135; Phys. Lett. B307 (1993) 147.\ J.C. Collins and R.K. Ellis, Nucl. Phys. B360, 3 (1991).\ L.V. Gribov, E.M. Levin, and M.G. Ryskin, Phys. Rep. 100, 1 (1983);\ E.M. Levin, M.G. Ryskin, Yu.M. Shabelsky and A.G. Shuvaev, Sov. J. Nucl. Phys. 53, 657 (1991). M. Nefedov, V. Saleev and A. V. Shipilova, Phys. Rev. D 87, no.9, 094030 (2013). M. Czech and A. Szczurek, J. Phys. [G32]{}, 1253 (2006). I. Babiarz, R. Pasechnik, W. Schäfer and A. Szczurek, J. High Energy Phys. [02]{} (2020) 037. I. Babiarz, R. Pasechnik, W. Schäfer and A. Szczurek, J. High Energy Phys. 06, 101 (2020). P. Lebiedowicz, R. Maciu[ł]{}a and A. Szczurek, Phys. Lett. [B806]{}, 135475 (2020). R. Aaij et al. (LHCb collaboration), Phys. Lett.[B718]{}, 902 (2013). K. Kutak, Phys. Rev. [D91]{}, 034021 (2015). V.P. Goncalves, R. Maciu[ł]{}a, R. Pasechnik and A. Szczurek, Phys. Rev. [D96]{} (2017) 094026. V.P. Goncalves, R. Maciu[ł]{}a and A. Szczurek, Phys. Lett. [B794]{} (2019) 29. R. Maciu[ł]{}a, A. Szczurek, J. Zaremba and I. Babiarz, J. High Energy Phys. [01]{} (2020) 116. A. Szczurek, arXiv:2006.12918 \[hep-ph\]. [^1]: also at University of Rzeszów, PL-35-959 Rzeszów, Poland [^2]: Here and in the following we concentrate only on the forward production mechanisms (with charm quark having positive-rapidity), but, the formalism for symmetric backward configuration is the same. [^3]: Technically, this could be done properly only if the parton level calculations are supplemented with a parton shower but it goes beyond the scope of the present study.	{ "pile_set_name": "ArXiv" }
--- abstract: \| We present night sky brightness measurements performed at the Vienna University Observatory and at the Leopold-Figl-Observatorium für Astrophysik, which is located about 35km to the southwest of Vienna. The measurements have been performed with Sky Quality Meters made by Unihedron. They cover a time span of roughly one year and have been carried out every night, yielding a night sky brightness value every 7 seconds and thus delivering a large amount of data. In this paper, the level of light pollution in Vienna, which ranges from 15 to 19.25 is presented for the very first time in a systematic way. We discuss the influence of different environmental conditions on the night sky brightness and implications for human vision. We show that the circalunar rhythm of night sky brightness is extinguished at our observatory due to light pollution. Additionally, we present spectra of the night sky in Vienna, taken with a 0.8m telescope. The goal of these spectroscopic measurements was to identify the main types of light sources and the spectral lines which cause the light pollution in Vienna. It turned out that fluorescent lamps are responsible for the strongest lines of the night sky above Vienna (e.g. lines at 546nm and at 611nm). address: 'Universität Wien, Institut für Astrophysik, Türkenschanzstra[ß]{}e 17, A-1180 Wien, Austria' author: - 'Johannes Puschnig, Thomas Posch, Stefan Uttenthaler' bibliography: - 'bibliography.bib' title: \| Night sky photometry and spectroscopy\ performed at the Vienna University Observatory --- atmospheric effects – site testing – light pollution – techniques: photometric – techniques: spectroscopic Introduction \[sec:intro\] ========================== Up to now, astronomers have mainly been measuring the brightness of the night sky (in magnitudes/arcsec$^2$) at dark sites, especially at modern mountain observatories, or at potential observatory sites as a part of “site-testing” and “site-monitoring”. Within the past few years, it has become evident that increasing night sky brightness and light pollution have far-reaching consequences for many branches of human life as well as wildlife (e.g. Posch et al. [@Posch2010]). Therefore, it is desirable to measure and monitor the night sky brightness not only at remote mountaintop observatory locations (as done, e.g., by Patat [@Patat2008]), but also close to the centers of modern civilization, and to do so every night, in a reproducible way, with the aim of performing long-term studies (such as in climate research). Only in this way can the impact of night sky brightness on biological rhythms on animal behaviour, and human health be assessed. This is the aim of our ongoing night sky brightness measurements in Vienna and at the “Leopold-Figl-Observatorium für Astrophysik” (LFOA) on Mount Mitterschöpfl. We complemented our measurements with spectroscopic studies of the night sky at the Vienna University Observatory, which hosts the “Institut für Astrophysik” (IfA). It should be noted that all our measurements refer to scattered light. We measured only brightness values and spectra of the sky background, while we explicitly avoided to measure any direct radiation from streetlamps. This is what we call “night sky brightness” (NSB). In other words: the brightness values and the spectra presented in this paper refer to the total backscattered light of the night sky. Its origin is the whole ensemble of streetlamps, facade illuminations, illuminated billboards etc. at the respective observing site and in its near and far surroundings. The natural nocturnal radiation from the Earth’s atmosphere, which is produced by different processes such as recombination of atoms that have been ionized by the Sun’s radiation during daytime, contributes very little to the NSB that we measured at our observing sites, since the latter is dominated by the influence of artificial light. Measurement sites and methods \[sec:methods\] ============================================= Measurement sites ----------------- First measurements began in November 2011 at the IfA. Since April 2nd, 2012 the measuring device (label “IFA”) is mounted at its ultimate place and points exactly to the zenith. About the same time (March 21st, 2012) we started measuring the NSB at our remote mountaintop site too. Currently we use three devices in total, two of which are located at LFOA (label “FOA” and “FOA2”). The geographical positions of our sites are given below. Location (altitude) Longitude Latitude --------------------- ------------------------ ------------------------ IfA (240m) 16$^{\circ}$ 20’ 03’ E 48$^{\circ}$ 13’ 54” N LFOA (880m) 15$^{\circ}$ 55’ 24” E 48$^{\circ}$ 05’ 03” N : Geographical coordinates of our measurement sites. The distances of the sites IfA and LFOA from the city center are 3.5km and 35km, respectively. \[t:sites\] Photometric measurements ------------------------ ### Units used for our photometric measurements The devices that we use for our NSB measurements yield data in a unit which is very widespread in astronomy, namely magnitudes per square arcsecond (mag/ arcsec$^2$). This unit corresponds to a luminance, but is a logarithmic measure derived from stellar photometry, where larger values correspond to fainter objects. In the same way, larger NSB values in mag/arcsec$^2$ indicate darker skies (with less light pollution). Equation (\[eq:unihedron\]) gives the conversion from magarcsec$^{-2}$ to and Table \[t:conversion\] lists selected pairs of corresponding values. $$Luminance\ [cd/m^2] = 10.8\times10^4 \times 10^{(-0.4\ \times\ [mag arcsec^{-2}])} \label{eq:unihedron}$$ mag/arcsec$^2$ comment ---------------- ------- ---------------------------------------- 14 271 15 108 16 43.0 17 17.1 18 6.81 19 2.71 20 1.08 21 0.430 21.75 0.215 value we assume for the [overall]{} natural clear sky brightness 22.0 0.172 value we assume for the [zenithal]{} clear sky brightness : Conversion between mag arcsec$^{-2}$ and . \[t:conversion\] For the purpose of monitoring the night sky brightness over a long period of time, we use Unihedron’s Sky Quality Meter with an integrated lensing system. The model is called “SQM-LE”, but hereafter we will refer to it as “SQM”. All our devices are connected via ethernet. According to the manufacturer’s manual, the integrated lens narrows down the field of view to only $\approx$20$^{\circ}$ full width at half maximum (FWHM), thus at an off-axis distance of $\approx$10$^{\circ}$, the sensitivity declines by a factor of 2. At higher angles the response decreases rapidly, so that a point source located $\approx$19$^{\circ}$ off-axis contributes a factor of 10 less to the measured brightness than on-axis. We have chosen the lensed version because we also use the devices in urban regions where the field of view is limited due to surrounding buildings. A wide beam width then could result in lower accuracy due to the possible influence of nearby lights. An advantage of the ethernet version is that with network cables far greater distances can be reached than with USB cables and by making use of the “Power over Ethernet” (PoE) technology also power supply at remote sites is easier to handle. Furthermore, the PoE device directly attached to the SQM leads to a sufficient amount of heating as to avoid the formation of dew. ### Spectral response function of the SQM and comparison to the Johnson V band Sky Quality Meters are equipped with a photo diode, a filter and a temperature sensor for thermal stabilization. The manufacturer provides information on the sensitivity curve of the photo diode and the filter, but not the resulting response function. Previous reports indicated that the night sky brightness values measured by SQMs do not differ strongly from those from systems equipped with a Johnson V filter. Nevertheless, given that SQM magnitudes are not the same as V magnitudes, we will use the unit “” throughout this paper in order to make this difference clear. The spectral sensitivity of the SQM was measured by Cinzano [@Cinzano2005]. A comparison between the SQM response function and the Johnson V band shows that the SQM is much more sensitive to light below $\approx$520nm. At a first glance the difference seems to make a calibration between the Johnson V band and the SQM response function difficult, but considering the spectral distribution of the night sky over Vienna which is dominated by emission lines above $\approx$525nm as seen in Fig. \[fig:skyspectrum\], we presume that a constant offset can be found for a comparison to the Johnson V band. Measurements by other groups indicate that the SQM actually underestimates the brightness of the night sky compared to V band measurements – precisely due to its enhanced sensitivity in the blue (which is less prominent in the night sky than e.g. in the twilight sky). This is supported also by our own preliminary comparative measurements with instruments equipped with a Johnson V filter. The difference may amount up to 0.6mag.for clouded skies, while it is smaller for clear skies (H. Spoelstra, priv. comm.). ![SQM Response function as measured by Cinzano [@Cinzano2005] in comparison to the Johnson V filter response function defined by the Johns Hopkins University [@Dobos2004].[]{data-label="fig:sqmresponse"}](Figure1.eps){width="13cm"} For the moment both SQM and Johnson V measurements seem to be suitable to quantify artificial NSB. But having in mind that light emitting diodes (LEDs) are developing fast, a change in the night sky brightness composition can be expected. Since future LEDs of color temperatures above 3000K may have a significant emittance in the blue part of the spectrum, Johnson V band measurements would not be able to register a change in color of the NSB. Since the SQM is sensitive to the blue end of the spectrum, it seems to be prepared for future measurements of the sky brightness in case of stronger contributions at the blue end of the spectrum. The best one can do without spectroscopy is attaching a set of SQMs parallel to each other, each equipped with a standard photometric filter (LRGB or UBVR). Such a configuration makes it possible to study the NSB development and to record possible changes in the color of the night sky at the same time (see Kyba et al. [@Kyba2012a]). ### Sampling rate We started our measurements with a sampling of 5 minutes, but soon it was noticed that with overcast skies, we miss some features within our light curves. Therefore in the next step the frequency was increased to a reading once per minute. This seems to be a good value for most applications and a good compromise between data storage, data handling and valuable information. But since we want to measure all possible variations that can occur, we even increased the sampling rate to a reading every 7 seconds which corresponds to a frequency of 0.143 Hz. The device measures accurately at that high frequency and so far we did not encounter any problems due to integration time limitations. We are able to detect lightning, fireworks and other short-term events. Moreover, with higher sampling it is possible to quantify the influence of the curfew in Vienna even in cloudy nights, when the light curves are much more disturbed. ### Automated data storage and publishing All our data are logged and stored automatically at each monitoring station. From there every five minutes a synchronization script uploads the data to the University’s web server for data mirroring and data publishing leading to “real-time” data visualization on our website\ <http://astro.univie.ac.at/en/institute/light-pollution/>. A tool is provided for calculating all relevant information about illuminated moon fraction, moon rise, moon set and dates for civil, nautical and astronomical twilights. Furthermore we calculate mean, best and worst values as well as statistical information for the defined time intervals (e.g. nautical twilight). Some other values like the “naked eye limiting magnitude” or the “clear sky factor” are calculated, but still in an experimental phase. A future discussion on that is planned. ### Intercomparison and relative metering precision of our two SQMs For checking the quality of the photometric data, we have made a parallel measurement with our two devices “IFA” and “FOA2” over a period of 33 days. Since an absolute calibration to any given standard such as Johnson V seems to be difficult for the moment, the question to be answered was, weather both of our devices give similar response under the same environmental conditions. Therefore, the devices were installed at exactly the same place at our institute and pointed to the zenithal region of the sky. The daily comparative light curves observed can be found in the appendix (Figs. \[fig:qualitylightcurves1\] to \[fig:qualitylightcurves3\]). Analysis of the data showed that for the substantial fraction of measurements the differences of both device data readings lie within an interval of only $\approx$0.1 . But it was also recognized that under certain circumstances scattered moonlight caused differences of up to $\approx$0.9. A discussion on that topic can be found in the Appendix. Neglecting scattered moonlight values, the results show that the initial calibration of our SQMs is offset by only $\approx$0.046 . Taking that offset into account the relative accuracy between the two devices is given by the standard deviation of all data points. Fig. \[fig:qualityhistogram\] shows the distribution of the calculated differences ($\Delta$) binned to 0.04 . It can be seen that after rejecting data with scattered moonlight for 87.5 % of all data points the difference in the response of our devices is equal to or lower than 0.08 leading to a standard deviation of $\approx$0.04 which corresponds to a relative error of only 4 %. Under the assumption of a 3 % growth of the NSB for Vienna, we will be able to detect a significant ($2\sigma$) change of the NSB level after 3 years (by the end of 2015). So far the SQM appears to be a handy and robust instrument for long term studies of the development of the nightsky brightness. Under some circumstances scattered moonlight can influence the readings, therefore some sort of shielding should be used for the lensed SQM version. ![Histogram showing the distribution of the differences between the measurements of 2 SQMs operated in parallel mode. Red lines mark distribution before correction for scattered moonlight. The bin size is 0.04 .[]{data-label="fig:qualityhistogram"}](Figure2.eps){width="13cm"} Spectroscopic measurements -------------------------- The spectroscopic observations, presented below, were done with a 80cm Cassegrain telescope in the North Dome of the Vienna University Observatory. Its focal length amounts to 6.7m. A spectrograph ’DSS-7’ produced by SBIG was attached to the telescope. DSS-7 is a grating spectrograph offering five slits with different widths, the narrowest of which provides for a dispersion of $\sim0.55$nm/pixel, corresponding to a spectral resolving power of $\lambda$/$\Delta$$\lambda$ $\approx$ 500. An ST-7 CCD camera, also made by SBIG, was connected to the spectrograph to acquire the spectra. With this camera, the wavelength coverage is approximately 450–850nm (with some small variations between the observing runs). The chip of the CCD camera was cooled down to -10$^{\circ}$C during the observations. To record the spectra of the night sky, the telescope was pointed to the south or to the south-east (i.e. towards the city center of Vienna) at an elevation of $\sim$45$^{\circ}$. Frames of 300s integration time were obtained. All spectroscopic observations were carried out during the evening hours. Some of our night sky spectra were obtained as dedicated observations, others were extracted as a by-product of stellar spectroscopy, in which case they were subtracted as background. The 2D raw spectra were dark-subtracted and flat-fielded with dome flats (using light bulbs mounted on the telescope). To extract the 1D spectra, all pixel rows that were illuminated by the narrowest slit were summed up along the slit direction. The wavelength calibration was done using the sky emission lines themselves by adopting a second-order polynomial to map pixel positions to wavelength values. We expect that night sky spectra will change worldwide during the next few decades due to increasing numbers of light emitting diodes (LEDs) used for street and facade lighting. It is desirable to monitor the resulting changes in the average night sky spectra in order to see whether indeed the relative flux at short wavelengths ($\lambda$ $<$ 500nm) will thereby increase (which would be unfortunate for many reasons). Data analysis ============= In the following all given daily mean values in and other statistical parameters for the NSB are calculated for time intervals between the astronomical twilights of the given date. As described in Section \[sec:methods\] the device “FOA2” mounted at our remote site was calibrated against our older SQM “IFA” located at our institute in Vienna. Detection of the natural circalunar rhythm ------------------------------------------ Our measurements meanwhile cover a period of one year. We were therefore interested in the behavior of the mean night sky brightness at IfA and at LFOA during this period (March 2012 – March 2013). At a rural site, a strong influence of the lunar phase on the NSB is expected: full moon nights should be recognizable by much brighter night skies, especially for clear sky conditions. In contrast, nights around new moon should be much darker. Since LFOA is a rural site, this pattern is indeed detected and a nice circalunar periodicity of the mean NSB becomes evident (see Fig. \[fig:lightcurve\]). The NSB data from Vienna show a completely different picture. Instead of the moon, the degree of cloudiness has the strongest influence on the NSB in Vienna. The more clouds over a big city, the stronger the backscattering of urban light. Of course, the degree of cloudiness does not follow any strict periodicity. Therefore, the mean NSB measured in Vienna seems to vary in a stochastic way. The circalunar rhythm is barely recognizable here. The lower half of Fig. \[fig:lightcurve\] also shows a circannual periodicity of the NSB close to full moon: winter nights with the full moon at higher elevations are significantly brighter (by $\approx$ 1.5) than summer nights. ![Mean daily NSB within astronomical twilight. Distribution over one year (date format: yyyy/mm/dd) obtained at both sites IfA (top) and LFOA (bottom). At LFOA the circalunar rhythm dominates the variations, whereas in Vienna the natural variation is barely recognizable. The LFOA data also show a circannual period depending on the maximum elevation of the full moon (see dashed lines).[]{data-label="fig:lightcurve"}](Figure3a.eps "fig:"){width="11.5cm"} ![Mean daily NSB within astronomical twilight. Distribution over one year (date format: yyyy/mm/dd) obtained at both sites IfA (top) and LFOA (bottom). At LFOA the circalunar rhythm dominates the variations, whereas in Vienna the natural variation is barely recognizable. The LFOA data also show a circannual period depending on the maximum elevation of the full moon (see dashed lines).[]{data-label="fig:lightcurve"}](Figure3b.eps "fig:"){width="11.5cm"} Range of NSB values under urban conditions ------------------------------------------ Figures \[fig:lightcurve\] (upper part) and \[fig:IFAdarkness\] clearly show that the NSB values we measure at our urban site IfA vary within a large range. The darkest skies reached at IfA are characterized by an NSB slightly above (= darker than) 19.25 , corresponding to a zenithal luminance of 2.15 . The brightest skies reached at IfA under overcast conditions have an NSB below (= brighter than) 15 , corresponding to a zenithal luminance of more than 100 , which is more than 0.1 and can easily be measured even with a simple luminance meter. As Fig. \[fig:IFAdarkness\] shows, not all NSB values occur equally often. The contour plot – which is based on more than 2 $\times$ 10$^{6}$ individual data points – shows two values which occur much more often than others: one is 19.1 – with a trend to still slightly larger (darker) values as the early morning hours (with less traffic and less artificial lights) are reached. These values are characteristic for clear skies (no clouds, little aerosol content of the air, relatively little amount of backscattering). The other range of frequently occurring NSBs is about 16.3, corresponding to 33. This value is about 13 times larger than the former one and corresponds to completely overcast conditions. The values between 16 and 19 are measured when the sky is partly cloudy, or cloudless but hazy, or clear but not moonless. Furthermore Fig. \[fig:IFAdarkness\] shows that the trend of NSB development in Vienna during one night typically follows a gradient of 0.1 per hour. It is also interesting to compare our “best” measured values for the urban conditions at IfA with the predited sky brightness that we extracted from Cinzano, Falchi & Elvidge [@Cinzano2001] in their “World Atlas of the artificial night sky brightness”. We find that our urban observatory IfA has a predicted clear-sky NSB of approximately 10 times the natural night sky brightness. Our above reported best value of 19.25 or 2.15 amounts to 12.5 times the natural zenithal night sky brightness, for which we assume 0.172 (Duriscoe, priv. comm.; see also Tab. \[t:conversion\]). ![Density plot of all measurements obtained at Vienna (IfA) from April 2012 to April 2013. A gradient of 0.1 per hour and two dominant NSB ranges can be identified.[]{data-label="fig:IFAdarkness"}](Figure4.eps){width="11.5cm"} Detecting the curfew(s) in Vienna (11 p.m., 12 p.m.) ---------------------------------------------------- The city of Vienna reduces a substantial fraction of its street lighting at 11 p.m. (curfew, in the following: ‘C’). Main roads, however, are still kept at constant illuminance throughout the whole night. At 12 p.m., most of the decorative facade lighting is switched off (this effect is called ‘F’ below). The sky brightness, therefore, shows two discontinuous steps towards lower values at 11 and 12 p.m., of which the first one is larger. This is shown in Fig. \[fig:curfew\]. On average, step C results in a night sky brightness decrease by 0.18 , while step F leads to a decrease (= improvement of the sky quality) by 0.09 . It would be worthwhile to monitor similar effects of curfews in other cities and communities, also to demonstrate this benefit for the nocturnal environment. ![Detection of the curfew at 11 p.m. (C) and reduction of facade lighting at 12 p.m. (F) in Vienna on March 21st, 2012.[]{data-label="fig:curfew"}](Figure5.eps){width="12cm"} Influence of snow and other seasonal variations ----------------------------------------------- Snow has a strong influence on the NSB, especially in urban regions. The reflectance of snow is clearly much larger than that of bitumen. Hence, streets, pavements and squares covered by (fresh) snow reflect a huge portion of the downward directed radiation of streetlamps back to the sky. Moreover, a night with snowfall is per se a cloudy night – hence, there is a second reason for such nights to be brighter, which is the backscattering of the upward light by the clouds. The combination of those two effects leads to extremely large values of the NSB during some winter nights. The fact that the mean NSB values shown in the upper half of Fig. \[fig:lightcurve\] reach values of 15 is due to the influence of clouded, snowy nights during the first weeks of 2013. This value corresponds to a zenithal luminance of 0.1, which is 461 times the natural NSB. Under the assumption of a uniform sky brightness of 0.1 (from the zenith to the horizon), an illuminance meter would display 0.31lx – which is more than in the case of full moon. In reality, the sky is always brighter near the horizon than at the zenith, such that our measurement of 0.1, according to our experiences, corresponds to about 0.5lx. A night sky brightness of 15 or 0.1 indicates very bright skies from the astronomical point of view, and yet we detected still larger NSB values at certain times – not as mean values for whole nights, but as records lasting at least for some minutes to hours. For example, during the night from the 6th to the 7th of February 2012 (see Fig. \[fig:IFArecord\]), up to 13.99 were recorded at 10:30 p.m. This spectacularly bright value corresponds to a luminance of 0.27 or 1570 times the natural zenithal sky brightness. In this case, which typically occurs when there is snow cover and ongoing snow fall, the diffuse backscattered light from the sky alone would illuminate a landscape with more than 1lx. Under such conditions, one could even consider dimming down or switching off local illuminations of small streets and paths, since any landscape which is uniformly illuminated at 1lx appears brighter than the same landscape with “punctual” illuminations even at peak values of several lux. ![During the night from February 6th to February 7th 2012, the NSB reached a peak value corresponding to 0.27 or 1570 times the natural zenithal sky brightness.[]{data-label="fig:IFArecord"}](Figure6.eps){width="10.5cm"} Special cases: thunderstorms, New Year’s Eve -------------------------------------------- Our NSB measurements not only show the strong influence of lighting, but also the strong (but short-term) influence of lightning. Given that thunderstorms occur much more frequently during the summer months in Austria, the detection rate of lightning is highest between June and September. Even with our sampling rate of one measurement every 7 seconds, bright lightning can have a strong impact in our night sky photometry. During two nights in June and August 2012, our NSB curves measured at Vienna (IfA) show “spikes” due to lightning with a baseline value of 16 and peak values up to 11, which corresponds to an increase by a factor of 100. At the LFOA site, in contrast, we find even an increase by a factor of 10000 (!) due to lightning during the night from June 20 to June 21, 2012 (see Fig. \[fig:FOAlightning\]). Since lightning has extremely high luminance, but cover only a tiny fraction of the sky, it might be more interesting to measure the illuminances (in lux) caused by lightning. To the best of our knowledge, this has not been done in a systematic way so far. A similar effect as the one caused by thunderstorms is observed – at least in Vienna – during New Year’s Eve. The effect of local fireworks is again clearly detected in the NSB measurements. In this case, we measured a short-term increase of the sky brightness from 18 to 16 around midnight. Again, this is only a lower limit of the influence of fireworks on the NSB, since our SQMs are pointed toward the zenith, and hardly any fireworks rocket would reach the local zenith at the observatory, which is surrounded by a large park. ![During the night from June 20th to June 21st 2012, a short term increase in the NSB from about 20 up to 10 is found at our mountaintop observing station LFOA – corresponding to an increase in the NSB by a factor of 10000 due to lightning.[]{data-label="fig:FOAlightning"}](Figure7.eps){width="10.5cm"} Implications of our results for the transition from mesopic to scoptopic vision ------------------------------------------------------------------------------- As shown in Figure \[fig:histo\], the mean zenithal NSB values in Vienna (location IfA) are in the range 15 to 19.5. The black vertical line represents a sky brightness of . This corresponds to a luminance of or 3 . We take this value as the lower limit of the cones in human vision. At smaller luminances, the color-insensitive rods entirely dominate our vision. If the NSB is brighter than , the adaptation of the human eye to darkness is very incomplete. The partial visibility of our Milky Way at least near the zenith requires a value which is reported to be close to 18.9 as well. According to Fig. \[fig:histo\], almost 90% of all nights at IfA are characterized by mean zenithal NSBs larger than the “” threshold: during the vast majority of all nights, therefore, entirely rod-dominated vision as well as Milky Way visibility is not reached. At our rural, 880m-above-sea-level observing station LFOA, in contrast, only 20% of all nights have mean zenithal NSB values brighter than 18.9 . The full range of mean NSBs at LFOA is 17 to 21 . Note that even the latter value is still about 1 magnitude (or a factor of 2.5) above the natural zenithal night sky brightness. ![Histogram showing the relative (top) and cumulative (bottom) distribution of the mean NSB measured at IfA in Vienna (black dashed bars) and LFOA (red solid bars). The vertical line indicates the limit between scotopic and mesopic vision at .[]{data-label="fig:histo"}](Figure8a.eps "fig:"){width="11.5cm"} ![Histogram showing the relative (top) and cumulative (bottom) distribution of the mean NSB measured at IfA in Vienna (black dashed bars) and LFOA (red solid bars). The vertical line indicates the limit between scotopic and mesopic vision at .[]{data-label="fig:histo"}](Figure8b.eps "fig:"){width="11.5cm"} Spectral lines detected in the Night Sky ---------------------------------------- Figure \[fig:skyspectrum\] shows one of our night sky spectra, taken on the 1st of February 2012. For comparison also laboratory spectra of common street lamps are shown. The strongest spectral line that we detect in the light-polluted night sky over Vienna is located at 546nm. An adjacent, but weaker line is detected at 544nm. Both lines are due to fluorescent lamps (FL), which are currently the most common type of lighting, especially in residential areas. The second strongest spectral feature that we find in our night sky spectra is centered at 611nm. Its origin is a composite of fluorescent lamps and high pressure sodium (HPS) lamps, which are used to illuminate suburban highways, crossroads and main streets. Between these two strongest spectral peaks, a series of lines are detected, most of which can be identified as sodium lines (outstanding peaks at 569nm, 588nm and 593nm). Another sodium line is detected in the photographic infrared, at 819nm. Since low-pressure sodium lines are uncommon in and around Vienna, the sodium peaks must be due to scattered light from high pressure sodium (HPS) lamps. More detailed information on the emission lines of HPS lamps can be found in Elvidge et al. [@Elvidge2010]. Furthermore, atmospheric absorption bands can be identified at 762nm (A band) and 688nm (B band) due to magnetic dipole electronic transitions of the molecule O$_2$ ([@Kiehl1985]). Peak position \[nm\] Assignment (see text) Remark ---------------------- ----------------------- ---------------- 544 FL 546 FL 569/588/598 HPS 611 FL blend with HPS 819 HPS : Selected (strong) spectral lines detected in Vienna’s night sky. \[t:spectrallines\] ![ Night sky spectrum (filled area) obtained in Vienna on February 1st, 2012 towards SE. It can be clearly seen that the most prominent features correlate with emission lines originating from FL (blue solid line) and HPS (orange solid line). Laboratory spectra of street lamps can be found on the website of the National Oceanic and Atmospheric Administration [@NOAA].[]{data-label="fig:skyspectrum"}](Figure9.eps){width="13cm"} According to the official statistics of the city of Vienna, 80% of all the lamps used for public illumination are fluorescent lamps, 16% are HPS, while other types of lamps add up to only 4% by number. This is in good accordance with our spectroscopic results. The total power used for street lighting in Vienna amounts to 14 MW (value from the official website <http://wien.gv.at> [@WIEN]. Assuming a lumen-to-Watt ratio of 70 for fluorescent lamps and 120 for HPS, we end up with a contribution of 784 Megalumen from the dominant source (= fluorescent lamps) and 269 Megalumen from HPS lamps. Together with the other 4% (among them high pressure mercury vapor lamps), the total light output of the city of Vienna, as far as it is caused by street lighting, amounts to about 1.1 Gigalumen, or about 622 lumen per capita. Conclusions =========== After one year of NSB measurements at an urban site (in Vienna) and at a rural mountaintop, we may draw the following conclusions: 1. Due to its high sensitivity covering the bulk of the visible spectrum, Sky Quality Meters are appropriate not only for long-term studies of the NSB, but also for short-term events. Nevertheless, an adequate site, weather proof housing and shade are crucial, such that the sensor is protected against scattered (moon) light. Once tested, the instrument will monitor the NSB development within a relative error of approximately 5 %. 2. The factor which is essentially determining the night sky brightness at light-polluted urban sites is the degree of cloudiness (see also Kyba et al. [@Kyba2012a]), since clouds strongly enhance the backscattering of artificial light in the atmosphere. 3. Typical NSB values at our urban measurement site IfA are 16.3 or 33 under overcast conditions. Note that without any artificial light, the overcast sky would be much darker than clear skies . The opposite is the case for cities nowadays. 4. Maximum values of the NSB at our urban measurement location are close to 14 or 0.3 . This value refers exclusively to the diffuse (scattered) light from the night sky. It does not include any influence of lights shining directly into our detector. 5. At our rural mountaintop site, the range of NSB values is typically 17 $\dots$ 21 . Here, the phase of the moon remains the determining factor for the NSB, which probably implies a strong zeitgeber function of the moonlight at this rural place. 6. In contrast, the circalunar illuminance pattern is extinguished in urban and, very likely, to a large extent also in suburban regions. This will also influence the chronobiological rhythms of animals living in local “islands of darkness” such as big parks. 7. By a spectroscopic analysis of the light scattered back from the sky, we were able to identify lines from high pressure sodium lamps (e.g. at 569nm, at 588nm and at 593nm) and from fluorescent lamps (e.g. at 546nm and at 611nm) as main “pollutants” of the night sky. 8. For studying the night sky color distribution, filter systems e.g. Johnson UBV, should be used in the future. As indicated by our spectroscopic analysis, currently the blue end of the spectrum below 525nm does not show significant contribution to the NSB. This could change in the future, if lamps of high brightness temperatures will be used in the future (which would be unfortunate for many reasons: e.g. insect attraction, melatonin suppression, and enhanced light pollution). [Acknowledgements]{} We thank Otto Beck, Erich Schäfer and Werner Zeilinger for their help with the installation of the SQMs in Vienna and especially at the Mitterschöpfl Observatory. We gratefully acknowledge Stefan Meingast’s help with the spectroscopic measurements. S. Uttenthaler acknowledges support from the Austrian Science Fund (FWF) under project P 22911-N16. Discussion on the origin of the scatter while parallel measurements of two SQMs =============================================================================== We performed parallel measurements of our two SQM devices over a period of 33 days. The daily light curves showing the measured NSB value and the difference between the readings of the devices are shown in Figs. \[fig:qualitylightcurves1\] to \[fig:qualitylightcurves3\]. It can be seen that both devices gave the same or similar response for most of the measurements and the calibration of the devices as delivered by the manufacturer is very accurate. However, within certain time intervals the differences increased up to 0.9. After checking for moon phase, altitude and azimuth, it was realized that these big differences always occur 1) at certain alt/azimuth coordinates of the moon, 2) when the moon illuminated fraction was greater than 50 % and 3) in the same “sine-wave-like” shape. Thus it is clear that the differences are a result of scattered moonlight. We assume that this affects only the SQM-LE, the lensed SQM version and suggest to use some sort of shielding or shading (see Fig. \[fig:sqms\]) when using the device. Since the lensed version has a very narrow field of view of only $\approx$20$^{\circ}$, some sort of shade can be installed easily. ![Parallel measurements of two SQMs. One of the devices is equipped with a simple shade for preventing scattered moon light.[]{data-label="fig:sqms"}](FigureA10.eps){width="9cm"} ![image](FigureA11.eps){width="13cm"} ![image](FigureA12.eps){width="13cm"} ![image](FigureA13.eps){width="13cm"}	{ "pile_set_name": "ArXiv" }
--- abstract: 'Given the apparent difficulty of learning models that are robust to adversarial perturbations, we propose tackling the simpler problem of developing adversarially robust features. Specifically, given a dataset and metric of interest, the goal is to return a function (or multiple functions) that 1) is robust to adversarial perturbations, and 2) has significant variation across the datapoints. We establish strong connections between adversarially robust features and a natural spectral property of the geometry of the dataset and metric of interest. This connection can be leveraged to provide both robust features, and a lower bound on the robustness of any function that has significant variance across the dataset. Finally, we provide empirical evidence that the adversarially robust features given by this spectral approach can be fruitfully leveraged to learn a robust (and accurate) model.' author: - \| Shivam Garg\ Vatsal Sharan[^1]\ Brian Hu Zhang\ Gregory Valiant Stanford University\ Stanford, CA 94305\ `{shivamgarg, vsharan, bhz, gvaliant}@stanford.edu ` - \| Shivam Garg\ Department of Computer Science\ Stanford University\ Stanford, CA 94305\ `[email protected]`\ Vatsal Sharan[^2]\ Department of Computer Science\ Stanford University\ Stanford, CA 94305\ `[email protected]` Brian Hu Zhang\ Department of Computer Science\ Stanford University\ Stanford, CA 94305\ `[email protected]` Gregory Valiant\ Department of Computer Science\ Stanford University\ Stanford, CA 94305\ `[email protected]`\ bibliography: - 'main.bib' title: A Spectral View of Adversarially Robust Features --- Introduction ============ While machine learning models have achieved spectacular performance in many settings, including human-level accuracy for a variety of image recognition tasks, these models exhibit a striking vulnerability to adversarial examples. For nearly every input datapoint—including training data—a small perturbation can be carefully chosen to make the model misclassify this perturbed point. Often, these perturbations are so minute that they are not discernible to the human eye. Since the initial work of [@szegedy2013intriguing] and [@goodfellow2014explaining] identified this surprising brittleness of many models trained over high-dimensional data, there has been a growing appreciation for the importance of understanding this vulnerability. From a conceptual standpoint, this lack of robustness seems to be one of the most significant differences between humans’ classification abilities (particularly for image recognition tasks), and computer models. Indeed this vulnerability is touted as evidence that computer models are not really learning, and are simply assembling a number of cheap and effective, but easily fooled, tricks. Fueled by a recent line of work demonstrating that adversarial examples can actually be created in the real world (as opposed to requiring the ability to edit the individual pixels in an input image) [@evtimov2017robust; @brown2017adversarial; @kurakin2016adversarial; @athalye2017synthesizing], there has been a significant effort to examine adversarial examples from a security perspective. In certain settings where trained machine learning systems make critically important decisions, developing models that are robust to adversarial examples might be a requisite for deployment. Despite the intense recent interest in both computing adversarial examples and on developing learning algorithms that yield robust models, we seem to have more questions than answers. In general, ensuring that models trained on high-dimensional data are robust to adversarial examples seems to be extremely difficult: for example, @athalye2018obfuscated claims to have broken six attempted defenses submitted to ICLR 2018 before the conference even happened. Additionally, we currently lack answers to many of the most basic questions concerning why adversarial examples are so difficult to avoid. What are the tradeoffs between the amount of data available, accuracy of the trained model, and vulnerability to adversarial examples? What properties of the geometry of a dataset determine whether a robust and accurate model exists? The goal of this work is to provide a new perspective on robustness to adversarial examples, by investigating the simpler objective of finding adversarially robust features. Rather than trying to learn a robust function that also achieves high classification accuracy, we consider the problem of learning any function that is robust to adversarial perturbations with respect to any specified metric. Specifically, given a dataset of $d$-dimensional points and a metric of interest, can we learn features—namely functions from ${\mathbb R}^d\rightarrow {\mathbb R}$—which 1) are robust to adversarial perturbations of a bounded magnitude with respect to the specified metric, and 2) have significant variation across the datapoints (which precludes the trivially robust constant function). There are several motivations for considering this problem of finding robust features: First, given the apparent difficulty of learning adversarially robust models, this is a natural first step that might help disentangle the confounding challenges of achieving robustness and achieving good classification performance. Second, given robust features, one can hope to get a robust model if the classifier used on top of these features is reasonably Lipschitz. While there are no a priori guarantees that the features contain any information about the labels, as we empirically demonstrate, these features seem to contain sufficient information about the geometry of the dataset to yield accurate models. In this sense, computing robust features can be viewed as a possible intermediate step in learning robust models, which might also significantly reduce the computational expense of training robust models directly. Finally, considering this simpler question of understanding robust features might yield important insights into the geometry of datasets, and the specific metrics under which the robustness is being considered (e.g. the geometry of the data under the $\ell_{\infty}$, or $\ell_2$, metric.) For example, by providing a lower bound on the robustness of any function (that has variance one across the datapoints), we trivially obtain a lower bound on the robustness of any classification model. Robustness to Adversarial Perturbations --------------------------------------- Before proceeding, it will be useful to formalize the notion of robustness (or lack thereof) to adversarial examples. The following definition provides one natural such definition, and is given in terms of a distribution $D$ from which examples are drawn, and a specific metric, $dist(\cdot,\cdot)$ in terms of which the magnitude of perturbations will be measured. \[def:rob\_dist\] A function $f: {\mathbb R}^d \rightarrow {\mathbb R}$ is said to be $(\eps,\delta,\gamma)$ robust to adversarial perturbations for a distribution $D$ over ${\mathbb R}^d$ with respect to a distance metric $dist: {\mathbb R}^d \times {\mathbb R}^d \rightarrow {\mathbb R}$ if, for a point $x$ drawn according to $D$, the probability that there exists $x'$ such that $dist(x,x')\le\eps$ and $\|f(x)-f(x')\| \ge \delta$, is bounded by $\gamma$. Formally, $$\Pr_{x \sim D}[\exists x' \text{ s.t. } dist(x,x')\le\eps \text{ and } \|f(x)-f(x')\| \ge \delta] \le \gamma.$$ In the case that the function $f$ is a binary classifier, if $f$ is $(\eps, 1, \gamma)$ robust with respect to the distribution $D$ of examples and a distance metric $d$, then even if adversarial perturbations of magnitude $\eps$ are allowed, the classification accuracy of $f$ can suffer by at most $\gamma$. Our approach will be easier to describe, and more intuitive, when viewed as a method for assigning feature values to an entire dataset. Here the goal is to map each datapoint to a feature value (or set of values), which is robust to perturbations of the points in the dataset. Given a dataset $X$ consisting of $n$ points in ${\mathbb R}^d$, we desire a function $F$ that takes as input $X$, and outputs a vector $F(X) \in {\mathbb R}^n$; such a function $F$ is robust for a dataset $X$ if, for all $X'$ obtained by perturbing points in $X$, $F(X)$ and $F(X')$ are close. Formally, let $\mathcal{X}$ be the set of all datasets consisting of $n$ points in ${\mathbb R}^d$, and $\norm{\cdot}$ denote the $\ell_2$ norm. For notational convenience, we will use $F_X$ and $F(X)$ interchangeably, and use $F_X(x)$ to denote the feature value $F$ associates with a point $x \in X$. We overload $dist(\cdot,\cdot)$ to define distance between two ordered sets $X = (x_1, x_2, \ldots, x_n)$ and $X' = (x_1', x_2', \ldots, x_n')$ as $dist(X, X') = max_{i \in [n]} \ dist(x_i, x_i')$. With these notations in place, we define a robust function as follows: \[def:rob\_dataset\] A function $F: \mathcal{X} \rightarrow {\mathbb R}^n$ is said to be $(\eps,\delta)$ robust to adversarial perturbations for a dataset $X$ with respect to a distance metric $dist(\cdot, \cdot)$ as defined above, if, for all datasets $X'$ such that $dist(X, X') \leq \epsilon$, $\norm{F(X)-F(X')} \leq \delta$. If a function $F$ is $(\epsilon, \delta)$ robust for a dataset $X$, it implies that feature values of $99\%$ of the points in $X$ will not vary by more than $\frac{10\delta}{\sqrt{n}}$ if we were to perturb all points in $X$ by at most $\epsilon$. As in the case of robust functions of single datapoints, to preclude the possibility of some trivial functions we require $F$ to satisfy certain conditions: 1) $F_{X}$ should have significant variance across points, say, $\sum_i (F_X(x_i) - \operatorname{\mathbb E}_{x \sim Unif(X)}[F_X(x)])^2 = 1$. 2) Changing the order of points in dataset X should not change $F_X$, that is, for any permutation $\sigma: {\mathbb R}^n \rightarrow {\mathbb R}^n$, $F_{\sigma(X)} = \sigma(F_X)$. Given a data distribution $D$, and a threshold $\epsilon$, the goal will be to find a function $F$ that is as robust as possible, in expectation, for a dataset $X$ drawn from $D$. We mainly follow Definition \[def:rob\_dataset\] throughout the paper as the ideas behind our proposed features follow more naturally under that definition. However, we briefly discuss how to extend these ideas to come up with robust features of single datapoints (Definition \[def:rob\_dist\]) in section \[subsec:feature\_extension\]. Summary of Results ------------------ In Section \[s:method\], we describe an approach to constructing features using spectral properties of an appropriately defined graph associated with a dataset in question. We show provable bounds for the adversarial robustness of these features. We also show a synthetic setting in which some of the existing models such as neural networks, and nearest-neighbor classifier are known to be vulnerable to adversarial perturbations, while our approach provably works well. In Section \[s:lower\_bound\], we show a lower bound which, in certain parameter regimes, implies that if our spectral features are not robust, then no robust features exist. The lower bound suggests a fundamental connection between the spectral properties of the graph obtained from the dataset, and the inherent extent to which the data supports adversarial robustness. To explore this connection further, in Section \[s:experiments\], we show empirically that spectral properties do correlate with adversarial robustness. In Section \[s:experiments\], we also test our adversarial features on the downstream task of classification on adversarial images, and obtain promising results. All the proofs have been deferred to the appendix. Shortcomings and Future Work ---------------------------- Our theory and empirics indicate that there may be fundamental connections between spectral properties of graphs associated with data and the inherent robustness to adversarial examples. A worthwhile future direction is to further clarify this connection, as it may prove illuminating and fruitful. Looking at the easier problem of finding adversarial features also presents the opportunity of developing interesting sample-complexity results for security against adversarial attacks. Such results may be much more difficult to prove for the problem of adversarially robust classification, since generalization is not well understood (even in the non-adversarial setting) for classification models such as neural networks. Our current approach involves computing distances between all pair of points, and performing an eigenvector computation on a Laplacian matrix of a graph generated using these distances. Both of these steps are computationally expensive operations, and future work could address improving the efficiency of our approach. In particular, it seems likely that similar spectral features can be approximated without computing all the pairwise distances, which would result in significant speed-up. We also note that our experiments for testing our features on downstream classification tasks on adversarial data is based on transfer attacks, and it may be possible to degrade this performance using stronger attacks. The main takeaway from this experiment is that our conceptually simple features along with a linear classifier is able to give competitive results for reasonable strong attacks. Future works can possibly explore using robustly trained models on top of these spectral features, or using a spectral approach to distill the middle layers of neural networks. Related Work ------------ One of the very first methods proposed to defend against adversarial examples was adversarial training using the [fast gradient sign method]{} (FGSM) [@goodfellow2014explaining], which involves taking a step in the direction of the gradient of loss with respect to data, to generate adversarial examples, and training models on these examples. Later, @madry2017towards proposed a stronger [projected gradient descent]{} (PGD) training which essentially involves taking multiple steps in the direction of the gradient to generate adversarial examples followed by training on these examples. More recently, @kolter2017provable, @raghunathan2018certified, and @sinha2017certifiable have also made progress towards training provably adversarially robust models. There have also been efforts towards proving lower bounds on the adversarial accuracy of neural networks, and using these lower bounds to train robust models [@hein2017formal; @peck2017lower]. Most prior work addresses the question of how to fix the adversarial examples problem, and there is less work on identifying why this problem occurs in the first place, or highlighting which geometric properties of datasets make them vulnerable to adversarial attacks. Two recent works specifically address the “why” question: @fawzi2018adversarial give lower bounds on robustness given a specific generative model of the data, and @schmidt2018adversarially and @bubeck2018adversarial describe settings in which limited computation or data are the primary bottleneck to finding a robust classifier. In this work, by considering the simpler task of coming up with robust features, we provide a different perspective on both the questions of “why” adversarial perturbations are effective, and “how” to ensure robustness to such attacks. Background: Spectral Graph Theory --------------------------------- Let $G = (V(G), E(G))$ be an undirected, possibly weighted graph, where for notational simplicity $V(G) = \{1, \dots, n\}$. Let $A = (a_{ij})$ be the adjacency matrix of $G$, and $D$ be the the diagonal matrix whose $i$th diagonal entry is the sum of edge weights incident to vertex $i$. The matrix $L = D - A$ is called the [Laplacian matrix]{} of the graph $G$. The quadratic form, and hence the eigenvalues and eigenvectors, of $L$ carry a great deal of information about $G$. For example, for any $v \in {\mathbb R}^n$, we have $$v^{T}L v = \sum_{(i, j) \in E(G) }a_{ij} \qty(v_i - v_j)^2.$$ It is immediately apparent that $L$ has at least one eigenvalue of $0$: the vector $v_1 = (1, 1, \dots, 1)$ satisfies $v^{T}Lv = 0$. Further, the second (unit) eigenvector is the solution to the minimization problem $$\label{eq:eigprob} \min_{v} \sum_{(i, j) \in E }a_{ij} \qty(v_i - v_j)^2 \qq{s.t.} \sum_i v_i = 0;\quad \sum_i v_i^2 = 1.$$ In other words, the second eigenvector assigns values to the vertices such that the average value is 0, the variance of the values across the vertices is 1, and among such assignments, minimizes the sum of the squares of the discrepancies between neighbors. For example, in the case that the graph has two (or more) connected components, this second eigenvalue is 0, and the resulting eigenvector is constant on each connected component. Our original motivation for this work is the observation that, at least superficially, this characterization of the second eigenvector sounds similar in spirit to a characterization of a robust feature: here, neighboring vertices should have similar value, and for robust features, close points should be mapped to similar values. The crucial question then is how to formalize this connection. Specifically, is there a way to construct a graph such that the neighborhood structure of the graph captures the neighborhood of datapoints with respect to the metric in question? We outline one such construction in Section \[s:method\]. We will also consider the [normalized]{} or [scaled Laplacian]{}, which is defined by $$L_\text{norm} = D^{-1/2}(D - A) D^{-1/2} = I - D^{-1/2} A D^{-1/2}.$$ The scaled Laplacian normalizes the entries of $L$ by the total edge weights incident to each vertex, so that highly-irregular graphs do not have peculiar behavior. For more background on spectral graph theory, we refer the readers to @spielman2007spectral and @chung1997spectral. Robust Features {#s:method} =============== In this section, we describe a construction of robust features, and prove bounds on their robustness. Let $X = (x_1,\ldots,x_n )$ be our dataset, and let $\eps > 0$ be a threshold for attacks. We construct a robust feature $F_X$ using the second eigenvector of the Laplacian of a graph corresponding to $X$, defined in terms of the metric in question. Formally, given the dataset $X$, and a distance threshold parameter $T>0$ which possibly depends on $\eps$, we define $F_X$ as follows: Define $G(X)$ to be the graph whose nodes correspond to points in $X$, i.e., $\{ x_1, \dots, x_n\}$, and for which there is an edge between nodes $x_i$ and $x_j$, if $dist(x_i, x_j) \le T$. Let $L(X)$ be the (un-normalized) Laplacian of $G(X)$, and let $\lambda_k(X)$ and $v_k(X)$ be its $k$th smallest eigenvalue and a corresponding unit eigenvector. In all our constructions, we assume that the first eigenvector $v_1(X)$ is set to be the unit vector proportional to the all-ones vector. Now define $F_X(x_i) = v_2\qty(X)_i$; i.e. the component of $v_2(X)$ corresponding to $x_i$. Note that $F_X$ defined this way satisfies the requirement of sufficient variance across points, namely, $\sum_i (F_X(x_i) - \operatorname{\mathbb E}_{x \sim Unif(X)}[F_X(x)])^2 = 1$, since $\sum_i v_2(X)_i = 0$ and $\norm{v_2(X)} = 1$. We now give robustness bounds for this choice of feature $F_X$. To do this, we will need slightly more notation. For a fixed $\eps > 0$, define the graph $G^+(X)$ to be the graph with the same nodes as $G(X)$, except that the threshold for an edge is $T + 2\eps$ instead of $T$. Formally, in $G^+(X)$, there is an edge between $x_i$ and $x_j$ if $dist(x_i, x_j) \le T + 2\epsilon$. Similarly, define $G^-(X)$ to be the graph with same set of nodes, with the threshold for an edge being $T - 2\eps$. Define $L^+(X), \lambda_k^+(X), v_k^+(X), L^-(X), \lambda_k^-(X), v_k^-(X)$ analogously to the earlier definitions. In the following theorem, we give robustness bounds on the function $F$ as defined above. [theorem]{}[thmmain]{}\[thm:main\] For any pair of datasets $X$ and $X'$, such that $dist(X, X') \le \eps$, the function $F: \mathcal{X}_n \rightarrow {\mathbb R}^n$ obtained using the second eigenvector of the Laplacian satisfies $$\label{eq:main} \min(\norm{F(X) - F(X')}, \norm{(-F(X)) - (F(X'))}) \le \qty(2\sqrt{2})\sqrt{ \frac{\lambda_2^+(X) - \lambda_2^-(X)}{\lambda_3^-(X) - \lambda_2^-(X)} }.