text
stringlengths 4
2.78M
| meta
dict |
|---|---|
---
abstract: 'The main result of this paper is that a pseudoperiodic automorphism on an oriented surface with boundary such that a power of it is a composition of positive Dehn twists around disjoint simple closed curves including all boundary components, admits itself a positive factorization. The main ingredients of the proof are a solution for the $\bar{\partial}$-problem in families of strongly pseudoconvex manifolds with bounds in the appropriate Lipschitz topology by Greene and Krantz, deformation theory for pseudoconvex manifolds as developed by Laufer, Bogomolov and De Oliveira, and a topological characterization of the monodromies on links of isolated complex surface singularities by Anne Pichon. This generalizes previous results by Honda, Kazez and Matić on the punctured torus and greatly narrows the difference between the monoid of right-veering mapping classes and the monoid of mapping classes admitting positive factorizations. We define a new invariant of automorphisms of surfaces that detects how far is an automorphism from admitting a positive factorization. Finally we apply the main theorem to give a sufficiency criterion for certain pseudoperiodic automorphisms to admit a positive factorization.'
author:
- Pablo Portilla Cuadrado
bibliography:
- 'bibliography.bib'
title: Positive factorizations of pseudoperiodic automorphisms
---
Introduction
============
Although the main result of this paper has its greatest impact in the theory of mapping class groups, its original motivation lies in the field of singularity theory. We describe this motivation first.
Let $f:{\mathbb{C}}^2 \to {\mathbb{C}}$ be a representative of a holomorphic map germ having an isolated singularity at the origin. Then for a generic linear form $\ell:{\mathbb{C}}^2 \to {\mathbb{C}}$ and $\epsilon'$ a real number with $|\epsilon'|$ small enough, we have that $$\tilde{f}:=f + \epsilon' \ell: {\mathbb{C}}^2 \to {\mathbb{C}}$$ is what is usually known as a [**morsification**]{} of $f$. That is, a holomorphic map that only has Morse-type singularities with distinct critical values all of them close to $0 \in {\mathbb{C}}$. On one hand, we have that $$\label{eq:locally_trivial_milnorle}
f_{|f^{-1}(\partial D_\delta) \cap B_{\epsilon}}: f^{-1}(\partial D_\delta) \cap B_{\epsilon} \to \partial D_\delta$$ is a locally trivial fibration for $B_\epsilon \subset {\mathbb{C}}^2$ a closed ball of small radius $\epsilon>0$ and $ D_\delta \subset {\mathbb{C}}$ a disk of small radius $0<\delta<<\epsilon$. This is known as the Milnor fibration on the tube [@Mil].
Since transversality is an open condition, $$\tilde{f}_{| \tilde{f}^{-1}(\partial D_\delta) \cap B_{\epsilon}}: \tilde{f}^{-1}(\partial D_\delta) \cap B_{\epsilon} \to \partial D_\delta$$ defines an equivalent fibration to that of \[eq:locally\_trivial\_milnorle\] for $\epsilon'$ small enough. Since $\tilde{f}$ has only singular points near the origin $0\in{\mathbb{C}}^2$, one can easily deduce that the monodromy associated with $f$ can be written as a composition of right-handed Dehn twists. This classical line of argument is known as Picard-Lefschetz theory and works well when the ambient space is a smooth Stein manifold.
Let us change ${\mathbb{C}}^2$ by a singular ambient space. Take $X'$ to be an isolated complex surface singularity. Let $f':X'\to {\mathbb{C}}$ be an element of its ring of functions, that is an holomorphic map. Assume that $f'$ is reduced as an element on the structure ring of $X'$. Associated with $f'$ there is also a locally trivial fibration $$f'_{|f'^{-1}(\partial D_\delta) \cap X' \cap B_\epsilon}: f'^{-1}(\partial D_\delta) \cap X' \cap B_\epsilon \to \partial D_\delta$$ on a [*tube*]{}. This one is known as the Milnor-Lê fibration [@Le]. Now we observe that a similar game does not work since a small perturbation $f' + \epsilon'\ell$ has some Morse singularities on the smooth part of $ (f'+\epsilon'\ell)^{-1}(\partial D_\delta) \cap X' \cap B_\epsilon$ [*and*]{} the singularity defined by $\ell$ at the singular point of $X'$. To see more on this we recommend the original work by Siersma [@Sier]. So, by these methods, we cannot conclude that the monodromy defined by $f'$ can be written as a composition of right-handed Dehn twists. To decide if these monodromies admit (or not) positive factorizations is the original motivation of this paper.
We argument as follows to solve this problem. First we consider the minimal resolution $X \to X'$ of $X'$. This gives us a strongly pseudoconvex surface $X$ with a non-trivial exceptional set without curves of the first kind. For these kind of surfaces, Laufer [@Lau; @Lau2] and later Bogomolov and De Oliveira [@Bog] developed methods to prove the existence of a deformation $\omega:{\mathcal{X}}\to Q$ of $X$ over a disk in such a way that generic fibers are Stein. As a ${\mathcal{C}}^\infty$-fibration, this deformation is a trivial fibration. We lift $f'$ to a holomorphic map on the resolution $f:X \to {\mathbb{C}}$ and extend this map to the fibers $X_t$ of the fibration as complex valued (but not holomorphic) smooth maps $f_t:X_t \to {\mathbb{C}}$.
Now we ask the question: Can we [*correct*]{} $f_t$ so that it becomes holomorphic? Note that $\bar{\partial}f_t$ is small in fibers close to the central fiber of the deformation. So we need to solve what is classically known as the $\bar{\partial}$-problem and get a nice enough bound for the solution. More concretely, we seek for $u_t:X_t \to {\mathbb{C}}$ such that $$\bar{\partial} f_t = \bar{\partial} u_t$$ and such that $u_t$ takes small values. It turns out that the PDE’s methods initiated by Kohn and Nirenberg [@Kohn] and followed Hörmander [@Hor] are not enough. These methods give only bounds of the $L^2$-norm of $u$ which does not assure us that $u$ takes small values. A few years later, Henkin [@Hen] developed [*integral representation*]{} methods to improve these results. His work was later generalized by Kerzman’s [@Ker] manifolds that can be covered by strongly pseudoconvex open domains by patching Henkin’s solution. This [*explicit*]{} integral solutions allowed Greene and Krantz to produce a version of this result for families of strongly pseudo-convex manifolds (where the complex structure varies smoothly) as in [@Greene Section 4]. This is the result that we invoke.
In the end we are able to construct such $u_t:X_t \to {\mathbb{C}}$ with $f_t-u_t$ a holomorphic map that defines a locally trivial fibration equivalent to the one defined by the original $f$ on $f'^{-1}(\partial D_\delta)\cap X' \cap B_\epsilon$. Since $X_t$ is Stein can apply Picard-Lefschetz theory to conclude that our original monodromy admits a positive factorization. This is what we prove in which is the main result.
In [@Pich], A. Pichon proved, using a previous result by Winters [@Win], a purely topological characterization of the monodromies that appear in links of isolated complex surface singularities associated with reduced holomorphic map germs. This class coincides with the class of pseudoperiodic diffeomorphisms of surfaces with boundary $\phi:{\Sigma}\to {\Sigma}$ such that $\phi^n$ is a composition of powers of right handed Dehn twist around disjoint simple closed curves including all boundary components ([*à torsades négatives*]{} in [@Pich]). So proving that these monodromies admit positive factorizations has highly unexpected consequences in the theory of mapping class groups. As a direct corollary we find, for example, that all freely periodic diffeomorphisms with positive fractional Dehn twist coefficients automatically admit positive factorizations. Observe that Honda, Kazez and Matić, [@KoII] proved this for the punctured torus.
Next we describe the organization of the paper.
### Outline of the paper
We start in \[sec:complex\_geometry\] by recalling some theory and definitions about pseudoconvex manifolds and their deformations. Then we discuss solutions to the $\bar{\partial}$-problem on Stein manifolds. First we introduce a theorem by Kerzman which proves existence and uniform bounds for the solution in the precise setting in which we are interested and finally we state a version of a theorem ( in this work) by Greene and Krantz that gives a solution to the e $\bar{\partial}$-problem for $(0,1)$ forms in smooth families of complex manifolds.
Next, in \[sec:mapping\_class\_groups\], we fix notation and conventions about certain topics of the theory of mapping class groups. This is always a necessary step if one wants to make precise statements because conventions highly vary from one author to another. We focus on the notions of fractional Dehn twist coefficient and screw numbers associated with a pseudoperiodic automorphism of a surface. We end the section by stating Anne Pichon’s characterization of monodromies on links of isolated complex surface singularities.
In a first read of this work, an expert reader might skip one or the previous two sections and go straight to the proof of the main theorem.
In \[sec:main\_thm\] we start by proving a transversality proposition and then apply the theory developed in the previous sections to prove our main result following the reasoning described previously in the introduction.
Finally, in \[sec:consequences\] we are devoted to explore the consequences of our main result. We start by exploting a result by Baykur, Monden and Van Horn-Morris in to prove that the pure mapping class group (diffeomorphisms up to isotopy free on the boundary) is generated as a semi-group by positive Dehn twists except in a few degenerate cases. This leads to the definition of two invariants () that measure the failure of a pseudoperiodic automorphism to be in ${\mathrm{Dehn}^+_{g,r}}$. We use this together with to give a sufficiency criterion (in ) that tells us, in many cases, if a pseudoperiodic automorphism admits a positive factorization. The criterion roughly says that if the fractional Dehn twist coefficients are very big with respect to the screw numbers of the separating curves of the automorphism, then the corresponding pseudoperiodic automorphism admits a positive factorization.
Acknowledgments {#acknowledgments .unnumbered}
===============
I wish to thank Xavier Gómez-Mont for inspiring conversations. I am also very thankful to Mohammad Jabbari who is an expert in the $\bar{\partial}$-problem and gave me many useful references that lead me to find the Greene and Krantz result that I ended up using.
Finally, I thank Baldur Sigursson who read carefully an early version of this manuscript and pointed out a gap in a lemma that was placed instead of current . His critics and comments have helped me greatly improve the final manuscript.
Preliminary theory on complex geometry and singularity theory {#sec:complex_geometry}
=============================================================
We start by reviewing some theory on pseudoconvex complex manifolds that we need to state the results. Let $X$ be a complex manifold. Denoting its tangent bundle by $TX$ we can see its complex structure as an endomorphism $$J:TX \to TX$$ satisfying $J^2=-{\mathop{\mathrm{id}}\nolimits}$ and an integrability condition. Let $d\rho$ denote the exterior derivative of $\rho$ for any smooth real valued function $\rho$. The complex structure $J$ allows us to define the complex exterior derivative as $d^{{\mathbb{C}}} \rho:= d\rho\circ J$. Let $\partial$ and $\bar{\partial}$ be the complex and complex conjugate parts of the exterior derivative. We can write $d= \partial + \bar{\partial}$ and $d^{{\mathbb{C}}}= i(\partial - \bar{\partial})$.
Let $\rho:X \to {\mathbb{R}}$ be a smooth real valued function. We define a $(1,1)$-form $\omega_\rho := -dd^{\mathbb{C}}\rho$, a symmetric bilinear form $g_\rho := \omega_\rho(\cdot, J\cdot)$ and a Hermitian form $h_\rho= g_\rho - i \omega_\rho $. Under suitable circumstances these turn, respectively, into a symplectic form, a Riemannian metric and Hermitian metric. This motivates the following definition.
\[def:spsh\_exh\] Let $X$ be a complex manifold and let $A\subset X$ be an open subset. We say that a smooth function $\rho:X \to {\mathbb{R}}$ is [**strongly plurisubharmonic**]{} (abbreviated [**spsh**]{}) on $A$ if the metric $g_\rho$ is positive-definite (a Riemannian metric) on $A$. If $g_\rho$ is positive-definite on all $X$, we simply say that $\rho$ is strongly plurisubharmonic.
We say that a smooth function $\rho:X \to {\mathbb{R}}$ is an [**exhaustion function**]{} if it is proper and bounded from below.
This leads us to the following important definitions.
\[def:spc\_stein\] We say that a complex manifold $X$ is [**strongly pseudoconvex**]{} if it admits an exhaustion function $\rho:X \to {\mathbb{R}}$ that is spsh outside a compact set. We say that $X$ is a [**Stein manifold**]{} if it admits an exhaustion function which is spsh on all $X$.
We say that $\bar X$ is a [**strongly pseudoconvex manifold with boundary**]{} if it is a compact complex manifold with smooth boundary that admits an exhaustion function $\rho: \bar X \to {\mathbb{R}}$ which is spsh outside a compact analytic set and such that $\partial \bar X$ is the level set $\{x \in \bar X:\rho = 0\}$ of $\rho$. We say that $\bar X$ is a [**Stein domain**]{} if it is a compact complex manifold with smooth boundary that admits an exhaustion spsh function $\rho: \bar X \to {\mathbb{R}}$ such that $\partial \bar X = \{x \in \bar X:\rho = 0\}$.
If a complex analytic space $\bar Y$ with boundary $\partial \bar Y$ satisfies that its boundary is the level set of a spsh function defined on a neighborhood of it, we say that $\bar Y$ has a [**strongly pseudoconvex boundary**]{} or that its boundary is strongly pseudoconvex.
Observe that if $X$ is a Stein manifold with spsh exhaustion function $\rho:X \to {\mathbb{R}}$ and $\{x \in X: \rho = 0\}$ is a regular level set, then $\bar X := \{x \in X: \rho (x) \leq 0\}$ is a Stein domain by definition. Actually, it can be shown that all Stein domains can be obtained like this.
\[rem:difference\] The difference between Stein manifolds and strongly pseudoconvex manifolds is that the latter might have a non-trivial exceptional set. And that is the only difference: a strongly pseudoconvex manifold that does not contain compact analytic sets of dimension greater than $0$ is a Stein manifold. A typical example of a strongly pseudoconvex manifold that is not a Stein manifold is the resolution space of an isolated singularity.
For each pair of non negative integers $p,q \geq 0$ we denote by ${\mathcal{C}}^\infty_{p,q}(X)$ the global differential forms of type $(p,q)$ with coefficients smooth complex valued functions. Then the operator $\bar{\partial}$ acts $$\bar{\partial}: {\mathcal{C}}^\infty_{p,q}(X) \to {\mathcal{C}}^\infty_{p,q+1}(X)$$ as the complex conjugate part of the exterior derivative and satisfies $\bar{\partial}^2 = 0$.
Observe that ${\mathcal{C}}^\infty_{0,0}(X)$ coincides with the set of smooth complex valued functions and that for an element $f \in {\mathcal{C}}^\infty_{0,0}(X)$ the equation $\bar{\partial}f=0$ is satisfied precisely when $f$ is a holomorphic function.
The $\bar{\partial}$-problem in families of strongly pseudoconvex surfaces {#the-barpartial-problem-in-families-of-strongly-pseudoconvex-surfaces .unnumbered}
--------------------------------------------------------------------------
Given an element $g \in {\mathcal{C}}^\infty_{p,q+1}(X)$ it is a classical problem to determine the existence (or lack thereof) of solutions $u \in {\mathcal{C}}^\infty_{p,q}(X)$ to the equation $$\bar{\partial} u = g$$ and to bound the value of some norm on $u$ by the value of some norm on $g$. This problem is known as the $\bar{\partial}$-problem.
In this work we wish to control the growth of the partial derivatives of $u$ under small smooth perturbations of the complex structure $J$. For this purpose it is necessary to introduce several norms that are used in the theorems that are invoked related to this questions. We warn the reader that there does not seem to be a consensus in the notation used for the different norms. We follow [@Greene] which is the work from where we cite the main result used. The definitions are made for domains in ${\mathbb{C}}^n$ and all of them have straightforward generalizations to complex valued functions defined on complex manifolds by patching local charts.
Let $D \subset {\mathbb{C}}^n$ be an open set. For a complex valued function $u:D \to {\mathbb{C}}$ we define the [*supremum norm*]{} by
$$||u||_\infty:=\sup_{x \in X} |u(x)|$$
where $|\cdot|$ denotes the usual complex modulus. The next is called the $C^j$ norm and controls the behavior of the partial derivatives of $u$ up to the $j$-th order.
$$||u||_{C^j}:= \sum_{|\alpha|+|\beta| \leq
j}\left|\left|\left(\frac{\partial}{\partial z}\right)^\alpha
\left(\frac{\partial}{\partial \bar{z}}\right)^\beta
u\right|\right|_\infty$$
where $\alpha$ and $\beta$ are meant in the usual multi-index notation and $|\alpha|$ and $|\beta|$ are their total orders. We define $C^j(D):=\{u:D \to
{\mathbb{C}}: ||u||_{C^j} < \infty\}.$
Kerzman, in [@Ker] proved the existence of solutions in the Stein manifold case with an uniform bound of the solution. He actually proved this bounds for the so-called Henkin solution previously discovered by Henkin in [@Hen]. Kerzman first proves the result for [*strongly pseudoconvex domains*]{} (which he defines to be open sets in ${\mathbb{C}}^n$ which are Stein manifolds) [@Ker Theorem 1.2.1]. Then he notes that this result is as well valid for the case of strictly pseudoconvex domains [*in*]{} Stein manifolds [@Ker Theorem 1.2.1”, pg. 309].
The original idea for that result emanates from Henkin’s work [@Hen] for $(0,1)$-forms in strongly pseudoconvex domains. A natural generalization of this estimate consists of the study of the regularity of the solutions to the $\bar{\partial}$-problem in terms of the regularity of a family of complex structures. In other words: if we consider a family of complex structures $J_t$ that moves continuously (in $t$) in a certain topology defined by some norm, in which topologies do the solutions (with respect to those complex structures) also move continuously?
This question is solved in different settings in [@Greene]. In this work we only use their estimates for the Henkin solution for $(0,1)$-forms with respect to a smooth family of complex structures. This is gathered in [@Greene Section 4]. As they explain in the introduction to Section 4 therein, the Henkin solution moves continuously in the $C^j$ topology provided that the complex structures varies sufficiently smoothly. From the results proven in that section we can extract the following useful result.
\[thm:continuous\_variation\] Let $\bar X$ be a strongly pseudoconvex manifold with boundary, let $g$ be a $\bar{\partial}$-closed $(0,1)$-form with $C^1$ coefficients and let $u$ be the Henkin solution to the problem $\bar{\partial} u = g$. Then, a small $C^k$ perturbation of the complex structure results in a $C^j$ variation of the solution $u$. Where $j$ can be expressed as a function depending only on $k$ and the dimension $n$ of $\bar{X}$ in such a way that if $k$ is big enough then $j$ is as big as we want. In particular, we can assure that $j \geq 1$ for $k>>1$.
Since the statement in the cited paper is expressed in a different language, we make the following observations.
The previous theorem is contained in the cited paper as follows. First, we have [@Greene Theorem 4.14, eq. 4.14.3’] which proves it for strongly pseudoconvex domains (relatively compact open sets with strongly pseudoconvex boundary) in ${\mathbb{C}}^n$. In the remark at the end of that same page it is explained that, with the help of Sobolev embedding theorems and similar techniques as the ones used by Kerzman in [@Ker], one can prove similar bounds for strongly pseudoconvex manifolds by patching the solutions (observe that a strongly pseudoconvex manifold can be defined as a manifold such that each point has a strictly convex open neighborhood).
Also note that [@Greene Theorem 4.14] is stated for [*perturbations*]{} of the boundary of the strongly pseudoconvex domain but at the end of the statement, the authors remark that similar bounds are true for variations of the complex structure (which they denote by $\prod_{1,0}$).
Finally, we make clear that we do not care about sharp bounds in this work. Actually, in our situation we only require that the first partial derivatives of the solution vary continuously (i.e. that the solution varies continuously in the $C^1$ topology) provided that the complex structure varies smoothly enough. So we do not investigate more in this direction and that is why we have just stated the weakest version of the result that is used here.
Singularity theory and deformations of strongly pseudoconvex surfaces {#singularity-theory-and-deformations-of-strongly-pseudoconvex-surfaces .unnumbered}
---------------------------------------------------------------------
We introduce the notion of [*deformation*]{} of a complex manifold and state the result on deformations of strongly pseudoconvex surfaces proven by Bogomolov and De Oliveira exploiting a previous result by Laufer.
Let $X$ be a strongly pseudoconvex manifold (maybe with smooth boundary). We say that a map $\omega:{\mathcal{X}}\to Q$ is a [**deformation**]{} of $X$ if
1. ${\mathcal{X}}$ and $Q$ are complex manifolds (maybe with smooth boundary).
2. The map $\omega$ is a flat submersion.
3. The central fiber $\omega^{-1}(0)$ is biholomorphic to $X$.
Suppose that we start with a strongly pseudoconvex manifold with smooth boundary $\bar X$. Then, it is known that if one wishes to keep track of the boundary of $X$, the theory of deformations of $X$ becomes infinite-dimensional [@Burn Corollary 4.2]. What Laufer did in his series of papers [@Lau2; @Lau] to avoid this fact is to fix a compact analytic set $A \subset X$ and study deformations of $X$ near $A$. This allowed him to apply Kodaira-Spencer techniques to prove his results. It turns out that this setting is enough for many purposes, in particular for proving the existence of a versal deformation space [@Lau2].
The next is a result obtained by Bogomolov and De Oliveira [@Bog Theorem (2)] that extends a previous result by Laufer [@Lau Theorem 3.6] by iteratively applying it.
\[thm:small\_stein\_deformation\] Let $\bar X$ be a strongly pseudoconvex surface with boundary that has a $1$-dimensional exceptional set without curves of the first kind. Then there exists a deformation $\omega:\bar {\mathcal{X}}\to Q$ over a small complex disk $Q\subset{\mathbb{C}}$ such that the fibers above $t \neq 0$ do not contain compact analytic curves.
The theorem above is another key ingredient for the proof of .
Preliminary theory on mapping class groups {#sec:mapping_class_groups}
==========================================
We turn now our attention to the theory of mapping class groups. The purpose of this section it to collect the results on mapping class group that we use throughout the rest of the text as well as to fix notation and conventions.
\[not:preliminaries\] The symbol ${\Sigma}_{g,r}$ denotes an oriented compact surface of genus $g$ and $r$ boundary components. When it is clear from the context or unnecessary in the discussion, the subindex is dropped and we just write ${\Sigma}$.
We write ${\mathrm{Mod}_{g,r}}$ for the [**mapping class group**]{} of the surface ${\Sigma}_{g,r}$, that is, the group of diffeomorphisms of ${\Sigma}_{g,r}$ that restrict as the identity to $\partial {\Sigma}$ up to isotopy fixing the boundary pointwise. We write ${\mathrm{PMod}_{g,r}}$ for what is usually known as the [**pure mapping class group**]{}, that is, the group of diffeomorphisms of a surface that leave each boundary component invariant, up to isotopy free on the boundary. Observe that there is a natural surjective homomorphism of groups $${\mathrm{Mod}_{g,r}}\to {\mathrm{PMod}_{g,r}}.$$
For $\phi\in {\mathrm{Mod}_{g,r}}$ we denote by $\hat{\phi}$ its image in ${\mathrm{PMod}_{g,r}}$ by the above projection. It is a known result that these mapping class groups do not depend on whether we consider homeomorphisms or diffeomorphisms (see for example [@Farb Section 1.4]). For this reason we choose to speak of [**automorphisms**]{} to refer to a particular representative of a mapping class.
The letter $\phi$ denotes either an automorphism or a mapping class. It is always clear from the context which is the case.
Given a simple closed curve $\gamma \subset {\Sigma}$ we denote by $$t_\gamma: {\Sigma}\to {\Sigma}$$ a [**right-handed Dehn twist**]{} around $\gamma$ or its mapping class in ${\mathrm{Mod}_{g,r}}$. The support of $t_\gamma$ is concentrated in a tubular neighborhood of $\gamma$. Also, the mapping class of $t_\gamma$ only depends on the isotopy class of the simple closed curve $\gamma$.
Given a boundary component $B \subset \partial {\Sigma}$ we denote by $t_{B}$ the [**right-handed boundary Dehn twist**]{} around $B$, which is by definition a right-handed Dehn twist around a simple closed curve parallel to $B$. The set of mapping classes that admit a factorization consisting only of right-handed Dehn twists is clearly a monoid and we denote it by $\mathbf{{\mathrm{Dehn}^+_{g,r}}}$. We also call these factorizations [**positive**]{}.
The following is a landmark result in mapping class group which we state to motivate the next definition.
\[thm:nt\_classification\] Let $\phi: {\Sigma}\to {\Sigma}$ be an orientation preserving automorphism that restricts to the identity on $\partial {\Sigma}$. Then there exists $\phi'$ isotopic to $\phi$ and a collection ${\mathcal{C}}$ of disjoint simple closed curves such that:
1. The collection of curves is invariant by $\phi'$, i.e. $\phi'({\mathcal{C}})= {\mathcal{C}}$.
2. The automorphism $\phi'$ restricted to each connected component of ${\Sigma}\setminus {\mathcal{C}}$ (and its iterations by $\phi'$) is either periodic or pseudo-Anosov.
The decomposition given by is called [**Nielsen-Thurston decomposition**]{} and is unique up to isotopy if the system of curves ${\mathcal{C}}$ is minimal. This decomposition leads to the following definition.
\[def:nt\_decomposition\] Let $\phi \in {\mathrm{Mod}_{g,r}}$, we say that $\phi$ is [**pseudoperiodic**]{} if only periodic pieces appear in its Nielsen-Thurston decomposition.
Pseudoperiodic automorphisms are of special importance in Singularity Theory: they are the only type of automorphism that appear in this world. This is almost equivalent to the fact the links of isolated complex surface singularities are graph manifolds. Actually, only a restricted type of pseudoperiodic automorphisms shows up in this area, to be able to describe precisely this class we introduce a couple of famous notions in mapping class group.
\[not:fractional\_dehn\_twist\] Let $\phi \in {\mathrm{Mod}_{g,r}}$ and let $B_1, \ldots, B_r$ be the connected components of $\partial {\Sigma}_{g,r}$. We denote the [**fractional Dehn twist coefficient**]{} of $\phi$ at $B_i$ by $\mathrm{fr}(\phi, B_i).$ It is a rational number and we use the sign convention (same as in [@KoI]) that a right-handed Dehn twist around a boundary parallel curve has fractional Dehn twist coefficient equal to $+1$.
If we assume that $\phi \in {\mathrm{Mod}_{g,r}}$ is pseudoperiodic, the fractional Dehn twist coefficient can be defined as follows. Let $B_i \subset \partial {\Sigma}_{g,r}$ and let $n \in {\mathbb{N}}$ be such that $\phi^n$ is freely isotopic to the identity on the periodic piece containing $B_i$. Then, in a collar neighborhood of $B_i$, the automorphism $\phi^n$ isotopic to $t_{B_i}^k$ for some $k \in {\mathbb{Z}}$. Then $$\mathrm{fr}(\phi, B_i):=\frac{k}{n}.$$ Note that $k$ might be negative meaning that $\phi^n$ is a power of a left-handed boundary Dehn twist.
Similarly we can define the notion of [**screw number**]{} for each orbit of separating curves of the Nielsen-Thurston decomposition of a pseudoperiodic automorphism $\phi$. Let $C_1, \ldots, C_a$ be an orbit of curves in ${\mathcal{C}}$ (as in ) so $\phi(C_i)=C_{i+1}$ for all $i=1, \ldots, a-1$ and $\phi(C_a)=C_1$. Let $n$ be such that $\phi^n$ is the identity on the two periodic pieces (which might be the same) adjacent to the orbit. Then near any of the curves $C_i$ in the orbit, $\phi$ is isotopic to $t^k_{C_i}$ for some $k\in {\mathbb{Z}}$. We define the screw number of $\phi$ at the orbit, as $$\mathrm{sc}(\phi,C_i):=\frac{k a }{n}.$$
This rational number measures the amount of twisting of $\phi^a$ around $C_i$ and it is the same for all curves in the same orbit so it is well defined the orbit itself or for any particular curve in ${\mathcal{C}}$.
We warn the reader that our conventions may differ from those of other authors. For example, Montesinos and Matsumoto [@Mat] have the same criterion that we do in the sign of fractional Dehn twist coefficients but have the opposite in screw numbers. Our choice of signs makes that [*turning to the right around a curve* ]{} is [*positive*]{} no matter if the curve is boundary parallel or a non separating simple closed curve in ${\mathcal{C}}$.
The notion of fractional Dehn twist coefficient has been present in the literature for a long time and it is difficult to attribute it to a particular author. For example, Gabai [@Gab] already used these quantities to measure the difference between an automorphism admitting a pseudo-Anosov representative and this representative.
More recently, Honda, Kazez and Matić [@KoI] used fractional Dehn twist coefficients as a topological tool to detect tight open books from the monodromy. More concretely, they [*recovered*]{} the notion of an automorphism being [**right-veering**]{} (see [@KoI Definition 2.1]) and defined the monoid $\mathbf{{\mathrm{Veer}_{g,r}}}$ consisting of all automorphisms of the surface ${\Sigma}_{g,r}$ that are right-veering. The following theorem contains just a couple of results of [@KoI] relevant for this work.
\[thm:veer\_honda\] Let $\phi \in {\mathrm{Mod}_{g,r}}$ then
1. ${\mathrm{Dehn}^+_{g,r}}\subset {\mathrm{Veer}_{g,r}}$ ([@KoI Lemma 2.5]).
2. A contact structure on a $3$-manifold is tight if and only if all the supporting open books have right-veering monodromy ([@KoI Theorem 1.1]).
Also, we have the following theorem.
\[thm:dehn\_fillable\] A contact structure is holomorphically fillable if and only if there exists an open book supporting it with monodromy in ${\mathrm{Dehn}^+_{g,r}}$.
The above two theorems emphasize the importance that it has in Contact Topology to be able to the determine the gap between ${\mathrm{Veer}_{g,r}}$ and ${\mathrm{Dehn}^+_{g,r}}$.
As we said in the introduction, this work is concerned with automorphisms that appear as monodromies of reduced holomorphic map germs defined on isolated complex surface singularities. Anne Pichon proved [@Pich Théorème 5.4], using a deep result by Winters [@Win Theorem 4.3], the following purely topological characterization of these automorphisms. This result is another key ingredient for the proof of .
\[thm:anne\_realization\] Let $\phi: {\Sigma}\to {\Sigma}$ be an automorphism of a connected and oriented surface with non-empty boundary. Then there exists an isolated complex surface singularity $X$ and a reduced holomorphic map germ $f: X \to {\mathbb{C}}$ whose associated monodromy is $\phi$ if and only if some power $\phi^n$ of the automorphism is a composition of powers of right-handed Dehn twists around disjoint simple closed curves including curves parallel to all boundary components. This property is equivalent to all fractional Dehn twists and screw numbers being strictly positive.
We name this important class of pseudoperiodic automorphisms once and for all:
\[def:mondoromies\] If $\phi \in {\mathrm{Mod}_{g,r}}$ is a pseudoperiodic automorphism with strictly positive fractional Dehn twist coefficients and screw numbers, we say that it is [**fully right-veering**]{}.
One can see in [@KoI] that the definition of right-veering automorphism refers only to the behavior of $\phi$ at the boundary of ${\Sigma}$. Roughly, a fully-right veering pseudoperiodic automorphism is a right-veering pseudoperiodic automorphism whose restrictions to each periodic piece are also right-veering.
Main theorem {#sec:main_thm}
============
In this section we prove that all fully right-veering pseudoperiodic automorphisms admit positive factorizations. We start by proving a proposition that has an interest of its own and that plays a key role in the proof of .
\[prop:extension\_to\_family\] Let $\omega:\bar{\mathcal{X}}\to Q$ be a deformation of a strongly pseudoconvex manifold $\bar X:=\omega^{-1}(0)$ with smooth boundary, where $Q \subset {\mathbb{C}}$ is a disk. And let $f:\bar X
\to
{\mathbb{C}}$ be a holomorphic map. Then there exists a neighborhood $Q'
\subset Q$ of $0$ and a $C^1$ map $F:\omega^{-1}(Q') \to {\mathbb{C}}$ such that $F|_{\omega^{-1}(t)}$ is holomorphic for all $t \in Q'$ and $F|_{\omega^{-1}(0)}=f$.
We observe that the map $\omega:\bar {\mathcal{X}}\to Q$ is actually a ${\mathcal{C}}^\infty$ locally trivial fibration and, since $Q$ is contractible, it is actually a trivial fibration. So there exists a diffeomorphism $ \bar X
\times Q \to \bar{{\mathcal{X}}}$ which induces diffeomorphisms $T_t:\bar{X}_t \to
\bar{X}$ with $T_0 = {\mathop{\mathrm{id}}\nolimits}$. This allows us to see the deformation $\omega:\bar{\mathcal{X}}\to Q$ as a family of complex structures $J_t$ on $\bar{X}$ varying smoothly on $t \in Q$. For instance, if $J_t'$ is the complex structure of $\bar{X}_t$, then one can consider $J_t := T_{t \ast} J_t'$ as the corresponding complex structure on $\bar{X}$. And by letting $\bar{\partial}_t'$ be the complex conjugate derivative in $\bar{X}_t$ and $\bar{\partial}_t$ be the corresponding one in $\bar X$ with the complex structure $J_t$, we find that if $g,f:X \to {\mathbb{C}}$ satisfy the equation $$\bar{\partial}_t g = \bar{\partial}_t f$$ then, by precomposing, the equation $$\bar{\partial}_t' (g\circ T_t) = \bar{\partial}_t' (f\circ T_t)$$ holds in the cohomology ring of $\bar{X}_t$ because $(\bar{X},J_t)$ and $(\bar{X}_t, J_t')$ are isomorphic as complex manifolds via $T_t$. This discussion tells us that we can solve our problem in the central fiber of our deformation equipped with different complex structures varying smoothly on $t$ and that this is equivalent to solving the extension problem in each fiber of the deformation as the statement of the proposition requires.
By definition, $\bar{\partial}_t f$ is a $\bar{\partial}$-closed $(0,1)$-form for all $t$. Since $f$ is smooth, we find that $\bar{\partial}_tf$ has in particular $C^1$ coefficients. Therefore, we can apply Greene and Krantz result () and get that there exists a family of solutions $\{u_t:\bar{X} \to {\mathbb{C}}\}$ for the $\bar{\partial}$ problems $$\bar{\partial}_t u_t = \bar{\partial}_t f.$$ Moreover, since the family of complex structures is ${\mathcal{C}}^\infty$, we can assure that our family of solutions is at least $C^1$ on the parameter $t$ for $t$ varying in a small disk $Q'\subset Q$.
The Henkin solution in the central fiber is $0$ since $f$ is holomorphic, i.e. $u_0 =
0$. Finally we define the sought $F:\omega^{-1}(Q')\to {\mathbb{C}}$ fiberwise by $F|_{\bar{X}_t}:=f \circ T_t - u_t \circ T_t = (f-u_t)\circ T_t$. By construction, $\bar{\partial}'_t F|_{\bar{X}_t}=0$ (so each restriction is holomorphic). Also by construction $F|_{\bar{X}_0}=f$ because $T_0={\mathop{\mathrm{id}}\nolimits}$ and $u_0=0$. And finally, $F$ is at least $C^1$ on the $t$ parameter because $u_t$ is $C^1$ and because $f$ and $T_t$ are smooth.
Now we prove the main theorem of the paper.
\[thm:positive\_factorization\] The following equivalent classes of pseudoperiodic automorphisms $\phi: {\Sigma}\to {\Sigma}$ admit positive factorizations:
1. \[it:i\] Fully right-veering automorphisms.
2. \[it:ii\] Monodromies associated with reduced holomorphic map germs defined on isolated complex surface singularities.
3. \[it:iii\] Pseudoperiodic automorphisms such that there exists $n\in {\mathbb{N}}$ such that $\phi^n$ is a composition of powers of right-handed Dehn twists around disjoint simple closed curves including all boundary components.
First of all we recall that, indeed, the three classes of automorphisms coincide. Class \[it:iii\] is the definition of \[it:i\] (see ). And that \[it:ii\] is the same class of automorphisms as \[it:iii\] is the content of Anne Pichon’s result ().
So given such a pseudoperiodic automorphism $\phi:{\Sigma}\to {\Sigma}$, there exists an isolated complex surface singularity $X'$ and a holomorphic function $f':X' \to {\mathbb{C}}$ such that the monodromy of the corresponding Milnor-Lê fibration on the tube $$\label{eq:milnor_le_fibration}
f'_{|f'^{-1}(\partial D_\delta) \cap X'\cap B_\epsilon}: f'^{-1}(\partial D_\delta) \cap X' \cap B_\epsilon \to \partial D_\delta$$ has fiber diffeomorphic to ${\Sigma}$ and monodromy conjugate to $\phi$ in ${\mathrm{Mod}_{g,r}}$ .
Let $\bar X'$ be a Milnor representative of $X'$, that is, $\bar X':= X'
\cap B_\epsilon$ where $B_\epsilon$ is a closed ball of small radius that intersects $X'$ transversely. We can assume as well that $\partial \bar{X}'$ is strongly pseudoconvex (recall ). The minimal resolution $X \to
X'$ induces a map $\pi:\bar X \to \bar X'$ where $\bar X$ is a strongly pseudoconvex surface with smooth boundary that contains a nontrivial $1$-dimensional exceptional set $A= \cup_j A_j$ such that $A_j\cdot A_j <
-1$ for all $j$. By Laufer and Bogomolov-De Oliveira’s result () there exists a deformation $\omega:\bar {\mathcal{X}}\to Q$ of $\bar X=\omega^{-1}(0)$ over a small disk such that each fiber $\bar X_t:=\omega^{-1}(t)$ does not contain compact analytic curves. By taking the deformation disk $Q$ small enough, we can ensure that the boundaries $\partial \bar X_t$ are strongly pseudoconvex. Indeed, the central fiber has a strongly pseudoconvex boundary and this is a property which is stable under small perturbations. Hence, we can assume that $\bar
X_t$ is a Stein domain for all $t \neq 0$.
Since the morphism $\pi:\bar X \to \bar X'$ induced by the minimal resolution is an isomorphism outside the singular point of $X'$, the Milnor-Lê fibration of \[eq:milnor\_le\_fibration\] lifts to an equivalent fibration $$\label{eq:milnor_le_res}
f_{|f^{-1}(\partial D_\delta) \cap \bar X}:f^{-1}(\partial D_\delta) \cap \bar X \to \partial D_\delta$$
Now we apply to our deformation $\omega$ and we get a smaller disk $Q' \subset Q$ and a map $F: \omega^{-1}(Q') \to Q'$ that extends $f:X \to {\mathbb{C}}$. Moreover, that same proposition tells us that the restriction of $F$ to $\bar{X}_t$ is holomorphic for all $t \in Q'$ and the first order partial derivatives of $F$ in directions tangent to the fibers of $\omega$ vary continuously with $t$ because $F$ is $C^1$.
Let $f_t:=F|_{\bar X_t}$. Since transversality is an open condition and the partial derivatives of the functions $f_t$ vary continuously on $t$ we find that there exists a (maybe smaller) disk $Q'' \subset Q$ containing $0$ such that the maps
$$\label{eq:milnor_t_fibration} f_{t|f_t^{-1}(\partial D_\delta) \cap \bar X_t}:f_t^{-1}(\partial D_\delta) \cap \bar X_t \to \partial D_\delta$$
are locally trivial fibrations equivalent to \[eq:milnor\_le\_res\] (and so to \[eq:milnor\_le\_fibration\]) for all $t \in Q''$.
Again, by maybe shrinking a little bit more $Q''$, we can assert that $f_t$ has no critical points on $\partial \bar X_t$ because $f$ did not have critical points on $\partial \bar{X}$ and being a submersion is a stable property. Since $\bar X_t$ is a Stein domain, it cannot contain compact analytic curves. So the map from \[eq:milnor\_t\_fibration\] only has isolated critical points in a smooth ambient space. Each of these isolated critical points can be locally seen as an isolated plane curve singularity. We conclude by a direct application of Picard-Lefschetz theory that the monodromy of \[eq:milnor\_t\_fibration\] admits a positive factorization since it is a composition of monodromies of isolated plane curve singularities. Then the monodromy of the equivalent fibrations \[eq:milnor\_le\_res\] and \[eq:milnor\_le\_fibration\] also admit positive factorizations.
In [@KoII], Honda, Kazez and Matić proved that every right-veering and freely periodic automorphism of the punctured torus admits a positive factorization. This proof is carried out by checking each case since the list of freely periodic automorphisms of the torus is short. Next we observe that the previous theorem is a vast generalization of this fact, in particular we obtain the following straightforward corollary.
Freely periodic automorphisms with positive fractional Dehn twist coefficients admit positive factorizations.
Consequences of the main theorem {#sec:consequences}
================================
In this section we define an invariant of a mapping class $\phi \in {\mathrm{Mod}_{g,r}}$ for all $g,r$ except for a small family of degenerate cases. This invariant is a non-negative integer that measures how far is a mapping class from being in ${\mathrm{Dehn}^+_{g,r}}$. This together with our main result () gives sufficient conditions for a pseudoperiodic automorphism to be in ${\mathrm{Dehn}^+_{g,r}}$.
For all $g \geq 0$ and $r=0$ all the elements in the pure mapping class group ${\mathrm{PMod}_{g,r}}$ (recall ) admit a positive factorization. A proof of this statement can be found in [@Farb page 124] under the name of “A strange fact”. In the next proposition we observe that a similar argument together with a recent result from [@Horn] implies a much more general result.
First we introduce a notion that we use in this section. For a surface ${\Sigma}_{g,r}$ denote by ${\mathcal{B}}$ the mapping class resulting of the composition of a single right-handed Dehn twist around each boundary component, that is, ${\mathcal{B}}:= t_{B_1} \cdots t_{B_r}$. We call ${\mathcal{B}}$ a [**boundary multitwist**]{}.
\[prop:pos\_fact\] Every element in the pure mapping class group ${\mathrm{PMod}_{g,r}}$ can be written as a composition of right-handed Dehn twists along non-separating simple closed curves for all $g \geq 1$ and $r \geq 1 $ except for the cases when $g=1$ and $r > 9$.
We show that one can use the same trick used by Farb and Margalit in [@Farb page 124]. Observe that [@Horn Theorem A and B] implies that for all $g \geq 1, r \geq 1$ (except in the exceptional cases mentioned in the statement) there exists in ${\mathrm{Mod}_{g,r}}$ a positive factorization of a power of a boundary multitwist.
A boundary multitwist ${\mathcal{B}}$ is the identity when seen in ${\mathrm{PMod}_{g,r}}$. So we find that the identity $\hat{{\mathop{\mathrm{id}}\nolimits}} \in {\mathrm{PMod}_{g,r}}$ can be expressed as a (non-empty) product of right-handed Dehn twists along non-separating simple closed curves $\gamma_1, \ldots, \gamma_k$. That is $${\hat{t}}_{\gamma_k} \cdots {\hat{t}}_{\gamma_1} = \hat{\mathop{\mathrm{id}}\nolimits}.$$ Therefore, multiplying both sides on the right by ${\hat{t}}_{\gamma_1}^{-1}$ yields a positive factorization of a left-handed Dehn twist around a non-separating simple closed curve. This together with the Change of Coordinates Principle [@Farb pg 37] which says that all non-separating simple closed curves are the same up to conjugation and the fact that ${\mathrm{PMod}_{g,r}}$ is generated by (right and left - handed) Dehn twists along non-separating simple closed curves gives the result. Just observe that we can substitute any left handed Dehn twist in a factorization in ${\mathrm{PMod}_{g,r}}$ by a composition of right-handed Dehn twists.
As a consequence we obtain the following corollary, which shows that composing with enough positive boundary Dehn twists stabilizes the whole mapping class group ${\mathrm{Mod}_{g,r}}$ into ${\mathrm{Dehn}^+_{g,r}}$.
\[cor:obstruc\_dehn\] Let $\phi \in {\mathrm{Mod}_{g,r}}$ and let $t_{B_{1}}, \ldots, t_{B_r}$ be right-handed boundary Dehn twists. Then for $n_1, \ldots, n_r \in {\mathbb{Z}}_{\geq0}$ big enough the element $$t_{B_{r}}^{n_r} \cdots t_{B_{1}}^{n_1}\phi$$ is in ${\mathrm{Dehn}^+_{g,r}}$ except in the cases $g=1$ and $r > 9$.
Suppose that $\phi$ cannot be written as a composition of right-handed Dehn twists. Take its image $\hat\phi$ in ${\mathrm{PMod}_{g,r}}$. By , $\hat\phi$ can be expressed as a product of right-handed Dehn twists in ${\mathrm{PMod}_{g,r}}$.
Take the composition of these right-handed Dehn twists in ${\mathrm{Mod}_{g,r}}$ and denote the resulting mapping class by $\tilde{\phi}$. Observe that $\phi \circ \tilde{\phi}^{-1}$ is a composition of powers of boundary Dehn twists and some of them are left-handed Dehn twists. Just compose $\phi$ with the inverse of these powers.
It is worth noting that it is a trivial observation that derives from the definition of right-veering automorphism in [@KoI] that composing with enough right-handed boundary Dehn twists, one can get any automorphism to be right-veering. The above corollary says that, except in the [*degenerate cases*]{}, composing with enough right-handed Dehn twists puts any monodromy in ${\mathrm{Dehn}^+_{g,r}}$!
The previous corollary contains the idea that there is an obstruction for an automorphism to be in ${\mathrm{Dehn}^+_{g,r}}$ localized at the fractional Dehn twist coefficients of $\phi$. So, by means of the following is well-defined.
\[def:correcting\] Let ${\Sigma}_{g,r}$ be a surface such that $g > 1$ or with $g=1$ and $r\leq 9$. Let $\phi \in {\mathrm{Mod}_{g,r}}$, we define $N_\phi$ as the minimal natural number such that $${\mathcal{B}}^{N_\phi}\phi \in {\mathrm{Dehn}^+_{g,r}}.$$ We call $N_\phi$ the [**correcting exponent**]{} of $\phi$.
\[def:essential\_element\] We say that a pseudoperiodic mapping class $\phi \in {\mathrm{Mod}_{g,r}}$ is an [**essential**]{} mapping class ${\mathrm{Mod}_{g,r}}$ if $|{\mathrm{fr}}(\phi,B_i)|<1$ and $|\mathrm{sc}(\phi, {\mathcal{C}}_i)| < 1$ for all boundary components $B_i$ and all orbits of invariant curves ${\mathcal{C}}$ in its Nielsen-Thurston decomposition.
\[lem:essential\_part\] Let ${\Sigma}_{g,r}$ be a surface with $g,r > 0$ and let $\phi \in {\mathrm{Mod}_{g,r}}$ be pseudoperiodic. Assume that $\phi$ has exactly $s$ different orbits of invariant curves in its Nielsen-Thurston decomposition and let $C_1, \ldots,C_s$ be a curve for each orbit. Then there exist unique integers $n_1, \ldots, n_r, m_1, \ldots, m_s$ such that
1. \[it:es\_i\]The mapping class $\tilde{\phi}:=\phi \cdot t_{B_{1}}^{n_1}\cdots
t_{B_{r}}^{n_r} t_{C_{1}}^{m_1}\cdots t_{C_{s}}^{m_s}$ is essential.
2. \[it:es\_ii\]If ${\mathrm{fr}}(\tilde\phi,B_i) \neq 0$, its sign is the same as ${\mathrm{fr}}(\phi,B_i)$.
3. \[it:es\_iii\] If $\mathrm{sc}(\tilde \phi,C_i) \neq 0$ its sign is the same as $\mathrm{sc}(\phi,C_i)$.
Let $x \in {\mathbb{Q}}$ and denote $Z(x)\in {\mathbb{Z}}$ its integer part. Then we just define $n_i:=-Z({\mathrm{fr}}(\phi,B_i))$ and $m_i:= -Z (\mathrm{sc}(\tilde \phi,C_i))$. From these it is straightforward to check that \[it:es\_i\], \[it:es\_ii\] and \[it:es\_iii\] are satisfied.
After the previous lemma, the following is a natural definition.
\[def:essential\_part\] For a pseudoperiodic mapping class $\phi$, we call the conjugacy class of the mapping class of $\tilde{\phi}$ defined by its [**essential part**]{}.
The screw numbers and fractional Dehn twist coefficients are conjugacy and isotopy invariants. Since the statement of does not impose conditions on the curves that we are selecting for each orbit, it only makes sense to define the essential part associated to a given pseudoperiodic automorphism as a conjugacy class.
Observe that $\phi$ is essential if and only if $\phi$ is conjugate to $\tilde{\phi}$. Putting together the results of this section yields the following sufficiency criterion that applies to certain pseudoperiodic mapping classes and decides if they are in ${\mathrm{Dehn}^+_{g,r}}$.
\[thm:criterion\] Let $\phi \in {\mathrm{Mod}_{g,r}}$ be a pseudoperiodic automorphism with ${\mathrm{fr}}(\phi,B_i) >0$ for all $i=1, \ldots,r$. Let us denote by ${\mathrm{sc}}_1, \ldots, {\mathrm{sc}}_s < 0$ its negative screw numbers and assume that the corresponding orbits of curves are not separating. Let $k =1$ if $g=1, r <9$ or $g \geq 2,r \leq 2g - 4$ and $k=2$ if $g \geq 2, r > 2g - 4$. If $$\sum_{i=1}^s k(|Z({\mathrm{sc}}_i)|+1) \leq \min_{j\in\{1, \ldots, r\}} {\mathrm{fr}}(\phi, B_j)$$ then $\phi$ admits a positive factorization.
First observe that composing $\phi$ with a right-handed Dehn twist around any curve of an orbit of $\mathcal{C}$ increases the screw number of that orbit by $1$.
Putting together the proof of with [@Horn Theorems (A) and (B))] we get that given a non-separating simple closed curve $\gamma$, there exist $c$ non-separating simple closed curves $\alpha_1, \ldots, \alpha_c$ such that $${\mathcal{B}}^k t_{\gamma}^{-1}= t_{\alpha_1} \cdots t_{\alpha_c}$$ holds in ${\mathrm{Mod}_{g,r}}$ where $k =1$ if $g=1, r <9$ or $g \geq 2,r \leq 2g - 4$ and $k=2$ if $g \geq 2, r \leq 2g - 4$ (this numerical conditions are contained in the cited theorems in [@Horn]). That is, we can say that a left-handed Dehn twist, [*costs*]{} a boundary multitwist if $g=1, r <9$ or $g \geq 2,r \leq 2g - 4$ and that it [*costs*]{} $2$ boundary multitwists if $g \geq 2, r \leq 2g - 4$.
Boundary multitwists commute with each other element in ${\mathrm{Mod}_{g,r}}$ because their support is disjoint from any other simple closed curve. So let $\gamma$ be a curve in ${\mathcal{C}}$ contained in an orbit with negative screw number and consider ${\mathcal{B}}^{-k}t_\gamma\phi$ (where $k$ is as in the hypothesis). This automorphism is [*like*]{} $\phi$ but with one fractional Dehn twist coefficient [*one smaller*]{} than $\phi$ and the screw number at the orbit of $\gamma$ [*one bigger*]{}. If the inequality of the hypothesis is satisfied we can iterate this process and get to a fully right-veering automorphism. Then applies.
We finish the article by introducing a natural refinement of the correcting exponent (recall ) which we think it could be useful for future exploration of the set ${\mathrm{Veer}_{g,r}}\setminus {\mathrm{Dehn}^+_{g,r}}$.
Let $\phi \in {\mathrm{Mod}_{g,r}}$ and let $\hat{\phi}$ be its image in ${\mathrm{PMod}_{g,r}}$. We define $$\Delta_\phi:=\left\{\prod t_{\gamma_i} | \prod \hat{t}_{\gamma_i} = \hat \phi \right\}$$ that is, the collection of all preimages in ${\mathrm{Mod}_{g,r}}$ of positive factorizations of $\hat{\phi}$ in ${\mathrm{PMod}_{g,r}}$. We know that this is non-empty in the majority of cases by .
Fix an order of the boundary components of ${\Sigma}_{g,r}$. That is, we have a labeling $B_1, \ldots, B_r$.
\[def:lattice\] Let $\phi \in {\mathrm{Mod}_{g,r}}$, we define the lattice $$L_{\phi}:= \left\{(a_1, \ldots, a_r) \in {\mathbb{Z}}^r | (a_1, \ldots, a_r) = {\mathrm{fr}}(\phi \circ \Delta^{-1}) \text{ for some } \Delta \in \Delta_\phi\right\}.$$ Which is well defined since the rational parts of ${\mathrm{fr}}(\Delta)$ and ${\mathrm{fr}}(\phi)$ coincide because $\hat{\Delta} = \hat{\phi}$. We call $L_{\phi}$ the [**correcting lattice**]{} of $\phi$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Genetic differences between individuals associated to quantitative phenotypic traits, including disease states, are usually found in non-coding genomic regions. These genetic variants are often also associated to differences in expression levels of nearby genes (they are “expression quantitative trait loci” or eQTLs, for short) and presumably play a gene regulatory role, affecting the status of molecular networks of interacting genes, proteins and metabolites. Computational systems biology approaches to reconstruct causal gene networks from large-scale omics data have therefore become essential to understand the structure of networks controlled by eQTLs together with other regulatory genes, as well as to generate detailed hypotheses about the molecular mechanisms that lead from genotype to phenotype. Here we review the main analytical methods and softwares to identify eQTLs and their associated genes, to reconstruct co-expression networks and modules, to reconstruct causal Bayesian gene and module networks, and to validate predicted networks *in silico*.'
author:
- 'Lingfei Wang$^{1}$ and Tom Michoel$^{1,\ast}$'
title: Detection of regulator genes and eQTLs in gene networks
---
$^1$Division of Genetics and Genomics, The Roslin Institute, The University of Edinburgh, Midlothian EH25 9RG, Scotland, United Kingdom
$^\ast$Corresponding author, E-mail: [email protected]
Introduction
============
![A flow chart for a typical systems genetics study and the corresponding softwares. Steps in light yellow are covered in this chapter.\[fig-flow\]](fig1.pdf)
Genetic differences between individuals are responsible for variation in the observable phenotypes. This principle underpins genome-wide association studies (GWAS), which map the genetic architecture of complex traits by measuring genetic variation at single-nucleotide polymorphisms (SNPs) on a genome-wide scale across many individuals [@mackay2009genetics]. GWAS have resulted in major improvements in plant and animal breeding [@goddard2009mapping] and in numerous insights into the genetic basis of complex diseases in human [@manolio2013bringing]. However, quantitative trait loci (QTLs) with large effects are uncommon and a molecular explanation for their trait association rarely exists [@mackay2009genetics]. The vast majority of QTLs indeed lie in non-coding genomic regions and presumably play a gene regulatory role [@hindorff2009potential; @schaub2012linking]. Consequently, numerous studies have identified *cis*- and *trans*-acting DNA variants that influence gene expression levels (i.e., “expression QTLs”; eQTLs) in model organisms, plants, farm animals and human (reviewed in [@rockman2006genetics; @georges2007mapping; @cookson2009mapping; @cheung2009genetics; @cubillos2012lessons]). Gene expression programmes are of course highly tissue- and cell-type specific, and the properties and complex relations of eQTL associations across multiple tissues are only beginning to be mapped [@dimas2009common; @foroughi2015; @greenawalt2011survey; @ardlie2015genotype]. At the molecular level, a mounting body of evidence shows that *cis*-eQTLs primarily cause variation in transcription factor (TF) binding to gene regulatory DNA elements, which then causes changes in histone modifications, DNA methylation and mRNA expression of nearby genes; *trans*-eQTLs in turn can usually be attributed to coding variants in regulatory genes or *cis*-eQTLs of such genes [@albert2015role].
Taken together, these results motivate and justify a systems biological view of quantitative genetics (“systems genetics”), where it is hypothesized that genetic variation, together with environmental perturbations, affects the status of molecular networks of interacting genes, proteins and metabolites; these networks act within and across different tissues and collectively control physiological phenotypes [@williams2006expression; @kadarmideen2006genetical; @rockman2008reverse; @schadt2009; @schadt2012new; @civelek2014systems; @bjorkegren2015genome]. Studying the impact of genetic variation on gene regulation networks is of crucial importance in understanding the fundamental biological mechanisms by which genetic variation causes variation in phenotypes [@chen2008], and is expected to lead to the discovery of novel disease biomarkers and drug targets in human and veterinary medicine [@schadt2009b]. Since direct experimental mapping of genetic, protein–protein or protein–DNA interactions is an immensely challenging task, further exacerbated by the cell-type specific and dynamic nature of these interactions [@walhout2006unraveling], comprehensive, experimentally verified molecular networks will not become available for multi-cellular organisms in the foreseeable future. Statistical and computational methods are therefore essential to reconstruct trait-associated causal networks by integrating diverse omics data [@rockman2008reverse; @schadt2009; @ritchie2015methods].
A typical systems genetics study collects genotype and gene, protein and/or metabolite expression data from a large number of individuals segregating for one or more traits of interest. After raw data processing and normalization, eQTLs are identified for each of the expression data types, and a co-expression matrix is constructed. Causal Bayesian gene networks, co-expression modules (i.e. clusters) and/or causal Bayesian module networks are then reconstructed. *In silico* validation of predicted networks and modules using independent data confirms their overall validity, ideally followed by experimental validation of the most promising findings in a relevant cell line or model organism (Figure \[fig-flow\]). Here we review the main analytic principles behind each of the steps from eQTL identification to *in silico* network validation, and present a selection of most commonly used methods and softwares for each step. Throughout this chapter, we tacitly assume that all data has been quality controlled, pre-processed and normalized to suit the assumptions of the analytic methods presented here. For expression data, this usually means working with log-transformed data where each gene expression profile is centred around zero with standard deviation one. We also assume that the data has been corrected for any confounding factors, either by regressing out known covariates and/or by estimating hidden factors [@stegle2012using].
Genetics of gene expression {#sec:genet-gene-expr}
===========================
A first step towards identifying molecular networks affected by DNA variants is to identify variants that underpin variations in eQTLs of transcripts [@cookson2009mapping], proteins [@foss2007] or metabolites [@nicholson2011genome] across individuals. When studying a single trait, as in GWAS, it is possible to consider multiple statistical models to explicitly account for additive and/or dominant genetic effects [@laird2011]. However, when the possible effects of a million or more SNPs on tens of thousands of molecular abundance traits need to be tested, as is common in modern genetics of gene expression studies, the computational cost of testing SNP-trait associations one-by-one becomes prohibitive. To address this problem, new methods have been developed to calculate the test statistics for the parametric linear regression and analysis of variance (ANOVA) models [@shabalin2012matrix] and the non-parametric ANOVA model (or Kruskal-Wallis test) [@qi2014] using fast matrix multiplication algorithms, implemented in the softwares **matrix-eQTL** (<http://www.bios.unc.edu/research/genomic_software/Matrix_eQTL/>) [@shabalin2012matrix] and **kruX** (<https://github.com/tmichoel/krux>) [@qi2014].
In both softwares, genotype values of $s$ genetic markers and expression levels of $k$ transcripts, proteins or metabolites in $n$ individuals are organized in an $s\times n$ genotype matrix ${\mathbf{G}}$ and $k\times n$ expression data matrix ${\mathbf{X}}$. Genetic markers take values $0,1,\dots,\ell$, where $\ell$ is the maximum number of alleles ($\ell=2$ for biallelic markers), while molecular traits take continuous values. In the linear model, a linear relation is tested between the expression level of gene $i$ and the genotype value (i.e. the number of reference alleles) of SNP $j$. The corresponding test statistic is the Pearson correlation between the $i$th row of ${\mathbf{X}}$ and the $j$th row of ${\mathbf{G}}$, for all values of $i$ and $j$. Standardising the data matrices to zero mean and unit variance, such that for all $i$ and $j$, $$\begin{aligned}
\sum_{l=1}^nX_{il} = \sum_{l=1}^nG_{jl} = 0 \quad\text{and}\quad
\sum_{l=1}^n X_{il}^2 = \sum_{l=1}^n G_{jl}^2 = n , \end{aligned}$$ it follows that the correlation values can be computed as $$\begin{aligned}
R_{ij} = \sum_{l=1}^nX_{il}G_{jl} = ({\mathbf{X}}{\mathbf{G}}^T)_{ij},\end{aligned}$$ where ${\mathbf{G}}^T$ denotes the transpose of ${\mathbf{G}}$. Hence, a single matrix multiplication suffices to compute the test statistics for the linear model for all pairs of traits and SNPs.
The ANOVA models test if expression levels in different genotype groups originate from the same distribution. Therefore, ANOVA models can account for both additive and dominant effects of a genetic variant on expression levels. In the parametric ANOVA model, suppose the test samples are divided into $\ell+1$ groups by the SNP $j$. The mean expression level for gene $i$ in each group $m$ can be written as $$\begin{aligned}
\overline{X_i^{(m,j)}} = \frac1{n^{(m,j)}} \sum_{\{l\colon G_{jl}=m\}} X_{il},\end{aligned}$$ where $n^{(m,j)}$ is the number of samples in genotype group $m$ for SNP $j$.
Again assuming that the expression data is standardised, the F-test statistic for testing gene $i$ against SNP $j$ can be written as $$\begin{aligned}
F_i^{(j)} = \frac{n-\ell-1}{\ell} \frac{SS_i^{(j)}}{n-SS_i^{(j)}},\end{aligned}$$ where $SS_i^{(j)}$ is the sum of squares between groups, $$\begin{aligned}
SS_i^{(j)} = \sum_{m=0}^\ell n^{(m,j)}\overline{X_i^{(m,j)}}^2.\end{aligned}$$
Let us define the $n\times s$ indicator matrix ${\mathbf{I}}^{(m)}$ for genotype group $m$, i.e. ${\mathbf{I}}_{lj}^{(m)} = 1$ if $G_{jl}=m$ and $0$ otherwise. Then $$\begin{aligned}
\sum_{\{l\colon G_{jl}=m\}} X_{il} = \left({\mathbf{X}}{\mathbf{I}}^{(m)}\right)_{ij}.\end{aligned}$$ Hence, for each pair of expression level $X_i$ and SNP $G_j$, the sum of squares matrix $SS_i^{(j)}$ can be computed via $\ell-1$ matrix multiplications [^1].
In the non-parametric ANOVA model, the expression data matrix is converted to a matrix ${\mathbf{T}}$ of data ranks, independently over each row. In the absence of ties, the Kruskal-Wallis test statistic is given by $$\begin{aligned}
S_{ij} = \frac{12}{n(n+1)} \sum_{m=0}^\ell n^{(m,j)}\,\overline{T_i^{(m,j)}}^2 - 3(n+1),\end{aligned}$$ where $\overline{T_i^{(m,j)}}$ is the average expression rank of gene $i$ in genotype group $m$ of SNP $j$, defined as $$\begin{aligned}
\overline{T_i^{(m,j)}}&=\frac1{n^{(m,j)}} \sum_{\{l\colon G_{jl}=m\}}T_{il},\end{aligned}$$ which can be similarly obtained from the $\ell-1$ matrix multiplications.
There is as yet no consensus about which statistical model is most appropriate for eQTL detection. Non-parametric methods were introduced in the earliest eQTL studies [@brem2002; @schadt2008] and have remained popular, as they are robust against variations in the underlying genetic model and trait distribution. More recently, the linear model implemented in matrix-eQTL has been used in a number of large-scale studies [@lappalainen2013transcriptome; @ardlie2015genotype]. A comparison on a dataset of 102 human whole blood samples showed that the parametric ANOVA method was highly sensitive to the presence of outlying gene expression values and SNPs with singleton genotype group. Linear models reported the highest number of eQTL associations after empirical False Discovery Rate (FDR) correction, with an expected bias towards additive linear associations. The Kruskal-Wallis test was most robust against data outliers and heterogeneous genotype group sizes and detected a higher proportion of non-linear associations, but was more conservative for calling additive linear associations than linear models [@qi2014].
In summary, when large numbers of traits and markers have to be tested for association, efficient matrix multiplication methods can be employed to calculate all test statistics at once, leading to a dramatic reduction in computation time compared to calculating these statistics one-by-one for every pair using traditional methods. Matrix multiplication is a basic mathematical operation which has been purposely studied and optimized for tens of years [@golub1996]. Highly efficient packages, such as **BLAS** (<http://www.netlib.org/blas/>) and **LAPACK** (<http://www.netlib.org/lapack/>), are available for use on generic CPUs, and are indeed employed in most mainstream scientific computing softwares and programming languages, such as Matlab and R. In recent years, Graphics Processor Unit (GPU)-accelerated computing, such as CUDA, has revolutionised scientific calculations that involve repetitive operations in parallel on bulky data, offering even more speedup than the existing CPU-based packages. The first applications of GPU computing in eQTL analysis have already appeared (e.g. [@hemani2014detection]), and more can be expected in the future.
Lastly, for pairs exceeding a pre-defined threshold on the test statistic, a $p$-value can be computed from the corresponding test distribution, and these $p$-values can then be further corrected for multiple testing by common procedures [@shabalin2012matrix; @qi2014].
Co-expression networks and modules
==================================
Co-expression gene networks\[sec-coex\]
---------------------------------------
The Pearson correlation is the simplest and computationally most efficient similarity measure for gene expression profiles. For genes $i$ and $j$, their Pearson correlation can be written as $$\label{eq:1}
C_{ij}=\sum_{l=1}^nX_{il}X_{jl}\,.$$ In matrix notation, this can be combined as the matrix multiplication [$${\mathbf{C}}={\mathbf{X}}{\mathbf{X}}^T.$$]{}
Gene pairs with large positive or negative correlation values tend to be up- or down-regulated together, due to either a direct regulatory link between them, or being jointly co-regulated by a third, often hidden, factor. By filtering for correlation values exceeding a significance threshold determined by comparison with randomly permuted data, a discrete co-expression network is obtained. Assuming that a high degree of co-expression signifies that genes are involved in the same biological processes, graph theoretical methods can be employed, for instance, to predict gene function [@sharan2007network].
One drawback of the Pearson correlation is that by definition it is biased towards *linear* associations. To overcome this limitation, other measures are available. The Spearman correlation uses expression data ranks (cf. Section \[sec:genet-gene-expr\]) in Equation , and will give high score to *monotonic* relations. Mutual information is the most general measure and detects both linear and non-linear associations. For a pair of discrete random variables $A$ and $B$ (representing the expression levels of two genes) taking values $a_l$ and $b_m$, respectively, the mutual information is defined as $$MI(A,B)=H(A)+H(B)-H(A,B),$$ where $$\begin{aligned}
H(A) &= -\sum_l P(a_l)\log P(a_l),\\
H(B) &= -\sum_m P(b_m)\log P(b_m),\\
H(A,B) &= \sum_{lm} P(a_l,b_m) \log P(a_l,b_m),\end{aligned}$$ are the individual and joint Shannon entropies of $A$ and $B$, and $P(a_l)=P(A=a_l)$, and likewise for the other terms. Since gene expression data are continuous, mutual information estimation is non-trivial and usually involves some form of discretisation [@daub2004]. Mutual information has been successfully used as a co-expression measure in a variety of contexts [@butte2000; @basso2005; @faith2007].
Clustering and co-expression module detection {#sec:clust-co-expr}
---------------------------------------------
It is generally understood that cellular functions are carried out by “modules”, groups of molecules that operate together and whose function is separable from that of other modules [@hartwell1999]. Clustering gene expression data (i.e. dividing genes into discrete groups on the basis of similarities in their expression profiles) is a standard approach to detect such functionally coherent gene modules. The literature on gene expression clustering is vast and cannot possibly be reviewed comprehensively here. It includes “standard” methods such as hierarchical clustering [@eisen1998cluster], $k$-means [@tavazoie1999systematic], graph-based methods that operate directly on co-expression networks [@sharan2000click], and model-based clustering algorithms which assume that the data is generated by a mixture of probability distributions, one for each cluster [@medvedovic2002bayesian]. Here we briefly describe a few recently developed methods with readily available softwares.
#### Modularity maximization
Modularity maximization is a network clustering method that is particularly popular in the physical and social sciences, based on the assumption that intra-module connectivity should be much denser than inter-module connectivity [@newman2004; @newman2006b]. In the context of co-expression networks, this method can be used to identify gene modules directly from the correlation matrix ${\mathbf{C}}$ [@Ayroles:2009]. Suppose the genes are grouped into $N$ modules $M_l,~l=1,\dots,N$. Each module $M_l$ is a non-empty set that can contain any combination of the genes $i=1,\dots,k$, but each gene is contained by exactly one module. Also define $M_0$ as the set containing all genes. The modularity score function is defined as [$$S(M)=\sum_{l=1}^N\left(\frac{W(M_l,M_l)}{W(M_0,M_0)}-\left(\frac{W(M_l,M_0)}{W(M_0,M_0)}\right)^2\right),$$]{} where $W(A,B) =\sum_{i\in A,j\in B,i\ne j}w(C_{ij})$ is a weight function, summing over all the edges that connect one vertex in $A$ with another vertex in $B$, and $w(x)$ is a monotonic function to map correlation values to edge strengths. Common functions are $w(x)=|x|$, $|x|^\beta$ (power law) [@Langfelder:2008], $e^{\beta |x|}$ (exponential) [@Ayroles:2009], or $1/(1+e^{\beta x})$ (sigmoid) [@lee2009learning].
A modularity maximization software particularly suited for large networks is **Fast Modularity** (<http://www.cs.unm.edu/~aaron/research/fastmodularity.htm>) [@Clauset:2004].
#### Markov Cluster algorithm
The Markov Cluster (MCL) algorithm is a graph-based clustering algorithm, which emulates random walks among gene vertices to detect clusters in a graph obtained directly from the co-expression matrix ${\mathbf{C}}$. It is implemented in the **MCL** software (<http://micans.org/mcl/>) [@Van-Dongen:2001; @Enright:2002]. The MCL algorithm starts with the correlation matrix ${\mathbf{C}}$ as the probability flow matrix of a random walk, and then iteratively suppresses weak structures of the network and performs a multi-step random walk. In the end, only backbones of the network structure remain, essentially capturing the modules of co-expression network. To be precise, the MCL algorithm performs the following two operations on ${\mathbf{C}}$ alternatingly:
- **Inflation:** The algorithm first contrasts stronger direct connections against weaker ones, using an element-wise power law transformation, and normalizes each column separately to sum to one, such that the element $C_{ij}$ corresponds to the dissipation rate from vertex $X_i$ to $X_j$ in a single step. The inflation operation hence updates ${\mathbf{C}}$ as ${\mathbf{C}}\rightarrow\mathbf{\Gamma}_\alpha{\mathbf{C}}$, where the contrast rate $\alpha>1$ is a predefined parameter of the algorithm. After operation $\mathbf{\Gamma}_\alpha$, each element of ${\mathbf{C}}$ becomes [$$C_{ij}\rightarrow\mathbf{\Gamma}_\alpha
C_{ij}=|C_{ij}|^\alpha/\sum_{p=1}^k|C_{pj}|^\alpha.$$]{}
- **Expansion:** The probability flow matrix ${\mathbf{C}}$ controls the random walks performed in the expansion phase. After some integer $\beta\ge2$ steps of random walk, gene pairs with strong direct connections and/or strong indirect connections through other genes tend to see more probability flow exchanges, suggesting higher probabilities of belonging to the same gene modules. The expansion operation for the $\beta$-step random walk corresponds to the matrix power operation [$${\mathbf{C}}\rightarrow{\mathbf{C}}^\beta.$$]{}
The MCL algorithm performs the above two operations iteratively until convergence. Non-zero entries in the convergent matrix ${\mathbf{C}}$ connect gene pairs belonging to the same cluster, whereas all inter-cluster edges attain the value zero, so that cluster structure can be obtained directly from this matrix [@Van-Dongen:2001; @Enright:2002].
#### Weighted Gene Co-expression Network Analysis
With higher than average correlation or edge densities within clusters, genes from the same cluster typically share more neighbouring (i.e. correlated) genes. The weighted number of shared neighbouring genes hence can be another measure of gene function similarity. This information is captured in the so-called topological overlap matrix ${\mathbf{\Omega}}$, first defined in [@ravasz2002] for binary networks as [$$\omega_{ij}=\frac{A_{ij}+\sum_u
A_{iu}A_{uj}}{\mathrm{min}(k_i,k_j)+1-A_{ij}},$$]{} where $A$ is the (binary) adjacency matrix of the network and $k_i=\sum_uA_{iu}$ is the connectivity of vertex $X_i$. The $\sum_uA_{iu}A_{uj}$ term represents vertex similarity through neighbouring genes, and the rest of terms normalise the output as $0\le\omega_{ij}\le1$. This concept was later extended onto networks with weighted edges by applying a “soft threshold” pre-process on the correlation matrix, for example as $$\begin{aligned}
A_{ij}&=\left|\frac{1+C_{ij}}{2}\right|^\alpha,\\
\intertext{or}
A_{ij}&=\left|C_{ij}\right|^\alpha,\end{aligned}$$ such that $0\le A_{ij}\le1$ [@zhang2005b]. Note that in the first case only positive correlations have high edge weight, whereas in the second case positive and negative correlations are treated equally. The parameter $\alpha>1$ is determined such that the weighted network with adjacency matrix $A$ has approximately a scale-free degree distribution [@zhang2005b].
In principle, any clustering algorithm (including the aforementioned ones) can be applied to the topological overlap matrix ${\mathbf{\Omega}}$. In the popular **WGCNA** software (<http://labs.genetics.ucla.edu/horvath/htdocs/CoexpressionNetwork/Rpackages/WGCNA/>) [@Langfelder:2008], which is a multi-purpose toolbox for network analysis, hierachical clustering with a dynamic tree-cut algorithm [@langfelder2008defining] is employed.
#### Model-based clustering
Model-based clustering approaches assume that the observed data is generated by a mixture of probability distributions, one for each cluster, and takes explicitly into account the noise of gene expression data. To infer model parameters and cluster assignments, techniques such as Expectation Maximization (EM) or Gibbs sampling are used [@liu2002]. A recently developed method assumes that the expression levels of genes in a cluster are random samples drawn from a mixture of normal distributions, where each mixture component corresponds to a clustering of samples for that module, i.e. it performs a two-way co-clustering operation [@joshi2008]. The method is available as part of the **Lemon-Tree** package (<https://github.com/eb00/lemon-tree>) and has been successfully used in a variety of applications [@bonnet2015].
The co-clustering is carried out by a Gibbs sampler which iteratively updates the assignment of each gene and, within each gene cluster, the assignment of each experimental condition. The co-clustering operation results the full posterior distribution, which can be written as $$p(\mathcal{C}\mid{\mathbf{X}}) \propto \prod_{l=1}^N \prod_{u=1}^{L_l} \iint
p(\mu,\tau) \prod_{i\in\mathcal{M}_l}\prod_{m\in \mathcal{E}_{l,u}} p (X_{im}\mid
\mu,\tau)\; d\mu d\tau,$$ where $\mathcal{C}=\{M_l, \mathcal{E}_{l,u}\colon l=1,\dots,N;
u=1,\dots, L_l\}$ is a co-clustering consisting of $N$ gene modules $M_l$, each of which has a set of $L_m$ sample clusters as $\mathcal{E}_{l,u}$; $p(X_{im}\mid \mu,\tau)$ is a normal distribution function with mean $\mu$ and precision $\tau$; and $p(\mu,\tau)$ is a non-informative normal-gamma prior. Detailed investigations of the convergence properties of the Gibbs sampler showed that the best results are obtained by deriving consensus clusters from multiple independent runs of the sampler. In the **Lemon-Tree** package, consensus clustering is performed by a novel spectral graph clustering algorithm [@michoel2012] applied to the weighted graph of pairwise frequencies with which two genes are assigned to the same gene module [@bonnet2015].
Causal gene networks
====================
Using genotype data to prioritize edge directions in co-expression networks\[sec-direction\]
--------------------------------------------------------------------------------------------
Pairwise correlations between gene expression traits define undirected co-expression networks. Several studies have shown that pairs of gene expression traits can be causally ordered using genotype data [@zhu2004; @chen2007harnessing; @aten2008using; @schadt2005integrative; @neto2008inferring; @neto2013modeling; @millstein2009disentangling]. Although varying in their statistical details, these methods conclude that gene $A$ is causal for gene $B$, if expression of $B$ associates significantly with $A$’s eQTLs and this association is abolished by conditioning on expression of $A$ and on any other known confounding factors. In essence, this is the principle of “Mendelian randomization”, first introduced in epidemiology as an experimental design to detect causal effects of environmental exposures on human health [@smith2003mendelian], applied to gene expression traits.
To illustrate how these methods work, let $A$ and $B$ be two random variables representing two gene expression traits, and let $E$ be a random variable representing a SNP which is an eQTL for gene $A$ and $B$. Since genotype cannot be altered by gene expression (i.e. $E$ cannot have any incoming edges), there are three possible regulatory models to explain the joint association of $E$ to $A$ and $B$:
1. $E\rightarrow A\rightarrow B$: the association of $E$ to $B$ is indirect and due to a causal interaction from $A$ to $B$.
2. $E\rightarrow B\rightarrow A$: idem with the roles of $A$ and $B$ reversed.
3. $A\leftarrow E\rightarrow B$: $A$ and $B$ are independently associated to $E$.
To determine if gene $A$ mediates the effect of SNP $E$ on gene $B$ (model 1), one can test whether conditioning on $A$ abolishes the correlation between $E$ and $B$, using the partial correlation coefficient $$\begin{aligned}
{\mathrm{cor}(E,B\mid A)} =
\frac{{\mathrm{cor}(E,B)}-{\mathrm{cor}(E,A)}{\mathrm{cor}(B,A)}}{\sqrt{(1-{\mathrm{cor}(E,A)}^2)(1-{\mathrm{cor}(B,A)}^2)}.} \end{aligned}$$ If model 1 is correct, then ${\mathrm{cor}(E,B\mid A)}$ is expected to be zero, and this can be tested for example using Fisher’s $Z$ transform to assess the significance of a sample correlation coefficient. The same approach can be used to test model 2, and if neither is significant, it is concluded that no inference on the causal direction between $A$ and $B$ can be made (using SNP $E$), i.e. that model 3 is correct. For more details, see [@aten2008using], who have implemented this approach in the **NEO** software (<http://labs.genetics.ucla.edu/horvath/htdocs/aten/NEO/>).
Other approaches are based on the same principle, but use statistical model selection to identify the most likely causal model, with the probability density functions (PDF) for the models below:
- $p(E,A,B)=p(E)p(A\mid E)p(B\mid A)$,
- $p(E,A,B)=p(E)p(B\mid E)p(A\mid B)$,
- $p(E,A,B)=p(E)p(A\mid E) p(B\mid E,A)$,
where the dependence on $A$ in the last term of the last model indicates that there may be a residual correlation between $B$ and $A$ not explained by $E$. The minimal additive model assumes the distributions are [@schadt2005integrative] [$$\begin{aligned}
E&\sim&\mathrm{Bernoulli}(q),\nonumber\\
A\mid E&\sim&\mathrm{N}(\mu_{A\mid E},\sigma_A^2),\nonumber\\
B\mid A&\sim&\mathrm{N}\left(\mu_B+\rho\frac{\sigma_B}{\sigma_A}(A-\mu_A),(1-\rho^2)\sigma_B^2\right),\nonumber\\
B\mid E,A&\sim&\mathrm{N}\left(\mu_{B\mid E}+\rho\frac{\sigma_B}{\sigma_A}(A-\mu_{A\mid E}),(1-\rho^2)\sigma_B^2\right),\nonumber\end{aligned}$$]{} so that $E$ fulfills a Bernoulli distribution, $A\mid E$ undergoes a normal distribution whose mean depends on $E$, and that $B\mid A$ has a conditional normal distribution whose mean and variance are contributed in part by $A$. For $(B\mid E,A)$, the mean of $B$ also depends on $E$. The parameters of all distributions can be estimated by maximum likelihood, and the model with the highest likelihood is selected as the most likely causal model. The number of free parameters can be accounted using penalties like the Akaike information criterion (AIC) [@schadt2005integrative].
The approach has been extended in various ways. In [@chen2007harnessing], likelihood ratio tests, comparison to randomly permuted data, and false discovery rate estimation techniques are used to convert the three model scores in a single probability value $P(A\to B)$ for a causal interaction from gene $A$ to $B$. This method is available in the **Trigger** software (<https://www.bioconductor.org/packages/release/bioc/html/trigger.html>). In [@millstein2009disentangling] and [@neto2013modeling], the model selection task is recast into a single hypothesis test, using $F$-tests and Vuong’s model selection test respectively, resulting in a significance $p$-value for each gene-gene causal interaction.
It should be noted that all of the above approaches suffer from limitations due to their inherent model assumptions. In particular, the presence of unequal levels of measurement noise among genes, or of hidden regulatory factors causing additional correlation among genes, can confuse causal inference. For example, excessive error level in the expression data of gene $A$, may mistake the true structure $E\rightarrow A\rightarrow B$ as $E\rightarrow B\rightarrow A$. These limitations are discussed in [@rockman2008reverse; @li2010critical].
Using Bayesian networks to identify causal regulatory mechanisms
----------------------------------------------------------------
Bayesian networks are probabilistic graphical models which encode conditional dependencies between random variables in a directed acyclic graph (DAG). Although Bayesian network cannot fully reflect certain pathways in gene regulation, such as self-regulation or feedback loops, they still serve as a popular method for modelling gene regulation networks, as they provide a clear methodology for learning statistical dependency structures from possibly noisy data [@Friedman:1999; @Friedman:2000; @koller2009].
We adopt our previous convention in Section \[sec:genet-gene-expr\], where we have the gene expression data ${\mathbf{X}}$ and genetic markers ${\mathbf{G}}$. The model contains a total of $k$ vertices (i.e. random variables), $X_i$ with $i=1,\dots,k$, corresponding to the expression level of gene $i$. Given a DAG ${\mathcal{G}}$, and denoting the parental vertex set of $X_i$ by ${\mathbf{Pa}}^{({\mathcal{G}})}(X_i)$, the acyclic property of ${\mathcal{G}}$ allows to define the joint probability distribution function as $$\label{eq:2}
p(X_1,\dots,X_k\mid{\mathcal{G}})
=\prod_{i=1}^kp(X_i \mid {\mathbf{Pa}}^{({\mathcal{G}})}(X_i)).$$ In its simplest form, we model the conditional distributions as $$p\bigl(X_i\mid {\mathbf{Pa}}^{({\mathcal{G}})}(X_i)\bigl) = N\biggl(\alpha_i +
\sum_{X_j\in{\mathbf{Pa}}^{({\mathcal{G}})}(X_i)} \beta_{ji}(X_j-\alpha_j),\sigma_i^2\biggr),$$ where $(\alpha_i,\sigma_i)$ and $\beta_{ji}$ are parameters for vertex $X_i$ and edge $X_j\rightarrow X_i$ respectively, as part of the DAG structure ${\mathcal{G}}$. Under such modelling, the Bayesian network is called a linear Gaussian network.
The likelihood of data ${\mathbf{X}}$ given the graph ${\mathcal{G}}$ is $$\begin{aligned}
p({\mathbf{X}}\mid{\mathcal{G}}) = \prod_{i=1}^k\prod_{l=1}^n
p(X_{il}\mid \{X_{jl}, X_j\in{\mathbf{Pa}}^{({\mathcal{G}})}(X_i)\}).\end{aligned}$$ Using Bayes’ rule, the log-likelihood of the DAG ${\mathcal{G}}$ based on the gene expression data ${\mathbf{X}}$ becomes [$$\log p({\mathcal{G}}\mid{\mathbf{X}})=\log
p({\mathbf{X}}\mid {\mathcal{G}})+\log p({\mathcal{G}})-\log p({\mathbf{X}}),$$]{} where $p({\mathcal{G}})$ is the prior probability for ${\mathcal{G}}$, and $p({\mathbf{X}})$ is a constant when the expression data is provided, so the follow-up calculations do not rely on it.
Typically, a locally optimal DAG is found by starting from a random graph and randomly ascending the likelihood by adding, modifying, or removing one directed edge at a time [@Friedman:1999; @Friedman:2000; @koller2009]. Alternatively, the posterior distribution $p({\mathcal{G}}\mid{\mathbf{X}})$ can be estimated with Bayesian inference using Markov Chain Monte Carlo (MCMC) simulation, allowing us to estimate the significance levels at an extra computational cost. The parameter values of $\alpha$, $\beta$, and $\sigma$, as part of ${\mathcal{G}}$, can be estimated with maximum likelihood.
When Bayesian network is modified by a single edge, only the vertices that receive a change would require a recalculation, whilst all others remain intact. This significantly reduces the amount of computation needed for each random step. A further speedup is achievable if we constrain the maximum number of parents each vertex can have, either by using the same fixed number for all nodes, or by pre-selecting a variable number of potential parents for each node using, for instance, a preliminary $L_1$-regularisation step [@schmidt2007learning].
Two DAGs are called Markov equivalent if they result in the same PDF [@koller2009]. Clearly, using gene expression data alone, Bayesian networks can only be resolved up to Markov equivalence. To break this equivalence and uncover a more specific causal gene regulation network, genotype data is incorporated in the model inference process. The most straightforward approach is to use any of the methods in the previous section to calculate the probability $P(X_i\to X_j)$ of a causal interaction from $X_i$ to $X_j$ [@zhu2004; @zhu2008; @zhu2012stitching; @zhang2013integrated], for example by defining the prior as $p({\mathcal{G}})=\prod_{X_i}\left(\prod_{X_j\in{\mathbf{Pa}}^{({\mathcal{G}})}(X_i)}P(X_j\to X_i)
\prod_{X_j\not\in{\mathbf{Pa}}^{({\mathcal{G}})}(X_i)}(1-P(X_j\to X_i))\right)$. A more ambitious approach is to jointly learn the eQTL associations and causal trait (i.e. gene or phenotype) networks. In [@neto2010causal], EM is used to alternatingly map eQTLs given the current DAG structure, and update the DAG structure and model parameters given the current eQTL mapping. In [@scutari2014multiple], Bayesian networks are learned where SNPs and traits both enter as variables in the model, with the constraint that traits can depend on SNPs, but not vice versa. However, the additional complexity of both methods means that they are computationally expensive and have only been applied to problems with a handful of traits [@neto2010causal; @scutari2014multiple].
A few additional “tips and tricks” are worth mentioning:
- First, when the number of vertices is much larger than the sample count, we may break the problem into independent sub-problems by learning a separate Bayesian network for each co-expression module (Section \[sec-coex\] and [@zhang2013integrated]). Dependencies between modules could then be learned as a Bayesian network among the module eigengenes [@langfelder2007eigengene], although this does not seem to have been explored.
- Second, Bayesian network learning algorithms inevitably result in locally optimal models which may contain a high number of false positives. To address this problem, we can run the algorithm multiple times and report an averaged network, only consisting of edges which appear sufficiently frequent.
- Finally, another technique that helps in distinguishing genuine dependencies from false positives is *bootstrapping*, where resampling with replacement is executed on the existing sample pool. A fixed number of samples are randomly selected and then processed to predict a Bayesian network. This process is repeated many times, essentially regarding the distribution of sample pool as the true PDF, and allowing to estimate the robustness of each predicted edge, so that only those with high significance are retained [@friedman1999data]. In theory, even the whole pipeline of Figure \[fig-flow\] up to the *in silico* validation could be simulated in this way. Although bootstrapping is computationally expensive and mostly suited for small datasets, it could be used in conjunction with the separation into modules on larger datasets.
Using module networks to identify causal regulatory mechanisms
--------------------------------------------------------------
Module network inference is a statistically well-grounded method which uses probabilistic graphical models to reconstruct modules of co-regulated genes and their upstream regulatory programs, and which has been proven useful in many biological case studies [@add1; @segal2003; @friedman2004; @bonnet2015]. The module network model was originally introduced as a method to infer regulatory networks from large-scale gene expression compendia, as implemented in the **Genomica** software (<http://genomica.weizmann.ac.il>)[@segal2003]. Subsequently the method has been extended to integrate eQTL and gene expression data [@lee2006; @lee2009learning; @zhang2010bayesian]. The module network model starts from the same formula as Equation . It is then assumed that genes belonging to the same module share the same parents and conditional distributions; these conditional distributions are parameterized as decision trees, with the parental genes on the internal (decision) nodes and normal distributions on the leaf nodes [@segal2003]. Recent algorithmic innovations decouple the module assignment and tree structure learning from the parental gene assignment and use Gibbs sampling and ensemble methods for improved module network inference [@joshi2008; @joshi2009]. These algorithms are implemented in the **Lemon-Tree** software (<https://github.com/eb00/lemon-tree>), a command line software suite for module network inference [@bonnet2015].
Illustrative example {#sec:illustrative-example}
--------------------
We have recently identified genome-wide significant eQTLs for 6,500 genes in seven tissues from the Stockholm Atherosclerosis Gene Expression (STAGE) study [@foroughi2015], and performed co-expression clustering and causal networks reconstruction [@talukdar2015]. To illustrate the above concepts, we show some results for a co-expression cluster in visceral fat (88 samples, 324 genes) which was highly enriched for tissue development genes ($P=5\times 10^{-10}$) and contained 10 genome-wide significant eQTL genes and 25 transcription factors, including eight members of the homeobox family (Figure \[fig-figs\]a).
![ **(a)** Heatmap of standardized expression profiles across 88 visceral fat samples for 10 eQTL genes and 25 TFs belonging to a co-expression cluster inferred from the STAGE data. **(b)** Co-expression of HAP1 and FOXG1 across 88 visceral fat samples. **(c)** Association between HAP1’s eQTL (rs1558285) and expression of HAP1 (red), FOXG1 (blue) and FOXG1 adjusted for HAP1 and FOXG1’s eQTL (green). **(d)** Association between FOXG1’s eQTL (rs7160881) and expression of FOXG1 (blue), HAP1 (red), and HAP1 adjusted for FOXG1 and HAP1’s eQTL (green). **(e)** Causal interactions inferred between the same genes as in (a) using Bayesian network inference. **(f)** Example of a regulatory module inferred by **Lemon-Tree** from the STAGE data. See Section \[sec:illustrative-example\] for further details.\[fig-figs\]](fig2.pdf)
A representative example of an inferred causal interaction is given by the co-expression interaction between HAP1 (huntingtin-associated protein 1, chr17 q21.2-21.3) and FOXG1 (forkhead box G1, chr14 q11-q13). The expression of both genes is highly correlated ($\rho=0.85$, $P=4.4\times 10^{-24}$, Figure \[fig-figs\]b). HAP1 expression shows a significant, non-linear association with its eQTL rs1558285 ($P=1.2\times 10^{-4}$); this SNP also associates significantly with FOXG1 expression in the cross-association test ($P=0.0024$), but not anymore after conditioning FOXG1 on HAP1 and its own eQTL rs7160881 ($P=0.67$) (Figure \[fig-figs\]c). In contrast, although FOXG1 expression is significantly associated with its eQTL rs7160881 ($P=0.0028$), there is no association between this SNP and HAP1 expression ($P=0.037$), and conditioning on FOXG1 and HAP1’s eQTL has only a limited effect ($P=0.19$) (Figure \[fig-figs\]d). Using conditional independence tests (Section \[sec-direction\]), this results in a high-confidence prediction that HAP1 $\to$ FOXG1 is causal.
A standard greedy Bayesian network search algorithm [@schmidt2007learning] was run on the aforementioned cluster of 324 genes. Figure \[fig-figs\]e shows the predicted consensus sub-network of causal interactions between the 10 eQTLs and 25 TFs. This illustrates how a sparse Bayesian network can accurately represent the fully connected co-expression network (all 35 genes have high-mutual co-expression, cf. Figure \[fig-figs\]a).
Figure \[fig-figs\]f shows a typical regulatory module inferred by the **Lemon-Tree** software, also from the STAGE data. Here a heatmap is shown of the genotypes of an eQTL (top), the expression levels of a regulatory gene (middle), predicted to regulate a co-expression module of 11 genes (bottom). The red lines indicate sample clusters representing separate normal distributions inferred by the model-based co-clustering algorithm (Section \[sec:clust-co-expr\]).
*In silico* validation of predicted gene regulation networks {#sec:silico-valid-pred}
============================================================
Gene regulation networks reconstructed from omics data represent hypotheses about the downstream molecular implications of genetic variations in a particular cell or tissue type. An essential first step towards using these networks in concrete applications (e.g. discovering novel candidate drug target genes and pathways) consists of validating them using independent data. The following is a non-exhaustive list of typical *in silico* validation experiments.
#### Model likelihood comparison and cross-validation.
When different algorithms are used to infer gene network models, their log-likelihoods can be compared to select the best one. (With the caveat that the same data that was used to learn the models is used to compare them, this comparison is meaningful only when the algorithms optimize *exactly* the same (penalized) log-likelihood functions.) In a $K$-fold cross-validation experiment, the available samples are divided into $K$ subsets of approximately equal size. For each subset, models are learned from a dataset consisting of the $K-1$ other subsets, and the model likelihood is calculated using only the unseen data subset. Thus, cross-validation is used to test the generalisability of the inferred network models to unseen data. For an example where model likelihood comparison and cross-validation were used to compare two module network inference strategies, see [@joshi2009].
#### Functional enrichment.
Organism-specific gene ontology databases contain structured functional gene annotations [@ashb00]. These databases can be used to construct gene signature sets composed of genes annotated to the same biological process, molecular function or cellular component. Reconstructed gene networks can then be validated by testing for enriched connectivity of gene signature sets using a method proposed by [@zhu2008]. For a given gene set, this method considers all network nodes belonging to the set and their nearest neighbours, and from this set of nodes and edges, the largest connected sub-network is identified. Then the enrichment of the gene set in this sub-network is tested using the Fisher exact test and compared to the enrichment of randomly selected gene sets of the same size.
#### Comparison with physical interaction networks.
Networks of transcription factor - target interactions based on ChIP-sequencing data [@furey2012chip] from diverse cell and tissue types are available from the **ENCODE** [@encode2012], **Roadmap Epigenomics** [@kundaje2015integrative] and **modENCODE** [@gerstein2010integrative; @roy2010identification; @yue2014comparative] projects, while physical protein-protein interaction networks are available for many organisms through databases such as the **BioGRID** [@chatr2014biogrid]. Due to indirect effects, networks predicted from gene expression data rarely show a significant overlap with networks of direct physical interactions. A more appropriate validation is therefore to test for enrichment for short connection paths in the physical networks between pairs predicted to interact in the reconstructed networks [@bonnet2015].
#### Gene perturbation experiments.
Gene knock-out experiments provide the ultimate gold standard of a causal network intervention, and genes differentially expressed between knock-out and control experiments can be considered as true positive direct or indirect targets of the knocked-out gene. Predicted gene networks can be validated by compiling relevant (i.e. performed in a relevant cell or tissue type) gene knock-out experiments from the **Gene Expression Omnibus** (<http://www.ncbi.nlm.nih.gov/geo/>) or **ArrayExpress** (<https://www.ebi.ac.uk/arrayexpress/>) and comparing the overlap between gene sets responding to a gene knock-out and network genes predicted to be downstream of the knocked-out gene. Overlap significance can be estimated by using randomized networks with the same degree distribution as the predicted network.
Future perspective: Integration of multi-omics data
===================================================
Although combining genotype and transcriptome data to reconstruct causal gene networks has led to important discoveries in a variety of applications [@civelek2014systems], important details are not incorporated in the resulting network models, particularly regarding the causal molecular mechanisms linking eQTLs to their target genes, and the relation between variation in transcript levels and protein levels, with the latter ultimately determining phenotypic responses. Several recent studies have shown that at the molecular level, *cis*-eQTLs primarily cause variation in transcription factor binding to gene regulatory DNA elements, which then causes changes in histone modifications, DNA methylation and mRNA expression of nearby genes (reviewed in [@albert2015role]). Although mRNA expression can be used as a surrogate for protein expression, due to diverse post-transcriptional regulation mechanisms, the correlation between mRNA and protein levels is known to be modest [@lu2007; @schwanhausser2011], and genetic loci that affect mRNA and protein expression levels do not always overlap [@foss2007; @wu2013variation]. Thus, an ideal systems genetics study would integrate genotype data and molecular measurements at all levels of gene regulation from a large number of individuals.
Human lymphblastoid cell lines (LCL) are emerging as the primary model system to test such a approach. Whole-genome mRNA and micro-RNA sequencing data are available for 462 LCL samples from five populations genotyped by the 1000 Genomes Project [@lappalainen2013transcriptome]; protein levels from quantitative mass spectrometry for 95 samples [@wu2013variation]; ribosome occupancy levels from sequencing of ribosome-protected mRNA for 50 samples [@cenik2015integrative]; DNA-occupancy levels of the regulatory TF PU.1, the RNA polymerase II subunit RBP2, and three histone modifications from ChIP-sequencing of 47 samples [@waszak2015population]; and the same three histone modifications from ChIP-sequencing of 75 samples [@grubert2015genetic]. These population-level datasets can be combined further with three-dimensional chromatin contact data from Hi-C [@rao20143d] and ChIA-PET [@grubert2015genetic], knock-down experiments followed by microarray measurements for 59 transcription-associated factors and chromatin modifiers [@cusanovich2014functional], as well as more than 260 ENCODE assays (including ChIP-sequencing of 130 TFs) [@encode2012] in a reference LCL cell line (GM12878). Although the number of samples where all measures are simultaneously available is currently small, this number is sure to rise in the coming years, along with the availability of similar measurements in other cell types. Despite the challenging heterogeneity of data and analyses in the integration of multi-omics data, web-based toolboxes, such as **GenomeSpace** (<http://www.genomespace.org>) [@add1] can prove helpful to non-programmer researchers.
Conclusions {#sec:conclusions}
===========
In this chapter we have reviewed the main methods and softwares to carry out a systems genetics analysis, which combines genotype and various omics data to identify eQTLs and their associated genes, reconstruct co-expression networks and modules, reconstruct causal Bayesian gene and module networks, and validate predicted networks *in silico*. Several method and software options are available for each of these steps, and by necessity a subjective choice about which ones to include had to be made, based largely on their ability to handle large datasets, their popularity in the field, and our personal experience of using them. Where methods have been compared in the literature, they have usually been performed on a small number of datasets for a specific subset of tasks, and results have rarely been conclusive. That is, although each of the presented methods will give somewhat different results, no objective measurements will consistently select one of them as the “best” one. Given this lack of objective criterion, the reader may well prefer to use a single software that allows to perform all of the presented analyses, but such an integrated software does not currently exist.
Nearly all of the examples discussed referred to the integration of genotype and transcriptome data, reflecting the current dominant availability of these two data types. However, omics technologies are evolving at a fast pace, and it is clear that data on the variation of TF binding, histone modifications, and post-transcriptional and protein expression levels will soon become more widely available. Developing appropriate statistical models and computational methods to infer causal gene regulation networks from these multi-omics datasets is surely the most important challenge for the field.
Acknowledgements {#acknowledgements .unnumbered}
================
The authors’ work is supported by the BBSRC \[BB/M020053/1\] and Roslin Institute Strategic Grant funding from the BBSRC \[BB/J004235/1\].
[100]{}
Mackay TF, Stone EA and Ayroles JF. The genetics of quantitative traits: challenges and prospects. *Nature Reviews Genetics* **10**:565–577 (2009).
Goddard ME and Hayes BJ. Mapping genes for complex traits in domestic animals and their use in breeding programmes. *Nature Reviews Genetics* **10**:381–391 (2009).
Manolio TA. Bringing genome-wide association findings into clinical use. *Nature Reviews Genetics* **14**:549–558 (2013).
Hindorff LA *et al.* Potential etiologic and functional implications of genome-wide association loci for human diseases and traits. *Proceedings of the National Academy of Sciences* **106**:9362–9367 (2009).
Schaub MA *et al.* Linking disease associations with regulatory information in the human genome. *Genome Research* **22**:1748–1759 (2012).
Rockman MV and Kruglyak L. Genetics of global gene expression. *Nature Reviews Genetics* **7**:862–872 (2006).
Georges M. Mapping, fine mapping, and molecular dissection of quantitative trait loci in domestic animals. *Annu Rev Genomics Hum Genet* **8**:131–162 (2007).
Cookson W *et al.* Mapping complex disease traits with global gene expression. *Nature Reviews Genetics* **10**:184–194 (2009).
Cheung VG and Spielman RS. Genetics of human gene expression: mapping dna variants that influence gene expression. *Nature Reviews Genetics* **10**:595–604 (2009).
Cubillos FA, Coustham V and Loudet O. Lessons from [eQTL]{} mapping studies: non-coding regions and their role behind natural phenotypic variation in plants. *Current Opinion in Plant Biology* **15**:192–198 (2012).
Dimas AS *et al.* Common regulatory variation impacts gene expression in a cell type–dependent manner. *Science* **325**:1246–1250 (2009).
Foroughi Asl H *et al.* Expression quantitative trait loci acting across multiple tissues are enriched in inherited risk of coronary artery disease. *Circulation: Cardiovascular Genetics* **8**:305–315 (2015).
Greenawalt DM *et al.* A survey of the genetics of stomach, liver, and adipose gene expression from a morbidly obese cohort. *Genome Research* **21**:1008–1016 (2011).
Ardlie KG *et al.* The genotype-tissue expression ([GTEx]{}) pilot analysis: Multitissue gene regulation in humans. *Science* **348**:648–660 (2015).
Albert FW and Kruglyak L. The role of regulatory variation in complex traits and disease. *Nature Reviews Genetics* **16**:197–212 (2015).
Williams RW. Expression genetics and the phenotype revolution. *Mammalian Genome* **17**:496–502 (2006).
Kadarmideen HN, von Rohr P and Janss LL. From genetical genomics to systems genetics: potential applications in quantitative genomics and animal breeding. *Mammalian Genome* **17**:548–564 (2006).
Rockman MV. Reverse engineering the genotype–phenotype map with natural genetic variation. *Nature* **456**:738–744 (2008).
Schadt EE. Molecular networks as sensors and drivers of common human diseases. *Nature* **461**:218–223 (2009).
Schadt EE and Bj[ö]{}rkegren JL. New: network-enabled wisdom in biology, medicine, and health care. *Science Translational Medicine* **4**:115rv1–115rv1 (2012).
Civelek M and Lusis AJ. Systems genetics approaches to understand complex traits. *Nature Reviews Genetics* **15**:34–48 (2014).
Bj[ö]{}rkegren JL *et al.* Genome-wide significant loci: How important are they?: Systems genetics to understand heritability of coronary artery disease and other common complex disorders. *Journal of the American College of Cardiology* **65**:830–845 (2015).
Chen Y *et al.* Variations in [DNA]{} elucidate molecular networks that cause disease. *Nature* **452**:429–435 (2008).
Schadt EE, Friend SH and Shaywitz DA. A network view of disease and compound screening. *Nat Rev Drug Disc* **8**:286–295 (2009).
Walhout AJ. Unraveling transcription regulatory networks by protein–[DNA]{} and protein–protein interaction mapping. *Genome Research* **16**:1445–1454 (2006).
Ritchie MD *et al.* Methods of integrating data to uncover genotype-phenotype interactions. *Nature Reviews Genetics* **16**:85–97 (2015).
Stegle O *et al.* Using probabilistic estimation of expression residuals (peer) to obtain increased power and interpretability of gene expression analyses. *Nature Protocols* **7**:500–507 (2012).
Foss EJ *et al.* Genetic basis of proteome variation in yeast. *Nature Genetics* **39**:1369–1375 (2007).
Nicholson G *et al.* A genome-wide metabolic [QTL]{} analysis in [E]{}uropeans implicates two loci shaped by recent positive selection. *PLoS Genetics* **7**:e1002270 (2011).
Laird N and Lange C. *The Fundamentals of Modern Statistical Genetics* (Springer2011).
Shabalin AA. Matrix [eQTL]{}: ultra fast [eQTL]{} analysis via large matrix operations. *Bioinformatics* **28**:1353–1358 (2012).
Qi J *et al.* : Matrix-based non-parametric [eQTL]{} discovery. *BMC Bioinformatics* **15**:11 (2014).
Brem RB *et al.* Genetic dissection of transcriptional regulation in budding yeast. *Science* **296**:752–755 (2002).
Schadt EE *et al.* . *PLoS Biol* **6**:e107 (2008).
Lappalainen T *et al.* Transcriptome and genome sequencing uncovers functional variation in humans. *Nature* **501**:506–511 (2013).
Golub GH and Van Loan CF. *Matrix computations* (The Johns Hopkins University Press1996), third edn.
Hemani G *et al.* Detection and replication of epistasis influencing transcription in humans. *Nature* **508**:249–253 (2014).
Sharan R, Ulitsky I and Shamir R. Network-based prediction of protein function. *Molecular Systems Biology* **3**:88 (2007).
Daub CO *et al.* Estimating mutual information using [B]{}-spline functions – an improved similarity measure for analysing gene expression data. *BMC Bioinformatics* **5**:118 (2004).
Butte A and Kohane I. Mutual information relevance networks: Functional genomic clustering using pairwise entropy measurements. *Pac Symp Biocomputing* **5**:415–426 (2000).
Basso K *et al.* Reverse engineering of regulatory networks in human b cells. *Nat Genet* **37**:382–390 (2005).
Faith JJ *et al.* Large-scale mapping and validation of *Escherichia coli* transcriptional regulation from a compendium of expression profiles. *PLoS Biol* **5**:e8 (2007).
Hartwell LH *et al.* From molecular to modular cell biology. *Nature* **402**:C47–C52 (1999).
Eisen MB *et al.* Cluster analysis and display of genome-wide expression patterns. *PNAS* **95**:14863–14868 (1998).
Tavazoie S *et al.* Systematic determination of genetic network architecture. *Nature Genetics* **22**:281–285 (1999).
Sharan R and Shamir R. : a clustering algorithm with applications to gene expression analysis. In *Proc Int Conf Intell Syst Mol Biol*, vol. 8, 16 (2000).
Medvedovic M and Sivaganesan S. Bayesian infinite mixture model based clustering of gene expression profiles. *Bioinformatics* **18**:1194–1206 (2002).
Newman MEJ and Girvan M. Finding and evaluating community structure in networks. *Phys Rev E* **69**:026113 (2004).
Newman MEJ. Modularity and community structure in networks. *PNAS* **103**:8577–8582 (2006).
Ayroles JF *et al.* Systems genetics of complex traits in drosophila melanogaster. *Nat Genet* **41**:299–307 (2009).
Langfelder P and Horvath S. Wgcna: an r package for weighted correlation network analysis. *BMC Bioinformatics* **9**:559 (2008).
Lee SI *et al.* Learning a prior on regulatory potential from eqtl data. *PLoS Genetics* **5**:e1000358 (2009).
Clauset A, Newman MEJ and Moore C. Finding community structure in very large networks. *Phys Rev E* **70**:066111 (2004).
Van Dongen SM. Graph clustering by flow simulation (2001).
Enright AJ, Van Dongen S and Ouzounis CA. An efficient algorithm for large-scale detection of protein families. *Nucleic Acids Research* **30**:1575–1584 (2002).
Ravasz E *et al.* Hierarchical organization of modularity in metabolic networks. *Science* **297**:1551–1555 (2002).
Zhang B and Horvath S. A general framework for weighted gene co-expression network analysis. *Stat Appl Genet Mol Biol* **4**:17 (2005).
Langfelder P, Zhang B and Horvath S. Defining clusters from a hierarchical cluster tree: the dynamic tree cut package for r. *Bioinformatics* **24**:719–720 (2008).
Liu JS. *[M]{}onte [C]{}arlo strategies in scientific computing* (Springer2002).
Joshi A, Van de Peer Y and Michoel T. Analysis of a [Gibbs]{} sampler for model based clustering of gene expression data. *Bioinformatics* **24**:176–183 (2008).
Bonnet E, Calzone L and Michoel T. Integrative multi-omics module network inference with [Lemon-Tree]{}. *PLoS Computational Biology* **11** (2015).
Michoel T and Nachtergaele B. Alignment and integration of complex networks by hypergraph-based spectral clustering. *Physical Review E* **86**:056111 (2012).
Zhu J *et al.* An integrative genomics approach to the reconstruction of gene networks in segregating populations. *Cytogenet Genome Res* **105**:363–374 (2004).
Chen LS, Emmert-Streib F and Storey JD. Harnessing naturally randomized transcription to infer regulatory relationships among genes. *Genome Biology* **8**:R219 (2007).
Aten JE *et al.* Using genetic markers to orient the edges in quantitative trait networks: the [NEO]{} software. *BMC Systems Biology* **2**:34 (2008).
Schadt EE *et al.* An integrative genomics approach to infer causal associations between gene expression and disease. *Nature Genetics* **37**:710–717 (2005).
Neto EC *et al.* Inferring causal phenotype networks from segregating populations. *Genetics* **179**:1089–1100 (2008).
Neto EC *et al.* Modeling causality for pairs of phenotypes in system genetics. *Genetics* **193**:1003–1013 (2013).
Millstein J *et al.* Disentangling molecular relationships with a causal inference test. *BMC Genetics* **10**:23 (2009).
Smith GD and Ebrahim S. ‘mendelian randomization’: can genetic epidemiology contribute to understanding environmental determinants of disease? *International Journal of Epidemiology* **32**:1–22 (2003).
Li Y *et al.* Critical reasoning on causal inference in genome-wide linkage and association studies. *Trends in Genetics* **26**:493–498 (2010).
Friedman N, Nachman I and Pe[é]{}r D. Learning bayesian network structure from massive datasets: The “sparse candidate” algorithm. In *Proceedings of the Fifteenth Conference on Uncertainty in Artificial Intelligence*, UAI’99, 206–215 (Morgan Kaufmann Publishers Inc., San Francisco, CA, USA1999).
Friedman N *et al.* Using bayesian networks to analyze expression data. *Journal of Computational Biology* **7**:601–620 (2000).
Koller D and Friedman N. *Probabilistic Graphical Models: Principles and Techniques* ([The MIT Press]{}2009).
Schmidt M, Niculescu-Mizil A and Murphy K. Learning graphical model structure using [L1]{}-regularization paths. In *AAAI*, vol. 7, 1278–1283 (2007).
Zhu J *et al.* Integrating large-scale functional genomic data to dissect the complexity of yeast regulatory networks. *Nature Genetics* **40**:854–861 (2008).
Zhu J *et al.* Stitching together multiple data dimensions reveals interacting metabolomic and transcriptomic networks that modulate cell regulation. *PLoS Biology* **10**:e1001301 (2012).
Zhang B *et al.* . *Cell* **153**:707–720 (2013).
Neto EC *et al.* Causal graphical models in systems genetics: a unified framework for joint inference of causal network and genetic architecture for correlated phenotypes. *The Annals of Applied Statistics* **4**:320 (2010).
Scutari M *et al.* Multiple quantitative trait analysis using [B]{}ayesian networks. *Genetics* **198**:129–137 (2014).
Langfelder P and Horvath S. Eigengene networks for studying the relationships between co-expression modules. *BMC Systems Biology* **1**:54 (2007).
Friedman N, Goldszmidt M and Wyner A. Data analysis with [B]{}ayesian networks: A bootstrap approach. In *Proceedings of the Fifteenth conference on Uncertainty in artificial intelligence*, 196–205 (Morgan Kaufmann Publishers Inc.1999).
Segal E *et al.* Module networks: identifying regulatory modules and their condition-specific regulators from gene expression data. *Nat Genet* **34**:166–167 (2003).
Friedman N. Inferring cellular networks using probabilistic graphical models. *Science* **308**:799–805 (2004).
Lee S *et al.* . *Proc Natl Acad Sci USA* **103**:14062–14067 (2006).
Zhang W *et al.* A [B]{}ayesian partition method for detecting pleiotropic and epistatic [eQTL]{} modules. *PLoS Computational Biology* **6**:e1000642 (2010).
Joshi A *et al.* Module networks revisited: computational assessment and prioritization of model predictions. *Bioinformatics* **25**:490–496 (2009).
Talukdar H *et al.* Cross-tissue regulatory gene networks in coronary artery disease. *Cell Systems* **2**:196–208 (2016).
Ashburner M *et al.* . *Nat Genet* **25**:25–29 (2000).
Furey TS. –seq and beyond: new and improved methodologies to detect and characterize protein–[DNA]{} interactions. *Nature Reviews Genetics* **13**:840–852 (2012).
. . *Nature* **489**:57–74 (2012).
Kundaje A *et al.* Integrative analysis of 111 reference human epigenomes. *Nature* **518**:317–330 (2015).
Gerstein M *et al.* Integrative analysis of the [C]{}aenorhabditis elegans genome by the [modENCODE]{} project. *Science* **330**:1775–1787 (2010).
Roy S *et al.* Identification of functional elements and regulatory circuits by [D]{}rosophila [modENCODE]{}. *Science* **330**:1787–1797 (2010).
Yue F *et al.* A comparative encyclopedia of [DNA]{} elements in the mouse genome. *Nature* **515**:355–364 (2014).
Chatr-aryamontri A *et al.* The [BioGRID]{} interaction database: 2015 update. *Nucleic acids research* gku1204 (2014).
Lu P *et al.* Absolute protein expression profiling estimates the relative contributions of transcriptional and translational regulation. *Nature Biotech* **25**:117–124 (2007).
Schwanhausser B *et al.* . *Nature* **473**:337–342 (2011).
Wu L *et al.* Variation and genetic control of protein abundance in humans. *Nature* **499**:79–82 (2013).
Cenik C *et al.* Integrative analysis of rna, translation and protein levels reveals distinct regulatory variation across humans. *Genome Research* doi:10.1101/gr.193342.115 (2015).
Waszak SM *et al.* Population variation and genetic control of modular chromatin architecture in humans. *Cell* **162**:1039–1050 (2015).
Grubert F *et al.* Genetic control of chromatin states in humans involves local and distal chromosomal interactions. *Cell* **162**:1051–1065 (2015).
Rao SS *et al.* A [3D]{} map of the human genome at kilobase resolution reveals principles of chromatin looping. *Cell* **159**:1665–1680 (2014).
Cusanovich DA *et al.* The functional consequences of variation in transcription factor binding. *PLoS Genetics* **10**:e1004226 (2014).
Qu K *et al.* Integrative genomic analysis by interoperation of bioinformatics tools in GenomeSpace. *Nature Methods* **13**:245–247 (2016).
[^1]: There are only $\ell-1$ matrix multiplications, because the data standardization implies that ${\mathbf{X}}{\mathbf{I}}^{(0)}=1-\sum_{m=1}^{\ell-1}{\mathbf{X}}{\mathbf{I}}^{(m)}$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We provide an extension of a recent approach to study non-equilibrium thermodynamics \[Phys. Rev. E **81**, 051130 (2010), to be denoted by I in this work\] to inhomogeneous systems by considering the latter to be composed of quasi-independent subsystems. The system $\Sigma$ along with the (macroscopically extremely large) medium $\widetilde{\Sigma}$ forms an isolated system $\Sigma_{0}$. Starting from the Gibbsian formulation of the entropy for $\Sigma_{0}$, which is valid even when $\Sigma_{0}$ is out of equilibrium, we derive the Gibbsian formulation of the entropy of $\Sigma$, which need not be in equilibrium. We show that the additivity of entropy requires *quasi-independence* of the subsystems, which requires that the interaction energies between different subsystems must be *negligible* so that the energy also becomes additive. The thermodynamic potentials of subsystems such as the Gibbs free energy that continuously decrease during approach to equilibrium are determined by the field parameters (temperature, pressure, etc.) of the medium and exist no matter how far the subsystems are out of equilibrium so that their field variables may not even exist. This and the requirement of quasi-independence make our approach different from the conventional approach due to de Groot and others as discussed in the text. As the energy depends on the frame of reference, the thermodynamic potentials and Gibbs fundamental relation, but not the entropy, depend on the frame of reference. The possibility of relative motion between subsystems described by their net linear and angular momenta gives rise to viscous dissipation. The concept of internal equilibrium introduced in I is developed further here and its important consequences are discussed for inhomogeneous systems. The concept of internal variables (various examples are given in the text) as variables that cannot be controlled by the observer for non-equilibrium evolution is also discussed. They are important because the internal equilibrium in the presence of internal variables is lost if internal variables are not used in thermodynamics. It is argued that their affinity vanishes only in equilibrium. Gibbs fundamental relation, thermodynamic potentials and irreversible entropy generation are expressed in terms of observables and internal variables. We use these relations to eventually formulate the non-equilibrium thermodynamics of inhomogeneous systems. We also briefly discuss the case when bodies form an isolated system without any medium to obtain their irreversible contributions and show that this case is no different than when bodies are in an extremely large medium.'
author:
- 'P.D. Gujrati'
title: 'Non-equilibrium thermodynamics. II: Application to inhomogeneous systems'
---
Introduction
=============
Nature of the Problem
---------------------
In an earlier paper [@GujratiNETI], which we will refer to as I in this work, we have considered some of the consequences of applying the second law of thermodynamics to an isolated system $\Sigma_{0}$, which consists of a macroscopic system of interest $\Sigma$ containing a fixed number $N$ of particles (atoms or molecules) surrounded by an extremely large medium $\widetilde{\Sigma}$; see Fig. \[Fig\_Systems\]. From now on, it will be implicitly assumed that the medium is extremely large to be unaffected by the system. (Later in this work, we will also consider $\Sigma$ to have a fixed volume $V$ instead of $N$ or fixed $V$ and $N$. Furthermore, we will also consider briefly the case of many similar size systems forming the isolated system $\Sigma_{0}$ without the extremely large medium $\widetilde{\Sigma}$.) The motivation has been to develop this approach to obtain a non-equilibrium thermodynamic description of the open system $\Sigma$ under various conditions. Throughout this work, we will use *body* to refer to any of the three kinds systems: the isolated system, the medium or the system. All quantities related to $\Sigma_{0},\widetilde{\Sigma}$ and $\Sigma$ will be denoted by a suffix $0$,$\sim$ over the top, and without any suffix, respectively. All quantities related to a body will be denoted without any suffix, as if we are dealing with an open system. Similarly, in this work, we will say that the system is *open* when it is in a medium. Even though it is not the common usage, this should not cause any confusion as the context will be clear.
\[ptb\]
[System\_Modified\_1.eps]{}
To avoid complications due to external shear, we had only considered $\Sigma$ under no external shear in I. This restriction is easily removed as we will do here. The isolated system will still have *no* external force acting on it to ensure its isolation; see below also. We will now allow forces acting at the surface $\partial V$ of the system or any of its subsystems; see Fig. \[Fig\_two\_boxes\_modified\_system\](a). These forces must balance the internal stress tensor for mechanical equilibrium. Thus, the force $t_{i}df$ acting on a surface element $df$ must equal the stress force $\tau_{ij}n_{j}df$ (summation over repeated indices implied), and we have [@Landau-FluidMechanics; @Landau-Elasticity] for the surface traction force$$t_{i}=\tau_{ij}n_{j}. \label{Force_stress_relation}$$
\[ptb\]
[Two\_Boxes\_Modified\_System.eps]{}
Here, $\mathbf{n}$ is the outward unit normal at the surface element $df$ surrounding a point on the surface. This condition must be satisfied at every point on the surface $\partial V.$ The net force and torque acting on the system are given by $$\mathbf{F}\equiv{\displaystyle\oint\limits_{\partial V}}
\mathbf{t}df,\ \ \mathbf{K}\equiv{\displaystyle\oint\limits_{\partial V}}
\mathbf{r}^{(\text{s})}\times\mathbf{t}df,$$ respectively; here, for convenience, $\mathbf{r}^{(\text{s})}$ is taken to be the radius vector of the surface element with respect to the center of mass of the system. The external forces are responsible for the deformation of the system, and also result in the translation and rotation of the system. Let us consider an infinitesimal volume element $dV$ of mass $dm$, which is moving with a velocity $\mathbf{v}(t)$ and rotating with an angular velocity $\boldsymbol{\omega}(t)$ and has an intrinsic angular momentum $\mathbf{m}(t)dV$. The linear and angular momenta of the system in some fixed frame are given by$$M\mathbf{V(}t\mathbf{)\equiv}{\textstyle\int\nolimits_{V(t)}}
\mathbf{v}(t)dm,\ \mathbf{M}(t)\mathbf{\equiv}M\mathbf{R(}t\mathbf{)\times
V(}t\mathbf{)+}{\textstyle\int\nolimits_{V(t)}}
\mathbf{r}(t)\times\mathbf{v}(t)dm+{\textstyle\int\nolimits_{V(t)}}
\mathbf{m}(t)dV,$$ respectively; here $M=m_{0}N$ is the mass of the system ( $m_{0}$ being the mass of a particle), which we consider fixed for fixed $N$, and $\mathbf{R(}t\mathbf{)}$ and $\mathbf{V}(t)$ are the location and the translational velocity of the center of mass in this frame. If the frame is taken to be the center of mass frame, then $\mathbf{R(}t\mathbf{)}$ and $\mathbf{V}(t)$ are zero. If the body as a whole is *stationary*, then the energy of the body is known as the internal energy. Such a stationary situation was considered in our previous work [@GujratiNETI], where no motions were considered. This limitation will be removed here so as to allow for relative motions (translation and rotation) between the system and the medium. These relative motions are the additional sources of viscous dissipation and give rise to additional irreversibility. The irreversibility due to temperature difference (such as between the system and the medium) has already been considered in I.
We should mention here the recent somewhat comprehensive investigation carried out by Bouchbinder and Langer [@Langer] who also consider a system under external shear; however, our approaches and emphases are very different. We should also mention earlier very different equilibrium-like attempts by Lubchenko and Wolynes [@Wolynes] and by Öttinger [@Ottinger]. Mention should also be made of a very interesting phenomenological approach by Oono and Paniconi [@Oono] on steady state thermodynamics, which was later advanced by Sasa and Tasaki [@Sasa]. The classical local non-equilibrium thermodynamics due to de Donder [@Donder; @deGroot; @Prigogine; @Prigogine-old] is close in spirit to our approach and that of Bouchbinder and Langer [@Langer]. It will be the standard formulation with which we will compare and contrast our approach initiated in [@GujratiNETI]. Therefore, for the sake of continuity and clarity, we briefly discuss the classical formulation involving local equilibrium in Sect. \[Marker\_Local\_Thermodynamics\]. We note that there are other versions of non-equilibrium thermodynamics usually known as the extended, rational and GENERIC non-equilibrium thermodynamics [@Demirel; @Muschik]; however, we do not discuss these formulations in this work.
It is well known that the second law of thermodynamics for the isolated system $\Sigma_{0}$ states that$$\frac{dS_{0}(t)}{dt}\geq0, \label{Second_Law_0}$$ where $S_{0}(t)$ denotes the entropy of $\Sigma_{0}$ at some instant $t$. It should be stressed that the isolation of $\Sigma_{0}$ requires that there be something outside of $\Sigma_{0}$ from which it is isolated. Therefore, we will assumes that $\Sigma_{0}$ is confined to a finite though extremely large volume $V_{0}$ [@Gujrati-Symmetry]. The isolation requires that we neglect all interactions, such as gravitational interactions, of $\Sigma_{0}$ with the outside that cannot be shielded. All interactions with the outside should be relatively very weak to be *negligible*. All relevant interactions must be confined inside the volume $V_{0}$. This to not to be taken as a weakness of our approach as including the interactions with outside will only make $\Sigma_{0}$ an open system, so that our investigation of an open system, which is our primary concern, can then be applied to it.
We should emphasize the following important point. The law of increase of entropy in Eq. (\[Second\_Law\_0\]) should give a pause to those readers who believe that the concept of entropy is meaningful only for an equilibrium state and that the entropy cannot be defined for a non-equilibrium state. This belief is unfounded. The mere fact that the second law dictates the approach to equilibrium clearly implies that the entropy exists even when the system is out of equilibrium. This issue and the related history, in particular the Gibbs formulation of the non-equilibrium entropy, has been reviewed recently [@Gujrati-Symmetry and references cited therein], and we refer the reader to this for further details. We should stress that the Gibbs formulation of the entropy requires that the dynamics of the system and that of the isolated system are not deterministic, but rather stochastic; see [@Gujrati-Symmetry] for further details.
Important Restrictions in I and Their Removal in the Current Work \[marker\_important\_assumptions\_I\]
-------------------------------------------------------------------------------------------------------
An assumption that was implicit, but not stated, in I was that there was no relative motion between the system and the medium and that the isolated system was stationary. This is normally the case in practice in which the various bodies $(\Sigma_{0},\widetilde{\Sigma}$ and $\Sigma)$ are stationary in the laboratory frame of reference; the latter we will denote by $\mathcal{L}$ in this work. In general, one can allow for a translation and rotation of a body, as done here because we now wish to study their effects on the deformation of the system. Because of the stationary assumption, there is no difference between the energy and the internal energy for $\Sigma_{0},\widetilde{\Sigma}$ and $\Sigma$. This is a very common but useful assumption as the entropies depend on the internal energies and not on the energies [@Landau]. The latter energies may contain the contribution from the translation and rotation of the system as a whole. In contrast, the internal energies are the energies systems have in a frame in which they are stationary. Whenever we discuss both energies together in the following, the internal energy will be denoted by a superscript “i” to distinguish it from the energy, which is denoted without the superscript; otherwise, it will be clear which energy we are considering. For a stationary system in the lab frame $\mathcal{L}$, the internal energy is the same as the energy. The translation of a body as a whole merely affects the energy, but not its thermodynamic properties. However, the rotation of a body as a whole gives rise to centrifugal potential energy that modifies the energies of microstates and has to be carefully incorporated in any thermodynamic investigation [@Landau see Sect. 34] as done in Sect. \[marker\_Rotating\_Systems\]; see also Appendices \[Appd\_Frames\] and \[Appd\_Rotating\_System\_0\].
Present Goal and the Layout
---------------------------
We had focused exclusively on the system as a whole without worrying about its internal parts in I. As entropy palys a central role in our development, we discuss the formulation of entropy and thermodynamic averages in Sect. \[marker\_averages\], where we show the entropy itself as a thermodynamic average. Under the assumption of internal equilibrium in the system, there were no relative motions between its various parts that could add additional irreversible processes inside the system. In this work, we will remove this limitation and treat the system $\Sigma$ as inhomogeneous as glasses normally exhibit both spatial and temporal inhomogeneity; see [@Wolynes2; @Chandler; @Gujrati-Sokolov] for some recent investigations. We do this by considering $\Sigma$ to be composed of a collection of a large number $N_{\text{S}}$ of *subsystems* $\sigma_{k}$, $k=1,2,\cdots,N_{\text{S}},$ which may be different from each other to allow for inhomogeneity and for relative motions and shear forces between different subsystems in terms of internal variables. Each subsystem is still *macroscopically* large so that we can introduce a legitimate entropy function $s_{k}$; see Sect. \[marker\_inhomogeneous\_arbitrary\] for further elaboration.
Apart from the *observables* that can be manipulated by the observer, there also appear *internal variables* often used in describing glasses, as is well known from the early works of Davies and Jones [@Jones; @Gujrati-book]. The latter variables cannot be manipulated by the observer and were briefly introduced in I, but not explored. We will also remedy this situation here and consider glass as an inhomogeneous system with internal variables. Specifically, we treat translations and rotations of various parts of a system as internal variables that are generated by surface traction forces. The alternative approach is to use the traction forces and the strains instead; see for example [@Langer]. The phenomenological ideology introduced by Davies and Jones [@Jones], which has been recently reviewed by Öttinger [@Ottinger], is by now standard and has been discussed in several textbooks; see for example [@Nemilov-Book; @Gutzow-Book]. The observables and internal variables will be collectively called *state variables*; see Sect. \[marker\_averages\] for proper definitions of these terms.
In Sect. \[marker\_internal\_equilibrium\], we discuss the consequences of internal equilibrium and its similarity with and differences from the concept of local equilibrium [@Donder; @deGroot; @Prigogine; @Prigogine-old] discussed in Sect. \[Marker\_Local\_Thermodynamics\]. In particular, we argue in the form of Theorem \[marker\_Uniform\_Motion\] that the system can only sustain a uniform translation and rotation in internal equilibrium. It is assumed here that there are no additional conditions (such as the potential flow in a superfluid) on the velocity. The proof is trivial but the theorem has far reaching consequences. For example, it implies that the uniform rotation must be about a principle axis of inertia. A simple way to understand internal equilibrium is to think of the system as follows. We first disconnect it form the medium with which it is not in equilibrium, and connect it to another medium whose conjugate field variables ($T,P$, etc.) are exactly the same as that of the system. The system remains in equilibrium with this medium so that there will be no irreversible process in the system. We discuss the generalization of equilibrium Maxwell relation to systems in internal equilibrium. The condition for additivity and quasi-independence is considered in Sect. \[marker\_entropy\_additivity\], where we prove that the Gibbs entropy formulation in Eq. (\[Gibbs entropy\]) is also applicable to an open system, which is assumed to be *quasi-independent* of the medium. Various thermodynamic potentials are identified in Sect. \[marker\_thermodynamic\_potentials\] that are in accordance with the second law. It is here that we see a clear distinction between our approach and the conventional non-equilibrium theory exploiting the local equilibrium concept [@Donder; @deGroot; @Prigogine; @Prigogine-old]. Internal variables are discussed in Sect. \[Marker\_Internal Variables\]. We prove that the chemical potential or the affinity associated with an internal variable must be zero when the system is in equilibrium. We also prove that the entropy expressed solely in terms of observables when there are independent internal variables must explicitly depend on time so that while the system is in internal equilibrium with respect to all state variables, it is not in internal equilibrium with respect to only observables. A system undergoing uniform translation and rotation is studied in Sect. \[marker\_Rotating\_Systems\] where we also develop the Gibbs fundamental relation for such a system. We then apply the results from this study to an isolated system in which the system and the medium undergo relative translational motion in Sect. \[marker\_homogeneous\], but the system is homogeneous. An inhomogeneous system with relative motions between its subsystems is studied in the next section. We also discuss in this section the case of several bodies, each in internal equilibrium but different from others, that form an isolated body together by themselves without a medium; see Fig. \[Fig\_two\_boxes\_modified\_system\](b). These bodies are macroscopic in size, but are not extensively as large as a medium. We discuss a direct method of calculating the irreversible entropy generation in each body in terms of the equilibrium state of all the bodies. We find that the same results are also obtained by bringing all bodies in contact with a medium as shown in Fig. \[Fig\_two\_boxes\_modified\_system\](a). This equivalence is used to prove Theorem \[Theorem\_Isolated\_Bodies\]. All these investigations are extended to include extra observables and internal variables. The final section contains concluding discussion and a brief summary of our results.
Local Non-equilibrium Thermodynamics: A Brief Review\[Marker\_Local\_Thermodynamics\]
=====================================================================================
As entropy is the central quantity appearing in the second law in Eq. (\[Second\_Law\_0\]), we will pay close attention to its determination, although this is usually not done in classical local non-equilibrium thermodynamics [@Donder; @deGroot; @Prigogine; @Prigogine-old], where its existence is taken as a postulate with an implicit assumption that it is always *additive* [@deGroot]. The entropy $S$ of a system is defined in terms of the *local* entropy density $s(\mathbf{r})$ per unit volume:$$S\equiv{\textstyle\int\nolimits_{V}}
s(\mathbf{r})dV. \label{Entropy_Additivity_Thermodynamics}$$ The local temperature $T$ and pressure $P$ are assumed continuous functions of the location $\mathbf{r}$ and time $t$, and are postulated to always exist. The *additivity* of the energy $E$ results in$$E\equiv{\textstyle\int\nolimits_{V}}
\left[ e(\mathbf{r})+\frac12
m_{0}\rho(\mathbf{r})\mathbf{v}^{2}(\mathbf{r})+\mathbf{m(\mathbf{r})\cdot
}\boldsymbol{\omega}(\mathbf{r})+\psi(\mathbf{r})\right] dV.
\label{Energy_Additivity_Thermodynamics}$$ Here, $e(\mathbf{r})dV$ is the internal energy, $\frac12
m_{0}\rho(\mathbf{r})\mathbf{v}^{2}(\mathbf{r})dV$ and $\mathbf{m(\mathbf{r})\cdot}\boldsymbol{\omega}(\mathbf{r})dV$ the translational and rotational kinetic energy density, respectively, and $\psi(\mathbf{r})dV$ the additional energy contribution due to interactions not included in the internal energy density in a volume $dV$ of the system; the local mass and angular momentum densities are given by $m_{0}\rho$ and $\mathbf{m(\mathbf{r})},$ respectively; compare with Eq. (\[subsystem\_energy\_relation0\]) derived later after limiting it to the volume element $dV$.
The functional form of the entropy density depends on the nature of the system. For example, for a simple system containing a fixed number of structureless particles, it is assumed to be a function only of the internal energy density $e$ and the local number density $\rho$; see for example, [@deGroot see Eq. (III.14)]$$s(\mathbf{r})=s(e(\mathbf{r}),\rho(\mathbf{r})).
\label{local_equilibrium_entropy}$$ The local Gibbs free energy density $\widehat{g}$ is given by $$\widehat{g}=e-Ts+P. \label{Local_Gibbs_Free_Energy}$$ whether local equilibrium exists or not. (The unconventional use of the symbol $\widehat{g}$ instead of $g$ will become clear later when we discuss the Gibbs free energy, which follows from the second law and which continuously decrease as the system approaches equilibrium.) However, no direct method of calculating the entropy is given in this approach except by assuming as another postulate the validity of Gibbs fundamental relation, which for a simple system with no internal variables reads $$Td(s/\rho)=d(e/\rho)+Pd(1/\rho); \label{Gibbs_fundamental_relation}$$ this postulate is a consequence of assuming *local equilibrium* [@deGroot see Eq. (III.15)]. The presence of the local temperature $T$ and pressure $P$ in the fundamental relation imposes strong conditions on the nature of the entropy in that its partial derivatives are related to the given $T$ and $P$ under local equilibrium, which follow from Eq. (\[Gibbs\_fundamental\_relation\]).
We avoid the above issues in the conventional non-equilibrium thermodynamics [@Donder; @deGroot; @Prigogine; @Prigogine-old] by first identifying the entropy of the system in terms of microstate probabilities as described below, see Eq. (\[Gibbs entropy\]), and then use the concept of *internal equilibrium* to introduce the temperature and pressure in terms of this entropy; the latter are defined only when there is internal equilibrium [@GujratiNETI]. This thus avoids the issues inherent in the conventional approach. Our approach is *not* local in that we always deal with quantities $S,E,$ etc. related to macroscopically large systems or subsystems as opposed to the conventional thermodynamics which deals with local quantities $s,e,$ etc. As a consequence, not only $S,E,$ etc. but also the temperatures, pressures, etc. associated with these systems or subsystems will not *always* be continuous functions of space at the interfaces. Thus, we will not impose continuity in space on these quantities, which makes our approach distinct from the traditional local non-equilibrium approach of de Donder [@Donder; @deGroot; @Prigogine; @Prigogine-old] where these quantities are always treated as continuous. In the latter approach, a system can be broken into subsystems, each sufficiently small, yet large enough to be in internal equilibrium to satisfy Gibbs fundamental relation. This hypothesis is known as the local equilibrium hypothesis. We add another requirement, that of quasi-independence of the subsystems in our approach, which we believe to be extremely important. Only this requirement ensures that the entropy retains the additivity property and also remains a state variable, as we will discuss later in Sect. \[marker\_entropy\_additivity\].
Helmholtz Theorem\[marker\_Helmholtz\_theorem\]
-----------------------------------------------
To accommodate relative motion, we need to allow for surface traction forces $t_{i}$, which then give rise to internal forces causing the deformation of the system. These forces can be related to the induced stress tensor $\tau_{ij}$ as shown in Eq. (\[Force\_stress\_relation\]), and will result in a motion of the system due to non-zero net force and torque acting on the system. It is well known that the local motions for a deformable system can be described as a combination of three distinct types of motions [@Sommerfeld; @Batchelor]:
1. a pure translation
2. a pure strain, and
3. a pure rotation
by expressing the instantaneous difference in the velocity $\delta\mathbf{v}$ at two nearby points separated by a displacement vector $\delta\mathbf{r}$ as $$\delta v_{i}=\psi_{ij}\delta x_{j}+\widehat{\omega}_{ij}\delta x_{j},
\label{velocity_decomposition}$$ where the symmetric tensor $$\psi_{ij}\equiv\frac12
(\frac{\partial v_{i}}{\partial x_{j}}+\frac{\partial v_{j}}{\partial x_{i}})$$ represents the rate of strain tensor and the antisymmetric tensor $$\widehat{\omega}_{ij}\equiv\frac12
(\frac{\partial v_{j}}{\partial x_{i}}-\frac{\partial v_{i}}{\partial x_{j}})\equiv e_{ijk}\widehat{\omega}_{k}$$ is the vorticity tensor, and represents the axial vector $\widehat
{\boldsymbol{\omega}}\boldsymbol{=}\frac12
\boldsymbol{\partial}\times\mathbf{v}$ associated with $\widehat{\omega}_{ij}.$ The second term $\widehat{\omega}_{ij}\delta x_{j}$ in Eq. (\[velocity\_decomposition\]) represents the components of the vector $\widehat{\boldsymbol{\omega}}\times\delta\mathbf{r}$**.** One should think of $\delta\mathbf{v}$ as the relative velocity between two neighboring points separated by $\delta\mathbf{r}$.
The first contribution represents a pure straining motion while the second contribution represents a rigid-body rotation. For example, a simple shearing motion in which plane layers of the system slide over each other can be treated as a combination of a pure strain (with no rate of volume change) and a rotation [@Batchelor].
Stress Tensor
-------------
The motion at the local level can also be studied directly by considering the stress tensor. The stress tensor is normally expressed as a sum of the non-dissipative and dissipative or viscous contributions [@Landau-Elasticity]:$$\tau_{ij}=\sigma_{ij}+\sigma_{ij}^{\prime},
\label{stress_tensor_decomposition}$$ in which the viscous contribution $\sigma_{ij}^{\prime}$ is some function that depends on the velocity gradients, i.e., on $$\partial v_{i}/\partial x_{j},\partial^{2}v_{i}/\partial x_{j}\partial
x_{k},\text{ etc.}$$ Thus, we can express it as $$\sigma_{ij}^{\prime}=A_{ijkl}f_{kl}(\partial v_{m}/\partial x_{n},\partial
^{2}v_{m}/\partial x_{n}\partial x_{p},...),$$ where $A_{ijkl}$ does not depend on the velocity distribution and $f_{kl}$ is a function of the velocity gradients$.$ For example, in a linear approximation using only $\partial v_{i}/\partial x_{j}$, $f_{kl}$ is taken to be $$f_{kl}=f_{0}\partial v_{k}/\partial x_{l}$$ with a constant $f_{0}$, which could be conveniently absorbed in $A_{ijkl}.$ In this approximation, we see from Eq. (\[velocity\_decomposition\]) that $\sigma_{ij}^{\prime}$ depends on both the rate of strain tensor and the vorticity $\widehat{\boldsymbol{\omega}}$.
In general, we can partition $\sigma_{ij}^{\prime}$ into a symmetric and an antisymmetric part, the latter due to the presence of intrinsic rotation of the system [@deGroot] and describing the role of the rotational viscosity. Thus, we can also partition $\tau_{ij}$ into a symmetric and an antisymmetric part. We refer the reader to Chapter 12 in the monograph of de Groot and Mazur [@deGroot]. Of course, there may be symmetry reasons such as the isotropy of the system that would forbid the dependence on vorticity, in which case there would be no antisymmetric part. It is possible to show [@deGroot] that the rate of change of the intrinsic angular momentum is determined solely by the antisymmetric part of $\tau_{ij}^{\prime}.$ Thus, the absence of this part will imply the conservation of the orbital and intrinsic angular momentum separately.
It can now be shown [@deGroot Eqs. (XII.18) and (XII.18)] that the antisymmetric part $$\tau_{ij}^{\text{a}}\equiv\frac12
\left( \tau_{ij}-\tau_{ji}\right)$$ contributes a term proportional to$$\boldsymbol{\tau}\cdot(\widehat{\boldsymbol{\omega}}\mathbf{-}\boldsymbol{\omega})$$ to the rate of change of the internal energy $e$ and to the entropy production. Here, $\boldsymbol{\tau}\ $ is the vector associated with $\tau_{ij}^{\text{a}}$ $$\tau_{ij}^{\text{a}}=e_{ijk}\tau_{k}$$ and $\boldsymbol{\omega}$ represents the angular velocity of rotation of the system; cf. Eq. (\[Energy\_Additivity\_Thermodynamics\]). Thus, this contribution vanishes for uniform rotation $\widehat{\boldsymbol{\omega}}\mathbf{=}\boldsymbol{\omega}$ as expected. For $\widehat{\boldsymbol{\omega
}}\mathbf{\neq}\boldsymbol{\omega}$, there would be precession of the local volume element [@Landau-Mechanics] about the direction of $\boldsymbol{\omega}$, so that the rotational viscosity would play an important role until $\widehat{\boldsymbol{\omega}}$ **** becomes equal to **** $\boldsymbol{\omega}$.
Energy Balance and the First Law
--------------------------------
The decomposition in Eq. (\[stress\_tensor\_decomposition\]) allows us to break the surface traction also in two terms related to the individual contribution:$$\mathbf{t}=\boldsymbol{\sigma}+\boldsymbol{\sigma}^{\prime},$$ where the two new terms are defined similar to that in Eq. (\[Force\_stress\_relation\]). The non-viscous contribution is sometimes expressed by taking out the isotropic pressure term as follows$$\boldsymbol{\sigma}=-P\mathbf{n}+\boldsymbol{\varepsilon},$$ assuming that the thermodynamic pressure can be defined. The rate at which the work is done *on* the system is given by the average (over all microstates) of the stress tensor at the surface $\partial V$ of the system:$$\frac{dW^{\prime}(t)}{dt}=\overline{{\displaystyle\oint\limits_{\partial V(t)}}
t_{j}v_{j}^{(\text{s})}df}=\overline{{\displaystyle\oint\limits_{\partial V(t)}}
\sigma_{j}v_{j}^{(\text{s})}df}+\overline{{\displaystyle\oint\limits_{\partial V(t)}}
\sigma_{j}^{\prime}v_{j}^{(\text{s})}df}. \label{Total_Work0}$$ where $df_{i}$ and $v_{j}^{(\text{s})}$ are the components of $d\mathbf{f}(t)$ and the velocity $\mathbf{v}^{(\text{s})}$ of the surface element, respectively. When the pressure can be defined, this rate can be expressed as$$\frac{dW^{\prime}(t)}{dt}=-P(t)\frac{dV(t)}{dt}+\overline{{\displaystyle\oint\limits_{\partial V(t)}}
\varepsilon_{j}v_{j}^{(\text{s})}df}+\overline{{\displaystyle\oint\limits_{\partial V(t)}}
\sigma_{j}^{\prime}v_{j}^{(\text{s})}df}. \label{Total_Work}$$
In this work, we will not consider latent heats and chemical or nuclear reactions within the body. In this case, the rate at which the heat is added to the system is obtained by considering the heat flux through the surface $\partial V$ and is given by$$\frac{dQ(t)}{dt},$$ where $dQ(t)$ is the heat added to the system in time $dt$. We can then write down for the rate of change of the energy due to the dynamics in the system as$$\frac{dE(t)}{dt}=\frac{dQ(t)}{dt}+\frac{dW^{\prime}(t)}{dt},
\label{Internal_Energy_Balance_Eq}$$ which is a restatement of the first law of thermodynamics; see also [@Langer].
It should be remarked that the balance equation (\[Internal\_Energy\_Balance\_Eq\]) is in an integral form for the entire system and should be contrasted with the differential form commonly stated in textbooks; see for example [@deGroot]. The latter formulation is valid for infinitesimal volumes. Here, we are not interested in such a local description. Our main focus is to consider regions of the system that are macroscopically large enough so that they can be treated as quasi-independent. Under this condition, the entropy of the system can be approximated to a high accuracy by adding the entropies of the subsystems. We discuss this point further in Sect. \[marker\_homogeneous\_arbitrary\].
Need for Internal Equilibrium\[marker\_internal\_equilibrium\_0\]
-----------------------------------------------------------------
If the system as a whole is stationary, then the average velocity $\overline{\mathbf{V(}t\mathbf{)}}=0$. In this case, the energy of the system is the internal energy. We will consider this case below in this section for simplicity.
The first law statement in Eq. (\[Internal\_Energy\_Balance\_Eq\]) does not by itself allow us to determine the way the entropy of the system changes. This law is valid even if the system is not in internal equilibrium. In the absence of internal equilibrium, there is no way to determine the change in the entropy from this law. We need to relate $dE(t)$ with the change $dS(t)$ in the entropy to determine the latter. However, if we now assume that the system is in internal equilibrium (see I and Sect. \[marker\_internal\_equilibrium\] below), then the entropy no longer has an explicit $t$-dependence. In this case, we can write down the differential of the entropy $S(E,V,N)$ of a monatomic system of neutral particles with fixed $N$ as [@GujratiNETI] $$dS=\frac{1}{T(t)}dE+\frac{P(t)}{T(t)}dV; \label{Internal_Equilibrium_Relation}$$ compare this equation with Eq. (\[Gibbs\_fundamental\_relation\]). We can now use Eq. (\[Internal\_Energy\_Balance\_Eq\]) in this equation to determine the rate of change of the entropy. However, as we discuss in Sect. \[marker\_internal\_equilibrium\], in this case there is no *additional* irreversible entropy production arising from viscous stress tensor. Thus, the last contribution in Eqs. (\[Total\_Work0\]) or (\[Total\_Work\]) vanishes [@Kuiken Sect. 7.4.2]. If there is no shearing at the surface, then the second term in Eq. (\[Total\_Work\]) also vanishes, and we obtain the standard relation $$T(t)\frac{dS(t)}{dt}=\frac{dQ(t)}{dt}; \label{Entropy_Production}$$ this identification was used in I [@GujratiNETI].
Reversible Entropy Change and Irreversible Entropy Production\[marker\_entropy\_production\]
--------------------------------------------------------------------------------------------
We wish to emphasize, as was done in the previous work [@GujratiNETI see Eqs. (16) and (18) in particular] that this assumption of internal equilibrium does not mean that the *irreversible entropy production* $d_{\text{i}}S(t)$ in the system is absent unless the system happens to be in equilibrium with the medium. To see this most easily, we recount from [@GujratiNETI] that $$d^{(\text{E})}S(t)=\frac{dQ(t)}{T(t)} \label{Entropy_Change}$$ is the change in the entropy due to heat transfer $dQ(t)$ to the system that is in internal equilibrium and depends on its instantaneous temperature $T(t)$. We have used the superscript E to indicate that the entropy change we are discussing is due to energy (heat) transfer for which the associated conjugate variable, see Eq. (\[Field\_Variables\]), is $y_{\text{E}}=$ ($\partial S/\partial E$)$_{V,N}=1/T(t)$ and takes the value $y_{0\text{E}}=$ $1/T_{0}$ in equilibrium$.$ It follows from this that the reversible entropy change $d_{\text{e}}^{(\text{E})}S(t)$ in the system depends on its *equilibrium* temperature $T_{0}$, which is also the *constant* temperature of the medium, and is given by$$d_{\text{e}}^{(\text{E})}S(t)=\frac{dQ(t)}{T_{0}},
\label{Reversible_Entropy_Change}$$ regardless of the instantaneous temperature of the system. This results in [@GujratiNETI] $$d_{\text{i}}^{(\text{E})}S(t)=dQ(t)\left( \frac{1}{T(t)}-\frac{1}{T_{0}}\right) =F_{\text{E}}(t)dQ(t)\geq0 \label{Irreversible_Entropy_Change}$$ in all cases. Here, $$F_{\text{X}}(t)\equiv F\left[ y(t)\right] \equiv y(t)-y_{0}, \label{Force_y}$$ with $y=\allowbreak y_{0}$ representing the equilibrium value of the conjugate field $y$ associated with the observable $X$, represents the *thermodynamic force* for the flow of $X$. Since $T(t)$ in Eq. (\[Entropy\_Change\]) undergoes a change $dT\propto dQ(t)$ due to the heat transfer, the heat transfer is not isothermal. Thus, it should not be surprising that there is an irreversible entropy generation as part of it. On the other hand, the determination of $d_{\text{e}}S(t)$ in Eq. (\[Reversible\_Entropy\_Change\]) requires the heat transfer to be *isothermal* for the process to be *reversible*. What is surprising is that $d_{\text{e}}S(t)$ is determined not by the current state of the system at $t$, but its equilibrium state in the future so that the concept of causality is inapplicable [@Gujrati-Symmetry]. It is this particular aspect of $d_{\text{e}}S(t)$ that is the cornerstone of the second law of thermodynamics: $d_{\text{i}}S(t)\geq0$ also depends on the future. We will make use of this observations later.
The situation with other extensive variables like volume $V$ is no different. As shown in [@GujratiNETI], the entropy change and the reversible entropy change due to a change $dV$ are given by $$d^{(\text{V})}S(t)=\frac{P(t)}{T(t)}dV(t);\text{\ }d_{\text{e}}^{(\text{V})}S(t)=\frac{P_{0}}{T_{0}}dV(t); \label{Entropy_Change_V}$$ It follows from this that the irreversible entropy generation $d_{\text{i}}^{(\text{V})}S(t)$ is given by an identical formulation as above for $d_{\text{i}}^{(\text{E})}S(t)$ $$d_{\text{i}}^{(\text{V})}S(t)=dV(t)\left( \frac{P(t)}{T(t)}-\frac{P_{0}}{T_{0}}\right) =dV(t)F_{\text{V}}(t)\geq0, \label{Entropy_Change_V_i}$$ with $F_{\text{V}}(t)$ given by Eq. (\[Force\_y\]) with $y_{\text{V}}=P(t)/T(t)$, see also Eq. (\[Field\_Variables\_0\]), represents the thermodynamic force for the “flow” of volume $V$.
It should be stressed that the validity of Eq. (\[Internal\_Equilibrium\_Relation\]) follows from the continuity of the entropy and the existence of the internal equilibrium. This is a general relation and is valid for all systems whose macrostate can be described by the three variables $E,V,$ and $N.$ On the other hand, the form of Eq. (\[Entropy\_Production\]) depends on particular form of the processes that go on in the system. Thus, it is process-specific; recall that we have eliminated various processes in deriving this equation. We can incorporate the missing additional contributions in Eq. (\[Entropy\_Production\]) by properly introducing internal variables to describe the microstates of the system and generalizing Eq. (\[Internal\_Equilibrium\_Relation\]) to include other extensive observables. The issue of internal variables is taken up in Sect. \[Marker\_Internal Variables\], and the generalization of Eq. (\[Internal\_Equilibrium\_Relation\]) is taken up in Sect.\[marker\_Rotating\_Systems\].
Entropy and Averages\[marker\_averages\]
========================================
Isolated System
---------------
The entropy of an isolated system such as $\Sigma_{0}$, whether in equilibrium or not, is given by the Gibbs formulation$$S_{0}(t)=-{\textstyle\sum\limits_{\alpha}}
p_{\alpha}(t)\ln p_{\alpha}(t), \label{Gibbs entropy}$$ where $p_{\alpha}(t)$ is the time-dependent probability of the $\alpha$-th microstate of the isolated system $\Sigma_{0}$; see a recent review [@Gujrati-Symmetry] of this formulation for an isolated system, regardless of how far it is from equilibrium. It is assumed that the dynamics of the system is stochastic and not deterministic, as the latter dynamics makes the above entropy a constant of motion in direct contradiction with the second law. In a deterministic, i.e. Hamiltonian dynamics, a microstate uniquely evolves into a microstate, while in a stochastic dynamics, a microstates evolves into several microstates with certain probabilities. The unique evolution is time-reversible, which results in the entropy being a constant of motion. In a stochastic dynamic, the entropy can only increase [@Gujrati-Symmetry]. Since the system is isolated, we do not allow external forces acting on it; we of course neglect weak stochastic forces acting on it. Thus, any deformation if it occurs is due to internal forces. We have defined the entropy as a $\emph{dimensionless}$ quantity, which is equivalent to setting the Boltzmann constant $k_{\text{B}}$ equal to $1$. The collection $\boldsymbol{\alpha}=\left\{ \alpha\right\} $ of these microstates along with their *non-zero* probabilities represents a *macrostate* $\mathcal{M}_{0}^{\boldsymbol{\alpha}}$ or simply $\mathcal{M}_{0}$ of $\Sigma_{0}$. As equilibrium and non-equilibrium thermodynamics is an experimental science, a macrostate of any body is specified in terms of a set of some extensive *observables* $\mathbf{X}$ that can be manipulated by the observer. The same set of observables are also used to identify the microstates. Most often, the microstates are identified by the energy, volume, and the number of particles that form the elements of $\mathbf{X}$, because of their special role in thermodynamics. If there are other *extensive* mechanical quantities (quantities that are not purely thermodynamic in nature) such as the total linear and angular momenta, then the microstates are characterized by all these extensive quantities, collectively denote by $\mathbf{X}$; see the discussion in Sect. \[Marker\_Internal Variables\] for more details. Apart from these observables, a body can also be specified by a set $\mathbf{I}$ ** of *internal variables* [@Donder; @deGroot; @Prigogine; @Prigogine-old] that have been found very useful in describing glassy dynamics and will be discussed later in Sect. \[Marker\_Internal Variables\]. Indeed, one needs internal variables to also explain the time-evolution of an isolated system towards equilibrium since all its observables remain constant. Thus, the macrostate of the isolated system can use the internal variables for its specification. We will take these internal variables to be also extensive and call both of them as *state variables* and collectively denote them by $\mathbf{Z}$. Taking internal variables as extensive allows us to deal all state variables on equal footing, so that generalization from observables to internal variables becomes almost trivial.
Let us continue with the discussion of the isolated system. In general, microstate probabilities $p_{\alpha}(t)$ are functions of the state variables $\mathbf{Z}_{0}$ along with $t$. As a consequence, the entropy $S_{0}(\mathbf{Z}_{0}(t),t)$ is also going to be a function of $\mathbf{Z}_{0}(t)$ and $t$. There are situations, when the entropy can also depend on some *external parameters* such as the angular velocity of the rotation of the frame of reference. These parameters need not necessarily be extensive just as $t$ is not. The number of state variables are too limited for a complete microscopic description of the system, but is sufficient to describe the macroscopic conditions of a system.
The observables remain fixed for the isolated system $\Sigma_{0}$. The internal variable $\mathbf{I}_{0}(t)$, if present in the isolated system, is normally a function of time; its time-dependence describes the temporal evolution of $\Sigma_{0}$. While the microstate $\alpha$, hence the value of the state variable $\mathbf{Z}_{0\alpha}$ in the microstate $\alpha$, does not vary with time, the *average* $\mathbf{Z}_{0}(t)$ for the macrostates varies with $t$:$$\mathbf{Z}_{0}(t)\equiv{\textstyle\sum\limits_{\alpha}}
p_{\alpha}(t)\mathbf{Z}_{0\alpha}. \label{Average 0}$$ It should be pointed out that the entropy, as formulated in Eq. (\[Gibbs entropy\]) is also an *average* of $(-\ln p)$ [@Gujrati-Symmetry], the negative of the index of probability $\ln p$. There will be times, when we will also use an overbar such as in $\overline{\mathbf{Z}}_{0}(t)$ to indicate such averages for the sake of clarity. For common thermodynamic quantities such as average energy, volume, etc. the normal practice is to not use the overbar (unless clarity is needed) as it is mostly these average quantities that we deal with.
An Arbitrary Body
-----------------
It should also be stressed that the microstates for a body remain the same whether the body is isolated or not. We can apply Eqs. (\[Gibbs entropy\]) and (\[Average 0\]) to determine the entropy and the average quantity for any body, isolated or not (such as the system $\Sigma$ or the medium $\widetilde{\Sigma}$, neither of which is isolated). In the following, we will always use $i$ to denote a microstate of a body but reserve $\alpha$ to denote the microstate of the isolated system. The entropy and the average energy of a macrostate of a body is given by
\[0\]$$\begin{aligned}
S(t) & \equiv-{\textstyle\sum\limits_{i}}
p_{i}(t)\ln p_{i}(t),\label{System_Entropy}\\
E(t) & \equiv{\textstyle\sum\limits_{i}}
p_{i}(t)E_{i}, \label{System_Energy}$$ where $i$ denotes one of its microstates, whose probability is denoted by $p_{i}(t)>0$. While we can certainly allow microstates with probabilities $p_{i}(t)=0$ in Eq. (\[0\]), we find it convenient to only allow microstates with non-zero probabilities in the sum. Microstates with non-zero probabilities will be identified as *allowed* [@Gujrati-Symmetry] in this work.
While there cannot be any doubt about the validity of Eq. (\[System\_Energy\]), one may feel some reservation about Eq. (\[System\_Entropy\]) for the entropy of an open system. Therefore, we will give a direct proof of Eq. (\[System\_Entropy\]) in Sect. \[marker\_entropy\_additivity\].
Internal Equilibrium Thermodynamics
===================================
Equiprobability Concept and Consequences: No Internal Variables
---------------------------------------------------------------
We will first consider the case when there are no internal variables. As is the normal practice (see I for details), we assume that the medium is in *internal equilibrium* even if the system and the medium are not in equilibrium. This assumption is similar to the assumption of local equilibrium in the conventional nonequilibrium thermodynamics noted in Sect. \[marker\_averages\]. As discussed in I and in the review [@Gujrati-Symmetry], the condition for the internal equilibrium to be met is that the entropy has the maximum possible value at each instant for the *instantaneous* average value $\mathbf{X}_{\text{IE}}\equiv
\mathbf{X}(t)\ $of the observable of the body. For the case considered in I, they are $\widetilde{E}_{\text{IE}}\equiv\widetilde{E}(t)$ and $\widetilde
{V}_{\text{IE}}\equiv\widetilde{V}(t)$ for the medium; the number of particles of the medium is not allowed to change at all, so that $\widetilde
{N}_{\text{IE}}\equiv\widetilde{N}$ is always kept fixed. It is easy to see from Gibbs’ formulation of the entropy in Eqs. (\[Gibbs entropy\]) or (\[System\_Entropy\]) that this happens *if and only if* all the microstates that are allowed ($p_{i}(t)>0$) at that instant are *equiprobable*. Another way to think about the internal equilibrium is to imagine isolating the medium by disconnecting it from the system. This will keep $\widetilde{\mathbf{X}}$ *fixed* at $\widetilde{\mathbf{X}}_{\text{IE}}$. Then the entropy of the isolated medium cannot change anymore. In other words, it is in *equilibrium*. We can apply the same idea to any body in internal equilibrium. The body will remain in equilibrium if it is isolated.
Let us follow the consequences of this concept beyond what was discussed in I.
1. We allow for the possibility of external forces including stresses acting on the system and produced by the medium; see Fig. \[Fig\_Systems\]. Under internal equilibrium, $$p_{i}(t)=1/W(t),\ \ \forall i, \label{Internal_equilibrium}$$ where $W(t)$ is the number of allowed microstates [@Gujrati-Symmetry] at that instant. This immediately leads to the Boltzmann entropy $$S(t)=\ln W(t) \label{Boltzmann entropy}$$ for a system in internal equilibrium, a common assumption in non-equilibrium thermodynamics; see for example Bouchbinder and Langer [@Langer]. Since the entropy is maximum at each instant $t$, it cannot increase further if all observables are kept *fixed* at their values $\mathbf{X}_{\text{IE}}$ at that moment $t$, when the internal equilibrium is achieved. To ensure that $\mathbf{X}$ is held fixed at $\mathbf{X}_{\text{IE}}$, we *isolate* the system from the medium so that $\Sigma$ turns into an isolated system. Let its entropy be denoted by $S_{\text{IS}}(t)$ after isolation. As the entropy is already at its maximum, it cannot change. In other words, $S_{\text{IS}}\equiv
S(\mathbf{X}_{\text{IE}})\ $at *fixed* $\mathbf{X}_{\text{IE}}$ must be independent of time.
The above argument proves that the entropy $S_{\text{IE}}\equiv S(\mathbf{X}_{\text{IE}})$ has no *explicit* $t$-dependence when the system is in internal equilibrium: $$S_{\text{IE}}\equiv S(E_{\text{IE}},V_{\text{IE}},\cdots,N)=S(E(t),V(t),\cdots
,N).$$
Its implicit time dependence when $\Sigma$ is treated as an open system comes from the temporal evolution of $\mathbf{X}(t)$. Thus, the (maximum) entropy in Eq. (\[Boltzmann entropy\]) will change with time as $\mathbf{X}(t)$ changes with time in the open system. For the isolated situation, $S_{\text{IS}}\equiv
S(\mathbf{X}_{\text{IE}})\ $at *fixed* $\mathbf{X}_{\text{IE}}$ will remain constant.
2. Since the entropy is maximum for fixed $\mathbf{X}_{\text{IE}}$, there cannot be any *additional* irreversible entropy production $d_{\text{i}}S^{(\text{IE})}$ anymore$$\left. d_{\text{i}}S_{\text{IE}}\right\vert _{\mathbf{X}_{\text{IE}}}=0\text{
in internal equilibrium (IE); } \label{IE_entropy_production}$$ here, the symbol $\left. {}\right\vert _{\mathbf{X}_{\text{IE}}}$means that $\mathbf{X}_{\text{IE}}$ are held fixed.
3. When the system is in internal equilibrium, its various parts must be in equilibrium with each other. Otherwise, there would be irreversible entropy generation.
4. It also follows from (3) that all the arguments that one uses to follow the consequences of equilibrium can be applied to different parts of $\Sigma$ that are in equilibrium. For example, the arguments that establish that a system in equilibrium can only sustain uniform translation and rotation [@Landau Sect. 10] can be applied without any change to a system in internal equilibrium. As this result is going to play an important role in our approach, we state it as a theorem.
\[marker\_Uniform\_Motion\]There cannot be any relative motion between different parts of $\Sigma$ for *fixed* $\mathbf{X}_{\text{IE}}$ in the state of internal equilibrium. Thus, a system in internal equilibrium can only sustain uniform translation and rotation [@Landau].
We refer the reader to Landau and Lifshitz [@Landau Sect. 10] for the details. Their argument applies without any change to an isolated system in equilibrium. We easily extend their argument by considering our system at some instant $t$ when it has $\mathbf{X}_{\text{IE}}=\mathbf{X}(t)$. We keep $\mathbf{X}(t)$ fixed at $\mathbf{X}_{\text{IE}}$ by isolating the system from the medium. The arguments now apply without any change to the system in internal equilibrium. This proves the theorem.
The axis of the uniform rotation must be a principal axis of the instantaneous moment of inertia of the system. Otherwise, the system will undergo precession in space [@Landau-Mechanics] and the rotation will not be uniform.
5. Even with internal equilibrium in the system, there are both elastic and inelastic or plastic deformations [@Kuiken Sect. 7.4.2], which result in viscoelasticity in the system.
If and only if the system is under internal equilibrium, the derivatives of $S(t)$ with respect to $\mathbf{X}(t)$ have the significance of the *field* variables, also called the *conjugate* variables, which we will denote by $\mathbf{y}(t)$ or $\mathbf{Y}(t):$ $$\mathbf{y}(t)\equiv\frac{\mathbf{Y}(t)}{T(t)}\equiv\left( \frac{\partial
S(t)}{\partial\mathbf{X}(t)}\right) _{\mathbf{X}^{\prime}(t)}
\label{Field_Variables}$$ where $\mathbf{X}^{\prime}(t)$ denotes all other elements of $\mathbf{X}(t)$ except the one used in the derivative; compare with Eq. (\[Medium\_Fields\]). The temperature $T(t)$ and the pressure $P(t)$ are defined in the customary way by $$y_{E}(t)\equiv\frac{1}{T(t)}=\left( \frac{\partial S(t)}{\partial
E(t)}\right) _{\mathbf{X}^{\prime}(t)},y_{V}(t)\equiv\ \ \frac{P(t)}{T(t)}=\left( \frac{\partial S(t)}{\partial V(t)}\right) _{\mathbf{X}^{\prime}(t)} \label{Field_Variables_0}$$ where $\mathbf{X}^{\prime}(t)$ denotes all other elements of $\mathbf{X}(t)$ except $E(t)$ or $V(t),$ respectively in the two derivatives. The pair of quantities $\mathbf{X}(t),\mathbf{Y}(t)$ or $\mathbf{X}(t),\mathbf{y}(t)$ are called *conjugate* to each other.
The definitions of the conjugate fields give us an alternative way to interpret the internal equilibrium. We imagine bringing the system in contact with another medium whose field variables are also $\mathbf{Y}_{\text{IE}}$, where $$\mathbf{y}_{\text{IE}}\equiv\frac{\mathbf{Y}_{\text{IE}}}{T_{\text{IE}}}\equiv\left. \left( \frac{\partial S(t)}{\partial\mathbf{X}(t)}\right)
\right\vert _{\mathbf{X}_{\text{IE}}};$$ here $\left. {}\right\vert _{\mathbf{X}_{\text{IE}}}$ denotes the value of the derivative at $\mathbf{X}_{\text{IE}}$. To distinguish this medium from the first medium that is characterized by $\mathbf{Y}_{0}=(T_{0},P_{0},\cdots)$, we denote the first medium by $\widetilde{\Sigma}(\mathbf{Y}_{0})$, and the new medium by $\widetilde{\Sigma}(\mathbf{Y}_{\text{IE}})$. The system $\Sigma$ in internal equilibrium with observables $\mathbf{X}_{\text{IE}}$ is in equilibrium with the new medium $\widetilde{\Sigma}(\mathbf{Y}_{\text{IE}})$. If the system is isolated by disconnecting it from $\widetilde{\Sigma
}(\mathbf{Y}_{\text{IE}})$, then there cannot be any change in the macrostate of the system as all its observables remain constant at $\mathbf{X}_{\text{IE}}$. In other words, the system $\Sigma$ after being isolated remains in equilibrium if it was originally in internal equilibrium, as noted earlier. This will not be true if there are internal variables, to which we will turn to in a moment.
Zeroth and the Second Law
-------------------------
All the above discussion has been for the entire system, but can be easily extended to a system consisting of various subsystems $\sigma_{k}$, each in internal equilibrium. Let us consider the system shown in Fig. \[Fig\_two\_boxes\_modified\_system\], which consists of two subsystems $\sigma_{1}$ and $\sigma_{2}$ whose internal temperatures are $T_{1}(t)$ and $T_{2}(t)>T_{1}(t)$, respectively. Let their respective energies be $E_{1}(t)$ and $E_{2}(t)$, with their sum denoted by $E(t)$. We consider all other observables fixed for both subsystems. We first consider the system to be isolated with no medium, as shown in Fig. \[Fig\_two\_boxes\_modified\_system\](b). Then $E$ remains constant, but not the individual energies. The irreversible entropy gain for the entire system is$$d_{\text{i}}^{(\text{E})}S=dQ\left( \frac{1}{T_{1}(t)}-\frac{1}{T_{2}(t)}\right) >0 \label{Isolated_S_Generation}$$ during an infinitesimal heat transfer $dQ$ from the hotter subsystem to the colder subsystem. Here, $$F_{\text{E}}(t)\equiv\frac{1}{T_{1}(t)}-\frac{1}{T_{2}(t)}$$ plays the role of the thermodynamic force $F_{\text{E}}(t)$ for the isolated system, and should be contrasted with the same quantity for an open system in Eq. (\[Force\_y\]). As the system is isolated, this is also the entropy change $d^{\left( \text{E}\right) }S=d^{\left( \text{E}\right) }S_{1}+d^{\left( \text{E}\right) }S_{2}$ for the system, with $$d^{\left( \text{E}\right) }S_{1}=\frac{dQ}{T_{1}(t)},\ \ \ d^{\left(
\text{E}\right) }S_{2}=-\frac{dQ}{T_{2}(t)}$$ for the two subsystems as follows from Eq. (\[Entropy\_Change\]). We now consider the system to be in a medium at a fixed temperature $T_{0}$, as shown in Fig. \[Fig\_two\_boxes\_modified\_system\](a). We take $T_{0}$ to be the equilibrium temperature of the isolated subsystems; it is intermediate between $T_{1}(t)$ and $T_{2}(t)$. The infinitesimal heat given out by the hotter subsystem is now $dQ=dQ^{\prime}+dQ_{2}.$ The heat gained $dQ^{\prime}+dQ_{1}$by the colder subsystem must be exactly the heat loss $dQ$, since we are dealing with an isolated system. Therefore, $dQ_{1}\equiv dQ_{2}$, so that the entropy of the medium does not change. As the entropy change of the isolated system is equal to that of the system, we have$$d^{\left( \text{E}\right) }S=\frac{dQ^{\prime}}{T_{1}(t)}-\frac{dQ^{\prime}}{T_{2}(t)}-\frac{dQ_{2}}{T_{2}(t)}+\frac{dQ_{1}}{T_{1}(t)}=dQ\left( \frac
{1}{T_{1}(t)}-\frac{1}{T_{2}(t)}\right) >0, \label{Medium_S_Generation}$$ since $dQ=dQ^{\prime}+dQ_{1}=dQ^{\prime}+dQ_{2}$. This is the same irreversible entropy gain in Eq. (\[Isolated\_S\_Generation\]) for the isolated system in Fig. \[Fig\_two\_boxes\_modified\_system\](b). This should not be surprising as none of the heat transfers is isothermal. Thus, bringing the isolated system $\Sigma$, which consists of two subsystems, in contact with a medium, characterized by the equilibrium temperature $T_{0}$ of $\Sigma$, does not affect the irreversible entropy production. It is easy to see that the arguments can be extended to many subsystems and to other field variables. We will not pause here to do that.
At this moment, it is important to follow another important consequence of the thermodynamic force $F_{\text{E}}(t)$, which vanishes if and only if the system has come to *thermal equilibrium*. This is the *zeroth* law of thermodynamics in terms of the internal temperatures of the two subsystems. Thus, the internal instantaneous temperature plays the role of a thermodynamic temperature in that the heat always flows from a hotter subsystem to a colder subsystem. The above result can be easily generalized to many subsystems.
Let us now consider the volumes of the two subsystems to adjust as they come to equilibrium. All other extensive observables are considered fixed. Then the same reasoniong as above results in$$d_{\text{i}}^{(\text{V})}S=dV(t)\left( \frac{P_{1}(t)}{T_{1}(t)}-\frac
{P_{2}(t)}{T_{2}(t)}\right) >0$$ where $dV(t)$ is change in the volume of $\sigma_{1}$; the volume of the isolated system remains unchanged. The corresponding thermodynamic force$$F_{\text{V}}(t)\equiv\frac{P_{1}(t)}{T_{1}(t)}-\frac{P_{2}(t)}{T_{2}(t)}$$ vanishes when the system comes to *mechanical equilibrium*. It usually happens that thermal equilibrium requires mechanical equilibrium in that the forces exerted on each other by any two subsystems must be equal and opposite. Thus, the conditions for the equilibrium is that not only the pressure $P(t)$ but also the temperature $T(t)$ have the same value in both subsystems.
Presence of Internal Variables
------------------------------
We can easily extend the above discussion to include internal variables $\mathbf{I}(t)$ by replacing $\mathbf{X}(t)$ by $\mathbf{Z}(t)$. In the context of internal variables, their conjugate variables are known as “*affinity*.” The general form of Eq. (\[Field\_Variables\]) is$$\mathbf{w}(t)\equiv\frac{\mathbf{W}(t)}{T(t)}\equiv\left( \frac{\partial
S(t)}{\partial\mathbf{Z}(t)}\right) _{\mathbf{Z}^{\prime}(t)},
\label{Field_Variables_1}$$ where $\mathbf{Z}^{\prime}(t)$ denotes all other elements of $\mathbf{Z}(t)$ except the one used in the derivative. The affinity $\mathbf{a}(t)$ is given by $$\mathbf{a}(t)\equiv\frac{\mathbf{A}(t)}{T(t)}\equiv\left( \frac{\partial
S(t)}{\partial\mathbf{I}(t)}\right) _{\mathbf{Z}^{\prime}(t)},
\label{Field_Variables_2}$$ so that $\mathbf{w}(t)$ consists of $$\mathbf{y}(t)\equiv\left( \frac{\partial S(t)}{\partial\mathbf{X}(t)}\right)
_{\mathbf{Z}^{\prime}(t)} \label{Field_Variables_3}$$ and $\mathbf{a}(t)$. The generalization of the thermodynamic force in Eq. (\[Force\_y\]) is given by$$F_{\text{Z}}(t)\equiv F\left[ w(t)\right] \equiv w(t)-w_{0}, \label{Force_w}$$ with $w_{0}$ representing the equilibrium value of $w$ corresponding to the state variable $Z$.
Let us now consider the system to be in internal equilibrium, while the medium containing it is $\widetilde{\Sigma}(\mathbf{Y}_{0},\mathbf{A}_{0})$, where $\mathbf{Y}_{0},\mathbf{A}_{0}$ characterize the medium. If we now disconnect the system from this medium, but bring it in contact with another medium $\widetilde{\Sigma}(\mathbf{Y}_{\text{IE}},\mathbf{A}_{\text{IE}})$, where $\mathbf{Y}_{\text{IE}}$ and $\mathbf{A}_{\text{IE}}$ are the field and affinity vectors associated with the system in internal equilibrium, then the system will remain in equilibrium with this medium. This is no different than what we have said above in the absence of any internal variable $\mathbf{I}$.
But the situation is very different when we try to keep the system isolated. Since internal variables are not under the control of the observer, they cannot be manipulated to remain constant after isolation and will continue to change. Thus, after the isolation, $\mathbf{X}_{\text{IS}}$, the value of $\mathbf{X}(t)$ at the instance of isolation, will remain constant, but $\mathbf{I}(t)$ will not remain fixed at its value $\mathbf{I}_{\text{IS}}$ at the instant it was isolated. This time variation in the internal variables is what drives this isolated system towards its equilibrium state during which its entropy will continuously increase. This is very different from the case above when there were no internal variables. Thus, a system in internal equilibrium cannot be isolated as an equilibrium system if there are internal variables present. It can only remain in equilibrium with the medium $\widetilde{\Sigma}(\mathbf{Y}_{\text{IS}},\mathbf{A}_{\text{IS}})$.
Maxwell Relations
-----------------
As the concept of internal equilibrium is not that different from the concept of equilibrium, it should not come as a surprise that there are analogs of Maxwell relations. We recall that in equilibrium thermodynamics, the standard Maxwell relations for a system characterized by only $S$ and $V$ (fixed $N$) are as follows in terms of Jacobians: $$\begin{aligned}
\frac{\partial(T_{0},S,N)}{\partial(V,S,N)} & =\frac{\partial(P_{0},V,N)}{\partial(V,S,N)},\ \ \ \frac{\partial(T_{0},S,N)}{\partial(P_{0},S,N)}=\frac{\partial(P_{0},V,N)}{\partial(P_{0},S,N)},\nonumber\\
\frac{\partial(P_{0},V,N)}{\partial(T_{0},V,N)} & =\frac{\partial
(T_{0},S,N)}{\partial(T_{0},V,N)},\ \ \ \frac{\partial(T_{0},S,N)}{\partial(P_{0},T_{0},N)}=\frac{\partial(P_{0},V,N)}{\partial(P_{0},T_{0},N)}.
\label{Maxwell_Jacobians}$$ All four Maxwell relations use the same numerators $\partial(T_{0},S,N)$ and $\partial(P_{0},V,N)$. They use different denominators. Thus, they can all be combined into one compact relation that can be simply written as $$\partial(T_{0},S,N)=\partial(P_{0},V,N). \label{Maxwell_Compact}$$ Here, the relation only has a meaning if each side is divided by one of the possible denominators $\partial(V,S,N),\partial(P_{0},S,N),\partial
(T_{0},V,N)$ and $\partial(P_{0},T_{0},N)$ on both sides.
We now consider a system. For simplicity, we assume that only one internal variable, which we denote by $\xi$, characterizes this system. We assume the system is in internal equilibrium. To simplify the notation, we will suppress $N$ but use the additional variable $\xi$ along with the other two variables. By considering the system in a medium $\widetilde{\Sigma}(\mathbf{Y}_{\text{IS}},\mathbf{A}_{\text{IS}})$, we recognize that the system is in equilibrium. Thus, the Maxwell relations now for fixed $\xi$ can be compactly represented by$$\partial(T,S,\xi)=\partial(P,V,\xi), \label{Maxwell_Relation_TSPV/N}$$ by replacing $T_{0},P_{0}$ by $T(t)=T_{\text{IS}},P(t)=P_{\text{IS}},$ which for simplicity have been written as $T,P$. The extension to the Maxwell relation in terms of the internal variable requires considering the pair $A,\xi$ in place of $T,S$ or $P,V$, where $A$ denotes the conjugate affinity to $\xi$. For fixed $V$, the Maxwell relation is$$\partial(T,S,V)=\partial(A,\xi,V), \label{Maxwell_Relation_TSPV/N_1}$$ and for fixed $S$, the Maxwell relation is$$\partial(P,V,S)=-\partial(A,\xi,S). \label{Maxwell_Relation_TSPV/N_2}$$ If we consider the system in the medium $\widetilde{\Sigma}(\mathbf{Y}_{\text{IS}},\mathbf{A}_{\text{IS}})$, then we have the standard equilibrium Maxwell relations similar to those in Eq. (\[Maxwell\_Jacobians\]). However, we are interested in the possible “non-equilibrium” Maxwell relations when the system is in the medium $\widetilde{\Sigma}(\mathbf{Y}_{0},\mathbf{A}_{0})$. We first consider fixed $\xi$. The Maxwell relation in Eq. (\[Maxwell\_Relation\_TSPV/N\]) turns into the identity
$$\frac{\partial(T,S,\xi)}{\partial(P_{0},S,\xi)}=\frac{\partial(P,V,\xi
)}{\partial(P_{0},S,\xi)} \label{Maxwell_Enthalpy/n}$$ that is $$\left( \frac{\partial T}{\partial P_{0}}\right) _{S,\xi}=\left(
\frac{\partial P}{\partial P_{0}}\right) _{S,\xi}\left( \frac{\partial
V}{\partial S}\right) _{P_{0},\xi}-\left( \frac{\partial V}{\partial P_{0}}\right) _{S,\xi}\left( \frac{\partial P}{\partial S}\right) _{P_{0},\xi}$$ for the enthalpy and $$\frac{\partial(T,S,\xi)}{\partial(T_{0},V,\xi)}=\frac{\partial(P,V,\xi
)}{\partial(T_{0},V,\xi)}, \label{Maxwell_Helmholtz/n}$$ that is $$\left( \frac{\partial P}{\partial T_{0}}\right) _{V,\xi}=\left(
\frac{\partial T}{\partial T_{0}}\right) _{V,\xi}\left( \frac{\partial
S}{\partial V}\right) _{T_{0},\xi}-\left( \frac{\partial S}{\partial P_{0}}\right) _{V,\xi}\left( \frac{\partial T}{\partial V}\right) _{T_{0},\xi}$$ for the Helmholtz free energy; the details will be given in a separate publications [@Gujrati-IV]. One can also obtain Maxwell relations at fixed $V$ or $S$. For example, we find the following Maxwell relation $$\frac{\partial(P,V,S)}{\partial(\xi,V,S)}=-\frac{\partial(A,\xi,S)}{\partial(\xi,V,S)},$$ that is $$\left( \frac{\partial P}{\partial\xi}\right) _{V,S}=\left( \frac{\partial
A}{\partial V}\right) _{\xi,S}.$$ Similarly, from the Maxwell relation$$\frac{\partial(T,S,V)}{\partial(\xi,S,V)}=\frac{\partial(A,\xi,V)}{\partial(\xi,S,V)},$$ we find$$\left( \frac{\partial T}{\partial\xi}\right) _{V,S}=-\left( \frac{\partial
A}{\partial S}\right) _{\xi,V}.$$ The Maxwell relations in Eq. (\[Maxwell\_Relation\_TSPV/N\]-\[Maxwell\_Relation\_TSPV/N\_2\]) contain the internal fields of the system and not of the medium [@Gujrati-IV] when the system is out of equilibrium. Obviously, the extensive variables in the relation must refer to the system.
Internal Equilibrium Thermodynamics versus Local Thermodynamics
---------------------------------------------------------------
We will argue in Sect. \[marker\_entropy\_additivity\] that the concept of internal equilibrium, which we adopt, is no different than the concept of local equilibrium used in Eq. (\[local\_equilibrium\_entropy\]) or in the Gibbs fundamental relation in Eq. (\[Gibbs\_fundamental\_relation\]). Despite this, the two approaches based on the concept of local and internal equilibrium, respectively, differ in some important ways that will be elaborated later. Here, we briefly mention these differences. The first important difference is that our approach is truly a statistical mechanical approach for non-equilibrium systems. Once the probabilities of microstates are known, the averages and other moments of all state variables that are used to identify the microstates and the entropy are determined for the macrostate. For example, the average fluctuation in $\mathbf{Z}$ for a body is given by$$\left[ \Delta\mathbf{Z}(t)\right] ^{2}\equiv{\textstyle\sum\limits_{i}}
p_{i}(t)\left[ \mathbf{Z}_{i}-\overline{\mathbf{Z}}(t)\right] ^{2},
\label{Fluctuations}$$ where$$\overline{\mathbf{Z}}(t)\equiv{\textstyle\sum\limits_{i}}
p_{i}(t)\mathbf{Z}_{i} \label{Average}$$ is the average $\overline{\mathbf{Z}}$ for the body; compare with Eq. (\[Average 0\]). As the microstate probabilities exist even when the system is out of equilibrium, these averages including the entropy exist at all times even if the system is not in internal equilibrium. Their temporal variations are controlled by the dynamics governing the system and give rise to various balance equations; see for example Eq. (\[Internal\_Energy\_Balance\_Eq\]). The use of a probabilistic approach in the determination of the entropy and other statistical properties means that the dynamics in the system must be *stochastic* and not deterministic, as the entropy remains constant in a deterministic dynamics [@Gujrati-Symmetry]. A consequence of the stochastic nature is that irreversible dissipation becomes an integral part of the statistical description of any system, which then results in the law of increase of entropy as captured by Eq. (\[Second\_Law\_0\]). It is this law that was the foundation of our approach in I, and which we develop further in this work.
The second difference is in the identification of the thermodynamic potentials for the open system, which has been discussed in I and will be further elaborated in Sect. \[marker\_thermodynamic\_potentials\] and again in Sect. \[marker\_inhomogeneous\_arbitrary\]. It is discovered in our approach that thermodynamic potentials contain the field parameters (temperature, pressure, chemical potentials,etc.) of the medium, which determine how far an open system is from its equilibrium with the medium. These thermodynamic potentials satisfy the second law in that they do *not* increase in a spontaneous process. In contrast, the form given in Eq. (\[Local\_Gibbs\_Free\_Energy\]) for the local Gibbs free energy density or its integral over the volume of the system does not always satisfy this requirement; see the discussion surrounding $\widehat{G}(t)$ in Eq. (22) of I. However, there is no discrepancy for the internal energy in the two approaches as both approaches give the same Gibbs fundamental relation. This is because the fundamental relation in both approaches does *not* depend on the field parameters of the medium, but include the instantaneous field parameters of the system.
The third difference is in the reversible and irreversible entropy changes, which depend on the equilibrium value $w_{0}$ ($y_{0}$ or $a_{0}$) of the conjugate field $w$ ($y$ or $a$), as is easily seen from Eqs. (\[Reversible\_Entropy\_Change\]), (\[Irreversible\_Entropy\_Change\]), (\[Entropy\_Change\_V\]) and (\[Entropy\_Change\_V\_i\]). In the localthermodynamics, these quantities are determined by the local conjugate fields.
The presence of the medium field variable in the thermodynamic potentials of the system does not mean that the situation would be any different if we consider the system or subsystems to form the isolated system $\Sigma_{0}$ without any medium, as shown in Fig. \[Fig\_two\_boxes\_modified\_system\](b). In this case, which we consider in Sect. \[marker\_comparable\_bodies\], we again find that the thermodynamic potentials do not depend on the field variables of the system or subsystems. The role of the field variables of the medium $\widetilde{\Sigma}(\mathbf{Y}_{0},\mathbf{A}_{0})$ are now played by the equilibrium conjugate variables $\mathbf{Y}_{0},\mathbf{A}_{0}$ of the system or subsystems. This then leads us to the following important theorem:
\[Theorem\_Isolated\_Bodies\]An isolated system $\Sigma$ is no different than the open system $\Sigma$ in an extensively large medium $\widetilde
{\Sigma}(\mathbf{Y}_{0},\mathbf{A}_{0})$, provided the medium is appropriately chosen to represent the equilibrium state (in terms of $\mathbf{Y}_{0},\mathbf{A}_{0}$) of the isolated system $\Sigma$.
In particular, the reversible entropy change and the irreversible entropy generation in the two cases (a) and (b) in Fig. \[Fig\_two\_boxes\_modified\_system\] are exactly the same, as they both depend on the equilibrium conjugate variables $\mathbf{Y}_{0},\mathbf{A}_{0}$ of the system or subsystems. An example of this is already seen in Eqs. (\[Reversible\_Entropy\_Change\]) and (\[Irreversible\_Entropy\_Change\]). We defer the proof of this theorem to Sect. \[marker\_comparable\_bodies\].
Additivity of Entropy and Quasi-independence \[marker\_entropy\_additivity\]
============================================================================
For simplicity of discussion, we consider all systems to be stationary in this section, so that we only deal with internal energies. It was noted in Sect. \[marker\_averages\] and recently reviewed in [@Gujrati-Symmetry], the entropy of an isolated body is given by the Gibbs formulation in Eq. (\[Gibbs entropy\]), regardless of whether it is in equilibrium or not. There is no reason to believe that this formulation also applies to an *open* body under all conditions, though its applicability in equilibrium is not in dispute [@Landau]. We now prove that this formulation also applies to an open body under a condition that is always taken for granted. We will specifically consider our system $\Sigma$ at some instant $t$, but the conclusion is valid for all bodies. Let us consider all allowed microstates of $\Sigma$ with fixed number of particles $N$; we index these microstates by $i=1,2,\cdots,W(t)$. Theses microstates correspond to all possible energies and volumes of the system. We use $\widetilde{\alpha}$ to denote the microstates of $\widetilde{\Sigma}$ whose number of particles $\widetilde{N}$ is also fixed. A specification of the microstates $i$ and $\widetilde{\alpha}$ gives a unique microstate specification $\alpha$ representing a microstate of the isolated system $\Sigma_{0}$. Hence, the number of microstates $W_{0}(t)\ $of the $\Sigma_{0}$ is the product $$W_{0}(t)=W(t)\widetilde{W}(t), \label{Microstate_Numbers}$$ where $W(t)$ and $\widetilde{W}(t)$ are respectively the number of all *allowed* microstates [@Gujrati-Symmetry] of $\Sigma$ and $\widetilde{\Sigma},$ respectively at that instant $t$.
As the concept of microstates does not depend on the nature of interactions (they exist even in the absence of interaction), the above equation is valid for all kinds of interactions. Let $E_{0},E(t)$ and $\widetilde{E}(t)$ denote the internal energies of $\Sigma_{0},\Sigma$ and $\widetilde{\Sigma}$, respectively. Let $E_{0}^{(\text{int})}(t)$ denote the mutual interaction energy between $\Sigma$ and $\widetilde{\Sigma}$ at that instant. For short-ranged interactions, this energy is determined by the surface $\partial
V(t)$ of $\Sigma$. For convenience, we assume that this entire area is exposed to the surrounding medium, as shown in Fig. \[Fig\_Systems\]. If long-ranged interactions are also present, or if the system size is very small, this energy may depend on the entire volume $V(t)$ of $\Sigma$. In all cases, this energy is defined by the following identity$$E_{0}\equiv E(t)+\widetilde{E}(t)+E_{0}^{(\text{int})}(t).
\label{Interaction_Energy}$$ Because of the smallness of $\Sigma$ relative to $\Sigma_{0}$, $E(t)<<\widetilde{E}(t)$. If it happens that $$E_{0}^{(\text{int})}(t)<<E(t), \label{Quasi-independence}$$ we call $\Sigma$ and $\widetilde{\Sigma}$ *quasi-independent*. For quasi-independence, the linear size of the system must be at least as large as, but hopefully larger than, the correlation length in the system. In this case, we can neglect their mutual interactions, which is a common practice in the discipline [@Landau]. The quasi-independence of the system and the medium holds to a very high degree of accuracy for all short-ranged interactions [@GujratiNETI], provided the system itself is macroscopically large so that the ratio of its surface to volume is insignificant. In most cases, this will also ensure that the correlation length is small compared to the size of the system. If there are also long-ranged interactions, then we can still have quasi-independence provided these interactions are relatively weak and shielding occurs and that Eq. (\[Quasi-independence\]) and the condition on the correlation length hold simultaneously.
Let us now assume that $\Sigma$ and $\widetilde{\Sigma}$ quasi-independent. In this case, the microstates of the two systems are independent of each other to a very high degree of accuracy and we have (we suppress all state variables for simplicity of notation) $$p_{\alpha}(t)=p_{i}(t)p_{\widetilde{\alpha}}(t).$$ Now, using $$\ln p_{i}(t)p_{\widetilde{\alpha}}(t)=\ln p_{i}(t)+\ln p_{\widetilde{\alpha}}(t),$$ and the sum rule$${\textstyle\sum\limits_{\widetilde{\alpha}}}
p_{\widetilde{\alpha}}(t)=1,\ \
{\textstyle\sum\limits_{i}}
p_{i}(t)=1,$$ we find that $$S_{0}(t)\equiv-{\textstyle\sum\limits_{i}}
p_{i}(t)\ln p_{i}(t)-{\textstyle\sum\limits_{\widetilde{\alpha}}}
p_{\widetilde{\alpha}}(t)\ln p_{\widetilde{\alpha}}(t),$$ where the two terms in the above equations represent the entropies of the system and the medium$$S(t)=-{\textstyle\sum\limits_{i}}
p_{i}(t)\ln p_{i}(t),\ \ \ \widetilde{S}(t)=-{\textstyle\sum\limits_{\widetilde{\alpha}}}
p_{\widetilde{\alpha}}(t)\ln p_{\widetilde{\alpha}}(t), \label{Entropies}$$ respectively. This demonstration justifies the additivity of entropies$$S_{0}(t)=S(t)+\widetilde{S}(t) \label{Entropy Sum}$$ as a consequence of *quasi-independence* so that$$E_{0}\equiv E(t)+\widetilde{E}(t) \label{Energy_Sum_Approx}$$ also holds to a very good approximation. Note that we have neither assumed the medium nor the system to be in internal equilibrium in the above demonstration.
If the system and the medium fail to be quasi-independent because their mutual interaction cannot be neglected, then Eq. (\[Quasi-independence\]) is violated. In this case, the presence of this interaction acts as a constraint on $\Sigma_{0}.$ Consequently, the entropy now will be strictly less than the above entropy in Eq. (\[Entropy Sum\]). We denote this difference by $S_{0}^{(\text{int})}(t)\leq0$ [@Gujrati-Athermal-Entropy], which is defined by the following identity$$S_{0}(t)\equiv S(t)+\widetilde{S}(t)+S_{0}^{(\text{int})}(t)\text{.}
\label{Entropy_Loss_Sum}$$ This identity reduces to the above additivity in Eq. (\[Entropy Sum\]) provided $$\left\vert S_{0}^{(\text{int})}(t)\right\vert <<S(t);
\label{Quasi-independence-Entropy}$$ compare with Eq. (\[Quasi-independence\]). This inequality will in general hold only if the interaction energy is also negligible.
If the strong inequality in Eq. (\[Quasi-independence-Entropy\]) is not satisfied, we have lost the additivity property of the entropy. Let us assume that the strong inequality is satisfied for some large size of the system $\Sigma$. Now, as the size of the system decreases, which is what will happen on the way to considering physically infinitesimal volume elements used in Eq. (\[Entropy\_Additivity\_Thermodynamics\]), there comes a point where the strong inequalities in Eqs. (\[Quasi-independence\]) and (\[Quasi-independence-Entropy\]) are violated. This will destroy the additivity of the entropy as exhibited in Eq. (\[Entropy Sum\]).
The violation of entropy additivity occurs at intermediate sizes of the system, somewhere between the macroscopic size where Eqs. (\[Quasi-independence\]) and (\[Quasi-independence-Entropy\]) are valid, and small local or microscopic size containing a small number $dN$ of particles. For example, for $dN\approx10^{18},$ the surface to volume ratio for the volume element $dV$ is about $10^{-6}$, implying an almost imperceptible error in neglecting the interaction entropy $S_{0}^{(\text{int})}(t)$, provided the linear size of this region is large compared to not only to the inter-particle separation [@Landau-FluidMechanics p.1] but also the correlation length in the system. Under these conditions, the integrand in Eq. (\[Entropy\_Additivity\_Thermodynamics\]) truly refers to a “physically” infinitesimal volume element containing a very large number of particles. In this sense, our starting premise is similar to that adopted in the conventional non-equilibrium thermodynamics [@Donder; @deGroot; @Prigogine; @Prigogine-old]. In particular, our concept of internal equilibrium is no different than the concept of local equilibrium in conventional non-equilibrium thermodynamics, as observed in I, except that we require quasi-independence, which imposes the strong condition that not only the interaction energy be small but also that the linear size be larger than the correlation length. At present, there is some evidence that the correlation length in a glass forming system appears to increasing as the system approaches the glass transition [@Ottinger-2].
The discussion above also clarifies that the additivity of entropy is a consequence of the additivity of energy of various parts of the body and that the interaction energies between them must be negligible. This additivity of the energy and entropy was adopted in I. However, as we will be interested in considering parts of $\Sigma$ as subsystems in this work, the additivity of their entropy requires that their mutual interaction energies be also negligibly small compared to their individual internal energies, and that their linear sizes be large compared to the correlation lengths [@Ottinger-2]. These requirements put a strong condition on the sizes of subsystems.
Thermodynamic Potentials\[marker\_thermodynamic\_potentials\]
=============================================================
Fixed Number of Particles $N$ of the System $\Sigma$
----------------------------------------------------
Despite similarities between our approach and that adopted in conventional thermodynamics [@Donder; @deGroot; @Prigogine; @Prigogine-old], there were important differences noted in I. One of these was the discovery that the differences of the temperature and pressure of $\Sigma$ and $\widetilde
{\Sigma},$ which are the same as that of $\Sigma$ and $\Sigma_{0},$ play the role similar to internal variables. The second difference was that the Gibbs free energy in our approach exists even if the system is not in internal equilibrium, and involves the temperature and pressure of $\Sigma_{0}$ or $\widetilde{\Sigma};$ of course,we assume$\ $that the medium is in internal equilibrium; see I and Sect. \[marker\_internal\_equilibrium\]. Under this very weak assumption for the medium, its field variables such as the temperature $T_{0}$, pressure $P_{0}$, etc. are well defined, and are unaffected by whatever processes happen to be going on within the system $\Sigma$ or whether $\Sigma$ is homogeneous or inhomogeneous. When the number of particles $N$ in the system $\Sigma$ is held fixed, the appropriate thermodynamic potential is the Gibbs free energy, which is identified as$$G(T_{0},P_{0},t)=E(t)-T_{0}S(t)+P_{0}V(t), \label{Gibbs_Free_Energy}$$ where the observables $E(t),S(t)$ and $V(t)$ have explicit time-dependence for fixed $N$. The particular form of the Gibbs free energy is in accordance with the second law in Eq. (\[Second\_Law\_0\]), and remains valid even if the system is so far out of equilibrium that its temperature and pressure cannot be defined; see also Landau and Lifshitz [@Landau see Sect. 20]. It is not surprising, therefore, that it contains the temperature and pressure of the medium, which are well defined. Theses quantities do not exists for the system unless it happens to be at least in internal equilibrium. Even then, the Gibbs free energy is given by Eq. (\[Gibbs\_Free\_Energy\]) and contains the temperature and pressure of $\widetilde{\Sigma}$. As such, it does not represent a state function of the system. In engineering context, this quantity is also known as *exergy* or *availability* [@Keenan].
However, our form in Eq. (\[Gibbs\_Free\_Energy\]) differs from the local form of the Gibbs free energy in Eq. (\[Local\_Gibbs\_Free\_Energy\]), which contains the local temperature and pressure. One can argue that the identification of the Gibbs free energy in I was for the entire system, but that once we account for the inhomogeneity by considering subsystems, the Gibbs free energy for each subsystem will somehow become consistent with that in Eq. (\[Local\_Gibbs\_Free\_Energy\]). This is a reasonable possibility and we need to investigate this possibility. This issue is deferred to Sect. \[marker\_inhomogeneous\_arbitrary\].
As the method to identify the Gibbs free energy or other thermodynamic potentials that follow from the second law in Eq. (\[Second\_Law\_0\]) is going to be employed here several times, we briefly sketch the derivation for the sake of continuity. The full details are given in I. We do not assume the existence of the temperature, pressure, etc. of the system to include the situation in our discussion when the system is far away form equilibrium so that they are not defined. For simplicity, we consider a monatomic system of structureless particles under no external shear as in the previous work [@GujratiNETI]. Accordingly, we only consider the energy $E$, volume $V$ and the number of particles $N$ to describe the macrostate of the system at any instant $t$. No internal variables will be considered at this moment. The system and the medium are assumed not to be in equilibrium.
We use the additivity in Eq. (\[Entropy Sum\]) to write the entropy $S_{0}(t)$ of $\Sigma_{0}$ as the sum of the entropies $S(t)$ of the system and $\widetilde{S}(t)$ of the medium. If we also assume that the latter is in internal equilibrium, then we have$$S_{0}(E_{0},V_{0},N_{0},t)=S(E,V,N,t)+\widetilde{S}(\widetilde{E},\widetilde{V},\widetilde{N}); \label{Total_Entropy}$$ there is no explicit $t$-dependence in $\widetilde{S}(\widetilde{E},\widetilde{V},\widetilde{N})$ due to its internal equilibrium. With $N$ and $\widetilde{N}$ fixed, we expand $S_{0}$ in terms of the small quantities $E(t)$ and $V(t)$ of the system $$\widetilde{S}(\widetilde{E},\widetilde{V},\widetilde{N})\simeq\widetilde
{S}(E_{0},V_{0},\widetilde{N})-\left. \left( \frac{\partial\widetilde{S}}{\partial\widetilde{E}}\right) \right\vert _{\mathbf{X}_{0}}E(t)-\left.
\left( \frac{\partial\widetilde{S}}{\partial\widetilde{V}}\right)
\right\vert _{\mathbf{X}_{0}}V(t),$$ where the derivatives are evaluated at $E_{0},V_{0},\widetilde{N}$. However, as $\widetilde{N}$ is very close to $N_{0}$, there is no harm in evaluating the derivatives at $E_{0},V_{0},N_{0}$. This is the reason that we have used $\mathbf{X}_{0}$ above for the derivative. This approximation will be made throughout in this work. The error is inconsequential when the system is a very small part of the isolated system. It follows from the internal equilibrium of $\widetilde{\Sigma}$ that$$\left. \left( \frac{\partial\widetilde{S}}{\partial\widetilde{E}}\right)
\right\vert _{\mathbf{X}_{0}}=\frac{1}{T_{0}},\ \ \left. \left(
\frac{\partial\widetilde{S}}{\partial\widetilde{V}}\right) \right\vert
_{\mathbf{X}_{0}}=\frac{P_{0}}{T_{0}}. \label{Medium_Fields}$$ We observe that $\widetilde{S}\equiv\widetilde{S}(E_{0},V_{0},\widetilde{N})$ is a constant, which is independent of the system $\Sigma$. Thus,$$S_{0}(t)-\widetilde{S}\simeq S(E,V,N,t)-E(t)/T_{0}-P_{0}V(t)/T_{0}.
\label{Total_Subtracted_Entropy_0}$$ In terms of $$G(t)\equiv H(t)-T_{0}S(t),\ H(t)\equiv E(t)+P_{0}V(t), \label{Free_Energies}$$ we finally have $$S_{0}(t)-\widetilde{S}=S(t)-H(t)/T_{0}=-G(t)/T_{0}.
\label{Gibbs_Free_Energy_Entropy_Relation}$$
It immediately follows from Eq. (\[Second\_Law\_0\]) that the Gibbs free energy $G(t)$ of the system in Eq. (\[Gibbs\_Free\_Energy\]) continuously decreases as the system relaxes towards equilibrium, a result quite well known in classical thermodynamics [@Landau]: $$\frac{dG(t)}{dt}\leq0. \label{Gibbs_Free_Energy_Variation}$$ The function $G(t)$ continues to decrease and finally becomes identical to the equilibrium Gibbs free energy at the current temperature and pressure $T_{0},P_{0}$. If we abruptly change the temperature $T^{\prime}$ and pressure $P^{\prime}$ of the system in some state A$^{\prime}$, where the system was in equilibrium, to a new state A where the temperature and pressure are $T_{0},P_{0},$ respectively, at time $t=0$, then the *initial* values of the energy, volume and entropy at the new temperature and pressure remain equal to their respective equilibrium values in the previous state A$^{\prime
}$ as the microstate probabilities $p_{\alpha}(t)$ at $t=0\ $have not had any time to change. Thus, initially $$G(0)=E_{\text{A}}^{\prime}-T_{0}S_{\text{A}}^{\prime}+P_{0}V_{\text{A}}^{\prime}$$ in the state A; the quantities with a prime are the equilibrium values in the state A$^{\prime}$. The Gibbs free energy decreases in accordance with Eq. (\[Gibbs\_Free\_Energy\_Variation\]) and eventually becomes equal to its new equilibrium value$$G_{\text{A}}=E_{\text{A}}-T_{0}S_{\text{A}}+P_{0}V_{\text{A}},$$ where the quantities with the subscript A denote the equilibrium values in the new state A.
It should be noted that the equilibrium Gibbs free energy in the state A$^{\prime}$ before the abrupt change is$$G_{\text{A}}^{^{\prime}}=E_{\text{A}}^{\prime}-T^{\prime}S_{\text{A}}^{\prime
}+P^{\prime}V_{\text{A}}^{\prime},$$ so that the Gibbs free energy undergoes a discontinuity at $t=0$ due to the abrupt change:$$\Delta G_{\text{A}}^{^{\prime}}=(T^{\prime}-T_{0})S_{\text{A}}^{\prime
}-(P^{\prime}-P_{0})V_{\text{A}}^{\prime}.$$ Its magnitude and sign has nothing to do with the second law as the abrupt change is not a spontaneous process.
A similar looking quantity $\widehat{G}(t)$, see Eq. (\[Local\_Gibbs\_Free\_Energy\]) for its local analog in the local non-equilibrium thermodynamics [@Donder; @deGroot; @Prigogine; @Prigogine-old], $$\widehat{G}(t)\equiv\widehat{H}(t)-T(t)S(t),\ \widehat{H}(t)\equiv
E(t)+P(t)V(t), \label{Free_Energies_Incorrect}$$ which can be defined *only* when the system is under internal equilibrium and not otherwise, was shown to increase with time [@GujratiNETI] during relaxation $$\frac{d\widehat{G}(t)}{dt}\geq0 \label{Gibbs_Free_Energy_Variation_Incorrect}$$ in a cooling process. Since it does not always decrease with time, it cannot be taken as the Gibbs free energy; the latter is supposed to never increase as the system equilibrates spontaneously as happens with $G(t)$; see Eq. (\[Gibbs\_Free\_Energy\_Variation\]).
Fixed Volume $V$ of the System $\Sigma$
---------------------------------------
Instead of keeping the number of particles in $\Sigma$ fixed, let us keep its volume $V$ fixed so that the volume of the medium is also kept fixed. The number of particles $\widetilde{N}$ of the medium is no longer fixed. The entropy $\widetilde{S}(\widetilde{E},\widetilde{V},\widetilde{N})$ of the medium in Eq. (\[Total\_Entropy\]) is expanded in terms of small quantities $E$ and $N$ of the system. We follow the steps similar to those above and obtain$$\widetilde{S}(\widetilde{E},\widetilde{V},\widetilde{N})\simeq\widetilde
{S}(E_{0},\widetilde{V},N_{0})-\left. \left( \frac{\partial\widetilde{S}}{\partial\widetilde{E}}\right) \right\vert _{\mathbf{X}_{0}}E(t)-\left.
\left( \frac{\partial\widetilde{S}}{\partial\widetilde{N}}\right)
\right\vert _{\mathbf{X}_{0}}N(t), \label{Medium_Expansion_V}$$ where $\mathbf{X}_{0}$ stands for $E_{0},V_{0},N_{0}$ for reasons explained above in deriving Eq. (\[Medium\_Fields\]). Let us now introduce the chemical potential $\mu_{0}$ of the particle in the medium by the standard definition $$\ \ \left. \left( \frac{\partial\widetilde{S}}{\partial\widetilde{N}}\right) \right\vert _{\mathbf{X}_{0}}=-\frac{\mu_{0}}{T_{0}},$$ We thus find that in terms of $\widetilde{S}\equiv\widetilde{S}(E_{0},\widetilde{V},N_{0})$ $$S_{0}(t)-\widetilde{S}\simeq S(E,V,N,t)-E(t)/T_{0}+\mu_{0}N(t)/T_{0}=-[E(t)-T_{0}S(t)-\mu_{0}N(t)]/T_{0}, \label{Total_Subtracted_Entropy}$$ which identifies a different thermodynamic potential in this case as $$\Omega(t)\equiv E(t)-T_{0}S(t)-\mu_{0}N(t); \label{Thrmo_Potential}$$ this thermodynamic potential also uses the field variables of the medium. We should emphasize again that no assumption about the internal equilibrium of the system has been made. The system may or may not be in internal equilibrium. The application of the second law in Eq. (\[Second\_Law\_0\]) now gives$$\frac{d\Omega(t)}{dt}\leq0. \label{Second_Law_P(t)}$$
Fixed $N$ and $V$ of the System $\Sigma$
----------------------------------------
If both $N$ and $V$ are kept fixed, it is easy to follow the above derivation to conclude that the Helmholtz free energy$$F(t)\equiv E(t)-T_{0}S(t)$$ must continuously decrease as the system reaches equilibrium:$$\frac{dF(t)}{dt}\leq0.$$ Thus, the second law for an open system is expressed in terms of different thermodynamic potentials depending on which variables are held fixed.
Internal variables\[Marker\_Internal Variables\]
================================================
As said above, a suitable equilibrium macrostate description of the system requires a set of *independent* macroscopic observables that can be controlled by an experimentalist and whose values will allow the experimentalist to differentiate between different macrostates of the same system. It normally happens that experimentalists have a far less number of external controls than the possible extensive variables that can be used to characterize the macrostates. Thus, one does not characterize a macrostate, especially an equilibrium macrostate, by specifying all of the relevant extensive system quantities. For example, for a single component system, one normally uses $E,V$ and $N$ to specify the macrostate if there are no external shearing forces. Let us for the moment consider a system without external shear. Usually, one considers a system with fixed $N$; then $E$ and $V$ can be controlled by the two external variables $T_{0}$ and $P_{0}$ associated with the medium. However, these external variables need not necessarily control the local or internal structures in the system at all times during its evolution towards equilibrium. As Frenkel has observed, the local structures can be important when considering the structural relaxation in a glass or other non-equilibrium systems [@Frenkel p. 208]. For example, one can consider the average numbers of neighbors and next-neighbors of a given particle to describe the local structure in the system. These quantities multiplied by the number $N$ can play the role of internal variables. The corresponding conjugate variables, normally identified as “chemical potentials” or “affinity” for these internal variables usually vanish in equilibrium. Frenkel goes on and calculates viscoelastic effects due to structural changes and compares them with Maxwell’s model of elastic relaxation ** or an RC-circuit. This investigation by Frenkel [@Frenkel] shows that internal variables can play an important role in the temporal evolution towards equilibrium in some systems such as glasses. As such, they become an integral part of the description of any non-equilibrium system and determine the relaxation of the system [@Landau-FluidMechanics Sect. 78]. The internal variables are also called hidden variables or internal order parameters.
To introduce the concept of internal variables, let us consider our isolated system $\Sigma_{0}$ for which one can identify a set of *conserved* quantities, i.e. integrals of motion. For a mechanical system of $s$ degrees of freedom, the number of such integrals of motion are $2s-1$ [@Landau-Mechanics]. Of these integrals of motion, those that are *additive* play an important role in thermodynamics and statistical mechanics. The notable ones are the energy, and linear and angular momenta of the system, among others such as the polarization, magnetization, etc. For the moment, let us consider $\Sigma_{0}$ to be stationary. Its macrostate $\mathcal{M}_{0}$ is characterized by fixed internal energy $E_{0}$, volume $V_{0}$, particle number $N_{0}$ and other extensive observables, collectively denoted by $\mathbf{X}_{0}$. All these observables are constant for $\Sigma_{0}$. Let us consider the energy $E_{0}$, which is an integral of motion. It usually happens (see below for an example) that there are many different components $E_{0}^{(k)}$ of the energy whose total sum is the energy of $\Sigma_{0}$: $$E_{0}\equiv\sum_{k=1}^{n+1}E_{0}^{(k)}(t), \label{Energy_Sum}$$ where $n+1>1\ $is the number of energy components. It is $E_{0}$ that is a constant of motion, not the individual components $E_{0}^{(k)}(t)$; the latter will continue to change as the system evolves in time while maintaining Eq. (\[Energy\_Sum\]). Let $W_{0}(\mathbf{X}_{0},t)$ denote the number of microstates corresponding to the macrostate $\mathcal{M}_{0}$ at time $t$. At each instant $t$, the microstates in $W_{0}(\mathbf{X}_{0},t)$ can be partitioned into groups according to the possible values of $E_{0}^{(k)}(t)$. Because of the sum rule in Eq. (\[Energy\_Sum\]), only $n$ of the components are independent for a given $E_{0}$, which we take to be given by $k=1,2,\cdots,n$. We will denote this set by an $n$-vector $\mathbf{I}_{0}(t)$ whose elements are$\left\{ I_{0}^{(k)}=E_{0}^{(k)}(t),k=1,2,\cdots,n\right\}
$. Then, $$E_{0}^{(n+1)}(t)\equiv E_{0}-\sum_{k=1}^{n}I_{0}^{(k)}(t).$$ Let $W_{0}(\mathbf{X}_{0},\mathbf{I}_{0}(t),t)$ denote the number of microstates for a given $\mathbf{Z}_{0}(t)$. These microstates define a new macrostate, which we denote by $\mathcal{N}_{0}$. Obviously, $$W_{0}(\mathbf{X}_{0},t)\equiv\sum_{\mathbf{I}_{0}(t)}W_{0}(\mathbf{X}_{0},\mathbf{I}_{0}(t),t); \label{NumberMicrostate_Sum}$$ the sum is over all possible $\mathbf{I}_{0}(t)$. As the system evolves, different components $E_{0}^{(k)}(t)$ of $\mathbf{I}_{0}(t)$ evolve in time $t$, but $E_{0}$ and $\mathbf{X}_{0}$ remain fixed. Thus, a better understanding of the evolution of the system can be obtained by monitoring how the various components $E_{0}^{(k)}(t)$ change in time. For this, it is better to use $\mathbf{Z}_{0}\equiv(\mathbf{X}_{0},\mathbf{I}_{0}(t))$ to identify the macrostate $\mathcal{N}_{0}$ even though individual $E_{0}^{(k)}(t)$ cannot be controlled by the observer. As $E_{0}$ can be controlled by the observer, it is still the choice observable to be used for identifying a macrostate. This is even more true for the isolated system for which $E_{0}$ is a constant of motion. The $n$ components of $\mathbf{I}_{0}(t)$ then play the role of *internal variables* in developing non-equilibrium thermodynamics of the isolated system.
It is evident that the same “extended” description should also be useful for an open system $\Sigma$. The only difference between an open system and an isolated system is that not all elements of $\mathbf{X}$ remain fixed. Some of the observables, denoted by $\mathbf{X}^{\prime}$ are controlled by external field parameters $\mathbf{Y}_{0}^{\prime}$ (such as $T_{0}$,$P_{0}$, etc.) of the medium so that they do not remain fixed but continue to fluctuate about their mean $\mathbf{X}^{\prime}(t)$ that keeps changing in time. However, at least one of the extensive observables such as $N$ must be kept *constant* to quantify the size of the system [@Gujrati-review:Fluctuations]. Thus, for an open system, these observables can be replaced by the fields $\mathbf{Y}_{0}^{\prime}$, with the remaining observables remaining constant. We will denote the latter observables by $\mathbf{C}$ to remind us that they are constant. The open system can be either specified by $\mathbf{X}^{\prime}(t),\mathbf{C}$ or $\mathbf{Y}_{0}^{\prime},\mathbf{C}$. However, for the sake of convenience, we will continue to use $\mathbf{X}(t)$ rather than $\mathbf{X}^{\prime
}(t),\mathbf{C}$ or $\mathbf{Y}_{0}^{\prime},\mathbf{C}$. Let us now consider $\Sigma$ which is not in internal equilibrium so that it undergoes internal deformation due to relative motions between its various parts. If there are external strains on the system, they can be controlled by us from the outside. Hence, they will not be considered as internal variables. However, internal stresses acting on various parts of the system when there are no external strains on the system are beyond our control and must be treated as internal variables in describing the system. As said earlier, we can describe the internal forces acting on each part in terms of translation and rotation of its various parts; see Sect. \[marker\_Helmholtz\_theorem\]. These motions must be described by the use of suitable internal variables such as the linear and angular momenta, as was discussed in Sect. \[marker\_important\_assumptions\_I\].
As the internal variables are uncontrollable, their affinity in equilibrium must vanish as we prove now as a theorem.
\[marker\_chemical\_potential\]The affinity of an internal variable must vanish in equilibrium.
It is sufficient to prove the theorem for an isolated system. Also, we will prove it for the energy components in Eq. (\[Energy\_Sum\]). The extension to the general case is a trivial extension and will not be done here. As we are dealing with equilibrium, we consider equilibrium values of all the quantities, which are going to be represented by suppressing the argument $t$ as they are stationary. We now construct the following partition function for the isolated system$$Z_{0}(\mathbf{X}_{0},\mathbf{A}_{0})\equiv\sum_{\mathbf{I}_{0}}W_{0}(\mathbf{X}_{0},\mathbf{I}_{0})\exp\left\{ -\sum_{k=1}^{n}a_{0}^{(k)}I_{0}^{(k)}\right\} ,$$ where $\mathbf{a}_{0}$ is the the $n$-vector $\left\{ a_{0}^{(k)}\right\} $ equilibrium affinity. Such a partition function correctly describes the situation in which the $n$ internal variables are not constant but keep changing from microstate to microstate. We now observe that this partition function reduces to the equilibrium value $W_{0}(\mathbf{X}_{0})$ in Eq. (\[NumberMicrostate\_Sum\]) (where we take the limit $t\rightarrow\infty$), provided $A_{0}^{(k)}\equiv0$: $$A_{0}^{(k)}\equiv0,\ k=1,2,\cdots,n,$$ for each of the internal variable in the set $\mathbf{I}_{0\text{,eq}}$. This proves the theorem.
The above theorem deals with equilibrium affinities, and says nothing about the affinities of the internal variables when the system is out of equilibrium.
Let $\alpha$ denote one of the microstates associated with the macrostate $\mathcal{M}_{0},$ and $\beta$ one of the microstates associated with the macrostate $\mathcal{N}_{0}$. Then, using their probabilities $p_{\alpha}(t)$ and $p_{\beta}(t),$ we can determine the entropy of the two macrostates using the Gibbs formulation in Eq. (\[Gibbs entropy\]):
\[Macrostate\_entropy\]$$\begin{aligned}
S_{0}(\mathbf{X}_{0},t) & \equiv-\sum_{\alpha}p_{\alpha}(t)\ln p_{\alpha
}(t),\label{M0_entropy}\\
S_{0}(\mathbf{X}_{0},\mathbf{I}_{0}(t),t) & \equiv-\sum_{\beta}p_{\beta
}(t)\ln p_{\beta}(t). \label{N0_entropy}$$ For a macroscopically large system, the following standard statistical mechanical arguments can be used to highlight the maximum of the summand in Eq. (\[NumberMicrostate\_Sum\]). Let the maximum of the summand be denoted by $M_{0}(t)$, which occurs for some particular value $\overline{\mathbf{I}}_{0}(t)$ of $\mathbf{I}_{0}(t):$
$$M_{0}(t)\equiv W_{0}(\mathbf{X}_{0},\overline{\mathbf{I}}_{0}(t),t)$$ We separate the maximum contribution from the sum and rewrite Eq. (\[NumberMicrostate\_Sum\]) as follows:$$W_{0}(\mathbf{X}_{0},t)\equiv M_{0}(t)\left[ 1+\sum_{\mathbf{I}_{0}(t)\neq\overline{\mathbf{I}}_{0}(t)}\frac{W_{0}(\mathbf{X}_{0},\mathbf{I}_{0}(t),t)}{M_{0}(t)}\right] ,$$ where the sum is over all remaining $\mathbf{I}_{0}(t)$. It is normally the case that the ratio in the above sum is vanishingly small for a macroscopic system and that the sum can be neglected. In this case, we have$$W_{0}(\mathbf{X}_{0},t)\approx W_{0}(\mathbf{X}_{0},\overline{\mathbf{I}}_{0}(t),t). \label{Equality_W0}$$ For a macroscopically large open system such as $\Sigma$, the above equation is formally valid, except that we must replace $\mathbf{X}_{0}$ by $\mathbf{X}(t),$which stands for $\mathbf{X}^{\prime}(t),\mathbf{C}$, and $\overline{\mathbf{I}}_{0}(t)$ by $\overline{\mathbf{I}}(t)$:$$W(\mathbf{X}(t),t)\approx W(\mathbf{X}(t),\overline{\mathbf{I}}(t),t).
\label{Equality_W}$$
We now prove an important theorem about the nature of the entropy.
\[marker\_absence\_of\_internal\_equilibrium\]The entropy expressed only in terms of the observables when (independent) internal variables are present must explicitly depend on $t.$
We first consider the isolated system $\Sigma_{0}$ and prove the theorem for it. For $\overline{\mathbf{I}}_{0}(t)$ to be independent of (fixed) $\mathbf{X}_{0}$, it must surely have an explicit dependence on time. In other words, $\overline{\mathbf{I}}_{0}(t)$ must be a function of $\mathbf{X}_{0}$ and $t$. Let us assume that there is an explicit $t$-dependence in both $W_{0}$-functions in Eq. (\[Equality\_W0\]). As the entropy of the macrostate $\mathcal{M}_{0}$ is given by the sum over all microstates $W_{0}(\mathbf{X}_{0},t)~$in Eq. (\[System\_Entropy\]), it must explicitly depend on $t$. Thus, the theorem is satisfied. If, however, neither of the $W_{0}$-functions in Eq. (\[Equality\_W0\]) have any explicit $t$-dependence, then this is possible only if $\overline{\mathbf{I}}_{0}(t)$ becomes a function of $\mathbf{X}_{0}$ as the left side is only a function of $\mathbf{X}_{0}$. Since $\mathbf{X}_{0}$ is constant, $\overline{\mathbf{I}}_{0}(t)$ itself must be constant. The latter is the situation in equilibrium: $$\overline{\mathbf{I}}_{0\text{,eq}}=\overline{\mathbf{I}}_{0}(\mathbf{X}_{0})\text{ a constant}. \label{Equilibrium_I}$$
It follows that $\overline{\mathbf{I}}_{0}\equiv\overline{\mathbf{I}}_{0\text{,eq}}$ is no longer an independent variable when the system is in equilibrium. Obviously, this case is not covered by the theorem since $\overline{\mathbf{I}}_{0}$ is not independent. The entropy in this case is given by the Boltzmann formulation, cf. Eq. (\[Boltzmann entropy\]), and we have from Eq. (\[Equality\_W0\])$$S_{0}(\mathbf{X}_{0})\approx S_{0}(\mathbf{X}_{0},\overline{\mathbf{I}}_{0})\ \text{a constant}.$$
Let us now consider the special case when the macrostate $\mathcal{N}_{0}$ satisfies the condition of internal equilibrium. In this case, $W_{0}(\mathbf{X}_{0},\overline{\mathbf{I}}_{0}(t),t)$ does not explicitly depend on $t$ and should be written as $W_{0}(\mathbf{X}_{0},\overline{\mathbf{I}}_{0}(t))$ with $\overline{\mathbf{I}}_{0}(t)~$having an explicit time dependence. This entropy is again given by Eq. (\[Boltzmann entropy\]):$$S_{0}(\mathbf{X}_{0},\overline{\mathbf{I}}_{0}(t))=\ln W_{0}(\mathbf{X}_{0},\overline{\mathbf{I}}_{0}(t)). \label{N0_entropy_Internal_Equilibrium}$$ It now follows from Eq. (\[Equality\_W0\]) that $W_{0}(\mathbf{X}_{0},t)$ must have an explicit time-dependence due to the explicit $t$-dependence of $\overline{\mathbf{I}}_{0}(t)~$in $W_{0}(\mathbf{X}_{0},\overline{\mathbf{I}}_{0}(t))$. This is because different values of $\overline{\mathbf{I}}_{0}(t)$ will result in different values of $W_{0}(\mathbf{X}_{0},\overline{\mathbf{I}}_{0}(t))$, which can be treated as $W_{0}(\mathbf{X}_{0},t)$ associated with the macrostate $\mathcal{M}_{0}$ at different times. This is the first case considered above. Thus, $S_{0}(\mathbf{X}_{0},t)$ will have an explicit $t$-dependence even though $S_{0}(\mathbf{X}_{0},\overline{\mathbf{I}}_{0}(t))$ does not.
This proves the theorem for an isolated system.
Let us consider an open system such as $\Sigma$. Again, $\overline{\mathbf{I}}(t)$ must be a function of $\mathbf{X}(t)$ and $t$ to remain independent of $\mathbf{X}(t)$. Let us assume that there is an explicit $t$-dependence in the $W$-functions in Eq. (\[Equality\_W\]). As the entropy of the macrostate $\mathcal{M}$ of $\Sigma$ is given by the sum over all microstates $W(\mathbf{X}(t),t)~$in Eq.(\[M0\_entropy\]), it must explicitly depend on $t$. If, however, neither of the $W$-functions have any explicit $t$-dependence, then $\overline{\mathbf{I}}(t)$ becomes a function of $\mathbf{X}(t)$. In this case, it is not independent of $\mathbf{X}(t)$. This situation is then not relevant for the theorem.
The possibility in which $\overline{\mathbf{I}}(t)$ is independent of $\mathbf{X}(t)$, but $W(\mathbf{X}(t),\overline{\mathbf{I}}(t))$ has no explicit $t$-dependence, when the system is under internal equilibrium, is very important. Fixing $\mathbf{X}_{\text{IS}}\equiv\mathbf{X}(t)$ allows us to think of the system as an isolated system. Now, we can use the argument given above for the isolated system to conclude that different values of $\overline{\mathbf{I}}(t)$ will result in different values of $W(\mathbf{X}_{\text{IS}},\overline{\mathbf{I}}(t))$, which can be treated as $W(\mathbf{X}_{\text{IS}},t)$ associated with the macrostate $\mathcal{M}$ at different times. In other words, the macrostate $\mathcal{M}$ does not represent an internal equilibrium state. Thus, we conclude that a macrostate $\mathcal{N}$ under internal equilibrium results in a macrostate $\mathcal{M}$; the latter is, however, not in internal equilibrium.
This proves the theorem.
It follows from the above discussion that a general thermodynamic state can be taken to be a function of internal variables along with other observables and time $t$ when we deal with non-equilibrium states. For an open system in which many of the observables are controlled by external field parameters $\mathbf{Y}_{0}$ (such as $T_{0}$,$P_{0}$, etc.) of the medium, we can express $\overline{\mathbf{I}}_{\text{eq}}$ either as $$\overline{\mathbf{I}}_{\text{eq}}=\overline{\mathbf{I}}(\mathbf{X}_{\text{eq}}^{\prime},\mathbf{C}),$$ or as$$\overline{\mathbf{I}}_{\text{eq}}\equiv\overline{\mathbf{I}}(\mathbf{Y}_{0}^{\prime},\mathbf{C}).$$ Away from equilibrium, the internal variable $\overline{\mathbf{I}}(\mathbf{X}^{\prime}(t),\mathbf{C})$ differs from its equilibrium values $\overline{\mathbf{I}}_{\text{eq}}$, and is normally treated as an independent variable and plays an important role in the dynamics of the system as the latter strives to reach equilibrium. Thus, it is not surprising that internal variables are employed to specify the macrostate of a glass. In non-equilibrium thermodynamics, this fact has been recognized for quite some time [@Donder; @deGroot; @Prigogine; @Prigogine-old].
Internal variables can also be related to the presence of internal degrees of freedom in the particles of interest. The internal degrees are more common in polymers but can also occur in small molecules in the form of rotation about some internal axes. An example will clarify the point much better. Consider a polymerization process resulting in a system of polydisperse linear polymer chains of average molecular weight $\overline{M}$ in a solution [@GujratiRC]. The model is defined on a lattice of $N$ sites and volume $V=Nv_{0}$, with $v_{0}$ a constant representing the volume occupied by a lattice site. One normally uses $E$, $V$, $\overline{M}$ defined below in Eq. (\[Molecular\_weight\]), and the number of chains $p$ as the standard observables that can be used to identify the macrostate (equilibrium or not) of the polymer solution. In turn, these quantities are controlled by the temperature, pressure and the initiation-termination and propagation rates; the last two can be related to the initiation-termination activity controlling the number of endgroups, two for each polymer, and the middlegroup activity. These activities determine the corresponding affinity or “chemical potentials.” Let $N_{\text{m}}\equiv N-N_{\text{v}}$ denote the number of monomers, each monomer occupying a lattice site, in terms of the number of voids or sites not covered by monomers $N_{\text{v}}$ so that $$\overline{M}\equiv\frac{N_{\text{m}}}{p}. \label{Molecular_weight}$$ In terms of the number of middle groups $N_{\text{M}}\equiv N_{\text{m}}-2p,$ or $N_{\text{m}}$, the number of chemical bonds in the $p$ polymers is given by$$N_{\text{B}}\equiv N_{\text{M}}+p=N_{\text{m}}-p.$$
There are two kinds of energy in the model [@GujratiRC]. One kind of energy is due to mutual interactions of voids (v) with the end (E) and middle (M) groups, and the mutual interactions between chemically unbonded M and E. Let $N_{ij},i,j=$v,M or E, denote the number of nearest-neighbor contacts $ij,i\neq j$, and $\varepsilon_{ij}$ the corresponding interaction energies, respectively. The other kind of energy is due to intrachain gauche bonds (g), and hairpin turns (hp). Their energies are $E_{\text{g}}\equiv
\varepsilon_{\text{g}}N_{\text{g}}$ for gauche bonds and $E_{\text{hp}}\equiv\varepsilon_{\text{hp}}N_{\text{hp}}$ for hairpin turns; here $N_{\text{g}},$ and $N_{\text{hp}}$ denote the number of gauche bonds and hairpin turns and parallel bonds and $\varepsilon_{\text{g}},$ and $\varepsilon_{\text{hp}}$ are their energies. In addition, there is a mutual interaction energy between two parallel (chemical) bonds, which may belong to the same or different polymers. Let $N_{\text{P}}\,$denote the number parallel bonds, each of energy $\varepsilon_{\text{P}}$. Thus,$$E\equiv{\displaystyle\sum\limits_{i\neq j\text{:v,M,E}}}
\varepsilon_{ij}N_{ij}+\varepsilon_{\text{g}}N_{\text{g}}+\varepsilon
_{\text{hp}}N_{\text{hp}}+\varepsilon_{\text{P}}N_{\text{P}}\equiv{\displaystyle\sum\limits_{i\neq j\text{:v,M,E}}}
E_{ij}+E_{\text{g}}+E_{\text{hp}}+E_{\text{P}}, \label{Energy_Partition}$$ where we have introduced $E_{ij},E_{\text{g}},E_{\text{hp}}$ and $E_{\text{P}}$ with obvious definitions. We thus observe that the energy can be partitioned into six extensive energies, five of which can be taken as internal variables.
To summarize, we conclude that the quantities that *cannot* be controlled by the observer can be identified as the *internal variables*. This statement should not be taken literary as what is considered uncontrollable today may not remain so in the future. Thus, to some degree, the decision to identify the internal variables is left to the observer. For us, any variable that cannot be controlled to have a fixed value when the system is out of equilibrium will be taken as an internal variable [@Maugin]. It should also be noted that the number of internal variables is not a unique number for a given system. For example, to describe local structures in a monatomic system [@Frenkel], one can consider any number of neighboring particles (neighbors, next-neighbors, next-to-next neighbors, and so on). Thus, a choice will have to be made to see how many of them are useful in a given experiment or investigation. This certainly gives rise to an additional complication in the study of non-equilibrium system.
Our approach allows us to associate affinity in a formal sense with all internal variables. This is how the classical non-equilibrium thermodynamics has been developed [@Donder; @deGroot; @Prigogine; @Prigogine-old]. As observed by Landau and Lifshitz [@Landau-FluidMechanics], the use of internal variables in a modern way can be traced to Mandelstam and Leontovich [@Mandelstam]; see also Pokrovski [@Pokrovski]. Under the internal equilibrium assumption, Prigogine addresses the issue of internal variables (orientation of a molecule, deformation due to flow, elastic deformation, etc.) in Sect. 11, Chapter III of his classic book [@Prigogine-old], or in Sect. 10.4 in the modern version [@Prigogine], and couples them to their “chemical potentials” or affinities. Indeed, Prigogine and Mazur were the first one to do this in their classic paper [@Prigogine-Mazur]; see also Coleman and Gurtin [@Coleman]. The issue of the internal variables is also discussed in Sect. 6, Ch. 10 in [@deGroot]. Pokrovski [@Pokrovski] provides a very illuminating discussion of internal variables and their role in determining the internal energy. Thus, we will treat internal variables as additional thermodynamic extensive quantities or “observables” similar to the number of chemical species in chemical reactions that can be controlled by affinities or chemical potentials. More recently, the idea has also been visited by Bouchbinder and Langer [@Langer].
Thermodynamics of a simple rotating body\[marker\_Rotating\_Systems\]
=====================================================================
General Case
------------
We will find it convenient for later use to consider observing a body in different frames of reference; see also Appendices \[Appd\_Frames\] and \[Appd\_Rotating\_System\_0\]. For concreteness, we consider the system $\Sigma$ and assume that no internal variables and no other observables besides the energy, volume and number of particles are present; the latter can be added easily as we will discuss later. We will consider three special frames: the lab frame denoted by $\mathcal{L}$, an intermediate frame $\mathcal{I}$, with its axes parallel to those of and moving with respect to $\mathcal{L}$ with a velocity $\mathbf{V}(t)$, and a frame $\mathcal{C}$ with its origin common with $\mathcal{I}$ and rotating with respect to it with an angular velocity $\boldsymbol{\Omega}(t)$. Let $\mathbf{R}(t)$ denote the location of the origins of $\mathcal{I}$ and $\mathcal{C}$ in the lab frame $\mathcal{L}$ at time $t$ with $\mathbf{R}(t=0)=0.$ Let $\mathbf{r}_{\mathcal{C}}(t)$ denote the coordinate of a particle of $\Sigma$ in the $\mathcal{C}$ frame, and $\mathbf{v}_{\mathcal{C}}(t)$ its velocity in this frame at time $t$. Its coordinate $\mathbf{r}_{\mathcal{L}}$ in the lab frame $\mathcal{L}$ is given by$$\mathbf{r}_{\mathcal{L}}\equiv\mathbf{R}+\mathbf{r}_{\mathcal{C}};
\label{L_C_r_Relation}$$ its velocity is given by Eq. (\[velocity\_Relation\]). As shown in the Appendix \[Appd\_Frames\], the energy of the particle in the two frames are related as shown in Eq. (\[Energy\_Transformation\_particle\]). Let us consider $\mathcal{I}$ to be the frame in which the center of mass of the body is at the origin. we will call it the center of mass frame for the body. Then, applying the above two relations to all the particles in the system and averaging over all allowed microstates [@Gujrati-Symmetry], which is carried out later in Sect. \[marker\_Statistical\_averaging\], we obtain that the energy of the system in the three frames are related as shown in Eqs. (\[Internal\_Energy\]) and (\[Internal\_Energy\_Iframe\]):
\[Internal\_Energy\_LICFrame\]$$\begin{aligned}
E_{\mathcal{C}} & =E_{\mathcal{L}}-\frac{\mathbf{P}^{2}}{2M}-\mathbf{M}\cdot\mathbf{\Omega=}E_{\mathcal{I}}\mathbf{-\mathbf{M}\cdot\mathbf{\Omega},}\label{Internal_Energy_LCFrame}\\
E_{\mathcal{I}} & =E_{\mathcal{L}}-\frac{\mathbf{P}^{2}}{2M},
\label{Internal_Energy_ICFrame}$$ where $\mathbf{P}$ and $\mathbf{M}$ are introduced in Eq. (\[System\_Definition\]). We have not used the overbar to express the statistical averages as explained in Sect. \[marker\_Statistical\_averaging\] but is implied.
We first prove the following theorem:
\[Theorem\_Equal\_Entropy\]The entropy of a system is the same in all three frames $\mathcal{L}$,$\mathcal{I}$, and $\mathcal{C}$.
To prove the theorem, we proceed as follows. Only for simplicity of the argument and presentation, we focus on a system with *fixed* $V=V(t)$ and $N$ at some instant $t$. The extension to considering other extensive variables is trivial. Consider observing the system simultaneously at $t$ in these frames. It is evident that corresponding to each pair $\mathbf{r}_{\mathcal{L}}\mathbf{,p}_{\mathcal{L}}$ of the coordinates and momenta of a given particle in the lab frame $\mathcal{L}$ at this moment, there is a unique pair $\mathbf{r}_{\mathcal{I}}\mathbf{,p}_{\mathcal{I}}$ and $\mathbf{r}_{\mathcal{C}}\mathbf{,p}_{\mathcal{C}}$ in the other two frames. This is true of all the particles. The collection of positions and momenta of all the particles defines a point in the phase space. In classical statistical mechanics, a microstate of the system is identified by a small volume element of size $(2\pi\hbar)^{3N}$ about a point in the phase space. Thus, corresponding to each microstate $i$ ($=i_{\mathcal{L}},i_{\mathcal{I}},$ or $i_{\mathcal{C}}$) in one frame, there exists a *unique* microstate in the other two frames. The uniqueness of microstate-mapping ensures that their probabilities in the three frames are also equal:$$p_{i_{\mathcal{L}}}=p_{i_{\mathcal{I}}}=p_{i_{\mathcal{C}}.}
\label{Equal_Microstate_Probabilities}$$
Let us consider all the microstates of the same energy $E_{\mathcal{C}}$ in the $\mathcal{C}$ frame at time $t$, and let $W(t)\equiv$ $W(E_{\mathcal{C}},t)$ denote their number and $\boldsymbol{p}_{\mathcal{C}}(t)$ the set of their probabilities (not to be confused with momenta $\mathbf{p}_{\mathcal{C}}$, etc.). Because of the uniqueness of the mapping of these microstates noted above, not only the number of microstates in the three frames are the same at that instant$$W(t)\equiv W(E_{\mathcal{C}},t)\equiv W(E_{\mathcal{L}},t)\equiv
W(E_{\mathcal{I}},t),$$ but also the set of their probabilities $$\boldsymbol{p}_{\mathcal{C}}(t)\equiv\boldsymbol{p}_{\mathcal{L}}(t)\equiv\boldsymbol{p}_{\mathcal{I}}(t);$$ however, their energies are different as given by Eqs. (\[Internal\_Energy\_LICFrame\])(\[Internal\_Energy\_LCFrame\]) and (\[Internal\_Energy\_ICFrame\]). This immediately shows that the entropies using the general Gibbs formulation in Eq. (\[Gibbs entropy\]) are equal in the three frames: $$S_{\mathcal{C}}(E_{\mathcal{C}},t)=S_{\mathcal{I}}(E_{\mathcal{I}},t)=S_{\mathcal{L}}(E_{\mathcal{L}},t), \label{Equal_Entropy}$$ whether the system is in internal equilibrium or not. This proves the theorem.
It should be noted that the center-of-mass kinetic energy $\mathbf{P}^{2}/2M$ is the same for all microstates in $W(t)$. Similarly, it follows from Eq. (\[Ec\_Mc\_Relation\]) that even $\mathbf{\mathbf{M}\cdot\mathbf{\Omega}}$ is the same for all microstates in $W(t);$ see also the discussion leading to Eq. (\[Av\_M\]). Thus, the three energies only differ by some constants at each instant $t$.
System under Internal Equilibrium
---------------------------------
We now specialize and assume the existence of the internal equilibrium, so that all microstates are equally probable$$p_{i}(t)=1/W(t),i=1,2,\cdots,W(t). \label{Instantaneous probabilities}$$ Hence, the three entropies are each equal to $$S(t)=\ln W(t). \label{Entropy_Frames}$$
It follows from Eq. (\[Equal\_Entropy\]) that there is no reason to use different subscripts to distinguish the entropies. Accordingly, we will use $S$ to represent the entropies in different frames; their energy arguments will of course depend on the frame of reference. The arguments $\mathbf{V}$ and $\mathbf{\mathbf{\Omega}}$ above are actually external parameters that are not extensive. We will show below that the entropies in the $\mathcal{I}$ and $\mathcal{L}$ frames are actually functions of extensive quantities $\mathbf{P}$ and $\mathbf{\mathbf{M}}$ that are conjugate to $\mathbf{V}$ and $\mathbf{\mathbf{\Omega}}$, respectively; cf. Eq. (\[Field\_Variables\]).
Statistical Averaging over Allowed Microstates\[marker\_Statistical\_averaging\]
--------------------------------------------------------------------------------
We now investigate the consequences of statistical averaging over microstates with non-zero probabilities [@Gujrati-Symmetry] and show that its consequences are the same as expressed in Eqs. (\[Internal\_Energy\_LICFrame\]) and (\[Equal\_Entropy\]). We first note that $\mathbf{M}$ in Eq. (\[System\_Definition\]) depends on the coordinates and momenta of the particles, but this is not the case with $\mathbf{P}$, even though both are extensive quantities. As $E_{\mathcal{C}}$ in Eq. (\[Internal\_Energy\]) or (\[Internal\_Energy\_LCFrame\]) is for a microstate determined by the coordinates and momenta of the particles, we need to average it using microstate probabilities in Eq. (\[Instantaneous probabilities\]). Averaging over various microstates relates the average energies in the two frame. We use an overbar, see Eqs. (\[Average 0\]) and (\[Average\]), to denote the average. We find that the same form also describes the desired relation between the average energies:$$\overline{E}_{\mathcal{C}}(t)=\overline{E}_{\mathcal{L}}(t)-\frac
{\mathbf{P}(t)^{2}}{2M}-\overline{\mathbf{M}}(t)\cdot\mathbf{\Omega
}(t)=\overline{E}_{\mathcal{I}}(t)\text{ }\mathbf{-\overline{\mathbf{M}}}(t)\mathbf{\cdot\mathbf{\Omega}}(t)\mathbf{,} \label{Internal_Energy_0}$$ where $$\overline{E}_{\mathcal{I}}(t)=\overline{E}_{\mathcal{L}}(t)-\frac
{\mathbf{P}^{2}}{2M}. \label{Internal_Energy_1}$$ The momentum $\mathbf{P}$, of course, does not require any averaging as noted above. Eq. (\[Internal\_Energy\_0\]) is valid at each instance $t$. We can also take the statistical average of Eq. (\[Angular\_momentum\_C\]) to obtain$$\overline{E}_{\mathcal{C}}(t)=\overline{E}_{\mathcal{L}}(t)-\frac
{\mathbf{P}(t)^{2}}{2M}-{\displaystyle\sum_{j}}
m_{j}\overline{\mathbf{r}_{j}\cdot\left( \mathbf{v}_{j}\times\mathbf{\Omega
}\right) }-\frac12
{\displaystyle\sum_{j}}
m_{j}\overline{(\mathbf{\Omega\times r}_{j})^{2}}, \label{Internal_Energy_2}$$ where the two sums are over all the particles in the system. Here $\mathbf{r}_{j}$ and $\mathbf{v}_{j}$ are the instantaneous position and velocity of the $j$th particle in a microstate with respect to the $\mathcal{C}$ frame; we have suppressed the subscript $\mathcal{C}$ from $\mathbf{r}_{j}$ and $\mathbf{v}_{j}$ for the sake of notational simplicity. In the last equation, the third contribution is due to the relative motion of the particles with respect to the $\mathcal{C}$ frame. Indeed, the average of Eq. (\[Angular\_Momentum\_Work\]) immediately yields$$\mathbf{\overline{\mathbf{M}}}(t)\mathbf{\cdot\mathbf{\Omega}}(t)={\displaystyle\sum_{j}}
m_{j}\overline{\mathbf{r}_{j}\cdot\left( \mathbf{v}_{j}\times\mathbf{\Omega
}\right) }+{\displaystyle\sum_{j}}
m_{j}\overline{(\mathbf{\Omega\times r}_{j})^{2}}.
\label{Angular_Momentum_Work_0}$$ The third contribution in Eq. (\[Internal\_Energy\_2\]) and the first contribution in Eq. (\[Angular\_Momentum\_Work\_0\]) vanish when the system is in internal equilibrium because of the absence of any relative motion in that case; see Theorem \[marker\_Uniform\_Motion\].
Since $\overline{E}_{\mathcal{I}}$ also does not depend on the velocity $\mathbf{V}$, a similar averaging of Eqs. (\[Ec\_V\_Relation\]) and (\[Ec\_Mc\_Relation\]) gives us
\[Av\_Derivatives\]$$\begin{aligned}
\left( \frac{\partial\overline{E}_{\mathcal{C}}(t)}{\partial\mathbf{V}(t)}\right) _{\overline{E}_{\mathcal{I}}\mathbf{,}V,N,\mathbf{\Omega}} &
=0\mathbf{,}\label{Av_P}\\
\left( \frac{\partial\overline{E}_{\mathcal{C}}(t)}{\partial\mathbf{\Omega
}(t)}\right) _{\overline{E}_{\mathcal{I}}\mathbf{,}V,N} & =-\overline
{\mathbf{M}}(t). \label{Av_M}$$ Comparing the above equations with the equations in the Appendix \[Appd\_Rotating\_System\_0\], we see that there is no reason to make a distinction between $\mathbf{\mathbf{M}}(t)$, used in the proof above, and $\mathbf{\overline{\mathbf{M}}}(t)$ or the average energies and the energy used above in the proof. This justifies not using overbars to indicate statistical averages in Eq. (\[Internal\_Energy\_LICFrame\]).
Since the entropy $S(t)$ in Eq. (\[Entropy\_Frames\]) is fixed for fixed $\overline{E}_{\mathcal{I}},V,N,$ and $\mathbf{\mathbf{\Omega,}}$ we can express the above two derivatives at fixed $S$ **** instead of **** fixed $\overline{E}_{\mathcal{I}}$:
$$\left( \frac{\partial\overline{E}_{\mathcal{C}}(t)}{\partial\mathbf{V}(t)}\right) _{S\mathbf{,}V,N,\mathbf{\Omega}}=0\mathbf{,}\left(
\frac{\partial\overline{E}_{\mathcal{C}}(t)}{\partial\mathbf{\Omega}(t)}\right) _{S\mathbf{,}V,N,\mathbf{V}}=-\overline{\mathbf{M}}(t).
\label{Av_EC_Derivatives}$$
The above equation is similar to the well known result [@Landau Sect. 11] in equilibrium statistical mechanics that the statistical average of the derivatives of the the energy with respect to external parameters ($\mathbf{V}$ and $\boldsymbol{\Omega}$) should be taken at *constant entropy* and other extensive quantities. We have extended this result to internal equilibrium now. Introducing the following standard derivatives $$\left( \frac{\partial\overline{E}_{\mathcal{C}}(t)}{\partial S(t)}\right)
_{V,N,\mathbf{V},\mathbf{\Omega}}=T(t),\left( \frac{\partial\overline
{E}_{\mathcal{C}}(t)}{\partial V(t)}\right) _{S\mathbf{,}N,\mathbf{V},\mathbf{\Omega}}=-P(t) \label{Av_EC_Derivatives_1}$$ defining the temperature and pressure of the system, we can write down the following differential identity$$d\overline{E}_{\mathcal{C}}=T(t)dS(t)-P(t)dV(t)\ \mathbf{-\ }\overline
{\mathbf{M}}\ (t)\cdot d\boldsymbol{\Omega}(t)\boldsymbol{.}
\label{EC_differential}$$ It should be noted that because of Eq. (\[Av\_P\]), the average energy $\overline{E}_{\mathcal{C}}(t)$ does not depend on the velocity of the frames $\mathcal{I~}$and $\mathcal{C}$. Thus, there is no reason to keep $\mathbf{V}$ fixed in the various derivatives in Eqs. (\[Av\_Derivatives\]-\[Av\_EC\_Derivatives\_1\]).
For $$\overline{E}_{\mathcal{I}}=\overline{E}_{\mathcal{C}}+\mathbf{\overline
{\mathbf{M}}\cdot\mathbf{\Omega,}}$$ we find that $$d\overline{E}_{\mathcal{I}}=T(t)dS(t)-P(t)dV(t)\ \mathbf{+\ }\boldsymbol{\Omega}(t)\cdot d\overline{\mathbf{M}}(t), \label{EI_differential}$$ which is an extension of the result given in Landau and Lifshitz [@Landau Sect. 26] to the internal equilibrium. The point to note is that the entropy $S(t)$ in the $\mathcal{I}$ frame is a function of the conjugate variable $\overline{\mathbf{M}}(t)$ instead of $\boldsymbol{\Omega}(t)$. However, for $$\overline{E}_{\mathcal{L}}(t)=\overline{E}_{\mathcal{C}}(t)+\frac
{\mathbf{P}(t)^{2}}{2M}+\overline{\mathbf{M}}(t)\cdot\mathbf{\Omega}(t),$$ we also find an additional contribution due to $\mathbf{V}$:$$d\overline{E}_{\mathcal{L}}(t)=T(t)dS(t)-P(t)dV(t)+\mathbf{V}(t)\cdot
d\mathbf{P}(t)\mathbf{+}\boldsymbol{\Omega}(t)\cdot d\overline{\mathbf{M}}(t)
\label{EL_differential}$$ given in terms of all extensive quantities. The additional contribution due to the momentum differential $d\mathbf{P}(t)$ is due to the velocity of the system as a whole and is important to include in the lab frame. For example, such a contribution is needed to describe the flow of a superfluid in which the normal and superfluid components have different velocities so that the superfluid cannot be considered at rest in any frame [@Landau-FluidMechanics see Eq. (130.9)]. We will need to allow for this possibility when we extend our approach of nonequilibrium thermodynamics to inhomogeneous systems where different subsystems will undergo relative motion. It follows form Eq. (\[EL\_differential\]) that the *drift velocity* of the center of mass of the system is given by$$\left( \frac{\partial\overline{E}_{\mathcal{L}}(t)}{\partial\mathbf{P}(t)}\right) _{S\mathbf{,}V,N,\overline{\mathbf{M}}}=\mathbf{V}(t).
\label{Drift_Velocity}$$ Similarly, the angular velocity is given by$$\left( \frac{\partial\overline{E}_{\mathcal{L}}(t)}{\partial\overline
{\mathbf{M}}(t)}\right) _{S\mathbf{,}V,N,\mathbf{P}}=\boldsymbol{\Omega}(t).
\label{Angular_Velocity}$$
We again observe that the entropy in the lab frame $\mathcal{L}$ is a function of the extensive conjugate quantities $\mathbf{P}(t)$ and $\overline
{\mathbf{M}}(t)$ rather than the external parameters $\mathbf{V}$ and $\mathbf{\mathbf{\Omega}}$.
From now on, we will not use the overbar to show statistical averages for the sake of notational simplicity.
It is clear from Eq. (\[EC\_differential\]) that we must treat $E_{\mathcal{C}}(t)$ as a function of $S(t),V(t)$ and $\boldsymbol{\Omega}(t)$ for constant $N$. Alternatively, we must treat $S(t)$ as a function of $E_{\mathcal{C}}(t),V(t)$ and $\boldsymbol{\Omega}(t):$$$S_{\mathcal{C}}(t)\equiv S_{\mathcal{C}}(E_{\mathcal{C}}(t),V(t),\boldsymbol{\Omega}(t),N), \label{Functional_dependence}$$ which is identical to the functional dependence shown in Eq. (\[Equal\_Entropy\]), except that we no longer have an explicit $t$-dependence because of internal equilibrium. The important point to observe is that the entropy is a function of not only the energy in the $\mathcal{C}$ frame, but is also a function of the angular velocity of the reference frame when rotation is involved.
Same Temperature and Pressure in Different Frames
-------------------------------------------------
We now make an important observation. It follows from Eqs. (\[EI\_differential\]) and (\[EL\_differential\]) that $$\left( \frac{\partial E_{\mathcal{I}}(t)}{\partial S(t)}\right)
_{V,N,\mathbf{M}}=\left( \frac{\partial E_{\mathcal{L}}(t)}{\partial
S(t)}\right) _{V,N,\mathbf{P},\mathbf{M}}=T(t), \label{EI_EL_Derivatives}$$ obtained by differentiating with respect to $S(t).$ Similar equations are obtained when we differentiate with respect to $V(t).$ $$\left( \frac{\partial E_{\mathcal{I}}(t)}{\partial V(t)}\right)
_{S\mathbf{,}N,\mathbf{M}}=\left( \frac{\partial E_{\mathcal{L}}(t)}{\partial
V(t)}\right) _{S\mathbf{,}N,\mathbf{V},\mathbf{M}}=-P(t).
\label{EI_EL_Derivatives_1}$$ These equations are identical to the derivatives in Eq. (\[Av\_EC\_Derivatives\_1\]) and show that the internal temperature $T(t)$ and the internal pressure $P(t)$ are the same in the three frames. Moreover, it is the same entropy function $S(t)$ that appears in Eq. (\[EC\_differential\]) also appears in Eqs. (\[EI\_differential\]) and (\[EL\_differential\]). In other words, the entropy is the same in all frames, except that the arguments are different.
Thermodynamics Potentials and Gibbs Fundamental Relation for a Homogeneous System with Translational Motion\[marker\_homogeneous\]
==================================================================================================================================
Thermodynamic Potentials for a System under Arbitrary Conditions\[marker\_homogeneous\_arbitrary\]
--------------------------------------------------------------------------------------------------
### Fixed number of particles $N$ of the system $\Sigma$
Before we discuss the inhomogeneous case, let us consider the homogeneous situation considered in I and revisited briefly in Sect. \[marker\_thermodynamic\_potentials\], and extend it to the case when the system $\Sigma$ moves as a whole with a linear momentum $\mathbf{P}$. We still assume that $\Sigma_{0}$ is at rest. Because of the linear momentum conservation, the linear momentum of the center of mass of $\widetilde{\Sigma
}$ is $-\mathbf{P}$. Thus, the centers of mass $\Sigma$ and $\widetilde
{\Sigma}$ are moving towards each other. For simplicity, we will assume the absence of overall intrinsic rotation for $\Sigma$ and $\widetilde{\Sigma}$ individually. This can easily be incorporated as we do in the next section. Then we only need to consider the orbital angular momentum $\mathbf{L}_{0}$ of $\Sigma_{0}$ in terms of the locations $\mathbf{R}$ and $\widetilde
{\mathbf{R}}$ of the centers of mass of $\Sigma$ and $\widetilde{\Sigma}$, respectively. It is clear that $\mathbf{L}_{0}$ always vanishes since the centers of mass of $\Sigma$ and $\widetilde{\Sigma}$ are moving towards or away from each other so that $\mathbf{P}$ and $\mathbf{R-}\widetilde
{\mathbf{R}}$ are colinear :$$\mathbf{R}\times\mathbf{P-}\widetilde{\mathbf{R}}\times\mathbf{P}=0.$$
The (average) internal energies of $\Sigma$ and $\widetilde{\Sigma}$ in their center of mass frames (the $\mathcal{C}$ frame) are $$\begin{aligned}
E^{\text{i}} & =E-\mathbf{P}^{2}/2M,\\
\widetilde{E}^{\text{i}} & =\widetilde{E}-\mathbf{P}^{2}/2\widetilde{M},\end{aligned}$$ while $E$ and $\widetilde{E}$ denote their total energies in the lab frame $\mathcal{L}$, respectively; see Eq. (\[Internal\_Energy\_LICFrame\]). However, because of the extreme large size of $\widetilde{\Sigma}$, its mass $\widetilde{M}$ satisfies the inequality $\widetilde{M}>>M$, so that we can replace $\widetilde{E}^{\text{i}}$ by$\ \widetilde{E}~$without any appreciable error. The entropy $S$ of $\Sigma$ is a function of the internal energy $\widetilde{E}^{\text{i}};$ however, this is not relevant for our argument here if we are only interested in identifying the appropriate thermodynamic potential for the system. The energy $$E_{0}=E+\widetilde{E}$$ of $\Sigma_{0}$ remains constant in time. As discussed above, the additivity of energy is valid under the assumption that the interaction energy $E_{0}^{(\text{int})}(t)$ between the system and the medium is negligible. This ensures that the entropies are also additive. In the lab frame $\mathcal{L}$, the entropies of the $\Sigma$ and $\widetilde{\Sigma}$ are obtained by considering their entropies in respective rest frames $\mathcal{C}$ and $\widetilde{\mathcal{C}};$ they are $S(E^{\text{i}},V,N,t)$ and $\widetilde{S}(\widetilde{E}^{\text{i}},\widetilde{V},\widetilde{N},t)$, respectively; recall that we have set $\mathbf{\Omega}=0$ for each of them$.$
Using the fact that the medium is under internal equilibrium, we modify Eq. (\[Total\_Entropy\]) to reflect the dependence on internal energies to obtain$$S_{0}(E_{0},V_{0},N_{0},t)=S(E^{\text{i}},V,N,t)+\widetilde{S}(\widetilde
{E}^{\text{i}},\widetilde{V},\widetilde{N})\simeq S(E^{\text{i}},V,N,t)+\widetilde{S}(\widetilde{E},\widetilde{V},\widetilde{N}).
\label{Total_Entropy_0}$$ We now expand and follow the steps in arriving at Eq. (\[Total\_Subtracted\_Entropy\_0\]); the steps are unaffected by the motion of $\Sigma$. We thus obtain$$S_{0}(t)-\widetilde{S}\simeq S(E^{\text{i}},V,N,t)-E(t)/T_{0}-P_{0}V(t)/T_{0}$$ in terms of the energy and volume of the system. We can now identify the Gibbs free energy and enthalpy in the lab frame $\mathcal{L}$ in terms of the energy of the system:$$G(t)=E(t)-T_{0}S(t)+P_{0}V(t),\ H(t)=E(t)+P_{0}V(t); \label{G-H_lab}$$ compare with the Gibbs free energy in Eq. (\[Gibbs\_Free\_Energy\]). Thus, the second law in terms of the Gibbs free energy remains unchanged and is given by Eq. (\[Gibbs\_Free\_Energy\_Variation\]).
In the center of mass frame $\mathcal{C}$ of the system, the Gibbs free energy and the enthalpy of the system are given by$$G^{\text{i}}(t)=E^{\text{i}}(t)-T_{0}S(t)+P_{0}V(t),\ H^{\text{i}}(t)=E^{\text{i}}(t)+P_{0}V(t). \label{G-H_CM}$$ Note that the above functions depend on the internal energy and not the energy of the system $\Sigma$. But they are not useful in the lab frame in which the system is being observed. Thus, we conclude that the overall motion of the system does not change the enthalpy and the Gibbs free energy; we must use the appropriate energy in the frame of observation; the temperature and the pressure of the medium are not affected by the choice of the frame as noted near the end of Sect. \[marker\_Rotating\_Systems\]. Similarly, the entropy of the system is unaffected by the choice of the frame as shown by Theorem \[Theorem\_Equal\_Entropy\].
### Fixed volume $V$ of the system $\Sigma$
Let us assume that instead of keeping $N$ fixed, we keep the volume of the system fixed. Then following the procedure given in Sect. \[marker\_thermodynamic\_potentials\], we find find that the correct thermodynamic potential now in the two frames $\mathcal{L}$ and $\mathcal{C}$ are given by$$\begin{aligned}
\Omega(t) & \equiv E(t)-T_{0}S(t)-\mu_{0}N(t),\\
\Omega^{\text{i}}(t) & \equiv E^{\text{i}}(t)-T_{0}S(t)-\mu_{0}N(t),\end{aligned}$$ respectively.
### Fixed $N$ and $V$ of the system $\Sigma$
Let us assume that we keep $N$ and $V$ fixed. Then following the procedure given in Sect. \[marker\_thermodynamic\_potentials\], we find find that the correct thermodynamic potential now in the two frames $\mathcal{L}$ and $\mathcal{C}$ are given by the Helmholtz free energy$$\begin{aligned}
F(t) & \equiv E(t)-T_{0}S(t),\\
F^{\text{i}}(t) & \equiv E^{\text{i}}(t)-T_{0}S(t),\end{aligned}$$ respectively.
### Extension to many state variables
From now on, we will list energy, volume and particle number for any body separately and use $\mathbf{X}$ and $\mathbf{Z}$ to denote the rest of the observables and state variables, respectively. We will only fix the number of particles $N,\widetilde{N}$ but allow all other state variables to fluctuate. In this case, $\widetilde{S}(\widetilde{E}^{\text{i}},\widetilde{V},\widetilde{N})$ in Eq. (\[Total\_Entropy\_0\]) must be written as $\widetilde{S}(\widetilde{E}^{\text{i}},\widetilde{V},\widetilde{N},\widetilde{\mathbf{Z}})$ and its expansion in terms of small quantities gives$$\widetilde{S}(\widetilde{E}^{\text{i}},\widetilde{V},\widetilde{N},\widetilde{\mathbf{Z}})\simeq\widetilde{S}(E_{0},V_{0},\widetilde
{N},\mathbf{Z}_{0})-\left. \left( \frac{\partial\widetilde{S}}{\partial\widetilde{E}}\right) \right\vert _{0}E(t)-\left. \left(
\frac{\partial\widetilde{S}}{\partial\widetilde{V}}\right) \right\vert
_{0}V(t)-\left. \left( \frac{\partial\widetilde{S}}{\partial\widetilde
{\mathbf{Z}}}\right) \right\vert _{0}\cdot\mathbf{Z}(t).$$ Here, $\left. {}\right\vert _{0}$ corresponds evaluating the derivative at $E_{0},V_{0},N_{0},\mathbf{Z}_{0}$ ($\mathbf{X}_{0}$ and $\mathbf{I}_{0})$ so that these derivatives are *constant*, independent of the properties of the system; see the discussion in deriving Eq. (\[Medium\_Fields\]). Introducing the corresponding “chemical potential vector” $\boldsymbol{\mu
}_{0}$ for $\mathbf{X}_{0}$ and the “affinity vector” $\mathbf{a}_{0}\equiv\mathbf{A}_{0}/T_{0}=0$ (see Sect. \[Marker\_Internal Variables\]) for $\mathbf{I}_{0}$ because of the internal equilibrium of $\widetilde{\Sigma}$ $$\left. \left( \frac{\partial\widetilde{S}}{\partial\widetilde{\mathbf{X}}}\right) \right\vert _{0}=-\frac{\boldsymbol{\mu}_{0}}{T_{0}},\left. \left(
\frac{\partial\widetilde{S}}{\partial\widetilde{\mathbf{I}}}\right)
\right\vert _{0}=\frac{\mathbf{A}_{0}}{T_{0}}=0,
\label{Chemical_Potential_Internal_Variables}$$ we can identify a new thermodynamic potential $G^{\mathbf{X}}(t)\equiv
-T_{0}[S_{0}(E_{0},V_{0},N_{0},\mathbf{Z}_{0},t)-\widetilde{S}(E_{0},V_{0},\widetilde{N},\mathbf{Z}_{0})]:$$$G^{\mathbf{X}}(t)=E(t)-T_{0}S(t)+P_{0}V(t)+\boldsymbol{\mu}_{0}\cdot
\mathbf{X}(t)=G(t)+\boldsymbol{\mu}_{0}\cdot\mathbf{X}(t)
\label{General_Thermodynamic_Potential}$$ in the lab frame $\mathcal{L}$. As $E_{0},V_{0},N_{0},\mathbf{X}_{0}$ and $\widetilde{N}$ are constant, we have$$\frac{d}{dt}\widetilde{S}(E_{0},V_{0},\widetilde{N},\mathbf{Z}_{0})=\left.
\left( \frac{\partial\widetilde{S}}{\partial\mathbf{I}_{0}}\right)
\right\vert _{0}\cdot\frac{d\mathbf{I}_{0}(t)}{dt}=0$$ because $\mathbf{a}_{0}=0$. Thus, $\widetilde{S}(E_{0},V_{0},\widetilde
{N},\mathbf{Z}_{0})$ is a constant, and the second law tells us that$$\frac{dS_{0}}{dt}=-\frac{1}{T_{0}}\frac{dG^{\mathbf{X}}}{dt}\geq0,
\label{General_Thermodynamic_Potential_Variation0}$$ as expected in any spontaneous process. In the $\mathcal{C}$ frame, we will instead have$$G^{\text{i}\mathbf{X}}(t)=E^{\text{i}}(t)-T_{0}S(t)+P_{0}V(t)+\boldsymbol{\mu
}_{0}\cdot\mathbf{X}(t)=G^{\text{i}}(t)+\boldsymbol{\mu}_{0}\cdot
\mathbf{X}(t). \label{General_Thermodynamic_Potential_i}$$ The important point to note is that chemical potential vector $\boldsymbol{\mu
}_{0}$ and the affinity vector $\mathbf{A}_{0}=0$ are associated with the medium, just as $T_{0},P_{0}$ are. The analogue of the thermodynamic potential $\Omega(t)$ and $F(t)$ are$$\begin{aligned}
\Omega^{\mathbf{X}}(t) & =\Omega(t)+\boldsymbol{\mu}_{0}\cdot\mathbf{X}(t),\Omega^{\text{i}\mathbf{X}}(t)=\Omega^{\text{i}}(t)+\boldsymbol{\mu}_{0}\cdot\mathbf{X}(t),\\
F^{\mathbf{X}}(t) & =F(t)+\boldsymbol{\mu}_{0}\cdot\mathbf{X}(t),F^{\text{i}\mathbf{X}}(t)=F^{\text{i}}(t)+\boldsymbol{\mu}_{0}\cdot\mathbf{X}(t),\end{aligned}$$ respectively. Again, it follows from the second law that$$\frac{d\Omega^{\mathbf{X}}}{dt}\leq0,\frac{dF^{\mathbf{X}}}{dt}\leq0,
\label{General_Thermodynamic_Potential_Variation}$$ as expected in any spontaneous process.
Gibbs Fundamental Relation for a System under Internal Equilibrium\[marker\_homogeneous\_internal\]
---------------------------------------------------------------------------------------------------
### No extra state variables $\mathbf{Z}(t)$
We will first assume that there are no internal variables, but the system $\Sigma$ satisfies the condition of internal equilibrium so that $S(E^{\text{i}},V,N)$ no longer depends explicitly on time. Then, we can identify the temperature, pressure, and the chemical potential of the system by
\[CM\_Derivatives\]$$\begin{aligned}
\left( \frac{\partial S}{\partial E^{\text{i}}}\right) & =\frac{1}{T(t)},\ \label{CM_Derivatives-1}\\
\left( \frac{\partial S}{\partial V}\right) & =\frac{P(t)}{T(t)},\ \label{CM_Derivatives-2}\\
\left( \frac{\partial S}{\partial N}\right) & =-\frac{\mu(t)}{T(t)}.
\label{CM_Derivatives-3}$$ This allows us to write down the Gibbs fundamental relation for constant $N$ as
$$T(t)dS=dE^{\text{i}}(t)+P(t)dV(t)-+\mu(t)dN(t),$$ which can be rearranged to write down the first law of thermodynamics$$dE^{\text{i}}(t)=T(t)dS-P(t)dV(t)+\mu(t)dN(t). \label{First_Law_int}$$ We now turn to Eq. (\[Internal\_Energy\_0\]). In terms of the momentum of the center of mass frame $\mathcal{C}$ of the system, we have$$dE^{\text{i}}=dE-\mathbf{V}\cdot d\mathbf{P}; \label{Diff_Energy_Relation}$$ recall that according to our assumption, the system has no intrinsic angular momentum. This allows us to use Eq. (\[First\_Law\_int\]) to write down the differential form$$dE=T(t)dS+\mathbf{V}\cdot d\mathbf{P}-P(t)dV(t)+\mu(t)dN(t)
\label{First-Law_Total}$$ for the first law of thermodynamics in terms of the energy $E(t)$ rather than the internal energy $E^{\text{i}}(t)$. This has some important consequences and will be extremely useful in the following. The first consequence is that it allows us to think of $S(E^{\text{i}},V,N)$ as a function of four variables $S(E,\mathbf{P},V,N)$:$$T(t)dS=dE-\mathbf{V}\cdot d\mathbf{P}+P(t)dV(t)-\mu(t)dN(t).$$ This equation is the Gibbs fundamental relation relating the entropy with $E(t)$ rather than $E^{\text{i}}(t)$. The second consequence, which follows from the Gibbs fundamental relation is that
\[Lab\_Derivatives\]$$\begin{aligned}
\left( \frac{\partial S}{\partial E}\right) & =\frac{1}{T(t)},\label{Lab_Derivatives-1}\\
\left( \frac{\partial S}{\partial\mathbf{P}}\right) & =-\frac
{\mathbf{V}(t)}{T(t)},\ \label{Lab_Derivatives-2}\\
\left( \frac{\partial S}{\partial V}\right) & =\frac{P(t)}{T(t)},\ \label{Lab_Derivatives-3}\\
\left( \frac{\partial S}{\partial N}\right) & =-\frac{\mu(t)}{T(t)}.
\label{Lab_Derivatives-4}$$ The *drift velocity* $\mathbf{V}$ (of the center of mass) of the system is given a thermodynamic interpretation in terms of the derivative in Eq. (\[Lab\_Derivatives-2\]) at fixed $E,V$ and $N$. For the case when the number of particles is held fixed, $dN(t)\equiv0$ and the last term in $dE^{\text{i}}(t),dE(t)$ and $dS(t)$ will be absent. For fixed $V$, the third term in $dE^{\text{i}}(t),dE(t)$ and $dS(t)$ will be absent, and so on.
### Inclusion of state variables $\mathbf{Z}(t)$
In the presence of internal variables $\mathbf{X}$ and $\mathbf{I}$, the extension of the above relations is quite obvious. Introducing the instantaneous chemical potential vector $\boldsymbol{\mu}(t)$ associated with $\mathbf{X}$ and the affinity vector $\mathbf{A}$ associated with $\mathbf{I}$ using
$$\left( \frac{\partial S}{\partial\mathbf{X}}\right) =-\frac{\boldsymbol{\mu
}(t)}{T(t)},\left( \frac{\partial S}{\partial\mathbf{I}}\right)
=\frac{\mathbf{A}(t)}{T(t)}, \label{Chemical_Potential_Affinity}$$
we generalize the entropy differential to
\[Gibbs\_Fundamental\_Relation\_System\]$$\begin{aligned}
T(t)dS & =dE^{\text{i}}(t)+P(t)dV(t)-\mu(t)dN(t)-\boldsymbol{\mu}(t)\cdot
d\mathbf{X}(t)+\mathbf{A}(t)\cdot d\mathbf{I}(t)\label{Gibbs_Fundamental_Relation_System_Internal}\\
& =dE(t)-\mathbf{V}(t)\cdot d\mathbf{P}(t)+P(t)dV(t)-\mu
(t)dN(t)-\boldsymbol{\mu}(t)\cdot\mathbf{X}(t)+\mathbf{A}(t)\cdot
d\mathbf{I}(t). \label{Gibbs_Fundamental_Relation_System_General}$$
Inhomogeneous System with Relative Motion
=========================================
Subsystems undergoing Relative Motion
-------------------------------------
We now consider the isolated system to be stationary so that it has no linear and angular momenta. However, $\Sigma$ and $\widetilde{\Sigma}$ may have relative motion so that they each may possess linear and angular momenta that individually must cancel out:
$$\mathbf{P}+\widetilde{\mathbf{P}}=0,\ \mathbf{M}+\widetilde{\mathbf{M}}=0.
\label{Net_Momenta}$$
Moreover, we will now treat $\Sigma$ as inhomogeneous and assume that it can be decomposed into a *collection* of a large number $N_{\text{S}}$ of *subsystems* $\sigma_{k}$, $k=1,2,\cdots,N_{\text{S}},$ which may be in different macrostates to allow for inhomogeneity and for relative motion between different subsystems and within each subsystem. Each subsystem is still *macroscopically* large so that we can not only introduce a legitimate entropy function $s_{k}$ for the macrostate $\mathcal{M}_{k}$ via Gibbs’s formulation$$s_{k}(e_{k}^{\text{i}},\boldsymbol{\Omega}_{k},n_{k},v_{k},\mathbf{z}_{k},t)\equiv-{\textstyle\sum\limits_{\alpha_{k}}}
p_{\alpha_{k}}(t)\ln p_{\alpha_{k}}$$ where $\alpha_{k}$ denotes one of the allowed microstates of the subsystem $\sigma_{k}$ corresponding to the macrostate $\mathcal{M}_{k}$ characterized by observables $e_{k}^{\text{i}},n_{k},v_{k}$ and $\mathbf{z}_{k}$, but we also have these entropies satisfy the *additive property*$$S(E^{\text{i}},\boldsymbol{\Omega},N,V,\mathbf{Z},t)={\textstyle\sum\limits_{k=1}^{N_{\text{S}}}}
s_{k}(e_{k}^{\text{i}},\boldsymbol{\Omega}_{k},n_{k},v_{k},\mathbf{z}_{k},t),
\label{Entropy_Sum_0}$$ which requires their *quasi-independence* as discussed in Sect. \[marker\_entropy\_additivity\] at each instant $t$. Using the entropy $s_{k},$ we can introduce the appropriate thermodynamic functions, but care must be exercised in identifying these functions in the lab frame $\mathcal{L}$, where the experiments are done. This is because the energies depend on the frame of reference, which will result in different values of the energies and thermodynamic potentials in different frames, such as the lab frame $\mathcal{L}$ and the rotating frame frame $\mathcal{C}_{k}$ attached to the center of mass of $\sigma_{k}$, which is translating with a linear velocity $\mathbf{v}_{k}(t)$ and rotating with an angular velocity $\Omega
_{k}(t)$.
To make further progress, we will make another assumption later as we did in I that each subsystem is in *internal equilibrium.* This occurs when all microstates contributing to the entropy are *equiprobable*$$p_{\alpha_{k}}(t)=1/W_{k},\ \ \alpha=1,2,,\cdots W_{k}(t);$$ here $W_{k}(t)$ represents the number of microstates of the subsystem $\sigma_{k}$ at time $t$. Under the equiprobability assumption or the internal equilibrium assumption, $$s_{k}(t)=\ln W_{k}(t),$$ which is what one would obtain by applying the Boltzmann formulation of the entropy [@Gujrati-Symmetry]. It also follows from Theorem \[marker\_Uniform\_Motion\] that the entire subsystem is uniformly translating with a linear velocity $\mathbf{v}_{k}(t)$ and rotating with an angular velocity $\Omega_{k}(t)$ so that the system in internal equilibrium is *stationary* in the frame $\mathcal{C}_{k}$.
System under Arbitrary Conditions\[marker\_inhomogeneous\_arbitrary\]
---------------------------------------------------------------------
The internal energy for each subsystem $\sigma_{k}$ is related to its energy $e_{k}$ in the lab frame $\mathcal{L}$ $$e_{k}^{\text{i}}\equiv e_{k}-\mathbf{p}_{k}^{2}/2m_{k}-\mathbf{m}_{k}\cdot\boldsymbol{\Omega}_{k}; \label{subsystem_energy_relation0}$$ see Eq. (\[Internal\_Energy\_0\]). Alternatively, we can use Eq. (\[Internal\_Energy\_2\]) to express this relation as $$e_{k}^{\text{i}}=e_{k}-\mathbf{p}_{k}^{2}/2m_{k}-{\displaystyle\sum_{j_{k}}}
m_{j_{k}}\mathbf{r}_{j_{k}}\cdot\left( \mathbf{v}_{j_{k}}\times
\mathbf{\Omega}_{k}\right) -\frac12
{\displaystyle\sum_{j_{k}}}
m_{j_{k}}(\mathbf{\Omega}_{k}\mathbf{\times r}_{j_{k}})^{2},
\label{subsystem_energy_relation1}$$ where each sum is over $n_{k}$ particles in the subsystem $k$. The third term in the last equation vanishes when the subsystem is in internal equilibrium as commented earlier. We also have the additivity laws
\[Conservation\_Laws\]$$\begin{aligned}
V & ={\displaystyle\sum}
v_{k},N={\displaystyle\sum}
n_{k},\mathbf{X}={\displaystyle\sum}
\mathbf{x}_{k},\mathbf{I}={\displaystyle\sum}
\mathbf{i}_{k},\label{Conservation_V}\\
E & ={\displaystyle\sum}
e_{k}={\displaystyle\sum}
(e_{k}^{\text{i}}+\mathbf{p}_{k}^{2}/2m_{k}+\mathbf{m}_{k}\cdot\Omega
_{k}),\label{Conservation_E}\\
\mathbf{P} & ={\displaystyle\sum}
\mathbf{p}_{k},\mathbf{M}={\displaystyle\sum}
(\mathbf{m}_{k}+\mathbf{r}_{k}\times\mathbf{p}_{k}), \label{Conservation_P}$$ at each instant $t$. The angular momentum $\mathbf{l}_{k}\equiv$ $\mathbf{r}_{k}\times\mathbf{p}_{k}$ is the orbital angular momentum of $\sigma_{k}$ with $\mathbf{r}_{k},$ $\mathbf{p}_{k}$ representing the location and momentum of the center of mass of $\sigma_{k}$, respectively, and should not be confused with its intrinsic angular momentum $\mathbf{m}_{k}$ introduced in Eq. (\[System\_Definition\]). We should emphasize that the additivity of the energy requires that the interaction energy between subsystems be negligible. This condition is necessary for the entropy to be additive as discussed in Sect. \[marker\_entropy\_additivity\]. As a consequence, we do not have the contribution analogous to $\psi(\mathbf{r})dV$ in Eq. (\[Energy\_Additivity\_Thermodynamics\]). This distinguishes our approach with that taken in local non-equilibrium thermodynamics [@Donder; @deGroot; @Prigogine; @Prigogine-old].
We keep $n_{k}$ fixed for simplicity so that $N$ is also fixed, and allow $e_{k}^{\text{i}},v_{k}$ and $\mathbf{z}_{k}$ to vary in time. We assume as above that $\widetilde{\Sigma}$ has well-defined field variables ($T_{0},P_{0}$, etc.) which do not change with time. This is ensured by assuming internal equilibrium for the medium. This is the only assumption we make regarding the isolated system $\Sigma_{0}$ for which we have
$$E_{0}=\widetilde{E}+E,V_{0}=\widetilde{V}+V,N_{0}=\widetilde{N}+N,\mathbf{X}_{0}=\widetilde{\mathbf{X}}+\mathbf{X,I}_{0}=\widetilde{\mathbf{I}}+\mathbf{I}.$$ Because the medium is taken to be in internal equilibrium, its energy is related to its internal energy according to Eq. (\[Internal\_Energy\_0\])$$\widetilde{E}^{\text{i}}\equiv\widetilde{E}-\mathbf{P}^{2}/2\widetilde
{M}-\widetilde{\mathbf{M}}\cdot\widetilde{\boldsymbol{\Omega}}=E-\mathbf{P}^{2}/2M-\mathbf{M}^{2}/2\widetilde{I}, \label{General_Ei_E_Relation_Medium}$$ where according to Eq. (\[Net\_Momenta\]), we have taken $\widetilde
{\mathbf{M}}=-\mathbf{M}$, and where $\widetilde{I}$ is the moment of inertia of the medium about its axis of rotation. The axis of rotation must be one of its principal axis of rotation; see the comment after the proof of Theorem \[marker\_Uniform\_Motion\] in Sect. \[marker\_internal\_equilibrium\]. The contribution coming from the internal motion, which is similar to the third contribution in Eq. (\[subsystem\_energy\_relation1\]) applied to the medium vanishes because of its uniform translation and rotation following Theorem \[marker\_Uniform\_Motion\] applied to the medium. The contribution similar to the last term in Eq. (\[subsystem\_energy\_relation1\]) is the standard rotational kinetic energy of the medium treated as a rigid body. The angular momentum $\widetilde{\mathbf{M}}$ is given by$$\widetilde{M}_{ij}=\widetilde{I}_{ij}\widetilde{\Omega}_{j}=-M_{ij}.$$ Assuming that the motions are finite, we conclude that $\widetilde{M}_{ij}$ must be finite. Therefore, for an extremely large medium, $\widetilde
{\mathbf{\Omega}}$ and $\widetilde{\mathbf{V}}$ must be extremely small, which ensure that the last two terms in Eq. (\[General\_Ei\_E\_Relation\_Medium\]) are extremely small. This allows us to approximate$$\widetilde{E}\simeq\widetilde{E}^{\text{i}}
\label{Medium_Energy_Internal_Energy}$$ without any appreciable error. As far as the system is concerned, the relationship between its internal energy and the energy is still given by Eq. (\[Internal\_Energy\]) $$E^{\text{i}}\equiv E-\mathbf{P}^{2}/2M-\mathbf{M}\cdot\boldsymbol{\Omega,}
\label{General_Ei_E_Relation}$$ except that the system may not be uniformly translating and rotating about any of its principal axis of inertia. We take $\widetilde{N}$ and $N$ as constants, but allow for $E,V,$ and $\mathbf{Z}$ ($\mathbf{X}$ and $\mathbf{I}$) to change with time. Now, we use the entropy additivity $$S_{0}(E_{0}^{\text{i}},N_{0},V_{0},\mathbf{Z}_{0},t)=S(E^{\text{i}},\boldsymbol{\Omega,}N,V,\mathbf{Z},t)+\widetilde{S}(\widetilde{E}^{\text{i}},\widetilde{\boldsymbol{\Omega}}\boldsymbol{,}\widetilde{N},\widetilde
{V},\widetilde{\mathbf{Z}})$$ and expand $$\begin{aligned}
\widetilde{S}(\widetilde{E}^{\text{i}},\widetilde{\boldsymbol{\Omega}}\boldsymbol{,}\widetilde{N},\widetilde{V},\widetilde{\mathbf{Z}}) &
=\widetilde{S}(E_{0},\widetilde{\boldsymbol{\Omega}}\boldsymbol{,}\widetilde{N},V_{0},\mathbf{Z}_{0})-E\left. \frac{\partial\widetilde{S}}{\partial\widetilde{E}}\right\vert _{0}-V\left. \frac{\partial\widetilde{S}}{\partial\widetilde{V}}\right\vert _{0}-\mathbf{Z\cdot}\left. \frac
{\partial\widetilde{S}}{\partial\widetilde{\mathbf{Z}}}\right\vert _{0}\\
& =\widetilde{S}(E_{0},\widetilde{\boldsymbol{\Omega}}\boldsymbol{,}\widetilde{N},V_{0},\mathbf{Z}_{0})-\frac{E}{T_{0}}-V\frac{P_{0}}{T_{0}}-\mathbf{X\cdot}\frac{\boldsymbol{\mu}_{0}}{T_{0}},\end{aligned}$$ by treating $E,V,\mathbf{X}$ and $\mathbf{I}$ as small parameters; here, $\left. {}\right\vert _{0}$ corresponds to evaluating the derivatives at $E_{0},\widetilde{\boldsymbol{\Omega}}\boldsymbol{,}V_{0},N_{0},\mathbf{Z}_{0}$ ($\mathbf{X}_{0}$ and $\mathbf{I}_{0}$). We have also used the definitions of the field variables of the medium in terms of the derivatives of the entropy $\widetilde{S}$ in Eqs. (\[Medium\_Fields\]) and (\[Chemical\_Potential\_Internal\_Variables\]), and set $\mathbf{A}_{0}=0$ as established in Sect. \[Marker\_Internal Variables\]. We thus finally obtain$$\begin{aligned}
S_{0}(E_{0},N_{0},V_{0},\mathbf{Z}_{0},t)-\widetilde{S}(E_{0}\widetilde
{\boldsymbol{\Omega}}\boldsymbol{,}\widetilde{N},V_{0},\mathbf{Z}_{0}) &
=S-\frac{E}{T_{0}}-V\frac{P_{0}}{T_{0}}=-\frac{G}{T_{0}}-\mathbf{X\cdot}\frac{\boldsymbol{\mu}_{0}}{T_{0}}\\
& =-\frac{G^{\text{i}}}{T_{0}}-\frac{\mathbf{P}^{2}}{2MT_{0}}-\frac
{\mathbf{M}\cdot\mathbf{\Omega}}{T_{0}}-\frac{\mathbf{X\cdot}\boldsymbol{\mu
}_{0}}{T_{0}},\end{aligned}$$ where $$G\equiv E-T_{0}S+P_{0}V,\ \ G^{\text{i}}\equiv E^{\text{i}}-T_{0}S+P_{0}V.$$ Observe that the expansion of $\widetilde{S}(\widetilde{E}^{\text{i}},\widetilde{\boldsymbol{\Omega}}\boldsymbol{,}\widetilde{N},\widetilde
{V},\widetilde{\mathbf{Z}})$ does not require any knowledge of what is happening with the system $\Sigma$. For example, we have *not* assumed internal equilibrium in the system or any of its subsystems. Since $\widetilde{S}(E_{0},\widetilde{N},V_{0},\mathbf{Z}_{0})$ is a constant as established earlier, the second law tells us that $$\frac{dS_{0}}{dt}=-\frac{1}{T_{0}}\frac{dG^{\mathbf{X}}}{dt}\geq0,
\label{Gibbs_Free_Energy_Variation_0}$$ the desired result for the system. The inhomogeneity of the system has no relevance in the above conclusion.
Let us now assume that $\mathbf{X}=0$ and introduce$$g_{k}\equiv e_{k}-T_{0}s_{k}+P_{0}v_{k},\ \ g_{k}^{\text{i}}\equiv
e_{k}^{\text{i}}-T_{0}s_{k}+P_{0}v_{k} \label{Inhomogeneous_Gibbs_Free_Energy}$$ for the subsystem $\sigma_{k}$. It is clear that$$G={\displaystyle\sum\limits_{k}}
g_{k},G^{\text{i}}={\displaystyle\sum\limits_{k}}
g_{k}^{\text{i}}.$$ Because of the quasi-independence of various subsystems, Eq. (\[Gibbs\_Free\_Energy\_Variation\_0\]) immediately leads to $$\frac{dg_{k}}{dt}\leq0. \label{subsystem_Gibbs_Free_Energy_Variation}$$ Thus, $g_{k}(t)$ can be identified as the Gibbs free energy of the subsystem $\sigma_{k}$ in the lab frame $\mathcal{L}$. Comparing this definition with the definition in Eq, (\[Local\_Gibbs\_Free\_Energy\]) used in the conventional non-equilibrium thermodynamics, we see that the discrepancy in the two Gibbs free energy has not disappeared by taking into account the inhomogeneity inherent in the system.
It may be argued that the above identification of $g_{k}$ in Eq. (\[Inhomogeneous\_Gibbs\_Free\_Energy\]) is based on considering the entire system in contact with the medium. One can alternatively consider a particular subsystem $\sigma$ of the system in contact with the medium and the remaining subsystems. However, a little bit of reflection shows that this will not affect the behavior of $g_{k}$ for the simple reason that the remaining subsystems still form a very small part of the isolated system so that they *cannot* affect the internal equilibrium of the medium. To see this more clearly, let us introduce a new medium $\widetilde{\Sigma}^{\prime}$ consisting of $\widetilde{\Sigma}$ and the remaining subsystems. Only for the sake of simplicity, we do not consider any additional state variables. The argument can be easily extended to include them. Let $\widetilde{S}^{\prime}$ denote the entropy and $\widetilde{E}^{\text{i}\prime}\simeq\widetilde
{E}^{\prime},$ $\widetilde{V}^{\prime}$ and $\widetilde{N}^{\prime}$ the internal energy, volume and the number of particles of $\widetilde{\Sigma
}^{\prime}$, the latter of which is kept fixed. Expanding this entropy in terms of the small quantities $e^{\text{i}}$ and $v$ of the chosen subsystem requires calculating the derivatives$$\left. \frac{\partial\widetilde{S}^{\prime}}{\partial\widetilde{E}^{\prime}}\right\vert _{E_{0},V_{0}}\text{ and }\left. \frac{\partial\widetilde
{S}^{\prime}}{\partial\widetilde{V}^{\prime}}\right\vert _{E_{0},V_{0}}.$$ Because of the small size of the system relative to the original medium $\widetilde{\Sigma}$, these derivatives are not different from$$\left. \frac{\partial\widetilde{S}}{\partial\widetilde{E}}\right\vert
_{E_{0},V_{0}}\text{ and }\left. \frac{\partial\widetilde{S}}{\partial
\widetilde{V}}\right\vert _{E_{0},V_{0}},$$ respectively. Thus, using $e=E_{0}-\widetilde{E}$ for the energy of the subsystem $\sigma_{k}$, we find that $$\widetilde{S}^{\prime}(\widetilde{E}^{\text{i}\prime},\widetilde
{\boldsymbol{\Omega}}\boldsymbol{,}\widetilde{N}^{\prime},\widetilde
{V}^{\prime})=\widetilde{S}(E_{0},\widetilde{\boldsymbol{\Omega}}\boldsymbol{,}\widetilde{N},V_{0})-e\left. \frac{\partial\widetilde
{S}^{\prime}}{\partial\widetilde{E}^{\prime}}\right\vert _{E_{0},V_{0}}-v\left. \frac{\partial\widetilde{S}^{\prime}}{\partial\widetilde{V}^{\prime}}\right\vert _{E_{0},V_{0}}=\widetilde{S}(E_{0},\widetilde{N},V_{0})-\frac{e}{T_{0}}-v\frac{P_{0}}{T_{0}}.$$ Using this in $S_{0}(E_{0}^{\text{i}},N_{0},V_{0},t)=s(e^{\text{i}},\boldsymbol{\Omega,}n,v,t)+\widetilde{S}^{\prime}(\widetilde{E}^{\text{i}\prime},\widetilde{\boldsymbol{\Omega}}\boldsymbol{,}\widetilde
{N}^{\prime},\widetilde{V}^{\prime})$, where $\boldsymbol{\Omega}$ is the angular velocity of the $\mathcal{C}$ frame of the selected subsystem, allows us to identify $$g\equiv e-T_{0}s+P_{0}v \label{Subsystem_Gibbs_free_energy}$$ as the Gibbs free energy of the particular subsystem $\sigma$. It now follows from Eq. (\[Second\_Law\_0\]) that$$\frac{dg}{dt}\leq0,$$ a property we expect from the Gibbs free energy of a system. Incidentally, this also provides an independent justification of the inequality in Eq. (\[subsystem\_Gibbs\_Free\_Energy\_Variation\]).
In terms of $g_{k}^{\text{i}}$, we immediately have$$G^{\text{i}}={\displaystyle\sum\limits_{k}}
(g_{k}^{\text{i}}+\frac{\mathbf{p}_{k}^{2}}{2m_{k}}+\mathbf{m}_{k}\cdot
\Omega_{k}),$$ which is expected in view of the sum rule in Eqs. (\[Conservation\_E\]) and (\[General\_Ei\_E\_Relation\]).
System under Internal Equilibrium\[marker\_inhomogeneous\_internal\]
--------------------------------------------------------------------
The above derivation only uses the second law, and the assumption that the medium satisfies the condition of internal equilibrium. The situation with the first law is very different. In general, the differential $ds_{k}$ of the entropy of the subsystem $\sigma_{k}$ is given by $$ds_{k}=\frac{\partial s_{k}}{\partial e_{k}}de_{k}^{\text{i}}+\frac{\partial
s_{k}}{\partial v_{k}}dv_{k}+\frac{\partial s_{k}}{\partial n_{k}}dn_{k}+\frac{\partial s_{k}}{\partial\boldsymbol{\Omega}_{k}}\cdot
d\boldsymbol{\Omega}_{k}+\frac{\partial s_{k}}{\partial\mathbf{z}_{k}}\cdot
d\mathbf{z}_{k}(t)+\frac{\partial s_{k}}{\partial t}dt;$$ cf. Eqs. (\[Equal\_Entropy\]) and (\[Functional\_dependence\]). If we now assume internal equilibrium so that $s_{k}$ does not have an explicit $t$-dependence, we can get rid of the last term above. In this case only, the remaining derivatives identify the field variables $1/T_{k},P_{k}/T_{k},$ $-\mu_{k}/T_{k}$ etc. of the subsystem $\sigma_{k}:$$$\frac{\partial s_{k}}{\partial e_{k}}=\frac{1}{T_{k}(t)},\frac{\partial s_{k}}{\partial v_{k}}=\frac{P_{k}(t)}{T_{k}(t)},\frac{\partial s_{k}}{\partial
n_{k}}=-\frac{\mu_{k}(t)}{T_{k}(t)}dn_{k},\frac{\partial s_{k}}{\partial
\boldsymbol{\Omega}_{k}}=\frac{\mathbf{m}_{k}(t)}{T_{k}(t)},\frac{\partial
s_{k}}{\partial\mathbf{x}_{k}}=-\frac{\boldsymbol{\mu}^{(k)}(t)}{T_{k}(t)},\frac{\partial s_{k}}{\partial\mathbf{i}_{k}}=\frac{\mathbf{A}^{(k)}(t)}{T_{k}(t)}. \label{subsystem_parameters0}$$ These derivatives then allows us to obtain the Gibbs fundamental relation$$de_{k}^{\text{i}}(t)=T_{k}(t)ds_{k}-P_{k}(t)dv_{k}+\mu_{k}(t)dn_{k}-\mathbf{m}_{k}(t)\cdot d\boldsymbol{\Omega}_{k}+\boldsymbol{\mu}^{(k)}(t)\cdot d\mathbf{x}_{k}(t)-\mathbf{A}^{(k)}(t)\cdot d\mathbf{i}_{k}(t).
\label{subsystem_First_Law_internal}$$ Using$$de_{k}(t)=de_{k}^{\text{i}}(t)+\mathbf{v}_{k}(t)\cdot d\mathbf{p}_{k}(t)+d[\mathbf{\Omega}_{k}(t)\cdot\mathbf{m}_{k}(t)],$$ we find the first law of thermodynamics can be expressed in terms of the energy as$$\begin{aligned}
de_{k} & =T_{k}(t)ds_{k}(t)+\mathbf{v}_{k}(t)\cdot d\mathbf{p}_{k}(t)+\mathbf{\Omega}_{k}(t)\cdot d\mathbf{m}_{k}(t)-P_{k}(t)dv_{k}(t)\label{First_Law_General}\\
& +\mu_{k}(t)dn_{k}(t)+\boldsymbol{\mu}^{(k)}(t)\cdot d\mathbf{x}_{k}(t)-\mathbf{A}^{(k)}(t)\cdot d\mathbf{i}_{k}(t);\nonumber\end{aligned}$$ compare with Eq. (\[EL\_differential\]). This allows us to think of the entropy $s_{k}((e_{k}^{\text{i}},\boldsymbol{\Omega}_{k},n_{k},v_{k},\mathbf{z}_{k},t)$ as a function$$s_{k}(e_{k},\mathbf{p}_{k},\mathbf{m}_{k},n_{k},v_{k},\mathbf{z}_{k},t)$$ so that the drift and the angular velocities in internal equilibrium are given by$$\frac{\mathbf{v}_{k}(t)}{T_{k}(t)}=-\frac{\partial s_{k}(t)}{\partial
\mathbf{p}_{k}(t)},\ \ \frac{\mathbf{\Omega}_{k}(t)}{T_{k}(t)}=-\frac{\partial
s_{k}(t)}{\partial\mathbf{m}_{k}(t)}. \label{subsystem_parameters1}$$ However, different subsystems will undergo relative motions with respect to each other as $\mathbf{v}_{k}(t)$ and $\mathbf{\Omega}_{k}(t)$ are different for them. In addition, their temperatures $T_{k}(t)$ and pressures $P_{k}(t)$ are also different for each other, and so are $\boldsymbol{\mu}^{(k)}(t)\ $and $\mathbf{A}^{(k)}(t)$. Thus, there would be viscous dissipation and, consequently, entropy production as they come to equilibrium with each other and with the medium. We now turn to this issue.
Reversible and Irreversible Contributions
=========================================
General Considerations: Bodies in a Medium
------------------------------------------
It is a well-known fact [@Landau Sect. 10] that in equilibrium, the system $\Sigma$ has a uniform translational motion as a whole with a constant velocity $\mathbf{V}_{0}$ and a uniform rotation of the whole system with a constant angular velocity $\boldsymbol{\Omega}_{0}$. Thus, there cannot be relative motions between different subsystems in equilibrium. Thus, no internal macroscopic motion is possible in equilibrium. In equilibrium, the coefficients of the differential quantities in Eq. (\[First\_Law\_General\]) take their constant values of the medium. Writing $$ds_{k}(t)\equiv d_{\text{e}}s_{k}(t)+d_{\text{i}}s_{k}(t)
\label{entropy_generation_0}$$ as a sum of the change in the entropy $d_{\text{e}}s_{k}(t)$ due to reversible exchange with the medium and the production of the entropy $d_{\text{i}}s_{k}(t)\geq0$ due to irreversible processes within the system, we have $$d_{\text{e}}s_{k}=\frac{1}{T_{0}}[de_{k}(t)-\mathbf{V}_{0}(t)\cdot
d\mathbf{p}_{k}(t)-\mathbf{\Omega}_{0}\cdot d\mathbf{m}_{k}(t)+P_{0}dv_{k}(t)-\mu_{0}dn_{k}(t)-\boldsymbol{\mu}_{0}\cdot d\mathbf{x}_{k}(t)]$$ and $$\begin{aligned}
d_{\text{i}}s_{k} & =F\left[ \frac{1}{T_{k}(t)}\right] ds_{k}(t)+F\left[
-\frac{\mathbf{v}_{k}(t)}{T_{k}(t)}\right] \cdot d\mathbf{p}_{k}(t)+F\left[
-\frac{\mathbf{\Omega}_{k}(t)}{T_{k}(t)}\right] \cdot d\mathbf{m}_{k}(t)+F\left[ \frac{P_{k}(t)}{T_{k}(t)}\right] dv_{k}(t)\nonumber\\
& +F\left[ -\frac{\mu_{k}(t)}{T_{k}(t)}\right] dn_{k}(t)+F\left[
-\frac{\boldsymbol{\mu}^{(k)}}{T_{k}(t)}\right] \cdot d\mathbf{x}_{k}(t)+F\left[ \frac{\mathbf{A}^{(k)}}{T_{k}(t)}\right] \cdot
d\mathbf{i}_{k}(t)\label{entropy_generation}\\
& \geq0, \label{entropy_generation_limit}$$ where $$F_{z}\equiv F\left[ w\right] =w(t)-w_{\text{eq}} \label{Irreversible_Change}$$ is the difference of the conjugate field $w(t)$ at time $t$ and its value in equilibrium, i.e. as $t\rightarrow\infty$, and represents the thermodynamic force associated with the conjugate extensive quantity $z(t)$; see also Eq. (\[Force\_y\]). For the internal variables, the equilibrium value of $\mathbf{A}_{0}$ is zero. According to the second law of thermodynamics, not only $d_{\text{i}}s_{k}\geq0,$ but each term in Eq. (\[entropy\_generation\]) is non-negative: $$F\left[ w\right] dz\geq0,$$ where $z$ and $w$ form a conjugate pair. In terms of these pairs, we can express the two parts of the entropy as follows:$$ds_{k}=\mathbf{w}_{k}\cdot d\mathbf{z}_{k}\mathbf{,\ \ }d_{\text{e}}s_{k}=\mathbf{w}_{0k}\cdot d\mathbf{z}_{k}\mathbf{,\ \ }d_{\text{i}}s_{k}=\mathbf{F}\left[ \mathbf{w}_{k}\right] \cdot d\mathbf{z}_{k}\mathbf{,}
\label{Generalized_Entropy_Change}$$ which is the general form of the entropy differenmtial and its two components.
Let us now turn back to the current case under investigation. We can express the generalized Gibbs fundamental relation as$$de_{k}\equiv d_{\text{e}}e_{k}(t)+d_{\text{i}}e_{k}(t),$$ where$$d_{\text{e}}e_{k}=T_{0}ds_{k}(t)+\mathbf{V}_{0}(t)\cdot d\mathbf{p}_{k}(t)+\mathbf{\Omega}_{0}\cdot d\mathbf{m}_{k}(t)-P_{0}dv_{k}(t)+\mu
_{0}dn_{k}(t)+\boldsymbol{\mu}_{0}\cdot d\mathbf{x}_{k}(t),$$ and $$\begin{aligned}
d_{\text{i}}e_{k} & =F\left[ T_{k}(t)\right] ds_{k}(t)+F\left[
\mathbf{v}_{k}(t)\right] \cdot d\mathbf{p}_{k}(t)+F\left[ \mathbf{\Omega
}_{k}(t)\right] \cdot d\mathbf{m}_{k}(t)-F\left[ P_{k}(t)\right]
dv_{k}(t)\\
& +F\left[ \mu_{k}(t)\right] dn_{k}(t)+F\left[ \boldsymbol{\mu}^{(k)}\right] \cdot d\mathbf{x}_{k}(t)-F\left[ \mathbf{A}^{(k)}\right]
\cdot d\mathbf{i}_{k}(t).\end{aligned}$$ The generalized form for the Gibbs fundamental relation is evidently $$de_{k}=T_{k}ds_{k}-\mathbf{W}_{k}\cdot d\mathbf{z}^{\prime}\mathbf{,\ \ }d_{\text{e}}e_{k}=T_{0}ds_{k}-\mathbf{W}_{0k}\cdot d\mathbf{z}^{\prime
}\mathbf{,\ \ }d_{\text{i}}e_{k}=F\left[ T_{k}\right] ds_{k}-\mathbf{F}\left[ \mathbf{W}_{k}\right] \cdot d\mathbf{z}^{\prime}\mathbf{,}$$ where $\mathbf{z}^{\prime}$ represents all state variables except the energy $e_{k}$, and where $\mathbf{W}_{k}=T\mathbf{w}$ and $\mathbf{W}_{0}\mathbf{=T}_{0}\mathbf{w}_{0}$.
It is easy to see that performing the Legendre transform to obtain the Gibbs free energy in Eq. (\[Inhomogeneous\_Gibbs\_Free\_Energy\]) only affects the form of $d_{\text{e}}s_{k},$ but leaves $d_{\text{i}}s_{k}$ unaffected. Thus, it is easy to see that $$dg_{k}(t)\equiv d_{\text{e}}g_{k}(t)+d_{\text{i}}g_{k}(t),$$ where $$d_{\text{e}}g_{k}=-s_{k}(t)dT_{0}+\mathbf{V}_{0}(t)\cdot d\mathbf{p}_{k}(t)+\mathbf{\Omega}_{0}\cdot d\mathbf{m}_{k}(t)+v_{k}(t)dP_{0}+\mu
_{0}dn_{k}(t)+\boldsymbol{\mu}_{0}\cdot d\mathbf{x}_{k}(t),$$ and $$d_{\text{i}}g_{k}\equiv d_{\text{i}}e_{k}.$$ The general form of $d_{\text{e}}g_{k}$ is $$d_{\text{e}}g_{k}=-s_{k}(t)dT_{0}+v_{k}(t)dP_{0}-\mathbf{W}_{k}\cdot
d\mathbf{z}^{^{\prime\prime}},$$ where $\mathbf{z}^{^{\prime\prime}}$ represents all state variables except the energy $e_{k}$ and the volume $v_{k}$.
Bodies forming a Finite Isolated System without a Medium\[marker\_comparable\_bodies\]
--------------------------------------------------------------------------------------
So far, we have considered a system or a collection of subsystems in a very large medium $\widetilde{\Sigma}$ so that its field variables are held fixed at $\mathbf{Y}_{0}$. We now consider a collection of bodies of comparable sizes forming an isolated system $\Sigma_{0}$. In this case, we cannot treat any of the bodies as a (macroscopically extensively large) medium with a fixed field at $\mathbf{Y}_{0}$. As the collection strives towards equilibrium, their field variables continue to change in time. This causes a problem in identifying reversible contributions to various quantities. To solve this problem, we discuss below a simple case; the generalization to more complex situation will be obvious.
### Two Bodies
Let us consider the simplest possible case of two comparable bodies $1$ and $2$ in internal equilibrium at temperature $T_{1}$ and $T_{2}>T_{1}$, respectively, that are brought in thermal contact at time $t=0.$ We will simplify the discussion by assuming that all other extensive observables besides the energy are held fixed. The case of two bodies in the shape of rectangular cuboid along the $x$-axis are shown schematically in Fig. \[Fig\_two\_boxes\_isolated\](a), with the rectangular interface of area $A$ lying in the $yz$-plane.
\[ptb\]
[Two\_Boxes\_Isolated\_System.eps]{}
Let $dQ=dE$ be the infinitesimal heat given to the body $1$ by the body $2$. As the amount of heat is infinitesimal, it does not alter the temperatures of the bodies in any significant way so that the entropy of the isolated system $\Sigma_{0}$ changes by$$dS_{0}=d_{\text{i}}S_{0}=dQ\left( \frac{1}{T_{1}}-\frac{1}{T_{2}}\right) >0$$ at $t=0$.; compare with Eq. (\[Isolated\_S\_Generation\]). This expression will not be correct as time goes on as the temperature continues to change. The general expression for the irreversible entropy generation is $$dS_{0}(t)=d_{\text{i}}S_{0}(t)=dQ(t)\left( \frac{1}{T_{1}(t)}-\frac{1}{T_{2}(t)}\right) >0, \label{Irreversible_Entropy_Change_Total}$$ where $dQ(t)$ is the infinitesimal heat given to the body $1$ by the body $2$ between $t$ and $t+dt$, and $T_{1}(t),T_{2}(t)$ are the instantaneous internal temperatures of the two bodies. It is clear from the discussion following Eq. (\[Isolated\_S\_Generation\]) that the irreversible entropy generation in Eq. (\[Irreversible\_Entropy\_Change\_Total\]) is the sum of entropy changes in the two bodies. Thus, Eq. (\[Irreversible\_Entropy\_Change\_Total\]) is exact for the isolated system consisting of the two bodies. However, the point to remember is that it is the irreversible entropy generation in the two bodies, and says nothing about the irreversible entropy generation within each body.
If the system above were not isolated, then Eq. (\[Irreversible\_Entropy\_Change\_Total\]) would still give the entropy change due to the direct flow of heat between the two bodies, but it not represent the irreversible entropy generation due to this heat flow. Moreover, there would also be changes in the entropy of the system and each of the bodies due to heat exchange with the medium or other bodies. These entropy changes will also have their own irreversible entropy generations. The presence of the medium at the *equilibrium* temperature $T_{0}$ of the isolated system is discussed below. The situation when $T_{0}$ is not the equilibrium temperature of the isolated system is very different, as discussed later; see the discussion after Eq. (\[Entropy\_Generation\_Each\_Body\_0\]).
What can we say about the irreversible entropy generation within each body? This question is, to the best of our knowledge, is not answered within the local thermodynamics approach. To answer this question in our approach, we proceed as follows. We know that the entropy generation in each body must vanish when the bodies come to equilibrium. To obtain the desired result, we introduce the common temperature $T_{0}$, when the two bodies come to equilibrium. We now discuss two alternative approaches to determine the individual entropy generation on the way to prove Theorem \[Theorem\_Isolated\_Bodies\].
#### Introduction of a Medium at constant $T_{0}$
Let us imagine that we insert the two bodies in an extensively large medium $\widetilde{\Sigma}$ kept at a fixed temperature $T_{0}$, with all three bodies now forming a new isolated system $\Sigma_{0}^{\prime}$. The situation is schematically shown in Fig. \[Fig\_two\_boxex\_in\_medium\]. There is no direct contact between the two bodies, as shown. Despite this, the heat loss $dQ(t)\geq0$ by the body $2$ to $\widetilde{\Sigma}$ is completely transferred to body $1$. An alternative is to insert the medium between the two bodies along the $x$-axis but not surrounding them from all sides with the same effect of transferring the entire heat $dQ(t)$ from body $2$ to body $1$. The width of the medium along the $x$-axis may be infinitesimally small of order $dx$, but must have a macroscopically large cross-sectional area in the $yz$-plane to ensure its constant temperature $T_{0}$ at all times. In either case, the entropy of the medium does not change so that $d\widetilde
{S}=0,d_{\text{e}}\widetilde{S}=0$, and $d_{\text{i}}\widetilde{S}=0.$ Therefore, the irreversible entropy generation in the new isolated system $\Sigma_{0}^{^{\prime}}$ is identical to the irreversible entropy generation in $\Sigma_{0}$. This artificial introduction of $\widetilde{\Sigma}$ can now be exploited to obtain the irreversible entropy generation in each smaller body. The method of calculation above can be easily applied to this case, see Eq. (\[entropy\_generation\]), to yield $$d_{\text{i}}S_{1}(t)=dQ(t)\left( \frac{1}{T_{1}(t)}-\frac{1}{T_{0}}\right)
,\ \ d_{\text{i}}S_{2}(t)=dQ(t)\left( \frac{1}{T_{0}}-\frac{1}{T_{2}(t)}\right) , \label{Entropy_Generation_Each_Body_0}$$ where $dQ(t)\geq0$ is the heat addede to body $1$ or given out by body $2$. By comparing with Fig. \[Fig\_two\_boxes\_modified\_system\], we note that $dQ(t)=dQ_{1}(t)+dQ^{\prime}(t)=$ $dQ_{2}(t)+dQ^{\prime}(t)$, giving $$dQ_{1}(t)=dQ_{2}(t).$$ It is evident that these entropy productions vanish in equilibrium, as required by the notion of equilibrium between the two bodies. If we had introduced a medium held at a constant temperature $T_{0}^{^{\prime}}\neq
T_{0}$, then the equilibrium will occur at $T_{0}^{^{\prime}}$, and not at $T_{0}$, and we would end up with a different final state of $\Sigma
_{0}^{^{\prime}}$ than that of $\Sigma_{0}$. Thus, the medium must have the constant temperature $T_{0}$.
\[ptb\]
[Two\_Boxex\_in\_Medium.eps]{}
The situation will be very different if the medium is taken to be at a themperature $T_{0}^{\prime}\neq T_{0}$. In this case, the heat given to body $1$ is different from the heat given out by body $2$. Comparing with Fig. \[Fig\_two\_boxes\_modified\_system\], we note for the total amount of heat $dQ_{2,\text{tot}}(t)$ given out by $2$ and the total amount of heat $dQ_{1,\text{tot}}(t)$ given to $1$ that $dQ_{1,\text{tot}}(t)\equiv
dQ_{1}(t)+dQ^{\prime}(t)\neq$ $dQ_{2,\text{tot}}(t)\equiv dQ_{2}(t)+dQ^{\prime}(t)$, giving $$dQ_{1}(t)\neq dQ_{2}(t).$$ The irreversible entropy generation in the two bodies are now givem by$$d_{\text{i}}S_{1}(t)=dQ_{1,\text{tot}}(t)\left( \frac{1}{T_{1}(t)}-\frac
{1}{T_{0}^{\prime}}\right) ,\ \ d_{\text{i}}S_{2}(t)=dQ_{2,\text{tot}}(t)\left( \frac{1}{T_{0}^{\prime}}-\frac{1}{T_{2}(t)}\right) ,$$ so that $d_{\text{i}}S(t)\equiv d_{\text{i}}S_{1}(t)+d_{\text{i}}S_{1}(t)$ is given by $$d_{\text{i}}S(t)=dQ^{\prime}(t)\left( \frac{1}{T_{1}(t)}-\frac{1}{T_{2}(t)}\right) +dQ_{1}(t)\left( \frac{1}{T_{1}(t)}-\frac{1}{T_{0}^{\prime}}\right) +dQ_{2}(t)\left( \frac{1}{T_{0}^{\prime}}-\frac{1}{T_{2}(t)}\right) . \label{entropy_generation_Diff_T0_Medium}$$ Notice that for $T_{0}^{\prime}>T_{2}(t)>T_{1}(t),$ $dQ_{1}(t)>0,$ and $dQ_{2}(t)<0$. Similarly, for $T_{0}^{\prime}<T_{1}(t),$ $dQ_{1}(t)<0,$ and $dQ_{2}(t)<0$. Thus, each of the last two irreversible entropy generation contributions above is non-negative, as expected. We will see below that the last two contributions are absent in the local theory.
The above approach can now be extended to many bodies at different initial temperatures $T_{k}$. We assume that the medium $\widetilde{\Sigma}$ surrounds all the bodies so that there is no direct contact between the bodies. If $T_{0}$ still denotes the final temperature of all the bodies, treated as an isolated system $\Sigma_{0}$, then the medium $\widetilde{\Sigma}$ must be chosen to be also at the same constant temperature $T_{0}$. It is obvious that $$d_{\text{i}}S_{k}(t)=dQ_{k}(t)\left( \frac{1}{T_{k}(t)}-\frac{1}{T_{0}^{\prime}}\right) , \label{Entropy_Generation_Each_Body_1}$$ where $dQ_{k}(t)$ denotes the infinitesimal heat added to the $k$-th body between $t$ and $t+dt,$ and $T_{k}(t)$ is its instantaneous internal temperature; compare with Eq. (\[entropy\_generation\]). This is a generalization of Eq. (\[Irreversible\_Entropy\_Change\]) to many bodies, and exploits the absence of causality inherent in $d_{\text{e}}S_{k}(t)$, as exemplified by Eq. (\[Reversible\_Entropy\_Change\]). The irreversible entropy generation in the isolated system $\Sigma_{0}^{^{\prime}}$ consisting only of the bodies (without the medium) is the sum of all these contributions:$$d_{\text{i}}S_{0}(t)\equiv{\textstyle\sum\limits_{k}}
d_{\text{i}}S_{k}(t)={\textstyle\sum\limits_{k}}
\frac{dQ_{k}(t)}{T_{k}(t)}={\textstyle\sum\limits_{k}}
dS_{k}(t)=dS(t),$$ as it must be for an isolated system; Obviusly, $d_{\text{i}}S_{0}(t)$ does not depend on the final temperature $T_{0}$. Here, we have used the fact that $${\textstyle\sum\limits_{k}}
dQ_{k}(t)\equiv0$$ because of the isolation of $\Sigma_{0}$ and $\Sigma_{0}^{^{\prime}}$. It is quite clear that the discussion is easily extended to include other extensive variables which results in the expression for $d_{\text{i}}S_{k}(t)$, which is identical in form to the expression $d_{\text{i}}s_{k}(t)$in Eq. (\[entropy\_generation\]).
One may feel uneasy that the introduction of the fictitious medium $\widetilde{\Sigma}$ has changed the problem. To see that this is not the case, we reemphasize that its introduction does not affect the heat $dQ_{k}(t)$. Since, it is the heat $dQ_{k}(t)$ that determines the entropy generation, and since the equilibrium state of each body is the correct equilibrium state, the result must be correct. To offer an even stronger argument, we now provide an alternative proof without the introduction of the medium so that we can feel comfortable in associating a medium in the case when finite-size bodies form an isolated system. The importance of this trick is that it allows us to use all the results we have found by the use of a medium.
#### Without any Medium
Let us revert to the simple case of two bodies in thermal contact. In reality, the interface or the contact region between the two bodies is not a sharp boundary with a discontinuity in temperature; rather, it is a narrow region of width $\Delta x$ over which the temperature rapidly changes from $T_{1}(t)$ to $T_{2}(t)$, as shown in Fig. \[Fig\_two\_boxes\_isolated\](b). Thus, there exist a point E over this region where the temperature is precisely $T_{0}$. It is possible that the location of the point E along the $x$-axis changes with time, but this point is irrelevant. The relevant point is that the temperature in a very narrow width $dx<<\Delta x$ around and including this point will remain constant in time. If we take the narrow neighborhood of a point whose temperature is different from $T_{0}$, its temperature will eventually change to $T_{0}$, as equilibrium is achieved. Thus, the temperature of the narrow region around this point will not be constant in time. We take the region where the temperature is greater than $T_{0}$ along with the half-width $dx/2$ around E on the side of $2$ to be the part of the body $2$, while the region with the temperature less than $T_{0}$ along with the half-width $dx/2$ around E on the side of $1$ to be the part of the body $1$. From the quasi-independence of the two bodies, it is clear that the inclusion of these regions will not significantly affect the internal temperatures $T_{1}(t)$ to $T_{2}(t).$ The heat lost $dQ(t)$ by the body $2$ is transferred to the body $1$ at a *constant* temperature $T_{0}$. Thus, while $$dS_{1}(t)=\frac{dQ(t)}{T_{1}(t)},$$ where $T_{1}(t)$ is the internal temperature of the body $1$,$$d_{\text{e}}S_{1}(t)=\frac{dQ(t)}{T_{0}},$$ so that $$d_{\text{i}}S_{1}(t)=dQ(t)\left( \frac{1}{T_{1}(t)}-\frac{1}{T_{0}}\right)
,$$ as discovered above; see Eq. (\[Entropy\_Generation\_Each\_Body\_0\]). In a similar fashion, we obtain $$d_{\text{i}}S_{2}(t)=dQ(t)\left( \frac{1}{T_{0}}-\frac{1}{T_{2}(t)}\right)$$ in accordance with Eq. (\[Entropy\_Generation\_Each\_Body\_0\]).
The above discussion justifies the use of $T_{0}$ as a temperature in the interfacial region. However, the important point is that the determination of $d_{\text{e}}S(t)$ for any body requires the use of the equilibrium temperature of the body in accordance with Eq. (\[Reversible\_Entropy\_Change\]). Thus, the discussion is equally valid for any number of subsystems in the system.
#### Comaprison with Local Thermodynamics
The same result as in Eq. (\[Irreversible\_Entropy\_Change\_Total\]) is also obtained in the local thermodynamics, as can be easily seen; see for example [@Prigogine-old Eq. (3.14)]. In the limit in which the interfacial region between the two subsystems is infinitesimal in width along the $x$-direction, we can obtain the continuum analog of the irreversible entropy generation *between* the two neighboring regions. Denoting $T_{1}(t)$ by $T(x,t)$ and $T_{2}(t)$ by $T(x+dx,t)\simeq T(x,t)+\left( \partial T/\partial
x\right) dx,$ we have $$dS_{0}(t)=d_{\text{i}}S_{0}(t)\simeq dQ(t)\left[ \partial(1/T)/\partial
x\right] dx$$ for conduction. Dividing and multiplying by the cross-sectional area $A$ of the interface, and using $Adx$ as the volume of the interfacial region, we have for the rate $\sigma$ of entropy production per unit interfacial volume$$\sigma(x,t)=\overset{.}{q}(t)\left[ \partial(1/T)/\partial x\right] ,
\label{Rate_of_Entropy_Production}$$ where $\overset{.}{q}(t)=$ $\overset{.}{Q}(t)/A$ denotes the heat flux. This expression (in three dimensions, the result will contain $\boldsymbol{\partial
}(1/T)$) in this limit is a standard result for the entropy production due to heat conduction in local thermodynamics.
It should be stressed that our derivation of $\sigma(x,t)$ above is independent of the how long the two subsystems are along the $x$-axis, but assumes implicitly that $T(x,t)$ has a Taylor series expansion. It is also obvious that $\sigma(x,t)$ must be zero outside the interfacial region. Thus, $\sigma(x,t)$ should be correctly identified as the rate of entropy per unit volume in the *interfacial region*. Therefore, as it follows from the discussion immediately following Eq. (\[Irreversible\_Entropy\_Change\_Total\]), the derivation says nothing about how much of the irreversible entropy is generated within each body. The issue is avoided in local thermodynamics by assuming that the entire volume of the system is composed only of such interfacial regions. This is contrary to the basic postulate of local equilibrium according to which each local region has a well-defined temperature $T(t)$, while the interfacial regions have non-zero gradients of the temperature.
We also observe that the form of $\sigma(x,t)$ in Eq. (\[Rate\_of\_Entropy\_Production\]) is only valid for the case when Eq. (\[Irreversible\_Entropy\_Change\_Total\]) is valid. It is not valid for the case covered in Eq. (\[entropy\_generation\_Diff\_T0\_Medium\]). This is the case when the equilibrium temperature $T_{0}^{\prime}$ of the system is different from the equilibrium temperature $T_{0}$ of the two bodies under consideration if treated as isolated bodies. Thus, the above expression $\sigma(x,t)$ will not be valid if our system consists of a large number of bodies so that the irreversible entropy generation between any two bodies in contact will be given by Eq. (\[entropy\_generation\_Diff\_T0\_Medium\]). In this case, the expresasion explicitly contains the equilibrium temperature of the system in the last two terms, which is not the case with $\sigma(x,t)$ in Eq. (\[Rate\_of\_Entropy\_Production\]), thus verifying the earlier statment made after Eq. (\[entropy\_generation\_Diff\_T0\_Medium\]).
### Several Bodies
Let us now consider a simple extension of the case shown in Fig. \[Fig\_two\_boxes\_isolated\](a) in which there are several bodies $1,2,3,\cdots,$B in thermal contact along the $x$-axis forming an isolated body $\Sigma_{0}$, with their temperatures in an increasing order: $T_{1}<T_{2}<\cdots<T_{k}<T_{k+1}<\cdots<T_{\text{B}}$. Let $T_{k-1}<T_{0}<T_{k}$ again denote the equilibrium temperature $T_{0}$ of the isolated body $\Sigma_{0}$, so that there exist a point E over their interface region where the temperature is precisely $T_{0}$. Let us consider the infinitesimal heat $dQ_{1}(t)$ received by $1$ to be the part of the heat $dQ_{1}(t)$ that was given out by the body $k$ and was transferred unaltered via $k-1,k-2,\cdots3,2$ to $1.$ Any heat transfer through the interface region at constant $T_{0}$ is isothermal. Therefore, the reversible entropy change due to $dQ_{1}(t)$ is precisely $d_{\text{e}}S_{1}=$ $-dQ_{1}(t)/T_{0}$. The entropy change of $1$ is $dS_{1}=dQ_{1}(t)/T_{1}(t)$ so that it immediately follows that$$d_{\text{i}}S_{1}(t)=dQ_{1}(t)\left( \frac{1}{T_{1}(t)}-\frac{1}{T_{0}}\right) ,$$ as above. Similarly, we find$$d_{\text{i}}S_{\text{B}}(t)=-dQ_{\text{B}}(t)\left( \frac{1}{T_{\text{B}}(t)}-\frac{1}{T_{0}}\right)$$ for the body B, where $dQ_{\text{B}}(t)$ is the heat rejected by B, which was transferred isothermally to the body $k$ unaltetred. For the body $2$, which receives an infinitesimal heat $dQ_{2}(t)$, we similarly find$$d_{\text{i}}S_{2}(t)=dQ_{2}(t)\left( \frac{1}{T_{2}(t)}-\frac{1}{T_{0}}\right) ,$$ and so on. All these results can be easily obtained by inserting a medium held at fixed temperature $T_{0}$ at the interface between each connsecutive pair of bodies. Thus, we conclude that we can consider an isolated inhomogeneous body as a body embedded in a medium without affecting any of the consequences. This then proves Theorem \[Theorem\_Isolated\_Bodies\].
General Discussion
------------------
We are now in a position to provide a proof of Theorem \[Theorem\_Isolated\_Bodies\]. We consider the reversible entropy change $d_{\text{e}}s^{p}$ in a given subsystem due to the $p$-th conjugate field $W^{p}(t)$ ($Y(t)$ or $A(t))$ due to the change $dz^{p}$. It is given by $$d_{\text{e}}s^{p}=\frac{W_{0}}{T_{0}}dz^{p},$$ while $$ds^{p}=\frac{W(t)}{T(t)}dz^{p},$$ so that $$d_{\text{i}}s^{p}=F\left[ \frac{W(t)}{T(t)}\right] dz^{p}\geq0$$ in terms of the quantity $F\left[ w\right] $ in Eq. (\[Irreversible\_Change\]). This thus proves the theorem.
Discussion and Conclusions
==========================
We have proposed a scheme to extend our previous work in I to an inhomogeneous system in which subsystems may undergo relative translational and rotational motions with respect to each other to contribute to irreversibility. Another source of irreversibility is the temperature variation, which was considered in I. Each subsystem or the medium is identified by a set of extensive state variables $\mathbf{Z}(t)$, and some constant parameters $\mathbf{C}$, some of which may be external parameters and need not be extensive. Examples of $\mathbf{C}$ may be the number of particles in the system that characterize the system, or the translational and the angular velocities of the frame of reference in which the observations are made. Examples of the state variables are mechanical quantities such as energy, volume, etc. and the internal variables are the translational and angular momenta of the various subsystems, etc.
Quasi-independence and Additivity of Entropy and Energy
-------------------------------------------------------
As our approach starts from the second law, the law of increase of entropy, the entropy as a state function plays the most important role in our approach. Accordingly, we need to ensure that this quantity can always be determined by or at least formally defined in terms of some basic quantities pertaining to the macrostate of the system of interest. We use the Gibbs formulation of the entropy such as that in Eq. (\[Gibbs entropy\]), which is applicable in all cases as discussed in a recent review [@Gujrati-Symmetry]. This formulation is well-established for an isolated system [@Gujrati-Symmetry], but we show in Sect. \[marker\_entropy\_additivity\] that it is also applicable to an open system even when it is not in equilibrium with its surrounding medium. This formulation of entropy for open system is a well known result in equilibrium thermodynamics [@Landau] and our demonstration generalizes this result to non-equilibrium systems so that this entropy can be used as the central quantity in developing a non-equilibrium thermodynamics with the second law as the starting point. It is required that the open system be quasi-independent of the medium. We have shown in Sect. \[marker\_entropy\_additivity\] that this quasi-independence is a very important property, which is required for the *additivity* of the entropy and of energy. For example, quasi-independence of the system and the medium ensures that their energy of interaction is negligible so that dropping it makes their energies additive, as seen from Eqs. (\[Interaction\_Energy\]) and (\[Energy\_Sum\_Approx\]). Indeed, without the additivity of the energy, the entropy cannot be additive as discussed in Sect. \[marker\_entropy\_additivity\]. The basic idea is very simple but profound. The entropy being a state function must only depend on state variables. By definition, all state variables of a body must be solely determined by what happens within the body. If there are other bodies, their influence on the body must not destroy the additivity of all extensive state variables. Let us consider the isolated system consisting of the system $\Sigma$ and the medium $\widetilde{\Sigma}$ it is in contact with, which is considered in Sect. \[marker\_entropy\_additivity\]; in particular, consider the energies in Eq. (\[Interaction\_Energy\]). We will assume no relative motion between $\Sigma$ and the medium $\widetilde{\Sigma}$ to make the discussion simple. The energies $E_{0},E(t)$ and $\widetilde{E}(t)\ $then represent the energies of the isolated system, the system and the medium, respectively. However, the presence of the interaction energy $E_{0}^{(\text{int})}(t)$ in this equation destroys the required property that the right side of Eq. (\[Interaction\_Energy\]) is the sum of internal energies of the system and the medium; the interaction energy is a property of both bodies. Accordingly, the entropy $S_{0}(E_{0})$ of $\Sigma_{0}$ cannot depend only on the energies of the two bodies separately. The difference represented by $S_{0}^{(\text{int})}(t)$ in Eq. (\[Entropy\_Loss\_Sum\]) depends on both bodies. Thus, the entropy of the isolated system cannot be a sum of entropies, each representing the entropy of one of the bodies alone. In other words, generalizing this result to a collection of bodies b$_{j}$ forming a body B, we conclude that the entropy $S($B$)$ of B *cannot* be expressed as a sum of entropies, each term $S($b$_{j})$ representing the entropy of the body b$_{j}$$$S(\text{B})\neq\sum_{j}S(\text{b}_{j}).$$ However, the additivity requires not only that the sum of the energies of various parts b$_{j}$ of a body must yield the energy of the body B itself$$E(\text{B})=\sum_{j}E(\text{b}_{j}),$$ which is satisfied if the linear sizes of the bodies are large compared with the range of interaction, but also requires that various bodies are quasi-independent, which is satisfied if the linear sizes of the bodies are large compared with thecorrelation length. Unless thess properties hold, the entropy cannot be expressed as a sum of the entropies of its parts, with each entropy depending only on the properties of the part alone. It is the latter property that makes entropy a state function.
Concept of Internal Equilibrium
-------------------------------
For a body (or its parts, such as subsystems) out of equilibrium, the entropy usually has an *explicit* dependence on time in addition to the implicit time-dependence$$S(\text{B})=S(\mathbf{X},\mathbf{I},t);$$ the latter arises from the dependence of the entropy on state variables $\mathbf{X}(t)$ and $\mathbf{I}(t)$ that explicitly depend on time; we will suppress this time-dependence in the following for notational simplicity. (We will assume, as discussed in [@Gujrati-review:Fluctuations], that at least one observable, which we take to be the number of particles $N$ is held fixed and is not contained in $\mathbf{X}$.) Thus, such an entropy will continue to change (increase if the body is isolated) for *fixed* state variables. For a homogeneous body, this will happen if the state variables in $\mathbf{X}$ and $\mathbf{I}$ do not uniquely specify its state. In other words, there may be other internal variables not contained in $\mathbf{I}$. This is a connsequence of a simple generalization of Theorem \[marker\_absence\_of\_internal\_equilibrium\]. The other possibility is that the body is not homogeneous. If the isolated body is not in equilibrium, its entropy continues to increase according to the second law in Eq. (\[Second\_Law\_0\]). However, if it happens that the entropy of the body does not explicitly depend on time$$S_{\text{IE}}(\text{B})=S(\mathbf{X},\mathbf{I}),$$ then its entropy must be at its *maximum* for fixed $\mathbf{X}$ and $\mathbf{I}$. This is how we had introduced the concept of internal equilibrium in I: in this state, the entropy of a system, which is not in equilibrium with the medium $\widetilde{\Sigma}(\mathbf{Y}_{0},0)$, has no explicit time-dependence. It can be brought in contact with another medium $\widetilde{\Sigma}^{\prime}$, representing the medium $\widetilde{\Sigma
}(\mathbf{Y}_{\text{IE}},\mathbf{A}_{\text{IE}})$, so that it remains in equilibrium with the new medium. In other words, there is no difference between a body in internal equilibrium with the medium $\widetilde{\Sigma
}(\mathbf{Y}_{0},0)$, and the body in equilibrium with the medium $\widetilde{\Sigma}(\mathbf{Y}_{\text{IE}},\mathbf{A}_{\text{IE}})$; in the latter case, the medium ensures to keep the averages $\mathbf{X}_{\text{IE}}$ and $\mathbf{I}_{\text{IE}}$ fixed. In both cases, the entropy is maximum for the fixed values of their state variables $\mathbf{X}_{\text{IE}}$ and $\mathbf{I}_{\text{IE}}$. Accordingly, all properties of a body in equilibrium also hold for a body in internal equilibrium at each instant. For example, Theorem \[marker\_Uniform\_Motion\] established in Sect. \[marker\_internal\_equilibrium\_0\] shows that there *cannot* be any relative motion within such a body. The only possible motion is a *uniform translation* and a *rigid-body rotation*. As a consequence, as discussed in Sect. \[marker\_internal\_equilibrium\_0\], there is no viscous dissipation *within* the body in internal equilibrium. Thus, while there is no relative motion between the system and $\widetilde{\Sigma}(\mathbf{Y}_{\text{IE}},\mathbf{A}_{\text{IE}})$, there will in general be relative motions between the system and the medium $\widetilde{\Sigma
}(\mathbf{Y}_{0},0)$. This result can be generalized to the following statement: The only source of viscous dissipation in a collection of bodies is then due to relative motions between different bodies, when each of which is postulated to be in internal equilibrium.
\[ptb\]
[Entropy\_Forms.eps]{}
It was assumed in I that the system $\Sigma$ at each instant remains in internal equilibrium as it goes from one equilibrium macrostate $\mathcal{M}^{\prime}$ in contact with a medium $\widetilde{\Sigma}(\mathbf{Y}_{0}^{^{\prime}},0)$ at time $t=0$ to another equilibrium macrostate $\mathcal{M}^{0}$ in contact with a medium $\widetilde{\Sigma}(\mathbf{Y}_{0},0)$ as $t\rightarrow\infty$. Let us try to understand the behavior of its entropy $S(t)$ during this process, which can then be extended to a subsystem in internal equilibrium. The field variables for the two macrostates are $\mathbf{W}^{\prime}\equiv(\mathbf{Y}^{^{\prime}}=\mathbf{Y}_{0}^{^{\prime}},\mathbf{A}^{^{\prime}}=0)$ and $\mathbf{W}^{0}\equiv(\mathbf{Y}=\mathbf{Y}_{0},\mathbf{A}=0)$, respectively. As discussed several times, the intermediate macrostate $\mathcal{N}(t)$ at some instant $t$ can be thought of as an equilibrium state of the system $\Sigma$ in contact with a medium $\widetilde{\Sigma}(\mathbf{Y}_{\text{IE}}=\mathbf{Y}(t),\mathbf{A}_{\text{IE}}=\mathbf{A}(t))$ after being disconnected from the medium $\widetilde{\Sigma}(\mathbf{Y}_{0},0)$. In this equilibrium state, the system has a well-defined field $\mathbf{W}(t)=(\mathbf{Y}(t),\mathbf{A}(t))$ and the entropy $S(t)$, which is shown by the solid curve in Fig. \[Fig\_entropy\]. However, this equilibrium state of the system does not represent the equilibrium state of the system in contact with either the medium $\widetilde{\Sigma}(\mathbf{Y}_{0},0)$ or the medium $\widetilde{\Sigma
}(\mathbf{Y}_{\text{IE}},0)$. If we bring the system in contact with the medium $\widetilde{\Sigma}(\mathbf{Y}_{\text{IE}},0),$ its entropy will increase as the internal variables relax towards the new equilibrium, to be denoted by $\mathcal{M}(t)$ with fixed $\mathbf{X}_{\text{IE}}$. The equilibrium entropy of the system in $\mathcal{M}(t)$ will be denoted by $S_{\text{eq}}(t)$, which is shown by the dotted curve in Fig. \[Fig\_entropy\].