$$ Theorem \[thm:main\] essentially guarantees that the features, as defined above, are robust up to sign-flip, as long as the eigengap between the second and third eigenvalues is large, and the second eigenvalue does not change significantly if we slightly perturb the distance threshold used to determine whether an edge exists in the graph in question. Note that flipping signs of the feature values of all points in a dataset (including training data) does not change the classification problem for most common classifiers. For instance, if there exists a linear classifier that fits points with features $F_X$ well, then a linear classifier can fit points with features $-F_X$ equally well. So, up to sign flip, the function $F$ is $(\eps, \delta_X)$ robust for dataset $X$, where $\delta_X$ corresponds to the bound given in Theorem \[thm:main\]. To understand this bound better, we discuss a toy example. Consider a dataset $X$ that consists of two clusters with the property that the distance between any two points in the same cluster is at most $4\epsilon$, and the distance between any two points in different clusters is at least $10 \epsilon$. Graph $G(X)$ with threshold $T = 6 \epsilon$, will have exactly two connected components. Note that $v_2(X)$ will perfectly separate the two connected components with $v_2(X)_i$ being $\frac{1}{\sqrt{n}}$ if $i$ belongs to component 1, and $\frac{-1}{\sqrt{n}}$ otherwise. In this simple case, we conclude immediately that $F_X$ is perfectly robust: perturbing points by $\epsilon$ cannot change the connected component any point is identified with. Indeed, this agrees with Theorem \[thm:main\]: $\lambda_2^+ = \lambda_2^- = 0$ since the two clusters are at a distance $> 10\eps$. Next, we briefly sketch the idea behind the proof of Theorem \[thm:main\]. Consider the second eigenvector $v_2(X')$ of the Laplacian of the graph $G(X')$ where dataset $X'$ is obtained by perturbing points in $X$. We argue that this eigenvector can not be too far from $v_2^-(X)$. For the sake of contradiction, consider the extreme case where $v_2(X')$ is orthogonal to $v_2^-(X)$. If the gap between the second and third eigenvalue of $G^-(X)$ is large, and the difference between $\lambda_2(X')$ and $\lambda_2^-(X)$ is small, then by replacing $v_3^-(X)$ with $v_2(X')$ as the third eigenvector of $G^-(X)$, we get a much smaller value for $\lambda_3^-(X)$, which is not possible. Hence, we show that the two eigenvectors in consideration can not be orthogonal. The proof of the theorem extends this argument to show that $v_2(X')$, and $v_2^-(X)$ need to be close if we have a large eigengap for $G^-(X)$, and a small gap between $\lambda_2(X')$ and $\lambda_2^-(X)$. Using a similar argument, one can show that $v_2(X)$ and $v_2^-(X)$, also need to be close. Applying the triangle inequality, we get that $v_2(X)$ and $v_2(X')$ are close. Also, since we do not have any control over $\lambda_2(X')$, we use an upper bound on it given by $\lambda_2(X+)$, and state our result in terms of the gap between $\lambda_2(X^+)$ and $\lambda_2(X^-)$. The approach described above also naturally yields a construction of a [set]{} of robust features by considering the higher eigenvectors of Laplacian. We define the $i^{\text{th}}$ feature vector for a dataset $X$ as $F^i_X = v_{i+1}(X)$. As the eigenvectors of a symmetric matrix are orthogonal, this gives us a set of $k$ diverse feature vectors $\{F^1_X, F^2_X, \dots, F^k_X\}$. Let $F_X(x) = (F_X^1(x),F_X^k(x), \dots, F_X^k(x))^T$ be a $k$-dimensional column vector denoting the feature values for point $x \in X$. In the following theorem, we give robustness bounds on these feature vectors. [theorem]{}[thmmainmulti]{}\[thm:main1\] For any pair of datasets $X$ and $X'$, such that $dist(X, X') \le \eps$, there exists a $k\times k$ invertible matrix $M$, such that the features $F_X$ and $F_{X'}$ satisfy $$\label{eq:main1} \sqrt{\sum_{i \in [n]} \norm{M F_X(x_i) - F_{X'}(x_i')}^2} \le \qty(2\sqrt{2 k}) \sqrt{ \frac{\lambda_{k+1}^+(X) - \lambda_2^-(X)}{\lambda_{k+2}^-(X) - \lambda_{2}^-(X)} }$$ Theorem \[thm:main1\] is a generalization of Theorem \[thm:main\], and gives a bound on the robustness of feature vectors $F_X$ up to linear transformations. Note that applying an invertible linear transformation to all the points in a dataset (including training data) does not alter the classification problem for models invariant under linear transformations. For instance, if there exists a binary linear classifier given by vector $w$, such that $sign(w^TF_X(x))$ corresponds to the true label for point $x$, then the classifier given by $(M^{-1})^Tw$ assigns the correct label to linearly transformed feature vector $MF_X(x)$. Extending a Feature to New Points {#subsec:feature_extension} --------------------------------- In the previous section, we discussed how to get robust features for points in a dataset. In this section, we briefly describe an extension of that approach to get robust features for points outside the dataset, as in Definition \[def:rob\_dist\]. Let $X = \{x_1,\ldots,x_n \} \subset {\mathbb R}^d$ be the training dataset drawn from some underlying distribution $D$ over ${\mathbb R}^d$. We use $X$ as a reference to construct a robust function $f_X: {\mathbb R}^d \rightarrow {\mathbb R}$. For the sake of convenience, we drop the subscript $X$ from $f_X$ in the case where the dataset in question is clear. Given a point $x \in {\mathbb R}^d$, and a distance threshold parameter $T>0$, we define $f(x)$ as follows: Define $G(X)$ and $G(x)$ to be graphs whose nodes are points in dataset $X$, and $\{x\} \cup X= \{ x_0 = x, x_1, \dots, x_n\}$ respectively, and for which there is an edge between nodes $x_i$ and $x_j$, if $dist(x_i, x_j) \le T$. Let $L(x)$ be the Laplacian of $G(x)$, and let $\lambda_k(x)$ and $v_k(x)$ be its $k$th smallest eigenvalue and a corresponding unit eigenvector. Similarly, define $L(X)$, $\lambda_k(X)$ and $v_k(X)$ for $G(X)$. In all our constructions, we assume that the first eigenvectors $v_1(X)$ and $v_1(x)$ are set to be the unit vector proportional to the all-ones vector. Now define $f(x) = v_2\qty(x)_0$; i.e. the component of $v_2(x)$ corresponding to $x_0 = x$. Note that the eigenvector $v_2(x)$ has to be picked “consistently” to avoid signflips in $f$ as $-v_2(x)$ is also a valid eigenvector. To resolve this, we select the eigenvector $v_2(x)$ to be the eigenvector (with eigenvalue $\lambda_2(x)$) whose last $\|X\|$ entries has the maximum inner product with $v_2(X)$. We now state a robustness bound for this feature $f$ as per Definition \[def:rob\_dist\]. For a fixed $\eps > 0$ define the graph $G^+(x)$ to be the graph with the same nodes and edges of $G(x)$, except that the threshold for $x_0 = x$ is $T + \eps$ instead of $T$. Formally, in $G^+(x)$, there is an edge between $x_i$ and $x_j$ if: 1. $i = 0$ or $j=0$, and $dist(x_i, x_j) \le T + \eps$; or 2. $i > 0$ and $j>0$, and $dist(x_i, x_j) \le T$. Similarly, define $G^-(x)$ to be the same graph with $T + \eps$ replaced with $T - \eps$. Define $L^+, \lambda_k^+, v_k^+, L^-, \lambda_k^-, v_k^-$ analogously to the earlier definitions. In the following theorem, we give a robustness bound on the function $f$ as defined above. [theorem]{}[thmptwise]{}\[thm:ptwise\] For a sufficiently large training set size $n$, if $\operatorname{\mathbb E}_{X \sim D} \qty[\qty(\lambda_3(X) - \lambda_2(X))^{-1}] \leq c$ for some small enough constant $c$, then with probability $0.95$ over the choice of $X$, the function $f_X: {\mathbb R}^d \rightarrow {\mathbb R}$ satisfies $\Pr_{x \sim D}[\exists x' \text{ s.t. } dist(x,x')\le\eps \text{ and } \abs{f_X(x) - f_X(x')} \geq \delta_x] \leq 0.05 $, for $$\label{eq:ptwise} \delta_x = \qty(6\sqrt{2})\sqrt{ \frac{\lambda_2^+(x) - \lambda_2^-(x)}{\lambda_3^-(x) - \lambda_2^-(x)}}.$$ This also implies that with probability $0.95$ over the choice of $X$, $f_X$ is $(\epsilon, 20 \operatorname{\mathbb E}_{x \sim D}[\delta_x], 0.1)$ robust as per Definition \[def:rob\_dist\]. This bound is very similar to bound obtained in Theorem \[thm:main\], and says that the function $f$ is robust, as long as the eigengap between the second and third eigenvalues is sufficiently large for $G(X)$ and $G^-(x)$, and the second eigenvalue does not change significantly if we slightly perturb the distance threshold used to determine whether an edge exists in the graph in question. Similarly, one can also obtain a set of $k$ features, by taking the first $k$ eigenvectors of $G(X)$ prepended with zero, and projecting them onto the bottom-$k$ eigenspace of $G(x)$. A Lower Bound on Adversarial Robustness {#s:lower_bound} ======================================= In this section, we show that spectral properties yield a lower bound on the robustness of any function on a dataset. We show that if there exists an $(\eps,\delta)$ robust function $F'$ on dataset $X$, then the spectral approach (with appropriately chosen threshold), will yield an $(\eps', \delta')$ robust function, where the relationship between $\eps, \delta$ and $\eps', \delta'$ is governed by easily computable properties of the dataset, $X$. This immediately provides a way of establishing a bound on the best possible robustness that dataset $X$ could permit for perturbations of magnitude $\eps$. Furthermore it suggests that the spectral properties of the neighborhood graphs we consider, may be inherently related to the robustness that a dataset allows. We now formally state our lower bound: [theorem]{}[thmlower]{}\[thm:lower\] Assume that there exists some $(\eps,\delta)$ robust function $F^$ for the dataset $X$ (not necessarily constructed via the spectral approach). For any threshold $T$, let $G_T$ be the graph obtained on $X$ by thresholding at $T$. Let $d_T$ be the maximum degree of $G_T$. Then the feature $F$ returned by the spectral approach on the graph $G_{2\eps/3}$ is at least $(\eps/6,\delta')$ robust (up to sign), for $$\delta'= \delta\sqrt{\frac{8(d_{\eps}+1)}{\lambda_3(G_{\eps/3})-\lambda_2(G_{\eps/3})}}.$$ The bound gives reasonable guarantees when the degree is small and the spectral gap is large. To produce meaningful bounds, the neighborhood graph must have some structure at the threshold in question; in many practical settings, this would require an extremely large dataset, and hence this bound is mainly of theoretical interest at this point. Still, our experimental results in Section \[s:experiments\] empirically validate the hypothesis that spectral properties have implications for the robustness of any model: we show that the robustness of an adversarially trained neural network on different data distributions correlates with the spectral properties of the distribution. Synthetic Setting: Adversarial Spheres {#s:synthetic} ====================================== [@gilmer2018adversarial] devise a situation in which they are able to show in theory that training adversarially robust models is difficult. The authors describe the “concentric spheres dataset”, which consists of—as the name suggests—two concentric $d$-dimensional spheres, one of radius $1$ and one of radius $R > 1$. The authors then argue that any classifier that misclassifies even a small fraction of the inner sphere will have a significant drop in adversarial robustness. We argue that our method, in fact, yields a near-perfect classifier—one that makes almost no errors on natural or adversarial examples—even when trained on a modest amount of data. To see this, consider a sample of $2N$ training points from the dataset, $N$ from the inner sphere and $N$ from the outer sphere. Observe that the distance between two uniformly chosen points on a sphere of radius $r$ is close to $r \sqrt{2}$. In particular, the median distance between two such points is exactly $r \sqrt{2}$, and with high probability for large $d$, the distance will be within some small radius $\eps$ of $r \sqrt{2}$. Thus, for distance threshold $\sqrt{2}+2\eps$, after adding a new test point to the training data, we will get a graph with large clique corresponding to the inner sphere, and isolated points on the outer sphere, with high probability. This structure doesn’t change by perturbing the test point by $\epsilon$, resulting in a robust classifier. We now formalize this intuition. Let the inner sphere be of radius one, and outer sphere be of some constant radius $R > 1$. Let $\eps = (R-1)/8$ be the radius of possible perturbations. Then we can state the following: [theorem]{}[thmadvspheres]{}\[thm:advspheres\] Pick initial distance threshold $T = \sqrt{2} + 2\eps$ in the $\ell_2$ norm, and use the first $N+1$ eigenvectors as proposed in Section \[subsec:feature\_extension\] to construct a $(N+1)$-dimensional feature map $f : {\mathbb R}^d \to {\mathbb R}^{N+1}$. Then with probability at least $1-N^2 e^{-\Omega(d)}$ over the random choice of training set, $f$ maps the entire inner sphere to the same point, and the entire outer sphere to some other point, except for a $\gamma$-fraction of both spheres, where $\gamma = Ne^{-\Omega(d)}$. In particular, $f$ is $(\eps, 0, \gamma)$-robust. The extremely nice form of the constructed feature $f$ in this case means that, if we use half of the training set to get the feature map $f$, and the other half to train a linear classifier (or, indeed, any nontrivial model at all) trained on top of this feature, this will yield a near-perfect classifier even against adversarial attacks. The adversarial spheres example is a case in which our method allows us to make a robust classifier, but other common methods do not. For example, nearest-neighbors will fail at classifying the outer sphere (since points on the outer sphere are generally closer to points on the inner sphere than to other points on the outer sphere), and @gilmer2018adversarial demonstrate in practice that training adversarially robust models on the concentric spheres dataset using standard neural network architectures is extremely difficult when the dimension $d$ grows large. Experiments {#s:experiments} =========== Image Classification: The MNIST Dataset --------------------------------------- While the main focus of our work is to improve the conceptual understanding of adversarial robustness, we also perform experiments on the MNIST dataset. We test the efficacy of our features by evaluating them on the downstream task of classifying adversarial images. We used a subset of MNIST dataset, which is commonly used in discussions of adversarial examples [@goodfellow2014explaining; @szegedy2013intriguing; @madry2017towards]. Our dataset has 11,000 images of hand written digits from zero to nine, of which 10,000 images are used for training, and rest for test. We compare three different models, the specifics of which are given below: ![Performance on adversarial data vs our upper bound.[]{data-label="fig:correlation"}](transfer_nn){width="100.00000%"} ![Performance on adversarial data vs our upper bound.[]{data-label="fig:correlation"}](correlation1){width="100.00000%"} Robust neural network (pgd-nn): We consider a fully connected neural network with one hidden layer having 200 units, with ReLU non-linearity, and cross-entropy loss. We use PyTorch implementation of Adam [@kingma2014adam] for optimization with a step size of $0.001$. To obtain a robust neural network, we generate adversarial examples using projected gradient descent for each mini-batch, and train our model on these examples. For projected gradient descent, we use a step size of $0.1$ for $40$ iterations. Spectral features obtained using scaled Laplacian, and linear classifier (unweighted-laplacian-linear): We use the $\ell_2$ norm as a distance metric, and distance threshold $T = 9$ to construct a graph on all 11,000 data points. Since the distances between training points are highly irregular, our constructed graph is also highly irregular; thus, we use the scaled Laplacian to construct our features. Our features are obtained from the 20 eigenvectors corresponding to $\lambda_2$ to $\lambda_{21}$. Thus each image is mapped to a feature vector in ${\mathbb R}^{20}$. On top of these features, we use a linear classifier with cross-entropy loss for classification. We train the linear classifier using 10,000 images, and test it on 1,000 images obtained by adversarially perturbing test images. Spectral features obtained using scaled Laplacian with weighted edges, and linear classifier (weighted-laplacian-linear): This is similar to the previous model, with the only difference being the way in which the graph is constructed. Instead of using a fixed threshold, we have weighted edges between all pairs of images, with the weight on the edge between image $i$ and $j$ being $\exp(-0.1\norm{x_i - x_j}_2^2)$. As before, we use 20 eigenvectors corresponding to the scaled Laplacian of this graph, with a linear classifier for classification. Note that generating our features involve computing distances between all pair of images, followed by an eigenvector computation. Therefore, finding the gradient (with respect to the image coordinates) of classifiers built on top of these features is computationally extremely expensive. As previous works [@papernot2016transferability] have shown that transfer attacks can successfully fool many different models, we use transfer attacks using adversarial images corresponding to robust neural networks (pgd-nn). The performance of these models on adversarial data is shown in figure \[fig:nn\_comparison\]. We observe that weighted-laplacian-linear performs better than pgd-nn on large enough perturbations. Note that it is possible that robustly trained deep convolutional neural nets perform better than our model. It is also possible that the performance of our models may deteriorate with stronger attacks. Still, our conceptually simple features, with just a linear classifier on top, are able to give competitive results against reasonably strong adversaries. It is possible that training robust neural networks on top of these features, or using such features for the middle layers of neural nets may give significantly more robust models. Therefore, our experiments should be considered mainly as a proof of concept, indicating that spectral features may be a useful tool in one’s toolkit for adversarial robustness. We also observe that features from weighted graphs perform better than their unweighted counterpart. This is likely because the weighted graph contains more information about the distances, while most of this information is lost via thresholding in the unweighted graph. Connection Between Spectral Properties and Robustness ----------------------------------------------------- We hypothesize that the spectral properties of the graph associated with a dataset has fundamental connections with its adversarial robustness. The lower bound shown in section \[s:lower\_bound\] sheds some more light on this connection. In Theorem \[thm:main\], we show that adversarial robustness is proportional to $\sqrt{ (\lambda_2^+ - \lambda_2^-)/(\lambda_3^- - \lambda_2^-)}$. To study this connection empirically, we created 45 datasets corresponding to each pair of digits in MNIST. As we expect some pairs of digits to be less robust to adversarial perturbations than others, we compare our spectral bounds for these various datasets, to their observed adversarial accuracies. Setup:* The dataset for each pair of digits has 5000 data points, with 4000 points used as the training set, and 1000 points used as the test set. Similarly to the previous subsection, we trained robust neural nets on these datasets. We considered fully connected neural nets with one hidden layer having 50 units, with ReLU non-linearity, and cross-entropy loss. For each mini-batch, we generated adversarial examples using projected gradient descent with a step size of 0.2 for 20 iterations, and trained the neural net on these examples. Finally, to test this model, we generated adversarial perturbations of size $1$ in $\ell_2$ norm to obtain the adversarial accuracy for all 45 datasets. To get a bound for each dataset $X$, we generated two graphs $G^-(X)$, and $G^+(X)$ with all $5000$ points (not involving adversarial data). We use the $\ell_2$ norm as a distance metric. The distance threshold $T$ for $G^-(X)$ is set to be the smallest value such that each node has degree at least one, and the threshold for $G^+(X)$ is two more than that of $G^-(X)$. We calculated the eigenvalues of the scaled Laplacians of these graphs to obtain our theoretical bounds. Observations: As shown in Figure \[fig:correlation\], we observe some correlation between our upper bounds and the empirical adversarial robustness of the datasets. Each dataset is represented by a point in Figure \[fig:correlation\], where the x-axis is proportional to our bound, and the y-axis indicates the zero-one loss of the neural nets on adversarial examples generated from that dataset. The correlation is 0.52 after removing the right-most outlier. While this correlation is not too strong, it suggests some connection between our spectral bounds on the robustness and the empirical robustness of certain attack/defense heuristics. Conclusion ========== We considered the task of learning adversarially robust features as a simplification of the more common goal of learning adversarially robust classifiers. We showed that this task has a natural connection to spectral graph theory, and that spectral properties of a graph associated to the underlying data have implications for the robustness of any feature learned on the data. We believe that exploring this simpler task of learning robust features, and further developing the connections to spectral graph theory, are promising steps towards the end goal of building robust machine learning models. Acknowledgments: This work was supported by NSF awards CCF-1704417 and 1813049, and an ONR Young Investigator Award (N00014-18-1-2295). Proof of Theorems \[thm:main\] and \[thm:main1\] {#proof-of-theorems-thmmain-and-thmmain1 .unnumbered} ================================================ Before proving the theorems, it will be useful to prove a few helper lemmas that we use at various places. \[lem:subset\_eigvec\] Let $G$ and $G^{-}$ be two graphs with $V(G^{-}) = V(G)$, $E(G^{-}) \subseteq E(G)$. Let $L$ and $L^{-}$ be their respective Laplacians. Then $v^TL^{-}v \leq v^T L v$ for all vectors $v$. The proof follows from the definition of Laplacian: $$\begin{gathered} \begin{split} v^T L v &= \sum_{\{i,j\} \in E(G)} (v_i - v_j)^2\\ &= \sum_{\{i,j\} \in E(G^-)} (v_i - v_j)^2 + \sum_{\{i,j\} \in E(G)\setminus E(G^-)} (v_i - v_j)^2\\ &\geq \sum_{\{i,j\} \in E(G^-)} (v_i - v_j)^2\\ &= v^T L^{-} v \end{split}\end{gathered}$$ \[lem:subset\_eigval\] Let $G$ and $G^{-}$ be two graphs with $V(G^{-}) = V(G)$, $E(G^{-}) \subseteq E(G)$. Let $L$ and $L^{-}$ be their respective Laplacians, with $k$th eigenvalue $\lambda_k$ and $\lambda_k^-$ respectively. Then $\lambda_k \geq \lambda_k^-$. From the min-max interpretation of eigenvectors, we have $$\lambda_{k} = \min_{S_k} \max_{v \in S_k} v^T L v$$ where $S_k$ ranges over all $k$-dimensional subspaces. From lemma \[lem:subset\_eigvec\], we get $$v^T L^- v \leq v^T L V$$ for all $v \in S_k$, and for all k-dimensional subspaces $S_k$. This gives us $$\max_{v \in S_k} v^T L^- v \leq \max_{v \in S_k} v^T L V$$ for all k-dimensional subspaces $S_k$. Since this is true for all $S_k$, we get $$\min_{S_k} \max_{v \in S_k} v^T L^- v \leq \min_{S_k} \max_{v \in S_k} v^T L V$$ From the min-max interpretation, we get $\lambda_k \geq \lambda_k^-$ For all the proofs, unless otherwise stated, assume the first eigenvector (corresponding to the smallest eigenvalue) to be the all-ones vector. For cleanliness of notation, let $G(X) = G$, $L(X) = L$, $\lambda_k(X) = \lambda_k, G^-(X) = G^-$, etc. Observe that $$\lambda_{2}^+ \ge \lambda_{2}(X') = v_{2}(X')^{T}L(X') v_{2}(X') \ge v_{2}(X')^{T}L^- v_{2}(X')$$ The first part of the inequality follows from lemma \[lem:subset\_eigval\], and second part follows from lemma \[lem:subset\_eigvec\]. Now write $$v_{2}(X') = \alpha v_{2}^- + \beta v_\perp$$ where $v_\perp$ is a unit vector perpendicular to $v_2^-$ and all-ones vector, and $\alpha$ and $\beta$ are scalars. Then $$\begin{aligned} \lambda_2^+ &\ge v_{2}(X')L^- v_{2}(X') \\&= \qty(\alpha v_{2}^- + \beta v_\perp)^{T}L^- \qty(\alpha v_{2}^- + \beta v_\perp) \\&= \alpha^2 v_2^{-{T}}L^- v_2^- + \beta^2 v_\perp^{T}L^- v_\perp^{T}\\&\ge \alpha^2 \lambda_2^- + \beta^2\lambda_{3}^-\end{aligned}$$ using the property that the $v_2^-$ and $v_\perp$ are orthogonal, and $v_2^-$ is an eigenvector of $L^-$. By rearranging, and observing $\alpha^2 + \beta^2 = 1$ (as $v_2(X')$ is a unit vector), we get $$\begin{aligned} \beta^2 (\lambda_{3}^- - \lambda_{2}^-) &\le \lambda_2^+ - \alpha^2 \lambda_2^- - \beta^2 \lambda_{2}^- = \lambda_2^+ - \lambda_2^- \\ \beta^2 &\le \frac{\lambda_2^+ - \lambda_2^-}{\lambda_{3}^- - \lambda_{2}^-}\end{aligned}$$ As $\abs{\alpha} = \sqrt{1 - \beta^2}$, we get $$\begin{gathered} \begin{split} \text{min}(\norm{v_{2}(X') - v_2^-}_2^2, \norm{-v_{2}(X') - v_2^-}_2^2 )&= \text{min}(2(1 - v_2(X')^T v_2^-), 2(1 + v_2(X')^T v_2^-))\\ &= 2(1 - \abs{\alpha})\\ &= 2\qty(1 - \sqrt{1 - \beta^2})\\ &\le 2\beta^2. \end{split}\end{gathered}$$ This gives us $$\text{min}(\norm{v_{2}(X') - v_2^-}_2, \norm{-v_{2}(X') - v_2^-}_2 ) \le \sqrt{2} \abs{\beta} \leq \sqrt{2} \ \sqrt{\frac{\lambda_2^+ - \lambda_2^-}{\lambda_{3}^- - \lambda_{2}^-}}$$ Notice that the above argument holds for any $X'$ such that $dist(X, X')\leq \eps$; in particular, it holds for $X' = X$. This gives us $$\text{min}(\norm{v_{2}(X) - v_2^-}_2, \norm{-v_{2}(X) - v_2^-}_2 ) \leq \sqrt{2} \ \sqrt{\frac{\lambda_2^+ - \lambda_2^-}{\lambda_{3}^- - \lambda_{2}^-}}.$$ Using triangle inequality, we get $$\text{min}(\norm{v_{2}(X) - v_2(X')}_2, \norm{v_{2}(X) - (-v_2(X'))}_2 ) \le 2\sqrt{2} \ \sqrt{\frac{\lambda_2^+ - \lambda_2^-}{\lambda_{3}^- - \lambda_{2}^-}}.$$ This finishes the proof as $F(X) = v_2(X)$ and $F(X')= v_2(X')$. This proof generalizes the proof of theorem \[thm:main\]. For cleanliness of notation, let $G(X) = G$, $L(X) = L$, $\lambda_k(X) = \lambda_k, G^-(X) = G^-$, etc. Let $v(X')$ be any unit vector in $S_k(X') = Span(v_2(X'), \dots, v_{k+1}(X'))$. We will prove a bound on distance of $v(X')$ from its closest unit vector in $S_k^- = Span(v_2^-, \dots, v_{k+1}^-)$. Write $$v(X') = \sum_{i=2}^{k+1} \alpha_i v_{i}^- + \beta v_\perp$$ where $v_\perp$ is a unit vector satisfying $v_\perp \perp v_i^-(x)$ for all $1 \le i \le k+1$, and $\alpha_i$ and $\beta$ are scalars. By lemma $\ref{lem:subset_eigval}$, we get $$\label{eqn:triangleineq} \lambda_{k+1}^+ \ge \lambda_{k+1}(X') = v_{k+1}(X')^{T}L(X') v_{k+1}(X')$$ and by lemma $\ref{lem:subset_eigvec}$ and by the definition of eigenvectors, we get $$\label{eqn:triangleineq} v_{k+1}(X')^{T}L(X') v_{k+1}(X') \ge v(X')^{T}L(X') v(X') \ge v(X')^{T}L^- v(X')$$ The first part of the inequality follows from lemma \[lem:subset\_eigval\], and second part follows from lemma \[lem:subset\_eigvec\]. Combining the two inequalities, we get $$\begin{aligned} \lambda_{k+1}^+ &\ge v(X')^{T}L^- v(X') \\&= \qty(\sum_{i=2}^{k+1} \alpha_i v_{i}^- + \beta v_\perp)^{T}L^- \qty(\sum_{i=2}^{k+1} \alpha_i v_{i}^- + \beta v_\perp) \\&= \sum_{i=2}^{k+1} \alpha_i^2 v_i^{-{T}}L^- v_i^- + \beta^2 v_\perp^{T}L^- v_\perp^{T}\\&\ge \sum_{i=2}^{k+1} \alpha_i^2 \lambda_i^- + \beta^2\lambda_{k+2}^-\end{aligned}$$ using the property that the $v_i^-$ and $v_\perp$ are all mutually orthogonal. Rearranging: $$\begin{aligned} \beta^2 (\lambda_{k+2}^- - \lambda_{2}^-) &\le \lambda_{k+1}^+ - \sum_{i=2}^{k+1} \alpha_i^2 \lambda_i^- - \beta^2 \lambda_{2}^- \\&\le \lambda_{k+1}^+ - \lambda_2^-\end{aligned}$$ which implies $$\beta^2 \le \frac{\lambda_{k+1}^+ - \lambda_2^-}{\lambda_{k+2}^- - \lambda_{2}^-}$$ For simplicity of notation, let $$\alpha = (\alpha_2, \dots, \alpha_{k+1}) \in {\mathbb R}^k,\qq{} v^- = \frac{\sum_{i=2}^{k+1} \alpha_i v_{i}^-}{\norm{\alpha}}$$ Then $$\begin{aligned} \label{eq:enough_for_thm1} \norm{v(X') - v^-}_2^2 &= 2\qty(1 - v(X')^T v^-) = 2(1 - \norm{\alpha}) = 2\qty(1 - \sqrt{1 - \beta^2}) \le 2\beta^2 \le 2 {\frac{\lambda_{k+1}^+ - \lambda_2^-}{\lambda_{k+2}^- - \lambda_{2}^-}}\end{aligned}$$ This shows that every unit vector in $S_k(X')$ is within $\sqrt{2} \ \sqrt{\frac{\lambda_{k+1}^+ - \lambda_2^-}{\lambda_{k+2}^- - \lambda_{2}^-}}$ of some unit vector in $S_k^-$, and, by symmetry, vice-versa. Notice that the above argument holds for any $X'$ such that $dist(X, X')\leq \eps$; in particular, it holds for $X' = X$. It thus follows from triangle inequality that every unit vector in $S_k(X')$ must be within $2 \sqrt{2} \ \sqrt{\frac{\lambda_{k+1}^+ - \lambda_2^-}{\lambda_{k+2}^- - \lambda_{2}^-}}$ of some unit vector $S_k(X)$. Let $F(X) = (F^1_X, F^2_X, \dots, F^k_X)$ be an $n \times k$ matrix. Define $F(X')$ similarly. Observe that $$\sqrt{\sum_{i \in [n]} \norm{M F_X(x_i) - F_{X'}(x_i')}^2} = \norm{F(X) M^T - F(X')} $$ Now, to prove the theorem we need to show the existence of an invertible matrix $M$ such that $$\norm{F(X) M^T - F(X')} \leq 2 \sqrt{2k} \ \sqrt{\frac{\lambda_{k+1}^+ - \lambda_2^-}{\lambda_{k+2}^- - \lambda_{2}^-}} $$ When $2 \sqrt{2} \ \sqrt{\frac{\lambda_{k+1}^+ - \lambda_2^-}{\lambda_{k+2}^- - \lambda_{2}^-}} \geq \sqrt{2}$, the desired bound is trivially true. To see this, set $M$ to be a diagonal matrix with diagonal entries $\pm 1$ such that $\langle M_{ii}F^i_X, F^i_{X'} \rangle \geq 0$ for all $i$. Since $F^i_X$ and $F^i_{X'}$ are unit vectors, we get $$\begin{gathered} \begin{split} \norm{F(X) M^T - F(X')} &= \sqrt{\sum_{i \in [k]} \norm{M_{ii}F^i_X - F^i_{X'}}^2}\\ &\leq \sqrt{2k}\\ &\leq 2 \sqrt{2k} \ \sqrt{\frac{\lambda_{k+1}^+ - \lambda_2^-}{\lambda_{k+2}^- - \lambda_{2}^-}} \end{split}\end{gathered}$$ where $\norm{\cdot}$ denotes the Frobenius norm for matrices, and $\ell_2$ norm for vectors. So now assume $2 \sqrt{2} \ \sqrt{\frac{\lambda_{k+1}^+ - \lambda_2^-}{\lambda_{k+2}^- - \lambda_{2}^-}} < \sqrt{2}$. Let $P$ be a projection map onto subspace $S_k(X)$. As $F^1_{X'}, \dots, F^k_{X'}$ form an orthonormal basis for $S_k(X')$, projecting them onto $S_k(X)$ must yield a basis for $S_k(X)$; if not, then we would have $Pu = 0$ for some unit vector $u \in S_k(X')$, so that $u$ would be orthogonal (i.e. have distance $\sqrt{2}$) to every unit vector in $S_k(X)$, which contradicts the assumption that $2 \sqrt{2} \ \sqrt{\frac{\lambda_{k+1}^+ - \lambda_2^-}{\lambda_{k+2}^- - \lambda_{2}^-}} < \sqrt{2}$. Now let $M \in {\mathbb R}^{k \times k}$ be an invertible matrix; such that $M^T$ corresponds to the change of basis matrix satisfying $PF(X') = F(X)M^T$ (this must exist since $PF(X')$ and $F(X)$ are both bases of $S_k(X)$).Then $$\norm{F(X) M^T - F(X')} = \norm{PF(X') - F(X')} \le 2 \sqrt{2k} \ \sqrt{\frac{\lambda_{k+1}^+ - \lambda_2^-}{\lambda_{k+2}^- - \lambda_{2}^-}}$$ where last inequality follows since for each column vector of $F(X')$, its projection onto $S_k(X)$ has $\ell_2$ distance of at most $2\sqrt{2} \ \sqrt{\frac{\lambda_{k+1}^+ - \lambda_2^-}{\lambda_{k+2}^- - \lambda_{2}^-}}$ from it. This is because for each unit vector in $S_k(X')$, there is a vector in $S_k(X)$ at a distance of at most $2\sqrt{2} \ \sqrt{\frac{\lambda_{k+1}^+ - \lambda_2^-}{\lambda_{k+2}^- - \lambda_{2}^-}}$ from it. This concludes the proof. Proof of Theorem \[thm:ptwise\] {#proof-of-theorem-thmptwise .unnumbered} =============================== We give the proof of this theorem as a series of lemmas. First, similar to theorem 1, we show the robustness of function $f$ up to sign. \[lem:f\_rob\_sign\_flip\] Given two points $x$ and $x'$, such that $dist(x, x') \le \eps$, the function $f: {\mathbb R}^d \rightarrow {\mathbb R}$ defined in the main text of the paper satisfies $$\label{eq:f_rob_sign_flip} \min(\abs{f(x) - f(x')},\abs{f(x) - (-f(x'))}) \le \qty(2\sqrt{2})\sqrt{ \frac{\lambda_2^+(x) - \lambda_2^-(x)}{\lambda_3^-(x) - \lambda_2^-(x)} }.$$ The proof follows from the same argument as Theorems \[thm:main\] and \[thm:main1\] above. For completeness, we include it again here. For cleanliness of notation, let $G(x) = G$, $L(x) = L$, $\lambda_k(x) = \lambda_k, G^-(x) = G^-$, etc. Observe that $$\lambda_{2}^+ \ge \lambda_{2}(x') = v_{2}(x')^{T}L(x') v_{2}(x') \ge v_{2}(x')^{T}L^- v_{2}(x')$$ The first part of the inequality follows from lemma \[lem:subset\_eigval\], and second part follows from lemma \[lem:subset\_eigvec\]. Now write $$v_{2}(x') = \alpha v_{2}^- + \beta v_\perp$$ where $v_\perp$ is a unit vector perpendicular to $v_2^-$ and all-ones vector, and $\alpha$ and $\beta$ are scalars. Then $$\begin{aligned} \lambda_2^+ &\ge v_{2}(x')L^- v_{2}(x') \\&= \qty(\alpha v_{2}^- + \beta v_\perp)^{T}L^- \qty(\alpha v_{2}^- + \beta v_\perp) \\&= \alpha^2 v_2^{-{T}}L^- v_2^- + \beta^2 v_\perp^{T}L^- v_\perp^{T}\\&\ge \alpha^2 \lambda_2^- + \beta^2\lambda_{3}^-\end{aligned}$$ using the property that the $v_2^-$ and $v_\perp$ are orthogonal, and $v_2^-$ is an eigenvector of $L^-$. By rearranging, and observing $\alpha^2 + \beta^2 = 1$ (as $v_2(x')$ is a unit vector), we get $$\begin{aligned} \beta^2 (\lambda_{3}^- - \lambda_{2}^-) &\le \lambda_2^+ - \alpha^2 \lambda_2^- - \beta^2 \lambda_{2}^- = \lambda_2^+ - \lambda_2^- \\ \beta^2 &\le \frac{\lambda_2^+ - \lambda_2^-}{\lambda_{3}^- - \lambda_{2}^-}\end{aligned}$$ As $\abs{\alpha} = \sqrt{1 - \beta^2}$, we get $$\begin{gathered} \begin{split} \text{min}(\norm{v_{2}(x') - v_2^-}_2^2, \norm{-v_{2}(x') - v_2^-}_2^2 )&= \text{min}(2(1 - v_2(x')^T v_2^-), 2(1 + v_2(x')^T v_2^-))\\ &= 2(1 - \abs{\alpha})\\ &= 2\qty(1 - \sqrt{1 - \beta^2})\\ &\le 2\beta^2. \end{split}\end{gathered}$$ This gives us $$\text{min}(\norm{v_{2}(x') - v_2^-}_2, \norm{-v_{2}(x') - v_2^-}_2 ) \le \sqrt{2} \abs{\beta} \leq \sqrt{2} \ \sqrt{\frac{\lambda_2^+ - \lambda_2^-}{\lambda_{3}^- - \lambda_{2}^-}}$$ Notice that the above argument holds for any $x'$ such that $dist(x, x')\leq \eps$; in particular, it holds for $x' = x$. This gives us $$\text{min}(\norm{v_{2}(x) - v_2^-}_2, \norm{-v_{2}(x) - v_2^-}_2 ) \leq \sqrt{2} \ \sqrt{\frac{\lambda_2^+ - \lambda_2^-}{\lambda_{3}^- - \lambda_{2}^-}}.$$ Using triangle inequality, we get $$\text{min}(\norm{v_{2}(x) - v_2(x')}_2, \norm{v_{2}(x) - (-v_2(x'))}_2 ) \le 2\sqrt{2} \ \sqrt{\frac{\lambda_2^+ - \lambda_2^-}{\lambda_{3}^- - \lambda_{2}^-}}.$$ Thus, we conclude $$\begin{aligned} &\min\qty(\abs{f(x) - f(x')},\abs{f(x) - (-f(x'))}) \\&\le \min\qty(\norm{v_2(x) - v_2(x')}, \norm{v_2(x) - (-v_2(x'))}) \\&\le 2\sqrt{2} \ \sqrt{\frac{\lambda_2^+ - \lambda_2^-}{\lambda_{3}^- - \lambda_{2}^-}}\end{aligned}$$ as desired. While flipping signs for the whole dataset is fine, we don’t want the features of some of the points to flip signs arbitrarily. As described in the main text, to resolve this, we select the eigenvector $v_2(x)$ to be the eigenvector (with eigenvalue $\lambda_2(x)$) whose last $\|X\|$ entries have the maximum inner product with $v_2(X)$. We show a bound on this inner product next. Let $v_2^(x)$ be a $\|X\|$ dimensional vector obtained by chopping off the first entry of $v_2(x)$. \[lem:innerprod\_train\] For $v_2^(x)$ defined as above, we have $$\label{eq:f_train_dist} \ev{ v_2^(x), v_2(X) } \geq \sqrt{1 - \frac{\lambda_2(x) - \lambda_2(X)\qty(1 - v_2(x)_0^2\frac{n+1}{n})}{\lambda_3(X) - \lambda_2(X)} - v_2(x)_0^2 \frac{n+1}{n}}$$ For cleanliness of notation, let $G(x) = G, L(x) = L, \lambda_k(x) = \lambda_k, v_2^(x) = v_2^$, etc. Write $v_2^ = \alpha v_2(X) + \beta w + \gamma \mathbf 1 / \sqrt{n}$ for scalars $\alpha, \beta, \gamma$ and vector $w$ orthogonal to $v_2(X)$ and $\mathbf 1$. Taking inner product of both sides with $\mathbf 1$, we get $$\langle v_2^, \mathbf 1 \rangle = \alpha \langle v_2(X) , \mathbf 1 \rangle + \beta \langle w , \mathbf 1 \rangle + \gamma \langle \mathbf 1 / \sqrt{n} , \mathbf 1 \rangle$$ As $\langle v_2(X) , \mathbf 1 \rangle =0$, $\langle w , \mathbf 1 \rangle =0$, and $\langle v_2^, \mathbf 1 \rangle = -v_2(x)_0$, this gives us $\gamma = -v_2(x)_0 / \sqrt{n}$. Now, as $w$ is orthogonal to the bottom two eigenvectors of $L(X)$, we get $w^T L(X) w \ge \lambda_3(X)$. Then $$\begin{aligned} \lambda_2 = v_2^T L v_2 &\ge v_2^{T} L(X) v_2^ \\&\ge \alpha^2 \lambda_2(X) + \beta^2 \lambda_3(X) \\&= \qty(1 - \gamma^2 - v_2(x)_0^2) \lambda_2(X) + \beta^2 \qty( \lambda_3(X) - \lambda_2(X))\end{aligned}$$ Putting $\gamma = -v_2(x)_0 / \sqrt{n}$. and rearranging, we get $$\begin{aligned} \label{eq:lem2_beta_bound} \beta^2 \le \frac{\lambda_2 - \lambda_2(X) \qty(1 - v_2(x)_0^2 \frac{n+1}{n}) }{\lambda_3(X) - \lambda_2(X)}\end{aligned}$$ Notice now that $\alpha^2 + \beta^2 + \gamma^2 = \norm{v_2^}^2 = 1 - v_2(x)_0^2$. Since we know $\gamma^2 + v_2(x)_0^2 = v_2(x)_0^2\frac{n+1}{n}$, and we know $\alpha \ge 0$ by definition, we conclude $$\ev{v_2^(x), v_2(X)} = \alpha = \sqrt{1 - \beta^2 - v_2(x)_0^2\frac{n+1}{n}}$$ whereupon substituting the bound on $\beta^2$ from Eqn. gives the desired result. Next, we give robustness bound on $f$ when $\langle v_2^(x), v_2(X) \rangle > \frac{1}{\sqrt{2}}$ \[lem:f\_rob\] Given two points $x$ and $x'$, such that $dist(x, x') \le \eps$, if $\langle v_2^(x), v_2(X) \rangle > \frac{1}{\sqrt{2}}$ the function $f_X: {\mathbb R}^d \rightarrow {\mathbb R}$ defined in the main text of the paper satisfies $$\abs{f_X(x) - f_X(x')} \le \qty(6\sqrt{2})\sqrt{ \frac{\lambda_2^+(x) - \lambda_2^-(x)}{\lambda_3^-(x) - \lambda_2^-(x)} }.$$ If $\ev{ v_2(x), v_2(x')} \geq 0$, then from the proof of Lemma \[lem:f\_rob\_sign\_flip\], we get $$\begin{aligned} \abs{f_X(x) - f_X(x')} &\le \norm{v_2(x) - v_2(x')} \\ &= \min\qty(\norm{v_2(x) - v_2(x')}, \norm{v_2(x) - (-v_2(x'))}) \\&\le 2\sqrt{2} \ \sqrt{\frac{\lambda_2^+ - \lambda_2^-}{\lambda_{3}^- - \lambda_{2}^-}}\end{aligned}$$ and we are done. Otherwise, let $v$ be the vector $v_2(X)$ with a zero prepended to it. Note that $\langle v_2(x), v \rangle = \langle v_2^(x), v_2(X) \rangle > \frac{1}{\sqrt{2}}$. Also, let $w$ be a unit vector in the direction of projection of $v$ on the subspace spanned by $v_2(x)$ and $v_2(x')$. And $w_\perp$ be a vector orthogonal to $w$ such that $v = \alpha w + \beta w_\perp$. This gives us $\langle v_2(x'), w \rangle = \frac{1}{\alpha} \langle v_2(x'), v \rangle \geq \langle v_2(x'), v \rangle \geq 0$ as $0< \alpha < 1$. Similarly, $\langle v_2(x), w \rangle > \frac{1}{\sqrt{2}}$. For three unit vectors $w, v_2(x), v_2(x')$ lying in a two dimensional subspace such that $\langle v_2(x), w \rangle > \frac{1}{\sqrt{2}}$ and $\langle v_2(x'), w \rangle \geq 0$, if $\langle v_2(x), v_2(x') \rangle < 0$, we get $\langle v_2(x), v_2(x') \rangle \geq \frac{-1}{\sqrt{2}}$. This implies $\langle v_2(x), -v_2(x') \rangle \leq \frac{1}{\sqrt{2}}$. From these inner product values, we get $\norm{v_2(x) - v_2(x')} \leq 1.85$, and $\norm{v_2(x) - (-v_2(x'))} \geq 0.75$. This gives us $$\norm{v_2(x) -v_2(x')} \leq 3 \min(\norm{v_2(x) - (-v_2(x'))}, \norm{v_2(x) - v_2(x')}) \le 3 \cdot 2\sqrt{2} \ \sqrt{\frac{\lambda_2^+ - \lambda_2^-}{\lambda_{3}^- - \lambda_{2}^-}},$$ where the second inequality now follows from the proof of Lemma \[lem:f\_rob\_sign\_flip\]. As $\|f_X(x) - f_X(x')\| \leq \norm{v_2(x) -v_2(x')} $, this finishes the proof. Next, we show that under the conditions mentioned in our theorem, $\langle v_2^(x), v_2(X) \rangle \geq \frac{1}{\sqrt{2}}$, for most $x \sim D$. \[lem:inner\_product\_big\] For a sufficiently large training set size $n$, if $\operatorname{\mathbb E}_{X \sim D} [\frac{1}{\lambda_3(X) - \lambda_2(X)}] \leq c$ for some small enough constant $c$, then with probability $0.95$ over the choice of $X$, $$\label{eq:inner_product_big} \Pr_{x \sim D}\qty[\langle v_2^(x), v_2(X) \rangle \geq \frac{1}{\sqrt{2}}] \geq 0.95$$ From lemma \[lem:innerprod\_train\], we know $$\label{eq:f_train_dist} \ev{ v_2^(x), v_2(X) } \geq \sqrt{1 - \frac{\lambda_2(x) - \lambda_2(X)\qty(1 - v_2(x)_0^2\frac{n+1}{n})}{\lambda_3(X) - \lambda_2(X)} - v_2(x)_0^2 \frac{n+1}{n}}$$ For $n$ large enough, $\frac{n+1}{n} \approx 1$, so we need to show $$\begin{aligned} \frac{\lambda_2(x) - \lambda_2(X)\qty(1 - v_2(x)_0^2)}{\lambda_3(X) - \lambda_2(X)} + v_2(x)_0^2 \leq \frac{1}{2}\end{aligned}$$ with probability 0.95 over the choice of $X$, and 0.95 over $x$. By Markov’s inequality, we get $$\begin{aligned} Pr_{X \sim D}[\lambda_3(X) - \lambda_2(X) \leq \frac{1}{100c}] \leq 0.01.\end{aligned}$$ For any size $n$ unit vector, if we pick one of its coordinates uniformly at random, it’s expected squared value is $\frac{1}{n}$. By this argument, we get $$\begin{aligned} \operatorname{\mathbb E}_{X \sim D, x \sim D}[v_2(x)_0^2] = \frac{1}{n+1}.\end{aligned}$$ By Markov’s inequality, we get that $$\begin{aligned} Pr_{X \sim D, x \sim D}[v_2(x)_0^2 \geq \frac{1000}{n+1}] \leq 0.001.\end{aligned}$$ This implies that with probability $0.98$ over the choice of $X$, $$\begin{aligned} Pr_{x \sim D}[v_2(x)_0^2 \geq \frac{1000}{n+1}] \leq 0.05.\end{aligned}$$ Also, note that $\lambda_2(x) \leq \lambda_2(X) + 1$, since the second eigenvalue of a graph can go up by at most one by adding a new vertex. We also have $\lambda_2(X) \leq n$ since the eigenvalues of an un-normalized Laplacian are bounded by the number of vertices. Putting all this together, and applying union bound, we get that with probability 0.97 over $X$, and $0.95$ over x, $$\begin{aligned} \frac{\lambda_2(x) - \lambda_2(X)\qty(1 - v_2(x)_0^2)}{\lambda_3(X) - \lambda_2(X)} + v_2(x)_0^2 \leq \frac{1+n(\frac{1000}{n+1})}{\frac{1}{100c}} + \frac{1000}{n}\end{aligned}$$ which is less than $\frac{1}{2}$ for small enough constant $c$, and $n$ large enough. Combining Lemmas \[lem:f\_rob\] and \[lem:inner\_product\_big\], we get that under the conditions stated in the theorem, with probability at least 0.95 over the choice of $X$ $$\begin{gathered} \label{eq:f_rob} \Pr_{x \sim D}\qty[\exists x' \text{ s.t. } dist(x,x')\le\eps \text{ and } \abs{f_X(x) - f_X(x')} \ge \delta_x] \leq 0.05 \\ \text{for } \delta_x = \qty(6\sqrt{2})\sqrt{ \frac{\lambda_2^+(x) - \lambda_2^-(x)}{\lambda_3^-(x) - \lambda_2^-(x)}}.\end{gathered}$$ which finishes the proof. By applying Markov inequality and a union bound, we also get that $f_X$ as defined above is $(\epsilon, 20 \operatorname{\mathbb E}_{x \sim D}[\delta_x], 0.1)$ robust with probability 0.95 over the choice of $X$. Proof of Theorem \[thm:lower\] {#proof-of-theorem-thmlower .unnumbered} ============================== Consider the graph $G_{2\eps/3}$ obtained by setting the threshold $T=2\eps/3$ on the dataset $X$. Let $G_{\eps}$ be the graph obtained by setting the threshold $T=\eps$ on the dataset $X$, and let $G_{\eps/3}$ be the graph obtained by setting the threshold $T=\eps/3$ on the dataset $X$. Note that if all datapoints in $X$ are perturbed by at most $\eps/6$ to get $X'$, then the inter point distances after perturbation are within $\eps/3$ of the original inter point distances. Hence by Theorem \[thm:main\], $$\min(\norm{F(X) - F(X')}, \norm{(-F(X)) - (F(X'))})\le 2\sqrt{2}\sqrt{\frac{\lambda_2(G_{\eps})-\lambda_2(G_{\eps/3})}{\lambda_3(G_{\eps/3})-\lambda_2(G_{\eps/3})}}.$$ As $\lambda_2(G_{\eps})-\lambda_2(G_{\eps/3})\le \lambda_2(G_{\eps})$, we will upper bound $\lambda_2(G_{\eps})$ to bound $\delta'$. Let $v$ be the vector such that the $i$th entry $v_i$ is $F^_{X}(x_i)$, the feature assigned by $F^$ to the datapoint $x_i$. Note that $\lambda_2(G_{\eps})\le \sum_{(i,j)\in G_{\eps}}^{}(v_i-v_j)^2$, by using $v$ as a candidate eigenvector. We claim that $\lambda_2(G_{\eps}) \le (d_{\eps}+1)\delta^2$. To show this, we partition all edges in $G_{\eps}$ into $t$ matchings $\{M_k,k\in [t]\}$. Note that for any matching $M_k$ , $\sum_{(i,j)\in M_k}^{}(v_i-v_j)^2\le \delta^2$. This follows by constructing the adversarial dataset $X'$ where each datapoint has been replaced by its matched vertex (if any) in the matching $M_k$, and by the fact that $F^$ is $(\eps,\delta)$ robust. By Vizing’s Theorem, the number of matchings $t$ required is at most $d_{\eps}+1$. Therefore $\lambda_2(G_{\eps}) \le (d_{\eps}+1)\delta^2$ and the theorem follows. Proof of Theorem \[thm:advspheres\] {#proof-of-theorem-thmadvspheres .unnumbered} =================================== We will explain the construction of the features $f$ in detail at the end of the proof, as they become relevant. We give the proof of this theorem as a series of lemmas. We use $S^{d-1}$ denote the unit sphere in ${\mathbb R}^d$ centered on the origin, and $rS^{d-1}$ denotes the sphere of radius $r$ centered on the origin. $\norm{\cdot}$ denotes the $\ell_2$ norm. \[lem:distance\] Let $A$ be any point on $r_1S^{d-1}$ and $B$ be a point chosen uniformly at random from $r_2S^{d-1}$. Then the median distance $\norm{A-B}$ is $M = \sqrt{r_1^2 + r_2^2}$. Further, for any fixed $\eps > 0$, we have $$\begin{aligned} \Pr[\abs{\norm{x-y}^2 - \qty(r_1^2 + r_2^2)} > \eps] \le \exp(-\Omega(d))+ \exp(-\Omega\qty(\frac{\eps^2 d}{r_1r_2})).\end{aligned}$$ For notational simplicity let $A = r_1 x$ and $B = r_2 y$ for $x, y \in S^{d-1}$. Assume WLOG that $x = (1, 0, \dots, 0)$. Suppose $y$ is drawn by drawing $z \sim N(0, I_d)$ and computing $y = z / \norm{z}$. Then $$\begin{aligned} \norm{A-B}^2 = \norm{r_1x - r_2y}^2 = r_1^2 + r_2^2 - 2r_1r_2 y_1.\end{aligned}$$ This immediately gives that the median value of $\norm{A-B}$ is $M = \sqrt{r_1^2 + r_2^2}$. Further, $$\begin{aligned} \abs{\norm{x-y}^2 - M^2} = 2r_1r_2 \abs{y_1} = 2r_1r_2\frac{\abs{z_1}}{\norm{z}}\end{aligned}$$ Let $z = (z_1, z_2, \dots, z_d)$, and consider $z' = (z_2, \dots, z_d) \in {\mathbb R}^{d-1}$. Then $\norm{z'}^2$, by definition, is a chi-square distributed random variable with $(d-1)$ degrees of freedom that is independent of $z_1$. The following Chernoff bounds for chi-square variables applies: $$\begin{aligned} \Pr[\norm{z'}^2 <{\frac{d-1}{2}}] &\le \qty(\frac{e}{4})^{(d-1)/4} = e^{-\Omega(d)} \\ \Pr[\norm{z'}^2 > 2(d-1)] &\le \qty(\frac{e}{4})^{(d-1)/4} = e^{-\Omega(d)}\end{aligned}$$ Thus, for any fixed $\eps,r_1, r_2$, we have $$\begin{aligned} \Pr[\abs{\norm{x-y}^2 - M^2} > \eps] &= \Pr[2r_1r_2\frac{\abs{z_1}}{\norm{z}} > \eps] \\&\le \Pr[\abs{z_1} > \frac{\eps \norm{z'}}{2r_1r_2}] \\&\le \Pr[\norm{z'}^2 < \frac{d-1}2] + \Pr[\abs{z_1} >\frac{\eps \norm{z'}}{2r_1r_2}\Bigm\vert \norm{z'}^2 \ge \frac{d-1}{2}] \\&\le e^{-\Omega(d)} + \Pr[z_1^2 > \frac{\eps^2 (d-1)}{8r_1r_2}] \\&\le e^{-\Omega(d)}+ \exp(-\Omega\qty(\frac{\eps^2 d}{r_1r_2})) \end{aligned}$$ using independence of $z_1$ and $z'$, and Gaussian tail bounds on $z_1$. Now let $x_1, \dots, x_{2N}$ be a training set, where $x_1, \dots, x_N$ are sampled i.i.d. uniform from $S^{d-1}$ and $x_{N+1}, \dots, x_{2N}$ are sampled i.i.d. uniform from $RS^{d-1}$. With probability at least $1 - O(N^2)e^{-\Omega(d)}$, both of these things hold: 1. For every pair $(x_i, x_j)$ of points on the inner sphere, we have $\norm{x_i - x_j} \le \sqrt{2}+\eps$ 2. For every pair $(x_i, x_j)$ of points at least one of which is on the outer sphere, we have $\norm{x_i - x_j} > \sqrt{2} + 3\eps$ By Lemma \[lem:distance\], for any constant $R > 1$ and $\eps = (R-1)/8$, the probability that any given pair of points $x_i, x_j$ for $i < j \le 2N$ satisfies our condition is $1 - e^{-\Omega(d)}$ (since $\sqrt{2} + 3\eps < \sqrt{1+R^2} - \eps$ for $\eps = (R-1)/8$). Thus, the probability that all pairs satisfy our condition is $1 - O(N^2) e^{-\Omega(d)}$ as desired. Thus with probability $1 - O(N^2)e^{-\Omega(d)}$, the graph that we construct at distance threshold $T = \sqrt{2} + 2\eps$ (with only the training points and no test point) has a very particular structure: one large connected component consisting of all the points on the inner sphere, and $N$ isolated points, one for each point on the outer sphere. \[lem:inner\] Let $x$ be a randomly drawn test point on the inner sphere. Then with probability at least $1 - O(N) e^{-\Omega(d)}$ over the choice of $x$, there is no $x'$ such that $\norm{x - x'} \le \eps$, and either (1) $\norm{x_i - x'} > \sqrt{2} + 2\eps$ for some $i \le N$, or (2) $\norm{x_i - x'} \le \sqrt{2} + 2\eps$ for some $i > N$. Certainly no such $i$ exists if $\norm{x_i - x} \le \sqrt{2} + \eps$ for all $i \le N$ and $\norm{x_i - x} > \sqrt{2} + 3\eps$ for all $i > N$. Using union bound and lemma \[lem:distance\], we get that a randomly drawn $x$ satisfies this with probability at least $1 - O(N) e^{-\Omega(d)}$. \[lem:outer\] Let $x$ be a randomly drawn test point on the outer sphere. Then with probability at least $1 - O(N)e^{-\Omega(d)}$ over the choice of $x$, there is no $x'$ such that $\norm{x - x'} \le \eps$ and $\norm{x_i - x'} \le \sqrt{2} + 2\eps$ for any $i \le 2N$. Similar to the previous lemma, no such $i$ exists if $\norm{x_i - x} > \sqrt{2} + 3\eps$ for all $i \le 2N$. Using union bound and lemma \[lem:distance\], we get that a randomly drawn $x$ satisfies this with probability at least $1 - O(N) e^{-\Omega(d)}$. Lemmas \[lem:inner\] and \[lem:outer\] together give us the result that for almost all points, even after adversarial perturbation, the graph $G(x)$ we construct with threshold $T = \sqrt{2} + 2\eps$ is identical for all points $x$ on the inner sphere, and identical for all points $x$ on the outer sphere (except a $\gamma = O(N)e^{-\Omega(d)}$ fraction): points on the inner sphere get connected to points on the inner sphere, and points on the outer sphere get connected to nothing. Since the map from graphs $G(x)$ to features $f(x)$ is deterministic, this means that $f$, in fact, maps all inner sphere points to one point and all outer sphere points to another point (except a $\gamma$-fraction); that is, $f$ is $(\eps, 0, Ne^{-\Omega(d)})$-robust. It only remains to show that $f$ maps inner-sphere and outer-sphere points to different outputs. Before proceeding further, we now fully explain the construction of the features $f$. Given a training set of $N$ points from the inner sphere and $N$ points from the outer sphere, construct the graph $G(X)$ and take the bottom-$(N+1)$ eigenvectors $v_1, v_2, \dots, v_{N+1}$ and prepend a 0 to each of them to yield $v_1',v_2', \dots, v_{N+1}' \in {\mathbb R}^{2N+1}$. To compute the feature $f(x)$, we first construct the graph $G(x)$. Then, we project the vectors $v_1', v_2', \dots, v_{N+1}'$ into the bottom-$(N+1)$ eigenspace of $G(x)$ to yield vectors $u_1, u_2, \dots, u_{N+1}$. The feature assigned to $G(x)$ is $f(x) := (u_{1,0},u_{2,0}, \dots, u_{N+1, 0}) = e_0^T U \in {\mathbb R}^{N+1}$, where $U$ is the matrix whose columns are the $u_i$, and $e_0 = (1, 0, 0, \dots, 0) \in {\mathbb R}^{2N+1}$ . Similarly, define the matrix $V'$ corresponding to vectors $v_i'$. Let $P$ be the projection matrix onto the bottom-$(N+1)$ eigenspace of $G(x)$ Assume WLOG that the first $N$ training examples are on the inner sphere, and the other $N$ are on the outer sphere. Then the vector $$v^* := (0, \underbrace{1, 1, \dots, 1}_N, \underbrace{0, 0, \dots}_N, 0) \in {\mathbb R}^{2N+1}$$ is in the span of the $v_i'$, since it consists of a 0 prepended to a vector in the zero-eigenspace of $G(X)$. Suppose $v^* = \sum_i {\alpha_i v_i'} = V' \alpha$. Notice that $e_0^T P V' \alpha = e_0^T P v^$ will be 0 when $x$ is on the outer sphere, since $v^$ is itself already in the zero-eigenspace of G(x) in this case. When $x$ is on the inner sphere, projecting $v^$ onto $G(x)$ will make its first component positive, since $v^$ has positive dot product with the vector $u^* = (1, 1, 1, \dots, 0, 0, \dots )$ (which is an eigenvector of $G(x)$), and is orthogonal to every zero eigenvector of $G(x)$ orthogonal to $u^*$—and thus $e_0^T PV'\alpha > 0$ for $x$ on the inner sphere. Thus, $e_0^T U = e_0^T P V'$ must take different values on the outer and inner spheres. This concludes the proof of Theorem \[thm:advspheres\]. [^1]: Equal contribution [^2]: Equal contribution	{ "pile_set_name": "ArXiv" }
--- author: - \| Andrey Malinin\ Department of Engineering\ University of Cambridge\ `{am969}@cam.ac.uk`\ Mark Gales\ Department of Engineering\ University of Cambridge\ `{mjfg}@eng.cam.ac.uk` bibliography: - 'bibliography.bib' title: Prior Networks for Detection of Adversarial Attacks ---	{ "pile_set_name": "ArXiv" }
--- abstract: 'This paper derives the Feynman rules for the diagrammatic perturbation expansion of the effective action around an arbitrary solvable problem. The perturbation expansion around a Gaussian theory is well known and composed of one-line irreducible diagrams only. For the expansions around an arbitrary, non-Gaussian problem, we show that a more general class of irreducible diagrams remains in addition to a second set of diagrams that has no analogue in the Gaussian case. The effective action is central to field theory, in particular to the study of phase transitions, symmetry breaking, effective equations of motion, and renormalization. We exemplify the method on the Ising model, where the effective action amounts to the Gibbs free energy, recovering the Thouless-Anderson-Palmer mean-field theory in a fully diagrammatic derivation. Higher order corrections follow with only minimal effort compared to existing techniques. Our results show further that the Plefka expansion and the high-temperature expansion are special cases of the general formalism presented here.' address: - 'Institute of Neuroscience and Medicine (INM-6) and Institute for Advanced Simulation (IAS-6) and JARA BRAIN Institute I, Jülich Research Centre, Jülich, Germany' - 'Institute of Neuroscience and Medicine (INM-6) and Institute for Advanced Simulation (IAS-6) and JARA BRAIN Institute I, Jülich Research Centre, Jülich, Germany' - 'Department of Physics, Faculty 1, RWTH Aachen University, Aachen, Germany' author: - Tobias Kühn - Moritz Helias title: 'Expansion of the effective action around non-Gaussian theories' --- Introduction\[sec:Introduction\] ================================ Many-particle systems are of interest in various fields of physics. Field theory offers a versatile unique language to describe such systems and powerful methods to treat diverse problems arising in classical statistical mechanics, quantum mechanics, quantum statistics, quantum field theory, and stochastic dynamical systems [@NegeleOrland98; @ZinnJustin96; @Amit84; @Altland01; @Vasiliev98]. Diagrammatic techniques in particular are convenient and efficient to organize practical calculations that arise in the context of systematic perturbation expansions and fluctuation expansions around a solvable problem. But most techniques rely on the solvable part being Gaussian: The basic connecting elements in Feynman diagrams, or Mayer graphs [@Mayer77], are lines. The purpose of this paper is to extend the diagrammatic computation of a central quantity, the effective action $\Gamma$, beyond this Gaussian case. The effective action, vertex generating functional, or Gibbs free energy, is the first Legendre transform of the generating functional of connected Green’s functions or cumulants. Originally developed in statistical mechanics to study non-ideal gases in the 1930s, its modern formulation as a Legendre transform was introduced in the context of quantum statistics [@DeDominicis64_14], but has soon been applied in quantum field theory [@JonaLasinio64_1790] (similar concepts had been developed in parallel in quantum field theory [@Goldstone62]). The effective action is central to the study of phase transitions, because its stationary points determine the values of the order parameters that characterize the states in which matter may exist; it therefore reduces the problem of a phase transition to finding bifurcation points in a variational problem [@Amit84; @ZinnJustin96; @Vasiliev98]. This formulation yields self-consistency equations for the mean value of a field that incorporate fluctuation corrections. At lowest order, at the tree level of diagrams, one obtains the mean-field theory of the Curie-Weiss type. Higher Legendre transforms generalize this concept to finding self-consistent solutions for renormalized higher order Green’s functions [@DeDominicis64_14][@Vasiliev98 chapter 6][@Berges02_223; @Senechal04]; the second transform leading, at lowest order, to a self-consistency equation for the propagator, the Hartree-Fock approximation [@Fock30][@Vasiliev98 i.p. chapter 6 for a review]. The second reason for the importance of the effective action is that it compactly encodes all dynamical or statistical properties of a system: Connected Green’s functions decompose into tree diagrams, whose vertices are formed by derivatives of $\Gamma$. Conversely, relatively fewer diagrams contribute to $\Gamma$, making its computation favorable: Its perturbative expansion around a Gaussian solvable theory, or around Gaussian fluctuations with regard to a background field, requires only one-line (particle) irreducible graphs (1PI) [@DeDominicis64_14; @Amit84; @Kleinert09; @ZinnJustin96]. Third, treating critical points by renormalization group methods in classical [@Wilson74_75; @Wilson75_773] and quantum systems [@Hertz76_1165] leads to the study of flow equations for the effective action; most explicitly shown in its functional formulation [@WETTERICH93_90; @Berges02_223; @Metzner12_299]. Perturbative calculations are essential for most applied problems. But practically all techniques in field theory are based on expansions around the Gaussian case. The topic of this article is a systematic, diagrammatic, and perturbative expansion of the effective action in the general case, where the solvable problem is not of Gaussian type; in statistics: the unperturbed problem has cumulants of orders three or higher; in field theory: the bare theory has non-vanishing connected Green’s functions of orders three or higher; in diagrams: we have bare propagators that tie together three or more field points instead of or besides the usual propagator lines. It may be obvious that the linked cluster theorem in this case remains intact without change, shown below for completeness. Also the converse problem, the decomposition of Green’s functions into connected Green’s functions by Wick’s theorem is immediately replaced by its general counterpart, the factorization of moments into all product sets of cumulants [@Gardiner85]. But diagrammatic rules for the direct expansion of the effective action do not follow as simply. Only for particular non-Gaussian cases, diagrammatic rules have been derived earlier: for the classical non-ideal gas [@Mayer77; @Uhlenbeck63] and for the Ising model [@Vasiliev74; @Vasiliev75] (compare [@Vasiliev98 end of chap. 6.3.1.]). For the classical gas, one finds that only so called star-graphs contribute [@Mayer77; @Vasiliev74; @Vasiliev98 chap. 6.3.2.], which is known as Mayer’s second theorem. Star-graphs are diagrams that do not fall apart by removing one cumulant. For the Ising model one has to introduce a special notation for the sums, the so called ${\cal N}$-form, which only includes contributions from star graphs[@Vasiliev75]. At the end of , we will give more details on this result and show its relation to our work. We first derive a recursive diagrammatic algorithm to compute the effective action perturbatively around a non-Gaussian solvable theory. Then we show that the correction terms produced by the iteration satisfy a set of Feynman rules: Only two topologically defined classes of diagrams remain. The first being irreducible diagrams in a more general sense than in the known Gaussian case. The second being diagrams that contain vertex functions of the solvable theory. The Feynman rules for the former directly follow from the linked cluster theorem. The rules for the latter graphs follow from a set of problem-independent skeleton diagrams that we present subsequently. This set of rules then allows the computation of corrections at arbitrary order in perturbation theory in a non-iterative manner. We find, however, that the iterative procedure itself proves highly efficient. Applied to the special case of the Ising model, it affords only a few more calculations than the method that is specific to the problem [@Vasiliev74; @Vasiliev75]. In this particular case, the iterative procedure is closely related to the known high temperature expansion [@Georges91]. The remainder of the paper is organized as follows: In we introduce the notation and the minimal extension of the diagrammatic language required. sets up the general perturbative problem and shows why the expansion of the partition function and the generating functional of connected Green’s functions straight forwardly extends to the non-Gaussian setting. These preparations are required to derive in our main result, a recursive algebraic equation that determines $\Gamma$. interprets the algebraic expressions in diagrammatic language, leading to a set of Feynman rules to calculate all graphs contributing to $\Gamma$ at arbitrary given order in a non-iterative manner. In we exemplify the algorithm for the simplest non-Gaussian extension of a $\varphi^{3}$ theory. In we demonstrate its application to the Ising model (which turns into the Sherrington-Kirkpatrick spin glass model for Gaussian random couplings [@Kirkpatrick1978]) to derive the Thouless-Anderson-Palmer mean-field theory [@Thouless77_593] and its higher order corrections [@Nakanishi97_8085; @Tanaka98_2302; @Georges91; @Vasiliev74; @Vasiliev75] in a purely diagrammatic manner. Results ======= We here chose a notation that should be transparent with regard to the nature of the problem (following to some extent [@Kleinert09]): our results easily transfer to classical statistical mechanics, quantum mechanics, quantum statistics, or quantum field theory. For clarity, we here stick to the language of classical statistical field theory, in particular to bosonic fields. In the following we set up the language and define elementary quantities. Definition of a field theory\[sub:Definition-field-theory\] ----------------------------------------------------------- We assume that the physical system has degrees of freedom $x\in\E$. We may think of $x$ as a scalar, a vector (as in the example in ) or a (possibly multi-component) field. The domain $\E$ must be chosen accordingly. The system is described by the action $S$ $$\begin{aligned} S:\E & \mapsto & \R.\end{aligned}$$ with the property that a particular configuration of $x$ appears with probability $p(x)\propto\exp(S(x))$. The partition function $\Z$ is then given as $$\begin{aligned} \Z(j) & := & \int_{x}\,\exp(S(x)+j^{\T}x),\qquad Z(j):=\Z(j)/\Z(0),\label{eq:def_partition}\end{aligned}$$ where we denote as $j^{\T}x$ the inner product on the space $\E$. The symbol $\int_{x}$ stands for the sum over all configurations of $x\in\E$, which technically may be a sum, an integral, or a path-integral, depending on the space $\E$. We call $j$ the source field. The properly normalized moment-generating function is denoted as $Z(j)$. Its logarithm is the cumulant generating function $$\begin{aligned} W(j) & := & \ln\,Z(j),\qquad{\savebox{\@brx}{$\m@th{\langle}$} \mathopen{\copy\@brx\kern-0.5\wd\@brx\usebox{\@brx}}}x^{n}{\savebox{\@brx}{$\m@th{\rangle}$} \mathclose{\copy\@brx\kern-0.5\wd\@brx\usebox{\@brx}}}(j):=W^{(n)}(j),\label{eq:def_W}\end{aligned}$$ also called generating functional of the connected Green’s functions in field theory; in statistical mechanics it is related to the Helmholtz free energy $F(j)=\ln\Z(j)$ by $W(j)=F(j)-F(0)$. We denote the $n$-th derivative of a function $f$ by $f^{(n)}$ and the $n$-th cumulant as ${\savebox{\@brx}{$\m@th{\langle}$} \mathopen{\copy\@brx\kern-0.5\wd\@brx\usebox{\@brx}}}x^{n}{\savebox{\@brx}{$\m@th{\rangle}$} \mathclose{\copy\@brx\kern-0.5\wd\@brx\usebox{\@brx}}}.