Because of our assumption of internal equilibrium, $S(t)$ represents the maximum possible entropy for the macrostate $\mathcal{N}(t)$ at $t$. The change in the entropy $\Delta S(t)$ is given by $$\Delta S(t)=S(t)-S^{\prime},$$ where $S^{\prime}$ is the equilibrium entropy in the macrostate $\mathcal{M}^{\prime}$ at $t=0$, and is shown by the left-most arrow in Fig. \[Fig\_entropy\]. The total entropy change in the entire process is $$\Delta S_{\infty}=S^{0}-S^{\prime},$$ where $S^{0}$ as the equilibrium entropy in the macrostate $\mathcal{M}^{0}$ as $t\rightarrow\infty$.
The entropy difference $\Delta S(t)$ can always be broken into two parts in analogy with the two terms in eq. (\[entropy\_generation\_0\]):$$\Delta S(t)=\Delta_{\text{e}}S(t)+\Delta_{\text{i}}S(t),$$ with $\Delta_{\text{e}}S(t),$ see the right-most arrow in Fig. \[Fig\_entropy\], representing the entropy exchange with the medium $\widetilde{\Sigma}(\mathbf{Y}_{0},0)$ and $\Delta_{\text{i}}S^{0}\geq0$ the irreversible entropy production within the system, see the middle small arrow in Fig. \[Fig\_entropy\]. It does not matter whether the process is reversible or irreversible, the entropy change is always $\Delta S(t)$. In the former process, $\Delta_{\text{e}}S(t)=\Delta S(t)$, $\Delta_{\text{i}}S(t)=0$ and $S(t)=S_{\text{eq}}(t)$. In the latter process, $$\Delta_{\text{e}}S(t)<\Delta S(t)$$ and there is irreversible entropy production during the passage from $\mathcal{M}^{\prime}$ to $\mathcal{N}(t)$, $$\Delta_{\text{i}}S(t)>0$$ in accordance with Eq. (\[entropy\_generation\_limit\]). We thus see that irreversible generation of entropy has been accounted for in our approach; see also I and Eq. (\[entropy\_generation\_limit\]). We can easily extend the above discussion to the subsystems, where we now also have the possibility of irreversible entropy generation through relative motions between them, which is also accounted for in our approach.
We have found (Theorem \[marker\_absence\_of\_internal\_equilibrium\]) that while a body is in internal equilibrium in the presence of internal variables, it is not in internal equilibrium when internal variables are not considered in its description. In other words, the entropy of a body, see the solid curve for $S(t)$ in Fig. \[Fig\_entropy\],$$S(\text{B})=S_{\text{IE}}(\text{B})=S(\mathbf{X},\mathbf{I})
\label{Body_Internal_Equilibrium}$$ has no explicit time-dependence; however, it becomes an explicit function of time if expressed only in terms of observables:$$S(\text{B})=S(\mathbf{X},t). \label{Body_Not_Internal_Equilibrium}$$ This theorem shows how important the internal variables are in describing glasses. Because of the presence of $t$, this entropy will continue to increase and approach $S_{\text{eq}}(t)$, shown in Fig. \[Fig\_entropy\], if we keep $\mathbf{X}$ fixed.
As experimentalists can only control the observables in $\mathbf{X}$, it is the expression of the entropy in Eq. (\[Body\_Not\_Internal\_Equilibrium\]) that is experimentally relevant. Because of the explicit $t$-dependence, the body will continue to relax even if the observables are held fixed by isolating the body from its surroundings and bringing it in contact with the medium $\widetilde{\Sigma}(\mathbf{Y}_{\text{IE}},0)$. This relaxation occurs because the internal variable $\mathbf{I}$ is no longer held fixed. It will continue to change with time until finally the entropy reaches its maximum possible value for fixed $\mathbf{X}$. We can apply Theorem \[marker\_chemical\_potential\] for an isolated system to our body. The conclusion is that the maximum of the entropy occurs when $\mathbf{A}$ for the internal variable $\mathbf{I}$ is identically zero, which is why we had selected this particular medium. The final value of $\mathbf{I}$ is some value $\mathbf{I}_{\text{eq}}(\mathbf{X})$. The body in this equilibrium state is the fully relaxed body with its entropy given by$$S_{\text{eq}}(\text{B})=S_{\text{eq}}(\mathbf{X});$$ this entropy is shown as $S_{\text{eq}}(t)$ in Fig. \[Fig\_entropy\]. The affinity $\mathbf{A}$ of the internal variable is identically zero, and the fields in this state are given by $\mathbf{Y\equiv Y}_{\text{IE}}$. This equilibrium state of the isolated body should not be confused with the state of the body in internal equilibrium in the presence of the internal variable $\mathbf{I}$, as shown in Eq.(\[Body\_Internal\_Equilibrium\]). This body has $\mathbf{I}$ different from $\mathbf{I}_{\text{eq}}(6\mathbf{X})$, and the corresponding affinity $\mathbf{A}$ different from zero.
Thermodynamic Potentials
------------------------
The additivity of the entropy and energy allows us to treat our system as a collection of quasi-independent subsystems so that we can develop the thermodynamics of an inhomogeneous system. Under the mildest possible assumption that the medium is in internal equilibrium so that its field variables $\widetilde{w}_{k}$ are defined via Eq. (\[Chemical\_Potential\_Internal\_Variables\]), but the system itself is (or the subsystems are) not necessarily in internal equilibrium, we are able to identify the thermodynamic potential for the system. The field variables of the system in internal equilibrium vary with time and, therefore, they differ from the medium’s field variables. They become identical only when the system is in equilibrium with the medium. It is obvious from Eq. (\[Chemical\_Potential\_Internal\_Variables\]) that the product $$\widetilde{w}_{k}\widetilde{Z}_{k},\text{ no summation implied}$$ is dimensionless because the entropy is defined as dimensionless in this work. Accordingly, the (dimensionless) entropy $\widetilde{S}(\widetilde{\mathbf{Z}})$ can be expressed in terms of these dimensionless products. It follows then that resulting thermodynamic potentials will contain these products, as shown in Sect. \[marker\_inhomogeneous\_arbitrary\]. The actual form of the thermodynamic potential function depends on the choice of the set $\mathbf{C}$, and requires field variables to maintain the dimensions of each term appearing in it. The fields variables in the thermodynamic potentials are the field variables of the medium, and not of the system as the latter are not even defined when the system in not in internal equilibrium; see for example, Eq. (\[Free\_Energies\]) or (\[Thrmo\_Potential\]). This is true whether we consider the system or any of its many subsystems; the latter are considered in Sect. \[marker\_inhomogeneous\_arbitrary\]. This is surprising since one knows that, in equilibrium thermodynamics, the thermodynamic potentials are state quantities. Therefore, they must contain all quantities related to the system. It just happens that in equilibrium the field variables are identical to the medium’s field variables. Therefore, our thermodynamic potentials reduce to the standard thermodynamic potentials in equilibrium. In this case, they become state functions. But this is not true when the system is not in equilibrium. What is surprising is that this result remains true even when the system or subsystem is in internal variable. The thermodynamic potentials always contain fields of the medium and not of the system or subsystem. These functions have the required thermodynamic property that they can never increase in any spontaneous process, as seen from Eq. (\[General\_Thermodynamic\_Potential\_Variation0\]) or (\[General\_Thermodynamic\_Potential\_Variation\]). One can easily see that this property is a consequence of the *convexity* property of the thermodynamic potentials. In contrast, a function obtained from thermodynamic potentials by replacing medium’s fields by the body’s field, when the latter is in internal equilibrium, does not have this required thermodynamic property. An example of this is the function $\widehat{g}(t)$ or $\widehat
{G}(t)$ from $g(t)$ or $G(t)$ given in Eq. (\[Local\_Gibbs\_Free\_Energy\]) or (\[Free\_Energies\_Incorrect\]).
While each subsystem is in internal equilibrium, so that there can be no irreversible processes inside them, the irreversibility occurs due to relative motions going on among them. Because of the internal equilibrium, each subsystem has its own internal temperature $T_{k}(t)$, pressure $P_{k}(t)$ or other conjugate field variables$.$ How are these temperatures and pressures related to the internal temperature $T(t)$ and pressure $P(t)$ introduced in I? To answer these questions, we consider their definitions in the $\mathcal{L}$ frame:$$\begin{aligned}
\frac{1}{T(t)} & =\frac{\partial S}{\partial E},\frac{P(t)}{T(t)}=\frac{\partial S}{\partial V}\\
\frac{1}{T_{k}(t)} & =\frac{\partial s_{k}}{\partial e_{k}},\frac{P_{k}(t)}{T_{k}(t)}=\frac{\partial s_{k}}{\partial v_{k}};\end{aligned}$$ see Eq. (\[subsystem\_parameters0\]) for the definition of $T_{k}(t)$ and $P_{k}(t)$. Let us first consider the temperature. Introducing$$r_{k}(t)\equiv\frac{\partial E}{\partial e_{k}},{\textstyle\sum}
r_{k}(t)=1,$$ we have$$\frac{\partial S}{\partial E}={\textstyle\sum\nolimits_{k}}
\frac{\partial s_{k}}{\partial e_{k}}/\frac{\partial E}{\partial e_{k}}={\textstyle\sum\nolimits_{k}}
\frac{r_{k}(t)}{T_{k}(t)}.$$ Thus, we have $$\frac{1}{T(t)}=\sum\nolimits_{k}\frac{r_{k}(t)}{T_{k}(t)}.$$ Introducing $$r_{k}^{\text{v}}(t)\equiv\frac{\partial V}{\partial v_{k}},{\textstyle\sum_{k}}
r_{k}^{\text{v}}(t)=1,$$ we find that$$\frac{P(t)}{T(t)}=\sum\nolimits_{k}\frac{r_{k}^{\text{v}}(t)P_{k}(t)}{T_{k}(t)}.$$ The same exercise can be carried out for other state variables. Introducing$$r_{k}^{l}(t)\equiv\frac{\partial Z_{l}}{\partial z_{lk}},{\textstyle\sum_{k}}
r_{k}^{l}(t)=1,$$ for the $l$-th state variable $Z_{l}$, we find that $$w_{l}(t)\equiv\frac{W_{l}(t)}{T(t)}=\sum\nolimits_{k}\frac{r_{k}^{l}(t)y_{k}(t)}{T_{k}(t)}.$$
Contrast with Local Non-equilibrium Thermodynamics
--------------------------------------------------
Our approach differs from the traditional local non-equilibrium thermodynamics due to de Groot [@Donder; @deGroot; @Prigogine; @Prigogine-old], briefly discussed in Sect. \[marker\_averages\], in three important ways.
1. The first one relates to the principle of additivity of energy. The energy in the local non-equilibrium thermodynamics is not the sum of the energies of its various part due to the presence of the mutual interaction energies expressed by $\psi(\mathbf{r})dV$ as seen from Eq. (\[Energy\_Additivity\_Thermodynamics\]). Indeed, usually one applies local equilibrium to a volume element $dV$, what is traditionally called a physically infinitesimal volume in that, while it contains a large number of atoms, the corresponding volume is still infinitesimally small. Such a volume is conventionally called a “particle” (not to be confused with our usage which refers to an atom or molecule). For all practical purposes, it is indeed considered as a limit $dV\rightarrow0$. It is evident that one must then consider the interaction energy $\psi(\mathbf{r})dV$ to account for the interaction of this volume with the rest of the system. Thus, the local non-equilibrium thermodynamics takes the additivity of entropy is taken as postulate even when the energy is not additive.
2. Because the volume element $dV$ is treated infinitesimal in the local theory, all thermodynamic quantities are also treated as continuous in space, while this continuity is not a prerequisite in our approach.
3. Our approach also differs from the local non-equilibrium theory in the form of the thermodynamic potentials. The volume element is considered to be in internal equilibrium from the start so that it has a well-defined temperature, pressure, etc. These fields are used in identifying the thermodynamic potentials. For example, the Gibbs free energy of the “particle” is taken to be $\widehat{g}(t)dV$even if the system consisting of such “particles” is not in equilibrium with the medium.
4. Another imporatnt difference between the two approaches is that the reversible entropy change and the irreversible entropy generation in each subsystem also depends on the equilibrium state of the system. The irreversible entropy generation in the local theory depends only on the current local properties.
These differences make our approach very different from the local non-equilibrium thermodynamics due to de Groot [@Donder; @deGroot; @Prigogine; @Prigogine-old].
**Acknowledgements** I would like to thank Arkady Leonov for a discussion on the local non-equilibrium thermodynamics and him and Peter Wolynes for their suggestion to apply the approach of I to inhomogeneous systems. I would also like to thank an anonymous referee of I who wanted the current approach to include internal dissipation.
Relation Between Lab and Body Frames \[Appd\_Frames\]
=====================================================
Let us consider a particle of mass $m$ in the lab frame $\mathcal{L}$, where it has a velocity $\mathbf{v}_{\mathcal{L}}$ and the potential energy $U$. The Lagrangian $L_{\mathcal{L}}$ in this frame is given by $$L_{\mathcal{L}}=\frac{1}{2}m\mathbf{v}_{\mathcal{L}}^{2}-U.$$ Let us transform to a rotating frame $\mathcal{C}$ which is moving with a velocity $\mathbf{V}$ and rotating with an angular velocity $\mathbf{\Omega.}$ The velocity of the particle in $\mathcal{C}$ is given by $\mathbf{v}_{\mathcal{C}}$ and is related to $\mathbf{v}_{\mathcal{L}}$ according to $$\mathbf{v}_{\mathcal{L}}=\mathbf{v}_{\mathcal{C}}\mathbf{+V+\Omega\times
r}_{\mathcal{C}}\mathbf{,} \label{velocity_Relation}$$ where $\mathbf{r}_{\mathcal{C}}$ is the coordinate of the particle in the $\mathcal{C}$ frame, and is related to the coordinate $\mathbf{r}_{\mathcal{L}}$ of the particle in the lab frame $\mathcal{L}$ by$$\mathbf{r}_{\mathcal{L}}\equiv\mathbf{R}+\mathbf{r}_{\mathcal{C}}.$$ In the following, it would also be convenient to consider a nonrotating frame $\mathcal{I}$, which is only moving with the velocity $\mathbf{V}$ with respect to the lab frame, but whose origin coincides with that of $\mathcal{C}$. The Lagrangian in the frame $\mathcal{C}$ is given by$$\begin{aligned}
L_{\mathcal{C}} & =\frac{1}{2}m\mathbf{(\mathbf{v}_{\mathcal{C}}+V+\Omega\times r}_{\mathcal{C}}\mathbf{)}^{2}-U\nonumber\\
& =\frac{1}{2}m\mathbf{v}_{\mathcal{C}}^{2}+\frac{1}{2}m\mathbf{(\Omega\times
r}_{\mathcal{C}}\mathbf{)}^{2}+m\mathbf{v}_{\mathcal{C}}\cdot\mathbf{V+}m\mathbf{v}_{\mathcal{C}}\cdot\mathbf{\Omega\times\mathbf{r}}_{\mathcal{C}}\mathbf{+}m\mathbf{V}\cdot\mathbf{\Omega\times\mathbf{r}}_{\mathcal{C}}\mathbf{-}U, \label{Energy_Relation}$$ in which we have omitted $\frac12
m\mathbf{V}^{2}$, which is a total time derivative.
The canonical momentum is obtained as $$\mathbf{p}_{\mathcal{C}}=\frac{\partial L_{\mathcal{C}}}{\partial
\mathbf{v}_{\mathcal{C}}}=m(\mathbf{v}_{\mathcal{C}}\mathbf{+V+\Omega
\times\mathbf{r}}_{\mathcal{C}}\mathbf{)},$$ so that the energy of the particle becomes$$E_{\mathcal{C}}=\mathbf{p}_{\mathcal{C}}\cdot\mathbf{v}_{\mathcal{C}}-L_{\mathcal{C}}=\frac{1}{2}m\mathbf{v}_{\mathcal{C}}^{2}-\frac{1}{2}m\mathbf{(\Omega\times\mathbf{r}}_{\mathcal{C}}\mathbf{)}^{2}-m\mathbf{V}\cdot\mathbf{\Omega\times\mathbf{r}}_{\mathcal{C}}\mathbf{+}U.
\label{Internal_Energy_particle}$$ Expressing $\mathbf{v}_{\mathcal{C}}$ in terms of $\mathbf{v}_{\mathcal{L}}$ from Eq. (\[velocity\_Relation\]), we find that $$E_{\mathcal{C}}=E_{\mathcal{L}}+\frac{1}{2}m\mathbf{V}^{2}-\mathbf{p}_{\mathcal{L}}\cdot\mathbf{V}-m\mathbf{r}_{\mathcal{C}}\times\mathbf{v}_{\mathcal{L}}\cdot\mathbf{\Omega=}E_{\mathcal{L}}+\frac{1}{2}m\mathbf{\mathbf{V}^{2}-\mathbf{p}}_{\mathcal{L}}\mathbf{\cdot\mathbf{V}-m}\cdot\mathbf{\Omega}, \label{Energy_Transformation_particle}$$ where $$E_{\mathcal{L}}=\frac{1}{2}m\mathbf{v}_{\mathcal{L}}^{2}+U,\ \mathbf{p}_{\mathcal{L}}=m\mathbf{v}_{\mathcal{L}},\ \ \mathbf{m=}m\mathbf{r}_{\mathcal{C}}\times\mathbf{v}_{\mathcal{L}}. \label{Definition}$$ Incidentally, we also note that $$\mathbf{p}_{\mathcal{C}}\mathbf{=p}_{\mathcal{L}}.$$ Thus, $$\mathbf{m}=\mathbf{r}_{\mathcal{C}}\times\mathbf{p}_{\mathcal{L}}=\mathbf{r}_{\mathcal{C}}\times\mathbf{p}_{\mathcal{C}},$$ which explains why there is no need to use $\mathcal{L}$ or $\mathcal{C}$ as a subscript in $\mathbf{m}$.
A Rotating and Translating System\[Appd\_Rotating\_System\_0\]
==============================================================
Let us now extend the previous calculation for a single particle to a system of particles of total mass $M$ at a given instant $t$. The system is also characterized by the number of particles and its volume. For specificity, we use $N$ and $V(t)$ to denote these quantities. The notation should not mean that we are only considering the system $\Sigma$ here. The system we have in mind is any generic system. We assume that this system is translating with a velocity $\mathbf{V}(t)$ and rotating with an angular velocity $\mathbf{\Omega
}(t)$ as a whole, both of which can change in time. For notational simplicity, we will suppress the explicit $t$-dependence of various quantities here. We focus on one particular instant $t$. We take the center of mass of this system to coincide with the origin of $\mathcal{C}$, so that $\mathcal{C}$ is fixed to the body and rotating with it. In this case, $\mathcal{C}$ represents the *center of mass frame*. We now sum Eq. (\[Energy\_Transformation\_particle\]) over all particles, with the result that it is replaced by$$E_{\mathcal{C}}=E_{\mathcal{L}}+\frac{\mathbf{P}^{2}}{2M}-\mathbf{P}_{\mathcal{L}}\cdot\mathbf{V}-\mathbf{M}\cdot\mathbf{\Omega},
\label{Energy_Transformation_System}$$ where we have introduced$$\mathbf{P=}M\mathbf{V},\ \ \mathbf{P}_{\mathcal{L}}={\displaystyle\sum}
m\mathbf{v}_{\mathcal{L}},\ \ \mathbf{M\equiv}{\displaystyle\sum}
\mathbf{r}_{\mathcal{C}}\times\mathbf{p}_{\mathcal{L}}\equiv{\displaystyle\sum}
\mathbf{r}_{\mathcal{C}}\times\mathbf{p}_{\mathcal{C}}\mathbf{.}
\label{System_Definition}$$ Here, $\mathbf{M}_{\mathcal{C}}$ is the *intrinsic* angular momentum of the system of particles in the $\mathcal{C}$ frame. The summation in the above formulas, which are also applicable to non-uniform rotation and translation of the center of mass frame $\mathcal{C}$, is a sum over all the particles in the system. The value of $E_{\mathcal{C}}$ in the $\mathcal{C}$ frame represents the internal energy $E^{\text{i}}$ of the system. The equation (\[Energy\_Transformation\_System\]) generalizes the result given by Landau and Lifshitz [@Landau-Mechanics see their Eq. (39.13)] to the case when $\mathbf{V}\neq0.$ The present generalization is not limited to constant rotation and translation.
We can express $\mathbf{M}$ as follows:$$\mathbf{M=}{\displaystyle\sum}
m\mathbf{r}_{\mathcal{C}}\times(\mathbf{v}_{\mathcal{C}}\mathbf{+V+\Omega
\times r}_{\mathcal{C}})={\displaystyle\sum}
m\mathbf{r}_{\mathcal{C}}\times\mathbf{v}_{\mathcal{C}}+\left(
{\displaystyle\sum}
m\mathbf{r}_{\mathcal{C}}\right) \times\mathbf{V+}{\displaystyle\sum}
m\mathbf{r}_{\mathcal{C}}\times(\mathbf{\Omega\times r}_{\mathcal{C}}),$$ in which the second sum on the right vanishes for $\mathcal{C}$, the center of mass frame. For the same reason, the third term in the second equation in Eq. (\[Internal\_Energy\_particle\]) does not contribute to the energy of the system. Thus, we find
$$\begin{aligned}
\mathbf{M} & \mathbf{=}{\displaystyle\sum}
m\mathbf{r}_{\mathcal{C}}\times\mathbf{v}_{\mathcal{C}}+{\displaystyle\sum}
m\mathbf{r}_{\mathcal{C}}\times(\mathbf{\Omega\times r}_{\mathcal{C}})\label{Angular_momentum_C}\\
E_{\mathcal{C}} & =\frac{1}{2}{\displaystyle\sum}
m\mathbf{v}_{\mathcal{C}}^{2}-\frac{1}{2}m\mathbf{(\Omega\times\mathbf{r}}_{\mathcal{C}}\mathbf{)}^{2}\mathbf{+}U. \label{Internal_Energy_C_system}$$
We see that
$$\mathbf{M\cdot\mathbf{\Omega}=}{\displaystyle\sum}
m\mathbf{r}_{\mathcal{C}}\cdot\left( \mathbf{v}_{\mathcal{C}}\times
\mathbf{\Omega}\right) +{\displaystyle\sum}
m(\mathbf{\Omega\times r}_{\mathcal{C}})^{2}. \label{Angular_Momentum_Work}$$
The first term vanishes when the motion is radial. Thus, it is the contribution to the energy from the relative transverse motion in the $\mathcal{C}$ frame and will vanish if the system is stationary in the this frame. The latter condition is met when the system is in internal equilibrium in accordance with Theorem \[marker\_Uniform\_Motion\].
For the center of mass frame $\mathcal{C},$ $\mathbf{P=P}_{\mathcal{L}},$ so that the energy of the system in the frame $\mathcal{C}$ is given by$$E_{\mathcal{C}}=E_{\mathcal{L}}-\frac{\mathbf{P}^{2}}{2M}-\mathbf{M}\cdot\mathbf{\Omega,} \label{Internal_Energy}$$ which can be rewritten as $$E_{\mathcal{L}}=E_{\mathcal{C}}+\frac{\mathbf{P}^{2}}{2M}+\mathbf{M}\cdot\mathbf{\Omega} \label{Energy_L_system0}$$ Using Eq. (\[Internal\_Energy\_C\_system\]), we find that $$E_{\mathcal{L}}=E_{\mathcal{C}}+\frac{\mathbf{P}^{2}}{2M}+\frac{1}{2}{\displaystyle\sum}
m(\mathbf{\Omega\times r}_{\mathcal{C}})^{2}+{\displaystyle\sum}
m\mathbf{r}_{\mathcal{C}}\cdot\left( \mathbf{v}_{\mathcal{C}}\times
\mathbf{\Omega}\right) , \label{Energy_L_system1}$$ in which the last term is the contribution of the transverse motion.
The energy of the system in the $\mathcal{I}$ frame, in which the system has only rotation, is given by $$E_{\mathcal{I}}=E_{\mathcal{L}}-\frac{\mathbf{P}^{2}}{2M}.
\label{Internal_Energy_Iframe}$$ It is obvious that $E_{\mathcal{I}}$ does not depend on the velocity $\mathbf{V}$. Thus,$$E_{\mathcal{C}}=E_{\mathcal{I}}-\mathbf{M}\cdot\mathbf{\Omega}$$ does not depend explicitly on $\mathbf{V}$. We can take $E_{\mathcal{C}}$ as a function of $E_{\mathcal{I}}$ and $\mathbf{\Omega,}$ from which it follows that$$\begin{aligned}
\left. \left( \frac{\partial E_{\mathcal{C}}}{\partial\mathbf{V}}\right)
\right\vert _{E_{\mathcal{I}},V,N,\mathbf{\Omega}} & =0\mathbf{,}\label{Ec_V_Relation}\\
\left. \left( \frac{\partial E_{\mathcal{C}}}{\partial\mathbf{\Omega}}\right) \right\vert _{E_{\mathcal{I}}\mathbf{,}V,N} & =-\mathbf{M}.
\label{Ec_Mc_Relation}$$ All the above results are for a particular microstate of the system undergoing a Hamiltonian dynamics. To obtain thermodynamics of the system, we need to average the above equations over all microstates using their probabilities, which will be taken up in Sect. \[marker\_Rotating\_Systems\]. The averaging takes into account the stochastic nature of the dynamics, which has not been considered in both appendices.
[99]{}
P.D. Gujrati, Phys. Rev. E **81**, 051130 (2010); P.D. Gujrati, arXiv:0910.0026.
L.D. Landau and E.M. Lifshitz, *Fluid Mechanics*, Pergamon Press, Oxford (1982).
L.D. Landau and E.M. Lifshitz, *Theory of Elasticity*, Third edition, Pergamon Press, Oxford (!986).
E. Bouchbinder and J.S. Langer, Phys. Rev. E **80**, 031131 (2009); *ibid.* 031132 (2009); *ibid.* 031133 (2009).
\(a) V. Lubchenko and P.G. Wolynes, J. Chem. Phys. **121**, 2852 (2004).
H.C. Öttinger, Phys. Rev. E **74**, 011113 (2006); see also H. C. Öttinger, *Beyond Equilibrium Thermodynamics,* Wiley, Hoboken, NJ (2005).
Y. Oono and M. Paniconi, Prog. Theor. Phys. Suppl. **130**, 29 (1998).
S-I. Sasa and H. Tasaki, J. Stat. Phys.**125**, 125 (2006).
S.R. de Groot and P. Mazur, *Non-Equilibrium Thermodynamics***,** First Edition, Dover, New York (1984).
Th. de Donder and P. van Rysselberghe, *Thermodynamic Theory of Affinity*, Stanford University, Stanford (1936).
D. Kondepudi and I. Prigogine, *Modern Thermodynamics*, John Wiley and Sons, West Sussex (1998).
I. Prigogine, *Introduction to Thermodynamics of Irreversible Processes*, Third Edition, Interscience Publishers, New York (1967).
W. Muschik,S. Gümbel, M. Kröger, and H.C. Öttinger, Physica A **285**, 448 (2000).
Y. Demirel, *Nonequilibrium Thermodynamics*, Elsevier, Amsterdam (2007).
P.D. Gujrati, Symmetry **2**, 1201 (2010).
L.D. Landau and E.M. Lifshitz, *Statistical Physics*, Part I, Third edition, Pergamon Press, Oxford (!986).
M. Merolle, J.P. Garrahan and D. Chandler, Proc. Nat. Acad. Sci. **102**, 10837 (2005).
P.G. Wolynes, Proc. Nat. Acad. Sci. **106**, 1353 (2009).
L. Hong, P.D. Gujrati, V.N. Novikov and A.P. Sokolov, J. Chem. Phys. **131**, 194511 (2009).
R.O. Davies and G.O. Jones, Adv. Phys. **2**, 370 (1953).
P.D. Gujrati in *Modeling and Simulation in Polymers,* edited by P.D. Gujrati and A.I Leonov, Wiley-VCH, Weinheim (2010).
Nemilov, S.V., *Thermodynamic and Kinetic Aspects of the Vitreous State*, CRC Press, Boca Raton (1995).
I. Gutzow and J.W.P. Schmelzer, *The Vitreous State: Thermodynamics, Structure, Rheology and Crystallization*, Springer, Berlin (1995).
A. Sommerfeld, *Mechanics of Deformable Bodies*, Academic Press, New York (1950).
G.K. Batchelor, *An Introduction to Fluid Dynamics*, Cambridge University, Cambridge (1967).
L.D. Landau and E.M. Lifshitz, *Mechanics*,Third edition, Pergamon Press, Oxford (!982).
G.D.C. Kuiken, *Thermodynamics of Irreversible Processes*, John Wiley and Sons, Chichester (1994).
P.D. Gujrati and P.P. Aung, arXiv:1101.0431.
P.D. Gujrati, *Phys. Rev. E* **51**, 957 (1995); P.D. Gujrati, *J. Chem. Phys.* **108**, 5089 (1998); *ibid.* **108**, 5104 (1998).
M. Mosayebi, E. Del Gado, P. Ilg, and H.C. Ottinger, Phys. Rev. Lett. **104**, 205704 (2010).
J.H. Keenan, *Thermodynamics*, MIT Press, Cambridge (1941).
J. Frenkel, Kinetic Theory of Liquids, Dover, N.Y. (1955).
P.D. Gujrati, Recent Res. Devel. Chem. Physics, **4**, 243 (2003).
P.D. Gujrati, S.S. Rane, and A. Corsi, Phys. Rev. E **67**, 052501 (2003).
G.A. Maugin, *The Thermomechanics of Nonlinear Irreversible Behaviors: An Introduction*, World Scientific, Singapore (1999).
L.I. Mandelstam and M.A. Leontovich, Zh. Eksp. Teor. Fiz. **7**, 438 (1937).
V.N. Pokrovski, Eur. J. Phys. **26**, 769 (2005).
I. Prigogine and P. Mazur, Physica **19**, 241 (1953).
B.D. Coleman and M.E. Gurtin, J. Chem. Phys. **47**, 597 (1967).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper, we introduce some explicit approximations for the summation $\sum_{k\leq n}\Omega(k)$, where $\Omega(k)$ is the total number of prime factors of $k$.'
address:
- 'Mohammad Avalin Charsooghi, Yousof Azizi and Laleh Mollazadeh-Beidokhti, Department of Physics, Institute for Advanced Studies in Basic Sciences, P.O. Box 45195-1159, Zanjan, Iran'
- 'Mehdi Hassani, Department of Mathematics, Institute for Advanced Studies in Basic Sciences, P.O. Box 45195-1159, Zanjan, Iran'
author:
- 'M. Avalin Charsooghi, Y. Azizi, M. Hassani and L. Mollazadeh-Beidokhti'
title: On a Result of Hardy and Ramanujan
---
Introduction
============
Let $\Omega(k)$ be the total number of prime factors of $k$. A result of Hardy and Ramanujan [@hardy-rama] asserts that $$\sum_{k\leq n}\Omega(k)=n\log\log n+M'n+O\Big(\frac{n}{\log n}\Big),$$ where $$M'=\gamma+\sum_{p}\left(\log\Big(1-p^{-1}\Big)+(p-1)^{-1}\right)\approx
1.0346538818.$$ More related results can be found in Chapter V of [@s-m-c]. The aim of this paper is to find an explicit version of this result. We proceed by $$n!=\prod_{p\leq n}p^{v_p(n!)},$$ standard factorization of $n!$ into primes. It is known that $$v_p(n!)=\sum_{k=1}^m\Big\lfloor\frac{n}{p^k}\Big\rfloor,$$ where $\lfloor x\rfloor$ is the largest integer less than or equal to $x$ (see for example [@nathan]) and $m=m_{n,p}=\lfloor\frac{\log n}{\log p}\rfloor$. First, we introduce some explicit (and neat) approximations for the summation $$\Upsilon(n)=\sum_{p\leq n}v_p(n!).$$ Then, considering $$\sum_{k\leq n}\Omega(k)=\Omega(n!)=\Upsilon(n),$$ we obtain the main result as follows.\
**Main Theorem.** *For every $n\geq 3$ we have* $$\left|\sum_{k\leq n}\Omega(k)-(n-1)\log\log(n-1)\right|<23(n-1).$$ Note that one can modify above result to the following one: $$\left|\sum_{k\leq n}\Omega(k)-n\log\log n\right|<23n,$$ which is an explicit version of the result of Hardy and Ramanujan.
Proof of the Main Theorem
=========================
Consider the inequality $$\label{vp-inqs}
\frac{n-p}{p-1}-\frac{\log n}{\log p}<v_p(n!)\leq \frac{n-1}{p-1},$$ (see [@hassani-vp] for a proof). To get to the main theorem, we need to approximate summations of the form $\sum_{p\leq n}f(p)$ with $f(p)=\frac{1}{\log p}$ and $f(p)=\frac{1}{p-1}$ (and more generally, for a given function $f\in C^1(\mathbb{R}^+)$). To do this, we use the reduction of a Riemann-Stieljes integral to a finite sum [@apos-anal], which allows us to get some ways to evaluate the summation $\sum_{p\leq n}f(p)$; two of them are:\
- Using $\vartheta(x)=\sum\limits_{p\leq x}\log p$, which ends to the approximation $$\sum_{p\leq n}f(p)=\int_{2^-}^n \frac{f(x)}{\log
x}d\vartheta(x)=\frac{f(n)\vartheta(n)}{\log
n}+\int_2^n\vartheta(x)\frac{d}{dx}\left(\frac{-f(x)}{\log
x}\right)dx,$$ and it is known that for $x>1$, we have $200\log^2x|\vartheta(x)-x|<793x$, and $\log^4x|\vartheta(x)-x|<1717433x$ (see [@dusart] for more details).
- Using $\pi(x)=\#\mathbb{P}\cap [2,x]$, which ends to the approximation $$\sum_{p\leq
n}f(p)=f(x)\pi(x)+\int_2^n\pi(x)\frac{d}{dx}\big(-f(x)\big)dx,$$ and we have some explicit bounds for $\pi(x)$ (again see [@dusart] for lots of them). In this paper we will use the following neat one: $$\label{pi-inq}
\pi(x)\leq\frac{x}{\log x}\Big(1+\frac{1.2762}{\log
x}\Big)\hspace{10mm}(x>1).$$
Both of these methods are applicable for the summation $\sum_{p\leq n}\frac{1}{p-1}$, while first method on the summation $\sum_{p\leq n}\frac{1}{\log p}$ ends to some integrals hard to approximate. Here, based on some known approximations for both of these summations, which are obtained using the second method, we give some neat bounds for them.
\[sum-1/p-1\] For every $n\geq 3$, we have $$\log\log(n-1)-14<\sum_{p\leq n}\frac{1}{p-1}<\log\log(n-1)+23.$$
It is known [@hassani-n!] that the inequality $$\log\log n+a+\frac{n}{(n-1)\log n}-\frac{1717433n}{(n-1)\log^5
n}<\sum_{p\leq n}\frac{1}{p-1},$$ holds for $n\geq 2$ with $a\approx -11.86870152$. But, for $n\geq 3564183$ we have $$\log\log(n-1)-14<\log\log n+a+\frac{n}{(n-1)\log
n}-\frac{1717433n}{(n-1)\log^5 n}.$$ Thus, for $n\geq 3564183$ we obtain $$\log\log(n-1)-14<\sum_{p\leq n}\frac{1}{p-1},$$ which is also true for $2\leq n\leq 3564182$, since for these values of $n$ the left hand side of the inequality is positive while the right hand side is negative. Also, we have [@hassani-n!] the inequality $$\sum_{p\leq n}\frac{1}{p-1}<\log\log(n-1)+b+\frac{n}{(n-1)\log
n}+\frac{1717433 n}{(n-1)\log^5 n},$$ for $n\geq 2$ with $b\approx 21.18095291$. On the other hand, for $n\geq
7126157$ we have $$b+\frac{n}{(n-1)\log n}+\frac{1717433 n}{(n-1)\log^5 n}<23.$$ So, for $n\geq 7126157$ we obtain $$\sum_{p\leq n}\frac{1}{p-1}<\log\log(n-1)+23.$$ To verify this inequality for $3\leq n\leq 7126156$, we note that because for $p_1\leq n<p_2$ where $p_1$ and $p_2$ are two successive primes, the left hand side is constant, while the right hand side is increasing, therefore we only need to check this inequality for $n$ equals to prime numbers. Appendix includes the Matlab program of doing this. The proof is completed.
\[sum-1/logp\] For every $n\geq 2$, we have $$\left|\sum_{p\leq n}\frac{1}{\log p}-\left\{\frac{n}{\log^2
n}+\frac{2n}{\log^3 n}+\frac{6n}{\log^4
n}\right\}\right|<271382\frac{n}{\log^5 n}.$$
In a similar process [@hassani-n!], we have $$\frac{n}{\log^2 n}+\frac{2n}{\log^3 n}+\frac{6n}{\log^4
n}+\frac{1607n}{100\log^5 n}-\frac{1717433n}{\log^6 n}+a<\sum_{p\leq
n}\frac{1}{\log p}\hspace{10mm}(n\geq 564),$$ where $a\approx -16.42613005$. Also, we have $$\sum_{p\leq n}\frac{1}{\log p}<\frac{n}{\log^2 n}+\frac{2n}{\log^3
n}+\frac{6n}{\log^4 n}+\frac{54281n}{800\log^5
n}+\frac{1717433n}{\log^6 n}+b\hspace{10mm}(n\geq 2),$$ where $b\approx 30.52238614$. Computation gives $$\frac{-271382n}{\log^5 n}<\frac{1607n}{100\log^5
n}-\frac{1717433n}{\log^6 n}+a\hspace{10mm}(n\geq 564).$$ Also $$\frac{54281n}{800\log^5 n}+\frac{1717433n}{\log^6
n}+b<\frac{271382n}{\log^5 n}\hspace{10mm}(n\geq 569).$$ Therefore, we obtain the following inequality: $$\left|\sum_{p\leq n}\frac{1}{\log p}-\left\{\frac{n}{\log^2
n}+\frac{2n}{\log^3 n}+\frac{6n}{\log^4
n}\right\}\right|<271382\frac{n}{\log^5 n}\hspace{10mm}(n\geq 569).$$ A computer program verifies the above inequality for $2\leq n\leq
568$, too. The proof is complete.
*Proof of the Main Theorem.* Considering the right hand side of (\[vp-inqs\]) and the Proposition \[sum-1/p-1\], for every $n\geq 3$ we have $$\Upsilon(n)\leq (n-1)\sum_{p\leq
n}\frac{1}{p-1}<(n-1)\log\log(n-1)+23(n-1).$$ On the other hand, considering the left hand side of (\[vp-inqs\]) and the Proposition \[sum-1/p-1\], for every $n\geq 3$ we have $$(n-1)\log\log(n-1)-14(n-1)-\mathcal{R}(n)<(n-1)\sum_{p\leq
n}\frac{1}{p-1}-\pi(n)-\log n\sum_{p\leq n}\frac{1}{\log
p}<\Upsilon(n),$$ where $$\mathcal{R}(n)=\pi(n)+\log n\sum_{p\leq n}\frac{1}{\log p},$$ and considering (\[pi-inq\]) and the Proposition \[sum-1/logp\], we have $$\mathcal{R}(n)\leq\frac{n}{\log n}\Big(1+\frac{1.2762}{\log
n}\Big)+\frac{n}{\log n}+\frac{2n}{\log^2 n}+\frac{6n}{\log^3
n}+\frac{271382n}{\log^4 n}=\frac{2n}{\log n}+\frac{3.2762n}{\log^2
n}+\frac{6n}{\log^3 n}+\frac{271382n}{\log^4 n}.$$ But, for $n\geq 563206$ the right hand side of this relation is strictly less than $9(n-1)$. So, we obtain $$(n-1)\log\log(n-1)-23(n-1)<\Upsilon(n),$$ for $n\geq 563206$, which holds true for $3\leq n\leq 563205$ too, because for these values of $n$, the left hand side is positive while the right hand side is negative. This completes the proof.
Remarks for Further Studies
===========================
Improving the Main Result
-------------------------
Of course the factor 23 in the main theorem is not optimal, and one can improve it. But, it is the best one with our methods and computational tools.
Explicit Approximation of the Function $\Omega(n)$
--------------------------------------------------
Concerning the main theorem, considering $n!=\Gamma(n+1)$, one can reform the above result as $$\left|\Omega(\Gamma(n))-(n-2)\log\log(n-2)\right|<23(n-2),$$ then replacing $n$ by $\Gamma^{-1}(n)$ (inverse of Gamma function), it yields to $$\left|\Omega(n)-(\Gamma^{-1}(n)-2)\log\log(\Gamma^{-1}(n)-2)\right|<23\left(\Gamma^{-1}(n)-2\right).$$ This suggests an explicit approximation for the function $\Omega(n)$ for some special values of $n$ in terms of the inverse of Gamma function, then by approximating $\Gamma^{-1}$, one can make it in terms of elementary functions.
An Extension of the Function $v_p(n!)$
--------------------------------------
The function $v_p(n!)$, defined by $$n!=\prod_{p\leq n}p^{v_p(n!)},$$ can be generalized for every positive integer $m\leq n$ instead of prime $p\leq n$. Fix $n$ and consider canonical decomposition $$m=\prod_{p\leq n}p^{v_p(m)}.$$ Same to $v_p(n!)$, we define $v_m(n!)$ in which $m^{v_m(n!)}\|n!$. So, $$m^{v_m(n!)}=\prod_{p\leq n}p^{v_p(m)v_m(n!)}\Big\|\prod_{p\leq
n}p^{v_p(n!)}.$$ Therefore, we must have $v_p(m)v_m(n!)\leq v_p(n!)$ for every prime $p\leq n$; that is $$v_m(n!)\leq\min_{\substack{
p\leq n\\
v_p(m)\neq 0}}\left\{\frac{v_p(n!)}{v_p(m)}\right\}.$$ This leads to the following definition:\
**Definition.** For positive integers $m,n$ with $m\leq n$, we set $$v_m(n!)=\left\lfloor\min_{\substack{
p\leq n\\
v_p(m)\neq 0}}\left\{\frac{v_p(n!)}{v_p(m)}\right\}\right\rfloor.$$ Note that in the above definition, $v_p(N)$ for a positive integer $N$ and prime $p$, is a well defined notation for the greatest power of $p$ dividing $N$. Related by this generalization, the following question arise to mind:\
**Question.** Find the function $\mathfrak{F}(n)$ such that $$\sum_{m=1}^n v_m(n!)=\mathfrak{F}(n)\sum_{p\leq n}v_p(n!).$$
Acknowledgment {#acknowledgment .unnumbered}
--------------
We would like to express our gratitude to the referee for valuable comments.
[99]{}
M. Abramowitz and I.A. Stegun, [*Handbook of Mathematical Functions: with Formulas, Graphs, and Mthematical Tables*]{}, Dover Publications, 1972.
Apostol, *Mathematical Analysis*, Addison-Wesley Publishing Company, Inc., 1957.
P. Dusart, Inégalités explicites pour $\psi(X)$, $\theta(X)$, $\pi(X)$ et les nombres premiers, *C. R. Math. Acad. Sci. Soc. R. Can.* **21** (1999), no. 2, 53-59.
G. Hardy and S. Ramanujan, The normal number of prime factors of a number $n$, *Quart. J. Math.* **48** (1917), 76-92.
M. Hassani, Equations and Inequalities Involving $v_p(n!)$, *Journal of Inequalities in Pure and Applied Mathematics (JIPAM)*, Volume **6**, Issue 2, Article 29, 2005.
M. Hassani, On the decomposition of $n!$ into primes, arXiv:math/0606316v5 \[math.NT\]. (submitted for publication)
Melvyn B. Nathanson, [*Elementary Methods in Number Theory*]{}, Springer, 2000.
József Sándor; Dragoslav S. Mitrinović; Borislav Crstici, [*Handbook of Number Theory I*]{}, Second printing of the 1996 original, Springer, Dordrecht, 2006.
------------------------------------------------------------------------
------------------------------------------------------------------------
------------------------------------------------------------------------
<span style="font-variant:small-caps;">Appendix. Matlab program of verifying the inequality $\sum_{p\leq n}\frac{1}{p-1}<\log\log(n-1)+23$ for prime values of $n$</span>
[`n=8000000; r=primes(n); s(1)=0; for i=2: length(r) s(i)=s(i-1)+1/(r(i)-1); end plot(r,s,’.’,r,log(log(r)))+23,’.’)`]{}\
\
Final step of program plots both sides of the inequality for comparison.
------------------------------------------------------------------------
------------------------------------------------------------------------
------------------------------------------------------------------------
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'It is shown that the collective nonlinear interactions between intense neutrino or anti-neutrino fluxes and a dense neutrino plasma are governed by a multi-dimensional coupled cubic Schrödinger equation in which the interaction potential is positive or negative depending on the neutrino type. The cubic Schrödinger equation describes the splitting and focusing of intense neutrino beams due to the nonlinear excitations associated with the modifications of the individual neutrino energies in a dense neutrino background.'
address:
- ' Department of Electromagnetics, Chalmers University of Technology, SE–412 96 Göteborg, Sweden'
- '§ Institut für Theoretische Physik IV, Fakultät für Physik und Astronomie, Ruhr-Universität Bochum, D–44780 Bochum, Germany'
- '$||$ Department of Plasma Physics, Ume[å]{} University, SE–901 87 Ume[å]{}, Sweden'
author:
- 'Mattias Marklund, Padma K. Shukla§[$||$]{} and Lennart Stenflo$||$'
title: Splitting and focusing of neutrino collective states
---
Keywords: neutrino properties, supernova neutrinos, ultra high energy photons and neutrinos
{#section .unnumbered}
Neutrinos and anti-neutrinos are produced in the Sun and in supernovae during violent processes. The propagation of intense neutrinos in ionised media (dense plasmas) is a topic of current interest in view of the fact that neutrinos may be massive [@Ahmad]. Because of their weak interaction with charged and neutral particles, neutrinos (or anti-neutrinos) can travel great distances without being affected appreciably by material obstacles. Massive neutrinos can account for neutrino oscillations and associated flavour change in solar plasmas [@John; @Raffelt], while in supernovae and cosmology they may play a decisive role in reviving shocks for supernovae explosions, as well as in accounting for the missing dark matter in our Universe [@Dolgov].
In the present paper, we employ a semi-classical approach to consider the nonlinear interaction between neutrinos and anti-neutrinos in neutrino plasmas. Using an effective field theory approach, a system of coupled nonlinear Schrödinger equations is derived, and it is shown that these equations admit splitting and focusing of intense neutrino (or anti-neutrino) beams due to the nonlinearities associated with the weak nuclear forces of the neutrinos/anti-neutrinos that are interacting with the background neutrinos.
Suppose a single neutrino (or anti-neutrino) moves in a neutrino–anti-neutrino admixture. The energy $E$ of the neutrino (anti-neutrino) is then given by (see, e.g. [@Bethe; @Kuo-Pantaleone; @Silva-etal]) $$\label{energy}
E = \sqrt{p^2c^2 + m^2c^4} + V_{\pm}(\bm{r},t) ,$$ where $\bm{p}$ is the neutrino (anti-neutrino) momentum, $c$ is the speed of light in vacuum, and $m$ is the neutrino mass. The effective potential for a neutrino moving on a background of it’s own flavor is [@Kuo-Pantaleone; @Notzhold-Raffelt] (see also [@Weldon; @Nunokawa-etal]) $
V_{\pm}(\bm{r},t) = \pm 2\sqrt{2}G_F (n - \bar{n}),
$ while the potential for a neutrino moving on a background of a different flavour takes the form $
V_{\pm}(\bm{r},t) = \pm \sqrt{2}G_F (n - \bar{n}).
$ Here $G_F/(\hbar c)^3 \approx 1.2\times 10^{-5}\,
\mathrm{GeV}^{-2}$, where $G_F$ is the Fermi constant, $n$ ($\bar{n}$) is the density of the background neutrinos (anti-neutrinos), and the $+$ ($-$) is for the propagating neutrino (anti-neutrino). From Eq. (\[energy\]) we can, using standard techniques [@Hasegawa], obtain a Schrödinger equation for a neutrino (anti-neutrino) wave packet ${{\mathnormal{\Psi}}}(\bm{r}, t)$ [@Marklund-etal] (see also Ref. [@Tsintsadze-etal] for a similar treatment of neutrino–electron interactions) $$\label{nlse}
i\frac{\partial{{\mathnormal{\Psi}}}}{\partial t} +
\frac{\hbar}{2m_{\mathrm{eff}}}\nabla^2{{\mathnormal{\Psi}}} -
\frac{V_{\pm}}{\hbar}{{\mathnormal{\Psi}}} = 0 ,$$ where $m_{\mathrm{eff}} = m \gamma$, $\gamma = (1 - v^2/c^2)^{-1/2}$, and $v$ is the magnitude of the group velocity of the neutrino wave packet.
In the case of self-interacting neutrinos and anti-neutrinos, moving with the same group velocity, we have $ n =
\langle|{{\mathnormal{\Psi}}}_{+}|^2\rangle$ and $\bar{n} =
\langle|{{\mathnormal{\Psi}}}_{-}|^2\rangle$, where ${{\mathnormal{\Psi}}}_{+}$ and ${{\mathnormal{\Psi}}}_{-}$ are the neutrino and anti-neutrino wave functions, respectively, and the angular bracket denotes the ensemble average. Thus, in this case Eq. (\[nlse\]) turns into the coupled system $$\begin{aligned}
i\frac{\partial{{\mathnormal{\Psi}}}_{\pm}}{\partial t} +
\frac{\alpha}{2}\nabla^2{{\mathnormal{\Psi}}}_{\pm} \mp
\beta(\langle|{{\mathnormal{\Psi}}}_{+}|^2\rangle - \langle|{{\mathnormal{\Psi}}}_{-}|^2\rangle
){{\mathnormal{\Psi}}}_{\pm} = 0 ,
\label{interaction}\end{aligned}$$ where $\alpha = \hbar/m_{\mathrm{eff}}$ and $\beta =
2\sqrt{2}G_F/\hbar$ for neutrinos of the same flavor.
Suppose now that the neutrino pulses are coherent and uni-directional. If one species, the pump species (denoted by the index $p$), dominates the number density as compared to the signal species (denoted by the index $s$), i.e. $n_p \gg n_s$, we obtain the equation $$i\frac{\partial{{\mathnormal{\Psi}}}_s}{\partial t} +
\frac{\alpha}{2}\frac{\partial^2{{\mathnormal{\Psi}}}_s}{\partial z^2} +
\beta|{{\mathnormal{\Psi}}}_p|^2{{\mathnormal{\Psi}}}_s = 0 ,
\label{split}$$ for the signal wave function ${{\mathnormal{\Psi}}}_s$, where the driving nonlinear term is determined from $$i\frac{\partial{{\mathnormal{\Psi}}}_p}{\partial t} +
\frac{\alpha}{2}\frac{\partial^2{{\mathnormal{\Psi}}}_p}{\partial z^2} -
\beta|{{\mathnormal{\Psi}}}_p|^2{{\mathnormal{\Psi}}}_p = 0 .
\label{dark}$$ \[system\] It is well known that Eq. (\[dark\]) have modulationally stable dark soliton solutions [@Kivshar], of which the fundamental soliton is $${{\mathnormal{\Psi}}}_p = {{\mathnormal{\Psi}}}_0\tanh\left(
\sqrt{\beta/\alpha}\,{{\mathnormal{\Psi}}}_0z
\right)\,\exp({-i\beta{{\mathnormal{\Psi}}}_0^2t}) .
\label{fundamental}$$
![Contour plot of $\ln[n_s(t,z)/n_{s0}]$ ($b = 1$) for the splitting case.[]{data-label="split_contour"}](split_contour3){width=".6\textwidth"}
![Contour plot of $\ln[n_s(t,z)/n_{s0}]$ ($b = 1$) for the focusing case.[]{data-label="focus_contour"}](focus_contour){width=".6\textwidth"}
Following Ref. [@Helczynski-etal], we perform a perturbation analysis of Eq. (\[split\]) for a given solution ${{\mathnormal{\Psi}}}_p(t,z)$ to Eq. (\[dark\]). Separating ${{\mathnormal{\Psi}}}_s(t,z)$ into its real amplitude $\sqrt{n_s(t,x)}$ and phase $\theta_s(t,z)$, we have, to lowest order in $t$, $$\theta_s(t,z) \approx \beta\int_0^t |{{\mathnormal{\Psi}}}_p(t',z)|^2\,dt' \approx
\beta |{{\mathnormal{\Psi}}}_p(0,z)|^2 t,$$ where the last equality follows if the interaction takes place over a time-scale short compared to the time over which the pump amplitude changes. The signal number density $n_s(t,z)$ is then $$\fl n_s(t,z) \approx n_s(0,z)\exp\bigg[ -\frac{\alpha\beta
t^2}{2}\bigg(
\frac{d^2 n_p(0,z)}{dz^2}
+ \frac{d n_p(0,z)}{dz}\frac{d}{dz}\ln
n_s(0,z) \bigg) \bigg] ,$$ where $n_p(t,z) = |{{\mathnormal{\Psi}}}_p(t,z)|^2$ is the pump number density. Using the fundamental soliton (\[fundamental\]) as the pump pulse, with ${{\mathnormal{\Psi}}}_0^2 = n_p(0,0) \equiv n_{p0}$, and a Gaussian distribution for the initial signal pulse, $$n_s(0,z) = n_{s0}\exp\left(-z^2/a^2\right),
\label{gauss}$$ where $a$ is the width and $n_{s0} = n_s(0,0)$, we find that $$\fl n_s(t,z) = n_{s0}\exp\Big\{ -b^2\zeta^2
-\Big[ \left(3{\,\mathrm{sech}}^2\zeta - 2
\right)\!{\,\mathrm{sech}}^2\zeta
-
2b^2\tanh\zeta{\,\mathrm{sech}}^2\zeta \Big]\tau^2 \Big\} ,
\label{solution}$$ where $\zeta = z/z_0$, $\tau = \alpha t/z_0^2$, $b = z_0/a$, and $z_0 =
(\beta n_{p0}/\alpha)^{1/2}$. We note that $b$, $\tau$ and $\zeta$ are dimensionless. The solution (\[solution\]) is plotted in Fig. \[split\_contour\], where the splitting of the collective neutrino state can clearly be seen.
In the case of neutrino–neutrino or anti-neutrino–anti-neutrino interaction, Eq. (\[interaction\]) will be modified according to $$\begin{aligned}
i\frac{\partial{{\mathnormal{\Psi}}}_{i}}{\partial t} +
\frac{\alpha}{2}\nabla^2{{\mathnormal{\Psi}}}_{i} -
\beta(\langle|{{\mathnormal{\Psi}}}_{1}|^2\rangle + \langle|{{\mathnormal{\Psi}}}_{2}|^2\rangle
){{\mathnormal{\Psi}}}_{i} = 0 ,
\label{interaction2}\end{aligned}$$ where $i= 1, 2$. For a dark soliton pump beam (\[fundamental\]) and a Gaussian signal beam (\[gauss\]), the approximate dynamics, analogous to the neutrino–anti-neutrino interaction, can be obtained from (\[solution\]) by letting $\tau^2 \rightarrow -\tau^2$. The result is seen in Fig. \[focus\_contour\], where the focusing is depicted.
The above discussion is based on the approximate solution (\[solution\]). However, numerical studies of the system (\[split\]) and (\[dark\]) show the same behaviour, when (\[fundamental\]) and (\[gauss\]) are used as initial conditions. Thus, a more complete solution supports the notion that the approximate analytical expression (\[solution\]) gives an accurate description of the dynamics of interacting neutrinos and anti-neutrinos.
In conclusion, we have considered the nonlinear interactions between neutrinos and anti-neutrinos, mediated by a dense neutrino background of the same flavour. It is shown that the self-interaction between neutrinos and anti-neutrinos comes about due to their coupling via the weak nuclear force. Such interactions are governed by coupled cubic Schrödinger equations for the slowly varying envelopes of the neutrino and anti-neutrino wave functions. Analytical and numerical analysis of the multi-dimensional cubic Schrödinger equations reveal that a collective neutrino state can be split or focused due to the presence of a dense soliton background. The present results suggest that nonlinear excitations in neutrino plasmas can significantly affect the propagation of intense neutrino and anti-neutrino beams. The neutrino beam splitting can account for the energy loss of neutrinos, while the self-focusing renders neutrino burst propagation over a long distance.
References {#references .unnumbered}
==========
[99]{}
Ahmad Q R, [*et al.*]{} 2002 *Phys. Rev. Lett.* [**89**]{} 011301
Bahcall J 1989 *Neutrino Astrophysics* (Cambridge: Cambridge University Press)
Raffelt G 1996 *Stars as Laboratories for Fundamental Physics* (Chicago: Chicago University Press)
Dolgov A D 2002 *Phys. Rep.* **370** 333
Sofue Y and Rubin V C 2001 *Ann. Rev. Astron. Astrophys.* **39** 137
Bethe H A 1987 *Phys. Rev. Lett.* **56** 1305; Bethe H A 1990 *Rev. Mod. Phys.***62**, 801
Kuo T K and Pantaleone J 1989 *Rev.Mod. Phys.* **61** 937
Silva L O, [*et al.*]{} 2000 *Astrophys. J.* **127** 481
Weldon H A 1982 *Phys. Rev. D* **26** 2789
Nötzhold D and Raffelt G 1988 *Nucl. Phys. B* **307** 924
Nunokawa H, *et al.* 1997 *Nucl. Phys. B* **501** 17
Hasegawa A 1975 *Plasma Instabilities and Nonlinear Effects* (Berlin: Springer-Verlag)
Marklund M, *et al.* 2003 submitted\[astro-ph/0306013\].
Tsintsadze N L, *et al.* 1998 *Phys. Plasmas* **5** 3512; Shukla P K, *et al.* 2002 *Phys. Plasmas* **9** 3625
Kivshar Y S and Luther–Davies B 1998 *Phys. Rep.* **298** 81
Helczynski L, *et al.* 2000 *Physica Scripta* **T84** 81
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'A solid-on-solid model is proposed to describe faceting of bcc(111) metal surface induced by a metal overlayer. It is shown that the first order phase transition occurs between faceted {211} or {110} and disordered phases. The ordered phases consist of large 3-sided pyramids with {211} facets or {110} facets. It is shown that the high-temperature disordered phase has not planar bcc(111) structure but faceted disordered structure. Hysteresis effects were observed when the system was warmed above the transition temperature and then cooled down. Temperature dependence of LEED patterns for faceted and disordered phase are calculated in kinematic approximation.'
author:
- 'Czes[ł]{}aw Oleksy'
date: 'March 4, 2003'
title: SOS model of overlayer induced faceting
---
Introduction
============
Recent experiments [@mad99a; @mad99b; @mad96; @mad94; @mad97; @mad95; @song95] have demonstrated that surfaces such as W(111) and Mo(111) covered by a single physical monolayer of certain metal, e.g., Pd, Pt, undergo massive reconstruction from planar morphology to microscopically faceted surface after annealing to $T>700 K$. The reconstructed surface consists of 3-sided pyramids with mainly $\{211\}$ facets, and pyramid dimensions range from $\sim 1$ to 100 nm. It has been shown that facets are composed of substrate atoms, and the monolayer of adsorbate remains on the outermost surface layer during the faceting transformation. Another type of massive reconstruction has been very recently found in STM and LEED experiment for Pd on Ta(111) system [@szuk02]. The reconstructed surface consists of $\{011\}$ facets which form large triangular pyramids. The third type of reconstructed structure, coexistence of small $\{011\}$ facets with large $\{211\}$ facets, has been observed [@mad99b] in Pd on W(111) for prolonged annealing time (for short annealing time only $\{211\}$ pyramids occurred). An important step in understanding the thermal stability of reconstructed surfaces was LEED experiment performed in high temperatures for Pd on Mo(111) by Song et al. [@song95]. They demonstrated existence of reversible phase transition between faceted $\{211\}$ and planar phases and they found that this transition has a large hysteresis, i.e., the transition temperature in cooling cycle ($T\approx 830$) is lower than that in the heating cycle ($T\approx 870$).
In theoretical studies of overlayer-induced faceting [@mad99b; @leung97; @leung98] the first principle method have been used to calculate the surface formation energy of fcc metals (Pd, Pt, Au, Ag, Cu) adsorbed on Mo and W. Results of these calculation performed for pseudomorphic adsorbate overlayer on (111), (211), and (011) flat surfaces show that (111) surface becomes unstable at the coverage of one physical monolayer, where the energy of (211) orientation has the lowest value. However, such energy calculation (at T=0) are not sufficient to explain why some of these metals do not induce faceting, e.g., Ag/Mo(111). On the other hand, the first principle calculation confirmed that critical coverage to induce faceting is approximately equal to one physical monolayer (PML) and that the growth mode is Stransky-Krastanov as the surface energy increases at coverage higher than 1 PML.
Theoretical studies of surface reconstruction and surface phase transitions in bimetalic system are complicated problems mainly due to occurrence of long range many-body interactions. One of approaches to such problems is application of simple solid-on-solid models [@beijeren77; @tosat95; @beijeren95; @kaseno97], in which the crystal is represented by two-dimensional array of columns. They were employed to study roughening transition [@beijeren77; @tosat95], missing-row reconstruction [@tosat96], growth of the surface [@tosat95], surface diffusion [@tosat95; @sieradz95], adsorption [@kaseno97], phase transitions in two component crystal [@beijeren95], etc.
In this paper we introduce a simple SOS model for bcc(111) surface covered by one physical monolayer of adsorbed atoms to describe surface reconstruction and phase transitions. Using Monte Carlo simulation we study change of the surface structure during the heating and cooling processes, phase transitions to a faceted phases, phase diagram, LEED patterns and temperature dependence of diffracted intensity calculated in kinematic approximation.
The SOS model {#s2}
=============
To study faceting in bimetalic system one needs to know interaction potentials. There are many-body potential for Mo-Mo and W-W interactions derived by Finnis and Sinclair [@finnis] but to our knowledge there is only Ni-Mo many-body potential [@zhang98] for interaction between (Mo, W) and fcc metals. However in Ni/Mo(111) system the faceting does not occur [@mad96]. Therefore we are going to study overlayer induced faceting using a simple solid-on-solid (SOS) model which describe a surface formation energy in agreement with result of the first principle calculation.
In order to construct a model of overlayer-induced faceting on bcc(111) surface (BCCSOS) we are taking into account some experimental evidences. There is a critical coverage, approximately equal to 1PML, to induce faceting on W(111) and Mo(111) [@mad94]. When the coverage exceeds 1PML, extra adsorbate atoms form 3D clusters on the top of faceted surfaces. Hence to simplify problem it will be assumed that coverage is constant and equal to 1PML. Such coverage is equivalent to 3, 2, and 1 geometric monolayers of adsorbate on bcc (111), (211), and (011) surfaces, respectively. A second assumption is that substrate atoms (B) can take position in discrete bcc lattice. It has been shown [@mad99b], in the case of Pd/W(111), that both $\{011\}$ facets and $\{112\}$ facets of W are covered by pseudomorphic monolayer of Pd. Therefore we assume that adsorbate atoms (A) are also located in positions of the bcc structure. There are three types of interaction energies in this SOS model: adsorbate-adsorbate interaction $\epsilon^{AA}_{i}$, adsorbate-substrate interaction $ \epsilon^{AB}_{i}$, and substrate-substrate interaction $\epsilon^{BB}_{i}$, with the range of interaction up to forth neighbors (i =1, 2, 3, 4).
![ Schematic view of the bcc(111) surface. Sites of the sublattices $l=0, 1, 2$ are denoted by squares, triangles, crosses, respectively. The Z axis is normal to the plane and $a_x=a\sqrt{2}/2$, $a_y=a\sqrt{6}/6$.[]{data-label="figure1"}](fig1){width="6cm"}
It is convenient to chose the coordinate system with the $Z$ axis parallel to $[1 1 1]$ direction and $X$, $Y$ axes along $[\bar{1} 0 1]$, $[1 \bar{2} 1]$ directions in the (111) plane. Atoms along the closed packed direction ( parallel to the z axis) form columns which positions in the (x, y) plane are described by three triangular sublattices $l=0, 1, 2$ shown in Fig. \[figure1\]. Column height $z_i$ at site [*i*]{} in the [*l*]{}th sublattice is defined as $z_i = h_i a\sqrt{3}/6$ where [*a*]{} is bcc lattice constant, $h_i=3 n_i +l$ and $n_i$ is the number of atoms in [*i*]{}th column. The assumption on constant coverage equal to 1PML means that there is exactly one A atom in each column placed on its top. As we are going to study reconstruction only, not desorption or crystal melting, the typical restriction of SOS models on columns heights will be assumed. The $h_i$ difference between nearest neighbor sites are forced to be $\pm 1, \pm 2$.
Let us define the Hamiltonian as surface formation energy of A material on B (111) surface $$\label{esurf}
H =E_{A/B}(N_A, N_B) -E^{bulk}_{A}(N_A)-E^{bulk}_{B}(N_B),$$ where $E_{A/B}(N_A, N_B)$ denotes energy of $N_A$ atoms A on $bcc(111)$ surface consisting of $N_B$ atoms B, $E^{bulk}_{A}(N_A)$, $E^{bulk}_{B}(N_B)$ are energies of A, B atoms in their bulk environments. Now taking into account above assumptions we can express Hamiltonian Eq. (\[esurf\]) in terms of column heights $h_i$ in the following form
$$\label{hsos}
\begin{array}{lll}
H &=&
\frac{1}{2}\sum\limits_i
\left\{
\sum\limits_{j_1}
{\left[
J_1 \delta
\left(
\left| h_i - h_{j_1 } \right| - 1
\right)
+ K_1 \delta
\left(
\left| h_i - h_{j_1 }\right|-2\right)
\right]
} \right.\\[3ex]
& &
+ \sum\limits_{j_2}
{\left[
2 J_2 \delta
\left(
\left| h_i - h_{j_2 }\right|
\right) + \left( 2 J_2 + K_2\right) \delta
\left(
\left| h_i - h_{j_2 } \right| - 3
\right)
\right] } \\
& &
+
\left.
J_2 \sum\limits_{j_3}
{\left[
\delta
\left( \left| h_i - h_{j_3 } \right| - 2 \right)
+ \delta
\left(
\left| h_i - h_{j_3}\right| - 4
\right)
\right]}
\right\}+ N J_0,
\end{array}$$
where
$$\begin{aligned}
J_0 &=& -\frac{1}{2} \epsilon^{BB}_{1}+ \epsilon_{1}, \\
J_1 &=& -\frac{1}{2} ( \epsilon^{BB}_{2} + \epsilon^{BB}_{3}+
2\epsilon^{BB}_{4} )+ \epsilon_{2} + \epsilon_{3}+ \epsilon_{4}, \\
K_1 &=& -\frac{1}{2} ( \epsilon^{BB}_{1} + 2\epsilon^{BB}_{3}+
\epsilon^{BB}_{4} )+ \epsilon_{1} + \epsilon_{3}+ \epsilon_{4}, \\
J_2 &=& -\frac{1}{2} \epsilon^{BB}_{4}+ \epsilon_{4}, \\
K_2 &=& -\frac{1}{2} \epsilon^{BB}_{3} + \epsilon_{3}-
\epsilon_{4}, \\
\epsilon_{i}&=& \epsilon^{AB}_{i}- \epsilon^{AA}_{i},\end{aligned}$$
and sums over $j_1$, $j_2$, and $j_3$ denotes summing over first, second, and third neighbors of a column at site $i$. In what follows, we will treat $J_1$, $J_2$, $K_1$, and $K_2$ as model parameters. It is worth noting that this BCCSOS Hamiltonian of overlayer-induced faceting at constant coverage can also be used to study a clean bcc(111) surface by setting $\epsilon^{AB}_{i}=\epsilon^{AA}_{i}=\epsilon^{BB}_{i}$.
Energies of (111), (211), and (110) faces at T=0 {#s2_stab}
-------------------------------------------------
First, we will check the stability of ideal surfaces (111), (211), and (110) covered with 3, 2, 1 geometric monolayers of A, similarly as it was performed in first principles calculations [@leung97]. This allows us to estimate values of model parameters for reconstruction, from (111) surface to $\{112\}$ or to $\{011\}$ faceted surface, under assumption that edge energies are neglected.
Using the Hamiltonian Eq. (\[hsos\]) we get the following expressions for surface energy per site for ideal face of orientation $(hkl)$
$ \displaystyle E_{111}= E_{r} + 7 J_2$,
$ \displaystyle E_{110}= E_{111}+ 2 J_2 + 2 K_2$,
$ \displaystyle E_{211}= E_{111}+ K_2$,
where
$ \displaystyle E_{r}= J_0 + K_1 + 2 J_1$.
Thus, the stability conditions are:
$K_2 \geq 0$, $ J_2 +K_2 \geq 0 \;\;$ for (111) surface,
$ J_2 +K_2 \leq 0$, $ 2 J_2 +K_2 \leq 0\;\;$ for (110) surface,
$K_2 \leq 0$, $ 2 J_2 +K_2 \geq 0\;\;$ for (211) surface.
It easy to see that transformation from (111) surface to {211} facets will be possible for negative values of $K_2$ and appropriate positive $J_2$. To satisfy these conditions, the range of interactions $\epsilon^{\alpha \beta}_{i}$ should be not smaller than $r_4=a \sqrt{11}/2$. On the other hand, it is easy to ensure stability of bcc(111) surface at zero coverage ( $\epsilon_{i} =0$) by choosing $\epsilon^{BB}_{3}<0$ and $\epsilon^{BB}_{4}<0$. Thus, there is general possibility to chose a set of interactions $\{\epsilon^{BB}_{3},
\epsilon^{BB}_{4}, \epsilon_{3}, \epsilon_{4}\}$ in such a way that at zero coverage the bcc(111) surface is stable whereas for coverage equal to one physical monolayer the bcc(211) surface become stable.
It is worth noting that surface energies per site $E_{hkl}$ are related to surface energies per surface atom $\sigma_{hkl}$(used in first principles calculation [@mad99b; @leung97]) in the following way $ \sigma_{111}= 3 E_{111}$, $ \sigma_{211}=2
E_{211}$, $\sigma_{110}=E_{110}$, as the number of surface atoms is $N/3$, $N/2$, and $N$, respectively. Moreover, above stability conditions correspond to faceting conditions [@mad99b; @leung97] expressed in terms of scaled $\sigma_{hkl}$, e.g., the conditions $E_{211} \leq E_{111}$, $E_{211} \leq E_{110}$ are equivalent to $3\sigma_{211}/2\leq
\sigma_{111}$, $3\sigma_{211}/2\leq 3\sigma_{110}$ for transformation $(111)\rightarrow \{112\}$.
Independence of stability condition on parameters $J_0$, $J_1$, and $K_1$ comes from assumptions on constant number of adsorbate atoms A and the restriction on column heights. Therefore we neglect these parameters in further analysis what corresponds to shifting of the surface energy by $-E_r$. Moreover, by choosing $J_2$ as the unit of energy we will work with dimensionless quantities: energy $\tilde H = H/J_2$, temperature $\tilde T =
k_BT/J_2$, and parameter $\tilde K = K_2/J_2$ (in what follows, the tilde will be omitted).
Simulation of faceting {#s3}
======================
We will discus here results of the annealing process investigated by Monte Carlo (MC) simulation in canonical ensemble. Let us start with some details concerning the simulation method. Configurations of BCCSOS model are generated via the classic Metropolis algorithm. A new configuration in Markov chain is generated from the previous configuration by moving one B atom from a site *i* to a site *j*, what is equivalent to the following change of heights: $(h_i, h_j)\rightarrow (h_i-3,
h_j+3)$. It is important to note that such change of heights is possible only if $h_i$ is the local maximum ( with respect to the nearest neighbors) and $h_j$ is the local minimum because of the restriction on column heights. Let us remember one of the model assumption that adsorbate atoms always stay on the column tops as we study reconstruction of B surface covered by the physical monolayer of A material.
![ Temperature dependence of the surface energy during the warming up cycles (triangles) and the cooling cycles (circles). []{data-label="figure2"}](fig2){width="7cm"}
MC simulation were carried out on rectangular lattice of linear size $L_x, L_y$ along x, y direction, respectively with periodic boundary conditions (PBC). To check the role of PBC additional simulations were performed with boundary atoms fixed at positions of the flat bcc(111) surface. The results are consistent with those obtained for calculation with PBC.
First, we investigate properties of the system with coupling constant $K=-1.25$ during warming up process. At each temperature $T_i=T_{i-1} + \Delta T$ system spends the same annealing time $\tau$ measured in Monte Carlo steps per site. The ideal bcc(111) surface is used as a starting configuration at $T_0$. We also investigate properties of the system during cooling down the sample. Results of simulations for the following parameters: $Lx=192$, $Ly=336$, $\Delta T=0.03$ and $\tau=
3\times 10^5$ are presented in Fig. \[figure2\]-Fig. \[figure5\]. Let us notice that these sizes of the lattice correspond to the area of $426\mbox{\AA} \times 431\mbox{\AA}$ on the bcc(111) surface with the lattice constant $a=3.14\mbox{\AA}$.
The surface formation energy decreases as temperature is elevated (see Fig. \[figure2\]) what indicates reconstruction of the surface. At $T\approx 0.4$ very small 3-sided {211} pyramids are formed on the surface and further warming up causes growth of pyramids sizes (see Fig. \[figure3\]). By the {211} pyramid we mean the pyramid built of (211), (121), and (112) facets.
![Snapshot of the surface at T=0.72 on warming.[]{data-label="figure3"}](fig3){width="8cm"}
A big jump of the surface energy is observed at $T= 1.26$ where the phase transition to disordered phase occurs. On the other hand, discontinuous change of the surface energy is observed at $T= 1.20$ when the system is cooling down. The hysteresis loop seen in Fig. \[figure2\] indicates that the phase transition is of first order. The surface is undercooled or overheated because system rest in metastable state separated from the stable state by the free energy barrier. The hysteresis effects, dependence of temperature of the phase transition on cooling and warming, has been observed in LEED experiment [@song95] for Pd/Mo(111). When system is cooling down we observe formation of large pyramids with defected faces and edges just below the temperature of the phase transition (see Fig. \[figure4\]). Facets of similar shapes and sizes were observed at this temperature on warming. During further cooling defects disappear and pyramids take nearly ideal shapes at $T\approx 1$. Therefore the dependence of surface energy on cooling is very week in the ordered faceted phase.
![Snapshot of the surface at T=1.17 on cooling. []{data-label="figure4"}](fig4){width="8cm"}
It is very interesting that in the disordered phase the surface is not planar as suggested in experimental works [@song95]. The structure of disordered phase contains randomly distributed small facets mainly of $\{112\}$ orientations what is shown in Fig. \[figure5\] where the cross section of the disordered phase is compared to cross sections of faceted {211} and ideal bcc(111) surfaces. Moreover, we calculated the average number of atoms on a facet of orientation (211) which in the disordered phase is nearly 10. So we will call this phase as disordered faceted phase (DFP). In very narrow temperature range, we observed a coexistence of a faceted $\{112\}$ and DFP, e.g., in simulation at $T=1.23$ denoted by the triangle in Fig. \[figure2\]. A surface structure in this case looks like surface of ordered phase where large pyramids are removed and on their place DFP is present. On the other hand, there are 3-sided pyramidal holes built of large facets of $\{112\}$ orientation. Sometimes, DFP is located also at the bottom of these holes. It is worth noting that coexistence of both faceted and disordered phases has been observed [@song95] in LEED experiment in Pd/Mo(111) system. However, it was suggested that coexistence is due to the inhomogeneity of the surface.
![ Plot of column height along the y direction for the surface: (a) below the phase transition, T=1.17, (b) above the phase transition, T=1.50, (c) ideal bc(111) - initial configuration. []{data-label="figure5"}](fig5){width="6cm"}
One of quantities measured in MC simulation is the square mean width of the surface, $${\delta h}^2 =
\left<
\frac{1}{N}
\sum\limits_{j}
{
\left( h_j - \left<{h}\right>\right)^2
}
\right>.$$ Behaviour of the $\delta h^2$ during warming up and cooling down processes allows us to study dependence of mean vertical sizes of pyramids on temperature. In Fig. \[figure6\] we present results of simulation for 3 different annealing times $\tau=
2\times 10^3,\;2\times 10^4,\;3\times 10^5$. In all cases the hysteresis loop is present near the phase transition and for shorter time a larger hysteresis is observed. It is easy to explain because it is more probable to overcome the free energy barrier in longer time. In the warming up process the size of pyramids does not depends on the annealing time $\tau$ up to $T\approx 0.5$. When temperature approaches $T\approx 1$ then a rapid growth is observed and for longer annealing time $\tau$ greater pyramids are formed on the surface. Pyramids reach maximal sizes just below the temperature of the phase transition where dependence of sizes on annealing time is the strongest. It worth noting that dependence of facets size on temperature qualitatively agrees with the experimental results [@mad94].
![ Temperature dependence of the $\delta h^2$ during the heating and the cooling cycles for 3 different annealing times. Each label contains symbols for heating, cooling and the number of MC steps[]{data-label="figure6"}](fig6){width="7cm"}
The value of $\delta h^2$ rapidly decreases above the phase transition where very weak dependence on temperature is observed. In the cooling down process large pyramids are formed just below the phase transition temperature, and further decreasing of temperature practically does not change the size of pyramids.
We check dependence of surface growth on the initial temperature $T_0$ and we find that the behaviour of $\delta h^2$ near the transition point is very similar to that in Fig. \[figure6\] for process which starts at $T_0=0.3$ or $T_0=1.0$. Finally we analyze the influence of the temperature increment, $\Delta T$, on the surface energy and the $\delta h^2$ near the phase transition for $\tau=3\times 10^5$. We found that for $\Delta
T=0.01$ the hysteresis loop slightly increases, there are more points where the faceted and disordered phases coexist but it seems that sizes of pyramids do not depend on the temperature increment.
We observe that the growth of the facets on warming base on formation of a larger pyramid from several smaller pyramids. Probability of such rebuilding of the surface is very small at low temperature. This is reflected in the acceptance ratio $f$, a fraction of accepted configurations in the Markov chain, which is very low at low temperatures, $f \approx 0.002$ below $T=0.3$ and $f=0.01$ at $T=0.7$. On the other hand, the acceptance ratio is of the same order for $T=1.0$, $f=0.1$, and close to the phase transition, $f=0.2$, but we observe difference in facets sizes at these temperatures. Looking at snapshots of the surface after annealing at temperature up to $T\approx 1.1$ on warming, we found that pyramids have nearly ideal shapes (see for example Fig. \[figure3\]). We think that all above observation can be explained in the following way. The system is passing through metastable states - local minima of the free energy - as temperature is increased. There are barriers between local minima hence the growth of facets depends on the annealing time.
Phase diagram {#s4}
=============
The stability analysis of flat (111), (211), and (110) surfaces in Sec. \[s2\_stab\] shows that at $T=0$ the (211) surface is stable for $-2\leq K<0$ whereas for $K<-2$ the (110) surface is stable. Whether these findings are generally correct we can check constructing phase diagram in the (T, K) plane by using MC simulations. Most of simulations were performed at constant K varying temperature. To check the results some simulations were performed at constant temperature. To identify different phases we calculated masses of clusters of different facet orientations (111), $\{110\}$, and $\{112\}$. A cluster of orientation (hkl) is defined in similar way as percolation cluster on the (hkl) plane. We also recorded equilibrium configurations at each investigated point of the (T, K) plane. We found that for $K<0$ the phase diagram ( see Fig. \[figure7\]) contains three phases
1. Faceted $\{112\}$ phase consisting of 3-sided $\{112\}$ pyramids.
2. Faceted $\{110\}$ phase consisting of 3-sided $\{110\}$ pyramids.
3. Disordered faceted phase.
The surface of the faceted $\{110\}$ phase is built of (110), (101), and (011) facets which form 3-sided pyramids on the bcc(111) surface. Such surface has been observed experimentally in Pd on Ta(111) system [@szuk02]. The phase transition between the disordered phase and one of ordered is of first order as surface energy, the square mean width of the surface $\delta
h^2$, and the structure factor (see Sec. \[s5\]) change discontinuously at the transition point.
![ Phase diagram: T versus K. []{data-label="figure7"}](fig7){width="7cm"}
The transition temperature between F{211} and disordered phase is bounded from above by maximum $T_c=1.20$ at $K=-1.50$. If we chose this point to estimate model parameters for Pd/Mo(111) with $T_c=850$ K we get $J_2=0.06$ eV, and $K_2=-0.09$ eV. On the other hand the transition temperature from F{110} to the disordered phases is increasing function of $|K|$ and it might reach large value. The present model does not include the desorption, thus we can expect that transition temperature might be above the desorption temperature in some systems especially where the faceted F{110} phase occurs. In such cases the phase transition to the disordered faceted phase could not be observed but rather deconstruction induced by desorption would be expected.
There is difficult to study phase transition between ordered phases using MC simulation because for $K\approx -2$ we observed coexistence of two types of facets $\{011\}$ and $\{112\}$. The mixed phase does not contain separated pyramids of two types but mixture of facets, e.g., on the (211) facet a smaller (011) facet can occur.
Simulation of LEED patterns {#s5}
===========================
Low energy electron diffraction (LEED) experiments play an important role in investigation of structures and phase transitions of faceted surfaces [@mad94; @song95]. Very recently, Song et al. [@song95] where able to study LEED patterns of Pd/Mo(111) system in high temperatures up to 1200K. They demonstrated the existence of reversible planar/faceted phase transition with transition temperature different for heating cycles ($T\approx 870K$) and for cooling cycles ($T\approx 830K$).
![ LEED patterns of (a) disordered phase where spots are represented by squares and also spots of ideal bcc(111) surface (crosses) are shown, (b) faceted {211} phase where circles, diamonds, and triangles denote spots coming from diffraction on facets (121), (211), and (112), respectively. []{data-label="figure8"}](fig8){width="6cm"}
In this section we employee the BCCSOS model to analyze the temperature dependence of LEED patterns. The diffracted intensity is proportional, in the kinematic approximation, to the structure factor
$$S({\bf k}) =
{\left<
\left|
\sum\limits_{j}
{
\exp(i {\bf k R}_j)
}
\right|^2
\right>},$$
where summing is over surface atoms, ${\bf k}= {\bf k}_f-{\bf
k}_i$, and ${\bf k}_i$, ${\bf k}_f$ is the wave vector of incident electron, scattered electron, respectively. Position of surface atoms ${\bf R}_j$ is expressed by column height $h_{r_j}$ in site ${r_j}=(x_j, y_j)$ in the following way ${\bf R}_j= a(3 \sqrt{2} x_j, \sqrt{6} y_j, \sqrt{3} h_{r_j})/6$. In a SOS model not all atoms from column tops are surface atoms. Therefore in calculation of diffracted intensity sometimes the shadowing factor is used [@tosat96]. In this paper we consider the contribution to $S$ only from surface atoms. As a surface atom in any configuration we choose the atom from the top of column which has at least four lower nearest neighbors. In the case when number of higher columns is equal to 3 then atom is regarded as surface atom if is not surrounded by the higher neighbors. Such case occurs for atoms on {110} facets. We will calculate structure factor only for a set of ${\bf k}_f$, chosen in such a way to describe diffraction on faces (111) and {211}. For each of these faces we construct a pair of inverse lattice vectors ${\bf g}_1, {\bf g}_2$ parallel to the face, e.g., for (121) ${\bf g}_1=2\pi/a(\sqrt{2}/2, 0, 0)$ and ${\bf
g}_2=2\pi/a(0, 4\sqrt{6}/9, 2\sqrt{3}/9)$. Then, the wave vectors of scattered electrons, ${\bf k_f}$, for given face are chosen to satisfy condition of constructive interference $({\bf
k}_f-{\bf k}_i)_{\parallel} =l{\bf g}_1+m{\bf g}_2 $, where $l,
m=0, \pm 1, \ldots$ and the symbol $\parallel$ denotes vector component parallel to the face.
We calculated structure factors for the wave vector of the incident electron, $ k_i =2\pi/a( 0, 0, 1.43)$, normal to the bcc(111) surface. Thus energy of incident electrons is equal to 31 eV for $a=3.14\mbox{\AA}$, similarly as in LEED experiment for Pd/Mo(111) [@song95]. Figure \[figure8\] shows LEED patterns obtained in computer simulation for the model parameter $K=-1.25$. In the ordered phase {211} we see besides 3 clovers (groups of 3 spots) observed experimentally, the forth clover in the center of the image, and six single spots more distant from the center. The pattern of disordered phase consists of three spots which positions correspond to the centers of external clovers in the diffracted pattern of the faceted phase. Apparent lack of spots from bcc(111) surface confirms that disordered phase has not the planar structure.
![ Temperature dependence of the structure factor for the wave vector corresponding to a spot of the disordered phase on warming (triangles) and on cooling (squares). The hysteresis loop for large lattice $Lx=192$, $Ly=336$, is shown in the inset. []{data-label="figure9"}](fig9){width="7cm"}
Temperature dependence of the structure factors were calculated using MC simulation and changing temperature in the way described in Sec. \[s3\]. Most of simulations were performed on the lattice with $Lx=90$, $Ly=132$. The spots of disordered phase have the same diffracted intensities. Their temperature dependence (see Fig. \[figure9\]) on warming and on cooling shows the presence of hysteresis loop in behaviour of $S(T)$ close to the phase transition what is in agreement with experimental results [@song95]. The hysteresis loop calculated for large lattice with $Lx=192$, $Ly=336$ is shown in the inset in Fig. \[figure9\]. There are several points where the structure factor takes intermediate values because at these temperatures coexistence of faceted and disordered phases has been observed. More coexistence points where found on warming than on cooling. We observe also discontinuity of the structure factor at the transition temperature. Thus hysteresis and discontinuity of the structure factors are another argument for the first-order phase transition between faceted and disordered phases. During the process of warming up the system the S decreases with temperature and becomes very small when on the surface large enough {211} facets are formed.
![ Temperature dependence of the structure factor for the wave vector corresponding to a leaf of the right clover on warming (triangles) and on cooling (squares). []{data-label="figure10"}](fig10){width="7cm"}
The diffracted intensities of a spot in the ordered phase was calculated for ${\bf k}_f$ fulfilling the condition of constructive interference and also for vectors from a very small surroundings of ${\bf k}_f$. Then maximal value of S were chosen to represent the intensity of spot at this ${\bf k}_f$. We use such procedure because the surface does not consist of ideal flat {211} facets during warming up and cooling processes, therefore spots can change shapes, intensities, and positions. The diffracted intensities of spots of faceted phase are different because areas of facets of different orientation are not equal. Moreover, their dependence on temperature during warming the system shows irregular behaviour (see Fig. \[figure10\]) because of facets grow i.e., from several small facets a bigger one arises. We observe rapid growth of S at temperature $T\approx 0.9$ when large pyramids are formed on the surface. During the cooling cycles, the S reaches nearly maximal value just below transition temperature what confirms formation of large {211} facets. The maximal values of S in the faceted phase are nearly ten times smaller than in the disordered phase because in former case only $\sim 1/3$ surface atoms contribute to S, i.e., surface atoms placed on facets of the same orientation.
Discussion
==========
We have presented here the BCCSOS model to study overlayer induced faceting. Although it might seem that model is very simple it gives many results in agreement with experiments. It has been shown that bcc(111) surface covered by a physical monolayer undergoes reconstruction upon annealing. We obtained two types of reconstructed surfaces: faceted {211} covered by 3-sided pyramids as observed in Pd/Mo(111) [@mad96] and faceted {011} surface found in Pd/Ta(111) [@szuk02]. The sizes of facets depend on annealing temperature and annealing time. At high temperature the phase transition to disordered phase occurs. We have shown that LEED diffraction patterns can be calculated, in kinematic approximation, in warming up and cooling down processes for different electron energies. In dependence of diffracted intensities on temperatures the hysteresis effect was observed close to temperature of the phase transition.
On the other hand, we can investigate some properties not reported in experimental papers. First of all the phase transition can be studied in details using this model. It is shown that the phase transition is of first order because quantities such as surface energy, the structure factor change discontinuously as temperature reaches the transition point. Moreover the hysteresis effects are observed when system is warming up and then cooling down. We can analyze the structure of the surface in disordered phase as well as in the coexistence of the faceted and disorder phase close to the transition temperature. We found that the disordered phase consists of many small {211} facets randomly distributed. Hence it is not planar bcc(111) surface. It would be interesting to check experimentally this prediction. It was suggested [@song95] that coexistence of faceted and disordered phases is due to the inhomogeneity of the adsorbate coverage but we have demonstrated that the coexistence could also exist for homogeneous coverage.
It seems that this model can be applied to study reconstruction of curved bcc surfaces where reconstructed surface has different structure than faceted bcc(111) surface. In case of Pd deposited on a needle-shaped tungsten [@sczep02], step like {211} microfacets has been observed. Using the present model we have not obtained the faceted phase with large {211} pyramids and small {011} pyramids as observed in Pd/W(111) [@mad99b] for prolonged annealing time. This might be due to lack of density fluctuation. So we are going to extend the model including dependence on coverage. The most general extension should base on replacing of the interactions $\epsilon^{\alpha \beta}_{i}$ by many-body potentials. To construct such potentials, e.g., for Mo-Pd, W-Pd systems, the results of the first-principle calculations could be used [@zhang98].
Acknowledgements {#acknowledgements .unnumbered}
----------------
I would like to thank prof. Jan Ko[ł]{}aczkiewicz and dr Andrzej Szczepkowicz for discussions. This work was supported by the Polish State Committee of Scientific Research (KBN) Grant No 2 P0 3B 107 19.
[99]{}
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The tracer equations are part of the primitive equations used in ocean modeling and describe the transport of tracers, such as temperature, salinity or chemicals, in the ocean. Depending on the number of tracers considered, several equations may be added to and coupled to the dynamics system. In many relevant situations, the time-step requirements of explicit methods imposed by the transport and mixing in the vertical direction are more restrictive than those for the horizontal, and this may cause the need to use very small time steps if a fully explicit method is employed. To overcome this issue, we propose an exponential time differencing (ETD) solver where the vertical terms (transport and diffusion) are treated with a matrix exponential, whereas the horizontal terms are dealt with in an explicit way. We investigate numerically the computational speed-ups that can be obtained over other semi-implicit methods, and we analyze the advantages of the method in the case of multiple tracers.'
author:
- 'S. Calandrini[^1]'
- 'K. Pieper[^2]'
- 'M. D. Gunzburger$^{*}$'
bibliography:
- 'references.bib'
title: '**[Exponential Time Differencing for the Tracer Equations Appearing in Primitive Equation Ocean Models]{}**'
---
Introduction
============
The primitive equations are the constitutive system of equations in ocean modeling. This system is composed of a momentum equation, an equation for the ocean thickness, equations for the transport of tracers, as well as an equation of state. The momentum and thickness equations describe the dynamics, i.e. the change in time of the velocity and thickness of the water. The tracer equation describes the transport of tracers, such as temperature, salinity or chemicals. Depending on the number of tracers considered, several equations may be added to and coupled to the dynamics system. The coupling between the tracers and the dynamics depends on the nature of tracers. Temperature and salinity impact the density of the layers of ocean water, influencing the dynamics, and for this reason they are called active tracers. Hence, between the dynamics and active tracers there is a two-way coupling. However, for most tracers the coupling is only one-way, with such tracers being called passive [@siberlin2011oceanic].
To this day, the numerical solution of the primitive equations remains a challenging task. One of the main issues is the presence of multiple time-scales, where different processes (e.g., external and internal gravity waves, eddies, biochemical reactions) take different times to be completed. Since the primitive equations arise from hyperbolic conservation laws ([@chiodaroli2017existence; @oliger1978theoretical]), explicit time integrators and Runge-Kutta schemes would hypothetically be good choices for solving it. However, these schemes are not capable of efficiently handling multiple time-scales, because the time-steps restrictions are too severe, resulting in a significant degradation of performance. To better handle multiple time-scales, methods have been developed with the help of mathematical and algorithmic techniques such as splitting strategies and semi-implicit approaches. Due to the different properties of the dynamics and the tracer equations, schemes have been created to separately deal with the two subsystems ([@ringler2013multi; @smith2010parallel]). Methods for the dynamics need to take into account the different wave speeds of the model, with the goal of having a model for which time step sizes are governed by the slow wave speeds or the speed of advection and not by the fast wave speeds as is the case for standard explicit schemes. Examples of time-stepping schemes for the dynamics include implicit ([@dukowicz1994implicit]) and split-explicit ([@higdon2005two; @ringler2013multi]) methods. More recently, exponential time differencing (ETD) methods, also known as exponential integrators, have gained attention in the ocean modeling community due to their stability properties that allow time steps considerably larger than those dictated by the CFL condition. In [@pieper2019exponential], an ETD scheme has been developed for the rotating shallow water equations with multiple horizontal layers, which correspond to a vertical discretization of the primitive equations dynamics in an isopycnal vertical coordinate system. The main idea behind exponential integrators is a splitting of the right-hand side term of an equation into a linear part and a remainder, i.e. $$\partial_t \theta = F(\theta) = A \theta + R(\theta)\;,$$ with an appropriate choice of the linear operator $A$. For a review of exponential integrators we refer to [@hochbruck2010exponential].
In this work, we devise an ETD method for the tracer equation. A tracer is supposed to satisfy a conventional advection-diffusion equation of the form, $$\label{conv_diff}
\partial_t \theta + \nabla_{\mathbf{x}}\cdot(\mathbf{u} \theta) + \mathcal{D}_{\mathbf{x}} \theta = q(\theta)$$ where $\theta$ is the tracer studied, $\mathbf{u} = (u, w) \in R^3$ is the velocity of water, which is usually split in to the horizontal velocity $u \in R^2$ and the vertical velocity $w$, $\mathcal{D}$ is a diffusion term, and $q(\theta)$ represents interior sources or sinks. The equation has to be solved on a three-dimensional domain, split into a two-dimensional horizontal and a vertical coordinate $\mathbf{x} = (x,z)$, and $\nabla_{\mathbf{x}} = (\nabla, \partial_z)$. We note that, the tracer equation can usually also be written as $$\label{conv_diff_split}
\partial_t \theta + \nabla_x \cdot(u \theta) + \mathcal{D}_x \theta + \partial_z(w \theta)
- \partial_z (\kappa_z \partial_z \theta) = q(\theta),$$ with the horizontal diffusion $\mathcal{D}_x$, and the vertical diffusion coefficient $\kappa_z$. To avoid problems with shocks, the velocity field $\mathbf{u}$ is usually chosen to be divergence free and tangential to the boundary.
When dealing with the tracer equation, a key point is appropriately including vertical mixing. Tracer vertical mixing usually occurs on small time-scales and can be induced by density differences and/or by turbulent motions. For explicit time stepping schemes, the time step requirement is usually set by the horizontal advective CFL condition, hence very small time steps may need to be used with explicit time stepping methods to include realistic vertical mixing of tracers. To avoid this issue, in the time-stepping schemes used by popular ocean models, the vertical diffusion term $\partial_z (\kappa_z \partial_z \theta)$ is treated implicitly. In POP [@smith2010parallel], the vertical tracer diffusion term is treated with an implicit Euler algorithm, whereas the remaining terms of the equation are treated with a leapfrog algorithm. In MPAS-Ocean [@ringler2013multi], tracer equations are stepped forward with the mid-time velocity values and this process is repeated in a predictor-corrector way. Implicit vertical mixing of tracers completes each time-step, where, as in POP, the vertical tracer diffusion term is treated with an implicit Euler algorithm. Another important phenomenon to deal with, besides vertical mixing, is when an inflow of cold water near the coast leads to cold water on top, scenario that creates density and pressure variations and downward motion. In an isopycnal configuration, there is no vertical transport, however, if a mostly Eulerian coordinate system is employed (z-star, z-level), rapid variations in the pressure/density may induce a fast vertical flow. This phenomenon is amplified by the usage of fine meshes in the vertical, with much smaller vertical than horizontal spacing, for example [@ringler2013multi] employs $1$–$15$km horizontal, $10$–$250$m vertical. As we said, many ocean models treat the vertical advection explicitly and so, when a fast transport of water among different layers occurs, the model may be unable to appropriately capture this behavior. Consequently, instabilities in the simulation may occur, causing the need to decrease the time-step. To resolve this issue, we propose an ETD solver where all the vertical terms, i.e. vertical advection and diffusion, are treated with a matrix exponential, whereas the horizontal are dealt with in an explicit way. This means that we are splitting the linear operator $A$ into two parts: $A^z$ that accounts for all vertical terms, and $A^x$ that contains all horizontal terms. $A^x$ is then incorporated in the remainder $R(\theta)$, so the actual linear operator we are working with is $A^z$. This operator splitting has two advantages. First, by treating exponentially terms related to fast time-scales, bigger time steps can be taken, and so computational speed-ups that can be obtained over other explicit methods. Compared to semi-implicit methods, we expect higher accuracy due to an exact treatment of the fast scales. Second, by including only the operator $A^z$ in the exponential, not the whole operator $A$, the computational cost is reduced, effectively lowering the total computational cost and time of the whole method.
Another important challenge to face when dealing with tracer equations, and in general with primitive equations, is in the numbers of tracers considered. The amount of tracers in an ocean simulation is usually around 40, but may increase up to 70, causing a significant computational load. Hence, efficiently solving multiple tracers equations is an important task in an ocean model. When an ETD scheme is used, efficiency corresponds to a low-cost evaluation of $\varphi_k$-functions. For a given matrix $B$, assembling $\varphi_k(B)$ is generally prohibitive in the large scale context; iterative methods are used instead, i.e. Krylov subspace algorithms ([@eiermann2006restarted; @saad1992analysis]).
Finally, our solver is compared with existing ocean models, to make sure that it is able to reproduce similar results under the same physical conditions. To do so, the whole primitive equations are solved and two tests from [@ilicak2012spurious] are performed. The tracer equations is coupled with the dynamics, and to solve this latter system a second ETD solver (Exponential Rosenbrock Euler) is used. The results obtained with our ETD solvers are compared with those obtained with three other codes: MPAS-Ocean, MITgcm and MOM.
The paper is organized as follows. Section \[tracer\_eq\_section\] describes the discretization of the tracer equation both in the vertical and in the horizontal. Section \[ETD\] focuses on exponential integrators and their implementation. For the computation of $\varphi_k$-functions, a restarted Krylov subspace method is described, and a scheme based on scaling and squaring relations is proposed. For the latter scheme, an error analysis is provided. In section \[solver\], the proposed ETD solver for the tracer equation is presented, and the properties of the operator splitting are described. The numerical tests are shown and discussed in section \[tests\], while the conclusions follow in section \[sec:conclusions\].
Discretization of the Tracer Equation {#tracer_eq_section}
=====================================
In this section, the discretization of the tracer equation in the vertical and in the horizontal is described. In most ocean models, an hydrostatic condition is assumed, leading to describing the primitive equation by the incompressible Boussinesq equations in hydrostatic balance. Following this assumption, the tracer equation in continuous form can be re-written as $$\dfrac{\partial (\widetilde{\rho} T)}{\partial t} = - \nabla \cdot (\widetilde{\rho} u T) -
\dfrac{\partial (\widetilde{\rho} T w)}{\partial z} + D^{T}_x + D^{T}_{z} + \mathcal{F^T}\;, \label{tracer}$$ where $T$ is the tracer, $u$ is the horizontal velocity, $w$ is the vertical velocity and $\widetilde{\rho}$ is the pseudo-density. The variable $z$ represents the vertical coordinate and it defined positive upward. $D^{T}_x$ and $D^{T}_{z}$ indicate the horizontal and vertical diffusion term, respectively. These terms are defined as $$\begin{aligned}
D^{T}_x & = \nabla \cdot (h\; \kappa_{x} \nabla T) \;, \\
D^{T}_{z} & = h \frac{\partial}{\partial z} \big(\kappa_{z} \frac{\partial T}{\partial z} \big) \;,\end{aligned}$$ where $\kappa_x$ and $\kappa_{z}$ are the horizontal and vertical diffusion, respectively. Without loss of generality, we assume that no forcing term is present, i.e. $\mathcal{F^T} = 0$. This assumption is equivalent to consider that no external factors have an influence on the tracer behavior.
To discretize equation , we employ z-level coordinates in the vertical [@petersen2015evaluation] and a finite-volume method using a C-grid staggering in the horizontal [@ringler2011exploring]. For more details about the discretization of all primitive equations, please refer to [@ringler2013multi].
Tracer Equation with vertical discretization
--------------------------------------------
As in MPAS-Ocean [@ringler2013multi; @petersen2015evaluation], the vertical coordinate we use is Arbitrary Lagrangian-Eulerian (ALE). With ALE, several coordinate systems can be specified depending on the application. Common choices for the vertical coordinates include z-level, where all layers have a fixed thickness except for the top layer, z-star, where all layer thicknesses vary in proportion to the sea surface height, and isopycnal, where there is no vertical transport between layers.
The tracer equation with vertical discretization is $$\begin{aligned}
& \dfrac{\partial (h_{k} T_k)}{\partial t} = - \nabla \cdot (h_k u_k T_k) -
\overline{T_k} w_k + \overline{T_{k+1}} w_{k+1} + [D^{T}_x]_k + [D^{T}_z]_k \;, \label{tracer_vert_discr}\end{aligned}$$ where $k$ indicates the vertical layer, $k = 1$ is the top layer and $k$ increases downward up to $N$; $z = 0$ is the mean elevation of the free surface, and the $z$ coordinate is positive upward. The variable $w_k$ indicates the transport of fluid from layer $k$ to $k-1$, i.e. across the top interface of layer $k$. The pseudo-density $\widetilde{\rho}$ has been replaced by the $h$, which is the layer-thickness. The operator $\overline{(\cdot)}$, on a generic variable $\psi_k$, is the vertical average between the layer $k$ and the above layer $k-1$, i.e., $$\overline{\psi_k} = \dfrac{\psi_{k-1} + \psi_k}{2}\;.$$ Finally, $[D^{T}_x]_k$ and $[D^{T}_{z}]_k$ indicate the discretized horizontal and vertical tracer diffusion terms, respectively, and are defined as $$[D^{T}_x]_k = \nabla \cdot (h_k \kappa_{x} \nabla T_k)\;, \quad \qquad
[D^{T}_{z}]_k = h_k \underline{\delta z_k}(\kappa_{z} \overline{\delta z_k}(T_k))\;.$$ The discrete operators $\overline{\delta z_k}(\cdot)$ and $\underline{\delta z_k}(\cdot)$, on a generic variable $\psi_k$, are defined as $$\begin{aligned}
\overline{\delta z_k} (\psi_k) = \dfrac{\psi_{k-1} - \psi_k}{{\overline{h_k}}}\;, \\
\underline{\delta z_k} (\psi_k) = \dfrac{\psi_{k} - \psi_{k+1}}{h_k}\;.\end{aligned}$$ The vertical tracer diffusion term can actually be rewritten without introducing the operator $\underline{\delta z_k}$ as $[D^{T}_{\nu}]_k = \kappa_{z} (\overline{\delta z_k}(T_k) - \overline{\delta z_k}(T_{k+1}))$.
The choice of the vertical coordinate system is enforced in the computation of the vertical transport. $w_k$ can be found by solving the thickness equation for $w_k$. The thickness equation discretized in the vertical has the form $$\dfrac{\partial h_k}{\partial t} + \nabla \cdot (h_k {\mathbf u}_k) + w_k - w_{k+1} = 0\;. \label{thickness}$$ For a Boussinesq fluid, this equation represents the continuity equation for the pseudo-density $\widetilde{\rho}$. To obtain $w_k$ from , all variables at the previous time step must be known, in particular the time derivative of $h$ at layer $k$, $\frac{\partial h_k}{\partial t}$, must be known. For this purpose, a new quantity, named $h_k^{ALE}$, is introduced. $h_k^{ALE}$ represents the desired thickness for the new time, and it is used to compute $\frac{\partial h_k}{\partial t}$ using a first-order finite difference approximation. In this way, $w_k$ can be found as $$w_k = w_{k+1} - \nabla \cdot (h_k {\mathbf u}_k) - \dfrac{h^{ALE}_k - h_k}{\Delta t}\;. \label{vertical_transport}$$ For isopycnal simulations, the vertical transport is set to zero in . For other coordinate systems, like z-level and z-star, the way $h_k^{ALE}$ is computed determines the type of coordinates chosen. For example, for z-level vertical coordinates $$\begin{aligned}
h_1^{ALE} & = h_1^{rest} + \zeta \;, \\
h_k^{ALE} & = h_k^{rest}, \mbox{ for } k > 1 \;,\end{aligned}$$ where $h_k^{rest}$ is the layer thickness when the ocean is at rest, and $\zeta$ is the sea surface height defined as $\sum_k h_k - \sum_k h_k^{rest}$. For z-level coordinates (and for z-type coordinates in general) the resting thickness is considered constant in each horizontal layer, but for other coordinate systems, like sigma coordinates, $h_k^{rest}$ varies horizontally in proportion to the column’s total depth. The simulations presented in section \[tests\] use z-level vertical coordinates. Please refer to [@petersen2015evaluation] for more details about the computation of $h_k^{ALE}$ for other coordinate systems.
Tracer Equation with horizontal discretization
----------------------------------------------
The horizontal discretization is a C-grid, finite-volume method applied to a spherical centroidal Voronoi tessellation (SCVT) mesh. Height, tracers, pressure and kinetic energy are defined at centers of the convex polygons, and the velocity is located at cell edges. Vorticity (curl of velocity) is defined at cell vertices. In the following, the subscripts $i$ and $e$ indicate the discretized variables through cell centers and edges, respectively. Since we are focusing on the discretization of the tracer equation only, we will not work with variables and operators defined at cell vertices.
The tracer equation with horizontal discretization is $$\begin{aligned}
\dfrac{\partial (h_{k,i} T_{k,i})}{\partial t} = & - [\nabla \cdot \big (\widehat{({h}_{k,: })}_e {\mathbf u}_{k,:} \widehat{({T}_{k,:})}_e \big)]_i
- \overline{T_{k,i}} w_{k,i} + \overline{T_{k+1,i}} w_{k+1,i} \nonumber \\
& + [D^{T}_x]_{k,i} + [D^{T}_{z}]_{k,i} \;, \label{tracer_hor_discr} \\
[D^{T}_x]_{k,i} & = [ \nabla \cdot \big( \widehat{({h}_{k,: })}_e \, \kappa_{x} \, [\nabla T_{k,:}]_e \big) ]_i\;, \quad \qquad
[D^{T}_{z}]_{k,i} = h_{k,i} \underline{\delta z_k}(\kappa_{z} \overline{\delta z_k}(T_{k,i}))\;. \end{aligned}$$ Each variable now has two sub-scripted indices, the first indicating the vertical layer, and the second indicating its position on the horizontal grid, namely either $i$ or $e$. Colons in subscripts may be places as second index to indicate that multiple edges or cell centers are used in computing the horizontal operator. For a generic variable $\psi_k$, the symbol $\widehat{(\psi_{k,:})}_e$ represents the averaging of the variable from two adjacent centers to the corresponding edge. We would like to point out that the vertical transport through the sea surface and at the bottom surface is zero, i.e. $w_{1,i} = 0$ and $w_{N+1,i} = 0$. Moreover, we consider $u_{k,e} = 0$ on all boundary edges.
For a generic vector field ${\mathbf Y}_k$ and variable $\psi_k$, the discrete horizontal operators $[\nabla \cdot {\mathbf Y}_{k,:}]_i$ and $[\nabla \psi_{k,:}]_e$ are defined as $$\begin{aligned}
[\nabla \cdot {\mathbf Y}_{k,:}]_i & = \frac{1}{A_i} \sum_{e \in E(i)} n_{e,i} Y_{k,e} l_e\;, \\
[\nabla \psi_{k,:}]_e & = \frac{1}{d_e} \sum_{i \in C(e)} - n_{e,i}\psi_{k,i} \;.\end{aligned}$$ $A_i$ indicates the Voronoi cell area, $d_e$ is the distance between cell centers, $l_e$ is edge length and $n_{e,i}$ represents the sign of the vector at edge $e$ with respect to cell $i$. The sets $E(i)$ are the edges about cell $i$, and the sets $C(e)$ are the cells neighboring edge $e$. Thus, the divergence moves from edges to cell-centered quantity, while the gradient moves from cell centers to edges.
Exponential Time Integration {#ETD}
============================
This section describes exponential time differencing methods, that later will be used to solve the tracer equation.
Exponential Integrators
-----------------------
Let $\partial_t T = F(T)$ be a system of partial differential equations (PDEs), where $T=T(t)$ denotes the vector of the solution variables for $t \in [t_n, t_{n+1}]$, and $F(T)$ is the right-hand-side term. The interval $[t_n, t_{n+1}]$ refers to one time step. The main idea behind exponential integrators is a splitting of the right-hand-side term into a linear part and a remainder, i.e. $$\partial_t T = F(T) = A_n T + R(T), \label{splitting1}$$ where $A_n$ represents a linear operator, and $R(T):=F(T)-A_n T$ denotes the remainder, which in general is nonlinear. Applying the variation of constants formula to equation , the solution at time $t_{n+1} = t_n + \Delta t$, i.e. $T_{n+1}=T(t_{n+1})$, is obtained as $$T_{n+1} = \exp(\Delta t A_n) T_n + \int_0^{\Delta t} \exp((\Delta t - \tau )A_n) R(T(t_n + \tau)) d\tau\;. \label{variation_of_consts}$$ At this point, to build a concrete exponential integrator, an approximation of $R(T(t_n + \tau))$ must be considered. By substituting $R(T(t_n + \tau))$ with its Taylor expansion truncated at $s$ ($s \in \mathbb{N}$), namely $$R(T(t_n + \tau)) = \sum_{k=1}^{s} \dfrac{\tau^{k-1}}{(k-1)!}
\left.
\dfrac{{\mathop{}\!\mathrm{d}}^{k-1} R(v(t_n + \tau))}{{\mathop{}\!\mathrm{d}}\tau^{k-1}}
\right\rvert_{\tau=0} \,,$$ the solution $T_{n+1}$ can be approximated by $$T_{n+1} \approx \exp(\Delta t A_n) T_n + \sum_{k=1}^s \dfrac{1}{(k-1)!} \Big[ \int_0^{\Delta t} \exp((\Delta t- \tau) A_n)\tau^{k-1}
d\tau \Big]
\left.
\dfrac{{\mathop{}\!\mathrm{d}}^{k-1} R(T(t_n + \tau))}{{\mathop{}\!\mathrm{d}}\tau^{k-1}}
\right\rvert_{\tau=0}.
$$ The above expression can be rewritten as $$T_{n+1} \approx \exp(\Delta t A_n) T_n + \sum_{k=1}^s \Delta t^k \varphi_k(\Delta t A_n)
\left.
\dfrac{{\mathop{}\!\mathrm{d}}^{k-1} R(T(t_n + \tau))}{{\mathop{}\!\mathrm{d}}\tau^{k-1}}
\right\rvert_{\tau=0} \,,
\label{exp_int}$$ by introducing the, so called, $\varphi$-functions defined by $$\begin{aligned}
\varphi_k(\Delta t A_n) & = \dfrac{1}{\Delta t^k(k-1)!} \int_0^{\Delta t} \exp((\Delta t- \tau) A_n)\tau^{k-1} d \tau \label{phi_def1}\\
& = \dfrac{1}{(k-1)!} \int_0^1 \exp((1-\sigma)\Delta t A_n) \sigma^{k-1} d\sigma, \quad k=1,2,\dots,s\;. \label{phi_def2} \end{aligned}$$ By performing the change of variable $\Delta t - \tau = (1 - \sigma)\Delta t$, can be obtained from . For $k=0$, we have that $\varphi_0(\Delta t A_n) = \exp(\Delta t A_n)$.
The parameter $s$ in indicates the number of stages of the method. By taking $s=1$, the exponential Euler method is obtained as $$T_{n+1} \approx \exp(\Delta t A_n) T_n + \Delta t \varphi_1(\Delta t A_n) R(T_n) = T_n + \Delta t \varphi_1(\Delta t A_n) F(T_n)\;, \label{exp_eul}$$ where $\varphi_1(\Delta t A_n) = \int_0^1 \exp((1-\sigma) \Delta t A_n) d\sigma =
(\Delta t A)^{-1}(\exp(\Delta t A_n) - I)$, with $I$ indicating the identity matrix. For a generic linear operator $A_n$, this method is first-order accurate, but if $A_n$ is the Jacobian matrix of the system evaluated at $t_n$, namely $A_n = F'(T_n)$, then the method becomes second-order accurate [@hochbruck2010exponential; @luan2019further]. In this case, the scheme is called Exponential Rosenbrock Euler. By taking $s=2$, a second-order single-step method with two stages is obtained as $$\begin{aligned}
& T_{n}^{(1^{st}\;stage)} = T_n + \Delta t \varphi_1(\Delta t A_n) F(T_n)\;, \label{stage1} \\
& T_{n+1} = T_{n}^{(1^{st}\;stage)} + \Delta t \varphi_2(\Delta t A_n) (R(T_n^{(1^{st}\;stage)}) - R(T_{n}))\;, \end{aligned}$$ where the first derivative of $N(T_n)$ is approximated with a first-order finite-difference approximation $$\left.\dfrac{{\mathop{}\!\mathrm{d}}R(T(t_n + \tau))}{{\mathop{}\!\mathrm{d}}\tau}\right\rvert_{\tau=0}
\approx
\dfrac{R(T_n^{(1^{st}\;stage)}) - R(T_{n})}{\Delta t}\;,$$ and $\varphi_2(\Delta t A_n) = \int_0^1 \exp((1-\sigma)\Delta t A_n) \sigma \, d \sigma
= (\Delta t A_n)^{-2} (\exp(\Delta t A_n) - \Delta t A_n - I)$. This scheme is known as ETD2-RK ([@hochbruck2005explicit; @hochbruck2010exponential]).
Computation of the $\varphi$-functions
--------------------------------------
The computation of the $\varphi$-functions is of major relevance for ETD methods. Let us assume that $A$ is in $\mathbb{R}^{n \times n}$, where $b \in \mathbb{R}^n$ and $k \in \mathbb{N}$ is the index of the $\varphi$-function. Fist of all, let us consider the special case $k=0$, for which we have $\varphi_0(A) = \exp(A)$. The most popular method for computing the matrix exponential $\exp(A)$ is the scaling and squaring algorithm [@moler1978nineteen; @moler2003nineteen]. In [@higham2005scaling], the scaling and squaring algorithm used by the MATLAB function *expm* is described, and a variation of this method that alleviates overscaling problems is presented in [@al2009new]. The Expokit package [@sidje1998expokit] uses a scaling and squaring method as well for computing matrix exponentials. An alternative to scaling and squaring methods is given by Krylov subspace projections [@hochbruck1997krylov; @saad1992analysis]. With Krylov schemes, matrix-vector products of the form $\varphi_k(A) b$ are computed, without the actual construction of the matrix $\varphi_k(A)$. To overcome some memory issues associated with standard Krylov methods, restarted schemes have been developed [@eiermann2006restarted]. The C/C++ library SLEPc [@roman2015slepc] uses the restarted algorithm presented in [@eiermann2006restarted].
In the following, a brief description of the algorithm presented in [@eiermann2006restarted] is given, since this method is used in the numerical results section, and scaling and squaring algorithm for indexes $k \ge 0$ is presented, for which an error estimate is derived.
### Krylov subspace approximation {#restarted_kry}
The Krylov subspace approximation has been found very efficient to compute matrix-vector products $\varphi_k(A) b$ due to the optimality of the matrix polynomials produced by Krylov methods. The idea behind a Krylov subspace approach is to approximate the vector $\varphi_k(A)b$, which lives in $\mathbb{R}^n$, in a smaller space of dimension $m$. The Krylov approximation of $\varphi_k(A)b$ is based on an Arnoldi decomposition of $A$ $$A V_m = V_{m+1} \widetilde{H}_m = V_m H_m + \eta_{m+1,m} v_{m+1} e^T_m \;,$$ where $V_m$ is a $m \times n$ matrix whose columns form an orthogonal basis for the Krylov subspace of dimension $m$ $$K_m(A, b) = \mbox{span}\{b, Ab, \dots, A^{m-1}b \}\;,$$ $\widetilde{H}_m = [\eta_{i,j}]$ is an $(m + 1) \times m$ upper Hessenberg matrix, $H_m = [I_m \; 0]\widetilde{H}_m$ and $e_m$ denotes the $m$th vector of the standard basis of $\mathbb{R}^m$. The matrices $V_m$ and $H_m$ are such that $$H_m = V_m^T A V_m\;, \label {arnoldi_consequence}$$ therefore $H_m$ can be seen as the projection of the action of $A$ to the Krylov subspace $K_m(A, b)$.
Standard Krylov methods requires the storage of $V_m$, i.e. the storage of $m$ vectors of size $n$, which may be costly for moderate to large values of $m$. Therefore, a standard Krylov subspace method may be impractical because of the storage requirements associated with $V_m$. To overcome this issue, the Arnoldi approximation could be modified in a way that allows the construction of successively better approximations of $\varphi_k(A)b$ based on a sequence of Krylov spaces of small dimension. Methods based on this approach are called restarted Krylov subspace methods, and they use an Arnoldi-like decomposition where a sequence of ascending (not necessarily orthonormal) basis vectors is introduced. We now focus on the restarted Krylov subspace algorithm presented in [@eiermann2006restarted], which can be summarized as follows. An Arnoldi-like decomposition of $A$ is constructed $$A \widehat{V}_p = \widehat{V}_p \widehat{H}_p + n_{p+1} v_{pm+1} e_{pm}^T \;, \label{arnoldi_like}$$ where $\widehat{V}_p=[V_1 \; V_2 \; \cdots \; V_p] \in \mathbb{R}^{n \times pm}$, $\widehat{H}_p = \begin{bmatrix}
H_1 & \\
E_2 & H_2 & \\
& \ddots & \ddots &\\
& & E_p & H_p
\end{bmatrix} \in \mathbb{R}^{pm \times pm}$, and\
$E_j=\eta_j e_1 e_m^T \in \mathbb{R}^{m \times m}$, $j=2, \dots, p$. Since is an Arnoldi-like decomposition, the columns of $\widehat{V}_k$ are only blockwise orthonormal. The matrices $V_1,V_2, \dots, V_p \in \mathbb{R}^{n \times m}$, $H_1,H_2, \dots, H_p \in \mathbb{R}^{m \times m}$ and the scalars $\eta_2, \eta_3, \dots, \eta_{p+1}$ are obtained from $p$ proper Arnoldi decompositions. Setting $$\varphi_k(\widehat{H}_p) = \begin{bmatrix}
\varphi_k^{1,1} & \vspace{0.05 cm}\\
\varphi_k^{2,1} & \varphi_k^{2,2} & \\
\vdots & \vdots & \ddots & \\
\varphi_k^{p,1} & \varphi_k^{p,2} & \dots & \varphi_k^{p,p}
\end{bmatrix}, \quad \mbox{where } \; \varphi_k^{j,j} = \varphi_k(H_j), \;\; j=1, 2, \dots, p \;,$$ the approximation $\varphi_k^p$ of $\varphi_k(A)b$ after $p$ restart cycles is given by $$\varphi_k^p = \widehat{V}_p \varphi_k(\widehat{H}_p)e_1 = [V_1 \; V_2 \; \cdots \; V_p] \varphi_k(\widehat{H}_p)e_1 =
\sum_{j=1}^p V_j \varphi_k^{j,1} e_1 = \varphi_k^{p-1} + V_p \varphi_k^{p,1} e_1 \;.$$ Therefore, the approximation $\varphi_k^{p}$ is obtained from the previous approximation $\varphi_k^{p-1}$ plus a correction term. Only $\varphi_k^{p-1}$ has to be stored from the previous cycle of the algorithm, and the matrix $V_{p-1}$ (together with $V_{p-2}, \dots ,V_1$) can be discarded after computing $\varphi_k^{p-1}$. An efficient implementation of this algorithm can be found in [@afanasjew2008implementation], where it is also shown how to stably compute the coefficient vector $\varphi_k^{p,1} e_1$.
### Scaling and Squaring {#scaling}
We now present a scaling and squaring method for the computation of $\varphi_k(A)$. We are going to base the scaling and squaring method on the recursive relations for the $\varphi_k$-functions; cf. [@beylkin1998new; @koikari2007error; @SkaflestadWright:2009]. They are given as (see [@SkaflestadWright:2009 Lemma 3]): $$\begin{aligned}
\label{eq:phi_simple}
&2^k\varphi_k(2A) = \exp(A)\varphi_k(A) + \sum_{j=0}^{k-1}\frac{\varphi_{k-j}(A)}{j!} \\
\label{eq:phi_BKV}
=&
\begin{cases}
\varphi_{k/2}(A)^2 + 2 \sum_{j=0}^{k/2} \frac{1}{j!} \varphi_{k-j}(A)
&\text{for } k \text{ even,} \\
\varphi_{(k-1)/2}(A)\varphi_{(k+1)/2}(A) + 2 \sum_{j=0}^{(k-1)/2} \frac{\varphi_{k-j}(A)}{j!}
+ \frac{\varphi_{((k+1)/2)}(A)}{((k+1)/2)!}
&\text{for } k \text{ odd.} \\
\end{cases}\end{aligned}$$ Note that in both expressions the matrix function of the matrix $2A$ is expressed as a product of two matrix functions of $A$ plus some correction terms. The first identity is simpler, but the second one reduces the number of correction terms, which is slightly more efficient for computations; however [@SkaflestadWright:2009] mentions that the first form is more stable in numerical experiments.
The first formula follows from the definition using $\Delta{t} = 2$, and a splitting of the integral into the intervals $(0,1)$ and $(1,2)$ $$\begin{aligned}
2^k\varphi_k(2A) &= \int_0^1 \exp((2- \tau)A) \frac{\tau^{k-1}}{(k-1)!} {\mathop{}\!\mathrm{d}}\tau
+ \int_1^2 \exp((2-\tau)A) \frac{\tau^{k-1}}{(k-1)!} {\mathop{}\!\mathrm{d}}\tau \\
&= \exp(A) \int_0^1 \exp((1-\tau)A) \frac{\tau^{k-1}}{(k-1)!}
+ \int_0^1 \exp((1-\tau)A) \frac{(1+\tau)^{k-1}}{(k-1)!} {\mathop{}\!\mathrm{d}}\tau,\end{aligned}$$ where the second integral was shifted to $(0,1)$. Now, using the binomial identity $(1+\tau)^{k-1}/(k-1)!
= \sum_{j=0}^{k-1} \tau^{k-j-1} / ((k-1-j)!\,j!)$ in the second term and the definition of $\varphi_k$ and $\varphi_{k-j}$ shows . Now, to derive the second identity, we use in a first step $$\begin{gathered}
\exp(A)\varphi_k(A) = (I + \varphi_1(A)A)\varphi_k(A)
= \varphi_k(A) + \varphi_1(A) (A\varphi_k(A)) \\
= \varphi_k(A) + \varphi_1(A) (\varphi_{k-1}(A) - I/(k-1)!)
= \varphi_1(A) \varphi_{k-1}(A) + \varphi_k(A) - \varphi_1(A)/(k-1)!.\end{gathered}$$ Proceeding iteratively $m$ times with the first term, we obtain $$\exp(A)\varphi_k(A) = \varphi_{m}(A)\varphi_{k-m}(A) + \sum_{j=1}^m\frac{\varphi_{k-j}(A)}{j!} - \sum_{j=k-m-1}^{k-1} \frac{\varphi_{k-j}(A)}{j!}.$$ using this identity for $m = k/2$ and $m = (k-1)/2$ respectively, we obtain .
In order to evaluate the matrix functions, we are going exploit these recursive relations to reduce the computation to matrix functions of a scaled matrix $A/2^M$. Then, we use a polynomial approximation for this matrix. In order to ensure accuracy and stability, we do this in the following way: First, let $$p_0^{0}(z) = T_{r}(z) + z^{r+1}q(z) = \exp(z) + \mathcal{O}(\abs{z}^{r+1})$$ be a polynomial of degree $N_p$ that approximates the exponential function up to order $r$. Here, $T_r(z) = 1+z + \dotsc + z^r/r!$ is the Taylor approximation to $\exp$, and $q(z)$ is a remainder. Then, for all $k \leq r+1$ we define the consistent polynomial approximations to the $\varphi_k$-functions as $$\label{eq:poly_init}
p_k^{0}(z) = z^{-k}\left(p_0^{M}(z) - T_k(z)\right)
= \sum_{j=0}^{r-k} \frac{z^{j}}{(j-k)!} + z^{r+1-k}q(z)$$ Now, we define the higher order recursive approximations for $M>0$ to $\varphi_k$ using as $$\label{eq:poly_rec}
p_k^{M}(z) = 2^{-k}\left(p^{M-1}_0(z/2)p^{M-1}_k(z/2)
+ \sum_{j=0}^{k-1} \frac{p_{k-j}^{M-1}(z/2)}{j!} \right).$$ This definition ensures that the resulting $p_k^{r,M}$ functions have similar properties as the original $\varphi$ functions.
For any $M \geq 0$ and $1\leq k\leq r+1$ there holds $$\begin{aligned}
\label{eq:pk_properties_1}
p_k^{M}(z) &= z^{-1}\left(p_{k-1}^M(z) - 1/(k-1)!\right),\\
\label{eq:pk_properties}
p_k^{M}(z) &= z^{-k}\left(p^M_0(z) - T_{k-1}(z)\right).\end{aligned}$$
For $M = 0$ this holds according to definition. For higher $M$ we follow an induction argument.
First, we show . We use the recursive definition to obtain $$z p_k^M(z) = \frac{1}{2^{k-1}}\left(p^{M-1}_0(z/2) (z/2) p^{M-1}_k(z/2) +
\sum_{j=0}^{k-1} \frac{(z/2) p^{M-1}_{k-j}(z/2)}{j!}\right)$$ and use the induction hypothesis to obtain $$\begin{aligned}
2^{k-1} z p_k^M(z) &=
p^{M-1}_0(z/2) \left(p_{k-1}^{M-1}(z/2) - \frac{1}{(k-1)!}\right)
+ \sum_{j=0}^{k-1} \frac{p^{M-1}_{k-j-1}(z/2) - 1/(k-j-1)!}{j!} \\
&= p^{M-1}_0(z/2) p^{M-1}_{k-1}(z/2) + \sum_{j=0}^{k-2} \frac{p^{M-1}_{k-1-j}(z/2)}{j!}
- \sum_{j=0}^{k-1}\frac{1}{j!(k-j-1)!} \\
&= 2^{k-1}p^{M}_{k-1}(z) - 2^{k-1}/(k-1)!,\end{aligned}$$ using again the recursive definition for $k-1$ and the well-known summation formula of binomial coefficients. Dividing by $2^{k-1}z$ yields . Concerning , we note that is suffices to repeatedly apply for $k$, $k-1$, ..., $1$.
The above result also implies that the recursion from is equivalent for the construction of $p_k^M$, since the equality can be used to convert between both versions.
In -, which is the ETD method we actually use to later solve the tracer equation, the only $\varphi_k$-function needed is $\varphi_1$. For this specific case, the recursion formula looks like $$\varphi_1(A) = \frac{1}{2} \big(\exp \big(\frac{1}{2} A \big) + I\big) \varphi_1(\frac{1}{2} A)\;.$$ and $$p_1^{M}(z) = z^{-1}(p_0^{M}(z)-1)\;.$$ Using the above relations, the following algorithm computes $\varphi_1(A)$ from $\varphi_1(A/2^M)$.
1. Define $${\textcolor{black}{p_1^{0}(A) = \frac{A^{-1}}{2^M}(p_0^{0}(A/2^M) - I)\;,\;\;\;\; p_0^{0}(A/2^M) = \exp(A/2^M) + \mathcal{O}(\abs{A/2^M}^{r+1}) \;.}}$$
2. For $0 \leq j \leq M$, Given $p_0^{j}(\frac{A}{2^{M+j-1}})$ and $p_1^{j}(\frac{A}{2^{M+j-1}})$, compute $p_1^{j}(\frac{A}{2^{M+j}})$ as $$p_1^{s,j}\Big(\frac{A}{2^{M+j}}\Big) = \frac{1}{2} \Big( p_0^{j}\Big(\frac{A}{2^{M+j-1}}\Big) + I \Big)
p_1^{j}\big(\frac{A}{2^{M+j-1}}\big)\;.$$
3. Given $p_0^{j}(\frac{A}{2^{M+j-1}})$, compute $p_0^{j}(\frac{A}{2^{M+j}})$ as $$p_0^{j}(\frac{A}{2^{M+j}}) = p_0^{j}(\frac{A}{2^{M+j-1}})) \; p_0^{j}(\frac{A}{2^{M+j-1}})\;.$$
Finally, we provide an error estimate for the approximation $\varphi_k \approx p_k^M$. First, we consider the polynomial approximation on a subset of the complex plane. To prepare for the general case, we let $\Sigma \subset \mathbb{C}^{-} + \rho_0 \subset \mathbb{C}$ be some compact subset of the negative complex plane shifted by $\rho_0 \geq 0$, and assume that the underlying polynomial fulfills the stability assumption $$\label{eq:assu_poly_stab}
\abs{p_0^0(z)} \leq \exp(\tau\rho_0) \quad\text{for all } z \in \tau\Sigma,
\text{ where } 0 \leq \tau \leq 1.$$ Moreover, since $\Sigma$ is compact, there exists $c_q > 0$ such that $$\label{eq:poly_approx}
\abs{p_0^0(z) - \exp(z)} \leq c_q \abs{z}^{r+1} \quad\text{for all } z \in \Sigma.$$ This will be the basis of the error estimates.
There are many situations where this assumption is fulfilled. In the case that $p_0^0(z)$ is the Taylor polynomial $T_r(z)$ and $\rho_0 = 0$, $\Sigma$ can be chosen as the intersection of the negative half-plane $\mathbb{C}^-$ and the well-known stability region of a Runge-Kutta scheme of order $r$. In this case, due to the relation $T_r(z) - \exp(z) = z^{r+1}\varphi_{r+1}(z)$ and $\abs{\varphi_{r+1}(z)} \leq
\varphi_{r+1}(\operatorname{Re}z) \leq 1/(r+1)!$ we have $c_q = 1/(r+1)!$. For an overview over known results on polynomials with optimal stability properties for various forms of $\Sigma$ and a computational approach to determine them, we refer to Ketcheson [@ketcheson2013optimal].
We first investigate the case of the matrix exponential $k=0$.
\[prop:err\_est\_scsq\] Let the polynomial $p_0^0$ fulfill . Then for all $M \geq 0$ we have the stability and approximation properties: $$\begin{aligned}
\label{eq:stab_M}
\abs{p_0^M(z)} &\leq \exp(\tau\rho_0), \\
\label{eq:approx_M}
\abs{p_0^M(z) - \exp(z)} &\leq c_q \exp(\tau\rho_0) \abs{z}^{r+1} 2^{-Mr},\end{aligned}$$ which hold for all $z \in \tau \Sigma$, for the enlarged stability radii $0 \leq \tau \leq 2^M$.
The first statement follows in a straightforward way from since $p_0^M(z) = p_0^0(z/2^M)^M$. For the second statement, we use an induction argument, noting that the case $M=0$ follows directly from . For $M > 0$, we let $z \in \tau\Sigma$ be arbitrary we use to obtain $$\begin{aligned}
\abs*{\exp(z) - p_0^M(z)}
&= \abs*{\left(\exp(z/2) + p_0^{M-1}(z/2)\right)\left(\exp(z/2) - p_0^{M-1}(z/2)\right)} \\
&\leq 2 \exp(\tau\rho_0/2)\, c_q\exp(\tau\rho_0/2) \abs*{z/2}^{r+1} 2^{-(M-1)r} \\
&= \exp(\tau\rho_0) \abs*{z}^{r+1} 2^{-Mr},\end{aligned}$$ where we have used $z/2 \in (\tau/2)\Sigma$ and the induction hypothesis together with and $\abs{\exp(z/2)} =
\exp(\operatorname{Re(z)/2}) \leq \exp(\tau\rho_0/2)$ in the second step.
A similar estimate follows now also for the higher $\varphi$-functions.
\[cor:err\_est\_scsq\] Let $p_0^0$ fulfill . Then for $k \leq r+1$ we have $$\abs{p_k^M(z) - \varphi_k(z)} \leq c_q \exp(\tau\rho_0) \abs{z}^{r+1-k} 2^{-Mr}.$$ for all $z \in \tau\Sigma$ and $0\leq \tau \leq 2^M$.
It suffices to use to write $$p_k^M(z) - \varphi_k(z) = z^{-k}(p_0^M - \exp(z)),$$ where we can apply Proposition \[prop:err\_est\_scsq\].
For the case of a diagonalizable matrix, one can now obtain an error estimate in the standard way.
\[cor:err\_est\_matr\] Let $A = V D V^{-1}$, where $D$ is a diagonal matrix with the eigenvalues $\sigma(A)
\subset 2^M \Sigma$ on the diagonal. Then if holds, we have $$\norm*{\varphi_k(A) - p_k^M(A)}_2 \leq c_q \operatorname{cond}(V) \exp(\rho_0)
\abs{A}^{r+1-k} \, 2^{-Mr},$$ where $\norm{\cdot}_2$ is the matrix norm induced by the Euclidean norm, $\abs{A}$ is the spectral radius of $A$, and $\operatorname{cond}(V) = \norm{V}_2\norm{V^{-1}}_2$ is the condition number of $V$.
ETD Solver with Operator Splitting for the Tracer Equation {#solver}
==========================================================
In this section, the ETD solver developed for the tracer equation is presented. We focus on the description of the linear operator chosen, whose structure strictly depends on the physics of the problem, i.e. processes occurring at different time-scales. In the ocean, most processes occur at large scales, as forcing by the wind or by currents in the upper layers of the ocean, but in many relevant situations (e.g., eddies, cold water flowing over warm water and vice versa), the vertical transport and mixing of tracers follow much faster time-scales. Hence, semi-implicit methods are needed for the appropriate inclusion of vertical transport and mixing in the model using time-steps that are not excessively small. Using an ETD scheme where the linear operator is split into a vertical and a horizontal part and the vertical part is treated exponentially, larger time steps can be taken and still have an appropriate description of the processes happening in the vertical. While models like MPAS-Ocean and POP treat only the tracer vertical diffusion implicitly, we treat with a matrix exponential all vertical terms of the tracer equation, i.e. vertical advection and diffusion.
Operator Splitting {#jac_splitting}
------------------
Let us rewrite the tracer equation as equation , that is $$\partial_t T = F(T) = J_n T + R(T), \label{tracer_split1}$$ where $T=T(t)$ denotes the vector of tracer values for $t \in [t_n,t_{n+1}]$, and the linear operator is the Jacobian of the system evaluated at $t_n$. Since we are assuming a zero forcing term, the tracer equation is linear in $T$, which implies that the nonlinear reminder $R(T)$ is zero, and so becomes $$\partial_t T = F(T) = J_n T\;.$$ At this point, we split the Jacobian $J_n$ into a vertical and a horizontal part, i.e. $$J_n = J_n^z + J_n^x\;, \label{oper_split}$$ where $J_n^z$ contains the derivatives of the vertical terms only, and $J_n^x$ contains the derivatives of the horizontal terms only. Thus, $$\partial_t T = F(T) = J_n^z T + J_n^x T\;. \label{tracer_split2}$$ An exponential integrator can be applied to solve . Both terms $J_n^z T$ and $J_n^x T$ are linear, so either one of them can be considered as the linear part of the equation or the remainder. Since the vertical processes are those occurring at fast time-scales, the vertical terms have to be treated with a matrix exponential. For this reason, the term $J_n^z T$ is interpreted as the linear part, while $J_n^x T$ is the remainder, which is linear as well in this case. Thus, we construct an exponential time differencing solver where the vertical terms are treated implicitly with a matrix exponential, whereas the horizontal are dealt with in an explicit way. The scheme has the following form: $$\begin{aligned}
& T_{n}^{1st\,stage} = T_n + \Delta t \varphi_1(\Delta t J_n^z)F(T_n)\;, \label{stage1_pc_tracer}\\
& T_{n+1} = T_{n}^{1st\,stage} +
\dfrac{1}{2}\Delta t \varphi_1(\Delta t J_n^z) (R^{1st\,stage}_{n}-R_n)\;, \label{stage2_pc_tracer}\end{aligned}$$ i.e. it is a two-stage ETD method following a predictor-corrector. The remainders are defined as $R_n = F(T_n) - J_n^z T_n = J_n^x T_n$ and $R^{1st\,stage}_{n} = F(T_{n}^{1st\,stage}) - J_n^z T_{n}^{1st\,stage} = J_n^x T_{n}^{1st\,stage}$. Both $R_n$ and $R^{1st\,stage}_n$ take into account only the contributions from the horizontal terms, but computationally $J_n^x$ is never built. $R_n$ is obtained for free from the construction of the right-hand side $F(T_n)$, which is needed in the first stage, and the construction of $R^{1st\,stage}_n$ is cheap because we do not need to construct the full right-hand side $F(T_n^{1st\,stage})$ but only its horizontal terms evaluated at $T_n^{1st\,stage}$.
Please note that we could have chosen to treat the full Jacobian $J_n$ with a matrix exponential, and this would have given us the exact solution (up to machine precision) of equation , since it is linear in $T$. From a computational point of view, this strategy is not appealing since the construction of $J_n$ might be very expensive in terms of time and computational cost. We think that the exponential Euler method would not have been a valid choice either, since it is likely too inaccurate to be useful.
Structure of the matrix $J_n^z$
-------------------------------
Another advantage of the operator splitting relies in the structure of the matrix $J_n^z$. This matrix contains the derivatives of the vertical terms only, and its dimension is $N_z$ $\times$ $N_x$, where $N_z$ indicates the total number of vertical layers, and $N_x$ is the total number of elements in the horizontal discretization, i.e. the total number of Voronoi cells. The entries of $J_n^z$ can be ordered so that the derivatives associated with the same horizontal element form a submatrix of dimension $N_z \times N_z$ . This is possible since, for every layer, there is no interaction between the derivatives of the vertical terms associated with two different Voronoi cells. Therefore, $J_n^z$ has a block diagonal structure, $$J_n^z = \begin{bmatrix}
J_n^{z,1} & 0 & \dots & 0\\
0 & J_n^{z,2} & \dots & 0\\
\vdots & \dots & \dots & \vdots\\
0 & 0 & \dots & J_n^{z,N_x}
\end{bmatrix}$$ where each block $J_n^{z,i}$ represents the contributions (derivatives of the vertical terms) given by a single element $i$ in the horizontal discretization, and the dimension of each block depends on the number of vertical layers. Figure \[block\_diag\] shows the structure of $J_n^z$ and each diagonal block in a simple case with 4 Voronoi cells and 4 vertical layers. Since we are dealing with a one dimensional domain in the horizontal, the diagonal blocks are banded matrices; in particular they are tridiagonal matrices because we use up-winding in the discretization.
![Block diagonal structure of $J_n^z$ for a simplified case with 4 horizontal elements and 4 vertical layers.[]{data-label="block_diag"}](images/block_matrix)
This block diagonal structure gives several advantages. First, $\varphi_1(\Delta t J_n^z)$ can be written as $$\varphi_1(\Delta t J_n^z) = \begin{bmatrix}
\varphi_1(\Delta t J_n^{z,1}) & 0 & \dots & 0\\
0 & \varphi_1(\Delta tJ_n^{z,2}) & \dots & 0\\
\vdots & \dots & \dots & \vdots\\
0 & 0 & \dots & \varphi_1(\Delta tJ_n^{z,N_h})
\end{bmatrix} \;. \label{parallel}$$ Hence, for every Voronoi cell $i$, the smaller matrices $\varphi_1(\Delta t J_n^{z,i})$ can be constructed one at the time, and the global matrix $\varphi_1(\Delta t J_n^z)$ is never assembled. The evaluation of small matrices $\varphi_1(\Delta t J_n^{z,i})$ instead of a large one can significantly speed up the calculations, especially in a parallel setting. Expression makes indeed the computation of $\varphi_1(\Delta t J_n^z)$ easy to implement in a parallel environment. Ideally, each matrix $\varphi_1(\Delta t J_n^{z,i})$ could be assigned to a different processor, if $N_x$ processors were available. This straightforward parallelization would be of great advantage to speed up the computational time, which is one of the major concerns in ocean modeling.
Expression holds for any $\varphi_k$ with $k \ge 0$ and for any block diagonal matrix, as shown in the following proposition.
Let $D$ be a block diagonal matrix $$D=\begin{bmatrix} D_1 & 0 & \dots & 0\\
0 & D_2 & \ddots & 0\\
\vdots & \ddots & \ddots & \vdots\\
0 & 0 & \dots & D_m \end{bmatrix}\;,$$ and let $\varphi_k$ be a $\varphi$-function of index $k$ defined as in (or equivalently ). Then, $$\label{eq:phi_blkdiag}
\varphi_k(D) = \begin{bmatrix}
\varphi_k(D_1) & 0 & \dots & 0\\
0 & \varphi_k(D_2) & \ddots & 0\\
\vdots & \ddots & \ddots & \vdots\\
0 & 0 & \dots & \varphi_k(D_m) \end{bmatrix} \;.$$
Using definition and the block diagonal structure of $D$, we have $$\begin{aligned}
\varphi_k(D) & = \dfrac{1}{(k-1)!} \int_0^1 \exp((1-\sigma)D) \sigma^{k-1} d\sigma \\
& = \dfrac{1}{(k-1)!} \int_0^1 \begin{bmatrix}
\exp((1-\sigma)D_1) & 0 & \dots & 0\\
0 & \exp((1-\sigma)D_2) & \ddots & 0\\
\vdots & \ddots & \ddots & \vdots\\
0 & 0 & \dots & \;\exp((1-\sigma)D_m)\\
\end{bmatrix}
\sigma^{k-1} {\mathop{}\!\mathrm{d}}\sigma.
$$ Applying the integral to every entry of the matrix we have that the diagonal blocks become $$\begin{aligned}
\frac{1}{(k-1)!} \int_0^1 \exp((1-\sigma) & D_i) \sigma^{k-1} {\mathop{}\!\mathrm{d}}\sigma = \varphi_k(D_i)\;,\end{aligned}$$ for all $i = 1,\ldots,N_h$. Thus, has been verified.
Another benefit given by relies in the different accuracy that can be chosen in the approximation of each matrix $\varphi_1(\Delta t J_n^{z,i})$. If a Krylov scheme is used, the dimension of the Krylov space chosen to approximate $\varphi_1(\Delta t J_n^{z,i})b_i$ can vary for every $i$, depending on the physics of the problem. A higher dimension (so more Krylov vectors) can be used for those horizontal elements that experience a great vertical mixing or transport, while a lower dimension can be adopted to approximate $\varphi_1(\Delta t J_n^{z,i})b_i$ for the remaining elements $i$. In most regions of the ocean, mixing is relatively small, and considering a lower subspace dimension for the cells that discretize those areas would considerably speed up the computational time without jeopardizing the accuracy of the approximation. Using a restarted Krylov method, the number of Krylov vectors can be significantly smaller than the dimension of $J_n^{z,i}$, i.e. $N_x$, this number can change at each restart. Similarly, if the scaling and squaring method described in section \[scaling\] is adopted, a different value of $M$ can be used for each matrix $\Delta t J_n^{z,i}$. For those Voronoi cells experiencing great vertical mixing or transport, the spectral radius of the corresponding $\Delta t J_n^{z,i}$ would be larger than for those cells impacted by relatively small mixing or transport. Hence, smaller values of $M$ are needed for the elements not experiencing processes occurring at fast time-scales, and since such elements are the great majority, significant speed ups can be obtained.
Computational Complexity {#com_com}
------------------------
We complete this section with a brief discussion of the computational cost depending on the approach chosen for computing the $\varphi_k$-functions. For dense matrices, the scaling and squaring approach is $\mathcal{O}(N^3_z)$ and probably expensive if $N_z$ is too large. for two banded matrices with bandwidth $b$, the matrix product will be banded with bandwidth $2b$, and the product will cost $\mathcal{O}((1+b)^2 N_z)$.
Numerical Tests {#tests}
===============
In this section, numerical tests are presented to investigate the performances of our ETD solver and compare it with other ocean models. The ETD time-stepping scheme is implemented using the two approaches for the computation of the $\varphi$-functions described in section \[ETD\]: the Krylov subspace scheme presented in [@eiermann2006restarted] and the scaling and squaring algorithm developed for $\varphi_k$ with $k \ge 0$. The performances given by the two implementations are investigated numerically in the case of one and multiple tracers. Comparisons with other semi-implicit schemes are also presented. Finally, , to make sure that the proposed method is able to reproduce similar results under the same physical conditions. To do so, the whole primitive equation system is solved. All the tests in this section are 2D, namely one dimension in the horizontal and one in the vertical. They have been implemented in the in-house the C++ library FEMuS ([@FEMuSlibrary]), and for the Krylov subspace scheme, the SLEPc library has been used.
For all the tests, up-winding is used to discretize the horizontal and vertical advection. For the performance tests, first order up-winding is used for both the vertical and the horizontal, whereas a third-order up-winding scheme for the horizontal advection and a first-order scheme for the vertical are employed for the comparisons with other ocean models. In section \[tracer\_eq\_section\], the discretization of the tracer equation was presented using a central difference scheme, but in general it is always possible to move from a central difference scheme to up-winding, either first order or higher, by making an appropriate choice of the diffusion operators. Computationally, up-winding was necessary since central differences cause instabilities in advection-diffusion problems ([@manzini2008finite]). The advection schemes used by other ocean models are discussed in section \[primitive\_eq\_tests\].
Performance
-----------
For the performance tests, the tracer equation is solved, considering $u$, $w$ and $h$ constant in time. The domain is a $10$ m $\times 10$ m box discretized with 12 elements in the horizontal and 100 layers in the vertical, hence $\Delta x = 0.83$ m and $\Delta z = 0.1$ m. The velocity field is a circular, . It is defined as $$(u, w) = (-\psi_1(x)\psi_2'(z),\psi_1'(x)\psi_2(z))\;,$$ where $$\psi_1(x) = 1 - \dfrac{(x - \frac{x_{max}}{2})^4}{(\frac{x_{max}}{2})^4}\;, \quad
\psi_2(z) = 1 - \dfrac{(z - \frac{z_{min}}{2})^2}{(\frac{-z_{min}}{2})^2}\;,$$ with $x_{max} = 10$ and $z_{min} = -10$. Figure \[velocity\] shows that, with this velocity field, the Voronoi cells close to the boundaries experience more vertical transport and mixing than those in the center.
![Velocity field.[]{data-label="velocity"}](images/velocity_second_test)
The two CFL numbers, CFL$_x$ and CFL$_z$ are defined as $$\mbox{CFL}_x = \dfrac{\max u \cdot dt}{\Delta x} \quad \mbox{and} \quad \mbox{CFL}_z = \dfrac{\max w \cdot dt}{\Delta z}\;.$$ With the given velocity field and discretization, the ratio $\dfrac{\mbox{CFL}_z}{\mbox{CFL}_x}$ is equal to 16.6. This implies that the transport and mixing in the vertical are significantly bigger than those in the horizontal. For the horizontal and vertical diffusion, the values chosen are $\kappa_{x} = 10^{-4}$ and $\kappa_{z} = 2.5 \cdot 10^{-5}$.
First, let us consider the case of a single tracer, and let us solve the equation with the second stage ETD method - and a semi-implicit method that uses implicit Euler to treat the vertical diffusion and RK4 for the rest of the terms. In the following, the latter scheme is denoted as RK4 + implicit Euler. The initial condition given for the tracer, which we assume to be temperature, $\varTheta(x,z)$, is $$\varTheta(x,z)=\begin{cases}
5, & \text{$x<5$ m,} \\
30, & \text{$x\ge 5$ m.}
\end{cases}$$ This is the same initial condition given for the temperature in the lock exchange test case analyzed later. Table \[speed\_up\_over\_RK\] shows the times employed by the two time-stepping schemes to reach the steady state. For the ETD method, both the restarted Krylov and the scaling and squaring algorithm have been used for the evaluation of the $\varphi_1$ functions. When the restarted Krylov algorithm is used, the method is denoted as ETD2 Restarted Krylov, whereas ETD2 Scaling & Squaring indicates the scheme using the scaling and squaring algorithm. With ETD2 Restarted Krylov, the vectors $\varphi(\Delta t J_n^{z,i})b_i$ are computed at each stage, without ever storing $\varphi_1(\Delta t J_n^{z,i})$. Taking advantage of the physics of the problem, i.e. larger vertical transport and mixing for the Voronoi cells close to the boundaries, 20 Krylov vectors have been used for the six external elements and 10 vectors for the central six with ETD2 Restarted Krylov. The highest possible dimension for the Krylov subspaces is 100, since the matrices $\Delta t J_n^{z,i}$ are $100 \times 100$. When the scaling and squaring method is applied, $2^4$ is used as scaling factor for all the matrices $\Delta t J_n^{z,i}$. As Table \[speed\_up\_over\_RK\] shows, with the ETD methods larger time steps can be taken than with RK4 + implicit Euler. The values of $\Delta t=3$ and $\Delta t=0.25$ were the largest values that could be taken for the two time-stepping schemes without compromising stability. The simulations end after 2000 time steps. The time employed by the ETD methods is, in all cases, smaller than the time required by RK4 + implicit Euler, because a time step 12 times larger than that of RK4 + implicit Euler could be used. The computational time is reduced by $5.29$ times using ETD2 Restarted Krylov and $5.94$ using ETD2 Scaling & Squaring. Another advantage of such methods over RK4 + implicit Euler is in the order of accuracy of the time-stepping scheme. The two-stage ETD method - is second order accurate, whereas RK4 + implicit Euler is only first order.
The results in Table \[speed\_up\_over\_RK\] are obtained computing the Jacobian $J_n$ at every time step, since in a realistic ocean model simulation, the values of $u$, $w$ and $h$ would change at every time step. Consequently, $J_n^z$ and $J_n^x$ would change. Let us recall that $J_n^x$ is actually never built since in $R_n$ is obtained for free, and $R_n^{1st\;stage}$ can be constructed by evaluating only the horizontal terms of $F(T_n^{1st\;stage})$.
If $w$ does not change too much from one step to the other, we could actually fix $J_n^z$ at some instant of time $t_m$, and use $J_m^z$ in the following iterations, . This choice might still give an advantage to the scaling and squaring method, since we would having a constant in time linear operator starting from $t_m$, without compromising accuracy. In section \[sec:conclusions\], we briefly discuss how this could be done as a future work.
\[new\_option\]
[|c|c|c|c|c|]{}\
& & &\
ETD2 Restarted Krylov & 3 & 2000 & 112.7933\
ETD2 Scaling & Squaring & 3 & 2000 & 100.3743\
RK4 + implicit Euler & 0.25 & 24000 & 596.6595\
Now, let us consider the case of multiple tracers. Having multiple tracers implies that multiple equations of the form need to be solved. The same test case as for the single tracer simulations is used. For simplicity, the same initial condition is given but with different numerical values for each tracer, so that the steady state changes for each tracer. In the upcoming tests, the time-stepping scheme RK4 + implicit Euler is not considered. Having multiple tracers may seem a straightforward task to deal with, but in ocean modeling the amount of tracers can be quite large, ranging from 1 to 70. The computational time employed to solve these equations can then drag down the run time of the whole simulation. For this reason, it is not ideal to just naively solve multiple tracer equations. When dealing with multiple equations of the form , one can take advantage of the fact that the Jacobian $J_n$ is the same for all of them, no matter how many they are. Therefore, the matrices $J_n^{z,i}$ and $\varphi_1(\Delta t J_n^{z,i})$ can be computed only for one tracer at every time step, and then used to solve all the tracer equations. What changes for each equation is just the right hand side $b_i$ that multiples $\varphi_1(\Delta t J_n^{z,i})$ for a given $i$. Combining this strategy with ETD2 Scaling & Squaring, we simply compute the matrices $\varphi_1(\Delta t J_n^{z,i})$ at the first stage for only one tracer. Table \[multiple\_tracers\] shows the computational times obtained with this approach in the case of 1, 2, 4 and 6 tracers . From the table, we see that an important advantage of lies in the increased time saving when more tracers are added. Using ETD2 Restarted Krylov, the matrices $\varphi_1(\Delta t J_n^{z,i})$ are re-computed for every tracer and this causes the times to roughly double when the number of tracers is doubled. The speed-up over the Restarted Krylov implementation is almost 50% with 2 tracers, and it keeps increasing up to 72.75% for 6 tracers. Therefore, a significant amount of time is saved by taking advantage of having the same linear part for all tracers.
For this test, we consciously considered a small number of elements in the horizontal because one of the biggest advantage of the ETD method - is to be heavily parallel, and so each process will have only a few numbers of Voronoi cells to deal with. Thus, using a fine grid in the horizontal is not relevant for this performance comparison.
Comparison with other ocean models {#primitive_eq_tests}
----------------------------------
We now compare the proposed ETD method to other ocean models under identical conditions (whenever the data is available). Two benchmark tests from [@ilicak2012spurious] are performed, addressing the solution of the primitive equations. These tests show, for different initial conditions, the temperature distribution at a given instant of time. Since hydrostaticity is assumed in this work, the primitive equations are described by the incompressible Boussinesq equations in hydrostatic balance. For the tests, the tracer equation is coupled to the dynamics. A second-order ETD solver is employed for the dynamics system, specifically Exponential Rosenbrock Euler. The tests presented are two-dimensional in $(x,z)$, namely one dimension in the horizontal and one dimension in the vertical, and a linear equation of state is used. The equation of state has the form $$\rho = \rho_{ref} - \alpha(\varTheta - \varTheta_{ref})\;,$$ where $\rho_{ref} = 1000$ kg m$^3$, $\alpha = 0.2$ kg m$^3$ C$^{-1}$ and $\varTheta_{ref} = 5^{\circ}$ C, so that density depends only on temperature. For both tests, our solutions are compared with those obtained by MPAS-Ocean, and for the second test, a comparison is made also with MITgcm and MOM. The advection schemes used by these three codes and our ETD scheme are different. The tracer advection scheme we use is a third-order upwinding scheme for the horizontal advection and a first-order upwinding scheme for the vertical, whereas MPAS-Ocean computes high- and low-order estimate of the tracer flux which are then blended using the flux-corrected transport scheme of Zalesak [@zalesak1979fully]. MITgcm uses a 7th-order monotonicity preserving advection scheme [@marshall1997finite], wheres MOM employs a third-order accurate scheme based on a multi-dimensionsal piecewise parabolic method [@griffies2009elements]. We remark that another different between the four solvers relies in the treatment of the vertical advection, which is explicit for the three ocean models and implicit in our method. The results shown in this section are obtained using the restarted Krylov subspace method for computing the $\varphi_1$ functions. Solutions obtained with the scaling and squaring method are nearly identical. The comparisons made are only qualitative, to show that our tracer solver is able to reproduce results comparable to those of other models.
### Lock Exchange Test Case
The lock exchange test case may be thought of as two basins of water with different temperatures that start interacting at time zero. The domain is a $64,000 \times 20$ rectangle, where $64,000$ is the $x$ size and $20$ is the $z$ size. All dimensions are in meters (m). The cell sizes are $\Delta x = 500$ and $\Delta z = 1$, i.e. 128 elements are considered in the horizontal, while twenty layers are considered in the vertical. The initial condition for temperature is $$\varTheta(x,z)=\begin{cases}
5, & \text{$x<32,000$ m,} \\
30, & \text{$x\ge 32,000$ m,}
\end{cases}$$ so, warm water flows over the cold water from right to left, and viceversa once the domain boundaries are touched. The initial condition for velocity is $u=0$ in every layer. The values for the horizontal and vertical viscosity are $100$ m$^2$ s$^{-1}$ and $0.0001$ m$^2$ s$^{-1}$, respectively, while all tracer diffusions are turned off. The simulation stops at 17 h, and the same $dt$ used by MPAS-Ocean in [@petersen2015evaluation] is adopted, i.e. $dt=60$ s. Figure \[lock\] shows the temperature distribution at 17 h obtained with our ETD solver and with MPAS-Ocean.
\(a) ![Temperature distribution for the lock exchange test case with $\nu_h=100$: (a) proposed ETD solver, (b) MPAS-Ocean ([@petersen2015evaluation]).[]{data-label="lock"}](images/pic_LE_third_crop "fig:"){height="3.0cm" width="6.4cm"}\
(b) ![Temperature distribution for the lock exchange test case with $\nu_h=100$: (a) proposed ETD solver, (b) MPAS-Ocean ([@petersen2015evaluation]).[]{data-label="lock"}](images/LE_ni100_MPAS "fig:")
The simulation performed by MPAS-Ocean uses z-star vertical coordinates, while we use z-level. This difference does not compromises the comparison between the two distributions, since, as reported in [@petersen2015evaluation], results for z-level and z-star vertical coordinate settings in MPAS-Ocean are nearly identical. As Figure \[lock\] shows, the two distributions are comparable. With z-type coordinates, the intermediate layers are expected to have temperatures in between $5^{\circ}$ C and $30^{\circ}$ C, and this behavior is visible in both Figure \[lock\] a) and Figure \[lock\] b). The right front location at 17 h is very similar: with MPAS-Ocean the front is at 62 km, while with our ETD solver is at 62.4 km. The location obtained with the proposed ETD solver coincides with the theoretical prediction for this test based on the speed of a gravity current in a rectangular channel [@benjamin1968gravity]. A difference between the two temperature distributions is in the amplitude of the mixing. The interface between the density layers is sharper in Figure \[lock\] b), and this is probably due to the different advection scheme used by MPAS-Ocean and to the exponential treatment of the vertical advection in our ETD solver.
### Internal Waves Test Case
\(a) ![Temperature distribution for the internal waves test case: (a) our ETD solver, (b) MPAS-Ocean ([@petersen2015evaluation]), (c) MITgcm ([@ilicak2012spurious]), (d) MOM ([@ilicak2012spurious]).[]{data-label="internal"}](images/pic_internal_waves "fig:") (b) ![Temperature distribution for the internal waves test case: (a) our ETD solver, (b) MPAS-Ocean ([@petersen2015evaluation]), (c) MITgcm ([@ilicak2012spurious]), (d) MOM ([@ilicak2012spurious]).[]{data-label="internal"}](images/internal_waves_MPAS "fig:"){height="4.2cm" width="5.0cm"}\
(c) ![Temperature distribution for the internal waves test case: (a) our ETD solver, (b) MPAS-Ocean ([@petersen2015evaluation]), (c) MITgcm ([@ilicak2012spurious]), (d) MOM ([@ilicak2012spurious]).[]{data-label="internal"}](images/MITgem "fig:") (d) ![Temperature distribution for the internal waves test case: (a) our ETD solver, (b) MPAS-Ocean ([@petersen2015evaluation]), (c) MITgcm ([@ilicak2012spurious]), (d) MOM ([@ilicak2012spurious]).[]{data-label="internal"}](images/MOM "fig:")
Internal waves are waves that oscillate within the interior of the ocean, rather than on its surface. They generate when the interface between layers of different water densities is disturbed. This test was chosen for further validation since linear internal waves tend to produce vertical mixing in ocean models, especially when z-level and z-star coordinates are employed [@gouillon2010internal]. The domain is a $250,000 \times 500$ rectangle, where the dimensions are given in meters. The cell sizes are $\Delta x = 5000$ and $\Delta z = 25$, i.e. 50 elements are considered in the horizontal, while twenty layers are considered in the vertical. The initial temperature distribution is $\varTheta_0(z) + \varTheta'(x,z)$, with $$\begin{aligned}
\varTheta_0(z) & = \varTheta_{bot}+(\varTheta_{top}-\varTheta_{bot})\dfrac{z_{bot}-z}{z_{bot}}\,,\;\;\mbox{and} \\
\varTheta'(x,z) & = -A \cos \Big(\frac{\pi}{2L}(x-x_0)\Big) \sin\Big(\pi\frac{z + 0.5 \Delta z}{z_{bot} + 0.5 \Delta z}\Big)\,,\end{aligned}$$ where $\varTheta_{bot} = 10.1^{\circ}$ C, $\varTheta_{top} = 20.1^{\circ}$ C, $z_{bot} = -487.5$ m, $L = 50$ km, $x_0 = 125$ km, $x_0 - L < x < x_0 + L$, $\Delta z = 25$ m, and $A = 2^{\circ}$ C. This means that we initially have a small temperature perturbation in each layer that induces wave propagation out from the center. This behavior is similar to that found in realistic global simulations. The initial condition for velocity is $u=0$ in every layer. The values for the horizontal and vertical viscosity are $0.01$ m$^2$ s$^{-1}$ and $0.0001$ m$^2$ s$^{-1}$, respectively, while all tracer diffusions are, again, turned off. Unlike the previous test, this simulation may proceed indefinitely, but we choose to stop it at 200 days. The time step used is $dt=300$ s, which, again, is the same adopted by MPAS-Ocean. Figure \[internal\] shows the temperature distribution at 200 days obtained with our ETD solver and three other models: MPAS-Ocean, MITgcm and MOM. In general, the four temperature distributions are comparable. The mixing is visible in all models, despite is different among the four solutions. Again, the four codes present four different advection schemes for the tracer equation, and this could explain the discrepancies in the solutions.
Concluding Remarks
==================
So far, in the analysis and in all tests performed, we assumed a zero forcing term for the tracer equations. Now, the more general case of a non-zero forcing term is considered, and we show that the analysis made above is still valid.
Forcing terms introduce the contributions given by external factors like penetrative solar radiation and surface boundary conditions [@madec2015nemo]. By treating these terms explicitly, i.e. excluding them from the matrix exponential, we can still follow the operator splitting procedure shown in section \[jac\_splitting\]. Let us assume that the forcing term $\mathcal{F^T}$ in is non-zero. Let us define the Jacobian $J$ as the Jacobian of the associate equation with a zero forcing term. In this way, the matrices $J_n^z$ and $J_n^x$ are the same as for the zero forcing term case and the reminder is now given by $\mathcal{F^T}$ plus $J_n^x T$. Hence, no matter the nature of $\mathcal{F^T}$ and no matter its form, we can still apply the method - with the same $J_n^z$ matrix as for the zero forcing term case. In this way, when multiple tracers are present, they will all have the same linear part, and so the matrices $\varphi_1(J_n^{z,i})$ just need to be computed for one tracer and re-used for the others, as we did in the numerical tests. The possibility of re-using the matrices $\varphi_1(J_n^{z,i})$ even in the more general case of a non-zero forcing term is a great feature of the method that gives a saving in computational time that increases with the number of tracers. What changes in the scheme - is the difference $R_n^{1st\;stage} - R_n$ that appears in the second stage, since in both reminders there is now the contribution of $\mathcal{F^T}$.
A possible issue that may occur with the introduction of forcing terms relies in the associated time-scales. By adding a forcing term, new physical or chemical processes are taken into account, in particular biochemical reaction terms may be included. If the time-scales associated with such processes are comparable with those associate with the vertical transport and mixing, then these terms must be included in the matrix exponential to correctly account for them. This scenario may occur for some tracers, but for many passive tracers the time-scales associated with their dissolution in the ocean are relatively long [@siberlin2011oceanic]. The dissolution of gases, for example, could take a long time to reach equilibrium, from decades to centuries.
Conclusions {#sec:conclusions}
===========
In this work, we developed an ETD solver for the tracer equations appearing in ocean modeling. The linear operator has been split in a vertical and horizontal part, the former was treated with a matrix exponential, whereas the latter was handled explicitly. The need to treat the vertical terms exponentially is due to the fast time scales that govern these terms for instance in the case of eddies or when two bodies of water with different temperatures meet. The ETD scheme was implemented using two methods to compute the $\varphi_k$-functions, i.e. the Krylov subspace method presented in [@eiermann2006restarted], and a scaling and squaring method that we developed to evaluate $\varphi_k$-functions with $k \ge 0$. The proposed time-stepping scheme has been compared with the semi-implicit scheme RK4 + implicit Euler and a significant speed-up were observed for both implementations, due to much bigger time-step sizes that could be taken with the ETD method. The case of multiple tracer equations was also addressed. Exploiting the fact that all these equations have the same linear part, the matrices $\varphi_1(J_n^{z,i})$ could be computed for one tracer and then re-used for the others. This approach resulted in a significant advantage in terms of computational time, even up to a $82.04$% gain over an implementation where the matrices were evaluated for every tracer. Finally, we coupled the tracer equations with the dynamics system to make comparisons with other ocean models. Two benchmark tests were performed and both showed that the results obtained with the proposed ETD scheme were comparable with the ones obtained with existing ocean models.
Future work will be on developing a local ETD time-stepping scheme, where different time-steps are used on different sub-domains, depending on their associated time-scales. As discussed in remark \[new\_option\], another interesting would be investigating another choice of the linear operator used in -, i.e. instead of having a time dependent $J_n^z$, a fix $J^z_{m}$ may be used starting at some instant of time $t_m$. A scaling and squaring method would particularly benefit from this choice, since in this context the assembly cost of matrices is not an issue and we can directly reuse the old propagator $\varphi_1(\Delta{t} J^z_{m}$). Then, the term $J_n^z - J^z_{m}$ is part of the remainder, which would be not stiff as long as $\Delta{t} < \abs{w(t_m) - w(t_n)} / \Delta{z}$.
Acknowledgments {#acknowledgments .unnumbered}
===============
This work was supported by the US Department of Energy Office of Science under grants DE-SC0016591 and DE-SC020418, by the Fog Research Institute under contract no. FRI-454, and in part, by UT-Battelle, LLC, under Contract No. DE-AC05-00OR22725 with the U.S. Department of Energy. Accordingly, the U.S. Government retains a non-exclusive, royalty-free license to publish or reproduce the published form of this contribution, or allow others to do so, for U.S. Government purposes.
[^1]: Department of Scientific Computing, Florida State University, Tallahassee FL 32306, USA.
[^2]: Computer Science and Mathematics Division, Oak Ridge National Laboratory, One Bethel Valley Road, P.O. Box 2008, MS-6211, Oak Ridge, TN 37831, USA.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
Feature construction can substantially improve the accuracy of Machine Learning (ML) algorithms. Genetic Programming (GP) has been proven to be effective at this task by evolving non-linear combinations of input features. GP additionally has the potential to improve ML explainability since explicit expressions are evolved. Yet, in most GP works the complexity of evolved features is not explicitly bound or minimized though this is arguably key for explainability. In this article, we assess to what extent GP still performs favorably at feature construction when constructing features that are (1) Of small-enough number, to enable visualization of the behavior of the ML model; (2) Of small-enough size, to enable interpretability of the features themselves; (3) Of sufficient informative power, to retain or even improve the performance of the ML algorithm. We consider a simple feature construction scheme using three different GP algorithms, as well as random search, to evolve features for five ML algorithms, including support vector machines and random forest. Our results on 21 datasets pertaining to classification and regression problems show that constructing only two compact features can be sufficient to rival the use of the entire original feature set. We further find that a modern GP algorithm, GP-GOMEA, performs best overall. These results, combined with examples that we provide of readable constructed features and of 2D visualizations of ML behavior, lead us to positively conclude that GP-based feature construction still works well when explicitly searching for compact features, making it extremely helpful to explain ML models.\
\
address:
- 'Life Sciences and Health Group, Centrum Wiskunde & Informatica, Amsterdam 1098 XG, the Netherlands'
- 'Department of Radiation Oncology, Amsterdam UMC, Amsterdam 1105 AZ, the Netherlands'
- 'Life Sciences and Health Group, Centrum Wiskunde & Informatica, Amsterdam 1098XG, the Netherlands'
- 'Algorithmics Group, Delft University of Technology, Delft 2628 XE, the Netherlands'
author:
- Marco Virgolin
- Tanja Alderliesten
- 'Peter A.N. Bosman'
bibliography:
- 'featureconstructiongomea.bib'
title: |
On Explaining Machine Learning Models\
by Evolving Crucial and Compact Features
---
feature construction,interpretable machine learning ,genetic programming ,GOMEA
Introduction
============
Feature selection and feature construction are two important steps to improve the performance of any Machine Learning (ML) algorithm [@liu1998feature; @friedman2001elements]. Feature selection is the task of excluding features that are redundant or misleading. Feature construction is the task of transforming (parts of) the original feature space into one that the ML algorithm can better exploit.
A very interesting method to perform feature construction automatically is Genetic Programming (GP) [@koza1992gp; @poli2008field]. GP can synthesize functions without many prior assumptions on their form, differently from, e.g., logistic regression or regression splines [@friedman1991multivariate; @hosmer2013applied]. Moreover, feature construction not only depends on the data at hand, but also on the way a specific ML algorithm can model that data. Evolutionary methods in general are highly flexible in their use due to the way they perform search (i.e., derivative free). This makes it possible, for example, to evaluate the quality of a feature for a specific ML algorithm by directly measuring what its impact is on the performance of the ML algorithm (i.e., by training and validating the ML algorithm when using that feature).
![Regression surface learned by SVM for the Yacht dataset (in blue), expressed as a 2D function of the two features (on the bottom axes) constructed by our approach. Circles are training samples, diamonds are test samples. The dataset has six features ($x^{(i)}$). Our approach constructs two new features (using GP-GOMEA, see Sec. \[sec:details-gp\]), which are non-linear transformations of the prismatic coefficient ($x^{(2)}$) and the Froude number ($x^{(6)}$). With only two features the SVM prediction surface can be visualized. Moreover, these new features are understandable. Finally, the modeling quality is actually improved over employing SVM directly on all six features. The coefficient of determination of SVM increased from 85% using the original features to 98% using the two new features.[]{data-label="fig:svm-yacht"}](SVMr_yacht_NEW){width="0.99\linewidth"}
Explaining what constructed features mean can shed light on the behavior of ML-inferred models that use such features. Reducing the number of features is also important to improve interpretability. If the original feature space is reduced to few constructed features (e.g., up to two for regression and up to three for classification), the function learned by the ML model can be straightforwardly visualized w.r.t. the new features. In fact, how to make ML models more understandable is a key topic of modern ML research, as many practical, sensitive applications exist, where explaining (part of) the behavior of ML models is essential to trust their use (e.g., in medical applications) [@lipton2018mythos; @guidotti2018survey; @adadi2018peeking; @goodman2017european]. Typically, GP for feature construction searches in a subspace of mathematical expressions. Adding to the appeal and potential of GP, these expressions can be human-interpretable if simple enough [@guidotti2018survey; @virgolin2019model].
Figure \[fig:svm-yacht\] presents an example of the potential held by such an approach: a multi-dimensional dataset transformed into a 2D one, where both the behavior of the ML algorithm and the meaning of the new features is clear, while the performance of the ML algorithm is not compromised w.r.t. the use of the original feature set (it is actually improved).
In this article we study whether GP can be useful to construct a *low* number of *small* features, to increase the chance of obtaining interpretable ML models, without compromising their accuracy (compared to using the original feature set). To this end, we design a simple, iterative feature construction scheme, and perform a wide set of experiments: we consider four types of feature construction methods (three GP algorithms and random search), five types of machine learning algorithms. We apply their combinations on 21 datasets between classification and regression to determine to what extent they are capable of effectively and efficiently finding crucial and compact features for specific ML algorithms.
The main original scientific contribution of this work is an investigation of whether GP can be used to construct features that are:
- Of small-enough number, to enable visualization of the behavior of the ML model;
- Of small-enough size, to enable interpretability of the features themselves;
- Of sufficient informative power, to retain or even improve the performance of the ML, compared to using the original feature set;
These aspects are assessed under different circumstances:
- We test different search algorithms, including modern model-based GP and random search;
- We test different ML algorithms.
The remainder of this article is organized as follows. Related work is reported in Section \[sec:related\]. The proposed feature construction scheme is presented in Section \[sec:scheme\]. The search algorithms to construct features, as well as the considered ML algorithms, are presented in Section \[sec:algos\]. The experimental setup is described in Section \[sec:experiments\]. Results related to performance are reported in Section \[sec:result-rq12\], while results concerning interpretability are reported in Section \[sec:result-rq3\]. Section \[sec:discussion\] discusses our findings, and Section \[sec:conclusion\] concludes this article.
Related work {#sec:related}
============
In this article, we consider GP for feature construction to achieve better explainable ML models. Different forms of GP to obtain explainable ML have been explored in literature, but they do not necessarily leverage feature construction. E.g., [@cano2013interpretable] introduced a form of GP for the automatic synthesis of interpretable classifiers, generated from scratch as self-contained ML models, made of IF-THEN rules. A very different paradigm for explainable ML by GP is considered in [@evans2019s], where the authors explore the use of GP to recover the behavior of a given unintelligible classifier by evolving interpretable approximation models. Other GP-based approaches and paradigms to synthesize interpretable ML models from scratch, or to approximate the behavior of pre-existing ML models by interpretable expressions, are reported in recent surveys on explainable artificial ingelligence such as [@guidotti2018survey; @adadi2018peeking].
Since in this article we particularly study what the potential of GP for feature construction is in terms of added value for explaining complex, not directly explainable models learned by various popular ML algorithms, the related work that follows describes GP approaches for feature construction. For readers interested in feature selection, we refer to a recent survey [@xue2016survey].
One of the first approaches of GP for feature construction is presented in [@krawiec2002genetic]. There, each GP solution is a set of $K$ features. The fitness of a set is the cross-validation performance of a decision tree [@breiman2017classification] using that set. The results on six classification datasets show that the approach is able to synthesize a feature set that is competitive with the original one, and can also be added to the original set for further improvements. No attention is however given to the interpretability of evolved features.
The work in [@muharram2005evolutionary] generates one feature with Standard, tree-based GP (SGP) [@koza1992gp], to be added to the original set. Feature importance metrics of decision trees such as information gain, Gini index and $Chi^2$ are used as fitness measure. An advantage of using such fitness measures over ML performance is that they can be computed very quickly. However, they are decision tree-specific. Results show that the approach can improve prediction accuracy, and, for a few problems, it is shown that decision trees that are simple enough to be reasonably interpretable, can be found.
Feature construction for high-dimensional datasets is considered in [@tran2016genetic], for eight bio-medical binary classification problems, with 2,000 to 24,188 features. This approach is different from the typical ones, as the authors propose to use SGP to evolve classifiers rather than features, and extract features from the components (subtrees) of such classifiers. These are then used as new features for an ML algorithm. Results on K-Nearest Neighbors [@altman1992introduction], Naive Bayes classifier [@russell2016artificial; @murphy2006naive], and decision tree show that a so-found feature set can be competitive or outperform the original one. The authors show an example where a single interpretable feature is constructed that enables linear separation of the classification examples.
Different from the aforementioned works, [@chen2017genetic] explores feature construction for regression. A SGP-based approach is designed to tackle regression problems with a large number of features, and is tested on six datasets. Instead of using the constructed features for a different ML algorithm, SGP dynamically incorporates them within an ongoing run, to enrich the terminal set. Every $\alpha$ generations of SGP, the subtrees composing the best solutions become new features by encapsulation into new terminal nodes. The approach is found to improve the ability of SGP to find accurate solutions. However, the features found by encapsulating subtrees are not interpretable because allowing subsequent encapsulations leads to an exponential growth of solution size.
A recent work that focuses on evolutionary dimensionality reduction and consequent visualization is [@cano2017multiobjective], where a multi-objective, grammar-based SGP approach is employed. $K$ feature transformations are evolved in synergy to enable, at the same time, good classification accuracy, and visualization through dimensionality reduction. The system is thoroughly tested on 42 classification tasks, showing that the algorithm performs well compared to state-of-the-art dimensionality reduction methods, and it enables visualization of the learned space. However, as trees are free to grow up to a height of 50, the constructed features themselves cannot be interpreted.
The most similar works to ours that we found are [@virgolin2018gecco] and [@tran2019genetic]. In [@virgolin2018gecco], which is our previous work, the possibility of using a modern model-based GP algorithm (which we also use in our comparisons) for feature construction is explored on four regression datasets. There, focus is put on keeping feature size small, to actively attempt to obtain readable features. These features are iteratively constructed to be added to the original feature set to improve the performance of the ML algorithm, and three ML algorithms are compared (linear regression, support vector machines [@cortes1995support], random forest [@breiman2001random]). Reducing the feature space to enable a better understanding of inferred ML models is not considered.
In [@tran2019genetic], different feature construction approaches are compared on gene-expression datasets that have a large number of features (thousands to tens of thousands) to study if evolving class-dependent features, i.e., features that are each targeted at aiding the ML algorithm detect one specific class, can be beneficial. Similarly to us, the authors show visualizations of feature space reduced to up to three constructed features, and an example of three features that are encoded as very small, easy-to-interpret trees. However, such small features are a rare outcome as the trees used to encode features typically had more than 75 nodes. These trees are therefore arguably extremely hard to read and interpret.
Our work is different from previous research in two major aspects. First, none of the previous work principally addresses the conflicting objectives of retaining good performance of an ML algorithm while attempting to explain both its behavior (by dimensionality reduction to allow visualization), and the meaning of the features themselves (by constraining feature complexity). Second, multiple GP algorithms within a same feature construction scheme, on multiple ML algorithms, are not compared in previous work. Most of the times, it is a different feature construction scheme that is tested, using arguably small variations of SGP. Here, we consider random search, two versions of SGP, as well as another modern GP algorithm. Furthermore, we adopt both “weak” ML algorithms such as ordinary least squares linear regression and the naive Bayes classifier, as well as “strong”, state-of-the-art ones, which are rarely used in literature for feature construction, such as support vector machine and random forest; on both classification and regression tasks.
Iterative evolutionary feature construction {#sec:scheme}
===========================================
We use a remarkably simple scheme to construct features. Our approach constructs $K \in \mathbb{N}^+$ features by iterating $K$ GP runs. The evolution of the $k$-th feature ($k \in \{1, \dots, K\}$) uses the previously constructed $k-1$ features.
Feature construction scheme
---------------------------
The dataset $D$ defining the problem at hand is split into two parts: the training $Tr$ and the test $Te$ set. This partition is kept fixed through the whole procedure. Only $Tr$ is used to construct features, while $Te$ is exclusively used for final evaluation to avoid positive bias in the results [@kohavi1997wrappers]. We use the notation $x^{(i)}_j$ to refer to the $i$-th feature value of the $j$-th example, and $y_j$ for the desired outcome (label for classification or target value for regression) of the $j$-th example.
The $k$-th GP run evolves the $k$-th feature. An example is shown in Figure \[fig:featureevolutionscheme\]. Each solution in the population competes to become the new feature $x^{(k)}$, that represents a transformation of the original feature set. In every run, the population is initialized at random.
We evaluate the fitness of a feature of the $k$-th run by measuring the performance of the ML algorithm on a dataset that contains that feature and the previously evolved $k-1$ features.
We only use original features (and random constants) as terminals. In particular, the features constructed by previous iterations are not used as terminal nodes in the $k$-th run. This prevents the generation of nested features, which could harm interpretability.
At the end of the $k$-th run, the best feature is stored and its values $x^{(k)}_j$ are added to $Tr$ and $Te$ for the next iterations.
Feature fitness {#sec:feature-fitness}
---------------
The fitness of a feature is computed by measuring the performance (i.e., error) of the ML algorithm when the new feature is added to $Tr$. We consider the $C$-fold cross-validation error rather than the training error to promote generalization and prevent overfitting. The pseudo code of the evaluation function is shown in Algorithm \[alg:feature-fitness-computation\].
Specifically, the $C$-fold cross-validation error is computed by partitioning $Tr$ into $C$ splits. For each $c = 1, \dots, C$ iteration, a different split is used for validation (set $V^c$), and the remaining $C-1$ splits are used for training (set $Tr^c$). The mean validation error is the final result.
For classification tasks, in order to take into account both multiple and possibly imbalanced class distributions, the prediction error is computed as 1 minus the *macro* $F1$ score, i.e., 1 minus the mean of the class-specific $F1$ scores:
$$\begin{aligned}
1 - F1 & = 1 - \frac{1}{ \# \textit{classes} } \sum_{ \gamma \in \textit{classes} } F1_\gamma
\\
& = 1 - \frac{2}{ \# \textit{classes} } \sum_{ \gamma \in \textit{classes} } \frac{ \frac{TP_\gamma}{TP_\gamma + FP_\gamma} \frac{TP_\gamma}{TP_\gamma + FN_\gamma} }{ \frac{TP_\gamma}{TP_\gamma + FP_\gamma} + \frac{TP_\gamma}{TP_\gamma + FN_\gamma} },\end{aligned}$$
where $TP_\gamma, FN_\gamma, FP_\gamma$ are the true positive, false negative, and false positive classifications for the class $\gamma$, respectively. If the computation of $F1_\gamma$ results in $\frac{0}{0}$, we set $F1_\gamma = 0$.
For regression, the prediction error is computed with the Mean Squared Error (MSE).
$Tr^\prime \gets $AddFeatureToCurrentTrainingSet($s$) $error \gets 0$ $T^c, V^c \gets $SplitSet($c,C,Tr^{\prime}$) $M \gets $TrainMLModel($T^c$) $error \gets error + $ComputeError($M, V^c$) Return$\left( \frac{error}{C} \right) $
Preventing unnecessary fitness computations {#sec:criteriaevaluation}
-------------------------------------------
Computing the fitness of a feature is particularly expensive, as it consists of a $C$-fold cross-validation of the ML algorithm. This limits the feasibility of, e.g., adopting large population sizes and large numbers of evaluations for the GP algorithms.
We therefore attempt to prevent unnecessary cross-validation calls, by assessing if features meet four criteria. Let $n$ be the number of examples in $Tr$. The criteria are the following:
1. *The feature is not a constant*. We avoid evaluating constant features as they are likely to be useless for many ML algorithms, which internally already compute an intercept.\
2. *The feature does not contain extreme values that may cause numerical errors*, i.e., with absolute value above a lower-bound $\beta_\ell$ or above an upper-bound $\beta_u$. Here, we set $\beta_\ell = 10^{-10}$, and $\beta_u = 10^{10}$ (none of the datasets considered here have values exceeding these bounds).\
3. *The feature is not equivalent to one constructed in the previous $k-1$ iterations*. Equivalence is determined by checking the values available in $Tr$, i.e., equivalence holds if:
$$\forall j \in Tr, \exists i \in \{1, \dots, k-1\} : x^{(k)}_j = x^{(i)}_{j}.$$
Note that a constructed feature that is equivalent to a feature of the original feature set can be valid, as long as no other previously constructed feature exists that is already equivalent. Thus, our approach can in principle perform pure feature selection.\
4. *The values of the feature in consideration have changed since the last time the feature was evaluated*. GP variation can change the syntax of a feature without necessarily affecting its behavior (e.g., inserting a multiplication by 1 will not change the final values a feature computes). If the values do not change, then the fitness of the feature will not change either (see Sec. \[sec:feature-fitness\]). We therefore avoid unnecessary re-computations of feature fitnesses, by caching the feature values prior to GP variation, and checking whether they have changed after variation.
The computational effort for each criterion is $O(n)$ (it is $O((k-1)n)$ for criterion 3, however in our experiments $k \ll n$). The fitness of a feature failing criterion 1, 2, or 3 is set to the maximum possible error value. If criterion 4 fails, the fitness remains the same (although performing cross-validation may lead to slightly different results when using stochastic ML algorithms like random forest).
Considered search algorithms and machine learning algorithms {#sec:algos}
============================================================
We consider SGP, Random Search (RS), and the GP instance of the Gene-pool Optimal Mixing Evolutionary Algorithm (GP-GOMEA) as competing search algorithms to construct features. SGP is widely used in feature construction (see related work in Sec. \[sec:related\]). RS is not typically considered, yet we believe it is important to assess whether evolution does bring any benefit over random enumeration within the confines of our study, i.e., when forcing to find small features. GP-GOMEA is a recently introduced GP algorithm that has proven to be particularly proficient in evolving accurate solutions of limited size [@virgolin2019model; @virgolin2017scalable; @virgolin2018gecco].
As ML algorithms, we consider the Naive Bayes classifier (NB), ordinary least-squares Linear Regression (LR), Support Vector Machines (SVM), Random Forest (RF), and eXtreme Gradient Boosting (XGB). NB is used only for classification tasks, LR only for regression tasks, SVM, RF, and XGB for both tasks. We provide more details in the following sections.
Details on the search algorithms {#sec:details-gp}
--------------------------------
All search algorithms use the fitness evaluation function. A feature $s$ is evaluated by first checking whether the four criteria of Section \[sec:criteriaevaluation\] are met, and then, if the outcome is positive, by running the ML algorithm over the feature-extended dataset.
For SGP, we use subtree crossover and subtree mutation, picking the depth of subtree roots uniformly randomly as proposed in [@pawlak2015semantic]. The candidate parents for variation are chosen with tournament selection. Since we are interested in constructing small features so as to increase the chances they will be interpretable, we consider two versions of SGP. The first is the classic one where solutions are free to grow to tree heights typically much larger than the one used for tree initialization. In the following, the notation SGP refers to this first version. The second one uses trees that are not allowed to grow past the initial maximum tree height. We call this version *bounded* SGP, and use the notation SGP~b~.
RS is realized by continuously sampling and evaluating new trees, keeping the best [@koza1992gp]. Like for SGP~b~, a maximum tree height is fixed during the whole run. If evolution is hypothetically no better than RS, then we expect that SGPb and GP-GOMEA will construct features that are no better than the ones constructed by RS.
GP-GOMEA is a recently introduced GP algorithm that has been found to deliver accurate solutions of small size on benchmark problems [@virgolin2017scalable], and to work well when a small size is enforced in symbolic regression [@virgolin2019model; @virgolin2018gecco]. GP-GOMEA uses a tree template fixed by a maximum tree height (which can include intron nodes to allow for unbalanced tree shapes) and performs homologous variation, i.e., mixed tree nodes come from the same positions in the tree. Each generation prior to mixing, a hierarchical model that captures interdependencies (*linkage*) between nodes is built (using mutual information). This model, called Linkage Tree (LT), drives variation by indicating what nodes should be changed *en block* during mixing, to avoid the disruption of patterns with large linkage.
The LT has been shown to enable GP-GOMEA to outperform subtree crossover and subtree mutation of SGP, as well as the use of a randomly-build LT, i.e., the Random Tree (RT), on problems of different nature [@virgolin2019model; @virgolin2017scalable]. However, the LT requires sufficiently large population sizes to be accurate and beneficial (e.g., several thousand solutions in GP for symbolic regression) [@virgolin2019model]. Because in the framework of this article fitness evaluations use the cross-validation of a ML algorithm, we cannot afford to use large population sizes. Accordingly, we found the adoption of the LT to not be superior to the adoption of the RT under these circumstances in preliminary experiments. Therefore, for the most part, we adopt GP-GOMEA with the RT (GP-GOMEA~RT~). This means we effectively compare random hierarchical homologous variation with subtree-based variation. An example of adopting the LT and large population sizes for feature construction is provided in Section \[sec:discussion\].
Details on the ML algorithms {#sec:details-ML algorithms}
----------------------------
We now briefly describe the ML algorithms used in this work: NB, LR, SVM, RF, and XGB. NB and LR are less computationally expensive compared to SVM, RF, and XGB. Details on the computational time complexity of these algorithms are reported at: https://bit.ly/2PG0xse.
NB is a classifier which assumes independence between features [@russell2016artificial; @murphy2006naive]. NB is often used as a baseline, as it is simple and fast to train. We use the *mlpack* implementation of NB [@mlpack2013] and assume the data to be normally distributed (default setting).
Similarly to NB, LR is often used as a baseline as it is simple and fast, for regression tasks. LR assumes that the target variable can be explained by a linear combination of the features [@russell2016artificial]. We use the *mlpack* implementation of LR [@mlpack2013].
SVM is a powerful ML algorithm that can be used for non-linear classification and regression [@cortes1995support; @CC01a]. We use the *libsvm* implementation [@CC01a]. We consider the Radial Basis Function (RBF) kernel, which works well in practice for many problems, with C-SVM for classification, and $\mathcal{E}$-SVM for regression.
RF is an ensemble ML algorithm which, like SVM, can be used for both classification and regression and can infer non-linear patterns [@breiman2001random]. RF builds an ensemble of (typically deep) decision trees, each trained on a sample of the training set (*bagging*). At prediction time, the mean (or maximum agreement) prediction of the decision trees is returned. We use the *ranger* implementation [@wright2015ranger].
XGB is, like RF, an ensemble ML algorithm, typically based on decision trees, and capable of learning non-linear models [@chen2016xgboost]. XGB works by boosting, i.e., stacking together multiple weak estimators (small decision tress) that fit the data in an incremental fashion. We use the *dmlc* implementation (https://bit.ly/34fBNeA).
Experiments {#sec:experiments}
===========
We perform 30 runs of our Feature Construction Scheme (FCS), with SGP, SGP~b~, RS, and GP-GOMEA~RT~, in combination with each ML algorithm (NB only for classification and LR only for regression), on each problem. Each run of the FCS uses a random train-test split of 80%-20%, and considers up to $K=5$ features construction rounds. We use a population size of 100 for the search algorithms, and assign a maximum budget of $10,000$ function evaluations to each FCS iteration. This results in relatively large running times for complex ML algorithms (see Sec. \[sec:running-time\]). An experiment including larger evolutionary budgets and the use of the LT in GP-GOMEA is presented in the discussion (Sec. \[sec:discussion\]). We use a limit on the total number of evaluations instead of a a limit on the total number of generations because GP-GOMEA~RT~ performs more evaluations than SGP per generation [@virgolin2017scalable].
For GP-GOMEA~RT~, SGP~b~, and RS, we consider two levels of maximum tree height $h$: 2 and 4. This choice yields a maximum solution size of 7 and 31 respectively (using function nodes with a maximum arity $r=2$). We choose these two height levels because we found features with $h=2$ to be arguably easy to read and interpret, whereas features with $h=4$ can already be very hard to understand. This indication is also reported in [@virgolin2019model] for the evolution of symbolic regression formulas. Note that using a tree height limit over a solution size limit prevents finding deep trees containing the nesting of the arguably more complicated to understand non-linear functions $\cdot^2, \sqrt{\cdot}, \log_p, \exp$. We do not consider bigger tree heights as resulting features may likely be impossible to interpret, defying a key focus of this work.
Other parameter settings used for the GP algorithms are shown in Table \[tab:ea-parameters\]. SGP~b~ uses the same settings as SGP, except for the maximum tree height (at initialization and along the whole run), which is set to the same of GP-GOMEA~RT~. In GP-GOMEA~RT~ we use the Half and Half (HH) tree initialization method instead of the Ramped Half and Half (RHH) [@koza1992gp] commonly used for SGP. This proved to be beneficial since GOM varies nodes instead of subtrees [@virgolin2019model; @virgolin2018gecco]. For both HH and RHH, syntactical uniqueness of solutions is enforced for up to 100 tries [@koza1992gp]. In GP-GOMEA~RT~ we additionally avoid sampling trees having a terminal node as root by setting the minimum tree height of the grow method to 1. This is not done for SGP and SGP~b~, because differently from GP-GOMEA~RT~ where homologous nodes are varied, subtree root nodes for subtree crossover (SX) and subtree mutation (SM) are chosen uniformly randomly. RS samples new trees using the same initialization method as SGP~b~, i.e., RHH.
The division operator $\div$ used in the function set is the analytic quotient operator ($ a \div b = a / \sqrt{1 + b^2}$), which was shown to lead to better generalization performance than protected division [@ni2013use]. The logarithm is protected $\log_p(\cdot) = \log( | \cdot | )$ and $\log_p(0)=0$, and so is the square root operator. The terminal set contains the original feature set, and an Ephemeral Random Constant (ERC) [@poli2008field] with values uniformly sampled between the minimum and maximum values of the features in the original training set, i.e., $[\min x^{(i)}_j, \max x^{(i)}_j], \forall i,j \in Tr$.
The hyperparameter settings for the SVM, RF and XGB are shown in Table \[tab:ML algorithm-parameters\], and are mostly default [@breiman2001random; @CC01a; @wright2015ranger] (for XGB, we referred to https://bit.ly/2JCM9x4). NB and LR implementations do not have hyperparameters.
We consider 10 classification and 10 regression benchmark datasets[^1] that can be considered traditional, i.e, they have small to moderate dimensionality (number of features). We mostly study this type of dataset because we seek to find small constructed features that can be interpreted. Hence, they can represent a transformation of only a limited number of original features. Details on the datasets are reported in Table \[tab:datasets\]. Rows with missing values are omitted. Most datasets are taken from the UCI Machine Learning repository[^2], with exception for Dow Chemical and Tower, which come from GP literature [@white2013better; @albinati2015effect].
We further consider a very high-dimensional dataset from UCI (https://bit.ly/334KbgW) to assess whether GP can still be useful to construct features in this type of scenario. The dataset in question concerns the classification of cancer type, given RNA-Seq gene expression levels as features. Five cancer class types are present, and class proportions in the data presents some unbalance: the class frequencies are 0.37, 0.18, 0.18, 0.17, 0.10. A total of $20,531$ features are considered, in $801$ examples. Since large computational resources are needed to handle this dataset, we consider only NB as ML algorithm for feature construction upon this data.
Results: performance on traditional datasets {#sec:result-rq12}
============================================
The results described in this section aim at assessing whether it is possible to construct few and small features that lead to an equal or better performance than the original set, and whether some search algorithms can construct better features than others.
General performance of feature construction {#sec:performancefcs}
-------------------------------------------
We begin by observing the dataset-wise aggregated performance of FCS for the different GP algorithms and the different ML algorithms, separately for classification and regression.
### Classification {#sec:result-rq12-class}
Figure \[fig:means\_median\_cls\] shows dataset-wise aggregated results obtained for NB, SVM, RF, and XGB, for the 10 traditional classification tasks. Each data point is the mean among the dataset-specific medians of macro F1 from the 30 runs.
In general, the use of only one constructed feature does not perform as good as the use of the original feature set. Constructing more features improves the performance, but with diminishing returns.
Specifically for NB, the use of two constructed features is already preferable to the use of the original feature set. This is likely due to the fact that NB assumes complete independence between the provided features, and this can be implicitly tackled by FCS. SGP (unbounded) is the best performing algorithm as it can evolve arbitrarily complex features, however, the magnitude of improvement of the macro F1 score with respect to GP-GOMEA~RT~ and SGP~b~ is limited. For $h=4$ and $K=5$, GP-GOMEA~RT~ reaches the performance of SGP. GP-GOMEA~RT~ is typically slightly better than SGP~b~, and RS has worse performance. Training and test F1 scores do not differ much for any feature construction algorithm, meaning that overfitting is not an issue for NB. Rather, compared to the other ML algorithms, NB underfits.
The performance of FCS for SVM has an almost identical pattern to the one observed for NB, except for the fact that the performance is found to be consistently better. However, for SVM it is preferable to use the original feature set rather than few constructed features. This is evident in terms of training performance, but less at test time. In fact, using only 5 constructed features leads to similar test performance compared to using the original set. The GP algorithms compare to each other similarly to when using NB. Compared to NB, it can be seen that SVM exhibits larger gaps between training and test results, suggesting that some overfitting takes place, especially when the original feature set is used.
The way performance improves for RF by constructing features is similar to the one observed for NB and SVM. However, for RF the differences between the search algorithms is particularly small: notice that using RS leads to close performance to the ones obtained by using the other GP algorithms, compared to the SVM case. Moreover, virtually no difference can be seen between GP-GOMEA~RT~ and SGP~b~. This suggests that RF already works well with less refined features. Now, the features constructed by SGP are no longer the best performing at test time. This is likely because SGP evolves larger, more complex features than the other algorithms (see Sec. \[sec:feature-size\]), making RF overfit. In fact, RF exhibits the largest difference between training and test results compared to NB and SVM, for any feature construction algorithm and $h$ limit. Still, the test results of RF are slightly better than the ones of SVM and markedly better than the ones of NB, meaning that the latter two are underfitting.
The training and test performance obtained when using XGB is similar to the one obtained when using RF, but the differences the between different search algorithms are even less marked than for RF. Some differences can be seen for $K=1$ on the training set (SGP better than GP-GOMEA~RT~, and GP-GOMEA~RT~ better than SGPb and RS), but this difference is much less marked on the test set. When more features are constructed, essentially all search algorithms deliver the same performance. XGB seems to be able to construct non-linear relationships even better than RF. As to potential overfitting, the trend of differences between training and test performance that can be observed for XGB mirrors the one visible for RF.
As to maximum tree height, allowing the constructed features to be bigger ($h=4$ vs $h=2$) moderately improves the performance. Interestingly, GP-GOMEA~RT~ with $h=4$ reaches competitive performance with SGP on all ML algorithms, despite the latter having no strict limitation on feature size.
### Regression {#sec:result-rq12-regr}
Results on the regression tasks are shown in Figure \[fig:means\_median\_regr\], dataset-wise aggregated for LR, SVM, RF, and XGB. We report the results in terms of coefficient of determination, i.e., $R^2(y, \bar{y}) = 1 - MSE(y, \bar{y})/var(y)$. For the four ML algorithms, results overall follow the same pattern. SGP is typically better, especially for LR and SVM, although constructing more features reduces the performance gap with the other GP algorithms. GP-GOMEA~RT~ is slightly, yet consistently, the best performing within the maximum tree height limitation of 2, while SGP~b~ is visibly preferable only when a single feature is constructed for LR and SVM, for $h=4$. Differently from the classification case, two features are typically enough to reach the performance of the original feature set for all ML algorithms except for XGB. Moreover, for LR, SVM, and RF, the performance between training and test is similar, meaning no considerable overfitting is taking place, no matter the feature construction algorithm used nor the limit of $h$. This however is not the case for XGB, where a large performance gap is encountered. Still, the test performance obtained when using XGB is ultimately slightly better than the obtained for RF.
As for classification, allowing for larger trees results in better performance overall, and reduces the gap between SGP and the other GP algorithms. With XGB, all search algorithms perform similarly.
### Feature size {#sec:feature-size}
Figure \[fig:size-all\] shows the aggregated feature size for the different GP algorithms and RS. The aggregated solution size is computed by taking the median solution size per run, then averaging over datasets, and finally averaging over ML algorithms (classification and regression are considered together). The picture shows how, overall, the known SGP tendency to bloat differs compared to the algorithms working with a strict tree height limitation. SGP features are so large that it is nearly impossible to interpret them (see Sec. \[sec:interpretability\]).
RS finds the smallest features for both height limits $h=2$ and $h=4$. Considering that GP-GOMEA~RT~ and SGP~b~ generate trees within the same height bounds of RS, we conclude that it is the variation operators that allow finding larger trees with improved fitness within the height limit. GP-GOMEA~RT~ seems to construct slightly, yet consistently, larger trees than SGP~b~.
For SGP, it can be seen that subsequently constructed features are smaller (this is barely visible for GP-GOMEA~RT~ and SGP~b~ as well). This is interesting because we do not use any mechanism to promote smaller trees. This result is likely linked to the diminishing returns in performance observed in Figure \[fig:means\_median\_cls\] and \[fig:means\_median\_regr\]: constructing new complex and informative features becomes harder with the number of FCS iterations.
=0.02mm
[lc]{}
Feature size
& ![Aggregated feature size for $k = 1,\dots,5$. Solid (dotted) lines represent solution size for maximum tree height $h=2$ ($h=4$). Shaded areas represent standard deviation. SGP is free to grow solutions up to $h=17$.[]{data-label="fig:size-all"}](size_all "fig:"){width="0.65\linewidth"}\
& ![Aggregated feature size for $k = 1,\dots,5$. Solid (dotted) lines represent solution size for maximum tree height $h=2$ ($h=4$). Shaded areas represent standard deviation. SGP is free to grow solutions up to $h=17$.[]{data-label="fig:size-all"}](legend2 "fig:"){width="0.65\linewidth"}
Statistical significance: comparing GP algorithms
-------------------------------------------------
The aggregated results of Section \[sec:performancefcs\] show moderate differences between GP-GOMEA~RT~ and SGP~b~. These are arguably the most interesting algorithms to compare in-depth, as they are able to construct small features that lead to good performance (RS typically constructs less informative features, while SGP constructs very large ones).
We perform statistical significance tests to compare GP-GOMEA~RT~ and SGP~b~. We consider their median performance on the test set $Te$, obtained by the FCS, and also compare it with the use of the original feature set, for each ML algorithm and each dataset. In our case, the *treatments* of our significance tests are the two search algorithms (i.e., GP-GOMEA~RT~ and SGP~b~) and the original feature set, while the *subjects* are the configurations given by pairing ML algorithms and datasets [@demvsar2006statistical].
We first perform a Friedman test to assess whether differences exists among the use of different treatments (GP algorithms and original feature set) upon multiple subjects (ML algorithm-dataset combinations). As post-hoc analysis, we use the pairwise Wilcoxon signed rank tests, paired by subject (ML algorithm-dataset combination), to see how the treatments compare to each other [@demvsar2006statistical]. We adopt the Holm correction method to prevent reporting false positive results that might have happened due to pure chance [@holm1979simple].
We consider both $h=2$ and $h=4$, and focus on $K=2$, since consideration of only two constructed features makes interpretation easier, and allows human visualization (see Sec. \[sec:interpretability\]).
### Classification {#classification}
For both $h=2,4$, the Friedman test strongly indicates differences between GP-GOMEA~RT~, SGP~b~, and the use of the original feature set ($p\text{-value} \ll 0.05$).
Figure \[fig:signiftests\] (top) shows the Holm-corrected $p$-values obtained by the pairwise Wilcoxon tests for classification, where the alternative hypothesis is that the row allows for larger macro F1 scores than the column. No significant differences between GP-GOMEA~RT~ and SGP~b~ are found for both $h=2,4$. Both the GP algorithms can deliver constructed features that are competitive with the use of the original feature set. The original feature set is not significantly better than using feature construction. Moreover, for GP-GOMEA~RT~ and $h=4$, the hypothesis that feature construction is *not* better than the original feature set can be rejected with a corrected $p$-value below 0.1. The latter result appears to be in contrast with the results from Fig. \[fig:means\_median\_cls\] for SVM, RF and XGB, where it can be seen that the construction of only two features does, on average, lead to slightly worse test results than using the original feature set. Nonetheless, the opposite is true for NB, and with rather large magnitude. A more in-depth analysis on this is provided in Sec. \[sec:statanal-per-mla\].
### Regression {#regression}
As for classification datasets, the Friedman test indicates that differences are presents between the treatments. Figure \[fig:signiftests\] (bottom) shows the Holm-corrected $p$-values obtained by the pairwise Wilcoxon tests for regression.
The statistical tests confirm the hypothesis that the algorithms are capable of providing constructed features that are more informative than the original feature set, as observed in Fig. \[fig:means\_median\_regr\] for the regression datasets. Now, GP-GOMEA~RT~ is significantly better than SGP~b~ when $h=2$. For $h=4$, instead, GP-GOMEA~RT~ is not found to be significantly better than SGP~b~.
=0.02mm
Statistical significance: two constructed features vs. the original feature set per ML algorithm {#sec:statanal-per-mla}
------------------------------------------------------------------------------------------------
Results presented in Sec. \[sec:performancefcs\] indicate that our FCS brings most benefit if used with the weak ML algorithms. We now report, for each ML algorithm, on how many datasets 2 features constructed using GP-GOMEA (with $h=2$ and $h=4$) lead to statistically significantly (using Holm-corrected pairwise Wilcoxon test, $p\text{-value} < 0.05$) better, equal, or worse results compared to using the original feature set on the test set. This is shown in Table \[tab:comp-mla-problems\].
These results confirm what seen in Figures \[fig:means\_median\_cls\] and \[fig:means\_median\_regr\]. Using FCS typically outperforms the use of the original feature set for the weak ML algorithms. For the strong ML algorithms, in most cases, using the original feature set is preferable. However, for some datasets reducing the space to two compact features without compromising performance is still possible.
The use of the original feature set is generally hardest to beat when adopting RF or XGB. For RF, in the regression case with $h=4$, FCS brings benefits on the datasets Airfoil, Energy Cooling, Energy Heating, and Yacht; and performs on par with the use of the original feature set on the datasets Boston Housing and Concrete. These datasets are the ones with the smallest number of original features. We find similar results for SVM and for XGB. In the latter case, FCS is, in terms of statistical significance, equal to the original feature set only on Energy Cooling, Energy Heating, and Yacht. It is reasonable to expect that FCS works well when few features can be combined.
In the classification case, findings are different. For RF and $h=4$, the datasets where using two constructed features bring similar or better results than using the original feature set are Breast Cancer Wisconsin and Iris. The latter does have a small number of original features (4), but the former has more than several other datasets (29). Furthermore, the datasets where FCS helps are different for SVM: FCS performs equally good to the original feature set on Iris and Cylinder Bands (39 features), and better on Madelon (500 features) and Image Segmentation (19 features). Regarding XGB, there is no dataset where FCS is superior to the original feature set, but it is also not worse on almost half of the datasets. For classification datasets, we cannot conclude that a small cardinality of the original feature set is a good indication feature construction will work well. Furthermore, feature construction influences different ML algorithms in different ways.
Results: performance on a highly-dimensional dataset {#sec:gene-expr}
====================================================
We further consider the RNA-Seq cancer gene expression dataset, comparing FCS by GP-GOMEA~RT~ with $h=4$ against the use of the original feature set, when using NB. Fig. \[fig:geneexpression\] shows that NB with the original feature set overfits: the training performance is maximal, while the test performance reaches an F1 of approximately 0.65. Even tough NB is typically considered a weak estimator, the system described by the data is so severely underdetermined (over $20,000$ features vs less than $1,000$ examples) that actual patterns cannot be retrieved. The use of FCS forces NB to use only a small number of constructed features, which, in turn, can contain only a small number of the original features. Essentially, FCS provides both the advantages of feature construction and feature selection. This leads to large F1 scores already when solely two features are constructed.
![Comparison between the use of the original feature set and FCS with GP-GOMEA~RT~ ($h=4$) on high-dimensional gene expression data. The vertical axis reports the median F1 score, the horizontal axis reports the number of features constructed by FCS. Stars indicate statistical significant superiority ($p\text{-value} < 0.05$) of one method w.r.t. the other.[]{data-label="fig:geneexpression"}](geneexpression_fcs){width="0.65\linewidth"}
Results: improving interpretability {#sec:result-rq3}
===================================
The results presented in Sec. \[sec:result-rq12\] and \[sec:gene-expr\] showed that the original feature set can be already outperformed by two small constructed features in many cases. We now aim at assessing whether constraining features size can enable interpretability of the features themselves, as well as if extra insight can be achieved by plotting and visualizing the behavior of a trained ML model in the new two-dimensional space.
Interpretability of small features {#sec:interpretability}
----------------------------------
Table \[tab:features-examples\] shows some examples of features constructed by GP-GOMEA~RT~, for $h=2$ and $h=4$. We report the first feature constructed for the $K=2$ case, with median test performance. We show the first feature as it is typically not smaller than the second (see Fig. \[fig:size-all\]). Analytic quotients and protected logarithms are replaced by their respective definitions. We remark that we do not check whether the meaning of the features is sound (e.g., ensuring a certain unit of measure is returned). Constraining feature meaning is problem-dependent, and outside the scope of this work.
For classification, we choose NB as it is the method which benefits most from feature construction. The dataset considered is Ecoli, where NB achieves the largest median test improvement when $K=2$: from $F1=0.51$ with the original set, to $F1=0.63$ for $h=2$, and to $F1=0.66$ for $h=4$.
For regression, we consider LR on the Concrete dataset, for the same aforementioned reasons. The test $R^2$ obtained with the original feature set is $0.59$, the one with two features constructed by GP-GOMEA~RT~ is $0.76$ ($0.78$) for $h=2$ ($h=4$).
For $h=2$, we argue that constructed features are mostly easy to interpret. For example, the feature shown for LR on Concrete tells us that aging ($x^{(8)}$) has a negative impact on concrete compressive strength, whereas using more water ($x^{(4)}$) than cement ($x^{(1)}$) has a positive effect (both features are in $\text{kg/cm}^3$). The impact of other features is less important (within the data variability of the dataset). For $h=4$, some features can be harder to read and understand, however many are still accessible. This is mostly because, even though the total solution size reachable with $h=4$ is 31, constructed features are typically half the size (see Fig. \[fig:size-all\]).
The features constructed for the RNA-Seq gene-expression dataset by GP-GOMEA~RT~ in Sec. \[sec:gene-expr\] are also not excessively complex to be understood. For example, the first two features for the median run are: $$\begin{aligned}
& 1\text{st}: \sqrt{ \left(x^{(18382)}\right)^2+x^{(8014)}+x^{(3885)}+x^{(17316)}}
\\
& 2\text{nd}: \left(x^{(7491)}+\sqrt{x^{(7296)}}+x^{(19333)}\right) \times\\ & {\tiny \left( \frac{x^{(5524)}+x^{(18053)}}{\sqrt{1 + \left(x^{(5579)}-x^{(4417)}\right)^2}}+x^{(14153)}+x^{(19751)}- \frac{ x^{(13744)} }{\sqrt{1 + \left( x^{(16581)} \right)^2 } }\right) }$$ Even tough the second feature is somewhat involved, it is arguably still possible to carefully analyze it and obtain a picture of how gene expression levels interact.
Overall, we cannot draw a strict conclusion on whether the features found by our approach are interpretable, as interpretability is a subjective matter and, to date, no clear-cut metric exists [@lipton2018mythos; @guidotti2018survey] (we discuss this more in Sec. \[sec:discussion\]). Yet, it appears evident that enforcing a restriction on their size is a necessary condition. We generally find that features using 15 or more nodes start to be hard to interpret w.r.t. our experimental settings, i.e., using our function set. Lastly, features constructed without a strict size limitation (by SGP) are in general very large, and thus extremely hard to understand. As an example, Figure \[fig:sgp-feature\] shows the first of the two features with median test performance constructed by SGP for LR on Concrete (this is smaller than the first feature found by SGP for NB on Ecoli).
Visualizing what the ML algorithm learns
----------------------------------------
The construction of a small number of interpretable features can enable a better understanding of the problem and of the learned ML models. The case where up to two features are constructed is particularly interesting, since it allows visualization.
We provide one example of classification boundaries and one of a regressed surface, inferred by SVM on a two dimensional feature space obtained with our approach using GP-GOMEA~RT~.
The classification dataset on which we find the best test improvement for $h=4$ is Image Segmentation, where the $F1$ score of SVM reaches 0.88, against 0.65 using the original feature set (median run). Figure \[fig:svm-segmentation\] shows the classification boundaries learned by SVM. The analytic quotient operator $\div$ and the protected log $\log_p$ are replaced by their definition for readability. The constructed features are rather complex here, yet readable. At the same time, it can be clearly seen how the training and test examples are distributed in the 2D space, and what classification boundaries SVM learned.
For regression, Figure \[fig:svm-yacht\] shows the surface learned by SVM on Yacht (median run), where GP-GOMEA~RT~ with $h=2$ constructs two features that lead to an $R^2$ of 0.98, against 0.85 obtained using the original feature set. The features are arguably easy to interpret, while it can be seen that the learned surface accurately models most of the data points.
![Classification boundaries learned by SVM with two features constructed by GP-GOMEA~RT~ ($h=4$) on the Image Segmentation dataset. The run with median test performance is shown. Circles are training samples, diamonds are test samples.[]{data-label="fig:svm-segmentation"}](SVMc_segmentation_NEW){width="0.89\linewidth"}
Running time {#sec:running-time}
============
Our results are made possible by evaluating the fitness of constructed features with cross-validation, a procedure which is particularly expensive. Table \[tab:running-time\] shows the (mean over 30 runs) serial running time to construct five features on the smallest and largest classification and regression datasets, using GP-GOMEA~RT~ with $h=4$ and the parameter settings of Sec. \[sec:experiments\], on the relatively old AMD Opteron Processor 6386 SE[^3]. Running time has a large variability, from seconds to dozens of hours, depending on dataset size and ML algorithm. For the traditional datasets and ML algorithms we considered, it can be argued that our approach can be used in practice. However, for very high-dimensional datasets, only fast ML algorithms can be used. The construction of 5 features for the RNA-Seq gene expression dataset took 25 minutes even tough NB was used. To use slower ML algorithms would easily require dozens to hundreds of hours.
As to memory occupation, it basically mostly depends on the way the chosen ML algorithm handles the dataset. Our runs required at most few hundreds of MBs when dealing with the larger traditional datasets, for SVM and RF. Handling the parallel execution of FCS experiments upon the gene expression dataset required a few GBs.
Discussion {#sec:discussion}
==========
We believe this is one of the few works on evolutionary feature construction where the focus is put on both improving the performance of an ML algorithm, and on human interpretability at the same time. The interpretability we aimed for is twofold: understanding the meaning of the features themselves, as well as reducing their number. GP algorithms are key, as they can provide constructed features as interpretable expressions given basic functional components, and a complexity limit (e.g., tree height).
We have run a large set of experiments, totaling more than 150,000 cpu-hours. Our results strongly support the hypothesis that the original feature set can be replaced by few (even solely $K=2$) features built with our FCS without compromising performance in many cases. In some cases, performance even improved. GP-GOMEA~RT~ and SGP~b~ achieve this result while keeping the constructed feature size extremely limited ($h=2,4$). SGP leads to slightly better performance than GP-GOMEA~RT~ and SGP~b~, but at the cost of constructing five to ten times larger features. RS proved to be less effective than the GP algorithms.
Our FCS is arguably most sensible to use for simpler ML algorithms, such as NB and LR. Constructed features change the space upon which the ML algorithm operates. SVM already includes the kernel trick to change the feature space. Similarly, the trees of RF and XGB effectively embody complex non-linear feature combinations to explain the variance in the data. NB and LR, instead, do not include such mechanisms. Rather, they have particular assumptions on how the features should be combined (NB assumes normality, LR linearity). The features constructed by GP can transform the input the ML algorithm operates upon, to better fit its assumptions.
We found that performance was almost always significantly better than compared to using the original feature set for NB and LR. As running times for these ML algorithms can be in the order of seconds or minutes (Sec. \[sec:running-time\]), feature construction has the potential to be routinely used in data analysis and machine learning practice. Furthermore, FCS (or a modification where the constructed features are added to the original set) can be used as an alternative way to tune simple ML algorithms which have limited or no hyper-parameters.
We have shown that our approach can also be helpful when dealing with high-dimensional data (on the RNA-Seq gene expression dataset), where system underdetermination can cause even simpler ML algorithms to overfit. This is because FCS essentially embodies feature selection, as we only construct a small number of small-sized features.
We remark that we did not adopt very popular high-dimensional datasets concerning image recognition such as MNIST [@lecun1998mnist], CIFAR [@krizhevsky2009learning], or ImageNet [@deng2009imagenet]. In these datasets, features represent pixels, and each pixel has no particular meaning. Consequently, constructing features as readable pixel transformations will likely carry no unhelpful information to explain the behavior of a ML model.
Regarding the comparison between the search algorithms, GP-GOMEA~RT~ was found to be slightly preferable to SGP~b~ (especially for $h=2,K=2$). We believe that significantly better results can be achieved if bigger population sizes and larger evaluations budgets can be employed (we kept the population size limited due to the computational expensiveness of SVM and RF).
Particularly for GP-GOMEA, previous work has shown that having sufficiently large population sizes enables the possibility to exploit linkage estimation and perform better-than-random mixing [@virgolin2019model; @virgolin2017scalable]. To validate this also within the framework of our proposed FCS, we scaled the population size and the budget of fitness evaluations, and compared the use of the LT with the use of the RT, on two traditional classification dataset: Image Segmentation (19 features) and Madelon (500 features), using NB. The outcome is shown in Figure \[fig:fos-comparison\]: the employment of big-enough population sizes (and of sufficient numbers of fitness evaluation) can lead to better performance, *if statistical metrics can be measured reliably*. For Image Segmentation, the number of terminals to be considered in the genotype is relatively small due to the use of 19 features. This allows the LT to estimate node interdependencies reliably, and deliver better-than-random performance. For Madelon, the large number of terminals (500 features) makes it hard for the LT to outperform the RT within a limited computational budget. All in all, we recommend the use of GP-GOMEA as feature constructor since it was not worse on classification and was statistically better for regression. Furthermore, we advice to use the LT if the population size can be of the order of thousands or more (or even better, if exponential population sizing schemes are used as in [@virgolin2019model; @virgolin2017scalable]). Otherwise, the RT should be preferred.
=0.00mm
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![Comparison between the use of the RT and of the LT in GP-GOMEA. Vertical axis: median F1 score of 30 runs, obtained by NB on Image Segmentation (left) and on Madelon (right) using the first constructed feature, with $h=4$ (note the different scale). Horizontal axis: population size / fitness evaluations budget. Stars indicate significant superiority ($p-\text{value} < 0.05$) of one method w.r.t. the other.[]{data-label="fig:fos-comparison"}](scaling_linkage_segmentation "fig:"){width="0.48\linewidth"} ![Comparison between the use of the RT and of the LT in GP-GOMEA. Vertical axis: median F1 score of 30 runs, obtained by NB on Image Segmentation (left) and on Madelon (right) using the first constructed feature, with $h=4$ (note the different scale). Horizontal axis: population size / fitness evaluations budget. Stars indicate significant superiority ($p-\text{value} < 0.05$) of one method w.r.t. the other.[]{data-label="fig:fos-comparison"}](scaling_linkage_madelon "fig:"){width="0.48\linewidth"}
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
To assess if small constructed features are interpretable and if it is possible to visualize what the behavior of learned ML models, we showed some examples, providing evidence that both requirements can be reasonably satisfied. However, we did not perform a thorough study on interpretability of the constructed features. Several metrics have been recently proposed to measure some form of interpretability for ML models, that could be used to measure the interpretability of features as well. E.g., in [@lipton2018mythos] two metrics called *simulatability* and *decomposability* are proposed. Simulatability represents the capability of humans to predict the output of a model given an input. Decomposability represents the capacity to intuitively understand the components of a model. Crucially, to measure this type of metrics, user studies need to be conducted. For example, experts of a field should be asked to provide feedback, on features constructed for datasets they are knowledgeable about (e.g., biochemists for data on gene expression, civil engineers for data on concrete strength). Nonetheless, we believe that enforcing features (and GP programs in general) to be small still remains a necessary condition to allow interpretability, although it is often ignored in GP literature [@virgolin2019model].
Considering the visualization examples proposed in Section \[sec:result-rq3\], it is natural to compare our approach with well-known dimensionality reduction techniques, such as Principal Component Analysis (PCA) [@wold1987principal] or t-Distributed Stochastic Neighbor Embedding (t-SNE) [@maaten2008visualizing]. We remark that those techniques and our FCS have very different objectives. In general, the sole aim of such techniques is to reduce the data dimensionality. PCA does so by detecting components that capture maximal variance. However, it does not attempt to optimize the transformation of the original feature set to improve an ML algorithm’s performance. Also, PCA does not focus on the interpretability of the feature transformations. FCS takes the performance of the ML algorithm and interpretability of the features into account, while dimensionality reduction comes from forcing the construction of few features. We compared using 2 features constructed with RS (the worst search algorithm) with maximum $h=2$, with using the first 2 PCs found by PCA. The use of constructed features over PCs resulted in significantly superior or equal test performance for all ML algorithms and for all problems. We remark, however, that PCA is extremely fast and independent from the ML algorithm.
Our FCS has several limitations. A first limitation regards the performance obtainable by the ML algorithm using the constructed features. FCS is iterative, and this can lead to suboptimal performance for a chosen $K$, compared to attempting to find $K$ features at once. This is because the contributions of multiple features to an ML algorithm are not necessarily perpendicular to each other [@tran2019genetic]. FCS could be changed to find at any given iteration, a synergistic set of $K$ features, that is independent from previous iterations. To this end, larger population sizes need to be employed, and the search algorithms need to be modified so that they can evolve sets of constructed features (a similar proposal for SGP was done in [@krawiec2002genetic]). Yet, it is reasonable to expect that if $K$ features need to be learned at the same time, larger population sizes may be needed compared to learning the $K$ features iteratively.
Another limitation of this work is that hyper-parameter tuning was not considered. To include hyper-parameter tuning within FCS could bring even higher performance scores, or help prevent overfitting. A possibility could be, for example, to evolve pairs of features and hyper-parameter settings, where every time a feature is evaluated, the optimal hyper-parameters are also searched for. Such a procedure may likely require strong parallelization efforts, as $C$-fold cross-validation should be carried out for each combination of hyper-parameter values.
Lastly, it would be interesting to extend our approach to other classification and regression settings, e.g., problems with missing data; or to unsupervised tasks, as simple features may lead to better clustering of the examples.
Conclusion {#sec:conclusion}
==========
With a simple evolutionary feature construction framework we have studied the feasibility of constructing few crucial and compact features with Genetic Programming (GP), towards improving the explainability of Machine Learning (ML) models without losing prediction accuracy. Within the proposed framework, we compared standard GP, random search, and the GP adaptation of the Gene-pool Optimal Mixing Evolutionary Algorithm (GP-GOMEA) as feature constructors, and found that GP-GOMEA is overall preferable when strict limitations on feature size are enforced. Despite limitations on feature size, and despite the reduction of problem dimensionality that we imposed by constructing only two features, we obtained equal or better ML prediction performance compared to using the original feature set for more than half the combinations of datasets and ML algorithms. In many cases, humans can understand what the feature means, and it is possible to visualize how trained ML models will behave. All in all, we conclude that feature construction is most useful and sensible for simpler ML algorithms, where more resources can be used for evolution (e.g., larger population sizes), which, in turn, unlock the added benefits of more advanced evolutionary mechanisms (e.g., using linkage learning in GP-GOMEA).
Acknowledgments {#acknowledgments .unnumbered}
===============
The authors acknowledge the Kinderen Kankervrij foundation for financial support (project \#187). The majority of the computations for this work were performed on the Lisa Compute Cluster with the support of SURFsara.
[^1]: The datasets are available at http://goo.gl/9D2z3b
[^2]: http://archive.ics.uci.edu/ml/
[^3]: http://cpuboss.com/cpu/AMD-Opteron-6386-SE
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this manuscript, we discuss the use of describing functions as a systematic approach to the analysis and design of oscillators. Describing functions are traditionally used to study the stability of nonlinear control systems, and have been adapted for analyzing LC oscillators. We show that they can be applied to other categories of oscillators too, including relaxation and ring oscillators. With the help of several examples of oscillators from various physical domains, we illustrate the techniques involved, and also demonstrate the effectiveness and limitations of describing functions for oscillator analysis.'
author:
- 'Tianshi Wang\'
bibliography:
- 'stringdefs.bib'
- 'jr.bib'
- 'PHLOGON-jr.bib'
- 'tianshi.bib'
title: Analyzing Oscillators using Describing Functions
---
*=***
| {
"pile_set_name": "ArXiv"
} |
[**M. Zra[ł]{}ek**]{}\
\
\
The problem if existing neutrinos are Dirac or Majorana particles is considered in a very pedagogical way. After a few historical remarks we recall the theoretical description of neutral spin 1/2 particles, emphasizing the difference between chirality and helicity which is important in our discussion. Next we describe the properties of neutrinos in the cases when their interactions are given by the standard model and by its extensions (massive neutrinos, right-handed currents, electromagnetic neutrino interaction, interaction with scalar particles). Various processes where the different nature of neutrinos could in principle be visible are reviewed. We clear up misunderstandings which have appeared in last suggestions how to distinguish both types of neutrinos.
Introduction.
==============
The main problem in neutrino physics is the one of the neutrino mass and mixing between different neutrino flavours. There are many indications that neutrinos are really massive particles (LSND experiment, problem of the solar and atmospheric neutrinos, dark matter).
If neutrinos are massive, the next problem is connected with their nature. Charged fermions are Dirac particles and it is a consequence of the electric charge conservation. Lepton number conservation is decidedly less fundamental than charge conservation and it does not govern the dynamics. Total lepton number can be broken, as it is predicted by many extensions of the Standard Model (SM). Then neutrinos do not hold any additive internal quantum numbers and can be identical to their own antiparticles. Such fermions are now generally known (not only for spin 1/2) as Majorana particles. The dilemma whether existing neutrinos are Dirac or Majorana particles is the subject of this paper. We would like to stress that it is not the point if Dirac and Majorana neutrino differ or not. Of course, they do. Majorana neutrinos are their own antiparticles, which is not the case for Dirac neutrino. The problem is whether there is some chance to distinguish them experimentally (within the Standard Model or beyond the SM neutrino interactions). These questions can also be divided into two parts. Firstly we can consider if they are distinguishable in principle and secondly what are technical possibilities to see different effects in real experiment for both types of neutrino. We would like to present a critical review of various efforts and suggestions how to distinguish Dirac from Majorana nature. It is still a ”hot” problem and there are many answers emerging here, both correct and wrong. After short historical remarks (Chapter 2) we remind the theoretical description of massless and massive, neutral spin 1/2 fermions (Chapter 3). Next, in Chapter 4, we describe the standard model interaction of neutrinos and analyze the others, beyond the standard model, neutrino properties which can give better chance to distinguish their nature. In Chapter 5 we give a review of various processes where it seems to be possible for light neutrinos to find some specific signal different for both characters of neutrinos. In case of wrong suggestions we indicate the place of errors. In Chapter 6 we summarize our main conclusions.
Historical remarks.
===================
After Wolfgang Pauli hypothesis \[1\] in 1930 neutrino was born as a Dirac fermion described by Paul Dirac equation known from 1927. Neutrino and antineutrino were distinct particles. Such Dirac particles were used in 1934 by Enrico Fermi in his model of neutrino interactions with nucleons in $%
\beta ^{-}$and $\beta^{+}$ decay \[2\]. Three years later, in 1937, Etore Majorana wrote his famous equation \[3\] in which neutrino was a neutral object, the same as its own antiparticle. Two years before, Maria Goeppert Meyer noticed that single $\beta $ decay was not allowed for even-even nuclei, but decay for such nuclei with emission of two electrons
$$\left( A,Z\right) \longrightarrow \left( A,Z+2\right) +2e^{-}+2\bar{\nu}%
_e,$$
was possible.
Already in 1939 Wendell Furry realized \[4\] that (if neutrinos have Majorana character) the neutrinoless double $\beta $ decay
$$\left( A,Z\right) \longrightarrow \left( A,Z+2\right) +2e^{-}$$
would be possible, too.
We will see that this process is also nowadays the best place where the nature of neutrinos is tested. In 1952 Raymond Davis found no evidence that antineutrinos from the reactor were absorbed in the chlorine detector by the reaction \[5\]
$$\bar{\nu}_e+_{17}^{37}Cl\longrightarrow e^{-}+_{18}^{37}Ar.$$
Four years later in 1956 (after neutrino discovering \[6\]) it was known that only neutrino $\nu_e$ could produce electron
$$\nu_e+_{17}^{37}Cl\longrightarrow e^{-}+_{18}^{37}Ar.$$
The results of both observations indicated that neutrino ($\nu_e)$ and antineutrino $\left( \bar{\nu}_e \right) $ were distinct particles and to describe the difference the electron lepton number was introduced $\left(
L_{\nu_e}=1,L_{\bar{\nu}_e}=-1\right) .$ After this discovery it was obvious that the neutrinos should be treated as the Dirac particle and not as the Majorana one.
In 1956 parity violation was discovered by Tsung Dao Lee and Chen-Ning Yang \[7\] and experimentally supported one year later by Chien-Shiung Wu et al. \[8\]. Immediately it was realized that breaking of the mirror symmetry is easy to understand if we assumed that neutrinos were massless particles \[9\]. Four component spinor, resolution of the Dirac equation with vanishing mass, decoupled for two independent two component spinors. The first one, which described particle with negative and antiparticle with positive helicity $\left(\nu_L,\bar{\nu}%
_R\right),$ and the second one with opposite helicities for particle and antiparticle $\left( \nu_R,\bar{\nu}_L\right) .$ If the neutrino interaction was of V-A type then only one particle should be visible, and experiments should decide which one. Such an experiment had been done in 1958 by Maurice Goldhaber et al. in Brookhaven \[10\]. The answer was clear. The neutrinos from $\beta ^{+}$ decay had negative helicity and that ones from $%
\beta ^{-}$ were positive helicity states. Only the first $\left( \nu_L,\bar{
\nu}_R\right) $ resolution of massless Dirac equation (known as Weyl equations) was realized in nature. After this discovery the Davis’ result could be interpreted in an alternative way. The chlorine experiment could only distinguish negative from positive helicity particle states; it couldnot tell the difference between Dirac and Majorana neutrinos. So, from the experimental point of view there was no way to distinguish
$$\begin{array}{c}
\nu_L\Leftrightarrow \nu \left( -\right) , \\
\\
\bar{\nu}_R\Leftrightarrow \bar{\nu}\left( +\right) .
\end{array}$$
In 1957-58 several papers appeared \[11\] which had shown that there was equivalence between Weyl and massless Majorana fermions. Then, for almost twenty years, there was practically no discussion in literature about neutrino’s nature. In the seventies unification theories appeared with massive neutrinos \[12\]. The so called ”see-saw” mechanism made it possible to understand why the mass of neutrinos was very small \[13\]. After first observations of the solar neutrino anomaly \[14\] the problem of neutrino mass became one of the most important subjects in particle physics (and later in astrophysics and cosmology). For massive neutrinos the problem of their nature once more began to be very important. Fifty years later the Majorana paper has become again famous, as it poses what Pontecorvo calls ”the central problem in neutrino physics”: is neutrino identical to its own antiparticle? From the beginning of eighties papers with different suggestions how to resolve this problem have been appearing continuously. Unfortunately, the very pessimistic observation made in 1982 \[15\] stating that all observable effects which differentiate Dirac and Majorana neutrinos disappear if neutrino mass goes to zero is still valid.
Dirac, Majorana, Weyl neutrinos their helicity, chirality and all that.
=======================================================================
For the future discussion it is worth presenting a short reminder of definitions of the basic properties of spin 1/2 fermion.
It is well known that Lorentz group L$_{+}^{\uparrow }$ has two nonequivalent two-dimensional representations. The objects which transform under Lorentz transformation are known as the van der Waerden spinors \[16\], right $\Psi _{R}$and left $\Psi _L.$
$$\Psi _{R\mbox{ }}
\stackrel{Lorentz\mbox{ }transformation}{\longrightarrow }%
\Psi _{R\mbox{ }}^{^{\prime }}=e^{\frac i2\theta \overrightarrow{n}%
\overrightarrow{\sigma }}e^{-\frac \lambda 2\overrightarrow{m}%
\overrightarrow{\sigma }}\Psi _{R\mbox{ }},$$
and
$$\Psi _{L\mbox{ }}
\stackrel{Lorentz\mbox{ }transformation}{\longrightarrow }%
\Psi _L^{^{\prime }}=e^{\frac i2\theta \overrightarrow{n}\overrightarrow{%
\sigma }}e^{\frac \lambda 2\overrightarrow{m}\overrightarrow{\sigma }}\Psi
_{L\mbox{ }},$$
where $\overrightarrow{n},\overrightarrow{m},\theta $ and $\lambda $ are proper characteristics of Lorentz transformation and $\overrightarrow{\sigma }
$ are Pauli matrices \[17\]. For zero mass objects these spinors satisfy the Weyl equations \[18\].
$$\left( \widehat{\sigma}^{\mu}\partial_{\mu} \right) \Psi _{R\mbox{ }}=0,
\mbox{ }\widehat{\sigma }^\mu =\left( \sigma ^0,
\overrightarrow{\sigma }\right) ,$$
$$\left( \sigma ^{\mu}\partial_{\mu}\right) \Psi _L=0,\mbox{ }\sigma ^{\mu}
=\left( \sigma
^0,-\overrightarrow{\sigma }\right) ,$$
and describe particle with positive $\left( \Psi _{R\mbox{ }}\right) $ and negative ($\Psi _L$) helicities. For massless particle the spin projection on momentum is Lorentz invariant. For particles with mass the Weyl equations are not satisfied and there are two possibilities. The first one, more fundamental was discovered by Majorana \[3\]. The fields $\Psi _{R(L)\mbox{ }}$satisfy the Majorana equations
$$i\left( \widehat{\sigma }^{\mu}\partial_{\mu}\right) \Psi _{R\mbox{ }}-m
\varepsilon
\Psi _{R\mbox{ }}^{*}=0,$$
and
$$i\left( \sigma^{\mu}\partial_{\mu}\right) \Psi _{L\mbox{ }}+m^{^{\prime
}}\varepsilon \Psi _{L\mbox{ }}^{*}=0,$$
where m, m$^{^{\prime }}$ are particle masses and $\varepsilon =\left(
\begin{array}{cc}
0 & 1 \\
-1 & 0
\end{array}
\right).$ Equations (10,11) describe two completely different objects with masses m and m$^{^{\prime }}$ which do not possess any additive quantum numbers and particles are their own antiparticles.
The second possibility of the field equation for massive fermion had been known before Majorana as the Dirac equation \[19\]
$$\begin{array}{c}
i\left(
\widehat{\sigma }^{\mu}\partial_{\mu}\right) \Psi _{R\mbox{ }}-m\Psi _{L\mbox{ }}=0,
\\ \\
i\left( \sigma^{\mu}\partial_{\mu}\right) \Psi _{L\mbox{ }}-m\Psi _{R\mbox{
}}=0,
\end{array}$$
and had described only one fermion with some additive quantum number (e.g. charge). Usually this equation is presented in four dimensional Dirac bispinor formalism as
$$\left( i\gamma ^\mu \partial _\mu -m\right) \Psi =0,$$
where $$%\,\,\,\,\,\,\,\,\mbox{
%where\thinspace \thinspace \thinspace \thinspace \thinspace }
\gamma ^\mu
=\left(
\begin{array}{cc}
0 & \sigma ^\mu \\
\widehat{\sigma }^\mu & 0
\end{array}
\right) ,\,\,\,\,\,\,\,\,\,\,\mbox{ and \thinspace \thinspace \thinspace
\thinspace \thinspace \thinspace \thinspace }\Psi =
\left( \matrix{\Psi _{R\mbox{ }} \cr
\Psi _{L\mbox{ }}} \right),$$ which is known as Weyl representation for Dirac $\gamma $ matrices. In this representation let us define
$$\begin{array}{c}
\gamma ^5=i\gamma ^0\gamma ^1\gamma ^2\gamma ^3=\left(
\begin{array}{cc}
1 & 0 \\
0 & -1
\end{array}
\right) ,\mbox{ P}_L=\frac 12\left( 1-\gamma _5\right) =\left(
\begin{array}{cc}
0 & 0 \\
0 & 1
\end{array}
\right) , \\
\\
\mbox{and \thinspace \thinspace \thinspace \thinspace \thinspace \thinspace }%
P_R=\frac 12\left( 1-\gamma _5\right) =\left(
\begin{array}{cc}
0 & 0 \\
0 & 1
\end{array}
\right) .
\end{array}$$
Then$$\Psi _{R\mbox{ }}\equiv \left( \matrix{ \Psi _{R\mbox{ }} \cr 0
} \right) \equiv P_R\Psi ,\mbox{
\thinspace \thinspace \thinspace \thinspace \thinspace \thinspace \thinspace
\thinspace \thinspace }\Psi _{L\mbox{ }}\equiv
\left( \matrix{ 0 \cr \Psi _{L\mbox{ }%
}} \right) \equiv P_L\Psi .$$ The spinors $\Psi _{R(L)\mbox{ }}$ are eigenvectors of $\gamma _5$
$$\gamma _5\Psi _R=\Psi _R,\mbox{ }\gamma _5\Psi _L=-\Psi _L,$$
and are known as chiral eigenvectors with eigenvalues + and – which have the name ”chirality”. For massless particles the chirality ”$\pm $” coincide with the helicity, $\pm \frac 12.$ For massive particles the chirality and helicity decouple. As we know from Eqs. (6,7) the chirality is Lorentz invariant, irrespective of whether particle is massive or massless. The helicity is Lorentz invariant only for massless particle. For a massive particle there always exist Lorentz frames in which the particle has opposite momentum. This means that helicity of this particle changes sign and cannot be a Lorentz invariant object. It is instructive to decompose the free fields $\Psi _{L(R)%
\mbox{ }}$for different kinds of particles in the helicity representation.
--------------------------------------------------------------------------------------------------------------------------------------------------
$m \neq 0$ m=0
---------- ------------------------------------------------------------------- -------------------------------------------------------------------
Dirac $\Psi_R(x)=\int \left[ A(+)e^{-ikx}-B^{\dagger}(-)e^{ikx} \right] $\Psi_R(x)=\int \left[ A(+)e^{-ikx}-B^{\dagger}(-)e^{ikx} \right]
\chi(+) \sqrt{E+k}$ \chi(+) \sqrt{2E}$
+ $ \int \left[ A(-)e^{-ikx}+B^{\dagger}(+)e^{ikx} \right]
\chi(-) \sqrt{E-k}$
and Weyl
$\Psi_L(x)=\int \left[ A(-)e^{-ikx}-B^{\dagger}(+)e^{ikx} \right] $\Psi_L(x)=\int \left[ A(-)e^{-ikx}-B^{\dagger}(+)e^{ikx} \right]
\chi(-) \sqrt{E+k}$ \chi(-) \sqrt{2E}$
fields + $ \int \left[ A(+)e^{-ikx}+B^{\dagger}(-)e^{ikx} \right]
\chi(+) \sqrt{E-k}$
Majorana $\Psi_R(x)=\int \left[ a(+)e^{-ikx}-a^{\dagger}(-)e^{ikx} \right] $\Psi_R(x)=\int \left[ a(+)e^{-ikx}-a^{\dagger}(-)e^{ikx} \right]
\chi(+) \sqrt{E+k}$ \chi(+) \sqrt{2E}$
+ $ \int \left[ a(-)e^{-ikx}+a^{\dagger}(+)e^{ikx} \right]
\chi(-) \sqrt{E-k}$
fields $\Psi_L(x)=\int \left[ a(-)e^{-ikx}-a^{\dagger}(+)e^{ikx} \right] $\Psi_L(x)=\int \left[ a(-)e^{-ikx}-a^{\dagger}(+)e^{ikx} \right]
\chi(-) \sqrt{E+k}$ \chi(-) \sqrt{2E}$
+ $ \int \left[ a(+)e^{-ikx}+a^{\dagger}(-)e^{ikx} \right]
\chi(+) \sqrt{E-k}$
--------------------------------------------------------------------------------------------------------------------------------------------------
[Table I. The fields $\Psi_{L(R)}$ for massive Dirac, massless Weyl and for both massive and massless Majorana neutrinos. See text for all denotations used in the Table. The integration is over three momentum: $\int=\int \frac{d^3k}{(2
\pi)^32E}$.]{}
This decomposition is shown in Table I where we use the following denotations:
$$\begin{array}{c}
\overrightarrow{k}=k\left( \sin \theta \cos \varphi ,\sin \theta \sin
\varphi ,\cos \theta \right) , \\ \\
E=\sqrt{m^2+k^2}
\end{array}$$
are momentum and energy of the particle;
$$\chi \left( \overrightarrow{k},+\right) =\left(
\begin{array}{cc}
e^{-i\varphi /2} & \cos \theta /2 \\
e^{i\varphi /2} & \sin \theta /2
\end{array}
\right) ,\mbox{ }\chi \left( \overrightarrow{k},-\right) =\left(
\begin{array}{cc}
-e^{-i\varphi /2} & \sin \theta /2 \\
e^{i\varphi /2} & \cos \theta /2
\end{array}
\right)$$
are Pauli spinors for helicity $+\frac 12$ and $-\frac 12$ respectively, and the $A^{\dagger}\left( A\right),$ $B^{\dagger}\left( B\right) $ are creation (annihilation) operators for Dirac, Weyl particle and antiparticle respectively, and the $a%
\left( a^{+}\right) $ are suitable operators for Majorana particles.
The careful analysis of the Table I is very instructive and it is worth making some comments.
1\. For the Dirac fields there are two distinct operators, one for particle $%
A^{\dagger}\left( A\right) ,$ and the other one for antiparticle $B^{\dagger}
\left( B\right) $. We see that for $E\gg m\neq 0,\;\sqrt{E-k}\approx \frac m{\sqrt{2E}}+{\cal
O}\left(
m^2\right) $ and for definite chirality L or R there are two helicity states, $h=\pm \frac 12.$ This fact is a consequence of the Lorentz invariance. However in that case both helicities have different weights; $\sqrt{E+k}\approx \sqrt{2E}$ for ”good helicity” and $%
\sqrt{E-k}\approx \frac m{\sqrt{2E}}$ for ”wrong helicity” states. For a pure left-handed interaction particles in the mixed helicity states will be produced. If helicity is not measured then the chiral particle state with energy E will be an incoherent superposition of two helicity states described by the statistical operator $\rho (E)$ $$\begin{array}{c}
\rho _{particle}\left( E\right) =\left(
\frac{E+k}{2E}\right) |\overrightarrow{k}, h =-\frac 12>_p<%
\overrightarrow{k},h =-\frac 12|+ \\ \left( \frac{E-k}{2E}\right) |%
\overrightarrow{k},h =+\frac 12>_p<\overrightarrow{k},h =+\frac
12|.
\end{array}$$ In such a state the neutrino e.g. in $\pi ^{+}\longrightarrow \mu
^{+}\nu{_\mu} $ decay will be produced. It is opposite for antiparticle
$$\begin{array}{c}
\rho _{antiparticle}\left( E\right) =\left(
\frac{E+k}{2E}\right) |\overrightarrow{k},h=+\frac 12>_a<%
\overrightarrow{k},h =+\frac 12|+ \\ \left( \frac{E-k}{2E}\right) |%
\overrightarrow{k},h=-\frac 12>_a<\overrightarrow{k},h=-\frac
12|
\end{array}$$
as e.g. for the neutrino in $\pi ^{-}\longrightarrow \mu^{-}\bar{\nu}%
_{\mu}$ decay. For relativistic particles the wrong helicity states $|\vec{k},h ^{}$ = +$\frac 12>_p$ and $|%
{\vec{k},h }$ = $-\frac 12>_a$ have very small weight $\left(
\frac m{2E}\right) ^2$ and even if, in principle, they can be produced, they have never been visible. Sometimes they are called ”sterile neutrino”.
2\. In the Majorana case there is only one operator which creates particle and its own antiparticle. Both states (18) and (19) describe the same object - the Majorana neutrino. There is no sterile neutrino; both helicity states can be produced with equal weights. The left-handed $\Psi _L$ and the right-handed $\Psi _R$ fields are connected $\Psi _R(x)=-\varepsilon \Psi
_L^{*}(x).$ The Majorana fields can be also written in the four component form
$$\Psi (x)=\left( \matrix{ -\varepsilon \Psi _L^{*}(x) \cr \Psi _L(x)} \right),$$
which satisfies the condition
$$i\gamma ^2\Psi ^{*}(x)=\Psi (x),\,\,\,\,\,\,\mbox{\rm where}\mbox{ \thinspace
\thinspace \thinspace \thinspace }i\gamma ^2=\left(
\begin{array}{cc}
0 & -\varepsilon \\
+\varepsilon & 0
\end{array}
\right) .$$
This relation is sometimes used as a definition of the field for the Majorana particle.
3\. In cases of both Weyl and massless Majorana neutrinos the limit m$%
\rightarrow $0 is smooth. From Dirac neutrino we obtain two independent Weyl fields $\Psi _L(x)$ and $\Psi _R(x).$ In the left-handed chiral field $\Psi
_L(x)$ there is particle with negative helicity and antiparticle with positive helicity. In the field $\Psi _R(x)$ it is just opposite: $A_{%
\overrightarrow{k}}(+)$ and $B_{_{\overrightarrow{k}}}(-)$.
For the massless Majorana neutrino two fields $\Psi _L(x)$ and $\Psi _R(x)$ are still connected $\left( \Psi _R(x)=-\varepsilon \Psi _L^{*}(x)\right) .$ In the statical case it was proved \[11\] that one Weyl neutrino e.g. $\Psi _L(x)$ (or separately $\Psi _R(x)$ ) is equivalent to massless Majorana neutrino described by two connected fields $\Psi _L$ and $\Psi _R$. This relation is known as Pauli-Gursey transformation \[11\] which, for annihilation operators, can be written in the form
$$\begin{array}{c}
U^{-1}A_{
\overrightarrow{k}}(-)U=a_{\overrightarrow{k}}(-), \\ \\
U^{-1}B_{\overrightarrow{k}}(+)U=a_{\overrightarrow{k}}(+),
\end{array}$$
where
$$\begin{aligned}
U&=&exp \left[ \frac{\pi}{4} \left( B_{\overrightarrow{k}}^{\dagger}(-)A_{%
\overrightarrow{k}}(-)-A_{\overrightarrow{k}}^{\dagger}(-)B_{\overrightarrow{k}%
}(-) \right. \right. \nonumber \\
&& \left. \left. -B_{\overrightarrow{k}}^{\dagger}(+)A_{\overrightarrow{k}}(+)+
A_{
\overrightarrow{k}}^{\dagger}(+)B_{\overrightarrow{k}}(+) \right) \right] , \end{aligned}$$
and Majorana operators are defined in the following way
$$\begin{array}{c}
a_{
\overrightarrow{k}}(-)=\frac 1{\sqrt{2}}\left[ A_{\overrightarrow{k}}(-)+B_{%
\overrightarrow{k}}(-)\right] , \\ \\
a_{\overrightarrow{k}}(+)=\frac 1{\sqrt{2}}\left[ A_{\overrightarrow{k}%
}(+)+B_{\overrightarrow{k}}(+)\right] .
\end{array}$$
It must be stressed that this equivalence theorem is valid only for not interacting fields. For interacting fields whether the theorem is valid or not depends on the type of interaction. We will see that the massless Weyl-Majorana particles are still indistinguishable if there is only left-handed V-A (or only right-handed V+A) interaction. But for other types of interactions this equivalence theorem is no longer true.
Real and hypothetical neutrino interactions.
============================================
Neutrinos in the Standard Model.
--------------------------------
It is only one case in which SM predicts masses of particles. The SM predicts that neutrinos are massless. There are three massless Weyl neutrinos $\nu _e,$ $\nu _\mu $ and $\nu _\tau .$ As a consequence there is no mixing between generations and (1) leptons have universal interactions, (2) both flavour $L_e,L_\mu ,$ $L_\tau $ and total $L=L_e+L_\mu +L_\tau $ lepton numbers are conserved, and (3) there is not CP violation in the lepton sector.
The massless neutrinos have only the left-handed interactions with the charged and neutral gauge bosons
$$L_{CC}=\frac g{2\sqrt{2}}\overline{N}\gamma ^\mu \left( 1-\gamma _5\right)
lW_\mu ^{+}+h.c.,$$
and $$L_{NC}=\frac g{4\cos \theta_W }\overline{N}\gamma ^\mu \left( 1-\gamma
_5\right) NZ_\mu .$$ The interaction of fermions with the Higgs particles is proportional to the fermion masses, so massless neutrinos do not interact with scalar particles.
This picture of neutrino interaction is confirmed by all terrestrial experiments (maybe LSND results are the first which contradict the presented picture but they still should be better confirmed (e.g. by CARMEN)). In frame of the SM there is not any chance to differentiate between Weyl and massless Majorana fermions.
The other possible neutrino interactions.
-----------------------------------------
If neutrinos are massive particles, the mixing between generations appears in the charged and neutral currents $$L_{CC}=\frac g{2\sqrt{2}}\overline{N_a}\gamma ^\mu \left( 1-\gamma _5\right)
K_{al}l_lW_\mu ^{+}+h.c,$$ and $$L_{NC}=\frac g{4\cos \theta _W}\overline{N_a}\gamma ^\mu \left( 1-\gamma
_5\right) \Omega _{ab}N_bZ_{\mu},$$ where $K_{al}\,$and $\Omega _{ab}$ are suitable mixing matrices resulting from diagonalization of a neutrino mass matrix. If there is a mixing in the lepton sector then the CP symmetry can be broken. It is the first place which differentiates the Dirac from Majorana neutrinos.
For the Dirac neutrinos situation looks like in the quark sector. Both charged leptons and Dirac neutrino fields have the phase transformation freedom
$$N_a\rightarrow N_a^{^{\prime }}=e^{i\alpha _a}N_a \mbox{ \thinspace
\thinspace and \thinspace \thinspace l}_l\rightarrow \mbox{l}_l^{^{\prime
}}\rightarrow e^{i\beta _l}l_l,$$
and substantial number of phases can be eliminated from mixing matrix K (matrix $\Omega $ is a function of K).
For the Majorana neutrinos the phase transformation (29) is not allowed. We can see it from the Majorana equation where the field and its complex conjugation are present simultaneously. Fewer number of phases can be eliminated, so greater amount of them break the CP symmetry. For example, if neutrinos are Dirac particles we need at least three families to break CP, for Majorana neutrinos the CP can be broken already for two families. This different number of CP violating phases for the Dirac and Majorana neutrino can have real physical consequences which could be potentially used to distinguish them experimentally. In practical calculation the difference in CP breaking effects is visible by the different number of Feynman diagrams. Let assume that the mass of muon neutrino is larger than the electron neutrino mass $m_{\nu _\mu } > m_{\nu
_e}.$ Then we can calculate the decay width for the process
$$\nu _\mu \rightarrow \nu _e+\gamma .$$
Let us assume that the mixing matrix in Eq. (27) has the form
$$K=\left(
\begin{array}{cc}
c & se^{i\delta } \\
-se^{-i\delta } & c
\end{array}
\right) .$$
If the neutrinos are Dirac fermions two Feynman diagrams will describe the process (30) at the one loop level.
We see from Fig.1 that the CP violating phase $\delta $ multiplies both diagrams in the same way and cancels after taking modulus square of the sum of both of them.
In the case of the Majorana neutrino there are four diagrams instead of two (Fig.2). We see that calculating the decay width (Fig.2) the CP violating phase will not disappear and we obtain \[21,22\]
$$\Gamma \left( \nu _\mu \rightarrow \nu _e+\gamma \right) \sim \left(
1+\left( \frac{m_e}{m_\mu }\right) ^2+2\left( \frac{m_e}{m_\mu }\right) \cos
2\delta \right) .$$
Even if the CP symmetry is conserved there is an important difference between the Dirac and Majorana case. For Dirac particles the creation operators for particle $(A$) and antiparticle $(B$) can be multiplied by different complex phases, and , as a result, any CP phase can be absorbed, so it is not physical \[23\].
For the Majorana neutrino there is only one operator $(A=B$) and the CP phase cannot be absorbed and it is a physical observable. The CP phase for the Majorana neutrino must be pure imaginary number, and for the helicity state there is \[23, 24, 25\]
$$CP\mbox{ }|\overrightarrow{p},\lambda \rangle \mbox{ }=\eta _{CP}e^{-i\frac
\pi 2}|-\overrightarrow{p},\lambda \rangle ,$$
where $\eta _{CP}=\pm i.$
This fact can also have real physical consequences. Let us consider, for example, the decay of $\pi^0$ into two identical Majorana neutrinos $\pi ^0\rightarrow \nu _M\nu _M$ \[25\]. The initial $\pi ^0$ state is $J^{PC}=0^{-+}$, so the possible final states are
$$J=0,\mbox{ }S=L=0,\,\,\,\,\,\,\,and\,\,\,\,\,\,\,\,J=0,\mbox{ }S=L=1,$$
so$$CP|\nu _M\nu _M\rangle =\left( \pm i\right) ^2\left( -\right) ^L|\nu _M\nu
_M\rangle \mbox{ = }-|\pi ^0\rangle .$$ From this we conclude that $L=0$ so $S=0$ and the emitted neutrino’s spins are antiparallel.
The next important differences between the Dirac and Majorana neutrinos are their electromagnetic structure. In general any spin 1/2 Dirac fermion can have four independent electromagnetic formfactors.
If we define the one photon interaction diagram with two fermions like in Fig.3 then the requirements that
\(i) initial and final fermions are on shell, and
\(ii) the current is conserved $(\Gamma ^\mu q_\mu =0)$ give the structure function $\Gamma ^\mu \left( P,q\right) $ \[26, 27\]
$$\begin{aligned}
\Gamma _D^\mu \left( P,q\right)& =&F_D\left( q^2\right) \gamma ^\mu
+iM_D\left( q^2\right) \sigma ^{\mu \nu }q_\nu \nonumber \\
&+& E_D\left( q^2\right) \sigma ^{\mu \nu }q_\nu \gamma _5+G_D\left( q^2\right)
\left( q^\mu 2m-q^2\gamma ^\mu \right) \gamma _5. \end{aligned}$$
The structure functions for $q^2\rightarrow 0$ correspond to;
$F_D\left( q^2\right) \stackrel{q^2\rightarrow 0}{\longrightarrow}$ $0,\;\;$ neutrino charge,
$\frac 1{2m}F_D\left( q^2\right) +M_D\left( q^2 \right)
\stackrel{q^2\rightarrow 0}{\longrightarrow}$ $\mu _m, \;\;$ magnetic moment,
$E_D\left( q^2\right) \stackrel{q^2\rightarrow 0}{\longrightarrow}$ $\mu
_e,\;\;$ electric moment, and
$G_D\left( q^2\right) \stackrel{q^2\rightarrow 0}{\longrightarrow}$ $T,\;\;$anapol moment \[28\].
For the Majorana particle only one electromagnetic formfactor survives. There are several ways to show it:
- the CPT invariance \[22, 29\],
- identity of fermions in the final state of the decay, $\gamma \rightarrow
\nu _M\nu
_M $ \[25, 30\], or
- from the Feynman rules - the effective coupling of Majorana fermion with a neutral vector boson is given by \[31\] $$\Gamma _M^\mu =\Gamma _D^\mu +C\Gamma _D^{\mu _T}C^{-1},$$
where
$$C= \left( \matrix{ -\varepsilon & 0 \cr
0 & \varepsilon } \right) .$$
Using any of the above methods we can show that only the anapol formfactor describes the electromagnetic structure of the Majorana neutrino
$$\begin{aligned}
\Gamma _M^\mu \left( P,q\right) &=&G_M\left( q^2\right) \gamma ^\mu \gamma
_{5,} \nonumber \\
\mbox{where} \hspace{4 cm} && \nonumber \\
G_M\left( q^2\right) &=&-2G_D\left( q^2\right)q^2.\end{aligned}$$
The electromagnetic structure differentiates the Dirac and Majorana fermions in the obvious way and we can expect to find some visible experimental effect connected with this difference.
Besides the diagonal moments (formfactors) which describe the electromagnetic structure there are also transition moments between different neutrinos (Fig.4)
Then the CPT symmetry does not give any restrictions for $\Gamma _{mn}^\mu \left(
P,q\right) $ and all four transition moments exist for the Dirac as well as for the Majorana neutrinos \[22,27\].
Up to now we have discussed the V-A interaction of massive neutrinos. Experimentally it is not excluded also that a right - handed current appears in the neutrino interaction with the charged and neutral gauge bosons like in the popular left-right symmetric model \[32\]. There are also models where scalar particles interact with neutrinos \[33\]. Generally it is worth remembering that $$\overline{\Psi }_a\Gamma \Psi _b=\Psi _a^{+}\left( \gamma ^0\Gamma \right)
\Psi _b,$$ from which we can find that for scalar and tensor interactions $\Gamma
=1,\gamma _5,\sigma ^{\mu \nu },\sigma ^{\mu \nu }\gamma _5$ the chirality of $%
\Psi _a$ and $\Psi _b$ must be opposite but for vector interactions $\Gamma
=\gamma ^\mu ,\gamma ^\mu \gamma _5$ the chirality is conserved. It follows very easily from the properties of the projection operators (Eq. (14))
$$P_LP_R=0,\mbox{ }P_L^2=P_L,\mbox{ }P_R^2=P_R,$$
then e.g.
$$\begin{array}{c}
P_L\gamma ^0\gamma ^\mu P_L\neq 0,
\mbox{ but }P_R\gamma ^0\gamma ^\mu P_L=0, \\ \\
P_L\gamma ^0\gamma _5P_R\neq 0,\mbox{ but }P_L\gamma ^0\gamma _5P_L=0.
\end{array}$$
For relativistic particles, the chirality is almost identical with the helicity, and we can transform the above rule for the helicity of incoming and outgoing particles.
Differences between the Dirac and Majorana neutrinos for various neutrino interactions.
---------------------------------------------------------------------------------------
First we consider the situation with vanishing neutrino mass. If there is only a left-handed interaction $\gamma ^\mu P_L$ (as in the SM), then particles with negative and antiparticles with positive helicities are produced and interact after production.
For Weyl particle: $A(-)$ and $B(+)$.
For Majorana object: $a(-)$ and $a(+)$. Our interaction will not change $A\longleftrightarrow B$ as well as it will not change $a(-)\longleftrightarrow a(+)$. We know that the $A$ and the $B$ objects interact differently but we have no possibility to check what is the reason for that. That means that two cases
1. the lepton number is conserved and different helicity in $A$ and $B$ operator has no meaning (so really $A\neq B$ and we have Weyl neutrino), or
2. the interaction depends on the helicity of particles, so $A=B$ and the particles interact differently because they have different helicities (Majorana neutrino)
are physically indistinguishable.
In order to answer the question if $A=B$ or not we must have possibility to compare particle with antiparticle in the same helicity states, so
$$A(-)\mbox{ with }B(-),$$
or
$$A(+)\mbox{ with }B(+).$$
But the operators $A(+)$ and $B(-)$ appear in right-handed chiral state $%
N_R. $ Such a field will appear if there are right-handed currents or scalar -neutrino interactions. Let us consider a simple example. Beam of massless neutrinos with negative helicity interacts with matter. We assume that there is the left-handed and the right-handed current in the neutrinos interaction with electrons, so
$$L_{CC}=\alpha \overline{l}\gamma ^\mu P_LNW_\mu ^{-}+\beta \overline{l}%
\gamma ^\mu P_RNW_\mu ^{-}+h.c.$$
Then
$\bullet $ for the Weyl neutrino only electrons will be produced in deep inelastic scattering (Fig.5)
$\bullet $ for the Majorana neutrino electrons and positrons will appear (Fig.6)
This simple (but unfortunately unrealistic) example shows us that if the left-handed and the right-handed currents are present we can distinguish the Weyl from the massless Majorana neutrinos (if there are the left and right-handed currents both fields $N_L$ and $N_R$ appear, so we already should talk about Dirac particle). There is also a more realistic example which convinces us that the massless Dirac and Majorana neutrinos are distinguishable (at least in principle) if both left-handed and right-handed currents are present. The magnetic moment calculated in frame of the L-R symmetric model for Dirac neutrino does not vanish even if neutrino mass is equal to zero \[22, 27, 34\]
$$\mu _{\nu_i}=\left( \frac{\sqrt{2}G_F}{\pi ^2}\sin \varphi \cos \varphi
m_e\sum_lm_lRe\left( V_{il}^{+}U_{il}\right) \right) \mu _B$$
(see references for precise denotation of all the parameters in Eq.(43)). So, with vanishing neutrino mass and only left-handed current interaction, there is no way to distinguish the Dirac from Majorana neutrinos. Such a possibility appears if neutrinos interact also by right-handed currents or if interactions with scalar particle are not proportional to the neutrino mass.
The situation will change if neutrinos have some tiny mass. Then even if there are only left-handed currents, there are ways (at least in principle) to distinguish both types of neutrino. It is so because the left-handed chiral states (Eqs. 18,19) are not exactly negative helicity states and, in principle, there is a possibility to compare interaction of particle and antiparticle in the same helicity state:
$$\begin{aligned}
\left( \frac m{\sqrt{2E}} \right) |\overrightarrow{k},\lambda & = &
+1/2\rangle _p \;\; \mbox{ for the particle} \nonumber \\
\mbox{\rm with} \;\;\;\;\;\;\;\;\;\; && \\
|\overrightarrow{k},\lambda & =& +1/2\rangle _a \;\; \mbox{ for
the antiparticle.} \nonumber\end{aligned}$$
The several sources of neutrinos are known (reactor, accelerator, the sun, supernova) but usually they are produced with relativistic energy $E \sim 0$ (MeV) {$m_{\nu_e}<3.5\;$eV $[35]$ $m_{\nu_{\mu} }<160\;$keV \[35\] $m_{\nu_{\tau}}<18.2\;$ MeV \[36\] but from Big-Bang Nucleosynthesis $m_{\nu_{\tau} }<0.95\;$ MeV \[37\] and from matter density of the Universe, $m_{\nu_{\tau}} <23\;$ eV \[37\]}. Only if we relax the astrophysical and cosmological information the weight factor $(\frac{m_\tau }{2E_\tau }%
)$ for $\tau \,\,$neutrino is large enough to have interesting value from the experimental point of view. But, unfortunately, up to now we have not produced the beam of $\tau $ neutrinos.
Let us consider the interaction of massive Dirac and Majorana neutrinos with a matter. In both cases the interaction Lagrangian is the same
$$L_{CC}=\frac g{\sqrt{2}}\left\{ \left( \overline{N}\gamma ^\mu P_Ll\right)
W_\mu ^{+}+\left( \overline{l}\gamma ^\mu P_LN\right) W_\mu ^{-}\right\} .$$
If the beam of Dirac neutrinos $\nu ^D$ with helicity $h_\nu $ interacts with a matter only electrons with helicities $h_e$ are produced (Fig.7).
The amplitudes for this process are proportional to the factor $\left[
\left( E_\nu -2h_\nu k_\nu \right) \left( E_e-2h_ek_\nu \right) \right]
^{1/2}$ where $E_\nu \left( E_e\right),$ $k_\nu \left( k_e\right) $ are energy and momentum of the neutrinos (electrons), respectively. Positrons will not be produced by the Dirac neutrinos even if neutrinos with positive helicity $\left( \nu
_{+}^D\right) $ exist in the beam. Neutrinos in both helicity states will produce electron only, but mainly $\nu _{-}^D$ will do it, $\nu _{+}^D$ produces e$^{-}\left( h_e=\pm 1/2\right) $ with the small weight $\left(
m_\nu ^2/4E_\nu ^2\right) .$
If the beam of massive Majorana neutrinos interacts with a matter the picture is different. Both electrons and positrons can be produced (Fig.8)
We see that electrons are produced mostly by neutrinos with negative helicity $\nu _{-}^M,$ contrary to positrons which are produced mainly by $%
\nu _{+}^M.\,$ The Diracand Majorana neutrinos are distinguishable (in principle) if the second diagram for the Majorana neutrino (absent in the Dirac case) gives any contribution to the neutrino interaction with a matter. It happens if the factor ($E_\nu +2h_\nu k_\nu )\neq 0,$ and we can conclude that both neutrinos are distinguishable if $m_\nu \neq 0$ or neutrinos with helicity $%
h_\nu =+1/2$ exist in the beam (even if $m_\nu =0$). The second case means that the massless Dirac and Majorana neutrinos are distinguishable in the charged current interaction.
>From presented considerations we could get the impression that massive Dirac neutrinos are very easily distinguishable from massive Majorana neutrinos. But it is not the case. In the relativistic beam of neutrinos produced by the chiral left-handed interaction the $\nu _{+\mbox{ }}$ neutrinos occur very seldom. So, it is very difficult with the tiny neutrino mass to distinguish Dirac from Majorana neutrinos using a process where the charged current dominates the interaction.
What about processes with the neutral currents? There are two things which are worth considering in this context.
\(i) At first sight the neutral current for the Majorana neutrino looks completely different in comparison with the Dirac neutrino. The Majorana neutrino has no vector current $\overline{\nu }_M\gamma ^\mu \nu _M=0$ \[15, 25\].
\(ii) In processes where we usually have Dirac neutrino and antineutrino in Majorana case two identical particles appear. For example in the $Z_0$ decay we get two identical Majorana neutrinos $Z_0\rightarrow \nu _M\nu _M$. Here it seems also that it is easy to distinguish both cases, because of symetrization procedure for identical particles.
Let us now consider the first problem. The second one, which was the cause of many mistakes, we will study in the next Chapter.
From the property (21) for Majorana neutrino it follows that
$$\overline{\nu }^M\left( x\right) \gamma ^\mu \nu ^M\left( x\right) =0,$$ and we have \[25, 38\]
$$\left\langle \nu _f^M\left| \overline{\nu }^M\gamma ^\mu (1-\gamma _5)\nu
^M\right| \nu _i^M\right\rangle _{\mid x=0}=-\left\langle \nu _f^M\left|
\overline{\nu }^M\gamma ^\mu \gamma _5\nu ^M\right| \nu _i^M\right\rangle
_{\mid x=0}.$$
If we decompose the neutrino fields in momentum representation (Table I) we get $$\begin{array}{c}
\langle \nu _f^M\left|
\overline{\nu }^M\gamma ^\mu \left( 1-\gamma _5\right) \nu ^M\right| \nu
_i^M\rangle _{\mid x=0}= \\ \\
\overline{u}\left( \stackrel{\rightarrow }{k}%
_f,h_f\right) \gamma ^\mu \gamma _5\,u\left( \stackrel{\rightarrow }{k%
}_i,h_i\right) +\overline{\upsilon }\left( \stackrel{\rightarrow }{k}%
_i,h_i\right) \gamma ^\mu \gamma _5\upsilon \left( \stackrel{%
\rightarrow }{k}_f,h_f\right) ,
\end{array}$$ and from the property $$\upsilon =C\overline{u}^T,\mbox{ }\overline{\upsilon }=-u^TC^{-1},$$ there is $$\left\langle \nu _f^M\left| \overline{\nu }^M\gamma ^\mu \left( 1-\gamma
_5\right) \nu ^M\right| \nu _i^M\right\rangle _{\mid _{x=0}}=-2\overline{u}%
_f\gamma ^\mu \gamma _5u_i.$$ For Dirac neutrino at first sight the result is different $$\left\langle \nu _f^D\left| \overline{\nu }^D\gamma ^\mu \left( 1-\gamma
_5\right) \nu ^D\right| \nu _i^D\right\rangle _{\mid x=0}=\overline{u}%
_f\gamma ^\mu \left( 1-\gamma _5\right) u_i.$$ But we have to our disposal relativistic neutrinos produced by the left-handed current so with precision $\left( \frac m{2E}\right) $ the negative helicity state $\left( h=-1/2\right) $ is the chiral left-handed state (see Eq. (18)) so for our initial spinors there is $$P_Lu_i\simeq u_i+{\cal O} \left( \frac m{2E}\right) ,\,\,\,\,\,\,\,\,\,P_Ru_i\simeq
{\cal O}\left( \frac m{2E}\right) ,$$ from which it follows that $$\gamma ^5u_i=-u_i+{\cal O}\left( \frac m{2E}\right) .$$ If we use this relation in (49) and (50) we see that with precision $%
{\cal O}\left( \frac m{2E}\right) $ both matrix elements are equal $$\left\langle \nu _f^M\left| \overline{\nu }^M\gamma ^\mu \left( 1-\gamma
_5\right) \nu ^M\right| \nu _i^M\right\rangle =\left\langle \nu _f^D\left|
\overline{\nu }^D\gamma ^\mu \left( 1-\gamma _5\right) \nu ^D\right| \nu
_i^D\right\rangle +{\cal O}\left( \frac m{2E}\right) .$$ For a tiny neutrino mass it was also shown that, if there are only the left-handed weak currents, the electromagnetic structure for the Dirac and Majorana neutrino smoothly becomes indistinguishable \[15, 39\]. As $\sigma ^{\mu \nu }$ and $\sigma ^{\mu \nu
}\gamma _5$ operators in Eq. (35) change chirality and the spinors in the electromagnetic currents (35) are, with precision $\left( \frac
m{2E}\right) $, the chiral eigenstates $\left( h=-1/2\Longleftrightarrow
P_Ru\cong 0\right) $ the electric and magnetic formfactors must vanish for $%
m_\nu \rightarrow 0$$$M_D\left( q^2\right) \stackrel{m_\nu \rightarrow 0}{\longrightarrow }%
0,\,\,\,\,\,\,\,\,\,\,\,\mbox{ }E_D\left( q^2\right) \stackrel{m_\nu
\rightarrow 0}{\longrightarrow }0.$$ For the other formfactors in (35) we have
$$\begin{aligned}
\left\langle \nu _{-}^D\left| J_\mu ^{EM}\right| \nu _{-}^D\right\rangle &=&
\overline{u}_f(-)\left[ F_D\gamma ^\mu -G_Dq^2\gamma ^\mu \gamma _5\right]
u_i\left( -\right) \nonumber \\
& \approx & \left( F_D+G_Dq^2\right) \overline{u}%
_f(-)\gamma ^\mu u_i\left( -\right) . \end{aligned}$$
For the same reasons (only the left-handed current is present) there is no transition for the $\nu _{+}^D$ Dirac states, so $$0\cong \left\langle \nu _{+}^D\left| J_\mu ^{EM}\right| \nu
_{+}^D\right\rangle =\left( F_D-G_Dq^2\right) \overline{u}_f(+)\gamma ^\mu
u_i\left( +\right) ,$$ and we have $$F_D\approx G_Dq^2.$$ At the same time for the Majorana neutrinos there is (see Eq. (37)) $$\left\langle \nu _{-}^M\left| J_\mu ^{EM}\right| \nu _{-}^M\right\rangle
=-2G_Dq^2\overline{u}_f(-)\gamma ^\mu \gamma ^5u_i\left( -\right) =+2G_Dq^2%
\overline{u}_f(-)\gamma ^\mu u_i\left( -\right) .$$ So, if we compare Eqs. (55) and (57) with (58) we see that in the limit $m_\nu \rightarrow 0$ both electromagnetic currents go smoothly to the same value $$\left\langle \nu _{-}^D\left| J_\mu ^{EM}\right| \nu _{-}^D\right\rangle
\stackrel{m_\nu \rightarrow 0}{\longrightarrow }\left\langle \nu _{-}^M\left|
J_\mu ^{EM}\right| \nu _{-}^M\right\rangle \stackrel{m_{\nu}\longrightarrow 0}{%
\rightarrow }2G_Dq^2\overline{u}_f\gamma ^\mu u_i.$$ In the next Chapter we will see that also the fact that Majorana particles are indistinguishable from their antiparticles will not help and, for the left-handed interacting neutrino, differences in all observables for the Dirac and Majorana neutrino smoothly vanish for $m_\nu \rightarrow 0.$ This statement was formulated in two papers by B. Kayser and R. Shrock \[15\] in 1982 and is known as ”Practical Dirac-Majorana Confusion Theorem”. Since that time many papers have appeared \[40-44\] and in the recent time many e-mail texts have become available on the hep-ph list \[45-50\]. They try to find observables where both neutrinos give the most visible different effects even if their masses are small \[40-42, 44\]. Some of them try to find effects in frame of extensions of the standard model. Some of them are technically correct. Some of them are not concerned about the practical value of the presented concept \[50\]. There are also simply wrong concepts \[43, 45, 48\].
If there are other neutrino interactions (right-handed currents, interaction with scalars) observable differences between the Dirac and Majorana neutrinos could be substantial even for small neutrino mass. But the SM works very well so effects of any of SM extensions must be small at least for presently attainable energies.
Review of various processes.
============================
Processes where the differences between Dirac and Majorana neutrinos are not seen.
----------------------------------------------------------------------------------
The main neutrino processes which measure their masses do not feel the differences between the Dirac and Majorana neutrinos
$\diamondsuit $ In all processes which give the bounds on the neutrino mass e.g. tritium $\beta $ decay $\left( H_1^3\rightarrow H_2^3\rightarrow e^{-}+%
\bar{\nu}_e\right);$ pion and tau decays $\pi ^{+}\rightarrow \mu
^{+}+\nu _\mu $, $\tau ^{-}\rightarrow 2\pi ^{+}3\pi ^{-}(\pi ^0)\nu _\tau $, there is only one neutrino which interacts by the charged current. In these circumstances all differences in measured quantities between the Dirac and Majorana neutrino disappear.
$\diamondsuit $ In the case of flavour neutrino oscillations, differences between both types of neutrinos disappear too \[51\]. It is very easy to see that. Probability for transition $\nu _\alpha \rightarrow \nu _\beta $ is given by $$P\left( \nu _\alpha \rightarrow \nu _\beta ,t\right) =\left|
\sum_{a=1}^nU_{\beta a}e^{-iE_at}U_{\alpha a}^{*}\right| ^2.$$ The Dirac and Majorana neutrinos give unique signals through the different structures of the mixing matrices U. There are more CP violating phases for the Majorana neutrinos. But the formula (60) does not feel these additional phases; they can be eliminated and remaining number of the CP phases is the same like for the Dirac neutrino.
Terrestrial experiments.
------------------------
There are many physical observables which feel the difference between Dirac and Majorana neutrinos. The problem is of course with the size of these effects.
$\diamondsuit $ In general, any process which violates the total lepton number (as for example $e^{-}e^{-}\rightarrow W^{-}W^{-}$, $K^{-}\rightarrow \pi
^{+}\mu ^{-}\mu ^{-},$ $K^{-}\rightarrow \pi ^{+}e^{-}e^{-},$ $\nu
\rightarrow \overline{\nu }$ oscillation, ...) will indicate that neutrinos have the Majorana character. Also the neutrinoless double $\beta $ decay violates the total lepton number. We will comment on it later.
$\diamondsuit $ There are also processes which do not violate the total lepton number and occur for Dirac as well as for Majorana neutrinos. But physical observables (cross sections, angular distributions, energy distributions, decay widths, polarizations) have specific properties which distinguish both types of neutrino.
For example the angular distributions for the processes like $\nu
e^{-}\rightarrow \nu e^{-},$ $\nu N\rightarrow \nu N,$ $e^{+}e^{-}%
\rightarrow \nu \nu $ look different for the Dirac and Majorana neutrinos \[15\].
If we could observe, for example, the photon polarization in the process of neutrino decay $\nu _i\rightarrow \nu _j+\gamma $ then the ratio of the left-handed $\left( M_L\right) $ to right-handed $\left( M_R\right) $ photon polarization distinguishes both types of neutrinos \[22\] $$\frac{M_L\left( \nu _i\rightarrow \nu _j+\gamma \right) }{M_R\left( \nu
_i\rightarrow \nu _j+\gamma \right) }= \left\{
\matrix{ \left(
\frac{m_{\nu _j}}{m_{\nu _i}}\right) ^2 \mbox{ for Dirac neutrinos,} \cr
\\ \mbox{
\thinspace \thinspace \thinspace \thinspace \thinspace \thinspace }%
1\,\,\,\,\,\, \mbox{ for Majorana neutrinos.}} \right.$$ There are many other processes where the difference can be written (but only written not observed). Now we consider two examples which were the places of wrong interpretation in the past \[45, 48\].
$\star $ Let us exam the scattering process $$\nu _\mu +e^{-}\rightarrow \nu _\mu +e^{-}$$ which is measured experimentally \[52\]. We assume that a beam of neutrinos is not a pure negative helicity state $\left(
h=-1/2\right) ,$ there is a mixture of $\left( h=+1/2\right) $ and the density matrix in the helicity basis is $$\rho =\left(
\begin{array}{cc}
\varepsilon & 0 \\
0 & 1-\varepsilon
\end{array}
\right) ,$$ where $0 \leq \varepsilon << 1.$
To be general, we take the coupling of neutrinos and electrons with the neutral boson $%
Z_0$ in the form $$\frac g{2\cos \theta }\left[ \overline{\nu }\gamma ^\mu \left( A_L^\nu
P_L+A_R^\nu P_R\right) \nu +\overline{e}\gamma ^\mu \left(
A_L^eP_L+A_R^eP_R\right) e\right] Z_\mu .$$ Let us define $$y=\frac{E_e^{LAB}}{E_\nu ^{LAB}}=\frac 12\left( 1-\cos \theta _{CM}\right)$$ where $E_{e\left( \nu \right) }^{LAB}$ is energy of outgoing electron (incoming neutrino) in the LAB system, $\theta _{CM}$ is the CM scattering angle. Then we can calculate the electron LAB energy distribution $\left( m_\nu
\approx 0,\mbox{ }m_e\approx 0\right) $$$\begin{aligned}
\frac{d\sigma }{dy}& = &\frac{G_F^2\rho s}{4\pi }\left\{ \left( A_R\right)
^2 \left[ \left( A_R^e\right) ^2+\left( A_L^e\right) ^2\left( 1-y\right)
^2\right] \varepsilon \right. \nonumber \\
&+& \left. \left( A_L\right) ^2 \left[ \left( A_L^e\right)
^2+\left( A_R^e\right) ^2\left( 1-y\right) ^2\right] \left( 1-\varepsilon
\right) \right\}, \end{aligned}$$ where$$A_R=\left\{
\begin{array}{c}
A_R^\nu \;\;
\mbox{ \thinspace \thinspace \thinspace for Dirac,} \\ A_R^\nu -A_L^\nu
\,\,\,\,\mbox{ for Majorana,}
\end{array}
\right.$$ and $$A_L=\left\{
\begin{array}{c}
A_L^\nu \,\,\,\,\;\;
\mbox{ for Dirac,} \\ A_L^\nu -A_R^\nu \mbox{ \thinspace \thinspace
\thinspace \thinspace for Majorana.}
\end{array}
\right.$$ Such distribution is measured by CHARM collaboration \[52\] and the result agrees with the SM $\left( A_L^l=-1+2\sin ^2\theta _W\mbox{ , }%
A_R^l=-1+2\sin ^2\theta _W\mbox{, }A_L^\nu =1\mbox{, }A_R^\nu =0\mbox{, }%
\varepsilon =0\right) .$ But let us assume that we have better data. Do we have any chance to distinguish (at least in principle) the Dirac from Majorana neutrino from the energy distribution (66)? The answer depends on the polarization of initial neutrino $\left( \varepsilon \right) $ and the existence of the right-handed currents $\left( A_R^\nu \right) .$
$\bullet $ If all initial $\nu _\mu $ neutrinos are in the pure state $\left( \varepsilon =0\right) $ and the interaction is pure left-handed $\left( A_R^\nu =0\right) $ then $\frac{d\sigma ^D}{dy}=\frac{%
d\sigma ^M}{dy}$. This situation we have in the SM.
$\bullet $ If $\varepsilon =0$ but the right-handed current exists, then there is a difference in normalization of both cross sections
$$\frac{d\sigma ^M}{dy}=\left( 1-\frac{A_R^\nu }{A_L^\nu }\right) ^2\frac{%
d\sigma ^D}{dy},$$ and the same is true for the total cross section so in principle the Dirac and Majorana neutrinos are distinguishable.
$\bullet $ If $\varepsilon >0$ but there are no the right-handed currents $%
\left( A_R^\nu =0\right) $ both neutrino types are in principle distinguishable. As $\left( A_L^e\right) ^2\neq \left( A_R^e\right) ^2$ the energy distribution is different for the Dirac and Majorana neutrinos.
$\bullet $ The best situation is for $\varepsilon >0$ and $A_R^\nu \neq 0$, there are two factors which change the angular distribution.
With present experimental precision there is no chance to measure such details of the energy distribution. The data agree with the SM $\left( \varepsilon =0\mbox{, }%
A_R^\nu =0\right) $ very well, and only the products of the neutrino and electron couplings are measured $$\left( A_L^\nu A_L^e\right) ^2\mbox{ and \thinspace \thinspace }\left(
A_L^\nu A_R^e\right) ^2.$$ Even if we know from other experiments the electron couplings, from neutrino energy distribution (66) we can find only the sum of the vector $g_V^{\nu _{}}$ and the axial vector $g_A^\nu \,$couplings $A_L^\nu =g_V^\nu +g_A^\nu $ .We can say nothing about $g_V^\nu \,$ and g$_A^\nu $ separately \[46, 47\]. Particularly we cannot conclude that $g_V^\nu \neq 0$ and because of this muon neutrino is a Dirac particle as it was wrongly suggested in \[45\].
$\star $ The other interesting problem arises for production of two neutrinos in the $e^{+}e^{-}$ collision. If the neutrinos are Dirac particles, neutrino and antineutrino are produced, for Majorana neutrino two indistinguishable particles appear in the final state. At first sight the angular distribution of two final neutrinos in the CM frame should look completely different in both cases. For two identical particles in the final state $\left( \nu \nu \right) $ the angular distribution must be forward-backward symmetric, for two different particles $\left( \nu
\overline{\nu }\right) $ there are no special reasons to have this symmetry. Two (three) Feynman diagrams give contribution to the process $%
e^{+}e^{-}\rightarrow \nu \overline{\nu }$ $\left( e^{+}e^{-}\rightarrow \nu
\nu \right) $ in the lowest order \[53\]. Let us specify the momenta and helicities of the particles $$e^{-}\left( p,\sigma \right) +e^{+}\left( \stackrel{}{\overline{p,}}%
\overline{\sigma }\right) \rightarrow \nu \left( k,\lambda \right) +%
\overline{\nu }\left( \overline{k},\overline{\lambda }\right) ,$$ and denote $\Delta \sigma =\sigma -\overline{\sigma }$, $\Delta \lambda
=\lambda -\overline{\lambda }.$
In what follows we use the same denotation for the Z-leptons couplings as in (Eq.64) but the charged current interaction we take in the form $$L_{CC}=\frac g{2\sqrt{2}}B_L\overline{\nu }\gamma ^\mu \left( 1-\gamma
_5\right) lW_\mu ^{+}+h.c.$$ If we neglect the lepton masses then only two (four) helicity amplitudes do not vanish for the Dirac (Majorana) neutrinos \[53\]. They are (we should remember to divide the Majorana amplitude by $\sqrt{2}$ for two identical particles in the final states: $\theta$ is the CM scattering angle) $$\begin{array}{c}
M_M\left( \Delta \sigma =1
\mbox{, }\Delta \lambda =\pm 1\right) =\pm \frac 12f_Re^{i\varphi }\left(
1\pm \cos \theta \right) , \\ \\
M_M\left( \Delta \sigma =-1\mbox{, }\Delta \lambda =\pm 1\right) =\pm \frac
12g_L\left( \mp \cos \theta \right) e^{-i\varphi }\left( 1\mp \cos \theta
\right)
\end{array}$$ for the Majorana neutrinos, and $$M_D\left( \Delta \sigma =\pm 1\mbox{, }\Delta \lambda =-1\right) =\sqrt{2}%
M_M\left( \Delta \sigma =\pm 1\mbox{, }\Delta \lambda =-1\right)$$ for the Dirac neutrinos ($\Delta \lambda =+1$ neutrinos are sterile in this case). All other amplitudes are equal to zero.
The $f_R$, $g_L$ parameters are defined by $$f_R=\sqrt{2}s\left( \frac e{2\sin \theta _W\cos \theta _W}\right) ^2\frac{%
A_R^eA_L^\nu }{s-M_W^2-i\Gamma _WM_W},$$ and $$\begin{aligned}
g_L\left( \cos \theta \right) & = & \sqrt{2}s \left\{
\left( \frac e{\sqrt{2}\sin
\theta _W}\right) ^2\frac{\left| B_L\right| ^2}{\frac s2s\cos \theta -\frac
s2+M_W^2} \right. \nonumber \\
&+& \left. \left( \frac e{2\sin \theta _W\cos \theta _W}\right) ^2
\frac{%
A_L^eA_R^\nu }{s-M_W^2-i\Gamma _iM_W}\right\} . \end{aligned}$$ The cross section is calculated from the formula $$\frac{d\sigma (\Delta \sigma ,\mbox{ }\Delta \lambda )}{d\cos \theta }=\frac
1{32\pi s}\left| M\left( \Delta \sigma ,\mbox{ }\Delta \lambda \right)
\right| ^2,$$ for the polarized particles, and$$\frac{d\sigma }{d\cos \theta }=\frac 1{128\pi s}\sum_{\Delta \sigma ,\Delta
\lambda }\left| M\left( \Delta \sigma ,\mbox{ }\Delta \lambda \right)
\right| ^2,$$ in the unpolarized case.
As the t and u amplitudes describe the scattering of Majorana neutrino, the unpolarized angular distribution is symmetric and completely differs from the Dirac neutrino distribution which is asymmetric. Does it mean that we have ”very easy” way of distinguishing the Dirac from Majorana neutrino \[48\]? Of course not \[49\], even if we could measure the angular distribution for the outgoing neutrinos. The point is that all detectors which we have to our disposal are not able to distinguish the nature of particles from their helicities. Detection of a particle in direction ($\theta ,\varphi )$ and an antiparticle in direction $\left( \pi -\theta ,\mbox{ }\pi +\varphi \right) $ in the CM frame is technically indistinguishable from the situation that one particle with helicity $h=-1/2$ travels to the solid angle ($\theta ,\varphi )$ and the other one to $\left( \pi -\theta ,\mbox{ }\pi +\varphi \right) .$
As an example let us calculate the angular distribution if neutrinos are Dirac particles (for simplicity for polarized electron and positron, $\Delta
\sigma =-1).$ Only one amplitude with $\Delta \lambda =-1$ is necessary, and $$\begin{aligned}
P_D\left( \Delta \sigma =-1
\mbox{, }\Delta \lambda =-1;\theta \mbox{,}\varphi \right) & = &
\left| M_D\left(
\Delta \sigma =-1\mbox{, }\Delta \lambda =-1\right) \right| ^2 \nonumber \\
&=& \left|
g_L(\cos \theta )\right| ^2\frac{\left( 1+\cos \theta \right) }2^2. \end{aligned}$$ In the Majorana case, the final particles are identical, and we have to add incoherently two experimentally indistinguishable situations $$\begin{array}{c}
P_M\left( \Delta \sigma =-1
\mbox{, }\Delta \lambda =-1\right) =\left| M_M\left( \Delta \sigma =-1\mbox{%
, }\Delta \lambda =-1;\theta \mbox{,}\varphi \right) \right| ^2+ \\ \left|
M_M\left( \Delta \sigma =-1
\mbox{, }\Delta \lambda =+1;\pi -\theta ,\mbox{ }\pi +\varphi \right)
\right| ^2 \\ =2\left| M_M\left( \Delta \sigma =-1\mbox{, }\Delta \lambda
=-1;\theta \mbox{,}\varphi \right) \right| ^2=\left| g_L(\cos \theta
)\right| ^2\frac{\left( 1+\cos \theta \right) }2^2,
\end{array}$$ where we use the relation
$$M_M\left( \Delta \sigma \mbox{, }\Delta \lambda ;\theta \mbox{,}\varphi
\right) =M_M\left( \Delta \sigma, - \Delta \lambda ;\pi -\theta ,%
\mbox{ }\pi +\varphi \right) ,$$
which follows from the identity of neutrinos. We see that really two distributions are identical. The same is true for other electron polarization (it simply follows from the relations (72) and (78)). So the total cross section for both electron polarizations are equal$$\sigma _D\left( \Delta \sigma \mbox{, }\Delta \lambda =-1\right) =\sigma
_M\left( \Delta \sigma \mbox{, }\Delta \lambda =-1\right) ,$$ where the angular distributions for the Dirac (76) and the Majorana (77) neutrinos are integrated over full solid angle.
In the Majorana case we can also calculate $$P_M\left( \Delta \sigma \mbox{, }\Delta \lambda =+1;\theta \mbox{,}\varphi
\right) =2\left| M_M\left( \Delta \sigma \mbox{, }\Delta \lambda =+1;\theta
\mbox{,}\varphi \right) \right| ^2$$ which does not exist for Dirac neutrino. But from (72) and (78) it follows that the distribution (79) is equal $$2\left| M_M\left( \Delta \sigma \mbox{, }\Delta \lambda =-1;\pi -\theta ,%
\mbox{ }\pi +\varphi \right) \right| ^2=P_D\left( \Delta \sigma \mbox{, }%
\Delta \lambda =-1;\pi -\theta ,\mbox{ }\pi +\varphi \right)$$ and corresponds to the case of the neutrino which is flying in the direction $%
\pi -\theta ,$ $\pi +\varphi $ and the antineutrino has opposite momentum. So, for the integrated cross section there is $$\sigma _M\left( \Delta \sigma \mbox{, }\Delta \lambda =+1\right) =\sigma
_D\left( \Delta \sigma \mbox{, }\Delta \lambda =-1\right) .$$ For the unpolarized cross section we have to sum (average) over final (initial) particle polarizations and we get $$\frac{d\sigma ^D}{d\cos \theta }=\frac 14\frac 1{32\pi s}\sum_{\Delta \sigma
}P_D\left( \Delta \sigma ,\mbox{ }\Delta \lambda =-1;\theta \right) ,$$ and for the Majorana neutrino where also the $\Delta \lambda =+1$ final neutrino polarization exists $$\frac{d\sigma ^M}{d\cos \theta }=\frac{d\sigma ^D}{d\cos \theta }\left(
\theta \right) +\frac{d\sigma ^D}{d\cos \theta }\left( \pi -\theta \right) ,$$ which is also obvious as the Majorana case is equivalent to $\nu \left(
h=-1/2\right) +\overline{\nu} \mbox{ }\left( h=+1/2\right) $ for the Dirac neutrinos. But for the total, unpolarized cross section we once more recover the equivalence between both types of neutrinos. In order not to take into account the same spin configuration two times, we have to integrate the Majorana cross section only over half of the solid angle and we have $$\sigma _{tot}\left( M\right) =\int\limits_{-1}^0d\cos \theta \frac{d\sigma ^M%
}{d\cos \theta }=\int\limits_{-1}^{+1}d\cos \theta \frac{d\sigma ^D}{d\cos
\theta }=\sigma _{tot}\left( D\right) .$$ We see that the Practical Dirac-Majorana Confusion Theorem still works.
The main problem in distinguishing Dirac from Majorana neutrino is the lack of neutrinos with positive helicity. There are two ways, discussed in literature, how to obtain the neutrino with reversed helicity
\(1) to overtake it \[54\]
\(2) to reverse the spin of the neutrino in an external magnetic field \[22\].
Let us assume that we have the beam of $\pi ^{+}$ with high energy (e.g. 600 GeV from Tevatron) in the laboratory system. In the rest frame of $\pi ^{+}$ the decay $\pi ^{+}\rightarrow \mu ^{+}+\nu _\mu $ looks like in Fig.10. All neutrinos with momentum in the backward direction (with respect to the $%
\pi ^{+}$ beam) will have forward momentum in the LAB frame. As a neutrino spin will not turn after Lorentz transformation, this means that helicity will change from $h_\nu =-1/2$ in the CM system to $h_\nu =+1/2$ in the LAB system.
This phenomenon is possible only if neutrino is massive and its efficiency depends on the neutrino mass. It was shown that m$_{\nu _\mu }\approx \,$10 keV will not be enough if all practical limitations are taken into account \[54\]. But there is also some chance that the muon neutrino will oscillate to the tau neutrino which can be much heavier. Now the result depends on the oscillation probability. It was shown that for U$_{\mu \tau }\approx 0.03$ and m$_{\nu _\tau }>1$ MeV the final results do not look like a completely wild scheme \[55\].
The other possibility considered in literature is to reverse the neutrino helicity by the influence of an external magnetic field \[22\]. The problem is that we need large neutrino magnetic moments and large magnetic fields to obtain visible effect. The SM predicts that for the Dirac neutrino with mass $m_\nu $ the magnetic moment is equal \[56\] $$\mu _\nu =\frac{3eG_Fm_\nu }{8\pi ^2\sqrt{2}}\approx 3\times 10^{-19}\left(
\frac{m_\nu }{1eV}\right) \mu _{Bohr}.$$
Then a magnetic field which is needed for feasible experiment is too large even for astronomical scale. But present limits on neutrino magnetic moments are not so small ($\mu _\nu <1.8\times 10^{-10}\mu _B,$ for reactor antineutrino \[57\]; $\mu _\nu <0.3\times 10^{-11}\mu _B,\,\,\,$from stellar cooling \[58\] ) and stellar magnetic field could have a chance to reverse neutrino helicity .
There is only one terrestrial experimental approach which currently promises to state whether neutrinos are Majorana or Dirac particles - it is the neutrinoless double $\beta $ decay,$\left( \beta \beta \right) _{0\nu }$\[4\]. The quantity measured in the $\left( \beta \beta \right) _{0\nu }$ is the average of neutrino masses \[59\] $$\left\langle m_\nu \right\rangle =\sum_nU_{en}^2m_n.$$ If we could find experimentally that $\left\langle m_\nu \right\rangle
>\kappa $ then we would obtain two important items of information, namely that (i) neutrinos are Majorana type and (ii) at least for one neutrino $%
m_\nu >\kappa .$
Experiments which try to find the amplitudes for the $\left( \beta \beta \right)
_{0\nu }$ decay $$\left( Z,A\right) \rightarrow \left( Z+2,A\right) +2e^{-}$$ are presently conducted in several places, using different even-even nuclei. Up to now the best limit on $\left\langle m_\nu \right\rangle $ has been obtained by the Heidelberg-Moscow collaboration from observing the half- time of $^{76}$Ge \[60\],$$T_{1/2}^{0\nu }>2.0\times 10^{25}\mbox{ year (68\% of C.L.),}$$ which gives$$\left\langle m_\nu \right\rangle <\left( 0.44-1.1\right) \mbox{eV (}68\%%
\mbox{ of C.L.) [61]}$$ depending on the method in which the nuclear matrix element is calculated. Future experiments which are planed will move the bounds \[61\]:
Heidelberg-Moscow coll. $$\begin{aligned}
^{76}Ge\rightarrow 5\times 10^{25}
\mbox{ year } & \Rightarrow & \left\langle m_\nu \right\rangle
\sim 0.2\mbox{ eV} \\
\mbox{NEMO coll.} \hspace{2.5 cm} && \nonumber \\
^{100}Mo\rightarrow 10^{25}\mbox{ year } & \Rightarrow & \left\langle m_\nu
\right\rangle \sim 0.16\mbox{ eV.} \end{aligned}$$
”Half-terrestrial” experiments.
-------------------------------
Experiments which use neutrinos from non terrestrial sources (the sun, supernova) were also considered. One such a possibility in which the solar neutrinos are used was recently proposed $\left[ 62\right] $ . If electron neutrino magnetic moment is in the range of present astrophysical limit $m_\nu <3\times
10^{-12}\mu _B$ there is some chance that strong magnetic fields in the sun will reverse the neutrino helicity. In \[62\] the energy distribution $\frac{d\sigma}{dT}$ for a final electron in the process $$\nu +e^{-}\rightarrow \nu +e^{-}$$ was calculated.As the $\nu _e\left( +\right) $ state is (is not) sterile for Dirac (Majorana) neutrinos the distribution $\frac{d\sigma }{dT}$ differs very much for both kinds of neutrinos (the size of the effect depends on the neutrino density matrix). The electron energy distribution is different for the Dirac and Majorana neutrino mostly for the low electron energy. So experiments which can measure the low energy of outgoing electrons are welcome. The HELLAZ experiment (with the threshold energy 100 keV) seems to be a good place \[62\].
It is also possible to find in literature the arguments which compare neutrino emitted from the supernova with the sun neutrinos \[63\]. The argument is as follows. The observation of neutrinos from the SN 1987 A explosion suggests that the diagonal and the transition moments for Dirac $\nu _e$ neutrino should be small $\mu _\nu <10^{-13}\mu _B$. Such magnetic moments are too small to explain hypothetical anticorrelation in the sun neutrino observation (visible in Davis experiment but not in Kamiokande). For explanation of such anticorrelations we need larger neutrino magnetic moment. So, if neutrinos are of the Majorana type, both observations could agree.
Conclusions.
============
The problem whether the neutrino is identical to its own antiparticle is the central problem in neutrino physics and very important in particle physics too.
If the neutrinos have no masses then
$\bullet \,\,$it was proved (Pauli-Gursey transformation) that without interactions each of Weyl neutrinos is equivalent with massless Majorana neutrinos,
$\bullet \,$if neutrinos interact with gauge bosons through left-handed currents (like in the SM) then only one Weyl neutrino appears in the theory and again we can never distinguish it from the massless Majorana neutrino,
$\bullet \,$if, besides the left-handed also the right-handed currents describe the gauge bosons –neutrino interaction, or massless neutrinos interact with scalar bosons then Weyl and Majorana neutrinos interact differently with the matter. Even if in such case there exist both Weyl neutrinos (equivalent with one massless Dirac neutrino), the beam of any kind of Weyl neutrinos behaves in a different way than the beam of Majorana neutrinos.
If the neutrinos are massive particles, by definition Dirac neutrino ($\nu
\neq \bar{\nu}$) differs from Majorana neutrino ($\nu =
\bar{\nu}$). They interact with the matter in different way even if only SM governs the neutrino interaction. But unfortunately in such a case (only left-handed current) all differences in physical observables disappear in a smooth way with vanishing neutrino mass (Kayser - Shrock theorem). As
- the SM works very well and any interaction signals beyond the SM have not been discovered,
- the neutrino interactions described by the SM are checked experimentally and are very weak ($\sigma _{\nu e}\approx 10^{-44}$cm$^2)$,
- the masses of light neutrinos are much smaller than the masses of charged fermions,
it is extremely difficult to find clear experimental evidence which informs us about the nature of existing light neutrinos.
There is only one terrestrial, experimental test that can reveal it, which is the neutrinoless double beta - decay. Several experimental groups, using different even - even nuclei, placed the upper limit on the lifetime of the $%
(\beta \beta )_{0\nu }$. Up to now, nobody has found the evidence for the $(\beta \beta )_{0\nu }$ decay. However there are plans to improve the upper limit for the $(\beta \beta )_{0\nu }$ by almost one order of magnitude and we can still have hope that the problem if neutrinos are of the Dirac or the Majorana type can be solved in the nearby future.
Acknowledgements {#acknowledgements .unnumbered}
================
This work was supported by Polish Committee for Scientific Researches under Grant No. 659/P03/95/08. I thank J. Gluza for reading the text and valuable remarks.
[99]{} W. Pauli, Letter to L. Meinter and other participants of the conference in Tubingen in 1930, a copy of the letter can be found in W. Pauli, ”Neutrino Physics” ed. by K. Winter, Cambridge Univ. Press, Cambridge 1991, p.4, the first publication was given in W. Pauli, Septieme Conseil de Physique Solvay 1933: ”Noyaux Atomiques”.
E. Fermi, Ricercha Scient. 2 No12 (1933); Zeit.f. Phys. 88 (1934) 161.
E. Majorana, Nuovo Cimento, 14 (1937) 171.
W.H. Furry, Phys. Rev. 56 (1939) 1184.
R. Davis, Phys. Rev. 86 (1952) 976.
F. Reines and C.L. Cowan, Jr., Nature 178 (1956) 446;\
C.L. Cowen et al., Science 124 (1956) 103.
T.D. Lee and C.N. Yang, Phys. Rev. 104 (1956) 254.
C.S. Wu et al., Phys. Rev. 107 (1957) 641.
T.D. Lee and C.N. Yang, Phys. Rev. 105 (1957) 1671;\
L.D. Landau, Nucl. Phys. 3 (1957) 127;\
A. Salam, Nuovo Cimento 5 (1957) 299.
M. Goldhaber, L. Grodzin and A.W. Sunyar, Phys. Rev. 109 (1958) 1015.
W. Pauli, Nuovo Cimento 6 (1957) 204;\
I.A. McLennan, Jr., Phys. Rev. 106 (1957) 821;\
K.M. Case, Phys. Rev. 107 (1957) 307;\
F. Gursey, NuovoCimento 7 (1958) 411.
J.C. Pati and A. Salam, Phys. Rev. Lett. 31 (1973) 661;\
H. Georgi and S.L. Glashow, Phys. Rev. Lett. 32 (1974) 438;\
H. Fritzsch and P. Minkowski, Ann. Phys. (N.Y.) 93 (1975) 193.
T. Yanagida, in ”Proc. of the Workshop on the Unified Theory of Baryon Number in the Univere”, ed. by O. Sawada and A. Sugamoto (KEK report 79-18, 1979), p. 95; M. Gell-Mann,P. Ramond and R. Slansky, in ”Supergravity”,ed. by P. van Nieuwenhuizen and D.Z. Freedman (North Holland, Amsterdam, 1979), p.315.
R. Davis, Jr., B.T. Cleveland and J.K. Rowley, in ”Proc. of the Neutrino Mass Miniconference”, Telemark, Wisconsin, ed. by V. Barger and D. Cline (Univ. of Wisconsin Report No. 186,1980), p. 38.
L.F. Li and F. Wilczek, Phys. Rev. D25 (1982) 143;\
B. Kayser and R.E. Shrock, Phys. Lett. B112 (1982) 137;\
B. Kayser, Phys. Rev. D26 (1982) 1662.
B.L. van der Waerden, Gottinger Nachrichten, 100(1929)18, see also ”Group Theory and Quantum Mechanics”, Springer, 1974.
See e.g. L.H. Ryder, Quantum Field Theory, Cambridge Univ. Press, Cambridge, 1985.
H. Weyl, Zeit. f. Phys. 56 (1929) 330.
P.M.A. Dirac, Proc. Roy. Soc. A, 117 (1928) 610.
S. Bilenky, J. Hosek and S. Petcov, Phys. Lett.B 94 (1980) 495;\
J. Schechter and J. Valle, Phys. Rev. D22 (1980) 2227;\
M. Doi et al., Phys. Lett.B102 (1981) 323;\
J. Kobzarev et al., Sov. J. Nucl. Phys. 32 (1980) 823.
P. Pal and L. Wolfenstein, Phys. Rev. D25 (1982) 766;\
J. Bernabeu, A. Pich and A. Santamaria, Zeit. f. Phys. C30 (1986) 213;\
B. Kayser, in ”CP Violation”, ed. C. Jarlskog (World Scientific, Singapore 1989), p. 334.
R. Shrock, Nucl. Phys. B206 (1982) 359.
L. Wolfenstein, Phys. Lett. B107 (1981) 77.
M. Doi at al., Prog. Theor. Phys. 70 (1983) 1331;\
B. Kayser, Phys. Rev. D30 (1984) 1023;\
B. Kayser, in ”Neutrino Physics”, ed. K. Winter (Cambridge Univ. Press, Cambridge 1989, p.115).
B.Kayser, F.Gibrat-Debu and F.Perrier,”The Physics of Massive Neutrinos”, (World Scietific,Singapore,1988), p.35.
I.Bernstein, M.Ruderman and G.Feinberg, Phys.Rev.132 (1963) 1227.
M.Fukugita and T.Yanagida in ”Physics and Astrophysics of Neutrinos”, ed. M.Fukugita and A.Suzuki (Springer-Verlag,Tokyo,1994), p.1.
The term ”anapol” was introduced for spin 1/2 particles by Y. B. Zeldovich, Zh. Eksp. Teor. Fiz. 33(1958) 1531.
B. Kayser and A.S. Goldhaber, Phys. Rev. D22 (1983) 1023.
F. Boudjema and C. Hamzaoui, Phys. Rev. D43 (1991) 3748.
J. Gluza and M. Zrałek, Phys. Rev. D45 (1992) 1693.
J.C. Pati and A. Salam, Phys. Rev. D10 (1974) 275;\
R.N. Mohapatra and J.C. Pati, Phys. Rev. D11 (1975) 2558;\
R.N. Mohapatra and G. Senjanović, Phys. Rev. D12 (1975) 1502; Phys. Rev. Lett. 44 (1980) 912; Phys. Rev. D23 (1981) 165.
For example: M. Fukugita and T. Yanagida, Phys. Rev. Lett. 58 (1987)1807; A. Zee, Phys. Lett. B93 (1980) 389; B161(1985) 141; S.M. Barr, E.M. Freire and A. Zee, Phys. Rev. Lett. 65 (1990) 2626.
J.E. Kim, Phys. Rev. D14 (1976) 3000;\
M.A. Beg, W.J. Marciano and M. Ruderman, Phys. Rev. D17 (1978) 1395.
Y. Suzuki, in ”Proceedings of the 28th International Conference of High Energy Phys.”, ed. Z. Ajduk and A.K. Wroblewski (World Scientific, 1997), p.273.
W. Li, ”[*Tau Physics*]{}”, Lepton-Photon Conference, Hamburg 1997.
J. Ellis, in ”[*Neutrino 96*]{}”, Proceedings of 17th International Conference in Neutrino Physics and Astrophysics, ed. K. Enqvist, K. Huita and J. Maalampi (World Scientific,1997) p. 541.
B. Kayser, Comm. Nucl. Part. Phys. 14 (1985) 69.
B. Kayser, in ”Proceedings of the Eight Moriond Workshop on Neutrino and Exotic Phenomena”, held in Les Arcs, France, January, 1988.
S.T. Petcov, Phys. Lett. B139 (1984) 421.
T. Garavaglia, Phys. Rev. D29 (1984) 387.
S.M. Bilenky, E.Ch. Christova and N.P. Nedelcheva, Phys. Lett. B161 (1985) 397.
G. Auriemma, Y. Strivastava and A. Widom, Phys. Lett. B195 (1987) 254.
T. Chhabra and P. Ram Babu, Phys. Rev. D46 (1992) 903.
R. Plaga, hep-ph/9610545 (1996).
S. Hannestad, hep-ph/9701216 (1997).
B. Kayser, hep-ph/9703294 (1997).
A. Hoefer, hep-ph/9705362 (1997).
S.H. Hausen, hep-ph/9708359 (1997).
Y.N. Srivastava and A. Widom, hep-ph/9708445 (1997).
S.M. Bilenky and S.T. Petcov, Rev. Mod. Phys. 59 (1987) 671.
CHARM II Collaboration, Phys. Lett. B281 (1992) 159; Phys. Lett. B320 (1994) 203; Phys. Lett. B335 (1994) 246.
J. Gluza and M. Zrałek, Phys. Rev. D48 (1993) 5093; Phys. Rev. D51 (1995) 4695.
B.Kayser,in Proc.Moriond Workshop (1984).
M. Carena, B. Lampe and C.E.M. Wagner, Phys. Lett. B317, 112 (1993).
M. Marciano and A. J. Sanda, Phys.Lett. B67(1977)303;
B.W.Lee and R.E Shrock, Phys. Rev. D16 (1977) 1444.
A.V. Kyuldjiev, Nucl. Phys. B243 (1984) 387.
G.G. Raffelt, Phys. Rev. Lett. 64 (1990) 2856.
M. Doi, T. Kotani and E. Takasugi, Prog. Theor. Phys. Supp. 83 (1985) 1.
Heidelberg-Moscow collab., Phys. Lett. B356 (1995) 450.
Y. Toshuka, ”[*Neutrino Physics-non Accelarator*]{}”, Lepton-Photon Conf., Hamburg, 1997.
V.B. Semikoz, Nucl. Phys. [**B498**]{}, 39 (1997).
S.P. Rosen, in ”Proceedings of the First International Symposium on Particle, Strings and Cosmology”, March, 1990.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
\
\
\
\
\
\
title: 'What season suits you best? Seasonal light changes and cyanobacterial competition.'
---
Nearly all living organisms, including some bacterial species, exhibit biological processes with a period of about 24 h called circadian (from the Latin circa, about, dies, day) rhythms. These rhythms allow living organisms to anticipate the daily alternation of light and darkness. Experiments carried out in cyanobacteria have shown the adaptive value of circadian clocks. In these experiments a wild type cyanobacterial strain (with a 24 h circadian rhythm) and a mutant strain (with a longer or shorter period) grow in competition. In different experiments the external light dark cycle was changed in order to match the circadian period of the different strains, revealing that the strain whose circadian period matches the light-dark has a larger fitness. As a consequence the initial population of one strain grows while the other decays.
These experiments were made under fixed light and dark intervals. However, in Nature this relationship changes according to the season. Therefore, seasonal changes in light could affect the results of the competition. Using a theoretical model we analyze how modulation of light can change the survival of the different cyanobacterial strains. Our results show that there is a clear shift in the competition due to the modulation of light, which could be verified experimentally.
Introduction
============
Circadian rhythms, oscillations with approximately 24 h period in many biological processes, are found in nearly all living organisms. Until the mid-1980s, it was thought that only eukaryotic organisms had a circadian clock, since it was assumed that an endogenous clock with a period of $\tau=24$ h would not be useful to organisms that often divide more rapidly [@Johnson]. However, in 1985, several research groups discovered that in cyanobacteria there was a daily rhythm of nitrogen fixation [@Stal; @Grobbelaar; @Mitsui]. Huang and co-workers were the first to recognize that a strain of Synechococcus, a unicellular cyanobacterium, had circadian rhythms [@Huang]. This transformed Synechococcus in one of the simplest models for studying the molecular basis of the circadian clock.
The ubiquity of circadian rhythms suggests that they confer an evolutionary advantage. The adaptive functions of biological clocks are divided into two hypotheses. The external advantage hypothesis supposes that circadian clocks allow living organisms to anticipate predictable daily changes, such as light/dark, so they can schedule their biological functions like feeding and reproduction at appropriate times. In contrast to this hypothesis, it has been suggested that circadian clocks confer adaptive benefit to organisms through temporal coordination of their internal physiology (intrinsic advantage) [@Sharma]. In this case, the circadian clock should be of adaptive value in constant conditions as well as in cyclic environments.
In order to study if circadian clocks provide evolutionary advantages Woelfle and co-workers tested the relative fitness under competition between various strains of cyanobacteria [@Woelfle]. They carried out experiments where a wild-type strain ($\tau=25$ h) of cyanobacteria and mutant strains, with shorter ($\tau=22$ h) and longer ($\tau=30$ h) periods, were subjected to grow in competition with each other under light-dark (LD) cycles of different periodicity. They found that the strain which won the competition was the one whose free-running period matched closely the period of the LD cycle. This difference in fitness was observed despite the fact that the growth rates were not significantly different when each strain was grown with no competition. Also, mutant strains could outcompete wild-type strains under continuous light (LL) conditions, suggesting that endogenous rhythms are advantageous only in rhythmic environments [@Woelfle]. This study provided one of the most convincing evidence so far in support of fitness advantages of synchronization between the endogenous period and the period of environmental cycles.
Ouyang et al. suggested an explanation for fitness differences: this could be due to competition for limiting resources or secretion of diffusible factors that inhibit the growth of other cyanobacterial strains [@Ouyang]. Roussel et al. proposed mathematical models in order to test which of these hypotheses was more plausible [@Roussel]. They found that the model based on mutual inhibition was consistent with the experimental observations of [@Ouyang]. In this model the mechanism of competition between cells involves the production of a growth inhibitor, which is produced only during the subjective day (sL) phase, and that growth occurs only in light phase.
Each of the experiments and computational simulations mentioned before had equal amounts of light and dark exposure. However, in Nature the relationship is not constant, and the duration of sunlight in a day changes according to the season and the latitude. The circadian system has to adapt to day length variation in order to have a functional role in optimizing seasonal timing and generating the capacity to survive at different latitudes [@Hut].
In this work we will test how day length variation plays a role in the competition between different strains of cyanobacteria.
The model
=========
For modelling the growth of each cyanobacterial strain we use the model introduced by Gonze et al. [@Gonze], that is based on a diffusible inhibitor with a light sensitive oscillator to represent the cellular circadian oscillator. The evolution equations of cell population $N_i$ and the level of inhibitor $I$ are:
$$\begin{array}{rl}
\displaystyle \frac{dN_i}{dt}=k_iN_i(1-\sum_{j=1}^n N_j), & \displaystyle \frac{dI}{dt}=\sum_{i=1}^n N_i (p_i- \frac{V_{max}I}{K_M+I})
\end{array}$$
$$\begin{array}{rl}
k_i=\begin{cases}
k & \mbox{in L and $i$ in sL or $I<I_c$}\\ 0 & \mbox{otherwise,}
\end{cases} & p_i=\begin{cases}
p & \mbox{if $i$ in sL}\\ 0 & \mbox{otherwise.}
\end{cases}
\end{array}$$
In these equations, $N_i$ is the number of cells of strain $i$, $k_i$ is the growth rate of each strain, $p$ measures the rate of inhibitor production, $V_{max}$ is the maximum rate and $K_M$ is the Michaelis constant characterizing the enzymatic degradation of the inhibitor.
Following the work of Gonze et al. we use a modified version of the Van der Pol oscillator to produce sustained circadian oscillations [@Gonze]. Also, we use the same parameters which were found to be in agreement with experimental data obtained by Woelfle et al. [@Woelfle].
In Fig. \[fig:model\] we present a schematic plot of the model, that shows how the growth of each population is coupled with the circadian oscillator. As can be seen in this figure, when we modify the length of the light (L) phase the overlaps 1 and 2 change, so the growth of each strain is altered affecting the competition.
![Following the schematic explanation of Gonze et al. [@Gonze], we show how the model works. The inhibitor $I$ is produced in 3, during the sL phase, and it is degraded during the entire day. Each strain grows in 1, if its sD phase overlaps with L and $I<I_c$, and in 2, when its SL phase overlaps with L.[]{data-label="fig:model"}](model_v3.eps){width="8"}
Results
=======
We want to test how day length variation can modify the competition between different cyanobacteria strains. In many organisms, a photoperiodic response is reflected in a physiological change. Photoperiodic responses are common amongst organisms from the equator to high latitudes and have been observed in different types of organisms, from arthropods to plants. Diapause (a suspension of development done by insects), migration and gonadal maturation are examples of these annual changes controlled by photoperiod. These biological processes are triggered as soon as the day length reaches certain duration, known as the critical photoperiod. Even near the equator, where day length changes are very small through the whole year, they are used to synchronize reproductive activities with annual events.
In Fig. \[fig:daylength\] we show how day length varies depending on the latitude. The figure compares the day length on the twenty-first day of each month in three cities in South America. We show Quito (Ecuador), which sits near the equator in latitude $0^{\circ}15'$, Jujuy (Argentina), which is located near the tropic of Capricorn in latitude $24^{\circ} 01'$ and Ushuaia, the southernmost city in Argentina which lies in latitude $54^{\circ} 48'$. We can see that during the equinoxes, all places receive 12 hours of daylight.
![Day length over the course of 2012 at different latitudes in South America.[]{data-label="fig:daylength"}](daylength.ps){width="6"}
We simulate the seasonal fluctuations in day length by adding or subtracting minutes of light-time every day to the external LD cycle. For example, if we add 12 minutes of light per day in a LD12:12 cycle, after five days the external cycle of LD has 13 hours of light and 11 of darkness. We initiate competition between equal fractions of wild-type strain and long-period mutant and equal amounts of light and dark. We dilute the culture after 8 light-dark cycles by dividing by a factor of 100 the variables $N_1$, $N_2$ and $I$. In this way we mimic experiments in cultures, that were diluted and sampled every 8 days [@Woelfle].
First, we analysed the case in which there was a phase of coexistence. For $T=28$, the period of the LD cycle has an intermediate value between the free-running period (FRP) of the two strains and both strains can coexist for a long time. However, when we allowed the days to become longer and the nights shorter, after some days the coexistence was broken, as we show in Fig. \[fig:fig2\]. This is due to the external period $T$ that starts to get closer to the FRP of the long-period mutant.
![(A) The outcome of competition between wild-type (continuous line) and long-period mutant (dashed line) shows coexistence between the two strains for $T=28$ h. (B) Coexistence ends after $\approx 8$ days of adding $3$ minutes of light-time per day. The long-period mutant can win competition as its free-running period becomes closer to the LD cycle.[]{data-label="fig:fig2"}](fig2_T28convivencia.ps){width="7"}
In Fig. \[fig:n1\] we show in the left panel the fraction of cells belonging to the wild-type strain as a function of time with fixed LD cycles (red continuous curve) and in the case in which we added 30 minutes of light time each day (blue dashed curve). The first days both curves are similar, but from the second day, the long-period strain has a FRP closest to the period of the LD cycle. The fraction of wild-type strain starts to decrease and is out-competed. In the right panel we show the corresponding fraction of long-period mutant cells in the same cases.
The results of our numerical simulations could be tested experimentally. This would be very simple, since the cultures do not need to be diluted. It is only needed to sample the culture at regular intervals to determine the composition of the population and verify a difference of about ten percent in the two cultures after eight days.
Starting from this simple test, we looked for non trivial effects in a longer experiment. We found an interesting effect that can be observed in Fig. \[fig:fig3\]. In this simulation we added 12 minutes of light time each day. In the first days, the growth was the expected. The wild-type strain could out-compete the long period mutant strain, since the external cycle was LD12:12. But after eight days, when the day was longer than 13 hours, a crossover was observed. The mutant strain started to win the competition because its endogenous period was closer now to the external cycle. This effect could also be tested experimentally, however, in this case the dilution of the culture every 8 days would be needed.
![Effect of modulation in light time (30 minutes per day) on the outcome of competition between strain 1 (wild-type, $\tau=25 h$) and strain 2 (long-period mutant, $\tau=30 h$) in (A) LD12:12 and (B) LD15:15 cycles. Fraction of strain 1 (left panel) and strain 2 (right panel) are shown as a function of time; red with modulation and blue with fixed cycles.[]{data-label="fig:n1"}](fig1.ps){width="9"}
![Competition between long-period mutant and wild-type strains in a LD12:12 cycle for the same parameters as in Fig. \[fig:n1\], but adding 12 minutes per day the light time. A crossover is observed after 8 days.[]{data-label="fig:fig3"}](fig3.ps){width="8"}
Conclusions
===========
The mechanisms underlying the enhancement of reproductive fitness remain still unknown. Despite numerous models have been tested, each has some evidence that supports it and none can be excluded at this time [@JohnsonReview]. In this work we used a diffusible inhibitor model, so our predictions in the growth rates changes could be useful to test the validity of this mechanism.
Our study is motivated by fluctuations in the day length throughout the year which are reflected in organisms behaviour. We studied how these fluctuations affect the competition between different strains of cyanobacteria. We found non-trivial effects which could be tested experimentally. In the first case we determinate the composition of two strains under competition after eight days when the light is modulated. The prediction of these numerical simulations can be tested in a simple experiment where no dilution is needed. We also propose a second experiment where dilution in the cultures is necessary, which allows for a non trivial effect such as a crossover to be observed.
[99]{}
Johnson, C.H., S.S. Golden, M. Ishiura, and T. Kondo. Circadian clocks in prokaryotes. *Mol. Microbiol.* 21:5-11 (1996).
Stal, L.J. and W.E. Krumbein. Nitrogenase activity in the non-heterocystous cyanobacterium Oscillatoria sp. grown under alternating light-dark cycles. *Arch. Microbiol.* 143:67-71 (1985).
Grobbelaar, N., T.C. Huang, H.Y. Lin and T.J. Chow. Dinitrogen-fixing endogenous rhythm in Synechococcus RF-1. *FEMS Microbiol. Lett.* 37:173-177 (1986).
Mitsui, A., S. Kumazawa, A. Takahashi, H. Ikemoto, and T. Arai. Strategy by which nitrogen-fixing unicellular cyanobacteria grow photoautotrophically. *Nature* 323:720-722 (1986).
Huang, T.C., and T.J. Chow. Characterization of the rhythmic nitrogen-fixing activity of Synechococcus RF-1 at the transcription level. *Curr. Microbiol.* 20:23-26 (1990).
Vaze, K.M. and V.K. Sharma. On the Adaptive Significance of Circadian Clocks for Their Owners. *Chronobiol. Int.* 30:413-433 (2013).
Woelfle, M.A., Y. Ouyang, K. Phanvijhitsiri and C.H. Johnson. The Adaptative Value of Circadian Clocks: An Experiment Assesment in Cyanobacteria *Curr. Biology* 14:1481-1486 (2004).
Ouyang,Y., C.R. Andersson, T. Kondo, S. Golden and C.H. Johnson. Resonating circadian clocks enhance fitness in cyanobacteria. *PNAS* 95:8660-8664 (1998).
Roussel, M., D. Gonze and A. Goldbeter. Modelling the Differential Fitness of Cyanobacterial Strains whose Circadian Oscillators have Different Free-running Periods: Comparing the Mutual Inhibition and Substrate Depletion Hypotheses. *J. Theor. Biol.* 205:321-340 (2000).
Hut, R.A. and G. M. Beersma. Evolution of time-keeping mechanisms: early emergence and adaptation to photoperiod. *Phil. Trans. R. Soc. B* 366:2141-2154 (2011).
Gonze, D., M. Roussel and A. Goldbeter. A Model for the Enhancement of Fitness in Cyanobacteria Based on Resonance of a Circadian Oscillator with the External Light-Dark Cycle. *J. Theor. Biol.* 214:577-597 (2002).
Ma, P., M. A. Woelfle, and C. H. Johnson. An evolutionary fitness enhancement conferred by the circadian system in cyanobacteria. *Chaos, Solitons & Fractals* 50:65-74 (2013).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Characteristic integrals of Toda field theories associated to simple Lie algebras are presented in the most explicit forms, both in terms of the formulas and in terms of the proofs.'
address: 'Department of Mathematics and Statistics, Utah State University, Logan, UT 84322-3900'
author:
- Zhaohu Nie
date: 'February 15, 2013'
title: On characteristic integrals of Toda field theories
---
Introduction
============
First consider the famous Liouville equation, for which we take the following version: $$\label{liouville}
u_{xy}=-e^{2u},$$ where $x$ and $y$ are the independent variables, and $u=u(x,y)$ is an unknown function. Let $$\label{cl for l}
I=u_{xx}-u_x^2.$$ Then $I_{y}=\tfrac{\p}{\p y}I=0$ for a solution $u(x,y)$ to . $I$ is thus called a *characteristic integral* of .
Toda field theories are generalizations of the Liouville equation . Let $\fg$ be a simple Lie algebra of rank $n$ with Cartan matrix $A=(a_{ij})_{i,j=1}^n$. For $1\leq i\leq n$, let $u^i=u^i(x,y)$ be $n$ unknown functions of the independent variables $x$ and $y$. The Toda field theory associated to $\fg$ is $$\label{toda}
u^i_{xy}=-e^{\rho_i}:=-\exp\bigg({\sum_{j=1}^n a_{ij}u^j}\bigg),\quad 1\leq i\leq n.$$ The Liouville equation is the Toda field theory associated to $A_1=\fs\fl_2$.
Toda field theories are important integrable systems and have rich properties. (For surveys on them, see for example [@BBT; @LS-book].) In this paper, we are concerned with the explicit forms of their characteristic integrals.
A differential polynomial in the $u^i(x,y)$ for $1\leq i\leq n$ is a polynomial in the $u^i$ and their partial derivatives with respect to $x$ of various orders $u^i_x, u^i_{xx},\cdots$.
A *characteristic integral* of the Toda field theory is a differential polynomial $I$ in the $u^i(x,y)$ for $1\leq i\leq n$ such that $I_y=0$ for solutions $u^i(x,y)$ to .
By the symmetry between $x$ and $y$ in Toda field theories , there are differential polynomials $\tilde I$ in the $u^i$ and their $y$-partial derivatives such that $\tilde I_x=0$ but they follow the same patterns as those in the above definition. For simplicity, we sometimes write $\p$ for $\p_x$.
For a differential mononomial in the $u^i$, we call by its *degree* the sum of the orders of differentiation multiplied by the corresponding algebraic degrees for the factors. Therefore the $I$ in has homogeneous degree $2$. Since a differential polynomial of some characteristic integrals is another such integral, the characteristic integrals form a differential algebra. The structure theorem (see for example [@FF-inv]\*[Theorem 1]{} for the affine version) asserts that it is a *polynomial* differential algebra generated by $n$ *primitive characteristic integrals*. Furthermore these generators are homogeneous and their degrees are equal to the degrees of the Lie algebra $\fg$. Recall that the algebra of adjoint-invariant functions on $\fg$ is a polynomial algebra on $n$ homogenous generators, whose degrees we call the degrees of $\fg$ [@K1].
Many works [@Leznov-where; @W-sym; @gen-W-sym] are devoted to the characteristic integrals, very often under different names such as *local conservation laws, chiral currents* or *intermediate integrals* and from different viewpoints. Furthermore many works [@LS-survey; @ho; @ZS; @ian's] are concerned with using these characteristic integrals to obtain explicit solutions to the original Toda field theories by the method of Darboux.
However to this author, the results about characteristic integrals for Toda field theories are not explicit enough, in terms of both the formulas and the proofs. Therefore we would like to present the concrete formulas to compute the primitive characteristic integrals together with new and self-contained proofs.
Our general formula in Theorem \[ds gauge\] employes the zero curvature representation [@LS] of the Toda field theories under a Drinfeld-Sokolov gauge [@DS]. When the Lie algebra has a non-branching representation (see ), we present a more concrete formula in Theorem \[quick\]. We stress that in both cases, we prove that we obtain characteristic integrals directly.
Let us introduce more terminology before we present our main results. Let $\fh\subset \fg$ be a Cartan subalgebra, and we denote the corresponding set of roots of $\fg$ by $\Delta$, the sets of positive/negative roots by $\Delta_\pm$, and the set of positive simple roots by $\pi=\{\a_i\}_{i=1}^n$. Let $\fg=\fh\oplus \bigoplus_{\a\in \Delta} \fg_\a$ be the root space decomposition.
For a root $\a$, define its height by $\on{ht}(\a)=\sum_{i=1}^n c_i$ if $\a=\sum_{i=1}^n c_i \a_i$. Also define the standard height gradation $$\label{height}
\fg=\bigoplus_{k} \fg_k, \quad \fg_k=\bigoplus_{\on{ht}(\a)=k} \fg_\a,\ \fg_0=\fh.$$ We also denote by $\fn=\bigoplus_{\a\in \Delta_+} \fg_\a=\bigoplus_{k>0} \fg_k$ the maximal nilpotent subalgebra, and by $N$ the corresponding unipotent group. For $\a\in \Delta_+$, let $e_\a$ and $e_{-\a}$ be root vectors in the root spaces $\fg_\a$ and $\fg_{-\a}$ such that for $H_\a=[e_\a,e_{-\a}]\in \fh$, we have $\a(H_\a)=2$. Then the Cartan matrix $A=(a_{ij})_{i,j=1}^n$ of $\fg$ is defined by $a_{ij}=\a_i(H_{\a_j})$.
Let us recall the zero curvature representation of following [@LS]. Let $$\label{players}
{\bf u}=\sum_{i=1}^n u^i_x H_{\a_i}, \quad
\e=\sum_{i=1}^n e_{-\a_i}, \quad
Y=\sum_{i=1}^n e^{\rho_i} e_{\a_i},
$$ where as in $\rho_i=\sum_{j=1}^n a_{ij} u^j$. Then the Toda field theory is equivalent to the following zero curvature equation $$\label{zero curv}
[-\p_x+\e+{\bf u},\, \p_y+Y]=0.$$
Now let us recall the definition of a Kostant slice $\fs\subset \fg$ [@K2], which is used in a Drinfeld-Sokolov gauge [@DS]. Let $\fs$ be a complement of $[\e,\fg]$ in $\fg$, that is, $$\label{split}
\fg\cong \fs\oplus [\e,\fg].$$ Then by [@K2], $\fs\subset \fn$, and $\dim(\fs)=n$ is equal to the rank. We call $\fs$ a Kostant slice, and let $\{s_j\}_{j=1}^n$ be a homogeneous basis of $\fs$ with respect to the height gradation . By [@DS], we can bring the first element in into its Drinfeld-Sokolov gauge. More precisely, there exists an element $g\in N$ (whose components are differential polynomials of the $u^i$) such that $$\label{bring up}
g^{-1} (-\p_x+\e+{\bf u}) g=-\p_x + \e + {\bf I},\quad {\bf I}=\sum_{j=1}^n I_j s_j\in \fs,$$ where the components $I_j$ are differential polynomials of the $u^i$.
\[ds gauge\] For the solutions $u^i$ to , the differential polynomials $I_j$ for $1\leq j\leq n$ defined in through a Drinfeld-Sokolov gauge are primitive characteristic integrals, that is, $\p_y I_j=0$.
The calculations of $g\in N$ and the $I_j$ in can be done by an inductive formula given in [@DS]. In practice, a computer software such as Maple solves it easily in simple cases. Furthermore very often there are more direct formulas for the characteristic integrals as we explain now.
Take an irreducible representation $\phi:\fg\to \text{End}\, V$. Let the $\beta_k\in \fh^*$ for $1\leq k\leq m$ be the weights of $\phi$, and $$\label{wt decomp}
V=\bigoplus_{k=1}^m V_{\beta_k},$$ the weight space decomposition. We assume that $\dim V_{\beta_k}=1$ and also that the representation $\phi$ does not branch. That is, for each weight $\beta_k$ there is one unique negative simple root $-\a_{i_k}$ such that $\beta_k-\a_{i_k}$ is another weight of $\phi$. Order our weights such that $\beta_{k+1}=\beta_k-\a_{i_k}$ for $1\leq k\leq m-1$ and we draw the following weight diagram $$\label{down}
\beta_m\os{-\a_{i_{m-1}}}\longleftarrow \cdots \os{-\a_{i_2}}\longleftarrow \beta_2\os{-\a_{i_1}}\longleftarrow \beta_1.$$ Such non-branching representations are the cases for the first fundamental representations of the Lie algebras $A_n, B_n, C_n$ and $\fg_2$.
\[quick\] If a non-branching representation $\phi$ as above exists with weight diagram , then we have $$\label{Pawel said}
[(\p-\beta_{{1}}(\bu))(\p-\beta_{2}(\bu))\cdots(\p-\beta_{m}(\bu)),\p_y]=0,$$ for a $\bu$ satisfying or equivalently the Toda filed theory . Here $\p=\p_x$, and the product is in the sense of composition for operators on functions of $x$ and $y$. The product is non-commutative, and we strictly follow the order. If by the Leibniz rule, we expand $$(\p-\beta_{{1}}(\bu))(\p-\beta_{2}(\bu))\cdots(\p-\beta_{m}(\bu))=\p^{m}+\sum_{j=1}^{m} I_j \p^{m-j},$$ then implies that $$\label{reason}
\p_y I_j=0,\quad 1\leq j\leq m.$$
We will prove the theorems in the next section. We present examples of Theorem \[quick\] in Section 3, and an example of Theorem \[ds gauge\] in Section 4.
[**Acknowledgment.**]{} The author thanks Ian Anderson for introducing this topic to him. He thanks Anderson and Pawel Nurowski for helps with Maple. He also thanks László Feh[é]{}r and Luen-Chau Li for their interests and useful correspondences.