$ To define the effective action $\Gamma(x^{\ast})$, we eliminate the dependence on the source field $j$ in favor of the mean value $\langle x\rangle$ $$\begin{aligned} x^{\ast}(j) & : & =\langle x\rangle={\savebox{\@brx}{$\m@th{\langle}$} \mathopen{\copy\@brx\kern-0.5\wd\@brx\usebox{\@brx}}}x{\savebox{\@brx}{$\m@th{\rangle}$} \mathclose{\copy\@brx\kern-0.5\wd\@brx\usebox{\@brx}}}=W^{(1)}(j),\label{eq:mean_of_j}\end{aligned}$$ by a Legendre-Fenchel transform $$\begin{aligned} \Gamma(x^{\ast}) & : & =\sup_{j}\,j^{\T}x^{\ast}-W(j).\label{eq:def_gamma}\end{aligned}$$ We provide the derivation of in , because we need an intermediate step as the starting point of the perturbative expansion. To this end we employ the reciprocity property of Legendre transforms $$\begin{aligned} j(x^{\ast}) & = & \Gamma^{(1)}(x^{\ast}),\label{eq:equn_of_state}\end{aligned}$$ which, as derived in the appendix (with ), leads to the integral-differential equation [@Kleinert09 Sec. 3.23.6] $$\begin{aligned} \exp\left(-\Gamma(x^{\ast})\right) & = & \Z(0)^{-1}\int_{\delta x}\,\exp\left(S(x^{\ast}+\delta x)+\Gamma^{(1)\T}(x^{\ast})\,\delta x\right).\label{eq:int_diff_gamma}\end{aligned}$$ The latter expression is the starting point for the perturbative expansion of $\Gamma$. It is also the natural starting point of a loopwise expansion. The derivative operators that lead to $W^{(1)}$ and $\Gamma^{(1)}$ are given according to the space $\E$ as either partial derivatives or, in the case of fields, functional derivatives. Perturbative problems\[sub:Perturbative-setting\] ------------------------------------------------- The equation of state provides us with a self-consistency equation for the mean value $x^{\ast}$. The general strategy is therefore to obtain an approximation of $\Gamma$ that includes fluctuation corrections and then use the equation of state to get an approximation for the mean value including these very corrections. We will here derive a perturbative procedure to calculate approximations of $\Gamma$ and will find the graphical rules for doing so. To solve a problem perturbatively we decompose the action as $$\begin{aligned} S(x) & = & S_{0}(x)+\epsilon V(x)\label{eq:def_S_pert}\end{aligned}$$ with a part $S_{0}$ that can be solved exactly, that is we know its corresponding cumulant generating function $W_{0}(j)$, and the remaining terms collected in $\epsilon V(x)$. For the special case of $S_{0}$ being quadratic in the fields $x$, this leads to the well known result that corrections to $\Gamma$ are composed of one-line irreducible graphs only; our algorithm includes but extends this case to non-quadratic $S_{0}$. An example of this situation is depicted in , where the solvable theory $W_{0}$ already contains cumulants of order one to three. We use the common notation for the interaction or bare vertices, which we will call just “vertices” in the following: the $n$-th power of the field $x$ is denoted as a vertex with $n$ emerging lines. The connecting elements in the graphs are the cumulants contained in $W_{0}$. These are symbolized by circles, where the number of legs corresponds to the order of the cumulant (following e.g. [@Amit84; @NegeleOrland98]). In ordinary Feynman diagrams, $W_{0}$ would only have second order cumulants, which are typically denoted by straight lines connecting the vertices. The proof for the diagrammatic expansion closely follows the proof of the linked cluster theorem (sometimes also called first Mayer theorem [@Mayer77]), the connectedness of contributions to the cumulant generating function $W$. We therefore briefly recapitulate the most important steps of the latter proof (adapted from [@ZinnJustin96 Sec. 6.1.1]) here and provide all details in for completeness and consistency of notation. The perturbative corrections $W_{V}(j)=W(j)-W_{0}(j)$ to the cumulant generating function obey the operator equation (following from ) $$\begin{aligned} & & \exp(W_{V}(j))\label{eq:W_V_operator_eq}\\ & = & \exp\left(-W_{0}(j)\right)\,\exp\left(\epsilon V(\partial_{j})\right)\,\exp\left(W_{0}(j)\right)\,\frac{\Z_{0}(0)}{\Z(0)},\nonumber \end{aligned}$$ where the latter constant factor is immaterial for the determination of cumulants. Expanding in $\epsilon$, the function $W$ can be constructed iteratively, where each step $$\begin{aligned} & & W_{l+1}(j)-W_{l}(j)\label{eq:expansion_W_iteration_main}\\ & = & \frac{\epsilon}{L}\,\left(\exp\left(-W_{l}(j)\right)\,V(\partial_{j})\,\exp\left(W_{l}(j)\right)\right)\nonumber \\ & & +\mathcal{O}\big(\epsilon^{2}/L^{2}\big)\nonumber \end{aligned}$$ produces additional diagrams. The proof of connectedness then proceeds by induction under the assumption that all graphs contained in $W_{l}$ are connected. In the following we will call a “component” a term that appears in $W_{V}$ that is composed of a product of vertices, the Taylor coefficients $V$, and of cumulants $W_{0}^{(n)}$, the derivatives of the cumulant generating function of the unperturbed theory. Representing $V$ in its Taylor series, every vertex of $V$ ties together components already contained in $W_{l}$. The cumulant generating function is then given as $$\begin{aligned} W & = & \lim_{L\to\infty}\,W_{L}.\end{aligned}$$ Combinatorics shows that only those graphs survive in the limit that pick up each of their $k$ vertices in a different step in , from which the factor $\left(\frac{\epsilon}{L}\right)^{k}\left(\begin{array}{c} L\\ k \end{array}\right)\;\stackrel{L\to\infty}{\to}\;\frac{\epsilon^{k}}{k!}$ arises (for details see ). So we see that $W_{l}$ at any intermediate step $l$ may contain connected components with arbitrary numbers of external legs. By the inductive nature of the proof it is therefore clear that the linked cluster theorem holds independently of the number of connected components in $W_{0}$; that is to say it is generally true, even if already $W_{0}$ contains connected components with arbitrary numbers of legs; if it is the cumulant generating function of a non-Gaussian theory. [test7]{} $$\begin{aligned} {\bf a)} \quad \epsilon V = \parbox{30mm}{ \begin{fmfgraph}(40,40) \fmfsurroundn{i}{3} \fmf{plain}{i1,v1,i2} \fmf{plain}{i3,v1} \end{fmfgraph} }\\ {\bf b)} \quad W_{0}= \parbox{15mm}{ \begin{fmfgraph}(30,30) \fmfsurroundn{i}{2} \fmf{phantom}{i1,v,i2} \fmfv{d.s=circle, d.filled=empty}{v} \end{fmfgraph} } \mkern-18mu = \; \quad \parbox{10mm}{ \begin{fmfgraph}(30,30) \fmfsurroundn{i}{2} \fmf{plain}{i2,v} \fmf{phantom}{v,i1} \fmfv{d.s=circle, d.filled=empty}{v} \fmflabel{j}{i2} \end{fmfgraph} } + {1 \over 2} \; \quad \parbox{15mm}{ \begin{fmfgraph}(30,30) \fmfsurroundn{i}{2} \fmf{plain}{i2,v} \fmf{plain}{v,i1} \fmfv{d.s=circle, d.filled=empty}{v} \fmflabel{j}{i2} \fmflabel{j}{i1} \end{fmfgraph} } + {1 \over 3!} \quad \parbox{15mm}{ \begin{fmfgraph}(30,30) \fmfsurroundn{i}{3} \fmf{plain}{i2,v,i3} \fmf{plain}{v,i1} \fmfv{d.s=circle, d.filled=empty}{v} \fmflabel{j}{i1} \fmflabel{j}{i2} \fmflabel{j}{i3} \end{fmfgraph} } \end{aligned}$$ Recursion for the effective action\[sub:Recursion-for-Gamma\] ------------------------------------------------------------- We now proceed to derive our main result, the diagrammatic expansion of the effective action $\Gamma$. We here explain the main steps and provide all details in and . To lowest order in perturbation theory, setting $\epsilon=0$ in , we get $W(j)=W_{0}(j)$; the leading order term in $\Gamma$ is the corresponding Legendre transform $$\begin{aligned} \Gamma_{0}(x^{\ast}) & = & \sup_{j}\,j^{\T}x^{\ast}-W_{0}(j).\label{eq:Gamma0_pert-1}\end{aligned}$$ We first need to derive a recursive equation to obtain approximations of the form $$\begin{aligned} \Gamma(x^{\ast}) & = & :\Gamma_{0}(x^{\ast})+\Gamma_{V}(x^{\ast}),\label{eq:Gamma_pert_decomposition}\end{aligned}$$ where we defined $\Gamma_{V}(x^{\ast})$ to contain all correction terms that have at least one interaction vertex due to the interaction potential $V$ to some order $\epsilon^{k}$ in perturbation theory. We will use the iteration in a second step to proof Feynman rules for the diagrams. It is a priori not clear that the decomposition is useful. We show in that this indeed so and, moreover, that it leads to the recursive operator equation for $\Gamma_{V}$ $$\begin{aligned} & & \exp(-\Gamma_{V}(x^{\ast}))\label{eq:Gamma_V_recursion}\\ & = & \exp(-W_{0}(j))\,\exp\big(\epsilon V(\partial_{j})+\Gamma_{V}^{(1)\T}(x^{\ast})(\partial_{j}-x^{\ast})\big)\,\exp(W_{0}(j))\big\|_{j=\Gamma_{0}^{(1)}(x^{\ast})},\nonumber \end{aligned}$$ which has a similar form as , defining the linked cluster expansion of $W$. The term $\Gamma_{V}^{(1)\T}(x^{\ast})(\partial_{j}-x^{\ast})$ can hence be seen as a monopole vertex. We want to solve the latter equation iteratively order by order in the number of vertices $k$, defining $\Gamma_{V,k}$. Analogous to the proof of the linked cluster theorem, we arrive at a recursion by writing the exponential of the differential operator in as a limit and by expanding the logarithm to obtain the iteration $$\begin{aligned} & & g_{l+1}(j)-g_{l}(j)\label{eq:iteration_g}\\ & = & \frac{\epsilon}{L}\,\exp(-W_{0}(j)-g_{l}(j))\,V(\partial_{j})\,\exp(W_{0}(j)+g_{l}(j))\label{eq:add_connected}\\ & + & \frac{1}{L}\,\exp(-W_{0}(j)-g_{l}(j))\,\Gamma_{V}^{(1)}(x^{\ast})\left(\partial_{j}-x^{\ast}\right)\,\exp(W_{0}(j)+g_{l}(j))\label{eq:add_reducible}\\ & + & \mathcal{O}(L^{-2}),\nonumber \end{aligned}$$ with initial condition $$\begin{aligned} g_{0} & = & 0.\label{eq:G_0}\end{aligned}$$ We obtain the perturbation correction to $\Gamma$ as the limit $$\begin{aligned} -\Gamma_{V}(x^{\ast}) & = & \lim_{L\to\infty}\,g_{L}(j)\Big\|_{j=\Gamma_{0}^{(1)}(x^{\ast})}=:g(j)\Big\|_{j=\Gamma_{0}^{(1)}(x^{\ast})}.\label{eq:Gamma_V_final}\end{aligned}$$ To obtain the final result , we need to express $j=\Gamma_{0}^{(1)}(x^{\ast})$ in $g_{l}(j)$. The latter step is crucial to obtain $\Gamma_{V}(x^{\ast})$ as a function of $x^{\ast}$. Because the mean value is given by $x^{\ast}=W_{0}^{(1)}(j_{0})$ for $j_{0}=\Gamma_{0}^{(1)}(x^{\ast})$ (following from and ), this step effectively expresses all cumulants ${\savebox{\@brx}{$\m@th{\langle}$} \mathopen{\copy\@brx\kern-0.5\wd\@brx\usebox{\@brx}}}x^{n}{\savebox{\@brx}{$\m@th{\rangle}$} \mathclose{\copy\@brx\kern-0.5\wd\@brx\usebox{\@brx}}}_{0}\equiv W_{0}^{(n)}$ of the unperturbed theory $$\begin{aligned} {\savebox{\@brx}{$\m@th{\langle}$} \mathopen{\copy\@brx\kern-0.5\wd\@brx\usebox{\@brx}}}x^{n}{\savebox{\@brx}{$\m@th{\rangle}$} \mathclose{\copy\@brx\kern-0.5\wd\@brx\usebox{\@brx}}}_{0}(x^{\ast}) & = & W_{0}^{(n)}(\Gamma_{0}^{(1)}(x^{\ast}))\label{eq:replacement_j}\end{aligned}$$ in terms of the first cumulant ${\savebox{\@brx}{$\m@th{\langle}$} \mathopen{\copy\@brx\kern-0.5\wd\@brx\usebox{\@brx}}}x{\savebox{\@brx}{$\m@th{\rangle}$} \mathclose{\copy\@brx\kern-0.5\wd\@brx\usebox{\@brx}}}_{0}(x^{\ast})=x^{\ast}$, where $x^{\ast}$ is the value of the first cumulant of the full theory. The relation $j_{0}(x^{\ast})$ can practically be obtained either by inverting $x^{\ast}=W_{0}^{(1)}(j_{0})$ or directly as $j_{0}=\Gamma_{0}^{(1)}(x^{\ast})$, if $\Gamma_{0}(x^{\ast})$ is known explicitly. From the last point follows that the differentiation in with $\Gamma_{V}^{(1)}(x^{\ast})\equiv\partial_{x^{\ast}}\Gamma_{V}(x^{\ast})=\partial_{x^{\ast}}(g\circ\Gamma_{0}^{(1)}(x^{\ast}))$ produces an inner derivative $\Gamma_{0}^{(2)}$ attached to a single leg of a any unperturbed cumulant $W_{0}^{(n)}$ contained in $g_{L}$. Defining the additional symbol $$\begin{aligned} \Gamma_{0}^{(2)}(x^{\ast})=:\quad & \Diagram{g!p{0}g}\end{aligned}$$ allows us to write these contributions as $$\begin{aligned} \partial_{x^{\ast}}(g\circ\Gamma_{0}^{(1)}) & \equiv & (g^{(1)}\circ\Gamma_{0}^{(1)})\,\Gamma_{0}^{(2)}\Diagram{=\Diagram{!c{g}fg!p{0}g} } .\label{eq:subgraph_g_1-2}\end{aligned}$$ This property will become important in the proof that the iteration generates only one-line irreducible graphs. To see this, we have to generalize the notion of irreducibility known from the Gaussian case, which we will do in the following section. Proof of a generalized irreducibility\[sub:Graphical-rules-for-Gamma\] ---------------------------------------------------------------------- The term “one-line irredicibility” in the Gaussian case refers to the absence of diagrams that can be disconnected by cutting a single second order bare propagator (a line in the original language of Feynman diagrams). In the slightly generalized graphical notation introduced in , these graphs would have the form $$\begin{aligned} & & \Diagram{!c{k^{\prime}}f!c{0}f!c{k^{\prime\prime}}\quad,}\end{aligned}$$ where two sub-graphs of $k$ and $k^{\prime}$ vertices, respectively, are joined by a bare second order cumulant $\Feyn{fcf}$. Before proceeding to the proof, we need to define irreducibility in a more general sense than used in the Gaussian case. If a graph can be decomposed into a pair of sub-graphs both containing vertices by disconnecting the end point of a single vertex, we call this graph reducible. In the Gaussian case, this definition is identical to one-line reducibility, because all end points of vertices necessarily connect to a second order propagator, a line. This is not necessarily the case if the bare theory has higher order cumulants. We may have components of graphs, such as [Example\_three\_point\_cumu\_three\_point\_int]{} $$\begin{aligned} &\parbox{100mm}{ \begin{fmfgraph}(150,75) \fmfpen{.75thin} \fmftop{ou1,og1,ou2,og2,ou3,og3,ou4,og4,ou5,og5} \fmfbottom{uu1,ug1,uu2,ug2,uu3,ug3,uu4,ug4,uu5,ug5} \fmf{phantom}{ou1,g1,G2,ug2} \fmf{phantom}{ou2,v1,g2,ug3} \fmf{plain}{v1,g2} \fmf{phantom}{ou3,G1,v2,ug4} \fmf{phantom}{ou4,g3,G3,ug5} \fmf{plain, tension=1.25}{v2,ug4} \fmf{phantom}{ou3,v1,G2,ug1} \fmf{plain, tension = 1.25}{ou3,v1} \fmf{phantom}{v1,G2,ug1} \fmf{phantom}{ou4,G1,g2,ug2} \fmf{plain}{g2,ug2} \fmf{phantom}{ou5,g3,v2,ug3} \fmf{phantom,tension=1.}{ou5,g3} \fmf{plain,tension=1.}{g3,v2} \fmf{plain}{g2,v2} \fmf{plain}{g1,v1} \fmfv{decor.shape=circle,decor.filled=empty, decor.size=12.thin}{g2} \end{fmfgraph} },& \end{aligned}$$ where the three-point cumulant (could also be of higher order) connects to two third (or higher) order interactions on either side. Disconnecting a single leg, either to the left or to the right, decomposes the diagram into two parts, each of which contains at least one interaction vertex. We call such a diagram reducible and diagrams without this property irreducible here. Note that a single leg of an interaction vertex that ends on a first order cumulant does not make a diagram reducible; both components need to contain at least one interaction. ### Cancellation of reducible diagrams\[sec:Cancellation-of-reducible\] We employ the following graphical notation: Since $$\begin{aligned} g_{l}(j) & = & :\feyn{!c{g_{l}}}\end{aligned}$$ depends on $j$ only indirectly by the $j$-dependence of the contained bare cumulants, we denote the derivative by attaching one leg, which is effectively attached to one of the cumulants of $W_{0}$ contained in $g_{l}$ $$\begin{aligned} \Diagram{\vertexlabel^{j}f!c{g_{l}}} & := & \partial_{j}\:\Diagram{!c{g_{l}}} :=\partial_{j}g_{l}(j).\end{aligned}$$ We first note that generates two kinds of contributions to $g_{l+1}$, corresponding to the lines and , respectively. The first line causes contributions that come from the vertices of $\epsilon V(\partial_{j})$ alone. These are similar as in the linked cluster theorem . Determining the first order correction yields with $g_{0}=0$ $$\begin{aligned} g_{1}(j) & = & \frac{\epsilon}{L}\,\exp(-W_{0}(j))\,V(\partial_{j})\,\exp(W_{0}(j))\label{eq:first_order_g1}\\ & + & \mathcal{O}(L^{-2}),\nonumber \end{aligned}$$ which contains all graphs with a single vertex from $V$ and connections formed by cumulants of $W_{0}$. These graphs are trivially irreducible, because they only contain a single vertex. The proof of the linked cluster theorem (see ) shows how the construction proceeds recursively: correspondingly the $l+1$-st step generates all connected graphs from components already contained in $W_{0}+g_{l}$. These are tied together with a single additional vertex from $\epsilon V(x)$. In each step, we only need to keep those graphs where the new vertex in joins at most one component from $g_{l}$ to an arbitrary number of components of $W_{0}$, hence we maximally increase the number of vertices in each component by one. This is so, because comparing the combinatorial factors and , contributions formed by adding more than a single vertex (joining two or more components from $g_{l}$ by the new vertex) in a single step are suppressed with at least $L^{-1}$, so they vanish in the limit . The second term is similar to with three important differences: - The differential operator appears in the form $\partial_{j}-x^{\ast}$. As a consequence, when setting $j_{0}=\Gamma_{0}^{(1)}(x^{\ast})$ in the end in , all terms cancel where $\partial_{j}$ acts directly on $W_{0}(j)$, because $W_{0}^{(1)}(j_{0})=x^{\ast}$; non-vanishing contributions only arise if the $\partial_{j}$ acts on a component contained in $g_{l}$. Since vertices and cumulants can be composed to a final graph in arbitrary order, the diagrams produced by $\partial_{j}-x^{\ast}$ acting on $g_{l}$ are the same as those in which $\partial_{j}-x^{\ast}$ first acts on $W_{0}$ and in a subsequent step of the iteration another $\partial_{j}$ acts on the produced $W_{0}^{(1)}$. So to construct the set of all diagrams it is sufficient to think of $\partial_{j}$ as acting on $g_{l}$ alone; the reversed order of construction, where $\partial_{j}$ first acts on $W_{0}$ and in subsequent steps of the iteration the remainder of the diagram is attached to the resulting $W_{0}^{(1)}$, is contained in the combinatorics. - The single appearance of the differential operator $\partial_{j}$ acts like a monopole vertex: the term therefore attaches an entire sub-diagram contained in $\Gamma_{V}^{(1)}$ by a single link to any component contained in $g_{l}$. - These attached sub-diagrams from $\Gamma_{V}^{(1)}(x^{\ast})$ do not depend on $j$; the $j$-dependence of all contained cumulants is fixed to the value $j=\Gamma_{0}^{(1)}(x^{\ast})$, as seen from . As a consequence, these sub-graphs cannot form connections to vertices in subsequent steps of the iteration. Considering that $\Gamma_{V}^{\left(1\right)}\left(x^{\ast}\right)$ is represented by , in step $l+1$ the line contributes graphs of the form $$\begin{aligned} g^{(1)}\,\Gamma_{0}^{(2)}\,g_{l}^{(1)} & =\Diagram{!c{g}fg!p{0}gf!c{g_{l}}} & \quad.\label{eq:reducible_general-1}\end{aligned}$$ Since by their definition as a pair of Legendre transforms we have $$\begin{aligned} 1 & =\Gamma_{0}^{(2)}W_{0}^{(2)} & =\Diagram{g!p{0}gf!c{0}f} \quad,\end{aligned}$$ we notice that the subtraction of the graphs may cancel certain connected graphs produced by the line . In the case of a Gaussian solvable theory $W_{0}$ this cancellation is the reason why only one-line irreducible contributions remain. We here obtain the more general result, that these contributions cancel all reducible components, according to the definition above. We will prove this central result in the following. To see the cancellation, we note that a reducible graph by our definition has at least two components joined by a single leg of a vertex. Let us first consider the case of a diagram consisting of exactly two irreducible sub-diagrams joined by a single leg, as it is generated by . This leg may either belong to the part $g^{(1)}$ or to $g_{l}^{(1)}$ in , so either to the left or to the right sub-diagram. In both cases, there is a second cumulant $W_{0}^{(2)}$ either left or right of $\Gamma_{0}^{(2)}$. This is because if the two components are joined by a single leg, this particular leg must have terminated on a $W_{0}^{(1)}$ prior to the formation of the compound graph; in either case this term generates $W_{0}^{(1)}\stackrel{\partial_{j}}{\to}W_{0}^{(2)}$. The second point to check is the combinatorial factor of graphs of the form . To construct a graph of order $k$, where the left component has $k^{\prime}$ bare vertices and the right has $k-k^{\prime}$, we can choose one of $L$ steps within the iteration in which we may pick up the left term by . The remaining $k-k^{\prime}$ vertices are picked up by , which are $\left(\begin{array}{c} L-1\\ k-k^{\prime} \end{array}\right)$ possibilities to choose $k-k^{\prime}$ steps from $L-1$ available ones. Every addition of a component to the graph comes with $L^{-1}$. Any graph in $\Gamma_{V}$ with $k^{\prime}$ vertices is $\propto\frac{\epsilon^{k^{\prime}}}{k^{\prime}!}$, so together we get $$\begin{aligned} \frac{L}{L}\,\frac{\epsilon^{k^{\prime}}}{k^{\prime}!}\,\left(\frac{\epsilon}{L}\right)^{k-k^{\prime}}\left(\begin{array}{c} L-1\\ k-k^{\prime} \end{array}\right) & \stackrel{L\to\infty}{\to} & \frac{\epsilon^{k}}{k^{\prime}!(k-k^{\prime})!}.\label{eq:comb_reducible-1}\end{aligned}$$ The symmetry factors $s_{1},s_{2}$ of the two sub-graphs generated by enter the symmetry factor $s=s_{1}\cdot s_{2}\cdot c$ of the composed graph as a product, where $c$ is the number of ways in which the two sub-graphs may be joined. But the factor $s$, by construction, excludes those symmetries that interchange vertices between the two sub-graphs. Assuming, without loss of generality, a single sort of interaction vertex, there are $s^{\prime}=\left(\begin{array}{c} k\\ k^{\prime} \end{array}\right)=\frac{k!}{k^{\prime}!(k-k^{\prime})!}$ ways of choosing $k^{\prime}$ of the $k$ vertices to belong to the left part of the diagram. Therefore the symmetry factor $s$ is smaller by the factor $s^{\prime}$ than the symmetry factor of the corresponding reducible diagram constructed by alone, because the latter exploits all symmetries, including those that mix vertices among the sub-graphs. Combining the defect $s^{\prime}$ with the combinatorial factor yields $\frac{1}{k^{\prime}!(k-k^{\prime})!}/s^{\prime}=\frac{1}{k!}$, which equals the combinatorial factor of the reducible graph under consideration. The generalization to the case of an arbitrary number of sub-diagrams of which $M$ are irreducible and connected to the remainder of the diagram by exactly one link, called “leaves”, is straightforward: we can always pick one of these sub-diagrams and replace it by its corresponding contribution to $\Gamma^{\left(1\right)}\left(x^{\ast}\right)$. Then, by the same arguments as before, we produce the same diagram with the same prefactor, only with opposite sign. In summary we conclude that all reducible graphs are canceled by . We therefore obtain the first part of our Feynman rules: All connected, irreducible diagrams that are also contained in the perturbation expansion of $W_{V}$ also contribute to $\Gamma_{V}$ with a negative sign and the same combinatorial factor. But there is a second sort of graphs produced by that does not exist in the Gaussian case: If the connection between the two sub-components by $\Feyn{gpg}$ ends on a third or higher order cumulant. These graphs cannot be produced by , so they remain with a minus sign. We show an example of such graphs in c). One might wonder why this contribution does not cancel while making a “reducible impression”, if one interprets $\Gamma_{0}^{\left(2\right)}$ as a kind of two-point interaction. The solution is that for diagrams of this kind, indeed contributions of opposite signs and same absolute values are generated, but the contribution with the one sign is generated once and the contribution with the other sign twice, therefore the total contribution does not vanish. We address this issue more thoroughly in . Moreover, subsequent application of $\partial_{x^{\ast}}$ on such components produces graphs that contain higher order derivatives of $\Gamma_{0}^{(n)}$. By the arguments given in the proof for the generalized irreducibility, we deduce that all diagrams cancel that have at least one leaf from $W_{V}^{\left(1\right)}$ (that includes only unperturbed cumulants and interactions, but no vertices $\Gamma^{\left(n\right)}\left(x^{\ast}\right)$) and that is connected to the remainder of the diagram by a single leg of an interaction (and not via a leg of $\Gamma^{\left(n\right)}\left(x^{\ast}\right)$, see ). Reducible non-standard diagrams that violate these prerequisites can occur. In the end of the following subsection, we will also give an example for this case. To enumerate all all non-standard diagrams and to find their Feynman rules, we also derive a set of skeleton diagrams in the following subsection that allow the computation of all contributions at any given order including their combinatorial factor. ### Perturbative diagrammatics derived from skeleton diagrams\[sec:Perturbative-diagrammatics-deriv\] To obtain a better understanding of the types of diagrams that contribute to the perturbation expansion, in particular the non-standard diagrams not contained in $W_{V}$, it is useful to derive a non-iterative approach based on a set of skeleton diagrams. We use the term “skeleton diagram” for a diagram containing dressed as opposed to bare cumulants and vertices. We start with a Taylor expansion around a adroitly chosen point $x_{1}$ $$\begin{aligned} \Gamma(x^{\ast}) & = & \Gamma\left(x_{1}\right)+\Gamma^{\left(1\right)}\left(x_{1}\right)^{\T}\left(x^{\ast}-x_{1}\right)\label{eq:Taylor_Gamma_Skeleton}\\ & & +\sum_{n=2}\,\sum_{i_{1}\cdots i_{n}}\frac{\Gamma_{i_{1}\cdots i_{n}}^{(n)}(x_{1})}{n!}\,\prod_{l=1}^{n}(x^{\ast}-x_{1})_{i_{l}},\nonumber \end{aligned}$$ where we wrote the first two terms explicitly. We now choose a particular point $x_{1}$ by making use of the involutive property of the Legendre transform, which is given for all smooth and convex cumulant generating functions, irrespective of the form of the underlying theory. Concretely, we choose $x_{1}:=W^{(1)}(\Gamma_{0}^{(1)}(x^{\ast})),$ which is equivalent to $$\begin{array}{lcr} j_{0} & = & \Gamma_{0}^{(1)}(x^{\ast})\\ j_{0} & = & \Gamma^{(1)}(x_{1}) \end{array}\Longleftrightarrow\begin{array}{lcr} x^{\ast} & = & W_{0}^{(1)}(j_{0})\\ x_{1} & = & W^{(1)}(j_{0}) \end{array}.\label{eq:Def_x1_via_x_star}$$ As an immediate consequence, we see that $$x^{\ast}-x_{1}=W_{0}^{(1)}(j_{0})-W^{(1)}(j_{0})=-W_{V}^{(1)}(j_{0}).\label{eq:W_V_x}$$ Using these relations and the definition of the Legendre transform, the two terms in the first line of can be rewritten as $$\begin{aligned} & & j_{0}^{\T}x_{1}-W\left(j_{0}\right)+j_{0}^{\T}\,\left(x^{\ast}-x_{1}\right)\nonumber \\ & = & j_{0}^{\T}x^{\ast}-W_{0}\left(j_{0}\right)-W_{V}\left(j_{0}\right)\nonumber \\ & = & \Gamma_{0}\left(x^{\ast}\right)-W_{V}\left(j_{0}\right).\label{eq:Vertex_Taylor_x1}\end{aligned}$$ We now see that the first two terms in , besides $\Gamma_{0}(x^{\ast})$, also contain all diagrams from $W_{V}\left(j_{0}\right)$, but with the opposite sign. We note that by their dependence on $x^{\ast}$ and $j_{0}$, respectively, these contributions can be calculated: The relation allows us to express all bare cumulants $W_{0}^{(n)}(j_{0})$ that appear in $-W_{V}$ in terms of the first cumulant $x^{\ast}\equiv W_{0}(j_{0})$. The second line in produces additional diagrams that are of second order in the interaction or higher; this is because the “leaves” of these tree diagrams are $-W_{V}$, which, by definition, contain at least one interaction vertex. These terms include those diagrams that cancel the reducible diagrams included in $W_{V}\left(j_{0}\right)$, as proven above. In addition, these terms contain the non-standard diagrams. We will describe in the following, how the exact form and combinatorial factors of these diagrams can be obtained. To express the dressed vertex functions $\Gamma^{(n)}(x_{1})$ that appear in the second line of , we first derive a perturbative expansion for $\Gamma^{(2)}(x_{1})$, which is basically Dyson’s equation. We here include the derivation to see that it holds beyond the Gaussian case. Decomposing $W$ into the solvable part and its perturbative corrections, the reciprocity relation $$\begin{aligned} W^{(2)}(j_{0})\,\Gamma^{(2)}(W^{(1)}(j_{0})) & = & 1\label{eq:reciprocity-1}\end{aligned}$$ takes the form $$\begin{aligned} (W_{0}^{(2)}(j_{0})+W_{V}^{(2)}(j_{0}))\,\Gamma^{(2)}(\underbrace{W^{(1)}(j_{0})}_{\equiv x_{1}}) & = & 1.\end{aligned}$$ Multiplying from left with $\Gamma_{0}^{(2)}(W_{0}^{(1)}(j_{0}))=\Gamma_{0}^{(2)}(x^{\ast})$, which is the inverse of $W_{0}^{(2)}(j_{0})$, we obtain by rearranging $$\begin{aligned} \Gamma^{(2)}(x_{1}) & = & \Gamma_{0}^{(2)}(x^{\ast})-\Gamma_{0}^{(2)}(x^{\ast})\,W_{V}^{(2)}(j_{0})\,\Gamma^{(2)}(x_{1}).\label{eq:Dyson_recursion}\end{aligned}$$ The latter term contains at least one interaction vertex in $W_{V}$, the former is of order zero. So we may iterate this equation to obtain an inverted Dyson’s equation for $\Gamma^{(2)}$ that reads $$\begin{aligned} \Gamma^{(2)}(x_{1}) & = & \Gamma_{0}^{(2)}(x^{\ast})-\Gamma_{0}^{(2)}(x^{\ast})\,W_{V}^{(2)}(j_{0})\,\Gamma_{0}^{(2)}(x^{\ast})\pm\ldots\label{eq:inversed_Dyson}\\ \nonumber \\ & = & \Diagram{g!p{0}g-g!p{0}g!c{V}g!p{0}g} \pm\ldots.\nonumber \end{aligned}$$ It is easily checked by insertion that solves . So we have expressed the second order vertex at $x_{1}$ by means of quantities that are directly accessible or can be calculated perturbatively in a straight-forward way, by their dependence on $x^{\ast}$ and $j_{0}$, respectively. Higher order vertices are determined in the standard way by differentiating the reciprocity relation multiple times with respect to $j_{0}$, and removing the legs $W^{(2)}(j_{0})$ that arise from inner derivatives by multiplication with the corresponding inverse $\Gamma^{(2)}(x_{1})$. The result is the well-known decomposition of vertex in terms of tree skeleton diagrams [@ZinnJustin96; @NegeleOrland98] $$\begin{aligned} \Gamma_{123}^{(3)}(x_{1}) & = & -\sum_{\{i_{l}\}}W_{i_{1}i_{2}i_{3}}^{(3)}(j_{0})\,\Gamma_{i_{1}1}^{(2)}(x_{1})\,\Gamma_{i_{2}2}^{(2)}(x_{1})\,\Gamma_{i_{3}3}^{(2)}(x_{1})\label{eq:skeleton}\\ \Gamma_{1234}^{(4)}(x_{1}) & = & -\sum_{\{i_{l}\}}W_{i_{1}i_{2}i_{3}i_{4}}^{(4)}(j_{0})\,\Gamma_{i_{1}1}^{(2)}(x_{1})\,\Gamma_{i_{2}2}^{(2)}(x_{1})\,\Gamma_{i_{3}3}^{(2)}(x_{1})\,\Gamma_{i_{4}4}^{(2)}(x_{1})\nonumber \\ & & +\sum_{\{i_{l}\}}W_{i_{1}i_{2}i_{3}}^{(3)}\,W_{i_{4}i_{5}i_{6}}^{(3)}(j_{0})\,\Gamma_{i_{1}1}^{(3)}(x_{1})\,\Gamma_{i_{2}2}^{(2)}(x_{1})\,\Gamma_{i_{3}i_{4}}^{(2)}(x_{1})\,\Gamma_{i_{5}3}^{(2)}(x_{1})\,\Gamma_{i_{6}4}^{(2)}(x_{1})+2\,\text{perm.}\nonumber \\ & \ldots & .\nonumber \end{aligned}$$ Diagrammatically, the first of these relations reads [treestructure\_of\_vertices\_2]{} $$\begin{aligned} \parbox{30mm}{ \begin{fmfgraph}(40,40) \fmfsurroundn{i}{3} \fmf{plain}{i1,v1,i2} \fmf{plain}{i3,v1,} \fmflabel{1}{i1} \fmflabel{2}{i2} \fmflabel{3}{i3} \fmfv{decor.shape=circle, d.filled=empty}{v1} \end{fmfgraph} } =& \; - \; \parbox{30mm}{ \begin{fmfgraph}(100,100) \fmfsurroundn{i}{3} \fmf{wiggly}{v1,v3} \fmf{wiggly}{v1,v5} \fmf{wiggly}{v1,v7} \fmf{plain}{i1,v2,v3} \fmf{plain}{i2,v4,v5} \fmf{plain}{i3,v6,v7} \fmflabel{1}{i1} \fmflabel{2}{i2} \fmflabel{3}{i3} \fmfv{label=$i_1$, l.a=90}{v3} \fmfv{label=$i_2$, l.a=-170}{v5} \fmfv{label=$i_3$, l.a=-10}{v7} \fmfv{decor.shape=circle, d.filled=shaded}{v1} \fmfv{decor.shape=circle, d.filled=empty}{v2} \fmfv{decor.shape=circle, d.filled=empty}{v4} \fmfv{decor.shape=circle, d.filled=empty}{v6} \end{fmfgraph} } \end{aligned}$$ We can hence express vertices of arbitrary order at the point $x_{1}$ by cumulants at the point $j_{0}$ (which is given by $x^{\ast}$ by means of ) and $\Gamma^{\left(2\right)}\left(x_{1}\right)$, which is again given by the same cumulants and $\Gamma_{0}^{(2)}(x^{\ast})$ by Dyson’s equation . The important point to note here is that the set of rules to translate skeleton diagrams into their perturbative expansion is problem independent; it just results from the properties of the Legendre transform and the decomposition into solvable part and perturbation. Because both factors in every term of the sum in the second line of depend on the interaction $V$, the order of each such term is the sum of the orders of its factors. We derive two additional rules from this view. First, since also $W_{0}(j_{0})$ and $\Gamma_{0}(x^{\ast})$ form a pair of Legendre transforms, their reciprocity relation $W_{0}^{(2)}(j_{0})\,\Gamma_{0}^{(2)}(x^{\ast})=1$ gives rise to a relation corresponding to , but with $\Gamma^{(n)}(x_{1})$ replaced by $\Gamma_{0}^{(n)}(x^{\ast})$ and $W^{(n)}(j_{0})$ by $W_{0}^{(n)}(j_{0})$. This means that the lowest order terms for the $n$-th derivative in resum to $\Gamma_{0}^{(n)}(x^{\ast})$. The final set of skeleton diagrams derived from , the main result of this section therefore reads $$\begin{aligned} \Gamma(x^{\ast}) & = & \Gamma_{0}(x^{\ast})-W_{V}(j_{0})\label{eq:Gamma_skeleton_final}\\ & & +\sum_{n=2}\,\frac{\Gamma_{0}^{(n)}(x^{\ast})+\bar{\Gamma}^{(n)}(x_{1})}{n!}\,(-W_{V}^{(1)}(j_{0}))^{n},\nonumber \end{aligned}$$ where we defined $\bar{\Gamma}^{(n)}(x_{1}):=\Gamma^{(n)}(x_{1})-\Gamma_{0}^{(n)}(x^{\ast})$, which contains all diagrams from and that have at least one interaction vertex, so which are of order $\mathcal{O}(\epsilon)$ or higher. The second rule follows from $W_{V}^{(1)}(j_{0})$ being of first order or higher, so that at the $n$-th order in the interaction, we have to consider at most the $n$-th term in the sum in . In this last term we must in addition drop the $\bar{\Gamma}$-term. The only term to consider at second order, for example, is $$\frac{1}{2!}\,W_{V}^{(1)}(j_{0})^{\T}\,\Gamma^{(2)}(x_{1})\,W_{V}^{(1)}(j_{0})=\frac{1}{2!}\,W_{V}^{(1)\T}(j_{0})\Gamma_{0}^{(2)}(x^{\ast})W_{V}^{(1)}(j_{0})+\mathcal{O}\left(\epsilon^{3}\right).$$ This term cancels the contributions of the diagrams b) and c) in . However, that all reducible diagrams cancel can be seen more systematically by the constructive arguments given in . But eq. is useful to get a handle on the non-standard diagrams. If we want to know all such diagrams of a certain order $k$ in the interaction, we first assign the number of interactions to be contained in each factor in any term $\Gamma^{(n)}(x_{1})/n!\,\left(-W_{V}^{(1)}(j_{0})\right)^{n}$, $2\leq n\leq k$ so that they sum up to $k$. Assigning $0$ interaction vertices to $\Gamma^{(n)}$ reduces it to $\Gamma_{0}^{(n)}(x^{\ast})$, as said above. For non-zero numbers of vertices in $\bar{\Gamma}^{(n)}$, we decompose the latter via and into an expression only containing full cumulants and $\Gamma_{0}^{\left(2\right)}\left(x^{\ast}\right)$ and distribute again the number of interactions assigned to $\bar{\Gamma}^{(n)}$ among these cumulants. Each full cumulant is then broken down into all its perturbative expansion, for which the standard Feynman rules hold according to the linked cluster theorem (see ). Throughout this expansion, we skip all diagrams that are also contained in $-W_{V}$, because these are just the reducible ones that we know to cancel each other. Another insight we gain by in combination with is that non-standard diagrams are not necessarily irreducible - in fact, they can even be reducible in the sense that they fall into two parts both containing interactions if we cut a second-order cumulant. We show an example for this phenomenon in a toy model at the end of . ### Summary of Feynman rules for the non-Gaussian case We now summarize the algorithmic rules derived from the above observations to obtain $\Gamma$: 1. \[enu:Calculate-Gamma0\]Calculate $\Gamma_{0}(x^{\ast})=\sup_{j}\,j^{\T}x^{\ast}-W_{0}(j)$ explicitly by finding $j_{0}$ that extremizes the right hand side. At this order $g_{0}=0$. 2. \[enu:recursion\]At order $k$ in the perturbation expansion: 1. \[enu:connected\]add all irreducible graphs in the sense of the definition above that have $k$ vertices; 2. \[enu:add-all-irreducible\]add all graphs containing derivatives $\Gamma_{0}^{(n)}$ as connecting elements that cannot be reduced to the form of a graph contained in the expansion of $W_{V}(j_{0})$; the graphs left out are the counterparts of the reducible ones in $W_{V}(j_{0})$. The topology and combinatorial factors of these non-standard contributions are generated iteratively by from the previous order in perturbation theory; this iteration, by construction, only produces diagrams, where at least two legs of each $\Gamma_{0}^{(n)}$ connect to a third or higher order cumulant. We can also directly leave out diagrams, in which a subdiagram contained in $W_{V}$ is connected to the remainder of the diagram by a single leg of an interaction vertex. Alternatively, the additional diagrams can be constructed directly for any order $k$ by instantiating the skeleton diagrams generated by to the desired order, as explained in . Note that here diagrams may appear in which $\Gamma^{(n>2)}$ couples to less than two third or higher order cumulants. After decomposing all $\Gamma^{(n>2)}$ into $\Gamma_{0}^{\left(2\right)}$ and diagrams from $W^{\left(n\right)}$ by and , we can discard diagrams, in which at least one subdiagram from $W_{V}$ forms a leave that is connected by a single link of an interaction vertex; these cancel by the same argument as given in the iterative approach. 3. \[enu:assign-the-factor\]assign the factor $\frac{\epsilon^{k}}{r_{1}!\cdots r_{l+1}!}$ to each diagram with $r_{i}$-fold repeated occurrence of vertex $i$; assign the combinatorial factor that arises from the possibilities of joining the connecting elements as usual in Feynman diagrams (see examples below and e.g. [@Amit84]); 4. \[enu:j0\_as\_x\_star\]express the $j$-dependence of the $n$-th cumulant ${\savebox{\@brx}{$\m@th{\langle}$} \mathopen{\copy\@brx\kern-0.5\wd\@brx\usebox{\@brx}}}x^{n}{\savebox{\@brx}{$\m@th{\rangle}$} \mathclose{\copy\@brx\kern-0.5\wd\@brx\usebox{\@brx}}}(x^{\ast})$ in all terms by the first cumulant $x^{\ast}={\savebox{\@brx}{$\m@th{\langle}$} \mathopen{\copy\@brx\kern-0.5\wd\@brx\usebox{\@brx}}}x{\savebox{\@brx}{$\m@th{\rangle}$} \mathclose{\copy\@brx\kern-0.5\wd\@brx\usebox{\@brx}}}=W_{0}^{(1)}(j_{0})$; this can be done, for example, by inverting the last equation or directly by using $j_{0}=\Gamma_{0}^{(1)}(x^{\ast})$; express the occurrence of $\Gamma_{0}^{(2)}$ by its explicit expression. This set of Feynman rules constitutes the central result of our work. Note that the rules include the well-known Gaussian case, because irreducibility in the here-defined sense is identical to one-line-irreduciblity in the expansion around a Gaussian theory: every leg of an interaction vertex necessarily connects to a line there. Non-standard diagrams hence cannot appear in this case. The Feynman rules hold for any order $k$. So one may compute corrections at order $k$ directly without having computed any lower order. The combinatorial factor is fixed unambiguously, too. To get an intuitive understanding of the cancellation, it is still useful to illustrate the recursive construction of graphs in in the application to a minimal, non-Gaussian setting. In we demonstrate the application to systems of Ising spins, recovering the TAP approximation [@Thouless77_593], the high temperature expansion [@Georges91], and the Plefka expansion [@Plefka82_1971]. It turns out that the recursive algorithmic procedure leads to a dramatic decrease of required computations. Illustrative example for the graphical rules\[sub:Illustrative-example\] ------------------------------------------------------------------------ As a first example let us consider a zero-dimensional field theory, that is to say, a probability distribution of a scalar variable. By assumption, $S_{0}$ constitutes the solvable theory, so that $W_{0}(j)$ can be determined exactly by . We here illustrate the method for a solvable theory that has non-vanishing cumulants of orders one, two, and three, hence $W_{0}$ is a polynomial of order three $$\begin{aligned} W_{0}(j) & = & \feyn{!c{0}}=\sum_{n=1}^{3}\,{\savebox{\@brx}{$\m@th{\langle}$} \mathopen{\copy\@brx\kern-0.5\wd\@brx\usebox{\@brx}}}x^{n}{\savebox{\@brx}{$\m@th{\rangle}$} \mathclose{\copy\@brx\kern-0.5\wd\@brx\usebox{\@brx}}}_{0}\,\frac{j^{n}}{n!},\end{aligned}$$ where the cumulants ${\savebox{\@brx}{$\m@th{\langle}$} \mathopen{\copy\@brx\kern-0.5\wd\@brx\usebox{\@brx}}}x^{n}{\savebox{\@brx}{$\m@th{\rangle}$} \mathclose{\copy\@brx\kern-0.5\wd\@brx\usebox{\@brx}}}_{0}$ with $n\in\{1,2,3\}$ appear as Taylor coefficients. Its graphical representation is shown in b. As perturbation we assume a three point vertex $\epsilon V(x)=\epsilon\,x^{3}$, shown in a. According to step () we determine $\Gamma_{0}(x^{\ast})$ by , which amounts to the calculation of the derivative $\partial_{j}W_{0}(j_{0})-x^{\ast}=0$ which determines $j_{0}(x^{\ast})$. We now calculate the diagrams recursively according to step (). At order $l=0$ we have $g_{0}=0$. At first order $l=1$, we therefore get from the graphs shown in . [test8]{} $1 \cdot$ $3 \cdot$ $1 \cdot$ Each graph has a single vertex, thus we get a factor $\frac{\epsilon}{1!}$. The combinatorial factors for each graph are stated explicitly above. So the diagrams in $\Gamma_{V}$ at first order are identical to all connected diagrams with a single vertex; in this regard, the expansion is identical to the well known Gaussian case, except that the diagram c) would not appear in the absence of third order cumulants of the solvable theory. At order $l=2$, the graphs in $W_{0}\cup g_{1}$ contain those that have been produced in the previous step. The step therefore produces additional diagrams out of these components, some of which are shown in . [test9a]{} $3 \cdot 3 \cdot 2 \cdot$ $3 \cdot 3 \cdot$\ [c)]{} $3 \cdot 3 \cdot$ The diagrams a) and b) in are composed of the diagram a) in contained in $g_{1}$ and one additional vertex. The diagram c) in is composed of diagram b) in ; its combinatorial factor $3\cdot3$ is the combined factor of the first order diagram and the factor $3$ due to three possibilities to select a leg of the vertex. We here skipped further diagrams; in particular all diagrams, that are generated by the first order diagram c) in . The line produces additional diagrams with negative sign. We here indicate the simplification $\Feyn{gpgfcf}=1$ by an overbrace. Some diagrams are shown in . [test10]{} $- 3 \cdot 3 \cdot$\ [b)]{} $- 3 \cdot 3 \cdot$\ [c)]{} $- 3 \cdot 3 \cdot 1 \cdot$\ We observe that the diagram a) in cancels the diagram b) in . This cancellation is of the ordinary type; the diagram b) in is one-line irreducible in the original sense: its two components are connected by a second order cumulant, which would be denoted by a line $\Feyn{ff0}$ in the original graphical language of Feynman diagrams; here denoted as $\Feyn{fcf}$. An example of the more general cancellation appears between diagram b) in and c) in : the two components in c) are not connected by a second order cumulant, hence the diagram is not one-line irreducible in the original sense, but it is reducible in the sense we defined above. We see that it is canceled by b), consistent with the general proof. If we want to obtain the diagrams of the next order, note that we have to keep reducible diagrams, because the cancellation occurs only after setting $j=\Gamma_{0}^{\left(1\right)}\left(x^{\ast}\right)$. The diagram c) produced by cannot be generated by , because the connecting proper (effective) vertex $\Gamma_{0}^{(2)}=\Feyn{gpg}$ connects to two third order cumulants. It is generated in two ways; with two $\Gamma^{\left(1\right)}\left(x^{\ast}\right)$-components or only one of them. These two contributions have opposite signs, but different prefactors, therefore they do not cancel (see also for details). We now derive the appearance of this latter diagram by the application of the method using skeleton diagrams. The relevant skeleton diagram is the one that appears at second order on the Taylor expansion in the second line of $$\begin{aligned} & & \frac{1}{2!}\,\Diagram{cfpfc} \quad.\label{eq:skeleton_order_two}\end{aligned}$$ We now instantiate this diagram with perturbative expansions of the desired order, so that the resulting diagram is of order two: The two point vertex function, by Dyson’s equation to lowest order is $\Gamma^{(2)}(x_{1})=\Gamma_{0}^{(2)}(x^{\ast})+\ldots$. The leaves $-W_{V}^{(1)}$ attached to either side, to first order, contain contributions of the form [First\_order\_cum\_for\_skeleton]{} $$\begin{aligned} \parbox{30mm}{ \begin{fmfgraph}(70,30) \fmfpen{0.75thin} \fmfleft{i1} \fmfright{o1} \fmf{plain, tension=3.5}{i1,v3} \fmfv{d.s=circle, d.filled=empty}{v3} \fmf{plain, left=.7}{v3,v4,v3} \fmf{plain,tension=3.5}{v4,o1} \fmfv{d.s=circle, d.filled=empty}{o1} \end{fmfgraph} }+\dots, \end{aligned}$$ where all elements on the right are cumulants of the unperturbed system and bare interaction vertices. The combinatorial factor of this latter sub-diagram is computed with the standard rules for connected diagrams, which yields a factor $3$ from the possible ways of attaching the first order cumulant to the interaction vertex. There is only a single possibility to attach the additional external leg to the third order cumulant and only a single way to join the elements in the skeleton diagram in , besides the explicit factor $1/2!$. Finally, each remaining diagram must be interpreted algebraically by steps () and (). These steps are straight forward here. We note that this step is possible at all in the technique using the skeleton diagrams, because all cumulants appearing are $W_{0}^{(n)}(j_{0})=W_{0}^{(n)}(\Gamma_{0}^{(1)}(x^{\ast}))$ and in Dyson’s equation in addition only $\Gamma_{0}^{(2)}(x^{\ast})$ appears; so all elements, by assumption of $S_{0}$ being the solvable theory, are known. There is yet another observation that we can make in the framework of of this toy model, namely that there are reducible non-standard diagrams that contribute to $\Gamma_{V}$. One example is the following fourth-order cumulant emerging from the $n=2$ term in the sum in and using the first-order correction from (with second order correction in the interaction for $W_{V}^{\left(2\right)}$): [Reducible\_nonstandard]{}\ Application to the Ising model\[sub:Application-to-Ising\] ---------------------------------------------------------- We here illustrate the method on a simple system, the classical Ising model with the action $$\begin{aligned} S(x) & = & \frac{\epsilon}{2}\,x^{\T}Jx+j^{\T}x,\label{eq:action_Ising}\end{aligned}$$ where $x\in\{-1,1\}^{N}$ is a vector of Ising spins and $J$ a symmetric matrix with $J_{ii}=0$. Note that there is no additional constraint on $J$; it could, for example, be drawn from a random distribution, as done in spin-glass models. In particular there is no restriction to nearest-neighbour interactions. We are here interested in the weak coupling limit with small $\epsilon$. There are essentially two different ways to arrive at an approximation of the effective action $\Gamma(x^{\ast})$. The first approach represents the pairwise coupling term as the result of a Gaussian average over auxiliary variables (see e.g. [@NegeleOrland98 Chapter 4.3 eq. 4.50a], [@Vasiliev98 chapter (5.2), eq. (5.49)], [@Sommers87_1268 eq. (4), (5)], or [@Horwitz61]). This Hubbard-Stratonovich transform reduces the calculation of $\Z$ to the summation of $n$-th moments of the Gaussian employing Wick’s theorem, weighted by the Taylor coefficients of $V(x)=\ln\sum_{x_{i}}e^{j_{i}x_{i}}=\sum_{i}\,\ln\,2\cosh\,j_{i}$; the latter play the role of vertices here. Consequently, standard Feynman diagrammatic rules apply. This approach is only applicable to this model because the interaction is pairwise; interactions of higher order cannot be written by help of Gaussian auxiliary fields. The second way, which we will follow here, considers as the solvable part the single-spin term of the action $$\begin{aligned} S_{0}(x) & = & j^{\T}x;\qquad W_{0}(j)=\sum_{i}\,\ln\,2\,\cosh(j_{i}),\label{eq:S0_W0_Ising}\end{aligned}$$ which is directly summable, yielding a cumulant-generating function $W_{0}$ whose decomposition into a sum over individual sites shows their statistical independence. For each $i$, the solvable theory therefore has the infinitely many non-vanishing cumulants of a binary variable. The interaction is in turn treated as the perturbation $$\begin{aligned} V(x) & = & \frac{\epsilon}{2}\,x^{\T}Jx.\label{eq:S_int_Ising}\end{aligned}$$ This approach has been followed in quite a number of variations over decades (see e.g. [@Vasiliev98 section 5.2, eq. 5.47], [@Bloch65; @Vasiliev74; @Bogolyubov1976; @Plefka82_1971; @Yedidia90; @Georges91; @Nakanishi97_8085; @Tanaka98_2302], [@Opper01 chapter 3]); these methods agree inasmuch as they all perform an expansion of $\Gamma$, where is treated as a perturbation and as the solvable part. In particular this approach is identical to Plefka’s method [@Plefka82_1971]. Both routes of course lead to identical results. To second order in $\epsilon$ one obtains the mean-field theory presented by Thouless, Anderson, and Palmer (TAP) [@Thouless77_593] as a “fait acompli”, without proof, but mentioning a previous diagrammatic derivation. Indeed, Vasiliev and Radzhabov ([@Vasiliev74; @Vasiliev75], summarized in [@Vasiliev98 section 6.3.1, 6.3.2, 6.3.4]) had derived this result diagrammatically with the help of the analog to the Dyson Schwinger equation for the effective action [@Dahmen67] before (see [@Vasiliev98 section 5.2] for a review of this method). If one performed a perturbation expansion directly on the level of $\Z$ $$\begin{aligned} \Z(j) & = & \exp(V(\partial_{j}))\,\exp(W_{0}(j))=\big\langle\sum_{n=0}^{\infty}\frac{1}{n!}(\frac{\epsilon}{2}\,x^{\T}Jx)^{n}\big\rangle_{0},\label{eq:brute_force}\end{aligned}$$ with $\langle f(x)\rangle_{0}=\sum_{x}f(x)\,\exp(j^{\T}x),$ one would quickly obtain unwieldy expressions; but most of these terms cancel by the subsequent transformations $\Z\stackrel{\ln}{\rightarrow}W\stackrel{\mathcal{L}}{\rightarrow}\Gamma$ in the final result, making a direct expansion of $\Gamma$ desirable. The calculation of $\Gamma$ at higher orders in $\epsilon$ by this direct approach has prompted for computer algebra systems [@Nakanishi97_8085]. Georges and Yedidia [@Yedidia90; @Georges91] developed a sequence of shortcuts and tricks for this problem by which they managed the approximation up to forth order (reviewed in [@Opper01 chapter 3]). As mentioned, $\Gamma$ has been derived diagrammatically by Vasiliev and Radzhabov [@Vasiliev74 eq. (21)]. Using Dyson-Schwinger equations, they explicitly presented all orders up to and including the third. We here chose this model to exemplify the application of the general framework exposed in previous sections. For one, because it allows the comparison to a wealth of methods developed over decades, as mentioned above. And also because the Ising model has proven useful in many other applications than micromagnetism, for example artificial neural networks [@Hopfield82]. Especially the TAP-approximation and its higher order corrections are employed to derive analytical approximations [@Gabrie15] to contrastive divergence [@CarreiraPerpinan05], a learning rule commonly employed to train restricted Boltzmann machines [@Smolensky86; @Hinton2006_504]. Another major field, in which the Ising model is applied, are inference problems, where the inverse Ising problem has to be solved, that is, the sources $h_{i}$ and couplings $J_{ij}$ have to be computed for given means and pairwise covariances. Depending on the problem, the spins represent activities of neural units [@Tkacik06; @Roudi09a; @Hertz11], active or inactive genes in the case of gene regulatory networks or participants in financial markets (see [@Nguyen17] for an excellent review of the inverse Ising problem). The most natural quantity for this type of problem would be the second Legendre transform, the usual entropy of the multi-variate binary distribution for given first two moments, because the couplings and sources turn out to be the extrema of this quantity. We present an iterative method to compute the second Legendre transform in . Still, the first Legendre seems to be commonly used also for the inverse problem [@Kappen98; @Tanaka98_2302]. The reason for this might be that calculating the second Legendre transform is technically challenging; nevertheless, specific approximations exist that are valid for small correlations, first computed by a technique specialized to the Ising model [@Sessak09], later obtained by techniques borrowed from the functional renormalization group [@Jacquin16], using the Wetterich equation [@WETTERICH93_90].. An extension to more than pairwise coupling is desirable in these fields [@Roudi09; @Schneidman06_1007; @Tkacik14_e1003408]. In particular the Hubbard-Stratonovich method mentioned above would not work in these cases, while the here proposed approach does. Performing the here presented diagrammatic expansion, we observe a drastic decrease of computations and their transparent organization by diagrams. The calculation up to fourth order indeed takes only minor effort. We start with the unperturbed theory, given by , in which all spins decouple and the cumulants up to fourth order read $$\begin{aligned} {\savebox{\@brx}{$\m@th{\langle}$} \mathopen{\copy\@brx\kern-0.5\wd\@brx\usebox{\@brx}}}x_{i}{\savebox{\@brx}{$\m@th{\rangle}$} \mathclose{\copy\@brx\kern-0.5\wd\@brx\usebox{\@brx}}}_{0} & = & \left\langle x_{i}\right\rangle _{0}=m_{i}:=\tanh\left(j_{i}\right)\\ {\savebox{\@brx}{$\m@th{\langle}$} \mathopen{\copy\@brx\kern-0.5\wd\@brx\usebox{\@brx}}}x_{i}x_{j}{\savebox{\@brx}{$\m@th{\rangle}$} \mathclose{\copy\@brx\kern-0.5\wd\@brx\usebox{\@brx}}}_{0} & = & \delta_{ij}\left(1-m_{i}^{2}\right)\\ {\savebox{\@brx}{$\m@th{\langle}$} \mathopen{\copy\@brx\kern-0.5\wd\@brx\usebox{\@brx}}}x_{i}x_{j}x_{k}{\savebox{\@brx}{$\m@th{\rangle}$} \mathclose{\copy\@brx\kern-0.5\wd\@brx\usebox{\@brx}}}_{0} & = & -\delta_{ij}\delta_{jk}2m_{i}\left(1-m_{i}^{2}\right)\\ {\savebox{\@brx}{$\m@th{\langle}$} \mathopen{\copy\@brx\kern-0.5\wd\@brx\usebox{\@brx}}}x_{i}x_{j}x_{k}x_{l}{\savebox{\@brx}{$\m@th{\rangle}$} \mathclose{\copy\@brx\kern-0.5\wd\@brx\usebox{\@brx}}}_{0} & = & -\delta_{ij}\delta_{jk}\delta_{kl}2\left(1-3m_{i}^{2}\right)\left(1-m_{i}^{2}\right).\end{aligned}$$ To zeroth order the Legendre transform of $W_{0}$ is the entropy of a binary variable $$\begin{aligned} \Gamma_{0}(m) & = & -\sum_{i}\frac{1+m_{i}}{2}\,\ln\left(\frac{1+m_{i}}{2}\right)+\frac{1-m_{i}}{2}\,\ln\left(\frac{1-m_{i}}{2}\right),\label{eq:Gamma_0}\end{aligned}$$ where $(1\pm m_{i})/2$ are the probabilities for $x_{i}=\pm1$. The result follows from step () of the algorithm by a short calculation from ; the form is moreover clear, because the distribution maximizes the entropy and the Legendre transform to $\Gamma$ just fixes $j$ so that the mean is $\langle x_{i}\rangle=m_{i}$. To obtain corrections to we represent the $n$-th cumulant by an empty circle with $n$ legs and the $n$-th derivative of $\Gamma_{0}$ by a hatched circle with $n$ legs. An interaction $J_{ij}$ is denoted by an edge: [Example\_feyn]{} $$\begin{aligned} \parbox{25mm}{ \begin{fmfgraph}(25,25) \fmfpen{0.5thin} \fmftop{o1,o2,o3} \fmfbottom{u1,u2,u3} \fmf{plain}{u1,o2} \fmf{plain}{u3,o2} \fmfv{label=$i$, label.angle=90, label.dist=-10pt}{u1} \fmfv{label=$j$, label.angle=90, label.dist=-10pt}{u3} \end{fmfgraph} } \mkern-72mu := J_{ij}, \parbox{25mm}{ \begin{fmfgraph}(25,25) \fmfpen{0.5thin} \fmftop{o1,o2,o3,o4,o5} \fmfbottom{u1,u2,u3,u4,u5} \fmf{phantom}{u1,m1,o1} \fmf{phantom}{u3,m3,o3} \fmf{phantom}{u5,m5,o5} \fmf{plain}{m3,o4} \fmf{plain}{m3,o5} \fmf{plain,tension=0}{m3,m5} \fmf{phantom, tension=10}{m1,m3} \fmf{plain,tension=2.5}{m3,u5} \fmfv{decor.shape=circle,decor.filled=empty, decor.size=6.5thin}{m3} \fmfv{label=$i_{1}$, label.angle=90, label.dist=4.5pt}{o4} \fmfv{label=$i_{2}$, label.angle=60, label.dist=2.pt}{o5} \fmfv{label=$i_{3}$, label.angle=20, label.dist=1.5pt}{m5} \fmfv{label=$...$, label.angle=0, label.dist=1.5pt}{u5} \end{fmfgraph} } \mkern-63mu := {\savebox{\@brx}{$\m@th{\langle}$} \mathopen{\copy\@brx\kern-0.5\wd\@brx\usebox{\@brx}}}x_{i_{1}}x_{i_{2}}x_{i_{3}}...{\savebox{\@brx}{$\m@th{\rangle}$} \mathclose{\copy\@brx\kern-0.5\wd\@brx\usebox{\@brx}}}, \; \parbox{25mm}{ \begin{fmfgraph}(25,25) \fmfpen{0.5thin} \fmftop{o1,o2,o3,o4,o5} \fmfbottom{u1,u2,u3,u4,u5} \fmf{phantom}{u1,m1,o1} \fmf{phantom}{u3,m3,o3} \fmf{phantom}{u5,m5,o5} \fmf{wiggly}{m3,o4} \fmf{wiggly}{m3,o5} \fmf{wiggly, tension=0}{m3,m5} \fmf{phantom, tension=10}{m1,m3} \fmf{wiggly,tension=2.5}{m3,u5} \fmfv{decor.shape=circle,decor.filled=shaded, decor.size=6.5thin}{m3} \fmfv{label=$i_{1}$, label.angle=90, label.dist=4.5pt}{o4} \fmfv{label=$i_{2}$, label.angle=60, label.dist=2.pt}{o5} \fmfv{label=$i_{3}$, label.angle=20, label.dist=1.5pt}{m5} \fmfv{label=$...$, label.angle=0, label.dist=1.5pt}{u5} \end{fmfgraph} } \mkern-63mu := \Gamma^{\left(n\right)}_{i_{1},i_{2},i_{3},...}.\\ \nonumber \end{aligned}$$ To evaluate the diagrams, we will only need the second derivative $\Gamma_{0}^{\left(2\right)}$, either directly given by differentiating twice or by using the relation $W_{0}^{\left(2\right)}\Gamma_{0}^{\left(2\right)}=1$. Both yields $\Gamma_{0,ij}^{\left(2\right)}=\frac{\delta_{ij}}{1-m_{i}^{2}}$. Within this language, the perturbative corrections up to the third order to be added to are readily constructed from steps \[enu:recursion\]-\[enu:j0\_as\_x\_star\] at the end of : [Gamma\_first\_three]{} $$\begin{aligned} \parbox{25mm}{ \begin{fmfgraph}(25,25) \fmfpen{0.5thin} \fmftop{o1,o2,o3} \fmfbottom{u1,u2,u3} \fmf{plain}{u1,o2} \fmf{plain}{u3,o2} \fmfv{decor.shape=circle,decor.filled=empty, decor.size=6.5thin}{u1,u3} \end{fmfgraph} } &\mkern-60mu= \frac{1}{2}\sum_{i\neq j} J_{ij}m_{i}m_{j}&\\ \mkern-40mu \parbox{25mm}{ \begin{fmfgraph}(75,25) \fmfpen{0.5thin} \fmftop{o1,o2,o3,o4,o5} \fmfbottom{u1,u2,u3,u4,u5} \fmf{phantom}{u1,v1,o3} \fmf{plain}{v1,o3} \fmf{phantom}{o1,v1,u3} \fmf{plain}{v1,u3} \fmf{phantom}{u3,v2,o5} \fmf{plain}{u3,v2} \fmf{phantom}{o3,v2,u5} \fmf{plain}{o3,v2} \fmfv{decor.shape=circle,decor.filled=empty, decor.size=6.5thin}{v1,v2} \end{fmfgraph} } & \mkern-60mu = \frac{1}{2!2^{2}} 2 \sum_{i\neq j} J_{ij}^{2}\left(1-m_{i}^{2}\right)\left(1-m_{j}^{2}\right) \label{Def_second_oder_diagram}&\\ \mkern-40mu \parbox{25mm}{ \begin{fmfgraph}(75,25) \fmfpen{0.5thin} \fmftop{o1,o2,o3} \fmfbottom{u1,u2,u3} \fmf{phantom,tension=100}{u1,dl,v1,o2} \fmf{plain}{dul,v1,o2} \fmf{phantom,tension=100}{u3,dr,v2,o2} \fmf{plain}{dur,v2,o2} \fmf{phantom}{u1,dul,u2,dur,u3} \fmf{plain}{dul,u2,dur} \fmf{phantom,tension=0.5}{dl,dum,dr} \fmf{phantom,tension=1}{dum,u2} \fmfv{decor.shape=circle,decor.filled=empty, decor.size=6.5thin}{v1,v2,u2} \end{fmfgraph} } & \mkern-60mu = \frac{1}{3!2^{3}} 2^{3} \sum_{i\neq j\neq k\neq i} J_{ij} J_{jk}J_{ki}\left(1-m_{i}^{2}\right)\left(1-m_{j}^{2}\right)\left(1-m_{k}^{2}\right)\label{Def_third_oder_ring_diagram}&\\ \parbox{25mm}{ \begin{fmfgraph}(25,25) \fmfpen{0.5thin} \fmftop{o1,o2,o3} \fmfbottom{u1,u2,u3} \fmf{phantom}{u2,du,do,o2} \fmf{phantom,tension=2}{du,do,o2} \fmf{plain}{u1,o2} \fmf{plain}{u3,o2} \fmf{plain}{u1,do} \fmf{plain}{u3,do} \fmf{plain}{u1,du} \fmf{plain}{u3,du} \fmfv{decor.shape=circle,decor.filled=empty, decor.size=6.5thin}{u1,u3} \end{fmfgraph} } & \mkern-60mu = \frac{1}{3!2^{3}} 2^{2} \sum_{i\neq j} J_{ij}^{3} \left(-2m_{i}\right)\left(1-m_{i}^{2}\right)\left(-2m_{j}\right)\left(1-m_{j}^{2}\right).\label{Def_third_oder_Citroen_diagram}& \end{aligned}$$ For the third order diagrams, the symmetry factors $2^{3}$ and $2^{2}$ are noted separately in front of the respective terms. For all diagrams, they are determined as usual for Feynman diagrams (see ). Until this order only the first sort of ordinary diagrams composed of vertices and cumulants contribute. All diagrams are irreducible in the general sense defined above. Going to fourth order, we first consider the four diagrams that are formed out of vertices and cumulants alone. These would also contribute to an expansion of $W$. [Gamma\_fourth\_ordinary]{} $$\begin{aligned} \mkern-150mu \parbox{25mm}{ \begin{fmfgraph}(75,25) \fmfpen{0.5thin} \fmftop{o1,o2,o3} \fmfbottom{u1,u2,u3} \fmf{phantom}{u1,dul,dml,v1,o2} \fmf{plain}{dml,v1,o2} \fmf{phantom}{u2,v2,dml,dol,o1} \fmf{plain}{u2,v2,dml} \fmf{phantom}{u2,v3,dmr,dor,o3} \fmf{plain}{u2,v3,dmr} \fmf{phantom}{u3,dur,dmr,v4,o2} \fmf{plain}{dmr,v4,o2} \fmfv{decor.shape=circle,decor.filled=empty, decor.size=6.5thin}{v1,v2,v3,v4} \end{fmfgraph} } & \mkern-30mu = \frac{S_{\mathrm{R}}}{4!2^{4}}\sum_{i\neq j\neq k \neq l\neq i} J_{ij}J_{jk}J_{kl}J_{li} \left(1-m_{i}^{2}\right)\left(1-m_{j}^{2}\right)\left(1-m_{k}^{2}\right)\left(1-m_{l}^{2}\right) \label{Fourth_order_ring}&\\ \mkern-150mu\parbox{25mm}{ \begin{fmfgraph}(75,25) \fmfpen{0.5thin} \fmftop{o1,o2,o3} \fmfbottom{u1,u2,u3} \fmf{phantom,tension=100}{u1,dl,v1,o2} \fmf{plain}{dul,v1,o2} \fmf{phantom,tension=100}{u3,dr,v2,o2} \fmf{plain}{dur,v2,o2} \fmf{phantom}{u1,dul,u2,dur,u3} \fmf{plain}{dul,u2,dur} \fmf{phantom,tension=0.5}{dl,dum,dr} \fmf{phantom,tension=1}{dum,u2} \fmf{plain}{v2,dum} \fmf{plain}{v1,dum} \fmfv{decor.shape=circle,decor.filled=empty, decor.size=6.5thin}{v1,v2,u2} \end{fmfgraph} } &\mkern-30mu =\frac{S_{\mathrm{TM}}}{4!2^{4}}\sum_{i\neq j\neq k\neq i} J_{ik}J_{ij}^2J_{kj} \left(-2m_{i}\right)\left(1-m_{i}^{2}\right)\left(-2m_{j}\right)\left(1-m_{j}^{2}\right)\left(1-m_{k}^{2}\right)\label{Fourth_order_TM}&\\ \mkern-150mu\parbox{25mm}{ \begin{fmfgraph}(75,25) \fmfpen{0.5thin} \fmftop{o0,o1,o2,o3,o4,o5,o6} \fmfbottom{u0,u1,u2,u3,u4,u5,u6} \fmf{phantom}{u1,v1,o3} \fmf{plain}{v1,o3} \fmf{phantom}{o1,v1,u3} \fmf{plain}{v1,u3} \fmf{phantom}{u3,v2,o5} \fmf{plain}{u3,v2} \fmf{phantom}{o3,v2,u5} \fmf{plain}{o3,v2} \fmf{phantom}{u0,v1,dummy1,o4} \fmf{plain}{v1,dummy1} \fmf{phantom}{u4,dummy2,v1,o0} \fmf{plain}{dummy2,v1} \fmf{phantom}{u2,dummy2,v2,o6} \fmf{plain}{dummy2,v2} \fmf{phantom}{u6,v2,dummy1,o2} \fmf{plain}{v2,dummy1} \fmfv{decor.shape=circle,decor.filled=empty, decor.size=6.5thin}{v1,v2} \end{fmfgraph} } & \mkern-30mu = \frac{S_{\mathrm{dC}}}{4!2^{4}}\sum_{i\neq j} J_{ij}^{4}\left(-2\right)\left(1-3m_{i}^{2}\right)\left(1-m_{i}^{2}\right)\left(-2\right)\left(1-3m_{j}^{2}\right)\left(1-m_{j}^{2}\right).