Proofs of the theorems
======================
We present direct proofs of our theorems in this section.
Suppose that for the $g\in N$ in , we have $$g^{-1}(\p_y + Y)g=\p_y + \tilde Y.$$ Since $g\in N$ and $Y\in \fg_1$ by , we have $\tilde Y\in \fn$. Suppose $\tilde Y=\sum_{i=1}^p Y_i$ by the height decomposition , where $p$ is the biggest height of $\fg$.
By the invariance of the zero-curvature equation under the adjoint action, from we get $$\label{conjged}
\bigg[-\p_x + \e + \sum_{j=1}^n I_j s_j, \p_y + \sum_{i=1}^{p} Y_i\bigg]=0.$$
We will prove by induction that all the $Y_i=0$. This then implies that $\p_y I_j=0$ for $1\leq j\leq n$.
First recall the basic result of Kostant [@K1] that $\ker(\on{ad}_\e) \cap (\fh\oplus \fn)=0$. The term on the left of with height zero is $[\e, Y_1]=0$, which then implies that $Y_1=0$.
Now assume $i\geq 2$ and that $Y_{j}=0$ for $j\leq i-1$. Then since the $s_j\in \fn$, the term on the left of with height $i-1$ is $$[\e, Y_i] - \sum_{\on{ht}(s_j)=i-1} (\p_y I_j) s_j =0.$$ By the decomposition , we get $[\e, Y_i]=0$ which then implies that $Y_i=0$.
In the weight decomposition , we let $v_{\beta_1}\in V_{\beta_1}$ be a weight vector for the highest weight, and inductively we define the other weight vectors by $v_{\beta_k}=\phi(e_{-\a_{i_{k-1}}})v_{\beta_{k-1}}$ for $2\leq k\leq m$ by . Therefore by $$\label{by ep}
\phi(\e) v_{\beta_{k-1}}=v_{\beta_{k}},\quad 2\leq k\leq m.$$
The zero curvature equation is the compatibility condition for the following system of equations. Let ${\psi}(x,y)=\sum_{k=1}^m \psi_k(x,y) v_{\beta_k}$ be a function of $x$ and $y$ with values in $V$. Then implies that the following system of equations has solutions $$\label{wave eqn}
\begin{cases}
(-\p_x+\phi({\bf u}+\e))\psi=0\\
(\p_y+\phi(Y))\psi=0.
\end{cases}$$
Then by and $\bu\in \fh$ , the first equation, at the weight vector $v_{\beta_k}$, means that $$\label{induction}
(\p_x - \beta_k (\bu)) \psi_k=\psi_{k-1},\quad 2\leq k\leq m.$$ When $k=1$, we actually have $$(\p_x-\beta_1 (\bu))\psi_1=0,$$ since $\beta_{1}$ is the highest weight. Therefore combining them, we have $$(\p-\beta_{1}(\bu))(\p-\beta_{2}(\bu))\cdots(\p-\beta_{m}(\bu)) \psi_m=0.$$ On the other hand, the second equation in , at the lowest weight vector $v_{\beta_m}$, quickly gives that $$\p_y \psi_m=0.$$ Therefore the above two equations have a solution $\psi_m=\psi_m(x,y)$, which is general. The implied compatibility condition is exactly . The rest of the theorem is clear.
Examples for Theorem \[quick\]
==============================
Non-branching representations occur for the first fundamental representations of the Lie algebras $A_n, B_n, C_n$ and $\fg_2$. Keeping our spirit of being as explicit as possible, we present the formulas in these cases. In this section we follow [@FH] for notation and choices of root vectors. The Cartan subalgebras $\fh$ always consist of diagonal matrices. We let $L_i\in \fh^*$ denote the linear function of taking the $i$th element on the diagonal. We also let $E_{ij}$ denote the matrix with a 1 at the $(i,j)$-position and zero everywhere else.
The simple roots for $A_n$ are $\a_i=L_i-L_{i+1}$ and the $H_{\a_i}=E_{i,i}-E_{i+1,i+1}$, $1\leq i\leq n$. The Cartan matrix is $$\begin{pmatrix}
2 & -1 & & &\\
-1 & 2 & -1 & & \\
& \ddots & \ddots & \ddots& \\
& &-1 & 2 & -1\\
& & & -1 & 2
\end{pmatrix}$$ The degrees of $A_n$ (that is, the degrees of primitive adjoint-invariant functions) are $2,3,\cdots,n+1$. The weight diagram for the first fundamental representation is $$L_{n+1}\os{-\a_n}\longleftarrow \cdots \os{-\a_2}\longleftarrow L_2\os{-\a_1}\longleftarrow L_1.$$ The $\bu$ in is $$\bu=\on{Diag}(u^1_x, u^2_x - u^1_x, \cdots, u^n_x - u^{n-1}_x, -u^n_x)$$
Consider the expansion $$\label{An det}
\begin{split}
&\quad(\p-u^1_x)(\p+u^1_x-u^2_x)\cdots(\p+u^{n-1}_x-u^n_x)(\p+u^n_x)\\
&=\p^{n+1}+\sum_{j=1}^n I_j\p^{n-j}.
\end{split}$$ Then the $I_j$ for $1\leq j\leq n$ are primitive characteristic integrals of the $A_n$ Toda field theory .
The simple roots for $C_n$ are $\a_i=L_i-L_{i+1}$ for $1\leq i\leq n-1$ and $\a_n=2L_n$. Also $H_{\alpha_i}=E_{i,i}-E_{i+1,i+1}-E_{n+i,n+i}+E_{n+i+1,n+i+1}$ for $1\leq i\leq n-1$ and $H_{\alpha_n}=E_{n,n}-E_{2n,2n}$. The Cartan matrix is $$\begin{pmatrix}
2 & -1 & & &\\
-1 & 2 & -1 & & \\
& \ddots & \ddots & \ddots& \\
& &-1 & 2 & -1\\
& & & -2 & 2
\end{pmatrix}$$ The degrees of $C_n$ are $2,4,\cdots,2n$. The weight diagram for the first fundamental representation is $$-L_1\os{-\a_1}\longleftarrow \cdots \os{-\a_{n-1}}\longleftarrow -L_n\os{-\a_n}\longleftarrow L_n\os{-\a_{n-1}}\longleftarrow \cdots\os{-\a_2}\longleftarrow L_2\os{-\a_1}\longleftarrow L_1.$$ The $\bu$ in is $$\bu=\on{Diag}(u^1_x, u^2_x - u^1_x, \cdots, u^n_x - u^{n-1}_x, -u^1_x, -u^2_x + u^1_x, \cdots, -u^n_x + u^{n-1}_x).$$
Consider the expansion $$\label{Cn det}
\begin{split}
&\quad (\p-u^1_x)(\p+u^1_x-u^2_x)\cdots(\p+u^{n-1}_x-u^n_x)\\
&\quad\quad (\p+u^{n}_x-u^{n-1}_x)
\cdots(\p+u^2_x-u^1_x)(\p+u^1_x)\\
&=\p^{2n}+\sum_{j=1}^n I_j\p^{2n-2j}+\sum_{j=1}^{n-1} J_j\p^{2n-2j-1}.
\end{split}$$ Then the $I_j$ for $1\leq j\leq n$ are primitive characteristic integrals of the $C_n$ Toda field theory, and the $J_j$ are some differential polynomials in them.
The simple roots for $B_n$ are $\a_i=L_i-L_{i+1}$ for $1\leq i\leq n-1$ and $\a_n=L_n$. Also $H_{\alpha_i}=E_{i,i}-E_{i+1,i+1}-E_{n+i,n+i}+E_{n+i+1,n+i+1}$ for $1\leq i\leq n-1$ and $H_{\alpha_n}=2E_{n,n}-2E_{2n,2n}$. The Cartan matrix is $$\begin{pmatrix}
2 & -1 & & &\\
-1 & 2 & -1 & & \\
& \ddots & \ddots & \ddots& \\
& &-1 & 2 & -2\\
& & & -1 & 2
\end{pmatrix}$$ The degrees of $B_n$ are $2,4,\cdots,2n$. The weight diagram for the first fundamental representation is $$\label{bn wt}
-L_1\os{-\a_1}\longleftarrow \cdots \os{-\a_{n-1}}\longleftarrow -L_n\os{-\a_n}\longleftarrow 0\os{-\a_n}\longleftarrow L_n\os{-\a_{n-1}}\longleftarrow \cdots\os{-\a_2}\longleftarrow L_2\os{-\a_1}\longleftarrow L_1.$$ The $\bu$ in is $$\bu=\on{Diag}(u^1_x, u^2_x - u^1_x, \cdots, 2 u^n_x - u^{n-1}_x, -u^1_x, -u^2_x + u^1_x, \cdots, -2u^n_x + u^{n-1}_x, 0).$$
Consider the expansion $$\label{Bn det}
\begin{split}
&\quad (\p-u^1_x)(\p+u^1_x-u^2_x)\cdots(\p+u^{n-1}_x-2u^n_x)\\
&\quad\quad \p(\p+2u^{n}_x-u^{n-1}_x)
\cdots(\p+u^2_x-u^1_x)(\p+u^1_x)\\
&=\p^{2n+1}+\sum_{j=1}^n I_j\p^{2n-2j+1}+\sum_{j=1}^n J_j\p^{2n-2j}.
\end{split}$$ Then the $I_j$ for $1\leq j\leq n$ are primitive characteristic integrals of the $B_n$ Toda field theory, and the $J_j$ are some differential polynomials in them.
The Cartan matrix is $$\begin{pmatrix}
2 & -1\\
-3 & 2
\end{pmatrix}$$ For simplicity we call the unknown functions $u^1, u^2$ by $u$ and $v$. The Toda system is $$\begin{cases}
u_{xy}=-e^{2u-v}\\
v_{xy}=-e^{-3u+2v}
\end{cases}$$
Although we can work abstractly with just the above Cartan matrix, we choose to use the embedding of $\fg_2\subset \fs\fo_7=B_3$ such that the two roots are $$\a_1=L_1-L_2,\quad \a_2:=L_2-L_3,$$ and $$\begin{aligned}
H_{\alpha_1}&=\on{Diag}(1,-1,2,-1,1,-2,0)\\
H_{\alpha_2}&=\on{Diag}(0,1,-1,0,-1,1,0). \end{aligned}$$ The $\bu$ in is $${\bf u}=
\on{Diag}(u_x, -u_x + v_x, 2u_x-v_x, -u_x, u_x - v_x, -2u_x+v_x, 0)
.$$
The first fundamental representation of $\fg_2$ is the restriction of that of $B_3$. Therefore we follow the weight diagram . The degrees of $\fg_2$ are 2 and 6.
Consider the expansion $$\begin{split}
&\quad(\p-u_x)(\p+u_x-v_x)(\p-2u_x+v_x)\\
&\quad\quad \p(\p+2u_x-v_x)(\p-u_x+v_x)(\p+u_x)\\
&=\p^{7}+I_1\p^5+I_2\p+\sum_{j=1}^{3} J_j\p^{5-j}+J_4.
\end{split}$$ Then the $I_1$ and $I_2$ are primitive characteristic integrals of the $\fg_2$ Toda field theory, and the $J_j$ are some differential polynomials in them.
We actually list the results easily computed by a computer. Also for simplicity, we use the notation $u_1=u_x,\ u_2=u_{xx},\ u_3=u_{xxx}$ and so on. The results are [$$\begin{aligned}
I_1&=6\,u_2+2\,v_2-6\,{u_1}^2+6\,u_1v_1-2\,{v_1}^2,\\
I_2&=5\,u_{{6}}+v_{{6}}+98\,u_{{2}}v_{{2}}v_{{1}}u_{{1}}-v_{{4
}}{u_{{1}}}^{2}-v_{{4}}{v_{{1}}}^{2}-{v_{{1}}}^{4}{u_{{1}}}^
{2}+21\,v_{{4}}u_{{2}}+30\,v_{{3}}u_{{3}}\\&
+3\,v_{{1}}u_{{5}
}+5\,v_{{5}}u_{{1}}-10u_{{5}}u_{{1}}-17u_{{4}}u_{{2}}
+
19\,v_{{2}}u_{{4}}-2\,v_{{5}}v_{{1}}-23\,u_{{4}}{u_{{1}}}^
{2}\\&
-7\,v_{{4}}v_{{2}}-7\,{v_{{1}}}^{2}u_{{4}}+46\,u_{{3}}{u_
{{1}}}^{3}-23\,v_{{3}}{u_{{1}}}^{3}-3\,{v_{{1}}}^{3}u_{{3}}+
2\,v_{{3}}{v_{{1}}}^{3}+28\,{v_{{2}}}^{2}{u_{{1}}}^{2}\\&
+114\,{u
_{{2}}}^{2}{u_{{1}}}^{2}+6\,{v_{{2}}}^{2}{v_{{1}}}^{2}+2\,{v_{
{1}}}^{4}u_{{2}}-13\,{v_{{1}}}^{2}{u_{{1}}}^{4}-10\,{v_{{2}}
}^{2}u_{{2}}+46\,v_{{2}}{u_{{2}}}^{2}+17\,{v_{{1}}}^{2}{u_{{2
}}}^{2}\\&
-2\,v_{{2}}{u_{{1}}}^{4}+12\,u_{{2}}{u_{{1}}}^{4}
+12\,v
_{{1}}{u_{{1}}}^{5}+6\,{v_{{1}}}^{3}{u_{{1}}}^{3}+27\,u_{{4}
}v_{{1}}u_{{1}}+63\,u_{{3}}v_{{1}}u_{{2}}\\&
-21\,v_{{2}}v_{{1
}}u_{{3}}-90\,{u_{{2}}}^{2}v_{{1}}u_{{1}}+29\,v_{{3}}v_{{1
}}{u_{{1}}}^{2}-11\,v_{{3}}{v_{{1}}}^{2}u_{{1}}
-22\,{v_{{2
}}}^{2}v_{{1}}u_{{1}}-126\,u_{{3}}u_{{2}}u_{{1}}\\&
-6\,v_{{3}
}v_{{2}}v_{{1}}+16\,v_{{2}}v_{{1}}{u_{{1}}}^{3}-v_{{3}}v_{
{2}}u_{{1}}+23\,v_{{3}}u_{{2}}u_{{1}}-18\,u_{{2}}{v_{{1}
}}^{3}u_{{1}}-122\,u_{{2}}v_{{2}}{u_{{1}}}^{2}\\&
+21\,u_{{3}}{v
_{{1}}}^{2}u_{{1}}-16\,v_{{2}}{v_{{1}}}^{2}{u_{{1}}}^{2}+42
\,u_{{3}}v_{{2}}u_{{1}}-57\,u_{{3}}v_{{1}}{u_{{1}}}^{2}+50
\,u_{{2}}{v_{{1}}}^{2}{u_{{1}}}^{2}-48\,u_{{2}}v_{{1}}{u_{{1
}}}^{3}\\&+4\,v_{{2}}{v_{{1}}}^{3}u_{{1}}
-22\,v_{{2}}{v_{{1}}
}^{2}u_{{2}}-v_{{4}}v_{{1}}u_{{1}}-12\,v_{{3}}v_{{1}}u_{{2
}}-10\,{u_{{3}}}^{2}-5\,{v_{{3}}}^{2}-42\,{u_{{2}}}^{3}\\&
-2\,{v_
{{2}}}^{3}-4\,{u_{{1}}}^{6}\end{aligned}$$ ]{} The other terms are $$\begin{aligned}
J_1&=\tfrac{5}{2}\,I_{1,x}\\
J_2&=3\,I_{1,xx}+\tfrac{1}{4}\,{I_1}^2\\
J_3&=2\, I_{1, xxx}+\tfrac{3}{4}\,I_1\cdot I_{1, x}\\
J_4&=\tfrac{1}{2}\,I_{2,x}-\tfrac{1}{4}\,I_{1,xxxxx}-\tfrac{3}{8}\,I_{1,x}\cdot I_{1,xx} -\tfrac{1}{8}\,I_1\cdot I_{1,xxx}\end{aligned}$$
An example of Theorem \[ds gauge\]
==================================
It can be said that Theorem \[quick\] is a quick way to compute a Drinfeld-Sokolov gauge when there is a non-branching representation, and this is the viewpoint in [@W-sym]. When such a simple representation does not exist, a Drinfeld-Sokolov gauge can still be computed through an inductive procedure (see [@DS; @gen-W-sym]). In the particular case of $D_n$, where an explicit matrix presentation of the Lie algebra is easy to write down, one can use a computer algebra system like Maple to solve a Drinfeld-Sokolov gauge easily and therefore obtain the characteristic integrals.
In the following example, we will show how this works for $D_4$. First recall that the Cartan matrix of $D_n$ is $$\begin{pmatrix}
2 & -1 & & & & \\
-1 & 2 & -1 & & & \\
& \ddots & \ddots & \ddots& & \\
& &-1 & 2 & -1 & -1\\
& & & -1 & 2 & \\
& & & -1 & & 2\\
\end{pmatrix},$$ and its first fundamental representation branches with the following weight diagram $$\label{splits}
\xymatrix{
& & & & L_n\ar[ld]_{-\a_n} & & & & \\
-L_1&-L_2\ar[l]_{-\a_1} &\cdots\ar[l]_{-\a_2} & -L_{n-1}\ar[l]_{-\a_{n-2}} & & L_{n-1}\ar[lu]_{-\a_{n-1}}\ar[ld]^{-\a_n} & \cdots\ar[l]_{-\a_{n-2}} & L_2\ar[l]_{-\a_2} & L_1\ar[l]_{-\a_1}\\
& & & & -L_n\ar[lu]^{-\a_{n-1}} & & & &
}$$ Therefore Theorem \[quick\] can not be applied to the Lie algebras $D_n$, nor to $F_4$ or the $E$’s.
For simplicity, we write $u, v, w, z$ for the unknown functions $u^1,\cdots,u^4$. The Toda field theory for $D_4$ is $$\begin{cases}
u_{xy}=-e^{2u-v}\\
v_{xy}=-e^{-u+2v-w-z}\\
w_{xy}=-e^{-v+2w}\\
z_{xy}=-e^{-v+2z}.
\end{cases}$$
The degrees of $D_4$ are $2, 4, 4, 6$, with the Pfaffian, the square root of the determinant, being an extra adjoint-invariant function of degree 4. In this example, we use the convention that the non-degenerate symmetric matrix preserved by the orthogonal matrices in $SO(8)$ is the anti-diagonal matrix $(\delta_{i, 9-j})_{i,j=1}^8$. Then the matrices in $D_4=\fs\fo_8$ are skew-symmetric with respect to the anti-diagonal. Using $u_1=u_x$ and so on, the $\e + \bu$ from is $${\small
\e+{\bf u}=
\begin{pmatrix}
u_1 & & & & & & & \\
1 & v_1-u_1 & & & & & & \\
& 1 & z_1+w_1-v_1 & & & & & \\
& & 1 & z_1-w_1 & & & & \\
& & 1 & & -z_1+w_1 & & & \\
& & & -1 & -1 & -z_1-w_1+v_1 & & \\
& & & & & -1 & -v_1+u_1& \\
& & & & & & -1 & -u_1
\end{pmatrix}}$$ We choose our slice in to be $${\small
\e+{\bf I}=
\begin{pmatrix}
0 & & & & I_2 & & I_4 & \\
1 & 0 & & & & I_3 & & -I_4 \\
& 1 & 0 & & I_1 & & -I_3 & \\
& & 1 & 0 & & -I_1 & & -I_2 \\
& & 1 & & 0 & & & \\
& & & -1 & -1 &0 & & \\
& & & & & -1 & 0& \\
& & & & & & -1 & 0
\end{pmatrix}}$$ A unipotent orthogonal matrix $B\in N_{SO(8)}$ is an upper-triangular matrix with $1$’s on the diagonal and element $b_{i,j}$ at the position $(i,j)$ for $i<j$ satisfying some extra conditions for the orthogonality. In particular we have that $b_{4,5}=0$ and $b_{3,4}+b_{5,6}=0$. Solving by Maple, we get the following expression for the primitive characteristic integrals of the $D_4$ Toda field theory. [$$\begin{aligned}
I_1&=-(u_{{2}}+v_{{2}}+z_{{2}}+w_{{2}}-{u_{{1}}}^{2}-{v_{{1}}}^{2}-{z_{{1}}}^{2}-{w_{{1}}}^{2}+u_{{1}}v_{{1}}+z_{{1}}v_{{1}}+
w_{{1}}v_{{1}})\\
I_2&=u_{{4}}+v_{{4}}+2\,w_{{4}}+z_{{2}}{v_{{1}}}^{2}+3\,w_{{2}}
v_{{2}}-w_{{2}}{v_{{1}}}^{2}+{w_{{1}}}^{2}v_{{2}}+v_{{2}}z
_{{2}}\\&
-v_{{2}}{z_{{1}}}^{2}+2\,v_{{2}}u_{{2}}-z_{{2}}u_{{2
}}-w_{{2}}{u_{{1}}}^{2}+z_{{2}}{u_{{1}}}^{2}+v_{{3}}u_{{1
}}\\&
-2\,u_{{3}}u_{{1}}+2\,v_{{3}}w_{{1}}-{w_{{1}}}^{2}u_{{2}
}+w_{{2}}u_{{2}}-4\,{w_{{2}}}^{2}-2\,{v_{{2}}}^{2}\\&
-2\,{u_{{2}}}^{
2}+{z_{{1}}}^{2}u_{
{2}}+w_{{2}}v_{{1}}u_{{1}}-z_{{1}}v_{{1}}u_{{2}}+w_{{1
}}v_{{1}}u_{{2}}\\&
-z_{{2}}v_{{1}}u_{{1}}-2\,v_{{1}}w_{{1}}
v_{{2}}+2\,w_{{2}}v_{{1}}w_{{1}}+2\,z_{{1}}v_{{1}}v_{{2}
}-2\,v_{{1}}z_{{1}}z_{{2}}\\&
+u_{{3}}v_{{1}}-2\,v_{{1}}v_{{3
}}+v_{{1}}z_{{3}}-4\,w_{{1}}w_{{3}}+v_{{1}}w_{{3}}-{z_{{1
}}}^{2}{u_{{1}}}^{2}+{w_{{1}}}^{2}{u_{{1}}}^{2}\\&
-{w_{{1}}}^{2}v
_{{1}}u_{{1}}-z_{{1}}{v_{{1}}}^{2}u_{{1}}-w_{{1}}v_{{1}}
{u_{{1}}}^{2}+{z_{{1}}}^{2}v_{{1}}u_{{1}}+z_{{1}}v_{{1
}}{u_{{1}}}^{2}+w_{{1}}{v_{{1}}}^{2}u_{{1}}\\
I_3 &= 2\,u_{{4}}+v_{{4}}+w_{{4}}+{w_{{1}}}^{2}{u_{{1}}}^{2}-2\,v_{{1}}v_{{
3}}+v_{{1}}z_{{3}}-2\,w_{{1}}w_{{3}}+v_{{1}}w_{{3}}\\&
+u_{{
3}}v_{{1}}+3\,v_{{2}}u_{{2}}-w_{{2}}{u_{{1}}}^{2}+w_{{2}
}u_{{2}}+2\,v_{{3}}u_{{1}}-4\,u_{{3}}u_{{1}}+v_{{3}}w_{{1
}}\\&
-{w_{{1}}}^{2}u_{{2}}+v_{{2}}z_{{2}}-v_{{2}}{z_{{1}}}^{2
}+z_{{2}}{v_{{1}}}^{2}+2\,w_{{2}}v_{{2}}-2\,{v_{{2
}}}^{2}-4\,{u_{{2}}}^{2}\\&
-2\,{w_{{2}}}^{2}-{w_{{1}}}^{2}v_{{1}}
u_{{1}}+w_{{1}}{v_{{1}}}^{2}u_{{1}}+2\,z_{{1}}v_{{1}}v_{{2
}}-2\,v_{{1}}z_{{1}}z_{{2}}+w_{{1}}v_{{1}}u_{{2}}\\&
+w_{{2
}}v_{{1}}u_{{1}}
-w_{{1}}v_{{1}}{u_{{1}}}^{2}-{v_{{1}}}^{2}
u_{{2}}+{z_{{1}}}^{2}w_{{2}}-{z_{{1}}}^{2}{w_{{1}}}^{2}+v_{{
2}}{u_{{1}}}^{2}-z_{{2}}w_{{2}}\\&
+z_{{2}}{w_{{1}}}^{2}
+{z_{{
1}}}^{2}v_{{1}}w_{{1}}-{v_{{1}}}^{2}z_{{1}}w_{{1}}+{w_{{1
}}}^{2}v_{{1}}z_{{1}}-z_{{2}}v_{{1}}w_{{1}}-w_{{2}}v_{{1
}}z_{{1}}\\&
+2\,v_{{1}}u_{{2}}u_{{1}}-2\,v_{{2}}v_{{1}}u_{{1
}}\\
I_4 &= -u_6 -\tfrac{1}{2}\, v_6 -w_6 +\text{a lot of other terms which we omit}\end{aligned}$$ ]{} Actually the usual Pfaffian of $\e+{\bf u}$, which is the product of the first 4 diagonal entries, is contained in the characteristic integral $I_2$ as the last terms involving only partial derivatives of the first order.
In [@W-sym], there is a procedure to apply an integration step in using the first fundamental representation of $D_n$ to get an analogous formula to Theorem \[quick\]. That integration step will cause us to lose one characteristic integral, which in the case of $D_4$ is exactly the above $I_2$ corresponding to the Pfaffian.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We used VLT/XShooter to target a sample of nearby analogs of Lyman Break Galaxies (LBGs). These Lyman Break Analogs (LBAs) are similar to the LBGs in many of their physical properties. We determine electron temperatures using the weak [\[O[iii]{}\]$\lambda$4363]{} emission line, and determine the oxygen abundance (O/H) using the direct and strong line methods. We show that the direct and strong line abundances are consistent with established relations within $\sim$0.2 dex. The analogs have nitrogen-to-oxygen ratios (N/O) and ionization parameters ($q$) that are, on average, offset with respect to typical local galaxies but similar to galaxies at $z\sim2$ and other analogs. The N/O and $q$ excesses correlate with the offsets observed in the strong line ratios, again similar to $z\sim2$. The star formation rate surface densities are consistent with the high electron density and ionization, indicating that the interstellar medium (ISM) pressure is set by feedback from the starbursts. For a given O/H, the apparent N/O excess arises due to the offset in O/H with respect to the local mass-metallicity relation. This can be explained by recent inflow of relatively metal-poor gas which lowers O/H while leaving N/O unchanged. The difficulties in determining even basic ISM parameters in these nearby analogs illustrates some of the challenges we face at much higher redshifts, where similar rest-frame optical diagnostics for large samples of galaxies can be accessed with JWST.'
author:
- 'Maryory Loaiza-Agudelo'
- 'Roderik A. Overzier'
- 'Timothy M. Heckman'
title: 'VLT/XShooter spectroscopy of Lyman Break Analogs: Direct method O/H abundances and nitrogen enhancements'
---
Introduction
============
Many of today’s most actively pursued open questions in galaxy evolution focus on the period between cosmic ‘dawn’ and cosmic ‘noon’. A complete picture of how galaxy evolution proceeded within this period must include answers as to how galaxies received and recycled their gas, formed their stellar populations, and how the ionizing radiation from the first galaxies reionized the universe. While new observatories have begun to address these questions by directly targeting the most distant galaxies, a complementary path is to study similar processes occurring in the more nearby universe using so-called ‘analogs’ in the hope of gaining insight into the processes occurring at less accessible wavelengths, resolutions or redshifts. One such class of objects are the Lyman Break Analogs (LBAs) first studied by @heckman05 and @hoopes07. The LBA project was designed to find and study relatively nearby starburst galaxies that share typical characteristics of Lyman Break Galaxies (LBGs) at high redshift. A crossmatch of the Sloan Digital Sky Survey (SDSS) with the GALEX UV imaging survey [@martin05] was used to select luminous ($L_{FUV}>10^{10.3}$ L$_\odot$) and compact ($I_{FUV} > 10^9$ L$_\odot$ kpc$^{-2}$) star-forming galaxies at $z<0.3$ having similar rest-frame far-UV properties as typical (i.e. $L_{FUV}\simeq L^*_{z=3}$) LBGs. These simple criteria select galaxies with relatively high SFRs that are furthermore relatively compact and experiencing modest extinction by dust similar to the high redshift LBGs. The sample of LBAs from SDSS-GALEX was significant because previous surveys lacked the depth and coverage in the UV to find significant numbers of such galaxies, making comparisons between local and distant galaxies less direct [e.g. @meurer99].
Several useful samples of local analogs have been constructed in recent years. Examples include galaxies having high equivalent width [Ly$\alpha$]{} selected to be good local analogs of [Ly$\alpha$]{} emitters (LAEs) at high redshift [e.g., the [Ly$\alpha$]{} Reference Sample (LARS); @ostlin14], high equivalent width [H$\alpha$]{} sources selected to be local analogs of [H$\alpha$]{} emitters (HAEs) at $z\sim4$ [@shim13], the Green Pea (GP) galaxies selected on the basis of strong optical emission lines that also span the properties of LAEs, HAEs and LBGs at high redshift [@cardamone09; @lofthouse17], high [\[O[iii]{}\]$\lambda$5007]{}/[\[O[ii]{}\]$\lambda$3727]{} ratio sources that are good candidate Lyman continuum leakers [e.g. @nakajima14; @schaerer16], and galaxies selected based on their offsets in the BPT diagram designed to match higher redshift samples [e.g. @bian16; @cowie16]. It is important to mention that there is typically significant overlap between all these samples and the UV-selected LBAs studied here [e.g. @cardamone09; @amorin12; @ostlin14].
The sample of LBAs share numerous other physical characteristics of star-forming galaxies at high redshift. The sizes, morphologies and gas kinematics of LBAs have been compared in detail with those of star-forming galaxies at $z>2$, finding general good agreement in the main parameter distributions [@overzier08; @overzier10; @goncalves10]. Although the triggers of star formation are likely to be different for the low and high redshift starbursts, at least the distributions of star formation, dust and gas appear to be comparable. The star formation in LBAs is dominated by luminous clumpy emission [@overzier08; @overzier09; @overzier10] that resembles that seen in the clumpy galaxies at intermediate redshifts [e.g. @elmegreen13; @garland15]. Some LBAs have luminous unresolved super-star clusters that are structurally similar to those seen in (lensed) sources at $z\gtrsim6$ [e.g. @bradley12; @bouwens17]. Although a small subset of LBAs also appear to host low-luminosity, obscured AGN at their SNe-feedback dominated centers [@jia11; @alexandroff12], the radio and X-ray luminosities associated with these AGN are well below the current detection threshold in high redshift star-forming galaxies [e.g. @habouzit17; @latif18]. These structural and energetic similarities further indicate that LBAs may be good local laboratories for probing the extreme physical conditions expected to govern the interstellar medium of young galaxies at high redshifts.
LBAs were important for testing the so-called $\beta$–IRX relation for dust-correcting UV-based star formation rate measurements of star-forming galaxies [@meurer99; @salim18]. Because LBAs are more similar to LBGs compared to typical local starburst galaxies that are often either highly obscured or have SFRs orders of magnitudes lower compared to their high redshift counterparts, dust-correction methods that are based on this sample should be less biased than those based on other types of local populations [@overzier11; @bouwens12]. For similar reasons, LBAs can be used to probe the conditions responsible for the emission of the far-infrared fine structure lines, molecular gas and dust continuum of ordinary galaxies on the main sequence at high redshifts [e.g. @overzier11; @goncalves14; @contursi17; @wu19].
![\[fig:sample\]Sample properties. Far-UV luminosity versus surface brightness for all sources in the SDSS-GALEX crossmatched catalog (grey shaded region). Large symbols indicate objects for which VLT/XShooter spectroscopy was obtained. Blue squares: LBAs from @overzier09. Magenta circles: new LBAs selected on the basis of a large offset in the BPT-diagram (9 objects) or a high HST/COS far-UV flux (1 object). Red pentagons: Objects selected on the basis of their relatively low \[S[ii]{}\]/H$\alpha$-ratios. See text and Table \[tab:sample\] for details. ](LUV-SB_new.png){width="\columnwidth"}
![\[fig:bpt\]BPT diagram of all sources in the SDSS-GALEX crossmatched catalog (grey shaded region) compared to LBAs. Small yellow squares indicate LBAs from the extended LBA catalog. Large symbols indicate objects with deep optical spectra discussed in this paper. Large blue squares are LBAs from @overzier09. Measurements for 7 objects with spectroscopic data from the literature are indicated with green pentagons (B14: @brown14; A12: @amorin12). Large magenta circles are new LBAs from the extended sample and selected for VLT/XShooter spectroscopy on the basis of a large offset in the BPT-diagram (9 objects) or a high HST/COS far-UV flux (1 object). The large red pentagons are objects selected on the basis of their relatively low \[S[ii]{}\]/H$\alpha$-ratios (3 objects). See text and Table \[tab:sample\] for details.](BPT_XSHOOTER-LBAs.png){width="\columnwidth"}
Far-UV spectroscopic observations with the Cosmic Origins Spectograph (COS) on the Hubble Space Telescope (HST) revealed that LBAs typically have strong outflows with some extreme cases up to $\sim1000-2000$ km s$^{-1}$ [@heckman11; @heckman15; @heckman16; @borthakur14]. These outflows correlate strongly with the SFR per unit area [@heckman15], as expected if the outflows are mainly driven by the momentum flux in compact starbursts. The LBA program also delivered a strong case of detection of Lyman continuum (LyC) photon escape [@borthakur14], showing for the first time a strong connection between the SNe-driven outflows in compact starbursts and LyC escape. @overzier09 showed that the extinction-corrected [H$\alpha$]{} was relatively weak compared to their far-UV and far-IR emission in some of the most compact LBAs, suggesting that ionizing radiation may be escaping (among several other explanations). @heckman11 noted that besides the strong outflows and relative weakness of [H$\alpha$]{}, some LBAs have significant residual intensities in the cores of the saturated interstellar low-ionization absorption lines tracing the neutral gas, and that these sources also tend to show significant blue-shifted [Ly$\alpha$]{} in emission. All these properties can be explained by a simple model in which compact starbursts drive powerful winds that remove the neutral gas along certain lines of sight, allowing LyC and [Ly$\alpha$]{} photons to escape. In a follow-up study, @alexandroff15 further found that objects either confirmed [@borthakur14] or suspected [@heckman11] of being LyC leakers based on the above-mentioned indicators, also have relatively weak [\[S[ii]{}\]$\lambda\lambda$6717,6731]{} emission line doublets, indicating the presence of matter-bounded [H[ii]{}]{} regions, while other diagnostics proposed to be indicators for LyC escape, such as relatively weak dust-corrected [H$\alpha$]{} or the [\[O[iii]{}\]$\lambda$5007]{}/[\[O[ii]{}\]$\lambda$3727]{} line flux ratio [e.g. @nakajima14], did not correlate with the other tracers. The indirect indicators of LyC escape based on the study of LBAs have thus offered a number of empirical probes that may be used as proxies for the escape fractions of galaxies during the EoR for which LyC emission cannot be measured directly [@borthakur14; @alexandroff15].
Analysis of optical spectra from SDSS has shown that LBAs lie, on average, below the local stellar mass-metallicity relation [@hoopes07; @overzier10], and above the star-forming sequence in the ‘BPT’ diagram [@overzier09; @bian16; @kojima17; @patricio18 and this paper], again two important features that are analogous to those of typical star-forming galaxies at high redshift [e.g. @erb06; @steidel14]. Given that both LBAs and LBGs are presumed to be galaxies undergoing a phase of rapid build-up of their stellar populations from recent influx of relatively metal-poor gas, LBAs could perhaps also aid in answering a number of open questions related to the chemical enrichment history of early galaxies. However, a closely related problem that has played a central role in recent years is that star-forming galaxies at high redshift occupy different locations in the main optical diagnostic emission line diagrams [e.g. @liu08; @overzier09; @steidel14; @bian16; @bian17; @bian18]. Specifically, @steidel14 showed that UV-selected galaxies at $z\sim2$ lie along a locus that is offset with respect to that of local star-forming galaxies in the [\[O[iii]{}\]$\lambda$5007]{}/[H$\beta$]{} versus [\[N[ii]{}\]$\lambda$6584]{}/[H$\alpha$]{} “BPT” diagram [@baldwin81]. Given that the conditions in the interstellar medium in typical star-forming galaxies at local and high redshifts are likely very different, this is perhaps not so surprising. Possible explanations include contributions from AGN photoionization, shocks, different N/O ratios, higher ionization parameters and harder radiation fields.
Understanding the BPT offsets and possible evolution is important, because they involve the same optical emission line ratios that are being employed to determine the nebular gas abundances of star-forming galaxies as well as to identify AGN [e.g. @kewley13; @hirschmann17]. This is especially important at high redshifts where the strong optical emission lines are often the only viable way of determining these ISM abundances. Proper local analogs can therefore aid in these studies as well. Strong line metallicities have been compared against direct method values in gravitationally-lensed galaxies with oxygen auroral lines detected out to $z\sim2.5$, showing that locally calibrated methods are reliable at high redshift within a dispersion of $\sim$0.2 dex [e.g. @patricio18; @gburek19]. @bian16 selected a new set of local analogs from SDSS solely on the basis of their proximity to the locus of $z\sim2$ galaxies from @steidel14 in the BPT diagram, showing that these objects have relatively high ionization parameters and electron densities that can only partly be explained by their increased (specific) star formation rates compared to typical star-forming galaxies. @bian17 show that at low stellar masses these analogs lie $\sim0.2$ dex below the local $M_*-Z$ relation (MZR) when using metallicities based on the N2 or O3N2 diagnostics [@pettini04], similar to samples at $z\sim2$ [e.g. @steidel14; @sanders15]. @bian18 stacked the SDSS spectra of their analogs sample to measure direct oxygen abundances, finding that local strong line calibrations underestimate the abundances by $\lesssim0.1$ dex, and providing updated calibrations. Various authors have found that the offsets are often related to an increased N/O abundance at high redshift compared to local galaxies at a fixed oxygen abundance [e.g. @masters14; @shapley15; @sanders16a; @kojima17]. This could be an effect of enhanced nitrogen production by WR stars, or, more likely, the consequence of rapid accretion of low metallicity gas which reduces O/H while leaving N/O unchanged [@koppen05; @amorin10; @amorin12; @masters16]. The N/O–O/H relation appears constant with redshift, and harder ionizing spectra may be required at high redshift to fully explain the BPT offsets [@steidel16; @strom17; @strom18; @shivaei18]. These harder spectra could be a result of the differences in star formation histories, and thus the chemical enrichment, between typical local and high redshift galaxies.
@andrews13 used large stacks of star-forming galaxies from the SDSS along the star-forming main sequence to compare the direct-method oxygen abundances with several calibrations derived from strong optical lines. They show that the MZR has a clear SFR-dependence, and that N/O correlates with O/H, star formation history and stellar mass. @brown16 improved local strong-line methods for use at high redshift by quantifying the dependence on specific star formation history. @brown14 tested the direct method using deep follow-up spectra obtained for four LBAs from the sample of @overzier09, finding that the strong line method abundances are in agreement with those from the temperature sensitive method. @amorin12 confirmed the low oxygen abundances but remarkably high nitrogen to oxygen ratios in three LBAs. @kojima17 show that local analogs and $z\sim2$ galaxies with BPT offsets can be explained by excesses in either N/O or ionization parameter, or combinations thereof, while their data did not allow them to test for changes in hardness of the radiation.
In this Paper, we analyze the oxygen and nitrogen abundances of LBAs determined through the direct and strong line methods, and investigate the cause of offsets in ionization, O/H and N/O in the context of typical star-forming galaxies at $z\sim2$. We use magnitudes in the AB system, a Chabrier initial mass function and the solar metallicity scale from @asplund09 where 12 + log(O/H)$_\odot=8.69$ and log(N/O)$_\odot=-0.86$. The cosmological parameters are set to $\Omega_m = 0.286$, $\Omega_\Lambda = 0.714$ and $H_0=69.6$ km s$^{-1}$ Mpc$^{-1}$.
Sample and observations
=======================
Sample selection
----------------
The targets studied as part of this paper were selected from a number of sources related to the LBA project, and by no means represent a complete sample. 17 targets (marked ‘LBA’ in Table \[tab:sample\]) were selected from the original sample of @heckman05 and studied in detail by @overzier09. Ten new LBAs (marked ‘LBA2’ in Table \[tab:sample\]) were taken from an extended sample of several hundred LBAs in the crossmatch between SDSS data release 7 [DR7; @abazajian09] and GALEX GR6. Of these ten targets, nine targets (marked ‘BPT’ in Table \[tab:sample\]) were further selected based on having large BPT offsets and observability, while one target (marked ‘COS’ in Table \[tab:sample\]) was selected because it has been observed with the HST/COS as part of a program to target LBAs having high far-UV fluxes within the COS aperture. Finally, three more targets (marked ‘S2-Deficit’ in Table \[tab:sample\]) were selected on the basis of their relative weakness of the \[SII\]6717,6731 optical emission-lines. This was motivated by previous work on LBAs that showed that objects with abnormally low \[SII\] 6717,6731/H$\alpha$ flux ratios compared to typical star-forming galaxies may have matter-bounded conditions in the ISM which can result in the escape of ionizing photons [e.g. @pellegrini12]. This has inspired a new category of local analogs that have potentially large escape fractions [@alexandroff15; @wang19]. All three ‘S2-Deficit’ objects satisfy the UV surface brightness criterion of LBAs, but one is of substantially lower FUV luminosity than LBAs, as defined by @heckman05. This object is a known low metallicity blue compact dwarf [e.g., it was part of Subsample 1 from @izotov07].
Originally part of the LBA sample, two objects (001054 and 005439) were found to be Type 1 AGN based on broad Mg II lines detected in the UVB-arm of VLT/XShooter. These two objects are considered to be contaminants of the LBA sample as the emission from the unobscured nucleus likely dominates the far-UV flux detected by GALEX. These two sources were removed from the sample. Object 210358 had problems with the observations, as the on-slit dither positions were chosen to be too close to eachother leading to problems with the background subtraction. This object is not studied further as part of this paper.
Several objects from the @overzier09 sample have been observed with deep spectroscopic data before. @amorin12 observed three of the sources with the GTC-OSIRIS spectrograph (004054, 113303, 232539), while @brown14 observed four sources with the LBT-MODS1 spectrograph (004054, 005527, 020356, 092600), one of which (004054) is in common with @amorin12. These six unique sources include two LBAs not covered by our spectroscopy (092600 and 113303), and four that overlap with our data. Although these authors have already derived many of the parameters similar to the ones that we will derive here (i.e. stellar populations, densities, temperatures and abundances), we will use the stellar absorption and dust-corrected line fluxes as reported by these authors in order to compare their results with our own data. Fig. \[fig:sample\] shows the rest-frame UV properties of the VLT/XShooter sample compared to the local galaxy population. The location of the objects in the BPT diagram as measured from the VLT/XShooter spectra or given by the literature data is shown in Fig. \[fig:bpt\].
{width="33.00000%"} {width="33.00000%"} {width="33.00000%"}\
{width="33.00000%"} {width="33.00000%"} {width="33.00000%"}\
{width="33.00000%"} {width="33.00000%"} {width="33.00000%"}\
{width="33.00000%"} {width="33.00000%"} {width="33.00000%"}\
{width="33.00000%"} {width="33.00000%"} {width="33.00000%"}\
{width="33.00000%"} {width="33.00000%"} {width="33.00000%"}\
{width="33.00000%"} {width="33.00000%"} {width="33.00000%"}\
{width="33.00000%"} {width="33.00000%"} {width="33.00000%"}\
{width="33.00000%"} {width="33.00000%"}\
Observations and data reduction
-------------------------------
Spectra were taken using the X-Shooter spectrograph installed at UT2 (Kueyen) of the ESO Very Large Telescope (VLT) in Chile during two separate observing campaings. The first sample of 17 objects was observed in service mode during 2010 (Program ID 085.B-0784(A)), and a second sample of 13 objects was observed in visitor mode during March 2016 (Program ID 096.B-0192(A)). The target list and log of observations is given in Table \[tab:sample\]. We used X-Shooter in slit-mode (11 slit) to obtain simultaneous spectra from the $U$- to $K$-band using the UVB ($1\farcs0$ wide slit, $R\approx5100$), VIS ($0\farcs9$ wide slit, $R\approx8800$) and NIR ($0\farcs9$, $R\approx5100$) arms. The exposure times for each object were $4\times690$ s, $8\times320$ s and $12\times240$ s in nodding-on-slit mode, except for very extended objects that were observed in offset mode and received half the exposure time. The slits were centered on the brightest starburst regions but we note that the effective radii of most objects are comparable to the slit width. The seeing was estimated from the telluric standard star spectra taken each night at the same airmass as the science observations.
We used the ESO X-Shooter pipeline [@modigliani10] and the EsoRex command-line recipes to reduce the data. The pipeline removes non-linear pixels, substracts the bias in the UVB and VIS arms and dark frames in the NIR arm. It then predicts the positions of arc-lines and order-edges on a format-check frame, determines the order positions and the two-dimensional wavelength solution to resample the orders, divides the raw frames by a master flat field, measures the instrument resolving power, computes the instrument flexures, the instrument response and the total system efficiency. The spectra are then flux calibrated using standard stars that were observed during each night. The pipeline combines the science frames and subtracts the sky background. One-dimentional spectra were extracted using the X-Shooter pipeline at the location of the continuum along the slit. The size of the apertures was set to the larger value of the seeing (FWHM) measured from the telluric standard in the UVB arm (averaged over all wavelengths) and the spatial extent of [H$\alpha$]{} in the VIS arm. The spectra extracted from each of the arms were joined, and small flux offsets were applied to the UVB and NIR spectra in order to correct for small flux normalization mismatches with overlapping wavelength ranges at either end of the VIS spectra. Although telluric standards were observed and reduced in the same way, ultimately the telluric absorption corrections were done using the ESO *Molecfit* tool [@kausch15; @smette15]. Molecfit allows one to fit a model of the tropo- and strato-spheric telluric features directly to the science spectra and subtract this model from the data in order to remove the telluric features. Molecfit also updates the input error spectra that are produced by the ESO pipeline.
![\[fig:Ne\]Electron densities estimated from [\[O[ii]{}\]$\lambda\lambda$3726,3729]{} and [\[S[ii]{}\]$\lambda\lambda$6717,6731]{} using the model fit results obtained by @sanders16a for a 5-level atom aproximation and assuming an electron temperature of 10,000 K. Red circles are LBAs observed with XShooter, while the green pentagons are the literature LBAs for which only the [\[S[ii]{}\]$\lambda\lambda$6717,6731]{}-doublet ratio is known. The black dotted line shows the one-to-one relation.](Ne.png){width="\columnwidth"}
Measurement methods
-------------------
The reduced, calibrated spectra are first corrected for foreground extinction assuming the @cardelli89 extinction curve and reddening from @schlegel98. Emission line fluxes are determined by fitting single Gaussians in the case of isolated lines or series of Gaussians in the case of groups of lines. In order to correct the main lines for underlying stellar continuum absorption we use the STARLIGHT code version 04 [@cid-fernandes05; @cid-fernandes11; @mateus06; @asari07]. We mask all emission lines, as well as the locations of possible WR bumps (4600–4700 Å and 5700–5900 Å). We apply an additional mask around the locations where the UVB and VIS arm spectra were joined due to increased noise, and possible residual offsets that exist in this wavelength range. The spectra are interpolated on a regular grid of 1 Å bins. We run the code with 115 single stellar populations (SSPs) with a @chabrier03 initial mass function for 5 different metallicities up to solar metallicity and ages up to 13 Gyr with the Padova 1994 tracks from @bruzual03. For objects with strong nebular continuum, we include a nebular continuum template. We use the @calzetti01 extinction curve with a single reddening parameter that is applied equally to all SSPs with $A(V)$ in the range 0–4 mag. We fit the stellar kinematics allowing for a systematic shift of $\pm500$ km s$^{-1}$ and a stellar velocity dispersion of up to 500 km s$^{-1}$. In order to obtain the best fits to the stellar continuum in the vicinity of the weakest, temperature-sensitive lines ([\[O[iii]{}\]$\lambda$4363]{}, [\[N[ii]{}\]$\lambda$5755]{}, and [\[O[ii]{}\]$\lambda\lambda$7319,7330]{}), we follow @andrews13 and limit the spectral range of the STARLIGHT fit to 400Å around these lines (200Å in either direction), which produced a good match to the observed continuum in all cases. For higher metallicity sources, in which [\[O[iii]{}\]$\lambda$4363]{} is weak, there can be substantial contamination of [\[O[iii]{}\]$\lambda$4363]{} from \[Fe[ii]{}\]$\lambda$4352.78 and \[Fe[ii]{}\]$\lambda$4359.34 [e.g. @andrews13; @curti17]. We took this into account by including the possible presence of the \[FeII\] lines into our fits and fixing the widths of these weak lines to that of [H$\gamma$]{}. The line fluxes were measured after subtracting the best STARLIGHT model, and corrected for dust extinction based on the Balmer decrement and assuming a @cardelli89 extinction curve. For [\[O[ii]{}\]$\lambda\lambda$7319,7330]{} we were able to estimate the fraction of recombination excitation that contributes to the auroral line using Equation 2 from @liu00 that uses the temperature, the [H$\beta$]{} line flux and the $\mathrm{O}^{2+}/\mathrm{H}^+$ ratio. In all cases in which [\[O[iii]{}\]$\lambda$4363]{} was detected (thus allowing a direct estimate of the temperature and $\mathrm{O}^{2+}/\mathrm{H}^+$), the recombination excitation fraction of the measured [\[O[ii]{}\]$\lambda\lambda$7319,7330]{} line flux was at the 2–4% level. We did not correct for this small contribution. The errors on the emission line fluxes were obtained by performing a monte carlo simulation of the line fitting process using the error spectrum. For the analysis involving the weak [\[O[iii]{}\]$\lambda$4363]{}, [\[N[ii]{}\]$\lambda$5755]{}, and [\[O[ii]{}\]$\lambda\lambda$7319,7330]{} lines we required a detection of at least 5$\sigma$. The continuum-subtracted spectra and the best-fit results in the spectral region around [\[O[iii]{}\]$\lambda$4363]{}, [\[N[ii]{}\]$\lambda$5755]{} and [\[O[ii]{}\]$\lambda\lambda$7319,7330]{} are shown in Fig. \[fig:4363\]. The measurements used and derived in this paper are summarized in Tables \[tab:fluxes\], \[tab:basic\_pars\], and \[tab:Z\_pars\].
![\[fig:R4363\]Electron temperature $T_e$(O[iii]{}) based on RO3 and the calibration of @proxauf14. Literature LBAs are marked with green pentagons. The median value of $T_e$(O[iii]{})$\approx$11,000 K is indicated by the dashed line.](R4363.png){width="\columnwidth"}
{width="33.00000%" height="30.00000%"}{width="33.00000%" height="30.00000%"}{width="33.00000%" height="30.00000%"}
![\[fig:ZOH\_bothmethods\]The difference in direct method oxygen abundances determined using our two temperatures $T_e$(O[ii]{}) (based on RO2) and $T_e$(O[iii]{}) (based on RO3), $12 + \mathrm{log(O/H)_{obs}}$, and that determined using $T_e$(O[ii]{}) inferred from $T_e$(O[iii]{}) and the electron density using Equation 14 from @perez-montero17, as function of the two-temperature value ($12 + \mathrm{log(O/H)_{obs}}$). The black dotted line indicates the zero offset relation.](ZOH_ZTO2_ZTO2_from_TO3.png){width="\columnwidth"}
![\[fig:ZN2\_vs\_ZO3N2\]Strong line method oxygen abundances determined through the N2 and O3N2 methods of @pettini04. Several best-fit relations from the literature are indicated for the range in O/H for which they were determined [@steidel14; @sanders15; @bian18]. The black dotted line indicates the one-to-one relation. Literature LBAs are indicated by the green pentagons.](ZN2_vs_ZO3N2_PP04.png){width="\columnwidth"}
Results
=======
Electron density and temperature {#sec:Ne}
--------------------------------
The electron density ($n_e$) plays an important role in abundance measurements, since this parameter affects the fluxes of collisionally excited lines in star-forming regions. We determine $n_e$ based on the sulfur [\[S[ii]{}\]$\lambda\lambda$6717,6731]{} and [\[O[ii]{}\]$\lambda\lambda$3726,3729]{} line ratios following @sanders16a. The results are shown in Fig. \[fig:Ne\] and show good agreement between the two values. However, given the wider separation of the [\[S[ii]{}\]$\lambda\lambda$6717,6731]{} doublet, in this paper we adopt the electron density based on [\[S[ii]{}\]$\lambda\lambda$6717,6731]{}. For two sources (040208 and 232539), the [\[S[ii]{}\]$\lambda\lambda$6717,6731]{} ratio produced an unphysical density and in the analysis below we used their [\[O[ii]{}\]$\lambda\lambda$3726,3729]{}-based densities instead. For the LBAs with literature spectroscopic data, we were only able to calculate densities on the basis of [\[S[ii]{}\]$\lambda\lambda$6717,6731]{}, showing that they are in the same range as the other objects (green triangles drawn as lower limits on $N_e$(O[ii]{}) in Fig. \[fig:Ne\]).
We measure electron temperatures, $T_e$(O[iii]{}), through the $RO3\equiv(I(4959)+I(5007))/I(4363)$ ratio that probes conditions in the innermost high-excitation zone in the [H[ii]{}]{} region, which is relatively insensitive to the density. The relative proximity of the LBAs allow us to detect the faint [\[O[iii]{}\]$\lambda$4363]{} line and hence determine RO3 (see the spectra in Fig. \[fig:4363\]). However, the electron temperature decreases with increasing metallicity, making the auroral lines too weak to be measured, especially for the more massive systems. In addition, these higher metallicity sources also often show an increased strength of \[Fe[ii]{}\]$\lambda$4359.34, which needs to be deblended from [\[O[iii]{}\]$\lambda$4363]{}. We detected [\[O[iii]{}\]$\lambda$4363]{} in 12 of the cases (S/N of [\[O[iii]{}\]$\lambda$4363]{} larger than 5). The remaining sources in which [\[O[iii]{}\]$\lambda$4363]{} was not detected were excluded from the parts of the analysis that require RO3. To estimate $T_e$(O[iii]{}) we use the fitting function from @proxauf14 that relates the temperature to RO3. The result is shown in Fig. \[fig:R4363\], and indicates that the temperatures are in the range 9,000 to 13,000 K, with the exception of S01\_2 which has 19,000 K. In order to derive total oxygen and nitrogen abundances, we also require the temperatures of the lower ionization states of O and N, $T_e$(O[ii]{}) and $T_e$(N[ii]{}). These can be estimated from the density-dependent $RO2\equiv(I(3726)+I(3729))/(I(7319)+I(7330))$ and $RN2\equiv(I(6548)+I(6584))/(I(5755)$ line ratios, respectively [see @perez-montero17]. Because these ratios depend on the usually weak auroral lines \[O[ii]{}\]$\lambda\lambda$7319,7330 and \[N[ii]{}\]$\lambda$5755, many studies rely on standard conversions between $T_e$(O[iii]{}) and $T_e$(O[ii]{}) [e.g. @garnett92; @proxauf14; @perez-montero17]. Requiring at least a 5$\sigma$ detection in each of the temperature sensitive lines, we were able to measure $T_e$(O[iii]{}) in 12, $T_e$(O[ii]{}) in 22, and $T_e$(N[ii]{}) in 4 sources. The results are compared in Fig. \[fig:temperatures\]. The discrepancies between $T_e$(N[ii]{}) and $T_e$(O[ii]{}) are up to a factor of $\sim$2, i.e. much larger than the formal measurement errors, although there are only 4 sources with both determinations (left panel). The scatter between $T_e$(O[iii]{}) and $T_e$(O[ii]{}) is a few thousand K, and the points lie around the $T_e$(O[iii]{})–$T_e$(O[ii]{}) relations typically assumed for an [H[ii]{}]{} region with an electron density of $\sim$100 cm$^{-3}$ (middle panel). Last, we compare the observed $T_e$(O[ii]{}) with that estimated from $T_e$(O[iii]{}) and the electron density using Equation 14 from @perez-montero17. The right panel of Fig. \[fig:temperatures\] shows the difference between the two values as funcion of the $T_e$(O[ii]{}) measured from RO2, with a full range of $\approx\pm$2000 K. Because our abundances results below are quite sensitive to the correct value for $T_e$(O[ii]{}), in this paper we will evaluate our results both using the RO2-based values and the values inferred from $T_e$(O[iii]{}). For the literature LBAs, we always estimate $T_e$(O[ii]{}) from $T_e$(O[iii]{}) in the same way as done for our own spectra.
{width="\columnwidth"} {width="\columnwidth"}
Direct and strong line oxygen abundance {#sec:direct}
---------------------------------------
Having determined electron densities and temperatures, we now use the formalism presented in @izotov06 to estimate the ‘direct method’ oxygen abundance, $Z_\mathrm{direct}=12 + \mathrm{log(O/H)}$ under the assumption that the total oxygen abundance can be approximated by $\mathrm{O/H}=(\mathrm{O}^++\mathrm{O}^{2+})/\mathrm{H}^+$ (see Eqns. 3 and 5 from @izotov06). This calculation depends on the electron temperatures $T_e$(O[iii]{}) and $T_e$(O[ii]{}), as well as the electron density for which we will use that determined from the [\[S[ii]{}\]$\lambda\lambda$6717,6731]{} doublet in Sect. \[sec:Ne\]. For the determination of $T_e$(O[ii]{}) we have two choices. In the right-hand panel of Fig. \[fig:temperatures\] we showed that the $T_e$(O[ii]{}) estimated from RO2 and that inferred from $T_e$(O[iii]{}) can show significant differences.
These different estimates for the same temperature can lead to significant differences in the derived oxygen abundances. In order to evaluate this effect, we show in Fig. \[fig:ZOH\_bothmethods\] the difference one would obtain when choosing one or the other method. The figure plots O/H calculated using the RO2-based $T_e$(O[ii]{}) and the RO3-based $T_e$(O[iii]{}) versus the difference between O/H estimated using this method and that using $T_e$(O[ii]{}) as inferred from $T_e$(O[iii]{}) (using the electron density and the TO3-TO2 relation from @perez-montero17. The discrepancy in O/H can be up to $\sim$0.4 dex. In the analysis below, we will adopt the $T_e$(O[ii]{}) values as inferred from $T_e$(O[iii]{}). However, in all relevant plots, we also include a second set of symbols (open circles) that give the results for the case where $T_e$(O[ii]{}) was estimated directly from RO2. While results on individual sources may vary, we show that our main conclusions are not affected by either choice.
We find values of 12 + log(O/H) in the range $\sim$8.0–8.6 (with the exception of the extremely metal-poor object S01\_2 which has 12 + log(O/H)$\sim$7.4). These direct method abundances can be compared to the strong line methods typically employed for objects at high redshift or objects with relatively shallow spectra at low redshift. We first compare two common strong line methods based on the $N2\equiv I(6584)/I(H\alpha)$ and $O3N2\equiv(I(5007)/I(H\beta))/(I(6584)/I(H\alpha))$ ratios. Using the @pettini04 calibrations, which relate these line ratios to abundances determined using the direct method applied to a local calibration data set, the oxygen abundances for N2 and O3N2 are in general agreement (Fig. \[fig:ZN2\_vs\_ZO3N2\]). We find slightly higher abundances based on N2 than on O3N2. A similar result was found for $z\sim2$ galaxies from the MOSDEF survey by @sanders15 (red line in Fig. \[fig:ZN2\_vs\_ZO3N2\]), which appears to be a good fit to the LBAs as well. This offset between N2 and O3N2 for LBAs and @sanders15 is about half the size of that found for $z\sim2$ KBSS galaxies by @steidel14, especially at high metallicities. We also indicate the relation found for the local BPT offset analogs from @bian18 (blue solid line). This relation was obtained by taking the best-fits between the direct method abundance and N2 and O3N2 given by @bian18, and converting these back to abundances in the @pettini04 calibration.
The comparison between the two strong line method abundance estimates shown in Fig. \[fig:ZN2\_vs\_ZO3N2\] does not necessarily tell us anything about the accuracy of these strong line methods. Having determined the direct method abundances for LBAs, we can now compare these to either strong line method results. The results are shown in Fig. \[fig:Zdirect\_vs\_ZO3N2\], which compare $Z(T_e)$ with $Z(O3N2)$ (left panel) and $Z(N2)$ (right panel). The direct oxygen abundances do not appear to correlate very strongly with those based on O3N2 or N2, although the dynamic range in O/H of our sample is quite small making it difficult to look for such trends. Overall, however, the direct method abundances appear in general agreement with the strong line methods within an intrinsic scatter of 0.1–0.2 dex and a small systematic offset of at most 0.1 dex. These findings are very similar to those found at $z\sim2$ by @patricio18. The LBAs lie close to the relations for typical SDSS galaxies and BPT offset analogs from @bian18, which have a very similar range of abundances (7.8$\lesssim$12+log(O/H)$_{\mathrm{direct}}$$\lesssim$8.4) as our sources (blue and red lines in Fig. \[fig:Zdirect\_vs\_ZO3N2\]).
In @hoopes07 and @overzier10 it was shown, based on strong line abundances, that LBAs lie below the local mass-metallicity relation of SDSS star-forming galaxies. We confirm this result using the direct method abundances in Fig. \[fig:mzr\]. At a given stellar mass, the LBAs for which direct method abundances could be determined are offset toward lower oxygen abundances compared to the local mass-metallicity relation (MZR). We compare the LBAs with the direct method abundances for stacked SDSS galaxies by @andrews13. The offset of the LBAs toward lower oxygen abundance is similar to the offsets observed for the SDSS stacks for high specific SFR objects (e.g., dashed black line in Fig. \[fig:mzr\]).
![\[fig:mzr\]The MZR relation based on stellar masses from SDSS and direct-method oxygen abundances determined in this paper. The red line shows the best-fit MZR for SDSS from @andrews13, while the green dotted and blue dashed lines show their best-fit MZRs for sub-samples of star-forming objects having $0<\mathrm{log(SFR)}<0.5$ and $1<\mathrm{log(SFR)}<1.5$, respectively. The direct-method abundances derived using the $T_e$(O[ii]{}) derived from RO2 are shown as open circles, while those derived using $T_e$(O[ii]{}) as inferred from $T_e$(O[iii]{}) are shown as filled circles. Literature LBAs are indicated by the green pentagons.](MZR.png){width="\columnwidth"}
Ionization parameter
--------------------
![\[fig:O32\_vs\_logU\]Dimensionless ionization parameter $U$ versus the $O_{32}$. The relation found for KBSS star-forming galaxies at $z\sim2$ from @strom18 is indicated by the red solid line. Objects for which the direct-method abundances were derived using the $T_e$(O[ii]{}) based on RO2 are shown as open circles, while those derived using $T_e$(O[ii]{}) as inferred from $T_e$(O[iii]{}) are shown as filled circles. Literature LBAs are indicated by the green pentagons.](O32_U.png){width="\columnwidth"}
The ionization state of the gas can be quantified by the ionization parameter $q\equiv Q(H^0)/n_H$, the ratio of the hydrogen ionizing photon rate at the surface of a plane-parallel slab and the density of hydrogen, also expressed as the dimensionless ionization parameter $U\equiv n_\gamma/n_H= q/c$, the ratio of the ionizing photon and hydrogen densities. We follow @kojima17 in determining the ionization parameter using the ionization-sensitive $O_{32}\equiv(I(4959)+I(5007))/I(3727)$ ratio. Although $O_{32}$ depends both on $q$ and the oxygen abundance [@kewley02], we can use our direct method estimate of the oxygen abundance to break this degeneracy. We calculate $q$ (and hence log(U)) using Eq. 13 from @kobulnicky04 based on the @kewley02 photoionization model grids. In Fig. \[fig:O32\_vs\_logU\] we show the relation between $O_{32}$ and log(U), finding that LBAs lie a small distance (0–0.1 dex) below the relation determined by @strom18 based on photoionization modeling of KBSS galaxies at $z\sim2$ (red line). @strom18 showed that the log(U) or $q$ in KBSS galaxies correlate well with $O_{32}$, Ne3O2 and O3, but most strongly with $O_{32}$.
In typical local star-forming galaxies, the ionization parameter is strongly anti-correlated with oxygen abundance. @kojima17 provide a best-fit relation based on the direct-method oxygen abundances measured in the SDSS stacks from @andrews13. This relation is indicated in Fig. \[fig:Z\_vs\_q\] along with the LBA data. Although the dynamic range of our sample of LBAs is too small to see whether they follow the same general trend, the majority of the LBAs lie above the local relation at a given oxygen abundance. This result is very similar to that obtained by @kojima17 for a sample of star-forming galaxies at $z=1.4-3.6$. They also noticed that this behavior is similar to that found for a sample of “green pea” galaxies [@amorin12; @jaskot13]. As pointed out above, these include the 3 objects from @amorin12 that are also in our LBA sample. The implications of this average offset in the ionization parameter $q$ at a fixed O/H will be discussed in §3.6 and §3.7 below.
![\[fig:Z\_vs\_q\]Ionization parameter $q$ versus the direct method oxygen abundance. The solid black line indicates the @kojima17 fit to local SDSS stacks from @andrews13. At a given 12 + log(O/H), LBAs tend to lie above the local relation. This is similar as that found for green pea galaxies and samples of $z\sim2$ star-forming galaxies studied by @kojima17. Objects for which the direct-method abundances were derived using the $T_e$(O[ii]{}) based on RO2 are shown as open circles, while those derived using $T_e$(O[ii]{}) as inferred from $T_e$(O[iii]{}) are shown as filled circles. Literature LBAs are indicated by the green pentagons.](ZOH_direct_q.png){width="\columnwidth"}
{width="\textwidth"}
ISM pressure and feedback from massive stars {#sec:SFRD}
--------------------------------------------
It is expected that the pressure in the interstellar medium (ISM) of starburst galaxies is set by the feedback from massive stars [@chevalier85; @heckman90]. The electron density is then the result of the compression of the shocked medium from the rate of momentum-injection from stellar winds and supernovae. The latter can be approximated by the SFR surface density. In the left-hand panel of Fig. \[fig:SFRD\] we plot the electron density estimated from the [\[S[ii]{}\]$\lambda\lambda$6717,6731]{} doublet against the SFR surface density. The SFR was estimated from [H$\alpha$]{} corrected for dust using the Balmer decrement, while the surface area was computed using the seeing-deconvolved half-light radius in the $r^\prime$-band from SDSS. The two quantities are indeed correlated, confirming our hypothesis that the star formation feedback sets the pressure in the ISM in these compact starburst galaxies. With knowledge of both the electron density and the ionization parameter we can furthermore estimate the total hydrogen-ionizing photon rate surface density, $N_ecU$. This is plotted in the right-hand panel of Fig. \[fig:SFRD\], again versus the SFR surface density. This ionizing photon rate surface density can be converted directly into a SFR surface density as well, assuming the standard relation between SFR and the ionizing photon rate of $\mathrm{SFR(Q(H}^0\mathrm{))}=7.4\times10^{-54}\mathrm{Q(H}^0\mathrm{)}$ [e.g. @calzetti13]. These estimates for the SFR surface densities can be read from the top axis of Fig. \[fig:SFRD\], and they are of very similar order of magnitude as the more direct measure of SFR surface density shown on the vertical axis. The dotted line shows the one-to-one relation, indicating that most objects lie within a factor of a few or better from the expected values. This analysis confirms that the conditions in the ISM are directly related to the star formation activity in these compact starburst galaxies.
Nitrogen-to-oxygen abundance ratio {#sec:NO}
----------------------------------
Nitrogen and oxygen have similar ionization potentials, making $\mathrm{N}^+/\mathrm{O}^+$ a good proxy of the nitrogen-to-oxygen abundance ratio, N/O. We use the electron temperatures derived in Sect. \[sec:Ne\] above together with the relations for N$^+$/H$^+$ and O$^+$/H$^+$ from @izotov06 (Eqns. 3 and 6) to estimate N/O. Besides this direct method for N/O, there also exist convenient strong line methods for measuring N/O, such as N2O2$\equiv$[\[N[ii]{}\]$\lambda$6584]{}/[\[O[ii]{}\]$\lambda$3727]{}, which can be easily measured in optical (low redshift) or near-infrared (high redshift) spectra of moderate signal-to-noise. Besides N2O2, N/O also correlates well with other strong line ratios involving nitrogen, such as N2S2$\equiv$[\[N[ii]{}\]$\lambda$6584]{}/[\[S[ii]{}\]$\lambda\lambda$6717,6731]{} and N2$\equiv$[\[N[ii]{}\]$\lambda$6584]{}/[H$\alpha$]{} [e.g. @perez-montero09; @kojima17; @strom17; @strom18].
![\[fig:N2O2\_N2S2\]The relation between strong line N/O abundance indicators N2O2 and N2S2. Fits from the literature are indicated. Black solid line: the median relation found for SDSS galaxies by @strom18. Black dotted line: the fit obtained for a sample of galactic and extragalactic [H[ii]{}]{} regions and galaxies from @perez-montero09. Blue dashed line: the fit to the sample of extra-galactic [H[ii]{}]{} regions by @strom17. Red solid line: the fit to KBSS galaxies at $z\sim2$ from @strom18. Literature LBAs are indicated by the green pentagons.](N2O2_vs_N2S2.png){height="\columnwidth"}
In Fig. \[fig:N2O2\_N2S2\] we first compare N2O2 with N2S2 with various relations from the literature. @strom18 showed that KBSS galaxies at $z\sim2$ lie along a relation that is more similar to that found for local [H[ii]{}]{} regions [@perez-montero09; @strom17] than typical star-forming galaxies from SDSS. LBAs occupy the region in between the relations for $z\sim2$ and [H[ii]{}]{} regions on one hand, and local SDSS galaxies on the other, with most LBAs being closer to the high redshift/local [H[ii]{}]{} lines. At a given N2O2, LBAs have higher N2S2 compared to typical SDSS galaxies. This behavior is very similar to that observed for local [H[ii]{}]{} regions and $z\sim2$ KBSS galaxies, and can be understood by the fact that LBAs and KBSS galaxies have star formation histories that are much simpler and more similar to [H[ii]{}]{} regions compared to typical SDSS star-forming galaxies [@sanders16a; @kashino17; @strom18]. Part of the \[S[ii]{}\] emission may arise from diffuse emission regions in between the [H[ii]{}]{} regions, lowering N2S2 in the case of more complex galaxies [@sanders17].
{width="\textwidth"}
In Fig. \[fig:NO\_stronglines\] we plot the strong line indicators for N/O (N2O2, N2S2, and N2) versus the direct N/O ratio. The majority of LBAs are concentrated in the region between $\mathrm{(N/O)_{direct}}\gtrsim-1.5$ and the solar value (horizontal dotted line). The smallest scatter is found for N2O2, which is not surprising because it is the most direct measure of N$^+$/O$^+$ among these three indicators. The LBAs lie along the relations found for KBSS galaxies and local [H[ii]{}]{} regions. The N2O2–N/O relation for [H[ii]{}]{} regions and galaxies from @perez-montero09 is known to be too steep, overestimating N/O for high N2O2 for both local [H[ii]{}]{} regions and high redshift galaxies [@strom18]. The relation between N/O and the strong line indicators found for the LBAs is more similar to the calibrations found for [H[ii]{}]{} regions and high redshift galaxies [@strom17; @strom18]. Using only those objects in our sample for which the direct method N/O ratio is available, we fit the relation between N/O and N2O2, finding $\mathrm{log(N/O)_{direct}}=0.73 \times\mathrm{N2O2} - 0.58$.
{width="\textwidth"}
{width="46.00000%"} {width="46.00000%"}
{width="70.00000%"}
We will now determine whether the nitrogen-to-oxygen ratios of LBAs are in any way markedly different compared to either typical local or high redshift sources. In Fig. \[fig:NO\_mass\_OH\] we show the direct method nitrogen-to-oxygen ratio log(N/O)$_{\mathrm{direct}}$ versus the stellar mass (left panel), and the direct method oxygen abundances. In both panels, we again give two sets of points, one calculated using T(O[ii]{}) as inferred from T(O[iii]{}) (filled circles) and the other using T(O[ii]{}) estimated from RO2 (open circles). In the left panel of Fig. \[fig:NO\_mass\_OH\], we are furthermore able to show a larger set of points by taking advantage of the best-fit relation between N/O and N2O2 determined for LBAs above (see left panel of Fig. \[fig:NO\_stronglines\]). This allows us to include LBAs for which no direct-method N/O abundances could be measured. These estimates are indicate by the open squares. In the left-hand panel, it can be seen that the LBAs scatter around the typical N/O found for stacks of SDSS galaxies of a similar stellar mass (solid black line), which was also determined through the direct method by @andrews13. @kojima17 find that their small sample of $z\sim2$ galaxies have N/O values comparable or smaller than this local relation, while @strom17 find a somewhat lower N/O on average by $\sim0.1-0.2$ dex for KBSS galaxies at $z\sim2$ compared to local SDSS galaxies with stellar masses in the range $\simeq10^{10}-10^{11}$ $M_\odot$ (red solid line in the left panel). However, in any case, these similar or somewhat lower values of N/O at fixed stellar mass are smaller than the typical downward offsets observed for LBAs in the O/H–$M_*$ relation (see Fig. \[fig:mzr\]). This implies that LBAs, and the $z\sim2$ samples from @kojima17 and @strom17, likely lie along a different N/O–O/H relation compared to typical local star-forming galaxies. We illustrate this in the right panel of Fig. \[fig:NO\_mass\_OH\], where we have indicated again the local sequence based on the SDSS stacks from @andrews13 as determined by @kojima17 (solid black lines). The LBAs are, on average, offset toward higher N/O at fixed O/H.
The N/O excesses observed are of similar order of magnitude as those found for galaxies at $z\sim2$ by @kojima17. The red dashed line in the right panel of Fig. \[fig:NO\_mass\_OH\] is the upper limit on the average N/O ($\mathrm{log(N/O)_{direct}}=-1.26$) for their $z\sim2$ sample, while the blue dashed line is the average N/O ($\mathrm{log(N/O)_{direct}}=-1.08$) they found for a sample of local galaxies selected to have similar (low) stellar masses and (high) star formation rates as their $z\sim2$ sample. @kojima17 also noted that a stacked spectrum of the @steidel16 sample of KBSS galaxies at $z\sim2$ falls close to their sample average (red dashed line), while green peas (which include several of the LBAs from our sample) span the range from no N/O excess to the average N/O for local, low mass, high star formation rate objects (blue dashed line). Our results for LBAs are in general agreement with those trends.
Origin of BPT offsets for Lyman Break Analogs
---------------------------------------------
It has been known for some time that intensely star-forming galaxies at high redshift such as Lyman break galaxies, star-forming BzK galaxies, and distant red galaxies are often offset toward higher line ratios with respect to the mean SDSS star-forming population at low redshift in a standard BPT diagram [e.g. @teplitz00; @shapley05; @erb06; @liu08; @lehnert09; @hayashi09]. Such offsets were also found in some populations of low redshift galaxies, such as the “warm” infrared-luminous galaxies [@kewley01] and Wolf–Rayet galaxies [@brinchmann08a]. These BPT offsets have been observed for LBAs as well [@hoopes07; @overzier08; @overzier09 and Fig. \[fig:bpt\] in this paper], further strengthening the conclusion that they are good local analogs of the high redshift star-forming population. @overzier09 showed that the BPT displacement is in the sense of enhancements in one or both of [\[N[ii]{}\]$\lambda$6584]{}/[H$\alpha$]{} and [\[O[iii]{}\]$\lambda$5007]{}/[H$\beta$]{}, and that the size of the perpendicular offset from the local star-forming ridge increases with both the luminosity of the brightest starburst clump and the electron density, indicating that higher star formation rate densities, interstellar densities and pressures, and ionization parameters may be related to the BPT offsets [e.g. @brinchmann08b; @liu08; @shirazi14].
In recent years, it has become feasible to address in more detail the physical origins of these BPT offsets, driven mainly by the wealth of near-infared spectroscopic data from new surveys of $z\sim2$ galaxies [e.g. @masters14; @steidel14; @shapley15; @kashino17]. Understanding how objects move through the BPT diagram, including the nature of the offsets, is extremely important, because the measurement of gas-phase abundances critically depends on the interpretation of ratios of the strong emission lines. This is particularly important at high redshift, where the temperature sensitive methods are often not available. There is therefore no guarantee that locally established calibrations are directly applicable. Various authors have found that high redshift objects display BPT offsets that are often larger than that can be explained by their increased ionization parameters and densities, and have shown that the BPT offset objects often have an increased N/O abundance and/or ionization parameter at high redshift compared to local galaxies for the same oxygen abundance [e.g. @masters14; @shapley15; @sanders16a; @kojima17]. @steidel16, @strom17 and @strom18 showed that $z\sim2$ galaxies indeed have a higher N/O (at fixed excitation), and that N/O varies with O/H in a similar manner as observed for local [H[ii]{}]{} regions. Besides a higher N/O, they furthermore show the need for a harder ionizing spectrum at high redshift to fully explain the BPT offsets. These harder spectra could be a result of the differences in the star formation (and thus chemical enrichment) histories between typical local and high redshift galaxies.
In the previous sections, we have shown that at fixed O/H, LBAs have higher ionization parameters (Fig. \[fig:Z\_vs\_q\]) and higher N/O (Fig. \[fig:NO\_mass\_OH\], right panel) compared to typical local star-forming galaxies. @kojima17 showed that galaxies at $z\sim2$, as well as a small sample of local analogs consisting of LBAs, green peas and low mass/high SFR objects, display BPT offsets when either one of N/O or $q$ are increased, or both are increased. Positions along the BPT diagram are further modulated by the different O/H of the galaxies. The change in N/O (at fixed O/H) only affects the [\[N[ii]{}\]$\lambda$6584]{}/[H$\alpha$]{} ratio by an amount that directly corresponds to that change (a 0.3 dex increase in N/O gives a 0.3 dex change in [\[N[ii]{}\]$\lambda$6584]{}/[H$\alpha$]{}). The increase in $q$ is in the opposite direction as N/O, with some component along [\[O[iii]{}\]$\lambda$5007]{}/[H$\beta$]{} that is roughly aligned along the star-forming ridge. At 12 + log(O/H)=8.10, a positive change of 0.3 dex in $q$ gives an increase in [\[O[iii]{}\]$\lambda$5007]{}/[H$\beta$]{} of $\sim0.1$ dex but a decrease in [\[N[ii]{}\]$\lambda$6584]{}/[H$\alpha$]{} of $\sim0.2$ dex. Combined with the 0.3 dex increase in N/O, the total offset in the BPT diagram is toward higer values of both [\[O[iii]{}\]$\lambda$5007]{}/[H$\beta$]{} and [\[N[ii]{}\]$\lambda$6584]{}/[H$\alpha$]{}, and in the direction observed for the $z\sim2$ galaxies studied.
In Fig. \[fig:BPT\_offsets\] we show again the location of LBAs in the BPT diagram, but with a color-coding determined by the offsets measured in N/O and $q$ relative to the local relations expectation given the O/H determined for each object through the direct method. The objects with the largest offsets in either parameter (or both) are expected to have the largest BPT offsets, which is indeed the case. The objects in our sample with the largest offsets lie around the average relation found by @steidel14 for galaxies at $z\sim2$. In the bottom right corner of the diagram, we find a number of massive and (likely) more metal-rich LBAs for which the [\[O[iii]{}\]$\lambda$4363]{} flux could not be determined, and thus were not included in the analysis. We show the offsets measured perpendicularly from the local star-forming ridge as a function of the offsets in ionization parameter and N/O in Fig. \[fig:DeltaBPT\]. Objects with the largest perpendicular offsets ($\Delta$BPT$_\perp$) indeed tend to have the largest offsets in $q$ (left panel) or log(N/O) (right panel) or both. From the analysis presented here, we thus conclude that the physical origin of the BPT offsets in LBAs are likely to be the same as those that explain the galaxies at $z\sim2$ and other types of local analogs thereof.
![\[fig:ewhb\]Offset in log(N/O) versus the equivalent width in [H$\beta$]{}. The black solid line shows the average N/O excess measured for WR galaxies compared to non-WR galaxies at the same equivalent width of [H$\beta$]{} from @brinchmann08a, which reaches a maximum value of $\Delta \mathrm{log(N/O)}\sim0.13\pm0.1$ dex at EW([H$\beta$]{}) $\lesssim$ 100 Å. Objects for which the direct-method abundances were derived using the $T_e$(O[ii]{}) based on RO2 are shown as open symbols, while those derived using $T_e$(O[ii]{}) as inferred from $T_e$(O[iii]{}) are shown as filled symbols. Objects marked with large pentagons are LBAs identified with WR features of class 2 or class 3 according to the classification scheme of @brinchmann08a.](EWHB.png){height="\columnwidth"}
Possible origin of the excess N/O
---------------------------------
As shown in Fig. \[fig:NO\_mass\_OH\], LBAs span a wide range of almost 1 dex in N/O values at a given O/H ratio. This wide scatter is similar to that observed for local [H[ii]{}]{} regions and the sample of KBSS galaxies at $z\sim2$ by @strom18. These N/O values furthermore lie, on average, several tenths of dex above the typical N/O expected for local galaxies. This is similar to the trends observed for $z\sim2$ galaxies and other local analogs [e.g. @kojima17; @strom18]. What could be responsible for this N/O excess? The two most straightforward explanations are that N/O is increased either due to the presence of WR stars that enhance nitrogen on a time-scale that is shorter than the SNe enrichment [e.g. @hawley12], or due to a decrease in O/H from the infall of a large quantity of low metallicity gas with the consequence that N/O appears spuriously high for the resulting O/H [e.g. @amorin10; @masters14].
### Evidence for WR stars
Galaxies with WR features often have enhanced N/O. @brinchmann08a compared the average N/O between WR-enhanced galaxies and galaxies without WR features as a function of the equivalent width of [H$\beta$]{} in order to control for difrerences that could arise due to secondary nitrogen production. In Fig. \[fig:ewhb\] we show the excess N/O as a function of the EW([H$\beta$]{}) for our sample. @brinchmann08a found that the WR phase can be responsible for about $0.15\pm0.1$ dex excess N/O for objects with EW([H$\beta$]{}) below $\sim$100 Å, while above that the starbursts are typically younger than the phase where enrichment by WR winds occurs (thick solid line in Fig. \[fig:ewhb\]). Several LBAs are known to have WR features. We have measured the strengths of the WR blue and red bumps in our sample, with two examples of objects with strong bumps shown in Fig. \[fig:wr\]. In total we find 6 objects with clearly identifiable WR features (005527, 015028, 020356, 040208, BPT10 and HST03). These are marked in Fig. \[fig:ewhb\] with the large pentagons. @brown14 also showed evidence for WR features in two additional LBAs (004054 and 092600). While some of the objects with the largest N/O excesses include objects with strong WR features in their spectra, the N/O excess is much larger than expected for WR galaxies. Furthermore, there appears to be no trend in Fig. \[fig:ewhb\] that relates the N/O excess with WR features or EW([H$\beta$]{}).
![\[fig:wr\]Strong Wolf-Rayet features detected in the LBAs SDSS005527 (left panel) and SDSS015028 (right panel). On the left hand side of the top row of panels we show the spectral regions used to measure the flux of the WR spectral features known as the blue (blue shaded region) and red bump (red shaded region). The spectral ranges marked in yellow on either side of the two bumps are used to subtract the continuum flux. On the right hand side of the top row of panels we show an SDSS image stamp of the sources. The bottom row of panels show the best-fit obtained for the blue (left) and red (right) bumps with the red lines showing the characteristic WR bump emission and the blue lines showing the nebular emission lines.](WR005527.png "fig:"){width="0.9\columnwidth"} ![\[fig:wr\]Strong Wolf-Rayet features detected in the LBAs SDSS005527 (left panel) and SDSS015028 (right panel). On the left hand side of the top row of panels we show the spectral regions used to measure the flux of the WR spectral features known as the blue (blue shaded region) and red bump (red shaded region). The spectral ranges marked in yellow on either side of the two bumps are used to subtract the continuum flux. On the right hand side of the top row of panels we show an SDSS image stamp of the sources. The bottom row of panels show the best-fit obtained for the blue (left) and red (right) bumps with the red lines showing the characteristic WR bump emission and the blue lines showing the nebular emission lines.](WR015028.png "fig:"){width="0.9\columnwidth"}
### Evidence for infall of metal-poor gas
Besides nitrogen enhancement from WR stars, the rapid inflow of relatively metal-poor gas is an attractive alternative (or additional) scenario, especially given that most LBAs appear to be starbursts triggered by a recent interaction event [@overzier08; @overzier09]. We showed that at a given stellar mass, LBAs have relatively low O/H with respect to the local MZR (Fig. \[fig:mzr\]), while they have normal N/O for their stellar mass (left panel of Fig. \[fig:NO\_mass\_OH\]). They also have a lower O/H than expected based on their N/O (from the right panel of Fig. \[fig:NO\_mass\_OH\]). This is exactly what would be expected for the accretion scenario. We can show this more clearly in Fig. \[fig:doh\], where we plot the offset in O/H with respect to the local MZR versus the offset in O/H with respect to the local average at a fixed N/O (from the right panel of Fig. \[fig:NO\_mass\_OH\]). The good one-to-one correspondence between them implies that these offsets are one and the same, and related to the infall of a large quantity of metal-poor gas. In order to decrease O/H by 0.3 dex, the galaxy would need to accrete a quantity of metal poor gas equal to the mass of the ISM prior to the infall. On the secondary axes of Fig. \[fig:doh\] we have translated the offsets in O/H observed to the quantities of accreted metal-poor gas in units of the pre-infall gas mass. We have shown that the effect of the presence of WR stars on N/O is expected to be relatively limited, and furthermore is at odds with the normal N/O versus stellar mass relation (left panel of Fig. \[fig:NO\_mass\_OH\]). We conclude that the recent gas accretion scenario can largely, or completely, explain the apparent N/O excesses observed.
![\[fig:doh\]The offset in O/H with respect to the local MZR (from Fig. \[fig:mzr\]) versus the offset in O/H with respect to the local average at a fixed N/O (from the middle panel of Fig. \[fig:NO\_mass\_OH\]). Objects for which the direct-method abundances were derived using the $T_e$(O[ii]{}) based on RO2 are shown as open symbols, while those derived using $T_e$(O[ii]{}) as inferred from $T_e$(O[iii]{}) are shown as filled symbols. The good one-to-one correspondence between them implies that these offsets are one and the same, and related to the infall of a large quantity of metal-poor gas. On the secondary axes we have translated the offsets into the quantities of accreted metal-poor gas needed (in units of the pre-infall gas mass).](dOH_vs_dOH.png){height="\columnwidth"}
Summary
=======
In this paper, we analyzed VLT/XShooter spectra of a sample of LBAs to study the physical parameters of the interstellar medium in luminous UV-selected sources that share many properties with star-forming galaxies at high redshift. Our main findings are summarized as follows.
$\bullet$ We estimated the electron densities and measured the electron temperatures T(\[OIII\]) and T(\[OII\]) for a subset of our sample, which allowed us to estimate direct oxygen abundances. The oxygen abundances of LBAs are in the range $12+\mathrm{log(O/H)}\simeq8.0-8.6$, confirming previous results based on strong-line methods that LBAs are, on average, offset from the local MZR (Fig. \[fig:mzr\]).
$\bullet$ Comparing the direct method abundance estimates with those based on the O3N2 and N2 strong line ratios, we find general agreement within a scatter of 0.1–0.2 dex. This suggests that the strong line methods can be used for LBA-like galaxies at higher redshifts, as shown by other authors.
$\bullet$ We determined the ionization parameter based on the $O_{32}$ index and the oxygen abundance, finding that LBAs have ionization parameters that are typically higher by up to 0.5 dex than typical star-forming galaxies of the same O/H (Fig. \[fig:Z\_vs\_q\]). We show that the SFR surface densities are correlated with the electron densities as expected for an ISM in which the pressure is regulated by the feedback from massive stars and supernovae. From the electron densities and the ionization parameter, we furthermore estimate the hydrogen-ionizing photon flux and show that it agrees remarkably well with the observed SFR surface density (Fig. \[fig:SFRD\]).
$\bullet$ We analyzed nitrogen abundance sensitive strong line ratios (N2O2, N2S2 and N2) and compare them with the N/O ratio determined through the temperature-sensitive method (Figs. \[fig:N2O2\_N2S2\] and \[fig:NO\_stronglines\]). The LBAs tend to follow the same relations between these parameters found for both typical star-forming galaxies at $z\sim2$ as well as local [H[ii]{}]{}-regions [e.g. @strom18].
$\bullet$ On avearge, LBAs lie on or close to the relation between N/O and stellar mass found for typical star-forming galaxies (left panel of Fig. \[fig:NO\_mass\_OH\]), but above the typical N/O expected for galaxies at fixed O/H (right panel of Fig. \[fig:NO\_mass\_OH\]).
$\bullet$ We show that the offsets observed for LBAs in the BPT diagram are linked to the excesses in $q$ and/or N/O (Figs. \[fig:BPT\_offsets\] and \[fig:DeltaBPT\]). The relatively high ionization parameters, relatively low oxygen abundances, excess N/O and BPT offsets are of a similar order of magnitude as those observed by other authors for star-forming galaxies at $z\sim2-3$ as well as previously studied local analogs [e.g. @amorin10; @brown14; @masters16; @sanders16a; @steidel14; @steidel16; @strom17; @kojima17; @bian18; @strom18].
$\bullet$ Finally, we explore the origin of the N/O excess considering the two main scenarios proposed in the literature. We show that WR features observed in the spectra of some LBAs can explain at most a small fraction of the nitrogen enhancement (Figs. \[fig:ewhb\] and \[fig:wr\]). The majority of the excess N/O, however, appears to be related to the recent inflow of large quantities of relatively metal-poor gas, which lowers O/H while leaving N/O unchanged. The relative decrease in O/H at fixed N/O is similar to the decrease in O/H at fixed stellar mass, and suggests that LBAs have experienced accretion of quantities of gas up to several times their original gas mass (Fig. \[fig:doh\]). This is consistent with the conclusions of @amorin10 based on strong line method abundances determinations for Green Pea galaxies and Lyman Break Analogs of @overzier09.\
The analysis performed in this paper highlights some of the difficulties of determining fundamental parameters such as ionization parameter, O/H, and N/O even in relatively nearby galaxies observed with 8-m telescopes. This illustrates some of the challenges we are faced with in determining these parameters and, more importantly, the physical processes that caused them, at much higher redshifts. This is important for upcoming observations with the [*James Webb Space Telescope*]{} which will give access to similar rest-frame optical emission line diagnostics for large samples of high redshift galaxies for the first time.
We are grateful to Andrew Humphrey, Fuyan Bian, Irene Shivaei, Masami Ouchi, Ricardo Amorin, Roberto Cid Fernandes, Tomoko Suzuki, and the anonymous referee for helpful comments. This study was financed in part by the Coordenação de Aperfeiçoamento de Pessoal de Nível Superior - Brasil (CAPES) - Finance Code 001. RAO is grateful for financial support from FAPERJ (202.876/2015), CNPq (400738/2014-7, 309456/2016-9), CAPES (88881.156185/2017-01) and FAPESP (2018/02444-7). The STARLIGHT project is supported by the Brazilian agencies CNPq, CAPES and FAPESP and by the France-Brazil CAPES/Cofecub program. Based on observations made with ESO Telescopes at the La Silla Paranal Observatory under programme IDs 085.B-0784 and 096.B-0192.
Abazajian, K. N., Adelman-McCarthy, J. K., Ag[ü]{}eros, M. A., et al. 2009, , 182, 543 Alexandroff, R., Overzier, R. A., Paragi, Z., et al. 2012, , 423, 1325 Alexandroff, R. M., Heckman, T. M., Borthakur, S., Overzier, R., & Leitherer, C. 2015, , 810, 104 Amor[í]{}n, R. O., P[é]{}rez-Montero, E., & V[í]{}lchez, J. M. 2010, , 715, L128 Amor[í]{}n, R., P[é]{}rez-Montero, E., V[í]{}lchez, J. M., & Papaderos, P. 2012, , 749, 185 Andrews, B. H., & Martini, P. 2013, , 765, 140 Asari N. V., et al., 2007, , 381, 263 Asplund, M., Grevesse, N., Sauval, A. J., & Scott, P. 2009, , 47, 481 Baldwin, J. A., Phillips, M. M., & Terlevich, R. 1981, , 93, 5 Bian, F., Kewley, L. J., Dopita, M. A., & Juneau, S. 2016, , 822, 62 Bian, F., Kewley, L. J., Dopita, M. A., & Blanc, G. A. 2017, , 834, 51 Bian, F., Kewley, L. J., & Dopita, M. A. 2018, , 859, 175 , M. R., [Dalcanton]{}, J., [Eisenstein]{}, D., [et al.]{} 2001, , 121, 2358 Borthakur, S., Heckman, T. M., Leitherer, C., & Overzier, R. A. 2014, Science, 346, 216 Bouwens, R. J., Illingworth, G. D., Oesch, P. A., et al. 2012, , 754, 83 Bouwens, R. J., Illingworth, G. D., Oesch, P. A., et al. 2017, arXiv:1711.02090 Bradley, L. D., Bouwens, R. J., Zitrin, A., et al. 2012, , 747, 3 , E. J., [Almaini]{}, O., [Hartley]{}, W. G., [et al.]{} 2013, , 433, 194 , J., [Kunth]{}, D., & [Durret]{}, F. 2008a, , 485, 657 Brinchmann, J., Pettini, M., & Charlot, S. 2008b, , 385, 769 Brown, J. S., Croxall, K. V., & Pogge, R. W. 2014, , 792, 140 Brown, J. S., Martini, P., & Andrews, B. H. 2016, , 458, 1529 Bruzual, G., & Charlot, S. 2003, , 344, 1000 , D., [Kinney]{}, A. L., & [Storchi-Bergmann]{}, T. 1994, , 429, 582 , D. 2001, , 113, 1449 Calzetti, D. 2013, Secular Evolution of Galaxies, 419 Cardamone, C., Schawinski, K., Sarzi, M., et al. 2009, , 399, 1191 Cardelli, J. A., Clayton, G. C., & Mathis, J. S. 1989, , 345, 245 Chabrier, G. 2003, , 115, 763 Chevalier, R. A., & Clegg, A. W. 1985, , 317, 44 , C. W., & [Vogt]{}, S. S. 2001, , 122, 679 Cid Fernandes R., Mateus A., Sodr[é]{} L., Stasi[ń]{}ska G., Gomes J. M., 2005, , 358, 363 Cid Fernandes, R., Mateus, A., Sodr[é]{}, L., Stasinska, G., & Gomes, J. M. 2011, Astrophysics Source Code Library, ascl:1108.006 , P. S. 1991, , 377, 115 Contursi, A., Baker, A. J., Berta, S., et al. 2017, , 606, A86 Cowie L. L., Barger A. J., Songaila A., 2016, , 817, 57 , M., [Cresci]{}, G., [Mannucci]{}, F., [et al.]{} 2017, , 465, 1384 , R., [Finlator]{}, K., & [Oppenheimer]{}, B. D. 2011, , 416, 1354 Elmegreen, B. G., Elmegreen, D. M., S[á]{}nchez Almeida, J., et al. 2013, , 774, 86 Erb, D. K., Shapley, A. E., Pettini, M., et al. 2006, , 644, 813 , I. F., [de Carvalho]{}, R., [Contini]{}, T., & [Gal]{}, R. R. 2004, , 355, 728 Garland, C. A., Pisano, D. J., Mac Low, M.-M., et al. 2015, , 807, 134 Garnett, D. R. 1992, , 103, 1330 Gburek T., et al., 2019, , submitted (arXiv:1906.11849) Gon[ç]{}alves, T. S., Basu-Zych, A., Overzier, R., et al. 2010, , 724, 1373 Gon[ç]{}alves, T. S., Basu-Zych, A., Overzier, R. A., P[é]{}rez, L., & Martin, D. C. 2014, , 442, 1429 Habouzit, M., Volonteri, M., & Dubois, Y. 2017, , 468, 3935 Hawley, S. A. 2012, , 124, 21 Hayashi, M., Motohara, K., Shimasaku, K., et al. 2009, , 691, 140 Heckman, T. M., Armus, L., & Miley, G. K. 1990, , 74, 833 Heckman, T. M., et al. 2005, , 619, L35 Heckman, T. M., Borthakur, S., Overzier, R., et al. 2011, , 730, 5 Heckman, T. M., Alexandroff, R. M., Borthakur, S., Overzier, R., & Leitherer, C. 2015, , 809, 147 Heckman, T. M., & Borthakur, S. 2016, , 822, 9 Hirschmann M., Charlot S., Feltre A., Naab T., Choi E., Ostriker J. P., Somerville R. S., 2017, , 472, 2468 Hoopes, C., et al. 2007, Accepted for publication in the GALEX special issue of (astro-ph/0609415) , Y. I., [Stasi[ń]{}ska]{}, G., [Meynet]{}, G., [Guseva]{}, N. G., & [Thuan]{}, T. X. 2006, , 448, 955 Izotov, Y. I., Thuan, T. X., & Guseva, N. G. 2007, , 671, 1297 Jaskot, A. E., & Oey, M. S. 2013, , 766, 91 Jia, J., Ptak, A., Heckman, T. M., et al. 2011, , 731, 55 Kashino, D., Silverman, J. D., Sanders, D., et al. 2017, , 835, 88 Kauffmann, G., Heckman, T. M., Tremonti, C., et al. 2003, , 346, 1055 Kausch, W., Noll, S., Smette, A., et al. 2015, , 576, A78 Kewley, L. J., Heisler, C. A., Dopita, M. A., & Lumsden, S. 2001, , 132, 37 Kewley, L. J., & Dopita, M. A. 2002, , 142, 35 Kewley, L. J., Groves, B., Kauffmann, G., & Heckman, T. 2006, , 372, 961 , L. J., & [Ellison]{}, S. L. 2008, , 681, 1183 Kewley L. J., et al., 2013, , 774, 100 , H. A., & [Kewley]{}, L. J. 2004, , 617, 240 , H. A., [Kennicutt]{}, Jr., R. C., & [Pizagno]{}, J. L. 1999, , 514, 544 Kojima, T., Ouchi, M., Nakajima, K., et al. 2017, , 69, 44 K[ö]{}ppen, J., & Hensler, G. 2005, , 434, 531 Latif, M. A., Volonteri, M., & Wise, J. H. 2018, , 476, 5016 Lehnert, M. D., Nesvadba, N. P. H., Le Tiran, L., et al. 2009, , 699, 1660 Leitherer, C., Schaerer, D., Goldader, J. D., et al. 1999, , 123, 3 Liu X.-W., Storey P. J., Barlow M. J., Danziger I. J., Cohen M., Bryce M., 2000, , 312, 585 Liu, X., Shapley, A. E., Coil, A. L., Brinchmann, J., & Ma, C.-P. 2008, , 678, 758 Lofthouse E. K., Houghton R. C. W., Kaviraj S., 2017, , 471, 2311 , R., [Nagao]{}, T., [Grazian]{}, A., [et al.]{} 2008, , 488, 463 Martin, D. C., Fanson, J., Schiminovich, D., et al. 2005, , 619, L1 Masters, D., McCarthy, P., Siana, B., et al. 2014, , 785, 153 Masters, D., Faisst, A., & Capak, P. 2016, , 828, 18 Mateus A., Sodr[é]{} L., Cid Fernandes R., Stasi[ń]{}ska G., Schoenell W., Gomes J. M., 2006, , 370, 721 Meurer, G. R., Heckman, T. M., & Calzetti, D. 1999, , 521, 64 Modigliani, A., Goldoni, P., Royer, F., et al. 2010, , 7737, 773728 Nakajima, K., & Ouchi, M. 2014, , 442, 900 , D. E., & [Cohen]{}, R. D. 1982, , 261, 64 stlin, G., Hayes, M., Duval, F., et al. 2014, , 797, 11 Overzier, R. A., Heckman, T. M., Kauffmann, G., et al. 2008, , 677, 37 Overzier, R. A., Heckman, T. M., Tremonti, C., et al. 2009, , 706, 203 Overzier, R. A., Heckman, T. M., Schiminovich, D., et al. 2010, , 710, 979 Overzier, R. A., Heckman, T. M., Wang, J., et al. 2011, , 726, L7 Patr[í]{}cio, V., Christensen, L., Rhodin, H., Ca[ñ]{}ameras, R., & Lara-L[ó]{}pez, M. A. 2018, , 481, 3520 Pellegrini, E. W., Oey, M. S., Winkler, P. F., et al. 2012, , 755, 40 P[é]{}rez-Montero, E., & Contini, T. 2009, , 398, 949 P[é]{}rez-Montero, E. 2017, , 129, 043001 , M., & [Pagel]{}, B. E. J. 2004, , 348, L59 , B., [[Ö]{}ttl]{}, S., & [Kimeswenger]{}, S. 2014, , 561, A10 , K. H. R., [Weiner]{}, B. J., [Koo]{}, D. C., [et al.]{} 2010, , 719, 1503 , K. H. R., [Prochaska]{}, J. X., [M[é]{}nard]{}, B., [et al.]{} 2011, , 728, 55 Salim, S., & Boquien, M. 2018, arXiv:1812.05606 Sanders, R. L., Shapley, A. E., Kriek, M., et al. 2015, , 799, 138 , R. L., [Shapley]{}, A. E., [Kriek]{}, M., [et al.]{} 2016a, , 816, 23 Sanders, R. L., Shapley, A. E., Kriek, M., et al. 2016b, , 825, L23 Sanders, R. L., Shapley, A. E., Zhang, K., & Yan, R. 2017, , 850, 136 , B. D., & [Sembach]{}, K. R. 1996, , 34, 279 , D., & [Vacca]{}, W. D. 1998, , 497, 618 Schaerer D., Izotov Y. I., Verhamme A., Orlitov[á]{} I., Thuan T. X., Worseck G., Guseva N. G., 2016, , 591, L8 Schlegel, D. J., Finkbeiner, D. P., & Davis, M. 1998, , 500, 525 Schneider, P. 2015, Extragalactic Astronomy and Cosmology (England: Springer Heidelberg New York Dordrecht London), 626, bibliografia: p. 510–571 Shapley, A. E., Coil, A. L., Ma, C.-P., & Bundy, K. 2005, , 635, 1006 Shapley, A. E., Reddy, N. A., Kriek, M., et al. 2015, , 801, 88 Shibuya, T., Ouchi, M., & Harikane, Y. 2015, , 219, 15 Shim, H., & Chary, R.-R. 2013, , 765, 26 Shirazi, M., Brinchmann, J., & Rahmati, A. 2014, , 787, 120 Shivaei I., et al., 2018, , 855, 42 Smette, A., Sana, H., Noll, S., et al. 2015, , 576, A77 , L. 1968, [Diffuse matter in space]{} , G., & [Leitherer]{}, C. 1996, , 107, 661 , C. C., [Rudie]{}, G. C., [Strom]{}, A. L., [et al.]{} 2014, , 795, 165 Steidel, C. C., Strom, A. L., Pettini, M., et al. 2016, , 826, 159 , A. L., [Steidel]{}, C. C., [Rudie]{}, G. C., [et al.]{} 2017, , 836, 164 Strom, A. L., Steidel, C. C., Rudie, G. C., Trainor, R. F., & Pettini, M. 2018, , 868, 117 Teplitz, H. I., McLean, I. S., Becklin, E. E., et al. 2000, , 533, L65 Wang B., Heckman T. M., Leitherer C., Alexandroff R., Borthakur S., Overzier R. A., 2019, , 885, 57 , B. J., [Coil]{}, A. L., [Prochaska]{}, J. X., [et al.]{} 2009, , 692, 187 , S. E., & [Bloom]{}, J. S. 2006, , 44, 507 Wu J. F., Baker A. J., Heckman T. M., Hicks E. K. S., Lutz D., Tacconi L. J., 2019, , in press (arXiv:1911.07854)
[lllllrllll]{} SDSS ID & ID$^a$ & Selection$^b$ & COS$^c$ & R.A. & Dec. & $z^d$ & Date & Seeing$^e$ & Mode$^f$\
& & & & (J2000) & (J2000) & & & FWHM () &\
J001009.97–004603.6 & 001009 & LBA & n & 00:10:09.97 & –00:46:03.66 & 0.243094 & 2010-11-07 & 1.7 & N\
J001054.85+001451.3$^\dagger$ & 001054 & LBA & n & 00:10:54.85 & 00:14:51.35 & 0.243141 & 2010-11-07 & 1.3 & O\
J004054.32+153409.6 & 004054 & LBA & n & 00:40:54.33 & 15:34:09.66 & 0.283241 & 2010-11-08 & 1.4 & N\
J005439.78+155446.9$^\dagger$ & 005439 & LBA & n & 00:54:39.80 & 15:54:46.93 & 0.236400 & 2010-11-06 & 2.0 & N\
J005527.45–002148.7 & 005527 & LBA & y & 00:55:27.46 & –00:21:48.71 & 0.167449 & 2010-11-07 & 1.1 & N\
J015028.39+130858.4 & 015028 & LBA & y & 01:50:28.41 & 13:08:58.40 & 0.146712 & 2010-11-06 & 1.3 & N\
J020356.91–080758.5 & 020356 & LBA & n & 02:03:56.91 & –08:07:58.51 & 0.188335 & 2010-11-07 & 1.1 & N\
J021348.53+125951.4 & 021348 & LBA & y & 02:13:48.54 & 12:59:51.46 & 0.218962 & 2010-09-07 & 1.7 & N\
J032845.99+011150.8 & 032845 & LBA & n & 03:28:45.99 & 01:11:50.85 & 0.142181 & 2010-10-05 & 1.8 & N\
J035733.99–053719.6 & 035733 & LBA & n & 03:57:34.00 & –05:37:19.70 & 0.203746 & 2010-10-05 & 1.8 & N\
J040208.86–050642.0 & 040208 & LBA & n & 04:02:08.87 & –05:06:42.06 & 0.139291 & 2010-10-05 & 1.9 & N\
J143417.15+020742.5 & 143417 & LBA & n & 14:34:17.16 & 02:07:42.58 & 0.180325 & 2010-04-27 & 1.1 & O\
J210358.74–072802.4 & 210358 & LBA & y & 21:03:58.75 & –07:28:02.45 & 0.136840 & 2010-08-10 & 1.6 & O\
J214500.25+011157.5 & 214500 & LBA & n & 21:45:00.26 & 01:11:57.58 & 0.204321 & 2010-11-07 & 1.4 & N\
J231812.99–004126.1 & 231812 & LBA & n & 23:18:13.00 & –00:41:26.10 & 0.251682 & 2010-11-06 & 1.7 & N\
J232539.22+004507.2 & 232539 & LBA & n & 23:25:39.23 & 00:45:07.25 & 0.277000 & 2010-11-06 & 2.4 & N\
J235347.69+005402.0 & 235347 & LBA & n & 23:53:47.69 & 00:54:02.08 & 0.223431 & 2010-10-08 & 1.4 & N\
J124423.37+021540.4 & BPT03 & LBA2, BPT & n & 12:44:23.28 & 02:15:40.40 & 0.238964 & 2016-03-13 & 1.1 & N\
J082247.66+224144.0 & BPT08 & LBA2, BPT & n & 08:22:47.75 & 22:41:44.10 & 0.216226 & 2016-03-12 & 2.1 & N\
J101629.88+073404.9 & BPT09 & LBA2, BPT & n & 10:16:30.00 & 07:34:04.90 & 0.182710 & 2016-03-13 & 0.8 & N\
J124509.05+104340.1 & BPT10 & LBA2, BPT & n & 12:45:09.12 & 10:43:40.00 & 0.165569 & 2016-03-12 & 1.4 & N\
J084034.10+134451.3 & BPT11 & LBA2, BPT & n & 08:40:34.07 & 13:44:51.30 & 0.226961 & 2016-03-13 & 1.3 & N\
J120735.77+082215.5 & BPT15 & LBA2, BPT & n & 12:07:35.76 & 08:22:15.50 & 0.204993 & 2016-03-13 & 1.3 & N\
J102355.73+232338.6 & BPT20 & LBA2, BPT & n & 10:23:55.67 & 23:23:38.70 & 0.254211 & 2016-03-12 & 2.1 & N\
J101009.90+205035.2 & BPT23 & LBA2, BPT & n & 10:10:09.83 & 20:50:35.50 & 0.209547 & 2016-03-12 & 2.1 & N\
J120721.44+021657.7 & BPT26 & LBA2, BPT & n & 12:07:21.35 & 02:16:57.70 & 0.221747 & 2016-03-15 & 0.9 & N\
J141612.87+122340.4 & HST03 & LBA2, COS & y & 14:16:12.96 & 12:23:40.50 & 0.123122 & 2016-03-12 & 2.1 & N\
J104457.79+035313.1 & S01$\_$2 & S2-Deficit & y & 10:44:57.84 & 03:53:13.10 & 0.012879 & 2016-03-13 & 1.3 & O\
J095343.89–000524.7 & S04$\_$1 & S2-Deficit & y & 09:53:43.91 & –00:05:24.60 & 0.083360 & 2016-03-15 & 1.4 & O\
J122627.93+094456.6 & S09$\_$1 & S2-Deficit & y & 12:26:27.84 & 09:44:56.70 & 0.090481 & 2016-03-15 & 0.9 & O\
\
\
\
\
\
\
[lrrrrrrrrrrrrr]{} ID &\
& & & & & & & & & & & & &\
BPT03 & 1762.1$\pm$ 8.6 & 75.8$\pm$1.4 & 867.9$\pm$ 6.7 & 1857.4$\pm$6.3 & 5412.4$\pm$14.8 & – & 38.6$\pm$1.0 & 2562.2$\pm$3.0 & 110.5$\pm$ 4.7 & 120.3$\pm$1.4 & 98.4$\pm$1.2 & 20.6$\pm$0.5 & 16.9$\pm$0.5\
BPT08 & 1879.5$\pm$40.5 & 60.9$\pm$1.6 & 899.4$\pm$ 9.6 & 1730.0$\pm$6.8 & 5229.3$\pm$13.7 & – & 85.5$\pm$1.6 & 2645.3$\pm$4.8 & 171.8$\pm$ 5.8 & 119.5$\pm$1.7 & 110.2$\pm$1.6 & 33.7$\pm$1.9 & 24.8$\pm$1.7\
BPT09 & 1655.3$\pm$11.4 & 42.5$\pm$1.7 & 728.8$\pm$10.8 & 1260.1$\pm$6.4 & 3816.3$\pm$13.8 & – & 51.7$\pm$1.1 & 2160.2$\pm$3.3 & 150.2$\pm$ 4.5 & 107.5$\pm$1.5 & 89.7$\pm$1.3 & 16.5$\pm$0.7 & 13.5$\pm$0.8\
BPT10 & 1534.9$\pm$ 6.3 & 24.6$\pm$0.7 & 864.8$\pm$ 5.5 & 1485.5$\pm$5.8 & 4429.9$\pm$10.6 & 2.7$\pm$0.5 & 63.1$\pm$1.0 & 2502.9$\pm$3.8 & 182.4$\pm$ 5.2 & 148.2$\pm$1.3 & 124.1$\pm$1.1 & 24.7$\pm$0.5 & 19.4$\pm$0.5\
BPT11 & 464.3$\pm$ 4.1 & 11.8$\pm$0.6 & 260.1$\pm$ 4.5 & 463.9$\pm$2.6 & 1392.3$\pm$ 5.3 & – & 20.3$\pm$0.6 & 758.4$\pm$1.8 & 62.6$\pm$ 2.3 & 47.9$\pm$0.9 & 38.1$\pm$0.8 & 7.2$\pm$0.3 & 5.7$\pm$0.5\
BPT15 & 557.8$\pm$ 4.8 & – & 213.8$\pm$ 4.1 & 171.0$\pm$2.7 & 507.9$\pm$ 4.6 & – & 49.3$\pm$1.1 & 626.8$\pm$2.0 & 153.3$\pm$ 2.6 & 59.9$\pm$1.1 & 52.8$\pm$1.0 & 6.1$\pm$0.6 & 5.1$\pm$0.8\
BPT20 & 341.2$\pm$ 7.3 & – & 206.2$\pm$ 4.9 & 124.9$\pm$3.0 & 369.9$\pm$ 4.7 & – & 104.6$\pm$1.6 & 606.0$\pm$2.2 & 280.9$\pm$ 2.5 & 51.6$\pm$1.5 & 47.7$\pm$1.5 & 7.3$\pm$1.2 & 4.8$\pm$0.9\
BPT23 & 605.6$\pm$70.6 & – & 237.4$\pm$ 8.9 & 53.9$\pm$4.8 & 159.7$\pm$ 7.4 & – & 122.3$\pm$2.5 & 703.6$\pm$3.5 & 357.8$\pm$ 3.4 & 107.4$\pm$2.4 & 87.1$\pm$2.2 & 1.2$\pm$1.8 & 13.2$\pm$1.6\
BPT26 & 380.8$\pm$29.3 & – & 304.7$\pm$17.7 & 43.3$\pm$8.0 & 158.6$\pm$14.8 & – & 204.2$\pm$4.0 & 934.6$\pm$4.9 & 569.2$\pm$ 5.7 & 98.8$\pm$3.0 & 102.3$\pm$3.1 & – & –\
HST03 & 7803.5$\pm$46.1 & – & 2149.8$\pm$13.7 & 1718.9$\pm$38.5& 5172.3$\pm$35.4 & 28.4$\pm$5.0 & 667.6$\pm$5.2 & 6465.2$\pm$8.0 & 1864.6$\pm$11.2 & 493.0$\pm$4.8 & 462.6$\pm$4.3 & 59.2$\pm$2.0 & 44.0$\pm$2.1\
S01$\_2$ & 876.5$\pm$10.1 & 522.0$\pm$4.0 & 2907.3$\pm$ 9.9 & 3820.4$\pm$8.9 &10006.9$\pm$ 18.0 & – & 9.7$\pm$1.2 & 8509.2$\pm$7.2 & – & 70.9$\pm$1.6 & 56.4$\pm$1.4 & 19.2$\pm$0.8 & 13.7$\pm$0.8\
S04$\_1$ & – & – & 48.7$\pm$ 3.0 & 13.5$\pm$1.9 & 30.6$\pm$ 3.0 & – & 35.0$\pm$1.6 & 106.7$\pm$1.7 & 106.7$\pm$ 1.9 & – & – & – & –\
S09$\_1$ & – & – & 681.0$\pm$11.0 & 65.4$\pm$5.5 & 170.9$\pm$ 9.2 & – & 716.5$\pm$8.1 & 2041.0$\pm$7.2 & 1718.7$\pm$ 9.0 & 110.3$\pm$4.2 & 134.5$\pm$4.4 & 8.6$\pm$2.4 & 12.4$\pm$2.5\
001009 & 122.3$\pm$ 2.7 & – & 47.6$\pm$ 2.3 & 33.6$\pm$1.4 & 90.7$\pm$ 2.2 & – & 7.7$\pm$0.5 & 138.5$\pm$1.1 & 24.0$\pm$ 1.0 & 19.2$\pm$1.1 & 14.8$\pm$0.9 & – & –\
004054 & 277.4$\pm$ 2.0 & 12.0$\pm$1.8 & 169.7$\pm$ 2.2 & 349.0$\pm$2.2 & 1044.6$\pm$ 4.7 & – & 5.8$\pm$0.5 & 493.5$\pm$1.3 & 19.2$\pm$ 1.7 & 23.7$\pm$0.7 & 18.6$\pm$0.6 & 3.9$\pm$0.8 & 3.3$\pm$1.0\
005527 & 3234.8$\pm$30.0 & 35.9$\pm$1.8 & 1611.3$\pm$10.1 & 1991.9$\pm$8.7 & 5918.5$\pm$16.2 & 9.9$\pm$1.8 & 230.0$\pm$2.1 & 4751.8$\pm$5.4 & 599.6$\pm$ 7.2 & 284.8$\pm$2.3 & 253.4$\pm$2.1 & 56.6$\pm$0.9 & 43.8$\pm$0.9\
015028 & 1344.9$\pm$10.3 & 10.0$\pm$1.2 & 605.9$\pm$16.1 & 449.8$\pm$4.4 & 1345.5$\pm$11.0 & – & 119.8$\pm$1.8 & 1786.2$\pm$3.4 & 366.9$\pm$ 4.1 & 155.7$\pm$1.6 & 133.6$\pm$1.5 & 11.4$\pm$0.5 & 9.5$\pm$0.8\
020356 & 732.9$\pm$ 4.1 & 6.7$\pm$0.6 & 248.3$\pm$ 4.3 & 338.2$\pm$2.8 & 1002.4$\pm$ 6.3 & 2.0$\pm$0.3 & 19.7$\pm$0.6 & 722.9$\pm$1.6 & 57.3$\pm$ 2.0 & 70.2$\pm$1.1 & 53.0$\pm$1.0 & 6.5$\pm$0.3 & 5.5$\pm$0.4\
021348 & – & – & 147.9$\pm$13.6 & – & 98.1$\pm$10.2 & – & 136.1$\pm$3.4 & 444.8$\pm$4.2 & 330.3$\pm$ 4.3 & 42.5$\pm$2.6 & 44.1$\pm$2.8 & – & –\
032845 & 605.7$\pm$ 3.3 & 3.6$\pm$0.5 & 281.2$\pm$ 2.6 & 206.2$\pm$1.9 & 645.0$\pm$ 4.0 & – & 35.7$\pm$0.9 & 815.4$\pm$1.6 & 115.6$\pm$ 2.1 & 80.1$\pm$1.0 & 61.6$\pm$0.8 & 6.5$\pm$0.3 & 5.4$\pm$0.4\
035733 & 847.2$\pm$ 8.1 & – & 233.1$\pm$ 5.4 & 118.4$\pm$3.9 & 363.9$\pm$ 5.1 & – & 40.5$\pm$1.2 & 690.9$\pm$2.6 & 118.8$\pm$ 2.8 & 78.4$\pm$1.4 & 64.2$\pm$1.3 & 5.3$\pm$0.8 & 5.0$\pm$1.2\
040208 & 373.9$\pm$ 2.8 & – & 104.0$\pm$ 2.1 & 89.3$\pm$1.7 & 259.1$\pm$ 3.4 & – & 11.6$\pm$0.6 & 303.0$\pm$1.2 & 36.6$\pm$ 1.4 & 41.3$\pm$1.1 & 28.3$\pm$0.8 & 3.0$\pm$0.4 & 1.9$\pm$0.6\
143417 & 533.8$\pm$ 8.6 & – & 297.8$\pm$10.2 & 63.0$\pm$3.2 & 183.7$\pm$ 6.5 & – & 109.3$\pm$2.1 & 880.5$\pm$2.9 & 324.5$\pm$ 3.9 & 98.7$\pm$2.0 & 77.8$\pm$1.6 & – & –\
214500 & 554.3$\pm$ 4.1 & – & 213.3$\pm$ 3.4 & 88.2$\pm$2.3 & 266.4$\pm$ 3.3 & – & 48.3$\pm$0.9 & 623.2$\pm$1.9 & 148.2$\pm$ 2.3 & 93.8$\pm$1.2 & 71.3$\pm$1.0 & 4.8$\pm$0.6 & 3.6$\pm$1.1\
231812 & 953.9$\pm$ 5.9 & – & 267.2$\pm$ 3.6 & 239.2$\pm$2.6 & 720.7$\pm$ 4.5 & – & 35.6$\pm$0.9 & 784.1$\pm$1.7 & 107.8$\pm$ 2.3 & 101.3$\pm$1.5 & 77.4$\pm$1.2 & 8.6$\pm$0.5 & 6.4$\pm$0.7\
232539 & 237.6$\pm$ 2.7 & – & 101.4$\pm$ 2.9 & 135.0$\pm$1.7 & 413.3$\pm$ 3.5 & – & 6.8$\pm$0.7 & 295.8$\pm$1.6 & 22.9$\pm$ 1.9 & 22.8$\pm$0.9 & 15.0$\pm$0.7 & – & –\
235347 & 352.6$\pm$ 3.9 & 9.1$\pm$0.6 & 133.6$\pm$ 4.4 & 233.6$\pm$2.0 & 664.5$\pm$ 4.3 & – & 5.8$\pm$0.5 & 390.6$\pm$1.4 & 16.1$\pm$ 1.7 & 30.6$\pm$0.8 & 22.6$\pm$0.7 & 4.2$\pm$0.7 & 2.9$\pm$0.4\
\
\
\
[lllllrll]{} SDSS ID & ID$^a$ & Selection$^b$ & COS$^c$ & R.A. & Dec. & $z^d$ & Reference$^e$\
& & & & (J2000) & (J2000) & &\
J004054.32+153409.6 & 004054 (L1) & LBA & n & 00:40:54.33 & +15:34:09.66 & 0.283241 & @amorin12\
J113303.80+651341.3 & 113303 (L2) & LBA & n & 11:33:03.80 & +65:13:41.31 & 0.241 & @amorin12\
J232539.22+004507.2 & 232539 (L3) & LBA & n & 23:25:39.23 & +00:45:07.25 & 0.277000 & @amorin12\
J004054.32+153409.6 & 004054 (L4) & LBA & n & 00:40:54.33 & +15:34:09.66 & 0.283241 & @brown14\
J005527.45–002148.7 & 005527 (L5) & LBA & y & 00:55:27.46 & –00:21:48.71 & 0.167449 & @brown14\
J020356.91–080758.5 & 020356 (L6) & LBA & n & 02:03:56.91 & –08:07:58.51 & 0.188335 & @brown14\
J092600.41+442636.1 & 092600 (L7) & LBA & y & 09:26:00.41 & +44:27:36.13 & 0.18072 & @brown14\
\
\
\
\
---------- -------------------- ------------------------ ------------------------ ------------------------ ------------------------ ------------------- -------------------
ID log($M_*/M_\odot$) log($L_{FUV}/L_\odot$) log($I_{FUV}/L_\odot$) log(\[OIII\]/H$\beta$) log(\[NII\]/H$\alpha$) log($N_e$\[SII\]) log($N_e$\[OII\])
(kpc$^{-2}$) (cm$^{-3}$) (cm$^{-3}$)
BPT03 9.59 10.7 9.66 0.79 -1.37 2.26 2.39
BPT08 9.58 10.5 9.15 0.77 -1.19 2.55 –
BPT09 9.46 10.5 9.28 0.72 -1.16 2.31 2.34
BPT10 9.58 10.7 9.42 0.71 -1.14 2.32 2.5
BPT11 9.29 10.4 9.20 0.73 -1.08 2.16 2.23
BPT15 10.4 10.5 9.81 0.38 -0.61 2.45 2.72
BPT20 10.1 10.6 9.66 0.25 -0.33 2.55 2.9
BPT23 10.7 10.6 9.74 -0.17 -0.30 2.23 –
BPT26 10.8 10.6 9.84 -0.28 -0.22 2.76 2.8
HST03 10.0 10.8 10.3 0.38 -0.54 2.58 2.86
S01$\_2$ 6.96 8.46 10.1 0.54 -2.50 2.17 2.64
S09$\_1$ 10.1 10.1 10.1 -0.60 -0.07 3.01 –
001009 10.5 10.5 9.47 0.28 -0.76 2.05 1.58
004054 9.27 10.3 9.07 0.79 -1.41 2.12 2.43
005527 9.69 10.6 10.1 0.57 -0.90 2.47 1.47
015028 10.3 10.6 9.44 0.35 -0.69 2.39 2.27
020356 9.41 10.6 9.30 0.61 -1.10 1.95 –
021348 10.4 10.7 10.1 -0.18 -0.13 2.76 –
032845 9.82 10.4 9.08 0.36 -0.85 2.03 2.19
035733 9.99 10.6 9.57 0.19 -0.77 2.26 2.19
040208 9.50 10.4 9.51 0.40 -0.92 1.99 1.99
143417 10.7 10.6 9.59 -0.21 -0.43 2.13 2.27
214500 9.98 10.5 9.93 0.10 -0.62 1.98 2.26
231812 10.0 11.0 9.46 0.43 -0.86 2.01 2.1
232539 9.27 10.6 9.72 0.61 -1.11 2.24 2.24
235347 9.47 10.5 9.04 0.70 -1.38 1.83 1.95
---------- -------------------- ------------------------ ------------------------ ------------------------ ------------------------ ------------------- -------------------
[lcccccccc]{} ID & $T_e$(O[iii]{}) & $T_e$(O[ii]{}) & $T_e$(N[ii]{}) & 12 + log(O/H) \[N2\] & 12 + log(O/H) \[O3N2\] & 12 + log(O/H) \[$T_e$\]$^a$ & log(N/O) \[$T_e$\]$^a$ & $q^b$\
& ($\times10^4$ K) & ($\times10^4$ K)& ($\times10^4$ K) & & & &\
BPT03 & 1.20$\pm$ 0.01 & 0.94$\pm$ 0.02 & – & 8.12$\pm$0.01 & 8.04$\pm$0.01 & 8.27$\pm$0.01 & -1.47$\pm$0.02 & 7.94$\pm$0.01\
BPT08 & 1.14$\pm$ 0.01 & 1.09$\pm$ 0.05 & – & 8.22$\pm$0.01 & 8.11$\pm$0.01 & 8.39$\pm$0.02 & -1.40$\pm$0.02 & 7.95$\pm$0.01\
BPT09 & 1.13$\pm$ 0.01 & 0.83$\pm$ 0.03 & – & 8.24$\pm$0.01 & 8.13$\pm$0.01 & 8.33$\pm$0.02 & -1.34$\pm$0.02 & 7.85$\pm$0.01\
BPT10 & 0.94$\pm$ 0.01 & 1.14$\pm$ 0.02 & 0.99$\pm$ 0.09 & 8.25$\pm$0.01 & 8.14$\pm$0.01 & 8.54$\pm$0.01 & -1.31$\pm$0.03 & 8.06$\pm$0.01\
BPT11 & 1.05$\pm$ 0.02 & 1.16$\pm$ 0.05 & – & 8.28$\pm$0.01 & 8.15$\pm$0.01 & 8.37$\pm$0.03 & -1.17$\pm$0.03 & 7.97$\pm$0.02\
BPT15 & – & 0.84$\pm$ 0.06 & – & 8.55$\pm$0.01 & 8.41$\pm$0.01 & – & – & –\
BPT20 & – & 1.21$\pm$ 0.16 & – & 8.71$\pm$0.01 & 8.54$\pm$0.01 & – & – & –\
BPT23 & – & 1.06$\pm$ 0.16 & – & 8.73$\pm$0.01 & 8.69$\pm$0.01 & – & – & –\
BPT26 & – & – & – & 8.78$\pm$0.01 & 8.75$\pm$0.02 & – & – & –\
HST03 & – & 0.60$\pm$ 0.02 & 0.99$\pm$ 0.08 & 8.59$\pm$0.01 & 8.44$\pm$0.01 & – & – & –\
S01$\_2$ & 1.91$\pm$ 0.01 & 1.47$\pm$ 0.06 & – & 7.47$\pm$0.15 & 7.76$\pm$0.08 & 7.41$\pm$0.01 & -1.62$\pm$0.25 & 8.15$\pm$0.01\
S04$\_1$ & – & – & – & 8.90$\pm$0.01 & 8.79$\pm$0.02 & – & – & –\
S09$\_1$ & – & 0.86$\pm$ 0.39 & – & 8.86$\pm$0.01 & 8.90$\pm$0.01 & – & – & –\
001009 & – & – & – & 8.47$\pm$0.01 & 8.40$\pm$0.01 & – & – & –\
004054 & 1.14$\pm$ 0.05 & 1.12$\pm$ 0.17 & – & 8.10$\pm$0.02 & 8.03$\pm$0.01 & 8.29$\pm$0.06 & -1.42$\pm$0.06 & 8.02$\pm$0.03\
005527 & 0.96$\pm$ 0.01 & 1.14$\pm$ 0.02 & 1.03$\pm$ 0.09 & 8.39$\pm$0.01 & 8.26$\pm$0.01 & 8.47$\pm$0.02 & -1.14$\pm$0.01 & 7.84$\pm$0.01\
015028 & 1.01$\pm$ 0.03 & 0.74$\pm$ 0.02 & – & 8.51$\pm$0.01 & 8.40$\pm$0.01 & 8.30$\pm$0.05 & -0.93$\pm$0.02 & 7.56$\pm$0.02\
020356 & 0.98$\pm$ 0.03 & 0.86$\pm$ 0.03 & 1.51$\pm$ 0.15 & 8.27$\pm$0.01 & 8.18$\pm$0.01 & 8.43$\pm$0.05 & -1.41$\pm$0.03 & 7.71$\pm$0.02\
021348 & – & – & – & 8.83$\pm$0.01 & 8.75$\pm$0.02 & – & – & –\
032845 & 0.94$\pm$ 0.03 & 0.94$\pm$ 0.03 & – & 8.42$\pm$0.01 & 8.34$\pm$0.01 & 8.30$\pm$0.06 & -1.06$\pm$0.02 & 7.58$\pm$0.02\
035733 & – & 0.69$\pm$ 0.05 & – & 8.46$\pm$0.01 & 8.42$\pm$0.01 & – & – & –\
040208 & – & 0.76$\pm$ 0.06 & – & 8.38$\pm$0.01 & 8.31$\pm$0.01 & – & – & –\
143417 & – & – & – & 8.65$\pm$0.01 & 8.66$\pm$0.01 & – & – & –\
214500 & – & 0.81$\pm$ 0.07 & – & 8.54$\pm$0.01 & 8.50$\pm$0.01 & – & – & –\
231812 & – & 0.83$\pm$ 0.03 & – & 8.41$\pm$0.01 & 8.32$\pm$0.01 & – & – & –\
232539 & – & – & – & 8.27$\pm$0.02 & 8.18$\pm$0.01 & – & – & –\
235347 & 1.19$\pm$ 0.03 & 0.98$\pm$ 0.08 & – & 8.11$\pm$0.03 & 8.06$\pm$0.02 & 8.19$\pm$0.04 & -1.54$\pm$0.06 & 7.73$\pm$0.02\
\
\
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We introduce an approach to computing the free energy of quasicrystals, which we use to calculate phase diagrams for systems of two-dimensional patchy particles with five regularly arranged patches that have previously been shown to form dodecagonal quasicrystals. We find that the quasicrystal is a thermodynamically stable phase for a wide range of conditions and remains a robust feature of the system as the potential’s parameters are varied. We also demonstrate that the quasicrystal is entropically stabilised over its crystalline approximants.'
author:
- Aleks Reinhardt
- Flavio Romano
- 'Jonathan P. K. Doye'
date: 8 April 2013
title: 'Computing phase diagrams for a quasicrystal-forming patchy-particle system'
---
Quasicrystals are a type of aperiodic crystal structure: they have well-ordered structures, but cannot be described by a periodic lattice, not even one with incommensurate lattice parameters [@Steurer2004; @*Steurer2012]. They were first reported by Shechtman *et al.* [@Shechtman1984], in whose work a rapidly cooled Al-Mn alloy was found to result in a metastable quasicrystal. Most quasicrystals discovered so far are metallic alloys [@Steurer2004; @*Steurer2012]; however, recently, an increasing number of examples have been reported in the field of soft condensed matter [@Zeng2005; @*Fischer2011; @*Dotera2011]. Quasicrystals have also been observed to form in computer simulations, in which not only binary, but also one-component quasicrystals have been seen [@Widom1987; @*Leung1989; @*Dzugutov1993; @*Skibinsky1999; @*Engel2007; @*Keys2007; @*Johnston2010b; @*Johnston2010c; @*ZhiWei2012; @*Dotera2012; @HajiAkbari2009; @HajiAkbari2011; @*HajiAkbari2011b; @Iacovella2011; @VanDerLinden2012]. While certain quasicrystals have been shown to be stable in experiment [@Tsai2003], in simulations, the relatively short timescales accessible mean that it is not necessarily clear whether quasicrystals are truly the stable phase, or rather a metastable phase that is more kinetically accessible. In this work, we address this important question of assessing the stability of quasicrystals and apply it to a soft matter system that has the potential to be realised experimentally.
To prove that quasicrystals are thermodynamically stable, their free energy must be computed. However, such a calculation is not straightforward: the principal problem is that there is no obvious reference state whose free energy is known and from which thermodynamic integration [@Vega2008; @Frenkel1984] could be used to calculate the free energy of the quasicrystalline phase. A phase transition intervenes when integrating from an ideal gas, which is normally used as a reference state for fluid phases, and an Einstein crystal, used as a reference state for crystalline phases, would also not be suitable because it would fail to capture the configurational entropy associated with the quasicrystal’s many possible structures.
One approach to assessing quasicrystal stability is to compute the free energy of an approximant crystalline phase. However, it is not immediately obvious whether quasicrystals are more or less stable than their approximants [@HajiAkbari2009; @HajiAkbari2011; @*HajiAkbari2011b]; whilst the enthalpy and vibrational properties of the quasicrystal and its approximant phases are likely to be similar, the free energy of a quasicrystal has a contribution from its configurational entropy that is not present for the approximant [@Iacovella2011]. Nevertheless, such calculations have provided some insight into the phase behaviour of systems such as hard tetrahedra and bipyramids [@HajiAkbari2011; @*HajiAkbari2011b] and spherical micelles [@Iacovella2011]. By contrast, Kiselev *et al.* recently estimated the free energy of the quasicrystal itself by combining a phonon contribution for a particular quasicrystal configuration from thermodynamic integration and a configurational contribution based on an approximation of uncorrelated phason flips [@Kiselev2012]. They confirmed that, within their approximation, the quasicrystal’s free energy is in certain conditions lower than the approximant’s.
Our solution to this conundrum is to note that thermodynamic integration is not the only way the melting point of a solid can be computed. Another method is to simulate the direct coexistence of two phases with an interface [@Vega2008]. By performing such simulations at a range of temperatures, we can bracket the regions where the quasicrystal melts ($T>T_\text{fus}$) and grows ($T<T_\text{fus}$). At $T=T_\text{fus}$, the chemical potential of the quasicrystal is equal to that of the fluid, which we can calculate using thermodynamic integration. Such an equilibrium approach by construction accounts for the quasicrystal’s configurational entropy. Once the quasicrystal’s free energy is known at one point, we can use thermodynamic integration to reach other points of interest on the phase diagram.
Here, we apply this approach to a two-dimensional system of patchy particles [@Glotzer2007; @*Pawar2010] with five regularly arranged attractive patches, which we model using a simple potential that has previously been used in simulations of self-assembly and crystallisation [@Wilber2007; @*Noya2007; @*Noya2010; @*Williamson2011b; @*Doppelbauer2012b; @Doye2007; @VanDerLinden2012; @Doppelbauer2010]. The potential is based on the Lennard-Jones (LJ) form, $V^{\text{LJ}}(r_{ij}) = 4 \varepsilon [(\sigma_\text{LJ}/r_{ij})^{2n}-
(\sigma_\text{LJ}/r_{ij} )^{n}]$, where $n=6$ for the standard LJ potential, where for $r_{ij}>\sigma_\text{LJ}$, this potential is multiplied by an angular modulation factor $$V^{\text{A}} =
\exp\left[\frac{-\theta_{k_\text{min}ij}^2}{2\,\sigma_\text{pw}^2}\right]
\exp\left[\frac{-\theta_{l_\text{min}ji}^2}{2\,\sigma_\text{pw}^2}\right],$$ where $\sigma_\text{pw}$ is the ‘patch width’ (in radians) and $\theta_{kij}$ is the angle between the patch vector of patch $k$ (chosen to minimise this angle) on particle $i$ and the interparticle vector $\mathbold{r}_{ij}$.
This interaction potential provides a simple model for the patchy colloidal particles that many experimental groups have, with progressively increasing success, been seeking to develop [@DeVries2007; @*Cho2007; @*Yang2008; @*Kraft2009; @*Wang2008; @*Mao2010; @*Duguet2011; @*Wang2012; @Chen2011; @Glotzer2007; @*Pawar2010]. The enhanced range of structural behaviour that such patchy interactions could facilitate has been extensively studied in computer simulation [@Wilber2007; @*Noya2007; @*Noya2010; @*Williamson2011b; @*Doppelbauer2012b; @Doye2007; @VanDerLinden2012; @Doppelbauer2010; @Kern2003; @*Zhang2004; @*Bianchi2006; @*Sciortino2009; @*Romano2011b; @*Bianchi2011].
![Examples of the main phases studied. (a) $\upsigma$ phase. (b) Approximant crystal. (c) Plastic hexagonal (Z) phase. (d) Quasicrystal, $N=2500$. (e) Neighbour classification of $\upsigma$, H and Z environments. The numbers of common neighbours are given for each of the central particle’s neighbours. These local environments are used to colour the particles in (a)–(d), where in addition, particles that do not adopt one of these environments are coloured green. (e) Diffraction pattern corresponding to the configuration shown in (a). (f) Dodecagonal motif characteristic of the quasicrystal.[]{data-label="fig-quasicrystal"}](fig-quasicrystal)
Dodecagonal quasicrystals were previously found to form on cooling systems of these five-patch particles in certain conditions, and structures with the lowest enthalpy were also identified [@VanDerLinden2012]. Some of the relevant phases studied are shown in Fig. \[fig-quasicrystal\](a–d); particles are classified based on a common neighbour analysis [@VanDerLinden2012] into $\upsigma$, H and Z environments (shown in Fig. \[fig-quasicrystal\](e)), by analogy to the Frank–Kasper phases [@Frank1958; @*Frank1959]. The hexagonal (Z) phase (Fig. \[fig-quasicrystal\](c)) was found to form at high pressures (as it is densest) and for wide patches (where the potential is closer to being isotropic). At low pressures and at reasonably narrow patch widths, the structure best satisfying the attractive patches is the enthalpically favoured phase. A crystal cannot exist with five-fold symmetry, and so no crystal with five perfectly aligned patches exists. As a compromise, each particle has five neighbours, but the angles between the neighbours do not perfectly match the five-fold symmetry of the particle itself. Local environments satisfying this requirement are the $\upsigma$ and H environments shown in Fig. \[fig-quasicrystal\](e); the lowest enthalpy structure at low pressure and reasonably narrow patch widths is the $\upsigma$ crystal shown in Fig. \[fig-quasicrystal\](a) [@VanDerLinden2012].
Quasicrystals were observed in some cooling runs in the region where the $\upsigma$ phase is the lowest in enthalpy, but near to the hexagonal ‘boundary’. The structure of a typical quasicrystal is shown in Fig. \[fig-quasicrystal\](d); its diffraction pattern (Fig. \[fig-quasicrystal\](f)) exhibits a characteristic twelve-fold symmetry. The local environments in the quasicrystal are predominantly of the $\upsigma$ type, but a significant fraction of Z environments can be seen: these are typically located at the centre of a dodecagonal motif (Fig. \[fig-quasicrystal\](g)). Much of the structure can be analysed in terms of packing of such dodecagons into triangular, square and rectangular arrangements [@VanDerLinden2012], which can readily be seen in Fig. \[fig-quasicrystal\](d); the various possible ways of arranging these dodecagons are likely to lead to substantial entropy. There is no translational order, but bonds can be orientated in any one of twelve directions, resulting in long-range orientational order of dodecagonal character.
We perform Monte Carlo simulations in the $NpT$ ensemble using a rectangular box with periodic boundaries [^1]. As a starting point in the determination of phase diagrams for this system, we chose a patch width, temperature and pressure at which the quasicrystal is known to form on cooling, and thus locate the fluid-quasicrystal (F-QC) equilibrium transition. This transition is mostly rapid and facile, with essentially no hysteresis, which allows us to determine the coexistence points directly by performing simulations in relatively large boxes starting from both phases. Whilst we initially performed direct coexistence simulations to determine the melting point of the quasicrystal, these simulations mostly did not prove to be any more efficient than direct brute-force simulations, in which no initial interface was introduced, as the growth of the phases was not restricted to the initial interface. This lack of hysteresis is most likely indicative of a particularly low F-QC interfacial free energy. Phase transitions can be observed from a kink in the potential energy or density plotted against the variable we are changing (such as the temperature or the pressure). However, we can further test for the nature of the phase by calculating appropriate diffraction patterns: the quasicrystal exhibits a distinctive 12-fold diffraction pattern (Fig. \[fig-quasicrystal\](f)), the Z phase exhibits a six-fold pattern, and the fluid phase does not exhibit a well-defined pattern.
Once a F-QC coexistence point is known, we can equate the chemical potential of the fluid (as calculated by thermodynamic integration) with that of the quasicrystal. From here, we can use thermodynamic integration to calculate the chemical potential of the quasicrystal at other temperatures and pressures. By also performing thermodynamic integration to find the chemical potential of the $\upsigma$ phase, we can locate the putative $\upsigma$-QC transition. We plot in Fig. \[fig-quasi-free-energies-BP1.5\] the free energies of the $\upsigma$, QC and F phases as a function of $T$ at a constant $p/k_\text{B}T$. From this figure, we can identify two clear transitions ($\upsigma$-QC and QC-F), and a temperature window in which the quasicrystal is thermodynamically stable. In the same figure, we also show the free energy of an approximant crystal phase that is based on the most common packing of dodecagons in the quasicrystal; it is never thermodynamically stable. To rationalise why the quasicrystal is stable, we also calculate the enthalpy (obtained directly from simulations) and hence obtain the entropy as a function of the temperature. These functions are also depicted in Fig. \[fig-quasi-free-energies-BP1.5\], and it is clear that in the region in which the quasicrystal is stable, it does indeed have a higher enthalpy than the $\upsigma$ phase, but its significantly higher entropy nevertheless allows it to become more stable. The origin of this entropy is a combination of the differing vibrational properties and the configurational disorder present in the quasicrystalline phase. If we assume the vibrational properties of the quasicrystal and its approximant are identical, the configurational component of this entropy can be estimated from the entropy difference between the quasicrystal and its approximant, $\upDelta s/k_\text{B}
\approx 0.113$ at $T_{\upsigma \leftrightarrow \text{QC}}$ [^2].
![Per-particle Gibbs energies, enthalpies and entropies, relative to the $\upsigma$ phase, as a function of $T$ at $\sigma_\text{LJ}^2
p/k_\text{B}T =1.5$, $\sigma_\text{pw}=0.49$ and $n=6$. Lines end at the limit of (meta)stability of the phases. Dashed lines denote coexistence points, $k_\text{B}T_{\upsigma\leftrightarrow\text{QC}}/\varepsilon=0.173$ and $k_\text{B}T_{\text{QC}\leftrightarrow\text{F}}/\varepsilon=0.1995$. The stable phase in each region is explicitly marked. The error in $\upDelta g/k_\text{B}T$ is $<0.012\,k_\text{B}T$ [@Note1]. []{data-label="fig-quasi-free-energies-BP1.5"}](fig-quasi-free-energies)
From Fig. \[fig-quasi-free-energies-BP1.5\], we can determine two coexistence points on the phase diagram. We construct the remainder of the phase diagram partly through the same procedure as above at, say, other pressures, but also through the use of Gibbs–Duhem integration [@Kofke1993; @*Kofke1993b; @Vega2008]. This method allows us to integrate Clapeyron equation analogues to obtain new coexistence points from already known coexistence points, but requires metastability and hysteresis in the transition so that the simulated phases retain their identity when simulated at and around the coexistence point. As a result, the method cannot be applied to the QC-F and QC-Z transitions for much of the phase diagram due to their relative reversibility.
![Phase diagrams. Markers show the technique used to obtain each point: Frenkel–Ladd and thermodynamic integration (filled circles), Gibbs–Duhem integration (crosses), hamiltonian Gibbs–Duhem integration (triangles) and direct simulation (open circles). Points obtained from previous phase diagrams are shown as squares. (a) $p/T$-$T$ phase diagram; $\sigma_\text{pw}=0.49$, $n=6$. (b) $T$-$\sigma_\text{pw}$ phase diagram; $\sigma_\text{LJ}^2 p/k_\text{B}T =0.5$, $n=6$. (c) $T$-$n$ phase diagram; $\sigma_\text{LJ}^2 p/k_\text{B}T=0.5$, $\sigma_\text{pw}=0.49$.[]{data-label="fig-quasicrystal-phasediags-all"}](fig-quasicrystal-phasediags-all)
The resulting phase diagram is shown in Fig. \[fig-quasicrystal-phasediags-all\](a). We note that the quasicrystal is stable for a wide range of pressures. At very high pressures, the Z phase becomes stable due to its greater density. The Z phase is a plastic crystal, meaning that particles are able to rotate relatively freely: this is understandable, as there is no obvious preferred way to orientate a five-fold particle in a six-fold environment [^3]. No orientational ordering ensues in the Z phase even at low temperatures. Whether it would form an orientationally-ordered crystal or an orientational glass at sufficiently low $T$ is not clear. It is noteworthy that for the QC-Z transition, $\mathrm{d}p/\mathrm{d}T<0$; since $\rho_\text{Z} > \rho_\text{QC}$, the Clapeyron equation implies that $S_\text{Z} > S_\text{QC}$: this is likely to stem from the orientational entropy of the plastic crystal. Interestingly, in addition to the two triple points ($\upsigma$-QC-Z and QC-Z-F), there is a range of pressures where the stable thermodynamic phase changes from F to Z to QC to $\upsigma$ as the system is cooled.
We also look at the effect of modifying potential parameters on the phase diagram, thus allowing us to assess the extent of quasicrystal stability and how finely-tuned the particle properties would need to be to observe a quasicrystalline phase in experiment. To determine the phase diagram as the potential parameters change, we additionally use hamiltonian Gibbs–Duhem integration, whereby we numerically integrate a Clapeyron equation that has been generalised to allow for changes to the potential itself [@Vega2008].
We first investigate the behaviour of the system as the patch width is varied. From the phase diagram in Fig. \[fig-quasicrystal-phasediags-all\](b), we see that the quasicrystal is only stable for a limited range of $\sigma_\text{pw}$, although the range may perhaps be wider at other pressures. As the patch width narrows, the quasicrystal becomes increasingly enthalpically destabilised with respect to the $\upsigma$ phase, as the many six-fold environments cannot satisfactorily fulfil patch-patch interactions. By contrast, as the patch width increases, the quasicrystal becomes more stable with respect to the $\upsigma$ phase, but the hexagonal phase is stabilised even more: as the patches are wider and closer to the isotropic case, six-fold environments are enthalpically preferred.
Another noteworthy feature of this phase diagram is that below $\sigma_\text{pw}\approx 0.35$, $T_{\upsigma\leftrightarrow\text{F}}$ begins to decrease: as the patches become narrower, there is an increasing energetic penalty for patches that do not point directly at each other. For very narrow patches, the $\upsigma$ phase transforms into a distorted $\upsigma$ phase (D$\upsigma$) [@Doppelbauer2010], in which three of the five patches point directly at the patches of neighbouring particles, thus breaking the square symmetry of the lattice. The reason for this behaviour is that at narrow patch widths, only ‘perfect’ connections count, as a slight misalignment of the patches rapidly reduces the interparticle attraction, and three perfect interactions become favoured over five imperfect ones.
Because colloidal interactions are often quite short-ranged, we also wish to investigate what happens to the system when we change the LJ exponent ($n$), which changes the range of the potential and the width of the potential well. We show the phase diagram as a function of $n$ in Fig. \[fig-quasicrystal-phasediags-all\](c). The effects of increasing $n$ on isotropic potentials are well understood [@Hagen1994; @*Vliegenthart1999]: the liquid phase is energetically destabilised because the intrinsic disorder in interparticle neighbour distances is penalised by the narrower potential [@Doye1996]. In Fig. \[fig-quasicrystal-phasediags-all\](c), the range of stability of the quasicrystal initially increases. The enthalpies of the QC and $\upsigma$ phases increase only slightly as $n$ increases, because their relative order means that most interparticle neighbour distances can be close to optimal, but do so more for the quasicrystal, as some disorder is typically present in the quasicrystal configurations (for example, particles with unclassified environments in Fig. \[fig-quasicrystal\](d)); the $\upsigma$-QC coexistence temperature does therefore increase slightly. By contrast, the fluid is initially much more enthalpically penalised: the fluid enthalpy increases rapidly, and the fluid density decreases as $n$ increases. At the maximum in $T_{\text{QC}\leftrightarrow\text{F}}$ ($n\approx 14$), the size of the temperature window of quasicrystal stability is three times that of the $n=6$ LJ potential. However, a narrower potential well associated with increasing $n$ also means that there is less vibrational entropy associated with the ordered structures, and beyond $n\approx 14$, this effect becomes dominant. The QC-F coexistence temperature decreases until the quasicrystal loses stability at $n\approx 53$.
The potential at the higher values of $n$ shown in Fig. \[fig-quasicrystal-phasediags-all\](c) is very short-ranged indeed. The fact that the quasicrystal maintains and even increases its window of stability for ranges of attraction typical of colloidal particles suggests that this is a robust, general result and that stable quasicrystals can be expected to form in experimental realisations of this system.
There are several potential approaches to such experimental realisations. If patchy colloidal particles similar to the ones studied here could be created, it would be reasonable to expect that a quasicrystal may form when they are confined in two dimensions. For example, Chen *et al.* observed the formation of a 2D kagomé lattice from tri-block Janus particles by introducing a density mismatch with the solvent to confine their colloidal system into two dimensions [@Chen2011]. A possible alternative to using colloidal patchy particles might involve the use of DNA multi-arm motifs [@Yan2003; @*He2005b; @*He2006; @*Zhang2008], for which a variety of 2D crystalline arrays have been observed for motifs with different numbers of arms, including a $\upsigma$ phase for five-arm motifs that is analogous to what we see in the current model when patches are narrow. However, such DNA motifs have a well-defined ‘valence’, and there is no equivalent of the patch width that could be tuned to make two co-ordination numbers compete. Therefore, a possible approach to forming a DNA quasicrystal might be to use a two-component mixture of five-arm and six-arm motifs of the appropriate composition.
In summary, we have shown how we can compute if and where a quasicrystalline phase is thermodynamically stable. To the best of our knowledge, this is the first such calculation of the chemical potential of a quasicrystal and the associated phase diagrams that has been obtained directly from simulations with no approximations. For our patchy particle system, we found that the quasicrystalline phase is stable over a significant portion of the phase diagram and is stabilised primarily by its configurational entropy. It is robust to parameter changes in the model, which inspires confidence that such a thermodynamically stable quasicrystal might be experimentally realised.
We wish to thank the Engineering and Physical Sciences Research Council for financial support.
[69]{}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
References {#references .unnumbered}
==========
[****, ()](\doibase 10.1524/zkri.219.7.391.35643) [****, ()](\doibase 10.1039/C2CS35063G) [****, ()](\doibase 10.1103/PhysRevLett.53.1951) [****, ()](\doibase 10.1016/j.cocis.2004.12.003) [****, ()](\doibase 10.1073/pnas.1008695108) [****, ()](\doibase 10.1002/ijch.201100146) [****, ()](\doibase 10.1103/PhysRevLett.58.706) [****, ()](\doibase 10.1103/PhysRevB.39.446) [****, ()](\doibase 10.1103/PhysRevLett.70.2924) [****, ()](\doibase
10.1103/PhysRevE.60.2664) [****, ()](\doibase 10.1103/PhysRevLett.98.225505) [****, ()](\doibase 10.1103/PhysRevLett.99.235503) [****, ()](\doibase 10.1021/jz101706k) [****, ()](\doibase 10.1063/1.3499323) [****, ()](\doibase 10.1088/0256-307X/29/5/050204) [****, ()](\doibase 10.1002/polb.22395) [****, ()](\doibase
10.1038/nature08641) [****, ()](\doibase 10.1103/PhysRevLett.107.215702) [****, ()](\doibase 10.1063/1.3651370) [****, ()](\doibase 10.1073/pnas.1019763108) [****, ()](\doibase 10.1063/1.3679653) [****, ()](\doibase 10.1021/ar010013x) [****, ()](\doibase
10.1088/0953-8984/20/15/153101) [****, ()](\doibase 10.1063/1.448024) [****, ()](\doibase 10.1103/PhysRevLett.109.225502) [****, ()](\doibase 10.1038/nmat1949) [****, ()](\doibase 10.1002/marc.200900614) [****, ()](\doibase
10.1063/1.2759922) [****, ()](\doibase
10.1063/1.2752155) [****, ()](\doibase
10.1063/1.3454907) [****, ()](\doibase 10.1039/C0SM01377C) [****, ()](\doibase 10.1088/0953-8984/24/28/284124) [****, ()](\doibase 10.1039/b614955c) [****, ()](\doibase 10.1088/0953-8984/22/10/104105) [****, ()](\doibase 10.1126/science.1133162) [****, ()](\doibase 10.1021/cm070051w) [****, ()](\doibase 10.1039/b716393b) [****, ()](\doibase
10.1021/ja8079803) [****, ()](\doibase
10.1002/anie.200801061) [****, ()](\doibase
10.1002/adfm.200902076) [****, ()](\doibase
10.1039/C0CS00048E) [****, ()](\doibase 10.1038/nature11564) [****, ()](\doibase 10.1038/nature09713) [****, ()](\doibase 10.1063/1.1569473) [****, ()](\doibase 10.1021/nl0493500) [****, ()](\doibase
10.1103/PhysRevLett.97.168301) [****, ()](\doibase 10.1103/PhysRevLett.103.237801) [****, ()](\doibase 10.1039/C0SM01494J) [****, ()](\doibase 10.1039/C0CP02296A) [****, ()](\doibase 10.1107/S0365110X58000487) [****, ()](\doibase 10.1107/S0365110X59001499) [****, ()](\doibase 10.1103/PhysRevB.48.6966) [****, ()](\doibase 10.1080/00268979300100881) [****, ()](\doibase 10.1063/1.465023) [****, ()](\doibase 10.1063/1.467526) [****, ()](\doibase 10.1016/S0378-4371(98)00515-9) [****, ()](\doibase 10.1126/science.271.5248.484) [****, ()](\doibase
10.1126/science.1089389) [****, ()](\doibase 10.1021/ja0541938) [****, ()](\doibase
10.1021/ja0665141) [****, ()](\doibase 10.1073/pnas.0803841105) [****, ()](\doibase 10.1063/1.1699114) @noop [**]{}, ed. (, , ) [****, ()](\doibase 10.1080/00268978400101951) [****, ()](\doibase 10.1016/0009-2614(77)85375-X) [****, ()](\doibase 10.1063/1.481457)
[**Supplementary material**]{}
Simulation details
==================
The simulations described in our work used the Metropolis Monte Carlo approach [@Metropolis1953] in the isobaric-isothermal ensemble [@Frenkel2002; @Eppenga1984] with periodic boundary conditions.
The potential as discussed in the manuscript is given more formally by $$V(\mathbold{r}_{ij},\varphi_i,\varphi_j)=\begin{cases}
V^{\text{LJ}}(r_{ij}) & r_{ij}<\sigma_\text{LJ}, \\
V^{\text{LJ}}(r_{ij}) V^{\text{A}}(\hat{\mathbold{r}}_{ij},\varphi_i,\varphi_j) & \sigma_\text{LJ}\le r_{ij} , \\
\end{cases}\label{quasi-eqn-potential}$$ where $\mathbold{r}_{ij}$ is the interparticle vector connecting the centres of the two particles $i$ and $j$, $r_{ij}$ is its magnitude, $\varphi_i$ and $\varphi_j$ are the orientations of the particles $i$ and $j$, the generalised Lennard-Jones potential is $$V^{\text{LJ}}(r_{ij}) =
4 \varepsilon \left[\left(\sigma_\text{LJ}/r_{ij}\right)^{2n}-\left(\sigma_\text{LJ}/r_{ij}\right)^{n}\right],\label{quasi-eqn-LJ}$$ where $n=6$ for the standard Lennard-Jones potential, and $$V^{\text{A}}(\hat{\mathbold{r}}_{ij},\,\varphi_i,\,\varphi_j) = \exp\left[\frac{-\theta_{k_\text{min}ij}^2}{2\,\sigma_\text{pw}^2}\right]\exp\left[\frac{-\theta_{l_\text{min}ji}^2}{2\,\sigma_\text{pw}^2}\right],$$ where $\sigma_\text{pw}$ is the patch width, $\theta_{kij}$ is the angle between the patch vector of patch $k$ on particle $i$ and the interparticle vector $\mathbold{r}_{ij}$, and $k_\text{min}$ is the patch that minimises the magnitude of this angle. Two particles therefore interact only through a single pair of patches. In the simulations reported here, we use a potential cutoff of $r_\text{cut}=3\sigma_\text{LJ}$, and the potential in Eq. is shifted so that it equals zero at $r_{ij}=r_\text{cut}$. The crossover to including angular modulation in the potential (Eq. ) is likewise adjusted so that it still occurs when the potential energy is identically zero.
Most simulations reported used relatively small system sizes (of the order of 500 particles), but the phase diagram in Fig. \[fig-quasicrystal-phasediags-all\](a) and the free energies in Fig. \[fig-quasi-free-energies-BP1.5\] were also calculated using a system of the order of 2500 particles, and certain transitions were calculated with systems of 5000 and 10000 particles, with no detectable differences in the behaviour of the systems, confirming that reasonably small simulations were sufficient to calculate phase diagrams.
Summary of methods
==================
Here, we summarise the methods we used in the manuscript in somewhat more detail. The approaches described here are all standard and are provided here as a quick summary for the benefit of the reader only.
Direct coexistence *vs* brute force simulation
----------------------------------------------
Direct coexistence simulations [@Ladd1977; @Vega2008] are simulations in which an explicit interface is set up between the two phases in question, and the thermodynamically stable phase is expected to grow into the thermodynamically unstable phase such that the interface between the two phases moves. In brute force simulations, no interface is set up; instead, a single phase is simulated and it spontaneously converts into the thermodynamically stable phase. Typically, this brute force simulation process is prone to hysteresis effects, where for example the quasicrystal will form at some temperature $T_1$, but when the quasicrystal is heated in a reverse brute force simulation, it will melt at temperature $T_2$, where $T_2>T_1$, because nucleation involves a free energy barrier that must be overcome. In direct coexistence simulations, this hysteresis does not occur, because both phases are present from the start, and the interface between them is pre-formed.
In our simulations, we established that there is essentially no hysteresis between the quasicrystal and the fluid under the conditions described in this paragraph, and so direct coexistence simulations offer no advantage over brute force simulations. Indeed, because nucleation is so facile, numerous phase interfaces are spontaneously formed in a direct coexistence simulation, and so direct coexistence simulations were not useful in phase diagram determination at this stage.
However, we have used direct coexistence simulations in certain cases as a check in confirming that the Frenkel–Ladd simulations yielded the correct coexistence points.
Thermodynamic integration
-------------------------
Thermodynamic integration [@Frenkel1984; @Vega2008] refers to a series of integration schemes of thermodynamic potentials. For integration along iso-($\beta p$) curves, where $\beta \equiv 1/k_\text{B} T$, we start from the product rule $${\ensuremath{\left( \frac{\partial (G/T)}{\partial T}\right)_{(\beta p)}}} = \frac{1}{T} {\ensuremath{\left( \frac{\partial G}{\partial T}\right)_{(\beta p)}}} - \frac{G}{T^2},\label{eqn-gibbsHelm-bp}$$ from where an application of simple thermodynamics allows us to write $${\ensuremath{\left( \frac{\partial (G/T)}{\partial T}\right)_{(\beta p)}}} = -\frac{U}{T^2}.$$ This can be integrated to give $$\frac{G_2}{N k_\text{B} T_2} = \frac{G_1}{N k_\text{B} T_1} - \int_{T_1}^{T_2} \frac{U}{N k_\text{B} T^2}\,{\ensuremath{\mathrm{d}}}T,$$ and this is the result we require for thermodynamic integration. Provided we know the Gibbs energy at some $T_1$, then we can find the Gibbs energy at $T_2$ along the iso-($\beta p$) curve by running several $NpT$ simulations for temperatures between $T_1$ and $T_2$ (at fixed $\beta p$), and finding a polynomial fit to the integrand, which can then by integrated analytically. Analogous results arise for integration along isobars [@Vega2008], $$\frac{G_2}{N k_\text{B} T_2} = \frac{G_1}{N k_\text{B} T_1} - \int_{T_1}^{T_2} \frac{H}{N k_\text{B} T^2}\,{\ensuremath{\mathrm{d}}}T,$$ and isotherms [@Vega2008], $$\frac{A_2}{N k_\text{B} T} = \frac{A_1}{N k_\text{B} T} + \int_{\rho_1}^{\rho_2} \frac{p}{k_\text{B} T\rho^2}\,{\ensuremath{\mathrm{d}}}\rho.$$
It is also possible to extend the scheme by coupling to a different hamiltonian, such that the potential energy is given by $U=U_\text{real} + \lambda U_\text{extra}$, and then integrating along $\lambda$; we find that [@Vega2008] $$A(\lambda=1) = A(\lambda=0) + \int_{0}^1 \avg{U_\text{extra}}\,{\ensuremath{\mathrm{d}}}\lambda.$$ This is known as hamiltonian thermodynamic integration [@Vega2008] and is one component of the Einstein crystal Frenkel–Ladd approach [@Frenkel1984].
The free energy of a fluid phase evaluated from the ideal gas by decreasing the pressure at constant temperature is given by [@Vega2008] $$\frac{A}{Nk_\text{B}T}=\ln(\rho \Lambda^3) - 1 + \int_0^\rho \left[ \frac{p}{k_\text{B} T\rho^2} - \frac{1}{\rho} \right] \,{\ensuremath{\mathrm{d}}}\rho,$$ where $\Lambda$ is the de Broglie thermal wavelength. It is usually beneficial to fit the entire integrand to a single polynomial to avoid diverging terms. As the density tends to zero, the integrand tends to the second virial coefficient [@Vega2008]; it is helpful to evaluate this explicitly to benchmark the simulation results.
Gibbs–Duhem integration
-----------------------
In Gibbs–Duhem integration [@Kofke1993; @*Kofke1993b; @Vega2008], we numerically integrate the Clapeyron equation, $${\ensuremath{\frac{\mathrm{d}p}{\mathrm{d}T}}} = \frac{\upDelta H}{T\upDelta V},$$ or an analogous expression, in order to obtain new coexistence points from already known coexistence points. In our simulations, we used the fourth-order Runge–Kutta algorithm to perform this integration.
Frenkel–Ladd Einstein crystal approach
--------------------------------------
The Einstein crystal approach [@Frenkel1984; @Vega2008] is a fairly involved way of calculating the free energy of a crystal from the Einstein crystal, for which the free energy is known. We direct the interested reader to Frenkel and Smit’s book (Ref. [@Frenkel2002]) and the review paper by Vega and co-workers (Ref. [@Vega2008]) for an in-depth discussion of the method. In the Frenkel–Ladd scheme as implemented for our model, we take the rotational energy to be $$u_{\text{Ein, or},\,i} = \Lambda_\text{rot} \sin^2 \left[\frac{p}{2} \left(\varphi_i - \varphi_{i,\,\text{orig}} \right)\right],$$ where $\Lambda_\text{rot}$ is a constant that we vary and $p=5$ is the number of patches. We thus ensure that the particle’s symmetry is accounted for: rotations into degenerate positions give the same Frenkel–Ladd energy.
Reliability of free energy calculations
=======================================
Computing absolute free energies from simulations requires a combination of numerical techniques, each of which can introduce errors, both systematic and random. Since it is difficult to keep track of how the errors propagate, the approach which is usually undertaken is to perform *a posteriori* checks on the whole process. As discussed by Vega and co-workers [@Vega2008], ‘consistency checks’ can be applied to simulations to ensure that the data obtained in free energy calculations are self-consistent and correspond to reality. There are several such checks that can be performed. For example, we can calculate the free energy at a particular point using several routes (*e.g.* by integrating along different isotherms, isobars or iso-($\beta p$) curves to the same point). Furthermore, direct coexistence simulations can be used to ensure that the transition temperature does in fact occur where free energy curves intersect. Finally, in simulations involving Gibbs–Duhem integration, we can integrate first in the ‘forward’ and then in the ‘reverse’ direction, ensuring that the method reproduces the starting point. We have performed all these checks to ensure that the results we have obtained are robust.
Gao and co-workers compared the results of Debye crystal reference state calculations with different force constants to obtain an estimate of the error in the free energy of hexagonal ice [@Gao2000]. We do something similar to estimate the error in free energy in Fig. \[fig-quasi-free-energies-BP1.5\] in the main paper: we calculate the Frenkel–Ladd free energy at several points ($k_\text{B}T/\varepsilon=0.1$, $k_\text{B}T/\varepsilon=0.15$, $k_\text{B}T/\varepsilon=0.17$ and $k_\text{B}T/\varepsilon=0.19$) along the iso-($\beta p$) curve, and then use thermodynamic integration to obtain free energies from each of the starting Frenkel–Ladd free energies. The largest difference in free energy between any pair of calculations along the entire curve can then serve as an estimate of the error in free energy in our simulations. The *largest* such deviation was $\upDelta (\upDelta g)/k_\text{B}T = 0.012$, which is sufficiently small not to affect any conclusion, and the error in practice is likely to be significantly smaller, as we were able to confirm phase transition temperatures in independent direct coexistence simulations.
[^1]: See supplementary material in the appendix for implementation details and a summary of simulation methods used.
[^2]: This value of $\upDelta s$ is significantly larger than the tiling entropy of a random Stampfli tiling, but somewhat smaller than that of a maximally random square-triangle tiling, where both have been evaluated at zero phason strain [@Oxborrow1993].
[^3]: We have also considered the additional non-plastic hexagonal crystal structures suggested by Doppelbauer *et al.* for pentavalent patchy particles [@Doppelbauer2010], but these were found quickly to lose their rotational specificity at the temperatures used in our simulations.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The effects of the sample’s boundaries in the magnetic response of the charged anyon fluid at finite temperature are investigated. For the case of an infinite-strip sample it is shown that the Meissner effect takes place at temperatures lower than the fermion energy gap $\omega _{c}$. The temperature dependence of the corresponding effective penetration depth is determined. At temperatures much larger than the scale $\omega _{c}$, a different phase is found, in which the external magnetic field penetrates the fluid.'
address: 'Dept. of Physics, State University of New York, Fredonia, NY 14063, USA'
author:
- '[*E. J. Ferrer and V. de la Incera*]{}'
title: 'BOUNDARY EFFECTS IN 2+1 DIMENSIONAL MAXWELL-CHERN-SIMONS THEORY'
---
Introduction
============
Recently much attention has been given to lower dimensional gauge theories. Such remarkable results as the chiral symmetry breaking [@1], quantum Hall effect [@2], spontaneously broken Lorentz invariance by the dynamical generation of a magnetic field [@3], and the connection between non-perturbative effects in low-energy strong interactions and QCD$%
_{2}$ [@4], show the broad range of applicability of these theories.
In particular, 2+1 dimensional gauge theories with fractional statistics -anyon systems [@4a]- have been extensively studied. One main reason for such an interest has been the belief that a strongly correlated electron system in two dimensions can be described by an effective field theory of anyons [@5], [@5a]. Besides, it has been claimed that anyons could play a basic role in high-T$_{C}$ superconductivity [@5a]-[@6b]. It is known [@a] that a charged anyon system in two spatial dimensions can be modeled by means of a 2+1 dimensional Maxwell-Chern-Simons (MCS) theory. An important feature of this theory is that it violates parity and time-reversal invariance. However, at present no experimental evidences of P and T violation in high-T$_{C}$ superconductivity have been found. It should be pointed out, nevertheless, that it is possible to construct more sophisticated P and T invariant anyonic models[@6a]. In any case, whether linked to high-T$_{C}$ superconductivity or not, the anyon system is an interesting theoretical model in its own right.
The superconducting behavior of anyon systems at $T=0$ has been investigated by many authors [@6]-[@15a]. Crucial to the existence of anyon superconductivity at $T=0$ is the exact cancellation between the bare and induced Chern-Simons terms in the effective action of the theory.
Although a general consensus exists regarding the superconductivity of anyon systems at zero temperature, a similar consensus at finite temperature is yet to be achieved [@8]-[@16]. Some authors (see ref. [@9]) have concluded that the superconductivity is lost at $T\neq 0$, based upon the appearance of a temperature-dependent correction to the induced Chern-Simons coefficient that is not cancelled out by the bare term. In ref. [@11] it is argued, however, that this finite temperature correction is numerically negligible at $T<200$ $K$, and that the main reason for the lack of a Meissner effect is the development of a pole $\sim \left( \frac{1}{{\bf k}%
^{2}}\right) $ in the polarization operator component $\Pi _{00}$ at $T\neq
0 $. There, it is discussed how the existence of this pole leads to a so called partial Meissner effect with a constant magnetic field penetration throughout the sample that appreciably increases with temperature. On the other hand, in ref. [@8], it has been independently claimed that the anyon model cannot superconduct at finite temperature due to the existence of a long-range mode, found inside the infinite bulk at $T\neq 0$. The long range mode found in ref. [@8] is also a consequence of the existence of a pole $\sim \left( \frac{1}{{\bf k}^{2}}\right) $ in the polarization operator component $\Pi _{00}$ at $T\neq 0$.
The apparent lack of superconductivity at temperatures greater than zero has been considered as a discouraging property of anyon models. Nevertheless, it may be still premature to disregard the anyons as a feasible solution for explaining high -T$_{c}$ superconductivity, at least if the reason sustaining such a belief is the absence of the Meissner effect at finite temperature. As it was shown in a previous paper [@16], the lack of a Meissner effect, reported in ref. [@11] for the case of a half-plane sample as a partial Meissner effect, is a direct consequence of the omission of the sample boundary effects in the calculations of the minimal solution for the magnetic field within the sample. To understand this remark we must take into account that the results of ref. [@11] were obtained by finding the magnetization in the bulk due to an externally applied magnetic field at the boundary of a half-plane sample. However, in doing so, a uniform magnetization was assumed and therefore the boundary effects were indeed neglected. Besides, in ref. [@11] the field equations were solved considering only one short-range mode of propagation for the magnetic field, while as has been emphasized in our previous letter [@16], there is a second short-range mode whose qualitative contribution to the solutions of the field equations cannot be ignored.
In the present paper we study the effects of the sample’s boundaries in the magnetic response of the anyon fluid at finite temperature. This is done by considering a sample shaped as an infinite strip. When a constant and homogeneous external magnetic field, which is perpendicular to the sample plane, is applied at the boundaries of the strip, two different magnetic responses, depending on the temperature values, can be identified. At temperatures smaller than the fermion energy gap inherent to the many-particle MCS model ($T\ll \omega _{c}$), the system exhibits a Meissner effect. In this case the magnetic field cannot penetrate the bulk farther than a very short distance ($\overline{\lambda }\sim 10^{-5}cm$ for electron densities characteristic of the high -T$_{c}$ superconductors and $T\sim 200$ $K$). On the other hand, as it is natural to expect from a physical point of view, when the temperatures are larger than the energy gap ($T\gg \omega
_{c} $) the Meissner effect is lost. In this temperature region a periodic inhomogeneous magnetic field is present within the bulk.
These results, together with those previously reported in ref. [@16], indicate that, contrary to some authors’ belief, the superconducting behavior (more precisely, the Meissner effect), found in the charged anyon fluid at $T=0$, does not disappear as soon as the system is heated.
As it is shown below, the presence of boundaries can affect the dynamics of the system in such a way that the mode that accounts for a homogeneous field penetration [@8] cannot propagate in the bulk. Although these results have been proved for two types of samples, the half-plane [@16] and the infinite strip reported in this paper, we conjecture that similar effects should also exist in other geometries.
Our main conclusion is that the magnetic behavior of the anyon fluid is not just determined by its bulk properties, but it is essentially affected by the sample boundary conditions. The importance of the boundary conditions in 2+1 dimensional models has been previously stressed in ref.[@b].
The plan for the paper is as follows. In Sec. 2, for completeness as well as for the convenience of the reader, we define the many-particle 2+1 dimensional MCS model used to describe the charged anyon fluid, and briefly review its main characteristics. In Sec. 3 we study the magnetic response in the self-consistent field approximation of a charged anyon fluid confined to an infinite-strip, finding the analytical solution of the MCS field equations that satisfies the boundary conditions. The fermion contribution in this approximation is given by the corresponding polarization operators at $T\neq 0$ in the background of a many-particle induced Chern-Simons magnetic field. Using these polarization operators in the low temperature approximation ($T\ll \omega _{c}$), we determine the system’s two London penetration depths. Taking into account that the boundary conditions are not enough to completely determine the magnetic field solution within the sample, an extra physical condition, the minimization of the system free-energy density, is imposed. This is done in Sec. 4. In this section we prove that even though the electromagnetic field has a long-range mode of propagation in the charged anyon fluid at $T\neq 0$ [@8], a constant and uniform magnetic field applied at the sample’s boundaries cannot propagate through this mode. The explicit temperature dependence at $T\ll \omega _{c}$ of all the coefficients appearing in the magnetic field solution, and of the effective London penetration depth are also found. In Sec. 5, we discuss how the superconducting behavior of the charged anyon fluid disappears at temperatures larger than the energy gap ($T\gg \omega _{c}$). Sec. 6 contains the summary and discussion.
MCS Many-Particle Model
=======================
The Lagrangian density of the 2+1 dimensional non-relativistic charged MCS system is
$${\cal L}=-\frac{1}{4}F_{\mu \nu }^{2}-\frac{N}{4\pi }\varepsilon ^{\mu \nu
\rho }a_{\mu }\partial _{\nu }a_{\rho }+en_{e}A_{0}+i\psi ^{\dagger
}D_{0}\psi -\frac{1}{2m}\left| D_{k}\psi \right| ^{2}+\psi ^{\dagger }\mu
\psi \eqnum{2.1}$$
where $A_{\mu }$ and $a_{\mu }$ represent the electromagnetic and the Chern-Simons fields respectively. The role of the Chern-Simons fields is simply to change the quantum statistics of the matter field, thus, they do not have an independent dynamics. $\psi $ represents non-relativistic spinless fermions. $N\ $ is a positive integer that determines the magnitude of the Chern-$%
%TCIMACRO{\func{Si}}
%BeginExpansion
\mathop{\rm Si}%
%EndExpansion
$mons coupling constant. The charged character of the system is implemented by introducing a chemical potential $\mu $; $n_{e}$ is a background neutralizing ‘classical’ charge density, and $m$ is the fermion mass. We will consider throughout the paper the metric $g_{\mu \nu }$=$(1,-%
\overrightarrow{1})$. The covariant derivative $D_{\nu }$ is given by
$$D_{\nu }=\partial _{\nu }+i\left( a_{\nu }+eA_{\nu }\right) ,\qquad \nu
=0,1,2 \eqnum{2.2}$$
It is known that to guarantee the system neutrality in the presence of a different from zero fermion density $\left( n_{e}\neq 0\right) $,$\ $a nontrivial background of Chern-Simons magnetic field $\left( \overline{b}=%
\overline{f}_{21}\right) $ is generated. The Chern-Simons background field is obtained as the solution of the mean field Euler-Lagrange equations derived from (2.1)
$$-\frac{N}{4\pi }\varepsilon ^{\mu \nu \rho }f_{\nu \rho }=\left\langle
j^{\mu }\right\rangle \eqnum{2.3}$$
$$\partial _{\nu }F^{\mu \nu }=e\left\langle j^{\mu }\right\rangle
-en_{e}\delta ^{\mu 0} \eqnum{2.4}$$
considering that the system formed by the electron fluid and the background charge $n_{e}$ is neutral
$$\left\langle j^{0}\right\rangle -n_{e}\delta ^{\mu 0}=0 \eqnum{2.5}$$
In eq. (2.5) $\left\langle j^{0}\right\rangle $ is the fermion density of the many-particle fermion system
$$\left\langle j^{0}\right\rangle =\frac{\partial \Omega }{\partial \mu },
\eqnum{2.6}$$
$\Omega $ is the fermion thermodynamic potential.
In this approximation it is found from (2.3)-(2.5) that the Chern-Simons magnetic background is given by
$$\overline{b}=\frac{2\pi n_{e}}{N} \eqnum{2.7}$$
Then, the unperturbed one-particle Hamiltonian of the matter field represents a particle in the background of the Chern-Simons magnetic field $%
\overline{b\text{,}}$
$$H_{0}=-\frac{1}{2m}\left[ \left( \partial _{1}+i\overline{b}x_{2}\right)
^{2}+\partial _{2}^{2}\right] \eqnum{2.8}$$
In (2.8) we considered the background Chern-Simons potential, $\overline{a}%
_{k}$, $(k=1,2)$, in the Landau gauge
$$\overline{a}_{k}=\overline{b}x_{2}\delta _{k1} \eqnum{2.9}$$
The eigenvalue problem defined by the Hamiltonian (2.8) with periodic boundary conditions in the $x_{1}$-direction: $\Psi \left( x_{1}+L,\text{ }%
x_{2}\right) =$ $\Psi \left( x_{1},\text{ }x_{2}\right) $,
$$H_{0}\Psi _{nk}=\epsilon _{n}\Psi _{nk},\qquad n=0,1,2,...\text{ }and\text{ }%
k\in {\cal Z} \eqnum{2.10}$$
has eigenvalues and eigenfunctions given respectively by
$$\epsilon _{n}=\left( n+\frac{1}{2}\right) \omega _{c}\qquad \eqnum{2.11}$$
$$\Psi _{nk}=\frac{\overline{b}^{1/4}}{\sqrt{L}}\exp \left( -2\pi
ikx_{1}/L\right) \varphi _{n}\left( x_{2}\sqrt{\overline{b}}-\frac{2\pi k}{L%
\sqrt{\overline{b}}}\right) \eqnum{2.12}$$
where $\omega _{c}=\overline{b}/m$ is the cyclotron frequency and $\varphi
_{n}\left( \xi \right) $ are the orthonormalized harmonic oscillator wave functions.
Note that the energy levels $\epsilon _{n}$ are degenerates (they do not depend on $k$). Then, for each Landau level $n$ there exists a band of degenerate states. The cyclotron frequency $\omega _{c}$ plays here the role of the energy gap between occupied Landau levels. It is easy to prove that the filling factor, defined as the ratio between the density of particles $%
n_{e}$ and the number of states per unit area of a full Landau level, is equal to the Chern-$%
%TCIMACRO{\func{Si}}
%BeginExpansion
\mathop{\rm Si}%
%EndExpansion
$mons coupling constant $N$. Thus, because we are considering that $N$ is a positive integer, we have in this MCS theory $N$ completely filled Landau levels. Once this ground state is established, it can be argued immediately [@6], [@6b], [@10a], [@15], that at $T=0$ the system will be confined to a filled band, which is separated by an energy gap from the free states; therefore, it is natural to expect that at $T=0$ the system should superconduct. This result is already a well established fact on the basis of Hartree-Fock analysis[@6] and Random Phase Approximation [@6b],[@10a]. The case at $T\neq 0$ is more controversial since thermal fluctuations, occurring in the many-particle system, can produce significant changes. As we will show in this paper, the presence in this theory of a natural scale, the cyclotron frequency $\omega _{c}$, is crucial for the existence of a phase at $T\ll \omega _{c}$, on which the system, when confined to a bounded region, still behaves as a superconductor.
The fermion thermal Green’s function in the presence of the background Chern-Simons field $\overline{b}$
$$G\left( x,x^{\prime }\right) =-\left\langle T_{\tau }\psi \left( x\right)
\overline{\psi }\left( x^{\prime }\right) \right\rangle \eqnum{2.13}$$
is obtained by solving the equation
$$\left( \partial _{\tau }-\frac{1}{2m}\overline{D}_{k}^{2}-\mu \right)
G\left( x,x^{\prime }\right) =-\delta _{3}\left( x-x^{\prime }\right)
\eqnum{2.14}$$
subject to the requirement of antiperiodicity under the imaginary time translation $\tau \rightarrow \tau +\beta $ ($\beta $ is the inverse absolute temperature). In (2.14) we have introduced the notation
$$\overline{D}_{k}=\partial _{k}+i\overline{a}_{k} \eqnum{2.15}$$
The Fourier transform of the fermion thermal Green’s function (2.13)
$$G\left( p_{4},{\bf p}\right) =\int\limits_{0}^{\beta }d\tau \int d{\bf x}%
G\left( \tau ,{\bf x}\right) e^{i\left( p_{4}\tau -{\bf px}\right) }
\eqnum{2.16}$$
can be expressed in terms of the orthonormalized harmonic oscillator wave functions $\varphi _{n}\left( \xi \right) $ as [@Efrain]
$$\begin{aligned}
G\left( p_{4},{\bf p}\right) &=&\int\limits_{0}^{\infty }d\rho
\int\limits_{-\infty }^{\infty }dx_{2}\sqrt{\overline{b}}\exp -\left(
ip_{2}x_{2}\right) \exp -\left( ip_{4}+\mu -\frac{\overline{b}}{2m}\right)
\rho \nonumber \\
&&\sum\limits_{n=0}^{\infty }\varphi _{n}\left( \xi \right) \varphi
_{n}\left( \xi ^{\prime }\right) t^{n} \eqnum{2.17}\end{aligned}$$
where $t=\exp \frac{\overline{b}}{m}\rho $, $\xi =\frac{p_{1}}{\sqrt{%
\overline{b}}}+\frac{x_{2}\sqrt{\overline{b}}}{2}$, $\xi ^{\prime }=\frac{%
p_{1}}{\sqrt{\overline{b}}}-\frac{x_{2}\sqrt{\overline{b}}}{2}$ and $%
p_{4}=(2n+1)\pi /\beta $ are the discrete frequencies $(n=0,1,2,...)$ corresponding to fermion fields.
Linear Response in the Infinite Strip
=====================================
Effective Theory at $\mu \neq 0$ and $T\neq 0$
----------------------------------------------
In ref.[@8] the effective current-current interaction of the MCS model was calculated to determine the independent components of the magnetic interaction at finite temperature in a sample without boundaries, i.e., in the free space. These authors concluded that the pure Meissner effect observed at zero temperature is certainly compromised by the appearance of a long-range mode at $T\neq 0$. Our main goal in the present paper is to investigate the magnetic response of the charged anyon fluid at finite temperature for a sample that confines the fluid within some specific boundaries. As we prove henceforth, the confinement of the system to a bounded region (a condition which is closer to the experimental situation than the free-space case) is crucial for the realization of the Meissner effect inside the charged anyon fluid at finite temperature.
Let us investigate the linear response of a charged anyon fluid at finite temperature and density to an externally applied magnetic field in the specific case of an infinite-strip sample. The linear response of the medium can be found under the assumption that the quantum fluctuations of the gauge fields about the ground-state are small. In this case the one-loop fermion contribution to the effective action, obtained after integrating out the fermion fields, can be evaluated up to second order in the gauge fields. The effective action of the theory within this linear approximation [@8],[@11] takes the form
$$\Gamma _{eff}\,\left( A_{\nu },a_{\nu }\right) =\int dx\left( -\frac{1}{4}%
F_{\mu \nu }^{2}-\frac{N}{4\pi }\varepsilon ^{\mu \nu \rho }a_{\mu }\partial
_{\nu }a_{\rho }+en_{e}A_{0}\right) +\Gamma ^{\left( 2\right) } \eqnum{3.1}$$
$$\Gamma ^{\left( 2\right) }=\int dx\Pi ^{\nu }\left( x\right) \left[ a_{\nu
}\left( x\right) +eA_{\nu }\left( x\right) \right] +\int dxdy\left[ a_{\nu
}\left( x\right) +eA_{\nu }\left( x\right) \right] \Pi ^{\mu \nu }\left(
x,y\right) \left[ a_{\nu }\left( y\right) +eA_{\nu }\left( y\right) \right]$$ Here $\Gamma ^{\left( 2\right) }$ is the one-loop fermion contribution to the effective action in the linear approximation. The operators $\Pi _{\nu }$ and $\Pi _{\mu \nu }$ are calculated using the fermion thermal Green’s function in the presence of the background field $\overline{b}$ (2.17). They represent the fermion tadpole and one-loop polarization operators respectively. Their leading behaviors for static $\left( k_{0}=0\right) $ and slowly $\left( {\bf k}\sim 0\right) $ varying configurations in the frame ${\bf k}=(k,0)$ take the form
$$\Pi _{k}\left( x\right) =0,\;\;\;\Pi _{0}\left( x\right) =-n_{e},\;\;\;\Pi
_{\mu \nu }=\left(
\begin{array}{ccc}
{\it \Pi }_{{\it 0}}+{\it \Pi }_{{\it 0}}\,^{\prime }\,k^{2} & 0 & {\it \Pi }%
_{{\it 1}}k \\
0 & 0 & 0 \\
-{\it \Pi }_{{\it 1}}k & 0 & {\it \Pi }_{\,{\it 2}}k^{2}
\end{array}
\right) , \eqnum{3.2}$$
The independent coefficients: ${\it \Pi }_{{\it 0}}$, ${\it \Pi }_{{\it 0}%
}\,^{\prime }$, ${\it \Pi }_{{\it 1}}$ and ${\it \Pi }_{\,{\it 2}}$ are functions of $k^{2}$, $\mu $ and $\overline{b}$. In order to find them we just need to calculate the $\Pi _{\mu \nu }$ Euclidean components: $\Pi
_{44} $, $\Pi _{42}$ and $\Pi _{22}$. In the Landau gauge these Euclidean components are given by[@11],
$$\Pi _{44}\left( k,\mu ,\overline{b}\right) =-\frac{1}{\beta }%
\sum\limits_{p_{4}}\frac{d{\bf p}}{\left( 2\pi \right) ^{2}}G\left( p\right)
G\left( p-k\right) , \eqnum{3.3}$$
$$\Pi _{4j}\left( k,\mu ,\overline{b}\right) =\frac{i}{2m\beta }%
\sum\limits_{p_{4}}\frac{d{\bf p}}{\left( 2\pi \right) ^{2}}\left\{ G\left(
p\right) \cdot D_{j}^{-}G\left( p-k\right) +D_{j}^{+}G\left( p\right) \cdot
G\left( p-k\right) \right\} , \eqnum{3.4}$$
$$\begin{aligned}
\Pi _{jk}\left( k,\mu ,\overline{b}\right) &=&\frac{1}{4m^{2}\beta }%
\sum\limits_{p_{4}}\frac{d{\bf p}}{\left( 2\pi \right) ^{2}}\left\{
D_{k}^{-}G\left( p\right) \cdot D_{j}^{-}G\left( p-k\right)
+D_{j}^{+}G\left( p\right) \cdot D_{k}^{+}G\left( p-k\right) \right.
\nonumber \\
&&\left. +D_{j}^{+}D_{k}^{-}G\left( p\right) \cdot G\left( p-k\right)
+G\left( p\right) \cdot D_{j}^{-}D_{k}^{+}G\left( p-k\right) \right\}
\nonumber \\
&&-\frac{1}{2m}\Pi _{4}, \eqnum{3.5}\end{aligned}$$
where the notation
$$\begin{aligned}
D_{j}^{\pm }G\left( p\right) &=&\left[ ip_{j}\mp \frac{\overline{b}}{2}%
\varepsilon ^{jk}\partial _{p_{k}}\right] G\left( p\right) , \nonumber \\
D_{j}^{\pm }G\left( p-k\right) &=&\left[ i\left( p_{j}-k_{j}\right) \mp
\frac{\overline{b}}{2}\varepsilon ^{jk}\partial _{p_{k}}\right] G\left(
p-k\right) , \eqnum{3.6}\end{aligned}$$
was used.
Using (3.3)-(3.5) after summing in $p_{4}$, we found that, in the $k/\sqrt{%
\overline{b}}\ll 1$ limit, the polarization operator coefficients ${\it \Pi }%
_{{\it 0}}$, ${\it \Pi }_{{\it 0}}\,^{\prime }$, ${\it \Pi }_{{\it 1}}$ and $%
{\it \Pi }_{\,{\it 2}}$ are
$${\it \Pi }_{{\it 0}}=\frac{\beta \overline{b}}{8\pi {\bf k}^{2}}%
\sum_{n}\Theta _{n},\;\qquad {\it \Pi }_{{\it 0}}\,^{\prime }=\frac{2m}{\pi
\overline{b}}\sum_{n}\Delta _{n}-\frac{\beta }{8\pi }\sum_{n}(2n+1)\Theta
_{n},$$
$${\it \Pi }_{{\it 1}}=\frac{1}{\pi }\sum_{n}\Delta _{n}-\frac{\beta \overline{%
b}}{16\pi m}\sum_{n}(2n+1)\Theta _{n},\qquad {\it \Pi }_{\,{\it 2}}=\frac{1}{%
\pi m}\sum_{n}(2n+1)\Delta _{n}-\frac{\beta \overline{b}}{32\pi m^{2}}%
\sum_{n}(2n+1)^{2}\Theta _{n},$$
$$\Theta _{n}=%
%TCIMACRO{\func{sech} }
%BeginExpansion
\mathop{\rm sech}%
%EndExpansion
\,^{2}\frac{\beta (\epsilon _{n}/2-\mu )}{2},\qquad \Delta _{n}=(e^{\beta
(\epsilon _{n}/2-\mu )}+1)^{-1} \eqnum{3.7}$$
The leading contributions of the one-loop polarization operator coefficients (3.7) at low temperatures $\left( T\ll \omega _{c}\right) $ are
$${\it \Pi }_{{\it 0}}=\frac{2\beta \overline{b}}{\pi }e^{-\beta \overline{b}%
/2m},\qquad {\it \Pi }_{{\it 0}}\,^{\prime }=\frac{mN}{2\pi \overline{b}}%
{\it \Lambda },\qquad {\it \Pi }_{{\it 1}}=\frac{N}{2\pi }{\it \Lambda }%
,\quad {\it \Pi }_{\,{\it 2}}=\frac{N^{2}}{4\pi m}{\it \Lambda },\qquad {\it %
\Lambda }=\left[ 1-\frac{2\beta \overline{b}}{m}e^{-\beta \overline{b}%
/2m}\right] \eqnum{3.8}$$
and at high temperatures $\left( T\gg \omega _{c}\right) $ are
$${\it \Pi }_{{\it 0}}=\frac{m}{2\pi }\left[ \tanh \frac{\beta \mu }{2}%
+1\right] ,\qquad {\it \Pi }_{{\it 0}}\,^{\prime }=-\frac{\beta }{48\pi }%
%TCIMACRO{\func{sech} }
%BeginExpansion
\mathop{\rm sech}%
%EndExpansion
\!^{2}\!\,\left( \frac{\beta \mu }{2}\right) ,\qquad {\it \Pi }_{{\it 1}}=%
\frac{\overline{b}}{m}{\it \Pi }_{{\it 0}}\,^{\prime },\qquad {\it \Pi }_{\,%
{\it 2}}=\frac{1}{12m^{2}}{\it \Pi }_{{\it 0}} \eqnum{3.9}$$
In these expressions $\mu $ is the chemical potential and $m=2m_{e}$ ($m_{e}$ is the electron mass). These results are in agreement with those found in refs.[@8],[@14].
MCS Linear Equations
--------------------
To find the response of the anyon fluid to an externally applied magnetic field, one needs to use the extremum equations derived from the effective action (3.1). This formulation is known in the literature as the self-consistent field approximation[@11]. In solving these equations we confine our analysis to gauge field configurations which are static and uniform in the y-direction. Within this restriction we take a gauge in which $A_{1}=a_{1}=0$.
The Maxwell and Chern-Simons extremum equations are respectively,
$$\partial _{\nu }F^{\nu \mu }=eJ_{ind}^{\mu } \eqnum{3.10a}$$
$$-\frac{N}{4\pi }\varepsilon ^{\mu \nu \rho }f_{\nu \rho }=J_{ind}^{\mu }
\eqnum{3.10b}$$
Here, $f_{\mu \nu }$ is the Chern-Simons gauge field strength tensor, defined as $f_{\mu \nu }=\partial _{\mu }a_{\nu }-\partial _{\nu }a_{\mu }$, and $J_{ind}^{\mu }$ is the current density induced by the anyon system at finite temperature and density. Their different components are given by
$$J_{ind}^{0}\left( x\right) ={\it \Pi }_{{\it 0}}\left[ a_{0}\left( x\right)
+eA_{0}\left( x\right) \right] +{\it \Pi }_{{\it 0}}\,^{\prime }\partial
_{x}\left( {\cal E}+eE\right) +{\it \Pi }_{{\it 1}}\left( b+eB\right)
\eqnum{3.11a}$$
$$J_{ind}^{1}\left( x\right) =0,\qquad J_{ind}^{2}\left( x\right) ={\it \Pi }_{%
{\it 1}}\left( {\cal E}+eE\right) +{\it \Pi }_{\,{\it 2}}\partial _{x}\left(
b+eB\right) \eqnum{3.11b}$$
in the above expressions we used the following notation: ${\cal E}=f_{01}$, $%
E=F_{01}$, $b=f_{12}$ and $B=F_{12}$. Eqs. (3.11) play the role in the anyon fluid of the London equations in BCS superconductivity. When the induced currents (3.11) are substituted in eqs. (3.10) we find, after some manipulation, the set of independent differential equations,
$$\omega \partial _{x}^{2}B+\alpha B=\gamma \left[ \partial _{x}E-\sigma
A_{0}\right] +\tau \,a_{0}, \eqnum{3.12}$$
$$\partial _{x}B=\kappa \partial _{x}^{2}E+\eta E, \eqnum{3.13}$$
$$\partial _{x}a_{0}=\chi \partial _{x}B \eqnum{3.14}$$
The coefficients appearing in these differential equations depend on the components of the polarization operators through the relations
$$\omega =\frac{2\pi }{N}{\it \Pi }_{{\it 0}}\,^{\prime },\quad \alpha =-e^{2}%
{\it \Pi }_{{\it 1}},\quad \tau =e{\it \Pi }_{{\it 0}},\quad \chi =\frac{%
2\pi }{eN},\quad \sigma =-\frac{e^{2}}{\gamma }{\it \Pi }_{{\it 0}},\quad
\eta =-\frac{e^{2}}{\delta }{\it \Pi }_{{\it 1}},$$
$$\gamma =1+e^{2}{\it \Pi }_{{\it 0}}\,^{\prime }-\frac{2\pi }{N}{\it \Pi }_{%
{\it 1}},\quad \delta =1+e^{2}{\it \Pi }_{\,{\it 2}}-\frac{2\pi }{N}{\it \Pi
}_{{\it 1}},\quad \kappa =\frac{2\pi }{N\delta }{\it \Pi }_{\,{\it 2}}.
\eqnum{3.15}$$
Distinctive of eq. (3.12) is the presence of the nonzero coefficients $%
\sigma $ and $\tau $. They are related to the Debye screening in the two dimensional anyon thermal ensemble. A characteristic of this 2+1 dimensional model is that the Debye screening disappears at $T=0$, even if the chemical potential is different from zero. Note that $\sigma $ and $\tau $ link the magnetic field to the zero components of the gauge potentials, $A_{0}$ and $%
a_{0}$. As a consequence, these gauge potentials will play a nontrivial role in finding the magnetic field solution of the system.
Field Solutions and Boundary Conditions
---------------------------------------
Using eqs.(3.12)-(3.14) one can obtain a higher order differential equation that involves only the electric field,
$$a\partial _{x}^{4}E+d\partial _{x}^{2}E+cE=0, \eqnum{3.16}$$
Here, $a=\omega \kappa $, $d=\omega \eta +\alpha \kappa -\gamma -\tau \kappa
\chi $, and $c=\alpha \eta -\sigma \gamma -\tau \eta \chi $.
Solving (3.16) we find
$$E\left( x\right) =C_{1}e^{-x\xi _{1}}+C_{2}e^{x\xi _{1}}+C_{3}e^{-x\xi
_{2}}+C_{4}e^{x\xi _{2}}, \eqnum{3.17}$$
where
$$\xi _{1,2}=\left[ -d\pm \sqrt{d^{2}-4ac}\right] ^{\frac{1}{2}}/\sqrt{2a}
\eqnum{3.18}$$
When the low density approximation $n_{e}\ll m^{2}$ is considered (this approximation is in agreement with the typical values $n_{e}=2\times
10^{14}cm^{-2}$, $m_{e}=2.6\times 10^{10}cm^{-1}$ found in high-T$_{C}$ superconductivity), we find that, for $N=2$ and at temperatures lower than the energy gap $\left( T\ll \omega _{c}\right) $, the inverse length scales (3.18) are given by the following real functions $$\xi _{1}\simeq e\sqrt{\frac{m}{\pi }}\left[ 1+\left( \frac{\pi ^{2}n_{e}^{2}%
}{m^{3}}\right) \beta \exp -\left( \frac{\pi n_{e}\beta }{2m}\right) \right]
\eqnum{3.19}$$ $$\xi _{2}\simeq e\sqrt{\frac{n_{e}}{m}}\left[ 1-\left( \frac{\pi n_{e}}{m}%
\right) \beta \exp -\left( \frac{\pi n_{e}\beta }{2m}\right) \right]
\eqnum{3.20}$$ These two inverse length scales correspond to two short-range electromagnetic modes of propagation. These results are consistent with those obtained in ref. [@8] using a different approach. If the masses of the two massive modes, obtained in ref. [@8] by using the electromagnetic thermal Green’s function for static and slowly varying configurations, are evaluated in the range of parameters considered above, it can be shown that they reduce to eqs. (319), (3.20).
The solutions for $B$, $a_{0}$ and $A_{0}$, can be obtained using eqs. (3.13), (3.14), (3.17) and the definition of $E$ in terms of $A_{0,}$
$$B\left( x\right) =\gamma _{1}\left( C_{2}e^{x\xi _{1}}-C_{1}e^{-x\xi
_{1}}\right) +\gamma _{2}\left( C_{4}e^{x\xi _{2}}-C_{3}e^{-x\xi
_{2}}\right) +C_{5} \eqnum{3.21}$$
$$a_{0}\left( x\right) =\chi \gamma _{1}\left( C_{2}e^{x\xi
_{1}}-C_{1}e^{-x\xi _{1}}\right) +\chi \gamma _{2}\left( C_{4}e^{x\xi
_{2}}-C_{3}e^{-x\xi _{2}}\right) +C_{6} \eqnum{3.22}$$
$$A_{0}\left( x\right) =\frac{1}{\xi _{1}}\left( C_{1}e^{-x\xi
_{1}}-C_{2}e^{x\xi _{1}}\right) +\frac{1}{\xi _{2}}\left( C_{3}e^{-x\xi
_{2}}-C_{4}e^{x\xi _{2}}\right) +C_{7} \eqnum{3.23}$$
In the above formulas we introduced the notation $\gamma _{1}=\left( \xi
_{1}^{2}\kappa +\eta \right) /\xi _{1}$, $\gamma _{2}=\left( \xi
_{2}^{2}\kappa +\eta \right) /\xi _{2}$.
In obtaining eq. (3.16) we have taken the derivative of eq. (3.12). Therefore, the solution of eq. (3.16) belongs to a wider class than the one corresponding to eqs. (3.12)-(3.14). To exclude redundant solutions we must require that they satisfy eq. (3.12) as a supplementary condition. In this way the number of independent unknown coefficients is reduced to six, which is the number corresponding to the original system (3.12)-(3.14). The extra unknown coefficient is eliminated substituting the solutions (3.17), (3.21), (3.22) and (3.23) into eq. (3.12) to obtain the relation
$$e{\it \Pi }_{{\it 1}}C_{5}=-{\it \Pi }_{{\it 0}}\left( C_{6}+eC_{7}\right)
\eqnum{3.24}$$
Eq. (3.24) has an important meaning, it establishes a connection between the coefficients of the long-range modes of the zero components of the gauge potentials $(C_{6}+eC_{7})$ and the coefficient of the long-range mode of the magnetic field $C_{5}$. Note that if the induced Chern-Simons coefficient ${\it \Pi }_{{\it 1}}$, or the Debye screening coefficient ${\it %
\Pi }_{{\it 0}}$ were zero, there would be no link between $C_{5}$ and $%
(C_{6}+eC_{7})$. This relation between the long-range modes of $B$, $A_{0}$ and $a_{0}$ can be interpreted as a sort of Aharonov-Bohm effect, which occurs in this system at finite temperature. At $T=0$, we have ${\it \Pi }_{%
{\it 0}}=0$, and the effect disappears.
Up to this point no boundary has been taken into account. Therefore, it is easy to understand that the magnetic long-range mode associated with the coefficient $C_{5}$, must be identified with the one found in ref. [@8] for the infinite bulk using a different approach. However, as it is shown below, when a constant and uniform magnetic field is perpendicularly applied at the boundaries of a two-dimensional sample, this mode cannot propagate (i.e. $C_{5}\equiv 0$) within the sample. This result is crucial for the existence of Meissner effect in this system.
In order to determine the unknown coefficients we need to use the boundary conditions. Henceforth we consider that the anyon fluid is confined to the strip $-\infty <y<\infty $ with boundaries at $x=-L$ and $x=L$. The external magnetic field will be applied from the vacuum at both boundaries ($-\infty
<x\leq -L$, $\;L\leq x<\infty $).
The boundary conditions for the magnetic field are $B\left( x=-L\right)
=B\left( x=L\right) =\overline{B}$ ($\overline{B}$ constant). Because no external electric field is applied, the boundary conditions for this field are, $E\left( x=-L\right) =E\left( x=L\right) =0$. Using them and assuming $%
L\gg \lambda _{1}$, $\lambda _{2}$ ($\lambda _{1}=1/\xi _{1}$, $\lambda
_{2}=1/\xi _{2}$), we find the following relations that give $C_{1,2,3,4}$ in terms of $C_{5}$,
$$C_{1}=Ce^{-L\xi _{1}},\quad C_{2}=-C_{1},\quad C_{3}=-Ce^{-L\xi _{2}},\quad
C_{4}=-C_{3},\quad C=\frac{C_{5}-\overline{B}}{\gamma _{1}-\gamma _{2}}
\eqnum{3.25}$$
Stability Condition for the Infinite-Strip Sample
=================================================
After using the boundary conditions, we can see from (3.25) that they were not sufficient to find the coefficient $C_{5}$. In order to totally determine the system magnetic response we have to use another physical condition from where $C_{5}$ can be found. Since, obviously, any meaningful solution have to be stable, the natural additional condition to be considered is the stability equation derived from the system free energy. With this goal in mind we start from the free energy of the infinite-strip sample
$${\cal F}=\frac{1}{2}\int\limits_{-L^{\prime }}^{L^{\prime
}}dy\int\limits_{-L}^{L}dx\left\{ \left( E^{2}+B^{2}\right) +\frac{N}{\pi }%
a_{0}b-{\it \Pi }_{{\it 0}}\left( eA_{0}+a_{0}\right) ^{2}\right.$$
$$\left. -{\it \Pi }_{{\it 0}}\,^{\prime }\left( eE+{\cal E}\right) ^{2}-2{\it %
\Pi }_{{\it 1}}\left( eA_{0}+a_{0}\right) \left( eB+b\right) +{\it \Pi }_{\,%
{\it 2}}\left( eB+b\right) ^{2}\right\} \eqnum{4.1}$$
where $L$ and $L^{\prime }$ determine the two sample’s lengths.
Using the field solutions (3.17), (3.21)-(3.23) with coefficients (3.25), it is found that the leading contribution to the free-energy density ${\it f}=%
\frac{{\cal F}}{{\cal A}}$ , (${\cal A}=4LL^{\prime }$ being the sample area) in the infinite-strip limit $(L\gg \lambda _{1}$, $\lambda _{2}$, $%
L^{\prime }\rightarrow \infty )$ is given by
$$f=C_{5}^{2}-{\it \Pi }_{{\it 0}}\left( C_{6}+eC_{7}\right) ^{2}+e^{2}{\it %
\Pi }_{\,{\it 2}}C_{5}^{2}-2e{\it \Pi }_{{\it 1}}\left( C_{6}+eC_{7}\right)
C_{5} \eqnum{4.2}$$
Taking into account the constraint equation (3.24), the free-energy density (4.2) can be written as a quadratic function in $C_{5}$. Then, the value of $%
C_{5}$ is found, by minimizing the corresponding free-energy density
$$\frac{\delta {\it f}}{\delta C_{5}}=\left[ {\it \Pi }_{{\it 0}}+e^{2}{\it %
\Pi }_{{\it 1}}^{\,}\,^{2}+e^{2}{\it \Pi }_{{\it 0}}{\it \Pi }_{\,{\it 2}%
}\right] \frac{C_{5}}{{\it \Pi }_{{\it 0}}}=0, \eqnum{4.3}$$
to be $C_{5}=0$.
This result implies that the long-range mode cannot propagate within the infinite-strip when a uniform and constant magnetic field is perpendicularly applied at the sample’s boundaries.
We want to point out the following fact. The same property of the finite temperature polarization operator component $\Pi _{00}$ that is producing the appearance of a long-range mode in the infinite bulk, is also responsible, when it is combined with the boundary conditions, for the non-propagation of this mode in the bounded sample. It is known that the nonvanishing of ${\it \Pi }_{{\it 0}}$ at $T\neq 0$ (or equivalently, the presence of a pole $\sim 1/k^{2}$ in $\Pi _{00}$ at $T\neq 0$) guarantees the existence of a long-range mode in the infinite bulk [@8]. On the other hand, however, once ${\it \Pi }_{{\it 0}}$ is different from zero, we can use the constraint (3.24) to eliminate $C_{6}+eC_{7}$ in favor of $C_{5%
\text{ }}$ in the free-energy density of the infinite strip. Then, as we have just proved, the only stable solution of this boundary-value problem, which is in agreement with the boundary conditions, is $C_{5}=0$. Consequently, no long-range mode propagates in the bounded sample.
In the zero temperature limit $\left( \beta \rightarrow \infty \right) $, because ${\it \Pi }_{{\it 0}}=0$, it is straightforwardly obtained from (3.24) that $C_{5}=0$ and no long-range mode propagates.
At $T\neq 0$, taking into account that $C_{5}=0$ along with eq. (3.25) in the magnetic field solution (3.21), we can write the magnetic field penetration as
$$B\left( x\right) =\overline{B}_{1}\left( T\right) \left( e^{-(x+L)\xi
_{1}}+e^{\left( x-L\right) \xi _{1}}\right) +\overline{B}_{2}\left( T\right)
\left( e^{-(x+L)\xi _{2}}+e^{\left( x-L\right) \xi _{2}}\right) \eqnum{4.4}$$
where,
$$\overline{B}_{1}\left( T\right) =\frac{\gamma _{1}}{\gamma _{1}-\gamma _{2}}%
\overline{B},\text{ \qquad \quad }\overline{B}_{2}\left( T\right) =\frac{%
\gamma _{2}}{\gamma _{2}-\gamma _{1}}\overline{B} \eqnum{4.5}$$
For densities $n_{e}\ll m^{2}$, the coefficients $\overline{B}_{1}$and $%
\overline{B}_{2}$ can be expressed, in the low temperature approximation $%
\left( T\ll \omega _{c}\right) $, as
$$B_{1}\left( T\right) \simeq -\frac{\left( \pi n_{e}\right) ^{3/2}}{m^{2}}%
\left[ 1/m+\frac{5}{2}\beta \exp -\left( \frac{\pi n_{e}\beta }{2m}\right)
\right] \overline{B},\qquad \eqnum{4.6}$$
$$B_{2}\left( T\right) \simeq \left[ 1+\frac{5\pi n_{e}}{2m^{2}}\sqrt{\pi n_{e}%
}\beta \exp -\left( \frac{\pi n_{e}\beta }{2m}\right) \right] \overline{B}
\eqnum{4.7}$$
Hence, in the infinite-strip sample the applied magnetic field is totally screened within the anyon fluid on two different scales, $\lambda _{1}=1/\xi
_{1}$ and $\lambda _{2}=1/\xi _{2}$. At $T=200K$, for the density value considered above, the penetration lengths are given by $\lambda _{1}\simeq
0.6\times 10^{-8}cm$ and $\lambda _{2}\simeq 0.1\times 10^{-4}cm$ . Moreover, taking into account that $\xi _{1}$ increases with the temperature while $\xi _{2}$ decreases (see eqs. (3.19)-(3.20)), and that $B_{1}\left(
T\right) <0$ while $B_{2}\left( T\right) >0$, it can be shown that the effective penetration length $\overline{\lambda }$ (defined as the distance $%
x$ where the magnetic field falls down to a value $B\left( \overline{\lambda
}\right) /\overline{B}=e^{-1}$) increases with the temperature as
$$\overline{\lambda }\simeq \overline{\lambda }_{0}\left( 1+\overline{\kappa }%
\beta \exp -\frac{1}{2}\overline{\kappa }\beta \right) \eqnum{4.8}$$
where $\overline{\lambda }_{0}=\sqrt{m/n_{e}e^{2}}$ and $\overline{\kappa }%
=\pi n_{e}/m$. At $T=200K$ the effective penetration length is $\overline{%
\lambda }\sim 10^{-5}cm$.
It is timely to note that the presence of explicit (proportional to $N$) and induced (proportional to ${\it \Pi }_{{\it 1}}$) Chern-Simons terms in the anyon effective action (3.1) is crucial to obtain the Meissner solution (4.4). If the Chern-Simons interaction is disconnected ($N\rightarrow \infty
$ and ${\it \Pi }_{{\it 1}}=0$), then $a=0,$ $d=1+e^{2}{\it \Pi }_{{\it 0}%
}{}^{\prime }\neq 0$ and $c=e^{2}{\it \Pi }_{{\it 0}}\,\neq 0$ in eq. (3.16). In that case the solution of the field equations within the sample are $E=0$, $B=\overline{B}$. That is, we regain the QED in (2+1)-dimensions, which does not exhibit any superconducting behavior.
High Temperature Non-Superconducting Phase
==========================================
We have just found that the charged anyon fluid confined to an infinite strip exhibits Meissner effect at temperatures lower than the energy gap $%
\omega _{c}$. It is natural to expect that at temperatures larger than the energy gap this superconducting behavior should not exist. At those temperatures the electron thermal fluctuations should make available the free states existing beyond the energy gap. As a consequence, the charged anyon fluid should not be a perfect conductor at $T\gg \omega _{c}$. A signal of such a transition can be found studying the magnetic response of the system at those temperatures.
As can be seen from the magnetic field solution (4.4), the real character of the inverse length scales (3.18) is crucial for the realization of the Meissner effect. At temperatures much lower than the energy gap this is indeed the case, as can be seen from eqs. (3.19) and (3.20).
In the high temperature $\left( T\gg \omega _{c}\right) $ region the polarization operator coefficients are given by eq. (3.9). Using this approximation together with the assumption $n_{e}\ll m^{2}$, we can calculate the coefficients $a$, $c$ and $d$ that define the behavior of the inverse length scales,
$$a\simeq \pi ^{2}{\it \Pi }_{{\it 0}}{}^{\prime }{\it \Pi }_{\,{\it 2}}
\eqnum{5.1}$$
$$c\simeq e^{2}{\it \Pi }_{{\it 0}}{} \eqnum{5.2}$$
$$d\simeq -1 \eqnum{5.3}$$
Substituting with (5.1)-(5-3) in eq. (3.18) we obtain that the inverse length scales in the high-temperature limit are given by
$$\xi _{1}\simeq e\sqrt{m/2\pi }\left( \tanh \frac{\beta \mu }{2}+1\right) ^{%
\frac{1}{2}} \eqnum{5.4}$$
$$\xi _{2}\simeq i\left[ 24\sqrt{\frac{2m}{\beta }}\cosh \frac{\beta \mu }{2}%
\left( \tanh \frac{\beta \mu }{2}+1\right) ^{-\frac{1}{2}}\right]
\eqnum{5.5}$$
The fact that $\xi _{2}$ becomes imaginary at temperatures larger than the energy gap, $\omega _{c}$, implies that the term $\gamma _{2}\left(
C_{4}e^{x\xi _{2}}-C_{3}e^{-x\xi _{2}}\right) $ in the magnetic field solution (3.21) ceases to have a damping behavior, giving rise to a periodic inhomogeneous penetration. Therefore, the fluid does not exhibit a Meissner effect at those temperatures since the magnetic field will not be totally screened. This corroborate our initial hypothesis that at $T\gg \omega _{c}$ the anyon fluid is in a new phase in which the magnetic field can penetrate the sample.
We expect that a critical temperature of the order of the energy gap ($T\sim
\omega _{c}$) separates the superconducting phase $\left( T\ll \omega
_{c}\right) $ from the non-superconducting one $\left( T\gg \omega
_{c}\right) $. Nevertheless, the temperature approximations (3.8) and (3.9) are not suitable to perform the calculation needed to find the phase transition temperature. The field solutions in this new non-superconducting phase is currently under investigation. The results will be published elsewhere.
Concluding Remarks
==================
In this paper we have investigated the magnetic response at finite temperature of a charged anyon fluid confined to an infinite strip. The charged anyon fluid was modeled by a (2+1)-dimensional MCS theory in a many-particle ($\mu \neq 0$, $\overline{b}\neq 0$) ground state. The particle energy spectrum of the theory exhibits a band structure given by different Landau levels separated by an energy gap $\omega _{c}$, which is proportional to the background Chern-Simons magnetic field $\overline{b}$. We found that the energy gap $\omega _{c}$ defines a scale that separates two phases: a superconducting phase at $T\ll \omega _{c}$, and a non-superconducting one at $T\gg \omega _{c}$.
The total magnetic screening in the superconducting phase is characterized by two penetration lengths corresponding to two short-range eigenmodes of propagation of the electromagnetic field within the anyon fluid. The existence of a Meissner effect at finite temperature is the consequence of the fact that a third electromagnetic mode, of a long-range nature, which is present at finite temperature in the infinite bulk [@8], does not propagate within the infinite strip when a uniform and constant magnetic field is applied at the boundaries. This is a significant property since the samples used to test the Meissner effect in high-$T_{c}$ superconductors are bounded.
It is noteworthy that the existence at finite temperature of a Debye screening (${\it \Pi }_{{\it 0}}\,\neq 0$) gives rise to a sort of Aharonov-Bohm effect in this system with Chern-Simons interaction ($N$ finite, ${\it \Pi }_{{\it 1}}\neq 0$). When ${\it \Pi }_{{\it 0}}\,\neq 0$, the field combination $a_{0}+eA_{0}$ becomes physical because it enters in the field equations in the same foot as the electric and magnetic fields (see eq. (3.12)). A direct consequence of this fact is that the coefficient $%
C_{5}$, associated to the long-range mode of the magnetic field, is linked to the coefficients $C_{6}$ and $C_{7}$ of the zero components of the potentials (see eq. (3.24)).
When $T=0$, since ${\it \Pi }_{{\it 0}}\,=0$ and ${\it \Pi }_{{\it 1}}\neq 0$, eq. (3.24) implies $C_{5}=0$. That is, at zero temperature the long-range mode is absent. This is the well known Meissner effect of the anyon fluid at $T=0$. When $T\neq 0$, eq. (3.24) alone is not enough to determine the value of $C_{5}$, since it is given in terms of $C_{6}$ and $C_{7}$ which are unknown. However, when eq. (3.24) is taken together with the field configurations that satisfy the boundary conditions for the infinite-strip sample (eqs. (3.17), (3.21)-(3.23) and (3.25)), and with the sample stability condition (4.3), we obtain that $C_{5}=0$. Thus, the combined action of the boundary conditions and the Aharonov-Bohm effect expressed by eq. (3.24) accounts for the total screening of the magnetic field in the anyon fluid at finite temperature.
Finally, at temperatures large enough ($T\gg \omega _{c}$) to excite the electrons beyond the energy gap, we found that the superconducting behavior of the anyon fluid is lost. This result was achieved studying the nature of the characteristic lengths (3.18) in this high temperature approximation. We showed that in this temperature region the characteristic length $\xi _{2}$ becomes imaginary (eq. (5.5)), which means that a total damping solution for the magnetic field does not exist any more, and hence the magnetic field penetrates the sample.
> Acknowledgments
The authors are very grateful for stimulating discussions with Profs. G. Baskaran, A. Cabo, E.S. Fradkin, Y. Hosotani and J. Strathdee. We would also like to thank Prof. S. Randjbar-Daemi for kindly bringing the publication of ref. [@b] to our attention. Finally, it is a pleasure for us to thank Prof. Yu Lu for his hospitality during our stay at the ICTP, where part of this work was done. This research has been supported in part by the National Science Foundation under Grant No. PHY-9414509.
R. D. Pisarski, Phys. Rev. [**D**]{} [**29**]{} (1984) 2423; T. W. Appelquist, M. J. Bowick, D. Karabali and L. C. R. Wijewardhana, Phys. Rev. [**D**]{} [**33**]{} (1986) 3704; D. Nash, Phys. Rev. Lett. [**62**]{} (1989) 3024; M. R. Pennington and D. Walsh, Phys. Lett. [**B 253**]{} (1991) 246; M. C. Diamantini, P. Sodano and G. W. Semenoff, Phys. Rev. Lett. [**70**]{} (1993) 3848.
K. Klitzing, G. Dorda and M. Pepper, Phys. Rev. Lett. [**45**]{} (1980) 494; D. C. Tsui, H. L. Strömer and A. C. Gossard, Phys. Rev. Lett. [**48**]{} (1982) 1559; K. Ishikawa, Phys. Rev. [**D**]{} [**31**]{} (1985) 1432; J. K. Jain, Phys. Rev. Lett. [**63**]{} (1989) 199; for a comprehensive reference see [*The Quantum Hall Effect,* ]{}edited by R. Prange and S. Girvin (Springer-Verlag, Berlin, 1987)[* *]{}and[* *]{}T. Chakraborty and P. Pietiläinen, [*The Fractional Quantum Hall Effect* ]{}(Springer-Verlag, Berlin, 1988).
Y. Hosotani, Phys. Lett [**B 319**]{} (1993) 199; Phys. Rev. [**D 51**]{} (1995) 2022; S. Kanemura and T. Matsushita, [*“Thermodynamics in 2+1 Dimensional QED with Chern-Simons Term,”* ]{}Osaka University preprint: [*OU-HET 212, hep-th/9505146 (1995).*]{}
See, for example, E. Abdalla, M. C. B. Abdalla and K. Rothe, [*Non-Perturbative Methods in Two-Dimensional Quantum Field Theory*]{}, (World Scientific, Singapore, 1985); E. Abdalla and M. C. B. Abdalla, Phys. Rep [**265**]{} (1996) 253.
J. M. Leinaas and J. Myrheim, Nuovo Cimento [**36b**]{} (1977) 1; G. A. Goldin, R. Menikoff and D. H. Sharp, J. Math. Phys. [**22**]{} (1981) 1664; Phys. Rev. [**D 28**]{} (1983) 830; F. Wilczek, Phys. Rev. Lett. [**48**]{} (1982) 1144; [**49**]{} (1982) 957.
P. W. Anderson, in: Physics of low-dimensional systems, Proc. Nobel symp. 73, eds. S. Lundquist and N. R. Nilsson (North Holland, Amsterdam, 1989); P. B. Wiegmann, Phys. Rev. Lett. [**60**]{} (1988) 821; I. Affleck and B. Marston, Phys. Rev. [**B 37**]{} (1988) 3774; X.G. Wen, F. Wilczek and A. Zee, Phys. Rev. [**B 39**]{}, (1989) 11413.
V. Kalmeyer and R.B. Laughlin, Phys. Rev. Lett. [**59**]{}, (1987) 2095.
R.B. Laughlin, Phys. Rev. Lett. [**60**]{}, (1988) 2677.
A.L. Fetter, C.B. Hanna and R.B. Laughlin, Phys. Rev. [**B 39**]{}, (1989) 9679.
D. P. Arovas, J. R. Schrieffer, F. Wilczek, and A. Zee, Nucl. Phys..[** B 251**]{} (1985) 117; A. Goldhaber, R. MacKenzie and F. Wilczek, Mod. Phys. Lett.[** A 4**]{} (1989) 21.
G. W. Semenoff and N. Weiss, Phys. Lett. [**B 250**]{} (1990) 117; N. Dorey and N. E. Mavromatos Phys. Lett. [**B 250**]{} (1990) 107; Nucl. Phys. [**B 386**]{} (1992) 614; Z. F. Ezawa and A. Iwazaki, Mod. Phys. Lett. [**B 7**]{} (1993) 109; J. L. Goity and J. Soto, Int. J. Mod. Phys. [**A 7**]{} (1993) 4595.
B. I. Halpering, J. March-Russel and F. Wilczek, Phys. Rev. [**B 40**]{} (1989) 8726.
Y.-H Chen, F. Wilczek, E. Witten and B. I. Halperin, Int. J. Mod. Phys. [**B3**]{} (1989) 1001.
X.G. Wen and A. Zee, Phys. Rev. [**B 41**]{} (1990) 240.
C. B. Hanna, R. B. Laughlin and A. L. Fetter, Phys. Rev. [**B 40**]{} (1989) 8745, [*ibid.* ]{}[**43**]{} (1991) 309; G. S. Canright, S. M. Girvin and A. Brass, Phys. Rev. Lett. [**63**]{} (1989) 2291, 2295; E. Fradkin, Phys. Rev. Lett. [**63**]{} (1989) 322, Phys. Rev. [**B 42**]{} (1990) 570; Y. Hosotani and S. Chakravarty, Phys. Rev. [**B 42**]{} (1990) 342; Phys. Rev. [**D 44**]{} (1991) 441; A. L. Fetter, and C. B. Hanna, Phys. Rev. [**B 45**]{} (1992) 2335.
S. Randjbar-Daemi, A. Salam and J. Strathdee, Nucl. Phys. [**B 340**]{} (1990) 403.
J.D. Lykken, J. Sonnenschein and N. Weiss, Phys. Rev. [**D 42**]{} (1990) 2161; Int. J. Mod. Phys. [**A 6**]{} (1991) 1335.
Y. Hosotani, Int. J. Mod. Phys. [**B7**]{} (1993) 2219; J.E. Hetric, Y. Hosotani and B.-H Lee, Ann. Phys [**209**]{} (1991) 151; J.E. Hetric and Y. Hosotani, Phys. Rev. [**B45**]{} (1992) 2981.
P.K. Panigrahi, R. Ray and B Sakita, Phys. Rev. [**B42**]{} (1990) 4036; J. Kapusta, M.E. Carrington, B. Bayman, D. Seibert and C.S. Song, Phys. Rev. [**B44**]{} (1991) 7519; Y. Georgelin, M. Knecht, Y. Leblanc, and J.C. Wallet, Mod. Phys. Lett. [**B5**]{} (1991) 211; Y. Leblanc, and J.C. Wallet, Mod. Phys. Lett. [**B6**]{} (1992) 1623; I.E. Aronov, E.N. Bogachek, I.V. Krive and S.A. Naftulin, JETP Lett. [**56**]{} (1992) 283; Y. Kitazawa and H. Murayama, Phys. Rev. [**B41**]{} (1990) 11101.
S.S. Mandal, S. Ramaswamy and V. Ravishankar; Mod. Phys. Lett. [**B8**]{} (1994) 561, Int. J. Mod. Phys. [**B8**]{} (1994) 3095.
E.J. Ferrer, R. Hurka and V. de la Incera, Mod. Phys. Lett. [**B11**]{} (1997) 1.
S. Randjbar-Daemi, A. Salam and J. Strathdee, Int. J. Mod. Phys. [**B5**]{} (1991) 845.
E. J. Ferrer and V. de la Incera, Int. J. Mod. Phys. [**B5**]{} (1995) 3585.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'At the beginning of the 4 November 2015 flare, in the 1300 – 2000 MHz frequency range, we observed a very rare slowly positively drifting burst. We searched for associated phenomena in simultaneous EUV observations made by IRIS, SDO/AIA, Hinode/XRT and in H$_\alpha$ observations. We found that this radio burst was accompanied with the bright blob, visible at transition region, coronal, and flare temperatures, falling down to the chromosphere along the dark loop with the velocity of about 280 km s$^{-1}$. The dark loop was visible in H$_\alpha$ but disappeared afterwards. Furthermore, we found that the falling blob interacted with the chromosphere as expressed by a sudden change of the H$_\alpha$ spectra at the location of this interaction. Considering different possibilities we propose that the observed slowly positively drifting burst is generated by the thermal conduction front formed in front of the falling hot EUV blob.'
author:
- Marian Karlický
- Alena Zemanová
- Jaroslav Dudík
- Krzysztof Radziszewski
title: SOLAR RADIO BURST ASSOCIATED WITH THE FALLING BRIGHT EUV BLOB
---
INTRODUCTION
============
It is commonly known that a plasma in solar flares is rapidly heated, e.g. by the magnetic field reconnection. In single loop models of solar flares or in loops of the flare arcade the plasma may be heated at tops of the loops. Such a plasma would expand along the loop to its footpoints. This plasma is collisionless, therefore a question arises if this plasma expansion is a free expansion of plasma particles without any interactions or the expansion generates some plasma waves making thus an obstacle for the free expansion.
[@Brown79] in their pioneering work answered this question and proposed the so called thermal conduction front formed between expanding hot and cold plasmas. Their idea was supported by observations made by [@Farnik83; @Rust85; @Mandrini96].
From that time several attempts to simulate the thermal front numerically have been made. The results of [@McKean90a; @McKean90b], based on the 1-D electrostatic particle-in-cell simulations, did not confirm an existence of the thermal front. They only showed a free expansion of electrons from the hot plasma region to the colder one. On the other hand, [@Arber09], using a Vlasov code, showed that a thermal front trapping of hot electrons can be really generated.
Recently, [@2015ApJ...814..153K] has studied a formation of the thermal front using a 3-D particle-in-cell fully electromagnetic code. He recognized the thermal front and importance of inductive effects for its formation. The thermal front was propagating with the sound speed of the hot plasma. The thermal front was associated not only with the strong ion-acoustic waves, but also with the plasma and electromagnetic ones. Just the electromagnetic waves at the thermal front led to an idea that the thermal front can be observed on radio waves.
In this article, we present a very rare example of the slowly positively drifting burst which we propose to be the radio emission of the thermal conduction front. Such an interpretation is supported by simultaneous EUV and H$_\alpha$ observations.
OBSERVATIONS
============
The studied phenomenon was observed at the beginning of the 4 November 2015 flare. The flare was classified as GOES M3.7 and started at 13:31UT, peaked at 13:52UT, and ended at 14:13UT and occurred in the active region NOAA 12443, located at this time close to the center of the solar disk.
Global temporal profiles of the flare in *GOES* 1-8 and 0.5-4 Å soft X-rays and their derivatives together with the radio flux observed by the Ondřejov radiometer [@JirickaK08] on 3 GHz are presented in Figure \[fig1\]. For global overview of this flare, see movies attached to the paper by [@Li2017], where they studied evolution of flare ribbons.
Radio Spectral Observations
---------------------------
Figure \[fig2\] shows the 800 – 2000 MHz radio spectrum observed by the Ondřejov radiospectrograph [@JirickaK08] at the beginning of the solar flare in the time interval 13:39:50 – 13:40:49 UT. As seen here, in the 900 – 1300 MHz frequency range and at 13:39:55 UT the flare radio emission started with the drifting pulsation structure [@Jiricka01], which drifted with the frequency drift -18 MHz s$^{-1}$. This drifting pulsation structure is then at about 13:40:20 UT followed by decimetric type III bursts, indicating the electron beams propagating upwards through the flare atmosphere. Their frequency drift is very high, greater than 3 GHz s$^{-1}$.
The most interesting and very rare burst in this radio spectrum is the slowly positively drifting burst (SPDB). It suddenly started at 1300 MHz, at 13:40:24 UT and drifted towards higher frequencies with the frequency drift of about 115 MHz s$^{-1}$. Its intensity was fading at higher frequencies. The drift indicates that this burst is generated by the plasma emission mechanism. Thus assuming that the frequency of SPDB corresponds to plasma frequency then the density in the radio source is $\geq$ 2.08 $\times$ 10$^{10}$ cm$^{-3}$. Note that SPDB was observed during the maximum peak of *GOES* 1-8 and 0.5-4 Å soft X-rays derivatives (Figure \[fig1\], part b).
After this burst, the Ondřejov and e-Callisto [@2013EGUGA..15.2027M] radio spectra show continua with fiber bursts and type II radio burst (signature of the flare shock) in the 45 – 80 MHz frequency range at 13:42:00 – 13:54:00 UT.
EUV Imaging Observations
------------------------
Using observations made by the Atmospheric Imaging Assembly [AIA, @Lemen12] onboard the Solar Dynamics Observatory [SDO, @Pesnell12], the X-ray Telescope [@Golub2007] onboard of Hinode satellite [@Kosugi07] and the Interface Region Imaging Spectrograph [IRIS, @DePontieu14] we searched for any EUV signature associated with the slowly positively drifting radio burst. The radio burst was very short (about 9s), so it was challenging to observe its counterpart, considering cadences of UV/EUV imaging observations. In search of the EUV counterpart we used the following criteria given by SPDB properties: a) the densities derived from SPDB should be $\geq$ 2.08 $\times$ 10$^{10}$ cm$^{-3}$, such densities are commonly observed in EUV [@Petkaki2012] and b) the EUV counterpart has to be observed at time of SPDB and with time duration at least as SPDB, and c) the positive frequency drift of SPDB indicates that the agent generating SPDB moves to higher densities, which means in the gravitationally stratified atmosphere a downward motion. Searching in all flare area at time of SPDB, we found one distinct bright blob that meets the above criteria. This bright blob is falling down to the chromosphere and is observed by IRIS, AIA, and XRT instruments.
Figure \[fig3\] contains images from AIA 131Å and 94Å filters, combined with one panel observed in XRT Be\_med filter. We observed the bright plasma blob (white arrows) falling along the dark loop structure within the bright flare arcade. The flare emission in AIA 94Å and 131Å originates at temperatures log($T$ \[K\])=6.85 and 7.15, respectively but the contribution form coronal plasma at log($T$ \[K\])$\sim$5.45 - 6.05 is present as well [@ODwyer2010; @DelZanna2013]. The Be\_med filter is sensitive to plasma at log($T$ \[K\])$\sim$6.5 - 6.9 [@ODwyer2014] while the coronal emission below log($T$ \[K\])$\lesssim$6.1 is suppressed [@Narukage2014]. Simultaneous observations in AIA 94, 131Å and XRT Be\_med filters point to a hot component within the blob, likely with log($T$ \[K\])$\geq$6.9.
Figure \[fig4\] shows the transition region and coronal counterpart of the hot emission in Figure \[fig3\]. This figure contains the IRIS 1400Å slit-jaw images (SJI) combined with AIA 211Å and AIA 304Å filter. There are no IRIS spectra of the blob since the slit did not reach it.
The 1400Å IRIS filter is dominated by transition region lines of formed at log($T$ \[K\])$\sim$4.9 [@Dudik14] but contains also photospheric continuum emission. The AIA 304Å filter is dominated by optically thick Ly-$\alpha$ line at 303.8Å formed at around log($T$ \[K\])$\approx$4.9 [@Dere1997; @DelZanna2015].
The emission at coronal temperatures should be seen by the AIA filters 171, 193, and 211Å. The 171 and 193Å observations were overexposed, thus we present AIA 211Å filter observations only. The emission seen in this passband is not entirely understood [@DelZanna2011; @DelZanna2013]. The coronal contribution of 211.3Å line at log($T$ \[K\])=6.3 can be present also under the flare conditions but the most significant contribution (in flare) comes from continuum emission [@ODwyer2010]. Thus a coronal emission is very likely present within the blob.
At transition region temperatures ribbons are well visible. The blob is also visible in AIA 304Å and 211Å observations (white arrows) and in IRIS 1400Å SJI. Yellow arrows in the first two panels of 304Å images show the position of dark loop structure along which the bright blob falls.
All these observations show that the falling bright blob has a multi-temperature structure spanning within two orders of magnitude in temperature.
We tried to estimate velocity of the falling blob in the plane of image from AIA filters in Figures \[fig3\] and \[fig4\]. The center position of the blob was marked manually at each image. From these positions and the cadence we estimated velocity, which is 279(+31/-33) km s$^{-1}$ during 13:40:06 – 13:40:24UT, when the blob moved through the largest arc of the dark loop.
The accompanying movie in 304Å filter shows that during 13:39:00UT - 13:40:00UT, the SPDB connected falling blob is preceded by other, but less bright blobs falling along the wider dark loop structure but at slightly different trajectories, perhaps along neighboring dark threads. They move on both sides of the dark loop and are immediately followed by the middle one, connected with SPDB. After this blob the dark loop disappeared. Only the blob connected with SPDB was visible in XRT Be\_med and Al\_thick filters indicating the highest temperature in this blob. This type of blobs resembles the flare-driven coronal rain clumps [@Scullion16]. However, our blobs are observed at the impulsive phase of the flare, contrary to the rain clumps that are observed at the end of the flare [@Scullion16]. The velocity of our blob is faster than that of the rain clumps (several tens km s$^{-1}$) and the blob temperature is higher than in rain clumps ($\sim$ 22000 K - $\sim$ 1 MK). In this sense, our observed blobs are unique, and it is unlikely that they occur due to coronal rain.
H$_\alpha$ Observations
-----------------------
The studied flare was also observed in the H$_\alpha$ (6562.8Å) by the Multi-channel Subtractive Double Pass (MSDP) imaging spectrograph and Large Coronograph (LC) at the Bia[ł]{}k[ó]{}w Observatory, Poland [@Mein91; @Romp94].
On November 4th, 2015 LC worked with artificial moon removed, enabling observations of the solar surface. Between 13:12UT and 14:15UT, 110 scans of NOAA 12443 were collected. The 2D spectra-images obtained from LC-MSDP system have pixel size $\sim$0.5” and the spatial resolution of LC is limited by seeing conditions to $\approx$ 1”. The nine-channel MSDP prism-box enables recording spectra-images at nine positions across the H$_\alpha$ line profile, with a total range of $\pm$ 1.6Å. After a standard dark current and flat-field reduction, two-dimensional, quasi-monochromatic images with pixel band-width of 0.06Å, separated by 0.2[Å]{} up to $\pm$ 1.2[Å]{} from the H$_\alpha$ line center are restored [@Rad06; @Rad07]. Thus the H$_\alpha$ profile in the range of $\pm$1.2[Å]{} from the line center is available for each pixel within the FOV.
We have studied H$_\alpha$ observations to look for any optical counterpart of radio/EUV event described in previous sections. Figure \[fig5\] shows two MSDP images in the center of H$_\alpha$ line at times close to that of SPDB. In the first panel the white arrow shows the position of the bright plasma blob visible at the same time (13:40:12UT) in AIA 94Å. There is no counterpart of the blob visible in H$_\alpha$ in this position at this time. Instead, a dark loop structure in H$_\alpha$ (log($T$ \[K\])$\approx$4) is already visible since about 13:35UT (see online movie in H$_\alpha$). It quickly evolved from a wider structure into narrow fibre at about 13:40:12UT (Figure \[fig5\], the first panel), and finally disappeared. A bright knot appeared in its place at about 13:40:42UT (Figure \[fig5\], the second panel), i.e., at the time when the UV/EUV blob reached the ribbon.
The H$_\alpha$ profiles (Figure \[fig5\], last panel) were taken from the same position, i.e. from the central part of the dark loop structure at 13:40:12UT and the bright knot at 13:40:42UT, marked by the white crosses in Figure \[fig5\]. The dark loop structure at 13:40:12UT is characterized by a small redshift corresponding to velocity of +11 km s$^{-1}$. After 30s, in the same position, the bright knot was visible, whose H$_\alpha$ profile was significantly increased. In red wing there is clear local maximum of intensity, which corresponds to plasma moving downward with velocity of +41 km s$^{-1}$ - i.e. almost three times more than earlier. Additionally, in blue wing, the smaller local maximum was visible, which implied the second component of upward moving plasma with velocity of -35 km s$^{-1}$. Such a profile suggests presence of both direction of plasma motion in the bright knot but with dominating downward component.
DISCUSSION AND CONCLUSIONS
==========================
The slowly positively drifting burst (SPDB) was observed at the beginning of solar flare and its drift (115 MHz s$^{-1}$) was much smaller than that of following decimetric type III bursts (several GHz s$^{-1}$). Because type III bursts are generated by electron beams with the velocity of about one third of speed of light, SPDB has to be generated by an agent having much smaller velocity.
The presented EUV observations show that at the time of SPDB the bright blob was propagating along the dark loop, visible in H$_\alpha$, downwards to the chromosphere. An interaction of this falling blob with the chromosphere was expressed by a sudden change of the H$_\alpha$ line spectrum, which shows a strong plasma heating and enhanced plasma velocities.
The fall velocity of the EUV blob is estimated as about 280 km s$^{-1}$. It is much smaller than the Alfvén speed at such atmospheric altitudes (about 1000 km s$^{-1}$), but greater than the velocities of flare-driven coronal rain clumps. As argued in Section 2.2, our blob is not due to the coronal rain phenomenon.
Based on these observations, firstly, we considered the model with the magnetic island (plasmoid) moving downwards. Plasmoids are observed in the flare impulsive phase. But this explanation is not very probable because the plasmoid has to move along the current sheet and no indications of the straight or distorted current sheet were found in and around the loop where the EUV blob was moving. Moreover, such a plasmoid would be magnetically open in the direction of the magnetic field lines, i.e., in the direction of the loop and thus it would need some structure like the thermal front to be thermally isolated from cold plasma. We also considered some dissipation process along the loop with the twisted magnetic field lines. But in this case the dissipation spreads with the Alfvén velocity that is higher than the observed one. The problem with the thermal isolation in the direction of the loop axis would be the same as in the plasmoid case. Thus on the results obtained by numerical simulations [@2015ApJ...814..153K], we propose that SPDB is generated by the thermal conduction front formed in front of the falling bright and hot blob. Namely, the thermal front can generate the electromagnetic (radio) waves by plasma emission mechanism.
In such an interpretation, the velocity of the thermal front corresponds to the the velocity of the falling blob and also to the sound speed of the hot plasma at the bottom edge of the falling blob. If we take the velocity of the falling blob (280 km s$^{-1}$) as the sound velocity, the corresponding plasma temperature is about 3.5 MK (log($T$ \[K\])$\approx$6.54.) The bright blob was observed in a broad range of temperatures from those in the transition region through coronal ones up to flare temperatures. But this temperature is lower than the temperature of the EUV lines dominating in 94Å and 131Å filters ($\approx$ 7 MK, i.e., log($T$ \[K\])$\approx$ 6.85). According to the theory the velocity of the thermal front corresponds to the sound velocity of the hot plasma close to the thermal front. Here the temperature is lower than that in central part of the bright blob, because the most energetic electrons from this region freely escape through the thermal front to the region of cold plasma and propagate towards the chromosphere. Just the flux of these energetic electrons can explain the maximum peak in the derivatives of *GOES* 1-8 and 0.5-4 Å soft X-rays (peak in hard X-ray emission considering the Neupert effect) (Figure \[fig1\] b) at the time of SPDB observation.
SPDB drifted from 1300 MHz to 2000 MHz during about 6 seconds. If in accordance with the assumed plasma emission mechanism of SPDB we assume that the radio frequencies correspond to the plasma frequencies, it means that the electron plasma density in front of the thermal front changed from 2.08 $\times$ 10$^{10}$ cm$^{-3}$ to 4.94 $\times$ 10$^{10}$ cm$^{-3}$ during 6 seconds.
Because there is no reliable density model for this case, in the following we consider the density dependance in the gravitationally stratified solar atmosphere: $n = n_0 \exp (-h/H)$, where $h$ is the height in the solar atmosphere and $H$ is the scale height (H\[m\] = 50 T\[K\]). The hot blob with the velocity 280 km s$^{-1}$ moves distance 1680 km during 6 seconds. If we assume that this motion is in vertical direction, we can estimate the scale height as $H$ = 1950 km, which gives the temperature in front of the thermal conduction front (in front of the falling blob) 39000 K. However, the blob falls along the cold loop structure, which is not vertical, therefore the scale height is probably shorter and temperature is smaller and thus to be in agreement with observations of the H$_\alpha$ cold loop (the temperature up to about 20000 K). However, we cannot exclude that the density gradient in this cold loop differs from that given by the gravitational equilibrium because plasma processes at this stage are very dynamic.
In the region of the thermal front the plasma waves were generated and then converted to the electromagnetic (radio) waves observed as SPDB. In this complex process there are several possibilities how to explain the sudden start and fading of SPDB. The most probable reason for this sudden start is sudden formation of the thermal front. Because the collisional optical depth for the plasma emission process increases as the second power of the radio frequency [@Benz93], the fading of SPDB towards higher frequencies can be explained by this optical depth increase. Nevertheless, there are other possibilities like a change of the level of the plasma waves and so on. In this emission process the so called ducting [@Benz93] probably plays a role, because the blob has a multithermal structure.
Just before the analyzed falling blob (which was associated with SPDB) we observed other falling blobs with no radio counterparts. But they were less bright and not visible in XRT Be\_med and Al\_thick filters. Note that for more bright and hotter blob the association with the radio burst and also formation of the thermal front is more probable. We think that in other cases maybe no thermal fronts were formed or the radio emission was very weak, because this emission is very sensitive to plasma conditions in the thermal front.
We acknowledge support from Grants 16-13277S and 17-16447S of the Grant Agency of the Czech Republic. K.R. was supported by the NCN grant no. UMO-2015/17/B/ST9/02073.
Arber, T. D., & Melnikov, V. F. 2009, , 690, 238
Bárta, M., Vr[š]{}nak, B., & Karlický, M., 2008, , 477, 649
Benz, A. O., 1993, Kinetic processes in solar and stellar coronae, Astrophysics and Space Science Library, Vol.184, Kluwer Acad. Publ., Dordrecht, The Netherlands, p. 263 and 277
Boerner, P., Christopher, E., Lemen, J. et al. 2012, , 275, 41
Brown, J.C., Melrose, D.B., & Spicer, D.S. 1979, , 228, 592
De Pontieu, B., Title, A.M., Lemen, J.R. et al. 2014, , 289, 2733
Del Zanna, G., Dere, K.P., Young, P.R. et al., , 582, 12
Del Zanna, G. 2013, , 558, A73
Del Zanna, G.,O’Dwyer, B., Mason, H.E. 2011, , 535, A46
Dere, K. P., Landi, E., Mason, H. E. et al. 1997, , 125,149
Dudík, J., Del Zanna, G., Dzifčáková, E. et al. 2014, , 780, L12
Fárník, F., Kaastra, J., Kálmán, B., Karlický, M., Slottje, C., & Valníček, B. 1983, , 89, 355
Golub, L., DeLuca, E., Austin, G. et al. 2007, , 243, 63
Jiřička, K., Karlick[ý]{}, M., [M[é]{}sz[á]{}rosov[á]{}]{}, H., & Sn[í]{}žek, V., 2001, , 375, 243
Jiřička, K. & Karlický, M. 2008, , 253, 95
, M. 2015, , 814, 153
Kosugi, T., Matsuzaki, K., Sakao, T. et al. 2007, , 243, 3
Lemen, J.R., Title, A.M., Akin, D.J. et al. 2012, , 275, 17
Li T., Zhang, J. and Hou, Y. 2017, , 848, 32. Mandrini, C. H., Demoulin, P., van Driel-Gesztelyi, L., Schmieder, B., Cauzzi, G., & Hofmann, A. 1996, , 168, 115
McKean, M.E., Winglee, R.M., & Dulk, G.A. 1990a, , 364, 295
McKean, M.E., Winglee, R.M., & Dulk, G.A. 1990b, , 364, 302
Mein, P., 1991, A&A, 248, 669
, C. 2013, EGU General Assembly Conference Abstracts 15, id. EGU2013-2027
Sakao, T., Kano, R., Shimojo, M., Winebarger, A. Weber, M., & Reeves, K. K. 2014, , 289, 1029
O’Dwyer, B., Del Zanna, G. & Mason, H. E., 2014, , 561, A20
O’Dwyer, B., Del Zanna, G., Mason, H. E.,Weber, M. A., & Tripathi, D. 2010, , 521, A21
Pesnell, W. D., Thompson, B. J., & Chamberlin, P. C. 2012, , 275, 3
Petkaki, P., Del Zanna, G. Mason, H. E. & Bradshaw, S. J., 2012, , 547, A25
Radziszewski, K., Rudawy, P., Phillips, K.J.H., 2007, A&A, 461, 303
Radziszewski, K., Rudawy, P., Phillips, K.J.H., Dennis, B.R., 2006, Adv. Space Res. 37, 1317
Rompolt, B., Mein, P., Mein, N., Rudawy, P., Berlicki, A., 1994, JOSO Annual Report 1993, 87
Rust, D. M., Simnett, G. M., & Smith, D. F. 1985, , 288, 401
Scullion, E., Rouppe van der Voort, L., Antolin, P., Wedemeyer, S., Vissers, G., Kontar, E. P., & Gallagher, P. T. 2016, , 833, 184
{width="8cm"} {width="8cm"}
![The 800-2000 radio spectrum observed during the 4 November 2015 flare by the Ondřejov radiospectrograph. The arrow shows the SPDB.[]{data-label="fig2"}](fig2.eps){width="18.0cm"}
{width="4.11cm"} {width="4.11cm"} {width="4.11cm"}
{width="4.11cm"} {width="4.11cm"} {width="4.11cm"}
{width="4.11cm"} {width="4.11cm"} {width="4.11cm"}
{width="4.11cm"} {width="4.11cm"} {width="4.11cm"}
{width="6.5cm"} {width="5.5cm"} {width="5.6cm"}
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Kinematical models are constrained by the latest observational data from geometry-distance measurements, which include 557 type Ia supernovae (SNIa) Union2 data and 15 observational Hubble data. Considering two parameterized deceleration parameter, the values of current deceleration parameter $q_{0}$, jerk parameter $j_{0}$ and transition redshift $z_{T}$, are obtained. Furthermore, we show the departures for two parameterized kinematical models from $\Lambda$CDM model according to the evolutions of jerk parameter $j(z)$. Also, it is shown that the constraint on jerk parameter $j(z)$ is weak by the current geometrical observed data.'
author:
- Jianbo Lu
- Lixin Xu
- Molin Liu
title: Constraints on kinematic models from the latest observational data
---
$\text{Introduction}$
=====================
The recently cosmic observations [@SNeRiess][@CMBSpergel][@LSSPope] suggest that the expansion of present universe is speeding up. The accelerated expansion of the universe is usually attributed to the fact that dark energy (DE) is an exotic component with negative pressure, such as cosmological constant, quintessence [@quintessence], phantom [@phantom], quintom [@quintom], generalized Chaplygin gas [@GCG], agegraphic dark energy [@agegraphic], etc [@other-models]. Also, the accelerating universe is related to the modification of the gravity theory at large scale, such as $f(R)$ modified gravity theory [@fR] and higher dimensional theory [@higherD], etc. These approaches correspond to the dynamics of the universe. For more information about dynamics of universe, please see review papers [@review].
Another route is the kinematical approach, which holds true regardless of the underlying cosmic dynamics [@kinematicalmodel], i.e., it depends neither on the validity of any particular metric theory of gravity nor on the matter-energy content of the observed universe [@09043550kinematical]. It is only related to the weaker assumption that space-time is homogeneous and isotropic so that the FRW metric is still valid [@09043550kinematical]. Then the kinematic approach is also called cosmokinetics [@cosmokineticas], cosmography [@cosmogrphy], or Friedmannless cosmology [@friedmannless]. For kinematic equations, there are Hubble parameter $H=\frac{\dot{a}}{a}$, deceleration parameter $q=-a\ddot{a}/\dot{a}^{2}$, and jerk parameter $j=-\dddot{a}a^{3}/(a\dot{a}^{3})$. It can be seen that parameters $H$, $q$ and $j$ are purely kinematical, since they are independent of any gravity theory, and all of them are only related to scale factor $a$ (or redshift $z$, since $a=\frac{1}{1+z}$).
The benefit of the kinematical analysis is that it has the fewer assumptions and a different set of models are explored for comparison with dynamical scenario. Since the origin of cosmic acceleration is unknown, the choice of parameterized kinematical model is essentially arbitrary. But inappropriate kinematical model could imply an unphysical universe at earlier time. For instance, for the model $q(z)=q_{0}+q_{1}z$, it has $q(z)>1/2$ at high redshift [@q0.5], which is not consistent with the matter dominated universe[^1]. The reason is simple that the model $q(z)=q_{0}+q_{1}z$, a expansion of $q(z)$ at low redshift, is not reliable at high redshift. As a complementarity to dynamical approach, in this paper we constrain two parameterized kinematical models by using the latest observational data: 557 type Ia supernovae (SNIa) Union2 dataset and 15 observational Hubble data.
$\text{The kinematical approach and models}$
============================================
The dimensionless Hubble and deceleration parameters are defined by the first and second derivative of scale factor $$H\equiv \frac{\dot{a}}{a}=-\frac{1}{1+z}\frac{dz}{dt},\label{e1}$$ $$q \equiv
-\frac{1}{H^{2}}(\frac{\ddot{a}}{a})=\frac{1}{2}(1+z)\frac{[H(z)^{2}]^{'}}{H(z)^{2}}-1,\label{e2}$$ where $^{"}$dot$^{"}$ denotes the derivative with respect to cosmic time $t$. $H$ and $q$ describe the rate of expansion and acceleration of universe. The relation between $H$ and $q$ is written as $$H=H_{0} \exp[\int^{z}_{0}[1+q(u)] d\ln(1+u)],\label{e3}$$ $H_{0}$ is Hubble constant. In this paper, subscript “0” denotes the current value of cosmological quantities.
Similar, the jerk parameter $j$ is defined as the dimensionless third derivative of scale factor with respect to cosmic time $$j \equiv
-\frac{1}{H^{3}}(\frac{\dot{\ddot{a}}}{a})=-[\frac{1}{2}(1+z)^{2}\frac{[H(z)^{2}]^{''}}{H(z)^{2}}-(1+z)\frac{[H(z)^{2}]^{'}}{H(z)^{2}}+1].\label{e4}$$ The use of the cosmic jerk parameter provides a more natural parameter space for kinematical studies, and transitions between phases of different cosmic acceleration are more naturally described by models incorporating a cosmic jerk [@jerkvalue]. Especially, since for flat $\Lambda$CDM model it has a constant jerk with $j(z)
= -1$, jerk parameter can provides us with a simple, convenient approach to search for departures from the cosmic concordance model, $\Lambda$CDM, just as deviations from $w = -1$ done in more standard dynamical analyses. The jerk parameter is related with the deceleration parameter by the following differential equation $$j=-[q+2q^{2}+(1+z)\frac{dq}{dz}].\label{e5}$$
From Eqs. (\[e3\]) and (\[e5\]), we can see that the expressions of Hubble parameter and jerk parameter can be given by deceleration parameter. Here we consider two parameterized deceleration parameters, $q(z)=q_{0}+\frac{q_{1}z}{1+z}$ ($M_{1}$) [@qa] and $q(z)=\frac{1}{2}+\frac{q_{1}+q_{2}z}{(1+z)^{2}}$ ($M_{2}$) [@qgong]. $M_{1}$ is one order expansion of scale factor $a$ at present $(a = 1)$, i.e., $q(a) = q_{0} + q_{1}(1-a)$. $M_{2}$ is an alternative parametrization to the three-epoch model[^2]. For these two models, the expressions of Hubble parameter, deceleration parameter and jerk parameter are shown in Table \[table1\].
For other parameterized deceleration parameter appeared in Ref. [@09043550kinematical], such as model $q(z)=q_{0}+q_{1}z$ and $q(z)=q_{0}$=constant, we will not discuss: since the former model is only interested at low redshift, not all the observed data from 557 SNIa Union2 dataset (redshift interval $0.015\leq z \leq 1.4$) and 15 observational Hubble data (redshift interval $0 \leq z \leq
1.75$) can be used to constrain its evolution; though the latter one indicates an accelerating universe [@09043550kinematical], it does not describe a transition of universe from decelerated expansion to accelerated expansion. Furthermore, for two-epoch model ($q=q_{0}$, $z\leq z_{T}$; $q=q_{1}$, $z > z_{T}$), the $q(z)$ function is not smooth, so it will not appeared in the paper, too.
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
$H(z)$ $q(z)$ $j(z)$
------------------------------------------------------- ---------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------
$H_{0}(1+z)^{1+q_{0}+q_{1}}\exp(-\frac{q_{1}z}{1+z})$ $q_{0}+\frac{q_{1}z}{1+z}$ $-q_{0}-\frac{q_{1}z}{1+z}-2(q_{0}+\frac{q_{1}z}{1+z})-(1+z)[\frac{q_{1}}{1+z}-\frac{q_{1}z}{(1+z)^{2}}]$
$H_{0}(1+z)^{\frac{3}{2}} $\frac{1}{2}+\frac{q_{1}+q_{2}z}{(1+z)^{2}}$ $-\frac{1}{2}-\frac{q_{1}+q_{2}z}{(1+z)^{2}}-2(\frac{1}{2}+\frac{q_{1}+q_{2}z}{(1+z)^{2}})^{2}-(1+z)[\frac{q_{2}}{(1+z)^{2}}-\frac{2(q_{1}+q_{2}z)}{(1+z)^{3}}]$
\exp[\frac{2q_{1}z+(q_{1}+q_{2})z^{2}}{2(1+z)^{2}}]$
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
: The expressions of Hubble parameter $H(z)$, deceleration parameter $q(z)$ and jerk parameter $j(z)$ for two models.[]{data-label="table1"}
$\text{Data and analysis Methods}$
==================================
Since SNIa behave as excellent standard candles, they can be used to directly measure the expansion rate of the universe up to high redshift for comparison with the present rate. Theoretical cosmic parameters are determined by minimizing the quantity [@chi2SNe] $$\chi^{2}_{SNIa}(H_{0}, \theta)=\sum_{i=1}^{N}\frac{(\mu_{obs}(z_{i})
-\mu_{th}(z_{i},H_{0},\theta))^2}{\sigma^2_{obs;i}},\label{e6}$$ where $N=557$ for SNIa Union2 data; $\sigma^2_{obs;i}$ are errors due to flux uncertainties, intrinsic dispersion of SNIa absolute magnitude and peculiar velocity dispersion, respectively; $\theta$ denotes the model parameters; $\mu_{obs}$ is the observed values of distance modulus and can be given by the SNIa dataset; the expression of theoretical distance modulus $\mu_{th}$ is related to the apparent magnitude of supernova at peak brightness $m$ and the absolute magnitude $M$, $$\mu_{th}(z_{i})\equiv
m_{th}(z_{i})-M=5log_{10}(D_{L}(z))+\mu_{0},\label{e7}$$ where luminosity distance $$D_{L}(z)=H_{0}d_{L}(z)=(1+z)\int_{0}^{z}\frac{H_{0}dz^{'}}{H(z^{'};\theta)},\label{e8}$$ and $$\mu_{0}=5log_{10}(\frac{H_{0}^{-1}}{Mpc})+25=42.38-5log_{10}h.\label{e9}$$ It should be noted that $\mu_{0}$ is independent of the data and the dataset, though it is a nuisance parameter. By expanding the $\chi^{2}$ of Eq. (\[e6\]) relative to $\mu_{0}$, the minimization with respect to $\mu_{0}$ can be made trivially [@SNeABC][@chi2sneli] $$\chi^{2}_{SNIa}(\theta)=A(\theta)-2\mu_{0}B(\theta)+\mu_{0}^{2}C,\label{e10}$$ where $$A(\theta)=\sum_{i=1}^{N}\frac{[\mu_{obs}(z_{i})-\mu_{th}(z_{i};\mu_{0}=0,\theta)]^{2}}{\sigma_{i}^{2}},\label{e11}$$ $$B(\theta)=\sum_{i=1}^{N}\frac{\mu_{obs}(z_{i})-\mu_{th}(z_{i};\mu_{0}=0,\theta)}{\sigma_{i}^{2}},\label{e12}$$ $$C=\sum_{i=1}^{N}\frac{1}{\sigma_{i}^{2}},\label{e13}$$ Evidently, Eq. (\[e6\]) has a minimum for $\mu_{0}=B/C$ at $$\widetilde{\chi}^{2}_{SNIa}(\theta)=A(\theta)-B(\theta)^{2}/C.\label{e14}$$ Since $\chi^{2}_{SNIa,min}=\widetilde{\chi}^{2}_{SNIa,min}$ and $\widetilde{\chi}^{2}_{SNIa}$ is independent of nuisance parameter $\mu_{0}$, here we utilize expression (\[e14\]) to displace (\[e6\]) for SNIa constraint. Alternatively, one can also perform a uniform marginalization over the nuisance parameter $\mu_{0}$ thus obtaining [@chi2sne-ln-refs1; @chi2sne-ln-refs2; @chi2sne-ln-refs3] $$\chi^{2}_{SNIa}(\theta)=A(\theta)-\frac{B(\theta)^{2}}{C}+\ln(\frac{C}{2\pi}).\label{chi2SNIa-ln}$$ Comparing Eq. (14) with Eq. (15), it can be seen that two $\chi^{2}$s are equivalent for using them to constrain cosmological models, since they are only different from a constant term $\ln(\frac{C}{2\pi})$, i.e. if one marginalize over all values of $\mu_{0}$, as in Ref. [@chi2sne-marginalization], that would just add a constant and would not change the constraint results.
Recently, Stern $et$ $al$ obtained the Hubble parameter $H(z)$ at 12 different redshifts from the differential ages of passively evolving galaxies [@12Hubbledata]. And in Ref. [@3Hubbledata], authors obtained $H(z = 0.24) = 79.69 \pm 2.32$, $H(z = 0.34) = 83.8
\pm 2.96$, and $H(z = 0.43) = 86.45 \pm 3.27$ by taking the BAO scale as a standard ruler in the radial direction. Using these data we can constrain cosmological models by minimizing $$\chi^2_{Hub}(H_{0},
\theta)=\sum_{i=1}^{N}\frac{\left[H_{th}(z_i)-H_{obs}(H_{0},\theta,
z_i)\right]^2}{\sigma^2_{obs;i}},\label{e15}$$ where $H_{th}$ is the predicted value for the Hubble parameter, $H_{obs}$ is the observed value, $\sigma^2_{obs;i}$ is the standard deviation measurement uncertainty. Here the nuisance parameter $H_{0}$ is marginalized in the following calculation with a Gaussian prior, $H_{0} = 74.2 \pm 3.6$ km s$^{-1}$ Mpc$^{-1}$ [@H0prior].
![The best-fit evolutions of $q(z)$ with $1\sigma$ confidence level constrained from 557 SNIa Union2 dataset and 15 observational Hubble data.[]{data-label="figureq"}](qa557-15.eps "fig:"){width="5cm"} ![The best-fit evolutions of $q(z)$ with $1\sigma$ confidence level constrained from 557 SNIa Union2 dataset and 15 observational Hubble data.[]{data-label="figureq"}](qg557-15.eps "fig:"){width="5cm"} ![The best-fit evolutions of $q(z)$ with $1\sigma$ confidence level constrained from 557 SNIa Union2 dataset and 15 observational Hubble data.[]{data-label="figureq"}](557+15qLCDM.eps "fig:"){width="5cm"}\
The likelihood function is written as $L\propto e^{-\chi^{2}/2}$, and the total $\chi^{2}$ equals, $$\chi^{2}_{total}=\tilde{\chi}^{2}_{SNIa}+\chi^{2}_{Hub},\label{e16}$$ where $\tilde{\chi}^{2}_{SNIa}$ and $\chi^{2}_{Hub}$ are the ones described in Eq. (\[e14\]) and Eq. (\[e15\]), respectively. It is easy to see that the matter density $\Omega_{m}$ are not contained explicitly in the $\chi^{2}_{total}$. Then, the constraint results may not depend on the dynamic variables $\Omega_{m}$, and gravitation theory.
model $\chi_{min}^{2}$ $\chi_{min}^{2}$/dof $q_{0}$ $(1\sigma)$ $j_{0}$ $(1\sigma)$ $z_{T}$ $(1\sigma)$
-------------- ------------------- ---------------------- --------------------------------------------------------- --------------------------------------------------------- -------------------------------------------------------
$M_{1}$ 554.335 (478.060) 0.973(1.163) $-0.701^{+0.089}_{-0.089}$ ($-0.653^{+0.092}_{-0.093}$) $-2.000^{+0.443}_{-0.442}$ ($-1.823^{+0.456}_{-0.457}$) $0.689^{+0.227}_{-0.117}$ ($0.674^{+0.257}_{-0.124}$)
$M_{2}$ 554.560 (478.334) 0.973(1.167) $-0.749^{+0.103}_{-0.103}$ ($-0.698^{+0.108}_{-0.108}$) $-2.619^{+0.602}_{-0.602}$ ($-2.386^{+0.619}_{-0.620}$) $0.687^{+0.198}_{-0.121}$ ($0.667^{+0.215}_{-0.126}$)
$\Lambda$CDM 557.008 (479.283) 0.975(1.166) $-0.598^{+0.028}_{-0.027}$ ($-0.575^{+0.030}_{-0.030}$) $-1$ $0.761^{+0.055}_{-0.055}$ ($0.716^{+0.056}_{-0.056}$)
: The values of the current deceleration parameter $q_{0}$, jerk parameter $j_{0}$, and transition redshift $z_{T}$ against the model, obtained by using 557 SNIa Union2 data and 15 observational Hubble data (the numerical results in brackets correspond to the constraints from 397 SNIa Constitution data and 15 observational Hubble data).[]{data-label="table2"}
![The best-fit evolutions of $j(z)$ with $1\sigma$ confidence level constrained from 557 SNIa Union2 dataset and 15 observational Hubble data.[]{data-label="figurej"}](ja557-15.eps "fig:"){width="5.5cm"} ![The best-fit evolutions of $j(z)$ with $1\sigma$ confidence level constrained from 557 SNIa Union2 dataset and 15 observational Hubble data.[]{data-label="figurej"}](jg557-15.eps "fig:"){width="5.5cm"}\
The evolutions of two kinematical parameters, $q(z)$ and $j(z)$, are plotted in Fig. \[figureq\] and \[figurej\] for two models, respectively. According to these figures, constraint results of cosmic parameters are obtained and shown in table \[table2\], with the constraints on model parameters shown in table \[table-modelparams\]. Relative to the evolutions of deceleration parameters $q(z)$ it can be seen that the constraints on jerk parameters $j(z)$, defined by the third derivative with respect to scale factor $a$, are weaker for these two models from the distance-measurement data. Furthermore, in Fig. \[figureq\] as a contrast we also plot the evolution of deceleration parameter $q(z)$ for $\Lambda$CDM model by the current geometry-distance observed data (and for this model, jerk parameter $j(z)=-1$ almost). From table \[table2\], it is shown that for the dynamical $\Lambda$CDM model, the constraint results tend to favor the bigger values of current deceleration parameter $q_{0}$ and transition redshift $z_{T}$ relative to the cases of two kinematical models.
model The values of model parameters $(1\sigma)$
-------------- -------------------------------------------------------------------------------------------------------------------------------------------------
$M_{1}$ $q_{0}=-0.701^{+0.089}_{-0.089}$ ($q_{0}=-0.653^{+0.092}_{-0.093}$), $q_{1}=1.718^{+0.475}_{-0.504} $ ($q_{1}=1.623^{+0.493}_{-0.529} $)
$M_{2}$ $q_{1}=-0.253^{+0.653}_{-0.673} $ ($q_{1}=-0.286^{+0.692}_{-0.714} $), $q_{2}=-1.249^{+0.106}_{-0.158} $ ($q_{2}=-1.198^{+0.165}_{-0.162} $)
$\Lambda$CDM $\Omega_{0m}=0.268^{+0.019}_{-0.018}$ ($\Omega_{0m}=0.283^{+0.021}_{-0.019}) $
: The $1\sigma$ confidence level of model parameters for the models: $q(z)=q_{0}+\frac{q_{1}z}{1+z}$, $q(z)=\frac{1}{2}+\frac{q_{1}+q_{2}z}{(1+z)^{2}}$ and $\Lambda$CDM model by using the 557 SNIa Union2 data and 15 observational Hubble data (the numerical results in brackets correspond to the constraints from 397 SNIa Constitution data and 15 observational Hubble data).[]{data-label="table-modelparams"}
$\text{Conclusions}$
====================
In this paper, kinematic models are constrained by the latest observational data: 557 SNIa Union2 dataset and 15 observational Hubble data. Generally, the expansion rhythm of current universe $q_{0}$ and transitional time from decelerated expansion to accelerated expansion $z_{T}$, depends on the parameterized form of kinematical equations. Here we consider two parameterized deceleration parameter. The best fit values of $z_{T}$, $q_{0}$ and $j_{0}$ with 1$\sigma$ confidence level are obtained. From table \[table2\] we can see that the values of $z_{T}$ indicated by these two models much approach each other. From Fig. \[figureq\] and \[figurej\] it can be seen that for the two kinematical models, the constraints on jerk parameters $j(z)$ are weak by the current observed data. In addition, we also can see the deviation of jerk parameter from $j =
-1$ according to the Fig. \[figurej\], with measuring the departures for kinematical models from $\Lambda$CDM model. Furthermore, considering Refs. [@09043550kinematical][@jerkvalue][@q0.5xu], where most models indicate that current data favors $j_{0} < -1$ case, which is consistent with our results. At last, for comparing the differences of constraint results on cosmic parameters between the different SNIa data, we also consider the case of displacing the 557 SNIa Union2 data[@557Union2][^3] with 397 SNIa Constitution data [@397Constitution][^4] in the above combined constraints, and the latter constraint results are listed in brackets in table \[table2\]. According to table \[table2\], it seems that the constraint results favor a bigger value of current deceleration parameter $q_{0}$ and jerk parameter $j_{0}$, and a smaller value of transition redshift $z_{T}$.
** Acknowledgments** The research work is supported by the National Natural Science Foundation (Grant No. 10875056), NSF (10703001) and NSF (No.11005088) of P.R. China.
[\*]{} A.G. Riess, et al., Astron. J. 116 (1998) 1009, \[arXiv:astro-ph/9805201\]; S. Perlmutter [*et al*]{}, 1999 [*Astrophys. J.*]{} [**517**]{} 565.
D.N. Spergel, et al., Astrophys. J. Suppl. 148 (2003) 175, \[arXiv:astro-ph/0302209\].
A.C. Pope, et al., Astrophys. J. 607 (2004) 655, \[arXiv:astro-ph/0401249\].
B. Ratra and P.J.E. Peebels, 1988 [*Phys. Rev. D.*]{} [**37**]{} 3406.
R.R. Caldwell, M. Kamionkowski and N. N. Weinberg, 2003 [*Phys. Rev. Lett.*]{} [**91**]{} 071301 \[arXiv:astro-ph/0302506\]; M.R. Setare, 2007 [*Eur. Phys. J. C* ]{}[**50**]{} 991.
B. Feng, X.L. Wang and X.M. Zhang, 2005 [*Phys. Lett. B*]{} [**607**]{} 35 \[arXiv:astro-ph/0404224\].
A.Y. Kamenshchik, U. Moschella and V. Pasquier, 2001 Phys. Lett. B 511 265 \[arXiv:gr-qc/0103004\]; M.C. Bento, O. Bertolami and A.A. Sen, Phys. Rev. D 66 (2002) 043507 \[arXiv:gr-qc/0202064\]; J.B. Lu, Y.X. Gui and L.X. Xu, 2009 Eur. phys. J. C 63 349.
R.G. Cai, Phys. Lett. B 657, 228 (2007) \[arXiv:hep-th/ 0707.4049\].
J.B. Lu, E.N. Saridakis, M.R. Setare and L.X. Xu. JCAP, 2010, 03: 031 \[arXiv:astro-ph/0912.0923\]; J.B. Lu et al, Gen. Relativ. Gravit. (2011) 43:819-832; J.B. Lu et al, 2008 Phys. Lett. B 662, 87; J.B. Lu, Phys. Lett. B 680, 404 (2009).
Barrow J D and Cotsakis S 1988 Phys. Lett. B 214 515.
L. Randall and R. Sundrum, Phys.Rev.Lett 83 3370 (1999) \[hep-ph/9905221\]; G. Dvali, G. Gabadadze and M. Porrati, Phys.Lett.B 485 208-214 (2000) \[hep-th/0005016\].
P.J.E. Peebles and B. Ratra, Rev. Mod. Phys. 75 (2003) 559 \[astro-ph/0207347\]; T. Padmanabhan, Phys. Rept. 380 (2003) 235 \[hep-th/0212290\]; S.M. Carroll, \[astro-ph/0310342\]; R. Maartens, \[arXiv:gr-gc/0312059\]; M. Trodden and S.M. Carroll, \[astro-ph/0401547\]; V. Sahni, \[astro-ph/0403324\]; R. Bean, S.M. Carroll and M. Trodden, \[astro-ph/0510059\]; V. Sahni and A.A. Starobinsky, Int. J. Mod. Phys. D 9 (2000) 373 \[astro-ph/9904398\]; S. Nojiri, S.D. Odintsov, \[hep-th/0601213\]; R.P. Woodard, \[hep-th/0601672\]; S. Bludman, \[astro-ph/0605198\]; E.J. Copeland, M. Sami and S. Tsujikawa, Int. J. Mod. Phys. D 15 (2006) 1753 \[hep-th/0603057\]; A.J. Albrecht et al., \[astro-ph/0609591\]; M. Kamionkowski, \[arXiv:0706.2986\]; B. Ratra and M.S. Vogeley, \[arXiv:0706.1565\]; E.V. Linder, Am. J. Phys. 76 (2008) 197 \[arXiv:0705.4102\]; J. Frieman, M. Turner and D. Huterer, Ann. Rev. Astron. Astrophys. 46 (2008) 385 \[arXiv:0803.0982\]; M.B. Altaie, \[arXiv:0907.4280\]; D. Baumann, \[arXiv:0907.5424\].
J.M. Virey et al., Phys. Rev. D 72 (2005) 061302 \[astro-ph/0502163\]; D. Rapetti, S.W. Allen, M.A. Amin and R.A. Daly et al., Astrophys. J. 677 (2008) 1; J.V. Cunha and J.A.S. Lima, Mon. Not. Roy. Astron. Soc. 390 (2008) 210 \[arXiv:0805.1261\].
A.C.C. Guimaraes, J.V. Cunha and J.A.S. Lima, \[arXiv:0904.3550\].
R.D. Blandford, M. Amin, V. Baltz, K. Mandel, P.J. Marshall, Astron. Soc. Pac. Conf. Series, 339, 27(2005) \[astro-ph/0408279\].
M.S. Turner, A.G. Riess, Astrophys. J. 569 18(2002) \[astro-ph/0106051\]; C. Shapiro, M. S. Turner, Astrophys. J. 649 563(2006) \[astro-ph/0512586\].
. Elgar${\o}$y, T. Multamäki, Mon. Not. Roy. Astron. Soc. 356 475(2005) \[astro-ph/0404402\]; [Ø]{}. Elgar[ø]{}y, T. Multamäki T JCAP 9 2(2006) \[astro-ph/0603053\].
J.V. Cunha, Phys. Rev. D 79 (2009) 047301 \[arXiv:0811.2379\].
R.D. Blandford, Mon. Not. Roy. Astron. Soc. 375 (2007) 1510 \[astro-ph/0605683\].
L.X. Xu and J.B. Lu, Mod. Phys. Lett. A 24 (2009) 369
Y.G. Gong, A.Z. Wang, Phys. Rev. D 73, 083506 (2006). astro-ph/0601453; Y.G. Gong, A.Z. Wang, Phys. Rev. D 75, 043520 (2007). \[astro-ph/0612196\]; J.B Lu, L.X Xu, M.L Liu and Y.X Gui, 2008 Eur. Phys. J. C 58 311 \[arXiv:astro-ph/0812.3209\].
M.S. Turner, A.G. Riess et al., Astrophys. J. 569, 18 (2002); C.A. Shapiro, M.S. Turner, astro-ph/0512586.
R. Amanullah et al. \[Supernova Cosmology Project Collaboration\], \[arXiv:astro-ph/1004.1711\]\
The numerical data of the full sample are available at http://supernova.lbl.gov/Union.
S. Nesseris and L. Perivolaropoulos, Phys. Rev. D 72 (2005) 123519 \[astro-ph/0511040\]; M. Szydlowski and W. Godlowski, 2006 [*Phys. Lett. B*]{} [**633**]{} 427 \[arXiv:astro-ph/0509415\]; S. Nesseris and L. Perivolaropoulos, 2007 [*JCAP*]{} [**0702**]{} 025, \[arXiv:astro-ph/0612653\]; R. Lazkoz, S. Nesseris, and L. Perivolaropoulos, J. Cosmol. Astropart. Phys. 07 (2008) 012, \[arXiv:astro-ph/0712.1232\].
L. Perivolaropoulos, 2005 [*Phys. Rev. D*]{} [**71**]{} 063503 \[arXiv:astro-ph/0412308\]; E. Di Pietro and J. F. Claeskens, 2003 [*Mon. Not. Roy. Astron. Soc.*]{} [**341**]{} 1299, \[arXiv:astro-ph/0207332\].
M. Li, X.D. Li, S. Wang and X. Zhang, \[arXiv:astro-ph/0904.0928\].
S. Nesseris, L. Perivolaropoulos, Phys.Rev.D 72:123519 (2005) \[arXiv:astro-ph/0511040\].
R. Lazkoz, S. Nesseris, L. Perivolaropoulos, JCAP 0511:010 (2005) \[arXiv:astro-ph/0503230\].
S. Nesseris and L. Perivolaropoulos, Phys. Rev. D 70, 043531 (2004) \[arXiv:astro-ph/0401556\].
A. G. Riess et al. \[Supernova Search Team Collaboration\], Astrophys. J. 607, 665 (2004).
D. Stern, R. Jimenez, L. Verde, M. Kamionkowski and S. A. Stanford, \[arXiv:astro-ph/0907.3149\]; J. Simon et al, 2005 Phys. Rev. D 71, 123001.
E. Gaztanñaga, A. Cabré and L. Hui, \[arXiv: 0807.3551\].
A. G. Riess [*et al.*]{}, arXiv:0905.0695\[astro-ph.CO\].
L.X. Xu W.B. Li and J.B. Lu, JCAP, 07 (2009) 31 \[arXiv:0905.4552\].
D. Rubin [*et al*]{}, \[arXiv:astro-ph/0807.1108\].
M. Hicken et al., \[arXiv:astro-ph/0901.4804\].
[^1]: For the matter dominated universe, it has $q(z)=1/2$, then there must be $q(z) \leq
1/2$ for any kinematical or dynamical model.
[^2]: The three-epoch model [@threeepoch] is a alternative scenario for the model $q(z)=q_{0}+q_{1}z$, but the problem is that this function $q(z)$ is not smooth [@qgong].
[^3]: The SNIa Union2 data are obtained, by adding new datapoints (including the high redshift SNIa) to the SNIa Union [@307Union] data, making a number of refinements to the Union analysis chain, refitting all light curves with the SALT2 fitter.
[^4]: The 397 Constitution data are obtained by adding 90 SNIa from CfA3 sample to 307 SNIa Union sample[@307Union]. CfA3 sample are all from the low-redshift SNIa, $z<0.08$, and these 90 SNIa are calculated with using the same Union cuts.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We consider the the $n$-dimensional generalisation of the nonholonomic Veselova problem. We derive the reduced equations of motion in terms of the mass tensor of the body and determine some general properties of the dynamics. In particular we give a closed formula for the invariant measure, we indicate the existence of steady rotation solutions, and obtain some results on their stability.
We then focus our attention on bodies whose mass tensor has a specific type of symmetry. We show that the phase space is foliated by invariant tori that carry quasi-periodic dynamics in the natural time variable. Our results enlarge the known cases of integrability of the multi-dimensional Veselova top. Moreover, they show that in some previously known instances of integrability, the flow is quasi-periodic without the need of a time reparametrisation.
author:
- 'Francesco Fassò, Luis C. García-Naranjo & James Montaldi'
title:
- 'Integrability and dynamics of the $n$-dimensional symmetric Veselova top[^1]'
- Lie Groups
---
Introduction
============
The $n$-dimensional generalisation of some classical nonholonomic systems considered recently by Fedorov and Jovanović [@FedJov; @FedJov2; @Jovan; @JovaRubber] provides a remarkable family of examples of nonholonomic systems whose dynamics, after a time reparametrisation, is quasi-periodic in large dimensional invariant tori. A crucial feature of these examples is that they possess an invariant measure and admit a Hamiltonization by Chaplygin’s multiplier method [@ChapRedMult]. After a time reparametrisation, the reduced equations of motion of these examples become Hamiltonian and, moreover, turn out to be related to classical integrable Hamiltonian systems. In this manner the authors establish the integrability of the systems and prove that the flow on the invariant, large dimensional tori is quasiperiodic in the new time. The Hamiltonization of these systems is hence central to their approach and is truly remarkable considering that Chaplygin’s method is only guaranteed to work for systems with $2$ degrees of freedom. This Hamiltonization relies on the very particular type of inertia operators considered by the authors that, as we explain below, are generally unphysical.
In this paper we analyse the dynamics of the $n$-dimensional Veselova top (described below) and treated before by Fedorov and Jovanović [@FedJov; @FedJov2]. A fundamental difference in our study of the problem with respect to these references is that we assume that the inertia tensor is [*physical*]{} (see below).
Our main contribution is to show that under certain symmetry assumptions on the mass distribution of the body, the dynamics takes place on invariant tori that carry quasi-periodic flow [*without the need of a time reparametrisation*]{}. We prove this by performing a detailed symmetry analysis of the problem, and in particular by performing a reduction to a system that is manifestly integrable and in fact periodic. The quasi-periodicity of the flow of the unreduced system (in the natural time variable) then follows from a theorem of Field [@Field80]. This type of analysis to establish integrability of nonholonomic systems had been previously followed in e.g. [@Hermans; @FassoGiacobbe; @FS2016].
Our approach allows us to recover all cases of integrability of the multi-dimensional Veselova top determined in [@FedJov; @FedJov2] that correspond to physical inertia tensors and we show that these correspond to [*axisymmetric*]{} bodies. We also determine new cases of integrability that we term [*cylindrical*]{} bodies. For this type of body, the system does not seem to allow a Chaplygin Hamiltonization (Remark \[rmk:hamiltonization\]), and the generic motion takes place in 4-dimensional invariant tori.
This paper also considers some properties of the motion of the general $n$-dimensional Veselova top (without any symmetry assumptions on the body). We prove the existence of steady rotation solutions that are periodic solutions with constant angular velocity, and determine some of their properties. In particular, we prove that some of these solutions are stable. The existence of these solutions appears not to have been observed before, not even in the 3D case. We also give the explicit form of the reduced equations of motion in terms of the mass tensor of the body, and a closed formula for the known invariant measure.
#### The $n$-dimensional Veselova top.
The 3D Veselova top, introduced by Veselova in her thesis [@Veselova], is a rigid body rotating under its own inertia (like Euler’s rigid body) and subject to a nonholonomic constraint which forces the projection of the angular velocity to a distinguished axis fixed in space, to vanish at all time (see also [@Veselov-Veselova-1988]).
This paper deals with a multi-dimensional generalization of this model that was introduced by Fedorov and Kozlov [@FedKoz]. They consider the motion of an $n$-dimensional rigid body subject to the nonholonomic constraint that requires the angular velocity matrix $\Omega \in {\mathfrak{so}}(n)$ to have rank two and to define a rotation in a plane containing a fixed axis in space.
The Veselova top and its $n$-dimensional generalisation are examples of [*LR systems*]{} which form a remarkable class of nonholonomic systems possessing an invariant measure [@Veselov-Veselova-1988].
#### The inertia tensor.
Throughout this paper, we work with the commonly accepted $n$-dimensional generalisation of rigid body dynamics (see e.g. [@Ratiu80]), where the inertia tensor is a linear operator ${\mathbb{I}}:{\mathfrak{so}}(n)\to {\mathfrak{so}}(n)$ of the form $$\label{eq:Phys-Inertia-Intro}
{\mathbb{I}}(\Omega) = {\mathbb{J}}\Omega + \Omega {\mathbb{J}},$$ where ${\mathbb{J}}$ is the [[***[mass tensor]{}***]{}]{}. ${\mathbb{J}}$ is a constant $n\times n$ matrix that depends on the mass distribution of the body which, by an appropriate choice of a body frame, may be assumed to be diagonal with positive entries (see Section\[sec:rigid body\] for details). We say that a linear operator ${\mathbb{I}}:{\mathfrak{so}}(n)\to {\mathfrak{so}}(n)$ is a [[***[physical inertia tensor]{}***]{}]{} if and only if it may be written as above for a certain mass tensor ${\mathbb{J}}$.
Summary of results {#summary-of-results .unnumbered}
------------------
The configuration space of the $n$-dimensional Veselova top problem is the orthogonal group[^2] ${\mathrm{O}}(n)$. The phase space for the $n$-dimensional Veselova top is a subbundle $D\subset T{\mathrm{O}}(n)$ of rank $n-1$ determined by the nonholonomic constraints.
Our first contribution is to show that the system admits steady rotation solutions on $D$. These are periodic solutions where the body steadily rotates in a principal plane of the body. Because of the constraint, the orientation of the body along these motions is such that the distinguished axis is contained in the plane of rotation. Along these solutions, the nonholonomic constraint forces vanish.
Next we perform the reduction of the system by the symmetry group $G_L={\mathrm{O}}(n-1)$ that corresponds to invariance of the system under rotations and reflections of the space frame that fix the distinguished axis (the subscript $L$ indicates that this action is by left multiplication in the configuration space ${\mathrm{O}}(n)$). As shown in [@FedJov; @FedJov2], this is a generalised Chaplygin reduction and the [[***[first reduced space]{}***]{}]{} $D/G_L$ is isomorphic to the cotangent bundle $T^*{\mathrm{S}}^{n-1}$. We give the reduced equations of motion and the explicit expression for the invariant measure in terms of the mass tensor ${\mathbb{J}}$. Moreover, with the reduced system at hand, we are able to show that the steady rotation solutions correspond to the singularities of an energy-momentum map and we show stability of those occurring on [*extremal* ]{} planes. In particular this proves that in 3D, the steady rotations about the smallest and largest axes of inertia of the body are stable.
We then focus on the study of symmetric $n$-dimensional rigid bodies whose mass tensor has just 2 distinct eigenvalues: $${\mathbb{J}}= {\mathop\mathrm{diag}\nolimits}[J_1,\dots,J_1,J_2,\dots,J_2],$$ with $J_1\neq J_2$. We assume that the multiplicities of $J_1$ and $J_2$ are $r$ and $r'=n-r$, with $1\leq r\leq r'$. Under this hypothesis, the mass tensor ${\mathbb{J}}$ has symmetry $G_R:={\mathrm{O}}(r)\times{\mathrm{O}}(r')\subset{\mathrm{O}}(n)$: if $$h=\left[\begin{matrix}
h_{11}&0\cr 0&h_{22}
\end{matrix}\right],$$ with $h_{11}\in{\mathrm{O}}(r)$ and $h_{22}\in{\mathrm{O}}(r')$, then $h{\mathbb{J}}h^T={\mathbb{J}}$. The subindex $R$ indicates that $G_R$ acts by right multiplication on the configuration space ${\mathrm{O}}(n)$. Since this action commutes with the left multiplication action of $G_L={\mathrm{O}}(n-1)$, it passes down to a proper action on $T^*{\mathrm{S}}^{n-1}$ that turns out not to be free.
The [[***[second reduced space]{}***]{}]{} ${\mathcal{R}}=T^*{\mathrm{S}}^{n-1} /G_R$ is not a smooth manifold, but a stratified space where, as we shall show, the dynamics is periodic. We apply a theorem of Field [@Field80] to conclude that the reconstructed motion, both on $D$ and on $T^*{\mathrm{S}}^{n-1}$, is quasi-periodic on invariant tori. We emphasise that this proves quasi-periodicity of the flow without a time reparametrisation.
In order to estimate the dimension of the invariant tori on $D$ and on $T^*{\mathrm{S}}^{n-1}$ it is necessary to determine the details of the stratification. For this, we distinguish two cases depending on the multiplicity $r$ of the eingenvalue $J_1$ as follows.
1. The mass tensor is $${\mathbb{J}}= {\mathop\mathrm{diag}\nolimits}[J_1,J_2,\dots,J_2],$$ that corresponds to an [[***[axisymmetric body]{}***]{}]{}. In this case the orbit space ${\mathcal{R}}$ is isomorphic to a singular semi-algebraic subspace in ${\mathbb{R}}^3$. We show that the dynamics on $D$ is essentially that of the 3D axisymmetric Veselova top (see Theorem \[thm:nD to 3D\] for a precise statement), and is generically quasi-periodic on tori of dimension $2$ both in $D$ and on $T^*{\mathrm{S}}^{n-1}$. We also give a physical description of the dynamics.
2. A body with this type of symmetry is called a [[***[cylindrical body]{}***]{}]{} and can only exist in dimension 4 or higher. For these bodies the orbit space ${\mathcal{R}}$ is isomorphic to a singular semi-algebraic subspace in ${\mathbb{R}}^4$. For $n>4$ we prove that the system evolves as a 4D cylindrical Veselova top (a precise statement is given in Theorem \[thm:nD to 4D\]) and we prove that the generic motion is quasi-periodic on tori of dimension 4 on $D$ and of dimension 3 on $T^*{\mathrm{S}}^{n-1}$. We also give an argument to prove that the generic motions do not take place in lower dimensional invariant tori.
#### Previous work.
Fedorov and Jovanović [@FedJov; @FedJov2] show that the $G_L$-reduction of the $n$-dimensional Veselova top to $T^*{\mathrm{S}}^{n-1}$ is Hamiltonizable and integrable under the assumption that the action of the inertia operator on rank 2 matrices[^3] $a\wedge b\in {\mathfrak{so}}(n)$ satisfies $$\label{eq:inertia-Fed-Jov}
{\mathbb{I}}(a\wedge b) = (Aa)\wedge (Ab),$$ for a diagonal matrix $A={\mathop\mathrm{diag}\nolimits}[A_1,\dots, A_n]$. In particular, this implies that the space of rank 2 matrices in ${\mathfrak{so}}(n)$ is invariant under ${\mathbb{I}}$. This condition is vacuous in dimension $3$ since every non-zero matrix in ${\mathfrak{so}}(3)$ has rank 2. However, for $n\geq 4$ it is a very restrictive assumption, as we discuss in Appendix \[A:Inertia\]. It turns out that an inertia tensor satisfying is physical as defined by Equation if and only if the body is axisymmetric. In this case $A={\mathop\mathrm{diag}\nolimits}[A_1,A_2,\dots, A_2]$, and the mass tensor ${\mathbb{J}}$ is determined by the relations $A_2^2=2J_2$ and $A_1A_2=J_1+J_2$.
As mentioned above, their approach only allows them to conclude quasi-periodicity of the flow after a time reparametrisation.
Structure of the paper {#structure-of-the-paper .unnumbered}
----------------------
In Section\[sec:Veselova\] we begin by recalling the kinematics of the rigid body in arbitrary dimensions, governed by a mass tensor ${\mathbb{J}}$. We then recall from [@FedKoz] the $n$-dimensional generalisation of the Veselova top and we discuss its symmetries, which depend on ${\mathbb{J}}$. We also describe the steady rotations which play an important role later. In Section \[sec:1st reduction\] we perform the reduction of the system by $G_L= {\mathrm{O}}(n-1)$. We obtain the reduced equations of motion in terms of ${\mathbb{J}}$ and determine general properties of the motion. Our analysis of axisymmetric and cylindrical bodies is respectively contained in Sections \[sec:axisymmetric\] and \[sec:cylindrical\]. In Section \[S:Question-Hamiltonisation\] we present some conclusions and open questions.
The paper ends with two appendices. The first recalls a well-known result about group actions, the reconstruction theorem of Field which we use in the proofs of quasiperiodicity, as well as some details on isotropy subgroups. The second appendix considers in detail the physical implications of the inertia tensor hypothesis that Fedorov and Jovanović make in [@FedJov; @FedJov2].
The Veselova system {#sec:Veselova}
===================
The $\boldsymbol n$-dimensional rigid body {#sec:rigid body}
------------------------------------------
We begin by recalling the model of an $n$-dimensional rigid body ${\mathcal{B}}$ that moves in ${\mathbb{R}}^n$ about a fixed point $O$; further details can be found in [@Ratiu80; @FedKoz] and the more recent work [@Izosimov]. The configuration of such a body is given by an element $g\in{\mathrm{O}}(n)$ called the attitude matrix that relates an inertial frame in ${\mathbb{R}}^n$ with a body frame rigidly attached to ${\mathcal{B}}$, with both frames having their origin in $O$. The velocity is given by the derivative $\dot g \in T_g{\mathrm{O}}(n)$. Given a motion of the body, a material point ${\mathbf{x}}\in{\mathcal{B}}$ moves along a curve $t\mapsto g(t){\mathbf{x}}\in{\mathbb{R}}^n$, whose velocity at time $t$ is $\dot g(t){\mathbf{x}}$. If the point has mass $m$ then its kinetic energy is $\frac12m\|\dot g{\mathbf{x}}\|^2$. We make use of the left-trivialization of the tangent bundle of ${\mathrm{O}}(n)$: $$\label{eq:left trivialization}
\begin{array}{rcl}
T{\mathrm{O}}(n) & \longrightarrow & {\mathrm{O}}(n)\times{\mathfrak{so}}(n) \\
(g,\dot g) &\longmapsto & (g,\Omega)
\end{array}$$ where $\Omega=g^{-1}\dot g$. The kinetic energy of the motion of the point ${\mathbf{x}}$ above becomes $$\frac12 m\|\dot g{\mathbf{x}}\|^2 = \frac12 m\|\Omega{\mathbf{x}}\|^2.$$
If the body ${\mathcal{B}}$ is a finite collection of particles, then the total kinetic energy is a sum over the constituent particles, while if it is a continuum this becomes an integral; in what follows we write it as an integral. This is conveniently done by introducing the [[***[mass tensor]{}***]{}]{} of the body $${\mathbb{J}}= \int_{\mathcal{B}}{\mathbf{x}}{\mathbf{x}}^T{\mathsf{d}}m({\mathbf{x}}),$$ which is a symmetric $n\times n$ matrix. We will assume the body’s geometry to be such that ${\mathbb{J}}$ is positive definite. The kinetic energy, and hence the Lagrangian, $L:T{\mathrm{O}}(n)\simeq{\mathrm{O}}(n)\times{\mathfrak{so}}(n)\to{\mathbb{R}}$, may be written in terms of the mass tensor as $$L(g,\Omega) = \tfrac12{\mathop\mathrm{tr}\nolimits}(\Omega{\mathbb{J}}\Omega^T).$$ Equivalently, we may write $$\label{eq:Lagrangian}
L(g,\Omega) = \tfrac12 \langle M, \Omega\rangle ,$$ where the [[***[angular momentum]{}***]{}]{} in the body frame $M\in {\mathfrak{so}}(n)$ is given by $$\label{eq:physical inertia}
M={\mathbb{I}}(\Omega) = {\mathbb{J}}\Omega+\Omega{\mathbb{J}},$$ and $\langle \cdot , \cdot \rangle$ denotes the usual invariant pairing on ${\mathfrak{so}}(n)$ $$\label{eq:pairing}
\left<\Lambda,\Omega\right> = \tfrac12{\mathop\mathrm{tr}\nolimits}(\Lambda^T\Omega),$$ that we use to identify $ {\mathfrak{so}}(n)^*$ with $ {\mathfrak{so}}(n)$. The symmetric, positive definite, linear map ${\mathbb{I}}:{\mathfrak{so}}(n)\to{\mathfrak{so}}(n)$ defined by is the [[***[inertia tensor]{}***]{}]{} (the Legendre transform). We will sometimes call ${\mathbb{I}}$ a [[***[physical inertia tensor]{}***]{}]{} when we wish to emphasise that it is not a generic symmetric and positive definite, linear operator on ${\mathfrak{so}}(n)$, but rather one that arises from the existence of a mass tensor ${\mathbb{J}}$ as in . As explained in the introduction, several works in nonholonomic mechanics [@FedJov; @FedJov2; @Jovan; @JovaRubber; @Jova18; @Gajic] depend on assumptions on ${\mathbb{I}}$ that are generally incompatible with it being physical.
A basis $\{f_1,\dots,f_n\}$ of the body frame that diagonalizes the symmetric mass tensor ${\mathbb{J}}$ is known as a [[***[principal basis]{}***]{}]{}, and for such a basis we write $${\mathbb{J}}= {\mathop\mathrm{diag}\nolimits}[J_1,J_2,\dots,J_n].$$ The quantities $J_i+J_j$ are called the [[***[principal moments of inertia]{}***]{}]{} of the body. Suppose $a,b$ are linearly independent vectors of ${\mathbb{R}}^n$: it is not hard to show that $a\wedge b$ is an eigenvector of the inertia operator ${\mathbb{I}}$ if and only if the plane $\Pi$ spanned by $a$ and $b$ is invariant under ${\mathbb{J}}$. Such planes are called the [[***[principal planes]{}***]{}]{} of the body, and it follows from ${\mathbb{J}}$ being symmetric that such a plane must contain two eigenvectors of ${\mathbb{J}}$, and if $J_i$ and $J_j$ are the corresponding eigenvalues, then ${\mathbb{I}}(a\wedge b)=(J_i+J_j)(a\wedge b)$. Given such a principal plane $\Pi$ we say the [[***[moment of inertia]{}***]{}]{} of that plane is ${\mathop\mathrm{in}\nolimits}(\Pi)=J_i+J_j$. If the eigenvalues of ${\mathbb{J}}$ are simple then there are $\frac12 n(n-1)$ such principal planes, but if they are not simple then there are infinitely many.
The Veselova constraint {#sec:constraint}
-----------------------
An $n$-dimensional version of the Veselova constraint was introduced by Fedorov and Kozlov [@FedKoz] and proceeds as follows. Choose a fixed axis in ${\mathbb{R}}^n$ (space), and the (generalized) Veselova constraint allows only rotations in planes containing that axis Note that [@FedJov] considers more general constraints. In the notation of that reference, we only consider the case $r=1$. We call this axis the [[***[distinguished axis]{}***]{}]{}.
In terms of the attitude matrix, let $e_1$ be a unit vector parallel to the distinguished axis, and complete to an orthonormal basis $\{e_1,\dots,e_n\}$ for ${\mathbb{R}}^n$ (space). Fix a principal basis for the body frame $\{f_1,\dots,f_n\}$, and as before denote by $g\in {\mathrm{O}}(n)$ the attitude matrix of the body and by $\Omega = g^{-1}\dot g \in {\mathfrak{so}}(n)$ its angular velocity expressed in these body coordinates. The matrix $\Omega_s:=g\Omega g^{-1}$ describes the angular velocity in the space frame, and the $n$-dimensional Veselova constraint is that $\Omega_s$ is of the form $e_1\wedge w$ for some $w\in{\mathbb{R}}^n$, or equivalently $$\label{eq:constraint-qv}
\Omega=q\wedge v \qquad \mbox{for some} \quad v\in{\mathbb{R}}^n, \quad \mbox{where} \quad q:=g^{-1}e_1.$$
Let us write $${\mathfrak{d}}= e_1\wedge{\mathbb{R}}^n =\mbox{span}\{ e_1\wedge e_2 , \dots, e_1\wedge e_n\}.$$ The nonholonomic constraint distribution $D\subset T{\mathrm{O}}(n)$ is then defined by $\Omega_s\in{\mathfrak{d}}$, or $\dot g \in D_g={\mathfrak{d}}g\subset T_g{\mathrm{O}}(n)$. By construction, $D$ is right invariant and is non-integrable since ${\mathfrak{d}}$ is not a subalgebra of ${\mathfrak{so}}(n)$. A vector $\dot g=g\Omega \in D_g$ if and only if ${\mathrm{Ad}}_g\Omega \in {\mathfrak{d}}$ which is equivalent to having $$\label{eq:constraints}
\langle E_r\wedge E_s , \Omega \rangle =0, \qquad 2\leq r <s \leq n,$$ where $E_r=g^{-1}e_r$ so that ${\mathrm{Ad}}_{g^{-1}}( e_r\wedge e_s)=E_r\wedge E_s$.
The left invariance of the Lagrangian , together with the right invariance of the constraint distribution, signifies that this is an example of an *LR system* as introduced by Veselova [@Veselova]. Following this reference, we write the equations of motion for the system in Euler-Poincaré-Arnold form: $$\label{eq:VesEqns}
\begin{split}
\frac{d}{dt} \left ( {\mathbb{I}}(\Omega) \right ) \ &= \ [{\mathbb{I}}(\Omega) , \Omega ] \ + \ \sum_{2\leq r <s \leq n} \lambda_{rs} E_r\wedge E_s, \\
\frac{d}{dt} \left (E_r\wedge E_s \right ) \ &= \ [E_r\wedge E_s , \Omega ].
\end{split}$$ Here, $[\cdot , \cdot ]$ denotes the matrix commutator in ${\mathfrak{so}}(n)$, and the multipliers $\lambda_{rs}$ are the unique solution to the linear system of equations $$\label{eq:ConstVesEqns}
\sum_{2\leq r <s \leq n}\langle {\mathbb{I}}^{-1} ( E_\rho \wedge E_\sigma ), E_r\wedge E_s \rangle \lambda_{rs} = -\langle {\mathbb{I}}^{-1} ( E_\rho \wedge E_\sigma ) , [{\mathbb{I}}(\Omega) , \Omega ] \rangle, \qquad 2\leq \rho <\sigma \leq n.$$ This choice of $\lambda_{rs}$ guarantees that $\langle E_r\wedge E_s , \Omega \rangle$ are first integrals of for all $2\leq r <s \leq n$. The dynamics of the Veselova system on $D$ is determined by considering the system restricted to the invariant manifold defined by , together with the kinematical equation $\dot g=g\Omega$.
Both the restriction of to , and the system on $D$, preserve the energy $H=\frac{1}{2}\langle {\mathbb{I}}(\Omega), \Omega \rangle$, and possess an invariant measure [@Veselova].
#### Symmetries
Let $G_L={\mathrm{O}}(n-1)$ be the subgroup of ${\mathrm{O}}(n)$ that fixes the distinguished axis, and let $G_R\subset{\mathrm{O}}(n)$ be the group of symmetries of ${\mathbb{J}}$, that is $$\label{eq:def-G_R}
g\in G_R \qquad \mbox{if and only if} \qquad g{\mathbb{J}}g^{-1}={\mathbb{J}}.$$ The natural action of $G=G_L\times G_R$ on the configuration space is, for $(A,B)\in G_L\times G_R$, $(A,B)\cdot g = A\,g\,B^{-1}$. The tangent lift of this to the tangent bundle is $$(A,B)\cdot (g,\dot g) = (A\,g\,B^{-1},\;A\,\dot g\,B^{-1}).$$ Equivalently, using the left trivialization , $$\label{eq:lifted action}
(A,B)\cdot (g,\Omega) = (A\,g\,B^{-1},\;B\,\Omega\,B^{-1}).$$ It is clear that the Lagrangian is invariant under this action, and furthermore, because $G_L$ fixes the distinguished axis, the constraint distribution $D$ is also invariant under this action. Since the only ingredients that determine the dynamics of the Veselova top are the Lagrangian and the distribution, it follows that the dynamics of the system is equivariant with respect to the action of $G=G_L\times G_R$.
Steady rotations {#sec:Steady-rotations}
----------------
We now show the Veselova system admits [[***[steady rotation solutions]{}***]{}]{}, which we define to be motions with constant angular velocity. We show such motions are necessarily periodic solutions in which the body steadily rotates in a principal plane of inertia of the body and the constraint forces vanish. In view of the nonholonomic constraint, the orientation of the body along these solutions is such that the principal plane of rotation contains the distinguished axis.
We recall that the $n$-dimensional free rigid body also allows steady rotation solutions, and that these are in general quasi-periodic if $n>3$. The periodicity of the steady rotations of the $n$-dimensional Veselova system is a consequence of the nonholonomic constraint that forces the angular velocity matrix to have rank 2.
\[prop:steady rotations on D\] The $n$-dimensional Veselova system admits steady rotation solutions. These solutions are all steady rotations in principal planes, and are characterised by the condition that the angular velocity matrix $\Omega$ satisfies $ [{\mathbb{I}}(\Omega) , \Omega ]=0$. The steady rotations in the principal plane $\Pi$ are periodic with period $2\pi/\|\Omega\|$, energy $H=\frac{1}{2}{\mathop\mathrm{in}\nolimits}(\Pi)\|\Omega\|^2$ and momentum $M$ satisfying $\|M\|^2=({\mathop\mathrm{in}\nolimits}(\Pi))^2\|\Omega\|^2$. Here $\| \cdot \|$ denotes the norm in ${\mathfrak{so}}(n)$ induced by the pairing .
It may seem the definition of steady rotation is ambiguous since it was not specified whether it is the angular velocity in the body or in space that should be constant. In fact the two are equivalent: write $\Omega$ and $\Omega_s$ for the two angular velocities. If $\Omega$ is constant, then the motion is given by $g(t)=g_0\exp(t\Omega)$ for some $g_0\in{\mathrm{O}}(n)$. Then $\Omega_s=\dot g g^{-1} = g_0\Omega g_0^{-1}$, which is constant. The argument is easily reversed to show the equivalence.
First we show the existence of such solutions. If $\Omega$ is a constant rank two matrix that defines a rotation on a principal plane then $[{\mathbb{I}}(\Omega),\Omega]=0$ and the multipliers $\lambda_{rs}$ defined by vanish, so the first equation in holds since $\Omega$ is constant. The motion on ${\mathrm{O}}(n)$ is then $g(t)=g_0\exp (\Omega t)$ for a constant $g_0\in {\mathrm{O}}(n)$, and is periodic with the stated period since $\Omega$ has rank 2. The formulas for $H$ and $\| M\|^2$ follow immediately from ${\mathbb{I}}(\Omega)={\mathop\mathrm{in}\nolimits}(\Pi) \Omega$.
To complete the proof that $(g(t),\Omega)$ is a solution we should check that the nonholonomic constraints are satisfied along the motion. We have ${\mathrm{Ad}}_{g(t)}\Omega= {\mathrm{Ad}}_{g_0}\Omega$ so the constraints are indeed satisfied provided that the initial orientation of the body $g_0$ satisfies ${\mathrm{Ad}}_{g_0}\Omega \in {\mathfrak{d}}$. This is accomplished in the following way. If $\Omega=a\wedge b$ then $g_0\in {\mathrm{O}}(n)$ should be such that $g_0^{-1}e_1$ lies on the plane spanned by $a$ and $b$; that is, the distinguished axis should lie on the plane of rotation.
Next, if $(g(t),\Omega(t))$ is a solution for which $[{\mathbb{I}}(\Omega(t)),\Omega(t)]=0$, then the multipliers $\lambda_{rs}$ defined by vanish, and the first equation in implies that $\Omega$ is constant.
Suppose now that $(g(t),\Omega)$ is a solution with constant $\Omega \neq 0$. It remains to show that $\Omega$ defines a rotation on a principal plane (the condition that $[{\mathbb{I}}(\Omega),\Omega]=0$ follows immediately from this). From , $\Omega=q\wedge v$ for some $v \in{\mathbb{R}}^n$ which can be chosen to be orthogonal to $q$. Recall that $q=g^{-1}e_1$. Then differentiating shows that $\dot q = -g^{-1}(g\Omega)q = -(q\wedge v)q=v$. On the other hand, given that $\Omega$ is constant we have $$0=\dot \Omega= \dot q \wedge v + q\wedge \dot v = q\wedge \dot v,$$ which shows that $\dot v$ is parallel to $q$. These observations, together with the boundedness of $q$, imply that the vectors $q(t)$ and $v(t)$ describe simple harmonic motion on the plane that they span, which remains constant throughout the motion. We will now prove that this plane is invariant under ${\mathbb{J}}$ which is equivalent to $\Omega$ defining a rotation on a principal plane.
Using the definition of ${\mathbb{I}}$ given in , one shows that $[{\mathbb{I}}(\Omega),\Omega] = [{\mathbb{J}},\Omega^2]$, and with $\Omega=q\wedge v$ this becomes $$\label{eq:commutator-steady rot}
[{\mathbb{I}}(\Omega),\Omega] = -\|v\|^2({\mathbb{J}}q q^T - qq^T{\mathbb{J}}) -({\mathbb{J}}vv^T-vv^T{\mathbb{J}}).$$ Since $\Omega$ is constant, then so is ${\mathbb{I}}(\Omega)$, and the first of the equations of motion becomes $$0 \ = \ [{\mathbb{I}}(\Omega),\Omega] \mod q^\perp\wedge q^\perp,$$ where $q^\perp$ is the subspace orthogonal to $q$. Applying both sides of the above relation to $q$ and using gives $$\label{eq:Jq-steady rotation}
0 \ = \ -\|v\|^2{\mathbb{J}}q - \|v\|^2(q\cdot{\mathbb{J}}q)q+(q\cdot {\mathbb{J}}v)v,$$ where $\cdot$ is the Euclidean scalar product in ${\mathbb{R}}^n$. This shows that ${\mathbb{J}}q$ is in the span of $q$ and $v$ for all time. The harmonic dynamics of $q(t)$ and $v(t)$ described above imply that the same is true about ${\mathbb{J}}v$.
First reduction of the general Veselova top {#sec:1st reduction}
===========================================
We now perform the reduction of the system by the group $G_L={\mathrm{O}}(n-1)$ introduced above. This corresponds to invariance of the system with respect to the rotations and reflections of the space frame that fix the distinguished axis. This reduction was first considered in [@FedJov Section 5] in a more general framework. Here we specialise some of their results and give their expressions in terms of the mass tensor ${\mathbb{J}}$ of the body, which is necessary for our purposes. The geometry is clarified by the decomposition introduced in [@FedJov], ${\mathfrak{so}}(n)= {\mathfrak{d}}\oplus {\mathfrak{h}}$ where ${\mathfrak{d}}$ is defined above and ${\mathfrak{h}}$ is given by $${\mathfrak{h}}= \mbox{span}\{ e_r\wedge e_s \, : \, 2\leq r< s\leq n\},$$ where $e_1$ is a unit vector along the distinguished axis. This is the orthogonal complement of ${\mathfrak{d}}$ with respect to the Killing metric given above. Note moreover that ${\mathfrak{h}}$ is the Lie algebra of the subgroup $G_L={\mathrm{O}}(n-1)$.
The splitting ${\mathfrak{so}}(n)={\mathfrak{d}}\oplus {\mathfrak{h}}$ implies that we have a generalized Chaplygin system [@Koi; @BKMM]. The reduced space $D/{\mathrm{O}}(n-1)$ is isomorphic to the tangent bundle $T{\mathrm{S}}^{n-1}$. An isomorphism may be constructed by viewing $D$ as the horizontal space of a connection on the principal ${\mathrm{O}}(n-1)$-bundle ${\mathrm{O}}(n)\to {\mathrm{S}}^{n-1}$, as described in [@Koi]. Explicitly, this map $D\to T{\mathrm{S}}^{n-1}$ is given by $$(g,\Omega) \longmapsto (q,v)$$ where $q=g^{-1}e_1$ and $v$ is the unique vector orthogonal to $q$ satisfying $$\label{eq:Omega and v}
\Omega=q\wedge v.$$ Note that with this definition, $(q,v)$ satisfies $\|q\|=1$ and $q\cdot v=0$ showing that $(q,v)$ is indeed in $T{\mathrm{S}}^{n-1}$.
The Legendre transform provides a further isomorphism of $T{\mathrm{S}}^{n-1}$ with $T^*{\mathrm{S}}^{n-1}$ and in our case is given as follows.
\[prop:Legendre\] For the Lagrangian given in , the Legendre transform $T{\mathrm{S}}^{n-1}\to T^*{\mathrm{S}}^{n-1}$ is given by $(q,v) \longmapsto (q,p)$, where $$\begin{aligned}
\label{eq:Legendre}
p=-{\mathbb{I}}(\Omega)q.\end{aligned}$$ and $\Omega$ is given in .
Note that since ${\mathbb{I}}(\Omega)$ is skew symmetric, it follows that $p\cdot q=0$, which means $T^*{\mathrm{S}}^{n-1}$ is realized as the submanifold of ${\mathbb{R}}^{2n}$ defined by the conditions $$\label{eq:qp-conds}
\|q\|=1, \qquad p\cdot q=0.$$
The fibre derivative of the Lagrangian $$L(q,v) = \frac12\left<{\mathbb{I}}(q\wedge v),(q\wedge v)\right>$$ in the direction $w$ is $$\frac12\left<{\mathbb{I}}(q\wedge v),(q\wedge w)\right>+\frac12\left<{\mathbb{I}}(q\wedge w),(q\wedge v)\right>.$$ By the symmetry of ${\mathbb{I}}$ this is equal to $\left<{\mathbb{I}}(q\wedge v),(q\wedge w)\right>$. Writing $\Omega=q\wedge v$ this is $$\begin{aligned}
-\frac12{\mathop\mathrm{tr}\nolimits}\left({\mathbb{I}}(\Omega)(qw^T-wq^T)\right) &=&
-\frac12{\mathop\mathrm{tr}\nolimits}\left({\mathbb{I}}(\Omega)(qw^T)\right) + \frac12{\mathop\mathrm{tr}\nolimits}\left({\mathbb{I}}(\Omega)(wq^T)\right) \\
&=& -\frac12{\mathop\mathrm{tr}\nolimits}\left(w^T{\mathbb{I}}(\Omega)q\right) + \frac12{\mathop\mathrm{tr}\nolimits}\left(q^T{\mathbb{I}}(\Omega)w\right) \\
&=& -{\mathbb{I}}(\Omega)q\cdot w.\end{aligned}$$ This shows that $p=-{\mathbb{I}}(\Omega)q$ as required.
Composing the two isomorphisms gives $$D/{\mathrm{O}}(n-1)\longrightarrow T^*{\mathrm{S}}^{n-1},\quad (g,\Omega) \longmapsto (q, p),$$ where $q=g^{-1}e_1$ and $p=-{\mathbb{I}}(\Omega)q$. This map is easily shown to be well-defined.
Following the action of $G=G_L\times G_R$ on $D$ (see Section\[sec:constraint\]) through the reduction process above it is easy to see that,
\[coroll:symmetry on T\^\*S\] The symmetry $G=G_L\times G_R$ of the Veselova top descends to the natural action of $G_R$ on the spaces $T{\mathrm{S}}^{n-1}$ and $T^*{\mathrm{S}}^{n-1}$; that is $g\in G_R$ acts by $g\cdot(q,v) = (gq,gv)$ and $g\cdot(q,p) = (gq,gp)$.
The following lemma is a particular instance of [@FedJov Theorem 5.4]:
\[lemma:q,p constant\] The vectors $q$ and $p$ are constant in the space frame, that is, they satisfy $$\label{eq:qdot}
\dot q=-\Omega q,\qquad \dot p = -\Omega p.$$
Since $q=g^{-1}e_1$, in space it becomes $e_1$ which is fixed by definition (alternatively, differentiating $q=g^{-1}e_1$ shows $\dot q = -\Omega q$). Next, starting with the relation $p=-{\mathbb{I}}(\Omega)q$ and using shows that $$\label{eq:pdot}
\dot p =\left ( -\frac{d}{dt} ({\mathbb{I}}(\Omega)) + {\mathbb{I}}(\Omega) \Omega \right ) q = \Omega {\mathbb{I}}(\Omega) q = -\Omega p,$$ where we have used $(E_r\wedge E_s)q=0$ for all $2\leq r <s \leq n$.
Combining the first equation in with Equation shows that along the motion $$\label{eq:Omqqdot}
\Omega= q\wedge \dot q.$$ We will use this repeatedly in what follows.
Let $C$ be the $n\times n$ matrix defined by, $$\label{eq:Cmatrix}
C=C(q) = \left({\mathbb{J}}+ (q\cdot{\mathbb{J}}q)\mathrm{Id}_n\right)^{-1},$$ where $\mathrm{Id}_n$ denotes the identity matrix. (It is easy to see the matrix in brackets is diagonal and positive definite and hence invertible.) The following proposition gives the explicit form of the reduced equations for a general physical inertia tensor (i.e., one derived from a mass tensor ${\mathbb{J}}$).
\[prop:eqns-nD\] The reduced equations of motion on $T^*{\mathrm{S}}^{n-1}$ are $$\begin{aligned}
\label{eq:RedEqnsGeneral}
\left\{\begin{aligned}
\dot q &= C\left[ p - \left(\frac{p\cdot Cq}{q\cdot Cq}\right)q\right], \\
\dot p &=-2H(q,p)q.
\end{aligned}\right.\end{aligned}$$ Here, $$\label{eq:energy-general}
H(q,p)=\frac{1}{2} \left (p\cdot Cp -\frac{(p\cdot Cq)^2}{ q\cdot Cq }\right ) ,$$ is the energy integral. Furthermore, this system is invariant under the action of $G_R$ described in Corollary \[coroll:symmetry on T\^\*S\].
In view of we have $$p=-{\mathbb{I}}(\Omega)q= -({\mathbb{J}}(q\wedge \dot q) + (q\wedge \dot q) {\mathbb{J}})q.$$ We would like to solve the above equation for $\dot q$ in terms of $q$ and $p$. This is not possible for general vectors in ${\mathbb{R}}^n$ since the right hand side is a linear expression in $\dot q$ that vanishes whenever $\dot q$ is parallel to $q$. However, on $T^*{\mathrm{S}}^{n-1}$ the equations hold and, since $\dot q\cdot q=0$, we may uniquely write $\dot q$ as in the first equation in .
Using that $p\cdot q=0$ along $T^*{\mathrm{S}}^{n-1}$, the equation $\dot p =-\Omega p$ (see lemma above) simplifies to $\dot p=- (\dot q\cdot p) q=-2H(q,p) q$ which is the second equation in .
It remains to show that $H(q,p)$ given by coincides with the energy $\frac{1}{2}\langle {\mathbb{I}}(\Omega), \Omega \rangle $. Using once again we obtain $$\tfrac{1}{2}\langle {\mathbb{I}}(\Omega), \Omega \rangle = -\tfrac{1}{2} {\mathop\mathrm{tr}\nolimits}( {\mathbb{J}}(q\wedge \dot q)^2)=\tfrac{1}{2}C^{-1}\dot q \cdot \dot q=\tfrac{1}{2}p \cdot \dot q =H(q,p).$$ Finally, to see the invariance, note that implies that $C(gq) = C(q)$ for all $g\in G_R$.
Lemma \[lemma:q,p constant\] above allows us to conclude the following proposition (which is an instance of a general result on kinetic LR systems given in [@FedJov Proposition 2.2]).
\[prop:conservation of P\] The square $\|p\|^2$ of the momentum is conserved under the dynamics on $T^*{\mathrm{S}}^{n-1}$.
This is clear from , since $p\cdot \dot p = -2H p\cdot q = 0$.
Another application of Lemma\[lemma:q,p constant\] is the following reconstruction formula, valid for the 3-dimensional system, and found in a slightly different form in [@Veselov-Veselova-1988] (not using $p$).
\[prop:reconstruction 3D\] In dimension 3, if $(q(t),p(t))$ is a solution of the reduced equations of motion on $T^*{\mathrm{S}}^2$, then for any $g_0\in G_L={\mathrm{O}}(2)$ the curve $$\label{eq:3d-reconst-g}
g(t)=g_0\begin{pmatrix}
q(t)^T\cr
\frac1{\|p(t)\|}p(t)^T \\
\frac1{\|p(t)\|}(q(t)\times p(t))^T
\end{pmatrix}$$ is a solution to the full Veselova system, and all solutions with $p\neq0$ can be written in this way.
We know $q,p$ are constant in space, whence $q\times p$ is also constant in space. Thus each row satisfies $\dot r = r\Omega$ and hence $\dot g = g\Omega$. Furthermore, for such $g_0$, the velocity $\dot g$ lies in the distribution $D$. This is because the angular velocity in space $\Omega_s$ satisfies $$\Omega_s = \dot g(t) g^{-1}(t) = g_0
\begin{pmatrix}
0&2H/\|p\| & (*)\\
-2H/\|p\| & 0 & 0 \\
(*) & 0 & 0
\end{pmatrix}g_0^{-1}$$ and since the lower left $2\times 2$ block vanishes, and this is preserved by $g_0$, $\Omega_s$ satisfies the constraint.
Reduced dynamics of the general Veselova top
--------------------------------------------
With the equations of motion above, we are able to discuss some aspects of the reduced dynamics of the general (physical) $n$-dimensional Veslova top.
Note that since the system on $D$ is an LR system it follows from a theorem of Fedorov and Jovanović [@FedJov Theorem 3.3] that Equations possess a smooth invariant measure on $T^*{\mathrm{S}}^{n-1}$; namely the push forward of the invariant measure on $D$ mentioned already. A formula for this volume form, valid in a more general framework than the one we consider, is given in [@FedJov Theorem 5.5]. The following proposition gives a non-trivial simplification of this formula to our setting.
\[prop:measure-nD\] The reduced equations of motion on $T^*{\mathrm{S}}^{n-1}$ have invariant measure $$\sqrt{\frac{\det (C)}{q\cdot C q}} \, \sigma,$$ where $\sigma$ is the Liouville volume form on $T^*{\mathrm{S}}^{n-1}$.
The Liouville form $\sigma$ on $T^*{\mathrm{S}}^{n-1}$ coincides with the volume form on $T^*{\mathrm{S}}^{n-1}$ inherited from the Euclidean measure $dq\wedge dp$ in the ambient space ${\mathbb{R}}^{2n}$. To prove the result we compute the divergence of the vector field on ${\mathbb{R}}^{2n}$ defined by the equations with respect to $\mu(q) \, dq\wedge dp$, where $\mu(q)=\sqrt{\frac{\det (C)}{q\cdot C q}}$, and show that its restriction to $T^*{\mathrm{S}}^{n-1}$ vanishes. Given that $\mu$ is independent of $p$, this is equivalent to showing that $$y(q,p):= \mu\sum_{i=1}^n \left ( \frac{\partial \dot q^i}{\partial q^i}+ \frac{\partial \dot p_i}{\partial p_i}\right ) + \frac{\partial \mu}{\partial q} \cdot \dot q$$ is zero along $T^*{\mathrm{S}}^{n-1}$. In the calculations below we will repeatedly use that the matrices ${\mathbb{J}}$ and $C$ are symmetric and commute (which is obvious since they are diagonal).
First notice that implies $$\sum_{i=1}^n \frac{\partial \dot p_i}{\partial p_i} = -2 \frac{\partial H}{\partial p}\cdot q = - 2\dot q \cdot q,$$ which vanishes along $T^*{\mathrm{S}}^{n-1}$. Next, by direct differentiation of $\mu$ and $\dot q^i$ defined by one obtains $$\begin{split}
\frac{\partial \mu}{\partial q}& = \mu \left (- X(q) {\mathbb{J}}q - \frac{Cq}{q \cdot Cq} \right ), \\
\sum_{i=1}^n \frac{\partial \dot q^i}{\partial q^i}&= \left ( C \left ( -2{\mathbb{J}}+ \frac{2(q \cdot {\mathbb{J}}Cq) -1 }{q \cdot Cq} \mbox{Id}_n \right ) q \right ) \cdot \dot q -
X(q) \frac{q \cdot Cp}{q \cdot Cq},
\end{split}$$ where $X(q):=\mbox{tr}(C) - \frac{q\cdot C^2 q}{q\cdot Cq}$. Therefore, along $T^*{\mathrm{S}}^{n-1}$ we may write $$\label{eq:aux-proof-inv-measure}
y(q,p)= - 2\mu \left ( C \left ( {\mathbb{J}}+ \frac{1- (q \cdot {\mathbb{J}}Cq) }{q \cdot Cq} \mbox{Id}_n \right ) q \right ) \cdot \dot q - \mu X(q)
\left ( {\mathbb{J}}q \cdot \dot q + \frac{q \cdot Cp}{q \cdot Cq} \right ).$$
We now make use of the equality $${\mathbb{J}}C k \cdot q +(Ck \cdot q)({\mathbb{J}}q \cdot q) = k\cdot q,$$ which holds for all $k\in {\mathbb{R}}^n$ and is easily established. By respectively applying it to $k=q$ and $k=p$ shows that along $T^*{\mathrm{S}}^{n-1}$ we have $$q \cdot {\mathbb{J}}Cq =1-(q \cdot Cq) (q \cdot {\mathbb{J}}q), \qquad \mbox{and} \qquad q \cdot {\mathbb{J}}C p = -( q \cdot Cp) (q \cdot {\mathbb{J}}q).$$ Using these identities and the expression for $\dot q$ given in one deduces the formulae $${\mathbb{J}}+ \frac{1- (q \cdot {\mathbb{J}}Cq) }{q \cdot Cq} \mbox{Id}_n = C^{-1}, \qquad {\mathbb{J}}q \cdot \dot q =- \frac{q \cdot Cp}{q \cdot Cq},$$ that are valid along $T^*{\mathrm{S}}^{n-1}$. Therefore, implies $$\left . y(q,p) \right |_{T^*{\mathrm{S}}^{n-1}} = - 2\mu q \cdot \dot q =0.$$
\[rmk:hamiltonization\] It is unknown whether the reduced system can be Hamiltonized by Chaplygin’s reducing multiplier method. This would amount to the introduction of a time reparametrization $dt = \nu(q) \, d\tau$, and a rescaling of the momenta $p\to \nu(q)p$, that would transform the equations into a Hamiltonian system. The factor $\nu(q)$, if it exists, is the so-called Chaplygin multiplier. The specific form of the invariant measure implies that, up to a constant factor, the multiplier $\nu(q)$ would be given by $$\nu(q) = \left ( \frac{\det C}{q\cdot Cq} \right )^{-\frac{1}{2(n-2)}}.$$ The details about the method may be found in e.g. [@FedJov; @EhlersKoiller]. For $n=3$ the Hamiltonization is possible and follows from Chaplygin’s last multiplier theorem [@ChapRedMult]. For $n\geq 4$, using the above expression for $\nu(q)$, we were able to verify the Hamiltonization of the system for axisymmetric tops (treated in Section \[sec:axisymmetric\]). This in fact follows from the work of Fedorov and Jovanović [@FedJov; @FedJov2]. On the other hand, some preliminary investigations of us suggest that the system is [*not*]{} Hamiltonizable for $n\geq 4$ and more general mass tensor ${\mathbb{J}}$.
Steady rotations {#steady-rotations}
----------------
\[prop:steady rotations on T\^\*S\] The (non-trivial) steady rotation solutions on $D$ described in Proposition \[prop:steady rotations on D\] project to periodic solutions of the reduced system on $T^*{\mathrm{S}}^{n-1}$, where $q$ and $p$ rotate uniformly in a principal plane of the body. The projected motion has the same period and the same energy as the motion on $D$.
We prove the stability of some of these steady rotations in Theorem\[thm:stability\] below.
Let ${\mathbb{J}}=\mbox{diag}[J_1,\dots, J_n]$ be the mass tensor and consider a steady rotation solution with angular velocity $\Omega$. In view of Proposition\[prop:steady rotations on D\], by rotating the body frame if necessary, we may assume that $\Omega=\omega f_i\wedge f_j$, where $\omega = \| \Omega \| \in {\mathbb{R}}$. Then $p=-{\mathbb{I}}(\Omega)q=-(J_i+J_j)\Omega q=(J_i+J_j)\dot q$. It follows from the second equation in that the motion of $q$, $p$, is simple harmonic with period $2\pi\sqrt{(J_i+J_j)/2H}=2\pi/|\omega|$. Moreover, $q$ and $p$ are contained in the plane spanned by $f_i$ and $f_j$ in view of . That the energies are equal is obvious, by the definition of $H$.
Consider the integral of motion $P=\| p\|^2$ and the resulting energy-momentum map $$(P,H) : T^*{\mathrm{S}}^{n-1}\to {\mathbb{R}}^2.$$ The following proposition refers to the singularities of this map. Since $H$ and $P$ are both non-negative, the image lies in the (closed) positive quadrant ${\mathbb{R}}_+^2\subset{\mathbb{R}}^2$. If $p=0$ then $(P,H)(q,0)=(0,0)$ and such points are all critical points of $(P,H)$. As we see in the following proposition, the other critical points occur at the steady rotations.
\[prop:EM critical rays\] Let ${\mathbb{J}}= {\mathop\mathrm{diag}\nolimits}[J_1, \dots, J_n]$ with $J_j>0$ for all $j$.
1. The critical points of the energy-momentum map $(P,H)$ with $p\neq0$ occur at the steady rotations described above.
2. The critical values of the energy-momentum map $(P,H)$ are the rays $L_{ij}$ given by $P=2(J_i+J_j)H$, for $i<j$.
The image and critical values of the energy-momentum map are illustrated in Figure\[fig:E-M for Veselova\]. Let us emphasize that, in the case of eigenvalues with multiplicity, this statement does not assume $J_i\neq J_j$, only that $i\neq j$. Moreover, recall from Section \[sec:rigid body\] that a rotation of the body in the principal plane $\Pi_{i,j}$ has angular velocity $\Omega$ equal to a multiple of $f_i\wedge f_j$, and this is an eigenvector of the inertia operator with eigenvalue ${\mathop\mathrm{in}\nolimits}(\Pi_{i,j})=J_i+J_j$.
(0,0) – (1.6,0) node\[anchor=north\][$P$]{}; (0,0) – (0,1.6) node\[anchor=east\][$H$]{}; (0,0) – (0.4,1.6) – (1.6,0.4); (0,0) – (0.4,1.6); (0,0) – (0.6,1.4); (0,0) – (1,1); (0,0) – (1.2,0.8); (0,0) – (1.5,0.5); (0,0) – (1.6,0.4);
\(i) Let $(q_0,p_0)\in T^*{\mathrm{S}}^{n-1}$ with $p_0\neq0$. Writing $H_0=H(q_0,p_0)$ and $P_0=\|p_0\|^2$, this implies $H_0>0$ and $P_0>0$. At such points, neither $dH=0$ nor $dP=0$. Thus, considering $H$ and $P$ as functions on the ambient space ${\mathbb{R}}^n\times{\mathbb{R}}^n$, the condition for a critical point at $(q_0,p_0)$ is that the gradients of $H$ and $P$ are linearly dependent (on $T^*{\mathrm{S}}^{n-1}$), and using Lagrange multipliers for the constraints defining this cotangent bundle gives $$\label{eq:LagMult1}
\nabla H = \lambda_1 \nabla P + \lambda_2 \nabla (p\cdot q)+\lambda_3 \nabla \|q\|^2,$$ and $(q_0,p_0)$ are such that $$\label{eq:CondExtrema}
q_0\cdot p_0=0, \qquad
\|q_0\|^2=1, \qquad
H(q_0,p_0)=H_0, \qquad \|p_0\|^2=P_0.$$ At such points, is equivalent to $$\label{eq:LagMult1bis}
\nabla H(q_0,p_0) = 2\lambda_1(0,p_0) + \lambda_2 (p_0,q_0)+2\lambda_3(q_0,0).$$ Denote by $$v (q,p):=\dot q= Cp- \frac{p\cdot Cq}{q\cdot Cq}Cq.$$ One calculates $$\frac{\partial H}{\partial p} =v, \qquad \frac{\partial H}{\partial q}=-\|v\|^2 {\mathbb{J}}q - \frac{(p\cdot Cq)}{(q\cdot Cq)}v.$$ Then can be rewritten as 2 equations,
\[eq:LagMult2\] $$\begin{aligned}
\label{eq:LagMult2a}
v_0 & = 2\lambda_1 p_0 + \lambda_2 q_0, \\
\label{eq:LagMult2b}
- \left(\frac{p_0\cdot Cq_0}{q_0\cdot Cq_0}\right)\,v_0 -\|v_0\|^2 {\mathbb{J}}q_0 &= \lambda_2 p_0 + 2\lambda_3 q_0,\end{aligned}$$
where $v_0=v(q_0,p_0)$. Taking the inner product on both sides of equation with $q_0$ and $p_0$, and using , leads to the relations: $$\label{eq:Lagrange multipliers 1,2}
\lambda_2=0, \qquad \lambda_1=\frac{H_0}{P_0}.$$ Substituting this into and taking the inner product with $q_0$ gives $\lambda_3 = -\frac12\|v_0\|^2(q_0\cdot{\mathbb{J}}q_0).$ Putting first $\lambda_2=0$ in shows that $v_0=2(H_0/P_0)p_0$, and secondly taking the inner product of with $p_0$ shows $$\lambda_3=-2\frac{H_0^2}{P_0}(q_0\cdot{\mathbb{J}}q_0),\quad\text{and}\quad
\frac{p_0\cdot Cq_0}{q_0\cdot Cq_0}=-\frac{2H_0}{P_0}(p_0\cdot {\mathbb{J}}q_0).$$ Now consider ${\mathbb{J}}q_0$ and ${\mathbb{J}}p_0$. First from , $v_0=(2H_0/P_0)p_0$ and then from we deduce that ${\mathbb{J}}q_0$ is a linear combination of $q_0$ and $p_0$. Again from , substituting for the definition of $v_0$, and using the definition of the matrix $C$ one finds $${\mathbb{J}}p_0 = \left(\frac{P_0}{2H_0}-(q_0\cdot{\mathbb{J}}q_0)\right)p_0 + \frac{P_0}{2H_0}\left(\frac{p_0\cdot Cq_0}{q_0\cdot Cq_0}\right)q_0,$$ which is also a linear combination of $p_0$ and $q_0$. In this way we have shown that the plane spanned by $q_0,p_0$ is invariant under ${\mathbb{J}}$, which is precisely the condition for it to be a principal plane.
Finally, if $p_0,q_0$ lie in a principal plane, and $p_0\neq0$, they define a steady rotation with $\Omega= \frac{2H_0}{P_0}q_0\wedge p_0$ as is seen from Equations , and .