\label{Fourth_order_Citroen}& \end{aligned}$$ In the appendix, , we show that the symmetry factors are given by $\SR=48$, $\ST=96$ and $\SdC=8$. The fourth standard diagram is shown below. It is similar to the first and second non-standard contribution. To indicate the origin of the different contributions, we denote sub-graphs originating from $\Gamma^{\left(1\right)}$ by filled circles - they are, however, translated in the same way as the empty ones: [Gans]{} $$\begin{aligned} \label{Fourth_order_extraordinary} & \underbrace{\mkern-18mu \parbox{20mm}{ \begin{fmfgraph}(75,25) \fmfpen{0.5thin} \fmftop{o1,o2,o3,o4,o5,o6,o7} \fmfbottom{u1,u2,u3,u4,u5,u6,u7} \fmfright{r1} \fmfleft{l1} \fmf{phantom}{v3,r1} \fmf{phantom}{l1,v1} \fmf{phantom}{u1,v1,o3} \fmf{plain}{v1,o3} \fmf{phantom}{o1,v1,u3} \fmf{plain}{v1,u3} \fmf{plain}{u3,v2,o5} \fmf{plain}{o3,v2,u5} \fmf{phantom}{u5,v3,o7} \fmf{plain}{u5,v3} \fmf{phantom}{o5,v3,u7} \fmf{plain}{o5,v3} \fmfv{decor.shape=circle,decor.filled=empty, decor.size=6.5thin}{v1,v2,v3} \end{fmfgraph} }\quad}_{a} - \underbrace{\mkern-18mu \parbox{20mm}{ \begin{fmfgraph}(100,25) \fmfpen{0.5thin} \fmftop{o1,o2,o3,o4,o5,o6,o7,o8,o9} \fmfbottom{u1,u2,u3,u4,u5,u6,u7,u8,u9} \fmfright{r1} \fmfleft{l1} \fmf{phantom}{v3,r1} \fmf{phantom}{l1,v1} \fmf{phantom}{u1,v1,o3} \fmf{plain}{v1,o3} \fmf{phantom}{o1,v1,u3} \fmf{plain}{v1,u3} \fmf{phantom}{u3,v2,o5} \fmf{plain}{u3,v2} \fmf{phantom}{o3,v2,u5} \fmf{plain}{o3,v2} \fmf{phantom}{u5,v4,o7} \fmf{plain}{v4,o7} \fmf{phantom}{u7,v4,o5} \fmf{plain}{u7,v4} \fmf{phantom}{u7,v3,o9} \fmf{plain}{u7,v3} \fmf{phantom}{o7,v3,u9} \fmf{plain}{o7,v3} \fmf{plain}{v2,g1} \fmf{wiggly}{g1,vv1} \fmf{wiggly}{vv1,g2} \fmf{plain}{g2,v4} \fmfv{decor.shape=circle,decor.filled=empty, decor.size=6.5thin}{v1,v2} \fmfv{decor.shape=circle,decor.filled=full, decor.size=6.5thin}{v3,v4} \fmfv{decor.shape=circle,decor.filled=shaded, decor.size=6.5thin}{vv1} \end{fmfgraph} } \quad \quad \quad}_{b} \; + \underbrace{\mkern-18mu \parbox{20mm}{ \begin{fmfgraph}(125,25) \fmfpen{0.5thin} \fmftop{o1,o2,o3,o4,o5,o6,o7,o8,o9,o10,o11} \fmfbottom{u1,u2,u3,u4,u5,u6,u7,u8,u9,u10,u11} \fmfright{r1} \fmfleft{l1} \fmf{phantom}{v3,r1} \fmf{phantom}{l1,v1} \fmf{phantom}{u1,v1,o3} \fmf{plain}{v1,o3} \fmf{phantom}{o1,v1,u3} \fmf{plain}{v1,u3} \fmf{phantom}{u3,v2,o5} \fmf{plain}{u3,v2} \fmf{phantom}{o3,v2,u5} \fmf{plain}{o3,v2} \fmf{phantom}{u7,v4,o9} \fmf{plain}{v4,o9} \fmf{phantom}{u9,v4,o7} \fmf{plain}{u9,v4} \fmf{phantom}{u9,v3,o11} \fmf{plain}{u9,v3} \fmf{phantom}{o9,v3,u11} \fmf{plain}{o9,v3} \fmf{plain}{v2,g1} \fmf{wiggly}{g1,vv1} \fmf{wiggly}{vv1,g2} \fmf{plain}{g2,v5} \fmf{plain}{v5,g3} \fmf{wiggly}{g3,vv2} \fmf{wiggly}{vv2,g4} \fmf{plain}{g4,v4} \fmfv{decor.shape=circle,decor.filled=empty, decor.size=6.5thin}{v5} \fmfv{decor.shape=circle,decor.filled=full, decor.size=6.5thin}{v1,v2,v3,v4} \fmfv{decor.shape=circle,decor.filled=shaded, decor.size=6.5thin}{vv1,vv2} \fmfv{label=$\overbrace{ \enskip\quad\quad }^{=1}$, label.angle=90, label.dist=4pt}{g3} \end{fmfgraph} } \quad\quad\quad\quad\quad\quad}_{c}&\\ \nonumber &=\frac{1}{2^{4}}\sum_{i\neq j\neq k} J_{ij}^{2}J_{jk}^{2}\\ \nonumber & \times \left[\frac{S_{a}}{4!}\left(1-m_{i}^{2}\right)\left(-2\right)\left(1-m_{j}^{2}\right)\left(1-3m_{j}^{2}\right)\left(1-m_{k}^{2}\right)\right.& \\ \nonumber & - \left(1-m_{i}^{2}\right) \frac{\left(-2m_{j}\right)\left(1-m_{j}^{2}\right)\left(-2m_{j}\right)\left(1-m_{j}^{2}\right)}{\left(1-m_{j}^{2}\right)} \left(1-m_{k}^{2}\right) \left. \left(\frac{S_{b}}{3!}-\frac{S_{c}}{2!}\right) \right] & \\ \nonumber \end{aligned}$$ With the symmetry factors given by $S_{a}=48$, $S_{b}=24$ and $S_{c}=4$ (see appendix, sec. for the derivation), we can add up the contributions of the three diagrams, which gives, leaving out factors that are equal in all diagrams for simplicity: $$\begin{aligned} \frac{1}{4!}48\left(-2\right)\left(1-3m_{j}^{2}\right)+\left(-\frac{1}{1!2!}8+\frac{1}{2}4\right)4m_{j}^{2} & = & -4\left(1-m_{j}^{2}\right).\end{aligned}$$ The last term looks like a second cumulant and, interestingly, this contribution is indeed exactly canceled by the contribution of the ring-diagram of fourth order (\[Fourth\_order\_ring\]) in the case that exactly one pair of indices, which belong to cumulants represented at opposite sites of the ring, are equal. This has to be so, as shown by Vasiliev and Radzhabov [@Vasiliev75], because, according to their terminology, diagram $a$ is a “nonstar graph”, that is, not a star graph - it decomposes into two parts by removing a cumulant (a “vertex” in their words). Note that the star graph property implies irreducibility, but not vice versa. As Vasiliev and Radzhabov have shown, only star diagrams contribute to the effective action of the Ising model, if we introduce so called compensating graphs. These graphs are constructed as follows: Take all star graphs of a certain order and draw contractions (in [@Vasiliev75] depicted by dashed lines) in all possible ways between circles that are not connected by an interaction and therefore represent indices that could take the same value (keep in mind that $J_{ii}=0$). If two or more circles are contracted, they are considered as a single circle, are associated with a common index and are translated as the product of all cumulants associated to this point. Then substract all compensating graphs that are non-star-graphs and sum up original and compensating graphs [@Vasiliev75 eq. (8)]. We may check that our result is consistent with this rule. In the ring-diagram (\[Fourth\_order\_ring\]), the only contraction that leads to a non-star graph is the one that identifies the indices of two opposite circles. This contribution therefore cancels from the final expression for the effective action, analogous to the summed up contribution of the diagrams $a$, $b$, and $c$. In our framework not only the double Citro[ë]{}n diagram (\[Fourth\_order\_Citroen\]) contributes to the sum over only two unequal indices, but also the ring diagram in the case that the indices of pairwise opposite circles and the diagrams $a$, $b$ and $c$ in the case that the indices of the outmost circles are equal (see for details). In summary, the general method that we present here is consistent with the specific result known for the Ising model. In total, we obtain the following expression for the effective action in an Ising model with coupling $J_{ij}$: $$\begin{aligned} & -\Gamma\left(\boldsymbol{m}\right)\label{eq:Effective_action_Ising}\\ = & -\sum_{i}\frac{1+m_{i}}{2}\ln\left(\frac{1+m_{i}}{2}\right)+\frac{1-m_{i}}{2}\ln\left(\frac{1-m_{i}}{2}\right)\nonumber \\ + & \frac{1}{2}\epsilon\sum_{i\neq j}J_{ij}m_{i}m_{j}+\frac{1}{4}\epsilon^{2}\sum_{i\neq j}J_{ij}^{2}\left(1-m_{i}^{2}\right)\left(1-m_{j}^{2}\right)\nonumber \\ + & \frac{1}{6}\epsilon^{3}\sum_{i\neq j\neq k\neq i}J_{ij}J_{jk}J_{ki}\left(1-m_{i}^{2}\right)\left(1-m_{j}^{2}\right)\left(1-m_{k}^{2}\right)\nonumber \\ + & \frac{1}{3}\epsilon^{3}\sum_{i\neq j}J_{ij}^{3}m_{i}\left(1-m_{i}^{2}\right)m_{j}\left(1-m_{j}^{2}\right)\nonumber \\ + & \frac{1}{8}\epsilon^{4}\sum_{\overset{i\neq j\neq k\neq l\neq i,}{i\neq k,j\neq l}}J_{ij}J_{jk}J_{kl}J_{li}\left(1-m_{i}^{2}\right)\left(1-m_{j}^{2}\right)\left(1-m_{k}^{2}\right)\left(1-m_{l}^{2}\right)\label{eq:fourth_order_I}\\ + & \epsilon^{4}\sum_{i\neq j\neq k\neq i}J_{ik}J_{ij}^{2}J_{kj}m_{i}\left(1-m_{i}^{2}\right)m_{j}\left(1-m_{j}^{2}\right)\left(1-m_{k}^{2}\right)\label{eq:fourth_order_II}\\ - & \frac{1}{24}\epsilon^{4}\sum_{i\neq j}J_{ij}^{4}\left(1-m_{i}^{2}\right)\left(1-m_{j}^{2}\right)\left(1+3m_{i}^{2}+3m_{j}^{2}-15m_{i}^{2}m_{j}^{2}\right)\nonumber \end{aligned}$$ This calculation reproduces (with $\epsilon=\beta$) eq. (11) of Georges and Yedidia [@Georges91] and is in line with equations (9), (10), (14), and (15) in Nakanishi and Takayama [@Nakanishi97_8085] [^1] and the equations (B2)-(B5) in Jacquin and Rançon [@Jacquin16], who derived this result using the Wetterich equation [@WETTERICH93_90]. Discussion ========== We presented a systematic diagrammatic scheme to calculate the effective action for any problem that can be decomposed into a solvable part with known correlators and a perturbing part. We have proved that corrections are composed of two types of diagrams: - irreducible diagrams in a more general sense than in the Gaussian case: those that cannot be decomposed by detaching a single leg of a vertex; - diagrams of special form that are neither contained in the perturbation expansions of $\Z$ nor $W$; they are composed of sub-graphs that are connected by either a vertex $\Gamma_{0}^{(n)},\ n\geq2$ that, by at least two legs, connects to a third or higher order bare cumulant. This set of diagrams is found by instantiating a known sequence of skeleton diagrams to the desired order in perturbation theory. The appearance of the latter diagrams can be regarded a as generalization of the amputation in the Gaussian case, because the inverse propagator $\Gamma_{0}^{(2)}$ may attach to cumulants of the solvable problem at any order, not only to second order cumulants; only in the latter case it “amputates” the lines. The presented inductive proof in addition yields an iterative equation which allows the algebraic construction of all graphs and their combinatorial factors from elementary rules of calculus. One may wonder why we derived the recursion for $\Gamma$ based on the ideas of the proof of the linked cluster theorem (see , extending the proof in [@ZinnJustin96 Sec. 6.1.1] to the non-Gaussian case). It might seem more direct to modify the corresponding proof of one-line irreducibility to the non-Gaussian setting considered here. The latter, however, appeared impossible to us: The elegant proof by Zinn-Justin [@ZinnJustin96 section 6.5] rests on the assumption that the underlying theory is Gaussian; in their eq. 6.59 each line is disconnected in all possible ways and the result is shown to remain connected. The proof hence requires that the only connecting elements of the bare theory be lines; this is precisely the restriction we lift here. The proof by Weinberg [@Weinberg05_II section 16.1] shows elegantly that $W$ decomposes into tree graphs, whose vertices - which one could call “effective vertices” in this case - are generated by $\Gamma$. This statement remains of course true also in the non-Gaussian case, because $W\stackrel{\mathcal{L}}{\leftrightarrow}\Gamma$ form a pair of Legendre transforms. We use this fact to derive the decomposition into skeleton diagrams. This decomposition also follows directly from the reciprocity relation $1=\Gamma^{(2)}W^{(2)}$ of the Hessians [@ZinnJustin96 section 6.2]. The connecting elements in these trees are the full propagators $W^{(2)}$, which Dyson’s equation expresses in terms of bare propagators and the self energy. The reciprocity relation implies the recursion for the self-energy $\Gamma_{V}^{(2)}=-(\Gamma_{0}^{(2)}+\Gamma_{V}^{(2)})\,W_{V}^{(2)}\,\Gamma_{0}^{(2)}.$ Since $W_{V}$ is composed of all connected graphs, including those containing higher order cumulants of $W_{0}$, we see that the terms $\cdots\Gamma_{0}^{(2)}W_{V}\Gamma_{0}^{(2)}\cdots$ appearing in the iteration produce those unusual terms of the form $W_{0}^{(n\ge2)}\Gamma_{0}^{(2)}W_{0}^{(m\ge2)}$ that we found in the general expansion (Eq. \[eq:reducible\_general-1\]). An alternative approach is the analog of the Dyson Schwinger equation for the effective action [@Dahmen67; @Vasiliev73; @Vasiliev74a], reviewed in [@Vasiliev98 sec. 1.8, 6.2]. It leads to equations of motion for $\Gamma$ that enable an iterative expansion, for example in the interaction strength. An iterative solution leads to a proof of the 1PI property in the Gaussian case [@Vasiliev98 sec. 1.8.3]. We believe that the latter approach could be generalized to obtain the same results as presented here. The equation of motion can be derived from the invariance of the integral measure with respect to translations [@Vasiliev98 sec. 1.8.3], but also extends to problems on discrete state spaces, such as the Ising model [@Vasiliev74; @Vasiliev75; @Bogolyubov1976; @Vasiliev98]. From this equation of motion, eq. (6.136) in [@Vasiliev98] is derived, similar to our with the difference that Vasiliev expands in terms of unperturbed cumulants and not full vertices. Probably, both equations are suitable starting points to rederive our results in . The Feynman rules would then be deduced from this step. Possibly these approaches could render the taxonomy of non-standard diagrams clearer. However, it will unlikely be as simple as in the Gaussian case because, as pointed out, reducible non-standard diagrams do not necessarily cancel. We leave this question for future work. In any case, more refined diagrammatic rules previously had to be obtained in a case by case manner so far, depending on the problem at hand (compare also the last paragraph of [@Vasiliev98 chap. 6.3.1.]). The algorithm presented in , in contrast, is the same for any model. One notes that eq. (21) in the work by Vasiliev and Radzhabov [@Vasiliev74] is identical to the third order derivation recovered much later by other means [@Plefka82_1971; @Georges91; @Nakanishi97_8085; @Tanaka98_2302]; in particular the result already includes the TAP approximation [@Thouless77_593], if only the first three terms are considered. In principle, these early works [@Vasiliev74; @Vasiliev75] also derive the Feynman rules for the Ising model; however, without giving any concrete expression for orders higher than three. In contrast to these results, the set of Feynman rules that we present here are applicable to general non-Gaussian theories. In case of the Ising model these had been sought for some time (see [@Opper01 p. 28]). We hope that the presented technique may prove useful in finding new approximations around known limiting cases. Examples may include expansions of the Hubbard model around the atomic limit [@Georges91]. An extension of our theory to higher order Legendre transforms, as broadly discussed in [@Vasiliev98], could lead to a diagrammatic formulation of the results derived in the context of the inverse Ising problem [@Sessak09; @Jacquin16] and to an alternative field-theoretic formulation of the extended Plefka-expansion for stochastic systems, that has recently been developed [@Bravi16]. The appendix presents an iterative algorithm as a first step towards this goal, an iterative equation to compute corrections to the second Legendre transform. Further work is required to derive a set of Feynman rules from this equation. The application of the here presented method is possible whenever a model admits a closed-form solution. An interesting regime of application may therefore be spherical models [@Berlin52_821]. In the thermodynamic limit, the free energy, the cumulant-generating function of the model, is known. Extensions of the spherical model that include four point coupling terms for example appear in the field of random lasers [@Antenucci15_053816]. If the quartic term is small compared to the quadratic term, the here proposed method could be applied to obtain approximate self-consistency equations. Such quartic spherical models, moreover, appear in inverse problems of diverse systems [@Marruzzo17]. Future work is needed to see if the perturbative results offered by the current work may help at obtaining approximate solutions to such inverse problems. We also note that the procedure as presented in yields a non-iterative way to generate all non-standard diagrams. We believe that this algorithm should be amenable to numerical implementation. The diagrammatic Monte Carlo technique, for example, presents an effective method to calculate the kernels of Schwinger-Dyson equations - for the one-particle Greens function this is the self-energy [@Greitemann18; @VanHoucke12_366; @VanHoucke08_arxiv]. One way to derive Schwinger-Dyson equations relies on multiple Legendre transforms, expressing the generating functionals in terms of correlation functions instead of potentials, as pioneered by de Dominicis and Martin [@DeDominicis64_14]. The equations of state then constitute self-consistency equations for these correlation functions, such as eq. 6.11 in [@Vasiliev98]. For example one needs to consider the second Legendre transform to determine the self-energy and the fourth to determine the two-particle Green’s function self-consistently. We do a first step towards developing diagrammatic rules for the second Legendre transform in the appendix , whereas higher order Legendre transforms are beyond the scope of the current work. Their diagrammatics is already quite involved if the underlying theory is Gaussian [@Vasiliev98 chap. 6.28.f]; the parquet equations form an approximation resulting from the fourth order Legendre transform. A different way to obtain a set of self-consistency equations, however, is by a direct resummation of diagrams, as in the derivation of the Bethe-Salpeter equation [@Salpeter51]. The diagrammatic rules derived here could be useful in such an approach. In general, the methods also seems particularly promising for hierarchical problems: assuming that a problem can be decomposed into small, but strongly interacting clusters that can be solved exactly, the method may be used to systematically expand in the interaction strength between such clusters. Appendix ======== Definition of the effective action\[app:Definition-effective-action\] --------------------------------------------------------------------- To define the effective action $\Gamma(x^{\ast})$ we eliminate the dependence on the source field $j$ in favor of the mean value $x^{\ast}(j):=\langle x\rangle=\partial_{j}W(j)$ by using and and by following the usual background field method [@Kleinert09 Chapter 3.23.6], briefly summarized here: We express $W$ as the integral $$\begin{aligned} \exp\left(W(j)\right) & = & \Z(0)^{-1}\int_{x}\,\exp\left(S(x)+j^{\T}x\right).\label{eq:W_as_Z-1}\end{aligned}$$ and then separate the fluctuations $\delta x=x-x^{\ast}$ from the background value $x^{\ast}$ to get $$\begin{aligned} \exp\left(W(j)-j^{\T}x^{\ast}\right) & = & \Z(0)^{-1}\int_{\delta x}\,\exp\left(S(x^{\ast}+\delta x)+j^{\T}\delta x\right).\label{eq:pre_Legendre}\end{aligned}$$ For given $x^{\ast}$, we now choose $j$ in a way that $x^{\ast}=\langle x\rangle(j)$ becomes the mean value of the field, so that $\delta x$ has vanishing mean $$\begin{aligned} 0\stackrel{!}{=}\langle\delta x\rangle & \equiv & \Z(0)^{-1}\int_{\delta x}\,\exp\left(S(x^{\ast}+\delta x)+j^{\T}\delta x\right)\,\delta x\\ & = & \Z(0)^{-1}\partial_{j}\,\int_{\delta x}\,\exp\left(S(x^{\ast}+\delta x)+j^{\T}\delta x\right)\\ & = & \partial_{j}\,\exp\left(W(j)-j^{\T}x^{\ast}\right),\end{aligned}$$ where we used in the last step. Since the exponential function is monotonic, $\exp(x)^{\prime}>0\quad\forall x$, the latter expression vanishes at the point where the exponent is stationary $$\begin{aligned} \partial_{j}\left(W(j)-j^{\T}x^{\ast}\right) & = & 0,\label{eq:stationary_j-1}\end{aligned}$$ The condition has the form of a Legendre transform from the function $W(j)$ to the effective action $\Gamma(x^{\ast})$. The supremum follows from stationarity and because $W$ is convex down, its Hessian $W^{(2)}$, the covariance matrix, is positive definite (cf. or [@ZinnJustin96 p. 166]). Therefore, $-W\left(j\right)+j^{T}x^{\ast}$, as a function of $j$, is convex up (concave), so we may define the Legendre transform by the supremum over $j$. Convexity of $W$\[sub:Convexity-of-W\] -------------------------------------- $W$ is convex, because $W^{(2)}$ is the covariance matrix: it is symmetric and therefore has real eigenvalues. For covariance matrices these are in addition always positive [@ZinnJustin96 p. 166]. This can be seen from the following argument. Let us define the bilinear form $$\begin{aligned} f(\eta) & : & =\eta^{\T}W^{(2)}\eta.\end{aligned}$$ A positive definite bilinear form has the property $f(\eta)>0\quad\forall\eta$. With $\delta x:=x-\langle x\rangle$ we can express the covariance as $W_{kl}^{(2)}=\langle\delta x_{k}\delta x_{l}\rangle$, so we may explicitly write $f(\eta)$ as $$\begin{aligned} f(\eta) & = & \sum_{k,l}\eta_{k}W_{kl}^{(2)}\eta_{l}\\ & = & \Z^{-1}(j)\,\eta^{\T}\int\,dx\,\delta x\,\delta x^{\T}\,\exp\left(S(x)+j^{\T}x\right)\eta\\ & = & \Z^{-1}(j)\,\int\,dx\,\left(\eta^{\T}\delta x\right)^{2}\,\exp\left(S(x)+j^{\T}x)\right)>0,\end{aligned}$$ which is the expectation value of a positive quantity. Linked cluster theorem\[app:Perturbative-cumulant-expansion\] ------------------------------------------------------------- The following proof of connectedness of all diagrammatic contributions to $W$ does not rely on the solvable part $S_{0}$ being Gaussian. We here start from the general expression to derive an expansion of $W(j)$, using the definition to write $$\begin{aligned} \exp(W(j))=Z(j) & = & \exp\left(\epsilon V(\partial_{j})\right)\,\exp\left(W_{0}(j)\right)\,\frac{\Z_{0}(0)}{\Z(0)}\label{eq:perturbation_general}\end{aligned}$$ Taking the logarithm, the latter factor turns into an additive constant $\ln\,\frac{\Z_{0}(0)}{\Z(0)}$ which ensures $W(0)=0$. Since we are ultimately interested in the derivatives of $W$, namely the cumulants, we may drop the constant and ensure $W(0)=0$ by finally dropping the zeroth order Taylor coefficient. The idea to prove connectedness follows to some extent [@ZinnJustin96]. The proof is by induction, dissecting the operator $\exp\left(\epsilon V(\partial_{j})\right)$ appearing in into infinitesimal operators using the definition of the exponential function as the limit $$\begin{aligned} \exp\left(\epsilon V(\partial_{j})\right) & = & \lim_{L\to\infty}(1+\frac{\epsilon}{L}\,V(\partial_{j}))^{L}.\label{eq:exponential_expansion}\end{aligned}$$ For large $L$ given and fixed and $0\leq l\leq L$, we define $$\begin{aligned} \exp\left(W_{l}(j)\right) & := & (1+\frac{\epsilon}{L}\,V(\partial_{j}))^{l}\,\exp\left(W_{0}(j)\right).\end{aligned}$$ It fulfills the trivial recursion $\exp(W_{l+1}(j))=(1+\frac{\epsilon}{L}\,V(\partial_{j}))\,W_{l}(j)$ from which follows an iteration by multiplication with $\exp(-W_{l}(j))$ and taking the logarithm $$\begin{aligned} & & W_{l+1}(j)-W_{l}(j)\label{eq:recursion_W_l}\\ & = & \ln\,\left[\exp\left(-W_{l}(j)\right)\,(1+\frac{\epsilon}{L}\,V(\partial_{j}))\,\exp\left(W_{l}(j)\right)\right].\nonumber \end{aligned}$$ The desired result $W(j)$ then follows as the limit $W(j)=\lim_{L\to\infty}\,W_{L}(j).$ Expanding $\ln(1+\frac{\epsilon}{L}x)=\frac{\epsilon}{L}\,x+\mathcal{O}\left((\frac{\epsilon}{L})^{2}\right)$ in we get $$\begin{aligned} & & W_{l+1}(j)-W_{l}(j)\label{eq:expansion_W_iteration}\\ & = & \frac{\epsilon}{L}\,\left(\exp\left(-W_{l}(j)\right)\,V(\partial_{j})\,\exp\left(W_{l}(j)\right)\right)+O\big(\big(\frac{\epsilon}{L}\big)^{2}\big).\nonumber \end{aligned}$$ We start the induction by noting that for $l=0$ we have $W_{l=0}=W_{0}$, the cumulant generating function of the solvable system. At this order, $W$ hence does not contain any diagrammatic corrections; so in particular no disconnected ones. We assume that the assumption is true until some $0\le l\le L$. Stated precisely, we assume that all perturbative corrections to $W_{l}(j)$ with $k$ vertices are connected and are $\propto\left(\frac{\epsilon}{L}\right)^{k}$; the sub-leading terms $O\big(\big(\frac{\epsilon}{L}\big)^{2}\big)$ in vanish in the limit $L\to\infty$, as shown below. Representing the potential $V$ as a Taylor series, we see that each step adds terms of the form shown in the second line $$\begin{aligned} W_{l+1}(j) & = & 1\cdot W_{l}(j)\label{eq:iterative_W}\\ & + & \frac{\epsilon}{L}\cdot\sum_{\{n_{i}\}}\frac{V^{(n_{1},\ldots,n_{N})}}{n_{1}!\cdots n_{N}!}\,\exp\left(-W_{l}(j)\right)\,\partial_{1}^{n_{1}}\cdots\partial_{N}^{n_{N}}\,\exp\left(W_{l}(j)\right),\nonumber \end{aligned}$$ where $\partial_{i}$ is used in short for $\partial_{j_{i}}$. The Taylor coefficient $\frac{V^{(n_{1},\ldots,n_{N})}}{n_{1}!\cdots n_{N}!}$ is graphically represented by a vertex (see ). Noting that the two exponential factors cancel each other after the differential operator has been applied to the latter factor, what remains is a set of connected components of $W_{l}(j)$ tied together by the vertex $\frac{V^{(n_{1},\ldots,n_{N})}}{n_{1}!\cdots n_{N}!}$. Disconnected components cannot appear, because there is only a single vertex; each of its legs belongs to one $\partial_{i}$, which, by acting on $W_{l}(j)$, attaches to one leg of the components in $W_{l}$. The iteration shows that the second term in each step adds to $W_{l}(j)$ a set of diagrams to obtain $W_{l+1}(j)$. It is clear from the single appearance of $V^{(n_{1},\ldots,n_{N})}$ that in each iteration only one such additional vertex is added. We will show now that we, moreover, only need to consider such additional diagrams, where the new vertex $V^{(n_{1},\ldots,n_{N})}$ connects to a perturbative correction contained in $W_{l}$ (with $k\ge1$ vertices), while all of its remaining legs connect to a cumulant of the unperturbed theory $W_{0}$ (with $k=0$ vertices). Stated differently, a perturbative correction with $k$ vertices picks up each of its vertices in a different iteration step $l$ in ; contributions where a single iteration step increases the number of vertices in a component by more than one vanish in the $L\to\infty$ limit. To understand why this is so, we consider the overall factor in front of a resulting diagram with $k$ vertices after $L$ iterations of . In each step of the first term copies all diagrams from $W_{l}$. The second term adds those formed by help of the additional vertex $V^{(n_{1},\ldots,n_{N})}$. Following how one particular graph is generated by the iteration, in each step we have the binary choice to either leave it as it is or to combine it with other components by help of an additional vertex. We first consider the case that each of the $k$ vertices is picked up in a different step (at different $l$) in the iteration. Each such step comes with a factor $\frac{\epsilon}{L}$ and in there are $\left(\begin{array}{c} L\\ k \end{array}\right)$ ways to select $k$ out of the $L$ iteration steps to pick up a vertex to construct this particular diagram. So in total we get a factor $$\begin{aligned} \left(\frac{\epsilon}{L}\right)^{k}\left(\begin{array}{c} L\\ k \end{array}\right) & = & \frac{\epsilon^{k}}{k!}\,\frac{L(L-1)\cdots(L-k+1)}{L^{k}}\;\stackrel{L\to\infty}{\to}\,\frac{\epsilon^{k}}{k!},\label{eq:factor_pick_up_one}\end{aligned}$$ which is independent of $L$. Now consider the case that we pick up the $k$ vertices along the iteration such that in one step we combined two sub-components with each one or more vertices already. Consequently, to arrive at $k$ vertices in the end, we only need $k^{\prime}<k$ iteration steps in which the second rather than the first term of acted on the component. The overall factor therefore is $$\begin{aligned} \left(\frac{\epsilon}{L}\right)^{k}\left(\begin{array}{c} L\\ k^{\prime} \end{array}\right) & = & \frac{\epsilon^{k}}{k^{\prime}!}\,\frac{L(L-1)\cdots(L-k^{\prime}+1)}{L^{k}}\label{eq:factor_pick_up_many}\\ & \stackrel{L\gg k^{\prime}}{=} & \frac{\epsilon^{k}}{k^{\prime}!}\,\frac{1}{L^{k-k^{\prime}}}\stackrel{L\to\infty}{=}0.\nonumber \end{aligned}$$ We can hence neglect the latter option and conclude that $W(j)$ is composed of all connected graphs, where a perturbative correction with $k$ vertices comes with the factor $\frac{\epsilon^{k}}{k!}$ given by . By the same reasoning we may neglect the $O\big(\big(\frac{\epsilon}{L}\big)^{2}\big)$ term in , because also here in a single step we would increase the order of the diagram by more than one factor $L^{-1}$. Operator equation for $\Gamma_{V}$\[app:operator\_equation\_Gamma\_V\] ---------------------------------------------------------------------- Let us first see why the decomposition into a sum in holds. To this end, we consider with $\delta x=x-x^{\ast}$ and use the decomposition of the action as well as the decomposition of $\Gamma$ to obtain $$\begin{aligned} & & \exp(-\Gamma_{0}(x^{\ast})-\Gamma_{V}(x^{\ast}))\\ & = & \Z^{-1}(0)\,\int_{x}\,\exp\big(S_{0}(x)+\epsilon V(x)+\big(\Gamma_{0}^{(1)\T}(x^{\ast})+\Gamma_{V}^{(1)\T}(x^{\ast})\big)(x-x^{\ast})\big).\end{aligned}$$ Collecting the terms depending on $x^{\ast}$ on the left hand side we get with $$\begin{aligned} & & \exp(\underbrace{-\Gamma_{0}(x^{\ast})+\Gamma_{0}^{(1)\T}(x^{\ast})\,x^{\ast}}_{W_{0}(j)\big\|_{j=\Gamma_{0}^{(1)}(x^{\ast})}}-\Gamma_{V}(x^{\ast}))\\ & = & \exp\big(\epsilon V(\partial_{j})+\Gamma_{V}^{(1)\T}(x^{\ast})(\partial_{j}-x^{\ast})\big)\,\Z^{-1}(0)\int_{x}\,\exp\big(S_{0}(x)+j^{\T}x)\big)\big\|_{j=\Gamma_{0}^{(1)}(x^{\ast})}\\ & = & \exp\big(\epsilon V(\partial_{j})+\Gamma_{V}^{(1)\T}(x^{\ast})(\partial_{j}-x^{\ast})\big)\,\exp\big(W_{0}(j)\big)\big\|_{j=\Gamma_{0}^{(1)}(x^{\ast})},\end{aligned}$$ where we moved the perturbing potential in front of the integral, making the replacement $x\to\partial_{j}$ and we identified the unperturbed cumulant generating function $\exp(W_{0}(j))=\Z^{-1}(0)\,\int_{x}\,\exp\big(S_{0}(x)+j^{\T}x\big)$ from the second to the third line. With the term $\Gamma_{0}^{(1)\T}(x^{\ast})x^{\ast}\equiv j_{0}^{\T}x^{\ast}$ on the left hand side, we get $-\Gamma_{0}(x^{\ast})+j^{\T}x^{\ast}=W_{0}(j)\big\|_{j=\Gamma_{0}^{(1)}(x^{\ast})}$, which follows from the definition . Multiplying with $\exp(-W_{0}(j))\big\|_{j=\Gamma_{0}^{(1)}(x^{\ast})}$ from left then leads to a recursive equation for $\Gamma_{V}$, which shows that our ansatz was indeed justified and that we may determine $\Gamma_{V}$ recursively, since $\Gamma_{V}$ appears again on the right hand side. Recursion for $g_{l}$\[app:Recusion-equation-for-Gamma\_V\] ----------------------------------------------------------- To construct the diagrams iteratively we write the perturbing term in as $$\begin{aligned} & & \exp\big(\epsilon V(\partial_{j})+\Gamma_{V}^{(1)\T}(x^{\ast})(\partial_{j}-x^{\ast})\big)\label{eq:pert_term_factors}\\ & = & \lim_{L\to\infty}\left(1+\frac{1}{L}\left(\epsilon V(\partial_{j})+\Gamma_{V}^{(1)\T}(x^{\ast})(\partial_{j}-x^{\ast})\right)\right)^{L}.\nonumber \end{aligned}$$ Inserted into we assume $L$ fixed but large and choose some $0\le l\le L$. We define $G_{l}(j)$ as the result after application of $l$ factors of the right hand side of $$\begin{aligned} \exp(G_{l}(j)) & := & \left(1+\frac{1}{L}\left(\epsilon V(\partial_{j})+\Gamma_{V}^{(1)\T}(x^{\ast})\,(\partial_{j}-x^{\ast})\right)\right)^{l}\,\exp(W_{0}(j)).\end{aligned}$$ Obviously we have $$\begin{aligned} G_{0} & \equiv & W_{0}.\label{eq:G_0_initial}\end{aligned}$$ For $l=L\to\infty$, we obtain the desired result as $$\begin{aligned} -\Gamma_{V}(x^{\ast}) & = & \lim_{L\to\infty}\,G_{L}(j)-W_{0}(j)\Big\|_{_{j=\Gamma_{0}^{(1)}(x^{\ast})}}.\end{aligned}$$ By its definition, $G_{l}(j)$ obeys the trivial iteration $$\begin{aligned} & & \exp(G_{l+1}(j))\\ & = & \left(1+\frac{1}{L}\left(\epsilon V(\partial_{j})+\Gamma_{V}^{(1)\T}(x^{\ast})\,(\partial_{j}-x^{\ast})\right)\right)\,\exp(G_{l}(j)).\end{aligned}$$ Multiplying from left with $\exp(-G_{l}(j))$ and taking the logarithm on both sides while using $\ln(1+\frac{1}{L}x)=\frac{1}{L}x+O(L^{-2})$ we get the recursion for the additional diagrams produced in step $l+1$ $$\begin{aligned} & & G_{l+1}(j)-G_{l}(j)\label{eq:iteration_G_app}\\ & = & \frac{\epsilon}{L}\,\exp(-G_{l}(j))\,V(\partial_{j})\,\exp(G_{l}(j))\nonumber \\ & + & \frac{1}{L}\,\exp(-G_{l}(j))\,\left(\Gamma_{V}^{(1)\T}(x^{\ast})\,(\partial_{j}-x^{\ast})\right)\,\exp(G_{l}(j))\nonumber \\ & + & \mathcal{O}(L^{-2}).\nonumber \end{aligned}$$ By the initial condition and the form of the additive iteration it is clear that all graphs of $W_{0}$ are also contained in $G_{l}$ for any step $l$. We may therefore define only the perturbative corrections as $g_{l}:=G_{l}-W_{0}$. We first note that indeed yields a closed iteration: Constructing the graphs of up to order $l+1$ in $G_{l+1}$ by we only need the graphs in $G_{l}$ on the right hand side of , which, by construction, are of order $\le l$ . This is because $V$ contains exactly one bare vertex and $\Gamma_{V}^{(1)}(x^{\ast})$ contains at least one. Taken together, we arrive at the central result of our work, Eq. . Second Legendre transform\[subsec:Second-Legendre-transform\] ------------------------------------------------------------- We here extend the iterative procedure to compute perturbative corrections to the second Legendre transform. To this end, we define an effective action that is a function of the first moment $x^{\ast}$ and the second moment $c^{\ast}=\langle x^{2}\rangle$. We express $W$ as the integral $$\begin{aligned} \exp\left(W(j,k)\right) & = & \Z(0)^{-1}\int_{x}\,\exp\left(S(x)+j^{\T}x+k^{\T}x^{2}\right).\label{eq:W_as_Z-1-1}\end{aligned}$$ Here $k^{\T}x^{2}$ must be understood as a bilinear form in $x$, hence $\sum_{il}k_{il}x_{i}x_{l}$. We have $$\begin{aligned} \frac{\partial W}{\partial j} & = & \langle x\rangle=:x^{\ast},\qquad\frac{\partial W}{\partial k}=\langle x^{2}\rangle=:c^{\ast}.\end{aligned}$$ We would like to define an effective action that is a function of these latter coordinates. So we define $\Gamma(x^{\ast},c^{\ast})$ as $$\begin{aligned} \exp\left(-\Gamma(x^{\ast},c^{\ast})\right) & = & \Z(0)^{-1}\int_{x}\,\exp\left(S(x)+j^{\T}(x-x^{\ast})+k^{\T}(x^{2}-c^{\ast})\right)\nonumber \\ \label{eq:pre_legendre-1}\\ \Gamma(x^{\ast},c^{\ast}) & = & j^{\T}x^{\ast}+k^{\T}c^{\ast}-W(j,k),\nonumber \end{aligned}$$ with the additional constraints that $\partial/\partial j$ and $\partial/\partial k$ of the right hand side vanishes. Consequently, the so-defined function $\Gamma$ fulfills the two equations of state $$\begin{aligned} \frac{\partial\Gamma}{\partial x^{\ast}} & = & j,\qquad\frac{\partial\Gamma}{\partial c^{\ast}}=k,\end{aligned}$$ which can be obtained by application of the chain rule. The second of these equations, for example, follows as $$\begin{aligned} \frac{\partial\Gamma(x^{\ast},c^{\ast})}{\partial c^{\ast}} & = & \frac{\partial j^{\T}}{\partial c^{\ast}}x^{\ast}+k+\frac{\partial k^{\T}}{\partial c^{\ast}}c^{\ast}-\underbrace{\frac{\partial W^{\T}}{\partial j}}_{\equiv x^{\ast}}\frac{\partial j}{\partial c^{\ast}}-\underbrace{\frac{\partial W^{\T}}{\partial k}}_{\equiv c^{\ast}}\frac{\partial k}{\partial c^{\ast}}=k.\end{aligned}$$ Hence we may write the definition of $\Gamma$ also as $$\begin{aligned} \exp\left(-\Gamma(x^{\ast},c^{\ast})\right) & = & \Z(0)^{-1}\int_{x}\,\exp\left(S(x)+\frac{\partial\Gamma^{\T}}{\partial x^{\ast}}\,(x-x^{\ast})+\frac{\partial\Gamma^{\T}}{\partial c^{\ast}}\,(x^{2}-c^{\ast})\right).\end{aligned}$$ We next decompose $S(x)=S_{0}(x)+\epsilon V(x)$, where we assume that we may compute the cumulant generating function $W_{0}(j,k)=\ln\int_{x}\,\exp(S_{0}(x)+j^{\T}x+k^{\T}x^{2})$ exactly. Correspondingly, we assume a decomposition of the effective action into the solvable part $\Gamma_{0}$ and the perturbative corrections $\Gamma_{V}$ as $$\begin{aligned} \Gamma(x^{\ast},c^{\ast}) & = & \Gamma_{0}(x^{\ast},c^{\ast})+\Gamma_{V}(x^{\ast},c^{\ast}).\end{aligned}$$ With the notation $\frac{\partial\Gamma}{\partial x^{\ast}}=:\Gamma^{(1,0)}$ and $\frac{\partial\Gamma}{\partial c^{\ast}}=:\Gamma^{(0,1)}$ we may express the integral equation as $$\begin{aligned} & & \exp(-\Gamma_{0}(x^{\ast},c^{\ast})-\Gamma_{V}(x^{\ast},c^{\ast}))\\ & = & \Z^{-1}(0)\,\int_{x}\,\exp\Big(S_{0}(x)+\epsilon V(x)\\ & & \phantom{\Z^{-1}(0)\,\int_{x}\,\exp}+\big(\Gamma_{0}^{(1,0)\T}+\Gamma_{V}^{(1,0)}\big)^{\T}(x-x^{\ast})\\ & & \phantom{\Z^{-1}(0)\,\int_{x}\,\exp}+\big(\Gamma_{0}^{(0,1)\T}+\Gamma_{V}^{(0,1)}\big)^{\T}(x^{2}-c^{\ast})\Big)\end{aligned}$$ or, bringing all terms that are independent of the integration variable $x$ to the left, we get with $j_{0}=\Gamma_{0}^{(1,0)}$ and $k_{0}=\Gamma_{0}^{(0,1)}$ $$\begin{aligned} & & \exp(\underbrace{-\Gamma_{0}(x^{\ast},c^{\ast})+j_{0}^{\T}x^{\ast}+k_{0}^{\T}c^{\ast}}_{W_{0}(j_{0},k_{0})}-\Gamma_{V}(x^{\ast},c^{\ast}))\\ & = & \exp\big(\epsilon V(\partial_{j})+\Gamma_{V}^{(1,0)\T}(\partial_{j}-x^{\ast})+\Gamma_{V}^{(0,1)\T}\,(\partial_{k}-c^{\ast})\big)\times\\ & & \times\Z^{-1}(0)\int_{x}\,\exp\big(S_{0}(x)+j^{\T}x+k^{\T}x^{2}\big)\big\|_{j=\Gamma_{0}^{(1,0)}(x^{\ast},c^{\ast}),\,k=\Gamma_{0}^{(0,1)}(x^{\ast},c^{\ast}).}\end{aligned}$$ We hence obtain the operator form of the equation as $$\begin{aligned} \exp(-\Gamma_{V}(x^{\ast},c^{\ast})) & = & \exp(-W_{0}(j,k))\times\\ & \times & \exp\big(\epsilon V(\partial_{j})+\Gamma_{V}^{(1,0)\T}(\partial_{j}-x^{\ast})+\Gamma_{V}^{(0,1)\T}(\partial_{k}-c^{\ast})\big)\times\\ & \times & \exp(W_{0}(j,k))\big\|_{j=\Gamma_{0}^{(1,0)}(x^{\ast},c^{\ast}),k=\Gamma_{0}^{(0,1)}(x^{\ast},c^{\ast})}.\end{aligned}$$ Using the representation of the exponential function as a series and the series expansion of the logarithm to first order, we get the iteration $$\begin{aligned} & & g_{l+1}(j,k)-g_{l}(j,k)\label{eq:iteration_g-1}\\ & = & \frac{\epsilon}{L}\,\exp(-W_{0}(j,k)-g_{l}(j,k))\,V(\partial_{j})\,\exp(W_{0}(j,k)+g_{l}(j,k))\nonumber \\ & + & \frac{1}{L}\,\exp(-W_{0}(j,k)-g_{l}(j,k))\,\Gamma_{V}^{(1,0)}\left(\partial_{j}-x^{\ast}\right)\,\exp(W_{0}(j,k)+g_{l}(j,k))\nonumber \\ & + & \frac{1}{L}\,\exp(-W_{0}(j,k)-g_{l}(j,k))\,\Gamma_{V}^{(0,1)}\left(\partial_{k}-c^{\ast}\right)\,\exp(W_{0}(j,k)+g_{l}(j,k))\nonumber \\ & + & \mathcal{O}(L^{-2}),\nonumber \end{aligned}$$ with the initial condition $g_{0}=0$. The perturbative corrections to the effective action then result as the limit $$\begin{aligned} -\Gamma_{V}(x^{\ast},c^{\ast})= & \lim_{L\to\infty} & \,g_{L}(j,k)\Big\|_{j=\Gamma_{0}^{(1,0)}(x^{\ast},c^{\ast}),\,k=\Gamma_{0}^{(0,1)}(x^{\ast},c^{\ast})}.\end{aligned}$$ One could use this iterative procedure to compute perturbative corrections. It seems that a proof of a generalized irreducibility should follow along similar lines as in the case of a first order Legendre transform. In particular, the term $\Gamma_{V}^{(0,1)}\left(\partial_{k}-c^{\ast}\right)$ establishes a double link between a component contained in $g_{l}$ and one in $\Gamma_{V}$, thus suggesting a more general 2PI property. We will leave a more careful consideration open for subsequent works. Taxonomy of reducible diagrams\[subsec:Taxonomy-of-reducible\] -------------------------------------------------------------- One might wonder why the diagram in c) remains in the perturbative expansion of $\Gamma$ despite being reducible in the sense that it is divided into two parts by cutting one of the legs of $\Gamma_{0}^{\left(2\right)}$, which otherwise acts similarly as a leg of a bare interaction vertex. We here provide an explanation in the framework of the iterative construction of $\Gamma_{V}$. There are four different types of diagrams that are reducible in the sense that they fall apart if one detaches one leg from a cumulant. We classify them by the type of subdiagrams (leaves) they are composed of and how these leaves are connected. We can distinguish five cases: 1. Two irreducible diagrams from $W_{V}$ connected by a single leg that belongs to a bare interaction vertex.\[enu:I-Ttwo–diagrams-Wv\] 2. Multiple irreducible diagrams from $W_{V}$ that are connected to the remainder by the same junction as in case I.\[enu:-Multiple-diagrams-Wv\] 3. Two irreducible diagrams from $W_{V}^{\left(1\right)}$ connected via a $\Gamma_{0}^{\left(2\right)}$-component.\[enu:-Two-diagramsWv-Gamma0\] 4. One diagram from $W_{V}^{\left(1\right)}$ and one non-standard component of $\Gamma_{V}$ (not contained in the expansion of $W_{V}$), necessarily connected by a $\Gamma_{0}^{\left(n\right)}$-component\[enu:-One-diagram-Wv\_One\_Gammav\] 5. Two non-standard diagrams of $\Gamma_{V}$ (not contained in the expansion of $W_{V}$), necessarily connected by a $\Gamma_{0}^{\left(n\right)}$-component\[enu:-Two-components\_Gammav\] In case I, in which both parts are irreducible, we may insert either one or two unities of the form $$1=\Diagram{gpgfcf} ,\ \label{eq:Decomposition_of_Unity}$$ where empty and hatched circles in this section always denote the unperturbed quantities $W_{0}^{\left(n\right)}$ and $\Gamma_{0}^{\left(2\right)}$, respectively. We therefore see that we can replace both leaves by a part of $\Gamma_{V}^{\left(1\right)}\left(x^{\ast}\right)$. So, together with the original diagram, we have four contributions that all contribute with the same magnitude, two of them positive, two of them negative: The sum is $0$. Schematically, the situation is as follows: [Example\_cancelling]{} $$\begin{aligned} \mkern-108mu 0=\quad\parbox{75mm}{ \begin{fmfgraph}(70,30) \fmfpen{0.75thin} \fmfleft{i1} \fmfright{o1} \fmf{plain, tension=.25, right=.5}{i1,v3,i1} \fmf{plain, tension=.25, right=.25}{i1,v3,i1} \fmf{plain, tension=1.}{v3,v4} \fmf{plain,tension=.33, right=.5}{v4,o1,v4} \fmf{plain,tension=.33}{v4,o1} \fmfv{decor.shape=circle, decor.filled=empty, decor.size = thin}{v3} \fmfv{d.s=circle, d.filled=hatched}{i1,o1} \end{fmfgraph} } \mkern-236mu & -2 \quad & \mkern-18mu \quad \parbox{125mm}{ \begin{fmfgraph}(120,30) \fmfpen{0.75thin} \fmfleft{i1} \fmfright{o1} \fmf{plain, tension=.25, right=.5}{i1,v3,i1} \fmf{plain, tension=.25, right=.25}{i1,v3,i1} \fmf{plain, tension=1.}{v3,g1} \fmf{plain, tension=1.}{g1,c1} \fmf{plain, tension=1.5}{c1,v4} \fmf{plain,tension=.33, right=.5}{v4,o1,v4} \fmf{plain,tension=.33}{v4,o1} \fmfv{decor.shape=circle, decor.filled=empty, decor.size = thin}{v3,c1} \fmfv{decor.shape=circle, decor.filled=shaded, decor.size = thin}{g1} \fmfv{d.s=circle, d.filled=hatched}{i1,o1} \fmfv{label=$\underbrace{\phantom{ist Phantomas!}}_{\subseteq \Gamma^{\left(1\right)}_{V}}$, label.angle=-90, label.dist=5pt}{v3} \fmfv{label=$\overbrace{\phantom{ente das ist Phantomas!}}^{\subseteq \Gamma^{\left(1\right)}_{V}}$, label.angle=110, label.dist=5pt}{v4} \end{fmfgraph} } \nonumber\\ & & \nonumber\\ &\mkern-18mu +& \parbox{125mm}{ \begin{fmfgraph}(180,30) \fmfpen{0.75thin} \fmfleft{i1} \fmfright{o1} \fmf{plain, tension=.25, right=.5}{i1,v3,i1} \fmf{plain, tension=.25, right=.25}{i1,v3,i1} \fmf{plain, tension=1.}{v3,g1} \fmf{plain, tension=1.}{g1,c1} \fmf{plain, tension=1.}{c1,g2} \fmf{plain, tension=1.}{g2,c2} \fmf{plain, tension=1.5}{c2,v4} \fmf{plain,tension=.33, right=.5}{v4,o1,v4} \fmf{plain,tension=.33}{v4,o1} \fmfv{decor.shape=circle, decor.filled=empty, decor.size = thin}{v3,c1,c2} \fmfv{decor.shape=circle, decor.filled=shaded, decor.size = thin}{g1,g2} \fmfv{d.s=circle, d.filled=hatched}{i1,o1} \fmfv{label=$\underbrace{\phantom{ist Phantomas!}}_{\subseteq \Gamma^{\left(1\right)}_{V}}$, label.angle=-90, label.dist=5pt}{v3} \fmfv{label=$\overbrace{\phantom{ente das ist Phantomas!}}^{\subseteq \Gamma^{\left(1\right)}_{V}}$, label.angle=110, label.dist=5pt}{v4} \end{fmfgraph} }\mkern-268mu.\nonumber\\ \nonumber \end{aligned}$$ Here the circles filled with squares denote some arbitrary continuation of the respective diagram. We conclude that in second order in the interactions, no diagrams with irreducible $W_{V}^{\left(1\right)}$-leaves contribute. We will proof by induction in the number of vertices $n$ that no diagram described in case Ia (with at least one reducible link) contributes to $\Gamma_{V}$. Case I is our induction assumption ($n=2$). Assume that the assumption holds $\forall\,k\leq n$. We observe that if the whole diagram contains $n+1$ interactions we can pick an arbitrary leaf that is connected to the rest by a single link, so that the (reducible) rest contains at most $n$ interactions. By the induction assumption, this rest is not contained in the diagrammatic expansion of $\Gamma_{V}$. Therefore, we have just one contribution, which can be replaced by a part of $\Gamma_{V}^{\left(1\right)}$ contributing with opposite sign. This contribution cancels the contribution constructed solely out of bare vertices. Diagrammatically this could be depicted by [Example\_cancelling\_single]{} $$\begin{aligned} \mkern-36mu 0=\quad\parbox{75mm}{ \begin{fmfgraph}(70,30) \fmfpen{0.75thin} \fmfleft{i1} \fmfright{o1} \fmf{plain, tension=.25, right=.5}{i1,v3,i1} \fmf{plain, tension=.25, right=.25}{i1,v3,i1} \fmf{plain, tension=1.}{v3,v4} \fmf{plain,tension=.33, right=.5}{v4,o1,v4} \fmf{plain,tension=.33}{v4,o1} \fmfv{decor.shape=circle, decor.filled=empty, decor.size = thin}{v3} \fmfv{d.s=circle, d.filled=hatched}{i1,o1} \end{fmfgraph} } \mkern-236mu & - \quad & \mkern-18mu \quad \parbox{125mm}{ \begin{fmfgraph}(120,30) \fmfpen{0.75thin} \fmfleft{i1} \fmfright{o1} \fmf{plain, tension=.25, right=.5}{i1,v3,i1} \fmf{plain, tension=.25, right=.25}{i1,v3,i1} \fmf{plain, tension=1.}{v3,g1} \fmf{plain, tension=1.}{g1,c1} \fmf{plain, tension=1.5}{c1,v4} \fmf{plain,tension=.33, right=.5}{v4,o1,v4} \fmf{plain,tension=.33}{v4,o1} \fmfv{decor.shape=circle, decor.filled=empty, decor.size = thin}{v3,c1} \fmfv{decor.shape=circle, decor.filled=shaded, decor.size = thin}{g1} \fmfv{d.s=circle, d.filled=hatched}{i1,o1} \fmfv{label=$\underbrace{\phantom{ist Phantomas!}}_{\subseteq \Gamma^{\left(1\right)}_{V}}$, label.angle=-90, label.dist=5pt}{v3} \end{fmfgraph} }\mkern-400mu. \nonumber \\ \nonumber \end{aligned}$$ This concludes the induction. We conclude from these observations that no diagrams with a single irreducible $W_{V}^{\left(1\right)}$-leaf contribute. Considering case II, it is obvious that we can choose one of the leaves to be provided by $\Gamma_{V}^{\left(1\right)}$ and the other one to be composed of bare vertices. We can also insert a unity to the left of the given $\Gamma_{0}^{\left(2\right)}$, so that we can interpret both leaves to come from $\Gamma_{V}^{\left(1\right)}\left(x^{\ast}\right)$. Because the diagram composed solely of bare vertices is missing, we only have three contributions. They all contribute with the same absolute value, but two of them with one sign, one with the other sign and therefore, the contributions do not sum to zero and the whole diagram contributes to the expansion of $\Gamma_{V}$, as claimed in the main text. Diagrammatically, this is illustrated by [Example\_non\_cancelling]{} $$\begin{aligned} \mkern-90mu 0\neq-2\cdot \quad \parbox{125mm}{ \begin{fmfgraph}(105,30) \fmfpen{0.75thin} \fmfleft{i1} \fmfright{o1} \fmf{plain, tension=.25, right=.5}{i1,v3,i1} \fmf{plain, tension=.25, right=.25}{i1,v3,i1} \fmf{plain, tension=1.}{v3,g1} \fmf{plain, tension=1.}{g1,v4} \fmf{plain,tension=.33, right=.5}{v4,o1,v4} \fmf{plain,tension=.33}{v4,o1} \fmfv{decor.shape=circle, decor.filled=empty, decor.size = thin}{v3,v4} \fmfv{decor.shape=circle, decor.filled=shaded, decor.size = thin}{g1} \fmfv{d.s=circle, d.filled=hatched}{i1,o1} \fmfv{label=$\underbrace{\phantom{ist Phantomas!}}_{\subseteq \Gamma^{\left(1\right)}_{V}}$, label.angle=-90, label.dist=5pt}{v3} \fmfv{label=$\overbrace{\phantom{ist Phantomas!}}^{\subseteq \Gamma^{\left(1\right)}_{V}}$, label.angle=110, label.dist=5pt}{v4} \end{fmfgraph} } \mkern-420mu & + \quad & \parbox{125mm}{ \begin{fmfgraph}(150,30) \fmfpen{0.75thin} \fmfleft{i1} \fmfright{o1} \fmf{plain, tension=.25, right=.5}{i1,v3,i1} \fmf{plain, tension=.25, right=.25}{i1,v3,i1} \fmf{plain, tension=1.}{v3,g1} \fmf{plain, tension=1.}{g1,c1} \fmf{plain, tension=1.}{c1,g2} \fmf{plain, tension=1.}{g2,v4} \fmf{plain,tension=.33, right=.5}{v4,o1,v4} \fmf{plain,tension=.33}{v4,o1} \fmfv{decor.shape=circle, decor.filled=empty, decor.size = thin}{v3,c1,v4} \fmfv{decor.shape=circle, decor.filled=shaded, decor.size = thin}{g1,g2} \fmfv{d.s=circle, d.filled=hatched}{i1,o1} \fmfv{label=$\underbrace{\phantom{ist Phantomas!}}_{\subseteq \Gamma^{\left(1\right)}_{V}}$, label.angle=-90, label.dist=5pt}{v3} \fmfv{label=$\overbrace{\phantom{ist Phantomas!}}^{\subseteq \Gamma^{\left(1\right)}_{V}}$, label.angle=110, label.dist=5pt}{v4} \end{fmfgraph} }\mkern-350mu. \nonumber\\ \end{aligned}$$ In point II, we have proved that there are diagrams contributing to $\Gamma_{V}$, but not to $W_{V}$ (for the example c)), therefore the situations of points III and IV can appear. The diagrams described under point III cancel, because the irreducible components within the $W_{V}^{\left(1\right)}$-like leaf can be replaced a $\Gamma_{V}^{\left(1\right)}$-part or can be left as it is. Both contribute with same absolute value, but opposite sign and cancel. Because the other part cannot be composed just by bare vertices and unperturbed cumulants, it occurs only once, namely as part of $\Gamma_{V}^{\left(1\right)}$. Therefore, the contribution from this diagram remains zero. Point IV is even simpler, because by construction we only have a unique way to construct the composed diagram, namely by joining two $\Gamma_{V}^{\left(1\right)}$-parts and therefore, the whole diagram is not canceled. Using these rules to determine if a given diagram contributes to $\Gamma_{V}$ or not, we conclude from case Ia that, as mentioned in the main text, diagrams with at least one component from $W_{V}$ do not contribute. However, one has to be careful applying case III to rule out the occurrence of a certain diagram, that can be decomposed according to III, because this decomposition is not necessarily unique and so the contribution of the whole diagram is actually not zero. Calculation of symmetry factors in the diagrammatic solution of the Ising model\[app:App\_Calc\_of\_sym\_factors\] ------------------------------------------------------------------------------------------------------------------ We determine here the symmetry factors of the diagrams given in the main text using the following scheme: First, label all legs of interactions by indices. Then, count the possible combinations of swaps of the two legs at an interaction and permutations of interactions that lead to a new labeled diagram. For the second-order-diagram (\[Def\_second\_oder\_diagram\]), this is just $2$ coming from the fact that flipping a single vertex produces a new labeled diagram, but flipping both vertices leaves the diagram invariant. For the ring diagram in third order (\[Def\_third\_oder\_ring\_diagram\]), permuting interactions produces the same diagram (“at most” it mirrors the diagram). The same holds for the other third order diagram (\[Def\_third\_oder\_Citroen\_diagram\]). Swapping legs in the last diagram (\[Def\_third\_oder\_Citroen\_diagram\]) yields a factor $2^{2}$, because switching all vertex legs does not yield a new labeled diagram, whereas for the ring diagram, it does, leading to the symmetry factor $2^{3}$. Proceeding to fourth order, we first compute the symmetry factor $\SR$ of the ring diagram in (\[Fourth\_order\_ring\]). For a given interaction, we have $3$ possibilities to choose another interaction not to pair it with and every vertex flip produces a new labeled diagram. This gives $S_{\mathrm{R}}=3\cdot2^{4}=48.$ For the second diagram (\[Fourth\_order\_TM\]), labeled by “TM” (because it is depicted in a hexagonal shape reminiscent of the elements of the game board of “Terra Mystica”), we may choose $\left(\begin{array}{c} 4\\ 2 \end{array}\right)$ vertices for the upper position and flip every of the four vertices: $\ST=6\cdot2^{4}=96$. The diagram (\[Fourth\_order\_Citroen\]) that looks like a doubled Citro[ë]{}n logo (dC), whose symmetry factor is given - analogous to its third-order-counterpart - by $\SdC=2^{3}=8$. Including the numerical factors contained in the third-oder-cumulants into the prefactor, we end up with the following contribution to the effective action: $$\frac{1}{16}\underbrace{\frac{8\cdot\left(-2\right)^{2}}{24}}_{=\frac{4}{3}}\sum_{i\neq j}J_{ij}^{2}J_{jk}^{2}\left(1-m_{i}^{2}\right)\left(1-m_{j}^{2}\right)\left(1-3m_{i}^{2}\right)\left(1-3m_{j}^{2}\right).\label{eq:Contr_double_Citroen_final}$$ Finally, let us determine the symmetry factors for the “non-standard” fourth-order-contributions in (\[Fourth\_order\_extraordinary\]), $S_{i}$ for $i\in\left\{ a,b,c\right\} $: For diagram $a$, which we term “glasses”-diagram for obvious reasons, there are three ways to pair, say, the first vertex with any of the three others. The other pair is then fixed, too. Then, every flip of a vertex produces a new labeled diagram because it matters which vertex is attached to the four-point-cumulant. Together, this gives $S_{a}=3\cdot2^{4}$. For diagram $S_{b}$, flipping each of the two single vertices in the left part produces a new labeled diagram, because it matters which vertex is attached to the three-point-cumulant ($2^{2}$). The leaf brings in a factor $2$, because also here, flipping both vertices brings in a factor $2$ each and a factor $1/2$ because the leaf is of second order in the interaction. We count the leaf as a different interaction type, therefore the prefactor is given by $\frac{2^{3}}{2!\cdot1!}$. Finally, there is only one way to construct diagram $c$ and we are left with the intrinsic symmetry factors of the clusters: $S_{c}=2^{2}$. Here again, $\Gamma_{V}^{\left(1\right)}$ acts as a special kind of interaction, which occurs twice, which leads to the prefactor $\frac{2^{2}}{2}$. Adding up the contributions of the three diagrams leads us to the following expression for the term inside the square bracket in (\[Fourth\_order\_extraordinary\]):$$\begin{aligned} &\mkern-95mu\left\{ \frac{1}{4!}3\cdot2^{4}\left(-2\right)\left(1-3m_{j}^{2}\right)+\left(-\frac{1}{2!1!}8+\frac{1}{2}4\right)4m_{j}^{2}\right\} \left(1-m_{i}^{2}\right)\left(1-m_{j}^{2}\right)\left(1-m_{k}^{2}\right)\nonumber\\ \mkern-115mu=&\mkern-95mu\left\{ -4\left(1-m_{j}^{2}\right)\right\} \left(1-m_{i}^{2}\right)\left(1-m_{j}^{2}\right)\left(1-m_{k}^{2}\right).\\ \nonumber\end{aligned}$$In conclusion, the diagrams in (\[Fourth\_order\_extraordinary\]) add up to $$\frac{-4}{2^{4}}\sum_{i\neq j\neq k}J_{ij}^{2}J_{jk}^{2}\left\{ 1-m_{j}^{2}\right\} \left(1-m_{i}^{2}\right)\left(1-m_{j}^{2}\right)\left(1-m_{k}^{2}\right).\label{eq:Fourth_order_extraord_raw}$$ The term in curly braces, despite emerging from a sum of three different diagrams, looks like a second cumulant. Interestingly, the contribution for the case $i\neq k$ is exactly canceled by that of the ring-diagram in the case that the indices of exactly two of the opposite cumulants in the ring are equal. We see this as follows: There are two ways how exactly two indices of opposite circles in the ring can be identified ($i=k$ or $j=l$) so that the symmetry factor of this diagram is multiplied by $2$, which gives the total prefactor $\frac{1}{4!}\frac{1}{2^{4}}\SR\cdot2=\frac{1}{4}$, which is indeed equal to the prefactor in (up to the sign). Similarly, we get a contribution to the case of only two different indices in addition to the “double Citro[ë]{}n-diagram”. For the diagrams $S_{a}$, $S_{b}$, and $S_{c}$ that means that the indices of the outer circles are equal, giving the contribution $$-\frac{1}{4}\sum_{i\neq j}J_{ij}^{2}J_{jk}^{2}\left(1-m_{i}^{2}\right)^{2}\left(1-m_{j}^{2}\right)^{2}.\label{eq:Fourth_order_extraord_outer_equal}$$ For the ring diagram, there is a unique way to identify indices of opposite circles, therefore we do not get the factor $2$ in addition to the symmetry factor such that this contribution does not cancel , but just reduces it to $-\frac{1}{16}2\sum_{i\neq j}J_{ij}^{2}J_{jk}^{2}\left(1-m_{i}^{2}\right)^{2}\left(1-m_{j}^{2}\right)^{2}$. The “quadruple Citro[ë]{}n-diagram” (\[Fourth\_order\_Citroen\]) gives us . Adding up the relevant terms of these two contributions, leaving out common factors and the summation, gives$$\begin{aligned} &\mkern-65mu-2\left(1-m_{i}^{2}\right)^{2}\left(1-m_{j}^{2}\right)^{2}+\frac{4}{3}\left(1-m_{i}^{2}\right)\left(1-m_{j}^{2}\right)\left(1-3m_{i}^{2}\right)\left(1-3m_{j}^{2}\right)\nonumber\\ \mkern-85mu=&\mkern-65mu\frac{2}{3}\left(1-m_{i}^{2}\right)\left(1-m_{j}^{2}\right)\left(-3\left(1-m_{i}^{2}\right)\left(1-m_{j}^{2}\right)+2\left(1-3m_{i}^{2}\right)\left(1-3m_{j}^{2}\right)\right)\nonumber\\ \mkern-85mu=&\mkern-65mu\frac{2}{3}\left(1-m_{i}^{2}\right)\left(1-m_{j}^{2}\right)\left(-1-3m_{i}^{2}-3m_{j}^{2}+15m_{i}^{2}m_{j}^{2}\right).\\ \nonumber\end{aligned}$$Taken together this leads to . Acknowledgements ================ We are indebted to Adam Rançon who commented on an earlier version of this manuscript and especially for his hints on Georges’ and Yedidia’s summation convention, which allowed us to see that our result is consistent with earlier works. Furthermore, we would like to acknowledge the two anonymous reviewers for many thoughtful comments that helped us to considerably improve the manuscript. In particular, we thank them for pointing out the importance of a non-iterative algorithm, that lead to the idea to use skeleton diagrams, and for pointing us towards literature on spherical models and laser physics as a potential field of application. [10]{} url\#1[[\#1]{}]{}urlprefix\[2\]\[\][[\#2](#2)]{} Negele J W and Orland H 1998 [Quantum Many-Particle Systems]{} (New York: Perseus Books) Zinn-Justin J 1996 [Quantum field theory and critical phenomena]{} (Clarendon Press, Oxford) Amit D J 1984 [Field theory, the renormalization group, and critical phenomena]{} (World Scientific) Altland A and Simons B 2010 [Concepts of Theoretical Solid State Physics]{} (Cambridge University Press) Vasiliev A 1998 [Functional Methods in Quantum Field Theory and Statistical Physics]{} (Gordon and Breach Science Publishers) Mayer J E and Geoppert-Mayer M 1977 [Statistical mechanics]{} 2nd ed (John Wiley & Sons, Inc.) De Dominicis C and Martin P C 1964 [Journal of Mathematical Physics]{} [5]{} 14 Lasinio J 1964 [Nuovo Cimento]{} [34]{} 1790 Goldstone J, Salam A and Weinberg S 1962 [Phys. Rev.]{} [127]{}(3) 965–970 <https://link.aps.org/doi/10.1103/PhysRev.127.965> Berges J, Tetradis N and Wetterich C 2002 [Physics Reports]{} [363]{} 223 – 386 ISSN 0370-1573 renormalization group theory in the new millennium. {IV} <http://www.sciencedirect.com/science/article/pii/S0370157301000989> Sénéchal D, Tremblay A and Bourbonnais C 2004 [Theoretical Methods for Strongly Correlated Electrons]{} (Springer) Fock V 1930 [Zeitschrift f[ü]{}r Physik]{} [61]{} 126–148 ISSN 0044-3328 <https://doi.org/10.1007/BF01340294> Kleinert H 2009 [Path Integrals in Quantum Mechanics, Statistics, Polymer Physics, and Financial Markets]{} 5th ed (World Scientific) Wilson K G and Kogut J 1974 [Physics Reports]{} [12]{} 75 – 199 ISSN 0370-1573 <http://www.sciencedirect.com/science/article/pii/0370157374900234> Wilson K G 1975 [Rev. Mod. Phys.]{} [47]{}(4) 773–840 <http://link.aps.org/doi/10.1103/RevModPhys.47.773> Hertz J A 1976 [Physical Review B]{} [14]{} 1165 Wetterich C 1993 [Physics Letters B]{} [301]{} 90 – 94 ISSN 0370-2693 <http://www.sciencedirect.com/science/article/pii/037026939390726X> Metzner W, Salmhofer M, Honerkamp C, Meden V and Schönhammer K 2012 [Rev. Mod. Phys.]{} [84]{}(1) 299–352 <http://link.aps.org/doi/10.1103/RevModPhys.84.299> Gardiner C W 1985 [Handbook of Stochastic Methods for Physics, Chemistry and the Natural Sciences]{} 2nd ed ([Springer Series in Synergetics]{} no 13) (Berlin: Springer-Verlag) ISBN 3-540-61634-9, 3-540-15607-0 Uhlenbeck G E and Ford G W 1963 [Lectures in Statistical Physics]{} (Providence, R.I.: American Mathematical Society) Vasiliev A N and Radzhabov R A 1974 [Theoretical and Mathematical Physics]{} [21]{} 963–970 ISSN 1573-9333 <https://doi.org/10.1007/BF01035593> Vasiliev A N and Radzhabov R A 1975 [Theoretical and Mathematical Physics]{} [23]{} 575–579 ISSN 1573-9333 <https://doi.org/10.1007/BF01041677> Georges A and Yedidia J S 1991 [Journal of Physics A: Mathematical and General]{} [24]{} 2173 <http://stacks.iop.org/0305-4470/24/i=9/a=024> Kirkpatrick S and Sherrington D 1978 [Phys. Rev. B]{} [17]{} 4384 Thouless D J, Anderson P W and Palmer R G 1977 [Philosophical Magazine]{} [35]{} 593–601 Nakanishi K and Takayama H 1997 [J. Phys. A: Math. Gen.]{} [30]{} 8085–8094 Tanaka T 1998 [Phys. Rev. E]{} [58]{} 2302–2310 Plefka T 1982 [J Phys A: Math Gen]{} [15]{} 1971–1978 Sommers H 1987 [Phys. Rev. Lett.]{} [58]{} 1268–1271 Horwitz G and Callen H B 1961 [Phys. Rev.]{} [124]{}(6) 1757–1785 <https://link.aps.org/doi/10.1103/PhysRev.124.1757> Bloch C and Langer J 1965 [Journal of Mathematical Physics]{} [6]{} 554–572 Bogolyubov N M, Brattsev V F, Vasiliev A N, Korzhenevskii A L and Radzhabov R A 1976 [Theoretical and Mathematical Physics]{} [26]{} 230–237 ISSN 1573-9333 <https://doi.org/10.1007/BF01032093> Yedidia J S and Georges A 1990 [Journal of Physics A: Mathematical and General]{} [23]{} 2165 <http://stacks.iop.org/0305-4470/23/i=11/a=037> Opper M and Saad D (eds) 2001 [Advanced Mean Field Methods - Theory and Practice]{} (The MIT Press) Dahmen H D and Jona-Lasinio G 1967 [Il Nuovo Cimento A (1965-1970)]{} [52]{} 807–838 ISSN 1826-9869 <https://doi.org/10.1007/BF02738845> Hopfield J J 1982 [PNAS]{} [79]{} 2554–2558 Gabri[é]{} M, Tramel E W and Krzakala F 2015 Training restricted boltzmann machines via the thouless-anderson-palmer free energy [Proceedings of the 28th International Conference on Neural Information Processing Systems - Volume 1]{} NIPS’15 (Cambridge, MA, USA: MIT Press) pp 640–648 <http://dl.acm.org/citation.cfm?id=2969239.2969311> Carreira-Perpiñán M A and Hinton G E 2005 On contrastive divergence learning [Artificial Intelligence and Statistics]{} p 8 Smolensky P 1986 Information processing in dynamical systems: Foundations of harmony theory [Parallel Distributed Processing: Explorations in the Microstructure of Cognition, Vol. 1]{} ed Rumelhart D E, McClelland J L and PDP Research Group C (Cambridge, MA, USA: MIT Press) pp 194–281 ISBN 0-262-68053-X <http://dl.acm.org/citation.cfm?id=104279.104290> Hinton G E and Salakhutdinov R R 2006 [Science]{} [313]{} 504–507 Tka[č]{}ik G, Schneidman E, Berry II M and Bialek W 2006 [arxiv]{} q–bio/0611072 Roudi Y, Tyrcha J and Hertz J 2009 [Phys. Rev. E]{} [79]{}(5) 051915 <https://link.aps.org/doi/10.1103/PhysRevE.79.051915> Hertz J, Roudi Y and Tyrcha J 2011 [arxiv]{} 1106.1752 Nguyen H C, Zecchina R and Berg J 2017 [Advances in Physics]{} [66]{} 197–261 (Preprint ) <https://doi.org/10.1080/00018732.2017.1341604> Kappen H and Rodríguez F B 1998 Boltzmann machine learning using mean field theory and linear response correction [Advances in Neural Information Processing Systems]{} (The MIT Press) pp 280–286 Sessak V and Monasson R 2009 [Journal of Physics A: Mathematical and Theoretical]{} [42]{} 055001 <http://stacks.iop.org/1751-8121/42/i=5/a=055001> Jacquin H and Ran çon A 2016 [Phys. Rev. E]{} [94]{}(4) 042118 <https://link.aps.org/doi/10.1103/PhysRevE.94.042118> Roudi Y, Aurell E and Hertz J A 2009 [Front. in Comput. Neurosc.]{} [3]{} 1–15 Schneidman E, Berry M J, Segev R and Bialek W 2006 [Nature]{} [440]{} 1007–1012 Tka[č]{}ik G, Marre O, Amodei D, Schneidman E, Bialek W and Berry II M J 2014 [PLoS Comp Biol]{} [10]{} e1003408 Weinberg S 2005 [The Quantum theory of fields - volume II]{} (Cambridge University Press) Vasiliev A N and Kazanskii A K 1973 [Theoretical and Mathematical Physics]{} [14]{} 215–226 ISSN 1573-9333 <https://doi.org/10.1007/BF01029302> Vasiliev A N, Kazanskii A K and Pis’mak Y M 1974 [Theoretical and Mathematical Physics]{} [19]{} 443–453 ISSN 1573-9333 <https://doi.org/10.1007/BF01035944> Bravi B, Sollich P and Opper M 2016 [Journal of Physics A: Mathematical and Theoretical]{} [49]{} 194003 <http://stacks.iop.org/1751-8121/49/i=19/a=194003> Berlin T H and Kac M 1952 [Phys. Rev.]{} [86]{}(6) 821–835 <https://link.aps.org/doi/10.1103/PhysRev.86.821> Antenucci F, Crisanti A and Leuzzi L 2015 [Phys. Rev. A]{} [91]{}(5) 053816 <https://link.aps.org/doi/10.1103/PhysRevA.91.053816> Marruzzo A, Tyagi P, Antenucci F, Pagnani A and Leuzzi L 2017 [Scientific Reports]{} [7]{} 1–8 <https://doi.org/10.1038/s41598-017-03163-4> Greitemann J and Pollet L 2018 [SciPost Phys. Lect. Notes]{} [1]{} 1–41 <https://scipost.org/10.21468/SciPostPhysLectNotes.2> Van Houcke K, Werner F, Kozik E, Prokofiev N, Svistunov B, Ku M J H, Sommer A T, Cheuk L W, Schirotzek A and Zwierlein M W 2012 [Nature Physics]{} [8]{} 366 <http://dx.doi.org/10.1038/nphys2273> Van Houcke K, Kozik E, Prokofiev N and Svistunov B 2008 [arxiv preprint: arXiv:0802.2923]{} <https://arxiv.org/abs/0802.2923> Salpeter E E and Bethe H A 1951 [Phys. Rev.]{} [84]{}(6) 1232–1242 <https://link.aps.org/doi/10.1103/PhysRev.84.1232> [^1]: In comparing the expressions to these works, note the different sum conventions: Georges et al. and Nakanishi et al. are only summing over those tuples of distinct indices that lead to different terms, while we allow multiple occurrence of terms already in the definition of the action and correct this by using $\frac{J_{ij}}{2}$ instead of $J_{ij}$ as the interaction. Note that the former sum convention in the second-to-last line of eq. (11) in Georges et al. amounts to a summation over all tuples of three different indices and all even permutations thereof, whereas in eq. (15) of Nakanishi et al, these permutations are written out explicitly, therefore the summation is only over the tuples with three different indices. Therefore, even if these two expressions seem to disagree by a factor $3$ at first sight, they do not, because, depending on the form of the sum, the interpretation of their sum convention changes. We thank Adam Rançon for clarifying this summation convention.	{ "pile_set_name": "ArXiv" }