\(ii) Without loss of generality, we may assume that the body frame $\{f_1, \dots , f_n\}$ is chosen such that ${\mathbb{J}}={\mathop\mathrm{diag}\nolimits}[J_1, \dots J_n]$ and $\mbox{span} \{q_0, p_0\} =\mbox{span} \{f_i, f_j\} $ ($i\neq j$). We may then write $$q_0=\cos \alpha f_i + \sin \alpha f_j, \qquad p_0= \sqrt{P_0}\left ( -\sin \alpha f_i + \cos \alpha f_j \right ),$$ for a certain $\alpha \in [0,2\pi)$. The relation $P_0=2H_0(J_i+J_j)=2H_0{\mathop\mathrm{in}\nolimits}(\Pi_{i,j})$ follows from Proposition\[prop:steady rotations on D\], since $P=\|M\|^2$.
#### Stability of steady rotations.
Recall that for a dynamical system, a compact subset $S$ of the phase space is Lyapunov stable if for each neighbourhood $U$ of $S$ there is a neighbourhood $V$ such that any trajectory that enters $V$ lies entirely in $U$. For the free $n$-dimensional rigid body, Izosimov [@Izosimov] has produced a fairly complete analysis of the stability of the relative equilibria. In particular, in Example 2.3 he shows that if ${\mathbb{J}}$ has simple eigenvalues, with $J_1<J_2<\dots<J_n$ then the 2-dimensional rotations in the principal plane $\Pi_{i,j}$ are (Lyapunov) stable if and only if $|i-j|=1$ (that is, if $J_i$ and $J_j$ are adjacent in the ordering). It is not known whether the same result holds with the Veselova constraint. However, we do have the following result which follows from the geometry of the energy-momentum map. We do not require the eigenvalues of ${\mathbb{J}}$ to be simple (motivated by the symmetric bodies considered in later sections).
Recall that if $\Pi$ is a principal plane of inertia then there is a corresponding moment of inertia, which we denote ${\mathop\mathrm{in}\nolimits}(\Pi)$; if $\Pi=\Pi_{i,j}$ for some principal basis then ${\mathop\mathrm{in}\nolimits}(\Pi_{i,j})=J_i+J_j$.
\[def:extremal\] A principal plane $\Pi$ is [[***[extremal]{}***]{}]{} if either ${\mathop\mathrm{in}\nolimits}(\Pi)\leq{\mathop\mathrm{in}\nolimits}(\Pi')$ for all principal planes $\Pi'$ (in which case it is minimal) or ${\mathop\mathrm{in}\nolimits}(\Pi)\geq{\mathop\mathrm{in}\nolimits}(\Pi')$ for all principal planes $\Pi'$ (in which case it is maximal).
If the eigenvalues of ${\mathbb{J}}$ are all simple and ordered as usual, then the only minimal principal plane is $\Pi_{1,2}$ and the only maximal one is $\Pi_{n-1,n}$. If some of the eigenvalues of ${\mathbb{J}}$ are not simple, then the dynamics has a further symmetry $G_R$ described by Equation, and one does not in that case expect a steady rotation to be Lyapunov stable. Instead the natural notion is $G_R$-Lyapunov stable, which is defined by saying that a $G_R$-invariant subset $S$ of phase space is $G_R$-Lyapunov stable if for every $G_R$-invariant neighbourhood $U$ of $S$ there is a $G_R$-invariant neighbourhood $V$ of $S$ such that any trajectory that intersects $V$ is entirely contained in $U$.
\[thm:stability\] Steady rotations in extremal principal planes are $G_R$-Lyapunov stable.
If as above, the eigenvalues are simple then the statement reduces to saying the steady rotations in extremal principal planes are Lyapunov stable. Before the proof of the theorem, we state the following immediate consequence.
\[cor:stability\] The steady rotations of the $3D$ Veselova top about the largest and smallest principal axes of inertia are Lyapunov stable (${\mathrm{O}}(2)$-Lyapunov stable for axisymmetric tops).
We shall prove this for minimal principal planes of inertia, and leave the maximal case to the reader. We define a function on the phase space which, under the hypotheses of the theorem, is a Lyapunov function. Let $(q_0,p_0) \in \Pi$ be a point of steady rotation, and let $f:T^*{\mathrm{S}}^{n-1}\to{\mathbb{R}}$ be defined by $$f(q,p) = H-\lambda P - \tfrac12(P-c)^2,$$ where $\lambda=(2{\mathop\mathrm{in}\nolimits}(\Pi))^{-1}$ and $c>0$ is a constant to be fixed. This function is clearly invariant under the dynamics, and is moreover invariant under the action of $G_R$. Let $S$ be the $G_R$-orbit of the trajectory through $(q_0,p_0)$. We show that (i) $f$ has a critical point at every point of $S$, and (ii) under the condition that $\Pi$ is minimal, the Hessian matrix of $f$ transverse to $S$ is negative definite. It then follows that for $\varepsilon>0$ the set $$V_\varepsilon := \{(q,p)\in T^*{\mathrm{S}}^{n-1} \mid f(q,p) > f(q_0,p_0)- \varepsilon\}$$ is a $G_R$-invariant neighbourhood of $S$, and that any $G_R$-invariant neighbourhood $U$ of $S$ contains $V_\varepsilon$ for some $\varepsilon>0$, and we can deduce the $G_R$-Lyapunov stability of $S$.
For (i), since $f$ is $G_R$-invariant, it suffices to prove it has a critical point at any point of the trajectory. Now, writing $T_0$ for the tangent space $T_0=T_{(q_0,p_0}(T^*{\mathrm{S}}^{n-1})$, we have $$df = (dH-\lambda dP)_{\restr{T_0}} - (P-c)dP_{\restr{T_0}}.$$ The first term vanishes by the proof of Proposition\[prop:EM critical rays\] above, and in particular the value of $\lambda$ is given by $\lambda_1$ in that proof—see . The second term vanishes when $P=c$, so from now on we assume $c=\|p_0\|^2.$
We now wish to compute the Hessian of $f$ at $(q_0,p_0)$ in directions transverse to the submanifold $S$. Let us choose a principal basis so that $\Pi=\mathrm{span}\{f_1,f_2\}$, in which case $$q_0=\cos\alpha f_1+\sin\alpha f_2,\quad p_0=\|p_0\|(-\sin\alpha f_1+\cos\alpha f_2).$$ We can further simplify the calculations by noting that since $f$ is invariant under the dynamics, the signature of the Hessian will also be invariant (by Sylvester’s law of inertia), so we can suppose $\alpha=0$, and we take $$\label{eq:q0 & p0}
q_0= f_1,\quad p_0=\|p_0\| f_2.$$
Now we consider a slice $V$ (of dimension $2n-3$) to the trajectory in the tangent space $T_0$. With $(q_0,p_0)$ as above, and putting $\hat p = (0,f_2)$ (the radial direction in $p$) we can choose the following basis for $V$: $$\left\{\hat p,\;(af_3,0),\;( 0, b_3f_3),\dots,\;(af_n,0),\;(0, b_nf_n)\right\}$$ (in ${\mathbb{R}}^n\oplus{\mathbb{R}}^n$), where $b_k = \sqrt{(J_1+J_2)(J_1+J_k)}$ and $a=\omega^{-1}$, where $\omega$ is the frequency of the steady rotation. With respect to this basis, one finds the Hessian matrix of $f$ at $(q_0,p_0)$ to be in block form, $$\left[\begin{array}{c|cccccc}
- 8\omega^2(J_1+J_2)^2 &0&0 &0& \dots &&0\cr
\hline
0&A_3&0&0&\dots&&0 \cr
0&0&A_4&0&\dots&&0 \cr
0 & 0&0&A_5&\dots &&0\cr
\vdots&\vdots&\vdots&&\ddots&&\vdots\cr
0&0&0&&\dots&&A_n
\end{array}\right],$$ where the $2\times2$ matrix $A_k$ is given simply by $$\label{eq:stability matrices}
A_k \ = \ \begin{pmatrix}J_1-J_k & 0 \\
0& J_2-J_k
\end{pmatrix}.$$ The remainder of the proof is simply combining the signs of these eigenvalues with the subspace of $V$ that suffices for a transversal to $S$.
Firstly, note that if all the eigenvalues $J_k$ are distinct, then the Hessian matrix is negative definite if and only if $J_k>\max\{J_1,J_2\}$ for all $k\geq3$, which is precisely the condition for a minimal principal plane. Indeed this holds more generally, provided $\max\{J_1,J_2\} <\min\{J_k\mid k\geq3\}$ (whether or not $J_1=J_2$ or some of the $J_k$ for $k\geq3$ are multiple).
Now assume $J_1\leq J_2 = J_3$ (which covers all remaining cases).
First consider the case where $J_1=J_2$ and suppose this has multiplicity $m\geq3$. Let $E_1$ be the $m$-dimensional eigenspace of ${\mathbb{J}}$ with eigenvalue $J_1$. The tangent space to the ${\mathrm{O}}(m)$-group orbit through $(q_0,p_0)=(f_1,\|p_0\|f_2)$ is $${\mathfrak{g}}\cdot(q_0,p_0) = \{(\xi f_1,\|p_0\|\xi f_2)\mid \xi\in{\mathfrak{so}}(m)\}.$$ Since $m\geq 3$ it follows that $${\mathfrak{g}}\cdot(q_0,p_0) = {\mathbb{R}}\{(f_2,-\|p_0\|f_1)\} \oplus (F_1\times\{0\}) \oplus (\{0\}\times F_1),$$ where $F_1$ is the span of $f_3,\dots,f_m$ in $E_1$. The first component here is the tangent space to the trajectory, so is not required in $V$, and then $$T_0={\mathbb{R}}(\dot q_0,\dot p_0) \oplus V = {\mathfrak{g}}\cdot(q_0,p_0) \oplus V_1$$ where $V_1$ has basis $$\left\{\hat p,\;(af_{r+1},0),\;( 0, b_{r+1}f_{r+1}),\dots,\;(af_n,0),\;(0, b_nf_n)\right\},$$ using the notation introduced above. It is clear that the restriction to this space of the Hessian matrix given above is again negative definite.
Finally suppose $J_1<J_2=J_3$, and let $m_2$ be the multiplicity of $J_2$, and $E_2$ the eigenspace. Since $J_1$ is simple there is only a finite group acting on that space, so that $${\mathfrak{g}}\cdot (q_0,p_0) = \{0\}\times E_2$$ where $E_2$ is the span of $\{f_3,\dots,f_{m+1}\}\subset E_2$. Again writing, $$T_0={\mathbb{R}}(\dot q_0,\dot p_0) \oplus V = {\mathbb{R}}(\dot q_0,\dot p_0) \oplus {\mathfrak{g}}\cdot(q_0,p_0) \oplus V_2$$ we have $V_2$ is spanned by $\hat p$ and $(f_r,0)$ ($r=3,\dots,m+1$) and by $(f_k,0),(0,f_k)$ for $k>m+1$. The eigenvalues of the Hessian above that are relevant are just the first (for the space spanned by $\hat p$) and the first element of each $A_k$ (for $k=3,\dots,m+1$, which are all negative), and the remaining $2\times 2$ blocks, which are all negative definite.
The axisymmetric Veselova top {#sec:axisymmetric}
=============================
This section considers [[***[axisymmetric tops]{}***]{}]{} for which the mass tensor relative to a frame $\{f_1, \dots f_n\}$ takes the form $$\label{eq:axisymmetric J}
{\mathbb{J}}= {\mathop\mathrm{diag}\nolimits}[J_1,J_2,\dots,J_2].$$ From its definition in , it follows that the symmetry group $G_R={\mathrm{O}}(n-1)$, which corresponds to rotations and reflections of the body frame that fix the symmetry axis.
In order to analyse the [[***[second reduced space]{}***]{}]{} ${\mathcal{R}}:=D/G = T^*{\mathrm{S}}^{n-1}/G_R$, we require the invariants for the action of $G_R={\mathrm{O}}(n-1)$ on $T^*{\mathrm{S}}^{n-1}$. Now, it is well-known that the ring of invariants for ${\mathrm{O}}(n-1)$ acting on $(q_2,p_2) \in {\mathbb{R}}^{n-1}\times{\mathbb{R}}^{n-1}$ is generated by three independent quadratic expressions: $$\|q_2\|^2,\quad \|p_2\|^2,\quad\text{and}\quad q_2\cdot p_2.$$ By the Cauchy-Schwarz inequality, these satisfy $|p_2\cdot q_2|^2 \leq\|q_2\|^2\|p_2\|^2$.
It follows that the ring of invariants for ${\mathrm{O}}(n-1)$ acting on $(q_1,q_2,p_1,p_2) \in {\mathbb{R}}\times{\mathbb{R}}^{n-1}\times{\mathbb{R}}\times{\mathbb{R}}^{n-1}$ is generated by the 5 invariants $$q_1,\quad p_1,\quad\|q_2\|^2,\quad \|p_2\|^2,\quad q_2\cdot p_2.$$
We now restrict to $T^*{\mathrm{S}}^{n-1}$ by requiring $\|q\|^2=q_1^2+\|q_2\|^2=1$ and $q\cdot p = q_1p_1+q_2\cdot p_2=0$. In place of 5 invariants, we now have 3: $$q_1,\quad p_1,\quad \text{and}\quad P:=\|p\|^2=p_1^2+\|p_2\|^2.$$
\[prop:axisymmetric R\] The orbit space $\mathcal{R}=T^*{\mathrm{S}}^{n-1}/{\mathrm{O}}(n-1)=D/G$ is a stratified semi-algebraic subspace of ${\mathbb{R}}^3$ given by $$\mathcal{R}=\{(q_1,p_1, P)\in {\mathbb{R}}^3 \, : \, p_1^2\leq (1-q_1^2)P, \;\; -1\leq q_1 \leq 1 \, \}.$$
The boundary of this space is often called the *canoe surface*. The reduced space $\mathcal{R}$ and the dynamics on it (to be discussed below) are illustrated in Figure \[fig:canoe\]. Basic notions of the stratification of orbits spaces are given in Appendix \[app:Field\]. Further details may be found for example in [@Duistermaat-Kolk].
(-1.5,0) – (1.5,0) node\[anchor=north east\][$q_1$]{}; (0,0) – (-0.6,-0.4) node\[anchor=north\][$p_1$]{}; (-1,1.5) – (-1,0) – (1,0) – (1,1.5); (0,1.5) ellipse \[x radius=1, y radius=0.24\]; (0,1.5) – (0,2) node\[anchor=south west\][$P$]{}; (0,0) – (0,1.5); (0,1.5) ellipse \[x radius=0.5, y radius=0.12\]; (0,1) ellipse \[x radius=1, y radius=0.2\]; (0,1) ellipse \[x radius=0.5, y radius=0.1\]; (0,0.5) ellipse \[x radius=1, y radius=0.15\]; (0,0.5) ellipse \[x radius=0.5, y radius=0.075\]; (-1,0) node\[anchor=north\][$-1$]{}; (1,0) node\[anchor=north\][$1$]{};
It was discussed above that the quantities $q_1,p_1$ and $P$ generate the invariants. The Cauchy-Schwarz inequality $(q_2\cdot p_2)^2\leq \|q_2\|^2\|p_2\|^2$ leads to the inequality $p_1^2 \leq P(1-q_1^2)$.
Th orbit space ${\mathcal{R}}$ has a natural stratification arising from the isotropy subgroups of the ${\mathrm{O}}(n-1)$ action on $T^*{\mathrm{S}}^{n-1}$. Accordingly, there are four strata of $\mathcal{R}$ as follows (recall that strata are by definition connected).
$S_0$, $S_0'$:
: If $q_2=p_2=0$, the point $(q,p)$ is fixed by ${\mathrm{O}}(n-1)$; this determines two zero-dimensional strata consisting of the points $(q_1,p_1,P)=(\pm 1,0, 0)$. This corresponds to the top being placed with its axis aligned with the distinguished axis, but stationary since $P=0$; from the dynamical point of view it is no more special than any other stationary configuration of the top.
$S_2$:
: This represents elements in $T^*{\mathrm{S}}^{n-1}$ for which $q_2$ and $p_2$ are parallel and do not both vanish. The isotropy group of such $(q,p)$ is conjugate to ${\mathrm{O}}(n-2)$. The image in ${\mathcal{R}}$ of this set is the canoe surface $p_1^2=(1-q_1^2)P$, with the points $(q_1,p_1,P)=(\pm 1,0, 0)$ removed.
$S_3$:
: This 3-dimensional stratum consists of the points $(q_1,p_1,P)$ corresponding to points of $T^*{\mathrm{S}}^{n-1}$ for which $q_2\wedge p_2\neq 0$ and their isotropy group is conjugate to ${\mathrm{O}}(n-3)$ (where, for $n=4$, ${\mathrm{O}}(1)={\mathbb{Z}}_2$, and for $n=3$, ${\mathrm{O}}(0)$ is the trivial group). This stratum corresponds to the interior of the orbit space, where $p_1^2<(1-q_1^2)P$.
Reduced dynamics {#sec:reduced dynamics axisymmetric}
----------------
The reduced equations on $\mathcal{R}$ can easily be obtained from . One finds, $$\label{eq:axisym-red-syst}
\dot q_1 = \frac{p_1}{J_1+J_2}, \qquad
\dot p_1 = -2H q_1, \qquad
\dot P =0 ,$$ where, after some calculations, $H$ is given by $$\label{eq:axisym Hred-nD}
H=\frac{1}{2(J_1+J_2)}\left ( \frac{(J_2-J_1)p_1^2+(J_1+J_2)P}{(J_1-J_2)q_1^2+2J_2}\right ),$$ and is a first integral. As expected, the flow of these equations leaves the strata of $\mathcal{R}$ invariant.
#### Dynamics:
Since $H$ is a first integral, the equations of motion above are linear, with solutions $$q_1(t) = A\cos(\omega t)+B\sin(\omega t),\qquad p_1(t)= (J_1+J_2)\omega (-A\sin(\omega t)+B\cos(\omega t)),$$ where $\omega^2 = 2h/(J_1+J_2)$ and $A,B$ are arbitrary constants of integration and $h$ is the constant value of $H$ along the solution. There are two classes of equilibria: along the $q_1$-axis (where $p_1=P=0$ and hence $H=0$) and along the $P$-axis (where $q_1=p_1=0$). In Figure\[fig:canoe\], the ellipses illustrate the trajectories; they lie in horizontal planes $P=\text{const}$ and are centred on the $P$-axis.
#### Steady rotations {#sec:steady-rot-axisymmetric}
Following Section\[sec:Steady-rotations\], we see that for an axisymmetric top there are two kinds of principal plane of inertia:
- Let $a\in{\mathbb{R}}^n$ be orthogonal to $f_1$. Then $\{f_1,a\}$ spans a principal plane containing the axis of symmetry, an [[***[axial]{}***]{}]{} principal plane, with associated moment of inertia ${\mathop\mathrm{in}\nolimits}(\Pi)=J_1+J_2$, and the corresponding steady rotation is simply a steady rotation in this plane.
- Now let $\Pi$ be any plane orthogonal to the axis of symmetry. This is also a principal plane of inertia, called an [[***[equatorial]{}***]{}]{} plane, and with ${\mathop\mathrm{in}\nolimits}(\Pi)=2J_2$. The resulting motion generalizes the 3D steady rotations where the body rotates about its symmetry axis.
\[prop:steady-rot-strata\]
1. The steady rotations in an axial principal plane are contained in $S_2$, and conversely all points in $S_2$ with $P>0$ correspond to such steady rotations.
2. Steady rotations in an equatorial plane of inertia are contained in $S_3$ and have $q_1=p_1=0$, and conversely, all points with $q_1=p_1=0$ correspond to such steady rotations.
3. Both the axial and equatorial steady rotations are Lyapunov stable relative to ${\mathrm{O}}(n-1)$.
First notice that if $\Omega=a\wedge b$ defines a steady rotation solution then, by Equation , $a\wedge b= q\wedge \dot q$ and it follows that $q(t)$ and $\dot q(t)$ are contained in the plane spanned by $a$ and $b$ at all time. We will make use of this in the following.
\(i) For a steady rotation in an axial plane we can write $\Omega=f_1\wedge a$ for a constant vector $a$ such that $f_1\cdot a=0$. Since $q$ lies on the plane spanned by $f_1$ and $a$, we have $q=q_1f_1\pm\frac{ \sqrt{1-q_1^2}}{\|a\|}a$. Hence, $$p=-{\mathbb{I}}(\Omega)q=-(J_1+J_2)\Omega q=(J_1+J_2)\left (\mp \sqrt{1-q_1^2}\,\|a\| f_1 + q_1a \right ).$$ The above formulae imply that $q_2$ and $p_2$ are parallel so the solution is indeed contained in $S_2$.
For the converse, recall that along $S_2$ the vectors $q_2$ and $p_2$ are parallel. Therefore, in view of we conclude that $\dot q_2$ is also parallel to $q_2$. Writing $\dot q_2=\lambda q_2$ we obtain $$\Omega = q\wedge \dot q = (q_1f_1 + (0,q_2))\wedge (\dot q_1f_1 + \lambda (0 , q_2))= f_1 \wedge a,$$ where $a=(\lambda q_1 -\dot q_1)\left (0, q_2 \right )$. Due to our symmetry assumptions on ${\mathbb{I}}$, this implies that ${\mathbb{I}}(\Omega)=(J_1+J_2)\Omega$ so $[{\mathbb{I}}\Omega, \Omega]=0$ which, by Proposition \[prop:steady rotations on D\], implies that the solution is a steady rotation. Therefore $a$ is constant and the steady rotation is of axial type. (Note that the condition $P>0$ implies $\Omega \neq 0$).
\(ii) If $\Omega=a\wedge b$ with $a, \, b, \, f_1,$ mutually perpendicular, then, since $q$ is contained in the plane spanned by $a$ and $b$, $q_1=q\cdot f_1=0$. But also $p_1=p\cdot f_1=0$ since $p=-{\mathbb{I}}(\Omega)q=-2J_2(a\wedge b) q$ is also contained in the plane spanned by $a$ and $b$.
For the converse, note that these points are equilibria on $\mathcal{R}$ satisfying $q_1=\dot q_1=0$. Considering that $\Omega=q\wedge \dot q$, and our symmetry assumptions on the inertia tensor, we have ${\mathbb{I}}(\Omega)=2J_2\Omega$. Therefore $[{\mathbb{I}}(\Omega),\Omega]=0$, and Proposition \[prop:steady rotations on D\] implies that the motion is a steady rotation which takes place on an equatorial plane since $f_1$ is orthogonal to $q$ and $\dot q$.
\(iii) This follows immediately from Theorem\[thm:stability\].
Reconstruction
--------------
We now give the global details of the reconstruction of the dynamics both to $T^*{\mathrm{S}}^{n-1}$ and to $D$. The first conclusion is that the dynamics is essentially that of the axisymmetric 3D Veselova top, in a sense that is made precise in the following theorem.
\[thm:nD to 3D\] Consider the $n$-dimensional axisymmetric Veselova top, with $n> 3$ and mass tensor ${\mathbb{J}}$ given in . Consider any solution to the reduced equation on ${\mathcal{R}}$, and a choice of initial lift to both $T^*{\mathrm{S}}^{n-1}$ and $D$. Then there exist choices of basis of ${\mathbb{R}}^n$ in the body and in space, the first containing $f_1$ and the second $e_1$, such that $$\label{eq:axisymmetric nD to 3D}
q(t) = \begin{pmatrix}
\bar q(t)\cr 0
\end{pmatrix},\quad p(t) = \begin{pmatrix}
\bar p(t)\cr 0
\end{pmatrix},\quad
g(t) = \begin{pmatrix}
\bar g(t)& 0\cr 0&\mathrm{Id}_{n-3}
\end{pmatrix},\quad
\Omega(t) = \begin{pmatrix}
\bar \Omega(t)&0\cr 0 &0_{n-3}
\end{pmatrix}$$ where $(\bar q(t),\bar p(t))\in T^*{\mathrm{S}}^2$ and $(\bar g(t), \bar\Omega(t))\in D(3)\subset T{\mathrm{O}}(3)$ satisfy the equations of motion of the axisymmetric 3D Veselova top (for $\bar g,\bar\Omega$) and its reduced equations (for $(\bar q,\bar p)$), with mass tensor $\bar{\mathbb{J}}= {\mathop\mathrm{diag}\nolimits}[J_1,J_2,J_2]$.
Suppose that the given solution in ${\mathcal{R}}$ is in the open stratum $S_3$ (the other cases follow similarly), and let $q_0,p_0,g_0,\Omega_0$ be corresponding initial values for the lifted dynamics. We describe the necessary changes of basis.
To begin, let $V_0$ be the subspace of ${\mathbb{R}}^n$ spanned by $q_0,p_0,f_1$ (which are necessarily linearly independent by virtue of being in the stratum $S_3$), and choose an orthonormal basis of $V_0$ consisting of $f_1$ and two further vectors. Finally complete this to an orthonormal basis $\{f_1,\dots,f_n\}$ of ${\mathbb{R}}^n$; being orthogonal to $f_1$, the remaining vectors will all be principal directions. The change of basis matrix $P_b\in G_R={\mathrm{O}}(n-1)$ maps $V_0$ to ${\mathbb{R}}^3\times \{0\}\subset{\mathbb{R}}^n$. Then $q_0, p_0$ has the form given in .
For the matrices, define now a basis in space as follows: let $V_0'=gV_0$. Clearly $e_1\in V_0'$ since $q_0\in V_0$. Extend $\{e_1\}$ to an orthonormal basis of $V_0'$ and then further extend to ${\mathbb{R}}^n$ by putting $e_j=gf_j$ for $j>3$. Let $P_s\in G_L={\mathrm{O}}(n-1)$ be the change of basis matrix. This maps $V_0'$ to ${\mathbb{R}}^3\times\{0\}$.
With this choice of basis, $q_0,p_0,g_0$ are of the form given in . It remains to show the same is true of $\Omega$ and then that the space is fixed by the dynamics. Now, from Equation , $\Omega=q\wedge \dot q$ and under the choice of basis transforms to $P_b\Omega P_b^T = (P_bq)\wedge(P_b \dot q)$. Since from the equations of motion , $\dot q\in V_0$ it follows that $P_b\Omega P_b^T$ is of the required form since both $P_bq$ and $P_b \dot q$ belong to ${\mathbb{R}}^3\times\{0\}$.
We need now to show that the space of such matrices and vectors is invariant under the dynamics. There are two ways to do this, either a symmetry argument or an inspection of the equations of motion. For the symmetry argument, first consider the $G_R={\mathrm{O}}(n-1)$ action on $T^*{\mathrm{S}}^{n-1}$ (Corollary \[coroll:symmetry on T\^\*S\]) and note that $V_0$ is the fixed point set of ${\mathrm{O}}(n-3)\subset{\mathrm{O}}(n-1)$. It follows that $(q(t),p(t))$ is fixed by ${\mathrm{O}}(n-3)$ for all $t$, and hence that it evolves on $(V_0\times V_0)\cap T^*{\mathrm{S}}^{n-1}=T^*{\mathrm{S}}^2$. Secondly, consider the action of $G=G_L\times G_R = {\mathrm{O}}(n-1)\times {\mathrm{O}}(n-1)$ on the pair of matrices $(g,\Omega)\in D$ (Equation ). Similar to the previous argument, the set matrix pairs $(g,\Omega)$ given in is the fixed point set of the diagonal subgroup $${\mathrm{O}}(n-3)_\Delta = \{ (k,k)\in G \, | \, k\in {\mathrm{O}}(n-3) \},$$ and hence that too is invariant under the dynamics.
From the equations of motion and it is easy to see that the resulting equations of motion on these fixed point sets are precisely those arising from the 3D axisymmetric Veselova top with the given mass tensor.
It follows from this theorem that in order to study the dynamics of the axisymmetric $n$-dimensional Veselova top, it suffices to reconstruct the dynamics for the 3D one. And because of the reconstruction formula given in Proposition\[prop:reconstruction 3D\], each motion on $T^*{\mathrm{S}}^2$ lifts to a similar motion on $D$. Recall that equatorial and axial planes are defined just prior to Proposition \[prop:steady-rot-strata\].
\[thm:reconstruction-axisymmetric\] Consider a solution on ${\mathcal{R}}$ of the reduced $3$-dimensional axisymmetric Veselova top, with mass tensor ${\mathbb{J}}= {\mathop\mathrm{diag}\nolimits}[J_1,J_2,J_2]$. The possible reconstructions to both $T^*{\mathrm{S}}^2$ and $D$ are given in the following table,
stratum reconstruction
------------------------------ --------------------------------------
$S_3$, with $(q_1,p_1)\neq0$ quasiperiodic motion on 2-tori
$(0,0,P)\in S_3$ steady rotations in equatorial plane
$S_2$ with $P>0$ steady rotations in axial plane
all remaining points equilibrium
The proofs of the statements in the second and third rows of the table were given in Proposition \[prop:steady-rot-strata\].
For the first row, the solutions through these points in the second reduced space ${\mathcal{R}}$ are periodic. Moreover, the action of $G_R={\mathrm{O}}(2)$ on $T^*{\mathrm{S}}^2$ restricts to a free action on $S_3$. Since ${\mathrm{O}}(2)$ has rank 1, Field’s theorem (Theorem\[thm:Field\]) implies that the motion is quasi-periodic of tori of dimension two on $T^*{\mathrm{S}}^{2}$.
In the remaining case $P=0$ in which case $\Omega=0$ and the kinetic energy vanishes. Since there is no potential, there is no motion.
\[rmk:Subtori\] We note the application of Field’s theorem to the reconstruction of orbits on $S_3$ to $D$ proves that the motion is quasi-periodic on invariant 3-tori since the group $G=G_L\times G_R={\mathrm{O}}(2)\times {\mathrm{O}}(2)$ has rank 2. Theorem \[thm:reconstruction-axisymmetric\] instead shows that the dynamics takes place in a finer foliation by invariant 2-dimensional sub-tori. A physical interpretation of this phenomenon is given in Section \[SS:physical description\] below. This scenario, where the invariant tori on $T^*{\mathrm{S}}^{n-1}$ and on $D$ have the same dimension, also holds for general non-physical inertia tensors satisfying (see [@FedJov Section 7]).
#### The energy-momentum map
The global description of the dynamics of the $n$-dimensional axi-symmetric Veselova top may now be conveniently illustrated by looking at the image of the energy-momentum map $(P,H)$. We assume that $J_1<J_2$ (the opposite inequality leads to a similar description). Using that $p_1^2\leq P(1-q_1^2)$, it is straightforward to show that $P/4J_2\leq H(q,p) \leq P/2(J_1+J_2)$. Therefore, the image of the energy-momentum map $(P,H)$ is the wedge shown in Figure \[fig:E-M for axisymmetric\], see also Proposition\[prop:EM critical rays\].
(0,0) – (1.4,0) node\[anchor=north\][$P$]{}; (0,0) – (0,1.4) node\[anchor=east\][$H$]{}; (0,0) – (0.4,1.2) – (1.2,0.4); (0,0) – (0.4,1.2) node\[anchor=south west\] [-5mm$P=2(J_1+J_2)H$]{}; (0,0) – (1.2,0.4) node\[anchor=west\] [$P=4J_2H$]{};
Moreover, for $(P,H)\neq(0,0)$, the equalities $H=P/2(J_1+J_2)$ and $H= P/4J_2$ are respectively attained on $S_2$ and along the points of $S_3$ where $q_1=p_1=0$ and $P>0$. Therefore, combining the conclusion of Theorem \[thm:nD to 3D\] with the different items of Theorem \[thm:reconstruction-axisymmetric\], we obtain:
1. The pre-image of the points in the interior of the wedge (the regular values of the map) is foliated by invariant tori of dimension $2$ on both $T^*{\mathrm{S}}^{n-1}$ and $D$ that carry quasiperiodic flow.
2. The pre-image of the lower boundary line $H=P/4J_2$, $P>0$, consists of steady rotations in equatorial principal planes of inertia.
3. The pre-image of the upper boundary line $H=P/2(J_1+J_2)$, $P>0$, consists of steady rotations in axial principal planes of inertia.
4. The pre-image of the vertex of the wedge consists of all the equilibria both on $D$ and on $T^*{\mathrm{S}}^{n-1}$.
Physical description of the dynamics {#SS:physical description}
------------------------------------
The dynamics of the $n$-dimensional axisymmetric Veselova top is now easily described. In view of Theorem \[thm:nD to 3D\], we restrict our attention to the 3D case. Consider a solution of the system with $H, P\neq 0$. Without loss of generality, we may choose the second axis $e_2$ of the space frame to be aligned with the space representation of the momentum vector $p(t)$. The attitude matrix $g(t)$ is then given by Equation with $g_0=\mbox{Id}_3$. It follows that the axis of symmetry of the body, as seen in space, traces the following path along the motion: $$g(t)f_1=\left (q_1(t),\frac{p_1(t)}{\sqrt{P}}, \pm \sqrt{L(t)} \right )^T,$$ where $L(t)=1-q_1(t)^2-\frac{p_1(t)^2}{\sqrt{P}}$. The above is a closed curve in space since all entries of the above vector are periodic with the same frequency.
Let us focus on the case where the solution is quasi-periodic on a 2-torus (see Theorem \[thm:reconstruction-axisymmetric\]). Then $0<L(t)<1$ and, as follows from the conservation of energy , the axis of the body rotates about the $e_3$ axis of the space frame along the curve given as the intersection of the cylinder $$\left \{ (x_1,x_2,x_3) \, \left | \, x_1^2+\frac{P}{2H(J_1+J_2)}x_2^2= \frac{4HJ_2-P}{2H(J_2-J_1)} \right . \right \}$$ with the unit sphere. Under our assumptions, the connected components of this curve are not contained on a plane with constant $x_3$. Therefore, apart from the periodic precession of the body about the $e_3$-axis, the body undergoes a periodic nutation with half its period (see Figure \[F:motion-axi\]). Apart from this periodic translational motion, the body rotates about its symmetry axis at an angular speed that is compatible with the nonholonomic constraint.
![Periodic motion of the axis of symmetry of the top. The distinguished axis that defines the nonholonomic constraint is $e_1$. The axis $e_2$ is aligned with the space representation of the momentum vector $p(t)$. []{data-label="F:motion-axi"}](motion-axi.jpg "fig:"){width="3cm"} (-85,20) [$e_1$]{} (-32,5) [$e_2$]{} (-43,80) [$e_3$]{}
Note that even though the system has the same symmetry group as the classical Lagrange top, whose motions are generically quasi-periodic with three frequencies (see e.g. [@MMCM]), the motion of the axisymmetric Veselova top only has two independent frequencies since the nutation and precession motions of the body are commensurate. This gives a physical interpretation of Remark \[rmk:Subtori\].
The cylindrical Veselova top {#sec:cylindrical}
============================
We now consider the more general case of [[***[cylindrical tops]{}***]{}]{}: that is, we assume $$\label{eq:cylindrical J}
{\mathbb{J}}= {\mathop\mathrm{diag}\nolimits}[J_1,\dots,J_1,J_2,\dots,J_2],$$ with $J_1\neq J_2$. Let $r$ be the multiplicity of eigenvalue $J_1$, and $r'=n-r$ the multiplicity of $J_2$, and we may assume $2\leq r\leq r'$ (the case $r=1$ was treated in the previous section). In particular the dimension $n$ of the body must be at least 4.
The symmetry group $G_R$ defined by Equation is $G_R:={\mathrm{O}}(r)\times{\mathrm{O}}(r')\subset{\mathrm{O}}(n)$. Indeed, if $$h=\left[\begin{matrix}
h_{11}&0\cr 0&h_{22}
\end{matrix}\right],$$ with $h_{11}\in{\mathrm{O}}(r)$ and $h_{22}\in{\mathrm{O}}(r')$, then $h{\mathbb{J}}h^T={\mathbb{J}}$. Corresponding to the splitting ${\mathbb{R}}^n={\mathbb{R}}^r\times{\mathbb{R}}^{r'}$, write $q=(q_1,q_2)$ and $p=(p_1,p_2)$. Then, the action of $h\in G_R$ on $(q,p)\in T^*{\mathrm{S}}^{n-1}$ (see Corollary \[coroll:symmetry on T\^\*S\]) is $$hq=(h_{11}q_1,\,h_{22}q_2)$$ and similarly for $p$.
Second reduction
----------------
As for the axisymmetric tops, denote the [[***[second reduced space]{}***]{}]{} $T^*{\mathrm{S}}^{n-1} /G_R$ by ${\mathcal{R}}$. To determine ${\mathcal{R}}$ for a cylindrical body, we find the invariant functions for the $G_R$ action on $T^*{\mathrm{S}}^{n-1}$. As mentioned before, the ring of invariants for ${\mathrm{O}}(k)$ acting on $(q,p)\in{\mathbb{R}}^k\times{\mathbb{R}}^k$ is generated by three independent quadratic expressions: $$\|q\|^2,\quad \|p\|^2,\quad\text{and}\quad q\cdot p.$$ By the Cauchy-Schwarz inequality, these satisfy $|q\cdot p|^2 \leq\|q\|^2\|p\|^2$.
It follows that the ring of invariants for $G_R={\mathrm{O}}(r)\times{\mathrm{O}}(r')$ acting on $(q_1,q_2,p_1,p_2) \in {\mathbb{R}}^r\times{\mathbb{R}}^{r'}\times{\mathbb{R}}^r\times{\mathbb{R}}^{r'}$ is generated by the 6 invariants $$\|q_1\|^2,\quad \|p_1\|^2,\quad q_1\cdot p_1,\quad\|q_2\|^2,\quad \|p_2\|^2,\quad q_2\cdot p_2.$$
We now restrict to $T^*{\mathrm{S}}^{n-1}$ by requiring $\|q\|^2=\|q_1\|^2+\|q_2\|^2=1$ and $q\cdot p = q_1\cdot p_1+q_2\cdot p_2=0$. In place of 6 invariants, we now have 4:
$$\label{eq:cylindrical invariants}
A=\|q_1\|^2,\quad B=\|p_1\|^2,\quad P=\|p_1\|^2+\|p_2\|^2,\quad\text{and}\quad D=p_1\cdot q_1.$$
The remaining invariants are given simply by $\|q_2\|^2=1-A$, $\|p_2\|^2=P-B$, and $q_2\cdot p_2=-D$. The reason for using $P$ rather than $\|p_2\|^2$ is that $P$ is conserved by the dynamics. (Note that $D$ refers both to this variable as well as the distribution in $T{\mathrm{O}}(n)$—we hope no confusion will be caused.)
\[prop:reduced space-cylindrical\] The second reduced space ${\mathcal{R}}$ is the semialgebraic variety given by the set consisting of those $(A,B,P,D)\in{\mathbb{R}}^4$ satisfying $$0\leq A\leq 1, \quad 0\leq B\leq P,\quad D^2\leq AB,\quad D^2 \leq(1-A)(P-B).$$ This reduced space ${\mathcal{R}}$ has a natural stratification arising from the isotropy groups for the action of $G_R$ on $T^*{\mathrm{S}}^{n-1}$. These subgroups and the associated strata, both in $T^*{\mathrm{S}}^{n-1}$ and their images in ${\mathcal{R}}$, are given in Table \[table:strata T\^\*S + R\].
A section of ${\mathcal{R}}$ with constant $P>0$ and the adjacencies between the strata are shown in Figure\[fig:R-cylindrical\].
$$\begin{array}{c||c|c||c}
& \multicolumn{2}{c||}{T^*{\mathrm{S}}^{n-1}} & {\mathcal{R}}\\
\hline
\text{stratum}& \text{the below are zero}& \text{ISG}& \text{equations of stratum} \\
\hline
\rule{0pt}{12pt}
S_0 & q_1, p_1, p_2 & {\mathrm{O}}(r)\times{\mathrm{O}}(r'-1) & A=B=P=D=0 \\[4pt]
S_0' & q_2, p_1, p_2 & {\mathrm{O}}(r-1)\times{\mathrm{O}}(r') & A=1, B=P=D=0\\[4pt]
S_1 & q_1,p_1 & {\mathrm{O}}(r)\times{\mathrm{O}}(r'-2) & A=B=D=0 \\[4pt]
S_1' & q_2,p_2 & {\mathrm{O}}(r-2)\times{\mathrm{O}}(r') & A=1, P-B=D=0 \\[4pt]
S_2 & \left\{\!\!\hbox{$\begin{array}{l}q_1\wedge p_1 \cr q_2\wedge p_2\end{array}$}\right. & {\mathrm{O}}(r-1)\times{\mathrm{O}}(r'-1) & \left\{\hbox{$\begin{aligned} D^2&=AB\cr &=(1-A)(P-B)\end{aligned}$}\right.\\[16pt]
S_3 & q_1\wedge p_1 & {\mathrm{O}}(r-1)\times{\mathrm{O}}(r'-2)& D^2=AB \\[4pt]
S_3' &q_2\wedge p_2 & {\mathrm{O}}(r-2)\times{\mathrm{O}}(r'-1) & D^2=(1-A)(P-B) \\[4pt]
S_4 &\text{--}& {\mathrm{O}}(r-2)\times{\mathrm{O}}(r'-2) & \text{the rest} \cr
\hline
\end{array}$$
The inequalities arise from the definition of $A$ and the Cauchy-Schwarz inequality.
For the stratification and isotropy subgroups, consider first the natural action of ${\mathrm{O}}(k)$ on $x=(a,b)\in{\mathbb{R}}^k\times{\mathbb{R}}^k$, with $k\geq2$. Clearly, the stabilizer of the point $x=(0,0)$ is ${\mathrm{O}}(k)$, and that is the only point fixed by the whole group. Next, if $a$ and $b$ are parallel ($a\wedge b=0$) and not both zero, then the stabilizer is (conjugate to) ${\mathrm{O}}(k-1)$ (if $k=2$ then ${\mathrm{O}}(1)={\mathbb{Z}}_2$ is generated by the reflection in the line containing $a,b$). Finally, if $a,b$ are linearly independent then the stabilizer is ${\mathrm{O}}(k-2)$, which is the group of all orthogonal transformations in the subspace orthogonal to the space spanned by $a,b$. (For the case of $k=2$, the group ${\mathrm{O}}(0)$ is the trivial group.)
These observations can now be applied to the action of $G_R={\mathrm{O}}(r)\times{\mathrm{O}}(r')$ on ${\mathbb{R}}^r\times{\mathbb{R}}^{r'}\times{\mathbb{R}}^r\times{\mathbb{R}}^{r'}$ and its restriction to $T^*{\mathrm{S}}^{n-1}$ to obtain the strata in Table \[table:strata T\^\*S + R\].
(0,0) – (2.92,1.48) – (2.92, -1.48) – (0,0); (6,0) – (3.08,1.48) – (3.08, -1.48) – (6,0); (3,-1.5) arc (-90:90:0.5 and 1.5); (3,1.5) arc (90:270:0.5 and 1.5); (0,0) node\[anchor=east\] [$S_1$]{}; (6,0) node\[anchor=west\] [$S_1'$]{}; (3,-1.5) node\[anchor=north\] [$S_2$]{}; (3,1.5) node\[anchor=south\] [$S_2$]{}; (1.5,-0.75) node\[anchor=north\] [$S_3$]{}; (4.5,0.75) node\[anchor=south\] [$S_3'$]{}; (2.3,0.5) node [$S_4$]{}; (0,0) – (3.364, -1.028) – (6,0); (0,0) – (3.479, -.4272) – (6,0); (0,0) – (3.479, .4272) – (6,0); (0,0) – (3.364, 1.028) – (6,0); (0,0) .. controls (3,-0.3) .. (6,0);
-4cm $$\begin{tikzcd}[column sep=small]
&&S_4 \arrow[dr] \arrow[dl] &&\\
&S_3 \arrow[dr]\arrow[dl] && S_3' \arrow[dr]\arrow[dl] \\
S_1\arrow[dr] && S_2 \arrow[dl] \arrow[dr] && S_1'\arrow[dl] \\
& S_0 &&S_0'
\end{tikzcd}$$
#### Description of strata
Here we give a brief description of the strata, in particular in the constraint distribution $D$; the statements about dynamics are proved in Proposition \[prop:steady-rot-strata-Cylindrical\] and Theorem\[thm:reconstruction-cylindrical\].
$S_0,S_0'$:
: These points in ${\mathcal{R}}$ correspond to points in $T^*{\mathrm{S}}^{n-1}$ where $p=0$, and either $q_1=0$ or $q_2=0$ respectively. This means the top is oriented in space in such a way that the distinguished axis is contained in one of the eigenspaces of ${\mathbb{J}}$, and the top is stationary. These strata do not appear in the cone in Figure\[fig:R-cylindrical\] since the cone represents a section of ${\mathcal{R}}$ with $P>0$.
$S_1,S_1'$:
: Here the top is such that again the distinguished axis lies in one of the eigenspaces of ${\mathbb{J}}$; we will see below that the corresponding motion is necessarily a steady rotation (of [*pure type*]{}; see below). These strata correspond to the vertices of the cones in Figure\[fig:R-cylindrical\].
$S_2$:
: The configuration is arbitrary, but we show below the motion is a steady rotation, here of [*mixed type*]{} (see below). In Figure\[fig:R-cylindrical\] the stratum $S_2$ is the circle where the two cones meet.
$S_3,S_3'$:
: In Figure\[fig:R-cylindrical\] these strata are the boundaries of the two solid cones. We show below that in fact the motion on these strata is ‘isomorphic’ to the motion of an axisymmetric top.
$S_4$:
: In the figure, this corresponds to the interior of the solid cone. This stratum contains the ‘genuine’ general motion of the cylindrical top, not reducible to other cases. For generic points, the solutions are quasiperiodic on tori of dimension 4; there are also some relative equilibria where the tori are of dimension 3, represented by the dashed curve in the figure.
\[prop:reducedEqMotion\] The reduced equations of motion for the cylindrical Veselova top on ${\mathcal{R}}$ are, $$\label{eq:reduced EoM}
\left\{\quad\begin{aligned}
\dot A \ &= \ \frac{2D}{J_1+J_2}, \\
\dot B \ &= \ -4HD, \\
\dot P \ &= \ 0, \\
\dot D \ &= \ -4HA + \frac{P-4J_2 H}{J_1-J_2}.
\end{aligned}\right.$$ Here $$\label{eq:Hamiltonian in ABPD}
H = \frac12\left((\beta_1-\beta_2)B+\beta_2P - \frac{(\beta_1-\beta_2)^2D^2}{\beta_1A+\beta_2(1-A)}\right),$$ which is an integral of motion, and $$\label{eq:betas}
\beta_1=((J_1-J_2)A+J_1+J_2)^{-1}, \quad\quad \beta_2=( (J_1-J_2)A+2J_2)^{-1}.$$
We derive these from the equations of motion given in Proposition\[prop:eqns-nD\]. Under our assumption on ${\mathbb{J}}$ we have $$q\cdot{\mathbb{J}}q = J_1\|q_1\|^2 +J_2\|q_2\|^2 = J_1A+J_2(1-A).$$ It follows that the matrix $C$ of satisfies $$C^{-1} = \begin{pmatrix}
\left(J_1(1+A)+J_2(1-A)\right)\mathrm{Id}_r &0\cr 0 & \left(J_1A+J_2(2-A)\right)\mathrm{Id}_{r'}
\end{pmatrix}$$ Then $C=\begin{pmatrix}
\beta_1\mathrm{Id}_r &0\cr 0&\beta_2\,\mathrm{Id}_{r'}
\end{pmatrix}$ where $\beta_1,\beta_2$ are given in . Using this notation, we find $$\begin{aligned}
q\cdot Cq &=& \beta_1A+\beta_2(1-A) \\
p\cdot Cp &=& \beta_1B+\beta_2C \\
q\cdot Cp &=& (\beta_1-\beta_2)D,\end{aligned}$$ and the Hamiltonian is indeed given by .
Now, $$\dot A = 2q_1\cdot\dot q_1 = 2q_1\cdot\left(\beta_1p_1 - \frac{(\beta_1-\beta_2)D}{\beta_1A+\beta_2(1-A)}\beta_1q_1\right) = 2\beta_1D\left(1-\frac{(\beta_1-\beta_2)A}{\beta_1A+\beta_2(1-A)}\right).$$ and after some simple calculations one finds the expression for $\dot A$ given above.
More simple is, $$\dot B = 2p_1\cdot\dot p_1=-4HD,\quad \dot P = 0,$$ and $$\dot D = -2HA+\beta_1B - \frac{\beta_1(\beta_1-\beta_2)D^2}{\beta_1A+\beta_2(1-A)},$$ which can be shown to be equal to the required expression.
\[rmk:reduced measure\] The invariant measure for the general (first reduced) $n$-dimensional Veslova top given in pushes forward to an invariant measure on ${\mathcal{R}}$ given simply by $$\beta_1\beta_2\,{\mathsf{d}}A\,{\mathsf{d}}B\,{\mathsf{d}}P\,{\mathsf{d}}D.$$
#### Reduced dynamics {#reduced-dynamics}
We can now describe the reduced dynamics on ${\mathcal{R}}$. Since $P,H$ are constant, the equations of motion are linear, and easy to integrate. First consider the equations for $(A,D)$: the general solution is $$\label{eq:AD solutions}
\begin{split}
A(t) &= C_1\cos(\omega t) + C_2\sin(\omega t) + A_* \\
D(t) &= \frac12(J_1+J_2)\omega \left(-C_1\sin(\omega t) + C_2\cos(\omega t)\right),
\end{split}$$ where $C_1,C_2$ and $A_*$ are arbitrary constants (if the constants are chosen so that the initial value lies within ${\mathcal{R}}$, then the entire solution must lie within ${\mathcal{R}}$). Here $$\label{eq:omega}
\omega^2 = \frac{8h}{J_1+J_2},$$ where $h$ is the constant value of $H$ on the solution. Substituting into the differential equation for $B$ shows, $$\label{eq:B solution}
B(t) = -2\,h (J_1+J_2) \left(
C_1\,\cos (\omega t) + C_2\,\sin(\omega t) \right) + B_*.$$ The quantities $A_*$ and $B_*$ (depending on the initial conditions) are the respective mean values of $A$ and $B$ over one period. The mean value for $D$ is $D_*=0$.
#### Equilibria
From $\dot A=0$ we require $D=0$ which then implies $\dot B=0$. There remains to consider the equation $\dot D=0$.
- As already noted above, along the $A$-axis, the points are equilibria (these correspond to $p=0$ so are the equilibria of the full system, and are the points where $H=0$). The end points of this axis are the points on $S_0$ and $S_0'$, while the other points belong to $S_2$.
- The strata $S_0,S_0',S_1,S_1'$ consist entirely of equilibrium points (for example, on $S_1$, if $A=B=D=0$ then $P=4J_2 H$ showing $\dot D=0$).
- More generally, equilibria occur with $H=h>0$, if $A=A_0$ and $B=B_0$, where $$\label{eq:equilibria}
A_0 = \frac{P-4J_2h}{4(J_1-J_2)h},\quad\text{and}\quad
B_0 = \frac{(P-4J_2h)(P+4J_1h)}{8(J_1-J_2)h}.$$ If $P=4J_2h$ then $A_0=B_0=0$ and the point lies in $S_1$; if $P=4J_1h$ then $A=1,B=P$ and the point lies in $S_1'$. All other equilibrium points lie in $S_4$, and are illustrated by the dashed curve in ${\mathcal{R}}$ in Figure\[fig:R-cylindrical\].
\[thm:dynamics on R\] The motion on the reduced space ${\mathcal{R}}$ has the following properties.
1. It is integrable, with the 3 first integrals $P,H$ and $$F:= 2(J_1+J_2)AH + B.$$
2. Each point with $H=0$ is an equilibrium.
3. For each $(h,P)$ with $h>0$ in the image of the energy-momentum map (see Figure\[fig:E-M for cylindrical\]) there is a unique equilibrium with $H=h$; this occurs at the point $(A,B,P,D) = (A_0,B_0,P,0)$, where $A_0$ and $B_0$ are the expressions given in .
4. All other solutions are periodic with period $\pi\sqrt{\frac{J_1+J_2}{2h}}$.
In particular the period of the periodic orbits given in (iv) depends only on the value of the energy; this is reminiscent of the period-energy relation in Hamiltonian systems, even though this system is not known to be Hamiltonian (not even Poisson-Hamilton systems in general satisfy this property).
\(i) For $P$ this is clear, for $H$ we know this already, while for $F$ it is a simple calculation: $$\begin{aligned}
\dot F &=& 2(J_1+J_2)\dot A H + \dot B \\
&=& 2H\left((J_1+J_2)\dot A -2D\right)\ =\ 0.\end{aligned}$$
\(ii) The set of points with $H=0$ is the $A$-axis in ${\mathcal{R}}$, and we have already seen above that these are equilibria.
\(iii) Solving the equation $\dot D=0$ with $H=h$ for $A$ shows immediately that $A=A_0$. Putting $A=A_0$ the equation $H=h$ can easily be solved for $B$ to find $B=B_0$.
\(iv) If a point is not an equilibrium then it lies on a periodic orbit, as shown in and , with period $2\pi/\omega$ with $\omega$ given in .
#### Conclusion:
In ${\mathcal{R}}$ each invariant set given by fixing values of $H$ and $P$ is a surface bounded by the intersection with a circle in the boundary of the cone in Figure\[fig:R-cylindrical\], with a unique equilibrium point surrounded by periodic orbits all of the same period.
\[rmk:Poisson structure\] Since the dynamics on ${\mathcal{R}}$ is entirely periodic, one may find Poisson structures of rank 2 for which the system is Poisson-Hamiltonian, with Hamiltonian $H$, and with Casimir functions $P$ and $F$ [@FassoRank2]. It would be interesting to know if any of these Poisson structures arises as the reduction of an almost Poisson structure on $D$ as described in [@LGN-JM17]. This possibility to express the dynamics on ${\mathcal{R}}$ in Hamiltonian form contrasts with the dynamics on $T^*{\mathrm{S}}^{n-1}$, c.f. Remark\[rmk:hamiltonization\].
#### Steady rotations {#steady-rotations-1}
It follows from Proposition\[prop:EM critical rays\] that there are three critical rays in the image of the momentum map; namely where $P=4J_1H$, $P=2(J_1+J_2)H$ and $P=4J_2H$. The image of the energy-momentum map is therefore as shown in Figure\[fig:E-M for cylindrical\].
(0,0) – (1.4,0) node\[anchor=north\][$P$]{}; (0,0) – (0,1.4) node\[anchor=east\][$H$]{}; (0,0) – (0.4,1.2) – (1.2,0.4); (0,0) – (0.4,1.2) node\[anchor=south west\] [$P=4J_1H$]{}; (0,0) – (0.8,0.8) node\[anchor=south west\] [$P=2(J_1+J_2)H$]{}; (0,0) – (1.2,0.4) node\[anchor=west\] [$P=4J_2H$]{};
Each critical ray in the image of the momentum map corresponds to a type of steady rotation of the cylindrical Veselova top:
- Any plane $\Pi$ contained in the $r$-dimensional $J_1$-eigenspace of ${\mathbb{J}}$ is a principal plane of inertia. Such a plane $\Pi$ satisfies ${\mathop\mathrm{in}\nolimits}(\Pi)=2J_1$.
- Any plane $\Pi$ contained in the $r'$-dimensional $J_2$-eigenspace of ${\mathbb{J}}$ is a principal plane of inertia. Such a plane $\Pi$ satisfies ${\mathop\mathrm{in}\nolimits}(\Pi)=2J_2$.
- Let ${\mathbb{J}}a =J_1a$ and ${\mathbb{J}}b=J_2b$, with $a,b$ non-zero. The plane $\Pi_{a,b}={\mathbb{R}}\{a,b\}$ is a principal plane in the body with moment of inertia ${\mathop\mathrm{in}\nolimits}(\Pi)=J_1+J_2$.
The inertial planes of the first two types we call [[***[pure inertial planes]{}***]{}]{}, while those of the third type we call [[***[mixed inertial planes]{}***]{}]{}. The steady rotations in pure inertial planes are relative equilibria, while the ones in mixed inertial planes are not (see Theorem\[thm:reconstruction-cylindrical\]).
Analogous to Proposition \[prop:steady-rot-strata\], we have:
\[prop:steady-rot-strata-Cylindrical\]
1. Steady rotations of mixed type are contained in $S_2$, and conversely, all points in $S_2$ with $P>0$ correspond to such steady rotations.
2. The steady rotations in a pure inertia plane associated to $J_1$ are contained in $S_1'$, and conversely all points in $S_1'$ correspond to such steady rotations. The same statement holds replacing $J_1$ by $J_2$ and $S_1'$ by $S_1$.
3. The pure steady rotations are Lyapunov stable relative to the action of $G_R$.
The proof is a straightforward adaptation of the proof of Proposition \[prop:steady-rot-strata\]. For example, let us show that all points in $S_2$ with $P>0$ correspond to steady rotations of mixed type. Recall from Table \[table:strata T\^\*S + R\] that along this stratum $q_j$ is parallel to $p_j$, $j=1,2$. In view of the cylindrical symmetry of ${\mathbb{J}}$, Equations then imply that $\dot q_j$ is parallel to $q_j$. Writing $\dot q_j=\lambda_j q_j$ we obtain $$\Omega = q\wedge \dot q = ((q_1,0)+ (0,q_2))\wedge (\lambda_1(q_1,0) + \lambda_2 (0 , q_2))=(\lambda_2-\lambda_1) (q_1,0)\wedge (0,q_2).$$ The cylindrical symmetry of ${\mathbb{J}}$ now implies that ${\mathbb{I}}(\Omega)=(J_1+J_2)\Omega$ so $[{\mathbb{I}}\Omega, \Omega]=0$ which, by Proposition \[prop:steady rotations on D\], implies that the solution is a steady rotation, that is clearly of mixed type. (Note that the condition $P>0$ implies $\Omega \neq 0$.)
The details of the proof of the other statements are left to the reader. The final statement follows from Theorem\[thm:stability\].
Reconstruction to $T^*{\mathrm{S}}^{n-1}$ and $D$ {#sec:lifting - cylindrical}
-------------------------------------------------
Recall that a Veselova top with cylindrical symmetry is necessarily of dimension at least 4. The analogous reduction to 3D for the axisymmetric top in this case is a reduction to 4D.
\[thm:nD to 4D\] Consider the $n$-dimensional cylindrical Veselova top, with $n>4$ and mass tensor ${\mathbb{J}}$ given in . Consider any solution to the reduced equation on ${\mathcal{R}}$, and a choice of initial lift to both $T^*{\mathrm{S}}^{n-1}$ and $D$. Then there exists choices of basis of ${\mathbb{R}}^n$ in the body and in space, such that $$\label{eq:cylindrical nD to 4D}
q(t) = \begin{pmatrix}
\bar q(t)\cr 0
\end{pmatrix},\quad p(t) = \begin{pmatrix}
\bar p(t)\cr 0
\end{pmatrix},\quad
g(t) = \begin{pmatrix}
\bar g(t)& 0\cr 0&\mathrm{Id}_{n-4}
\end{pmatrix},\quad
\Omega(t) = \begin{pmatrix}
\bar \Omega(t)&0\cr 0 &0_{n-4}
\end{pmatrix},$$ where $(\bar q(t),\bar p(t))\in T^*{\mathrm{S}}^3$ and $(\bar g(t), \bar\Omega(t))\in D(4)\subset T{\mathrm{O}}(4)$ satisfy the equations of motion of the axisymmetric 4D Veselova top (for $(\bar g,\bar\Omega)$) and its reduced equations (for $(\bar q,\bar p)$), with mass tensor $\bar{\mathbb{J}}= {\mathop\mathrm{diag}\nolimits}[J_1,J_1,J_2,J_2]$.
This is similar to the proof of Theorem\[thm:nD to 3D\], and we highlight the differences. Here we consider a point in the stratum $S_4$. Let $V=V_1\oplus V_2$ be the 4-dimensional subspace of ${\mathbb{R}}^n$, where $$V_1 \ = \ {\mathbb{R}}\,\left\{\begin{pmatrix}q_1\cr0\end{pmatrix},\;\begin{pmatrix}p_1\cr0\end{pmatrix}\;\,\right\},\quad\text{and}\quad V_2 \ = \ {\mathbb{R}}\,\left\{\begin{pmatrix}0\\ q_2\end{pmatrix},\; \begin{pmatrix}0\cr p_2\end{pmatrix} \,\right\}.$$ The assumption that the initial point is in $S_4$ implies these 4 vectors are linearly independent. We now construct an orthonormal body basis $\{f_1,\dots, f_n\}$ as follows. First choose $\{f_1,f_2\}$ and $\{f_3,f_4\}$ as orthonormal bases of $V_1$ and $V_2$ respectively. Next choose $\{f_5,\dots, f_{r+2}\}$ to be an orthonormal basis of the complement of $V_1$ in ${\mathbb{R}}^r\times \{0\}$ and the remaining vectors $\{f_{r+3},\dots, f_{n}\}$ to be a basis of the complement of $V_2$ in $\{0\}\times {\mathbb{R}}^{r'}$.
With this choice of body frame, $q_0$ and $p_0$ are of the form given in . However, contrary to the situation encountered in the proof of Theorem\[thm:nD to 3D\], the corresponding change of basis matrix $P_b\notin G_R$. With the new choice of body frame, the mass tensor ${\mathbb{J}}$ transforms to $${\mathbb{J}}'= {\mathop\mathrm{diag}\nolimits}[J_1,J_1, J_2, J_2, \underbrace{J_1,\dots,J_1}_{r-2},\underbrace{J_2,\dots,J_2}_{r'-2}].$$
Now, for the space frame, $e_1=g_0q$ so is contained in the image $g_0(V)$, and complete $\{e_1\}$ to an orthonormal basis of $g_0(V)$. Finally let $e_j=g_0 f_j$ for $j>4$. The change of basis matrix $P_s\in G_L$ and the matrices $g_0$ and $\Omega_0$ are in the forms given in .
To conclude invariance of the set under the dynamics we apply a symmetry argument as in the proof of Theorem\[thm:nD to 3D\]. This time we consider the group $H$ with elements: $$h= \left[\begin{array}{ccc}
\mbox{Id}_4& 0& 0 \\
0 & h_{11} & 0\\
0 & 0 &h_{22} \end{array}\right], \qquad h_{11} \in {\mathrm{O}}(r-2), \quad h_{22}\in {\mathrm{O}}(r'-2).$$ Note that for $h\in H$ we clearly have $h^T{\mathbb{J}}'h={\mathbb{J}}'$ so $H \subset G_R$. Similar to the argument given in the proof of Theorem\[thm:nD to 3D\], all vectors $(q,p)$ in are fixed by the $H$ action on $T^*{\mathrm{S}}^{n-1}$ and the matrices $(g,\Omega)$ are fixed by the action of the diagonal subgroup $H_\Delta$ on $D$ where $$H_\Delta = \{ (h,h)\in G= G_L\times G_R \, | \, h\in H\}.$$
It can readily be seen that the equations of motion on these fixed point subspaces is the same as those of the 4D cylindrical Veselova top.
We now describe the dynamics of the cylindrical Veselova top in $D$, as well as its first reduction in $T^*{\mathrm{S}}^{n-1}$. Since in ${\mathcal{R}}$ the motion is entirely periodic, the lifted motions are known as *relative periodic orbits*. We take advantage of the previous theorem and assume, without loss of generality, that $n=4$.
\[thm:reconstruction-cylindrical\] The reconstruction to $T^*{\mathrm{S}}^3$ and to $D$ of the solutions of the reduced equations are as follows.
1. In $S_0,S_0'$ the solutions are equilibria.
2. In $S_1,S_1'$, the relative equilibria are steady rotations in pure inertial planes (for $S_1$ they rotate in a principal plane $\Pi$ with ${\mathop\mathrm{in}\nolimits}(\Pi)=2J_2$ while in $S_1'$ it has ${\mathop\mathrm{in}\nolimits}(\Pi)=2J_1$).
3. In $S_2$ we have two following possibilities. If $P=0$ the top is in equilibrium, whereas if $P>0$, the solutions are steady rotations in mixed inertial planes, and they are not relative equilibria.
4. In $S_3,S_3'$ the motion lies on tori of dimension at most 2 in both $T^*{\mathrm{S}}^3$ and $D$ and behaves in the same way as the 3D axisymmetric Veselova top.
5. In $S_4$ we have two possibilities. In the (relative) equilibria in $S_4$ the motion lies on tori of dimension at most 2 in $T^*{\mathrm{S}}^3$ and at most 3 in $D$. These relative equilibria are $G_R$-Lyapunov stable. Finally, over the periodic orbits in $S_4$ the motion lies on tori of dimension at most 3 in $T^*{\mathrm{S}}^3$ and at most 4 in $D$.
\(i) $S_0$ and $S_0'$ have $p=0$ and hence are equilibria.
\(ii) and (iii): the proofs are entirely analogous to the proof of Proposition \[prop:steady-rot-strata\] (the case where $P=0$ in (iii) is not, but is immediate).
\(iv) Suppose $(q,p)\in S_3$ ($S_3'$ is similar). Then their projections $q_1$ and $p_1$ are parallel, say to a vector $u$, while $q_2$ and $p_2$ are linearly independent. The line ${\mathbb{R}}u$ is the fixed point set of the isotropy group ${\mathrm{O}}(1)={\mathbb{Z}}_2$, and hence the dynamics will take place on the 3-dimensional subspace spanned by $u,q_2$ and $p_2$. The resulting equations of motion are precisely those of the 3-dimensional axisymmetric top with ${\mathbb{J}}={\mathop\mathrm{diag}\nolimits}[J_1,J_2,J_2]$, and the conclusions about the invariant tori are to be found in Theorem\[thm:reconstruction-axisymmetric\].
\(v) Denote by $\widehat{S}_4$ the stratum in $T^*{\mathrm{S}}^3$ corresponding to the stratum $S_4$ in ${\mathcal{R}}$. It follows from Table \[table:strata T\^\*S + R\] that the $G_R={\mathrm{O}}(2)\times{\mathrm{O}}(2)$ action on $\widehat{S}_4$ is free. The conclusion about the dynamics on $T^*{\mathrm{S}}^3$ then follows from Field’s theorem (Theorem\[thm:Field\]) since ${\mathrm{O}}(2)\times{\mathrm{O}}(2)$ has rank 2. Similarly, if $\widehat{\widehat{S}}_4$ denotes the corresponding stratum on $D$, then the action of $G=G_L\times G_R={\mathrm{O}}(3)\times{\mathrm{O}}(2)\times {\mathrm{O}}(2)$ on $\widehat{\widehat{S}}_4$ is free and the conclusion follows from the same theorem (since $G$ has rank 3). The stability statement follows from the Lyapunov stability of the equilibria on ${\mathcal{R}}$.
Note that the above theorem only gives upper bounds for the dimension of the invariant tori corresponding to the reconstruction of the solutions on the stratum $S_4$. We will now present some evidence to show that these bounds are generically attained and the dynamics does not take place in lower dimensional sub-tori (compare with Remark \[rmk:Subtori\]). We begin with the following proposition that considers the reconstruction of the non-trivial relative equilibria in $S_4$.
Let $(h,P)$ with $h>0$ be a point in the image of the energy-momentum map (see Figure\[fig:E-M for cylindrical\]) and consider the corresponding relative equilibrium described by item (iii) of Theorem \[thm:dynamics on R\]. A reconstruction of this solution to $T^*{\mathrm{S}}^3$ and to $D$ is respectively determined by: $$\begin{split}
q(t)=\left ( \sqrt{A_0}\cos \omega_1 t, - \sqrt{A_0}\sin \omega_1 t, \sqrt{1-A_0}\cos \omega_2 t, -\sqrt{1-A_0}\sin \omega_2 t \right )^T, \\
p(t)=\left ( -\sqrt{B_0}\sin \omega_1 t, -\sqrt{B_0}\cos \omega_1 t, - \sqrt{P-B_0}\sin \omega_2 t, - \sqrt{P-B_0}\cos \omega_2 t \right )^T,
\end{split}$$ and $$g(t)= \left[\begin{array}{cccc}
1&0 & 0 & 0\\
0 &1& 0 & 0\\
0 & 0 & \cos \omega_3 t & -\sin \omega_3 t \\
0 & 0 &\sin \omega_3 t & \cos \omega_3 t
\end{array}\right] \, g_0 \, \left[\begin{array}{cccc}
\cos \omega_1 t &- \sin \omega_1 t & 0 & 0\\
\sin \omega_1 t & \cos \omega_1 t& 0 & 0\\
0 & 0 & \cos \omega_2 t & -\sin \omega_2 t \\
0 & 0 &\sin \omega_2 t & \cos \omega_2 t
\end{array}\right],$$ where $$g_0=\left[\begin{array}{cccc}
\sqrt{A_0}&0 & \sqrt{1-A_0} & 0\\
0 &\sqrt{B_0/P}& 0 & \sqrt{(P-B_0)/P} \\
\sqrt{1-A_0} & 0 & -\sqrt{A_0} & 0 \\
0 & \sqrt{(P-B_0)/P} &0 & -\sqrt{B_0/P}
\end{array}\right].$$ The frequencies in the above formulae are given by $$\label{eq:Reconstruction frequencies}
\omega_1=\frac{2\sqrt{2}h}{\sqrt{4J_1h+P}}, \qquad \omega_2=\frac{2\sqrt{2}h}{\sqrt{4J_2h+P}}, \qquad
\omega_3=\frac{-4h\sqrt{P}}{\sqrt{4J_1h+P}\sqrt{4J_2h+P}}.$$
The proof for $T^*{\mathrm{S}}^3$ is a direct verification that $(q(t),p(t))$ is a solution of Equations . For $D$, one may compute $\Omega=g^{-1}\dot g$ and verify by a direct calculation that the nonholonomic constraints and relations such as hold. For generic $(h,P)$ the reconstruction frequencies are non-resonant so these relative equilibria indeed reconstruct to motions on invariant 3-tori on $D$ and on an invariant 2-tori on $T^*{\mathrm{S}}^3$.
Now consider the motion along a relative periodic orbit on ${\mathcal{R}}$ that is close to the relative equilibrium treated above. Both the reconstructed solutions on $T^*{\mathrm{S}}^3$ and on $D$ are quasi-periodic. They will posses the natural frequency given by Equation (that only depends on $h$) and reconstruction frequencies that are close to those given in . The joint set of these frequencies will generically be non-resonant and the motion will take place on a 4-torus on $D$ and on a 3-torus on $T^*{\mathrm{S}}^3$.
Conclusions and open question {#S:Question-Hamiltonisation}
=============================
Our study shows that the dynamics of the axisymmetric and cylindrical Veselova tops is quasiperiodic in the natural time variable. More precisely, at the level of the first reduced space $D/G_L=T^*{\mathrm{S}}^{n-1}$ the flow possesses an invariant measure and:
1. for the axisymmetric case (we may take $n=3$), the generic motion is quasi-periodic on invariant $2$-tori on $T^*{\mathrm{S}}^2$ and there are two integrals of motion $H$ and $P$;
2. for the cylindrical case (we may take $n=4$), the generic motion is quasi-periodic on invariant $3$-tori on $T^*{\mathrm{S}}^3$ and there are three integrals of motion $H$, $P$ and $F$.
These scenarios are remarkably similar to the situation found in Liouville-Arnold integrable Hamiltonian systems. We leave it as an open problem to find a Hamiltonian structure for these systems on these cotangent bundles (one that does not involve a time reparametrisation) or to identify an obstruction to its existence. The interest in this question arises in view of the great quantity of research that in recent years has been devoted to identifying the differences between nonholonomic and Hamiltonian systems.
Field’s theorem on reconstruction {#app:Field}
=================================
We briefly recall the language and main properties of (compact) group actions and equivariant vector fields, needed for the proofs of quasiperiodicity.
Suppose a (compact Lie) group $G$ acts smoothly on a manifold $M$. If two points lie in the same orbit, their isotropy subgroups are conjugate. This motivates the partition or stratification of $M$ into “orbit types,” where the orbit type stratum $M_{(H)}$ consists of those points with isotropy subgroup (ISG) conjugate to $H$, or more precisely a stratum is a connected component of such points. Moreover, if $M_H$ is (a connected component of) the set of points with ISG equal to $H$, then, up to connected components, $M_{(H)} = G.M_H$ (the image of $M_H$ under the group action). The orbit space $M/G$ is similarly stratified by the orbit type; one denotes its corresponding stratum by $(M/G)_H$. All of these spaces $M_{(H)}, M_H$ and $(M/G)_H$ are manifolds.
If in addition there is an equivariant dynamical system on $M$, then it descends to a dynamical system on $M/G$ which respects each stratum. The question of reconstruction is, if one knows some properties of the dynamics on $M/G$ what does this imply about the dynamics on $M$. We use the following simplified version of an important theorem due to Field [@Field80] (see also [@Field-book Chapter 8] or [@CDS-book]).
\[thm:Field\] Consider a smooth equivariant dynamical system, with symmetry group $G$ acting freely on the phase space $M$. The dynamics passes down to the smooth orbit space $M/G$.
1. Let $x\in M/G$ be an equilibrium point of the reduced equations. Then the inverse image of $x$ in $M$ is a group orbit which is foliated by invariant tori of dimension at most the rank of $G$.
2. Let $\gamma$ be a periodic orbit of the reduced dynamics in $M/G$. Then the inverse image of this curve is also foliated by invariant tori, but now of dimension at most $\mathrm{rk}(G)+1$.
In both cases the dynamics in the invariant tori is quasiperiodic.
In case (i) the dynamics on $M$ is called a *relative equilibrium*, while in (ii) it is a a *relative periodic orbit*. If the dynamics on $M/G$ include an invariant quasiperiodic torus, then it is unknown except in special cases, see [@FGNG], what the corresponding reconstructed dynamics may be. The more general version of Field’s theorem does not require the action to be free, but this suffices for our purposes.
Physical inertia tensors satisfying the hypothesis of Fedorov-Jovanović {#A:Inertia}
=======================================================================
Recall that Fedorov and Jovanović [@FedJov; @FedJov2] work under the assumption that there is an $n\times n$ diagonal matrix $A$ such that the inertia tensor ${\mathbb{I}}:{\mathfrak{so}}(n)\to{\mathfrak{so}}(n)$ satisfies $$\label{eq:inertia-tensor}
{\mathbb{I}}(a\wedge b) = (Aa)\wedge(Ab),\quad (\forall a,b\in{\mathbb{R}}^n).$$ In this appendix we examine the feasibility of this condition within the family of physical inertia tensors. The following proposition, that was already suggested in [@FedJov2; @Jovan], shows that for $n\geq 4$ a physical inertia tensor that satisfies necessarily corresponds to an axisymmetric rigid body.
\[P:phys-inertia\] Let ${\mathbb{I}}$ be a physical inertia tensor defined by ${\mathbb{I}}(\Omega)={\mathbb{J}}\Omega+\Omega{\mathbb{J}}$ (see ).
1. If $n=3$ then the identity holds for arbitrary $a, b\in {\mathbb{R}}^3$ with $$A_i=\sqrt{\frac{(J_i+J_j)(J_i+J_k)}{J_j+J_k}},$$ for $\{i,j,k\}=\{1,2,3\}$, where $\mathbb{J}={\mathop\mathrm{diag}\nolimits}[J_1,J_2,J_3]$ (with $J_i+J_j>0$ for $i\neq j$).
2. If $n\geq 4$, there exists a diagonal matrix $A$ satisfying if and only if ${\mathbb{I}}$ is the inertia tensor of an axisymmetric body. In this case, the body frame can be chosen in such way that ${\mathbb{J}}$ is given by $$\label{eq:Lagrange-J}
{\mathbb{J}}=\mbox{diag}\left[A_1A_2-\tfrac12 A_2^2,\tfrac12A_2^2,\dots, \tfrac12 A_2^2\right]$$ and $A={\mathop\mathrm{diag}\nolimits}[A_1, A_2, \dots, A_2]$, where $A_1,A_2\in {\mathbb{R}}$ satisfy $A_1\geq A_2/2>0$.
\(i) is a calculation.
\(ii) “Only if”: by selecting the body frame with $f_1$ parallel to the symmetry axis of the body we have $\mathbb{J}=(J_1,J_2,\dots,J_2)$ with $J_1\geq 0$ and $J_2>0$. The statement follows by putting $A_1=(J_1+J_2)/\sqrt{2J_2}$ and $A_2=\sqrt{2J_2}$
“If": suppose that holds and ${\mathbb{I}}$ is a physical inertia tensor with corresponding diagonal matrix ${\mathbb{J}}$. Then there are real numbers $A_1,\dots, A_n$ such that for each $i\neq j$, $$J_i+J_j=A_iA_j.\eqno(E_{ij})$$ We claim that if the $J_i$ satisfy these conditions, then at least $n-1$ of the $A_i$ coincide. In that case suppose (w.l.o.g.) $A_2=\cdots=A_n$. Then comparing the equations $E_{1i}$ shows that all $J_1+J_i$ coincide for $i>1$, and hence $J_2=\cdots=J_n$ as required.
To prove the claim, note that for any collection $\{i,j,k,\ell\}$ of 4 distinct indices, $$E_{ij}+E_{ik}-E_{jk} \text{ becomes } 2J_i = A_iA_j+A_iA_k-A_jA_k,$$ and similarly $$E_{ij}+E_{i\ell}-E_{j\ell} \text{ becomes } 2J_i = A_iA_j+A_iA_\ell-A_jA_\ell.$$ Subtracting these shows $$(A_i-A_j)(A_k-A_\ell)=0.$$ Suppose $J_1\neq J_2$ and apply this to the indices $1,2,i,j$ with $i,j>2$. It follows that $A_3=A_4=\dots= A_n$. Finally, if $A_1=A_3$ we are done, while if $A_1\neq A_3$ apply the previous reasoning to the indices $1,3,2,4$ and one can conclude that $A_2=A_4$, and hence indeed $A_2=\cdots=A_n$.
The usefulness of a condition of type for the developments in [@FedJov; @FedJov2] (also [@Jovan; @JovaRubber]) seems to arise from the fact that an inertia tensor satisfying maps the set of rank 2 matrices in ${\mathfrak{so}}(n)$ into itself. For interest, although we make no use of this, we show that for $n\geq4$, the only physical inertia tensors with this property are axisymmetric.
Let ${\mathbb{I}}$ be a physical inertia tensor defined by . If $n\geq 4$, then ${\mathbb{I}}$ maps the set of rank two matrices in ${\mathfrak{so}}(n)$ into itself if and only if the body is axisymmetric.
Let us consider the case $n=4$. Suppose ${\mathbb{I}}$ maps the space of rank two matrices in ${\mathfrak{so}}(4)$ into itself and that the body is not axisymmetric. Choose the body frame so that ${\mathbb{J}}={\mathop\mathrm{diag}\nolimits}[J_1, \dots, J_4]$ with $J_1\neq J_3$, $J_2\neq J_4$. A direct calculation shows that $$\det({\mathbb{I}}((f_1+f_3)\wedge (f_2+f_4)))=(J_1-J_3)^2(J_2-J_4)^2\neq 0,$$ which shows ${\mathbb{I}}((f_1+f_3)\wedge (f_2+f_4))$ has rank 4, reaching a contradiction. A similar argument shows this implication for $n>4$.
The converse statement follows from part (ii) of the Proposition\[P:phys-inertia\].
**Acknowledgements:** This research was made possible by a Newton Advanced Fellowship from the Royal Society, ref: NA140017. LGN and JM are grateful to the hospitality of the Department of Mathematics Tullio Levi-Civita of the Univeristy of Padova, during its 2018 intensive period “Hamiltonian Systems”. LGN acknowledges the Alexander Von Humboldt Foundation for a Georg Forster Research Fellowship that funded a visit to TU Berlin where the last part of this work was completed.
The authors are thankful to Božidar Jovanović for comments on an early draft of this paper and for sharing the recent preprint [@Gajic] with us.
[99]{}\
,
[<span style="font-variant:small-caps;">Arnold, V.I.</span>\
]{} *Mathematical Methods of Classical Mechanics* 2nd ed., Springer, 1988.
[<span style="font-variant:small-caps;">Bloch, A.M., P.S. Krishnaprasad, J.E. Marsden and R.M. Murray</span>\
]{} Nonholonomic mechanical systems with symmetry. *Arch. Ration. Mech. Anal.* [[**[136]{}**]{}]{} (1996), 21–99.
[<span style="font-variant:small-caps;">Cushman R, J.J. Duistermaat and J. Śniatycki</span>\
]{} *Geometry of nonholonomically constrained systems*. Advanced Series in Nonlinear Dynamics, 26. World Scientific Publishing, 2010.
[<span style="font-variant:small-caps;">Chaplygin, S.A.</span>\
]{} On the theory of the motion of nonholonomic systems. The reducing-multiplier theorem. *Regul. Chaotic Dyn.* [[**[13]{}**]{}]{}, 369–376 (2008) \[Translated from *Matematicheskiǐ Sbornik* (Russian) [[**[28]{}**]{}]{} (1911), by A. V. Getling\]
[<span style="font-variant:small-caps;">Duistermaat, J.J. and J.A.C. Kolk</span>\
]{}
. Springer (2000).
[<span style="font-variant:small-caps;">Ehlers, K., J. Koiller, R. Montgomery and P.M. Rios</span>\
]{} Nonholonomic Systems via Moving Frames: Cartan Equivalence and Chaplygin Hamiltonization. in *The breath of Symplectic and Poisson Geometry*, Progress in Mathematics Vol. [[**[232]{}**]{}]{} (2004), 75–120.
[<span style="font-variant:small-caps;">Fassò, F., A. Giacobbe and N. Sansonetto</span>\
]{} Periodic flows, rank-two Poisson structures, and nonholonomic mechanics. *Regul. Chaotic Dyn.*, [[**[10]{}**]{}]{} (2005), 267–284.
[<span style="font-variant:small-caps;">Fassò F. and A. Giacobbe</span>\
]{} Geometry of invariant tori of certain integrable systems with symmetry and an application to a nonholonomic system. *SIGMA Symmetry Integrability Geom. Methods Appl.* [[**[3]{}**]{}]{} (2007), Paper 051, 12 pp.
[<span style="font-variant:small-caps;">Fassò, F., L.C. García-Naranjo and A. Giacobbe</span>\
]{} Quasiperiodicity in relative quasiperiodic tori. *Nonlinearity* [[**[28]{}**]{}]{} (2015), 4281–4301.
[<span style="font-variant:small-caps;">Fassò and N. Sansonetto</span>\
]{} Conservation of ‘moving’ energy in nonholonomic systems with affine constraints and integrability of spheres on rotating surfaces. *J. Nonlinear Sci.* [[**[26]{}**]{}]{} (2016), 519–544.
[<span style="font-variant:small-caps;">Fedorov, Y. N., and V.V. Kozlov</span>\
]{} Various aspects of $n$-dimensional rigid body dynamics. *Amer. Math. Soc. Transl. (2)* [[**[168]{}**]{}]{} (1995), 141–171.
[<span style="font-variant:small-caps;">Fedorov, Y. N. and B. Jovanović</span>\
]{} Nonholonomic LR systems as generalized Chaplygin systems with an invariant measure and flows on homogeneous spaces. *J. Nonlinear Sci.* [[**[14]{}**]{}]{} (2004), 341–381.
[<span style="font-variant:small-caps;">Fedorov, Y. N. and B. Jovanović</span>\
]{} Hamiltonization of the generalized Veselova LR system. *Regul. Chaot. Dyn.* [[**[14]{}**]{}]{} (2009), 495–505.
[<span style="font-variant:small-caps;">Field, M.J.</span>\
]{} Equivariant dynamical systems, *Trans. Am. Math. Soc.* [[**[259]{}**]{}]{} (1980), 185–205.
[<span style="font-variant:small-caps;">Field, M.J.</span>\
]{} *Dynamics and Symmetry*. Imperial College Press, 2007.
[<span style="font-variant:small-caps;">Gajić B. and B. Jovanović.</span>\
]{} Nonholonomic connections, time reparametrizations, and integrability of the rolling ball over a sphere. Preprint.
[<span style="font-variant:small-caps;">García-Naranjo, L.C. and J. Montaldi</span>\
]{} Gauge momenta as Casimir functions of nonholonomic systems. *Arch. Ration. Mech. Anal.* **228** (2018), 563–602.
[<span style="font-variant:small-caps;">Hermans J.</span>\
]{} A symmetric sphere rolling on a surface, *Nonlinearity* **8** (1995), 493–515.
[<span style="font-variant:small-caps;">Izosimov, A</span>\
]{} Stability of relative equilibria of multidimensional rigid body. *Nonlinearity* [[**[27]{}**]{}]{} (2014), 1419–1443.
[<span style="font-variant:small-caps;">Jovanović, B.</span>\
]{} LR and L+R systems. *J. Phys. A* [**42**]{} (2009), 18 pp.
[<span style="font-variant:small-caps;">Jovanović, B.</span>\
]{} Hamiltonization and integrability of the Chaplygin sphere in ${\mathbb{R}}^n$. *J. Nonlinear Sci.* [[**[20]{}**]{}]{} (2010), 569–593.
[<span style="font-variant:small-caps;">Jovanović, B.</span>\
]{} Rolling balls over spheres in ${\mathbb{R}}^n$. arXiv:1804.03697 (2018).
[<span style="font-variant:small-caps;">Koiller, J.</span>\
]{} Reduction of some classical nonholonomic systems with symmetry. *Arch. Ration. Mech. Anal.* [[**[118]{}**]{}]{} (1992), 113–148.
[<span style="font-variant:small-caps;">Ratiu, T.S.</span>\
]{} The motion of the free n-dimensional rigid body. *Indiana Univ. Math. J.* [[**[29]{}**]{}]{} (1980), 609–629.
[<span style="font-variant:small-caps;">Veselov, A.P. and L.E. Veselova</span>\
]{} *Integrable Nonholonomic Systems on Lie Groups*. Mat. Notes [[**[44]{}**]{}]{} (5-6) (1988), 810-819. \[Russian original in *Mat. Zametki* [[**[44]{}**]{}]{} (1988), no. 5, 604–619.\]
[<span style="font-variant:small-caps;">Veselova, L.</span>\
]{} *New cases of integrability of the equations of motion of a rigid body in the presence of a nonholonomic constraint.* Geometry, Differential Equations, and Mechanics (in Russian), Moscow State Univ. (1986), pp. 64-68.
FF: Dipartimento di Matematica ‘Tullio Levi Civita’, Università di Padova, Italy. [email protected]
LGN: IIMAS, UNAM, Mexico City, Mexico. [email protected]
JM: School of Mathematics, University of Manchester, UK. [email protected]
[^1]: This research was made possible by a Newton Advanced Fellowship from the Royal Society, no. NA140017
[^2]: Readers may be more familiar with using the connected component ${\mathrm{SO}}(n)$ as the configuration space, however allowing both components simplifies our exposition in Section \[sec:cylindrical\]; moreover Arnold [@MMCM p.133] suggests that ${\mathrm{O}}(n)$ is the ‘correct’ configuration space.
[^3]: For $a, b\in {\mathbb{R}}^n$ we denote $a\wedge b =ab^T-ba^T \in {\mathfrak{so}}(n)$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
For a lattice $L$ of ${\ensuremath{\mathbb{R}}}^n$, a sphere $S(c,r)$ of center $c$ and radius $r$ is called [*empty*]{} if for any $v\in L$ we have $\Vert v - c\Vert \geq r$. Then the set $S(c,r)\cap L$ is the vertex set of a [*Delaunay polytope*]{} $P=\operatorname{conv}(S(c,r)\cap L)$. A Delaunay polytope is called [*perfect*]{} if any affine transformation $\phi$ such that $\phi(P)$ is a Delaunay polytope is necessarily an isometry of the space composed with an homothety.
Perfect Delaunay polytopes are remarkable structure that exist only if $n=1$ or $n\geq 6$ and they have shown up recently in covering maxima studies. Here we give a general algorithm for their enumeration that relies on the Erdahl cone. We apply this algorithm in dimension $7$ which allow us to find that there are only two perfect Delaunay polytopes: $3_{21}$ which is a Delaunay polytope in the root lattice $\mathsf{E}_7$ and the Erdahl Rybnikov polytope.
We then use this classification in order to get the list of all types Delaunay simplices in dimension $7$ and found $11$ types.
address: 'Mathieu Dutour Sikirić, Rudjer Bosković Institute, Bijenicka 54, 10000 Zagreb, Croatia, Fax: +385-1-468-0245'
author:
- Mathieu Dutour Sikirić
bibliography:
- 'LatticeRef.bib'
title: The seven dimensional perfect Delaunay polytopes and Delaunay simplices
---
[^1]
Introduction
============
A lattice $L$ is a set of the form $L={\ensuremath{\mathbb{Z}}}v_1 + \dots + {\ensuremath{\mathbb{Z}}}v_n\subset {\ensuremath{\mathbb{R}}}^n$ with $(v_1, \dots, v_n)$ being independent. For such $L$ a sphere $S(c,r)$ of center $c$ and radius $r$ is called [*empty*]{} if for any $v\in L$ we have $\Vert v - c\Vert \geq r$. A polytope $P$ is called a [*Delaunay polytope*]{} if it is full-dimensional and if the vertex-set of $P$ is $S(c,r)\cap L$ with $S(c,r)$ an empty sphere. A Delaunay polytope is called [*perfect*]{} if any affine transformation $\phi$ such that $\phi(P)$ is a Delaunay polytope is necessarily an isometry of the space composed with an homothety.
In [@Erdahl1992; @DGL92] it was proved that for dimension $n\leq 5$ the only possible perfect Delaunay polytope is the interval $[0,1]$. Also in [@DGL92] it is proved that the Gosset’s polytopes $2_{21}$ and $3_{21}$ which are Delaunay polytopes of $\mathsf{E}_6$ and $\mathsf{E}_7$ are perfect. From the construction of infinite sequences in [@Dutour2005; @Grishukhin2006; @InfiniteER2002; @SecInfinite] we know that for any dimension $n\geq 6$ there exist perfect Delaunay polytopes. In [@Dutour2005] for any $n\geq 6$ we define a Delaunay polytope $ED_n$ of a lattice $LD_n$. The lattice $LD_n$ is formed by lamination over the root lattice $\mathsf{D}_{n-1}$ and we prove in [@InhomogeneousExtremeForms] that $ED_n$ is the unique Delaunay polytope of maximum circumradius of $LD_n$ and compute its covering density.
In [@Hyp7] we proved that the $ED_6 = 2_{21}$ is the unique perfect Delaunay polytopes in dimension $6$. This work uses a new approach in order to prove the following theorem:
\[Theorem3\_21\_ER7\] The $7$-dimensional perfect Delaunay polytopes are the Gosset polytope $ED_7 = 3_{21}$ and the Erdahl and Rybnikov polytope $ER_7$ [@35tope; @InfiniteER2002].
See Section \[DimSeven\_perfectDelaunay\].
Perfect Delaunay polytope are of importance for the theory of Covering Maxima. A [*covering maximum*]{} is a lattice $L$ such that its covering density is reduced if it is perturbed. In [@InhomogeneousExtremeForms] it is proved that a lattice $L$ is a covering maximum if and only if the Delaunay polytopes of maximum circumradius are perfect and [*eutactic*]{} (see [@InhomogeneousExtremeForms] for the definition). This characterization echoes Voronoi’s theorem [@VoronoiI] for the characterization of lattices of maximum density in terms of perfection and eutacticity. In [@InhomogeneousExtremeForms] we proved that $LD_n$ is one such covering maxima. Based on Theorem \[Theorem3\_21\_ER7\] and partial enumerations in dimension $8$, $9$ and $10$ we state the following conjecture:
\[UpperBoundConj\] For each $n\geq 6$, the lattice $LD_n$ defined in [@Dutour2005] has maximal covering density among all covering maxima.
The Minkowski conjecture [@Koksma p. 18] on the product of inhomogeneous forms has inspired a lot of research. Recently, it has been proved for $n\leq 8$ in [@HansGill1; @HansGill2; @HansGill3; @HansGill4] by computational methods based on Korkine-Zolotarev reduction theory. Other theoretical approaches have been attempted in [@McMullenCurtis; @StableLattices] by Dynamical System Theory. In particular the following theorem is proved in [@StableLattices Corollary 1.3]:
If Conjecture \[UpperBoundConj\] holds for a dimension $n\geq 1$ then Minkowski’s conjecture holds for dimension $n$.
As a consequence of the work of this paper we have that Minkowski’s conjecture is correct in dimension $7$, thereby confirming [@HansGill4].
We prove Theorem \[Theorem3\_21\_ER7\] by using the [*Erdahl cone*]{} which is defined as the set of polynomial functions $f$ of degree at most $2$ such that $f(x)\geq 0$ for $x\in {\ensuremath{\mathbb{Z}}}^n$. We already used this cone in [@InhomogeneousExtremeForms] for the study of covering maxima. We have then to do a kind of dual description computation with the problem that the number of defining inequalities is infinite, we have no local polyhedrality result as in the perfect form case (see [@bookschurmann] for details) and we are interested in only a subset of the extreme rays (see Theorem \[ClassifKindExtRay\] for the list of possible kinds of extreme rays of the Erdahl cone).
In [@Hyp7] we used a different approach, i.e. hypermetrics that allowed us to find all the $6$-dimensional perfect Delaunay polytopes. But this approach relied on previous work [@Ba; @BaRy] on $6$-dimensional Delaunay simplices that we could not extend easily to dimension $7$. Thus, it appears that the only way to classify the perfect $7$-dimensional Delaunay is to use the Erdahl cone. Moreover, we are able to use this classification in order to get the classification of Delaunay simplices:
\[EnumerationSimplices\] Up to arithmetic equivalence there are $11$ types of $7$-dimensional Delaunay simplices. The full list is given in Table \[TableFundamentalSimplices\].
See Section \[DimSeven\_simplices\].
In contrast to perfect Delaunay polytopes, the lattices simplices of this list (except the trivial simplex) had not been discovered before. A similar study has been undertaken in [@MinkowskianSublattices] for the set of shortest vectors of lattices. In view of this work it seems reasonable to think that the classification of Delaunay simplices is possible in dimension $8$. Equally importantly the classification of perfect Delaunay polytopes in dimension $8$ could be done and a conjectural list of the $27$ known possibilities is available in [@PerfectDelaunayLowDim].
$i$ Representative $S_i$ $\operatorname{vol}(S_i)$ $\left\vert \operatorname{Stab}(S_i)\right\vert$ Nb interval
----- ----------------------------------- --------------------------- -------------------------------------------------- -------------
1 $e_0, \dots, e_6,e_7$ 1 40320 127
2 $e_0, \dots, e_6,(1,1,1,1,1,1,2)$ 2 40320 63
3 $e_0, \dots, e_6,(0,0,1,1,1,1,2)$ 2 1440 63
4 $e_0, \dots, e_6,(1,2,2,2,2,2,3)$ 3 5040 42
5 $e_0, \dots, e_6,(1,1,1,2,2,2,3)$ 3 1152 42
6 $e_0, \dots, e_6,(0,1,1,2,2,2,3)$ 3 240 41
7 $e_0, \dots, e_6,(1,1,1,1,1,3,4)$ 4 1440 27
8 $e_0, \dots, e_6,(1,1,1,1,2,2,4)$ 4 240 31
9 $e_0, \dots, e_6,(1,1,1,2,2,3,4)$ 4 144 31
10 $e_0, \dots, e_6,(1,1,3,3,3,4,5)$ 5 72 24
11 $e_0, \dots, e_6,(1,1,1,1,2,3,5)$ 5 48 24
: Representative of Delaunay simplices in dimension $7$. $e_1, \dots, e_7$ is the standard basis of ${\ensuremath{\mathbb{Z}}}^7$ and $e_0=0$. $\operatorname{vol}(S)$ is $n!$ times the Euclidean volume of $S$. $\left\vert\operatorname{Stab}(S)\right\vert$ is the size of the lattice automorphism group preserving $S$. “Nb interval” is the number of Delaunay polyhedra of the type $\{0,1\}\times {\ensuremath{\mathbb{Z}}}^6$ in which $S_i$ is contained[]{data-label="TableFundamentalSimplices"}
In Section \[Sec:DelaunayPolyhedra\] Delaunay polyhedra are considered, their basic structure and relation to the Erdahl cone are introduced here. The facial structure of the Erdahl cone is reviewed in Section \[FacialErdahlCone\], in particular not all extreme rays of the Erdahl cone are related to Delaunay polyhedra [@Erdahl1992]. We also explain how the hypercube $[0,1]^n$ correspond to the cut polytope in the Erdahl cone. Section \[GeometryErdahlCone\] is not used in later sections. In it we construct a retraction of the Erdahl cone on the faces defined by Delaunay polyhedra. In Section \[SectionLtype\] we establish the link between the Erdahl cone and the classic $L$-type theory. In Section \[SectionHypermetric\] we do the same for the hypermetric cone. In Section \[SectionConn\] we give the connectivity and finiteness results on which our enumeration algorithm relies. Then we present in Section \[SectionAlgo\] our enumeration method, which is modelled on the Voronoi algorithm for perfect forms [@martinet] and on the adjacency decomposition method [@symsurvey]. In Section \[DimSeven\_perfectDelaunay\] we give the obtained results in the classification of $7$-dimensional perfect Delaunay polyhedra. In Section \[DimSeven\_simplices\] we use this classification to classify the $7$-dimensional types of Delaunay simplices.
Delaunay polyhedra {#Sec:DelaunayPolyhedra}
==================
Denote by $E_2(n)$ the vector space of polynomials of degree at most $2$ on ${\ensuremath{\mathbb{R}}}^n$ and by $\operatorname{AGL}_n({\ensuremath{\mathbb{Z}}})$ the group of affine integral transformations on ${\ensuremath{\mathbb{Z}}}^n$. The [*Erdahl cone*]{} is defined as $$Erdahl(n)=\{f\in E_2(n)\mbox{~such~that~}f(x)\geq 0\mbox{~for~}x\in {\ensuremath{\mathbb{Z}}}^n\}.$$ It is a convex cone of dimension $(n+1)(n+2)/2$ on which the group $\operatorname{AGL}_n({\ensuremath{\mathbb{Z}}})$ acts. All defining inequalities $f(x)\geq 0$ are equivalent under $\operatorname{AGL}_n({\ensuremath{\mathbb{Z}}})$ and therefore $Erdahl(n)$ is not polyhedral.
We denote $\cdot$ the standard scalar product on ${\ensuremath{\mathbb{R}}}^n$ defined by $x\cdot y = x^{T} y$. For a symmetric matrix $A$ and $x\in {\ensuremath{\mathbb{R}}}^n$ we define $A[x] = x^T A x$. We write any $f\in E_2(n)$ in the form $$f(x)=\operatorname{Cst}(f) + 2\operatorname{Lin}(f) \cdot x + \operatorname{Quad}(f)[x]$$ with $\operatorname{Cst}(f)\in {\ensuremath{\mathbb{R}}}$, $\operatorname{Lin}(f)\in {\ensuremath{\mathbb{R}}}^n$ and $\operatorname{Quad}(f)$ a $n\times n$ symmetric matrix. We define $S^n$ to be the set of symmetric matrices, $S^n_{>0}$ the set of positive definite matrices, $S^{n}_{\geq 0}$ the set of positive semidefinite matrices. We also define $Erdahl_{>0}(n)$ to be the set of $f\in Erdahl(n)$ with $\operatorname{Quad}(f)\in S^n_{>0}$. If $f\in Erdahl(n)$ then $\operatorname{Quad}(f)\in S^n_{\geq 0}$.
A [*sublattice*]{} of ${\ensuremath{\mathbb{Z}}}^n$ is a subgroup of ${\ensuremath{\mathbb{Z}}}^n$. An [*affine sublattice*]{} is one of the form $x_0 + L$ with $x_0\in {\ensuremath{\mathbb{Z}}}^n$ and $L$ a sublattice of ${\ensuremath{\mathbb{Z}}}^n$. A lattice $L\subset {\ensuremath{\mathbb{Z}}}^n$ is called [*saturated*]{} if $(L\otimes {\ensuremath{\mathbb{R}}})\cap {\ensuremath{\mathbb{Z}}}^n = L$. If $L_1$ and $L_2$ are two sublattices of ${\ensuremath{\mathbb{Z}}}^n$ we write $L_1\oplus_{{\ensuremath{\mathbb{Z}}}} L_2={\ensuremath{\mathbb{Z}}}^n$ if $L_1\cap L_2=\{0\}$ and $L_1\cup L_2$ generates ${\ensuremath{\mathbb{Z}}}^n$ over ${\ensuremath{\mathbb{Z}}}$. In that case both $L_1$ and $L_2$ are saturated.
Let us fix $n\geq 1$ and define:
\(i) A [*Delaunay polyhedron*]{} $D$ to be a set of the form $D=P_{L'}(D) + L(D)\subset {\ensuremath{\mathbb{Z}}}^n$ with
- $\operatorname{conv}(P_{L'}(D))$ a Delaunay polytope of an affine sublattice $L'$ of ${\ensuremath{\mathbb{Z}}}^n$,
- $L(D)$ a sublattice of ${\ensuremath{\mathbb{Z}}}^n$ and
- $L'\oplus_{{\ensuremath{\mathbb{Z}}}} L(D) = {\ensuremath{\mathbb{Z}}}^n$.
\(ii) A [*Delaunay simplex set*]{} $S$ to be a Delaunay polyhedron with $\vert S\vert = n+1$.
\(iii) A [*repartitioning set*]{} $R$ to be a Delaunay polyhedron with $\vert S\vert = n+2$.
The [*isotropy lattice*]{} $L(D)$ is uniquely determined by $D$ and its dimension is called the [*degeneracy rank*]{} of $D$ denoted $\operatorname{degrk}(D)$. Note that $D$ is the vertex set of a convex body only when $L(D)=0$. Also Delaunay polyhedra is full-dimensional, i.e. the smallest affine saturated lattice containing $D$ is ${\ensuremath{\mathbb{Z}}}^n$ itself.
The set $P_{L'}(D)$ is included in $L'$ and depends on $L'$. For any two choices $L'_1$ and $L'_2$ there exist a bijective affine map $\phi:L'_1\rightarrow L'_2$ with $\phi(P_{L'_1}(D)) = P_{L'_2}(D)$. When we consider properties that do not depend on the integral representation, we drop the lattice and write $P(D)$.
For $f\in Erdahl(n)$ we write $$Z(f)=\left\{x \in {\ensuremath{\mathbb{Z}}}^n\mbox{~such~that~} f(x)=0\right\}.$$
In classical geometry of numbers the essential tool is the quadratic form $Q$ instead of the quadratic function. The following establish a direct link between both:
For a Delaunay simplex set $S \subset {\ensuremath{\mathbb{Z}}}^n$ and $Q\in S^n$ there exists a unique function $f\in E_2(n)$ such that:
- $f(x) = 0$ for $x\in S$
- and $Q = \operatorname{Quad}(f)$.
This function is denoted $f_{S,Q}$ and depends linearly on $Q$.
The key reason for using Delaunay polyhedra is the following theorem:
\(i) If $D$ is a Delaunay polyhedron then there exist a function $f\in Erdahl(n)$ such that $D=Z(f)$.
\(ii) If $f\in Erdahl(n)$ then either $Z(f)$ is empty or there exist a $k$-dimensional saturated affine lattice $L\subset {\ensuremath{\mathbb{Z}}}^n$ such that $Z(f)$ is a Delaunay polyhedron of $L$.
\(i) Let us take a Delaunay polyhedron $D = P_{L'}(D) + L(D)$ with $P_{L'}(D)$ being a Delaunay polyhedron of a lattice $L'$ with $L' \oplus_{{\ensuremath{\mathbb{Z}}}} L(D) = {\ensuremath{\mathbb{Z}}}^n$. Let us denote by $S(c,r)$ the sphere around $P_{L'}(D)$ and write $L' = {\ensuremath{\mathbb{Z}}}v_1+\dots+{\ensuremath{\mathbb{Z}}}v_k$. The function $$\begin{array}{rcl}
f':{\ensuremath{\mathbb{Z}}}^n & \rightarrow & {\ensuremath{\mathbb{R}}}\\
x=(x_1,\dots,x_n) & \mapsto & \left\Vert \sum_{i=1}^n x_i v_i -c \right\Vert^2-r^2
\end{array}$$ belongs to $Erdahl(k)$. More precisely, $f'(x)=0$ if and only if $\sum_{i=1}^{k} x_i v_i \in P_{L'}(D)$. For $x\in {\ensuremath{\mathbb{Z}}}^n$ we write $x= x_1 + z$ with $x_1\in L'$ and $z\in L(D)$ and write $f(x)=f'(x_1)$. It is easy to prove that $f\in Erdahl(n)$ and $Z(f) = D$.
\(ii) This is [@Erdahl1992 Corollary 2.5].
We define the [*rational closure*]{} $S^{n}_{rat, \geq 0}$ to be the set of positive semidefinite forms whose kernel is defined by rational equalities.
If $f\in Erdahl(n)$ is such that $Z(f)$ is a Delaunay polyhedron then $\operatorname{Quad}(f)\in S^{n}_{rat, \geq 0}$.
Let us write $D=Z(f)$ and take a lattice $L' \subset {\ensuremath{\mathbb{Z}}}^n$ with $L'\oplus_{{\ensuremath{\mathbb{Z}}}} L(D) = {\ensuremath{\mathbb{Z}}}^n$. We write any $x\in {\ensuremath{\mathbb{Z}}}^n$ as $x = x_1 + z$ with $x_1\in L'$ and $z\in L(D)$. There is a quadratic function $f_1$ on $L'$ such that $f(x) = f_1(x_1)$ and $Z(f_1) = P_{L'}(D)$. Thus $\operatorname{Quad}(f_1)$ is positive definite and since $L(D)$ is an integral lattice the matrix $\operatorname{Quad}(f)$ belongs to $S^{n}_{rat, \geq 0}$.
Given a set $V\subset {\ensuremath{\mathbb{Z}}}^n$ we will need to be able to test if it is a Delaunay polyhedron or not. Algorithm \[TestRealizability\] does this iteratively for a finite point set by solving larger and larger linear programs until conclusion is reached. The algorithm can be easily adapted to the case of a point set of the form $R + L$ with $L$ a lattice and $R$ finite. The corresponding algorithm for perfect form is given in [@MinkowskianSublattices Algorithm 1].
\[TestRealizability\] $S_{vert} \leftarrow V$.\
$S_{vect} \leftarrow \emptyset$.\
If $D$ is a $n$-dimensional Delaunay polyhedron, then we define $$\operatorname{Aut}(D)=\{\phi\in \operatorname{AGL}_n({\ensuremath{\mathbb{Z}}})\mbox{~:~}\phi(D)=D\}.$$ When using Algorithm \[TestRealizability\] it is best to impose that the sought function $f$ is invariant under $\operatorname{Aut}(D)$ since it simplifies the search and a Delaunay polyhedron admits an invariant function (see Corollary \[InvFct\] below).
Before stating our result on the description of $\operatorname{Aut}(D)$, we remind on the notion of semidirect product. Given a group $G$, we call $G$ a semidirect product and write $G=N\times H$ if $N$ is a normal subgroup of $G$, $H$ a subgroup $G=NH$ and $N\cap H=\{e\}$.
\[AutomAut\_D\] If $D$ is a $n$-dimensional Delaunay polyhedron of degeneracy degree $d$ then we have the isomorphism $$\operatorname{Aut}(D)=\left\langle ({\ensuremath{\mathbb{Z}}}^d)^{1+n-d} \rtimes \operatorname{GL}_d({\ensuremath{\mathbb{Z}}})\right\rangle \rtimes \operatorname{Aut}(P(D)).$$
Let us take a basis $v=(v_i)_{1\leq i\leq d}$ of $L(D)$. Any automorphism of $D$ will send $v$ to another basis of $L(D)$ and this determines a component $\operatorname{GL}_d({\ensuremath{\mathbb{Z}}})$ of the automorphism group. Let us write $L'$ for an affine sublattice of ${\ensuremath{\mathbb{Z}}}^n$ such that $L'\oplus_{{\ensuremath{\mathbb{Z}}}} L(D) = {\ensuremath{\mathbb{Z}}}^n$. Such affine sublattice are determined by $1+n-d$ vectors in $L(D)$ and this determines the component $({\ensuremath{\mathbb{Z}}}^d)^{1+n-d}$ of the automorphism group. The last component comes from the fact that the automorphisms have to preserve the polytope $\operatorname{conv}(P(D))$.
We denote by $\operatorname{Aff}(D)$ the normal subgroup $({\ensuremath{\mathbb{Z}}}^d)^{1+n-d} \rtimes \operatorname{GL}_d({\ensuremath{\mathbb{Z}}})$ given in the above theorem.
\[InvFct\] A Delaunay polyhedron $D$ admits a function $f\in Erdahl(n)$ with $Z(f)=D$ that is invariant under $\operatorname{Aut}(D)$.
Let us write $L' \oplus_{{\ensuremath{\mathbb{Z}}}} L(D) = {\ensuremath{\mathbb{Z}}}^n$. A vector $x\in {\ensuremath{\mathbb{Z}}}^n$ is decomposed as $x=x_1 + z$ with $x_1\in L'$ and $z\in L(D)$. Any function $f$ with $Z(f) = D$ must have $\operatorname{Quad}(f)[z] = 0$ and $\operatorname{Lin}(f) \cdot z=0$. Therefore, there exist a function $f_1$ on $L'$ such that $f(x) = f_1(x_1)$. So, $f$ is invariant under $\operatorname{Aff}(D)$. The group acting on $f_1$ is $\operatorname{Aut}(P_{L'}(D))$, which is finite. The function $$f'_1(x_1) = \sum_{u\in \operatorname{Aut}(P_{L'}(D))} f(u(x_1))$$ is $\operatorname{Aut}(P_{L'}(D))$ invariant. Thus we get a function $f'(x) = f'_1(x_1)$ invariant under $\operatorname{Aut}(D)$.
If $D$ and $D'$ are two Delaunay polyhedra such that $D\subset D'$ then we define the stabilizer group $$\begin{array}{rcl}
\operatorname{Stab}(D, D') &=& \left\{\phi \in \operatorname{AGL}_n({\ensuremath{\mathbb{Z}}})\mbox{~:~}\phi(D)=D\mbox{~and~}\phi(D')=D'\right\}\\
&=& \operatorname{Aut}(D) \cap \operatorname{Aut}(D').
\end{array}$$
We have the following results
\[InclusionAff\] Let $D$ and $D'$ be two Delaunay polyhedra satisfying $D\subset D'$. Then:
1. we have $L(D)\subset L(D')$ and $\operatorname{Aff}(D) \subset \operatorname{Aff}(D')$.
2. There exist a finite group $G_1\subset \operatorname{Aut}(P(D))$ such that $$\operatorname{Stab}(D, D') = \operatorname{Aff}(D) \rtimes G_1$$ in particular $\operatorname{Aff}(D)$ is a finite index subgroup of $\operatorname{Stab}(D, D')$
We have $L(D)\subset L(D')$. Let us take a ${\ensuremath{\mathbb{Z}}}$-basis $\{e_1, \dots, e_d\}$ of $L(D)$ and complement it by adding $\{e_{d+1}, \dots, e_{d'}\}$ to a basis of $L(D')$. We can find $f_1, \dots, f_{n-d'}$ so that the $e_i$ and $f_j$ form a basis of ${\ensuremath{\mathbb{Z}}}^n$. The group $\operatorname{Aff}(D')$ is generated by translations along $f_1, \dots, f_{n-d'}$ and $\operatorname{GL}_{d'}({\ensuremath{\mathbb{Z}}})$. The group generated by translations along $\{e_{d+1}, \dots, e_{d'}\}$ and $\operatorname{GL}_d({\ensuremath{\mathbb{Z}}})$ directly embeds into $\operatorname{GL}_{d'}({\ensuremath{\mathbb{Z}}})$ and this determine the group inclusion. So (i) holds.
By (i) we have the inclusion $\operatorname{Aff}(D) \subset \operatorname{Stab}(D,D')$. Let us choose a lattice $L'$ with $L'\oplus_{{\ensuremath{\mathbb{Z}}}} L(D) = {\ensuremath{\mathbb{Z}}}^n$. Then if $u\in \operatorname{Stab}(D,D')\subset \operatorname{Aut}(D)$ we can find an unique element $n\in \operatorname{Aff}(D)$ such that $u n^{-1}$ stabilizes $L'$. Thus $u n^{-1}$ belongs to $\operatorname{Aut}(P_{L'}(D))$. The image defines the group $G_1$. the finite index property follows.
Facial structure of the Erdahl cone {#FacialErdahlCone}
===================================
The standard scalar product on $S^n$ is $\langle A, B\rangle = Tr(AB)$. We equip $Erdahl(n)$ with the inner product $$(f, g) = \operatorname{Cst}(f) \operatorname{Cst}(g) + 2 \operatorname{Lin}(f) \cdot \operatorname{Lin}(g) + \langle \operatorname{Quad}(f), \operatorname{Quad}(g) \rangle$$ and for each $x\in {\ensuremath{\mathbb{Z}}}^n$ we define the evaluation function $ev_x$ by $ev_x(y) = (1 + x\cdot y)^2$ such that $$(f, ev_x) = f(x).$$
A *convex cone* ${{\mathcal C}}$ is defined as a set invariant under addition and multiplication by positive scalars. ${{\mathcal C}}$ is called *full-dimensional* if the only vector space containing it is ${\ensuremath{\mathbb{R}}}^m$. ${{\mathcal C}}$ is called *pointed* if no linear subspace of positive dimension is contained in it. Let ${{\mathcal C}}$ be a full-dimensional pointed convex polyhedral cone in ${\ensuremath{\mathbb{R}}}^m$. Given $f \in ({\ensuremath{\mathbb{R}}}^{m})^{*}$, the inequality $f(x)\geq 0$ is said to be *valid* for ${{\mathcal C}}$ if it holds for all $x\in {{\mathcal C}}$. A *face* of ${{\mathcal C}}$ is a pointed polyhedral cone $\{ x \in {{\mathcal C}}\mbox{~}:\mbox{~} f(x) = 0 \}$, where $f(x) \geq 0$ is a valid inequality.
A face of dimension $1$ is called an *extreme ray* of ${{\mathcal C}}$; a face of dimension $m-1$ is called a *facet* of ${{\mathcal C}}$. The set of faces of ${{\mathcal C}}$ forms a partially ordered set under inclusion. We write $F\lhd G$ if $F\subset G$ and $\dim F=\dim G-1$. Two extreme rays of ${{\mathcal C}}$ are said to be *adjacent* if they generate a two-dimensional face of ${{\mathcal C}}$. Two facets of ${\mathcal C}$ are said to be *adjacent* if their intersection has dimension $m - 2$. Any $(m-2)$-dimensional face of ${{\mathcal C}}$ is called a *ridge* and it is the intersection of exactly two facets of ${{\mathcal C}}$.
By the Farkas-Minkowski-Weyl Theorem (see e.g. [@schrijver Corollary 7.1a]) a convex cone ${{\mathcal C}}$ is *polyhedral* if and only it is defined either by a finite set of *generators* $\{v_1,\ldots, v_N\} \subseteq {\ensuremath{\mathbb{R}}}^m$ or by a finite set of linear functionals $\{f_1, \ldots, f_M\}\subseteq ({\ensuremath{\mathbb{R}}}^m)^*$: $${{\mathcal C}}=
\bigg\{\sum_{i=1}^N \lambda_i v_i \mbox{~}:\mbox{~} \lambda_i\geq 0\bigg\} =
\bigg\{x\in {\ensuremath{\mathbb{R}}}^m \mbox{~}:\mbox{~} f_i(x)\geq 0\bigg\}.$$
Every minimal set of generators $\{v_1, \ldots, v_{N'}\}$ defining a polyhedral cone ${{\mathcal C}}$ has the property $$\{{\ensuremath{\mathbb{R}}}_+ v_1,\ldots, {\ensuremath{\mathbb{R}}}_+ v_{N'}\}=\{e \mbox{~}:\mbox{~} \mbox{$e$ extreme ray of ${{\mathcal C}}$}\}.$$ Every minimal set of linear functionals $\{f_1, \ldots, f_{M'}\}$ defining ${{\mathcal C}}$ has the property that $\{F_1, \ldots, F_{M'}\}$ with $F_i=\{x\in {{\mathcal C}}\mbox{~}:\mbox{~} f_i(x)=0\}$ is the set of facets of ${{\mathcal C}}$. The problem of transforming a minimal set of generators into a minimal set of linear functionals (or vice versa) is called the *dual description problem*.
In our work, we have to deal with Delaunay polyhedra with an infinite number of vertices and we cannot apply the Farkas-Minkowski-Weyl theorem to them nor of course existing dual-description software [@lrs; @cdd].
Let $D\subset {\ensuremath{\mathbb{Z}}}^n$ be a Delaunay polyhedron
\(i) We define the vector space $$Space(D)=\{f\in E_2(n)\mbox{~such~that~} f(x)=0\mbox{~for~}x\in D\}.$$
\(ii) The dimension of $Space(D)$ is called the [*perfection rank*]{} $\operatorname{rankperf}(D)$ and $D$ is [*perfect*]{} if $\operatorname{rankperf}(D)=1$.
Let $D\subset {\ensuremath{\mathbb{Z}}}^n$ be a Delaunay polyhedron
\(i) $Space(D) \cap Erdahl(n)$ is a face of $Erdahl(n)$ of dimension $\dim Space(D)$.
\(ii) If $D$ is perfect then $Space(D) \cap Erdahl(n)$ is an extreme ray of $Erdahl(n)$.
Let $p=\dim Space(D)$ and $g_1$, …, $g_p$ be a basis of $Space(D)$. Since $D$ is a Delaunay polyhedron there exist a function $f\in Erdahl(n)$ such that $D=Z(f)$. For each $1\leq i\leq p$ there exist $\lambda_i > 0$ such that $\lambda_i f + g_i\in Erdahl(n)$ and so (i) follows. (ii) follows directly from (i).
For a perfect Delaunay polyhedron $D$ we denote $f_D$ a generator of the extreme ray $Space(D)\cap Erdahl(n)$.
\[ClassifKindExtRay\]([@Erdahl1992 Theorem 2.1]) The generators of extreme rays of $Erdahl(n)$ are:
1. The constant function $f=1$
2. The functions of the form $f_{a,\beta}(x)=(a_1 x_1 + \dots a_n x_n + \beta)^2$ with $(a_1, \dots, a_n)$ not colinear to an integral vector.
3. The functions of the form $f_D$ with $D$ a perfect Delaunay polyhedron.
This theorem indicates that the structure of the extreme rays of $Erdahl(n)$ is more complicated than for a polytope. Since we are interested only in the third class of extreme rays, some reduction will be necessary and it turns out that we can work out everything with Delaunay polyhedra.
In this paper we will work with both spaces of functions in $E_2(n)$ and with point sets of Delaunay polyhedra:
Given a Delaunay polyhedron $D$ we define
\(i) The cone of admissible functions is defined as $$Erdahl_{supp}(D)=\left\{g\in E_2(n)\;:\; g(x)\geq 0 \mbox{~for~all~} x\in D\right\}.$$
\(ii) The cone of evaluation functions is defined as $$Erdahl_{supp}^*(D)= \left\{ \sum_{x\in D} \lambda_x ev_x \mbox{~with~} \lambda_x\geq 0\right\}.$$
\[BasicResultErdahlCone\] Let $D$ be a Delaunay polyhedron of perfection rank $r$.
\(i) $Erdahl_{supp}(D)$ is the product of a pointed convex cone $C_D$ with $Space(D)$.\
The dual of $C_D$ is $Erdahl_{supp}^*(D)$.
\(ii) Any Delaunay polyhedron $D'\subset D$ of perfection rank $r+1$ gives a facet of $Erdahl_{supp}^*(D)$.
\(iii) If $L(D) = 0$ then facets of $Erdahl_{supp}^*(D)$ correspond to Delaunay polyhedra $D'\subset D$ of perfection rank $r+1$.
\(i) If $f$ and $-f$ both belong to $Erdahl_{supp}(D)$ then $f(x)=0$ for $x\in D$ and so $f\in Space(D)$. So, $Erdahl_{supp}(D)$ is the sum of $Space(D)$ and a closed convex cone. The duality result follows from [@barvinok Part IV.5]) for closed full-dimensional convex cones.
\(ii) Let $f\in Erdahl(n)$ such that $Z(f) = D'$. Thus we have $f(x)=0$ on $D'$ and $f(x) > 0$ on $D - D'$ and so $f$ defines a facet of $Erdahl_{supp}^*(D)$.
\(iii) If $L(D) = 0$ then for any facet of $Erdahl_{supp}^*(D)$ we can find a set $D'\subset D$ and a function $f\in E_2(n)$ with $f(x)=0$ for $x\in D'$ and $f(x)>0$ for $x\in D - D'$. There exist a function $g\in Erdahl(n)$ such that $D=Z(g)$. Then we can find $\lambda > 0$ such that $f + \lambda g\in Erdahl(n)$. Then we have $D'=Z(f + (\lambda+1) g)$ and so $D'$ is a Delaunay polyhedron.
Theorem [@Erdahl1992 Theorem 2.1] shows that if $L(D)\not= 0$ there are other facets of $Erdahl_{supp}^*(D)$ than the ones from Delaunay polyhedra.
\[LowerBoundNrVertices\] If $D$ is a $n$-dimensional perfect Delaunay polyhedron with degeneracy degree $d$ then $P(D)$ has at least $${n-d + 2 \choose 2} -1$$ points.
The isotropy lattice $L(D)$ has dimension $d$ and we can choose a complement lattice $L'$ such that $L' \oplus_{{\ensuremath{\mathbb{Z}}}} L(D) = {\ensuremath{\mathbb{Z}}}^n$. If $f$ is a quadratic function with $Z(f) = D$ then $f$ is determined by its restriction to $L'$. Hence $f$ belongs to a vector space of dimension ${n-d + 2 \choose 2}$ and this gives the minimal number of determining inequalities.
A surprising relation has been found between the Erdahl cone of the hypercube $\{0,1\}^n$ and the cut polytope, which is classic polytope of Combinatorial Optimization [@DL]: Write $N=\{1,\dots, n\}$; if $S\subset N$, then the cut metric $\delta_S$ on $N$ is defined as follows: $$\delta_S(i,j)=\left\{\begin{array}{rl}
1 & \mbox{ if } |\{ij\}\cap S|=1,\\
0 & \mbox{ otherwise}.
\end{array}\right.$$ We have $\delta_S=\delta_{N-S}$ and the [*cut polytope*]{} $\operatorname{CUTP}_{n}$ is defined as the convex hull of the cut metrics $\delta_S$. The cone defined by the cut polytope is defined as $$\operatorname{CCUTP}_n = \left\{ \sum_{S\subset \{1,\dots,n\}} \lambda_S (1, \delta_S) \mbox{~with~} \lambda_S \geq 0\right\}.$$ The facets of the cone $\operatorname{CCUTP}_n$ are in one-to-one correspondence with the facets of the polytope $\operatorname{CUTP}_n$.
\[CUTn\_and\_ErdahlHn\] The polyhedral cone $Erdahl_{supp}^*(\{0,1\}^n)$ is linearly equivalent to $\operatorname{CCUTP}_{n+1}$.
The hypercube $[0,1]^n$ is defined as the convex hull of $2^n$ vectors $v=(v_1, \dots, v_n)$ with $v_i\in \{0,1\}$. For every such vector the evaluation function is $$ev_v(x)=1 + 2(v\cdot x) + (v\cdot x)^2.$$ Thus we can associate to $ev_v$ the vector $$(1, (v_i)_{1\leq i\leq n}, (v_iv_j)_{1\leq i\leq j\leq n}).$$ Since $v_i^2=v_i$ for $v_i\in \{0,1\}$ this vector family is linearly equivalent to $$(1, (v_i)_{1\leq i\leq n}, (v_i v_j)_{1\leq i < j\leq n}) \mbox{~for~}v\in \{0,1\}^n .$$ To the same $v$ we can associate the vector $\overline{v}=(0,v_1,\dots,v_n)$ and the set $S=\{i\in \{0,\dots, n\}, \mbox{~s.t.~} \overline{v}_i=1\} \subset \{0,\dots,n\}$. The cut metric $\delta_S$ on $\{0,\dots, n\}$ is characterized by $$(\delta_S(i,j) )_{0\leq i<j\leq n}$$ and since $\delta_S(i,j)=(\overline{v}_i - \overline{v}_j)^2$ the family of pairs $(1, \delta_S)$ is linearly equivalent to the family of evaluation map $ev_v$.
It is interesting to note that the symmetry group of the hypercube $[0,1]^n$ is of size $2^n n!$ but that the symmetry group of the cut polytope $\operatorname{CUTP}_{n+1}$ is of size $2^n (n+1)!$ for $n\not=3$ [@Relatives; @DL].
The Delaunay polyhedra retract {#GeometryErdahlCone}
==============================
For $f\in Erdahl(n)$, we define $Vect\,Z(f)$ to be the vector space spanned by difference of elements of $Z(f)$. We define $V_f=Vect\, Z(f) + \operatorname{Ker}\, \operatorname{Quad}(f)$. For a given $f\in Erdahl(n)$ we call [*proper pair*]{} a pair $(g,h)\in E_2(n)^2$ such that $g\in Erdahl(n)$, $h(x)\geq 0$ for $x\in {\ensuremath{\mathbb{R}}}^n$ and $f=g + h$.
\[KernelsOfH\] Let $f\in Erdahl(n)$.
\(i) For a proper pair $(g,h)$, $Z(f)\subset Z(g)$ and $\operatorname{Ker}\,\operatorname{Quad}(h)\subset V_f$.
\(ii) There exist a proper pair $(g,h)$ with $\operatorname{Ker}\,\operatorname{Quad}(h) = V_f$
If $x\in Z(f)$ then one sees that necessarily $h(x)=g(x)=0$. So, $h(x)=0$ for $x\in \operatorname{conv}(Z(f))$ which implies $Vect\, Z(f)\subset Ker\, \operatorname{Quad}(h)$. Also, it is clear that for any vector $v\in \operatorname{Ker}\,\operatorname{Quad}(f)$ we have $\operatorname{Quad}(g)[v]=\operatorname{Quad}(h)[v]=0$. Hence, (i) holds.
Let us denote by ${\ensuremath{\mathbb{Z}}}\operatorname{Ker}\,\operatorname{Quad}(f)$ the smallest subspace of ${\ensuremath{\mathbb{R}}}^n$ having an integral basis containing $\operatorname{Ker}\,\operatorname{Quad}(f)$. By the Decomposition Lemma 3.1 in [@Erdahl1992] there exist a $g\in Erdahl(n)$ with $\operatorname{Ker}\,\operatorname{Quad}(g) = {\ensuremath{\mathbb{Z}}}\operatorname{Ker}\,\operatorname{Quad}(f)$ and a positive semidefinite form $Q_1$ such that $$f(x) = g(x) + Q_1[x].$$ Let us denote by $V$, respectively $W$, an integral supplement of $\operatorname{Ker}\, \operatorname{Quad}(f)$ in ${\ensuremath{\mathbb{Z}}}\operatorname{Ker}\,\operatorname{Quad}(f)$, respectively ${\ensuremath{\mathbb{Z}}}\operatorname{Ker}\,\operatorname{Quad}(f)$ in ${\ensuremath{\mathbb{Z}}}^n$. Denote by $\phi_1$, …, $\phi_m$ some affine functions on ${\ensuremath{\mathbb{Z}}}^n$ such that $\phi_i({\ensuremath{\mathbb{Z}}}\operatorname{Ker}\,\operatorname{Quad}(f))=0$ and $$\{x\in W\mbox{~s.t.~} \phi_1(x)=\dots=\phi_m(x)=0\} = Vect\,Z(f) \cap W.$$ Then for $\epsilon>0$ small enough the function $g_1 - \epsilon\sum_{i=1}^m \phi_i(x)^2$ is still in $Erdahl(n)$. So, one gets that the pair $(f-h,h)$ with $h(x) = Q_1[x] + \epsilon\sum_{i=1}^m \phi_i(x)^2$ is proper and (ii) is true.
Let us call $W$ an integral supplement of $V_f$. Denote by $\operatorname{Quad}(f)\vert_{W}$ the quadratic form $\operatorname{Quad}(f)$ restricted to $W$. A proper pair $(g,h)$ is called [*extremal*]{} if $\det\,\operatorname{Quad}(h)|_{W}$ is maximal among all proper pairs. Lemma \[KernelsOfH\].(i) implies that the notion of being extremal is independent of the chosen subspace $W$, while Lemma \[KernelsOfH\].(ii) implies that there is at least one form of non-zero determinant.
\[UniqueDecomposition\] Let $f\in Erdahl(n)$.
\(i) An extremal proper pair has $Z(g)$ being a Delaunay polyhedron.
\(ii) There exist a unique extremal proper pair $(g,h)$.
Let us take an integral supplement $W$ as above and suppose to avoid trivialities that $W\not=\emptyset$. So, by restricting $f$ to $W$, we can assume that $\operatorname{Quad}(h)$ is positive definite. $$f(x) = g(x) + h(x)$$ with $h(x)\geq 0$ for $x\in {\ensuremath{\mathbb{R}}}^n$. This condition on $h$ is equivalent to $$A_h=\left(\begin{array}{cc}
\operatorname{Cst}(h) & \operatorname{Lin}(h)^T\\
\operatorname{Lin}(h) & \operatorname{Quad}(h)
\end{array}\right)$$ being positive definite. Hence we consider the following semidefinite programming problem: find the $A_h\in S^{n+1}_{\geq 0}$ maximizing $\det\,\operatorname{Quad}(h)$ and satisfying for all $x\in {\ensuremath{\mathbb{Z}}}^n$ $$f(x) \geq A_h[(1,x)] = h(x).$$ We also write $g=f-h$.
Suppose that $Z(g)$ is not a Delaunay polyhedron and that $\operatorname{Quad}(h)$ is positive definite. Then $Z(g)$ does not generates ${\ensuremath{\mathbb{R}}}^n$ as an affine space and so there exists an affine function $\phi$ such that $\phi( Z(g) )=0$. Then there exist $\alpha > 0$ such that the pair $(f-h', h')$ with $h'= h + \alpha \phi^2$ is still proper. Since $\det\,\operatorname{Quad}(h') > \det\,\operatorname{Quad}(h)$ the pair is not extremal and (i) holds.
Let us take $N=n(n+1)/2$ points $v_i\in {\ensuremath{\mathbb{Z}}}^n$ such that the family $\left\{(1,v_i)(1,v_i)^T\right\}_{1\leq i\leq N}$ is of full rank. The inequalities $g(v_i) \geq A[(1,v_i)] \geq 0$ implies that all coefficients of $A$ are bounded. Thus the problem is actually to minimize the convex function $h\mapsto -\log\det\,\operatorname{Quad}(h)$ over a compact convex set, hence existence follows.
Since $-\log\, \det$ is a strictly convex function, we know that if we have two optimal solutions $h_1$ and $h_2$ then $\operatorname{Quad}(h_1) = \operatorname{Quad}(h_2)$. Let us denote by $D_1=Z(g_1)$ and $D_2=Z(g_2)$ the corresponding Delaunay polyhedra. The function $h_{mid} =(h_1 + h_2)/2$ is also an optimal solution of the problem. We have $Z(g_{mid}) = D_1\cap D_2$. The set $D_1\cap D_2$ is necessarily a Delaunay polyhedron since otherwise we could still increase the determinant by the above construction and this would contradict the optimality. But if $Z(f)$ is a Delaunay polyhedron then the terms $\operatorname{Cst}(f)$ and $\operatorname{Lin}(f)$ are determined by $\operatorname{Quad}(f)$. So, one gets $h_1=h_2$ and the uniqueness is proved on the restriction to $W$. But Lemma \[KernelsOfH\](i) implies that once $\operatorname{Quad}(h)$ is known on $W$ then it is known on ${\ensuremath{\mathbb{Z}}}^n$. By the condition $h\geq 0$ the linear part is known as well.
Note that in the above determinant maximization problem a finite set of inequalities suffices to determine the optimal solution. This follows from the fact that since we are maximizing the determinant we can assume that the lowest eigenvalue of $\operatorname{Quad}(h)$ is bounded away from $0$, i.e. that there exist $c>0$ such that $\operatorname{Quad}(h) \geq c I_n$.
For $f\in Erdahl(n)$, we write $proj(f)=g$ and $proj'(f)=h$ with $(g,h)$ the unique extremal pair associated to $f$. From the unicity of extremal pairs we also get that $proj$ and $proj'$ commute with the action of $\operatorname{AGL}_n({\ensuremath{\mathbb{Z}}})$.
The function $proj$ is continuous.
Let us define $Erdahl_{dp}(n)$ to be the set of $f\in Erdahl(n)$ such that $Z(f)$ is a Delaunay polyhedron. The above conjecture if true implies that the set $Erdahl_{dp}(n)$ is simply connected and this could be of interest for topological applications. However, we were not able to prove the conjecture and instead we prove in later sections connectedness results which are sufficient for our purposes.
Relation with $L$-types theory {#SectionLtype}
==============================
In this section we reframe classical $L$-type theory from [@VoronoiII] (see also [@bookschurmann] for a modern account) in term of Erdahl cone and state several key lemmas.
Let $Q \in S^n_{rat,\geq 0}$. The Delaunay polyhedra tessellation ${\mathcal DPT}(Q)$ defined by $Q$ is the set of Delaunay polyhedra $D$ such that there exist a $f\in Erdahl(n)$ with
- $Z(f) = D$
- and $\operatorname{Quad}(f) = Q$.
If $Q$ is positive definite then the Delaunay polyhedra tessellation is the classical Delaunay polytope tessellation, i.e. all Delaunay polyhedra occurring are actually vertex sets of Delaunay polytopes. The number of translation classes of Delaunay polyhedra is always finite. These Delaunay polyhedra tessellations were considered in [@EquivariantLtypeDSV Section 2.2]. Efficient algorithm for the enumeration of Delaunay polytope tessellations are given in [@ComplexityVoronoiDSV].
From this one can define the $L$-type which are parameter spaces of Delaunay polytope tessellations:
Let us take a Delaunay polyhedra tessellation ${\mathcal T}$. Then the $L$-type $LT({\mathcal T})$ is defined as the closure of the set of quadratic forms $Q$ such that ${\mathcal DPT}(Q) = {\mathcal T}$. It is well known (see [@bookschurmann] and [@VoronoiII] for proofs) that $L$-types are polyhedral cones.
A $L$-type is called [*primitive*]{} if it is of maximal dimension, this is equivalent to say that all its Delaunay polyhedra are Delaunay simplex sets.
The set of all $L$-types for all possible Delaunay tessellations defines a tessellation of the cone $S^{n}_{rat, \geq 0}$.
Given two Delaunay polyhedra tessellation ${\mathcal T}$ and ${\mathcal T}'$ we say that [*${\mathcal T}'$ is a refinement of ${\mathcal T}$*]{} if every Delaunay polyhedron of ${\mathcal T}'$ is included in a single Delaunay polyhedron of ${\mathcal T}$. ${\mathcal T}'$ is a [*simplicial refinement*]{} if all its Delaunay polyhedra are Delaunay simplex sets.
\[ExistenceSimplicialRefinement\] Any Delaunay polyhedra tessellation ${\mathcal T}$ admits at least one simplicial refinement.
Let us denote by $L({\mathcal T})$ the space $L(D)$ of the Delaunay polyhedra $D$ occurring in the tessellation and by $Q \in S^n_{rat, \geq 0}$ the form realizing it. Let us take a lattice $L'$ such that $L' \oplus_{{\ensuremath{\mathbb{Z}}}} L({\mathcal T}) = {\ensuremath{\mathbb{Z}}}^n$. We write $x\in {\ensuremath{\mathbb{Z}}}^n$ as $x=x_1 + z$ with $x_1\in L'$ and $z\in L({\mathcal T})$. $D$ is a Delaunay polyhedron for the quadratic function $f\in Erdahl(n)$. Necessarily $f$ is of the form $f(x) = f_1(x_1)$ with $f_1$ a quadratic function on $L'$. Let us denote $(D_i)_{i\in I}$ the Delaunay polyhedra occurring in the tessellation. Let us take a basis $w_1, \dots, w_m$ of $L({\mathcal T})$ and define linear forms $\phi_i$ on ${\ensuremath{\mathbb{Z}}}^n$ such that $\phi_i(w_j) = \delta_{ij}$ and $\phi_i(L') = 0$. The quadratic form $$Q'[x] = Q[x] + \sum_{i=1}^m \left(\phi_i(x)\right)^2$$ is positive definite and the Delaunay polyhedra tessellation ${\mathcal DT}$ corresponding to $Q'$ is formed by the Delaunay polyhedra $$D_i + \sum_{k=1}^m \{a_k, a_k+1\} w_k \mbox{~with~}1\leq i\leq r \mbox{~and~} a_k\in {\ensuremath{\mathbb{Z}}}.$$ In particular ${\mathcal T}_2$ is a refinement of ${\mathcal T}$.
Since the $L$-type domain form a tiling of $S^n_{rat, \geq 0}$ the form $Q'$ belongs to at least one primitive $L$-type $LT({\mathcal T}_2)$. This $L$-type defines a Delaunay polyhedra tessellation by simplices which is a refinement of ${\mathcal T}_2$ and so of ${\mathcal T}$.
Let us take a primitive $L$-type ${\mathcal T}$. Any facet $F$ of ${\mathcal T}$ is determined by a pair of Delaunay simplex sets $S_1$ and $S_2$ in the Delaunay tessellation that determine a repartitioning set. We say that two facet defining repartitioning sets are in the same class if they define the same facet $F$ of ${\mathcal T}$. If $R$ is a repartitioning set then $\operatorname{conv}(R)$ admits exactly two triangulations (see [@bookschurmann Section 4.3.2]). One says that two primitive $L$-types are adjacent if their intersection is a codimension $1$ face in the cone $S^n_{>0}$. When we move from one $L$-type to another $L$-type, the Delaunay tessellation is changed and this is done combinatorially by the repartitioning sets. That is some Delaunay simplex sets are merged into repartitioning sets and the triangulation is changed to the other triangulation, thus yielding another $L$-type.
Given a Delaunay polyhedron $D$ a Delaunay polyhedra tessellation ${\mathcal T}$ is called [*$D$-proper*]{} if $D$ is the union of the Delaunay polyhedra $D'$ contained in $D$. We have following lemma:
\[ConvexityDproper\] Let $D$ be a Delaunay polyhedron. The graph formed by the primitive $L$-types whose corresponding Delaunay polyhedra tessellations is primitive and $D$-proper is connected.
Let us consider a function $f_D\in Erdahl(n)$ such that $Z(f_D) = D$. We can consider the triangulations induced by positive definite quadratic forms on $D$ itself. This set is connected by the theory of regular triangulations (see [@TriangulationBook] for an account).
Any triangulation ${\mathcal T}_{part}$ on $D$ induced by a positive definite quadratic form $Q$ can be extended to a triangulation ${\mathcal T}$ of ${\ensuremath{\mathbb{Z}}}^n$: It suffices to replace $Q$ by $Q + \lambda \operatorname{Quad}(f_D)$ for $\lambda$ sufficiently large. The reason is that $\operatorname{Quad}(f_D)$ will not change the Delaunay triangulation for Delaunay simplex sets contained in $D$.
Now given a primitive $L$-type $LT$ whose Delaunay polyhedra tessellation ${\mathcal T}$ is $D$-proper, we denote by ${\mathcal S}$ its set of Delaunay simplex sets included in $D$. We consider the following cone ${\mathcal C}({\mathcal S})$: $${\mathcal C}({\mathcal S}) = \left\{ Q\in S^n_{rat,\geq 0} \mbox{~s.t.~} f_{S, Q} (x) \geq 0 \mbox{~for~} S\in {\mathcal S}\mbox{~and~} x\in {\ensuremath{\mathbb{Z}}}^n - S\right\}.$$ This cone is convex and is an union of primitive $L$-types. Thus this set of $L$-types is connected. The connectedness follows by combining above results.
Delaunay polytopes restricted to the set $D$. When doing so, by the connectivity
\[ConnectivitySimplices\] Let us take $D$ a Delaunay polyhedron and two Delaunay simplex sets $S$ and $S'$ in $D$. Then there exist a sequence of Delaunay simplex sets $\{S=S_0,S_1, \dots, S_m=S'\}$ with $S_i\subset D$ for $0\leq i\leq m$ such that $S_i\cup S_{i+1}$ is a repartitioning set for $0\leq i\leq m-1$.
Let us take $f_D\in Erdahl(n)$ a function such that $Z(f_D)=D$. Take $f_S$, $f_{S'}$ the corresponding functions for $S$ and $S'$. Denote by ${\mathcal T}$ the Delaunay polyhedra tessellation defined by $\operatorname{Quad}(f_D)$, which obviously has $D$ as one of its component.
When we perturb $\operatorname{Quad}(f_D)$, we are changing the Delaunay tessellation. However, if we take $\epsilon > 0$ small enough, we can ensure that the Delaunay polyhedra tessellations ${\mathcal DPT}(\operatorname{Quad}(f_D + \epsilon f_S))$ and ${\mathcal DPT}(\operatorname{Quad}(f_D + \epsilon f_S'))$ are $D$-proper. By applying Proposition \[ExistenceSimplicialRefinement\] we can find simplicial refinement of those two tessellations which we name ${\mathcal TR}$ and ${\mathcal TR}'$ and are both $D$-proper. We call $LT$ and $LT'$ the corresponding primitive $L$-types.
By Lemma \[ConvexityDproper\] there exist a path between $LT$ and $LT'$ that uses only $D$-proper $L$-types. By following this path, we can change $S$ into another Delaunay simplex set $S_2$ in ${\mathcal TR}'$.
Denote by $f_1$, …, $f_r$ the facets of $LT'$. Every such facet corresponds to a family of repartitioning sets. We say that two Delaunay simplex sets included in $D$ are adjacent if their union is a repartitioning set which gives a facet of $LT'$. The Delaunay polyhedron $D$ is a coarsening obtained by merging all simplices, so the above defined graph is connected. This means that we can find a path from $S_2$ to $S'$.
Relation with hypermetric theory {#SectionHypermetric}
================================
We define the [*volume*]{} $\operatorname{vol}(S)$ of a Delaunay simplex set $S$ to be $n! Vol(\operatorname{conv}(S))$ with $Vol$ the Euclidean volume. This rescaled volume is an integer and satisfies $\operatorname{vol}(S)\leq n!$. The possible rescaled volumes $PossVol(n)$ are given in Table \[TableRescaledVolume\] for $n\leq 7$ and a super-exponential lower bound on $\max\, PossVol(n)$ is proven in [@SuperExpVolume]. The best known upper bound [@DL Proposition 14.2.4] is $$\label{UpperBound_PossVol}
\max\,PossVol(n) \leq n! \frac{2^n}{{2n\choose n}}.$$
\[TableRescaledVolume\]
$n$ $1$ $2$ $3$ $4$ $5$ $6$ $7$
-------------- --------- --------- --------- --------- ----------- ------------- -----------------
$PossVol(n)$ $\{1\}$ $\{1\}$ $\{1\}$ $\{1\}$ $\{1,2\}$ $\{1,2,3\}$ $\{1,2,3,4,5\}$
: Possible volume of lattice Delaunay simplices. See [@Ba] for the proof for $n\leq 6$ and Section \[DimSeven\_simplices\] for the proof for $n=7$.
Let us take two $n$-dimensional Delaunay polyhedron $D$, $D'$ with $D\subset D'$. We can define the [*generalized hypermetric cone*]{} $$\operatorname{Hyp}(D,D')=\left\lbrace f\in E_2(n) \mbox{~s.t~} f(x)=0\mbox{~if~}x\in D
\mbox{~and~}f(x)\geq 0\mbox{~if~}x\in D'\right\rbrace .$$
We have the inclusion $\operatorname{Hyp}(D,D')\subset Erdahl_{supp}(D')$ and $\operatorname{Hyp}(D,D')$ is a priori defined by an infinity of inequalities.
As a direct application we can express the $L$-type domains as intersection of generalized hypermetric cones:
Let ${\mathcal T}$ be a Delaunay polyhedra tessellation. Then we have $$LT({\mathcal T}) = \cap_{D\in {\mathcal T} } \operatorname{Quad}\operatorname{Hyp}(D, {\ensuremath{\mathbb{Z}}}^n).$$
The polyhedral cone $\operatorname{Hyp}(D,D')$ is polyhedral.
Let us take a Delaunay simplex set $S=\{v_0, \dots, v_n\} \subset D$ which exists by Proposition \[ExistenceSimplicialRefinement\]. If we prove the polyhedrality of $\operatorname{Hyp}(S,D')$ then $\operatorname{Hyp}(D,D')$ is polyhedral as well since it is obtained from $\operatorname{Hyp}(S, D')$ by adding equalities $f(x)=0$ for $x\in D - S$.
Suppose that a $v\in D'$ defines a relevant inequality. Then there exist a function $f$ such that $f(x)=0$ for $x\in S\cup \{v\}$ and $f(x) >0$ for $x\in D' - S \cup \{v\}$. Since $D'$ is a Delaunay polyhedron, there exist a function $g$ such that $g(x)=0$ for $x\in D'$ and $g(x) > 0$ for $x\in {\ensuremath{\mathbb{Z}}}^n - D'$. Then we can find $\lambda>0$ such that $f(x) + \lambda g(x) > 0$ for $x\in {\ensuremath{\mathbb{Z}}}^n - S \cup \{v\}$. As a consequence the polytope $\operatorname{conv}(S \cup \{v\})$ is a Delaunay polytope. This implies that for any $i\in \{0,\dots, n\}$ the Delaunay simplex set $S_{v,i} = \{v, v_0,\dots,v_{i-1},v_{i+1},\dots,v_n\}$ has $\operatorname{vol}(S_{v,i}) \leq n!$ (see proof of Theorem 14.2.1 in [@DL]). Hence the coefficients of $v$ are bounded by a bound depending only on $S$ and this proves that $\operatorname{Hyp}(S,D')$ is polyhedral.
For a Delaunay simplex set $S$ of volume $1$, the cone $\operatorname{Hyp}(S,{\ensuremath{\mathbb{Z}}}^n)$ is called the [*hypermetric cone*]{} and is studied in [@DL]. For other simplices, they are called [*Baranovski cone*]{} in [@bookschurmann]. The facets of the Baranovski cones are determined up to dimension $6$ in [@BaRy]. There is a correspondence between facets of $\operatorname{Hyp}(S,D)$ and repartitioning sets $P$ with $S\subset P\subset D$. That is the inequality $f(x)\geq 0$ defines a facet of $\operatorname{Hyp}(S,D)$ if and only if $S\cup\{x\}$ is a repartitioning set.
Connectivity results {#SectionConn}
====================
For a given Delaunay polyhedron $D$, let us write $${\mathcal C}_{r,d}(D)=\left\{\begin{array}{l}
D'\subset D\mbox{~s.t.~} D' \mbox{~Delaunay~polyhedron}\\
\operatorname{rankperf}(D')=r\mbox{~and~}\dim L(D')\leq d
\end{array}\right\}.$$
If ${\mathcal A}$ and ${\mathcal B}$, are sets of Delaunay polyhedra then the graph $Gr({\mathcal A},{\mathcal B})$ is the graph on ${\mathcal A}$ with two Delaunay polyhedra $D_1, D_2\in {\mathcal A}$ adjacent if and only if $D_1\cap D_2\in {\mathcal B}$.
If $D$ is a Delaunay polyhedron of perfection rank $r$ and degeneracy degree $d$ then ${\mathcal C}_{r+1, d}(D)$ is decomposed into a finite number of orbits under $\operatorname{Aut}(D)$.
Without loss of generality, we can write $D = P + {\ensuremath{\mathbb{Z}}}^d$. Let us write $D'\subset D$ as $D' = P' + L'$ with $k=\dim\,L'$. By applying an element of $\operatorname{Aut}(D)$, we can assume that $L'={\ensuremath{\mathbb{Z}}}^k$. So, without loss of generality, we can assume that $L'=0$. Let us take a Delaunay simplex set $S$ in $D'$; its volume is bounded by $\max\,PossVol(n)$. Again by using $\operatorname{Aut}(D)$ we can find a constant $C'$ such that the absolute value of the coordinates of $S$ in $P + {\ensuremath{\mathbb{Z}}}^d$ are bounded by $C'$. The polyhedrality of the cones $\operatorname{Hyp}(S, D)$ implies the finiteness.
\[TheCriticalLemma\] If ${\mathcal C}$ is a polyhedral cone, $F$ a face of ${\mathcal C}$ and $e$, $e'$ are two extreme rays which are not contained in $F$, then $e$ and $e'$ are connected by a path which does not intersect $F$.
By taking the intersection ${\mathcal C}\cap H$ with $H$ a suitable hyperplane, we can transform ${\mathcal C}$ into a polytope $P$ and $e$, $e'$ into vertices of $P$. We can find an affine function $\phi$ such that $\phi(x)\geq 0$ is a valid inequality on $P$ and $\phi(x)=0$ defines the face $F\cap H$ of $P$. By maximizing the function $\phi$ over $P$ and using the simplex algorithm (see [@schrijver; @zieglerLectureOnPolytopes]), we can find paths $p(v,v_{opt})$, $p(v',v_{opt})$ from $v$, $v'$ to an optimal vertex $v_{opt}$ such that $\phi$ is monotone on both paths. Since $\phi(v)>0$ and $\phi(v')>0$, such paths avoid the face $F$ and put together gives the required path.
\[ErdahlConnectivityResult\] If $D$ is a Delaunay polyhedron of perfection rank $r$ and degeneracy degree $d\geq 1$ then $$Gr\left({\mathcal C}_{r+1,d}(D), {\mathcal C}_{r+2,d-1}(D)\right)$$ is connected.
Let us take two Delaunay polyhedra $D_a$ and $D'_a$ in $D$ of perfection rank $r+1$. Let us take two Delaunay simplex sets $S$, $S'$ contained in $D_a$, $D'_a$. By using Theorem \[ConnectivitySimplices\] we can find a chain of simplices $(S_i)_{0\leq i\leq m}$ with $S_i\subset D$, $S_0=S$ and $S_m=S'$. Denote by $R_i=S_i\cup S_{i+1}$ the repartitioning set. Write $e_{-1}$, $e_{m}$ to be the extreme rays in $\operatorname{Hyp}(S_0,D)$, $\operatorname{Hyp}(S_m,D)$ corresponding to $D_a$, $D'_a$. For each Delaunay simplex set $S_i$ we consider the cone $\operatorname{Hyp}(S_i,D)$. The extreme rays correspond to Delaunay polyhedra of rank $r+1$. If a Delaunay polyhedron $D'\subset D$ has degeneracy degree $d$ then necessarily $L(D')=L(D)$. We define the restricted trace function to be $$\phi(f) = Tr\left(\left. \operatorname{Quad}(f)\right\vert_{L(D)}\right).$$ A function $f\in Erdahl_{supp}(D)$ has $Z(f)\cap D$ of degeneracy degree $d$ if and only if $\phi(f)=0$. The hyperplane $\phi(f)=0$ determine a face $F_i$ of the cone $\operatorname{Hyp}(S_i,D)$. The intersection is $$\operatorname{Hyp}(S_i,D) \cap \operatorname{Hyp}(S_{i+1},D) = \operatorname{Hyp}(R_i, D).$$ Thus we can find a ray $e_i$ in $\operatorname{Hyp}(R_i,D)$, which is not contained in $F_i$. Since $\operatorname{Hyp}(S_i,D)$ is polyhedral, by Lemma \[TheCriticalLemma\] there exists a path from $e_{i-1}$ to $e_i$ in $\operatorname{Hyp}(S_i,D)$ that avoids the face $F_i$. So, by putting all the paths together, we got the required connectivity result.
\[DiamondOnErdahl\] If $D_1$ and $D_3$ are two Delaunay polyhedra of perfection rank $r$ and $r+2$ with $D_3\subset D_1$ then there exist exactly two Delaunay polyhedra $D_{2,1}$ and $D_{2,2}$ with $$D_3 \subset D_{2,i}\subset D_1$$ with $\operatorname{rankperf}(D_{2,i})=r+1$.
Since $D_3$ is a Delaunay polyhedron there exist a Delaunay simplex set $S\subset D_3$. The Delaunay polyhedra $D_3$, $D_1$ correspond to faces $F_3$, $F_1$ of dimension $r+2$, $r$ in the cone $\operatorname{Hyp}(S, {\ensuremath{\mathbb{Z}}}^n)$. It is well known from polytope theory [@zieglerLectureOnPolytopes Theorem 2.7.(iii)] that there are exactly two faces $F_{2,1}$, $F_{2,2}$ containing $F_1$ and contained in $F_3$. Those gives the corresponding Delaunay polyhedron.
By using this theorem, we are able to compute inductively the Delaunay polyhedra in ${\ensuremath{\mathbb{Z}}}^n$. The property with the degeneracy degree ensures that we are able to effectively reduce the complexity of the computation at each step and thus we are reduced in the end to computation with Delaunay polyhedra of degeneracy $0$, i.e. polytopes for which polytopal methods exist.
Algorithms {#SectionAlgo}
==========
In [@symsurvey] a general survey of methods for computing dual description of highly symmetric polytopes with many facets are presented. Among the method presented there, we want to adapt the Recursive Adjacency Decomposition Method to our situation, i.e. to a case with an infinite group and an infinity of defining inequalities.
Computing $\operatorname{Aut}(D)$ {#AlgoComputAutD}
---------------------------------
In this subsection we explain the techniques needed to compute $\operatorname{Aut}(D)$, compute $\operatorname{Stab}_{D'}(\operatorname{Aut}(D))$ and split orbits. In the decomposition of Theorem \[AutomAut\_D\] the only component that is not clear is $\operatorname{Aut}(D_1)$, i.e. the computation of the automorphism group of a Delaunay polytope. For that purpose the methods of [@ComplexityVoronoiDSV] can be used. The one that we are using is the method of isometry groups.
Let $D$ be a Delaunay polyhedron. Take a lattice $L'$ with $L'\oplus_{{\ensuremath{\mathbb{Z}}}} L(D) = {\ensuremath{\mathbb{Z}}}^n$. Take a basis $w_1, \dots, w_r$ of $L'$. Denote by $v_1, \dots, v_m$ the expression of the vertices of $P_{L'}(D)$ in the basis $(w_i)$.\
We define the matrix $Q$ by $$Q = \sum_{i=1}^m \left(\begin{array}{c}
1\\
v_i
\end{array}\right) (1, v_i^t).$$ From then we define the distance function $f_D: {\ensuremath{\mathbb{Z}}}^n\times {\ensuremath{\mathbb{Z}}}^n \mapsto {\ensuremath{\mathbb{R}}}$ by
- $f_D(x, x') = \phi(x) Q^{-1} \phi(x')^T$
- with $\phi(x) = (1, u_1, \dots, u_r)$ if $$x = u_1 w_1 + \dots + u_r w_r + z \mbox{~and~} z\in L(D).$$
The construction of the matrix $Q$ and its inverse above is relatively standard. We used it first in [@PerfectDim8] and further work on this are done in [@symsurvey; @GroupPolytopeLMS].
The interest of this construction is that it allows to compute automorphism groups.
\[ScalProdPreservation\] Let $D$ be a Delaunay polyhedron. It holds:
\(i) If $u\in \operatorname{Aut}(D)$ then we have $f_D(u(x), u(y) ) = f_D(x,y)$ for $x,y\in D$.
\(ii) If $L'$ is a sublattice such that $L'\oplus_{{\ensuremath{\mathbb{Z}}}} L(D) = {\ensuremath{\mathbb{Z}}}^n$ and $u$ is a permutation of $P_{L'}(D)$ such that $f_D(u(x), u(y) ) = f_D(x,y)$ for $x,y\in D$ then $u$ is induced by an affine rational transformation of $L'\otimes {\ensuremath{\mathbb{R}}}$.
\(i) By its construction if $z\in L(D)$ then we have $f_D(x + z,y) = f_D(x,y)$. Thus if $u\in \operatorname{Aff}(D)$ then $u$ preserve $f_D$. On the other hand if $u\in \operatorname{Aut}(P_{L'}(D))$ then we can see by summation that $u$ preserve $f_D$. The proof is available for example in [@PerfectDim8; @symsurvey].
\(ii) The reverse implication is also available from [@PerfectDim8; @symsurvey].
Let us denote by $\operatorname{Aut}_{{\ensuremath{\mathbb{Q}}}}(P(D))$ the group of rational transformations preserving $P(D)$. By Theorem \[ScalProdPreservation\].(ii) we have $\operatorname{Aut}(P(D)) = \operatorname{AGL}_{r}({\ensuremath{\mathbb{Z}}}) \cap \operatorname{Aut}_{{\ensuremath{\mathbb{Q}}}}(P(D))$ with $r$ the dimension of $L'$. The computation of $\operatorname{Aut}_{{\ensuremath{\mathbb{Q}}}}(P(D))$ is done efficiently by using known partition backtracking software such as [@nauty]; see [@symsurvey; @GroupPolytopeLMS] for more details. In the cases considered in this paper the number of vertices is quite small and this computation is very easy.
A Delaunay polytope is called [*generating*]{} if difference between its vertices generate ${\ensuremath{\mathbb{Z}}}^n$. If a Delaunay polytope is non-generating, then it is actually a Delaunay polytopes for more than one lattice. If $P(D)$ is generating then we have $\operatorname{Aut}(P(D)) = \operatorname{Aut}_{{\ensuremath{\mathbb{Q}}}}(P(D))$ and we are done. Otherwise, we can apply some of the strategies listed in [@GroupPolytopeLMS Section 3.1]. Here, the situation is particularly simple and the simplest strategy of iterating over the group elements and keeping the integral ones works very well. Also note that the above methods with only slight modifications work for testing equivalence of Delaunay polyhedra.
Computing stabilizers
---------------------
We now give methods for computing stabilizers of Delaunay polyhedra, more precisely of the transformations preserving two polyhedra $D\subset D'$ which occurs in our computations.
Let us select a lattice $L'$ such that $L'\oplus_{{\ensuremath{\mathbb{Z}}}} L(D) = {\ensuremath{\mathbb{Z}}}^n$. Denote by $G_1$ the group occurring in Theorem \[InclusionAff\].
Let us define the following function on $P_{L'}(D)$ by $$f_{D,D'}(x, y) = \left( f_D(x,y), f_{D'}(x,y) \right).$$ By Theorem \[ScalProdPreservation\] the elements of the group $G_1$ must preserve the function $f_{D,D'}$. Thus we may use the partition backtrack algorithm of Section \[AlgoComputAutD\] to get the group $\operatorname{Aut}(f_{D,D'})$ of permutations preserving $f_{D,D'}$.
Then we obtain the group $G_1$ by keeping only the elements that are in $\operatorname{AGL}_n({\ensuremath{\mathbb{Z}}})$ and preserve $D'$. This is possible since the group $\operatorname{Aut}(f_{D,D'})$ is finite and of moderate size in most cases. All the algorithms above have equivalents for testing equivalence and of course what has been done for pairs $D\subset D'$ of Delaunay polyhedra can be extended to triples $D\subset D'\subset D''$.
Splitting orbits
----------------
Suppose that we have an orbit $Gx$ of an element $x$ under a group $G$. For a subgroup $H\subset G$ we wish to decompose $G x$ into orbits $H x_i$. Such an orbit splitting decomposition $$G x = \cup_{i=1}^m H x_i \mbox{~with~}x_i=g_i x$$ is equivalent to a double coset decomposition $$G = \cup_{i=1}^m H g_i \operatorname{Stab}_G(x).$$
In the case of interest to us we have $G=\operatorname{Aut}(D)$, $H=\operatorname{Aut}(D)\cap \operatorname{Aut}(D')$ for $D$, $D'$ Delaunay polyhedra with $D\subset D'$ and $x$ a Delaunay polyhedron included in $D$. Since a priori $\operatorname{Aut}(D)$ is infinite we cannot apply standard tools from computer algebra software such as [GAP]{} [@gap]. By the finiteness result Theorem \[UniqueDecomposition\].(ii) we can find a coset decomposition $$G = \cup_{i=1}^m H g_i.$$ However, it is not a double coset decomposition, i.e. we can have $H g_i \not= H g_j$ but still have $Hg_i x = Hg_j x$. Therefore, we need to eliminate duplicate in order to do the orbit splitting.
The flipping algorithm
----------------------
Suppose that $D_1\subset D_2\subset D_3$ are Delaunay polyhedra having $$\operatorname{rankperf}\, D_1=1+\operatorname{rankperf}\, D_2=2+\operatorname{rankperf}\, D_3.$$ By Lemma \[DiamondOnErdahl\] we know that there exist a unique Delaunay polyhedron $D'_2$ having $D_1\subset D'_2\subset D_3$ and $D_2\not= D'_2$.
We can find functions $f_i\in Erdahl(n)$ such that $Z(f_i)=D_i$. We can also assume that $C_{D_i}= {\ensuremath{\mathbb{R}}}f_i\oplus C_{D_{i+1}}$ for $i=1$, $2$. We need to find $f'_2\in Erdahl(n)$ such that $Z(f'_2)=D'_2$.
If $L(D_3)=0$ then $D_3$ is a polytope, i.e. it has a finite number of vertices and the algorithm is called the [*gift wrapping*]{} procedure [@symsurvey; @PerfectDim8; @CR]. If $L(D_3)\not= 0$ we have to modify the algorithm in order to take care of the fact that we have an infinity of vertices by writing an iterative algorithm. This is quite similar to the flipping in the Voronoi algorithm [@bookschurmann].
\[Flipping\_algorithm\] ${\mathcal V} \leftarrow \emptyset$\
$f'_2\leftarrow f'_2 + f_3$
The non-negativity test for $f'_2$ on ${\ensuremath{\mathbb{Z}}}^n$ is done by solving a closest vector problem. The non-negativity test on $D_3$ is done by decomposing it into $\{v_1 + L(D_3)\} \cup \dots \cup \{v_m + L(D_3) \}$. The non-negativity is tested by $m$ closest vector problems. The final operation on $f'_2$ is done to ensure that $f'_2(x)>0$ if $x\notin Z(f_3)$.
The recursive adjacency decomposition method
--------------------------------------------
Given a Delaunay polyhedron $D$ of perfection rank $r$ the algorithm of this section will give the orbits of Delaunay polyhedra $D'\subset D$ of perfection rank $r+1$. If $\operatorname{degrk}(D)=0$ then the computation of the orbits of Delaunay polyhedra can be achieved by Theorem \[BasicResultErdahlCone\].(iii).
If $\operatorname{degrk}(D)>0$ we have to proceed differently. By Theorem \[ErdahlConnectivityResult\] we can limit ourselves to Delaunay polyhedra with $\operatorname{degrk}(D)\leq \operatorname{degrk}(D)-1$. The algorithm takes one initial Delaunay polyhedron of perfection rank $r+1$ and computes the adjacent Delaunay polyhedron of perfection rank $r+1$. If an obtained Delaunay polyhedron is not equivalent to an existing one then we insert it into the list. We iterate until all orbits have been treated. The computation of the adjacent Delaunay polyhedra adjacent to a Delaunay polyhedron $D'$ requires the computation of orbits the Delaunay polyhedra contained in $D'$. Thus we have a recursive call to the algorithm. Fortunately the degeneracy degree diminish by at least $1$ so there is no infinite recursion.
The mapping from ${{\mathcal F}}_1$ to ${{\mathcal F}}_2$ is done using the orbit splitting procedure. Finding $D''$ from $D_2$, $D'$ and $D$ is done using the flipping procedure.
There is a degree of choice in the initial Delaunay polyhedron $D_{init}$. The standard choice is if $D=P(D) + L(D)$ with $L(D)={\ensuremath{\mathbb{Z}}}v_1 + \dots + {\ensuremath{\mathbb{Z}}}v_{\operatorname{degrk}(D)}$ to take $D_{init}$ a Delaunay polyhedron of the form $$D_{init} = P(D) + \{0, v_1\} + {\ensuremath{\mathbb{Z}}}v_2 + \dots + {\ensuremath{\mathbb{Z}}}v_{\operatorname{degrk}(D)} .$$ The schematic of the algorithm is shown in Algorithm \[Enumeration\_inequivalent\].
\[Enumeration\_inequivalent\]
Perfect Delaunay polytopes in dimension $7$ {#DimSeven_perfectDelaunay}
===========================================
In the enumeration of inhomogeneous perfect form in dimension $7$ we need to describe the Delaunay polytopes that will occur. The list of perfect Delaunay polyhedra in dimension $7$ is thus
1. $\{0,1\}\times {\ensuremath{\mathbb{Z}}}^6$.
2. $2_{21} \times {\ensuremath{\mathbb{Z}}}$ with $2_{21}$ the Schläfli polytope
3. $3_{21}$ the Gosset polytope [@DL].
4. $ER_7$ the polytopes discovered by Erdahl and Rybnikov [@35tope; @InfiniteER2002].
The geometry of the Schläfli and Gosset polytope are described in more details in [@DL; @PerfectDelaunayLowDim].
An [*affine basis*]{} of a $n$-dimensional Delaunay polytope $D$ is a family of $n+1$ vertices $v_0$, …, $v_n$ such that for any vertex $v$ of $D$ there exist $\lambda_i\in {\ensuremath{\mathbb{Z}}}$ such that $v=\sum_{i=0}^n \lambda_i v_i$. The perfect Delaunay polytopes of dimension $7$ have an affine basis but it is possible that in higher dimension there are perfect Delaunay polytopes without affine basis. It is known that in dimension at least $12$, there are Delaunay polytopes with no affine basis [@RankComput]. Also the perfect Delaunay polytopes of dimension $7$ are generating. Note that in [@SecInfinite] we found some non-generating perfect Delaunay polytopes for $n\geq 13$. We have $\operatorname{rankperf}(\{0,1\}^n)=n$ (see for example [@DelaunaySix; @DL]).
In terms of computation, the overwhelming majority of the time is spent computing the rank $2$ faces of $\{0,1\}\times {\ensuremath{\mathbb{Z}}}^6$. By the recursive approach chosen, the method requires the computation of the facets of $Erdahl_{supp}^*(\{0,1\}^7)$ and so by Theorem \[CUTn\_and\_ErdahlHn\] of the facets of $\operatorname{CUTP}_8$. We actually computed the list of orbits of facets of $\operatorname{CUTP}_8$ (and some other graph cut polytopes) in [@CUT8_facet]. In dimension $8$ the partial enumeration algorithm of [@ANewAlgorithm] found $27$ perfect Delaunay polytopes and it is likely that the list is complete. But to prove its completeness by using the method of this work would require the determination of all facets of $\operatorname{CUTP}_9$ and this is very hard [@CR]. In [@ANewAlgorithm] a partial enumeration of perfect Delaunay polytopes was done with only Delaunay polyhedra with $L(D)=0$ being considered. The two perfect Delaunay polytopes of dimension $7$ were determined in this work and our enumeration proves that the list is complete.
The implementation is available from [@Polyhedral] and uses the [GAP]{} computer algebra system [@gap].
Classification of Delaunay simplices in dimension $7$ {#DimSeven_simplices}
=====================================================
Formula gives $187$ as an upper bound on the volume of Delaunay simplex sets. With this upper bound we can devise an algorithm for enumeration of Delaunay simplex sets, which will unfortunately prove inefficient:
\[ExpansionAlgorithm\] Suppose we have a list of types of Delaunay simplex sets in dimension $n-1$. If $S = \{v_0, \dots, v_{n-1}, v_n\}$ is a Delaunay simplex set of dimension $n$ then $\{v_0, \dots, v_{n-1}\}$ is a $n-1$ dimensional Delaunay simplex sets of the lattice $v_1 - v_0$, …, $v_{n-1} - v_0$.
If we have a $n-1$ dimensional Delaunay simplex set $v_0=0$, $v_i\in {\ensuremath{\mathbb{Z}}}^{n-1}$ for $1\leq i\leq n-1$ and is of volume $v$. We write the $n$-dimensional simplex as $v'_0 = (v_0, 0)$, $v'_1 = (v_1, 0)$, …, $v'_{n-1} = (v_{n-1}, 0)$ and $v'_n = (x_1, \dots, x_{n-1}, x_n)$ with $x_n > 0$. The volume of the simplex defined by $(v'_i)$ is $v x_n$.
For a fixed $x_n$ the number of possibilities for $(x_1, \dots, x_{n-1})$ is $x_n^{n-1} / v$.
The Delaunay simplices of dimension $6$ were classified in [@Ba] and so we could apply the algorithm of Remark \[ExpansionAlgorithm\]. Unfortunately the number of possibilities to be applied is very large, of the order of $187^6$, on which we have to apply Algorithm \[TestRealizability\].
Therefore, we need a different method:
\[Simplices\_NotInPerfect\] Let $S$ be a Delaunay simplex set that is not contained in any perfect Delaunay polyhedron different from $\{0,1\}\times {\ensuremath{\mathbb{Z}}}^{n-1}$. The possibilities are:
1. For $n\leq 4$ the Delaunay simplices of volume $1$.
2. For $n=5$ the Delaunay simplex of volume $1$ or $2$.
3. For $n=6$ or $7$ there are no possibilities.
Let us take a Delaunay simplex set $S = \{v_0, \dots, v_n\}$. We can assume that $v_0$ is located at the origin by using translation if necessary. For each $1\leq i\leq n$ let us define $\ell_i$ the linear form on ${\ensuremath{\mathbb{R}}}^n$ such that $\ell_i(v_i)=1$ and $\ell_i(v_j)=0$ for $i\not=i$. Any Delaunay polyhedron $D$ isomorphic to $\{0,1\}\times {\ensuremath{\mathbb{Z}}}^{n-1}$ and such that $S \subset D$ corresponds to a linear form $\ell$ on ${\ensuremath{\mathbb{R}}}^n$ such that $\ell(v_i)\in \{0,1\}$ and $$D=\{x\in {\ensuremath{\mathbb{Z}}}^n \mbox{~$\vert$~} \ell(x) = 0\mbox{~or~} 1\}.$$ The linear form $\ell$ is then called [*admissible*]{} and the corresponding quadratic function is $q_{\ell}(x)= \ell(x)( \ell(x) - 1)$.
Let us denote by $S\subset \{1,\dots,n\}$ the set of points $i$ such that $\ell(x)=1$. Clearly, one can write $$\ell = \sum_{i\in S} \ell_i.$$ A function $\ell$ is admissible if and only if $\ell$ is integral valued on ${\ensuremath{\mathbb{Z}}}^n$. If it is not integral valued then there exists a $v\in {\ensuremath{\mathbb{Z}}}^n$ such that $0 < \ell(v) < 1$ which implies that $q_{\ell}(v) < 0$ which is not allowed. If it is integral valued then $D=Z(q_{\ell})$ is equivalent to $\{0,1\}\times {\ensuremath{\mathbb{Z}}}^{n-1}$. In the following, for a set $S\subset \{1, \dots, n\}$ we write $$\ell_S = \sum_{i\in S} \ell_i \mbox{~and~} v_S = \sum_{i\in S} v_i.$$ Let us define $${\mathcal S} = \left\{ S\subset \{1, \dots, n\} \mbox{~s.t.~} \ell_S\mbox{~is~integral~valued}\right\} .$$ Let us denote by ${\ensuremath{\mathbb{Z}}}{\mathcal S}$ the ${\ensuremath{\mathbb{Z}}}$-span of the elements $v_S$ for $S\in {\mathcal S}$. This defines a lattice ${\mathcal L}$ of ${\ensuremath{\mathbb{Z}}}^n$. Since $S$ is contained only in Delaunay polyhedra isomorphic to $\{0,1\}\times {\ensuremath{\mathbb{Z}}}^{n-1}$ the set of the function $q_{\ell_S}$ is full-dimensional. This implies that $\left\vert{\mathcal S}\right\vert \geq n(n+1)/2$ and that the lattice ${\mathcal L}$ is actually full dimensional. Denote by $h$ its index.
The set of the function $\ell_S$ is also full-dimensional in $({\ensuremath{\mathbb{Z}}}^n)^*$. Its index is $$h/\operatorname{vol}(S)\geq 1.$$ We are interested in the point sets of the form $\{0,1\}^n\cap L$ with $L$ an affine subspace of ${\ensuremath{\mathbb{Z}}}^n$. By direct enumeration we obtain the full list of $3363$ orbits of such points for $n=7$. By selecting the point sets whose cone of functions $q_{\ell_S}$ is full-dimensional we get an upper bound of $3$ on the index $h$ and so an upper bound of $3$ on the possible volumes of such simplices.
With volume at most $3$ we can apply the algorithm implied by Remark \[ExpansionAlgorithm\]. Each facet of such a Delaunay simplex is also a Delaunay simplex of one dimension lower. Therefore, we can use previous enumeration result to get a list of $796$ possible candidates of $7$-dimensional Delaunay simplices. We then use Algorithm \[TestRealizability\] for checking which ones of them are indeed Delaunay simplices. This gives $6$ cases (the ones of Table \[TableFundamentalSimplices\] of volume at most $3$). Each one of them is also contained in a Delaunay polytope $ER_7$ and so there is no such Delaunay simplices in dimension $7$.
Dimension $n\leq 6$ follows from known results.
[**Proof of Theorem \[EnumerationSimplices\]**]{}: If $S$ is a Delaunay simplex set then $\operatorname{Hyp}(S,{\ensuremath{\mathbb{Z}}}^7)$ is a full-dimensional polyhedral cone, i.e. defined by a finite number of inequalities and having a finite number of extreme rays. Any such extreme ray corresponds to a perfect Delaunay polyhedron $D$. We have $|S|=8$ and $S$ defines a face of the cone $Erdahl_{supp}^*(D)$.
By Lemma \[Simplices\_NotInPerfect\], $S$ has to be contained in a Delaunay polyhedron of type $3_{21}$, $ER_7$ or $2_{21}\times {\ensuremath{\mathbb{Z}}}$.
The perfect Delaunay polyhedron $ER_7$ has $35$ vertices which matches the lower bound given by Proposition \[LowerBoundNrVertices\]. As a consequence any $8$-element subset of $ER_7$ defines a face of $Erdahl_{supp}^*(ER_7)$. The automorphism group of $ER_7$ has size $1440$ and by using it one can get easily the $9434$ orbits of $8$-element subsets of $ER_7$. Actually all $11$ types of simplices occur this way.
The Gosset polytope $3_{21}$ has $56$ vertices and the automorphism group is equal to the Weyl group of the root lattice $\mathsf{E}_7$. We found $521$ orbits of $8$-element sets in $3_{21}$, $474$ of them correspond to faces of $Erdahl_{supp}^*(3_{21})$.
For the perfect Delaunay polyhedron $2_{21}\times {\ensuremath{\mathbb{Z}}}$ we have to proceed differently since the number of points to be considered is infinite. We have to enumerate the possible $8$-point subsets of $2_{21}\times {\ensuremath{\mathbb{Z}}}$ of volume at most $187$ (Formula ) up to the action of $\operatorname{Aut}(2_{21}\times {\ensuremath{\mathbb{Z}}})$. The $8$ points are expressed in the form $v_i=(w_i, h_i)$ with $w_i\in 2_{21}$ and $h_i\in {\ensuremath{\mathbb{Z}}}$. The set of points $(w_i)_{1\leq i\leq 8}$ must define a $6$-dimensional affine space. Thus $7$ of them, say $(w_i)_{1\leq i\leq 7}$, must be sufficient to define a $6$-dimensional Delaunay simplex set $S_{Sch}$. An exhaustive enumeration on the $27$ vertices of $2_{21}$ gives $31$ types up to isomorphism. The volume $\operatorname{vol}(S_{Sch})$ can be $1$, $2$ or $3$. If the volume is $1$, then we can use an element of $\operatorname{Aff}(2_{21}\times {\ensuremath{\mathbb{Z}}})$ and obtain $h_i=0$ for $1\leq i\leq 7$. For higher volumes, the situation is more complicate but by using $\operatorname{Aff}(2_{21}\times {\ensuremath{\mathbb{Z}}})$ and linear algebra we we can reduce to $\operatorname{vol}(S_{Sch})$ possibilities, i.e. $2$ or $3$. For the last point $(v_8, h_8)$ we have $27$ possibilities for $v_8$ and a finite number for $h_8$ due to the upper bound of $187$. We then apply Algorithm \[TestRealizability\] to test realizability of the finite list of possible cases. This gives us the $11$ possible simplices.
Acknowledgments {#acknowledgments .unnumbered}
===============
The author thanks Viacheslav Grishukhin and Achill Schürmann for useful discussion on this work.
[^1]: The author gratefully acknowledge support from the Alexander von Humboldt foundation
| {
"pile_set_name": "ArXiv"
} |
---
author:
- Yoritaka Iwata
title: Phase diagram of dilute cosmic matter
---
Introduction
============
Nuclear pasta, whose existence and role were suggested in association with the collapse driven supernova explosions [@watanabe; @sowa], is of great importance with respect to the astrophysical synthesis of chemical elements. Indeed nuclear pasta is known to change the transparency rate of neutrinos [@ravenhall; @sato], which play a role of trigger in supernova explosions giving rise to a wide variety of chemical elements. Therefore the existence of nuclear pasta at low densities is expected to have a great impact on clarifying the historical evolution of chemical elements and also on revising the low-density and low-temperature part of the nuclear phase diagram [@sonoda; @lamb]. In particular, here we are concerned with its existence in the outermost part of protoneutron stars (for the review of supernova mechanisms, see [@bethe]).
In this paper, based on three-dimensional time-dependent density functional calculations, a dynamical formation of nuclear pasta in the low-density and low-temperature situations is demonstrated by 3-nucleus simultaneous heavy-ion collision. This calculation simulates the situation outside of the known low-density existence limit of nuclear pasta phase. In order to examine the validity of this calculation, the possibility of multi-nucleus simultaneous collisions is studied by a systematic three-dimensional Monte-Carlo calculation, and the mean free path of a nucleus at a given density and temperature is obtained. Note that for this kind of nuclear pasta formation, the presence or absence of the fast charge equilibration process [@iwata] determines the upper boundary of nuclear pasta phase with respect to the temperature [@iwata2].
Heavy-ion collisions in cosmic matter
=====================================
Dilute cosmic matter
--------------------
Cosmic nuclear matter, which consists of both neutrons and protons, is studied. Here we are concerned with the dilute cosmic matter whose density is less than 10 % of the standard nuclear density $\rho_0$. Nuclear matter of this kind is expected to exist in the outermost part of the protoneutron star. In the following the dilute cosmic matter means the matter whose density is less than 10 % of the standard nuclear density.
Although the existence of nuclear pasta was suggested in more dense situations, its existence on the outermost part of protoneutron star has never been understood. Its existence, however, is shown in this paper.
Multiple nucleus collisions
---------------------------
We are concerned with heavy-ion collisions appearing in the dilute cosmic matter. Here the mean free path of a nucleus is important, and multi-nucleus simultaneous collisions might play a role. Indeed, rather dense situations, which are not achieved on the earth, appear in the universe, so that the presence of multi-nucleus simultaneous collisions is expected. For reference, the density of targets in accelerators is roughly estimated by $$\rho_{\rm Lab} = 10^{-12} \rho_0$$ where the radii of nuclei and atoms are assumed to be 10$^{-14}$ m and 10$^{-10}$ m, respectively. In the following the density is denoted by $\rho$, and the terminology of multi-nucleus simultaneous collision exactly means simultaneous collisions between more than two nuclei.
Calculations of 3-nucleus simultaneous collisions
=================================================
A three-nucleus simultaneous collision is shown based on three-dimensional time-dependent density functional calculations employing the Skyrme force parameter set SLy6 [@Chabanat-Bonche]. The calculation is carried out in a spatial box $64 \times 32 \times 64$ fm$^3$ with a spatial grid spacing of 1.0 fm, where periodic boundary condition is applied. Three $^{78}$Zn are initially located at (8,0,-8), (-8,0,-8) and (0,0,8), respectively (cf. the first frame in Fig. \[fig1\]). The two nuclei initially located at (8,0,-8) and (-8,0,-8) are not given velocities in the center-of-mass frame, while the nucleus initially located at (0,0,8) has the initial velocity along the z-axis. Accordingly, the total system consists of 90 protons and 144 neutrons, and the total incident energy is set to 100 MeV in the center-of-mass frame.
Figure \[fig1\] demonstrates a 3-nucleus simultaneous collision resulting in nuclear pasta. The temperature and density of this calculation are 0.5 MeV (see following discussion leading to Eq. (\[sb\])) and 0.039 $\rho_0$, respectively. This situation corresponds to the state with a density lower than the known low-density existence limit of nuclear pasta phase [@sonoda]. If we consider collisions with a higher energy ($T =$ 7.5 MeV), nuclear pasta cannot be formed. It corresponds to the energy above the upper-limit energy of fast charge equilibration [@iwata]. For comparison, 2-nucleus collisions require more dense situations to form a nuclear pasta. A calculation with exactly the same setting except for the force parameter set (SKI3 [@reinhard]) results in the same conclusion. Consequently the low-density existence limit of nuclear pasta phase is suggested to be extended by taking into account multi-nucleus simultaneous collision dynamics. It is important to note that the boundary condition is not essential. Therefore Fig. \[fig1\] should not be interpreted as the formation of infinitely connected matter, but simulates the favorable situation of very long finite spaghetti-like structure. Pasta formation by 3-nucleus simultaneous collisions has been calculated for several different settings of the initial condition [@iwata2].
Monte Carlo calculation
=======================
Although multi-nucleus collisions are suggested to enhance nuclear pasta formation, it is worthless if such a situation is quite rare in the cosmic matter.
Let us confirm the validity of multi-nucleus collisions in the cosmic matter. The numbers of simultaneous collisions are calculated utilizing the Monte Carlo method. A cubic space with side length $L$ fm is given. Spheres (corresponding to nucleus) with the same radius $r$ fm are introduced in the cubic space, where those positions are chosen randomly (The left panel of Fig. \[fig2\]). Accounting for the effect from the Pauli principle (Fermionic feature) and the Coulomb force, a sphere overlapping with the other sphere is discarded from the beginning. Together with periodic boundary condition, thermal equilibrium of nuclear matter is simulated. The density of a sphere is fixed to $\rho_0$, so that the density of matter is given by adjusting the values of $L$ ($L$ is fixed in actual calculations), $r$ and the number of spheres. Another sphere (nucleus) with exactly the same feature (radius: $r$ fm, density: $\rho_0$), which corresponds to the projectile, is coming from the outside of the cube. The temperature of the system is determined by the relative velocities between spheres. Here it is given by the relative velocity of the projectile to the cubic space. The projectile experiences collisions with spheres, when it goes though this periodic space.
Let the typical time interval of low-energy heavy-ion collisions be 1000 fm/c = $33.3 \times 10^{-22}$ s. Collisions are defined to be simultaneous, if they occur within this time interval. The statistical values for the number of simultaneous collisions is obtained by $10^4$ trial calculations. Let $A$ denote the nuclear mass number. Three cases with different radii of the sphere ($A=$ 50, 100, 200): $$r = 1.2 \times A^{1/3} ~{\rm fm},$$ are calculated for several densities and temperatures. In particular, with respect to low-energy heavy-ion collisions, we consider three different temperatures $T = 1, 5, 10 ~{\rm MeV}$. According to the Stefan-Boltzmann relation, the temperature of matter in heavy-ion reactions is estimated by the relative kinetic energy, $$\label{sb}
\sigma T^4 = E^*,$$ where $T$ and $E^*$ mean the temperature and the incident energy, respectively.
\(a) $A$ = 50\
$T~ \backslash ~ \rho$ 0.02 $\rho_0$ 0.10 $\rho_0$ 0.20 $\rho_0$
------------------------ ------------------- ------------------- ----------------- -- -- --
10 MeV 19.5 4.6 2.7
5 MeV 30.9 6.9 4.1
1 MeV 54.4 14.2 8.2
: Mean free path \[fm\] of a nucleus in nuclear matter, which consists of nuclei with mass number $A$.
\
(b) $A$ = 100\
$T~ \backslash ~ \rho$ 0.02 $\rho_0$ 0.10 $\rho_0$ 0.20 $\rho_0$
------------------------ ------------------- ------------------- ----------------- -- -- --
10 MeV 23.6 5.9 3.4
5 MeV 40.2 10.4 5.6
1 MeV 70.1 20.1 10.8
: Mean free path \[fm\] of a nucleus in nuclear matter, which consists of nuclei with mass number $A$.
\
(c) $A$ = 200\
$T~ \backslash ~ \rho$ 0.02 $\rho_0$ 0.10 $\rho_0$ 0.20 $\rho_0$
------------------------ ------------------- ------------------- ----------------- -- -- --
10 MeV 31.2 7.4 4.3
5 MeV 42.5 11.1 6.4
1 MeV 89.7 34.9 16.1
: Mean free path \[fm\] of a nucleus in nuclear matter, which consists of nuclei with mass number $A$.
\[table1\]
Mean free path of nucleus in the matter
---------------------------------------
Table \[table1\] shows the mean free path of a nucleus at given temperature and density. The mean free path tends to be larger for matter including heavier nuclei. It seems to be reasonable; for a fixed density and volume of the cubic space, there are more nuclei in lighter case. The mean free path becomes smaller for dense and high temperature situations. Comparing two cases in $\rho = 0.02 \rho_0$ and $T= 1$ MeV, almost 35 fm difference of mean free path is noticed depending only on the size of the nuclei. Note that the cases with $T = 1$ MeV roughly correspond to heavy-ion reactions at the incident energy slightly higher than the Coulomb barrier. This table implies that multi-nucleus simultaneous collisions commonly appear in the dilute cosmic matter satisfying $\rho > 0.01 \rho_0$.
Transition to the situation dominated by multi-nucleus collisions
-----------------------------------------------------------------
The right panel of Fig. \[fig2\] shows the statistics of simultaneous collisions for $T = 1$ MeV and A = 100. It demonstrates the density-dependent transition from dominance of 2-nucleus collisions to that of multi-nucleus collisions. For the lowest density case with $\rho = 0.02 \rho_0$, the situation is dominated by 2-nucleus collisions. However, depending on the increase of the density, 2-nucleus collisions lose their dominance below $\rho = 0.10 \rho_0$, and multi-nucleus collisions becomes dominant instead. It is also seen that three-nucleus simultaneous collisions are not expected (their expectation values are negligibly small), if we are only interested in collisions in the laboratory ($\rho = \rho_{\rm Lab}$).
$T~ \backslash ~ A$ 50 100 200
----------------------- --------------- --------------- ---------------- -- -- --
10 MeV 1.5 $\to$ 2.0 1.5 $\to$ 2.0 1.5 $\to$ 2.0
5 MeV 1.5 $\to$ 2.0 1.5 $\to$ 2.0 1.5 $\to$ 2.0
1 MeV 2.0 $\to$ 4.0 2.0 $\to$ 4.0 4.0 $\to$ 10.0
: Transition density \[$10^{-2} \rho_0$\] from dominance of 2-nucleus collisions to that of multi-nucleus collisions. The dominance of 2-nucleus collisions is defined to be true, if 2-nucleus collisions are the most probable among other finite-nucleus collisions (not taking into account collisionless cases). The density dominated by the 2-nucleus collisions (the left side of arrow) and that dominated by the multi-nucleus collisions (the right side of arrow) are shown in each case.
\[table2\]
Table \[table2\] summarizes the transition from the situation dominated by 2-nucleus collisions to that dominated by multi-nucleus collisions. The transition density is found to be of the order of $10^{-2} \rho_0$, where its temperature dependence is small. In addition simultaneous collisions between many nuclei are expected for the nuclear matter with 1 % of the standard density. Indeed, for $\rho = \rho_0$ and $T= 1$ MeV, the most frequent numbers of collisions are 77, 59, 47 for $A =$ 50, 100, 200, respectively. Note again that this calculation is carried out by giving a typical time interval (1000 fm/c), so that the transition density becomes larger if we assume the smaller time interval. Consequently, multi-nucleus simultaneous collisions become important and are expected to play a role in the dilute cosmic matter with density higher than $0.01 \rho_0$, and the validity of the calculation shown in Fig. \[fig1\] is confirmed.
Summary
=======
Enhancement of nuclear pasta formation due to 3-nucleus simultaneous collisions has been presented based on time-dependent density functional calculations (using SLy6 and SKI3 force parameter sets). Multi-nucleus simultaneous collisions between more than two nuclei have been confirmed to occur in cosmic matter with density higher than $0.01 \rho_0$. Therefore, nuclear pasta formation is possible in the outermost part of protoneutron stars, and neutrino radiation is not as easy as expected. Eventually, the low-density existence limit of nuclear pasta phase is suggested to be lower than believed up to now.\
The author thanks to Profs. K. Iida, N. Itagaki, J. A. Maruhn and T. Otsuka. This work was supported by the Helmholtz Alliance HA216/EMMI.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'F. M. Dunlop, P. A. Ferrari, L. R. G. Fontes'
title: |
A dynamic one-dimensional interface\
interacting with a wall
---
¶[[P]{}]{} [**H**]{} &=&
[**Abstract.**]{} We study a symmetric randomly moving line interacting by exclusion with a wall. We show that the expectation of the position of the line at the origin when it starts attached to the wall satisfies the following bounds: $$c_1t^{1/4} \le\E\xi_t(0) \le c_2 t^{1/4}\log t$$ The result is obtained by comparison with a “free” process, a random line that has the same behavior but does not see the wall. The free process is isomorphic to the symmetric nearest neigbor one-dimensional simple exclusion process. The height at the origin in the interface model corresponds to the integrated flux of particles through the origin in the simple exclusion process. We compute explicitly the asymptotic variance of the flux and show that the probability that this flux exceeds $Kt^{1/4}\log
t$ is bounded above by const. $t^{2-K}$. We have also performed numerical simulations, which indicate $\E\xi_t(0)^2 \sim t^{1/2}\log t$ as $t\to\infty$.
[**Key words:**]{} Interface motion, entropic repulsion, particle flux, simple exclusion process.
[**AMS Classification:**]{} Primary: 60K35 82B 82C
Introduction
============
We consider a process $\xi_t$ on $$\X=\{\xi\in\N^\Z: |\xi(x)- \xi(x+1)|=1,\, \xi(0) \hbox{ even}\}$$ the space of trajectories of nearest neighbor random walks that stay non negative and such that at even “times” the walk visits even integers.
The generator of the process is given by $$\label{gen}
{\cal L}f(\xi) = \frac12 \sum_x \one\{\xi + \Delta
\xi(x) \,\delta_x \ge 0\}\, [f(\xi + \Delta
\xi(x)\,\delta_x) - f(\xi)]$$ where $\delta_x$ is the infinite vector having $1$ in the $x^{\rm th}$ coordinate and zero on the others. The sum $\xi+ a\delta_x$ is understood coordinatewise. The discrete Laplacian $\Delta$ is defined by $$\label{p28}
\Delta \xi(x) := \xi(x+1) - 2 \xi(x) + \xi(x-1).$$ In words we can describe the dynamics as follows. The discrete Laplacian assumes only three values, $-2$, $0$ and $2$. When the Laplacian is zero, the interface does not move. When it is $-2$ or $2$, at rate $\frac12$ it makes a jump of length $2$ in the same direction as the sign of the Laplacian. Over this motion we impose a restriction to keep the process in $\X$: the interface cannot be negative, so we simply prohibit the jumps which violate the restriction. This is the meaning of the indicator function $\one\{\xi + \Delta
\xi(x) \,\delta_x \ge 0\}$ in the generator. We can think the prohibition of becoming negative as the interaction by exclusion of the interface with a wall at $-1$. For shortness we call $\xi_t$ the *wall process*.
Our main result is the following
\[98\] Let $\xi_t$ be the process with generator and initial *flat configuration*: $$\label{99}
\xi_0(x) := x\,(\hbox {\rm mod } 2).$$ Then there exist positive constants $c_1$ and $c_2$ such that $$\label{100}
c_1t^{1/4} \le\E\xi_t(0) \le c_2 t^{1/4}\log t$$ for sufficiently large $t$.
Theorem \[98\] catches the effect of the “entropic repulsion” in a stochastically moving interface interacting with a wall by exclusion.
The line induced in $\R^2$ by joining $(x,\xi_t(x))$ to $(x+1,\xi_t(x+1))$ for all $x\in \Z$ has the same behavior as the interface between $-1$’s and $1$’s in a zero-temperature two-dimensional nearest-neighbors Ising model with a positive external field in the semiplane below the diagonal $x=y$ and with initial condition “all plus” below the diagonal and “all minus” above it. See Section \[ising\] for details.
The equilibrium statistical mechanics of this model is well known. If one considers the generator ${\cal L}$ restricted to the box $[-L,L]$ with boundary conditions $\xi_t(-L)=\xi_t(L)=0$, the invariant distribution is the uniform distribution in the set of nearest neighbors random walk trajectories starting at time $-L$ at the origin, finishing at time $L$ at the origin and being non negative for all intermediate times. Actually the uniform measure is even reversible for the process. But the uniform measure in this set corresponds to the law of a symmetric nearest neighbors random walk $X_i$ conditioned to the set $\{X_{-L}=X_L=0,\, X_i\ge 0, i\in [-L,L]\}$. Hence, the typical height of a configuration $\xi$ with the invariant law in the bulk of the box is $$\xi([rL]) \;\sim\; O(\sqrt{rL})$$ More precisely, the normalized process process $(L^{1/2}\xi([Lr]),
r\in[-1,1])$ converges as $L\to\infty$ to Brownian excursion on $[-1,1]$; see Theorem 2.6 of Kaigh (1976).
Many papers deal with the problem of entropic repulsion in Equilibrium Statistical Mechanics. The role of the entropic repulsion in the Gaussian free field was studied by Lebowitz and Maes (1987), Bolthausen, Deuschel and Zeitouni (1995), Deuschel (1996) and Deuschel and Giacomin (1999).
Entropic repulsion for Ising, SOS and related models was discussed in Bricmont, El Mellouki and Fröhlich (1986), Bricmont (1990), Holický and Zahradník (1993), Cesi and Martinelli (1996), Lebowitz, Mazel and Suhov (1996), Dinaburg and Mazel (1994) and Ferrari and Martínez (1998).
The exponent $1/4$ for dynamic entropic repulsion was predicted by Lipowsky (1985) using scaling arguments. This exponent was then found numerically by Mon, Binder, Landau (1987), Albano, Binder, Heermann, Paul (1989-1992), see Binder (1990), De Coninck, Dunlop and Menu (1993). It has also been observed in real experiments by Bartelt, Goldberg, Einstein, Williams, Heyraud, Métois (1993). Further theoretical investigations of dynamics of lines, in relation to experiments can be found in Blagojevic, Duxbury (1999).
Dynamic entropic repulsion for a line of finite extension $L$ when $t,L\to\infty$ strongly depends on the ratio $t/L^2$. The present paper deals with $L=\infty$ (analytical) or $t/L^2\to 0$ (numerical). The case $t/L^2={\cal O}(1)$ has been studied by Funaki and Olla (2001).
The exponent $1/4$ also applies to the growth of fluctuations of an initially straight interface not interacting with the wall (see below). For the Gaussian case, explicit computations were made by Abraham, Upton (1989), Abraham, Collet, De Coninck, Dunlop (1990). It was observed numerically in the two-dimensional Ising model by Stauffer, Landau (1989).
The strategy to show Theorem \[98\] is to compare the wall process $\xi_t$ with a *free* process $\zeta_t$ having the same local dynamics as $\xi_t$ but not interacting with the wall. The free process lives in $$\X_0 = \{\zeta\in \Z^\Z: |\zeta(x)-\zeta(x+1)|=1, \zeta(0)= \hbox{ even}\}$$ and its generator is $$\label{gen0}
{\cal L}_0f(\zeta)\; =\; \frac12\sum_x [f(\zeta + \Delta
\zeta(x)\,\delta_x) - f(\zeta)]$$
In the next section we prove that with flat initial condition the variance of the height at the origin for the free process behaves as $t^{1/2}$: $$\label{vfp}
\lim_{t\to\infty}t^{-1/2}\V \zeta_t(0)
\;=\;\frac{1}{\sqrt\pi}.$$
We then couple the wall process and the free process in such a way that $$\label{cou}
\zeta_t(x) \le \xi_t(x)$$ for all $x$ and $t$. The free process has enough symmetry and, properly rescaled, has uniformly bounded in time exponential moments. With these, and , we get the lower bound in .
The idea for the upperbound is to consider a family of free processes with initial condition depending on $t$: $$\zeta^{a_t}_0 (x) = \zeta_0(x) + a_t$$ (a flat interface of height $a_t$). Then we fix $a_t = c t^{1/4}\log t$, the constant $c$ to be determined later and exhibit a coupling under which $$\label{eq:c01}
\xi_s(0) \le \zeta^{a_t}_s(0)$$ for all $s\le t$ with large probability. Combined with , inequality is the key for the upperbound in . The existence of exponential moments (mentioned above) yields the moderate deviations result needed here.
The control of the fluctuations of the position at the origin of the free process is obtained by an isomorphism between the free process and the one-dimensional symmetric nearest-neighbor simple exclusion process. Under this map, $\zeta_t(0)=2J_t$, where $J_t$ is the integrated flux of particles at the origin for the exclusion process. We compute explicitly the asymptotic variance of the integrated flux for the flat initial condition in Theorem \[104\] below and obtain $$\label{vfp6}
\lim_{t\to\infty}{\V J_t \over \sqrt t}
=\frac{1}{4\sqrt{\pi}}.$$ De Masi and Ferrari (1985) proved that the asymptotic variance of the integrated flux when the initial configuration is distributed according to a product measure with density $1/2$ is given by $$\label{105a}
\lim_{t\to\infty}{\V J_t \over \sqrt t}
=\frac{1}{2\sqrt{2\pi}}.$$ which is strictly bigger than . When the initial density is $\rho$, the asymptotic variance is given by $\rho(1-\rho)\sqrt{2/\pi}$. The method to show and is based on duality and comparison with systems of independent particles and it is inspired by Arratia (1983), who used these tools to compute the variance of a tagged particle for the process starting with an (invariant) product measure. However a modification of Arratia’s proof is needed in due to the deterministic character of the initial configuration.
The study of the flux in the simple exclusion process is done in Section 2. In Section 3 we prove Theorem \[98\]. Section 4 is devoted to numerical simulations. Section 5 shows the equivalence to the dynamics of a particular zero temperature Ising model interface.
The free process and simple exclusion
======================================
The *simple exclusion process* lives in $\{0,1\}^\Z$ and its generator is $$\label{101}
{\cal L}^{ex}f(\eta) = \frac12\sum_{x\in\Z} [f(\eta^{x,x+1}) - f(\eta)]$$ where $$\label{130}
\eta^{x,x+1}(y) = \cases{\eta(y) &if $y\neq x, x+1$\cr
\eta(x+1) &if $y= x$\cr
\eta(x) &if $y= x+1$\cr}$$ It is convenient to construct the processes using the Harris graphical construction.
[**Harris graphical construction**]{} Let $(N_t(x):x\in\Z)$ be a family of independent Poisson processes of rate $\frac12$. For each $x$, $N_t(x)$ counts the number of Poisson events associated to $x$ in the time interval $[0,t]$. Denote $d N_t(x) =
\one\{$there is a Poisson event associated to $x$ at time $t\}=\lim_{h\to0}
(N_{t}(x) - N_{t-h}(x))$. Let $\eta_t$ be the process defined by $$\label{111}
d \eta_t(x) = (\eta_t(x-1) - \eta_t(x)) dN_t(x) + (\eta_t(x+1) -
\eta_t(x)) dN_t(x+1).$$ The process is well defined because for each finite time $t$ the value of the process in a finite box can be determined by looking at only a finite but random number of Poisson events and initial values. See for instance Arratia (1983). In words, the motion can be described as follows. The Poisson marks of $N_t(x)$ are associated to the bond $(x,x+1)$ and each time a Poisson mark occurs, the contents of the associated bond are interchanged. It is immediate to show that this process has generator .
[**Stirring particles.**]{} To introduce the notion of duality and to deal with the flux of particles it is convenient to follow the “stirring particles” as defined by Arratia (1983). Let $X^x_t$ be the position at time $t$ determined by $X^x_0=x$ and the equations $$\label{117}
d X^x_t = dN_t(X^x_t)- dN_t(X^x_t-1)$$ So that, each time a Poisson mark associated to one of the neighboring bonds of a particle occurs, the particle jumps across the bond. Of course, if both extremes of a bond are occupied, the particles jump simultaneously, respecting the exclusion condition “at most one particle per site”. For each $t\ge 0$ the (random) map $$\label{119}
x\mapsto X^x_t$$ is a bijection of $\Z$ in $\Z$. The (marginal) law of $X^x_t$ is a symmetric nearest neighbor random walk starting at $x$.
[**Duality.**]{} Let $y\mapsto D^y_t$ be the inverse map defined by $x = D^y_t$ if and only if $y =
X^x_t$. The following “duality formula” holds immediately $$\label{118}
\eta_t(y) = \eta_0(D^y_t)$$ So, $$\label{120}
\prod_{y\in A}\eta_t(y) = \prod_{y\in A}\eta_0(D^y_t)$$ Notice that for a finite set of sites $A$, $\{D^y_t:y\in A\}$ has the same one-time marginal as a simple exclusion process with initial condition $A$ (here we are identifying the configuration $\eta$ with the set $\{x:\eta(x)=1\}$). When $A=\{y\}$ (contains only one site), the one-time marginal $D^y_t$ has the same law as $X^y_t$ for all $t\ge 0$.
#### Integration by parts formula
Consider $(Y^i_t,Z^j_t)$ independent random walks with the same marginals as the stirring process $(X^i_t,X^j_t)$. The generator of the process $(Y^i_t,Z^j_t)$ is the following: $$\label{110a}
Uf(i,j) = \frac12 \sum_{e\in\Z^2:|e|=1}[f((i,j)+e) - f(i,j)]$$ and the generator of the process $X_t$ is $$\label{110b}
Vf(i,j) = \cases{\frac12 \sum_{e\in\Z^2:|e|=1}[f((i,j)+e) - f(i,j)] &if
$i-j>1$\cr
\frac12 f(j,i) + \frac12 f(i+1,j)+ \frac12 f(i,j-1)- \frac32 f(i,j)
&if
$i-j=-1$ \cr
\frac12 f(j,i) + \frac12 f(i-1,j)+ \frac12 f(i,j+1)- \frac32 f(i,j)
&if
$i-j=1$ }$$ Hence, for $i\neq j$, $$\label{110c}
Uf(i,j)- Vf(i,j) = -\,\frac12\one\{|i-j|=1\} (f(i,j)
+f(j,i)
-f(i,i) -f(j,j))$$ Let $U_t$ and $V_t$ be the semigroups generated by $U$ and $V$ respectively. Let $f:\Z^2\to\R$. Then $$\begin{aligned}
\label{110d}
\E f(X^i_t,X^j_t) - \E f(Y^i_t,Z^j_t) = [V_t - U_t] f(i,j)
\;=\; \int_0^t V_s [U-V] U_{t-s}f(i,j) ds\end{aligned}$$ where the last identity is the integration by parts formula (see Liggett (1985) Proposition 8.1.7). Now, using to compute we get for $i\neq j$: $$\begin{aligned}
\label{110e}
\lefteqn{\E f(X^i_t,X^j_t) - \E f(Y^i_t,Z^j_t) }\nonumber\\
&=&-\frac12 \int_0^t \,ds\;
\E\Bigl(\one\{|X^i_s-X^j_s|=1\}\\
&&\qquad\times\,
[f(Y^{X^{i}_{t-s}}_t,Z^{X^{j}_{t-s}}_t)+
f(Y^{X^{j}_{t-s}}_t,Z^{X^{i}_{t-s}}_t) -
f(Y^{X^{i}_{t-s}}_t,Z^{X^{i}_{t-s}}_t)
- f(Y^{X^{j}_{t-s}}_t,Z^{X^{j}_{t-s}}_t)]\Bigr).\nonumber\end{aligned}$$ This identity will be used in the sequel.
[**Flux.**]{} Let $J_t$ be the integrated flux of $\eta$ particles through the point $-1/2$ in the exclusion process: $$\label{eq:1}
J_t := \sum_{x< 0} \eta_0(x)\one\{X^{x}_t\ge 0\}\, - \,\sum_{x\ge 0}
\eta_0(x)\one\{X^{x}_t< 0\}$$ where $X^{x}_t$ is the position at time $t$ of the exclusion particle that at time zero was at position $x$.
Replacing in , we write $$\label{115}
J_t := \sum_{i< 0} \one\{X^{2i}_t\ge 0\}\, - \,\sum_{i\ge 0} \one\{X^{2i}_t<
0\}.$$
$J_t$ is almost symmetric. Let $$\label{1101}
H_t := \sum_{i<0} \one\{X^{2i}_t\ge 0\};\,\,\,
H'_t := \,\sum_{i\ge 0} \one\{X^{2i}_t<-1\};\,\,\,
I_t := \,\sum_{i\ge 0} \one\{X^{2i}_t<0\}.$$ Then, clearly, $$\label{1102}
J_t= H_t-I_t;\,\,\,H_t\sim H'_t;\,\,\,|H'_t-I_t|\leq1,$$ where $\sim$ means identity in distribution and is justified in this case by spatial and distributional symmetry.
\[104\] For the simple exclusion process with generator and initial condition $\eta_0 = 1-2\xi_0$, as defined in , $$\label{105}
\lim_{t\to\infty}{\V J_t \over \sqrt t}
=\frac{1}{4\sqrt\pi}.$$
Working from , we get $$\begin{aligned}
\nonumber
\E (J_t)^2 \= \sum_{i< 0} \P(X^{2i}_t\ge 0)\;
+\; \sum_{i\ge 0} \P(X^{2i}_t< 0)
\; +\;2\,\sum_{i<0}\sum_{i<j<0}\, \P(X^{2i}_t\ge 0,\,X^{2j}_t\ge 0) \\
\label{eq:2}
&+&\;2\,\sum_{i\ge 0}\sum_{i>j\ge 0}\, \P(X^{2i}_t< 0,\,X^{2j}_t< 0)
\;-\;2\sum_{i<0}\sum_{j\ge 0}\, \P(X^{2i}_t\ge 0,\,X^{2j}_t< 0)\\ \nonumber\\
\nonumber
(\E J_t)^2 \= \sum_{i< 0} \P^2(X^{2i}_t\ge 0)\; +\; \sum_{i\ge 0}
\P^2(X^{2i}_t< 0)
\; +\;2\,\sum_{i<0}\sum_{i<j<0}\, \P(X^{2i}_t\ge 0)\,\P(X^{2j}_t\ge 0) \\
\label{eq:3}
&+&\;2\,\sum_{i\ge 0}\sum_{i>j\ge 0}\, \P(X^{2i}_t< 0)\,\P(X^{2j}_t< 0)
\;-\;2\sum_{i<0}\sum_{j\ge 0}\, \P(X^{2i}_t\ge 0)\,\P(X^{2j}_t< 0)\end{aligned}$$ Immediately we have: $$\label{eq:4}
\sum_{i< 0} \P(X^{2i}_t\ge 0) + \sum_{i\ge 0} \P(X^{2i}_t< 0) = \sum_{i< 0}
\P(X^{i}_t\ge 0)$$ and analogously for the $\P^2$ terms in (\[eq:3\]). Using $$\label{eq:5}
\P(A\,B) - \P(A)\,\P(B) \,=\, - (\P(A\,B^c) - \P(A)\,\P(B^c))
\,=\, \P(A^c\,B^c) - \P(A^c)\,\P(B^c)$$ we get $$\label{eq:6}
\V J_t \,=\, {\cal V}_t\, +\, {\cal E}_t,$$ where $$\label{v0}
{\cal V}_t=\sum_{i< 0}\P(X^{i}_t\ge 0)-\sum_{i< 0}\P^2(X^{i}_t\ge 0)$$ and $$\begin{aligned}
\nonumber
{\cal E}_t \= \, \Bigl(\sum_{i,j<0,\,i\ne j}\,+\,\sum_{i,j\ge0,\,i\ne j}
\,+\,2\sum_{i<0,j\ge0}\Bigr)\Bigl(\P(X^{2i}_t\ge
0,\,X^{2j}_t\ge 0) - \P(X^{2i}_t\ge 0)\,\P(X^{2j}_t\ge 0)\Bigr)\\
\label{eq:7}
\= \, \sum_{i\ne j}\Bigl(\P(X^{2i}_t\ge
0,\,X^{2j}_t\ge 0) - \P(X^{2i}_t\ge 0)\,\P(X^{2j}_t\ge 0)\Bigr)\end{aligned}$$
Since $\P(X^{i}_t\ge 0)\P(X^{j}_t\ge 0)= \P(Y^{i}_t\ge 0,Z^{j}_t\ge 0)$, we can use with $f(i,j) =\one\{i\ge 0, j\ge 0\}$ to get $$\begin{aligned}
\nonumber
\lefteqn{\P(X^{i}_t\ge
0,\,X^{j}_t\ge 0) - \P(X^{i}_t\ge 0)\,\P(X^{j}_t\ge 0)}\\
\label{eq:8}
&=&-\frac12 \,\int_0^t \sum_{y} \P(
\{X^{i}_s, \,X^{j}_s\} = \{y,y +1\} )\,
\Bigl(\P(Y^y_{t-s}\ge 0 )-\P(Y^{y+1}_{t-s}\ge 0 )\Bigr)^2 \,ds\end{aligned}$$ See also Theorem 2 of Ferrari, Galves and Landim (2000) for a probabilistic proof of the previous identity. Translation invariance and self-duality of $(X^{i}_s, \,X^{j}_s)$ implies that equals $$\begin{aligned}
-\frac12\,\int_0^t \sum_{y} \P(\{X^0_s, \,X^{1}_s\} = \{i-y,j-y\} )
\P^2(Y^0_{t-s}=y)\,ds \label{8}\end{aligned}$$ [From]{} (\[eq:7\]) and (\[eq:8\])–, $$\label{eq:9}
{\cal E}_t\;=\; -\frac12\,\int_0^t \sum_{y} \P^2(Y^0_{t-s}=y)
\sum_{i\ne j}\P(\{X^0_s, \,X^{1}_s\} = \{2i-y,2j-y\} )\,ds.$$ Since $X^0_t\neq X^1_t$, $$\label{eq:10}
\sum_{i\ne j}\P(\{X^0_s, \,X^{1}_s\} = \{2i-y,2j-y\})
\;=\;2\,\P(X^{1}_s(\mod\, 2)=X^0_s(\mod\, 2)= y (\mod\, 2))$$ Let $A_y= \{(i,j)\in\Z^2: i(\mod\, 2)=j(\mod\,
2)=y(\mod\, 2)\}$. We show below that $$\label{eq:11}
\lim_{s\to\infty} \P((X^{0}_s,X^1_s)\in A_y) \,=\,
1/4\,.$$ uniformly in $y$. Hence, $$\label{eq:15}
\lim_{t\to\infty}t^{-1/2}{\cal E}_t
= -\frac14\lim_{t\to\infty}t^{-1/2}\int_0^t\sum_{y}\P^2(Y^0_{s}=y)\,ds$$ Let $Z^0_t$ be an independent copy of $Y^0_t$. Since $\sum_{y}\P^2(Y^0_{s}=y)=
\P(Y^0_{t}-Z^0_{t}=0)$, changing variables the above limit equals $$\begin{aligned}
\label{16c}
\lim_{t\to\infty}\int_0^1(st)^{1/2}\P(Y^0_{st}-Z^0_{st}=0)\,\frac{ds}{\sqrt s}
&=&\int_0^1\lim_{t\to\infty}(st)^{1/2}\P(Y^0_{st}-Z^0_{st}=0)\,\frac{ds}{\sqrt
{s}},\end{aligned}$$ where the interchange of the limit and the integral are guaranteed by the local central limit theorem for $(Y^0_t-Z^0_t)$, which is a symmetric random walk of rate $2$. This also implies that equals $$\begin{aligned}
\label{eq:16}
\int_0^1\frac{1}{\sqrt{2\pi}\sqrt2}\,\frac{ds}{\sqrt {s}}\;=\; \frac1{\sqrt \pi}.\end{aligned}$$ We conclude that $$\label{eq:17}
\lim_{t\to\infty}t^{-1/2}{\cal E}_t
\;=\;-\frac{1}{4\sqrt\pi}.$$
To compute ${\cal V}_t$ notice that $$\begin{aligned}
\label{eq:12}
\sum_{i< 0} \P(X^{i}_t\ge 0)\=\sum_{i>0} \P(X^0_t\ge i)\;=\;\E((X^0_t)^+);\\
\label{eq:13}
\sum_{i< 0} \P^2(X^{i}_t\ge 0)\=\sum_{i>0} \P^2(X^0_t\ge i)
\;=\;\sum_{i>0} \P(Y^0_t\wedge Z^0_t\ge i)\;=\;\E[(Y^0_t\wedge Z^0_t)^+],\end{aligned}$$ Thus $$\label{eq:14}
\lim_{t\to\infty}t^{-1/2}{\cal V}_t
\;=\;\E(X^+)-\E[(X\wedge X')^+]\;=\;\frac1{2\sqrt\pi},$$ where $X$ and $X'$ are i.i.d. standard normals.
Finally , substituting (\[eq:17\]) and (\[eq:14\]) in (\[eq:6\]) we get .
#### Proof of
The continuous time Markov chain $(Y^0_t(\mod\, 2), Y^1_t(\mod\, 2))$ converges exponentially fast to the uniform distribution in $\{(0,1),(1,0),(0,0),(1,1)\}$. This implies that there exist positive constants $C_1$, $C_2$ such that $$\label{11a}
|\P((Y^{i}_s,Z^j_s)\in A_y) \,-\,
1/4| \le C_1 e^{-C_2 t}\,.$$ uniformly in $i,j,y$. Writing $f_y(i,j):=\one\{(i,j)\in A_y\}$ and using we get $$\begin{aligned}
\nonumber
\lefteqn{ |\P((X^0_t,X^1_t)\in A_y) - \P((Y^0_t,Z^1_t)\in A_y)|}\nonumber\\
&\le&\frac12 \int_0^t \,ds\,\E\Bigl(\one\{|X^0_s-X^1_s|=1\}\\
&&\qquad\times\,
[f_y(Y^{X^{0}_{t-s}}_t,Z^{X^{1}_{t-s}}_t)+
f_y(Y^{X^{1}_{t-s}}_t,Z^{X^{0}_{t-s}}_t) -
f_y(Y^{X^{0}_{t-s}}_t,Z^{X^{0}_{t-s}}_t)
- f_y(Y^{X^{1}_{t-s}}_t,Z^{X^{1}_{t-s}}_t)]\Bigr)\nonumber\\
\label{11d}
&\le& 2 \int_0^t \,ds\,\P(|X^1_s-X^0_s|=1)\, C_1\, e^{-C_2 (t-s)}\end{aligned}$$ (using ). Now $|X^1_s-X^0_s|$ is a Markov chain in $\{1,2,\ldots\}$ with rates $p(1,2)=p(x,x+1)=p(x,x-1)=1/2,\, x>1$. It can be easily coupled to a a Markov chain in $\{0,1,2,\ldots\}$ starting in $0$, say $R^0_t$, with rates $p(0,1)=1,\, p(x,x+1)=p(x,x-1)=1/2,\, x>0$ in such a way that $|X^1_s-X^0_s|\geq R^0_s$ for all $s$. Since $R^0_t$ is a simple symmetric random walk reflected at the origin, we get that $\lim_{s\to\infty}\P(|X^1_s-X^0_s|=1)\leq\lim_{s\to\infty}\P(|R^0_s|\leq1)=0$ and thus, from and dominated convergence (after a change of variables $s\to t-s'$), it follows that $\lim_{t\to\infty}\P((X^0_s,X^1_s) \in
A_y)=\lim_{t\to\infty}\P((Y^0_s,Z^1_s) \in A_y)=1/4$.
\[1103\] Let $H_t$ be as in and $\tilde H_t=t^{-1/4}[H_t-\E (H_t)].$ Then for all $\la\in\R$ $$\label{eq:expmom}
\limsup_{t\to\infty}\E(e^{\la \tilde H_t})\;\leq \;e^{s\la^2/2},$$ where $s=1/\sqrt{2\pi}$.
[**Proof.**]{} $$\label{eq:l6}
\E(e^{\la \tilde H_t})\;=\;
\frac{\E\left[\exp\left(\la t^{-1/4}\sum_{i<0}\one\{X^{2i}_t\ge0\}\right)
\right]}
{\exp\left(\la t^{-1/4}\sum_{i<0}\P(X^{2i}_t\ge0)\right)}$$
We will show that the quotient in (\[eq:l6\]) is bounded above by a constant. For that, we need to evaluate the expected value in that equation. Let $\la\geq0$. We will argue below that $$\label{eq:l7}
\E\left[\exp\left(\la t^{-1/4}\sum_{i<0}\one\{X^{2i}_t\ge0\}\right)\right]
\;\leq\;\prod_{i<0}\E\left[\exp\left(\la t^{-1/4}\one\{X^{2i}_t\ge0\}\right)\right].$$
The last expectation equals $$\label{eq:l8}
1+[\exp(\la t^{-1/4})-1]\P(X^{2i}_t\ge0)= 1+[\la t^{-1/4}+(\la^2 t^{-1/2}/2)
+ o(t^{-1/2})]\P(X^{2i}_t\ge0),$$ for all $t$ large enough. The last expression is bounded above by $$\label{eq:l9}
\exp\left\{[\la t^{-1/4}+(\la^2 t^{-1/2}/2) + o(t^{-1/2})]\P(X^{2i}_t\ge0)\right\}.$$ Substituting into the right hand side of (\[eq:l7\]), we get $$\label{eq:l10}
\exp\left\{[(\la^2 t^{-1/2}/2) + o(t^{-1/2})]\sum_{i<0}\P(X^{2i}_t\ge0)\right\}.$$ as an upper bound for the quotient in (\[eq:l6\]). It is not difficult to see that the expression on the exponent in (\[eq:l10\]) converges to $e^{s\la^2/2}$.
To finish the argument for $\la\geq0$, we have to justify the inequality (\[eq:l7\]). That follows from taking limits as $M\to-\infty$ (and using monotone convergence) on the respective inequalities gotten by replacing the infinite sums by $\sum_{M<i<0}$. These are justified by the fact that the functions $\exp(t^{-1/4}\sum_{M<i<0}\one\{X^{2i}_t>0\})$ are bounded, symmetric and positive definite for all $M<0$. The inequalities then follow from Proposition 1.7, Chapter VIII of Liggett (1985).
For the case $\la<0$, we use the identity $$\label{eq:l1}
\sum_{i< 0} \one\{X^{2i}_t\ge 0\} -\sum_{i< 0} \P(X^{2i}_t\ge 0)
=-\left[\sum_{i< 0} \one\{X^{2i}_t< 0\} -\sum_{i< 0} \P(X^{2i}_t< 0)\right]$$ and a similar argument as above.
\[110\] Let $\eta_0$ be given by the flat condition as in Theorem \[104\]. Then $$\label{eq:expabs}
\sup_{t\geq0}\E(e^{|J_t/t^{1/4}|})<\infty.$$ Furthermore for all $K>0$ and all $t$ large enough $$\label{eq:lema}
\P(|J_t|>Kt^{1/4}\log t)\;\leq \;c t^{-K},$$ where $c$ is a constant.
[**Proof.**]{} The bound follows straightforwardly from Lemma and relations .
[From]{} the relations , to show it is enough to prove the result with $|H_t-\E (H_t)|$ replacing $|J_t|$ in (the constant of course does not need to be the same). We have $$\label{1105}
\P(|H_t|>Kt^{1/4}\log t)=\P(|\tilde H_t|>\log t^K)\leq c' t^{-K}$$ where the last inequality follows from the exponential Markov inequality and $c'=\sup_{t\geq0}\E(e^{|\tilde H_t|})$ is finite by .
[**Graphical construction of free process.**]{} Let $\zeta_t$ be the process defined by $$\label{112}
d \zeta_t(x) = \Delta\zeta_t(x)\, dN_t(x),$$ where the discrete Laplacian $\Delta$ was defined in . In words, each time a Poisson mark of the process $N_t(x)$ occurs, the height at $x$ at time $t$ decreases or increases two unities, according to the value of the Laplacian at this point at this time; if the Laplacian vanishes, no jump occurs. This process has generator .
\[103\] Let $\eta_0(x) = \zeta_0(x+1) -\zeta_0(x)$. Then $$\label{113}
\eta_t(x) = \zeta_t(x+1)-\zeta_t(x)$$ where the processes $\zeta_t$ and $\eta_t$ are defined by and and have initial conditions $\eta_0$ and $\zeta_0$ respectively. Furthermore, $$\label{102}
\zeta_t (0) -\zeta_0(0) = 2J_t\,.$$
Notice that from , $$\begin{aligned}
\label{x1}
\Delta\zeta_t(0) =2 (\eta_t(-1)- \eta_t(0))\end{aligned}$$ Assume that there is a mark of the process $N_t(-1)$ at time $t$. Then , and imply that if $\eta_t(-1)- \eta_t(0)=0$ no changes occur neither for $\eta_t(-1),\eta_t(0)$ nor for $\zeta_t(0)$; if $\eta_t(-1)- \eta_t(0)=1$, an exclusion particle jumps from $-1$ to $0$ and the free process at the origin jumps two units up; if $\eta_t(-1)-
\eta_t(0)=-1$, an exclusion particle jumps from $0$ to $-1$ and the free process at the origin jumps two units down. Identity follows from .
\[110q\] Let $\zeta_t$ be the free process with flat initial condition . Then $$\label{105p}
\lim_{t\to\infty}{\V \ze_t(0) \over \sqrt t}
=\frac{1}{\sqrt\pi};$$ $$\label{110r}
\sup_{t\geq0}\E(e^{|\ze_t(0)/t^{1/4}|})<\infty$$ and for all $K>0$ and all $t$ large enough $$\label{110p}
\P(|\ze_t(0)|>Kt^{1/4}\log t)\;\leq \;c t^{-K}\,.$$
It follows from identity , the limit and the bounds and .
Coupling the wall and the free processes
========================================
We construct graphically the wall process which simultaneously provides another graphical construction for the free process. Under this construction the wall process dominates the free one. We consider two independent families of Poisson processes with the same law as $N_t(x)$ called $N^+_t(x)$ and $N^-_t(x)$, to be used for upwards and downwards jumps, respectively. The process satisfying the equations $$\label{121}
d \xi_t(x) \;=\; \Delta
\xi_t(x)\,\one\{\Delta
\xi_t(x) > 0\}\, dN^+_t(x)\;+\;\Delta
\xi_t(x)\,\one\{\Delta
\xi_t(x) < 0,\,\xi_t(x) + \Delta
\xi_t(x) \ge 0\}\, dN^-_t(x)$$ has generator . The process $\zeta_t$ satisfying $$\label{121a}
d \ze_t(x) = \Delta
\xi_t(x)\,\one\{\Delta
\xi_t(x) > 0\}\, dN^+_t(x)+\Delta
\xi_t(x)\,\one\{\Delta
\xi_t(x) < 0\}\, dN^-_t(x).$$ has generator .
In words, when a time event of the process $N^+_t(x)$ occurs at time $t$, the process $\xi_t$ at site $x$ and time $t$ jumps two units upwards if $\Delta
\xi_t(x)>0$. When a time event of the process $N^-_t(x)$ occurs at time $t$, the process $\xi_t$ at site $x$ and time $t$ jumps two units downwards if $\Delta \xi_t(x)<0$ *and* the wall condition $\xi_t(x) + \Delta \xi_t(x)
\ge 0$ holds. The process satisfying follows the same marks in the same manner but ignoring the wall condition. The difference with the process satisfying is that in this case the Poisson events $N_t$ are used for *both* upwards and downwards jumps; this construction is not attractive in the sense that it does not satisfy below.
Let $r$ be a non negative integer and $\xi^r_t$ and $\zeta^r_t$ be the processes defined by and but with initial condition $$\label{p13}
\xi^r_0(x) \,=\, \zeta^r_0(x) \,=\, r + x(\mod 2)$$ Notice that $\ze^0_t$ and $\ze_t$ as defined in have the same law but are *different* processes. The processes $\xi_t$ and $\xi^r_t$ defined by and the same initial condition satisfy $$\label{p12}
\xi_t(0) \le \xi^r_t(0)$$ for all $r\ge 0$. This joint construction corresponds to what Liggett (1985) calls *basic coupling*.
\[p10\] There exists a constant $c>0$ such that for any $K>0$ and $t\geq0$ $$\label{p11}
\P(\xi_t(0) > 2K t^{1/4}\log t) \le c t^{2-K}$$
Let $a_t = 2K t^{1/4}\log t$. Take $r\ge 0$ and write $$\begin{aligned}
\label{p14}
\P(\xi_t(0)\ge a_t) &\le& \P(\xi^r_t(0)\ge a_t) \nonumber\\
&=& \P(\xi^r_t(0)\ge a_t\,,\,\xi^r_t(0) = \zeta^r_t(0))\,+\, \P(\xi^r_t(0)\ge
a_t\,,\,\xi^r_t(0) \neq \zeta^r_t(0) )\nonumber\\
&\le& \P(\zeta^r_t(0)\ge a_t)\,+\, \P(\xi^r_t(0) \neq \zeta^r_t(0) )\end{aligned}$$ The first term in will be bounded using Corollary \[110q\]. To bound the second term notice that if the interacting process and the free process differ at the origin this is due to a collision of the interacting process with the wall at some point $x$ that separate the two processes at $x$ at some time $s$; the discrepancy then propagates and arrives to zero by time $t$. We fix an $\alpha>0$ and separate the discrepancies in two classes: those that come from the interval $[-\alpha t,\alpha t]$ and those that come from outside this interval. If in the time interval $[0,t]$ the free process does not touch the wall in the space interval $[-\alpha t,\alpha t]$, then the discrepancy must come from outside. Hence, $$\begin{aligned}
\nonumber
\{\xi^r_t(0)\neq \zeta^r_t(0) \}\!\!\!\!&\subset&\!\!\!\!
\{\zeta^r_t(x)<0\,\hbox{ for some }s\in[0,t],\, x\in[-\alpha t,\alpha t]\}\\
\label{p15}
&\,\,\,\,\,\,\,\,\,\,\,\,\cup&\!\!\!\! \{\hbox{ a discrepancy from }[-\alpha
t,\alpha t]^c\hbox{ reaches } 0 \hbox{ up to time }t\}\end{aligned}$$ Observe that the law of $\zeta^r_t(x)-r$ is the same as the law of $\ze_t(0)$ and that $\P(\ze_s(0) <-r)\le \P(\ze_s(0) > r)$ due to the initial condition being non-negative. Hence, fixing $$\label{1234}
r= a_t/2\,,$$ the probability of the first event in the right hand side of is bounded by $$\begin{aligned}
\label{p23}
(2\alpha t+1)\, \P(\ze_s(0) > a_t/2\, \hbox{ for
some } s\in[0,t])\end{aligned}$$ [From]{} and the exponential Markov inequality, we have that $\sup_{s\leq t}\P(\ze_s(0)>r)\le ct^{-K}$ for some constant $c$, so we can bound with $$\label{p24}
(2\alpha t+1)\, \int_0^t c t^{-K} ds \;\le\; (2\alpha+1) \, c\, t^{2-K}$$
To bound the probability of the second event in the right hand side of notice that discrepancies cannot travel faster than $N_t$, a Poisson process of parameter 1. Hence $$\begin{aligned}
\label{p17}
\P(\hbox{a discrepancy from }[-\alpha
t,\alpha t]^c\hbox{ reaches } 0 \hbox{ up to time }t) \le
2\P(N_t > \alpha t)\le 2\,e^{-t (\alpha+1-e)}\end{aligned}$$ using the exponential Chebyshev inequality. Fixing $\alpha=2$, and using the bounds and , the probability of is bounded by $$\label{p31}
4 \, c\, t^{2-K} \,+\,2\,e^{-t (3-e)}\le
c't^{2-K}$$ for some constant $c'$ and sufficiently large $t$.
[**Proof of Theorem \[98\].**]{} It follows straightforwardly [from]{} (\[110r\]) that ${\tilde\ze_t^2}$ is uniformly integrable, where ${\tilde\ze_t}=t^{-1/4}\ze_t(0)$. This, together with , implies the lower bound in , as we will see now. Indeed, implies that $t^{-1/4}\E\xi_t(0)\geq\E|\tilde\ze_t|$. Now, $$\begin{aligned}
\label{p25}
\V\tilde\ze_t\leq\E\tilde\ze_t^2=
\E(\tilde\ze_t^2;\,\tilde\ze_t^2\leq M^2)
+\E(\tilde\ze_t^2;\,\tilde\ze_t^2>M^2)
\leq M\E|\tilde\ze_t|+\e_M\end{aligned}$$ uniformly in $t$, where $M$ is an arbitrary positive number and $\e_M\to0$ as $M\to\infty$. Thus $t^{-1/4}\E\xi_t(0)\geq(\V\tilde\ze_t-\e_M)/M$. We conclude that $$\begin{aligned}
\label{p26}
\liminf_{t\to\infty}t^{-1/4}\E\xi_t(0)
\geq\sup_{M>0}\left(\frac1{\sqrt\pi}-\e_M\right)/M>0.\end{aligned}$$
For the upperbound, we use \[110p\] to obtain $$\begin{aligned}
\label{p27}
{\E \xi_t(0) \over t^{1/4}\log t} = \sum_{k\ge 0} \P(\xi_t(0) > k
t^{1/4}\log t)
\le 4+\sum_{k\ge 5}c't^{2-k/2} \;\le c_2 \;<\;\infty\end{aligned}$$ for some constant $c_2<\infty$.
Numerical simulation
====================
We have simulated the processes $\zeta_t$ and $\xi_t$ numerically, using various pseudo-random number generators. The interface is of length $L$ with periodic boundary conditions, so that the processes live in $$\X_{0L}=\{\zeta\in\Z^{\Z/L\Z}: |\zeta(x)- \zeta(x+1)|=1,\, \zeta(0)
\hbox{ even}\}$$ or $$\X_L=\{\xi\in\N^{\Z/L\Z}: |\xi(x)- \xi(x+1)|=1,\, \xi(0) \hbox{ even}\}$$ Time is an integer multiple of $2L^{-1}$, i.e. $t\in (2L^{-1})\N$. For each time step, a site $x$ is chosen randomly according to the uniform measure on $\Z/L\Z$, and the interface is updated with the same rules as in the continuous time version of the processes. The transition operator corresponding to a time step $\delta t=2L^{-1}$ is $$T_0f(\zeta)=f(\zeta)+L^{-1}\sum_x \, [f(\zeta + \Delta
\zeta(x)\,\delta_x) - f(\zeta)]$$ or $$Tf(\xi)=f(\xi)+ L^{-1}\sum_x \one\{\xi + \Delta
\xi(x) \,\delta_x \ge 0\}\, [f(\xi + \Delta
\xi(x)\,\delta_x) - f(\xi)]$$ $Tf(\xi)$ is the expected value of the function $f$ evaluated at time $2L^{-1}$ (after one step) when the initial configuration is $\xi$ for the discretized version of the process. The same interpretation is valid for $T_0f(\zeta)$. By abuse of notation we call the discrete time versions of the process $\xi_t$ and $\zeta_t$ as we did for the continuous time versions. Notice that $$\lim_{L\to\infty} {Tf(\xi)-f(\xi))\over\delta t} = {\cal L}f(\xi)\,;\qquad
\lim_{L\to\infty} {T_0f(\zeta)-f(\zeta))\over\delta t} = {\cal L}_0f(\zeta)\,.$$ The processes $\xi_t$ and $\zeta_t$ are coupled in the simplest possible way: the same random sequence of sites are used for both. Notice however that this coupling is different from the one described in Section 3 (in particular it is not attractive in the sense that it does not necessarily satisfy but it is faster). For $L$ finite the discrete time and continuous time processes can be identified up to a time change, using the ordered sequence of updated sites. The random time involved in the time change has fluctuations which should be negligible for our purposes.
The numerical samples for the data shown below were drawn using either the Mersenne Twister pseudorandom integer generator, see Matsumoto, Nishimura (1998), or the R250 pseudorandom generator, see Kirkpatrick, Stoll (1981). The length $L$ is $10^6$ or $2^{20}=1024^2$ and time runs up to $2.10^6$ or $2^{21}$. The number of calls to the generator for the realization of one sample of length $L$ up to time $t$ is $L.t/2\le10^{12}$, which of course is much less than the period of the generator (a necessary but not sufficient condition for reliability). We compute empirical averages $$\overline{\xi_t^2}=L^{-1}\sum_x\xi_t(x)^2\ ,\quad
\overline{\zeta_t^2}=L^{-1}\sum_x\zeta_t(x)^2$$ and empirical distribution functions $$f_t(n)=L^{-1}\sum_x\one\{\xi_t(x)=n\}\ ,\quad
f_{0,t}(n)=L^{-1}\sum_x\one\{\zeta_t(x)=n\}\ ,\quad n\in\Z$$ scaled into $$\phi_t(s)=t^{1/4}L^{-1}\sum_x\one\{t^{-1/4}\xi_t(x)=s\}\ ,\quad
\phi_{0,t}(s)=t^{1/4}L^{-1}\sum_x\one\{t^{-1/4}\zeta_t(x)=s\}\ ,\quad
s\in t^{-1/4}\Z$$ which, extended to $s\in\R$ approximate the Schwartz distributions $$\tilde\phi_t(s)=L^{-1}\sum_x\delta(t^{-1/4}\xi_t(x)-s)\ ,\quad
\tilde\phi_{0,t}(s)=L^{-1}\sum_x\delta(t^{-1/4}\zeta_t(x)-s)\ ,\quad
s\in\R$$ where $L^{-1}\sum_x$ is an ersatz for the expectation over a real random variable, limit of $t^{-1/4}\xi_t(x)$.
The processes were studied for time $t\le 2L$, whereas the effect of finite size with periodic boundary conditions is expected to be visible only after a time of order $L^2$, the relaxation time of an interface of length $L$. The law of large numbers in empirical averages as above is believed to be at work with an effective number of weakly dependent variables of order $L/t^{1/2}$: the interface at time $t$ can be thought of as a collection of $L/t^{1/2}$ segments of length $t^{1/2}$, the different segments being weakly dependent. For $t=L$ and one sample of the process, we have only $t^{1/2}$ independent segments, hence an expected relative statistical error of order $t^{-1/4}$. This explains the more erratic behaviour at larger times in Fig. 1.
The numerical experiment clearly favors an asymtotic behavior $\E\xi_t(0)^2\sim t^{1/2}\log t$ as $t\to\infty$.
Figure 2 shows the scaled empirical distribution functions at various large times. Clearly $t^{-1/4}\zeta_t(x)$ converges to a centered Gaussian random variable as expected. The distributon function of $t^{-1/4}\xi_t(x)$ is markedly asymmetrical. Zooming around $s=0$ indicates $\phi_\infty(0)=\phi_\infty'(0)=0$ and $\phi_\infty''(0)>0$, and $$f_t(0)=L^{-1}\sum_x\one\{\xi_t(x)=0\}\sim t^{-1/2}$$
Interface of the Ising model at zero temperature {#ising}
================================================
In this section we explain the relation of our model with the interface of a particular Ising model at zero temperature. Let the “inverse temperature” $\beta\ge 0$ and $\sigma_t\in\{-1,+1\}^{\Z^2}$ be the Ising model with generator $${\cal L}_{\beta}f(\sigma) = \frac12\sum_{x\in\Z^2}
c_\beta(x,\sigma)\, [f(\sigma^x)-f(\sigma)]$$ with $\sigma^x (z) = \sigma (z)$ for $z\neq x\in\Z^2$, $\sigma^x(x) =
-\sigma(x)$ and $c_\beta(x,\sigma)$ are the Glauber rates $$c_\beta(x,\sigma) = \frac{e^{-\beta H(\sigma^x)}}{ e^{-\beta H(\sigma^x)}+e^{-\beta H(\sigma)}}$$ with Hamiltonian $$H(\sigma) =- \sum_x\sum_{y:|y-x|=1}\sigma(x)\sigma(y) - h \sum_{x:x_1> x_2}
\sigma(x)$$ for some positive magnetic field $h>0$. Consider the case $\beta=\infty$ and assume that the starting configuration $\sigma_0$ is “all ones” below the diagonal and “all minus ones” above or in the diagonal: $$\label{s00}
\sigma_0(x_1,x_2) = \cases{ +1, &if $x_1>x_2$; \cr
-1, &if $x_1\leq x_2$.}$$ In this case for all $t$ the configuration $\sigma_t$ has the property that all sites have either exactly two or no neighbor with opposite sign; furthermore, only sites above or in the diagonal may be negative. As a consequence, for all $t\ge 0$ the rates $c(x,\sigma_t)$ are positive only for sites $x$ above or in the diagonal for which there are exacly two neighboring sites with different sign: under initial condition , $$c_\infty(x,\sigma_t) = \cases {
1/2 &if $\sum_{y:|y-x|=1}
\one\{\sigma_t(y)\neq\sigma_t(x)\}=2$ and $x_1\le x_2$\cr
0 &otherwise}$$ To get the wall process of Theorem \[98\] from the above dynamics with initial condition , we first rotate the lattice by $-45^0$ and multiply by $\sqrt 2$, that is, we perform the transformation $R:\Z^2\to{\Z}^2_2;R(x,y)=(x+y,x-y)$, where ${\Z}^2_2:=\{(x,y)\in\Z^2:x+y\mbox{ is even}\}$ is the even sublattice of $\Z^2$. The above dynamics then induces a dynamics in $\{-1,+1\}^{\Z^2_2}$ given by $\tilde\sigma_t(z)=\sigma_t(R^{-1}z)$, $z\in{\Z}^2_2,t\geq0$. Defining $\tilde\xi_t(x):= \min\{y:(x,y)\in{\Z}^2_2\mbox{ and
}\tilde\sigma_t(x,y)=-1\}$, $x\in\Z,t\geq0$, we have that the wall process $\xi_\cdot(\cdot)$ with generator and initial configuration has the same law as $\tilde\xi_\cdot(\cdot)$.
D.B. Abraham, P. Upton (1989) Dynamics of Gaussian interface models. [*Phys. Rev B **39***]{}, 736.
D.B. Abraham, P. Collet, J. De Coninck, F. Dunlop (1990) Langevin Dynamics of an Interface near a Wall. [*J. Stat. Phys. **42***]{}, 509-532.
E.V. Albano, K. Binder, D.W. Heermann, W. Paul (1992) [*Physica A **183***]{}, 130
R. Arratia (1983) The motion of a tagged particle in the simple symmetric exclusion system on ${\Z}$. [*Ann. Probab. **11***]{}, no. 2, 362–373.
N. C. Bartelt, J. L. Goldberg, T. L. Einstein, Ellen D. Williams, J. C. Heyraud and J. J. Métois (1993) Brownian motion of steps on Si(111). [*Phys. Rev B **48***]{}, 15453-15456.
K. Binder (1990) Growth kinetics of wetting layers at surfaces. pp 31-44 in [*Kinetics of Ordering and Growth at Surfaces*]{}. Edited by M.G. Lagally. Plenum Press, New-York.
B. Blagojevic, P.M. Duxbury (1999) Atomic diffusion, step relaxation, and step fluctuations. [*Phys. Rev E **60***]{}, 1279-1291.
E. Bolthausen, J.-D. Deuschel, O. Zeitouni (1995) Entropic repulsion of the lattice free field. [*Comm. Math. Phys. **170***]{}, no. 2, 417–443.
J. Bricmont (1990) Random surfaces in statistical mechanics (1990) in Wetting phenomena. [*Proceedings of the Second Workshop held at the University of Mons*]{}, Mons, October 17–19, 1988. Edited by J. De Coninck and F. Dunlop. [*Lecture Notes in Physics, **354***]{}. Springer-Verlag, Berlin.
J. Bricmont, A. El Mellouki, J. Fröhlich (1986) Random surfaces in Statistical Mechanics - roughening, rounding, wetting [*J. Stat. Phys. **42***]{}: (5-6) 743-798
F. Cesi, F. Martinelli (1996) On the layering transition of an SOS surface interacting with a wall. I. Equilibrium results. [*J. Stat. Phys. **82***]{}, no. 3-4, 823–913.
J. De Coninck, F. Dunlop, F. Menu (1993) Spreading of a Solid-On-Solid drop. [*Phys. Rev. E **47***]{}: (3) 1820-1823.
A. De Masi, P. A. Ferrari (2002) Flux fluctuations in the one dimensional nearest neighbors symmetric simple exclusion process. To appear in [ *J. Stat. Phys.*]{} http://front.math.ucdavis.edu/math.PR/0103233.
J.-D. Deuschel (1996) Entropic repulsion of the lattice free field. II. The $0$-boundary case. [*Comm. Math. Phys. **181***]{}, no. 3, 647–665.
J.-D. Deuschel, G. Giacomin (1999) Entropic repulsion for the free field: pathwise characterization in $d\geq3$. [*Comm. Math. Phys. **206***]{}, no. 2, 447–462.
E. Dinaburg, A.E. Mazel (1994) Layering transition in SOS model with external magnetic-field. [*J. Stat. Phys. **74***]{}: (3-4) 533-563.
P. A. Ferrari, S. Martínez (1998) Hamiltonians on random walk trajectories. [*Stochastic Process. Appl. **78***]{}, no. 1, 47–68.
P. A. Ferrari, A. Galves, C. Landim (2000) Rate of convergence to equilibrium of symmetric simple exclusion processes. [*Markov Processes and Related Fields **6***]{}, 73-88.
T. Funaki, S. Olla (2001) Fluctuations for $\nabla\phi$ interface model on a wall. Stochastic Process. Appl. 94, no. 1, 1–27.
P. Holický, M. Zahradník (1993) On entropic repulsion in low temperature Ising models. [*Cellular automata and cooperative systems*]{} (Les Houches, 1992), 275–287, NATO Adv. Sci. Inst. Ser. C Math. Phys. Sci., 396, Kluwer Acad. Publ., Dordrecht.
W. D. Kaigh (1976) An invariance principle for random walk conditioned by a late return to zero. [*Ann. Probability **4***]{}, no. 1, 115–121.
S. Kirkpatrick, E. Stoll (1981) [*J. Computational Physics*]{} [**40**]{}, 517-526.
J.L. Lebowitz, A.E. Mazel, Y.M. Suhov (1996) An Ising interface between two walls: Competition between two tendencies. [*Reviews In Mathematical Physics **8***]{}: (5) 669-687.
J.L. Lebowitz, C. Maes (1987) The effect of an external field on an interface, entropic repulsion. [*J. Stat. Phys. **46***]{}, no. 1-2, 39–49.
T.M. Liggett (1985) [*Interacting particle systems. Grundlehren der Mathematischen Wissenschaften \[Fundamental Principles of Mathematical Science\], 276.*]{} Springer-Verlag, New York-Berlin.
R. Lipowsky (1985) Nonlinear growth of wetting layers. [*J. Phys. A **18***]{}, L585-L590
M. Matsumoto, T. Nishimura (1998) Mersenne Twister: A 623-dimensionally equidistributed uniform pseudorandom number generator, ACM Trans. on Modeling and Computer Simulation [**8**]{}, 3-30. (http://www.math.keio.ac.jp/ matumoto/emt.html).
K. K. Mon, K. Binder, and D. P. Landau (1987) Monte Carlo simulation of the growth of wetting layers [*Phys. Rev.*]{} B [**35**]{}: (7) 3683 – 3685.
D. Stauffer and D. P. Landau (1989) Interface growth in a two-dimensional Ising model. [*Phys. rev. B **39***]{}, 9650-9651.
François Dunlop Laboratoire de Physique Théorique et Modélisation (CNRS-ESA8089) Université de Cergy-Pontoise, 95031 Cergy-Pontoise, France [[email protected]]{} http://www.u-cergy.fr/rech/labo/equipes/ptm/Dunlop Pablo A. Ferrari, Luiz Renato G. Fontes IME–USP, Caixa Postal 66281, 05315-970 - São Paulo, Brazil [[email protected], [email protected]]{} http://www.ime.usp.br/$\widetilde{\phantom m} $pablo, http://www.ime.usp.br/$\widetilde{\phantom m} $lrenato
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The interaction between a photon and a qubit in the Janeys-Cummings (JC) model generates a kind of quasiparticle called polariton. While they are widely used in quantum optics, difficulties in engineering controllable coupling of them severely limit their applications to simulate spinful quantum systems. Here we show that, in the superconducting quantum circuit context, polariton states in the single-excitation manifold of a JC lattice can be used to simulate a spin-1/2 system, based on which tunable synthetic spin-orbit coupling and novel topological polaritons can be generated and explored. The lattice is formed by a sequence of coupled transmission line resonators, each of which is connected to a transmon qubit. Synthetic spin-orbit coupling and effective Zeeman field of the polariton can both be tuned by modulating the coupling strength between neighbouring resonators, allowing for the realization of a large variety of polaritonic topological semimetal bands. Methods for detecting the polaritonic topological edge states and topological invariants are also proposed. Therefore, our work suggests that the JC lattice is a versatile platform for exploring spinful topological states of matter, which may inspire developments of topologically protected quantum optical and information processing devices.'
author:
- 'Feng-Lei Gu$^1$, Jia Liu$^1$, Feng Mei$^{2,3}$, Suotang Jia$^{2,3}$, Dan-Wei Zhang$^1$ and Zheng-Yuan Xue$^1$'
title: 'Synthetic spin-orbit coupling and topological polaritons in Janeys-Cummings lattices'
---
Introduction
============
The Janeys-Cummings (JC) model proposed in 1963 [@JCModel] is a seminal theoretical model treating light-matter interaction with full quantum theory, i.e., the interaction of a quantized electromagnetic field with a two-level atom. This model has been widely applied to many quantum platforms for studying the interaction of a quantized bosonic field with a qubit, which now has become the cornerstone in quantum optics and quantum computation [@Ion; @CQED; @SC3; @SC5; @Nano; @Opto]. Furthermore, an interconnected array of multiple JC systems can form a JC lattice [@Hartmann; @Fazio; @Nori; @Koch], which provides an innovative quantum optical platform for studying condensed matter physics. This is highlighted by previous works which show that coupled JC systems can be used to realize the Bose-Hubbard model and investigate superfluid-to-Mott-insulator phase transition [@Hartmann2006; @Greentree; @Angelakis]. However, spinful lattices have not been simulated in this platform due to the difficulty in engineering a tunable coupling between different cavities.
On the other hand, the search of topological states of matter in artificial systems recently has become a rapidly growing field of research [@add-njp; @add-prl109; @TPCA1; @TPCA2; @TPCA3; @TPPho1; @TPPho2; @TPPhonon; @TPPolariton; @TPOpto]. Topological states are characterized by topological invariants which are robust to the smooth changes in system parameters and disorders, where topological edge states can be employed for robust quantum transport [@Kane; @Zhang]. Therefore, they hold tremendous promise for fundamental new states of matter as well as for dissipationless quantum transport devices and topological quantum computation [@TPQC]. One of the key ingredients for generating such states is to realize tunable spin-orbit coupling (SOC). Significant theoretical and experimental progress on realizing synthetic SOC recently have been achieved in ultracold atom systems [@SOC1; @SOC2; @SOC3]. This progress stimulate great research interests to explore topological states with ultracold atoms trapped in optical lattices [@TPExp2; @TPExp3; @TPExp4; @TPExp6]. However, the limited trapping time and the site addressing difficulty increase the experimental complexity, i.e., it is generally difficult to have modulable coupling between two neighbouring sites in an optical lattice.
Here, we find that the JC lattice system can be used to realize various topological spin lattice models, where synthetic polaritonic SOC and Zeeman field can be induced with *in situ* tunability, which provides a flexible platform to explore topological states of matter with great controllability. Specifically, we consider realizing the JC lattice in the context of superconducting quantum circuits, where each JC lattice site is constructed by a transmission line resonator (TLR) coupled to a two-level transmon qubit. We find that the dressed polariton states in the single-excitation manifold in each JC lattice site can simulate a spin-1/2 system. Particularly, synthetic SOC and Zeeman field for polaritons can be induced and manipulated by only engineering the coupling strength between neighbouring resonators. Meanwhile, we show that, based on tunable synthetic SOC and Zeeman field, nodal-loop semimetal bands [@NLS; @NodalLoop1; @NodalLoop2] and topological polaritons can be realized and explored in the simulated JC lattice. Moreover, through calculating the topological winding number, we find that this tunable system has a rich topological phase diagram.
Our proposal to explore the topological states in the JC lattice system is different from previous ones based on the optical lattices [@SOC1; @SOC2; @SOC3]. In particular, our proposal has a number of advantages. (i) Unlike ultracold atoms, polaritons are quasiparticles which are hybrids of photons and qubit excitations. Topological polaritons emerge from the topological structure of light-matter interaction, where photons and qubit excitations are topologically trivial by themselves, but combining together, they become hybrid topological states. Therefore using polariton for quantum simulation enriches our controlling methods – both photonic and atomic means take effects. (ii) The systematic parameters in JC lattice systems can be tuned at a single-site level, which allows us to generate a wide variety of SOC forms. (iii) The geometry of JC lattice can be artificially designed and lattice boundaries are easy to be created for observing topological edge states, thus various topological lattice models and topological effects can be constructed and probed. (iv) The particle number putting in a JC lattice can be deterministically controlled. With such an advantage, we present a method using single-particle quantum dynamics to probe topological winding numbers and topological polariton edge states. (v) In the quantum optics platform, JC lattice systems previously have generated multiple important applications, including masers, lasers, photon transistors, and quantum information processors. Meanwhile, there are indeed several disadvantages in our proposed JC lattice system, such as the limited system size, parameter fluctuations, and decoherence. However, the essential physics of the simulated topological polariton states, such as the topological invariants and edge states, can still be detected under these realistic circumstances. Therefore, the topological JC lattice system in superconducting quantum circuits offers the possibility to develop functional topological spin quantum devices.
![**The proposed superconducting circuit implementation of spin-1/2 lattice models.** (A) The “spin-1/2" polariton lattice with two types of unit cells, A-type (red) and B-type (blue), arranged alternately. Each unit cell has two pseudo-spin-1/2 states simulated by the two single-excitation eigenstates of the JC model. The two types of unit cells are of the different qubit and photon eigenfrequencies and JC coupling strengths. The zoom-in figure details the equivalent superconducting circuits of two neighbouring unit cells and their coupling circuit, which is a combination of a SQUID and an inductor $L$ in series, to induce the tunable inter-cell photon hopping. (B) The resonant and detuning couplings of inter-cell spin states. Since the alternate A- and B- type unit cells arrangement, two sets of driving, $J_{\text{AB}}$ and $J_{\text{BA}}$, have to be adopted to ensure the translation symmetry in the rotating frame defined by $U$. (C) The levels and designed hopping of the polariton lattice in the rotating frame, where the A-type and B-type unit cells can be treated as the same, so that the proposed circuit simulates a 1D spin-1/2 tight-binding lattice model. (D) The Rabi oscillation of two-unit-cell system to justify the treatment of the proposed inter-cell coupling. The considered transition $|\uparrow \rangle_{\text{A}}\leftrightarrow |\downarrow \rangle_{\text{B}}$ is of the worst meeting the RWA requirement among the 4 possible transitions, thus the fidelity obtained is the least one, but it still reaches a very high value of $0.9979$ in the third Rabi cycle. All the numerical simulations are based on the Hamiltonian in Eq. (\[H\_JC\]) without RWA.[]{data-label="setup"}](figure1){width="6cm"}
Results
=======
Janeys-Cummings lattice
-----------------------
The method for implementing a one-dimensional (1D) JC lattice in superconducting quantum circuits [@SC3] is as follow. As shown in Fig. 1A, every unit cell consists of a TLR resonantly coupled with a transmon, forming a JC model [@Nori-rew-Simu2-JC]. The neighbouring TLRs are connected by a combination of a SQUID and a small inductor $L$ in series, which can actually be regarded as the counterpart of a semitransparent mirror in the cavity QED system, allowing photons to hop across neighbouring cavities (see Methods). As a result, setting $\hbar= 1$ hereafter, the system Hamiltonian of this JC lattice is $$\label{H_JC}
H_{\text{JC}}=\sum_{l=1}^{N}h_l+\sum_{l=1}^{N-1}J_{l}(t)\left(\hat{a}_l^\dagger \hat{a}_{l+1}+\text{h.c.}\right)+H_{\text{c}},$$ where $N$ is the number of the unit cells; $h_l= \omega_l (\sigma_l^+\sigma_l^- + \hat{a}_l^\dagger \hat{a}_l) + g_l\left(\sigma_l^+ \hat{a}_l + \sigma_l^- \hat{a}_l^\dagger\right)$ is the JC type interacting Hamiltonian in $l$th unit cell. The condition $g_l \ll \omega_l$ has to be met for justifying the JC coupling. $\sigma_l^{+}=|\text{e}\rangle\langle \text{g}|$ and $\sigma_l^{-}=|\text{g} \rangle\langle \text{e}|$ are the raising and lowering operators of the $l$th transmon qubits. $\hat{a}_l$ and $\hat{a}_l^\dagger$ are the annihilation and creation operators of the photon in $l$th TLR. And $J_{l}(t)$ is the inter-TLR hopping strengths between $l$th and $(l+1)$th unit cells. Different from optical cavities, the time-dependence of $J_{l}(t)$ here can be induced by adding a time-varying external magnetic flux threading through the SQUIDs (see Methods). $H_{\text{c}}=\sum_{l=1}^{N}g_l\left(\sigma_l^+ \hat{a}_l^\dagger + \sigma_l \hat{a}_l \right)+\sum_{l=1}^{N-1}J_{l}(t)\left(\hat{a}_l^\dagger \hat{a}_{l+1}^\dagger+\text{h.c.}\right)$ is the counter-rotating term that can be neglected by using the rotating wave approximation (RWA).
The lowest three eigenstates of the JC Hamiltonian $h_l$ for the $l$th unit are $|0\text{g}\rangle_l $, $|$$\uparrow\rangle_l =\left(|0\text{e}\rangle_l + |1\text{g}\rangle_l \right)/\sqrt{2}$ and $|$$\downarrow \rangle_l =\left(|0\text{e}\rangle_l - |1\text{g}\rangle_l \right)/\sqrt{2}$, where $|n\text{g}\rangle_l$ and $|n\text{e}\rangle_l \; (n=0, 1, 2, \cdots)$ are the states containing $n$ photons while the transmon is at the ground and excited state, respectively. And their eigen-energies are $E_{l,0g}=0$ and $E_{l,\uparrow(\downarrow)}= \omega_l +(-) g_l$. Here, we exploit the two single-excitation eigenstates $|\uparrow \rangle_l$ and $|\downarrow \rangle_l$ to simulate the effective electronic spin-up and spin-down state. As each of the two states consists of half “photon" and half “atom", they are regarded as a whole and was termed as “polariton”.
Polaritonic spin-orbit coupling
-------------------------------
We proceed to show that a spin-1/2 chain model with tunable Zeeman field and SOC can be simulated with the JC lattice by only adjusting the pulse shape of the coupling strengths $J_{l}(t)$ between neighbouring TLRs. Firstly, since each cell contains two pseudo-spin states, there are totally four inter-cell neighbouring hopping. In order to control each hopping separately, selective frequency addressing is employed (see Methods), i.e., we assign each of the four hopping with its unique hopping frequency. To achieve this, we adopt two sets of unit cells, A-type and B-type, which are different in the sense that they have different eigen-frequencies and coupling strengths of the JC model. Then we arrange them in an alternate way, as shown in Fig. 1A. Setting started with an A-type one, when $l$ is odd (even), $\omega_l=\omega_{\text{A}}(\omega_{\text{B}})$ and $g_l=g_{\text{A}}(g_{\text{B}})$. Then, based on the current experimental reaches [@SC3], we set $\omega_{\text{A}}/2\pi= 6$ GHz, $\omega_{\text{B}}/2\pi= 5.65$ GHz, $g_{\text{A}}/2\pi=300$ MHz, and $g_{\text{B}}/2\pi=270$ MHz. With this, the energy intervals of the four hopping are $\{|E_{l,\alpha}-E_{l+1,\alpha'}|/2\pi\}=\{ 220, 320, 380, 920\}$ MHz with $\alpha, \alpha' \in \{\uparrow, \downarrow\}$. The differences between every two of them are no less than 20 times of the effective hopping strength $t_0/2\pi=3$ MHz, thus they can be selective addressed in frequency. Correspondingly, the $J_l(t)$ contains four tunes and can be written as $$\label{coupling2}
J_{l}(t)=\sum_{\alpha, \alpha'} 4t_{0,l\alpha\alpha'} \cos\left(\omega_{l\alpha\alpha'}^{\text{d}} t+ s_{l\alpha\alpha'} \varphi_{l\alpha\alpha'}\right),$$ where $$s_{l\alpha\alpha'}= \sgn(E_{l,\alpha}-E_{l+1,\alpha'})
\label{sign1}$$ is the sign of the hopping phase, $4t_{0,l\alpha\alpha'},\; \omega_{l\alpha\alpha'}^{\text{d}}$ and $s_{l\alpha\alpha'} \varphi_{l\alpha\alpha'}$ are the amplitudes, frequencies and phases, corresponding to the hopping $|\alpha\rangle_l \rightarrow |\alpha'\rangle_{l+1}$, respectively. Note that due to the alternate arrangement, based on the definition above, the $J_l(t)$s take only two different forms – when $l$ is odd (even), $J_l=J_{\text{AB}} (J_{\text{BA}})$ where the subscript “AB" (“BA") refers to that the A-type (B-type) unit cell is on the left. Experimentally, this time-dependent coupling strength $J_l(t)$ can be realized by adding external magnetic fluxes with both dc and ac components threading through the SQUIDs (see Methods). In this way, selective hopping can be induced, i.e., only when a frequency of the driving flux matches a particular hopping energy interval, that hopping can be triggered, otherwise it will not take into effect. Meanwhile, both the strengths and phases of the hopping can be controlled by the amplitudes and phases of the ac magnetic fluxes. However, only controlling these two is not enough for realizing the topological states, we still need to induce and adjust the spin splitting. Then we theoretically find out, and numerically prove that this spin splitting can be induced by just adding a detuning to the spin-flipped transition tunes. Concretely, while the spin-preserved transition frequencies are set as $\omega_{l\alpha\alpha}^{\text{d}}= |E_{l,\alpha}-E_{l+1,\alpha}|$, we set the spin-flipped transition frequencies with a detuning $2m$ as $\omega_{l\alpha\alpha'}^{\text{d}}= |E_{l,\alpha}- E_{l+1,\alpha'}| -2m$, where $\alpha\ne \alpha'$, as shown in Fig. 1B. Thus, in the rotating frame (explained later), there will be a spin splitting $m$ for each cell, as shown in Fig. 1C.
We now show how the time-dependent coupling strength in Eq. (\[coupling2\]) can induce a designable spin transition process in a certain rotating frame. First, we map the Hamiltonian in Eq. (\[H\_JC\]) into the single excitation direct product subspace span$\{|0\text{g}, \cdots , 0\text{g}, \underset{l\text{th}}{\alpha}, 0\text{g}, \cdots, 0\text{g} \rangle \}_{l=1,\cdots,N;\, \alpha =\uparrow, \downarrow}$. Hereafter, when there is no ambiguity, we use $|\alpha \rangle_l$ to denote $|0\text{g}, \cdots, 0\text{g}, \underset{l\text{th}}{\alpha}, 0\text{g}, \cdots, 0\text{g}\rangle$, and $|G\rangle$ to denote $|0\text{g}, \cdots, 0\text{g} \rangle$. Then, we define a rotating frame by a unitary operator $U=\exp \{-\text{i}\left[\sum_{l} h_l-m(|\right.$$\uparrow$$\rangle_l\langle $$\uparrow$$ |- |$$\downarrow$$\rangle_l\langle $$\downarrow$$\left. | )\right] t \}$, which leads the Hamiltonian in Eq. (\[H\_JC\]) to $H'_{\text{JC}}=U^{\dag}H_{\text{JC}}U+\text{i}\dot{U}^{\dag}U$. Neglecting the fast rotating terms (see Methods), one obtains $$\label{eq.simu}
H'_{\text{JC}}=\sum_{l}^{N} m \bm{S^z}_{l}
+\sum_{l=1}^{N-1} \sum_{\alpha, \alpha'} \left(t_{0,l\alpha\alpha'} \text{e}^{\text{i}\varphi_{l\alpha\alpha'}}
\hat{c}^\dagger_{l,\alpha} \hat{c}_{l+1,\alpha'} +\text{h.c.}\right),$$ where $\bm{S^z}_{l}=|$$ \uparrow$$ \rangle_l\langle$$\uparrow$$|- |$$\downarrow$$\rangle_l\langle $$ \downarrow $$ |$, $\,t_{0,l\alpha\alpha'}$ is the effective coupling strength and $\hat{c}^\dagger_{l,\alpha}=|\alpha\rangle_l \, \langle G | $ is the creation operator for polariton with “spin” $\alpha$ in $l$th unit cell. As a result, the Hamiltonian (\[eq.simu\]) represents a general 1D spin-1/2 tight-binding lattice model with $m$, $t_{0,l\alpha\alpha'}$ and $\varphi_{l\alpha\alpha'}$ being the equivalent Zeeman energy, hopping strength and hopping phase, respectively. These three variables can all be experimentally tuned in wide ranges by the frequencies, amplitudes and phases of the external ac magnetic fluxes. Notably, the effective on-site potential $m$ can be tuned as either positive or negative depending on the detuning direction.
Meanwhile, it is worth noticing that although we have introduced two kinds of unit cells (A-type and B-type) in the laboratory frame, by adjusting the coupling parameters $J_{\text{AB}}$ and $J_{\text{BA}}$, the translation symmetry in the rotating frame is still preserved. In other words, the smallest repeating unit in the rotating frame contains only one unit cell as shown in Fig. 1C. By now, the adjustable spin-preserved tunneling ($\alpha=\alpha'$) and SOC terms ($\alpha \ne \alpha'$) can both be induced, hence our superconducting quantum circuit setup can naturally be used to simulate a tunable SOC topological polariton insulator.
In order to justify the individual frequency addressing of the inter-cell transitions, we numerically simulate the dynamics of a system containing only two unit cells, with an A-type one on the left and a B-type one on the right. We test every hopping of the four transitions $\{|\alpha\rangle_{\text{A}} \leftrightarrow |\alpha'\rangle_{\text{B}}\}$ one by one, adding only one corresponding frequency $\omega_{1\alpha\alpha'}^{\text{d}}$ in $J_{\text{AB}}(t)$. As expected, when we pick out a resonant frequency $\omega_{1\alpha\alpha'}^{\text{d}}$ with $m=0$, there will be a Rabi oscillation between the two corresponding target states. One example of these Rabi oscillations was shown in Fig. 1D, which is of the least fidelity among the four. Even in this worst case, and in the third Rabi cycle, the fidelity still reaches a high value of $0.9979$, which justify the RWA. In addition, our numerical simulation also shows that, in the present of the unmatched driving, all the initial non-target states remain almost unchanged, thus justify our individual frequency addressing method.
Nodal-loop topological polaritons
---------------------------------
Our protocol provides a tunable platform using polaritons to study topological matters. Here, we take the nodal-loop semimetal as an application sample to demonstrate how to simulate a specific condensed matter model in our proposed setup. To fit our simulation setup, we reform the Hamiltonian of the original 3D nodal-loop model in Ref. [@NodalLoop1], without losing any topological properties. Firstly, we relabel the coordinates to set the hopping terms with SOC to be along the $x$ axis. Secondly, we consider the Fourier transformations along $y$ and $z$ directions with quasi-momenta $k_y$ and $k_z$ and treat them as system parameters. Then, according to Eq. (\[eq.simu\]), we can simulate this 3D nodal-loop model in our 1D system with the other two dimensions being the parametric dimensions. In this direct simulation, to engineer the four transitions between different paloriton states, we need four different tunes in the inter-cell coupling strength $J_{l}^{\text{nod}}(t)$.
The above implementation can further be simplified as following. We first make a unitary transformation to the original Hamiltonian so that $H_\text{{nod}}=V^\dag H_\text{{ori}} V$, where $V=\sum_{l=1}^{N} (-\text{i})^{l-1} \bm{I}_l$ with $\bm{I}_l=|$$\uparrow$$\rangle_l\langle $$\uparrow$$ |+ |$$\downarrow$$\rangle_l\langle $$\downarrow$$|$. The transformed 1D lattice Hamiltonian from Eq. (\[eq.simu\]), without losing any physical properties, is $$\begin{aligned}
\label{H_nodal}
H_\text{{nod}}=&\sum_{l=1}^{N} m'(k_y, k_z)\bm{S^z}_{l}
+\sum_{l=1}^{N-1}\sum_{\alpha, \alpha'} \left(\text{i}t'_{0}\hat{c}^\dagger_{l,\uparrow} \hat{c}_{l+1,\uparrow}\right.\\
&\left.-\text{i}t'_{0}\hat{c}^\dagger_{l,\downarrow} \hat{c}_{l+1,\downarrow}+\text{i}t'_{0}\hat{c}^\dagger_{l,\uparrow} \hat{c}_{l+1,\downarrow}
-\text{i}t'_{0}\hat{c}^\dagger_{l,\downarrow} \hat{c}_{l+1,\uparrow} +\text{h.c.}\right),\notag\end{aligned}$$ where $m'(k_y, k_z)=M+2d(\cos k_y+\cos k_z)$ with $M$ being the effective Zeeman energy and $d$ being the effective hopping energy along $y$ and $z$ directions. Then, we set the parameters of the JC model to be $ \omega_{\text{A}}= \omega_{\text{B}}=2\pi \times 6$ GHz, $g_{\text{A}}/2\pi=200$ MHz and $g_{\text{B}}/2\pi=100$ MHz. Correspondingly, the coupling strengths can be set to contain only two tunes as $$\label{J_nod}
J_{l}^{\text{nod}}(t)=4t_0\cos\left[\omega_{1}^{\text{d}} t+\frac{(-1)^{l+1}\pi}{2}\right]+ 4t_0\cos\left(\omega_{2}^{\text{d}} t+ \frac{\pi}{2}\right),$$ where $\omega_{1}^{\text{d}}/2\pi=100$ MHz and $\omega_{2}^{\text{d}}/2\pi=(300-2m)$ MHz. In this way, when transforming into the rotating frame of $U$ and applying the RWA, the Hamiltonian (\[H\_JC\]) of our JC-lattice system will takes the form of Eq. (\[H\_nodal\]), which accomplishes the quantum simulation of the topological nodal-loop semimetals (see Methods).
Experimentally, one can choose $t_0/2\pi=3$ MHz and set the detuning within the range of $m \in 2\pi \times [-20, 20] $ MHz, such that any value of the variable tunes $\omega_{l\alpha\alpha'}^d (\alpha \ne \alpha')$ still maintain a frequency difference no less than $20t_0$ away from other tunes, and thus the crosstalk caused by unwanted tunes will be negligible. Last, for testing the validity of our theoretical protocol, numerical simulations will be given in the Experimental detection section after a brief introduction of the characteristics of the topological nodal-loop polaritons.
{width="11.5cm"}
Characteristics of the topological polaritons
---------------------------------------------
To investigate the bulk characteristic of the topological nodal-loop polaritons, we first consider the periodic boundary condition to obtain the Hamiltonian in Eq. (\[H\_nodal\]) in the momentum space as $$H=b_y \bm{S^y}+ b_z\bm{S^z},$$ where $b_y=-2t'_0 \cos k_x$, $b_z=2t'_0 \sin k_x+m'(k_y, k_z)$ and $\bm{S^{y}}=\text{i}( |\downarrow\rangle\langle\uparrow| -|\uparrow\rangle\langle\downarrow|)$. This system has two energy bands $$E=\pm\sqrt{b_y^{2}+b_z^2},
\label{energy}$$ which will touch when $E=0$. The touching points form closed lines, the so-called nodal-loops, in momentum space as shown in Fig. 2A. These two loops appear in the $k_x=\pi/2$ and/or $k_x=-\pi/2$ planes. By fixing $k_x=\pm\pi/2$, we plot these two energy bands in the $k_y$-$k_z$ space as shown in Fig. 2B and 2C, where one can see that the touching is right along the nodal-loops.
The topological index characterizing each nodal loop is a winding number defined as $$\begin{split}
\nu(k_y, k_z) &=\frac{1}{2\pi }\int_{-\pi}^{\pi} dk_x \; \mathbf{v}\times \partial _{k_{x}}\mathbf{v}\\
&=\frac{1}{2}\big[\sgn\left( m'+2t'_0\right)
-\sgn\left( m'-2t'_0\right) \big],
\end{split}
\label{gamma}$$ where $\mathbf{v}=(v_{y},v_{z})=(b_{y},b_{z})/\sqrt{b_{y}^{2}+b_{z}^{2}}$. This shows that the quantized winding number is either $1$ or $0$, corresponding to the cases whether a nodal loop is enclosing the straight line, which is along $k_x$ direction of the fixed $(k_x, k_y)$ point, or not [@NodalLoop1; @NodalLoop2]. Hence, the nodal-loops divide the $k_y$-$k_z$ space into regions with different winding numbers, as shown in Fig. 2D. For all straight lines along $k_x$ inside the nodal loop, each of them can be regarded as being corresponding to a topological 1D gapped subsystem with winding number $1$.
Two striking topological characters of nodal-loop semimetal are the zero-energy modes inside the energy gap and their corresponding edge states. We take a slice of $k_y=0$, indicated by the green line in Fig. 2D, as an example to plot the energy spectrum with various $k_z$ for a finite chain with $N=20$, as shown in Fig. 2E. The numerical result shows that there are two mid-gap degenerated zero-energy modes appear in the range of $\nu=1$ (the red area in Fig. 2D). The quantum states corresponding to the two mid-gap energies are edge states localized in the left and right end of the lattice, respectively. When $N$ is large enough for ignoring the finite-size effects, their wave-functions can be expressed as
\[edge\_state\] $$\psi_{\text{L}} =\sum_{l=1}^{N} \text{i}^{l-1} A\text{e}^{-\lambda (l-1)} \left(|\uparrow\rangle_l+|\downarrow\rangle_l\right),$$ $$\psi_{\text{R}} =\sum_{l=1}^{N} \text{i}^{l-1} A\text{e}^{-\lambda (N-l)}\left(|\uparrow\rangle_l-|\downarrow\rangle_l\right),$$
where $A=\sqrt{(1-q^2)/2(1-q^{2N})}$ and $\lambda=\ln(1/q)$ with $q=\tan(m'\pi/8t'_0)$. The phase factor $\text{i}^{l-1}$ stems from the unitary transformation $V$ to get the nodal-loop Hamiltonian in Eq. (\[H\_nodal\]) that simplified our simulation, from the original Hamiltonian in Eq. (\[eq.simu\]). These analytical results of wave-functions are in very good agreement with the numerical results for $N=20$, as shown in Fig. 2F, thus justifies that we can use the JC lattice of experimentally capable size to simulate the topological features.
There are several phases in our simulated Hamiltonian where the phase transition is indicated by the emerging or vanishing of the nodal loops. Inferring from Eq. (\[energy\]), the critical conditions are obtained as $k_x=-\pi/2$ and $k_x=\pi/2$, which are corresponding to one nodal loop in each of the two regions of $-2t'_0-4d<M<-2t'_0+4d$ and $2t'_0-4d<M<2t'_0+4d$ with $d>0$ in the $M$-$d$ plane. Therefore, in the area where the two regions overlap, there are two nodal loops. But in the area outside these two regions, there is no nodal loop so that the whole Brillouin zone will be in a purely trivial or nontrivial phase. Consequently, there are totally five different phases of different winding number configurations $\nu(k_y, k_z)$ in the $M$-$d$ plane, as shown in Fig. 3. According to the chosen parameters, $t_0$ and $m$, the area $\{M,d |-6t'_0+4d<M<6t'_0-4d \}$ will include all the the five phases in the phase diagram of Fig. 3, i.e., all the possible phases can all be simulated in our proposed system.
![**Phase diagram of nodal-loop semimetal bands with the corresponding winding number configurations.** Each color denotes a different phase. The dark blue region is of a trivial gapped phase without nodal loops in momentum space; the yellow and green region is of only one nodal loop in the $k_y$-$k_z$ plane; the light blue region is of two nodal loops in the $k_y$-$k_z$ plane; the red region is of a non-trivial gapped phase without nodal loops but with winding number $\nu=1$ in the whole $k_y$-$k_z$ plane. The three insets are the configuration of winding number in the $k_y$-$k_z$ plane (the red regions are of $\nu=1$, while blue regions are of $\nu=0$), corresponding to points A (-2.5, 0.5), B (0, 1) and C (2.5, 0.5), respectively.[]{data-label="phase"}](figure3){width="7.8cm"}
Experimental detection methods
------------------------------
*Polaritonic topological edge state detection.* According to Eq. (\[edge\_state\]), or as shown in Fig. 2F, the polariton in the left or right edge state is maximally distributed in the leftmost or rightmost JC lattice site. Their internal spins are in the superposition states $\left(|\uparrow\rangle_l+|\downarrow\rangle_l \right)/\sqrt{2}$ and $\left(|\uparrow\rangle_l-|\downarrow\rangle_l \right)/\sqrt{2}$, respectively. Therefore one can find that, in the beginning, the left and right polaritonic edge states only have qubit excitation and photon components, respectively. Taking the detection of left edge state as an example, initially, the polariton in the leftmost JC lattice site is prepared into $|0\text{e}\rangle_1$, i.e., the leftmost qubit (resonator) has been prepared in the excited (vacuum) state. The qubits and resonators in the other sites are prepared into the ground and vacuum states which means the initial systematic state is $|\psi(t=0)\rangle=|0\text{e}\rangle_1|0\text{g}\rangle_2\cdots|0\text{g}\rangle_N$. After that, we let the above initial state evolve for a time about 0.5 $\mu$s. If the JC lattice is in the topological nontrivial phase supporting the left edge states, the final density distribution of the polaritons will maximally populate the leftmost site. The reason is that the initial state $|\psi(t=0)\rangle$ has a large overlap with the left edge state. It will evolve mainly via the edge state wave packet and maximally localized in the leftmost site. While if the system is in the topological trivial phase and has no edge states, the initial state will be a superposition of different bulk sates. The final density distribution will not have maximal distribution in the leftmost site. Similarly, one also can prepare the JC lattice into $|\psi(t=0)\rangle=|0\text{g}\rangle_1\cdots|0\text{g}\rangle_{N-1}|1\text{g}\rangle_N$ and detect [@Lei-measu-wanghh] the right polaritonic topological edge state based on observing its time evolution.
In Fig. 4A and 4B, we have numerically calculated the time evolution of the polaritonic density when the JC lattice is in the topological trivial and nontrivial nodal-loop semimetal phases, respectively. For the trivial case, the wave packet has a ballistic spread versus time, which is a typical feature of bulk Bloch state. It shows that there is no edge state localization and the system is in topological trivial phase. For the nontrivial case, the density of the polaritons will always maximally localize in the leftmost JC lattice site, which indicates the existence of left topological edge states demonstrating that the system is in topological nontrivial phase. The time evolution of the qubit excitation and the photon population for the topological nontrivial case are also numerically calculated in Fig. 4C and 4D, which shows that the localized qubit excitation and the photon in the leftmost site have a Rabi-like oscillation feature inherited from the JC model.
![**Dynamical detection of polaritonic topological edge states.** Time evolution of polaritonic density distribution $\langle \sigma^+\sigma^- + \hat{a}^\dagger \hat{a}\rangle$ when the JC lattice is in (A) topological trivial phase of $k_z=0.3\pi$ and (B) topological nontrivial phase of $k_z=0.7\pi$ (see Fig. 2E). Time evolution of (C) qubit excitation distribution $\langle \sigma_l^+\sigma_l^- \rangle$ and (D) photon distribution $\langle \hat{a}_l^\dagger \hat{a}_l\rangle$ for the topological nontrivial case. All the numerical simulations are based on the Hamiltonian in Eq. (\[H\_JC\]) without RWA. Other parameters are the same as that of in Fig. 2E.[]{data-label="dynamic"}](figure4){width="\columnwidth"}
*Polaritonic topological invariant detection*. Another important hallmark for topological nontrivial nodal-loop polaritonic semimetal phase is the nontrivial polaritonic topological winding number. Here we show that such polaritonic topological invariant also can be dynamically detected. Our method is based on a previous work which shows that the topological winding number rotted in the momentum space can be detected through measuring the dynamical chiral center in the real space [@Mei]. The chiral operator for our topological polaritonic model is $\bm{S^x}_l=|$$\uparrow$$\rangle_l\langle$$\downarrow$$|+ |$$\downarrow$$\rangle_l\langle$$\uparrow$$|$. Then the chiral center operator for the JC lattice is defined as $\hat{P}_{\text{d}}=\sum_{l=1}^{N}l \bm{S^x}_l$. The polaritonic topological winding number can be related with the time-averaged dynamical chiral center associated with the single-polariton dynamics, i.e., $$\nu ={\lim_{T\rightarrow \infty }}\frac{2}{T}\int_{0}^{T}\text{d}t\, \langle \psi_{\text{c}}(t)| \hat{P}_{\text{d}} |\psi_{\text{c}}(t)\rangle,
\label{pdt2}$$ where $T$ is the evolution time, $|\psi_{\text{c}}(t)\rangle =\exp(-\text{i}H_{\text{nod}}t)|\psi_{\text{c}}(0)\rangle$ is the time evolution of the initial single-polariton state $|\psi_{\text{c}}(0)\rangle=|0\text{g}\rangle_1\cdots |\uparrow\rangle_{\lceil N/2 \rceil} \cdots|0\text{g}\rangle_N$, where one of the middle JC lattice site has been put one polariton in, with its spin prepared in the state $|\uparrow\rangle$.
In Fig. 5A and 5B, we have numerically calculated the dynamical chiral center $\bar{P}_{\text{d}}(t)=\langle\psi_{\text{c}}(t)|\hat{P}_{\text{d}}|\psi_{\text{c}}(t)\rangle$ for topological trivial and nontrivial cases, respectively. According to Eq. (\[pdt2\]), one can find that the topological winding number is equal to twice the oscillation center of $\bar{P}_{\text{d}}(t)$. As shown in Fig. 5A, $\bar{P}_{\text{d}}(t)$ oscillates around the average value 0, which gives the polaritonic topological winding number $\nu=0$. In contrast, the result for the topological nontrivial case in Fig. 5B shows that $\bar{P}_{\text{d}}(t)$ oscillates around 0.5, which yields the polaritonic topological winding number $\nu=1$. Experimentally, the states of qubits and resonators can be measured with fidelity higher than $0.99$ in superconducting circuits. In our case, one only need to measure the qubit excitation and photon populations for getting their imbalance and deriving the chiral center, without requiring full quantum state tomography. In this way, the topological winding number can be easily and unambiguously detected based on monitoring single-polariton quantum dynamics in a JC lattice.
![**Dynamical detection of polaritonic topological invariants.** Time evolution of the chiral center $\bar{P}_{\text{d}}$ when the JC lattice is in (A) the topological trivial phase of $k_z=0.3\pi$ and (B) non-trivial phase of $k_z=0.7\pi$ (see Fig. 2E). The red dashed line denotes the oscillation center. All the numerical simulations are based on the Hamiltonian in Eq. (\[H\_JC\]) without RWA. Other parameters are chosen the same as that of in Fig. 4.[]{data-label="chiral"}](figure5){width="\columnwidth"}
Discussion {#discussion .unnumbered}
==========
We further investigate how our construction and detection methods are influenced by the quantum decoherence effects. The Lindblad master equation is adopted to take three main decoherence factors including the losses of the photon and the decay and dephasing of the transmon into account. The Lindblad master equation of our system can be written as $$\dot {\rho }=-{\text{i}}[H_{\text{JC}},\rho ]+\sum_{l=1}^{N}\sum _{i=1}^{3}\gamma\left(\Gamma_{l,i}\,\rho \Gamma_{l,i}^{\dagger }-{\frac {1}{2}}\left\{\Gamma_{l,i}^{\dagger }\Gamma_{l,i},\rho \right\}\right),$$ where $\rho$ is the density operator of the whole system, $\gamma$ is the decay rate or noise strength which are set to be the same here, $\Gamma_{l,1}=a_{l},\;\Gamma_{l,2}=\sigma^-_{l}$ and $\Gamma_{l,3}=\sigma^z_{l}$ are the photon-loss, transmon-loss and the transmon-dephasing operators in the $l$th lattice, respectively. In Fig. 6A, we plot the edge-site population $P_1(t)=\text{tr}\left[\rho(t)\left(a_1^\dag a_1+\sigma_1^+\sigma_1^-\right)\right]$ after 0.5 $\mu$s and the oscillation center $\nu/2$ of the trivial and nontrivial cases for different decay rates. It shows that the edge state population and the chiral center smoothly decrease when the decay rate increase. However, our detection method can tolerate the decay rate up to the order of $2\pi\times 100$ kHz, while the typical decay rates are only $2\pi\times 5$ kHz.
![**The influences of the decoherence and the lattice site on the detection of the topological effects.** (A) The edge-site population $P_1(t)=\text{tr}\left[\rho(t)\left(a_1^\dag a_1+\sigma_1^+\sigma_1^-\right)\right]$ at 0.5 $\mu$s and the oscillation center $\nu/2$ of the topological trivial ($k_z=0.3\pi$) and nontrivial ($k_z=0.7\pi$) cases for different decay rate $\gamma$. (B) The oscillation center of the topological trivial ($k_z=0.3\pi$) and nontrivial ($k_z=0.7\pi$) cases varying over the number of lattice sites. The decay rates are all set to be $2\pi\times 5$ kHz. All the numerical simulations are based on the Hamiltonian in Eq. (\[H\_JC\]) without RWA. Other parameters are chosen the same as that of in Fig. 4.[]{data-label="influences"}](figure6){width="\columnwidth"}
We also investigate the minimum sites that are needed for the experimental detecting the oscillation-center. In the presence of the decay rates of $2\pi\times 5$ kHz, we plot the oscillation center for lattice with different number of sites in both the trivial and nontrivial cases, as shown in Fig. 6B. It shows that the four sites case corresponding to an oscillation center value about 0.40, is big enough for distinguishing the topologically trivial and nontrivial cases with current state-of-the-art technologies.
To conclude, we have introduced the concept of topological states into the JC lattice which is one of the most important building blocks in quantum optics and quantum information processing. We have studied the topological structure of light-matter interaction and shown that SOC physics and topological polaritons can be pursued in the JC lattice. Different from synthetic topological states in ultracold atomic, photonic and acoustic systems, topological polaritons are topological superposition states of photons and qubits. Tunable synthetic polaritonic SOC is induced by engineering the JC lattice couplings, which provides the basic ingredient for realizing spinful topological states of matter. We have also provided a method using single-particle quantum dynamics in real space to directly observe the polaritonic topological edge states and topological invariants.
Our work has a broad generalization and opens the door for exploring spinful topological states of matter using polaritons in JC lattice system. (i) In addition to mimicking spin-1/2, polariton states in multiple-excitation manifold have an extended spin degree of freedom and can also be used to mimic larger spin. It is challenging to realize SOC for the larger spin case in solid state materials and ultracold atoms. However, the method proposed in our work can be generalized to realize synthetic SOC for large-spin polaritons. This allows us to explore a large variety of topological states, including triple point topological states of matter [@TPS1]; (ii) Polariton states, the eigenstates of a JC model, can be referred as a synthetic dimension where each pair of states mimic a spin in a spin-lattice model and the coupling between polariton states provides the hopping in the synthetic dimension. With such synthetic dimension, high-dimensional topological states of matter can be explored in a low-dimensional JC lattice, including topological states beyond three dimensions [@4DTP1; @4DTP2]; (iii) Polaritons have tunable strong nonlinear interaction, which allows us to study polaritonic fractional topological states of matter [@FTP1]; (iv) Besides superconducting circuit, JC lattices can also be realized in many quantum optical systems, including trapped ions [@Ion], cavity quantum electrodynamics [@CQED], nanoscopic lattice [@Nano], optomechanical systems [@Opto] and so on.
Methods
=======
SQUID induced time-dependent photon hopping
-------------------------------------------
We here present how to induce the time-dependent photon hopping strength between two TLRs. As shown in Fig. 1A, neighbouring TLRs are connected by a common SQUID and an inductor $L$ in series then to the ground. The SQUID actually serves as a single Josephson junction (JJ) but with effective Josephson inductance tunable by the external flux. Concretely, by applying external magnetic flux $\Phi_{\text{ext}}=\Phi_{\text{dc}}+\Phi_{\text{ac}}$ threading through the SQUID, when $\Phi_{\text{ac}} \ll \phi_0 $ with $\phi_{0}$ being the reduced flux quanta, the effective inductance of the SQUID reads [@Lei-induc3-prl] $$\label{L_S}
L_S(\Phi_{\text{ext}})=\frac{\phi_{0}}{2 I_{c}\cos \left(\Phi_{\text{ext}}/2\phi_{0} \right)},$$ where $I_{\text{c}}$ is the shared critical current of the two JJ in each SQUID. On the other hand, comparing to the inductance of the TLRs, both the SQUID and the inductor $L$ have far smaller inductances, thus there is a voltage node but a current peak at both ends of each TLR. For these boundary conditions, after the conventional quantization of the TLRs [@Lei-induc3-prl; @Nori-rew-Simu2-JC], the flux density and the charge density wave-function of the lowest-energy mode in $l$th TLR can be expressed as
$$\hat{\phi}_l (x_l,t)=\frac{\sqrt{ \omega_l L_l} }{d_l} \cos\left(\frac{\pi}{d_l} x_l\right) \left[\hat{a}_l^\dagger (t)+ \hat{a}_l(t)\right],$$
$$\hat{q}_l(x_l,t)=\text{i}\frac{\sqrt{ \omega_l L_l} }{d_l} \sin\left(\frac{\pi}{d_l} x_l\right) \left[\hat{a}_l^\dagger(t)-\hat{a}_l(t) \right],$$
where $\omega_l=\pi/\sqrt{L_lC_l}$ is the frequency of the photon with $L_l$ and $C_l$ being the inductance and capacitance, respectively. $d_l$ is the length and $x_l$ is the coordinate of the $l$th TLR. Meanwhile, because of the relatively low-impedances of the SQUID and the inductor, the currents from the ends of every TLR will flow directly through them to the ground, without crossing to their neighbouring TLRs. Hence, the interaction Hamiltonian between the $l$th and $(l+1)$th TLR is just the summation energy of the SQUID and the inductor $L$ $$\begin{aligned}
\label{H_S}
H_{\text{int}}^l&=&\frac{1}{2}(L_S+L)(I_l^{\text{ri}}+I_{l+1}^{\text{le}})^2 \notag\\
&=& \sum_{j=l}^{l+1} \frac{\omega_j}{2L_j} \left( L_S+L \right) \left(\hat{a}_j^\dagger+\hat{a}_j\right)^2\\
&& - \sqrt{\frac{ \omega_l \omega_{l+1}}{L_l L_{l+1}}} \left( L_S+L \right) \left(\hat{a}_l^\dagger+\hat{a}_l\right)\left(\hat{a}_{l+1}^\dagger+\hat{a}_{l+1}\right),\notag
\end{aligned}$$ where $I_l^{\text{ri(le)}}=\hat{\phi}_l (x_l,t) d_l /L_l|_{x_l=d_l(0)}$ is the current of right(left)-end of the $l$th TLR. Moreover, If we set
$$\Phi_{\text{ac}}=2\phi_{0}\arccos\frac{-1}{1+\sum_{j}^{n'} \Omega_j \left[\cos\left(\omega_j^d t+\varphi_j\right)+1\right]} ,$$
$$t_{0,j}=\frac{ \phi_{0} \Omega_j}{8 I_{\text{c}} } \sqrt{ \frac{\omega_l\omega_{l+1}} {L_lL_{l+1}} },$$
$$L =\frac{\phi_{0}(\sum_{j}^{n'} \Omega_j+1)}{2 I_{\text{c}}},$$
where $n'$ is the number of the tunes in $\Phi_{\text{ac}}$, $\Phi_{\text{dc}}=4\pi n''\phi_0$ where $n''$ is an arbitrary positive integer, and choose resonant or detuned frequencies of $\omega_j$, after the RWA, we obtain $$\begin{aligned}
\label{HS3}
H_{\text{int}}^l= 4\sum_{j}^{n'} t_{0,j} \cos\left(\omega_j^d t+\varphi_j\right) \left(\hat{a}_l^\dagger\hat{a}_{l+1}+\text{h.c.}\right),
\end{aligned}$$ which right meets the form of Eqs. (\[H\_JC\]) and (\[coupling2\]). This equation can be interpreted as describing the photons hopping between neighbouring unit cells, which means that the SQUID-L combination can actually serve as a counterpart of the semitransparent mirror in the optical cavity system.
Frequency addressing control
----------------------------
We now show how the selective control of individual hopping in the JC lattice can be achieved by adjusting $J_l(t)$ in Eq. (\[H\_JC\]) via the ac flux. We first map the Hamiltonian in Eq. (\[H\_JC\]) into the single excitation subspace span$\{|\alpha\rangle_l\}_{\l=1, 2, \cdots, N; \; \alpha =\uparrow, \downarrow}$ and get $$H_{\text{JC}}=\sum_{l,\alpha} E_{l,\alpha}\hat{c}_{l\alpha}^\dag \hat{c}_{l\alpha} +\frac{1}{2} \sum_{l=1}^{N-1} \sum_{\alpha, \alpha'} J_l(t) \hat{c}_{l,\alpha}^\dag \hat{c}_{l+1,\alpha'} + \text{h.c.},$$ where $\hat{c}_{l,\alpha}^\dag=|\alpha\rangle_l \langle G |$. Then in every $J_l(t)$, we add four tunes, corresponding to the four inter-cell hopping, cf. Eq. (\[HS3\]), each of which contains its independent tunable amplitude, frequency and phase as $$J_l(t)=\sum_{\alpha,\alpha'} 4t_{0,l\alpha\alpha'} \cos\left(\omega_{l\alpha\alpha'}^d t+ s_{l\alpha\alpha'}\varphi_{l\alpha\alpha'}\right),$$ where $s_{l\alpha\alpha'}$ is defined in Eq. (\[sign1\]) being the sign of each phase and the frequencies are set as $$\omega_{l\alpha\alpha'}^d=
\begin{cases}
|E_{l,\alpha}-E_{l+1,\alpha'}| \qquad \qquad(\alpha=\alpha'),\\
|E_{l,\alpha}-E_{l+1,\alpha'}|-2m \quad \;\,(\alpha \ne \alpha'),
\end{cases}$$ where $2m$ is a detuning. By now, the form of $J_l(t)$ is determined leaving $m$, $t_{0,l\alpha\alpha'}$ and $\varphi_{l\alpha\alpha'}$ to be chosen arbitrarily depending on the topological insulator model one simulates. Eventually, the target topological insulator model is in the rotating frame transformed by $$U=\text{e}^{-\text{i}\left[\sum_{l} h_l-m(|\uparrow\rangle_l\langle \uparrow|- |\downarrow\rangle_l\langle \downarrow | )\right] t},$$ After the picture transformation $H'_{\text{JC}}=U^{\dag}H_{\text{JC}}U+\text{i}\dot{U}^{\dag}U$, one gets
$$\begin{aligned}
&H'_{\text{JC}}=\sum_{l=1}^{N} m \bm{S^z}_l +\sum_{l=1}^{N-1} \sum_{\alpha,\alpha'} t_{0,l\alpha\alpha'}\notag \\
&\times \left[\text{e}^{\text{i}(\omega_{l\alpha\alpha'}\,t+s_{l\alpha\alpha'}\varphi_{l\alpha\alpha'})} \right. + \left.
\text{e}^{-\text{i}(\omega_{l\alpha\alpha'}\,t+s_{l\alpha\alpha'}\varphi_{l\alpha\alpha'})} \right] \notag \\
&\times \left[\text{e}^{\text{i}(E_{l+1,\uparrow}-E_{l,\uparrow})t}\hat{c}_{l,\uparrow} \hat{c}^\dag_{l+1,\uparrow}+ % &
\text{e}^{\text{i}(E_{l+1,\downarrow}-E_{l,\downarrow})t}\hat{c}_{l,\downarrow} \hat{c}^\dag_{l+1,\downarrow}\right.\notag \\
& \quad + \left. \text{e}^{\text{i}(E_{l+1,\uparrow}-E_{l,\downarrow}-2m)t}\hat{c}_{l,\downarrow} \hat{c}^\dag_{l+1,\uparrow}\right. \notag \\
&\quad \left. + \text{e}^{\text{i}(E_{l+1,\downarrow}-E_{l,\uparrow}+2m)t}\hat{c}_{l,\uparrow} \hat{c}^\dag_{l+1,\downarrow} \right]+\text{h.c.},\end{aligned}$$
where $\bm{S^z}_{l}=|$$ \uparrow$$ \rangle_l\langle$$\uparrow$$|- |$$\downarrow$$\rangle_l\langle $$ \downarrow $$ |$. After doing the multiplication in this equation, the four resonant terms and their Hermitian conjugates will be absent of time $t$ (frequency addressing). Meanwhile, if the conditions $\{ t_{0,l\alpha\alpha'}\}_{\alpha,\alpha'=\uparrow,\downarrow}\ll \{ \omega_{l\alpha\alpha'},\; \omega_{l\alpha\alpha'}- \omega_{l\beta\beta'} \}_{\alpha,\alpha',\beta,\beta' =\uparrow,\downarrow; \; \omega_{l\alpha\alpha'}\ne \omega_{l\beta\beta'} }$ are satisfied, all the other terms are fast rotating term that can be dropped within RWA. As a result, we obtain the tight-binding model with tunable SOC in Eq. (\[eq.simu\]).
Simplified method of implementation
-----------------------------------
We here explain why the four hopping terms in the nodal-loop semimetal modal can be induced by the coupling strength in Eq. (\[J\_nod\]) using only two tunes. Firstly, when we set parameters of the JC model to be $ \omega_{\text{A}}= \omega_{\text{B}}=2\pi \times 6$ GHz, $g_{\text{A}}/2\pi=200$ MHz and $g_{\text{B}}/2\pi=100$ MHz, the energy intervals of the four hopping overlap and reduce into two set of intervals, i.e., a spin-conserved hopping interval $$|E_{l,\uparrow}-E_{l+1,\uparrow}|=|E_{l,\downarrow}-E_{l+1,\downarrow}|=2\pi \times 100 \text{MHz},$$ and a spin-flipped hopping interval $$|E_{l,\uparrow}-E_{l+1,\downarrow}|= |E_{l,\downarrow}-E_{l+1,\uparrow}|= 2\pi \times 300 \text{MHz}.$$ Therefore, in this way, one tune in $J_{l}^{\text{nod}}(t)$ can induce two hopping while the two spin-conserved hopping still remain being controlled separated from the two spin-flipped hopping, thus they can be of different detuning.
Secondly, note that within this setting, according to the definition in Eq. (\[sign1\]), we have $s_{l\uparrow\uparrow}= -s_{l\downarrow\downarrow}$ and $s_{l\uparrow\downarrow}= -s_{l\downarrow\uparrow}$, thus both the two spin-conserved hopping terms and spin-flipped hopping terms can be induced by a same tune, but with opposite signs, i.e., $\text{i}$ and $-\text{i}$. This is possible because the RWA selects different terms. These opposite signs are ideal for realizing the wanted SOC in our protocol. Anyway, this is only possible when we do the unitary transformation $V$ to transform the hopping phase from the original $1$ and $-1$ into the pure imaginary numbers $\text{i}$ and $-\text{i}$. For the former case, one still has to use four tunes and induce them separately. Therefore, this transformation simplifies our simulation.
ACKNOWLEDGEMENTS
================
This work is supported by the NSFC (Grants No. 11874156, No. 11604103 and No. 11604392), the Key R&D Program of Guangdong province (Grant No. 2018B0303326001), the National Key R&D Program of China (Grant No. 2016YFA0301803 and No. 2017YFA0304203), the NSF of Guangdong Province (Grant No. 2016A030313436), the Startup Foundation of South China Normal University, the PCSIRT (Grant No. IRT\_17R70), the 1331KSC, and the 111 Project (Grant No. D18001).
[usrt]{}
Cummings, F. W. & Jaynes, E. T. Comparison of quantum and semiclassical radiation theories with application to the beam maser, *Proceedings of the IEEE* **51** 89-109 (1963).
Leibfried, D., Blatt, R., Monroe, C. & Wineland, D. Quantum dynamics of single trapped ions, *Rev. Mod. Phys.* **75**, 281-324 (2003).
Raimond, J. M., Brune, M. & Haroche, S. Manipulating quantum entanglement with atoms and photons in a cavity, *Rev. Mod. Phys.* **73**, 565-582 (2001).
Devoret, M. H. & Schoelkopf, R. J. Superconducting Circuits for Quantum Information: An Outlook, *Science* **339**, 1169-1174 (2013).
Xiang, Z. L., Ashhab, S., You, J. Q. & Nori, F. Hybrid quantum circuits: Superconducting circuits interacting with other quantum systems, *Rev. Mod. Phys.* **85**, 623-653 (2013).
Chang, D. E., Douglas, J. S., González-Tudela, A., Hung, C.-L. & Kimble, H. J. Quantum matter built from nanoscopic lattices of atoms and photons, *Rev. Mod. Phys.* **90**, 031002 (2018)
Aspelmeyer, M., Kippenberg, T. J. & Marquardt, F. Cavity optomechanics, *Rev. Mod. Phys.* **86**, 1391-1452 (2014).
Hartmann, M. J., Brandão, F. G. S. L. & Plenio, M. B. Quantum many-body phenomena in coupled cavity arrays, *Laser & Photonics Rev.* **2**, 527-556 (2008).
Tomadin, A. & Fazio, R. Many-body phenomena in QED-cavity arrays, *J. Opt. Soc. Am. B* **27**, A130-A136 (2010).
Buluta, I. & Nori, F. Quantum simulators, *Science* **326**, 108-111 (2009).
Houck, A. A., Türeci, H.E. & Koch, J. On-chip quantum simulation with superconducting circuits, *Nat. Phys.* **8**, 292-299 (2012).
Hartmann, M. J., Brandão, F. G. S. L. & Plenio, M. B. Strongly interacting polaritons in coupled arrays of cavities, *Nat. Phys.* **2**, 849-855 (2006).
Greentree, A. D., Tahan, C., Cole, J. H. & Hollenberg, L. C. L. Quantum phase transitions of light, *Nat. Phys.* **2**, 856-861 (2006).
Angelakis, D. G., Santos, M. F. & Bose, S. Photon-blockade-induced Mott transitions and XY spin models in coupled cavity arrays, *Phys. Rev. A* **76**, 031805(R) (2007).
Nunnenkamp, A., Koch, J. & Girvin, S. M. Synthetic gauge fields and homodyne transmission in Jaynes¨CCummings lattices, *New J. Phys.* **13**, 095008 (2011).
Schiró, M., Bordyuh, M., Ö ztop, B. & Türeci, H. E. Phase Transition of Light in Cavity QED Lattices, *Phys. Rev. Lett.* **109**, 053601 (2012).
Lu, L., Joannopoulos, J. D. & Soljačić, M. Topological photonics, *Nat. Photonics* **8**, 821-829 (2014).
Karzig, T., Bardyn, C. E., Lindner, N. H. & Refael, G. Topological Polaritons, *Phys. Rev. X* **5**, 031001 (2015).
Peano, V., Brendel, C., Schmidt, M. & Marquardt, F. Topological phases of sound and light, *Phys. Rev. X* **5**, 031011 (2015).
Goldman, N., Budich, J. C. & Zoller, P. Topological quantum matter with ultracold gases in optical lattices, *Nat. Phys.* **12**, 639-645 (2016).
Huber, S. D. Topological mechanics, *Nat. Phys.* **12**, 621-623 (2016).
Cooper, N. R., Dalibard, J. & Spielman, I. B. Topological bands for ultracold atoms, Rev. Mod. Phys. [**91**]{}, 015005 (2019)..
Zhang, L. & Liu, X. J. Spin-orbit coupling and topological phases for ultracold atoms, *arXiv:*1806.05628 (2018).
Ozawa, T. *et al.* Topological photonics, Rev. Mod. Phys. [**91**]{}, 015006 (2019).
Hasan, M. Z. & Kane, C. L. Topological insulators, *Rev. Mod. Phys.* **82**, 3045-3067 (2010).
Qi, X. L. & Zhang, S. C. Topological insulators and superconductors, *Rev. Mod. Phys.* **83**, 1057-1110 (2011).
Nayak, C., Simon, S. H., Stern, A., Freedman, M. & Sarma, S. D. Non-Abelian anyons and topological quantum computation, *Rev. Mod. Phys.* **80**, 1083-1159 (2008).
Dalibard, J., Gerbier, F., Juzeliūnas, G., & Öhberg, P. Artificial gauge potentials for neutral atoms, *Rev. Mod. Phys.* **83**, 1523-1543 (2011).
Galitski, V. & Spielman, I. B. Spin-orbit coupling in quantum gases, *Nature (London)* **494**, 49-54 (2013).
Goldman, N., Juzeliūnas, G., Öhberg, P. & Spielman, I. B. Light-induced gauge fields for ultracold atoms, *Rep. Prog. Phys.* **77**, 126401 (2014).
Jotzu, G. *et al.* Experimental realization of the topological Haldane model with ultracold fermions, *Nature (London)* **515**, 237-240 (2014).
Aidelsburger, M. *et al.* Measuring the Chern number of Hofstadter bands with ultracold bosonic atoms, *Nat. Phys.* **11**, 162-166 (2015).
Wu, Z. *et al.* Realization of two-dimensional spin-orbit coupling for Bose-Einstein condensates, *Science* **354**, 83-88 (2016).
Fläschner, N. *et al.* Observation of dynamical vortices after quenches in a system with topology, *Nat. Phys.* **14**, 265-268 (2018).
Bzdušek, T., Wu, Q., Rüegg, A., Sigrist, M. & Soluyanov, A. A. Nodal-chain metals, *Nature (London)* **538**, 75-78 (2016).
Burkov, A. A., Hook, M. D. & Balents, L. Topological nodal semimetals, *Phys. Rev. B* **84**, 235126 (2011).
Zhang, D.-W. *et al.* Quantum simulation of exotic PT-invariant topological nodal loop bands with ultracold atoms in an optical lattice, *Phys. Rev. A* [**93**]{}, 043617 (2016).
Gu, X., Kockum, A. F., Miranowicz, A., Liu, Y.-X. & Nori, F. Microwave photonics with superconducting quantum circuits, *Phys. Rep.* **718-719**, 1-102 (2017).
LinPeng, X. Y. *et al.* Joint quantum state tomography of an entangled qubit-resonator hybrid, *New J. Phys.* **15**, 125027 (2013).
Mei, F., Chen, G., Tian, L., Zhu, S. L. & Jia, S. Topology-dependent quantum dynamics and entanglement-dependent topological pumping in superconducting qubit chains, *Phys. Rev. A* **98**, 032323 (2018).
Bradlyn, B. *et al.* Beyond Dirac Weyl fermions: Unconventional quasiparticles in conventional crystals, *Science* **353**, aaf5037 (2016).
Lohse, M., Schweizer, C., Price, H. M., Zilberberg, O. & Bloch, I. Exploring 4D quantum Hall physics with a 2D topological charge pump, *Nature (London)* **553**, 55-58 (2018).
Zilberberg, O. *et al.* Photonic topological boundary pumping as a probe of 4D quantum Hall physics, *Nature (London)* **553**, 59-62 (2018).
Hayward, A. L. C., Martin, A. M. & Greentree, A. D. Fractional Quantum Hall Physics in Jaynes-Cummings-Hubbard Lattices, *Phys. Rev. Lett.* **108**, 223602 (2012).
Felicetti, S. *et al.* Dynamical Casimir Effect Entangles Artificial Atoms, *Phys. Rev. Lett.* **113**, 093602 (2014).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'A proof based on the Chern-Gauss-Bonnet Theorem is given to Hopf Theorem concerning the degree of the Gauss map of a hypersurface in ${{\mathbb R}}^n$.'
address: 'Universidade Federal do Ceará, Fortaleza, CE, Brasil'
author:
- Daniel Cibotaru
title: A short proof of a Theorem by Hopf
---
Introduction
============
Let $H\subset {{\mathbb R}}^n$ be a smooth, compact, embedded hypersurface without boundary. It is well-known that $H$ separates Euclidean space into two connected components. Let $W$ be the bounded component of ${{\mathbb R}}^n$ with $\partial W=H$. Orient $H$ as the boundary of $W$ using the outer normal first convention. Hopf [@H] proved the following:
\[HT\] The degree of the Gauss map $\mathscr{G}_{H}:H{\rightarrow}S^{n-1}$ is equal to $\frac{\chi(H)}{2}$ when $n$ is odd and to $\chi(W)$ when $n$ is even.
In the classical textbook [@GP], Guillemin and Pollack show that the case of odd $n$ is a consequence of Poincaré-Hopf Theorem. Using Morse Theory, Sakkalis [@S] gave a particularly simple proof covering both cases of Hopf’s Theorem \[HT\]. In this note, we give another short proof, using the Chern-Gauss-Bonnet Theorem and its two versions: for manifolds with and without boundary [@Ch1; @Ch2]. While Guillemin and Pollack use Hopf’s Theorem to prove Chern-Gauss-Bonnet for even-dimensional hypersurfaces, we explain why we believe it might be as natural to go the other way around and consider Theorem \[HT\] to be an application of this other famous result.
One can look at Chern-Gauss-Bonnet as a combination of two fundamental facts in topology:
- a Poincaré Duality statement which says that the Pfaffian of the Levi-Civita connection (properly normalized) is Poincaré dual to the degree $0$ current induced by a generic vector field with only isolated stationary points; in simple terms the later is represented by a bunch of points with appropriate signs attached;
- the homotopy invariance of the intersection number which takes the form of the Poincaré-Hopf Theorem, masterly explained in [@GP]; admitedly one needs also a particular construction of a nice vector field in order to identify the index with the Euler characteristic.
Without entering into historical details we will say that the first intrinsic proof of the generalization to the classical Gauss-Bonnet Theorem is due to Chern in [@Ch1]. This is where the transgression class (see below) first appears. A year later, Chern generalized his result to the case of a manifold with boundary in [@Ch2].
In regard to item (a), we recommend a fairly elementary treatment of a general version of Chern-Gauss-Bonnet for oriented vector bundles of even rank in [@Ni]. For a direct proof, based on a general transgression formula see [@Ci]. We remark that, notwithstanding the difficulty of any complete proof, the Chern-Gauss-Bonnet Theorem is not fundamentally based on Theorem \[HT\] but is indeed a manifestation of Poincaré Duality. On the other hand, the latter is a direct consequence of the former as we will see next.
The proof
=========
We use the notation of the previous section. The following is well-known to geometers.
The pull-back via the Gauss map $\mathscr{G}_H:H{\rightarrow}S^{n-1}$ of the Levi-Civita connection $\nabla^{S^{n-1}}$ on $S^{n-1}$ endowed with the round metric is the Levi-Civita connection $\nabla^{H}$ on $H$.
First of all, the pull-back (via the Gauss map) of the tangent bundle to the sphere is canonically isomorphic to the tangent bundle to $H$. Let now $X:S^{n-1}{\rightarrow}TS^{n-1}$ be a vector field on the sphere. Then at a point $p\in H$ $$\mathscr{G}_H^*\nabla^{S^{n-1}}_{Y_p}(X\circ \mathscr{G}_H):=\nabla^{S^{n-1}}_{d_p\mathscr{G}_H(Y_p)}X=P_{T_{\mathscr{G}_H(p)}S^{n-1}}dX(d_p\mathscr{G}_H(Y_p))=$$$$=P_{T_pH}d_p(X\circ \mathscr{G}_H)(Y_p)=\nabla_{Y_p}^H(X\circ \mathscr{G}_H).$$ The notation $P_V$ represents orthogonal projection onto the linear subspace $V$. The pull-back connection is uniquely determined by its action on “pull-back” sections, i.e. sections of type $\widetilde{X}= X\circ \mathscr{G}_H$ and therefore we conclude that $\mathscr{G}_H^*\nabla^{S^{n-1}}=\nabla^H.$
This result is a particular case of the celebrated Narasimhan Ramanan Theorem [@NR] which says that every pair vector bundle plus connection over a compact manifold is isomorphic with the pull-back of a universal bundle plus universal connection via a classifying map. Indeed this holds even more generally for principal bundles with connections. In the Lemma above the sphere plays the role of the base space of the universal bundle and the Gauss map of a hypersurface plays the role of the classifying map.
In order to compute the degree of the Gauss map when $n$ is odd, instead of using the volume form on $S^n$ we use the Pfaffian of the Levi-Civita connection $\nabla^{S^{n-1}}$. By the Gauss-Bonnet Theorem the integral over the sphere of the Pfaffian $\operatorname{Pf}(\nabla^{S^{n-1}})$ (properly normalized) equals $2$. Hence the degree of the Gauss map is $$\label{parta} \deg{\mathscr{G}_{H}}=\frac{1}{2}\int_{H}\mathscr{G}_{H}^*\operatorname{Pf}(\nabla^{S^{n-1}})=\frac{1}{2}\int_{H}\operatorname{Pf}(\nabla^{H})=\frac{\chi(H)}{2}.$$ where we used Chern-Gauss-Bonnet Theorem on $H$ in the last equality.
When $n$ is even and hence $S^{n-1}$ is odd-dimensional, the Pfaffian of $\nabla^{S^{n-1}}$ is zero by definition. However, there exists another natural form built with the help of the Levi-Civita connection that plays the role of the Pfaffian. This is a so-called transgression form. It can be described as follows. The trivial bundle ${{\mathbb R}}^{n}\bigr|_{S^{n-1}}$ splits orthogonally as a direct sum of the normal bundle $\tau{\rightarrow}S^{n-1}$ and $TS^{n-1}{\rightarrow}S^{n-1}$. It can therefore be endowed with two connections: the trivial one $d$ and the direct sum connection $d\oplus \nabla^{S^{n-1}}$.
The Pfaffians of both $\overline{\nabla}:=d$ and $\hat{\nabla}:=d\oplus \nabla^{S^{n-1}}$ are obviously identically zero as they are of degree $n>\dim{S^{n-1}}$. However, a standard result of the theory of characteristic classes (see for example the article by Chern and Simons, [@CS]) says that the difference of the two Pfaffians is equal to an exact form, which can be explicitly described. We write: $$0=\operatorname{Pf}(d)-\operatorname{Pf}(d\oplus \nabla^{S^{n-1}})=d\operatorname{TPf}(\overline\nabla,\hat{\nabla})$$ where $$\label{transclass} \operatorname{TPf}(\overline\nabla,\hat{\nabla})=\int_{[0,1]}\operatorname{Pf}(\widetilde{\nabla}).$$ Above $\widetilde{\nabla}:=\frac{d}{dt}+(1-t)\hat{\nabla}+t\overline{\nabla}$ is a connection on the trivial bundle $\underline{{{\mathbb R}}^{n}}{\rightarrow}[0,1]\times S^{n-1}$ resulting by taking the affine combination of $\overline{\nabla}$ and $\hat{\nabla}$. The operator $\frac{d}{dt}$ signals that the covariant derivative of $\widetilde{\nabla}$ in the $\frac{\partial}{\partial t}$ direction is just derivation with respect to this field. Finally, the integral is really integration of a form over the fiber of the projection $$\pi:[0,1]\times S^{n-1}{\rightarrow}S^{n-1}.$$ Before one starts believing that we just wrote $0$ in a complicated manner we show the following.
For $n$ even the folowing holds: $$\int_{S^{n-1}}\operatorname{TPf}(\overline\nabla,\hat{\nabla})=-1.$$
Apply Chern-Gauss-Bonnet Theorem [@Ch2] to the manifold with boundary $D^{n}\subset {{\mathbb R}}^{n}$: $$1=\chi(D^{n})=\int_{D^{n}}\operatorname{Pf}(\overline{\nabla})-\int_{S^{n-1}}\operatorname{TPf}(\overline\nabla,\hat{\nabla})=-\int_{S^{n-1}}\operatorname{TPf}(\overline\nabla,\hat{\nabla}).$$
For a computational proof and also for a local description of the transgression class the reader can take a look at Lemma 8.3.18 in [@Ni]. This starts by noting that that $\operatorname{TPf}(\overline\nabla,\hat{\nabla})$ is a constant multiple of the volume form of $S^{n-1}$ since it enjoys the same symmetry property as the volume form. To see which multiple exactly, one can relate the two quantities at a single preferred point on the sphere.
It is easily seen that the transgression class satisfies a naturality property similar with that of the Pfaffian. Namely: $$\mathscr{G}_H^*\operatorname{TPf}(\overline\nabla,\hat{\nabla})=\operatorname{TPf}(\mathscr{G}_H^*\overline\nabla,\mathscr{G}_H^*\hat{\nabla})=\operatorname{TPf}(\overline{\nabla}^H,\hat{\nabla}^H),$$ where $\overline\nabla^H$ is the trivial connection on $\underline{{\mathbb R}}^n{\rightarrow}H$ and $\hat{\nabla}^H$ is the direct sum of the trivial connection and the Levi-Civita connection on $H$. We conclude that $$\label{partb} \deg{ \mathscr{G}_H}=-\int_{H}\mathscr{G}_H^*\operatorname{TPf}(\overline\nabla,\hat{\nabla})=\int_{W}\operatorname{Pf}(\overline\nabla^{H})-\int_{\partial W}\operatorname{TPf}(\overline{\nabla}^H,\hat{\nabla}^H)=\chi(W),$$ where, in the last line we used Chern-Gauss-Bonnet [@Ch2] on the oriented, compact, *flat* manifold with boundary $W$. The lines (\[parta\]) and (\[partb\]) finish the proof of this note.
[99]{} S.-S. Chern, *A Simple Intrinsic Proof of the Gauss-Bonnet Formula for Closed Riemannian Manifolds*, Ann. of Math., [**45**]{}, no. 4 (1944), 747-752. S.-S. Chern, *On the Curvatura Integra in a Riemannian Manifold*, Ann. of Math., [**46**]{}, no. 4 (1945), 674-684. S.-S. Chern, J. Simons, *Characteristic Forms and Geometric Invariants*, Ann. of Math., Sec. Series, [**99**]{}, no. 1 (1974), 48-69. D. Cibotaru, *Vertical flows and a general currential homotopy formula*, http://arxiv.org/abs/1405.0952. V. Guillemin, A. Pollack, *Differential Topology*, Prentice-Hall Inc., New Jersey, 1974. H. Hopf, *Über die Curvatura integra geschlossener Hyperflächen*, Math. Ann., [**95**]{} (1925), 340-365. L. Nicolaescu, *Lectures on the Geometry of Manifolds*, Sec. Edition, World Scientific, Singapore, 2007. M.S. Narasimhan, S. Ramanan, *Existence of Universal Connections*, Amer. J. Math., [**83**]{}, No. 3 (1961), 563-572. T. Sakkalis, *On a Theorem by Hopf*, Internat. J. Math. & Math. Sci, [**13**]{}, no. 4 (1990), 813-816.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present two classes of stars with yet unknown evolutionary phase: the B\[e\] supergiants and the so-called unclassified B\[e\] stars. While the B\[e\] supergiants are luminous post-main sequence stars with high mass progenitors, not much is known of the unclassified stars. We discuss how it might be possible to determine their evolutionary phases in the area of new large telescopes.'
author:
- Michaela Kraus
title: 'The evolutionary phase of B\[e\] supergiants and unclassified B\[e\] stars'
---
[ address=[Astronomical Institute, Utrecht University, Princetonplein 5, 3584 CC Utrecht, The Netherlands]{} ]{}
Introduction
============
Stars showing the B\[e\] phenomenon have several characteristics of their optical spectrum in common: they are of spectral type B, have strong Balmer emission lines, and many permitted and forbidden emission lines of predominantly low ionization metals like e.g. Fe[ii]{} and O[i]{}. In addition, these stars are known to possess a strong near- or mid-infrared excess due to hot circumstellar dust. Since most of the characteristics are based purely on the optical spectrum, these stars can be of quite different evolutionary phase, and @l1998 grouped all stars with the B\[e\] phenomenon accordingly. The different classes they found are: B\[e\] supergiants, Herbig Ae/B\[e\] stars, and compact planetary nebula B\[e\] stars. The biggest group, however, is formed by stars of unknown evolutionary phase, the so-called unclassified B\[e\] stars. In this paper we want to concentrate on two classes of B\[e\] stars: the supergiants and the unclassified stars. Both classes have in common, that their evolutionary phase is still unknown and we want to strengthen the need of large telescopes for the determination of the evolutionary phases of these stars.
The B\[e\] supergiants
======================
B\[e\] supergiants are luminous ($4 \leq \log L/L_{\odot} \leq 6$) post-main sequence objects. Their progenitors must have been massive ($7 < M/M_{\odot}
< 85$) and probably rapidly rotating stars. The best known B\[e\] supergiants are located in the Magellanic Clouds. In the Milky Way only some candidates have been found. Here, due to large interstellar extinction the proper determination of distances and therefore luminosities is difficult. The optical spectra of B\[e\] supergiants show a hybrid character, i.e. a co-existence of very broad Balmer lines and extremely narrow forbidden and permitted low ionization metal lines. This hybrid character has led to the interpretation of two different winds: a normal line-driven (CAK-type) fast wind of low density in polar direction and a slow, high density disk-forming wind in equatorial direction where the hot dust is located (see @z1985). @k2003 performed ionization structure calculations in such non-spherical winds of B\[e\] supergiants using a latitude dependent mass flux that increases from pole to equator. They found that with such a model hydrogen can recombine close to the stellar surface in the equatorial region leading to a hydrogen neutral disk-like structure around these luminous stars. High-resolution optical spectra of B\[e\] supergiants revealed strong emission of O[i]{} forbidden lines (see e.g. @kb2004) which indicates the existence of a huge amount of neutral hydrogen since H and O have the same ionization potential. The additional detection of CO band emission in the near-infrared in several B\[e\] supergiants (e.g. @Mc1; [@Mc2]; [@Mc3]) completes the picture of the circumstellar disk leading to the following schematic representation of its temperature structure (from inside out): \[O[i]{}\] emission (10000K $\leq T \leq$ 6000) – CO band emission (5000K $\leq T \leq$ 2000) – dust emission ($T \leq 2\,000$K).
In the HR diagram, B\[e\] supergiants share their location with other massive post-main sequence stars like Wolf-Rayet stars and Luminous Blue Variables (LBVs). But the connection of B\[e\] supergiants with these other evolutionary states is still an unsolved problem and even newest stellar evolution calculations of massive stars cannot predict the B\[e\] supergiant phase (see e.g. @m2000; @m2004).
Unclassified B\[e\] stars: the example of Hen 2-90
==================================================
The enigmatic galactic star Hen 2-90 belongs to the group of unclassified B\[e\] stars. It has been classified in the literature either as a compact planetary nebula (e.g. @c1993; @l1998) or as a symbiotic object (@s2000; @g2001). On an HST image (@s2002) a non-spherical wind structure is visible: a high-ionized polar wind, a low-ionized wind at intermediate latitudes and a (neutral) disk structure in equatorial direction.
We took high- and low-resolution optical spectra with slit centered on this non-spherical wind structure. The spectra contain very strong Balmer lines and lots of forbidden emission lines from e.g. S, N, O, Ar, Cl, Fe in different ionization states. We detected neither emission from He[ ii]{} nor TiO absorption bands, which are both the main characteristics of a symbiotic object. In addition, we performed a detailed analysis of almost all observed forbidden emission lines and found that C and O are depleted, which means that Hen 2-90 should be an evolved object. Our conlusions are therefore, that Hen 2-90 is most probable a compact planetary nebula (@k2004).
There are, however, a few puzzling details that have to date not been solved: (i) Hen 2-90 shows a jet like structure with many knots being ejected regularly on both sides of the star and perpendicular to the disk. (ii) The velocity profiles of the different emission lines hint to a much more complex structure of the circumstellar material than might be explained with a simple wind model. (iii) We also found from our modeling that N is depleted, which cannot be understood in terms of a normal single star evolution so that probably a binary nature of this star needs to be taken into account (@kba2004). More observations are certainly needed to disentangle the nature of this fascinating object.
How can we determine evolutionary phases ?
==========================================
The determination of the evolutionary phases of B\[e\] supergiants and unclassified B\[e\] stars strongly depends on the interplay of theory with observations. Here, we want to mention some of the methods we are using to fix stellar and circumstellar parameters necessary for the classification.
- [**Abundance determinations:**]{}
The determination of elemental abundances is one key project when dealing with stellar evolution. It is thereby necessary to determine abundances in the circumstellar matter as well as on the stellar surface. The circumstellar material mirrors the surface abundance at the time of matter ejection, while the surface itself gives the actual abundance of the star. The surface abundance can be quite different from the circumstellar abundance if (i) the star was ejecting its complete hydrogen rich outer layers leaving a He rich surface behind, or (ii) the star is rapidly rotating leading to rotationally induced mixing of the internal material which results in a continuous surface enrichement of processed material. Since rapidly rotating stars have high mass loss rates, the mixing leads to an abundance gradient in the wind material.
- [**Determination of the geometry and kinematics of the circumstellar matter:**]{}
The geometry of the system is a crucial point, especially if no direct imaging is available or possible. As discussed in the previous section, the appearance and strength of specific (e.g. forbidden) emission lines can hint to a non-spherical density distributions and even to neutral material close to the hot star. A detailed analysis of forbidden (and permitted) emission lines is needed to derive the ionization structure in non-spherical winds and disks. Especially in the case of B\[e\] stars we know that in polar direction highly-ionized, line-driven winds of rather high velocity emanate, while from equatorial directions the narrow emission lines of low-ionized or even neutral metals are observed. Both types of profiles will be present in high-resolution spectra.
Besides the atomic and ionic lines also molecular emission, like e.g. the CO bands can be used to determine the kinematics of the emitting gas. Especially for several B\[e\] supergiants the CO bands which arise in the near-infrared have been observed. High-resolution observations of the CO $2\longrightarrow 0$ band head display the complete velocity information of the CO emitting gas. This band head shows e.g. a red peak and a blue shoulder in case of rotation. By modeling the high-resolution, high signal-to-noise $2\longrightarrow 0$ band head structure one can discriminate contributions coming from rotation and/or outflow (see e.g. @k2000).
- [**Determination of the mass loss history:**]{}
The analysis of the forbidden emission lines can also be used to derive (non-spherical) mass fluxes and therefore the total mass loss rate of a star as has been shown by @k2004 for the unclassified B\[e\] star Hen 2-90. Mass loss rates of the B\[e\] supergiants and unclassified B\[e\] stars can then be compared with those available in the literature for stars of different initial conditions and in different evolutionary phases to find agreements and therefore a possible evolutionary stage of each star.
Observations needed
===================
For the modeling of the CO bands high-resolution NIR spectra are needed. PHOENIX, the new high-resolution near-infrared spectrograph at Gemini-South is most suitable for these observations.
For the determination of terminal velocities, the mass loss history and the surface and circumstellar matter abundances we need mainly high-resolution optical (and UV) observations. Since many of the unclassified B\[e\] stars are southern objects and most of the nowadays known B\[e\] supergiants are located in the Magellanic Clouds, the new South African Large Telescope (SALT) will be the ideal tool to guarantee the success of our projects. Its major targets will be the Magellanic Clouds (see contributions by David and William in this volume) allowing us to retrieve the high-resolution optical spectra which are needed for a proper classification of B\[e\] stars.
This research was supported by the Nederlandse Organisatie voor Wetenschappelijk Onderzoek grant No.614.000.310.
H. J. G. L. M. Lamers, F.-J. Zickgraf, D. de Winter, L. Houziaux and J. Zorec, [*Astron. Astroph.*]{} [**340**]{}, 117-128 (1998). F.-J. Zickgraf, B. Wolf, O. Stahl, C. Leitherer and G. Klare, [*Astron. Astroph.*]{} [**143**]{}, 421-430 (1985). M. Kraus and H. J. G. L. M. Lamers, [*Astron. Astroph.*]{} [**405**]{}, 165-174 (2003). M. Kraus and M. Borges Fernandes, in [*The Nature and Evolution of Disks Around Hot Stars*]{}, edited by R. Ignace and K. Gayley, ASP Conference Series, San Francisco: Astronomical Society of the Pacific, 2004, astro-ph/0408073. P. J. McGregor, D. J. Hillier and A. R. Hyland, [*Astroph. J.*]{} [**334**]{}, 639-656 (1988a). P. J. McGregor, A. R. Hyland and D. J. Hillier, [*Astroph. J.*]{} [**324**]{}, 1071-1098 (1988b). P. J. McGregor, A. R. Hyland and M. T. McGinn, [*Astron. Astroph.*]{} [**223**]{}, 237-240 (1989). A. Maeder and G. Meynet, [*Ann. Rev. Astron. Astroph.*]{} [**38**]{}, 143-190 (2000). A. Maeder, G. Meynet and R. Hirschi, in: [*The Fate of the Most Massive Stars*]{}, edited by R. Humphreys and K. Stanek, ASP Conference Series, San Francisco: Astronomical Society of the Pacific, 2004, astro-ph/0408320. R. D. D. Costa, J. A. de Freitas Pacheco and W. J. Maciel, [*Astron. Astroph.*]{} [**276**]{} 184-186 (1993). R. Sahai and L.-Å Nyman, [*Astroph. J.*]{} [**537**]{}, L145-149 (2000). M. A. Guerrero, L. F. Miranda, Y. H. Chu, M. Rodriguez and R. M. Williams, [*Astroph. J.*]{} [**563**]{}, 883-888 (2001). R. Sahai, S. Brillant, M. Livio, E. K. Grebel, W. Brandner, S. Tingay and L.-Å Nyman, [*Astroph. J.*]{} [**573**]{}, L123-127 (2002). M. Kraus, M. Borges Fernandes, F. X. de Araújo and H. J. G. L. M. Lamers, [*Astron. Astroph.*]{} (2004) [ *submitted*]{}. M. Kraus, M. Borges Fernandes and F. X. de Araújo, in: [*Massive Stars in Interacting Binaries*]{}, edited by A. Moffat and N. St-Louis, ASP Conference Series, San Francisco: Astronomical Society of the Pacific, 2004, astro-ph/0410196 M. Kraus, E. Krügel, C. Thum and T. R. Geballe, [*Astron. Astroph.*]{} [**362**]{}, 158-168 (2000).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The paper gives a comprehensive presentation of a framework, embedded into the simply typed higher-order logic, and aimed at providing a sound assistance in formal reasoning about models of imperative programs with interleaved computations. As a case study, a model of the Peterson’s mutual exclusion algorithm will be scrutinised in the course of the paper illustrating applicability of the framework.'
address: |
Department of Software Science\
Tallinn University of Technology
author:
- Maksym Bortin
title: 'A Framework for Modelling, Verification and Transformation of Concurrent Imperative Programs'
---
[^1]
Introduction {#S:intro .unnumbered}
============
The behaviours of programs running in parallel can become entangled to such an extent that providing resilient assurances about these simply escapes the capabilities of mere reviewing and testing the code. On the other hand, such assurances are hardly dispensable, for example when safe functioning of some nearly invaluable autonomous devices is at stake. A fundamentally rigorous and systematic reasoning about program properties in presence of interleaving is unavoidable to this end.
From the specific perspective of safety- and mission-critical software, more recent standards encourage application of formal methods in certification processes. On the other hand, they also demand to explicitly outline the limits of the chosen method regarding compliance with the actual requirements specification, as the formal methods supplement to DO-178C [@standard:dod178c] in particular stresses. Indeed, like with outcomes of any other measure, applicability of formally derived conclusions to requirements on eventual implementations is bound to series of assumptions that one shall attempt to keep as short and reasonable as possible.
In comparison to the sequential case, development of concurrent software has significantly more sources for this kind of assumptions. For example, it is a common practice to model and verify sequential programs at a conveniently structured level and to rely on a compiler that translates these to an assembly code with some corresponding behaviour. However, in the context of interleaved computations one can observe that such compilations in particular tend to fraction the granularity, modelled at the structured level, meaning that their outcomes could potentially be exposed to interferences that were basically ruled out when reasoning about their structured counterparts or, in other words, were implicitly assumed to be nonexistent.
To address these kinds of problems in a cohesive manner, the paper presents a framework geared towards verification and transformation of models of imperative concurrent programs encompassing state abstractions as well as high- and low-level language features simultaneously. It shall enable ways to reach detailed representations by chains of transformations reusing verified properties along these chains. In this sense, the approach is particularly related to the Event-B [@abrial_2010] method, which offers a homogeneous framework for transformation of models of reactive systems, gradually addressing more detailed system requirements. In contrast to Event-B, however, the presented framework is less generic as it is specifically tailored to concurrent programs. Moreover, it is a conservative extension to the simply typed higher-order logic and is thus as sound as the underlying logic.
A sound framework is surely a must, but addresses only a part of the problem: derivations of properties of programs that exceed the size of what is usually called a ‘toy example’ typically require an impressive amount of successive applications of various logical rules which also need to be correctly instantiated at each particular step. Even a tiniest mistake during this process is likely to compromise all efforts to reach a sound conclusion. And this is the point where proof assistants ultimately become relevant, being capable to accomplish these complex, yet largely mechanisable, tasks in a highly reliable and efficient way. Based on Robin Milner’s influential work [@Milner72; @Milner411], there are tools such as [@Nipkow_PW:Isabelle], [@DBLP:conf/tphol/Harrison09a], [@DBLP:conf/tphol/Gordon91] that provide the sought assistance.
Regarding related work, the paper brings together and further elaborates on the following approaches:
- the Hoare-style rely/guarantee program logic development essentially follows the systematic extension [@STIRLING1988347] of the Owicki-Gries method [@Owicki_Gries_76];
- the technique of blending deep and shallow embeddings with abstraction over state and atomic state transformations has been adopted from *SIMPL* (Sequential Imperative Programming Language) [@Schirmer:PhD];
- the light-weight approach to enable state relations as postconditions using the power of the underlying logic has been taken from [@DBLP:conf/mpc/Staden15] and slightly adjusted.
The paper is structured as follows. Next section contains a superficial introduction to the basic decisions behind the framework’s design. Section \[S:lang\] gives a formal presentation of the framework’s language accompanied by a computational model featuring interleaving. Section \[S:pcorr\] is devoted to program correspondences and their properties. Section \[S:pcs-props\] introduces the relevant conditions on potential computations, and based on these, Section \[S:prog-log\] presents a Hoare-style program logic. Section \[S:PM1\] applies this to verify properties of a model of Peterson’s mutual exclusion algorithm. Section \[S:gener\] shows how the program logic can be lifted to state relations as postconditions, whereas Section \[S:PM2\] utilises this extension to strengthen the results of Section \[S:PM1\]. Section \[S:live\] describes an approach to verification of basic liveness properties, such that Section \[S:PM3\] can conclude the case study presenting how termination of the mutex model can be established. Furthermore, in Section \[S:split\] an approach to splitting of atomic program steps will be sketched. Finally, Section \[S:concl\] concludes the paper.
Outlining Basic Ideas {#S:concept}
=====================
This section contains a brief overview of some basic ideas behind the framework’s ‘low-level’ commands and program transformations involving these. Although reaching some lower level representations of a concurrent program does not need to be a goal of a development process, enabling this option nonetheless mints the framework’s design in certain ways, outlined below. This will be accomplished without delving into much detail, appealing rather to the intuition. The more so as assembly level languages and their execution models are not directly in the formal scope of the paper: the claim is merely that programs having an appropriate form would be evaluated by the framework’s computational model (Section \[Sb:sem\]) basically in the same way assembly interpreters would do.
One of the most striking differences between a structured and an assembly-level language is that in the former case computational effects are conveniently defined for each of the language constructors by viewing these as nodes of an abstract syntax tree. Thus, for instance, a program ${{\mathsf{cjump}}\:C\;{\mathsf{to}}\:i\;{\mathsf{otherwise}}\;p\;{\mathsf{end}}};q$ is evaluated (Section \[Sb:sem\]) by first picking the node ${{\mathsf{cjump}}\:C\;{\mathsf{to}}\:i\;{\mathsf{otherwise}}\;p\;{\mathsf{end}}}$, checking if the condition $C$ holds, and if so, retrieving the code ${\rho\:i}$ associated to the label $i$ and ‘pasting’ it back, which results in ${\rho\:i};q$. Likewise, if $C$ does not hold, we would continue with $p;q$. Whereas evaluations of the node ${{\mathsf{cjump}}\:C\;{\mathsf{to}}\:i\;{\mathsf{otherwise}}\;p\;{\mathsf{end}}}$ essentially mimic processing of a conditional jump by assembly interpreters, evaluations of ${{\mathsf{cjump}}\:C\;{\mathsf{to}}\:i\;{\mathsf{otherwise}}\;p\;{\mathsf{end}}};q$ do not: they follow the abstract syntax tree structure and simply advance to the same $q$ in both branches $\rho{\:i};q$ and $p;q$ once $\rho{\:i}$ or $p$ are done with their computations. However, such behaviour can generally be achieved at the assembly level only when either ${\rho\:i}$ or $p$ concludes with an explicit jump to the entry label of $q$. These considerations shall underline that invocation of jumps in tree-structured programs demands certain preparations to make sense from the assembly language perspective and the transformations, sketched below, take care of that.
First transformation ought to replace all conditional statements by conditional jumps across a program, and will be backed by the observation that ${{\mathsf{if}}\:{\mathit{C}}\;{\mathsf{then}}\;p_1\;{\mathsf{else}}\;p_2\;{\mathsf{fi}}}$ is essentially the same as ${{\mathsf{cjump}}\:{\neg C}\;{\mathsf{to}}\:i\;{\mathsf{otherwise}}\;p_1\;{\mathsf{end}}}$ when ${\rho\:i} = p_2$ (${\neg C}$ will denote the complement of the condition $C$). As stressed above, applied to ${{\mathsf{if}}\:{\mathit{C}}\;{\mathsf{then}}\;p_1\;{\mathsf{else}}\;p_2\;{\mathsf{fi}}};q$ the transformation would result in ${{\mathsf{cjump}}\:{\neg C}\;{\mathsf{to}}\:i\;{\mathsf{otherwise}}\;p_1\;{\mathsf{end}}};q$ with ${\rho\:i} = p_2$. To repair that, ${{\mathsf{if}}\:{\mathit{C}}\;{\mathsf{then}}\;p_1\;{\mathsf{else}}\;p_2\;{\mathsf{fi}}};q$ is first replaced by the equivalent ${{\mathsf{if}}\:{\mathit{C}}\;{\mathsf{then}}\;p_1;q\;{\mathsf{else}}\;p_2;q\;{\mathsf{fi}}}$, distributing all subsequent code to both branches. Now, replacing ${{\mathsf{if}}\:{\mathit{C}}\;{\mathsf{then}}\;p_1;q\;{\mathsf{else}}\;p_2;q\;{\mathsf{fi}}}$ yields ${{\mathsf{cjump}}\:{\neg C}\;{\mathsf{to}}\:i\;{\mathsf{otherwise}}\;p_1;q\;{\mathsf{end}}}$ with ${\rho\:i} = p_2;q$, whose evaluations do not make use of any implicit jumps and simply halt when $p_1;q$ or $p_2;q$ has been processed.
With ${{\mathsf{while}}\:{\mathit{C}}\;{\mathsf{do}}\;{\mathit{p}}\;{\mathsf{od}}}$ we would have a similar situation: evaluations of ${{\mathsf{while}}\:{\mathit{C}}\;{\mathsf{do}}\;{\mathit{p}}\;{\mathsf{od}}};{\mathit{q}}$ make use of the syntax tree structure to advance to $q$ when $C$ does not hold. In order to enable a normalisation similar to the conditional statements, the syntax of ${\mathsf{while}}$-statements will be extended to ${{\mathsf{while}}\:{\mathit{C}}\;{\mathsf{do}}\;{\mathit{p}}\;{\mathsf{subsequently}}\;{\mathit{q}}\;{\mathsf{od}}}$, such that ${{\mathsf{while}}\:{\mathit{C}}\;{\mathsf{do}}\;{\mathit{p}}\;{\mathsf{subsequently}}\;q_1\;{\mathsf{od}}};q_2$ can be transformed to the equivalent ${{\mathsf{while}}\:{\mathit{C}}\;{\mathsf{do}}\;{\mathit{p}}\;{\mathsf{subsequently}}\;q_1;q_2\;{\mathsf{od}}}$. Thus, having access to all of subsequent code, ${{\mathsf{while}}\:{\mathit{C}}\;{\mathsf{do}}\;{\mathit{p}}\;{\mathsf{subsequently}}\;{\mathit{q}}\;{\mathsf{od}}}$ can safely be replaced by ${{\mathsf{cjump}}\:{\neg {\mathit{C}}}\;{\mathsf{to}}\:j\;{\mathsf{otherwise}}\;{\mathit{p}};{\mathsf{jump}}\:i\;{\mathsf{end}}}$ with ${\rho\:i} = {{\mathsf{cjump}}\:{\neg {\mathit{C}}}\;{\mathsf{to}}\:j\;{\mathsf{otherwise}}\;{\mathit{p}};{\mathsf{jump}}\:i\;{\mathsf{end}}}$ and ${\rho\:j} = q$. Note that the extra unconditional jump following $p$ is not completely ‘for free’ and will appear in form of an extra ${{\mathsf{skip}}}$ when defining how ${{\mathsf{while}}\:{\mathit{C}}\;{\mathsf{do}}\;{\mathit{p}}\;{\mathsf{subsequently}}\;{\mathit{q}}\;{\mathsf{od}}}$ shall be evaluated.
Successive applications of the above transformations using consistent labelling eventually reach an equivalent ‘while-free’ program which essentially corresponds to flat lists of labelled instructions. For instance, $p \parallel q$ with some sequential $p$ and $q$ would yield two such lists: one corresponding to $p$ and one – to $q$.
To sum up, this section outlined questions that arise when viewing programs, where conditional and ${\mathsf{while}}$-statements are replaced by jumps, from the perspective of an assembly interpreter, and how these questions are addressed by the framework. All that will reappear in the remainder of the paper which, by contrast, is kept formal.
A Generic Concurrent Imperative Language {#S:lang}
========================================
The notions of shallow and deep embeddings arise quite naturally when a language needs to be modelled in a formal system such as a proof assistant ([@DBLP:conf/tphol/Harrison09a]). The technique applied here ([@Schirmer:PhD]) aims to combine the merits of both. At the first glance the syntax of the framework’s language is deeply embedded by means of a free construction. However, all indivisible state transformations as well as control flow conditions (such as the variable $C$ in conditional statements mentioned in the previous section) are embedded shallowly or, in other words, the language is generated over the logical terms having an interpretation so that no extra evaluations are needed.
The syntax {#Sb:lang}
----------
The parameterised type ${\mathtt{\mathcal{L}}_{\alpha}}$, where $\alpha$ is a type parameter representing underlying states and $\mathcal{L}$ – a type constructor giving a new type for any actual type supplied for $\alpha$, captures the syntax. For example, ${\mathtt{\mathcal{L}}_{{\mathtt{int}} \times {\mathtt{int}}}}$ would be an instance acting on the states with two integer variables. Note that such instantiations will appear in specific modellings only. Otherwise, states will be kept abstract for the sake of uniform reasoning.
Terms of type ${\mathtt{\mathcal{L}}_{\alpha}}$ are constructed by the following grammar: $$\begin{array}{l c l}
{\mathtt{\mathcal{L}}_{\alpha}} & \mbox{::=} & {{\mathsf{skip}}}\;\; | \;\; {{\mathsf{basic}}}\:f \;\; |\\
& & {{\mathsf{cjump}}\:C\;{\mathsf{to}}\:i\;{\mathsf{otherwise}}\;{\mathtt{\mathcal{L}}_{\alpha}}\;{\mathsf{end}}}\;\; |\\
& & {{\mathsf{while}}\:C\;{\mathsf{do}}\;{\mathtt{\mathcal{L}}_{\alpha}}\;{\mathsf{subsequently}}\;{\mathtt{\mathcal{L}}_{\alpha}}\;{\mathsf{od}}}\;\;| \\
& & {{\mathsf{if}}\:C\;{\mathsf{then}}\;{\mathtt{\mathcal{L}}_{\alpha}}\;{\mathsf{else}}\;{\mathtt{\mathcal{L}}_{\alpha}}\;{\mathsf{fi}}} \;\;| \\
& & {\mathtt{\mathcal{L}}_{\alpha}} ; {\mathtt{\mathcal{L}}_{\alpha}} \;\;|\\
& & {\parallel\!\!\mathcal{L}^+_\alpha} \;\;| \\
& & {{\mathsf{await}}\:C\;{\mathsf{do}}\;{\mathtt{\mathcal{L}}_{\alpha}}\;{\mathsf{od}}}
\end{array}$$ where
- $i$ is called a *label*, and ranges over the natural numbers (in principle, any infinitely countable set of identifiers, such as strings, might also be employed to this end);
- $f$ is called a *state transformer*, and ranges over the values of type $\alpha \Rightarrow \alpha$;
- $C$ is called a *state predicate*, and ranges over the values of type $\alpha \Rightarrow {\mathtt{bool}}$.
Furthermore, $\mathcal{L}^+_\alpha$ stands for a finite, non-empty sequence of elements of type ${\mathtt{\mathcal{L}}_{\alpha}}$. So, for instance, ${\parallel\!\!p_1, p_2, p_3}$ constructs a new value of type ${\mathtt{\mathcal{L}}_{\alpha}}$ out of the three values of this type. The term ${\parallel\!\!p_1, p_2, p_3}$ is then called the *parallel composition* of $p_1$, $p_2$, $p_3$. In the sequel we will more conveniently write $p_1 \parallel p_2$ in place of ${\parallel\!\!p_1, p_2}$. The reasons for having ${\parallel\!\!p_1, \ldots, p_n}$ as a primitive, as opposed to the binary operator with nesting, will be outlined in the following section. Regarding sequential composition, $p_1;p_2;p_3$ will be used as a shorthand for $p_1;(p_2;p_3)$. Moreover, ${{\mathsf{if}}\:C\;{\mathsf{then}}\;p\;{\mathsf{fi}}}$ abbreviates ${{\mathsf{if}}\:C\;{\mathsf{then}}\;p\;{\mathsf{else}}\;{{\mathsf{skip}}}\;{\mathsf{fi}}}$ and, accordingly, ${{\mathsf{while}}\:C\;{\mathsf{do}}\;p\;{\mathsf{od}}}$ stands for ${{\mathsf{while}}\:C\;{\mathsf{do}}\;p\;{\mathsf{subsequently}}\;{{\mathsf{skip}}}\;{\mathsf{od}}}$.
A term of type ${\mathtt{\mathcal{L}}_{\alpha}}$ will be called *jump-free* if it does not contain any ${\mathsf{cjump}}$, and *locally sequential* if the parallel operator does not occur in it. The motivation behind this explicit locality will also be given below.
The computational model {#Sb:sem}
-----------------------
This comprises two components: program steps and environment steps. To ensure fine-grained interleaving for program steps, we follow the principles of ‘Structural Operational Semantics’ [@PLOTKIN20043]. Inheriting the state abstraction from ${\mathtt{\mathcal{L}}_{\alpha}}$, let $\alpha$ be a fixed arbitrary type throughout this section.
### Program steps
Let $\rho$ be a code retrieve function, a function of type ${\mathtt{nat}} \Rightarrow {\mathtt{\mathcal{L}}_{\alpha}}$. Then in the context of $\rho$, the effects of performing a program step are captured by means of the relation ${\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}$ of type ${\mathtt{\mathcal{L}}_{\alpha}} \times \alpha \Rightarrow {\mathtt{\mathcal{L}}_{\alpha}} \times \alpha \Rightarrow {\mathtt{bool}}$, connecting two *configurations* that comprise a program part of type ${\mathtt{\mathcal{L}}_{\alpha}}$ and a state part of type $\alpha$. We will use the notation ${\rho \vdash (p, \sigma) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(q, \sigma{^{\scriptscriptstyle\prime}})}$ to indicate that $(p, \sigma)$ can be transformed via ${\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}$ to $(q, \sigma{^{\scriptscriptstyle\prime}})$ using $\rho$. This transformation is defined inductively by the rules listed in Figure \[fig:pstep\] and explained in more detail below.
$$\begin{array}{l l}
\mathrm{Basic} & {\dfrac{}{{\rho \vdash ({{\mathsf{basic}}}\:f, \sigma) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}({{\mathsf{skip}}}, f\:\sigma)}}} \\
& \\
\mbox{CJump-True} & {\dfrac{\sigma \in C}{{\rho \vdash ({{\mathsf{cjump}}\:C\;{\mathsf{to}}\:i\;{\mathsf{otherwise}}\;p\;{\mathsf{end}}}, \sigma) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}({\rho\:i}, \sigma)}}} \\
& \\
\mbox{CJump-False} & {\dfrac{\sigma \notin C}{{\rho \vdash ({{\mathsf{cjump}}\:C\;{\mathsf{to}}\:i\;{\mathsf{otherwise}}\;p\;{\mathsf{end}}}, \sigma) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(p, \sigma)}}} \\
& \\
\mathrm{Await} & {\dfrac{\sigma \in C \quad\quad {\rho \vdash (p, \sigma) {\stackrel{\scriptscriptstyle\star}{\mathop{\rightarrow}}_{\scriptscriptstyle{\mathcal{P}}}}({{\mathsf{skip}}}, \sigma{^{\scriptscriptstyle\prime}})}}{{\rho \vdash ({{\mathsf{await}}\:C\;{\mathsf{do}}\;p\;{\mathsf{od}}}, \sigma) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}({{\mathsf{skip}}}, \sigma{^{\scriptscriptstyle\prime}})}}} \\
& \\
\mbox{Conditional-True} & {\dfrac{\sigma \in C}{{\rho \vdash ({{\mathsf{if}}\:C\;{\mathsf{then}}\;p_1\;{\mathsf{else}}\;p_2\;{\mathsf{fi}}}, \sigma) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(p_1, \sigma)}}} \\
& \\
\mbox{Conditional-False} & {\dfrac{\sigma \notin C}{{\rho \vdash ({{\mathsf{if}}\:C\;{\mathsf{then}}\;p_1\;{\mathsf{else}}\;p_2\;{\mathsf{fi}}}, \sigma) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(p_2, \sigma)}}} \\
& \\
\mbox{While-True} & {\dfrac{\sigma \in C \quad\quad x = {{\mathsf{while}}\:C\;{\mathsf{do}}\;p_1\;{\mathsf{subsequently}}\;p_2\;{\mathsf{od}}}}{{\rho \vdash (x, \sigma) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(p_1 ; {{\mathsf{skip}}}; x, \sigma)}}} \\
& \\
\mbox{While-False} & {\dfrac{\sigma \notin C}{{\rho \vdash ({{\mathsf{while}}\:C\;{\mathsf{do}}\;p_1\;{\mathsf{subsequently}}\;p_2\;{\mathsf{od}}}, \sigma) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(p_2, \sigma)}}} \\
& \\
\mbox{Sequential} & {\dfrac{{\rho \vdash (p_1, \sigma) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(p{^{\scriptscriptstyle\prime}}_1, \sigma{^{\scriptscriptstyle\prime}})}}{{\rho \vdash (p_1 ; p_2, \sigma) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(p{^{\scriptscriptstyle\prime}}_1 ; p_2, \sigma{^{\scriptscriptstyle\prime}})}}} \\
& \\
\mbox{Sequential-Skip} & {\dfrac{}{{\rho \vdash ({{\mathsf{skip}}}; p, \sigma) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(p, \sigma)}}} \\
& \\
\mathrm{Parallel} & {\dfrac{{\rho \vdash (p_i, \sigma) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(p{^{\scriptscriptstyle\prime}}_i, \sigma{^{\scriptscriptstyle\prime}})} \quad\quad 1 \le i \le m \quad\quad m > 0}{{\rho \vdash ({\parallel\!\!p_1,\ldots p_i \ldots, p_m}, \sigma) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}({\parallel\!\!p_1,\ldots p{^{\scriptscriptstyle\prime}}_i \ldots, p_m}, \sigma{^{\scriptscriptstyle\prime}})}}} \\
& \\
\mbox{Parallel-Skip} & {\dfrac{\forall i\in\{1, \ldots, m\}.\: p_i = {{\mathsf{skip}}}\quad\quad m > 0}{{\rho \vdash ({\parallel\!\!p_1,\ldots, p_m}, \sigma) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}({{\mathsf{skip}}}, \sigma)}}}
\end{array}$$
1. The rule ‘Basic’ declares one indivisible computation step given by a state transformer $f$. Note that ‘indivisibility’ is actually a modelling decision: we might have a configuration $({{\mathsf{basic}}}\:(g \circ h), \sigma)$ transforming $\sigma$ to $g(h\: \sigma)$ in one atomic step but, conceptually, nothing keeps us from breaking it in two atomic steps $({{\mathsf{basic}}}\:h ; {{\mathsf{basic}}}\: g, \sigma)$. This, however, would lead to a different model with a finer granularity (Section \[S:split\]).
2. The rules ‘CJump-True’ and ‘CJump-False’ declare how a conditional jump is handled in dependence on $C$. There and throughout the paper, $\sigma \in C$ denotes that $\sigma$ satisfies the state predicate $C$. Furthermore, from now on let $\top_\tau$ denote the universal predicate on the type $\tau$, the one which is true for any element of type $\tau$. For the sake of brevity we will just write $\top$ whenever $\tau$ is clear from the context. In particular, considering configurations over a fixed arbitrary type of states $\alpha$, we just write $\top$ for $\top_\alpha$. Now, the *unconditional jump* command can be defined by $${{\mathsf{jump}}\:i} {\mathrel{\stackrel{\scriptscriptstyle{\mathit{def}}}{=}}}{{\mathsf{cjump}}\:\top\;{\mathsf{to}}\:i\;{\mathsf{otherwise}}\;{{\mathsf{skip}}}\;{\mathsf{end}}}$$ such that ${\rho \vdash ({{\mathsf{jump}}\:i}, \sigma) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}({\rho\:i}, \sigma)}$ holds with $({\rho\:i}, \sigma)$ as the only possible successor of $({{\mathsf{jump}}\:i}, \sigma)$.
It is worth to be highlighted once more that the framework does not impose any restrictions on $C$, except its type $\alpha \Rightarrow {\mathtt{bool}}$, whereas according to the rules it always takes one step to evaluate the condition, regardless how complex it is. Such liberty is useful in abstract modelling, but one has to keep in mind that certain interferences could be ruled out this way. In other words, when attempting to reach some low-level representations, $C$ shall comprise a condition that indeed can be evaluated in one atomic step.
3. The rule ‘Await’ declares that if $C$ is satisfied and $p$ can terminate without being interrupted in some state then the entire ${\mathsf{await}}$-statement terminates in this state by means of a single, indivisible step. By the inductive construction, ${{\mathsf{await}}\:C\;{\mathsf{do}}\;p\;{\mathsf{od}}}$ is blocked, cannot perform any step, in cases when the current state does not satisfy $C$. Thus, ${{\mathsf{await}}\:\top\;{\mathsf{do}}\;p\;{\mathsf{od}}}$ never blocks but creates an *atomic section*, which allows us to model situations where one process forces all others to hold on with their computations upon termination of $p$. As this particular construction will be used quite frequently, we just write $\langle p \rangle$ for it.
4. The rule ‘While-True’ has largely the standard form, except that it additionally accounts for transformation to a representation using jumps, sketched in Section \[S:concept\]. This amounts to the additional (and otherwise obsolete) ${{\mathsf{skip}}}$ in $(p_1 ; {{\mathsf{skip}}}; x, \sigma)$ that will act as a sort of placeholder for a jump in the proof of Proposition \[thm:while-norm2\]. Here, however, this ${{\mathsf{skip}}}$ has no other effect than to delay computations, introducing an additional ‘break’ between $p_1$ and $x$ which opens more possibilities for interleaving.
Notice that the computational model has more of such technical delays. It takes for example two steps to reach $p$ from $({{\mathsf{skip}}};{{\mathsf{skip}}});p$. In other words, the model is geared towards enabling more interleaving which is basically safer than to cut it.
5. The rule ‘Parallel’ states that a computation step of $m$ components composed in parallel is accomplished by random picking of a component that can perform its step. Note that this rule makes ${\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}$ in general non-deterministic.
Now we can justify the choice of having ${\parallel\!\!p_1, \ldots, p_n}$ as a primitive operator. Suppose $p$ is not ${{\mathsf{skip}}}$ and also not blocked by an ${\mathsf{await}}$-statement then $p \parallel ({{\mathsf{skip}}}\parallel {{\mathsf{skip}}})$ can make program steps either to $p \parallel {{\mathsf{skip}}}$ or to $p{^{\scriptscriptstyle\prime}}\parallel ({{\mathsf{skip}}}\parallel {{\mathsf{skip}}})$ with some successor $p{^{\scriptscriptstyle\prime}}$, whereas $(p \parallel {{\mathsf{skip}}}) \parallel {{\mathsf{skip}}}$ can do the step to $(p{^{\scriptscriptstyle\prime}}\parallel {{\mathsf{skip}}}) \parallel {{\mathsf{skip}}}$ only. This observation in principle reveals that the binary $\parallel$ is not associative at the level of small-step computations. Although the deviations might be regarded as small and technical, they nonetheless would lead to unnecessary complications if we were not able to resort to the ‘flat’ composition ${\parallel\!\!p_1, \ldots, p_n}$ instead.
### Environment steps
In addition to ${\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}$, the second integral component of the computational model is the already mentioned *environment* which can also perform indivisible steps via the ${\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{E}}}\:}$ relation defined as follows: $(p, \sigma) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{E}}}\:}(p{^{\scriptscriptstyle\prime}}, \sigma{^{\scriptscriptstyle\prime}})$ holds iff $p = p{^{\scriptscriptstyle\prime}}$. That is, an environment can arbitrarily modify the state of a configuration, leaving the program part unchanged.
Programs
--------
In line with the computational model, a *program* is constituted by a pair $(\rho, p)$ where $\rho$ is a retrieve function and $p$ is a term of type ${\mathtt{\mathcal{L}}_{\alpha}}$. It should be clear that if $p$ is jump-free then it does not actually need a retrieve function to perform any of its computation steps or, more precisely, performs same steps independently of how $\rho$ has been defined. Therefore we will mostly omit it in such cases for the sake of brevity. By contrast, if $p$ contains a jump, say ${{\mathsf{jump}}\:i}$, then the choice of a retrieve function can surely affect the behaviour: we could have ${\rho \vdash (p, \sigma) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(p_1, \sigma)}$ and ${\rho{^{\scriptscriptstyle\prime}}\vdash (p, \sigma) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(p_2, \sigma)}$ with potentially arbitrarily unrelated $p_1$ and $p_2$ if ${\rho\:i} \neq \rho{^{\scriptscriptstyle\prime}}\:i$.
Furthermore, note that with certain pairs $(\rho, p)$ we could achieve something intuitively comparable to programs comprising infinite source code. So, for instance, setting $p = {{\mathsf{jump}}\:0}$ and ${\rho\:i} = {{\mathsf{jump}}\:(i + 1)}$ in principle corresponds to the ‘program’, written below in a sort of pseudo assembly, $$\begin{array}{l l}
& {{\mathsf{jump}}\:0}\\
0: & {{\mathsf{jump}}\:1}\\
1: & {{\mathsf{jump}}\:2}\\
\ldots & \\
\end{array}$$ (that is, $i:{{\mathsf{jump}}\:i+1}$ are supposed to form separate blocks labelled by $i$) doing nothing except perpetual jumping forward through the code. Such creations can surely be regarded as a byproduct, and in the sequel we will only focus on pairs $(\rho, p)$ that are well-formed. Intuitively, $(\rho, p)$ shall be considered well-formed if all labels that are invoked in the process of its evaluation can be computed beforehand. Formally, we can first note that for any $p$ we can define the finite set of labels ${\mathcal{J}\!\mathit{umps}_{\mathit{local}}}\:p$ that occur in $p$. Further, given a set of labels $L$, we can define ${{\mathcal N}}_\rho \:L$ to be the set $\{i\:|\: \exists j\in L.\:i \in {\mathcal{J}\!\mathit{umps}_{\mathit{local}}}({\rho\:j})\}$, and consider the closure $${\mathcal{J}\!\mathit{umps}}(\rho, p) {\mathrel{\stackrel{\scriptscriptstyle{\mathit{def}}}{=}}}\bigcup_{n\ge0}{{\mathcal N}}^n_\rho({\mathcal{J}\!\mathit{umps}_{\mathit{local}}}\:p)$$ A program $(\rho, p)$ is then called *well-formed* if ${\mathcal{J}\!\mathit{umps}}(\rho, p)$ is finite.
Further, if there is some $\rho{^{\scriptscriptstyle\prime}}$ that coincides with $\rho$ on ${\mathcal{J}\!\mathit{umps}}(\rho, p)$ then ${\rho \vdash (p, \sigma) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(p{^{\scriptscriptstyle\prime}}, \sigma{^{\scriptscriptstyle\prime}})}$ implies ${\rho{^{\scriptscriptstyle\prime}}\vdash (p, \sigma) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(p{^{\scriptscriptstyle\prime}}, \sigma{^{\scriptscriptstyle\prime}})}$ for any $p{^{\scriptscriptstyle\prime}}, \sigma$ and $\sigma{^{\scriptscriptstyle\prime}}$. In other words, $\rho$ can arbitrarily be altered outside of ${\mathcal{J}\!\mathit{umps}}(\rho, p)$ without any effect on evaluations of $(\rho, p)$. In the particular case when $p$ is jump-free, it follows immediately from the definition that ${\mathcal{J}\!\mathit{umps}}(\rho,p) = \emptyset$ holds for any $\rho$, highlighting once more that a retrieve function does not matter in such cases.
Lastly, taking up the question of sequentiality, a program $(\rho, p)$ is called *sequential* if $p$ is locally sequential and, moreover, ${\rho\:i}$ is locally sequential for any label $i \in {\mathcal{J}\!\mathit{umps}}(\rho, p)$. Clearly, if $p$ is jump-free then $(\rho, p)$ is sequential iff $p$ is locally sequential, regardless which $\rho$ we take. Also note that by contrast to local sequentiality, program steps retain sequentiality, ${\rho \vdash (p, \sigma) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(p{^{\scriptscriptstyle\prime}}, \sigma{^{\scriptscriptstyle\prime}})}$ implies that $(\rho, p{^{\scriptscriptstyle\prime}})$ is sequential if $(\rho, p)$ is.
Potential computations of a program {#sub:pcs}
-----------------------------------
Having defined all possible one-step transformations on configurations, an *infinite potential computation* of a program $(\rho, p)$ comprises a sequence of configurations $${\mathit{sq}}= (p_0, \sigma_0), (p_1, \sigma_1),\ldots$$ where $p_0 = p$ and either ${\rho \vdash (p_{i}, \sigma_{i}) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(p_{i+1}, \sigma_{i+1})}$ or $(p_{i}, \sigma_{i}) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{E}}}\:}(p_{i+1}, \sigma_{i+1})$ holds for each $i \in \mathbb{N}$. Note that any program has at least one such computation, since any type has at least one inhabitant such that some environment step can always be performed.
In the sequel ${\mathit{sq}}_i$ will denote the $i$-th configuration of ${\mathit{sq}}$, whereas ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_i)}$ and ${{\mathsf{\mathcal{S}}}({\mathit{sq}}_i)}$ – the program and the state of this configuration, respectively.
Furthermore, let ${_{\scriptscriptstyle n}|{\mathit{sq}}}$ and ${^{\scriptscriptstyle n}|{\mathit{sq}}}$ denote the sequences obtained by taking the first $n$ configurations from ${\mathit{sq}}$, and by removing the first $n$ configurations from ${\mathit{sq}}$, respectively. If ${\mathit{sq}}$ is an infinite computation and $n > 0$ then ${_{\scriptscriptstyle n}|{\mathit{sq}}}$ yields a *finite potential computation* of $p$ $${\mathit{sq}}{^{\scriptscriptstyle\prime}}= (p_0, \sigma_0), \ldots, (p_{n-1}, \sigma_{n-1})$$ of length $n$, denoted by $|{\mathit{sq}}{^{\scriptscriptstyle\prime}}|$. In case a computation ${\mathit{sq}}$ is already finite, the prefix ${_{\scriptscriptstyle n}|{\mathit{sq}}}$ is a finite potential computation as well, provided $0<n\le|{\mathit{sq}}|$ holds. Similarly, the condition $i<|{\mathit{sq}}|$ must be provided when accessing the $i$-th configuration ${\mathit{sq}}_i$ in finite cases.
Moreover, when ${\mathit{sq}}$ is finite, for ${^{\scriptscriptstyle m}|{\mathit{sq}}}$ to be well-defined $m\le|{\mathit{sq}}|$ must be provided, but it yields a potential computation of ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_m)}$ only if $m<|{\mathit{sq}}|$ holds, as it would be empty otherwise. In case ${\mathit{sq}}$ is an infinite sequence, ${^{\scriptscriptstyle m}|{\mathit{sq}}}$ is one of its infinite suffixes. Generally, we have ${^{\scriptscriptstyle m}|{\mathit{sq}}_i} = {\mathit{sq}}_{i + m}$, the $i$-th configuration of the suffix is the $(i+m)$-th configuration of the original sequence.
Two finite computations ${\mathit{sq}}$ and ${\mathit{sq}}{^{\scriptscriptstyle\prime}}$ can be composed to a single computation of length $|{\mathit{sq}}| + |{\mathit{sq}}{^{\scriptscriptstyle\prime}}| - 1$ provided the last configuration of ${\mathit{sq}}$ is the same as the first of ${\mathit{sq}}{^{\scriptscriptstyle\prime}}$. In such cases we can drop the first configuration of ${\mathit{sq}}{^{\scriptscriptstyle\prime}}$ taking the suffix ${^{\scriptscriptstyle 1}| {\mathit{sq}}{^{\scriptscriptstyle\prime}}}$ (which could also be empty, just meaning that ${\mathit{sq}}$ is already the composition) and attach it to ${\mathit{sq}}$ by means of the first transition of ${\mathit{sq}}{^{\scriptscriptstyle\prime}}$.
By ${\lsemanticp\rsemantic^\omega_{\rho}}$ and ${\lsemanticp\rsemantic_{\rho}}$ we summarise the sets of all infinite and finite potential computations of $(\rho, p)$, respectively. Since computation steps of a jump-free $p$ do not depend on the choice of $\rho$, we will write ${\lsemanticp\rsemantic^\omega}$ and ${\lsemanticp\rsemantic}$ in such cases. It is worth noting that the explicit distinction between finite and infinite computations will be useful later on: the Hoare-style program logic (Section \[S:prog-log\]) will conveniently focus on finite computations, taking accounts of the infinite ones only implicitly. However, ${\lsemanticp\rsemantic^\omega}$ will be inevitable for reasoning about liveness (Section \[S:live\]).
Next section is devoted to program correspondences: a generic approach to semantic relations between programs, covering in particular all transformations, sketched in Section \[S:concept\].
Stepwise Correspondence between Programs {#S:pcorr}
========================================
Let $X$ be a relation of type ${\mathtt{\mathcal{L}}_{\alpha}} \times {\mathtt{\mathcal{L}}_{\beta}} \Rightarrow {\mathtt{bool}}$, where $\alpha$ and $\beta$ are state type abstractions. Further, let $r$ be of type $\alpha \times \beta \Rightarrow {\mathtt{bool}}$, a relation between the underlying states, and $\rho$ and $\rho{^{\scriptscriptstyle\prime}}$ – two retrieve functions of type ${\mathtt{nat}} \Rightarrow {\mathtt{\mathcal{L}}_{\alpha}}$ and ${\mathtt{nat}} \Rightarrow {\mathtt{\mathcal{L}}_{\beta}}$, respectively. Then $X$ is a *simulation* w.r.t. $\rho, \rho{^{\scriptscriptstyle\prime}}$ and $r$ if the following conditions hold:
1. if $(p, q) \in X$ and $(\sigma_1, \sigma_2) \in r$ then for any program step ${\rho{^{\scriptscriptstyle\prime}}\vdash (q, \sigma_2) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(q{^{\scriptscriptstyle\prime}}, \sigma{^{\scriptscriptstyle\prime}}_2)}$ there is a program step ${\rho \vdash (p, \sigma_1) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(p{^{\scriptscriptstyle\prime}}, \sigma{^{\scriptscriptstyle\prime}}_1)}$ such that $(p{^{\scriptscriptstyle\prime}}, q{^{\scriptscriptstyle\prime}}) \in X$ and $(\sigma{^{\scriptscriptstyle\prime}}_1, \sigma{^{\scriptscriptstyle\prime}}_2) \in r$,
2. if $({{\mathsf{skip}}}, q) \in X$ then $q = {{\mathsf{skip}}}$,
3. and if $(p, {{\mathsf{skip}}}) \in X$ then $p = {{\mathsf{skip}}}$.
\[def:pcorr\] Two programs $(\rho, p)$ and $(\rho{^{\scriptscriptstyle\prime}}, q)$ *correspond* w.r.t. $r$ if there exists some $X$ that contains the pair $(p, q)$ and is a simulation w.r.t. $\rho, \rho{^{\scriptscriptstyle\prime}}$ and $r$. This will be denoted by ${\rho, \rho{^{\scriptscriptstyle\prime}}\models p\sqsupseteq_{r}q}$. Furthermore, we say that $(\rho, p)$ and $(\rho{^{\scriptscriptstyle\prime}},q)$ *mutually correspond* w.r.t. $r$ if additionally ${\rho{^{\scriptscriptstyle\prime}}, \rho \models q\sqsupseteq_{{r^{\circ}}}p}$ holds, where ${r^{\circ}}$ is the converse of $r$. This, in turn, will be denoted by ${\rho, \rho{^{\scriptscriptstyle\prime}}\models p\approx_{r}q}$.
The notation ${\rho \models p\sqsupseteq_{r}q}$ will be used whenever the retrieve functions $\rho$ and $\rho{^{\scriptscriptstyle\prime}}$ are the same, and write ${\rho, \rho{^{\scriptscriptstyle\prime}}\models p\sqsupseteq_{}q}$ when $r$ is the identity relation on the underlying states. In line with the correspondence, we also write ${\rho, \rho{^{\scriptscriptstyle\prime}}\models p\approx_{}q}$ whenever $r$ is an identity, and ${\rho \models p\approx_{r}q}$ when $\rho$ and $\rho{^{\scriptscriptstyle\prime}}$ are the same.
Properties of correspondences {#Sb:corr_props}
-----------------------------
Firstly, notice that the singleton set $\{({{\mathsf{skip}}}, {{\mathsf{skip}}})\}$ is trivially a simulation with respect to any (appropriately typed) $\rho, \rho{^{\scriptscriptstyle\prime}}$ and $r$. Hence, ${\rho, \rho{^{\scriptscriptstyle\prime}}\models {{\mathsf{skip}}}\sqsupseteq_{r}{{\mathsf{skip}}}}$ holds, the correspondence relation is not empty for any $\rho, \rho{^{\scriptscriptstyle\prime}}$ and $r$. It also follows immediately from the definition that the inequalities ${\rho, \rho{^{\scriptscriptstyle\prime}}\models x\sqsupseteq_{r}{{\mathsf{skip}}}}$ and ${\rho, \rho{^{\scriptscriptstyle\prime}}\models {{\mathsf{skip}}}\sqsupseteq_{r}x}$ possess only one solution $x = {{\mathsf{skip}}}$.
\[thm:pcorr\_steps\] Assume ${\rho, \rho{^{\scriptscriptstyle\prime}}\models p\sqsupseteq_{r}q}$, $(\sigma_1, \sigma_2) \in r$ and ${\rho{^{\scriptscriptstyle\prime}}\vdash (q, \sigma_2) {\stackrel{\scriptscriptstyle n}{\mathop{\rightarrow}}_{\scriptscriptstyle{\mathcal{P}}}} (q{^{\scriptscriptstyle\prime}}, \sigma{^{\scriptscriptstyle\prime}}_2)}$ with $n \in \mathbb{N}$. Then there exist $p{^{\scriptscriptstyle\prime}}, \sigma{^{\scriptscriptstyle\prime}}_1$ such that ${\rho \vdash (p, \sigma_1) {\stackrel{\scriptscriptstyle n}{\mathop{\rightarrow}}_{\scriptscriptstyle{\mathcal{P}}}} (p{^{\scriptscriptstyle\prime}}, \sigma{^{\scriptscriptstyle\prime}}_1)}$ with ${\rho, \rho{^{\scriptscriptstyle\prime}}\models p{^{\scriptscriptstyle\prime}}\sqsupseteq_{r}q{^{\scriptscriptstyle\prime}}}$ and $(\sigma{^{\scriptscriptstyle\prime}}_1, \sigma{^{\scriptscriptstyle\prime}}_2) \in r$.
${\rho, \rho{^{\scriptscriptstyle\prime}}\models p\sqsupseteq_{r}q}$ gives us a simulation $X$ w.r.t. $\rho, \rho{^{\scriptscriptstyle\prime}}, r$. By induction on $n$ we obtain some $p{^{\scriptscriptstyle\prime}}, \sigma{^{\scriptscriptstyle\prime}}_1$ such that ${\rho \vdash (p, \sigma_1) {\stackrel{\scriptscriptstyle n}{\mathop{\rightarrow}}_{\scriptscriptstyle{\mathcal{P}}}} (p{^{\scriptscriptstyle\prime}}, \sigma{^{\scriptscriptstyle\prime}}_1)}$ with $(p{^{\scriptscriptstyle\prime}}, q{^{\scriptscriptstyle\prime}}) \in X$ and $(\sigma{^{\scriptscriptstyle\prime}}_1, \sigma{^{\scriptscriptstyle\prime}}_2) \in r$.
Note that the above statement is in general not true with ${\rho, \rho{^{\scriptscriptstyle\prime}}\models p\approx_{r}q}$ in place of ${\rho, \rho{^{\scriptscriptstyle\prime}}\models p\sqsupseteq_{r}q}$.
Next proposition shows how correspondences can be composed. There and in what follows, ${r \diamond s}$ denotes the relational composition $\{(a, b) \:|\: \exists c. \: (a, c) \in r \wedge (c, b) \in s\}$.
\[thm:ctrans\] If ${\rho_1, \rho_2 \models p_1\sqsupseteq_{r}p_2}$ and ${\rho_2, \rho_3 \models p_2\sqsupseteq_{s}p_3}$ then ${\rho_1, \rho_3 \models p_1\sqsupseteq_{{r \diamond s}}p_3}$.
The assumptions give us two simulations: $X_1$ w.r.t. $\rho_1, \rho_2, r$ containing $(p_1, p_2)$, and $X_2$ w.r.t. $\rho_2, \rho_3, s$ containing $(p_2, p_3)$. We use their composition ${X_1 \diamond X_2}$ containing $(p_1, p_3)$, as a witness to establish the conclusion: that ${X_1 \diamond X_2}$ is a simulation w.r.t. $\rho_1, \rho_3, {r \diamond s}$ follows from the simulation properties of $X_1$ and $X_2$.
With identity in place of $r$, the above statement assumes the following form.
\[thm:ctrans2\] If ${\rho_1, \rho_2 \models p_1\sqsupseteq_{}p_2}$ and ${\rho_2, \rho_3 \models p_2\sqsupseteq_{}p_3}$ then ${\rho_1, \rho_3 \models p_1\sqsupseteq_{}p_3}$.
The following proposition states that $\sqsupseteq$ is also reflexive and, hence, a preorder.
\[thm:crefl\] ${\rho \models p\sqsupseteq_{}p}$.
The relevant set is the identity on ${\mathtt{\mathcal{L}}_{\alpha}}$, which is a simulation with respect to the identity on $\alpha$.
Thus, ${\rho \models p\approx_{}q}$ is an equivalence relation and, in particular, allows us to exchange $p$ and $q$ arbitrarily when reasoning with correspondences.
Regarding more specific equivalences, one would expect the sequential composition to be associative. As mentioned in the previous section, the question of associativity of the parallel operator does not arise since we can apply it to any finite non-empty sequence of components. However, one would expect the parallel composition to be commutative.
\[thm:seq-assoc\] ${\rho \models p_1 ; p_2 ; p_3\approx_{}(p_1 ; p_2) ; p_3}$.
We show ${\rho \models p_1 ; p_2 ; p_3\sqsupseteq_{}(p_1 ; p_2) ; p_3}$, whereas the opposite direction follows similarly. The set $X {\mathrel{\stackrel{\scriptscriptstyle{\mathit{def}}}{=}}}\{(u ; p_2 ; p_3,(u ; p_2) ; p_3) \:|\: u \in \top\} \cup {\mathsf{id}}$ contains the pair $p_1 ; p_2 ; p_3$ and $(p_1 ; p_2) ; p_3$. Further, suppose ${\rho \vdash ((u ; p_2) ; p_3, \sigma) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(x, \sigma{^{\scriptscriptstyle\prime}})}$ for some $u$. If $u = {{\mathsf{skip}}}$ then $x = p_2 ; p_3$ and $\sigma{^{\scriptscriptstyle\prime}}= \sigma$, such that we can match this by ${\rho \vdash ({{\mathsf{skip}}}; p_2 ; p_3, \sigma) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(p_2 ; p_3, \sigma)}$. If $u \neq {{\mathsf{skip}}}$ then there is a step ${\rho \vdash (u, \sigma) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(u{^{\scriptscriptstyle\prime}}, \sigma{^{\scriptscriptstyle\prime}})}$ with some $u{^{\scriptscriptstyle\prime}}$ such that $x = (u{^{\scriptscriptstyle\prime}}; p_2) ; p_3$. Also in this case we have ${\rho \vdash (u ; p_2 ; p_3, \sigma) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(u{^{\scriptscriptstyle\prime}}; p_2 ; p_3, \sigma{^{\scriptscriptstyle\prime}})}$ with $(u{^{\scriptscriptstyle\prime}}; p_2 ; p_3, (u{^{\scriptscriptstyle\prime}}; p_2) ; p_3) \in X$.
\[thm:par-comm1\] Let $\pi$ be a permutation on $\{1, \ldots, m\}$ with $m > 0$, and assume $q_i = p_{\pi(i)}$ for all $i \in \{1, \ldots, m\}$. Then ${\rho \models \:{\parallel\!\!p_1, \ldots, p_m}\sqsupseteq_{}\:{\parallel\!\!q_1, \ldots, q_m}}$.
Let $X {\mathrel{\stackrel{\scriptscriptstyle{\mathit{def}}}{=}}}\{({\parallel\!\!\!u_{1}, \ldots, u_{m}}, {\parallel\!\!\!v_1, \ldots, v_m}) \:|\: \forall i \in \{1, \ldots, m\}.\: v_i = u_{\pi(i)}\} \cup
\{({{\mathsf{skip}}}, {{\mathsf{skip}}})\}$. Further, suppose ${\rho \vdash ({\parallel\!\!v_1, \ldots, v_m}, \sigma) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(x, \sigma{^{\scriptscriptstyle\prime}})}$ and $({\parallel\!\!u_1, \ldots, u_m}, {\parallel\!\!v_1, \ldots, v_m}) \in X$ hold. If $x$ is ${{\mathsf{skip}}}$ so are all of $v_1, \ldots, v_m$ and we are done. Otherwise there is some $i \in \{1, \ldots, m\}$ such that ${\rho \vdash ({\parallel\!\!v_1, \ldots, v_i, \ldots, v_m}, \sigma) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(x, \sigma{^{\scriptscriptstyle\prime}})}$ holds with $x =\: {\parallel\!\!v_1, \ldots, w, \ldots, v_m}$ due to a step ${\rho \vdash (v_i, \sigma) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(w, \sigma{^{\scriptscriptstyle\prime}})}$. Since $v_i = u_{\pi(i)}$, we can match it by ${\rho \vdash ({\parallel\!\!u_1, \ldots, u_{\pi(i)}, \ldots, u_m}, \sigma) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}({\parallel\!\!u_1, \ldots, w, \ldots, u_m}, \sigma{^{\scriptscriptstyle\prime}})}$.
\[thm:par-comm\] Let $\pi$ be a permutation on $\{1, \ldots, m\}$ with $m > 0$, and assume $q_i = p_{\pi(i)}$ for all $i \in \{1, \ldots, m\}$. Then ${\rho \models \:{\parallel\!\!p_1, \ldots, p_m}\approx_{}\:{\parallel\!\!q_1, \ldots, q_m}}$.
The direction ${\rho \models \:{\parallel\!\!\!p_1, \ldots, p_m}\sqsupseteq_{}\:{\parallel\!\!\!q_1, \ldots, q_m}}$ follows straight from the previous proposition. Turning the things around, we once more apply Proposition \[thm:par-comm1\], but now with $\pi^{-1}$ for $\pi$. This is sound since $p_i = p_{\pi(\pi^{-1}(i))} = q_{\pi^{-1}(i)}$ holds for all $i \in \{1, \ldots, m\}$.
Closure properties {#Sb:corr-rules}
------------------
One of the central goals of this section is to enable derivations of program correspondences in a syntax-driven manner.
\[thm:parallel-corr\] Assume $m > 0$ and ${\rho, \rho{^{\scriptscriptstyle\prime}}\models p_i\sqsupseteq_{r}q_i}$ for all $i$ with $1 \le i \le m$. Then ${\rho, \rho{^{\scriptscriptstyle\prime}}\models \:{\parallel\!\!p_1,\ldots, p_m}\sqsupseteq_{r}\:{\parallel\!\!q_1, \ldots,q_m}}$.
From the assumption we obtain simulations $X_1, \ldots, X_m$ w.r.t. $\rho, \rho{^{\scriptscriptstyle\prime}}, r$ such that $(p_i, q_i) \in X_i$ for all $i \in \{1, \ldots, m\}$. Then let $$X {\mathrel{\stackrel{\scriptscriptstyle{\mathit{def}}}{=}}}\{({\parallel\!\!u_1, \ldots, u_m}, {\parallel\!\!v_1, \ldots, v_m}) \:|\: \forall i \in \{1, \ldots, m\}. \: (u_i, v_i) \in X_i \} \cup
\{({{\mathsf{skip}}}, {{\mathsf{skip}}})\}$$ which contains the pair $({\parallel\!\!p_1, \ldots, p_m}, {\parallel\!\!q_1, \ldots, q_m})$. Further, to show that $X$ is a simulation w.r.t. $\!\rho, \rho{^{\scriptscriptstyle\prime}}, r$ suppose we have a step ${\rho{^{\scriptscriptstyle\prime}}\vdash ({\parallel\!\!v_1, \ldots, v_m}, \sigma_2) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(x, \sigma{^{\scriptscriptstyle\prime}}_2)}$ with $(\sigma_1, \sigma_2) \in r$ and $({\parallel\!\!u_1, \ldots, u_m}, {\parallel\!\!v_1, \ldots, v_m}) \in X$. If $x = {{\mathsf{skip}}}$ then $v_i = {{\mathsf{skip}}}$ for all $i \in \{1, \ldots, m\}$, and therefore $u_i = {{\mathsf{skip}}}$ for all $i \in \{1, \ldots, m\}$ as well, since $(u_i, v_i) \in X_i$. If $x \neq {{\mathsf{skip}}}$ then for some $i \in \{1, \ldots, m\}$ we have a step ${\rho{^{\scriptscriptstyle\prime}}\vdash (v_i, \sigma_2) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(v{^{\scriptscriptstyle\prime}}_i, \sigma{^{\scriptscriptstyle\prime}}_2)}$ such that $x = \:{\parallel\!\!\!v_1, \ldots, v{^{\scriptscriptstyle\prime}}_i, \ldots, v_m}$ holds. Since $(u_i, v_i) \in X_i$ and $(\sigma_1, \sigma_2) \in r$, we obtain some $u{^{\scriptscriptstyle\prime}}_i$ and $\sigma{^{\scriptscriptstyle\prime}}_1$ such that ${\rho \vdash (u_i, \sigma_1) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(u{^{\scriptscriptstyle\prime}}_i, \sigma{^{\scriptscriptstyle\prime}}_1)}$ holds with $(u{^{\scriptscriptstyle\prime}}_i, v{^{\scriptscriptstyle\prime}}_i) \in X_i$ and $(\sigma{^{\scriptscriptstyle\prime}}_1, \sigma{^{\scriptscriptstyle\prime}}_2) \in r$. This enables the transition $
{\rho \vdash ({\parallel\!\!\!u_1, \ldots, u_i, \ldots, u_m}, \sigma_1) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}({\parallel\!\!\!u_1, \ldots, u{^{\scriptscriptstyle\prime}}_i, \ldots, u_m}, \sigma{^{\scriptscriptstyle\prime}}_1)}$ with $({\parallel\!\!u_1, \ldots, u{^{\scriptscriptstyle\prime}}_i, \ldots, u_m}, {\parallel\!\!v_1, \ldots, v{^{\scriptscriptstyle\prime}}_i, \ldots, v_m}) \in X$.
\[thm:seq-corr\] If ${\rho, \rho{^{\scriptscriptstyle\prime}}\models p_1\sqsupseteq_{r}q_1}$ and ${\rho, \rho{^{\scriptscriptstyle\prime}}\models p_2\sqsupseteq_{r}q_2}$ then ${\rho, \rho{^{\scriptscriptstyle\prime}}\models p_1 ; p_2\sqsupseteq_{r}q_1 ; q_2}$.
From the assumptions we get simulations $X_1, X_2$ w.r.t. $\rho, \rho{^{\scriptscriptstyle\prime}}, r$ with $(p_1, q_1) \in X_1$ and $(p_2, q_2) \in X_2$. Then let $X$ be defined by $\{(u ; p_2, v ; q_2) \:|\: (u, v) \in X_1\} \cup X_2$, such that $(p_1 ; p_2, q_1 ; q_2) \in X$. Further, let ${\rho{^{\scriptscriptstyle\prime}}\vdash (v ; q_2, \sigma_2) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(x, \sigma{^{\scriptscriptstyle\prime}}_2)}$ with $(\sigma_1, \sigma_2) \in r$ and $(u, v) \in X_1$. If $v = {{\mathsf{skip}}}$ then $u = {{\mathsf{skip}}}$ and $x = q_2$ and $\sigma{^{\scriptscriptstyle\prime}}_2 = \sigma_2$, such that we are done, since $(p_2, q_2) \in X_2$ and $X_2 \subseteq X$.
If $v \neq {{\mathsf{skip}}}$ then we have a step ${\rho{^{\scriptscriptstyle\prime}}\vdash (v, \sigma_2) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(v{^{\scriptscriptstyle\prime}}, \sigma{^{\scriptscriptstyle\prime}}_2)}$ such that $x = v{^{\scriptscriptstyle\prime}}; p{^{\scriptscriptstyle\prime}}_2$. Since $(u, v) \in X_1$, we also have a step ${\rho \vdash (u, \sigma_1) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(u{^{\scriptscriptstyle\prime}}, \sigma{^{\scriptscriptstyle\prime}}_1)}$ with $(\sigma{^{\scriptscriptstyle\prime}}_1, \sigma{^{\scriptscriptstyle\prime}}_2) \in r$ and $(u{^{\scriptscriptstyle\prime}}, v{^{\scriptscriptstyle\prime}}) \in X_1$. Therefore, ${\rho \vdash (u ; p_2, \sigma_1) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(u{^{\scriptscriptstyle\prime}}; p_2, \sigma{^{\scriptscriptstyle\prime}}_1)}$ with $(u{^{\scriptscriptstyle\prime}}; p_2, x) \in X$.
In the following three propositions we also have to take accounts of branching and blocking conditions.
\[thm:cond-corr\] Assume
1. ${\rho, \rho{^{\scriptscriptstyle\prime}}\models p_1\sqsupseteq_{r}p_2}$,
2. ${\rho, \rho{^{\scriptscriptstyle\prime}}\models q_1\sqsupseteq_{r}q_2}$,
3. $\sigma_1 \in C_1$ iff $\sigma_2 \in C_2$ for any $(\sigma_1, \sigma_2) \in r$.
Then ${\rho, \rho{^{\scriptscriptstyle\prime}}\models {{\mathsf{if}}\:C_1\;{\mathsf{then}}\;p_1\;{\mathsf{else}}\;q_1\;{\mathsf{fi}}}\sqsupseteq_{r}{{\mathsf{if}}\:C_2\;{\mathsf{then}}\;p_2\;{\mathsf{else}}\;q_2\;{\mathsf{fi}}}}$.
From (1) and (2) we get simulations $X_p, X_q$ w.r.t. $\rho, \rho{^{\scriptscriptstyle\prime}}, r$ with $(p_1, p_2) \in X_p$ and $(q_1, q_2) \in X_q$. Then let $X {\mathrel{\stackrel{\scriptscriptstyle{\mathit{def}}}{=}}}X_p \cup X_q \cup \{({{\mathsf{if}}\:C_1\;{\mathsf{then}}\;p_1\;{\mathsf{else}}\;q_1\;{\mathsf{fi}}},{{\mathsf{if}}\:C_2\;{\mathsf{then}}\;p_2\;{\mathsf{else}}\;q_2\;{\mathsf{fi}}})\}$.
First, note that as the union of simulations, $X_p \cup X_q$ is again a simulation. Then for the remaining case, suppose we have ${\rho{^{\scriptscriptstyle\prime}}\vdash ({{\mathsf{if}}\:C_2\;{\mathsf{then}}\;p_2\;{\mathsf{else}}\;q_2\;{\mathsf{fi}}}, \sigma_2) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(x, \sigma{^{\scriptscriptstyle\prime}}_2)}$ and $(\sigma_1, \sigma_2) \in r$. If $\sigma_2 \in C_2$ then $x = p_2$ and, since $\sigma_1 \in C_1$ by (3), we get the transition ${\rho \vdash ({{\mathsf{if}}\:C_1\;{\mathsf{then}}\;p_1\;{\mathsf{else}}\;q_1\;{\mathsf{fi}}}, \sigma_1) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(p_1, \sigma{^{\scriptscriptstyle\prime}}_1)}$ with $(\sigma{^{\scriptscriptstyle\prime}}_1, \sigma{^{\scriptscriptstyle\prime}}_2) \in r$. Likewise with $\sigma_2 \notin C_2$.
\[thm:while-corr\] Assume
1. ${\rho, \rho{^{\scriptscriptstyle\prime}}\models p_1\sqsupseteq_{r}p_2}$,
2. ${\rho, \rho{^{\scriptscriptstyle\prime}}\models q_1\sqsupseteq_{r}q_2}$,
3. $\sigma_1 \in C_1$ iff $\sigma_2 \in C_2$ for any $(\sigma_1, \sigma_2) \in r$.
Then ${\rho, \rho{^{\scriptscriptstyle\prime}}\models {{\mathsf{while}}\:C_1\;{\mathsf{do}}\;p_1\;{\mathsf{subsequently}}\;q_1\;{\mathsf{od}}}\sqsupseteq_{r}{{\mathsf{while}}\:C_2\;{\mathsf{do}}\;p_2\;{\mathsf{subsequently}}\;q_2\;{\mathsf{od}}}}$.
From (1) and (2) we get simulations $X_p, X_q$ w.r.t. $\rho, \rho{^{\scriptscriptstyle\prime}}, r$ with $(p_1, p_2) \in X_p$ and $(q_1, q_2) \in X_q$. The relevant set of pairs is, however, slightly more involved than in the previous proof: $$X {\mathrel{\stackrel{\scriptscriptstyle{\mathit{def}}}{=}}}\{(L,R), ({{\mathsf{skip}}}\: ;\: L, {{\mathsf{skip}}}\: ;\: R)\} \cup \{(u\: ;\: {{\mathsf{skip}}}\: ;\: L, v\: ;\: {{\mathsf{skip}}}\: ;\: R) \:|\: (u, v) \in X_p\} \cup X_q$$ where $L$ is ${{\mathsf{while}}\:C_1\;{\mathsf{do}}\;p_1\;{\mathsf{subsequently}}\;q_1\;{\mathsf{od}}}$, whereas $R$ – ${{\mathsf{while}}\:C_2\;{\mathsf{do}}\;p_2\;{\mathsf{subsequently}}\;q_2\;{\mathsf{od}}}$. Further, assume $(\sigma_1, \sigma_2) \in r$ and consider the following cases.
First suppose we have a step ${\rho{^{\scriptscriptstyle\prime}}\vdash (R, \sigma_2) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(x, \sigma{^{\scriptscriptstyle\prime}}_2)}$. If $\sigma_2 \in C_2$ then $x = p_2\: ;\: {{\mathsf{skip}}}\: ;\: R$ and $\sigma{^{\scriptscriptstyle\prime}}_2 = \sigma_2$. This can be matched by ${\rho \vdash (L, \sigma_1) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(p_1\: ;\: {{\mathsf{skip}}}\: ;\: L, \sigma_1)}$, since $\sigma_1 \in C_1$ by (3), and $((p_1 \: ; \: {{\mathsf{skip}}}\: ; \: L, p_2 \: ; \: {{\mathsf{skip}}}\: ; \: R) \in X$. If $\sigma_2 \notin C_2$ then $x = q_2$ and $\sigma{^{\scriptscriptstyle\prime}}_2 = \sigma_2$, which can be matched by ${\rho \vdash (L, \sigma_1) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(q_1, \sigma_1)}$.
Further, the case ${\rho{^{\scriptscriptstyle\prime}}\vdash ({{\mathsf{skip}}}\: ;\: R, \sigma_2) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(R, \sigma_2)}$ is clear, since the step can be matched by ${\rho \vdash ({{\mathsf{skip}}}\: ; \: L, \sigma_1) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(L, \sigma_1)}$.
Finally, suppose we have ${\rho{^{\scriptscriptstyle\prime}}\vdash (v\: ; \: {{\mathsf{skip}}}\: ;\: R, \sigma_2) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(x, \sigma{^{\scriptscriptstyle\prime}}_2)}$ and $(u, v) \in X_p$. If $v = {{\mathsf{skip}}}$ then $u = {{\mathsf{skip}}}$, which leads to essentially the same situation as in the previous case. Otherwise there exists some $v{^{\scriptscriptstyle\prime}}$ such that ${\rho{^{\scriptscriptstyle\prime}}\vdash (v, \sigma_2) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(v{^{\scriptscriptstyle\prime}}, \sigma{^{\scriptscriptstyle\prime}}_2)}$ and $x = v{^{\scriptscriptstyle\prime}}\: ; \: {{\mathsf{skip}}}\: ;\: R$. Since $X_p$ is a simulation, we also have a transition ${\rho \vdash (u, \sigma_1) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(u{^{\scriptscriptstyle\prime}}, \sigma{^{\scriptscriptstyle\prime}}_1)}$ with $(u{^{\scriptscriptstyle\prime}}, v{^{\scriptscriptstyle\prime}}) \in X_p$ and $(\sigma{^{\scriptscriptstyle\prime}}_1, \sigma{^{\scriptscriptstyle\prime}}_2) \in r$. This, in turn, enables the matching step ${\rho \vdash (u\: ; \: {{\mathsf{skip}}}\: ;\: L, \sigma_1) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(u{^{\scriptscriptstyle\prime}}\: ;\: {{\mathsf{skip}}}\: ; \:L, \sigma{^{\scriptscriptstyle\prime}}_1)}$.
\[thm:await-corr\] Assume
1. ${\rho, \rho{^{\scriptscriptstyle\prime}}\models p_1\sqsupseteq_{r}p_2}$,
2. $\sigma_2 \in C_2$ implies $\sigma_1 \in C_1$ for any $(\sigma_1, \sigma_2) \in r$.
Then ${\rho, \rho{^{\scriptscriptstyle\prime}}\models {{\mathsf{await}}\:C_1\;{\mathsf{do}}\;p_1\;{\mathsf{od}}}\sqsupseteq_{r}{{\mathsf{await}}\:C_2\;{\mathsf{do}}\;p_2\;{\mathsf{od}}}}$.
In order to establish ${\rho, \rho{^{\scriptscriptstyle\prime}}\models {{\mathsf{await}}\:C_1\;{\mathsf{do}}\;p_1\;{\mathsf{od}}}\sqsupseteq_{r}{{\mathsf{await}}\:C_2\;{\mathsf{do}}\;p_2\;{\mathsf{od}}}}$ let $X$ be defined by $\{({{\mathsf{await}}\:C_1\;{\mathsf{do}}\;p_1\;{\mathsf{od}}}, {{\mathsf{await}}\:C_2\;{\mathsf{do}}\;p_2\;{\mathsf{od}}}), ({{\mathsf{skip}}}, {{\mathsf{skip}}})\}$. Further, suppose we have some $(\sigma_1, \sigma_2) \in r$ and a step ${\rho{^{\scriptscriptstyle\prime}}\vdash ({{\mathsf{await}}\:C_2\;{\mathsf{do}}\;p_2\;{\mathsf{od}}}, \sigma_2) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(x, \sigma{^{\scriptscriptstyle\prime}}_2)}$. This entails $\sigma_2 \in C_2$, ${\rho{^{\scriptscriptstyle\prime}}\vdash (p_2, \sigma_2) {\stackrel{\scriptscriptstyle\star}{\mathop{\rightarrow}}_{\scriptscriptstyle{\mathcal{P}}}}({{\mathsf{skip}}}, \sigma{^{\scriptscriptstyle\prime}}_2)}$ and $x = {{\mathsf{skip}}}$. Then using (1), from Proposition \[thm:pcorr\_steps\] we obtain some $\sigma{^{\scriptscriptstyle\prime}}_1$ such that ${\rho \vdash (p_1, \sigma_1) {\stackrel{\scriptscriptstyle\star}{\mathop{\rightarrow}}_{\scriptscriptstyle{\mathcal{P}}}}({{\mathsf{skip}}}, \sigma{^{\scriptscriptstyle\prime}}_1)}$ and $(\sigma{^{\scriptscriptstyle\prime}}_1, \sigma{^{\scriptscriptstyle\prime}}_2) \in r$. Since $\sigma_1 \in C_1$ holds by (2), this enables the matching step ${\rho \vdash ({{\mathsf{await}}\:C_1\;{\mathsf{do}}\;p_1\;{\mathsf{od}}}, \sigma_1) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}({{\mathsf{skip}}}, \sigma{^{\scriptscriptstyle\prime}}_1)}$.
Lastly, the remaining two propositions show how correspondences between indivisible steps can be established.
\[thm:basic-corr\] If $(f\: \sigma_1, g\:\sigma_2) \in r$ for all $(\sigma_1, \sigma_2) \in r$ then ${\rho, \rho{^{\scriptscriptstyle\prime}}\models {{\mathsf{basic}}}\:f\sqsupseteq_{r}{{\mathsf{basic}}}\:g}$.
From the assumption we can infer that the set $\{({{\mathsf{basic}}}\;f, {{\mathsf{basic}}}\;g), ({{\mathsf{skip}}}, {{\mathsf{skip}}})\}$ is a simulation w.r.t. $\rho, \rho{^{\scriptscriptstyle\prime}}, r$.
\[thm:await-corrL\] Assume that for any $(\sigma_1, \sigma_2) \in r$ there exists $\sigma{^{\scriptscriptstyle\prime}}_1$ such that $(\sigma{^{\scriptscriptstyle\prime}}_1, f\:\sigma_2) \in r$ and ${\rho \vdash (p, \sigma_1) {\stackrel{\scriptscriptstyle\star}{\mathop{\rightarrow}}_{\scriptscriptstyle{\mathcal{P}}}}({{\mathsf{skip}}}, \sigma{^{\scriptscriptstyle\prime}}_1)}$ hold. Then ${\rho, \rho{^{\scriptscriptstyle\prime}}\models \langle p \rangle\sqsupseteq_{r}{{\mathsf{basic}}}\:f}$.
Let $X {\mathrel{\stackrel{\scriptscriptstyle{\mathit{def}}}{=}}}\{(\langle p \rangle, {{\mathsf{basic}}}\:f), ({{\mathsf{skip}}}, {{\mathsf{skip}}})\}$. In order to establish that $X$ is a simulation w.r.t. $\rho, \rho{^{\scriptscriptstyle\prime}}, r$, suppose we have some $(\sigma_1, \sigma_2) \in r$, such that from the assumption we further obtain $\sigma{^{\scriptscriptstyle\prime}}_1$ with $(\sigma{^{\scriptscriptstyle\prime}}_1, f\:\sigma_2) \in r$ and ${\rho \vdash (p, \sigma_1) {\stackrel{\scriptscriptstyle\star}{\mathop{\rightarrow}}_{\scriptscriptstyle{\mathcal{P}}}}({{\mathsf{skip}}}, \sigma{^{\scriptscriptstyle\prime}}_1)}$. Since ${\rho \vdash (p, \sigma_1) {\stackrel{\scriptscriptstyle\star}{\mathop{\rightarrow}}_{\scriptscriptstyle{\mathcal{P}}}}({{\mathsf{skip}}}, \sigma{^{\scriptscriptstyle\prime}}_1)}$ implies ${\rho \vdash (\langle p \rangle, \sigma_1) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}({{\mathsf{skip}}}, \sigma{^{\scriptscriptstyle\prime}}_1)}$, we are done.
Sequential normalisation
------------------------
Section \[S:concept\] pointed out the necessity of code restructurings prior to transformations of conditional and ${\mathsf{while}}$-statements into corresponding representations using jumps. Next two propositions justify these restructurings semantically.
\[thm:cond-norm1\] The following equivalence holds: $${\rho \models {{\mathsf{if}}\:C\;{\mathsf{then}}\;p_1\;{\mathsf{else}}\;p_2\;{\mathsf{fi}}}; q\approx_{}{{\mathsf{if}}\:C\;{\mathsf{then}}\;p_1; q\;{\mathsf{else}}\;p_2; q\;{\mathsf{fi}}}}.$$
We establish ${\rho \models {{\mathsf{if}}\:C\;{\mathsf{then}}\;p_1\;{\mathsf{else}}\;p_2\;{\mathsf{fi}}}\: ; q\sqsupseteq_{}{{\mathsf{if}}\:C\;{\mathsf{then}}\;p_1 ; q\;{\mathsf{else}}\;p_2 ; q\;{\mathsf{fi}}}}$ by means of the set $\{({{\mathsf{if}}\:C\;{\mathsf{then}}\;p_1\;{\mathsf{else}}\;p_2\;{\mathsf{fi}}}\: ; q, {{\mathsf{if}}\:C\;{\mathsf{then}}\;p_1; q\;{\mathsf{else}}\;p_2;q\;{\mathsf{fi}}})\} \cup {\mathsf{id}}$. Suppose we have a step ${\rho \vdash ({{\mathsf{if}}\:C\;{\mathsf{then}}\;p_1; q\;{\mathsf{else}}\;p_2;q\;{\mathsf{fi}}}, \sigma) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(x, \sigma{^{\scriptscriptstyle\prime}})}$. If $\sigma \in C$ then $x = p_1;q$ and $\sigma{^{\scriptscriptstyle\prime}}= \sigma$. This enables the matching step ${\rho \vdash ({{\mathsf{if}}\:C\;{\mathsf{then}}\;p_1\;{\mathsf{else}}\;p_2\;{\mathsf{fi}}}\: ; q, \sigma) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(p_1;q, \sigma)}$. The case $\sigma \notin C$ is symmetric and the opposite direction ${\rho \models {{\mathsf{if}}\:C\;{\mathsf{then}}\;p_1;q\;{\mathsf{else}}\;p_2;q\;{\mathsf{fi}}}\sqsupseteq_{}{{\mathsf{if}}\:C\;{\mathsf{then}}\;p_1\;{\mathsf{else}}\;p_2\;{\mathsf{fi}}}\: ;q}$ can be concluded likewise.
\[thm:while-norm1\] The following equivalence holds: $${\rho \models {{\mathsf{while}}\:C\;{\mathsf{do}}\;p_1\;{\mathsf{subsequently}}\;p_2\;{\mathsf{od}}} ; q\approx_{}{{\mathsf{while}}\:C\;{\mathsf{do}}\;p_1\;{\mathsf{subsequently}}\;p_2 ; q\;{\mathsf{od}}}}.$$
As in the previous proof, we first show the direction ${\rho \models L\sqsupseteq_{}R}$ where $L$ abbreviates ${{\mathsf{while}}\:C\;{\mathsf{do}}\;p_1\;{\mathsf{subsequently}}\;p_2\;{\mathsf{od}}}\: ; q$ and $R$ – ${{\mathsf{while}}\:C\;{\mathsf{do}}\;p_1\;{\mathsf{subsequently}}\;p_2;q\;{\mathsf{od}}}$. This is done by means of the set $X {\mathrel{\stackrel{\scriptscriptstyle{\mathit{def}}}{=}}}\{({{\mathsf{skip}}}; L, {{\mathsf{skip}}}; R), (L, R)\} \cup \{(v ; {{\mathsf{skip}}}; L, v ; {{\mathsf{skip}}}; R) \:|\: v \in \top \} \cup {\mathsf{id}}$.
The case with $({{\mathsf{skip}}}; L, {{\mathsf{skip}}}; R)$ is clear. Next, suppose ${\rho \vdash (R, \sigma) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(x, \sigma{^{\scriptscriptstyle\prime}})}$. If $\sigma \in C$ then $x = p_1 ;{{\mathsf{skip}}};R$ and $\sigma{^{\scriptscriptstyle\prime}}= \sigma$. Then the step from $(L, \sigma)$ to the configuration $(p_1;{{\mathsf{skip}}};L, \sigma)$ matches that with $(p_1;{{\mathsf{skip}}};L, x) \in X$. If $\sigma \notin C$ then $x = p_2 ; q$ and $\sigma{^{\scriptscriptstyle\prime}}= \sigma$, such that we can make a step from $(L, \sigma)$ to exactly the same configuration.
Finally, suppose ${\rho \vdash (v ; {{\mathsf{skip}}}; R, \sigma) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(x, \sigma{^{\scriptscriptstyle\prime}})}$. If $v = {{\mathsf{skip}}}$ then $x = {{\mathsf{skip}}}; R$ and $\sigma{^{\scriptscriptstyle\prime}}= \sigma$, and we accordingly have ${\rho \vdash ({{\mathsf{skip}}}; {{\mathsf{skip}}}; L, \sigma) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}({{\mathsf{skip}}}; L, \sigma)}$. If $v \neq {{\mathsf{skip}}}$ then there is some $v{^{\scriptscriptstyle\prime}}$ with ${\rho \vdash (v, \sigma) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(v{^{\scriptscriptstyle\prime}}, \sigma{^{\scriptscriptstyle\prime}})}$ such that $x = v{^{\scriptscriptstyle\prime}}; {{\mathsf{skip}}}; R$. This enables the step ${\rho \vdash (v ; {{\mathsf{skip}}}; L, \sigma) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(v{^{\scriptscriptstyle\prime}}; {{\mathsf{skip}}}; L, \sigma{^{\scriptscriptstyle\prime}})}$.
The opposite direction, ${\rho \models R\sqsupseteq_{}L}$, follows in a symmetric way.
Note that these equivalences subsume ${\rho \models {{\mathsf{if}}\:C\;{\mathsf{then}}\;p\;{\mathsf{fi}}}\: ; q\approx_{}{{\mathsf{if}}\:C\;{\mathsf{then}}\;p \:; q\;{\mathsf{else}}\;{{\mathsf{skip}}}\:; q\;{\mathsf{fi}}}}$ and ${\rho \models {{\mathsf{while}}\:C\;{\mathsf{do}}\;p\;{\mathsf{od}}}\: ; q\approx_{}{{\mathsf{while}}\:C\;{\mathsf{do}}\;p\;{\mathsf{subsequently}}\;{{\mathsf{skip}}}\:; q\;{\mathsf{od}}}}$.
Replacing conditional and while-statements by jumps
---------------------------------------------------
\[thm:cond-norm2\] The following equivalence holds: $${\rho \models {{\mathsf{if}}\:C\;{\mathsf{then}}\;p\;{\mathsf{else}}\;{\rho\:j}\;{\mathsf{fi}}}\approx_{}{{\mathsf{cjump}}\:{\neg C}\;{\mathsf{to}}\:j\;{\mathsf{otherwise}}\;p\;{\mathsf{end}}}}.$$
Taking the set of pairs $\{({{\mathsf{if}}\:C\;{\mathsf{then}}\;p\;{\mathsf{else}}\;{\rho\:j}\;{\mathsf{fi}}},{{\mathsf{cjump}}\:{\neg C}\;{\mathsf{to}}\:j\;{\mathsf{otherwise}}\;p\;{\mathsf{end}}})\} \cup {\mathsf{id}}$ for the $\sqsupseteq$-direction and $\{({{\mathsf{cjump}}\:{\neg C}\;{\mathsf{to}}\:j\;{\mathsf{otherwise}}\;p\;{\mathsf{end}}}, {{\mathsf{if}}\:C\;{\mathsf{then}}\;p\;{\mathsf{else}}\;{\rho\:j}\;{\mathsf{fi}}})\} \cup {\mathsf{id}}$ for the opposite.
Next proof is the only motivation for the extra ${{\mathsf{skip}}}$ in the rule ‘While-True’ (Figure \[fig:pstep\]).
\[thm:while-norm2\] Assume
1. ${\rho\:i} = {{\mathsf{cjump}}\:{\neg C}\;{\mathsf{to}}\:j\;{\mathsf{otherwise}}\;p;{{\mathsf{jump}}\:i}\;{\mathsf{end}}}$,
2. ${\rho\:j} = q$.
Then ${\rho \models {{\mathsf{while}}\:C\;{\mathsf{do}}\;p\;{\mathsf{subsequently}}\;q\;{\mathsf{od}}}\approx_{}{{\mathsf{cjump}}\:{\neg C}\;{\mathsf{to}}\:j\;{\mathsf{otherwise}}\;p;{{\mathsf{jump}}\:i}\;{\mathsf{end}}}}$.
A set of pairs to establish the $\sqsupseteq$-direction is $$X {\mathrel{\stackrel{\scriptscriptstyle{\mathit{def}}}{=}}}\{(L, {\rho\:i}), ({{\mathsf{skip}}}; L, {{\mathsf{jump}}\:i})\} \cup \{(u ; {{\mathsf{skip}}}; L, u ; {{\mathsf{jump}}\:i}\: | \: u \in \top \} \cup {\mathsf{id}}$$ where $L$ abbreviates ${{\mathsf{while}}\:C\;{\mathsf{do}}\;p\;{\mathsf{subsequently}}\;q\;{\mathsf{od}}}$.
First, suppose we have a transition ${\rho \vdash ({\rho\:i}, \sigma) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(x, \sigma{^{\scriptscriptstyle\prime}})}$ with some $x, \sigma$ and $\sigma{^{\scriptscriptstyle\prime}}$. If $\sigma \in C$ then $x = p;{{\mathsf{jump}}\:i}$ and $\sigma{^{\scriptscriptstyle\prime}}= \sigma$. We can match that by ${\rho \vdash (L, \sigma) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(p ; {{\mathsf{skip}}}; L, \sigma)}$ since $(p ; {{\mathsf{skip}}}; L, p; {{\mathsf{jump}}\:i}) \in X$. If $\sigma \notin C$ then $x = {\rho\:j} = q$ and $\sigma{^{\scriptscriptstyle\prime}}= \sigma$, such that we can make a step from $(L, \sigma)$ to exactly the same configuration.
Next, if we have a transition ${\rho \vdash ({{\mathsf{jump}}\:i}, \sigma) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}({\rho\:i}, \sigma)}$ then this is matched by the only possible transition ${\rho \vdash ({{\mathsf{skip}}}; L, \sigma) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(L, \sigma)}$, since $(L, \rho(i)) \in X$.
Finally, suppose ${\rho \vdash (u; {{\mathsf{jump}}\:i}, \sigma) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(x, \sigma{^{\scriptscriptstyle\prime}})}$ with some $u$. If $u = {{\mathsf{skip}}}$ then $x = {{\mathsf{jump}}\:i}$ and $\sigma{^{\scriptscriptstyle\prime}}= \sigma$, which is matched by ${\rho \vdash ({{\mathsf{skip}}}; {{\mathsf{skip}}}; L, \sigma) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}({{\mathsf{skip}}}; L, \sigma)}$. Otherwise, there is a step ${\rho \vdash (u, \sigma) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(u{^{\scriptscriptstyle\prime}}, \sigma{^{\scriptscriptstyle\prime}})}$ with some $u{^{\scriptscriptstyle\prime}}$ such that $x = u{^{\scriptscriptstyle\prime}}; {{\mathsf{jump}}\:i}$. This enables the transition ${\rho \vdash (u ; {{\mathsf{skip}}}; L, \sigma) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(u{^{\scriptscriptstyle\prime}}; {{\mathsf{skip}}}; L, \sigma{^{\scriptscriptstyle\prime}})}$ with $(u{^{\scriptscriptstyle\prime}}; {{\mathsf{skip}}}; L, u{^{\scriptscriptstyle\prime}};{{\mathsf{jump}}\:i}) \in X$.
The opposite direction can be shown likewise.
Note that the aim of the above proposition is merely to convey the basic principle behind the transformation. In particular, the statement does not take into account that $p;{{\mathsf{jump}}\:i}$ on the [*rhs*]{} can, in turn, become subject to the sequential normalisation.
Conditions on Potential Computations {#S:pcs-props}
====================================
This section takes up the topic of potential computations and puts program correspondence in that context. We start with the conditions allowing us to systematically restrict the set of potential computations of a program.
Let $R$ be a state relation. Then the *environment condition* ${\mathcal{E}\!\mathit{nv}^\omega\:R}$ comprises the set of *infinite* potential computations ${\mathit{sq}}$ satisfying $({{\mathsf{\mathcal{S}}}({\mathit{sq}}_{i})}, {{\mathsf{\mathcal{S}}}({\mathit{sq}}_{i+1})}) \in R$ for any $i \in \mathbb{N}$ with ${\mathit{sq}}_{i} {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{E}}}\:}{\mathit{sq}}_{i+1}$.
We need to recast the definition to finite computations.
Let $R$ be a state relation. Then the *environment condition* ${\mathcal{E}\!\mathit{nv}\:R}$ comprises the set of *finite* potential computations ${\mathit{sq}}$ satisfying $({{\mathsf{\mathcal{S}}}({\mathit{sq}}_{i})}, {{\mathsf{\mathcal{S}}}({\mathit{sq}}_{i+1})}) \in R$ for any $i < |{\mathit{sq}}| -1$ with ${\mathit{sq}}_{i} {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{E}}}\:}{\mathit{sq}}_{i+1}$.
The reason for this distinction is the following. The set of *actual computations* ( [@STIRLING1988347]) of a program is a subset of its potential computations and comprises those without any environment steps. Now, consider ${\mathcal{E}\!\mathit{nv}^\omega\:\bot}$ and ${\mathcal{E}\!\mathit{nv}\:\bot}$, where $\bot$ denotes the empty relation – the opposite to the already introduced $\top$. Then ${\mathcal{E}\!\mathit{nv}^\omega\:\bot}$ specifies all infinite actual computations, whereas ${\mathcal{E}\!\mathit{nv}\:\bot}$ – all finite. There are, however, terminating programs that do have only finite actual computations, which consequently form a subset of ${\mathcal{E}\!\mathit{nv}\:\bot}$ but not of ${\mathcal{E}\!\mathit{nv}^\omega\:\bot}$. In other words, we do not get ${\mathcal{E}\!\mathit{nv}\:\bot}$ just by prefixing ${\mathcal{E}\!\mathit{nv}^\omega\:\bot}$. Nonetheless, for the sake of brevity, in the sequel only ${\mathcal{E}\!\mathit{nv}\:R}$ will be used, as it will be clear from the context which version is meant.
Let $G$ be a state relation. Then the *program condition* ${\mathcal{P}\!\mathit{rog}\:G}$ comprises the set of infinite potential computations ${\mathit{sq}}$ satisfying $({{\mathsf{\mathcal{S}}}({\mathit{sq}}_{i})}, {{\mathsf{\mathcal{S}}}({\mathit{sq}}_{i+1})}) \in G$ for any $i \in \mathbb{N}$ with ${\rho \vdash {\mathit{sq}}_{i} {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}{\mathit{sq}}_{i+1}}$, as well as all finite non-empty prefixes thereof.
Let $C$ be a state predicate. A (finite or infinite) potential computation ${\mathit{sq}}$ satisfies the *input condition* ${\mathcal{I}\!\mathit{n}\:C}$ if ${{\mathsf{\mathcal{S}}}({\mathit{sq}}_0)} \in C$. Furthermore, a (finite or infinite) potential computation ${\mathit{sq}}$, where for the first index $i$ with ${\mathit{sq}}_i = ({{\mathsf{skip}}}, \sigma)$, if such exists, we have $\sigma \in C$, satisfies the *output condition* ${\mathcal{O}\!\mathit{ut}\:C}$.
Note that the definition of output conditions is, in essence, the same as in [@STIRLING1988347].
Program correspondence and finite potential computations {#Sb:corr_pcs}
--------------------------------------------------------
The following proposition resorts to environment conditions in order to describe how finite potential computations can be replayed along program correspondences.
\[thm:corr-sim\] Assume ${\rho, \rho{^{\scriptscriptstyle\prime}}\models p\sqsupseteq_{r}q}$, ${\mathit{sq}}^q \in {\lsemanticq\rsemantic_{\rho{^{\scriptscriptstyle\prime}}}}$, ${\mathit{sq}}^q \in {\mathcal{E}\!\mathit{nv}\:R{^{\scriptscriptstyle\prime}}}$, ${r \diamond R{^{\scriptscriptstyle\prime}}} \subseteq {R \diamond r}$. Moreover, let $\sigma$ be a state with $(\sigma, {{\mathsf{\mathcal{S}}}({\mathit{sq}}^q_0)}) \in r$. Then there exists a computation ${\mathit{sq}}^p \in {\lsemanticp\rsemantic_{\rho}}$ with the same length as ${\mathit{sq}}^q$ such that the following conditions hold:
1. ${{\mathsf{\mathcal{S}}}({\mathit{sq}}^p_0)} = \sigma$,
2. ${\mathit{sq}}^p \in {\mathcal{E}\!\mathit{nv}\:R}$,
3. $({{\mathsf{\mathcal{S}}}({\mathit{sq}}^p_i)}, {{\mathsf{\mathcal{S}}}({\mathit{sq}}^q_i)}) \in r$ for any $i < |{\mathit{sq}}^q|$,
4. ${\rho, \rho{^{\scriptscriptstyle\prime}}\models {{\mathsf{\mathcal{P}}}({\mathit{sq}}^p_i)}\sqsupseteq_{r}{{\mathsf{\mathcal{P}}}({\mathit{sq}}^q_i)}}$ for any $i < |{\mathit{sq}}^q|$,
5. ${\mathit{sq}}^p_i {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{E}}}\:}{\mathit{sq}}^p_{i+1}$ iff ${\mathit{sq}}^q_i {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{E}}}\:}{\mathit{sq}}^q_{i+1}$ for any $i < |{\mathit{sq}}^q| - 1$.
We proceed by induction on the length of ${\mathit{sq}}^q$. The case $|{\mathit{sq}}^q| = 0$ is trivial, since ${\mathit{sq}}^q \in {\lsemanticq\rsemantic_{\rho{^{\scriptscriptstyle\prime}}}}$ implies that ${\mathit{sq}}^q$ is not empty. Next, if $|{\mathit{sq}}^q| = 1$ then we can take the singleton sequence $(p, \sigma) \in {\lsemanticp\rsemantic_{\rho}}$ which satisfies the conditions (1)–(5).
Further, suppose ${\mathit{sq}}^q = {\mathit{sq}}^q_0, \ldots, {\mathit{sq}}^q_{n-1},{\mathit{sq}}^q_{n}$ with $n > 1$, such that ${_{\scriptscriptstyle n}|{\mathit{sq}}^q} = {\mathit{sq}}^q_0, \ldots, {\mathit{sq}}^q_{n-1}$ is a non-empty prefix of ${\mathit{sq}}^q$. Hence, ${_{\scriptscriptstyle n}|{\mathit{sq}}^q}\in {\lsemanticq\rsemantic_{\rho{^{\scriptscriptstyle\prime}}}} \cap{\mathcal{E}\!\mathit{nv}\:R{^{\scriptscriptstyle\prime}}}$ holds, such that by the induction hypothesis we obtain some computation ${\mathit{sq}}^p \in {\lsemanticp\rsemantic_{\rho}}$ with $|{\mathit{sq}}^p| = n$ satisfying the conditions (1)–(5). In particular, we have
1. $({{\mathsf{\mathcal{S}}}({\mathit{sq}}^p_{n-1})}, {{\mathsf{\mathcal{S}}}({\mathit{sq}}^q_{n-1})}) \in r$ and
2. ${\rho, \rho{^{\scriptscriptstyle\prime}}\models {{\mathsf{\mathcal{P}}}({\mathit{sq}}^p_{n-1})}\sqsupseteq_{r}{{\mathsf{\mathcal{P}}}({\mathit{sq}}^q_{n-1})}}$.
In case of an environment step ${\mathit{sq}}^q_{n-1} {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{E}}}\:}{\mathit{sq}}^q_{n}$, $({{\mathsf{\mathcal{S}}}({\mathit{sq}}^q_{n-1})}, {{\mathsf{\mathcal{S}}}({\mathit{sq}}^q_{n})}) \in R{^{\scriptscriptstyle\prime}}$ follows from the assumption ${\mathit{sq}}^q \in {\mathcal{E}\!\mathit{nv}\:R{^{\scriptscriptstyle\prime}}}$. Hence, with (a) we can infer $({{\mathsf{\mathcal{S}}}({\mathit{sq}}^p_{n-1})}, {{\mathsf{\mathcal{S}}}({\mathit{sq}}^q_{n})}) \in {r \diamond R{^{\scriptscriptstyle\prime}}}$, and from ${r \diamond R{^{\scriptscriptstyle\prime}}} \subseteq {R \diamond r}$ consequently obtain some $\sigma{^{\scriptscriptstyle\prime}}$ such that $({{\mathsf{\mathcal{S}}}({\mathit{sq}}^p_{n-1})}, \sigma{^{\scriptscriptstyle\prime}}) \in R$ and $(\sigma{^{\scriptscriptstyle\prime}}, {{\mathsf{\mathcal{S}}}({\mathit{sq}}^q_{n})}) \in r$ hold. Thus, we can extend ${\mathit{sq}}^p$ to $
{\mathit{sq}}^p_0, \ldots, {\mathit{sq}}^p_{n-1}, ({{\mathsf{\mathcal{P}}}({\mathit{sq}}^p_{n-1})}, \sigma{^{\scriptscriptstyle\prime}})
$, which is a computation in ${\lsemanticp\rsemantic_{\rho}}$ satisfying the conditions (1)–(5).
Otherwise, assume a program step ${\rho{^{\scriptscriptstyle\prime}}\vdash {\mathit{sq}}^q_{n-1} {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}{\mathit{sq}}^q_{n}}$. Then (a) and (b) entail a corresponding program step ${\rho \vdash {\mathit{sq}}^p_{n-1} {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(p{^{\scriptscriptstyle\prime}}, \sigma{^{\scriptscriptstyle\prime}})}$, such that ${\rho, \rho{^{\scriptscriptstyle\prime}}\models p{^{\scriptscriptstyle\prime}}\sqsupseteq_{r}{{\mathsf{\mathcal{P}}}({\mathit{sq}}^q_{n})}}$ and $(\sigma{^{\scriptscriptstyle\prime}}, {{\mathsf{\mathcal{S}}}({\mathit{sq}}^q_{n})}) \in r$ hold. Using this, ${\mathit{sq}}^p$ can be extended to $
{\mathit{sq}}^p_0, \ldots, {\mathit{sq}}^p_{n-1}, (p{^{\scriptscriptstyle\prime}}, \sigma{^{\scriptscriptstyle\prime}})
$ which is, in turn, a computation in ${\lsemanticp\rsemantic_{\rho}}$ satisfying the conditions (1)–(5).
A Hoare-style Rely/Guarantee Program Logic {#S:prog-log}
==========================================
Having the environment, program, input and output conditions on potential computations defined, this section is devoted to a Hoare-style ( [@Hoare]) logic for reasoning about parallel programs by means of *extended Hoare triples* ${\rho\models\{R,\:P\}\:p\:\{Q,\:G\}}$, where $R$ is called the rely, $P$ – the precondition, $Q$ – the postcondition, and $G$ – the guarantee. Following [@STIRLING1988347], the word ‘extended’ emphasises that such triples arise by generalisation of the Owicki-Gries method [@Owicki_Gries_76], moving annotated assertions to the level of state relations: the rely and the guarantee, namely.
\[def:rgval\] Let $(\rho, p)$ be a program, $R$ and $G$ – state relations, whereas $P$ and $Q$ – state predicates. Then ${\rho\models\{R,\:P\}\:p\:\{Q,\:G\}}$ holds iff ${\mathcal{E}\!\mathit{nv}\:R} \cap {\mathcal{I}\!\mathit{n}\:P} \cap {\lsemanticp\rsemantic_{\rho}} \subseteq {\mathcal{O}\!\mathit{ut}\:Q} \cap {\mathcal{P}\!\mathit{rog}\:G}$ does.
Next definition subsumes this, encompassing also infinite computations.
The extended Hoare triple ${\rho \models^\omega\{R,\:P\}\:p\:\{Q,\:G\}}$ holds iff
1. ${\mathcal{E}\!\mathit{nv}\:R} \cap {\mathcal{I}\!\mathit{n}\:P} \cap {\lsemanticp\rsemantic^\omega_{\rho}} \subseteq {\mathcal{O}\!\mathit{ut}\:Q} \cap {\mathcal{P}\!\mathit{rog}\:G} $ and
2. ${\rho\models\{R,\:P\}\:p\:\{Q,\:G\}}$
hold.
That the program logic rules can conveniently focus on finite computations without losing anything regarding ${\rho \models^\omega\{R,\:P\}\:p\:\{Q,\:G\}}$ is due to the following statement.
${\rho \models^\omega\{R,\:P\}\:p\:\{Q,\:G\}}$ iff ${\rho\models\{R,\:P\}\:p\:\{Q,\:G\}}$.
The if-direction amounts to showing that ${\mathcal{E}\!\mathit{nv}\:R} \cap {\mathcal{I}\!\mathit{n}\:P} \cap {\lsemanticp\rsemantic^\omega_{\rho}} \subseteq {\mathcal{O}\!\mathit{ut}\:Q} \cap {\mathcal{P}\!\mathit{rog}\:G} $ is implied by ${\rho\models\{R,\:P\}\:p\:\{Q,\:G\}}$. To this end, let ${\mathit{sq}}$ be an infinite computation in ${\lsemanticp\rsemantic^\omega_{\rho}}$ satisfying ${\mathcal{E}\!\mathit{nv}\:R}$ and ${\mathcal{I}\!\mathit{n}\:P}$.
We first show ${\mathit{sq}}\in {\mathcal{O}\!\mathit{ut}\:Q}$. Assume ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_i)} = {{\mathsf{skip}}}$ for some $i \in \mathbb{N}$ and consider the prefix ${_{\scriptscriptstyle i+1}|{\mathit{sq}}}$. Then ${_{\scriptscriptstyle i+1}|{\mathit{sq}}} \in {\mathcal{E}\!\mathit{nv}\:R} \cap {\mathcal{I}\!\mathit{n}\:P} \cap {\lsemanticp\rsemantic_{\rho}}$ holds, such that ${_{\scriptscriptstyle i+1}|{\mathit{sq}}} \in {\mathcal{O}\!\mathit{ut}\:Q}$ follows from the assumption ${\rho\models\{R,\:P\}\:p\:\{Q,\:G\}}$. Since the last element of ${_{\scriptscriptstyle i+1}|{\mathit{sq}}}$ is the ${{\mathsf{skip}}}$-configuration ${\mathit{sq}}_i$, we get some $j \le i$ with ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_j)} = {{\mathsf{skip}}}$ and ${{\mathsf{\mathcal{S}}}({\mathit{sq}}_j)} \in Q$.
In order to establish ${\mathit{sq}}\in {\mathcal{P}\!\mathit{rog}\:G}$ a similar argumentation can be applied. Suppose ${\rho \vdash {\mathit{sq}}_{i} {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}{\mathit{sq}}_{i+1}}$ holds for some $i \in \mathbb{N}$. Now we take the prefix ${_{\scriptscriptstyle i+2}|{\mathit{sq}}}$. From the assumption ${\rho\models\{R,\:P\}\:p\:\{Q,\:G\}}$ we can first conclude ${_{\scriptscriptstyle i+2}|{\mathit{sq}}} \in {\mathcal{P}\!\mathit{rog}\:G}$, and then $({{\mathsf{\mathcal{S}}}({\mathit{sq}}_{i})}, {{\mathsf{\mathcal{S}}}({\mathit{sq}}_{i+1})}) \in G$.
The opposite direction is immediate.
The program correspondence rule {#Sb:pcorr-rule}
-------------------------------
To start with the program logic, next proposition shows how extended Hoare triples can be moved along program correspondences. From now on, let ${R\cdot X}$ denote the image of a set $X$ under a relation $R$, $\{b \:|\:\exists a\in X. \:(a, b) \in R\}$.
\[thm:pcorr-rule\] Assume ${\rho, \rho{^{\scriptscriptstyle\prime}}\models p\sqsupseteq_{r}q}$, ${\rho\models\{R,\:P\}\:p\:\{Q,\:G\}}$ and
1. ${r \diamond R{^{\scriptscriptstyle\prime}}} \subseteq {R \diamond r}$,
2. $P{^{\scriptscriptstyle\prime}}\subseteq {r\cdot P}$,
3. ${r\cdot Q} \subseteq Q{^{\scriptscriptstyle\prime}}$,
4. ${{{r^{\circ}} \diamond G} \diamond r} \subseteq G{^{\scriptscriptstyle\prime}}$.
Then ${\rho{^{\scriptscriptstyle\prime}}\models\{R{^{\scriptscriptstyle\prime}},\:P{^{\scriptscriptstyle\prime}}\}\:q\:\{Q{^{\scriptscriptstyle\prime}},\:G{^{\scriptscriptstyle\prime}}\}}$.
Let ${\mathit{sq}}^q \in {\lsemanticq\rsemantic_{\rho{^{\scriptscriptstyle\prime}}}}$ be a finite potential computation such that ${\mathit{sq}}^q \in {\mathcal{E}\!\mathit{nv}\:R{^{\scriptscriptstyle\prime}}} \cap {\mathcal{I}\!\mathit{n}\:P{^{\scriptscriptstyle\prime}}}$ holds. Then from (2) we obtain some $\sigma \in P$ with $(\sigma, {{\mathsf{\mathcal{S}}}({\mathit{sq}}^q_0)}) \in r$. Thus, Proposition \[thm:corr-sim\] provides the existence of a potential computation ${\mathit{sq}}^p \in {\lsemanticp\rsemantic_{\rho}}$ such that $|{\mathit{sq}}^p| = |{\mathit{sq}}^q|$ and
1. ${\mathit{sq}}^p \in {\mathcal{E}\!\mathit{nv}\:R} \cap {\mathcal{I}\!\mathit{n}\:P}$,
2. $({{\mathsf{\mathcal{S}}}({\mathit{sq}}^p_i)}, {{\mathsf{\mathcal{S}}}({\mathit{sq}}^q_i)}) \in r$ for any $i < |{\mathit{sq}}^q|$,
3. ${\rho, \rho{^{\scriptscriptstyle\prime}}\models {{\mathsf{\mathcal{P}}}({\mathit{sq}}^p_i)}\sqsupseteq_{r}{{\mathsf{\mathcal{P}}}({\mathit{sq}}^q_i)}}$ for any $i < |{\mathit{sq}}^q|$,
4. ${\mathit{sq}}^p_i {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{E}}}\:}{\mathit{sq}}^p_{i+1}$ iff ${\mathit{sq}}^q_i {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{E}}}\:}{\mathit{sq}}^q_{i+1}$ for any $i < |{\mathit{sq}}^q| - 1$.
Further, the assumption ${\rho\models\{R,\:P\}\:p\:\{Q,\:G\}}$ and (a) entail ${\mathit{sq}}^p \in {\mathcal{O}\!\mathit{ut}\:Q} \cap {\mathcal{P}\!\mathit{rog}\:G}$, which we will use below to establish ${\mathit{sq}}^q \in {\mathcal{O}\!\mathit{ut}\:Q{^{\scriptscriptstyle\prime}}} \cap {\mathcal{P}\!\mathit{rog}\:G{^{\scriptscriptstyle\prime}}}$.
First, suppose ${{\mathsf{\mathcal{P}}}({\mathit{sq}}^q_i)} = {{\mathsf{skip}}}$ for some $i < |{\mathit{sq}}^q|$. Then ${{\mathsf{\mathcal{P}}}({\mathit{sq}}^p_i)} = {{\mathsf{skip}}}$ follows from (c). Since ${\mathit{sq}}^p \in {\mathcal{O}\!\mathit{ut}\:Q}$, there must be some $j \le i$ with ${\mathit{sq}}^p_j = ({{\mathsf{skip}}}, \sigma)$ and $\sigma \in Q$. Then from (b) and (c) we can infer ${\mathit{sq}}^q_j = ({{\mathsf{skip}}}, \sigma{^{\scriptscriptstyle\prime}})$ with $(\sigma, \sigma{^{\scriptscriptstyle\prime}}) \in r$. That is, $\sigma{^{\scriptscriptstyle\prime}}$ is in the image of $Q$ under $r$ such that $\sigma{^{\scriptscriptstyle\prime}}\in Q{^{\scriptscriptstyle\prime}}$ follows from (3).
Second, in order to show ${\mathit{sq}}^q \in {\mathcal{P}\!\mathit{rog}\:G{^{\scriptscriptstyle\prime}}}$ assume ${\rho{^{\scriptscriptstyle\prime}}\vdash {\mathit{sq}}^q_i {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}{\mathit{sq}}^q_{i+1}}$ with some $i < |{\mathit{sq}}^q| - 1$. Then (d) ensures a program step ${\rho \vdash {\mathit{sq}}^p_i {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}{\mathit{sq}}^p_{i+1}}$ at the same position, so that ${\mathit{sq}}^p \in {\mathcal{P}\!\mathit{rog}\:G}$ implies $({{\mathsf{\mathcal{S}}}({\mathit{sq}}^p_i)}, {{\mathsf{\mathcal{S}}}({\mathit{sq}}^p_{i+1})}) \in G$. Furthermore, from (b) we can conclude $({{\mathsf{\mathcal{S}}}({\mathit{sq}}^p_i)}, {{\mathsf{\mathcal{S}}}({\mathit{sq}}^q_{i})}) \in r$ and $({{\mathsf{\mathcal{S}}}({\mathit{sq}}^p_{i+1})}, {{\mathsf{\mathcal{S}}}({\mathit{sq}}^q_{i+1})}) \in r$. Thus, $({{\mathsf{\mathcal{S}}}({\mathit{sq}}^q_i)}, {{\mathsf{\mathcal{S}}}({\mathit{sq}}^q_{i+1})}) \in G{^{\scriptscriptstyle\prime}}$ follows from (4).
In particular, assuming $p = q$ and taking ${\mathsf{id}}$ for $r$, Proposition \[thm:pcorr-rule\] yields the canonical rule of consequence. Another result of Proposition \[thm:pcorr-rule\] is the following
\[thm:peqv-rg\] Assume ${\rho, \rho{^{\scriptscriptstyle\prime}}\models p\approx_{}q}$. Then ${\rho\models\{R,\:P\}\:p\:\{Q,\:G\}}$ iff ${\rho{^{\scriptscriptstyle\prime}}\models\{R,\:P\}\:q\:\{Q,\:G\}}$.
With ${\mathsf{id}}$ in place of $r$, the conditions (1)–(4) of Proposition \[thm:pcorr-rule\] hold due to the reflexivity of $\subseteq$.
Rules for ${\mathsf{basic}}$ and ${\mathsf{skip}}$
--------------------------------------------------
\[thm:skip-rule\] $
{\rho\models\{R,\:P\}\:{{\mathsf{skip}}}\:\{P,\:G\}}.
$
Any computation ${\mathit{sq}}\in {\lsemantic{{\mathsf{skip}}}\rsemantic_{\rho}}$ trivially satisfies ${\mathcal{P}\!\mathit{rog}\:G}$, whereas ${\mathit{sq}}\in {\mathcal{O}\!\mathit{ut}\:P}$ is provided by the initial configuration ${\mathit{sq}}_0$.
\[thm:basic-rule\] Assume
1. ${R\cdot P} \subseteq P$,
2. $P \subseteq \{\sigma \:|\: f\:\sigma \in Q \wedge (\sigma, f\:\sigma) \in G \}$.
Then ${\rho\models\{R,\:P\}\:{{\mathsf{basic}}}\:f\:\{Q,\:G\}}$.
Let ${\mathit{sq}}\in {\mathcal{E}\!\mathit{nv}\:R} \cap {\mathcal{I}\!\mathit{n}\:P} \cap {\lsemantic{{\mathsf{basic}}}\:f\rsemantic_{\rho}}$. If ${\mathit{sq}}$ does not contain a program step then both, ${\mathit{sq}}\in {\mathcal{O}\!\mathit{ut}\:Q}$ and ${\mathit{sq}}\in {\mathcal{P}\!\mathit{rog}\:G}$, hold trivially. Otherwise, let ${\rho \vdash {\mathit{sq}}_n {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}{\mathit{sq}}_{n+1}}$ be the first program step, such that ${\mathit{sq}}_n = ({{\mathsf{basic}}}\:f, \sigma_n)$ and ${\mathit{sq}}_{n+1} = ({{\mathsf{skip}}}, f\:\sigma_n)$ with some $\sigma_n$ hold. Furthermore, for any $m < n$ we have ${\mathit{sq}}_m {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{E}}}\:}{\mathit{sq}}_{m+1}$ with ${\mathit{sq}}_m = ({{\mathsf{basic}}}\:f, \sigma_m)$ and ${\mathit{sq}}_{m+1} = ({{\mathsf{basic}}}\:f, \sigma_{m+1})$. Using the stability assumption (1) and ${{\mathsf{\mathcal{S}}}({\mathit{sq}}_0)} \in P$ we can conclude that $\sigma_n \in P$ holds. Hence, $f\:\sigma_n \in Q$ and $(\sigma_n, f\:\sigma_n) \in G$ by (2). Since ${^{\scriptscriptstyle n+1}|{\mathit{sq}}} \in {\lsemantic{{\mathsf{skip}}}\rsemantic_{\rho}}$, there are no other program steps apart from ${\rho \vdash {\mathit{sq}}_n {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}{\mathit{sq}}_{n+1}}$, we have ${\mathit{sq}}\in {\mathcal{P}\!\mathit{rog}\:G}$. Finally, ${\mathit{sq}}\in {\mathcal{O}\!\mathit{ut}\:Q}$ holds, since ${\mathit{sq}}_{n+1}$ is the first ${{\mathsf{skip}}}$-configuration.
The stability assumption (1) in the above statement ensures that any environment step retains the initial condition $P$, and is a general way to systematic reasoning about interleaving. It is also worth noting that ${{\mathsf{skip}}}$ is the only language constructor that can get away without any stability assumptions.
A rule for the parallel composition {#Sb:parallel-rule}
-----------------------------------
The following rule and the proof are, in essence, extensions of the rule and the proof given in [@STIRLING1988347] for the binary operator $p \parallel q$.
\[thm:parallel-rule\] Let $I = \{1,\ldots,m\}$ with $m > 0$. Furthermore, assume $G$ is a reflexive state relation and
1. $P \subseteq \bigcap\nolimits_{k \in I} P_i $,
2. $R \subseteq \bigcap\nolimits_{k \in I} R_i$,
3. ${\rho\models\{R_k,\:P_k\}\:p_k\:\{Q_k,\:\bigcap\nolimits_{l \in I \setminus \{k\}} R_l \cap G\}}$ for any $k \in I$,
4. ${R_k\cdot Q_k} \subseteq Q_k $ for any $k \in I$,
5. $\bigcap\nolimits_{k \in I} Q_k \subseteq Q$.
Then ${\rho\models\{R,\:P\}\:\!{\parallel\!\!p_1,\ldots,p_m}\:\{Q,\:G\}}$.
Suppose ${\mathit{sq}}\in {\mathcal{E}\!\mathit{nv}\:R} \cap {\mathcal{I}\!\mathit{n}\:P} \cap {\lsemantic\:{\parallel\!\!p_1,\ldots,p_m}\:\rsemantic_{\rho}}$, and let ${\mathit{sq}}{^{\scriptscriptstyle\prime}}$ be the longest prefix of ${\mathit{sq}}$ that does not contain a ${{\mathsf{skip}}}$-configuration. First, notice that if ${\mathit{sq}}$ does not contain such a configuration at all then ${\mathit{sq}}{^{\scriptscriptstyle\prime}}= {\mathit{sq}}$. Second, since ${\mathit{sq}}{^{\scriptscriptstyle\prime}}$ is a non-empty prefix of ${\mathit{sq}}$, we have ${\mathit{sq}}{^{\scriptscriptstyle\prime}}\in {\mathcal{E}\!\mathit{nv}\:R} \cap {\mathcal{I}\!\mathit{n}\:P} \cap {\lsemantic\:{\parallel\!\!p_1,\ldots,p_m}\:\rsemantic_{\rho}}$ as well. Third, the program part of each configuration of ${\mathit{sq}}{^{\scriptscriptstyle\prime}}$ has the form ${\parallel\!\!x_1,\ldots,x_m}$.
Thus, ${\mathit{sq}}{^{\scriptscriptstyle\prime}}$ can be ‘split’ into $m$ computations ${\mathit{sq}}^k \in {\lsemanticp_k\rsemantic_{\rho}}$ for $k\in I$, all having the same length $|{\mathit{sq}}{^{\scriptscriptstyle\prime}}|$ such that
1. ${{\mathsf{\mathcal{P}}}({\mathit{sq}}{^{\scriptscriptstyle\prime}}_i)} =\; {\parallel\!{{\mathsf{\mathcal{P}}}({\mathit{sq}}^{\scriptscriptstyle 1}_i)}, \ldots, {{\mathsf{\mathcal{P}}}({\mathit{sq}}^m_i)}}$ for any $i < |{\mathit{sq}}{^{\scriptscriptstyle\prime}}|$,
2. ${{\mathsf{\mathcal{S}}}({\mathit{sq}}{^{\scriptscriptstyle\prime}}_i)} = {{\mathsf{\mathcal{S}}}({\mathit{sq}}^k_i)}$ for any $i < |{\mathit{sq}}{^{\scriptscriptstyle\prime}}|$ and $k \in I$,
3. ${\mathit{sq}}{^{\scriptscriptstyle\prime}}_i {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{E}}}\:}{\mathit{sq}}{^{\scriptscriptstyle\prime}}_{i+1}$ iff for all $k \in I$ we have ${\mathit{sq}}^k_i {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{E}}}\:}{\mathit{sq}}^k_{i+1}$, for any $i < |{\mathit{sq}}{^{\scriptscriptstyle\prime}}| - 1$.
By this construction, ${\mathit{sq}}^k \in {\mathcal{I}\!\mathit{n}\:P_k}$ follows immediately from (1) for any $k \in I$.
Next, we show by contradiction that ${\mathit{sq}}^k \in {\mathcal{E}\!\mathit{nv}\:R_k}$ holds for all $k \in I$ as well. Assuming the opposite, the set $\bigcup_{k\in I}M_k$ where $$M_k {\mathrel{\stackrel{\scriptscriptstyle{\mathit{def}}}{=}}}\{i \:|\: i < |{\mathit{sq}}{^{\scriptscriptstyle\prime}}| - 1 \wedge {\mathit{sq}}^k_i {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{E}}}\:}{\mathit{sq}}^k_{i+1} \wedge ({{\mathsf{\mathcal{S}}}(sq_i)}, {{\mathsf{\mathcal{S}}}(sq_{i+1})}) \notin R_k \}$$ is not empty. Then let $\mu$ be the least index in $\bigcup_{k\in I}M_k$ and let $k_\mu \in I$ be chosen such that $\mu \in M_{k_\mu}$. Note that there must be some $n \in I$ with a program step at $\mu$, ${\rho \vdash {\mathit{sq}}^n_\mu {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}{\mathit{sq}}^n_{\mu+1}}$, for otherwise we could first conclude ${\mathit{sq}}{^{\scriptscriptstyle\prime}}_\mu {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{E}}}\:}{\mathit{sq}}{^{\scriptscriptstyle\prime}}_{\mu+1}$ by (c), then $({{\mathsf{\mathcal{S}}}(sq_\mu)}, {{\mathsf{\mathcal{S}}}(sq_{\mu+1})}) \in R$ from ${\mathit{sq}}\in {\mathcal{E}\!\mathit{nv}\:R}$, and finally $({{\mathsf{\mathcal{S}}}(sq_\mu)}, {{\mathsf{\mathcal{S}}}(sq_{\mu+1})}) \in R_{k_\mu}$ by (2). Further, note that the prefix ${_{\scriptscriptstyle \mu+2}|{\mathit{sq}}^n}$ must satisfy ${\mathcal{E}\!\mathit{nv}\:R_n}$. Otherwise there would be some $j < \mu$ with ${\mathit{sq}}^n_j {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{E}}}\:}{\mathit{sq}}^n_{j+1}$ and $({{\mathsf{\mathcal{S}}}(sq_j)}, {{\mathsf{\mathcal{S}}}(sq_{j+1})}) \notin R_n$, $j \in M_n$ contradicting to $\mu \le j$. Thus, ${_{\scriptscriptstyle \mu+2}|{\mathit{sq}}^n} \in {\mathcal{I}\!\mathit{n}\:P_n} \cap {\mathcal{E}\!\mathit{nv}\:R_n}$ such that ${_{\scriptscriptstyle \mu+2}|{\mathit{sq}}^n} \in {\mathcal{P}\!\mathit{rog}\:(\bigcap\nolimits_{l \in I \setminus \{n\}} R_l \cap G)}$ follows from (3). Since $k_\mu \neq n$, for the program step ${\rho \vdash {\mathit{sq}}^n_\mu {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}{\mathit{sq}}^n_{\mu+1}}$ this yields $({{\mathsf{\mathcal{S}}}(sq_\mu)}, {{\mathsf{\mathcal{S}}}(sq_{\mu+1})}) \in R_{k_\mu}$, in contradiction to $\mu \in M_{k_\mu}$.
The intermediate result is that ${\mathit{sq}}^k \in {\lsemanticp_k\rsemantic_{\rho}} \cap {\mathcal{I}\!\mathit{n}\:P_k} \cap {\mathcal{E}\!\mathit{nv}\:R_k}$ holds for all $k\in I$, such that ${\mathit{sq}}^k \in {\mathcal{O}\!\mathit{ut}\:Q_k} \cap {\mathcal{P}\!\mathit{rog}\:(\bigcap\nolimits_{l \in I \setminus \{k\}} R_l \cap G)}$ follows from (3) for all $k \in I$. This will be utilised in the remainder of the proof showing ${\mathit{sq}}\in {\mathcal{O}\!\mathit{ut}\:Q} \cap {\mathcal{P}\!\mathit{rog}\:G}$.
First, in order to establish ${\mathit{sq}}\in {\mathcal{O}\!\mathit{ut}\:Q}$, let $n = |{\mathit{sq}}{^{\scriptscriptstyle\prime}}|$ and assume some $j < |{\mathit{sq}}|$ with ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_j)} = {{\mathsf{skip}}}$. Then $n \le j$ since ${\mathit{sq}}{^{\scriptscriptstyle\prime}}$ does not contain any ${{\mathsf{skip}}}$-configurations. Further, the step of ${\mathit{sq}}$ at the position $n-1$ must be a program step to the first ${{\mathsf{skip}}}$-configuration, for otherwise ${\mathit{sq}}{^{\scriptscriptstyle\prime}}$ would not be the longest prefix without reaching a ${{\mathsf{skip}}}$-configuration. That is, we have ${\rho \vdash {\mathit{sq}}_{n-1} {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}{\mathit{sq}}_n}$ with ${\mathit{sq}}_{n-1} = ({\parallel\!\!{{\mathsf{skip}}}, \ldots, {{\mathsf{skip}}}}, \sigma)$ and ${\mathit{sq}}_n = ({{\mathsf{skip}}}, \sigma)$ for some state $\sigma$ and need to show $\sigma \in Q$. Since ${\mathit{sq}}^k \in {\mathcal{O}\!\mathit{ut}\:Q_k}$ and ${\mathit{sq}}^k_{n-1} = ({{\mathsf{skip}}}, \sigma)$ hold for all $k \in I$, we obtain ${\mathit{sq}}^k_{i_k} = ({{\mathsf{skip}}}, \sigma_{i_k})$ with $\sigma_{i_k} \in Q_k$ and $i_k \le n - 1$ for all $k \in I$. The assumption (4) ensures that each $Q_k$ is retained by all steps of ${\mathit{sq}}^k$ from $i_k$ on, since these can comprise environment steps only. Therefore, we can first infer $\sigma \in \bigcap\nolimits_{k \in I} Q_k$, and then $\sigma \in Q$ by (5).
Regarding the program condition, assume ${\rho \vdash {\mathit{sq}}_{i} {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}{\mathit{sq}}_{i+1}}$ with $i < |{\mathit{sq}}| - 1$ and let $n = |{\mathit{sq}}{^{\scriptscriptstyle\prime}}|$. If $i < n - 1$ then from (c) follows that there exists some $k \in I$ with ${\rho \vdash {\mathit{sq}}^k_{i} {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}{\mathit{sq}}^k_{i+1}}$ and, since ${\mathit{sq}}^k \in {\mathcal{P}\!\mathit{rog}\:G}$ holds, we can conclude $({{\mathsf{\mathcal{S}}}({\mathit{sq}}_i)}, {{\mathsf{\mathcal{S}}}({\mathit{sq}}_{i+1})}) = ({{\mathsf{\mathcal{S}}}({\mathit{sq}}^k_i)}, {{\mathsf{\mathcal{S}}}({\mathit{sq}}^k_{i+1})}) \in G$. Further, if $i + 1 = n$ then ${\mathit{sq}}_{i} = ({\parallel\!\!{{\mathsf{skip}}}, \ldots, {{\mathsf{skip}}}}, \sigma)$ and ${\mathit{sq}}_{i+1} = ({{\mathsf{skip}}}, \sigma)$ with some state $\sigma$ must hold, since otherwise ${\mathit{sq}}{^{\scriptscriptstyle\prime}}$ would not be the longest prefix without a ${{\mathsf{skip}}}$-configuration. This step is covered by the reflexivity assumption on $G$. Finally, if $i \ge n$ then ${\mathit{sq}}_{i}$ must be a ${{\mathsf{skip}}}$-configuration, which contradicts the assumption ${\rho \vdash {\mathit{sq}}_{i} {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}{\mathit{sq}}_{i+1}}$.
By setting $m = 2$, the statement of Proposition \[thm:parallel-rule\] assumes the following form.
\[thm:parallel-rule2\] Assume $G$ is reflexive and
1. $P \subseteq P_1 \cap P_2 $,
2. $R \subseteq R_1 \cap R_2$,
3. ${\rho\models\{R_1,\:P_1\}\:p_1\:\{Q_1,\:R_2 \cap G\}}$,
4. ${\rho\models\{R_2,\:P_2\}\:p_2\:\{Q_2,\:R_1 \cap G\}}$,
5. ${R_1\cdot Q_1} \subseteq Q_1 $,
6. ${R_2\cdot Q_2} \subseteq Q_2 $,
7. $Q_1 \cap Q_2 \subseteq Q$.
Then ${\rho\models\{R,\:P\}\:p_1 \parallel p_2\:\{Q,\:G\}}$.
These rules enable structured reasoning about interleaved computations, but are far from being complete: the famous ‘parallel increment’ $${\mathsf{parallel\mbox{-}inc}} {\mathrel{\stackrel{\scriptscriptstyle{\mathit{def}}}{=}}}{\mathbf{x}} := {\mathbf{x}} + 1 \parallel {\mathbf{x}} := {\mathbf{x}} + 1$$ acts as a witness for that, which will be explicitly highlighted below. To this end, we first exactly describe how actual programs working on actual states can be represented.
In what follows, any state $\sigma$ on which a program operates comprises a mapping that sends each state variable that occurs in the program (such as ${\mathbf{x}}$ in ${\mathsf{parallel\mbox{-}inc}}$ above) to a value of a fixed type assigned to the variable, $\sigma{\mathbf{x}}$ is an integer in ${\mathsf{parallel\mbox{-}inc}}$. Then an assignment of an appropriately typed term $t$ to a state variable ${\mathbf{a}}$ amounts to the indivisible step ${{\mathsf{basic}}}(\lambda\sigma. \:\sigma_{[{\mathbf{a}} := t]})$, where $\sigma_{[{\mathbf{a}} := t]}$ denotes the updated state sending ${\mathbf{a}}$ to $t$ and any ${\mathbf{b}} \neq {\mathbf{a}}$ – to $\sigma{\mathbf{b}}$. Whenever a state variable occurs in $t$, such as ${\mathbf{x}}$ in ${\mathbf{x}} + 1$, it actually stands for $\sigma{\mathbf{x}}$. That is, the assignment ${\mathbf{x}} := {\mathbf{x}} + 1$ is a shorthand for ${{\mathsf{basic}}}(\lambda\sigma. \:\sigma_{[{\mathbf{x}} := \sigma{\mathbf{x}} + 1]})$. Similar conventions apply to state predicates: a condition like ${\mathbf{a}} = 0$ is a shorthand for $\{\sigma \:|\: \sigma{\mathbf{a}} = 0\}$. Moreover, with state relations we will resort to prime variables: for instance ${\mathbf{a}} = {\mathbf{a}}{^{\scriptscriptstyle\prime}}\wedge {\mathbf{b}} = {\mathbf{a}}{^{\scriptscriptstyle\prime}}$ stands for $\{(\sigma, \sigma{^{\scriptscriptstyle\prime}}) \:|\: \sigma{\mathbf{a}} = \sigma{^{\scriptscriptstyle\prime}}{\mathbf{a}} \wedge \sigma{\mathbf{b}} = \sigma{^{\scriptscriptstyle\prime}}{\mathbf{a}}\}$.
Now, consider the following extended Hoare triple: $$\label{eq:pinc}
{\models\{\bot,\:{\mathbf{x}} = 0\}\:{\mathsf{parallel\mbox{-}inc}}\:\{{\mathbf{x}} = 2,\:\top\}}$$ where the rely condition $\bot$ rules out any environment that could interleave with ${\mathsf{parallel\mbox{-}inc}}$. It is actually not too difficult to directly verify that (\[eq:pinc\]) is indeed a property of ${\mathsf{parallel\mbox{-}inc}}$: this amounts to checking the postcondition on seven computations starting from $\sigma_{[{\mathbf{x}} := 0]}$ and performing up to three program steps.
On the other hand, it is impossible to derive the same property immediately using the parallel rule in Corollary \[thm:parallel-rule2\] because, assuming the opposite, we would have found some $R_i, P_i, Q_i$ with $i \in \{1, 2\}$ such that
1. ${\models\{R_1,\:P_1\}\:{\mathbf{x}} := {\mathbf{x}} + 1\:\{Q_1,\:R_2\}}$,
2. ${\models\{R_2,\:P_2\}\:{\mathbf{x}} := {\mathbf{x}} + 1\:\{Q_2,\:R_1\}}$
and, moreover, $\sigma_{[{\mathbf{x}} := 0]} \in P_1 \cap P_2$ and $Q_1 \cap Q_2 \subseteq {\mathbf{x}} = 2$ hold. Then from (1) we could infer $\sigma_{[{\mathbf{x}} := 1]} \in Q_1$, provided by the computation $({\mathbf{x}} := {\mathbf{x}} + 1, \sigma_{[{\mathbf{x}} := 0]}) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}({{\mathsf{skip}}}, \sigma_{[{\mathbf{x}} := 1]})$ that trivially satisfies ${\mathcal{E}\!\mathit{nv}\:R_1}$. Since $\sigma_{[{\mathbf{x}} := 1]} \in Q_2$ similarly follows from (2), we have $\sigma_{[{\mathbf{x}} := 1]} \in Q_1 \cap Q_2$ and therefore a contradiction in form of $\sigma_{[{\mathbf{x}} := 1]} \in ({\mathbf{x}} = 2)$.
However, this shall not lead to the hasty conclusion that (\[eq:pinc\]) is not derivable by the rules at all, for it is: in order to apply the parallel rule we first have to introduce two auxiliary variables, each local to one thread, and then show a generalised property. Subsequently, showing (\[eq:pinc\]) in principle amounts to discarding the auxiliaries by means of the program correspondence rule.
To sum up, placing auxiliary variables is often required for a successful application of the parallel rule, and the case study, starting in Section \[S:PM1\], will appeal to this technique too.
A rule for the sequential composition
-------------------------------------
\[thm:seq-rule\] Assume $G$ is a reflexive state relation and
1. ${\rho\models\{R,\:P\}\:p\:\{S,\:G\}}$,
2. ${\rho\models\{R,\:S\}\:q\:\{Q,\:G\}}$,
3. $q = {{\mathsf{skip}}}$ implies ${R\cdot Q} \subseteq Q$.
Then ${\rho\models\{R,\:P\}\:p;q\:\{Q,\:G\}}$.
Suppose ${\mathit{sq}}\in {\mathcal{E}\!\mathit{nv}\:R} \cap {\mathcal{I}\!\mathit{n}\:P} \cap {\lsemanticp;q\rsemantic_{\rho}}$ and let ${\mathit{sq}}{^{\scriptscriptstyle\prime}}$ be the longest prefix of ${\mathit{sq}}$ that does not contain a configuration of the form $({{\mathsf{skip}}}; q, \sigma)$. Then we can construct a computation ${\mathit{sq}}^p \in {\lsemanticp\rsemantic_{\rho}}$ with the length $|{\mathit{sq}}{^{\scriptscriptstyle\prime}}|$ such that
1. ${{\mathsf{\mathcal{P}}}({\mathit{sq}}{^{\scriptscriptstyle\prime}}_i)} = {{\mathsf{\mathcal{P}}}({\mathit{sq}}^p_i)} ; q$ for any $i < |{\mathit{sq}}{^{\scriptscriptstyle\prime}}|$,
2. ${{\mathsf{\mathcal{S}}}({\mathit{sq}}{^{\scriptscriptstyle\prime}}_i)} = {{\mathsf{\mathcal{S}}}({\mathit{sq}}^p_i)}$ for any $i < |{\mathit{sq}}{^{\scriptscriptstyle\prime}}|$,
3. ${\mathit{sq}}{^{\scriptscriptstyle\prime}}_i {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{E}}}\:}{\mathit{sq}}{^{\scriptscriptstyle\prime}}_{i+1}$ iff ${\mathit{sq}}^p_i {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{E}}}\:}{\mathit{sq}}^p_{i+1}$ for any $i < |{\mathit{sq}}{^{\scriptscriptstyle\prime}}| - 1$.
Thus, ${\mathit{sq}}^p \in {\mathcal{E}\!\mathit{nv}\:R} \cap {\mathcal{I}\!\mathit{n}\:P}$ holds, so that ${\mathit{sq}}^p \in {\mathcal{O}\!\mathit{ut}\:S} \cap {\mathcal{P}\!\mathit{rog}\:G}$ follows from (1).
If ${\mathit{sq}}$ does not contain a configuration of the form $({{\mathsf{skip}}}; q, \sigma)$ at all then ${\mathit{sq}}\in {\mathcal{O}\!\mathit{ut}\:Q}$ holds trivially, whereas ${\mathit{sq}}\in {\mathcal{P}\!\mathit{rog}\:G}$ holds due to ${\mathit{sq}}^p \in {\mathcal{P}\!\mathit{rog}\:G}$.
Otherwise, let $n = |{\mathit{sq}}{^{\scriptscriptstyle\prime}}|$ and assume $n < |{\mathit{sq}}|$. Then we have $${\rho \vdash ({{\mathsf{\mathcal{P}}}({\mathit{sq}}_{n-1})}, \sigma_{n-1}) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}({{\mathsf{\mathcal{P}}}({\mathit{sq}}_n)}, \sigma_n)}$$ with ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_{n-1})} = {{\mathsf{\mathcal{P}}}({\mathit{sq}}^p_{n-1})};q$ and ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_n)} = {{\mathsf{skip}}};q$. Thus, we can extend ${\mathit{sq}}^p$ to ${\mathit{sq}}^{p_1}$ by appending the configuration $({{\mathsf{skip}}}, \sigma_n)$ such that ${\mathit{sq}}^{p_1} \in {\mathcal{E}\!\mathit{nv}\:R} \cap {\mathcal{I}\!\mathit{n}\:P} \cap {\lsemanticp\rsemantic_{\rho}}$, and hence ${\mathit{sq}}^{p_1} \in {\mathcal{O}\!\mathit{ut}\:S} \cap {\mathcal{P}\!\mathit{rog}\:G}$ hold. In particular, note that ${\mathit{sq}}^{p_1} \in {\mathcal{O}\!\mathit{ut}\:S}$ implies $\sigma_n \in S$.
Further, in case there is no $m > n$ with ${\rho \vdash {\mathit{sq}}_{m-1} {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}{\mathit{sq}}_{m}}$ we once more have that ${\mathit{sq}}\in {\mathcal{O}\!\mathit{ut}\:Q}$ holds trivially, whereas ${\mathit{sq}}\in {\mathcal{P}\!\mathit{rog}\:G}$ holds because ${\mathit{sq}}^p \in {\mathcal{P}\!\mathit{rog}\:G}$ does and $G$ is assumed to be reflexive.
Otherwise, assume ${\mathit{sq}}$ makes a program step at $m$ with $n \le m < |{\mathit{sq}}| -1$ and let $m_0$ be the least such index. That is, we have ${\rho \vdash {\mathit{sq}}_{m_0} {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}{\mathit{sq}}_{m_0 + 1}}$ with ${\mathit{sq}}_{m_0} = ({{\mathsf{skip}}};q, \sigma_{m_0})$ and ${\mathit{sq}}_{m_0 + 1} = (q, \sigma_{m_0})$. Now consider the subsequence ${\mathit{sq}}_n {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{E}}}\:}\ldots {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{E}}}\:}{\mathit{sq}}_{m_0}$ where for each configuration ${\mathit{sq}}_k$ with $n \le k \le m_0$ we have ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_k)} = {{\mathsf{skip}}};q$. From this we obtain the computation ${\mathit{sq}}^{q_0} = (q, \sigma_n) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{E}}}\:}\ldots {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{E}}}\:}(q, \sigma_{m_0})$. Then composing ${\mathit{sq}}^{q_0}$ with the suffix ${^{\scriptscriptstyle m_0+1}|{\mathit{sq}}}$ yields the computation ${\mathit{sq}}^q \in {\mathcal{E}\!\mathit{nv}\:R} \cap {\mathcal{I}\!\mathit{n}\:S} \cap {\lsemanticq\rsemantic_{\rho}}$ such that ${\mathit{sq}}^q \in {\mathcal{O}\!\mathit{ut}\:Q} \cap {\mathcal{P}\!\mathit{rog}\:G}$ follows from (2).
Next, to show ${\mathit{sq}}\in {\mathcal{O}\!\mathit{ut}\:Q}$, let $j < |{\mathit{sq}}|$ be an index with ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_j)} = {{\mathsf{skip}}}$. Then $j > m_0$ must hold, whereas $j = m_0 + 1$ would imply $q = {{\mathsf{skip}}}$. If $q = {{\mathsf{skip}}}$ then ${{\mathsf{\mathcal{P}}}({\mathit{sq}}^q_{m_0 - n})} = {{\mathsf{skip}}}$, and from ${\mathit{sq}}^q \in {\mathcal{O}\!\mathit{ut}\:Q}$ we obtain some $i \le m_0 - n$ with ${{\mathsf{\mathcal{P}}}({\mathit{sq}}^q_i)} = {{\mathsf{skip}}}$ and ${{\mathsf{\mathcal{S}}}({\mathit{sq}}^q_i)} \in Q$. Using (3) we can extend this to ${{\mathsf{\mathcal{S}}}({\mathit{sq}}^q_{m_0 - n})} = {{\mathsf{\mathcal{S}}}({\mathit{sq}}_{m_0})} = \sigma_{m_0} \in Q$. If $q \neq {{\mathsf{skip}}}$ then by the remark above we have $m_0 + 1 < j$. Therefore, ${{\mathsf{\mathcal{P}}}({\mathit{sq}}^q_{j - n - 1})} = {{\mathsf{\mathcal{P}}}({\mathit{sq}}_j)} = {{\mathsf{skip}}}$ such that from ${\mathit{sq}}^q \in {\mathcal{O}\!\mathit{ut}\:Q}$ we obtain some $i \le j - n - 1$ with ${{\mathsf{\mathcal{P}}}({\mathit{sq}}^q_i)} = {{\mathsf{skip}}}$ and ${{\mathsf{\mathcal{S}}}({\mathit{sq}}^q_i)} \in Q$. Notice that $i \le m_0 - n$ would imply $q = {{\mathsf{skip}}}$, $i > m_0 - n$ must hold and hence $j \ge i + n + 1 > m_0 + 1$. Thus, $ {{\mathsf{\mathcal{P}}}({\mathit{sq}}_{i+n+1})} = {{\mathsf{\mathcal{P}}}({\mathit{sq}}^q_i)} = {{\mathsf{skip}}}$ and $ {{\mathsf{\mathcal{S}}}({\mathit{sq}}_{i+n+1})} = {{\mathsf{\mathcal{S}}}({\mathit{sq}}^q_i)} \in Q$.
Lastly, ${\mathit{sq}}\in {\mathcal{P}\!\mathit{rog}\:G}$ follows from ${\mathit{sq}}^{p_1} \in {\mathcal{P}\!\mathit{rog}\:G}$ and ${^{\scriptscriptstyle m_0+1}|{\mathit{sq}}} \in {\mathcal{P}\!\mathit{rog}\:G}$, if taking also into account that the ‘omitted’ transition ${\rho \vdash {\mathit{sq}}_{m_0} {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}{\mathit{sq}}_{m_0 + 1}}$ is covered by $G$ as we assumed it to be reflexive.
It must be stressed that the assumptions (1) and (2) in the above proposition are not sufficient to remain sound. To substantiate that, suppose Proposition \[thm:seq-rule\] would hold without the assumption (3). Then the statement ${\models\{{\mathbf{a}} = {\mathbf{a}}{^{\scriptscriptstyle\prime}},\:{\mathbf{a}}=0\}\:{\mathbf{b}} := {\mathbf{a}} + 1;{{\mathsf{skip}}}\:\{{\mathbf{b}}=1,\:\top\}}$ would follow from ${\models\{{\mathbf{a}} = {\mathbf{a}}{^{\scriptscriptstyle\prime}},\:{\mathbf{a}}=0\}\:{\mathbf{b}} := {\mathbf{a}} + 1\:\{{\mathbf{b}}=1,\:\top\}}$ and ${\models\{{\mathbf{a}} = {\mathbf{a}}{^{\scriptscriptstyle\prime}},\:{\mathbf{b}}=1\}\:{{\mathsf{skip}}}\:\{{\mathbf{b}}=1,\:\top\}}$, where the former is an instance of Proposition \[thm:basic-rule\], and the latter – of Proposition \[thm:skip-rule\] since this does not have any stability assumptions. On the other hand, starting in some state $\sigma$ with $\sigma{\mathbf{a}} = 0$, the computation $$\!\!\!\!\!\!\!\!\!\!\!\!({\mathbf{b}} := {\mathbf{a}} + 1;{{\mathsf{skip}}}, \sigma) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}({{\mathsf{skip}}};{{\mathsf{skip}}}, \sigma_{[{\mathbf{b}} := 1]}) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{E}}}\:}({{\mathsf{skip}}};{{\mathsf{skip}}}, \sigma_{[{\mathbf{b}} := 0]}) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}({{\mathsf{skip}}},\sigma_{[{\mathbf{b}} := 0]})$$ clearly satisfies ${\mathcal{E}\!\mathit{nv}\:\!({\mathbf{a}} = {\mathbf{a}}{^{\scriptscriptstyle\prime}})} \cap {\mathcal{I}\!\mathit{n}\:\!({\mathbf{a}} = 0)} \cap {\lsemantic{\mathbf{b}} := {\mathbf{a}} + 1;{{\mathsf{skip}}}\rsemantic}$ but not ${\mathcal{O}\!\mathit{ut}\:\!({\mathbf{b}}=1)}$.
This and similar argumentations fail with Proposition \[thm:seq-rule\]: the assumption (3) forces us to provide stability of the postcondition ${\mathbf{b}}=1$ under the rely ${\mathbf{a}} = {\mathbf{a}}{^{\scriptscriptstyle\prime}}$, which is impossible.
A rule for ${\mathsf{while}}$-statements
----------------------------------------
\[thm:while-rule\] Assume $G$ is a reflexive state relation and
1. ${R\cdot P} \subseteq P$,
2. ${\rho\models\{R,\:P \cap C\}\:p\:\{P,\:G\}}$,
3. ${\rho\models\{R,\:P \cap {\neg C}\}\:q\:\{Q,\:G\}}$.
Then ${\rho\models\{R,\:P\}\:{{\mathsf{while}}\:C\;{\mathsf{do}}\;p\;{\mathsf{subsequently}}\;q\;{\mathsf{od}}}\:\{Q,\:G\}}$.
Let $p_{{\mathit{while}}}$ abbreviate ${{\mathsf{while}}\:C\;{\mathsf{do}}\;p\;{\mathsf{subsequently}}\;q\;{\mathsf{od}}}$. Then we have to show that ${\mathit{sq}}\in {\mathcal{E}\!\mathit{nv}\:R} \cap {\mathcal{I}\!\mathit{n}\:P} \cap {\lsemanticp_{{\mathit{while}}}\rsemantic_{\rho}}$ implies ${\mathit{sq}}\in {\mathcal{O}\!\mathit{ut}\:Q} \cap {\mathcal{P}\!\mathit{rog}\:G}$, and proceed by induction on the length of ${\mathit{sq}}$. Thus, the induction hypothesis is
1. *[if ${\mathit{sq}}{^{\scriptscriptstyle\prime}}\in {\mathcal{E}\!\mathit{nv}\:R} \cap {\mathcal{I}\!\mathit{n}\:P} \cap {\lsemanticp_{{\mathit{while}}}\rsemantic_{\rho}}$ then ${\mathit{sq}}{^{\scriptscriptstyle\prime}}\in {\mathcal{O}\!\mathit{ut}\:Q} \cap {\mathcal{P}\!\mathit{rog}\:G}$ for any ${\mathit{sq}}{^{\scriptscriptstyle\prime}}$ with $|{\mathit{sq}}{^{\scriptscriptstyle\prime}}| < |{\mathit{sq}}|$]{}*
which, in turn, implies the statement
1. *[if ${\mathit{sq}}{^{\scriptscriptstyle\prime}}_0 = (p_{{\mathit{while}}}, \sigma)$ and $\sigma \in P$ then ${\mathit{sq}}{^{\scriptscriptstyle\prime}}\in {\mathcal{O}\!\mathit{ut}\:Q} \cap {\mathcal{P}\!\mathit{rog}\:G}$ for any proper and non-empty suffix ${\mathit{sq}}{^{\scriptscriptstyle\prime}}$ of ${\mathit{sq}}$]{}*
that we will additionally assume.
In case there is no $n < |{\mathit{sq}}| - 1$ with ${\rho \vdash {\mathit{sq}}_n {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}{\mathit{sq}}_{n+1}}$, we are done. Otherwise, let $n_0$ be the first such index. That is, ${\rho \vdash {\mathit{sq}}_{n_0} {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}{\mathit{sq}}_{n_0+1}}$ holds with ${\mathit{sq}}_{n_0} = (p_{{\mathit{while}}}, \sigma_{n_0})$ and ${\mathit{sq}}_{n_0 + 1} = (u, \sigma_{n_0})$ for some $u$. Using the stability assumption (1) we can further infer $\sigma_{n_0} \in P$.
First, if $\sigma_{n_0} \notin C$ then $u = q$, and hence ${^{\scriptscriptstyle n_0 + 1}|{\mathit{sq}}} \in {\mathcal{E}\!\mathit{nv}\:R} \cap {\mathcal{I}\!\mathit{n}\:(P \cap {\neg C})} \cap {\lsemanticq\rsemantic_{\rho}}$. Then ${^{\scriptscriptstyle n_0 + 1}|{\mathit{sq}}} \in {\mathcal{O}\!\mathit{ut}\:Q} \cap {\mathcal{P}\!\mathit{rog}\:G}$ follows from (3), and therefore ${\mathit{sq}}\in {\mathcal{O}\!\mathit{ut}\:Q} \cap {\mathcal{P}\!\mathit{rog}\:G}$ holds, since $n_0$ is the first index with a program step and $G$ is reflexive.
If $\sigma_{n_0} \in C$ then $u = p;{{\mathsf{skip}}};p_{{\mathit{while}}}$, and hence ${^{\scriptscriptstyle n_0 + 1}|{\mathit{sq}}} \in {\mathcal{E}\!\mathit{nv}\:R} \cap {\mathcal{I}\!\mathit{n}\:(P \cap C)} \cap {\lsemanticu\rsemantic_{\rho}}$. Let ${\mathit{sq}}{^{\scriptscriptstyle\prime}}$ be the longest prefix of the suffix ${^{\scriptscriptstyle n_0 + 1}|{\mathit{sq}}}$ that does not contain a configuration of the form $({{\mathsf{skip}}}; {{\mathsf{skip}}}; p_{{\mathit{while}}}, \sigma)$ with some state $\sigma$. Then we can construct a computation ${\mathit{sq}}^p \in {\lsemanticp\rsemantic_{\rho}}$ with the length $|{\mathit{sq}}{^{\scriptscriptstyle\prime}}|$ such that
1. ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_{n_0 + i + 1})} = {{\mathsf{\mathcal{P}}}({\mathit{sq}}^p_i)} ; {{\mathsf{skip}}}; p_{{\mathit{while}}}$ for all $i < |{\mathit{sq}}{^{\scriptscriptstyle\prime}}|$,
2. ${{\mathsf{\mathcal{S}}}({\mathit{sq}}_{n_0 + i + 1})} = {{\mathsf{\mathcal{S}}}({\mathit{sq}}^p_i)}$ for all $i < |{\mathit{sq}}{^{\scriptscriptstyle\prime}}|$,
3. ${\mathit{sq}}_{n_0 + i + 1} {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{E}}}\:}{\mathit{sq}}_{n_0 + i + 2}$ iff ${\mathit{sq}}^p_i {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{E}}}\:}{\mathit{sq}}^p_{i+1}$ for any $i < |{\mathit{sq}}{^{\scriptscriptstyle\prime}}| - 1$.
Thus, ${\mathit{sq}}^p \in {\mathcal{E}\!\mathit{nv}\:R} \cap {\mathcal{I}\!\mathit{n}\:\!(P \cap C)}$ holds, and ${\mathit{sq}}^p \in {\mathcal{O}\!\mathit{ut}\:P} \cap {\mathcal{P}\!\mathit{rog}\:G}$ follows from (2).
If ${^{\scriptscriptstyle n_0 + 1}|{\mathit{sq}}}$ does not reach a configuration of the form $({{\mathsf{skip}}}; {{\mathsf{skip}}}; p_{{\mathit{while}}}, \sigma)$ then ${\mathit{sq}}\in {\mathcal{O}\!\mathit{ut}\:Q}$ holds trivially, whereas ${\mathit{sq}}\in {\mathcal{P}\!\mathit{rog}\:G}$ holds because ${\mathit{sq}}^p \in {\mathcal{P}\!\mathit{rog}\:G}$ does and ${\rho \vdash {\mathit{sq}}_{n_0} {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}{\mathit{sq}}_{n_0+1}}$ is covered by $G$ due to the reflexivity assumption.
Otherwise, let $n_1 = |{\mathit{sq}}{^{\scriptscriptstyle\prime}}|$. Then $n_0 + n_1 < |{\mathit{sq}}| - 1$ and $${\rho \vdash ({{\mathsf{\mathcal{P}}}({\mathit{sq}}_{n_0 + n_1})}, \sigma_{n_0 + n_1}) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}({{\mathsf{\mathcal{P}}}({\mathit{sq}}_{n_0 + n_1 + 1})}, \sigma_{n_0 + n_1+ 1})}$$ with ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_{n_0 + n_1})} = {\mathit{sq}}^p_{n_1-1};{{\mathsf{skip}}}; p_{{\mathit{while}}}$ and ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_{n_0 + n_1 + 1})} = {{\mathsf{skip}}};{{\mathsf{skip}}}; p_{{\mathit{while}}}$. Using this, ${\mathit{sq}}^p$ can be extended to ${\mathit{sq}}^{p_1}$ by appending the configuration $({{\mathsf{skip}}}, \sigma_{n_0 + n_1+ 1})$. Since ${\mathit{sq}}^{p_1}$ differs from ${\mathit{sq}}^{p}$ only in one program step, ${\mathit{sq}}^{p_1} \in {\mathcal{E}\!\mathit{nv}\:R} \cap {\mathcal{I}\!\mathit{n}\:\!\!(P \cap C)} \cap {\lsemanticp\rsemantic_{\rho}}$ holds and hence ${\mathit{sq}}^{p_1} \in {\mathcal{O}\!\mathit{ut}\:P} \cap {\mathcal{P}\!\mathit{rog}\:G}$ follows from (2). In particular, ${\mathit{sq}}^{p_1} \in {\mathcal{O}\!\mathit{ut}\:P}$ implies $\sigma_{n_0 + n_1+ 1} \in P$. For the suffix ${^{\scriptscriptstyle n_0 + n_1+ 1}|{\mathit{sq}}}$ this yields ${^{\scriptscriptstyle n_0 + n_1+ 1}|{\mathit{sq}}}\in {\mathcal{E}\!\mathit{nv}\:R} \cap {\mathcal{I}\!\mathit{n}\:P} \cap {\lsemantic{{\mathsf{skip}}};{{\mathsf{skip}}}; p_{{\mathit{while}}}\rsemantic_{\rho}}$.
If ${^{\scriptscriptstyle n_0 + n_1+ 1}|{\mathit{sq}}}$ does not contain two program steps (from ${{\mathsf{skip}}};{{\mathsf{skip}}}; p_{{\mathit{while}}}$ to ${{\mathsf{skip}}}; p_{{\mathit{while}}}$ and from ${{\mathsf{skip}}}; p_{{\mathit{while}}}$ to $p_{{\mathit{while}}}$) then ${\mathit{sq}}\in {\mathcal{O}\!\mathit{ut}\:Q}$ holds once more trivially, whereas ${\mathit{sq}}\in {\mathcal{P}\!\mathit{rog}\:G}$ holds since ${\mathit{sq}}^{p_1} \in {\mathcal{P}\!\mathit{rog}\:G}$ does and $G$ is reflexive.
Otherwise, we have some $n_2$ with $n_0 + n_1+ 2 < n_2 < |{\mathit{sq}}|$ and ${\mathit{sq}}_{n_2} = (p_{{\mathit{while}}}, \sigma_{n_2})$. Using the stability assumption (1), we can also infer $\sigma_{n_2} \in P$ from $\sigma_{n_0 + n_1+ 1} \in P$. Thus, the induction hypothesis (4) applies to ${^{\scriptscriptstyle n_2}|{\mathit{sq}}}$, yielding ${^{\scriptscriptstyle n_2}|{\mathit{sq}}} \in {\mathcal{O}\!\mathit{ut}\:Q} \cap {\mathcal{P}\!\mathit{rog}\:G}$.
To show the output condition for ${\mathit{sq}}$, let $j < |{\mathit{sq}}|$ be an index with ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_j)} = {{\mathsf{skip}}}$. From the above considerations follows that $n_2 < j$ holds. Since this implies ${{\mathsf{\mathcal{P}}}({^{\scriptscriptstyle n_2}|{\mathit{sq}}}_{j - n_2})} = {{\mathsf{skip}}}$, from ${^{\scriptscriptstyle n_2}|{\mathit{sq}}} \in {\mathcal{O}\!\mathit{ut}\:Q}$ we obtain some $i \le j - n_2$ such that ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_{i + n_2})} = {{\mathsf{\mathcal{P}}}({^{\scriptscriptstyle n_2}|{\mathit{sq}}}_{i})} = {{\mathsf{skip}}}$ and ${{\mathsf{\mathcal{S}}}({\mathit{sq}}_{i + n_2})} = {{\mathsf{\mathcal{S}}}({^{\scriptscriptstyle n_2}|{\mathit{sq}}}_{i})} \in Q$, ${\mathit{sq}}\in {\mathcal{O}\!\mathit{ut}\:Q}$.
Finally, ${\mathit{sq}}\in {\mathcal{P}\!\mathit{rog}\:G}$ follows from ${\mathit{sq}}^{p_1} \in {\mathcal{P}\!\mathit{rog}\:G}$ and ${^{\scriptscriptstyle n_2}|{\mathit{sq}}} \in {\mathcal{P}\!\mathit{rog}\:G}$.
A rule for ${\mathsf{await}}$-statements
----------------------------------------
The particular difficulty with ${{\mathsf{await}}\:C\;{\mathsf{do}}\;p\;{\mathsf{od}}}$ is that one has to express that a state before entering $p$ is related by means of the given guarantee to a state where $p$ terminates, which, by contrast to Proposition \[thm:basic-rule\], can be connected by a computation involving several program steps. Therefore, in the assumption (2) below, $\sigma$ is used as a reference to states before entering $p$, such that in the postcondition we can specify that inputs and outputs must be related by the guarantee $G$. The same approach can be encountered in [@10.1007/3-540-36575-3_24]. It is also worth noting that appealing to the assumption $(P \cap C) \times Q \subseteq G$ instead, is in principle possible but would yield a significantly weaker rule, since $(P \cap C) \times Q$ in general gives only a rough overapproximation of the actual input/output behaviour of $p$.
\[thm:await-rule\] Assume
1. ${R\cdot P} \subseteq P$,
2. ${\rho\models\{\bot,\:P \cap C \cap \{\sigma\}\}\:p\:\{\{\sigma{^{\scriptscriptstyle\prime}}\:|\: \sigma{^{\scriptscriptstyle\prime}}\in Q \wedge (\sigma, \sigma{^{\scriptscriptstyle\prime}}) \in G\},\:\top\}}$ for any state $\sigma$.
Then ${\rho\models\{R,\:P\}\:{{\mathsf{await}}\:C\;{\mathsf{do}}\;p\;{\mathsf{od}}}\:\{Q,\:G\}}$.
Let ${\mathit{sq}}\in {\mathcal{E}\!\mathit{nv}\:R} \cap {\mathcal{I}\!\mathit{n}\:P} \cap {\lsemantic{{\mathsf{await}}\:C\;{\mathsf{do}}\;p\;{\mathsf{od}}}\rsemantic_{\rho}}$. If ${\mathit{sq}}$ does not make a program step then we are done. Otherwise, assume ${\rho \vdash {\mathit{sq}}_i {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}{\mathit{sq}}_{i+1}}$ with $i < |{\mathit{sq}}| - 1$ being the first such index. Then ${\mathit{sq}}_i = ({{\mathsf{await}}\:C\;{\mathsf{do}}\;p\;{\mathsf{od}}}, \sigma_i)$ and ${\mathit{sq}}_{i+1} = ({{\mathsf{skip}}}, \sigma_{i+1})$ with $\sigma_i \in C$ and, moreover, $\sigma_i \in P$ due to the stability assumption (1). Furthermore, ${\rho \vdash (p, \sigma_{i}) {\stackrel{\scriptscriptstyle n}{\mathop{\rightarrow}}_{\scriptscriptstyle{\mathcal{P}}}} ({{\mathsf{skip}}}, \sigma_{i + 1})}$ holds for some $n \in \mathbb{N}$. That is, we have a computation ${\mathit{sq}}^p \in {\mathcal{E}\!\mathit{nv}\:\bot} \cap {\mathcal{I}\!\mathit{n}\:\{\sigma_i\}} \cap{\lsemanticp\rsemantic_{\rho}}$ with the length $n+1$ such that ${\mathit{sq}}^p_0 = (p, \sigma_{i})$ and ${\mathit{sq}}^p_n = ({{\mathsf{skip}}}, \sigma_{i+1})$ hold.
Hence, by instantiation of $\sigma$ by $\sigma_i$ in (2) we can first infer ${\mathit{sq}}^p \in {\mathcal{O}\!\mathit{ut}\:A}$, where $A$ is $\{\sigma{^{\scriptscriptstyle\prime}}\:|\: \sigma{^{\scriptscriptstyle\prime}}\in Q \wedge (\sigma_{i}, \sigma{^{\scriptscriptstyle\prime}}) \in G\}$, and then $\sigma_{i+1} \in Q$ and $(\sigma_{i}, \sigma_{i+1}) \in G$. The former provides ${\mathit{sq}}\in {\mathcal{O}\!\mathit{ut}\:Q}$, since ${\mathit{sq}}_{i+1}$ is the first ${{\mathsf{skip}}}$-configuration on ${\mathit{sq}}$, whereas the latter – ${\mathit{sq}}\in {\mathcal{P}\!\mathit{rog}\:G}$, since ${\rho \vdash {\mathit{sq}}_i {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}{\mathit{sq}}_{i+1}}$ is also the last program step on ${\mathit{sq}}$.
A rule for conditional statements
---------------------------------
\[thm:cond-rule\] Assume $G$ is a reflexive state relation and
1. ${R\cdot P} \subseteq P$,
2. ${\rho\models\{R,\:P \cap C\}\:p\:\{Q,\:G\}}$,
3. ${\rho\models\{R,\:P \cap {\neg C}\}\:q\:\{Q,\:G\}}$.
Then ${\rho\models\{R,\:P\}\:{{\mathsf{if}}\:C\;{\mathsf{then}}\;p\;{\mathsf{else}}\;q\;{\mathsf{fi}}}\:\{Q,\:G\}}$.
Let ${\mathit{sq}}\in {\mathcal{E}\!\mathit{nv}\:R} \cap {\mathcal{I}\!\mathit{n}\:P} \cap {\lsemantic{{\mathsf{if}}\:C\;{\mathsf{then}}\;p\;{\mathsf{else}}\;q\;{\mathsf{fi}}}\rsemantic_{\rho}}$. If ${\mathit{sq}}$ does not make a program step then we are done. Otherwise, assume ${\rho \vdash {\mathit{sq}}_i {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}{\mathit{sq}}_{i+1}}$ with $i < |{\mathit{sq}}| - 1$ being the first such index. Then ${\mathit{sq}}_i = ({{\mathsf{if}}\:C\;{\mathsf{then}}\;p\;{\mathsf{else}}\;q\;{\mathsf{fi}}}, \sigma)$ and ${\mathit{sq}}_{i+1} = (x, \sigma)$ with some $x$ and $\sigma \in P$ due to the stability assumption (1). If $\sigma \in C$ then $x=p$ and ${^{\scriptscriptstyle i+1}|{\mathit{sq}}}\in {\mathcal{E}\!\mathit{nv}\:R} \cap {\mathcal{I}\!\mathit{n}\:P} \cap {\lsemanticp\rsemantic_{\rho}}$. Hence, ${^{\scriptscriptstyle i+1}|{\mathit{sq}}}\in {\mathcal{O}\!\mathit{ut}\:Q} \cap {\mathcal{P}\!\mathit{rog}\:G}$ by (2), from which ${\mathit{sq}}\in {\mathcal{O}\!\mathit{ut}\:Q} \cap {\mathcal{P}\!\mathit{rog}\:G}$ follows using reflexivity of $G$. The case $\sigma \notin C$ is symmetric.
A brief summary
---------------
Starting with the definition of extended Hoare triples, this section presented the program correspondence rule as well as a program logic rule for each of the language constructors, except ${{\mathsf{cjump}}\:\:C\;{\mathsf{to}}\:j\;{\mathsf{otherwise}}\;p\;{\mathsf{end}}}$, which is surely not because it is impossible: the equivalence in Proposition \[thm:cond-norm2\] combined with Proposition \[thm:cond-rule\] gives a rule for ${\mathsf{cjump}}$, which, however, would have ${\rho\models\{R,\:P \cap C\}\:{\rho\:j}\:\{Q,\:G\}}$ as an assumption about a jump destination.
Case Study: Peterson’s Mutual Exclusion Algorithm {#S:PM1}
=================================================
This section embarks on the first part of the case study employing the techniques developed so far.
$$\begin{array}{l l l}
{\mathsf{thread}}_0 \; {\mathit{cs}}_0 & {\mathrel{\stackrel{\scriptscriptstyle{\mathit{def}}}{=}}}& {\mathbf{flag}}_0 := {\mathsf{True}};\\
& & {\mathbf{turn}} := {\mathsf{True}};\\
& & {{\mathsf{while}}\:{\mathbf{flag}}_1 \wedge {\mathbf{turn}}\;{\mathsf{do}}\;{{\mathsf{skip}}}\;{\mathsf{od}}};\\
& & {\mathit{cs}}_0; \\
& & {\mathbf{flag}}_0 := {\mathsf{False}}\\
& & \\
& & \\
{\mathsf{thread}}_1 \; {\mathit{cs}}_1 & {\mathrel{\stackrel{\scriptscriptstyle{\mathit{def}}}{=}}}& {\mathbf{flag}}_1 := {\mathsf{True}};\\
& & {\mathbf{turn}} := {\mathsf{False}};\\
& & {{\mathsf{while}}\:{\mathbf{flag}}_0 \wedge \neg{\mathbf{turn}}\;{\mathsf{do}}\;{{\mathsf{skip}}}\;{\mathsf{od}}};\\
& & {\mathit{cs}}_1;\\
& & {\mathbf{flag}}_1 := {\mathsf{False}}\\
\end{array}$$
The definitions of ${\mathsf{thread}}_0$ and ${\mathsf{thread}}_1$, shown in Figure \[fig:pm\], shall represent the threads in the well-known Peterson’s mutual exclusion algorithm [@PETERSON1981115]. They are parameterised by ${\mathit{cs}}_0$ and ${\mathit{cs}}_1$, which can be instantiated by any term of type $\mathcal{L}_\tau$ for any specific type $\tau$, and regarded as placeholders for whatever needs to be performed within the critical sections. Thus, the definition $${\mathsf{mutex}}\; {\mathit{cs}}_0 \; {\mathit{cs}}_1 {\mathrel{\stackrel{\scriptscriptstyle{\mathit{def}}}{=}}}{\mathsf{thread}}_0 \; {\mathit{cs}}_0 \parallel {\mathsf{thread}}_1 \; {\mathit{cs}}_1$$ models the entire protocol, featuring abstraction over the contents of critical sections. Note that this model is not tailored to be transformed to some low-level representations, in particular due to the compound conditions ${\mathbf{flag}}_1 \wedge {\mathbf{turn}}$ and ${\mathbf{flag}}_0 \wedge \neg{\mathbf{turn}}$ in the ${\mathsf{while}}$-statements. However, in this form it suits more the actual purpose of the case study: to elicit the logical principles behind the algorithm without delving into technical details. More specifically, the central goal of this section is to establish that ${\mathsf{mutex}}$ enables certain privileged conditions to ${\mathit{cs}}_0$ and ${\mathit{cs}}_1$ regarding interference, despite that ${\mathsf{thread}}_0$ and ${\mathsf{thread}}_1$ can run in an arbitrarily interleaved way.
In order to pinpoint what these privileged conditions are, let ${\mathbf{shared}}$ be a variable representing some resource that both ${\mathit{cs}}_0$ and ${\mathit{cs}}_1$ are entitled to access for reading and writing. Moreover, let $P_0, Q_0$ and $P_1, Q_1$ be predicates that can depend on the values of ${\mathbf{shared}}$ such that, $P_0 \;{\mathbf{shared}}$ can be used to specify some conditions on these values before entering the critical section ${\mathit{cs}}_0$, whereas $Q_0 \;{\mathbf{shared}}$ – when ${\mathit{cs}}_0$ terminates. Thus, specifications of the critical sections can be sketched by means of the following extended Hoare triples: $$\begin{array}{l}
{\models\{{\mathbf{shared}} = {\mathbf{shared}}{^{\scriptscriptstyle\prime}}\; \ldots,\:P_0 \:{\mathbf{shared}}\}\:{\mathit{cs}}_0\:\{Q_0 \:{\mathbf{shared}},\:\ldots\}}\\
{\models\{{\mathbf{shared}} = {\mathbf{shared}}{^{\scriptscriptstyle\prime}}\; \ldots,\:P_1 \:{\mathbf{shared}}\}\:{\mathit{cs}}_1\:\{Q_1 \:{\mathbf{shared}},\:\ldots\}}
\end{array}$$ Among the relies, ${\mathbf{shared}} = {\mathbf{shared}}{^{\scriptscriptstyle\prime}}$ essentially captures a half of the privileged conditions mentioned above: each critical section can treat ${\mathbf{shared}}$ as its local variable. Another half of the privileged conditions will consequently amount to *the absence* of ${\mathbf{shared}} = {\mathbf{shared}}{^{\scriptscriptstyle\prime}}$ among the yet unspecified guarantees: its presence there would prohibit any modification of ${\mathbf{shared}}$ within ${\mathit{cs}}_0$ and ${\mathit{cs}}_1$, making the whole abstraction straight away useless.
The parallel increment example in Section \[Sb:parallel-rule\] has stressed the occasional necessity to appeal to auxiliary variables in order to make the parallel composition rule applicable. And they will play a rôle here too: the verification process requires access to the information whether a thread has performed its assignment to ${\mathbf{turn}}$ or not. This can be indicated only by means of an auxiliary variable. The definitions of ${\mathsf{thread\_aux}}_0$ and ${\mathsf{thread\_aux}}_1$,
$$\begin{array}{l l l}
{\mathsf{thread\_aux}}_0 \; {\mathit{cs}}_0 & {\mathrel{\stackrel{\scriptscriptstyle{\mathit{def}}}{=}}}& {\mathbf{flag}}_0 := {\mathsf{True}};\\
& & \langle {\mathbf{turn}} := {\mathsf{True}};\\
& & \;\; {\mathbf{turn\_aux}}_0 := {\mathsf{True}}\rangle;\\
& & {{\mathsf{while}}\:{\mathbf{flag}}_1 \wedge {\mathbf{turn}}\;{\mathsf{do}}\;{{\mathsf{skip}}}\;{\mathsf{od}}}\\
& & {\mathit{cs}}_0; \\
& & {\mathbf{flag}}_0 := {\mathsf{False}}\\
& & \\
& & \\
{\mathsf{thread\_aux}}_1 \; {\mathit{cs}}_1 & {\mathrel{\stackrel{\scriptscriptstyle{\mathit{def}}}{=}}}& {\mathbf{flag}}_1 := {\mathsf{True}};\\
& &\langle {\mathbf{turn}} := {\mathsf{False}};\\
& & \;\; {\mathbf{turn\_aux}}_1 := {\mathsf{True}}\rangle;\\
& & {{\mathsf{while}}\:{\mathbf{flag}}_0 \wedge \neg{\mathbf{turn}}\;{\mathsf{do}}\;{{\mathsf{skip}}}\;{\mathsf{od}}}\\
& & {\mathit{cs}}_1; \\
& & {\mathbf{flag}}_1 := {\mathsf{False}}\\
\end{array}$$
shown in Figure \[fig:pm-aux\], employ ${\mathbf{turn\_aux}}_0$ and ${\mathbf{turn\_aux}}_1$ to this end. Then the auxiliary model of the mutual exclusion algorithm is defined by $${\mathsf{mutex\_aux}}\; {\mathit{cs}}_0 \; {\mathit{cs}}_1 {\mathrel{\stackrel{\scriptscriptstyle{\mathit{def}}}{=}}}{\mathsf{thread\_aux}}_0 \; {\mathit{cs}}_0 \parallel {\mathsf{thread\_aux}}_1 \; {\mathit{cs}}_1$$
The states, on which ${\mathsf{mutex}}$ and ${\mathsf{mutex\_aux}}$ operate, are shaped as follows. Firstly, ${\mathbf{flag}}_0, {\mathbf{flag}}_1,$ and ${\mathbf{turn}}$ carry the Boolean values used for communication between threads. Secondly, we have the Boolean values ${\mathbf{turn\_aux}}_0$ and ${\mathbf{turn\_aux}}_1$ used for verification purposes explained above. Thirdly, there are ${\mathbf{shared}}, {\mathbf{local}}_0, {\mathbf{local}}_1$ whose exact type can remain open yet. The anticipated purpose of ${\mathbf{local}}_0$ and ${\mathbf{local}}_1$ is to allow ${\mathit{cs}}_0$ and ${\mathit{cs}}_1$ to make local copies of the shared resource and modify these before submitting the results back to ${\mathbf{shared}}$.
Now, from the logical perspective the conditions $$\begin{array}{l c l}
{\mathbf{cond}}_0 & {\mathrel{\stackrel{\scriptscriptstyle{\mathit{def}}}{=}}}& {\mathbf{flag}}_0 \wedge (\neg{\mathbf{turn\_aux}}_1 \vee \neg{\mathbf{flag}}_1 \vee \neg{\mathbf{turn}})\\
{\mathbf{cond}}_1 & {\mathrel{\stackrel{\scriptscriptstyle{\mathit{def}}}{=}}}& {\mathbf{flag}}_1 \wedge (\neg{\mathbf{turn\_aux}}_0 \vee \neg{\mathbf{flag}}_0 \vee {\mathbf{turn}})
\end{array}$$ are at the core of the algorithm: if ${\mathbf{turn\_aux}}_0$ and ${\mathbf{turn\_aux}}_1$ hold in a state then ${\mathbf{cond}}_0$ and ${\mathbf{cond}}_1$ cannot both be true in this state.
With all these preparations, we can formulate the following statement for ${\mathsf{thread}}_0$: $$\label{eq:pm1}
\!\!\!{\models\{R_0,\:\neg{\mathbf{turn\_aux}}_0 \wedge P_0\: {\mathbf{shared}}\}\:{\mathsf{thread\_aux}}_0 \; {\mathit{cs}}_0\:\{Q_0\: {\mathbf{shared}},\:R_1\}}$$ where $R_0$ abbreviates the rely condition $$\begin{array}{l r}
\!\!\!\!\! {\mathbf{flag}}_0 ={\mathbf{flag}}{^{\scriptscriptstyle\prime}}_0 \wedge {\mathbf{turn\_aux}}_0 = {\mathbf{turn\_aux}}{^{\scriptscriptstyle\prime}}_0 \wedge {\mathbf{local}}_0 = {\mathbf{local}}{^{\scriptscriptstyle\prime}}_0\; \wedge & \hspace{2.7cm}\textrm{(a)} \\
\!\!\!\!\! ({\mathbf{turn\_aux}}_0 \wedge {\mathbf{cond}}_0 \longrightarrow {\mathbf{shared}} = {\mathbf{shared}}{^{\scriptscriptstyle\prime}})\; \wedge & \textrm{(b)}\\
\!\!\!\!\! (\neg{\mathbf{turn\_aux}}_1 \wedge {\mathbf{turn\_aux}}{^{\scriptscriptstyle\prime}}_1 \longrightarrow \neg{\mathbf{turn}}{^{\scriptscriptstyle\prime}})\; \wedge & \textrm{(c)}\\
\!\!\!\!\! ({\mathbf{turn\_aux}}_1 \longrightarrow {\mathbf{turn}} = {\mathbf{turn}}{^{\scriptscriptstyle\prime}}) \; \wedge & \textrm{(d)}\\
\!\!\!\!\! ({\mathbf{turn\_aux}}_1 \wedge {\mathbf{flag}}{^{\scriptscriptstyle\prime}}_1 \longrightarrow {\mathbf{flag}}_1)\; \wedge & \textrm{(e)}\\
\!\!\!\!\! (P_0\: {\mathbf{shared}} \longrightarrow P_0\: {\mathbf{shared}}{^{\scriptscriptstyle\prime}}) \wedge (Q_0\: {\mathbf{shared}} \longrightarrow Q_0\: {\mathbf{shared}}{^{\scriptscriptstyle\prime}}) & \textrm{(f)}
\end{array}$$ specifying that ${\mathsf{thread\_aux}}_0$ can rely on
1. ${\mathsf{thread\_aux}}_1$ never modifies ${\mathbf{flag}}_0$, ${\mathbf{turn\_aux}}_0$ and ${\mathbf{local}}_0$;
2. if ${\mathsf{thread\_aux}}_1$ makes a step from a state where ${\mathbf{turn\_aux}}_0$ and ${\mathbf{cond}}_0$ hold then ${\mathbf{shared}}$ will retain its value to the next state;
3. if ${\mathsf{thread\_aux}}_1$ makes a step from $\sigma$ to $\sigma{^{\scriptscriptstyle\prime}}$ such that $\neg \sigma{\mathbf{turn\_aux}}_1$ and $\sigma{^{\scriptscriptstyle\prime}}{\mathbf{turn\_aux}}_1$ hold then $\neg\sigma{^{\scriptscriptstyle\prime}}{\mathbf{turn}}$ must hold as well: this condition addresses exactly the transition where ${\mathsf{thread\_aux}}_1$ makes its assignment to ${\mathbf{turn}}$;
4. ${\mathsf{thread\_aux}}_1$ will not modify ${\mathbf{turn}}$ after ${\mathbf{turn\_aux}}_1$ was set;
5. ${\mathsf{thread\_aux}}_1$ can only reset ${\mathbf{flag}}_1$ after ${\mathbf{turn\_aux}}_1$ was set;
6. the conditions $P_0\: {\mathbf{shared}} \longrightarrow P_0\: {\mathbf{shared}}{^{\scriptscriptstyle\prime}}$ and $Q_0\: {\mathbf{shared}} \longrightarrow Q_0\: {\mathbf{shared}}{^{\scriptscriptstyle\prime}}$ capture the point that ${\mathsf{thread\_aux}}_1$ shall not compromise what ${\mathsf{thread\_aux}}_0$ is trying to accomplish; as a result, this can further restrict actions of ${\mathsf{thread\_aux}}_1$ on the shared resource, but is necessary in order to achieve some meaningful cooperation of the two threads.
The guarantee $R_1$ in (\[eq:pm1\]) is symmetric to $R_0$: $$\begin{array}{l}
{\mathbf{flag}}_1 ={\mathbf{flag}}{^{\scriptscriptstyle\prime}}_1 \wedge {\mathbf{turn\_aux}}_1 = {\mathbf{turn\_aux}}{^{\scriptscriptstyle\prime}}_1 \wedge {\mathbf{local}}_1 = {\mathbf{local}}{^{\scriptscriptstyle\prime}}_1\; \wedge\\
({\mathbf{turn\_aux}}_1 \wedge {\mathbf{cond}}_1 \longrightarrow {\mathbf{shared}} = {\mathbf{shared}}{^{\scriptscriptstyle\prime}})\; \wedge \\
(\neg{\mathbf{turn\_aux}}_0 \wedge {\mathbf{turn\_aux}}{^{\scriptscriptstyle\prime}}_0 \longrightarrow {\mathbf{turn}}{^{\scriptscriptstyle\prime}})\; \wedge \\
({\mathbf{turn\_aux}}_0 \longrightarrow {\mathbf{turn}} = {\mathbf{turn}}{^{\scriptscriptstyle\prime}})\; \wedge \\
({\mathbf{turn\_aux}}_0 \wedge {\mathbf{flag}}{^{\scriptscriptstyle\prime}}_0 \longrightarrow {\mathbf{flag}}_0)\; \wedge \\
(P_1\: {\mathbf{shared}} \longrightarrow P_1\: {\mathbf{shared}}{^{\scriptscriptstyle\prime}}) \wedge (Q_1\: {\mathbf{shared}} \longrightarrow Q_1\: {\mathbf{shared}}{^{\scriptscriptstyle\prime}})
\end{array}$$ and will accordingly serve as the rely for ${\mathsf{thread\_aux}}_1$ in the triple $${\models\{R_1,\:\neg{\mathbf{turn\_aux}}_1 \wedge P_1\: {\mathbf{shared}}\}\:{\mathsf{thread\_aux}}_1 \; {\mathit{cs}}_1\:\{Q_1\: {\mathbf{shared}},\:R_0\}}$$ Further, the extended Hoare triple $$\label{eq:pm3}
{\models\{{\mathbf{shared}} = {\mathbf{shared}}{^{\scriptscriptstyle\prime}}\wedge {\mathbf{local}}_0 = {\mathbf{local}}{^{\scriptscriptstyle\prime}}_0,\:P_0 \:{\mathbf{shared}}\}\:{\mathit{cs}}_0\:\{Q_0 \:{\mathbf{shared}},\:G_0\}}$$ specifies the behaviour of ${\mathit{cs}}_0$, where $G_0$ stands for the guarantee condition $$\begin{array}{l r}
\!\!\!{\mathbf{flag}}_0 ={\mathbf{flag}}{^{\scriptscriptstyle\prime}}_0 \wedge {\mathbf{flag}}_1 ={\mathbf{flag}}{^{\scriptscriptstyle\prime}}_1 \wedge {\mathbf{turn}} ={\mathbf{turn}}{^{\scriptscriptstyle\prime}}\wedge & \hspace{0.7cm}\mbox{(a)}\\
\!\!\!{\mathbf{local}}_1 = {\mathbf{local}}{^{\scriptscriptstyle\prime}}_1 \wedge {\mathbf{turn\_aux}}_0 = {\mathbf{turn\_aux}}{^{\scriptscriptstyle\prime}}_0 \wedge {\mathbf{turn\_aux}}_1 = {\mathbf{turn\_aux}}{^{\scriptscriptstyle\prime}}_1 \wedge & \mbox{(b)}\\
\!\!\!(P_1\: {\mathbf{shared}} \longrightarrow P_1\: {\mathbf{shared}}{^{\scriptscriptstyle\prime}}) \wedge (Q_1\: {\mathbf{shared}} \longrightarrow Q_1\: {\mathbf{shared}}{^{\scriptscriptstyle\prime}}) & \mbox{(c)}
\end{array}$$ Note once more the presence of ${\mathbf{shared}} = {\mathbf{shared}}{^{\scriptscriptstyle\prime}}$ among the rely conditions and its absence in $G_0$ which, however, does not leave indivisible steps of ${\mathit{cs}}_0$ completely unrestricted requiring from ${\mathit{cs}}_0$
1. not to modify ${\mathbf{flag}}_0, {\mathbf{flag}}_1, {\mathbf{turn}}$, as these are used for communication between threads;
2. not to modify the variable ${\mathbf{local}}_1$, which is supposed to be local to ${\mathsf{thread\_aux}}_1$, as well as the auxiliary variables;
3. to retain the conditions $P_1\: {\mathbf{shared}}$ and $Q_1\: {\mathbf{shared}}$ in line with the guarantees of ${\mathsf{thread\_aux}}_0$.
The specification of ${\mathit{cs}}_1$ consequently assumes the following form: $${\models\{{\mathbf{shared}} = {\mathbf{shared}}{^{\scriptscriptstyle\prime}}\wedge {\mathbf{local}}_1 = {\mathbf{local}}{^{\scriptscriptstyle\prime}}_1,\:P_1 \:{\mathbf{shared}}\}\:{\mathit{cs}}_1\:\{Q_1 \:{\mathbf{shared}},\:G_1\}}$$ where $G_1$ abbreviates $$\begin{array}{l}
{\mathbf{flag}}_0 ={\mathbf{flag}}{^{\scriptscriptstyle\prime}}_0 \wedge {\mathbf{flag}}_1 ={\mathbf{flag}}{^{\scriptscriptstyle\prime}}_1 \wedge {\mathbf{turn}} ={\mathbf{turn}}{^{\scriptscriptstyle\prime}}\wedge \\
{\mathbf{local}}_0 = {\mathbf{local}}{^{\scriptscriptstyle\prime}}_0 \wedge {\mathbf{turn\_aux}}_0 = {\mathbf{turn\_aux}}{^{\scriptscriptstyle\prime}}_0 \wedge {\mathbf{turn\_aux}}_1 = {\mathbf{turn\_aux}}{^{\scriptscriptstyle\prime}}_1 \\
(P_0\: {\mathbf{shared}} \longrightarrow P_0\: {\mathbf{shared}}{^{\scriptscriptstyle\prime}}) \wedge (Q_0\: {\mathbf{shared}} \longrightarrow Q_0\: {\mathbf{shared}}{^{\scriptscriptstyle\prime}})
\end{array}$$ the counterpart to $G_0$.
Processing ${\mathsf{thread\_aux}}_0$
-------------------------------------
In order to establish (\[eq:pm1\]), for the first part of ${\mathsf{thread\_aux}}_0$ up to the critical section $$\begin{array}{l l l}
{\mathsf{thread\_aux}}^{{\mathit{pre}}}_0 & {\mathrel{\stackrel{\scriptscriptstyle{\mathit{def}}}{=}}}& {\mathbf{flag}}_0 := {\mathsf{True}};\\
& & \langle {\mathbf{turn}} := {\mathsf{True}};\\
& & \;\; {\mathbf{turn\_aux}}_0 := {\mathsf{True}}\rangle;\\
& & {{\mathsf{while}}\:{\mathbf{flag}}_1 \wedge {\mathbf{turn}}\;{\mathsf{do}}\;{{\mathsf{skip}}}\;{\mathsf{od}}}
\end{array}$$ we can derive $$\begin{array}{l}
{\models\{R_0,\:\neg{\mathbf{turn\_aux}}_0 \wedge P_0\: {\mathbf{shared}}\}\:\\\quad{\mathsf{thread\_aux}}^{{\mathit{pre}}}_0\\\quad\:\{{\mathbf{flag}}_0 \wedge {\mathbf{turn\_aux}}_0 \wedge P_0\: {\mathbf{shared}} \wedge \neg({\mathbf{flag}}_1 \wedge {\mathbf{turn}}),\:R_1\}}
\end{array}$$ using ${\mathbf{flag}}_0 \wedge \neg{\mathbf{turn\_aux}}_0 \wedge P_0\: {\mathbf{shared}}$ and ${\mathbf{flag}}_0 \wedge {\mathbf{turn\_aux}}_0 \wedge P_0\: {\mathbf{shared}}$ as intermediate conditions. Note that both are stable under $R_0$, and the latter is invariant to the ${\mathsf{while}}$-statement. This yields the following goal for the critical section: $$\label{eq:pm2}
\!\!\!\!\!\!\!\!\! {\models\{R_0,\:{\mathbf{flag}}_0 \wedge {\mathbf{turn\_aux}}_0 \wedge P_0\: {\mathbf{shared}} \wedge \neg({\mathbf{flag}}_1 \wedge {\mathbf{turn}})\}\:{\mathit{cs}}_0\:\{Q_0\: {\mathbf{shared}},\:R_1\}}$$ According to the assumption (\[eq:pm3\]), ${\mathit{cs}}_0$ can rely on the condition ${\mathbf{shared}} = {\mathbf{shared}}{^{\scriptscriptstyle\prime}}$. On the other hand, $R_0$ provides only ${\mathbf{turn\_aux}}_0 \wedge {\mathbf{cond}}_0 \longrightarrow {\mathbf{shared}} = {\mathbf{shared}}{^{\scriptscriptstyle\prime}}$, making plain rely-weakening impossible. Hence, to apply the program correspondence rule (Proposition \[thm:pcorr-rule\]) a suitable state relation is sought.
To this end, notice that ${\mathbf{flag}}_0 \wedge \neg({\mathbf{flag}}_1 \wedge {\mathbf{turn}})$ implies ${\mathbf{cond}}_0$ and is present in the precondition of (\[eq:pm2\]). Then we define $$\begin{array}{l c l}
{\mathsf{r}}_0 & {\mathrel{\stackrel{\scriptscriptstyle{\mathit{def}}}{=}}}& \{(\sigma, \sigma{^{\scriptscriptstyle\prime}}) \:|\: \sigma = \sigma{^{\scriptscriptstyle\prime}}\wedge \sigma{\mathbf{turn\_aux}}_0 \wedge \sigma{\mathbf{cond}}_0 \}
\end{array}$$ such that ${\mathsf{r}}_0 \subseteq {\mathsf{id}}$, ${\mathsf{r}}_0$ is a coreflexive state relation which shall act as an invariant for the critical section ${\mathit{cs}}_0$. The additional assumption $$\label{eq:pm4}
{\models {\mathit{cs}}_0\sqsupseteq_{{\mathsf{r}}_0}{\mathit{cs}}_0}$$ captures this partly, as it does not take accounts of interleaving, stability of ${\mathsf{r}}_0$ under $R_0$. This part will be addressed in (1) below. Also note that it is reasonable to assume (\[eq:pm4\]) as it holds for any instance of ${\mathit{cs}}_0$ which does not modify ${\mathbf{flag}}_0$, ${\mathbf{flag}}_1$, ${\mathbf{turn}}$ and the auxiliary variables.
With (\[eq:pm3\]) and (\[eq:pm4\]) the program correspondence rule becomes applicable because
1. ${{\mathsf{r}}_0 \diamond R_0} \subseteq {({\mathbf{shared}} = {\mathbf{shared}}{^{\scriptscriptstyle\prime}}\wedge {\mathbf{local}}_0 = {\mathbf{local}}{^{\scriptscriptstyle\prime}}_0) \diamond {\mathsf{r}}_0}$ holds, as explained in detail below;
2. ${\mathbf{flag}}_0 \wedge {\mathbf{turn\_aux}}_0 \wedge P_0\: {\mathbf{shared}} \wedge \neg({\mathbf{flag}}_1 \wedge {\mathbf{turn}}) \subseteq {{\mathsf{r}}_0\cdot (P_0\: {\mathbf{shared}})}$ holds since ${\mathbf{flag}}_0 \wedge \neg({\mathbf{flag}}_1 \wedge {\mathbf{turn}})$ implies ${\mathbf{cond}}_0$;
3. ${{\mathsf{r}}_0\cdot (Q_0 \:{\mathbf{shared}})} \subseteq Q_0 \:{\mathbf{shared}}$ holds due to the coreflexivity of ${\mathsf{r}}_0$;
4. ${{{\mathsf{r}}_0 \diamond G_0} \diamond {\mathsf{r}}_0} \subseteq R_1$ holds since ${\mathbf{turn\_aux}}_0 \wedge {\mathbf{cond}}_0$ and ${\mathbf{turn\_aux}}_1 \wedge {\mathbf{cond}}_1$ are mutually exclusive and therefore trivially imply ${\mathbf{shared}} = {\mathbf{shared}}{^{\scriptscriptstyle\prime}}$.
That ${{\mathsf{r}}_0 \diamond R_0} \subseteq {({\mathbf{shared}} = {\mathbf{shared}}{^{\scriptscriptstyle\prime}}\wedge {\mathbf{local}}_0 = {\mathbf{local}}{^{\scriptscriptstyle\prime}}_0) \diamond {\mathsf{r}}_0}$ holds is one of the most crucial points in this context. In essence, it amounts to showing that $\sigma{\mathbf{shared}} = \sigma{^{\scriptscriptstyle\prime}}{\mathbf{shared}}$ and $\sigma{^{\scriptscriptstyle\prime}}{\mathbf{turn\_aux}}_0 \wedge \sigma{^{\scriptscriptstyle\prime}}{\mathbf{cond}}_0$ follow from $\sigma{\mathbf{turn\_aux}}_0 \wedge \sigma{\mathbf{cond}}_0$ and $(\sigma, \sigma{^{\scriptscriptstyle\prime}}) \in R_0$.
Firstly, note that $(\sigma, \sigma{^{\scriptscriptstyle\prime}}) \in R_0$ simply subsumes that $\sigma{\mathbf{shared}} = \sigma{^{\scriptscriptstyle\prime}}{\mathbf{shared}}$ is implied by $\sigma{\mathbf{turn\_aux}}_0 \wedge \sigma{\mathbf{cond}}_0$, whereas deriving the condition $\sigma{^{\scriptscriptstyle\prime}}{\mathbf{turn\_aux}}_0 \wedge \sigma{^{\scriptscriptstyle\prime}}{\mathbf{cond}}_0$ corresponds to showing stability of ${\mathsf{r}}_0$ under $R_0$, which complements (\[eq:pm4\]) as mentioned above.
To this end, note that $\sigma{^{\scriptscriptstyle\prime}}{\mathbf{turn\_aux}}_0$ and $\sigma{^{\scriptscriptstyle\prime}}{\mathbf{flag}}_0$ follow immediately from $\sigma{\mathbf{turn\_aux}}_0$, $\sigma{\mathbf{flag}}_0$ and $(\sigma, \sigma{^{\scriptscriptstyle\prime}}) \in R_0$, such that only the condition $\neg\sigma{^{\scriptscriptstyle\prime}}{\mathbf{turn\_aux}}_1 \vee \neg\sigma{^{\scriptscriptstyle\prime}}{\mathbf{flag}}_1 \vee \neg\sigma{^{\scriptscriptstyle\prime}}{\mathbf{turn}}$ remains to be shown. Assuming $\sigma{^{\scriptscriptstyle\prime}}{\mathbf{turn\_aux}}_1, \sigma{^{\scriptscriptstyle\prime}}{\mathbf{flag}}_1$ and $\sigma{^{\scriptscriptstyle\prime}}{\mathbf{turn}}$, we consider the cases if $\sigma{\mathbf{turn\_aux}}_1$ holds.
If it does then from $(\sigma, \sigma{^{\scriptscriptstyle\prime}}) \in R_0$ we can first obtain $\sigma{\mathbf{flag}}_1$, since $\sigma{\mathbf{turn\_aux}}_1$ and $\sigma{^{\scriptscriptstyle\prime}}{\mathbf{flag}}_1$ hold. Hence, from $\sigma{\mathbf{cond}}_0$ we can infer $\neg\sigma{\mathbf{turn}}$. Furthermore, $(\sigma, \sigma{^{\scriptscriptstyle\prime}}) \in R_0$ also subsumes that $\sigma{\mathbf{turn\_aux}}_1$ implies $\sigma{\mathbf{turn}} = \sigma{^{\scriptscriptstyle\prime}}{\mathbf{turn}}$, from which $\neg\sigma{^{\scriptscriptstyle\prime}}{\mathbf{turn}}$ follows.
Lastly, assume $\neg\sigma{\mathbf{turn\_aux}}_1$. This situation arises exactly at the point when ${\mathsf{thread\_aux}}_1$ makes its assignment to ${\mathbf{turn}}$, which is captured in $R_0$ by means of the condition that $\neg\sigma{\mathbf{turn\_aux}}_1$ and $\sigma{^{\scriptscriptstyle\prime}}{\mathbf{turn\_aux}}_1$ imply $\neg\sigma{^{\scriptscriptstyle\prime}}{\mathbf{turn}}$.
Summarising the intermediate results
------------------------------------
What has been achieved so far was to derive the extended Hoare triple $${\models\{R_0,\:\neg{\mathbf{turn\_aux}}_0 \wedge P_0\: {\mathbf{shared}}\}\:{\mathsf{thread\_aux}}_0 \; {\mathit{cs}}_0\:\{Q_0\: {\mathbf{shared}},\:R_1\}}$$ under the assumptions $${\models\{{\mathbf{shared}} = {\mathbf{shared}}{^{\scriptscriptstyle\prime}}\wedge {\mathbf{local}}_0 = {\mathbf{local}}{^{\scriptscriptstyle\prime}}_0,\:P_0 \:{\mathbf{shared}}\}\:{\mathit{cs}}_0\:\{Q_0 \:{\mathbf{shared}},\:G_0\}}$$ and $
{\models {\mathit{cs}}_0\sqsupseteq_{{\mathsf{r}}_0}{\mathit{cs}}_0}.
$
In an entirely symmetric way we can also derive $${\models\{R_1,\:\neg{\mathbf{turn\_aux}}_1 \wedge P_1\: {\mathbf{shared}}\}\:{\mathsf{thread\_aux}}_1 \; {\mathit{cs}}_1\:\{Q_1\: {\mathbf{shared}},\:R_0\}}$$ under the assumptions $${\models\{{\mathbf{shared}} = {\mathbf{shared}}{^{\scriptscriptstyle\prime}}\wedge {\mathbf{local}}_1 = {\mathbf{local}}{^{\scriptscriptstyle\prime}}_1,\:P_1 \:{\mathbf{shared}}\}\:{\mathit{cs}}_1\:\{Q_1 \:{\mathbf{shared}},\:G_1\}}$$ and $
{\models {\mathit{cs}}_1\sqsupseteq_{{\mathsf{r}}_1}{\mathit{cs}}_1}
$ where we accordingly define $$\begin{array}{l c l}
{\mathsf{r}}_1 & {\mathrel{\stackrel{\scriptscriptstyle{\mathit{def}}}{=}}}& \{(\sigma, \sigma{^{\scriptscriptstyle\prime}}) \:|\: \sigma = \sigma{^{\scriptscriptstyle\prime}}\wedge \sigma{\mathbf{turn\_aux}}_1 \wedge \sigma{\mathbf{cond}}_1 \}
\end{array}$$ The following proposition summarises the intermediate result.
\[thm:mutex-aux\] Assume
1. ${\models\{{\mathbf{shared}} = {\mathbf{shared}}{^{\scriptscriptstyle\prime}}\wedge {\mathbf{local}}_0 = {\mathbf{local}}{^{\scriptscriptstyle\prime}}_0,\:P_0 \:{\mathbf{shared}}\}\:{\mathit{cs}}_0\:\{Q_0 \:{\mathbf{shared}},\:G_0\}}$,
2. ${\models\{{\mathbf{shared}} = {\mathbf{shared}}{^{\scriptscriptstyle\prime}}\wedge {\mathbf{local}}_1 = {\mathbf{local}}{^{\scriptscriptstyle\prime}}_1,\:P_1 \:{\mathbf{shared}}\}\:{\mathit{cs}}_1\:\{Q_1 \:{\mathbf{shared}},\:G_1\}}$,
3. ${\models {\mathit{cs}}_0\sqsupseteq_{{\mathsf{r}}_0}{\mathit{cs}}_0}$,
4. ${\models {\mathit{cs}}_1\sqsupseteq_{{\mathsf{r}}_1}{\mathit{cs}}_1}$.
Then $${\begin{array}{l l}
\models & \{{\mathsf{id}},\: P_0\: {\mathbf{shared}} \wedge P_1\: {\mathbf{shared}} \wedge \neg{\mathbf{turn\_aux}}_0 \wedge \neg{\mathbf{turn\_aux}}_1\} \\
& {\mathsf{thread\_aux}}_0 \; {\mathit{cs}}_0 \parallel {\mathsf{thread\_aux}}_1 \; {\mathit{cs}}_1 \\
& \{ Q_0\: {\mathbf{shared}} \wedge Q_1\: {\mathbf{shared}},\:\top\}
\end{array}}$$
By the parallel composition rule (Corollary \[thm:parallel-rule2\]).
The next step will be to derive the corresponding rule for ${\mathsf{mutex}}$.
Removing the auxiliaries
------------------------
We once more resort to the program correspondence rule, now using the state relation $$\begin{array}{l c l l}
{\mathsf{r}}_{{\mathsf{eqv}}} & {\mathrel{\stackrel{\scriptscriptstyle{\mathit{def}}}{=}}}& \{ (\sigma, \sigma{^{\scriptscriptstyle\prime}}) \:|\: & \sigma{\mathbf{flag}}_0 = \sigma{^{\scriptscriptstyle\prime}}{\mathbf{flag}}_0 \wedge \\
& & & \sigma{\mathbf{flag}}_1 = \sigma{^{\scriptscriptstyle\prime}}{\mathbf{flag}}{^{\scriptscriptstyle\prime}}_1 \wedge \\
& & & \sigma{\mathbf{turn}} = \sigma{^{\scriptscriptstyle\prime}}{\mathbf{turn}}{^{\scriptscriptstyle\prime}}\wedge \\
& & & \sigma{\mathbf{shared}} = \sigma{^{\scriptscriptstyle\prime}}{\mathbf{shared}}{^{\scriptscriptstyle\prime}}\wedge \\
& & & \sigma{\mathbf{local}}_0 = \sigma{^{\scriptscriptstyle\prime}}{\mathbf{local}}{^{\scriptscriptstyle\prime}}_0 \wedge \\
& & & \sigma{\mathbf{local}}_1 = \sigma{^{\scriptscriptstyle\prime}}{\mathbf{local}}{^{\scriptscriptstyle\prime}}_1 \}\\
\end{array}$$ the equivalence on states up to the values of auxiliary variables.
\[thm:mutex\] Assume
1. ${\models\{{\mathbf{shared}} = {\mathbf{shared}}{^{\scriptscriptstyle\prime}}\wedge {\mathbf{local}}_0 = {\mathbf{local}}{^{\scriptscriptstyle\prime}}_0,\:P_0 \:{\mathbf{shared}}\}\:{\mathit{cs}}_0\:\{Q_0 \:{\mathbf{shared}},\:G_0\}}$,
2. ${\models\{{\mathbf{shared}} = {\mathbf{shared}}{^{\scriptscriptstyle\prime}}\wedge {\mathbf{local}}_1 = {\mathbf{local}}{^{\scriptscriptstyle\prime}}_1,\:P_1 \:{\mathbf{shared}}\}\:{\mathit{cs}}_1\:\{Q_1 \:{\mathbf{shared}},\:G_1\}}$,
3. ${\models {\mathit{cs}}_0\sqsupseteq_{{\mathsf{r}}_0}{\mathit{cs}}_0}$,
4. ${\models {\mathit{cs}}_1\sqsupseteq_{{\mathsf{r}}_1}{\mathit{cs}}_1}$,
5. ${\models {\mathit{cs}}_0\sqsupseteq_{{\mathsf{r}}_{{\mathsf{eqv}}}}{\mathit{cs}}_0}$,
6. ${\models {\mathit{cs}}_1\sqsupseteq_{{\mathsf{r}}_{{\mathsf{eqv}}}}{\mathit{cs}}_1}$.
Then ${\models\{{\mathsf{id}},\: P_0\: {\mathbf{shared}} \wedge P_1\: {\mathbf{shared}}\}\:{\mathsf{thread}}_0 \; {\mathit{cs}}_0 \parallel {\mathsf{thread}}_1 \; {\mathit{cs}}_1\:\{ Q_0\: {\mathbf{shared}} \wedge Q_1\: {\mathbf{shared}},\:\top\}}$.
Firstly, the extended triple $${\begin{array}{l l}
\models & \{{\mathsf{id}},\: P_0\: {\mathbf{shared}} \wedge P_1\: {\mathbf{shared}} \wedge \neg{\mathbf{turn\_aux}}_0 \wedge \neg{\mathbf{turn\_aux}}_1\} \\
& {\mathsf{thread\_aux}}_0 \; {\mathit{cs}}_0 \parallel {\mathsf{thread\_aux}}_1 \; {\mathit{cs}}_1 \\
& \{ Q_0\: {\mathbf{shared}} \wedge Q_1\: {\mathbf{shared}},\:\top\}
\end{array}}$$ follows from Proposition \[thm:mutex-aux\] using (1) – (4). Secondly, the correspondence $${\models {\mathsf{thread\_aux}}_0 \; {\mathit{cs}}_0 \parallel {\mathsf{thread\_aux}}_1 \; {\mathit{cs}}_1\;\sqsupseteq_{{\mathsf{r}}_{{\mathsf{eqv}}}}\;{\mathsf{thread}}_0 \; {\mathit{cs}}_0 \parallel {\mathsf{thread}}_1 \; {\mathit{cs}}_1}$$ follows from (5) and (6), since $${\models \langle {\mathbf{turn}} := {\mathsf{True}};\; {\mathbf{turn\_aux}}_0 := {\mathsf{True}}\rangle\sqsupseteq_{{\mathsf{r}}_{{\mathsf{eqv}}}}{\mathbf{turn}} := {\mathsf{True}}}$$ and $${\models \langle {\mathbf{turn}} := {\mathsf{False}};\; {\mathbf{turn\_aux}}_1 := {\mathsf{True}}\rangle\sqsupseteq_{{\mathsf{r}}_{{\mathsf{eqv}}}}{\mathbf{turn}} := {\mathsf{False}}}$$ can be derived, using Proposition \[thm:await-corrL\].
Although not present in the conclusion of Proposition \[thm:mutex\], the auxiliary variables still persist in its premises: as the condition ${\mathbf{turn\_aux}}_0 = {\mathbf{turn\_aux}}{^{\scriptscriptstyle\prime}}_0 \wedge {\mathbf{turn\_aux}}_1 = {\mathbf{turn\_aux}}{^{\scriptscriptstyle\prime}}_1$ in both guarantees $G_0$ and $G_1$, and in the state relations ${\mathsf{r}}_0$, ${\mathsf{r}}_1$. This essentially only rules out instances of ${\mathit{cs}}_0$ and ${\mathit{cs}}_1$ which modify or depend on the auxiliaries.
Deploying the mutex {#sub:mutex-inst}
-------------------
In order to illustrate applicability of the rule in Proposition \[thm:mutex\], ${\mathit{cs}}_0$ and ${\mathit{cs}}_1$ will be instantiated by programs, each adding an element to a shared set.
Recall that the type of ${\mathbf{shared}}, {\mathbf{local}}_0$ and ${\mathbf{local}}_1$ was left open, as it did not matter for the actual mutual exclusion protocol. Now, however, we instantiate this type by ${\mathtt{int}} \Rightarrow {\mathtt{bool}}$, sets of integers. Then ${\mathit{cs}}_0$ shall add $0$ to ${\mathbf{shared}}$, whereas ${\mathit{cs}}_1$ – $1$, such that the purpose of the mutual exclusion is to ensure that when the critical sections have accomplished their respective tasks, ${\mathbf{shared}}$ always contains both: $0$ and $1$.
Formally, this is modelled by the definitions $$\begin{array}{l c l}
{\mathsf{shared\_update}}_0 & {\mathrel{\stackrel{\scriptscriptstyle{\mathit{def}}}{=}}}& {\mathbf{local}}_0 := {\mathbf{shared}}; \\
& & {\mathbf{local}}_0 := \{0\} \cup {\mathbf{local}}_0; \\
& & {\mathbf{shared}} := {\mathbf{local}}_0\\
& & \\
{\mathsf{shared\_update}}_1 & {\mathrel{\stackrel{\scriptscriptstyle{\mathit{def}}}{=}}}& {\mathbf{local}}_1 := {\mathbf{shared}}; \\
& & {\mathbf{local}}_1 := \{1\} \cup {\mathbf{local}}_1; \\
& & {\mathbf{shared}} := {\mathbf{local}}_1\\
\end{array}$$ such that the property we want to establish is accordingly $${\models\{{\mathsf{id}},\:\top\}\:{\mathsf{mutex}} \;{\mathsf{shared\_update}}_0 \; {\mathsf{shared\_update}}_1\:\{ 0 \in {\mathbf{shared}} \wedge 1 \in {\mathbf{shared}},\:\top\}}.$$ This conjecture matches the conclusion of Proposition \[thm:mutex\] by setting $P_0, P_1$ to $\top$ and $Q_0, Q_1$ to $0 \in {\mathbf{shared}}, 1 \in {\mathbf{shared}}$, respectively. To match its premises we first need to establish $${\models\{{\mathbf{shared}} = {\mathbf{shared}}{^{\scriptscriptstyle\prime}}\wedge {\mathbf{local}}_0 = {\mathbf{local}}{^{\scriptscriptstyle\prime}}_0,\:\top\}\:{\mathsf{shared\_update}}_0 \:\{0 \in {\mathbf{shared}},\:G{^{\scriptscriptstyle\prime}}_0\}}$$ where $G{^{\scriptscriptstyle\prime}}_0$ is the following instance of $G_0$: $$\begin{array}{l}
{\mathbf{flag}}_0 ={\mathbf{flag}}{^{\scriptscriptstyle\prime}}_0 \wedge {\mathbf{flag}}_1 ={\mathbf{flag}}{^{\scriptscriptstyle\prime}}_1 \wedge {\mathbf{turn}} ={\mathbf{turn}}{^{\scriptscriptstyle\prime}}\wedge \\
{\mathbf{local}}_1 = {\mathbf{local}}{^{\scriptscriptstyle\prime}}_1 \wedge {\mathbf{turn\_aux}}_0 = {\mathbf{turn\_aux}}{^{\scriptscriptstyle\prime}}_0 \wedge {\mathbf{turn\_aux}}_1 = {\mathbf{turn\_aux}}{^{\scriptscriptstyle\prime}}_1 \wedge \\
(1 \in {\mathbf{shared}} \longrightarrow 1 \in {\mathbf{shared}}{^{\scriptscriptstyle\prime}})
\end{array}$$ The most interesting part here is clearly the condition that $1 \in {\mathbf{shared}}$ implies $1 \in {\mathbf{shared}}{^{\scriptscriptstyle\prime}}$, whose purpose was to ensure that ${\mathsf{shared\_update}}_0$ does not compromise what ${\mathsf{shared\_update}}_1$ is trying to accomplish. And indeed, it would be challenging to guarantee this if ${\mathsf{shared\_update}}_0$ would, for instance, randomly remove elements from ${\mathbf{shared}}$. Similar observations can be made about the statement for ${\mathsf{shared\_update}}_1$: $${\models\{{\mathbf{shared}} = {\mathbf{shared}}{^{\scriptscriptstyle\prime}}\wedge {\mathbf{local}}_1 = {\mathbf{local}}{^{\scriptscriptstyle\prime}}_1,\:\top\}\:{\mathsf{shared\_update}}_1 \:\{1 \in {\mathbf{shared}},\:G{^{\scriptscriptstyle\prime}}_1\}}$$ where $G{^{\scriptscriptstyle\prime}}_1$ is the respective instance of $G_1$.
Thus, the remaining obligations $$\begin{array}{l}
{\models {\mathsf{shared\_update}}_0\sqsupseteq_{{\mathsf{r}}_0}{\mathsf{shared\_update}}_0}\\
{\models {\mathsf{shared\_update}}_1\sqsupseteq_{{\mathsf{r}}_1}{\mathsf{shared\_update}}_1}\\
{\models {\mathsf{shared\_update}}_0\sqsupseteq_{{\mathsf{r}}_{{\mathsf{eqv}}}}{\mathsf{shared\_update}}_0}\\
{\models {\mathsf{shared\_update}}_1\sqsupseteq_{{\mathsf{r}}_{{\mathsf{eqv}}}}{\mathsf{shared\_update}}_1}
\end{array}$$ are concerned with program correspondences and amount to routine tasks since the critical sections are ‘well-defined’: ${\mathsf{shared\_update}}_0$ and ${\mathsf{shared\_update}}_1$ neither modify nor depend on ${\mathbf{flag}}_0, {\mathbf{flag}}_1, {\mathbf{turn}}, {\mathbf{turn\_aux}}_0$ or ${\mathbf{turn\_aux}}_1$.
State Relations as Postconditions {#S:gener}
=================================
A closer look at Definition \[def:rgval\] and the final result of the previous section $${\models\{{\mathsf{id}},\:\top\}\:{\mathsf{mutex}} \;{\mathsf{shared\_update}}_0 \; {\mathsf{shared\_update}}_1\:\{ 0 \in {\mathbf{shared}} \wedge 1 \in {\mathbf{shared}},\:\top\}}$$ reveals that the postcondition $0 \in {\mathbf{shared}} \wedge 1 \in {\mathbf{shared}}$ gives us a property of the values of ${\mathbf{shared}}$ in all final states without any accounts to its initial values. So, for instance, ${\mathbf{shared}}$ could initially hold the value $\{2\}$, but $\{0, 1\}$ and $\{0, 1, 3\}$ and the like would perfectly satisfy the postcondition although each of the critical sections merely adds its entry to the shared resource. Would postconditions comprise state relations rather than predicates then ${\mathbf{shared}} \subseteq {\mathbf{shared}}{^{\scriptscriptstyle\prime}}\wedge 0 \in {\mathbf{shared}}{^{\scriptscriptstyle\prime}}\wedge 1 \in {\mathbf{shared}}{^{\scriptscriptstyle\prime}}$ would be a significantly more accurate specification of the ${\mathsf{mutex}} \;{\mathsf{shared\_update}}_0 \; {\mathsf{shared\_update}}_1$ behaviours.
The insight is surely not new, and [@DBLP:conf/mpc/Staden15] in particular addresses this point by means of an elegant approach: having a single state $\sigma$ in a precondition allows us to use its image under a state relation $Q$ in place of a postcondition, ${\rho\models\{R,\:\{\sigma\}\}\:p\:\{{Q\cdot \{\sigma\}},\:G\}}$. In line with that, instead of ${\rho\models\{R,\:P\}\:p\:\{Q,\:G\}}$ one considers a set of triples ${\rho\models\{R,\:\{\sigma\}\}\:p\:\{Q,\:G\}}$ for each $\sigma \in P$:
\[thm:pcondF\] An extended triple ${\rho\models\{R,\:P\}\:p\:\{Q,\:G\}}$ holds iff ${\rho\models\{R,\:\{\sigma\}\}\:p\:\{Q,\:G\}}$ holds for any $\sigma \in P$.
A direct consequence of Definition \[def:rgval\].
The only deficiency in specifying behaviours of some $(\rho, p)$ by ${\rho\models\{R,\:\{\sigma\}\}\:p\:\{{Q\cdot \{\sigma\}},\:G\}}$ for all $\sigma \in P$ is that $P$ remains a state predicate which creates a certain disbalance posing particular challenges for the sequential composition. This motivates the following
\[def:rgval2\] Let $(\rho, p)$ be a program, whereas $R, P, Q, G$ – state relations. Then ${\rho\models_{\scriptscriptstyle2}\{R,\:P\}\:p\:\{Q,\:G\}}$ holds iff ${\rho\models\{R,\:{P\cdot \{\sigma_0\}}\}\:p\:\{{Q\cdot \{\sigma_0\}},\:G\}}$ holds for any state $\sigma_0$.
A remarkable effect of the definition is that each of the program logic rules from Section \[S:prog-log\] can seamlessly be lifted to ${\rho\models_{\scriptscriptstyle2}\{R,\:P\}\:p\:\{Q,\:G\}}$, which largely amounts to consequent replacement of relational images by relational compositions. This principle will be illustrated below by means of the program correspondence rule and the rule for ${\mathsf{basic}}$.
\[thm:pcorr-rule2\] Assume ${\rho, \rho{^{\scriptscriptstyle\prime}}\models p\sqsupseteq_{r}q}$, ${\rho\models_{\scriptscriptstyle2}\{R,\:P\}\:p\:\{Q,\:G\}}$ and
1. ${r \diamond R{^{\scriptscriptstyle\prime}}} \subseteq {R \diamond r}$,
2. $P{^{\scriptscriptstyle\prime}}\subseteq {P \diamond r}$,
3. ${Q \diamond r} \subseteq Q{^{\scriptscriptstyle\prime}}$,
4. ${{{r^{\circ}} \diamond G} \diamond r} \subseteq G{^{\scriptscriptstyle\prime}}$.
Then ${\rho{^{\scriptscriptstyle\prime}}\models_{\scriptscriptstyle2}\{R{^{\scriptscriptstyle\prime}},\:P{^{\scriptscriptstyle\prime}}\}\:q\:\{Q{^{\scriptscriptstyle\prime}},\:G{^{\scriptscriptstyle\prime}}\}}$.
According to Definition \[def:rgval2\] we have to show ${\rho{^{\scriptscriptstyle\prime}}\models\{R{^{\scriptscriptstyle\prime}},\:{P{^{\scriptscriptstyle\prime}}\cdot \{\sigma_0\}}\}\:q\:\{{Q{^{\scriptscriptstyle\prime}}\cdot \{\sigma_0\}},\:G{^{\scriptscriptstyle\prime}}\}}$ for any $\sigma_0$. Then consider the instance of Proposition \[thm:pcorr-rule\] with ${P{^{\scriptscriptstyle\prime}}\cdot \{\sigma_0\}}$ for $P{^{\scriptscriptstyle\prime}}$ and ${Q{^{\scriptscriptstyle\prime}}\cdot \{\sigma_0\}}$ for $Q{^{\scriptscriptstyle\prime}}$ as well as ${P\cdot \{\sigma_0\}}$ for $P$ and ${Q\cdot \{\sigma_0\}}$ for $Q$.
First, ${\rho\models\{R,\:{P\cdot \{\sigma_0\}}\}\:p\:\{{Q\cdot \{\sigma_0\}},\:G\}}$ follows immediately from the assumption ${\rho\models_{\scriptscriptstyle2}\{R,\:P\}\:p\:\{Q,\:G\}}$ by Definition \[def:rgval2\]. Next, the condition ${P{^{\scriptscriptstyle\prime}}\cdot \{\sigma_0\}} \subseteq {r\cdot ({P\cdot \{\sigma_0\}})}$ is the same as ${P{^{\scriptscriptstyle\prime}}\cdot \{\sigma_0\}} \subseteq {({P \diamond r})\cdot \{\sigma_0\}}$ and follows from (2). Similarly, ${r\cdot ({Q\cdot \{\sigma_0\}})} \subseteq {Q{^{\scriptscriptstyle\prime}}\cdot \{\sigma_0\}}$ is the same as ${({Q \diamond r})\cdot \{\sigma_0\}} \subseteq {Q{^{\scriptscriptstyle\prime}}\cdot \{\sigma_0\}}$ and follows from (3).
\[thm:basic-rule2\] Assume
1. ${P \diamond R} \subseteq P$,
2. $P \subseteq \{(\sigma, \sigma{^{\scriptscriptstyle\prime}}) \:|\: (\sigma, f\:\sigma{^{\scriptscriptstyle\prime}}) \in Q \wedge (\sigma{^{\scriptscriptstyle\prime}}, f\:\sigma{^{\scriptscriptstyle\prime}}) \in G \}$.
Then ${\rho\models_{\scriptscriptstyle2}\{R,\:P\}\:{{\mathsf{basic}}}\:f\:\{Q,\:G\}}$.
We have to show ${\rho\models\{R,\:{P\cdot \{\sigma_0\}}\}\:{{\mathsf{basic}}}\:f\:\{{Q\cdot \{\sigma_0\}},\:G\}}$ for any $\sigma_0$. Then consider the instance of Proposition \[thm:basic-rule\] with ${P\cdot \{\sigma_0\}}$ for $P$ and ${Q\cdot \{\sigma_0\}}$ for $Q$, such that the condition ${R\cdot ({P\cdot \{\sigma_0\}})} \subseteq {P\cdot \{\sigma_0\}}$ follows from (1) whereas ${P\cdot \{\sigma_0\}} \subseteq \{\sigma \:|\: f\:\sigma \in {Q\cdot \{\sigma_0\}} \wedge (\sigma, f\:\sigma) \in G \}$ – from (2).
Abstracting code fragments
--------------------------
The language ${\mathtt{\mathcal{L}}_{}}$ does not provide any specific built-in mechanism for abstraction of frequently used code fragments by means of a procedure, and there are reasons for that. First of all, procedure is a high-level concept, and probably the most distinct feature of structured programming. But just to stress it explicitly: any procedure call within a program ultimately amounts to certain series of low-level operations and could therefore be prone to additional interferences. This was already mentioned in the introduction, pointing out that compilations tend to fraction the granularity of computations. In other words, correct behaviour of each procedure call would be subject to a number of side conditions such as interference freedom during the phase of arguments loading which, by the way, would also impose a specific structure on the underlying states, typically in form of stacks, violating the basic idea of state abstraction. Nonetheless, what can be avoided in this framework is that certain reusable fragments, whose input/output properties can be captured by postconditions, need to be verified at each invocation. The following small example illustrates that.
Let ${\mathsf{inc}} {\mathrel{\stackrel{\scriptscriptstyle{\mathit{def}}}{=}}}{\mathbf{x}} := {\mathbf{x}} + 1$ such that $$\label{eq:inc}
{\models_{\scriptscriptstyle2}\{{\mathbf{x}} = {\mathbf{x}}{^{\scriptscriptstyle\prime}},\:{\mathsf{id}}\}\:{\mathsf{inc}}\:\{{\mathbf{x}}{^{\scriptscriptstyle\prime}}= {\mathbf{x}} + 1,\:{\mathbf{x}} \le {\mathbf{x}}{^{\scriptscriptstyle\prime}}\}}$$ clearly holds. So, whenever the variable ${\mathbf{x}}$ needs to be incremented, we can apply ${\mathsf{inc}}$ and use (\[eq:inc\]). For instance, we shall write ${\mathsf{inc}};{\mathsf{inc}}$ rather than ${\mathbf{x}} := {\mathbf{x}} + 1; {\mathbf{x}} := {\mathbf{x}} + 1$ when we want to increment ${\mathbf{x}}$ twice, and consequently appeal to (\[eq:inc\]) twice in order to establish $${\models_{\scriptscriptstyle2}\{{\mathbf{x}} = {\mathbf{x}}{^{\scriptscriptstyle\prime}},\:{\mathsf{id}}\}\:{\mathsf{inc}};{\mathsf{inc}}\:\{{\mathbf{x}}{^{\scriptscriptstyle\prime}}= {\mathbf{x}} + 2,\:{\mathbf{x}} \le {\mathbf{x}}{^{\scriptscriptstyle\prime}}\}}$$ However, this treatment demands additional considerations since for the second invocation of ${\mathsf{inc}}$ we would be prompted to show $$\label{eq:inc2}
{\models_{\scriptscriptstyle2}\{{\mathbf{x}} = {\mathbf{x}}{^{\scriptscriptstyle\prime}},\:{\mathbf{x}}{^{\scriptscriptstyle\prime}}= {\mathbf{x}} + 1\}\:{\mathsf{inc}}\:\{{\mathbf{x}}{^{\scriptscriptstyle\prime}}= {\mathbf{x}} + 2,\:{\mathbf{x}} \le {\mathbf{x}}{^{\scriptscriptstyle\prime}}\}}$$ which is not an immediate instance of (\[eq:inc\]), motivating the following
\[thm:ldiv\] Assume
1. ${\rho\models_{\scriptscriptstyle2}\{R,\:{\mathsf{id}}\}\:p\:\{Q{^{\scriptscriptstyle\prime}},\:G\}}$,
2. ${P \diamond Q{^{\scriptscriptstyle\prime}}} \subseteq Q$.
Then ${\rho\models_{\scriptscriptstyle2}\{R,\:P\}\:p\:\{Q,\:G\}}$.
We have to show ${\rho\models\{R,\:{P\cdot \{\sigma_0\}}\}\:p\:\{{Q\cdot \{\sigma_0\}},\:G\}}$ for any $\sigma_0$. Using Proposition \[thm:pcondF\], this can be done by showing ${\rho\models\{R,\:\{\sigma\}\}\:p\:\{{Q\cdot \{\sigma_0\}},\:G\}}$ for any state $\sigma$ with $(\sigma_0, \sigma) \in P$. To this end, note that ${\rho\models\{R,\:\{\sigma\}\}\:p\:\{{Q{^{\scriptscriptstyle\prime}}\cdot \{\sigma\}},\:G\}}$ follows from (1), and consequently only ${Q{^{\scriptscriptstyle\prime}}\cdot \{\sigma\}} \subseteq {Q\cdot \{\sigma_0\}}$ remains to be established. Assuming $(\sigma, \sigma{^{\scriptscriptstyle\prime}}) \in Q{^{\scriptscriptstyle\prime}}$ with some $\sigma{^{\scriptscriptstyle\prime}}$, we can first derive $(\sigma_0, \sigma{^{\scriptscriptstyle\prime}}) \in {P \diamond Q{^{\scriptscriptstyle\prime}}}$, and then $(\sigma_0, \sigma{^{\scriptscriptstyle\prime}}) \in Q$ by (2).
Thus, with (\[eq:inc\]) as (1) and $
{({\mathbf{x}}{^{\scriptscriptstyle\prime}}= {\mathbf{x}} + 1) \diamond ({\mathbf{x}}{^{\scriptscriptstyle\prime}}= {\mathbf{x}} + 1)} \subseteq {\mathbf{x}}{^{\scriptscriptstyle\prime}}= {\mathbf{x}} + 2
$ as (2), the above statement yields (\[eq:inc2\]).
Case Study: Strengthening Critical Sections’ Specifications {#S:PM2}
===========================================================
In comparison to the development in Section \[S:PM1\], the only difference is that the parameters $P_0, Q_0, P_1, Q_1$, capturing the behaviours of critical sections, can now depend on values of ${\mathbf{shared}}$ in two states. So, for instance, the triple for ${\mathsf{thread\_aux}}_0 \; {\mathit{cs}}_0$ (\[eq:pm1\] in Section \[S:PM1\]) assumes the following form: $$\!\!\!\!\!\!\!\!\!{\models_{\scriptscriptstyle2}\{R_0,\:\neg{\mathbf{turn\_aux}}_0 \wedge P_0\: {\mathbf{shared}}\: {\mathbf{shared}}{^{\scriptscriptstyle\prime}}\}\:{\mathsf{thread\_aux}}_0 \; {\mathit{cs}}_0\:\{Q_0\: {\mathbf{shared}}\: {\mathbf{shared}}{^{\scriptscriptstyle\prime}},\:R_1\}}$$ where $R_0$ now abbreviates the rely $$\begin{array}{l}
{\mathbf{flag}}_0 ={\mathbf{flag}}{^{\scriptscriptstyle\prime}}_0 \wedge {\mathbf{turn\_aux}}_0 = {\mathbf{turn\_aux}}{^{\scriptscriptstyle\prime}}_0 \wedge {\mathbf{local}}_0 = {\mathbf{local}}{^{\scriptscriptstyle\prime}}_0\; \wedge \\
({\mathbf{turn\_aux}}_0 \wedge {\mathbf{cond}}_0 \longrightarrow {\mathbf{shared}} = {\mathbf{shared}}{^{\scriptscriptstyle\prime}})\; \wedge \\
(\neg{\mathbf{turn\_aux}}_1 \wedge {\mathbf{turn\_aux}}{^{\scriptscriptstyle\prime}}_1 \longrightarrow \neg{\mathbf{turn}}{^{\scriptscriptstyle\prime}})\; \wedge \\
({\mathbf{turn\_aux}}_1 \longrightarrow {\mathbf{turn}} = {\mathbf{turn}}{^{\scriptscriptstyle\prime}}) \; \wedge \\
({\mathbf{turn\_aux}}_1 \wedge {\mathbf{flag}}{^{\scriptscriptstyle\prime}}_1 \longrightarrow {\mathbf{flag}}_1)\; \wedge \\
(\forall v.\: P_0\:v\: {\mathbf{shared}} \longrightarrow P_0\:v\: {\mathbf{shared}}{^{\scriptscriptstyle\prime}}) \wedge (\forall v.\:Q_0\:v\: {\mathbf{shared}} \longrightarrow Q_0\:v\: {\mathbf{shared}}{^{\scriptscriptstyle\prime}})
\end{array}$$ which deviates from Section \[S:PM1\] consequently only in the conditions ensuring that ${\mathsf{thread\_aux}}_1$ does not compromise what ${\mathsf{thread\_aux}}_0$ is trying to accomplish. More precisely, the condition $
\forall v.\: P_0\:v\: {\mathbf{shared}} \longrightarrow P_0\:v\: {\mathbf{shared}}{^{\scriptscriptstyle\prime}}$ is a part of the rely $R_0$, such that ${\mathbf{shared}}$ and ${\mathbf{shared}}{^{\scriptscriptstyle\prime}}$ refer here to the respective values in two consecutive states of any environment step, as opposed to, $P_0\:{\mathbf{shared}}\: {\mathbf{shared}}{^{\scriptscriptstyle\prime}}$ in the precondition, where ${\mathbf{shared}}$ refers to some initial state and ${\mathbf{shared}}{^{\scriptscriptstyle\prime}}$ – to a state before entering ${\mathsf{thread\_aux}}_0$. Therefore, if ${\mathsf{thread\_aux}}_1$ makes a step from $\sigma_1$ to $\sigma_2$ then $\forall v.\: P_0\:v\: {\mathbf{shared}} \longrightarrow P_0\:v\: {\mathbf{shared}}{^{\scriptscriptstyle\prime}}$ requires that any value ${\sigma_0{\mathbf{shared}}}$, which is linked to ${\sigma_1{\mathbf{shared}}}$ by $P_0$, will be linked to ${\sigma_2{\mathbf{shared}}}$ by $P_0$ as well. Similar applies to $\forall v.\:Q_0\:v\: {\mathbf{shared}} \longrightarrow Q_0\:v\: {\mathbf{shared}}{^{\scriptscriptstyle\prime}}$ and the postcondition $Q_0\:{\mathbf{shared}}\: {\mathbf{shared}}{^{\scriptscriptstyle\prime}}$.
Since these additional capabilities of $P_0, Q_0, P_1, Q_1$ do not affect anything regarding the actual process of acquiring exclusive access to the shared resource, the new statement for ${\mathsf{mutex}}$ is almost the same as in Proposition \[thm:mutex\], except that, for instance, $P_0 \:{\mathbf{shared}}$ is replaced by $P_0 \:{\mathbf{shared}}\: {\mathbf{shared}}{^{\scriptscriptstyle\prime}}$ in all preconditions, whereas the guarantee $G_0$ changes to $$\begin{array}{l}
{\mathbf{flag}}_0 ={\mathbf{flag}}{^{\scriptscriptstyle\prime}}_0 \wedge {\mathbf{flag}}_1 ={\mathbf{flag}}{^{\scriptscriptstyle\prime}}_1 \wedge {\mathbf{turn}} ={\mathbf{turn}}{^{\scriptscriptstyle\prime}}\wedge \\
{\mathbf{local}}_1 = {\mathbf{local}}{^{\scriptscriptstyle\prime}}_1 \wedge {\mathbf{turn\_aux}}_0 = {\mathbf{turn\_aux}}{^{\scriptscriptstyle\prime}}_0 \wedge {\mathbf{turn\_aux}}_1 = {\mathbf{turn\_aux}}{^{\scriptscriptstyle\prime}}_1 \wedge\\
(\forall v.\: P_1\:v\: {\mathbf{shared}} \longrightarrow P_1\:v\: {\mathbf{shared}}{^{\scriptscriptstyle\prime}}) \wedge (\forall v.\:Q_1\:v\: {\mathbf{shared}} \longrightarrow Q_1\:v\: {\mathbf{shared}}{^{\scriptscriptstyle\prime}})
\end{array}$$ The actual benefit of these modifications becomes apparent when we instantiate ${\mathit{cs}}_0$, ${\mathit{cs}}_1$ and the conditions $P_0, Q_0, P_1, Q_1$. Recall that in Section \[S:PM1\] we had $${\models\{{\mathsf{id}},\:\top\}\:{\mathsf{mutex}} \;{\mathsf{shared\_update}}_0 \; {\mathsf{shared\_update}}_1\:\{ 0 \in {\mathbf{shared}} \wedge 1 \in {\mathbf{shared}},\:\top\}}$$ $P_0, P_1$ were just $\top$, $Q_0$ was instantiated by $0 \in {\mathbf{shared}}$, whereas $Q_1$ – by $1 \in {\mathbf{shared}}$. By contrast, now we can additionally take accounts of the property that ${\mathsf{shared\_update}}_0$ and ${\mathsf{shared\_update}}_1$ do not remove elements from the shared resource. To this end we instantiate both, $P_0$ and $P_1$, by ${\mathbf{shared}} \subseteq {\mathbf{shared}}{^{\scriptscriptstyle\prime}}$, $Q_0$ – by ${\mathbf{shared}} \subseteq {\mathbf{shared}}{^{\scriptscriptstyle\prime}}\wedge 0 \in {\mathbf{shared}}{^{\scriptscriptstyle\prime}}$, whereas $Q_1$ – by ${\mathbf{shared}} \subseteq {\mathbf{shared}}{^{\scriptscriptstyle\prime}}\wedge 1 \in {\mathbf{shared}}{^{\scriptscriptstyle\prime}}$. This yields the following statement $$\label{eq:mutex-ext1}
\begin{array}{l}
{\models_{\scriptscriptstyle2}\{{\mathsf{id}},\:{\mathbf{shared}} \subseteq {\mathbf{shared}}{^{\scriptscriptstyle\prime}}\}\:\\ \quad {\mathsf{mutex}} \;{\mathsf{shared\_update}}_0 \; {\mathsf{shared\_update}}_1 \\ \quad \;\:\{ {\mathbf{shared}} \subseteq {\mathbf{shared}}{^{\scriptscriptstyle\prime}}\wedge 0 \in {\mathbf{shared}}{^{\scriptscriptstyle\prime}}\wedge 1 \in {\mathbf{shared}}{^{\scriptscriptstyle\prime}},\:\top\}}
\end{array}$$ such that for ${\mathsf{shared\_update}}_0$ we need to show $$\label{eq:mutex-ext2}
\begin{array}{l}
{\models_{\scriptscriptstyle2}\{{\mathbf{shared}} = {\mathbf{shared}}{^{\scriptscriptstyle\prime}}\wedge {\mathbf{local}}_0 = {\mathbf{local}}{^{\scriptscriptstyle\prime}}_0,\:{\mathbf{shared}} \subseteq {\mathbf{shared}}{^{\scriptscriptstyle\prime}}\}\:\\ \quad {\mathsf{shared\_update}}_0 \\ \quad\:\{{\mathbf{shared}} \subseteq {\mathbf{shared}}{^{\scriptscriptstyle\prime}}\wedge 0 \in {\mathbf{shared}}{^{\scriptscriptstyle\prime}},\:G{^{\scriptscriptstyle\prime}}_0\}}
\end{array}$$ where $G{^{\scriptscriptstyle\prime}}_0$ is accordingly $$\label{eq:mutex-ext3}
\begin{array}{l}
{\mathbf{flag}}_0 ={\mathbf{flag}}{^{\scriptscriptstyle\prime}}_0 \wedge {\mathbf{flag}}_1 ={\mathbf{flag}}{^{\scriptscriptstyle\prime}}_1 \wedge {\mathbf{turn}} ={\mathbf{turn}}{^{\scriptscriptstyle\prime}}\wedge \\
{\mathbf{local}}_1 = {\mathbf{local}}{^{\scriptscriptstyle\prime}}_1 \wedge {\mathbf{turn\_aux}}_0 = {\mathbf{turn\_aux}}{^{\scriptscriptstyle\prime}}_0 \wedge {\mathbf{turn\_aux}}_1 = {\mathbf{turn\_aux}}{^{\scriptscriptstyle\prime}}_1 \wedge \\
(\forall v. \: v \subseteq {\mathbf{shared}} \longrightarrow v \subseteq {\mathbf{shared}}{^{\scriptscriptstyle\prime}}) \\
(\forall v. \: (v \subseteq {\mathbf{shared}} \wedge 1 \in {\mathbf{shared}}) \longrightarrow (v \subseteq {\mathbf{shared}}{^{\scriptscriptstyle\prime}}\wedge 1 \in {\mathbf{shared}}{^{\scriptscriptstyle\prime}}))
\end{array}$$ Note once more that ${\mathbf{shared}}$ and ${\mathbf{shared}}{^{\scriptscriptstyle\prime}}$ serve different purposes in the rely and in the pre- and postcondition of (\[eq:mutex-ext2\]). In the former, ${\mathbf{shared}}$ and ${\mathbf{shared}}{^{\scriptscriptstyle\prime}}$ refer to the values in two consecutive states of an environment step whereas in, ${\mathbf{shared}} \subseteq {\mathbf{shared}}{^{\scriptscriptstyle\prime}}\wedge 0 \in {\mathbf{shared}}{^{\scriptscriptstyle\prime}}$, ${\mathbf{shared}}$ refers to an input and ${\mathbf{shared}}{^{\scriptscriptstyle\prime}}$ – to an output value of ${\mathit{cs}}_0$. In this setting, by contrast to Section \[S:PM1\] where critical sections did not have any specific preconditions, relying on ${\mathbf{shared}} = {\mathbf{shared}}{^{\scriptscriptstyle\prime}}$ is essential as it provides stability for the precondition ${\mathbf{shared}} \subseteq {\mathbf{shared}}{^{\scriptscriptstyle\prime}}$, namely $
{({\mathbf{shared}} \subseteq {\mathbf{shared}}{^{\scriptscriptstyle\prime}}) \diamond ({\mathbf{shared}} = {\mathbf{shared}}{^{\scriptscriptstyle\prime}})} \subseteq ({\mathbf{shared}} \subseteq {\mathbf{shared}}{^{\scriptscriptstyle\prime}})
$. Moreover, both guarantees $$\forall v. \: v \subseteq {\mathbf{shared}} \longrightarrow v \subseteq {\mathbf{shared}}{^{\scriptscriptstyle\prime}}$$ and $$\forall v. \: (v \subseteq {\mathbf{shared}} \wedge 1 \in {\mathbf{shared}}) \longrightarrow (v \subseteq {\mathbf{shared}}{^{\scriptscriptstyle\prime}}\wedge 1 \in {\mathbf{shared}}{^{\scriptscriptstyle\prime}})$$ follow straight from the property that ${\mathsf{shared\_update}}_0$ only adds its entry to the shared resource and, in particular, cannot remove $1$ from ${\mathbf{shared}}$ whenever ${\mathsf{shared\_update}}_1$ has added it to ${\mathbf{shared}}$ first. All that symmetrically applies to ${\mathsf{shared\_update}}_1$.
Finally, we can strengthen the precondition of (\[eq:mutex-ext1\]) to bring it into the form $$\begin{array}{l}
{\models_{\scriptscriptstyle2}\{{\mathsf{id}},\:{\mathsf{id}}\}\:\\ \quad {\mathsf{mutex}} \;{\mathsf{shared\_update}}_0 \; {\mathsf{shared\_update}}_1 \\ \quad \;\:\{ {\mathbf{shared}} \subseteq {\mathbf{shared}}{^{\scriptscriptstyle\prime}}\wedge 0 \in {\mathbf{shared}}{^{\scriptscriptstyle\prime}}\wedge 1 \in {\mathbf{shared}}{^{\scriptscriptstyle\prime}},\:\top\}}
\end{array}$$ That is, for any initial state, the value of ${\mathbf{shared}}$ in any final state not only contains $0$ and $1$, but is also a superset of the value of ${\mathbf{shared}}$ in this initial state.
Verifying Basic Liveness Properties {#S:live}
===================================
Liveness covers a variety of questions that is usually summarised by ‘will something good eventually happen?’, answering which demands very involved considerations in many instances. Keeping it slightly down, we recast the above question to the current context by restricting ‘something good’ to *certain computations of a given program reach a state satisfying some given predicate*. Making this point clear, let ${\mathsf{p}}_1$ be the program $$\begin{array}{l}
{{\mathsf{while}}\:{\mathbf{a}} > 0\;{\mathsf{do}}\;{{\mathsf{skip}}}\;{\mathsf{od}}};\\
{\mathbf{b}} := {\mathsf{True}}
\end{array}$$ whereas ${\mathsf{p}}_2$ shall comprise the assignment ${\mathbf{a}} := 0$ only. In this setting, one possible liveness question is if *certain computations* of ${\mathsf{p}}_1 \parallel {\mathsf{p}}_2$, starting in a state $\sigma_0 \in ({\mathbf{a}} = 1 \wedge \neg{\mathbf{b}})$, will eventually reach a state satisfying ${\mathbf{b}}$. Intuitively, *certain computations* should address the largest set for which the answer to the question is *yes*. More precisely, consider the following potential candidates:
1. ${\lsemantic{\mathsf{p}}_1 \parallel {\mathsf{p}}_2\rsemantic}$, all finite potential computations of ${\mathsf{p}}_1 \parallel {\mathsf{p}}_2$;
2. ${\lsemantic{\mathsf{p}}_1 \parallel {\mathsf{p}}_2\rsemantic^\omega} \cap {\mathcal{E}\!\mathit{nv}\:\bot}$, all infinite potential computations of ${\mathsf{p}}_1 \parallel {\mathsf{p}}_2$ with no environment steps;
3. ${\lsemantic{\mathsf{p}}_1 \parallel {\mathsf{p}}_2\rsemantic^\omega} \cap {\mathcal{E}\!\mathit{nv}\:{\mathsf{id}}}$, all infinite potential computations of ${\mathsf{p}}_1 \parallel {\mathsf{p}}_2$ in presence of a ‘stuttering’ environment.
With (1) the answer is *no*: ${\lsemantic{\mathsf{p}}_1 \parallel {\mathsf{p}}_2\rsemantic}$ includes all prefixes, such as the initial configuration $({\mathsf{p}}_1 \parallel {\mathsf{p}}_2, \sigma_0)$ as the simplest instance. With (2) the answer is also *no*: we basically take into consideration all infinite actual computations of ${\mathsf{p}}_1 \parallel {\mathsf{p}}_2$, thereby ruling out those that reach a state satisfying ${\mathbf{b}}$ and terminate. The option (3) not only covers all relevant computations reaching ${\mathbf{b}}$, but also those with infinitely long runs of the ‘stuttering’ environment having $\neg{\mathbf{b}}$ as an invariant, as well as with infinitely long runs of ${{\mathsf{while}}\:{\mathbf{a}} > 0\;{\mathsf{do}}\;{{\mathsf{skip}}}\;{\mathsf{od}}}$. Excluding these would result in an adequate set of computations of ${\mathsf{p}}_1 \parallel {\mathsf{p}}_2$ for which the answer to the above question would be *yes*.
These considerations ineluctably lead to the conclusion that some form of fairness among threads and the environment must be assumed in order to achieve rational reasoning about liveness properties, which is by no means a new insight: this subject is particularly well-explored in the contexts of process algebras and temporal logics ( [@VANGLABBEEK2019100480; @10.1007/3-540-61474-5_84; @10.5555/903616] to name only a few). Therefore, a notion of fairness for this framework will be elaborated next.
Fair computations {#Sb:faircomp}
-----------------
The starting point is that any term in ${\mathtt{\mathcal{L}}_{\alpha}}$ gives rise to a finite set of *redexes*, subprograms in different parallel components that can perform the next step. In the example above, ${\mathsf{p}}_1 \parallel {\mathsf{p}}_2$ has two such redexes: ${\mathsf{p}}_1$ and ${\mathsf{p}}_2$. We can identify redexes by their positions ( [@baader_nipkow_1998] for positions of terms in general) which comprise strings over the alphabet of natural numbers. That is, the function ${\mathit{{\mathcal P}\!os}}$ maps any element of ${\mathtt{\mathcal{L}}_{\alpha}}$ to a set of strings over natural numbers, as defined by the following recursive equations using pattern matching: $$\begin{array}{l c l}
{\mathit{{\mathcal P}\!os}}\: {{\mathsf{skip}}}& = & \emptyset \\
{\mathit{{\mathcal P}\!os}}({{\mathsf{skip}}};q) & = & \{ 0 \} \\
{\mathit{{\mathcal P}\!os}}(p ;q) & = & \{0x \:|\: x \in {\mathit{{\mathcal P}\!os}}\:p \} \\
{\mathit{{\mathcal P}\!os}}({\parallel\!\!{{\mathsf{skip}}}, \ldots, {{\mathsf{skip}}}}) & = & \{ 0 \} \\
{\mathit{{\mathcal P}\!os}}({\parallel\!\!p_1, \ldots, p_m}) & = & \bigcup_{i=1}^{m}\{ix \:|\: x \in {\mathit{{\mathcal P}\!os}}\:p_i \} \\
{\mathit{{\mathcal P}\!os}}\: p & = & \{ 0 \} \\
\end{array}$$ For instance, in the above example ${\mathit{{\mathcal P}\!os}}({\mathsf{p}}_1 \parallel {\mathsf{p}}_2)$ yields $\{100, 20\}$. Note that ${\mathit{{\mathcal P}\!os}}\:p$ contains more than one element only if the parallel operator, applied to at least two arguments different from ${{\mathsf{skip}}}$, occurs in $p$. In particular, any locally sequential $p$ has at most one $x \in {\mathit{{\mathcal P}\!os}}\:p$.
Functions, providing access to the subprogram associated with a position, satisfy the equations: $$\begin{array}{l c l}
{p |_{0}} & = & p \\
{(p ;q) |_{0x}} & = & {p |_{x}} \\
{({\parallel\!\!p_1, \ldots, p_m}) |_{ix}} & = & {p_i |_{x}}
\end{array}$$ such that ${p |_{x}}$ is well-defined for any $x \in {\mathit{{\mathcal P}\!os}}\:p$. Moreover, at any position $x \in {\mathit{{\mathcal P}\!os}}\:p$ we can substitute some $p{^{\scriptscriptstyle\prime}}$, denoted by ${p[p{^{\scriptscriptstyle\prime}}]_{x}}$: $$\begin{array}{l c l}
{p[p{^{\scriptscriptstyle\prime}}]_{0}} & = & p{^{\scriptscriptstyle\prime}}\\
{(p ;q)[p{^{\scriptscriptstyle\prime}}]_{0x}} & = & ({p[p{^{\scriptscriptstyle\prime}}]_{x}});q \\
{({\parallel\!\!p_1, \ldots, p_m})[p{^{\scriptscriptstyle\prime}}]_{ix}} & = & {\parallel\!p_1, \ldots, {p_i[p{^{\scriptscriptstyle\prime}}]_{x}}, \ldots, p_m}
\end{array}$$ such that the following proposition puts all that together.
\[thm:rpos-pstep\] ${\rho \vdash (p, \sigma) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(q, \sigma{^{\scriptscriptstyle\prime}})}$ holds iff there exist a position $x \in {\mathit{{\mathcal P}\!os}}\:p$ and $p{^{\scriptscriptstyle\prime}}$ such that ${\rho \vdash ({p |_{x}}, \sigma) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(p{^{\scriptscriptstyle\prime}}, \sigma{^{\scriptscriptstyle\prime}})}$ and $q = {p[p{^{\scriptscriptstyle\prime}}]_{x}}$ hold.
The if-direction can be shown by induction on $p$, whereas the opposite – by structural induction on ${\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}$.
It is worth noting that $x \in {\mathit{{\mathcal P}\!os}}\:p$ does not imply that for any state $\sigma$ there exist some $p{^{\scriptscriptstyle\prime}}$ and $\sigma{^{\scriptscriptstyle\prime}}$ such that ${\rho \vdash ({p |_{x}}, \sigma) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(p{^{\scriptscriptstyle\prime}}, \sigma{^{\scriptscriptstyle\prime}})}$ holds, which is due to certain ${\mathsf{await}}$-statements. Firstly, $({{\mathsf{await}}\:C\;{\mathsf{do}}\;p\;{\mathsf{od}}}, \sigma)$ would be blocked when $\sigma \notin C$. Secondly, $p$ also does not need to terminate starting from $\sigma$ otherwise. In other words, ${{\mathsf{await}}\:C\;{\mathsf{do}}\;p\;{\mathsf{od}}}$ is in general not always available to perform a step, as opposed to any other redex. In order to simplify the matter, in what follows the focus will be on positions that do not require some specific states to perform a step.
A position $x \in {\mathit{{\mathcal P}\!os}}\:p$ is *always available* if ${p |_{x}} = {{\mathsf{await}}\:C\;{\mathsf{do}}\;p{^{\scriptscriptstyle\prime}}\;{\mathsf{od}}}$ implies $C = \top$ and for any $\sigma$ there exists $\sigma{^{\scriptscriptstyle\prime}}$ such that ${\rho \vdash (p{^{\scriptscriptstyle\prime}}, \sigma) {\stackrel{\scriptscriptstyle\star}{\mathop{\rightarrow}}_{\scriptscriptstyle{\mathcal{P}}}}({{\mathsf{skip}}}, \sigma{^{\scriptscriptstyle\prime}})}$ holds.
That is, if $x \in {\mathit{{\mathcal P}\!os}}\:p$ is always available then for any $\sigma$ there exist some $p{^{\scriptscriptstyle\prime}}$ and $\sigma{^{\scriptscriptstyle\prime}}$ such that ${\rho \vdash ({p |_{x}}, \sigma) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(p{^{\scriptscriptstyle\prime}}, \sigma{^{\scriptscriptstyle\prime}})}$ as well as ${\rho \vdash (p, \sigma) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}({p[p{^{\scriptscriptstyle\prime}}]_{x}}, \sigma{^{\scriptscriptstyle\prime}})}$ hold.
Regarding fairness we should expect that whenever an always available position is present in a configuration then any fair computation starting from it eventually performs a program step at this position, which is captured formally as follows.
A computation ${\mathit{sq}}\in {\lsemanticp\rsemantic^\omega_{\rho}}$ is called *fair* if for any $i \in\mathbb{N}$ and any always available $x \in {\mathit{{\mathcal P}\!os}}({{\mathsf{\mathcal{P}}}(sq_i)})$ there exists $j \ge i$ with $x \in {\mathit{{\mathcal P}\!os}}({{\mathsf{\mathcal{P}}}(sq_j)})$ and ${{{\mathsf{\mathcal{P}}}(sq_j)} |_{x}} = {{{\mathsf{\mathcal{P}}}(sq_i)} |_{x}}$, and there exists $p{^{\scriptscriptstyle\prime}}$ such that ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_{j+1})} = {{{\mathsf{\mathcal{P}}}({\mathit{sq}}_j)}[p{^{\scriptscriptstyle\prime}}]_{x}}$ with $${\rho \vdash ({{{\mathsf{\mathcal{P}}}(sq_j)} |_{x}}, {{\mathsf{\mathcal{S}}}(sq_j)}) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(p{^{\scriptscriptstyle\prime}}, {{\mathsf{\mathcal{S}}}({\mathit{sq}}_{j+1})})}$$ holds. This will be denoted by ${\mathit{sq}}\in {\lsemanticp\rsemantic^{\scriptscriptstyle{\mathit{fair}}}_{\rho}}$.
Note that the notion of fairness leaves open whether some ${{\mathsf{await}}\:C\;{\mathsf{do}}\;p\;{\mathsf{od}}}$ where $C \subset \top$ will ever be executed, for otherwise we would simply force fair computations to reach a state satisfying $C$, which could easily and inconspicuously lead to vacuous liveness properties even when $\bot \subset C$.
As already mentioned, if $p$ is locally sequential then there exists at most one $x \in {\mathit{{\mathcal P}\!os}}\:p$ and, hence, at most one always available position. Therefore, assuming a program $(\rho, p)$ to be sequential, ${\mathit{sq}}\in {\lsemanticp\rsemantic^{\scriptscriptstyle{\mathit{fair}}}_{\rho}}$ becomes equivalent to the following condition: ${\mathit{sq}}\in {\lsemanticp\rsemantic^\omega_{\rho}}$ and if $x \in {\mathit{{\mathcal P}\!os}}({{\mathsf{\mathcal{P}}}(sq_i)})$ is always available then there exists $j \ge i$ with a program step ${\rho \vdash {\mathit{sq}}_j {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}{\mathit{sq}}_{j+1}}$ for any $i$ and $x$. In other words, if ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_i)}$ can perform a program step that does not require some particular state condition then there cannot be only environment steps from ${\mathit{sq}}_i$ on.
Lastly, we can sketch how fairness enables the liveness property in the above example. Any computation in ${\lsemantic{\mathsf{p}}_1 \parallel {\mathsf{p}}_2\rsemantic^{\scriptscriptstyle{\mathit{fair}}}} \cap {\mathcal{E}\!\mathit{nv}\:{\mathsf{id}}}$ starting with $\sigma_0$ will eventually perform ${\mathbf{a}} := 0$ since the position $20$ is available in ${\mathsf{p}}_1 \parallel {\mathsf{p}}_2$, leaving $100$ as the only available position. Thus, the states of all subsequent configurations will satisfy ${\mathbf{a}} = 0$. Appealing to the fairness once more, the environment, prior to keep on ‘stuttering’ *ad infinitum*, will let the first component exit ${{\mathsf{while}}\:{\mathbf{a}} > 0\;{\mathsf{do}}\;{{\mathsf{skip}}}\;{\mathsf{od}}}$, assign ${\mathsf{True}}$ to ${\mathbf{b}}$ and terminate.
A refutational approach
-----------------------
To show that any computation in ${\lsemanticp\rsemantic^{\scriptscriptstyle{\mathit{fair}}}} \cap {\mathcal{I}\!\mathit{n}\:P} \cap {\mathcal{E}\!\mathit{nv}\:{\mathsf{id}}}$ reaches a state satisfying some given condition $Q$, one can attempt to refute the opposite, namely that there exists a computation ${\mathit{sq}}\in {\lsemanticp\rsemantic^{\scriptscriptstyle{\mathit{fair}}}} \cap {\mathcal{I}\!\mathit{n}\:P} \cap {\mathcal{E}\!\mathit{nv}\:{\mathsf{id}}}$ having $\neg Q$ as an invariant.
Let ${{\mathit{sq}}\Vdash P {\smalltriangleright}p {\smalltriangleright}Q}$ denote the condition ${\mathit{sq}}\in {\lsemanticp\rsemantic^{\scriptscriptstyle{\mathit{fair}}}} \cap {\mathcal{I}\!\mathit{n}\:P}$ with ${{\mathsf{\mathcal{S}}}({\mathit{sq}}_n)} \notin Q$ for all $n \in \mathbb{N}$. Further, let $$\begin{array}{l c l}
\!\!\!\!\!\!\!\!\!{{\mathit{sq}}\Vdash_{\scriptscriptstyle{\mathcal T}} P {\smalltriangleright}p {\smalltriangleright}Q {\smalltriangleright}n} & {\mathrel{\stackrel{\scriptscriptstyle{\mathit{def}}}{=}}}& {{\mathit{sq}}\Vdash P {\smalltriangleright}p {\smalltriangleright}Q} \wedge {{\mathsf{\mathcal{P}}}({\mathit{sq}}_n)} = {{\mathsf{skip}}}\wedge (\forall m < n.\: {{\mathsf{\mathcal{P}}}({\mathit{sq}}_m)} \neq {{\mathsf{skip}}}) \\
\!\!\!\!\!\!\!\!\!{{\mathit{sq}}\Vdash_{\scriptscriptstyle{\mathcal N}} P {\smalltriangleright}p {\smalltriangleright}Q} & {\mathrel{\stackrel{\scriptscriptstyle{\mathit{def}}}{=}}}& {{\mathit{sq}}\Vdash P {\smalltriangleright}p {\smalltriangleright}Q} \wedge (\forall n.\: {{\mathsf{\mathcal{P}}}({\mathit{sq}}_n)} \neq {{\mathsf{skip}}}) \\
\end{array}$$ That is, ${{\mathit{sq}}\Vdash P {\smalltriangleright}p {\smalltriangleright}Q}$ can be split in two exclusive cases: the condition ${{\mathit{sq}}\Vdash_{\scriptscriptstyle{\mathcal T}} P {\smalltriangleright}p {\smalltriangleright}Q {\smalltriangleright}n}$ additionally captures that ${\mathit{sq}}$ reaches the first ${{\mathsf{skip}}}$-configuration by $n$ transitions, whereas ${{\mathit{sq}}\Vdash_{\scriptscriptstyle{\mathcal N}} P {\smalltriangleright}p {\smalltriangleright}Q}$ – that ${\mathit{sq}}$ does not reach a ${{\mathsf{skip}}}$-configuration at all.
Furthermore, note that ${{\mathit{sq}}\Vdash P {\smalltriangleright}p {\smalltriangleright}\bot}$ is the same as ${\mathit{sq}}\in {\lsemanticp\rsemantic^{\scriptscriptstyle{\mathit{fair}}}} \cap {\mathcal{I}\!\mathit{n}\:P}$, whereas refuting ${{\mathit{sq}}\Vdash_{\scriptscriptstyle{\mathcal N}} P {\smalltriangleright}p {\smalltriangleright}\bot}$ means that any computation in ${\lsemanticp\rsemantic^{\scriptscriptstyle{\mathit{fair}}}} \cap {\mathcal{I}\!\mathit{n}\:P}$ eventually reaches a ${{\mathsf{skip}}}$-configuration.
In order to systematically draw conclusions about a computation ${\mathit{sq}}$ from statements like ${{\mathit{sq}}\Vdash_{\scriptscriptstyle{\mathcal N}} P {\smalltriangleright}p {\smalltriangleright}\bot}$, a set of rules for ${{\mathit{sq}}\Vdash_{\scriptscriptstyle{\mathcal N}} P {\smalltriangleright}p {\smalltriangleright}Q}$ and ${{\mathit{sq}}\Vdash_{\scriptscriptstyle{\mathcal T}} P {\smalltriangleright}p {\smalltriangleright}Q {\smalltriangleright}n}$, where $p$ is a specific language constructor, will be presented in the remainder of this section. To start with the simplest, the statement ${{\mathit{sq}}\Vdash_{\scriptscriptstyle{\mathcal N}} P {\smalltriangleright}{{\mathsf{skip}}}{\smalltriangleright}Q}$ is already a contradiction for any ${\mathit{sq}}, P$ and $Q$, whereas from ${{\mathit{sq}}\Vdash_{\scriptscriptstyle{\mathcal T}} P {\smalltriangleright}{{\mathsf{skip}}}{\smalltriangleright}Q {\smalltriangleright}n}$ we can infer $n = 0$, thus gaining an additional piece of information.
Similarly to ${{\mathsf{skip}}}$, ${{\mathit{sq}}\Vdash_{\scriptscriptstyle{\mathcal N}} P {\smalltriangleright}{{\mathsf{basic}}}\:f {\smalltriangleright}Q}$ immediately leads to a contradiction. However, fairness is required to this end, for otherwise ${\mathit{sq}}$ could also comprise an infinite sequence of environment steps. By contrast, without any fairness assumptions we can infer from ${{\mathit{sq}}\Vdash_{\scriptscriptstyle{\mathcal T}} P {\smalltriangleright}{{\mathsf{basic}}}\:f {\smalltriangleright}Q {\smalltriangleright}n}$ that $n > 0$, ${\mathit{sq}}_{n-1} = ({{\mathsf{basic}}}\:f, \sigma)$ and ${\mathit{sq}}_n = ({{\mathsf{skip}}}, f\:\sigma)$ must hold for some $\sigma$. Since ${{\mathit{sq}}\Vdash_{\scriptscriptstyle{\mathcal T}} P {\smalltriangleright}{{\mathsf{basic}}}\:f {\smalltriangleright}Q {\smalltriangleright}n}$ subsumes ${{\mathsf{\mathcal{S}}}({\mathit{sq}}_n)} \notin Q$, this might give a direct argument for a contradiction.
Next two propositions handle the cases with conditional and ${\mathsf{await}}$-statements.
\[thm:await-live\] The following implications hold:
1. if ${{\mathit{sq}}\Vdash_{\scriptscriptstyle{\mathcal N}} P {\smalltriangleright}\langle p \rangle {\smalltriangleright}Q}$ then there exists a state $\sigma$ such that $p$ does not terminate starting from this;
2. if ${{\mathit{sq}}\Vdash_{\scriptscriptstyle{\mathcal T}} P {\smalltriangleright}{{\mathsf{await}}\:C\;{\mathsf{do}}\;p\;{\mathsf{od}}} {\smalltriangleright}Q {\smalltriangleright}n}$ then $n > 0$, ${{\mathsf{\mathcal{S}}}({\mathit{sq}}_{n-1})} \in C$ and $${\rho \vdash (p, {{\mathsf{\mathcal{S}}}({\mathit{sq}}_{n-1})}) {\stackrel{\scriptscriptstyle\star}{\mathop{\rightarrow}}_{\scriptscriptstyle{\mathcal{P}}}}({{\mathsf{skip}}}, {{\mathsf{\mathcal{S}}}({\mathit{sq}}_{n})})}.$$
In (1) we appeal to fairness as follows. Assume that for all states $\sigma$ there exists $\sigma{^{\scriptscriptstyle\prime}}$ such that ${\rho \vdash (p, \sigma) {\stackrel{\scriptscriptstyle\star}{\mathop{\rightarrow}}_{\scriptscriptstyle{\mathcal{P}}}}({{\mathsf{skip}}}, \sigma{^{\scriptscriptstyle\prime}})}$. Then the position $0$ in ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_0)}$, $\langle p \rangle$, becomes always available, and since ${\mathit{sq}}$ is fair we obtain some $j$ such that ${\rho \vdash (\langle p \rangle, {{\mathsf{\mathcal{S}}}(sq_j)}) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(p{^{\scriptscriptstyle\prime}}, {{\mathsf{\mathcal{S}}}({\mathit{sq}}_{j+1})})}$ holds. This implies $p{^{\scriptscriptstyle\prime}}= {{\mathsf{skip}}}$, contradicting the assumption that ${\mathit{sq}}$ does not terminate.
As with ${\mathsf{basic}}$, fairness is not required for (2) such that we can handle any possible ${\mathsf{await}}$-statement and not only atomic sections: from ${{\mathit{sq}}\Vdash_{\scriptscriptstyle{\mathcal T}} P {\smalltriangleright}{{\mathsf{await}}\:C\;{\mathsf{do}}\;p\;{\mathsf{od}}} {\smalltriangleright}Q {\smalltriangleright}n}$ we can infer that ${\mathit{sq}}_n$ is the first ${{\mathsf{skip}}}$-configuration, such that $n > 0$ and $${\rho \vdash ({{\mathsf{await}}\:C\;{\mathsf{do}}\;p\;{\mathsf{od}}}, {{\mathsf{\mathcal{S}}}({\mathit{sq}}_{n-1})}) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}({{\mathsf{skip}}}, {{\mathsf{\mathcal{S}}}({\mathit{sq}}_{n})})}$$ must hold, which implies ${{\mathsf{\mathcal{S}}}({\mathit{sq}}_{n-1})} \in C$ and ${\rho \vdash (p, {{\mathsf{\mathcal{S}}}({\mathit{sq}}_{n-1})}) {\stackrel{\scriptscriptstyle\star}{\mathop{\rightarrow}}_{\scriptscriptstyle{\mathcal{P}}}}({{\mathsf{skip}}}, {{\mathsf{\mathcal{S}}}({\mathit{sq}}_{n})})}$.
\[thm:cond-live\] The following implications hold:
1. if ${{\mathit{sq}}\Vdash_{\scriptscriptstyle{\mathcal N}} P {\smalltriangleright}{{\mathsf{if}}\:C\;{\mathsf{then}}\;p\;{\mathsf{else}}\;q\;{\mathsf{fi}}} {\smalltriangleright}Q}$ then there exists some $n$ with $n > 0$ such that either ${{^{\scriptscriptstyle n}|{\mathit{sq}}} \Vdash_{\scriptscriptstyle{\mathcal N}} C {\smalltriangleright}p {\smalltriangleright}Q}$ or ${{^{\scriptscriptstyle n}|{\mathit{sq}}} \Vdash_{\scriptscriptstyle{\mathcal N}} \neg C {\smalltriangleright}q {\smalltriangleright}Q}$;
2. if ${{\mathit{sq}}\Vdash_{\scriptscriptstyle{\mathcal T}} P {\smalltriangleright}{{\mathsf{if}}\:C\;{\mathsf{then}}\;p\;{\mathsf{else}}\;q\;{\mathsf{fi}}} {\smalltriangleright}Q {\smalltriangleright}n}$ then there exists some $m$ with $0 < m \le n$ such that either ${{^{\scriptscriptstyle m}|{\mathit{sq}}} \Vdash_{\scriptscriptstyle{\mathcal T}} C {\smalltriangleright}p {\smalltriangleright}Q {\smalltriangleright}n-m}$ or ${{^{\scriptscriptstyle m}|{\mathit{sq}}} \Vdash_{\scriptscriptstyle{\mathcal T}} \neg C {\smalltriangleright}q {\smalltriangleright}Q {\smalltriangleright}n-m}$.
In (1) the fairness assumption ensures that a step at the position $0$ in ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_0)}$ will eventually be performed. That is, there must be some $n > 0$ with ${\rho \vdash {\mathit{sq}}_{n-1} {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}{\mathit{sq}}_n}$ and ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_{n-1})} = {{\mathsf{if}}\:C\;{\mathsf{then}}\;p\;{\mathsf{else}}\;q\;{\mathsf{fi}}}$. If ${{\mathsf{\mathcal{S}}}(sq_{n-1})} \in C$ then ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_n)} = p$ and ${{\mathsf{\mathcal{S}}}(sq_{n})} = {{\mathsf{\mathcal{S}}}(sq_{n-1})} $ such that the suffix ${^{\scriptscriptstyle n}|{\mathit{sq}}}$
1. starts in a state satisfying $C$,
2. does not terminate and does not reach a state satisfying $Q$,
3. last but not least, is fair as a suffix of a fair computation, ${^{\scriptscriptstyle n}|{\mathit{sq}}} \in {\lsemanticp\rsemantic^{\scriptscriptstyle{\mathit{fair}}}}$.
The case with ${{\mathsf{\mathcal{S}}}(sq_{n-1})} \notin C$ is entirely symmetric.
In (2) ${{\mathit{sq}}\Vdash_{\scriptscriptstyle{\mathcal T}} P {\smalltriangleright}{{\mathsf{if}}\:C\;{\mathsf{then}}\;p\;{\mathsf{else}}\;q\;{\mathsf{fi}}} {\smalltriangleright}Q {\smalltriangleright}n}$ implies that there must be some $i < n$ with ${\rho \vdash {\mathit{sq}}_i {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}{\mathit{sq}}_{i+1}}$ and ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_i)} = {{\mathsf{if}}\:C\;{\mathsf{then}}\;p\;{\mathsf{else}}\;q\;{\mathsf{fi}}}$. Thus, the suffix ${^{\scriptscriptstyle i+1}|{\mathit{sq}}}$ starts either with ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_{i+1})} = p$ or with ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_{i+1})} = q$ and reaches the first ${{\mathsf{skip}}}$-configuration in $n - i - 1$ steps.
Let ${{\mathit{sq}}\sim {\mathit{sq}}{^{\scriptscriptstyle\prime}}}$ from now on denote that computations ${\mathit{sq}}$ and ${\mathit{sq}}{^{\scriptscriptstyle\prime}}$ run through the same states, ${{\mathsf{\mathcal{S}}}({\mathit{sq}}_i)} = {{\mathsf{\mathcal{S}}}({\mathit{sq}}{^{\scriptscriptstyle\prime}}_i)}$ for any $i\in\mathbb{N}$.
\[thm:seq-live\] The following implications hold:
1. if ${{\mathit{sq}}\Vdash_{\scriptscriptstyle{\mathcal N}} P {\smalltriangleright}p;q {\smalltriangleright}Q}$ then there exists ${\mathit{sq}}{^{\scriptscriptstyle\prime}}$ with ${{\mathit{sq}}\sim {\mathit{sq}}{^{\scriptscriptstyle\prime}}}$ and either ${{\mathit{sq}}{^{\scriptscriptstyle\prime}}\Vdash_{\scriptscriptstyle{\mathcal N}} P {\smalltriangleright}p {\smalltriangleright}Q}$ or there exist $m$ and $n$ such that $m < n$, ${{\mathit{sq}}{^{\scriptscriptstyle\prime}}\Vdash_{\scriptscriptstyle{\mathcal T}} P {\smalltriangleright}p {\smalltriangleright}Q {\smalltriangleright}m}$, and ${{^{\scriptscriptstyle n}|{\mathit{sq}}} \Vdash_{\scriptscriptstyle{\mathcal N}} \top {\smalltriangleright}q {\smalltriangleright}Q}$;
2. if ${{\mathit{sq}}\Vdash_{\scriptscriptstyle{\mathcal T}} P {\smalltriangleright}p;q {\smalltriangleright}Q {\smalltriangleright}n}$ then there exist ${\mathit{sq}}{^{\scriptscriptstyle\prime}}$, $k, l$ such that $k < l \le n$, ${{\mathit{sq}}\sim {\mathit{sq}}{^{\scriptscriptstyle\prime}}}$, ${{\mathit{sq}}{^{\scriptscriptstyle\prime}}\Vdash_{\scriptscriptstyle{\mathcal T}} P {\smalltriangleright}p {\smalltriangleright}Q {\smalltriangleright}k}$ and ${{^{\scriptscriptstyle l}|{\mathit{sq}}} \Vdash_{\scriptscriptstyle{\mathcal T}} \top {\smalltriangleright}q {\smalltriangleright}Q {\smalltriangleright}n-l}$.
In (1) we have a non-terminating ${\mathit{sq}}\in {\lsemanticp;q\rsemantic^{\scriptscriptstyle{\mathit{fair}}}_{\rho}}$. First, in case ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_i)} \neq {{\mathsf{skip}}};q$ holds for all $i$, we can conclude that for any $i$ there is some $u_i$ such that ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_i)} = u_i;q$ holds, and define ${\mathit{sq}}{^{\scriptscriptstyle\prime}}_i {\mathrel{\stackrel{\scriptscriptstyle{\mathit{def}}}{=}}}(u_i, {{\mathsf{\mathcal{S}}}({\mathit{sq}}_i)})$. Thus, we have ${{\mathit{sq}}\sim {\mathit{sq}}{^{\scriptscriptstyle\prime}}}$ by construction. Furthermore, ${\mathit{sq}}{^{\scriptscriptstyle\prime}}$ is fair: let $x \in {\mathit{{\mathcal P}\!os}}\:u_i$ be an always available position with some $i$ to this end. Then $0x$ is an always available position of $u_i;q$. Hence, by fairness of ${\mathit{sq}}$ we obtain some $j\ge i$ such that ${\rho \vdash ({{{\mathsf{\mathcal{P}}}(sq_j)} |_{0x}}, {{\mathsf{\mathcal{S}}}(sq_j)}) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(p{^{\scriptscriptstyle\prime}}, {{\mathsf{\mathcal{S}}}({\mathit{sq}}_{j+1})})}$ holds with ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_{j+1})} = {{{\mathsf{\mathcal{P}}}({\mathit{sq}}_j)}[p{^{\scriptscriptstyle\prime}}]_{0x}}$. This is the same as ${\rho \vdash ({{{\mathsf{\mathcal{P}}}({\mathit{sq}}{^{\scriptscriptstyle\prime}}_j)} |_{x}}, {{\mathsf{\mathcal{S}}}(sq_j)}) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(p{^{\scriptscriptstyle\prime}}, {{\mathsf{\mathcal{S}}}({\mathit{sq}}_{j+1})})}$ with ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_{j+1})} = {u_j[p{^{\scriptscriptstyle\prime}}]_{x}};q$, ${{\mathsf{\mathcal{P}}}({\mathit{sq}}{^{\scriptscriptstyle\prime}}_{j+1})} = {{{\mathsf{\mathcal{P}}}({\mathit{sq}}{^{\scriptscriptstyle\prime}}_j)}[p{^{\scriptscriptstyle\prime}}]_{x}}$, establishing ${{\mathit{sq}}{^{\scriptscriptstyle\prime}}\Vdash_{\scriptscriptstyle{\mathcal N}} P {\smalltriangleright}p {\smalltriangleright}Q}$.
Second, suppose there is some $k$ such that ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_k)} = {{\mathsf{skip}}};q$ and let $m$ be the least such index. Then we can define ${\mathit{sq}}{^{\scriptscriptstyle\prime}}_i {\mathrel{\stackrel{\scriptscriptstyle{\mathit{def}}}{=}}}(u_i, {{\mathsf{\mathcal{S}}}({\mathit{sq}}_i)})$ for all $i\le m$, and ${\mathit{sq}}{^{\scriptscriptstyle\prime}}_i {\mathrel{\stackrel{\scriptscriptstyle{\mathit{def}}}{=}}}({{\mathsf{skip}}}, {{\mathsf{\mathcal{S}}}({\mathit{sq}}_i)})$ with ${\mathit{sq}}{^{\scriptscriptstyle\prime}}_{i-1} {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{E}}}\:}{\mathit{sq}}{^{\scriptscriptstyle\prime}}_{i}$ for all $i> m$, such that ${{\mathit{sq}}\sim {\mathit{sq}}{^{\scriptscriptstyle\prime}}}$ and ${{\mathit{sq}}{^{\scriptscriptstyle\prime}}\Vdash_{\scriptscriptstyle{\mathcal T}} P {\smalltriangleright}p {\smalltriangleright}Q {\smalltriangleright}m}$ hold by construction. Further, fairness of ${\mathit{sq}}$ provides that starting from ${\mathit{sq}}_{m}$ we eventually reach some $n>m$ with ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_n)} = q$. Note that the state ${{\mathsf{\mathcal{S}}}({\mathit{sq}}_n)}$ does not need to possess any property other than ${{\mathsf{\mathcal{S}}}({\mathit{sq}}_n)} \notin Q$, such that we can conclude ${{^{\scriptscriptstyle n}|{\mathit{sq}}} \Vdash_{\scriptscriptstyle{\mathcal N}} \top {\smalltriangleright}q {\smalltriangleright}Q}$.
In (2) we assume that ${\mathit{sq}}\in {\lsemanticp;q\rsemantic^{\scriptscriptstyle{\mathit{fair}}}_{\rho}}$ reaches ${{\mathsf{skip}}}$ in $n$ steps. Hence, there also must be the least index $k$ with ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_{k})} = {{\mathsf{skip}}};q$. Then for any $i\le k$ there is some $u_i$ such that ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_i)} = u_i;q$. As above, we define ${\mathit{sq}}{^{\scriptscriptstyle\prime}}_i {\mathrel{\stackrel{\scriptscriptstyle{\mathit{def}}}{=}}}(u_i, {{\mathsf{\mathcal{S}}}({\mathit{sq}}_i)})$ for all $i\le k$, and ${\mathit{sq}}{^{\scriptscriptstyle\prime}}_i {\mathrel{\stackrel{\scriptscriptstyle{\mathit{def}}}{=}}}({{\mathsf{skip}}}, {{\mathsf{\mathcal{S}}}({\mathit{sq}}_i)})$ with ${\mathit{sq}}{^{\scriptscriptstyle\prime}}_{i-1} {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{E}}}\:}{\mathit{sq}}{^{\scriptscriptstyle\prime}}_{i}$ for all $i> k$. Further, there must be some $l$ with $k < l \le n$ and ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_{l})} = q$, for otherwise we would have ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_{n})} \neq {{\mathsf{skip}}}$. Thus, ${{^{\scriptscriptstyle l}|{\mathit{sq}}} \Vdash_{\scriptscriptstyle{\mathcal T}} \top {\smalltriangleright}q {\smalltriangleright}Q {\smalltriangleright}n-l}$.
For the sake of brevity, next proposition handles the cases for the parallel composition of two components only. However, the same principle can be applied to ${\parallel\!\!p_1, \ldots, p_m}$ by splitting it into $p_1$ and ${\parallel\!\!p_2, \ldots, p_m}$ which would consequently lead to $2^m - 1$ cases.
\[thm:parallel2-live\] The following implications hold:
1. if ${{\mathit{sq}}\Vdash_{\scriptscriptstyle{\mathcal N}} P {\smalltriangleright}p \parallel q {\smalltriangleright}Q}$ then there exist ${\mathit{sq}}^p$ with ${{\mathit{sq}}\sim {\mathit{sq}}^p}$ and ${\mathit{sq}}^q$ with ${{\mathit{sq}}\sim {\mathit{sq}}^q}$ such that either
1. ${{\mathit{sq}}^p \Vdash_{\scriptscriptstyle{\mathcal N}} P {\smalltriangleright}p {\smalltriangleright}Q}$ and ${{\mathit{sq}}^q \Vdash_{\scriptscriptstyle{\mathcal N}} P {\smalltriangleright}q {\smalltriangleright}Q}$, or
2. ${{\mathit{sq}}^p \Vdash_{\scriptscriptstyle{\mathcal N}} P {\smalltriangleright}p {\smalltriangleright}Q}$ and there exists some $n$ with ${{\mathit{sq}}^q \Vdash_{\scriptscriptstyle{\mathcal T}} P {\smalltriangleright}q {\smalltriangleright}Q {\smalltriangleright}n}$, or
3. ${{\mathit{sq}}^q \Vdash_{\scriptscriptstyle{\mathcal N}} P {\smalltriangleright}q {\smalltriangleright}Q}$ and there exists some $n$ with ${{\mathit{sq}}^p \Vdash_{\scriptscriptstyle{\mathcal T}} P {\smalltriangleright}p {\smalltriangleright}Q {\smalltriangleright}n}$;
2. if ${{\mathit{sq}}\Vdash_{\scriptscriptstyle{\mathcal T}} P {\smalltriangleright}p \parallel q {\smalltriangleright}Q {\smalltriangleright}n}$ then there exist ${\mathit{sq}}^p$ with ${{\mathit{sq}}\sim {\mathit{sq}}^p}$, ${\mathit{sq}}^q$ with ${{\mathit{sq}}\sim {\mathit{sq}}^q}$, $n_1 < n$ and $n_2 < n$ such that ${{\mathit{sq}}^p \Vdash_{\scriptscriptstyle{\mathcal T}} P {\smalltriangleright}p {\smalltriangleright}Q {\smalltriangleright}n_1}$ and ${{\mathit{sq}}^q \Vdash_{\scriptscriptstyle{\mathcal T}} P {\smalltriangleright}q {\smalltriangleright}Q {\smalltriangleright}n_2}$.
In (1) we can assume that ${\mathit{sq}}\in {\lsemanticp\parallel q\rsemantic^{\scriptscriptstyle{\mathit{fair}}}_{\rho}}$ does not terminate. First, note that if there would be some $u, v, k, l$ with ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_k)} = {{\mathsf{skip}}}\parallel v$ and ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_l)} = u \parallel {{\mathsf{skip}}}$ then with $m$ as the maximum of $k$ and $l$ we would have ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_m)} = {{\mathsf{skip}}}\parallel {{\mathsf{skip}}}$. Furthermore, fairness of ${\mathit{sq}}$ would provide some $n > m$ with ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_n)} = {{\mathsf{skip}}}$, contradicting non-termination. Hence, we particularly have that for all $i$ there are some $u_i$ and $v_i$ such that ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_i)} = u_i\parallel v_i$. Then we can define ${\mathit{sq}}^p_i {\mathrel{\stackrel{\scriptscriptstyle{\mathit{def}}}{=}}}(u_i, {{\mathsf{\mathcal{S}}}({\mathit{sq}}_i)})$ such that ${\mathit{sq}}^p_i {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}{\mathit{sq}}^p_{i+1}$ if ${\mathit{sq}}_i {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}{\mathit{sq}}_{i+1}$ is due to a program step from $u_i$ to $u_{i+1}$ and ${\mathit{sq}}^p_i {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{E}}}\:}{\mathit{sq}}^p_{i+1}$ otherwise. This construction provides ${{\mathit{sq}}\sim {\mathit{sq}}^p}$. Regarding fairness of ${\mathit{sq}}^p$, let $x \in {\mathit{{\mathcal P}\!os}}\:u_i$ be an always available position with some $i$. Then $1x$ is an always available position of $u_i \parallel v_i$, and by fairness of ${\mathit{sq}}$ we obtain some $j\ge i$ such that ${\rho \vdash ({{{\mathsf{\mathcal{P}}}(sq_j)} |_{1x}}, {{\mathsf{\mathcal{S}}}(sq_j)}) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(p{^{\scriptscriptstyle\prime}}, {{\mathsf{\mathcal{S}}}({\mathit{sq}}_{j+1})})}$ holds with ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_{j+1})} = {{{\mathsf{\mathcal{P}}}({\mathit{sq}}_j)}[p{^{\scriptscriptstyle\prime}}]_{1x}}$. This is the same as ${\rho \vdash ({{{\mathsf{\mathcal{P}}}({\mathit{sq}}^p_j)} |_{x}}, {{\mathsf{\mathcal{S}}}(sq_j)}) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(p{^{\scriptscriptstyle\prime}}, {{\mathsf{\mathcal{S}}}({\mathit{sq}}_{j+1})})}$ with ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_{j+1})} = {u_j[p{^{\scriptscriptstyle\prime}}]_{x}} \parallel v_{j+1}$, ${{\mathsf{\mathcal{P}}}({\mathit{sq}}^p_{j+1})} = {{{\mathsf{\mathcal{P}}}({\mathit{sq}}^p_j)}[p{^{\scriptscriptstyle\prime}}]_{x}}$. A computation ${\mathit{sq}}^q\in {\lsemanticq\rsemantic^{\scriptscriptstyle{\mathit{fair}}}_{\rho}}$ with ${{\mathit{sq}}\sim {\mathit{sq}}^q}$ can be constructed likewise.
Thus, if $u_i \neq {{\mathsf{skip}}}$ and $v_i \neq {{\mathsf{skip}}}$ hold for all $i$ then ${{\mathit{sq}}^p \Vdash_{\scriptscriptstyle{\mathcal N}} P {\smalltriangleright}p {\smalltriangleright}Q}$ and ${{\mathit{sq}}^q \Vdash_{\scriptscriptstyle{\mathcal N}} P {\smalltriangleright}q {\smalltriangleright}Q}$ follow from the above construction. In case $u_i \neq {{\mathsf{skip}}}$ holds for all $i$ but there is some $j$ with $v_j = {{\mathsf{skip}}}$, we have ${{\mathit{sq}}^p \Vdash_{\scriptscriptstyle{\mathcal N}} P {\smalltriangleright}p {\smalltriangleright}Q}$ and ${{\mathit{sq}}^q \Vdash_{\scriptscriptstyle{\mathcal T}} P {\smalltriangleright}q {\smalltriangleright}Q {\smalltriangleright}n}$ where $n$ is the least index with $v_n = {{\mathsf{skip}}}$. The remaining case with $v_i \neq {{\mathsf{skip}}}$ for all $i$, and some $j$ with $u_j = {{\mathsf{skip}}}$ is symmetric.
In (2) we have ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_n)} = {{\mathsf{skip}}}$ with $n>0$ and some $u_i$ and $v_i$ such that ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_i)} = u_i\parallel v_i$ holds for all $i<n$. Since $n$ is the least such index, ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_{n-1})} = {{\mathsf{skip}}}\parallel {{\mathsf{skip}}}$ must hold as well. Hence, there exists the least index $n_1<n$ with $u_{n_1} = {{\mathsf{skip}}}$ and the least index $n_2<n$ with $v_{n_2} = {{\mathsf{skip}}}$. Then we can define ${\mathit{sq}}^p {\mathrel{\stackrel{\scriptscriptstyle{\mathit{def}}}{=}}}(u_i, {{\mathsf{\mathcal{S}}}({\mathit{sq}}_i)})$ for all $i\le n_1$, and ${\mathit{sq}}^p {\mathrel{\stackrel{\scriptscriptstyle{\mathit{def}}}{=}}}({{\mathsf{skip}}}, {{\mathsf{\mathcal{S}}}({\mathit{sq}}_i)})$ with ${\mathit{sq}}^p_{i-1} {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{E}}}\:}{\mathit{sq}}^p_{i}$ for all $i> n_1$, such that ${{\mathit{sq}}\sim {\mathit{sq}}^p}$ and ${{\mathit{sq}}^p \Vdash_{\scriptscriptstyle{\mathcal T}} P {\smalltriangleright}p {\smalltriangleright}Q {\smalltriangleright}n_1}$ follow by construction. Finally, ${\mathit{sq}}^q$ can be defined likewise to derive ${{\mathit{sq}}^q \Vdash_{\scriptscriptstyle{\mathcal T}} P {\smalltriangleright}q {\smalltriangleright}Q {\smalltriangleright}n_2}$.
Regarding ${\mathsf{while}}$-statements, the case ${{\mathit{sq}}\Vdash_{\scriptscriptstyle{\mathcal T}} P {\smalltriangleright}{{\mathsf{while}}\:C\;{\mathsf{do}}\;p\;{\mathsf{subsequently}}\;q\;{\mathsf{od}}} {\smalltriangleright}Q {\smalltriangleright}n}$ does not entail that ${\mathit{sq}}$ ever visits a state satisfying $C$. What we can infer is that there must be some $i < n$ with the step ${\mathit{sq}}_i {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}{\mathit{sq}}_{i+1}$ such that ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_i)} = {{\mathsf{while}}\:C\;{\mathsf{do}}\;p\;{\mathsf{subsequently}}\;q\;{\mathsf{od}}}$, ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_{i+1})} = q$ and ${{\mathsf{\mathcal{S}}}({\mathit{sq}}_{i})} = {{\mathsf{\mathcal{S}}}({\mathit{sq}}_{i+1})} \notin C$ hold. That is, ${{^{\scriptscriptstyle i+1}|{\mathit{sq}}} \Vdash_{\scriptscriptstyle{\mathcal T}} \neg C {\smalltriangleright}q {\smalltriangleright}Q {\smalltriangleright}n - i - 1}$ must hold.
By contrast, from ${{\mathit{sq}}\Vdash_{\scriptscriptstyle{\mathcal N}} P {\smalltriangleright}{{\mathsf{while}}\:C\;{\mathsf{do}}\;p\;{\mathsf{subsequently}}\;q\;{\mathsf{od}}} {\smalltriangleright}Q}$ follows that ${\mathit{sq}}$ visits certain states satisfying $C$ infinitely often, provided $p$ and $q$ terminate when starting from $C$ and $\neg C$, respectively.
\[thm:while-live\] If ${{\mathit{sq}}\Vdash_{\scriptscriptstyle{\mathcal N}} P {\smalltriangleright}{{\mathsf{while}}\:C\;{\mathsf{do}}\;p\;{\mathsf{subsequently}}\;q\;{\mathsf{od}}} {\smalltriangleright}Q}$ then either
1. there exists a strictly ascending infinite sequence of natural numbers $\phi$ such that for any $i$ there exist $n< \phi_{i+1} - \phi_i$ and ${\mathit{sq}}{^{\scriptscriptstyle\prime}}$ with ${{^{\scriptscriptstyle \phi_i}|{\mathit{sq}}} \sim {\mathit{sq}}{^{\scriptscriptstyle\prime}}}$ and ${{\mathit{sq}}{^{\scriptscriptstyle\prime}}\Vdash_{\scriptscriptstyle{\mathcal T}} C {\smalltriangleright}p {\smalltriangleright}Q {\smalltriangleright}n}$, or
2. there exist $n>0$ and ${\mathit{sq}}{^{\scriptscriptstyle\prime}}$ such that ${{^{\scriptscriptstyle n}|{\mathit{sq}}} \sim {\mathit{sq}}{^{\scriptscriptstyle\prime}}}$ and ${{\mathit{sq}}{^{\scriptscriptstyle\prime}}\Vdash_{\scriptscriptstyle{\mathcal N}} C {\smalltriangleright}p {\smalltriangleright}Q}$, or
3. there exists some $n > 0$ with ${{^{\scriptscriptstyle n}|{\mathit{sq}}} \Vdash_{\scriptscriptstyle{\mathcal N}} \neg C {\smalltriangleright}q {\smalltriangleright}Q}$.
Let $p_{{\mathsf{while}}}$ abbreviate ${{\mathsf{while}}\:C\;{\mathsf{do}}\;p\;{\mathsf{subsequently}}\;q\;{\mathsf{od}}}$ and assume ${\mathit{sq}}\in {\lsemanticp_{{\mathsf{while}}}\rsemantic^{\scriptscriptstyle{\mathit{fair}}}_{\rho}} \cap {\mathcal{I}\!\mathit{n}\:P}$ has $\neg Q$ as an invariant and does not reach a ${{\mathsf{skip}}}$-configuration. Further, we will also assume the respective negations of (2) and (3), which do not need to take accounts of $Q$:
1. *for any ${\mathit{sq}}{^{\scriptscriptstyle\prime}}\in {\lsemanticp\rsemantic^{\scriptscriptstyle{\mathit{fair}}}_{\rho}} \cap {\mathcal{I}\!\mathit{n}\:C}$ and $n>0$ such that ${{^{\scriptscriptstyle n}|{\mathit{sq}}} \sim {\mathit{sq}}{^{\scriptscriptstyle\prime}}}$ there exists some $j$ with ${{\mathsf{\mathcal{P}}}({\mathit{sq}}{^{\scriptscriptstyle\prime}}_j)} = {{\mathsf{skip}}}$*,
2. *for any $n > 0$ such that ${^{\scriptscriptstyle n}|{\mathit{sq}}} \in {\lsemanticq\rsemantic^{\scriptscriptstyle{\mathit{fair}}}_{\rho}} \cap {\mathcal{I}\!\mathit{n}\:\neg C}$ there exists some $j$ with ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_{n+j})} = {{\mathsf{skip}}}$*,
in order to establish (1). To this end, we first show that for any $i$ such that ${\rho \vdash {\mathit{sq}}_i {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}{\mathit{sq}}_{i+1}}$ with ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_i)} = p_{{\mathsf{while}}}$ and ${{\mathsf{\mathcal{S}}}({\mathit{sq}}_i)} \in C$ there exists $j > i$ such that ${\rho \vdash {\mathit{sq}}_j {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}{\mathit{sq}}_{j+1}}$ with ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_j)} = p_{{\mathsf{while}}}$ and ${{\mathsf{\mathcal{S}}}({\mathit{sq}}_j)} \in C$. Note that ${\mathit{sq}}_{i+1} = (p;{{\mathsf{skip}}};p_{{\mathsf{while}}},{{\mathsf{\mathcal{S}}}({\mathit{sq}}_i)})$ and suppose there is no $k > i$ with ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_{k})} = {{\mathsf{skip}}};{{\mathsf{skip}}};p_{{\mathsf{while}}}$. Consequently, for all $k > i$ there is some $u_k$ such that ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_{k})} = u_k;{{\mathsf{skip}}};p_{{\mathsf{while}}}$ holds, and we can define ${\mathit{sq}}{^{\scriptscriptstyle\prime}}_k {\mathrel{\stackrel{\scriptscriptstyle{\mathit{def}}}{=}}}(u_{i+k+1}, {{\mathsf{\mathcal{S}}}({\mathit{sq}}_{i+k+1})})$ for all $k \in \mathbb{N}$. Thus, ${\mathit{sq}}{^{\scriptscriptstyle\prime}}\in {\lsemanticp\rsemantic^{\scriptscriptstyle{\mathit{fair}}}_{\rho}} \cap {\mathcal{I}\!\mathit{n}\:C}$ and ${{^{\scriptscriptstyle i+1}|{\mathit{sq}}} \sim {\mathit{sq}}{^{\scriptscriptstyle\prime}}}$ such that (a) yields some $j$ with ${{\mathsf{\mathcal{P}}}({\mathit{sq}}{^{\scriptscriptstyle\prime}}_j)} = {{\mathsf{skip}}}$, ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_{i+j+1})} = {{\mathsf{skip}}};{{\mathsf{skip}}};p_{{\mathsf{while}}}$, which is a contradiction.
Then let $k > i$ with ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_{k})} = {{\mathsf{skip}}};{{\mathsf{skip}}};p_{{\mathsf{while}}}$. Fairness of ${\mathit{sq}}$ ensures that there are also some $l > k$ with ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_{l})} = {{\mathsf{skip}}};p_{{\mathsf{while}}}$, and some $m > l$ with ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_{m})} = p_{{\mathsf{while}}}$. Once more, since the position $0$ of ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_{m})}$ is always available, by fairness of ${\mathit{sq}}$ we obtain some $n \ge m$ with ${\rho \vdash {\mathit{sq}}_n {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}{\mathit{sq}}_{n+1}}$ and ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_n)} = p_{{\mathsf{while}}}$. In case ${{\mathsf{\mathcal{S}}}({\mathit{sq}}_n)} \in C$ we are done. Otherwise, we have ${\mathit{sq}}_{n+1} = (q,{{\mathsf{\mathcal{S}}}({\mathit{sq}}_n)})$, Hence, ${^{\scriptscriptstyle n+1}|{\mathit{sq}}} \in {\lsemanticq\rsemantic^{\scriptscriptstyle{\mathit{fair}}}_{\rho}} \cap {\mathcal{I}\!\mathit{n}\:\neg C}$, such that (b) provides some $j$ with ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_{n+1+j})} = {{\mathsf{skip}}}$, contradicting the assumption that ${\mathit{sq}}$ does not reach a ${{\mathsf{skip}}}$-configuration.
The intermediate result is a function $\psi$ that sends each $i \in \mathbb{N}$ with ${\rho \vdash {\mathit{sq}}_i {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}{\mathit{sq}}_{i+1}}$, ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_i)} = p_{{\mathsf{while}}}$ and ${{\mathsf{\mathcal{S}}}({\mathit{sq}}_i)} \in C$ to $\psi\:i > i$ with ${\rho \vdash {\mathit{sq}}_{\psi\:i} {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}{\mathit{sq}}_{\psi\:i+1}}$, ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_{\psi\:i})} = p_{{\mathsf{while}}}$ and ${{\mathsf{\mathcal{S}}}({\mathit{sq}}_{\psi\:i})} \in C$. Furthermore, fairness of ${\mathit{sq}}$ ensures that there exists some $i_0$ with ${\rho \vdash {\mathit{sq}}_{i_0} {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}{\mathit{sq}}_{i_0+1}}$ and ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_{i_0})} = p_{{\mathsf{while}}}$. Note that ${{\mathsf{\mathcal{S}}}({\mathit{sq}}_{i_0})} \in C$ must hold, for otherwise we would have ${^{\scriptscriptstyle i_0+1}|{\mathit{sq}}} \in {\lsemanticq\rsemantic^{\scriptscriptstyle{\mathit{fair}}}_{\rho}} \cap {\mathcal{I}\!\mathit{n}\:\neg C}$ and, hence, a contradiction due to (b) as above. Then we define a strictly ascending infinite sequence $\phi_n$ by $\psi^n\:i_0 + 1$.
Let $i$ be an arbitrary natural number in the remainder of the proof. The definition of $\phi$ induces ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_{\phi_i})} = p;{{\mathsf{skip}}};p_{{\mathsf{while}}}$, ${{\mathsf{\mathcal{S}}}({\mathit{sq}}_{\phi_i})} \in C$ and, moreover, ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_{\phi_{i+1} - 1})} = p_{{\mathsf{while}}}$. Hence, there must be the least index $n$ with $\phi_i \le n < \phi_{i+1}$ and ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_n)} = {{\mathsf{skip}}};{{\mathsf{skip}}};p_{{\mathsf{while}}}$. Furthermore, for any $l$ with $\phi_i \le l \le n$ there is some $u_l$ with ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_l)} = u_l;{{\mathsf{skip}}};p_{{\mathsf{while}}}$ such that we can define a computation ${\mathit{sq}}{^{\scriptscriptstyle\prime}}$ by ${\mathit{sq}}{^{\scriptscriptstyle\prime}}_l {\mathrel{\stackrel{\scriptscriptstyle{\mathit{def}}}{=}}}(u_{\phi_i + l}, {{\mathsf{\mathcal{S}}}({\mathit{sq}}_{\phi_i + l})})$ for any $l\le n - \phi_i$, and ${\mathit{sq}}{^{\scriptscriptstyle\prime}}_l {\mathrel{\stackrel{\scriptscriptstyle{\mathit{def}}}{=}}}({{\mathsf{skip}}}, {{\mathsf{\mathcal{S}}}({\mathit{sq}}_{\phi_i + l})})$ with ${\mathit{sq}}{^{\scriptscriptstyle\prime}}_{l-1} {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{E}}}\:}{\mathit{sq}}{^{\scriptscriptstyle\prime}}_{l}$ if $l > n - \phi_i$. Thus, for ${\mathit{sq}}{^{\scriptscriptstyle\prime}}$ we have ${{^{\scriptscriptstyle \phi_i}|{\mathit{sq}}} \sim {\mathit{sq}}{^{\scriptscriptstyle\prime}}}$ and ${{\mathit{sq}}{^{\scriptscriptstyle\prime}}\Vdash_{\scriptscriptstyle{\mathcal T}} C {\smalltriangleright}p {\smalltriangleright}Q {\smalltriangleright}n - \phi_i}$.
Program correspondence and fair computations {#Sb:fair-corr}
--------------------------------------------
Proposition \[thm:corr-sim\] shows how finite potential computations of a program can be replayed by computations of some corresponding program, and this statement can be extended to infinite potential computations. However, things get significantly more involved when attempting to do likewise with fair computations. The major difficulty thereby is that apart from program correspondence there also must be some kind of matching between positions of the configurations on the corresponding computations. To sketch that, suppose $p$ has the form ${\parallel\!\!u_1, \ldots, u_m}$ whereas $q$ – ${\parallel\!\!v_1, \ldots, v_m}$ and we additionally have ${\rho, \rho{^{\scriptscriptstyle\prime}}\models u_i\sqsupseteq_{r}v_i}$ for all $i \in \{1, \ldots, m\}$. Now, if $q$ performs a program step which is due to a position $ix$ with $x\in {\mathit{{\mathcal P}\!os}}\:v_i$ then it can be matched by a step of $p$ in $u_i$, based on some position $x{^{\scriptscriptstyle\prime}}\in {\mathit{{\mathcal P}\!os}}\:u_i$. Intuitively, in this setting positions of $p$ and $q$ can be matched componentwise and an exact definition for that will be elaborated below.
A program $(\rho, p)$ is called *locally non-blocking* if for any ${{\mathsf{await}}\:C\;{\mathsf{do}}\;p{^{\scriptscriptstyle\prime}}\;{\mathsf{od}}}$ occurring in $p$ we have $C = \top$ and for any $\sigma$ there exists $\sigma{^{\scriptscriptstyle\prime}}$ such that ${\rho \vdash (p{^{\scriptscriptstyle\prime}}, \sigma) {\stackrel{\scriptscriptstyle\star}{\mathop{\rightarrow}}_{\scriptscriptstyle{\mathcal{P}}}}({{\mathsf{skip}}}, \sigma{^{\scriptscriptstyle\prime}})}$. Furthermore, $(\rho, p)$ will be called *non-blocking* if it is locally non-blocking and for any label $i \in {\mathcal{J}\!\mathit{umps}}(\rho, p)$ the program $(\rho, \rho\:i)$ is locally non-blocking as well.
In other words, $(\rho, p)$ is locally non-blocking iff all $x \in {\mathit{{\mathcal P}\!os}}\:p$ are always available. Note also that as with sequential programs, if $(\rho, p)$ is non-blocking then $(\rho, p{^{\scriptscriptstyle\prime}})$ is non-blocking for any program step ${\rho \vdash (p, \sigma) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(p{^{\scriptscriptstyle\prime}}, \sigma{^{\scriptscriptstyle\prime}})}$.
Any $p$ and $q$ will be called *corresponding componentwise* w.r.t. $m, r, \rho, \rho{^{\scriptscriptstyle\prime}}$ if there exist $p_1, \ldots, p_m$ and $q_1, \ldots, q_m$ such that $p =\: {\parallel\!\!p_1,\ldots, p_m}$, $q = \:{\parallel\!\!q_1,\ldots, q_m}$, and
1. ${\rho, \rho{^{\scriptscriptstyle\prime}}\models p_i\sqsupseteq_{r}q_i}$,
2. $(\rho, p_i)$ is sequential,
3. $(\rho{^{\scriptscriptstyle\prime}}, q_i)$ is non-blocking
hold for all $i \in \{1, \ldots, m\}$.
Moreover, in this setting we need to know which position in ${\mathit{{\mathcal P}\!os}}\:p$ has actually been ‘fired’ by a program step ${\rho \vdash (p, \sigma) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(q, \sigma{^{\scriptscriptstyle\prime}})}$. Proposition \[thm:rpos-pstep\] does not provide unique existence of such position due to the following technical complication. Consider ${{\mathsf{jump}}\:i} \parallel {{\mathsf{jump}}\:j}$ with ${\rho\:i} = {{\mathsf{jump}}\:i}$ and ${\rho\:j} = {{\mathsf{jump}}\:j}$. Thus, ${\rho \vdash ({{\mathsf{jump}}\:i} \parallel {{\mathsf{jump}}\:j}, \sigma) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}({{\mathsf{jump}}\:i} \parallel {{\mathsf{jump}}\:j}, \sigma)}$ is the only possible program step for any $\sigma$, regardless which of the positions $10$ or $20$ we choose. This ambiguity poses a particular difficulty to replaying fair computations. Furthermore, it is worth noting that any computation ${\mathit{sq}}\in {\lsemantic{{\mathsf{jump}}\:i} \parallel {{\mathsf{jump}}\:j}\rsemantic^\omega_{\rho}} \cap {\mathcal{E}\!\mathit{nv}\:\bot}$ has the form $({{\mathsf{jump}}\:i} \parallel {{\mathsf{jump}}\:j}, {{\mathsf{\mathcal{S}}}({\mathit{sq}}_0)}), ({{\mathsf{jump}}\:i} \parallel {{\mathsf{jump}}\:j}, {{\mathsf{\mathcal{S}}}({\mathit{sq}}_0)}), \ldots$ and within the framework one can neither prove nor disprove that ${\mathit{sq}}$ is fair: this would require access to the information which of the components has been chosen at any given step to apply the computational rule ‘Parallel’ in Figure \[fig:pstep\].
Avoiding the ambiguity, only those retrieve functions $\rho$ will be considered in the following for which we have $\rho\:i \neq {{\mathsf{cjump}}\:C\;{\mathsf{to}}\:i\;{\mathsf{otherwise}}\;u\;{\mathsf{end}}}$ for any $C, i$ and $u$. With this assumption, any transition ${\rho \vdash (p, \sigma) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(q, \sigma{^{\scriptscriptstyle\prime}})}$ entails a unique position $x \in {\mathit{{\mathcal P}\!os}}\:p$ with ${\rho \vdash ({p |_{x}}, \sigma) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(p{^{\scriptscriptstyle\prime}}, \sigma{^{\scriptscriptstyle\prime}})}$ and $q = {p[p{^{\scriptscriptstyle\prime}}]_{x}}$. Such $x$ will be called *the fired position of* ${\rho \vdash (p, \sigma) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(q, \sigma{^{\scriptscriptstyle\prime}})}$.
\[thm:fair-corr-sim1\] Assume $p =\: {\parallel\!\!p_1,\ldots, p_m}$ and $q = \:{\parallel\!\!q_1,\ldots, q_m}$ correspond componentwise w.r.t. $m, r, \rho, \rho{^{\scriptscriptstyle\prime}}$, and ${\mathit{sq}}^q \in {\lsemanticq\rsemantic^\omega_{\rho{^{\scriptscriptstyle\prime}}}} \cap {\mathcal{E}\!\mathit{nv}\:{\mathsf{id}}}$. Further, let $\sigma$ be a state with $(\sigma, {{\mathsf{\mathcal{S}}}({\mathit{sq}}^q_0)}) \in r$. Then there exists a computation ${\mathit{sq}}^p \in {\lsemanticp\rsemantic^\omega_{\rho}} \cap {\mathcal{E}\!\mathit{nv}\:{\mathsf{id}}}$ such that ${{\mathsf{\mathcal{S}}}({\mathit{sq}}^p_0)} = \sigma$ as well as the following conditions hold for all $i \in \mathbb{N}$*:*
1. either ${{\mathsf{\mathcal{P}}}({\mathit{sq}}^p_i)}$ and ${{\mathsf{\mathcal{P}}}({\mathit{sq}}^q_i)}$ correspond componentwise w.r.t. $m, r, \rho, \rho{^{\scriptscriptstyle\prime}}$ or both are ${{\mathsf{skip}}}$,
2. $({{\mathsf{\mathcal{S}}}({\mathit{sq}}^p_i)}, {{\mathsf{\mathcal{S}}}({\mathit{sq}}^q_i)}) \in r$,
3. if $i>0$ and we have ${\rho{^{\scriptscriptstyle\prime}}\vdash {\mathit{sq}}^q_{i-1} {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}{\mathit{sq}}^q_{i}}$ with the fired position $jx$ where $1\le j \le m$ then we have the program step ${\rho \vdash {\mathit{sq}}^p_{i-1} {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}{\mathit{sq}}^p_{i}}$ which is due to the fired position $jx{^{\scriptscriptstyle\prime}}$ with some $x{^{\scriptscriptstyle\prime}}$.
We show that any finite computation ${\mathit{sq}}\in {\lsemanticp\rsemantic_{\rho}} \cap {\mathcal{E}\!\mathit{nv}\:{\mathsf{id}}}$ of length $n$ that satisfies the conditions (1)–(3) for all $i < n$ can be extended to ${\mathit{sq}}{^{\scriptscriptstyle\prime}}\in {\lsemanticp\rsemantic_{\rho}} \cap {\mathcal{E}\!\mathit{nv}\:{\mathsf{id}}}$ of length $n + 1$ satisfying the conditions (1)–(3) for all $i \le n$. This way we can construct a computation ${\mathit{sq}}^p \in {\lsemanticp\rsemantic^\omega_{\rho}} \cap {\mathcal{E}\!\mathit{nv}\:{\mathsf{id}}}$ by assigning $(p, \sigma)$ to ${\mathit{sq}}^p_0$ first, so that for any $n \in \mathbb{N}$ a configuration extending ${\mathit{sq}}^p_0, \ldots, {\mathit{sq}}^p_n$ can be assigned to ${\mathit{sq}}^p_{n+1}$.
Then assume ${\mathit{sq}}\in {\lsemanticp\rsemantic_{\rho}} \cap {\mathcal{E}\!\mathit{nv}\:{\mathsf{id}}}$ satisfies the conditions (1)–(3) for all $i < n$ with $n = |{\mathit{sq}}|$. Since ${\mathit{sq}}^q \in {\mathcal{E}\!\mathit{nv}\:{\mathsf{id}}}$, in case of an environment step ${\mathit{sq}}^q_{n-1} {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{E}}}\:}{\mathit{sq}}^q_n$ we have ${\mathit{sq}}^q_{n} = {\mathit{sq}}^q_{n-1}$, such that we can extend ${\mathit{sq}}$ by an environment step to ${\mathit{sq}}_0, \ldots, {\mathit{sq}}_{n-1}, {\mathit{sq}}_n$ where ${\mathit{sq}}_{n} = {\mathit{sq}}_{n-1}$.
Otherwise, we have a program step ${\rho{^{\scriptscriptstyle\prime}}\vdash {\mathit{sq}}^q_{n-1} {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}{\mathit{sq}}^q_{n}}$, ${{\mathsf{\mathcal{P}}}({\mathit{sq}}^q_{n-1})}$ cannot be ${{\mathsf{skip}}}$. Hence, ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_{n-1})}$ and ${{\mathsf{\mathcal{P}}}({\mathit{sq}}^q_{n-1})}$ must correspond componentwise w.r.t. $m, r, \rho, \rho{^{\scriptscriptstyle\prime}}$, such that ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_{n-1})} =\: {\parallel\!\!u_1,\ldots, u_m}$ and ${{\mathsf{\mathcal{P}}}({\mathit{sq}}^q_{n-1})} = \:{\parallel\!\!v_1,\ldots, v_m}$ particularly hold with some $u_1, \ldots, u_m$ and $v_1, \ldots, v_m$.
If ${\mathit{sq}}^q_{n} \neq {{\mathsf{skip}}}$ then the transition ${\rho{^{\scriptscriptstyle\prime}}\vdash {\mathit{sq}}^q_{n-1} {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}{\mathit{sq}}^q_{n}}$ has the fired position of the form $jx$ where $1\le j \le m$ and $x \in {\mathit{{\mathcal P}\!os}}\:v_j$, such that ${\rho{^{\scriptscriptstyle\prime}}\vdash ({v_j |_{x}}, {{\mathsf{\mathcal{S}}}({\mathit{sq}}^q_{n-1})}) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(w, {{\mathsf{\mathcal{S}}}({\mathit{sq}}^q_{b})})}$ and ${{\mathsf{\mathcal{P}}}({\mathit{sq}}^q_n)}= \:{\parallel\!\!v_1,\ldots, {v_j[w]_{x}}, \ldots, v_m}$ hold with some $w$. For the component $v_j$ we thus have the program step ${\rho{^{\scriptscriptstyle\prime}}\vdash (v_j, {{\mathsf{\mathcal{S}}}({\mathit{sq}}^q_{n-1})}) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}({v_j[w]_{x}}, {{\mathsf{\mathcal{S}}}({\mathit{sq}}^q_{n})})}$, and using the correspondence ${\rho, \rho{^{\scriptscriptstyle\prime}}\models u_j\sqsupseteq_{r}v_j}$ we obtain a step ${\rho \vdash (u_j, {{\mathsf{\mathcal{S}}}({\mathit{sq}}_{n-1})}) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(u{^{\scriptscriptstyle\prime}}_j, \sigma)}$ with some $u{^{\scriptscriptstyle\prime}}_j$ and $\sigma$ satisfying ${\rho, \rho{^{\scriptscriptstyle\prime}}\models u{^{\scriptscriptstyle\prime}}_j\sqsupseteq_{r}{v_j[w]_{x}}}$ and $(\sigma, {{\mathsf{\mathcal{S}}}({\mathit{sq}}^q_i)}) \in r$. Hence, there must be a position $x{^{\scriptscriptstyle\prime}}\in {\mathit{{\mathcal P}\!os}}\:u_j$ and some $w{^{\scriptscriptstyle\prime}}$ such that ${\rho \vdash ({u_j |_{x{^{\scriptscriptstyle\prime}}}}, {{\mathsf{\mathcal{S}}}({\mathit{sq}}^p_{n-1})}) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}(w{^{\scriptscriptstyle\prime}}, \sigma)}$ with $u{^{\scriptscriptstyle\prime}}_j = {u_j[w{^{\scriptscriptstyle\prime}}]_{x{^{\scriptscriptstyle\prime}}}}$ holds. This further implies the program step ${\rho \vdash ({\parallel\!\!u_1,\ldots, u_j, \ldots, u_m}, {{\mathsf{\mathcal{S}}}({\mathit{sq}}_{n-1})}) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}({\parallel\!\!u_1,\ldots, u{^{\scriptscriptstyle\prime}}_j, \ldots, u_m}, \sigma)}$, such that we can extend ${\mathit{sq}}$ to ${\mathit{sq}}_0, \ldots, {\mathit{sq}}_{n-1}, ({\parallel\!\!u_1,\ldots, u{^{\scriptscriptstyle\prime}}_j, \ldots, u_m}, \sigma)$ with the fired position $jx{^{\scriptscriptstyle\prime}}$.
Lastly, in case ${\mathit{sq}}^q_{n} = {{\mathsf{skip}}}$ we have $v_j = {{\mathsf{skip}}}$ and $u_j = {{\mathsf{skip}}}$ for all $j \in \{1, \ldots, m\}$, such that we can extend ${\mathit{sq}}$ to ${\mathit{sq}}_0, \ldots, {\mathit{sq}}_{n-1},({{\mathsf{skip}}}, {{\mathsf{\mathcal{S}}}({\mathit{sq}}_{n-1})})$.
Next proposition shows that computations retain the fairness property when replayed as described above.
\[thm:fair-corr-sim2\] Let ${\mathit{sq}}^p \in {\lsemanticp\rsemantic^\omega_{\rho}}$ and ${\mathit{sq}}^q \in {\lsemanticq\rsemantic^\omega_{\rho{^{\scriptscriptstyle\prime}}}}$ be computations satisfying the conditions *(1)–(3)* of Proposition *\[thm:fair-corr-sim1\]* for all $i \in \mathbb{N}$. Then ${\mathit{sq}}^p$ is fair whenever ${\mathit{sq}}^q$ is.
Let $i$ be a natural number and $y$ – an always available position of ${{\mathsf{\mathcal{P}}}({\mathit{sq}}^p_i)}$. Then by the condition (1), ${{\mathsf{\mathcal{P}}}({\mathit{sq}}^p_i)}$ and ${{\mathsf{\mathcal{P}}}({\mathit{sq}}^q_i)}$ must correspond componentwise w.r.t. $m, r, \rho, \rho{^{\scriptscriptstyle\prime}}$, such that ${{\mathsf{\mathcal{P}}}({\mathit{sq}}^p_{i})} =\: {\parallel\!\!u_1,\ldots, u_m}$ and ${{\mathsf{\mathcal{P}}}({\mathit{sq}}^q_{i})} = \:{\parallel\!\!v_1,\ldots, v_m}$ particularly hold with some $u_1, \ldots, u_m$ and $v_1, \ldots, v_m$.
If $y = 0$ then we have $u_k = {{\mathsf{skip}}}$ for all $k \in \{1, \ldots, m\}$ and, hence, $v_k = {{\mathsf{skip}}}$ for all $k \in \{1, \ldots, m\}$ due to the componentwise correspondence. By fairness of ${\mathit{sq}}^q$ we obtain some $j \ge i$ with ${{\mathsf{\mathcal{P}}}(sq^q_j)} = \:{\parallel\!\!{{\mathsf{skip}}}, \ldots, {{\mathsf{skip}}}}$ and ${{\mathsf{\mathcal{P}}}(sq^q_{j+1})} = {{\mathsf{skip}}}$. Thus, $sq^q_{j+1}$ is the first ${{\mathsf{skip}}}$-configuration on ${\mathit{sq}}^q$ and, consequently, $sq^p_{j+1}$ must be the first ${{\mathsf{skip}}}$-configuration on ${\mathit{sq}}^p$. That is, ${{{\mathsf{\mathcal{P}}}(sq^p_j)} |_{0}} = {{{\mathsf{\mathcal{P}}}(sq^p_i)} |_{0}} = \:{\parallel\!\!{{\mathsf{skip}}}, \ldots, {{\mathsf{skip}}}}$ as well as ${\rho \vdash ({{{\mathsf{\mathcal{P}}}(sq^p_j)} |_{0}}, {{\mathsf{\mathcal{S}}}(sq^p_j)}) {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}({{\mathsf{skip}}}, {{\mathsf{\mathcal{S}}}({\mathit{sq}}^p_{j+1})})}$ hold with ${{\mathsf{\mathcal{P}}}({\mathit{sq}}^p_{j+1})} = {{\mathsf{skip}}}= {{{\mathsf{\mathcal{P}}}({\mathit{sq}}^p_j)}[{{\mathsf{skip}}}]_{0}}$.
Next, assume $y = kx$ with some $k \in \{1, \ldots, m\}$ and $x \in {\mathit{{\mathcal P}\!os}}\:u_k$. Then $v_k \neq {{\mathsf{skip}}}$ due to ${\rho, \rho{^{\scriptscriptstyle\prime}}\models u_k\sqsupseteq_{r}v_k}$. Furthermore, since $(\rho{^{\scriptscriptstyle\prime}}, v_k)$ is locally non-blocking there must be some $x{^{\scriptscriptstyle\prime}}\in {\mathit{{\mathcal P}\!os}}\:v_k$ which is always available, $kx{^{\scriptscriptstyle\prime}}$ is an always available position of ${\mathit{sq}}^q_i$ and fairness of ${\mathit{sq}}^q$ provides some $j \ge i$ such that the following conditions hold for some $v{^{\scriptscriptstyle\prime}}_1,\ldots, v{^{\scriptscriptstyle\prime}}_m$:
1. ${{\mathsf{\mathcal{P}}}(sq^q_j)} = \:{\parallel\!\!v{^{\scriptscriptstyle\prime}}_1,\ldots, v{^{\scriptscriptstyle\prime}}_m}$,
2. $x{^{\scriptscriptstyle\prime}}\in {\mathit{{\mathcal P}\!os}}\:v{^{\scriptscriptstyle\prime}}_k$,
3. ${\rho{^{\scriptscriptstyle\prime}}\vdash {\mathit{sq}}^q_j {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}{\mathit{sq}}^q_{j+1}}$ has the fired position $kx{^{\scriptscriptstyle\prime}}$.
Thus, we have ${\mathit{sq}}^p_j = \:{\parallel\!\!u{^{\scriptscriptstyle\prime}}_1,\ldots, u{^{\scriptscriptstyle\prime}}_m}$ and the condition (3) of Proposition [\[thm:fair-corr-sim1\]]{} provides some $x{^{\scriptscriptstyle\prime\scriptscriptstyle\prime}}\in {\mathit{{\mathcal P}\!os}}\:u{^{\scriptscriptstyle\prime}}_k$ that has been fired by ${\rho \vdash {\mathit{sq}}^p_j {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}{\mathit{sq}}^p_{j+1}}$. Now, the position $kx$ in ${\mathit{sq}}^p_i$ could have been fired prior to $j$ by some step ${\rho \vdash {\mathit{sq}}^p_l {\mathop{\rightarrow}\nolimits_{\scriptscriptstyle{\mathcal{P}}}\:}{\mathit{sq}}^p_{l+1}}$ with $i \le l < j$. If this is the case then we are done. Otherwise, $kx$ remains available up to $j$ and, moreover, we have ${{{\mathsf{\mathcal{P}}}({\mathit{sq}}^p_i)} |_{kx}} = {{{\mathsf{\mathcal{P}}}({\mathit{sq}}^p_j)} |_{kx}}$. That is, we have $x, x{^{\scriptscriptstyle\prime\scriptscriptstyle\prime}}\in {\mathit{{\mathcal P}\!os}}\:u{^{\scriptscriptstyle\prime}}_k$, and since by the condition (1) of Proposition [\[thm:fair-corr-sim1\]]{} $u{^{\scriptscriptstyle\prime}}_k$ is locally sequential, $x$ must be equal to $x{^{\scriptscriptstyle\prime\scriptscriptstyle\prime}}$.
As a preparation to the next section, the following corollary explicitly states how termination can be established by means of componentwise correspondence.
\[thm:fair-corr-sim3\] Assume $p =\: {\parallel\!\!p_1,\ldots, p_m}$ and $q = \:{\parallel\!\!q_1,\ldots, q_m}$ correspond componentwise w.r.t. $m, r, \rho, \rho{^{\scriptscriptstyle\prime}}$ and
1. $P{^{\scriptscriptstyle\prime}}\subseteq {r\cdot P}$,
2. for any computation ${\mathit{sq}}\in {\lsemanticp\rsemantic^{\scriptscriptstyle{\mathit{fair}}}_{\rho}} \cap {\mathcal{E}\!\mathit{nv}\:{\mathsf{id}}} \cap {\mathcal{I}\!\mathit{n}\:P}$ there exists some $i$ with ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_i)} = {{\mathsf{skip}}}$.
Then for any computation ${\mathit{sq}}\in {\lsemanticq\rsemantic^{\scriptscriptstyle{\mathit{fair}}}_{\rho{^{\scriptscriptstyle\prime}}}} \cap {\mathcal{E}\!\mathit{nv}\:{\mathsf{id}}} \cap {\mathcal{I}\!\mathit{n}\:P{^{\scriptscriptstyle\prime}}}$ there exists some $i$ with ${{\mathsf{\mathcal{P}}}({\mathit{sq}}_i)} = {{\mathsf{skip}}}$.
Let ${\mathit{sq}}^q \in {\lsemanticq\rsemantic^{\scriptscriptstyle{\mathit{fair}}}_{\rho{^{\scriptscriptstyle\prime}}}} \cap {\mathcal{E}\!\mathit{nv}\:{\mathsf{id}}} \cap {\mathcal{I}\!\mathit{n}\:P{^{\scriptscriptstyle\prime}}}$. Since (1) provides some $\sigma \in P$ with $(\sigma, {{\mathsf{\mathcal{S}}}({\mathit{sq}}^q_0)}) \in r$, from Proposition \[thm:fair-corr-sim1\] we obtain a computation ${\mathit{sq}}^p \in {\lsemanticp\rsemantic^\omega_{\rho}} \cap {\mathcal{E}\!\mathit{nv}\:{\mathsf{id}}} \cap {\mathcal{I}\!\mathit{n}\:P}$ that is fair by Proposition \[thm:fair-corr-sim2\]. Then from (2) follows that there must be some $i$ with ${{\mathsf{\mathcal{P}}}({\mathit{sq}}^p_i)} = {{\mathsf{skip}}}$, and with ${\rho, \rho{^{\scriptscriptstyle\prime}}\models {{\mathsf{\mathcal{P}}}({\mathit{sq}}^p_i)}\sqsupseteq_{r}{{\mathsf{\mathcal{P}}}({\mathit{sq}}^q_i)}}$ we can conclude that ${{\mathsf{\mathcal{P}}}({\mathit{sq}}^q_i)}$ must be ${{\mathsf{skip}}}$ as well.
Case Study: Termination of Mutex {#S:PM3}
================================
Recall the model of the Peterson’s mutual exclusion algorithm shown in Figure \[fig:pm\]. The developments in Section \[S:PM1\] and Section \[S:PM2\] have implicitly utilised its property that the threads cannot be simultaneously in their critical sections to show that some specified task will be accomplished correctly upon termination. However, in this sense any other model that simply keeps one of the threads from entering its critical section would exhibit the same property and deviate from ${\mathsf{mutex}}$ only regarding liveness. Therefore, the question of termination will be addressed in this section, ultimately establishing correctness of ${\mathsf{mutex}}$.
Fairness is central to this end: it is not overly difficult to come up with a computation in ${\lsemantic{\mathsf{thread}}_0 \; {\mathit{cs}}_0 \parallel {\mathsf{thread}}_1 \; {\mathit{cs}}_1\rsemantic^\omega}$ that does not terminate without any environment steps whatsoever. For example, one may consider the following scenario: when both ${\mathbf{flag}}_0$ and ${\mathbf{flag}}_1$ have been set, ${\mathsf{thread}}_0$ also sets ${\mathbf{turn}}$, enters its ‘busy waiting’ phase, and is busy to such an extent that it keeps ${\mathsf{thread}}_1$ from performing ${\mathbf{turn}} := {\mathsf{False}}$. This could go on perpetually, violating the fairness condition, though. Hence, the goal of this section shall be to establish that any computation in ${\lsemantic{\mathsf{thread}}_0 \; {\mathit{cs}}_0 \parallel {\mathsf{thread}}_1 \; {\mathit{cs}}_1\rsemantic^{\scriptscriptstyle{\mathit{fair}}}} \cap {\mathcal{E}\!\mathit{nv}\:{\mathsf{id}}}$ will eventually terminate, provided that ${\mathit{cs}}_0$ and ${\mathit{cs}}_1$ terminate. Also for this task the auxiliary variables ${\mathbf{turn\_aux}}_0$ and ${\mathbf{turn\_aux}}_1$ are essential, and once more we will consider the model in Figure \[fig:pm-aux\] first.
According to the approach from the previous section, this goal can be achieved by a refutation of the statement $$\!\!\!\!\!\!\!\!\!\!\!\!\exists{\mathit{sq}}\in {\mathcal{E}\!\mathit{nv}\:{\mathsf{id}}}. \:{{\mathit{sq}}\Vdash_{\scriptscriptstyle{\mathcal N}} \neg{\mathbf{turn\_aux}}_0 \wedge \neg{\mathbf{turn\_aux}}_1 {\smalltriangleright}{\mathsf{thread\_aux}}_0\: {\mathit{cs}}_0 \parallel {\mathsf{thread\_aux}}_1\:{\mathit{cs}}_1 {\smalltriangleright}\bot}$$ Thus, for the remainder of this section, let ${{\mathit{sq}}^{\scriptscriptstyle{\mathit init}}}\in {\mathcal{E}\!\mathit{nv}\:{\mathsf{id}}}$ be a fixed computation such that $$\label{eq:pm32}
{{{\mathit{sq}}^{\scriptscriptstyle{\mathit init}}}\Vdash_{\scriptscriptstyle{\mathcal N}} \neg{\mathbf{turn\_aux}}_0 \wedge \neg{\mathbf{turn\_aux}}_1 {\smalltriangleright}{\mathsf{thread\_aux}}_0\: {\mathit{cs}}_0 \parallel {\mathsf{thread\_aux}}_1\:{\mathit{cs}}_1 {\smalltriangleright}\bot}$$
Sketching a rough refutation plan for (\[eq:pm32\]), application of the rule for the parallel composition (Proposition \[thm:parallel2-live\]) would lead to the three cases: either both, ${\mathsf{thread\_aux}}_0$ and ${\mathsf{thread\_aux}}_1$, do not terminate, or ${\mathsf{thread\_aux}}_0$ terminates but ${\mathsf{thread\_aux}}_1$ does not, or vice versa.
In the first case with non-terminating ${\mathsf{thread\_aux}}_0$ and ${\mathsf{thread\_aux}}_1$ we can infer that ${\mathbf{flag}}_1 \wedge {\mathbf{turn}}$ as well as ${\mathbf{flag}}_0 \wedge \neg{\mathbf{turn}}$ must hold infinitely often on ${{\mathit{sq}}^{\scriptscriptstyle{\mathit init}}}$. In particular, this means that the value of ${\mathbf{turn}}$ alternates *ad infinitum* on ${{\mathit{sq}}^{\scriptscriptstyle{\mathit init}}}$. Furthermore, from a certain point both, ${\mathbf{turn\_aux}}_0$ and ${\mathbf{turn\_aux}}_1$, will hold perpetually. Then taking also into account the ${\mathsf{mutex\_aux}}$ guarantee condition ${\mathbf{turn\_aux}}_0 \wedge {\mathbf{turn\_aux}}_1 \longrightarrow {\mathbf{turn}} = {\mathbf{turn}}{^{\scriptscriptstyle\prime}}$, from a certain point ${\mathbf{turn}}$ will remain constant, contradicting the previous conclusion that ${\mathbf{turn}}$ flips infinitely often.
In the second case, termination of ${\mathsf{thread\_aux}}_0$ entails that at some point ${\mathbf{turn\_aux}}_0$ has been set on ${{\mathit{sq}}^{\scriptscriptstyle{\mathit init}}}$ and, at some later point, ${\mathbf{flag}}_0$ has been reset. Thus, using the guarantees ${\mathbf{turn\_aux}}_0 \longrightarrow {\mathbf{turn\_aux}}{^{\scriptscriptstyle\prime}}_0$ and ${\mathbf{turn\_aux}}_0 \wedge {\mathbf{flag}}{^{\scriptscriptstyle\prime}}_0 \longrightarrow {\mathbf{flag}}_0$, we could infer that $\neg{\mathbf{flag}}_0$ holds perpetually starting from the point where it has been reset. On the other hand, non-termination of ${\mathsf{thread\_aux}}_1$ entails that the condition ${\mathbf{flag}}_0 \wedge \neg{\mathbf{turn}}$ holds infinitely often. Altogether, this would yield the existence of a state where both, $\neg{\mathbf{flag}}_0$ and ${\mathbf{flag}}_0$, hold.
To start rendering all that more accurately, consider the following extended Hoare triple, derivable using the program logic rules from Section \[S:prog-log\]: $$\label{eq:pm33}
\!\!\!\!\!\!\!\!\!\!\!{\models^\omega\{{\mathsf{id}},\:\neg{\mathbf{turn\_aux}}_0 \wedge \neg{\mathbf{turn\_aux}}_1\}\:{\mathsf{thread\_aux}}_0\: {\mathit{cs}}_0 \parallel {\mathsf{thread\_aux}}_1\: {\mathit{cs}}_1 \:\{\top,\:G_{{\mathit{global}}}\}}$$ provided ${\models\{\top,\:\top\}\:{\mathit{cs}}_0 \:\{\top,\:G_{{\mathit{cs}}}\}}$ and ${\models\{\top,\:\top\}\:{\mathit{cs}}_1 \:\{\top,\:G_{{\mathit{cs}}}\}}$ hold, where $G_{{\mathit{cs}}}$ abbreviates the guarantee condition $$\begin{array}{l}
{\mathbf{turn\_aux}}_0 ={\mathbf{turn\_aux}}{^{\scriptscriptstyle\prime}}_0 \wedge {\mathbf{turn\_aux}}_1 ={\mathbf{turn\_aux}}{^{\scriptscriptstyle\prime}}_1 \; \wedge \\
{\mathbf{flag}}_0 ={\mathbf{flag}}{^{\scriptscriptstyle\prime}}_0 \wedge {\mathbf{flag}}_1 ={\mathbf{flag}}{^{\scriptscriptstyle\prime}}_1 \wedge {\mathbf{turn}} ={\mathbf{turn}}{^{\scriptscriptstyle\prime}}\end{array}$$ which simply forces critical sections not to modify the auxiliary and the protocol variables. As opposed to Section \[S:PM1\], where guarantees of ${\mathsf{thread\_aux}}_0\: {\mathit{cs}}_0 \parallel {\mathsf{thread\_aux}}_1\: {\mathit{cs}}_1$ remained unspecified, in (\[eq:pm33\]) $G_{{\mathit{global}}}$ stands for $$\begin{array}{l}
({\mathbf{turn\_aux}}_0 \longrightarrow {\mathbf{turn\_aux}}{^{\scriptscriptstyle\prime}}_0) \; \wedge \; ({\mathbf{turn\_aux}}_1 \longrightarrow {\mathbf{turn\_aux}}{^{\scriptscriptstyle\prime}}_1) \; \wedge \\
({\mathbf{turn\_aux}}_0 \wedge {\mathbf{flag}}{^{\scriptscriptstyle\prime}}_0 \longrightarrow {\mathbf{flag}}_0) \; \wedge \; ({\mathbf{turn\_aux}}_1 \wedge {\mathbf{flag}}{^{\scriptscriptstyle\prime}}_1 \longrightarrow {\mathbf{flag}}_1) \; \wedge \\
({\mathbf{turn\_aux}}_0 \wedge {\mathbf{turn\_aux}}_1 \longrightarrow {\mathbf{turn}} = {\mathbf{turn}}{^{\scriptscriptstyle\prime}})
\end{array}$$ and contains the rules which ${\mathsf{thread\_aux}}_0\: {\mathit{cs}}_0 \parallel {\mathsf{thread\_aux}}_1\: {\mathit{cs}}_1$ follows when modifying the auxiliary and protocol variables. Now, ${{\mathit{sq}}^{\scriptscriptstyle{\mathit init}}}\in {\mathcal{P}\!\mathit{rog}\:G_{{\mathit{global}}}}$ is the essential consequence of ${{\mathit{sq}}^{\scriptscriptstyle{\mathit init}}}\in {\mathcal{E}\!\mathit{nv}\:{\mathsf{id}}}$, (\[eq:pm32\]) and (\[eq:pm33\]), such that the property ${{\mathit{sq}}^{\scriptscriptstyle{\mathit init}}}\in {\mathcal{E}\!\mathit{nv}\:{\mathsf{id}}} \cap {\mathcal{P}\!\mathit{rog}\:G_{{\mathit{global}}}}$ enables the following two propositions, where $\sigma_i$ denotes the state of the $i$-th configuration on ${{\mathit{sq}}^{\scriptscriptstyle{\mathit init}}}$, ${{\mathsf{\mathcal{S}}}({{\mathit{sq}}^{\scriptscriptstyle{\mathit init}}}_i)}$.
\[thm:PM3a\] Assume $\sigma_n{\mathbf{turn\_aux}}_0$ and $\sigma_m{\mathbf{turn\_aux}}_1$. Then there exists some $d$ such that $\sigma_i{\mathbf{turn}} = \sigma_j{\mathbf{turn}}$ holds for all $i \ge d$ and $j \ge d$.
Let $d$ be the maximum of $n$ and $m$. Due to the guarantees ${\mathbf{turn\_aux}}_0 \longrightarrow {\mathbf{turn\_aux}}{^{\scriptscriptstyle\prime}}_0$ and ${\mathbf{turn\_aux}}_1 \longrightarrow {\mathbf{turn\_aux}}{^{\scriptscriptstyle\prime}}_1$ we have $\sigma_i{\mathbf{turn\_aux}}_0 \wedge \sigma_i{\mathbf{turn\_aux}}_1$ for all $i\ge d$. Furthermore, ${\mathbf{turn\_aux}}_0 \wedge {\mathbf{turn\_aux}}_1 \longrightarrow {\mathbf{turn}} = {\mathbf{turn}}{^{\scriptscriptstyle\prime}}$ implies $\sigma_i{\mathbf{turn}} = \sigma_{i+1}{\mathbf{turn}}$ for all $i\ge d$, which induces $\sigma_i{\mathbf{turn}} = \sigma_{j}{\mathbf{turn}}$ for all $i \ge d$ and $j \ge d$.
\[thm:PM3b\] Assume $\sigma_m{\mathbf{turn\_aux}}_0$ and $\neg\sigma_n{\mathbf{flag}}_0$ with $m \le n$. Then $\sigma_i{\mathbf{turn\_aux}}_0$ and $\neg\sigma_i{\mathbf{flag}}_0$ hold for all $i \ge n$.
The assumption $\sigma_m{\mathbf{turn\_aux}}_0$ and the guarantee ${\mathbf{turn\_aux}}_0 \longrightarrow {\mathbf{turn\_aux}}{^{\scriptscriptstyle\prime}}_0$ induce $\sigma_i{\mathbf{turn\_aux}}_0$ for all $i\ge m$, and for all $i\ge n$ in particular. Then $\neg\:\sigma_n{\mathbf{flag}}_0$ and the guarantee ${\mathbf{turn\_aux}}_0 \wedge {\mathbf{flag}}{^{\scriptscriptstyle\prime}}_0 \longrightarrow {\mathbf{flag}}_0$ additionally induce $\neg\sigma_i{\mathbf{flag}}_0$ for all $i\ge n$.
Regarding critical sections, we will assume termination of ${\mathit{cs}}_0$ and ${\mathit{cs}}_1$: each of the statements $$\label{eq:pm3-asm1}
{{\mathit{sq}}\Vdash_{\scriptscriptstyle{\mathcal N}} P {\smalltriangleright}{\mathit{cs}}_0 {\smalltriangleright}\bot}$$ and $$\label{eq:pm3-asm2}
{{\mathit{sq}}\Vdash_{\scriptscriptstyle{\mathcal N}} P {\smalltriangleright}{\mathit{cs}}_1 {\smalltriangleright}\bot}$$ shall lead to a contradiction for any ${\mathit{sq}}$ and $P$.
With all these preparations we can advance to processing (\[eq:pm32\]) by successive application of the rules from the previous section. In the first step, Proposition \[thm:parallel2-live\] yields two computations ${\mathit{sq}}^0$ with ${{{\mathit{sq}}^{\scriptscriptstyle{\mathit init}}}\sim {\mathit{sq}}^0}$ and ${\mathit{sq}}^1$ with ${{{\mathit{sq}}^{\scriptscriptstyle{\mathit init}}}\sim {\mathit{sq}}^1}$ such that either $$\label{eq:pm34}
{{\mathit{sq}}^0 \Vdash_{\scriptscriptstyle{\mathcal N}} \neg{\mathbf{turn\_aux}}_0 \wedge \neg{\mathbf{turn\_aux}}_1 {\smalltriangleright}{\mathsf{thread\_aux}}_0\: {\mathit{cs}}_0 {\smalltriangleright}\bot}$$ and $$\label{eq:pm35}
{{\mathit{sq}}^1 \Vdash_{\scriptscriptstyle{\mathcal N}} \neg{\mathbf{turn\_aux}}_0 \wedge \neg{\mathbf{turn\_aux}}_1 {\smalltriangleright}{\mathsf{thread\_aux}}_1\: {\mathit{cs}}_1 {\smalltriangleright}\bot}$$ or $$\label{eq:pm36}
{{\mathit{sq}}^0 \Vdash_{\scriptscriptstyle{\mathcal T}} \neg{\mathbf{turn\_aux}}_0 \wedge \neg{\mathbf{turn\_aux}}_1 {\smalltriangleright}{\mathsf{thread\_aux}}_0\: {\mathit{cs}}_0 {\smalltriangleright}\bot {\smalltriangleright}n}$$ and $$\label{eq:pm37}
{{\mathit{sq}}^1 \Vdash_{\scriptscriptstyle{\mathcal N}} \neg{\mathbf{turn\_aux}}_0 \wedge \neg{\mathbf{turn\_aux}}_1 {\smalltriangleright}{\mathsf{thread\_aux}}_1\: {\mathit{cs}}_1 {\smalltriangleright}\bot}$$ hold with some $n$. Note once more that the third case is symmetric to (\[eq:pm36\]) and (\[eq:pm37\]) and therefore not treated here.
Processing (\[eq:pm34\]) and (\[eq:pm35\]) leads to several trivially refutable branches with statements such as $${{\mathit{sq}}{^{\scriptscriptstyle\prime}}\Vdash_{\scriptscriptstyle{\mathcal N}} \neg{\mathbf{turn\_aux}}_0 \wedge \neg{\mathbf{turn\_aux}}_1 {\smalltriangleright}{\mathbf{flag}}_0 := {\mathsf{True}} {\smalltriangleright}\bot} \mbox{, or}$$ $${{\mathit{sq}}{^{\scriptscriptstyle\prime\scriptscriptstyle\prime}}\Vdash_{\scriptscriptstyle{\mathcal N}} \top {\smalltriangleright}\langle {\mathbf{turn}} := {\mathsf{True}};{\mathbf{turn\_aux}}_0 := {\mathsf{True}}\rangle {\smalltriangleright}\bot}$$ for some ${\mathit{sq}}{^{\scriptscriptstyle\prime}}$ and ${\mathit{sq}}{^{\scriptscriptstyle\prime\scriptscriptstyle\prime}}$. Also notice that all of the subcases claiming terminating ${\mathsf{while}}$-statements are similarly simply refutable since none of the subsequent code can be non-terminating due to (\[eq:pm3-asm1\]) and (\[eq:pm3-asm2\]). The remaining essential case comprises
1. some $m_0, m_1$ with $\sigma_{m_0}{\mathbf{turn\_aux}}_0$ and $\sigma_{m_1}{\mathbf{turn\_aux}}_1$,
2. two strictly ascending sequences of natural numbers $\phi, \psi$ and some $d_0>m_0, d_1>m_1$ such that $\sigma_{\phi_{i} + d_0}{\mathbf{flag}}_1 \wedge \sigma_{\phi_{i} + d_0}{\mathbf{turn}}$ and $\sigma_{\psi_{i} + d_1}{\mathbf{flag}}_0 \wedge \neg\sigma_{\psi_{i} + d_1}{\mathbf{turn}}$ hold for all $i \in \mathbb{N}$,
where we continue to write $\sigma_i$ for the state of the $i$-th configuration on ${{\mathit{sq}}^{\scriptscriptstyle{\mathit init}}}$. Then Proposition \[thm:PM3a\] and (a) imply that there exists some $d$ with $\sigma_i{\mathbf{turn}} = \sigma_j{\mathbf{turn}}$ for all $i, j\ge d$. In particular, we have $\sigma_{\phi_{d} + d_0}{\mathbf{turn}} = \sigma_{\psi_{d} + d_1}{\mathbf{turn}}$, since $\phi_{d} + d_0 \ge d$ and $\psi_{d} + d_1 \ge d$, whereas $\sigma_{\phi_{d} + d_0}{\mathbf{turn}}$ and $\neg\sigma_{\psi_{d} + d_1}{\mathbf{turn}}$ follow from (b).
Next, backtracking to the case with (\[eq:pm36\]) and (\[eq:pm37\]), the termination assumption (\[eq:pm36\]) provides $\neg\sigma_{n}{\mathbf{flag}}_0$ as well as some $m<n$ with $\sigma_{m}{\mathbf{turn\_aux}}_0$. Then Proposition \[thm:PM3b\] yields $\neg\sigma_i{\mathbf{flag}}_0$ for all $i\ge n$. On the other hand, (\[eq:pm37\]) entails some strictly ascending sequence $\phi$ and some $d$ with $\sigma_{\phi_{i} + d}{\mathbf{flag}}_0 \wedge \neg\sigma_{\phi_{i} + d}{\mathbf{turn}}$ for all $i \in \mathbb{N}$. Thus, we can derive both, $\sigma_{\phi_{n} + d}{\mathbf{flag}}_0$ and $\neg\sigma_{\phi_{n} + d}{\mathbf{flag}}_0$, since $\phi_{n} + d \ge n$.
The following proposition summarises what has been derived so far.
\[thm:PM3-aux\] Assume
1. ${\models\{\top,\:\top\}\:{\mathit{cs}}_0 \:\{\top,\:G_{{\mathit{cs}}}\}}$,
2. ${\models\{\top,\:\top\}\:{\mathit{cs}}_1 \:\{\top,\:G_{{\mathit{cs}}}\}}$,
3. ${{\mathit{sq}}\Vdash_{\scriptscriptstyle{\mathcal N}} P {\smalltriangleright}{\mathit{cs}}_0 {\smalltriangleright}\bot}$ leads to a contradiction for any ${\mathit{sq}}$ and $P$,
4. ${{\mathit{sq}}\Vdash_{\scriptscriptstyle{\mathcal N}} P {\smalltriangleright}{\mathit{cs}}_1 {\smalltriangleright}\bot}$ leads to a contradiction for any ${\mathit{sq}}$ and $P$.