--- abstract: 'We study the effect of Kerr type nonlinear medium in quantum state transfer. We have investigated the effect of different coupling schemes and Kerr medium parameters $p$ and $\omega_{{K}}$. We found that, the Kerr medium introduced in the connection channel can act like a controller for quantum state transfer. The numerical simulations are performed without taking the adiabatic approximation. Rotating wave approximation is used in the atom-cavity interaction only in the lower coupling regime.' author: - \| Manosh T.M., Muhammed Ashefas C.H. and Ramesh Babu Thayyullathil\ Department of Physics\ Cochin University of Science and Technology, India - 682022\ $^{\dagger}[email protected] bibliography: - 'reference.bib' date: title: Effects of Kerr medium in coupled cavities on quantum state transfer --- Keywords: Quantum optics, quantum information processing, coupled cavities, Kerr medium. Introduction ============ Quantum information processing (QIP) need an effective implementation of quantum state transferring schemes. The implementation of such systems with optical cavity has been a state of art. The simplest quantum description of cavity system are well described in the literature[@1443594; @2012arXiv1203.2410H; @PhysRevA.50.1785]. Modification on the Jaynes Cummings model (JCM) has been an active field of research ever since. The two level system (TLS) was then extended to multilevel, multi-cavity, multi atom TLS, [@0954-8998-3-2-005; @Seke:85; @Seke:92]etc. Optical cavities are very good candidate for quantum information processing (QIP)[@2004quant.ph..5030C]. The nonlinear optical behavior due to $\chi^{(2)}$, Kerr nonlinearity ($\chi^{(3)}$) etc. has also been used to modify these optical cavities and are extensively studied both theoretically as well as experimentally [@scully_zubairy_1997; @10.1007/978-3-540-68484-8]. Implementation of such optical cavities and its use in QIP has been an area of active research since late 1960s [@PhysRev.155.980; @Puri2017]. State transfer in coupled cavities appears to be a reliable platform for data transfer in QIP [@2004quant.ph..5030C; @Man2015; @PhysRevA.75.022310; @PhysRevA.75.032331; @Nohama2007; @PhysRevA.93.062339]. Here we discuss coupled cavity systems with and without Kerr type nonlinearity. In this paper we discuss the quantum state transfer (QST) in a linearly coupled cavity array with and without Kerr medium. We introduce the Kerr medium in the connection channel alone. We found that, the presence of a Kerr medium can be used as a controller in the QST. Linearly coupled cavities ========================= A quantum state carries an information, which has to be transfered from one place to another without any loss or modification. As quantum mechanics follows no cloning theorem[@Wootters1982], it is not possible to send an exact copy of the information[@Park1970]. Here we consider a quantum state in the cavity 1 which has to be transfered to a cavity 3, through an intermediate cavity 2. The system is illustrated in the figure (\[figure1\]). ![Three linearly coupled cavity with one qubit at either ends.[]{data-label="figure1"}](figure1){width="8cm" height="3cm"} Here in all the 3 cavities we have single mode photon field with a two level atom (or qubit) in 1st and 3rd cavities. The system can be described by the Hamiltonian, $$\hat{H}=\hat{H}_0+\hat{H}_I, \label{equation1}$$ where, $\hat{H}_0$ is the free Hamiltonian and $\hat{H}_I$ is the interaction Hamiltonian and we have $$\hat{H}_0=\hbar\omega_ca^{\dagger}_1a_1+\hbar\omega_ca^{\dagger}_2a_2+\hbar\omega_ca^{\dagger}_3a_3+\frac{\hbar}{2}\omega_a\sigma_z^{(1)}+\frac{\hbar}{2}\omega_a\sigma_z^{(3)},$$ $$\begin{split} \hat{H}_I=&\lambda_1\left(a^{\dagger}_1\sigma_-^{(1)}+a_1\sigma_+^{(1)}\right)+\lambda_3\left(a^{\dagger}_3\sigma_-^{(3)}+a_3\sigma_+^{(3)}\right)\\ &+J_{12}\left(a_1^{\dagger}a_2+a_1a_2^{\dagger}\right)+J_{23}\left(a_2^{\dagger}a_3+a_2a_3^{\dagger}\right), \end{split}$$ Here $\lambda_i\ (i=1,3)$ are the atom field coupling constant, $J_{lm}$ are the coupling strength between the cavities $l$ and $m$, $a_i\ (i=1,2,3)$ denotes the field annihilation operator and $\sigma_z^{i},\sigma_+^{i}$ and $\sigma_-^{i}\ (i=1,3)$ are the atomic operators for the $i$th cavity. A tensor product state of the system can be written as, $$\ket{\psi}=\ket{k_1n_1n_2k_3n_3}, \label{equation4}$$ where $k_i=0$ and $k_i=1$ correspond to ground and excited state respectively of the atom (qubit) in the $i$th cavity and $n_i$ represents the number of photons in the $i$th cavity. Thus, if we consider a state with maximum of one excitation at a time, the corresponding general state may be written as $$\begin{split} \ket{\psi(t)}=& q_1(t)\ket{10000}+f_1(t)\ket{01000}+f_2(t)\ket{00100}\\&+q_3(t)\ket{00010}+f_3(t)\ket{00001}, \end{split}$$ where $q_i(t)$ and $f_i(t)$ respectively are the atomic and field excitation coefficients in the $i$th cavity. The dynamics of the system can now be studied by solving the corresponding Schrödinger equation. For convenience, we can take the atomic transition frequency, $\omega_a$ and the field frequency, $\omega_c$ as the same. Thus the detuning, $\Delta=\omega_a-\omega_c=0$ and we denote, $\omega_a=\omega_c=\omega$. The state vector in the interaction picture, $$\ket{\psi(t)}_I=e^{\frac{i}{\hbar}\hat{H}_0t}\ket{\psi(t)}_S,$$ satisfies the evolution equation, $$\hat{\mathcal{H}}_I\ket{\psi(t)}_I=i\hbar\frac{\partial}{\partial t}\ket{\psi(t)}_I,$$ where we have $$%\begin{split} \hat{\mathcal{H}}_I=e^{\frac{i}{\hbar}\hat{H}_0t}\hat{H}_Ie^{-\frac{i}{\hbar}\hat{H}_0t}=\hat{H}_I, %\end{split}$$ since $\left[\hat{H}_0,\hat{H}_I\right]=0$ because the detuning is set as zero. The differential equations for $q_i(t)$ and $f_i(t)$ can be obtained as, $$\begin{aligned} i\hbar\frac{\partial}{\partial t}q_1(t) &= \lambda_1f_1(t), \label{equation9a}\\ i\hbar\frac{\partial}{\partial t}f_1(t) &= \lambda_1q_1(t)+J_{12}f_2(t),\label{equation9b}\\ i\hbar\frac{\partial}{\partial t}f_2(t) &= J_{12}f_1(t)+J_{23}f_3(t),\label{equation9c}\\ i\hbar\frac{\partial}{\partial t}q_3(t) &= \lambda_3f_3(t),\label{equation9d}\\ i\hbar\frac{\partial}{\partial t}f_3(t) &= \lambda_3q_3(t)+J_{23}f_2(t). \label{equation9e}\end{aligned}$$ These equations can be solved numerically and we can investigate how the coupling parameters affects the quantum state transfer. Analytical approach ------------------- The Laplace transform of the equations (\[equation9a\]) to (\[equation9e\]) can be written as, ($\hbar=1$) $$\begin{aligned} i\left[q_1(0)+sQ_1(s)\right] &= \lambda_1F_1(s)\label{equation10a},\\ i\left[f_1(0)+sF_1(s)\right] &= \lambda_1Q_1(s)+J_{12}F_2(s)\label{equation10b},\\ i\left[f_2(0)+sF_2(s)\right] &= J_{12}F_1(s)+J_{23}F_3(s)\label{equation10c},\\ i\left[q_3(0)+sQ_3(s)\right] &= \lambda_3F_3(s)\label{equation10d},\\ i\left[f_3(0)+sF_3(s)\right] &= \lambda_3Q_3(s)+J_{23}F_2(s). \label{equation10e}\end{aligned}$$ For further simplicity we can take the value of $\lambda_1=\lambda_3=\lambda$ and $J_{12}=J_{23}=J$. Now solving the equations (\[equation10a\]) to (\[equation10e\]) for an initial condition, $q_1(0)=1$ and $q_2(0)=f_1(0)=f_2(0)=f_3(0)=0$, results in, $$\begin{aligned} Q_1(s) &=- \frac{J^{2} \lambda^{2} + 2.0 J^{2} s^{2} + \lambda^{2} s^{2} + s^{4}}{s \left(2.0 J^{2} \lambda^{2} + 2.0 J^{2} s^{2} + \lambda^{4} + 2.0 \lambda^{2} s^{2} + s^{4}\right)}\label{equation11a},\\ F_1(s) &= \frac{i \lambda \left(J^{2} + \lambda^{2} + s^{2}\right)}{2.0 J^{2} \lambda^{2} + 2.0 J^{2} s^{2} + \lambda^{4} + 2.0 \lambda^{2} s^{2} + s^{4}}\label{equation11b},\\ F_2(s) &= \frac{J \lambda}{s \left(2.0 J^{2} + \lambda^{2} + s^{2}\right)}\label{equation11c},\\ Q_3(s) &= - \frac{J^{2} \lambda^{2}}{s \left(2.0 J^{2} \lambda^{2} + 2.0 J^{2} s^{2} + \lambda^{4} + 2.0 \lambda^{2} s^{2} + s^{4}\right)}\label{equation11d},\\ F_3(s) &= - \frac{i J^{2} \lambda}{2.0 J^{2} \lambda^{2} + 2.0 J^{2} s^{2} + \lambda^{4} + 2.0 \lambda^{2} s^{2} + s^{4}}. \label{equation11e}\end{aligned}$$ Now taking the inverse Laplace transform of equations (\[equation11a\]) to (\[equation11e\]), we obtain, $$\begin{aligned} q_1(t)&=-\frac{1}{2}\left( \cos (\lambda t)+\frac{\lambda ^2 \cosh \left(t \sqrt{-2 J^2-\lambda ^2}\right)+2 J^2}{4 J^2+2\lambda ^2}\right)\label{equation12a},&\\ q_3(t)&=\frac{1}{2} \left(\cos (\lambda t)-\frac{\lambda ^2 \cosh \left(t \sqrt{-2 J^2-\lambda ^2}\right)+2 J^2}{2 J^2+\lambda ^2}\right)\label{equation12b},&\\ f_1(t)&=\frac{1}{2} i \left(\sin (\lambda t)+\frac{\lambda \sinh \left(t \sqrt{-2 J^2-\lambda ^2}\right)}{\sqrt{-2 J^2-\lambda ^2}}\right)\label{equation12c},&\\ f_3(t)&=-\frac{1}{2} i \left(\sin (\lambda t)-\frac{\lambda \sinh \left(t \sqrt{-2 J^2-\lambda ^2}\right)}{\sqrt{-2 J^2-\lambda ^2}}\right)\label{equation12d},&\\ f_2(t)&=-\frac{J \lambda \left(\cosh \left(t \sqrt{-2 J^2-\lambda ^2}\right)-1\right)}{2 J^2+\lambda ^2}\label{equation12e}.& \end{aligned}$$ The probability can now be calculated as $$\begin{split} \left\|q_1(t)\right\|^2=&\frac{1}{4}\left\| \frac{\left(2 J^2+\left(2 J^2+\lambda ^2\right) \cos (t \lambda )+\lambda ^2 \cosh \left(t \sqrt{-2 J^2-\lambda ^2}\right)\right) }{\left(2 J^2+\lambda ^2\right) \left(2 \left(J^\right)^2+\left(\lambda ^\right)^2\right)}\right\|\times \\ &\left\| \frac{\left(2 \left(\cos (t \lambda )^+1\right) \left(J^\right)^2+\left(\lambda ^\right)^2 \left(\cos (t \lambda )^+\cosh \left(\left(t \sqrt{-2 J^2-\lambda ^2}\right)^\right)\right)\right)}{\left(2 J^2+\lambda ^2\right) \left(2 \left(J^\right)^2+\left(\lambda ^\right)^2\right)}\right\| \label{equation13}. \end{split}$$ $$\begin{aligned} \left\|q_3(t)\right\|^2&=\frac{1}{4}\left\| \frac{\left(2 J^2-\left(2 J^2+\lambda ^2\right) \cos (t \lambda )+\lambda ^2 \cosh \left(t \sqrt{-2 J^2-\lambda ^2}\right)\right)^2}{\left(2 J^2+\lambda ^2\right)^2}\right\| \label{equation14}.\\ \left\|f_1(t)\right\|^2&= \frac{1}{4}\left\| \frac{\left(\sqrt{-2 J^2-\lambda ^2} \sin (t \lambda )+\lambda \sinh \left(t \sqrt{-2 J^2-\lambda ^2}\right)\right)^2}{2 J^2+\lambda ^2}\right\| \label{equation15}.\\ \left\|f_2(t)\right\|^2&=\frac{1}{4} e^{-2 \Re\left(t \sqrt{-2 J^2-\lambda ^2}\right)}\left\| \frac{\left(-1+e^{t \sqrt{-2 J^2-\lambda ^2}}\right)^2 \left(-1+e^{\left(t \sqrt{-2 J^2-\lambda ^2}\right)^}\right)^2 J^2 \lambda ^2}{\left(2 J^2+\lambda ^2\right) \left(2 \left(J^\right)^2+\left(\lambda ^\right)^2\right)}\right\| \label{equation16}.\\ \left\|f_3(t)\right\|^2&=\frac{1}{16} \left\| 2 \sin (t \lambda )-\frac{2 \lambda \sinh \left(t \sqrt{-2 J^2-\lambda ^2}\right)}{\sqrt{-2 J^2-\lambda ^2}}\right\| ^2 \label{equation17}.\end{aligned}$$ using equation (\[equation13\]) to (\[equation17\]), we can define the population inversion of each qubit as, $$\braket{\sigma_z^{(1)}}=\left\|q_1(t)\right\|^2-\left(\left\|q_3(t)\right\|^2+\left\|f_1(t)\right\|^2+\left\|f_2(t)\right\|^2+\left\|f_3(t)\right\|^2\right) \label{equation18},$$ $$\braket{\sigma_z^{(3)}}=\left\|q_3(t)\right\|^2-\left(\left\|q_1(t)\right\|^2+\left\|f_1(t)\right\|^2+\left\|f_2(t)\right\|^2+\left\|f_3(t)\right\|^2\right) \label{equation19}.$$ Linearly coupled cavities with Kerr medium ========================================== Nonlinear effects in optical cavities has been studied in the literature [@10.1007/978-3-540-68484-8; @PhysRev.155.980]. This nonlinear effects can be used to construct effective transfer mechanism in quantum engineering[@PhysRevA.85.023826; @Puri2017]. Here we consider the 3rd order nonlinearity, widely known as Kerr nonlinearity and we introduce such a nonlinear effect in the second cavity of the previously described system shown in the figure (\[figure1\]) and we get the modified system as in figure (\[figure2\]). ![3 linearly coupled cavity with one qubit at either ends and a Kerr medium in the second cavity[]{data-label="figure2"}](figure2){width="8cm" height="3cm"} Hamiltonian with a Kerr type nonlinear medium in the second cavity is described by [@doi:10.1080/09500349114550961; @PhysRevA.39.2969] $$\hat{H}_{{K}}=\omega_{{K}}b^{\dagger}b+qb^{\dagger2}b^2+p\left(a_2^{\dagger}b+a_2b^{\dagger}\right).$$ where $b$ is the annihilation operator of the Kerr medium, $\omega_{{K}}$ denotes the anharmonic Kerr field frequency, $q$ is the anharmonicity parameter and $p$ represents the field-Kerr medium coupling strength. In the adiabatic limit the field frequency $\omega$ and medium frequency $\omega_{{K}}$ are very different. Now the Hamiltonian of the new system takes the form, $$\hat{\mathcal{H}}=\hat{H}_0+\hat{H}_I+\hat{H}_{{K}}=\hat{\mathcal{H}}_0+\hat{\mathcal{H}}_I. \label{equation21}$$ where the new free and interaction Hamiltonians are respectively given as, $$\hat{\mathcal{H}}_0=\hat{H}_0+\omega_{{K}}b^{\dagger}b+qb^{\dagger 2}b^2,$$ $$\hat{\mathcal{H}}_I=\hat{H}_I+p\left(a_2^{\dagger}b+a_2b^{\dagger}\right).$$ With the Kerr medium operator $b$ and $b^{\dagger}$, we need to extend the Hilbert space of states given in equation(\[equation4\]) and it get modified to to accommodate the new state, which can be defined as, $$\ket{\psi}=\ket{k_1n_1n_2k_3n_3n_b}. \label{equation24}$$ where $n_b$ is the bosonic number of the Kerr medium. Here also we can find the Hamiltonian in the interaction picture and we can show that $$\left[\hat{\mathcal{H}}_0,\hat{\mathcal{H}}_I\right]=0.$$ The dynamics can be studied by obtaining the differential equations similar to equations (\[equation9a\]) to (\[equation9e\]) and solving it. Analytical approach ------------------- If we take only a maximum of one excitation in the cavity we may write the general state as $$\begin{split} \ket{\Psi(t)}=& q_1(t)\ket{100000}+f_1(t)\ket{010000}+f_2(t)\ket{001000}\\&+q_3(t)\ket{000100}+f_3(t)\ket{000010}+k(t)\ket{000001}. \end{split}$$ The differential equations for $q_i(t)$, $f_i(t)$ and $k(t)$ can be obtained as, $$\begin{aligned} i\hbar\frac{\partial}{\partial t}q_1(t) &= \lambda_1f_1(t)\label{equation27a},\\ i\hbar\frac{\partial}{\partial t}f_1(t) &= \lambda_1q_1(t)+J_{12}f_2(t)\label{equation27b},\\ i\hbar\frac{\partial}{\partial t}f_2(t) &= J_{12}f_1(t)+J_{23}f_3(t)+pk(t)\label{equation27c},\\ i\hbar\frac{\partial}{\partial t}q_3(t) &= \lambda_3f_3(t)\label{equation27d},\\ i\hbar\frac{\partial}{\partial t}f_3(t) &= \lambda_3q_3(t)+J_{23}f_2(t)\label{equation27e},\\ i\hbar\frac{\partial}{\partial t}k(t) &=\left(\omega_k-\omega-q\right)k(t)+pf_2(t)\label{equation27f}.\end{aligned}$$ The Laplace transform of equations (\[equation27a\]) to (\[equation27f\]) are, ($\hbar=1$) $$\begin{aligned} i\left[q_1(0)+sQ_1(s)\right] &= \lambda_1F_1(s)\label{equation28a},\\ i\left[f_1(0)+sF_1(s)\right] &= \lambda_1Q_1(s)+J_{12}F_2(s)\label{equation28b},\\ i\left[f_2(0)+sF_2(s)\right] &= J_{12}F_1(s)+J_{23}F_3(s)+pK(s)\label{equation28c},\\ i\left[q_3(0)+sQ_3(s)\right] &= \lambda_3F_3(s)\label{equation28d},\\ i\left[f_3(0)+sF_3(s)\right] &= \lambda_3Q_3(s)+J_{23}F_2(s)\label{equation28e},\\ i\left[k(0)+sK(s)\right] &= \left(\omega_k-\omega-q\right)K(s)+pF_2(s)\label{equation28f}.\end{aligned}$$ For Kerr medium analytical solutions of can be more rigorous to handle. We do not take the adiabatic approximation even when the $\omega_{{K}}$ is far away from $\omega$ and we investigated the evolution of the system numerically. [@JOHANSSON20131234] Results and discussion ====================== Quantum state transfer without Kerr medium ------------------------------------------ First we consider quantum state transfer for different coupling schemes, without an intermediate Kerr medium and we can see that, there is a transfer of state from the qubit 1 to qubit 2 . This time can be controlled by controlling the coupling parameter. With $J_{lm}=0.1\lambda_i$ and $J_{lm}=0.2\lambda_i$, the results are shown in figures (\[figure3\_3a\]) and (\[figure3\_3b\]). The effect of coupling scheme is evident from these simulations. The photon number, $\braket{n}$ in each cavities are also estimated for $J_{lm}=0.2\lambda_i$. Figure (\[figure4\]) shows the corresponding results. Here we have taken $\omega_a=\omega_c=\omega=2\pi f$ and $f$ is taken to be $1$GHz for the simulations. The graphs are plotted against scaled time. Here in all the plotting the unit of scaled time is in nanoseconds. All other parameters are scaled with respect to the unit of $\omega$. The photon number also follows a pattern similar to the population inversion, such that whenever the qubit is in the excited state, photon number becomes zero inside the respective cavity. However, the intermediate cavity shows a rise in the photon number as the coupling in the system is increased so that, the photon number in the first and last cavity has reduced. This suggest that the probability of inversion is not 100% in any case with an intermediate coupling cavity with a non zero cavity-cavity coupling. So the expense of a controlled transmission with an intermediate coupling cavity is the quality of the transmission. The value of qubit-cavity coupling is kept at, $\lambda=0.1\omega$, which allows us to use RWA [@Tur2000]. ![Photon number in 3 cavities $\braket{n_1}$, $\braket{n_2}$ and $\braket{n_3}$ for $J_{lm}=0.2\lambda_i$.The range of time differs in the plots on two columns.[]{data-label="figure4"}](figure4){width="\linewidth"} The analytical solutions including detuning is very cumbersome. The effect of detuning on the state transfer is shown in figures (\[figure5\]) and (\[figure6\]). We can clearly see that detuning affects the state transfer. ![Population inversion of qubit 1 and qubit 2 for $J_{lm}=0.2\lambda_i$ and $\omega_c=0.99 \omega_a$ ($\Delta=0.01\omega_a$)[]{data-label="figure5"}](figure5){width="\linewidth"} ![Photon number in 3 cavities $\braket{n_1}$, $\braket{n_2}$ and $\braket{n_3}$ for $J_{lm}=0.2\lambda_i$ and $\omega_c=0.99 \omega_a$($\Delta=0.01\omega_a$). The range of time differs in the plots on two columns.[]{data-label="figure6"}](figure6){width="\linewidth"} Quantum state transfer with Kerr medium --------------------------------------- The presence of a Kerr medium in the intermediate cavity can affect the state transfer. Numerical simulations are done for different Kerr-cavity coupling value, $p$ and $\omega_{{K}}$. The results are shown in figures (\[figure7\_7a\]), (\[figure7\_7b\]), (\[figure7\_7c\]) and (\[figure7\_7d\]). Here we do not take the adiabatic approximation for the Kerr Hamiltonian. For $p\approx0.5 \omega_c$, $\omega_{{K}}\approx\omega_c$ and $J_{lm}=0.5\lambda_i$, the nature of population inversion is almost equivalent to the case where there is no Kerr medium in the second cavity and $J_{lm}=0.1\lambda_i$. The results are shown in figures (\[figure8\_8a\]) and (\[figure8\_8b\]) Thus with higher coupling between the cavities, we can have a controlled state transfer between two qubits, by means of a Kerr medium in the intermediate cavity. Conclusion ========== In the present work we have numerically studied a system of 3 linearly coupled cavities with one qubit in either ends of the cavity and an intermediate cavity in between them. Our study focused on the presence of a Kerr medium in the second cavity and how it affect the quantum state transfer. We found that the presence of Kerr medium can affect the transmission and hence can be used as a quantum state transfer controller in quantum information processing. Without taking the adiabatic approximation in the Kerr medium, there can be a controlled state transfer. All the plotting are done with a scaling corresponds to the cavity frequency, which set at $1$ GHz. We have only taken the the RWA in the appropriate limit. Acknowledgments {#acknowledgments .unnumbered} =============== The authors, MTM and RBT would like to thank the financial support from KSCSTE, Government of Kerala State, under Emeritus Scientist scheme.	{ "pile_set_name": "ArXiv" }
--- abstract: 'We use the remark that, through Bargmann-Fock representation, diagonal operators of the Heisenberg-Weyl algebra are scalars for the Hadamard product to give some properties (like the stability of periodic fonctions) of the Hadamard product by a rational fraction. In particular, we provide through this way explicit formulas for the multiplication table of the Hadamard product in the algebra of rational functions in $\C[[z]]$.' author: - \| [G. H. E. Duchamp, S. Goodenough]{} [^1]\ [and ]{}[K. A. Penson]{} [^2] title: \| Rational Hadamard products via\ Quantum Diagonal Operators --- Heisenberg-Weyl algebra and Bargmann-Fock\ representation ========================================== Let $\H=l^2(\N,\C)$, the standard separable Hilbert space of complex-valued sequences $(\al_n)_{n\in \N}$ such that $\sum_{n\in\N}\|\al_n\|^2<\infty$ and $e_n=(\delta_{m,n})_{m\in \N}$ be its the canonical Hilbert basis with $$\H_0=\bigoplus_{n\in \N} \C e_n$$ the standard dense subspace of $\H$.\ Using the classical [bra]{} and [ket]{} denotations, one defines the following operators $$a=\sum_{n\geq 1}\sqrt{n}\ \|n-1\rangle\langle n\|\ ;\ a^\dag=\sum_{n\geq 0}\sqrt{n+1}\ \|n+1\rangle\langle n\|$$ where “$\|n\rangle$” (resp. “$\langle n\|$”) stand for “$\|e_n\rangle$” (resp. “$\langle e_n\|$”). By definition, $a,a^\dag$ are hermitian conjugates and densely defined (because $\H_0$ is a subspace of $dom(a)$ and $dom(a^\dag)$) and one verifies easily that $[a,a^\dag]=Id_{\H_0}$. Due to their proeminent importance in quantum mechanics, there is a huge litterature on these operators and it can be proved that - the algebra (in $End(\H_0)$) generated by $a,a^\dag$, can be presented as $$HW_\C=\Big< a,a^\dag\ ;\ aa^+-a^+a=1\Big>_{\C-AAU}$$ where $\C-AAU$ is the category of $\C$-associative algebras with unit. - there is no non-zero representation of $HW_\C$ in a Banach algebra - the representation $\be$ on the space of complex formal power series $\C[[z]]$ by operators $\be(a),\be(a^\dag)$ such that $\be(a)(S)=\frac{d}{dz}S\ ;\ \be(a^\dag)(S)=zS$, known as Bargmann-Fock representation, is faithful - due to the (only) relator $aa^+-a^+a=1$, the family $$\Big((a^\dag)^k a^l\Big)_{k,\, l\geq\, 0}$$ is a basis of $HW_\C$ (the expressions w. r. t. this basis are called “normally ordered”) - denoting $HW^{(e)}_\C$ the subspace generated by the “monomials” $\Big((a^\dag)^k a^l\Big)_{k-l=e}$, one has $$\label{gr_alg} HW^{(e_1)}_\C.HW^{(e_2)}_\C\subset HW^{(e_1+e_2)}_\C$$ so that the algebra $HW_\C$ is $\Z$-graded (the parameter $e$ is called “degree” by algebraists and “excess/defect” -according to its sign - by physicists). Despite of its simple definition $HW_\C$ supports very rich combinatorial studies [@BBM; @GOF7; @GOF8; @GOF4; @FPSAC07; @OPG; @GOF6; @D20; @GOF2; @GOF3; @GOF5; @Varvak]. Problem of the rational Hadamard table ====================================== The Hadamard product on generating functions was introduced by the French mathematician Jacques Hadamard [@Ha] as a shifted convolution product on the one-dimensional torus ($S^1$, the commutative group of angles). This product (denoted $\odot$ in the sequel) amounts to performing the pointwise product on the coefficients of the expanded sequences $$\Big(\sum_{n=0}^\infty a_n z^n\Big)\odot \Big(\sum_{n=0}^\infty b_n z^n\Big):=\Big(\sum_{n=0}^\infty a_nb_n z^n\Big).$$ In [@Ju], it is proved that the Hadamard product of an algebraic and a rational series is algebraic (this theorem was extended later to the noncommutative case by Schützenberger [@Sc1] and the latter could be used as a crucial result in [@D21]). Hence the algebraic series are a module over the algebra of rational series and it remains the combinatorial problem of giving explicit formulas for the multiplication of two rational series i. e. elements of $K[[z]]$ of the form $\frac{P(z)}{Q(z)}$ with $Q(0)\not=0$. This problem is called here the “Rational Hadamard product problem”. We can make this problem very precise in the case when the coefficients are taken in an algebraically closed field (for the sake of readability, it will be here taken equal to $\C$, the field of complex numbers) as the algebra of rational series admits $$\label{rat_lin_basis} (z^n)_{n\in \N} ; \Big(\frac{1}{(1-\al z)^m}\Big)_{\al\in \C^\atop m\in \N^}$$ as linear basis. The aim of this paper is to show how to compute explicitely the multiplication table of this algebra with respect to the former basis using the classical creation/annihilation operators. To this end, we will use the algebra generated by the two operators on $\C[[z]]$ $$\be(a^+): S\ra zS\ ;\ \be(a): S\ra \frac{d}{dz} S\ .$$ These formulas define (on the vector space of complex series) a faithful representation of the Heisenberg-Weyl algebra $HW_\C$.\ It follows from Eq that $HW^{(0)}_\C$ is a subalgebra of $HW_\C$, called the algebra of diagonal operators. In order to compute the Hadamard products of the elements of the basis , one remarks that the diagonal operators are scalars for the Hadamard product. Indeed, as $HW^{(0)}_\C$ is the linear span of the monomials $(a^+)^na^n$, it is sufficient to prove that, for all $S,T\in \C[[z]]$, $$\be((a^+)^na^n)[S]\odot T=S\odot \be((a^+)^na^n)[T]$$ which is easily seen by direct computation $$\begin{aligned} \be((a^+)^na^n)[S]\odot T&=& \sum_{m=0}^\infty \scal{\be((a^+)^na^n)[S]}{z^m}\scal{T}{z^m}z^m= \sum_{m=0}^\infty \frac{n!}{m!}\scal{S}{z^m}\scal{T}{z^m}z^m\cr &=&\sum_{m=0}^\infty \scal{S}{z^m}\scal{\be((a^+)^na^n)[T]}{z^m}z^m=S\odot \be((a^+)^na^n)[T] \end{aligned}$$ Results ======= In this section, we prove that the Hadamard multiplication operator $$f\mapsto \frac{1}{(1-z)^{k+1}}\odot f$$ is a diagonal operator and hence, the corresponding multiplication operator by $\frac{1}{(1-\al z)^{k+1}}$ is the composition of a diagonal operator and a dilation (i. e. a substitution $z\mapsto \al z$) which is itself a Hadamard multiplication operator as $$f(\al z)=\frac{1}{1-\al z}\odot f(z)\ .$$ For $k\in\N,\ \al\in \C^$, let $D_{(\al,k)}$ be the operator “Hadamard multiplication by $\frac{1}{(1-\al z)^{k+1}}$” $$f\mapsto \frac{1}{(1-\al z)^{k+1}}\odot f$$ Then, one has\ 1) i) $D_{(1,k)}$ is a diagonal operator through the Bargmann-Fock repesentation $\beta$. More precisely $$D_{(1,k)}=\be\Big(\sum_{j=0}^k \frac{\binomial{k}{j}}{j!} (a^+)^ja^j\Big)$$ ii) $D_{(\al,0)}$ is the substitution $z\ra \al z$ (automorphism for $\al\not=0$).\ iii) Due to the associativity of the Hadamard product, one has $$D_{(\al,k)}=D_{(\al,0)}\circ D_{(1,k)}=D_{(1,k)}\circ D_{(\al,0)}$$ These formulas on diagonal operators yield following theorem. The multiplication table of the algebra of rational power series is the following $$\label{mult_by_monomials} z^n\odot z^m=\delta_{m,n}z^n\ ;\ z^n\odot \frac{1}{(1-\al z)^{m+1}}=\binomial{-(m+1)}{n}\al^nz^n$$ and $$\begin{aligned} \label{mult_by_inverses} \frac{1}{(1-\al z)^{k+1}}\odot \frac{1}{(1-\al z)^{l+1}}= D_{(\al\be,0)}\Big[\sum_{j=0}^{k} \frac{\binomial{k}{j}}{j!} (l+1)...(l+j)\frac{z^j}{(1-z)^{l+j+1}}\Big]=\cr D_{(\al\be,0)}\Big[\sum_{j=0}^{k} \frac{\binomial{k}{j}}{j!} (l+1)^{\bar j} \sum_{s=0}^j \binomial{j}{s} \frac{(-1)^{j-s}}{(1-z)^{l+s+1}}\Big] \end{aligned}$$ The algebra of rational power series $\ncp{\C}{z}^\circ$ is closed under the action of $HW_\C$ Concluding remarks ================== In a theorem of G. Cauchon (reproved by other means in [@D21], see also [@BR]), one has to consider the algebra of rational series in the sense of Neumann-Malcev algebras. In one variable, it is the field of fractions $$\frac{P(z)}{Q(z)}\ ;\ P,Q\in \C[z],\ Q\not=0$$ within $\C((z,z^1))$, the algebra of Laurent series. This algebra being a subfield of functions $\Z\ra \C$, the Hadamard product (which is in fact the pointwise product) extends at once.\ It is immediate to see that this field of frations admits the linear basis $$\label{full_rat_lin_basis} (z^n)_{n\in \Z}\ ; \Big(\frac{1}{(1-\al z)^m}\Big)_{\al\in \C^\atop m\in \N^}$$ and the multiplication table (formulas and ) have to be extended by zero as regards the products $z^n\odot \frac{1}{(1-\al z)^{m+1}}$ with $n<0$. As a consequence, this algebra again is closed under Hadamard products. [ABC]{} , [Rational series and their languages EATCS Monographs on Theoretical Computer Science]{}, Springer (1988).\ New electronic version\ [http://www-igm.univ-mlv.fr/ berstel/LivreSeries/LivreSeries.html]{} , Quantum field theory of partitions, J. Math. Phys. Vol 40 (1999) , [Boson normal ordering via substitutions and Sheffer-Type Polynomials]{}, Phys. Lett. A [338]{} (2005) 108 , [Some useful formula for bosonic operators]{}, Jour. Math. Phys. [46]{} 052110 (2005). , Feynman graphs and related Hopf algebras, Journal of Physics: Conference Series, SSPCM’05, Myczkowce, Poland. arXiv : cs.SC/0510041 , [Hopf algebras of diagrams]{}, submitted to FPSAC07. , [One-parameter groups and combinatorial physics]{}, Proceedings of the Symposium Third International Workshop on Contemporary Problems in Mathematical Physics (COPROMAPH3) (Porto-Novo, Benin, Nov. 2003), J. Govaerts, M. N. Hounkonnou and A. Z. Msezane (eds.), p.436 (World Scientific Publishing 2004)\ arXiv: [quant-ph/04011262]{} , [Un critère de rationalité provenant de la géométrie noncommutative]{} Invent. Math. [128]{} 613-622. (1997). , [A product formula and combinatorial field theory]{}, Proceedings of the XI International Conference on Symmetry Methods in Physics (SYMPHYS-11) (Prague, Czech Republic, June 2004), C. Burdik, O. Navratil, and S. Posta (eds.) (JINR Publishers, Dubna)\ arXiv:[quant-ph/0409152]{} , [Théorème sur les séries entières]{}, Acta Math., Uppsala, t. 22, 1899, p. 55-63. , Comment. math. Helvet., t. 3 1931, p. 266-306 (Thèse Sc. math.). , [Ordering relations for q-boson operators, continued fractions techniques, and the q-CBH enigma]{}. Journal of Physics A [28]{} 7209-7225 (1995). , On a theorem of R. Jungen, Proc. American Math. Society (1962) p. 189-197. , [Combinatorial Physics, Normal Order and Model Feynman Graphs]{}, Proceedings of the Symposium ’Symmetries in Science XIII’, Bregenz, Austria, 2003, B. Gruber, G. Marmo and N. Yoshinaga (eds.), p.527 (Kluwer Academic Publishers 2004)\ arXiv : [quant-ph/0310174]{} , [Normal Order: Combinatorial Graphs Quantum Theory and Symmetries]{}, Proceedings of the 3rd International Symposium P.C. Argyres, T.J. Hodges, F. Mansouri, J.J. Scanio, P. Suranyi, and L.C.R. Wijewardhana (eds.), p.398 (World Scientific Publishing 2004)\ arXiv:[quant-ph/0402082]{} , [Partition functions and graphs: A combinatorial approach]{}, Proceedings of the XI International Conference on Symmetry Methods in Physics (SYMPHYS-11) (Prague, Czech Republic, June 2004), C. Burdik, O. Navratil, and S. Posta (eds.) (JINR Publishers, Dubna, 2004)\ arXiv:[quant-ph/0409082]{} , [Rook numbers and the normal ordering problem*]{}, J. Comb. Theory, Ser. A 112(2): 292-307 (2005) [^1]: LIPN - UMR 7030 CNRS - Université Paris 13 F-93430 Villetaneuse, France [^2]: Laboratoire de Physique Théorique de la Matière Condensée, Université Pierre et Marie Curie, CNRS UMR 7600 Tour 24 - 2ième ét., 4 pl. Jussieu, F 75252 Paris Cedex 05, France	{ "pile_set_name": "ArXiv" }
--- author: - 'J.-F. Desmurs, R. Soria-Ruiz, F. Colomer, K.B. Marvel, V. Bujarrabal, J. Alcolea, P.J. Diamond, D. Boboltz' - 'A. Kemball' title: The Relative Spatial Distribution of SiO Masers in AGB Stars at 43 and 86 GHz --- Introduction ============ SiO masers emission at 7 mm wavelength ($v$=1 and $v$=2, J=1–0 transition near 43 GHz) has been observed in AGB stars with very hight resolution by means of VLBI techniques, yielding important results in relation with their not yet well understood pumping mechanism. The 7 mm maser emission regions are found to be distributed in a number of spots forming a ring-like structure at about 2-3 stellar radii [@col92; @dia94; @gre95; @des00] which are assumed to be centered on the stellar position. This ring-like flux distribution arise naturally in the framework of the radiative pumping mechanism of SiO masers (see Bujarrabal et al. 1994), but may also be explained by a collisional modelisation. Recent simultaneous observations of the $v$=1 and $v$=2, J=1–0 (see Desmurs et al. 2000) show that the $v$=1 and $v$=2 maser spots are often close, but appear systematically shifted by a few mas and are only rarely coincident; a result that would argue in favor of radiative pumping models. In an attempt to pursue the comparison between the predictions of the different models and observational data, we have measured the relative spatial distribution of SiO maser emission between the J=1–0 and J=2–1 transitions. In that case, models using either radiative or collisional pumping predict “maser chains” across the excitated vibrational states, in such a way that the inversion of the different-J transition in the same $v$ state are mutually reinforced. Models predict that the rotational masers in the same vibrational state should appear under the same physical conditions and, therefore, that the 43GHz and 86GHz masers emission originate from the same condensations in the circumstellar envelope (CSE) [see @buj94; @hum02]. Observations and data analysis ============================== We have performed quasi-simultaneous observations (separated only by few hours) of the J=1–0 $v$=1 and $v$=2 (at 7mm), J=2–1 $v$=1 and $v$=2 lines of $^{28}$SiO, and the J=1–0 $v$=0 line of $^{29}$SiO (at 3mm) with NRAO Very Long Baseline Array, on May 9th, 2001 of three AGB stars: TX Cam, R Cas and $\chi$ Cyg. This was possible with the new capabilities of the VLBA at 3 mm. The system was setup to record 4 MHz at 7mm in dual polarization and 8 MHz at 3mm in single polarization. The correlation was produced at the VLBA correlator in Socorro (NM, USA) providing 256 and 512 spectral channels respectively for a final spectral resolution of $\sim$0.1 and $\sim$0.05km s$^{-1}$. The calibration was performed using the standard scheme in the Astronomical Image Processing System (AIPS) for spectral line experiments; the amplitude calibration was performed applying the template spectra method. In this paper, we report only on the first results we obtained for the source $\chi$ Cyg in the $v$=1, J=1–0 (at 43122.080 MHz see Fig \[chi\_cyg\_7mm\]) and $v$=1, J=2–1 (at 86243.442 MHz see Fig \[chi\_cyg\_3mm\]) lines. Maps of the two transitions were independently produced by solving the residual fringe-rates on the line source. This correction is determined by selecting a channel containing a simple feature with a simple structure and a high signal to noise ratio, that is used as phase reference for all other channels. This fixed the position of that maser spot as the map origin, in both cases. ![Integrated intensity maps of the $v$=1, J=1–0 line (rest frequency 43122.080 MHz) of SiO masers towards $\chi$ Cyg. Contours are multiples by 10% of the peak flux ($\sim$28 Jy). The resolution beam is 0.8x0.21 mas (PA=-12.4$^{\circ}$). []{data-label="chi_cyg_7mm"}](desmurs_fig1.ps) Results and discussion ====================== We produced the first VLBA maps with milliarcsecond resolution of the SiO maser emission from an S-type star, $\chi$ Cyg. We obtained maps of the transitions $v$=1, J=1–0 at 7mm (see Fig \[chi\_cyg\_7mm\]) and of the $v$=1, J=2–1 at 3mm (see Fig \[chi\_cyg\_3mm\]). The map size is approximately $\sim$80 by 80 mas. The gaussian restoring beam has a FWHM of 0.8x0.2 mas (with a position angle PA=-12.4$^{\circ}$) and 0.75x0.05 mas (with a position angle PA=-15.8$^{\circ}$), respectively, for the data at 7 and 3mm. Both figures are using the same scale to ease a direct comparison, but do not share spatial origin as explained before. Our preliminary results shows that the SiO maser emission distribution in both transition occur more or less at a similar radii from the center. In particular, we observed that masers for $v$=1, J=2–1 (3mm) arise from a ring-like structure, as it has been reported in several other AGB stars at 43 GHz [@dia94; @gre95; @des00; @phi01]. The radii of this ring structure is of the order of $\sim$ 28mas which, adopting the Hipparcos distance (ESA 1997) of 106$\pm$15 pc, is equivalent to 4.510$^{13}$cm. One of the most surprising result of these observations is that the emission distribution between the two lines is completly different. Whatever the choosed alignment, it is impossible to make coincident more than one maser spot at the same time. Even the regions emitting in one transition or the other are very different at large scale (see Figs \[chi\_cyg\_7mm\] and \[chi\_cyg\_3mm\]). This is in complete contradiction with all theoretical predictions. ![Integrated intensity maps of the $v$=1, J=2–1 line (rest frequency 86243.442 MHz) of SiO masers towards $\chi$ Cyg. Contours are multiples by 10% of the peak flux ($\sim$36 Jy). The resolution beam is 0.75x0.052 mas (PA=-15.8$^{\circ}$). []{data-label="chi_cyg_3mm"}](desmurs_fig2.ps) Our conclusions are only preliminary. But in the case that they would be confirmed by future observations and in other sources, this result would cast serious doubts on the present models of the SiO maser excitation. Bujarrabal, V., 1994, A&A 285, 953 Colomer, F., Graham, D. A., Krichbaum, T. P., Ronnang, B. O., de Vicente, P., Witzel, A., Barcia, A., Baudry, A., Booth, R. S., Gomez-Gonzalez, J., Alcolea, J., Daigne, G, 1992, A&A 254, L17 Diamond, P., Kemball, A. J., Junor, W., Zensus, A., Benson, J., Dhawan, V., 1994, ApJ Letter 430, L61 Desmurs, J.-F., Bujarrabal, V., Colomer, F., & Alcolea, J., 2000, A&A 360, 189 Greenhill,L.J, Colomer, F., Moran, J. M., Backer, D. C., Danchi, W. C., Bester, M., 1995, ApJ 449, 365 Humphreys, E.M.L., Gray, M.D., Yates, J.A., Field, D., Bowen, G.H. & Diamond, P.J., 2002, A&A 386, 256 Phillips, R.B., Sivakoff, G.R., Lonsdale, C.J., Doeleman, S.S., 2001, AJ 122, 2674.	{ "pile_set_name": "ArXiv" }
--- bibliography: - 'mybib.bib' --- MANUSCRIPT\ Authors:Alexander Tank and Alan A. Stocker\ \ Affiliation: Department of Psychology and\ Department of Electrical and Systems Engineering\ University of Pennsylvania\ \ Correspondence: Dr. Alan A. Stocker\ Computational Perception and Cognition Laboratory\ 3401 Walnut Street 313C\ Philadelphia, PA 19104-6228\ U.S.A.\ [email protected]\ phone: +1 215 573 9341\ \ Journal: arXiv\ \ Classification: Quantiative Biology (Neurons and Cognition)\ \ \ Biased perception leads to biased action: Validating a Bayesian model of interception {#biased-perception-leads-to-biased-action-validating-a-bayesian-model-of-interception .unnumbered} ===================================================================================== [Alexander Tank and Alan A. Stocker]{} Abstract {#abstract .unnumbered} ======== We tested whether and how biases in visual perception might influence motor actions. To do so, we designed an interception task in which subjects had to indicate the time when a moving object, whose trajectory was occluded, would reach a target-area. Subjects made their judgments based on a brief display of the object’s initial motion at a given starting point. Based on the known illusion that slow contrast stimuli appear to move slower than high contrast ones, we predict that if perception directly influences motion actions subjects would show delayed interception times for low contrast objects. In order to provide a more quantitative prediction, we developed a Bayesian model for the complete sensory-motor interception task. Using fit parameters for the prior and likelihood on visual speed from a previous study we were able to predict not only the expected interception times but also the precise characteristics of response variability. Psychophysical experiments confirm the model’s predictions. Individual differences in subjects’ timing responses can be accounted for by individual differences in the perceptual priors on visual speed. Taken together, our behavioral and model results show that biases in perception percolate downstream and cause action biases that are fully predictable. Motivation ========== Bayesian models of perceptual inference explain many perceptual biases [@Knill07; @alan2006; @hedges2011]. By leveraging prior knowledge, Bayesian inference provides a principled and optimal strategy for an observer to infer the state of a perceptual variable from observed noisy evidence. Perceptual biases arise in this context because prior beliefs about the environmental statistics influence the perceptual process. Well documented examples include biases cue combination [@Knill07], low contrast biases in perceived speed [@alan2006], and biases towards the cardinals in orientation perception [@Odelia06; @Girshick11], to name a few. A separate line of work suggests that many visual illusions fail to lead to equivalent biases in motor behavior [@GooddaleREV]. For example, the Ponzo illusion, where lines of the same length are perceived differently due to receding distance, does not translate into differences in motor commands for grasping these lines [@Ganel08]. From a Bayesian perspective, the disconnect between perception and action appears puzzling. If a perceptual illusion can be explained in the context of optimal Bayesian inference, then we would expect the illusion to translate into biased behavior because the Bayesian framework is grounded in the statistics of the physical world [@Giesler03]. Furthermore, why allocate costly resources for optimal perception if these perceptions are not used in guiding actions? Speed perception has been extensively studied within the Bayesian framework [@alan2006] and provides an entry point for dissecting the influence of Bayesian perception upon action. Speed perception is traditionally studied with a two alternative forced choice task (2AFC), where subjects must choose the faster of two motion stimuli. Under this setup, low contrast stimuli are perceived to move slower relative to high contrast stimuli of the same speed [@Thompson81]. A prior distribution favoring slower speeds combined with a wider likelihood width for low contrast qualitatively explains the slow speed illusion and provides a tight fit to 2AFC data [@alan2006; @hedges2011]. ![[Interception task: task structure and generative model.]{} (A) (top) Subjects fixate on the small red dot on the bottom of the screen while viewing a drifting motion stimulus [@alan2006] for 1s at an eccentricity of $6^{\circ}$. (middle) At the moment the motion stimulus disappears, the fixation dot turns white to alert the subjects that the stimulus has been released. The stimulus moves, occluded, at a constant speed across the screen. (bottom) Subject makes a button press to indicate when the stimulus would have reach the center of the target area. For medium contrast conditions, feedback was provided indicating the end position of the stimulus at the time of the bottom press. Position was shown in green if the stimulus is within the target area and in red otherwise. (B) The generative model for perception shown as a directed graph. The values of the latent physical variables, stimulus velocity, $v$, and length, $l$, are combined using a simple physics model (deterministic) to define a time, $t$, for the stimulus to reach the target area. Observed variables represent the sensory measurements for speed, $m_v$, and length, $m_l$. The width of the noise distribution for $m_v$ is contrast dependent and changes based on the contrast from trial to trial. (C) The generative model for the motor control assumes that the time of the key press (action), $t_p$, is drawn from a Gaussian distribution with mean equal to the perceptual time estimate, $\hat{t}$. Solid arrows indicate stochastic dependencies whereas dotted lines indicate deterministic dependencies.[]{data-label="fig:task"}](figure1){width="\linewidth"} We developed a sensory-motor interception task in order to assess whether or not these perceptual speed biases translate to biases in timing actions. In this interception task subjects briefly view a moving grating stimuli contained in a circular window (Figure \[fig:task\]). The stimuli then disappears at which point it moves occluded with constant velocity across the screen. Subjects are instructed to make a button press when they believe the stimuli has crossed to the other side of the screen and coincides with some target a fixed distance away from stimulus presentation. Importantly, the moving stimuli are of differing contrasts to induce the slow speed illusion. We also present different speeds to obtain timing data for multiple speed contrast pairs. To formalize a clear hypothesis, we developed a Bayesian model for this task that assumes that a subject’s timing response is optimal in order to best most accurately intercept the object in the target area. Such an optimal model necessarily predicts that the perceptual biases (because they are optimal) translate to the corresponding motor biases. In what follows we first derive the Bayesian model for our the described interception task. We then explain how we perform the task with human subjects and present the measured behavioral results. Finally we compare the data to predictions of our model based on prior and likelihood parameters obtained from psychophysical experiments in a previous study of speed perception. Interception model ================== We assume each subject observes both the speed of the moving stimulus and the distance between the stimulus and the target (see Fig. \[fig:task\]A). These separate perceptual measurements must be integrated and decoded in some sensible way in order to determine an adequate time estimate. We formalize perception, intuitive physics, and action in our task using generative models and then derive an optimal time estimator based on assumptions of rationality, the independence assumptions of our perceptual model, and the motor noise in timing actions. Generative model for perception ------------------------------- We model each trial independently, and in what follows we avoid trial subscripts for convenience. On each trial the stimulus speed, $v$, and length between stimulus center and target, $l$, are each drawn independently from their respective natural prior distributions, $p(v)$ and $p(l)$. The time, $t$, the stimulus needs to reach the center of the target is dependent only on the physical speed and length and is thus drawn from the conditional distribution, $p(t\|v,l)$. Speed, length, and time are the latent variables that we combine in a vector ${\bf h} = \{v, l, t\}$. The observable variables ${\bf m} = \{m_v, m_l\}$, are the noisy sensory measurements of the true speed and length and are each drawn independently from their conditional distributions, $p(m_v\|v)$ and $p(m_l\|l)$. Taken together, these conditional and prior distributions define a generative model over the observable variables, ${\bf m}$, and latent physical variables, ${\bf h}$. The joint distribution for each trial becomes: $$p({\bf m}, {\bf h}) = p(v)p(l)p(m_v\|v)p(m_l\|l)p(t\|v,l)$$ and is displayed graphically in Figure \[fig:task\]. The observer has access to only the measurements and must infer the physical state. The posterior distribution of the physical variables given the measurements is: $$p({\bf h}\|{\bf m}) = \frac{p(m_v\|v)p(v)p(m_l\|l)p(l)p(t\|v,l)}{p(m_v)p(m_l)} \nonumber \\ = p(v\|m_v) p(l\|m_l)p(t\|v,l)$$ where the marginal $p(m_v,m_l)$ = $p(m_v)p(m_l)$ because the physical sources of the measurements are independent. Furthermore, we see that the marginal distribution for latent physical variables is separable into independent posterior distributions for speed and length and the conditional distribution for time. Physics model ------------- We embed the physical relationship between time, velocity and length in the conditional distribution $p(t\|v,l)$. If we assume that length and velocity are constant, rather than fluctuating due to some unknown latent forces, the conditional density becomes deterministic: $$p(t\|v,l) = \delta(t,q(v,l))$$ where $\delta$ is the Dirac delta measure that assigns all of its probability mass to the second argument. The function $q(v,l)$ is the physics model which maps speed and length to time. Because we assume constant speed and length, the physics model is the simple Newtonian relationship: $q(v,l) = \frac{l}{v}$. Action model ------------ The timing decision, $\hat{t}({\bf m})$, is a deterministic function of both length and speed measurements. We assume the subject has only indirect control over its motor commands, in the sense that the produced motor output, $t_p$, stochastically depends on the perceptual estimate $\hat{t}({{\bf m}})$ via the distribution $p(t_p\|\hat{t}({{\bf m}}))$ (Figure \[fig:task\]). We model $p(t_p\|\hat{t}({{\bf m}}))$ as a Gaussian distribution with mean $\mu_p = \hat{t}({{\bf m}})$ and standard deviation that grows proportionally with the mean, $\sigma_p = w_p \mu_p$ where $w_p$ is the time production Weber fraction [@JazayeriNat]. Estimator of travel time ------------------------ We model timing action choice in our task using decision theory [@JuliaCOG], which allows us to identify an optimal estimator from a class of functions, $f({\bf m})$, that maps the set of measurements directly onto a timing decision. We assume subjects operate rationally and try to accrue as much task dependent reward as possible. Thus our optimal estimator, $\hat{t}({\bf m})$, is one which minimizes the posterior loss between the timing actions, $t_p$, that depend stochastically on the timing estimate, and the latent physical variables: $$\hat{t} ({\bf m}) = \underset{f({\bf m})}{\operatorname{argmin}}\int \mathcal{L}({\bf h},f({\bf m})) p({\bf h} \| {\bf m}) \, d {\bf h}$$ where $\mathcal{L}$ is the the expected penalty for making the timing decision $f({\bf m})$ in physical state ${\bf h}$ and $p({\bf h}\|\bf{m})$ is the posterior distribution of our generative model. Of course, in our model the timing actions depend stochastically on the timing decision so $\mathcal{L}$ is determined by integrating over all possible timing actions: $$\mathcal{L}({\bf h}, f(m)) = \int L({\bf h},t_p) p(t_p\|f(m)) \, d t_p$$ where $L({\bf h},t_p)$ is the exact loss for performing action $t_p$ in latent state ${\bf h}$ and $p(t_p\|f({\bf m}))$ is the probability of performing motor command $t_p$ under time decision $f({\bf m})$, as defined above. While interception actions are performed in time, actions are chosen in reference to the distance moved by the object. We thus decompose the loss function, $L$, into two separate functions that measure loss along different dimensions: $$L({\bf h}, t_p)= L_L({\bf h}, t_p) + \eta L_T({\bf h}, t_p)$$ Where $L_L$ measures loss in distance, $L_T$ measures loss in time, and $\eta$ parameterizes the weight of loss in time with respect to loss in distance. A natural form for loss in distance is the squared error between the true length, $l$, and the distance traveled at time press $t_p$: $$L_L({\bf h}, t_p) = \big(l - v t_p \big)^2$$ where $v t_p$ is the distance traveled by the stimulus moving at speed $v$ at time $t_p$. This loss function captures the motivational thrust of our task: to land the moving stimulus as close to the target center as possible. We also assume that subjects prefer shorter timing actions to finish the experiment quickly. The loss function in time then takes a linear form: $$L_T({\bf h}, t_p) = t_p$$ Under these loss functions $\eta$ controls the trade off between task performance and completion time. With these model specifications, one can show that the optimal estimator that minimizes the posterior loss is given by: $$\hat{t}({\bf m}) = f_{opt} ({\bf m}) = \frac {E (l\|m_l) E (v\|m_v) - \frac{\eta}{2}}{(w_p^2 + 1) E (v^2\|m_v)}$$ where $E(l \|m_l)$ and $E(v\|m_v)$ are the expectations of the posterior distributions for $l$ and $v$ respectively and $w_p$ is the timing noise fraction. To derive this expression we plug the full loss function and posterior distribution into the expected posterior loss and then find the minima of the resulting expression. We first integrate through each term: $$\begin{aligned} Loss &= \iint \Big( l^2 - 2 l v t_p + v^2 t_p^2 + \eta t_p\Big) p(t_p\|f({\bf m})) d t_p p({\bf h}\|{\bf m}) \, d {\bf h} \\ &= \int l^2 p(l\|m_v) dl - 2 \iiint l v t_p p(t_p\|f({\bf m})) p(v\|m_v) p(l\|m_l) \, dv \, dl \, d t_p \\ & + \iint v^2 t_p^2 p(v\|m_v) p(t_p\|f({\bf m})) \, dv \, dt_p + \int \eta t_p p(t_p\|f({\bf m})) d t_p \\ &= E(l^2\|m_l) + \eta f({\bf m}) - 2 f({\bf m}) E(l\|m_l) E(v\|m_v) + E(v^2\|m_v)(w_p^2 f({\bf m})^2 + f({\bf m})^2) \end{aligned}$$ where we have used the fact that probability distributions integrate to one and that the first and second moments of a distribution are the mean and variance plus mean squared, respectively. We note that the final expression is quadratic in $f({\bf m})$ and thus take the derivative of the above equation with respect to $f({\bf m})$, set it equal to zero, and solve for $f({\bf m})$ to obtain the desired result. We now describe our experiment and behavioral results before returning to model specifications needed to predict behavior. Psychophysical experiment ========================= On each trial subjects are presented with a circular patch which contains a horizontal drifting grating. See Figure \[fig:task\] for example stimuli and [@alan2006] for stimulus specifics. The patch is $3^{\circ}$ in radius and is located $5.2^{\circ}$ to the right of the center of the screen. Subjects are told to fixate on a small red circle $3^{\circ}$ below the center of the screen. A large target ring, $4^{\circ}$ in radius, is located $5.2^{\circ}$ to the left of screen center with a small cross in its center. These distances were chosen to ensure that the eccentricity between fixation and stimulus are identical to previous work, $6^{\circ}$. After one second of motion, the grating disappears and the red fixation circle turns white to indicate that the stimuli trajectory has begun. Subjects make a button press when they believe the center of the stimuli coincides with the center of the target (Figure \[fig:task\]). Subjects first complete two training blocks with feedback at .3 contrast level to adapt to the task. The remaining blocks contain seven different speeds, equally tiled between 3 to 10 $\frac{deg}{s}$, at three contrast levels, high (.8), medium (.3), and low (.1). Feedback remains for medium contrast trials. Three male subjects ages, 23 to 40, completed the task. Subject 1 completed 48 trails for each contrast speed pair while Subjects 2 and 3 completed 90 trials. To assess the effects of contrast and speed on both the bias and noise in timing actions we first pool trials of the same contrast speed pair to obtain a sample distribution over timing actions, ${\bf t}_{v,c} = \{{t_{1:v,c},...,t_{n:v,c}\}}$, where $v$ and $c$ indicate the speed and contrast and $n$ is the number of trials for each condition. We then compute three sample statistics for each contrast speed pair: the mean, the variance, and the skewness. By comparing mean estimates we can determine the bias in timing actions due to contrast. The variance provides a measure of how the variability in motor behavior is affected by the perceptual noise due to contrast. Skewness, measured as the third standardized moment, is another statistic we apply to compare model and data characteristics. Finally, we calculate sample statistics for an average subject by pooling timing actions across subjects. We do this for both statistical power and for comparison to model predictions with average parameter values. For the average subject there is significant difference in timing actions ($p \ll .05$) between medium and low contrast and between medium and high contrast stimuli at all speed levels. Lower contrast stimuli lead to longer times than medium contrast and medium contrast leads to longer times than high contrast (Figure \[fig:avg\]). Individual subject data portrays a similar result (Figure \[fig:datsub\]). Subject 1 shows a significant difference between times for medium and low contrast at the slower speeds but not for the two fastest tests speeds. Inspection of the timing plots shows that there is a contraction in the bias between medium and low contrast at higher speeds. Subject 2 shows no contraction in bias and maintains significant difference between conditions at higher speeds. For all subjects the variance decreases as speed increases and is higher for lower contrast than medium and high contrast. These effects are reflected in the average subject plots (Figure \[fig:avg\]). For skewness, no clear trend between contrasts is discernible yet overall skewness increases as speed increases. ![ [Interception model qualitatively matches average subject data across timing distribution statistics.]