Then any ${\mathit{sq}}\in {\lsemantic{\mathsf{thread\_aux}}_0\: {\mathit{cs}}_0 \parallel {\mathsf{thread\_aux}}_1\:{\mathit{cs}}_1\rsemantic^{\scriptscriptstyle{\mathit{fair}}}} \cap {\mathcal{E}\!\mathit{nv}\:{\mathsf{id}}}$, additionally satisfying the input condition ${\mathcal{I}\!\mathit{n}\:\!(\neg{\mathbf{turn\_aux}}_0 \wedge \neg{\mathbf{turn\_aux}}_1)}$, eventually reaches a ${{\mathsf{skip}}}$-configuration.
Removing the auxiliaries
------------------------
As in the first part of the case study in Section \[S:PM1\], the actual goal of this section is to prove liveness of the mutual exclusion model that does not make use of auxiliary variables, ${\mathsf{mutex}}$ shown in Figure \[fig:pm\]. Consequently, the correspondences ${\models {\mathsf{thread\_aux}}_0 \; {\mathit{cs}}_0\;\sqsupseteq_{{\mathsf{r}}_{{\mathsf{eqv}}}}\;{\mathsf{thread}}_0 \; {\mathit{cs}}_0}$ and ${\models {\mathsf{thread\_aux}}_1 \; {\mathit{cs}}_1\;\sqsupseteq_{{\mathsf{r}}_{{\mathsf{eqv}}}}\;{\mathsf{thread}}_1 \; {\mathit{cs}}_1}$, established in the proof of Proposition \[thm:mutex\] can be reused to show that ${\mathsf{thread\_aux}}_0 \parallel {\mathsf{thread\_aux}}_1$ corresponds componentwise to ${\mathsf{thread}}_0 \parallel {\mathsf{thread}}_1$ since ${\mathsf{thread\_aux}}_0$ and ${\mathsf{thread\_aux}}_1$ are sequential, whereas ${\mathsf{thread}}_0$ and ${\mathsf{thread}}_1$ – non-blocking, provided ${\mathit{cs}}_0$ and ${\mathit{cs}}_1$ are both: sequential and non-blocking. The proof of the following proposition utilises the componentwise correspondence.
\[thm:PM3\] Assume
1. ${\models\{\top,\:\top\}\:{\mathit{cs}}_0 \:\{\top,\:G_{{\mathit{cs}}}\}}$,
2. ${\models\{\top,\:\top\}\:{\mathit{cs}}_1 \:\{\top,\:G_{{\mathit{cs}}}\}}$,
3. ${{\mathit{sq}}\Vdash_{\scriptscriptstyle{\mathcal N}} P {\smalltriangleright}{\mathit{cs}}_0 {\smalltriangleright}\bot}$ leads to a contradiction for any ${\mathit{sq}}$ and $P$,
4. ${{\mathit{sq}}\Vdash_{\scriptscriptstyle{\mathcal N}} P {\smalltriangleright}{\mathit{cs}}_1 {\smalltriangleright}\bot}$ leads to a contradiction for any ${\mathit{sq}}$ and $P$,
5. ${\models {\mathit{cs}}_0\sqsupseteq_{{\mathsf{r}}_{{\mathsf{eqv}}}}{\mathit{cs}}_0}$,
6. ${\models {\mathit{cs}}_1\sqsupseteq_{{\mathsf{r}}_{{\mathsf{eqv}}}}{\mathit{cs}}_1}$,
7. ${\mathit{cs}}_0$ and ${\mathit{cs}}_1$ are sequential and non-blocking.
Then any ${\mathit{sq}}\in {\lsemantic{\mathsf{thread}}_0\: {\mathit{cs}}_0 \parallel {\mathsf{thread}}_1\:{\mathit{cs}}_1\rsemantic^{\scriptscriptstyle{\mathit{fair}}}} \cap {\mathcal{E}\!\mathit{nv}\:{\mathsf{id}}}$ eventually reaches a ${{\mathsf{skip}}}$-configuration.
Using the assumptions (1)–(4), from Proposition \[thm:PM3-aux\] follows that any computation ${\mathit{sq}}\in {\lsemantic{\mathsf{thread\_aux}}_0\: {\mathit{cs}}_0 \parallel {\mathsf{thread\_aux}}_1\:{\mathit{cs}}_1\rsemantic^{\scriptscriptstyle{\mathit{fair}}}} \cap {\mathcal{E}\!\mathit{nv}\:{\mathsf{id}}} \cap {\mathcal{I}\!\mathit{n}\:\!(\neg{\mathbf{turn\_aux}}_0 \wedge \neg{\mathbf{turn\_aux}}_1)}$ eventually reaches a ${{\mathsf{skip}}}$-configuration. Thus, in order to apply Corollary \[thm:fair-corr-sim3\] it remains to show $\top \subseteq {{\mathsf{r}}_{{\mathsf{eqv}}}\cdot (\neg{\mathbf{turn\_aux}}_0 \wedge \neg{\mathbf{turn\_aux}}_1)}$, for any $\sigma$ there must be $\sigma{^{\scriptscriptstyle\prime}}$ with $\neg\sigma{^{\scriptscriptstyle\prime}}{\mathbf{turn\_aux}}_0$, $\neg\sigma{^{\scriptscriptstyle\prime}}{\mathbf{turn\_aux}}_1$ and $(\sigma{^{\scriptscriptstyle\prime}}, \sigma) \in{\mathsf{r}}_{{\mathsf{eqv}}}$. This state is $\sigma_{[{\mathbf{turn\_aux}}_0 := {\mathsf{False}}, {\mathbf{turn\_aux}}_1 := {\mathsf{False}}]}$.
Finally, instantiating critical sections ${\mathit{cs}}_0$ and ${\mathit{cs}}_1$ in Proposition \[thm:PM3\] by ${\mathsf{shared\_update}}_0$ and ${\mathsf{shared\_update}}_1$, defined in Section \[sub:mutex-inst\], leads to the conclusion that any computation ${\mathit{sq}}\in {\lsemantic{\mathsf{mutex}} \;{\mathsf{shared\_update}}_0 \; {\mathsf{shared\_update}}_1\rsemantic^{\scriptscriptstyle{\mathit{fair}}}} \cap {\mathcal{E}\!\mathit{nv}\:{\mathsf{id}}}$ reaches a ${{\mathsf{skip}}}$-configuration.
Dividing Indivisible Steps {#S:split}
==========================
Program correspondences have been defined stepwise such that it is possible to establish connections like ${\models {{\mathsf{basic}}}\:{\mathsf{t}}_1\sqsupseteq_{}{{\mathsf{basic}}}\:({\mathsf{t}}_3 \circ {\mathsf{t}}_2)}$ with some state transformers ${\mathsf{t}}_1, {\mathsf{t}}_2, {\mathsf{t}}_3$, but impossible to extend the chain to ${{\mathsf{basic}}}\;{\mathsf{t}}_2 ; {{\mathsf{basic}}}\;{\mathsf{t}}_3$.
Taking the particular perspective of the rely/guarantee program logic, one can explain that by means of Proposition \[thm:pcorr-rule\]: the program correspondence rule would imply that any property ${\models\{R,\:P\}\:{{\mathsf{basic}}}\:({\mathsf{t}}_3 \circ {\mathsf{t}}_2)\:\{Q,\:G\}}$ is also a property of ${{\mathsf{basic}}}\;{\mathsf{t}}_2 ; {{\mathsf{basic}}}\;{\mathsf{t}}_3$ if assuming ${\models {{\mathsf{basic}}}\:({\mathsf{t}}_3 \circ {\mathsf{t}}_2)\sqsupseteq_{}{{\mathsf{basic}}}\:{\mathsf{t}}_2 ; {{\mathsf{basic}}}\:{\mathsf{t}}_3}$. This cannot be the case in general, since ${{\mathsf{basic}}}\;{\mathsf{t}}_2 ; {{\mathsf{basic}}}\;{\mathsf{t}}_3$ could be subject to more interferences in first place. Moreover, that ${{\mathsf{basic}}}\:({\mathsf{t}}_3 \circ {\mathsf{t}}_2)$ complies with the guarantee $G$ does not imply that ${{\mathsf{basic}}}\;{\mathsf{t}}_2$ and ${{\mathsf{basic}}}\;{\mathsf{t}}_3$ do, meaning that not only anything that might run in parallel to a thread containing ${{\mathsf{basic}}}\:({\mathsf{t}}_3 \circ {\mathsf{t}}_2)$ can potentially affect its behaviours if ${{\mathsf{basic}}}\:({\mathsf{t}}_3 \circ {\mathsf{t}}_2)$ becomes replaced by ${{\mathsf{basic}}}\;{\mathsf{t}}_2 ; {{\mathsf{basic}}}\;{\mathsf{t}}_3$, but also behaviours of anything that might run in parallel to the thread could be affected too. Nonetheless, quite frequently many properties can be retained replacing some ${{\mathsf{basic}}}\:(g \circ f)$ by ${{\mathsf{basic}}}\:f; {{\mathsf{basic}}}\:g$ across a program, and the goal of the section is, in principle, to give sufficient conditions for that.
Regarding extended Hoare triples, an approach can be to focus on their derivations for sequential program components by means of the rules presented in Section \[S:prog-log\]. More precisely, let ${\vdash_{\scriptscriptstyle{\mathit{seq}}}\{R,\:P\}\:p\:\{Q,\:G\}}$ be the smallest set that is closed under the rules of Section \[S:prog-log\], excluding the parallel composition rule (Proposition \[thm:parallel-rule\]). As we want to keep the focus on the essentials in this section, suppose that the program correspondence rule (Proposition \[thm:pcorr-rule\]) is omitted as well, with the sole aim to make the induction and the rule inversion for ${\vdash_{\scriptscriptstyle{\mathit{seq}}}\{R,\:P\}\:p\:\{Q,\:G\}}$ simpler.
This construction exhibits the following two key properties. First, any derivation ${\vdash_{\scriptscriptstyle{\mathit{seq}}}\{R,\:P\}\:p\:\{Q,\:G\}}$, having ${\vdash_{\scriptscriptstyle{\mathit{seq}}}\{R{^{\scriptscriptstyle\prime}},\:P{^{\scriptscriptstyle\prime}}\}\:q\:\{Q{^{\scriptscriptstyle\prime}},\:G{^{\scriptscriptstyle\prime}}\}}$ as a subderivation, can be replayed with ${\vdash_{\scriptscriptstyle{\mathit{seq}}}\{R{^{\scriptscriptstyle\prime}},\:P{^{\scriptscriptstyle\prime}}\}\:q{^{\scriptscriptstyle\prime}}\:\{Q{^{\scriptscriptstyle\prime}},\:G{^{\scriptscriptstyle\prime}}\}}$ in place of ${\vdash_{\scriptscriptstyle{\mathit{seq}}}\{R{^{\scriptscriptstyle\prime}},\:P{^{\scriptscriptstyle\prime}}\}\:q\:\{Q{^{\scriptscriptstyle\prime}},\:G{^{\scriptscriptstyle\prime}}\}}$, provided the latter implies the former. This would yield a derivation of ${\vdash_{\scriptscriptstyle{\mathit{seq}}}\{R,\:P\}\:p{^{\scriptscriptstyle\prime}}\:\{Q,\:G\}}$ where $p{^{\scriptscriptstyle\prime}}$ is obtained by replacing $q$ by $q{^{\scriptscriptstyle\prime}}$ within $p$. Second, the rely $R$ and the guarantee $G$ remain fixed across any derivation of ${\vdash_{\scriptscriptstyle{\mathit{seq}}}\{R,\:P\}\:p\:\{Q,\:G\}}$, unless an ${\mathsf{await}}$-statement occurs in $p$. In such cases, however, the rule for ${\mathsf{await}}$ (Proposition\[thm:await-rule\]) merely alters $R$ and $G$ to $\bot$ and $\top$, respectively. Therefore, if we make global assumptions about $R$ and $G$, which particularly hold with $\bot$ and $\top$ in place of $R$ and $G$, then these would apply to any subderivation of ${\vdash_{\scriptscriptstyle{\mathit{seq}}}\{R,\:P\}\:p\:\{Q,\:G\}}$. The conditions (1) and (2) in the following proposition, where ${f^{\mathcal G}}$ denotes the graph of a function $f$, act as such global assumptions.
Assume ${\vdash_{\scriptscriptstyle{\mathit{seq}}}\{R,\:P\}\:{{\mathsf{basic}}}\:(g \circ f)\:\{Q,\:G\}}$ and
1. ${{f^{\mathcal G}} \diamond R} \subseteq {R \diamond {f^{\mathcal G}}}$,
2. $(\sigma, f\:\sigma) \in G$ and $(f\:\sigma, (g \circ f)\:\sigma) \in G$ for any state $\sigma$ with $(\sigma, (g \circ f)\:\sigma) \in G$.
Then ${\vdash_{\scriptscriptstyle{\mathit{seq}}}\{R,\:P\}\:{{\mathsf{basic}}}\:f ; {{\mathsf{basic}}}\:g\:\{Q,\:G\}}$.
Note that ${\vdash_{\scriptscriptstyle{\mathit{seq}}}\{R,\:P\}\:{{\mathsf{basic}}}\:(g \circ f)\:\{Q,\:G\}}$ implies
1. ${R\cdot P} \subseteq P$ and
2. $P \subseteq \{\sigma \:|\: (g \circ f)\:\sigma \in Q \wedge (\sigma, (g \circ f)\:\sigma) \in G \}$
by inversion of the rule for ${{\mathsf{basic}}}$ (Proposition \[thm:basic-rule\]).
As the first step, we show ${\vdash_{\scriptscriptstyle{\mathit{seq}}}\{R,\:P\}\:{{\mathsf{basic}}}\:f\:\{{{f^{\mathcal G}}\cdot P},\:G\}}$. To this end, notice that $(\sigma, f\:\sigma) \in G$ holds for any $\sigma \in P$, since for any such $\sigma$ we can obtain $(\sigma, (g \circ f)\:\sigma) \in G$ from (4), and then $(\sigma, f\:\sigma) \in G$ by (2). Thus, ${\vdash_{\scriptscriptstyle{\mathit{seq}}}\{R,\:P\}\:{{\mathsf{basic}}}\:f\:\{{{f^{\mathcal G}}\cdot P},\:G\}}$ follows from Proposition \[thm:basic-rule\] and (3).
Second, we show ${\vdash_{\scriptscriptstyle{\mathit{seq}}}\{R,\:{{f^{\mathcal G}}\cdot P}\}\:{{\mathsf{basic}}}\:g\:\{Q,\:G\}}$. In that case, however, in order to apply Proposition \[thm:basic-rule\] we also need to establish the stability condition ${R\cdot ({{f^{\mathcal G}}\cdot P})} \subseteq {{f^{\mathcal G}}\cdot P}$, which holds since ${R\cdot ({{f^{\mathcal G}}\cdot P})} = {({{f^{\mathcal G}} \diamond R})\cdot P} \subseteq^{\mbox{\tiny by (1)}} {({R \diamond {f^{\mathcal G}}})\cdot P} = {{f^{\mathcal G}}\cdot ({R\cdot P})} \subseteq^{\mbox{\tiny by (3)}} {{f^{\mathcal G}}\cdot P}$. Furthermore, $(f\:\sigma, (g \circ f)\:\sigma) \in G$ holds for any $\sigma \in P$, since we first derive $(\sigma, (g \circ f)\:\sigma) \in G$ from (4) and then $(f\:\sigma, (g \circ f)\:\sigma) \in G$ from (2).
Altogether, with ${{f^{\mathcal G}}\cdot P}$ as $S$ in the rule for the sequential composition (Proposition \[thm:seq-rule\]) we can infer ${\vdash_{\scriptscriptstyle{\mathit{seq}}}\{R,\:P\}\:{{\mathsf{basic}}}\:f ; {{\mathsf{basic}}}\:g\:\{Q,\:G\}}$, since ${{\mathsf{basic}}}\:g \neq {{\mathsf{skip}}}$.
As a result, replacing any ${{\mathsf{basic}}}\:(g \circ f)$ by ${{\mathsf{basic}}}\:f ; {{\mathsf{basic}}}\:g$ across $p$ yields a program $p{^{\scriptscriptstyle\prime}}$ such that in any derivation ${\vdash_{\scriptscriptstyle{\mathit{seq}}}\{R,\:P\}\:p\:\{Q,\:G\}}$ we can also replace any subderivation ${\vdash_{\scriptscriptstyle{\mathit{seq}}}\{R,\:P{^{\scriptscriptstyle\prime}}\}\:{{\mathsf{basic}}}\:(g \circ f)\:\{Q{^{\scriptscriptstyle\prime}},\:G\}}$ by ${\vdash_{\scriptscriptstyle{\mathit{seq}}}\{R,\:P{^{\scriptscriptstyle\prime}}\}\:{{\mathsf{basic}}}\:f ; {{\mathsf{basic}}}\:g\:\{Q{^{\scriptscriptstyle\prime}},\:G\}}$ in order to obtain ${\vdash_{\scriptscriptstyle{\mathit{seq}}}\{R,\:P\}\:p{^{\scriptscriptstyle\prime}}\:\{Q,\:G\}}$, provided $R, G, f$ and $g$ satisfy the conditions (1) and (2) of the above statement. Note that since we can infer ${\models\{R,\:P\}\:p\:\{Q,\:G\}}$ from ${\vdash_{\scriptscriptstyle{\mathit{seq}}}\{R,\:P\}\:p\:\{Q,\:G\}}$ but not the other way around, in order to utilise this method, transformational developments would have to focus on sequential components in separate first, and assemble these by means of the parallel composition only when all necessary splits have been done.
Shifting now the attention to liveness properties, such splits can produce different effects. In the simplest situations when $p$ does not contain any ${\mathsf{while}}$- and ${\mathsf{await}}$-statements and therefore terminates on all inputs, any split in $p$ will clearly retain this property. For instance, such situation was present in the previous section with ${\mathsf{shared\_update}}_0$ and ${\mathsf{shared\_update}}_1$. Regardless which splits will be performed to obtain some ${\mathsf{shared\_update}}{^{\scriptscriptstyle\prime}}_0$ and ${\mathsf{shared\_update}}{^{\scriptscriptstyle\prime}}_1$, the statement ${{\mathit{sq}}\Vdash_{\scriptscriptstyle{\mathcal N}} P {\smalltriangleright}{\mathsf{shared\_update}}{^{\scriptscriptstyle\prime}}_0 {\smalltriangleright}\bot}$ would still lead to a contradiction for any ${\mathit{sq}}$ and $P$.
Generally, however, that ${{\mathit{sq}}\Vdash_{\scriptscriptstyle{\mathcal N}} P {\smalltriangleright}p {\smalltriangleright}\bot}$ yields a contradiction for any ${\mathit{sq}}$ does not need to imply that ${{\mathit{sq}}{^{\scriptscriptstyle\prime}}\Vdash_{\scriptscriptstyle{\mathcal N}} P {\smalltriangleright}p{^{\scriptscriptstyle\prime}}{\smalltriangleright}\bot}$ can be refuted as well. For instance, refutation of the statement (\[eq:pm32\]) in the previous section was in particular based on the guarantees $G_{{\mathit{global}}}$ derived via the extended Hoare triple (\[eq:pm33\]). As described above, splitting some atomic step in ${\mathsf{thread\_aux}}_0\: {\mathit{cs}}_0 \parallel {\mathsf{thread\_aux}}_1\: {\mathit{cs}}_1$ does not need to retain the property (\[eq:pm33\]). But without this argument, termination of the transformed model would, in turn, become questionable.
Conclusion and Outlook {#S:concl}
======================
The paper gave a comprehensive presentation of a framework, built as a conservative extension to the simply typed higher-order logic and geared towards modelling, verification and transformation of concurrent imperative programs. The essential points were:
- a concise computational model encompassing fine-grained interleaving, state abstraction and jumps;
- stepwise program correspondence relations, in particular capturing program equivalence and providing transformations to low-level representations;
- a Hoare-style rely/guarantee program logic, independent from the type of the underlying states, and featuring the program correspondence rule as a generalised rule of consequence;
- a light-weight extension of the logic to enable state relations in place of postconditions;
- a notion of fair computations;
- a refutational approach to verification of basic liveness properties using the fairness notion;
- and, last but not least, conditions that enable splits of indivisible steps without loss of program properties, complementing transformations by program correspondence.
Many of the presented methods could be enhanced. This applies in first place to verification of liveness properties. A challenging task would be to extend the notion of fairness towards fairer treatment of ${\mathsf{await}}$-statements in general and not only in the very special case of atomic sections. As pointed out in Section \[Sb:faircomp\], this would demand lots of care in order to avert inconspicuous reasoning with non-existent computations. The refutational approach could also be further refined to address more sophisticated questions such as whether fair computations not just reach certain states, but visit these infinitely often. Furthermore, the approach could be enhanced towards more modularity as well as more advanced reasoning about effects of atomicity splits, as sketched in the previous section.
The case study described a verification process of a model of the Peterson’s mutual exclusion algorithm, featuring abstraction from the actual contents of critical sections. It surely did not aim at proclaiming correctness of something that is well-known to be correct anyway, but to illustrate how the framework can assist in proving properties of intricate models with interleaved computations. Apart from that, it also attempted to highlight the logical principles behind the algorithm that achieves mutual exclusion and termination in a remarkable way without any restrictions on interleaving, save fairness, of course.
What the case study might also have conveyed is how entangled the behaviours of programs running in parallel can become. And maybe a concept or a line of argumentation, presented in the course of the paper, could contribute to the development of provably safe and secure concurrent software systems.
[KPRS01]{}
J-R. Abrial. . Cambridge University Press, 2010.
F. Baader and T. Nipkow. . Cambridge University Press, 1998.
M. J. C. Gordon. Introduction to the [HOL]{} system. In Myla Archer, Jeffrey J. Joyce, Karl N. Levitt, and Phillip J. Windley, editors, [*Proceedings of the 1991 International Workshop on the [HOL]{} Theorem Proving System and its Applications, August 1991, Davis, California, [USA]{}*]{}, pages 2–3. [IEEE]{} Computer Society, 1991.
J. Harrison. light: An overview. In Stefan Berghofer, Tobias Nipkow, Christian Urban, and Makarius Wenzel, editors, [*Theorem Proving in Higher Order Logics, 22nd International Conference, TPHOLs 2009, Munich, Germany, August 17-20, 2009. Proceedings*]{}, volume 5674 of [*Lecture Notes in Computer Science*]{}, pages 60–66. Springer, 2009.
C. A. R. Hoare. An axiomatic basis for computer programming. , 12(10):576–580, October 1969.
Y. Kesten, A. Pnueli, L. Raviv, and E. Shahar. Model checking with strong fairness. Technical report, ISR, 2001.
O. Kupferman and M. Y. Vardi. Verification of fair transition systems. In Rajeev Alur and Thomas A. Henzinger, editors, [*Computer Aided Verification*]{}, pages 372–382, Berlin, Heidelberg, 1996. Springer Berlin Heidelberg.
R. Milner. Implementation and applications of Scott’s logic for computable functions. , 7(1):1–6, January 1972.
R. Milner. The use of machines to assist in rigorous proof. , 312(1522):411–422, 1984.
L. P. Nieto. The rely-guarantee method in Isabelle/HOL. In Pierpaolo Degano, editor, [*Programming Languages and Systems*]{}, pages 348–362, Berlin, Heidelberg, 2003. Springer Berlin Heidelberg.
T. Nipkow, L. Paulson, and M. Wenzel. , volume 2283 of [*Lecture Notes in Computer Science*]{}. Springer, 2002.
S. Owicki and D. Gries. An axiomatic proof technique for parallel programs. , 6:319–340, 1976.
G. L. Peterson. Myths about the mutual exclusion problem. , 12(3):115 – 116, 1981.
G. D. Plotkin. The origins of structural operational semantics. , 60-61:3 – 15, 2004. Structural Operational Semantics.
Radio Technical Commission for Aeronautics (RTCA). , 2011.
N. Schirmer. . PhD thesis, Technische Universität München, 2006.
C. Stirling. A generalization of [O]{}wicki-[G]{}ries’s [H]{}oare logic for a concurrent while language. , 58(1):347 – 359, 1988.
R. van Glabbeek. Ensuring liveness properties of distributed systems: Open problems. , 109:100480, 2019.
S. van Staden. On rely-guarantee reasoning. In Ralf Hinze and Janis Voigtl[ä]{}nder, editors, [*Mathematics of Program Construction - 12th International Conference, [MPC]{} 2015, K[ö]{}nigswinter, Germany, June 29 - July 1, 2015. Proceedings*]{}, volume 9129 of [*Lecture Notes in Computer Science*]{}, pages 30–49. Springer, 2015.
[^1]: Research partly supported by the ESF funded Estonian IT Academy measure 2014-2020.4.05.19-0001.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this work we consider general fermion systems in two spatial dimensions, both with and without charge conservation symmetry, which realize a non-trivial fermionic topological order with only Abelian anyons. We address the question of precisely how these quantum phases differ from their bosonic counterparts, both in terms of their edge physics and in the way one would identify them in numerics. As in previous works, we answer these questions by studying the theory obtained after gauging the global fermion parity symmetry, which turns out to have a special and simple structure. Using this structure, a minimal scheme is outlined for how to numerically identify a general Abelian fermionic topological order, without making use of fermion number conservation. Along the way, some subtleties of the momentum polarization technique are discussed. Regarding the edge physics, it is shown that the gauged theory, which is bosonic, can have a gapped boundary to the vacuum if and only if the ungauged fermion theory does as well.'
address: 'Department of Physics, University of California, Berkeley, CA 94720, USA'
author:
- Nick Bultinck
bibliography:
- 'bibliography.bib'
title: Numerical identification and gapped boundaries of Abelian fermionic topological order
---
November 2019
Introduction
============
Since the experimental discovery of the fractional quantum Hall effect, a tremendous effort has been devoted to the study of the possible topologically ordered phases that can be realized in strongly correlated quantum many-body systems. In the past decade, substantial progress has been made in the characterization of topologically ordered phases in spin or boson systems, and a unified algebraic framework called ‘modular tensor categories’ has been identified to describe these phases [@Kitaev]. From a physical point of view, a modular tensor category simply describes how the anyonic excitations fuse and braid with each other. Phrasing the properties of topologically ordered phases in the rigid algebraic language of modular tensor categories has allowed theorists to make significant progress in the study of bosonic topological phases.
Not only has our theoretical understanding of topologically ordered phases vastly improved, in recent years many important results have been obtained on how to numerically identify the type of topological order realized by a particular microscopic Hamiltonian. For example, it was realized that a non-trivial topological order leaves an imprint on the entanglement entropy of a spatial region in the ground state via the ‘topological entanglement entropy’ term [@KitaevPreskill; @LevinWen]. A later refinement of the topological entanglement entropy showed that the complete spectrum of the reduced density matrix of a spatial region contains information about the universal edge physics of the topological phase [@LiHaldane]. When the system of interest is put on a torus, it will necessarily have a ground state degeneracy if a a non-trivial topological order is realized [@WenNiu]. In Refs. [@ZhangGrover; @Cincio], a useful basis for the ground state subspace, the so-called Minimally-Entangled State (MES) basis, was identified and it was shown that this basis can be used to obtain the $S$-matrix, which contains information about the anyon braiding statistics, and the $T$-matrix, which gives access to the chiral central charge and the topological spins of the anyons. Later works showed that these MES also give access to the topological spins of the anyons via a quantity called the ‘momentum polarization’ [@ZaletelMong; @TuZhang].
In this work, we focus on fermion systems with Abelian topological order, meaning that the anyons form an Abelian group under fusion. In the algebraic languague, the most important difference between bosonic and fermionic topological orders is that the latter don’t satisfy the same strict requirement of modularity as bosonic systems do. We explain this in more detail in the main text, where the algebraic frameworks for both bosonic and fermionic Abelian topological orders are reviewed. A fermionic system necessarily has a fermion parity symmmetry, and a useful tool in the study of fermionic topological orders is to gauge this symmetry [@Kitaev] (the approach of gauging a global symmetry has also proven to be very useful in the study of symmetry-protected phases [@LevinGu]). Importantly, since the microscopic fermion becomes a non-local gauge charge after gauging, the resulting theory is purely bosonic. Below, two main questions regarding Abelian fermionic topological orders (AfTO) are addressed: (1) How does one uniquely identify the most general AfTO in numerics?, and (2) Can the fermion parity gauging procedure have a non-trivial effect on the edge physics?
Regarding the first question, we note that a lot of the above mentiond numerical techniques for identifying a topological order have been successfully applied to fractional quantum Hall systems and fractional Chern insulators [@Neupert; @ShengGu; @Regnault; @ZaletelMong; @Rezayi; @Grushin], which are of course fermionic in nature. These approaches relied crucially on the presence of a fermion number convervation symmetry, which allows for the definition of a Hall conductivity $\sigma_{xy}$. In fractional quantum Hall systems one knows the value of $\sigma_{xy}$ exactly from Galilean invariance (if there is no disorder), and in lattice systems it can be computed numerically as the many-body Chern number [@Niu; @ShengHaldane], or from the entanglement eigenvalues by an adiabatic flux insertion procedure [@Zaletel; @Grushin] (see also Ref. [@Alexandradinata]). In this work, we outline an alternative numerical detection scheme which does not rely on computing the Hall conductance, and which can identify the most general Abelian fermionic topological order. In formulating this numerical scheme, we will make heavy use of the special structure of gauged AfTOs. Another important ingredient for our detection scheme is the momentum polarization technique [@ZaletelMong; @TuZhang], and we point out some properties of this numerical probe which –to the best of our knowledge– have not been discussed previously in the literature and which are relevant for fermionic systems.
As for the second question regarding the edge physics of topological phases, we show that the bosonic theory which is obtained by gauging the fermion parity symmetry in an AfTO can have a gapped edge with the trivial vacuum if and only if the ungauged fermionic theory can have a gapped edge. As argued below, one direction of this statement, namely that an ungauged AfTO with a gapped edge implies a gauged bosonic topological order with a gapped edge, is expected on physical grounds. However, the inverse implication is less obvious. To show that gauging fermion parity does not change whether a system admits a gapped edge or not, we use the bulk-boundary correspondence formulated in terms of Lagrangian subgroups as introduced in Ref. [@Levin].
For completeness, we also mention some previous works which have studied fermionic topological orders and are relevant for the present work. Refs. [@GuWangWen; @TianLan1; @TianLan2] worked out an algebraic framework for general fermionic topological orders (both Abelian and non-Abelian), which is a generalization of the modular tensor categories for bosonic topological orders. Some ideas of these works will be used below. Refs. [@Belov; @Cano] have studied Abelian fermionic topological orders with fermion number conservation symmetry using multi-component U$(1)$ Chern-Simons theories. The algebraic approach adopted in this work agrees with the U$(1)$ Chern-Simons approach of Refs. [@Belov; @Cano] where the results overlap. And finally, Ref. [@Wang] has studied general gauged fermionic topological orders from a mathematical perspective, and conjectured that Kitaev’s $16$-fold way [@Kitaev] has a natural generalization to the most general fermionic topological order. This being said, let us now turn to a short review of the algebraic framework behind bosonic and fermionic Abelian topological order.
Abelian bosonic topological order
=================================
In this section, we briefly review the properties of Abelian bosonic topological orders (AbTO) that are relevant for the main discussion below. For more details, the reader is referred to Ref. [@Kitaev]. In mathematical terms, an AbTO is equivalent to an Abelian modular tensor category. To specify such an Abelian modular tensor category, we need to provide a list $\mathcal{A}=\{a,b,c,\dots\}$ of $N$ anyon types, together with the corresponding Abelian fusion rules $a\times b = \sum_c N_{a,b}^c c$ such that $N_{a,b}^c\in \{0,1\}$, and topological spins $\theta_a$ [^1]. Every AbTO has a unique trivial anyon, denoted as $1$, which has the properties $1\times a = a$ and $\theta_1=1$. Often, we will write the fusion of two anyons $a$ and $b$ simply as $ab = a\times b$. In principle, we also need to provide the $F$-symbols, but they will not be important here so we omit them. We denote the braiding phase associated with moving an anyon $b$ counter-clockwise around anyon $a$ as $M_{a,b}$. By definition, the braiding phases are symmetric: $M_{a,b}=M_{b,a}$. The ribbon identity allows $M_{a,b}$ to be expressed in terms of the topological spins:
$$M_{a,b}=\frac{\theta_{ab}}{\theta_a\theta_b}$$
The $S$ matrix is related to the braiding phases by $S_{a,b}=M_{a,b}\mathcal{D}^{-1}$, where the total quantum dimension $\mathcal{D}$ of an Abelian modular tensor category is given by $\mathcal{D}=\sqrt{N}$. Modularity requires that $S^\dagger S=\mathds{1}$.
An important connection between the bulk anyons and the boundary theory of an AbTO is given by the following relation [@Kitaev]:
$$\label{kit}
\frac{1}{\sqrt{N}}\sum_a \theta_a = e^{2\pi i c_-/8}\, ,$$
where $c_-$ is the chiral central charge of the boundary theory. This implies that the bulk anyons determine the boundary chiral central charge up to a multiple of $8$. This is the best one can do, as there exists an invertible bosonic topological phase, the so-called $E_8$ state [@KitaevE8], which has no anyons but a chiral edge with $c_-=8$. One can thus always stack an $E_8$ state on top of the system of interest, which does not affect the bulk anyons but changes the boundary chiral central charge by $8$.
Gapped boundaries
-----------------
As was shown by Levin, an AbTO admits a gapped edge iff (1) $c_-=0$ and (2) the bulk anyons have a bosonic Lagrangian subgroup [@Levin]. A bosonic Lagrangian subgroup $\mathcal{L}_b$ is a subgroup of anyons which have the following properties:
- Every anyon in $\mathcal{L}_b$ has trivial topological spin,
- All anyons in $\mathcal{L}_b$ braid trivially with each other,
- Every anyon which is not in $\mathcal{L}_b$ braids non-trivially with at least one anyon in the Lagrangian subgroup.
Using Eq. (\[kit\]), we will show that the existence of a Lagrangian subgroup implies that the chiral central charge is a multiple of $8$. It thus follows that an AbTO allows for a gapped edge iff it has a Lagrangian subgroup, and the separate requirement of zero chiral central charge is redundant, provided that we are allowed to stack $E_8$ states on top of our system. Combined with the results of Ref. [@Levin2], this implies that every AbTO with a Lagrangian subgroup has a string-net representation [@stringnet].
To derive $c_-=0$ mod $8$ from the existence of a bosonic Lagrangian subgroup $\mathcal{L}_b$, we write the AbTO as $\mathcal{A}=\{\mathcal{L}_b,\mathcal{L}_b\times n_1, \mathcal{L}_b\times n_2,\dots\}$, where $n_i$ are a set of arbitrary anyons not in $\mathcal{L}_b$. Starting from Eq. (\[kit\]), we can now do the following manipulations:
$$\begin{aligned}
e^{2\pi ic_-/8} & = & \frac{1}{\sqrt{N}}\sum_{a\in\mathcal{A}} \theta_a \\
& = & \frac{1}{\sqrt{N}}\sum_{l \in \mathcal{L}_b}\theta_l + \frac{1}{\sqrt{N}} \sum_{n_i}\sum_{l \in \mathcal{L}_b}\theta_{n_il} \\
& = & \frac{1}{\sqrt{N}}\sum_{l \in \mathcal{L}_b}\theta_l + \frac{1}{\sqrt{N}} \sum_{n_i}\theta_{n_i} \sum_{l \in \mathcal{L}_b}\theta_{l}M_{l,n_i} \\
& = & \frac{N_{L}}{\sqrt{N}}+ \frac{1}{\sqrt{N}} \sum_{n_i}\theta_{n_i} \sum_{l \in \mathcal{L}_b}M_{l,n_i} \label{sum}\\
& = & \frac{N_{L}}{\sqrt{N}}\, ,\label{Lb}\end{aligned}$$
where $N_{L}$ is the number of anyons in $\mathcal{L}_b$. In the third line we have used the ribbon identity, and in the fourth line we relied on the definition of a bosonic Lagrangian subgroup wich states that $\theta_l=1$ if $l\in \mathcal{L}_b$. To see why the second term in (\[sum\]) is zero, note that $M_{l_1,n_i}M_{l_2,n_i}=M_{l_1l_2,n_i}$, such that $M_{l,n_i}$ for fixed $n_i$ forms a representation of $\mathcal{L}_b$. Because there is at least one $l\in\mathcal{L}_b$ for which $M_{l,n_i}\neq 1$, this representation cannot be the trivial representation. Schur’s orthogonality relations then imply that the sum of $M_{l,n_i}$ over all $\l\in\mathcal{L}_b$ is zero. We have thus obtained the desired result that $c_-=0$ mod $8$ if there exists a Lagrangian subgroup. As a side-result, we also found that $N_{L}^2=N$, such that only AbTOs where the number of anyons is a square number can have a Lagrangian subgroup (this relation between $N_{L}$ and $N$ was also obtained previously in Ref. [@Levin]).
Abelian fermionic topological order
===================================
The main difference between a fermionic topological order (fTO) and a bTO, is that a fTO has a distinguished fermion excitation $f$ with the properties $f^2=f\times f=1$ and $\theta_f=-1$, which is ‘transparent’, i.e. it braids trivially with all other particles. This is not allowed in a bTO because the existence of such a particle is a violation of modularity. In Refs. [@Cano; @Cheng], it was shown that every Abelian fermionic topological order (AfTO) $\mathcal{A}_f$ can be written as
$$\label{prod}
\mathcal{A}_f = \mathcal{A}_b\times \{1,f\}\, ,$$
where $\mathcal{A}_b$ is an AbTO. This result also follows from corollary A.19 of Ref. [@Drinfeld]. Note that the above factorization property does not hold for non-Abelian fTO, as is known from explicit counter examples [@TianLan1; @TianLan2].
It is important to keep in mind that in general, the factorization in Eq. (\[prod\]) is not unique. In particular, let us define the homomorphism $\beta: \mathcal{A}_b\rightarrow \mathbb{Z}_2$, where $\beta(a)$ takes values in $\{0,1\}$. The fact that $\beta$ is a homomorphism then implies that $\beta(a b) = \beta(a)+\beta(b)$ mod $2$. Using $\beta$, we can rewrite the factorization in Eq. \[prod\] in a different way as $\mathcal{A}_f = \mathcal{A}_b^\beta\times\{1,f\}$, where $\mathcal{A}_b^\beta = \{a f^{\beta(a)}|a\in\mathcal{A}_b\}$. So the number of different factorizations is given by the number of homomorphisms from $\mathcal{A}_b$ to $\mathbb{Z}_2$.
A fermionic system necessarily has fermion parity symmetry. This implies that we can introduce a corresponding fermion parity flux or $\pi$-flux into the system. In the absence of U$(1)$ fermion number symmetry, there is no unique way of defining such a fermion parity flux. In particular, for a given parity flux $\phi_i$ we can always obtain a different parity flux by attaching an anyon in $\mathcal{A}_f$ to it. This of course assumes that the parity flux cannot ‘absorb’ the anyons in $\mathcal{A}_f$, i.e. that $\phi_i \times a\neq \phi_i$ for $a \in \mathcal{A}_f$. While it is possible for $\phi_i$ to absorb the transparent fermion $f$, we will show in the next section that the parity fluxes cannot absorb the anyons in $\mathcal{A}_b$. The case where the fermion parity fluxes can absorb $f$ occurs in superconducting systems where a Majorana mode binds to the parity flux, like in the weak pairing phase of spinless $p-$wave superconductors [@Volovik; @ReadGreen]. When this happens, the parity fluxes are non-Abelian defects [@Ivanov].
An important property of fermion parity fluxes is that they have a well-defined topological spin. This is different from $\mathbb{Z}_2$ fluxes in bosonic systems, which have a topological spin that is only defined up to a minus sign. The reason is that bosonic systems admit trivial particles with both even and odd $\mathbb{Z}_2$ charge. So if we attach such a trivial charge one object to a $\mathbb{Z}_2$ flux, we don’t change the superselection sector, but we do change its topological spin by $-1$. In fermionic systems, however, all charge one particles have topological spin $\theta_f=-1$, so by attaching them to a parity flux we don’t change the topological spin.
Gapped boundaries
-----------------
Similarly to the bosonic case, Levin showed that an AfTO admits a gapped edge if and only if $c_-=0$, and there exists a fermionic Lagrangian subgroup $\mathcal{L}_f$, which is defined to have the following properties [@Levin]:
- All anyons in $\mathcal{L}_f$ braid trivially with each other,
- Every anyon in $\mathcal{A}_b$ which is not in $\mathcal{L}_f$ braids non-trivially with at least one anyon in the Lagrangian subgroup.
The only difference between the definitions of $\mathcal{L}_b$ and $\mathcal{L}_f$ is that in the latter we do not require the anyons in $\mathcal{L}_f$ to have trivial topological spin. Note, however, that because $\theta_l^2=M_{l,l}=1$ for all $l\in\mathcal{L}_f$, it follows that the definition of a fermionic Lagrangian subgroup only leaves a sign ambiguity in the topological spins of the anyons in $\mathcal{L}_f$.
Gauging fermion parity: modular extensions of an AfTO
=====================================================
There exists a well-defined microscopic prescription to gauge the fermion parity symmetry in any fermionic lattice Hamiltonian [@Kogut]. After gauging, the system realizes a bosonic topological order in the bulk. This is because the gauging procedure promotes the parity fluxes to deconfined anyonic excitations, which braid non-trivially with $f$. This implies that the gauged theory is modular, and can be realized in a bosonic system. In mathematical terms, a GfTO is called a ‘modular extension’ of the original fTO [@TianLan1; @TianLan2]. In this section, we will show that for AfTO such modular extensions have a special structure. For this we consider the process where one creates an $a-\bar{a}$ anyon pair from the vacuum and braids one of them, say $a$, around a fermion parity flux. It is well-known from bosonic symmetry-enriched topological orders that braiding around a symmetry defect $g$ can permute the anyon types [@Bombin; @Barkeshli; @Tarantino; @Chen]. This means that after braiding an anyon $a$ around $g$, it is possible for $a$ not to come back as itself, but as a different anyon $\pi_g(a)$. After gauging, the parity fluxes become deconfined anyonic excitations. Because braiding of anyons cannot change the anyon type, this implies that after gauging the anyons $a$ and $\pi_g(a)$ have to be identified as the same anyon [@Barkeshli; @Tarantino; @Chen]. However, we now argue that this cannot happen for a fermion parity flux, i.e. for fermion parity we always have $\pi_\phi(a)=a$. The reason is that fermionic Hilbert spaces have a superselection rule which states that every physical state needs to have well-defined fermion parity [@Wick]. So if fermion parity could permute anyons, then starting from an excited state with some localized anyons it would be possible to create an orthogonal state by acting with fermion parity, but this clearly violates the superselection rule.
As already anticipated above, we can now also argue that fermion parity fluxes cannot absorb anyons in $\mathcal{A}_b$. To see this, assume that a parity flux $\phi_i$ could absorb an anyon $a\in\mathcal{A}_b$. The only way this can happen consistently, is if all anyons $b$ with $M_{a,b}\neq 1$ get permuted when they braid with the parity flux. But as we just argued, this is impossible. Note that $f$ escapes this argument since it braids trivially with all anyons in $\mathcal{A}_f$.
When gauging a $\mathbb{Z}_2$ symmetry in bosonic systems, the trivial anyon $1$ before gauging splits into a trivial anyon and a non-trivial anyon after gauging. This is because the original symmetry-enriched topological order has trivial anyons with both even and odd $\mathbb{Z}_2$ charge. Under gauging, the trivial anyons with even charge remain trivial, but the trivial anyons with odd charge become the gauge charge anyons of the gauged theory. For fermionic systems, this does not happen because a fTO has no trivial anyons with odd fermion parity.
The above two arguments show that the anyons of a fTO do not get identified and do not split after gauging fermion parity, which implies that the original fTO is a subcategory of the GfTO. For AfTO, if $\mathcal{A}_f\times\{1,f\}$ is a subcategory of the GfTO, then $\mathcal{A}_b$ is obviously also a subcategory of the gauged theory. We can now use theorem 3.13 from Ref. [@Drinfeld], which says the following:\
\
**Theorem (Ref. [@Drinfeld]).** Consider a MTC $\mathcal{K}$, and assume that it is a fusion subcategory of $\mathcal{C}$, i.e. $\mathcal{K}\subset\mathcal{C}$. If $\mathcal{C}$ is a MTC, then $\mathcal{C}=\mathcal{K}\times\mathcal{K}'$, where $\mathcal{K}'$ is also a MTC.\
\
Because $\mathcal{A}_b$ is modular, we can directly apply the above theorem to conclude that the GfTO takes the form
$$\label{factor1}
GfTO = \mathcal{A}_b \times\{1,f,\phi,f\phi\}\, ,$$
if the fermion parity fluxes are Abelian, and
$$\label{factor2}
GfTO = \mathcal{A}_{b}\times\{1,f,\phi\}\, ,$$
if the fermion parity fluxes are non-Abelian. Both $\phi$ and $f\phi$ are anyons which correspond to deconfined fermion parity fluxes.
In the present context, it is not hard to prove the factorization of the GfTO, as given in Eqs. (\[factor1\]) and (\[factor2\]), without invoking theorem 3.13 of Ref. [@Drinfeld]. To see this, note that since fermion parity fluxes $\phi_i$ cannot permute anyons, the process of making an $a-\bar{a}$ pair, braiding $a$ around $\phi_i$, and subsequently annihilating the anyon pair again is a well-defined adiabatic process for every parity flux $\phi_i$ and anyon $a\in\mathcal{A}_b$. Therefore, we can associate a Berry phase to it. Because there are no trivial particles with odd fermion parity, the Berry phases depend only on the anyon type and it makes sense to write them as $e^{i\gamma_i(a)}$ and $e^{i\gamma_i(af)}$ for every flux $\phi_i$ and $a\in\mathcal{A}_b$. The Berry phases satisfy the obvious properties $e^{i\gamma_i(a b)}=e^{\gamma_i(a)}e^{i\gamma(b)}$ and $e^{i\gamma_i(f)}=-1$. Because $\mathcal{A}_b$ is modular, we can use lemma 3.31 from Ref. [@Drinfeld] (see also Ref. [@Barkeshli], page 11), which states that for every function $e^{i\gamma_i(\cdot)}:\mathcal{A}_b\rightarrow U(1)$ that satisfies $e^{i\gamma_i(a b)}=e^{\gamma_i(a)}e^{i\gamma_i(b)}$, there exists a corresponding unique anyon $a_i\in \mathcal{A}_b$ such that
$$\label{braidingthm}
e^{i\gamma_i(a)}=M_{a,a_i}\,,\;\;\forall a \in \mathcal{A}_b$$
This implies that for every parity flux $\phi_i$, we can find an anyon $a_i\in\mathcal{A}_b$ such that $\phi_i \bar{a}_i$ braids trivially with all anyons in $\mathcal{A}_b$. Because the different $\phi_i$ are related by fusion with anyons in $\mathcal{A}_b$, and for every $\phi_i$ there is a unique anyon $a_i$ such that (\[braidingthm\]) holds, we conclude that $\phi_i \bar{a}_i$ is independent of $i$. The parity flux in Eqs. (\[factor1\]) and (\[factor2\]) is then simply defined as $\phi = \phi_i\bar{a}_i$ (a similar argument for the factorization of GAfTOs with fermion number symmetry was recently given in Ref. [@Lapa]).
As mentioned previously, the factorization of an AfTO $\mathcal{A}_f=\mathcal{A}_b\times\{1,f\}$ is not unique and we can obtain an equivalent factorization $\mathcal{A}_f=\mathcal{A}_b^\beta\times\{1,f\}$ using a homorphism $\beta:\mathcal{A}_b\rightarrow \mathbb{Z}_2$. So as consistency check, we show that for every $\mathcal{A}_b^\beta$, there exists a corresponding factorization of the GfTO as in Eqs. (\[factor1\]) and (\[factor2\]). To see this, we can use the same lemma from Ref. [@Drinfeld] to conclude that for every homomorphism $\beta$, there must exist an anyon $\tilde{b}\in\mathcal{A}_b$ such that
$$(-1)^{\beta(a)}=M_{a,\tilde{b}}\,,\;\; \forall a \in\mathcal{A}_b$$
So we see that now the GfTO factorizes as $\mathcal{A}_b^\beta\times \{1,\tilde{b}\phi ,f, \tilde{b}\phi f\}$ or $\mathcal{A}_b^\beta\times \{1,\tilde{b}\phi ,f\}$.
### Example: U$(1)_4\times$ $\overline{IQH}$
Let us give an example to illustrate the factorization property of GAfTO. We consider the multi-component U$(1)$ Chern-Simons theory
$$\mathcal{L}=\frac{1}{4\pi}K_{IJ}\epsilon^{\mu\nu\lambda}a_\mu^I \partial_\nu a_\lambda^J + \frac{1}{2\pi}t_I \epsilon^{\mu\nu\lambda} A_\mu \partial_\nu a_{\lambda}^I\, ,$$
with $K$-matrix $$K = \left( \begin{array}{cc} 4 & \\ & -1 \end{array}\right)$$ This describes a $\mathcal{A}_f=\mathcal{A}_b\times \{1,f\}=\mathbb{Z}_4\times \{1,f\}$ fermionic topological order, where the transparant fermion $f$ corresponds to the vector $l_f = \left(0,1\right)^T$, and $\mathbb{Z}_4=\{a,a^2,a^3,a^4=1\}$ is generated by anyon $a$ corresponding to vector $l_a=\left(1,0\right)^T$. The topological spins are given by $\theta_{a^p}=e^{ip^2\pi l_a^TK^{-1}l_a}=e^{ip^2\pi/4}$ and $\theta_f=e^{i\pi l_fK^{-1}l_f}=-1$. $A_\mu$ is a probe gauge field for the global U$(1)$ particle number symmetry. If we require that the fermion has charge one, then this implies that $q_f=1=l_f^TK^{-1}t= -t_2$. So the only freedom left is the first component from the charge vector $t=(t_1, -1)^T$. This freedom determines the Hall conductance, which is given by $\sigma_{xy}=t^TK^{-1}t=t_1^2/4-1$, and the U$(1)$ charges of the anyons in $\mathbb{Z}_4$: $q_a = l_a^TK^{-1}t=t_1/4$. Let us take $t_1=1$, such that $\sigma_{xy}=-3/4$ and $q_a=1/4$. If a system has U$(1)$ fermion number symmetry, then there is a preferred way to create a parity flux by adiabatically inserting $\pi$ flux of the U$(1)$ particle number symmetry. Let us denote the fermion parity flux obtained via this adiabatic procedure as $\phi_A$. As is well-known, the topological spin of $\phi_A$ is fixed by the Hall conductance. In particular, it holds that $\theta_{\phi_A} = e^{i\pi\sigma_{xy}/4}$ [@Goldhaber; @ChengZaletel]. If we apply this formula to our example with $t_1=1$, we learn that $\theta_{\phi_A}=e^{-i3\pi/16}$ and therefore $\theta_{\phi_A^2} = e^{-i3\pi/4}$. This implies that $\phi_A\times\phi _A= af$ or $\phi_A\times \phi_A = a^3f$, which is at odds with the proposed factorization property of the GAfTO because we cannot find a parity flux $\phi= \phi_A a^p$ such that $\phi\times\phi= f$. However, the choice $t_1=1$ is not allowed. This is because if $q_a=1/4$, then $q_{a^4}=q_1= 1$. This is not possible if $t$ is the charge vector of U$(1)$ fermion number symmetry, because the trivial anyon should always have even fermion parity. So $t_1=2t'$ has to be even. With this property correctly incorporated, we find $\theta_{\phi_A} = e^{i\pi (t'^2-1)/4}$ and $\theta_{\phi_A^2}=e^{i\pi(t'^2-1)}=(-1)^{t'+1}$. Under a shift $t'\rightarrow t'+4$, the anyon charges change as $q_a\rightarrow q_a + 2$, and $\theta_{\phi_A}$ remains invariant. So the only four remaining cases we have to consider are $t'=0,1,2,3$, corresponding to respectively $q_a = 0,1/2,1,3/2$. Let us work through the case where $t'=1$. With $t'=1$, it holds that $\phi_A\times\phi_A = a^2f$. We can now define $\phi = \phi_A a^3$, such that $\phi \times\phi = f$ and $M_{a,\phi}=M_{a,a^3}M_{a,\phi_A} = e^{i3\pi/2}e^{i\pi q_a} = 1$. Therefore, the GfTO factorizes as $\{a,a^2,a^3,1\}\times \{1,\phi_Aa^3,f,\phi_Aa^3f\}$. The factorization for other choices of $t'$ can be obtained in a similar way.
Numerical identification of Abelian fermionic topological order
===============================================================
An important question is how one can numerically identify the type of topological order realized by a particular microscopic lattice Hamiltonian. In the past decade, it has become clear that the ground state wavefunctions contain a lot of (if not all) information about the anyonic excitations. The first example of a ground-state property that can be used to diagnose topological order is the topological entanglement entropy [@KitaevPreskill; @LevinWen], which gives access to the total quantum dimension. By looking not only at the entanglement entropy, but at the entire spectrum of the reduced density matrix corresponding to some spatial region in the ground state wavefunction one also obtains universal information about possible gapless edge modes of the system [@LiHaldane]. This correspondence between the entanglement spectrum and edge spectrum has been worked out in full detail for free fermion systems in Ref. [@Alexandradinata; @Fidkowski]. For systems on a torus, Ref. [@ZhangGrover] identified the ground states with a definite anyon flux through one of the holes of the torus as those which are ‘minimally entangled states’ (MES) with respect to cuts wrapping the hole under consideration. In the MES basis, one can obtain both the $S$ and $T$ matrices by taking certain wave function overlaps [@ZhangGrover; @ZaletelMong]. Finally, using the same MES on the cylinder, one can also find the $T$ matrix by calculating the ‘entanglement polarization’ [@ZaletelMong; @TuZhang; @Cincio; @Zaletel; @He; @Wen]. In this section, we first clarify/review some subtleties of entanglement polarization in both boson and fermion systems. After the discussion of momentum polarization, we address the question of how to uniquely determine a fermionic topological order in numerics.
Momentum polarization in bosonic systems
----------------------------------------
Before turning to fermion systems, we first review the concept of momentum polarization in boson or spin systems. Consider a MES on a cylinder with an anyon flux of the type $a$ through the hole of the cylinder. Let us denote this MES as $|\psi[a]\rangle$. We will call the direction along the axis of the cylinder the $x$-direction, and the direction wrapping the hole the $y$-direction. The size of the cylinder is given by $N_x\times N_y$ unit cells. We now choose a cut along the $y$-direction close to the middle of the cylinder, dividing the cylinder in two. The length of the left half is then $N_x^L$, while the length of the right half is $N_x^R$, such that $N_x^L+N_x^R=N_x$. With this cut, the translation operator in the $y$-direction can be written as a tensor product between the translation operator on the left half and the translation operator on the right half: $T_y = T_y^L\otimes T_y^R$. With these definitions in place, momentum polarization was defined in Refs. [@ZaletelMong; @TuZhang; @Cincio] as the expectation value $\langle\psi[a]|T_y^L|\psi[a]\rangle$. It was found that this expectation value scales with $N_y$ as
$$\label{mompol}
\langle\psi[a]|T_y^L|\psi[a]\rangle = \exp\left(\frac{2\pi i}{N_y}\left( h_a -\frac{c_-}{24}\right) - \alpha N_y \right)\, ,$$
where $\theta_a = e^{2\pi i h_a}$ is the topological spin of anyon $a$, and $c_-$ is the chiral central charge as before. The complex number $\alpha$ is non-universal.
In the most general case, Eq. (\[mompol\]) needs to be generalized to the form
$$\label{mompolgeneral}
\langle\psi_b[C,a]|T_y^{C,L}|\psi_b[C,a]\rangle = \exp\left(\frac{2\pi i}{N_y}\left( h_a -\frac{c_-}{24}\right) + i\Theta_{a,b}N_x^{C,L} - \alpha_C N_y \right)\, ,$$
where the notation for the MES $|\psi_b[C,a]\rangle$ now depends on two anyon labels $a$ and $b$, and on an integer $C$ which denotes the position of the entanglement cut. The anyon $a$ is again the anyon flux through the hole of the cylinder, measured along cut $C$, and $b$ is a ‘background anyon’ which sits inside every unit cell [@Zaletel3; @ChengZaletel]. The translation operator $T^{C,L}_y$ acts on the left of the cut labeled by $C$. On the right hands side, $N_x^{C,L}$ is an integer which corresponds to the length of the left half of the cylinder, and $\alpha_C$ is a non-universal complex number also depending on the cut. Although $\alpha_C$ is non-universal, we will argue that the difference $\Delta\alpha_C = \alpha_{C+1}-\alpha_C$ between two neighboring cuts is universal. The interpretation of $\Theta_{a,b}$ will be explained in the next paragraph.
To explain the general form of the momentum polarization formula Eq. (\[mompolgeneral\]), it is useful to decompose the momentum polarization in an ‘eigenvalue contribution’ and an ‘entanglement contribution’. Let us consider translationally invariant systems, such that $|\psi_b[C,a]\rangle$ is an eigenstate of $T_y$. The $y$-momentum of the MES is then determined by the anyon flux $a$ and the background anyon $b$ as follows:
$$\label{Teig}
T_y|\psi_b[C,a]\rangle = e^{ i \Theta_{a,b} N_x}|\psi_a \rangle\, ,$$
where $e^{i\Theta_{a,b}} = M_{a,b}$ [@Zaletel3; @ChengZaletel]. This background anyon has to be non-trivial in systems where a Lieb-Schultz-Mattis-Oshikawa-Hastings (LSMOH) obstruction to a gapped trivial featureless phase is present [@LSM; @Oshikawa1; @Hastings; @Oshikawa2; @Zaletel3]. Equation (\[Teig\]) holds irrespective of whether the system is bosonic or fermionic, although in general the types of topological orders which can satisfy the LSMOH obstruction are different in both cases [@FillingC]. From Eq. (\[Teig\]) we immediately identify the eigenvalue contribution to the momentum polarization as $e^{i\Theta_{a,b}N_x^{C,L}}$.
To identify the entanglement contribution to momentum polarization, we write the MES as
$$\begin{aligned}
|\psi_b[C,a]\rangle & = & \sum_{\alpha,\beta} \Psi_{C,a,b}^{\alpha,\beta}|\alpha\rangle_L\otimes|\beta\rangle_R \\
& = & \sum_{\mu} s_{C,a,b}^\mu\, |\mu\rangle_L\otimes |\mu\rangle_R\end{aligned}$$
In the first line, we have used an arbitrary basis $|\alpha\rangle_L$ ($|\beta\rangle_R$) for the left (right) half of the cylinder. In the second line, the state is decomposed in the Schmidt basis. In the Schmidt basis, the action of translation in the $y$-direction can be written as
$$\begin{aligned}
T_y|\psi_b[C,a]\rangle & = & \sum_\mu s_{C,a,b}^\mu\, T^{C,L}_y|\mu\rangle_L\otimes T_y^{C,R}|\mu\rangle_R \\
& = & e^{ i \Theta_{a,b} N_x}\sum_{\mu\lambda\sigma} s_{C,a,b}^\mu [U^{*C,L}_{a,b}]_{\mu\lambda} [U^{C,R}_{a,b}]_{\mu\sigma}|\lambda\rangle_L\otimes|\sigma\rangle_R\, ,\end{aligned}$$
where in the second line, we have re-expanded $T^{C,L/R}_y|\mu\rangle_{L/R}$ in the Schmidt basis. Note that since $T_y^{C,L}$ and $T_y^{C,R}$ are unitary, so are $U^{C,L}_{a,b}$ and $U^{C,R}_{a,b}$. We have also separated out the eigenvalue factor $e^{i\Theta_{a,b}N_x}$. The MES $|\psi_b[C,a]\rangle$ is an eigenstate of $T_y$ if
$$\left(U_{a,b}^{C,L}\right)^\dagger S_{C,a,b}U_{a,b}^{C,R} = S_{C,a,b}\, ,$$
where $S_{C,a,b} =$diag$(s_{C,a,b}^\mu)$. This implies that $U^{C,L}_{a,b}=U^{C,R}_{a,b}=U^C_{a,b}$, where $U^C_{a,b}$ commutes with $S_{C,a,b}$. The entanglement contribution to momentum polarization is then entirely given in terms of the Schmidt values and the unitary matrix $U^C_{a,b}$, and takes the form
$$\tr(S^2_{C,a,b}U^{C}_{a,b}) = \exp\left(\frac{2\pi i}{N_y}\left( h_a -\frac{c_-}{24}\right) - \alpha_C N_y \right) \, ,$$
where we have again used the notation $\alpha_C$ to emphasize that $\alpha_C$ depends on the choice of cut. To understand how $\alpha_C$ depends on the cut, we note that because of the background anyon $b$ per unit cell, the anyon flux through the hole of the cylinder is not the same for every cut. The anyon fluxes for two neighboring cuts differ by the total background anyon charge enclosed by the two cuts, which is $b^{N_y}$. In other words, it holds that
$$|\psi_b[C,a]\rangle = |\psi_b[C+1,ab^{N_y}]\rangle$$
The ribbon identity allows us to write $\theta_{ab^{Ny}} = \theta_a\theta_{b}^{N_y^2}M_{a,b}^{N_y}$. So given the momentum polarization for a particular cut, we can obtain the momentum polarization for the neighboring cut by shifting
$$h_a \rightarrow h_a + h_b N_y^2 + \frac{\Theta_{a,b}}{2\pi}N_y$$
From this we conclude that $\alpha_C$ depends on the cut as $\alpha_{C+1}=\alpha_C +2\pi i h_b$. Note in particular that this implies that even though the anyon flux through the hole of the cylinder depends on the choice of entanglement cut, the momentum polarization nevertheless gives a topological spin $h_a$ which is independent of the choice of cut because it relies on a scaling in the cylinder circumference $N_y$.
Momentum polarization in fermion systems
----------------------------------------
### Anti-periodic sector –
For fermion systems on a cylinder, there are two types of boundary conditions: periodic and anti-periodic. For each type of boundary condition, one can find a set of MES. Let us start by considering MES in the anti-periodic sector. It is important to note that these MES have anyon fluxes through the hole of the cylinder which are labeled by the anyons in $\mathcal{A}_f$, and *not* by the different types of parity fluxes. The latter label MES in the *periodic* sector, which we discuss below. So let us write a MES in the anti-periodic sector as
$$|\psi_{f^\sigma b}[C,f^\lambda a]\rangle\, , \;\;\;\sigma,\lambda\in \{0,1\}\,, \;\;a,b \in \mathcal{A}_b\, ,$$
where $f^\sigma b$ again denotes the background anyon and $f^\lambda a$ the anyon flux through the hole of the cylinder measured at cut $C$. Note that both the background anyon and the flux through the hole of the cylinder are labeled with anyons in $\mathcal{A}_f=\mathcal{A}_b\times\{1,f\}$, even though the ground state degeneracy on the torus is only given by $|\mathcal{A}_b|$, i.e. the number of anyons in $\mathcal{A}_b$.
In the anti-periodic sector, the translation operator along the $y$-direction is defined as
$$\begin{aligned}
\tilde{T}_y: & c^\dagger_{(x,y)}\rightarrow c^\dagger_{(x,y+1)}\,,\qquad y\neq N_y-1\\
& c^\dagger_{(x,N_y-1)}\rightarrow -c^\dagger_{(x,1)}\end{aligned}$$
Using this twisted translation operator, the momentum polarization is defined as
$$\label{mompolAP}
\langle\psi_{f^\sigma b}[C,f^\lambda a]|\tilde{T}^{C,L}_y|\psi_{f^\sigma b}[C,f^\lambda a] \rangle = \exp\left(\frac{2\pi i}{N_y}\left( h_{f^\lambda a} -\frac{c_-}{24}\right) + i\Theta_{a,b}N_x^{C,L} - \alpha_C N_y \right)$$
We again have the following identity for the MES:
$$|\psi_{f^\sigma b}[C,f^\lambda a]\rangle = |\psi_{f^\sigma b}[C+1,f^{\lambda+\beta N_y} ab^{N_y}]\rangle\, ,$$
which via the replacement $h_{f^\lambda a}\rightarrow h_{f^\lambda a} + N_y^2 (\sigma h_f +h_b) + \Theta_{a,b}N_y/2\pi$ implies that $\alpha_{C+1}-\alpha_C = 2\pi (\sigma h_f +h_b) = \sigma \pi + 2\pi h_b$. The dependence of $\Delta\alpha_C$ on $\sigma$ arises from the eigenvalue contribution to the entanglement polarization. To see this, consider the situation where $\mathcal{A}_b=1$. In this case, $\sigma$ corresponds to the fermion parity per unit cell, which means that on the torus the fermion parity of $|\psi_{f^\sigma}[C,1]\rangle$ is given by $(-1)^{\sigma N_x N_y}$. It then holds that the momentum in the $y$-direction on the torus (with periodic boundary conditions in the $x$-direction) is given by
$$\label{transeig1}
\tilde{T}_y|\psi_{f^\sigma}[C,1]\rangle = (-1)^{\sigma N_x N_y}|\psi_{f^\sigma}[C,1]\rangle$$
This property can readily be checked for band insulators, where $\sigma$ is the number of filled bands modulo $2$, and also follows from fermionic tensor network descriptions of gapped ground states [@fMPS; @fPEPS]. From Eq. (\[transeig1\]) we identify the eigenvalue contribution to the momentum polarization as $e^{i\sigma\pi N_x^{C,L}N_y}$, which indeed leads to the dependence of $\alpha_C$ on the cut as described above.
The last aspect of the MES in the anti-periodic sector that we need to comment on is the role of $f^\gamma$. The choice of $\gamma = 0,1$ is a non-universal property of the MES, as can be seen by noting that the value of $\gamma$ can be flipped by adding a single electron in a $k_y = \pi/N_y$ momentum state on the left of the cut, which changes the (eigenvalue contribution to the) momentum polarization accordingly by a factor of $e^{i\pi/N_y}$. This reflects the fact that the ground state degeneracy on the torus is given by $|\mathcal{A}_b|$, and not $|\mathcal{A}_f|$.
### Periodic sector –
In the periodic sector, we write the MES as
$$|\psi_{f^\sigma b}[C,\phi a f^\lambda]\rangle\, , \;\;\;\sigma,\lambda\in \{0,1\}\,, \;\;a,b \in \mathcal{A}_b\, ,$$
where, as before, $\phi$ is the fermion parity flux which braids trivially with all anyons in $\mathcal{A}_b$. The momentum polarization in the periodic sector is then given by
$$\label{mompolP}
\langle\psi_{f^\sigma b}[C,\phi a f^\lambda]| T_y^{C,L}|\psi_{f^\sigma b}[C,\phi a f^\lambda]\rangle = \exp\left(\frac{2\pi i}{N_y}\left( h_{\phi a} -\frac{c_-}{24}\right) + i\Theta_{\phi a,bf^\sigma}N_x^{C,L} - \alpha_C N_y \right)$$
The by now familiar property of the MES
$$|\psi_{f^\sigma b}[C,\phi a f^\lambda]\rangle = |\psi_{f^\sigma b}[C+1,\phi ab^{N_y} f^{\lambda+\sigma N_y}]\rangle$$
implies that $\alpha_C$ depends on the cut as $\alpha_{C+1}-\alpha_C = \sigma \pi + 2\pi h_b$. As a consistency check, let us again take $\mathcal{A}_b = 1$ such that $\sigma$ is the fermion number per site. This means that if we define the state on the torus with periodic boundary conditions along both cycles, then it holds that
$$(-1)^{\hat{F}}|\psi_{f^\sigma }[C,\phi ]\rangle = (-1)^{\eta+\sigma N_x N_y}|\psi_{f^\sigma }[C,\phi ]\rangle \, ,$$
where $\hat{F}$ is the fermion number operator, and $\eta=1$ if $\phi$ is non-Abelian [@ReadGreen] and $\eta=0$ otherwise. With this definition, one finds that on a torus the momentum in the $y$-direction is given by
$$\label{neighboringcuts}
T_y|\psi_{f^\sigma }[C,\phi ]\rangle = (-1)^{ \sigma N_x(N_y + 1)} |\psi_{f^\sigma }[C,\phi ]\rangle\, ,$$
Again, this property can easily be checked for band insulators, and can also be seen in the fermionic tensor network formalism [@fMPS; @fPEPS]. Eq. (\[neighboringcuts\]) implies that the momentum polarizations for two neighboring cuts indeed differ by a factor $(-1)^{\sigma(N_y +1)} = e^{i\left(\Theta_{\phi,f^\sigma}+\sigma \pi N_y\right)}$, which arises entirely from the eigenvalue contribution to the momentum polarization.
Finally, we note that the momentum polarization with periodic boundary conditions is independent of $f^\lambda$. This agrees with the fact that the topological spins of the fermion parity fluxes are invariant under the addition of a transparent fermion.
### Example –
Let us now give an illustration of the application of momentum polarization to fermion systems. We consider a system with $\mathcal{A}_b=1$, described by the translationally invariant, spinless free fermion Hamiltonian on the square lattice
$$\label{freefermion}
H = \sum_{\textbf{k}}\psi^\dagger_{\textbf{k}} h(\textbf{k})\psi_{\textbf{k}} \, ,$$
where the single particle Hamiltonian is given by
$$h(\textbf{k}) = \textbf{d}(\textbf{k})\cdot \mathbf{\sigma}=\left( \begin{array}{cc} \cos(k_x)+\cos(k_y)-1 & \sin(k_x)+i\sin(k_y) \\ \sin(k_x)-i\sin(k_y) & -\cos(k_x)-\cos(k_y)+1\end{array}\right)$$
Importantly, $|\textbf{d}(\textbf{k})|$ is non-zero in the entire Brillouin zone, and the map $\textbf{k}\rightarrow \textbf{d}(\textbf{k})/|\textbf{d}(\textbf{k})|$ covers the unit sphere once. This implies that the gapped bands of $h(\textbf{k})$ have Chern number $\pm 1$.
a\) ![(a) The momentum polarization for a Chern insulator on a cylinder of size $N_x\times N_y$ with fixed length $N_x =30$. $N_y \theta_C(N_y)$ is plotted as a function of $N_y^2$, where $\theta_C(N_y)$ is defined as $\langle \psi_f[C,1]|\tilde{T}^{C,L}_y|\psi_f[C,1]\rangle = |\langle \psi_f[C,1]|\tilde{T}^{C,L}_y|\psi_f[C,1]\rangle| \exp(i\theta_C(N_y))$. $N_y\theta_C(N_y)$ was determined numerically as $N_y\theta_C(N_y) = \ln( \exp(iN_y\theta_C(N_y))$ (see the discussion in the main text for why this is important). The intercept $2\pi p$ of the linear fit to the data points gives access to the chiral central charge via the relation $p=-c_-/24 = -1/24 \approx -0.041666$ mod 1. The entanglement cut was chosen such that $N_x^{C,L}=16$. (b) Same as in (a), but now in the periodic sector. The intercept $2\pi p$ is determined by $p = h_\phi -c_-/24 = 1/8 - 1/24 = 1/12 \approx 0.08333$ mod 1. (c) The slope of the linear fits in $(a)$ and $(b)$ as a function of $N_x^{C,L}$.[]{data-label="intercept"}](Chern_AP.pdf "fig:") b) ![(a) The momentum polarization for a Chern insulator on a cylinder of size $N_x\times N_y$ with fixed length $N_x =30$. $N_y \theta_C(N_y)$ is plotted as a function of $N_y^2$, where $\theta_C(N_y)$ is defined as $\langle \psi_f[C,1]|\tilde{T}^{C,L}_y|\psi_f[C,1]\rangle = |\langle \psi_f[C,1]|\tilde{T}^{C,L}_y|\psi_f[C,1]\rangle| \exp(i\theta_C(N_y))$. $N_y\theta_C(N_y)$ was determined numerically as $N_y\theta_C(N_y) = \ln( \exp(iN_y\theta_C(N_y))$ (see the discussion in the main text for why this is important). The intercept $2\pi p$ of the linear fit to the data points gives access to the chiral central charge via the relation $p=-c_-/24 = -1/24 \approx -0.041666$ mod 1. The entanglement cut was chosen such that $N_x^{C,L}=16$. (b) Same as in (a), but now in the periodic sector. The intercept $2\pi p$ is determined by $p = h_\phi -c_-/24 = 1/8 - 1/24 = 1/12 \approx 0.08333$ mod 1. (c) The slope of the linear fits in $(a)$ and $(b)$ as a function of $N_x^{C,L}$.[]{data-label="intercept"}](Chern_P.pdf "fig:") c) ![(a) The momentum polarization for a Chern insulator on a cylinder of size $N_x\times N_y$ with fixed length $N_x =30$. $N_y \theta_C(N_y)$ is plotted as a function of $N_y^2$, where $\theta_C(N_y)$ is defined as $\langle \psi_f[C,1]|\tilde{T}^{C,L}_y|\psi_f[C,1]\rangle = |\langle \psi_f[C,1]|\tilde{T}^{C,L}_y|\psi_f[C,1]\rangle| \exp(i\theta_C(N_y))$. $N_y\theta_C(N_y)$ was determined numerically as $N_y\theta_C(N_y) = \ln( \exp(iN_y\theta_C(N_y))$ (see the discussion in the main text for why this is important). The intercept $2\pi p$ of the linear fit to the data points gives access to the chiral central charge via the relation $p=-c_-/24 = -1/24 \approx -0.041666$ mod 1. The entanglement cut was chosen such that $N_x^{C,L}=16$. (b) Same as in (a), but now in the periodic sector. The intercept $2\pi p$ is determined by $p = h_\phi -c_-/24 = 1/8 - 1/24 = 1/12 \approx 0.08333$ mod 1. (c) The slope of the linear fits in $(a)$ and $(b)$ as a function of $N_x^{C,L}$.[]{data-label="intercept"}](alpha.pdf "fig:")
The free fermion Hamiltonian (\[freefermion\]) has two possible interpretations. Either we interpret it as a charge conserving model with two orbitals $A$ and $B$ on each site, in which case the vector of annihilation operators is given by
$$\begin{aligned}
\psi_{\textbf{k}} & = & \left(\begin{array}{c} c_{A,\textbf{k}} \\ c_{B,\textbf{k}} \end{array}\right)\, ,\end{aligned}$$
or we interpret $H$ as a superconductor, in which case $\psi_{\textbf{k}}$ is a Nambu spinor given by
$$\begin{aligned}
\psi_{\textbf{k}} & = & \left(\begin{array}{c} c_{\textbf{k}} \\ c^\dagger_{-\textbf{k}} \end{array}\right)\end{aligned}$$
In the first interpretation, the non-zero Chern number of the bands implies that the system is an anomalous Hall (or Chern) insulator. In the superconducting case, the system is a topological or weak-pairing $p+ip$ superconductor [@ReadGreen].
For free fermion systems there is a straightforward way to obtain the eigenvalues of the reduced density matrix corresponding to some spatial region [@Peschel]. First, we define the one-particle reduced density matrix as
$$C_{(j,k_y),(j',k_y')} = \langle c^\dagger_{(j,k_y)} c_{(j',k_y')}\rangle = \delta_{k_y,k_y'}\langle c^\dagger_{(j,k_y)} c_{(j',k_y)}\rangle\, ,$$
where $j$ is a spatial index along the axis of the cylinder, and $k_y$ is the momentum along the periodic direction. Next, we take the subblock of $C_{(j,k_y),(j',k_y)}$ where both $j$ and $j'$ lie on the left of the entanglement cut, and calculate the eigenvalues $\zeta_{k_y,n}$ of that subblock. The eigenvalues of the reduced density matrix corresponding to the left half of the cylinder are then labeled by a set of occupation numbers $n_{k_y,n}\in \{0,1\}$, and are given by [@Peschel]
$$\lambda[\{n_{k_y,n}\}] = \prod_{k_y,n} \left(\zeta_{k_y,n}\right)^{n_{k_y,n}}(1-\zeta_{k_y,n})^{1-n_{k_y,n}}$$
Using this expression for the eigenvalues of the reduced density matrix, the momentum polarizations for the Chern insulator in the anti-periodic and periodic sectors can readily be obtained as [@TuZhang; @Alexandradinata]
$$\begin{aligned}
\langle \psi_f[C,1]|\tilde{T}^{C,L}_y|\psi_f[C,1]\rangle & = & \prod_{-\pi< k_y \leq\pi }\prod_n\left((e^{ik_y}-1)\zeta_{k_y,n}+1 \right)\\
\langle \psi_f[C,\phi]|T^{C,L}_y|\psi_f[C,\phi]\rangle & = & \prod_{-\pi< k_y \leq\pi }\prod_n\left((e^{ik_y}-1)\zeta_{k_y,n}+1 \right)\, ,\end{aligned}$$
where in the anti-periodic (periodic) sector $k_y = \frac{2\pi}{N_y}\left(j+\frac{1}{2}\right)-\pi, \, j\in \{0,\dots,N_y-1\}$ for $N_y$ even (odd), and $k_y = \frac{2\pi}{N_y}j - \pi, \, j\in \{0,\dots,N_y-1\}$ for $N_y$ odd (even). For the $p+ip$ superconductor, the expressions for the momentum polarizations are very similar:
$$\begin{aligned}
\langle \psi_f[C,1]|\tilde{T}^{C,L}_y|\psi_f[C,1]\rangle & = & \prod_{0\leq k_y\leq \pi }\prod_n \left((e^{ik_y}-1)\zeta_{k_y,n}+1 \right) \\
\langle \psi_f[C,\phi]|T^{C,L}_y|\psi_f[C,\phi]\rangle & = & \prod_{0\leq k_y\leq \pi }\prod_n \left((e^{ik_y}-1)\zeta_{k_y,n}+1 \right)\, ,\end{aligned}$$
and the only difference is in the range of $k_y$. Using the particle-hole symmetry $\sigma^xh(-\textbf{k})^*\sigma^x = -h(\textbf{k})$, which implies that
$$\label{PH}
\zeta_{k_y,n} = 1-\zeta_{-k_y,n}\, ,$$
one can find an exact relation between the momentum polarizations of the Chern insulator and the $p+ip$ superconductor. From Eq. (\[PH\]), one finds that
$$\prod_{0<k_y<\pi ,n}\left( (e^{-ik_y} -1)\zeta_{-k_y,n} + 1\right) = \prod_{0<k_y<\pi,n} e^{-ik_y}\left( (e^{ik_y} -1)\zeta_{k_y,n} + 1\right)$$
This implies that the momentum polarizations of the Chern insulator are the square of those of the $p+ip$ superconductor, up to a factor $\prod_{0<k_y<\pi,n} e^{-ik_y}$. This factor only changes the term in the exponent of the momentum polarization which is linear in $N_y$, implying that both the chiral central charge and the topological spin of the fermion parity flux differ by a factor of two between the Chern insulator and the $p+ip$ superconductor. This is of course consistent with the known values $c_- = 1$ and $h_\phi = 1/8$ for the Chern insulator and $c_-=1/2$ and $h_\phi = 1/16$ for the $p+ip$ superconductor.
Let us write the momentum polarization in the anti-periodic sector as $\langle \psi_f[C,1]|\tilde{T}^{C,L}_y|\psi_f[C,1]\rangle = |\langle \psi_f[C,1]|\tilde{T}^{C,L}_y|\psi_f[C,1]\rangle| \exp(i\theta_C(N_y))$. In Fig. \[intercept\] (a), $N_y\theta(N_y)=\ln(\exp(iN_y\theta_C(N_y))$ is plotted as a function of $N_y^2$ for the Chern insulator. According to the discussion above, $\theta_C(N_y)$ is given by
$$\theta_C(N_y) = -\frac{2\pi i}{N_y} \frac{c_-}{24} - \alpha_C N_y\, ,$$
where $\Delta\alpha_C = \pi$ because the Chern insulator has an odd fermion parity per site ($\sigma = 1$). This means that $N_y\theta_C(N_y)$ has a linear dependence on $N_y^2$, and the intercept is determined by the chiral central charge. Because $\Delta\alpha_C = \pi$, the slopes for two neighboring cuts differ by $\pi$. This dependence of the slope on the choice of cut is also found numerically, as shown in Fig. \[intercept\] (c).
In the periodic sector, $\theta_C(N_y)$ is defined as $\langle \psi_f[C,\phi]|T^{C,L}_y|\psi_f[C,\phi]\rangle = |\langle \psi_f[C,\phi]|T^{C,L}_y|\psi_f[C,\phi]\rangle| \exp(i\theta_C(N_y))$, and it takes the form
$$\label{thetac}
\theta_C(N_y) = \frac{2\pi i}{N_y} \left(h_{\phi } -\frac{c_-}{24}\right) -\pi N_x^{C,L}- \alpha_C N_y \, ,$$
where again $\Delta\alpha_C=\pi$. From this equation it is clear that $e^{i\theta_C(N_y)}$ differs by a factor of $(-1)^{N_y+1}$ for two neighboring cuts. So if we compute $N_y\theta_C(N_y)$ numerically by taking the logarithm of $e^{iN_y\theta_C(N_y)}$, then the dependence on the cut of the terms in Eq. (\[thetac\]) which are constant and linear in the circumference $N_y$ will not show up. This means that $N_y\theta_C(N_y)$ computed in his way has a linear dependence on $N_y^2$, with a slope which does not depend on the choice of cut. In Fig. \[intercept\] (b), this linear dependence of $N_y\theta_C(N_y)$ on $N_y^2$ is plotted. From the intercept of this line we can find the topological spin of the parity flux. In Fig. \[intercept\] (c), the independence of the slope on the choice of cut is shown. However, if we would instead calculate $N_y\theta_C(N_y)$ as $N_y\theta_C(N_y)= N_y \ln(\exp(i\theta_C(N_y))$, then it follows from Eq. (\[thetac\]) that $N_y\theta_C(N_y)$ would not be linear in $N_y^2$ for odd $N_x^{C,L}$. We have verified that this is indeed the case, and that for odd $N_x^{C,L}$, $N_y\theta_C(N_y) = N_y \ln (\exp(i\theta_C(N_y)+\pi(N_y+1)))$ is linear in $N_y^2$, with the same slope as for even $N_x^{C,L}$. This provides a non-trivial consistency check on the generalized expressions for the momentum polarizations in Eqs. (\[mompolAP\]) and (\[mompolP\]).
Numerically determining the topological order
---------------------------------------------
At this point we have collected all the necessary ingredients to outline how one would uniquely identify an AfTO in numerics. First, by calculating the momentum polarizations for the different MES on the cylinder in the anti-periodic sector we obtain the topological spins of the anyons in $\mathcal{A}_b$, up to a minus sign ambiguity. Secondly, for the MES on the torus with anti-periodic boundary conditions along both cycles, we can use the formalism of Refs. [@ZaletelMong; @ZhangGrover; @Cincio] to obtain the unitary $S$-matrix of the MTC corresponding to $\mathcal{A}_b$, just as one does for bosonic systems. From the $S$-matrix, one obtains the fusion rules (i.e. the group structure) of $\mathcal{A}_b$ via the Verlinde formula [@Verlinde]
$$N_{a,b}^c = \sum_d \frac{S_{a,d}S_{b,d}S_{d,c}^*}{S_{0,d}}\, ,$$
Once the group structure of $\mathcal{A}_b$ is obtained, this can be used to partially fix the sign ambiguity in the topological spins of the anyons in $\mathcal{A}_b$, by imposing that $S_{a,b} = N^{-1/2}M_{a,b} = N^{-1/2}\theta_{ab}/(\theta_a\theta_b)$, where $N$ is the number of MES (= the number of anyons in $\mathcal{A}_b$). This fixes the topological spins up to a homomorphism $\beta$ from $\mathcal{A}_b$ to $\mathbb{Z}_2$. This is the same homomorphism as discussed above, and different $\beta$ correspond to different ways of writing $\mathcal{A}_f = \mathcal{A}_b^\beta\times \{1,f\}$, where $\mathcal{A}_b^\beta = \{af^{\beta(a)}|a\in\mathcal{A}_b\}$. At this point, one has to choose a particular homomorphism to fix the topological spins.
To completely determine the topological quantum phase of the system of interest one does not only need to know $\mathcal{A}_f$, but also the complete algebraic data corresponding to the modular extension $\mathcal{A}_b\times \{1,f,\phi,f\phi\}$ (if $\phi$ is Abelian) or $\mathcal{A}_b\times\{1,f,\phi\}$ (if $\phi$ is non-Abelian). Let us first consider the case where the fermion parity fluxes are Abelian. Since $\phi$ is Abelian, it holds that either $\phi \times \phi = 1$ or $\phi\times\phi=f$. This in turn implies that $\theta_\phi^4=1$ or $\theta_\phi^4=-1$, which gives us eight different possible values for $\theta_\phi$. Once we know the topological spin of the Abelian parity flux, we have completely fixed which of the eight possible modular extensions is realized. When $\mathcal{A}_b=1$, this was shown by Kitaev as a part of his ‘$16$-fold way’ [@Kitaev].
A non-Abelian parity flux satisfies $\phi\times\phi=1+f$. If $\mathcal{A}_b=1$, then Kitaev has shown that there exist exactly eight different modular extensions with non-Abelian parity fluxes [@Kitaev]. These eight different modular extensions correspond to the eight different Ising MTC’s, and are uniquely identified by the topological spin of the fermion parity flux. In Ref. [@Wang], it was conjectured that Kitaev’s 16-fold way generalizes to all fTO’s (Abelian and non-Abelian), i.e. it was conjectured that every fTO has exactly 16 different modular extensions, 8 of which have Abelian fermion parity flux and 8 which have non-Abelian fermion parity flux. When the fTO is Abelian, it was shown above that the GfTO factorizes as $\mathcal{A}_b\times\{1,f,\phi\}$, so there are indeed eight different non-Abelian modular extensions which again correspond the eight different Ising categories.
From the above discussion, we learn that to complete the numerical identification of an Abelian fermionic topological order we only need to know $\theta_\phi$. To access $\theta_\phi$, one first calculates the momentum polarizations of the MES on the cylinder with periodic boundary conditions. This provides a set of topological spins, which correspond to $\theta_{a\phi}$. Because $\phi$ braids trivially with all anyons in $\mathcal{A}_b$, these topological spins factorize as $\theta_{a\phi} = \theta_a \theta_\phi$. This implies that we can organize the topological spins in the periodic sector in an ordered vector, which is proportional to the ordered vector of topological spins in the anti-periodic sector. The proportionality constant obtained in this way is unique and corresponds to $\theta_\phi$, such that one can simply try all the possible permutations of the topological spins in the periodic sector until one finds one where the required proportionality is realized. To see why the proportionality constant is unique we need to show that we cannot permute the vector of topological spins in the periodic sector to obtain a vector that is proportional to (and different from) the original, unpermuted one. Since we are only interested in permutations that do not leave the vector invariant, there can be no element in the vector that is fixed under the permutation. This means that the permutation acts as $\theta_{a\phi} \rightarrow \theta_{ad\phi}$, where $d$ is an anyon in $\mathcal{A}_b$ which is not the trivial anyon. Because $\theta_{ad\phi} = \theta_{a\phi}\theta_d M_{a,d}$, the resulting vector is proportional to the unpermuted one if and only if $M_{a,d}$ is independent of $a$. But this is a violation of modularity, and therefore this cannot happen. This completes the procedure of how to uniquely characterize an AfTO in numerics.
Before concluding the discussion on numerical identification of fermionic topological orders, let us consider what happens if one would have made a different choice of homomorphism $\beta$. In that case, $\mathcal{A}_b$ becomes $\mathcal{A}_b^\beta = \{af^{\beta(a)}|a\in \mathcal{A}_b\}$ and one would interpret the set of topological spins obtained from momentum polarization in the periodic sector as $\theta_{af^{\beta(a)}\tilde{\phi}} = \theta_{af^{\beta(a)}\phi \tilde{b}}$, where $\tilde{b}$ is the unique anyon that satisfies $(-1)^{\beta(a)} = M_{a,\tilde{b}}$. Using this property, one can factorize the topological spins in the periodic sector as
$$\theta_{af^{\beta(a)}\tilde{\phi}}=\theta_{af^{\beta(a)}\phi \tilde{b}} = \theta_{af^{\beta(a)}}\theta_{\phi \tilde{b}} = \theta_{af^{\beta(a)}}\theta_{\tilde{\phi}} \, ,$$
so we can again permute them in such a way that they become proportional –as a vector– to the vector of topological spins $\theta_{af^{\beta(a)}}$ in the anti-periodic sector. Of course, the set of topological spins in the periodic sector obtained from momentum polarization is independent of our choice of $\beta$, as can easily be verified:
$$\begin{aligned}
\theta_{af^{\beta(a)}\tilde{\phi}} = \theta_{af^{\beta(a)}\phi \tilde{b}}= \theta_{a\tilde{b}\phi}\end{aligned}$$
So we find that the final identification of the AfTO is independent of our choice of $\beta$, as it should be of course. One only needs to keep in mind that when comparing two different AfTO with the same $\mathcal{A}_f$, one should always use the same of choice of $\beta$ to compare the topological spins of the fermion parity fluxes $\phi$.
Gauging fermion parity with boundaries
======================================
The physical picture
--------------------
In this section, we address the question of what happens at the boundary of an AfTO after fermion parity is gauged. Physically, one expects the gauged topological order to have a gapped edge if the ungauged fTO has a gapped edge. To see this, consider a fTO which has a gapped boundary $\mathcal{B}_0$ separating it from the trivial phase. Now imagine gauging the fermion parity everywhere in the bulk, except in a narrow strip along the edge. After gauging, the GfTO in the bulk is separated by a boundary $\mathcal{B}_1$ from a narrow strip of the original ungauged fTO, which itself is separated from the trivial phase by the gapped boundary $\mathcal{B}_0$. See figure \[fig:B1\](a) for an illustration. If $\mathcal{B}_1$ is also gapped, then this construction gives a gapped boundary separating the GfTO from the trivial phase.
To argue why we can always take $\mathcal{B}_1$ to be gapped, let us first consider the case of bTO. In bosonic systems, the ‘ungauging’ procedure corresponds to condensing the gauge charges [@Barkeshli], which are always bosonic and braid trivially with each other. Condensing the gauge charges results in confinement of the gauge fluxes, which means that after condensation the energy of a flux pair grows linearly with the spatial separation between the pair. Condensation of the gauge charges thus transforms the gauge fluxes into the symmetry defects of the ungauged phase [@Barkeshli]. Using the general relation between anyon condensation and gapped boundaries [@Levin; @KitaevKong], we can then always construct the gapped boundary $\mathcal{B}_1$ between the gauged and ungauged bTO as a domain wall where the gauge charges get condensed.
In GfTOs, the $\mathbb{Z}_2$ gauge charge $\tilde{f}$ is by definition a fermion. Because $\tilde{f}$ has non-trivial topological spin, it is impossible to construct a gapped boundary where $\tilde{f}$ gets condensed. However, at a domain wall between the GfTO and the original fTO, we can condense the bound state $\tilde{f}f$, where $f$ is the transparant fermion of the fTO. Because $\tilde{f}f$ is a boson, this will result in a gapped boundary. See Refs. [@Aasen; @Wan] for more details.
A more microscopic way to understand why $\mathcal{B}_1$ is gapped, is to recall that the lattice gauging procedure maps local commuting projector Hamiltonians to local commuting projector Hamiltonians (if the gauging is done at zero gauge coupling). It is generally believed that every phase with a gapped boundary admits a commuting projector Hamiltonian representation (see e.g. Ref. [@Levin2] for a proof for AbTO). So this implies that gauging discrete symmetries (with both bosonic and fermionic gauge charges) maps systems with gapped boundaries to systems with gapped boundaries.
![A GfTO obtained by gauging a fTO with a gapped boundary $\mathcal{B}_0$ to the trivial phase. The gauging is done such that a narrow strip along the boundary is unaffected and remains in the original fTO phase. The boundary separating the GfTO from the ungauged fTO is denoted as $\mathcal{B}_1$.[]{data-label="fig:B1"}](B1.pdf)
If the fTO under consideration does not admit a gapped edge to the vacuum, i.e. if $\mathcal{B}_0$ is gapless, then this does not necessarily imply that the GfTO does not admit a gapped edge. To see this, again consider the situation shown in figure \[fig:B1\] and assume that the boundary $\mathcal{B}_0$ between the fTO and the trivial phase is gapless. In this case, it might be possible to gap out $\mathcal{B}_0$ by coupling it to the gapped edge $\mathcal{B}_1$. Or, there might exist a completely different way of constructing a gapped edge, which does not involve introducing two neighboring boundaries $\mathcal{B}_0$ and $\mathcal{B}_1$.
As mentioned previously, the GfTO is modular, and so we can think of it as a bosonic topological phase. Therefore, a natural question is whether the GfTO also admits a gapped edge $\mathcal{B}$ without making use of an intervening fTO. In other words, we would like to know whether the GfTO also has a gapped edge if we completely gauge the system such that there are no transparent fermions anymore, neither in the bulk nor at the boundary. Based on the fact that gauging maps commuting projector Hamiltonians to commuting projector Hamiltonians, we indeed expect the GfTO to admit a gapped edge if the ungauged fTO has a gapped edge. Below, we will show that this is indeed the case for AfTOs. We will also prove the less obvious opposite implication that the GAfTO has a gapless edge if the ungauged AfTO has a gapless edge.
Abelian parity flux
-------------------
When the parity fluxes are Abelian it is not difficult to see that if we apply Eq. (\[kit\]) to the modular GAfTO, we get the following expression:
$$\label{GM1}
e^{2\pi ic_-/8} = \frac{\theta_\phi}{\sqrt{N_b}}\sum_{a\in\mathcal{A}_b}\theta_a \,,$$
where $N_b$ is the number of anyons in $\mathcal{A}_b$. Because gauging a discrete symmetry cannot change the chiral central charge, Eq. (\[GM1\]) not only determines $c_-$ of the GAfTO (mod 8), but also of the original ungauged fTO [@TianLan1; @TianLan2].
If we apply Eq. (\[GM1\]) to a different factorization $\mathcal{A}_b^\beta\times\{1,\tilde{b}\phi,f,\tilde{b}\phi f\}$ of the GAfTO, we get
$$\begin{aligned}
e^{2\pi ic_-/8} & = & \frac{\theta_{\tilde{b}\phi}}{\sqrt{N_b}}\sum_{a\in\mathcal{A}_b^\beta}\theta_a \\
& = & \frac{\theta_{\phi}\theta_{\tilde{b}}}{\sqrt{N_b}} \sum_{a\in\mathcal{A}_b}\theta_a (-1)^{\beta(a)} \label{GM2}\end{aligned}$$
Because $M_{a,\tilde{b}}=(-1)^{\beta(a)}$ for all $a\in\mathcal{A}_b$, it follows that $\tilde{b}^2$ braids trivially with all anyons in $\mathcal{A}_b$, and must therefore be the trivial anyon. From $\tilde{b}^2=1$, we know that $\theta_{\tilde{b}}^4=1$. So by equating Eqs. (\[GM1\]) and (\[GM2\]), we find that the insertion of the minus signs $(-1)^{\beta(a)}$ in the sum of the topological spins of $\mathcal{A}_b$ changes the value of that sum by a multiplicative factor which is a fourth root of unity, and equals $\theta_{\tilde{b}}^*$. Now we are equipped to show the following result:
*A GAfTO has a bosonic Lagrangian subgroup $\mathcal{L}_b$ if and only if the corresponding ungauged AfTO has a fermionic Lagrangian subgroup $\mathcal{L}_f$, and an integer chiral central charge which is a multiple of eight.*
Using the bulk-boundary connection reviewed in the previous sections, this result then implies that a GfTO has a gapped edge if and only if the ungauged fTO has a gapped edge.
We first show the ‘if’ direction and assume that the fTO has a fermionic Lagrangian subgroup $\mathcal{L}_f$ and zero chiral central charge (mod 8). To start, we observe that because the anyons in $\mathcal{L}_f$ braid trivially with each other, it follows from the ribbon identity that their topological spins form a $\mathbb{Z}_2$ valued representation of $\mathcal{L}_f$. In other words, $\theta_{l_i}=(-1)^{\alpha(l_i)}$ such that $\alpha(\cdot):\mathcal{L}_f\rightarrow \mathbb{Z}_2=\{0,1\}$ is a homomorphism. In general this homomorphism cannot be extended to a homomorphism from $\mathcal{A}_b$ to $\mathbb{Z}_2$. Using this homomorphism, we define $\mathcal{L}'_f = \{l_if^{\alpha(l_i)}|l_i\in\mathcal{L}_f\}$ such that all anyons in $\mathcal{L}_f'$ have trivial topological spin. We now want to extend $\mathcal{L}'_f$ in such a way that it becomes a bosonic Lagrangian subgroup of the GAfTO. Because all anyons in $\mathcal{L}_f'$ braid trivially with $f$ and because $f$ itself cannot be in $\mathcal{L}_b$, the extended Lagrangian subgroup $\mathcal{L}_b$ will have to contain a fermion parity flux. At this point, it is clear that we can find a bosonic Lagrangian subgroup $\mathcal{L}_b=\mathcal{L}'_f\times\{1,\phi c\}$ of the GfTO, provided that there exists an anyon $c\in \mathcal{A}_b$ such that: (1) $M_{c,l}= (-1)^{\alpha(l)}$ for all $l\in\mathcal{L}_f$, and (2) $\theta_\phi = \theta_c^*$. Here, we have used the factorization property of GAfTO to define $\phi$ as the parity flux which braids trivially with all anyons in $\mathcal{A}_b$.
Let us first show that there exists an anyon $c$ satifying property (1), i.e. $M_{c,l}=(-1)^{\alpha(l)}$ for all $l\in\mathcal{L}_f$. If the homomorphism $\alpha: \mathcal{L}_f\rightarrow \mathbb{Z}_2$ is trivial, then property (1) is also trivial and $c$ is simply the identity anyon. Let us therefore focus on the case where $\alpha$ is non-trivial and write $\mathcal{A}_b = \{\mathcal{L}_f, \mathcal{L}_f\times c_1, \mathcal{L}_f\times c_2,\dots,\mathcal{L}_f\times d_1,\mathcal{L}_f\times d_2,\dots \}$, where $c_i, d_i$ are a set of anyons not in $\mathcal{L}_f$ of which the $c_i$ satisfy criterion (1), and the $d_i$ do not. Using Eq. (\[GM1\]), we find
$$\begin{aligned}
\sqrt{N_b}\theta_\phi^* e^{2\pi ic_-/8} & = & \sum_{l\in\mathcal{L}_f}\theta_l + \sum_{i}\sum_{l\in\mathcal{L}_f}\theta_{lc_i}+\sum_{i}\sum_{l\in\mathcal{L}_f}\theta_{ld_i} \\
& = & \sum_{i}\theta_{c_i}\sum_{l\in\mathcal{L}_f}(-1)^{\alpha(l)}M_{l,c_i}+\sum_{i}\theta_{d_i}\sum_{l\in\mathcal{L}_f}(-1)^{\alpha(l)}M_{l,d_i} \\
& = & N_{L_f}\sum_{i}\theta_{c_i}\label{ci}\, ,\end{aligned}$$
where $N_{L_f}$ is the number of anyons in $\mathcal{L}_f$. In the second and third line we have applied Schur’s orthogonality relations to the 1D irreps of $\mathcal{L}_f$. From this result, we see that there must exist at least one anyon $c_i$.
To show that we can find an anyon $c$ satisfying both properties $(1)$ ($M_{c,l}=(-1)^{\alpha(l)}$), and $(2)$ ($\theta_c=\theta_\phi^*$), we proceed as follows. Because the $c_ic_j$ braid trivially with all anyons in $\mathcal{L}_f$, it follows from the definition of a fermionic Lagrangian subgroup that $c_ic_j \in \mathcal{L}_f$. This implies that
$$\theta_{c_ic_j} = (-1)^{\alpha(c_ic_j)} = M_{c_i,c_ic_j}$$
Applying the ribbon identity to the left-hand side of this equation, we find
$$\theta_{c_i}\theta_{c_j} M_{c_i,c_j} = M_{c_i,c_i}M_{c_i,c_j} = \theta_{c_i}^2 M_{c_i,c_j} \Rightarrow \theta_{c_i} = \theta_{c_j}$$
Because the topological spins of all the $c_i$ are the same, expression (\[ci\]) for the chiral central charge becomes
$$e^{2\pi ic_-/8} = \frac{N_{L_f}}{\sqrt{N_b}}N_c \theta_\phi \theta_c \, ,$$
where $N_c$ is the number of $c_i$. From the assumption that the chiral central charge $c_-$ is a multiple of eight, we find that the topological spin of the $c_i$ indeed satisfies $\theta_c = \theta_\phi^*$. Because $N_{L_f}=\sqrt{N_b}$, it also follows that $N_c=1$, i.e. the anyon $c$ satisfying $\theta_c=\theta_\phi^*$ and $M_{l,c}=\theta_l$ for all $l\in\mathcal{L}_f$ is unique.
The ‘only if’ direction is almost trivial to show. If we assume that the GAfTO = $\mathcal{A}_b\times\{1,\phi,f,\phi f\}$ has a bosonic Lagrangian subgroup, we know that there exists a group of anyons $\mathcal{M}\subset\mathcal{A}_b$ that braid trivially with each other, and have the property that every $a\in\mathcal{A}_b$ which is not in $\mathcal{M}$ braids non-trivially with at least one anyon in $\mathcal{M}$. Because the anyons in $\mathcal{M}$ have trivial mutual braiding, it follows from the ribbon identity that their topological spins form a representation of $\mathcal{M}$: $\theta_{m_1}\theta_{m_2} = \theta_{m_1m_2}$ for all $m_1,m_2\in\mathcal{M}$. From the relation $\theta_{m}^2=M_{m,m}=1$ between topological spin and self-braiding, we also know that $\theta_{m}=\pm 1$ for all $m\in\mathcal{M}$. But this implies that $\mathcal{M}=\mathcal{L}_f$ is a fermionic Lagrangian subgroup of $\mathcal{A}_f=\mathcal{A}_b\times\{1,f\}$.
### Example: gauging U$(1)_4\times\overline{IQH}$
To illustrate the general result we revisit the example discussed above, i.e. the $\mathcal{A}_f=\mathbb{Z}_4\times\{1,f\}=\{a,a^2,a^3,1\}\times\{1,f\}$ AfTO. Regardless of how we choose the charge vector $t=(2t',-1)$, this AfTO always has a fermionic Lagrangian subgroup given by $\mathcal{L}_f=\{1,a^2\}$. Recall that $\theta_{a^2}=-1$, so this would not be a valid Lagrangian subgroup if the system were bosonic. We first consider the case where the charge vector is given by $t=\left(2,-1\right)^T$. This implies that $\sigma_{xy}=0$, and $q_a=1/2$. This model is known to have a gapped edge, as it is equivalent to a fermionic $\mathbb{Z}_2$ gauge theory [@FillingC] (i.e. a fermionic toric code [@GuWang]). Let us again denote the fermion parity flux obtained by adiabatic flux insertion as $\phi_{A}$. Because $\sigma_{xy}=0$, $\phi_{A}$ is a boson. Because $a^2$ has charge one (recall that $q_{a^p} = pt'/2$), it also follows that $M_{\phi_{A},a^2}=-1$. From this we see that $\mathcal{L}_b=\{1,a^2f,\phi_{A},\phi_{A}a^2f\}=\{1,a^2f\}\times\{1,\phi_A\}=\mathbb{Z}_2\times\mathbb{Z}_2$ is a bosonic Lagrangian subgroup of the GAfTO.
Let us now repeat this analysis for the case where $t=\left(0,-1\right)$. This choice of charge vector implies that $\sigma_{xy}=-1$ and $q_a=0$. This fTO is just the stacking of a $\sigma_{xy}=-1$ IQH state, and a purely bosonic U$(1)_4$ topological order (because all anyons in U$(1)_4$ have trivial fermion parity charge). The parity flux obtained by adiabatic flux insertion now has topological spin $\theta_{\phi_{A}}=e^{i\pi\sigma_{xy}/4}=e^{-i\pi/4}$. The flux $\phi_A$ also braids trivially with all anyons in $\mathcal{A}_b=\{1,a,a^2,a^3\}$ because they have zero charge. It is now easy to check that $\mathcal{L}_b=\{1,a\phi_{A},a^2f,a^3\phi_{A}f\}=\mathbb{Z}_4$ is the Lagrangian subgroup of the GAfTO. So in this example, the anyon $a$ plays the role of the special anyon $c$ which occured in the general proof above.
Non-Abelian parity flux
-----------------------
If the GfTO is non-Abelian, we have to use a generalization of Eq. (\[kit\]) to determine the chiral central charge modulo eight from the bulk data. Writing the quantum dimensions of the anyons in the GfTO order as $d_a$, the general expression for $c_-$ becomes [@Kitaev]:
$$\label{nonAb}
e^{2\pi i c_-/8} = \frac{1}{\mathcal{D}} \sum_a d_a^2 \theta_a\, ,$$
where the total quantum dimension is given by $\mathcal{D}=\sqrt{\sum_a d_a^2}$. In a GAfTO $\mathcal{A}_b\times\{1,\phi,f\}$, the only non-Abelian anyons are $\mathcal{A}_b\times \phi$, and these are therefore the only anyons which have a quantum dimension not equal to one. From the fusion rule $\phi\times\phi = 1+f$, it follows that $d_\phi = \sqrt{2}$. Applying Eq. (\[nonAb\]) to a non-Abelian GAfTO, we find
$$\label{eqdouble}
e^{2\pi i c_-/8} = \frac{\theta_\phi}{\sqrt{N_b}}\sum_{a\in\mathcal{A}_b} \theta_a$$
This is exactly the same expression as the one we obtained for Abelian fermion parity fluxes.
We can now easily argue that an AfTO with non-Abelian fermion parity fluxes can never have a gapped edge. First, we note that $N_b^{-1/2}\sum_{a\in\mathcal{A}_b}\theta_a$ is always an eighth root of unity. This is because Abelian topological orders have a multi-component U$(1)$ Chern-Simons description, such that the corresponding edge theories are chiral Luttinger liquids with integer chiral central charge [@XGWen]. On the other hand, if $\phi$ is non-Abelian, it has a topological spin $\theta_\phi = e^{2\pi i (2n+1)/16}$, where $n$ is an integer [@Kitaev]. So from Eq. (\[eqdouble\]), it follows that $c_-$ is a half odd integer, which means that the edge is always chiral and cannot be gapped.
Conclusions
===========
In this work we have explored the special structure of gauged Abelian fermionic topological orders, and we have exploited this structure to study both the numerical detection of such phases and the fate of the edge physics under the gauging process. We have outlined a minimal scheme to uniquely identify the AfTO realized by a microscopic lattice Hamiltonian, which does not make use of fermion number conservation symmetry. We have also shown that gauging the global fermion parity symmetry in a system with Abelian anyons will not remove or introduce gapless degrees of freedom on the boundary of the system.
An obvious question is of course how to generalize these results to systems with non-Abelian anyons. The mathematical framework required to address non-Abelian fermionic topological orders is developed and discussed in Refs. [@GuWangWen; @TianLan1; @TianLan2; @Wang; @Aasen], and it is substantially more involved than the simple arguments used in this work. However, the understanding of Abelian systems provides a clear, intuitive picture of the physics involved, and hopefully this will be helpful for a rigorous study of non-Abelian systems. We leave such a study for future work.
### Acknowledgements
I am grateful to Meng Cheng for pointing out theorem 3.13 from Ref. [@Drinfeld] to me, and for a collaboration on a previous project which was very useful for the present paper. I also want to thank Mike Zaletel for inspiring discussions which formed the motivation for the present work, and Johannes Motruk for discussions and for pointing me to some important references. During the completion of this work I was supported by the DOE, office of Basic Energy Sciences under contract no. DE-AC02-05-CH11231.
References {#references .unnumbered}
==========
[^1]: This is a slight abuse of terminology which is common in the literature. If we write $\theta_a=e^{2\pi i h_a}$, then it is more appropriate to call $h_a$ the topological spin. However, in this work it will be more convenient to simply refer to $\theta_a$ as the topological spin.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- Jianqiang Zhao
title: Standard Relations of Multiple Polylogarithm Values at Roots of Unity
---
[Department of Mathematics, Eckerd College, St. Petersburg, FL 33711]{}
0.6cm
[**Abstract.**]{} Let $N$ be a positive integer. In this paper we shall study the special values of multiple polylogarithms at $N$th roots of unity, called multiple polylogarithm values (MPVs) of level $N$. These objects are generalizations of multiple zeta values and alternating Euler sums, which was studied by Euler, and more recently, many mathematicians and theoretical physicists.. Our primary goal in this paper is to investigate the relations among the MPVs of the same weight and level by using the regularized double shuffle relations, regularized distribution relations, lifted versions of such relations from lower weights, and seeded relations which are produced by relations of weight one MPVs. We call relations from the above four families *standard*. Let $d(w,N)$ be the $\Q$-dimension of $\Q$-span of all MPVs of weight $w$ and level $N$. Then we obtain upper bound for $d(w,N)$ by the standard relations which in general are no worse or no better than the one given by Deligne and Goncharov depending on whether $N$ is a prime-power or not, respectively, except for 2- and 3-powers, in which case standard relations seem to be often incomplete whereas Deligne shows that their bound should be sharp by a variant of Grothedieck’s period conjecture. This suggests that in general there should be other linear relations among MPVs besides the standard relations, some of which are written down in this paper explicitly with good numerical verification. We also provide a few conjectures which are supported by our computational evidence.
0.6cm
Introduction
============
In recent years, there is a revival of interest in multi-valued classical polylogarithms (polylogs) and their generalizations. For any positive integers $s_1,\dots, s_\ell$, Goncharov [@Gicm] defines the multiple polylogs of complex variables as follows: $$\label{equ:polylog}
Li_{s_1,\dots, s_\ell}(x_1,\dots,x_\ell)=\sum_{k_1>\dots>k_\ell>0}
\frac{x_1^{k_1}\cdots x_\ell^{k_\ell}}{ k_1^{s_1}\cdots k_\ell^{s_\ell}}.$$ Conventionally one calls $\ell$ the [*depth*]{} (or [*length*]{}) and $s_1+\dots+s_\ell$ the [*weight*]{}. When the depth $\ell=1$ the function is nothing but the classical polylog. When the weight is also 1 we get the MacLaurin series of $-\log(1-x)$. Another useful expression of the multiple polylogs is given by the following iterated integral: $$\label{equ:iteratedForm}
Li_{s_1,\dots,s_\ell}(x_1,\dots,x_\ell)=(-1)^\ell
\int_{0}^1 \left(\frac{dt}{t}\right)^{\circ(s_1-1)}\circ\frac{d t}{t-a_1}
\circ\cdots \circ\left(\frac{dt}{t}\right)^{\circ(s_\ell-1)}\circ\frac{d t}{t-a_\ell}$$ where $a_i=1/(x_1\dots x_i)$ for $1\le i\le \ell$. Here, we define the iterated integrals recursively by $\int_a^b f(t)\circ w(t) =\int_a^b (\int_a^x f(t)) w(x)$ for any 1-form $w(t)$ and concatenation of 1-forms $f(t)$. We may think the path lies in $\C$; however, it is more revealing to use iterated integrals in $\C^\ell$ to find the analytic continuation of this function (see [@Zanamp]).
It is well-known that special values of polylogs have significant applications in arithmetic such as Zagier’s conjecture [@Zag p.622]. On the other hand, the multiple zeta values (MZV) appear naturally in the study of the fundamental group of ${{\mathbb P}}^1-\{0,1\infty\}$ which is closely related to the absolute Galois group $\operatorname{{Gal}}(\overline{\Q}/\Q)$ according the Grothendieck [@D1]. As pointed out by Goncharov, higher cyclotomy theory should study the multiple polylogs at roots of unity, not only those of the classical ones. Moreover, theoretical physicists have already found out that such values appear naturally in the study of Feynmen diagrams ([@Br1; @Br2]).
Starting from early 1990s Hoffman [@H1; @H2] has constructed some quasi-shuffle (we will call “stuffle”) algebras reflecting the essential combinatorial properties of MZVs. Recently he [@H3] extends this to incorporate the multiple polylog values (MPVs) at roots of unity, although his definition of $*$-product is different from ours. Our approach here is a quantitative comparison between the the results obtained by Racinet [@Rac] who considers MPVs from the motivic viewpoint of Drinfeld associators, and those by Deligne and Goncharov [@DG] who study the motivic fundamental groups of ${{\mathbb P}}^1-(\{0,\infty\}\cup\mmu_N)$ by using the theory of mixed Tate motives over $S$-integers of number fields, where $\mmu_N$ is the group of $N$th root of unity.
Fix an $N$th root of unity $\mu=\mu_N:=\exp(2\pi
\sqrt{-1}/N)$. The *level* $N$ MPVs are defined by $$\label{equ:z}
L_N(s_1,\dots,s_\ell|i_1,\dots,i_\ell):=
Li_{s_1,\dots,s_\ell}(\mu^{i_1},\dots,\mu^{i_\ell}).$$ We will always identify $(i_1,\dots,i_\ell)$ with $(i_1,\dots,i_\ell)
\pmod{N}.$ It is easy to see from that a MPV converges if and only if $(s_1,\mu^{i_1})\ne (1,1).$ Clearly, all level $N$ MPVs are automatically of level $Nk$ for any positive integer $k$. For example when $i_1=\cdots=i_\ell=0$ or $N=1$ we get the multiple zeta values $\zeta(s_1,\dots,s_\ell)$. When $N=2$ we recover the alternating Euler sums studied in [@BBB2; @Zesum]. To save space, if a substring $S$ repeats $n$ times in the list then $\{S\}^n$ will be used. For example, $L_N(\{2\}^2|\{0\}^2)=\zeta(2,2)=\pi^4/120$.
Standard conjectures in arithmetic geometry imply that $\Q$-linear relations among MVPs can only exist between those of the same weight. Let $\MPV(w,N)$ be the $\Q$-span of all the MPVs of weight $w$ and level $N$ whose dimension is denoted by $d(w,N)$. In general, to determine $d(w,N)$ precisely is a very difficult problem because any nontrivial lower bound would provide some nontrivial irrational/transcendental results which is related to a variant of Grothendieck’s period conjecture (see [@Del]). For example, we can easily show that $\MPV(2,4)=\langle \log^2 2, \pi^2, \pi \log 2\sqrt{-1},
(K-1)\sqrt{-1}\rangle,$ where $K=\sum_{n\ge 0} (-1)^n/(2n+1)^2$ is the Catalan’s constant. From Grothendieck’s conjecture we know $d(2,4)=4$ (see op. cit.) but we don’t have a unconditional proof yet. On the other hand, we may obtain upper bound of $d(w,N)$ by finding as many linear relations in $\MPV(w,N)$ as possible. As in the cases of MZVs and the alternating Euler sums the regularized double shuffle relations (RDS) play important roles in revealing the relations among MPVs. We shall study this theory for MPVs in section \[sec:RDS\] by generalizing some results of [@IKZ] (also cf. [@Bigotte]). It is commonly believed that in levels one and two all linear relations among MPVs are consequences of RDS.
From the point of view of Lyndon words and quasi-symmetric functions Bigotte et al. [@Bigotte] have studied MPVs (they call them *colored MZVs*) primarily by using double shuffle relations. However, when the level $N\ge 3$, these relations are not complete in general, as we shall see in this paper.
If the level $N>3$ then by a theorem of Bass [@Bass] there are many non-trivial linear relations (regarded as *seeds*) in $\MPV(1,N)$ whose structure is clear to us. Multiplied by MPVs of weight $w-1$ these relations can produce non-trivial linear relations among MPVs of weight $w$ which we call the *seeded relations.* Similar to these relations we may produce new relations by multiplying MPVs on RDS of lower weights. We call such relations *lifted relations*. We conjecture that when level $N=3$ all linear relations among MPVs are consequences of the RDS and the lifted RDS with $d(w,3)=2^w$.
Among MPVs we know that there are the so-called finite distribution relations (FDT). Racinet [@Rac] considers further the regularization of these relations by regarding MPVs as the coefficients of some group-like element in a suitably defined pro-Lie-algebra of motivic origin. Our computation shows that the regularized distribution relations (RDT) do contribute to new relations not covered by RDS and FDT. But they are not enough yet to produce all the lifted RDS.
We call a $\Q$-linear relation between MPVs *standard* if it can be produced by combinations of the following four families of relations: regularized double shuffle relations (RDS), regularized distribution relations (RDT), seeded relations, and lifted relations from the above. Otherwise, it is called a *non-standard* relation.
The main goal of this paper is to provide some numerical evidence concerning the (in)completeness of the standard relations. Namely, these relations in general are not enough to cover all the $\Q$-linear relation between MPVs (see Remark \[rem:Ta:bN\] and Remark \[rem:wt2\]); however, when the level is a prime $\le 47$ and weight $w=2$ using a result of Goncharov we can show that the standard relations are complete under the assumption of Grothendieck’s period conjecture (see [@Zocta]). We further find that when weight $w=2$ and $N=25$ or $N=49$, the standard relations are complete. However, when $N$ is a 2-power or 3-power or has at least two distinct prime factors, we know that the standard relations are often incomplete by comparing our results with those of Deligne and Goncharov [@DG]. Moreover, we don’t know how to obtain the non-standard relations except that when $N=4$, we discover recently that octahedral symmetry of ${{\mathbb P}}^1-(\{0,\infty\}\cup \mmu_4)$ can produce some (presumably all) new relations not covered by the standard ones (see op. cit.)
Most of the MPV identities in this paper are discovered with the help of MAPLE using symbolic computations. We have verified almost all relations by GiNaC [@GiNac] with an error bound $<10^{-90}$.
This work was started while I was visiting Chern Institute of Mathematics at Nankai University and the Morningside Center of Mathematics at Beijing, China in the summer of 2007. I would like to thank both institutions and my hosts Chengming Bai and Fei Xu for their hospitality and the ideal working environment. I also want to thank Jens Vollinga for answering some of my questions regarding the numerical computation of the multiple polylog values. The paper was revised later while I was visiting the Institute for Advanced Study and thanks are due to Prof. Deligne for his patient explanation of [@DG] and many insightful remarks on the paper. This work was partially supported by a faculty development fund from Eckerd College.
The double shuffle relations and the algebra $\fA$ {#sec:RDSoverC}
==================================================
It is Kontsevich [@K1] who first noticed that MZVs can be represented by iterated integrals (cf. [@Rac]). We now extend this to MPVs. Set $$a=\frac{dt}{t},\qquad b_i=\frac{\mu^i dt}{1-\mu^i t}\quad
\text{ for } i=0,1,\dots,N-1.$$ For every positive integer $n$ define $$y_{n,i}:=a^{n-1} b_i.$$ Then it is straight-forward to verify using that if $(s_1,\mu^{i_1})\ne (1,1)$ then (cf. [@Rac (2.5)]) $$\label{equ:mzv}
L_N(s_1,\dots,s_n|i_1,i_2,\dots,i_n)=\int_0^1
y_{s_1,i_1}y_{s_2,i_1+i_2}\cdots
y_{s_n,i_1+i_2+\dots+i_n}.$$ We now define an algebra of words as follows:
Set $A_0=\{\bfone\}$ to be the set of the empty word. Define $\fA=\Q\langle A\rangle$ to be the graded noncommutative polynomial $\Q$-algebra generated by letters $a$ and $b_i$ for $i\equiv 0,\dots,N-1\pmod{N}$, where $A$ is a locally finite set of generators whose degree $n$ part $A_n$ consists of words (i.e., a monomial in the letters) of depth $n$. Let $\fA^0$ be the subalgebra of $\fA$ generated by words not beginning with $b_0$ and not ending with $a$. The words in $\fA^0$ are called *admissible words.*
Observe that every MPV can be expressed uniquely as an iterated integral over the closed interval $[0,1]$ of an admissible word $w$ in $\fA^0$. Then we denote this MPV by $$\label{equ:Z}
Z(w):=\int_0^1 w.$$ Therefore we have (cf. [@Rac (2.5) and (2.6)]) $$\begin{aligned}
\label{equ:1-1LZ}
L_N(s_1,\dots,s_n|i_1,i_2,\dots,i_n)=&Z(
y_{s_1,i_1}y_{s_2,i_1+i_2}\cdots
y_{s_n,i_1+i_2+\dots+i_n}),\\
Z(y_{s_1,i_1}y_{s_2,i_2}\cdots
y_{s_n,i_n})=&L_N(s_1,\dots,s_n|i_1,i_2-i_1,\dots,i_n-i_{n-1}).
\label{equ:1-1ZL}\end{aligned}$$ For example $L_3(1,2,2|1,0,2) =Z(y_{1,1}y_{2,1}y_{2,0}).$ On the other hand, during 1960s Chen developed a theory of iterated integral which can be applied in our situation.
\[chen’s\] *([@Chen (1.5.1)])* Let $\om_i$ $(i\ge 1)$ be $\C$-valued 1-forms on a manifold $M$. For every path $p$, $$\int_p \om_1\cdots \om_r\int_p \om_{r+1}\cdots \om_{r+s}=
\int_p (\om_1\cdots \om_r) \sha (\om_{r+1}\cdots \om_{r+s})$$ where $\sha$ is the shuffle product defined by $$(\om_1\cdots \om_r) \sha (\om_{r+1}\cdots
\om_{r+s}):=\sum_{\substack{\gs\in
S_{r+s},\gs^{-1}(1)<\cdots<\gs^{-1}(r)\\
\gs^{-1}(r+1)<\cdots< \gs^{-1}(r+s)}}
\om_{\gs(1)}\cdots \om_{\gs(r+s)}.$$
For example, we have $$\begin{aligned}
\ &L_N(1|1)L_N(2,3|1,2)=Z(y_{1,1})Z(y_{2,1}y_{3,3})=Z(b_1\sha
(ab_1a^2b_3))\\
=&Z(b_1ab_1a^2b_3+2ab_1^2a^2b_3+(ab_1)^2ab_3+ab_1a^2b_1b_3+ab_1a^2b_3b_1)\\
=&Z(y_{1,1}y_{2,1}y_{3,3}+2y_{2,1}y_{1,1}y_{3,3}+y_{2,1}^2y_{2,3}
+y_{2,1}y_{3,1}y_{1,3}+y_{2,1}y_{3,3}y_{1,1})\\
=&L_N(1,2,3|1,0,2)+2L_N(2,1,3|1,0,2)+L_N(2,2,2|1,0,2)\\
\ &\hskip3cm +L_N(2,3,1|1,0,2)+L_N(2,3,1|1,2,N-2).\end{aligned}$$
Let $\fA_\sha$ be the algebra of $\fA$ together with the multiplication defined by shuffle product $\sha$. Denote the subalgebra $\fA^0$ by $\fA_\sha^0$ when we consider the shuffle product. Then we can easily prove
\[shahomo\] The map $Z: \fA_\sha^0\lra \C$ is an algebra homomorphism.
On the other hand, it is well known that MPVs also satisfy the series stuffle relations. For example $$L_N(2|5)L_N(3|4)=L_N(2,3|5,4)+L_N(3,2|4,5)+L_N(5|9).$$ because $$\sum_{j>0}\sum_{k>0}=\sum_{j>k>0}+\sum_{k>j>0}+\sum_{j=k>0}.$$ To study such relations in general we need the following definition.
Denote by $\fA^1$ the subalgebra of $\fA$ which is generated by words $y_{s,i}$ with $s\in \Z_{>0}$ and $i\equiv 0,\dots,N-1\pmod{N}$. Equivalently, $\fA^1$ is the subalgebra of $\fA$ generated by words not ending with $a$. For any word $w=y_{s_1,i_1}y_{s_2,i_2}\cdots
y_{s_n,i_n}\in \fA^1$ and positive integer $j$ we define the exponent shifting operator $\tau_j$ by $$\tau_j(w)=y_{s_1,j+i_1}y_{s_2,j+i_2}\cdots y_{s_n,j+i_n}.$$ For convenience, on the empty word we have the convention that $\tau_j(\bfone)=\bfone.$ We then define a new multiplication $*$ on $\fA^1$ by requiring that $*$ distribute over addition, that $\bfone*w=w*\bfone=w$ for any word $w$, and that, for any words $\om_1,\om_2$, $$\begin{gathered}
\label{equ:defnstuffle}
y_{s,j}\om_1*y_{t,k}\om_2 = y_{s,j}\Big(\tau_j\big(\tau_{-j}(\om_1)*y_{t,k}\om_2\big)\Big)
+ y_{t,k}\Big(\tau_k\big( y_{s,j}\om_1*\tau_{-k}(\om_2)\big)\Big) \\
+y_{s+t,j+k}\Big(\tau_{j+k}\big(\tau_{-j}(\om_1)*\tau_{-k}(\om_2)\big)\Big).\end{gathered}$$ We call this multiplication the *stuffle product*.
Our $\fA$, $\fA^0$ and $\fA^1$ are related to $\QX$, $\QX_\cv$ and $\QY$ of [@Rac], respectively. See section \[sec:Rac\].
If we denote by $\fA_*^1$ the algebra $(\fA^1,*)$ then it is not hard to show that (cf. [@H2 Thm. 2.1])
The polynomial algebra $\fA_*^1$ is a commutative graded $\Q$-algebra.
Now we can define the subalgebra $\fA_*^0$ similar to $\fA_\sha^0$ by replacing the shuffle product by stuffle product. Then by induction on the lengths and using the series definition we can quickly check that for any $\om_1,\om_2\in \fA_*^0$ $$Z(\om_1)Z(\om_2)=Z(\om_1\ast \om_2).$$ This implies that
\[\*homo\] The map $Z: \fA_*^0 \lra \C$ is an algebra homomorphism.
For $\om_1,\om_2\in \fA^0$ we will say that $$Z(\om_1\sha \om_2-\om_1*\om_2)=0$$ is a finite double shuffle (FDS) relation. It is known that even in level one these relations are not enough to provide all the relations among MZVs. However, it is believed that one can remedy this by considering RDS produced by the following mechanism. This was explained in detail in [@IKZ] when Ihara, Kaneko and Zagier considered MZVs where they call these extended double shuffle relations.
Combining Propositions \[\*homo\] and \[shahomo\] we can prove easily (cf. [@IKZ Prop. 1]):
\[prop:eDS\] We have two algebra homomorphisms: $$Z^*: (\fA_*^1,*)\lra \C[T],\quad \text{and}\quad Z^\sha: (\fA_\sha^1,\sha)\lra \C[T]$$ which are uniquely determined by the properties that they both extend the evaluation map $Z:\fA^0\lra \C$ by sending $b_0=y_{1,0}$ to $T$.
In order to establish the crucial relation between $Z^*$ and $Z^\sha$ we can adopt the machinery in [@IKZ]. For any $(\bs|\bi)=(s_1,\dots,s_n|i_1,\dots,i_n)$ where $i_j$’s are integers and $s_j$’s are positive integers, let the image of the corresponding words in $\fA^1$ under $Z^*$ and $Z^\sha$ be denoted by $Z_{(\bs|\bi)}^*(T)$ and $Z_{(\bs|\bi)}^\sha(T)$ respectively. For example, $$\begin{aligned}
TL_N(2|3)=&Z_{(1|0)}^*(T)Z_{(2|3)}^*(T)=Z^*(y_{1,0}*y_{2,3})\\
=&Z_{(1,2|0,3) }^*(T)+Z_{(2,1|3,3) }^*(T)+Z_{(3|3)}^*(T),\end{aligned}$$ while $$\begin{aligned}
TL_N(2|3)=&Z_{(1|0)}^\sha(T)Z_{(2|3)}^\sha(T)=Z^\sha(y_{1,0}\sha
y_{2,3})=Z^\sha(b_0\sha ab_3)\\
=&Z_{(1,2|0,3)}^\sha(T)+Z_{(2,1|0,3)}^\sha(T) +Z_{(2,1|3,0)}^\sha(T) .\end{aligned}$$ Hence we find the following RDS by the next Theorem: $$L_N(2,1|3,0) +L_N(3|3)=L_N(2,1|3,N-3)+L_N(2,1|0,3).$$
\[thm:RDSoverC\] Define a $\C$-linear map $\rho:\C[T]\to \C[T]$ by $$\rho(e^{Tu})=\exp\left(\sum_{n=2}^\infty\frac{(-1)^n}{n}\zeta(n)u^n\right)e^{Tu},\qquad |u|<1.$$ Then for any index set $(\bs|\bi)$ we have $$Z_{(\bs|\bi)}^\sha(T)= \rho(Z_{(\bs|\bi)}^*(T)).$$
This is a the generalization of [@IKZ Thm. 1] to the higher level MPV cases. The proof is essentially the same. One may compare Cor. 2.24 in [@Rac]. The above steps can be easily transformed to computer codes which are used in our MAPLE programs.
Finite and regularized double shuffle relations (FDS & RDS) {#sec:RDS}
===========================================================
It is generally believed that all the linear relations between MZVs can be derived from RDS. Although the naive generalization of this to arbitrary levels is wrong the idea in [@IKZ] to formalize this via some universal objects is still very useful. We want to generalize this idea to MPVs in this section.
Keep the same notation as in the preceding sections. Let $R$ be a commutative $\Q$-algebra with 1 and $Z_R:\fA^0\lra R$ such that the “finite double shuffle” (FDS) property holds: $$Z_R(\om_1\sha \om_2)=Z_R(\om_1* \om_2)=Z_R(\om_1)Z_R(\om_2).$$ We then extend $Z_R$ to $Z_R^\sha$ and $Z_R^*$ as before. Define an $R$-module $R$-linear automorphism $\rho_R$ of $R[T]$ by $$\rho_R(e^{Tu})=A_R(u)e^{Tu}$$ where $$A_R(u)=\exp\left(\sum_{n=2}^\infty \frac{(-1)^n}{n}
Z_R(a^{n-1} b_0)u^n\right) \in R[\![u]\!].$$ Similar to the situation for MZVs, we may define the $\fA^0$-algebra isomorphisms $$\operatorname{{reg}}_\sha^T:\fA_\sha^1 =\fA_\sha^0[b_0]\lra\fA_\sha^0[T],\qquad
\operatorname{{reg}}_*^T:\fA_*^1 =\fA_*^0[b_0]\lra\fA_*^0[T],$$ which send $b_0$ to $T$. Composing these with the evaluation map $T=0$ we get the maps $\operatorname{{reg}}_\sha$ and $\operatorname{{reg}}_*$.
\[thm:RDS\] Let $(R,Z_R)$ be as above with the FDS property. Then the following are equivalent:
- $(Z_R^\sha-\rho_R\circ Z_R^*)(w)=0$ for all $w\in \fA^1$.
- $(Z_R^\sha-\rho_R\circ Z_R^*)(w)|_{T=0}=0$ for all $w\in \fA^1$.
- $ Z_R^\sha(\om_1\sha \om_0-\om_1*\om_0)=0$ for all $\om_1\in
\fA^1$ and all $\om_0\in \fA^0$.
- $ Z_R^*(\om_1\sha \om_0-\om_1*\om_0)=0$ for all $\om_1\in
\fA^1$ and all $\om_0\in \fA^0$.
- $ Z_R(\operatorname{{reg}}_\sha(\om_1\sha \om_0-\om_1*\om_0))=0$ for all $\om_1\in
\fA^1$ and all $\om_0\in \fA^0$.
- $ Z_R(\operatorname{{reg}}_*(\om_1\sha \om_0-\om_1*\om_0))=0$ for all $\om_1\in
\fA^1$ and all $\om_0\in \fA^0$.
- $ Z_R(\operatorname{{reg}}_\sha(b_0^m*w))=0$ for all $m\ge 1$ and all $w\in \fA^0$.
- $ Z_R(\operatorname{{reg}}_*(b_0^m\sha w-b_0^m*w))=0$ for all $m\ge 1$ and all $w\in \fA^0$.
If $Z_R$ satisfies any one of these then we say that $Z_R$ has the *regularized double shuffle* (RDS) property.
Notice that RDS automatically implys FDS. The proof of the theorem is almost the same as that of [@IKZ Thm. 2] but for completeness we give the most important details in the following because there is some subtle difference for MPVs of arbitrary level.
Denote by $\caS$ the set of the $y_{s,j}$ ($s\in \Z_{>0},
j=0,\dots, N-1$). For convenience we write $\tau_z=\tau_j$ if $z=y_{s,j}\in\caS$. If $w=y_{s_1,i_1}\dots y_{s_n,i_n}\in \fA^1$ then we put $\tau_w=\tau_{i_1+\dots+i_n}$ and $\tau_{-w}=\tau_{-i_1-\dots-i_n}$. Then (cf. [@IKZ Prop. 2])
\[prop:gdz\] We have
1. For $z\in \caS$ the map $\gd_z:\fA^1\to \fA^1$ defined by $$\gd_z(w):=z*w-z\tau_z(w)$$ is a “twisted derivation” in the sense that $$\gd_z(ww')=\gd_z(w)\tau_z(w')+w\tau_w\Big(\gd_z\big(\tau_{-w}(w')\big)\Big).$$ Moreover, all these twisted derivations commute.
2. The above twisted derivations extend to a twisted derivation on all of $\fA$ after setting $\tau_a=\text{id}$, with values on the letters $a,
b_j$ given by $$\gd_z(a)=0, \quad \gd_z(b_j)=(a+b_j)\tau_j(z)\quad (z\in \caS, j=0,\dots,N-1).$$ In particular, $\gd_z$ preserves $\fA^0$.
Easy computation by Definition .
\[cor:gdz\] Denote by $\zz$ the $\Q$-linear span of the $y_{s,0}$ ($s\in
\Z_{>0}$). Then for $z\in \zz$ the map $\gd_z$ is a derivation on $\fA^1$ which preserves $\fA^0$. Moreover, $\gd_z$ can be extended to a derivation on $\fA$ by Prop. \[prop:gdz\](ii).
Define $\gd'$ as in Prop. \[prop:gdz\](ii). For any $z=y_{s,0}\in\zz$ and $y_{t,i}$ a generator of $\fA^1$ we have $$\gd'_z(y_{t,i})=a^{t-1}\gd'_z(b_i)=a^{t-1}(a+b_i)y_{s,i}
=y_{s,0}*y_{t,i}-y_{s,0}y_{t,i}=\gd_z(y_{t,i}).$$ Note that for $z\in \zz$ and $w, w'\in\fA$ we have $\tau_w\Big(\gd_z\big(\tau_{-w}(w')\big)\Big)=\gd_z(w')$. So indeed $\gd_z$ can be extended to a derivation on $\fA$. It’s obvious that $\gd_z$ fixes both $\fA^0$ and $\fA^1$.
We can define another operation on $\caS$ by $y_{s,i}\circ y_{t,j}=y_{s+t,i+j}.$ We can then restrict this to $\{y_{s,0}: s\in\Z_{>0}\}$ then extend linearly to $\zz$. The following result is then straight-forward.
The vector space $\zz$ becomes a commutative and associative algebra with respect to the multiplication $\circ$ defined by $$z\circ z'=z*z'-zz'-z'z.$$
The following proposition is one of the keys to the proof of Theorem \[thm:RDS\] (cf. [@IKZ Prop. 4]).
\[circ\] Let $u$ be a formal variable. For $z\in \caS$ we have $$\exp(zu\tau_z) (\bfone) =(2-\exp_{\circ}(z u\tau_z) )^{-1} (\bfone).$$ (The inverse on the right is with respect to the concatenation product.)
Define power series $$f(u \tau_z)=\exp_\circ(z u\tau_z )-1=z u\tau_z +z\circ z \frac{u^2\tau_z^2}{2}+\cdots$$ Then taking derivative with respect to $u$ we get $f'(u\tau_z)=z\circ(1+f(u \tau_z))\tau_z.$ Now for $z,\om_i\in\caS$ we have by Prop. \[prop:gdz\] and Prop. \[circ\] $$z*(\om_1\om_2\cdots \om_n)=\sum_{i=0}^n \om_1\cdots
\om_i\tau_{\om_i}(z)\tau_z(\om_{i+1}\cdots \om_n)
+\sum_{i=1}^n \om_1\cdots
\om_{i-1}( z\circ \om_i)\tau_z(\om_{i+1}\cdots \om_n).$$ This yields $$\begin{aligned}
\ & z*\left(z^{\circ n_1}\frac{(u \tau_z)^{n_1}}{n_1!}\cdots
z^{\circ n_d}\frac{(u \tau_z)^{n_d}}{n_d!}\right)\\
=& \sum_{i=0}^d \left(z^{\circ n_1}\frac{(u
\tau_z)^{n_1}}{n_1!}\cdots
z^{\circ n_i}\frac{(u \tau_z)^{n_i}}{n_i!}\right)z\tau_z \left(z^{\circ n_{i+1}}\frac{(u
\tau_z)^{n_{i+1}}}{n_{i+1}!}\cdots
z^{\circ n_d}\frac{(u \tau_z)^{n_d}}{n_d!}\right)\\
+&\sum_{i=1}^d
\left(z^{\circ n_1}\frac{(u \tau_z)^{n_1}}{n_1!}\cdots
z^{\circ (n_i+1)}\frac{(u \tau_z)^{n_i}}{n_i!}\right)\tau_z\left(z^{\circ n_{i+1}}\frac{(u
\tau_z)^{n_{i+1}}}{n_{i+1}!}\cdots
z^{\circ n_d}\frac{(u \tau_z)^{n_d}}{n_d!}\right).\end{aligned}$$ Hence $$z *\big(1-f(u \tau_z)\big)^{-1}\tau_z(\bfone)=\frac{d}{du}\Big(\big(1-f(u
\tau_z)\big)^{-1} \Big) (\bfone).$$ This implies that $$\exp_*(zu \tau_z ) (\bfone)=\big(1-f(u\tau_z)\big)^{-1} (\bfone)$$ as desired.
\[cor:key\] For all $z\in \caS$ we have $$\exp_*(\log_\circ (1+z\tau_z)) (\bfone)=\big(1-z\tau_z\big)^{-1} (\bfone).$$
If $z\in \zz$ then $\tau_z=\text{id}$ and therefore we have
\[cor:circ\] For $z\in \zz$ we have $$\exp(zu)=(2-\exp_{\circ}(z u) )^{-1},\quad \ \exp_*(\log_\circ
(1+z))=(1-z)^{-1}.$$
Let’s consider a non-trivial example of Cor. \[cor:key\]. Let $N=2$ and $z=y_{k,1}$ then we have $$\exp\left(\sum_{n=1}^\infty (-1)^{n-1}
\zeta\big(nk;(-1)^k\big)\frac{u^n}{n}\right)=1+\sum_{n=1}^\infty
\zeta\big(\{k\}^n;\{-1\}^n\big) u^n$$ where $\zeta(s_1,\dots,s_d;(-1)^{\gs_1},\dots,(-1)^{\gs_d})=L_2(
s_1,\dots,s_d|\gs_1,\dots,\gs_d)$ are the alternating Euler sums. For example, by comparing the coefficients of $u^2$ and $u^3$ we get $2\zeta(\ol{k},\ol{k})=\zeta(\ol{k})^2-\zeta(2k),$ and $6\zeta(\ol{k},\ol{k},\ol{k})=\zeta(\ol{k})^3-3\zeta(\ol{k})\zeta(2k)+2\zeta(\ol{3k}).$ Here $\ol{s_j}$ means that the corresponding $\gs_j$ is odd.
The following two propositions are generalizations of Prop. 5-6 of [@IKZ] respectively whose computational proofs are mostly omitted since nothing new happens in our situation.
\[prop:zz’\] For $z,z'\in \zz$ and $w\in \fA^1$ we have $$\begin{aligned}
\label{equ:gdzz'}
\exp(\gd_z)(z')=&\big(\exp_\circ(z)\circ z'\big)\exp_*(z),\\
\exp(\gd_z)(w)=&(\exp_*(z))^{-1}\big(\exp_*(z)* w\big).\end{aligned}$$
\[prop:Phi\] For $z\in \zz$ define $\Phi_z:\fA^1\to \fA^1$ by $$\label{equ:Phi}
\Phi_z(w):=(1-z)\left(\big(1-z\big)^{-1}*w\right)\qquad (w\in
\fA^1).$$ Then $\Phi_z$ is an automorphism of $\fA^1$ and we have $$\label{equ:Phiexp}
\Phi_z(w)=\exp(\gd_t)(w), \qquad\text{where }t=\log_\circ(1+z)\in
\zz.$$ All the $\Phi_z$ commute. Moreover, after restricting the derivation $\gd_t$ to $\fA^0$ we can regard $\Phi_z$ as an automorphism of $\fA^0$, If we extend the derivation $\gd_t$ to the whole $\fA$ as in Cor. \[cor:gdz\] then we can regard $\Phi_z$ as an automorphism of $\fA$.
The key point is that $\gd_t$ sends $\fA^0$ to $\fA^0$ as a derivation by Cor. \[cor:gdz\]. Hence $\exp(\gd_t)$ is an automorphism on $\fA^1$ as well as on $\fA^0$.
The next three results are generalizations of Prop. 7, its corollary, and Prop. 8 of [@IKZ], respectively. The proofs there can be easily adapted into our situation because the $\sha$-product is essentially the same (note that the the only essentially new phenomenon in the higher level MPV cases is that there are exponent shiftings on the roots of unity in our stuffle product.)
\[prop:d\] Define the map $d:\fA\to \fA$ by $d(w)=b_0\sha w-b_0 w$. Then $d$ is a derivation and by setting $u$ as a formal parameter we have $$\exp(du)(w)=(1-b_0u)\left(\big(1-b_0u\big)^{-1}\sha w\right)\qquad (w\in
\fA^1).$$ On the generators we have $$\label{equ:dgens}
\exp(du)(a)=a\big(1-b_0u\big)^{-1},\qquad\exp(du)(b_j)=b_j\big(1-b_0u\big)^{-1},\
j=0,\dots,N-1.$$
In fact, we can replace the whole $\fA$ by $\fA^1$ in the first part of Prop. \[prop:d\]. We can do the same in the next corollary. However, in the proof of Theorem \[thm:RDS\] we only need this weaker version.
\[cor:Delta\] Let $u$ be a formal parameter. Let $\gD_u=\exp(-du)\circ \Phi_{b_0u}\in {\rm Aut}(\fA)[\![u]\!]$ (here $\circ$ means the composition). Then $$(1-b_0u)^{-1}*w=(1-b_0u)^{-1} \sha \gD_u(w), \qquad \forall
w\in\fA^1.$$ In particular, for $w\in \fA^0$ by taking $\operatorname{{reg}}$ on both sides of the above equation we get $$\operatorname{{reg}}_\sha\Big((1-b_0u)^{-1}*w\Big)= \gD_u(w).$$
\[prop:reg\] For $\om_0=a\om_0' \in\fA^0$ we have $$\operatorname{{reg}}_\sha^T\left(\big(1-b_0u\big)^{-1} \om_0\right)=
\exp(-du)(\om_0)e^{Tu}=a\left(\big(1+b_0u\big)^{-1}\sha \om_0'\right)e^{Tu}.$$
Theorem \[thm:RDS\] now follows easily from a detailed computation as in [@IKZ]. As a matter of fact, the same argument shows that [@IKZ Prop. 10] and its Cor. are both valid in our general setup if we replace $\HH^0$ there by $\fA^0$.
Seeded (or weight one) relations {#sec:seed}
================================
When $N\ge 4$ there exist linear relations among MPVs of weight one by a theorem of Bass [@Bass]. These relations are important because by multiplying any MPV of weight $w-1$ by such a relation we can get a relation between MPVs of weight $w$ which we call a *seeded relation.* This is one of the key ideas in finding the formula in [@DG 5.25] concerning $d(w,N)$.
First, we know there are $N-1$ MPVs of weight 1 and level $N$: $$L_N(1|j)=-\log(1-\mu^j), \qquad 0<j<N ,$$ where $\mu=\mu_N=\exp(2\pi\sqrt{-1}/N)$ as before. Here we have taken $\C-(-\infty,0]$ as the principle domain of the logarithm. Further, it follows from the motivic theory of classical polylogs developed by Deligne and Beilinson and the Borel’s theorem (see [@G1 Thm. 2.1]) that the $\Q$-dimension of $\MPV(1,N)$ is $$d(1,N)=\dim K_1(\Z[\mu_N][1/N])\otimes \Q+1=\varphi(N)/2+\nu(N),$$ where $\varphi$ is the Euler’s totient function and $\nu(N)$ is the number of distinct prime factors of $N$. Hence there are many linear relations among $L_N(1|j)$. For instance, if $j<N/2$ then we have the symmetric relation $$-\log(1-\mu^j)=-\log(1-\mu^{N-j})-\log(-\mu^j)=
-\log(1-\mu^{N-j})+\frac{N-2j}{N}\pi\sqrt{-1}.$$ Thus for all $1<j<N/2$ $$\label{equ:symrel}
(N-2)(L_N(1|j)-L_N(1|N-j))=(N-2j)(L_N(1|1)-L_N(1|N-1)).$$ Further, from [@Bass (B)] for any divisor $d$ of $N$ and $1\le a<d':=N/d$ we have the distribution relation $$\label{equ:prodrel}
\sum_{0\le j<d} L_N(1|a+jd')= L_N(1|ad).$$ It follows from the main result of Bass [@Bass] corrected by Ennola [@En] that all the linear relations between $L_N(1|j)$ are consequences of and . Hence the seeded relations have the following forms in words: for all $w\in \fA^0$ $$\label{equ:seed}
\left\{\aligned
(N-2)Z(y_{1,j}*w-y_{1,-j}*w)=&(N-2j)(Z(y_{1,1}*w-y_{1,-1}*w),\\
\sum_{0\le j<d} Z(y_{1,a+jd'}*w)=& Z(y_{1,ad}*w).
\endaligned\right.$$
Regularized distribution relations {#sec:Rac}
==================================
Multiple polylogs satisfy the following distribution formula (cf. [@Rac Prop. 2.25]): $$\label{equ:dist}
Li_{s_1,\dots,s_n}(x_1,\dots,x_n) = d^{s_1+\dots+s_n-n}
\sum_{y_j^d=x_j,1\le j\le n} Li_{s_1,\dots,s_n}(y_1,\dots,y_n),$$ for all positive integer $d$. When $s_1=1$ we need to exclude the case of $x_1=1$. We call these *finite distribution relations* (FDT). Racinet further considers the regularized version of these relations, which we now recall briefly.
Fix an embedding $\mmu_N\hookrightarrow \C$ and denote by $\gG$ its image. Define two sets of words $$\bfX:=\bfX_\gG=\{x_\gs: \gs\in \gG\cup\{0\}\}, \quad\text{and}\quad
\bfY:=\bfY_\gG=\{y_{n,\gs}=x_0^{n-1}x_\gs: n\in \N, \gs\in \gG\}.$$ Then one may consider the coproduct $\gD$ of $\Q\langle\bfX\rangle$ defined by $\gD x_\gs=1\otimes x_\gs+x_\gs\otimes 1$ for all $\gs\in \gG\cup\{0\}$. For every path $\gam\in {{\mathbb P}}^1(\C)-(\{0,\infty\}\cup\gG)$ Racinet defines the group-like element $\calI_\gam\in \CXX$ by $$\calI_\gam:=\sum_{p\in \N,\gs_1,\dots,\gs_p\in \gG\cup\{0\}}
\calI_\gam(\gs_1,\dots,\gs_p)x_{\gs_1}\cdots x_{\gs_p},$$ where $\calI_\gam(\gs_1,\dots,\gs_p)$ is the iterated integral $\int_\gam \om(\gs_1) \cdots \om(\gs_p)$ with $$\om(\gs)(t)=\left\{
\begin{array}{ll}
\gs\,dt/(1-\gs t), & \hbox{if $\gs\ne 0$;} \\
dt/t, & \hbox{if $\gs=0$.}
\end{array}
\right.$$ This $I_\gam$ is essentially the same element denoted by $\operatorname{{dch}}$ in [@DG]. Note that $\QY$ is the sub-algebra of $\QX$ generated by words not ending with $x_0$. We let $\pi_\bfY:\QX\ra \QY$ be the projection. As $x_0$ is primitive one knows that $(\QY,\gD)$ has a graded co-algebra structure.
Let $\QX_\cv$ be the sub-algebra of $\QX$ not beginning with $x_1$ and not ending with $x_0$. Let $\pi_\cv:\QX\ra \QX_\cv$ be the projection. Passing to the limit one get:
*([@Rac Prop.2.11])* The series $\calI_\cv:=\lim_{a\to 0^+,b\to 1^-}\pi_\cv(\calI_{[a,b]})$ is group-like in $(\CXX_\cv,\gD)$.
Let $\calI$ be the unique group-like element in $(\CXX,\gD)$ whose coefficients of $x_0$ and $x_1$ are 0 such that $\pi_\cv(\calI)=\calI_\cv$. In order to do the numerical computation we need to find out explicitly the coefficients for $\calI$. Put $$\calI=\sum_{p\in \N, ,\gs_1,\dots,\gs_p\in \gG\cup\{0\}}
C(\gs_1,\dots,\gs_p)x_{\gs_1}\cdots x_{\gs_p}.$$
\[prop:CoeffI\] Let $p,$ $m$ and $n$ be three non-negative integers. If $p>0$ then we assume $\gs_1\ne 1$ and $\gs_p\ne 0$. Set $(\gs_1,\dots,\gs_p,\{0\}^n) =(\gs_1,\dots,\gs_q)$. Then we have $$\begin{aligned}
\notag
\ C(\{1\}^m,\gs_1,& \dots,\gs_p,\{0\}^n)\\
=& \left\{
\begin{array}{ll}
0, {\displaystyle \phantom{\frac1m}} & \hbox{if $mn=p=0$;} \\
Z\big(\pi_\bfY(x_{\gs_1}\cdots x_{\gs_p})\big) , {\displaystyle \phantom{\frac1m}} & \hbox{if $m=n=0$;} \\
{\displaystyle -\frac1m \sum_{i=1}^q C(\{1\}^{m-1},\gs_1,\dots,\gs_i,1,\gs_{i+1},\dots,\gs_q),} & \hbox{if $m>0$;} \\
{\displaystyle -\frac1n \sum_{i=1}^p C(\gs_1,\dots,\gs_{i-1},0,\gs_i,\dots,\gs_p,\{0\}^{n-1}),} & \hbox{if $m=0,n>0$.}
\end{array}
\right. \label{equ:CoeffI}
\end{aligned}$$ Here $Z$ is defined by .
This proposition provides the recursive relations we may use to compute all the coefficients of $\calI$.
Since $\calI$ is group-like we have $$\label{equ:Igam}
\gD \calI=\calI\otimes\calI.$$ The first case follows from this immediately since $C(0)=C(1)=0$. The second case is essentially the definition of $Z$. If $m>0$ then we can compare the coefficient of $x_1\otimes x_1^{m-1}x_{\gs_1}\cdots x_{\gs_q}$ of the two sides of and find the relation . Finally, if $m=0$ and $n>0$ then we may similarly consider the coefficient of $x_{\gs_1}\cdots x_{\gs_p}x_0^{n-1}\otimes x_0$ in . This finishes the proof of the proposition.
For any divisor $d$ of $N$ let $\gG^d=\{\gs^d:\gs\in \gG\}$, $i_d:\gG^d\hookrightarrow \gG$ the embedding, and $p^d:\gG\twoheadrightarrow \gG^d$ the $d$th power map. They induce two algebra homomorphisms: $$\begin{aligned}
{6}
p^d_\ast: \Q\langle\bfX_\gG\rangle & \lra\Q\langle\bfX_{\gG^d}\rangle & {} & &\hskip1cm
i_d^\ast: \Q\langle\bfX_\gG\rangle & \lra\Q\langle\bfX_{\gG^d}\rangle \\
x_\gs &\lms \begin{cases} d x_0, \ &\text{ if }\gs=0,\\
x_{\gs^d}, \ &\text{ if }\gs\in \gG.\end{cases}
&\qquad\overset{\displaystyle\text{ and }}{\phantom{\sum}} & &
x_\gs &\lms \begin{cases} x_0, \ &\text{ if }\gs=0,\\
x_\gs, \ &\text{ if }\gs\in \gG^d,\\
0, \ &\text{ otherwise}.\end{cases}\end{aligned}$$
It is easy to see that both $i_d^\ast$ and $p^d_\ast$ are $\gD$-coalgebra morphisms such that $i_d^\ast(\calI)$ and $p^d_\ast(\calI)$ have the same image under the map $\pi_\cv$. By the standard Lie-algebra mechanism one has
\[prop:RDist\] *([@Rac Prop.2.26])* For every divisor $d$ of $N$ $$\label{equ:RDist}
p^d_\ast(\calI)=\exp\left(\sum_{\gs^d=1,\gs\ne 1} Li_1(\gs) x_1\right) i_d^\ast(\calI).$$
Combined with Proposition \[prop:CoeffI\] the above result provides the so-called *regularized distribution relations* (RDT) which of course include all the FDT of MPVs given by .
Computation suggests the the following conjecture concerning a special class of distribution relations.
\[conj:distr\] Let $d$ be a positive integer. Then all the distribution relations in , where $x_j=1$ for all $j$, are consequences of RDS of MPVs of level $d$.
We are able to confirm this conjecture in the special case that $w=2$, $n=1$, and $d$ is a prime.
Write $L(i,j)=L_p(1,1|i,j)$ and $D(i)=L_p(2|i)$. Define for $1\le i,j<p$: $$\begin{aligned}
\ & \text{FDT}:=
D(0)-p\sum_{j=0}^{p-1}D(j)=0, \qquad
\text{RDS$(i)$}:=
D(i)+L(i,0)-L(i,-i)=0,\\
\ &\text{FDS$(i,j)$}:=
D(i+j)+L(i,j)+L(j,i)-L(i,j-i)-L(j,i-j)=0.\end{aligned}$$ Then $$\label{equ:distNull}
{\rm FDT}=\sum_{1\le i<p} {\rm FDS}(i,i)
+2\sum_{1\le j<i<p} {\rm FDS}(i,j)+2\sum_{i=1}^{p-1} {\rm RDS}(i,i).$$
By changing the order of summation we see that $$\begin{aligned}
2\sum_{1\le j<i<p} D(i+j)=& \sum_{i=2}^{p-1}\sum_{j=1}^{i-1} D(i+j) +
\sum_{j=1}^{p-2}\sum_{i=j+1}^{p-1} D(i+j)\\
=&\sum_{i=2}^{p-2}\sum_{i\ne j=1}^{p-1} D(i+j)
+ \sum_{j=1}^{p-2} D(j-1) +\sum_{i=2}^{p-1} D(i+1)\\
=&(p-3)\sum_{j=0}^{p-1} D(j)-\sum_{i=2}^{p-2} D(i)-\sum_{i=1}^{p-1} D(2i)
+ \sum_{j=1}^{p-2} D(j) +\sum_{j=2}^{p-1} D(j)+2D(0)\\
=&(p-1)D(0)+(p-3)\sum_{j=1}^{p-1} D(j)\end{aligned}$$ since $\sum_{j=0}^{p-1} D(i+j)=\sum_{j=1}^{p-1} D(j)$ for all $i$ and $\sum_{i=1}^{p-1} D(2i)=\sum_{i=1}^{p-1} D(i)$. This implies that the dilogarithms on the right hand side of exactly add up to FDT. Thus we only need to show that all the double logarithms on the right hand side of cancel.
First we note that $L(i,0)$ in ${\rm FDS}(i,i)$ and ${\rm RDS}(i,i)$ cancel. Now let us consider the lattice points $(i,j)$ of $\Z^2$ corresponding to $L(i,j)$. The points $(i,j)$ corresponding to $L(i,j)$ with positive signs fill in exactly the area inside the square $[1,p-1]\times [1,p-1]$ (boundary inclusive): $L(i,i)$ in ${\rm FDS}(i,i)$ provides the diagonal $y=x$, $\sum_{1\le j<i<p} L(i,j)$ (resp. $\sum_{1\le j<i<p} L(j,i)$) form the lower right (resp. upper left) triangular region.
For the negative terms of the double logs, $L(i,-i)$ in ${\rm RDS}(i)$ provides the diagonal $x+y=p$, $\sum_{1\le j<i<p} L(i,j-i)=\sum_{i=2}^{p-1}\sum_{j=p+1-i}^{p-1}L(i,j)$ form the upper right triangular region. Similarly, by changing the order of summation $\sum_{1\le j<i<p} L(j,i-j)=\sum_{i=1}^{p-2}\sum_{j=i+1}^{p-1}L(i,j-i)
=\sum_{i=2}^{p-1}\sum_{j=1}^{p-1-i}L(i,j)$ fills the lower left region.
Further, numerical evidence up to level $N=49$ supports the following
\[conj:regdistr\] In weight two, all RDT are consequences of RDS and FDT.
Lifted relations from lower weights
===================================
Note that when $N=3$ there are no seeded relations nor (regularized) distribution relations. When we deal with MZVs and alternating Euler sums we expect that all the linear relations come from RDS. Are these enough when $N=3$? Surprisingly, the answer is no.
The first counterexample is in weight four, i.e., $(w,N)=(4,3)$. Easy computation shows that there are 144 MPVs in this case among which there are 239 nontrivial RDS which include 191 FDS. Using these relations we get 127 independent linear relations among the 144 MPVs. But the upper bound of $d(4,3)$ by [@DG 5.25] is 16, so there must be at least one more linearly independent relation. Where else can we find it? It is easy to verify that all the seven RDT (including four FDT) can be derived from RDS. However, we know that a product of two weight two MPVs is of weight four. So on each of the five RDS (including two FDS) in $\MPV(2,3)$ we can multiply any one of the nine MPVs of $(w,N)=(2,3)$ to get a relation in $\MPV(4,3)$. For instance, we have a FDS $$Z(y_{1,1}\ast y_{1,1}-y_{1,1}\sha y_{1,1})
=L_3(2|2)+2L_3(1,1|1,1)-L_3(1,1|1,0)=0.$$ Multiplying by $L_3(1,1|1,1)=Z(y_{1,1}y_{1,2})$ we have a new relation not derivable from RDS in $\MPV(4,3)$: $$\begin{aligned}
& Z\big(y_{1,1}y_{1,2} \sha[y_{2,0}+2y_{1,1}y_{1,2}-2y_{1,1}y_{1,0}] \big)\\
=&L_3(1,1,2|1,1,0)+2L_3(1,2,1|1,1,0)+2L_3(2,1,1|1,1,0)+L_3(2,1,1|2,2,1)
+4L_3(\{1\}^4|1,1,2,1)\\
+&8L_3(\{1\}^4|1,0,1,0)-6L_3(\{1\}^4|1,0,0,1)
-4L_3(\{1\}^4|1,0,1,2)-2L_3(\{1\}^4|1,1,2,0)=0.\end{aligned}$$ Such relations coming from the lower weights are called *lifted relations (from lower weights).* In this way, when $(w,N)=(4,3)$ we can produce 45 lifted RDS relations from weight two, 58 from weight three. We may also lift RDT and obtain nine and six relations from weight two and three, respectively. However, all the lifted relations together only produce one new linearly independent relation, as expected. Hence we find totally 128 linearly independent relations among the 144 MPVs of $(w,N)=(4,3)$. This implies that $d(4,3)\le 16$ which is the same bound obtained by [@DG 5.25] and is proved to be exact under a variant of Grothendieck’s period conjecture by Deligne [@Del].
For general levels $N$ we may lift not only RDS and RDT but also the seeded relations. But a moment reflection tells us that the lifted seeded relations are seeded so we don’t need to consider these after all.
We call a $\Q$-linear relation between MPVs *standard* if it can be produced by combinations of the following four families of relations: regularized double shuffle relations (RDS), regularized distribution relations (RDT), seeded relations, and lifted relations from the above. Otherwise, it is called a *non-standard* relation.
There are no seeded relations if $N=3$. In this case we believe that all the linear relations among MPVs come from RDS and the lifted relations (see Conjecture \[conj:mainconj\]). Moreover, computation in small weight cases supports the following
\[conj:level3\] Suppose $N=3$ or $4$. Every MPV of level $N$ is a linear combination of MPVs of the form $L(\{1\}^w|t_1,\dots,t_w)$ with $t_j\in
\{1,2\}$. Consequently, the $\Q$-dimension of the MPVs of weight $w$ and level $N$ is given by $d(w,N)=2^w$ for all $w\ge 1$.
\[rem:notenough\] Even adding all the lifted relations from lower weights does not provide all the linear relations among MPVs. A quick look at the Table \[Ta:dbzeta\] in §\[sec:comp\] tells us that if $(w,N)=(3,4)$ even though we know $d(3,4)\le 8$ and $d(4,4)\le 16$ by [@DG 5.25], and the equality should hold by Conjecture \[conj:level3\] or by a variant of Grothendieck’s period conjecture (see Remark \[rem:Del\]), we cannot produce enough relations by using the standard ones. Instead, we can only show that $d(3,4)\le 9$ and $d(4,4)\le 21$. More recently, by using octahedral symmetry of ${{\mathbb P}}^1-(\{0,\infty\}\cup \mmu_4)$ we find (presumably all) the non-standard relations in these two cases (see [@Zocta]).
\[rem:Del\] Let $N=2,3,4$ or $8$. Assuming a variant of Grothendieck’s period conjecture, Deligne [@Del] constructs explicitly a set of basis for $\MPV(w,N)$. His results would imply that $d(w,2)$ is given by the Fibonacci numbers, $d(w,3)=d(w,4)=2^w$, and $d(w,8)=3^w.$
Some conjectures of FDS and RDS
===============================
Recall that if a map $Z_R:\fA^0\lra R$ satisfies the FDS and any one of the equivalent conditions in Theorem \[thm:RDS\] then we say that $Z_R$ has the *regularized double shuffle* (RDS) property. Let $R_{RDS}$ be the universal algebra (together with a map $Z_{RDS}:\fA^0\lra R_{RDS}$) such that for every $\Q$-algebra $R$ and a map $Z_R:\fA^0\lra R$ satisfying RDS there always exists a map $\varphi_R$ to make the following diagram commutative: $$\text{$\diagramcompileto{diagwt2}
\fA^0 \drto_{Z_R}\rto^-{Z_{RDS}}& R_{RDS} \dto^-{\varphi_R}\\
\ & R
\enddiagram$}$$
When $N=3$ computation shows that the lifted relations contribute non-trivially when the weight $w=5$: we can only get $d(5,3)\le 33$ instead of the conjecturally correct dimension 32 without using lifted relations. We may say that $Z_R$ has the *lifted regularized double shuffle* (LRDS) property if it satisfies RDS and for all $\om_1\in \fA^1$ and $\om_0,\om_0',\om_0''\in \fA^0$ $$Z_R(Z_R^{-1}\circ \rho_R\circ Z_R(\om_1)*\om_0-\om_1*\om_0)=Z_R(
(\om_0*\om_0')*\om_0''-(\om_0\sha \om_0')*\om_0'')=0.$$ We can define $Z_{SR}$ and $R_{SR}$ corresponding to the standard relations similar to $Z_{RDS}$ and $R_{RDS}$ such that for every $\Q$-algebra $R$ and a map $Z_R:\fA^0\lra R$ satisfying the standard relations there always exists a map $\varphi_R$ to make the following diagram commutative: $$\label{equ:cm}
\text{$\diagramcompileto{diagwt2}
\fA^0 \drto_{Z_R}\rto^-{Z_{SR}}& R_{SR} \dto^-{\varphi_R}\\
\ & R
\enddiagram$}$$
\[conj:mainconj\] Let $(R,Z_R)=(\R, Z)$ if $N=1,2$ and $(R,Z_R)=(\C, Z)$ if $N=p$ is a prime $\ge 3$, where $Z$ is given by . If $N=1$ or $2$ then the map $\varphi_\R$ is injective, namely, the algebra of MPVs is isomorphic to $R_{RDS}$. If $N=p$ is a prime $\ge 3$ then the map $\varphi_\C$ is injective so the algebra of MPVs of level $p$ is isomorphic to $R_{SR}$. Moreover, if $N=3$ then $Z_{LRDS}=Z_{SR}$ and $R_{LRDS}=R_{SR}$.
From Conjecture \[conj:mainconj\] all the linear relations among MPVs can be produced by RDS when $N=1$ or $2$, and by the standard ones when $N=p$ is prime $\ge 3$. When $p\ge 5$ this is proved in [@Zocta] under the assumption of Grothendieck’s period conjecture.
Computation in many cases such as those listed in Remark \[rem:Ta:bN\] and \[rem:wt2\] show that MPVs must satisfy some other relations besides the standard ones when $N$ has more than two distinct prime factors, so a naive generalization of Conjecture \[conj:mainconj\] to all levels does not exist at present. However, when $N=4$ we find that octahedral symmetry of ${{\mathbb P}}^1-(\{0,\infty\}\cup \mmu_4)$ may provide all the non-standard relations (see [@Zocta]). But since we only have numerical evidence in weight 3 and 4 it may be a little premature to form a conjecture for level four at present.
The structure of MPVs and some examples {#sec:comp}
=======================================
In this section we concentrate on RDS between MPVs of small weights. Most of the computations in this section are carried out by MAPLE. We have checked the consistency of these relations with many known ones and verified our results numerically using GiNac [@GiNac] and EZ-face [@EZface].
By considering all the admissible words we see easily that the number of distinct MPVs of weight $w\ge 2$ and level $N$ is $N^2(N+1)^{w-2}$ and there are at most $N(N+1)^{w-2}$ RDS (but not FDS). If $w\ge 4$ then the number of FDS is given by $$(N-1)N^2(N+1)^{w-3}+\Big(\Big[\frac
w2\Big]-1\Big)N^4(N+1)^{w-4}=\Big(N^2\Big[\frac
w2\Big]-1\Big)N^2(N+1)^{w-4}.$$ If $w=2$ (resp. $w=3$) then the number of FDS is $(N-1)^2$ (resp. $N^2(N-1)$).
Weight one. {#sec:wt1}
-----------
From §\[sec:seed\] we know that all relations in weight one follow from and , and no RDS exists. The relations in weight one are crucial for higher level cases because they provide the seeded relations considered in §\[sec:seed\]. Moreover, easy computation by and shows that there is a hidden integral structure, namely, in each level there exists a $\Q$-basis consisting of MPVs such that every other MPV is a $\Z$-linear combination of the basis elements. This fact is proved by Conrad [@Conrad Theorem 4.6]. Similar results should hold for higher weight cases and we hope to return to this in a future publication [@Zint].
Weight two.
-----------
There are $N^2$ MPVs of weight 2 and level $N$: $$L_N(1,1|i,j),\quad L_N(2|j), \qquad 1\le i\le N-1, 0\le j\le N-1.$$ For $1\le i,j<N$ the FDS $Z^*(y_{1,i}*y_{1,j})=Z^\sha(y_{1,i}\sha
y_{1,j})$ yields $$\label{equ:wt2FDS}
L_N(2|i+j)+L_N(1,1|i,j)+L_N(1,1|j,i)=L_N(1,1|i,j-i)+L_N(1,1|j,i-j).$$ Now from RDS $\rho(Z^*(y_{1,0}*y_{1,i}))= Z^\sha(y_{1,0}\sha
y_{1,i})$ we get for $1\le i<N$ $$\label{equ:wt2RDS}
L_N(1,1|i,0)+L_N(2|i)=L_N(1,1|i,-i).$$ The FDT in yields: for every divisor $d$ of $N$, and $1\le a,b< d':=N/d$ $$\begin{aligned}
\label{equ:wt2dist1}
L_N(2|ad)=&d\sum_{j=0}^{d-1}L_N(2|a+jd'), \\
L_N(1,1|ad,bd)=&\sum_{j,k=0}^{d-1} L_N(1,1|a+jd',b+k d').\label{equ:wt2dist2}\end{aligned}$$ To derive the RDT we can compare the coefficients of $x_1x_{\mu^{ad}}$ in and use Prop. \[prop:CoeffI\] to get: for every divisor $d$ of $N$, and $1\le a< d'$ $$\begin{gathered}
\label{equ:wt2RDT}
L_N(1|ad)\sum_{j=1}^{d-1}L_N(1|jd')
=\sum_{j=1}^{d-1}\sum_{k=0}^{d-1} L_N(1,1|jd',a+kd')\\
-\sum_{k=0}^{d-1} L_N(1,1|a+kd',-a-kd')-L_N(1,1|ad,-ad).\end{gathered}$$ By definition, the seeded relations are obtained from and . For example, if $N=p$ is a prime then is trivial and is equivalent to: for all $1\le j<h:=(p-1)/2$ $$\label{equ:symrelprime}
L_N(1|j)-L_N(1|-j)=(p-2j)(L_N(1|h)-L_N(1|h+1)).$$ Thus multiplying by $L_N(1|i)$ ($1\le i<p$) and applying the shuffle relation $L_N(1|a)L_N(1|b)=L_N(1,1|a,b-a)+L_N(1,1|b,a-b)$ we get: $$\begin{gathered}
\label{equ:wt2N=pseed}
L_N(1,1|i,j-i)+L_N(1,1|j,i-j)-L_N(1,1|i,-j-i)-L_N(1,1|-j,i+j)\\
=(p-2j)\big(L_N(1,1|i,h-i)+L_N(1,1|h,i-h)
-L_N(1,1|i,-i-h)-L_N(1,1|-h,i+h) \big).\end{gathered}$$
Computation shows that the following conjecture should hold.
The RDT follows from the combination of the following relations: the seeded relations, the RDS and , and the FDT and .
Weight three.
-------------
Apparently there are $N^2(N+1)$ MPVs of weight 3 and level $N$: for each choice $(i,j,k)$ with $1\le i\le N-1, 0\le j,k\le N-1$ we have four MPVs of level $N$: $$L_N(1,1,1|i,j,k), \quad L_N(1,2|i,j), \quad L_N(2,1|j,k), \quad L_N(3|k).$$ For $1\le i,j,k<N$ the FDS $Z^*\big(y_{1,i}*(y_{1,j}y_{1,k})\big)=Z^\sha\big(y_{1,i}\sha(y_{1,j}y_{1,k})\big)$ yields $$\begin{gathered}
\label{equ:wt3DS1}
L_N(\{1\}^3,i,j-i,k)+L_N(\{1\}^3,j,i-j,k+j-i)+L_N(\{1\}^3,j,k,i-k-j)\\
=L_N(2,1,i+j,k)+L_N(1,2,j,i+k)\hskip4cm\ \\
+L_N(\{1\}^3,i,j,k)+L_N(\{1\}^3,j,i,k)+L_N(\{1\}^3,j,k,i).\end{gathered}$$ For $1\le i,j<N$ the FDS $Z^*(y_{1,i}*y_{2,j})=Z^\sha(y_{1,i}\sha
y_{2,j})$ yields $$\label{equ:wt3DS2}
\aligned \ & L_N(3,i+j)+L_N(1,2,i,j)+L_N(2,1,j,i)\\
=&L_N(1,2,i,j-i)+L_N(2,1,i,j-i)+L_N(2,1,j,i-j).
\endaligned$$ Moreover, there are three ways to produce RDS. Since $\gb(T)=T$ the first family of RDS come from $Z^*\big(y_{1,0}*(y_{1,i}y_{1,i+j})\big)=Z^\sha\big(y_{1,0}\sha
(y_{1,i}y_{1,i+j})\big)$ for $1\le i\le N-1, 0\le j\le N-1$: $$\begin{aligned}
\ &y_{1,0}*(y_{1,i}y_{1,i+j})
=y_{1,0}y_{1,i}y_{1,i+j}+y_{1,i}
\tau_i(y_{1,0}*y_{1,j})+y_{2,i}y_{1,i+j}\\
=&y_{1,0}y_{1,i}y_{1,i+j}+y_{1,i}y_{1,i}y_{1,i+j}+y_{1,i}y_{1,i+j}y_{1,i+j}
+y_{1,i}y_{2,i+j}+y_{2,i}y_{1,i+j}\end{aligned}$$ On the other hand, $$y_{1,0}\sha y_{1,i}y_{1,i+j}
=y_{1,0}y_{1,i}y_{1,i+j}+y_{1,i}y_{1,0}y_{1,i+j}+y_{1,i}y_{1,i+j}y_{1,0}.$$ Hence $$\begin{gathered}
\label{equ:wt3RDS1}
L_N(\{1\}^3|i,0,j)+L_N(\{1\}^3|i,j,0)+L_N(1,2|i,j)+L_N(2,1|i,j)\\
=L_N(\{1\}^3|i,-i,i+j)+L_N(\{1\}^3|i,j,-i-j).$$\end{gathered}$$
The second family of RDS follow from $\gb(Z^*(y_{1,0}*y_{2,i}))=
Z^\sha(y_{1,0}\sha y_{2,i})$: $$y_{1,0}y_{2,i}+y_{2,i}y_{1,i}+y_{3,i}=y_{1,0}y_{2,i}+y_{2,0}y_{1,i}+y_{2,i}y_{1,0}$$ which implies that $$\label{equ:wt3RDS2}
L_N(2,1,i,0)+L_N(3,i)=L_N(2,1,i,-i)+L_N(2,1,0,i).$$
Now we consider the last family of RDS. By the definition of stuffle product: $$\begin{aligned}
y_{1,0}*y_{1,0}*y_{1,i}=&(2y_{1,0}^2+y_{2,0})*y_{1,i}\\
=&2y_{1,0}(y_{1,0}*y_{1,i})+2y_{1,i}^3+2y_{2,i}y_{1,i}+y_{2,0}*y_{1,i} \\
=&2y_{1,0}^2y_{1,i}+2y_{1,0}y_{1,i}^2+2y_{1,0}y_{2,i}+2y_{1,i}^3
+2y_{2,i}y_{1,i}+y_{2,0}*y_{1,i}.\end{aligned}$$ Applying $\gb\circ Z^*$ and noticing that $Z^\sha_{(2|0)}(T)=\zeta(2)$ we get $$\begin{gathered}
\label{equ:Z*wt3}
(T^2+\zeta(2))Z^\sha_{(1|i)}(T)=
2Z^\sha_{(1,1,1|0,0,i)}(T)+2Z^\sha_{(1,1,1|0,i,i)}(T)+2Z^\sha_{(1,2|0,i)}(T)\\
+2Z^\sha_{(1,1,1|i,i,i)}(T)
+2Z^\sha_{(2,1|i,i)}(T)+Z^\sha_{(2|0)}(T)Z^\sha_{(1|i)}(T).\end{gathered}$$ On the other hand by the definition of shuffle product $$\begin{aligned}
\om_0\sha \om_0\sha y_{1,i} =& 2\om_0^2 \sha \om_i
=2 \om_0^2 \om_i+2 \om_0 \om_i \om_0+ 2\om_i \om_0^2\\
=&2y_{1,0}^2y_{1,i}+2y_{1,0}y_{1,i}y_{1,0}+2y_{1,i}y_{1,0}^2\end{aligned}$$ Applying $Z^\sha$ we get $$\label{equ:Zshawt3}
T^2Z^\sha_{(1|i)}(T)=2Z^\sha_{(1,1,1|0,0,i)}(T)+2Z^\sha_{(1,1,1|0,i,0)}(T)
+2Z^\sha_{(1,1,1|i,0,0)}(T).$$ We further have $$\begin{aligned}
\ & Z^\sha(y_{1,0}y_{1,i}^2+y_{1,0}y_{2,i}-y_{1,0}y_{1,i}y_{1,0})\\
=&Z^\sha{(1,1,1|0,i,i)}(T)+Z^\sha_{(1,2|0,i)}(T)-Z^\sha_{(1,1,1|0,i,0)}(T)\\
=&2Z^\sha_{(1,1,1|i,0,0)}(T)-Z^\sha_{(2,1|i,0)}(T)-Z^\sha_{(2,1|0,i)}(T)
-Z^\sha_{(1,1,1|i,0,i)}(T) -Z^\sha_{(1,1,1|i,i,0)}(T)\end{aligned}$$ where we have used the facts that $$\begin{aligned}
Z^\sha_{(1,2|0,i)}(T)=&TZ^\sha_{(2|i)}(T)-Z^\sha_{(2,1|i,0)}(T)-Z^\sha_{(2,1|0,i)}(T)\\
Z^\sha_{(1,1,1|0,i,i)}(T)=&TZ^\sha_{(1,1|i,i)}(T)-Z^\sha_{(1,1,1|i,0,i)}(T)
-Z^\sha_{(1,1,1|i,i,0)}(T)\\
Z^\sha_{(1,1,1|0,i,0)}(T)=&TZ^\sha_{(1,1|i,0)}-2Z^\sha_{(1,1,1|i,0,0)}(T)\\
Z^\sha_{(1,1|i,0)} = &Z^\sha_{(2|i)}(T)+Z^\sha_{(1,1|i,i)}(T).\end{aligned}$$ Hence for $1\le i<N$ we have by subtracting from $$\begin{gathered}
\label{equ:wt3RDS3}
L_N(\{1\}^3|i,0,0)+L_N(2,1|i,0)+L_N(\{1\}^3|i,-i,0)=\\
L_N(2,1|i,-i)+L_N(2,1|0,i)+L_N(\{1\}^3|i,-i,i)+L_N(\{1\}^3|i,0,-i).\end{gathered}$$ Setting $j=0$ in and subtracting from we get $$\begin{gathered}
\label{equ:wt3RDS4}
L_N(\{1\}^3|i,-i,0)=
L_N(2,1|i,-i)+L_N(2,1|0,i)+L_N(\{1\}^3|i,0,0)+L_N(1,2|i,0).\end{gathered}$$
Upper bound of $d(w,N)$ by Deligne and Goncharov
------------------------------------------------
By using the theory of motivic fundamental groups of ${{\mathbb P}}^1-(\{0,\infty\}\cup \mmu_N)$ Deligne and Goncharov [@DG 5.25] show that $d(w,N)\le D(w,N)$ where $D(w,N)$ are defined by the formal power series $$\label{equ:DGbound}
1+\sum_{w=1}^\infty D(w,N)t^w=
\left\{
\begin{array}{ll}
(1-t^2-t^3)^{-1}, & \hbox{if $N=1$;} \\
(1-t-t^2)^{-1}, & \hbox{if $N=2$;} \\
\big(1-\big(\frac{\varphi(N)}2+\nu(N)\big)t+
\big(\nu(N)-1 \big)t^2 \big)^{-1}, & \hbox{if $N\ge 3$.} \\
\end{array}
\right.$$ Here $\varphi$ is the Euler’s totient function and $\nu(N)$ is the number of distinct prime factors of $N$. Set $a=a(N):=\varphi(N)/2+\nu(N)$ and $b=b(N):=\nu(N)-1$. If $N>2$ then we have $$\sum_{j=1}^\infty D(w,N) t^n
= at+(a^2-b)t^2+(a^3-2ab)t^3+(a^4-3a^2b+b^2)t^4+(3ab^2-4a^3b+a^5)t^5
+ \cdots$$ We will compare the bound obtained by standard relations to $D(w,N)$ in the next two sections.
Computational results in weight two {#sec:wt2comp}
===================================
In this section we combine the analysis in the previous sections and the theory developed by Deligne and Goncharov [@DG] to present a detailed computation in weight two and level $N\le 49$.
Let $\calG:=\iota(\operatorname{{Lie}}U_\om)$ be the motivic fundamental Lie algebra (see [@DG (5.12.2)]) associated to the motivic fundamental group of ${{\mathbb P}}^1-(\{0,\infty\}\cup \mmu_N)$. As pointed out in §6.13 of op. cit. one may safely replace $\calG(\mmu_N)^{(\ell)}$ by $\calG$ throughout [@G1]. Then it follows from the proof of [@DG 5.25] that if a variant of Grothendieck’s period conjecture (see 5.27(c) of op. cit.) is true, which we assume in the following, then $$\label{equ:d(2,N)}
d(2,N)=D(2,N)-\dim \ker(\gb_N),$$ where $\gb_N: \bigwedge^2 \calG_{-1,-1} \lra \calG_{-2,-2}$ is given by Ihara’s bracket $\gb_N(a \wedge b)=\{a,b\}$ defined by (5.13.6) of op. cit. Here $\calG_{\bullet,\bullet}$ is the associated graded of the weight and depth gradings of $\calG$ (see [@G1 §2.1]). Let $k(N):=\dim \ker(\gb_N)$. Then $$\label{equ:delta1}
\gd_1(N):=\dim \calG_{-1,-1}=\varphi(N)/2+\nu(N)-1$$ by [@G1 Thm. 2.1]. Thus $$\label{equ:iN}
i(N):=\dim \operatorname{{Im}}(\gb_N)=\gd_1(N)(\gd_1(N)-1)/2-k(N).$$ Since $\dim \calG_{-2,-1}=\varphi(N)/2$ if $N>2$ and $0$ otherwise the dimension of the degree 2 part of $\calG$ is $$\label{equ:delta2}
\gd_2(N):=\dim \calG_{-2,-1}+\dim\calG_{-2,-2}=
\left\{
\begin{array}{ll}
i(N), & \hbox{if $N=1$ or 2;} \\
\varphi(N)/2+i(N), & \hbox{if $N\ge 3$.}
\end{array}
\right.$$ Let $sr(N)$ be the upper bound of $\gd_2(N)$ obtained by the standard relations. This can be computed by the method described in [@Zocta]. Let $SR(N)$ be the upper bound of $d(2,N)$ similarly obtained by standard relations. In Table \[Ta:dbzeta22\] we use MAPLE to provide the following data: $k(N)$, $sr(N)$, and $SR(N)$. Then we can calculate $\gd_1$, $\gd_2$ and $i(N)$ from to . From we can check the consistency by verifying $$sr(N)-\gd_2(N)=SR(N)-d(2,N)=SR(N)-D(2,N)+k(N)$$ which gives the number of linearly independent non-standard relations (assuming Grothendieck’s period conjecture). To save space we use $D=D(2,N)$ and $d=d(2,N)$.
$N$ 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19
--------- --- --- --- --- --- --- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- -----
$\gd_1$ 0 1 1 1 2 2 3 2 3 3 5 3 6 4 5 4 8 4 9
$i$ 0 0 0 0 0 1 1 1 3 3 5 3 8 6 10 6 16 6 21
$k$ 0 0 0 0 1 0 2 0 0 0 5 0 7 0 0 0 12 0 15
$\gd_2$ 0 1 1 1 2 2 4 3 6 5 10 5 14 9 14 10 24 9 30
$sr$ 0 1 1 1 2 2 4 4 6 6 10 8 14 12 16 16 24 19 30
$D$ 1 2 4 4 9 8 16 9 16 15 36 15 49 24 35 25 81 24 100
$SR$ 1 2 4 4 8 8 14 10 16 16 31 18 42 27 37 31 69 34 85
$d$ 1 2 4 4 8 8 14 9 16 15 31 15 42 24 35 25 69 24 85
: Upper bound of $d(2,N)$ obtained by standard relations and [@DG 5.25].[]{data-label="Ta:dbzeta22"}
$N$ 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35
--------- ---- ---- ---- ----- ---- ----- ---- ----- ---- ----- ---- ----- ----- ----- ----- -----
$\gd_1$ 5 7 6 11 5 10 7 9 7 14 6 15 8 11 9 13
$i$ 10 21 15 33 10 40 21 36 21 56 15 65 28 55 36 78
$k$ 0 0 0 22 0 5 0 0 0 35 0 40 0 0 0 0
$\gd_2$ 14 27 20 44 14 50 27 45 27 70 19 80 36 65 44 90
$sr$ 24 32 35 44 32 50 42 54 48 70 48 80 64 77 72 96
$D$ 35 63 48 144 35 121 63 100 63 225 47 256 81 143 99 195
$SR$ 45 68 58 122 53 116 78 109 84 190 76 216 109 158 127 201
$d$ 35 63 48 122 35 116 63 100 63 190 47 216 81 143 99 195
: Upper bound of $d(2,N)$ obtained by standard relations and [@DG 5.25].[]{data-label="Ta:dbzeta22"}
$N$ 36 37 38 39 40 41 42 43 44 45 46 47 48 49
--------- ----- ----- ----- ----- ----- ----- ----- ----- ----- ----- ----- ----- ----- -----
$\gd_1$ 7 18 10 13 9 20 8 21 11 13 12 23 9 21
$i$ 21 96 45 78 36 210 28 133 55 78 66 161 36 175
$k$ 0 57 0 0 0 70 0 77 0 0 0 92 0 35
$\gd_2$ 27 114 54 90 44 140 34 154 65 90 77 184 44 196
$sr$ 72 114 89 112 96 140 96 154 120 144 132 184 128 196
$D$ 63 361 120 195 99 441 79 484 143 195 168 576 99 484
$SR$ 108 304 156 217 151 371 141 407 198 249 223 484 183 449
$d$ 63 304 120 195 99 371 79 407 143 195 168 484 99 449
: Upper bound of $d(2,N)$ obtained by standard relations and [@DG 5.25].[]{data-label="Ta:dbzeta22"}
\[rem:wt2\] We now make the following observations in weight two case.
\(a) If the level $N$ is a prime then the standard relations provide all the $\Q$-linear relations under the assumption of a variant of Grothendieck’s period conjecture. This is proved in [@Zocta Thm. 1].
\(b) Notice that when $p\ge 11$ the vector space $\ker\beta_p$ contains a subspace isomorphic to the space of cusp forms of weight two on $X_1(p)$ which has dimension $(p-5)(p-7)/24$ (see [@G1 Lemma 2.3 & Theorem 7.8]). So it must contain another piece which has dimension $(p-3)/2$ since $\dim(\ker\beta_p)=(p^2-1)/24$ by [@Zocta (5)]. What is this missing piece?
\(b) If $N$ is a $2$-power or a $3$-power then $D(2,N)$ should be sharp by the conjecture mentioned in (a). See Remark \[rem:Del\].
\(c) If $N$ has at least two distinct prime factors then $D(2,N)$ seems to be sharp, though we don’t have any theory to support it.
\(d) Suppose Grothendieck’s period conjecture is true. Then by [@DG 5.27], (b) and (c) is equivalent to saying that the kernel of $\gb_N$ is trivial if $N$ is a 2-power or a 3-power, or has at least two distinct prime factors. We believe this condition on $N$ for $\gb_N$ to be trivial is necessary, too.
\(e) If the level $N$ is a $p$-power for some prime $p\ge 5$ then $\gb_N$ is unlikely to be injective (the prime square case is proved in [@Zocta Prop. 5.3]). We conjecture that non-standard relation doesn’t exist (i.e., $SR(N)$ is sharp), though we only have verified the first two prime squares, $N=25$ and $N=49$.
\[rem:Ta:bN\] In the three cases $(w,N)=(2,8), (2,10)$ and $(2,12)$ we see that $d(w,N)>D(w,N)$. By numerical computation we conjecture that the bounds given by $D(w,N)$ are sharp in these cases and the following relations are the non-standard ones: let $L_N(-)=L_N(1,1|-)$ and $L_N^{(2)}(-)=L_N(2|-)$, then [$$\begin{aligned}
37L_8(1,1)=&34L_8^{(2)}(5)+112L_8(3,1)+11L_8(3,0)+37L_8^{(2)}(1)-2L_8(2,6)\notag \\
& +3L_8(7,3)-111L_8(5,7)+38 L_8(7,7)-8L_8(5,5),\label{equ:conj2}\\
7L_{10}(5,2)=&72L_{10}^{(2)}(1)+265L_{10}^{(2)}(7)-7L_{10}(2,5)-467L_{10}(4,2)+467L_{10}(8,6)\notag \\
&+14L_{10}(5,6)+64L_{10}(9,8)-164L_{10}(9,4)+166L_{10}(7,9)\notag \\
& -260L_{10}(8,1)-66L_{10}(3,9) -7L_{10}(6,9)+7L_{10}(6,5).\label{equ:conj3}\\
L_{12}(8,7)=&5L_{12}^{(2)}(5)+8L_{12}(8,10)-6L_{12}(10,11)-8L_{12}(9,11)+L_{12}(10,9)\notag \\
\ & -15L_{12}(8,1)+5L_{12}(9,10)+5L_{12}(6,1) -L_{12}(1,1)\notag \\
\ &+6L_{12}(8,11)-11L_{12}(6,11)+8L_{12}(8,3)-L_{12}(11,8),\\
60L_{12}(8,11)=&38L_{12}(8,7)+348L_{12}(10,11)+502L_{12}(9,11)-492L_{12}(10,9)\notag \\
\ & +600L_{12}(8,1)-552L_{12}(9,10)-154L_{12}(11,10)+20L_{12}(6,1)\notag \\
\ & +261L_{12}(6,11)-502L_{12}(8,3)+221L_{12}(11,8)-319L_{12}(8,10),\\
221L_{12}(1,1)=&1854L_{12}(8,10)+562L_{12}(8,7)-1018L_{12}(10,11)-2416L_{12}(9,11)\notag \\
\ & +319L_{12}(10,9)-4270L_{12}(8,1)+2293L_{12}(9,10)+956L_{12}(11,10)\notag \\
\ &+1110L_{12}(6,1)+2416L_{12}(8,11)-3305L_{12}(6,11)+2416L_{12}(8,3).\end{aligned}$$]{}
We call the level $N$ *standard* if either (i) $N=1, 2$ or $3$, or (ii) $N$ is a prime power $p^n$ ($p\ge 5$). Otherwise $N$ is called *non-standard*.
When $N$ is a non-standard level we find that very often there are non-standard relations among MPVs. For examples, the five relations in Remark \[rem:Ta:bN\] are discovered only through numerical computation. But are the standard relations enough to produce all the linear relations when $N$ *is* standard? In weight two, when $N$ is a prime the answer is affirmative if one assumes a variant of Grothendieck’s period conjecture [@Zocta]. Computations above give strong support for it and, in fact, is the primary motivation of it.
Computational results in other weights {#sec:wt3comp}
======================================
In this last section we briefly discuss our results in weight 3,4 and 5. Since the computational complexity increases exponentially with weight we cannot do as many cases as we have done in weight two.
Combining the FDS , , RDS -, and the seeded relations we have verified the following facts by MAPLE: $d(3,1)= 1, d(3,2)\le 3, d(3,3)\le 8$.... We have done similar computation in other small weight and low level cases and listed the results in Table \[Ta:dbzeta\].
We list some values of $D(w,N)$ in Table \[Ta:dbzeta\] to compare with the bound $SR(w,N)$ obtained by standard relations.
$N $ 1 2 3 4 5 6 7 8 9 10 11 12 13
----------- --- --- ---- ---- ----- ----- ------ ----- ------ ----- ------ ----- -------
$SR(3,N)$ 1 3 8 9 22 23 50 38 67 70 157 94 246
$D(3,N)$ 1 3 8 8 27 21 64 27 64 56 216 56 343
$SR(4,N)$ 1 5 16 21 61 69
$D(4,N)$ 1 5 16 16 81 55 256 81 256 209 1296 209 2401
$SR(5,N)$ 2 8 32
$D(5,N)$ 2 8 32 32 243 144 1024 243 1024 780 7776 780 16807
: Upper bound of $d(w,N)$ obtained by standard relations and [@DG 5.25].[]{data-label="Ta:dbzeta"}
Note that $d(3,4)=D(3,4)+1$. By numerical computation we find the following non-standard relation: $$\begin{aligned}
5L_4(1,2|2,3)=&46L_4(1,1,1|1,0,0)-7L_4(1,1,1|2,2,1)-13L_4(1,1,1|1,1,1) +13L_4(1,2|3,1) \notag \\
-L_4(1,1,1|&3,2,0)+25L_4(1,1,1|3,0,0)-8L_4(1,1,1|1,1,2) +18L_4(2,1|3,0), \label{equ:conj1}\end{aligned}$$ Recently, we prove that by using the octahedral symmetry of ${{\mathbb P}}^1-(\{0,\infty\}\cup \mmu_4)$ one can deduce equation (see [@Zocta]).
From the available data in Table \[Ta:dbzeta\] we can formulate the following conjecture.
Let $N=p$ be a prime $\ge 5$. Then $$d(3,p)\le \frac {p^3+4p^2+5p+14}{12}.$$ Moreover, equality hold if standard relations produce all the linear relations.
We obtained this conjecture under the belief that the upper bound of $d(3,p)$ produced by the standard relations should be a polynomial of $p$ of degree 3. Then we find the coefficients by the bounds of $d(3,p)$ for $p=5,7,11,13$ in Table \[Ta:dbzeta\].
When $w>2$ it’s not too hard to improve the bound of $d(w,p)$ given in [@DG 5.25] by the same idea as used in the proof of [@DG 5.24] (for example, decrease the bound by $(p^2-1)/24$). But they are often not the best. We conclude our paper with the following conjecture.
\[conj:stand\] If $N$ is a standard level then the standard relations always provide the sharp bounds of $d(w,N)$, namely, all linear relations can be derived from the standard ones. If $N$ is a non-standard level then the bound in *[@DG Cor. 5.25]* is sharp and the non-standard relations exist in $\MPV(w,N)$ for all $w\ge 3$ (and in $\MPV(2,N)$ if $N\ge 10$).
[9]{}
H. Bass, *Generators and relations for cyclotomic units,* Nagoya Math. J. **27** (2)(1966), 401–407.
M. Bigotte, , G. Jacob, N.E. Oussous and M. Petitot, *Lyndon words and shuffle algebras for generating the coloured multiple zeta values relations tables,* Theoretical Computer Science, **273** (1-2)(2002), 271–282.
J. M. Borwein, D. J. Broadhurst, and D. M. Bradley, *Evaluations of $k$-fold Euler/Zagier sums: a compendium of results for arbitrary $k$*, Electronic J. Combinatorics, **4** (2)(1997), \#R5. Wilf Festschrift.
J. Borwein, P. Lisonek, and P. Irvine, *An interface for evaluation of Euler sums*, available online at http://oldweb.cecm.sfu.ca/cgi-bin/EZFace/zetaform.cgi
D. J. Broadhurst, *Massive 3-loop Feynman diagrams reducible to SC\* primitives of algebras of the sixth root of unity*, European Phys. J. C (Fields) **8** (1999), 311–333
D. J. Broadhurst, C*onjectured enumeration of irreducible multiple zeta values, from knots and Feynman diagrams,* preprint hep-th9612012.
K.-T.-Chen, *Algebras of iterated path integrals and fundamental groups*, Trans. Amer. Math. Soc. **156** (1971), 359–379.
M. Conrad, *Construction of Bases for the Group of Cyclotomic Units*, Journal of Number Theory **81**(2000), 1-15.
P. Deligne, *Le groupe fondamental de la droite projective moins trois points*, Galois groups over $\Q$, (Berkeley, CA, 1987), Springer, New York, 1989, p. 79–297.
P. Deligne, *Le groupe fondamental de la $\G_m-\mmu_N$*, unpublished manuscript.
P. Deligne and A. Goncharov, *Groupes fondamentaux motiviques de Tate mixte*, Annales Scientifiques de l’École Normale Supérieure, **38** (1)(2005), 1–56. Also available math.NT/0302267.
V. Ennola, *On relations between cyclotomic units*, J. Number Theory **4** (1972), 236–-247.
A. B. Goncharov, *Polylogarithms in arithmetic and geometry*, in: Proc. ICM, Zürich, 374–387, Vol. I, Birkhäuser, 1994.
A. Goncharov, *The dihedral Lie algebras and [G]{}alois symmetries of $\pi_1^{(l)}(\mathbb{P}_1-(\{0,\infty\}\cup\mu_{N}))$*, Duke Math. J. **110** (3)(2001), 397–487.
M. E. Hoffman, *Multiple harmonic series,* Pacific J. Math., **152** (2)(1992), 275–290.
M. E. Hoffman, *Quasi-shuffle products,* J. Algebraic Combin. **11** (2000), 49–68.
M. E. Hoffman, *Algebra of Multiple Zeta Values and Euler Sums,* Mini-Conference on Zeta Functions, Index, and Twisted K-Theory: Interactions with Physics, Oberwolfach, Germany, May 2, 2006. Available online www.usna.edu/Users/math/meh.
K. Ihara, M. Kaneko, and D. Zagier, *Derivation and double shuffle relations for multiple zeta values*, Comp. Math. **142** (2006), 307–338.
M. Kontsevich, *Vassiliev’s knot invariants,* I. M. Gel’fand Seminar, Avd. Soviet Math., **16** (2)(1993), 137-150.
G. Racinet, *Doubles mélanges des polylogarithmes multiples aux racines de l’unité*, Publ. Math. IHES **95** (2002), 185–231.
J. Vollinga, S. Weinzierl, *Numerical evaluation of multiple polylogarithms*, arXiv:hep-ph/0410259
D. Zagier, *Values of Zeta Function and Their Applications,* Proceedings of the First European Congress of Mathematics, **2**, (1994), 497–512.
J. Zhao, Analytic continuation of multiple polylogarithms. [*Analysis Mathematica*]{}, **33** (2007). arXiv: math.AG/0302054
J. Zhao, *Multiple polylogarithm values at roots of unity*, *C. R. Acad. Sci. Paris, Ser. I.* 2008. DOI: 10.1016/j.crma.2008.09.011
J. Zhao, *Double shuffle relations of Euler sums*, arXiv: 0705.2267.
J. Zhao, *An integral structure of multi-polylog values at roots of unity*, in preparation.
| {
"pile_set_name": "ArXiv"
} |
There has been a tremendous amount of work on mean field calculations for disordered systems in the past 20 years, in part driven by the exact solution provided by Parisi’s replica symmetry breaking Ansatz. Although his solution was developed in the context of spin glasses, the formalism has been extremely useful for understanding other disordered systems. Generally, one expects mean field to give exact results as the dimensionality goes to infinity. One can then ask whether mean field leads to “acceptable” errors for systems of interest, e.g., in three dimensions, and whether it is possible to compute Euclidean corrections to the mean field formulae. Such computations might correspond to a $1/d$ expansion for the thermodynamic functions of interest. In frustrated disordered systems, however, this has turned out to be intractable. To date, such Euclidean corrections have been pushed furthest [@MezardParisi_88] for the Minimum Matching Problem (MMP) [@PapadimitriouSteiglitz]. In what follows, we determine the accuracy of the mean field approximation and the effectiveness of the corrections thereto in the MMP by comparing with the actual properties of the $d$-dimensional Euclidean model. First, we find that the relative error introduced by the mean field approximation for the zero temperature energy density is less than $4 \%$ at $d=2$ and $3 \%$ at $d=3$. Second, the inclusion of the “leading” Euclidean corrections to the mean field approximation reduces the error by a factor of about $10$ at $d=2$ and $d=3$ Third, we argue that the large $d$ behavior of systems such as the MMP depends on arbitrarily high order correlations and is thus beyond all orders of the expansion proposed by Mézard and Parisi [@MezardParisi_88].
Consider $N$ points ($N$ even) and a specified set of link lengths $l_{ij}=l_{ji}$ separating the points, for $1 \le i,j \le N$. One defines a matching (a dimerization) of these points by combining them pairwise so that each point belongs to one and only one pair. Define also the energy or length of a matching as the sum of the lengths of the links associated with each matched pair. The Minimum Matching Problem is the problem of finding the matching of minimum energy. One can also consider the thermodynamics of this system, as proposed by Orland [@Orland] and Mézard and Parisi [@MezardParisi_85], by taking [*all*]{} matchings but weighting them with the Boltzmann factor associated with their energy. Here we concentrate on the $T=0$ properties because there is no effective numerical method for extracting thermodynamical functions in this system, but exact energy minima can be obtained quite easily for any given instance. Indeed, the MMP belongs to the algorithmic class P of polynomial problems, and there are standard algorithms which solve any instance of size $N$ using on the order of $N^3$ steps [@BallDerigsMBM_Net].
Physically, one is not interested in the properties of any particular instance of the MMP; more relevant are typical and ensemble properties such as the average energy when the lengths $l_{ij}$ are random variables with a given distribution. One then speaks of the stochastic MMP. There are two frequently used ensembles for the $l_{ij}$, corresponding to the [*Euclidean*]{} MMP and the [*random link*]{} MMP. In the first, the $N$ random points lie in a $d$-dimensional Euclidean volume and the $l_{ij}$ are the usual Euclidean distances between pairs of points. The points are independent and identically distributed, so one speaks of a random point problem. In the second ensemble, it is the link lengths $l_{ij}$ which are [*independent*]{} and identically distributed random variables. A connection between these two systems was first given by Mézard and Parisi [@MezardParisi_86b]: they pointed out that the one and two-link distributions could be made identical in the two problems. A consequence is that the “Cayley tree” approximation for the random point and random link problems are the same. Mézard and Parisi were able to solve the random link MMP using an approach based on replicas [@MezardParisi_85; @MezardParisi_86b]. One may then consider the random link MMP to be a “mean field model” for the Euclidean MMP. The mean field [*approximation*]{} consists of using the thermodynamic functions of the random link model as estimators for those of the Euclidean model. This approximation is applicable to all link based combinatorial optimization problems, such as the assignment problem and the traveling salesman problem; hereafter we shall refer to it as the [*random link approximation*]{} [@CBBMP]. Finally, Mézard and Parisi have shown how to derive corrections systematically to the random link approximation using a connected-correlation link expansion. In [@MezardParisi_88], they have computed the leading corrections, these being associated with the triangle inequality (3-link correlations) in the Euclidean model. How accurate are these approximations? To answer this, we first give our results for the Euclidean problem, and then compare with the predictions of the random link approximation and of the link expansion method. In the Euclidean MMP, let $L_{MM}^E$ be the energy or length of the minimum matching. Taking the points to be independent and uniformly distributed in a unit volume, Steele [@Steele_AP] has shown that as $N \to \infty$, $L_{MM}^E / N^{1-1/d}$ converges with probability one to a non-random, $N$-independent constant $\beta_{MM}^E(d)$. In physics language, this result shows that $L_{MM}^E$ is self-averaging and that the zero-temperature energy density has an infinite volume limit when the density of points is kept fixed. To date, little has been done to compute $\beta_{MM}^E(d)$. The best estimates seem to be due to Smith [@Smith_Thesis]: $\beta_{MM}^E(2) \approx 0.312$ and $\beta_{MM}^E(3) \approx 0.318$. Let us use a systematic procedure [@CBBMP] to obtain $\beta_{MM}^E(d)$ with quantifiable errors. First, in order to have a well defined dependence on $N$, we have used the ensemble average, $\langle L_{MM}^E \rangle / N^{1-1/d}$. Second, in order to reduce corrections to scaling in the extrapolation to the large $N$ limit, we have placed the points randomly in the $d$-dimensional unit hypercube with periodic boundary conditions. This removes surface effects and empirically leads to the finite size scaling law $$\label{eq_fss}
{ \langle L_{MM}^E \rangle \over N^{1-1/d} } =
\beta_{MM}^E(d) ( 1 + {A(d) \over N} + { B(d) \over N^2} + \cdots).$$ Finally, in order to reduce statistical fluctuations, we have used a variance reduction trick [@CBBMP]. The improved estimator has the effect of reducing the variance of our estimates by more than a factor $4$, and thus saves us a considerable amount of computer time. Using these methods, we have extracted from our numerical data $\beta_{MM}^E(d)$ and its associated statistical error. The fits to Eq.(\[eq\_fss\]) are good, with $\chi^2$ values confirming the form of the finite size scaling law. The error bars on the extrapolated value $\beta_{MM}^E(d)$ are obtained in the standard way by requiring that $\chi^2$ increase by one from its minimum. We find in particular $\beta_{MM}^E(2) = 0.3104 \pm 0.0002$, and $\beta_{MM}^E(3) = 0.3172 \pm 0.00015$; values at higher dimensions are given in Table \[tab\_betas\]. We have checked that these results are not significantly modified when using another random number generator to produce the instances, and that the fits are stable to truncation of the data.
Now we discuss how to use the random link model to approximate $\beta_{MM}^E(d)$. For any two points $(i,j)$ placed at random in the unit $d$-dimensional hypercube, the density distribution of $l_{ij}$ is given at short distances by $P_d(l_{ij}=r) = d B_d r^{d-1}$, where $B_d = \pi^{d/2}/(d/2)!$ is the volume of the $d$-dimensional ball with unit radius. If we take the random link model where link lengths are independent and have the individual distribution $P_d(l)$, then the Euclidean and random link MMP have the same one and two-link distributions [@MezardParisi_86b] because two Euclidean distances are independent. If correlations among three or more link lengths are weak, then the properties of the two systems should be quantitatively close. Thus an analytic approximation to $\beta_{MM}^E(d)$ is obtained by computing its analogue $\beta_{MM}^{RL}(d)$ in the random link MMP. In references [@MezardParisi_85; @MezardParisi_86b], Mézard and Parisi solved these random link models under the replica symmetry hypothesis. They showed further [@MezardParisi_87b] that the replica symmetric solution is stable (at least for $d=1$), and thus is most likely exact unless a first order phase transition occurs in this system. Their solution gives $\beta_{MM}^{RL}(d)$ in terms of a function $G_d$ related to the probablility distribution of link lengths for matched pairs. In our Euclidean units their result can be written $$\label{eq_mf_beta}
\beta_{MM}^{RL}(d) = {D_1(d)\over 2} {d\over (1/d)!}
\int_{-\infty}^{+\infty} G_d(x) e^{-G_d(x)} dx$$ where $G_d$ satisfies the integral equation $$\label{eq_mf_G}
G_d(x) = d \int_{-x}^{+\infty} (x+y)^{d-1} e^{-G_d(y)} dy$$ and where $$D_1(d) = \lim_{N\to \infty} \langle L_1 \rangle /N^{1-1/d} = (1/d)!
B_d^{-1/d}$$ is the average (rescaled) link length of the nearest neighbor graph in the limit $N\to \infty$.
Brunetti [*et al.*]{} [@BKMP] have used direct numerical simulations of these random link models to confirm the predictions to the level of $0.2 \%$ at $d=1$ and $2$, and we have done the same to the level of $0.1 \%$ at $1 \le d \le 10$, giving further evidence that the replica symmetric solution is exact. From the analytical side, solving the integral equation for $G_d$ leads to $\beta_{MM}^{RL}(1)=\pi^2/24 = 0.4112335$…, $\beta_{MM}^{RL}(2) = 0.322580$…, and $\beta_{MM}^{RL}(3)=0.326839$…; values at higher dimensions are given in Table \[tab\_betas\]. If we consider $\beta_{MM}^{RL}(d)$ as a mean field prediction for $\beta_{MM}^E(d)$, the accuracy is surprisingly good. Including the trivial value $\beta_{MM}^E(1) = 0.5$, we see that the random link approximation leads to a relative error of $17.8 \%$ at $d=1$, of $3.9 \%$ at $d=2$, and of $3.0 \%$ at $d=3$. Also, the error decreases with increasing dimension. It can be argued, for the MMP as well as for other link-based combinatorial optimization problems [@CBBMP], that the random link approximation not only has a relative error tending towards $0$ as $d \to \infty$, but that in fact this error is at most of order $1/d^2$. Given our high quality estimates, we are able to confirm this property numerically. In Figure \[fig\_rl\_error\] we plot the quantity $d (\beta_{MM}^{RL}-\beta_{MM}^E)/ \beta_{MM}^{E}$ along with a quadratic fit given to guide the eye. As expected, the data scales as $1/d$. Thus the random link approximation gives both the leading and $1/d$ subleading dependence of $\beta_{MM}^E(d)$. In order to obtain analytic expressions for these coefficients, we have derived the $1/d$ expansion for $\beta_{MM}^{RL}$ from Eqs.(\[eq\_mf\_beta\],\[eq\_mf\_G\]). We used two methods to do this. The first, straightforward but computationally lengthy, consists of setting ${\tilde G_d}(x) = G_d(\tilde x = x/d + 1/2)$ and then writing ${\tilde G_d}(x)$ as a power series in $1/d$. From this we find $$\label{eq_betaRL}
\beta_{MM}^{RL}(d) = {D_1(d)\over 2} (1 + {1-\gamma \over d}
+ O(1/d^2) )$$ where $\gamma = 0.577\ldots$ is Euler’s constant. If, as claimed, the random link approximation gives an error of order $1/d^2$, Eq.(\[eq\_betaRL\]) gives an analytic expression for the leading and first subleading terms in the $1/d$ expansion of $\beta_{MM}^E(d)$. This claim is strongly supported by the numerical results: performing a fit of our $\beta_{MM}^E(d)$ values to a truncated $1/d$ series leads to $0.424 \pm 0.008$ for the coefficient of the $1/d$ term; this is to be compared to the theoretical prediction of $1-\gamma = 0.422784\ldots$
We have been able to obtain the next coefficient of the series in $1/d$ for $\beta_{MM}^{RL}$ by using a second method. We introduce a modified random link model where the links are shifted and rescaled in such a way that the leading term of the $1/d$ expansion for this new model is exactly the $1/d$ coefficient for the initial one [@Boutet_Thesis]. In fact it is possible to introduce a sequence of such “rescaled” models, where the $k^{th}$ model is designed to produce the $1/d^k$ term of the expansion. We have computed the leading terms predicted by a replica symmetric analysis of these models for $k=1$ and $2$, from which we find that the order $1/d^2$ coefficient in Eq.(\[eq\_betaRL\]) is $\pi^2/12 + \gamma^2/2 - \gamma$.
We now come to the final point of the paper: how well can one predict $\beta_{MM}^E(d)$ by incorporating Euclidean corrections to the random link approximation? It is necessary here to review the work of Mézard and Parisi; for greater detail, we refer the reader to their article [@MezardParisi_88]. They begin with the partition function $Z$ for an arbitrary stochastic MMP and write the quenched average for $n$ replicas. In the Euclidean model, the $l_{ij}$ have three and higher-link correlations. Mézard and Parisi keep the three-link correlations (arising only when the three links make a triangle) and neglect higher connected correlations. Note that it is not clear [*a priori*]{} whether these “higher order” terms are negligeable compared to the three-link term. The resulting expression for the quenched average becomes $$\begin{aligned}
\label{eq_Z_EC}
{\overline {Z^n}} = \prod_{j=1}^N \prod_{\alpha=1}^n \big(
\int_0^{2 \pi} {d \lambda_j^{\alpha} \over 2 \pi} ~ e^{i \lambda_j^{\alpha} }
\big)
\nonumber \\
\times ~~ e^{ \sum_{(ij)} {\overline {u_{ij}} } +
\sum_{(ij)(kl)(mn)}^{\prime} {\overline {u_{ij} u_{kl} u_{mn} } }^{~C} }\end{aligned}$$ where $u_{ij}$ is a complicated nonlinear function of the link length $l_{ij}$. They then compute the limit $N \to \infty$, $n \to 0$ using the saddle point method while assuming that replica symmetry is not broken. In the zero temperature limit, just as in the standard random link model, the saddle point equations can be expressed in terms of $G_d$, but now $G_d$ satisfies a more complicated integral equation (Eq.(34) in their paper). From this, one can calculate new estimates for $\beta_{MM}^E(d)$, which we shall denote $\beta_{MM}^{EC}$, where $EC$ stands for Euclidean corrections.
We have solved the equations numerically for this modified $G_d$, and have computed $\beta_{MM}^{EC}(d)$ for $2 \le d \le 10$. We find $\beta_{MM}^{EC}(2)=0.30915$ and $\beta_{MM}^{EC}(3)=0.31826$. The results for $d \ge 4$ are given in Table \[tab\_betas\]. Comparing with $\beta_{MM}^E(d)$ and $\beta_{MM}^{RL}(d)$, we see that the new estimates are considerably more accurate. At $d=2$, the random link approximation leads to an error of $3.9 \%$; this error is decreased by nearly a factor $10$ by incorporating these leading Euclidean corrections. Similarly at $d=3$, the error is reduced from $3.0 \%$ to less than $0.4\%$. At larger $d$, the error continues to decrease, but the effect is less dramatic.
To interpret this last result, consider how the difference $\beta_{MM}^{EC} - \beta_{MM}^{RL}$ scales with $d$. Using Eq.(\[eq\_Z\_EC\]), we see that it is sufficient to estimate the size of the $3$-link correction term. Its dimensional dependence follows that of the probability of finding nearly equilateral triangles as $d \to \infty$. Since this probability goes to zero exponentially with $d$, the $3$-link correlations give tiny corrections at large $d$ (as confirmed by the numreics), and also the power series expansion in $1/d$ of $\beta_{MM}^{EC}$ is [*identical*]{} to that of $\beta_{MM}^{RL}$. This property continues to hold if one includes 4, 5, or any [*finite*]{} number of multi-link correlations in Eq.(\[eq\_Z\_EC\]). This is due to the fact that the Euclidean and random link graphs have [*local*]{} properties that are identical up to exponentially small terms in $d$. In particular, the statistics of fixed sized ($N$-independent) loops connecting near neighbors are nearly identical.
Although this reasoning was given for the MMP, it applies equally well to other link-based problems. In such statistical mechanics systems, if the thermodynamic functions depend only on the local properties of the (short) link graph, then the random link approximation applied to the Euclidean system will have an error which is exponentially small in $d$. However, for combinatorial optimization problems such as the MMP, the assignment problem, and the traveling salesman problem, the $N \to \infty$ limit and the $k$-link expansion do not commute: $k$-link correlations with $k$ growing with $N$ remain important as $N \to \infty$. In particular, arbitrarily large loops matter and contribute to the thermodynamics at order $1/d^2$. In a polymer picture, we can say that the random link approximation is exponentially good in the dilute phase, while it leads to $1/d^2$ errors in the collapsed phase. The power corrections in this phase are beyond all orders in a $k$-link correlation expansion such as Eq.(\[eq\_Z\_EC\]).
In summary, we have estimated by numerical simulation $\beta_{MM}^E(d)$, the zero energy density in the Euclidean Minimum Matching problem at dimensions $2 \le d \le 10$. We have then computed two analytical estimates for these energy densities, namely $\beta_{MM}^{RL}(d)$ and $\beta_{MM}^{EC}(d)$. The first method uses the random link approximation where all link correlations are neglected. Using the “exact” mean field solution of Mézard and Parisi, we find that even at low dimensions, the error introduced by this approximation is small: $3.9 \%$ at $d=2$, $3.0 \%$ at $d=3$, and $2.0 \%$ at $d=4$. In the second method, the connected three-link correlations are taken into account while higher ones are neglected. Using Mézard and Parisi’s expressions, we find that this modification to the random link model gives excellent predictions at $d=2$ and $3$, with the error there being divided by almost $10$ compared to the random link approximation. This provides a stringent quantitative test of a systematic expansion which goes beyond uncorrelated disorder variables, and suggests that even the leading such correction is enough to get predictions for thermodynamic functions precise to better than one percent. Finally, at high dimensions, we have seen that the $k$-link correlation expansion leads to corrections which vanish exponentially with $d$; this expansion thus misses important $1/d$ power law corrections for problems such as the MMP. This leaves open the determination of the $1/d^2$ term in the expansion of the constants $\beta_{MM}^E(d)$. We have performed a fit on our data, imposing the leading and the $1/d$ term to be those of the random link model. We find the $1/d^2$ coefficient to be very small (smaller than $0.01$ in absolute value). Clearly, it would be of major interest to obtain an analytical value for this term.
We are grateful to C. De Dominicis, J. Houdayer, W. Krauth, M. Mézard, and H. Orland for their interest and suggestions. JBdM acknowledges a fellowship from the MENESR, and OCM acknowledges support from the Institut Universitaire de France. The Division de Physique Théorique is an Unité de Recherche des Universités Paris XI et Paris VI associée au CNRS.
[\*]{} Electronic address: [email protected]
Electronic address: [email protected]
[10]{}
M. O. Ball and U. Derigs. An analysis of alternate strategies for implementing matching algorithms. , 13:517–549, 1983.
J. [Boutet de Monvel]{}. . PhD thesis, Institut de Physique Nucléaire, Université Paris-Sud, Orsay, 1996.
R. Brunetti, W. Krauth, M. Mézard, and G. Parisi. Extensive numerical solutions of weighted matchings: Total length and distribution of links in the optimal solution. , 14(4):295–301, 1991.
N. J. Cerf, J. [Boutet de Monvel]{}, O. Bohigas, O. C. Martin, and A. G. Percus. The random link approximation for the [E]{}uclidean traveling salesman problem. , 7(1):117–136, 1997.
M. Mézard and G. Parisi. Replicas and optimization. , 46:L771–L778, 1985.
M. Mézard and G. Parisi. Mean-field equations for the matching and the travelling salesman problems. , 2:913–918, 1986.
M. Mézard and G. Parisi. On the solution of the random link matching problems. , 48:1451–1459, 1987.
M. Mézard and G. Parisi. The [E]{}uclidean matching problem. , 49:2019–2025, 1988.
H. Orland. Mean-field theory for optimization problems. , 46:L763–L770, 1985.
C. H. Papadimitriou and K. Steiglitz. . Prentice Hall, Englewood Cliffs, NJ, 1982.
W. D. Smith. . PhD thesis, Princeton University, Princeton, NJ, 1989.
M. Steele. Subadditive [E]{}uclidean functionals and nonlinear growth in geometric probability. , 9(3):365–376, 1981.
(300,300)
$d$ $\beta^E(d)$ $\beta^{RL}(d)$ $d {\beta^{RL} - \beta^E\over \beta^E}$ $\beta^{EC}(d)$ ${\beta^{EC}-\beta^E\over \beta^E}$
----- --------------------- ----------------- ----------------------------------------- ----------------- -------------------------------------
4 0.3365$\pm$ 0.0003 0.343227 +0.080 0.33756 +0.30%
5 0.3572$\pm$ 0.00015 0.362175 +0.070 0.35818 +0.27%
6 0.3777$\pm$ 0.0001 0.381417 +0.059 0.37849 +0.21%
7 0.3972$\pm$ 0.0001 0.400277 +0.054 0.39807 +0.22%
8 0.4162$\pm$ 0.0001 0.418548 +0.045 0.41685 +0.17%
9 0.4341$\pm$ 0.0001 0.436185 +0.042 0.43485 +0.17%
10 0.4515$\pm$ 0.0001 0.453200 +0.037 0.45214 +0.14%
: Comparison of the MMP constants for the three models: Euclidean, random link, and random link including 3-link Euclidean corrections ($4\le d\le 10$).[]{data-label="tab_betas"}
| {
"pile_set_name": "ArXiv"
} |
---
address: |
LACL, UMR-FRE 2673, Université Paris 12, Route forestière Hurtault, F-77300 Fontainebleau, France,\
[email protected]\
LIAFA, UMR 7089 and Université Paris 6, 2 Place Jussieu, 75254 Paris Cedex 5, France; send correspondence to [email protected]\
Steklov Institute of Mathematics, Fontanka 27, St. Petersburg, Russia. [email protected]\
Support by INTAS grant 04-77-7173 is gratefully acknowledged.
author:
- Patrick Cégielski Irène Guessarian Yuri Matiyasevich
title: Multiple serial episodes matching
---
-1cm
Introduction
============
The recent development of datamining induced the development of computing techniques, among them is episode searching and counting. An example of frequent serial episode search is as follows: let $t$ be a text consisting of requests to a university webserver ; assume we wish to count how many times, within at most 10 time units, the sequence $e_1e_2e_3e_4$ appears, where $e_1=$ ‘Computer Science’, $e_2=$ ‘Master’, $e_3=$ ‘CS318 homepage’, $e_4=$ ‘Assignment’. It suffices to count the number of 10-windows of $t$ containing the subsequence $p=e_1e_2e_3e_4$. If $e_1,e_2,e_3,e_4$ must appear in that same order in the window, the episode is said to be [*serial*]{}, if they can appear in any order, the episode is said to be [*parallel*]{}; a partial order can also be imposed on the events composing an episode (see [@mtv], which proposes several algorithms for episode searching). Searching serial episodes is more complex than searching parallel episodes. Of course, if one has to scan a $\log$ file, it is better to do it for several episodes $e_1e_2\ldots e_n$, $f_1f_2\ldots f_m$, $g_1g_2\ldots g_p$ simultaneously. We will hence investigate the search of several serial episodes in the same window: each serial episode is ordered, but no order is imposed among occurrences of the episodes in the window.
The problem we address is the following: given a text $t$ of length $n$, patterns $m_1,\ldots, m_q$ on the same alphabet $A$ and an integer $w$, we wish to determine the number of size $w$ windows of text containing all $q$ patterns as serial episodes, [*i.e.*]{} the letters of each $m_i$ appear in the window, in the same order as in $m_i$, but they need not be consecutive because other letters can be interleaved. When searching for a single pattern $m$, this problem with arguments the window size $w$, the text $t$ and pattern $m$ is called [*serial episode matching problem* ]{} in [@mtv], [*episode matching*]{} in [@dfggk] and [*subsequence matching*]{} in [@ahu]; a related problem is the [*matching with don’t cares*]{} of [@mby; @kr].
This problem is an interesting generalisation of [*pattern-matching*]{}. Without the window size restriction, it is easy to find in linear time whether $p$ occurs in the text: if $p=p_1\ldots p_k$, a finite state automaton with $k+1$ states $s_0,s_1,\ldots,s_k$ will read the text; the initial state is $s_0$; after reading letter $p_1$ we go to state $s_1$, then after reading letter $p_2$ we go to state $s_2$, …; the text is accepted as soon as state $s_k$ is reached. Episode matching within a $w$-window is harder; its importance is due to potential applications to datamining [@m; @mtv] and molecular biology[@mby; @kr; @nr].
For the problem with a single episode in $w$-windows, a standard algorithm is described in [@dfggk; @mtv]. It is close to the algorithms of [*pattern-matching*]{} [@a; @ahu] and its time complexity is $O(nk)$. Another [*on-line*]{} algorithm is described in [@dfggk]: the idea is to slice the pattern in $k/\log k$ well-chosen pieces organised in a [*trie*]{}; its time complexity is $O(nk/\log k)$. We gave an [*on-line*]{} algorithm reading the text $t$, each text symbol being read only once and whose time complexity is $O(n)$ [@bcgm].
In this paper, we describe two efficient algorithms (Section \[sect.plusmot\]) for solving the problems of simultaneous search of multiple episodes. These algorithms use the [*MP–RAM*]{}, that we introduced in [@bcgm], to model microprocessor basic operations, using only the fast operations on bits ([*shifts*]{}), and bit-wise addition; this gives an on-line algorithm in time $O(nq)$ (theorem \[mult.thm\]). In practice, this algorithm based on MP–RAMs and a new implementation of [*tries*]{}, is much faster as shown in section \[sect.exp\]. We believe that other algorithms can be considerably improved if programmed on MP–RAMs.
Our algorithm relies upon two ideas: 1) preprocess patterns and window size to obtain a finite automaton solving the problem as in Knuth, Morris, and Pratt algorithm [@kmp] (the solutions preprocessing the text [@t; @mby; @sliss; @u] are prohibitive here because of their space complexity) and 2) code the states of this automaton to compute its transitions very quickly on MP-RAMs, without precomputing, nor storing the automaton: using the automaton itself is also prohibitive, not the least because of the number of states; we emulate the behaviour of the automaton without computing the automaton. We study: (a) the case when the patterns have no common part and (b) the case when they have similar parts. In each case, an appropriate preprocessing of the set of patterns enables us to build an automaton solving the problem and we show that the behaviour of this automaton can be emulated on-line on MP-RAMs. Moreover, the time complexity of the preprocessing is insignificant because it is smaller than the text size by several orders of magnitude: typically, window and patterns will consist of a few dozen characters while the text will consist of several million characters.
The paper is organised as follows: in section 2, we define the problem, in section \[sect.plusmot\] we describe the algorithms searching multiple episodes in parallel; we present the experimental results in section \[sect.exp\].
The problem
===========
The (multiple) episode problem
------------------------------
An [*alphabet*]{} is a finite non-empty set $A$. A [*length*]{} $n$ [*word*]{} on $A$ is a mapping $t$ from $\{1,\ldots,n\}$ to $A$. The only length zero word is the [*empty word*]{}, denoted by $\varepsilon$. A non-empty word $t \; : \; i \mapsto t_i$is denoted by $t_1t_2 \cdots t_n$. A [*language*]{} on alphabet $A$ is a set of words on $A$.
Let $t = t_1t_2 \cdots t_n$ be a word which will be called the [*text*]{} in the paper. The word $p = p_1p_2 \cdots p_k$ is a [*factor*]{} of $t$ iff, there exists an integer $j$ such that $t_{j+i} = p_i$ for $1 \leq i \leq k$. A size $w$ [*window*]{} of on $t$, in short $w$-window, is a size $w$ factor $t_{i+1}t_{i+2} \cdots t_{i+w}$ of $t$; there are $n-w+1$ such windows in $t$. The word $p$ is an [*episode*]{} (or [*subsequence*]{}) of $t$ iff there exist integers $1 \leq i_1 < i_2 < \cdots < i_k \leq n$ such that $t_{i_j} = p_j$ for $1 \leq j \leq k$. If moreover, $i_k - i_1 < w$, $p$ is an [*episode*]{} of $t$ [*in a $w$-window*]{}.
\[ex1\] [If $t={}$“dans ville il y a vie" (a French advertisement, see figure \[monoprix.fig\]) then “vie" is a factor and hence a subsequence of $t$. “vile" is neither a factor, nor a subsequence of $t$ in a $4$-window, but it is a subsequence of $t$ in a $5$-window. See figure \[monop.fig\].]{}
Given an alphabet $A$, and words $t,m_1,\ldots, m_q$ on $A$:
the simultaneous [*pattern-matching*]{} problem consists in finding whether $m_1,\ldots,m_q$ are factors of $t$,
given moreover a window size $w$:
the [*subsequence existence*]{} problem consists in finding whether $m_1,\ldots,m_q$ are subsequences of $t$ in a $w$-window;
the [*multiple episode search*]{} problem consists in counting the number of $w$-windows in which all of $m_1,\ldots,m_q$ are subsequences of $t$.
For the simultaneous search of several subsequences $m_1,\ldots,m_q$, we have various different problems:
- either we count the number of occurrences of each $m_i$ in a $w$-window (not necessarily the same): this case will be useful for searching in parallel, with a single scan of the text, a set of patterns which are candidates for being frequent.
- or we count the number of windows containing all the $m_i$s: this case will be useful for trying to verify association rules. For example, the association rule $m_2,\ldots,m_q \Longrightarrow m_1$ will be useful if the number of $w$-windows containing all the $ m_2,\ldots,m_q$ is high enough, and to check that, we will count the $w$-windows containing [*all of*]{} $m_2,\ldots,m_q$. Our method will enable us to verify more easily both the validity of the association rule (“among the windows containing $ m_2,\ldots,m_q$ many contain also $m_1$”) and the fact that it is interesting enough (“many windows contain $m_2,\ldots,m_q$”): it will suffice to count simultaneously the windows containing $m_2,\ldots,m_q$ and the windows containing $m_1, m_2,\ldots,m_q$.
A naive solution exists for [*pattern-matching*]{}. Its time complexity on RAM is $O(nk)$, where $k$ is the pattern size. Knuth, Morris, and Pratt [@kmp] gave a well-known algorithm solving the problem in linear time $O(n + k)$. A solution in $O(nk)$ is given in [@mtv] for searching a single size $k$ episode. We gave in [@bcgm] an algorithm with time complexity $O(n)$ (on MP–RAM) for searching a single episode.
The notation $o(nk)$
--------------------
Let us first make precise the meaning of the notation $o(nk)$.
The notation $o(h(n))$ was introduced to compare growth rates of functions with one argument; for comparing functions with several arguments, various non-equivalent interpretations $o(h(n,m,...))$ are possible. Consider a function $t(n,k)$; $t(n,k)=o(nk)$ could mean:
either $\displaystyle \lim_{n+k\to+\infty} t(n,k)/nk =0$;
or $\displaystyle\lim_{n\to+\infty\atop k\to+\infty} t(n,k)/nk =0$, i.e. $\forall \epsilon, \exists N, \ \forall n,\forall k\allowbreak
\big(n>N \hbox{ and }k>N\allowbreak \Longrightarrow t(n,k)<\epsilon nk\big)$.
With meaning 1, no algorithm can solve the [*single episode within a window*]{} problem in time $o(nk)$. Indeed, any algorithm for the [*episode within a window*]{} problem must scan the text at least once, hence $t(n,k)\geq n$. For a given $k$, for example $k=2$,we have $t(n,k)/nk\geq 1/2$. Hence $\displaystyle \lim_{n+k\to+\infty} t(n,k)/nk =0$ is impossible. We thus have to choose meaning 2.
Algorithms on MP–RAM {#section.algo3}
--------------------
Given a window size $w$ and $q$ patterns, we preprocess (patterns + window size $w$) to build a virtual finite state automaton $\A$; we will then emulate on-line the behaviour of $\A$ to scan text $t$ and count in time $nq$ the number of windows containing our patterns as episodes. Note that our method is different from both: 1) methods preprocessing the text [@t; @mby; @sliss; @u] (we preprocess the pattern) and 2) methods using suffixes of the pattern [@croch; @mby; @kr; @u] (we use prefixes of the patterns). We encode the subset of states of $\A$ needed to compute the transitions on-line on an MP-RAM. Indeed, $\A$ has $O(w+1)^k$ state, where $k$ is the size of the structure encoding the $q$ patterns $m_1,\ldots,m_q$; for $w$ and $q$ large, the time and space complexity for computing the states of $\A$ becomes prohibitive, whence the need to compute the states on-line quickly without having to precompute nor store them. We introduced MP-RAMs to this end.
[*Pattern-matching*]{} algorithms are often given on RAMs. This model is not good when there are too many different values to be stored, for example ${O(w+1)^ k}$ states for $\A$. As early as 1974, the motivation of [@prs] for introducing “vector machines” was the remark that boolean bit-wise operations and shifts which are implemented on computers are faster and better suited for many problems. This work was the starting point of a series of papers: [@trl; @bg] comparing the complexities of computations on various models of machines allowing for boolean bit-wise operations and shifts with computation complexities on classical machines, such as Turing machines, RAMs etc. The practical applications of this technique to various [*pattern-matching*]{} problems start with [@byg; @wm]: they are known as [*bit-parallelism*]{}, or [*shift-OR*]{} techniques. We follow this track with the episode search problem, close to the problems studied in [@byg; @wm; @byn], albeit different from these problems.
In the sequel, we use a variant of RAMs, which is a more realistic computation model in some aspects, and we encode $\A$ to ensure that (i) each state of $\A$ is stored in a single memory cell and (ii) only the most basic microprocessor operations are used to compute the transitions of $\A$. Our RAMs have the same control structures as classical RAMs[^1], but the operations are enriched by allowing for boolean bit-wise operations and shifts, which we will preferably use whenever possible. Such RAMs are close to microprocessors, this is why we called them MP–RAMs.
An MP–RAM is a RAM extended by allowing new operations:
the bit-wise [and]{}, denoted by $\&$,
the [*left shift*]{}, denoted by $\sl$ or $shl$, and
the [*right shift*]{}, denoted by $\sr$ or $shr$.
The new operations are low-level operations, executable much faster than the more complex [*MULT, DIV*]{} operations.
Assume our have unbounded memory cells. We will have for example: $(10110 \ \& \ 01101) = 100$, $(10110 \sl 4) =101100000$ and $(10110\sr 3) = 10$. If memory cells have at most 8 bits, we will have: $(10110 \sl 4) = 1100000$, that will be written as $(00010110 \sl 4) =01100000$.
Parallel search of several patterns {#sect.plusmot}
===================================
Let us recall the problems. Given patterns $m_1, m_2,\ldots,m_q$, we can:
- either count the number of occurrences of each $m_i$ in a $w$-window (not necessarily the same one);
- or count the number of $w$-windows containing $m_1,m_2,\ldots,m_q$.
The algorithm we described in [@bcgm] for counting the number of $w$-windows containing [*a single*]{} pattern $m$ can be adapted to all these cases, only the acceptance or counting condition will change.
To search simultaneously several patterns $m_1,\ldots,m_q$, [@wm] propose a method concatenating all the patterns. To search simultaneously several episodes $m_1,\ldots,m_q$, we generalise our algorithm [@bcgm]: we use $q$ counters $c_1,\ldots,c_q$ initially set to 0, and we define an appropriate multiple counting condition such that each time $m_i$ is in a $w$-window, the corresponding counter $c_i$ is incremented. This method has a drawback: if the patterns are too long, it will need more than one memory cell for coding the states of the automaton. For searching multiple patterns the method proposed by [@dfggk] to optimise the search, when words $m_1,\ldots,m_q$ have common prefixes, is to organise $m_1,\ldots,m_q$ in a [*trie*]{} [@k] before applying the standard algorithm. We apply our algorithm on MP-RAMs in a similar way, and implement [*tries*]{} in a new way. We thus can encode the set of patterns compactly, and then encode the states of the automaton on a single memory cell.
Representing patterns by a trie
-------------------------------
Consider for example episodes $m_1=tu$, $m_2=tue$, and $m_3=tutu$. We choose this example because it illustrates most of the difficulties in encoding the automaton: episode [*taie* ]{} is very simple because all letters are different, [*tati* ]{} is less simple because there are two occurrences of $t$ which must be distinguished, [*tutu* ]{} a bit more complex (the first occurrence of [*tu*]{} must be distinguished from the second one), [*turlututu*]{} would be even more complex. We represent these three episodes by the trie $t$ pictured in figure \[trie.fig\].
We implement this trie $t$ by the three tables below:
[r|c|c|c|c|c|]{}
------------------------------------------------------------------------
$tr=$&t&u&e&t&u\
[r|c|c|c|c|c|]{}
------------------------------------------------------------------------
$pr=$&0&1&2&2&4\
[r|c|c|c|]{}
------------------------------------------------------------------------
$f=$&2&3&5\
Table $tr$ represents the “flattened” trie. Predecessors are in table $pr$: $pr[i]$ gives the index in $tr$ of the parent of $tr[i]$ in the trie; 0 means there is no predecessor and hence it is a pattern start[^2]. Finally $f$ marks patterns ends: $f[i]$ is the index in $tr$ of the end of pattern $i$.
Preprocessing the trie and algorithm
------------------------------------
We preprocess the trie of patterns and this gives us a finite state automaton $\A$. Its alphabet is $A$. The states are the $k$-tuples of integers $\langle l_1,\ldots,l_k\rangle$ with $l_j$ belonging to $\{1,\ldots,w, +\infty\}$, where $k$ is the size of table $tr$ and $w$ the window size.
We describe informally the behaviour of $\A$. $\A$ scans $t$, it will be in state $\langle l_1,\ldots,l_k\rangle$ after scanning $t_1\ldots t_m$ iff $l_i$ is the length of the shortest suffix[^3] of $t_1\ldots t_m$ shorter than $w$ and containing $tr[j_i]\ldots tr[i]$ as subsequence for $i= 1,\ldots,k$, where $tr[j_i]$ $\ldots$ $tr[i]$ is the sequence of letters labelling the path going from the root of the trie to the node represented by $tr[i]$. If no suffix (of length less than $w$) of $t_1\ldots t_m$ contains $tr[j_i]\ldots tr[i]$ as a subsequence, we let $l_i=+\infty$.
Let us now describe our algorithm. Let $\Omega$ be the least integer such that $w+2\leq 2^\Omega$. The rôle of $+\infty$ is played by $2^\Omega-1$, whose binary encoding is a sequence of $\Omega$ ones. We define the function Next$_{\Omega}$ by: $${\rm Next}_{\Omega}(l)=\cases{l+1,&if $l<2^\Omega-1$;\cr
2^\Omega-1,&else.\cr}$$
State $\langle l_1,\ldots,l_k\rangle$ is encoded by integer: $$L=\sum_{i=1}^k l_i(2^{\Omega+1})^{i-1}
=\sum_{i=1}^k \Big( l_i\sl
\big((\Omega+1)(i-1)\big)\Big).\label{codstate.eq}$$
[$$\overbrace{0\,\vdots\!\underbrace{\raise 2pt\hbox{\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_}}_{\hbox{\footnotesize binary expansion of $l_k$}}\!|\ \ \ \ \ldots\ \ \ \ 0\,\vdots\!\underbrace{\raise 2pt\hbox{\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_}}_{\hbox{\footnotesize binary expansion of $l_2$}}\! |\ \ 0\,\vdots\!\underbrace{\raise 2pt\hbox{\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_}}_{\hbox{\footnotesize binary expansion of $l_1$}}\!|}^{\hbox{\footnotesize binary expansion of $L$}}$$ ]{}
Let $\overline{l_i}$ denote the binary expansion of $l_i$, $i=1,\ldots,k$, prefixed by zeros in such a way that $\overline{l_i}$ occupies $\Omega$ bits (all $l_i$s are smaller than $2^\Omega-1$, hence they will fit in $\Omega$ bits). The binary expansion of $L$ is obtained by concatenating the $\overline{l_i}$s, each prefixed by a zero (figure \[es.fig\]). These initial zeros are needed for implementing function [Next]{}$_{\Omega}$ to indicate overflows. Every integer smaller than $2^{k(\Omega+1)}$ can be written as $k$ [*big blocks*]{} of $(\Omega+1)$ bits, the first bit of each big block is 0 (and is called the [*overflow bit*]{}) and the $\Omega$ remaining bits constitute a [*small block*]{}. The blocks are numbered 1 to $k$ from right to left (the rightmost block is block 1, the leftmost block is block $k$).
\[ex.fig\]
[r|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|]{}
------------------------------------------------------------------------
$L=$&0&$\overline {l_5}$&0&$\overline {l_4}$&0&$\overline
{l_3}$&0&$\overline {l_2}$&0&$\overline {l_1}$\
By the definition in equation (\[codstate.eq\]), the initial state $\langle +\infty,\ldots,+\infty\rangle$ is encoded by: $$I_0=\sum_{i=1}^k
(2^\Omega-1)2^{(\Omega+1)(i-1)}
=\sum_{i=1}^k \Big(\big((1\sl \Omega )-1\big)\sl
{\big((\Omega+1)(i-1)\big)\Big)}.$$
One might see a multiplication here. In fact we will need a loop for $i=1$ to $k$. We will execute each time we go through the loop a shift of $\Omega +1$, and the multiplication will disappear. All equations below are treated in the same way.
Assume that the window size is $w=13$ hence $\Omega = 4$. With the notations of figure \[fig0\], state $l=\langle 2,5,\infty,5,\infty\rangle$ is encoded by:
[r|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|]{}
------------------------------------------------------------------------
$L=$&0&$\overline
{15}$&0&$\overline {5}$&0&$\overline
{15}$&0&$\overline {5}$&0&$\overline
{2}$\
The initial state is represented by:
[r|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|]{} $I_0=$&0&1111&0&1111&0&1111&0&1111&0&1111\
or, writing $\underline {1}$ instead of the $\Omega$ ones representing $\infty$:
[r|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|]{}
------------------------------------------------------------------------
$I_0=$&0&$\underline {1}$&0&$\underline
{1}$&0&$\underline {1}$&0&$\underline {1}$&0&$\underline
{1}$\
In transition $l=\langle l_1,\ldots,l_k\rangle \drv \sigma l'=\langle
l'_1,\ldots,l'_k\rangle$, the $l'_i$ component of the new state $l'$ is either [Next]{}$_\Omega(l_{pr[i]})$ or [Next]{}$_\Omega(l_{i})$ according to whether the scanned letter $\sigma$ is equal to $tr[i]$ or not. The cases $l'_i= {Next}_\Omega(l_{pr[i]})$ and $l'_i= {Next}_\Omega(l_{i})$ respectively yield a [*first type computation*]{} and a [*second type computation*]{}.
To generalise the algorithm of [@bcgm], we must define several [*masks*]{} $M_\sigma$ for each letter $\sigma$ of alphabet A. If $\sigma$ has several occurrences in table $tr$, we will need as many masks $M_\sigma$ as occurrences $tr[i]$ and $tr[i']$ of $\sigma$ with $j=i-pr[i]\not=i'-pr[i']=j'$ (a single mask will suffice for the set of all occurrences such that $i-pr[i]$ has the same value $j$, because they correspond to the same shift of $j$ big blocks). The $M_\sigma^j$ are the masks preparing first type computations. Precisely, if $tr[i]=\sigma$ and $i-pr[i]=j$, the operation $\big(L\sl j(\Omega+1)\big)\& M_\sigma^j$ will shift everything of $j$ big blocks leftwards and will erase the blocks for which $\sigma\not= p_i$ or $i-pr[i]\not=j$. For $i>1$, the $i$-th block will thus contain $\overline{l_{pr[i]}}$ iff $tr[i]=\sigma$ and $i-pr[i]=j$. It will contain 0 otherwise.
In our example ($m_1=tu$, $m_2=tue$, and $m_3=tutu$), we will need two masks $M_t$ but a single mask $M_u$ will suffice:
[r|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|]{} $M_t^1=$&0&0&0&0&0&0&0& 0&0&1\
[r|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|]{} $M_t^2=$&0&0&0&1&0&0&0& 0&0&0\
[r|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|]{} $M_u^1=$&0&1&0&0&0&0&0& 1&0&0\
[r|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|]{} $M_e^1=$&0&0&0&0&0&1&0& 0&0&0\
where $\underline 0=0000$ and $\underline 1=1111$.
Mask $N_\sigma$ is the complement of $\sum_j M_\sigma^j$, preparing second type computations. The operation $L\& N_\sigma$ will erase the blocks for which $\sigma= tr[i]$. For our example, we have:
[r|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|]{} $N_t=$&0&1&0&0&0&1&0&1&0 &0\
[r|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|]{} $N_u=$&0&0&0&1&0&1&0&0&0 &1\
[r|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|]{} $N_e=$&0&1&0&1&0&0&0&1&0 &1\
Generally, if $k$ is table $tr$ size, $$M_\sigma^j=\sum_{\scriptstyle tr[i]=\sigma \hbox{ and } pr[i]=i-j\atop
1\leq i\leq k}
\Big(\big((1\sl \Omega)-1\big)\sl {\big((\Omega+1)(i-1)\big)\Big)}.$$ and $$N_\sigma=\sum_{\scriptstyle p_i\not=\sigma\atop
1\leq i\leq k}
\Big(\big((1\sl \Omega)-1\big)\sl
{\big((\Omega+1)(i-1)\big)\Big)}.$$ $N_\sigma$ is the complement of $\sum_j M_\sigma^j$.
Transition $l=\langle l_1,\ldots,l_k\rangle \drv \sigma l'=\langle
l'_1,\ldots,l'_k\rangle$ is computed by:
$$T=\sum_j \big((L\sl j({\Omega+1}))\& M_\sigma^j\big)+(L\& N_\sigma) + E_1$$ where:
[l|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l| r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|]{} $E_1=$&0&$0001$&0&$0001$&0&$0001$&0&$0001$&0&0001\
Adding $E_1$ amounts to add 1 to each small block.
In our example, if we scan letter $t$, the transition is computed by: $$T=\big((L\sl 2({\Omega+1}))\& M_t^2\big)+
\big((L\sl ({\Omega+1}))\& M_t^1\big)+(L\& N_t) + E_1$$ yielding for $l=\langle 2,5,\infty,5,\infty\rangle$, encoded by:
[r|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|]{}
------------------------------------------------------------------------
$L=$&0&$\overline
{15}$&0&$\overline {5}$&0&$\overline
{15}$&0&$\overline {5}$&0&$\overline
{2}$\
the result:
[r|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|]{}
------------------------------------------------------------------------
$T=$&1&$\overline
{0}$&0&$\overline {6}$&1&$\overline
{0}$&0&$\overline {6}$&0&$\overline
{1}$\
All the blocks contain the correct result, except for the leftmost block and the middle block where an overflow occurred. To treat blocks where overflow occurred it suffices of initialise again these blocks by replacing $T$ with $L'=T-\big((T\& E_2)\sr\Omega\big)$, where:
[l|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|]{} $E_2=$&1&0&1&0&1&0&1&0&1&0\
We find:
[l|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|]{} $T\& E_2=$&1&0&0&0&1&0&0&0&0&0\
Hence:
[l|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|]{} $(T\& E_2)\sr\Omega=$&0&$\overline 1$&0&0&0&$\overline 1$&0&0&0&0\
and finally:
[r|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|r@[ $\vdots$ ]{}l|]{}
------------------------------------------------------------------------
$L'=T-\big((T\& E_2)\sr\Omega\big)=$&0&$\overline
{15}$&0&$\overline {6}$&0&$\overline
{15}$&0&$\overline {6}$&0&$\overline
{1}$\
Last we define a counter $c_i$ for each pattern $m_i$, and increment it whenever $l_{f[i]}<w+1$, which is implanted by: $M_i\& L < (w+1) 2^{(\Omega+1)(f[i]-1)}$, for $i=1,\ldots,k$, where $M_i= \big((1\sl \Omega )-1\big)\sl {\big((\Omega+1)(f[i]-1)\big)}$.
Our algorithm treats the more complex case where we demand that all episodes appear in a same window, a case that cannot be treated by the separate counting of the number of windows containing each episode. A simple modification of the counting condition enables us to also count [*with a single scan of the text*]{} the number of windows containing each individual episode, in a more efficient way than if the text were to be scanned for each episode.
\[mult.thm\] There exists an on-line algorithm in time $O(nq)$ solving the parallel search of $q$ serial episodes in a size $n$ text (assuming the episode alphabet has at most $\sqrt n/q$ letters) on MP–RAM.
Let $\alpha$ be the number of letters of the alphabet. As in [@dfggk], we treat in the same way all letters not occurring in the patterns; this leads to defining two masks $M_{\it other}$ and $N_{\it other}$ common to all such letters. Let $|w|$ be the length of the binary expansion of $w$. The algorithm consists of four steps:
compute (at most) $q\times (k+1)$ integers representing the masks $M_\sigma^j$, $(k+1)$ integers representing the masks $N_\sigma$ and the integers $\Omega,\Delta, I_0,F,E_1,E_2$; all these integers are of size $k(|w|+2)$ and are computed [*simultaneously*]{} in $k$ iterations at most. The integer $k$ is the size of the trie representing the patterns: $k \leq \sum_{i=1}^q |m_i| \leq\sqrt n$.
let $c=0$ ($c$ is the number of $w$-windows containing all the patterns).
let $L=I_0$.
scan text $t$; after scanning $t_i$, compute the new state $L$ ([*on-line and without preprocessing*]{} with an MP–RAM) and if $c_i<w$ for $i=1,\ldots,q$, increment $c$ by 1.
Our algorithm uses only the simple and fast operations $\&$, together with a careful implementation of $\sl,\sr$ and addition. Step 1 of preprocessing is in time $qk(k+1)+q(k+1)+\log(w)
\leq q(\sqrt n)^2 + 2q\sqrt n + q + \log(w)=O(nq)$; in general, $k$, $q$ and $w$ are smaller than $n$ by several orders of magnitude and we will have: $qk(k+1)+q(k+1)+\log(w)=o(n)$. In step 4 we scan text $t$ linearly in time $O(n)$ and perform $q$ comparisons (one for each counter $c_i$). Complexity is thus in time $nq$, hence finally a time complexity $O(nq)$ for the algorithm.
Experimental results {#sect.exp}
====================
The algorithm on MP–RAM has a better complexity than the standard algorithm, however, the underlying computation models being different, we checked experimentally that the MP–RAM algorithm is faster. We implemented all algorithms in C++. Experiments were realised on a PC (256 Mo, 1Ghz) with Linux. The text was a randomly generated file. We measured the time with machine clock ticks.
For searching multiple patterns, we took 3 to 5 patterns of length 2 to 4; in figure \[patternmult.fig\], case (a) is the case of patterns having no common prefix, and case (b) is the case of patterns having common prefixes. In case (a), the MP–RAM algorithm where we concatenate the patterns is at least twice as fast as the standard “naive” algorithm where patterns are concatenated; both standard algorithms (with patterns concatenated or organised in a trie) are equivalent, the algorithm with concatenation being slightly faster; this was predictable since a trie organisation will not give a significant advantage in that case; the MP–RAM algorithm where the patterns are organised in a trie is 30 to 50% faster than the standard algorithm with trie, and 10 to 15% slower than the MP–RAM algorithm where the patterns are concatenated. However, as soon as the total length of the patterns is larger than 7 or 8, or the window size is larger than 30, if patterns are concatenated, the automaton state can no longer be encoded in a single 32 bits memory cell, and it is better to use the MP–RAM algorithm with trie (figure \[patternmult.fig\] case (b)). Figure \[patternmult.fig\] case (b) shows that, for patterns having common prefixes, the MP–RAM algorithm with trie is 1.3 to 1.5 times faster than the standard algorithm with trie, itself 1.4 to 1.6 times faster than the standard algorithm with concatenation.
Conclusion
==========
We presented new algorithms for multiple episode search, much more efficient than the standard algorithms. This was confirmed by our experimental analysis. Note that with our method, counting the number of windows containing several episodes is not harder than checking the existence of one window containing these episodes. This is not true with most other problems; usually counting problems are much harder than the corresponding existence problems: for example, for the “matching with don’t cares” problem, the existence problem is in linear time while the counting problem is in polynomial time [@kr] and in the particular case of [@mby], the existence problem is in logarithmic time while the counting problem is in sub-linear time.
[abcde]{}
A. Aho, Algorithms for Finding Patterns in Strings, in Handbook of Theoretical Computer Science, Vol. 1, van Leeuwen Ed., North-Holland, Amsterdam (1990), pp. 255–300.
A. Aho, J. Hopcroft, J. Ullman, Design and Analysis of Computer Algorithms, Addison-Wesley, London (1974).
R. Baeza-Yates, B. Ribeiro-Neto, Modern Information Retrieval, ACM Press Books, New-York (1999).
R. Baeza-Yates, G. Gonnet, A new approach to text searching, Communications of the ACM, Vol 35 (1992), 74–82.
R. Baeza-Yates, G. Navarro, A faster algorithm for approximate string matching, Proc. 1996 Combinatorial Pattern Matching Conf., LNCS 1075, Springer-Verlag, Berlin (1996), pp. 1–23.
A. Ben-Amram, Z. Galil, On the power of the shift instruction, Inf. Comput. Vol 117 (1995), pp. 19–36.
L. Boasson, P. Cegielski, I. Guessarian, Y. Matiyasevich, Window Accumulated Subsequence Matching is linear, Annals of Pure and Applied Logic Vol. 113 (2001), pp. 59-80.
M. Crochemore, String-matching with constraints, Proc. MFCS’88, LNCS 324, Springer-Verlag, Berlin (1988), pp. 44–58.
M. Crochemore, W. Rytter, Text Algorithms, Oxford University Press, Oxford (1994).
M. Crochemore, C. Hancart, T.. Lecroq, Algorithmique du text, Vuibert, Paris (2001).
G. Das, R. Fleischer, L. Gasienic, D. Gunopoulos, J. Kärkkäinen, Episode Matching, Proc. 1997 Combinatorial Pattern Matching Conf., LNCS 1264, Springer-Verlag, Berlin (1997), pp. 12–27.
Z. Galil, String matching in real time, J. Assoc. Comput. Mac. Vol 28, (1981), pp. 134–149.
D. Knuth, The art of computer programming, Vol. 1, Fundamental algorithms, Addison-Wesley, Reading (1997).
D. Knuth, J. Morris, V. Pratt, Fast Pattern Matching in Strings, SIAM Journal of Comput. Vol 6(2), (1977), pp. 323–350.
G. Kucherov, M. Rusinovitch, Matching a Set of Strings with variable Length Don’t Cares, Theor. Comput. Sc. Vol 178, (1997), pp. 129–154.
U. Manber, R. Baeza-Yates, An Algorithm for String Matching with a Sequence of Don’t Cares, Inform. Proc. Letters Vol 37, (1991), pp. 133–136.
H. Mannila, Local and Global Methods in Data Mining: Basic Techniques and open Problems, Proc. ICALP 2002, LNCS 1186, Springer-Verlag, Berlin (2002).
H. Mannila, Methods and Problems in Data Mining, Proc. 1997 ICDT Conf., LNCS 1186, Springer-Verlag, Berlin (1997), pp. 41–55.
H. Mannila, H. Toivonen, A. Verkamo, Discovering Frequent Episodes in Sequences, Proc. 1995 KDD Conf., (1995), pp. 210–215.
Y. Matiyasevich, Real-time recognition of the inclusion relation, Zapiski Nauchnykh Leningradskovo Otdeleniya Mat. Inst. Steklova Akad. Nauk SSSR, Vol. 20, (1971), pp. 104–114. Translated into English, Journal of Soviet Mathematics, Vol. 1, (1973), http://logic.pdmi.ras.ru/\~ -0.5em yumat/Journal, pp. 64–70.
G. Navarro, M. Raffinot, Flexible Pattern Matching in Strings Practical on-line search algorithms for texts and biological sequences, Cambridge University Press, Cambridge (2002).
V. Pratt, M. Rabin, L. Stockmeyer, A charaterization of the power of vector machines, Proc. SToC 74, pp. 122-134.
A. Slissenko, String-matching in real time, LNCS 64, Springer-Verlag, Berlin (1978), pp. 493–496.
J. Trahan, M. Loui, V. Ramachandran, Multiplication, division and shift instructions in parallel random access machines, Theor. Comput. Sc. Vol. 100, (1992), pp.. 1–44.
Z. Tronicek, Episode matching, 12th Annual Symposium, Combinatorial Pattern Matching 2001, Jerusalem, LNCS 2089, Springer-Verlag, Berlin (2002), pp. 143-146.
E. Ukkonen, On-line construction of suffix-trees, Algorithmica, Vol. 14, (1995), pp.. 249–260.
S. Wu, U. Manber, Fast text searching, Communications of the ACM, Vol 35 (1992), 83–91.
[^1]: See [@ahu] pages 5–11, for a definition of classical RAMs.
[^2]: Numbering of indices starts at 1 in order to indicate pattern starts by 0.
[^3]: Word $s$ is a [*prefix*]{} (resp. [*suffix*]{}) of word $t$ iff there exists a word $v$ such that $t=sv$ (resp. $t=vs$).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
In this paper, we propose two analytical preconditioning strategies. One is parameter independent and easy to implement, the other one has the traditional affinity with respect to the parameters which allows for efficient implementation through an offline-online decomposition. Overall, the preconditioning improves the quality of the error estimation uniformly on the parameter domain, and speeds up the convergence of the reduced solution to the truth approximation. 0.5
[**Résumé**]{} 0.5On ' etend dans cette note la méthode de collocation réduite récemment introduite dans [@ChenGottlieb] au cas non linéaire et on propose deux stratégies de préconditionnement dont une est indépendante des paramètres et facile a mettre en oeuvre et l’autre possède la propriété classique de décomposition affine qui permet une mise en oeuvre rapide en-ligne/hors-ligne. Ces stratégies améliorent la qualité de l’approximation et la vitesse de convergence.
address:
- '1 - Department of Mathematics, University of Massachusetts Dartmouth, 285 Old Westport Road, North Dartmouth, MA 02747, USA. The research of the first author was partially supported by National Science Foundation grant DMS-1216928. The research of the second author was partially supported by AFOSR grant FA-9550-12-1-0224.'
- '2 - Sorbonne Universités, UPMC Univ Paris 06, UMR 7598, Laboratoire Jacques-Louis Lions & Institut Universitaire de France, F-75005, Paris, France'
- '3 - Division of Applied Mathematics, Brown University, 182 George St, Providence, RI 02912, USA'
author:
- 'Yanlai Chen$^{1}$'
- 'Sigal Gottlieb$^{1}$'
- 'Yvon Maday$^{2,3}$'
title: Parametric Analytical Preconditioning and its Applications to the Reduced Collocation Methods
---
,
,
,
[Presented by Olivier Pironneau]{}
Version française abrégée {#version-française-abrégée .unnumbered}
=========================
La méthode de base réduite classique (RBM)[@Barrett_Reddien; @Noor_Peters; @Peterson; @Prudhomme_Rovas_Veroy_Maday_Patera_Turinici] pour l’approximation de la solution d’équations aux dérivées partielles (EDP) paramétrées du type $[\mathbb{L} (\mu)\, u_\mu] (x) = f(x; \mu), \,\, x \in \Omega \,\,\,{\textcolor{blck}{\subset \mathbb{R}^n}}$ repose sur la définition d’un espace de discrétisation ad’hoc, engendré par des solutions particulières de l’EDP en certain paramètres bien choisis. Ces solutions particulières doivent être préalablement approchées par une méthode traditionnelle spectrale ou d’éléments finis par exemple. Elle est principalement développée dans le cadre variationnel et permet la résolution en temps bien plus faible que des méthodes traditionnelles. Dans certain cadres, néanmoins, l’approche de collocation est préférable à l’approche variationnelle, en particulier lorsque la physique est complexe. La méthode de collocation réduite récemment introduite dans [@ChenGottlieb] permet de poser le problème de cette façon. Ainsi lorsque la méthode traditionnelle est de type spectrale collocation où, après avoir définit un opérateur discret $\mathbb{L}_\calN (\mu)$, on cherche un polynôme $u^\calN_\mu$ tel que $[\mathbb{L}_\calN (\mu)\, u^\calN_\mu] (x_j) = f(x_j; \mu) $ est vérifié exactement sur un ensemble de $\calN$ points de collocation $C^\calN = \{x_j\}_{j=1}^\calN$, l’approche de collocation réduite propose une approximation $u^{(N)}_{\mu^*} $: $u^{(N)}_{\mu^*} = \sum_{j=1}^{N} c_j(\mu^*) u^\calN_{\mu^j}$ vérifiant (\[eq:rbmpde\]) soit au sens des moindre carré (LSRCM), puisqu’il y a plus de point $x_k$ que de coefficients $c_j$ (en effet $N<\!<\calN$) soit seulement en certain points bien choisis $ x \in C_R^N$ (ERCM).
Les méthodes de collocation sont connues pour être moins stables que les méthodes variationnelles. Pour rectifier cet inconvévient, nous proposons deux types de préconditionnement analytique, basés sur la définition d’un opérateur de prćonditionnement $P$ et sur une approximation de $P \mathbb{L}_\calN (\mu)\, u^\calN_\mu (x_j) \simeq P f(x_j; \mu)$ dans les deux sens précédents. Les deux opérateurs de préconditionnement analytiques sont : une version indépendante du paramètre $P^{\mu^c} :=\mathbb{L}_\calN(\mu^c)^{-1}$, qui améliore l’approximation surtout au niveau de la valeur barycentrale $\mu^c$ et une version paramétrée qui, dans le cas où l’ensemble des paramètres est le carré $[0,1]^2$ repose sur une interpolation $Q_1$ entre les 4 valeurs de $\mathbb{L}_\calN(\mu)^{-1}$ aux coins du domaine paramétrique: $P^I(\mu) = P_{00} (1 - \mu^1) (1 - \mu^2) + P_{01} (1 - \mu^1) \mu^2 + P_{10} \mu^1 (1 - \mu^2) + P_{11} \mu^1 \mu^2$. Les illustrations numériques des performance de ces deux préconditionnement analytiques sont proposées dan les figures 2 et 3. La figure 2 illustre la comparaison des trois opérateurs analytiques de préconditionnement : sur la gauche sont tracés les indices d’effectivité de l’estimation de l’erreur ÊÊ sur le système avec ces opérateurs de préconditionnement. Sur la droite sont tracés ÊÊÊÊ la pire des convergences selon ces scénarii. La figure 3 illustre l’histoire de la convergence selon les opérateurs analytiques de préconditionnement pour l’approche des moindres carrés (à gauche) et l’approche empirique de collocation (à droite).
Enfin une extension de l’approche de collocation réduite empirique aux cas d’EDP non linéaire est aussi proposée et consiste naturellementt en la vérification de l’EDP non linéaire en des points de collocation choisis de façon empirique.
Introduction
============
The Reduced Basis Method (RBM)[@Barrett_Reddien; @Noor_Peters; @Peterson; @Prudhomme_Rovas_Veroy_Maday_Patera_Turinici] has been developed to numerically solve PDEs in scenarios that require a large number of numerical solutions to a parametrized PDE, and in which we are ready to expend significant computational time to pre-compute data that can be later used to compute accurate solutions in real-time. The RBM splits the solution procedure into two parts: an offline part where a greedy algorithm is utilized to judiciously select $N$ parameter values for pre-computation; and an online part where the solution for many new parameter is efficiently approximated by a Galerkin projection onto the low-dimensional space spanned by these $N$ pre-computed solutions.
While Galerkin methods (that are mostly used for RBM) are derived by requiring that the projection of the residual onto a prescribed space is zero, collocation methods require the residual to be zero at some pre-determined collocation points. They are attractive for their ease of implementation, particularly for time-dependent nonlinear problems [@HesthavenGottlieb2007; @TrefethenSpecBook]. In [@ChenGottlieb], two of the authors developed the so-called Reduced Collocation Method (RCM). It adopts the RBM idea for collocation methods providing a strategy to practitioners who prefer a collocation, rather than Galerkin, approach. Our current implementation of this new method uses collocation for both the truth solver and the online reduced solver, but the offline part could be based on a variational approach as well. Furthermore, one of the two approaches in [@ChenGottlieb], the empirical reduced collocation method (ERCM) allows to eliminate a potentially costly online procedure that is needed for non-affine problems with a Galerkin approach. The method’s efficiency matches (or, for non-affine problems, exceeds) that of the traditional RBM in the Galerkin framework.
However, collocation methods may suffer from bad conditioning. In this paper, we propose and test two analytical preconditioning strategies to address this issue in the parametric setting of RCM. One strategy is parameter independent, which is advantageous for ease of implementation. The other one is parameter dependent, but has the traditional affinity with respect to the parameters which allows extremely efficient implementation through an offline-online decomposition. Overall, we show that the preconditioning uniformly improves the quality of the approximation, and speeds up the convergence of the solution process without adversely impacting the efficiency of the method in any significant way. In Section \[sec:alg\], we briefly review RCM and describe our analytical preconditioners. Numerical results are provided in Section \[sec:numerical\].
The Algorithms {#sec:alg}
==============
We begin with a linear parametrized PDE depending on a parameter $\mu \in \calD\subset \R^d$, written in a strong form as $[\mathbb{L} (\mu)\, u_\mu] (x) = f(x; \mu), \,\, x \in \Omega \,\,\,{\textcolor{blck}{\subset \mathbb{R}^n}}$ with appropriate boundary conditions. We approximate the solution to this equation using a collocation approach: for any $\mu \in \calD$, we define a discrete differentiation operator $\mathbb{L}_\calN (\mu)$, and a discrete (polynomial) solution $u^\calN_\mu$ such that $[\mathbb{L}_\calN (\mu)\, u^\calN_\mu] (x_j) = f(x_j; \mu) $ on a given set of collocation points $C^\calN = \{x_j\}_{j=1}^\calN$, usually taken as a tensor product of $\calN_x$ collocation points for each dimension that is allowed by rectangular domains. We assume that the resulting approximate solution $u^\calN_\mu$ is highly accurate and refer to it as the “truth approximation”.
[*2.1 Online algorithms*]{}
For completeness, we briefly outline the RCM [@ChenGottlieb]. The idea is that when the solution for any parameter value $\mu^* \in \calD $ is needed, instead of solving for the costly truth approximation $u^\calN_{\mu^*}$, we somehow combine $N$ pre-computed truth approximations $u^\calN_{\mu^1},\, \dots,\, u^\calN_{\mu^N}$ to produce a surrogate solution $u^{(N)}_{\mu^*} $: $u^{(N)}_{\mu^*} = \sum_{j=1}^{N} c_j(\mu^*) u^\calN_{\mu^j}.$ The key feature of the algorithm is the requirement that the surrogate solution $u^{(N)}_{\mu^*} $ will satisfy the discretized differential equation in some sense $\mathbb{L}_\calN(\mu^*) u^{(N)}_{\mu^*} \approx f({\textcolor{blck}{\cdot}}; \mu^*).$ Exploiting the linearity of the operator, we observe that the system of equations we to solve is : find $\vec{c}(\mu^*)$ such that $$\vec{c}(\mu^*) = (c_1(\mu^*), c_2(\mu^*), \dots, c_N(\mu^*))^T,\quad
\sum_{j=1}^{N} c_j(\mu^*) [\mathbb{L}_\calN(\mu^*) u^\calN_{\mu^j}](x_k) \approx f(x_k; \mu^*) \; \; \; \;
.
\label{eq:rbmpde}$$
Satisfying the above equation exactly for $k=1, . . ., \calN$ is usually an overdetermined system since we have only $N$ unknowns, but $\calN >> N$ equations.
[**Least squares approach.**]{} When faced with an overdetermined system, we can determine the coefficients by satisfying the equation in a least squares sense: we define, for any $\mu^*$, an $\calN \times N$ matrix $\mathbb{A}_N (\mu^*)$ with $j^{\rm th}$ column $\mathbb{L}_\calN(\mu^*)\, u_{\mu^j}^\calN$, and vector of length $\calN$, $\boldf^\calN_j (\mu^*)= f(x_j ; \mu^*) \; \; x_j \in C^\calN,$ and solve $\mathbb{A}_N^T(\mu^*)\,\mathbb{A}_N(\mu^*) \,\vec{c}(\mu^*) = \mathbb{A}_N^T(\mu^*)\, \boldf^\calN(\mu^*)$ to obtain $\vec{c}(\mu^*)$.
[**Reduced Collocation approach.**]{} Our second approach is more natural from the collocation point-of-view. We determine the coefficients ${\mathbf{c}}(\mu^*)$ by enforcing at a reduced set of collocation points $C_R^N$. In other words, we solve $\sum_{j=1}^{N} c_j(\mu^*) [ \mathbb{L}_\calN(\mu^*) u^\calN_{\mu^j}] (x) = f(x;\mu^*)$, for $ x \in C_R^N$, which can also be written as $\mathbb{{\textcolor{blck}{I}}}_\calN^N \mathbb{A}_N(\mu^*) \,\vec{c}(\mu^*) = \mathbb{{\textcolor{blck}{I}}}_\calN^N \boldf^\calN$, where $ \mathbb{{\textcolor{blck}{I}}}_\calN^N$ is a $N \times \calN$ matrix, that extracts the $N$ values of a $\calN-$vector associated to the indices of the reduced set of collocation points. Later we will demonstrate how this set of points can be determined, together with the choice of basis functions, through the greedy algorithm.
[*2.2 Offline-online decomposition*]{}
The size of the matrix we need for solving $\vec{c}(\mu^*)$ for each new $\mu^*$ is $N \times N$, but its assembly is not obviously independent of $\calN$. For that purpose, we need the affine assumption[^1] on the operator as in the Galerkin framework. Thus, the overall online component is independent of $\calN$ after a preparation stage where all the parameter independent quantities are precomputed [@ChenGottlieb]. We remark that there are remedies available when the parameter-dependence is not affine [@Barrault_Nguyen_Maday_Patera].
[*2.3 Analytical Preconditioning*]{}
Collocation methods are frequently ill-conditioned. The situation is exacerbated when we form the normal equation in the Least Squares approach. In this section, we propose two analytical preconditioning techniques. One is parameter-independent and the other is parameter-dependent. Both will provide an operator $P$ such that the discretization is based on the minimization of $P \mathbb{L}_\calN (\mu)\, u^\calN_\mu (x_j) - P f(x_j; \mu)$ and the reduced problem in, e.g. the second approach becomes $\sum_{j=1}^{N} c_j(\mu^*) \mathbb{{\textcolor{blck}{I}}}_\calN^N \left(P \mathbb{L}_\calN(\mu^*) u^\calN_{\mu^j}\right) = \mathbb{I}_\calN^N (P f).$ [*Parameter-independent approach:*]{} We propose using $P^{\mu^c} :=\mathbb{L}_\calN (\mu^c)^{-1}$ as a preconditioning operator. Here $\mu^c$ is the center of the parameter domain $\calD$. We remark that this preconditioner is in general, ideal for $\mu = \mu^c$ (making condition number exactly $1$). Moreover, it is affordable in the parametric setting since we can perform the $\calN-$dependent operations for the offline preconditioning once for all.
$P^{\mu^c}$ works well. However, it is more effective when $\mu$ is close to $\mu^c$. To have a preconditioning operator working well uniformly on the parameter domain, we need a parameter-dependent one. In addition, for the preconditioning to be meaningful in our parametric setting, a key requirement is that it satisfies an affine property similar to those for the operator $\mathbb{L}_N(\mu)$.
Assuming our ($d-$dimensional) parameter domain is rectangular, we form $2^d$ operators at the vertices of the domain: $ \mathbb{L}_\calN(\mu^{V_i})$ for $i = 1, \dots, 2^d$, find their inverses $P_{V_i} = \left(\mathbb{L}_\calN(\mu^{V_i}) \right)^{-1}$, and define the preconditioning operator through interpolation. In the case $d=2$ (we assume $\mu = (\mu^1,\mu^2) \in [0,1]^2$ without loss of generality), this is a $Q_1$ interpolation defined as below: $P^I(\mu) = P_{00} (1 - \mu^1) (1 - \mu^2) + P_{01} (1 - \mu^1) \mu^2 + P_{10} \mu^1 (1 - \mu^2) + P_{11} \mu^1 \mu^2,$ where $P_{ij}$ is $P_{V} $ with $V$ being the $(i,j)-$corner.
[*2.4 Offline algorithms*]{}
In this section we describe the two greedy algorithms for the least squares and the reduced collocation approaches for choosing the reduced basis set $\{ u_{\mu^1}^\calN, \dots,
u_{\mu^N}^\calN \}$. We assume that given $\{ u_{\mu^1}^\calN, \dots,
u_{\mu^i}^\calN \}$ we can compute an upper bound $\Delta_i(\mu)$ for the error of the reduced solution $u_\mu^{(i)}$ for any parameter $\mu$ [@ChenGottlieb].
Discretize the parameter domain $\calD$ by $\Xi$, and denote the center of $\calD$ by $\mu^c$. Randomly select $\mu^1$, solve $\mathbb{L}_\calN (\mu^1)\, u^\calN_{\mu^1}
(x) = f(x; \mu^1)$ for $x \in C^\calN$ and let $\xi_1^\calN = u^\calN_{\mu^1}$. For $i = 2, \dots, N$ do
- For all $\mu \in \Xi$, form $\mathbb{A}_{i-1}(\mu) = \left(\mathbb{L}_\calN(\mu)\, \xi_{1}^\calN, \mathbb{L}_\calN(\mu)\, \xi_{2}^\calN, \, \dots, \,
\mathbb{L}_\calN(\mu)\, \xi_{{i-1}}^\calN\right)$.
- For all $\mu \in \Xi$, solve $\mathbb{A}_{i-1}(\mu)^T\,\mathbb{A}_{i-1}(\mu) \,\vec{c} = \mathbb{A}_{i-1}^T(\mu)\,\boldf^\calN$ to obtain $u^{(i-1)}_\mu
= \sum_{j = 1}^{i - 1} c_j \xi^\calN_{j}$ and $\Delta_{i-1}(\mu)$.
- Set $\mu^i = argmax_{\mu}\,\,\Delta_{i-1}(\mu)$, and solve $\mathbb{L}_\calN (\mu^i)\, u^\calN_{\mu^i}(x) = f(x; \mu^i)$ for $x \in C^\calN$.
- Apply a modified Gram-Schmidt transformation, with inner product defined by\
$(u,v) \equiv \left(\mathbb{L}_\calN (\mu^c)u, \mathbb{L}_\calN (\mu^c) v\right)_{L^2(\Omega)}$, on $\left\{\xi^\calN_{1}, \xi^\calN_{2}, \dots, \xi^\calN_{i-1}, u^\calN_{\mu^i}\right\}$ to obtain $\left\{\xi_1^\calN, \xi_2^\calN, \dots, \xi_i^\calN\right\}$.
Randomly select $\mu^1$, solve $\mathbb{L}_\calN (\mu^1)\, u^\calN_{\mu^1}
(x) = f(x; \mu^1)$, let $x^1 = argmax_{x \in X} \,\, \left|u^\calN_{\mu^1}(x)\right|,\quad \xi^\calN_1 =
\frac{u^\calN_{\mu^1}}{u^\calN_{\mu^1}(x^1)}$. For $i = 2, \dots, N$ do
- Let $C_R^{i-1} = \left\{x^1,\dots,x^{i-1}\right\}$.
- For all $\mu \in \Xi$, solve $\sum_{j=1}^{i-1} c_j \mathbb{I}_\calN^N
\left(\mathbb{L}_\calN(\mu) u^\calN_{\mu^j}\right) = f(x; \mu) \,\, {\rm for} \,\, x \in C_R^{i-1}$ to obtain $u^{(i-1)}_\mu = \sum_{j = 1}^{i - 1} c_j u^\calN_{\mu^j}$.
- Set $\mu^i = argmax_{\mu \in \Xi}\,\,\Delta_{i-1}(\mu)$ and solve $\mathbb{L}_\calN (\mu^i)\, u^\calN_{\mu^i} (x) = f(x; \mu^i)$.
- Find $\alpha_1, \dots, \alpha_{i-1}$ such that, if we define $\xi^\calN_i = u^\calN_{\mu^i} - \sum_{j=1}^{i-1}
\alpha_j\,\xi^\calN_j$, we have $\xi^\calN_i(x^j) = 0$ for $j = 1, \dots, i-1$.
- Set $x^i = argmax_x \,\, \left|\xi^\calN_{i}\right|$ and $\xi^\calN_i = \frac{\xi^\calN_i}{\xi^\calN_i(x^i)}$.
-
[*2.5 Extension to the nonlinear case*]{}
Numerical Results {#sec:numerical}
=================
![Comparison of the three analytical preconditioning operators: On the left is the plot for the effectivity indices for the error estimate on the system with these preconditioning operators. On the right are the worst case scenario convergence plots.[]{data-label="fig:L2d_H1u"}](EllipSample.eps){width="\textwidth"}
![Comparison of the three analytical preconditioning operators: On the left is the plot for the effectivity indices for the error estimate on the system with these preconditioning operators. On the right are the worst case scenario convergence plots.[]{data-label="fig:L2d_H1u"}](EffIndx_Compare3.eps "fig:"){width="48.00000%"} ![Comparison of the three analytical preconditioning operators: On the left is the plot for the effectivity indices for the error estimate on the system with these preconditioning operators. On the right are the worst case scenario convergence plots.[]{data-label="fig:L2d_H1u"}](L2Down_H1Up_Compare3.eps "fig:"){width="48.00000%"}
{width="24.00000%"} {width="24.00000%"} {width="24.00000%"}
{width="35.00000%"} {width="35.00000%"}
In this section, we demonstrate the accuracy and efficiency of the proposed methods on a 2D diffusion-type problem with zero Dirichlet boundary condition [@ChenGottlieb]: $(1+ \mu_1 x) u_{xx} + (1+ \mu_2 y) u_{yy}= e^{4 x y}
\,\, \mbox{ on } \Omega = [-1,1] \times [-1,1].$ Our truth approximations are generated by a spectral Chebyshev collocation method [@TrefethenSpecBook; @HesthavenGottlieb2007]. For $C^\calN$, we use the Chebyshev grid based on points in each direction . We consider the parameter domain $\calD$ for $(\mu_1,\mu_2)$ to be $[-0.99,0.99]^2$. For $\Xi$, they are discretized uniformly by a $64 \times 64$ Cartesian grid. One sample solution for this problem is plotted in Figure \[fig:truthsample\].
In Figure \[fig:L2d\_H1u\] (Left), we show that while the non-parametric preconditioning $P^{\mu^c}$ give non-uniform improvement, the parametric preconditioning $P^I(\mu)$ improves effectivity indices by one order of magnitude. Figure \[fig:L2d\_H1u\] (Right) shows that the $P^{\mu^c}$ preconditioning operator improves the $L_2$ norm of the error but worsens the $H^1$ norm. In comparison, $P^I(\mu)$ improves the error in $L^2$ norm without significantly degrading (in some cases improving) the error in $H^1$ norm. Finally, we plot the stability constant of these preconditioned operators as a function of the parameter in Figure \[fig:howtoprec\]. We clearly see that $P^I(\mu)$ is most efficient in terms of enforcing the parametric stability number uniformly close to $1$.
For the preconditioned RCM, we can see, from Figures \[fig:IvsFD\_LS\_Ellip\] that the error for the least squares approach is around one order of magnitude better in the worst case scenario. For the empirical reduced collocation approach, the error is smaller and, more importantly, converges much more stably.
Concluding Remarks {#sec:conclude}
==================
We propose and test two analytical preconditioning strategies in the context of reduced collocation method. The parameter dependent one is shown to be capable of offline-online decomposition, improving both the quality of error estimation uniformly on the parameter domain, and enabling the preconditioned reduced collocation method to converge much faster and more stably than the non-preconditioned version.
[99]{}
M. Barrault, N. C. Nguyen, Y. Maday, and A. T. Patera. An “empirical interpolation” method: Application to efficient reduced-basis discretization of partial differential equations. , 339:667–672, 2004.
A. Barrett and G. Reddien. On the reduced basis method. , 75(7):543–549, 1995.
Y. Chen and S. Gottlieb. Reduced collocation methods: Reduced basis methods in the collocation framework. , 55(3):718–737, 2013.
J. S. Hesthaven, S. Gottlieb, and D. Gottlieb. , volume 21 of [ *Cambridge Monographs on Applied and Computational Mathematics*]{}. Cambridge University Press, Cambridge, 2007.
Y. Chen and S. Gottlieb. Reduced collocation methods: Reduced basis methods in the collocation framework. , 55(3):718–737, 2013.
A. K. Noor and J. M. Peters. Reduced basis technique for nonlinear analysis of structures. , 18(4):455–462, April 1980.
J. S. Peterson. The reduced basis method for incompressible viscous flow calculations. , 10(4):777–786, 1989.
C. Prud’homme, D. Rovas, K. Veroy, Y. Maday, A. T. Patera, and G. Turinici. Reliable real-time solution of parametrized partial differential equations: Reduced-basis output bound methods. , 124(1):70–80, March 2002.
L. N. Trefethen. , volume 10 of [*Software, Environments, and Tools*]{}. Society for Industrial and Applied Mathematics (SIAM), Philadelphia, PA, 2000.
[^1]: $\mathbb{L} (\mu)$ can be written as a linear combination of parameter-dependent coefficients and parameter-independent operators: $\mathbb{L} (\mu) = \sum_{q = 1}^{Q_a} a^{\mathbb{L}}_q(\mu) \mathbb{L}_q.$ Similarly, for $f$: $f(x; \mu) = \sum_{q = 1}^{Q_f} a^f_q(\mu) f_q (x).$
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We consider two varieties of labeled rooted trees, and the probability that a vertex chosen from all vertices of all trees of a given size uniformly at random has a given rank. We prove that this probability converges to a limit as the tree size goes to infinity.'
address: |
M. Bóna, Department of Mathematics, University of Florida, 358 Little Hall, PO Box 118105, Gainesville, FL 32611–8105 (USA)\
I. Mező, School of Mathematics and Statistics, Nanjing University of Information Science and Technology, Nanjing, 210044, (P.R. China)
author:
- Miklós Bóna and István Mező
title: Limiting probabilities for vertices of a given rank in rooted trees
---
Introduction
============
Let $\cal T$ be a class of rooted labeled trees. If $v$ is a vertex of a tree $T\in \cal T$, then let the *rank* of $v$ be the number of edges in the shortest path from $v$ to a leaf of $T$ that is a descendant of $v$. So leaves are of rank 0, neighbors of leaves are of rank 1, and so on. For a fixed $n$, consider all vertices of all trees in $\cal T$ that have $n$ vertices, and choose one vertex uniformly at random. Let $a_{n,k}$ be the probability that the chosen vertex is of rank $k$. It is then natural to ask whether the limiting probability $$a_k = \lim_{n\rightarrow \infty} a_{n,k}$$ exists.
For one tree variety, [*decreasing binary trees*]{}, it has been shown [@protected], [@janson] that these limits $a_k$ exist, and the values of $a_k$ were explicitly computed in [@bona-pittel] for $k\leq 6$. Recursive trees are discussed in [@holmgren]. However, the methods that were successful for these trees are often unsuccessful for other tree varieties if $k>1$. This is because many of the relevant differential equations cannot be solved, or even, explicitly stated, caused by the fact that many of the relevant functions lack an elementary antiderivatives. We will explain this phenomenon in Section \[nonplane12\].
This raises the intriguing question whether we can prove that $a_k$ exists for some of these tree varieties, [*even though we cannot explicitly compute its value*]{}. In this paper we will answer that question in the affirmative for two labeled tree varieties, [*non-plane 1-2 trees*]{}, and [*plane 1-2 trees*]{}. For $k=0$ and $k=1$, we are able to compute the exact values of $a_k$.
Non-plane 1-2 trees {#nonplane12}
===================
Our first example is the class of [*labeled non-plane 1-2 trees*]{}. In such trees, every non-leaf vertex has at most two children, the vertices are bijectively labeled with the elements of $[n]=\{1,2,\cdots ,n\}$ so that the label of each vertex is less than that of its parent, and the set of children of any given vertex is unordered. See Figure \[fig:fivetrees\] for the five non-plane 1-2 trees on vertex set $[4]$. In this section, when we write [*tree*]{}, we will always mean a labeled non-plane 1-2 tree on vertex set $[n]$.
![The five rooted non-plane 1-2 trees on vertex set $[4]$. []{data-label="fig:fivetrees"}](root1-2.eps){width="70mm"}
It is well-known [@FL] that the number of labeled non-plane 1-2 trees on vertex set $[n]$ is the $n$th Euler number $E_n$, and that the identity $$\label{euler} E(z)=\sum_{n\geq 0} E_n \frac{z^n}{n!} = \sec z + \tan z$$ holds, where we set $E_0=1$.
The first values of $E_n$ are as follows:
------- --- --- --- --- --- ---- ---- ----- ------ ------ ------- --
$n$ 0 1 2 3 4 5 6 7 8 9 10
$E_n$ 1 1 1 2 5 16 61 272 1385 7936 50521
------- --- --- --- --- --- ---- ---- ----- ------ ------ ------- --
It follows from (\[euler\]) that $E(z)$ has two singularities of smallest modulus, at $z=\pi/2$ and at $z=-\pi/2$. Therefore, the exponential order of growth of the Euler numbers is $2/\pi$. In order to find the growth rate of the Euler numbers more precisely, note that at both of these singularities, we can find the residue of $E(z)$ by the following well-known formula.
\[residues\] Let $H(z)=f(z)/g(z)$ be a function so that $f(z)$ and $g(z)$ are analytic functions at $z_0$, and $f(z_0)\neq 0$, while $g(z)=0$ and $g'(z)\neq 0$. Then $${\mathrm{Res } }H(z) \mid_{z_0} =\frac{f(z_0)}{g'(z_0)}.$$
We can apply Proposition \[residues\] to $E(z)$ if we note that $E(z)=\frac{1+\sin z}{\cos z}$. Then Proposition \[residues\] implies that ${\mathrm{Res } }E(z)\Bigr|_{\pi/2}=\frac{2}{-1}=-2$, and ${\mathrm{Res } }E(z) \mid_{-\pi/2}=\frac{0}{1}=0$.
Now observe that $$\begin{aligned}
\frac{R}{z-a} & = & \frac{R}{-a} \cdot \frac{1}{1-\frac{z}{a}}
\\
& = & \frac{R}{-a} \sum_{n\geq 0} \frac{z^n}{a^n}.\end{aligned}$$
Applying this observation to $E(z)$ with $a=\pi/2$ and $R=-2$, we get that the dominant term of $E(z)$ is of the form $\frac{4}{\pi}\sum_{n\geq 0} z^n (2/\pi)^n$, so $$\label{eulerprecise}
\frac{E_n}{n!} \sim \frac{4}{\pi} \cdot \left(\frac{2}{\pi} \right)^n .$$
Now we proceed to determine $a_0$ and $a_1$. For these small values of $k$, we can explicitly determine $a_k$, but we will also see why the same approach fails for larger values of $k$.
Leaves
------
Now let $A_{0,n}$ denote the total number of [*leaves*]{} in all non-plane 1-2 trees on vertex set $[n]$. Then $A_{0,0}=0$, $A_{0,1}=A_{0,2}=1$, while $A_{0,3}=3$, $A_{0,4}=9$, $A_{0,5}=35$, and $A_{0,6}=155$.
Let $A_0(z)=\sum_{n\geq 0}A_{0,n} \frac{z^n}{n!}$. Then $$\label{formofb} A_0(z)=\frac{z-1+\cos z}{1-\sin z}.$$
Let $(v,T)$ be an ordered pair in which $T$ is a non-plane 1-2 tree on vertex set $[n]$ and $v$ is a leaf of $T$. Then $A_0(z)$ is the exponential generating function counting such pairs. Let us first assume that $n>1$, and let us remove the root of $T$. On the one hand, this leaves a structure that is counted by $A_0'(z)$. On the other hand, this leaves an ordered pair consisting of a non-plane 1-2 tree with a leaf marked, and a non-plane 1-2 tree. By the Product formula of exponential generating functions (see [@WalkThroughComb Theorem 8.21]), such ordered pairs are counted by the generating function $A_0(z)E(z)$. Finally, if $n=1$, then no such ordered pair is formed, while $A_0'(z)$ has constant term 1. This leads to the linear differential equation $$A_0'(z)=E(z)A_0(z) +1,\label{deA0}$$ with the initial condition $A_0(0)=0$. Solving this equation we get formula (\[formofb\]) for $A_0(z)$.
In order to determine the growth rate of the numbers $A_{0,n}$, we will need the following lemma, which is an enhanced version of Proposition \[residues\].
\[doubleres\] Let $H(z)=\frac{f(z)}{g(z)}$ be a function so that $f$ and $g$ are analytic functions at $z_0$, and $f(z_0)\neq 0$, while $g(z_0)=g'(z_0)=0$, and $g''(z)\neq 0$. Then $$H(z)=\frac{2f(z_0)}{g''(z_0)} \cdot \frac{1}{(z-z_0)^2} + \frac{h_{-1}}{z-z_0}
+h_0+\cdots.$$
The conditions directly imply that $g$ has a double root, and hence $H$ has a pole of order two, at $z_0$. In order to find the coefficient that belongs to that pole, let $g(z)=q(z)(z-z_0)^2$. Now differentiate both sides twice with respect to $z$, to get $$g''(z)=q''(z)(z-z_0)^2 + 4q'(z)(z-z_0) + 2q(z).$$ Setting $z=z_0$, we get $$\label{fandg} g''(z_0)=2q(z_0).$$ By our definitions, in a neighborhood of $z_0$, the function $H(z)$ behaves like $$\frac{f(z)}{q(z)(z-z_0)^2},$$ and our claim follows by (\[fandg\]).
\[rank0\] The equality $$a_0=\lim_{n\rightarrow \infty} \frac{A_{0,n}}{nE_n} = 1-\frac{2}{\pi} \approx 0.3633802278$$ holds. In other words, for large $n$, the average non-plane 1-2 tree on vertex set $[n]$ has about $(n+1)\cdot ( 1-\frac{2}{\pi})$ leaves.
Note that $A_{0}(z)$ has a unique singularity of smallest modulus, at $z=\pi/2$, hence the exponential growth rate of its coefficients is $2/\pi$. Also note that at that point, the denominator of $A_0(z)$ has a double root. Therefore, Lemma \[doubleres\] applies and we get that the coefficient of the $(z-\pi/2)^{-2}$ term in the Laurent series of $A_0(z)$ is $$2\cdot \frac{(\pi)/2 -1 +\cos (\pi/2)}{\sin(\pi/2)}=\pi-2.$$ Now observe that $$\begin{aligned}
\frac{D}{(z-a)^2} & = & \frac{D}{a^2} \cdot
\frac{1}{(1-\frac{z}{a})^2} \\
& = & \frac{D}{a^2} \cdot \sum_{n\geq 0}(n+1)\frac{z^n}{a^n}.\end{aligned}$$ Applying this to the dominant term of $A_0(z)$ with $D=\pi-2$ and $a=\pi/2$, we get that $$\label{leafprecise}
\frac{A_{0,n}}{n!} \sim (n+1) (\pi -2) \cdot \left(\frac{2}{\pi}\right)^{n+2}.$$ The proof of our claim is now immediate by comparing formulas (\[leafprecise\]) and (\[eulerprecise\]).
It is worth pointing out that $\int A_0(z)= 1-(1-z)(\tan z+ \sec z)$, which implies the identity $A_{0,n}=(n+1)E_{n} -E_{n+1}$. See sequence A034428 in the Online Encyclopedia of Integer Sequences [@sloane].
Vertices of rank 1 {#sectionrank1}
------------------
Let $A_{1,n}$ be the total number of vertices in all non-plane 1-2 trees on vertex set $[n]$ that are of rank 1. Note that such vertices are neighbors of a leaf. If $n>1$, then each leaf has exactly one rank-1 vertex as a neighbor, while some rank-1 vertices have not only one, but two leaves as neighbors.
The first few members of the sequence $A_{1,n}$ are $A_{1,0}=0$, $A_{1,1}=0$, $A_{1,2}=1$, $A_{1,3}=2$, $A_{1,4}=8$, $A_{1,5}=30$, and $A_{1,6}=135$.
Let $A_1(z)=\sum_{n\geq 0}A_{1,n}\frac{z^n}{n!}$. Let $(v,T)$ be an ordered pair in which $T$ is a non-plane 1-2 tree on vertex set $[n]$ and $v$ is a vertex of $T$ that is of rank 1. If $v$ is not the root of $T$, then removing the root of $T$ decomposes $(v,T)$ into two structures, one of which is again a non-plane 1-2 tree with a vertex of rank 1 marked, and the other one of which is simply a non-plane 1-2 tree. If $n>1$, and $v$ is the root of $T$, then removing $v$, we get two structures, one of which is a leaf, and the other one is a non-plane 1-2 tree. These two structures are distinguishable unless the original tree had three vertices, and its root had two children. That tree contributed $z^3/6$ to the generating function $A_1(z)$, but that contribution was counted twice. This leads to the linear differential equation $$\label{formfora1} A_1'(z)= A_1(z)\cdot E(z) + zE(z)-\frac{z^2}{2!},$$ with the initial condition $A_1(0)=0$.
Solving this equation we get $$\label{formofv}
A_1(z)= \frac{1}{6} \cdot
\frac{12z\sin z +12\cos z -12 -3z^2\cos z -z^3}{1-\sin z}.$$ The above formula for $A_1(z)$ shows that $A_1(z)$ has a unique singularity of smallest modulus, at $z=\pi/2$. Therefore, the exponential growth rate of the coefficients of $A_1(z)$ is $2/\pi$. At $z=\pi/2$, the power series $A_1(z)$ has a pole of order two, since the denominator has a double root at that point, while the numerator is non-zero there.
Therefore, we can apply Lemma \[doubleres\] with $f(z)=12z\sin z +12\cos z -12 -3z^2\cos z -z^3$ and $g(z)=6(1-\sin z)$. Then $f(\pi/2)=6\pi-\frac{\pi^3}{8} -12$, while $g''(\pi/2)=6$. Hence Lemma \[doubleres\] shows that the dominant term of $A_1(z)$ is of the form $$\frac{2\pi - \frac{\pi^3}{24} -4}{(z-\frac{\pi}{2})^2}=
\frac{2\pi - \frac{\pi^3}{24} -4}{(\pi/2)^2} \cdot
\sum_{n\geq 0}(n+1)\frac{z^n}{(\pi/2)^n}.$$ This implies that $$\frac{A_{1,n}}{n!} \sim (n+1)\cdot \frac{2\pi -
\frac{\pi^3}{24} -4}{(\frac{\pi}{2})^{n+2}}.$$ Comparing this to (\[eulerprecise\]), we get the following theorem.
\[rank1\] The equality $$a_1=\lim_{n\rightarrow \infty} \frac{A_{1,n}}{(n+1)E_n} = 2-\frac{\pi^2}{24}-
\frac{4}{\pi}\approx 0.3155269391$$ holds.
\[nointegral\] Note that it directly follows from the argument we used to prove (\[formfora1\]) that if $R_1(z)$ is the exponential generating function for the number of non-plane 1-2 trees on vertex set $[n]$ whose root is of rank 1, then $$R_1'(z)= zE(z) -\frac{z^2}{2}=z\tan z + z\sec z -\frac{z^2}{2}.$$ Unfortunately, this closed form for $R_1'(z)$ does not lead to a closed form for $R_1(z)$, since $R_1(z)$ does not have an elementary antiderivative.
Vertices of rank $\ge2$ {#higher-nonplane}
-----------------------
The methods that we used to enumerate vertices of rank 0 and rank 1 will fail for vertices of higher rank, because we are not able to solve the linear differential equations analogous to (\[formfora1\]), since the relevant functions have no elementary antiderivatives. Remark \[nointegral\] shows how early these kind of problems start occuring; we are not even able to state the equation analogous to (\[formfora1\]) in an explicit form.
Therefore, we apply a new method to prove that the limit $a_k=\lim_{n\rightarrow \infty}a_{n,k}$ exists. We will then be able to approximate $a_k$ from above and below. Our first simple notion is the following. Each vertex of a tree is the root of a unique subtree, which we will call [*the subtree of the vertex*]{}. In other words, the subtree of a vertex $v$ consists of all descendants of $v$, including $v$ itself. The subtrees of leaves consist of one vertex only.
For a fixed positive integer $r$, let $V_{n,r}$ be the probability that a randomly selected vertex in a random non-plane 1-2 tree of size $n$ is the root of a subtree of size exactly $r$. For instance, if $n=3$, then $V_{n,1}=1/2$, $V_{n,2}=1/6$, and $V_{n,3}=1/3$.
Vertices of a given subtree size are much easier to enumerate than vertices of a given rank, because the number of ways in which a vertex can have a subtree of size $r$ is a fixed number, namely the Euler number $E_n$, once the set of labels in that subtree is selected.
\[fixedr\] For all positive integers $r$, the limit $$v_r = \lim_{n\rightarrow \infty} V_{n,r}$$ exists.
Let $V_r(z)=\sum_{n\geq 0} V_{n,r}z^n/n!$. Then by the Product formula, we have $$V_r'(z)=V_r(z)E(z) + f_r'(z),$$ where $f_r(z)$ is the generating function for the number of trees in which [*the root*]{} has a “subtree” of size $r$. That is, $f_r(z)=E_r z^r /r!$, so $f_r'(z)=E_rz^{r-1}/(r-1)!$, and the last displayed equation becomes $$\label{treesize} V_r'(z)=V_r(z)E(z) + E_r\frac{z^{r-1}}{(r-1)!} .$$ This is a first order linear differential equation with initial condition $V_r(0)=0$. Its solution is $$V_r(z)=\frac{E_r}{(r-1)!} \frac{\int z^r(1-\sin z) \ dz }{1-\sin z} +\frac{C}{1-\sin z},$$ where $C$ is to be selected so that $V_k(0)=0$ holds.
The integral in the numerator can be explicitly computed using the well-known formula $$\int z^r \sin z \ dz = \cos z \sum_{i=0}^{r/2} (-1)^{i+1}z^{r-2i} \frac{r!}{(r-2i)!}$$ $$+ \sin z\sum_{i=0}^{(r-1)/2}
(-1)^i z^{r-2i-1} \frac{r!}{(r-2i-1)!}.$$ This means that $V_r(z)$ has a unique singularity of smallest modulus (a double pole) at $z_0=\pi/2$. The rest of the argument uses Lemma \[doubleres\] at $z_0=\pi/2$ to determine $v_r$, in the same way as we did in the proofs of Theorems \[rank0\] and \[rank1\].
\[polynomial\] Note that we are able to explicitly solve the linear differential equation (\[treesize\]) because its “correction term”, that is, the summand that does not contain $V'_r(z)$ or $V_r(z)$, is a [*polynomial*]{}. The same argument used here would work for any polynomial instead of $f_r'(z)E_rz^{r-1}/(r-1)!$.
Proposition \[fixedr\] shows that the limit $v_r$ exists for every fixed positive integer $r$. As the $v_r$ are all positive real numbers, and $\sum_r v_r\leq 1$, the sum $\sum_{r=1}^\infty v_r $ is convergent. However, [*what*]{} is the value of that sum? The exact formulas we obtain for each $v_r$ from Proposition \[fixedr\] are too complicated to be useful for the computation of that sum. Note that it is not true in every tree variety that the analogously defined sum is equal to 1. A simple counterexample is the family of rooted trees in which every non-leaf vertex has exactly one child. However, for our non-plane 1-2 trees, the sum turns out to be 1, although not in a trivial way. This is the content of the following Theorem that has been conjectured by the present authors, and has recently been proved by Svante Janson [@jansonpers].
\[tight\] The equality $$\sum_{r=1}^\infty v_r=1$$ holds.
In order to prove Theorem \[tight\], we first need the following fact.
\[every\] For all $n$, the expected number of leaves in a random non-plane 1-2 tree on vertex set $[n]$ is is at least $1/4$.
Note that in Theorem \[rank0\], we proved that much more is true for large $n$. However, the statement of Proposition \[every\] is true for [*every*]{} $n$. With a little bit of additional work, it is possible to prove that in fact, the expected number of leaves is at least $n/3$ for every $n$, but the weaker claim of Proposition \[every\] suffices for our purposes.
Every such tree contains exactly one more leaves than vertices with two children. Therefore, it suffices to show that the expected number of vertices with one child is not more than $n/2$. We prove this by induction on $n$. Let $M_n$ be the expected number of vertices with one child in a random tree on vertex set $[n]$. Then $M_1=0$, $M_2=1$, and $M_3=1$, so the statement holds if $n\leq 3$. Now let us assume that $n>3$. Let $p_n$ be the probability that the [*root*]{} of a non-plane 1-2 tree on $[n]$ has exactly one child. Then by conditional expectations, we have $$M_n \leq (1-p_n) \cdot \frac{n-1}{2} +p_n \left(\frac{n-1}{2}+1\right)=\frac{n-1}{2}+p_n\leq \frac{n}{2}.$$ The last inequality follows, since $p_n=E_{n-1}/E_{n}$, the ratio of two consecutive Euler numbers, and the Euler numbers are known to be log-convex [@Liu]. So the sequence of the numbers $p_n$ is decreasing. As $p_3=1/2$, it follows that $p_n\leq 1/2$ if $n\geq 3$.
Let $Z_n$ be a random variable defined on the set of all vertices of all non-plane 1-2 trees on vertex set $[n]$, so that $Z_n(v)$ is the size of the subtree rooted at $v$.
\[estimate\] For all $n$, the inequality $E(\sqrt{Z_n}) \leq 100$ holds.
We will use strong induction to prove the stronger inequality $$E(\sqrt{Z_n}) \leq 100 -\frac{90}{\sqrt{n}}.$$ This inequality clearly holds for $n=1$. Now let us assume that it holds for all positive integers $r<n$ and prove it for $n$.
Let $T$ be a tree of size $n$, selected uniformly at random. Let $v$ be a vertex of $T$, selected in the same way.
Then by Proposition \[every\], there is an at least $1/4$ chance of $v$ being a leaf. There is an $1/n$ chance of $v$ being the root. There is a less than $3/4$ chance of $v$ being another vertex, in which case the induction hypothesis applies to the subtree of $v$, with some $r<n$ playing the role of $n$. Therefore, $$E(\sqrt{Z_n}) \leq \frac{3}{4} \cdot \left (100-\frac{90}{\sqrt{n}} \right ) + \frac{1}{4}+ \frac{1}{n}\cdot\sqrt{ n}$$ $$=75.25-\frac{66.5}{\sqrt{n}}$$ $$\leq 100 -\frac{90}{\sqrt{n}},$$ where the last estimate holds as $n\geq 1$.
Now we are ready to prove Theorem \[tight\].
(of Theorem \[tight\]) By Lemma \[estimate\] and by Markov’s inequality, we know that for all positive constants $C$, and for all $n$, we have $$Pr(\sqrt{Z_n}> 100C ) \leq 1/C,$$ so $$\label{markovineq} Pr(Z_n>10000C^2 ) \leq 1/C.$$
Let us now assume that $\sum_{k=1}^\infty v_k = \alpha < 1$. That means that for all $N$, the inequality $\sum_{k=1}^N v_k < \alpha $ holds. In other words, if $n$ is large enough, then in an average non-plane 1-2 tree on $[n]$, there are at least $(1-\alpha)n/2$ vertices whose subtree is of size more than $N$. That is, $$\label{largeenough} Pr(Z_n>N) \geq \frac{1-\alpha}{2}.$$ Now select $C$ to be a positive integer so that $1/C < (1-\alpha)/2$, then select $N=10000C^2$. Then inequality (\[markovineq\]) forces $$Pr(Z_n>N) \leq 1/C <(1-\alpha)/2$$ for all $n$, while inequality (\[largeenough\]) forces $Pr(Z_n>N)\geq (1-\alpha)/2$ for $n$ sufficiently large, which is clearly a contradiction.
The following is an obvious corollary of Theorem \[tight\] that we will need soon.
\[ucor\] Let $U_{n,r}=1-\sum_{i=1}^r V_{n,i}$ be the probability that a random vertex of a random non-plane 1-2 tree has a subtree of size larger than $r$. Then clearly, $$u_r:=\lim_{n\rightarrow \infty} U_{n,r}= 1-\sum_{i=1}^r v_i.$$ Furthermore, and this is where Theorem \[tight\] is needed, $$\label{ulimit} \lim_{r\rightarrow \infty} u_r = 1 - \lim_{r\rightarrow \infty} \sum_{i=1}^r v_i=0.$$
We now return to our main goal, that is, to proving that the limit $a_k=\lim_{n\rightarrow \infty}a_{n,k}$ exists. For the rest of this section, we fix the rank $k$ of vertices we are studying, and, to alleviate notation, we do [*not add*]{} the index $k$ to all parameters related to these vertices.
Our main idea is the following. The set $\mathcal R_k$ of all vertices of all trees of size $n$ [*contains*]{} the set $\cup_{i=1}^r \mathcal W_{n,i}$, where $\mathcal W_{n,i}$ is the set of all vertices of all trees of size $n$ that are of rank $k$ and have a subtree of size $i$. On the other hand, $\mathcal R_k$ [*is contained*]{} in the set $\left(\cup_{i=1}^r \mathcal W_{n,i} \right) \cup \left( \cup_{i>r} \mathcal V_{n,i} \right)$, where $ \mathcal V_{n,i} $ is the set of all vertices in all trees of size $n$ whose subtree is of size $i$ (but are of any rank).
Let $W_{n,i}$ be the probability that a random vertex of a random tree of size $n$ is of rank $k$ and is the root of a subtree of size $i$. Let $$w_i=\lim_{n\rightarrow \infty} W_{n,i} .$$ The limits $w_i$ exist, because the exponential generating functions of the numbers $W_{n,i}$ satisfy a linear differential equation like (\[treesize\]), and, as explained in Remark \[polynomial\], we can explicitly solve those differential equations, since their “correction term” is a polynomial. Indeed, there are only a finite number of ways that a subtree of a vertex can be of rank $k$ and have a subtree of size $i$, once the set of labels going into that subtree is selected.
As the $w_i$ are positive real numbers, and for all $r$, the inequality $\sum_{i=1}^r w_i \leq 1$ holds, the sum $$w=\sum_{i=1}^{\infty} w_i$$ exists.
Now we are ready to state and prove our main theorem.
\[maintheorem\] For all positive integers $k$, the limit $$a_k:= \lim_{n\rightarrow \infty} a_{n,k}$$ exists. Furthermore, $$a_k= w.$$
First notice that for all $n$ and $r$, the inequality $$\sum_{i=1}^r W_{n,i}\leq a_{n,k}$$ holds, since the left-hand side is the probability of a random vertex having a more restrictive property (rank $k$, subtree size at most $r$) than the property represented on the right-hand side (rank $k$). Therefore, $$\label{lowerineq} \sum_{i=1}^r w_r\leq \liminf_{n} a_{n,k} ,$$ and so $$\label{liminf} w \leq \liminf_n a_{n,k} .$$
Now notice that for all $n$ and $r$, the inequality $$a_{n,k} \leq \sum_{i=1}^r W_{n,i}+ \sum_{i>r} V_{n,i} = \left( \sum_{i=1}^r W_{n,i} \right )+ U_{n,r}$$ holds. Indeed, the right-hand side is the probability of a random vertex being of rank $k$ and having a subtree of size at most $r$, or simply having a subtree of size more than $r$ (and any rank). A particular way of this occuring is when the random vertex is of rank $k$, which is the event whose probability is represented on the left-hand side.
This implies that for all $r$, we have $$\label{upperineq} \limsup_n a_{n,k} \leq \sum_{i=1}^r w_i + \left(1-\sum_{i=1}^r v_i\right).$$ As $r$ goes to infinity, the first sum on the right-hand side goes to $w$, while the second sum goes to 0, as we saw in Corollary \[ucor\]. This proves that $$\label{limsup} \limsup_n a_{n,k} \leq w.$$
Comparing inequalities (\[liminf\]) and (\[limsup\]), we see that $$\limsup_n a_{n,k} \leq w \leq \liminf_n a_{n,k} ,$$ completing the proof of the theorem.
For numerical approximations, one can use the following corollary, which is an immediate consequence of Theorem \[maintheorem\] that we have just proved, and inequalities (\[lowerineq\]) and (\[upperineq\]) that we have used in the proof of that theorem.
\[numerical\] For all $r$, the chain of inequalities $$\sum_{i=1}^r w_i \leq a_k \leq \sum_{i=1}^r w_i + \left ( 1 - \sum_{i=1}^r v_i \right )$$ holds.
Plane 1-2 trees
===============
The next tree class we study is the class of [*plane*]{} 1-2 trees on vertex set $[n]$. These are similar to the trees of the previous section, except that now the order of the children of each vertex matters. See Figure \[fig:threetrees\] for an illustration. We denote the number of such trees on $[n]$ by $b_n$. Our goal is to show that Theorem \[maintheorem\] can be proved for these trees as well. Most steps are similar to what we saw in Section \[nonplane12\], but there will be one step that requires a separate argument.
![The three rooted plane 1-2 trees on vertex set $[3]$. []{data-label="fig:threetrees"}](planetrees.eps){width="70mm"}
The first few values of the sequence $b_n$, starting with $b_1$, are 1, 1, 3, 9, and 39. This is sequence A080635 in OEIS [@sloane]. Setting $b_0=1$, the exponential generating function $$B(z)=\sum_{n=0}^\infty b_n\frac{z^n}{n!}$$ satisfies the differential equation $$B'(z)=1-B(z)+B^2(z).$$ Solving this equation yields $$B(z)=\frac{1}{2} + \frac{\sqrt{3}}{2} \tan \left(\frac{\sqrt{3}}{2}z +\frac{\pi }{6}\right).$$ The power series form of $B(z)$ leads to the asymptotic formula $$\frac{b_n}{n!}\sim\frac{3^{3/2}}{2\pi}\left(\frac{3^{3/2}}{2\pi}\right)^n.\label{bnasymp}$$
Just as it was the case for non-plane trees, we can determine the values of $a_0$ and $a_1$ for plane 1-2 trees as well.
Leaves {#leaves-plane}
------
We can find the ratio of leaves among all vertices in a way that is analogous to that for non-plane 1-2 trees.
The exponential generating function of the numbers of leafs in decreasing plane 1-2 trees is $$B_0(z)=\sum_{n=0}^\infty b_{0,n}\frac{z^n}{n!}=\frac{6 z+\sqrt{3} \sin \left(\sqrt{3} z\right)+3 \cos \left(\sqrt{3} z\right)-3}{-3 \sqrt{3} \sin \left(\sqrt{3} z\right)+3 \cos \left(\sqrt{3} z\right)+6},$$ which satisfies the differential equation $$B_0'(z)=2B_0(z)(B(z)-1)+B_0(z)+1.$$
Just like in our proofs for analogous results in Section \[nonplane12\], we count ordered pairs $(v,T)$, where $v$ is a leaf of the tree $T$. Let us remove the root of $T$. Then there are two cases, namely, either the removed root was $v$, or it was not.
If $v$ is not the root, and we got two trees, one with the marked vertex, then the Product formula yields the generating function $2B_0(z)(B(z)-1)$, as the order of the components matters.
If $v$ is not the root, and the root has only one child, then removing the root, we got only one tree, with a marked vertex, which contributes the generating function $B_0(z)$.
Finally, if $v$ is the root, the only possible tree is the one-point graph. The removal of that root leads to the empty graph, represented by $1=\left(\frac{z^1}{1!}\right)'$ in the differential equation.
By the generating function we can determine the first values of $b_{0,n}$:
----------- --- --- --- --- ---- ---- ----- ------ ------- -------- ---------
$n$ 0 1 2 3 4 5 6 7 8 9 10
$b_{0,n}$ 0 1 1 5 17 93 513 3477 25569 212733 1929393
----------- --- --- --- --- ---- ---- ----- ------ ------- -------- ---------
A simple application of Lemma \[doubleres\] yields that $$\frac{b_{0,n}}{(n+1)n!}\sim\frac{\frac{4\pi}{9\sqrt3}-\frac13}{\left(\frac{2\pi}{3\sqrt3}\right)^{n+2}},$$ and, by recalling , $$\lim_{n\to\infty}\frac{b_{0,n}}{(n+1)b_n}=\frac23-\frac{\sqrt3}{2\pi}\approx0.391.$$
Neighbors of leaves {#neighbors-plane}
-------------------
This case is similar to that of leaves, with some subtle differences. Let $b_{1,n}$ denote the number of all vertices of rank 1 in all trees on vertex set $[n]$, and let $B_{1}(z)$ be the exponential generating function of these numbers. Let us count ordered pairs $(v,T)$, where $v$ is a vertex of rank 1 in a tree $T$. Let us remove the root of $T$. The case when $v$ is not the root, is the same as in Section \[leaves-plane\], contributing the term $2B_1(z)(B(z)-1)+B_1(z)$. When $v$ is the root, then removing it we obtain a leaf and a tree. If this tree is not empty, we must distinguish whether it was on the left or right hand side, so we must add a term $2z(B(z)-1)$. If, in turn, the subtree is empty, we must add the term representing the path of length one, that is $\left(\frac{z^2}{2!}\right)'=z$. Finally, we must realize that the terms $2B_1(z)(B(z)-1)$ and $2z(A(z)-1)$ both contain the two trees on three points where the root has two children. Therefore we must subtract $2\left(\frac{z^3}{3!}\right)'=z^2$. This proves the following.
The generating function $B_1(z)$ satisfies the differential equation $$B_1'(z)=2B_1(z)(B(z)-1)+B_1(z)+2z(B(z)-1)+z-z^2\quad(B_1(0)=0.$$ Therefore, $$B_1(z)=\frac{6 z^3+\sqrt{3} \left(3 z^2-15 z-5\right) \sin \left(\sqrt{3} z\right)+3 \left(3 z^2+5 z-5\right) \cos \left(\sqrt{3} z\right)+15}{9 \left(\sqrt{3} \sin \left(\sqrt{3} z\right)-\cos \left(\sqrt{3} z\right)-2\right)}.$$
The first values of $b_{1,n}$ are as follows.
----------- --- --- --- --- ---- ---- ----- ------ ------- -------- ---------
$n$ 0 1 2 3 4 5 6 7 8 9 10
$b_{1,n}$ 0 0 1 3 15 75 435 2883 21447 177435 1613835
----------- --- --- --- --- ---- ---- ----- ------ ------- -------- ---------
The asymptotic expression for the total number of rank one vertices can be found easily: $$\frac{b_{1,n}}{(n+1)n!}\sim\frac{540 \sqrt{3} \pi -16 \sqrt{3} \pi ^3-1215}{2187\left(\frac{2\pi}{3\sqrt3}\right)^{n+2}},$$ and $$\lim_{n\to\infty}\frac{b_{1,n}}{(n+1)b_n}=\frac{10}{9}-\frac{5}{2\sqrt3\pi}-\frac{8\pi^2}{243}\approx 0.3267.$$
Vertices of higher rank
-----------------------
If we try to apply the method of Sections \[leaves-plane\] and \[neighbors-plane\] for vertices of rank $k$, for $k\geq 2$, we fail, because yet again, the relevant generating functions will not have elementary antiderivatives. However, the method that we used in Section \[higher-nonplane\] to prove that the limits $a_k$ exist will work again, as we will show.
Let us define the limits $v_r$ and $w_r$ exactly as we did in Section \[higher-nonplane\], except that now the trees are plane. If we try to follow the argument of the non-plane case, we see that the first step towards proving the existence of $v_r$ and $w_r$ is to show that we can explicitly solve the linear differential equation $$f'(z)=2f(z)(B(z)-1) +f(z) +P(z),$$ where $P(z)$ is a [*polynomial*]{} function. Indeed, we get differential equations of the above kind when we attempt to find the probabilities $V(n,r)$ or $W(n,r)$.
Bringing the above differential equation to standard form, we get $$\label{standard} f'(z)+(1-2B(z))f(z) = P(z).$$
In order to solve (\[standard\]), we multiply both sides by the integrating factor $$Q(z)=\exp \left (\int (1-2B(z)) \ dz \right)=\frac{1}{2}+\frac{\cos \left (\sqrt{3}z \right )}{4} -\frac{\sqrt{3}\sin \left (\sqrt{3}z \right )}{4}.$$
Multiplying both sides of (\[standard\]) by $Q(z)$, we get the equation $$(f(z)Q(z))' = Q(z)P(z),$$ which we can explicitly solve as long as we can integrate $Q(z)P(z)$. In the present case, we can certainly do that, since $P(z)$ is a polynomial function of $z$, hence it is a polynomial function of $\sqrt{3}z$ as well, so a substitution $t=\sqrt{3}z$ will result in a function consiting of the sums of summands in the form $t^m\sin t$ and $t^n \cos t$. In the end, we obtain $$\label{generalf} f(z)=\frac{\int Q(z)P(z)}{Q(z)}=\frac{K(z)}{Q(z)},$$ a meromorphic function. The asymptotic behavior of meromorphic functions is well understood. See Theorem IV.10 of [@FL] for the most important results. In our case, the numerator $K(z)$ of $f(z)$ in (\[generalf\]) is an entire function, while the denominator has a zero at $z=2\sqrt{3}\pi /9$ that has multiplicity two. Therefore, the asymptotics of the coefficients of $f(z)$ can be computed using Lemma \[doubleres\]. Therefore, the existence of $v_r$ and $w_r$ can also be proved in the same way as it was in Section \[nonplane12\] for non-plane 1-2 trees.
The next step is to prove Theorem \[tight\], that is, the equality $\sum_{r=1}^{\infty}v_rs=1$ for plane 1-2 trees. There is one step in that proof that needs an argument that is different from its non-plane analogue, which is Proposition \[every\]. Therefore, we announce and prove it separately as follows.
\[pevery\] For all $n$, the expected number of leaves in a random plane 1-2 tree on vertex set $[n]$ is at least $1/4$.
We prove that the expected number of leaves in plane 1-2 trees on vertex set $[n]$ is at least as large as the expected number of leaves in non-plane 1-2 trees on vertex set $[n]$. As the latter has been proved to be at least $n/4$ in Proposition \[every\], this will be sufficient.
Clearly, in both tree varieties, the number of vertices with two children is one less than the number of leaves. Therefore, it suffices to prove that average the number of vertices with one child is [*at most as large*]{} in plane 1-2 trees on $[n]$ as it is on non-plane 1-2 trees on $[n]$. We use a well-known inequality, known as the [*Chebyshev sum inequality*]{} or (a special case of) the [*rearrangement inequality*]{}.
\[cheb\] Let $r_1\leq r_2\leq \cdots \leq r_u$ and $t_1\geq t_2\geq \cdots \geq t_u$ be nonnegative real numbers. Then the inequality $$\frac{r_1+r_2+\cdots +r_u}{u} \geq \frac{r_1t_1+r_2t_2+\cdots +r_ut_u}{t_1+t_2+\cdots +t_u}$$ holds.
See [@hardy] or [@holstermann] for a proof.
Let us return to the proof of Proposition \[pevery\]. Consider all $E_n$ non-plane trees on vertex set $[n]$. Let $s_1,s_2,\cdots ,s_{E_n}$ denote the number of vertices with exactly one child in each of these trees, and let us order the set of these $E_n$ trees so that the sequence of the $s_i$ is non-decreasing, that is, $s_1\leq s_2\leq \cdots \leq s_{E_n}$. Then the average number of vertices with one child in all non-plane 1-2 trees on $[n]$ is $$\label{firstav} M_n = \frac{\sum_{i=1}^{E_n} s_i}{E_n} .$$ On the other hand, if such a tree $T$ has $s_i$ vertices with one child, then it has $(n-1-s_i)/2$ vertices with two children, (and, though we will need this only later, $T$ has $(n+1-s_i)/2$ leaves). Therefore, there are exactly $2^{(n-1-s_i)/2}$ [*plane*]{} 1-2 trees on vertex set $[n]$ that are identical to $T$ as non-plane trees, and each of those trees has $s_i$ vertices with one child. This proves that the average number of vertices with exactly one child in all [*plane*]{} 1-2 trees on vertex set $[n]$ is $$\label{secondav} m_n =
\frac{\sum_{i=1}^{E_n} s_i 2^{(n-1-s_i)/2}}{\sum_{i=1}^{E_n} 2^{(n-1-s_i)/2}} .$$
Finally, note that the sequences $s_1\leq s_2\leq \cdot \leq s_{E_n}$ and $2^{(n-1-s_1)/2}\geq 2^{(n-1-s_2)/2}\geq
\cdots \geq 2^{(n-1-s_{E_n})/2}$ satisfy the requirements of Proposition \[cheb\], so $M_n \geq m_n$ holds. So the average non-plane 1-2 tree has at least as many vertices with one child as the average plane 1-2 tree of the same size. Therefore, the average [*plane*]{} 1-2 tree has at least as many leaves as the average non-plane 1-2 tree of the same size. The proof of our claim is now immediate, since we saw in Proposition \[every\] that the average non-plane tree on $[n]$ has at least $n/4$ leaves.
All remaining steps of Theorem \[maintheorem\] can be carried out without any extra effort, showing that the limits $a_k$ exist for all $k$, for the variety of plane 1-2 trees as well.
Approximations
==============
Corollary \[numerical\] makes numerical approximations of $a_k$ possible. As the upper bound provided by (\[upperineq\]) was obtained by a rather crude estimate, it is reasonable to assume that the lower bound in that corollary is a better estimate for $w$ than the upper bound. It follows from our methods that both the upper and the lower bounds will be of the form $\pi^{-2}F(\pi)$, where $F$ is a polynomial function with rational coefficients. For instance, selecting $k=2$ and $r=12$ leads to a lower bound of $ 0.188285\leq a_2$. On the other hand, less rigorous, but more extensive computations carried out by Jay Pantone [@pantone] suggest the approximate values $a_2 \approx 0.20278137$, $a_3 \approx 0.0893474$, and $ a_4 \approx 0.0243854$.
[**Acknowledgement**]{}
The authors are indebted to Svante Janson for the proof of Theorem \[tight\]. They are also grateful to Jay Pantone who helped them obtaining the numerical results of the last section.
[99]{}
\(2011) A Walk Through Combinatorics, third edition, World Scientific.
\(2014) $k$-protected vertices in binary search trees, *Adv. in Appl. Math*. [**53,**]{} 1–11.
\(2017) On a random search tree: asymptotic enumeration of vertices by distance from leaves, *Adv. in Appl. Probab*. [**49**]{} (2017), no. 3, 850–876.
(2014). Protected nodes and fringe subtrees in some random trees, *Electron. Commun. Probab*. [**19,**]{} no. 6, 10 pages.
\(2009) Analytic Combinatorics, Cambridge University Press.
(1988). Inequalities. Cambridge Mathematical Library. [*C*ambridge University Press]{}.
(2017), A Generalization of the Rearrangement Inequality. *Mathematical Reflections* [**5**]{}, 4 pages.
(2017). Fringe trees, Crump-Mode-Jagers branching processes and $m$-ary search trees. *Probability Surveys* [**14**]{} 53–154.
Personal communication, February, 2018.
On the log-convexity of combinatorial sequences. [*Adv. in Appl. Math.*]{} [**39**]{} (2007), no. 4, 453–476.
Personal communication, January, 2018.
, online database, [www.oeis.org](www.oeis.org).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
The purpose of this article is to give a rather thorough understanding of the compact support property for measure-valued diffusion processes corresponding to semi-linear equations of the form $$\begin{aligned}& u_t=Lu+\beta u-\alpha u^p \ \ \text{in} \ R^d\times (0,\infty),
\ p\in(1,2];\\
&u(x,0)=f(x) \ \ \text{in}\ R^d;\\
&u(x,t)\ge0 \ \ \text{in} \ R^d\times[0,\infty).
\end{aligned}$$ In particular, we shall investigate how the interplay between the underlying motion (the diffusion process corresponding to $L$) and the branching affects the compact support property. In [@EP99], the compact support property was shown to be equivalent to a certain analytic criterion concerning uniqueness of the Cauchy problem for the semilinear parabolic equation related to the measured valued diffusion. In a subsequent paper [@EP03], this analytic property was investigated purely from the point of view of partial differential equations. Some of the results obtained in this latter paper yield interesting results concerning the compact support property. In this paper, the results from [@EP03] that are relevant to the compact support property are presented, sometimes with extensions. These results are interwoven with new results and some informal heuristics. Taken together, they yield a fairly comprehensive picture of the compact support property. *Inter alia, we show that the concept of a measure-valued diffusion *hitting a point can be investigated via the compact support property, and suggest an alternate proof of a result concerning the hitting of points by super-Brownian motion.**
address: |
Department of Mathematics\
Technion—Israel Institute of Technology\
Haifa, 32000\
Israel
author:
- 'Ross G. Pinsky'
title: 'The compact support property for measure-valued diffusions'
---
Introduction and Statement of Results {#S:intro}
=====================================
The purpose of this article is to give a rather thorough understanding of the compact support property for measure-valued diffusion processes. In particular, we shall investigate how the interplay between the underlying motion and the branching affects the compact support property. In [@EP99], the compact support property was shown to be equivalent to a certain analytic criterion concerning uniqueness of the Cauchy problem for the semilinear parabolic equation related to the measured valued diffusion. In a subsequent paper [@EP03], this analytic property was investigated purely from the point of view of partial differential equations. Some of the results obtained in this latter paper yield interesting results concerning the compact support property. In this paper, the results from [@EP03] that are relevant to the compact support property are presented, sometimes with extensions. These results are interwoven with new results and some informal heuristics. Taken together, they yield a fairly comprehensive picture of the compact support property. *Inter alia, we show that the concept of a measure-valued diffusion *hitting a point can be investigated via the compact support property and suggest an alternate proof of a result concerning the hitting of points by super-Brownian motion.**
We begin by defining the measure-valued diffusions under study. Let $$L=\frac12\sum_{i,j=1}^d a_{i,j}\frac{\partial^2}{\partial x_i\partial
x_j}+\sum_{i=1}^d b_i\frac\partial{\partial x_i},$$ where $a_{i,j},b_i\in C^{\alpha}(R^d)$ and $\{a_{i,j}\}$ is strictly elliptic; that is, $\sum_{i,j=1}^da_{i,j}(x)\nu_i\nu_j>0$, for all $x\in R^d$ and $\nu\in R^d-\{0\}$. Let $Y(t)$ denote the diffusion process corresponding to the generalized martingale problem for $L$ on $R^d$ [@P95], and denote corresponding probabilities by $\mathcal P_\cdot$. Denote the lifetime of $Y(t)$ by $\tau_\infty$. One has $\tau_\infty=\lim_{n\to\infty}\tau_n$, where $\tau_n=\inf\{t\ge0:|Y(t)|\ge n\}$. Recall that $Y(t)$ is called *non-explosive (or conservative) if $\mathcal P_x(\tau_\infty<\infty)=0$, for some, or equivalently all, $x\in R^d$; otherwise $Y(t)$ is called *explosive.The process $Y(t)$ serves as the underlying motion of the measure-valued diffusion.**
The branching mechanism is of the form $\Phi(x,z)= \beta(x)z-\alpha(x)z^p$, where $p\in(1,2]$, $\beta$ is bounded from above, $\alpha>0$, and $\alpha, \beta\in C^\kappa(R^d)$, for some $\kappa\in (0,1]$. A finite measure-valued diffusion $X(t)=X(t,\cdot)$ is the Markov process defined uniquely via the following log-Laplace equation: $$\label{Def:MVD}
E(\exp(-<f, X(t)>)|X(0)=\mu)=\exp(-<u_f(\cdot,t),\mu>),$$ for $f\in C^+_c(R^d)$, the space of compactly supported, nonnegative, continuous functions on $R^d$, and for finite initial measures $\mu$, where $u_f$ is the minimal positive solution to the evolution equation
$$\label{evoequ}
\begin{aligned}& u_t=Lu+\beta u-\alpha u^p \ \ \text{in} \ R^d\times (0,\infty);\\
&u(x,0)=f(x) \ \ \text{in}\ R^d;\\
&u(x,t)\ge0 \ \ \text{in} \ R^d\times[0,\infty).
\end{aligned}$$
The measure for the process started from $\mu$ will be denoted by $P_\mu$, and its expectation operator will be denoted by $E_\mu$.
We recall the compact support property.
Let $\mu\in \mathcal M_F(D)$ be compactly supported. The measure-valued process corresponding to $P_\mu$ possesses the compact support property if $$P_\mu(\bigcup_{0\le s\le t}\{\text{supp}\ \ X(s)\ \text{ is bounded}\})=1, \
\text{for all}\ t\ge0.$$
The parameter $\beta$ may be thought of as the mass creation parameter (see the discussion of the particle process approximation to the measure-valued diffusion at the end of this section). *Without further mention, it will always be assumed in this paper that $\beta$ is bounded from above. It is possible to extend the construction of the measure-valued process to certain $\beta$ which are unbounded from above; namely, to those $\beta$ for which the generalized principal eigenvalue of the operator $L+\beta$ is finite. However, the resulting process has paths which are mutually absolutely continuous with respect to another measure-valued diffusion whose mass creation parameter $\beta$ *is bounded from above [@EP99]. Thus, the compact support property will hold for the former process if and only if it holds for the latter one.**
There are four objects, corresponding to four different underlying probabilistic effects, which can influence the compact support property:
1. $L$, the operator corresponding to the underlying motion;
2. $\beta$, the mass creation parameter of the branching mechansim;
3. $\alpha$, the nonlinear component of the branching mechanism, which can be thought of as the variance parameter if $p=2$;
4. $p$, the power of the nonlinearity, which is the scaling power and is connected to the fractional moments of the offspring distribution in the particle process approximation to the measure-valued diffusion.
(For (2), (3) and (4) above, see the discussion of the particle process approximation to the measure-valued diffusion at the end of this section.)
We shall see that both $L$ and $\alpha$ play a large role in determining whether or not the compact support property holds; $\beta$ and $p$ play only a minor role.
In [@EP99], the compact support property was shown to be equivalent to a uniqueness property for solutions to .
The compact support property holds for one, or equivalently all, nonzero, compactly supported initial measures $\mu$ if and only if there are no nontrivial solutions to with initial data $f\equiv0$.
We emphasize that the uniqueness property in Theorem EP1 concerns *all classical solutions to , with no growth restrictions. If one restricts to mild solutions—solutions which solve an integral equation involving the linear semigroup corresponding to the operator $L$—then, for example, uniqueness holds in this class if $\alpha, \beta$ and the coefficients of $L$ are bounded [@P83]; yet, these conditions certainly do not guarantee uniqueness in the class of all positive, classical solutions (see Theorem \[smallalpha\] below).*
In fact, the proof of Theorem EP1 shows that there exists a maximal solution $u_{max}$ to with initial data $f=0$, and $$P_\mu(\bigcup_{0\le s\le t}\{\text{supp}\ \ X(s)\ \text{ is bounded}\})=
\exp(-\int_{R^d}u_{max}(x,t)\mu(dx)),$$ for compactly supported $\mu$. This shows that when the compact support property fails, the onset of the failure is gradual; that is, as a function of $t$,$P_\mu(\bigcup_{0\le s\le t}\{\text{supp}\ X(s)\ \text{ is bounded}\})$ is continuous and equal to 1 at time $t=0$. This behavior is in contrast to the behavior of the measure-valued process corresponding to the semilinear operator $u_t=\Delta u-u\log u$, investigated recently in [@FS04]. This process is obtained as a weak limit as $p\to1$ of the processes corresponding to the semilinear operators $u_t=\Delta u+\frac1{p-1}u-\frac1{p-1}u^p$. Unlike the measure-valued diffusions defined above, this process is immortal; that is, $P_\mu(X(t)=0)=0$, for all $t\ge0$. Furthermore, it is shown that $P_\mu(\bigcup_{0\le s\le t}\{\text{supp}\ \ X(s)\ \text{ is bounded}\})=0$, for all $t>0$. Thus, the onset of the failure of the compact support property is instantaneous. Theorem EP1 is not valid for this process. Indeed, the proof of Theorem EP1 requires the fact that a maximal solution exists for with initial condition $f=0$. The existence of such a maximal solution is essentially equivalent to the existence of a universal, a priori upper bound on all solutions to ; that is, the existence of a finite function $M(x,t)$ on $R^d\times(0,\infty)$ such that $u(x,t)\le M(x,t)$, for all $(x,t)\in R^d\times(0,\infty)$, and all solutions $u$ to . Such a universal a priori upper bound does not exist for the equation $u_t=\Delta u-u\log u$. In [@P04], a more or less necessary and sufficient condition on the nonlinear term (independent of the operator $L$) is given for the existence of such a bound.
Theorem EP1 suggests a parallel between the compact support property for measure-valued diffusions and the non-explosion property for diffusion processes. Indeed, the non-explosion property for the diffusion process $Y(t)$ corresponding to the operator $L$ is equivalent to the nonexistence of nontrivial, *bounded positive solutions to the linear Cauchy problem with 0 initial data: $$\label{linequ}
\begin{aligned}&u_t=Lu\ \ \text{in}\ R^d\times(0,\infty);\\
&u(x,0)=0\ \ \text{in}\ R^d;\\
&u\ge0 \ \ \text{in} \ R^d\times(0,\infty).
\end{aligned}$$ (For one direction of this result, note that $u(x,t)\equiv\mathcal P_x(\tau_\infty\le t)$ serves as a nontrivial solution to in the explosive case.) It is natural for bounded, positive solutions to be the relevant class of solutions in the linear case and for positive solutions to be the relevant class of solutions in the semilinear case. Indeed, by Ito’s formula, the probabilities for certain events related to $Y(t)$ are obtained as bounded, positive solutions to the linear equation, and by the log-Laplace equation, the negative of the logarithm of the probability of certain events related to $X(t)$ can be obtained as positive solutions to the semilinear equation.*
The class of operators $L$ satisfying the following assumption will play an important role.
\[quadlin\] For some $C>0$,
1. $\sum_{i,j=1}^na_{ij}(x)\nu_i\nu_j
\le C|\nu|^2(1+|x|^2), \ x,\nu\in R^d$;
2. $|b(x)|\le C(1+|x|),\ x\in R^d$.
The next theorem culls some results from [@EP03] and applies them to the probabilistic setting at hand.
Let $p\in(1,2]$ and let the coefficients of $L$ satisfy Assumption \[quadlin\].
1. There is no nontrivial solution to ; thus, the diffusion process $Y(t)$ does not explode.
2. If $$\inf_{x\in R^d}\alpha(x)>0,$$ then there is no nontrivial solution to with initial data $f=0$; thus, the compact support property holds for $X(t)$.
\[Proof of part (1)\] For the proof that there are no nontrivial solutions to if Assumption \[quadlin\] holds, see [@EP03 Proposition 5 and Remark 1 following it]. Non-explosiveness of $Y(t)$ then follows from the parenthetical sentence following above.
**Proof of part (2). For the proof that there are no nontrivial solutions to , see [@EP03 Theorem 2]. The fact that the compact support property holds then follows from Theorem EP1 above.**
The conditions in Assumption 1 are classical conditions which arise frequently in the theory of diffusion processes. Theorem EP2 shows that if the coefficients of $L$ obey this condition and if the branching coefficient $\alpha$ is bounded away from zero, then everything is well behaved—neither can the underlying diffusion process explode nor can the measure-valued diffusion fail to possess the compact support property.
The following result shows that the compact support property can fail if $\inf_{x\in R^d}\alpha(x)=0$. It also demonstrates that the effect of $\alpha$ on the compact support property cannot be studied in isolation, but in fact depends on the underlying diffusion.
\[smallalpha\] Let $p\in(1,2]$ and let $$\begin{aligned}
&L=A(x)\Delta,\
\text{where} \ C_0^{-1}(1+|x|)^m\le A(x)\le C_0(1+|x|)^m,\ m\in[0,2],\\
& \text{for some}\ C_0>0.
\end{aligned}$$
1. If $$\alpha(x)\ge C_1\exp(-C_2|x|^{2-m}),$$ for some $C_1,C_2>0$, then the compact support property holds for $X(t)$.
2. If $$\alpha(x)\le C\exp(-|x|^{2-m+\epsilon}) \ \text{and}\
\beta(x)\ge-C(1+|x|)^{2-m+2\delta},$$ for some $C,\epsilon>0$ and some $\delta<\epsilon$, then the compact support property does not hold for $X(t)$.
By Theorem EP1, to prove Theorem 1, it is necessary and sufficient to show that if $\alpha$ is as in part (1) of the theorem, then there is no nontrivial solution to with initial data $f=0$, while if $\alpha$ is as in part (2) of the theorem, then there is such a nontrivial solution. In the case that $L=\frac12\Delta$, and for part (2), $\beta\ge0$, this result was obtained in [@EP03 Theorem 7]. An alternative, more purely probabilistic proof which does not rely on Theorem EP1 can be found in [@R04] for the case $L=\frac12\Delta$ and $\beta=0$.
A heuristic, qualitative understanding of Theorem \[smallalpha\] is given at the end of this section.
As a complement to Theorem \[smallalpha\], we note the following result [@E00; @EP03].
Let $p\in(1,2]$.
1. Let $d\ge2$ and let $$L=A(x)\Delta,\
\text{where} \ A(x)\ge C(1+|x|)^m,\
\text{for some}\ C>0\
\text{and}\ m>2.$$ Assume that $$\sup_{x\in R^d}\alpha(x)<\infty\ \text{and}\
\beta\ge0.$$ Then the compact support property does not hold for $X(t)$.
2. Let $d=1$ and let $$L=A(x)\frac{d^2}{dx^2},\
\text{where} \ A(x)\ge C(1+|x|)^m,\
\text{for some}\ C>0\
\text{and}\ m>1+p.$$ Assume that $$\sup_{x\in R^d}\alpha(x)<\infty\ \text{and}\
\beta\ge0.$$ Then the compact support property does not hold for $X(t)$.
3. Let $d=1$ and let $$L=A(x)\frac{d^2}{dx^2},\
\text{where} \ A(x)\le C(1+|x|)^m,\
\text{for some}\ C>0,\
\text{and}\ m\le1+p.$$ Assume that $$\inf_{x\in R^d}\alpha(x)>0\ \text{and}\
\beta\le0.$$ Then the compact support property holds for $X(t)$.
It follows from Theorem EP3 that if $d=1$ and $L=(1+|x|)^m\frac{d^2}{dx^2}$, with $m\in(2,3]$, and say $\alpha=1$ and $\beta=0$, then the compact support property will depend on the particular choice of $p\in(1,2]$.
By Theorem EP1, it is necessary and sufficient to show that under the conditions of parts (1) and (2), there exists a nontrivial solution to , while under the conditions of part (3) there does not. In the case that $\alpha=1$ (or equivalently, any positive constant) and $\beta=0$, this follows from [@EP03 Theorem 5]. To extend this to $\alpha$ and $\beta$ as in the statement of the theorem, one appeals to a comparison result which we state below [@EP03 Proposition 4].
\[comparison\] Assume that $$\beta_1\le \beta_2$$ and $$0<\alpha_2\le\alpha_1.$$ If uniqueness holds for with initial data $f=0$ when $\beta=\beta_2$ and $\alpha=\alpha_2$, then uniqueness also holds when $\beta=\beta_1$ and $\alpha=\alpha_1$. Thus, from Theorem EP1, if the compact support property holds for $\beta=\beta_2$ and $\alpha=\alpha_2$, then it also holds for $\beta=\beta_1$ and $\alpha=\alpha_1$.
Theorem \[smallalpha\] and Theorem EP3 demonstrate the effect of the underlying diffusion $Y(t)$ on the compact support property in the case that $L$ is a scalar function times the Laplacian. We now consider more generally the effect of the underlying diffusion process on the compact support property. We begin with the following result which combines [@EP03 Theorem 3] with Theorem EP1.
Let $p\in(1,2]$ and assume that the underlying diffusion process $Y(t)$ explodes. Assume in addition that $$\inf_{x\in R^d}\frac{\beta(x)}{\alpha(x)}>0.$$ Then there exists a nontrivial solution to with initial data $f=0$; thus, by Theorem EP1, the compact support property does not hold.
In particular, if $Y(t)$ is explosive and $\sup_{x\in R^d}\alpha<\infty$, then a sufficient condition for the compact support property to fail is that $$\inf_{x\in R^d}\beta (x)>0.$$
What about the case that $\beta=0$? This is the case of critical branching (see the discussion of the particle process approximation to the measure-valued diffusion at the end of this section). We will prove the following result.
\[strongexp\] Let $p\in(1,2]$ and assume that the diffusion process $Y(t)$ is “strongly” explosive in the sense that $$\label{finexp}
\sup_{x\in R^d}E_x\tau_\infty<\infty.$$ Assume in addition that $\sup_{x\in R^d}\alpha(x)<\infty$ and that $$\label{lowerboundbeta}
\beta>-\frac1{\sup_{x\in R^d}E_x\tau_\infty}.$$ Then the measure-valued diffusion $X(t)$ does not possess the compact support property.
The next result shows that the restriction $\sup_{x\in R^d}\alpha(x)<\infty$ in Theorem \[strongexp\] is essential.
Let $p\in(1,2]$. Let $m\in (-\infty,\infty)$, $$L=(1+|x|)^m\Delta \ \text{in}\ R^d,$$ $\beta=0$ and $$\alpha(x)=(1+|x|)^{m-2}.$$ Then the compact support property holds for the measure-valued diffusion $X(t)$. However, if $m>2$ and $d\ge3$, the diffusion process $Y(t)$ explodes and $\sup_{x\in R^d}E_x\tau_\infty<\infty$.
There is no nontrivial, nonnegative solution to the elliptic equation $\Delta W-\frac1{(1+|x|)^2}W^p=0$ in $R^d$ (see, for example [@EP03 Theorem 6]). From this it follows that uniqueness holds for with initial data $f=0$; indeed, if uniqueness did not hold, and $u(x,t)$ were a nontrivial solution, then it would follow from the maximum principle that $u(x,t)$ is increasing in $t$. Then $W(x)=\lim_{t\to\infty}u(x,t)$ would be a nontrivial solution to the above stationary elliptic equation. (For details, see [@EP03 Theorem 4-(ii)].) Since uniqueness holds for , it follows from Theorem EP1 that the compact support property holds.
For a proof that $Y(t)$ explodes if $m>2$ and $d\ge3$, see, for example, [@EP03 Proposition 5 and Remark 1 following it]). The function $g(x)=E_x\tau_\infty\le\infty$ is the minimal positive solution to the equation $Lg=-1$ in $R^d$. It satisfies $g(x)=\lim_{n\to\infty}g_n(x)$, where $g_n(x)=E_x\tau_n$ solves $Lg_n=-1$, for $|x|<n$, and $g_n(x)=0$, for $|x|=n$. Since everything is radial, a direct calculation can be made, which shows that $\sup_{x\in R^d}g(x)<\infty$.
In the sequel, we will show that condition is not necessary in order that the compact support property fail when, say $\beta=0$ and $\alpha=1$. We will also demonstrate how $p$ and $b$ can affect the compact support property. In order to accomplish this, we first need to discuss how the concept of a measure-valued diffusion *hitting a point can be formulated and understood in terms of the compact support property. This last point is of independent interest.*
Let $\mathcal
R_t=\text{cl}\big(\cup_{s\in[0,t]}~\text{supp}(X(s))\big)$ and let $\mathcal R=\text{cl}\big(\cup_{s\ge0}~\text{supp}(X(s))\big)$. The random set $\mathcal
R$ is called the *range of $X=X(\cdot)$. A path of the measure-valued diffusion is said to *hit a point $x_0\in R^d$ if $x_0\in \mathcal R$. If $X(t)$ becomes extinct with probability one, that is, $P_\mu(X(t)=0 \ \text{for all large
}\ t)=1$, or more generally, if $X(t)$ becomes locally extinct with probability one, that is, $P_\mu(X(t,B)=0 \ \text{for all large}\ t)=1$, for each bounded $B\subset R^d$, then $x_0\in \mathcal R$ if and only if $x_0\in
\mathcal R_t$ for sufficiently large $t$. Thus, we have: $$\label{hitpoint}
\begin{aligned}
&\text{If }\ X(t)\ \text{suffers local extinction with probability one,
then}\\
&P_\mu(X \ \text{hits} \ x_0)>0\ \text{if and only if there exists
a } \ t>0
\ \text{such that}\\
&P_\mu(\cup_{0\le s\le t}~\{ \text{supp}( X(s)) \ \text{is not
compactly embedded in}\ R^d-\{x_0\}\})>0.
\end{aligned}$$**
Now although we have assumed in this paper that the underlying state space is $R^d$, everything goes through just as well on an arbitrary domain $D\subset R^d$ [@EP99]. Of course now, the compact support property is defined with respect to the domain $D$, and the underlying diffusion will explode if it hits $\partial D$ in finite time. In particular, Theorem EP1 still holds with $R^d$ replaced by $D$ [@EP99].
In light of the above observations, consider a measure-valued diffusion $X(t)$ corresponding to the log-Laplace equation on $R^d$ with $d\ge2$. The underlying diffusion process $Y(t)$ on $R^d$ corresponds to the operator $L$ on $R^d$. Let $\hat Y(t)$ denote the diffusion process on the domain $D=R^d-\{x_0\}$ with absorption at $x_0$ and corresponding to the same operator $L$. (Note that if $x_0$ is polar for $Y(t)$, then $Y(t)$ and $\hat Y(t)$ coincide when started from $x\neq x_0$. In fact, $x_0$ is always polar under the assumptions we have placed on the coefficients of $L$ [@D98].) Let $\hat X(t)$ denote the measure-valued diffusion corresponding to the log-Laplace equation , but with $R^d$ replaced by $D=R^d-\{x_0\}$. *It follows from that if $X(t)$ suffers local extinction with probability one, then the measure-valued diffusion $X(t)$ hits the point $x_0$ with positive probability if and only if $\hat X(t)$ on $R^d-\{x_0\}$ does not possess the compact support property.Furthermore, the above disccussion shows that even if $X(t)$ does not suffer local extinction with probability one, a sufficient condition for $X(t)$ to hit the point $x_0$ with positive probability is that $\hat X(t)$ on $R^d-\{x_0\}$ does not possess the compact support property.*
Consider now the following semilinear equation in the punctured space $R^d-\{0\}$. $$\label{semilinear}
\begin{aligned}
&u_t=\frac12\Delta u-u^p\ \text{in}\ (R^d-\{0\})\times(0,\infty);\\
& u(x,0)=0 \ \text{in}\ R^d-\{0\};\\
&u\ge0\ \text{in}\ (R^d-\{0\})\times[0,\infty).\\
\end{aligned}$$
By Theorem EP1, the measure-valued diffusion corresponding to the semilinear equation $u_t=\frac12\Delta u-u^p$ in $R^d-\{0\}$ will possess the compact support property if and only if has a nontrivial solution. The following theorem was recently proved in [@P04a].
Let $p>1$ and $d\ge2$.
1. If $d<\frac{2p}{p-1}$, then there exists a nontrivial solution to .
2. If $d\ge\frac{2p}{p-1}$, then there is no nontrivial solution to .
Using the method employed in [@P04a] to prove Theorem P, we will prove the following extension in the case $d<\frac{2p}{p-1}$.
\[beta\] Let $p>1$ and $d\ge2$. Let $X(t)$ denote the measure-valued diffusion on the punctured space $R^d-\{0\}$ corresponding to the semilinear equation $$\label{singular}
\begin{aligned}
&u_t=\frac12\Delta u+\beta u-u^p\ \text{in}\ (R^d-\{0\})\times(0,\infty);\\
& u(x,0)=0 \ \text{in}\ R^d-\{0\};\\
&u\ge0\ \text{in}\ (R^d-\{0\})\times[0,\infty).\\
\end{aligned}$$ Assume that $$d<\frac{2p}{p-1}.$$ Let $$\beta_0=\frac{d(p-1)-2p}{(p-1)^2}<0.$$
1. If $$\beta(x)\ge\frac{\beta_0+\kappa}{|x|^2},\ \text{for some}\ \kappa\in(0,-\beta_0],$$ then there exists a nontrivial solution to ; hence, the compact support property does not hold for $X(t)$.
2. If $$\limsup_{x\to0}|x|^2\beta(x)<\beta_0,$$ then there is no nontrivial solution to ; hence, the compoact support property holds for $X(t)$.
The restriction $\kappa\le-\beta_0$ is made to insure that $\beta$ is bounded from above.
We have the following corollary of Theorem P and Theorem 3.
\[hitting\] Let $X(t)$ denote the measure-valued diffusion on all of $R^d$ corresponding to the semilinear equation $$u_t=\frac12\Delta u+\beta u-u^p \ \text{in}\ R^d\times(0,\infty).$$
1. If $\beta\ge0$ and $d<\frac{2p}{p-1}$, then $X(t)$ hits any point $x_0$ with positive probability;
2. If $\beta\le0$ and $d\ge\frac{2p}{p-1}$, then $X(t)$ hits any point $x_0$ with probability 0;
3. If $\beta\le 0$, $d<\frac{2p}{p-1}$ and $\beta$ has a singularity at the origin such that $$\limsup_{x\to0}|x|^2\beta(x)<\frac{d(p-1)-2p}{(p-1)^2},$$ then $X(t)$ hits 0 with probability 0.
When $\beta=0$, part (1) follows immediately from Theorem P, Theorem EP1 and the discussion preceding Theorem P. For $\beta\ge0$, one appeals to the comparison result appearing above after the proof of Theorem EP3.
When $\beta=0$, it is well-known that the super-Brownian motion in the statement of the corollary suffers extinction with probability one. By comparison, this also holds when $\beta\le0$ [@EP99]. Thus, part (2) follows from Theorem P, Theorem EP1 and the discussion preceding Theorem P, while part (3) follows from Theorem 3, Theorem EP1 and the discussion preceding Theorem P.
When $\beta=0$, the results in parts (1) and (2) of Corollary \[hitting\] state that critical, super-Brownian motion hits a point with positive probability if $d<\frac{2p}{p-1}$, and with zero probability if $d\ge\frac{2p}{p-1}$. This result can be found in [@DIP89] for the case $p=2$. We are unaware of a reference when $p\in(1,2)$, although the result is certainly “known”; results in a similar vein for $p\in(1,2)$ can be found in [@I88 section 5]. See also [@F88]. The proof of parts (1) and (2) of Corollary \[hitting\] (via Theorem P) is quite different from the proof in [@DIP89].
Theorem P gives an example where the measure-valued diffusion does not possess the compact support property even though the underlying diffusion process does not explode. Indeed, the underlying diffusion is Brownian motion in $R^d-\{0\}$. Since singletons are polar for multi-dimensional Brownian motion, this process does not explode. However for $d<\frac{2p}{p-1}$, the compact support property fails for the measure-valued process.
In order to obtain an example of the phenomenon occurring in Remark 2 when the state space is the whole space, and in order to see how the parameter $p$ and the mass creation parameter $\beta$ can affect the compact support property when the state space is the whole space, we convert the set-up in Theorem \[beta\] and Theorem P to the state space $R=(-\infty,\infty)$ by considering just the radial variable, $r=|x|$, and then making a change of variables, say, $z=\frac1r-r$. One obtains an operator $L=\frac12a(z)\frac{d^2}{dz^2}+b(z)\frac d{dz}$ on $R$, where $$\begin{aligned}&a(z)\sim z^4 \ \text{as}\ \ z\to\infty, \
\lim_{z\to-\infty}a(z)=1;\\
& b(z)\sim \frac{3-d}2z^3,\ \text{as}\ \ z\to\infty, \
\lim_{z\to-\infty}b(z)=0.
\end{aligned}$$ Also, one has $\alpha=1$. One has $\beta=0$ in Theoerm P and one has
\[joint\] $$\label{part1}
\beta(z)\ge\frac{4(\beta_0+\kappa)}{\big((z^2+4)^\frac12-z)\big)^2}
\ \ \text{in part (1) of Theorem \ref{beta}},$$ $$\label{part2}
\limsup_{z\to\infty}\frac14\big((z^2+4)^\frac12-z)\big)^2\beta(z)
<\beta_0 \ \ \text{in part (2)
of Theorem \ref{beta}}.$$
It follows then that $\liminf_{z\to\infty}\frac{\beta(z)}{z^2}\ge\beta_0+\kappa$ in part (1) of Theorem 3 and $\limsup_{z\to\infty}\frac{\beta(z)}{z^2}<\beta_0$ in part (2) of Theorem 3.
Consider first the phenomenon mentioned in Remark 2 with regard to Theorem P. After the change of variables, the state space is $R$ and the operator $L$ depends on the parameter $d$. The diffusion corresponding to $L$ is nonexplosive for all $d$, as it inherits this property from the original process before the change of variables. Also $\alpha=1$ and $\beta=0$. However, if $d<\frac{2p}{p-1}$, then the compact support property fails.
With the same setup as in the previous paragraph, we also see how the parameter $p$ affects the compact support property—the property will hold if and only if $d\ge\frac{2p}{p-1}$.
Now consider the above change of variables applied to Theorem 3. We see that if $d<\frac{2p}{p-1}$, then $\beta$ affects the compact support property—it holds if $\beta$ satisfies and does not hold if $\beta$ satisfies . Note that demonstrates that an unbounded change in $\beta$ is necessary here to affect the compact support property.
The discussion at the end of Remark 3 leads us to pose the following question.
Can a bounded change in the mass creation parameter $\beta$ affect the copact support property?
The particle process approximation to the measure-valued diffusion, which we discuss in some detail below, along with the remarks on the probabilistic intuition for the role of branching in Theorem 1, which follows that discussion, might suggest that in the case that $\alpha=1$ and $\beta=0$, the compact support property for the measure-valued diffusion is equivalent to the non-explosion property for the underlying diffusion. Remark 2 following Corollary \[hitting\] shows that one of these inclusions is not true. Theorem \[strongexp\] shows that if the explosion requirement is replaced by the strong explosion condition , the other inclusion becomes true. Whether this inclusion is true in its original form we leave as an open question.
If $\alpha=1$, $\beta=0$ and the underlying diffusion process explodes, does the compact support property necessarily fail?
This is an interesting question from the point of view of partial differential equations. In light of Theorem EP1, the question is whether nonuniqueness for the linear Cauchy problem $u_t=Lu$ guarantees nonuniqueness for the semilinear Cauchy problem $u_t=Lu-u^p$ with vanishing initial data.
Note that if the answer to Open Question 1 is negative, then by Theorem EP4, the answer to Open Question 2 is affirmative.
Theorems \[smallalpha\]-\[beta\] will be proved successively in the sections that follow.
We now turn to an intuitive probabilistic understanding of the role of the branching in Theorem 1. We recall the particle process approximation to the measure-valued diffusion in the case that $\alpha$ and $\beta$ are bounded. Consider first the case $p=2$, the case in which the offspring distribution has finite variance. Let $ \hat R^d=R^d\cup \{\Delta\}$ denote the one-point compactification of $R^d$. One may consider the diffusion process $Y(t)$ to live on $\hat R^d$; if $Y(t)$ does not explode, then it never reaches $\Delta$, while if $Y(t)$ does explode, then it enters the state $\Delta$ upon leaving $R^d$, and remains there forever. For each positive integer n, consider $N_{n}$ particles, each of mass $\frac{1}n$, starting at points $y_{i}^{(n)}(0)\in
R^d,i=1,2,\dots,N_{n},$ and performing independent branching diffusion according to the process $Y(t)$, with branching rate $cn,c>0$, and spatially dependent branching distribution $\{p_{k}^{(n)}(y)\}_{k=0}^{\infty}$, where
$$\sum_{k=0}^{\infty}kp_{k}^{(n)}(y)=1+\frac{\gamma(y)}n+o(\frac1n),\
\text{as}\ n\to\infty;$$ $$\sum_{k=0}^{\infty}(k-1)^{2}p_{k}^{(n)}(y)=m(y)+o(1),\ \text{as}\ n\to
\infty, \ \text{uniformly in}\ y,$$ with $m,\gamma \in C^\alpha( R^d)$ and $m(y)>0$. Let $N_{n}(t)$ denote the number of particles alive at time t and denote their positions by $\{Y_{i}^{(n)}(t)\}_{i=1}^{N_{n}(t)}$. Denote by $\mathcal{M}_{F}( R^d)$ ($\mathcal{M}_{F}(\hat R^d)$) the space of finite measures on $ R^d$ ($\hat R^d$). Define an $\mathcal{M}_{F}(\hat R^d)$- valued process $X_{n}(t)$ by $X_{n}(t)=\frac{1}n\sum_{1}^{N_{n}(t)}\delta_{Y_{i}^{(n)}(t)}(\cdot)$. Note that $X_n$ is càdlàg. Denote by $P^{(n)}$ the probability measure corresponding to $\{X_{n}(t), 0\le t<\infty\}$ on $D([0,\infty),\mathcal M_F(\hat R^d))$, the space of càdlàg paths with the Skorohod topology,
Assume that $m(y)$ and $\gamma(y)$ are bounded from above. One can show that if $w-\lim_{n\to\infty} X_n(0)=\mu\in \mathcal M_F( R^d)$, then $P^*_\mu=w-\lim_{n\to\infty}P^{(n)}$ exists in $D([0,\infty),\mathcal M_F(\hat R^d))$. Furthermore, the measure $P_\mu^*$ restricted to $D([0,\infty),\mathcal M_F(R^d))$ satisfies and with $\beta(y)=c\gamma(y)$, $\alpha(y)=\frac{1}{2}cm(y)$ and $p=2$ (see [@EP99]). Denoting this restriction by $P_\mu^*|_{R^d}$, it then follows that $P_\mu=P_\mu^*|_{R^d}$. In fact, one can show that $P_\mu$ is supported on the space of continuous paths, $ C([0,\infty),\mathcal M_F( R^d))$.
One should think of $\beta$ and $\alpha$ as the *mass creationand the *variance parameters respectively of the branching.**
For the case that $p\in(1,2)$, one cooks up a sequence of distributions $\{p_k^{(n)}(y)\}_{n=1}^\infty$ for which the generating functions $\Phi^{(n)}(s;y)=\sum_{k=0}^\infty
p_k^{(n)}(y)s^y$ satisfy $\lim_{n\to\infty}n^p\big(\Phi^{(n)}(1-\frac\lambda n;y)-(1-\frac\lambda n)
\big)=\alpha(y)\lambda^p-\beta(y)\lambda$. These offspring distributions $\{p_k^{(n)}(y)\}_{n=1}^\infty$ will possess all moments smaller than $p$. Again, $\beta$ can be thought of as the mass creation parameter. As above, each particle is given mass $\frac1n$, but in the present case, the branching rate is $n^{p-1}$. The same construction and conclusion as above holds, although in this case the paths are not continuous, but only càdlàg [@D93].
With the above set-up, we can now give some intuition concerning Theorem 1. Consider two particular cases of the above construction. In both cases we will assume that at time 0 there are $n$ particles, all positioned at $y=0$; that is, $N_n=n$ and $y_i^{(n)}=0$, for $i=1,2,\ldots n$. Then the initial measure, both for the approximating process and the limiting one, will be $\mu=\delta_0$. We will also assume that the diffusion $Y(t)$ does not explode.
The first case is the completely trivial case in which there is *no branching at all. This degenerate case corresponds to $\beta=\alpha=0$ (and thus does not actually fit into the above set-up). In this case, $X_n(t)$ is a random probability measure with $n$ atoms of mass $\frac1n$ positioned at $n$ IID points, distributed according to the distribution of $Y(t)$. Thus, by the law of large numbers, $X(t)=w-\lim_{n\to\infty}X_n(t)$ is the deterministic measure $\text{dist}(Y(t))$. Since $\text{dist}(Y(t))$ is not compactly supported for $t>0$, it follows that the compact support property does not hold for $X(t)$ in this trivial case.*
Now consider the case of critical, binary branching; that is, $p_0^{(n)}=p_2^{(n)}=\frac12$. Letting $c=1$, it then follows that $\beta=0$ and $\alpha=\frac12$. In such a case, discrete, deterministic branching may be used in place of the exponentially distributed branching described above: the branching will occur deterministically at integral multiples of $\frac1{n}$ instead of according to an exponential random variable with parameter $n$. This deterministic mechanism will be more convenient in what follows.
Consider the process at, say, time $t=1$. At this time, the process has undergone $n$ generations of branching. The $N_n(1)$ particles $\{Y_i^{(n)}(1)\}_{i=1}^{N_n(1)}$ may be grouped in the following way: for each $j=1,2,...n$, let $M_j^{(n)}(1)$ denote the number of descendants of $y_j^{(n)}$ and let $\{Z_{j;i}^{(n)}(1)\}_{i=1}^{M_j^{(n)}(1)}$ denote their positions. Then $N_n(1)=\sum_{j=1}^n M_j^{(n)}(1)$ and $\cup_{j=1}^n\cup_{i=1}^{M_j^{(n)}(1)} Z_{j;i}^{(n)}(1)=\{Y_i^{(n)}(1)\}_{i=1}^{N_n(1)}$. Let $$W_j^{(n)}(1)=\frac{1}n\sum_{1}^{M_j^{(n)}(t)}\delta_{Z_{j;i}^{(n)}(1)}(\cdot),
\ j=1,\cdots n.$$ Note that the $\{W_j^{(n)}(1)\}_{j=1}^n$ are independent point measures, and that the approximate measure valued diffusion, $X_n(1)$, is given by $$X_n(1)=\frac1n\sum_{j=1}^n W_j^{(n)}(1).$$ The random variable $M_j^{(n)}(1)$ is just the number of individuals alive in the $n$-th generation of a critical, binary branching process starting from a single individual at the zeroth generation. Denoting such a branching process by $\{U_n\}_{n=0}^\infty$, standard results [@AN72] give $$\label{condbranch}
\begin{aligned}
&P(U_n>0)\sim\frac2n,\ \text{ as}\ n\to\infty;\\
&\lim_{n\to\infty}P(\frac{U_n}n>z|U_n>0)=\exp(-2z),\ \text{ for}\ z\ge0.
\end{aligned}$$ In particular, from it follows that the distribution of the number of nonzero $W_j^{(n)}(1)$ is approximately $Bin(n,\frac2n)$; thus, by the Poisson approximation, as $n\to\infty$, the contribution to $X_n(1)$ will come from approximately a $Poiss(2)$ number of nonzero measures from among the $\{W_j^{(n)}(1)\}_{j=1}^n$. In particular, for example, with probability about $2\exp(-2)$ there will be exactly one nonzero random variable from among $\{W_j^{(n)}(1)\}_{j=1}^n$. Conditioned on this event, denote the unique nonzero random variable by $W_{j_n}^{(n)}(1)$. It follows easily from that $$\label{condbranch1}
\liminf_{n\to\infty}P(\frac{M_{j_n}(s)}n\ge C|\frac{M_{j_n}(1)}n\ge K)>0,\
\text{for all}\ C,K>0 \ \text{and}\ s\in(0,1].$$ Now is perhaps misleading, as it may suggest that if one fixes a time $s<1$ and some $c_1>0$, then the probability that at time 1 there are at least $c_1n$ particles with a common ancestor at time $s$ remains bounded away from 0 as $n\to\infty$. Note that if this were to occur, then if one looked at the contribution to $X_n(1)$ coming from these particular particles with a commmon ancestor at time $s$, and let $n\to\infty$, the law of large numbers would come into play as it did above in the degenerate case and cause $X_n(1)$ to be dominated by $\delta \text{dist}(Y(1-s))$, for some $\delta>0$. Since the distribution of $\delta \text{dist}(Y(1-s))$ is not compactly supported, regardless of the choice of $L$, one would conclude that the compact support property does not hold when $\beta=0$ and $\alpha=\frac12$, for all $L$. This is of course false. The above fallacy is a well-known paradox in the theory of branching processes—if a critical, binary, branching process has $n$ individuals alive at the $n$-th generation, then the probability that they had a common ancestor at time $\rho n$ goes to 0 at $n\to\infty$, for all $\rho<1$.
The above discussion suggests that one way for the compact support property to break down is for the branching mechanism to be spatially dependent and to decay sufficiently fast as $|x|\to\infty$ so that the law of large numbers will come into play. Furthermore, the faster the diffusion is, the more quickly individual particles that begin together become statistically uncorrelated, so one might expect that the stronger the diffusion, the weaker the threshold on the decay rate in order for the compact support property to break down. If the diffusion process $Y(t)$ corresponds to the operator $L=\frac12(1+|x|)^m\Delta$, for $m\in[0,2]$, then for $m=0$ one obtains Brownian motion, while for $m\in(0,2]$, one obtains a time-changed Brownian motion with the diffusion sped up, the speed increasing in $m$. Theorem one shows that for such an underlying diffusion, for any $\epsilon>0$, the rate $\exp(-|x|^{2-m+\epsilon})$ is sufficiently fast to cause the compact support property to fail, but the rate $\exp(-|x|^{2-m})$ is not fast enough.
Proof of Theorem \[smallalpha\] {#S:smallalpha}
===============================
.
Let $u(x,t)$ be any solution of with initial data $g=0$. By Theorem EP1, we need to show that $u\equiv0$. Define $\hat U(x,t)$ through the equality $u(x,t)=\hat U(x,t)\exp(\lambda(1+|x|^2)^{\frac{2-m}2}(t+\delta))$, for some $\lambda,\delta>0$. Then $$\label{Laplacian}
\begin{aligned}
&\exp(-\lambda(1+|x|^2)^{\frac{2-m}2}(t+\delta))A(x)\Delta u=\\
&A(x)\Big(\Delta \hat U+2\lambda(2-m)(1+|x|^2)^{-\frac m2}(t+\delta) x\cdot\nabla
\hat U\Big)\\
&+A(x)\Big((t+\delta)^2\lambda^2(2-m)^2(1+|x|^2)^{-m}|x|^2\\
&+(t+\delta)\lambda d(2-m)(1+|x|^2)^{-\frac m2}-(t+\delta)\lambda(2-m)m(1+|x|^2)^{-\frac m2-1}
|x|^2)\Big)\hat U
\end{aligned}$$ Also, using the bound on $\alpha$ in the statement of the theorem, we have $$\label{otherterms}
\begin{aligned}
&\exp(-\lambda(1+|x|^2)^{\frac{2-m}2}(t+\delta))\Big(\beta u-\alpha u^p-u_t\Big)=\\
&\Big(\beta-\lambda(1+|x|^2)^{\frac {2-m}2}\Big)\hat U-
\alpha\exp\big((p-1)\lambda(1+|x|^2)^{\frac{2-m}2}(t+\delta)\big)
\hat U^p.
\end{aligned}$$
Since $u$ is a solution to , the sum of the left hand sides of and is equal to 0. Consider now the sum of the terms on the right hand sides of and with the variable $t$ restricted by $0\le t\le \delta$. By assumption, $\alpha(x)\ge C_1\exp(-C_2|x|^{2-m})$. Thus, the coefficient of $\hat U^p$ will be bounded away from 0 if $$\label{lambdadelta1}
\lambda\delta=\frac{C_2}{p-1}.$$ In the case that $m=2$, the coefficient of $\hat U$ is bounded from above. Otherwise, the two unbounded terms in the coefficient of $\hat U$ are $A(x)(t+\delta)^2\lambda^2(2-m)^2(1+|x|^2)^{-m}|x|^2$ and $-\lambda(1+|x|^2)^{\frac{2-m}2}$. By assumption, there exists a $C_4>0$ such that $A(x)\le C_4(1+|x|^2)^\frac m2$. Thus, in order to guarantee that the coefficient of $\hat U$ is bounded from above, it suffices to have $\lambda= C_4(2\delta)^2\lambda^2(2-m)^2$. Using to substitute for $(\lambda\delta)^2$ on the right hand side above, we have $$\label{lambdadelta2}
\lambda=4C_4(2-m)^2C_2^2(p-1)^{-2}.$$ With $\lambda$ and $\delta$ chosen as in and , it then follows that $$\label{subsol}
\begin{aligned}
&A(x)\Big(\Delta \hat U+2\lambda(2-m)(1+|x|^2)^{-\frac m2}(t+\delta) x\cdot\nabla
\hat U\Big)\\
&+c_1\hat U-c_2\hat U^p\ge0,\ \text{for}\ (x,t)\in R^n\times[0,\delta],
\ \text{for some} \ c_1,c_2>0.
\end{aligned}$$
Let $M_{R,K}(x,t)=(1+|x|)^{\frac2{p-1}}
(R-|x|)^{-\frac2{p-1}}\exp(K(t+1))$, for $(x,t)\in B_R\times(0,\infty)$. In [@EP03 proof of Theorem 2], it was shown that for any operator $\mathcal L$ satisfying the conditions of Theorem EP2, there exists a $K>0$ such that for all $R>0$ $$\label{testfunction}
\mathcal LM_{R,K}(x,t)\le0,
\ (x,t)\in B_R\times(0,\infty),$$ where $B_R$ denotes the ball of radius $R$ centered at the origin. In particular the operator $\mathcal A=
A(x)\Big(\Delta +2\lambda(2-m)(1+|x|^2)^{-\frac m2}(t+\delta) x\cdot\nabla
\Big)$ satisfies the conditions of Theorem EP1, except for the fact that it is time inhomogeneous. The time inhomogeneity causes no problem since we are only considering $t\in[0,\delta]$ and since for fixed $x$, everything in sight is uniformly bounded for $t\in[0,\delta]$. Thus the proof of shows that $$\label{supersol}
\mathcal AM_{R,K}(x,t)\le0,
\ (x,t)\in B_R\times(0,\delta].$$ Since $0=\hat U(x,0)\le M_{R,K}(x,0)$ and $\hat U(y,t)\le M_{R,K}(y,t)=\infty$, for $y\in \partial B_R$, it follows from , and the maximum principle for semilinear equations [@EP03 Proposition 1] that $$\hat U(x,t)\le M_{R,K}(x,t), \text{for}\ (x,t)\in B_R\times[0,\delta].$$ Letting $R\to\infty$, we conclude that $\hat U(x,t)=0$, for $(x,t)\in R^d
\times(0,\delta]$. Thus, we also have $u(x,t)=0$, for $(x,t)\in R^d\times[0,\delta]$. Since $u$ satisfies a time homogeneous equation, we conclude that in fact $u(x,t)=0$, for $(x,t)\in R^d\times[0,\infty)$. This completes the proof of part (1).
*Proof of part (2).By Theorem EP1, we need to show that there is a nontrivial solution to with initial data $g=0$. Let $u(x,t)$ be any solution of with initial data $g=0$. Define $\hat U(x,t)$ through the equality $u(x,t)=\hat U(x,t)\exp(\lambda(1+|x|^2)^{\frac{2-m+\kappa}2})$, where $\kappa\in(\delta,\epsilon)$, and $\epsilon$ and $\delta$ are as in the statement of the theorem. Now and hold with the following changes: (i) $m$ is replaced by $m-
\kappa$; (ii) $(t+\delta)$ is replaced by 1; (iii) the term $\lambda(1+|x|^2)^{\frac{2-m}2}$ in is deleted. As before, the sum of the left hand sides of and is equal to 0. By assumption, $\alpha(x)\le C\exp(-|x|^{2-m+\epsilon})$. Thus, the coefficient of $\hat U^p$ in the amended version of is bounded from above, since $\kappa<\epsilon$. The coefficient of $\hat U$ in the amended version of is $\beta$, and by assumption, there exists a $C_5>0$ such that $\beta\ge-C_5(1+|x|^2)^{\frac{2-m+2\delta}2}$. The coefficient of $\hat U$ in the amended version of is $A(x)\Big(\lambda^2(2-m+\kappa)^2(1+|x|^2)^{-m+\kappa}|x|^2
+\lambda d(2-m+\kappa)(1+|x|^2)^{-\frac {m-\kappa}2}-
\lambda(2-m+\kappa)(m-\kappa)(1+|x|^2)^{-\frac {m-\kappa}2-1}
|x|^2)\Big)$. By assumption, there exists a $C_3>0$ such that $A(x)\ge C_3(1+|x|^2)^{\frac m2}$. It is easy to check that by choosing $\lambda$ sufficiently large, the factor multiplying $A(x)$ above will be bounded from below by $\frac{2C_5}{C_3}(1+|x|^2)^{\frac{2-2m+2\kappa}2}$. (For $|x|\le\frac12$, use the second term in the parentheses, and for $|x|>\frac12$ use the first and third terms.) Thus, the coefficient of $\hat U$ in the amended version of is greater or equal to $2C_5(1+|x|^2)^{\frac{2-m+2\kappa}2}$. Since $\kappa>\delta$, it follows that the coefficient of $\hat U$ from the sum of the right hand sides of and is bounded below by a positive constant.*
The above analysis shows that $\hat U$ satisfies the equation $$\label{hatU}
\begin{aligned}
&\hat U_t=A(x)
\Big(\Delta \hat U+2\lambda(2-m+\kappa)(1+|x|^2)^{-\frac {m-\kappa}2} x\cdot\nabla
\hat U\Big)+\hat \beta\hat U-\hat \alpha \hat U^p;\\
&\hat U(x,0)=0,
\end{aligned}$$ where $\hat \beta\ge C_6>0$ and $\hat \alpha\le C_7$, and that uniqueness for the original equation is equivalent to uniqueness for . We will show below that the diffusion process corresponding to the operator $A(x)
\Big(\Delta +2\lambda(2-m+\kappa)(1+|x|^2)^{-\frac {m-\kappa}2} x\cdot\nabla
\Big)$ explodes. Thus, it follow from Theorem EP4 that uniqueness does not hold for . Consequently, uniqueness does not hold for the original equation with initial data $g=0$; thus, by Theorem EP1, the compact support property does not hold.
It remains to show that the diffusion corresponding to the operator $A(x)
\Big(\Delta +2\lambda(2-m+\kappa)(1+|x|^2)^{-\frac {m-\kappa}2} x\cdot\nabla
\Big)$ explodes. The diffusion in question is a time change of the diffusion corresponding to $\Delta+2\lambda(2-m+\kappa)(1+|x|^2)^{-\frac {m-\kappa}2} x\cdot\nabla$. It is not hard to show that since $A(x)\ge C_3(1+|x|^2)^{\frac m2}$, explosion will occur for the diffusion in question if it occurs for the diffusion corresponding to the operator $C_3(1+|x|^2)^{\frac m2}
\Big(\Delta +2\lambda(2-m+\kappa)(1+|x|^2)^{-\frac {m-\kappa}2} x\cdot\nabla
\Big)$. This latter operator is radially symmetric, and its radial component is of the form $p(r)\frac{d^2}{dr^2}+q(r)\frac d{dr}$, with $p(r)$ satisfying $c_1(1+r)^m\le p(r)\le c_2(1+r)^m$, for all $r\ge0$, and $q(r)$ satisfying $c_3(1+r)^{1+\kappa}\le q(r)\le c_4(1+r)^{1+\kappa}$, for $r>1$, where $c_1, c_2,c_3,c_4>0$. This diffusion explodes by the Feller criterion [@P95].
Proof of Theorem \[strongexp\]
==============================
We recall the $h$-transform theory for measure-valued diffusions, developed in [@EP99 Section 2]. Let $X(t)$ be the measure-valued diffusion corresponding to the triplet $(L,\beta,\alpha)$, and let $h(x)$ be a twice differentiable, positive function. The $h$-transform $L^h$ of the operator $L$ is defined by $L^hf=\frac1hL(fh)$. Equivalently, $L^h=L+a\frac{\nabla h}h+\frac{Lh}h$. Let $L_0^h=L+a\frac{\nabla h}h$ so that $$\label{L_0L^h}
L^h=L_0^h+\frac{Lh}h$$ Then the measure-valued diffusion $X^h(t)$ corresponding to the triplet $(L^h_0,\beta+\frac{Lh}h,\alpha h^{p-1})$ is almost surely, mutually absolutely continuous with respect to $X(t)$ and the Radon-Nikodym derivative is equal to the deterministic, time-independent function $h$; that is $\frac{dX^h(t)}{d X(t)}(x)=h(x)$. From this, it follows in particular that $X(t)$ possesses the compact support property if and only if $X^h(t)$ does.
We utilize this $h$ transform as follows. Let $X(t)$ correspond to the triplet $(L,\beta,\alpha)$, as in the statement of the theorem. Let $g(x)=E_x\tau_\infty$. By assumption, $g$ is finite. In fact, $g$ is the minimal positive solution to the equation $Lg=-1$ in $R^d$. (To see this, let $g_n$ solve $Lg_n=-1$ in the ball of radius $n$ centered at the origin, with $g_n=0$ on the boundary of the ball. By Ito’s formula, $g_n(x)=E_x\tau_n$. Then $g(x)=\lim_{n\to\infty}g_n(x)$, and the minimality of $g$ follows from the maximum principal and the vanishing Dirichlet boundary data on the boundary of the ball.) Let $M=\sup_{x\in R^d}g(x)$. By assumption, $M<\infty$. For $\delta>0$, define $h(x)=M+\delta-g(x)$. Let $X^h(t)$ be the measure-valued diffusion corresponding to the triplet $(L_0^h,\beta+\frac{Lh}h,\alpha h)$. We have $\sup_{x\in R^d}h(x)= M+\delta$ and $Lh=1$. Since $\alpha$ is bounded, so is $\alpha h$. Also, by , it follows that if $\delta$ is chosen sufficiently small, then $$\inf_{x\in R^d}(\beta(x)+\frac{Lh(x)}{h(x)})
=\inf_{x\in R^d}(\beta(x)+\frac1{M+\delta})>0.$$ Thus, if we show that the diffusion process corresponding to $L_0^h$ explodes, then it follows from Theorem EP4 that $X^h(t)$ does not possess the compact support property. Consequently, by the above $h$-transform theory for measure-valued diffusions, neither does $X(t)$ possess the compact support property. Thus, to complete the proof of the theorem, it suffices to show that the diffusion process corresponding to $L_0^h$ explodes.
Let $p(t,x,y)=\mathcal P_x(Y(t)\in dy)$ denote the transition kernel for the semigroup corresponding to $L$. Let $Y_0(t)$ denote the diffusion process corresponding to the operator $L_0^h$, and denote the corresponding probabilities and expectations for the process starting from $x$ by $\mathcal P_x^0$ and $\mathcal E_x^0$. Let $p_0(t,x,y)=\mathcal P_x^0(Y_0(t)\in dy)$ denote the transition kernel for the semigroup corresponding to $L^h_0$. Finally, let $p^h(t,x,y)$ denote the transition kernel for the semigroup corresponding to $L^h$. In light of and the fact that $\frac{Lh}h=\frac1h$, the Feynman-Kac formula gives $$\label{FK}
p^h(t,x,y)=\mathcal E_x^0\exp(\int_0^t\frac1{h(Y^0(s))}ds) 1_{dy}(Y_0(t)).$$ In particular then, $$\label{p^hp_0}
p^h(t,x,y)\ge p_0(t,x,y).$$ On the other hand, it follows from $h$-transform theory for diffusions [@P95 Theorem 4.1.1] that $$\label{htransform}
p^h(t,x,y)=p(t,x,y)\frac{h(y)}{h(x)}.$$ From and along with the fact that $h$ is bounded and bounded away from 0, we conclude that $$\label{pp_0}
p_0(t,x,y)\le Cp(t,x,y), \ \text{for some}\ C>0.$$ Since $$E_x\tau_\infty=E_x\int_0^\infty 1_{R^d}(Y(s))ds=
\int_0^\infty P_x(Y(s)\in R^d)=\int_0^\infty\int_{R^d}p(t,x,y)dy,$$ it follows from that $$\label{finiteintegral}
\int_0^\infty\int_{R^d}p(t,x,y)dy<\infty.$$ Letting $\tau^0_\infty$ denote the explosion time for $Y_0(t)$, we conclude from - that $$E_x^0\tau_\infty^0=\int_0^\infty\int_{R^d}p_0(t,x,y)dy<\infty.$$ In particular then, the process $Y^0(t)$ explodes.
Proof of Theorem \[beta\]
=========================
We will show that uniqueness does not hold for . Then by Theorem EP1, the compact support property does not hold. By the comparison result stated after the proof of Theorem EP3, we may assume that $\beta=\frac{\beta_0+\kappa}{|x|^2}$. Since the problem is now radially symmetric, it suffices to show that there exists a nontrivial solution to the radially symmetric equation $$\label{onedim}
\begin{aligned}
&u_t=u_{rr}+\frac {d-1}ru_r-u^p,\ r\in(0,\infty), \ t>0;\\
&
u(r,0)=0, \ r\in(0,\infty);\\
&u\ge0,\ r\in(0,\infty), \ t\ge0.
\end{aligned}$$ The function $W(x)=\kappa^{\frac1{p-1}}r^{-\frac2{p-1}}$ is a positive, stationary solution of the parabolic equation $u_t=u_{rr}+\frac {d-1}ru_r+\frac{\beta_0+\kappa}{r^2}u-u^p$ in $(0,\infty)$. By [@P04 Theorem 2-ii], the fact that there exists a nontrivial positive, stationary solution guarantees that uniqueness does not hold for the corresponding parabolic equation with initial data 0; that is, uniqueness does not hold for . Actually, the result in [@P04] is for equations with domain $R^d$, $d\ge1$, whereas the domain here is $(0,\infty)$. One can check that the proof also holds in a half space, but more simply, one can make the change of variables $z=\frac1x-x$, which converts the problem to all of $R$.
*Proof of part (2).We will show that uniqueness holds for . Then by Theorem EP1, the compact support property holds. For $\epsilon$ and $R$ satisfying $0<\epsilon<1$ and $R>1$, and for some $l\in(0,1]$, define $$\label{phi}
\phi_{R,\epsilon}(x)=((|x|-\epsilon)(R-|x|))^{-\frac2{p-1}}(1+|x|)^\frac2{p-1}
(1+\frac{\epsilon^l}{ |x|^l} R^\frac2{p-1}).$$ Also, for $R$ and $\epsilon$ as above, and some $\gamma>0$, define $$\label{psi}
\psi_{R,\epsilon}(x,t)=\phi_{R,\epsilon}(x)\exp(\gamma(t+1)).$$ Note that $\psi_{R,\epsilon}(x,0)>0$, for $|x|\in(\epsilon, R)$, and $\psi_{R,\epsilon}(x,t)
=\infty$, for $|x|=\epsilon$ and $|x|=R$. We will show that for all sufficiently large $R$ and all sufficiently small $\epsilon$, and for $\gamma$ sufficiently large and $l$ sufficiently small, independent of those $R$ and $\epsilon$, one has $$\label{supersolution}
\frac12\Delta\psi_{R,\epsilon}
+\beta \psi_{R,\epsilon}-\psi^p_{R,\epsilon}
-(\psi_{R,\epsilon})_t\le0,\ \ \text{for} \
\epsilon<|x|<R\ \ \text{and}\ t>0.$$ It then follows from the maximum principle for semi-linear equations [@EP03 Proposition 1] that every solution $u(x,t)$ to satisfies $$\label{upsi}
u(x,t)\le\psi_{R,\epsilon}(x,t), \ \ \text{for}\ \epsilon<|x|<R
\ \ \text{and} \ t\in[0,\infty).$$ Substituting and in , letting $\epsilon\to0$, and then letting $R\to\infty$, we conclude that $u(x,t)\equiv0$. Thus, it remains to show .*
From now on we will use radial coordinates, writing $\phi(r)$ for $\phi(x)$ with $|x|=r$ and similarly for $\psi$. We have $$\label{first}
\begin{aligned}
&\exp(-\gamma(t+1))(\psi_{R,\epsilon})_r=\\
&-(\frac2{p-1})((r-\epsilon)(R-r))^{-\frac2{p-1}-1}
(R+\epsilon-2r)(1+r)^\frac2{p-1}(1+\frac{\epsilon^l}{r^l}R^\frac2{p-1})\\
&+(\frac2{p-1})((r-\epsilon)(R-r))^{-\frac2{p-1}}(1+r)^{\frac2{p-1}-1}
(1+\frac{\epsilon^l}{r^l}R^\frac2{p-1})\\
&-l((r-\epsilon)(R-r))^{-\frac2{p-1}}(1+r)^{\frac2{p-1}}
\frac{\epsilon^l}{r^{l+1}}R^\frac2{p-1},
\end{aligned}$$ and $$\label{second}
\begin{aligned}
&\exp(-\gamma(t+1))\big((r-\epsilon)(R-r)\big)^{-\frac2{p-1}-2}
(\frac12\psi_{R,\epsilon})_{rr}=\\
&(\frac1{p-1})(\frac2{p-1}+1)(R+\epsilon-2r)^2(1+r)^\frac2{p-1}
(1+\frac{\epsilon^l}{r^l}R^\frac2{p-1})\\
&+(\frac2{p-1})(r-\epsilon)
(R-r)(1+r)^{\frac2{p-1}}(1+\frac{\epsilon^l}{r^l}R^\frac2{p-1})\\
&-(\frac2{p-1})^2(r-\epsilon)(R-r)(R+\epsilon-2r)(1+r)^{\frac2{p-1}-1}
(1+\frac{\epsilon^l}{r^l}R^\frac2{p-1})\\
&+l(\frac2{p-1})(r-\epsilon)(R-r)(R+\epsilon-2r)(1+r)^{\frac2{p-1}}
\frac{\epsilon^l}{r^{l+1}}R^{\frac2{p-1}}\\
&+(\frac1{p-1})(\frac2{p-1}-1)((r-\epsilon)(R-r))^2
(1+r)^{\frac2{p-1}-2}
(1+\frac{\epsilon^l}{r^l}R^\frac2{p-1})\\
&-l(\frac2{p-1})((r-\epsilon)(R-r))^2
(1+r)^{\frac2{p-1}-1}
\frac{\epsilon^l} {r^{l+1}}R^\frac2{p-1}\\
&+\frac{l(l+1)}2((r-\epsilon)(R-r))^2(1+r)^\frac2{p-1}
\frac\epsilon{r^{l+2}}R^\frac2{p-1}.
\end{aligned}$$ Using , and the fact that $\frac2{p-1}+2=\frac{2p}{p-1}$, we have $$\label{other}
\begin{aligned}
&\exp(-\gamma(t+1))\big((r-\epsilon)(R-r)\big)^{-\frac2{p-1}-2}\times\\
&\Big(\frac {d-1}{2r}(\psi_{R,\epsilon})_r+\beta\psi_{R,\epsilon}-\psi_{R,\epsilon}^p
-(\psi_{R,\epsilon})_t\Big)=\\
&-(\frac2{p-1})\frac {d-1}{2r}(r-\epsilon)(R-r)(R+\epsilon-2r)
(1+r)^{\frac2{p-1}}(1+\frac{\epsilon^l}{r^l}R^\frac2{p-1})\\
&+(\frac2{p-1})\frac {d-1}{2r}((r-\epsilon)(R-r))^2
(1+r)^{\frac2{p-1}-1}(1+\frac{\epsilon^l}{r^l}R^\frac2{p-1})\\
&-l\frac {d-1}{2r}((r-\epsilon)(R-r))^2
(1+r)^{\frac2{p-1}}\frac{\epsilon^l} {r^{l+1}}R^\frac2{p-1}\\
&+(\beta-\gamma)((r-\epsilon)(R-r))^2(1+r)^{\frac2{p-1}}
(1+\frac{\epsilon^l}{r^l}R^\frac2{p-1})\\
&-(1+r)^{\frac{2p}{p-1}}(1+\frac{\epsilon^l}{r^l}R^\frac2{p-1})^p
\exp((p-1)\gamma(t+1)).
\end{aligned}$$ We will show that for all sufficiently large $R$ and sufficiently small $\epsilon$, and for sufficiently large $\gamma$ and sufficiently small $l$, independent of those $R$ and $\epsilon$, the sum of the right hand sides of and is non-positive. This will then prove .
We will denote the seven terms on the right hand side of by $J_1-J_7$, and the five terms on the right hand side of by $I_1-I_5$. Note that the terms that are positive are $J_1, J_2, J_4, J_5, J_7$ and $I_2$. (Since $\beta$ is bounded from above, $I_4$ is negative for $\gamma$ sufficiently large.) In what follows, $M$ will denote a positive number that can be made as large as one desires by choosing $\gamma$ sufficiently large. Consider first those $r$ satisfying $r\ge cR$, where $c$ is a fixed positive number. For $r$ in this range, we have $|I_5|\ge MR^{\frac2{p-1}+2}(1+\epsilon^l R^{\frac2{p-1}-l})$, It is easy to see that for $M$ sufficiently large, $|I_5|$ dominates each of the positive terms, uniformly over large $R$ and small $\epsilon$, and thus (since $M$ can be made arbitrarily large) also the sum of all of the positive terms. Now consider those $r$ for which $\delta_0\le r\le C$, for some constants $0<\delta_0<C$. For $r$ in this range, we have $|I_4|\ge MR^2(1+\epsilon^l R^{\frac2{p-1}})$, and it is easy to see that for $M$ sufficiently large, $|I_4|$ dominates each of the positive terms, uniformly over large $R$ and small $\epsilon$, and thus, also the sum of all of the positive terms. One can also show that the transition from $r$ of order unity to $r$ of order $R$ causes no problem. Thus, we conclude that for any fixed $\delta_0>0$, for all $l\in(0,1]$ and $\gamma$ sufficiently large, the sum of the right hand sides of and is negative for all large $R$ and small $\epsilon$.
We now turn to the case that $\epsilon\le r\le \delta_0$. (Note that at $r=\epsilon$, all the terms vanish except $J_1$ and $I_5$. Using the fact that $\frac2{p-1}+2=\frac{2p}{p-1}$, it is easy to see that for sufficiently large $\gamma$, $|I_5(\epsilon)|$ dominates $J_1(\epsilon)$, uniformly over all large $R$ and small $\epsilon$. However, when $r$ is small, but on an order larger than $\epsilon$, the analysis becomes a lot more involved.) In the sequel, whenever we say that a condition holds for $\gamma$ or $M$ sufficiently large, or for $l$ sufficiently small, we mean that it holds independent of $R$ and $\epsilon$.
Clearly, $J_5\le |I_4|$ if $\gamma$ is sufficiently large. We now show that for $\gamma$ sufficiently large, $J_2\le |I_4|+|I_5|$, for $\epsilon\le r\le\delta_0$. (We are reusing $|I_4|$ here. Later we will reuse $|I_5|$. This is permissible because $\gamma$ can be chosen as large as we like.) To show this inequality, it suffices to show that for $M$ sufficiently large, $$\label{J_2}
(r-\epsilon)R
\le M(r-\epsilon)^2R^2
+M(1+\frac{\epsilon^l}{r^{l+1}}R^\frac2{p-1})^{p-1},\
\text{for}\ r\in[\epsilon,\delta_0].$$ A trivial calculation shows that the left hand side of is less than the first term on the right hand side if $r\ge\epsilon+\frac1{RM}$. If $r\in[\epsilon,\epsilon+\frac1{RM}]$, then the left hand side of is less than or equal to $\frac1M$ while the second term on the right hand side is greater than $M$. We conclude that holds with $M\ge1$.
Since $I_2$ has the factor $(r-\epsilon)^2$, while $I_1$ has the factor $(r-\epsilon)$, and since $\frac{R-r}{R+\epsilon-2r}$ can be made arbitrarily close to one by choosing $R$ sufficiently large, it follows that for any $\eta>0$, we can guarantee that $I_2\le\eta|I_1|$, for $r\in[\epsilon,\delta_0]$, if we choose $\delta_0$ sufficiently small. Note that $J_4\le\frac {2l}{d-1}|I_1|$. Thus, given any $\zeta>0$, if we choose $\delta_0$ and $l$ sufficiently small, we will have $I_2+J_4\le\zeta |I_1|$.
To complete the proof, we will show that $$\label{J_1J_7I_4I_5}
J_1+J_7+(1-\zeta)I_1+I_4+I_5\le 0, \ \text{for}\ r\in[\epsilon,\delta_0],$$ for sufficiently large $\gamma$ and sufficiently small $l$ and $\delta_0$, uniformly over large $R$ and small $\epsilon$. By the assumption on $\beta$, there exists an $\eta_0>0$ such that if $\delta_0$ is chosen sufficiently small, then $$\label{I_4}
\begin{aligned}
&I_4\le(\beta_0-\eta_0)(1-\frac\epsilon r)^2(R-r)^2
(1+r)^{\frac2{p-1}}(1+\frac{\epsilon^l}{r^l}R^\frac2{p-1}),\\
&\text{for}\ r\in[\epsilon,\delta_0],
\end{aligned}$$ where $\beta_0$ is as in the statement of the theorem. Since $\frac{R+\epsilon-2r}{R-r}$ can be made arbitrarily close to one by choosing $R$ sufficiently large, we also have $$\label{I_1}
I_1\le-(\frac{d-1}{p-1}-\frac{\eta_0}2)(1-\frac\epsilon r) (R-r)^2
(1+r)^{\frac2{p-1}}(1+\frac{\epsilon^l}{r^l}R^\frac2{p-1}),
\ \text{for}\ r\in[\epsilon,\delta_0].$$ Since $$\begin{aligned}
&J_7=\frac{l(l+1)}2((r-\epsilon)(R-r))^2(1+r)^\frac2{p-1}
\frac{\epsilon^l}{r^{l+2}}R^\frac2{p-1}\\
&\le
\frac{l(l+1)}2(R-r)^2(1+r)^\frac2{p-1}
\frac{\epsilon^l}{r^l}R^\frac2{p-1},
\end{aligned}$$ and since $\frac{R-r}{R+\epsilon-2r}$ can be made arbitrarily close to one by choosing $R$ sufficiently large, it follows that for any $\tau>0$, we can choose $l$ sufficiently small so that $$\label{J_7J_1}
J_7\le\tau J_1, \ r\in[\epsilon,\delta_0],$$ uniformly over large $R$ and small $\epsilon$. Using - along with the fact that $$J_1\le\frac{p+1}{(p-1)^2}(R-r)^2(1+r)^\frac2{p-1}
(1+\frac{\epsilon^l}{r^l}R^\frac2{p-1})$$ and that $\beta_0=\frac{d(p-1)-2p}{(p-1)^2}$, it follows that $$\label{J_1J_7I_4}
\begin{aligned}
&J_1+J_7+(1-\zeta)I_1+I_4\le
\Big(\frac{\tau(p+1)+\zeta(d-1)(p-1)}{(p-1)^2}-\eta_0(\frac12+\frac\zeta2)
+C\frac\epsilon r\Big)\times\\
&(R-r)^2(1+r)^\frac2{p-1}
(1+\frac{\epsilon^l}{r^l}R^\frac2{p-1}), \ \text{for}\ r\in[\epsilon,\delta_0],
\end{aligned}$$ for some $C>0$, uniformly over large $R$ and small $\epsilon$. By picking $\tau$ and $\zeta$ sufficiently small, we have $\frac{\tau(p+1)+\zeta(d-1)(p-1)}{(p-1)^2}-\eta_0(\frac12+\frac\zeta2)<0$. Thus, in order to show , it suffices to show that $$\label{final}
\frac\epsilon r R^2\le M(1+\frac{\epsilon^l}{r^l}R^\frac2{p-1})^{p-1},
\ r\in[\epsilon,\delta_0],$$ for sufficiently large $M$. But since $l(p-1)\le1$, the right hand side of is greater or equal to $M\frac\epsilon r R^2$.
**Acknowledgement. The author thanks Robert Adler for a helpful conversation.**
[99]{}
Athreya, K. and Ney, P. *Branching Processes* Springer-Verlag, 1972.
Dawson, D. A. *Measure-valued Markov processes*, École d’Été de Probabilités de Saint-Flour XXI—1991, 1–260, Lecture Notes in Math., **1541**, Springer, Berlin, 1993.
Dawson, D. A., Iscoe, I. and Perkins, E. A. *Super-Brownian motion: path properties and hitting probabilities*, Probab. Theory Related Fields **83** (1989), 135–205.
DeBlassie, R. D. *On hitting single points by a multidimensional diffusion*, Stochastics Stochastics Rep. **65** (1998), 1–11.
Engländer, J. *Criteria for the existence of positive solutions to the equation $\rho\Delta u=u^2$ in $R^d$ for all $d\ge1$– a new probabilistic approach*, Positivity **4** (2000), 327-337.
Engländer, J. and Pinsky, R. *On the construction and support properties of measure-valued diffusions on $D\subset R^d$ with spatially dependent branching*, Ann. of Probab. **27** (1999), 684-730.
Englander, J. and Pinsky, R. *Uniqueness/nonuniqueness for nonnegative solutions of second-order parabolic equations of the form $u\sb t=Lu+Vu-\gamma u\sp p$ in $\bold R\sp n$*, J. Differential Equations **192** (2003), 396–428.
Fleischmann, K. *Critical behavior of some measure-valued processes*, Math. Nachr. **135** (1988), 131-147.
Fleischmann, K. and Sturm, A. *A super-stable motion with infinite mean branching*, Ann. Inst. H. Poincaré Probab. Statist., **40** (2004), 513–537.
Iscoe, I. *On the supports of measure-valued critical branching Brownian motion*, Ann. Probab. **16** (1988), 200–221.
Pazy, A. *Semigroups of linear operators and applications to partial differential equations*, Springer-Verlag, New York, 1983.
Pinsky, R. *Positive Harmonic Functions and Diffusion*, Cambridge University Press, 1995.
Pinsky, R. *Positive Solutions of Reaction diffusion equations with super-linear absorption: universal bounds, uniqueness for the Cauchy problem, boundedness of stationary solutions*, submitted.
Pinsky, R. *Uniqueness/nonuniqueness for nonnegative solutions of the Cauchy problem for $u_t=\Delta u-u^p$ in a punctured space*, submitted.
Ren, Y. *Support properties of super-Brownian motions with spatially dependent branching rate*, Stochastic Process. Appl. **110** (2004), 19–44.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present formulae for computing the Yamada polynomial of spatial graphs obtained by replacing edges of plane graphs, such as cycle-graphs, theta-graphs, and bouquet-graphs, by spatial parts. As a corollary, it is shown that zeros of Yamada polynomials of some series of spatial graphs are dense in a certain region in the complex plane, described by a system of inequalities. Also, the relation between Yamada polynomial of graphs and the chain polynomial of edge-labelled graphs is obtained.'
address:
- 'School of Mathematical Sciences, Dalian University of Technology, Dalian 116024, P. R. China'
- 'School of Mathematical Sciences, Dalian University of Technology, Dalian 116024, P. R. China'
- 'School of Mathematical Sciences, Dalian University of Technology, Dalian 116024, P. R. China'
- 'Novosibirsk State University and Sobolev Institute of Mathematics Novosibirsk, 630090, Russia'
author:
- Miaowang LI
- Fengchun LEI
- Fengling LI
- Andrei VESNIN
title: On Yamada polynomial of spatial graphs obtained by edge replacements
---
[^1]
Introduction
============
Spatial graph theory on intrinsic knotting and linking of graphs in $S^3$ developed in the 1980s since J.H. Conway and C. Gordon [@CC] proved that any embedding of the complete graph $\mathbf K_7$ in $\mathbb R^3$ contains a knotted cycle and any embedding of the complete graph $\mathbf K_{6}$ in $\mathbb R^{3}$ contains a pair of linked cycles. Being motivated by problems on knotting and linking of DNA and chemical compounds, the study of spatial graphs is in the center of interest for last decades. Thus, J. Simon [@S] discussed the chirality of an embedding of the complete graph $\mathbf K_5$, which answered a question raised by D.M. Walba [@W]. C. Ernst and D.W. Sumners [@ES] introduced tangle theory of site specific recombination and gave an approach to study the behavior of DNA. X.-S. Cheng, Y. Lei and W. Yang [@CLY] considered invariants of double crossover links characterize topological properties of double crossover DNA polyhedra.
The modern study of spatial graphs and their generalizations combines topological and graph-theoretical methods. The powerful of polynomial invariants of knot as well as polynomial invariants for graphs was a natural motivation for investigation of polynomial invariants of spatial graphs initiated by L.H. Kauffman [@k]. There are a number of invariants which connect spatial graphs and knots. Y. Ohyama and K. Taniyama [@OT] explored relations among the Vassiliev invariants of knots contained in certain graphs. In 2015, N. Chbili [@NC] gave criteria for a spatial graph to be (p,q)-lens graph. In 2017, A. Henrich and L.H. Kauffman [@HK] provided a topological invariant for pseudoknots and four-valent rigid vertex spatial graphs.
It is well known [@K1; @K2] that the Alexander ideal and Alexander polynomial are invariants of spatial graphs which are determined by the fundamental groups of the complements of spatial graphs. In 1989, S. Yamada [@YA] introduced Yamada polynomial of spatial graphs in $R^3$, which can distinguish some non-isotopy spatial graphs with the same fundamental group. The Yamada polynomial is an concise and useful ambient isotopy invariant for graphs with maximal degree less than four. There are many interesting results on Yamada polynomial and its generalizations. J. Murakami [@M] investigated the two-variable extension $\mathbf{Z}_S$ of the Yamada polynomial and gave an invariant related to the HOMFLY polynomial. In 1994, the crossing number of spatial graphs in terms of the reduced degree of Yamada polynomial has been studied by T. Motohashi, Y. Ohyama and K. Taniyama [@MO]. In 1996, A. Dobrynin and A. Vesnin [@DV] studied properties of the Yamada polynomial of spatial graphs. For any graph $G$, V. Vershinin and A. Vesnin [@VV] defined bigraded cohomology groups whose graded Euler characteristic is a multiple of the Yamada polynomial of $G$. A polynomial invariant of virtual graphs was constructed by Y. Miyazawa [@MY] as an extension of the Yamada polynomial in 2006. Another invariant of spatial graphs associated with $U_{q}(sl(2,C))$ was introduced by S. Yoshinaga [@YO]. See [@Kobe] about the relation between Yamada polynomial and Yoshinaga polynomial. Nice results on the structure of the Yamada and flows polynomials of cubic graphs are established by I. Agol and V. Krushkal [@AK].
Zeros of polynomial invariants of knots and graphs is a question of special interest studied by A.D. Sokal [@SO] and P. Csikvári, P.E. Frenkel, J. Hladký, T. Hubai [@CFHH] for chromatic polynomial; by O.T. Dashbach, T.D. Le, X.-S. Lin [@DLL] and X. Jun, F. Zhang, F. Dong, E.G. Tay [@XF] for Jones polynomial.
This paper consists of three parts. In the first part (Sections 2 and 3) we recall properties of Yamada polynomial of graphs and obtain some formulae for computing the Yamada polynomial of graphs by edge replacements via the chain polynomial (see Theorem \[tger\]). In the second part (Sections 4 and 5) we give formulae for computing the Yamada polynomial of spatial graphs obtained by replacing edges of cycle graphs, theta-graphs, or bouquet graphs by spatial parts (see Theorem \[tcn\]). In the last part, Section 6, we prove that zeros of the Yamada polynomial of spatial graphs are dense in a certain region in the complex plane, described by a system of inequalities (see Theorem \[tmain3\]).
Yamada polynomial of a graph
============================
We consider a graph $G$, admitting loops and multiple edges. Let us use standard notations $p(G)$ and $q(G)$ for number of vertices and number of edges of it.
Before defining the Yamada polynomial of a graph, we recall a graph invariant which is a special case of the Negami polynomial invariant [@N].
[@YA] Define 2-variable Laurent polynomial $h(G) = h(G) (x,y)$ of graph $G=(V,E)$, where $V = V(G)$ is the vertex set and $E = E(G)$ is the edge set of $G$, by the rule $$h(G)(x,y) = \sum_{F\subset E}(-x)^{-|F|} f(G-F),$$ with $f(G)=x^{\mu(G)}y^{\beta(G)}$, where $\mu(G)$ and $\beta(G)$ is the number of connected components of $G$ and the first Betti number of $G$.
[@YA] [*The Yamada polynomial*]{} of a graph $G$ is a Laurent polynomial in one variable, obtained by the following substitution into $h(G)(x,y)$: $$H(G)(A) = h(G)(-1,-A-2-A^{-1}).$$
The following properties of $H(x,y)$ hold (see [@YA] for details):
- $H(\cdot)= -1$.
- Let $e$ be a non-loop edge of a graph $G$. Then $H(G)=H(G/e)+H(G-e)$, where $G/e$ is the graph obtained by contracting the edge $e$, and $G-e$ is the graph obtained by deleting of the edge $e$.
- Let $e$ be a loop edge of a graph $G$. Then $H(G)=-\sigma H(G-e)$, where $\sigma= A+1+A^{-1}$.
- Let $G_{1}\cup G_{2}$ be a disjoint union of graphs $G_{1}$and $G_{2}$, then $H(G_{1}\cup G_{2})= H(G_{1})H(G_{2})$.
- Let $G_{1}\cdot G_{2}$ be a union of graphs $G_{1}$ and $G_{2}$ having one common point, then $$H(G_{1}\cdot G_{2})= -H(G_{1})H(G_{2}).$$
- If $G$ has an isthmus, then $H(G)= 0$.
It is easy to find directly (see also [@DV]) polynomial $H(G)$ for some simple classes of graphs.
\[lemma-dv\] The following properties holds with $\sigma = A + 1 + A^{-1}$.
- Let $T_q$ be a tree with $q$ edges. Then $H(T_q)=0$ for all $q$.
- Let $C_{n}$ be the cycle of length $n$. Then $H(C_{n})= \sigma$ for all $n$.
- Let $B_q$ be the one-vertex graph with $q$ loops, also known as “$q$-bouquet”. Then $H(B_q) \, = \,(-1)^{q-1} \sigma^q$.
- Let $\Theta_s$ be the graph consisting of two vertices and $s$ edges between them, also known as “$s$-theta-graph” (see Fig. \[fig1\]). Then $$H(\Theta_s) \, = \, \frac{1}{\sigma+1} [\sigma +(-\sigma)^s].$$
\[ptv\] Let $G_{1}:G_{2}$ be the union of two graphs $G_{1}$ and $G_{2}$ having only two common vertices $u$ and $v$. Let $K_{1}$ and $K_{2}$ be graphs obtained from $G_{1}$ and $G_{2}$, respectively, by identifying $u$ and $v$. Then $$\begin{aligned}
\label{twovert}
H(G_{1}:G_{2}) & = & \frac{1}{\sigma} \Big[ H(K_{1}) H(K_{2}) + (\sigma+1) H(G_{1}) H(G_{2}) \cr
& & \qquad + H(K_{1}) H(G_{2}) + H(K_{2}) H(G_{1}) \Big].\end{aligned}$$
Denote $G = G_{1}:G_{2}$ and prove the statement using induction by $q(G)$, the number of edges of the graph $G$.
If $q(G)=0$, then $G_{1}$ and $G_{2}$ consist of isolated vertices. Suppose that $p(G_{1}) =p_{1}\geq 2$ and $p (G_{2}) = p_{2}\geq 2$. Then $G$ consists of $p_{1}+p_{2}-2$ isolated vertices, that is a disjoint union of $p_{1}+p_{2}-2$ one-vertex graphs. By properties (1) and (4), $H(G) = (-1)^{p_{1}+p_{2}-2}$. Since $p(K_{1})=p_{1}-1$ and $p(K_{2}) = p_{2}-1$, for the right-hand side of Eq.(\[twovert\]) we get: $$\frac{1}{\sigma} \Big[ (-1)^{p_{1} + p_{2}-2} + (\sigma + 1) (-1)^{p_{1} + p_{2}} + (-1)^{p_{1} -1 + p_{2}} + (-1)^{p_{2}-1+p_{1}} \Big] = (-1)^{p_{1}+p_{2}}.$$ Hence, the equality Eq.(\[twovert\]) holds.
Now suppose that the statement holds for all graphs which $q(G) < k$, where $k\geq 1$. Consider $G=G_{1}:G_{2}$ such that $q(G) = k$. Let $e$ be an edge of $G$. Without loss of generality, we can suppose $e \in E(G_{1})$. There are three cases to be considered.
Let $e$ be a loop of $G_{1}$. Then $e$ is also a loop of $K_{1}$. Hence, we have $$\begin{aligned}
H(G_{1}:G_{2}) & = & -\sigma H(G_{1} - e : G_{2}) \\
& = & - \Big[ H(K_{1} - e) H(K_{2}) \ + (\sigma+1) H(G_{1} - e) H(G_{2}) \\
& & \quad + H(K_{1} -e) H(G_{2}) + H(K_{2})H(G_{1}-e) \Big] \\
&=& \frac{1}{\sigma} \Big[ H(K_{1}) H(K_{2}) + (\sigma+1)H(G_{1}) H(G_{2}) \\
& & \quad + H(K_{1}) H(G_{2}) + H(K_{2}) H(G_{1}) \Big].\end{aligned}$$
Let $e$ be neither a loop nor an edge with two end vertices $u$ and $v$, where $u$ and $v$ are common vertices of $G_1$ and $G_2$. Then $e$ is a non-loop edge in $K_{1}$. Hence, $$\begin{aligned}
H(G/e) & = & H(G_{1}/e : G_{2}) \\
& = & \frac{1}{\sigma} \Big[ H(K_{1}/e )H(K_{2}) + (\sigma+1) H((G_{1}/e) H(G_{2}) \\
& & \qquad + H(K_{1}/e) H(K_{2}) + H(K_{2}) H(G_{1}/e) \Big]\end{aligned}$$ and $$\begin{aligned}
H(G-e) &= & H(G_{1}-e:G_{2}) \\
& = &\frac{1}{\sigma} \Big[ H(K_{1}-e) H(K_{2}) + (\sigma+1) H(G_{1}-e) H(G_{2}) \\
& & \qquad + H(K_{1}-e) H(G_{2}) + H(K_{2}) H(G_{1}-e) \Big].\end{aligned}$$ Therefore, $$\begin{aligned}
H(G)&= & H(G/e) + H(G-e) \\
&= &\frac{1}{\sigma} \Big[ H(K_{1}/e) H(K_{2}) + (\sigma+1) H(G_{1}/e) H(G_{2}) \\
& &\quad + H(K_{1}/e)H(G_{2}) + H(K_{2})H(G_{1}/e) + H(K_{1}-e) H(K_{2}) \\
& & \quad + (\sigma+1) H(G_{1}-e) H(G_{2})+ H(K_{1}-e) H(G_{2}) \\
& & \quad +H(K_{2}) H(G_{1}-e) \Big] .\end{aligned}$$ Hence $$\begin{aligned}
H(G) &=& \frac{1}{\sigma} \Big[ \Big( H(K_{1}/e) + H(K_{1}-e) \Big) H(K_{2}) \\
& & \quad + (\sigma+1) \Big( H((G_{1}/e) + H(G_{1}-e) \Big) H(G_{2}) \\
& & \quad + \Big( H(K_{1}/e) + H(K_{1}-e) \Big) H(G_{2}) \\
& & \quad + H(K_{2}) \Big( H(G_{1}/e)+H(G_{1}-e) \Big) \Big] \\
&=& \frac{1}{\sigma} \Big[ H(K_{1})H(K_{2})+(\sigma+1)H(G_{1})H(G_{2})+H(K_{1})H(G_{2}) \\
& & \quad +H(K_{2})H(G_{1}) \Big] .\end{aligned}$$
Let $e$ be an edge with two end vertices $u$ and $v$, where $u$ and $v$ are common vertices of $G_1$ and $G_2$. Then $e$ is a loop in $K_{1}$. Hence, by properties (5) and (3) we have $$H(G/e) = H(G_{1}/e:G_{2}) = -H(K_{1}-e)H(K_{2}) = \frac{1}{\sigma}H(K_{1})H(K_{2})$$ and $$\begin{aligned}
H(G-e) & =& H(G_{1}-e:G_{2}) \\
& = &\frac{1}{\sigma} \Big[ H(K_{1}-e) H(K_{2}) + (\sigma+1) H(G_{1}-e) H(G_{2}) \\
& & \qquad + H(K_{1}-e) H(G_{2}) + H(K_{2}) H(G_{1}-e) \Big] \\
&=&\frac{1}{\sigma} \Big[ -\frac{1}{\sigma} H(K_{1}) H(K_{2}) + (\sigma+1) \Big( H(G_{1})+\frac{1}{\sigma}H(K_{1}) \Big) H(G_{2}) \\
& & \qquad -\frac{1}{\sigma} H(K_{1}) H(G_{2}) + H(K_{2}) \Big( H(G_{1}) + \frac{1}{\sigma} H(K_{1}) \Big) \Big] .\end{aligned}$$ Therefore, $$\begin{aligned}
H(G) &= & H(G/e) + H(G-e) = \frac{1}{\sigma} H(K_{1}) H(K_{2}) \\ & & + \frac{1}{\sigma} \Big[ - \frac{1}{\sigma} H(K_{1}) H(K_{2}) + (\sigma+1) H(G_{2}) \Big( H(G_{1}) + \frac{1}{\sigma} H(K_{1}) \Big) \\
& & \qquad -\frac{1}{\sigma} H(K_{1}) H(G_{2}) + H(K_{2}) \Big( H(G_{1}) + \frac{1}{\sigma} H(K_{1}) \Big) \Big] \\
&=& \frac{1}{\sigma} \Big[ H(K_{1}) H(K_{2}) + (\sigma+1) H(G_{1}) H(G_{2}) + H(K_{1}) H(G_{2})\\
& & \quad + H(K_{2}) H(G_{1}) \Big] .\end{aligned}$$ This completes the proof.
Yamada polynomials of a graph obtained by edge replacements
===========================================================
We recall some properties of the chain polynomial introduced by R.C. Read and E.G. Whitehead [@RE], see also [@XF]. The chain polynomial is defined on edge-labelled graphs with labels are elements of a commutative ring with unity. We will denote the edges by the labels associated with them.
[*The chain polynomial*]{} $\operatorname{Ch}(G)$ of a labelled graph $G$ is defined as $$\operatorname{Ch}(G) = \sum_{Y \subset E} F_{G-Y} (1-w) \prod_{a \in Y} a,$$ where the sum is taken over all subsets of the edge set $E$ of $G$, $F_{G-Y} (1-w)$ denotes the flow polynomial of the subgraph $G - Y$ calculated at $1-w$, and $\prod_{a\in Y}$ denotes the product of edge-labels of $Y$.
The chain polynomial can be also defined in the following recursive form.
\[def3.2\] [*The chain polynomial*]{} $\operatorname{Ch} (G) (w)$ in a variable $w$ of a labelled graph $G$ is defined by following rules.
- If $G$ is edgeless, then $\operatorname{Ch}(G)=1$.
- Otherwise, suppose $a$ is an edge of $G$ labelled by $a$. Then
- If $a$ is a loop, then $\operatorname{Ch}(G) = (a-w) \operatorname{Ch}(G-a)$.
- If $a$ is not a loop, then $\operatorname{Ch}(G) = (a-1) \operatorname{Ch} (G-a) + \operatorname{Ch}(G/a)$.
\[Ecn\] Let $C_{n}$ be the $n$-cycle with edges labelled by $a_{1}, a_{2}, \cdots, a_{n}$, then $$\operatorname{Ch}(C_{n}) = \prod_{i=1}^n a_i \, - \, w.$$
Let $\Theta_{s}$ be the $s$-theta-graph with edges labelled by $a_{1}, a_{2}, \cdots, a_{s}$, then $$\operatorname{Ch} (\Theta_{s}) = \frac{1}{1-w} \Big[ \prod^{s}_{i=1}(a_{i}-w) - w \prod ^s_{i=1}(a_{i}-1) \Big].$$
Let $B_{q}$ be the $q$-bouquet with $q$ loops labelled by $a_{1}, a_{2}, \ldots, a_{q}$, then $$\operatorname{Ch} ( B_{q}) = \prod^{И}_{i=1}(a_{i}-w).$$
To explore the relation between the chain polynomial and the Yamada polynomial, inspired by [@XF], let us give the following notation. Let $G$ be a connected labelled graph. Denote by $\widetilde{G}$ the graph obtained from $G$ by replacing each edge $a= uv$ of $G$ by a connected graph $K_{a}$ with two attached vertices $u$ and $v$ that has only the vertices $u$ and $v$ in common with $\widetilde{G-a}$.
Let $K'_{a}$ be the graph obtained from $K_{a}$ by identifying $u$ and $v$, the two attached vertices of $K_{a}$. Denote $$\alpha_{a }= \alpha(K_{a}) := \frac{1}{\sigma}[(\sigma+1)H(K_{a})+H(K'_{a})],$$ $$\beta_{a} = \beta(K_{a}) := \frac{1}{\sigma}[H(K_{a})+H(K'_{a})],$$ and $$\gamma_{a} = \gamma(K_{a}) : =1-\frac{\alpha (K_{a})}{\beta (K_{a})}.$$ It is easy to see that $$H(K'_{a})=(\sigma+1)\beta_{a}-\alpha_{a} \qquad \text{and} \qquad H(K_{a})=\alpha_{a}-\beta_{a}.$$
The following result gives the relation between Yamada polynomial and chain polynomial.
\[tger\] Let $G$ be a connected labelled graph, and $\widetilde{G}$ be the graph obtained from $G$ by replacing the edge $a$ by a connected graph $K_{a}$ for every edge $a$ of $G$. If we replace $w$ by $-\sigma$, and replace $a$ by $\gamma_{a}$ for every label $a$ in $\operatorname{Ch}(G)$, then we get $$H(\widetilde{G})=\frac{\prod_{a\in E(G)}\beta_{a}}{(-1)^{q(G)-p(G)}} \operatorname{Ch}(G),$$ where $p(G)$ and $q(G)$ are the numbers of vertices and edges of $G$, respectively.
Denote $$T(G) = \frac{(-1)^{q(G)-p(G)} H(\widetilde{G})}{\prod_{a\in E(G)}\beta_{a}}.$$ To prove the statement we will show that $T(G)$ satisfies conditions from Definition 3.2 of $\operatorname{Ch}(G)$ if we identify $w$ with $-\sigma$ and $a$ with $\gamma_{a}$. Let us check these conditions.
\(1) If $G$ is edgeless, then $E(G) = \emptyset$, $E(\widetilde{G}) = \emptyset$ and $H(\widetilde{G}) = (-1)^{p(\widetilde{G})} = (-1)^{p(G)}$. Then $$T(G)=\frac{(-1)^{q(G)-p(G)} H(\widetilde{G}) }{\prod_{a\in E(G)}\beta_{a}}= (-1)^{-p(G)} (-1)^{p(G)} = 1.$$
(2a) If $a$ is a loop of $G$, then $$\begin{aligned}
T(G) & = & \frac{(-1)^{q(G)-p(G)} H(\widetilde{G})}{\prod_{a\in E(G)}\beta_{a}} \\
& = & \frac{(-1)^{q(G-a)+1-p(G-a)} \Big[-H(K_{a'}) H(\widetilde{G-a})\Big]}{\beta_{a}\prod_{a\in E(G-a)}\beta_{a}} \\
& = & \frac{(-1)^{q(G-a)+1-p(G-a)} \Big[ \Big( \alpha_{a} - (\sigma+1) \beta_{a} \Big) H(\widetilde{G-a}) \Big]}{\beta_{a} \prod_{a\in E(G-a)}\beta_{a}} \\
&=& \frac{- \Big( \alpha_{a}-(\sigma+1)\beta_{a}) \Big)}{\beta_{a}}T(G-a) =(\gamma_{a}+\sigma)T(G-a).\end{aligned}$$ Compare with $\operatorname{Ch}(G) = (a - w) \operatorname{Ch} (G-a)$ from Definition 3.2.
(2b) If $a \in E(G)$ is not a loop, then by Proposition \[ptv\] $$\begin{aligned}
H(\widetilde{G}) &=& H(\widetilde{G-a}:K_{a}) \\
& = & \frac{1}{\sigma} \Big[ H(\widetilde{G/a}) H(K'_{a} )+ (\sigma+1) H(\widetilde{G-a}) H(K_{a}) \\
& & \qquad \qquad \qquad \quad + H(K'_{a}) H(\widetilde{G-a}) + H(K_{a}) H(\widetilde{G/a})\Big] \\
&= & \frac{1}{\sigma} \Big[ H(\widetilde{G/a}) \Big( H(K'_{a})+H(K_{a}) \Big) \\
&& \qquad \qquad \qquad \quad+ H(\widetilde{G-a})\Big( (\sigma+1)H(K_{a})+H(K'_{a}) \Big) \Big] \\
&=& \beta_{a}H(\widetilde{G/a}) + \alpha_{a}H(\widetilde{G-a}).\end{aligned}$$ Therefore, $$\begin{aligned}
T(G) & = & \frac{(-1)^{q(G)-p(G)}H(\widetilde{G})}{\prod_{a\in E(G)}\beta_{a}} \\
& = & \frac{(-1)^{q(G-a)+1-p(G-a)} \Big[ \beta_{a} H(\widetilde{G/a}) + \alpha_{a} H(\widetilde{G-a}) \Big]}{\beta_{a}\prod_{a\in E(G-a)}\beta_{a}}\\
&=&\frac{(-1)^{q(G/a)+1-p(G/a)-1} \beta_{a} H(\widetilde{G/a}) }{\beta_{a}\prod_{a\in E(G/a)}\beta_{a}} \\
& & \quad +
\frac{(-1)^{q(G-a)+1-p(G-a)} \alpha_{a}H(\widetilde{G-a}) }{\beta_{a}\prod_{a\in E(G-a)}\beta_{a}}\\
&=&T(G/a) + \Big(-\frac{\alpha_{a}}{\beta_{a}} \Big) T(G-a) .\end{aligned}$$ Thus, $$T(G) = T(G/a) +(\gamma_{a}-1) T(G-a).$$ Compare with $\operatorname{Ch} (G) = \operatorname{Ch} (G/a) + (a-1) \operatorname{Ch} (G-a)$ from Definition 3.2.
Yamada polynomial of spatial graphs
===================================
Next we will consider the Yamada polynomial of spatial graphs. Through the paper we work in the piecewise-linear category. Let $G$ be a graph embedded in $\mathbb R^3$, and $g$ be a diagram of $G$. For any double point, S. Yamada [@YA] defined the spin of $+1$, $-1$ and $0$, which are denoted by $S_+$, $S_-$ and $S_0$, as shown in Fig. \[fig2\].
Let $S$ be the plane graph obtained from $g$ by replacing each double point with a spin. $S$ is called [*a state*]{} on $g$. Denote the set of all states by $U(g)$. Put $$c(g|S)= A^{m_1-m_2},$$ where $m_1$ and $m_2$ are the numbers of double points with spin $S_+$ and $S_-$, respectively, used to obtain $S$ from $g$.
[@YA]\[dsg\] [*The Yamada polynomial*]{} of a diagram $g$ of a spatial graph $G$ is a Laurent polynomial in variable $A$ defined as follows $$R[g] = R[g](A) : =\sum_{S\in U(g)} c(g|S)H(S),$$ where $H(S) = h(S)(-1,-A-2-A^{-1})$.
For interesting properties of the Yamada polynomial and its invariantness under transformations of spatial graph diagrams see [@YA]. Here we recall only some of them.
[@YA]\[pr\] The following properties hold.
- Let $g_{1}\cup g_{2}$ be a disjoint union of diagrams $g_{1}$ and $g_{2}$, then $$R[g_{1}\cup g_{2}]= R[g_{1}]R[g_{2}].$$
- Let $g_{1}\cdot g_{2}$ be a union of diagrams $g_{1}$ and $g_{2}$ having one common point, then $$R[g_{1}\cdot g_{2}] = -R[g_{1}]R[g_{2}].$$
- If $g$ has an isthmus, then $R[g]= 0$.
If a diagram $g$ of $G$ does not have double points, then $R[g]= H(G)$.
Proposition \[ptv\] about $H(G)$ implies the similar formula for polynomial $R[g]$.
\[cdtv\] Let $g_{1}:g_{2}$ be the union of two diagrams $g_{1}$ and $g_{2}$ having only two common vertices $u$ and $v$. Let $k_{1}$ and $k_{2}$ be diagrams obtained from $g_{1}$ and $g_{2}$, respectively, by identifying $u$ and $v$. Then $$\begin{gathered}
R[g_{1}:g_{2}] = \frac{1}{\sigma} \Big( R[k_{1}] R[k_{2}] + (\sigma+1) R[g_{1}] R[g_{2}] \\ + R[k_{1}] R[g_{2}] + R[k_{2}] R[g_{1}] \Big).
\end{gathered}$$
By Definition \[dsg\] and Proposition \[ptv\], we obtain $$\begin{aligned}
R[g_{1}:g_{2}] & = & \sum_{S\in U(g_{1}:g_{2})} c(g_{1}:g_{2}|S) H(S) \\
&=& \sum_{S_{1}\in U(g_{1})} \sum_{S_{2}\in U(g_{2})} c(g_{1}|S_1) c(g_{2}|S_2) H(S_{1}:S_2) \\
&=&\sum_{S_{1}\in U(g_{1})} \sum_{S_{2}\in U(g_{2})} c(g_{1}|S_1) c(g_{2}|S_2) \frac{1}{\sigma} \Big[ H(S_{1}')H(S_{2}') \\
& & \quad + (\sigma+1) H(S_{1}) H(S_{2}) +H(S_{1})H(S_{2}')+H(S_{1}')H(S_{2}) \Big]\\
&=&\frac{1}{\sigma} \Big( \sum_{S_{1} \in U(g_{1})} \Big[ c(g_{1}|S_1) H(S_{1}') \Big] \sum_{S_{2}\in U(g_{2})} \Big[ c(g_{2}|S_2) H(S_{2}') \Big] \\
& & + (\sigma+1) \sum_{S_{1} \in U(g_{1})} \Big[ c(g_{1}|S_1) H(S_{1}) \Big] \sum_{S_{2}\in U(g_{2})} \Big[ c(g_{2}|S_2) H(S_{2}) \Big] \\
&& + \sum_{S_{1}\in U(g_{1})} \Big[ c(g_{1}|S_1)H(S_{1}') \Big] \sum_{S_{2}\in U(g_{2})} \Big[ c(g_{2}|S_2) H(S_{2}) \Big] \\
& & +\sum_{S_{1}\in U(g_{1})} \Big[ c(g_{1}|S_1) H(S_{1}) \Big] \sum_{S_{2} \in U(g_{2})} \Big[c(g_{2}|S_2) H(S_{2}') \Big] \Big) \\
&=&\frac{1}{\sigma} \Big[ R[k_{1}]R[k_{2}] + (\sigma+1) R[g_{1}] R[g_{2}] + R[k_{1}] R[g_{2}] \\ & & \qquad \qquad + R[k_{2}] R[g_{1}] \Big].\end{aligned}$$ This completes the proof.
Yamada Polynomials of diagrams obtained by edge replacements
============================================================
Let $G$ be a connected plane labelled graph, where each edge $a$ with terminal vertices $u$ and $v$ is labelled by a diagram $g_{a}$ having two vertices $u_{a*}$ and $v_{a*}$ indicated. We define $G(g_{a})$ to be the spatial graph (its diagram) obtained from $G$ by replacing an edge $a= u_{a}v_{a}$ of $G$ by a connected diagram $g_{a}$ with identification $u_{a}$ with $u_{a*}$ and $v_{a}$ with $v_{a*}$ in such a way that $g_{a}$ has only vertices $u_{a}=u_{a*}$ and $v_{a}=v_{a*}$ in common with $G-a$. We define $G(g_{a}, g_{b})$ to be the spatial graph (its diagram) obtained from $G(g_{a})$ by replacing an edge $b= u_{b}v_{b}$ of $G(g_{a})\cap (G-a)$ by a connected diagram $g_{b}$ with identification $u_{b}$ with $u_{b*}$ and $v_{b}$ with $v_{b*}$ in such a way that $g_{b}$ has only vertices $u_{b}=u_{b*}$ and $v_{b}=v_{b*}$ in common with $G(g_{a})$. With the same construction, we can obtain $G(g_{a_1}, g_{a_2}, g_{a_3}, \ldots)$. In this context we denote by $g_{a}'$ the diagram obtained from $g_{a}$ by identifying vertices $u_{a*}$ and $v_{a*}$ in such a way that no new intersections appear.
\[tcn\] The following properties hold.
- Let $C_{n}$ be the $n$-cycle graph with edges labelled by $a_{1}, a_{2}, \ldots, a_{n}$, then $$\label{ecn}
R[C_{n}(g_{a_1}, g_{a_2}, \cdots, g_{a_n})] = \prod_{i=1}^n(-R[g_{a_i}])+\sigma\prod_{i=1}^n \Big( \frac{R[g_{a_i}]+R[g_{a_i}']}{\sigma} \Big).$$
- Let $\Theta_{s}$ be the $s$-theta-graph with edges labelled by $a_{1}, a_{2}, \ldots, a_{s}$, then $$\label{eths}
\begin{gathered}
R[\Theta_{s}(g_{a_1}, g_{a_2}, \cdots, g_{a_s})] \quad \qquad \qquad \qquad \qquad \qquad \qquad \qquad \qquad \\
= \frac{(-1)^{s}}{1+\sigma} \prod_{i=1}^sR[g_{a_i}'] + \frac{\sigma}{1+\sigma} \prod_{i=1}^s \Big(\frac{\sigma+1}{\sigma}R[g_{a_i}]+\frac{1}{\sigma}R[g_{a_i}'] \Big).
\end{gathered}$$
- Let $B_{q}$ be the $q$-bouquet with loops labelled by $a_{1}, a_{2}, \ldots, a_{q}$, then $$\label{ebp}
R[B_{q}(g_{a_1}, g_{a_2}, \cdots, g_{a_q})] = (-1)^{q-1}\prod_{i=1}^q R[g_{a_i}'].$$
\(1) Firstly, we prove Eq.(\[ecn\]) by induction on $n$. Recall, that by Example \[Ecn\], the chain polynomial of the $n$-cycle graph is $$\operatorname{Ch} (C_n) = \prod_{i=1}^n a_i-w.$$
If $n=1$, then $R[C_1(g_{a_1})]=R[g_{a_1}']$, so Eq.(\[ecn\]) holds.
Suppose Eq.(\[ecn\]) holds for all $n \leq k-1$, where $k \geq 2$. Observe that spatial graph $C_{k}(g_{a_1}, g_{a_2}, \ldots, g_{a_k})$ can be presented as the union $g_{a_k} : L_{k-1}(g_{a_1}, g_{a_2}, \ldots, g_{a_{k-1}})$, where $L_{k-1}$ is a graph with $k$ vertices connected by $k-1$ edges one by one, as shown in Fig. \[fig3\].
Then $$R[C_{k}(g_{a_1}, g_{a_2}, \cdots, g_{a_k})]=R[g_{a_k} : L_{k-1}(g_{a_1}, g_{a_2}, \cdots, g_{a_{k-1}})].$$ By Proposition \[cdtv\], we have $$\begin{aligned}
& & R[g_{a_k} : L_{k-1}(g_{a_1}, g_{a_2}, \cdots, g_{a_{k-1}})]\\
& & \qquad \qquad \qquad = \frac{1}{\sigma} \Big[ R[L_{k-1}(g_{a_1}, g_{a_2}, \cdots, g_{a_{k-1}})']R[g_{a_k}'] \\
& & \qquad \qquad \qquad \qquad + (1+\sigma)R[L_{k-1}(g_{a_1}, g_{a_2}, \cdots, g_{a_{k-1}})]R[g_{a_k}]\\
& & \qquad \qquad \qquad \qquad + R[L_{k-1}(g_{a_1}, g_{a_2}, \cdots, g_{a_{k-1}})']R[g_{a_k}] \\
& & \qquad \qquad \qquad \qquad +R[L_{k-1}(g_{a_1}, g_{a_2}, \cdots, g_{a_{k-1}})]R[g_{a_k}'] \Big].\end{aligned}$$
Applying Proposition \[pr\], we get $$R[L_{k-1}(g_{a_1}, g_{a_2}, \cdots, g_{a_{k-1}})] = (-1)^{k}\prod_{i=1}^{k-1}R[g_{a_i}]$$ and $$\begin{gathered}
R[L_{k-1}(g_{a_1}, g_{a_2}, \cdots, g_{a_{k-1}})'] = R[C_{k}(g_{a_1}, g_{a_2}, \cdots, g_{a_{k-1}})] \\ =\prod_{i=1}^{k-1} \Big( -R[g_{a_i}] \Big) + \sigma \prod_{i=1}^{k-1} \Big( \frac{R[g_{a_i}] + R[g_{a_i}']}{\sigma} \Big).
\end{gathered}$$ Therefore, $$\begin{aligned}
&& R[C_{k}(g_{a_1}, g_{a_2}, \cdots, g_{a_k})]\\
& & = \frac{1}{\sigma} \Big[ \Big( \prod_{i=1}^{k-1}(-R[g_{a_i}]) + \sigma \prod_{i=1}^{k-1}\frac{R[g_{a_i}]+R[g_{a_i}']}{\sigma} \Big) R[g_{a_k}'] \\
& & \qquad \qquad +(1+\sigma)(-1)^{k} \Big( \prod_{i=1}^{k-1} R[g_{a_i}] \Big) R[g_{a_k}] \\
& & \qquad \qquad + \Big( \prod_{i=1}^{k-1}(-R[g_{a_i}]) + \sigma \prod_{i=1}^{k-1}\frac{R[g_{a_i}] + R[g_{a_i}']}{\sigma} \Big) R[g_{a_k}] \\
& & \qquad \qquad +(-1)^{k} \Big( \prod_{i=1}^{k-1}R[g_{a_i}] \Big) R[g_{a_k}'] \Big] \\
& & = \prod_{i=1}^k (-R[g_{a_i}]) + \sigma\prod_{i=1}^k \Big( \frac{R[g_{a_i}] +R[g_{a_i}']}{\sigma} \Big).\end{aligned}$$
Thus, Eq.(\[ecn\]) holds.
\(3) Similarly to the above case, Eq.(\[ebp\]) is equivalent to $$\begin{gathered}
R[B_{q}(g_{a_1}, g_{a_2}, \ldots, g_{a_q})] \qquad \qquad \qquad \qquad \qquad \qquad \qquad \qquad \\
=R[g_{a_1}] \, R[g_{a_2}] \cdots R[g_{a_{q-1}}] \, R[g_{a_q}]= (-1)^{q-1}\prod_{i=1}^qR[g_{a_i}'].
\end{gathered}$$
\(2) Now, we prove Eq.(\[eths\]) by induction on $s$. If $s=1$, then $R[\Theta_{1}(g_{a_1})]= R[g_{a_1}]$ and Eq.(\[eths\]) holds.
Suppose that Eq.(\[eths\]) holds for all $s \leq k-1$, where $k\geq 2$. Observe that the left-hand side of Eq.(\[eths\]) can be presented as the union $R[\Theta_{s}(g_{a_1}, g_{a_2}, \cdots, g_{a_{s-1}}): g_{a_s}]$.
Since $$R[\Theta_{s}(g_{a_1}, g_{a_2}, \ldots, g_{a_s})'] = R[B_{s}(g_{a_1}, g_{a_2}, \ldots, g_{a_s})],$$ by Proposition \[cdtv\] we have $$\begin{aligned}
\begin{gathered}
R[\Theta_{s}(g_{a_1}, g_{a_2}, \ldots, g_{a_{s}})] \qquad \qquad \qquad \qquad \qquad \qquad \qquad \\
= \frac{1}{\sigma} \Big( R[\Theta_{s}(g_{a_1}, g_{a_2}, \ldots, g_{a_{s-1}})']R[g_{a_s}'] \\
\qquad + (1+\sigma) R[\Theta_{s}(g_{a_1}, g_{a_2}, \ldots, g_{a_{s-1}})]R[g_{a_s}] \\
\qquad +R[\Theta_{s}(g_{a_1}, g_{a_2}, \ldots, g_{a_{s-1}})'] R[g_{a_s}]
+ R[\Theta_{s}(g_{a_1}, g_{a_2}, \ldots, g_{a_{s-1}})]R[g_{a_s}'] \Big) \\
= \frac{(-1)^{s}}{1+\sigma}\prod_{i=1}^sR[g_{a_i}'] + \frac{\sigma}{1+\sigma}\prod_{i=1}^s \Big( \frac{\sigma+1}{\sigma}R[g_{a_i}] + \frac{1}{\sigma}R[g_{a_i}'] \Big).
\end{gathered}\end{aligned}$$ The proof is competed.
If every edge $a$ of $G$ is labelled by the same diagram $g_{a} = g$, we denote $G(g, \ldots, g)$ shortly by $G(g)$.
If all edges of a graph $G$ are labelled by the same diagram $g$, then $$R[C_{n}(g)]=(-R[g])^{n}+\sigma \Big( \frac{R[g]+R[g']}{\sigma} \Big)^{n},$$ $$R[\Theta_{s}(g)] = \frac{(-1)^{s}}{1+\sigma}R[g']^{s} + \frac{\sigma}{1+\sigma} \Big( \frac{\sigma+1}{\sigma}R[g]+\frac{1}{\sigma}R[g'] \Big)^{s},$$ and $$R[B_{q}(g)] = (-1)^{q-1} R[g']^q.$$
Remark that $C_{n}(g_{a_1}, g_{a_2}, \ldots, g_{a_n})$ is the “ring of beads” discussed in [@xj] and [@RE]. By Eq.(\[ecn\]), we get the following property.
The Yamada polynomial $R[C_{n}(g_{a_1}, g_{a_2}, \ldots, g_{a_n})]$ of a “ring of beads” graph $C_{n}(g_{a_1}, g_{a_2}, \ldots, g_{a_n})$ is independent of the order in which the beads occur.
The following example illustrates the applying of Theorem \[tcn\].
\[E4\] Let $\infty_{+}$ be the spatial graph diagram with two vertices and one double point signed by “$+$”, see Fig. 2. We denote the mirror image of diagram $\infty_{+}$ by $\infty_{-}$. Direct calculations give Yamada polynomials $R[\infty_{+}]=A^{-2}\sigma$ and $R[\infty_{+}']=\sigma$, where $\sigma = A + 1 + A^{-1}$. By Theorem \[tcn\], the Yamada polynomial of $C_{n}(\infty_{+})$ is as follows: $$R[C_{n}(\infty_{+})]=(-A^{-2}\sigma)^{n}+\sigma(A^{-2}+1)^{n};$$ the Yamada polynomial of $\Theta_{s}(\infty_{+})$ is as follows: $$R[\Theta_{n}(\infty_{+})]=\frac{(-\sigma)^s+\sigma\big((\sigma+1)A^{-2}+1\big)^s}{1+\sigma};$$ and the Yamada polynomial of $B_{q}(\infty_{+})$ is as follows: $$R[B_{q}(\infty_{+})]=(-1)^{q-1} \sigma^q.$$
Zeros of Yamada polynomial
==========================
In this section, we investigate the density of zeros of Yamada polynomial of two classes of spatial graphs. One is the “ring of beads” graphs $C_{n}(\Theta_{s})$ obtained from $C_{n}$ by replacing the edge $a$ by a connected plane $s$-theta graph $\Theta_{s}$ for every edge $a$ of $C_{n}$. The other is the class of spatial graphs $C_{n}(\Theta_{s}(\infty_{+}))$ and their mirror images $C_{n}(\Theta_{s}(\infty_{-}))$.
To study zeros of families of polynomials we will base on the following results by S. Beraha, J. Kahane, N.J. Weiss [@BJ] and by A.D. Sokal [@SO].
\[tbj\] [@BJ] If $\{f_n(x)\}$ is a family of polynomials such that $$f_n(x) = \alpha_1(x) \lambda_1(x)^n + \alpha_2(x) \lambda_2(x)^n +\ldots + \alpha_l(x) \lambda_l(x)^n,$$ where the $\alpha_i(x)$ and $\lambda_i(x)$ are fixed non-zero polynomials, such that no pair $i\neq j$ has $\lambda_i(x) \equiv \omega \lambda_j(x)$ for some complex number $\omega$ of unit modulus. Then $z$ is a limit of zeros of $\{f_n(x)\}$ if and only if
- two or more of the $\lambda_i(z)$ are of equal modulus, and strictly greater in modulus than the others; or
- for some j, the modulus of $\lambda_j(z)$ is strictly greater than those of the others, and $\alpha_i(z)=0$.
[@SO Lemma 1.6] \[lso\] Let $F_{1}$, $F_{2}$, and $G$ be analytic functions on a disc $\{ z \in \mathbb C : |z|<R \}$ such that $|G(0)|\leq 1$ and G not constant. Then, for each $\epsilon >0$ there exists $s_{0}<\infty$ such that for all integers $s\geq s_{0}$ the equation $$|1+F_{1}(z)G(z)^s|=|1+F_{2}(z)G(z)^s|$$ has a solution in the disc $|z|<\epsilon$.
Now, let us turn to families of Yamada polynomials.
\[tmain\] Zeros of the Yamada polynomials $\{R[C_{n}(\Theta_{s})]\}_{n=1,s=1}^{\infty, \infty}$ are dense in the region $\{z\in \mathbb{C} \, : \, |z+1+z^{-1}|\geq 1\}$.
By calculating the Yamada polynomial of $C_{n}(\Theta_{s})$ we obtain $$R[C_{n}(\Theta_{s})]=H(C_{n}(\Theta_{s})) = \Big( -H(\Theta_{s}) \Big)^n + \sigma \Big( \frac{H(\Theta_{s})+H(\Theta_{s}')}{\sigma} \Big)^n.$$ We will apply Theorem \[tbj\] in the case $l=2$ and use condition (1). Thus, numbers $A$ satisfying the equation $$\label{e1}
|-H(\Theta_{s})| = \Big| \frac{H(\Theta_{s}) + H(\Theta_{s}')}{\sigma} \Big|,$$ are exactly limits of zeros of Yamada polynomials $\{ R[C_{n} (\Theta_{s})] \}_{n=1,s=1}^{\infty, \infty}$.
Observe that $\Theta_{s}'$ is the $s$-bouquet $B_{s}$. Since polynomials $H(\Theta_{s})$ and $H(B_{s})$ were calculated in Lemma \[lemma-dv\], we have $$H(\Theta_{s})=\frac{\sigma-(-1)^{s-1}\sigma^s}{1+\sigma} \qquad \text{and} \qquad H(\Theta_{s}')=(-1)^{s-1}\sigma^s.$$ Therefore, Eq.(\[e1\]) is equivalent to $$\Big| \frac{\sigma-(-1)^{s-1}\sigma^s}{1+\sigma} \Big| = \Big| \frac{1}{\sigma} \Big[ \frac{\sigma-(-1)^{s-1}\sigma^s}{1+\sigma}+(-1)^{s-1}\sigma^s \Big] \Big| ,$$ whence $$|\sigma-(-1)^{s-1}\sigma^s| = |1+(-1)^{s-1}\sigma^s|$$ and then $$\label{e2}
|1+\sigma(-\sigma)^{-s}|=|1-(-\sigma)^{-s}|.$$
Recall that $\sigma = A + 1 + A^{-1}$. Let $A_0$ be a complex number with $|A_{0}+1+A_{0}^{-1}|\geq 1$. Denoting $a=A-A_0$ we get that Eq.(\[e2\]) is equivalent to $$\begin{gathered}
\Big| 1+ \Big( (a+\!A_0) + (a+\!A_0)^{-1} +1 \Big) \Big(-(a+A_0) - (a +A_0)^{-1} - 1 \Big)^{-s} \Big| \\
= \Big| 1- \Big( - (a +A_0)-(a +A_0)^{-1} -1\Big)^{-s} \Big|.
\end{gathered} \label{e3}$$
Consider function $$G(a) = \Big(-(a+A_0) - (a+ A_0)^{-1} -1 \Big)^{-1}.$$ Then Eq.(\[e3\]) is equivalent to $$\Big| 1+ \Big( (a+\!A_0) + (a+\!A_0)^{-1} +1 \Big) G(a)^{s} \Big| \\
= \Big| 1- G(a)^{s} \Big|$$ and $|G(0)|=\frac{1}{|A_0+1+A_0^{-1}|} \leq 1$. Hence by Lemma \[lso\], for any positive real number $\varepsilon > 0$ there exists $s_0$ such that for any $s\geq s_0$ the equation Eq.(\[e3\]) has a solution in the disk $|a| < \varepsilon$. Thus, for $A_{0}$ there exists $A'$ such that $|A'-A_0|<\frac{\varepsilon}{2}$ and $A'$ satisfies the equation Eq.(\[e1\]).
Since zeros of Eq.(\[e1\]) are the limits of zeros of Yamada polynomials $\{ R [C_{n}(\Theta_{s})] \}_{n=1,s=1}^{\infty, \infty}$, there exist integers $n>0$ and $s>0$ such that $R[C_{n}(\Theta_{s})]$ has zero $A^{*}$ with $|A^{*}-A'|< \frac{\varepsilon}{2}$.
So there exists a zero $A^{*}$ of $R[C_{n}(\Theta_{s})]$ with $|A^{*}-A_0|\leq \varepsilon$, where complex number $A_0$ satisfies $|A_{0}+1+A_{0}^{-1}|\geq 1$.
We remark that another proof of Theorem \[tmain\] follows from [@SO] by the relations between graph polynomials. Indeed, Therem 7.2 of [@SO] states that if $|q_{0} -1| \geq 1$, then for each $\varepsilon >0$ there exists $s_{0} < \infty$ and $n_{0}(s) < \infty$ such that for all $s \geq s_{0}$ and $n \geq n_{0}$, the flow polynomial $F_{C_{n(\Theta_{s})}} (q)$ has a zero in the disc $| q - q_{0}| < \varepsilon$. Recall, see for example [@AK], that for planar graphs $G$ the Yamada polynomial coincides with a renormalization of the flow polynomial: $$F_{G}(q) = (-1)^{V(G) - E(G)} R_{G} (A), \quad \text{where} \quad q = A +2 + A^{-1}$$ with $|q_{0} - 1| \geq 1$ equivalent to $|A_{0} + 1 + A^{-1}_{0}| \geq 1$.
Next we consider zeros of Yamada polynomial of the second class of spatial graphs.
\[tmain2\] Zeros of Yamada polynomials $$\{R[C_{n}(\Theta_{s}(\infty_{+}))]\}_{n=1, s=1}^{\infty, \infty}$$ are dense in the region $$\{z\in \mathbb{C} \, : \, |z^{-3}+2z^{-2}+z^{-1}+1|\leq |z^{-1}+1+z|\}.$$ Zeros of the Yamada polynomials $\{ R[C_{n}(\Theta_{s}(\infty_{-}))] \}_{n=1, s=1}^{\infty, \infty}$ are dense in the region $$\{z\in \mathbb{C} \, : \, |z^{3}+2z^{2}+z^1+1|\leq |z^{-1}+1+z|\}.$$
By calculating the Yamada polynomial of $C_{n}(\Theta_{s}(\infty_{+}))$ we obtain $$R[C_{n}(\Theta_{s}(\infty_{+}))]= \Big( -R[\Theta_{s}(\infty_{+})] \Big)^n + \sigma \Big( \frac{R[\Theta_{s}(\infty_{+})]+R[\Theta_{s}(\infty_{+})']}{\sigma} \Big)^n.$$ By Theorem \[tbj\], numbers $A$ satisfying the equation $$\label{e21}
|-R[\Theta_{s}(\infty_{+})]| = \Big| \frac{R[\Theta_{s}(\infty_{+})] + R[\Theta_{s}(\infty_{+})']}{\sigma} \Big|,$$ are exactly limits of zeros of Yamada polynomials $\{ R[C_{n} (\Theta_{s}(\infty_{+}))] \}_{n=1,s=1}^{\infty, \infty}$.
Observe that $\Theta_{s}'$ is the $s$-bouquet $B_{s}$. By Example \[E4\] we have $$R[\Theta_{s}(\infty_{+})]=\frac{(-\sigma)^s+\sigma\big((\sigma+1)A^{-2}+1\big)^s}{1+\sigma}$$ and $$R[\Theta_{s}(\infty_{+})']=(-1)^{s-1}\sigma^s.$$ Therefore, Eq.(\[e21\]) is equivalent to $$\begin{gathered}
\left| \frac{(-\sigma)^s+\sigma\big((\sigma+1)A^{-2}+1\big)^s}{1+\sigma} \right| \qquad \qquad \qquad \qquad \qquad \qquad \qquad \qquad \\
= \left| \frac{1}{\sigma} \left( \frac{(-\sigma)^s+\sigma\big((\sigma+1)A^{-2}+1\big)^s}{1+\sigma}+(-1)^{s-1}\sigma^s \right) \right| ,
\end{gathered}$$ whence $$\label{e22}
\left|1+\sigma\Big( \frac{(\sigma+1)A^{-2}+1}{-\sigma} \Big)^{s}\right|=\left|1-\Big( \frac{(\sigma+1)A^{-2}+1}{-\sigma} \Big)^{s}\right|.$$
Recall that $\sigma = A + 1 + A^{-1}$. Let $A_0$ be a complex number with $$|A_0^{-3}+2A_0^{-2}+A_0^{-1}+1|\leq |A_0^{-1}+1+A_0|.$$ Denoting $a=A-A_0$ we get that Eq.(\[e22\]) is equivalent to $$\begin{gathered}
\left| 1+ \Big( (a+\!A_0)+1+(a+\!A_0)^{-1}\Big) \right. \qquad \qquad \qquad \qquad \\
\cdot \left. \left( \frac{\Big( (a+\!A_0)+2+(a+\!A_0)^{-1}\Big)(a+\!A_0)^{-2}+1}{-\Big( (a+\!A_0)+1+(a+\!A_0)^{-1}\Big)}\right)^s \right| \\
= \left| 1- \left( \frac{\Big( (a+\!A_0)+2+(a+\!A_0)^{-1}\Big)(a+\!A_0)^{-2}+1}{-\Big( (a+\!A_0)+1+(a+\!A_0)^{-1}\Big)}\right)^s \right|.
\end{gathered} \label{e23}$$
Consider the function $$G(a) = \frac{\Big( (a+\!A_0)+2+(a+\!A_0)^{-1}\Big)(a+\!A_0)^{-2}+1}{-\Big( (a+\!A_0)+1+(a+\!A_0)^{-1}\Big)} .$$ Then Eq.(\[e23\]) is equivalent to $$\Big| 1+ \Big( (a+\!A_0) + (a+\!A_0)^{-1} +1 \Big) G(a)^{s} \Big| \\
= \Big| 1- G(a)^{s} \Big|$$ and $|G(0)|=\Big| \frac{(A_0+2+A_0^{-1})A_0^{-2}+1}{A_0+1+A_0^{-1}} \Big| \leq 1$. Hence by Lemma \[lso\], for any positive real number $\varepsilon > 0$ there exists $s_0$ such that for any $s\geq s_0$ the equation Eq.(\[e23\]) has a solution in the disk $|a| < \varepsilon$. Thus, for $A_{0}$ there exists $A'$ such that $|A'-A_0|<\frac{\varepsilon}{2}$ and $A'$ satisfies the equation Eq.(\[e21\]).
Since zeros of Eq.(\[e21\]) are the limits of zeros of Yamada polynomials $\{ R [C_{n}(\Theta_{s}(\infty_+))] \}_{n=1,s=1}^{\infty, \infty}$, there exist integers $n>0$ and $s>0$ such that $R[C_{n}(\Theta_{s}(\infty_+))]$ has zero $A^{*}$ with $|A^{*}-A'|< \frac{\varepsilon}{2}$.
So there exists a zero $A^{*}$ of $R[C_{n}(\Theta_{s}(\infty_+))]$ with $|A^{*}-A_0|\leq \varepsilon$, where complex number $A_0$ satisfies $|A_0^{-3}+2A_0^{-2}+A_0^{-1}+1|\leq |A_0^{-1}+1+A_0|$.
By Proposition $6$ in [@YA], the relation between the Yamada polynomials of a graph $g$ and its mirror image of a diagram $\widehat{g}$ is $R[g](A)=R[\widehat{g}](A^{-1})$.
Since $C_{n}(\Theta_{s}(\infty_-))$ is the mirror image of diagram $C_{n}(\Theta_{s}(\infty_+))$, the zeros of $R[C_{n}(\Theta_{s}(\infty_-))]$ are dense in the following region $\{z\in \mathbb{C} \, : \, |z^{3}+2z^{2}+z^1+1|\leq |z^{-1}+1+z|\}$.
Combining Theorem \[tmain\] and Theorem \[tmain2\], we have
\[tmain3\] Zeros of the Yamada polynomial of spatial graphs are dense in the following region: $$\Omega =
\big\{ z \in \mathbb C \, : \, |z+1+z^{-1}| \geq \min \{1, |z^{3}+2z^{2}+z+1|, |1 + z^{-1} + 2 z^{-2} + z^{-3} | \} \big\}.$$
Acknowledgments {#acknowledgments .unnumbered}
===============
The authors are thankful to Prof. Andrey Dobrynin for the helpful Maple’s visualization of regions, discussed in the paper.
[10]{}
I. Agol, V. Krushkal, Structure of the flow an Yamada polynomial of cubic graphs, preprint available at arXiv:1801.00502v1.
S. Beraha, J. Kahane, N.J. Weiss, Limits of zeros of recursively defined families of polynomials. [*Studies in Foundations and Combinatorics*]{}, Advances in Mathematics Supplementary Studies [**1**]{} (1978), 212–232.
J.H. Conway, C. Gordon, Knots and links in spatial graphs, [*J. Graph Theory*]{} [**7**]{} (1983), 445–453.
N. Chbili, The Yamada polynomial of lens spatial graphs, [*Asian-European Journal of Mathematics*]{} [**8.02**]{} (2015), 1550029.
X.-S. Cheng, Y. Lei, W. Yang, The Homfly polynomial of double crossover links, [*J. Math. Chem.*]{} [**52**]{} (2014), 23–41.
P. Csikvári, P. Frenkel, J. Hlafký, T. Hubai, Chromatic roots and limit of dense graphs, [*Discrete Mathematics*]{} [**340**]{} (2017), 1129–1135.
O.T. Dashbach, T.D. Le, X.-S. Lin, Quantum Morphism and the Jones Polynomial, [*Commun. Math. Phys.*]{} [**224**]{} (2001), 427–442.
A. Dobrynin, A. Vesnin, On Yamada polynomial for graphs embedded knot-wise in the three-dimensional space, [*Vychisl. Sistemy*]{} [**155**]{} (1996), 37–86. (in Russian). The translation is available at https://www.researchgate.net /publication/266336562.
A. Dobrynin, A. Vesnin, On the Yoshinaga polynomial of spatial graphs, [*Kobe J. Math.*]{} [**20:1-2**]{} (2003), 31–37.
C. Ernst, D.W. Sumners, A calculus for rational tangles: applications to DNA recombination, [*Mathematical Proceedings of the Cambridge Philosophical Society*]{} [**108(3)**]{} (1990), 489–515.
A. Henrich, L.H. Kauffman, Tangle insertion invariants for pseudoknots, singular knots, and rigid vertex spatial graphs. In [*Knots, Links, Spatial Graphs, and Algebraic Invariants*]{}, Contemporary Mathematics [**689**]{} (2017), 177–189.
X. Jin, F. Zhang, On computing Kauffman bracket polynomial of Montesinos links, [*Journal of Knot Theory and Its Ramifications*]{} [**19(08)**]{} (2010), 1001–1023.
X. Jin, F. Zhang, F. Dong, E.G.Tay, Zeros of the Jones Polynomial are Dense in the Complex Plane, [*Electronic Journal of Combinatorics*]{} [**17(1)**]{} (2010), 2493–2503.
L.H. Kauffman, A Tutte polynomial for signed graphs, [*Discrete Applied Mathematics*]{} [**25(1)**]{} (1989), 105–127.
L.H. Kauffman, Invariants of graphs in three-space, [*Trans. Amer. Math. Soc.*]{} [**311(2)**]{} (1989), 697–710.
S. Kinoshita, Alexander polynomials as isotopy invariants. I, [*Osaka Math. J.*]{} [**10(2)**]{} (1958), 263–271.
S. Kinoshita, Alexander polynomials as isotopy invariants. II, [*Osaka Math. J.* ]{} [**11**]{} (1959), 91–94.
J. Murakami, The Yamada polynomial of spacial graphs and knit algebras, [*Communications in Mathematical Physics*]{} [**155(3)**]{} (1993), 511–522.
Y. Miyazawa, The Yamada polynomial for virtual graphs, in [*Intelligence of Low Dimensional Topology, 2006*]{} – The International Conference Hiroshima, Japan, 22–26 July 2006. Eds. J.S. Carter, S. Kamada, L.H. Kauffman, A. Kawauchi, T. Kohno. (2007), 205–212.
T. Motohashi, Y. Ohyama, K. Taniyama, Yamada polynomial and crossing number of spatial graphs, [*Rev. Mat. Univ. Complut. Madrid*]{} [**7(2)**]{} (1994), 247–277.
S. Negami, Polynomial invariants of graphs, [*Trans. Amer. Math. Soc.*]{} [**299**]{} (1987), 601–622.
Y. Ohyama, K. Taniyama, Vassiliev invariants of knots in a spatial graph, [*Pacific J. Math.*]{} [**200(1)**]{} (2001), 191–205.
R.C. Read, E.G. Whitehead, Chromatic polynomials of homeomorphism classes of graphs, [*Discrete. Math.*]{} [**204**]{} (1999), 337–356.
A.D. Sokal, Chromatic roots are dense in the whole complex plane, [*Comb. Probab. Comput.*]{} [**13**]{} (2004), 221–261.
J. Simon, Topological chirality of certain molecules, [*Topology*]{} [**25(2)**]{} (1986), 229–235.
V. Vershinin, A. Vesnin, Yamada Polynomial and Khovanov Cohomology, in [*Intelligence of Low Dimensional Topology, 2006*]{} – The International Conference Hiroshima, Japan, 22–26 July 2006. Eds. J.S. Carter, S. Kamada, L.H. Kauffman, A. Kawauchi, T. Kohno. (2007), 337–346.
D. M. Walba, Stereochemical topology, in [*Chemical applications of topology and graph theory: A Collection of Papers from a Symposium Held at University of Georgia, Athens, Georgia, USA, 18-22 April 1983*]{}, ed. R.B. King (Elsevier, 1983), 17–32.
S. Yamada, An invariant of spatial graphs, [*J. Graph Theory*]{} [**13**]{} (1989), 537–551.
S. Yoshinaga, An invariant of spatial graphs associated with $U(q)(sl(2, C)$, [*Kobe J. Math.*]{} [**8**]{} (1991), 25–40.
[^1]: The second author is supported in part by grants ( No.11329101 and No.11431009 ) of NSFC; the third author is supported in part by grants ( No.11671064 and No.11471151 ) of NSFC; the fourth author is supported by the Laboratory of Topology and Dynamics, Novosibirsk State University (a grant no. 14.Y26.31.0025 of the government of the Russian Federation) and a grant RFBR-16-01-00414.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- Chandrayee Basu
- Mukesh Singhal
- 'Anca D. Dragan'
bibliography:
- 'featurequery.bib'
title: 'Learning from Richer Human Guidance: Augmenting Comparison-Based Learning with Feature Queries'
---
| {
"pile_set_name": "ArXiv"
} |
[**Proof of two supercongruences by the Wilf-Zeilberger method**]{}
Guo-Shuai Mao
[$^1$Department of Mathematics, Nanjing University of Information Science and Technology, Nanjing 210044, People’s Republic of China\
[[email protected] ]{} ]{}
0.7cm [**Abstract.**]{} In this paper, we prove two supercongruences by the Wilf-Zeilberger method. One of them is, for any prime $p>3$, $$\begin{aligned}
\sum_{n=0}^{(p-1)/2}\frac{3n+1}{(-8)^n}\binom{2n}n^3\equiv p\left(\frac{-1}p\right)+\frac{p^3}4\left(\frac2p\right)E_{p-3}\left(\frac14\right)\pmod{p^4},\end{aligned}$$ where $\left(\frac{\cdot}p\right)$ stands for the Legendre symbol, and $E_{n}(x)$ are the Euler polynomials. This congruence confirms a conjecture of Sun [@sun-numb-2019 (2.18)] with $n=1$.
[*Keywords*]{}: Supercongruence; Binomial coeficients; Wilf-Zeilberger method; Euler polynomials.
0.2cm [*AMS Subject Classifications:*]{} 11B65, 11A07, 11B68, 33F10, 05A10.
Introduction
============
Recall that the Euler numbers $\{E_n\}$ and the Euler polynomials $\{E_n(x)\}$ are defined by $$\frac{2e^t}{e^{2t}+1}=\sum_{n=0}^\infty E_n\frac{t^n}{n!}\ (|t|<\frac{\pi}2)\ \mbox{and}
\ \frac{2e^{xt}}{e^{t}+1}=\sum_{n=0}^\infty E_n(x)\frac{t^n}{n!}\ (|t|<\pi),$$ the Bernoulli numbers $\{B_n\}$ and the Bernoulli polynomials $\{B_n(x)\}$ are defined as follows: $$\frac x{e^x-1}=\sum_{n=0}^\infty B_n\frac{x^n}{n!}\ \ (0<|x|<2\pi)\ \mbox{and}\ B_n(x)=\sum_{k=0}^n\binom nkB_kx^{n-k}\ \ (n\in\mathbb{N}).$$
In the past decade, many researchers studied supercongruences via the Wilf-Zeilberger (WZ) method (see, for instance, [@gl-arxiv-2019; @CXH-rama-2016; @he-jnt-2015; @hm-rama-2017; @mz-rama-2019; @oz-jmaa-2016; @sun-ijm-2012; @zudilin-jnt-2009]). For instance, W. Zudilin [@zudilin-jnt-2009] proved several Ramanujan-type supercongruences by the WZ method. One of them, conjectured by van Hamme [@vhamme], says that for any odd prime $p$, $$\begin{aligned}
\label{wzpr}
\sum_{k=0}^{(p-1)/2}(4k+1)(-1)^k\left(\frac{\left(\frac12\right)_k}{k!}\right)^3\equiv(-1)^{(p-1)/2}p\pmod{p^3},\end{aligned}$$ where $(a)_n=a(a+1)\ldots(a+n-1) (n\in\{1,2,\ldots\})$ with $(a)_0=1$ is the raising factorial for $a\in\mathbb{C}$.
For $n\in\mathbb{N}$, define $$H_n:=\sum_{0<k\leq n}\frac1k,\ H_n^{(2)}:=\sum_{0<k\leq n}\frac1{k^2},\ H_0=H_0^{(2)}=0,$$ where $H_n$ with $n\in\mathbb{N}$ are often called the classical harmonic numbers. Let $p>3$ be a prime. J. Wolstenholme [@wolstenholme-qjpam-1862] proved that $$\begin{aligned}
\label{hp-1}
H_{p-1}\equiv0\pmod{p^2}\ \mbox{and}\ H_{p-1}^{(2)}\equiv0\pmod p,\end{aligned}$$ which imply that $$\begin{aligned}
\binom{2p-1}{p-1}\equiv1\pmod{p^3}.\label{2p1p}\end{aligned}$$
Z.-W. Sun [@sun-ijm-2012] proved the following supercongruence by the WZ method, for any odd prime $p$, $$\label{sun}
\sum_{k=0}^{p-1}\frac{4k+1}{(-64)^k}\binom{2k}k^3\equiv(-1)^{\frac{(p-1)}2}p+p^3E_{p-3}\pmod{p^4}.$$
Guo and Liu [@gl-arxiv-2019] showed that for any prime $p>3$, $$\label{glp4}
\sum_{k=0}^{(p+1)/2}(-1)^k(4k-1)\frac{\left(-\frac12\right)_k^3}{(1)_k^3}\equiv p(-1)^{(p+1)/2}+p^3(2-E_{p-3})\pmod{p^4}.$$ Guo also researched $q$-analogues of Ramanujan-type supercongruences and $q$-analogues of supercongruences of van Hamme (see, for instance, [@g-jmaa-2018; @guo-rama-2019]).
Long [@long-2011-pjm] and Chen, Xie and He [@CXH-rama-2016] proved that, for any odd prime $p$, $$\sum_{n=0}^{(p-1)/2}\frac{6n+1}{(-512)^n}\binom{2n}n^3\equiv p\left(\frac{-2}p\right)\pmod{p^2}.$$
Recently, the author [@mao-arxiv-2019] proved a conjecture of Z.-W. Sun which says that: Let $p>3$ be a prime. Then $$\sum_{n=0}^{(p-1)/2}\frac{6n+1}{(-512)^n}\binom{2n}n^3\equiv p\left(\frac{-2}p\right)+\frac{p^3}4\left(\frac2p\right)E_{p-3}\pmod{p^4}.$$ Chen, Xie and He [@CXH-rama-2016] confirmed a supercongruence conjetured by Z.-W. Sun [@sun-scm-2011], which says that for any prime $p>3$, $$\sum_{k=0}^{p-1}\frac{3k+1}{(-8)^k}\binom{2k}k^3\equiv p(-1)^{(p-1)/2}+p^3E_{p-3}\pmod{p^4}.$$ In this paper we prove the following result:
\[Thsun\] Let $p>3$ be a prime. Then $$\sum_{n=0}^{(p-1)/2}\frac{3n+1}{(-8)^n}\binom{2n}n^3\equiv p\left(\frac{-1}p\right)+\frac{p^3}{4}\left(\frac{2}p\right)E_{p-3}\left(\frac14\right)\pmod{p^4}.$$
This congruence confirms a conjecture of Sun [@sun-numb-2019 (2.18)] with $n=1$. And this congruence with [@mao-arxiv-2019 Theorem 1.5] yields that $$\sum_{n=0}^{(p-1)/2}\frac{3n+1}{(-8)^n}\binom{2n}n^3\equiv4\left(\frac{2}p\right)\sum_{n=0}^{p-1}\frac{6n+1}{(-512)^n}\binom{2n}n^3-3p\left(\frac{-1}p\right)\pmod{p^4},$$ which is a conjecture of Sun [@sun-scm-2011 Conjecture 5.1].
Guo [@g-jmaa-2018] proved that $$\sum_{k=0}^{(p^r-1)/2}\frac{4k+1}{(-64)^k}\binom{2k}k^3\equiv(-1)^{\frac{(p-1)r}2}p^r\pmod{p^{r+2}},$$ and in the same paper he proposed a conjecture as follows:
\[Guo-2018\][([@g-jmaa-2018 Conjecture 5.1])]{} $$\sum_{k=0}^{p^r-1}\frac{4k+1}{(-64)^k}\binom{2k}k^3\equiv(-1)^{\frac{(p-1)r}2}p^r\pmod{p^{r+2}}.$$
Guo and zudilin proved this Conjecture by founding its $q$-anology, (see [@GZ-aim-2019]). And the author [@mao-2019-arxiv] gave a new proof of it by the WZ method.
\[Th10ngz\] For any prime $p>3$ and integer $r>0$, we have $$\begin{aligned}
\sum_{n=0}^{p^r-1}\frac{3n+1}{(-8)^n}\binom{2n}n^3&\equiv(-1)^{(p^r-1)/2}p^r\pmod{p^{r+2}}.\end{aligned}$$
Our main tool in this paper is the WZ method. We shall prove Theorem \[Thsun\] in the next Section. And Theorem \[Th10ngz\] will be proved in Section 3.
Proof of Theorem \[Thsun\]
==========================
We will use the following WZ pair which appears in [@CXH-rama-2016] to prove Theorem \[Thsun\]. For nonnegative integers $n, k$, define $$F(n,k)=\frac{(-1)^{n}(3n-2k+1)\binom{2n}n\binom{2n-2k}{n-k}\binom{2n-2k}{n}}{2^{3n-2k}}$$ and $$G(n,k)=\frac{(-1)^{n+1}n\binom{2n}n\binom{2n-2k}{n-k}\binom{2n-2k}{n-1}}{2^{3n-2k}}.$$ Clearly $F(n,k)=G(n,k)=0$ if $n<k$ and $F(n,n)=0$ for any positive integer $n$. It is easy to check that $$\label{FG}
F(n,k-1)-F(n,k)=G(n+1,k)-G(n,k)$$ for all nonnegative integer $n$ and $k>0$.
Summing (\[FG\]) over $n$ from $0$ to $(p-1)/2$ we have $$\sum_{n=0}^{(p-1)/2}F(n,k-1)-\sum_{n=0}^{(p-1)/2}F(n,k)=G\left(\frac{p+1}2,k\right)-G(0,k)=G\left(\frac{p+1}2,k\right).$$ Furthermore, summing both side of the above identity over $k$ from $1$ to $(p-1)/2$, we obtain $$\begin{aligned}
\label{wz1}
\sum_{n=0}^{(p-1)/2}F(n,0)=\sum_{k=1}^{(p-1)/2}G\left(\frac{p+1}2,k\right).\end{aligned}$$
\[mor\][([@Mor])]{} For any prime $p>3$, we have $$\binom{p-1}{(p-1)/2}\equiv(-1)^{(p-1)/2}4^{p-1}\pmod{p^3}.$$
By the definition of $G(n,k)$ we have $$\begin{aligned}
G\left(\frac{p+1}2,k\right)&=(-1)^{\frac{p-1}2}\frac{\frac{p+1}2\binom{p+1}{\frac{p+1}2}\binom{p+1-2k}{\frac{p+1}2-k}\binom{p+1-2k}{\frac{p-1}2}}{2^{\frac{3p+3}2-2k}}\notag\\
&=\frac{2p(-1)^{\frac{p-1}2}\binom{p-1}{\frac{p-1}2}}{2^{\frac{3p+3}2}}\binom{p+1-2k}{\frac{p+1}2-k}\binom{p+1-2k}{\frac{p-1}2}4^k.\end{aligned}$$ It is easy to see that $$\frac{\binom{2n-2k}{n-k}}{4^{n-k}}=\frac{\left(\frac12\right)_{n-k}}{(n-k)!}, \ \ \left(\frac12\right)_{n-k}\left(\frac12+n-k\right)_{k-1}=\left(\frac12\right)_{n-1}$$ and $$\binom{2n-2k}{n-1}=\frac{(2n-2k)!}{(n-1)!(n+1-2k)!}=\frac{\binom{2n-2k}{n-k}(n-k)!^2}{(n-1)!(n+1-2k)!}.$$ So we have $$\begin{aligned}
\label{Gpk}
G\left(\frac{p+1}2,k\right)&=\frac{(-1)^{\frac{p-1}2}2p\binom{p-1}{\frac{p-1}2}4^{p+1}}{\left(\frac{p-1}2\right)!2^{\frac{3p+3}2}}\frac{\left(\frac12\right)_{(p+1)/2-k}^2}{\left(\frac{p+3}2-2k\right)!4^k}\notag\\
&=\frac{32p(-1)^{\frac{p-1}2}\binom{p-1}{\frac{p-1}2}^3}{2^{\frac{3p+3}2}}\frac{\left(\frac{p-1}2\right)!}{\left(\frac{p+3}2-2k\right)!\left(\frac p2+1-k\right)_{k-1}^24^k}.\end{aligned}$$
\[p-12k\] Let $p>3$ be a prime. Then $$2^{\frac{9p-9}2}\sum_{k=0}^{\frac{p-3}2}\frac{\binom{\frac{p-1}2}{2k}\binom{2k}k}{4^k}\equiv (-1)^{(p-1)/2}\left(1+6pq_p(2)+15p^2q_p(2)^2\right)\pmod{p^3}.$$
First we have the following two identities: $$\sum_{k=0}^n\frac{\binom{2n}k\binom{2n-k}k}{4^k}=\frac{\binom{4n}{2n}}{4^n}\ \ \mbox{and}\ \ \sum_{k=0}^n\frac{\binom{2n+1}k\binom{2n+1-k}k}{4^k}=\frac{\binom{4n+1}{2n+1}}{4^n}.$$ So when $p\equiv1\pmod4$, by Lemma \[mor\] we have $$\begin{aligned}
2^{\frac{9p-9}2}\sum_{k=0}^{\frac{p-3}2}\frac{\binom{\frac{p-1}2}{2k}\binom{2k}k}{4^k}&=2^{\frac{9p-9}2}\sum_{k=0}^{\frac{p-1}4}\frac{\binom{\frac{p-1}2}{k}\binom{\frac{p-1}2-k}k}{4^k}=2^{4p-4}\binom{p-1}{\frac{p-1}2}\equiv(-1)^{\frac{p-1}2}2^{6p-6}\\
&\equiv(-1)^{(p-1)/2}\left(1+6pq_p(2)+15p^2q_p(2)^2\right)\pmod{p^3}.\end{aligned}$$ And when $p\equiv3\pmod4$, we have $$\begin{aligned}
2^{\frac{9p-9}2}\sum_{k=0}^{\frac{p-3}2}\frac{\binom{\frac{p-1}2}{2k}\binom{2k}k}{4^k}&=2^{\frac{9p-9}2}\sum_{k=0}^{\frac{p-3}4}\frac{\binom{\frac{p-1}2}{k}\binom{\frac{p-1}2-k}k}{4^k}=2^{4p-3}\binom{p-2}{\frac{p-1}2}\equiv(-1)^{\frac{p-1}2}2^{6p-6}\\
&\equiv(-1)^{(p-1)/2}\left(1+6pq_p(2)+15p^2q_p(2)^2\right)\pmod{p^3}.\end{aligned}$$ Therefore the proof of Lemma \[p-12k\] is finished.
\[p-12khk\] For any prime $p>3$, we have $$2^{\frac{9p-9}2}\sum_{k=0}^{\frac{p-3}2}\frac{\binom{\frac{p-1}2}{2k}\binom{2k}kH_k}{4^k}\equiv-3(-1)^{\frac{p-1}2}\left(2q_p(2)+11pq_p(2)^2\right)\pmod{p^2}.$$
First we have the following two identities: $$\sum_{k=0}^n\frac{\binom{2n}k\binom{2n-k}kH_k}{4^k}=\frac{\binom{4n}{2n}}{4^n}(3H_{2n}-2H_{4n}),$$ $$\sum_{k=0}^n\frac{\binom{2n+1}k\binom{2n+1-k}kH_k}{4^k}=\frac{\binom{4n+1}{2n+1}}{4^n}\left(3H_{2n+1}-2H_{4n+2}\right).$$ So when $p\equiv1\pmod4$, by Lemma \[mor\], [@sun-jnt-2008 Theorem 3.2] and (\[hp-1\]) we have $$\begin{aligned}
&2^{\frac{9p-9}2}\sum_{k=0}^{\frac{p-3}2}\frac{\binom{\frac{p-1}2}{2k}\binom{2k}kH_k}{4^k}=2^{\frac{9p-9}2}\sum_{k=0}^{\frac{p-1}4}\frac{\binom{\frac{p-1}2}{k}\binom{\frac{p-1}2-k}kH_k}{4^k}=2^{4p-4}\binom{p-1}{\frac{p-1}2}(3H_{\frac{p-1}2}-2H_{p-1})\\
&\equiv3(-1)^{\frac{p-1}2}2^{6p-6}H_{\frac{p-1}2}\equiv3(-1)^{(p-1)/2}\left(1+6pq_p(2)\right)(-2q_p(2)+pq_p(2))\\
&\equiv-3(-1)^{\frac{p-1}2}\left(2q_p(2)+11pq_p(2)^2\right)\pmod{p^2}.\end{aligned}$$ And when $p\equiv3\pmod4$, we have $$\begin{aligned}
&2^{\frac{9p-9}2}\sum_{k=0}^{\frac{p-3}2}\frac{\binom{\frac{p-1}2}{2k}\binom{2k}kH_k}{4^k}=2^{\frac{9p-9}2}\sum_{k=0}^{\frac{p-3}4}\frac{\binom{\frac{p-1}2}{k}\binom{\frac{p-1}2-k}kH_k}{4^k}=2^{4p-4}\binom{p-1}{\frac{p-1}2}(3H_{\frac{p-1}2}-2H_{p-1})\\
&\equiv3(-1)^{\frac{p-1}2}2^{6p-6}H_{\frac{p-1}2}\equiv3(-1)^{(p-1)/2}\left(1+6pq_p(2)\right)(-2q_p(2)+pq_p(2))\\
&\equiv-3(-1)^{\frac{p-1}2}\left(2q_p(2)+11pq_p(2)^2\right)\pmod{p^2}.\end{aligned}$$ Therefore the proof of Lemma \[p-12khk\] is completed.
\[p-12khkhk2\] For any prime $p>3$, we have $$2^{\frac{9p-9}2}\sum_{k=0}^{\frac{p-3}2}\frac{\binom{\frac{p-1}2}{2k}\binom{2k}k\left(H_k^2+H_{k}^{(2)}\right)}{4^k}\equiv36(-1)^{\frac{p-1}2}q_p(2)^2\pmod{p}.$$
First we have the following two identities: $$\sum_{k=0}^n\frac{\binom{2n}k\binom{2n-k}k\left(H_k^2+H_k^{(2)}\right)}{4^k}=\frac{\binom{4n}{2n}}{4^n}\left(5H_{2n}^{(2)}-4H_{4n}^{(2)}\right)+\frac{\binom{4n}{2n}}{4^n}\left(3H_{2n}-2H_{4n}\right)^2,$$ $$\sum_{k=0}^n\frac{\binom{2n+1}k\binom{2n+1-k}k\left(H_k^2+H_k^{(2)}\right)}{4^k}=\frac{\binom{4n+1}{2n+1}}{4^n}\left(\left(5H_{2n+1}^{(2)}-4H_{4n+2}^{(2)}\right)+\left(3H_{2n+1}-2H_{4n+2}\right)^2\right).$$ So when $p\equiv1\pmod4$, by Lemma \[mor\], [@sun-jnt-2008 Theorem 3.5] and (\[hp-1\]) we have $$\begin{aligned}
&2^{\frac{9p-9}2}\sum_{k=0}^{\frac{p-3}2}\frac{\binom{\frac{p-1}2}{2k}\binom{2k}k\left(H_k^2+H_k^{(2)}\right)}{4^k}=2^{\frac{9p-9}2}\sum_{k=0}^{\frac{p-1}4}\frac{\binom{\frac{p-1}2}{k}\binom{\frac{p-1}2-k}k\left(H_k^2+H_k^{(2)}\right)}{4^k}\\
&=2^{4p-4}\binom{p-1}{\frac{p-1}2}\left(\left(5H_{(p-1)/2}^{(2)}-4H_{p-1}^{(2)}\right)+\left(3H_{(p-1)/2}-2H_{p-1}\right)^2\right)\\
&\equiv(-1)^{\frac{p-1}2}9H_{\frac{p-1}2}^2\equiv36(-1)^{(p-1)/2}q_p(2)^2\pmod{p}.\end{aligned}$$ And when $p\equiv3\pmod4$, we have $$\begin{aligned}
&2^{\frac{9p-9}2}\sum_{k=0}^{\frac{p-3}2}\frac{\binom{\frac{p-1}2}{2k}\binom{2k}k\left(H_k^2+H_k^{(2)}\right)}{4^k}=2^{\frac{9p-9}2}\sum_{k=0}^{\frac{p-3}4}\frac{\binom{\frac{p-1}2}{k}\binom{\frac{p-1}2-k}k\left(H_k^2+H_k^{(2)}\right)}{4^k}\\
&=2^{4p-4}\binom{p-1}{\frac{p-1}2}\left(\left(5H_{(p-1)/2}^{(2)}-4H_{p-1}^{(2)}\right)+\left(3H_{(p-1)/2}-2H_{p-1}\right)^2\right)\\
&\equiv(-1)^{\frac{p-1}2}9H_{\frac{p-1}2}^2\equiv36(-1)^{(p-1)/2}q_p(2)^2\pmod{p}.\end{aligned}$$ Therefore the proof of Lemma \[p-12khkhk2\] is complete.
\[mos\][\[See [@MOS]\]]{} Let $x$ and $y$ be variables and $n\in\mathbb{N}$. Then $$\begin{aligned}
&{\rm (i)}.\ B_{2n+1}=0.\\
&{\rm (ii)}.\ B_n(1-x)=(-1)^nB_n(x).\\
&{\rm (iii)}.\ B_n(x+y)=\sum_{k=0}^n\binom{n}kB_{n-k}(y)x^k.\\
&{\rm (iv)}.\ E_{n-1}(x)=\frac{2^n}n\left(B_n\left(\frac{x+1}2\right)-B_n\left(\frac x2\right)\right).\end{aligned}$$
\[p-12khk2\] For any prime $p>3$, we have $$\sum_{k=0}^{\frac{p-3}2}\frac{\binom{\frac{p-1}2}{2k}\binom{2k}kH_{k}^{(2)}}{4^k}\equiv\sum_{k=1}^{\lfloor\frac{p-1}4\rfloor}\frac{\binom{4k}{2k}\binom{2k}kH_{k}^{(2)}}{{64}^k}\equiv-E_{p-3}\left(\frac14\right)\pmod{p}.$$
It is easy to see that $$\sum_{k=0}^{\frac{p-3}2}\frac{\binom{\frac{p-1}2}{2k}\binom{2k}kH_{k}^{(2)}}{4^k}=\sum_{k=0}^{\lfloor\frac{p-1}4\rfloor}\frac{\binom{\frac{p-1}2}{2k}\binom{2k}kH_{k}^{(2)}}{4^k}\equiv\sum_{k=1}^{\lfloor\frac{p-1}4\rfloor}\frac{\binom{4k}{2k}\binom{2k}kH_{k}^{(2)}}{{64}^k}\pmod p$$ since $\binom{\frac{p-1}2}{2k}\equiv\binom{4k}{2k}/{16}^k\pmod p$ for each $k\in\{0,1,\ldots,\lfloor\frac{p-1}4\rfloor\}$ and $H_0^{(2)}=0$. So we just need to verify that $$\sum_{k=1}^{\lfloor\frac{p-1}4\rfloor}\frac{\binom{4k}{2k}\binom{2k}kH_{k}^{(2)}}{{64}^k}\equiv-E_{p-3}\left(\frac14\right)\pmod{p}.$$ We have three identities: $$\sum_{k=1}^n\binom{n}k\binom{-3/4}kH_k^{(2)}=(-1)^n\binom{-1/4}n\left(H_n^{(2)}-\sum_{k=1}^n\frac{(-1)^k}{k^2\binom{-1/4}k}\right),$$ $$\sum_{k=1}^n\binom{n}k\binom{-1/4}kH_k^{(2)}=(-1)^n\binom{-3/4}n\left(H_n^{(2)}-\sum_{k=1}^n\frac{(-1)^k}{k^2\binom{-3/4}k}\right)$$ and $$\sum_{k=1}^n\frac{(-1)^k}{k^2\binom{n}k}=H_n^{(2)}+2\sum_{k=1}^n\frac{(-1)^k}{k^2}.$$ It is known that $\frac{\binom{4k}{2k}\binom{2k}k}{{64}^k}=\binom{-1/4}k\binom{-3/4}k$. Thus, if $p\equiv1\pmod4$, then $$\begin{aligned}
\sum_{k=1}^{\lfloor\frac{p-1}4\rfloor}\frac{\binom{4k}{2k}\binom{2k}kH_{k}^{(2)}}{{64}^k}&\equiv\sum_{k=1}^{\frac{p-1}4}\binom{\frac{p-1}4}k\binom{-3/4}kH_k^{(2)}\equiv(-1)^{\frac{p-1}4}\left(H_{\frac{p-1}4}^{(2)}-\sum_{k=1}^{\frac{p-1}4}\frac{(-1)^k}{k^2\binom{\frac{p-1}4}k}\right)\\
&=-2(-1)^{\frac{p-1}4}\sum_{k=1}^{\frac{p-1}4}\frac{(-1)^k}{k^2}\pmod p.\end{aligned}$$ If $p\equiv3\pmod 4$, then $$\begin{aligned}
\sum_{k=1}^{\lfloor\frac{p-1}4\rfloor}\frac{\binom{4k}{2k}\binom{2k}kH_{k}^{(2)}}{{64}^k}&\equiv\sum_{k=1}^{\frac{p-3}4}\binom{\frac{p-3}4}k\binom{-1/4}kH_k^{(2)}\equiv(-1)^{\frac{p-3}4}\left(H_{\frac{p-3}4}^{(2)}-\sum_{k=1}^{\frac{p-1}4}\frac{(-1)^k}{k^2\binom{\frac{p-3}4}k}\right)\\
&=-2(-1)^{\frac{p-3}4}\sum_{k=1}^{\frac{p-3}4}\frac{(-1)^k}{k^2}\pmod p.\end{aligned}$$ Hence $$\sum_{k=1}^{\lfloor\frac{p-1}4\rfloor}\frac{\binom{4k}{2k}\binom{2k}kH_{k}^{(2)}}{{64}^k}\equiv-2(-1)^{\lfloor\frac{p-1}4\rfloor}\sum_{k=1}^{\lfloor\frac{p-1}4\rfloor}\frac{(-1)^k}{k^2}\pmod p.$$ In view of [@sun-jnt-2008 Lemma2.1], for any $p,m\in\mathbb{N}$ and $k,r\in\mathbb{Z}$ with $k\geq0$, we have $$\sum_{\substack{x=0\\{x\equiv r\pmod{m}}}}^{p-1}x^k=\frac{m^k}{k+1}\left(B_{k+1}\left(\frac{p}{m}+\left\{\frac{r-p}{m}\right\}\right)-B_{k+1}\left(\left\{\frac{r}{m}\right\}\right)\right).$$ By using this identity, (i) and (iii) of Lemma \[mos\], we have $$\begin{aligned}
\sum_{j=1}^{\lfloor\frac{p-1}8\rfloor}\frac1{j^2}=64\sum_{\substack{k=1\\k\equiv0\pmod8}}^{p-1}\frac{1}{k^2}\equiv64\sum_{\substack{x=0\\x\equiv0\pmod8}}^{p-1}x^{p-3}\equiv-\frac12B_{p-2}\left(\left\{-\frac p8\right\}\right)\pmod p\end{aligned}$$ and $$\begin{aligned}
\sum_{j=1}^{\lfloor\frac{p-1}8\rfloor}\frac1{(2j-1)^2}&=16\sum_{j=1}^{\lfloor\frac{p-1}8\rfloor}\frac1{(8j-4)^2}=16\sum_{\substack{j=1\\j\equiv4\pmod8}}^{p-1}\frac1{j^2}\equiv16\sum_{\substack{x=0\\j\equiv4\pmod8}}^{p-1}x^{p-3}\\
&\equiv-\frac18B_{p-2}\left(\left\{\frac{4-p}8\right\}\right)\pmod p.\end{aligned}$$ So $$\begin{aligned}
\sum_{k=1}^{\lfloor\frac{p-1}4\rfloor}\frac{(-1)^k}{k^2}&=\frac14\sum_{k=1}^{\lfloor\frac{p-1}8\rfloor}\frac{1}{k^2}-\sum_{k=1}^{\lfloor\frac{p-1}8\rfloor}\frac{1}{(2k-1)^2}\\
&\equiv\frac18\left(B_{p-2}\left(\left\{\frac{4-p}8\right\}\right)-B_{p-2}\left(\left\{-\frac p8\right\}\right)\right)\pmod p.\end{aligned}$$ Hence $$\sum_{k=1}^{\lfloor\frac{p-1}4\rfloor}\frac{\binom{4k}{2k}\binom{2k}kH_{k}^{(2)}}{{64}^k}\equiv-\frac14(-1)^{\lfloor\frac{p-1}4\rfloor}\left(B_{p-2}\left(\left\{\frac{4-p}8\right\}\right)-B_{p-2}\left(\left\{-\frac p8\right\}\right)\right)\pmod p.$$ Then we immediately get the desired result by (iv) of Lemma \[mos\].
\[G12\] For any primes $p>3$, we have $$\sum_{k=1}^{(p-1)/2}G\left(\frac{p+1}2,k\right)\equiv p\left(\frac{-1}p\right)+\frac{p^3}4\left(\frac{2}p\right)E_{p-3}\left(\frac14\right)\pmod{p^4}.$$
It is easy to check that $$\begin{aligned}
\left(\frac p2+1-k\right)_{k-1}^2&=\left(\frac p2+1-k\right)^2\ldots\left(\frac p2-1\right)^2\\
&\equiv(k-1)!^2\left(1-pH_{k-1}+\frac{p^2}4\left(2H_{k-1}^2-H_{k-1}^{(2)}\right)\right)\pmod{p^3}.\end{aligned}$$ This with (\[Gpk\]) yields that, modulo $p^4$ we have $$\begin{aligned}
G\left(\frac{p+1}2,k\right)&\equiv\frac{32p(-1)^{\frac{p-1}2}\binom{p-1}{\frac{p-1}2}^3}{2^{\frac{3p+3}2}}\frac{\binom{\frac{p-1}2}{2k-2}\binom{2k-2}{k-1}}{\left(1-pH_{k-1}+\frac{p^2}2H_{k-1}^2-\frac{p^2}4H_{k-1}^{(2)}\right)4^k}\\
&\equiv\frac{32p(-1)^{\frac{p-1}2}\binom{p-1}{\frac{p-1}2}^3}{2^{\frac{3p+3}2}}\frac{\binom{\frac{p-1}2}{2k-2}\binom{2k-2}{k-1}}{4^k}\left(1+pH_{k-1}+\frac{p^2}2H_{k-1}^2+\frac{p^2}4H_{k-1}^{(2)}\right).\end{aligned}$$ So by Lemma \[mor\], we have $$\begin{aligned}
\sum_{k=1}^{(p-1)/2}G\left(\frac{p+1}2,k\right)&\equiv 4p2^{\frac{9p-9}2}\sum_{k=1}^{\frac{p-1}2}\frac{\binom{\frac{p-1}2}{2k-2}\binom{2k-2}{k-1}}{4^k}\left(1+pH_{k-1}+\frac{p^2}2H_{k-1}^2+\frac{p^2}4H_{k-1}^{(2)}\right)\\
&=p2^{\frac{9p-9}2}\sum_{k=0}^{\frac{p-3}2}\frac{\binom{\frac{p-1}2}{2k}\binom{2k}{k}}{4^k}\left(1+pH_{k}+\frac{p^2}2H_{k}^2+\frac{p^2}4H_{k}^{(2)}\right)\pmod{p^4}.\end{aligned}$$ By Lemmas \[p-12k\]–\[p-12khkhk2\], \[p-12khk2\] and $$2^{(p-1)/2}\equiv\left(\frac2p\right)\left(1+\frac p2q_p(2)-\frac{p^2}8q_p(2)^2\right)\pmod {p^3},$$ we immediately obtain the desired result of Lemma \[G12\].
[*Proof of Theorem \[Thsun\]*]{}. Combining (\[wz1\]) with Lemma \[G12\], we immediately get that for any prime $p>3$ $$\sum_{n=0}^{(p-1)/2}\frac{3n+1}{(-8)^n}\binom{2n}n^3\equiv p\left(\frac{-1}p\right)+\frac{p^3}4\left(\frac{2}p\right)E_{p-3}\left(\frac14\right)\pmod{p^4}.$$ Therefore the proof of Theorem \[Thsun\] is complete. [$\Box$]{}
Proof of Theorem \[Th10ngz\]
============================
By the same WZ pair in Section 2, we have $$\begin{aligned}
\label{wz2}
\sum_{n=0}^{p^r-1}F(n,0)=\sum_{k=1}^{p^r-1}G\left(p^r,k\right).\end{aligned}$$
\[smor\][([@long-2011-pjm Lemma 2.4])]{} For any prime $p>3$, we have $$\binom{p^r-1}{(p^r-1)/2}\equiv(-1)^{(p^r-1)/2}4^{p^r-1}\pmod{p^3}.$$
\[Gprpr+12\] For any prime $p>3$ and integer $r>0$, we have $$G\left(p^r,\frac{p^r+1}2\right)\equiv(-1)^{(p^r-1)/2}p^r\pmod{p^{r+2}}.$$
By the definition of $G(n,k)$, we have $$G\left(p^r,\frac{p^r+1}2\right)=\frac{p^r\binom{2p^r}{p^r}\binom{p^r-1}{(p^r-1)/2}}{2^{2p^r-1}}.$$ It is easy to see that $$\begin{aligned}
\label{mao}
\binom{2p^r}{p^r}=2\prod_{k=1}^{p^r-1}\left(1-\frac{2p^r}k\right)\equiv2-4p^rH_{p^r-1}\equiv2-4pH_{p-1}\equiv2\pmod {p^2}\end{aligned}$$ with Wolstenholme’s result $H_{p-1}\equiv0\pmod{p^2}$ as mentioned in the introduction.
So by Lemma \[smor\] and (\[mao\]), we have $$G\left(p^r,\frac{p^r+1}2\right)\equiv(-1)^{(p^r-1)/2}p^r\pmod{p^{r+2}}.$$ Now the proof of Lemma \[Gprpr+12\] is complete.
\[Gprk\] For any prime $p>3$ and integer $r>0$, we have $$\sum_{k=1}^{(p^r-1)/2}G\left(p^r,k\right)\equiv0\pmod{p^{r+2}}.$$
By the definition of $G(n,k)$, we have $$\begin{aligned}
\sum_{k=1}^{(p^r-1)/2}G\left(p^r,k\right)&=\frac{p^r\binom{2p^r}{p^r}}{2^{3p^r}}\sum_{k=1}^{(p^r-1)/2}\binom{2p^r-2k}{p^r-k}\binom{2p^r-2k}{p^r-1}4^k\\
&=\frac{p^r\binom{2p^r}{p^r}}{2^{3p^r}}\sum_{k=1}^{(p^r-1)/2}\binom{2p^r-2k}{p^r-k}\binom{-p^r}{p^r+1-2k}4^k\\
&=-\frac{p^{2r}\binom{2p^r}{p^r}}{2^{3p^r}}\sum_{k=1}^{(p^r-1)/2}\binom{2p^r-2k}{p^r-k}\binom{-p^r-1}{p^r-2k}\frac{4^k}{p^r+1-2k},\end{aligned}$$ where we used $\binom{n}{k}=\binom{n}{n-k}$ and $\binom{n}{k}=(-1)^k\binom{-n+k-1}k$.
In view of the paper [@PS], let $k$ and $l$ be positive integers with $k+l=p^r$ and $0<l<p^r/2$, we have $$l\binom{2l}l\binom{2k}k\equiv-2p^r\pmod{p^{r+1}},\ \ \frac{-2p^r}{l\binom{2l}l}\equiv\binom{2k}k\pmod{p^{2}}$$ and $$\begin{aligned}
\label{2kk}
\binom{2p^r-2l}{p^r-l}\equiv0\pmod{p}.\end{aligned}$$ Thus, $$\sum_{k=1}^{(p^r-1)/2}G\left(p^r,k\right)\equiv0\pmod{p^{r+2}},$$ since $$\binom{-p^r-1}{p^r-2k}=-\prod_{j=1}^{p^r-2k}\left(1+\frac{p^r}j\right)\equiv1\pmod p$$ and $$\binom{2p^r-2k}{p^r-k}\equiv0\pmod p,\ \ \ \ \frac{p^r}{p^r+1-2k}\equiv0\pmod p$$ for each $1\leq k\leq(p^r-1)/2$.
Therefore the proof of Lemma \[Gprk\] is completed.
[*Proof of Theorem \[Th10ngz\]*]{}. By the definition of $G(n,k)$, we have $$\sum_{k=1}^{p^r-1}G\left(p^r,k\right)=\sum_{k=1}^{(p^r+1)/2}G\left(p^r,k\right).$$ So combining (\[wz2\]), Lemmas \[Gprpr+12\] and \[Gprk\], we immediately obtain that for any prime $p>3$ and integer $r>0$, $$\sum_{n=0}^{p^r-1}\frac{3n+1}{(-8)^n}\binom{2n}n^3\equiv(-1)^{(p^r-1)/2}p^r\pmod{p^{r+2}}.$$ Therefore the proof of Theorem \[Th10ngz\] is complete.[$\Box$]{}
[**Acknowledgments.**]{} The author is funded by the Startup Foundation for Introducing Talent of Nanjing University of Information Science and Technology (2019r062).
[99]{} Y. G. Chen, X. Y. Xie and B. He, On some congruences of certain binomial sums, Raman. J. 40 (2016), 237–244.
V.J.W. Guo, A $q$-analogue of a Ramanujan-type supercongruence involving central binomial coefficients, J. Math. Anal. Appl. 458 (2018), 590–600.
V.J.W. Guo, $q$-Analogues of the (E.2) and (F.2) supercongruences of Van Hamme, Raman. J. 49 (2019), 531–544.
V.J.W. Guo and J.-C. Liu, Some congruences related to a congruence of Van Hamme, preprint, arxiv:1903.03766.
V.J.W. Guo and W. Zudilin, A $q$-microscope for supercongruences, adv. math, 346 (2019), 329–358.
B. He, On the divisibility properties of certain binomial sums, J. Number Theory, 147 (2015), 133–140.
D.-W. Hu and G.-S. Mao, On an extension of a van Hamme supercongruence, Raman. J. 42 (2017), 713–723.
L. Long, Hypergeometric evaluation identities and supercongruences, Pacific J. Math. 249 (2011), no 2, 405–418.
W. Magnus, F. Oberhettinger and R.P. Soni, Formulas and theorems for the special functions of mathematical physics, 3rd ed., Sringer-Verlag, New York, 1966, pp. 25–32.
G.-S. Mao, Proof of some supercongruences via the Wilf-Zeilberger method, preprint, arxiv:1909.13173.
G.-S. Mao, Proof of two supercongruences conjectured by Z.-W. Sun via the Wilf-Zeilberger method, preprint, arxiv:1910.09983.
G.-S. Mao and T. Zhang, [*Proof of Sun’s conjectures on super congruences and the divisibility of certain binomial sums*]{}, Ramanujan. J., [**50**]{} (2019), 1–11.
F. Morley, Note on the congruence $2^{4n}\equiv(-1)^n(2n)!/(n!)^2$, where $2n+1$ is a prime, Ann. Math. 9 (1895), 168–170.
R. Osburn, W. Zudilin, On the (K.2) supercongruence of Van Hamme, J. Math. Anal. Appl. 433 (2016), 706–711.
H. Pan and Z.-W. Sun, Proof of three conjectures on cogruences, Sci. China Math. 57 (2014), no. 10, 2091–2102.
Z.-H. Sun, Congruences involving Bernoulli and Euler numbers, J. Number Theory, 128 (2008), 280–312.
Z.-W. Sun, Super congruences and Euler numbers, Sci. China Math. 54 (2011), 2509–2535.
Z.-W. Sun, A refinement of a congruence result by van Hamme and mortenson, Illinois J. Math. 56 (2012), no. 3, 967–979.
Z.-W. Sun, Open conjectures on congruences, J. Nanjing Uni. math. Biquart., 36 (2019), 1–99.
L. van Hamme, Some conjectures concerning partial sums of generalized hypergeometric series, in:“$p$-adic functional analysis” (Nijmegen, 1996), 223–236, Lecture Notes in Pure and Appl. Math. 192, Dekker, 1997.
J. Wolstenholme, On certain properties of prime numbers, Quart. J. Pure Appl. Math. 5 (1862), 35–39.
W. Zudilin, Ramanujan-type supercongruences, J. Number Theory, 129 (2009), 1848–1857.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We study some properties of the randomized series and their applications to the geometric structure of Banach spaces. For $n\ge
2$ and $1<p<\infty$, it is shown that $\ell_\infty^n$ is representable in a Banach space $X$ if and only if it is representable in the Lebesgue-Bochner $L_p(X)$. New criteria for various convexity properties in Banach spaces are also studied. It is proved that a Banach lattice $E$ is uniformly monotone if and only if its $p$-convexification $E^{(p)}$ is uniformly convex and that a Köthe function space $E$ is upper locally uniformly monotone if and only if its $p$-convexification $E^{(p)}$ is midpoint locally uniformly convex.
address: 'Mathematics Department 202 Mathematical Sciences Bldg University of Missouri Columbia, MO 65211 USA '
author:
- Han Ju Lee
title: Randomized series and Geometry of Banach spaces
---
[^1]
Randomized series
=================
Let $\{r_n\}_{n=1}^\infty$ be a sequence of mutually independent, symmetric and integrable random variables on a probability space $(\Omega, \mathcal{F}, P)$ and $\{x_n\}_{n=0}^\infty$ an arbitrary sequence in a Banach space $X$. A [*randomized series*]{} $S_n$ is a vector-valued random variable defined by $$S_n = x_0 + r_1 x_1 + \cdots + r_n x_n,\ \ \ (n=0, 1,\ldots)$$
Let $\mathcal{F}_0$ be the trivial $\sigma$-algebra $\{\emptyset,
\Omega\}$ and $\mathcal{F}_k$, $(k\ge 1)$ the $\sigma$-algebra generated by random variables $\{r_i\}_{i=1}^k$. It is easy to see that the sequence of randomized series $\{S_n\}_{n=0}^\infty$ is a martingale with respect to the filtration $\{\mathcal{F}_n\}_{n=0}^\infty$. In this paper, $\mathbb{E}$ stands for the expectation with respect to the probability $P$.
We begin with the basic properties of the randomized series. For more properties of random series, see [@K].
\[propsubmartingale\] Let $\varphi$ be a convex function on $\mathbb{R}$. Then $\{\varphi(\|S_n\|)\}_{n=0}^\infty$ is a submartingale with respect to $\{\mathcal{F}_{n}\}^\infty_{n=0}$. In particular, $$\mathbb{E}[\varphi(\|S_0\|)]\le \mathbb{E}[\varphi(\|S_1\|)]\le
\mathbb{E}[\varphi(\|S_2\|)]\le \cdots.$$
In the proof, the notation $\mathbb{E}[\;\cdot\;
|\mathcal{G}]$ means the conditional expectation with respect to the sub-$\sigma$-algebra $\mathcal{G}$ and we need to show that $\mathbb{E}[\varphi(\|S_{n+1}\|)|\mathcal{F}_n]\ge \varphi(\|S_n\|)$ almost surely. By the independence and symmetry of $\{r_k\}$ we get, for almost all $\omega\in \Omega$, $$\begin{aligned}
\mathbb{E}[\varphi&(\|S_{n+1}\|)|\mathcal{F}_n](\omega) =
\int_{\Omega} \varphi(\|S_n(\omega) + r_{n+1}(t)x_{n+1}\|) \;dP(t)\\
&= \int_{\Omega} \frac{\varphi(\|S_n(\omega) + r_{n+1}(t)x_{n+1}\|)
+ \varphi(\|S_n(\omega) - r_{n+1}(t)x_{n+1}\|)}2\;dP(t)\\
&\ge \int_{\Omega} \varphi(\|S_n(\omega)\|)\; dP(t) =
\varphi(\|S_n(\omega)\|),\end{aligned}$$ which completes the proof.
\[propmonotone\] Let $\varphi$ be a convex increasing function on $[0,\infty)$ and $x, y\in X$. Then the function $\psi$ on $\mathbb{R}$ defined by $$\psi(\lambda) = \mathbb{E}[\varphi(\|x + \lambda r_1 y\|)]$$ is an increasing convex function on $[0,\infty)$ with $\psi(\lambda) =
\psi(|\lambda|)$ for every $\lambda\in \mathbb{R}$.
By the convexity of $\varphi$, we get $\psi(\lambda) = \frac12[\psi(\lambda) + \psi(-\lambda) ] \ge
\psi(0)$ for every real $\lambda$, which implies that $\lambda\mapsto \psi(\lambda)$ is increasing on $[0,\infty)$. Clearly $\lambda\mapsto \psi(\lambda)$ is an even function on $\mathbb{R}$ since $r_1$ is symmetric.
Since random variables $\{r_i\}_i$ are independent, Proposition \[propmonotone\] shows that for any two real sequences $\{\lambda_i\}_{i=1}^n$ and $\{\xi_i\}_{i=1}^n$ satisfying $|\lambda_i| \le |\xi_i|$ for $i=1, \cdots, n$, and for any $x_0,
x_1, \ldots, x_n$ in $X$, we get $$\mathbb{E}[ \varphi\|x_0+ \lambda_1 r_1x_1 + \cdots + \lambda_n r_n x_n\|] \le
\mathbb{E}[\varphi\|x_0+\xi_1 r_1 x_1 + \cdots + \xi_n r_nx_n \|
].$$
A convex function $\varphi:[0,\infty) \to \mathbb{R}$ is said to be [*strictly convex*]{} if $$\varphi(\frac{s+t}2)<\frac{\varphi(s)+\varphi(t)}2$$ holds for all distinct positive numbers $s,t$. Notice that if $\varphi$ is strictly convex and increasing on $[0,\infty)$ and $a, b$ are real numbers with $b\neq0$, then $$\mathbb{E}[\varphi(|a + r_2 b|)] =
\mathbb{E}[ \frac{\varphi(|a + r_2 b|) + \varphi(|a - r_2 b|)}2] >
\varphi(|a|).$$
Now we state our main theorem concerning the randomized series.
\[thmbaisics\] Let $\{x_i\}_{i=1}^n$ be a finite sequence in a Banach space $X$, $\varphi$ a strictly convex, increasing function with $\varphi(0)=0$ and $\{r_i\}_{i=1}^\infty$ symmetric independent random variables with $\|r_i\|_\infty =1$, $(i=1, 2, \cdots)$.
Suppose that there is a constant $\rho>0$ such that the following holds: $$\sup_{\epsilon_1 =\pm 1, \ldots, \epsilon_n =\pm 1} \|
\epsilon_1 x_1 + \cdots + \epsilon_n x_n \|\ge \|x_1\| + \rho.$$ Then there is a constant $\delta=\delta(\rho) >0$ such that $$\mathbb{E}[ \varphi(\|x_1 + r_2x_2 + \cdots + r_n x_n \|) ] \ge \varphi(\|x_1\|) +
\delta.$$ In particular, if we take $\rho_1 = \min\{\rho, 1/2\}$, then $$\delta = \min\left\{ \varphi(\frac{\rho_1}3)\prod_{i=2}^nP\{|r_i-1|< \frac{\rho_1}{3n}\},
\min_{2\le j\le n}\mathbb{E}\left[\varphi\left(\|x_1\|+r_j
\frac{\rho_1}{3n}\right)\right]-\varphi(\|x_1\|)\right\}.$$
We adapt the argument in the proof of Proposition 2.2 in [@D]. We assume that there exist $0<\rho<1/2$ and signs $\epsilon_2,
\ldots, \epsilon_n$ such that $$\|x_1 + \epsilon_2 x_2 + \cdots +\epsilon_n x_n \| \ge \|x_1\|+\rho.$$ Select a unit element $x^*$ in $X^*$ such that $x^*(x_1) =\|x_1\|$ and let $\lambda_i = x^*(x_i)$ for $1\le i\le n$. Now we shall consider two cases according to the size of $|\lambda_i|$. In the first case we suppose that $\max_{2\le i\le n} |\lambda_i|\le
\frac{\rho}{3n}$. If $|\eta_i - \epsilon_i| \le \frac{\rho}{3n}$, $(2\le i \le n)$, then $$\|x_1 +\eta_2x_2 + \cdots + \eta_n x_n \| \ge
\|x_1 +\epsilon_2 x_2+\cdots \epsilon_n x_n\| -\frac{\rho}3\ge
\|x_1\|+\frac{2\rho}3.$$ Since $|\lambda_1+\eta_2\lambda_2 + \cdots
+ \eta_n\lambda_n | \le |\lambda_1| + \frac\rho3 =\|x_1\| +
\frac{\rho}3$, we get $$\|x_1 +\eta_2x_2 + \cdots + \eta_n x_n \| \ge
|\lambda_1 +\eta_2 \lambda_2+\cdots \eta_n
\lambda_n|+\frac{\rho}3,$$ if $|\eta_i - \epsilon_i| \le
\frac{\rho}{3n}$, $(2\le i \le n)$. Let $$F = \bigcap_{j=2}^n \left\{w\in \Omega: |r_j(w) - \epsilon_j|< \frac{\rho}{3n}\right\}$$ and take $T_n = x_1 + r_2x_2 + \cdots + r_n x_n$. Then we have $$\mathbb{E}[\varphi(\|T_n\|)] = \mathbb{E}[\varphi(\|T_n\|)\chi_F] +
\mathbb{E}[\varphi(\|T_n\|)\chi_{F^c}].$$ Since $\varphi(a+b) \ge
\varphi(a) + \varphi(b)$, $(a,b\ge 0)$, $$\begin{aligned}
\mathbb{E}[\varphi(\|T_n\|)\chi_F] &\ge
\mathbb{E}[\varphi(|\lambda_1 + r_2\lambda_2 + \cdots + r_n
\lambda_n |+\rho/3)\chi_F]\\&\ge \mathbb{E}[\varphi(|\lambda_1 +
r_2\lambda_2 + \cdots + r_n \lambda_n |)\chi_F] +
\varphi(\rho/3)P(F).\end{aligned}$$ Hence $$\begin{aligned}
\mathbb{E}[\varphi(\|T_n\|)]& \ge
\mathbb{E}[\varphi(|\lambda_1 + r_2\lambda_2 + \cdots + r_n
\lambda_n |)] + \varphi(\rho/3)P(F)\\&\ge \varphi(\lambda_1)+
\varphi(\rho/3)P(F) =\varphi(\|x_1\|) +
\varphi(\rho/3)P(F).\end{aligned}$$ Notice that $P(F)=\prod_{i=2}^nP\{w\in \Omega: |r_i(w) - 1|<
\frac{\rho}{3n}\}>0$ for $r_i$’s are independent symmetric random variables with $\|r_i\|_\infty=1$, $(i=1,2,\cdots)$. In the second case we suppose that there exists $i_0$ $(2\le i_0\le n)$ such that $|\lambda_{i_0}|\ge \frac{\rho}{3n}$. It follows from Proposition \[propsubmartingale\], \[propmonotone\] and strict convexity of $\varphi$ that $$\begin{aligned}
\mathbb{E}[\varphi(\|T_n \|)] &\ge
\mathbb{E}[\varphi(|\lambda_1 + r_2 \lambda_2 +\cdots + r_n
\lambda_n |)]
\\&\ge \mathbb{E}[\varphi(|\lambda_1+ r_{i_0}\lambda_{i_0}|)]
\\&\ge \mathbb{E}\left[\varphi\left(\left|\lambda_1+
r_{i_0}\frac{\rho}{3n}\right|\right)\right]\\&>\varphi(\lambda_1)=
\varphi(\|x_1\|).\end{aligned}$$ The proof is complete.
Representability of $\ell^n_\infty$
===================================
A Banach space $Y$ is said to be [*representable*]{} in $X$ if, for each $\lambda>1$, there is a bounded linear map $T: Y \to X$ such that $\|x\| \le \|Tx\|\le \lambda \|x\|$ for every $x\in Y$. A Banach space $Y$ is said to be [*finitely representable*]{} in $X$ if every finite dimensional subspace of $Y$ is representable in $X$.
It is well-known due to the work of B. Mauray and G. Pisier [@MP] that $c_0$ is finitely representable in $X$ if and only if $c_0$ is finitely representable in $L_p(X)$ for all $1\le p
<\infty$. S. J. Dilworth considered the quatitative version of this theorem in [@D], where he showed that if $X$ is a complex Banach space, $n\ge 2$ and $0<p<\infty$, then $\ell_\infty^n(\mathbb{C})$ is representable in $X$ if and only if it is representable in $L_p(X)$. As we see in the next example it is not true for real Banach spaces.
Let $X$ be a nontrivial real Banach space. Then $\ell_\infty^2$ is representable in $L_1([0,1];X)$. Indeed, if we choose the Rademacher sequence $\{r_n = {\rm sign}( \sin2^n\pi t)\}_{n=1}^\infty$ in $L_1[0,1]$ and $x_0 \in S_X$, then $x= r_1 x_0$ and $y = r_2 x_0$ are the elements of unit sphere of $L_1(X)$ and they satisfy $$\|x+y\|_{L_1(X)} = \|x- y\|_{L_1(X)}=1,$$ which means that $\ell_\infty^2$ is representable in $L_1(X)$.
The subharmonicity of absolute value of holomorphic functions on $\mathbb{C}$ plays the crucial role in the proof in [@D]. In this paper, the strict convexity of $\varphi(t) = |t|^p$ $( 1< p
<\infty)$ on $\mathbb{R}$ plays the analogous role, and thus it is shown here that $\ell_\infty^n$ is representable in $X$ if and only if it is representable in $L_p(X)$ for every $1<p<\infty$.
The following proposition is a real version of Proposition 2.2 in [@D].
\[criterion-rep\] Let $\{r_i\}_{i=1}^\infty$ be symmetric independent random variables with $\|r_i\|_\infty =1$, $(i=1, 2, \cdots)$. Suppose that $X$ is a real Banach space and that $n\ge 2$. The following properties are equivalent:
1. $\ell_\infty^n$ is not representable in $X$.
2. There exists $\rho>0$ such that whenever $x_1,\ldots, x_n$ are unit vectors in $X$, then there exist signs $\epsilon_1, \ldots,
\epsilon_n$ such that $$\|\epsilon_1 x_1 + \cdots + \epsilon_n x_n \| \ge 1+ \rho.$$
3. There exist strictly convex, increasing function $\varphi$ on $[0,\infty)$ with $\varphi(0)=0$ and $\rho>0$ such that whenever $x_1, \ldots, x_n$ are unit vectors in $X$ then $$\mathbb{E}[ \varphi(\|x_1 + r_2x_2\cdots + r_n
x_n\|)] \ge \varphi(1)+ \rho.$$
4. For each strictly convex, increasing function $\varphi$ on $[0,\infty)$ with $\varphi(0)=0$, there is $\rho>0$ such that whenever $x_1, \ldots, x_n$ are unit vectors in $X$ then $$\mathbb{E}[ \varphi(\|x_1 + r_2x_2\cdots + r_n
x_n\|)] \ge \varphi(1)+ \rho.$$
The implications $(4)\Rightarrow(3)\Rightarrow(2)\Rightarrow(1)$ are clear. To show $(1)\Rightarrow (2)$, suppose that $(2)$ fails, so for any $\rho>0$ there exist unit vectors $x_1, \ldots, x_n$ in $X$ such that for all signs $\epsilon_1, \ldots, \epsilon_n$, we have $$\|\epsilon_1 x_1 + \cdots + \epsilon_n x_n \| < 1+ \rho.$$ It follows that for all $\lambda_1, \ldots, \lambda_n$ with $1=|\lambda_{i_0}| = \max_{1\le i\le n} |\lambda_i|$, we have $$\|\lambda_1 x_1 + \cdots + \lambda_n x_n \| < 1+\rho.$$ Hence $$\begin{aligned}
1= \|\lambda_{i_0}x_{i_0}\| &\le \frac12 \|\lambda_1 x_1 +
\cdots + \lambda_n x_n \| + \frac12 \|\lambda_1x_1 + \cdots +
\lambda_n x_n -2\lambda_{i_0}x_{i_0}\|\\ &\le \frac12 \|\lambda_1x_1
+ \cdots + \lambda_n x_n \| + \frac12(1+\rho).\end{aligned}$$ Thus, $\|\lambda_1x_1 +\cdots + \lambda_n x_n\|\ge 1-\rho$. Since $\rho$ is arbitrary, it follows that $\ell_\infty^n$ is representable in $X$. Now we have only to show that $(2)\Rightarrow
(4)$. Suppose that $(2)$ holds and that $\varphi$ is a strictly convex, increasing function on $[0, \infty)$ with $\varphi(0)=0$. There is $0<\rho<1/2$ such that whenever $x_1, \ldots, x_n$ are unit vectors in $X$, there exist signs $\epsilon_2, \ldots, \epsilon_n$ such that $$\|x_1 + \epsilon_2 x_2 + \cdots \epsilon_n x_n \| \ge 1+\rho.$$ Then Theorem \[thmbaisics\] shows that $(2)$ implies $(4)$.
Notice that in the case of the Rademacher sequence $\{r_n\}$, for every $x_1, \ldots, x_n$ in $X$, $$\mathbb{E}[\varphi(\|r_1 x_1 + \cdots + r_n x_n \|)] =
\mathbb{E}[\varphi(\|x_1 + r_2 x_2 + \cdots + r_n x_n\|)].$$ The following theorem shows the lifting property of representability of $\ell_\infty^n$.
Suppose $X$ is a Banach space, $(M, \mathfrak{M}, \mu)$ is a measure space with a measurable subset $A$ satisfying $0<\mu(A)<\infty$, and $1<p<\infty$, $n\ge2$. Then $\ell_\infty^n$ is representable in $X$ if and only if it is representable in $L_p(M, \mathfrak{M}, \mu;
X)$.
One implication is clear. To prove the other implication, suppose that $\ell_\infty^n$ is not representable in $X$ and let $\{r_n\}$ be the Rademacher sequence. By Proposition \[criterion-rep\], there exits $0<\rho<1/2$ such that whenever $x_1, \ldots, x_n$ are unit vectors in $X$, we have $$\mathbb{E}\|r_1x+r_2 x_2 + \cdots + r_n x_n \|^p \ge (1+\rho)^p.$$ Suppose that $f_1, \ldots, f_n$ are unit vectors in $L_p(X)$. We define the following functions on $M$. For $w\in M$, let $$\begin{aligned}
q(w)^p &=
\mathbb{E}\|r_1f_1(w) + \cdots + r_n f_n(w)\|^p,\\
M(w) &= \max\{ \|f_i(w) \| : 1\le i\le n\},\\
m(w) &= \min\{ \|f_i(w) \| : 1\le i\le n\}.\end{aligned}$$ By Proposition \[propsubmartingale\], $q(w) \ge M(w)$ for all $w\in M$. Now the argument divides into two cases according to the relative sizes of $M(w)$ and $m(w)$. In the first case we suppose that $(1-\rho/3)M(w)\ge m(w)$. Then $$\begin{aligned}
\frac1n \sum_{i=1}^n \|f_i(w)\|^p &\le \frac{n-1}n M(w)^p + \frac 1n
m(w)^p\\& \le \left( 1- \frac \rho{3n} \right) M(w)^p\\&\le q(w)^p -
\frac \rho{3n} \left( \frac 1n \sum_{i=1}^n
\|f_i(w)\|^p\right)\end{aligned}$$ and so $$q(w)^p \ge \left( 1+ \frac \rho{3n}\right) \frac{1}n \sum_{i=1}^n
\|f_i(w)\|^p.$$ In the second case, we suppose that $(1-\rho/3)M(w)<
m(w)$. Then $$\begin{aligned}
q(w)^p &\ge (1+\rho)^p m^p(w)\\&\ge
(1+\rho)^p\left(1-\frac\rho3\right)^p \frac 1n \sum_{i=1}^n
\|f_i(w)\|^p\\&\ge \left(1+\frac\rho2\right)^p\frac 1n \sum_{i=1}^n
\|f_i(w)\|^p.\end{aligned}$$ Hence by the Fubini theorem, $$\mathbb{E}\|rf_1+\cdots+r_nf_n\|_{L^P(X)}^p=\int_{M} q(w)^p\
d\mu \ge \min\left\{\left(1+\frac \rho2\right)^p, \left(1+
\frac\rho{3n}\right)\right\},$$ which shows that $\ell_\infty^n$ is not representable in $L_p(X)$ by Proposition \[criterion-rep\]. The proof is completed.
Applications to the convexity of Banach spaces
==============================================
Recall that a point $x$ in $S_X$ is an [*extreme point*]{} of $B_X$ if $\max\{\|x+y\|, \|x-y\|\} =1$ for some $y\in X$ implies $y=0$. A point $x\in S_X$ is called a [*strongly extreme point*]{} of $B_X$ if, given $\epsilon>0$, there is a $\delta= \delta(x,\epsilon)>0$ such that $$\inf\{ \max\{\|x+y\|, \|x-y\|\} : \|y\|\ge \epsilon\} \ge 1 + \delta.$$ A Banach space is said to be [*strictly convex*]{} (resp. [*midpoint locally uniformly convex*]{}) if every point of $S_X$ is (resp. strongly) extreme point of $B_X$. A Banach space is called [*uniformly convex*]{} if, given $\epsilon>0$, there is a $\delta=\delta(\epsilon)>0$ such that $$\inf\{ \max\{\|x+y\|, \|x-y\|\} : \|y\|\ge \epsilon, \|x\|=1\} \ge 1 +
\delta.$$
Theorem \[thmbaisics\] gives the following criteria for the various convexity properties.
\[convexity\] Let $X$ be a real Banach space and $\varphi$, a strictly convex increasing function on $[0, \infty)$ with $\varphi(0)=0$ and $r$, a symmetric random variable with $\|r\|_\infty=1$. Then
1. A point $x$ in $S_X$ is an extreme point of $B_X$ if and only if $\mathbb{E}[\varphi(\|x+ r y\|)] =\varphi(1)$ for $y\in X$ implies $y=0$.
2. A point $x$ in $S_X$ is a strongly extreme point of $B_X$ if and only if for every $\epsilon>0$ there is a $\delta>0$ such that whenever $\|y\|\ge \epsilon$, we get $$\mathbb{E}[\varphi(\|x+
r y\|)] \ge \varphi(1) +\delta.$$
3. \[3uniformconvex\] $X$ is uniformly convex if and only if the modulus $\delta_\varphi(\epsilon)>0$ for every $\epsilon>0$, where $
\delta_\varphi(\epsilon) = \inf\{ \mathbb{E}[\varphi(\|x+ r
y\|)]-\varphi(1) : x\in S_X, \|y\|\ge \epsilon \}.$
We prove only (\[3uniformconvex\]) because the proof of the others are similar. Suppose that $X$ is uniformly convex. Given $\epsilon>0$, there is a $\rho>0$ such that for any $x\in S_X$ and $y\in X$ with $\|y\|\ge \epsilon$, we have $$\max\{\|x+y\|, \|x-y\|\}\ge 1 + \rho.$$ Then Theorem \[thmbaisics\] shows that there is $\delta>0$ such that for any $x\in S_X$ and $y$ with $\|y\|\ge \epsilon$, $$\mathbb{E}[\varphi(\|x+r y\|)]\ge\varphi(1) + \delta.$$ Conversely, suppose that $\delta_\varphi(\epsilon)>0$ for every $\epsilon>0$. Then given $\epsilon>0$ for any $x\in S_X$ and $y\in
X$ with $\|y\|\ge \epsilon$, we have $$\max\{\varphi(\|x+y\|), \varphi(\|x-y\|)\} \ge \max_{-1\le t\le 1}
\{ \varphi(\|x+ty\|)\} \ge \mathbb{E}[\varphi(\|x+r y\|)].$$ Since $\varphi$ is strictly increasing we get, for any $x\in S_X$ and $y\in X$ with $\|y\|\ge \epsilon$, $$\max\{\|x+y\|, \|x-y\|\}\ge
\varphi^{-1}(1+\delta_\varphi(\epsilon))>1.$$ Therefore $X$ is uniformly convex and this completes the proof.
It is worthwhile to notice that Theorem \[convexity\] does not hold if we consider the general increasing convex function $\varphi$ with $\varphi(0)=0$. Indeed, it is easily checked that if $\varphi(t)=|t|$ and $\{r_n\}_n$ is the Rademacher sequence then for any nontrivial Banach space $X$, $$\mathbb{E}\|x + r_1 x\|=1 \ \ \ (x\in S_X).$$ Consequently, we cannot characterize the extreme point of $B_X$ with $\varphi(t)=t$.
We shall discuss the uniform convexity of $p$-convexification $E^{(p)}$ for uniformly monotone Banach lattice $E$. For more details on Banach lattices, order continuity and Köthe function spaces, see [@LT]. For the definition of $p$-convexification $E^{(p)}$ of $E$ and the addition $\oplus$ and multiplication $\odot$ there, see [@Lee2; @LT]. A Banach lattice is said to be [*uniformly monotone*]{} (resp. [*upper locally uniformly monotone*]{}) if given $\epsilon>0$ $$M_p(\epsilon) =\inf \{ \|(|x|^p + |y|^p)^{1/p}\|-1: \|y\|\ge
\epsilon, \|x\|=1\}>0$$ $$(\text{resp.}\ \ \ \ N_p(\epsilon;x)=\inf\{\| (|x|^p +
|y|^p)^{1/p} \|-1:\|y\|\ge \epsilon\}>0\ \ \ )$$ for some $1\le
p<\infty$. It is shown in [@Lee2; @Lee] that, given $\epsilon>0$ and $1\le p<\infty$ there is a $C_p>0$ such that for every $\epsilon>0$, $$\label{eqlocalmonotone1}C_p^{-1}
M_1(C_p^{-1}\epsilon^p) \le M_p(\epsilon) \le
M_1(\epsilon).$$
In the case when $E$ is an order continuous Banach lattice or a Köthe function space, we also get the following relations by Lemma 2.3 in [@Lee]: There is a $C_p>0$ such that every $x\in S_X$ and $\epsilon>0$, $$\label{eqlocalmonotone2}C_p^{-1}
N_1(C_p^{-1}\epsilon^p;x) \le N_p(\epsilon;x) \le
N_1(\epsilon;x).$$
Notice that relations (\[eqlocalmonotone1\]), (\[eqlocalmonotone2\]) show that if a Banach lattice $E$ is uniformly monotone then $E^{(1/p)}$ is uniformly monotone quasi-Banach lattice for $1<p<\infty$. Similarly, if $E$ is upper locally uniformly monotone order continuous Banach lattice or Köthe function space, then $E^{(1/p)}$ is also upper locally uniformly monotone quasi-Banach lattice for $1<p<\infty$ (cf. [@Lee2]). The characterizations of local uniform monotonicity of various function spaces have been discussed in [@FK].
In [@HKM], H. Hudzik, A. Kamińska and M. Mastyło showed that if a Köthe function space $E$ is uniformly monotone then its $p$-convexification $E^{(p)}$ is uniformly convex for $1<p<\infty$. A partial generalization of this result has been studied by the author in [@Lee2], where it was shown that if a Banach lattice is uniformly monotone then $E^{(p)}$ is uniformly convex for all $2\le p<\infty$. In the next theorem, the gap is completed.
\[thmmonotonetoconvex\] Let $E$ be a Banach lattice. The following statements are equivalent.
1. \[item1\] $E$ is uniformly monotone.
2. \[item2\] $E^{(p)}$ is uniformly convex for all $1<p<\infty$.
3. \[item3\] $E^{(p)}$ is uniformly convex for some $1<p<\infty$.
Proposition 4.4 in [@Lee2] shows that uniformly convex Banach lattice is uniformly monotone. So if we assume (\[item3\]), then $E^{(p)}$ is uniformly monotone and $E$ is uniformly monotone. Hence (\[item3\]) $\Rightarrow$ (\[item1\]) is proved. The implication (\[item2\]) $\Rightarrow$ (\[item3\]) is clear. So we have only to show that (\[item1\]) implies (\[item2\]).
We shall use Theorem \[convexity\] (3) with the Rademacher function $|r|=1$. Let $\epsilon>0$ and let $f, g\in E^{(p)}$ with $\|f\|_{E^{(p)}}=\|f\|^{1/p}_E= 1$ and $\|g\|_{E^{(p)}}=
\|g\|_E^{1/p}\ge \epsilon$. Recall the following well-known inequality (cf. Lemma 4.1 [@Lee2]) : for any $1<p<\infty$ there is $C=C(p)$ such that for any reals $s,t$, $$\left( \left| \frac{s-t}{C}\right|^2 + \left|
\frac{s+t}{2}\right|^2\right)^{\frac 12} \le \left(\frac{|s|^p+
|t|^p}2\right)^{\frac 1p}.$$ Then applying the Krivine functional calculus to the inequality above, we get $$\begin{aligned}
\label{eqlatticeconvex1} \mathbb{E}\left[\|f \oplus (r\odot
g)\|_{E^{(p)}}^p \right]&= \mathbb{E}[\|\ |f^{1/p}+ r g^{1/p}|^p
\|_{E}]\\&= \frac{\mathbb{E}\|\ |f^{1/p}+rg^{1/p}|^p\|_E +
\mathbb{E}\|\ |f^{1/p}-rg^{1/p}|^p\|_E}2\nonumber\\
&\ge \mathbb{E}\left[ \left\|\frac{|f^{1/p}+rg^{1/p}|^p +
|f^{1/p}-rg^{1/p}|^p}2\right\|_{E}\right]\nonumber\\& \ge
\mathbb{E}\left\|\left(|f|^{2/p} + \frac
{|2g|^{2/p}}{C^{2/p}}\right)^{p/2}\right\|_{E}\nonumber\\&=
\left\|\left(|f|^{2/p} + \frac
{|2g|^{2/p}}{C^{2/p}}\right)^{p/2}\right\|_{E}\nonumber.\end{aligned}$$
By (\[eqlatticeconvex1\]), if $1<p\le 2$, then $\mathbb{E}\left[\|f \oplus (r\odot g)\|_{E^{(p)}}^p \right]\ge 1 +
M_{2/p}(2\epsilon^p/C)$. In the case of $2<p<\infty$, (\[eqlatticeconvex1\]) shows that $$\begin{aligned}
\mathbb{E}\left[\|f \oplus (r\odot g)\|_{E^{(p)}}^p
\right] &\ge \left\|\left(|f|^{2/p} + \frac
{|2g|^{2/p}}{C^{2/p}}\right)^{p/2}\right\|_{E}\\
&\ge \left\|\left(|f| + \frac {|2g|}{C}\right)\right\|_{E}\ge 1+
M_1(2\epsilon^p/C).\end{aligned}$$ Hence $$\mathbb{E}[ \|f\oplus (r\odot g) \|_{E^{(p)}}^p] \ge 1+M_{\max\{1, 2/p \}}(2\epsilon^p/C)$$ completes the proof.
Now we discuss the the local version of Theorem \[thmmonotonetoconvex\]. A point $x\in S_X$ in a complex Banach space $X$ is said to be a [*complex strongly extreme point*]{} if there is $0<p<\infty$ such that given $\epsilon>0$, $$H_p(\epsilon; x) =\inf \left\{\left( \frac{1}{2\pi} \int_0^{2\pi} \|x+
e^{i\theta}y\|^p\ d\theta\right)^{1/p}-1: \|y\|\ge \epsilon \right\}
> 0.$$
It is known in [@DHM] that $x\in S_X$ is a complex strongly extreme point if and only if for every $\epsilon>0$, $$H_\infty(\epsilon; x) = \inf\{ \max_{0\le \theta\le
2\pi}\|x+e^{i\theta}y\|-1: \|y\|\ge \epsilon \}>0.$$ For more details about these moduli, see [@DGT; @D; @DHM]. A complex Banach space $X$ is said to be [*locally uniformly complex convex*]{} if every point of $S_X$ is a complex strongly extreme point.
Let $E$ be an order continuous Banach lattice or a Köthe function space. Then the following are equivalent:
1. $E$ is upper locally uniformly monotone.
2. $E^{\mathbb{C}}$ is locally uniformly complex convex.
3. $E^{(p)}$ is midpoint locally uniformly convex for all $1<p<\infty$.
4. $E^{(p)}$ is midpoint locally uniformly convex for some $1<p<\infty$.
First we prove the equivalence of (1) and (2). We shall use a similar argument as in the proof of [@CHL Proposition 3.7] in the sequence space. Suppose that $E$ is locally uniformly complex convex. Then for each $x\in S_E$ and $\epsilon>0$ there is $\delta=\delta(x,\epsilon)>0$ such that for all $y \in X$ with $\|y\|\ge \epsilon$ $$\|\ |x| + |y|\ \|\ge \frac 1{2\pi} \int_0^{2\pi} \|x+ e^{i\theta} y\| \; d\theta \ge 1 + \delta.$$ So $X$ is upper locally uniformly monotone.
For the converse, suppose that $E$ is upper locally uniformly monotone. Now, if we use Theorem 7.1 in [@DGT], then we have for every pair $x,y$ in $E$, $$\frac{1}{2\pi}\int_0^{2\pi}\|x+e^{i\theta} y\|\; d\theta
\ge \left\| \left(|x|^2 + \frac12|y|^2\right)^{1/2}\right\|.$$ Hence for every $x\in S_X$ and $\epsilon>0$, we get $$\frac{1}{2\pi}\int_0^{2\pi}\|x+e^{i\theta} y\|\; d\theta
\ge 1+ N_2(\epsilon/\sqrt{2};x).$$ Therefore, the upper local uniform monotonicity of $E$ implies the local uniform complex convexity of $E$.
For (1)$\Rightarrow$(3), fix $f$ with $\|f\|_{E^{(p)}}=
\|f\|^{1/p}=1$ and for any $g\in E^{(p)}$ with $\|g\|_{E^{(p)}}=
\|g\|^{1/p}\ge\epsilon$, (\[eqlatticeconvex1\]) holds. Hence $$\mathbb{E}[ \|f\oplus (r\odot g) \|_{E^{(p)}}^p] \ge 1+N_{\max\{1, 2/p \}}(2\epsilon^p/C;f),$$ which shows that (1)$\Rightarrow$(3) holds.
The implication (3)$\Rightarrow$(4) is clear. Finally assume that (4) holds. Note that every midpoint locally uniformly convex Banach lattice is upper locally uniformly monotone. Indeed, if $x\in S_X$ and $\epsilon$, there is $\delta>0$ such that $$1+\delta \le \max\{ \|x+y\|, \|x-y\|\} \le \| \ |x| + |y|\ \|.$$ Since the midpoint local uniform convexity of $E^{(p)}$ implies the upper local uniform monotonicity of $E^{(p)}$, $E$ is upper locally uniformly monotone. This completes the proof.
Let $X$ be a real Banach space and $\Delta$ be the open unit disk in $\mathbb{C}$. Let $(f_n)$ be a sequence of continuous functions from $\Delta$ into $X$ and $f:\Delta\to X$ be continuous. We say that $(f_n)$ converges to $f$ with respect to the topology of norm uniform convergence on compact subsets of $\Delta$ if $\lim_{n\to
\infty} \sup\{\|f_n(z)- f(z)\|: z\in K \}=0$ for all compact subsets $K$ of $\Delta$. We will denote by $\beta$ the topology of norm uniform convergence on compact subsets of $\Delta$.
A Banach space $X$ is said to have [*Kadec-Klee property*]{} with respect to topology $\tau$ $(KK(\tau))$ if whenever $(x_n)$ is a sequence in $X$ and $x\in X$ satisfy $\|x_n\|=\|x\|=1$ for all $n\in
\mathbb{N}$ and $\tau$-$\lim_n x_n = x$, then $\lim_n\|x_n - x\|=
0$.
A function $f:\Delta\to X$ is [*harmonic*]{} if $f$ is twice continuously differentiable and if the Laplacian of $f$ is zero. It is known [@H] that $f:\Delta\to X$ is harmonic if and only if $x^*f$ is harmonic for all $x^*\in X^*$ if and only if there is a sequence $\{a_n\}_n\subset X$ so that for all $0\le r<1$, $\theta\in
\mathbb{R}$, $$f(re^{i\theta}) = \sum_{n=-\infty}^\infty a_n r^{|n|}e^{in \theta}
,$$ where the series is absolutely and locally uniformly convergent.
We now define $h^p(\Delta; X)$ for $1<p<\infty$ by $$h^p(\Delta; X) = \{ f:\Delta\to X : f\ \text{is harmonic and}\
\|f\|_p<\infty\},$$ where $$\|f\|_p = \sup_{0\le r<1} \left( \frac{1}{2\pi}\int_0^{2\pi} \|f(re^{i\theta})\|^p
\;d\theta \right)^{1/p}.$$ It is easy to see that on $h^p(\Delta;
X)$, $\|\cdot\|_p$ is $\beta$-lower semicontinuous function.
It is shown by P. N. Dowling and C. J. Lennard [@DL] that if $h^p(\Delta; X)$ has $KK(\beta)$, then $X$ is strictly convex and has the Radon-Nikodým property.
In fact, the following proposition is a consequence of the results in [@DHS]. We present an easy proof.
If $h^p(\Delta; X)$ has $KK(\beta)$, then $X$ is midpoint locally uniformly convex.
Suppose that $X$ is not locally uniformly convex. Then applying Theorem \[convexity\] with $r(\theta) = \cos\theta$, there exist an $\epsilon>0$, a sequence $(x_n)$ in $X$ and $x\in S_X$ such that $\|x_n\|\ge \epsilon$ and $$\lim_{n\to \infty} \frac{1}{2\pi}\int_0^{2\pi}
\|x+(\cos \theta) x_n\|^p \;d\theta= 1.$$ Define $f_n: \Delta\to X$ by $$f_n(z) = x + \frac 12 (z^n + \bar{z}^n)x_n$$ and $f:\Delta\to X$ by $f(z)=x$. Then it is easy to see that $\beta$-$\lim_n f_n(z) = f(z)$. Notice that $$\begin{aligned}
\|f_n\|^p_p & =\sup_{0\le r<1} \frac{1}{2\pi}\int_0^{2\pi} \|x +
r^n\cos (n\theta)x_n \|^p \;d\theta \\ &=
\frac{1}{2\pi}\int_0^{2\pi} \|x + \cos\theta x_n\|^p \;d
\theta.\end{aligned}$$ Then $\lim_n \|f_n\|_p = 1=\|f\|_p$. However $$\begin{aligned}
\|f_n - f\|_p^p & =\sup_{0\le r<1} \frac{1}{2\pi}\int_0^{2\pi} \|r^n\cos (n\theta)x_n
\|^p \;d \theta \\ &= \frac{\|x_n\|^p}{2\pi} \int_0^{2\pi}
|\cos\theta|^p \; d \theta\ge \frac{\epsilon^p}{2\pi} \int_0^{2\pi}
|\cos\theta|^p \; d \theta.\end{aligned}$$ Hence $h^p(\Delta; X)$ fails to have $KK(\beta)$. The proof is complete.
It is worthwhile to remark here that it has been shown in [@DHS] that $h^p(\Delta; X)$ has $KK(\beta)$ if and only if $X$ has the Radon-Nikodým property and every element of $S_X$ is a denting point of $B_X$, which is called [*property $(G)$*]{}. It is easy to see that a Banach space with property $(G)$ is midpoint locally uniformly convex.
Acknowledgment {#acknowledgment .unnumbered}
--------------
The author thanks A. Kamińska for useful comments.
[10]{}
Y. S. Choi, K. H. Han and H. J. Lee, *Boundaries for algebras of holomorphic functions on Banach spaces*, Illinois J. Math. to appear.
W. Davis, D. J. H. Garling and N. Tomczak-Jagermann, *The complex convexity of quasi-normed linear spaces*, J. Funct. Anal. **55** (1984), 110–150.
S. J. Dilworth, *Complex convexity and the geometry of Banach spaces*, Math. Proc. Cambridge Philos. Soc. **99** (1986), 495–506.
P. N. Dowling, Z. Hu and D. Mupasiri, *Complex convexity in Lebesgue-Bochner function spaces*, Trans. Amer. Math. Soc., **348** (1996), 127–139.
P. Dowling, Z. Hu and M. A. Smith, *Geometry of spaces of vector-valued harmonic functions*, Canad. J. Math. **46** (1994), no. 2, 274–283.
P. N. Dowling and C. J. Lennard, *Kadec-Klee properties of vector-valued Hardy spaces*. Math. Proc. Cambridge Philos. Soc. **111** (1992), no. 3, 535–544.
P. Foralewski and P. Kolwicz, *Local uniform rotundity in Calderón-Lozanovskiĭ spaces*, Preprint.
W. Hensgen,*Hardy-Raüme vektorwertiger Funktionen*. Dissertaion, Munich 1986.
H. Hudzik, A. Kamińska and M. Mastyło, *Geometric properties of some Calderon-Lozanovskiĭ spaces and Orlicz-Lorentz spaces*, Houston J. Math. **22** (1996), no. 3, 639–663.
J. P. Kahane, *Some random series of functions*. Second edition. Cambridge Studies in Advanced Mathematics, 5. Cambridge University Press, Cambridge, 1985.
H. J. Lee, *Complex convexity and monotonicity in quasi-Banach lattices*, Israel J. Math. **159** (2007), 57–91.
H. J. Lee, *Monotonicity and complex convexity in Banach lattices*, J. Math. Anal. Appl. **307** (2005), 86–101.
J. Lindenstrauss and L. Tzafriri, *Classical Banach spaces II*, Springer-Verlag, 1979.
B. Maurey and G. Pisier, *Séries de variables aléatoires vectorielles indépendantes et propriétés géométriques des espaces de Banach*, Studia Math. **58** (1976), no. 1, 45–90.
[^1]: This work was supported by grant No. R01-2004-000-10055-0 from the Basic Research Program of the Korea Science & Engineering Foundation
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
Tyler Seacrest[^1]\
`[email protected]`
bibliography:
- '2ndNeighborhood.bib'
title: 'The Arc-Weighted Version of the Second Neighborhood Conjecture'
---
Introduction
============
Unless otherwise noted, all digraphs in this paper are oriented simple graphs, and thus do not contain loops or two-cycles. We wil use $V(D)$ to denote the set of vertices of a digraph $D$, $A(D)$ to denote the set of arcs or edges.
Given a digraph $D$ and vertices $u$ and $v$, we call $u$ an *$n$th out-neighbor* of $v$ if the shortest directed path connecting $v$ to $u$ has $n$ edges. Let $N_n^+(v)$ be the set of all $n$th out-neighbors of $v$. We will focus on $N_1^+(v)$ and $N_2^+(v)$, and we note that these are disjoint. We will use $N_1^-(v)$ and $N_2^-(v)$ to refer to the sets of first and second in-neighbors, defined analogously to out-neighbors. If not specified, the term *neighbors* refers to first out-neighbors.
If $|N_1^+(v)| \leq |N_2^+(v)|$, we will call $v$ a *weakly expanding vertex* or a *Seymour vertex*. If $v$ is not Seymour vertex, then $|N_1^+(v)| > |N_2^+(v)|$ and we say $v$ is *strongly contracting*.
Seymour made the following conjecture, which has become known as Seymour’s Second Neighborhood Conjecture.
\[conj:original\] Every digraph without loops or two-cycles contains a Seymour vertex.
Early work on this conjecture focused on tournaments, which was known as Dean’s Conjecture. This important special case was proven by Fisher [@Fisher96] using the losing density of a digraph. A weight function $\ell: V \to [0, 1]$ is a *losing density* if the weights sum to one and every vertex has more weight on its out-neighbors than its in-neighbors. That is, $\sum_{v \in V} \ell(v) = 1$ and for every vertex $u$, $\sum_{v \to u} \ell(v) \leq \sum_{u \to v} \ell(v)$. He first proved that every digraph (not just tournaments) contained a losing density using Farka’s lemma (see [@BertsimasTsitsiklis97]) from linear programming. He then used the existence of a losing density on a tournament to show that every tournament has a Seymour vertex.
Later, Havet and Thomassé [@HavetThomasse00] gave a short and elegant proof of Fisher’s Theorem using median orders. A *median order* is an ordering of the vertices $v_1, \ldots, v_n$ of a digraph that minimizes the number of pairs $v_i \to v_j$ where $j < i$. They proved that for a tournament, $v_n$ in a median order is a Seymour vertex. They went on to show the existence of two Seymour vertices in the case the tournament had no sink.
Several other partial results are known. Kaneko and Locke [@KanekoLocke01] proved Seymour’s conjecture for graphs with minimum out-degree six or less. Chen, Shen, and Yuster [@ChenShenYuster03] show every oriented graph contains a vertex $v$ such that $|N_1^+(v)| > \gamma |N_2^+(v)|$ for $\gamma \approx 0.657298$ is the real root of the equation $2 x^3 + x^2 - 1 = 0$. Fidler and Yuster [@FidlerYuster07] introduced the vertex-weighted version of the conjecture, and proved, among other results, that the conjecture is true for orientations of a complete graph minus a matching, which generalized Fisher’s Theorem. Ghazal [@Ghazal12], also used vertex-weighted digraphs, and also extended Fisher’s theorem. In this case, the author showed that the conjecture holds for orientations of generalized stars or threshold graphs.
We examine the second neighborhood conjecture for arc-weighted digraphs, defined in Section \[sec:arc-weighted\]. The arc-weighted version is equivalent to the original conjecture, as shown in Proposition \[prop:equiv\]. For tournaments, however, the arc-weighted version and original version are not equivalent, because the auxiliary graph used in the proof of Proposition \[prop:equiv\] is not a tournament. Our main result, given in Section \[sec:tournaments\], shows that the arc-weighted tournaments do in fact satisfy the second neighborhood conjecture. In fact, we do not explicitly use Fisher’s theorem in our proof, and thus our proof can be seen as an alternate proof of Fisher’s Theorem, albeit one that is not as radically different as Havet and Thomassé’s proof. Along the way, in Section \[sec:vertex-weighted\], we will develop some results related to vertex-weighted digraphs.
Vertex-Weighted Digraphs {#sec:vertex-weighted}
========================
Seymour’s second neighborhood conjecture can be expanded to vertex-weighted digraphs, which are digraphs that have a weight function $\eta$ which assigns each vertex a nonnegative real number. This extends naturally to a weight function on sets of vertices $S$, where $\eta(S) = \sum_{v \in S} \eta(v)$. We define the *$n$th neighborhood weight* of a vertex $v$ to be $\eta(N_n^+(v))$. In a weighted digraph, $v$ is *weakly expanding* if $\eta(N_1^+(v)) \leq \eta(N_2^+(v))$, and is *strongly contracting* if $\eta(N_1^+(v)) > \eta(N_2^+(v))$. A weakly expanding vertex is called a *Seymour vertex*.
Seymour’s conjecture then becomes
Every weighted digraph without loops or two-cycles contains a Seymour vertex.
Note that while this may first appear to be a stronger conjecture, it is implied by the original version.
Like Fisher, we will use Farkas’ Lemma.
\[lemma:farkas\] For any matrix $A$ and vector $\mathbf{b}$, exacly one of the following holds.
1. $A \mathbf{x} = \mathbf{b}$ for some $\mathbf{x}$ such that $\mathbf{x} \geq 0$.
2. $\mathbf{p}^T A \geq 0$ for some $\mathbf{p}$ such that $\mathbf{p}^T \mathbf{b} < 0$.
Let the *reverse* of $D$, denoted $\overleftarrow{D}$, be the digraph with vertex set $V(D)$ and edge set $\{ vu \mid uv \in A(D)\}$. Using Farkas’ lemma we show the following.
\[thm:expanding-or-contracting\] Given any digraph $D$ and its reverse $\overleftarrow{D}$, there is either a vertex-weighting of $D$ where each vertex is weakly expanding or a vertex-weighting of $\overleftarrow{D}$ where each vertex is strongly contracting, and not both.
Let $D$ be a digraph on vertex set $v_1 \ldots v_n$. Let $N$ be a matrix with entry $n_{ij}$ in the $i$th row and $j$th column, where $n_{ij}$ is $-1$ if $v_j$ is a first out-neighbor of $v_i$, and $n_{ij} = 1$ if $v_j$ is a second out-neighbor of $v_i$. Let $A$ be the matrix $$A = \begin{bmatrix} N & -I \\ 1 \cdots 1 & 0 \cdots 0\end{bmatrix}.$$ Let $\mathbf{x}^T = (x_1, \ldots, x_n, s_1, \ldots, s_n)$ and let $\mathbf{b}^T = (0, \ldots, 0, 1)$. Then $A \mathbf x = \mathbf b$ for $\mathbf x \geq 0$ has a solution if and only if there are non-trivial vertex-weights $x_1, \ldots, x_n$ such that every vertex has as much weight in its second neighborhood then its first. In other words, every vertex is weakly expanding. Let $\mathbf{p}^T = (p_1, \ldots, p_{n+1})$. Then the system $\mathbf{p}^T A \geq 0$, $\mathbf p^T \mathbf b < 0$ means that there are nonpositive weights $p_1, \ldots, p_n$ of the vertices of $D$ such that $\eta(N_1^-(v)) < \eta(N_2^-(v))$ for all $v$. If we use the nonnegative weights $-p_1, \ldots, -p_n$ instead, we have $\eta(N_1^-(v)) > \eta(N_2^-(v))$. In other words, there is a weighting of the vertices such that every vertex in $\overleftarrow{D}$ is strongly contracting. Farkas’ lemma said that exactly one of these must hold for every digraph, which gives the result.
The second neighborhood conjecture says it is impossible for every vertex of $\overleftarrow{D}$ to be strongly contracting, and hence we have that the following is equivalent to the second neighborhood conjecture.
Every digraph without loops or two-cycles has a vertex-weighting such that every vertex is weakly expanding.
Theorem \[thm:expanding-or-contracting\], while not explicitly stated or proved previously, was strongly inspired by Fisher [@Fisher96], especially his use of the Farkas’ lemma. Also, Fisher’s main result can be restated as
\[thm:expanding\] Every tournament $D$ has a vertex-weighting such that every vertex in $\overleftarrow{D}$ is weakly expanding. (In particular, this is achieved by a losing density.)
Fisher then concluded by a simple calculation that $D$ has a Seymour vertex. We could also go from Theorem \[thm:expanding\] to the existence of a Seymour vertex immediately by applying Theorem \[thm:expanding-or-contracting\].
Another consequnce of Theorem \[thm:expanding-or-contracting\] is that a minimum counterexample to the vertex-weighted second neighborhood conjecture must have a dual counterexample. That is, consider a minimum digraph $D$ such that $D$ has a vertex-weight function $\eta$ so that it is a counterexample to the second neighborhood conjecture. There must also be a weight function $\eta'$ so that $\overleftarrow{D}$ is a counterexample: otherwise, one could find an expansion weighting of $D$ by Theorem \[thm:expanding-or-contracting\] and subtract it from the counterexample weighting until a vertex reaches weight zero. Once there is a vertex of weight zero, deleting that vertex yields a smaller counterexample.
We will end this section with a result that we will need later on, showing the second neighborhood conjecture is true for vertex-weighted digraphs where if every arc is contained in a directed triangle. This result is along the same lines as one of the results by Brantner, Brockman, Kay, and Snively [@BrantnerBrockmanKaySnively09], who showed the second neighborhood conjecture holds for digraphs without a transitive triangle.
\[prop:triangle-full\] Let $D$ be a vertex-weighted digraph such that every edge is contained in a directed triangle. Then $D$ satisfies the second neighborhood conjecture.
By way of contradiction, suppose every edge of $D$ is contained in a directed triangle, and yet $D$ does not satisfy the second neighborhood conjecture. For any $v \in V(D)$, $N_1^-(v) \subseteq N_2^{+}(v)$. Thus we have the string of inequalitites $$\sum_{v \in V(D)} \eta(N_1^-(v)) \leq \sum_{v \in V(D)} \eta(N_2^{+}(v)) < \sum_{v \in V(D)} \eta(N_1^+(v)) = \sum_{v \in V(D)} \eta(N_1^-(v)),$$ which yields a contradiction.
Arc-Weighted Digraphs {#sec:arc-weighted}
=====================
Seymour’s second neighborhood conjecture can also be expanded to arc-weighted digraphs, which are digraphs with a weight function $w$ assigning each arc and nonnegative real number. Given a vertex $v$, the *first neighborhood weight* of $v$ is denoted $\alpha^D_v$ and is defined as $\sum_{u \in N_1^+(v)} w(vu)$.
We will now define the second neighborhood weight of $v$, which will be denoted $\beta^D_v$. Let $s$ be a vertex at the end of a path of length 2 starting at $v$. Define $\beta_v(s)$ to be the maximum over all $u$ such that $v \to u \to s$ of $w(us) - w(vs)$, and $\beta_v(s)$ is $0$ if this maximum is negative. Notationally, $$\beta_v(s) = \max \left( \{0\} \cup \bigcup_{u: v \to u \to s} \{w(us) - w(vs)\} \right).$$ For the purpose of this definition, if $vs$ is not an edge, take $w(vs) = 0$. The *second neighborhood weight* $\beta^D_v$ is given by $\sum_{s} \beta_v(s)$. Finally, the *neighborhood weight difference* of a vertex $v$, denoted by $\delta^D_v$, is $\beta^D_v - \alpha^D_v$. When $D$ is clear from context, we will use $\alpha_v$, $\beta_v$, and $\delta_v$ instead of $\alpha^D_v$, $\beta^D_v$, and $\delta^D_v$. For an arc-weighted digraph, a vertex is *weakly expanding* if $\delta_v \geq 0$. Similarly, a vertex is *strongly contracting* if $\delta_v < 0$. If every vertex in the graph is weakly expanding, the graph is weakly expanding, and the same is true for strongly contracting.
Consider the following example.
Here, the first neighborhood weight of $v$ is the sum of the out-going arcs, and hence $\alpha_v = 3 + 6 = 9$. To compute the second neighborhood weight, we need the value $\beta_v(s)$ of every neighbor $s$ at the end of a path of length $2$. Note that the possibilities for $s$ are $u_1$, $u_2$, and $u_4$. We see $\beta_v(u_2) = 2$ because there is only one way to reach $u_2$, and it ends on an arc of weight $2$. There are two ways to get to $u_4$, one that ends on an arc of length $5$, and another that ends on an arc of weight $1$. Taking the maximum of these two, we get $\beta_v(u_4) = 5$. Finally, there is only one way to get to $u_1$, and it ends on an arc of weight $4$. But since $v u_1$ is an arc of weight $3$, we subtract $3$ from $4$ to get $\beta_v(u_1) = 1$. The total second neighborhood weight is then $\beta_v = \beta_v(u_1) + \beta_v(u_3) + \beta_v(u_4) = 1 + 5 + 2 = 8$. Thus we have $\delta_v = \beta_v - \alpha_v = 8 - 9 = -1$, which shows that $v$ is strongly contracting.
Notice that an arc of weight zero is not the same as a missing arc. If there is an arc of weight zero from $u$ to $v$, $u$ will still count all of $v$’s neighbors as potential second neighbors, where as a missing arc from $u$ to $v$ means the neighbors of $v$ do not count as potential second neighbors of $u$.
The arc-weighted second neighborhood conjecture is as follows.
\[conj:arc-weighted\] Every arc-weighted digraph $D$ without loops or two-cycles contains a vertex $v$ such that $\delta_v \geq 0$.
The following proposition shows the equivalence to the original conjecture via an auxiliary graph construction.
\[prop:equiv\] The arc-weighted version of the second neighborhood conjecture (Conjecture \[conj:arc-weighted\]) is equivalent to the orignal conjecture (Conjecture \[conj:original\]).
If we assume the arc-weighted version, then the original follows by applying the arc-weighted version to a digraph where each arc has weight 1.
Now suppose there is a counterexample $D$ to the arc-weighted version of the conjecture. Thus every vertex is strongly contracting, which means $\delta_v < 0$ for every vertex $v$. We can assume $D$ has no arcs of weight zero, as removing these cannot make $\delta_v$ larger. By the fact that the rationals are dense in the reals, we can assume the arc-weights on $D$ are rational. By scaling the weights with a large enough multiple, we can then assume the arc-weights on $D$ are positive integers.
We now create an auxiliary digraph $D'$ without arc-weights that will be a counterexample to the original conjecture. To form $D'$ from $D$, replace each vertex $v$ of $D$ with a set of vertices $S_v$ such that $|S_v| = \max_{u \in N_1^-(v)} w(uv)$, and arbitrarily order the vertices of each $S_v$. If there is an edge from $u$ to $v$ in $D$, place arcs between every vertex of $S_u$ to the first $w(uv)$ vertices of $S_v$ in $D'$.
We now show that $D'$ is a counterexample to the original conjecture. Given any vertex $x \in S_v$, $x$ has as out neighbors the first $w(vu)$ vertices in $S_u$ for all $u \in N_1^+(v)$. Thus we see that $|N_1^+(x)| = \sum_{u \in N_1^+(v)} w(vu)$, which is the first neighborhood weight of $v$, and hence $|N_1^+(x)| = \alpha_v$.
$N_2^+(x)$ consists of vertices in $S_s$ such that there exists a $u$ with $v \to u \to s$ in $D$. In particular, $S_s \cap N_2^+(x)$ consists of the first $w(us)$ elements of $S_s$, where $u$ is chosen such that $v \to u \to s$ and $w(us)$ is maximized. However, there may be some elements of $S_s$ that are actually first neighbors, so the number of neighbors is reduced by $w(vs)$. Putting this all together, we get $$|N_2^+(x)| = \sum_{s \in N_1^+(v) \cup N_2^+(v)} \max_{v \to u \to s} \{w(us) - w(vs), 0\}.$$ This is the second neighborhood weight of $v$, and hence $|N_2^+(x)| = \beta_v$.
Since $D$ is a counterexample to the arc-weighted conjecture, we know $\alpha_v > \beta_v$. Hence $|N_1^+(x)| > |N_2^+(x)|$ for all $x \in V(D')$, which shows that $D'$ is a counterexample to the original conjecture.
Notice that if $D$ has vertex-weights $\eta$ and is a counterexample to the second neighborhood conjecture, define an arcweight function $w$ where $w(uv) = \eta(v)$. Then $D$ with arc-weights $w$ is a counterexample to the arc-weighted version of the conjecture.
Arc-Weighted Tournaments {#sec:tournaments}
========================
While the arc-weighted version of the conjecture follows from the original conjecture, it is not true that the arc-weighted version for tournaments is a simple consequence of Fisher’s Theorem. This is because if an arc-weighted tournament $D$ were to undergo the the transformation to $D'$ from the proof of Proposition \[prop:equiv\], $D'$ would likely not be a tournament.
Nor do the proofs of Fisher or Havet and Thomassé easily extend to the case of arc-weighted tournaments.
To extend Fisher’s proof, one would need to extend the idea of losing density to an arc-weighted tournament. One natural candidate is as follows: given a digraph $D$ with arc-weights given by $w$, we say $\ell$ is a *losing density* if, for all $v$, $$\sum_{xv \in A(D)} w(xv) \ell(x) \leq \sum_{vy \in A(D)} w(vy) \ell(y).$$ Fisher’s main result for tournaments without arc-weights is that the losing density, as a vertex-weighting of $\overleftarrow{D}$, is weakly expanding at every vertex. However, a losing density is not necessarily weakly expanding at every vertex of $\overleftarrow{D}$ once arc-weights are introduced. In the example below (adapted from [@Fisher96]), $D$ is shown with an arc-weighted losing density, but the vertex in the upper left corner is not weakly expanding in $\overleftarrow{D}$.
Simiarly, Havet and Thomassé’s proof, using median orders, does not seem to easily generalize either. Generalizing this to arc-weighted digraphs, a median order $v_1, \ldots, v_n$ minimizes the total arc-weight of arcs $v_i \to v_j$ where $j < i$. For non-arc-weighted tournaments, they proved $v_n$ is a Seymour vertex. However, that is not the case for arc-weighted tournaments, by taking Havet and Thomassé’s example for non-tournaments and adding arcs of weight zero.
This vertex-weighted digraph is meant to represent a digraph *without* vertex-weights. Instead, each vertex with a vertex-weight $k$ represents a tournament of $k$ vertices with $0$-weighted arcs between every pair of vertices, and an arc between two weighted vertices $u \to v$ represents a complete bipartite graph of edges from every vertex in the tournament $u$ to every vertex in the tournament $v$. Here, some vertex in the tournament labeled $4$ will contain $v_n$, but $v_n$ will not be a Seymour vertex.
However, by extending the work of Fisher, we can generalize his proof to arc-weighted tournaments. It depends on a nice property of arc-weighted diagraphs, where in some situations, one can contract along an arc and maintain the property of vertices being strongly contracting.
\[lemma:contract\] Let $D$ be an arc-weighted digraph, and let $u$ and $v$ be two vertices such that, for any vertex $x$, if $x \to u$ is an arc of nonzero weight, then $x \to v$ is also an arc (possibly weight zero). Then by removing $u$ and adding all of the weight of $x \to u$ to $x \to v$ for all $x$, a new digraph $D'$ is created and $\delta^{D'}_y \leq \delta^D_y$ for all vertices $y \in V(D)$. Such a maneuver is called a *contraction* of $u$ to $v$.
Notice that in this lemma statement, $x$ may equal $v$, and therefore the lemma does not apply if $v \to u$ is an arc of nonzero weight. We typically we will apply the lemma when $u \to v$ is an arc.
Let the property P be the property that for any $x$ such that $x \to u$ is an arc of nonzero weight, we have $x \to v$ is also an arc (possibly of weight zero).
Let $y$ be any vertex. We will show that $\alpha^{D'}_y = \alpha^D_y$ and $\beta^{D'}_y \leq \beta^D_y$, which will prove the lemma.
We first show that $\alpha^{D'}_y = \alpha^D_y$. If $y \to u$ is an arc of nonzero weight, then by property P, we have that $y \to v$ is an arc, and the only change in $\alpha^D_y$ is the weight transfer from $y \to u$ to $y \to v$ and the removal of $u$. In this case, $\alpha^{D'}_y = \alpha^D_y$. If $y \to u$ is not an arc or has zero weight, then the first neighborhood of $y$ either remains unchanged or loses an arc of weight zero, and hence $\alpha^{D'}_y = \alpha^D_y$.
We next show that $\beta^{D}_y \geq \beta^{D'}_y$. Going from $D$ to $D'$, $y$ potentially loses second neighborhood weight on $u$ since $u$ is deleted, and potentially gains second neighborhood weight on $v$ because arcs into $v$ may gain weight. Let $\gamma$ be the sum of all the other unaffected second neighborhood weight. In other words, $\gamma = \sum_{z \neq u, v} \beta_y(z)$. Let $x$ be the vertex such that $y \to x$ is an arc and the second neighborhood weight of $v$ in $D'$ is $\max(w_{D'}(x v) - w_{D'}(y v), 0)$. The second neighborhood weight of $u$ and $v$ in $D$ is at least $\max(w_D(x v) - w_D(y v), 0) + \max(w_D(x u) - w_D(y u, 0)$. Thus, we have $$\begin{aligned}
\beta^{D}_y & \geq & \max(w_D(x v) - w_D(y v), 0) + \max(w_D(x u) - w_D(y u), 0) + \gamma \\
& \geq & \max((w_D(x v) - w_D(y v)) + (w_D(x u) - w_D(y u)), 0) + \gamma \\
& = & \max((w_D(x v) + w_D(x u)) - (w_{D}(y v) + w_D(y u)), 0) + \gamma \\
& = & \max(w_{D'}(x v) - w_{D'}(y v), 0) + \gamma \\
& = & \beta^{D'}_y.\end{aligned}$$ as desired.
Now the main result.
\[thm:arc-tournaments\] Every arc-weighted tournament $D$ contains a Seymour vertex.
Suppose $D$ with arc-weights $w$ is a counterexample, meaning every vertex is strongly contracting. Suppose further that $D$ and $w$ are chosen to minimize the number of vertices, and subject to this constraint, to maximize the number of arcs of weight zero. There are three cases.
**There is a vertex-weighting $\eta$ of $\overleftarrow{D}$ such that every vertex of $\overleftarrow{D}$ is strongly contracting.** By Proposition \[prop:triangle-full\], there must be an arc $u \to v$ in $\overleftarrow{D}$ not contained in any directed triangles. Thus, for any vertex $x$, if $x \to u \in A(\overleftarrow{D})$, then $v \to x$ cannot be an arc, since that would be a directed triangle. Since this is a tournament, $x \to v \in A(\overleftarrow{D})$. Now apply arc-weights $w'$ to $\overleftarrow{D}$ where $w'(uv) = \eta(v)$. We can then apply Lemma \[lemma:contract\] to $\overleftarrow{D}$ with arc-weights $w'$ to obtain a counterexample with fewer vertices than $D$, contradicting that $D$ had the minimum number of vertices.
Hence, we can assume $\overleftarrow{D}$ does not have a vertex-weighting such that every vertex is strongly contracting. Therefore, by Theorem \[thm:expanding-or-contracting\], $D$ has a vertex-weighting $\eta^*$ such that every vertex is weakly expanding.
**For every vertex $v$ and every second neighbor $x$, there is an arc of nonzero weight from $N_1^+(v)$ to $x$.** Create new arc-weights $w^*$ where $w^*(uv) = w(uv) - \epsilon \eta^*(v)$ if $w(uv) > 0$, and $w^*(uv) = 0$ if $w(uv) = 0$. Here, $\epsilon$ is chosen so that no arc changes to negative weight and at least one arc changes to weight zero.
Notice that $D$ with arc-weights $w^*$ is still a counterexample. For any vertex $v$, its second neighborhood weight is decreased by at least $\epsilon \eta^*(N_2^+(v))$, which doesn’t count any lowering of second neighborhood weight within $N_1^+(v)$. The first neighborhood weight is lowered by $\epsilon \eta^*(N_1^+(v))$. By the definition of $\eta^*$, we have $\eta^*(N_1^+(v)) \leq \eta^*(N_2^+(v))$, and hence the second neighborhood weight of every vertex decreased by at least as much as the first neighborhood weight. Since $D$ with arc-weights $w^*$ is a counterexample with more arcs of weight zero than $w$, we have contradicted the fact that $w$ was chosen to maximize the number of arcs of weight $0$.
**There exists a vertex $v$ and a second neighbor $x$ of $v$ such that all the arcs from $N_1^+(v)$ to $x$ are weight zero.** If $y \to x$ is an arc of nonzero weight, we cannot have $v \to y$ as an arc, since then $y \to x$ would be an arc of nonzero weight from $N_1^+(v)$ to $x$. Since $D$ is a tournament, $y \to v$ is an arc. Then by Lemma \[lemma:contract\], we can contract along $x \to v$ to obtain a counterexample with fewer vertices, contradicting the minimality of $D$.
We remark that the proof of Theorem \[thm:arc-tournaments\] relies on the fact that for a tournament, either Proposition \[prop:triangle-full\] applies or Lemma \[lemma:contract\] applies. This is not true for general digraphs, and it seems difficult to extend this proof method beyond tournaments.
Acknowledgements {#acknowledgements .unnumbered}
----------------
The author would like to thank Stephen Hartke for introducing him Fisher’s work on this problem and Debbie Seacrest for her feedback and discussions.
[^1]: The Universty of Montana Western, 710 S Atlantic St, Dillon, MT 59725, United States
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Let $n\ge 5$ and $k\ge 4$ be positive integers. We determine the maximum size of digraphs of order $n$ that avoid distinct walks of length $k$ with the same endpoints. We also characterize the extremal digraphs attaining this maximum number when $k\ge 5$.'
author:
- 'Zejun Huang,[^1] Zhenhua Lyu,[^2] Pu Qiao[^3]'
title: problems for digraphs avoiding distinct walks of a given length with the same endpoints
---
[**Key words:**]{} digraph, problem, transitive tournament, walk
[**AMS subject classifications:**]{} 05C35, 05C20
Introduction
============
problems concern the study of the maximum number, called number, of edges in graphs containing no given subgraphs and the extremal graphs realizing that maximum. Mantel’s theorem determines the maximum number of edges of triangle-free simple graphs as well as the unique graph attaining that maximum. Paul [@PT; @PT2] generalized Mantel’s theorem by determining the maximum number of edges of $K_r$-free graphs on $n$ vertices and the unique graph attaining that maximum, where $K_r$ denotes the complete graph on $r$ vertices. ’s theorem initiated the development of a major branch of graph theory, known as extremal graph theory [@BB; @VN]. Most of the previous results in extremal graph theory concern undirected graphs and only a few extremal problems on digraphs have been investigated; see [@BB; @BES; @BES2; @BH; @BS; @JM]. In this paper we study extremal problems on digraphs.
We consider strict digraphs, i.e., digraphs without loops and parallel arcs. For digraphs, we abbreviate directed walks and directed cycles as walks and cycles, respectively. The number of vertices in a digraph is called its [*order*]{} and the number of arcs its [*size*]{}. We use $\overrightarrow{K}_r$ and $\overrightarrow{C}_r$ to denote the complete digraph and the directed cycle on $r$ vertices.
One natural problem on digraphs is determining the maximum size of a $\overrightarrow{K}_r$-free strict digraph of a given order, which has been solved in [@JM].
Note that the $k$-cycle is a generalization of the triangle when we view a triangle as a 3-cycle in undirected graphs. Another generalization of Mantel’s Theorem is the problem for $k$-cycle-free graphs. However, this problem is difficult even for $C_4$-free graphs [@FK; @TT]. An alternative direction on this problem is considering the orientations of $C_k$-free graphs. For example, $C_4$ has the following orientations.
Note that the $k$-cycles are generalizations of the triangle when we view a triangle as a 3-cycle in undirected graphs. When considering problems on digraphs, another natural problem is the problem on $k$-cycle-free digraphs. An alternative direction is considering the problem on digraphs containing no orientations of $C_k$. For example, if we consider strict digraphs without loops and parallel arcs, $C_4$ has the following orientations.
{width="1in"} {width="1in"} {width="1in"} {width="1in"}\
$C_4^{(1)}$$C_4^{(2)}$ $C_4^{(3)}$$C_4^{(4)}$
It is clear that a graph is $C_4$-free if and only if any of its orientation contains no copy of the above four digraphs. Hence the number for $C_4$-free graphs is equal to that for $\{C_2,C_4^{(1)},C_4^{(2)},C_4^{(3)},C_4^{(4)}\}$-free digraphs. For $t\in\{1,2,3,4\}$, the problem for $C_4^{(t)}$-free digraphs has independent interests; see [@BH]. We will consider a generalization of the problem for $C_4^{(4)}$-free digraphs.
Given a positive integer $k$, we denote by $\mathscr{F}_k$ the family of digraphs consisting of two different walks of length $k$ with the same initial vertex and the same terminal vertex, which have the following diagram
{width="4in"}
where the vertices $u,v,u_1,u_2,\ldots, u_{k-1},w_1,w_2,\ldots,w_{k-1}$ can be duplicate but $$(u_1,u_2,\ldots,u_{k-1})\ne (w_1,w_2,\ldots,w_{k-1}).$$
We say a digraph $D$ is [*$\F$-free*]{} if $D$ contains no subgraph from $\F$. For any digraph $D$ on the vertices $1,2,\ldots,n$, $D $ is $\F$-free if and only if there is at most one walk of length $k$ from $i$ to $j$ for every pair of vertices $i,j$. Let $ex(n,\F)$ be the maximum size of $\F$-free strict digraphs of order $n$ and $Ex(n,\F)$ be the set of $\F$-free strict digraphs of order $n$ with size $ex(n,\F)$. We study the following problem on strict digraphs.
\[pro1\] Given positive integers $n $ and $k$, determine $ex(n,\F)$ and $Ex(n,\F)$.
When $k=1$, it is clear that $ex(n,\mathscr{F}_1)=n(n-1)$ and the unique extremal digraph attaining $ex(n,\mathscr{F}_1)$ is the complete digraph of order $n$. When $k=2$, $\mathscr{F}_2$ consists of a unique digraph $C_4^{(4)}$. and the following four digraphs.
{width="1in"} {width="1in"} {width="1in"} {width="1in"}
This case has been solved by Wu [@WU], whose result states $$ex(n,\mathscr{F}_2)=\begin{cases}\frac{n^2+4n-1}{4}, & \text{if $n$ is odd,}\\
\frac{n^2+4n-4}{4},& \text{if $n$ is even and $n\not=4,$}\\
8, & \text{if $n=4$}.\end{cases}$$ The digraphs attaining $ex(n,\mathscr{F}_2)$ are also determined in [@WU].
In this paper, we always assume the order $n\ge 5$. We will determine $ex(n,\F)$ for $k\ge 4$ and characterize $Ex(n,\F)$ for $k\ge 5$. The paper is organized as follows. Section 2 presents our main result Theorem \[thh2\], which determines $ex(n,\F)$ for $n\ge k+4\ge 8$ and characterizes $Ex(n,\F)$ for $n\ge k+5\ge 10$; section 3 presents the characterization of $Ex(n,\F)$ for $k\ge 4$ and $n=k+2,k+3,k+4$; section 4 presents the proof of Theorem \[thh2\]; section 5 gives a discussion of the unsolved cases.
Main result
===========
In order to present our main result, we need the follow notations and definitions. Let $A$ be an $n\times n$ matrix and $\alpha=\{i_1,i_2,\ldots,i_k\}\subseteq \{1,2,\ldots,n\}$. We denote by $A[\alpha]$ or $A[i_1,i_2,\ldots,i_k]$ the principal submatrix of $A$ lying on its $i_1$-th, $i_2$-th, $\ldots$, $i_k$-th rows and columns, and denote by $A(\alpha)$ or $A(i_1,i_2,\ldots,i_k)$ the principal submatrix of $A$ obtained by deleting its $i_1$-th, $i_2$-th, $\ldots$, $i_k$-th rows and columns.
Let $D=(\mathcal{V},\mathcal{A})$ be a digraph with vertex set $\mathcal{V}=\{v_1,v_2,\ldots,v_n\}$ and arc set $\mathcal{A}$. Its [*adjacency matrix*]{} $A_D=(a_{ij})$ is defined by $$\label{eqh1}
a_{ij}=\left\{\begin{array}{ll}
1,& (v_i,v_j)\in \mathcal{A};\\
0,&\textrm{otherwise}.\end{array}\right.$$ Conversely, given an $n\times n$ 0-1 matrix $A=(a_{ij})$, we can define its digraph $D(A)=(\mathcal{V},\mathcal{A})$ on vertices $v_1,v_2,\ldots,v_n$ by (\[eqh1\]), whose adjacency matrix is $A$.
Let $A=(a_{ij})$ and $B=(b_{ij})$ be matrices of order $m$ and $n$, respectively. $A\otimes B=(a_{ij}B)$ is the tensor product of $A$ and $B$, whose order is $mn$. Denote by $J_{m,n}$ and $J_n$ the $m\times n$ and $n\times n$ matrices with all entries equal to one, $$T_n=\begin{bmatrix}
0&1&\cdots&1\\
&\ddots&\ddots&\vdots\\
&&0&1\\
&&&0
\end{bmatrix}$$ the upper triangular tournament matrix of order $n$, and $$\Pi_{m, n}=J_m\otimes T_n=\begin{bmatrix}
T_{n}&\cdots&T_{n}\\
\vdots&\ddots&\vdots\\
T_{n}&\cdots&T_{n}
\end{bmatrix}.$$
Two matrices $A$ and $B$ are said to be permutation similar if there is a permutation matrix $P$ such that $B=PAP^T$, where $P^T$ denotes the transpose of $P$. For two digraphs $D_1$ and $D_2$, $A_{D_1}$ and $A_{D_2}$ are permutation similar if and only if $D_1$ and $D_2$ are isomorphic.
A digraph of order $n$ is called a [*transitive tournament*]{} if its adjacency matrix is permutation similar to $T_n$. Suppose $m$ and $t<n$ are nonnegative integers. We say a digraph of order $mn+t$ is an [*$(m,n,t)$-completely transitive tournament*]{} if its adjacency matrix is permutation similar to $\Pi_{m+1, n}(\alpha)$, where $\Pi_{m+1, n}(\alpha)$ is an $ (mn+t)\times (mn+t)$ principal submatrix of $\Pi_{m+1, n}$ with $\alpha\subseteq\{mn+1,mn+2,\ldots,mn+n\}$ and $|\alpha|=n-t$. When $t=0$, we see that a digraph of order $mn$ is an $(m,n,t)$-completely transitive tournament if and only if its adjacency matrix is permutation similar to $\Pi_{m, n}$. Moreover, an $(m,n,t)$-completely transitive tournament is a subgraph of the $(m+1,n,0)$-completely transitive tournament.
Now we are ready to state our main result.
\[thh1\] Let $k\ge 4$ and $n=sk+t$, where $s$, $t$ are nonnegative integers with $t<k$. Then $$ex(n,\F)=\frac{n(n-1)}{2}-\frac{(s-1)n}{2}-\frac{(s+1)t}{2}.$$ Moreover, when $k\ge 5$, a digraph $D$ in $ \F $ has size $ex(n,\F)$ if and only if $A_D$ is permutation similar to $\Pi_{k,s+1}(\beta)$, where $\beta\in S(sk+k,sk+1,k-t)$.
\[thh2\] Let $n=sk+t$ with $s, k, t$ being nonnegative integers such that $t<k$. If $ n\ge k+4\ge 8$, then $$\label{eq22}
ex(n,\F)= {n\choose 2} -{s\choose 2}k-st.$$Moreover, if $n\ge k+5\ge 10$, then a digraph $D$ is in $Ex(n,\F)$ if and only if $D$ is an $(s,k,t)$-completely transitive tournament.
We will also determine $Ex(n,\F)$ for $k\ge 4$ and $n=k+2,k+3,k+4$, while $ex(n,\F)$ for $ n\le k+3$ and $Ex(n,\F)$ for $n\le k +1$ can be easily deduced from [@HZ1].
From now on we deal with digraphs with no parallel arcs but allowing loops, and we use the same notations $\F$, $ex(n,\F)$ and $Ex(n,\F)$ for digraphs allowing loops as for strict digraphs. We will give solutions to Problem \[pro1\] for digraphs allowing loops. The same results for strict digraphs follow straightforward, since there is no loop in these extremal digraphs from $Ex(n,\F)$.
$ex(n,\F)$ and $Ex(n,\F)$ for $
n\le k+4$
===============================
For given integers $n$ and $k$, denote by $M_n\{0,1\}$ the set of $n\times n$ 0-1 matrices, $f(A)$ the number of ones in a 0-1 matrix $A$, $$\Gamma(n,k)=\left\{A\in M_n\{0,1\}: A^k\in M_n\{0,1\}\right\},$$ $$\theta(n,k)=\max_{A\in \Gamma(n,k)}f(A) \quad \textrm{and} \quad \Theta(n,k)=\left\{A\in \Gamma(n,k): f(A)=\theta(n,k)\right\}.$$ Let $A\in M_n\{0,1\}$, $B$ an $m\times m$ principal submatrix of $A$, then it is clear that $A\in \Gamma(n,k)$ implies $B\in \Gamma(m,k)$. Moreover, given any $n\times n$ permutation matrix $P$, $A\in \Gamma(n,k)$ if and only if $P^TAP\in \Gamma(n,k)$.
To determine $\theta(n,k)$ and $\Theta(n,k)$ is an interesting problem posed by Zhan (see [@ZH page 234]), which has been partially solved by Wu [@WU], Huang and Zhan [@HZ1].
Given a digraph $D$, the $(i,j)$-entry of $(A_D)^k$ equals $t$ if and only if there are exactly $t$ distinct directed walks of length $k$ from vertex $v_i$ to vertex $v_j$ in $D$. Hence, a digraph $D$ is $\F$-free if and only if its adjacency matrix $A_D$ is in $\Gamma(n,k)$. Moreover, $$\label{eq03}
\theta(n,k) =ex(n,\F).$$ It should be noticed that (\[eq03\]) is not necessarily true for strict digraphs.
For digraphs allowing loops, Huang and Zhan determined $ex(n,\mathscr{F}_k)$ and $Ex(n,\mathscr{F}_k)$ for $k\ge n-1\ge 4$ as follows.
\[thh1\] Let $n,k$ be given integers such that $k\ge n-1\ge 4.$ Then $ex(n,\F)=n(n-1)/2$ and a digraph $D$ is in $Ex(n,\F)$ if and only if $D$ is a transitive tournament of order $n$.
They also determine $ex(n, \mathscr{F}_{n-2})=n(n-1)/2-1$ for $n\ge 6$ and $ex(n, \mathscr{F}_{n-3})=n(n-1)/2-2$ for $n\ge 7$. Hence, when $k\ge 4$, $ex(n,\F)$ for $5\le n\le k+3$ and $Ex(n,\F)$ for $5\le n\le k+1$ have been determined.
In the following of this section, we will determine $ex(k+4,\F)$ and $Ex(n,\F)$ for $k\ge 4$ and $n=k+2,k+3,k+4$. We need the following lemmas.
\[le3\] Let $n \ge 3$, $p$ and $q$ be nonnegative integers such that $(p,q)\ne (0,0)$, and let $A\in M_{n}\{0,1\}$. If $$f(A(i))\le \frac{(n-1)(n-2)}{2}-p\frac{n-1}{2}-q \textrm{ for all } 1\le i\le n,$$ then $$\label{eqle1}
f(A)\le \frac{n(n-1)}{2}-p\frac{n+1}{2}-q-1.$$
Using the same idea as in the proof of [@HZ1 Corollary 10], we count the number of ones in the principal submatrices $A(1),\ldots,A(n)$. Note that each diagonal entry of $A$ appears $n-1$ times and each off-diagonal entry of $A$ appears $n-2$ times in these submatrices. Suppose $A$ has $d$ nonzero diagonal entries. Then $$\begin{aligned}
(n-1)d+(n-2)[f(A)-d]&=&\sum_{i=1}^nf(A(i))
\le n\left[\frac{(n-1)(n-2)}{2}-p\frac{n-1}{2}-q \right].
\end{aligned}$$ It follows that $$f(A)\le \frac{n(n-1)}{2}-p\frac{n+1}{2}-q-\frac{p+2q}{n-2}-\frac{d}{n-2}.$$ Since $(p+2q)/(n-2)>0$, $d/(n-2)\ge 0$ and $f(A)$ is an integer, we get (\[eqle1\]).
For the sake of convenience, we will always use $\{1,2,\ldots,n\}$ to denote the vertex set of a digraph $D$ of order $n$ and use the notation $i\rightarrow j$ to denote the arc $(i,j)$.
\[le4\] Let $m=k+t+s+1$ with $s\ge1$, $k\ge1$, $t\ge 3$ being integers, and let $x_{1},y_{1}\in \mathbb{R}^{k}$, $x_{2},y_{2}\in \mathbb{R}^{t}$, $x_{3},y_{3}\in \mathbb{R}^{s}$. If $$(a_{ij})=\begin{bmatrix}
0&J_{k,t}&J_{k,s}&x_{1}\\
0&T_{t}&J_{t,s}&x_{2}\\
0&0&0&x_{3}\\
y_{1}^T&y_{2}^T&y_{3}^T&\alpha
\end{bmatrix}\in \Gamma(m,t+1)$$ and $$\sum\limits_{i=1}^{3}[f(x_{i})+f(y_{i})]+\alpha\ge s+k+2,$$ then $$\alpha=0,~~
y_{1}=0, ~~x_{3}=0,\textrm{ and }
a_{im}a_{mj}=0 \textrm{ for all }j\le i+2, 1\le i,j\le n.$$
Denote $A=(a_{ij})$. First we claim that $x_{3}=0$ and $y_{1}=0$. Otherwise suppose $x_3\ne 0$ or $y_1\ne 0$. Then $a_{im}=1$ for some $i\in \{k+t+1,\ldots,k+t+s$} or $a_{mj}=1$ for some $j\in \{1,2,\ldots,k\}$. It follows that $D(A)$ has two distinct walks of length $t+1$ from $k$ to $m$ or from $m$ to $k+t+1$: $$\left\{\begin{array}{l}
k\rightarrow k+1 \rightarrow k+3 \rightarrow\cdots\rightarrow k+t\rightarrow i\rightarrow m,\\
k\rightarrow k+2 \rightarrow k+3 \rightarrow\cdots\rightarrow k+t\rightarrow i\rightarrow m,\\
\end{array}\right.$$ $$\left\{\begin{array}{l}
m\rightarrow j\rightarrow k+1\rightarrow k+3\rightarrow k+4\rightarrow \cdots\rightarrow k+t+1,\\
m\rightarrow j\rightarrow k+1\rightarrow k+2\rightarrow k+4\rightarrow \cdots\rightarrow k+t+1,\\
\end{array}\right.$$ which contradicts $A\in \Gamma(m,t+1)$. Hence, $x_{3}$ and $y_{1}$ are zero vectors.
Next we assert that $\alpha=0$. Otherwise, $\alpha=1$. Since $$\sum\limits_{i=1}^{3}(f(x_{i})+f(y_{i}))\ge s+k+1,$$ we have either $$\textrm{}\sum\limits_{i=1}^{3}f(x_{i})\ge k+1 \textrm{ or } \sum\limits_{i=1}^{3}f(y_{i})\ge s+1.$$ If $\sum\limits_{i=1}^{3}f(x_{i})\ge k+1$, then the last column of $A$ has at least two nonzero entries $a_{im}=a_{jm}=1$ with $1\le i<j\le k+t$. Hence $D(A)$ has the following two distinct walks of length $t+1$ from $i$ to $m$: $$\left\{\begin{array}{l}
i\rightarrow m\rightarrow m\rightarrow m\rightarrow\cdots\rightarrow m,\\
i\rightarrow j\rightarrow m\rightarrow m\rightarrow\cdots\rightarrow m.
\end{array}\right.$$ If $\sum\limits_{i=1}^{3}f(y_{i})\ge s+1$, then the last row of $A$ has at least two nonzero entries $a_{mi}=a_{mj}=1$ with $k+1\le i<j\le m$ and $i\le k+t$. It follows that $D(A)$ has the following two distinct walks of length $t+1$ from $m$ to $j$: $$\left\{\begin{array}{l}
m\rightarrow m\rightarrow \cdots\rightarrow m\rightarrow i\rightarrow j, \\
m\rightarrow m\rightarrow \cdots\rightarrow m\rightarrow m\rightarrow j.
\end{array}\right.$$ In both cases we get contradictions. Therefore, $\alpha=0$.
Next we claim $a_{im}a_{mj}=0$ for $j\le i+2$. Otherwise suppose $a_{im}=a_{mj}=1$ with $1\le i,j\le m-1$ and $j\le i+2$. Since $x_{3}=y_{1}=0$, we have $i\le k+t$ and $j\ge k+1$. We distinguish the following cases to find two distinct walks of length $t+1$ with the same endpoints in $D(A)$, which contradicts $A\in \Gamma(m,t+1)$. If $i\le k$, then $j\le k+2$ and $D(A)$ has $$\left\{\begin{array}{l}
i\rightarrow k+1\rightarrow k+2\rightarrow \cdots\rightarrow k+t+1 ,\\
i\rightarrow m\rightarrow j\rightarrow k+3\rightarrow \cdots\rightarrow k+t+1.
\end{array}\right.$$ If $k<i\le k+t-2$, then $j\le k+t$ and $D(A)$ has $$\left\{\begin{array}{l}
k\rightarrow k+1\rightarrow k+2\rightarrow \cdots\rightarrow k+t\rightarrow k+t+1,\\
k\rightarrow k+1\rightarrow
\cdots\rightarrow i\rightarrow m\rightarrow j\rightarrow i+3\rightarrow \cdots\rightarrow k+t+1.
\end{array}\right.$$ If $i= k+t-1$, then $j\le k+t+1$ and $D(A)$ has $$\left\{\begin{array}{l}
k\rightarrow k+1\rightarrow k+2\rightarrow \cdots\rightarrow i\rightarrow k+t\rightarrow k+t+1,\\
k\rightarrow k+1\rightarrow k+2\rightarrow \cdots\rightarrow i\rightarrow m\rightarrow k+t+1,\texttt{ \it if } j=k+t+1,\\
k\rightarrow k+2\rightarrow \cdots\rightarrow i\rightarrow m\rightarrow j\rightarrow\cdots\rightarrow k+t+1, \texttt{ \it if } j\le k+t.
\end{array}\right.$$ If $i= k+t$, then $j\le k+t+2$ and $D(A)$ has $$\left\{\begin{array}{l}
k\rightarrow k+2\rightarrow k+3\rightarrow \cdots\rightarrow i\rightarrow m\rightarrow j,\\
k\rightarrow k+1\rightarrow k+3\rightarrow \cdots\rightarrow i\rightarrow m\rightarrow j.
\end{array}\right.$$ Therefore, $a_{im}a_{mj}=0$ for all $j\le i+2$.
\[co6\] Let $x,y\in \mathbb{R}^{n-1}$ with $n\ge 6$. If $$\label{eqh2}
\begin{bmatrix}
T_{n-1}&x\\y^T&\beta
\end{bmatrix}\in \Gamma(n,n-2),$$ and $$f(x)+f(y)+\beta=n-2,$$ then one of the following holds:
- $x=( 1,\ldots,1,0)^T,y=0$ and $\beta=0$;
- $y=(0,1,\ldots,1 )^T,x=0$ and $\beta=0$.
Denote the matrix in (\[eqh2\]) by $A=(a_{ij})$. Applying Lemma \[le4\] with $k=s=1$, we have $$\label{eqh3}
\beta=a_{n1}=a_{n-1,n}=0,\textrm{ and } a_{in}a_{nj}=0 \textrm{ for all }j\le i+2.$$
We assert $f(x)=0$ or $f(y)=0$. Otherwise, assume that $a_{i_0n}$ is the last nonzero component in $x$, and $a_{nj_0}$ is the first nonzero component in $y$. Since $f(x)+f(y)=n-2\le i_0+n-1-(j_0-1)$, we have $j_0-i_0\le 2$, and $a_{i_0n}a_{nj_0}=0$ follows from (\[eqh3\]), which contradicts the assumption that $a_{i_0n}a_{nj_0}=1$. Therefore, $x=0$ or $y=0$. It follows that either (1) or (2) holds.
Now we are ready to characterize $Ex(k+2,\F)$ and $Ex(k+3,\F)$ for $k\ge 4$.
\[le6\] Let $n\ge 6$ be an integer. Then $$\label{eqh4}
ex(n,\mathscr{F}_{n-2})=\frac{n(n-1)}{2}-1.$$ Moreover, a digraph $D$ is in $Ex(n,\mathscr{F}_{n-2})$ if and only if $A_D$ is permutation similar to $$\label{eq2}
K_n\equiv\begin{bmatrix}
T_{n-2}&J_{n-2,2}\\0&0
\end{bmatrix}\textrm{\quad or \quad}
K'_n\equiv\begin{bmatrix}
0&J_{2,n-2}\\0&T_{n-2}
\end{bmatrix}.$$
By [@HZ1 Corollary 10] we get (\[eqh4\]). Suppose $D$ is a digraph in $Ex(n,\mathscr{F}_{n-2})$ and $A\equiv A_D$. Applying Lemma \[le3\], there exists some $i\in\{1,2,\ldots,n\}$ such that $f(A(i))\ge \frac{(n-1)(n-2)}{2}$. Since $A(i)\in \Gamma(n-1,n-2)$, applying Theorem \[thh1\] we get $f(A(i))=\frac{(n-1)(n-2)}{2}$ and $A(i)$ is permutation similar to $T_{n-1}$. Using permutation similarity if necessary, without loss of generality we assume $i=n$ and $$A=\begin{bmatrix}
T_{n-1}&x\\y^T&\alpha
\end{bmatrix}$$ with $x,y\in \mathbb{R}^{n-1}$. It follows that
$$f(x)+f(y)+\alpha=f(A)-f(A(n))=ex(n,\mathscr{F}_{n-2})-\frac{(n-1)(n-2)}{2}=n-2.$$
Applying Corollary \[co6\], one of the following holds.
- $x=( 1,\ldots,1,0)^T,y=0$ and $\alpha=0.$ Then $A=K_n$;
- $y=(0,1,\ldots,1 )^T,x=0$ and $\alpha=0.$ Then $PAP^T=K'_n$, where $$P=\begin{bmatrix}
0&1\\
I_{n-1}&0
\end{bmatrix}.$$
Therefore, $A_D$ is permutation similar to $K_n$ or $K_n'$.
Conversely, if the adjacency matrix $A$ of a digraph $D$ is permutation similar to $K_n$ or $K'_n$, by direct computation we can verify $f(A)=ex(n, \mathscr{F}_{n-2})$ and $A^{n-2}\in M_{n}\{0,1\}$. Hence $D\in Ex(n,\mathscr{F}_{n-2})$.
\[leh6\] Let $n\ge 7$ be an integer. Then $$\label{eqh5}
ex(n,\mathscr{F}_{n-3})=\frac{n(n-1)}{2}-2.$$ Moreover, a digraph $D$ is in $Ex(n,\mathscr{F}_{n-3})$ if and only if $A_D$ is permutation similar to $$F_n\equiv\begin{bmatrix}
0&J_{2,n-4}&J_{2,2}\\
0&T_{n-4}&J_{n-4,2}\\
0&0&0
\end{bmatrix}.$$
From [@HZ1 Corollary 11] we get $(\ref{eqh5})$. Suppose $D\in Ex(n,\mathscr{F}_{n-3})$ and $A\equiv A_D$. Applying \[le3\] we see that $A$ contains a submatrix $A(i)$, say $A(n)$, such that $f(A(n))\ge \frac{(n-1)(n-2)}{2}-1$. By Theorem \[le6\], $f(A(n))=\frac{(n-1)(n-2)}{2}-1$ and $A(n)$ is permutation similar to $K_{n-1}$ or $K'_{n-1}$.
First we consider the case that $A(n)$ is permutation similar to $K_{n-1}$. Without loss of generality we can assume $A(n)=K_{n-1}$ and $$A=\begin{bmatrix}
0&J_{1,n-4}&J_{1,2}&x_{1}\\
0&T_{n-4}&J_{n-4,2}&x_{2}\\
0&0&0&x_{3}\\
y_{1}^T&y_{2}^T&y_{3}^T&\alpha
\end{bmatrix},$$ where $x_{1},y_{1}\in \mathbb{R}$, $x_{2},y_{2}\in \mathbb{R}^{n-4}$, and $x_{3},y_{3}\in \mathbb{R}^{2}$.
Let $x=(x^T_{1}, x_{2}^T, x_{3}^T) $ and $y=(y_{1}^T,y_{2}^T,y_{3}^T) $. Then $$\alpha+f(x)+f(y)=f(A)-f(A(n))=n-2.$$ Applying Lemma \[le4\], we get $x_{3}=0$, i.e., $a_{n-2,n}=a_{n-1,n}=0$.
Let $i=n-2$ or $n-1$. Then $$\label{eqq37}
f(A(i))=f(A)-(n-3)-a_{ni}= \frac{(n-1)(n-2)}{2}-a_{ni}.$$ On the other hand, since $A(i)\in \Gamma(n-1,n-3)$, by Theorem \[le6\] we have $$\label{eqh6}
f(A(i))\le\frac{(n-1)(n-2)}{2}-1.$$ Combining (\[eqq37\]) and (\[eqh6\]) we get $a_{ni}=1$.
Now applying Corollary \[co6\] to $A(n-1)$ we have $y=(0,1,\ldots,1 ),x=0$, $\alpha=0,$ and $$A= \begin{bmatrix}
0&J_{1,n-4}&J_{1,2}&0\\
0&T_{n-4}&J_{n-4,2}&0\\
0&0&0&0\\
0&J_{1,n-4}&J_{1,2}&0
\end{bmatrix}=P^T F_nP$$ where $$P=\begin{bmatrix}
0&1\\I_{n-1}&0
\end{bmatrix}.$$
Next suppose $A(n)$ is permutation similar to $K'_{n-1}$. Without loss of generality we can assume $A(n)=K'_{n-1}$ and $$A=\begin{bmatrix}
0&J_{2,n-4}&J_{2,1}&x_{1}\\
0&T_{n-4}&J_{n-4,1}&x_{2}\\
0&0&0&x_{3}\\
y_{1}^T&y_{2}^T&y_{3}^T&\alpha
\end{bmatrix},$$ where $x_{1},y_{1}\in \mathbb{R}^2$, $x_{2},y_{2}\in \mathbb{R}^{n-4}$, and $x_{3},y_{3}\in \mathbb{R}$. Applying the same argument as above by counting $f(A(1))$, $f(A(2))$ and applying Corollary \[co6\] to $A(1)$, we get $A=F_{n}$.
Conversely, if the adjacency matrix $A$ of a digraph $D$ is permutation similar to $F_n$, by direct computation we can verify $f(A)=ex(n, \mathscr{F}_{n-3})$ and $A^{n-3}\in M_{n}\{0,1\}$. Hence $D\in Ex(n,\mathscr{F}_{n-3})$.\
Next we determine $ex(k+4,\F)$ for $k\ge 4$ and $Ex(k+4,\F)$ for $k\ge 5$.
\[le7\] Let $x,y\in \mathbb{R}^{n-1}$ with $n\ge 6$, and $$A=\begin{bmatrix}
T_{n-1}&x\\y^T&\alpha
\end{bmatrix}.$$
- If $f(x)+f(y)+\alpha=n-2$, then $A\in \Gamma(n,n-1)$ if and only if $$\label{eqh39}
\alpha=0, ~~ x=(a^T,0)^T,~~ y=(0,b^T)^T,$$ where $a\in \mathbb{R}^s$, $b\in \mathbb{R}^{n-s-1}$ with $ s\in\{0,1,\ldots,n-2\}$. Here $s=0$ means $x=0$.
- If $f(x)+f(y)+\alpha=n-1$, then $A\in \Gamma(n,n-1)$ if and only if $$\label{eqh310}
\alpha=0, ~~x=(J_{1, s},0)^T, ~~y=(0,J_{1,n-s-1})^T,$$ where $s\in\{0,1,\ldots,n-2\}$.
[*(i)*]{} Suppose $A\in \Gamma(n,n-1)$. First we claim $\alpha=0$. Otherwise, since $f(x)+f(y)=n-3\ge 3$, we have either $f(x)\ge 2$ or $f(y)\ge 2$. If $f(x)\ge 2$, say, $a_{in}=a_{jn}=1$ with $1\le i<j\le n-1$, then $D(A)$ has the following two distinct walks from $i$ to $n$ with the same length $n-1$: $$\left\{\begin{array}{l}
i\rightarrow n\rightarrow n\rightarrow n\rightarrow\cdots\rightarrow n,\\
i\rightarrow j\rightarrow n\rightarrow n\rightarrow\cdots\rightarrow n.
\end{array}\right.$$ If $f(y)\ge 2$, say, $a_{ni}=a_{nj}=1$ with $1\le i<j\le n-1$, then $D(A)$ has the following two distinct walks from $n$ to $j$ with the same length $n-1$: $$\left\{\begin{array}{l}
n\rightarrow n\rightarrow \cdots\rightarrow n\rightarrow i\rightarrow j, \\
n\rightarrow n\rightarrow \cdots\rightarrow n\rightarrow n\rightarrow j.
\end{array}\right.$$ In both cases we get contradictions. Hence $\alpha=0$.
Next we claim $$\label{eqn311}
a_{in}a_{nj}=0 \textrm{~~~for all~~~}i\ge j.$$ If $x=0$ or $y=0$, the claim is clear. Suppose $x$, $y$ are nonzero, and there exist $i\ge j$ such that $a_{in}a_{nj}=1$. Then we have the following cases and in each of these cases $D(A)$ has two different walks of length $n-1$ with the same endpoints, which contradicts $A\in \Gamma(n,n-1)$.
[*Case 1.*]{} $i\le 2$. $D(A)$ has $$\left\{\begin{array}{l}
1\rightarrow i\rightarrow n\rightarrow j\rightarrow i+2\rightarrow i+3\rightarrow\cdots \rightarrow n-1,\\
1\rightarrow i\rightarrow n\rightarrow j\rightarrow i+1\rightarrow i+3\rightarrow\cdots \rightarrow n-1,\\
\end{array}\right.$$ where the arc $1\rightarrow i$ does not appear if $i=1$.
[*Case 2.*]{} $i=3$. $D(A)$ has $$\left\{\begin{array}{l}
1\rightarrow 3\rightarrow n\rightarrow j\rightarrow i+1\rightarrow\cdots \rightarrow n-1,\\
1\rightarrow 2\rightarrow 3\rightarrow n\rightarrow j\rightarrow i+2\rightarrow\cdots \rightarrow n-1.\\
\end{array}\right.$$
[*Case 3.*]{} $4\le i\le n-1$. $D(A)$ has $$\left\{\begin{array}{l}
1\rightarrow 3\rightarrow 4\rightarrow\cdots \rightarrow i\rightarrow n\rightarrow j\rightarrow i+1\rightarrow\cdots \rightarrow n-1,\\
1\rightarrow 2\rightarrow 4\rightarrow\cdots \rightarrow i\rightarrow n\rightarrow j\rightarrow i+1\rightarrow\cdots \rightarrow n-1,\\
\end{array}\right.$$ where the walk $j\rightarrow i+1\rightarrow\cdots \rightarrow n-1$ does not appear if $i=n-1$.
Let $a_{sn}$ be the last nonzero component in $x$ and $a_{nt}$ be the first nonzero component in $y$. Since $f(x)+f(y)=n-2$, by (\[eqn311\]) we have $t-s=1$ or 2. Hence (\[eqh39\]) holds.
Conversely, suppose $A$ satisfies (\[eqh39\]). Let $$\label{eqq12}
B =\begin{bmatrix}
T_{s}&J_{s,n-s-1}&J_{s,1}\\
0&T_{n-s-1}&0\\
0&J_{1,n-s-1}&0
\end{bmatrix}$$ with $0\le s\le n-1$. To prove $A\in \Gamma(n,n-1)$, it suffices to verify $B \in \Gamma(n,n-1),$ since $B\ge A$, where the notation $\ge$ is to be understood entrywise.
If $s=n-1$, then $B=T_{n}\in \Gamma(n,n-1)$. If $s<n-1$, then $$B =\begin{bmatrix}
T_{s}&J_{s,1}&J_{s,n-s-2}&J_{s,1}\\
0&0&J_{1,n-s-2}&0\\
0&0&T_{n-s-2}&0\\
0&1&J_{1,n-s-2}&0
\end{bmatrix}$$ is permutation similar to $$\begin{bmatrix}
T_{s}&J_{s,1}&J_{s,1}&J_{s,n-s-2}\\
0&0&1&J_{1,n-s-2}\\
0&0&0&J_{1,n-s-2}\\
0&0&0&T_{n-s-2}
\end{bmatrix}=T_{n}.$$ Therefore, $B \in \Gamma(n,n-1)$. This completes the proof for (i).
\(ii) For the sufficiency part, if (\[eqh310\]) holds, then $A=B\in \Gamma(n,n-1)$. For the necessity part, let $a_{sn}$ be the last nonzero component in $x$ and $a_{nt}$ be the first nonzero component in $y$. Since $f(x)+f(y)=n-1$, applying the same arguments as above we get $\alpha=0$ and $t-s=1$. It follows that (\[eqh310\]) holds.
\[le8\] Let $n\ge 8$ be an integer. Then $$\label{eq316}
ex(n,\mathscr{F}_{n-4})=\frac{n(n-1)}{2}-4.$$
Let $A$ be the adjacency matrix of any $\mathscr{F}_{n-4}$-free digraph $D$ of order $n$. Then $A\in \Gamma(n,n-4)$ and $A(i)\in \Gamma(n-1,n-4)$ for all $1\le i\le n$. Hence $$f(A(i))\le ex(n-1,\mathscr{F}_{n-4})=\frac{(n-1)(n-2)}{2}-2$$ for all $1\le i\le n$. By Lemma \[le3\], we have $$\label{eq313}
f(A)\le \frac{n(n-1)}{2}-3.$$
Suppose equality in (\[eq313\]) holds. Then by Lemma \[le3\], $A$ contains a submatrix $A(i)$ with $\frac{(n-1)(n-2)}{2}-2$ nonzero entries. Using permutation similarity if necessary, without loss of generality we assume $f(A(n))=\frac{(n-1)(n-2)}{2}-2$. Since $A(n)\in \Gamma(n-1,n-4)$, by Theorem \[leh6\], we may further assume $$\label{eq311}
A=(a_{ij})=\begin{bmatrix}
0&J_{2,n-5}&J_{2,2}&x_{1}\\
0&T_{n-5}&J_{n-5,2}&x_{2}\\
0&0&0&x_{3}\\
y^{T}_{1}&y^{T}_{2}&y^{T}_{3}&\alpha
\end{bmatrix},$$ where $x_1,x_3,y_1, y_3\in \mathbb{R}^{2}$, $x_2,y_2\in \mathbb{R}^{n-5}$.
Let $x=(x_1^T,x_2^T,x_3^T)$ and $y=(y_1^T,y_2^T,y_3^T)$. Then $$f(x)+f(y)+\alpha=f(A)-f(A(n))=n-2.$$ Applying Lemma \[le4\] to $A$ we know $y_{1}=x_{3}=0$ and $\alpha=0$.
Since $a_{n-1,n}=a_{n1}=0$ and $A(1,n-1)\in \Gamma(n-2,n-4)$, by (\[eq311\]) we have $$\begin{aligned}
f(A(1,n-1))&=&f(A)-2(n-5)-3-a_{1n}-a_{n1}-a_{n-1,n}-a_{n,n-1}\\
&=&\frac{(n-2)(n-3)}{2}+1-a_{1n}-a_{n,n-1}\\
&\le&\frac{(n-2)(n-3)}{2}-1
\end{aligned}$$ where the inequality follows from Theorem \[le6\]. Hence, $$\label{eq312}
a_{1n}=a_{n,n-1}=1.$$ On the other hand, applying Corollary \[co6\] to $A(1,n-1)$ we have either $x=0$ or $y=0$, which contradicts (\[eq312\]). Hence, (\[eq313\]) is a strict inequality and we have $$\label{eq314}
ex(n,\mathscr{F}_{n-4})\le \frac{n(n-1)}{2}-4.$$
Now let $D$ be the digraph with adjacency matrix $$B=\begin{bmatrix}
0&J_{3,n-5}&J_{3,2}\\
0&T_{n-5}&J_{n-5,2}\\
0&0&0\\
\end{bmatrix}.$$ By direct computation, we have $$B^{n-4}=\begin{bmatrix}
0&0&J_{3,2}\\
0&0&0\\
0&0&0\\
\end{bmatrix},$$ and hence $D$ is $\mathscr{F}_{n-4}$-free. Therefore, $$\label{eq315}
ex(n,\mathscr{F}_{n-4})\ge f(B)=\frac{n(n-1)}{2}-4.$$ Combining (\[eq314\]) and (\[eq315\]) we get (\[eq316\]).
\[le10\] Let $k\ge 5$ and $s$ be positive integers, let $x_i,y_i\in \mathbb{R}^k$ with components from $\{0,1\}$ for $i=1,2,\ldots,s$, and let $$A=(a_{ij})=\begin{bmatrix}
T_{k}&x_1&x_2&x_3&\cdots&x_s\\
y_1^T&0&1&0&\cdots&0\\
y_2^T&0&0&1&\ddots&\vdots\\
y_3^T&0&0&0&\ddots&0\\
\vdots&\vdots&\vdots&\vdots&\ddots&1\\
y_s^T&1&0&\cdots&\cdots&0
\end{bmatrix}$$.
- If there is some $i\in\{1,2,\ldots,s\}$ such that $f(x_i)\ge 3$ or $f(y_{i})\ge 3$, then $A \notin \Gamma(k+s,p)$ for any integer $p\ge 2$.
- If $s=2$ and there is some $i\in\{1,2\}$ such that $f(x_i)=f(y_{i})= 2$, then $A \notin \Gamma(k+s,p)$ for any integer $p\ge 5$.
[*(i)*]{} If there is some $t$ such that $f(x_t)\ge 3$, then we have $a_{i,k+t}=a_{j,k+t}=a_{m,k+t}=1$ for $1\le i<j<m\le k$. Without loss of generality we assume $t=1$. For any $p\in \{2,3,\ldots,k\}$, we can find two distinct walks of length $p$ between the same endpoints in the following walks in $D(A)$. $$\left\{\begin{array}{l}
i\rightarrow j\rightarrow k+1\rightarrow k+2\rightarrow\cdots\rightarrow k+s \rightarrow k+1\rightarrow k+2\rightarrow\cdots \\
i\rightarrow m\rightarrow k+1\rightarrow k+2\rightarrow\cdots\rightarrow k+s \rightarrow k+1\rightarrow k+2\rightarrow\cdots
\end{array}\right.$$
If there is some $t$, say $t=1$, such that $f(y_t)\ge 3$, then we have $a_{k+1,i}=a_{k+1,j}=a_{k+1,m}=1$ for $1\le i<j<m\le k$. For any $p\in \{2,3,\ldots,k\}$, we can find two distinct walks of length $p$ between the same endpoints in the following walks in $D(A)$ $$\left\{\begin{array}{l}
\cdots\rightarrow k+s\rightarrow k+1\rightarrow k+2\rightarrow\cdots\rightarrow k+s\rightarrow k+1\rightarrow i\rightarrow m,\\
\cdots\rightarrow k+s\rightarrow k+1\rightarrow k+2\rightarrow\cdots\rightarrow k+s\rightarrow k+1\rightarrow j\rightarrow m.
\end{array}\right.$$ Therefore, $A\notin \Gamma(k+s,p)$ for any $p\in \{2,3,\ldots,k\}$.\
[*(ii)*]{} Without loss of generality, we assume $f(x_1)=f(y_1)=2$ and $a_{p_1,k+1}=a_{p_2,k+1}=a_{k+1,q_1}=a_{k+1,q_2}=1$ with $1\le p_1<p_2\le k$ and $1\le q_1<q_2\le k$. If $p\ge 5$ is odd, then $D(A)$ has the following two distinct walks of length $p$ from $p_1$ to $q_2$: $$\left\{\begin{array}{l}
p_1 \rightarrow k+1\rightarrow k+2\rightarrow\cdots \rightarrow k+1 \rightarrow q_1 \rightarrow q_2,\\
p_1\rightarrow p_2 \rightarrow k+1\rightarrow k+2\rightarrow\cdots \rightarrow k+1 \rightarrow q_2.
\end{array}\right.$$ If $p\ge 5$ is even, then $D(A)$ has the following two distinct walks of length $p$ from $p_1$ to $q_2$: $$\left\{\begin{array}{l}
p_1 \rightarrow p_2\rightarrow k+1\rightarrow k+2\rightarrow\cdots \rightarrow k+1 \rightarrow q_1 \rightarrow q_2,\\
p_1\rightarrow k+1 \rightarrow k+2\rightarrow k+1\rightarrow\cdots\rightarrow k+2 \rightarrow k+1 \rightarrow q_2.
\end{array}\right.$$
Therefore, $A\not\in \Gamma(k+s,p)$ for any integer $p\ge 5$.
\[th9\] Let $n\ge 9$ be an integer. Then a digraph $D$ is in $Ex(n,\mathscr{F}_{n-4})$ if and only if $A_D$ is permutation similar to one of the following matrices $$F_{1}(n)\equiv\begin{bmatrix}
0&J_{3,n-5}&J_{3,2}\\
0&T_{n-5}&J_{n-5,2}\\
0&0&0\\
\end{bmatrix},\quad
F_{2}(n)\equiv\begin{bmatrix}
0&J_{2,n-5}&J_{2,3}\\
0&T_{n-5}&J_{n-5,3}\\
0&0&0\\
\end{bmatrix},$$$$F_{3}(n)\equiv\begin{bmatrix}
0&J_{2,n-5}&J_{2,2}&J_{2,1}\\
0&T_{n-5}&J_{n-5,2}&U_{m}\\
0&0&0&0\\
0&U'_{m}&J_{1,2}&0
\end{bmatrix},\quad
F_{4}(n)\equiv\begin{bmatrix}
T_{n-4}&w_{1}&w_{2}&w_{3}&w_{4}\\
u_{1}&0&1&1&1\\
u_{2}&0&0&1&1\\
u_{3}&0&0&0&1\\
u_{4}&0&0&0&0
\end{bmatrix},$$ where $$U_{m}=(J_{1,m},0)^T,~~ U'_{m}=(0,J_{1,n-m-7}) \textrm{~~ with~~} 0\leq m\leq n-7$$ and $$w_{j}=(J_{1,k_j},0)^T, ~~u_{j}=(0,J_{1,n-k_j-5}) \textrm{~~for~~} i=1,2,3,4$$ with $0\leq k_1<k_2<k_3<k_4\leq n-5$.
Suppose $A\equiv A_D$ is permutation similar to one of $F_1(n),F_2(n),F_3(n)$ and $F_4(n)$. It is clear that $f(A)=\frac{n(n-1)}{2}-4$. To prove $D\in Ex(n,\mathscr{F}_{n-4})$, it suffices to prove $F_i^{n-4}(n)\in M_n\{0,1\}$ for $i=1,2,3,4$.
By direct computation we know $$F_{1}^{n-4}(n)=\begin{bmatrix}
0&J_{3,2}\\
0&0
\end{bmatrix},\quad F_{2}^{n-4}(n)=\begin{bmatrix}
0&J_{2,3}\\
0&0
\end{bmatrix},\quad F_{3}^{n-4}(n)=\begin{bmatrix}
0&J_{2,2}&O_{2\times 1}\\
0&0&0
\end{bmatrix}$$ are all 0-1 matrices, where $O_{2\times 1}$ is the $2\times 1$ zero matrix.
For $i=4$, let $\alpha=\{1,2,\ldots,n-4,n-3+k_1,n-3+k_2,n-3+k_3,n-3+k_4\}$ and $$A' \equiv J_2\otimes T_{n-4}=\begin{bmatrix}
T_{n-4}&T_{n-4}\\
T_{n-4}&T_{n-4}
\end{bmatrix}.$$ Then $F_4(n)=A'[\alpha]$ is a principal submatrix of $A'$. Moreover, $$(A')^{n-4}=(J_2\otimes T_{n-4})^{n-4}=J_2^{n-4}\otimes T_{n-4}^{n-4}=0$$ implies $F_4^{n-4}(n)=0$. Thus we get the sufficiency of Theorem \[th9\].
Next we prove the necessity part of Theorem \[th9\]. Suppose $D\in Ex(n,\mathscr{F}_{n-4})$. Denote by $A\equiv A_D$ the adjacency matrix of $D$. Then $f(A)=ex(n,\mathscr{F}_{n-4})$, $A\in \Gamma(n,n-4)$ and $A(i)\in \Gamma(n-1,n-4)$ for all $1\le i\le n$. By (\[eqh5\]) we have $$f(A(i))\le \frac{(n-1)(n-2)}{2}-2 \textrm{~~for all~~} 1\le i\le n.$$ We distinguish two cases.
[*Case 1.*]{} $f(A(q))=\frac{(n-1)(n-2)}{2}-2$ for some $q\in\{1,2,\ldots,n\}$. By Theorem \[leh6\], without loss of generality, we assume $q=n$ and $$A=\begin{bmatrix}
0&J_{2,n-5}&J_{2,2}&x_{1}\\
0&T_{n-5}&J_{n-5,2}&x_{2}\\
0&0&0&x_{3}\\
y^{T}_{1}&y^{T}_{2}&y^{T}_{3}&\alpha
\end{bmatrix},$$ where $x_1,x_3,y_1, y_3\in \mathbb{R}^{2}$, $x_2,y_2\in \mathbb{R}^{n-5}$. Let $x=(x_1^T,x_2^T,x_3^T)$ and $y=(y_1^T,y_2^T,y_3^T)$. Since $$\label{eq319}
f(x)+f(y)+\alpha=f(A)-f(A(n))=n-3\ge 6,$$ applying Lemma \[le4\], we have $$\label{eq318}
\alpha=0, ~y_{1}=x_{3}=0,\textrm{~~and~~ }a_{in}a_{nj}=0\textrm{~~for~~}j\le
i+2.$$
If $x=0$, then $A$ is permutation similar to $F_{1}(n)$. If $y=0$, then $A=F_{2}(n)$. If both $x$ and $y$ are nonzero, let $a_{sn}$ be the last nonzero component in $x$ and $a_{nt}$ be the first nonzero component in $y$. By (\[eq319\]) and (\[eq318\]) we have $$\label{eq320}
t-s=3,~~x=(J_{1,s},0)\textrm{~~and~~}y=(0,J_{1,n-s-3}).$$ By exchanging row 1 and row 2 of $A$, and exchanging column 1 and column 2 of $A$ simultaneously, we obtain a new matrix $A'=(a'_{ij})$. Applying Lemma \[le4\] to $A'$ we have $a'_{in}a'_{nj}=0$$j\le
i+2$. Hence $$a_{1n}a_{n4}=a'_{2n}a'_{n4}=0.$$ Similarly, by interchanging the roles of the indices $n-1$ and $n-2$, we get $$a_{n-4,n}a_{n,n-1}=0.$$ Therefore, in (\[eq320\]) we have $s\ne 1$ and $t\ne n-1$. Hence $A=F_3(n)$ with $m=s-2$.\
Case 2. $f(A(i))\le\frac{(n-1)(n-2)}{2}-3$ for all $i$. Denote by $$\delta_{i}=\sum_{j=1}^na_{ij}+\sum_{j\ne i}a_{ji}$$ the number of nonzero entries in the $i$-th row and the $i$-th column of $A$. Then $$\label{eq321}
\delta_{i}= f(A)-f(A(i))\ge n-2 \textrm{~~for all~~} 1\le i\le n.$$ Applying Lemma \[le3\] to $A$, there exists some $i_0$ such that $$f(A(i_0))=\frac{(n-1)(n-2)}{2}-3.$$ Without loss of generality, we assume $i_0=n$. For any $i\in\{1,\ldots,n-1\}$, since $A(i,n)\in \Gamma(n-2,n-4)$, by Theorem \[le6\] we have $$\label{eq322}
f(A(i,n))\le\frac{(n-2)(n-3)}{2}-1.$$
Next we prove the following claim.
[*[**Claim 1.**]{} Let $i\in\{1,2,\ldots,n-1\}$. Then $$\label{eq3323}
f(A(i,n))\le\frac{(n-2)(n-3)}{2}-2.$$ Moreover, there exists some $i\in\{1,2,\ldots,n-1\}$ such that equality holds in (\[eq3323\]).* ]{}
Suppose equality in (\[eq322\]) holds for some $i_1$, say, $i_1=n-1$. Then by Theorem \[le6\], $A(n-1,n)$ is permutation similar to $K_{n-2}$ or $K'_{n-2}$. By (\[eq321\]) we have $$\begin{aligned}
\delta_{n-1}&=&f(A(n))-f(A(n-1,n))+a_{n-1,n}+a_{n,n-1}\\
&=&n-4+a_{n-1,n}+a_{n,n-1}\\
&\ge& n-2.
\end{aligned}$$ It follows that $a_{n-1,n}=a_{n,n-1}=1$.
If $A(n-1,n)$ is permutation similar to $K'_{n-2}$, without loss of generality, we may assume $$A=\begin{bmatrix}
0&J_{2,n-4}&x_{1}&x_{3}\\
0&T_{n-4}&x_{2}&x_{4}\\
y^{T}_{1}&y^{T}_{2}&\alpha&1\\
y^{T}_{3}&y^{T}_{4}&1&\alpha'
\end{bmatrix},$$ where $x_1,x_3,y_1, y_3\in \mathbb{R}^{2}$, $x_2,y_2,x_4,y_4\in \mathbb{R}^{n-4}$. By [@HZ1 Lemma 1], we have $\alpha=\alpha'=0$ and $$a_{in}+a_{ni}\le 1\textrm{~~for all~~}1\le i\le n-2.$$ Thus $f(x_3)+f(y_3)\le 2$ and $$\label{eq324}
f(x_3)+f(x_4)+f(y_3)+f(y_4)=\delta_n-2=n-4\ge 5.$$
If $x_4$ has two nonzero entries, say, $a_{i_1,n}=a_{i_2,n}=1$ with $3\le i_1,i_2\le n-2$, then $D$ has the following distinct walks of length $n-4$ between the same endpoints $$\left\{\begin{array}{l}
2\rightarrow i_1\rightarrow n\rightarrow n-1\rightarrow n\rightarrow\cdots\rightarrow n(\rightarrow n-1),\\
2\rightarrow i_2\rightarrow n\rightarrow n-1\rightarrow n\rightarrow\cdots\rightarrow n(\rightarrow n-1).
\end{array}\right.$$ Hence $f(x_4)\le 1$ and $$\label{eq325}
f(y_4)=n-4-f(x_3)-f(y_3)-f(x_4)\ge n-7\ge 2.$$
If $y_3$ and $y_4$ have three nonzero entries, say, $a_{n, j_1}=a_{n, j_2}=a_{n, j_3}=1$ with $j_1<j_2<j_3\le n-2$ and $j_2\ge 3$, then $D$ has the following distinct walks of length $n-4$ between the same endpoints $$\left\{\begin{array}{l}
(n-1\rightarrow) n\rightarrow n-1\rightarrow n\rightarrow\cdots\rightarrow n\rightarrow n-1\rightarrow n\rightarrow j_3,\\
(n-1\rightarrow) n\rightarrow n-1\rightarrow n\rightarrow\cdots\rightarrow n \rightarrow j_1\rightarrow j_2\rightarrow j_3.
\end{array}\right.$$ Combining this with (\[eq324\]) and (\[eq325\]) , we have $$f(y_4)=2,~~y_3=0,~~f(x_3)=2 \textrm{~~and~~}f(x_4)=1.$$ Suppose the nonzero entries in $y_4$ are $a_{n,j_1}$ and $a_{n,j_2}$ with $3\le j_1<j_2\le n-2$. Then $D$ has two distinct walks of length $n-4$ between the same endpoints in the following walks: $$\left\{\begin{array}{l}
1 \rightarrow 3\rightarrow 4\rightarrow\cdots\rightarrow n-2,\\
1 \rightarrow n \rightarrow n-1\rightarrow n\rightarrow\cdots\rightarrow n \rightarrow j_2 \rightarrow n-2 ,\\
1 \rightarrow n \rightarrow n-1\rightarrow n\rightarrow\cdots\rightarrow n \rightarrow j_1\rightarrow j_2 \rightarrow n-2 ,
\end{array}\right.$$ where $j_2\rightarrow n-2$ does not appear when $j_2=n-2$. This contradicts the condition that $A\in \Gamma(n,n-4)$.
If $A(n-1,n)$ is permutation similar to $K_{n-2}$, then we may assume $$A=\begin{bmatrix}
T_{n-4}&J_{n-4, 2}&y_{2}&y_{4}\\
0&0&y_{1}&y_{3}\\
x^{T}_{2}&x^{T}_{1}&\alpha&1\\
x^{T}_{4}&x^{T}_{3}&1&\alpha'
\end{bmatrix},$$ where $x_1,x_3,y_1, y_3\in \mathbb{R}^{2}$, $x_2,y_2,x_4,y_4\in \mathbb{R}^{n-4}$. Applying similar arguments as above we can deduce $A\not\in \Gamma(n,n-4)$, which contradicts $D\in Ex(n,\mathscr{F}_{n-4})$. Hence (\[eq322\]) is a strict inequality and we get (\[eq3323\]).
On the other hand, applying Lemma \[le3\] to $A(n)$, there exists some $i\in\{1,2,\ldots,n-1\}$ such that equality in (\[eq3323\]) holds. Thus we get Claim 1.\
By Lemma 1 of \[1\], we have $\alpha=\alpha'=0$ and $$a_{in}+a_{ni}\le 1\textrm{ for all }1\le i\le n-2.$$ Thus $f(x_3)+f(y_3)\le 2$ and $$\label{eq3324}
f(x_3)+f(x_4)+f(y_3)+f(y_4)=\delta_n-2=n-4\ge 5.$$
If $x_4$ has two nonzero entries, say, $a_{n, i_1}=a_{n, i_2}=1$ with $1\le i_1,i_2\le n-4$. Then $D(A)$ has the following distinct walks of length $n-4$ between the same endpoints $$\left\{\begin{array}{l}
n\rightarrow n-1\rightarrow n\rightarrow\cdots\rightarrow n \rightarrow i_1\rightarrow n-2,\\
n\rightarrow n-1\rightarrow n\rightarrow\cdots\rightarrow n \rightarrow i_2\rightarrow n-2.
\end{array}\right.$$ Hence $f(x_4)\le 1$ and $$\label{eq3325}
f(y_4)=n-4-f(x_3)-f(y_3)-f(x_4)\ge n-7\ge 2.$$
If $y_3$ and $y_4$ have three nonzero entries, say, $a_{j_1, n}=a_{j_2, n }=a_{j_3, n }=1$ with $j_1<j_2<j_3\le n-2$ and $j_2\le n-4$. Then $D(A)$ has the following distinct walks of length $n-4$ between the same endpoints $$\left\{\begin{array}{l}
j_3\rightarrow n\rightarrow n-1\rightarrow n\rightarrow\cdots\rightarrow n\rightarrow n-1\rightarrow n(\rightarrow n-1),\\
j_1\rightarrow j_2\rightarrow j_3 \rightarrow n\rightarrow n-1\rightarrow n\rightarrow\cdots\rightarrow n (\rightarrow n-1).
\end{array}\right.$$ Combining this with (\[eq3324\]) and (\[eq3325\]) , we have $$f(y_4)=2, y_3=0, f(x_3)=2 \textrm{ and }f(x_4)=1.$$ Suppose the nonzero entries in $y_4$ are $a_{j_1, n}$ and $a_{j_2, n}$ with $1\le j_1<j_2\le n-4$. Then $D(A)$ has two distinct walks of length $n-4$ between the same endpoints in the following walks: $$\left\{\begin{array}{l}
1 \rightarrow 2\rightarrow\cdots\rightarrow n-4\rightarrow n-3,\\
1\rightarrow j_1 \rightarrow n \rightarrow n-1\rightarrow n\rightarrow\cdots\rightarrow n \rightarrow n-3,\\
1\rightarrow j_1\rightarrow j_2 \rightarrow n \rightarrow n-1\rightarrow n\rightarrow\cdots\rightarrow n \rightarrow n-3
\end{array}\right.$$ where $1\rightarrow j_1$ does not appear when $j_1=1$. Therefore, we always find two distinct walks with length $n-4$ between the same endpoints, which contradicts the condition that $A\in \Gamma(n,n-4)$.
Now without loss of generality we assume $f(A(n-1,n))=\frac{(n-2)(n-3)}{2}-2$. Then $A(i,n-1,n)\in \Gamma(n-3,n-4)$ and $$\label{eqq26}
f(A(i,n-1,n))\le\frac{(n-3)(n-4)}{2}\textrm{~~for all~~} 1\le i\le n-2.$$ Next we prove the following claim.\
[*[**Claim 2.**]{} Let $i\in\{1,2,\ldots,n-2\}$. Then $$\label{eq326}
f(A(i,n-1,n))\le\frac{(n-3)(n-4)}{2}-1.$$ Moreover, there exists some $i\in\{1,2,\ldots,n-2\}$ such that equality in (\[eq326\]) holds.*]{}\
Suppose equality in (\[eqq26\]) holds for some $i$, say, $i=n-2$. By Theorem \[thh1\], $A(n-2,n-1,n)$ is permutation similar to $T_{n-3}$. Without loss of generality, we assume $$A=\begin{bmatrix}
T_{n-3}&x_{1}&x_{2}&x_{3}\\
y^{T}_{1}&a_{n-2,n-2}&a_{n-2,n-1}&a_{n-2,n}\\
y^{T}_{2}&a_{n-1,n-2}&a_{n-1,n-1}&a_{n-1,n}\\
y^{T}_{3}&a_{n,n-2}&a_{n,n-1}&a_{n,n}
\end{bmatrix},$$where $x_i,y_i\in \mathbb{R}^{n-3}$, for $i=1,2,3$. Since $\delta_{n-2}\ge n-2$ and $$f(x_{1})+f(y_{1})+a_{n-2,n-2}=f(A(n-1,n))-f(A(n-2,n-1,n))=n-5,$$ we have $$\sum\limits_{i=n-1,n}(a_{n-2,i}+a_{i,n-2})=\delta_{n-2}-[f(x_{1})+f(y_{1})+a_{n-2,n-2}]\ge 3.$$ Then either $a_{n-2,n-1}+a_{n-1,n-2}=2$ or $a_{n-2,n}+a_{n,n-2}=2$. Without loss of generality, we assume $a_{n-1,n-2}=a_{n-2,n-1}=1$. By Lemma 1 (ii) of [@HZ1], we obtain $a_{n-1,n-1}=a_{n-2,n-2}=0$ and $$f(x_1)+f(y_1)=n-5\ge 4.$$ Then we have $f(x_1)\ge 3$ or $f(y_1)\ge 3$, or $f(x_1)=f(y_1)=2$. Applying Lemma \[le10\] to $A(n)$, we get $D\notin Ex(n,\mathscr{F}_{n-4})$, a contradiction.
Therefore, (\[eqq26\]) is a strict inequality and we have (\[eq326\]). Moreover, applying Lemma \[le3\] to $A(n-1,n)$ we get the second part of Claim 2.\
Without loss of generality we assume $$f(A(n-2,n-1,n))=\frac{(n-3)(n-4)}{2}-1.$$ For any $i\in \{1,2,\ldots,n-3\}$, since $A(i,n-2,n-1,n)\in \Gamma(n-4,n-4)$, by Theorem \[thh1\] we have $$\label{eq327}
f(A(i,n-2,n-1,n))\le\frac{(n-4)(n-5)}{2}.$$ Applying Lemma \[le3\] to $A(n-2,n-1,n)$, there is some $i$, say, $i=n-3$ such that equality in (\[eq327\]) holds. It follows that $A(n-3,n-2,n-1,n)$ is permutation similar to $T_{n-4}$ and we may assume $$A=(a_{ij})=\begin{bmatrix}
T_{n-4}&x_{n-3}&x_{n-2}&x_{n-1}&x_{n}\\
y_{n-3}^T&a_{n-3,n-3}&a_{n-3,n-2}&a_{n-3,n-1}&a_{n-3,n}\\
y_{n-2}^T&a_{n-2,n-3}&a_{n-2,n-2}&a_{n-2,n-1}&a_{n-2,n}\\
y_{n-1}^T&a_{n-1,n-3}&a_{n-1,n-2}&a_{n-1,n-1}&a_{n-1,n}\\
y_{n}^T&a_{n,n-3}&a_{n,n-2}&a_{n,n-1}&a_{n,n}
\end{bmatrix}$$ where $x_{i},y_{i}\in \mathbb{R}^{n-4}$ for $i=n,n-1,n-2,n-3$.
Since $$\begin{aligned}
f(x_{n-3})+f(y_{n-3})+a_{n-3,n-3}
=f(A(n-2,n-1,n))-f(A(n-3,n-2,n-1,n))
=n-5,\end{aligned}$$ applying Lemma \[le7\] to $A(n-2,n-1,n)$ we have $a_{n-3,n-3}=0$ and $$\begin{aligned}
\label{eqn328}
f(x_{n-3})+f(y_{n-3})
=n-5\ge 4.\end{aligned}$$
We assert $$\label{eq328}
a_{n-3,n-2}a_{n-2,n-3}=0.$$ Otherwise $a_{n-3,n-2}=a_{n-2,n-3}=1$. Applying Lemma \[le10\] to $A(n-1,n)$ we can deduce $A\not\in \Gamma(n,n-4)$, a contradiction.
Similarly, we have $$a_{n-3,i}a_{i,n-3}=0 \textrm{~~for~~} i=n-1,n.$$ It follows that $$\sum\limits_{i=n-2}^{n}(a_{n-3,i}+a_{i,n-3})\le 3.$$ On the other hand, $$\sum\limits_{i=n-2}^{n}(a_{n-3,i}+a_{i,n-3})=\delta_{n-3}-f(x_{n-3})-f(y_{n-3})\ge n-2-(n-5)=3.$$ Hence, we have $\sum\limits_{i=n-2}^{n}(a_{n-3,i}+a_{i,n-3})= 3$ and $$a_{n-3,i}+a_{i,n-3}=1\textrm{~~for~~} i=n-2,n-1,n.$$ Applying Lemma \[le7\] to $A(n-3,n-1,n)$ we obtain $a_{n-2,n-2}=0$ and $$\begin{aligned}
&& f(x_{n-2})+f(y_{n-2})\\
&& =f(A(n-1,n))-f(A(n-2,n-1,n))-a_{n-2,n-2}-a_{n-3,n-2}-a_{n-2,n-3}\\
&&=n-5.
\end{aligned}$$
Repeating the above arguments, we get $$a_{n-2,i}+a_{i,n-2}=1\textrm{~~for~~} i=n-1,n$$ and $$a_{n-1,n-1}=a_{nn}=0,~~a_{n-1,n}+a_{n,n-1}=1.$$ Moreover, we have $$\label{eq330}
f(x_{n-i})+f(y_{n-i})=n-5\textrm{~~for~~}i=0,1,2,3.$$
Now we verify\
[ *[**Claim 3.**]{} $B\equiv A[n-3,n-2,n-1,n]$ is permutation similar to $T_{4}$.*]{}\
It is well known that the adjacency matrix of an acyclic digraph is permutation similar to a strictly upper triangular matrix. Suppose Claim 3 does not hold. Then the digraph $D(B)$ has at least one cycle. Note that $a_{ii}=0$ and $a_{ij}a_{ji}=0$ for $i,j=n-3,\ldots,n$. $D(B)$ has no loop or 2-cycle. If $D(B)$ has a 4-cycle, then $B$ is permutation similar to $$B'=\begin{bmatrix}
0&1&b_{13}&0\\
0&0&1&b_{24}\\
b_{31}&0&0&1\\
1&b_{42}&0&0
\end{bmatrix}.$$ Now $b_{24}=1$ implies $D(B)$ has a cycle $1\rightarrow 2\rightarrow 4\rightarrow 1$ of length $3$; $b_{42}=1$ implies $D(B)$ has a cycle $2\rightarrow 3\rightarrow 4\rightarrow 2$ of length $3$. Therefore, $D$ always has a 3-cycle with vertices from $\{n-3,n-2,n-1,n\}.$ Without loss of generality, we assume the 3-cycle is $n-3\rightarrow n-2\rightarrow n-1\rightarrow n-3$.
Applying Lemma \[le10\] to $A(n)$, we get $$f(x_{i})\le 2 \textrm{~~and~~} f(y_{i})\le 2 \textrm{~~ for~~} i=n-3,n-2,n-1.$$ By (\[eq330\]) we have $n=9$ and $$f(x_{i})=f(y_{i})=2 \textrm{~~for~~} i=n-3,n-2,n-1.$$ Suppose $a_{j_1,6}=a_{j_2,6}=a_{8,j_3}=a_{8,j_4}=1$ with $1\le j_1<j_2\le n-4=5$ and $1\le j_3<j_4\le 5$. Then $D$ has two distinct walks from $j_1$ to $j_4$ of length $n-4$: $$\left\{\begin{array}{l}
j_1\rightarrow 6\rightarrow 7\rightarrow 8\rightarrow j_{3}\rightarrow j_4,\\
j_1\rightarrow j_2\rightarrow 6\rightarrow 7\rightarrow 8\rightarrow j_4,
\end{array}\right.$$ a contradiction. Therefore, $D(B)$ is acyclic and Claim 3 holds.\
Without loss of generality, we assume $$A=\begin{bmatrix}
T_{n-4}&x_{n-3}&x_{n-2}&x_{n-1}&x_{n}\\
y_{n-3}^T&0&1&1&1\\
y_{n-2}^T&0&0&1&1\\
y_{n-1}^T&0&0&0&1\\
y_{n}^T&0&0&0&0
\end{bmatrix}.$$
For $i=n,n-1,n-2,n-3$, let $a_{s_{i},i}$ be the last nonzero component in $x_{i}$ and $a_{i,t_{i}}$ be the first nonzero component in $y_{i}$, where $s_{i}\equiv 0$ if $x_{i}=0$ and $t_{i}\equiv n-3$ if $y_i=0$. Applying Lemma \[le7\] to $A(n-2,n-1,n)$, $A(n-3,n-1,n)$, $A(n-3,n-2,n)$ and $A(n-3,n-2,n-1)$, we have $$\label{eqn331}
x_i=(a_i^T,0)^T,~~y_i=(0,b_i^T)^T \textrm{~~for~~} i=n-3,n-2,n-1,n,$$ where $a_i\in \mathbb{R}^{s_i}$, $b_i\in \mathbb{R}^{n-s_i-4}$. Moreover, by (\[eq330\]) we have $$\label{eq331}
s_{i}<t_{i}\le s_{i}+2\textrm{~~for~~}i=n,n-1,n-2,n-3.$$
Next, we verify the following claim.\
[*[**Claim 4.**]{} If $n-3\le i<j\le n$, then $$\label{eq332}
t_{j}>s_{i}+1.$$*]{}
We assert $$\label{eq34}
a_{ij}=0 \textrm{~~for~~}n-5\le i\le n-4, n-3\le j\le i+3.$$ Otherwise, $D$ has the following two distinct walks of length $n-4$ from $1$ to $j+1$ or $j+2$: $$\left\{\begin{array}{l}
1\rightarrow 2\rightarrow 4\rightarrow\cdots\rightarrow i\rightarrow j\rightarrow j+1 (\rightarrow j+2), \\
1\rightarrow 3\rightarrow 4\rightarrow\cdots \rightarrow i\rightarrow j\rightarrow j+1 (\rightarrow j+2).
\end{array}\right.$$ Similarly, we have $a_{i1}=0$ for $i=n-2,n-1,n$. Otherwise, $D$ has the following two distinct walks of length $n-4$ from $i-1$ to $n-4$: $$\left\{\begin{array}{l}
i-1\rightarrow i\rightarrow 1\rightarrow 2\rightarrow 4\rightarrow\cdots\rightarrow n-4,\\
i-1\rightarrow i\rightarrow 1\rightarrow 3\rightarrow 4\rightarrow\cdots\rightarrow n-4.
\end{array}\right.$$
Given any $n-3\le i<j\le n$, we have $$t_j\ge 2 \textrm{~~for~~} n-2\le j\le n$$ and $$s_{i}\le n-5 \textrm{~~for~~} n-3\le i\le n-1.$$ If $s_{i}=0$, then $t_{j}>s_{i}+1$ and (\[eq332\]) holds. For $s_{i}\ge 1$, if (\[eq332\]) does not hold, we can distinguish the following cases to find two distinct walks of length $n-4$ between the same endpoints to deduce contradictions.
[*Subcase 1.*]{} $t_{j}= s_{i}+1$. If $s_{i}=1$, then $t_j=2$ and $D$ has $$\left\{\begin{array}{l}
1\rightarrow i\rightarrow j\rightarrow t_{j}\rightarrow 4\rightarrow 5\rightarrow \cdots\rightarrow n-4,\\
1\rightarrow i\rightarrow j\rightarrow t_{j}\rightarrow 3\rightarrow 5\rightarrow \cdots\rightarrow n-4.
\end{array}\right.$$ If $s_{i}=2$ and $a_{1i}=0$, then by (\[eq330\]) and (\[eqn331\]), $t_j=3$, $t_{i}=s_{i}+1=3$ and $D$ has $$\left\{\begin{array}{l}
1\rightarrow 2\rightarrow i\rightarrow j\rightarrow t_{j}\rightarrow 5\rightarrow 6\rightarrow \cdots\rightarrow n-4,\\
1\rightarrow 2\rightarrow i\rightarrow t_{i}\rightarrow 4\rightarrow 5\rightarrow \cdots\rightarrow n-4.
\end{array}\right.$$ If $s_{i}=2$ and $a_{1i}=1$, then $D$ has $$\left\{\begin{array}{l}
1\rightarrow 2\rightarrow i\rightarrow j\rightarrow t_{j}\rightarrow 5\rightarrow 6\rightarrow \cdots\rightarrow n-4,\\
1\rightarrow i\rightarrow j\rightarrow t_{j}\rightarrow 4\rightarrow 5\rightarrow 6\rightarrow \cdots\rightarrow n-4.
\end{array}\right.$$ If $s_{i}=3$ and $a_{2i}=0$, then $t_{i}=s_{i}+1=4$ and $D$ has $$\left\{\begin{array}{l}
1\rightarrow 3\rightarrow i\rightarrow j\rightarrow t_{j}\rightarrow 5\rightarrow 6\rightarrow \cdots\rightarrow n-4,\\
1\rightarrow 2\rightarrow 3\rightarrow i\rightarrow t_{i}\rightarrow 5\rightarrow 6\rightarrow \cdots\rightarrow n-4.
\end{array}\right.$$ If $s_{i}=3$ and $a_{2i}=1$, then $D$ has $$\left\{\begin{array}{l}
1\rightarrow 3\rightarrow i\rightarrow j\rightarrow t_{j}\rightarrow 5\rightarrow 6\rightarrow \cdots\rightarrow n-4,\\
1\rightarrow 2\rightarrow i\rightarrow j\rightarrow t_{j}\rightarrow 5\rightarrow 6\rightarrow \cdots\rightarrow n-4.
\end{array}\right.$$ If $ 4\le s_{i}\le n-5$, then $D$ has $$\left\{\begin{array}{l}
1\rightarrow 3\rightarrow 4\rightarrow \cdots \rightarrow s_{i}\rightarrow i\rightarrow j\rightarrow t_{j} \rightarrow t_{j}+1\rightarrow \cdots\rightarrow n-4 ,\\
1\rightarrow 2\rightarrow 4\rightarrow \cdots \rightarrow s_{i}\rightarrow i\rightarrow j\rightarrow t_{j} \rightarrow t_{j}+1\rightarrow \cdots\rightarrow n-4 ,
\end{array}\right.$$ where the walk $t_j\rightarrow t_{j}+1\rightarrow \cdots\rightarrow n-4$ does not appear when $s_i=n-5$.
[*Subcase 2.*]{} $t_{j}<s_{i}+1$. Since $t_{j}\ge 2$ for $j=n-2,n-1,n$, we have $s_{i}\ge 2$. If $s_i=2$ or $3$, $D$ has the same walks as in the previous subcase. If $4\le s_{i}\le n-5$, the walks $$\left\{\begin{array}{l}
1\rightarrow 3\rightarrow 4\rightarrow \cdots\rightarrow s_{i}\rightarrow i\rightarrow j\rightarrow t_{j}\rightarrow s_{i}+1\rightarrow \cdots\rightarrow n-4 \\
1\rightarrow 2\rightarrow 4\rightarrow \cdots\rightarrow s_{i}\rightarrow i\rightarrow j\rightarrow t_{j}\rightarrow s_{i}+1\rightarrow \cdots\rightarrow n-4
\end{array}\right.$$ contain two walks of length $n-4$ with the same endpoints.
Thus we obtain Claim 4.\
Now we show
[*[**Claim 5.**]{} $s_i\ne s_{j}$$n-3\le i<j\le n$.*]{}\
Suppose $s_k=s_{l}$ for some $n-3\le k<l\le n$. Then by the definition of $s_{i}$ we have $$\label{eq335}
a_{s_{k}+1,k}=a_{s_{k}+1,l}=0.$$ From (\[eqn331\]) we have $$\label{eq336}
a_{s_{k}+1,j}+a_{j,s_{k}+1}\le 1\textrm{~~for~~ }j=n-3,n-2,n-1,n.$$ By (\[eq321\]), $$\delta_{s_{k}+1}=n-5+\sum\limits_{j=n-3}^{n}(a_{s_{k}+1,j}+a_{j,s_{k}+1})\ge n-2.$$ Combining this with (\[eq335\]) and (\[eq336\]) we have $$a_{k,s_{k}+1}+a_{l,s_{k}+1}\ge1.$$ By (\[eq331\]) and (\[eq332\]), we have $t_{l}=s_{l}+2$. Hence, $$a_{l,s_{k}+1}=0\textrm{~~and~~ } a_{k,s_{k}+1}=1.$$ It follows that $$\label{eq337}
t_{k}=s_{k}+1.$$ If $s_{k}=0$, then $D$ has two distinct walks of length $n-4$ from $k$ to $n-4$: $$\left\{\begin{array}{l}
k\rightarrow 1\rightarrow 2\rightarrow \cdots \rightarrow n-4,\\
k\rightarrow l\rightarrow 2\rightarrow \cdots \rightarrow n-4,
\end{array}\right.$$ a contradiction. If $1\le s_{k}\le n-6$, then $D$ has two distinct walks of length $n-4$ from $1$ to $n-4$: $$\left\{\begin{array}{l}
1\rightarrow 2\rightarrow \cdots \rightarrow s_{k}\rightarrow k\rightarrow l\rightarrow t_{l}\rightarrow t_{l}+1\rightarrow\cdots\rightarrow n-4,\\
1\rightarrow 2\rightarrow \cdots \rightarrow s_{k}\rightarrow k\rightarrow t_{k}\rightarrow t_{k}+1\rightarrow\cdots\rightarrow n-4,
\end{array}\right.$$ a contradiction. Hence we have $s_k\ge n-5$.
On the other hand, by (\[eq34\]) we have $$s_{n-3}\le n-6, s_{n-2}\le n-6\textrm{ and }s_{n-1}\le n-5.$$ It follows that $$s_{n-3}\ne s_{n-2},~~k=n-1,~~l=n\textrm{ ~~and~~} s_k=n-5.$$ Moreover, there exists $s_m\ge 1$ with $m\in \{n-3,n-2\}$. Now we can distinguish the following cases to find distinct walks of length $n-4$ with the same endpoints in $D$ to deduce contradictions.
If $t_{m}=s_{m}+1$, then by (\[eq337\]) $D$ has $$\left\{\begin{array}{l}
1\rightarrow 2\rightarrow \cdots \rightarrow s_{m}\rightarrow m\rightarrow s_{m}+1\rightarrow\cdots\rightarrow n-4,\\
1\rightarrow 2\rightarrow \cdots \rightarrow s_{k}\rightarrow k\rightarrow n-4.
\end{array}\right.$$ If $s_{m}+2=t_{m}\le n-5$, then $D$ has $$\left\{\begin{array}{l}
1\rightarrow 2\rightarrow \cdots \rightarrow s_{m}\rightarrow m\rightarrow t_{m}\rightarrow \cdots\rightarrow s_k\rightarrow k\rightarrow n-4,\\
1\rightarrow 2\rightarrow \cdots \rightarrow s_{k}\rightarrow k\rightarrow n-4.
\end{array}\right.$$ If $s_{m}+2=t_{m}= n-4$, then $s_{m}=n-6$ and $D$ has $$\left\{\begin{array}{l}
1\rightarrow 2\rightarrow \cdots \rightarrow s_m\rightarrow m\rightarrow k\rightarrow n-4,\\
1\rightarrow 2\rightarrow \cdots \rightarrow s_{k}\rightarrow k\rightarrow n-4.
\end{array}\right.$$ Hence we get Claim 5.\
Combining Claim 4 and Claim 5 we obtain $$s_{j}>s_{i} \textrm{~~for~~} n-3\le i<j\le n.$$ Otherwise, $s_{j}<s_{i}$ leads to $t_{j}\le s_{j}+2< s_{i}+2$, which contradicts Claim 4. Therefore, we have $$\label{eq338}
0\le s_{n-3}<s_{n-2}<s_{n-1}<s_{n}\le n-4.$$
Finally, we verify
[*[**Claim 6.**]{} $t_{i}=s_{i}+2$ for $i\in\{n-3,n-2,n-1,n\}$.*]{}\
Suppose $t_{n}=s_{n}+1$. By (\[eq338\]) and (\[eq331\]) we have $1\le s_{n-2}\le n-6$ and $s_{n}\ge t_{n-2}$. We can distinguish the following cases to find two distinct walks of length $n-4$ from $1$ to $n-4$ or $n$ in $D$, which contradicts $D\in Ex(n,\mathscr{F}_{n-4})$. If $t_{n-2}=s_{n-2}+2$, then $D$ has $$\left\{\begin{array}{l}
1\rightarrow 2\rightarrow \cdots\rightarrow s_{n}\rightarrow n \rightarrow s_{n}+1
\rightarrow s_{n}+2\rightarrow \cdots\rightarrow n-4 , \\
1\rightarrow 2\rightarrow \cdots\rightarrow s_{n-2}\rightarrow n-2\rightarrow s_{n-2}+2\rightarrow \cdots
\rightarrow s_{n}\rightarrow n \rightarrow s_{n}+1\rightarrow \cdots\rightarrow n-4 ,
\end{array}\right.$$ where the walk $n \rightarrow s_{n}+1\rightarrow \cdots\rightarrow n-4$ does not appear when $s_n=n-4$. If $t_{n-2}=s_{n-2}+1$ and $s_n\le n-5$, then $D$ has $$\left\{\begin{array}{l}
1\rightarrow 2\rightarrow \cdots\rightarrow s_{n}\rightarrow n\rightarrow s_{n}+1
\rightarrow s_{n}+2\rightarrow \cdots\rightarrow n-4,\\
1\rightarrow 2\rightarrow \cdots\rightarrow s_{n-2}\rightarrow n-2\rightarrow s_{n-2}+1
\rightarrow s_{n-2}+2\rightarrow \cdots\rightarrow n-4.
\end{array}\right.$$ If $t_{n-2}=s_{n-2}+1$ and $s_n= n-4$, then $a_{n-4,n}=1 $ and $D$ has $$\left\{\begin{array}{l}
1\rightarrow 2\rightarrow \cdots\rightarrow n-5\rightarrow n-4\rightarrow n,\\
1\rightarrow 2\rightarrow \cdots\rightarrow s_{n-2}\rightarrow n-2\rightarrow t_{n-2}\rightarrow t_{n-2}+2\rightarrow \cdots\rightarrow n-4\rightarrow n, \textrm{ if }s_{n-2}\le n-7,\\
1\rightarrow 3\rightarrow \cdots\rightarrow s_{n-2}\rightarrow n-2\rightarrow t_{n-2}\rightarrow t_{n-2}+1\rightarrow \cdots\rightarrow n-4\rightarrow n, \textrm{ if }s_{n-2}= n-6.
\end{array}\right.$$ Hence, $t_{n}=s_{n}+2$ and $s_n\le n-5$.
Next suppose $t_{i}=s_{i}+1$ for $i\in\{n-2,n-1\}$ and $j\in \{n-2,n-1\}\setminus\{i\}$. Then $s_i,s_j>0$. If $t_j=s_j+1$, then $D$ has the following two distinct walks of length $n-4$ from $1$ to $n-4$: $$\left\{\begin{array}{l}
1\rightarrow 2 \rightarrow \cdots \rightarrow s_{i}\rightarrow i\rightarrow s_{i}+1\rightarrow \cdots\rightarrow n-4,\\
1\rightarrow 2 \rightarrow \cdots \rightarrow s_{j}\rightarrow j\rightarrow s_{j}+1\rightarrow \cdots\rightarrow n-4,
\end{array}\right.$$ a contradiction. If $t_j=s_j+2$, then $D$ has the following two distinct walks of length $n-4$ from $1$ to $n-4$: $$\left\{\begin{array}{l}
1 \rightarrow \cdots \rightarrow s_{i}\rightarrow i\rightarrow s_{i}+1\rightarrow \cdots\rightarrow n-4,\\
1 \rightarrow \cdots \rightarrow s_{i}\rightarrow i\rightarrow s_{i}+1\rightarrow \cdots\rightarrow s_{j}\rightarrow j\rightarrow s_{j}+2\rightarrow \cdots\rightarrow n-4,~~\textrm{if}~~i<j,\\
1 \rightarrow \cdots \rightarrow s_{j}\rightarrow j\rightarrow s_{j}+2\rightarrow \cdots\rightarrow s_{i}\rightarrow i\rightarrow s_{i}+1\rightarrow \cdots\rightarrow n-4,~~\textrm{if}~~i>j \textrm{ and } s_i\ge s_j+2\\
1 \rightarrow \cdots \rightarrow s_{j}\rightarrow j\rightarrow i\rightarrow s_{i}+1\rightarrow \cdots\rightarrow n-4,~~\textrm{if}~~i>j \textrm{ and } s_i=s_j+1,
\end{array}\right.$$ a contradiction. Hence, we get $$t_{n-2}=s_{n-2}+2,\quad t_{n-1}=s_{n-1}+2.$$
Now we conclude $t_{n-3}=s_{n-3}+2$. Otherwise $D$ has the following two distinct walks of length $n-4$ from $1$ or $n-3$ to $n-4$: $$\left\{\begin{array}{l}
1\rightarrow 2\rightarrow \cdots \rightarrow s_{n-3}\rightarrow n-3\rightarrow t_{n-3}\rightarrow \cdots\rightarrow n-4,\\
1\rightarrow 2\rightarrow \cdots \rightarrow s_{n-3}\rightarrow n-3\rightarrow t_{n-3}\rightarrow \cdots\rightarrow s_{n-1}\rightarrow n-1\rightarrow s_{n-1}+2\rightarrow \cdots\rightarrow n-4,
\end{array}\right.$$ where the walk $1\rightarrow 2 \rightarrow \cdots \rightarrow s_{n-3} \rightarrow n-3$ does not appear when $s_{n-3}=0$. Thus we get Claim 6.\
Finally, combining (\[eq330\]) and Claim 6 we have $A=F_4(n)$. This completes the proof.
From the proof of Theorem \[th9\], we have the following corollary.
\[co1\] Let $k\ge 5$ be an integer, $n=k+4$ and $$A=(a_{ij})=\begin{bmatrix}
T_{k}&B\\
C&E
\end{bmatrix}\in \Gamma(n,k).$$ If $$\begin{aligned}
f(A)=\frac{n(n-1)}{2}-4, &&f(A(n))=\frac{(n-1)(n-2)}{2}-3,\\
f(A(n-1,n))=\frac{(n-2)(n-3)}{2}-2, &&f(A(n-2,n-1,n))=\frac{(n-3)(n-4)}{2}-1,
\end{aligned}$$ then $A$ is permutation similar to $F_4(n)$ by permuting its last 4 rows and columns.
Proof of Theorem \[thh2\]
=========================
In this section we give the proof of Theorem \[thh2\]. We will use induction on $n$. First we need the following lemma to show that Theorem \[thh2\] holds for $k=5$.
\[le14\] Let $n\ge 10$ be an integer. Then $$\label{eq339}
ex(n,\mathscr{F}_{n-5})=\frac{n(n-1)}{2}-5.$$ Moreover, a digraph $D$ is in $Ex(n,\mathscr{F}_{n-5})$ if and only if $D$ is a $(1,n-5,5)$-completely transitive tournament.
Let $D$ be any $ \mathscr{F}_{n-5}$-free digraph of order $n$. Denote by $A\equiv A_D$. Given any $i\in \{1,2,\ldots,n\}$, since $A(i)\in \Gamma(n-1,n-5)$, by Theorem \[le8\] we have $$\label{eqq41}
f(A(i))\le \frac{(n-1)(n-2)}{2}-4.$$ Applying Lemma \[le3\] to $A$, we have $$f(A)\le \frac{n(n-1)}{2}-5.$$ Hence, $$ex(n,\mathscr{F}_{n-5})\le \frac{n(n-1)}{2}-5.$$
Let $D$ be any $(1,n-5,5)$-completely transitive tournament. Then $A_D$ is a principal submatrix of $A'=
J_2\otimes T_{n-5}
$. Since $A'(\alpha)\in\Gamma(n,n-5)$, the digraph $D(A')$, and hence $D$ is $\mathscr{F}_{n-5}$-free. It is clear that $D$ has size $\frac{n(n-1)}{2}-5$. Thus we get (\[eq339\]) and the sufficiency of the second part.\
Let $D\in Ex(n,\mathscr{F}_{n-5})$ and $A\equiv A_D$. Again, denote by $\delta_i$ the number of nonzero entries lying in the $i$-th row and the $i$-th column of $A$. Then by (\[eqq41\]), $$\label{eq40}
\delta_i=f(A)-f(A(i))\ge n-2 \textrm{~~for all~~}1\le i\le n.$$
Applying Lemma \[le3\] we get $$f(A(i_0))\ge
\frac{(n-1)(n-2)}{2}-4$$ for some $i_0\in \{1,2,\ldots,n\}$. Without loss of generality, we assume $i_0=n$. By Theorem \[th9\], $A(n)$ is permutation similar to $F_t(n-1)$ with $t\in\{1,2,3,4\}$. Now we distinguish four cases.
[*Case 1.*]{} $A(n)$ is permutation similar to $F_1(n-1)$. Without loss of generality, we assume $$A=\begin{bmatrix}
0&J_{3,n-6}&J_{3,2}&x_{1}\\
0&T_{n-6}&J_{n-6,2}&x_{2}\\
0&0&0&x_{3}\\
y_{1}^{T}&y_{2}^{T}&y_{3}^{T}&\alpha
\end{bmatrix},$$ where $x_{1},y_{1}\in \mathbb{R}^{3}$, $x_{2},y_{2}\in \mathbb{R}^{n-6}$, $x_{3}, y_{3}\in \mathbb{R}^{2}$.
Since $$\delta_{n}=\sum\limits_{i=1}^{3}[f(x_{i})+f(y_{i})]+\alpha
=f(A)-f(A(n))=n-2,$$ applying Lemma \[le4\] to $A$ we have $$y_{1}=0\textrm{ and }x_{3}=0.$$ Then $\delta_1\le n-3$, which contradicts (\[eq40\]).
[*Case 2.*]{} $A(n)$ is permutation similar to $F_2(n-1)$. Applying the same arguments as in Case 1 we get $\delta_{n-2}\le n-3$, a contradiction.
[*Case 3.*]{} $A(n)$ is permutation similar to $F_3(n-1)$. Without loss of generality, we assume $$A=\begin{bmatrix}
0&J_{2,n-6}&J_{2,2}&J_{2,1}&x_1\\
0&T_{n-6}&J_{n-6,2}&U_{m}&x_2\\
0&0&0&0&x_3\\
0&U'_{m}&J_{1,2}&0&a_{n-1,n}\\
y_1^T&y_2^T&y_3^T&a_{n,n-1}&a_{nn}
\end{bmatrix}$$ where $x_1,x_3,y_1,y_3\in \mathbb{R}^2$, $x_2,y_2\in \mathbb{R}^{n-6}$, $U_{m}=(J_{1,m},0)^T$, $U'_{m}=(0,J_{1,n-m-8})$, $0\le m\le n-8$. By (\[eq40\]), $$\delta_{n-1}=n-4+a_{n-1,n}+a_{n,n-1}\ge n-2$$ implies $a_{n-1,n}=a_{n,n-1}=1$. Applying Lemma \[le4\] to $A(n-1)$ we get $x_3=0$ and $y_1=0$. Hence, $\delta_1\ge n-2$ and $\delta_2\ge n-2$ force $x_1=J_{2,1}$. Then $D$ has the following two distinct walks of length $n-5$ from 1 to $n-1$ or $n$: $$\left\{\begin{array}{l}
1\rightarrow 2\rightarrow n-1 \rightarrow n \rightarrow n-1\rightarrow\cdots \rightarrow n-1(\rightarrow n),\\
1\rightarrow n\rightarrow n-1 \rightarrow n \rightarrow n-1\rightarrow\cdots \rightarrow n-1(\rightarrow n),
\end{array}\right.$$ a contradiction.
[*Case 4.*]{} $A(n)$ is permutation similar to $F_4(n-1)$. Without loss of generality we assume $$A=(a_{ij})=\begin{bmatrix}
T_{n-5}&w_{4}&w_{3}&w_{2}&w_{1}&x\\
u_{4}&0&1&1&1&a_{n-4,n}\\
u_{3}&0&0&1&1&a_{n-3,n}\\
u_{2}&0&0&0&1&a_{n-2,n}\\
u_{1}&0&0&0&0&a_{n-1,n}\\
y^{T}&a_{n,n-4}&a_{n,n-3}&a_{n,n-2}&a_{n,n-1}&a_{n,n}
\end{bmatrix}\equiv \begin{bmatrix} T_{n-5}&B\\C&E\end{bmatrix},$$ where $x,y\in \mathbb{R}^{n-5}$, $$w_{i}=(J_{1,q_i},0)^T,~~ u_{i}=(0,J_{1,n-q_i-6}) \textrm{~~for~~} i\in \{1,2,3,4\}$$ with $0\leq q_4<q_3<q_2<q_1\leq n-6$.
We claim $$\label{eq41}
a_{n-i,n}a_{n,n-i}=0 \textrm{~~for~~}i=1,2,3,4.$$ Otherwise suppose $a_{n-i,n}=a_{n,n-i}=1$ for some $i\in \{1,2,3,4\}$. Set $$\alpha=\{1,2,\ldots,n-5,n-i,n\}.$$ Applying Lemma \[le10\] to $A[\alpha]$ we get $A\not\in \Gamma(n,n-5)$, a contradiction. Thus we obtain (\[eq41\]).
On the other hand, by (\[eq40\]) we have $$a_{n,n-i}+a_{n-i,n}=\delta_{n-i}-f(w_i)-f(u_i)-3\ge 1\textrm{~~for~~} i\in \{1,2,3,4\}.$$ Hence, we have $a_{n,n-i}+a_{n-i,n}=1$, and $$f(A(n-i))=f(A)-\delta_{n-i}=\frac{(n-1)(n-2)}{2}-4\textrm{~~for~~}i=1,2,3,4.$$
Given $i\in\{1,2,3,4,5\}$, applying Corollary \[co1\] to $A(n-i)$ we know each $4\times 4$ principal submatrix of $E$ is permutation similar to $T_4$. Let $w_5=x$ and $u_5=y^T$. By Lemma 9 of [@HZ1], $E$ is permutation similar to $T_5$ and $A$ is permutation similar to $$H=\begin{bmatrix}
T_{n-5}&w_{\sigma_1}&w_{\sigma_2}&w_{\sigma_3}&w_{\sigma_4}&w_{\sigma_5}\\
u_{\sigma_1}&0&1&1&1&1\\
u_{\sigma_2}&0&0&1&1&1\\
u_{\sigma_3}&0&0&0&1&1\\
u_{\sigma_4}&0&0&0&0&1\\
u_{\sigma_5}&0&0&0&0&0
\end{bmatrix}$$ with $\sigma$ a permutation of $\{1,2,3,4,5\}$. Applying Corollary \[co1\] to each $A(n-i)$ again we get $$w_{\sigma_i}=(J_{1,k_i},0)^T,~~ u_{\sigma_i}=(0,J_{1,n-k_i-6})\textrm{~~for~~} i=1,2,3,4,5$$ with $0\le k_{1}<k_{2}<k_{3}<k_{4}<k_{5}\le n-6$.
Denote $G=J_2\otimes T_{n-5}$ and $$\beta=\{n-4,n-5,\ldots,2(n-5)\}
\setminus\{n-4+k_1,n-4+k_2,n-4+k_3,n-4+k_4,n-4+k_5\}.$$ Then $H=G(\beta)$ and $D$ is a $(1,n-5,5)$-completely transitive tournament.
Applying the same proof of Lemma \[le3\], we have the following lemma.
\[le13\] Let $n \ge 3$, $p\ge 1$, $q$ be nonnegative integers, and let $A\in M_{n}\{0,1\}$. If $$f(A(i))\le \frac{(n-1)(n-2)}{2}-p\frac{n-1}{2}-q \textrm{ for all } 1\le i\le n,$$ then $$f(A)\le \frac{n(n-1)}{2}-(p\frac{n-1}{2}+q)-(p+1).$$
Now we are ready to present the proof of Theorem \[thh2\].\
[**Proof of Theorem \[thh2\].**]{} We use induction on $n$. By Theorem \[le8\] and Lemma \[le14\] we know the results hold for $n=k+4$ and $n=k+5$. Assume the results hold for $n=k+5,\ldots,sk+t$, where $0\le t<k$ and $s>0$ are integers.
Now suppose $n=sk+t+1.$ Let $u,v$ be integers such that $v<k$ and $ n=uk+v$. Then $u=s$, $v=t+1$ when $t<k-1$, and $u=s+1$, $v=0$ when $t=k-1$.
to prove $$\label{eq342}
ex(n,\F)\le\frac{n(n-1)}{2}-\frac{(s-1)n}{2}-\frac{(s+1)t}{2}.$$ By Theorem \[le8\] we know (\[eq342\]) holds for $n=k+4$.
Given any $ \mathscr{F}_{k}$-free digraph $D$ of order $n$, denote by $A\equiv A_D$ its adjacency matrix. For any $ i\in\{1,2,\ldots,n\}$, since the digraph of $A(i)$ is an $\mathscr{F}_{k}$-free digraph of order $n-1$, by the induction hypothesis we have $$\begin{aligned}
\label{eq342}
\nonumber f(A(i))&\le& ex(n-1,\mathscr{F}_{k})\\
\nonumber &= &{n-1 \choose 2} -{s\choose 2}k-st\\ &=&\frac{(n-1)(n-2)}{2}-\frac{(s-1)(n-1)}{2}-\frac{(s+1)t}{2}.\end{aligned}$$ Applying Lemma \[le3\] we have $$\begin{aligned}
f(A)&\le& \frac{n(n-1)}{2}-\frac{(s-1)(n-1)}{2}-\frac{(s+1)t}{2}-s\\
&=&\frac{n(n-1)}{2}-\frac{(s-1)n}{2}-\frac{(s+1)(t+1)}{2}\\
&=&\frac{n(n-1)}{2}-\frac{(u-1)n}{2}-\frac{(u+1)v}{2}.
\end{aligned}$$ Hence, $$\begin{aligned}
ex(n,\mathscr{F}_{k})\le
\frac{n(n-1)}{2}-\frac{(u-1)n}{2}-\frac{(u+1)v}{2}.