{} (top) Combined timing data for all three subjects. For each contrast speed pair the sample mean, sample variance, and sample skewness were determined from the pooled data. (bottom) Model timing results using an average extracted speed prior distribution and likelihood widths (see [@alan2006]). Model and data both display a consistent contrast dependent timing bias, whereby lower contrast leads to longer times. Variability is also contrast dependent and the model predicts both the general shape and the range of variance values. The model reproduces the upward trend in skewness and the range of skewness values. Parameters: $s = .65$ $\eta = 6$.[]{data-label="fig:avg"}](figure2.pdf){width="\linewidth"} Model simulations ================= Our interception model is general and can be applied to a variety of tasks. In what follows we adapt this model to predict behavior in our psychophysical experiment. The model accounts for both the differences in timing bias at low, medium, and high contrasts and the individual subject differences in bias contraction at high speeds. As mentioned earlier, the prior and conditional distribution for speed have been previously fit to five different subjects [@alan2006]. We predict behavior for these subjects by plugging in the fit parameter values to our model. Overall we make behavioral predictions for two of the subjects from the Stocker and Simoncelli 2006 paper, referred to as Subject 1S and Subject 2S, and predictions for an average subject with average parameters taken across all five subjects. We then compare predicted behavior to human performance by computing the expected value, variance, and skewness of the simulated timing distributions. We first flesh out the specifics of the prior and conditional distribution for speed. Speed prior and conditional distribution ---------------------------------------- While the physical speed distribution across the human retina has yet to be measured, human prior distributions for speed have previously been extracted from subjects and show a power law shape [@alan2006]. We parameterize this speed prior by assuming the density in physical space is approximately constant for slow speeds, drops as a power law for medium to high speeds, and then transitions back to a constant regime at very high speeds (adapted from [@hedges2011]): $$p(v) \propto \frac {1} {(\|v\|^2 + b^2)^{d}} + r \vspace{-.0099in}$$where $b$ controls at what speed the density transitions from a constant function of speed to a power law, $d$ controls the rate of decay, and $r$ controls at what speed the prior transitions back to constant density. We assume this prior distribution is fixed in advance and is not modulated by the speeds viewed in our experiment. To obtain the parameters for our simulations we fit this parameterized form of the speed prior to the nonparametric prior distributions extracted for each of the five subjects. We now specify an appropriate speed representation for the observer. We choose a mapping function, $f(v)$, which maps the linear physical speed to a normalized logarithmic speed, $\tilde{v}$ which is in the same space as the perceptual measurement for speed, $m_v$. The speed transformation is: $ \tilde{v} = f(v) = \log(\frac{v}{v_o} + 1)$ where $v_o$ is a small normalization constant. Working in this space allows us to model the conditional speed observation function, $p(m_v\|v)$, as a Gaussian distribution with mean $\tilde{v}$. The standard deviation at each speed is separable into functions for the contrast of the stimuli, $c$, and speed: $\sigma(c,\tilde{v}) = g(c)h(\tilde{v})$. The speed and contrast likelihood parameters are either set to those previously obtained for each subject or linearly interpolated from these values. Our experiment uses the same distance for every trial so we assume that the subject posterior distribution over length can be approximated by a delta function, $\delta(l,l_o)$, where $l_o$ is the true distance between stimuli and target. The expectation over length becomes $l_o$. Motor adaptation and model output --------------------------------- Subjects were trained with feedback on stimuli of medium contrast. Due to the slow speed illusion, this leads them (and our model) to overestimate the time it takes for the stimuli to reach the target. After sufficient feedback, timing behavior for medium contrast shifts to faster, more correct, times. To model this motor adaptation on a fast time scale we introduce a scaling factor, $s$, that shrinks the time estimate appropriately to obtain a new corrected time estimate, $\hat{t}_c ({{\bf m}})$. Namely, $\hat{t}_c({{\bf m}}) = s \, \hat{t}({{\bf m}})$. Because this adaptation happens quickly it should only affect motor output and not the estimator, which we assume has been optimized by the observer over a lifetime of interacting with moving objects and making time judgments. For simulations the $s$ parameter is chosen such that the mean timing predictions at medium contrast overlap with the veridical time to contact. The timing noise fraction, $w_p$, was previously fit to subject timing behavior [@JazayeriNat] and we set it to the average of these maximum likelihood fits, $w_p$ = $.07$. In order to compare the timing distributions of the data to our model we computed the probability of making action $t_p$ under speed $v$ by marginalizing out the measurements: $$p(t_p\|v) = \int p(t_p\|\hat{t}_c(m_v)) p(m_v\|v) d m_v$$ where $p(t_p\|\hat{t}_c(m_v))$ and $p(m_v\|v)$ are as defined above. The expected value, variance, and skewness of the timing distribution are all calculated using numerical integration. ![[ Variance in previously extracted subject priors explains between subject variance in timing actions.]{} ([A.]{}) Mean timing actions for Subjects 1 and 2 as functions of speed for each contrast. Subject 2 maintains a roughly constant difference in bias between the contrast levels across speeds while Subject 1 shows a contraction in timing bias. ([B.]{}) The extracted parameterized speed prior distribution for both model subjects. The Subject 1S prior slope becomes shallow at high speeds while the Subject 2S prior maintains a constant slope. ([C.]{}) Mean timing model predictions for the same two subjects. Again, we see a contraction in timing bias for Subject 1S and not for Subject 2S which can be partially explained by differences in prior shape. Parameters: 1S $s = .9$ $\eta = 8$ 2S: $s = .8$ $\eta = 6.4$ []{data-label="fig:datsub"}](figure3.pdf){width="\linewidth"} Model results and comparison to data ------------------------------------ The predictions for the average subject show a clear contrast dependent timing bias qualitatively similar to results from the average human subject. Furthermore, both model and data show no contraction at higher speeds. The model variance predictions across contrasts show a steep initial decline followed by a leveling off at higher speeds. Lower contrasts induce more variability in timing response, as expected due to the wider width of the conditional distribution for speed measurements. This prediction is also qualitatively similar to our human data and within the same variance range. At slow speeds lower contrast stimuli predict smaller skewness while as the speeds increase we see a gradual cross over whereby at high speeds this pattern is reversed. While the raw human data is hard to interpret, both data and model show a shallow but consistent increase in skewness as speed increases. Predictions for individual subjects reveals key variation in model behavior that is lost when we look at predictions for an average subject. Timing predictions for Subject 1S show a strong contraction in the timing bias at higher speeds, quite similar to the experimental data from Subjects 1 and 3. Predictions for Subject 2S shows no such contraction, and appears qualitatively similar to Subject 2 from our experiment. The difference in subject timing behavior at high speeds can be accounted for by differences in the shape of the prior distribution. Stocker and Simoncelli [@alan2006] show that under some conditions the speed bias is proportional to the slope of the prior distribution in the normalized logarithmic space. The slope of the prior for Subject 1S flattens out at faster speeds while the Subject 2S prior maintains a constant slope across speeds (Figure \[fig:datsub\]). We see that this difference in prior is reflected by a contraction in timing bias for Subject 1S and no contraction in bias for Subject 2S. This suggests that the qualitative differences in human timing response at high speeds can be partially explained by differences in the prior shape. Discussion ========== Experimentally, we show that the contrast speed visual illusions is reflected in timing actions. To support our conclusion we constructed an interception model grounded in a detailed Bayesian model for speed and showed that our model produces similar biases in timing actions. We argued that differences in timing biases across subjects can be partially explained by differences in the shape of the prior distribution, further strengthening the link between Bayesian perception on the one hand and actions on the other. The qualitative link between the model variance and subject variance and the similar upward trend in skewness shows that most of the behavioral noise is explained by the noise parameters in our model: the motor noise that grows proportionally with timing magnitude and the perceptual noise due to contrast. These results challenge the notion that visual illusions are not reflected in action [@GooddaleREV] and demands a more nuanced view. Our results suggest that visual illusions that can be explained in terms of Bayesian inference will similarly affect action. Our modeling framework for the interception task draws inspiration from recent Bayesian models of human intuitive physics [@Battaglia11; @Sanborn09]. These models combine uncertainty over physical variables, based on perception and prior knowledge, with a deterministic Newtonian physics model to infer the most likely behavior of the physical world. We expand on this work by investigating a simple speed timing task in detail and thus show that biases in one component of a physics model can percolate through to affect inference about other physical variables, such as time. In future work we intend to teach subjects novel physical relationships and see if speed biases continue to influence judgments about other physical variables. Over all, we see this work as part of an ongoing effort to probe the degree to which Bayesian models generalize across different behavioral tasks. Our results indicate that the Bayesian model for speed perception generalizes to novel task domains that are different from the one it was originally conceived and that the speed prior transfers across these domains. Acknowledgments {#acknowledgments .unnumbered} --------------- We thank all the subjects who participated in the experiments. This work was made possible by the Office of Naval Research (grant N00014-11-1-0744).	{ "pile_set_name": "ArXiv" }
--- abstract: \| The classical moduli space $\mc$ of a supersymmetric gauge theory with trivial superpotential can be stratified according to the unbroken gauge subgroup at different vacua. We apply known results about this stratification to obtain the $W \neq 0$ theory classical moduli space $\mc ^W \subset \mc$, working entirely with the composite gauge invariant operators $\hp$ that span $\mc$, assuming we do not know their elementary matter chiral field content. In this construction, the patterns of gauge symmetry breaking of the $W \neq 0$ zero theory are determined, Higgs flows in these theories show important differences from the $W=0$ case. The methods here introduced provide an alternative way to construct tree level superpotentials that lift all classical flat directions leaving a candidate theory for dynamical supersymmetry breaking, and are also useful to identify heavy composite fields to integrate out from effective superpotentials when the elementary field content of the composites is unknown. We also show how to recognize the massless singlets after Higgs mechanism at a vacuum $\hp \in \mc^W$ among the moduli $\delta \hp$ using the stratification of $\mc$, and establish conditions under which the space of non singlet massless fields after Higgs mechanism (unseen as moduli $\delta \hp$) is null. A small set of theories with so called “unstable” representations of the complexified gauge group is shown to exhibit unexpected properties regarding the dimension of their moduli space, and the presence of non singlet massless fields after Higgs mechanism at all of their vacua. address: \| FaMAF, Universidad Nacional de Córdoba,\ Ciudad Universitaria, 5000, Córdoba, ARGENTINA author: - Gustavo Dotti date: June 2000 title: 'Higgs Mechanism and Luna Strata in ${\cal N}=1$ Gauge Theories' --- = 6in [C]{} \#1[[Lie]{}( \#1 )]{} \#1[ \^[-1]{}( ( \#1 ) ) ]{} u v \#1[[S]{}\_[\#1]{}]{} §\#1[\_[\#1]{}]{} \#1[[T]{}\_[\#1]{}]{} \#1[[N]{}\_[\#1]{}]{} \#1[[N]{}\_[\#1]{}]{} \#1\#2[[- \#2pt]{}]{} Introduction {#intro} ============ The construction of the classical moduli space $\mc$ of a supersymmetric gauge theory with trivial superpotential is well known [@plb; @gatto; @gatto2; @luty]: starting from the elementary chiral matter fields $\phi \in \c^n$, a basic set $\hp^i(\phi), i=1,...,s$ of holomorphic gauge invariant composites is obtained. Generically, the basic invariants are constrained, there are polynomials $p_{\a}(\hp)$ such that $p_{\a}(\hp(\phi))$ vanishes identically. The classical moduli space $\mc$, defined to be the set of $\df$ points mod the gauge group action, can be shown to be parameterized by the subset of $\c^s$ defined by the constraints among the invariants, $\mc = \{ \hp \in \c^s \| p_{\a}(\hp)=0 \}$ [@plb; @luty]. It is worth recalling at this point that $\mc$ agrees with the [quantum]{} moduli space of the theory if the Dynkin index of the gauge group representation on the elementary field space is greater than the index of the adjoint representation [@sc]. $\mc$ also has a geometrical interpretation [@plb; @gatto]: if $G^c$ is the complexification of the gauge group $G$, then $G^c$ is non-compact and some of the $G^c$ orbits in $\phi$ space $\c^n$ are not closed. $\mc$ is shown to parameterize the set of [closed]{} $G^c$ orbits, denoted $\c^n//G$ to distinguish it from orbit space $\c^n/G$. The relation $\mc = \c^n//G$ is due to the fact that there is precisely one $G$ orbit of points per closed $G^c$ orbit, and no points in non-closed $G^c$ orbits [@plb; @gatto].\ Now suppose we add a tree level superpotential $W(\phi)$. To ensure gauge invariance, we must have $W(\phi)=\hw(\hp(\phi))$, where $\hw: \c^s \to \c$ is an arbitrary function on the basic invariants (the distinction of the superpotential $\hw$ as a function of the basic invariants from the superpotential $W$ as a function of the elementary fields is crucial in what follows.) The classical moduli space $\mc ^W \subset \mc$ of the theory with the added superpotential is the image under $\pi: \phi \to \hp(\phi)$ of the set $dW=0$ of $F-$flat points in $\c^n$. In [@luty] it is shown that $\mc^W \subset \mc \subseteq \c^s$ can be obtained by adding to the algebraic constraints $p_{\a}(\hp)=0$ among the invariants the gauge invariant constraints resulting from $dW=0$. A natural question to ask is the following: suppose we are given $\mc$ (i.e., the number $s$ of basic invariants and the constraints $p_{\a}: \c^s \to \c$) and $\hw(\hp)$, but we [do not know]{} the elementary field composition $\hp(\phi)$ of the basic invariants (in particular, we do not know $W(\phi)=\hw(\hp(\phi))$). Is it possible to construct $\mc^W$ from this information? This would give us what we may call a “low energy description” of $\mc^W$, since only the composite fields are involved in the construction. At first sight, we may think that knowledge of the constraints linking the basic invariants $\hp$, the ones that define $\mc$, is enough. For example, if $\hw = m \hp^1$ is a mass term and we know the constraints linking $\hp ^1$ to the other composite superfields $\hp $, we may think we should be able to deduce which composite superfields are made heavy by $\hw$. Unfortunately, this is not the case, a “low energy" description is not possible unless further input is given. The following is probably the simplest example illustrating this fact: consider an $SO(N)$ theory with two flavors of vector fields, $\{ \phi \} = \{Q_i^{\a}, \a=1,...,N, i=1,2\}\simeq \c^{2N}$. The basic invariants are $\{ \hp \} = \{ S_{ij} \equiv Q^{\a}_i Q^{\a}_j \}$, and $\mc = \{ S_{ij} \} \simeq \c^3$, as there are no constraints among the invariants. Although the directions $S_{11}, S_{12}$ and $S_{22}$ in $\c^3$ are completely equivalent, $\mc^W=\{ (0,0,0) \}$ if $\hw = mS_{12}$, whereas $\mc^W = \{(S_{11},0,0)\} \simeq \c^1$ if $\hw =mS_{22}$. The example shows that knowledge of the invariants $\hp$, their constraints, and $\hw(\hp)$ is not enough to obtain $\mc^W$, an extra piece of information is required. The zero superpotential moduli space $\mc$ can be stratified according to the conjugate class $(H)$ of the unbroken gauge subgroup $H \subseteq G$ at each vacuum. The stratum $\S {(H)} \subset \mc$ contains all vacua with unbroken gauge subgroup conjugate to $H$. It turns out that the stratification $\mc = \cup_{(H)} \Sigma_{(H)}$ is precisely the extra piece of information required to accomplish the desired low energy description. The relation between the stratification of $\mc$ and the low energy construction of $\mc^W$ comes from the equality $\mc^W \cap \Sigma _{(H)} = \{ \hp \in \Sigma _{(H)} \| d \hw_{(H)} (\hp) = 0 \}$, $\hw_{(H)}$ being the restriction ${\hw}_{\|_{\S {(H)}}}$ of $\hw$ to $\Sigma_{(H)}$. $\mc^W \subset \mc$ can be constructed in steps by finding the stationary points of the restriction of $\hw$ to $\Sigma _{(H)}$, one stratum at a time. This useful fact, pointed out in [@plb], follows from results of Luna [@luna], Abud and Sartori [@as], Procesi and Schwarz [@plb; @schwarz1]. In this paper we elaborate further on these results and obtain an algorithm to construct $\mc^W$ which, in some cases, saves us the job of looking for critical points in every stratum, but only on some carefully chosen ones. These techniques are applied to recognize heavy composites (of unknown elementary field content) to integrate out from an effective superpotential $W_{eff}(\hp)$ [@out; @susy]. They are also used to construct tree level superpotentials $\hw$ that lift all non-trivial flat directions, reducing the classical moduli space to a point. In all cases the input is the stratification of $\mc$, where the calculations are performed, the composition $\hp(\phi)$ of the basic invariants in terms of the elementary fields is not required. Theories lifting all non trivial flat directions are interesting as candidates for dynamical supersymmetry breaking [@dsb]. We finally use the results in [@plb] to investigate the relationship (in the classical theory) between the massless modes $\delta \phi$ at a vacuum $\phi$ in unitary gauge, and the moduli $\delta \hp$ obtained by linearizing at $\hp(\phi)$ the constraints among the $\hp$’s. The expected isomorphism between these two sets holds (in most theories) only at the so called principal stratum $\Sigma_{(G_P)}$, where the gauge group $G$ is maximally broken. Yet, some exceptional theories are found for which the isomorphism does not hold even a the principal stratum. This is the same set of theories for which the equation $\text{dim } \mc = \text{ dim } \{\phi\} - (\text{ dim}_{\Bbb R}\, G - \text{ dim}_{\Bbb R}\, G_P)$ does not hold[^1], they are characterized by the fact that the bulk of the configuration space $\{ \phi \} \simeq \c^n$ is filled with non closed orbits of the complexification $G^c$, case in which the $G^c$ action on $\phi$ space is termed “unstable". Since the $G$ representation on $\c^n$ must be anomaly free, most anomaly free representations are real, and real representations are stable, unstable theories are rare.\ The paper is organized as follows. In Section \[ls\] we introduce the stratification of $\mc$ and an order relation between strata. The important results of Luna, Procesi and Schwarz are integrated in Theorem I in Section \[lt\], examples are given in Section \[exs\]. In Section \[app\] we apply Theorem I to a number of problems. The low energy construction of $\mc^W$, is treated in Section \[lec\], in Section \[comp\] we show the usefulness of breaking $\mc^W$ up into its irreducible components, and study the patterns of gauge symmetry breaking in $W \neq 0$ theories, the problem of identifying heavy composites, and that of constructing superpotentials that lift all non trivial vacua. In Section \[mf\] we study the relation between massless fields after Higgs mechanism (MFHM) at a vacuum $\hp_0 \in \mc^W$ and the space of moduli tangent to $\mc^W$ at $\hp_0$. A number of examples is given, many of them were constructed to illustrate the subtleties involved in the given results. Section \[conc\] contains the conclusions. We defer to Appendix \[appen\] some technical aspects in the derivation of the results in Section \[app\]. Luna’s Stratification of the Moduli Space {#ls} ========================================= Let $\{\phi\}\simeq \c^n$ be the set of matter chiral fields of a supersymmetric gauge theory with gauge group $G$ and zero superpotential, $\hp^i(\phi), i=1,...,s$ a basic set of holomorphic $G$ invariant operators, $p_{\a}(\hp)=0, \a=1,...,l$ the algebraic constraints among the basic invariants. The moduli space of the theory is $\mc = \{ \hp \in \c^s \| p_{\a}(\hp)=0 \}$. This means that for every $\hp_0$ satisfying $p_{\a}(\hp_0)=0$ there is precisely one $G$ orbit $G \phi_0$ of points satisfying $\hp(\phi_0)=\hp_0$. Note that $G \phi$ denotes the $G$ orbit through $\phi$, whereas $G_{\phi}$ denotes the unbroken gauge subgroup at $\phi$. Since points in the same $G$ orbit have conjugate little groups, $G_{g\phi} = gG_{\phi}g^{-1} \forall g \in G$, a conjugate class $G_{\hp_0}$ can be associated to $\hp_0 \in \mc$, namely, $(G_{\hp_0}) \equiv (G_{\phi_0})$, where $\phi_0$ is any point satisfying $\hp(\phi_0)=\hp_0$. The definition makes sense since any two points $\phi_0, \phi_1$ satisfying $\hp(\phi_0)=\hp_0=\hp(\phi_1)$ are $G$ related. A stratum $\S {(H)}$ is the set of $\hp$’s in $\mc$ satisfying $(G_{\hp})=(H)$, $\mc = \cup_{(H)}\S {(H)}$ is the disjoint union of its strata. The strata are complex manifolds of different dimensions, $\mc$ instead is an [algebraic set]{} [@clo], the zero set of a family of polynomials. The tangent space at a point $x \in X$, $X$ an algebraic set or a complex manifold, is denoted $T_xX$. For an algebraic set $X = \{ x \in \c^s \| p_{\a}(x)=0, \a=1,...,l\}$, $T_xX$ is defined to be the kernel of the matrix $\partial p_{\a}/ \partial x^i(x)$, i.e., the $\delta x'$s allowed by the linearized constraints.[^2] Generically, the dimension of the tangent space of an algebraic set may change from point to point. If $X$ is an algebraic set satisfying dim $T_xX=n \; \forall x \in X$, then $X$ is a complex manifold of dimension $n$ [@kodaira]. The projection map $\pi: \phi \to \hp(\phi)$ sends $\c^n$ onto $\mc$. Its differential at $\phi$, $\pi'_{\phi}: T_{\phi}\c^n \simeq \c^n \to T_{\pi(\phi)}\mc$ relates the $\delta \phi$ at $\phi$ with the moduli $\delta \hp$ at $\hp$, $\pi'_{\phi}: \delta \hp \to \partial \hp^i(\phi) / \partial \phi^j \delta \phi^j$. An order relation can be introduced in the set of isotropy classes, we say that $(H_1) < (H_2)$ if $H_1$ is conjugate to a subgroup of $H_2$. This order relation is [partial]{}, it is not true that given any two classes $(H_1) \neq (H_2)$ either $(H_1) < (H_2)$ or $(H_1) > (H_2)$, there are unrelated classes. The partial order relation among conjugate classes induces a partial order relation among the strata: $\S {(H_1)} > \S {(H_2)}$ whenever $(H_1) < (H_2)$. A theorem on the stratification of the moduli space {#lt} --------------------------------------------------- The important results in [@plb; @im] are the following (see also [@luna; @as; @schwarz1]):\ [Theorem I:]{} - There are only finitely many strata of $\mc$. The strata are complex manifolds, their closures are algebraic subsets of $\mc$. - The closure of $\S {(H)}$ is $$\label{closure} \overline{\S {(H)}} = \bigcup_{(L) \geq (H)} \S {(L)},$$ i.e., the boundary of $\S {(H)}$ is the union of the strata that are strictly smaller than $\S {(H)}$. - There is a unique minimal isotropy class $(G_P)$, called [principal isotropy class]{}, $\S {(G_P)}$ is called principal stratum. $(G)$ is a unique maximal isotropy class. - Assume $\phi$ is and let $\t {\phi} \equiv \lie {G^c} {\phi} \simeq T_{\phi}G^c \phi$, the tangent at $\phi$ of the $G^c$ orbit through $\phi$. $\t {\phi} \subset \c^n$ is a $G_{\phi}$ invariant subspace , and it has a $G_{\phi}$ invariant complement ${\t {\phi}}^{\perp}$. The theory with gauge group $G_{\phi}$ and matter content ${\t {\phi}}^{\perp}$ is called [slice representation]{}. The stratification of the moduli space of the slice representation contains precisely the $(H) \leq (G_{\phi})$ classes of the original theory. - Let $\s {\phi} \subseteq {\t {\phi}}^{\perp}$ be the subspace of $G_{\phi}$ singlets, $\n {\phi}$ a $G_{\phi}$ invariant complement of $\s {\phi}$ in ${\t {\phi}}^{\perp}$, then $\c^n = \t{\phi} \oplus \s {\phi} \oplus \n {\phi}$. The differential $\pi'_{\phi}$ of the projection map $\pi$ at $\phi$ has kernel $\t {\phi} \oplus \n {\phi}$, its rank is $T_{\pi(\phi)} \S {(G_{\phi})}$, the tangent to the stratum through $\pi(\phi)$. - Assume the point $\phi$ satisfies $\pi(\phi) \in \S {(G_P)}$. Then $\n {\phi} = \{0\}$ if and only if the $G^c$ representation on $\c^n$ is stable. If the representation is unstable, the theory with gauge group $G_P$ and matter content $\n {\phi}$ (i.e., the slice theory without the singlets) has no holomorphic $G_P$ invariants. Some explanations are in order. Regarding point (c) note that in a partially ordered set $U$ there may be more than one maximal element. Generically, there is a subset $M \subset U$ of maximal elements. Any two elements in $M$ are unrelated under $<$, whereas $m>p$ for all $m \in M, p \in U \setminus M$. Analogously, there is a subset of minimal elements of $U$ . Regarding point (d) note that the “slice representation" is just the supersymmetric gauge theory obtained by Higgs mechanism at energies below the masses of the broken generators. An interesting observation in [@im] is that $G_{\phi}$ determines entirely the slice representation, i.e., there cannot be two different points leading to theories with the same (class of) $G$ subgroup as gauge group but having different matter content. This is a consequence of the following identity of direct sums of $G_{\phi}$ representations ($\rho$ stands for the $G$ representation on $\{ \phi \} = \c^n$, whereas $\rho_{\|_{H}}$ means its restriction to the $G$ subgroup $H$.) $$\s {\phi} \oplus \n {\phi} \oplus (Ad \; G)_{\|_{G_{\phi}}} = \ \rho_{\|_{G_{\phi}}} \oplus Ad \; G_{\phi},$$ Theorem I.c-d guarantees that [any]{} pattern of symmetry breaking from $G$ to subsequently smaller $G$ subgroups lead to the theory with maximally broken gauge subgroup $G_P$. According to Theorem I.f this theory contains only $G_P$ singlets, except in those cases where $\rho$ is unstable. As explained above, the complexification $G^c$ of the gauge group is non-compact, and some of its orbits are not closed. $\rho$ is said to be unstable if there is a $G^c$ invariant subset of $\c^n$, open in the Zariski topology, containing only non-closed $G^c$ orbits. The Zariski topology on $\c^n$ [@clo] is the one whose closed sets are algebraic sets, i.e., zeroes of a family of polynomials, it is coarser than the usual $\c^n \simeq {\Bbb R}^{2n}$ topology. This topology is useful in studying representations of algebraic groups, of which the complexification $G^c$ of the compact Lie group $G$ is an example. Zariski open subsets of a vector space $\c^n$ are (Zariski) dense, we may therefore view unstable theories as those for which the bulk of the elementary field space $\c^n$ is filled with non-closed $G^c$ orbits, i.e., orbits without points. It was shown in [@schwarz1] that if the $G$ representation $\rho$ on $\c^n$ is real then it is stable. As physical theories must be free of gauge anomalies, and most anomaly free representations are real, unstable supersymmetric gauge theories are rare. In fact, the only unstable theories based on a simple gauge group are $SU(2N+1)$ with ${\raisebox{-3.5pt}{{\hbox{\rule{0.4pt}{6.5pt}\hskip-0.4pt\rule{6.5pt}{0.4pt}\hskip-6.5pt\rule[6.5pt]{6.5pt}{0.4pt}}\rule[6.5pt]{0.4pt}{0.4pt}\hskip-0.4pt\rule{0.4pt}{6.5pt}}}\hskip-6.9pt \raisebox{3pt}{{\hbox{\rule{0.4pt}{6.5pt}\hskip-0.4pt\rule{6.5pt}{0.4pt}\hskip-6.5pt\rule[6.5pt]{6.5pt}{0.4pt}}\rule[6.5pt]{0.4pt}{0.4pt}\hskip-0.4pt\rule{0.4pt}{6.5pt}}}}+ (2N-3) \overline{{\raisebox{-.5pt}{{\hbox{\rule{0.4pt}{6.5pt}\hskip-0.4pt\rule{6.5pt}{0.4pt}\hskip-6.5pt\rule[6.5pt]{6.5pt}{0.4pt}}\rule[6.5pt]{0.4pt}{0.4pt}\hskip-0.4pt\rule{0.4pt}{6.5pt}}}}}, N \geq 2,$ and $SO(10)$ with a spinor. These theories exhibit some curious properties, as we will see.\ Note from (b-c) that $\mc = \overline{\S {(G_P)}}$, this leads to the definition dim $\mc=$ dim $\S {(G_P)}$ (in agreement with the standard definition of dimension of an irreducible algebraic set [@clo]). The dimension of an algebraic set may change from point to point, generically there are [singular points]{} $\hp \in \mc$ at which dim $T_{\hp} \mc > \text{dim }\mc$, they belong to smaller strata. As stressed in [@luty], however, it is not true that all vacua $\hp$ satisfying $(G_{\hp})> (G_P)$ are singular, a trivial counterexample being offered by those theories with unconstrained basic invariants, for which all points of $\mc \simeq \c^s$ are non-singular, including those with enhanced gauge symmetry.\ From Theorem I we can show that $$\label{int} \S {(H')} \cap \overline{\S {(H)}} \neq \emptyset \Rightarrow \S {(H')} \leq \S {(H)} \;\; \text{ (equivalently } \S {(H')} \subseteq \overline{\S {(H)}}).$$ This is proved by taking $\phi \in \S {(H')} \cap \overline{\S {(H)}}$, then $(G_{\phi}) = (H')$ and also, using Theorem I.b, $(G_{\phi}) \geq (H)$, from where equation (\[int\]) follows. Another straightforward consequence of the theorem is that, for stable actions (only!), $\text{dim } \mc = n - \text{ dim } G^c + \text{dim } {G_P}^c $. This is proved by picking a point $\phi$ satisfying $\pi(\phi) \in \S {G_P}$. We have the following (in)equalities from (b,e) of Theorem I:[^3] $\text{dim } \mc \equiv \text{dim } \S {(G_P)} = \text{rank } \pi'_{\phi} = n - \text{ dim ker } \pi'_{\phi} = n - \text{ dim } \t {\phi} - \text {dim } \n {\phi} = n -(\text { dim } G^c - \text{ dim } {G_P}^c) - \text {dim } \n {\phi} \leq n -(\text { dim } G^c - \text{ dim } {G_P}^c)$. According to Theorem I.f, equality holds only if $\rho$ is stable. For unstable theories the dimension of $\mc$ is smaller than the expected value $ n - \text{ dim } G + \text{dim}_{\Bbb R} G_P$, this is consistent with the statement above that “the bulk of $\phi$ space" (a Zariski dense subset) contains no point. Unstable theories [do]{} have $G^c$ orbits of dimension equal to $n - \text{dim }\mc > \text{dim } G^c - \text{dim } {G_P}^c$ [@ela], however, there is no point in these highest dimensional orbits. In other words, unstable theories are characterized by the impossibility of breaking $G^c$ to the smallest isotropy $G^c$ subgroup by a point. Examples {#exs} -------- In the following, we will arrange partially ordered sets $U$ in columns in this way: the first column (from left to right) contains the subset $C_ 1 \subset U$ of maximal elements in $U$, the second column contains the subset $C_2$ of maximal elements in $U \setminus C_1$, the third column $C_3$ contains the maximal elements in $U \setminus (C_1 \cup C_2)$, and so on. We will also draw a line linking the elements in adjacent columns which are related under $<$. Note that, by construction, any element in $C_{i+1}$ is smaller than at least one element in $C_i$. Note also from Theorem I.c that if $U$ is the set of strata $\S {(H)}$ or conjugate classes $(H)$, then the first and last column contain a single element. For totally ordered sets there is a single entry per column.\ Our first example is a theory with a smooth moduli space $\mc \simeq \c^s$ and totally ordered strata.\ [Example \[exs\].1:]{} Consider $F$ flavor, $N$ color SQCD with quarks $Q^{\a}_i$ and antiquarks $\q_{\b}^j$, $\a,\b = 1,...,N$; $i,j = 1...F$, $F<N$. The basic invariants are $M^j_i = \q ^j_{\a} Q^{\a}_i$, they are unconstrained and so $\c^{F^2} \simeq \mc = {\Bbb M}^F $, the set of $F \times F$ complex matrices. The classical global non-R symmetries are $K = U(F)_Q \times U(F)_{\q}$. A generic point can be $G \times K$ rotated onto $$\label{dfqcd} Q^{\a}_i = {({\q}^{\dagger})}_{\beta}^j = \left( \begin{array}{cc} V & 0 \\ 0 & 0 \end{array} \right), \hspace{1cm} V = \left( \begin{array}{ccccc} v_1 & 0 & 0 & \cdots & 0\\ 0 & v_2 & 0 & \cdots &0\\ \vdots & \vdots & & & \vdots\\ 0 & 0 & \cdots & v_{r-1} & 0 \\ 0 & 0 & \cdots & 0 & v_r \\ \end{array} \right), \;\;\ v_i \neq 0, \;\; r \leq F.$$ As isotropy $G$ subgroups are $K$ invariant and $G$ conjugate we only need consider the points eq.(\[dfqcd\]) to obtain Luna’s stratification of $\mc$. The unbroken $G$ subgroup at $(Q,\q)$ of equation (\[dfqcd\]) is $SU(N-r)$ ($SU(1)$ meaning the trivial group). There are $F+1$ strata, $\S {SU(N-r)}$, $r=0,1,...,F$, and there is a [complete]{} order relation $\S {SU(N)} < \S {SU(N-1)}~<~\cdots~< \S {SU(N-F)}$, then we arrange the strata as $$\S {SU(N-F)} - \S {SU(N-F-1)} - \cdots - \S {SU(N-1)} - \S {SU(N)}.$$ From (\[dfqcd\]) follows that $\S {SU(N-r)}$ is the set of $K$ orbits of points $M = \text{diag} (\|v_1\|^2,...,\|v_{r}\|^2,0,...,0),\; \|v_i\| \neq 0$, which is the set ${\Bbb M}^{F}_r$ of rank $r$ complex $F \times F$ matrices. The determinantal variety [@harris] ${\Bbb M}^F_{\leq r}$ of $F \times F$ matrices of rank less than or equal to $r$ is the algebraic set $$\label{detva} {\Bbb M}^F_{\leq r} = \{ M \in {\Bbb M}^F \| M^{[j_1}_{i_1}M^{j_2}_{i_2} \cdots M^{j_{r+1} ]}_{i_{r+1}} = 0 \}.$$ As ${\Bbb M}^F_r = {\Bbb M}^F_{\leq r} \setminus {\Bbb M}^F_{\leq r-1}$, equation (\[detva\]) defines the smallest Zariski closed (i.e., algebraic) set containing ${\Bbb M}^F_{r}$, i.e., ${\Bbb M}^F_{\leq r}=\overline{{\Bbb M}^F_r}$. This verifies Theorem I.b: $\overline{\S {SU(N-r)}} = \cup_{j \leq r} \S {SU(N-j)}$. It is instructive to see what the tangent space $T_M {\Bbb M}^F_{\leq r}$ is (for an alternative derivation see [@harris]). As the equations defining ${\Bbb M}^F_{\leq r}$ in (\[detva\]) satisfy the requirement in footnote 2, the tangent space at $M$ of ${\Bbb M}^F_{\leq r}$ is obtained by linearizing (\[detva\]), $$\label{tanqcd} T_M {\Bbb M}^F_{\leq r} = \{ \delta M \in {\Bbb M}^{F} : M^{[j_1}_{i_1}M^{j_2}_{i_2} \cdots M^{j_r}_{i_r} \delta M^{j_{r+1} ]}_{i_{r+1}} = 0 \}.$$ To understand the condition eq. (\[tanqcd\]) contract $M^{[j_1}_{i_1}M^{j_2}_{i_2} \cdots M^{j_r}_{i_r} \delta M^{j_{r+1} ]}_{i_{r+1}} = 0 $ with $r+1$ linearly independent vectors $t^{i_k}_{(k)}, k=1,...,r+1$. If rank $M < r$ at least two of the $t$ vectors belong to ker $M$, (\[tanqcd\]) is trivially satisfied for any matrix $\delta M$, $T_M {\Bbb M}^F_{\leq r} \simeq {\Bbb M}^F$, dim $T_M {\Bbb M}^F_{\leq r}=F^2$. If rank $M = r$ we get a nontrivial condition if we choose the $t_{(j)}$ such that only one of them, say $t_{(r+1)}$, belongs to the kernel of $M$. The condition is $M^{[j_1}_{i_1}M^{j_2}_{i_2} \cdots M^{j_r}_{i_r} \delta M^{j_{r+1} ]}_{i_{r+1}} t_{(1)}^{i_1} \cdots t^{i_{r+1}}_{(r+1)} = 0$, meaning that $\delta M$ must send the kernel of $M$ onto the rank of $M$, the dimension of the tangent space at $M$, the space of allowed $\delta M$’s, being $F^2-(F-r)^2$. We conclude that $\S {SU(N-r)} = {\Bbb M}_{r}^F$ is the subset of non singular points of ${\Bbb M}_{\leq r}^F = \overline{\S {SU(N-r)}}$, the dimension of the complex manifold $\S {SU(N-r)} = {\Bbb M}_{r}^F$ being $F^2 - (F-r)^2$.\ The complexification of $G$ is $SU(N)^c=SL(N,\c)$, and $T \in \lie {SL(N,\c)}$ can be written as $$\lie {SL(N,\c)} \ni T = \left(\begin{array}{c\|c} t_1 & t_2 \\ \hline t_3 & t_4 \end{array} \right), \;\;\;\; t_4 \in \lie {GL(N-r,\c)}, \tr {t_1} + \tr {t_4} = 0.$$ The (Lie algebra of the) isotropy group ${G^c}_{(Q,\q)}={G_{(Q,\q)}}^c$ of (\[dfqcd\]) is obtained by setting $t_1=t_2=t_3=0$, $t_4 \in SL(n,\c)$. We also split $Q$ and $\q$ as $$Q^{\a}_i = \left(\begin{array}{c\|c} q_1 & q_2 \\ \hline q_3 & q_4 \end{array} \right), \;\;\;\; \q^j_{\a} = \left(\begin{array}{c\|c} \tilde{q}_1 & \tilde{q}_2 \\ \hline \tilde{q}_3 & \tilde{q}_4 \end{array} \right),$$ where $q_1$ and $\tilde{q}_1$ are $r \times r$ blocks. The tangent space to the $G^c$ orbit of (\[dfqcd\]) is obtained by acting with $\lie {SL(n,\c)}$ on $(Q,\q)$ $$\t {(Q,\q)}: \delta Q ^{\a}_i = \left( \begin{array}{c\|c} t_1 V & 0 \\ \hline t_3 V & 0 \end{array} \right), \;\;\;\; \delta \q ^j_{\b} = \left( \begin{array}{c\|c} - V^{\dagger} t_1 & - V^{\dagger} t_2 \\ \hline 0 & 0\end{array} \right).$$ An $SU(N-r)$ invariant complement is given by $\n {(Q,\q)} \oplus \s {(Q,\q)}$, where $$\begin{aligned} \n {(Q,\q)} &:& \delta Q^{\a}_i = \left(\begin{array}{c\|c} 0 & 0 \\ \hline 0 & \delta q_4 \end{array} \right), \;\;\;\; \delta \q^j_{\a} = \left(\begin{array}{c\|c} 0 & 0 \\ \hline 0 & \delta \tilde{q}_4 \end{array} \right). \\ \s {(Q,\q)} &:& \delta Q^{\a}_i = \left(\begin{array}{c\|c} 0 & \delta q_2 \\ \hline 0 & 0 \end{array} \right), \;\;\;\; \delta \q^j_{\a} = \left(\begin{array}{c\|c} \delta \tilde{q}_1 & 0 \\ \hline \delta \tilde{q}_3 & 0 \end{array} \right).\end{aligned}$$ The slice representation at $(\ref{dfqcd})$ is $\n {(Q,\q)} \oplus \s {(Q,\q)}$, the $SU(N-r)$ theory with $(F-r) ({\raisebox{-.5pt}{{\hbox{\rule{0.4pt}{6.5pt}\hskip-0.4pt\rule{6.5pt}{0.4pt}\hskip-6.5pt\rule[6.5pt]{6.5pt}{0.4pt}}\rule[6.5pt]{0.4pt}{0.4pt}\hskip-0.4pt\rule{0.4pt}{6.5pt}}}}+ \overline{{\raisebox{-.5pt}{{\hbox{\rule{0.4pt}{6.5pt}\hskip-0.4pt\rule{6.5pt}{0.4pt}\hskip-6.5pt\rule[6.5pt]{6.5pt}{0.4pt}}\rule[6.5pt]{0.4pt}{0.4pt}\hskip-0.4pt\rule{0.4pt}{6.5pt}}}}}) + (2Fr-r^2) {\bf 1}$, as is well known. The configuration point $(Q,\q)$ of eq. (\[dfqcd\]) is sent by $\pi$ to the following point of $\mc = {\Bbb M}^{F}$: $$M = \pi(Q,\q) = \left( \begin{array}{cc} V^{\dagger} V & 0 \\0 & 0\end{array} \right).$$ It is easily verified that $\pi'_{(Q, \q)}$ annihilates $\t {(Q,\q)} \oplus \n {(Q,\q)}$, whereas $$\label{rankqcd} \text{rank} \; \pi'_{(Q,\q)} = \pi'_{(Q,\q)} (\s {SU(N-r)}) = \biggl\{ \delta M^i_j \in {\Bbb M}^{F} : \delta M^i_j = \left( \begin{array}{c\|c} \delta \tilde{q}_1 V & V^{\dagger} \delta q_2 \\ \hline \delta \tilde{q}_3 V & 0 \end{array} \right) \biggr\}.$$ As $V$ is invertible, (\[rankqcd\]) agrees with the set of matrices sending ker $M$ onto rank $M$, which is the tangent space $T_M {\Bbb M}_r^F$ at $M$ of the stratum through $M$. This verifies Theorem I.e.\ The moduli space ${\cal M}$ of the following example contains singular points. Its strata are totally ordered, and $\S {(G_P)}$ equals the set of non-singular points of $\mc$, a property that is not generic. [Example \[exs\].2:]{} Consider $F=N$ SQCD. points can be $G \times K$ rotated onto $Q^{\a}_i = \text{diag } \;(q_1,...,q_N)$, $\q^{j}_{\b} = \text{diag } \;(\tilde{q}_1,...,\tilde{q}_N)$ subject to $$\label{cond} \|q_i\|^2 - \|\tilde{q}_i\|^2 = c, \;\; \text{ independent of } i.$$ The invariants are $M^j_i = Q_i^{\a} \q ^j_{\a}, B = \text{det }\; Q$, and $\tilde{B} = \text{ det } \q$, they satisfy $$\label{f=n} \text{det } M - B \tilde{B} = 0.$$ If $B = \prod_i q_i \neq 0$ or $\tilde{B} = \prod \tilde{q}_i \neq 0$, $G$ is completely broken. If some of the $q$’s are zero, then the same set of $\tilde{q}$’s must be zero, otherwise we get both $c > 0$ and $c < 0$ in equation (\[cond\]) . Let $r$ be the number of zero $q$’s. If $r=1$, $SU(N)$ is completely broken, rank $M = N-1$, and $B=\tilde{B}=0$. If $r > 1$, $SU(N)$ is broken to $SU(r)$, rank $M = N-r$, and $B=\tilde{B}=0$. We conclude that the principal stratum is $\S {e} = \{ (M,B,\tilde{B}) \| B \neq 0, \text { or } \tilde{B} \neq 0, \text{ or cofactor } M \neq 0 \}$. The other strata are $\S {SU(r)} = \{ (M,B,\tilde{B}) \| B=\tilde{B}=0 \text{ and rank } M = N-r \}, r > 1$. By linearizing eq (\[f=n\]) we see that $\S {e}$ agrees with the set of non singular points of $\mc$. The $N-1$ strata are completely ordered: $$\S {e} - \S{SU(2)} - \cdots - \S { SU(N)}.$$ We now present examples where the set of strata is only partially ordered.\ [Example \[exs\].3:]{} Consider $G=SU(N)$ with an ($SL(N,\c)$) adjoint field $A^{\a}_{\b}$. The basic invariants are $t_j=\tr A^{j+1}, j=1,...,N-1$, they are unconstrained and so $\mc = \c^{N-1}$. The $D-$flatness condition is $\tr T [A,A^{\dagger}]=0, \forall T \in SU(N)$, then $[A,A^{\dagger}] \propto {\Bbb I}$, and so $[A,A^{\dagger}] = 0$. This implies that $A$ can be $G$ rotated onto a diagonal complex matrix. The residual gauge symmetry, the group of permutations of the diagonal entries, can be used to bring $A^{\a}_{\b}$ to the following form: $$\label{dfadj} A = \text{diag} \; (\stackrel{m_1}{\overbrace{v_1,v_1,...,v_1}}, \stackrel{m_2}{\overbrace{v_2,v_2,...,v_2}}, \cdots , \stackrel{m_j}{\overbrace{v_j,v_j,...,v_j}}),$$ where $$\label{dfadj1} m_1 \geq m_2 \geq \cdots m_j \geq 1,\;\; \sum_{k=1}^j m_k = N, \;\; \text{and}\;\; \sum_{k=1}^j m_k v_k = 0.$$ The configuration point above breaks $SU(N)$ to $S(U(m_1) \times U(m_2) \times \cdots \times U(m_{j-1}) \times U(m_j))$, (block diagonal matrices of the form diag$(g_1,...,g_j)$, $g_k \in U(m_k)$ and $\prod_{i=1}^j \text{det} \, g_i = 1$). In some particular cases this is a direct product group, for example, if $m_j=1$ then $S(U(m_1) \times U(m_2) \times \cdots \times U(m_{j-1}) \times U(m_j)) = U(m_1) \times U(m_2) \times \cdots \times U(m_{j-1})$. The isotropy groups are in one to one correspondence with the partitions ${\cal P}$ of $N$, a partition being a decomposition $N = m_1 + m_2 + \cdots m_j$ where $m_1 \geq m_2 \geq \cdots \geq m_j \geq 1$. The partial order in the set of isotropy groups induces the following partial order relation in the set of partitions of $N$: ${\cal P}_1$ is smaller than ${\cal P}_2$ if ${\cal P}_2$ is obtained from ${\cal P}_1$ by summing some of its terms and ordering the resulting terms. We give some $N=5$ examples: $2+1+1+1=2+(1+1+1)=3+2$, then $2+1+1+1 < 3+2$, also $3+2=(3+2)=5$ then $3+2 < 5$; finally, $3+1$ and $2+2$ are unrelated. It is easy to see that the partitions of $N$ (and therefore the isotropy groups and strata of the $SU(N)$ theory with an adjoint) are totally ordered if $N=2,3$, but only partially ordered if $N\geq 4$. There is exactly one point of the form (\[dfadj\]-\[dfadj1\]) in a $G$ orbit of points, this implies that $\{v_1,...,v_{j-1}\}$ can be taken as a set of local coordinates of $\S {S(U(m_1) \times \cdots U(m_j))}$. In particular, $\S {S(U(m_1) \times \cdots U(m_j))}$ has (complex) dimension $j-1$. Starting $N=4$ we have distinct strata of the same dimension. According to Theorem I.b, two such strata must be unrelated under $<$, as none of them can lie in the boundary of the other one. Write $$A^{\a}_{\b} = \left( \begin{array}{cccc} t_{11} & t_{12} & \cdots & t_{1j} \\ t_{21} & t_{22} & \cdots & t_{2j} \\ \vdots & \vdots & & \vdots \\ t_{j1} & t_{j2} & \cdots & t_{jj} \end{array} \right),$$ $t_{ik}$ is an $m_i \times m_k$ matrix, $\sum_k \tr t_{kk} = 0$. The tangent space at $(\ref{dfadj})$ breaks up into $$\begin{aligned} \t {A} &=& \{ \delta A \| \delta t_{kk} = 0, \; k=1,...,j \} \\ \label{asing} \s {A} &=& \{ \delta A \| \delta t_{ij} = \delta_{ij} a_i {\Bbb I}_{m_i \times m_i}, \; \sum_{i=1}^j m_i a_i = 0 \} \\ \n {A}&=& \{ \delta A \| \delta t_{ij} = \delta_{ij} t_{ii}, \tr t_{kk}=0 \; \text{for} \; k=1,...,j \}\end{aligned}$$ It is readily verified that $\pi'_{A}$ annihilates $\t {A} \oplus \n {A}$. The easiest way to see that $\pi'_{A}$ sends $\s {A}$ isomorphically onto $T_{\pi(A)} \S {S(U(m_1) \times \cdots U(m_j))}$ is by noting that the linear coordinates $a_i$ of $\s {A}$ in (\[asing\]) correspond to variations $\delta v_i$ of the local coordinates $v_i$ of $\S {S(U(m_1) \times \cdots U(m_j))}$ in equation (\[dfadj\]). Theorem I.e is therefore verified in this case.\ We give more details for the special cases $N=3$ and $N=4$.\ $SU(3)$ [with an adjoint field:]{} The partitions of $N=3$ are completely ordered: $$3 > 2+1 > 1+1+1$$ Equivalently, we have the following ordered set of isotropy groups: $$SU(3) > U(2) > U(1) \times U(1)$$ leading to the arrangement $$\S {U(1) \times U(1)} - \S {U(2)} - \S {SU(3)}$$ of the strata, which have complex dimensions $2,1$ and $0$. The equations defining the strata of $\mc \simeq \c^2$ can be obtained by finding the relations among the invariants $t_j$ at points $A_H$ of the form (\[dfadj\]-\[dfadj1\]) with isotropy group $H$: $$\begin{aligned} \left(A^{\a}_{\b}\right)_{SU(3)} &=& 0; \;\;\; \left(A^{\a}_{\b}\right)_{U(2)} = \text{diag} (x,x,-2x), \; x \neq 0; \\ \nonumber \left(A^{\a}_{\b}\right)_{U(1) \times U(1)} &=& \text{diag} (x,y,-x-y), y \neq x, -2x, -x/2. \end{aligned}$$ For example, at $\left(A^{\a}_{\b}\right)_{U(2)}$ we have $t_1=6x^2, t_2=-6x^3, x \neq 0$, this defines the algebraic set $t_1^3-6t_2^2=0$ with the point $(0,0)$ removed. Proceeding in this way we arrive at $$\begin{aligned} \S {U(1) \times U(1)} &=&\{(t_1,t_2) \in \c^2 \| t_1^3-6t_2^2 \neq 0 \}, \\ \nonumber \S {U(2)} &=& \{(t_1,t_2) \in \c^2 \| t_1^3-6t_2^2 = 0 \; \text{and} \; (t_1,t_2) \neq (0,0) \}, \\ \nonumber \S {SU(3)} &=& \{(0,0)\}. \end{aligned}$$ $SU(4)$ [with an adjoint:]{} we have the following partitions of $4$: $$\begin{array}{ccccccc} & & 3+1 & & &&\\ & \diagup & & \diagdown & & & \\ 4 & & & & 2+1+1 & \; \; - \;\; & 1+1+1+1, \\ & \diagdown && \diagup &&&\\ && 2+2 &&&& \end{array}$$ corresponding to the following patterns of symmetry breaking $$\begin{array}{ccccccc} & & U(3) & & &&\\ & \diagup & & \diagdown & & & \\ SU(4) & & & & U(2) \times U(1) & \;\; - \;\; & U(1) \times U(1) \times U(1). \\ & \diagdown && \diagup &&&\\ && S(U(2) \times U(2)) &&&& \end{array}$$ Following branches from left to right be have two decreasing sequences of isotropy groups, or two increasing sequence of strata of dimensions $0,1,2$ and $3$. There is no order relation between the one dimensional $U(3)$ and $S(U(2) \times U(2))$ strata. Generic diagonal elements at different strata have the forms $$\begin{aligned} \label{stru4} &&\left(A^{\a}_{\b}\right)_{SU(4)} = 0; \\ \nonumber && \left(A^{\a}_{\b}\right)_{U(3)} = \text{diag} (x,x,x,-3x),\; x \neq 0; \\ \nonumber &&\left( A^{\a}_{\b}\right)_{S(U(2) \times U(2))} = \text{diag} (x,x,-x,-x), \; x \neq 0; \\ \nonumber && \left(A^{\a}_{\b}\right)_{U(2) \times U(1)} = \text{diag} (x,x,y,-2x-y), \; y \neq \pm x, -3x; \\ \nonumber &&\left(A^{\a}_{\b}\right)_{U(1) \times U(1) \times U(1)} = \text{diag} (x,y,z,-x-y-z), x,y,z \;\; \text{and} \;\; -x-y-z \;\; \text{all different}.\end{aligned}$$ From the above equations we get $t_1=2x^2+y^2+(2x+y)^2, t_2=2x^3+y^3-(2x+y)^3,$ and $t_3=2x^4+y^4+(2x+y)^4$ at $\S {U(2) \times U(1)}$. If $x$ and $y$ are unrestricted, these are parametric equations for $\overline{\S {U(2) \times U(1)}} \subset \c^3$. An equivalent implicit equation, obtained by using Gröebner basis [@clo], is $288t_3t_1^2+144t_3t_1t_2^2 -90t_3t_1^4-288t_3^3+9t_1^6-68t_2^2t_1^3-24t_2^4=0$. The equations defining the strata are $$\begin{aligned} \nonumber \S {U(1) \times U(1) \times U(1)} &=& \{(t_1,t_2,t_3) \| 288t_3t_1^2+144t_3t_1t_2^2 -90t_3t_1^4-288t_3^3+9t_1^6-68t_2^2t_1^3-24t_2^4 \neq 0 \} \\ \nonumber \S {U(2) \times U(1)} &=& \{(t_1,t_2,t_3) \| 288t_3t_1^2+144t_3t_1t_2^2 -90t_3t_1^4-288t_3^3+9t_1^6-68t_2^2t_1^3-24t_2^4 = 0 \\ \nonumber &&\;\; \text{ and } ( t_2 \neq 0 \; \text{ or } t_1^2-4t_3 \neq 0 ) \text { and } ( t_1^3-3t_2^2 \neq 0 \text{ or } \frac{7}{12}t_1^2-t_3 \neq 0 ) \} \\ \nonumber \S {S(U(2) \times U(2))} &=&\{(t_1,t_2,t_3) \| t_2 = 0, t_1^2-4t_3=0, \text { and } t_3 \neq 0 \}, \\ \nonumber \S {U(3)} &=& \{(t_1,t_2,t_3) \| t_1^3-3t_2^2 = 0, \frac{7}{12}t_1^2-t_3 = 0, \; \text{ and } t_3 \neq 0 \}, \\ \nonumber \S {SU(4)} &=& \{(0,0,0)\}. \end{aligned}$$ $\overline{\S {U(2) \times U(1)}}$ is a two dimensional complex surface on which the complex curves $\overline{\S {U(3)}}$ and $\overline{\S {S(U(2) \times U(1)}}$ lie. These two curves meet at $\S {SU(4)}$.\ Our final example is a theory with an unstable representation of the complexified gauge group.\ [Example \[exs\].4:]{} Let $G=SU(2N+1), \rho = \overline{{\raisebox{-3.5pt}{{\hbox{\rule{0.4pt}{6.5pt}\hskip-0.4pt\rule{6.5pt}{0.4pt}\hskip-6.5pt\rule[6.5pt]{6.5pt}{0.4pt}}\rule[6.5pt]{0.4pt}{0.4pt}\hskip-0.4pt\rule{0.4pt}{6.5pt}}}\hskip-6.9pt \raisebox{3pt}{{\hbox{\rule{0.4pt}{6.5pt}\hskip-0.4pt\rule{6.5pt}{0.4pt}\hskip-6.5pt\rule[6.5pt]{6.5pt}{0.4pt}}\rule[6.5pt]{0.4pt}{0.4pt}\hskip-0.4pt\rule{0.4pt}{6.5pt}}}}} + (2N-3) \overline{{\raisebox{-.5pt}{{\hbox{\rule{0.4pt}{6.5pt}\hskip-0.4pt\rule{6.5pt}{0.4pt}\hskip-6.5pt\rule[6.5pt]{6.5pt}{0.4pt}}\rule[6.5pt]{0.4pt}{0.4pt}\hskip-0.4pt\rule{0.4pt}{6.5pt}}}}}$, the classical flavor symmetry group is $K = U(1) \times U(2N-3)$. If $N=2$, the only point is the trivial one, and $\mc$ is a zero dimensional vector space. Actually, the $SU(5)$ with an antifundamental and an antisymmetric tensor, together with $SO(10)$ with a spinor, are the only theories based on a simple gauge group with only trivial points, and therefore a single stratum. If $N \geq 3$, $\mc$ is the vector space of $U(2N-3)$ unconstrained antisymmetric tensors $V^{ij} = A^{\a \b} Q^i_{\a} Q^j_{\b} = \pi(Q,A)$. The $D-$flatness condition reads tr $ [T(2A A^{\dagger} - Q^{\dagger} Q)] = 0$. A generic point can be $G \times K$ rotated to $$\begin{aligned} \label{dfunst} Q^i_{\a} &=& \left( \begin{array}{cc} q & 0 \\ 0 & 0 \end{array} \right), \;\;\; q = diag(q_1,q_2,...,q_{2k}), \\ \nonumber A^{\a \b} &=& \left( \begin{array}{cc} v & 0 \\0&0 \end{array} \right), \;\; v = \text{diag}(v_1 \sigma, v_2 \sigma, ..., v_k \sigma), \;\; k \leq N-2, \;\; \sigma = \left( \begin{array}{rr} 0 & 1 \\ -1 & 0 \end{array} \right), \end{aligned}$$ with $\|q_{2j-1}\| = \|q_{2j}\| = \|v_j\| \neq 0$. This point breaks $G$ to $SU(2(N-k)+1)$, the set of strata $\S {SU(2(N-k)+1)}, k=0,...,N-1$ being totally ordered. Under $\pi$, (\[dfunst\]) goes to $$\label{p} V^{ij} = \text{diag} ( q_1q_2v_1 \sigma, q_3q_4v_2 \sigma,...,q_{2k-1}q_{2k} v_k \sigma,0,0,...,0).$$ The $K$ orbits of the points (\[p\]) generate the $SU(2(N-k)+1)$ stratum. $\S {SU(2(N-k)+1)}$ is the $4kN-2k^2-7k$ dimensional complex manifold of $(2N-3) \times (2N-3)$ antisymmetric matrices $V^{ij}$ of rank $2k$.\ Under $SU(2(N-k)+1)$, the configuration space $\c ^{(2N+1)(3N - 3)} \simeq T_{(A,Q)} \c ^{(2N+1)(3N - 3)} $ breaks into $\t {(A,Q)} \oplus \s {(A,Q)} \oplus \n {(A,Q)}$. Using (\[dfunst\]) and writing a $\lie {G^c}$ element as $$T = \left( \begin{array}{cc} t_1 & t_2 \\ t_3 & t_4 \end{array} \right), \hspace{1cm} t_4 \in \lie {SL(2(N-k)+1)},$$ we obtain $$\t {(A,Q)}: \delta Q = \left( \begin{array}{cc} -q t_1 & -q t_2 \\ 0 & 0 \end{array} \right), \hspace{1cm} \delta A = \left( \begin{array}{cc} t_1v + vt_1^T & v t_3^T \\ t_3 v & 0 \end{array} \right).$$ A possible choice for $\n {(A,Q)} \oplus \s {(A,Q)}$ is $$\begin{aligned} \label{sliceunst} \s {(A,Q)}: \;\;\; \delta Q = \left( \begin{array}{cc} 0&0 \\ \delta q_3 & 0 \end{array} \right), \hspace{1cm} \delta A = \left( \begin{array}{cc} \delta A_1 & 0 \\ 0 & 0 \end{array} \right) \\ \nonumber \n {(A,Q)}: \;\;\; \delta Q = \left( \begin{array}{cc} 0 & 0 \\ 0 & \delta q_4 \end{array} \right), \hspace{1cm} \delta A = \left( \begin{array}{cc} 0 & 0 \\ 0 & \delta A_4 \end{array} \right)\end{aligned}$$ The special feature of this example is that the $G^c$ action is [unstable]{}. Although $G^c$ applied to (\[dfunst\]) with $k=N-1$ gives a highest dimensional $G^c$ orbit containing points, there are $G^c$ orbits of higher dimension. An example of a highest dimensional orbit is that of the configuration point $$\begin{aligned} \label{nco} Q^i_{\a} &=& \left( \begin{array}{ccc} {\bf 0}_{(2N-3) \times 3}& q & {\bf 0}_{(2N-3) \times 1} \end{array} \right), \;\;\; q = diag(q_1,q_2,...,q_{2N-3}), \\ \nonumber A^{\a \b} &=& \text{diag}(v_1 \sigma, v_2 \sigma, ..., v_N \sigma, 0), \hspace{1cm} \sigma = \left( \begin{array}{rr} 0 & 1 \\ -1 & 0 \end{array} \right).\end{aligned}$$ The $G^c$ isotropy group at (\[nco\]), ${G^c}_0$, is different from ${G_0}^c$, a common situation for the $G$ and $G^c$ isotropy groups at $G^c$ orbits of non points. We can readily check that $\lie {{G^c}_0}$ is the set of $T \in {\frak sl}(2N+1,\c)$ having the form $$\label{nrunst} T = \left( \begin{array}{rrrrrrrr} x & y & 0 & a & 0 & \cdots & 0 & d \\ z & -x & 0 & b&0& \cdots & 0&e\\ \frac{-v_2}{v_1}b &\frac{v_2}{v_1}a&0&c&0&\cdots&0&f \\ 0&0&0&0&0&\cdots&0&0 \\ \vdots & \vdots & \vdots & \vdots & \vdots & & \vdots& \vdots \\ 0&0&0&0&0&\cdots&0&0 \end{array} \right).$$ $x,y$ and $z$ span an ${\frak sl}(2,\c)$ non-invariant Lie subalgebra of the isotropy subalgebra, whereas $a,b,c,d,e,f$ span a six dimensional unipotent (a Lie algebra of nilpotent matrices) Lie algebra ${\frak u}_6$ which is an ideal of $\lie {{G^c}_0}$. In other words $$\lie{{G^c}_0} = {\frak sl}(2,\c) \oplus {\frak u}_6 \; \text{(direct sum of vector spaces)}, \;\; [\lie{{G^c}_0}, {\frak{u}}_6] \subseteq {\frak u}_6.$$ After exponentiating we get a semidirect product: $G^c_0 = SL(2,\c) \ltimes {\frak U}_6$.\ The slice representation (\[sliceunst\]) at the point eq.(\[dfunst\]) is $SU(2(N-k)+1)$ with $[2(N-k)-3] \overline{{\raisebox{-.5pt}{{\hbox{\rule{0.4pt}{6.5pt}\hskip-0.4pt\rule{6.5pt}{0.4pt}\hskip-6.5pt\rule[6.5pt]{6.5pt}{0.4pt}}\rule[6.5pt]{0.4pt}{0.4pt}\hskip-0.4pt\rule{0.4pt}{6.5pt}}}}} + {\raisebox{-3.5pt}{{\hbox{\rule{0.4pt}{6.5pt}\hskip-0.4pt\rule{6.5pt}{0.4pt}\hskip-6.5pt\rule[6.5pt]{6.5pt}{0.4pt}}\rule[6.5pt]{0.4pt}{0.4pt}\hskip-0.4pt\rule{0.4pt}{6.5pt}}}\hskip-6.9pt \raisebox{3pt}{{\hbox{\rule{0.4pt}{6.5pt}\hskip-0.4pt\rule{6.5pt}{0.4pt}\hskip-6.5pt\rule[6.5pt]{6.5pt}{0.4pt}}\rule[6.5pt]{0.4pt}{0.4pt}\hskip-0.4pt\rule{0.4pt}{6.5pt}}}}+ (4kN-2k^2-7k) {\Bbb I}$. At the main stratum, $k=N-2$, the slice is $SU(5)$ with $ \overline{{\raisebox{-.5pt}{{\hbox{\rule{0.4pt}{6.5pt}\hskip-0.4pt\rule{6.5pt}{0.4pt}\hskip-6.5pt\rule[6.5pt]{6.5pt}{0.4pt}}\rule[6.5pt]{0.4pt}{0.4pt}\hskip-0.4pt\rule{0.4pt}{6.5pt}}}}} + {\raisebox{-3.5pt}{{\hbox{\rule{0.4pt}{6.5pt}\hskip-0.4pt\rule{6.5pt}{0.4pt}\hskip-6.5pt\rule[6.5pt]{6.5pt}{0.4pt}}\rule[6.5pt]{0.4pt}{0.4pt}\hskip-0.4pt\rule{0.4pt}{6.5pt}}}\hskip-6.9pt \raisebox{3pt}{{\hbox{\rule{0.4pt}{6.5pt}\hskip-0.4pt\rule{6.5pt}{0.4pt}\hskip-6.5pt\rule[6.5pt]{6.5pt}{0.4pt}}\rule[6.5pt]{0.4pt}{0.4pt}\hskip-0.4pt\rule{0.4pt}{6.5pt}}}}+ (2N-3)(N-2) {\Bbb I}$. Taking out the singlets we get $SU(5)$ with $ \overline{{\raisebox{-.5pt}{{\hbox{\rule{0.4pt}{6.5pt}\hskip-0.4pt\rule{6.5pt}{0.4pt}\hskip-6.5pt\rule[6.5pt]{6.5pt}{0.4pt}}\rule[6.5pt]{0.4pt}{0.4pt}\hskip-0.4pt\rule{0.4pt}{6.5pt}}}}} + {\raisebox{-3.5pt}{{\hbox{\rule{0.4pt}{6.5pt}\hskip-0.4pt\rule{6.5pt}{0.4pt}\hskip-6.5pt\rule[6.5pt]{6.5pt}{0.4pt}}\rule[6.5pt]{0.4pt}{0.4pt}\hskip-0.4pt\rule{0.4pt}{6.5pt}}}\hskip-6.9pt \raisebox{3pt}{{\hbox{\rule{0.4pt}{6.5pt}\hskip-0.4pt\rule{6.5pt}{0.4pt}\hskip-6.5pt\rule[6.5pt]{6.5pt}{0.4pt}}\rule[6.5pt]{0.4pt}{0.4pt}\hskip-0.4pt\rule{0.4pt}{6.5pt}}}}$, a theory with a zero dimensional moduli space, theorem I.f is verified. To show that $SU(5)$ with $ \overline{{\raisebox{-.5pt}{{\hbox{\rule{0.4pt}{6.5pt}\hskip-0.4pt\rule{6.5pt}{0.4pt}\hskip-6.5pt\rule[6.5pt]{6.5pt}{0.4pt}}\rule[6.5pt]{0.4pt}{0.4pt}\hskip-0.4pt\rule{0.4pt}{6.5pt}}}}} + {\raisebox{-3.5pt}{{\hbox{\rule{0.4pt}{6.5pt}\hskip-0.4pt\rule{6.5pt}{0.4pt}\hskip-6.5pt\rule[6.5pt]{6.5pt}{0.4pt}}\rule[6.5pt]{0.4pt}{0.4pt}\hskip-0.4pt\rule{0.4pt}{6.5pt}}}\hskip-6.9pt \raisebox{3pt}{{\hbox{\rule{0.4pt}{6.5pt}\hskip-0.4pt\rule{6.5pt}{0.4pt}\hskip-6.5pt\rule[6.5pt]{6.5pt}{0.4pt}}\rule[6.5pt]{0.4pt}{0.4pt}\hskip-0.4pt\rule{0.4pt}{6.5pt}}}}$ has a zero dimensional moduli space we specialize the above equations to the $N=2$ case. The orbit of (\[nco\]) has dimension $15$, as its isotropy group (\[nrunst\]) has dimension $9$. Taking the closure of this orbit we a get a fifteen dimensional algebraic subset of $\c^{15} \simeq \overline{{\raisebox{-.5pt}{{\hbox{\rule{0.4pt}{6.5pt}\hskip-0.4pt\rule{6.5pt}{0.4pt}\hskip-6.5pt\rule[6.5pt]{6.5pt}{0.4pt}}\rule[6.5pt]{0.4pt}{0.4pt}\hskip-0.4pt\rule{0.4pt}{6.5pt}}}}} + {\raisebox{-3.5pt}{{\hbox{\rule{0.4pt}{6.5pt}\hskip-0.4pt\rule{6.5pt}{0.4pt}\hskip-6.5pt\rule[6.5pt]{6.5pt}{0.4pt}}\rule[6.5pt]{0.4pt}{0.4pt}\hskip-0.4pt\rule{0.4pt}{6.5pt}}}\hskip-6.9pt \raisebox{3pt}{{\hbox{\rule{0.4pt}{6.5pt}\hskip-0.4pt\rule{6.5pt}{0.4pt}\hskip-6.5pt\rule[6.5pt]{6.5pt}{0.4pt}}\rule[6.5pt]{0.4pt}{0.4pt}\hskip-0.4pt\rule{0.4pt}{6.5pt}}}}$, the only possibility being the whole $ \overline{{\raisebox{-.5pt}{{\hbox{\rule{0.4pt}{6.5pt}\hskip-0.4pt\rule{6.5pt}{0.4pt}\hskip-6.5pt\rule[6.5pt]{6.5pt}{0.4pt}}\rule[6.5pt]{0.4pt}{0.4pt}\hskip-0.4pt\rule{0.4pt}{6.5pt}}}}} + {\raisebox{-3.5pt}{{\hbox{\rule{0.4pt}{6.5pt}\hskip-0.4pt\rule{6.5pt}{0.4pt}\hskip-6.5pt\rule[6.5pt]{6.5pt}{0.4pt}}\rule[6.5pt]{0.4pt}{0.4pt}\hskip-0.4pt\rule{0.4pt}{6.5pt}}}\hskip-6.9pt \raisebox{3pt}{{\hbox{\rule{0.4pt}{6.5pt}\hskip-0.4pt\rule{6.5pt}{0.4pt}\hskip-6.5pt\rule[6.5pt]{6.5pt}{0.4pt}}\rule[6.5pt]{0.4pt}{0.4pt}\hskip-0.4pt\rule{0.4pt}{6.5pt}}}}= \{ \phi \}$ vector space. If $\hp(\phi)$ is a holomorphic invariant, then $\hp(\phi)$ is constant in the closure of this orbit, i.e., the only holomorphic invariants of this theory are the constants, $\mc$ is a zero dimensional vector space. Applications {#app} ============ Low energy construction of $\mc^W$ and Lagrange multipliers {#lec} ----------------------------------------------------------- A holomorphic $G$ invariant superpotential $W: \c^n \to \c$ can always be written in terms of a basic set of holomorphic invariants $\hp^i(\phi), i=1,...,s$, as $W(\phi) = \hw (\hp(\phi))$, $\hw$ being an arbitrary $\c^s \to \c$ function. The $W=0$ classical moduli space $\mc$ is parameterized by the subset of $\c^s$ defined by the algebraic constraints $p_{\a}(\hp)=0, \a=1,...,l$ among the basic invariants $\hp(\phi)$. The moduli space $\mc^W$ of the supersymmetric gauge theory with the added superpotential is usually obtained by first solving for the $F$-flat point set $\c^n_W = \{ \phi \in \c^n \| dW(\phi)=0 \}$, then projecting $\c^n_W$ down to $\c^s$ using the map $\pi: \phi \to \hp(\phi)$, i.e., $\mc^W = \pi(\c^n_W)$. It can be shown [@luty] that $\mc^W \subset \mc \subseteq \c^s$ is the the algebraic set defined by the gauge invariant polynomial constraints $p_{\a}(\hp)=0, \a=1,...,l; w_{\b}(\hp)=0, \b=1,...,r$, where $w_{\b}(\hp)=0, \b=1,...,r$ are the gauge invariant constraints resulting from $dW=0$ [@luty]. In this section we elaborate further on the results in [@plb] on methods to obtain from $\hw$ and $p_{\a}(\hp)=0$ the equations $w_{\b}(\hp)=0$ defining ${\cal M}^W \subset \mc \subseteq \c^s$, working entirely in the space $\c^s$ of composite superfields $\hp$, assuming we do not know the functions $\hp(\phi)$, i.e., how the composite superfields are made out of the elementary fields. In Section \[intro\] we used an $SO(N)$ theory with two ${\raisebox{-.5pt}{{\hbox{\rule{0.4pt}{6.5pt}\hskip-0.4pt\rule{6.5pt}{0.4pt}\hskip-6.5pt\rule[6.5pt]{6.5pt}{0.4pt}}\rule[6.5pt]{0.4pt}{0.4pt}\hskip-0.4pt\rule{0.4pt}{6.5pt}}}}$ to show that knowledge of $\hw$ and the constraints among the basic invariants is not enough to obtain $\mc^W$, and claimed that the required additional information was the stratification of the moduli space. This last assertion follows from Theorem I: the differential at the point $\phi$ of the map $\pi: \phi \to \hp(\phi)$, $\pi'_{\phi} = \partial \hp^j(\phi) / \partial \phi^i$, annihilates the subspace $\t {\phi} \oplus \n {\phi}$ of $\c^n = \t {\phi} \oplus \n {\phi} \oplus \s {\phi}$, (Theorem I.e) and so $$\label{w'} \frac{\partial W}{\partial \phi^i} \delta \phi^i = \frac{\partial \hw}{\partial \hp ^j} \left( \frac{\partial \hp ^j}{\partial \phi^i} \delta \phi^i \right), \;\;\;\; \left( \frac{\partial \hp ^j}{\partial \phi^i} = \pi' \right),$$ is zero if $\delta \phi \in \t {\phi} \oplus \n {\phi}$. On the other hand, again by Theorem I.e, $\partial \hp^i(\phi) / \partial \phi^j \, \delta \phi^j$ does not span the whole tangent space $T_{\hp(\phi)}\mc$ of $\mc$ at $\hp(\phi)$, but only the subspace $T_{\hp(\phi)} \S {(G_{\phi})} \subseteq T_{\hp(\phi)} \mc$ tangent to the stratum through $\hp(\phi)$. Therefore, $dW=0$ is equivalent to $$\label{dw} \frac{\partial \hw}{\partial \hp^i} {\Bigg\vert_{\hp(\phi)}} \delta \hp^i = 0, \;\;\; \forall \delta \hp^i \in T_{\hp(\phi)} \S {G_{\phi}},$$ In other words, $dW(\phi)=0$ if and only if $\pi(\phi)$ is a stationary point of the restriction $\hw_{(G_{\phi})} \equiv \hw\|_{\S {(G_{\phi})}}$ of $\hw$ to the stratum passing through $\pi(\phi)$. This fact, pointed out in [@plb] gives an answer to the problem of finding $\mc^W$ working entirely with gauge invariant operators: first find, for each stratum $\S {(H)}$, the critical points of the restriction of $\hw$ to $\S {(H)}$, then take the union of the resulting sets. We will see in the following section that it is not always necessary to solve the stationary point equations at [every]{} stratum. There are two ways of finding the stationary points of $\hw_{(H)} \equiv \hw_{\|_{\S {(H)}}}$. We can use the fact that $\S {(H)}$ is a complex manifold, cover it with local coordinate charts $\{ x^i, i=1,...,\text{dim } \S {(H)} \}$, and find the critical points $\partial W_{(H)} / \partial x^i=0$ in every chart. Alternatively, we can use Lagrange multipliers and find the critical points of $\hw_{(H)} + c^{\beta} K^{(H)}_{\beta}$. Here $K^{(H)}_{\beta}(\hp)=0$ are the equations (partially) defining $\S {(H)}$. In fact the $K^{(H)}_{\beta}(\hp)$ are polynomials, their zero set is the smallest algebraic set containing $S {(H)}$, i.e., the Zariski closure $\overline{\S {(H)}}$ which, according to Theorem I.b, is the union of $\S {(H)}$ and the smaller dimensional strata in its boundary. Any stationary point of $\hw_{(H)} + c^{\beta} K^{(H)}_{\beta}$ outside $\S {(H)}$ has to be discarded. The Lagrange multiplier method is “safe" because it only requires that the constraints $K_{\b}^{(H)}(\hp)$ satisfy the condition rank $\partial K^{(H)}_{\b} / \partial \hp ^j =$ maximal. As $\S {(H)}$ is a complex manifold, points in $\S {(H)} \subseteq \overline{\S {(H)}}$ are smooth, and the rank condition is met at the stationary points that are not discarded. This guarantees the validity of applying Lagrange multipliers to this problem.\ [Example \[lec\].1:]{} Assume a given theory contains no $G$ singlets, then $\S {(G)} = \{\hp=0 \}$ is zero dimensional and $d \hw_{\|_{\S {(G)}}}=0$ is trivially satisfied, thus $\S {(G)} \subseteq \mc^W$. In a microscopic description we prove $0 = \hp(0) \in \mc^W$ by noting that, since there are no gauge singlets, $\hp(\phi)$ is at least quadratic in $\phi$ and so $dW$ eq. (\[dw\]) equals zero at the point $\phi=0$.\ [Example \[lec\].2:]{} Consider the $SO(N)$ with $2{\raisebox{-.5pt}{{\hbox{\rule{0.4pt}{6.5pt}\hskip-0.4pt\rule{6.5pt}{0.4pt}\hskip-6.5pt\rule[6.5pt]{6.5pt}{0.4pt}}\rule[6.5pt]{0.4pt}{0.4pt}\hskip-0.4pt\rule{0.4pt}{6.5pt}}}}$ theory. The basic invariants are $S_{ij} = Q^{\a}_{i} Q^{\a}_{j}$, $\mc = \{S_{ij} \} = \c^3$. There are three strata: $$\begin{aligned} \label{sostrat} \S {SO(N-2)} &=& \{ S \| \;\text{det}\; S \neq 0 \}, \\ \nonumber \S {SO(N-1)} &=& \{ S \neq 0 \|\; \text{det}\; S \; = 0 \}, \\ \nonumber \S {SO(N)} &=& \{ S = 0 \}. \end{aligned}$$ The polynomials $K^{(H)}_{\b}$ in the definition of the strata are $K_1^{SO(N-1)}= S_{11}S_{22} - {S_{12}}^2; \;\; K_{ij}^{SO(N)}= S_{ij}, (i,j)=(1,1),(1,2),(2,2)$, no constraints for $\S {SO(N-2)}$. The equation $\text{det } S = 0$ actually defines the [closure]{} of $\S {SO(N-1)}$ where $\partial (\text{det } S)/ \partial S_{ij}$ fails to have constant rank because of the included boundary point $S=0$. The additional condition $S \neq 0$ in the definition of $\S {SO(N-1)}$ excludes the boundary, problematic point that would invalidate the Lagrange multipliers method.\ Assume $\hw(S_{ij}) = mS_{22}$. We will find $\mc^W$ using the two methods described above. (i)\ [Vacua at $\S {SO(N-2)}$:]{} $\S {SO(N-2)}$ is an open subset of $\c^3$, $\{(S_{11},S_{12},S_{22})\}$ is an appropriate set of (global) coordinates. There are no critical points of $\hw_{SO(N-2)}(S_{11},S_{12},S_{22}) = mS_{22}$, there is no vacuum at the principal stratum.\ [Vacua at $\S {SO(N-1)}$:]{} $\S {SO(N-1)} $ can be covered with two coordinate patches: $\S {SO(N-1)}^{(A)}$, the set defined by $S_{11} \neq 0$ and $\S {SO(N-1)}^{(B)}$, the open subset where $S_{22} \neq 0$. The coordinates are $$S_{ij} = \left( \begin{array}{cc} x & y \\ y & y^2/x \end{array} \right) \;\;x \neq 0 \; \text{on } \S {SO(N-1)}^{(A)}, \; \;\; S_{ij} = \left( \begin{array}{cc} y^2/z & y \\ y & z \end{array} \right) \;\; z \neq 0 \; \text{on } \S {SO(N-1)}^{(B)}.$$ We find that ${\hw} _{SO(N-1)}(y,z) = mz$ at the $B$ chart, $d {\hw} _{SO(N-1)} = 0$ has no solutions there. At $\S {SO(N-1)}^{(A)}$, $\hw _{SO(N-1)}(x,y) = my^2/x$, and we find the solutions $S_{ij}= \text{diag} (x,0), x \neq 0$.\ [Vacua at $\S {SO(N)}$:]{} the only point of this zero dimensional manifold is a vacuum.\ Taking the union of the solution sets we arrive at: $$\label{mod} \mc ^W = \{ S_{ij} \| S_{12} = S_{22} = 0 \}.$$ (ii)\ [Vacua at $\S {SO(N-2)}$:]{} we find the extrema of $f(S_{11},S_{12}, S_{22}) = m S_{22}$ and keep only the solutions satisfying det $S \neq 0$. There are no solutions.\ [Vacua at $\S {SO(N-1)}$:]{} we find the extrema of $ f(S_{11},S_{12},S_{22}) = m S_{22} + \alpha (S_{11} S_{22} - S_{12}^2)$ and discard $S=0$ as a solution. The solutions are $\a \neq 0, S_{ij} = \text{diag}\; (-m/\a,0)$.\ [Vacua at $\S {SO(N)}$:]{} we look for stationary points of $f(S_{11},S_{12},S_{22}) = mS_{22}+\a S_{11} + \b S_{12} + \gamma S_{22}$ and find $S_{ij}=\a=\b=m+\gamma=0$.\ Taking the union of the solution sets we recover (\[mod\]). Irreducible components of $W \neq 0$ moduli spaces {#comp} -------------------------------------------------- An algebraic set is said to be irreducible if it is not the union of two distinct algebraic sets. Every algebraic set $X$ can be uniquely decomposed as $X = \cup_{i=1}^r X_i$, with $X_i$ irreducible and $r$ minimal. As an example, the set $X \subset \c^2 = \{(x,y) \}$ defined by the equation $xy=0$ has two irreducible components: $X = \{(x,y)\| x=0\} \cup \{ (x,y)\| y=0 \}$. The moduli space $\mc$ of a supersymmetric gauge theory with zero superpotential is irreducible, because is the image under the regular (polynomial) map $\pi$ of the irreducible set $\c^n$ [@clo], the vector space of elementary fields. However, when a superpotential is added, $\mc^W$ is generically reducible. We will see that complete irreducible components of $\mc^W$ can be obtained by finding their vacua just at the maximal stratum intersecting the component, instead of searching in every stratum. This is particularly useful if $\mc^W$ is known a priori to be irreducible, case in which we will only need to solve the equation $d \hw _{(H)}=0$ in a single stratum. A trivial example of an irreducible moduli space $\mc^W$ is when $\mc^W$ consists a single point. Such theories are interesting because they may lead to dynamical supersymmetry breaking in the quantum regime [@dsb]. Another example arises in the process of integrating out heavy composites from an effective superpotential $W_{eff}$. A tree level mass term $W_{mass}=m\hp^1$ is added to a supersymmetric gauge theory whose low energy effective superpotential $W_{eff}(\hp)$ is known. The effective superpotential of the resulting theory is obtained by integrating out the heavy composites $\hp^i, i=1,2,...,r \leq s$ from $W_{eff}$, usually identified from the elementary field content of $\hp^1$ and the other invariants. The heavy composites can also be identified using the stratification of the zero superpotential classical moduli space $\mc = \{ \hp \in \c^s \| p_{\a}(\hp)=0 \}$, without knowing the elementary quark content of the invariants. The light elementary fields $\phi$ span the vector space $\c^n_{W_{mass}} = \{ \phi \in \c^n \| \partial W_{mass} / \partial \phi^i = 0\}$, which is irreducible, then $\mc^{W_{mass}}= \pi(\c^n_{W_{mass}})=\{ \hp \in \c^s \| p_{\a}(\hp)=0, \; \text{ and } \; \hp^j=0, j=1,...,r \}$ is also irreducible, and excludes precisely the heavy fields to integrate out from $W_{eff}$. The problem of identifying heavy composites reduces to finding the irreducible classical moduli space $\mc^{W_{mass}}$, which can be done using the stratification of $\mc$. For irreducible moduli spaces $\mc^W$, important simplification arise in the methods described in [@plb].\ Let $$\mc^W = \bigcup_i {\mc^W}_{(i)}$$ be the decomposition of $\mc^W$ into irreducible components. As proved in Appendix A, the set of strata intersecting ${\mc^W}_{(i)}$ contains a unique maximal element $\S {(H_i)}$. Furthermore (eq (\[muf\])) $$\label{icms} {\mc^W}_{(i)} = \overline{{\mc^W}_{(i)} \cap \S {(H_i)}}.$$ The above equation tells us that once the maximal set intersecting ${\mc^W}_{(i)}$ is found, we only need to find the stationary points of $\hw_{(H_i)}$ and take the closure of the resulting set. In taking the closure, we are actually incorporating all the other vacua in the smaller strata intersecting ${\mc^W}_{(i)}$ without solving the corresponding stationary point equations. If $\mc^W$ is irreducible, we only need to solve the equation $d \hw _ {(H)} = 0$ on a single stratum (the maximal stratum intersecting $\mc^W$), then take the closure of the critical point set, otherwise we follow the procedure described below. ### Procedure to obtain $\mc^W$ This procedure is based on the fact that the set of strata intersecting an irreducible component ${\mc^W}_{(i)}$ of the moduli space contains a single maximal element $\S {(H_i)}$ and eq (\[icms\]) holds. It stops after a few steps if $\mc^W$ is irreducible.\ [Procedure to obtain ${\mathbf \mc^W}$:]{} $\mc^W \subset \mc \subseteq \c^s$ can be obtained, one (subset of) irreducible component(s) at a time, by means of the following procedure:\ - Arrange the partially ordered set of strata of $\mc$ as explained at the beginning of Section \[exs\]. By Theorem I.c the first and last columns contain a single entry ($\S {(G_P)}$ and $\S {(G)}$ respectively). The set of paths through linked strata give all the different patterns of gradual symmetry breaking from $G$ to $G_P$. - Look for solutions of $d {\hw}_{(G_P)} = 0$. If there are solutions, take the closure of the solution set $\{ \hp \in \S {(G_P)} \| d {\hw}_{(G_P)}(\hp) = 0 \}$, this yields one or more complete irreducible components of $\mc^W$. - Look for new solutions in the strata in the next column, if there are new solutions, say in $\S {(H)}$, go to \[iv\], otherwise repeat \[iii\]. - Take the closure of the solution set to obtain further irreducible components of $ \mc ^W$. - Look for new solutions in the other strata in the column of $(H)$, if any, go to \[iv\], otherwise go to \[iii\] Solutions to $d \hw _{(H)} =0$ can be found either by covering the stratum with local coordinates or by using Lagrange multipliers, as explained above. Step iv saves us some work, in taking the closure we obtain some solutions $dW_{(H')}=0, (H') > (H)$ without actually performing explicit computations. However, if $\mc^W$ is reducible, $\overline{\mc^W \cap \S {(H)}}$ does not necessarily exhaust the solution set $\bigcup_{(H') \geq (H)} (\mc^W \cap \S {(H')})$. The following example exhibits some of these subtleties.\ [Example \[comp\].1:]{} [$SO(13)$ with a spinor (Figure 1):]{} A complete classification of the $G^c$ orbits of this theory can be found in ref [@gv]. Theorem I in [@gv] states that there are two invariants, $p$ and $q$ (of degrees $4$ and $8$ in the elementary spinor) which are unconstrained, i.e., $\mc = \c^2$. There are four strata (as there are four types of closed $G^c$ orbits, the ones that contain points, see Table 1 in [@gv]), we order them as in step \[i\] of the procedure above: $$\label{sspin13} \begin{array}{cccccc} & & \S {G_2 \times SU(3)} & & &\\ & \diagup & & \diagdown & & \\ \S {SU(3) \times SU(3)}& & & && \S {SO(13)}. \\ & \diagdown && \diagup &&\\ && \S {SU(6)} &&& \end{array}$$ The equations defining the strata are the following $$\begin{aligned} \label{ssspin13} \S {SU(3) \times SU(3)} &=& \{(p,q) \| p^2-4q \neq 0 \; \text{and} \; q \neq 0 \} \\ \nonumber \S {G_2 \times SU(3)} &=& \{(p,q) \| p^2-4q = 0 \; \text{and} \; p \neq 0 \} \\ \nonumber \S {SU(6)} &=& \{(p,q) \| q = 0 \; \text{and} \; p \neq 0 \} \\ \nonumber \S {SO(13)} &=& \{(0,0)\}.\end{aligned}$$ The real section $(p,q) \in {\Bbb R}^2$ of $\mc \simeq \c^2$ and its strata is depicted in Figure 1.a. The dimensions of the strata in the first, second and third column of (\[sspin13\]) are respectively two, one and zero. We will not use Lagrange multipliers but local coordinates on the strata. $\{ (p,q) \| q \neq 0, p^2/4 \}$ is a good set of (global) coordinates on the principal stratum, whereas $p \neq 0$ can be taken as a (global) coordinate of $\S {SU(6)}$ and also of $\S {G_2 \times SU(3)}$. We apply the procedure above to solve for $\mc ^W$ in the following three cases (step [\[i\]]{} is already done in equation (\[sspin13\])):\ [(i) $ \hw (p,q) = f(p)$ (Figure 1.b).]{}\ [step [\[ii\]]{}:]{} $\hw_{SU(3) \times SU(3)}(p,q) = f(p), q \neq 0, p^2/4$. The set of critical points is $\mc^W \cap \S {SU(3) \times SU(3)} = \{ (p_i,q) \| q \neq 0, p_i^2/4 \text{ and } f'(p_i)=0, i=1,...,k \}$, $k$ the number of distinct roots of the polynomial $f'$. The closure of this set is $\{(p_i,q) \| q \in \c, i=1,...,k \}$, which is the union of $k$ irreducible sets.\ [step [\[iii\]]{}:]{} No [new]{} solution arises in $\S {G_2 \times SU(3)}$ or $\S {SU(6)}$ but those already found in taking the closure in step [\[ii\]]{}.\ [step [\[iii\]]{}:]{} If $0$ is among the $p_i$’s, there is not any new solution in $\S {SO(13)}$, otherwise we add the solution $(p,q)=(0,0)$.\ $$\label{dec-i} \mc ^W = \cup_{i=1}^k \{ (p_i,q) \| q \in \c \} \cup \{(0,0)\},$$ has $k+1$ irreducible components if $f'(0) \neq 0$, $k$ components if $f'(0)=0$.\ [(ii) $ \hw (p,q) = (p^2-4q-m^8)^2/M^{13}$, $m\neq 0$ (Figure 1.c).]{}\ [step [\[ii\]]{}:]{} $\hw _{SU(3) \times SU(3)} = \hw(p,q)$ with the restrictions $q \neq 0, p^2/4$, $d \hw _{SU(3) \times SU(3)} = 0$ gives $\mc ^W \cap \S {SU(3) \times SU(3)} = \{(p,q) \| q = (p^2-m^8)/4) \text{ and } p \neq \pm m^4 \}$. The closure of this set is $\{ (p,q) \| q = (p^2-m^8)/4 \}$.\ [step [\[iii\]]{}:]{} $\hw_{SU(6)} (p) = (p^2-m^8)^2/M^{13}$, $p \neq 0$. $d \hw_{SU(6)} = 0$ only at $p=\pm m^4$. These two solutions correspond to $\overline{\left(\mc ^W \cap \S {SU(3) \times SU(3)}\right) } \cap \S {SU(6)}$, they are not [new]{} solutions, we are still seeing the irreducible component of $\mc^W$ found in step [\[ii\]]{}. Contrast with what happens at $\S {G_2 \times SU(3)}$. $\hw_{G_2 \times SU(3)} = m^{16}/M^{13} = constant$, then $d \hw_{G_2 \times SU(3)} \equiv 0$. $\S {G_2 \times SU(3)} \subset \mc^W$ is an entire new set of solutions! In fact $\overline{\mc ^W \cap \S {SU(3) \times SU(3)}} \cap \S {G_2 \times SU(3)} = \emptyset$.\ [step [\[iv\]]{}:]{} In taking the closure of $\mc^W \cap \S {G_2 \times SU(3)}$ we add the solution $(0,0)$ that completes the $q=p^2/4$ parabola.\ [step [\[v\]]{}:]{} We go back to step [\[iii\]]{} and find the trivial solution at $\S {SO(13)}$, which is not new.\ $\mc ^W$ has two irreducible components: $$\label{dec-ii} {\mc ^W}_{(1)} = \{ (p,q) : q = (p^2-m^8)/4 \}, \;\;\; {\mc^W}_{(2)} = \{ (p,q) \| q = p^2/4 \}.$$ [(iii) $\hw (p,q) = [p(p-\a)-q]^2/M^{13}$ (Figure 1.d)]{}.\ This example is somewhat intermediate between [(i)]{} and [(ii)]{} in the sense that the closure of the solution set in a given stratum intersects smaller strata, where also new solutions arise. The superpotentials and solution sets at different strata are: $$\begin{aligned} \begin{array}{ll} \nonumber \hw _{SU(3) \times SU(3)} = \frac{[p(p-\a)-q]^2}{M^{13}}, & \mc^W \cap \S {SU(3) \times SU(3)} = \{ (p,q) \| q=p(p-\a)), \; q \neq 0,p^2/4 \}; \\ \nonumber \hw _{SU(6)} = \frac{[p(p-\a)]^2}{M^{13}}, & \mc^W \cap \S {SU(6)} = \{ (\a,0), (\a /2,0) \}; \\ \nonumber \hw _{G_2 \times SU(3)} = \frac{[\frac{3}{4}p^2-p \a]^2}{M^{13}}, & \mc^W \cap \S {G_2 \times SU(3)} = \{ (2 \a /3,\a ^2 /9), (4 \a /3, 4 \a^2 /9) \}; \\ \nonumber \hw _{SO(13)} = 0, & \mc \cap \S {SO(13)} = \{ (0,0) \}. \end{array} \label{iii} \end{aligned}$$ One of the two solutions in $\S {SU(6)}$ ($\S {G_2 \times SU(3)}$) comes from $\overline{\mc^W \cap \S {SU(3) \times SU(3)}}$, the other one belongs to a different irreducible component containing a single point. The decomposition of $\mc$ into irreducible components is $$\label{dec-iii} \mc = \{(p,q=p(p-\a)) \} \cup \{ (\a /2,0) \} \cup \{ (2 \a /3, \a ^2 /9) \}.$$ (110,90)(-20,-50) (-20,0)[(1,0)[40]{}]{} (0,-10)[(0,1)[40]{}]{} (20,-2)[$p$]{} (2,30)[$q$]{} (-20,30)[[a)]{}]{} (40,30)[[b)]{}]{} (-20,-20)[[c)]{}]{} (40,-20)[[d)]{}]{} (-17,0)[(1,0)[34]{}]{} (-10,25,0,0,10,25) (9,20)[ ${}^{\Sigma_{G_2 \times SU(3)}}$ ]{} (0,0) (-4,-2)[${}_{\Sigma_{SO(13)}}$]{} (-20,0)[${}^{\Sigma_{SU(6)}}$]{} (40,0)[(1,0)[40]{}]{} (49,30.25,60,0,71,30.25) (69,20)[ ${}^{\Sigma_{G_2 \times SU(3)}}$]{} (60,0) (56,-2)[${}_{\Sigma_{SO(13)}}$]{} (40,0)[${}^{\Sigma_{SU(6)}}$]{} (-20,-50)[(1,0)[40]{}]{} (-11,-19.75,0,-50,11,-19.75) (-11.5,-24)[ ${}^{\Sigma_{G_2 \times SU(3)}}$]{} (0,-50) (-5,-52)[${}_{\Sigma_{SO(13)}}$]{} (-20,-50)[${}^{\Sigma_{SU(6)}}$]{} (40,-50)[(1,0)[40]{}]{} (49,-19.75,60,-50,71,-19.75) (69,-30)[ ${}^{\Sigma_{G_2 \times SU(3)}}$]{} (60,-50) (50,-52)[${}_{\Sigma_{SO(13)}}$]{} (40,-50)[${}^{\Sigma_{SU(6)}}$]{} (67,-10)[(0,1)[40]{}]{} (68,-7)[${}_{{\mc^W}_{(1)}}$]{} (60,0) (-10,-25,0,-50,10,-25) (-12.6,-20.5,0,-60.75,12.6,-20.5) (11.5,-30)[${}_{{\mc^W}_{(1)}}$]{} (-10,-30)[ ${}^{{\mc^W}_{(2)}}$]{} (56.5,-20.25,62.5,-56.25,68.5,-20.25) (62.5,-50) (63.3,-47.2) \ [Figure 1: a) The real section $(p,q) \in {\Bbb R}^2$ of the moduli space $\c^2$ of the $SO(13)$ theory with a spinor analyzed in example \[comp\].1. The figure shows the strata $\S {G_2 \times SU(3)}, \S {SU(6)}$ and $\S {SO(13)}$, removing them from the plane we obtain the principal stratum $\S {SU(3) \times SU(3)}$. b) Moduli space of example \[comp\].1(i), assuming $f'(p)$ has a single (real positive) root, in which case $\mc^W$ has two irreducible components, the line ${\mc^W}_{(1)}$ and the point ${\mc^W}_{(2)}= \S {SO(13)}$. c) The two irreducible components of the moduli space of example \[comp\].1(ii) are parabolas, one of them agrees with the stratum $\S {G_2 \times SU(3)}$. d) The three irreducible components of the moduli space of example \[comp\].1(iii) are a parabola and two isolated points, one of them lying on $\S {G_2 \times SU(3)}$, the other on $\S {SU(6)}$ .]{}\ ### Integrating out heavy fields The procedure described above simplifies if $\mc ^W$ is known a priori to be irreducible: order the strata as in \[i\], then look for solutions in the first column, then the second one, etc, until solutions are found. If this first happens at $\S {(H)}$ and the solution set is $s \subseteq \S {(H)}$, then $\mc ^{W} = \overline{s}$. As an application, consider the problem of identifying composites made heavy by a mass superpotential $\hw_{mass} = m \hp$, a first step in the process of integrating out fields from an effective superpotential [@out; @susy]. The set $\c^n_{W_{mass}}$ of critical points of $W_{mass}(\phi) = \hw_{mass}(\hp(\phi))$ is a vector space, therefore an irreducible $\c^n$ algebraic subset, and so is $\mc ^{W_{mass}} = \pi(\c^n_{W_{mass}})$. If $\S {(H)}$ is [the]{} highest dimensional stratum intersecting $\mc ^{W_{mass}}$, then $\mc ^{W_{mass}} = \overline{ \{ \hp \in \S {(H)}\| dW^{mass}_{(H)}(\hp)=0 \} }$.\ [Example \[comp\].2:]{} Consider $\hw = m M^F_F$ in $F < N$ SQCD (refer to Example \[ls\].1). There are no solutions at the main stratum $\S {SU(N-F)} = {\Bbb M}^F_F$, the set of rank $F, F \times F$ matrices. We look for solutions at the only stratum in the second column, which is $\S {SU(N-F+1)} = {\Bbb M}^F_{F-1}$. We use Lagrange multipliers and look for critical points of $m M^F_F + \a \; \text{det} \; M$ satisfying cofactor $M \neq 0$. The solution set is $\mc ^W \cap \S {SU(N-F+1)} = \{ M \| M = \text{diag} \; (M_L,0) \; M_L \in {\Bbb M}^{F-1}_{F-1} \} \simeq {\Bbb M}^{F-1}_{F-1}$, taking its closure we obtain $\mc^W = \{ M \| M = \text{diag} \; (M_L,0) \} = {\Bbb M}^{F-1}$. This tells us that the heavy fields are $M^F_i$ and $M^i_F, i = 1,...,F$.\ In the special case of an irreducible $\mc ^W$ intersecting the main stratum $\S {(G_P)}$ all we need to know are the constraints defining $\mc = \overline{ \S {(G_P)}}$, as these are the ones used in the Lagrange multiplier method. method.\ [Example \[comp\].3:]{} $W=0$, $N=2, F=3$ SQCD contains six $SU(2)$ fundamentals $Q^{\a}_i, i=1,...,6$. The basic invariants are $V_{ij} = Q^{\a}_i Q^{\b}_j \epsilon_{\a \b}$. The moduli space is $\mc = \{ V \| \epsilon^{i_1i_2i_3i_4i_5i_6} V_{i_1i_2} V_{i_3i_4} = 0 \}$ and has two strata: $\S {1} = \{ V \in \mc \| V \neq 0 \}$, and $\S {SU(2)} = \{ V=0 \}$. The quantum theory develops the effective superpotential $\hw _{eff} = \epsilon^{i_1i_2i_3i_4i_5i_6} V_{i_1i_2} V_{i_3i_4} V_{i_5i_6} / \Lambda_{(F=3)}^3$, $\mc$ is the set of stationary points of $W_{eff}$. Adding a tree level superpotential $\hw=m V_{56}$ and integrating out the heavy composite fields $V_{5i}, V_{6i}$ from $\hw_{eff}+\hw_{tree}$ we obtain the quantum deformed $F=N=2$ moduli space $Pf \; V = \Lambda_{(F=2)}^4$. Suppose we want a “low energy description" of the integrating out procedure. We do not know the elementary quark composition of the $V_{ij}$’s and need to find out which fields are made heavy by $\hw = m V_{56}$. Following the above recipe, we first look for the set stationary points of the restriction of $W_{tree}$ to the main stratum of $\mc$, then take the closure of the solution set. The stationary points of $m V_{56} + \epsilon^{i_1i_2i_3i_4i_5i_6} V_{i_1i_2} V_{i_3i_4} \lambda_{i_5 i_6}$ ($\lambda_{ij}= -\lambda_{ji}$ are Lagrange multipliers) satisfy the following conditions: $\lambda \neq 0, \lambda_{5i}=\lambda_{6i}=0; V \neq 0, V_{5i}=V_{6i}=0, \epsilon^{i_1i_2i_3i_4 5 6} V_{i_1i_2} V_{i_3i_4} = 0$, and $\epsilon^{i_1i_2i_3i_456} V_{i_1i_2}\lambda_{i_3i_4} = -m/2$. We conclude the light fields are $V_{ij}, i,j \neq 5,6$, classically constrained by $\epsilon^{i_1i_2i_3i_4 5 6} V_{i_1i_2} V_{i_3i_4} = 0$. Thus, the fields to integrate out are $V_{5i}$ and $V_{6i}$. ### Potentials lifting flat directions The fact that $\mc^W \cap \S {(H)}$ is the set of stationary points of ${\hw}_{(H)}$ can be applied to a systematic search of superpotentials $\hw$ lifting the non trivial classical flat directions of a theory with given gauge group $G$ and matter content $\phi$. The interest in finding superpotentials satisfying this condition lies in the fact that the resulting theory is a candidate for dynamical supersymmetry breaking [@dsb]. If the theory contains no singlets, $d {\hw}_{(G)}=0$ is trivially satisfied, since $\S {(G)}$ is zero dimensional, and the problem in hand is finding all $\hw$ for which the equation $d {\hw}_{(H)}=0$ has no solution if $(H) < (G)$.\ [Example \[comp\].4:]{} Let us look for all superpotentials lifting flat directions in the $SO(13)$ with a spinor theory above, which are at most quadratic in the invariants $(p,q)$[^4], $\hw = Ap+Bq+Cp^2/2+Dq^2/2+Epq$. We have $$\begin{aligned} {\hw}_{SU(6)} &=& Ap+Cp^2/2, p \neq 0, \label{6} \\ {\hw}_{G_2 \times SU(3)} &=& Ap + (B/4+C/2)p^2+Ep^3/4+Dp^4/32, p\neq 0 \label{g2}.\end{aligned}$$ There are two possibilities:\ (i) The complex polynomial $Ap + (B/4+C/2)p^2+Ep^3/4+Dp^4/32$ has no zeroes, then $B+2C=D=E=0, A \neq 0$. The condition that $d{\hw}_{SU(6)}/dp =(A+Cp)$ has no $p \neq 0$ zeroes adds $C=0$, then $\hw = Ap$ and $d{\hw}_{SU(3) \times SU(3)}$ is automatically non-zero.\ (ii) The polynomial $A + (B/2+C)p + 3Ep^2/4+Dp^3/8$ has zero as its only root, then $A=0$ and only one of of $B+2C, E$ or $D$ is non-zero. Adding $d{\hw}_{(H)} \neq 0$ for $H=SU(6)$ and $SU(3) \times SU(3)$ gives $A=E=D=0$, $B,C$ and $B+2C$ non-zero.\ In conclusion, the only superpotentials at most quadratic in the invariants that lift all classical flat directions are $\hw = Ap$ and $\hw = Bq+Cp^2/2$ with $B,C,$ and $B+2C$ all different from zero.\ [Example \[comp\].5:]{} Consider the $SU(3) \times SU(2)$ model of Affleck, Dine and Seiberg [@dsb]. The matter content is a field $Q$ in the $({\bf 3}, {{\bf 2}})$, fields $\overline{u}$ and $\overline{d}$ in the $(\overline{{\bf 3}}, {\bf 1})$ and a field $L$ in the $({\bf 1}, {\bf 2})$. The basic invariants are $x^1=Q\overline{u}L, x^2=Q\overline{d}L$ and $x^3=Q\overline{u} Q\overline{d}$. They are unconstrained, then $\mc = \c^3$. The strata are readily seen to be $\S {1}~=~\{ (x^1,x^2,x^3) \| x^3 \neq 0 \}$, $\S {SU(2)}~=~\{ (x^1,x^2,x^3) \| x^3=0 \text{ and } (x^1,x^2) \neq (0,0) \}$, and $\S {SU(3) \times SU(2)}~=~\{(0,0,0)\}$. Assume $\hw$ is less than cubic in the composites, $\hw = A_i x^i + B_{ij} x^i x^j/2$. The supersymmetric vacua in $\S {1}$ and $\S {SU(2)}$ are respectively the solutions to the equations $$\begin{aligned} d {\hw}_{1} &=& B_{ij}x^j + A_i = 0, x^3 \neq 0, \label{c1} \\ d {\hw}_{SU(2)} &=& B_{i'j'}x^{j'} + A_{i'} = 0, (x^1,x^2) \neq (0,0), \label{c2}\end{aligned}$$ where $i,j=1,2,3$ and $i',j'=1,2$. Requiring that $\hw$ lifts all non trivial flat points is equivalent to demanding that the only possible solution to the linear system in (\[c1\]) be the trivial one [^5] and also that the only possible solution of the linear system in (\[c2\]) be trivial. This leads to the following three possibilities: (i) neither $B_{ij}x^j + A_i = 0$ nor $B_{i'j'}x^{j'} + A_{i'} = 0$ has a solution, (ii) $B_{ij}x^j + A_i = 0$ has no solution and $B_{i'j'}x^{j'} + A_{i'} = 0$ only for $(x^1,x^2)=(0,0)$, which implies $A_1=A_2=0$ and det $(B_{i'j'}) \neq 0$; and (iii) each linear system has the trivial solution as the only one, i.e, $A_i=0$, det $(B_{ij}) \neq 0$ and det $(B_{i'j'}) \neq 0$. As an example, $B_{ij}=0$ and $(A_1,A_2) \neq (0,0)$ is a possible solution, and choosing $A_3=0$ we obtain the only renormalizable gauge invariant superpotential lifting all flat directions.[^6] A $B_{ij} \neq 0$ example is $\hw = Bx^1x^2+Cx^3$.\ ### Patterns of gauge symmetry braking in $W \neq 0$ theories Theorem I.a,b,c gives a well defined pattern for the breaking of the gauge symmetry $G$ in theories with zero superpotential. There is an order relation in the set ${\Bbb S}$ of (classes of) unbroken subgroups of $G$ at different vacua, namely $(H) < (H')$ if $H$ is conjugate to a proper subgroup of $H'$. ${\Bbb S}$ contains a unique maximal class $(G)$ and a unique minimal isotropy group $(G_P)$, and, when ${\Bbb S}$ is arranged as explained at the beginning of Section (\[exs\]), all patterns of gauge symmetry breaking of the $W=0$ theory from $G$ to $G_P$ are exhibited. If a superpotential $W$ is turned on, the resulting moduli space will intersect [some]{} of the strata $\S {(H)}$ of the $W=0$ theory. From the stratification $\mc = \cup_{(H)} \S {(H)}$ of $\mc$, and the fact that $\mc^W \subset \mc$, we obtain the stratification of $\mc^W$: $$\label{ws} \mc^W = \bigcup_{(H) \in \ws} \left( \mc^W \cap \S {(H)} \right),$$ $\ws$ being the set of (classes of) unbroken subgroups at vacua in the theory with superpotential $W$, i.e., the set of strata intersecting $\mc^W$. As $W$ lifts flat directions, some of the unbroken subgroups of the $W=0$ theory are missing in $\ws$. The partial order relation in ${\Bbb S}$ is inherited by $\ws$, this is used to order the $\mc^W$ strata $\mc^W \cap \S {(H)}$. It is then natural to ask if some of the conditions in Theorem Ia,b,c subsist in the theory with superpotential. Consider first Theorem I.a, the stratification (\[ws\]) is finite, but it is easy to see that, generically, the strata are not manifolds. Consider e.g. the $SO(13)$ theory with a spinor of Example \[comp\].1 with a superpotential $\hw(p,q) = (p-p_0)^2(q-q_0)^2, q_0 \neq 0, p_0^2/4$. The $SU(3) \times SU(3)$ stratum of this theory, being singular at $(p_0,q_0)$, is not a manifold. Point (b) in Theorem I does not hold if $W \neq 0$, the three superpotentials in Example. \[comp\].1 illustrate this fact. Most important, point (c) in Theorem I is no longer true either. Generically, the set of minimal (classes of) unbroken subgroups contains more than one element. A simple example is the $SO(13)$ theory with a spinor and superpotential $\hw(p,q)=q(q-p^2/4)$, which exhibits the following pattern of symmetry breaking: $$\label{p1} \begin{array}{ccc} & & G_2 \times SU(3) \\ & \diagup & \\ SO(13) & & \\ & \diagdown &\\ && SU(6) \end{array}$$ Although dim $G_2 \times SU(3) < \text{ dim } SU(6)$, $G_2 \times SU(3)$ is not conjugate to an $SU(6)$ subgroup, there is no Higgs flows between these two unrelated theories. A unique maximal unbroken gauge subgroup (minimal stratum) exists if the theory contains no $G$ singlets, this is $(G)$ ($\S {(G)}$). Yet, theories with a gauge singlet may not even have a maximal unbroken gauge subgroup when a superpotential is turned on. As an example, add an $SO(13)$ singlet $r$ to the $SO(13)$ theory with a spinor. The moduli space is $\mc = \{(p,q,r)\} = \c^3$ and the strata are the sets of $(p,q,r)$ constrained by the same equations in (\[ssspin13\]). Take $\hw(p,q,r)=r(p-p_0), p_0 \neq 0$, then $\mc^W$ is the line $\{ (p_0,q,0), q \in \c \}$ which does not intersect $\S {SO(13)} = \{ (0,0,r) \}$. The pattern of gauge symmetry breaking of this theory, $$\label{p2} \begin{array}{ccc} G_2 \times SU(3) & & \\ & \diagdown & \\ & & SU(3) \times SU(3) \\ & \diagup &\\ SU(6) && \end{array}$$ has [two]{} maximal $SO(13)$ subgroups (minimal strata) from where to start flowing down to smaller subgroups by Higgs mechanism. The reader can check that the superpotential $\hw = q(q-p_0^2/2)+r(p-p_0)^2, p_0 \neq 0$ lifts all $SO(13)$ and $SU(3)\times SU(3)$ vacua, then the moduli space of this theory has two maximal (minimal) unbroken gauge subgroups.\ The situation gets better if we consider instead [irreducible components]{} ${\mc^W}_{(i)} \subseteq \mc^W$. According to the results in Appendix A, there is a unique maximal stratum $\S {(H_i)}$ intersecting ${\mc^W}_{(i)}$ and equation (\[icms\]) holds. This is analogous to equation (\[closure\]) in Theorem I.b when applied to the maximal stratum (only). Irreducible moduli spaces share this important property with the $W=0$ (irreducible) moduli spaces.\ The results in Section \[comp\] are gathered below.\ [Corollary 1 of Theorem I:]{} Let $\hp ^i(\phi), i=1,...,s$, be a basic set of holomorphic $G$ invariants of the theory with matter content $\{ \phi \}$ and gauge group $G$, $p_{\a}(\hp(\phi))\equiv 0$ the algebraic constraints among the basic invariants, $\mc = \{ \hp \in \c^s \| p_{\a}(\hp)=0 \}$ the moduli space of the $W=0$ theory. Let $\S {(H)} \subseteq \mc$ be the stratum of vacua with (classes of) unbroken gauge subgroups conjugate to $H \subseteq G$, $K_{\b}^{(H)}(\hp)=0$ the polynomial equations defining (the closure of) $\S {(H)}$. Let $W(\phi) = \hw (\hp (\phi))$, be a superpotential and $\hw _{(H)}$ the restriction of $\hw$ to the complex manifold $\S {(H)}$. - The set of vacua in $\S {(H)}$, $\mc ^W \cap \S {(H)}$, is the set of critical points $d \hw _{(H)}~=~0$ [@plb; @as]. This can be obtained (i) by covering the complex manifold $\S {(H)}$ with local coordinates $x^i$ and solving $\partial \hw_{(H)}(x) / \partial x^i =0$, or (ii) by using Lagrange multipliers to find the stationary points of $\hw(\hp)+ C^{\b} K^{(H)}_{\b}(\hp)$, and then discarding the solutions not in $\S {(H)}$. - Generically, if $W \neq 0$ the strata $\S {(H)} \cap \mc^W$ are [not]{} manifolds, $\overline{\mc^W \cap \S {(H)}} \neq \cup_{(L) \geq (H)} (\mc^W \cap \S {(L)})$, and the sets of maximal and minimal classes of unbroken gauge subgroups contain more than one element. - If $\mc^W = \cup_i {\mc^W}_{(i)}$ is the decomposition of $\mc^W$ into irreducible components, then for each $i$ there is a maximal stratum $\S {(H_i)}$ intersecting ${\mc^W}_{(i)}$, and ${\mc^W}_{(i)} = \overline{{\mc^W}_{(i)} \cap \S {(H_i)}}$. Massless fields after Higgs mechanism {#mf} ------------------------------------- The differential $\pi'_{\phi_0}$ of the map $\pi: \phi \to \hp(\phi)$ at the point $\phi_0$ is given by the matrix $\partial \hp^i(\phi_0) / \partial \phi^j, \; \pi'_{\phi_0}: \delta \phi^j \to \delta \hp^i = (\partial \hp^i(\phi_0)/ \partial \hp^j) \delta \hp^j$. Note that $\pi: \c^n \to \mc = \{ \hp \in \c^s \| p_{\a}(\hp)=0 \}$, then $\pi'_{\phi_0}: T_{\phi_0}\c^n \to T_{\hp_0} \mc$, $\hp_0 \equiv \hp(\phi_0)$. The tangent at $\phi_0$ of $\c^n$ is $T_{\phi_0}\c^n \simeq \c^n$, and the tangent $T_{\hp_0}\mc$ is the space of moduli $\delta \hp$ consistent with the linearized constraints, $(\partial p_{\a}(\hp_0)/ \partial \hp^j) \delta \hp^j = 0$ (assuming the constraints satisfy the requirement in footnote 2.) A natural question to ask is whether $\pi'_{\phi_0}$ makes $T_{\hp_0}\mc^W \subseteq T_{\hp_0} \mc$ isomorphic to the space of massless modes at a supersymmetric vacuum $\phi_0$ in the classical regime. We devote this section to answering this question.\ \ The space $\{ \delta \phi \} = T_{\phi_0} \c^n = \t {\phi} \oplus \n {\phi} \oplus \s {\phi}$, $\delta \phi$ uniquely decomposes as $\delta \phi = \delta t + \delta n + \delta s$. The fields $\delta t$ in $\t {\phi}$ are eaten by the broken gauge generators (two real fields per heavy vector superfield). Thus, if $W=0$, the light fields in unitary gauge, i.e., the massless fields after Higgs mechanism (MFHM) are those in $\n {\phi} \oplus \s {\phi} \equiv {\rm NMFHM} \oplus {\rm SMFHM}$, where (N)SMFHM is a short for (non)singlet massless fields after Higgs mechanism. According to Theorem I.e $\pi'_{\phi_0}$ annihilates $\n {\phi_0}$, [the [NMFHM]{} are not represented in $T_{\hp_0} \mc$]{}. On the other hand, the rank of $\pi'_{\phi_0}$ is not the whole $T_{\hp_0}\mc$ but the tangent to the stratum $\S {(G_{\phi_0})} \equiv \S {\hp_0}$ through $\hp_0$, and so [there are spurious fields $C_{\hp_0} \subseteq T_{\hp_0}\mc$, unrelated to the]{} MFHM. The situation is illustrated in the following diagram: $$\begin{aligned} \label{w=0} \begin{array}{ccccc} T_{\hp_0}\mc &=& T_{\hp_0} \S {\hp_0}& \oplus & C_{\hp_0} \\ && \, \\| & &\\ {\rm MFHM} &=& \s {\phi_0}& \oplus &\n {\phi_0}\\ \end{array}\end{aligned}$$ We would like to know when $C _{\hp_0}$ and $\n {\phi_0}$ are null. We consider separately the following two cases:\ (i) $\hp_0 \in \S {(G_P)}$ ($\S {\hp_0} = \S {(G_P)})$: From Theorem I.b-c $\mc = \overline{\S {\hp_0}}$, then $T_{\hp_0} \S {\hp_0} = T_{\hp_0}\mc$ and $C_{\phi_0}$ is null. From Theorem I.f $\n {\phi_0}$ is null if and only if the theory is stable.\ (ii) $\hp_0 \notin \S {(G_P)}, (\S {\hp_0} < \S {(G_P)})$: From Theorem I.b $\S {\hp_0}$ lies in the boundary of the principal stratum, dim $\S {\hp_0} < \text{dim } \S {(G_P)} = \text{dim } \mc \leq \text{dim } T_{\hp_0} \mc$, and so $T_{\hp_0} \S {\hp_0} \subsetneq T_{\hp_0} \mc$, $C _{\hp_0}$ is non trivial. In this case also $(G_{\phi_0}) > (G_P)$ i.e., $G_P$ is conjugate to a proper subgroup of $G_{\phi_0}$, as follows from the definition of the order relation among strata and isotropy classes, and so dim $G_{\phi_0} > \text{ dim } G_P$. We can use this information together with Theorem I.e to show that $\n {\phi_0}$ is not null. Pick any point $\phi_1$ such that $\hp(\phi_1) \in \S {(G_P)}$, then (see footnotes 1 and 2) $$\begin{aligned} \label{dn0} \text{dim } \n {\phi_0} &=& n - (\text{dim}_{\Bbb R} G - \text{dim}_{\Bbb R} G_{\phi_0}) - \text{dim } \S {(G_{\phi_0})}\\ \nonumber & > & n - (\text{dim}_{\Bbb R} G - \text{dim }_{\Bbb R} G_P) - \text{dim } \S {(G_P)} = \text{dim } \n {\phi_1} \geq 0,\end{aligned}$$ In other words, Higgs mechanism at a vacuum $\phi_0$ with $(G_{\phi_0}) > (G_P)$ always leaves a theory with fields transforming non trivially under $G_{\phi_0}$.\ In conclusion, for any $W=0$ theory, spurious fields in $T_{\hp_0} \mc$ are always present unless $\hp_0$ belongs to the principal stratum. $\pi'_{\phi_0}$ is an isomorphism between the space of SMFHM and $T_{\hp_0} \S {\hp_0}$. The NMFHM are unseen as moduli $\delta \hp$, they are always present, except at vacua in the principal stratum of a stable theory.\ \ The space of massless fields at the supersymmetric vacuum $\phi_0$ is the kernel of $W_{ij}(\phi_0) = \partial^2 W(\phi_0)/\partial \phi^i \partial \phi^j$. The kernel includes the eaten fields $\t {\phi_0}$, as follows from the $G^c$ invariance of $W$ $$W_i(\phi)T^i_k \phi^k = \frac{d}{ds}{\Bigg\vert_{s=0}} W(e^{sT}\phi) = 0, \;\;\; \forall \; T \in \lie {G^c}, \phi \in \c^n,$$ by taking a $\phi$ derivative an using the $F-$flatness of $\phi_0$: $$0 = \frac{\partial}{\partial \phi^j}{\Bigg\vert_{\phi=\phi_0}} W_i(\phi) T^i_k \phi^k = W_{ij}(\phi_0) T^i_k \phi_0^k, \;\;\; \forall \; T \in \lie {G^c}.$$ As $W_{ij}(\phi_0)$ is $G_{\phi_0}$ invariant, it cannot mix $\n {\phi_0}$ and $\s {\phi_0}$, otherwise, we could write a $G_{\phi_0}$ invariant mass term $W_{ij}(\phi_0) \delta \phi^i \delta \phi ^j$ mixing singlets $\delta s$ with non singlets $\delta n$. We conclude that, under $\c^n = \t {\phi_0} \oplus \n {\phi_0} \oplus \s {\phi_0}$, $W_{ij}$ is block diagonal: $$\begin{aligned} \label{wij} && \begin{array}{ccc} \t {\phi_0} \n {\phi_0} & \s {\phi_0} \end{array} \\ \nonumber W_{ij}(\phi_0) = \begin{array}{c} {\t {\phi_0}} ^* \\ {{\n {\phi_0}}}{}^* \\ {\s {\phi_0}} ^* \end{array} && \left( \begin{array}{ccc} 0 &0&0 \\ 0 & N_{ij} & 0 \\ 0&0& S_{ij} \end{array} \right) \end{aligned}$$ After Higgs mechanism we are left with $\n {\phi_0} \oplus \s {\phi_0}$ and so MFHM $= \text{ker } S_{ij} \oplus \text{ ker } N_{ij} \equiv {\rm SMFHM} \oplus {\rm NMFHM}$. We consider the SMFHM space first. In view of equation (\[w=0\]), $\pi'_{\phi_0}$ makes $\s {\phi_0}$ isomorphic to $T_{\hp_0} \S {\hp_0}$. From this isomorphism and the inverse function theorem follows that a neighborhood of the origin of $\s {\phi_0}$ can be used as a coordinate patch of the complex manifold $\S {\hp_0}$ around $\hp_0$. Note that if $x^j$ and $y^k$ are any two local coordinate sets of $\S {\hp_0}$ with $x=y=0$ at $\hp_0$, and $\hp_0 \in \mc^W$, then $\partial {\hw}_{(G_{\phi_0})} / \partial y^k = 0$ at $y=0$ (Corollary 1.a in Section (\[comp\])), and $$[\hw_{(G_{\phi_0})}]_{ij}(\hp_0) \equiv \frac{\partial^2 \hw _{(G_{\phi_0})}}{\partial x^i \partial x^j} {\Bigg\vert_{x=0}} = \left( \frac{\partial^2 \hw _{(G_{\phi_0})}}{\partial y^k \partial y^l} {\Bigg\vert_{y=0}}\right) \left( \frac{\partial y^k}{ \partial x^i}{\Bigg\vert_{x=0}} \right) \left(\frac{\partial y^l}{ \partial x^j }{\Bigg\vert_{x=0}}\right)$$ transforms as a $(0,2)$ tensor at $\hp_0$ [^7], then $$\text{ker } [\hw_{(G_{\phi_0})}]_{ij}(\hp_0) = \left\{ \delta x^i \, \| \,\frac{\partial^2 {\hw}_{(G_{\phi_0})}(\hp_0)}{\partial x^i \partial x^j} \delta x^j =0 \right\}$$ is a well defined (coordinate independent) subspace of $T_{\hp_0} \S {\hp_0}$ with complement $D_{\hp_0}^W$. This subspace is obtained by linearizing at $\hp_0$ the constraints $\partial \hw _{(G_{\phi_0)}} / \partial x^i = 0$ defining $\mc^W \cap \S {(G_{\phi_0})}$ (Corollary 1.a), then is the tangent space $T_{\hp_0} (\mc^W \cap \S {(G_{\phi_0})})$ [^8]. In the coordinates $\delta s$ of $\S {\hp_0}$, $[\hw_{(G_{\phi_0})}]_{ij}=S_{ij}$, the $W \neq 0$ analogous of eq.(\[w=0\]) is $$\begin{aligned} \label{wneq0} \begin{array}{ccccccc} T_{\hp_0}\mc &=& \text{ ker } [\hw_{(G_{\phi_0})}]_{ij} &\oplus & D_{\hp_0}^W & \oplus & C_{\hp_0} \\ && \\| & &&&\\ {\rm MFHM} &=& \text{ker } S_{ij} & \oplus &\text{ker } N_{ij} && \\ \end{array}\end{aligned}$$ Among the MFHM, the SMFHM ker $S_{ij} \simeq \text{ ker } [\hw_{(G_{\phi_0)}}]_{ij}$ are represented as moduli, whereas the NMFHM ker $N_{ij}$ are not. The moduli in $ D_{\hp_0}^W \oplus C_{\hp_0}$ are spurious. We establish conditions under which the space ker $N_{ij}$ of NMFHM is null:\ (i) $\hp_0 \in \S {(G_P)}$: If the theory is stable, $\n {\hp_0}$ is null (Theorem I.f) and so is ker $N_{ij}$. If the theory is unstable, $\n {\phi_0}$ is non trivial and the theory with gauge group $G_P = G_{\phi_0}$ and matter content $\{ \delta n \} = \n {\phi_0}$ has no holomorphic $G_{\phi_0}$ invariants. In particular, $N_{ij} \delta n^i \delta n^j$, being holomorphic and $G_{\phi_0}$ invariant, must be zero, then $N_{ij}=0$ and ker $N_{ij}=\n {\phi_0}$ is not null.\ (ii) $\hp_0 \notin \S {(G_P)}$: According to eq.(\[dn0\]) dim $\n {\phi_0} > 0$. However, no general statement can be made about ker $N_{ij} \subseteq \n {\phi_0}$ if $W$ is unknown. An exception is when the theory with gauge group $G_{\phi_0}$ and matter content $\n {\phi_0}$ is known to be chiral (no quadratic holomorphic invariants), case in which we can repeat the argument above to show that $N_{ij}=0$ and so ker $N_{ij}=\n {\phi_0}$ is not null.\ These results are gathered in the Corollary below:\ [Corollary 2 of Theorem I:]{} The space MFHM of massless fields after Higgs mechanism at a vacuum with residual gauge group $H$ is the direct sum of the $H$ singlet space SMFHM and the non-singlet space NMFHM - Let $x^i$ be any set of local coordinates of $\S {(H)}$ around a vacuum $\hp_0$. SMFHM is isomorphic to the subspace $\{ \delta x^i \, \| (\partial^2 \hw_{(H)} (\hp_0)/ \partial x^i \partial x^j)\ \, \delta x^j = 0 \} \subseteq T_{\hp_0} \S {(H)}$. ${\rm SMFHM} = T_{\hp_0}(\mc ^W \cap \S {H})$ (see however footnote 8). - The NMFHM are annihilated by $\pi'_{\phi_0}$, and so they are missing (unseen as moduli $\delta \hp$) in the moduli space. For any $W$, this set is trivial if $\hp_0$ belongs to the principal stratum of a stable theory, non-trivial if $\hp_0$ is in the principal stratum of an unstable theory. - At vacua in non principal strata there are (potentially) missing NMFHM if $W=0$ ($W \neq 0$). [Example \[mf\].1.]{} Coming back to Example \[lec\].2, at $\S {SO(N-1)} ^{(A)}$ is $\hw_{SO(N-1)} = my^2/x, x \neq 0$, then the vacuum condition $d \hw _{SO(N-1)} = (-my^2/x^2,2my/x) = 0$ implies $y=0$ and $$\left( \hw_{SO(N-1)}\right)_{ij} = \frac{2m}{x} \left( \begin{array}{cc} y^2/x^2 & -y/x \\ -y/x & 1 \end{array} \right) = \frac{2m}{x} \left( \begin{array}{cc} 0 & 0 \\ 0 & 1 \end{array} \right),$$ giving a single massless $SO(N-1)$ singlet after Higgs mechanism, a fact that can be readily verified in a microscopic field description.\ [Example \[mf\].2]{} We continue the analysis of the three different cases of Example \[comp\].1.\ [(i) $ \hw (p,q) = f(p) \equiv (p-p_0)^2, p_0 \in {\Bbb R}^{>0}$ (Figure 1.b).]{}\ Using coordinate charts as in Example \[comp\].1 we get $$\begin{aligned} \label{rank} [ \hw_{SU(3) \times SU(3)} ]_{ij} &=&\text{diag }~(f''(p), 0), \\ \nonumber [\hw_{SU(6)}] _{ij} &=& f''(p), \\ \nonumber [\hw_{G_2 \times SU(3)}] _{ij} &=& f''(p),\end{aligned}$$ the dimensions of the SMFHM space at $SU(3) \times SU(3), G_2 \times SU(3), SU(6)$ and $SO(13)$ vacua equal $1,0,0$ and $0$ respectively. Note that there is no problem of the kind mentioned in footnote 8. We can use Corollary 2.a, SMFHM $ = T_{\hp_0} ( \mc^W \cap \S {(G_{\phi_0})})$, and the dimension of SMFHM can easily be read off from fig. 1.b. At the $\hp=0$ vacuum we have the original theory, for which the space of SMFHM is null, that is why dim $T_{0}(\S {(G)} \cap \mc^W)=$ dim $\S {(G)}=0$. The (real section) $(p,q) \in {\Bbb R}^2$ of the component $p=p_0$ of $\mc^W$ is a vertical line intersecting all strata but $\S {SO(13)}$ (fig.1.b). The line intersects $\S {SU(3) \times SU(3)}, \S {G_2 \times SU(3)}$, and $\S {SU(6)}$ at sets of dimension $1,0$ and $0$, these are the dimensions of the SMFHM spaces for vacua in these strata. All vacua in the main stratum have a null NNMFHM space, because the theory is stable. At vacua in smaller strata there could be NMFHM, unseen as moduli $\delta \hp$.\ [(ii) $ \hw (p,q) = (p^2-4q-m^8)^2/M^{13}$ (Figure 1.c).]{}\ We use again Corollary 2 to read from figure 1.c the dimension of the SMFHM space at each vacuum. $\mc^W$ has two irreducible components: $\mc^W = {\mc^W}_{(1)} \cup {\mc^W}_{(2)}$, the two parabolas in figure 1.c. Although ${\mc^W}_{(1)}$ is one dimensional, its intersection with $\S {SU(6)}$ is zero dimensional, and so there is a single massless singlet at each $SU(3) \times SU(3)$ vacuum in ${\mc^W}_{(1)}$, no massless $SU(6)$ singlet at any of the two $\S {SU(6)}$ vacua. A similar analysis holds for the one dimensional manifold ${\mc^W}_{(2)}$. ${\mc^W}_{(2)} \cap \S {G_2 \times SU(3)} = {\mc^W}_{(2)} \setminus \{ \hp=0 \}$ is one dimensional, whereas ${\mc^W}_{(2)} \cap \S {SO(13)} = \{ \hp =0 \}$ is zero dimensional. Correspondingly, SMFHM is one (zero) dimensional for ${\mc^W}_{(2)}$ vacua with residual $G_2 \times SU(3)$ ($SO(13)$) gauge symmetry.\ [(iii) $ \hw (p,q) = [p(p-\a)-q]^2/M^{13}$ (Figure 1.d).]{}\ Refer to figure 1.d. The moduli space has three irreducible components: a parabola ${\mc^W}_{(1)}$ intersecting all four strata, a one point component ${\mc^W}_{(2)}$ in $\S {SU(6)}$ and a single vacuum component ${\mc^W}_{(3)}$ with residual gauge symmetry $G_2 \times SU(3)$. Every vacuum in ${\mc^W}_{(1)}$ has a one dimensional space of massless singlets except for the three vacua with residual gauge symmetry $G_2 \times SU(3)$, $SU(6)$ and $SO(13)$, which have no massless singlets in their spectra. This is so because ${\mc^W}_{(1)} \cap \S{H}$ is zero dimensional for $H = SO(13), SU(6)$ and $G_2 \times SU(3)$, whereas ${\mc^W}_{(1)} \cap \S{SU(3) \times SU(3)} = {\mc^W}_{(1)} \setminus \{ \text{ three isolated points } \}$ is one dimensional. There are no SMFHM at vacua in the other two components.\ We should stress here that the results in this section all refer to the classical regime. Although for theories with a simple gauge group $G$, matter fields $\phi$ in a $G$ representation with Dynkin index $\mu$ greater than the index $\mu_G$ of the adjoint, and $W=0$ the classical moduli space $\mc$ and the quantum moduli space are equal, it is generally [not]{} true that classical and quantum spectra of massless fields agree at every vacuum $\hp \in \mc$. As an example, consider the s-confining theories in [@sc]. These theories have an effective superpotential $W_{eff}(\hp)$ whose set of stationary points is $\mc$. In the classical theory, at the $\hp=0$ vacuum we have gauge group $G$ and matter content $\phi$, without singlets. Quantum mechanically, evidence indicates that $G$ is completely broken and the massless spectrum are the unconstrained moduli $\delta \hp$ [@sc; @susy]. A second $\mu > \mu_G$ example are the theories with a low energy dual [@susy; @sei], they have equal classical and quantum moduli spaces, but the classical and quantum massless spectra are completely different. Conclusions {#conc} =========== A low energy description of the moduli space $\mc^W$ of a $W \neq 0$, ${\cal N} = 1$ gauge theory, one in which $\mc^W$ is constructed entirely in the space spanned by the basic holomorphic invariants $\hp$ without knowing their elementary field content $\hp(\phi)$, is possible. The construction requires knowledge of the constraints among the basic invariants $\hp$ that define the $W=0$ moduli space $\mc$, and also of the stratification $\mc = \cup_H \S {(H)}$ according to the unbroken gauge subgroups class $(H)$ at different vacua. Some shortcuts are possible when searching for isolated irreducible components of $\mc^W$, a fact that is useful to identify heavy composite fields to integrate out from an effective superpotential, and to construct superpotentials that lift all flat directions, leaving a candidate theory for dynamical symmetry breaking. The stratification of $\mc$, together with the low energy construction of $\mc^W$, allows a systematic study of the patterns of gauge symmetry breaking. When $W$ is trivial, there is theory with a minimal unbroken gauge subgroup $G_P$ to which flow by Higgs mechanism leads in many different ways. A non zero superpotential, on the contrary, may leave a set of vacua with no unique minimal unbroken subgroup, then different Higgs flows end up at different theories.\ Among the massless fields after Higgs mechanism (MFHM) at a vacuum $\hp \in \mc^W$, the singlets (SMFHM) are represented by moduli $\delta \hp$, whereas the non singlet (NMFHM) are not. Being gauge invariant, $W(\phi) = \hw (\hp)$. $\mc^W \cap \S {(H)}$ is the set of critical points of the restriction $\hw_{\mid_{\S {(H)}}}$ of $\hw$ to the stratum $\S {(H)}$, whereas the space of SMFHM at a vacuum $\hp \in \S {(H)}$ is the kernel of the tensor $\nabla_i \nabla_j \hw_{\mid_{\S {(H)}}}$ at $\hp$, $\nabla$ any covariant derivative on the complex manifold $\S {(H)}$. In looking for critical points $d {\hw}_{(H)} = 0 $ local coordinates on the complex manifold $\S {H}$ can be used. An alternative is using Lagrange multipliers, adding to $\hw$ terms containing the polynomial constraints in the definition of $\S {(H)}$. The Lagrange multipliers method is safe in all cases. The space of NMFHM is null for vacua in the principal stratum (where the gauge group is broken to the minimal subgroup $G_P$) of a stable theory. In unstable theories, on the contrary, even for vacua $\hp$ at the principal stratum there are NMFHM, unseen as moduli $\delta \hp$. Unstable theories are characterized by the impossibility of breaking the complexified gauge group to a minimum dimension subgroup by a configuration. Another distinguishing feature of unstable theories is that the dimension of their $W = 0$ moduli space $\mc$ violates the rule $\text{dim}\; \mc = \text{ dim microscopic matter field space } - \text{ dim gauge group } + \text{ dim } G_P$. Theories with matter fields in a real representation of the gauge group are stable, and this is the case for most (but not all) of the allowed representations, since they must be anomaly free. Unstable theories, therefore, are rare. This work was partially supported by Fundación Antorchas, Conicet and Secyt-UNC. I thank Witold Skiba for useful comments on the manuscript and for suggesting the application of the techniques here introduced to dynamical supersymmetry breaking. Derivation of equation (40) {#appen} =========================== Let $$\label{ic} \mc^W = \bigcup_{i} {\mc^W}_{(i)}$$ be the decomposition of $\mc^W$ into irreducible components. As $\mc$ is the disjoint union of its strata $\S {(H)}$ we have $$\label{1} {\mc^W}_{(i)} = \bigcup_{\S {(H)} \in \sigma_i} \left( {\mc^W}_{(i)} \cap \S {(H)} \right),$$ where $\sigma_i$ is the set of strata intersecting ${\mc^W}_{(i)}$. Let $\sigma_i^{max}$ be the subset of maximal strata in $\sigma_i$, i.e., $\S {(H)} \in \sigma_i^{max}$ if and only if any other stratum $\S {(H')} \in \sigma_i$ is either smaller than or unrelated to $\S {(H)}$. From Theorem I.b , any stratum in $\sigma_i$ lies in the closure of a $\sigma_i^{max}$ stratum, then the union of the strata in $\sigma_i$ equals the union of the closures of the strata in $\sigma_i^{max}$ and $${\mc^W}_{(i)} = \bigcup_{\S {(H)} \in \sigma_i^{max}} \left( {\mc^W}_{(i)} \cap \overline{\S {(H)}} \right).$$ ${\mc^W}_{(i)}$ being irreducible means that one of the closed sets in the union above contains the others, i.e., there is a $\S {(H_i)} \in \sigma_i^{max}$ such that $$\label{3} {\mc^W}_{(i)} = {\mc^W}_{(i)} \cap \overline{\S {(H_i)}}.$$ Equation (\[3\]) implies that $\sigma_i^{max}$ contains a single element, namely, $\S {(H_i)}$. In fact, assuming there is a $\sigma_i^{max} \ni \S {(H)} \neq \S {(H_i)}$ leads to a contradiction: $$\label{cont} \emptyset \neq {\mc^W}_{(i)} \cap \S {(H)} = {\mc^W}_{(i)} \cap \overline{\S {(H_i)}} \cap \S {(H)} \; \Rightarrow \; \overline{\S {(H_i)}} \cap \S {(H)} \neq \emptyset.$$ From equations (\[int\]) and (\[cont\]) we get $\S {(H_i)} > \S {(H)}$, contradicting the assumption that $\S {(H)}$ is maximal. We conclude that there is a single maximal element $\S {(H_i)}$ in the set $\sigma_i$ of strata intersecting the irreducible component ${\mc^W}_{(i)}$. We will show now that we can replace ${\mc^W}_{(i)} = {\mc^W}_{(i)} \cap \overline{\S {(H_i)}}$ by the more useful formula $$\label{muf} {\mc^W}_{(i)} = \overline{{\mc^W}_{(i)} \cap \S {(H_i)}}.$$ Equation (\[muf\]) has the advantage (over equation(\[3\])) of requiring only the determination of the critical points $d \hw _{H_i} = 0$, saving us the work of explicitly finding the ${\mc^W}_{(i)}$ points in smaller strata. To prove (\[muf\]) we start by taking the closure of equation (\[1\]): $$\label{decomp} {\mc^W}_{(i)} = \bigcup_{\S {(H)} \in \sigma_i} \overline{\left( {\mc^W}_{(i)} \cap \S {(H)} \right)}.$$ Again, ${\mc^W}_{(i)}$ being irreducible means that one of the sets in the union, say $ \overline{\left( {\mc^W}_{(i)} \cap \S {(H_i')} \right)}$, contains the others. To show that $\S {(H_i')} = \S {(H_i)}$ we start from $\emptyset \neq {\mc^W}_{(i)} \cap \S {(H_i)} = \overline{\left( {\mc^W}_{(i)} \cap \S {(H_i')} \right)} \cap \S {(H_i)} \subseteq {\mc^W}_{(i)} \cap \overline{\S {(H_i')}} \cap \S {(H_i)}$ (here we use that for any two sets $A$ and $B$, $\overline{A \cap B} \subseteq \overline{A} \cap \overline{B}$. This implies $\overline{\S {(H_i')}} \cap \S {(H_i)} \neq \emptyset$ and, from equation (\[int\]), $\S {(H_i')} \geq \S {(H_i)}$. As $\S {(H_i)}$ is the maximal set intersecting ${\mc^W}_{(i)}$ it must be $\S {(H_i')} = \S {(H_i)}$, and equation (\[muf\]) follows.\ C. Procesi and G. Schwarz, Phys. Lett B 161 (1985) 117. R. Gatto and G. Sartori, Commun. Math. Phys 109 (1987) 327. R. Gatto and G. Sartori, Phys. Lett. B 157 (1985) 389, Phys. Lett. B 118 (1982) 79, F. Buccella, J. P. Derendinger and C. A. Savoy, Phys. Lett. B 115 (1982) 375. M. A. Luty, W. Taylor IV, Phys. Rev. D 53 (1996) 3399. C. Csaki, M. Schmaltz and W. Skiba, Phys. Rev. D 55 (1997) 7840, hep-th 9612207. D. Luna, Bull. Soc. Math. France, Mémoire 33 (1973) 81. M. Abud and G. Sartori, Phys. Lett. B 104 (1981) 147, Ann. of Phys. 150 (1983) 307. G. Schwarz, Publ. Math. IHES 51 (1980) 37. K. Intriligator, R.G. Leigh and N. Seiberg, Phys. Rev. D 50 (1994) 1092, hep-th/9403198. K. Intriligator and N. Seiberg, Nuc. Phys B (Proc Suppl.) 45BC (1996) 1, hep-th/9509066; M.E. Peskin, TASI 96 hep-th/9702094. I. Affleck, M. Dine and N. Seiberg, Nuc. Phys. B 256 (1985) 557; W. Skiba, Mod. Phys. Lett.A 12 (1997) 737, hep-th/9703159; E. Poppitz and S. Trivedi, Phys. Lett.B 365 (1996) 125, hep-th/9507169; K. Intriligator, N. Seiberg and S. Shenker, Phys. Lett.B 342 (1995) 152, hep-ph/9410203; Y. Shadmi and Y. Shirman, Rev. Mod. Phys. 72 (2000) 25, hep-th/9907225, and references therein. D. Cox, J. Little and D. O’Shea, [Ideals, Varieties, and Algorithms]{}, 2nd edition, Springer-Verlag, 1997. , J. Morrow and K. Kodaira, Holt, Rinehart and Winston, Inc, 1971. G. Schwarz, Inv. Math 49 (1978), 167. A.G.Élashvili, Funct. Anal. Appl. 1 (1968), 267. J. Harris, [Algebraic Geometry]{}, Springer-Verlag, 1992. V. Gatti (V. Kac) and E. Viniberghi (E. B. Vinberg), Adv. Math. 30 (1978), 137. N. Seiberg, Nuc. Phys. B 435 (1995) 129. [^1]: dim denotes complex dimension, whereas $\text{dim}_{\Bbb R}$ means real dimension, then dim $G^c = \text{dim}_{\Bbb R}\, G$. [^2]: Note however that different sets of polynomials define the same algebraic set, $\{p_{\a}\}$ must be chosen such that any polynomial $p$ vanishing on $X$ admits an expansion $p(x) = \sum_{\a} q^{\a}(x) p_{\a}(x)$ with polynomials $q^{\a}$ [@clo]. Otherwise, the span of the linearized constraints may be larger than the tangent space. As an example, the line $x_2=0$ in $\c^2 = \{ (x_1,x_2) \}$ can also be defined as the zero set of the polynomial $(x_2)^2=0$, but this second choice leads to a wrong definition of tangent space. [^3]: For a point $\phi$, ${G^c}_{\phi} = {G_{\phi}}^c$ [@gatto], then the complex dimension $\text{dim } {G^c}_{\phi}$ equals the real dimension $\text{dim}_{\Bbb R}\, G_{\phi}$. In particular, if $\pi(\phi)$ is in the principal stratum, $\text{dim } {G^c}_{\phi} = \text{dim}_{\Bbb R}\, G_{\phi} = \text{dim}_{\Bbb R} \, G_P = \text{dim } {G_P}^c.$ [^4]: Note that there is no renormalizable gauge invariant superpotential for this theory, since $p=S^4$ and $q=S^8$, $S$ the spinor field. [^5]: Any $x^3=0$ solution would also be a solution of equation (\[c2\]) unless $x^1=x^2=0$. [^6]: The Affleck, Dine and Seiberg theory corresponds to the choice $B_{ij}=0$, $A_2=A_3=0$. [^7]: This tensor can be written more covariantly as $\nabla _i \nabla_j \hw _{(G_{\phi_0})}= \partial_i \partial_j \hw _{(G_{\phi_0})} + \Gamma^k_{ij} \partial_k \hw _{(G_{\phi_0})}$, $\nabla_i$ an arbitrary covariant derivative on the manifold $\S {\hp_0}$, as the second term vanishes when evaluated at a vacuum. [^8]: It might actually be bigger than $T_{\hp_0} (\mc^W \cap \S {(G_{\phi_0})})$ if there is problem of the type indicated in footnote 2. This may happen if $\hw(\hp)$ is of high degree in the invariants (therefore non renormalizable), or the constraints defining the strata are high degree polynomials.	{ "pile_set_name": "ArXiv" }
LTH 740\ \ [ ]{} Thomas Mohaupt Theoretical Physics Division\ Department of Mathematical Sciences\ University of Liverpool\ Liverpool L69 3BX, UK\ [[email protected]]{} .2in [ABSTRACT]{} > Supersymmetric black holes provide an excellent theoretical laboratory to test ideas about quantum gravity in general and black hole entropy in particular. When four-dimensional supergravity is interpreted as the low-energy approximation of ten-dimensional string theory or eleven-dimensional M-theory, one has a microscopic description of the black hole which allows one to count microstates and to compare the result to the macroscopic (geometrical) black hole entropy. Recently it has been conjectured that there is a very direct connection between the partition function of the topological string and a partition for supersymmetric black holes. We review this idea and propose a modification which makes it compatible with electric-magnetic duality. Supersymmetric Black Holes ========================== Our setup for constructing supersymmetric black hole solutions is $N=2$ supergravity couled to $n$ vector multiplets. This arises[^1] as the effective field theory of heterotic string compactifications on $K3 \times T^2$ and of type-II string theory on Calabi-Yau threefolds. The field equations are invariant under $Sp(2n+2,\mathbbm{R})$ rotations, which generalize the electric-magnetic duality rotations of Einstein-Maxwell theory.[^2] As a consequence, all vector multiplet couplings are encoded in a single holomorphic function called the prepotential $F$. This function must be homogenous of degree 2 in its variables $Y^I$, which provide homogenous coordinates on the scalar manifold $M_{\rm VM}$. The Kähler potential for the metric on $M_{\rm VM}$ can be expressed in terms of the holomorphic prepotential. The resulting geometry is known as special (Kähler) geometry [@deWit:1984pk; @deWit:1984px]. It is possible to include a certain class of higher derivative terms involving the square of the Riemann tensor and arbitrary powers of gauge field strengths, by giving the prepotential an explicit dependence on the so-called Weyl multiplet. Associated to these terms is an infinite series of field-dependent couplings. In type-II compactifications these couplings can be computed in terms of the free energy of topologically twisted string theory [@Bershadsky:1993cx; @Antoniadis:1993ze]. As long as we neglect the higher derivative terms, we are dealing with a generalized Einstein-Maxwell theory with several abelian gauge fields and field-dependent couplings, plus a scalar sigma-model. The supersymmetric black hole solutions of such a theory are natural generalizations of the extremal Reissner-Nordström black hole. Besides that the black hole now carries several electric and magnetic charges, the new feature is that we have scalar fields which vary non-trivially as a function of the radial variable.[^3] At infinity, the solutions are asymptotically flat and the scalars can take arbitrary values in $M_{\rm VM}$. The behaviour at the horizon is radically different: the scalars cannot take arbitrary values but must take fixed point values which are determined by the electric and magnetic charges of the black hole. This is the so-called black hole attractor mechanism [@Ferrara:1995ih], which generalizes to the case where higher derivative terms are included [@LopesCardoso:1998wt; @LopesCardoso:2000qm]. Since both metric and gauge fields are determined by the scalar fields through supersymmetry, it follows that the area of the event horizon is a function of the electric and magnetic charges, and does not depend on the values of the scalar fields at infinity. Once higher curvature terms are included in the action, the black hole entropy is no longer given by one quarter of the area of the event horizon[^4] but is given by the surface charge of the Killing vector field which becomes null on the horizon [@Wald:1993nt]. When evaluating the surface charge for supersymmetric black holes in $N=2$ supergravity, one sees that the entropy is given by the sum of two symplectic functions of the charges [@LopesCardoso:1998wt]. While the first term is the area of the horizon divided by 4, the second term depends only on the couplings of the higher derivative terms. Therefore the black hole entropy is modified in two ways: first through the modification of the area itself, second by the deviation from the area law. The microscopic state degeneracy[@Maldacena:1997de; @Vafa:1997gr] agrees with black hole entropy if and only if both corrections are taken into account [@LopesCardoso:1998wt]. Black Hole Partition Functions ============================== If one performs a partial Legendre transformation of the black hole entropy, which replaces the electric charges by the associated electrostatic potentials, one obtains the imaginary part of the ‘generalized prepotential’ [@Ooguri:2004zv]. This is a power series in the Weyl multiplet which has as its coefficients the prepotential (determining the two-derivative couplings) and the coupling functions of the higher derivative terms. By the relation between couplings in the effective action and the topological string, this function is proportional to the real part of the (holomorphic) free energy of the topologically twisted type-II string. This suggests to interprete the imaginary part of the generalized prepotential as the free energy of the black hole, and one obtains the ‘OSV-relation’ [@Ooguri:2004zv] $Z_{\rm BH} = \|Z_{\rm top}\|^2$, which relates the black hole partition function (exponential of the free energy) to the partition function of the topological string. However, many details of this proposal need to be made more precise. One is whether the relation is meant to be an exact statement (strong version) or as an asymptotic statement in the limit of large charges, which corresponds to the semi-classical limit (weak version). Before reviewing the evidence supporting the weak version, we need to address another point. By definition, the black hole free energy is a function of the magnetic charges and of the electrostatic potential. Thermodynamically this corresponds to a mixed ensemble, where the magnetic charges have been fixed, while electric charges fluctuate and the corresponding chemical potentials are fixed [@Ooguri:2004zv]. This implies that a fundamental property, namely covariance with respect to symplectic transformations is not manifest. As a consequence, it is not clear whether the proposal is compatible with string dualities. In fact, discrepancies between the actual microscopic state degeneracy and the state degeneracy predicted by the OSV conjecture show that the OSV-relation must be modified[@Shih:2005he; @LopesCardoso:2006bg]. A natural way of deriving the modification is based on the observation that the full Legendre transformation of the black hole entropy, where both electric and magnetic charges are replaced by the corresponding potentials has a natural meaning: the resulting function is a Hesse potential for the metric on the scalar manifold [@LopesCardoso:2006bg]. Moreover, the relations between entropy, free energy (mixed ensemble), Hesse potential and attractor equations can be formulated in terms of a variational principle [@Behrndt:1996jn; @LopesCardoso:2006bg]. This suggests to interprete the Hesse potential as the free energy of the black hole, but now with respect to a canonical instead of a mixed ensemble. One can show that this proposal leads to a specific correction factor in the OSV-relation. Explicit tests can be performed in compactifications with $N=4$ supersymmetry, which can be treated within the $N=2$ formalism explained in this article [@LopesCardoso:1999ur]. Subleading corrections to the state degeneracy have been computed [@Dijkgraaf:1996it; @LopesCardoso:2004xf; @Jatkar:2005bh] and the result agrees with the canonical black hole partition function proposed in [@LopesCardoso:2006bg] in the semi-classical limit. The agreement is impressive as it involves an infinite series of non-perturbative corrections to the effective action.[^5] The precise relation between the canonical black hole partition function and the topological string remains to be clarified. [10]{} B. de Wit and A. Van Proeyen, [Nucl. Phys.]{} [B245]{}, p. 89 (1984). B. de Wit, P. G. Lauwers and A. Van Proeyen, [Nucl. Phys.]{} [B255]{}, p. 569 (1985). M. Bershadsky, S. Cecotti, H. Ooguri and C. Vafa, [Commun. Math. Phys.]{} [165]{}, 311 (1994). I. Antoniadis, E. Gava, K. S. Narain and T. R. Taylor, [Nucl. Phys.]{} [ B413]{}, 162 (1994). S. Ferrara, R. Kallosh and A. Strominger, [Phys. Rev.]{} [D52]{}, 5412 (1995). G. Lopes Cardoso, B. de Wit and T. Mohaupt, [Phys. Lett.]{} [B451]{}, 309 (1999). G. Lopes Cardoso, B. de Wit, J. Kappeli and T. Mohaupt, [JHEP]{} [12]{}, p. 019 (2000). R. M. Wald, [Phys. Rev.]{} [D48]{}, 3427 (1993). J. M. Maldacena, A. Strominger and E. Witten, [JHEP]{} [12]{}, p. 002 (1997). C. Vafa, [Adv. Theor. Math. Phys.]{} [2]{}, 207 (1998). H. Ooguri, A. Strominger and C. Vafa, [Phys. Rev.]{} [D70]{}, p. 106007 (2004). D. Shih and X. Yin, [JHEP]{} [04]{}, p. 034 (2006). G. Lopes Cardoso, B. de Wit, J. Kappeli and T. Mohaupt, [JHEP]{} [03]{}, p. 074 (2006). K. Behrndt [et al.]{}, [Nucl. Phys.]{} [B488]{}, 236 (1997). G. Lopes Cardoso, B. de Wit and T. Mohaupt, [Nucl. Phys.]{} [B567]{}, 87 (2000). R. Dijkgraaf, E. P. Verlinde and H. L. Verlinde, [Nucl. Phys.]{} [B484]{}, 543 (1997). G. Lopes Cardoso, B. de Wit, J. Kappeli and T. Mohaupt, [JHEP]{} [12]{}, p. 075 (2004). D. P. Jatkar and A. Sen, [JHEP]{} [04]{}, p. 018 (2006). [^1]: together with further matter multiplets which are irrelevant for our purposes. [^2]: If the supergravity action is the low energy effective action of a string compactification, then string dualities, such as S-duality and T-duality, are embedded into the symplectic group. [^3]: We only consider spherically symmetric solutions here. [^4]: We are using Planckian units. [^5]: These corrections are world-sheet instantons from the point of view of the type-II string but space-time instantons for the dual heterotic string.	{ "pile_set_name": "ArXiv" }
--- abstract: 'All articles [must]{} contain an abstract. This document describes the preparation of a conference paper to be published in using and the class file. The abstract text should be formatted using 10 point font and indented 25 mm from the left margin. Leave 10 mm space after the abstract before you begin the main text of your article. The text of your article should start on the same page as the abstract. The abstract follows the addresses and should give readers concise information about the content of the article and indicate the main results obtained and conclusions drawn. As the abstract is not part of the text it should be complete in itself; no table numbers, figure numbers, references or displayed mathematical expressions should be included. It should be suitable for direct inclusion in abstracting services and should not normally exceed 200 words. The abstract should generally be restricted to a single paragraph. Since contemporary information-retrieval systems rely heavily on the content of titles and abstracts to identify relevant articles in literature searches, great care should be taken in constructing both.' address: 'Production Editor, , , Dirac House, Temple Back, Bristol BS1 6BE, UK' author: - Jacky Mucklow title: Preparing a paper using for publication in --- Introduction ============ These guidelines show how to prepare articles for publication in using LaTeX so they can be published quickly and accurately. Articles will be refereed by the s but the accepted PDF will be published with no editing, proofreading or changes to layout. It is, therefore, the author’s responsibility to ensure that the content and layout are correct. This document has been prepared using so serves as a sample document. The class file and accompanying documentation are available from `http://jpcs.iop.org`. Preparing your paper ==================== `jpconf` requires and can be used with other package files such as those loading the AMS extension fonts `msam` and `msbm` (these fonts provide the blackboard bold alphabet and various extra maths symbols as well as symbols useful in figure captions); an extra style file `iopams.sty` is provided to load these packages and provide extra definitions for bold Greek letters. Headers, footers and page numbers --------------------------------- Authors should [not]{} add headers, footers or page numbers to the pages of their article—they will be added by as part of the production process. [ ]{}package options -------------------- The class file has two options ‘a4paper’ and ‘letterpaper’: \documentclass[a4paper]{jpconf} or \documentclass[letterpaper]{jpconf} [@\[7]{}[l]{}]{} Option&Description\ `a4paper`&Set the paper size and margins for A4 paper.\ `letterpaper`&Set the paper size and margins for US letter paper.\ The default paper size is A4 (i.e., the default option is [a4paper]{}) but this can be changed to Letter by using `\documentclass[letterpaper]{jpconf}`. It is essential that you do not put macros into the text which alter the page dimensions. The title, authors, addresses and abstract ========================================== The code for setting the title page information is slightly different from the normal default in LaTeX but please follow these instructions as carefully as possible so all articles within a conference have the same style to the title page. The title is set in bold unjustified type using the command `\title{#1}`, where `#1` is the title of the article. The first letter of the title should be capitalized with the rest in lower case. The next information required is the list of all authors’ names followed by the affiliations. For the authors’ names type `\author{#1}`, where `#1` is the list of all authors’ names. The style for the names is initials then surname, with a comma after all but the last two names, which are separated by ‘and’. Initials should [not]{} have full stops. First names may be used if desired. The command `\maketitle` is not required. The addresses of the authors’ affiliations follow the list of authors. Each address should be set by using `\address{#1}` with the address as the single parameter in braces. If there is more than one address then a superscripted number, followed by a space, should come at the start of each address. In this case each author should also have a superscripted number or numbers following their name to indicate which address is the appropriate one for them. Please also provide e-mail addresses for any or all of the authors using an `\ead{#1}` command after the last address. `\ead{#1}` provides the text Email: so `#1` is just the e-mail address or a list of emails. The abstract follows the addresses and should give readers concise information about the content of the article and should not normally exceed 200 words. [All articles must include an abstract]{}. To indicate the start of the abstract type `\begin{abstract}` followed by the text of the abstract. The abstract should normally be restricted to a single paragraph and is terminated by the command `\end{abstract}` Sample coding for the start of an article {#startsample} ----------------------------------------- The code for the start of a title page of a typical paper might read: \title{The anomalous magnetic moment of the neutrino and its relation to the solar neutrino problem} \author{P J Smith$^1$, T M Collins$^2$, R J Jones$^{3,}$\footnote[4]{Present address: Department of Physics, University of Bristol, Tyndalls Park Road, Bristol BS8 1TS, UK.} and Janet Williams$^3$} \address{$^1$ Mathematics Faculty, Open University, Milton Keynes MK7~6AA, UK} \address{$^2$ Department of Mathematics, Imperial College, Prince Consort Road, London SW7~2BZ, UK} \address{$^3$ Department of Computer Science, University College London, Gower Street, London WC1E~6BT, UK} \ead{[email protected]} \begin{abstract} The abstract appears here. \end{abstract} The text ======== The text of the article should should be produced using standard LaTeX formatting. Articles may be divided into sections and subsections, but the length limit provided by the should be adhered to. Acknowledgments --------------- Authors wishing to acknowledge assistance or encouragement from colleagues, special work by technical staff or financial support from organizations should do so in an unnumbered Acknowledgments section immediately following the last numbered section of the paper. The command `\ack` sets the acknowledgments heading as an unnumbered section. Appendices ---------- Technical detail that it is necessary to include, but that interrupts the flow of the article, may be consigned to an appendix. Any appendices should be included at the end of the main text of the paper, after the acknowledgments section (if any) but before the reference list. If there are two or more appendices they will be called Appendix A, Appendix B, etc. Numbered equations will be in the form (A.1), (A.2), etc, figures will appear as figure A1, figure B1, etc and tables as table A1, table B1, etc. The command `\appendix` is used to signify the start of the appendixes. Thereafter `\section`, `\subsection`, etc, will give headings appropriate for an appendix. To obtain a simple heading of ‘Appendix’ use the code `\section{Appendix}`. If it contains numbered equations, figures or tables the command `\appendix` should precede it and `\setcounter{section}{1}` must follow it. References ========== In the online version of references will be linked to their original source or to the article within a secondary service such as INSPEC or ChemPort wherever possible. To facilitate this linking extra care should be taken when preparing reference lists. Two different styles of referencing are in common use: the Harvard alphabetical system and the Vancouver numerical system. For , the Vancouver numerical system is preferred but authors should use the Harvard alphabetical system if they wish to do so. In the numerical system references are numbered sequentially throughout the text within square brackets, like this \[2\], and one number can be used to designate several references. Using ------ We highly recommend the [***iopart-num]{} package by Mark A Caprio [@iopartnum], which is included with this documentation. Reference lists --------------- A complete reference should provide the reader with enough information to locate the article concerned, whether published in print or electronic form, and should, depending on the type of reference, consist of: - name(s) and initials; - date published; - title of journal, book or other publication; - titles of journal articles may also be included (optional); - volume number; - editors, if any; - town of publication and publisher in parentheses for [books]{}; - the page numbers. Up to ten authors may be given in a particular reference; where there are more than ten only the first should be given followed by ‘[et al]{}’. If an author is unsure of a particular journal’s abbreviated title it is best to leave the title in full. The terms [loc. cit. ]{}and [ibid. ]{}should not be used. Unpublished conferences and reports should generally not be included in the reference list and articles in the course of publication should be entered only if the journal of publication is known. A thesis submitted for a higher degree may be included in the reference list if it has not been superseded by a published paper and is available through a library; sufficient information should be given for it to be traced readily. Formatting reference lists -------------------------- Numeric reference lists should contain the references within an unnumbered section (such as `\section{References}`). The reference list itself is started by the code `\begin{thebibliography}{<num>}`, where `<num>` is the largest number in the reference list and is completed by `\end{thebibliography}`. Each reference starts with `\bibitem{<label>}`, where ‘label’ is the label used for cross-referencing. Each `\bibitem` should only contain a reference to a single article (or a single article and a preprint reference to the same article). When one number actually covers a group of two or more references to different articles, `\nonum` should replace `\bibitem{<label>}` at the start of each reference in the group after the first. For an alphabetic reference list use `\begin{thereferences}` ... `\end{thereferences}` instead of the ‘thebibliography’ environment and each reference can be start with just `\item` instead of `\bibitem{label}` as cross referencing is less useful for alphabetic references. References to printed journal articles -------------------------------------- A normal reference to a journal article contains three changes of font (see table \[jfonts\]) and is constructed as follows: - the authors should be in the form surname (with only the first letter capitalized) followed by the initials with no periods after the initials. Authors should be separated by a comma except for the last two which should be separated by ‘and’ with no comma preceding it; - the article title (if given) should be in lower case letters, except for an initial capital, and should follow the date; - the journal title is in italic and is abbreviated. If a journal has several parts denoted by different letters the part letter should be inserted after the journal in Roman type, e.g. [Phys. Rev.]{} A; - the volume number should be in bold type; - both the initial and final page numbers should be given where possible. The final page number should be in the shortest possible form and separated from the initial page number by an en rule ‘– ’, e.g. 1203–14, i.e. the numbers ‘12’ are not repeated. A typical (numerical) reference list might begin [9]{} Strite S and Morkoc H 1992 [J. Vac. Sci. Technol.]{} B [10]{} 1237 Jain S C, Willander M, Narayan J and van Overstraeten R 2000 [J. Appl. Phys]{}. [87]{} 965 Nakamura S, Senoh M, Nagahama S, Iwase N, Yamada T, Matsushita T, Kiyoku H and Sugimoto Y 1996 [Japan. J. Appl. Phys.]{} [35]{} L74 Akasaki I, Sota S, Sakai H, Tanaka T, Koike M and Amano H 1996 [Electron. Lett.]{} [32]{} 1105 O’Leary S K, Foutz B E, Shur M S, Bhapkar U V and Eastman L F 1998 [J. Appl. Phys.]{} [83]{} 826 Jenkins D W and Dow J D 1989 [Phys. Rev.]{} B [39]{} 3317 which would be obtained by typing \begin{\thebibliography}{9} \item Strite S and Morkoc H 1992 {\it J. Vac. Sci. Technol.} B {\bf 10} 1237 \item Jain S C, Willander M, Narayan J and van Overstraeten R 2000 {\it J. Appl. Phys}. {\bf 87} 965 \item Nakamura S, Senoh M, Nagahama S, Iwase N, Yamada T, Matsushita T, Kiyoku H and Sugimoto Y 1996 {\it Japan. J. Appl. Phys.} {\bf 35} L74 \item Akasaki I, Sota S, Sakai H, Tanaka T, Koike M and Amano H 1996 {\it Electron. Lett.} {\bf 32} 1105 \item O'Leary S K, Foutz B E, Shur M S, Bhapkar U V and Eastman L F 1998 {\it J. Appl. Phys.} {\bf 83} 826 \item Jenkins D W and Dow J D 1989 {\it Phys. Rev.} B {\bf 39} 3317 \end{\thebibliography} [@l\[15]{}[l]{}]{} Element&Style\ Authors&Roman type\ Date&Roman type\ Article title (optional)&Roman type\ Journal title&Italic type\ Volume number&Bold type\ Page numbers&Roman type\ References to articles ----------------------- Each conference proceeding published in will be a separate [volume]{}; references should follow the style for conventional printed journals. For example: Douglas G 2004 J. Phys.: Conf. Series* 1 23–36 References to preprints ----------------------- For preprints there are two distinct cases: 1. Where the article has been published in a journal and the preprint is supplementary reference information. In this case it should be presented as: 2. Kunze K 2003 T-duality and Penrose limits of spatially homogeneous and inhomogeneous cosmologies [Phys. Rev.]{} D [68]{} 063517 ([Preprint]{} gr-qc/0303038) 3. Where the only reference available is the preprint. In this case it should be presented as 4. Milson R, Coley A, Pravda V and Pravdova A 2004 Alignment and algebraically special tensors [Preprint]{} gr-qc/0401010 References to electronic-only journals -------------------------------------- In general article numbers are given, and no page ranges, as most electronic-only journals start each article on page 1. - For [New Journal of Physics]{} (article number may have from one to three digits) - Fischer R 2004 Bayesian group analysis of plasma-enhanced chemical vapour deposition data [New. J. Phys.]{} [6]{} 25 - For SISSA journals the volume is divided into monthly issues and these form part of the article number - Horowitz G T and Maldacena J 2004 The black hole final state [J. High Energy Phys.]{} JHEP02(2004)008 - Bentivegna E, Bonanno A and Reuter M 2004 Confronting the IR fixed point cosmology with high-redshift observations [J. Cosmol. Astropart. Phys.]{} JCAP01(2004)001 References to books, conference proceedings and reports ------------------------------------------------------- References to books, proceedings and reports are similar to journal references, but have only two changes of font (see table \[book\]). [@l\[15]{}[l]{}]{} Element&Style\ Authors&Roman type\ Date&Roman type\ Book title (optional)&Italic type\ Editors&Roman type\ Place (city, town etc) of publication&Roman type\ Publisher&Roman type\ Volume&Roman type\ Page numbers&Roman type\ Points to note are: - Book titles are in italic and should be spelt out in full with initial capital letters for all except minor words. Words such as Proceedings, Symposium, International, Conference, Second, etc should be abbreviated to [Proc.]{}, [Symp.]{}, [Int.]{}, [Conf.]{}, [2nd]{}, respectively, but the rest of the title should be given in full, followed by the date of the conference and the town or city where the conference was held. For Laboratory Reports the Laboratory should be spelt out wherever possible, e.g. [Argonne National Laboratory Report]{}. - The volume number, for example vol 2, should be followed by the editors, if any, in a form such as ‘ed A J Smith and P R Jones’. Use [et al]{} if there are more than two editors. Next comes the town of publication and publisher, within brackets and separated by a colon, and finally the page numbers preceded by p if only one number is given or pp if both the initial and final numbers are given. Examples taken from published papers: Kurata M 1982 [Numerical Analysis for Semiconductor Devices]{} (Lexington, MA: Heath) Selberherr S 1984 [Analysis and Simulation of Semiconductor Devices]{} (Berlin: Springer) Sze S M 1969 [Physics of Semiconductor Devices]{} (New York: Wiley-Interscience) Dorman L I 1975 [Variations of Galactic Cosmic Rays]{} (Moscow: Moscow State University Press) p 103 Caplar R and Kulisic P 1973 [Proc. Int. Conf. on Nuclear Physics (Munich)]{} vol 1 (Amsterdam: North-Holland/American Elsevier) p 517 Cheng G X 2001 [Raman and Brillouin Scattering-Principles and Applications]{} (Beijing: Scientific) Szytula A and Leciejewicz J 1989 [Handbook on the Physics and Chemistry of Rare Earths]{} vol 12, ed K A Gschneidner Jr and L Erwin (Amsterdam: Elsevier) p 133 Kuhn T 1998 [Density matrix theory of coherent ultrafast dynamics Theory of Transport Properties of Semiconductor Nanostructures]{} (Electronic Materials vol 4) ed E Schöll (London: Chapman and Hall) chapter 6 pp 173–214 Tables and table captions ========================= Tables should be numbered serially and referred to in the text by number (table 1, etc, [rather than]{} tab. 1). Each table should be a float and be positioned within the text at the most convenient place near to where it is first mentioned in the text. It should have an explanatory caption which should be as concise as possible. The basic table format ---------------------- The standard form for a table is: \begin{table} \caption{\label{label}Table caption.} \begin{center} \begin{tabular}{llll} \br Head 1&Head 2&Head 3&Head 4\\ \mr 1.1&1.2&1.3&1.4\\ 2.1&2.2&2.3&2.4\\ \br \end{tabular} \end{center} \end{table} The above code produces table \[ex\]. [llll]{} Head 1&Head 2&Head 3&Head 4\ 1.1&1.2&1.3&1.4\ 2.1&2.2&2.3&2.4\ Points to note are: 1. The caption comes before the table. 2. The normal style is for tables to be centred in the same way as equations. This is accomplished by using `\begin{center}` … `\end{center}`. 3. The default alignment of columns should be aligned left. 4. Tables should have only horizontal rules and no vertical ones. The rules at the top and bottom are thicker than internal rules and are set with `\br` (bold rule). The rule separating the headings from the entries is set with `\mr` (medium rule). These commands do not need a following double backslash. 5. Numbers in columns should be aligned as appropriate, usually on the decimal point; to help do this a control sequence `\lineup` has been defined which sets `\0` equal to a space the size of a digit, `\m` to be a space the width of a minus sign, and `\-` to be a left overlapping minus sign. `\-` is for use in text mode while the other two commands may be used in maths or text. (`\lineup` should only be used within a table environment after the caption so that `\-` has its normal meaning elsewhere.) See table \[tabone\] for an example of a table where `\lineup` has been used. [\[7]{}[l]{}]{} $\0\0A$&$B$&$C$&$D$&$E$&$F$&$\0G$ 23.5&60 &0.53&$-20.2$&$-0.22$ &1.7&14.539.7&60&0.74&$-51.9$&$-0.208$&47.2 &146 123.7 &0 &0.75&$-57.2$&— &— &— 3241.56 &60 &0.60&$-48.1$&$-0.29$ &41 &15 Figures and figure captions =========================== Figures must be included in the source code of an article at the appropriate place in the text not grouped together at the end. Each figure should have a brief caption describing it and, if necessary, interpreting the various lines and symbols on the figure. As much lettering as possible should be removed from the figure itself and included in the caption. If a figure has parts, these should be labelled ($a$), ($b$), ($c$), etc. gives the definitions for describing symbols and lines often used within figure captions (more symbols are available when using the optional packages loading the AMS extension fonts). [lllll]{} Control sequence&Output&&Control sequence&Output\ `\dotted`&&&`\opencircle`&\ `\dashed`&&&`\opentriangle`&\ `\broken`&&&`\opentriangledown`&\ `\longbroken`&&&`\fullsquare`&\ `\chain`&&&`\opensquare`&\ `\dashddot`&&&`\fullcircle`&\ `\full`&&&`\opendiamond`&\ Authors should try and use the space allocated to them as economically as possible. At times it may be convenient to put two figures side by side or the caption at the side of a figure. To put figures side by side, within a figure environment, put each figure and its caption into a minipage with an appropriate width (e.g. 3in or 18pc if the figures are of equal size) and then separate the figures slightly by adding some horizontal space between the two minipages (e.g. `\hspace{.2in}` or `\hspace{1.5pc}`. To get the caption at the side of the figure add the small horizontal space after the `\includegraphics` command and then put the `\caption` within a minipage of the appropriate width aligned bottom, i.e. `\begin{minipage}[b]{3in}` etc (see code in this file used to generate figures 1–3). Note that it may be necessary to adjust the size of the figures (using optional arguments to `\includegraphics`, for instance `[width=3in]`) to get you article to fit within your page allowance or to obtain good page breaks. ![\[label\]Figure caption for second of two sided figures.](name.eps){width="14pc"} ![\[label\]Figure caption for second of two sided figures.](name.eps){width="14pc"} ![\[label\]Figure caption for a narrow figure where the caption is put at the side of the figure.](name.eps){width="14pc"} Using the graphicx package figures can be included using code such as: \begin{figure} \begin{center} \includegraphics{file.eps} \end{center} \caption{\label{label}Figure caption} \end{figure} References {#references-1 .unnumbered} ========== [9]{} IOP Publishing is to grateful Mark A Caprio, Center for Theoretical Physics, Yale University, for permission to include the [iopart-num]{} package (version 2.0, December 21, 2006) with this documentation. Updates and new releases of [iopart-num]{} can be found on `www.ctan.org` (CTAN).	{ "pile_set_name": "ArXiv" }
--- abstract: '$\pi^+\pi^-$ transitions of heavy quarkonia, especially the $\Upsilon(3S)\to\Upsilon(1S)\pi^+\pi^-$ decay process, are revisited. In the framework of the Chiral Unitary Theory (ChUT), the $S$ wave $\pi\pi$ final state interaction (FSI) is included. It is found that when an additional intermediate state with $J^P=1^+$ and $I=1$ is introduced, not only the $\pi\pi$ invariant mass spectrum and the $cos\theta_\pi^$ distribution in the $\Upsilon(3S)\to\Upsilon(1S)\pi^+\pi^-$ process can simultaneously be well-explained, but also a consistent description for other bottomonia $\pi^+\pi^-$ transitions can be obtained. As a consequence, the mass and the width of the intermediate state are predicted. From the quark content analysis, this state should be a $b\bar bq\bar q$ state.' author: - \| F.-K. Guo$^{1,6}$[^1], P.-N. Shen$^{2,1,4,5}$, H.-C. Chiang$^{3,1}$, R.-G. Ping$^{2,1}$\ [$^1$Institute of High Energy Physics, Chinese Academy of Sciences,]{}\ [P.O.Box 918(4), Beijing 100049, China]{}\ [$^2$CCAST(World Lab.), P.O.Box 8730, Beijing 100080, China]{}\ [$^3$South-west Normal University, Chongqing 400715, China]{}\ [$^4$Institute of Theoretical Physics, Chinese Academy of Sciences, P.O.Box 2735, China]{}\ [$^5$Center of Theoretical Nuclear Physics, National Laboratory of Heavy Ion Accelerator,]{}\ [Lanzhou 730000, China]{}\ [$^6$Graduate School of the Chinese Academy of Sciences, Beijing 100049, China]{} --- =-2.5cm =-1cm =1.0cm [PACS]{} numbers: 13.25.Gv, 12.39.Fe, 12.39.Mk\ Introduction ============ In recent years, more and more decay data of heavy quarkonia have been accumulated, and new information on hadron physics has been extracted. Many investigations along this line, for instance, the decay properties of the $\Psi(2S)\to J/\Psi\pi^+\pi^-$, $\Upsilon(2S)\to\Upsilon(1S)\pi^+\pi^-$, $\Upsilon(3S)\to\Upsilon(2S)\pi^+\pi^-$, and $\Upsilon(3S)\to\Upsilon(1S)\pi^+\pi^-$ processes have been carried out. A commonly used method for such studies is the QCD multipole expansion method proposed by Gottfried [gott78]{} and further developed by others [yan80,vz80,ns81,ky81,mo89,zk91]{}. It was shown that although most data of the mentioned processes can be well-reproduced, the $\pi\pi$ invariant mass spectrum and the angular distribution of the $\Upsilon(3S)\to\Upsilon(1S)\pi^+\pi^-$ decay cannot satisfactorily be described. Phenomenological models [@lt88; @bc75; @mu97; @ywz99; @ue03] are also used in such studies. For instance, in Ref. [@mu97], Mannel et al. constructed an effective Lagrangian on the basis of the chiral perturbation theory (ChPT) and the heavy quark non-relativistic expansion. Under the approximation in the limits of the chiral symmetry and the heavy quark mass [@ckk94], the measured $\pi\pi$ invariant mass spectra of the $\Psi(2S)\to J/\Psi\pi^+\pi^-$, $\Upsilon(3S)\to\Upsilon(2S)\pi^+\pi^-$ and $ \Upsilon(2S)\to\Upsilon(1S)\pi^+\pi^-$ decay processes were very well fitted, but not of the decay $\Upsilon(3S)\to\Upsilon(1S)\pi^+\pi^-$. In order to explain the data of $d\Gamma(\Upsilon(3S)\to \Upsilon(1S)\pi^+\pi^-)/dm_{\pi\pi}$, the $\pi\pi$ S-wave FSI was included by using a parameterization. However, although the coupling constant ratios $(g_2/g_1)_{b \bar{b}}$ in the $\Upsilon(2S)\to\Upsilon(1S)\pi^+\pi^-$ and $\Upsilon(3S)\to\Upsilon(2S)\pi^+\pi^-$ decays are approximately equal to each other, they are much smaller than that in the $\Upsilon(3S)\to\Upsilon(1S)\pi^+\pi^-$ decay. The later one is about 10 times larger than the former. It seems unnatural. Moreover, M.-L. Yan et al. [@ywz99] pointed out that the parametrization of the S-wave FSI there was not properly carried out because in the $g_2$ term in the amplitude, there are also $D$ wave components. In Ref. [@ue03], according to a unitarized chiral theory, an $S$ wave effective Lagrangian and an effective scalar form factor were adopted. As a result, the $\pi\pi$ invariant mass spectra of heavy quarkonium decays can be reproduced, but the $\cos\theta_\pi^$ distribution still cannot be explained. In the $\Upsilon(3S)\to\Upsilon(1S)\pi^+\pi^-$ decay, the two-peak structure of the $\pi^+\pi^-$ invariant mass spectrum has been considered as a consequence of $\pi\pi$ FSI [@bdm89; @ck93; @ckk94; @gmp04; @lr02] or the additional contribution from the $D$ wave component of $\Upsilon(3S)$ [@ckk93]. Ignoring the contribution from the higher order pion momentum, Chakravarty [et al.]{} [@ckk94] explained the $\pi\pi$ invariant mass spectrum of $\Upsilon(3S)\to\Upsilon(1S)\pi^+\pi^-$ with $\chi^2/N_{d.f.}$=11.0/7 or C.L.=14.0%, but only quantitatively gave the $\cos\theta_\pi^$ distribution. Gallegos et al. [@gmp04] parametrized a more generalized amplitude in which both $S$ and $D$ wave contributions are included by fitting to the invariant mass distributions and the angular distributions of the decays mentioned at the beginning of this section. It is clear that the result of the parametrization should be consistent with the existing $\pi\pi$ scattering data. Whether this condition has been satisfied in that calculation remains a question. Lähde et al. [@lr02] showed that in order to better describe the $\pi\pi$ invariant mass spectra, the contribution of the pion re-scattering should be small in the $\Psi(2S)\to J/\Psi\pi^+\pi^-$ or $\Upsilon(2S)\to \Upsilon(1S)\pi^+\pi^-$ decay, but dominant in the $\Upsilon(3S)\to\Upsilon(1S)\pi^+\pi^-$ decay. Why it is so is still a puzzle. On the other hand, by introducing a $b\bar{b}q\bar{q}$ resonance with $J^P=1^+$ and mass of 10.4-10.8 GeV, Anisovich et al. [@absz95] explained the $\pi^+ \pi^-$ invariant mass spectra of above mentioned bottomonium decays, but not the angular distribution of $\Upsilon(3S)\to\Upsilon(1S)\pi^+\pi^-$. To treat meson-meson $S$ wave FSI properly, a chiral nonperturbative approach, called chiral unitary theory (ChUT), was recently proposed by Oller and Oset [@oo97] and later developed by themselves and others [@oop98; @oo98; @oond; @ol98; @mo01; @ue03; @lov04; @prov03; @rpoc04; @lo98; @oo01; @ol05] (for details, refer to the review article Ref. [@oor00]). In this theory [@oo97], the coupled channel Bethe-Salpeter equations (BSE), in which the lowest order amplitudes in the ChPT are employed as the kernel, are solved to resum the contributions from the s-channel loops of the re-scattering between pseudoscalar mesons. By properly choosing the three-momentum cutoff, the only free parameter in the theory, ChUT can well-describe the data of the $S$ wave meson-meson interaction up to $\sqrt{s}\simeq1.2GeV$ [@oo97] which is much larger than the energy in the region where the standard ChPT is still valid, and can dynamically generate the scalar resonances $\sigma$ ($f_0(600)$), $ f_0(980)$ and $a_0(980)$ [@oo97]. In this work, we adopt an amplitude used in Ref. [@mu97], which includes both $S$ and $D$ wave contributions, and consider the $S$ wave $\pi\pi$ FSI in the framework of ChUT to study the $\Psi(2S)\to J/\Psi\pi^+\pi^-$, $\Upsilon(2S)\to\Upsilon(1S)\pi^+\pi^-$, $\Upsilon(3S)\to\Upsilon(2S)\pi^+\pi^-$ and $\Upsilon(3S)\to\Upsilon(1S)\pi^+\pi^-$ decays. The paper is organized as follows: In Section \[sec:for\], the effective Lagrangian and ChUT are briefly introduced. In terms of the $t$-matrix written in Section \[sec:for\], the $\Psi(2S)\to J/\Psi\pi^+\pi^-$ decay is discussed in Section \[sec:psi\]. Section \[sec:upsilon\] is dedicated to the bottomonium $\pi^+ \pi^-$ transitions, and a brief summary is given in Section \[sec:sum\]. Brief formalism {#sec:for} =============== In the heavy quarkonium $\pi^+ \pi^-$ transition process, the Lagrangian in the lowest order, which appropriately incorporates the chiral expansion with the heavy quark expansion, can be written as [mu97]{} $${\cal L}= {\cal L}_0+{\cal L}_{S.B.}$$ $$\begin{aligned} {\cal L}_0 &\!\!\!\!=&\!\!\!\! g_1A_{\mu}^{(v)}B^{(v)\mu}Tr[(\partial_{\nu}U)(\partial^{\nu}U)^\dag] +g_2A_{\mu}^{(v)}B^{(v)\mu}Tr[(v\cdot{\partial}U)(v\cdot{\partial}U)^\dag] \nonumber \\ &\!\!\!\!+&\!\!\!\! g_3A_{\mu}^{(v)}B_{\nu}^{(v)}Tr[(\partial^{\mu}U)(\partial^{\nu}U)^\dag +(\partial^{\mu}U)^\dag(\partial^{\nu}U)]+h.c.\end{aligned}$$ $$\begin{aligned} {\cal L}_{S.B.} &\!\!\!\!=&\!\!\!\! g_4A_{\mu}^{(v)}B^{(v)\mu}Tr[M(U+U^{\dag}-2)] \nonumber \\ &\!\!\!\!+&\!\!\!\! ig_{^{\prime}}\varepsilon^{\mu\nu\alpha\beta}[v_{\mu}A_{\nu}^{(v)}\partial_{\alpha}B_{\beta}^{(v)} -(\partial_{\mu}A_{\nu}^{(v)})v_{\alpha}B_{\beta}^{(v)}]Tr[M(U-U^{\dag})]+h.c.\end{aligned}$$ where $g_i$ denotes the coupling constants, $U$ is a $3{\times}3$ matrix that contains the pseudoscalar Goldstone fields, and $M= diag\{m_u, m_d, m_s\}$ is the quark mass matrix with $m_u$, $m_d$, $m_s$ being the masses of current quarks $u$, $d$ and $s$, respectively. $A_\mu^{(v)}$ and $B_\mu^{(v)}$ are the fields of the initial and final states of heavy vector quarkonia, respectively, and $v$ is the velocity vector of $A$. The tree diagram amplitude for the decay of a vector meson into two pseudoscalar mesons and one vector meson in the rest frame of decaying particle can be expressed as [@ck93; @ckk94; @mu97] $$\label{eq:mannel} t=-\frac{4}{f^{2}_\pi}[(g_1p_1\cdot p_2+g_{2}p_1^0p_2^0 +g_3m_\pi^2)\varepsilon^\cdot\varepsilon^{^{\prime}}+g_4(p_{1\mu}p_{2\nu}+p_{1\nu}p_{2\mu}) \varepsilon^{\mu}\varepsilon^{^{\prime}\nu}]$$ where $f_\pi=93MeV$ is the decay constant of pion, $p_1$ and $p_2$ are the four-momenta of $\pi^+$ and $\pi^-$, respectively, $p_1^0$ and $p_2^0$ denote the energies of $\pi^+$ and $\pi^-$ in the lab frame, and $\varepsilon $ and $\varepsilon^{^{\prime}}$ are the polarization vectors of the heavy quarkonia, respectively. It can be verified by the CLEO data [@ckk94] that by considering the chiral symmetry breaking scale and the heavy quark mass, the contribution from the last term ($g_4$-term) is strongly suppressed [@mu97]. Thus, the $g_4$-term in Eq. (\[eq:mannel\]) can be ignored and the amplitude can further be written as $$\begin{aligned} \label{eq:v0} V_{0}=-\frac{4}{f^{2}_\pi}(g_1p_1\cdot p_2+g_{2}p_1^0p_2^0 +g_3m_\pi^2)\varepsilon^\cdot\varepsilon^{^{\prime}}.\end{aligned}$$ It is noted that the $D$ wave component exists in the $g_2$ term [@bc75; @ywz99]. Under Lorentz transformation, $p_1^0$ and $p_2^0$ can be expressed as the functions of the momenta of pions in the center of mass (c.m.) frame of the $\pi\pi$ system: $$\begin{aligned} p_1^0 &\!\!\!\!=&\!\!\!\! \frac{1}{\sqrt{1-{\bm\beta}^2}}(p_1^{0}+\|{\bm \beta}\|\|{\mathbf p}_1^{}\|\cos\theta_{\pi}^), \\ p_2^0 &\!\!\!\!=&\!\!\!\! \frac{1}{\sqrt{1-{\bm\beta}^2}}(p_1^{0}-\|{\bm \beta}\|\|{\mathbf p}_1^{}\|\cos\theta_{\pi}^),\end{aligned}$$ where ${\bm\beta}$ is the velocity of the $\pi\pi$ system in the rest frame of the initial particle, $p_1^$=($p_1^{0}$, ${\mathbf p}_1^$) and $p_2^ $=($p_2^{0}$, ${\mathbf p}_2^$) are the four-momenta of $\pi^+$ and $\pi^-$ in the c.m. frame of the $\pi\pi$ system, respectively. So $p_1^0p_2^0$ can be decomposed as $$\begin{aligned} p_1^0p_2^0=\frac{1}{1-{\bm \beta}^2}[(p_1^{02}-\frac{{\bm \beta}^2{\mathbf p}_1^{2}}{3})P_0(\cos\theta_{\pi}^)-2{\bm \beta}^2{\mathbf p}_1^{2}P_2(\cos\theta_{\pi}^)]\end{aligned}$$ where $P_0(\cos\theta_{\pi}^)=1$ and $P_2(\cos\theta_{\pi}^)=\frac{1}{2}(\cos^2\theta_{\pi}^-\frac{1}{3})$ are the Legendre functions of the 0-th order and 2-nd order, respectively. Furthermore, the $S$ wave $\pi \pi$ FSI which is important in this energy region should properly be included into the theoretical calculation. ChUT [@oo97] is one of the suitable approaches for this job, because by using this theory, the $S$ wave $\pi -\pi $ scattering data up to 1.2GeV can be well reproduced. However, ChPT amplitudes in the $O(p^{2})$ order are adopted as the kernel of the coupled-channel BSE [@oo97], so the $D$ wave FSI cannot be included. In the decays considered, the kinematical region is below $0.9GeV$ for the decay $\Upsilon(3S)\to\Upsilon(1S)\pi^+\pi^-$ and below $0.6GeV$ for the others. One can see that they are far below the $D$ wave resonant region ($>1.2GeV$). So as a primary consideration, the $D$ wave contribution comes only from the $D$ wave terms appeared in Eq. (\[eq:v0\]). The basic diagrams for the $V^{\prime}\to VPP$ decay, where $V^{\prime}(V)$ and $P$ denote the vector and pseudoscalar mesons, respectively, are shown in Fig. \[fig:feyn1\]. In the figure, (a) represents the $V^{\prime}\to VPP $ decay without FSI, namely the tree diagram or Born term, and (b) describes the decay with $\pi\pi$ FSI. In Fig. \[fig:feyn1\](b), the $\pi\pi\to \pi^+\pi^-$ $t$-matrix described by the solid black circle is obtained by the loop resummation [@oo97], namely by a set of coupled-channel BSEs and both $\pi\pi$ and $K{\bar K}$ channels are included (for details, see Ref. [@oo97]). To factorize the $\pi\pi$ FSI from the direct $V'\to VPP$ decay part in Fig. \[fig:feyn1\](b), the on-shell approximation is adopted. Thus, only $\pi^+$, $\pi^-$, $\pi^0$ exist in the first loop which is directly linked to the $V'\to VPP$ vertex. The off-shell effects can be absorbed into the phenomenological coupling constants of the vertex. In fact, this approximation is often used in parameterizing the $S$ wave FSI with phase shift data [@ckk94]. The full $S$ wave $\pi\pi\to \pi^+\pi^-$ $t$-matrix can be expressed as $$\label{eq:fsi} \langle \pi^+\pi^- + \pi^-\pi^+ +\pi^0\pi^0\|t\|\pi^+\pi^-\rangle=2t^{I=0}_{\pi\pi,\pi\pi},$$ where $t^{I=0}_{\pi\pi,\pi\pi}$ denotes the full $S$ wave $\pi\pi\to \pi\pi$ $t$-matrix in the isospin $I=0$ channel, which is the solution of a set of on-shell coupled channel BSEs [@oo97]. Now, the $t$-matrix for the $V^{'}\to V\pi^+\pi^-$ decay can be expressed as $$t=V_0 + V_{0S}\cdot G\cdot 2t^{I=0}_{\pi\pi,\pi\pi},$$ where $V_0$ is the amplitude of the tree diagram (a), $V_{0S}$ is the $S$ wave component of $V_0$ and $G$ is the two meson loop propagator $$G=i\int\frac{d^4q}{(2\pi)^4}\frac{1}{q^2-m_{\pi}^2+i \varepsilon} \frac{1}{(p^{^{\prime}}-p-q)^2-m_{\pi}^2+i\varepsilon}. \label{eq:2loop}$$ The numerical calculation is done by introducing a three-momentum cutoff $q_{max}$. The value of the cutoff is taken as that used in [@oo97], where the $\pi\pi$ scattering data can be well reproduced up to 1.2 GeV, namely, the $\pi\pi$ FSI in our model is consistent with the $\pi\pi$ scattering data. Moreover, this cutoff is also consistent with the dimensional regularization [@oop98]. The analytic expression of the loop integral in Eq. (\[eq:2loop\]) can be given as $$\label{eq:g} G=\frac{1}{8\pi^2}\{\sigma \arctan{\frac{1}{\lambda\sigma}} - \ln[\frac{q_{max}}{m_{\pi}}(1+\lambda)]\},$$ where $\sigma=\sqrt{\frac{4m_{\pi}^2}{s}-1}$ and $\lambda=\sqrt{1+\frac{m_{\pi}^2}{q^2_{max}}}$. The differential decay width with respect to the $\pi\pi$ invariant mass and $\cos{\theta}^_\pi$ reads $$\frac{d\Gamma}{dm_{\pi\pi}dcos{\theta}^_\pi}=\frac{1}{8M^2(2\pi)^3} \overline{\sum}\sum\|t\|^2\|{\bf p_1^}\|\|{\bf p_3}\|$$ where $\overline{\sum}\sum$ describes the average over initial states and the sum over final states, and ${\bf p_3}$ is the 3-momentum of the final vector meson in the lab frame. Results for the $\Psi(2S)\to J/\Psi\pi^+\pi^-$ decay {#sec:psi} ==================================================== In the model, the parameters involved are the coupling constants $g_1$, $g_2$ and $g_3$. The values of the parameters can be determined by fitting the experimental data of the $\Psi(2S)\to J/\Psi\pi^+\pi^-$ process. It is shown that the resultant $g_3$ value is so small that we can safely take $g_3=0$. The remaining coupling constants $g_1$ and $g_2$ are obtained by fitting the total decay rate and the $\pi^+\pi^-$ invariant mass spectrum simultaneously. The decay data of the $\Psi(2S)\to J/\Psi\pi^+\pi^-$ process are taken from ref.[@bes00]. These BES data are normalized by using $\Gamma_{\Psi(2S)}= 277keV$ and $B(\Psi(2S)\to J/\Psi\pi^+\pi^-)=31.7\%$ [@pdg04]. The resultant coupling constants are $$g_1=0.106, ~~~~~~g_2/g_1=-0.319, ~~~~~~g_3/g_1=0.$$ 1.cm -1.5cm Our best fit to the $\pi^+\pi^-$ invariant mass spectrum and the $\cos\theta_{\pi}^$ distribution are shown in Fig. \[fig:psi\]. It is found that the $\pi^+\pi^-$ invariant mass spectrum is well-fitted, but the theoretical angular distribution is somewhat too flat. Note that in fitting angular distribution, we only consider $\cos\theta_{\pi}^$ from -0.8 to 0.8, because the efficiency correction to the data at large $\|\cos\theta_{\pi}^\|$ is not accurate enough [@bes00]. The deviation in angular distribution implies that the $D$ wave contribution is somehow too small. In fact, as discussed in Section \[sec:for\], in our calculation, the $D$ wave FSI is not included. It is also found that the $S$ wave FSI enhances the invariant mass spectrum considerably. It should be noted that due to $P_2(\cos\theta_{\pi}^)= 1/2(\cos^2\theta_{\pi}^-1/3)$, integrating over $\cos\theta_{\pi}^$ will result that the $D$ wave contribution is not so important in the invariant mass spectrum. However, in the angular distribution, the $\cos\theta_{\pi}^$-dependence, and consequently the $D$ wave effect, will explicitly show up. Thus, we deem that the deviation in angular distribution may be due to lack of $D$ wave FSI. In fact, it can be confirmed in the following way: In Ref.[@bes00; @mu97], without FSI, the authors can well-reproduce both the $\pi\pi$ invariant mass spectrum and the angular distribution with $g_1=0.30{\pm 0.01}$, $g_2/g_1$$=-0.35\pm0.03$ and $g_3=0$. However, with the $S$ wave FSI, in the best data fitting to the $\pi\pi$ invariant mass spectrum, the resultant $g_1$ is 0.106 which is 3 times less than that in Ref. [@mu97], while $g_2/g_1$ keeps almost the same value as that in Ref. [@mu97]. This means that the effect of the $S$ wave FSI is so large that $g_1$ has to be much smaller to explain the $\pi\pi$ invariant mass spectrum. As a consequence, the $D$ wave component is also greatly reduced. If we naively multiply a factor of 3 to the $D$ wave component, the angular distribution can also be reproduced better. Unfortunately, as mentioned above, the $D$ wave FSI can not be treated properly in the simple ChUT approach [@oo97]. The bottomonia $\pi^+\pi^-$ transitions {#sec:upsilon} ======================================= Similar to the case of $\Psi(2S)\to J/\Psi\pi^+\pi^-$, in the $\Upsilon(2S)\to\Upsilon(1S)\pi^+\pi^-$ or $\Upsilon(3S)\to\Upsilon(2S)\pi^+\pi^-$ decay, $g_3=0$ can be adopted. But for the $\Upsilon(3S)\to\Upsilon(1S)\pi^+\pi^-$ decay, the $S$ wave FSI is no longer a main contributor, and a finite value of $g_3$ is requested. With this consideration, Mannel et al. showed that the resultant values of $g_2/g_1$ for the $\Upsilon(2S)\to\Upsilon(1S)\pi^+\pi^-$ and $\Upsilon(3S)\to\Upsilon(2S)\pi^+\pi^-$ processes are very close, but quite different from that for the $\Upsilon(3S)\to\Upsilon(1S)\pi^+\pi^-$ process. The latter one is about ten times larger than the former [@mu97]. This is somewhat unnatural. Suffice to say, the pions involved in $\Upsilon(3S)\to\Upsilon(1S)\pi^+\pi^-$ are somewhat harder than in the other bottomonium transitions, and in principle the values of $g_2/g_1$ for these processes should not be the same for different dynamical regions. However, in the decay processes considered, the vector mesons involved are all in the $S$ wave state, and the particles involved are in the same mass scale, and the difference among kinematical regions is not too large. Thus we deem that the values of $g_2/g_1$ for these processes should be very close. To reduce the number of free parameters, we take the same $g_2/g_1$ value for different $\Upsilon(nS)$ decays. 2.cm -1.5cm 2.cm 1.cm -1.5cm The decay data for $\Upsilon(2S)\to\Upsilon(1S)\pi^+\pi^-$ are taken from [@cleo98] and for $\Upsilon(3S)\to\Upsilon(2S)\pi^+\pi^-$ and $\Upsilon(3S)\to\Upsilon(1S)\pi^+\pi^-$ from [@cleo94]. To get the physical coupling constants, the data for $\Upsilon(2S)\to\Upsilon(1S)\pi^+\pi^-$ [@cleo98] is normalized by $\Gamma_{\Upsilon(2S)}= 43keV$ and $B(\Upsilon(2S)\to\Upsilon(1S)\pi^+\pi^-)=19.2\%$ [@pdg04]. Our calculated results are plotted in Figs. \[fig:noxfit\]-\[fig:nox31fit\]. It is shown that the resultant $\pi\pi$ invariant mass spectra for both $\Upsilon(2S)\to\Upsilon(1S)\pi^+\pi^-$ and $\Upsilon(3S)\to\Upsilon(2S)\pi^+\pi^-$ decays agree with the data values, but the angular distribution for the former one is somewhat flat, which might also be due to the same reason discussed in Section \[sec:psi\]. On the other hand, there is almost no way to fit the $\pi^+\pi^-$ invariant mass spectrum and the $\cos\theta_{\pi}^$ distribution of the $\Upsilon(3S)\to\Upsilon(1S)\pi^+\pi^- $ process simultaneously, even if $g_2/g_1$ is further released as a free parameter. The resultant parameters are listed in Tab. \[tab:nox\]. Decay $g_1$ $g_2/g_1$ $g_3/g_1$ ----------------------------------------- -------- ----------- ----------- $\Upsilon(2S)\to\Upsilon(1S)\pi^+\pi^-$ 0.0944 -0.230 0 $\Upsilon(3S)\to\Upsilon(2S)\pi^+\pi^-$ 0.768 -0.230 0 $\Upsilon(3S)\to\Upsilon(1S)\pi^+\pi^-$ 0.0123 0.564 -13.602 : Resultant parameters through data fitting. \[tab:nox\] Sequential decay mechanism -------------------------- In order to explain the decay data of the $\Upsilon(3S)\to\Upsilon(1S)\pi^+\pi^-$ decay, we propose an additional sequential decay mechanism where an intermediate state, called $X$, is introduced. Additional Feynman diagrams for $\Upsilon(nS)\to\Upsilon(mS)\pi^+\pi^-$ are shown in Fig. \[fig:feyn2\], where (a) depicts the tree diagram and (b) the diagram including the $\pi\pi$ $S$ wave FSI. =4.cm We adopt a simple $S$ wave coupling for $\Upsilon(nS)\to\pi X$. The quantum numbers of $X$ should be $J^P=1^+$ and $I=1$. The decay amplitude of Fig. \[fig:feyn2\](a) can be written as $$V_X^{tree}=g_{nm} \epsilon^{^{\prime}}_{\mu}\epsilon^{}_{\nu}(\frac{-g^{\mu\nu}+p_{X^+}^{\mu}p_{X^+}^{\nu}/m_X^2}{p_{X^+}^2-m_X^2+i m_X\Gamma_X}+\frac{-g^{\mu\nu}+p_{X^-}^{\mu}p_{X^-}^{\nu}/m_X^2}{p_{X^-}^2-m_X^2+i m_X\Gamma_X}) \label{eq:xtree}$$ where $p_{X^+}$ and $p_{X^-}$ are the momenta of $X^+$ and $X^-$ respectively, $g_{nm}$ is an effective coupling constant among $\Upsilon(nS)$, $\Upsilon(mS)$, $\pi^+$ and $\pi^-$ via an intermediate resonant state $X$. In fact, $g_{nm}$ is the product of two coupling constants $g_{nX}$ and $g_{mX}$ where $g_{kX}(k=n,m)$ denotes the coupling constant for the $\Upsilon(kS)X\pi$ vertex. To further consider the effect of the $\pi\pi$ $S$ wave FSI, the contribution of Fig. \[fig:feyn2\](b) should be included. In this figure, the three-propagator loop can be expressed as $$G_X^{\mu\nu}=i\int\frac{dq^4}{(2\pi)^4} \frac{-g^{\mu\nu}+p_{X}^{\mu}p_{X}^{\nu}/m_X^2}{p_X^2-m_X^2+i\varepsilon} \frac{1}{q^2-m_{\pi}^2+i \varepsilon} \frac{1}{(p^{^{\prime}}-p-q)^2-m_{\pi}^2+i\varepsilon} \label{eq:3loop}$$ where $p_X=p^{^{\prime}}-q$ is the four-momentum of $X$. The calculation is carried out in the c.m. frame of the $\pi\pi$ system with the same cutoff value used in the two-meson loop calculation. As argued in ref.[@absz95], terms with $\epsilon^{^{\prime}}_{\mu}\epsilon^{}_{\nu}p_X^{\mu}p_X^{\nu}/m_X^2$ in Eqs. \[eq:xtree\] and \[eq:3loop\] can be neglected, because of the expected heavy mass of $X$. Then the total $t$-matrix can finally be written as $$\label{eq:total} t=V_0 + V_{0S}\cdot G\cdot 2t^{I=0}_{\pi\pi,\pi\pi} + V_X^{tree} + g_{nm}\epsilon^{^{\prime}}_{\mu}\epsilon^{}_{\nu} G_X^{\mu\nu}\cdot 2t^{I=0}_{\pi\pi,\pi\pi}.$$ Results for bottomonium $\pi^+\pi^-$ transitions {#subsec:Xre} ------------------------------------------------ In terms of the $t$-matrix in Eq. (\[eq:total\]), we calculate the $\pi^+\pi^-$ invariant mass spectra and the $\cos\theta_{\pi}^$ distributions of the bottomonium $\pi^{+}\pi^{-}$ transitions. Similar to the argument given in the former sections, we take $g_3=0$ for the $\Upsilon(2S)\to\Upsilon(1S)\pi^+\pi^-$ and $\Upsilon(3S)\to\Upsilon(2S)\pi^+\pi^-$ decays and $g_3$ as a free parameter for the $\Upsilon(3S)\to\Upsilon(1S)\pi^+\pi^-$ process. We also demand the values of $g_2/g_1$ to be the same for all three decays for reducing the number of free parameters. The values of $g_1$ and $g_2$ are determined by fitting the experimental decay data [@cleo98; @cleo94]. 2.cm -1.5cm The calculated results are plotted in Fig. \[fig:fit\]. It is shown that not only both the $\pi\pi$ invariant mass spectrum and the $\cos\theta_\pi^$ distribution of the $\Upsilon(3S)\to\Upsilon(1S)\pi^+\pi^-$ process can simultaneously be well explained, but also a consistent description of other bottomonium $\pi^+\pi^-$ transitions can be obtained. The resultant parameters are tabulated in Tab. \[tab:fit\]. Decay $g_1$ $g_2/g_1$ $g_3/g_1$ $g_{nm}$(GeV$^2$) $m_X$(GeV) $\Gamma_X$(GeV) ---------- --------- ----------- ----------- ------------------- ------------ ----------------- $2\to1$ 0.0886 -0.230 0 -2.316 $3\to 2$ 0.769 -0.230 0 -0.00418 10.080 0.655 $3\to 1$ 0.00546 -0.230 4.949 4.712 : \[tab:fit\] Resultant parameters in the data fitting. In the first column, $n\to m$ denotes the $\Upsilon(nS)\to\Upsilon(mS)\pi^+\pi^-$ decay. 2.cm -1.5cm To understand thoroughly the roles of different terms in the $\pi^{+}\pi^{-}$ invariant mass spectrum and the $\cos\theta_{\pi}^{\ast}$ distribution of the $\Upsilon(3S)\to\Upsilon(1S)\pi^{+}\pi^{-}$ decay, it is necessary to analyze their individual contributions. The results are shown in Fig. \[fig:analyze\]. In the figure, the solid curves represent our best fitted results, and the dotted, dashed, and dash-dotted curves describe the contributions from the terms without $X$ and with $X$ only and the interference term respectively, and the dash-dot-dotted curves represent the tree level contributions with $X$ only. The calculated $\pi\pi$ invariant mass spectrum (Fig. \[fig:analyze\] (a)) shows that the contribution from $X$ plays a dominant role, the contribution from the terms without $X$ can qualitatively but not quantitatively give the two-peak feature, and the interference term contributes constructively in the smaller $\pi\pi$ invariant mass region but destructively in the larger invariant mass region. The resultant $\cos\theta_{\pi}^{\ast}$ distribution (Fig. \[fig:analyze\] (b)) further shows that the contribution from $X$, even in the tree level, produces almost the whole angular distribution structure. Although the scalar meson $\sigma$ dynamically generated by the $S$ wave $\pi\pi$ FSI in ChUT [@oo97] can make a peak around its pole position at about 450 MeV in the $\pi\pi$ invariant mass spectrum, the contribution from the diagrams without $X$ is not dominant due to smaller values of coupling constants. Thus, an additional $D$ wave FSI which provides a flat contribution in the invariant mass region considered will not be an important contributor. These indicate that the intermediate state $X$ is very important in reproducing not only the $\pi^{+}\pi^{-}$ invariant mass spectrum but also the $\cos\theta_{\pi }^{\ast}$ distribution in the $\Upsilon(3S)\to\Upsilon(1S)\pi^{+}\pi^{-}$ decay. If we further consider the quark structures of the particles involved, the intermediate state should contain $b,~\bar{b},~q$ and $\bar{q}$. This state might be a tetraquark state, for instance, $b\bar{b}u\bar{d}$, with $J^P=1^+$ and $I=1$ for $X^+$ or a $B\bar{B}$ bound state, for instance, $B^+\bar{B}^0$, for $X^+$. =6.cm It should be mentioned that similar mechanism was also proposed by V.V. Anisovich [et al.]{}[@absz95]. In their paper, a trivial $S$ wave coupling was used in the effective vertex $\Upsilon(nS)\Upsilon(mS)\pi\pi$ which is described by Eq. (\[eq:v0\]) in our model with both $S$ wave and $D$ wave components. With that mechanism, they successfully reproduced the $\pi\pi$ invariant mass spectra of the $\Upsilon(nS)\to\Upsilon(mS)\pi^+\pi^-$ decays, but did not give reasonable $\cos\theta_{\pi}^$ distributions due to the dominance of the $\pi\pi$ $S$ wave in their model. As a result, the estimated mass of the additional intermediate state is in the range of 10.4-10.8 GeV which is located outside the data area of the Dalitz plot where the data points in the direction of $m_{\Upsilon\pi}$ are located from 9.6 GeV to 10.2 GeV. Thus the effect of the state does not show up in the Dalitz plot. The mass and width of the intermediate state in our work are different from those in Ref. [@absz95]. We also present the Dalitz plot for the $\Upsilon(3S)\to\Upsilon(1S)\pi^+\pi^-$ decay in Fig. \[fig:dalitz\]. It is shown that although the estimated mass of $X$ in our model ($M_X =10.08$ GeV) is inside the data area in the Dalitz plot, the signal of $X$ in the Dalitz plot is not very clear due to its large width of 0.655 GeV. It does not conflict with the CLEO experiment [@cleo98]. Moreover, we would mention that with the typical values, $M_X$=10.5 GeV and $\Gamma_X$=0.15 GeV given in Ref. [@absz95], we cannot produce a $\cos\theta_\pi^$ distribution that is consistent with the experimental data[@cleo98]. Summary {#sec:sum} ======= Starting from an effective Lagrangian and further employing ChUT to include the $\pi\pi$ $S$ wave FSI properly, the $\pi^+\pi^-$ transitions of heavy quarkonia are intensively studied. In order to consistently explain the $\pi^+\pi^-$ invariant mass spectra and angular distributions in the mentioned processes simultaneously, especially in the $\Upsilon(3S)\to\Upsilon(1S)\pi^+\pi^-$ decay process, an additional sequential process, where an intermediate state $X$ is introduced, is further considered in the bottomonium transitions. With such a process included, all the $\pi^+\pi^-$ transition data can be well-explained, especially the angular distribution of the $\Upsilon(3S)\to\Upsilon(1S)\pi^+\pi^-$ decay. As a consequence, the newly introduced intermediate state should have quantum numbers of $J^P=1^+$ and $I=1$, a mass of about 10.08 GeV and a width of about 0.655 GeV. The quark content of the state should be $b\bar{b}q\bar{q}$. It might be a tetraquark state or a $B\bar{B}$ bound state. The detailed inner structure of the state should be carefully studied both theoretically and experimentally. Acknowledgements {#acknowledgements .unnumbered} ================ We would like to thank B.S. Zou for his valuable discussions and suggestions. We are also benefit from fruitful discussions and constructive comments given by E. Oset and D.O. Riska. We should also thank F.A. Harris for providing us the BES data used in [@bes00]. This project is partially supported by the NSFC grant Nos. 90103020, 10475089, 10435080, 10447130 and CAS Knowledge Innovation Key-Project grant No. KJCX2SWN02. [99]{} CLEO Collaboration, F. Butler [et al.]{}, Phys. Rev. D 49 (1994) 40 K. Gottfried, Phys. Rev. Lett. 40 (1978) 598 T.-M. Yan, Phys. Rev. D 22 (1980) 1652 M. Voloshin and V. Zakharov, Phys. Rev. Lett. 45 (1980) 688 V. A. Novikov and M. A. Shifman, Z. Phys. C 8 (1981) 43 Y.-P. Kuang and T.-M. Yan, Phys. Rev. D 24 (1981) 2874 P. Moxhay, Phys. Rev. D 39 (1989) 3497 H.-Y. Zhou and Y.-P. Kuang, Phys. Rev. D 44 (1991) 756 H. J. Lipkin and S. F. Tuan, Phys. Lett. B 206 (1988) 349 L. Brown and R. Cahn, Phys. Rev. Lett. 35 (1975) 1 T. Mannel and R. Urech, Z. Phys. C 73 (1997) 541 M.-L. Yan, Y. Wei and T.-L. Zhuang, Eur. Phys. J. C 7 (1999) 61 M. Uehara, Prog. Theo. Phys. 109 (2003) 265 G. Bélanger, T. DeGrand and P. Moxhay, Phys. Rev. D 39 (1989) 257 S. Chakravarty, S. M. Kim and P. Ko, Phys. Rev. D 48 (1993) 1205 S. Chakravarty and P. Ko, Phys. Rev. D 48 (1993) 1212 S. Chakravarty, S. M. Kim and P. Ko, Phys. Rev. D 50 (1994) 389 A. Gallegos, J. L. Lucio M and J. Pestieau, Phys. Rev. D 69 (2004) 074033 T. A. Lähde and D. O. Riska, Nucl. Phys. A 707 (2002) 425; V. V. Anisovich, D. V. Bugg, A. V. Sarantsev and B. S. Zou, Phys. Rev. D 51 (1995) R4619 J. A. Oller and E. Oset, Nucl. Phys. A 620 (1997) 438; (Erratum) [ibid.]{} 652 (1999) 407 J. A. Oller, E. Oset and J. R. Peláez, Phys. Rev. Lett. 80 (1998) 3452; Phys. Rev. D 59 (1999) 074001, (Erratum) [ibid.]{} 60 (1999) 099906 J. A. Oller and E. Oset, Nucl. Phys. A 629 (1998) 739 J. A. Oller and E. Oset, Phys. Rev. D 60 (1999) 074023 J. A. Oller, Phys. Lett. B 426 (1998) 7 Ulf-G. Mei$\beta$ner and J. A. Oller, Nucl. Phys. A 679 (2001) 671 C. Li, E. Oset and M. J. Vicente Vacas, Phys. Rev. C 69 (2004) 015201 J. E. Palomar, L. Roca, E. Oset and M. J. Vicente Vacas, Nucl. Phys. A 729 (2003) 743 L. Roca, J. E. Palomar, E. Oset and H. C. Chiang, Nucl. Phys. A 744 (2004) 127 T. S. H. Lee, J. A. Oller, E. Oset, A. Ramos, Nucl. Phys. A 643 (1998) 402 J. A. Oller, E. Oset, J. E. Palomar, Phys. Rev. D 63 (2001) 114009 J. A. Oller, Phys. Rev. D 71 (2005) 054030 J. A. Oller, E. Oset and A. Ramos, Prog. Part. Nucl. Phys. 45 (2000) 157 BES Collaboration, J. Z. Bai [et al.]{}, Phys. Rev. D 62 (2000) 032002 Particle Data Group, S. Eidelman [et al.]{}, Phys. Lett. B 592 (2004) 1 CLEO Collaboration, J. P. Alexander [et al.]{}, Phys. Rev. D 58 (1998) 052004 [^1]:	{ "pile_set_name": "ArXiv" }

Subsets and Splits

No community queries yet

The top public SQL queries from the community will appear here once available.