\end{aligned}$$
On the other hand, if $D$ is a $(u,k,v)$-completely transitive tournament, then there exist $k-t$ numbers $j_1,j_2,\ldots,j_{k-t}\in \{uk+1,uk+2,\ldots,(u+1)k\}$ such that $$PAP^T=\Pi_{u+1,k}(j_1,j_2,\ldots,j_{k-v})$$ for some permutation matrix $P$. Since $(\Pi_{u+1,k})^k=0$, we have $(A_D)^k=0$, and hence $D$ is $\mathscr{F}_{k}$-free. Moreover, the size of $D$ is $$f(A)=\frac{n(n-1)}{2}-\frac{(u-1)n}{2}-\frac{(u+1)v}{2}.$$ Hence, $$\begin{aligned}
\label{eq43}
ex(n,\mathscr{F}_{k})=
\frac{n(n-1)}{2}-\frac{(u-1)n}{2}-\frac{(u+1)v}{2}={n \choose 2}-{u \choose 2}k-uv
\end{aligned}$$ and any $(u,k,v)$-completely transitive tournament is in $Ex(n,\mathscr{F}_{k})$.\
Conversely, suppose $D\in Ex(n,\mathscr{F}_{k})$ and $A\equiv A_D$. Applying Lemma \[le3\] to $A$, by (\[eq43\]) we know there is some $i_0\in \{1,2,\ldots,n\}$ such that equality in (\[eq342\]) holds. Without loss of generality, we assume $i_0=n$. We distinguish two cases.
[*Case 1.*]{} $t=0$. Then $u=s\ge 2$ and $v=1$. By the induction hypothesis we may assume $$\label{eq344}
A=\begin{bmatrix}
T_{k}&T_{k}&\cdots&T_{k}&x_1\\
T_{k}&T_{k}&\cdots&T_{k}&x_2\\
\vdots&\vdots&\ddots&\vdots&\vdots\\
T_{k}&T_{k}&\cdots&T_{k}&x_s\\
y_1^T&y_2^T&\ldots&y_s^T&a_{n,n}
\end{bmatrix}\in M_{sk+1}\{0,1\},$$where $x_{i},y_{i}\in \mathbb{R}^{k}$ for $i=1,2,\cdots,s$.
Let $a_{s_{i},n}$ be the last nonzero component in $x_{i}$, and $a_{n,t_{i}}$ be the first nonzero component in $y_{i}$ for $i=1,\ldots, s$. Here we define $s_i=(i-1)k$ when $x_{i}=0$, and $t_{i}=ik+1$ when $y_i=0$.
We claim $$\label{eq44}
t_{i}\ne s_{i}+1 \textrm{~~for~~} 1\le i\le s.$$ Note that we can change the role of each $(x_i,y_i)$ by permutation similarity. To prove (\[eq44\]) it suffices to verify the case $i=1$. Suppose $t_1=s_1+1$. If $s_{1}\le 2$, then $D$ has two distinct walks of length $k$ from 1 or $n$ to $k$: $$\left\{\begin{array}{l}
1\rightarrow \cdots\rightarrow s_{1}\rightarrow n\rightarrow s_{1}+1\rightarrow s_{1}+2\rightarrow s_{1}+3\rightarrow \cdots \rightarrow k,\\
1\rightarrow \cdots\rightarrow s_{1}\rightarrow n\rightarrow s_{1}+1\rightarrow k+s_{1}+2\rightarrow s_{1}+3\rightarrow \cdots\rightarrow k,
\end{array}\right.$$ where the walk $1\rightarrow \cdots\rightarrow s_{1}$ does not appear when $s_1=0$. If $3\le s_{1}\le k$, then $D$ has two distinct walks of length $k$ from 1 to $n$ or $k$: $$\left\{\begin{array}{l}
1\rightarrow 2\rightarrow 3\rightarrow \cdots\rightarrow s_{1}\rightarrow n\rightarrow s_{1}+1\rightarrow s_{1}+2\rightarrow \cdots\rightarrow k,\\
1\rightarrow k+2\rightarrow 3\rightarrow \cdots\rightarrow s_{1}\rightarrow n\rightarrow s_{1}+1\rightarrow s_{1}+2\rightarrow \cdots\rightarrow k,
\end{array}\right.$$ where the walk $n\rightarrow s_{1}+1\rightarrow s_{1}+2\rightarrow \cdots\rightarrow k$ does not appear when $s_1=k$. This contradicts $D\in Ex(n,\F)$ and (\[eq44\]) follows.
Denote by $$B_i=\begin{bmatrix}T_k&x_i\\y_i^T&a_{nn}\end{bmatrix} \textrm{~~for~~} 1\le i\le s.$$ Given any $i\in \{1,2,\ldots,s\}$, since $B_i$ is a principal submatrix of $A$, then $B_i\in \Gamma(k+1,k)$. By Theorem \[thh1\] we have $f(B_i)\le k(k+1)/2$ and $$\label{eq45}
f(x_i)+f(y_i)+a_{nn}=f(B_i)-f(T_k)\le k.$$ If equality in (\[eq45\]) holds, then applying Lemma \[le7\] to $B_i$ we get $t_i=s_i+1$, which contradicts (\[eq44\]). Therefore, we have $$f(x_i)+f(y_i)+a_{nn}\le k-1 \textrm{~~for~~} 1\le i\le s.$$
Since $$\begin{aligned}
n-s-1= s(k-1)&\ge& \sum_{i=1}^s[f(x_i)+f(y_i)+a_{nn}]\\
&\ge& \sum_{i=1}^s[f(x_i)+f(y_i)]+a_{nn}\\
& =&f(A)-s^2f(T_k)=n-s-1\end{aligned}$$ we get $a_{nn}=0$ and $$f(x_i)+f(y_i)= k-1 \textrm{~~for~~} 1\le i\le s.$$ By (\[eq44\]), applying Lemma \[le7\] to each $B_i$ again we have $$t_i=s_i+2 \textrm{~~and~~}
x_i=(J_{1, q_i},0)^T,~~y_i=(0,J_{1, k-q_i-1})^T \textrm{~~with~~} q_i\in\{0,1,\ldots,k-1\}$$ for all $1\le i\le s$. Now we assert $$q_1=q_2=\cdots=q_s=q \textrm{~~for some~~} q\in\{0,1,2,\ldots,k-1\}.$$ Otherwise, without loss of generality we assume $q_1< q_2$. Then $$s_1=q_1, s_2=k+q_2.$$ Denote by $P$ the permutation matrix obtained by interchanging row $q_1+1$ and row $k+q_1+1$ of the $n\times n$ identity matrix. Let $A'=(a'_{ij})=PAP^T$. Then $A'\in Ex(n,\F)$ has the same form (\[eq344\]) as $A$. Moreover, the last nonzero component in $x_1$ is $a'_{q_1+1,n}=a_{k+q_1+1,n}$; the first nonzero component in $y_1$ is $a'_{n,q_1+2}=a_{n,q_1+2}$. If we define $s'_i$ and $t'_i$ for $A'$ the same as $s_i$ and $t_i$ for $A$, then $$s'_1=q_1+1, \textrm{~~and~~}
t'_1= q_1+2.$$ On the other hand, using the same arguments as above we have $t'_1=s'_1+2$, a contradiction.
Therefore, we have $x_{i}=(J_{1,q},0)^T$, $y_{i}=(0,J_{1,k-q-1})^T$ for $i= 1,\ldots,s$, and $A$ is permutation similar to $\Pi_{s+1 ,k}(\beta)$ with $\beta=\{sk+1,sk+2,\ldots,sk+k\}\setminus\{q+1\}$. Hence, $D$ is a $(u,k,v)$-completely transitive tournament.
Case 2. $t\ne 0$. By the induction hypothesis we may assume $$A=\begin{bmatrix}
T_{k}&T_{k}&\cdots&T_{k}&w_1&w_2&\cdots&w_{t}&x_1\\
T_{k}&T_{k}&\cdots&T_{k}&w_1&w_2&\cdots&w_{t}&x_2\\
\vdots&\vdots&&\vdots&\vdots&\vdots&&\vdots&\vdots\\
T_{k}&T_{k}&\cdots&T_{k}&w_1&w_2&\cdots&w_{t}&x_{s}\\
u_1&u_1&\cdots&u_1&0&1&\cdots&1&a_{sk+1,n}\\
u_2&u_2&\cdots&u_2&0&0&\ddots&\vdots&\vdots\\
\vdots&\vdots&&\vdots&\vdots&\vdots&\ddots&1&a_{sk+t-1,n}\\
u_{t}&u_{t}&\cdots&u_{t}&0&0&\cdots&0&a_{sk+t,n}\\
y_1^T&y_2^T&\ldots&y_{s}^T&a_{n,sk+1}&a_{n,sk+2}&\cdots&a_{n,sk+t}&a_{nn}
\end{bmatrix}\in M_{s k+t+1}\{0,1\},$$where $x_{i},y_{i}\in \mathbb{R}^{k}$ for $i=1,2,\ldots,s$, $w_j=(J_{1,k_j},0)^T$, $u_j=(0,J_{1,k-1-k_j})$, and $0\leq k_1<k_2<\cdots<k_{t}\leq k-1$.
Let $i\in\{1,2,\ldots,t\}$. We claim $$\label{eq47}
a_{sk+i,n}+a_{n,sk+i}=1.$$ Since $A(sk+i)\in \Gamma(n-1,k)$, we have $$\begin{aligned}
\label{eq348}
a_{sk+i,n}+a_{n,sk+i}&=&f(A)-f(A(sk+i))- s[f(w_i)+f(u_i)]-
(t-1) \nonumber\\
&\ge& f(A)-ex(n-1,\F)- s(k-1)-(t-1)=1.\end{aligned}$$ If $a_{sk+i,n}=a_{n,sk+i}=1$, setting $$\alpha=\{1,2,\ldots,k,sk+i,n\}$$ and applying Lemma \[le10\] to $A[\alpha]$ we get $A[\alpha]\not\in \Gamma(k+2,k)$, which contradicts $A\in \Gamma(n,k)$. Then $A[\alpha]\in \Gamma(k+2,k)$. Applying Lemma \[le10\] to $A[\alpha]$ we get $$\begin{aligned}
f(w_i)\le 2\textrm{ and } f(u_i)\le 2.\end{aligned}$$ Hence, $$5\le k=f(w_i)+f(u_i)+1\le 5$$ implies $k=5$ and $f(w_i)=f(u_i)=2$. Then $$A[\alpha]=\begin{bmatrix}
0&1&1&1&1&1&a_{1n}\\
0&0&1&1&1&1&a_{2n}\\
0&0&0&1&1&0&a_{3n}\\
0&0&0&0&1&0&a_{4n}\\
0&0&0&0&0&0&a_{5n}\\
0&0&0&1&1&0&a_{sk+i,n}\\
a_{n1}&a_{n2}&a_{n3}&a_{n4}&a_{n5}&a_{n,sk+i}&a_{nn}
\end{bmatrix}.$$ It follows that $D$ has two walks of length 5 from 1 to 5: $$\left\{\begin{array}{l}
1\rightarrow sk+i\rightarrow n\rightarrow sk+i\rightarrow 4\rightarrow 5,\\
1\rightarrow 2\rightarrow sk+i\rightarrow n\rightarrow sk+i\rightarrow 5,\\
\end{array}\right.$$ a contradiction. Thus we get (\[eq47\]).
Combining (\[eq47\]) and (\[eq348\]) we get $$f(A(sk+i))=ex(n-1,\F).$$ By the induction hypothesis, $A(sk+i)$ is permutation similar to a submatrix of $\Pi_{s+1,k}$. Therefore, $a_{nn}=0$ $$\sum_{i=1}^s[f(x_i)+f(y_i)]=f(A)-f(A(n))-a_{nn}-
\sum_{i=1}^t(a_{sk+i,n}+a_{n,sk+i})=s(k-1).$$
Next we distinguish two subcases.
[*Subcase 1.*]{} $s>1$. Let $\alpha
=\{sk+1,sk+2\ldots,sk+t\}$. We consider $A(\alpha)$. Then $$\begin{aligned}
f(A(\alpha))&=&f(A)-s\sum_{i=1}^t[f(w_i)+f(u_i)]-t(t-1)/2-\sum_{i=1}^t(a_{sk+i,n}+a_{n,sk+i})-a_{nn}\\
&=&\frac{n(n-1)}{2}-\frac{(s-1)n}{2}-\frac{(s+1)(t+1)}{2}-st(k-1)-t(t-1)/2-t\\
&=& \frac{(n-t)(n-t-1)}{2}-\frac{(s-1)(n-t)}{2}-\frac{s+1}{2}\\
&=&ex(n-t,\F).\end{aligned}$$ Applying Case 1 to $A(\alpha)$, we have $$x_j=(J_{1,q},0)^T \textrm{~~and~~} y^T_j=(0,J_{1,k-1-q})^T \textrm{~~for all~~}j\in \{1,\ldots,s\}$$ with $ 0\le q\le k-1$.
We assert $q\not\in \{k_1,k_2,\ldots,k_t\}$. Otherwise suppose $q=k_i$ for some $i$. Since $$a_{q+1,n}=a_{n,q+1}=a_{q+1,sk+i}=a_{sk+i,q+1}=0,$$ we have $$\delta_{q+1}\le s(k-1)+t-1=n-s-2$$ and $$f(A(q+1))=f(A)-\delta_{q+1}>ex(n-1,\F),$$ which contradicts $A(q+1)\in \Gamma(n-1,k)$.
Next we show that $$\label{eq49}
a_{sk+i,n}=1 \textrm{~~when~~}q>k_i \textrm{~~for~~} i=1,\ldots,t.$$ Otherwise suppose $q>k_i$ and $a_{sk+i,n}=0$. Then by (\[eq47\]) we have $a_{sk+i,q+1}=a_{n,sk+i}=1$. If $q\le 2$, then $D$ has two distinct walks of length $k$ from $1$ to $k$: $$\left\{\begin{array}{l}
1\rightarrow \cdots\rightarrow q\rightarrow n\rightarrow sk+i\rightarrow q+1\rightarrow q+2\rightarrow q+3\rightarrow \cdots\rightarrow k-1\rightarrow k,\\
1\rightarrow \cdots\rightarrow q\rightarrow n\rightarrow sk+i\rightarrow q+1\rightarrow k+q+2\rightarrow q+3\rightarrow \cdots\rightarrow k,
\end{array}\right.$$ a contradiction. If $q\ge 3$, then $D$ has two distinct walks of length $k$ from $1$ to $k$: $$\left\{\begin{array}{l}
1\rightarrow 3\rightarrow \cdots\rightarrow q\rightarrow n\rightarrow sk+i\rightarrow q+1\rightarrow \cdots\rightarrow k-1\rightarrow k,\\
1\rightarrow 2\rightarrow \cdots\rightarrow q-1\rightarrow n\rightarrow sk+i\rightarrow q+1\rightarrow \cdots\rightarrow k-1\rightarrow k,
\end{array}\right.$$ a contradiction.
Using similar arguments as above, we can deduce $a_{sk+i,n}=0$ when $q<k_i$. Let $$\beta=\{sk+1,sk+2,\ldots,n\}\setminus\{sk+k_1+1,sk+k_2+1,\ldots,sk+k_t+1,sk+q+1\}.$$ If $q>k_t$, then $A=\Pi_{s+1,k}(\beta)$. Otherwise let $$\begin{aligned}
P=\left\{\begin{array}{ll}
\begin{bmatrix}
0&1\\
I_{t}&0\end{bmatrix},&\textrm{ if }q<k_1,\\
\begin{bmatrix}
I_{j}&0&0\\
0&0&1\\
0&I_{t-j}&0\end{bmatrix},&\textrm{ if } k_j<q<k_j+1,
\end{array}\right.\end{aligned}$$ and $Q=I_{sk}\oplus P$. Then $QAQ^T=\Pi_{s+1,k}(\beta)$. Therefore, $A$ is permutation similar to $\Pi_{s+1,k}(\beta)$ and $D$ is a $(u,k,v)$-completely transitive tournament.
[*Subcase 2.*]{} $s=1$. Then $t\ge 5$ and $$A=(a_{ij})=\begin{bmatrix}
T_{k}&w_{1}&w_{2}&\cdots&w_{t}&x\\
u_{1}&0&1&\cdots&1&a_{k+1,n}\\
u_{2}&0&0&\ddots&\vdots&\vdots\\
\vdots&0&0&\ddots&1&\vdots\\
u_{t}&0&0&0&0&a_{k+t,n}\\
y^{T}&a_{n,k+1}&a_{n,k+2}&\cdots&a_{n,k+t}&0
\end{bmatrix}\in M_{n}\{0,1\},$$ where $x,y\in \mathbb{R}^{n-5}$, $w_{i}=(J_{1,k_i},0)^T$, $u_{i}=(0,J_{1,k-1-k_i})$, and $0\leq k_1<k_2<\cdots<k_t\leq k-1$.
Choose any three distinct numbers $p,q,r\in \{1,2,\ldots,t\}$. Denote $\alpha=\{1,2,\ldots,k,k+p, k+q,k+r,n\}$ and $B=A[\alpha]$. Then $$f(B)=f(T_k)+3(k-1)+3+f(x)+f(y)+3=(k+4)(k+3)/2-4=ex(k+4,\F).$$ Applying Corollary \[co1\] to $B$, we have $$x=(J_{1,h},0)^T, ~~y=(0,J_{1,k-h-l})^T,$$ where $0\le h\le k-1$. Moreover, we have $h\not\in\{k_p,k_q,k_r\}$ and $$\label{eq50}
~~a_{k+i,n}=\left\{\begin{array}{ll}
1,& \textrm{ if } h>k_i, \\
0,& \textrm{ if } h<k_i,
\end{array}
\right.$$ $i=p,q,r$.
Since $p,q,r$ are arbitrarily chosen from $\{1,2,\ldots,t\}$, we have $$h\not\in\{k_1,k_2,\ldots,k_t\}$$ and (\[eq50\]) holds $i=1,2,\ldots,t.$ Using the same arguments as in the previous subcase, we can prove that $D$ is a $(u,k,v)$-completely transitive tournament.
This completes the proof. $\Box$
Further discussion
==================
In this section we discuss the unsolved cases for Problem \[pro1\]. We focus our attention on strict digraphs. The second part of Theorem \[thh2\] may not be true for $n=k+5$ when $k=4$. For example, let $D$ be the digraph with adjacency matrix $$A=\begin{bmatrix}
0&0&1&1&1&1&1&1&1\\
0&0&1&1&1&1&1&1&1\\
0&0&0&1&1&1&1&0&1\\
0&0&0&0&1&1&1&0&0\\
0&0&0&0&0&1&1&0&0\\
0&0&0&0&0&0&0&0&0\\
0&0&0&0&0&0&0&0&0\\
0&0&0&0&1&1&1&0&1\\
0&0&0&0&0&1&1&0&0
\end{bmatrix}.$$ Then $f(A)=30$ and $A^4\in M_{9}\{0,1\}$. But $A$ is not permutation similar to any principal submatrix of $\Pi_{4,3}$, since $A^4\ne 0$. However, when $k=4$ and $n$ is sufficiently large, we conjecture the second part of Theorem \[thh2\] is still true.
When $k=3$, Theorem \[thh2\] does not hold, which is shown by the following example. Let $$T'_3=\begin{bmatrix}
0&1&1\\
0&1&1\\
0&0&0
\end{bmatrix}$$ and $$A=\begin{bmatrix}
T_{3}&T'_{3}&T_{3}&\cdots&T_{3}\\
T_{3}&T_{3}&T_{3}&\cdots&T_{3}\\
T_{3}&T_{3}&T_{3}&\cdots&T_{3}\\
\vdots&\vdots&\vdots&\ddots&\vdots\\
T_{3}&T_{3}&T_{3}&\cdots&T_{3}
\end{bmatrix}\in M_{3t}\{0,1\}.$$ Denote $(b_{ij})=A^3$. Then $$b_{ij}=
\begin{cases}
1,& \textrm{if~~} i=3s+1,j=3l \textrm{~~for~~} s\in\{0,\ldots,t-1\}, l\in\{1,\ldots,t\};\\
0,& \textrm{otherwise}.
\end{cases}$$ Therefore, $D(A)$ is an $\mathscr{F}_3$-free digraph and $$ex(3t,\mathscr{F}_3)\ge f(A)>f(\Pi_{t,3}).$$
When $k=2$, by [@WU Theorem 2] we can deduce $$ex(n,\mathscr{F}_2)=\begin{cases}\frac{n^2+4n-5}{4}, & \text{if $n$ is odd,}\\
\frac{n^2+4n-8}{4},& \text{if $n$ is even and $n\not=4,$}\\
7, & \text{if $n=4$}.\end{cases}$$ The set of extremal digraphs $Ex(n,\mathscr{F}_2)$ is not known.
We leave the problems of determining $ex(n,\mathscr{F}_3)$ and $Ex(n,\mathscr{F}_k)$ for $k=2,3,4$ for future work.
Acknowledgement {#acknowledgement .unnumbered}
===============
The first author is grateful to Professor Yuejian Peng for helpful discussions on extremal graph theory. The research of Huang was supported by the NSFC grant 11401197 and a Fundamental Research Fund for the Central Universities.
[WWW]{}
B. Bollobás, Extremal graph theory, Handbook of combinatorics, Vol. 2, 1231-1292, Elsevier, Amsterdam, 1995. J.A. Bondy, U.S.R. Murty, Graph Theory, in: GTM, vol. 244, Springer, 2008. W.G. Brown, P. Erdős, M. Simonovits, Extremal problems for directed graphs, J. Combinatorial Theory Ser. B 15 (1973) 77-93. W.G. Brown, P. Erdős, M. Simonovits, Algorithmic solution of extremal digraph problems, Trans. Amer. Math. Soc. 292 (1985) 421-449. W.G. Brown and F. Harary, Extremal digraphs, Combinatorial Theory and its Applications, Colloq. Math. Soc. Janos Bolyai 4 (1970) I, 135-198. W. G. Brown, M. Simonovits, Extremal multigraph and digraph problems, Paul Erdős and his mathematics, II (Budapest, 1999), 157-203, Bolyai Soc. Math. Stud., 11, J¨¢nos Bolyai Math. Soc., Budapest, 2002.
F. A. Firke, P. M. Kosek, E. D. Nash, J. Williford, Extremal graphs without 4-cycles, J. Combin. Theory Ser. B 103 (2013) 327-336.
Z. Huang, X. Zhan, Digraphs that have at most one walk of a given length with the same endpoints, Discrete Math. 311 (2011) 70-79.
H. Jacob, H. Meyniel, Extension of ’s and Brooks’ theorems and new notions of stability and coloring in digraphs, Annals of Discrete Mathematics, 17 (1983) 365-370. V. Nikiforov, Some new results in extremal graph theory, Surveys in combinatorics 2011, 141-181, London Math. Soc. Lecture Note Ser. 392, Cambridge Univ. Press, Cambridge, 2011. M. Tait, C. Timmons, Sidon sets and graphs without 4-cycles, J. Comb. 5 (2014) 155-165.
P. , Egy grafelmeletsi zelsbertekfe ladatrbl, Mat. Fiz. Lapok 48 (1941),4 36-452. P. , On the theory of graphs, Colloq. Math. 3 (1954) 19-30.
H. Wu, On the 0-1 matrices whose squares are 0-1 matrices, Linear Algebra Appl. 432 (2010) 2909-2924.
X. Zhan, Matrix theory, Graduate Studies in Mathematics 147, American Mathematical Society, Providence, RI, 2013.
[^1]: Institute of Mathmatics, Hunan University, Changsha 410082, P.R. China. ([email protected])
[^2]: College of Mathematics and Econometrics, Hunan University, Changsha 410082, P.R. China. ([email protected])
[^3]: Department of Mathematics, East China Normal University, Shanghai 200241, China. ([email protected])
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We argue that the inference of CP violation in experiments involving the $K^0 - \bar{K^0}$ system in weak interactions of particle physics is facilitated by the assumption of particle trajectories for the decaying particles and the decay products. A consistent explanation in terms of such trajectories is naturally incorporated within the Bohmian interpretation of quantum mechanics.'
---
=17 truecm =20 truecm =-1 truecm =-2 truecm
0.2in
0.5in
Dipankar Home[^1]
0.2in
Bose Institute, Calcutta 700009, India
0.2in
A.S.Majumdar[^2]
0.2in
S.N.Bose National Centre for Basic Sciences
0.1in
Block-JD, Sector III, Salt Lake, Calcutta 700091, India.
2.0in
I. Introduction
The Bohm model is able to provide a causal interpretation of quantum mechanics in a consistent manner \[1\]. At the same time, the predictions of Bohmian mechanics are in exact agreement with the standard quantum mechanical predictions for observable probabilities in all usual experimental situations. In this paper we shall be concerned with examining the possible importance of the Bohmian approach in interpreting certain experiments whose understanding in terms of the standard interpretation is rather ambiguous.
For the purpose of [*reinterpreting*]{} the standard quantum formalism using the Bohmian scheme, a wave function $\psi$ is [*not*]{} taken to provide a complete specification of the state of an individual system; an [*additional*]{} ontological “position” coordinate (an objectively real “position” existing irrespective of any external observation) is ascribed to an individual particle. The “position” coordinate of the particle evolves with time obeying an equation which can be derived from the Schrodinger equation (considering the one dimensional case) $$\begin{aligned}
i\hbar {\partial\psi \over \partial t} = H\psi \equiv - {\hbar^2
\over 2m} {\partial^2 \psi \over \partial x^2} + V(x)\psi\end{aligned}$$ by writing $$\begin{aligned}
\psi = Re^{iS/\hbar}\end{aligned}$$ and using the continuity equation $$\begin{aligned}
{\partial \over \partial x} (\rho v) + {\partial\rho \over \partial
t} = 0\end{aligned}$$ for the probability distribution $\rho(x,t)$ given by $$\begin{aligned}
\rho = \vert \psi \vert^2.\end{aligned}$$ It is important to note that $\rho$ is ascribed an [*ontological*]{} significance by regarding it as representing the probability density of “particles” occupying [*actual*]{} positions. In contrast, in the standard formulation $\rho$ is interpreted as the probability density of [*finding*]{} a particle around a certain position. Setting ($\rho v$) equal to the quantum probability current leads naturally to the Bohmian interpretation whrere the particle velocity $v(x,t)$ is given by $$\begin{aligned}
v \equiv {dx \over dt} = {1\over m}{\partial S \over \partial x}\end{aligned}$$ The particle “trajectory” is completely deterministic and is obtained by integrating (5) with the appropriate initial conditions. The essential significance of Bohm’s model lies in providing an elegant solution to the measurement problem (which has been described by Weinberg \[2\] as “the most important puzzle in the interpretation of quantum mechanics”) without requiring wave function collapse, since according to the Bohmian interpretation, in any measurement a definite outcome is singled out by the relevant ontological position coordinate.
In view of the importance of the Bohm model in providing not only an internally consistent [*alternative*]{} interpretation of the standard quantum formalism, but also perhaps the neatest solution to the measurement problem \[1\], it should be worthwhile to look for specific situations where the [*conceptual superiority*]{} of Bohm’s model over the standard interpretation may become easily transparent. To this end, we now proceed to examine the analysis of a fundamentally important experiment of particle physics, namely, the discovery of CP-violation \[3\].
II\. The CP-violation experiment
C(charge conjugation) and P(parity) are two of the fundamental discrete symmetries of nature, the violations of which have not been empirically detected in phenomena other than weak interactions. If a third discrete symmetry T(time reversal) is taken into account, there exists a fundamental theorem of quantum field theory, viz., the CPT theorem which states that all physical processes are invariant under the combined operation of CPT. Nevertheless, there is no theorem forbidding the violation of CP symmetry, and indeed, there have been several experiments to date \[4\], starting from the pioneering observation of Christenson, Cronin, Fitch and Turlay \[3\], that have revealed the occurrence of CP violation through weak interactions of particle physics involving the particles $K^0$ and $\bar{K^0}$. The eigenstates of strangeness $K^0$ 0.1in $(s=+1)$ and its CP conjugate $\bar{K^0}$ 0.1in $(s=-1)$ are produced in strong interactions, for example, the decay of $\Phi$ particles. Weak interactions do not conserve strangeness, whereby $K^0$ and $\bar{K^0}$ can mix through intermediate states like $2\pi, 3\pi, \pi\mu\nu, \pi e\nu$, etc. The observable particles, which are the long lived $K$-meson $K_L$, and the short lived one $K_S$, are linear superpositions of $K^0$ and $\bar{K^0}$, i.e., $$\begin{aligned}
\vert K_L \rangle & = & (p\vert K^0\rangle - q\vert
\bar{K^0}\rangle )/ \sqrt{\vert p\vert^2 + \vert q\vert^2} \\
\vert K_S \rangle & = & (p\vert K^0\rangle + q\vert
\bar{K^0}\rangle )/ \sqrt{\vert p\vert^2 + \vert q\vert^2}\end{aligned}$$ which obey the exponential decay law $\vert K_L\rangle \rightarrow
\vert K_L\rangle exp(-\Gamma_L t/2)exp(-im_Lt)$ and analogously for $\vert K_S\rangle$, where $\Gamma_L$ and $m_L$ are the decay width and mass respectively of the $K_L$ particle. It follows from (6) and (7) that $$\begin{aligned}
\langle K_L\vert K_S\rangle = {|p|^2-|q|^2 \over |p|^2+|q|^2}\end{aligned}$$
CP violation takes place if the states $\vert K_L\rangle$ and $\vert K_S\rangle$ are not orthogonal. Through weak interactions the $K_S$ particle decays rapidly into channels such as $K_S
\rightarrow \pi^{+}\pi^{-}$ and $K_S \rightarrow 2\pi^{0}$ with a mean lifetime of $10^{-10}s$, whereas, the predominant decay modes of $K_L$ are $K_L \rightarrow \pi^{\pm}e^{\pm}\nu$ (with branching ratio $\sim 39\%$), $K_L \rightarrow \pi^{\pm}\mu^{\pm}\nu (\sim
27\%)$, and $K_L \rightarrow 3\pi ( \sim 33\%)$ \[4\]. The CP violating decay mode $K_L \rightarrow 2\pi$ is extremely rare (with branching ratio $\sim 10^{-3}$) in the background of the other large decay modes. Considering the Schrodinger evolution, if the analysis of the term corresponding to $K_s$ in the relevant initial wave function shows that it cannot contribute significantly to the emission of two pions with suitable momenta and locations, then one can infer the occurrence of CP violation in this particular situation. In other words such $2\pi$ can only arise through the $K_L$ decay mode. The momenta and locations of the emitted pions are important since the key experimental issue is to detect the $2\pi$ particles coming from the decay of $K_L$ and identify them as coming from $K_L$ and [*not*]{} $K_S$.
In a typical experiment to detect CP violation, an initial state of the type $$\begin{aligned}
\vert\psi_i\rangle = (a\vert K_L\rangle + b\vert K_S\rangle )\end{aligned}$$ is used which is a coherent superposition of the $K_L$ and $K_S$ states. Such a state has been produced by the technique of ‘regeneration’ \[5\] which has been used in a large number of experiments \[6\]. The common feature of all these experiments is the measurement of the vector momenta $\stackrel{\rightarrow}{p_i}$ of the charged decay products $\pi^{+}\pi^{-}$ or $2\pi^0$ from the decaying pions. It is only the [*type*]{} of instrument used for actually measuring the momenta that varies from experiment to experiment.
III\. Bohmian trajectories
To see how the Bohmian interpretation helps in drawing the relevant inference from this experiment, we concentrate on the analysis of a single event in which the two emitted pions from a decaying kaon are detected by two detectors respectively along two different directions. From the measured momenta $\stackrel{\rightarrow}{p_1}$ and $\stackrel{\rightarrow}{p_2}$, the “trajectories” followed by the individual pions are [*retrodictively*]{} inferred [*assuming*]{} that they have followed [*linear “trajectories”*]{}. The point of intersection of these retrodicted “trajectories” is inferred to be the point from which the decay products have emanated from the decaying system; in other words, what is technically known as the “decay vertex” is determined in this way. The value of the momentum of the decaying kaon is obtained by $\stackrel{\rightarrow}{p_k} =
\stackrel{\rightarrow}{p_1} + \stackrel{\rightarrow}{p_2}$. Once the decay vertex and the kaon momentum is known, one estimates the time taken by the kaon to reach the decay vertex from the source, [*again*]{} using at this stage the idea of a [*linear “trajectory”*]{}. If this time turns out to be much larger than the $K_S$ mean lifetime ($\sim 10^{-10}s$), one infers that the detected $2\pi$ pair must have come from $K_L$, which, as already mentioned, is the signature of CP violation.
It is thus evident from the above discussion that the assumption of a [*linear “trajectory”*]{} of a freely evolving particle (kaon or pion) provides a consistent explanation in support of CP violation in such an experiment. Within the standard interpretationm of quantum mechanics, there is no way one can justify assigning a “trajectory” to a freely evolving particle. Moreover, assuming such a “trajectory” to be [*linear*]{} is an additional [*ad hoc*]{} input. One possible argument could be to assign localized wave packets to emitted pions and kaons, and to use the fact that their peaks follow classical trajectories in the case of a free evolution. However, in the standard quantum mechanical description of decay processes, the decay products are regarded as asymptotically free, and hence should be represented by plane wave states. Moreover, even if they are approximated in some sense by localized wave packets, there would be inevitable spreading of the wave packets. Even if this spreading is regarded as negligible within the time interval concerned, a ‘literal identification’ of the wave packet with the particle is conceptually impermissible without an [*additional*]{} input at the fundamental level in the form of the notion of a “particle” with a definite position even when unobserved (“particle” ontology).
On the other hand, the assumption of [*linear “trajectories”*]{} followed by the decaying particles and the decay products is amenable to a natural explanation within the Bohmian framework. The decaying kaons as well as the asymptotically free decay products are represented by plane waves $$\begin{aligned}
\psi \sim e^{ikx}.\end{aligned}$$ Hence it follows that in the Bohmian scheme the velocity equation (5) is in this case given by $$\begin{aligned}
{dx \over dt} = {\hbar k \over m}\end{aligned}$$ which when integrated provides the linear “trajectories” of the particles. These trajectories are ontological and deterministic. Therefore, in this interpretation, the exact position coordinates of the “decay vertex” can be assigned in a natural way by retrodicting the pion “trajectories” without any inconsistencies of the type inherent in the standard interpretation. Hence, it seems necessary that the standard formalism of quantum mechanics needs to be supplemented with the Bohmian interpretation of ontological particle “trajectory” (in the sense that the particle has traversed a well defined path even when unobserved) to enable for the consistent inference of the observation of CP violation in the actual experiments involving kaon decays.
IV\. Concluding remarks
The main reasons for choosing, in particular, the CP violation experiment for this purpose are the following. First, [*unlike*]{} other common high energy experiments this particular experiment involves [*not*]{} merely the measurement of some physical quantities but [*inferring*]{} from the measured quantities the violation of a fundamental symmetry property of the pertinent physical interactions. Secondly, again unlike other common high energy experiments, the effects of particle creation and annihilation are not relevant for the important part of the experiment involved with the prediction of CP violation, and no second quantized treatment is required for the theoretical framework. The crucial phenomena of particle decays which this experiment is concerned with, is appropriately described in terms of the Schrodinger equation (see \[4\] and references therein) for which there exists a consistent Bohmian interpretation. Note that ignoring interpretational nuances, if one tries to follow a very pragmatic approach and [*approximates*]{} the plane wave states of the decay products by wave packets whose peaks follow classical trajectories with finite speeds, careful estimates need to be done to [*quantify*]{} the resulting errors or fluctuations due to spreading of wave packets by taking into account the [*actual*]{} distances involved in the performed experiments. (Of course, the estimates of these distances related to the particle trajectories are fundamental from the Bohmian perspective.) This is important because the CP violation effect is exceedingly [*small*]{}; the branching ratio of the CP violating decay mode $K_L
\rightarrow 2\pi$ is $10^{-3}$. In [*none*]{} of the CP violation experiments performed to date has this point been considered in the relevant analysis.
We conclude by noting that this analysis suggests that it should be worthwhile to look for more such appropriate examples where the inadequacy or ambiguity of the standard formalism in comprehending the results of the concerned experiments can be avoided by using the Bohmian interpretation. It should be appreciated that since there is no measurement problem in the Bohmian interpretation, a Bohmian analysis is useful for all experiments in quantum mechanics, and in particular scattering experiments where it is required to know why particles are detected where they are at the end of the experiment. The answer to this is clear from the Bohmian perspective—the particles are detected where they [*actually are*]{}. However, from the viewpoint of the standard interpretation the explanation is rather obscure, as long as the Schrodinger wave function is regarded as the complete description of the physical system. In this context it has been recently argued \[7\] that the concept of quantum probability current, a full understanding of which is provided by Bohmian mechanics, is fundamental for a genuine understanding of scattering phenomena. Apart from this, it has been claimed \[8\] that a special significance of Bohmian mechanics lies in experiments related to the measurement of time of flight of particles, and tunelling time in particular for which it is difficult to find a consistent or unambiguous definition within the standard framework of quantum mechanics. All this is of course [*different*]{} from empirically verifying a new consequence, if any exists, of the Bohmian interpretation which is not obtainable from the standard interpretation. Nevertheless, such investigations like the one reported in this paper could be helpful in understanding more clearly the relative merits of the standard and Bohmian interpretations.
0.5in
This work was supported by the Department of Science and Technology, India.
[**REFERENCES**]{}
0.5in
\[1
: \] P.R.Holland, “The Quantum Theory of Motion”, (Cambridge University Press, London, 1993); D.Bohm, Phys. Rev. [**85**]{} (1952) 166; D.Bohm and B.J.Hiley, “The Undivided Universe”, (Routledge, London, 1993); J.T.Cushing, “Quantum Mechanics - Historical Contingency and the Copenhagen Hegemony”, (University of Chicago Press, Chicago, 1994).
\[2
: \] S.Weinberg, “Dreams of a final theory”, (Vintage, London, 1993) p. 64.
\[3
: \] J.H.Christenson, J.W.Cronin, V.L.Fitch and R.Turlay, Phys. Rev. Lett. [**13**]{} (1964) 138.
\[4
: \] For a review, see for instance, K.Kleinknecht, in “CP violation”, edited by C.Jarlskog, (World Scientific, Singapore, 1989) pp. 41 -104.
\[5
: \] A.Pais and O.Piccioni, Phys. Rev. [**100**]{} (1955) 1487.
\[6
: \] For eample, see, C.Geweniger et al., Phys. Lett. B [**48**]{} (1974) 487; V.Chaloupka et al., Phys. Lett. B [**50**]{} (1974) 1; W.C.Carithers et al., Phys. Rev. Lett. [**34**]{} (1975) 1244; N.Grossmann, et al., Phys. Rev. Lett., [**59**]{} (1987) 18.
\[7
: \] M.Daumer, D.Duerr, S.Goldstein and N.Zanghi, J. Stat. Phys. [**88**]{} (1997) 967.
\[8
: \] C.R.Leavens, Phys Lett. A[**197**]{} (1995) 88; in “Bohmian Mechanics and Quantum Theory: An Appraisal”, eds. J.T.Cushing, A.Fine and S.Goldstein (Kluwer, Dordrecht, 1996).
[^1]: e-mail:[email protected]
[^2]: e-mail:[email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present an overview of some of our recent results on the existence of rays of minimal growth for elliptic cone operators and two new results concerning the necessity of certain conditions for the existence of such rays.'
address:
- |
3000 Ivyside Park\
Penn State Altoona\
Altoona, PA 16601\
U.S.A.
- |
Institut für Mathematik\
Universität Potsdam\
D-14415 Potsdam\
Germany
- |
Department of Mathematics\
Temple University\
Philadelphia, PA 19122\
U.S.A.
author:
- 'Juan B. Gil'
- Thomas Krainer
- 'Gerardo A. Mendoza'
title: |
On rays of minimal growth for\
elliptic cone operators
---
Introduction
============
The aim of this article is twofold. On the one hand, we present an overview of some of the results contained in [@GKM1; @GKM2] on the subject in the title, and of the geometric perspective we developed in the course of the investigations leading to the aforementioned papers. We illustrate the main ideas of our approach by means of examples concerning Laplacians on a compact $2$-manifold. Already this simple situation exhibits the structural richness and complexity of the general theory.
On the other hand, we offer some improvements, cf. Theorems \[GeometricNecAndSuff\] and \[NecessityConstCoeff\], regarding necessary and sufficient conditions for a closed sector $\Lambda\subset \C$ to be a sector of minimal growth for a certain class of elliptic cone operators $A$ and for the associated model operator $A_{\wedge}$.
Recall that a closed sector of the form $$\label{Sector}
\Lambda = \{\lambda \in {\C}: \lambda = re^{i\theta} \; \text{for} \;
r \geq 0, \; \theta \in {\R}, \; |\theta - \theta_0| \leq a\}$$ is called a *sector of minimal growth* (or of maximal decay) for a closed operator $$A:\Dom\subset H\to H,$$ where $H$ is a Hilbert space and $\Dom$ is dense in $H$, if there is a constant $R>0$ such that $A-\lambda$ is invertible for every $\lambda\in \Lambda_R=\{\lambda\in\Lambda: |\lambda|\ge R\}$, and the resolvent $(A-\lambda)^{-1}$ satisfies either of the equivalent estimates $$\label{NormEstimates}
\bigl\|(A-\lambda)^{-1}\bigr\|_{\L(H)} \le C/|\lambda|, \quad
\bigl\|(A-\lambda)^{-1}\bigr\|_{\L(H,\Dom)} \le C$$ for some $C>0$ and all $\lambda\in\Lambda_R$.
We are interested in cone operators on smooth manifolds with boundary. Specifically, let $M$ be a smooth $n$-manifold with boundary $Y=\partial
M$ and let $E\to M$ be a Hermitian vector bundle over $M$. Fix a defining function $x$ for $Y$. A differential *cone operator* of order $m$ acting on sections of $E\to M$ is an operator of the form $A= x^{-m}P$ with $P$ in the class $\operatorname{Diff}_b^m(M;E)$ of totally characteristic differential operators of order $m$, cf. Melrose [@RBM1]. We write $A\in x^{-m}\operatorname{Diff}_b^m(M;E)$. More explicitly, in the interior $\open M$ of $M$, $A$ is a differential operator with smooth coefficients, and near the boundary, in local coordinates $(x,y)\in (0,\eps)\times Y$, it is of the form $$\label{cone-operator}
A=x^{-m}\sum_{k+|\alpha|\le m} a_{k\alpha}(x,y)(xD_x)^k D_y^\alpha$$ with coefficients $a_{k\alpha}$ smooth up to $x=0$; here $D_x=-i\partial/\partial x$ and likewise $D_{y_j}$. We will say that $A$ (or $P$) has *coefficients independent of $x$ near $Y$*, if the coefficients $a_{k\alpha}$ in do not depend on $x$ (this notion depends on the choice of tubular neighborhood map, defining function $x$, and connection on $E$. For a precise definition see [@GKM1]).
This paper consists of 5 sections. In Section 2 we review some basic properties of cone operators while in Section 3 we discuss the associated model operators. The new results on rays of minimal growth can be found in Sections 4 and 5. Apart from the works explicitly cited in the text, our list of references contains additional items referring to related works on resolvents and rays of minimal growth for elliptic operators.
Preliminaries on cone operators {#sec-Preliminaries}
===============================
Let $A$ be a differential cone operator. As introduced in [@GKM1], the principal symbol of $A$, $\csym(A)$, is defined on the $c$-cotangent ${}^cT^*M$ of $M$ rather than on the cotangent itself. Over $\open M$ it is essentially the usual principal symbol, and equal to $$\sum_{k+|\alpha|= m} a_{k\alpha}(x,y)\xi^k \eta^\alpha$$ near the boundary $Y$, see .
Let $M$ be a compact $2$-manifold with boundary $Y=S^1$. Let $g_{Y}(x)$ be a smooth family of Riemannian metrics on $S^1$ such that $g_{Y}(0)$ is the standard metric, $dy^2$. We equip $M$ with a “cone metric” $g$ that near $Y$ takes the form $g=dx^2 + x^2 g_{Y}(x)$ ($g$ is a regular Riemannian metric in the interior of $M$). Then, near $Y$, the Laplace-Beltrami operator $\Delta$ has the form $$\label{Laplacian}
x^{-2}\big((xD_x)^2 + a(x,y)(xD_x) + \Delta_{Y}(x)\big),$$ where $a(x,y)$ is a smooth function with $a(0,y)=0$ and $\Delta_Y(x)$ is the nonnegative Laplacian on $S^1$ associated with $g_{Y}(x)$. In this case, near the boundary, we have $$\csym(\Delta) = \xi^2 + \operatorname{ \sigma\!\!\!\sigma}(\Delta_Y(x)).$$
Ellipticity and boundary spectrum {#ellipticity-and-boundary-spectrum .unnumbered}
---------------------------------
An operator $A\in x^{-m}\operatorname{Diff}_b^m(M;E)$ is *$c$-elliptic* if $\csym(A)$ is invertible on ${}^cT^*M\backslash 0$. Moreover, the family $A-\lambda$ is said to be *$c$-elliptic with parameter* $\lambda\in\Lambda\subset\C$ if $\csym(A)-\lambda$ is invertible on $({}^cT^*M\times\Lambda) \backslash 0$.
Associated with $A=x^{-m}P$ there is an operator-valued polynomial $$\C\ni \sigma\mapsto \hat P(\sigma)\in \operatorname{Diff}^m(Y;E|_Y)$$ called the *conormal symbol* of $P$ (and of $A$). If we write $A$ as in , then $$\hat P(\sigma) = \sum_{k+|\alpha|\le m}
a_{k\alpha}(0,y)\sigma^k D_y^\alpha.$$ If $A$ is $c$-elliptic, then $\hat P(\sigma)$ is invertible for all $\sigma\in\C$ except a discrete set $\operatorname{spec}_b(A)$, the *boundary spectrum of $A$*, cf. [@RBM1]; $\hat P(\sigma)$ is a holomorphic family of elliptic operators on $Y$ and $\sigma\to \hat P(\sigma)^{-1}$ is a meromorphic operator-valued function on $\C$.
The Laplacian is clearly $c$-elliptic. If $y$ is the angular variable on $S^1$, then $$\hat P(\sigma) = \sigma^2 + \Delta_Y(0) =\sigma^2+D_y^2,$$ and the boundary spectrum of $\Delta$ is given by $$\operatorname{spec}_b(\Delta) = \{\pm ik: k\in\mathbb{N}_0\}.$$
Closed extensions {#closed-extensions .unnumbered}
-----------------
Let $\mathfrak{m}$ be a positive $b$-density on $M$, that is, $x\mathfrak{m}$ is a smooth everywhere positive density on $M$. Let $L^2_b(M;E)$ be the $L^2$ space of sections of $E$ with respect to the Hermitian form on $E$ and the density $\mathfrak{m}$. Consider $A$ initially defined on $C_0^\infty(\open M;E)$ and look at it as an unbounded operator on the Hilbert space $$x^{-m/2}L^2_b(M;E) = L^2(M;E;x^{2m} \mathfrak m).$$ The particular weight $x^{-m/2}$ is just a convenient normalization and represents no loss. If we are interested in $A$ on $x^{\mu}L^2_b(M;E)$ for $\mu\in\R$, we can base all our analysis on the space $x^{-m/2}L^2_b(M;E)$ by considering the operator $x^{-\mu-m/2} A\,x^{\mu+m/2}$.
Typically, $A$ has a large class of closed extensions $$\label{Ext}
A_{\Dom}:\Dom\subset x^{-m/2}L^2_b(M;E)\to x^{-m/2}L^2_b(M;E).$$ There are two canonical closed extensions, namely the ones with domains $$\begin{aligned}
\Dom_{\min}(A) &= \text{closure of } C^\infty_0(\open M;E)
\text{ with respect to } \|\cdot\|_A, \\
\Dom_{\max}(A) &=\{u\in x^{-m/2}L^2_b(M;E): Au\in x^{-m/2}L^2_b(M;E)\},\end{aligned}$$ where $\|u\|_A=\|u\|+\|Au\|$ is the graph norm in $\Dom_{\max}(A)$. Both domains are dense in $x^{-m/2}L^2_b(M;E)$, and for any closed extension , $$\Dom_{\min}(A)\subseteq \Dom\subseteq \Dom_{\max}(A).$$ Let $$\mathfrak D(A)=\{\Dom\subset\Dom_{\max}(A): \Dom \text{ is a vector space
and } \Dom_{\min}(A)\subset\Dom\}.$$ The elements of $\mathfrak D(A)$ are in one-to-one correspondence with the subspaces of $\Dom_{\max}(A)/\Dom_{\min}(A)$. If the operator $A$ is fixed and there is no possible ambiguity, we will omit $A$ from the notation and will write simply $\Dom_{\min}$, $\Dom_{\max}$, and $\mathfrak D$.
\[LeschTheorem\] If $A\in x^{-m}\operatorname{Diff}_b^m(M;E)$ is $c$-elliptic, then $$\dim \Dom_{\max}/\Dom_{\min}<\infty$$ and all closed extensions of $A$ are Fredholm. Moreover, $$\label{RelIndexA}
\operatorname{ind}A_{\Dom}= \operatorname{ind}A_{\Dom_{\min}}+ \dim\Dom/\Dom_{\min}.$$
Modulo $\Dom_{\min}$, the elements of $\Dom_{\max}$ are determined by their asymptotic behavior near the boundary of $M$. The structure of these asymptotics depends on the conormal symbols of $A$ and on the part of $\operatorname{spec}_b(A)$ in the strip $\{|\Im\sigma|<m/2\}$. More details will be discussed in the next section.
If $A$ is $c$-elliptic and symmetric (formally selfadjoint), then $$\operatorname{ind}A_{\Dom_{\max}} = -\operatorname{ind}A_{\Dom_{\min}} \;\;\text{and}\;\;
\operatorname{ind}A_{\Dom_{\min}} = -\frac12 \dim\Dom_{\max}/\Dom_{\min}.$$
Consider the cone Laplacian $\Delta$, cf. . Then and imply $$\dim\Dom_{\max}(\Delta)/\Dom_{\min}(\Delta)=2$$ and thus, by the previous corollary, $$\label{IndexLaplacian}
\operatorname{ind}\Delta_{\min}=-1 \;\;\text{ and }\;\; \operatorname{ind}\Delta_{\max}=1.$$
If $A\in x^{-m}\operatorname{Diff}_b^m(M;E)$ is $c$-elliptic, the embedding $\Dom_{\max}\hookrightarrow x^{-m/2}L^2_b(M;E)$ is compact. Therefore, for every $\Dom\in\mathfrak D$ and $\lambda\in\C$, the operator $A_{\Dom}-\lambda$ is also Fredholm with $\operatorname{ind}(A_{\Dom}-\lambda)=
\operatorname{ind}A_{\Dom}$. Consequently, if $\operatorname{spec}(A_\Dom)\not=\C$, then we necessarily have $\operatorname{ind}A_\Dom=0$. For this reason, we will primarily be interested in the set of domains $$\label{GrasmannianA}
\mathfrak G=\{\Dom\in \mathfrak D: \operatorname{ind}A_\Dom = 0\}$$ which is empty unless $\operatorname{ind}A_{\Dom_{\min}}\le 0$ and $\operatorname{ind}A_{\Dom_{\max}}\ge 0$. Let $d''=-\operatorname{ind}A_{\Dom_{\min}}$.
Using that the map $\mathfrak D\ni \Dom\mapsto \Dom/\Dom_{\min}$ is a bijection, we identify $\mathfrak G$ with the complex Grassmannian of $d''$-dimensional subspaces of $\Dom_{\max}/\Dom_{\min}$.
For $\Delta$ we have $$\mathfrak G(\Delta) \cong \mathbb{CP}^1 = S^2.$$ Note that by and , $\operatorname{ind}\Delta_{\Dom}=0$ if and only if $\dim\Dom/\Dom_{\min}=1$.
We finish this section with the following proposition that gives a first glimpse of the complexity of the spectrum of elliptic cone operators.
If $A$ is $c$-elliptic and $\dim \mathfrak G>0$, then for any $\lambda\in\C$ there is a domain $\Dom\in\mathfrak G$ such that $\lambda\in\operatorname{spec}(A_\Dom)$. If, in addition, $A$ is symmetric on $\Dom_{\min}$, then for any $\lambda\in\R$ there is a $\Dom\in\mathfrak G$ such that $A_\Dom$ is selfadjoint and $\lambda\in\operatorname{spec}(A_\Dom)$.
A proof is given in [@GKM1 Propositions 5.7 and 6.7]. A surprising consequence of the second statement is that for any arbitrary negative number $\lambda$ there is always a selfadjoint extension of $A$ having $\lambda$ as eigenvalue, even if $A$ is positive on $\Dom_{\min}$.
The model operator {#sec-ModelOperator}
==================
Let $A\in x^{-m}\operatorname{Diff}_b^m(M;E)$ be $c$-elliptic. The *model operator $A_\wedge$* associated with $A$ is an operator on $N_+Y$, the closed inward normal bundle of $Y$, that in local coordinates takes the form $$A_\wedge=x^{-m}\sum_{k+|\alpha|\le m}
a_{k\alpha}(0,y)(xD_x)^k D_y^\alpha,$$ if $A$ is written as in . A Taylor expansion in $x$ (at $x=0$) of the coefficients of the operator $A$ induces a decomposition $$\label{bOperatorTaylor}
x^m A = \sum_{k=0}^{N-1} P_k x^k + x^N \tilde P_N
\quad\text{for every } N\in\mathbb{N},$$ where each $P_k$ has coefficients independent of $x$ near $Y$. Thus the model operator can be written, near $Y$, as $A_\wedge =x^{-m}P_0$. In other words, $A_\wedge$ can be thought of as the “most singular” part of $A$.
We trivialize $N_+Y$ as $Y^\wedge=[0,\infty)\times Y$. The operator $A_\wedge\in x^{-m}\operatorname{Diff}_b^m(Y^\wedge;E)$ acts on $C^\infty_0(\open Y^\wedge;E)$ and can be extended as a densely defined closed operator in $x^{-m/2}L^2_b(Y^\wedge;E)$. The space $L^2_b(Y^\wedge;E)$ is the $L^2$ space with respect to a density of the form $\frac{dx}{x}\otimes \pi^*\mathfrak{m}_Y$ and the canonically induced Hermitian form on $\pi^*(E|_Y)$, where $\pi : Y^{\wedge} \to Y$ is the projection on the factor $Y$. The density $\mathfrak{m}_Y$ is related to $\mathfrak{m}$ and, by abuse of notation, we denote $\pi^*(E|_Y)$ by $E$, cf. [@GKM1]. Again, there are two canonical domains $\Dom_{\wedge,\min}$ and $\Dom_{\wedge,\max}$ and we denote by $\mathfrak D_\wedge$ the set of subspaces of $\Dom_{\wedge,\max}$ that contain $\Dom_{\wedge,\min}$. There is a natural (and useful) linear isomorphism $$\theta: \Dom_{\max}/\Dom_{\min} \to \Dom_{\wedge,\max}/\Dom_{\wedge,\min},$$ cf. Section \[sec-Necessity\]. As a consequence we have $$\label{EqualDimension}
\dim \Dom_{\wedge,\max}/\Dom_{\wedge,\min} = \dim \Dom_{\max}/\Dom_{\min}$$ which by Theorem \[LeschTheorem\] is finite. It is known (cf. Lesch [@Le97]) that $\Dom_{\wedge,\max}/\Dom_{\wedge,\min}$ is isomorphic to a finite dimensional space $\Sing_{\wedge,\max}$ consisting of functions of the form $$\label{SingularFunctions}
\varphi=\sum_{\substack{\sigma\in\operatorname{spec}_b(A) \\|\Im\sigma|<m/2}}
\left(\sum_{k=0}^{m_\sigma} c_{\sigma,k}(y) \log^k x\right) x^{i\sigma}$$ where $c_{\sigma,k}\in C^\infty(Y;E)$. More precisely, for every $u\in\Dom_{\wedge,\max}$ there is a function $\varphi\in \Sing_{\wedge,\max}$ such that $u(x,y)-\omega(x)\varphi(x,y)\in \Dom_{\wedge,\min}$ for some (hence any) cut-off function $\omega \in C_0^\infty([0,1))$, $\omega=1$ near $0$. The function $\varphi$ is uniquely determined by the equivalence class $u+\Dom_{\wedge,\min}$.
We identify $$\Sing_{\wedge,\max} = \Dom_{\wedge,\max}/\Dom_{\wedge,\min}$$ and let $$\pi_{\wedge,\max}:\Dom_{\wedge,\max}\to \Sing_{\wedge,\max}$$ be the canonical projection.
Contrary to the situation in Theorem \[LeschTheorem\], the closed extensions of $A_\wedge$ do not need to be Fredholm. However, if $A-\lambda$ is $c$-elliptic with parameter, then the canonical extensions $A_{\wedge,\min}-\lambda$ and $A_{\wedge,\max}-\lambda$ are both Fredholm for $\lambda\not=0$, cf. [@GKM2 Remark 5.26]. Moreover, we have $$\label{IndexAwedgemin}
\operatorname{ind}(A_{\wedge,\min}-\lambda)=\operatorname{ind}A_{\Dom_{\min}},$$ cf. Corollary [5.35]{} in [@GKM2].
On $Y^\wedge=[0,\infty)\times S^1$ with the cone metric $dx^2+x^2 dy^2$, the Laplace-Beltrami operator is given by $$\label{ModelLaplacian}
\Delta_\wedge = x^{-2}\big((xD_x)^2 + \Delta_Y\big),$$ where $\Delta_Y$ is the nonnegative Laplacian on $S^1$. $\Delta_\wedge$ is precisely the model operator associated with the cone Laplacian $\Delta$ discussed in the previous section, cf. . It is easy to check that for any cut-off function $\omega\in
C_0^\infty([0,1))$, the functions $$\omega(x)\cdot 1,\; \Delta_\wedge(\omega(x)\cdot 1), \;
\omega(x)\log x,\, \text{ and } \Delta_\wedge(\omega(x)\log x)$$ are all in the space $x^{-1}L^2_b(Y^\wedge)$. Thus $\omega(x)\cdot 1$ and $\omega(x)\log x$ are elements of $\Dom_{\wedge,\max}$. In fact, $$\label{QuotientLaplacian}
\Sing_{\wedge,\max}=\operatorname{span}\{1,\log x\}.$$ Observe that $\Delta-\lambda$ is $c$-elliptic with parameter $\lambda\in \C\backslash \R_+$ and therefore the closed extensions of $\Delta_\wedge-\lambda$ are Fredholm for every $\lambda\in \C\backslash \overline{\R}_+$.
The model operator has a dilation/scaling property that can be exploited to analyze its closed extensions and their resolvents from a geometric point of view. In order to describe this property we first introduce the one-parameter group of isometries $$\R_+\ni \varrho \mapsto
\kappa_\varrho:x^{-m/2}L^2_b(Y^\wedge;E)\to x^{-m/2}L^2_b(Y^\wedge;E)$$ which on functions is defined by $$\label{kapparho}
(\kappa_\varrho f)(x,y)= \varrho^{m/2} f(\varrho x,y).$$ It is easily verified that the operator $A_\wedge$ satisfies the relation $$\kappa_\varrho A_\wedge =\varrho^{-m} A_\wedge\kappa_{\varrho}.$$ This implies $$\label{kappaHomogeneous}
A_\wedge-\lambda = \varrho^{m} \kappa_\varrho (A_\wedge-\lambda/\varrho^m)
\kappa_{\varrho}^{-1}$$ for every $\varrho>0$ and $\lambda\in\C$. This homogeneity property, called *$\kappa$-homogeneity*, will be used systematically to describe the closed extensions of $A_\wedge$ with nonempty resolvent sets.
It is convenient to introduce the set $$\operatorname{bg-res}A_\wedge=\{\lambda\in\C: A_{\wedge,\min}-\lambda \text{ is
injective and } A_{\wedge,\max}-\lambda \text{ is surjective}\},$$ the *background resolvent set* of $A_\wedge$, cf. [@GKM1].
If $\lambda\in \operatorname{bg-res}A_\wedge$ and $\Dom\in \mathfrak D_\wedge$, then $A_{\wedge,\Dom} -\lambda$ is Fredholm. The set $\operatorname{bg-res}A_\wedge$ is a disjoint union of open sectors, $$\operatorname{bg-res}A_\wedge = \bigcup_{\alpha\in \mathfrak I\subset \mathbb N}
\open \Lambda_\alpha.$$
This lemma follows immediately from .
For $\lambda\in \operatorname{bg-res}A_\wedge$ and $\Dom\in \mathfrak D_\wedge$ we have $$\label{RelIndexAwedge}
\operatorname{ind}(A_{\wedge,\Dom}-\lambda)=
\operatorname{ind}(A_{\wedge,\min}-\lambda)+\dim\Dom/\Dom_{\wedge,\min}.$$ Moreover, the map $$\open\Lambda_\alpha\ni \lambda\mapsto \operatorname{ind}(A_{\wedge,\Dom}-\lambda)$$ is constant since the embedding $\Dom\hookrightarrow
x^{-m/2}L^2_b(Y^\wedge;E)$ is continuous. Now, in analogy with we define $$\mathfrak G_{\wedge,\alpha}=\{\Dom\in\mathfrak D_\wedge:
\operatorname{ind}(A_{\wedge,\Dom}-\lambda)=0 \text{ for } \lambda\in
\open\Lambda_\alpha\}$$ and let $d''_\alpha= -\operatorname{ind}(A_{\wedge,\min}-\lambda)$ for $\lambda\in
\open\Lambda_\alpha$. We identify $\mathfrak G_{\wedge, \alpha}$ with the complex Grassmannian of $d''_\alpha$-dimensional subspaces of $\Sing_{\wedge,\max}$.
The canonical domains $\Dom_{\wedge,\min}$ and $\Dom_{\wedge,\max}$ are both $\kappa$-invariant. Thus the group action $\kappa_{\varrho}$ induces an action on $\Sing_{\wedge,\max}$. In general, $\kappa_{\varrho}$ does not preserve the elements of $\mathfrak D_\wedge$. In fact, the set of $\kappa$-invariant domains in $\mathfrak D_\wedge$ is an analytic variety because it consists of the stationary points of a holomorphic flow, cf. Section [7]{} in [@GKM1]. To better analyze the resolvents of the closed extensions of $A_\wedge$ over the open sector $\open \Lambda_\alpha$, we will consider the manifold $\mathfrak G_{\wedge,\alpha}$ together with the flow generated by the induced action of $\kappa_{\varrho}$ given by $\kappa_{\varrho}(\Dom/\Dom_{\wedge,\min})=
\kappa_{\varrho}(\Dom)/\Dom_{\wedge,\min}$.
The background resolvent set of $\Delta_\wedge$ is the open sector $\C\backslash \overline{\R}_+$; this is easily seen after noting that $\Delta_\wedge$ is the standard Laplacian in $\R^2$ written in polar coordinates. Moreover, since $\operatorname{ind}(\Delta_{\wedge,\min}-\lambda)= -1$ for every $\lambda\in \C\backslash \overline{\R}_+$, we have that $\Dom\in\mathfrak D_{\wedge}$ belongs to $\mathfrak G_\wedge$ if and only if $\dim \Dom/\Dom_{\wedge,\min}=1$. Thus $$\mathfrak G_\wedge \cong \mathbb{CP}^1 = S^2.$$ We identify $\Sing_{\wedge,\max}$ with $\Dom_{\wedge,\max}/\Dom_{\wedge,\min}$ and use to write $$\Sing_{\wedge,\max} = \operatorname{span}\left\{1,\log x\right\}.$$ For $\Dom\in \mathfrak G_\wedge$ we then have $$\label{FiniteDomainLaplace}
\pi_{\wedge,\max}\Dom = \operatorname{span}\left\{\zeta_0\cdot 1+\zeta_1\log x\right\}
\;\text{ for some } \zeta_0,\zeta_1\in\C, \, (\zeta_0,\zeta_1)\not=0.$$ Hence, with $\kappa$ as defined in , we get $$\label{FinitekappaD}
\pi_{\wedge,\max}\kappa_{\!\varrho}^{-1}\Dom
=\operatorname{span}\{(\zeta_0-\zeta_1\log\varrho)\cdot 1 + \zeta_1\log x\}.$$ Clearly, the only $\kappa$-invariant domain in $\mathfrak G_\wedge$ is the domain $\Dom_{F}$ such that $$\pi_{\wedge,\max}\Dom_{F}=\operatorname{span}\{1\};$$ $\Dom_F$ is precisely the domain of the Friedrichs extension of $\Delta_\wedge$, cf. [@GiMe01]. Every domain $\Dom\in
\mathfrak G_\wedge$ with $\zeta_1\not=0$ in generates a nontrivial orbit as given by . In order to describe the flow of $\kappa$ on these nonstationary points, rewrite as $$\pi_{\wedge,\max}\Dom =
\operatorname{span}\left\{\tfrac{\zeta_0}{\zeta_1}\cdot 1+\log x\right\}.$$ Then the projection to $\Sing_{\wedge,\max}$ of the dilation $\kappa_{\varrho}^{-1}\Dom$ is given by $$\label{FinitekappaD2}
\pi_{\wedge,\max}\kappa_{\!\varrho}^{-1}\Dom
=\operatorname{span}\left\{\big(\tfrac{\zeta_0}{\zeta_1}-\log\varrho\big)\cdot 1
+ \log x\right\}.$$ If $[\zeta_0:\zeta_1]\in\mathbb{CP}^1$ is the point corresponding to $\Dom$, then $\kappa_{\varrho}^{-1}\Dom$ is represented by $[\zeta_0-\zeta_1\log\varrho:\zeta_1]$. In other words, in the situation at hand, the flow generated by $\kappa$ on $\mathfrak G_\wedge \cong
\mathbb{CP}^1$ consists of curves that in projective coordinates are lines parallel to the real axis, see Figure \[figOrbits\].
(100,130) (-30,-10) (64,21) (74,18)[$\zeta_0/\zeta_1$]{} (60,59) (-20,75)[$\C$]{}
Observe that the Friedrichs extension corresponds to the point $[1:0]\in
\mathbb{CP}^1$. Using $$\pi_{\wedge,\max}\kappa_{\!\varrho}^{-1}\Dom
=\operatorname{span}\left\{1+\tfrac{\zeta_1}{\zeta_0-\zeta_1\log\varrho}
\log x\right\}\;\text{ if } \varrho\not=e^{\zeta_0/\zeta_1}$$ we see that $$\label{DomainConvergence}
\pi_{\wedge,\max}\kappa_{\!\varrho}^{-1}\Dom \to
\operatorname{span}\{1\} = \pi_{\wedge,\max}\Dom_F
\;\text{ as $\varrho\to\infty$ or $\varrho\to 0$.}$$
Let $\Dom_0\in\mathfrak G_\wedge$ be such that $\pi_{\wedge,\max}\Dom_0
=\operatorname{span}\{\log x\}$. This domain gives a selfadjoint extension of $\Delta_\wedge$ which on the sphere corresponds to the point $[0:1]$. The circle consisting of the orbit of $\Dom_0$ together with $\Dom_F$ is the set of domains of selfadjoint extensions of $\Delta_\wedge$.
Ray conditions on the model cone
================================
In this section we will discuss the existence of sectors of minimal growth for the model operator $A_\wedge\in x^{-m}\operatorname{Diff}^m_b(Y^\wedge;E)$ associated with a $c$-elliptic cone operator. We fix a component $\open\Lambda_\alpha$ of $\operatorname{bg-res}A_\wedge$ and let $\mathfrak G_{\wedge}=\mathfrak G_{\wedge,\alpha}$.
Let $\Lambda$ be a closed sector such that $\Lambda\backslash 0\subset
\open\Lambda_\alpha$, cf. , let $\operatorname{res}A_{\wedge,\Dom}$ be the resolvent set of $A_{\wedge,\Dom}$. The $\kappa$-invariant domains are the simplest domains to analyze.
Suppose $\Dom\in\mathfrak G_{\wedge}$ is $\kappa$-invariant. If there exists $\lambda_0\in\open\Lambda_\alpha$ such that $A_{\wedge,\Dom}-
\lambda_0$ is invertible, then $\open\Lambda_\alpha\subset
\operatorname{res}A_{\wedge,\Dom}$ and $\Lambda$ is a sector of minimal growth for $A_{\wedge,\Dom}$.
If $\Dom\in\mathfrak G_{\wedge}$ is not $\kappa$-invariant, the situation is more complicated. Nonetheless, in [@GKM1] we found a condition necessary and sufficient for a sector $\Lambda$ to be a sector of minimal growth for $A_{\wedge,\Dom}$. This condition is expressed in terms of finite dimensional spaces and projections that we proceed to discuss briefly.
For $\lambda\in\operatorname{bg-res}A_\wedge$ we let $$\K_{\wedge,\lambda}=\ker(A_{\wedge,\max}-\lambda).$$ Then $$\operatorname{res}A_{\wedge,\Dom}=\operatorname{bg-res}A_\wedge\cap \{\lambda:
\K_{\wedge,\lambda}\cap \Dom=0\},$$ and for $\lambda\in \operatorname{res}A_{\wedge,\Dom}$ we have $$\label{Dmax=KplusD}
\Dom_{\wedge,\max} = \K_{\wedge,\lambda}\oplus \Dom.$$ Projecting on $\Sing_{\wedge,\max}$, this direct sum induces the decomposition $$\label{EwedgeDecomposition}
\Sing_{\wedge,\max}= \pi_{\wedge,\max}\K_{\wedge,\lambda}
\oplus \pi_{\wedge,\max}\Dom,$$ and the projection on $\pi_{\wedge,\max} \K_{\wedge,\lambda}$ according to is given by the map $$\label{KmaxProjection}
\begin{gathered}
\hat\pi_{\K_{\wedge,\lambda}, \Dom}:
\Sing_{\wedge,\max} \to\Sing_{\wedge,\max} \\
(u+\Dom_{\wedge,\min}) \mapsto \pi_{\K_{\wedge,\lambda}, \Dom}u
+ \Dom_{\wedge,\min},
\end{gathered}$$ where $\pi_{\K_{\wedge,\lambda},\Dom}$ is the projection on $\K_{\wedge,\lambda}$ according to .
The following theorem gives a condition on the operator norm of for a sector $\Lambda$ to be a sector of minimal growth for $A_{\wedge,\Dom}$. Define $$\|u\|^2_\lambda=\|u\|^2+|\lambda|^{-2}\|A_\wedge u\|^2$$ for $\lambda\not=0$ and $u\in\Dom_{\wedge,\max}$.
\[NecAndSuffWedge\] Let $\Dom\in \mathfrak G_{\wedge}$, let $\Lambda$ be a closed sector with $\Lambda\backslash 0\subset \open\Lambda_\alpha$. Then $\Lambda$ is a sector of minimal growth for $A_{\wedge,\Dom}$ if and only if there are $C$, $R>0$ such that $\Lambda_R\subset \operatorname{res}A_{\wedge,\Dom}$, and $$\label{NewFormWedgeCond}
\big\|\hat\pi_{\K_{\wedge,\lambda},\Dom}
\big\|_{\L(\Sing_{\wedge,\max},\|\cdot\|_\lambda)}\leq C
\;\text{ for } \lambda\in \Lambda_R,$$ where $\hat\pi_{\K_{\wedge,\lambda},\Dom}$ is the projection .
This theorem is a rephrasing of [@GKM1 Theorem [8.7]{}]. There the condition appears in the equivalent form $$\label{NecAndSuffWedgeCond} \tag{\ref{NewFormWedgeCond}$'$}
\big\|\hat\pi_{\K_{\wedge,\hat\lambda},
\kappa_{|\lambda|^{1/m}}^{-1}\Dom}
\big\|_{\L(\Sing_{\wedge,\max})}\leq C \;\text{ for } \lambda\in \Lambda_R,$$ where $\hat\lambda=\lambda/|\lambda|$, and $\hat\pi_{\K_{\wedge,\hat\lambda},
\kappa_{|\lambda|^{1/m}}^{-1}\Dom}$ is the projection on $\K_{\wedge,\hat\lambda}$ induced (following the steps –) by the direct sum $$\label{Dmax=kappaKplusD}
\Dom_{\wedge,\max}=
\K_{\wedge,\varrho^{-m}\lambda}\oplus \kappa_\varrho^{-1}\Dom$$ for $\lambda\in\operatorname{res}A_{\wedge,\Dom}$ and $\varrho>0$. This decomposition is a consequence of and the $\kappa$-invariance of $\Dom_{\wedge,\max}$, as follows. First, the $\kappa$-homogeneity of $A_\wedge-\lambda$, cf. , implies $$\kappa_\varrho^{-1}(\K_{\wedge,\lambda})=
\K_{\wedge,\varrho^{-m}\lambda} \;\text{ for } \varrho>0.$$ Furthermore, if $\Dom\in\mathfrak G_\wedge$ and $\lambda\in\operatorname{bg-res}A_\wedge$, then $$\varrho^{-m}\lambda\in \operatorname{res}A_{\wedge,\kappa_\varrho^{-1}\Dom}
\Longleftrightarrow
\lambda\in \operatorname{res}A_{\wedge,\Dom}.$$ In particular, $$\K_{\wedge,\varrho^{-m}\lambda}\cap \kappa_\varrho^{-1}\Dom=\{0\}
\Longleftrightarrow \K_{\wedge,\lambda}\cap \Dom=\{0\},$$ as claimed.
The equivalence of and follows immediately from the identity $$\kappa_{|\lambda|^{1/m}}^{-1} \pi_{\K_{\wedge,
\lambda},\Dom}\, \kappa_{|\lambda|^{1/m}}= \pi_{\K_{\wedge,\hat\lambda},
\kappa_{|\lambda|^{1/m}}^{-1}\Dom}$$ using the relation and the fact that $\kappa$ is an isometry on $x^{-m/2}L^2_b$. The virtue of is that the norm is fixed, while the advantage of lies in that it gives a more explicit dependence on $\lambda$ and deals with a projection on a subspace of $\Sing_{\wedge,\max}$ with [*fixed*]{} complement $\pi_{\wedge,\max}\Dom$.
In [@GKM1 Corollary [8.22]{}] it is proved that $\Lambda$ is a sector of minimal growth for $A_{\wedge,\Dom}$ if and only if there are constants $C$, $R>0$ such that $\Lambda_R\subset \operatorname{res}A_{\wedge,\Dom}$ and $$\big\|\kappa_{|\lambda|^{1/m}}^{-1}(A_{\wedge,\Dom}-
\lambda)^{-1}\big\|_{\L(x^{-m/2}L^2_b,\Dom_{\wedge,\max})}
\leq C/|\lambda|,\quad \lambda\in \Lambda_R.$$ It can be shown that this estimate is equivalent to and .
We consider again the model Laplacian $\Delta_{\wedge}$ from the previous section. Recall that $\operatorname{bg-res}\Delta_{\wedge}=\C\backslash \overline{\R}_+$. For $\lambda\in \C\backslash\overline{\R}_+$, we have $$\pi_{\wedge,\max}\K_{\wedge,\lambda}=
\operatorname{span}\left\{-k_0 \log(-\lambda)+k_1\log x\right\}
\;\text{ for some } k_0, k_1>0,$$ where $\log$ means the principal branch of the logarithm. Moreover, by , $$\pi_{\wedge,\max}\kappa_{\!\varrho}^{-1}\Dom
=\operatorname{span}\left\{\big(\tfrac{\zeta_0}{\zeta_1}-\log\varrho\big)\cdot 1
+ \log x\right\}.$$ The projection in can be computed explicitly. Namely, if $u=\alpha_0+\alpha_1\log x\in\Sing_{\wedge,\max}
=\operatorname{span}\{1,\log x\}$ and $\lambda=\varrho^{m}\lambda_0$, then $$\label{DeltaKmaxProjection}
\hat\pi_{\K_{\wedge,\lambda_0}, \kappa_\varrho^{-1}\Dom}u
=\frac{-\alpha_0+\alpha_1(\frac{\zeta_0}{\zeta_1}-\log\varrho)} {k_0\log(-\lambda_0)+k_1(\frac{\zeta_0}{\zeta_1}-\log\varrho)}
\left(-k_0\log(-\lambda_0)+k_1\log x\right).$$
Let $\Lambda$ be a closed sector in $\C\backslash \R_+$ containing the half-plane $\{\Re\lambda<0\}$. Since the family of projections is bounded as $\varrho\to\infty$, uniformly for $|\lambda_0|=1$ in $\Lambda$, regardless of the specific choice of $\alpha_0$, $\alpha_1$, Theorem \[NecAndSuffWedge\] implies that every closed extension $\Delta_{\wedge,\Dom}$, $\Dom\in\mathfrak G_\wedge$, of the model Laplacian admits $\Lambda$ as a sector of minimal growth.
Equivalent geometric condition {#equivalent-geometric-condition .unnumbered}
------------------------------
We identify $\mathfrak G_{\wedge}$ with the Grassmannian $\Gr_{d''}(\Sing_{\wedge,\max})$ where $d''=-\operatorname{ind}(A_{\wedge,\min}-\lambda)$ for $\lambda\in\open\Lambda_\alpha\subset \operatorname{bg-res}A_\wedge$. Let $d'=\dim\K_{\wedge,\lambda}$. The condition that in the Grassmannian $\Gr_{d''}(\Sing_{\wedge,\max})$, the curve $$[R,\infty)\ni \varrho\mapsto \pi_{\wedge,\max}\kappa_\varrho^{-1}\Dom$$ does not approach the set $$\mathscr V_{\K_{\wedge,\lambda}}=\{D\in \Gr_{d''}(\Sing_{\wedge,\max}):
D\cap \pi_{\wedge,\max}\K_{\wedge,\lambda}\ne 0\}$$ as $\varrho\to\infty$, is sufficient for the validity of . This is [@GKM1 Theorem 8.28]. The following theorem states that the condition is also necessary.
For $D\in \Gr_{d''}(\Sing_{\wedge,\max})$ let $$\begin{aligned}
\Omega^{-}(D)=\big\{D'\in \Gr_{d''}(\Sing_{\wedge,\max})
:& \;\exists\, \{\varrho_k\}_{k=1}^\infty\subset \R_+ \text{ such that}\\
&\; \varrho_k\to\infty \text{ and }
\kappa_{\varrho_k}^{-1}D\to D' \text{ as } k\to\infty\big\}. \end{aligned}$$
\[GeometricNecAndSuff\] Let $\lambda_0\in\open\Lambda_\alpha$. The ray through $\lambda_0$ is a ray of minimal growth for $A_{\wedge,\Dom}$ if and only if $\Omega^{-}(\pi_{\wedge,\max}\Dom) \cap \mathscr V_{\K_{\wedge,\lambda_0}}
=\varnothing$.
Let $\lambda_0\in\open\Lambda_\alpha$ and $\Dom\in\mathfrak G_\wedge$. For simplicity, we use the notation $$\begin{gathered}
D=\pi_{\wedge,\max}\Dom, \quad
\mathscr V=\mathscr V_{\K_{\wedge,\lambda_0}}, \quad
\K=\pi_{\wedge,\max}\K_{\wedge,\lambda_0} \;\text{ and }\;
\pi_{\K,D}=\hat\pi_{\K_{\wedge,\lambda_0},\Dom}.\end{gathered}$$ Suppose $\Omega^{-}(D)\cap \mathscr V=\varnothing$. Since $\Omega^{-}(D)$ and $\mathscr V$ are closed sets, there are a neighborhood $\mathcal U$ of $\mathscr V$ and a constant $R>0$ such that if $\varrho>R$ then $\kappa_\varrho^{-1}D\not\in\mathscr V$. Then Lemma 5.24 in [@GKM1] gives that $\big\|\pi_{\K,\kappa_\varrho^{-1}D}\big\|$ is uniformly bounded as $\varrho\to\infty$, and therefore, by Theorem \[NecAndSuffWedge\] the ray through $\lambda_0$ is a ray of minimal growth for $A_{\wedge,\Dom}$.
Assume now that there are $C$, $R>0$ such that $\Lambda_R\subset \operatorname{res}A_{\wedge,\Dom}$ and the condition is satisfied. Suppose $\Omega^{-}(D)\cap \mathscr V \not=\varnothing$ and let $D_0\in \Omega^{-}(D)\cap \mathscr V$. Since $D_0\in
\mathscr V$, we have $D_0\cap\K\not=\{0\}$. On the other hand, $D_0\in
\Omega^{-}(D)$ implies that there is a sequence $\{\varrho_k\}_{k=1}^\infty\subset\R_+$ such that $\varrho_k\to\infty$ and $D_k=\kappa_{\varrho_k}^{-1}D\to D_0$ as $k\to\infty$. Note that for $\varrho_k$ large we have $\varrho_k^m\lambda_0 \in\operatorname{res}A_{\wedge,\Dom}$, so $\lambda_0\in \operatorname{res}A_{\wedge,\kappa_{\varrho_k}^{-1}\Dom}$ and therefore, $D_k\not\in \mathscr V$.
Pick $v\in D_0\cap \K$ with $\|v\|=1$. Let $\pi_{D_k}$ denote the orthogonal projection on $D_k$. Since $D_k\to D_0$ as $k\to\infty$, we have $\pi_{D_k}
\to \pi_{D_0}$, so $v_k=\pi_{D_k}v\to \pi_{D_0}v=v$ as $k\to\infty$. Since $D_k\not\in \mathscr V$, $v_k-v\not=0$ and $\pi_{\K,D_k}v_k=0$. Hence $$\pi_{\K,D_k}\bigg(\frac{v-v_k}{\|v-v_k\|}\bigg) =
\frac{v}{\|v-v_k\|} \to \infty \text{ as } k\to\infty,$$ since $\|v\|=1$ and $v_k\to v$ as $k\to\infty$. But this implies that $\|\pi_{\K,D_k}\|\to\infty$ contradicting the boundedness of the norm in . Thus $\Omega^{-}(D)\cap
\mathscr V =\varnothing$.
Let $\Delta_{\wedge}$ be the model Laplacian and let $\Dom\in\mathfrak G_{\wedge}$. In this case, the limiting set $\Omega^{-}(\pi_{\wedge,\max}\Dom)$ consists of the one element of $\mathbb{CP}^1$ corresponding to the Friedrichs extension of $\Delta_{\wedge}$, cf. . From this new perspective, it is evident that every closed extension of $\Delta_{\wedge}$ must admit a sector of minimal growth.
Rays of minimal growth {#sec-Necessity}
======================
We continue to assume that $A\in x^{-m}\operatorname{Diff}_b^m(M;E)$ is $c$-elliptic.
Unlike the case of a differential operator with smooth coefficients on a closed manifold, that a ray $\Gamma$ is a ray of minimal growth for the principal symbol $\csym(A)$ of $A$ is not expected to imply that $\Gamma$ is a ray of minimal growth for $A$. In this context, it is useful to think of $A_\wedge$ as a symbol (the wedge symbol) associated with $A$, cf. Schulze [@Sz89], so that it is natural to impose ray conditions on $A_\wedge$. For this to work, however, we need to transfer the information about the given domain $\Dom$ of $A$ on $M$ to equivalent information for $A_\wedge$ on $Y^\wedge$, and vice versa.
\[thetaIsomorphism\] There is a natural isomorphism $$\theta^{-1}:\Dom_{\wedge,\max}/\Dom_{\wedge,\min}\to \Dom_{\max}/\Dom_{\min}$$ given by a finite iterative procedure that involves the boundary spectrum of $A$ and the decomposition . In particular, if $A$ has coefficients independent of $x$ near $Y$, then $\theta$ is the identity map.
Let $M$ be a compact $2$-manifold with boundary $Y=S^1$. Let $A$ be a differential operator in $x^{-2}\operatorname{Diff}^2_b(M)$ that over the interior of $M$ coincides with some Laplacian, and near $Y$, is of the form $$A = x^{-2}\big((xD_x)^2 + q(x)\Delta_{Y}\big),$$ where $\Delta_Y$ is the standard nonnegative Laplacian on $S^1$ and $q$ is a smooth function. We assume $q$ to have the form $$q(x)= \alpha^2 + \beta x + x^2\gamma(x),$$ where $\alpha$, $\beta$ are constants such that $\frac12<\alpha<1$, $\beta\not=0$, and $\gamma(0)=1$. The associated model operator is then given by $$A_\wedge = x^{-2}\big((xD_x)^2 + \alpha^2\Delta_{Y}\big),$$ and $\operatorname{spec}_b(A)=\{\pm i\alpha k: k\in\mathbb N_0\}$. Since $\frac12<\alpha<1$, only the set $\{-i\alpha,0,i\alpha\}$ is relevant for the spaces $\Sing_{\max}$ and $\Sing_{\wedge,\max}$, cf. . Here, similar to $\Sing_{\wedge,\max}$, the space $\Sing_{\max}$ consists of singular functions and is isomorphic to the quotient $\Dom_{\max}/\Dom_{\min}$. If $y$ denotes the angular variable on $S^1$, $$\begin{gathered}
\Sing_{\wedge,\max}=
\operatorname{span}\{1,\log x, e^{iy} x^{\alpha}, e^{-iy} x^{\alpha},
e^{iy} x^{-\alpha},e^{-iy} x^{-\alpha}\}, \\
\Sing_{\max}=
\operatorname{span}\Big\{1,\log x, e^{\pm iy} x^{\alpha},
e^{\pm iy} x^{-\alpha}\big(1-\tfrac{\beta}{2\alpha-1}x\big)\Big\}. \end{gathered}$$ In this case, $\theta:\Sing_{\max}\to \Sing_{\wedge,\max}$ acts as the identity on $\operatorname{span}\{1,\log x, e^{\pm iy} x^{\alpha}\}$, but $$\theta\Big(e^{\pm iy}x^{-\alpha}
\big(1-\tfrac{\beta}{2\alpha-1}x\big)\Big) = e^{\pm iy}x^{-\alpha}.$$
The map $\theta$ induces an isomorphism $$\Theta:\mathfrak{D} \to \mathfrak{D}_\wedge$$ that we use to define $\Dom_\wedge=\Theta\Dom$ for any given $\Dom\in \mathfrak{D}$. The operator $A_{\wedge,\Dom_\wedge}$ is the closed extension of $A_\wedge$ in $x^{-m/2}L_b^2(Y^\wedge;E)$ uniquely associated with $A_\Dom$.
As in [@GKM2 Section 6], and motivated by the importance of $\kappa_\varrho$ in studying the model operator $A_\wedge$, we introduce on $\Dom_{\max}(A)/\Dom_{\min}(A)$ the one-parameter group $$\tilde \kappa_\varrho = \theta^{-1} \kappa_\varrho \theta
\;\text{ for } \varrho>0.$$
Similar to the situation on the model cone, the spectrum and resolvent of the closed extensions of $A$ can be geometrically analyzed by considering the manifold $\mathfrak G$, cf. , together with the flow generated by $\tilde\kappa_{\varrho}$.
An interesting consequence of Theorem \[thetaIsomorphism\] is the following.
\[IndexAwedgeD\] If $A-\lambda$ is $c$-elliptic with parameter $\lambda\not=0$, then $$\operatorname{ind}(A_{\wedge,\Dom_\wedge}-\lambda)=\operatorname{ind}A_\Dom.$$
The existence of $\theta$ implies $\dim \Dom_\wedge/\Dom_{\wedge,\min}=
\dim \Dom/\Dom_{\min}$. Now, the proposition follows by combining this identity with the relative index formulas and , together with the equation .
The following theorem describes the pseudodifferential structure of the resolvent of a cone operator $A$ and gives tangible conditions over a given sector $\Lambda$ on the symbols $\csym(A)$ and $A_\wedge$ for $A$ to have $\Lambda$ as a sector of minimal growth.
\[MainPaper2\] Let $A\in x^{-m}\operatorname{Diff}_b^m(M;E)$ be such that $A-\lambda$ is $c$-elliptic with parameter $\lambda\in \Lambda$. If $\Lambda$ is a sector of minimal growth for $A_{\wedge,\Dom_\wedge}$, then it is a sector of minimal growth for $A_\Dom$. Moreover, $$(A_\Dom-\lambda)^{-1}=B(\lambda) + G_{\Dom}(\lambda),$$ where $B(\lambda)$ is a parametrix of $A_{\Dom_{\min}}-\lambda$ with $B(\lambda)(A_{\Dom_{\min}}-\lambda)=1$ for $\lambda$ sufficiently large, and $G_{\Dom}(\lambda)$ is a pseudodifferential regularizing operator of finite rank.
The following lemma gives further information about the behavior at large of the resolvent along a sector of minimal growth.
Given two cut-off functions $\omega_0$ and $\omega_1$, the notation $\omega_1 \prec \omega_0$ will indicate that $\omega_0=1$ in a neighborhood of the support of $\omega_1$.
\[localizations\] Let $A \in x^{-m}\operatorname{Diff}_b^m(M;E)$ be $c$-elliptic and let $\Lambda$ be a sector of minimal growth for $A_{\Dom}$. For every pair of cut-off functions $\omega_1\prec\omega_0$, supported near the boundary, we have $$(1-\omega_0)(A_{\Dom}-\lambda)^{-1}\omega_1
\in \S\bigl(\Lambda,\L(x^{-m/2}L_b^2,\Dom_{\max})\bigr),$$ where $\S$ stands for Schwartz (rapidly decreasing as $|\lambda|\to\infty$).
Since $\Lambda$ is a sector of minimal growth for $A_{\Dom}$, the family $A-\lambda$ must be $c$-elliptic with parameter $\lambda\in \Lambda$, and $A_{\wedge,\min}-\lambda$ must be injective for every $\lambda\in\Lambda$, $\lambda\not=0$. A proof of this can be found in [@GKM2 Theorem 4.1].
As a consequence (cf. [@GKM2 Section 5]), there is a parametrix $B(\lambda)$ such that $B(\lambda)(A_{\Dom_{\min}}-\lambda)=1$ for large $\lambda\in \Lambda$, and $$\label{LocalizedParam}
(1-\omega_0) B(\lambda) \omega_1 \in
\S\bigl(\Lambda,\L(x^{-m/2}L_b^2,\Dom_{\max})\bigr)$$ for all cut-off functions $\omega_1\prec\omega_0$ supported near the boundary. We now make use of the identity $$(A_{\Dom}-\lambda)^{-1} = B(\lambda) +
(1-B(\lambda)(A-\lambda))(A_{\Dom}-\lambda)^{-1}.$$ Multiplying by $(1-\omega_0)$ from the left and by $\omega_1$ from the right, proves the assertion for the first term involving $B(\lambda)$. On the other hand, since $1-B(\lambda)(A-\lambda)$ vanishes on $\Dom_{\min}$ for large $\lambda$, we have for such $\lambda$, $$\begin{aligned}
(1-\omega_0)(1-B(\lambda)(A-\lambda))
&= (1-\omega_0)(1-B(\lambda)(A-\lambda))\omega_2 \\
&= -(1-\omega_0)B(\lambda)(A-\lambda)\omega_2 \\
&= -(1-\omega_0)B(\lambda)\omega_1(A-\lambda)\omega_2\end{aligned}$$ whenever $\omega_2\prec\omega_1$. Thus, by , $$(1-\omega_0)(1-B(\lambda)(A-\lambda)): \Dom_{\max} \to \Dom_{\max}$$ is rapidly decreasing as $|\lambda| \to \infty$. Finally, the assertion of the lemma can be completed using the fact that $(A_\Dom-\lambda)^{-1}
\omega_1: x^{-m/2}L_b^2 \to \Dom_{\max}$ is uniformly bounded.
Necessity of the conditions {#necessity-of-the-conditions .unnumbered}
---------------------------
The converse of Theorem \[MainPaper2\] involves proving that the minimal growth of the resolvent $(A_\Dom-\lambda)^{-1}$ over a sector $\Lambda$ implies a corresponding behavior for the inverse of $\csym(A)-
\lambda$ and for the resolvent $(A_{\wedge,\Dom_{\wedge}}-\lambda)^{-1}$.
While in [@GKM2 Theorem 4.1] we established the necessity of the condition on $\csym(A)$, we did not address the question whether $\Lambda$ must necessarily be a sector of minimal growth for $A_{\wedge,\Dom_{\wedge}}$. In the next theorem we prove that this is indeed the case when $A$ has coefficients independent of $x$ near $Y=\partial M$.
\[NecessityConstCoeff\] Let $A\in x^{-m}\operatorname{Diff}_b^m(M;E)$ be $c$-elliptic with coefficients independent of $x$ near $Y$. If $\Lambda$ is a a sector of minimal growth for $A_\Dom$, then $A-\lambda$ is $c$-elliptic with parameter $\lambda\in\Lambda$, and $\Lambda$ is a sector of minimal growth for $A_{\wedge,\Dom_\wedge}$.
As stated in the proof of Lemma \[localizations\], the assumption on the resolvent of $A_{\Dom}$ implies that $A-\lambda$ is $c$-elliptic with parameter $\lambda\in \Lambda$ and that $A_{\wedge,\min}-\lambda$ is injective for every $\lambda\not=0$. Thus we only need to prove the statement about $A_{\wedge,\Dom_\wedge}$.
By Proposition \[IndexAwedgeD\], and since $\operatorname{ind}A_\Dom=0$, we have $\operatorname{ind}(A_{\wedge,\Dom_\wedge}-\lambda)=0$ for $\lambda\not=0$. For this reason, in order to show that $\Lambda$ is a sector of minimal growth for $A_{\wedge,\Dom_\wedge}$, it suffices to find (for large $\lambda\in\Lambda$) a right-inverse of $A_{\wedge,\Dom_\wedge}-\lambda$ that is uniformly bounded in $\L\big(x^{-m/2}L_b^2,\Dom_\wedge\big)$ as $|\lambda|\to\infty$.
Since $A$ is assumed to have coefficients independent of $x$ near the boundary, there is a cut-off function $\omega_0$ such that $$A\omega_0=A_\wedge \omega_0 \quad\text{and}\quad
\omega_0\Dom=\omega_0\Dom_\wedge.$$ Let $\omega_1$, $\omega_2$ be cut-off functions with $\omega_2\prec
\omega_1\prec \omega_0$. Then the operator $$B(\lambda)=\omega_1 (A_\Dom-\lambda)^{-1} \omega_2$$ can be regarded as an operator on $M$ with values in $\Dom$ or as an operator on $Y^\wedge$ with values in $\Dom_\wedge$. Depending on the context we will write $B(\lambda)$ as $$B_\Dom(\lambda):x^{-m/2}L^2_b(M;E)\to\Dom \quad\text{or}\quad
B_{\Dom_\wedge}(\lambda):x^{-m/2}L^2_b(Y^\wedge;E)\to\Dom_{\wedge}.$$ On $M$ we consider $$\begin{aligned}
(A_\Dom-\lambda)B_\Dom(\lambda)
&=\omega_0(A_\Dom-\lambda)\omega_1 (A_\Dom-\lambda)^{-1} \omega_2 \\
&=\omega_2 -\omega_0(A_\Dom-\lambda)(1-\omega_1)
(A_\Dom-\lambda)^{-1}\omega_2 \\
&=\omega_2 + R(\lambda)\end{aligned}$$ with $R(\lambda)=-\omega_0(A_\Dom-\lambda)(1-\omega_1)(A_\Dom-\lambda)^{-1}
\omega_2$. By Lemma \[localizations\], $R(\lambda)$ is rapidly decreasing in the norm as $|\lambda|\to \infty$.
Because of the presence and nature of the cut-off functions $\omega_0$ and $\omega_2$, $R(\lambda)$ can also be regarded as an operator on $Y^\wedge$, say $R_\wedge(\lambda)\in \S\big(\Lambda,\L(x^{-m/2}L^2_b)\big)$. Now, using that $(A_\Dom-\lambda)\omega_1= (A_{\wedge,\Dom_\wedge}-\lambda)
\omega_1$, we get on $Y^\wedge$ the identity $$\label{ParamNearBoundary}
(A_{\wedge,\Dom_{\wedge}}-\lambda)B_{\Dom_\wedge}(\lambda)=
\omega_2 + R_\wedge(\lambda).$$ Furthermore, we have $$\|B_{\Dom_\wedge}(\lambda)\|_{\L(x^{-m/2}L^2_b,
\Dom_{\wedge,\max})}= O(1) \;\text{ as } |\lambda|\to\infty,$$ since, by assumption, $\|B_{\Dom}(\lambda)\|_{\L(x^{-m/2}L^2_b,\Dom_{\max})}$ has the same asymptotic behavior.
On the other hand, as $A-\lambda$ is $c$-elliptic with parameter, by [@GKM2 Theorem 5.24] there is a family of pseudodifferential operators $B_{2,\wedge}(\lambda):x^{-m/2}L^2_b\to \Dom_{\wedge,\min}$ (uniformly bounded in $\lambda$) such that $(A_\wedge-\lambda)
B_{2,\wedge}(\lambda)-1$ is regularizing, and for $\omega_3\prec \omega_2$, the families $\omega_3 B_{2,\wedge}(\lambda) (1-\omega_2)$ and $\big[(A_\wedge-\lambda)B_{2,\wedge}(\lambda)-1\big](1-\omega_2)$ are rapidly decreasing in the norm as $|\lambda| \to \infty$. Thus, as $A_{\wedge,\Dom_\wedge}(1-\omega_3)=A_{\wedge}(1-\omega_3)$, $$\label{ParamAwayBoundary}
(A_{\wedge,\Dom_\wedge}-\lambda)(1-\omega_3)B_{2,\wedge}(\lambda)(1-\omega_2)
= (1-\omega_2) + S_\wedge(\lambda)$$ with $S_\wedge(\lambda)\in \S\big(\Lambda,\L(x^{-m/2}L^2_b)\big)$. Finally, the operator family $$Q_\wedge(\lambda)
=B_{\Dom_\wedge}(\lambda)+(1-\omega_3)B_{2,\wedge}(\lambda)(1-\omega_2)
:x^{-m/2}L^2_b\to \Dom_{\wedge,\max}$$ is bounded in the norm as $|\lambda|\to\infty$ and by and we have $$(A_{\wedge,\Dom_\wedge}-\lambda)Q_\wedge(\lambda)-1 \in
\S\big(\Lambda,\L(x^{-m/2}L^2_b)\big).$$ By a Neumann series argument, it follows that $A_{\wedge,\Dom_\wedge}-\lambda:\Dom_\wedge\to x^{-m/2}L^2_b$ has a uniformly bounded right-inverse for large $\lambda\in\Lambda$.
[99]{}
M. Agranovich and M. Vishik, *Elliptic problems with a parameter and parabolic problems of general type*, Russ. Math. Surveys **19** (1963), 53–159.
S. Agmon, *On the eigenfunctions and on the eigenvalues of general elliptic boundary value problems*, Comm. Pure Appl. Math. **15** (1962), 119–147.
J. Brüning and R. Seeley, *The expansion of the resolvent near a singular stratum of conical type*, J. Funct. Anal. **95** (1991), no. 2, 255–290.
J. Gil, *Full asymptotic expansion of the heat trace for non-self-adjoint elliptic cone operators*, Math. Nachr. **250** (2003), 25–57.
J. Gil, T. Krainer, and G. Mendoza, *Geometry and spectra of closed extensions of elliptic cone operators*, preprint math.AP/0410178 at arXiv.org, to appear in Canadian Journal of Mathematics.
J. Gil, T. Krainer, and G. Mendoza, *Resolvents of elliptic cone operators*, preprint math.AP/0410176 at arXiv.org.
J. Gil and G. Mendoza, *Adjoints of elliptic cone operators*, Amer. J. Math. **125** (2003) 2, 357–408.
G. Grubb, *Functional calculus of pseudodifferential boundary problems*, 2nd ed., Progress in Mathematics, vol. 65. Birkh[ä]{}user, Basel, 1996.
V. Kondrat’ev, *Boundary problems for elliptic equations in domains with conical or angular points*, Trans. Mosc. Math. Soc. **16** (1967), 227–313.
T. Krainer, *Resolvents of elliptic boundary problems on conic manifolds*, preprint math.AP/0503021 at arXiv.org.
M. Lesch, *Operators of [F]{}uchs type, conical singularities, and asymptotic methods*, Teubner-Texte zur Math. vol 136, B.G. Teubner, Stuttgart, Leipzig, 1997.
P. Loya, *On the resolvent of differential operators on conic manifolds*, Comm. Anal. Geom. **10** (2002), no. 5, 877–934.
R. Melrose, *Transformation of boundary value problems*, Acta Math. **147** (1981), no. 3-4, 149–236.
E. Schrohe and J. Seiler, *The resolvent of closed extensions of cone differential operators*, Canad. J. Math. **57** (2005), no. 4, 771–811.
B.-W. Schulze, *Pseudo-differential operators on manifolds with edges*, Proc. Symp. ‘Partial Differential Equations’, Holzhau 1988 (Leipzig), Teubner-Texte zur Math., vol. 112, Teubner, 1989, pp. 259–288.
R. Seeley, *Complex powers of an elliptic operator*, Singular Integrals, AMS Proc. Symp. Pure Math. X, 1966, Amer. Math. Soc., Providence, 1967, pp. 288–307.
Acknowledgment {#acknowledgment .unnumbered}
--------------
The new results contained herein reflect part of work carried out by the three authors at the Mathematisches Forschungsinstitut Oberwolfach under their “Research in Pairs” program. They gratefully acknowledge the Institute’s support and hospitality.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We generalize a theorem of Chifan and Ioana by proving that for any, possibly type III, amenable von Neumann algebra $A_0$, any faithful normal state $\varphi_0$ and any discrete group $\Gamma$, the associated Bernoulli crossed product von Neumann algebra $M=(A_0,\varphi_0)^{{\mathbin{\overline{\otimes}}}\Gamma}\rtimes \Gamma$ is solid relatively to $\mathcal{L}(\Gamma)$. In particular, if $\mathcal{L}(\Gamma)$ is solid then $M$ is solid and if $\Gamma$ is non-amenable and $A_0 \neq \mathbb{C}$ then $M$ is a full prime factor. This gives many new examples of solid or prime type $\mathrm{III}$ factors. Following Chifan and Ioana, we also obtain the first examples of solid non-amenable type $\mathrm{III}$ equivalence relations.'
address:
- |
École Normale Supérieure\
45 rue d’Ulm 75230 Paris Cedex 05\
France
- |
Laboratoire de Mathématiques d’Orsay\
Université Paris-Sud\
Université Paris-Saclay\
91405 Orsay\
France
author:
- Amine Marrakchi
bibliography:
- 'database.bib'
title: Solidity of type III Bernoulli crossed products
---
[^1]
Introduction
============
In [@TakaSolid], Ozawa discovered a remarkable rigidity property of von Neumann algebras that he called *solidity*. One of the many possible definitions is the following one: a von Neumann algebra $M$ is *solid* if for every subalgebra with expectation $Q \subset M$ there exists a sequence of projections $z_n \in \mathcal{Z}(Q)$ with $\sum_n z_n =1$ such that $Qz_0$ is amenable and $Qz_n$ is a non-amenable factor for all $n \geq 1$. The main interest of this notion is that any solid non-amenable factor is automatically *full* (every centralizing sequence is trivial) and *prime* (not a tensor product of two non type $\mathrm{I}$ factors). Ozawa’s celebrated result [@TakaSolid] states that the group von Neumann algebra $\mathcal{L}(\Gamma)$ of any hyperbolic group $\Gamma$ is solid.
A closely related property in the context of *equivalence relations* was discovered by Chifan and Ioana in [@ChifanIoanaBern]. An equivalence relation $\mathcal{R}$ on a probability space $(X,\mu)$ is *solid* (or *solidly ergodic* [@GaboriauSurvey Definition 5.4]) if for every subequivalence relation $\mathcal{S} \subset \mathcal{R}$ there exists a measurable partition $X=\bigsqcup_n X_n$ by $\mathcal{S}$-invariant subsets such that $\mathcal{S}_{|X_0}$ is amenable and $\mathcal{S}_{|X_n}$ is non-amenable and ergodic for all $n \geq 1$. A solid ergodic non-amenable equivalence relation is automatically *strongly ergodic* (every sequence of almost invariant subsets is trivial) and *prime* (not a product of two non type $\mathrm{I}$ equivalence relations). The main theorem of Chifan and Ioana [@ChifanIoanaBern Theorem 1] states that the orbital equivalence relation of a Bernoulli action of *any* countable group $\Gamma$ is solid. In fact, they deduce this solidity result from a stronger theorem [@ChifanIoanaBern Theorem 2] which essentially says that for any *tracial* amenable von Neumann algebra $(A_0,\tau_0)$ and any discrete group $\Gamma$, the *Bernoulli crossed product von Neumann algebra* $M=(A_0,\tau_0)^{{\mathbin{\overline{\otimes}}}\Gamma}\rtimes \Gamma$ is *solid relatively to* $\mathcal{L}(\Gamma)$, a notion which we define precisely in section 3. Unlike the method of Ozawa which is based on $C^*$-algebraic techniques and requires the group $\Gamma$ to be exact [@TakaAmenAct Proposition 4.5 and 4.6], the approach of Chifan and Ioana is based on Popa’s *deformation/rigidity theory* [@PopaSurvey] and his *spectral gap rigidity* principle [@PopaSpectralGap] so that they do not need to make any assumption on $\Gamma$. Using the same approach, R. Boutonnet was able to generalize their results to *Gaussian actions* [@RemiGaussian].
Our main theorem generalizes the relative solidity result of Chifan and Ioana to non-tracial, possibly type $\mathrm{III}$, Bernoulli crossed products.
\[solid\] Let $A_0$ be an amenable von Neumann algebra with separable predual, $\varphi_0$ a faithful normal state on $A_0$ and $\Gamma$ any countable group. Let $M=(A_0,\varphi_0)^{{\mathbin{\overline{\otimes}}}\Gamma}\rtimes \Gamma$ be the associated Bernoulli crossed product von Neumann algebra. For any subalgebra with expectation $Q \subset 1_QM1_Q$ such that $Q \nprec_M \mathcal{L}(\Gamma)$ (see Section 2 for Popa’s intertwining symbol $\prec_M$) there exists a sequence of projections $z_n \in \mathcal{Z}(Q)$ with $\sum_n z_n =1_Q$ such that
- $Qz_0$ is amenable.
- $Qz_n$ is a full prime factor for all $n \geq 1$.
In particular, if $\Gamma$ is non-amenable and $A_0 \neq \mathbb{C}$ then $M$ is a full prime factor and if $\mathcal{L}(\Gamma)$ is solid then $M$ is solid.
The fullness of type $\mathrm{III}$ Bernoulli crossed products by non-amenable groups was already established by S.Vaes and P.Verraedt in [@Vaes2015296]. However, Theorem \[solid\] provides many new examples of prime or solid type $\mathrm{III}$ factors. In fact, in [@ConnesAlmostPeriodic], A. Connes introduced two new invariants for type $\mathrm{III}$ factors: the $\mathrm{Sd}$ invariant and the $\tau$ invariant. He used the noncommutative Bernoulli crossed products to construct type $\mathrm{III}$ factors with prescribed invariants $\mathrm{Sd}$ and $\tau$. By combining these constructions with Theorem \[solid\], we obtain the following corollary:
\[invariants\] For every countable subgroup $\Lambda \subset \mathbb{R}^*_+$, there exists a solid non-amenable type $\mathrm{III}$ factor with separable predual and with a Cartan subalgebra such that its $\mathrm{Sd}$ invariant is $\Lambda$. For any topology $\tau_0$ on $\mathbb{R}$ induced by an injective continuous separable unitary representation of $\mathbb{R}$, there exists a solid non-amenable type $\mathrm{III}_1$ factor with separable predual and with a Cartan subalgebra such that its $\tau$ invariant is $\tau_0$.
Note that there was previously no known example of a non-amenable solid type $\mathrm{III}$ factor with a Cartan subalgebra. Using Theorem \[solid\] we can generalize [@ChifanIoanaBern Theorem 7] by removing the assumption that the base equivalence relation is measure preserving and hence we also obtain the first examples of non-amenable solid type $\mathrm{III}$ equivalence relations.
\[equivalence\] Let $(X_0,\mu_0)$ be a probability space, $\mathcal{R}_0$ an arbitrary amenable non-singular equivalence relation on $X_0$ and $\Gamma$ any countable group. Let $\mathcal{R}=\mathcal{R}_0 \wr \Gamma$ be the *wreath product equivalence relation* on $(X_0,\mu_0)^{\Gamma}$ defined by $(x_i) \sim_{\mathcal{R}} (y_i)$ if and only if there exists $g \in \Gamma$ and a finite subset $F \subset \Gamma$ such that
- $\forall i \in F, \; x_i \sim_{\mathcal{R}_0} y_{gi}$
- $\forall i \in \Gamma \setminus F, \; x_i=y_{gi}$
Then $\mathcal{R}$ is solid, i.e. for any subequivalence relation $\mathcal{S} \subset \mathcal{R}$ there exists a countable partition of $(X_0,\mu_0)^{\Gamma}$ into $\mathcal{S}$-invariant components $Z_n, n \in \mathbb{N}$ such that
- $\mathcal{S}_{| Z_0}$ is amenable.
- $\mathcal{S}_{| Z_n}$ is strongly ergodic and prime for all $n \geq 1$.
This article is organized in the following way. Section 2 is devoted to preliminaries. We recall in particular the theory of ultraproducts for non-tracial von Neumann algebras. We also recall *Popa’s intertwining theorem* [@PopaBetti] and its recent generalization to arbitrary von Neumann algebras by C. Houdayer and Y. Isono [@HoudIsoUPF]. In Section 3, we introduce the notion of *relative solidity* and its main properties. In Section 4, we recall the notion of *s-malleable deformations* [@PopaMalleable1] and prove an abstract non-tracial version of Popa’s *spectral gap rigidity* argument [@PopaSpectralGap] by using ultraproduct techniques. Finally in Section 5, we prove our main theorems. Our proof follows the lines of Chifan and Ioana’s original proof. Indeed, we introduce the same s-malleable deformation of the Bernoulli crossed product and we show that the same bimodule computation, and hence the key spectral gap property, still holds in the non-tracial situation. Therefore we can apply the spectral gap argument of Section 4. Houdayer and Isono’s generalization of Popa’s intertwining theorem is then used to prove a crucial dichotomy for rigid subalgebras, similar to the one used by Chifan and Ioana, from which the conclusion follows easily. Altogether, the proof does not use Takesaki’s continuous decomposition of type $\mathrm{III}$ factors.
### Acknowledgment {#acknowledgment .unnumbered}
We are very grateful to our advisor C. Houdayer for suggesting this problem and for his help during this work. We also thank R. Boutonnet for his useful comments.
Preliminaries
=============
All mentioned von Neumann algebras $M$ are supposed to be *countably decomposable* (or *$\sigma$-finite*). This means that $M$ admits a faithful normal state. However, we do not assume that the predual $M_*$ is separable unless it is explicitly stated. We say that $M$ is *diffuse* if it has no minimal projection, i.e. $pMp \neq \mathbb{C}p$ for every non-zero projection $p \in M$. We say that $M$ is *properly non-amenable* if $Mz$ is non-amenable for every non-zero projection $z \in \mathcal{Z}(M)$. Equivalently $M$ is properly non-amenable if and only if $pMp$ is non-amenable for every non-zero projection $p \in M$.
An inclusion $N \subset M$ of two von Neumann algebras is always assumed to be unital. For non-unital inclusions we will use the notation $N \subset 1_NM1_N$ where the projection $1_N \in M$ is the unit of $N$. An inclusion $N \subset M$ is said to be *with expectation* if there is a faithful normal conditional expectation $E_N: M \rightarrow N$.
For every von Neumann algebra $M$, the Hilbert space $L^{2}(M)$ denotes the *standard form* of $M$ [@TakesakiII Chapter $\mathrm{IX}$, Section $1$]. For every $\xi \in L^2(M)$ we use the $M$-$M$-bimodule notation $a\xi b$ for the left and right action of $a,b \in M$. For every normal faithful semi-finite weight $\psi$ on $M$, there is a canonical map, denoted $x \mapsto x \psi^{1/2}$, from $\mathfrak{n}_\psi = \{ x \in M \mid \psi(x^*x) < +\infty \}$ to $L^2(M)$ such that $\langle x \psi^{1/2} , y \psi^{1/2} \rangle = \psi(y^*x)$ for all $x,y \in \mathfrak{n}_\psi$. This map identifies $L^2(M)$ with the completion of $\mathfrak{n}_\psi$ with respect to the inner product $(x,y) \mapsto \psi(y^*x)$. Moreover $\mathfrak{n}_\psi$ is a left ideal in $M$ and the map $x \mapsto x \psi^{1/2}$ is compatible with the left multiplication: $a(x\psi^{1/2})=(ax)\psi^{1/2}$ for every $a \in M, x \in \mathfrak{n}_\psi$. Similarly there is a map $x \mapsto \psi^{1/2}x$ from $\mathfrak{n}_\psi^*=\{ x \in M \mid \psi(xx^*) < + \infty \}$ to $L^2(M)$ with the similar properties. If $\psi$ is finite then $1 \in \mathfrak{n}_\psi=\mathfrak{n}_\psi^*=M$ and $\psi^{1/2}$ makes sense as a vector in $L^2(M)$ in this case.
For a von Neumann algebra $M$ with a faithful normal semi-finite weight $\psi$, let $t \mapsto \sigma_t^\psi$ denote the modular flow of $\psi$. An element $x \in M$ is said to be $\psi$-*analytic* if the function $t \mapsto \sigma_t^\psi(x)$ extends to an analytic function defined on the entire complex plane $\mathbb{C}$, which will be still denoted $z \mapsto \sigma_z^\psi(x)$. The $\psi$-analytic elements form a dense $*$-algebra in $M$ (see [@TakesakiII Chapter VIII, Lemma 2.3]). We let $M^\psi$ denote the centralizer of $\psi$, i.e. the fixed point algebra of $\sigma^\psi$. For every $\psi$-analytic element $x \in M$ we have $x\psi^{1/2}=\psi^{1/2} \sigma_{i/2}^\psi(x)$.
If $E_N: M \rightarrow N$ is a faithful normal conditional expectation and $\varphi$ is a faithful normal state on $N$ then for the faithful normal state on $M$ defined by $\widetilde{\varphi}=\varphi \circ E_N$ we have that $$\forall x \in N, \; \sigma_t^\varphi(x)=\sigma_t^{\widetilde{\varphi}}(x).$$ In particular $N$ is globally invariant by $\sigma_t^\psi$. Conversely, if $\varphi$ is a faithful normal state on $M$ such that $N$ is globally invariant by $\sigma_t^\varphi$ for all $t \in \mathbb{R}$, then by a theorem of Takesaki [@TakesakiII Chapter IX, Theorem 4.2] there exists a unique faithful normal conditional expectation from $M$ to $N$ which preserves $\varphi$. In particular, any subalgebra of $M^\varphi$ is with expectation in $M$. Also, we see that if $A \subset M$ is with expectation then $A' \cap M$ and $A \vee (A' \cap M)$ are also subalgebras with expectation in $M$.
Another important consequence is the following lemma which will be used frequently, without explicit reference. It already appears in [@HoudUedaFreeRig Proposition 2.2] but for the reader’s convenience, we reproduce it here with a simpler proof.
\[expectation\] Let $A \subset M$ be an inclusion with expectation and let $e$ be a projection in $A$ or in $A' \cap M$. Then $eAe \subset eMe$ is also with expectation.
Let $e \in A$ be a projection and $E: M \rightarrow A$ a faithful normal conditional expectation. Then $E(eMe) \subset eAe$, hence $E$ restricts to a faithful normal conditional expectation from $eMe$ to $eAe$.
Now let $e \in A' \cap M$. First we replace $E$ by a new faithful normal conditional expectation $\tilde{E}: M \rightarrow A$ defined by $$\forall x \in M, \; \tilde{E}(x)=E(exe+(1-e)x(1-e)).$$ Now, pick $\varphi$ a faithful normal state on $A$ and let $\psi = \varphi \circ \tilde{E}$. Then we have $\sigma^\varphi=\sigma^\psi_{|A}$. Also, by construction $e$ is in the centralizer of $\psi$. Now define a new faithful normal state $\phi = \psi_{|eMe}$. Since $e \in M^\psi$, we have $\forall x \in eMe, \; \sigma_t^\phi(x)=\sigma_t^\psi(x)$. In particular, it is then clear that $eAe$ is globally invariant by $\sigma^\phi$ in $eMe$. Hence it is with expectation by Takesaki’s theorem.
Ultraproducts {#ultraproducts .unnumbered}
-------------
In this section we recall the construction of ultraproducts for non-tracial von Neumann algebras. A general reference on this topic is [@AndoHaagUltra].
Let $M$ be any von Neumann algebra. Let $\omega$ be any free ultrafilter $\omega \in \beta \mathbb{N} \setminus \mathbb{N}$. In the von Neumann algebra $\ell^\infty(\mathbb{N},M)$ we define a $C^*$-subalgebra $$\mathcal{I}^\omega(M)=\{ (x_n)_{n} \in \ell^\infty(\mathbb{N},M) \mid \lim_{n \rightarrow \omega} x_n = 0 \; *\text{-strongly} \}.$$ When $M$ is finite, $\mathcal{I}^\omega(M)$ is an ideal of $\ell^\infty(\mathbb{N},M)$ and one defines the *ultraproduct* algebra $M^\omega$ as the quotient $\ell^\infty(\mathbb{N},M)/\mathcal{I}^\omega(M)$.
When $M$ is not finite, $\mathcal{I}^\omega(M)$ is no longer an ideal and one introduces instead its *multiplier algebra* $$\mathcal{M}^\omega(M)=\{ (x_n)_{n} \in \ell^\infty(\mathbb{N},M) \mid (x_n)_{n}\mathcal{I}^\omega(M) \subset \mathcal{I}^\omega(M) \text{ and } \mathcal{I}^\omega(M)(x_n)_{n} \subset \mathcal{I}^\omega(M) \}.$$ The quotient $M^\omega=\mathcal{M}^\omega(M)/\mathcal{I}^\omega(M)$ is always a von Neumann algebra [@OcneanuUltra Theorem 5.1] called the *Ocneanu ultraproduct* of $M$. If $(x_n)_n \in \mathcal{M}^\omega(M)$ we denote by $(x_n)^\omega_n$ its image in the quotient $M^\omega$. One can identify $M$ with the algebra of constant sequences in $\mathcal{M}^\omega(M)$ and hence and we have a natural inclusion $M \subset M^\omega$. This inclusion is with expectation since we have a canonical faithful normal conditional expectation $E^\omega: M^\omega \rightarrow M$ defined by $$E^\omega((x_n)^\omega_n)=\lim_{n \rightarrow \omega} x_n \; \text{ in the weak* topology.}$$ If $\varphi$ is a faithful normal state on $M$, we denote by $\varphi^\omega$ the faithful normal state on $M^\omega$ defined by $\varphi^\omega= \varphi \circ E^\omega$. It holds for the modular flow of $\varphi^\omega$ that $$\sigma_t^{\varphi^\omega}((x_n)_n^\omega)=(\sigma_t^{\varphi}(x_n))_n^\omega.$$
Let $N \subset M$ be a subalgebra. Then $\ell^\infty(\mathbb{N},N) \subset \ell^\infty(\mathbb{N},M)$ and $\mathcal{I}^\omega(N) \subset \mathcal{I}^\omega(M)$. If $N \subset M$ is with expectation then we also have $\mathcal{M}^\omega(N) \subset \mathcal{M}^\omega(M)$. Hence, in this case, we can identify $N^\omega$ canonically with a von Neumann subalgebra of $M^\omega$. Moreover, the inclusion $N^\omega \subset M^\omega$ is with expectation. Indeed if $E_N: M \rightarrow N$ is a faithful normal conditional expectation, then we can define a faithful normal conditional expectation $E_{N^\omega}: M^\omega \rightarrow N^\omega$ by $$E_{N^\omega}((x_n)_n^\omega)=(E_N(x_n))_n^\omega.$$
Popa’s intertwining theory {#popas-intertwining-theory .unnumbered}
--------------------------
In this section we recall the powerful method for intertwining subalgebras developed by S. Popa [@PopaBetti Appendix] and [@PopaMalleable1 Theorem 2.1, Corollary 2.3]. We will also need the recent generalization of this method to arbitrary von Neumann algebras by C. Houdayer and Y. Isono [@HoudIsoUPF].
The following lemma will be used a lot, without explicit reference. A proof can be found in [@TakesakiIII Lemma 5.5, Chapter $\mathrm{XIV}$].
\[corner\] Let $A$ be a subalgebra of a von Neumann algebra $M$ and $e$ a projection in $A$ or in $A' \cap M$. We have $e(A' \cap M)e=(eAe)' \cap eMe$ and $e(A \vee (A' \cap M))e=eAe \vee e(A' \cap M)e$.
\[def\_semi\_conj\] Let $M$ be a von Neumann algebra and let $A \subset 1_AM1_A$ and $B \subset 1_BM1_B$ be two subalgebras with expectations. For $h \in M$, we say that $A$ and $B$ are *semi-conjugated* by $h \in M$ if $$\forall x \in A, \; hx=0 \Rightarrow x=0$$ and there exists an (onto) $*$-isomorphism $\psi: A \rightarrow B$ such that $$\forall x \in A, \; \psi(x)h=hx.$$ We denote this relation by $A \sim_h B$. We will also write $A \sim_M B$ if there exists an $h \in M$ such that $A \sim_h B$.
\[semi-conj\] We make the following easy remarks:
- The $*$-isomorphism $\psi$ is completely determined by $h$.
- $A \sim_h B$ if and only if $B \sim_{h^*} A$.
- $A \sim_h B$ implies that $A \sim_{1_B h1_A} B$ so we can always suppose that $h \in 1_BM1_A$.
- If $h=v|h|$ is the polar decomposition of $h$ then $A \sim_h B$ implies that $A \sim_{v} B$ so we can always suppose that $h$ is a partial isometry in $1_BM1_A$.
However, we note a very important difficulty with the relation $\sim_M$: it is *not transitive*. Indeed if $A \sim_h B$ and $B \sim_g C$, then nothing guarantees that $gh \neq 0$. For this, one has to control the relative commutants of the involved subalgebras. For example, an interesting special case is when $A' \cap 1_AM1_A \subset A$ and $B' \cap 1_BM1_B \subset B$ (e.g. $A$ and $B$ maximal abelian). Indeed, in this case, if a partial isometry $v \in 1_BM1_A$ satisfies $A \sim_v B$ then we get a genuine conjugacy, i.e. $v^*v=1_A$, $vv^*=1_B$ and $B=vAv^*$. In the general case however, we only know that $e=v^*v \in A' \cap 1_AM1_A$ and $f=vv^* \in B' \cap 1_BM1_B$. Using Lemma \[corner\], we see that the isomorphism $\mathrm{Ad}(v): eMe \rightarrow fMf$ sends $Ae$ onto $Bf$ and $(Ae)' \cap eMe=e(A'\cap 1_AM1_A)e$ onto $(Bf)' \cap fBf = f(B' \cap 1_BM1_B)f$. In particular, $e(A \vee (A' \cap 1_AM1_A))e$ is conjugated to $f(B \vee (B' \cap 1_BM1_B))f$ by $v$.
Let $M$ be a von Neumann algebra and let $A \subset 1_AM1_A$ and $B \subset 1_BM1_B$ be two subalgebras with expectation. We say that *a corner of $A$ embeds with expectation into a corner of $B$ inside $M$* if there exists non-zero projections $p \in A$ and $q \in B$ and a subalgebra with expectation $C \subset qBq$ such that $pAp \sim_M C$.
We will use the notation $A \prec_M B$ to say that a corner of $A$ embeds with expectation into a corner of $B$ inside $M$ and $A \nprec_M B$ to say that no corner of $A$ embeds with expectation into a corner of $B$ inside $M$.
We note that in order to check that $A \prec_M B$ it suffices to find projections $p \in A$, $q \in B$, a normal $*$-morphism $\psi: pAp \rightarrow qBq$ and a non-zero $h \in qMp$ such that $\psi(pAp) \subset qBq$ is with expectation and $\forall x \in pAp, \psi(x)h=hx$. Indeed, in this case, there is a projection $p' \in \mathcal{Z}(pAp)$ and an injective $*$-morphism $\psi': p'Ap' \rightarrow qBq$ such that $\psi(x)=\psi'(p'x)$ for all $x \in pAp$. And since we have $h(1-p')=\psi(1-p')h=0$ we will have $h \in qMp'$ and $\forall x \in p'Ap', \psi'(x)h=hx$. Therefore, we may always suppose that $\psi$ is injective. Now, we can see $\psi$ as an $*$-isomorphism from $pAp$ to $C$ where $C \subset qBq$ is a subalgebra with expectation. However, we don’t have yet the condition $\forall x \in pAp, hx=0 \Rightarrow x=0$. To fix this, note that $\{ x \in pAp \mid hx=0 \}$ is a two-sided ideal in $pAp$ so there exists a unique projeciton $p' \in \mathcal{Z}(pAp)$ such that $hx=0 \Leftrightarrow p'x=0$. We have $p' \neq 0$ because $h \neq 0$. Let $q'=\psi(p') \in \mathcal{Z}(C)$. Let $C'=Cq' \subset q'Bq'$ which is a subalgebra with expectation. Finally let $\psi': p'Ap' \rightarrow C'$ be the restriction of $\psi$. Now we still have $\forall x \in p'Ap', \psi'(x)h=hx$ and the condition $hx=0 \Rightarrow x=0$ is satisfied for all $x \in p'Ap'$ which means that $p'Ap' \sim_h C'$. Hence we have indeed $A \prec_M B$.
The following nontrivial proposition will be used frequently in the sequel.
\[sub\_embed\] Let $M$ be a von Neumann algebra and $A \subset 1_AM1_A$, $B \subset 1_BM1_B$ two subalgebras with expectation. Let $D \subset A$ be a subalgebra with expectation. If $A \prec_M B$ then $D \prec_M B$.
The following generalization of the powerful intertwining theorem of S. Popa will be needed for Lemma \[alternative\] which is crucial for the proof of Theorem \[solid\].
\[criterion\_intertwine\] Let $M$ be any von Neumann algebra and $A \subset 1_AM1_A$, $B \subset 1_BM1_B$ two subalgebras with expectation. Suppose that $A$ is finite and choose a faithful normal conditional expectation $E_B: 1_BM1_B \rightarrow B$. Then the following are equivalent:
- $A \nprec_M B$.
- There exists a net of unitaries $(u_i)_{i \in I}$ in $\mathcal{U}(A)$ such that $$\forall x,y \in 1_BM1_A, \; E_B(xu_iy^*) \rightarrow 0$$ in the $*$-strong topology.
\[density\] A useful fact is that it is sufficient to check the condition $E_B(xu_iy^*) \rightarrow 0$ only for $x,y$ in a dense subset of $1_BM1_A$ (see [@HoudIsoUPF Theorem 4.3 (5)]). Another useful trick is that if we have a family of subalgebras with expectation $(B_j)_{j \in J}$ such that $A \nprec_M B_j$ for all $j \in J$, then we can construct a net of unitaries $(u_i)_{i \in I}$ satisfying the condition simultaneously for all the subalgebras $B_j, j\in J$. To see this, one can first reduce to the case where $J$ is finite and then apply Theorem \[criterion\_intertwine\] to $A$ and $B=\bigoplus_{j \in J} B_j$, viewed as subalgebras of $\bigoplus_{j \in J} M$. This idea goes back to [@Ioana2008 beginning of the proof of Theorem 3.4].
Finally, we consider an absorption property which was very prominent in Popa’s work on deformation/rigidity, mainly because it helps to solve the relative commutant issue explained in Remark \[semi-conj\].
Let $M$ be a von Neumann algebra and $N \subset M$ a subalgebra with expectation. We say that $N \subset M$ is *absorbing* if for every diffuse subalgebra with expectation $Q \subset 1_QN1_Q$, we have $Q' \cap 1_QM1_Q \subset 1_QN1_Q$.
Obviously, the trivial inclusions $M \subset M$ and $\mathbb{C} \subset M$ are always absorbing (since $\mathbb{C}$ has no diffuse subalgebras). As it was observed by S. Popa, more interesting examples of absorbing inclusions are given by *mixing* actions:
\[mixing\] Let $(A,\varphi)$ be a von Neumann algebra with a faithful normal state $\varphi$ and $\sigma: \Gamma \rightarrow \mathrm{Aut}(A,\varphi)$ a $\varphi$-preserving action of a discrete group $\Gamma$, let $M=A \rtimes_\sigma \Gamma$ be the crossed product von Neumann algebra. Suppose that the action $\sigma$ is *mixing*, i.e. $$\forall a,b \in A, \; \lim_{g \rightarrow \infty} \varphi(a\sigma_g(b))=\varphi(a)\varphi(b).$$
Then the inclusion $\mathcal{L}(\Gamma) \subset M$ is absorbing.
Other examples of this absorption phenomenon come from some group inclusions as well as free products [@popa1983orthogonal] and amalgamated free products [@Ioana2008 Theorem 1.1].
Later on, we will need the following lemma:
\[abs\_com\] Let $N \subset M$ be an absorbing inclusion. Let $Q \subset 1_QM1_Q$ be a diffuse subalgebra with expectation. If $Q \prec_M N$ then there exists a non-zero partial isometry $v \in M$ such that $v^*v \in Q \vee (Q' \cap 1_QM1_Q)$, $vv^* \in N$ and $v(Q \vee (Q' \cap 1_QM1_Q))v^* \subset N$. In particular, we have $Q \vee (Q' \cap 1_QM1_Q) \prec_M N$.
Take non-zero projections $p \in Q$, $q \in N$, a subalgebra with expectation $C \subset qNq$ and a partial isometry $v \in qMp$ such that $pQp \sim_v C$. Then by Remark \[semi-conj\], we know that $v^*v \in (pQp)' \cap pMp$, $vv^* \in C' \cap qMq$ and $$v(pQp \vee ((pQp)' \cap pMp))v^* \subset C \vee (C' \cap qMq)$$ We have $pQp \vee ((pQp)' \cap pMp)=p(Q \vee (Q' \cap 1_QM1_Q))p$ so that $v^*v \in Q \vee (Q' \cap 1_QM1_Q)$ and $v(Q \vee (Q' \cap 1_QM1_Q))v^* \subset C \vee (C' \cap qMq)$. Since $Q$ is diffuse, $C$ is also diffuse and because $N$ is absorbing we have $C \vee (C' \cap qMq) \subset qNq$ so that we get $vv^* \in N$ and $v(Q \vee (Q' \cap 1_QM1_Q))v^* \subset N$.
Relative solidity
=================
In [@TakaSolid], N. Ozawa introduced the notion of *solid* von Neumann algebras. It is easy to check that in the $\mathrm{II}_1$ case, the definition of solidity given in [@TakaSolid] is equivalent to the following one (see Proposition \[formulation\]).
Let $M$ be a von Neumann algebra. We say that $M$ is *solid* if every properly non-amenable subalgebra with expectation $Q \subset 1_QM1_Q$ has discrete center.
Equivalently, $M$ is *solid* if and only if every subalgebra with expectation $Q \subset M$ is a direct sum of an amenable von Neumann algebra and a family (possibly empty) of non-amenable factors. This clearly shows the analogy with the notion of *solid ergodicity* for equivalence relation discovered in [@ChifanIoanaBern] and formally introduced in [@GaboriauSurvey Definition 5.4].
In this section, we are interested in a *relative* version of solidity:
Let $M$ be a von Neumann algebra and $N \subset M$ a subalgebra with expectation. We say that $M$ is *solid relatively to* $N$ if every properly non-amenable subalgebra with expectation $Q \subset 1_QM1_Q$ such that $Q \nprec_M N$ has discrete center.
Clearly, a von Neumann algebra $M$ is solid if and only if it is solid relatively to $\mathbb{C}$. The following property justifies the terminology.
\[relative\_absolute\] Let $P \subset N \subset M$ be inclusions of von Neumann algebras with expectations. If $M$ is solid relatively to $N$ and $N$ is solid relatively to $P$ then $M$ is solid relatively to $P$. In particular, if $M$ is solid relatively to $N$ and $N$ is solid then $M$ is solid.
Take a properly non-amenable subalgebra with expectation $Q \subset 1_QM1_Q$ such that $Q$ has diffuse center. We have to show that $Q \prec_M P$. Since $M$ is solid relatively to $N$, we already know that $Q \prec_M N$. Hence there exist a non-zero projection $p \in Q$, a subalgebra with expectation $C \subset 1_CN1_C$ and a partial isometry $v \in 1_CMp$ such that $pQp \sim_v C$. Let $\psi: pQp \rightarrow C$ be the $*$-isomorphism by $v$. Let $q$ be the smallest projection in $N$ which is greater then $vv^*$ so that we have $yv=0 \Leftrightarrow yq=0$ for all $y \in N$. Since $vv^* \in C' \cap 1_CM1_C$, we have $q \in C' \cap 1_CN1_C$. Hence, we have a $*$-morphism $\phi: pQp \rightarrow Cq$ defined by $\phi(x)=\psi(x)q$. In fact, it is a $*$-isomorphism because $\psi(x)q=0 \Leftrightarrow \psi(x)v=0 \Leftrightarrow vx=0 \Leftrightarrow x=0$. Now $Cq \subset qNq$ is a sublagebra with expectation which is properly non-amenable and it has diffuse center (because it is isomorphic to $pQp$). Since $N$ is solid relatively to $P$, this implies that $Cq \prec_N P$. Hence there exists a subalgebra with expectation $D \subset 1_DP1_D$, a non-zero projection $r \in Cq$ and a partial isometry $w \in 1_DNr$ such that $rCr \sim_w D$. Let $\theta: rCr \rightarrow D$ be the $*$-isomorphism implemented by $w$. Let $p'=\phi^{-1}(r)$, let $\phi': p'Qp' \rightarrow rCr$ be the $*$-isomorphism obtained by restriction of $\phi$ to $p'Qp'$ and let $h:=wv$. Then we have a $*$-isomorphism $\alpha:=\theta \circ \phi': p'Qp' \rightarrow D$ and for all $x \in p'Qp'$ we have $$\alpha(x)h=\theta(\phi'(x))wv=w\phi'(x)v=wvx=hx$$ Moreover, if $hx=0$ then $\alpha(x)wv=0$ which means that $\alpha(x)w=\alpha(x)wq=0$ by definition of $q$. This implies that $\theta(\phi'(x))=0$ by definition of $w$ and hence $x=0$. Thus we proved that $p'Qp' \sim_h D$ which means that $Q \prec_M P$ as we wanted.
Next we present other possible formulations of relative solidity.
\[formulation\] Let $M$ be a von Neumann algebra and $N \subset M$ a subalgebra with expectation. The following are equivalent:
1. $M$ is solid relatively to $N$.
2. For every diffuse subalgebra with expectation $Q \subset 1_QM1_Q$ such that $Q' \cap 1_QM1_Q$ is non-amenable we have $Q' \cap 1_QM1_Q \prec_M N$.
3. For every non-amenable subalgebra with expectation $Q \subset 1_QM1_Q$ such that $Q' \cap 1_QM1_Q$ is diffuse we have $Q \prec_M N$.
$(1) \Rightarrow (2)$. Suppose that $Q' \cap 1_QM1_Q \nprec_M N$. Take $A \subset Q$ a diffuse abelian subalgebra with expectation (see [@HoudUedaAsympFree Lemma 2.1]). Then $P = A \vee (Q' \cap 1_QM1_Q)$ has diffuse center and $P \nprec_M N$ by Proposition \[sub\_embed\]. Hence $P$ is amenable by $(1)$. Therefore $Q' \cap 1_QM1_Q$ is also amenable.
$(2) \Rightarrow (3)$. Let $P = Q' \cap 1_QM1_Q$. Then $P$ is diffuse and since $Q$ is non-amenable and $Q \subset P' \cap 1_QM1_Q$ we have that $P' \cap 1_QM1_Q$ is non-amenable. Hence $P' \cap 1_QM1_Q \prec_M N$ by $(2)$. Therefore $Q \prec_M N$ by Proposition \[sub\_embed\].
$(3) \Rightarrow (1)$. Let $Q \subset 1_QM1_Q$ be a properly non-amenable subalgebra with expectation such that $Q$ has diffuse center. Then $Q' \cap 1_QM1_Q$ is also diffuse. Hence by $(3)$, we have $Q \nprec_M N$. Therefore $M$ is solid relatively to $N$.
Note that [@ChifanIoanaBern Theorem 2] as well as [@RemiGaussian Theorem B] and [@RemiGaussian Theorem C] are examples of relative solidity results. Relative solidity can be useful to prove fullness or primeness in situations where true solidity fails. Recall that a factor $M$ is *full* ([@ConnesAlmostPeriodic]) if for every bounded net $x_i \in M, i \in I$ such that $||[x_i,\varphi] || \rightarrow 0$ for all $\varphi \in M_*$ there exists a net $\lambda_i \in {\mathbb{C}}, i \in I$ such that $x_i - \lambda_i \rightarrow 0$ in the $*$-strong topology. A factor $M$ is *prime* if it is not of type $\mathrm{I}$ and $M \simeq P_1 {\mathbin{\overline{\otimes}}}P_2$ implies that $P_1$ or $P_2$ is of type $\mathrm{I}$.
\[prime\_full\] Let $M$ be a von Neumann algebra and $N \subset M$ a subalgebra with expectation. Suppose that $M$ is solid relatively to $N$ and the inclusion $N \subset M$ is absorbing. Let $P \subset 1_PM1_P$ be any non-amenable factor with expectation such that $P \nprec_M N$. Then $P$ is prime. If $P$ has moreover a separable predual then it is full.
Suppose that $P=P_1 {\mathbin{\overline{\otimes}}}P_2$ where $P_1$ and $P_2$ are two diffuse factors. Since $P$ is non-amenable then one of them, say $P_1$, is non-amenable. Since $M$ is solid relatively to $N$, by Proposition \[formulation\], we must have $P_2 \prec_M N$. Since $N \subset M$ is absorbing, this implies that $P=P_2 \vee (P_2'\cap P) \prec_M N$ by Lemma \[abs\_com\]. Contradiction.
Now we suppose that $P$ has separable predual and we show that $P$ is full. On the contrary, suppose that $P$ is not full. Then, by [@HoudUedaFreeRig Theorem 3.1], there exists a decreasing sequence of diffuse abelian subalgebras $Q_i \subset P$ with expectation such that $P = \bigvee_{i \in \mathbb{N}} (Q_i' \cap P)$. Suppose that for some $i$, we have $Q_i'\cap 1_PM1_P \prec_M N$. Note that $Q_i \subset Q_i'\cap 1_PM1_P$. So by Lemma \[abs\_com\], we know that there exists a non-zero partial isometry $v \in M$ such that $e=v^*v \in Q_i' \cap 1_PM1_P$, $f=vv^* \in N$ and $v(Q_i' \cap qMq)v^* \subset fNf$ with expectation. Note that for all $j \geq i$, we have $Q_j \subset Q_i \subset Q_i' \cap 1_PM1_P$ with expectation and $Q_j$ is diffuse. Hence, since $N$ is absorbing we have $v(Q'_j \cap 1_PM1_P)v^* \subset (vQ_jv^*)' \cap fMf \subset fNf$. Therefore for all $j \geq i$, we have $v(Q'_j \cap 1_PM1_P)v^* \subset fNf$. Thus $v(\bigvee_{i \in \mathbb{N}} (Q_i' \cap 1_PM1_P))v^* \subset fNf$. In particular $\bigvee_{i \in \mathbb{N}} (Q_i' \cap 1_PM1_P) \prec_M N$. Since $P \subset \bigvee_{i \in \mathbb{N}} (Q_i' \cap 1_PM1_P)$ with expectation we get $P \prec_M N$ by Proposition \[sub\_embed\]. But this is not possible by assumption on $P$. Hence we must have $Q_i' \cap 1_PM1_P \nprec_M N$ for all $i$. By relative solidity and using Proposition \[formulation\], this implies that $Q_i' \cap 1_PM1_P$ is amenable for all $i$. In particular, $Q_i' \cap P$ is amenable for all $i$. Hence $P=\bigvee_{i \in \mathbb{N}} (Q_i' \cap P)$ is also amenable. From this contradiction we conclude that $P$ is full.
Spectral gap rigidity for non-tracial von Neumann algebras
==========================================================
In this section, we prove an abstract non-tracial version of Popa’s spectral gap rigidity principle [@PopaSpectralGap Lemma 5.1] and [@PopaSpectralGap Lemma 5.2]. The idea is that in the presence of a *spectral gap* property, a subalgebra with properly non-amenable commutant will behave as a *rigid* subalgebra, making it easy to locate. This principle will be used to prove Theorem \[solid\] in the next section.
Bimodules and the spectral gap property {#bimodules-and-the-spectral-gap-property .unnumbered}
---------------------------------------
Let $M$ be a von Neumann algebra. As usual, an *$M$-$M$-bimodule* is a pair $(H,\pi_H)$ where $H$ is a Hilbert space and $\pi_H: M \otimes_{alg} M^{op} \rightarrow B(H)$ is a $*$-representation which is *binormal* (i.e. the restrictions of $\pi_H$ to $M$ and $M^{op}$ are normal). Recall that if $H,K$ are $M$-$M$-bimodules with bimodule representations $\pi_H: M \otimes_{alg} M^{op} \rightarrow B(H)$ and $\pi_K: M \otimes_{alg} M^{op} \rightarrow B(K)$ then $K$ is *weakly contained* in $H$ if and only if there exists a unital completely positive map $\Phi: B(H) \rightarrow B(K)$ such that $\Phi \circ \pi_H = \pi_K$ (Note that $\Phi$ will restrict to a morphism from the $C^*$-algebra generated by $\pi_H(M \otimes_{alg} M^{op})$ to the $C^*$-algebra generated by $\pi_K(M \otimes_{alg} M^{op})$, and conversely, such a morphism can be extended to a u.c.p map from $B(H)$ to $B(K)$ by Arveson’s theorem). Recall also that a von Neumann algebra $M$ is amenable if and only if the *identity* $M$-$M$-bimodule $L^2(M)$ is weakly contained in the *coarse* $M$-$M$-bimodule $L^2(M) {\mathbin{\overline{\otimes}}}L^2(M)$. See [@NCG Appendix B] for more information on bimodules.
Let $M \subset N$ be an inclusion of von Neumann algebras with expectation. Then any choice of a faithful normal conditional expectation $E_M: N \rightarrow M$ gives rise to an inclusions of $M$-$M$-bimodules $L^2(M) \subset L^2(N)$. We say that the inclusion $M \subset N$ is *coarse*[^2] if, for some choice of a faithful normal conditional expectation $E_M: N \rightarrow M$, the $M$-$M$-bimodule $$L^2(N) \ominus L^2(M)=\{ \xi \in L^2(N) \mid \xi \perp L^2(M) \}$$ is weakly contained in the coarse $M$-$M$-bimodule $L^2(M) {\mathbin{\overline{\otimes}}}L^2(M)$.
The following lemma is an abstract non-tracial version of an argument used in [@PopaSpectralGap Lemma 5.1]. See also [@HoudIsoStrongErg Theorem 4.1].
\[ultra\_gap\] Let $M \subset N$ be a coarse inclusion. Let $P \subset M$ be a subalgebra with expectation and suppose that $P$ is properly non-amenable. Let $\omega$ be any free ultrafilter on $\mathbb{N}$. Then we have $P' \cap N^\omega \subset M^\omega$.
Let $E_M: N \rightarrow M$ be a faithful normal conditional expectation as in the definition of a coarse inclusion. Let $E^{\omega}: N^\omega \rightarrow N$ the canonical conditional expectation and $E_{M^\omega}: N^\omega \rightarrow M^\omega$ the conditional expectation induced by $E_M$.
Now, suppose, by contradiction, that there is $Y \in P' \cap N^\omega$ with $Y \neq 0$ and such that $E_{M^\omega}(Y)=0$. We have $E_M(Y^*Y) \in P' \cap M$. Let $c \in P'\cap M$ be an element such that $q=E_M(Y^*Y)^{\frac{1}{2}}c$ is a non-zero projection in $P' \cap M$. Then $Yc \in P' \cap N^\omega$ and $E_{M^\omega}(Yc)=0$ and we have $E_M((Yc)^*(Yc))=q$. So, without loss of generality, we can directly suppose that $q=E_M(Y^*Y) \in P' \cap M$ is a non-zero projection. We will show that the $P$-$P$-bimodule $qL^2(M)$ is weakly contained in the $P$-$P$-bimodule $L^2(N) \ominus L^2(M)$. Let $V : L^{2}(M) \rightarrow L^{2}(N)$ and $W : L^2(N) \ominus L^2(M) \rightarrow L^{2}(N)$ be the inclusion of bimodules. Note that $VV^{*}=1-WW^{*}=e_M$. Pick a sequence $(y_n)_{n \in \mathbb{N}}$ representing $Y$ and define a completely positive map $$\Phi: B(L^2(N) \ominus L^2(M)) \rightarrow B(L^2(M))$$ $$T \mapsto \lim_{n \rightarrow \omega} \left( V^{*} y_n^* W T W^{*}y_n V \right) \text{ in the weak* topology.}$$ We have $$\Phi(1)=\lim_{n \rightarrow \omega} \left( V^{*} y_n^* (1-e_M) y_n V \right)=E_M(Y^{*}Y)-E_M(Y^{*})E_M(Y)=q$$ Hence $\Phi$ takes its values in $qB(L^{2}(M)q \simeq B(qL^{2}(M))$ which means that we can view $\Phi$ as a unital completely positive map from $B(L^2(N) \ominus L^2(M))$ to $B(qL^2(M))$. And since $Y \in P'\cap N^\omega$ we see that $\Phi$ preserves the $P$-$P$-bimodule representations. Hence the $P$-$P$- bimodule $qL^2(M)$ is weakly contained in $L^2(N) \ominus L^2(M)$. Since $P$ is with expectation in $M$, we have an inclusion of $P$-$P$-bimodules $L^2(P) \subset L^2(M)$. Hence the $P$-$P$-bimodule $qL^2(P)$ is weakly contained in $L^2(N) \ominus L^2(M)$. By the coarse inclusion property, this implies in particular that $qL^2(P)$ is weakly contained in the coarse $P$-$P$-bimodule. We conclude easily that $qP$ is amenable and this contradicts the assumption that $P$ is properly non-amenable.
Malleable deformations and spectral gap rigidity {#malleable-deformations-and-spectral-gap-rigidity .unnumbered}
------------------------------------------------
Symmetric malleable deformations, or *s-malleable* deformations, where introduced by S. Popa in [@PopaMalleable1] as a very useful tool for obtaining intertwining relations. We start by recalling this notion. Let $M$ be a von Neumann algebra. A *malleable deformation* of $M$ is a pair $(\widetilde{M},\theta)$ where $M \subset \widetilde{M}$ is an inclusion with expectation and $\theta: \mathbb{R} \rightarrow \mathrm{Aut}(\widetilde{M})$ is a continuous action of $\mathbb{R}$. The deformation $(\widetilde{M},\theta)$ is said to be *symmetric* if there exists $\beta \in \mathrm{Aut}(\widetilde{M})$ such that $\beta_{|M}=\mathrm{Id}$ and for all $t \in \mathbb{R}$, $\beta \circ \theta_t \circ \beta= \theta_{-t}$. We will say that a subalgebra $Q \subset M$ with expectation is *rigid relatively to* the deformation $(\widetilde{M},\theta)$ if $\theta$ converges uniformly on the unit ball of $Q$: for every $*$-strong neighborhood $\mathcal{V}$ of $0$ in $\widetilde{M}$ there exists $t_0 > 0$ such that $$\forall t \in [-t_0,t_0], \forall x \in (Q)_1, \; \theta_t(x)-x \in \mathcal{V}.$$
Now we can state the main theorem of this section. It is an abstract non-tracial version of an argument due to S. Popa [@PopaSpectralGap Lemma 5.2].
\[def\_rigidity\] Let $M$ be a von Neumann algebra, $(\widetilde{M},\theta)$ a symmetric malleable deformation of $M$ and $Q \subset M$ a subalgebra with expectation. Suppose that
- The inclusion $M \subset \widetilde{M}$ is coarse.
- $Q$ is finite.
- $Q' \cap 1_QM1_Q$ is properly non-amenable.
Then $Q$ is rigid relatively to the deformation $(\widetilde{M},\theta)$ and $Q \prec_{\widetilde{M}} \theta_1(Q)$.
We proceed in 4 steps following the lines of Popa’s original argument. In fact, only step 1 is different from the tracial case.
*<span style="font-variant:small-caps;">Step 1</span> - The subalgebra $Q$ is rigid relatively to $(\widetilde{M},\theta)$.*
Suppose that $Q$ is not relatively rigid. Then we can find a $*$-strong neighborhood of $0$ in $\widetilde{M}$ denoted by $\mathcal{V}$, a sequence $x_n \in (Q)_1$ and a sequence of reals $t_n \rightarrow 0$ such that $\theta_{t_n}(x_n)-x_n \notin \mathcal{V}$ for all $n$. Let $\omega$ be any free ultrafilter on $\mathbb{N}$. Since $Q$ is finite the sequence $(x_n)_{n \in \mathbb{N}}$ defines an element $x$ in the ultraproduct $Q^\omega \subset M^\omega \subset \widetilde{M}^\omega$. Also, since $t_n \rightarrow 0$ we have $||\varphi \circ \theta_{t_n} - \varphi || \rightarrow 0$ for all $\varphi \in \widetilde{M}_*$. Using this, we check that the automorphism of $\widetilde{M}$ defined by $$\Theta((y_n)_n)=(\theta_{t_n/2}(y_n))_n$$ preserves the ideal $\mathcal{I}^\omega(\widetilde{M})$. Hence $\Theta$ induces an automorphism of $\widetilde{M}^\omega$, still denoted $\Theta$, such that $$\Theta((y_n)^\omega)=(\theta_{\frac{t_n}{2}}(y_n))^\omega.$$ Note that the choice of $x_n$ and $t_n$ we made implies that $\Theta^2(x) \neq x$. Now, observe that $\Theta(y)=y$ for all $y \in \widetilde{M}$ because $t_n \rightarrow 0$. In particular, if we let $P:=Q' \cap 1_QM1_Q$ then we have $\Theta(P)=P$ and therefore $\Theta(P' \cap 1_Q \widetilde{M}^\omega 1_Q)=P' \cap 1_Q \widetilde{M}^\omega 1_Q$. Since $x \in P' \cap 1_Q \widetilde{M}^\omega 1_Q$, we get $\Theta(x) \in P' \cap 1_Q \widetilde{M}^\omega 1_Q$. Lemma \[ultra\_gap\] applies to the coarse inclusion $1_QM1_Q \subset 1_Q \widetilde{M}1_Q$ and shows that $$P' \cap 1_Q \widetilde{M}^\omega 1_Q \subset 1_Q M^\omega 1_Q.$$ Therefore we get $\Theta(x) \in M^\omega$. Now, choose a symmetry $\beta$ for $(\widetilde{M},\theta)$ and extend it naturally to an automorphism $\beta \in \mathrm{Aut}(\widetilde{M}^\omega)$. Then $\beta$ fixes $M^\omega$ and we have $(\beta \circ \Theta \circ \beta)(x) = \Theta^{-1}(x)$. Since $x \in M^\omega$ and $\Theta(x) \in M^\omega$, we conclude that $\Theta(x)=\Theta^{-1}(x)$. And this contradicts the fact that $\Theta^2(x) \neq x$. Therefore $Q$ is rigid relatively to the deformation $(\widetilde{M},\theta)$.
*<span style="font-variant:small-caps;">Step 2</span> - For sufficiently small $t$ there exists a non-zero $\theta_t(Q)$-$Q$-intertwiner.*
Since $Q$ is finite and with expectation, we can take $\psi$ a faithful normal state on $\widetilde{M}$ such that $Q$ is in the centralizer $\widetilde{M}^\psi$. By Step 1, $Q$ is rigid relatively to $(\widetilde{M},\theta)$. Hence we can find $t_0$ small enough so that for all $|t| \leq t_0$ we have $$\forall u \in \mathcal{U}(Q), \; \mathfrak{Re}( \psi(\theta_t(u)u^*)) \geq \frac{\psi(1_Q)}{2} > 0.$$ Now take $\mathcal{C} \subset (\widetilde{M})_1$ the weak$^*$ closed convex hull of $\{ \theta_t(u)u^* \mid u \in \mathcal{U}(Q) \}$ and let $w_t \in \mathcal{C}$ the unique element which minimizes $||w_t||_\psi$. Then, we have $w_t \in \theta_t(1_Q)\widetilde{M}1_Q$ and by the above inequality $w_t \neq 0$. Since $Q \subset \widetilde{M}^\psi$ we have $||\theta_t(u)w_tu^*||_\psi = ||w_t||_\psi$ for all $u \in \mathcal{U}(Q)$. Therefore, by the uniqueness of $w_t$ we have $$\forall x \in Q, \; \theta_t(x)w_t = w_t x.$$ So $w_t$ is indeed a non-zero $\theta_t(Q)$-$Q$-intertwiner.
*<span style="font-variant:small-caps;">Step 3</span> - If there exists a non-zero $\theta_t(Q)$-$Q$-intertwiner then there exists a non-zero $\theta_{2t}(Q)$-$Q$-intertwiner.*
Take a non-zero $\theta_t(Q)$-$Q$-intertwiner $w_t \in \theta_t(1_Q)\widetilde{M}1_Q$ so that $$\forall x \in Q, \; \theta_t(x)w_t = w_t x.$$ Take a symmetry $\beta \in \mathrm{Aut}(\widetilde{M})$ for $(\widetilde{M},\theta)$. Let $P= Q' \cap 1_QM1_Q$. Note that for all $d \in P$, the element $w_t d \beta(w_t^*)$ is a $\theta_{t}(Q)$-$\theta_{-t}(Q)$-intertwiner. Indeed, for all $x \in Q$, we have $$\theta_{t}(x)w_t d \beta(w_t^*)=w_t d x\beta(w_t^*)=w_t d \beta(xw_t^*)=w_t d \beta(w_t^*\theta_t(x))=w_t d \beta(w_t^*)\theta_{-t}(x).$$ Hence, $\theta_t(w_t d \beta(w_t^*))$ is $\theta_{2t}(Q)$-$Q$-intertwiner. Therefore, we just need to find $d$ such that $w_t d \beta(w_t^*) \neq 0$. Suppose that $w_t d \beta(w_t^*) = 0$ for all $d \in P$. Let $q \in \widetilde{M}$ be the unique projection such that $\widetilde{M}q$ is the weak$^*$ closure of the left ideal $\widetilde{M}w_tP$. Then we have $q\beta(q)=0$. However, note that $q \in P' \cap 1_Q \widetilde{M}1_Q$ (because $\widetilde{M}q$ is invariant by the right multiplication by elements of $P$ so that $qx=qxq$ for all $x \in P$). Hence by Lemma \[ultra\_gap\], we get that $q \in M$. Thus, we have $q=q\beta(q)=0$. Since $w_t \in \widetilde{M}q$, this contradicts the fact that $w_t \neq 0$ and we are done.
*<span style="font-variant:small-caps;">Step 4</span> - Conclusion.*
Take $t=\frac{1}{2^n}$ sufficiently small and choose a non-zero $\theta_t(Q)$-$Q$-intertwiner. Then build recursively non-zero $\theta_{2^k t}(Q)$-$Q$-intertwiners until $k=n$. This gives a non-zero $\theta_1(Q)$-$Q$-intertwiner as we wanted.
Bernoulli crossed products
==========================
In this section, we prove our main results using the spectral gap rigidity principle of the previous section.
Fix $(A_0,\varphi_0)$ any von Neumann algebra with a faithful normal state $\varphi_0$. Let $(A,\varphi)=(A_0,\varphi_0)^{{\mathbin{\overline{\otimes}}}\Gamma}$ be the infinite tensor product indexed by $\Gamma$ where $\Gamma$ is any infinite discrete group. Let $\sigma: \Gamma \rightarrow \mathrm{Aut}(A)$ be the *Bernoulli action* of $\Gamma$ on $A$ obtained by shifting the tensors. The crossed product von Neumann algebra $M=A \rtimes_\sigma \Gamma$ is called the *noncommutative Bernoulli crossed product* of $\Gamma$ with base $(A_0,\varphi_0)$. If $A_0 \neq \mathbb{C}$ and $\Gamma$ is infinite, it is well known that the Bernoulli action $\sigma$ is *ergodic* (i.e. the fixed point algebra $A^\sigma$ is trivial) and *properly outer* (i.e. if $\sigma_g(x)v=vx$ for all $x \in A$ and some non-zero $v \in A$ then $g=1$) and so, in this case, $M=A \rtimes_\sigma \Gamma$ is a factor (see [@Vaes2015296] for proofs of this facts). There is a canonical faithful normal conditional expectation $E_A: M \rightarrow A$ allowing us to extend $\varphi$ to a faithful normal state $\varphi$ on $M$ by the formula $\varphi \circ E_A = \varphi$. The action $\sigma$ preserves the state $\varphi$ so that $\mathcal{L}(\Gamma)$ is contained in the centralizer $M^\varphi$. An important fact for us is that the Bernoulli action is *mixing*: $$\forall a,b \in A, \; \varphi(a\sigma_g(b)) \rightarrow \varphi(a)\varphi(b) \text{ when } g \rightarrow \infty.$$ Therefore, the inclusion $\mathcal{L}(\Gamma) \subset M$ is *absorbing* (Example \[mixing\]).
Now, following [@ChifanIoanaBern], we define a symmetric malleable deformation of $M$. Let $(\widetilde{A}_0,\psi_0)=(A_0,\varphi_0) \ast ( \mathcal{L}(\mathbb{Z}),\tau)$ be the free product von Neumann algebra where $\tau$ is the Haar trace of $\mathcal{L}(\mathbb{Z})$. As before, let $(\widetilde{A},\psi) = (\widetilde{A}_0, \psi_0)^{{\mathbin{\overline{\otimes}}}\Gamma}$ be the infinite tensor product and $\widetilde{M}=\widetilde{A} \rtimes \Gamma$ the crossed product with respect to the Bernoulli action. The von Neumann algebra $\widetilde{M}$ contains $M$ with a normal conditional expectation $E_M: \widetilde{M} \rightarrow M$ such that $\psi =\varphi \circ E_M$. Now, take $v$ the canonical Haar unitary generating $\mathcal{L}(\mathbb{Z})$ and let $h \in \mathcal{L}({\mathbb{Z}})$ be a self-adjoint element such $v=e^{ih}$. For every $t \in {\mathbb{R}}$, let $v_t = e^{ith}$ and define an automorphism $\theta^0_t \in \mathrm{Aut}(\widetilde{A}_0)$ by $\theta^0_t(x)=v_txv_t^*$. Then $\theta^0_t$ induces an automorphism $\theta_t$ of $\widetilde{M}$ that fixes the elements of $\mathcal{L}(\Gamma)$ and preserves $\psi$. Moreover, the action $\theta: t\mapsto \theta_t \in \mathrm{Aut}(\widetilde{M})$ is continuous. Let $\beta_0 \in \mathrm{Aut}(\widetilde{A}_0)$ be the automorphism defined by $\beta_0(a)=a$ for all $a \in A_0$ and $\beta_0(v)=v^*$. Then $\beta_0$ induces naturally an automorphism $\beta$ of $\widetilde{M}$ such that $\beta \circ \theta_t \circ \beta =\theta_{-t}$. Therefore $(\widetilde{M},\theta)$ is indeed a symmetric malleable deformation of $M$. In fact, it is malleable *over* $\mathcal{L}(\Gamma)$ (see [@PopaSurvey Section 6.2]) meaning that we have the following *commuting square relation* for the $\psi$-preserving conditional expectations: $$E_M \circ E_{\theta_1(M)} = E_{\theta_1(M)} \circ E_M = E_{\mathcal{L}(\Gamma)}.$$ This fact is important since in many cases it allows one to obtain an intertwining relation in $M$ from an intertwining relation in $\widetilde{M}$. In our specific case we get the following dichotomy which was obtained in the tracial case by S. Popa [@PopaMalleable1 Theorem 4.1 and 4.4] and completed by A. Ioana [@IoanaWreath Theorem 3.3 and 3.6] (see also [@bourbakiHoud Theorem 7.3, step 3]). In order to generalize it to the non-tracial case we use Theorem \[criterion\_intertwine\].
\[alternative\] Let $Q \subset 1_QM1_Q$ be a finite subalgebra with expectation. If $Q \prec_{\widetilde{M}} \theta_1(M)$ then one of the following holds:
- $Q \prec_M \mathcal{L}(\Gamma)$.
- $Q \prec_M {\mathbin{\overline{\otimes}}}_F A_0$ for some finite subset $F \subset \Gamma$.
Suppose that $Q \nprec_M \mathcal{L}(\Gamma)$ and $Q \nprec_M {\mathbin{\overline{\otimes}}}_F A_0$ for every finite subset $F \subset \Gamma$. Then by Theorem \[criterion\_intertwine\] and Remark \[density\] we can find a net $u_i \in \mathcal{U}(Q)$ such that $$\forall x,y \in M, \; E_{\mathcal{L}(\Gamma)}(xu_iy^*) \rightarrow 0$$ $$\forall F \subset \Gamma \text{ finite }, \forall x,y \in M, \; E_{{\mathbin{\overline{\otimes}}}_F A_0}(xu_iy^*) \rightarrow 0$$ in the $*$-strong topology. We will contradict the assumption that $Q \prec_{\widetilde{M}} \theta_1(M)$ by showing that $$\forall x,y \in \widetilde{M}, \; E_{\theta_1(M)}(xu_iy^*) \rightarrow 0.$$ First, we can suppose that $x,y \in \widetilde{A}$ since $\mathcal{L}(\Gamma) \subset \theta_1(M)$ and $\widetilde{M}=\widetilde{A} \rtimes \Gamma$. By Remark \[density\], we can suppose that $x=\otimes_{g \in K} x_g$ and $y=\otimes_{g \in K'} y_g$ where $K,K' \subset \Gamma$ are finite subsets and $x_g,y_g \in \widetilde{A_0}$. Also the result is obvious if both $x,y \in \theta_1(A)A$ since $E_{\theta_1(M)} \circ E_M = E_{\mathcal{L}(\Gamma)}$. So we can suppose that $x$ is orthogonal to $\theta_1(A)A$ for example. Now, note that we have the following relation for all $g \in \Gamma$ $$E_{\theta_1(A)}(E_{\theta_1(M)}(xu_iy^*)u_g^*)=E_{\theta_1(A)}(xE_{{\mathbin{\overline{\otimes}}}_{K \cup gK'} A_0}(u_i u_g^*)\sigma_g(y^*)).$$ This shows first that if $g$ is outside some finite set $F \subset \Gamma$, so that the support of $x$ and $\sigma_g(y)$ are disjoint, then $$E_{\theta_1(A)}(E_{\theta_1(M)}(xu_iy^*)u_g^*)=0$$ because $x$ is orthogonal to $\theta_1(A)A$. Hence we have a finite sum $$E_{\theta_1(M)}(xu_iy^*) = \sum_{g \in F} E_{\theta_1(A)}(E_{\theta_1(M)}(xu_iy^*)u_g^*) u_g$$ and each term of this sum converges to $0$ because $E_{{\mathbin{\overline{\otimes}}}_{K \cup gK'} A_0}(u_i u_g^*) \rightarrow 0$. Therefore we have $ E_{\theta_1(M)}(xu_iy^*) \rightarrow 0 $ for all $x,y \in \widetilde{M}$ as we wanted.
In order to apply our deformation/rigidity principle, we still need to show, as in the original proof of Chifan and Ioana, that the inclusion $M \subset \tilde{M}$ is coarse, i.e. that the $M$-$M$-bimodule $L^{2}(\tilde{M}) \ominus L^{2}(M)$ is weakly contained in the coarse $M$-$M$-bimodule (see Section 4).
\[bernoulli\_gap\] Suppose that $A_0$ is amenable. Then the inclusion $M \subset \widetilde{M}$ is coarse.
We will compute the bimodule $L^2(\widetilde{M})\ominus L^2(M)$ following [@ChifanIoanaBern Lemma 5]. The computations still hold even though $\psi$ is not a trace. We will use the bimodule notation $x \xi y \in L^2(M)$ for $x,y \in M$ and $\xi \in L^2(M)$. We will denote by $\psi^{\frac{1}{2}}$ the unique cyclic vector in $L^2(M)$ such that $\langle x \psi^{\frac{1}{2}}, \psi^{\frac{1}{2}} \rangle=\psi(x)$ for all $x \in M$.
Let $\mathcal{A}_0 \subseteq A_0$ be a $\varphi_0$-orthonormal base of $A_0$ with $1 \in \mathcal{A}_0$. In $\widetilde{A}_0=A_0 \ast \mathcal{L}(\mathbb{Z})$ consider the set $$\widetilde{\mathcal{A}}_0 = \{ v^{n_0}a_1 \cdots v^{n_{k-1}}a_k v^{n_k} \mid k \geq 0, \quad n_i \in \mathbb{Z} \setminus 0, \quad a_i \in \mathcal{A}_0 \setminus 1 \}.$$
Then it is easy to check that the subspaces $A_0 \tilde{a}_0 A_0$ are pairwise $\psi_0$-orthogonal for $\tilde{a}_0 \in \widetilde{\mathcal{A}}_0$. Thus, we have $$L^2(\widetilde{A}_0) \ominus L^2(A_0) = \bigoplus_{\tilde{a}_0 \in \widetilde{\mathcal{A}}_0} \overline{A_0 \tilde{a}_0 A_0 \psi_0^{\frac{1}{2}}}.$$
Now, let $\widetilde{\mathcal{A}}$ be the set of elements of $\widetilde{A}$ of the form $$\tilde{a}=\otimes_{g \in \Gamma} \tilde{a}_g$$ where $\tilde{a}_g \in \widetilde{\mathcal{A}}_0$ for finitely many (and at least one) $g$ and $\tilde{a}_g=1$ otherwise. Then we have $$L^2(\widetilde{A}) \ominus L^2(A) = \bigoplus_{ \tilde{a} \in \widetilde{\mathcal{A}}} \overline{A \tilde{a} A\psi^{\frac{1}{2}}}.$$
Now, focus on the $M$-$M$-bimodules $H_{\tilde{a}}=\overline{M\tilde{a}M\psi^{\frac{1}{2}}} \subset L^2(\widetilde{M})$ for $\tilde{a} \in \widetilde{\mathcal{A}}$. We note that $H_{\tilde{a}}=H_{\sigma_g(\tilde{a})}$ for all $g \in \Gamma$ while $H_{\tilde{a}}$ and $H_{\tilde{a}'}$ are orthogonal when $\tilde{a}$ and $\tilde{a'}$ are not in the same $\Gamma$-orbit. So let $\Omega$ be the set of $\Gamma$-orbits of $\widetilde{\mathcal{A}}$ and for every $\pi \in \Omega$ define $$H_\pi:= H_{\tilde{a}}$$ where $\tilde{a}$ is any element of the orbit $\pi$. Then we have an $M$-$M$-bimodules decomposition $$L^2(\widetilde{M}) \ominus L^2(M) = \bigoplus_{\pi \in \Omega} H_\pi .$$ In order to conclude, it suffices to show that for each $\pi \in \Omega$, $H_\pi$ is weakly contained in $L^2(M) {\mathbin{\overline{\otimes}}}L^2(M)$. So let $\pi$ be such an orbit, represented by $\tilde{a} \in \widetilde{\mathcal{A}}$. Write $\tilde{a}=\otimes_{g \in \Gamma}\tilde{a}_g$ and let $F$ the non-empty finite set of elements $g \in \Gamma$ such that $\tilde{a}_g \neq 1$. The stabilizer of $\tilde{a}$ inside $\Gamma$ is denoted by $$S = \{ g \in \Gamma \mid \sigma_g(\tilde{a})=\tilde{a} \}.$$ It is a finite subgroup since it must leave the support $F$ invariant. Now let $$K = A_0^{{\mathbin{\overline{\otimes}}}\Gamma \setminus F} \rtimes S \subset M.$$ The von Neumann algebra $K$ is globally invariant by the modular flow of $\varphi$. So there exists a unique normal conditional expectation $E_{K}$ from $M$ to $K$ that preserves $\varphi$. We will show that there is an isomorphism of $M$-$M$-bimodules $$H_\pi=H_{\tilde{a}} \simeq L^2(\langle M, K \rangle)$$ where $\langle M, K \rangle=(J_M K J_M)' \subset B(L^2(M))$ is the basic construction. Let $e_{K} \in \langle M, K \rangle$ be the Jones projection associated to $E_K$. It is known that the $*$-subalgebra $Me_{K}M$ is dense in $\langle M, K \rangle$. Let $\hat{\varphi}$ be the unique normal faithful semi-finite weight on $\langle M, K \rangle$ which satisfies $$\forall x,y \in M, \hat{\varphi}(xe_{K}y)=\varphi(xy).$$ We have $$L^2(\langle M, K \rangle) = \overline{Me_{K}M \hat{\varphi}^{\frac{1}{2}}}.$$ Denote by $U: H_{\tilde{a}} \rightarrow L^2(\langle M, K \rangle)$ the linear map densely defined by $$U(x\tilde{a}y\psi^{\frac{1}{2}})=xe_{K}y\hat{\varphi}^{\frac{1}{2}}$$ for all $x,y \in M$ which are $\varphi$-analytic. We claim that $U$ extends to a unitary map which is a $M$-$M$-bimodule isomorphism. In fact $U$ clearly commutes with the left action, and it also commutes with the right action because for $z \in M$ analytic we have $$U(x\tilde{a}y\psi^{\frac{1}{2}}z)=U(x\tilde{a}yz'\psi^{\frac{1}{2}})=xe_{K}yz'\hat{\varphi}^{\frac{1}{2}}=xe_{K}y\hat{\varphi}^{\frac{1}{2}}z= U(x\tilde{a}y\psi^{\frac{1}{2}})z.$$ where $z' = \sigma^\varphi_{\frac{-i}{2}}( z)=\sigma^\psi_{\frac{-i}{2}}( z)=\sigma^{\hat{\varphi}}_{\frac{-i}{2}}( z)$ (Note that the modular flows of $\psi$ and $\hat{\varphi}$ coincide on $M$ with the modular flow of $\varphi$). So it only remains to check that $U$ defines indeed a unitary, i.e. that $$\psi(y_1^*\tilde{a}^*x_1^*x_2\tilde{a}y_2)=\hat{\varphi}(y_1^*e_{K}x_1^*x_2e_{K}y_2)$$ for all $x_1,x_2,y_1,y_2 \in M$ which are $\varphi$-analytic. Once again, since the modular flow of $\hat{\varphi}$ and $\psi$ coincide on $M$, we can pass $y_1^*$ to the other side $$\psi(y_1^*\tilde{a}^*x_1^*x_2\tilde{a}y_2)=\psi(\tilde{a}^*x_1^*x_2\tilde{a}y_2\sigma^\varphi_{-i}(y_1^*))$$ $$\hat{\varphi}(y_1^*e_{K}x_1^*x_2e_{K}y_2)=\hat{\varphi}(e_{K}x_1^*x_2e_{K}y_2\sigma^\varphi_{-i}(y_1^*))$$ and so we just need to check that $$\forall x,y \in M, \psi(\tilde{a}^*x\tilde{a}y)=\hat{\varphi}(e_{K}xe_{K}y)=\varphi(E_{K}(x)y).$$ In order to prove this, we can suppose, by density, that $x$ and $y$ are of the form $$x=(\otimes_{g}x_g)u_{\gamma}$$ $$y=(\otimes_{g}y_g)u_{\delta}$$ with $\gamma, \delta \in \Gamma$, $x_g,y_g \in A_0$ for all $g$ and $x_g=y_g=1$ except for finitely many $g$. We also have $\tilde{a}=\otimes_g \tilde{a}_g$ with $\tilde{a}_g \in \widetilde{\mathcal{A}}_0$ for finitely many (not zero) $g$ and $\tilde{a}_g=1$ otherwise. Recall that $\widetilde{\mathcal{A}}_0$ is our orthonormal base.
Now we compute $\psi(\tilde{a}^*x\tilde{a}y)$. First, if $\delta \gamma \neq 1$ then $\psi(\tilde{a}^*x\tilde{a}y)=0$. So suppose that $\delta=\gamma^{-1}$. Then we have $$\psi(\tilde{a}^*x\tilde{a}y)=\psi(\otimes_g (\tilde{a}_g^*x_g\tilde{a}_{\gamma^{-1}g}y_{\gamma^{-1}g}))=\prod_{g \in \Gamma} \psi_0(\tilde{a}_g^*x_g\tilde{a}_{\gamma^{-1}g}y_{\gamma^{-1}g}).$$ For this product to be non-zero, we must have $\tilde{a}_g=\tilde{a}_{\gamma^{-1}g}$ for all $g$. This means that $\sigma_\gamma(\tilde{a})=\tilde{a}$, i.e. $\gamma \in S$. In this case, the usual computation of free probability gives $$\psi(\tilde{a}^*x\tilde{a}y)=\prod_{g \in F} \varphi_0(x_g)\varphi_0(y_{\gamma^{-1}g}) \prod_{g \in \Gamma \setminus F} \varphi_0(x_g y_{\gamma^{-1}g}).$$ So we have shown that $\psi(\tilde{a}^*x\tilde{a}y)$ is given by the formula above when $\delta=\gamma^{-1} \in S$ and is equal to $0$ otherwise.
Now, in order to compute $\varphi(E_{K}(x)y)$, we just need to check the formula $$E_{K}(x)=E_{K}((\otimes_g x_g)u_{\gamma})=1_{S}(\gamma)\prod_{g \in F} \varphi_0(x_g) ( \otimes_{g \in \Gamma \setminus F} x_g ) u_\gamma$$ and we conclude that the equality $$\psi(\tilde{a}^*x\tilde{a}y)=\varphi(E_{K}(x)y)$$ is true in all cases.
Hence, we have shown that there is an isomorphism of $M$-$M$-bimodules $$H_\pi=H_{\tilde{a}} \simeq L^2(\langle M, K \rangle).$$ Finally, since $K=A_0^{{\mathbin{\overline{\otimes}}}\Gamma \setminus F} \rtimes S$ is the crossed product of an amenable von Neumann algebra by a finite group $S$ then $K$ is also amenable. Therefore, its commutant $\langle M, K \rangle$ is also amenable. In particular, this implies that $H_\pi$ is weakly contained in $L^2(M) {\mathbin{\overline{\otimes}}}L^2(M)$ as an $M$-$M$-bimodule. Since this is true for all $\pi \in \Omega$, we conclude that the $M$-$M$-bimodule $$L^2(\widetilde{M}) \ominus L^2(M) = \bigoplus_{\pi \in \Omega} H_\pi$$ is weakly contained in $L^2(M) {\mathbin{\overline{\otimes}}}L^2(M)$ as we wanted.
We first show that $M$ is solid relatively to $\mathcal{L}(\Gamma)$. Suppose we have a properly non-amenable subalgebra with expectation $Q \subset 1_QM1_Q$ such that $Z=\mathcal{Z}(Q)$ is diffuse. We have to show that $Q \prec_M \mathcal{L}(\Gamma)$. Since $\mathcal{L}(\Gamma) \subset M$ is absorbing and $Q \subset Z' \cap 1_QM1_Q$, then using Lemma \[abs\_com\] and Proposition \[sub\_embed\], we see that it is enough to show that $Z \prec_M \mathcal{L}(\Gamma)$. By Theorem \[def\_rigidity\] we know that $Z \prec_{\widetilde{M}} \theta_1(M)$. Hence by Lemma \[alternative\] we must have $Z \prec_M \mathcal{L}(\Gamma)$ or $Z \prec_M {\mathbin{\overline{\otimes}}}_F A_0$ for some finite subset $F \subset \Gamma$. So we just need to show that the case where $Z \prec_M {\mathbin{\overline{\otimes}}}_F A_0$ leads to a contradiction. Take a subalgebra with expectation $C \subset 1_C \left( {\mathbin{\overline{\otimes}}}_F A_0 \right) 1_C$ and a non-zero projection $p \in Z$ such that $Zp \sim_M C$. Note that $Z$ and $C$ are abelian so, by Remark \[semi-conj\], we know that $Z' \cap 1_QM1_Q \prec_M C' \cap 1_CM1_C$. Hence, by Proposition \[sub\_embed\], we get $Q \prec_M C' \cap 1_CM1_C$. Recall that, by assumption, $Q$ is properly non-amenable. Therefore, we just need to show that $C' \cap 1_CM1_C$ is amenable in order to get a contradiction. Let $x \in C' \cap 1_C M 1_C$. We claim that $x_g=E_{A}(xu_g^*)=0$ for $g \notin FF^{-1}$. In fact, since $C$ is abelian and diffuse (because $Z$ is abelian and diffuse), there exists a net of unitaries $u_i \in \mathcal{U}(C)$ which tends weakly to $0$. Take $g \notin FF^{-1}$. Then we have $\sigma_g(u_i) \in {\mathbin{\overline{\otimes}}}_{\Gamma \setminus F} A_0$. Hence $$\forall a,b\in A, \; E_{{\mathbin{\overline{\otimes}}}_F A_0}(a\sigma_g(u_i)b) \rightarrow 0$$ in the $*$-strong topology. Since $u_i x_g = x_g \sigma_g(u_i)$ we get $$u_iE_{{\mathbin{\overline{\otimes}}}_F A_0}(x_gx_g^*)= E_{{\mathbin{\overline{\otimes}}}_F A_0}(x_g\sigma_g(u_i)x_g^*) \rightarrow 0$$ in the $*$-strong topology. Thus $x_g=0$. Therefore we have shown that $$C' \cap 1_CM1_C \subset \sum_{g \in FF^{-1}} Au_g$$ We will show that this implies that $C' \cap 1_CM1_C$ is amenable. Let $D=C' \cap 1_CM1_C \oplus (1-1_C){\mathbb{C}}$. Let $E_D: M \rightarrow D$ be a faithful normal conditional expectation. Since $A$ is amenable, there is a conditional expectation $\Psi: B(L^2(A)) \rightarrow A$. Define a map $\Phi: B(L^2(M)) \rightarrow D$ by $$\Phi(T) = \sum_{g \in FF^{-1}} E_D(\Psi(e_ATu_g^*e_A)u_g)$$ where $e_A: L^2(M) \rightarrow L^2(A)$ is the Jones projection. Since $D \subset \sum_{g \in FF^{-1}} Au_g$, a little computation shows that $\Phi(x)=x$ for all $x \in D$. Moreover $\Phi$ is completely bounded (compositions and finite sums of completely bounded maps are still completely bounded). Therefore, using [@Pisier Corollaire 5], we have that $D$ is amenable which means that $C' \cap 1_C M 1_C$ is amenable as we wanted.
For the second part of the theorem we can apply Proposition \[relative\_absolute\] and Proposition \[prime\_full\] to get the desired conclusion. Note that when $\Gamma$ is non-amenable and $A_0 \neq \mathbb{C}$ then $M$ is a non-amenable factor and $M \nprec_M \mathcal{L}(\Gamma)$. Indeed, since $A_0 \neq {\mathbb{C}}$ then $A$ is diffuse and we can find a diffuse abelian subalgebra with expectation $B \subset A$. Let $u_i \in B$ a net of unitaries wich tends weakly to $0$. Then we have $E_{\mathcal{L}(\Gamma)}(x u_i y) \rightarrow 0$ in the $*$-strong topology for all $x,y \in M$ (by density, it suffices to check it for $x,y$ of the form $a u_g$ with $g \in \Gamma$ and $a \in A$). Therefore, by Theorem \[criterion\_intertwine\], we know that $B \nprec_M \mathcal{L}(\Gamma)$. By Proposition \[sub\_embed\], we get $M \nprec_M \mathcal{L}(\Gamma)$.
Note that the separability and countability assumptions in Theorem \[solid\] are only needed for the fullness property in Proposition \[prime\_full\]. The proof of the relative solidity in itself does not require any separability assumption.
See ([@ConnesAlmostPeriodic], Proposition 3.9) and the constructions in ([@ConnesAlmostPeriodic], Corollary 4.4) for the first part and ([@ConnesAlmostPeriodic], Theorem 5.2) for the second part. In both cases, the examples are obtained by taking Bernoulli crossed products $M=(A_0,\varphi_0) \rtimes \mathbb{F}_2$ where $\mathbb{F}_2$ is the free group on $2$ generators and $A_0$ is some non-trivial amenable algebra with separable predual. Since $\mathcal{L}(\mathbb{F}_2)$ is solid and non-amenable then by Theorem \[solid\] we know that $M$ is solid and non-amenable. If $B_0$ is a Cartan subalgebra of $A_0$, then it is not hard to check (using the fact that the Bernoulli action is properly outer) that ${\mathbin{\overline{\otimes}}}_{\mathbb{F}_2} B_0$ is again a Cartan subalgebra of $M$.
Before the proof of the last corollary, we refer to [@FeldMooreII] for the construction of the von Neumann algebra $\mathcal{L}(\mathcal{R})$ of a non-singular equivalence relation $\mathcal{R}$ and its properties.
We can suppose that $X_0$ is not a point and that $\Gamma$ is infinite (otherwise the result is obvious). Let $X=(X_0,\mu_0)^\Gamma$. Let $A_0=\mathcal{L}(\mathcal{R}_0)$ and let $\varphi_0$ be the faithful normal state on $A_0$ induced by $\mu_0$. Let $B_0=L^\infty(X_0) \subset A_0$ be the canonical Cartan subalgebra. Since $A_0 \neq \mathbb{C}$ and $\Gamma$ is infinite, the Bernoulli action of $\Gamma$ on $(A_0,\varphi_0)^{{\mathbin{\overline{\otimes}}}\Gamma}$ is properly outer. Using this, it is not hard to check that $(B_0,\varphi_0)^{{\mathbin{\overline{\otimes}}}\Gamma} \subset M=(A_0,\varphi_0)^{{\mathbin{\overline{\otimes}}}\Gamma} \rtimes \Gamma$ is again a Cartan subalgebra and that we can identify canonically the Cartan pair $(M,(B_0,\varphi_0)^{{\mathbin{\overline{\otimes}}}\Gamma})$ with the Cartan pair $(\mathcal{L}(\mathcal{R}),L^\infty(X))$.
Let $\mathcal{S} \subset \mathcal{R}$ be a subequivalence relation. Then with the preceding identification we have that $(B_0,\varphi_0)^{{\mathbin{\overline{\otimes}}}\Gamma} \subset \mathcal{L}(\mathcal{S}) \subset M$ with expectations. Since the subalgebra $(B_0,\varphi_0)^{{\mathbin{\overline{\otimes}}}\Gamma} \subset (A_0,\varphi_0)^{{\mathbin{\overline{\otimes}}}\Gamma}$ is diffuse, it is easy to check that $(B_0,\varphi_0)^{{\mathbin{\overline{\otimes}}}\Gamma} \nprec_M \mathcal{L}(\Gamma)$. Therefore $\mathcal{L}(\mathcal{S}) \nprec_M \mathcal{L}(\Gamma)$. Hence, Theorem \[solid\] applies and we get a sequence of projections $z_n \in \mathcal{Z}(\mathcal{L}(\mathcal{S}))$ with $\sum_n z_n=1$ such that $\mathcal{L}(\mathcal{S})z_0$ is amenable and $\mathcal{L}(\mathcal{S})z_n$ is a full prime factor for all $n \geq 1$. By identifying the projections $z_n$ with $\mathcal{S}$-invariant measurable subsets $Z_n \subset X$ we get the desired conclusion since fullness and primeness of $\mathcal{L}(\mathcal{S}_{|Z_n})=\mathcal{L}(\mathcal{S})z_n$ imply easily strong ergodicity and primeness of $\mathcal{S}_{|Z_n}$.
[^1]: A. Marrakchi is supported by ERC Starting Grant GAN 637601
[^2]: We thank R. Boutonnet for suggesting this name.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
In this work we present a novel unsupervised framework for hard training example mining. The only input to the method is a collection of images relevant to the target application and a meaningful initial representation, provided by pre-trained CNN. Positive examples are distant points on a single manifold, while negative examples are nearby points on different manifolds. Both types of examples are revealed by disagreements between Euclidean and manifold similarities. The discovered examples can be used in training with any discriminative loss.
The method is applied to unsupervised fine-tuning of pre-trained networks for fine-grained classification and particular object retrieval. Our models are on par or are outperforming prior models that are fully or partially supervised.
author:
- |
Ahmet Iscen$^1$ Giorgos Tolias$^1$ Yannis Avrithis$^2$ Ond[ř]{}ej Chum$^{1}$\
[$^1$VRG, FEE, CTU in Prague $^2$Inria Rennes]{}\
bibliography:
- 'egbib.bib'
title: 'Mining on Manifolds: Metric Learning without Labels'
---
The authors were supported by the MSMT LL1303 ERC-CZ grant.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'For simple hypergraphs, we characterize minimal vertex covers of maximum cardinality in terms of bouquet partial hypergraphs. We apply this to resolutions of edge ideals, giving a new combinatorial characterization on the bounds of projective dimension of monomial ideals.'
author:
- 'Nursel Erey[^1]'
bibliography:
- '<your-bib-database>.bib'
title: 'Bouquets, vertex covers and the projective dimension of hypergraphs'
---
Introduction
============
Let $\Bbbk$ be a fixed field and $S=\Bbbk[x_1,...,x_n]$. The *edge ideal* of a simple hypergraph $\mathcal{H}$ is a squarefree monomial ideal $I(\mathcal{H})\subseteq S$ given by $$I(\mathcal{H})=(x_{i_1}... x_{i_t} : \{x_{i_1},...,x_{i_t}\}\in E(\mathcal{H})).$$
A current research topic in commutative algebra is to express or bound the invariants of minimal free resolution of a simple hypergraph in terms of its combinatorial properties (see, for example, [@dao; @schweig] – [@zheng]). Many authors introduced new graph parameters and notions in this context.
A subset $C$ of vertices of $\mathcal{H}$ is called a *vertex cover* of $\mathcal{H}$ if every edge in $\mathcal{H}$ contains an element of $C$. A vertex cover $C$ is *minimal* if no proper subset of $C$ is a vertex cover of $\mathcal{H}$. We write $\alpha_0^\prime (\mathcal{H})$ for the maximum possible cardinality of a minimal vertex cover of $\mathcal{H}$. There is a one-to-one correspondence between the minimal vertex covers of $\mathcal{H}$ and the *minimal prime ideals* of $I(\mathcal{H})$ given by $$C \text{ is a minimal vertex cover of } \mathcal{H} \Leftrightarrow (x_i : i\in C) \text{ is a minimal prime ideal of } I(\mathcal{H}).$$ Therefore the parameter $\alpha_0^\prime (\mathcal{H})$ coincides with the *big height* of $I(\mathcal{H})$, which is the maximum height of the minimal prime ideals of $I(\mathcal{H})$. It is known that the big height of $I(\mathcal{H})$ is a lower bound for the projective dimension $\operatorname{pd}(\mathcal{H})$ of $\mathcal{H}$ (which is the length of the minimal free resolution of $S/I(\mathcal{H}))$, see, for example, Corollary 3.33 of [@morey; @villarreal] for a proof. In fact, it is a sharp bound in the following cases:
For a simple hypergraph $\mathcal{H}$, the equality $\operatorname{pd}(\mathcal{H})=\alpha_0^\prime (\mathcal{H})$ holds if
$(a)$ [@morey; @villarreal; @faridi] $S/I(\mathcal{H})$ is sequentially Cohen-Macaulay.
$(b)$ [@dao; @schweig; @francisco; @van; @tuyl; @c-5; @free; @Khosh; @Moradi; @kimura] $\mathcal{H}$ is a chordal graph.
$(c)$ [@khosh; @moradi] $\mathcal{H}$ is a vertex decomposable graph.
In this work, we will give a new characterization of $\alpha_0^\prime (\mathcal{H})$, or equivalently, the big height of $I(\mathcal{H})$, in terms of bouquet partial hypergraphs of $\mathcal{H}$.
Definitions
===========
A *simple hypergraph* $\mathcal{H}$ on a finite set $V(\mathcal{H})$ is a family $\mathcal{H}=(\mathcal{E}_1,...,\mathcal{E}_d)$ of subsets of $V(\mathcal{H})$ such that
- $\mathcal{E}_i\neq \emptyset$ for all $i=1,...,d$
- $\cup_{i=1}^d\mathcal{E}_i=V(\mathcal{H})$
- $\mathcal{E}_i\subseteq \mathcal{E}_j \Longrightarrow i=j$.
The elements of $V(\mathcal{H})$ are called the *vertices* and the sets $\mathcal{E}_1,...,\mathcal{E}_d$ are the *edges* of $\mathcal{H}$. We write $E(\mathcal{H})$ for the set of edges of $\mathcal{H}$. If every edge of a simple hypergraph consists of two elements, then it is called a *simple graph*.
A simple hypergraph $\mathcal{K}$ is said to be a *partial hypergraph* of $\mathcal{H}$ if $\mathcal{K}=(\mathcal{E}_j : j\in J)$ for some $J\subseteq \{1,...,d\}$. For a set $A\subseteq V(\mathcal{H})$, we define the *partial hypergraph of* $\mathcal{H}$ *on* $A$ as $\mathcal{H}|_A=(\mathcal{E} \in E(\mathcal{H}) : \mathcal{E}\subseteq A )$. We say that $A$ is *independent* in $\mathcal{H}$ if $\mathcal{E} \nsubseteq A$ for any $\mathcal{E}\in E(\mathcal{H})$.
(Compare to Definition 1.7 [@zheng]) A *bouquet* is a simple hypergraph ${\mathcal{B}}=(\mathcal{E}_1,...,\mathcal{E}_d)$ together with a designated set of *flowers* $F(\mathcal{B})=\{\ell_1,...,\ell_d\}$ such that
- $ \cap_{i=1}^d\mathcal{E}_i \neq \emptyset $
- $\ell_i \in \mathcal{E}_j \Leftrightarrow i=j $.
Observe that one can designate flowers to a simple hypergraph in different ways to make it a bouquet. However if the bouquet $\mathcal{B}$ is a simple graph with at least two edges, then its flowers are automatically determined.
(v1) at (0,0) ; (v2) at (-4,4) ; (v3) at (0,5) ; (v4) at (4,4) ; (v5) at (3,2.3) ; (v6) at (1,2) ; (v7) at (1.3,1.5) ; (v8) at (-3,1.8) ; (v9) at (0.2, 0.6) ; (v10) at (-1.2, 1.5) ; (v11) at (-2.7, 2.7) ;
($(v1)+(0,-0.7)$) to\[out=180,in=180\] ($(v3) + (0,1)$) to\[out=0,in=0\] ($(v1) + (0,-0.7)$);
($(v1)+(0,-0.7)$) to\[out=180,in=180\] ($(v4) + (0,1)$) to\[out=0,in=0\] ($(v1) + (0,-0.7)$); ($(v1)+(0,-0.7)$) to\[out=180,in=180\] ($(v2) + (-1,1)$) to\[out=0,in=0\] ($(v1) + (0,-0.7)$)
;
ǐn [1,2,...,11]{} [ (v) circle (0.1); ]{}
(v2) circle (0.1) node \[left\] [$\ell_1$]{}; (v3) circle (0.1) node \[left\] [$\ell_2$]{}; (v4) circle (0.1) node \[left\] [$\ell_3$]{};
If a partial hypergraph $\mathcal{B}$ of $\mathcal{H}$ is a bouquet, then we simply say that $\mathcal{B}$ is a bouquet of $\mathcal{H}$. Suppose that $\bm{\mathcal{B}}=\{\mathcal{B}_1,...,\mathcal{B}_j\}$ is a set of bouquets of $\mathcal{H}$. Then we set $$F(\bm{\mathcal{B}}):= \bigcup_{i=1}^{j}F(\mathcal{B}_i), \ \ E(\bm{\mathcal{B}}):= \bigcup_{i=1}^{j}E(\mathcal{B}_i) \text{ and } V(\bm{\mathcal{B}}):= \bigcup_{i=1}^{j}V(\mathcal{B}_i).$$
We call $F(\bm{\mathcal{B}}), E(\bm{\mathcal{B}}) $ and $V(\bm{\mathcal{B}})$ the *flower set*, the *edge set* and the *vertex set* of $\bm{\mathcal{B}}$ respectively.
In [@kimura] Kimura introduced the notion of semi-strongly disjoint bouquets for simple graphs in order to study Betti numbers of chordal graphs. The following is a natural generalization of semi-strongly disjoint bouquets to hypergraphs.
(Compare to Definition 5.1, [@kimura]) \[semi strongly disjoint hypergraph\] A set $\bm{\mathcal{B}}=\{\mathcal{B}_1,\mathcal{B}_2,...,\mathcal{B}_j\}$ of bouquets of a simple hypergraph $\mathcal{H}$ is said to be *semi-strongly disjoint* in $\mathcal{H}$ if the following conditions hold.
- If $\mathcal{E}\in E(\mathcal{B}_p)$ then $\mathcal{E}\cap F(\mathcal{B}_q)=\emptyset$ for all $q\neq p$.
- $V(\bm{\mathcal{B}})\setminus F(\bm{\mathcal{B}})$ is independent in $\mathcal{H}$.
And we set $$d_{\mathcal{H}}^\prime= \operatorname{max}\{|E(\bm{\mathcal{B}})| : \bm{\mathcal{B}} \text{ is a semi-strongly disjoint set of bouquets of } \mathcal{H}\}.$$
Proof of the theorem
====================
\[l: max cardinality min vertex cover number\] Let $\mathcal{H}$ be a simple hypergraph. If $\mathcal{K}=\mathcal{H}|_U$ for some $U\subseteq V(\mathcal{H})$, then any minimal vertex cover of $\mathcal{K}$ can be extended to a minimal vertex cover of $\mathcal{H}$. In particular, $\alpha_0^\prime (\mathcal{K})\leq \alpha_0^\prime (\mathcal{H})$.
Suppose that $V(\mathcal{H})\setminus V(\mathcal{K})=A$ and $C$ is a minimal vertex cover of $\mathcal{K}$. If $A=\emptyset$ then $\mathcal{K}=\mathcal{H}$ and there is nothing to prove. So we assume that $A\neq \emptyset$. Clearly $A\cap C=\emptyset$ and $A\cup C$ covers $\mathcal{H}$. By removing the redundant elements from $A \cup C$, one can get a minimal vertex cover $C' \subseteq A\cup C$ of $\mathcal{H}$. But then $C'\setminus A$ is a minimal vertex cover of $\mathcal{K}$. Since $C'\setminus A\subseteq C$ we get $C'\setminus A=C$ by minimality of $C$. Thus $C\subseteq C'$ is the desired extension.
The Lemma above is not necessarily true if $\mathcal{K}$ is an arbitrary partial hypergraph of $\mathcal{H}$. See for example Figures \[fig:minipage1\] and \[fig:minipage2\].
\[scale=0.8, vertices/.style=[draw, fill=black, circle, inner sep=1.7pt]{}\] (a) at (-2,1) ; (b) at (-1,0) ; (c) at (-2,-1) ; (d) at (2,1) ; (e) at (1,0) ; (f) at (2,-1) ; /in [a/b,b/c,e/d,e/f]{} ()–();
\[scale=0.8, vertices/.style=[draw, fill=black, circle, inner sep=1.7pt]{}\] (a) at (-2,1) ; (b) at (-1,0) ; (c) at (-2,-1) ; (d) at (2,1) ; (e) at (1,0) ; (f) at (2,-1) ; /in [a/b,b/c,e/d,e/f, b/e]{} ()–();
We now prove the main result of this paper.
\[l: equality max min vertex cover versus max semi-strongly disjoint\] For any simple hypergraph $\mathcal{H}$, the flower set of a semi-strongly disjoint set of bouquets of $\mathcal{H}$ can be extended to a minimal vertex cover of $\mathcal{H}$. Conversely, for any minimal vertex cover $C$ of $\mathcal{H}$, there exists a semi-strongly disjoint set of bouquets of $\mathcal{H}$ with the flower set $C$. In particular, the equality $\alpha_0^\prime (\mathcal{H})= d_\mathcal{H}^\prime $ holds.
First suppose that $\bm{\mathcal{B}}=\{\mathcal{B}_1,...,\mathcal{B}_j\}$ is a semi-strongly disjoint set of bouquets of $\mathcal{H}$. Consider $\mathcal{H}|_{V(\bm{\mathcal{B}})}$, the partial hypergraph of $\mathcal{H}$ on $V(\bm{\mathcal{B}})$. Then $F(\bm{\mathcal{B}})$ is a vertex cover of $\mathcal{H}|_{V(\bm{\mathcal{B}})}$ by independence of $V(\bm{\mathcal{S}})\setminus F(\bm{\mathcal{S}})$ and it is minimal by condition $(1)$ of Definition \[semi strongly disjoint hypergraph\]. Thus, by Lemma \[l: max cardinality min vertex cover number\] the first part of the given statement is verified.
Next, suppose that $C$ is a minimal vertex cover of $\mathcal{H}$. We will construct a set $\bm{\mathcal{B}}$ of semi-strongly disjoint bouquets of $\mathcal{H}$ such that $ F(\bm{\mathcal{B}})= C$. Let $C_0=\{c\in C : \{c\} \text{ is an edge of } \mathcal{H} \}$. Let $\bm{\mathcal{B}_0}=\{\mathcal{H}|_{\{c\}} : c\in C_0\}$ be a set of bouquets of $\mathcal{H}$. Now if $C=C_0$, then clearly $\bm{\mathcal{B}_0}=\bm{\mathcal{B}}$ is the desired semi-strongly disjoint bouquets as $V(\bm{\mathcal{B}_0})\setminus F(\bm{\mathcal{B}_0})=\emptyset$ is independent in $\mathcal{H}$. So we assume that $C_1=C\setminus C_0 \neq \emptyset$. Now we will construct another semi-strongly disjoint set $\bm{\mathcal{B}_1}$ of bouquets such that $\bm{\mathcal{B}_0} \cup \bm{\mathcal{B}_1}=\bm{\mathcal{B}}$.
First note that by minimality of $C$, for every $v\in C$ there exists an edge $\mathcal{E}_v$ of $\mathcal{H}$ such that $\mathcal{E}_v \cap C =\{v\} $. Pick an element $\ell_{1}^1\in C_1$. Then there exists an edge $\mathcal{E}_{1}^1$ of $\mathcal{H}$ such that $C_1 \cap \mathcal{E}_{1}^1=\{\ell_{1}^1\}$. As $\ell_{1}^1\notin C_0$ there exists $r_1\in \mathcal{E}_{1}^1\setminus\{\ell_{1}^1\}$. Suppose that $\ell_{1}^1,\ell_{2}^1,...,\ell_{d_1}^1$ are the elements of $C_1$ that satisfy the property $$\label{property for basis case}\text{there exists } \mathcal{E}_{i}^1\in E(\mathcal{H}) \text{ such that } \mathcal{E}_{i}^1\cap C_1 =\{\ell_{i}^1\} \text{ and } r_1\in \mathcal{E}_{i}^1$$ for every $1\leq i\leq d_1$. Let $\mathcal{E}_1^1,...,\mathcal{E}_{d_1}^1$ be fixed edges that satisfy the property above. Consider the partial hypergraph $\mathcal{B}_1=(\mathcal{E}_{1}^1,...,\mathcal{E}_{d_1}^1)$ of $\mathcal{H}$ with the designated flowers $\ell_{1}^1,...,\ell_{d_1}^1$.
Now if $F(\mathcal{B}_1)=C_1$, then $\bm{\mathcal{B}_1}=\{\mathcal{B}_1\}$ and we are done. Otherwise we keep constructing new bouquets inductively as follows. Suppose that we have semi-strongly disjoint bouquets {$\mathcal{B}_1,...,\mathcal{B}_t$} such that $\cup_{i=1}^tF(\mathcal{B}_i)$ is a proper subset of $C_1$. Pick an element $\ell_{1}^{t+1}\in C_1\setminus\cup_{i=1}^tF(\mathcal{B}_i)$ and an edge $\mathcal{E}_{1}^{t+1}$ of $\mathcal{H}$ such that $\mathcal{E}_1^{t+1}\cap C_1=\{\ell_{1}^{t+1}\}$. Fix $r_{t+1}\in\mathcal{E}_1^{t+1}\setminus \{\ell_1^{t+1}\}$ and let $\ell_{1}^{t+1},\ell_{2}^{t+1},...,\ell_{d_{t+1}}^{t+1}$ be the elements of $C_1 \setminus \cup_{i=1}^tF(\mathcal{B}_i) $ that satisfy the property $$\label{property for inductive case}\text{there exists } \mathcal{E}_{i}^{t+1}\in E(\mathcal{H}) \text{ such that } \mathcal{E}_{i}^{t+1}\cap C_1 =\{\ell_{i}^{t+1}\} \text{ and } r_{t+1}\in \mathcal{E}_{i}^{t+1}$$ for every $1\leq i\leq d_{t+1}$. Let $\mathcal{E}_{1}^{t+1},...,\mathcal{E}_{d_{t+1}}^{t+1}$ be fixed edges that satisfy the property above. Consider the partial hypergraph $\mathcal{B}_{t+1}=(\mathcal{E}_{1}^{t+1},...,\mathcal{E}_{d_{t+1}}^{t+1})$ of $\mathcal{H}$ as a bouquet with flowers $\ell_{1}^{t+1},\ell_{2}^{t+1},...,\ell_{d_{t+1}}^{t+1}$. Now we check if $\bm{\mathcal{B}_0}\cup\{\mathcal{B}_1,...,\mathcal{B}_{t+1}\}$ is semi-strongly disjoint. Condition $(1)$ of Definition \[semi strongly disjoint hypergraph\] clearly holds by construction. To see that the second condition holds, observe that $$V(\bm{\mathcal{B}_0}\cup\{\mathcal{B}_1,...,\mathcal{B}_{t+1}\})\setminus F(\bm{\mathcal{B}_0}\cup\{\mathcal{B}_1,...,\mathcal{B}_{t+1}\}) = V(\bm{\mathcal{B}_0}\cup\{\mathcal{B}_1,...,\mathcal{B}_{t+1}\})\setminus C$$ is not independent in $\mathcal{H}$ since $C$ is a vertex cover, so every edge intersects $C$.
Having verified that this construction yields semi-strongly disjoint bouquets at every step, we know that it will terminate as $\mathcal{H}$ has finitely many vertices. In that case, $C_1=\cup_{i=1}^pF(\mathcal{B}_i)$ for some $p\geq 1$ and $\bm{\mathcal{B}_1}=\{\mathcal{B}_1,...,\mathcal{B}_p\}$ is as desired.
The following is an immediate consequence of Theorem \[l: equality max min vertex cover versus max semi-strongly disjoint\].
For a simple hypergraph $\mathcal{H}$ we have the followings.
- If $\bm{\mathcal{B}}$ is a semi-strongly disjoint set of bouquets of $\mathcal{H}$ such that $\left \vert F(\bm{\mathcal{B}})\right \vert =d_\mathcal{H}^\prime$, then $F(\bm{\mathcal{B}})$ is a minimal vertex cover of $\mathcal{H}$ of maximum cardinality.
- If $C$ is a minimal vertex cover of $\mathcal{H}$ of maximum cardinality, then there exist a semi-stronly disjoint set $\bm{\mathcal{B}}$ of bouquets of $\mathcal{H}$ such that $F(\bm{\mathcal{B}})=C$ and $\left\vert F(\bm{\mathcal{B}})\right \vert = d_\mathcal{H}^\prime . $
Note that for a simple graph $G$ the inequality $\alpha_0^\prime (G)\geq d_G^\prime$ is not unknown, see, for example, Proposition 2.7 of [@c-5; @free; @Khosh; @Moradi] for another proof. Also $\alpha_0^\prime (G)= d_G^\prime$ was known for the special case of vertex decomposable graphs (see, Theorem 3.8 of [@khosh; @moradi]).
Lastly, we state the following Corollary regarding projective dimension of edge ideals.
If $\mathcal{H}$ is a simple hypergraph, then $\operatorname{pd} S/I(\mathcal{H}) \geq d_\mathcal{H}^\prime$.
By Corollary 3.33 of [@morey; @villarreal], we have $\operatorname{pd} S/I(\mathcal{H}) \geq \alpha_0^\prime (\mathcal{H}). $ Hence the proof is immediate from Theorem \[l: equality max min vertex cover versus max semi-strongly disjoint\].
0.4in [**Acknowledgments:**]{} I would like to thank my supervisor Sara Faridi for very useful discussions which improved the presentation of this paper. I also thank Fahimeh Khosh-Ahang Ghasr for pointing out some flaws in the early version of this work.
[10]{}
C. Berge, *Hypergraphs*, North-Holland 1989. H. Dao, J. Schweig, *Projective dimension, graph domination parameters, and independence complex homology*, J. Combin. Theory Ser. A 120 (2013), no. 2, 453–469.
S. Faridi, *The projective dimension of sequentially Cohen-Macaulay monomial ideals*, math.AC/1310.5598, 2013.
C. A. Francisco, A. Van Tuyl, *Sequentially Cohen-Macaulay edge ideals*, Proc. Amer. Math. Soc. 135 (2007), 2327–2337.
H. T. Hà, A. Van Tuyl, *Monomial ideals, edge ideals of hypergraphs, and their graded Betti numbers*, J. Algebraic Combin. 27 (2008), 215–245.
F. Khosh-Ahang, S. Moradi, *Codismantlable hypergraphs, projective dimension and regularity of edge ideal of special hypergraphs*, math.AC/1305.5954v1, 2013.
F. Khosh-Ahang, S. Moradi, *Regularity and projective dimension of edge ideal of $C_5$-free vertex decomposable graphs*, to appear in Proc. Amer. Math. Soc. K. Kimura, *Non-vanishingness of Betti numbers of edge ideals*, Harmony of Gröbner bases and the modern industrial society, 153–168, World Sci. Publ., Hackensack, NJ, 2012. S. Morey, R.H. Villarreal, *Edge ideals: algebraic and combinatorial properties*, Progress in commutative algebra 1, 85–126, de Gruyter, Berlin, 2012. X. Zheng, *Resolutions of facet ideals*, Comm. Algebra 32 (2004) 2301–2324.
[^1]: Department of Mathematics and Statistics, Dalhousie University, Halifax, Nova Scotia, Canada B3H 4R2 e-mail: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We assess and develop techniques to remove contaminants when calculating the 3D galaxy power spectrum. We separate the process into three separate stages: (i) removing the contaminant signal, (ii) estimating the uncontaminated cosmological power spectrum, (iii) debiasing the resulting estimates. For (i), we show that removing the best-fit contaminant ([ *mode subtraction*]{}), and setting the contaminated components of the covariance to be infinite (*mode deprojection*) are mathematically equivalent. For (ii), performing a Quadratic Maximum Likelihood (QML) estimate after *mode deprojection* gives an optimal unbiased solution, although it requires the manipulation of large $N_{\rm mode}^2$ matrices [ ($N_{\rm mode}$ being the total number of modes)]{}, which is unfeasible for recent 3D galaxy surveys. Measuring a binned average of the modes for (ii) as proposed by @Feldman [FKP] is faster and simpler, but is sub-optimal and gives rise to a biased solution. We present a method to debias the resulting FKP measurements that does not require any large matrix calculations. We argue that the sub-optimality of the FKP estimator compared with the QML estimator, caused by contaminants is less severe than that commonly ignored due to the survey window.'
author:
- |
B. Kalus$^{1}$[^1], W. J. Percival$^{1}$, D. J. Bacon$^{1}$ and L. Samushia$^{2,1,3}$\
$^{1}$Institute of Cosmology & Gravitation, Dennis Sciama Building, University of Portsmouth, Portsmouth, PO1 3FX, UK\
$^{2}$Department of Physics, Kansas State University, 116, Cardwell Hall, Manhattan, KS, 66506, USA\
$^{3}$National Abastumani Astrophysical Observatory, Ilia State University, 2A Kazbegi Ave., GE-1060 Tbilisi, Georgia
bibliography:
- 'systpap2.bib'
date: 'Accepted . Received ; in original form '
title: Unbiased contaminant removal for 3D galaxy power spectrum measurements
---
\[firstpage\]
methods: statistical – cosmology: large-scale structure of Universe.
Introduction
============
Galaxy surveys provide a rich store of information about the nature of the Universe, allowing us to constrain cosmological models with baryon acoustic oscillations (BAO), gravitational models with redshift space distortions (RSD) and inflationary models with primordial non-Gaussianity. A basic statistic containing large-scale structure information is the galaxy power spectrum $P(k)$, which is the 2-point function of the Fourier transformed density field. Future large-scale structure surveys, such as the Dark Energy Spectroscopic Instrument survey [@Schlegel:2011zz; @Levi:2013gra DESI], Euclid [@Laureijs:2011gra][^2] and the Square Kilometre Array (SKA) [^3], will probe larger volumes, therefore allowing us to measure more Fourier modes of the galaxy density field.
The observed galaxy field can be contaminated with fluctuations of non-cosmological origin, such as variations due to the galactic extinction and the stellar density. Often the contaminants are not known exactly (e.g. we may know the shape of the spurious mode but may not know its exact amplitude) which makes their exact removal impossible. These modes have the potential to strongly bias cosmological constraints derived from the clustering measurements, so we need to correct or suppress these misleading modes in a responsible way.
We now introduce the basic mathematical problem that we wish to solve and introduce the main methods of removing contaminants discussed in literature. We assume that we have measured the galaxy density field as real numbers in configuration space, which we (fast) Fourier transform to obtain a Hermitian density field $F({{\bmath{k}}})$. Furthermore, we assume that the contamination can be described by another Hermitian field $f({{\bmath{k}}})$, such that the true density field is given by $$D({{\bmath{k}}})=F({{\bmath{k}}})-{ \varepsilon_\mathrm{true}}f({{\bmath{k}}}),
\label{eq:Fassum}$$ [ with $\varepsilon_\mathrm{true}$ unknown.]{} In cases with multiple contaminants (which we label with capital Latin indices), we extend Eq. to $$D({{\bmath{k}}})=F({{\bmath{k}}})-\sum_A{ \bvarepsilon^\mathrm{(true)}_{A}} f_A({{\bmath{k}}}).$$ Furthermore, we assume that $F({{\bmath{k}}})$ and $f({{\bmath{k}}})$ are uncorrelated, which is a valid assumption for most sources of systematics since they originate from our Galaxy or due to telescope effects. Large scale surveys will reduce the current sample variance limitation on the power spectrum on scales where the systematic errors have a significant impact. As a consequence, having control of these systematics is a key requirement to provide accurate cosmological measurements.
In order to investigate techniques for estimating the power spectrum in the presence of contaminants, we separate the process into three separate stages: (i) removing the contaminant signal, (ii) estimating the uncontaminated cosmological power spectrum, (iii) debiasing the resulting estimates. Two techniques are in common usage for removing the contaminant signal (i): The first is [ *mode subtraction*]{} (cf. Sec. \[sec:systrem\] and Sec. \[sec:debias\]), where contaminants are removed by fitting the amplitude of the contaminant field $f({{\bmath{k}}})$ to the data and simply subtracted off from $F({{\bmath{k}}})$. The second is *mode deprojection* [@Rybicki], which is based on assigning infinitely large covariances to contaminated modes, thus removing them from any analysis. [ In our nomenclature, a mode is a linear combination of Fourier modes rather than a single ${{\bmath{k}}}$-mode. This is reflected in the naming of *mode subtraction* and *mode deprojection*. This choice of names shall distinguish the *mode subtraction* technique from a third technique for removing the contaminant signal, called *template subtraction*, where the observed power spectra are corrected using best-fit amplitudes derived via cross-correlations between the data and the templates. @Elsner have shown that this method provides a biased estimate of the power and we will not consider it further in this article.]{} For (ii), the power spectrum P(k) is commonly estimated by the FKP estimator [@Feldman], which is an approximation to the Quadratic Maximum Likelihood (QML) estimator [@Tegmark:1997rp]. As well as being optimal, the QML estimator has the advantage of producing unbiased power spectrum estimates. However, when applying this methodology [ to data with $N_\mathrm{mode}$ modes]{}, one has to calculate, for each bin, a $N_\mathrm{mode} \times N_\mathrm{mode}$ matrix, [ and then, after binning the data into $N_\mathrm{bin}$ bins,]{} an overall $N_\mathrm{bin} \times N_\mathrm{bin}$ normalisation matrix, which makes the application of this methodology unfeasible for future surveys with increased number of modes $N_\mathrm{mode}$. In this work, we suggest a modified FKP-style [ *mode subtraction*]{} approach. We show that this technique can be made unbiased and, on a mode-by-mode basis, is mathematically identical to *mode deprojection*. The FKP estimator with debiased [ *mode subtraction*]{} is not optimal in that it discards more information than the full QML estimator, but we expect that, in realistic cases, this loss of information will be small.
The outline of this paper is as follows: We provide an introduction to power spectrum estimation in Sec. \[sec:notation\], introducing the QML and FKP estimators. We introduce the systematics removal techniques, *mode deprojection* and [ *mode subtraction*]{}, in sections \[sec:BMP\] and \[sec:systrem\], respectively, and we show that before normalisation their resulting power spectra are the same. These are extended to multiple contaminants in Appendices \[sec:MPmultderiv\] &\[sec:TSmultideriv\], respectively. We introduce a new normalisation factor in Sec. \[sec:debias\] for a single contaminant and compare it to the normalisation of the quadratic maximum likelihood (QML) estimator of @Tegmark:1997rp. This derivation is extended to allow for a non-diagonal covariance in Appendix \[sec:nondiag\]. We show that we can apply our methodology also to multiple contaminants in Sec. \[sec:multisyst\] and test the different methods on simulations in Sec. \[sec:eg\]. We conclude in Sec. \[sec:conclusion\].
Power spectrum estimators {#sec:notation}
=========================
In this section, we review two basic power spectrum estimators: the *quadratic maximum likelihood* (QML) estimator [@Tegmark:1997rp] and the simplified FKP estimator [@Feldman], to which QML reduces in the limit of uncorrelated modes with equal noise per mode in each bin. Even without considering any contaminants, the FKP estimator is easier to implement and is used in most recent analyses of large-scale structure, while the QML estimator is optimal but difficult to implement especially on smaller scales.
The *quadratic maximum likelihood* (QML) estimator [@Tegmark:1997rp] is given by $$\widehat P(k_i)=\sum_{j}{\mathbf{N}}_{ij}{^{-1}}\mathbf{p}_j,
\label{eq:QuadEst}$$ where the power is a convolution of the inverse of a normalisation matrix ${\mathbf{N}}_{ij}$ and a weighted two-point function $$\mathbf{p}_j\equiv \sum_{\alpha, \beta} F^\ast({{\bmath{k}}}_\alpha) {\mathbf{E}}_{\alpha\beta}(k_j) F({{\bmath{k}}}_\beta).
\label{eq:pundef}$$ The weight is given by the estimator matrix $${\mathbf{E}}(k_j)=-\frac{\partial {\mathbf{C}^{-1}}}{\partial P(k_j)},$$ which describes how the inverse of the density field covariance matrix ${\mathbf{C}}$ changes with respect to the prior of the power spectrum of the respective bin. If the QML normalisation is proportional to the Fisher information, i.e. $$\begin{aligned}
{\mathbf{N}}_{ij}&={\operatorname{tr}}\left\lbrace {\mathbf{C}^{-1}}\frac{\partial{\mathbf{C}}}{\partial P(k_i)}{\mathbf{C}^{-1}}\frac{\partial{\mathbf{C}}}{\partial P(k_j)}\right\rbrace,
\label{eq:Ndef}\end{aligned}$$ the QML estimator is the optimal maximum likelihood estimator of the variance of a field that obeys a multivariate Gaussian distribution [@Tegmark:1997rp]. Assuming a Gaussian density field $D({{\bmath{k}}})$, the QML estimator therefore provides an estimate of the power spectrum with minimal errors.
Under the assumption that all modes are independent, the covariance of the density field is given by the power spectrum (and the Kronecker delta $\delta_{\mu\nu}$): $${\mathbf{C}}_{\mu\nu}=\delta_{\mu\nu}P(k_{\mu}).
\label{eq:DiagCov}$$ We assume for the derivative of ${\mathbf{C}}$ with respect to $P(k_i)$ that it is unity if the modes ${{\bmath{k}}}_\alpha$ and ${{\bmath{k}}}_\beta$ are equal and contained in the bin $\Bbbk_i$, and zero otherwise, which we write using the Heaviside function $\Theta$ as: $$\frac{\partial{\mathbf{C}}_{\alpha\beta}}{\partial P(k_i)}=\delta_{\alpha\beta}\Theta({{\bmath{k}}}_\alpha\in \Bbbk_i)\equiv\delta_{\alpha\beta}\Theta_{\alpha i}.
\label{eq:Cderiv}$$ Given Eq. and , we find $$\begin{aligned}
{\mathbf{E}}_{\alpha\beta}(k_j)=\frac{\delta_{\alpha\beta}}{P^2(k_\alpha)}\Theta_{\alpha j}\end{aligned}$$ and $$\begin{aligned}
{\mathbf{N}}_{ij}=\frac{N_{\Bbbk_i}}{P^2(k_i)}\delta_{ij},
\label{eq:FKPNorm}\end{aligned}$$ where $N_{\Bbbk_i}$ is the total number of modes in a given bin $\Bbbk_i$. Hence the QML estimator of Eq. reduces to the FKP estimator [@Feldman] under the assumption that the covariance is constant within the ${{\bmath{k}}}$-bin, where several modes (labelled with Greek indices) are combined into bins (denoted by $\Bbbk_i$ and distinguished with lower case Latin indices) and the absolute values squared of the density field of each bin are summed: $$\widehat{P}(k_i)=\frac{1}{N_{\Bbbk_i}}\sum_{{{\bmath{k}}}_\alpha\in\Bbbk_i}\left\vert F({{\bmath{k}}}_\alpha)\right\vert^2.
\label{eq:FKP}$$ The difference here is that the QML estimator uses a prior of the power spectrum $P(k_\alpha)$ to weight contributions from each mode optimally, which means that the covariance of the power spectrum is minimal. The FKP estimator is commonly applied even when the assumptions of Eq. to are not valid.
Removing Contaminants: mode deprojection {#sec:BMP}
========================================
We now describe how *mode deprojection* can be applied to estimate the 3D galaxy power spectrum. The method was first suggested in @Rybicki in the context of noisy, irregularly sampled data. Applications and extensions to angular power spectra can be found for WMAP data in @Slosar:2004fr, for SDSS-III data in @Ho:2012vy, for photometric quasars of the XDQSOz catalogue in @Leistedt:2014wia and @Leistedt:2014zqa and for 2D galaxy clustering in general in @Elsner. We use the notation of @Elsner for consistency.
Suppose we estimate the power spectrum using QML and that there is only a single contaminant. Then one can suppress contaminated modes in the covariance matrix updating the covariance matrix as [@Elsner] $${\mathbf{C}}_{\alpha\beta}\rightarrow\tilde{\mathbf{C}}_{\alpha\beta}={\mathbf{C}}_{\alpha\beta}+\lim_{\sigma\rightarrow\infty}\sigma f({{\bmath{k}}}_\alpha)f^\ast({{\bmath{k}}}_\beta),
\label{eq:MP}$$ i.e. letting the covariances of contaminated modes tend to infinity. Making use of the Sherman-Morrison matrix inversion lemma [@Sherman], one can see that (if $f({{\bmath{k}}})\neq 0\;\forall{{\bmath{k}}}$) the inverse updated covariance matrix converges to $$\tilde{\mathbf{C}}{^{-1}}_{\alpha\beta}={\mathbf{C}}{^{-1}}_{\alpha\beta}-\frac{\sum_{\mu\nu}{\mathbf{C}}_{\alpha\mu}{^{-1}}f({{\bmath{k}}}_\mu)f^\ast({{\bmath{k}}}_\nu){\mathbf{C}}_{\nu\beta}{^{-1}}}{\sum_{\mu\nu} f^\ast({{\bmath{k}}}_\mu){\mathbf{C}}{^{-1}}_{\mu\nu}f({{\bmath{k}}}_\nu)}.
\label{eq:ShermanMorrisonLimit}$$ Now supposing that the modes are independent, i.e. Eq. holds, we can insert it into Eq. so that $$\tilde{\mathbf{C}}{^{-1}}_{\alpha\beta}=\frac{\delta_{\alpha\beta}}{P(k_\alpha)}-\frac{1}{R_P} \frac{f({{\bmath{k}}}_\alpha)f^\ast({{\bmath{k}}}_\beta)}{P(k_\alpha)P(k_\beta)}
\label{eq:Ctildeinv}$$ where we have defined $$R_P\equiv \sum_\mu\frac{\vert f({{\bmath{k}}}_\mu)\vert^2}{P(k_\mu)},$$ for simplicity. Taking the derivative of Eq. with respect to $P(k_i)$, we obtain the updated estimator matrix[^4] $$\tilde{\mathbf{E}}_{\alpha\beta}(k_j)=\frac{\delta_{\alpha\beta}}{P^2_\alpha}\Theta_{\alpha j}
-\frac{1}{R_P} \frac{f_\alpha f^\ast_\beta}{P_\alpha P_\beta}\left(\frac{\Theta_{\alpha j}}{P_\alpha}+\frac{\Theta_{\beta j}}{P_\beta}-\frac{t_j}{R_P}\right),
\label{eq:Ealphabetaj}$$ where $$t_i\equiv \sum_{{{\bmath{k}}}_\alpha\in\Bbbk_i}\frac{\vert f({{\bmath{k}}}_\alpha)\vert^2}{P^2(k_\alpha)}.$$ After inserting Eq. into Eq. , we obtain for the two point function $$\begin{aligned}
\mathbf p_i
=& \sum_{{{\bmath{k}}}_\alpha\in\Bbbk_i}\left\lbrace\frac{\vert F({{\bmath{k}}}_\alpha)\vert^2}{P^2(k_\alpha)}-\frac{2}{R_P}{\operatorname{Re}}\left[S_P \frac{F^\ast({{\bmath{k}}}_\alpha)f({{\bmath{k}}}_\alpha)}{P^2(k_\alpha)}\right]\right.\nonumber\\
&\left.+\frac{\vert S_P\vert^2}{R_P^2}\frac{\vert f({{\bmath{k}}}_\alpha)\vert^2}{P^2(k_\alpha)}\right\rbrace\nonumber\\
=& \sum_{{{\bmath{k}}}_\alpha\in\Bbbk_i}\frac{\left\vert F({{\bmath{k}}}_\alpha)-\frac{S_P}{R_P}f({{\bmath{k}}}_\alpha)\right\vert^2}{P^2(k_\alpha)},
\label{eq:punnorm}\end{aligned}$$ where we have defined $$S_P\equiv\sum_{{{\bmath{k}}}_\alpha}\frac{F^\ast({{\bmath{k}}}_\alpha)f({{\bmath{k}}}_\alpha)}{P_\alpha}.$$ $S_P$ is real, because $F({{\bmath{k}}})$ and $f({{\bmath{k}}})$ are Hermitian fields with real Fourier transforms.
Eq. is in a considerably simpler form than Eq. and does not require calculating many matrix elements of the estimator matrix ${\mathbf{E}}$. We show in the next section that we can consider this equation as a best-fit of the contaminants in the data.
We can normalise the updated mode deprojected QML estimator by replacing ${\mathbf{C}}$ by $\tilde{\mathbf{C}}$ in Eq. . As the term that suppresses contaminated modes from the covariance matrix in Eq. does not depend on the power $P(k)$, we have $\frac{\partial\tilde{\mathbf{C}}_{\alpha\beta}}{\partial P(k_i)}=\frac{\partial{\mathbf{C}}_{\alpha\beta}}{\partial P(k_i)}$ and hence the normalisation is $$\begin{aligned}
\tilde{\mathbf{N}}_{ij}=&\sum_{\alpha\mu\nu\rho}\tilde{\mathbf{C}}_{\alpha\mu}{^{-1}}\delta_{\mu\nu}\Theta_{\mu i}\tilde{\mathbf{C}}{^{-1}}_{\nu\rho}\delta_{\rho\alpha}\Theta_{\alpha j}\nonumber\\
=&\sum_{\alpha\mu}\vert\tilde{\mathbf{C}}_{\alpha\mu}{^{-1}}\vert^2\Theta_{\mu i}\Theta_{\alpha j}\nonumber\\
=&\sum_{\alpha\mu}\Theta_{\alpha i}\Theta_{\mu j}\left[\frac{\delta_{\alpha\mu}}{P^2(k_\alpha)}\left(1-\frac{2\vert f({{\bmath{k}}}_\alpha)\vert^2}{R_P P(k_\alpha)}\right)\right.\nonumber\\
&\left.+\frac{1}{R_P^2}\frac{\vert f({{\bmath{k}}}_\alpha)f({{\bmath{k}}}_\mu)\vert^2}{P^2(k_\alpha)P^2(k_\mu)}\right]
\label{eq:MPNormLong}\end{aligned}$$ where we have used the Hermitian property of $\tilde{{\mathbf{C}}}{^{-1}}$ in the third equality. In the first term in the square brackets, ${{\bmath{k}}}_\alpha$ has to be in both $\Bbbk_i$ and $\Bbbk_j$, hence we can replace one $\Theta$ with $\delta_{ij}$, such that Eq. can be written as a diagonal matrix with diagonal elements $\tilde n$ and the outer product of a vector with itself: $$\begin{aligned}
\tilde{\mathbf{N}}_{ij}
=&\sum_{{{\bmath{k}}}_\alpha\in\Bbbk_i}\frac{\delta_{i j}}{P^2(k_\alpha)}\left(1-\frac{2\vert f({{\bmath{k}}}_\alpha)\vert^2}{R_P P(k_\alpha)}\right)+\frac{t_it_j}{R_P^2}\nonumber\\
\equiv &\tilde n_i\delta_{ij}+\frac{t_it_j}{R_P^2},
\label{eq:MPNorm}\end{aligned}$$ This means that we can apply the Sherman-Morrison matrix inversion lemma [@Sherman]: $$\begin{aligned}
\tilde{\mathbf{N}}{^{-1}}_{ij}
&=\frac{\delta_{ij}}{\tilde n_i}-\frac{1}{R_P^2+\sum_\ell \frac{t_\ell^2}{\tilde n_\ell}}\frac{t_i}{\tilde n_i}\frac{t_j}{\tilde n_j}.
\label{eq:Ndecomp}\end{aligned}$$ As $\tilde{\mathbf{N}}{^{-1}}$ is not diagonal, it does not reduce to a simple FKP style estimator, i.e. if we have $N_\mathrm{bin}$ bins, we have to calculate for each bin the $N_\mathrm{mode}\times N_\mathrm{mode}$ estimator matrix ${\mathbf{E}}$ and we have to invert the $N_\mathrm{bin}\times N_\mathrm{bin}$ normalisation matrix. This is not feasible for 3D clustering, because of the large number of modes to be considered, especially if we want to choose narrow bins. Including several contaminants makes it even more costly.
One way around this is a new framework introduced by @Leistedt:2014wia which they call *extended mode projection* and that [ selectively removes modes based on cross correlations with the data. However, this procedure]{} reintroduces a small bias [@Elsner].
Another possibility is using [ the methodology of]{} the SDSS-III Baryon Oscillation Spectroscopic Survey (BOSS)-collaboration[ , which is similar to that described in the next section, but applied]{} at the power spectrum level. However, this method is also biased [@Elsner]. Although @Rossinprep show that, for the Completed SDSS-III Baryon Oscillation Spectroscopic Survey (BOSS DR12), the bias is much smaller than the statistical uncertainty, it was shown in the appendix of @Ross:2012qm that the bias is significant when one attempts to correct for many systematics. Furthermore, we expect smaller statistical uncertainties with future surveys, so in the next two sections we consider a computationally cheaper way of removing this small bias.
Removing Contaminants: [ mode]{} subtraction {#sec:systrem}
============================================
Here we will consider [ *mode subtraction*]{} and its link to *mode deprojection*. In order to remove contaminants we start by [ treating the true, but unknown, amplitude of the contamination $\varepsilon_{\rm true}$ in Eq. as]{} a free parameter $\varepsilon$, so that [ an estimate of the true density field $D({{\bmath{k}}})$]{} reads $${ \widehat D}({{\bmath{k}}})=F({{\bmath{k}}})-\varepsilon f({{\bmath{k}}}).
\label{eq:vartemp}$$ Note that this is different to the *template subtraction* method introduced by @Ho:2012vy, which is used by the BOSS collaboration and works entirely at the level of power spectra, whereas Eq. works at the map level. We can write a simplified model of the Gaussian likelihood whose maximum is given by the QML (cf. Eq. and @Tegmark:1997rp) in the approximation of a diagonal covariance matrix, with a small contaminant that does not affect the covariance. This is given by $$-2\ln\mathcal L=\ln\left(\prod_{{\bmath{k}}}P(k)\right)+\sum_{{\bmath{k}}}\frac{\vert F({{\bmath{k}}})-\varepsilon f({{\bmath{k}}})\vert^2}{P(k)}.
\label{eq:chi2}$$ We can therefore find $\varepsilon$ by minimising Eq. , which is equivalent to simultaneously fitting $\varepsilon$ and the model parameters entering the model power spectrum. The derivative of $\ln\mathcal L$ with respect to $\varepsilon$ reads $$\begin{aligned}
\frac{\partial\ln\mathcal L}{\partial\varepsilon}&=\sum_{{\bmath{k}}}\frac{{\operatorname{Re}}\left[F_f({{\bmath{k}}})F^\ast({{\bmath{k}}})\right]-\varepsilon\vert F_f({{\bmath{k}}})\vert^2}{P(k)}.\end{aligned}$$ This expression is equal to zero and the likelihood maximised if $$\varepsilon^\mathrm{(BF)}=\frac{S_P}{R_P}.$$
The uncontaminated estimate of the density field is hence given by $${ \widehat D}({{\bmath{k}}})=F({{\bmath{k}}})-\frac{S_P}{R_P}f({{\bmath{k}}}),
\label{eq:FsignalBF}$$ and we can estimate the power as $$\widehat{P}(k_i)=\frac{1}{N_{\Bbbk_i}}\sum_{{{\bmath{k}}}_\alpha}\left\vert F({{\bmath{k}}}_\alpha)-\frac{S_P}{R_P}f({{\bmath{k}}}_\alpha)\right\vert^2.
\label{eq:PTS}$$ This is similar to the *mode deprojection* result of Eq. with a bias, missing the inverse noise matrix convolution of Eq. . [ The bias of this estimate comes about because $S_P$ is correlated with the true density field $D({{\bmath{k}}})$. This correlation is similar to that created by the internal linear combination (ILC) method [e.g. @Bennett:2003ca] for the analysis of cosmic microwave background (CMB) data.]{} Based on this knowledge, we build an unbiased FKP-style estimator in the next section.
An Unbiased FKP-Style Estimator {#sec:debias}
===============================
We present in this section a simple, although sub-optimal, way to [ remove the bias on]{} the power spectrum estimate [ resulting from imperfectly removing systematics using]{} either Eq. or . A straightforward way to remove the bias consists of calculating the expectation value of the power from each mode analytically, assuming Eq. , and divide out the bias. We start with calculating some useful expectations which we need for the final result, summarised in Table \[tab:expectations\].
---------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------
$\langle F({{\bmath{k}}}_\alpha) F^\ast({{\bmath{k}}}_\beta)\rangle$ $\delta_{\alpha\beta}P(k_\alpha)+{ \varepsilon_\mathrm{true}^2} f({{\bmath{k}}}_\alpha) f^\ast({{\bmath{k}}}_\beta)$
$\langle\varepsilon_\mathrm{BF}\rangle$ $\frac{\langle S_P\rangle}{R_P}={ \varepsilon_\mathrm{true}}$
$\langle\varepsilon^2_\mathrm{BF}\rangle$ $\frac{1}{R_P}+{ \varepsilon_\mathrm{true}^2}$
$\langle S_P F({{\bmath{k}}}_\alpha)\rangle$ $f({{\bmath{k}}}_\alpha)+R_P f({{\bmath{k}}}_\alpha){ \varepsilon_\mathrm{true}^2}$
---------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------
: Expectation values of quantities entering Eq. .[]{data-label="tab:expectations"}
With these equations at hand, we can calculate the expectation of Eq. and , i.e. the two-point function of Eq. : $$\begin{aligned}
&\langle\vert F({{\bmath{k}}}_\alpha)-\frac{S_P}{R_P} f({{\bmath{k}}}_j)\vert^2\rangle\nonumber\\
&=\langle\vert F({{\bmath{k}}}_\alpha)\vert^2\rangle-\frac{2}{R_P}\langle S_P F({{\bmath{k}}}_\alpha)\rangle f^\ast({{\bmath{k}}}_\alpha)+{ \langle\varepsilon^2_\mathrm{BF}\rangle}\vert f({{\bmath{k}}}_\alpha)\vert^2\nonumber\\
&=P(k_\alpha)-\frac{\vert f({{\bmath{k}}}_\alpha)\vert^2}{R_P},\end{aligned}$$ hence, we can build an unbiased estimator of the power by dividing each mode in Eq. and by $$1-\frac{1}{R_P}\frac{\vert f({{\bmath{k}}}_\alpha)\vert^2}{P(k_\alpha)}.
\label{eq:debiasfactor}$$ If we want to debias the two-point function using this factor, we have to assume a prior power spectrum. Note that the QML approach also requires the prior knowledge of the power spectrum. We will see in Sec. \[sec:eg\] that the impact of adopting a slightly wrong prior is indeed small. Our final estimator of the power spectrum is then $$\widehat{P}(k_i)=\frac{1}{N_{\Bbbk_i}}\sum_{{{\bmath{k}}}_\alpha}\frac{\left\vert F({{\bmath{k}}}_\alpha)-\frac{S_P}{R_P}f({{\bmath{k}}}_\alpha)\right\vert^2}{1-\frac{1}{R_P}\frac{\vert f({{\bmath{k}}}_\alpha)\vert^2}{P(k_\alpha)}}.
\label{eq:ubPTS}$$ Eq. is one of the key results of this article: this is an extension of the FKP estimator that removes potential contaminants from the data in an unbiased way, without the need for large matrices. Moreover, as it is in the same form as the well established FKP estimator, this can easily be folded into estimators for redshift-space clustering such as those by @Bianchi:2015oia and @Scoccimarro:2015bla.
The same debiasing factor can also be derived from the QML Fisher information matrix ${\mathbf{N}}$, which in the QML approach performs both the debiasing and optimisation effects. Without binning, Eq. simplifies to $$\begin{aligned}
\tilde{\mathbf{N}}_{\alpha\beta}
=&\frac{\delta_{\alpha \beta}}{P^2(k_\alpha)}\left(1-\frac{2\vert f({{\bmath{k}}}_\alpha)\vert^2}{R_P P(k_\alpha)}\right)+\frac{1}{R_P^2}\frac{\vert f({{\bmath{k}}}_\alpha)\vert^2}{P^2(k_\alpha)}\frac{\vert f({{\bmath{k}}}_\beta)\vert^2}{P^2(k_\beta)}.
\label{eq:Nnobins}\end{aligned}$$ The difference between the two approaches is that QML provides an unbiased optimal power estimate, whereas Eq. has been constructed such that it is only unbiased, i.e. the powers in the denominators of Eq. act as optimal weights to each mode. If we allow for some information loss within bins, by assuming the expected power is constant within each bin, we can replace $P^2(k_\beta)$ by $P(k_\alpha)P(k_\beta)$, such that $$\begin{aligned}
\tilde{\mathbf{N}}_{\alpha\beta}
=&\frac{\delta_{\alpha \beta}}{P^2(k_\alpha)}\left(1-\frac{2\vert f({{\bmath{k}}}_\alpha)\vert^2}{R_P P(k_\alpha)}\right)+\frac{1}{R_P^2}\frac{\vert f({{\bmath{k}}}_\alpha)\vert^2}{P^3(k_\alpha)}\frac{\vert f({{\bmath{k}}}_\beta)\vert^2}{P(k_\beta)}.\end{aligned}$$ This normalisation is proportional to the Fisher information matrix [@Tegmark:1997rp], from which we marginalise out contributions from other modes by summing over all modes ${{\bmath{k}}}_\beta$: $$\begin{aligned}
\sum_\beta\tilde{\mathbf{N}}_{\alpha\beta}
=&\frac{1}{P^2(k_\alpha)}\left(1-\frac{2\vert f({{\bmath{k}}}_\alpha)\vert^2}{R_P P(k_\alpha)}\right)+\frac{1}{R_P}\frac{\vert f({{\bmath{k}}}_\alpha)\vert^2}{P^3(k_\alpha)}\nonumber\\
=&\frac{1}{P^2(k_\alpha)}\left(1-\frac{\vert f({{\bmath{k}}}_\alpha)\vert^2}{R_P P(k_\alpha)}\right).\end{aligned}$$ This is exactly Eq. with a factor of $\frac{1}{P^2(k_\alpha)}$ that cancels out the difference between Eq. and Eq. . We have therefore shown that Eq. is a non-optimal, but unbiased, approximation to using the QML normalisation with *mode deprojection*. In the limit of narrow bins, when the power spectrum does not change significantly within the bin, Eq. is mathematically identical to the QML result. We shall study the impact of this sub-optimality in examples later in Sec. \[sec:eg\]. In fact, we will argue later that this is actually a weaker effect than many common approximations applied when using the FKP estimator, such as ignoring large-scale window effects in the QML approach, when averaging large scale modes.
Note that[ , in the absence of systematics,]{} we have assumed a diagonal covariance matrix in the derivation of both the [ *mode subtraction*]{} and the debiasing step. In practic[ e t]{}he covariance matrix has off-diagonal terms due to the effect of the survey window. However, this is usually not included when calculating the data power spectrum but, instead, it is included as a convolution in the model power spectrum. We show in Appendix \[sec:nondiag\] that Eq. still holds in the general case of having a non-diagonal covariance matrix, as long as $R_P$ is generalised as in Eq. . [ This generalised $R_P$ requires the inversion of the full $N_{\rm mode}^2$ covariance matrix. However, we show in Appendix \[sec:diagCapprox\] that the effect of assuming a diagonal covariance matrix is either small, or can be corrected for using the covariance matrix, without inversion.]{}
Removing multiple Contaminants {#sec:multisyst}
==============================
We have shown the equivalence between *mode deprojection* and *debiased [ mode]{} subtraction* for one contaminant. A realistic survey has several sources of potential contaminants, so we show here this equivalence holds for an arbitrary number of templates. For *mode deprojection*, we have to update the covariance matrix with a sum over all templates, and thus we have to replace Eq. with $$\tilde {\mathbf{C}}_{\alpha\beta}={\mathbf{C}}_{\alpha\beta}+\lim_{\sigma\rightarrow\infty}\sigma\sum_{A=1}^{N_\mathrm{sys}}f_A({{\bmath{k}}}_\alpha)f^\ast_A({{\bmath{k}}}_\beta).
\label{eq:MPmult}$$ Starting from Eq. , we derive in Appendix \[sec:MPmultderiv\] the unbinned *mode deprojection* power spectrum $$\widehat P(k_\alpha)=\left\vert F({{\bmath{k}}}_\alpha)-\sum_{AB}\mathbf{S}_A\mathbf{R}{^{-1}}_{AB}f_B({{\bmath{k}}}_\alpha)\right\vert^2,
\label{eq:multiBMPpower}$$ where $\mathbf{R}_{AB}\equiv\sum_{\mu}\frac{f_A^\ast({{\bmath{k}}}_\mu)f_B(k_\mu)}{P(k_\mu)}$ and $\mathbf{S}_A\equiv \sum_\alpha\frac{f_A({{\bmath{k}}}_\alpha)F^\ast({{\bmath{k}}}_\alpha)}{P(k_\alpha)}$ are matrix and vector equivalents of $R_P$ and $S_P$, respectively, in contaminant space.
To apply multiple [ mode]{} subtraction, we extend the likelihood given in Eq. to $$-2\ln\mathcal{L}=\sum_\alpha\frac{\vert F({{\bmath{k}}}_\alpha)-\sum_A \mathbf{\varepsilon}_A f_A({{\bmath{k}}}_\alpha)\vert^2}{P(k_\alpha)}.$$ Writing $\bvarepsilon$ as a vector, the joint maximum likelihood solution fitting all contaminants is given by (cf. Appendix \[sec:TSmultideriv\]) $$\bvarepsilon^\mathrm{(BF)}=\mathbf{R}{^{-1}}\mathbf{S}.$$ Note that this would require fitting the amplitude of all contaminants simultaneously. The absolute value squared of the best fitting signal is hence equal to Eq. . Hence, we also do not need large $N_\mathrm{mode}\times N_\mathrm{mode}$ matrices when we have to remove several potential contaminants.
We can calculate the debiasing factor $$\begin{aligned}
\sum_{j}\tilde{\mathbf{N}}_{ij}P^2(k_i)=1-\sum_{AB}\frac{f_A({{\bmath{k}}}_i)\mathbf{R}{^{-1}}_{AB}f_B^\ast({{\bmath{k}}}_i)}{P(k_i)}\end{aligned}$$ analogously to Sec. \[sec:systrem\] from the *mode deprojection* normalisation matrix without binning.
Testing Contaminant Removal {#sec:eg}
===========================
In this section we show how simple contaminants can be removed in power spectrum measurements from simulated density fields, using the hitherto described methodologies.
Gaussian Spike Contaminant {#sec:GaussSpike}
--------------------------
![A slice through a realisation of a Gaussian random field contaminated with a Gaussian spike used in Sec. \[sec:GaussSpike\]. The top panel shows the “clean" Gaussian random field (corresponding to $D({{\bmath{k}}})$ through Fourier transform) in configuration space. In the central panel, we have plotted the contaminated field (Fourier pair of $F({{\bmath{k}}})$) with an obvious Gaussian overdensity in the centre. The bottom panel shows [ the residual, i.e. the difference of the field after *mode subtraction* (i.e. the Fourier transform of $F({{\bmath{k}}})-{\varepsilon}^\mathrm{(BF)} f({{\bmath{k}}})$, cf. Eq. ) and the input field]{}. The best-fitting ${\varepsilon}^\mathrm{(BF)}$ for this particular realisation amounts to 1.078. Although differences between the top and bottom panels are hard to spot by eye, in Fourier space the differences correspond to the bias in the [ *mode subtraction*]{} estimator.[]{data-label="fig:spikefield"}](SpikeContours.pdf){width="\linewidth"}
As a first test, we generate 3-dimensional Gaussian random fields according to an input power spectrum that we calculate using CAMB [@Lewis:1999bs]. Each of these fields consists of a $16\times 16\times 16$ grid, in a box of length $3136h^{-1}\;\mathrm{Mpc}$. An example of such a field is shown in the left panel of Fig. \[fig:spikefield\]. We contaminate these Gaussian random fields by adding a real Gaussian spike in $k$-space with width $\sigma^2=10^{-5}h\;\mathrm{Mpc}$, centred around $k=0.01h\;\mathrm{Mpc}^{-1}$, such that its maximum lies within a bin with sufficiently good statistics. The Fourier transform of this contaminant field is again a Gaussian spike in the centre of the box with some long wavelength fluctuations around it. The amplitude of the real part over-density in ${{\bmath{k}}}$-space is 100, thus having the same order of magnitude as the “true" density field. An example of this setup can be seen in the central panel of Fig. \[fig:spikefield\]. We calculate four different power spectra:
1. We do not account for the contaminants and just average the absolute values squared of the density field in each bin (cf. Eq. ).
2. We perform a naïve [ *mode subtraction*]{}, i.e. we subtract off the template, but do not debias the two-point function (cf. Eq. ).
3. We debias the previous power spectrum by applying Eq. .
4. We use the full QML estimator with *mode deprojection*.
In the cases (ii) to (iv), we have to assume a prior power spectrum, which we take as equal to the input power. We shall test the effect of this assumption with the next example. As each bin contains modes with a range of different ${{\bmath{k}}}$-values, we have to clarify what we mean by the prior power spectrum $P(k_i)$ for a specific bin. We find that the power spectrum measurements are closest to the input values, when we assume that the input power spectrum $P(k_i)$ is given by the average of the prior power spectrum values for each mode in the respective bin, i.e. $$P(k_i)\equiv \frac{1}{N_{\Bbbk_i}}\sum_{{{\bmath{k}}}_\alpha\in\Bbbk_i}P(k_\alpha).$$
In Fig. \[fig:spike\], we can clearly see an increase of power in the bins around $k=0.01$ in case (i). Subtracting off the template in the naïve way (method (ii)) is biased in the bins affected by the spike. However, this bias is only a 1 part in a thousand effect. Methods (iii) and (iv) both reproduce the input power spectrum well, removing the bias. A significant difference between their error bars cannot be observed. It is therefore sufficient in this case to use the FKP-style estimator we introduced in Sec. \[sec:debias\].
![Means and standard deviations of the power spectra of 70,000 realisations of Gaussian random fields contaminated with a real Gaussian spike. The top panel shows the input power spectrum as a solid blue line, as well as the power spectra obtained with methods (i)-(iv) as described in Sec. \[sec:GaussSpike\]. In the lower panel, we plot fractional errors for methods (ii)-(iv). []{data-label="fig:spike"}](test_fBMP_eval_spike.pdf){width="\linewidth"}
Single Contaminated Mode {#sec:scm}
------------------------
![This plot is similar to Fig. \[fig:spikefield\], but shows a slice through a field with a single contaminated mode as described in Sec. \[sec:scm\]. The best-fitting ${\varepsilon}^\mathrm{(BF)}$ for this particular realisation amounts to 1.005. All panels appear very similar; this is quantified in Fig. \[fig:highlybiased\].[]{data-label="fig:ex2field"}](Ex2Contours.pdf){width="\linewidth"}
As a second example we use Eq. to construct a contaminant that would lead to a strong bias in the recovered $P(k)$ without the debiasing step. Eq. only contains positive quantities and is normalised such that the bias is a value between 0 and 1. 1 corresponds to an unbiased estimate, hence 0 is the maximal bias. This extreme case would be fulfilled if $f$ is large for one mode and 0 otherwise. Therefore, we construct a contaminant that is a large number at the modes corresponding to ${{\bmath{k}}}=\pm (0.003,0.003,0.003)h\;\mathrm{Mpc}^{-1}$. An example of this setup can be found in Fig. \[fig:ex2field\]. The top panel again shows an uncontaminated Gaussian random field, the central panel shows the same field with the contaminant added. The contaminant itself is not as prominent as the one in Fig. \[fig:spikefield\], because this single contaminated mode just adds a long wavelength contribution in real space. The bottom panel shows the field after subtracting the template.
We measure the same cases (i)-(iv) as in the previous subsection, which we plot in Fig. \[fig:highlybiased\]. The prior power is again the input power. If we were to apply this to a real survey, we would not know the true power, so we perform a few runs, where we first assume a flat prior power spectrum $P(k)=1\;\forall k$, and then iteratively compute the power with the power from the previous run as the prior power spectrum. The effect of the prior power spectrum is negligible, because the result in the first step provides the same result as assuming the input power as prior.
The data points for all cases (i)-(iv) are close to the input power in all bins but the second. In the second bin, the power spectrum for case (i) extends beyond the plotted range, chosen to highlight differences between the other approaches. In case (ii), the power is significantly underestimated. The bias amounts to about 2 per cent, i.e. it highly affects measurements where small-${{\bmath{k}}}$ modes are crucial, such as $f_\mathrm{NL}$-measurements. The difference between the cases (iii) and (iv) is much smaller, even in this extreme example.
![Means and standard deviations of the power spectra of 1000 realisations of Gaussian random fields contaminated with Hermitian Gaussian spikes. The red dots represent measurements, where the contamination has not been taken into account. For the blue dots, *mode deprojection* has been used to remove the spikes. For the green dots, we used *debiased [ mode]{} subtraction*. The solid blue line shows the input power spectrum.[]{data-label="fig:highlybiased"}](test_fBMP_eval_highlybiased.pdf){width="\linewidth"}
Conclusions {#sec:conclusion}
===========
We have considered methods to remove contaminants when measuring the 3D galaxy power spectrum from a given density field, focussing on *mode deprojection* and [ *mode subtraction*]{}. In order to understand how these are related, we have decomposed the problem into separate steps. In particular we have separated *mode deprojection* from power spectrum estimation - they are often considered together - arguing that this split makes sense given the mathematical equivalence of *mode deprojection* and [ *mode subtraction*]{}. We argue that the QML estimation is not practical for modern surveys with large numbers of observed modes, but that we can apply *mode deprojection* to the FKP-estimator, using the mathematical equivalence of *mode deprojection* and [ *mode subtraction*]{}, thus avoiding having to create large estimator and covariance matrices for all modes. The resulting estimate is biased, but can easily be made unbiased with a simple correction, again that can be implemented without the [ inversion]{} of large matrices. This correction is easily extended to the case of multiple contaminants and is not affected if the modes are correlated even without the effects of contaminants. The final result of our short paper is the suggestion that 3D galaxy power spectrum should be estimated using Eq. ,
$$\widehat{P}(k_i)=\frac{1}{N_{\Bbbk_i}}\sum_{{{\bmath{k}}}_\alpha}\frac{\left\vert F({{\bmath{k}}}_\alpha)-\frac{S_P}{R_P}f({{\bmath{k}}}_\alpha)\right\vert^2}{1-\frac{1}{R_P}\frac{\vert f({{\bmath{k}}}_\alpha)\vert^2}{P(k_\alpha)}}.$$
While theoretically it is sub-optimal, in practice the degradation of signal is expected to be less than ignoring window effects in the optimisation of mode averaging when using the standard FKP estimator.
Acknowledgments {#acknowledgments .unnumbered}
===============
The authors would like to thank Franz Elsner, Héctor Gil-Marín, Ashley Ross and the unknown referee for valuable comments.
We used matplotlib [@Hunter:2007] to generate plots. The CAMB package [@Lewis:1999bs] has been used to generate model and prior power spectra. We made use of the facilities and staff of the UK Sciama High Performance Computing cluster supported by the ICG, SEPNet and the University of Portsmouth.
BK thanks the Faculty of Technology of the University of Portsmouth for support during his PhD studies. WJP and DB acknowledge support from UK STFC through the consolidated grant ST/K0090X/1, WJP also acknowledges support from the European Research Council through the Darksurvey grant and the UK Space Agency through grant ST/N00180X/1. LS is grateful for support from SNSF grant SCOPES IZ73Z0-152581, GNSF grant FR/339/6-350/14, and DOE grant DEFG 03-99EP41093.
Derivation of Mode Deprojection with multiple templates {#sec:MPmultderiv}
=======================================================
In this appendix we want to derive Eq. from Eq. . We start by rewriting Eq. in matrix notation $$\tilde{\mathbf{C}}={\mathbf{C}}+\lim_{\sigma\rightarrow\infty}\sigma{\mathbf{f}}\mathbf{I}_{N_\mathrm{sys}}{\mathbf{f}}^\dagger,$$ defining an $N_\mathrm{mode}\times N_\mathrm{sys}$ matrix ${\mathbf{f}}_{\alpha A}\equiv f_A({{\bmath{k}}}_\alpha)$, such that we can invert $\tilde{\mathbf{C}}$ using the Woodbury matrix identity $$\begin{aligned}
\tilde{\mathbf{C}}{^{-1}}&={\mathbf{C}}{^{-1}}-{\mathbf{C}}{^{-1}}\lim_{\sigma\rightarrow\infty}\sigma{\mathbf{f}}\left(\mathbf{I}{^{-1}}_{N_\mathrm{sys}}+{\mathbf{f}}^\dagger{\mathbf{C}}{^{-1}}\sigma{\mathbf{f}}\right){^{-1}}{\mathbf{f}}^\dagger{\mathbf{C}}{^{-1}}\nonumber\\
&={\mathbf{C}}{^{-1}}-{\mathbf{C}}{^{-1}}{\mathbf{f}}\left({\mathbf{f}}^\dagger{\mathbf{C}}{^{-1}}{\mathbf{f}}\right){^{-1}}{\mathbf{f}}^\dagger{\mathbf{C}}{^{-1}}\nonumber\\
&\equiv{\mathbf{C}}{^{-1}}-{\mathbf{C}}{^{-1}}{\mathbf{f}}\mathbf{R}{^{-1}}{\mathbf{f}}^\dagger{\mathbf{C}}{^{-1}}.\end{aligned}$$ If we assume ${\mathbf{C}}_{\alpha\beta}=\delta_{\alpha\beta}P(k_\alpha)$, $\mathbf{R}\equiv{\mathbf{f}}^\dagger{\mathbf{C}}{^{-1}}{\mathbf{f}}$ becomes a matrix equivalent to the factor $R_P$ in previous sections: $$\mathbf{R}_{AB}=\sum_{\mu\nu}f_A^\ast({{\bmath{k}}}_\mu)\frac{\delta_{\mu\nu}}{P(k_\mu)}f_B(k_\nu)=\sum_{\mu}\frac{f_A^\ast({{\bmath{k}}}_\mu)f_B(k_\mu)}{P(k_\mu)}.$$ The inverse updated covariance matrix then reads $$\tilde{\mathbf{C}}{^{-1}}_{\alpha\beta}=\frac{\delta_{\alpha\beta}}{P(k_\alpha)}-\sum_{AB}\frac{f_A({{\bmath{k}}}_\alpha)\mathbf{R}{^{-1}}_{AB}f_B^\ast({{\bmath{k}}}_\beta)}{P(k_\alpha)P(k_\beta)}.
\label{eq:CinvtildeMulti}$$ If we do not bin, but apply mode deprojection to each mode separately, the matrix $\tilde{\mathbf{E}}$ simplifies to $$\tilde{\mathbf{E}}_{\alpha\beta}(k_j)=\sum_{\mu\nu}\tilde{\mathbf{C}}{^{-1}}_{\alpha\mu}\delta_{\mu j}\delta_{\mu\nu}\tilde{\mathbf{C}}{^{-1}}_{\nu\beta}=\tilde{\mathbf{C}}{^{-1}}_{\alpha j}\tilde{\mathbf{C}}{^{-1}}_{j\beta}.
\label{eq:Etildemulti}$$ After inserting Eq. into Eq. , we obtain $$\begin{aligned}
P^2&(k_j)\sum_{\alpha\beta} F^\ast({{\bmath{k}}}_\alpha)\tilde{\mathbf{E}}_{\alpha\beta}(k_j)F({{\bmath{k}}}_\beta)\nonumber\\
=&\vert F({{\bmath{k}}}_j)\vert^2
\nonumber\\
&-\sum_{AB\alpha}\frac{F^\ast({{\bmath{k}}}_\alpha)f_A({{\bmath{k}}}_\alpha)}{P(k_\alpha)}\mathbf{R}{^{-1}}_{AB}f_B^\ast({{\bmath{k}}}_j)F({{\bmath{k}}}_j)\nonumber\\
&-\sum_{AB\beta}F^\ast({{\bmath{k}}}_j)f_A({{\bmath{k}}}_j)\mathbf{R}{^{-1}}_{AB}\frac{f_B^\ast({{\bmath{k}}}_\beta)F({{\bmath{k}}}_\beta)}{P(k_\beta)}\nonumber\\
&+\sum_{ABCD\alpha\beta}\frac{F^\ast({{\bmath{k}}}_\alpha)f_A({{\bmath{k}}}_\alpha)}{P(k_\alpha)}\mathrm{R}{^{-1}}_{AB}f^\ast_B({{\bmath{k}}}_j)f_C({{\bmath{k}}}_j)\mathrm{R}{^{-1}}_{CD}\frac{f_D^\ast({{\bmath{k}}}_\beta)F({{\bmath{k}}}_\beta)}{P(k_\beta)}\nonumber\\
=&\vert F({{\bmath{k}}}_j)\vert^2
\nonumber\\
&-2{\operatorname{Re}}\left[\sum_{AB}\mathbf{S}_A\mathbf{R}{^{-1}}_{AB}f_B^\ast({{\bmath{k}}}_j)F({{\bmath{k}}}_j)\right]\nonumber\\
&+\left\vert\sum_{AB}\mathbf{S}_A\mathbf{R}{^{-1}}_{AB}f_B({{\bmath{k}}}_j)\right\vert^2\nonumber\\
=&\left\vert F({{\bmath{k}}}_j)-\sum_{AB}\mathbf{S}_A\mathbf{R}{^{-1}}_{AB}f_B({{\bmath{k}}}_j)\right\vert^2,\end{aligned}$$ where we defined $\mathbf{S}_A\equiv \sum_\alpha\frac{f_A({{\bmath{k}}}_\alpha)F^\ast({{\bmath{k}}}_\alpha)}{P(k_\alpha)}$ analogous to $S_P$.
Derivation of [ mode]{} subtraction with multiple templates {#sec:TSmultideriv}
===========================================================
Here we derive the best-fitting $\bvarepsilon^{(BF)}$ from the likelihood $$-2\ln\mathcal{L}=\sum_\alpha\frac{\vert F({{\bmath{k}}}_\alpha)-\sum_A \mathbf{\varepsilon}_A f_A({{\bmath{k}}}_\alpha)\vert^2}{P(k_\alpha)}$$ to find the same result as in the previous appendix. Taking the derivative with respect to $\bvarepsilon_B$ yields $$\frac{\partial\chi^2}{\partial\bvarepsilon_B}=-2\bvarepsilon_B\sum_\alpha\frac{f_B({{\bmath{k}}}_\alpha)F^\ast({{\bmath{k}}}_\alpha)-\sum_A\bvarepsilon_A f_B({{\bmath{k}}}_\alpha)f^\ast_A({{\bmath{k}}}_\alpha)}{P(k_\alpha)}.$$ This derivative is zero if $$\sum_\alpha\frac{f_B({{\bmath{k}}}_\alpha)F^\ast({{\bmath{k}}}_\alpha)}{P(k_\alpha)}=\sum_{A\alpha}\frac{\bvarepsilon_A f_B({{\bmath{k}}}_\alpha)f^\ast_A({{\bmath{k}}}_\alpha)}{P(k_\alpha)},$$ which reads $$\mathbf{S}=\mathbf{R}\bvarepsilon$$ in matrix notation. The best fitting $\bvarepsilon$ value is therefore given by $$\bvarepsilon^\mathrm{(BF)}=\mathbf{R}{^{-1}}\mathbf{S}.$$ The absolute value squared of the best fitting signal is hence equal to Eq. : $$\begin{aligned}
\left\vert F({{\bmath{k}}}_\alpha)-\sum_{A}\bvarepsilon^\mathrm{(BF)}_A f_A({{\bmath{k}}}_\alpha)\right\vert^2=\left\vert F({{\bmath{k}}}_\alpha)-\sum_{AB}\mathbf{R}{^{-1}}_{AB}\mathbf{S}_B f_A({{\bmath{k}}}_\alpha)\right\vert^2.\end{aligned}$$
[ mode]{} subtraction and the Debiasing Step for a non-diagonal Covariance Matrix
=================================================================================
Including the covariance in the calculation of the best-fit mode to subtract {#sec:nondiag}
----------------------------------------------------------------------------
Here we show a derivation similar to the one in Sec. \[sec:systrem\] and \[sec:debias\] for the more general case of a non-diagonal covariance matrix. We show that the debiasing works in the same way as in Sec. \[sec:debias\], just with a generalised definition of $R_P$.
Defining the covariance matrix of the true density $${\mathbf{C}}_{\alpha\beta}\equiv\left\langle D_\alpha D_\beta^\ast\right\rangle$$ and assuming that the true signal and the contaminant are uncorrelated, we can write $$\left\langle F_\alpha F_\beta^\ast\right\rangle={\mathbf{C}}_{\alpha\beta}+{ \varepsilon_\mathrm{true}^2} f_\alpha f_\beta^\ast.
\label{eq:Cdef}$$ As we did in Sec. \[sec:systrem\], we introduce a free parameter $\varepsilon$, such that $${ \widehat D}_\alpha\equiv F_\alpha-\varepsilon f_\alpha.$$ Assuming that the true density field is Gaussian, its log-likelihood reads $$-2\ln\mathcal{L}=\sum_{\alpha\beta}\left(F_\alpha-\varepsilon f_\alpha\right)^\ast {\mathbf{C}}_{\alpha\beta}{^{-1}}\left(F_\beta-\varepsilon f_\beta\right)+\mathrm{const}.$$ To find the best-fitting $\varepsilon^\mathrm{(BF)}$, we take the derivative of the log-likelihood with respect to $\varepsilon$: $$\begin{aligned}
-2\frac{\partial\ln\mathcal{L}}{\partial \varepsilon}=&-\sum_{\alpha\beta}f_\alpha^\ast {\mathbf{C}}_{\alpha\beta}{^{-1}}\left(F_\beta-\varepsilon f_\beta\right)\nonumber\\
&-\sum_{\alpha\beta}\left(F_\alpha-\varepsilon f_\alpha\right)^\ast {\mathbf{C}}_{\alpha\beta}{^{-1}}f_\beta\nonumber\\
=&2\varepsilon\sum_{\alpha\beta}f_\alpha^\ast {\mathbf{C}}_{\alpha\beta}{^{-1}}f_\beta\nonumber\\
&-\sum_{\alpha\beta}\left[f_\alpha^\ast {\mathbf{C}}_{\alpha\beta}{^{-1}}F_\beta+F_\alpha^\ast {\mathbf{C}}_{\alpha\beta}{^{-1}}f_\beta\right].
\label{eq:logLderiv}\end{aligned}$$ As ${\mathbf{C}}$ is a covariance matrix of complex random variables, it is Hermitian positive-semidefinite, such that the second sum can be written as $$\sum_{\alpha\beta}\left[f_\alpha^\ast {\mathbf{C}}_{\alpha\beta}{^{-1}}F_\beta+F_\alpha^\ast {\mathbf{C}}_{\alpha\beta}{^{-1}}f_\beta\right]=2{\operatorname{Re}}\left[\sum_{\alpha\beta}f_\alpha^\ast {\mathbf{C}}_{\alpha\beta}{^{-1}}F_\beta\right].$$ For shortness and in analogy to Sec. \[sec:systrem\], we call this sum $$S_P\equiv \sum_{\alpha\beta}{\operatorname{Re}}\left[f_\alpha^\ast {\mathbf{C}}_{\alpha\beta}{^{-1}}F_\beta\right]$$ and the first sum in Eq. we call $$R_P\equiv \sum_{\alpha\beta}f_\alpha^\ast {\mathbf{C}}_{\alpha\beta}{^{-1}}f_\beta.
\label{eq:RPdef}$$ We obtain the best-fitting, i.e. maximum likelihood, value $$\varepsilon^\mathrm{(BF)}=\frac{S_P}{R_P}$$ by equating Eq. to zero.
Now we want to calculate the expectation value $$\begin{aligned}
\left\langle\left\vert F_\alpha-\frac{S_P}{R_P} f_\alpha\right\vert^2\right\rangle=&\left\langle\left\vert F_\alpha\right\vert^2\right\rangle-\frac{2}{R_P}{\operatorname{Re}}\left[\left\langle S_P F_\alpha^\ast f_\alpha\right\rangle\right]\nonumber\\
&+\left\langle\frac{S_P^2}{R_P^2}\right\rangle\left\vert f_\alpha\right\vert^2.
\label{eq:ExpVal}\end{aligned}$$ We calculate each term separately:
1. The first term $\langle\vert F_\alpha\vert^2\rangle={\mathbf{C}}_{\alpha\alpha}+{ \varepsilon_\mathrm{true}^2}\vert f_\alpha\vert^2$ is a special case of Eq. .
2. To calculate the second term, we reexpand $S_P$ and use the fact that ${\operatorname{Re}}\left[F_\alpha f_\alpha^\ast\right]={\operatorname{Re}}\left[F_\alpha^\ast f_\alpha\right]$: $$\begin{aligned}
2{\operatorname{Re}}\left[\left\langle S_P F_\alpha^\ast f_\alpha\right\rangle\right]=2{\operatorname{Re}}\left[\sum_{\gamma\beta}f_\gamma^\ast {\mathbf{C}}_{\gamma\beta}{^{-1}}\left\langle F_\beta F_\alpha^\ast\right\rangle f_\alpha\right] \end{aligned}$$ After reinserting Eq. , we get $$\begin{aligned}
2{\operatorname{Re}}\left[\left\langle S_P F_\alpha^\ast f_\alpha\right\rangle\right]=&2{\operatorname{Re}}\left[\sum_{\gamma\beta}f_\gamma^\ast {\mathbf{C}}_{\gamma\beta}{^{-1}}{\mathbf{C}}_{\beta\alpha} f_\alpha\right.\nonumber\\
&\left.+{ \varepsilon_\mathrm{true}^2}\sum_{\gamma\beta}f_\gamma^\ast {\mathbf{C}}_{\gamma\beta}{^{-1}}f_\beta f_\alpha^\ast f_\alpha\right].
\end{aligned}$$ In the first term we have $\sum_{\beta}{\mathbf{C}}_{\gamma\beta}{^{-1}}{\mathbf{C}}_{\beta\alpha}=\delta_{\gamma\alpha}$, and in the second term we find the definition of $R_P$. Thus, the second term of Eq. is $$2{\operatorname{Re}}\left[\left\langle S_P F_\alpha^\ast f_\alpha\right\rangle\right]=2\vert f_\alpha\vert^2 \left(1+{ \varepsilon_\mathrm{true}^2}R_P\right).$$
3. In the third term, we can again make use of Eq. : $$\begin{aligned}
\left\langle S_P^2\right\rangle&=\sum_{\alpha\beta\gamma\delta} {\operatorname{Re}}\left[f_\alpha^\ast{\mathbf{C}}_{\alpha\beta}{^{-1}}f_\gamma {\mathbf{C}}_{\gamma\delta}{^{-1}}\left\langle F_\beta F_\delta^\ast\right\rangle\right]\nonumber\\
&=\sum_{\alpha\beta\gamma\delta} {\operatorname{Re}}\left[f_\alpha^\ast{\mathbf{C}}_{\alpha\beta}{^{-1}}f_\gamma {\mathbf{C}}_{\gamma\delta}{^{-1}}{\mathbf{C}}_{\beta\delta}+{ \varepsilon_\mathrm{true}^2}f_\alpha^\ast{\mathbf{C}}_{\alpha\beta}{^{-1}}f_\gamma {\mathbf{C}}_{\gamma\delta}{^{-1}}f_\beta f_\delta^\ast\right]
\end{aligned}$$ In the first term, we have again $\sum_{\beta}{\mathbf{C}}_{\alpha\beta}{^{-1}}{\mathbf{C}}_{\beta\delta}=\delta_{\alpha\delta}$, and the second term is equal to $R_P^2$, such that $$\begin{aligned}
\left\langle S_P^2\right\rangle&=\sum_{\alpha\gamma} {\operatorname{Re}}\left[f_\alpha^\ast f_\gamma {\mathbf{C}}_{\gamma\alpha}{^{-1}}\right]+{ \varepsilon_\mathrm{true}^2}R_P^2=R_P+{ \varepsilon_\mathrm{true}^2}R_P^2
\end{aligned}$$
Recollecting 1.-3. and inserting into Eq. yields $$\begin{aligned}
\left\langle\left\vert F_\alpha-\frac{S_P}{R_P} f_\alpha\right\vert^2\right\rangle=&{\mathbf{C}}_{\alpha\alpha}+{ \varepsilon_\mathrm{true}^2}\vert f_\alpha\vert^2\nonumber\\
&-2\vert f_\alpha\vert^2 \left({ \varepsilon_\mathrm{true}^2}+\frac{1}{R_P}\right)+\left({ \varepsilon_\mathrm{true}^2}+\frac{1}{R_P}\right)\left\vert f_\alpha\right\vert^2\nonumber\\
&={\mathbf{C}}_{\alpha\alpha}-\frac{\vert f_\alpha\vert^2}{R_P}.\end{aligned}$$ As the power spectrum $P(k_\alpha)={\mathbf{C}}_{\alpha\alpha}$ is defined as the diagonal elements of the covariance matrix, the debiasing step is the same for a non-diagonal covariance matrix as for a diagonal one (cf. Sec. \[sec:debias\]), we just have to use the generalised definition of $R_P$ as in Eq. .
The independent mode approximation {#sec:diagCapprox}
----------------------------------
We have seen in Appendix \[sec:nondiag\] that [ mode]{} subtraction also works when the covariance matrix is non-diagonal. However, to compute the generalised $R_P$, one has to invert the full $N_\mathrm{mode}\times N_\mathrm{mode}$ covariance matrix, which makes this approach computationally almost as expensive as using the QML estimator. We will argue that, in most cases, Eq. provides a good estimate of the power, even in the presence of covariant modes, and we will provide a further correction term that corrects for using Eq. when off-diagonal covariances are important.
Suppose we apply Eq. assuming a diagonal covariance matrix, even though there are covariances between different modes. Then, we find a best fitting $$\varepsilon_\mathrm{BF}^\prime=\frac{\sum_\alpha\frac{F^\ast_\alpha f_\alpha}{P_\alpha}}
{\sum_\mu\frac{\vert f_\mu\vert^2}{P_\mu}}$$ instead of the true $$\varepsilon_\mathrm{BF}=\frac{\sum_{\alpha\beta}f_\alpha^\ast {\mathbf{C}}_{\alpha\beta}{^{-1}}F_\beta}{\sum_{\alpha\beta}f_\alpha^\ast {\mathbf{C}}_{\alpha\beta}{^{-1}}f_\beta}.$$ The expectations are the same $\left\langle\varepsilon_\mathrm{BF}^\prime\right\rangle=\left\langle\varepsilon_\mathrm{BF}\right\rangle=\varepsilon_\mathrm{true}$, but their variances are different. For the approximate estimate we have $$\begin{aligned}
\left\langle\varepsilon_\mathrm{BF}^{\prime 2}\right\rangle=&\frac{\left\langle \sum_{\alpha\beta}\frac{F^\ast_\alpha f_\alpha F_\beta f^\ast_\beta}{P_\alpha P_\beta}\right\rangle}{R_P^{\prime 2}}\nonumber\\
=&\frac{\sum_{\alpha\beta}\frac{ f_\alpha{\mathbf{C}}_{\alpha\beta} f^\ast_\beta}{P_\alpha P_\beta}}{R_P^{\prime 2}}+\frac{\sum_{\alpha\beta}\frac{\varepsilon_\mathrm{true}^2 \left\vert f_\alpha\right\vert^2 \left\vert f_\beta\right\vert^2}{P_\alpha P_\beta}}{R_P^{\prime 2}}\nonumber\\
=&\frac{1}{R_P^{\prime 2}}\sum_{\alpha\beta}\frac{ f_\alpha{\mathbf{C}}_{\alpha\beta} f^\ast_\beta}{P_\alpha P_\beta}+\varepsilon_\mathrm{true}^2.\end{aligned}$$ Unlike in the previous estimates, the covariance matrix does not cancel in the first term. Similarly, $$\begin{aligned}
\left\langle\varepsilon_\mathrm{BF}^\prime F_\alpha^\ast f_\alpha\right\rangle=&\frac{1}{R_P^\prime}\left\langle\sum_\beta\frac{f_\beta^\ast F_\beta F_\alpha^\ast f_\alpha}{P_\beta}\right\rangle\nonumber\\
=&\frac{1}{R_P^{\prime}}\sum_{\beta}\frac{ f_\alpha{\mathbf{C}}_{\alpha\beta} f^\ast_\beta}{P_\beta}+\varepsilon_\mathrm{true}^2\left\vert f_\alpha\right\vert^2.\end{aligned}$$ Combining the previous two equations, we obtain $$\begin{aligned}
\left\langle\left\vert F_\alpha-\varepsilon_\mathrm{BF}^\prime f_\alpha\right\vert^2\right\rangle=C_{\alpha\alpha}-\frac{2}{R_P^\prime}\sum_{\beta}\frac{ f_\alpha{\mathbf{C}}_{\alpha\beta} f^\ast_\beta}{P_\beta}+\frac{\vert f_\alpha\vert^2}{R_P^{\prime 2}}\sum_{\gamma\beta}\frac{ f_\gamma{\mathbf{C}}_{\gamma\beta} f^\ast_\beta}{P_\gamma P_\beta}.\end{aligned}$$ Splitting the covariance matrix $$C_{\alpha\beta}=P_\beta\left(\delta_{\alpha\beta}+\Delta_{\alpha\beta}\right)$$ into a diagonal and off-diagonal elements yields $$\begin{aligned}
\left\langle\left\vert F_\alpha-\varepsilon_\mathrm{BF}^\prime f_\alpha\right\vert^2\right\rangle=P_\alpha-\frac{\vert f_\alpha\vert^2}{R_P^\prime}\left[1+\sum_{\gamma\beta} f_\gamma\Delta_{\gamma\beta} f^\ast_\beta\left(\frac{2\delta_{\alpha\gamma}}{\vert f_\alpha\vert^2}
-\frac{1}{R_P^{\prime}P_\gamma}\right)\right].\end{aligned}$$ Hence, one can perform [ mode]{} subtraction assuming a diagonal covariance matrix and then apply another correction term which is linear in its off-diagonal elements. The advantage of this procedure is that it does not require any inversion of the $N_\mathrm{mode}^2$ covariance matrix. If the off-diagonal elements are small, then the bias correction reverts back to the form of Eq. .
\[lastpage\]
[^1]: E-mail: [email protected]
[^2]: www.euclid-ec.org
[^3]: www.skatelescope.org
[^4]: writing $f_\alpha\equiv f({{\bmath{k}}}_\alpha)$ and $P_\alpha\equiv P(k_\alpha)$ to save space
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Two high performance coronagraphic approaches compatible with segmented and obstructed telescope pupils are described. Both concepts use entrance pupil amplitude apodization and a combined phase and amplitude focal plane mask to achieve full coronagraphic extinction of an on-axis point source. While the first concept, named Apodized Pupil Complex Mask Lyot Coronagraph (APCMLC), relies on a transmission mask to perform the pupil apodization, the second concept, named Phase-Induced Amplitude Apodization complex mask coronagraph (PIAACMC), uses beam remapping for lossless apodization. Both concepts theoretically offer complete coronagraphic extinction (infinite contrast) of a point source in monochromatic light, with high throughput and sub-$\lambda$/D inner working angle, regardless of aperture shape. The PIAACMC offers nearly 100% throughput and approaches the fundamental coronagraph performance limit imposed by first principles. The steps toward designing the coronagraphs for arbitrary apertures are described for monochromatic light. Designs for the APCMLC and the higher performance PIAACMC are shown for several monolith and segmented apertures, such as the apertures of the Subaru Telescope, Giant Magellan Telescope (GMT), Thirty Meter Telescope (TMT), the European Extremely Large Telescope (E-ELT) and the Large Binocular Telescope (LBT). Performance in broadband light is also quantified, suggesting that the monochromatic designs are suitable for use in up to 20% wide spectral bands for ground-based telescopes.'
author:
- Olivier Guyon
- 'Philip M. Hinz'
- Eric Cady
- Ruslan Belikov
- Frantz Martinache
bibliography:
- 'ms.bib'
title: High Performance Lyot and PIAA Coronagraphy for Arbitrarily shaped Telescope Apertures
---
Introduction {#sec:intro}
============
Direct imaging of exoplanets with ground-based telescopes is becoming possible thanks to advances in adaptive optics, as demonstrated by several recent direct imaging exoplanet discoveries [@2010Sci...329...57L; @2008Sci...322.1348M; @2013ApJ...763L..32C]. While current ground-based instruments are most sensitive to relatively massive and young planets at large angular separation (typically beyond a few tenths of an arcsecond), recent developments in coronagraphic techniques, “extreme” Adaptive Optics and calibration techniques are pushing detection limits deeper in contrast and closer in angular separation, soon providing access to the planet-rich inner parts of planetary systems [@2008SPIE.7015E..31M; @2008SPIE.7014E..41B; @2009SPIE.7440E..20M; @2011ApJ...729..132C]. Direct imaging of the inner part (1 to 5 AU) of young planetary systems is of especially high scientific value to constrain and understand planetary systems formation and evolution near the habitable zone, and requires the combination of an efficient coronagraph offering small inner working angle and a high level of wavefront correction and calibration.
High contrast imaging from space allows access to considerably better contrast than possible with ground-based telescopes, thanks to the absence of atmospheric turbulence. Laboratory coronagraphy systems have demonstrated that raw contrasts of about 1e-9 can be achieved in a stable environment with a deformable mirror and a coronagraph (see for example [@2007Natur.446..771T]). At such high contrast, coronagraphic imaging can allow characterization of potentially habitable planets through spectroscopy from space [@2009arXiv0911.3200L].
While most high performance coronagraphs are designed for unobstructed circular pupils, current and future large ground-based telescopes are centrally obscured, and also segmented above 8.4-m diameter. Future large space-based telescopes will also likely be centrally obscured and/or segmented, although a telescope dedicated to high contrast imaging could be built off-axis if required for coronagraphy [@2009arXiv0911.3200L]. The scientific return of an exoplanet direct imaging mission or instrument is a steep function of telescope diameter: larger telescopes allow access to exoplanets at smaller angular separations, which are brighter in reflected light (apparent luminosity scales as inverse square of angular separation in reflected light), more numerous (the number of planets of a given type accessible with a telescope scales as the third power of telescope diameter), and more relevant to exoplanet systems habitability than widely separated planets. Larger ground-based telescope size also allows higher contrast observation by better concentrating planet light over the speckle halo background, and the gain in collecting area enables spectroscopic characterization. It is therefore essential to identify and develop coronagraph concepts which can deliver high performance on centrally obscured and/or segmented apertures.
Coronagraph designs for centrally obscured and/or segmented apertures have been proposed by several authors, offering a wide range of solutions and approaches:
- [[**Lyot-type coronagraphs with amplitude masks.**]{} Most studies of coronagraph designs for obscured and/or segmented apertures considered Lyot-type coronagraph optimized for high contrast by either apodization of the entrance pupil (APLC concept introduced by [@2003AA...397.1161S]) or apodization of the focal plane mask (Band-limited coronagraph concept introduced by [@2002ApJ...570..900K]). For the apodized pupil Lyot coronagraph (APLC), [@2005ApJ...618L.161S] and [@2009ApJ...695..695S] showed that the entrance pupil apodizer can be optimized for centrally obscured pupils. Using this technique, [@2007AA...474..671M] studied the APLC for ELTs, finding high throughput solutions offering better than 1e-5 contrast at and beyond 3 $\lambda$/D separation. [@2010AA...519A..61M] proposed using a multistage apodized pupil Lyot coronagraph (APLC) to mitigate central obstruction limitations. While central obstruction can be mitigated in the Lyot-type coronagraph design, [@2005ApJ...633..528S; @2005ApJ...626L..65S] showed that spiders and gaps in APLC and band-limited Lyot coronagraphs diffract light within the geometrical aperture, making it impossible to achieve very high contrast on segmented apertures. Moderate-contrast band-limited Lyot coronagraphs have been designed for the NIRCAM instrument [@2009SPIE.7440E..28K] on the James Webb Space Telescope, but the aggressive Lyot stop designs, which remove much of the residual diffraction from the secondary structure and the segmented primary, come at the cost of significant companion throughput.]{}
- [[**Phase mask coronagraphs.**]{} Coronagraphs using phase focal plane masks are also affected by central obscuration and spiders/gaps. [@2003SPIE.4860..171L] showed that the 4-quadrant phase mask can only achieve full coronagraphic suppression on unobstructed pupils free of gaps or spiders, as any obscuration diffracts light outward in the Lyot plane. The optical vortex coronagraph is similarly affected by obscurations, although [@2011OptL...36.1506M] showed that central obstruction can be mitigated by a dual-stage approach, where the second stage rejects most of the light diffracted by the central obstruction. For both the vortex and the 4 quadrant phase mask coronagraphs, no solution has been found to eliminate the light diffracted by spiders and gaps. ]{}
- [[**Shaped Apertures.**]{} For moderate contrast level and relatively large IWA, shaped apertures can be designed for centrally obscured and segmented pupils. [@2006PASJ...58..627T] designed shaped apertures delivering 1e-7 contrast at 4 $\lambda$/D. Similarly, [@2011arXiv1108.4050C] showed that shaped pupil can be designed for 1e-6 contrast and $\approx 4 \lambda$/D inner working angle for a variety of centrally obscured and segmented apertures. ]{}
A different approach to this problem is to remap the entrance aperture to remove central obstruction and/or spiders. [@2005PASP..117..295M] propose a 2-mirror system to remove central obstruction and spiders for a four-quadrant coronagraph. [@2006PASP..118..860G] propose a high efficiency nulling coronagraph concept adapted to central obstruction and spiders by performing destructive interferences between pairs of unobstructed off-axis subapertures. [@2009PASP..121.1232L] demonstrate that a prism-like transmissive device and aspheric optics can be used to remove both central obstruction and spiders from the Subaru Telescope pupil, theoretically allowing high performance coronagraphy with the full telescope aperture. These remapping solutions are complex, challenging to implement and align, and extremely sensitive to tip-tilt and stellar angular size [@2009ApJ...702..672C] at high contrast: when points on either size of an obstruction are brought next to each other in the remapped pupil, a small tip-tilt in the entrance beam leads to a phase discontinuity in the remapped beam. When due to finite stellar angular diameters, diffraction due to this discontinuity cannot be mitigated or controlled by wavefront control, as it is incoherent (opposite sides of the stellar disk produce diffracted light components of opposite signs). [@2007ApJ...658.1386S] chose to avoid entirely the problem by using an unobstructed 1.5-m diameter off-axis part of the 5-m Palomar telescope to perform high contrast imaging with the optical vortex coronagraph. While this allows the use of high performance coronagraphs designed for unobstructed apertures, the performance loss due to the use of an aperture considerably smaller than the full telescope is significant.
The solutions previously proposed to mitigate the effects of central obstruction, spiders and gaps are generally suitable for ground-based coronagraphy at a few $\lambda$/D IWA, with a raw contrast around $10^{-5}$, as reported by [@2008AA...492..289M] who performed a study of coronagraphic performance on ELTs including realistic assumptions on the level of residual wavefront error after an extreme-AO system. For most of the coronagraphs, central obstruction and spiders were found to have a major impact on performance, limiting the achievable contrast to $10^{-4}$ in the 1 to 4 $\lambda$/D separation range. The notable exceptions to this rule were the Achromatic Interfero Coronagraph , which is insensitive to centro-symmetric pupil features (such as a central obstruction or a set of four radial spiders at 90 deg), and the APLC, which could be designed to take into account central obstruction and was found to be quite robust to spiders at the $10^{-5}$ contrast level. The coronagraphs concepts for which ground-based designs compatible with central obscuration have been proposed (shaped aperture, APLC, band-limited Lyot coronagraph) are unfortunately not able to offer IWA less than $\approx 2 \lambda/D$, and also do not enable high contrast (approximately $10^{9}$) coronagraphy on centrally obscured or segmented apertures.
The work presented in this paper is aimed at demonstrating that high performance coronagraphy is possible in monochromatic light on centrally obscured and/or segmented pupils for both ground-based and space-based telescopes. The Apodized Pupil Complex Mask Lyot Coronagraph (APCMLC) and Phase-Induced Amplitude Apodization complex mask coronagraph (PIAACMC) concepts, previously described for circular unobstructed apertures in [@2010ApJS..190..220G], are here adapted to arbitrarily shaped apertures. Section \[sec:APCMLC\] describes how the APCMLC can be adapted to non-circular apertures, and a step by step process to design a APCMLC for any aperture shape is proposed and examples are shown. In Section \[sec:PIAACMC\], the PIAACMC is shown to offer performance superior to the APCMLC, and its design for centrally obscured and segmented apertures is discussed, with examples representative of current and future large telescopes shown. High performance APCMLC and PIAACMC for pupils with strong aspect ratios is briefly discussed in Section \[sec:aspectratio\]. Chromaticity of the concepts is discussed in Section \[sec:chrom\]. Results are discussed in Section \[sec:conclusion\].
[lcccc]{}
\
Subaru APCMLC \#1 & 0.596 $\lambda/D$ & 99.62% & 68.88% & 0.71 $\lambda/D$\
Subaru APCMLC \#2 & 0.8 $\lambda/D$ & 24.89% & 54.65% & 0.90 $\lambda/D$\
Subaru APCMLC \#3 & 1.2 $\lambda/D$ & 8.57% & 39.19% & 1.30 $\lambda/D$\
\
GMT APCMLC \#1 & 0.666 $\lambda/D$ & 99.64% & 64.50% & 0.78 $\lambda/D$\
GMT APCMLC \#2 & 0.7 $\lambda/D$ & 79.47% & 61.99% & 0.81 $\lambda/D$\
GMT APCMLC \#3 & 1.2 $\lambda/D$ & 35.16% & 11.39% & 1.28 $\lambda/D$\
GMT APCMLC \#4 & 1.5 $\lambda/D$ & 28.59% & 9.68% & 1.58 $\lambda/D$\
\
TMT APCMLC \#1 & 0.764 $\lambda/D$ & 99.72% & 55.67% & 0.86 $\lambda/D$\
TMT APCMLC \#2 & 0.8 $\lambda/D$ & 85.48% & 53.23% & 0.90 $\lambda/D$\
TMT APCMLC \#3 & 1.2 $\lambda/D$ & 36.08% & 34.26% & 1.27 $\lambda/D$\
TMT APCMLC \#4 & 1.5 $\lambda/D$ & 40.94% & 28.41% & 1.57 $\lambda/D$\
\
E-ELT APCMLC \#1 & 0.825 $\lambda/D$ & 99.85% & 54.26% & 0.93 $\lambda/D$\
E-ELT APCMLC \#2 & 0.9 $\lambda/D$ & 78.76% & 49.92% & 1.00 $\lambda/D$\
E-ELT APCMLC \#3 & 1.2 $\lambda/D$ & 50.56% & 38.43% & 1.29 $\lambda/D$\
Apodized pupil complex mask Lyot Coronagraph (APCMLC) for apertures of arbitrary shape {#sec:APCMLC}
======================================================================================
Principle {#ssec:APCMLCprinc}
---------
In this section, it is shown that the APCMLC is compatible with non-circular apertures, as illustrated in Figure \[fig:APCMLCprinciple\], and a description of how it can be designed for arbitrarily shaped pupils is provided. While the APCMLC description provided here remains qualitative and focused on aspects relevant to non-circular apertures, a more complete analytical description is provided in [@2010ApJS..190..220G] for circular unobstructed apertures. We describe the APCMLC by following how electric field (also refered to as complex amplitude in this paper) from an on-axis point source propagates through the coronagraph system.
The APCMLC, illustrated in Figure \[fig:APCMLCprinciple\], uses a circular focal plane mask to partially transmit and phase shift the on-axis point spread function (PSF) core (complex amplitude B on Figure \[fig:APCMLCprinciple\]). The transmission and phase shift are uniform within the mask radius, and the mask is fully transmissive, with no phase shift, outside this radius. This produces a destructive interference within the geometric pupil, between the light that passes around the focal plane mask disk and the phase-shifted light passing through the focal plane phase-shifting disk. With a Lyot mask (Lmask) selecting only the geometric pupil, a coronagraphic effect is achieved. The concept is thus an intermediate point between the conventional Lyot coronagraph or Apodized Pupil Lyot Coronagraph (APLC) [@2003AA...397.1161S], which use a large fully opaque focal plane mask, and the phase mask coronagraph [@1997PASP..109..815R; @2000SPIE.4006..377G; @1999PASP..111.1321G; @2010AA...509A...8N] which uses a small size fully transmissive phase-shifting focal plane mask. In the APCMLC, the focal plane mask size can be chosen anywhere between these two extremes, and defines the ratio between the amount of light within the circular mask and outside the mask. As the focal plane mask radius decreases, a smaller fraction of the light is within the mask radius, and its transmission must increase to maintain the flux balance between the “inside focal plane mask” (corresponding to low spatial frequencies in the pupil) and “outside focal plane mask” (high spatial frequencies in the pupil) components, a necessary condition to achieve destructive interference.
Full destructive interference within the geometric pupil also requires that the two components are equal in amplitude for every point within the pupil. Since this match does not naturally occur, all three concepts (APLC, Roddier phase mask coronagraph and APCMLC) require the entrance pupil to be amplitude apodized to enforce this match. Qualitatively, for small focal plane mask size, the apodization mostly changes the pupil light distribution for the “outside focal plane mask” component, while the light distribution for the “inside focal plane mask” component is mostly driven by the size of the focal plane mask. The entrance pupil apodization can therefore be iteratively derived to force the “outside focal plane mask” component to match the “inside focal plane mask component”, using the following steps:
1. [Adopt a focal plane mask diameter $a$]{}
2. [Compute the on-axis complex amplitude PSF for the aperture. This is the Fourier transform of the pupil complex amplitude P]{}
3. [Clip the PSF: values outside the focal plane mask radius are forced to zero]{}
4. [Inverse-Fourier transform the clipped PSF, and adopt this function as the apodized pupil plane amplitude function A, after multiplication by a factor $\Lambda_a$ so that its maximum value across the pupil is be equal to 1 (full transmission)]{}
5. [Return to step (2), with the output of step (4) as the pupil complex amplitude function]{}
This iterative algorithm is a generalization of the iterative algorithm used by [@2000SPIE.4006..377G; @2002AA...391..379G] and detailed in @guyonPhD to compute optimal apodization for the phase mask coronagraph (for which the mask is fully transmissive), and the iterative algorithm used to compute optimal apodization for the APLC (for which the mask is fully opaque) on centrally obscured circular apertures [@2005ApJ...618L.161S; @2010AA...520A.110M] and on more complex pupil shapes [@2009ApJ...695..695S]. showed that the apodization solutions obtained for rectangular and circular apertures are Prolate functions for which analytical expressions exist. Apodization functions can also be computed for centrally obscured apertures [@2005ApJ...618L.161S], and for arbitrary non circular symmetric pupils [@2009ApJ...695..695S]. A remarkable property of the iterative algorithm described above is that it converges for a wide range of pupil shapes and focal plane mask diameters [@guyonPhD]. For small focal plane mask diameters, convergence is due to the fact that modifying the entrance aperture light distribution predominantly affects the “outside focal plane mask” light component. Exact apodization solutions for the APLC and APCMLC therefore exist for most aperture geometries and focal plane mask diameters. An example APCMLC design on a non-circular aperture, for which the entrance pupil apodization function was computed using the iterative algorithm described in this section, is shown in Figure \[fig:APCMLCprinciple\].
The APCMLC is described here analytically for monochromatic light using notations shown in Figure \[fig:APCMLCprinciple\]. The entrance pupil shape is defined by the real function $P(\mathbf{r})$, with $\mathbf{r}$ the 2-D position vector in the pupil plane, and $P(\mathbf{r})=1$ for points within the pupil and $P(\mathbf{r})=0$ outside of the pupil. The apodizer function $Apo(\mathbf{r})$ is applied to the pupil, yielding the following complex amplitude in plane $A$:
$$\Psi_A(\mathbf{r}) = Apo(\mathbf{r}) P(\mathbf{r})$$
The $\mathbf{r}$-dependence is dropped in subsequent equations. The iterative algorithm previously described is used to numerically compute the apodization function $Apo(\mathbf{r})$, which will converge to a pupil function $\psi_a$ which is the eigenvector of the “truncate (by P), Fourier Transform, truncate (by $|\mathbf{r}|<a$), and inverse Fourier Transform” operator, with eigenvalue equal to the scaling factor $\Lambda_a$ used in step (4) of the iterative algorithm given previously.
$$\label{equ:La}
( \psi_a P ) \otimes \widetilde{M_a} = \Lambda_a \psi_a$$
where $\otimes$ is the convolution operator, $M_a$ is is equal to 1 within a disk of diameter $a$ and is equal to 0 outside it, and $\widetilde{M_a}$ is the Fourier Transform of $M_a$. In the APCMLC, the apodizer function is chosen equal to $\psi_a$:
$$\Psi_A = \psi_a P$$
The focal plane mask complex amplitude is :
$$F_{mask} = 1 - (1-t) M_a$$
where $t$ is the complex amplitude transmission within the circular focal plane mask. The complex amplitude in plane B is :
$$\Psi_B = F_{mask} \widetilde{\Psi_A} = \widetilde{\Psi_A} - (1-t) M_a \widetilde{\Psi_A}$$
The complex amplitude in plane C is obtained by Fourier transform of $\Psi_B$: $$\label{equ:PsiC0}
\Psi_C = \Psi_A - (1-t) (\psi_a P) \otimes \widetilde{M_a}$$
By combining equations \[equ:La\] and \[equ:PsiC0\], and multiplying by $P(r)$, the complex amplitude in plane C within the geometrical pupil is:
$$\label{equ:PsiC}
\Psi_C P(\mathbf{r}) = P(\mathbf{r}) \times (1- (1-t) \Lambda_a) \psi_a$$
This equation shows that, if $t = 1-\Lambda_a^{-1}$ (this value is now noted $t_a$), then $\Psi_C$ is equal to zero within the geometric pupil. This is the condition for a APCMLC, which completely removes light from an on-axis point source, provided that a Lyot pupil plane mask $L_{mask}(\mathbf{r})=P(\mathbf{r})$ is used to only select light within the geometric pupil. Since $\Lambda_a < 1$, $t_a$ is negative: the focal plane mask is both partially transmissive and π-phase shifting. A coronagraphic solution requires $t_a>-1$, and therefore exists only if $\Lambda_a > 0.5$: the focal plane size needs to be sufficiently large so that light going through the mask can be balanced with light going outside the mask.
The same pupil apodization technique is used in the Apodized Pupil Lyot Coronagraph (APLC) to optimize the pupil entrance complex amplitude to the hard edged opaque focal plane mask [@2003AA...397.1161S]. In the APLC, $t=0$ in equation \[equ:PsiC\], and the coronagraphic extinction is therefore not total for an on-axis point source. Equation \[equ:PsiC\] shows that the on-axis PSF in the final focal plane mask is an exact copy of the non-coronagraphic PSF, scaled by $(1-\Lambda_a)^2$ in intensity. For large focal plane masks diameter $a$, $\Lambda_a$ is close to 1, and the coronagraphic extinction is satisfactory. The APLC concept has been adopted for the Palomar Observatory high contrast imaging program [@2011PASP..123...74H] and the Gemini Planet Imager [@2008SPIE.7015E..31M] and has been validated in laboratory demonstrations [@2011AJ....142..119T]. The APCMLC concept is very similar to the APLC, the only fundamental difference being that its focal plane mask transmission is allowed to be non-zero, therefore allowing full coronagraphic extinction for any focal plane mask size $a$ for which $\Lambda_a > 0.5$.
APCMLC designs for segmented apertures {#ssec:APCMLCdesigns}
--------------------------------------
Apodized pupil complex mask Lyot coronagraphs (APCMLCs) were designed for the Subaru Telescope, Giant Magellan Telescope (GMT), Thirty Meter Telescope (TMT) and European Extremely Large Telescope (E-ELT) pupil geometries, following the process described in the previous section. For each pupil geometry, several focal plane mask sizes were chosen. The designs with the smallest possible focal plane mask sizes use full transmission $\pi$-phase shifting circular focal plane masks, and are referred to as optimal IWA APCMLC designs in this paper. As the focal plane mask size increases, it also becomes more opaque, the system throughput (which is equal to the apodizer throughput) decreases and the IWA increases. Results are summarized in Table \[tab:APCMLC\], and show that optimal IWA designs offer IWAs around 0.9 $\lambda/D$ and throughputs around 60%. For all designs, the IWA is approximately equal to the focal plane mask radius, and the throughput decreases rapidly with increasing focal plane mask size: with a 1.2 $\lambda/D$ radius, the throughput ranges from approximately 10% to 35% depending on the pupil geometry. The performance of the optimal IWA design is largely independent of pupil geometry, and is similar for segmented apertures to the performance previously reported for a full unobstructed circular pupil [@2010ApJS..190..220G]. However, as the mask size increases, pupil geometry has a larger impact on performance, as the range of pupil plane spatial frequencies accessed by the focal plane mask begins to overlap with the low spatial frequency components of the pupil geometry (central obstruction, large segments, thick spider vanes). This difference is most noticeable between the GMT pupil with few large segments and the TMT or EELT geometries with numerous small segments.
Selected examples of apodization functions and Lyot plane intensity images are shown in Figure \[fig:APCMLCexamples\]. In each case, the apodization function is smooth and free of high spatial frequencies, and no light is left within the geometric pupil in the Lyot pupil plane, as all residual starlight is diffracted outside of the pupil and in the gaps between segments.
Table \[tab:APCMLC\] gives for several APCMLC designs the key design parameters (focal plane mask size $a$, focal plane mask transmission) as well as the coronagraph performance (throughput and IWA). For each pupil shape considered, the first design (design \# 1) is the most aggressive in IWA, with a nearly fully transmissive focal plane mask. This aggressive design is also the one with the highest throughput, as the apodization strength needs to increase for larger focal plane mask sizes. The APCMLC throughput never exceeds 70% due to the need for a pupil apodization. Transmission curves are given in Figure \[fig:APCMLCtransm\] for the APCMLC designs listed in Table \[tab:APCMLC\].
Phase Induced Amplitude Apodization Complex Mask Coronagraph (PIAACMC) for apertures of arbitrary shape {#sec:PIAACMC}
=======================================================================================================
Lossless Phase-Induced Amplitude Apodization (PIAA)
---------------------------------------------------
While the APCMLC described in Section \[sec:APCMLC\] achieves full on-axis coronagraphic extinction for almost any pupil shape, its throughput is limited by the entrance apodization required to achieve perfect destructive interference in the output pupil plane. The system throughput decreases as the focal plane mask size increases, with a maximum throughput equal to 72% for a 0.64 $\lambda/D$ radius purely phase-shifting transparent mask on a circular unobstructed aperture. Throughput, and consequently angular resolution, degrade rapidly with increased focal plane mask size: it is 18% for a 2 $\lambda/D$ radius mask. The results obtained in Section \[ssec:APCMLCdesigns\] also show that the APCMLC maximum throughput (achieved for the designs with the most aggressive IWA) is lower on segmented pupils than it is for an unobstructed circular pupil (“Throughput” column of Table \[tab:APCMLC\]). Moreover, throughput, angular resolution and IWA are significantly degraded when the focal plane mask size is increased - while mitigation of undesired chromatic effects at the focal plane mask may require a larger and more opaque mask.
Phase-induced Amplitude Apodization (PIAA) uses aspheric mirrors to achieve a lossless beam apodization [@2003AA...404..379G], and can therefore produce a highly apodized beam suitable for high contrast imaging without the angular resolution loss and throughput loss of a conventional apodizer. PIAA can also be used to replace the entrance apodization in the APCMLC design described in Section \[sec:APCMLC\], as previously proposed for unobstructed circular pupils [@2010ApJS..190..220G]. The resulting coronagraph, denoted Phase-induced Amplitude Apodization Complex mask coronagraph (PIAACMC), offers simultaneously full throughput, small inner working angle and total on-axis extinction.
An example PIAACMC design is shown in Figure \[fig:PIAACMCprinc\] for a segmented centrally obscured pupil. The entrance pupil P (image shown in the lower left of the figure) is apodized with lossless aspheric PIAA optics. Because the PIAA optics perform apodization by remapping instead of selective transmission, the resulting pupil P1 shape is modified. A conventional apodizing mask may be used to fine-tune the apodization if the PIAA optics do not exactly produce the required amplitude distribution (this will be addressed in the following section). The resulting pupil A is shown in the second image from the lower left corner. The image of an on-axis point source is shown in the center image, where the phase-shifting partially transmissive focal plane mask is inserted. In the output pupil plane C, all light within the pupil has been removed, while diffracted starlight fills the gap and obstructions of the segmented pupil. A Lyot mask (noted Lmask) can then select only the geometric pupil area (after remapping) to fully block on-axis starlight while fully transmitting the light from distant off-axis source. A well-documented side-effect of apodization with PIAA optics is that off-axis PSFs are highly distorted, and corrective optics (inverse PIAA) are required at the output of the coronagraph to maintain diffraction limited sharp PSFs over a scientifically useful field of view [@2009PASP..121.1232L]. Except for PIAA and inverse PIAA optics, the PIAACMC architecture is functionally identical to the APCMLC architecture described in Section 2.1: between planes P1 (output of the PIAA optics) and the plane immediately after the pupil plane Lyot mask, the architecture is an APCMLC. The main difference between APCMLC and PIAACMC is that the lossless apodization allows increased performance by maintaining full throughput and angular resolution, regardless of the focal plane mask size adopted.
Designing a PIAACMC for a non circular aperture
-----------------------------------------------
We consider in this work PIAACMC designs that perform a lossless PIAA apodization of the pupil to produce a generalized prolate function for the aperture. We note that other apodization functions could be adopted, and could potentially lead to superior performance, but this is not explored in this paper. In the unobstructed circular pupil case [@2010ApJS..190..220G], designing the PIAACMC is relatively simple, as PIAA apodization using a radial remapping function preserves the circular aperture shape. The prolate function can thus be first computed, and then realized with a radial PIAA apodization.
Designing a PIAACMC for complex shaped apertures is considerably more challenging because the PIAA apodization modifies the aperture shape, which itself changes the generalized prolate function. In addition to this circular problem, if the aperture is not circularly symetric, the generalized prolate is also not symmetric, and the required remapping function therefore cannot be written as a radial function. While PIAA optics can be designed for any radial remapping [@2003AA...404..379G], an arbitrarily chosen 2D remapping function can almost never be realized with a set of two PIAA optics.
To overcome the two challenges listed above (circular design problem due to effect of PIAA on aperture shape, and complexity/impossibility of designing PIAA optics for non-circular symmetric remapping), a hybrid PIAACMC design is adopted, which includes a conventional apodizer after the remapping to produce the required prolate function. Thanks to this post-apodizer, the output of the PIAA apodization does not need to exactly match the generalized prolate function, allowing radial remapping functions to be used on non-circular symmetric apertures. The goal of the design optimization is to bring the PIAA apodization and generalized prolate functions close, in order to minimize the strength of the post-apodizer and thus maintain a high system throughput.
The proposed steps for designing a PIAACMC for complex shaped apertures are :
1. [Choose radial remapping function $r_1=f_b(r_0)$, where $r_0$ and $r_1$ are the radial coordinate in the input (before remapping) and output(after remapping) pupils respectively. For convenience, the remapping function is selected among a pre-computed set of functions used to produce prolate spheroidal apodizations on circular apertures. The focal plane mask diameter corresponding to the prolate spheroidal function is denoted $b$, and the corresponding remapping function and apodization intensity profile are respectively $f_b$ and $I_b$.]{}
2. [Apply the remapping function to the entrance pupil. The remapping transforms the entrance pupil intensity P(x,y) into P1(x,y).]{}
3. [Choose a focal plane mask diameter $a$.]{}
4. [Compute the generalized prolate function $Prola(x,y)$ for the remapped aperture shape defined by $P1(x,y)>0$, using the focal plane mask diameter $a$. This is done iteratively as described in section \[ssec:APCMLCprinc\]]{}
5. [Compute the amplitude ratio $Apo(x,y)=Prola(x,y)/P1(x,y)$. This is the post-apodizer amplitude transmission function. $Apo(x,y)$ is then scaled to ensure that its maximum value is equal to 1. The intensity-weighted average of $Apo(x,y)^2$ defines the coronagraph throughput for off-axis sources.]{}
Steps (3) to (5) are repeated for different values of the focal plane mask size $a$. The off-axis coronagraph throughput is computed for each choice of $a$, and the final focal plane mask size is chosen to maximize throughput. This optimization links the choice of the remapping function (step (1)) to a value of the focal plane mask radius a. For a circular unobstructed aperture, the solution would be $a=b$, for which the PIAA apodization would perfectly match the generalized prolate function. On arbitrarily shaped pupils, the focal plane mask radius is usually close to, but not equal to, $b$. Stronger apodization functions correspond to larger values for $a$ and $b$.
![\[fig:PIAACMC\_Subaru\_PSFs\] Simulated Subaru PIAACMC images of 5 point sources of equal brightness. The point sources are at coordinates (0;0), (1;0), (0;2), (-4;0) and (0;-8) in $/lambda/D$ units. A non-coronagraphic images (left) shows all five point sources. The partially transmissive central focal plane mask is visible in the intermediate focal plane image (center), where off-axis PSFs are distorted by the PIAA remapping. In the output focal plane image (right), the central source is fully canceled and the off-axis PSFs images are sharp thanks to the inverse PIAA optics. ](fig06.eps)
PIAACMC design examples
-----------------------
### Centrally obscured pupils: Subaru Telescope pupil
The Subaru telescope pupil is representative of current large aperture astronomical telescopes, with a large central obstruction and thick spiders. Both features must be taken into account for the design of a high performance coronagraph.
Figure \[fig:PIAACMC\_Subaru\] shows two PIAACMC designs for the Subaru Telescope pupil. The small IWA design (left) was computed from the $b/2=0.6 \lambda/D$ beam remapping, and uses a small sub-$\lambda/D$ radius focal plane mask with high transmission. The large IWA design was computed from $b/2=1.2 \lambda/D$, adopts a larger mostly opaque focal plane mask, and relies on a stronger PIAA remapping. Both designs offer throughput above 97%, and their throughput could be further increased by slightly elongating the focal plane mask, which was kept circular for simplicity in this study. The large IWA design, by relying on a stronger PIAA remapping, introduces a large pupil deformation, as visible in the figure. The post-focal plane mask pupil images demonstrate the PIAACMC’s ability to diffract all of the light from a central source outside the geometrical pupil, including within the gaps of the pupil (here, central obstruction and spiders).
Figure \[fig:PIAACMC\_Subaru\_PSFs\] shows intensity images of a field consisting of five equally bright point sources. The left images are obtained without a coronagraph, and simply show the imaging quality of the Subaru pupil in the absence of wavefront aberrations. The center column shows images in plane B of Figure \[fig:PIAACMCprinc\], immediately after the focal plane mask. The focal plane mask in the low IWA design (top) is more transmissive, and is also physically smaller. The large IWA design (bottom) introduces large off-axis aberrations due to the stong remapping. In the final coronagraphic images (right column), the central source is perfectly removed, and the images of the off-axis sources are sharp thanks to the inverse-PIAA optics.
### Segmented pupils: Giant Magellan Telescope (GMT)
The Giant Magellan Telescope (GMT) consists of one central 8.4-m circular segment, surrounded by a ring of six 8.4-m diameter segments. While the outer segments are unobscured, the central segment includes a central obstruction due to the secondary mirrors and its support structure.
Figure \[fig:PIAACMC\_GMT\] show three PIAACMC designs for the GMT pupil: a small IWA design computed for $b/2=0.7 \lambda/D$ (design \#1), a medium IWA design computed for $b/2=1.2 \lambda/D$ (design \#2), and a large IWA design computed for $b/2=1.5 \lambda/D$ (design \#3). As $b$ increases, the PIAA remapping becomes stronger, and the physical size of the focal plane mask increases. In each case, the PIAACMC achieves complete suppression of the on-axis point source, and its light is diffracted outside the geometrical aperture in plane C, including between the seven subapertures and within the secondary mirror obstruction and support structure.
The imaging quality of the GMT PIAACMC designs is illustrated in the right panel of Figure \[fig:PIAACMC\_GMT\], which shows that for each of the three designs, the final coronagraphic image maintains high thoughput and largely uncompromized imaging quality outside the central $\approx 1 \lambda/D$ region. The images also show that off-axis aberrations are stronger as the design relies more on PIAA remapping, although these aberrations are well corrected by the inverse PIAA system.
### Highly segmented pupils: European Extremely Large Telescope (E-ELT) and Thirty Magellan Telescope (TMT)
Figures \[fig:PIAACMC\_EELT\] and \[fig:PIAACMC\_TMT\] each show three PIAACMC designs for the European Extremely Large Telescope (E-ELT) and the Thirty Meter Telescope (TMT) pupil geometries. Both pupils consist of a large number of small segments, a central obstruction and spider vanes. Each of the six designs achieves total rejection of a central point souce with high system throughput (between 97.8 % and 99.98%) for off-axis sources. The inner working angle ranges from $\approx$ 0.8 $\lambda$/D for the most aggressive designs (designs \#1) to $\approx$ 1.0 $\lambda$/D for the more conservative designs (designs \#3). Figures \[fig:PIAACMC\_EELT\] and \[fig:PIAACMC\_TMT\] show that thanks to the phase-shifting focal plane mask, light from an on-axis source is diffracted between the small segments of the pupil, within the spider vane shadows, within the central obstruction and outside the overall pupil: in the output pupil plane, no light is present within the geometric pupil.
For these designs, the Lyot mask must mask the gaps between the segements while transmitting light within the segments, and it must therefore be carefully aligned with the pupil. A Lyot mask for which the masked zones are slightly oversized may be used to accomodate pupil alignment errors at the cost of system throughput.
Discussion
----------
[lcccccc]{}\
Subaru PIAACMC \#1 & 0.6 $\lambda/D$ & 2.42 & 0.603 $\lambda/D_{syst}$ & 84.24% & 99.91% & 0.67 $\lambda/D$\
Subaru PIAACMC \#2 & 0.9 $\lambda/D$ & 6.79 & 0.916 $\lambda/D_{syst}$ & 8.57% & 99.39% & 0.88 $\lambda/D$\
Subaru PIAACMC \#3 & 1.2 $\lambda/D$ & 26.83 & 1.33 $\lambda/D_{syst}$ & 2.06% & 97.04% & 1.11 $\lambda/D$\
\
GMT APCMLC \#1 & 0.7 $\lambda/D$ & 3.30 & 0.693 $\lambda/D_{syst}$ & 98.55% & 99.98% & 0.72 $\lambda/D$\
GMT APCMLC \#2 & 1.2 $\lambda/D$ & 26.83 & 1.12 $\lambda/D_{syst}$ & 20.71% & 99.47% & 0.89 $\lambda/D$\
GMT APCMLC \#3 & 1.5 $\lambda/D$ & 124.09 & 1.32 $\lambda/D_{syst}$ & 16.64% & 99.14% & 0.92 $\lambda/D$\
\
TMT APCMLC \#1 & 0.8 $\lambda/D$ & 4.69 & 0.797 $\lambda/D_{syst}$ & 85.51% & 99.80% & 0.78 $\lambda/D$\
TMT APCMLC \#2 & 1.2 $\lambda/D$ & 26.83 & 1.16 $\lambda/D_{syst}$ & 32.46% & 98.51% & 0.94 $\lambda/D$\
TMT APCMLC \#3 & 1.5 $\lambda/D$ & 124.09 & 1.394 $\lambda/D_{syst}$ & 27.73% & 98.71% & 0.99 $\lambda/D$\
\
E-ELT APCMLC \#1 & 0.8 $\lambda/D$ & 4.69 & 0.816 $\lambda/D_{syst}$ & 99.87% & 97.77% & 0.81 $\lambda/D$\
E-ELT APCMLC \#2 & 1.2 $\lambda/D$ & 26.83 & 1.15 $\lambda/D_{syst}$ & 45.58% & 99.50% & 0.93 $\lambda/D$\
E-ELT APCMLC \#3 & 1.5 $\lambda/D$ & 124.09 & 1.37 $\lambda/D_{syst}$ & 37.85% & 99.42% & 0.98 $\lambda/D$\
Table \[tab:PIAACMC\] summarizes the PIAACMC designs discussed in this section. For each design, the circular remapping function was first chosen, and is represented in the table by the parameter $b$, which is the diameter of the focal plane mask used to iteratively compute the generalized prolate function for a circular aperture. A small value of $b$ indicates a weak apodization. The PIAA strength listed in the table is the surface brightness ratio between the brightest and faintest parts of the remapped beam, and is a function of only $b$. This ratio is a good indicator for both the level of distortions of the off-axis PSFs in the intermediate focal plane, and for the difficulty in making the PIAA optics. Current PIAA optics for conventional PIAA coronagraphs have a strength around 100, and any value below 100 therefore corresponds to PIAA optics that can be manufactured to nm-level surface accuracy without technological advances. For PIAA strength values above 100, a hybrid scheme where some of the edge apodization is offloaded to a conventional apodizer should be adopted, at the cost of lower throughput (typically up to 10% throughput loss) and loss of angular resolution and IWA (by typically up to 5%).
Pupils with strong aspect ratio {#sec:aspectratio}
===============================
Challenges
----------
The APCMLC and PIAACMC coronagraphs described in the previous section achieve full starlight suppression by performing, for each point in the output pupil, perfect destructive interference between the light that passes through the circular focal plane mask and the light that passes around it. To offer $\lambda/D$-level inner working angle, these concepts therefore require that the telescope’s non coronagraphic point spread function consist of a central diffraction spot within which a disk containing approximately half of the total PSF flux can be drawn, surrounded by other fainter diffractive features (rings, spikes). The examples given in the previous sections (Subaru, GMT, TMT, E-ELT) fulfill this requirement, as these pupil shapes are sufficiently close to a disk.
While very sparse or elongated apertures are not compatible with the APCMLC and PIAACMC concepts as described so far, simple geometric transformations can extend the concepts to a wider range of pupil shapes. The Large Binocular Telescope (LBT) pupil is used in this section as an example of a sparse aperture with a strong aspect ratio: with its two centrally obscured 8.4-m diameter circular subapertures separated by 14.4m (center to center), the LBT pupil has a strong aspect ratio (8.4-m x 22.8-m). The corresponding non-coronagraphic PSF consists of three bright interference fringes within an envelope defined by the single aperture PSF. No circular mask can be drawn within the central bright fringe that contains half of the total PSF flux.
Using non-circular focal plane masks
------------------------------------
Stretching the LBT pupil along its narrow direction by a factor four would create a pupil sufficiently close to circular for the APCMLC and PIAACMC concepts as presented above. This stretch is equivalent to using an elliptical focal plane mask, which is four times longer in the direction running along the fringe in the PSF. Figure \[fig:LBT\_APCMLC\_design\] shows an APCMLC design for the LBT pupil using an elliptical focal plane mask. The design shown does produce total extinction of an on-axis point source, and its inner working angle is close to 1 $\lambda$/D along the long axis of the pupil (here, D is defined as the diameter of the circle enscribing the pupil, and is equal to 22.8 m), while it is $\approx$ 3.5 $\lambda$/D along the short axis (fundamentally limited by the telescope pupil diffraction along this axis, rather than by the coronagraph). A focal plane mask consisting of three separate zones covering part of the three central fringes may also be adopted to further improve system throughput, although this has not been numerically tested.
The same elliptical focal plane mask scheme can also applied for the PIAACMC concept on the LBT pupil. Interestingly, the pupil remapping which is part of the PIAACMC concept may be chosen to also bring the two aperture closer to approach the circular pupil case.
The elliptical focal plane mask approach may also be adopted to improve the APCMLC and PIAACMC performance for other non-circular pupil geometries: the focal plane mask shape should ideally be chosen to best match the non-coronagraphic PSF in order to maximize the conventional apodizer’s transmission. For example, the generalized prolate function for the Subaru Telescope PIAACMC design \#1 is slightly elongated due to the off-axis spider vanes. This produces a slight mismatch with the circular symmetric remapping function, which is absorbed by the conventional apodizer. Most of the conventional apodizer’s light loss (0.1% total) is due to this mismatch. For this example, using an slightly elliptical focal plane mask would only improve throughput by at most 0.1% since the pupil is very close to being circular. More importantly, the elliptical focal plane mask may allow high performance operation of the PIAACMC without an apodizer. Adopting a hexagonal shaped focal plane mask would offer similar benefits for hegagonal-shaped pupils such as the one shown in Figures \[fig:APCMLCprinciple\] and \[fig:PIAACMCprinc\].
Pupil remapping
---------------
With extremely sparse pupil geometries, the re-design of the focal plane mask geometry may not be sufficient to adapt the pupil shape to the APCMLC and PIAACMC requirements. In this case, geometrical transformation of the sparse entrance pupil into a more compact geometry can be achieved through pupil remapping. This scheme was explored to implement coronagraphy on sparse apertur [@2002AA...396..345R; @2002AA...391..379G], and commonly referred to as the hypertelescope concept.
Even if pupil remapping is not required, it may be useful to improve the APCMLC and PIAACMC system throughput. With sparse apertures, the apodizer becomes less transmissive: for example, the LBT pupil APCMLC design given in this section offers a 41% throughput, which is significantly less than the $\approx$60% throughput of comparable APCMLC designs for the Subaru, GMT, TMT and E-ELT pupils. Bringing the LBT subapertures closer together with periscope-like optics would allow for higher throughput in the coronagraph. In order to maintain a good image quality over a wide field of view, the original pupil geometry should be re-created prior to the final imaging focal plane: the compact pupil is only an intermediate step required for efficient removal of the central source’s light.
[lcccccc]{} Monochromatic & 1568.34 nm & 0 & 0 & 0 & 0\
2% band & 1568.25 nm & 5.29e-5 & 9.56e-11 & 8.31e-7 & 3.10e-7\
4% band & 1568.03 nm & 1.89e-4 & 2.15e-10 & 3.35e-6 & 1.16e-6\
10% band & 1566.54 nm & 1.08e-3 & 2.19e-7 & 2.13e-5 & 7.07e-6\
20% band & 1561.24 nm & 4.30e-3 & 7.68e-7 & 8.57e-5 & 2.80e-5\
40% band & 1543.88 nm & 1.52e-2 & 6.09e-7 & 3.24e-4 & 1.05e-4\
Chromaticity {#sec:chrom}
============
Sensitivity to chromatic effects
--------------------------------
All coronagraph systems discussed in this paper were designed for monochromatic light operation. While design of polychromatic APCMLC and PIAACMC systems is outside the scope of this paper (this will be discussed in a future publication), we describe qualitatively in this section how the monochromatic designs perform in broadband light.
Several effects result in a loss of performance in broadband light :
1. [The physical size of the focal plane mask is adjusted for a single wavelength. While the mask size is independent of wavelength, it should ideally scale linearly with wavelength.]{}
2. [The phase shift introduced by the mask may vary as a function of wavelength, while it should ideally be constant across the spectral band.]{}
3. [The transmission of the mask may vary as a function of wavelength, while it should ideally be constant across the band]{}
The amplitude of the last two effects is a function of how the focal plane mask is manufactured. In this section, we assume that no attempt to achromatize the mask phase shift has been made, and that it consists of a single material deposited on a substrate, with the material thickness adjusted for monochromatic light operation.
The sensitivity to chromatic effects is mostly driven by the focal plane transmission for both APCMLC and PIAACMC systems. Designs with large nearly opaque focal plane masks are more tolerant to chromatic effects, since the mask’s role becomes close to a simple light block, and he mask size relative to the on-axis source image increases. To illustrate broadband performance, we adopt in the next section a monochromatic PIAACMC design with partial ($0<|t|<1$) focal plane mask transmission.
Example: PIAACMC design for a centrally obscured pupil
------------------------------------------------------
We consider the PIAACMC design 2 for the Subaru Telescope pupil in this section. It is assumed that the focal plane mask size is optimized for monochromatic light at $\lambda = 1.65 \mu m$, and that the mask is a disk of material ($SiO_2$). The mask transmission is fixed to the ideal monochromatic value, and is not assumed to change with wavelength. Several scenarios are considered: monochromatic, 2%, 4%, 10%, 20% and 40% wide bands (all centered at 1.65 $\mu$m).
The mask thickness is a free parameter, and is adjusted for each case to yield the best broadband on-axis extinction, as measured by the total light in the final focal plane. Results are shown in table \[tab:chrom\]. The last 3 columns of the table show spatially averaged contrast values between the coronagraph’s inner working angle ($0.88 \lambda / D$) and $3.6 \lambda/D$.
This particular design delivers better than 1e-4 averaged raw contrast in a 20% wide specral band, and is therefore valuable for ground-based use behind adaptive optics. Pupil and focal plane images and contrast radial profiles are shown in Figure \[fig:chrom\] across a 40% wide band centered at $1.65 \mu m$, illustrating that raw contrast is best at the center of the band.
Conclusions {#sec:conclusion}
===========
The APCMLC and PIAACMC concepts, previously proposed for unobstructed circular apertures, are also applicable to telescopes with arbitrary pupil shapes. Their performance is largely unaffected by aperture shape, and full throughput low-IWA coronagraphy is therefore theoretically possible on any pupil shape with the PIAACMC. The demonstration that the coronagraph with the highest known theoretical performance can be applied on any pupil may remove the requirement that a future space-based exoplanet direct imaging mission should use an off-axis telescope. On ground-based telescopes, which adopt optical designs which are generally not driven by coronagraphy, high efficiency coronagraphy at and within 1 $\lambda/D$ is possible, potentially allowing direct imaging of habitable planets around nearby M and K type main sequence stars for which the planet-to-star contrast is favorable but the angular separation is extremely challenging and requires $\approx \lambda/D$ IWA even on a 30-m class telescope.
Manufacturing and implementation challenges have not been addressed in this paper. Manufacturing an achromatic focal plane mask for the APCMLC or PIAACMC is challenging, as its size should scale linearly with wavelength, and its complex amplitude transmission should be achromatic. Similar challenges have been previously addressed for other coronagraphs [@2011ApJ...729..144S], using carefully designed multilayer coatings of variable thickness and/or sub-$\lambda/D$ sized features optimized to produced the required chromatic dependence within the geometric pupil [@2003AA...403..369S; @2012AA...538A..55N]. The PIAA optics required for the PIAACMC are however not as challenging to manufacture as PIAA optics previously made for hard edged opaque focal plane masks, because the PIAACMC’s entrance apodization is milder. As any high performance low IWA coronagraph, the PIAACMC performance is highly sensitive to residual wavefront errors, which must be actively sensed and controlled. The PIAACMC’s high throughput is an asset for achieving the required wavefront quality, as wavefront sensing can be performed rapidly, using all incoming light.
Small IWA high contrast coronagraphy requires exquisite control of tip-tilt and low order wavefront errors. The central star angular size may also impose limits on the achievable performance [@2009ApJ...702..672C]. These issues have not been addressed or quantified in this paper, but may drive the optimal coronagraph design for a particular scientific application. We note that both sub-$\lambda/D$ IWA coronagraphs designs described in this paper can also be designed for IWA equal to or larger than $\lambda$/D if required, offering lower performance but improved resilience against pointing errors and stellar angular size. While the APCMLC design with larger IWA has a lower throughput (due to the stronger apodization), for the PIAACMC, the large-IWA designs maintain full throughput and total on-axis extinction, offering a wide range of practical high performance coronagraphic options. $\:$ $\:$ $\:$ $\:$ $\:$ $\:$ -
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'Weikai Yang[^1], Zhen Li$^{*}$, Mengchen Liu[^2], Yafeng Lu[^3], Kelei Cao$^{*}$, Ross Maciejewski[^4], Shixia Liu$^{*}$'
bibliography:
- 'DriftVis-VIS2020.bib'
title: Diagnosing Concept Drift
---
[^1]: e-mail:{{ywk19, liz11, ckl17}@mails., shixia@}tsinghua.edu.cn. Weikai Yang and Zhen Li are joint first authors. Shixia Liu is the corresponding author.
[^2]: e-mail:[email protected]
[^3]: e-mail:[email protected]
[^4]: e-mail:[email protected]
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'A. Suárez Mascareño'
- 'J. I. González Hernández'
- 'R. Rebolo'
- 'N. Astudillo-Defru'
- 'X. Bonfils'
- 'F. Bouchy'
- 'X.Delfosse'
- 'T. Forveille'
- 'C. Lovis'
- 'M. Mayor'
- 'F. Murgas'
- 'F. Pepe'
- 'N. C. Santos'
- 'S. Udry'
- 'A. Wünsche'
- 'S. Velasco'
bibliography:
- 'RHK\_ref.bib'
date: 'Written July-2016'
title: 'A super-Earth orbiting the nearby M dwarf GJ 536 [^1]'
---
[We report the discovery of a super-Earth orbiting the star GJ 536 based on the analysis of the radial-velocity time series from the HARPS and HARPS-N spectrographs. GJ 536 b is a planet with a minimum mass M sin $i$ of 5.36 $\pm$ 0.69 M$_{\oplus}$; it has an orbital period of 8.7076 $\pm$ 0.0025 days at a distance of 0.066610(13) AU, and an orbit that is consistent with circular. The host star is the moderately quiet M1 V star GJ 536, located at 10 pc from the Sun. We find the presence of a second signal at 43 days that we relate to stellar rotation after analysing the time series of Ca II H&K and H$_{\alpha}$ spectroscopic indicators and photometric data from the ASAS archive. We find no evidence linking the short period signal to any activity proxy. We also tentatively derived a stellar magnetic cycle of less than 3 years. ]{}
Introduction
============
Several surveys have attempted to take advantage of the low masses of M dwarfs and therefore of the stronger radial-velocity signals induced for the same planetary mass – and their closer habitable zones to detect rocky habitable planets [@Bonfils2013; @Howard2014; @Irwin2015; @Berta-Thompson2015; @AngladaEscude2016]. While surveying M dwarfs has advantages, it also has its own drawbacks. Stellar activity has been one of the main difficulties when trying to detect planets trough Doppler spectroscopy. Not only does it introduce noise, but also coherent signals that can mimic those of planetary origin [@Queloz2001; @Bonfils2007; @Robertson2013]. M dwarfs tend to induce signals with amplitudes comparable to those of rocky planets [@Howard2014; @Robertson2013]. While these kinds of stars allow for the detection of smaller planets, they also demand a more detailed analysis of the radial-velocity signals induced by activity. In addition, these low-mass stars offer valuable complementary information on the formation mechanisms of planetary systems. For instance, giants planets are known to be rare around M dwarfs, while super-Earths appear to be more frequent [@Bonfils2013; @Dressing2013; @Dressing2015].
In spite of the numerous exoplanets detected by *Kepler* [@Howard2012] and by radial-velocity surveys [@Howard2009; @Mayor2011] the number of known small rocky planets is still comparably low. There are around 1500 confirmed exoplanets and more than 3000 Kepler candidates, but only about a hundred of the confirmed planets have been reported on M dwarfs and only a fraction of them are rocky planets. The first discovery of a planet around an M dwarf dates back to 1998 [@Delfosse1998; @Marcy1998]. Since then several planetary systems have been reported containing Neptune-mass planets and super-Earths [@Udry2007; @Delfosse2013; @Howard2014; @AstudilloDefru2015] and even some Earth-mass planets [@Mayor2009; @BertaThompson2015; @Wright2015; @Affer2016; @AngladaEscude2016]. However, the frequency of very low-mass planets around M dwarfs is not well established. In particular, as noted by @Bonfils2013, the frequency of rocky planets at periods shorter than 10 days is $0.36^{+0.24}_{-0.10}$; it is $0.41^{+0.54}_{-0.13}$ for the habitable zone of the stars. On the other hand @Gaidos2013 estimated that the frequency of habitable rocky planets is $0.46^{+0.20}_{-0.15}$ on a wider spectral sample of Kepler dwarfs, and @Kopparapu2013 gave a frequency of $0.48^{+0.12}_{-0.24}$ for habitable planets around M dwarfs. The three measurements are compatible, but uncertainties are still big making it important to continue the search for planets around this star type in order to refine the statistics.
We present the discovery of a super-Earth orbiting the nearby star GJ 536, which is a high proper motion early M dwarf at a distance of 10 pc from the Sun [@vanLeeuwen2007; @Maldonado2015]. Because of its high proper motion and its closeness, this star shows a secular acceleration of 0.24 m s$^{-1}$ yr$^{-1}$ [@Montet2014]. Table \[parameters\] shows the stellar parameters. Its moderately low activity combined with its long rotation period of more than 40 days [@Masca2015] makes it a very interesting candidate to search for rocky planets.
Parameter GJ 536 Ref.
------------------------------------- ----------------------- ------ -- -- -- -- -- -- -- -- --
RA (J2000) 14:01:03.19 1
DEC (J2000) -02:39:17.52 1
$\delta$ RA($mas$ yr$^{-1}$) -823.47 1
$\delta$ DEC ($mas$ yr$^{-1}$) 598.19 1
Distance \[pc\] 10.03 1
$m_{B}$ 11.177 2
$m_{V}$ 9.707 2
$m_{V}$ ASAS 9.708 0
Spectral Type M1 3
T$_{\rm eff}$ \[K\] 3685 $\pm$ 68 3
$[Fe/H]$ -0.08 $\pm$ 0.09 3
$M_\star$ $[M_{\odot}]$ 0.52 $\pm$ 0.05 3
$R_\star$ $[R_{\odot}]$ 0.50 $\pm$ 0.05 3
log $g$ (cgs) 4.75 $\pm$ 0.04 3
log($L_{\star}/L_{\odot}$) -1.377 3
$\log_{10}(R'_\textrm{HK})$ -5.12 $\pm$ 0.05 0
$P_{\rm rot}$ 45.39 $\pm$ 1.33 0
$v \sin i$ (km s$^{-1}$) $\textless$ 1.2$^{*}$ 0
Secular acc. (m s$^{-1}$ yr$^{-1}$) 0.24 4
\[parameters\]
: Stellar parameters of GJ 536 \[tab:parameters\]
**References:** 0 - This work, 1 - @vanLeeuwen2007, 2 - @Koen2010, 3 -@Maldonado2015, 4 - Calculated following @Montet2014.\
$^{*}$ Estimated using the Radius estimated by @Maldonado2015 and our period determination.\
The star GJ 536 is part of the @Bonfils2013 sample and has been extensively monitored since mid-2004. We have used 146 HARPS spectra taken over 11.7 yr along with 12 HARPS-N spectra taken during April and May 2016. HARPS [@Mayor2003] and HARPS-N [@Cosentino2012] are two fibre-fed high-resolution echelle spectrographs installed at the 3.6 m ESO telescope in La Silla Observatory (Chile) and at the Telescopio Nazionale Galileo in the Roque de los Muchachos Observatory (Spain), respectively. Both instruments have a resolving power greater than $R\sim 115\,000$ over a spectral range from $\sim$380 to $\sim$690 nm and have been designed to attain very high long-term radial-velocity accuracy. Both are contained in vacuum vessels to avoid spectral drifts due to temperature and air pressure variations, thus ensuring their stability. HARPS and HARPS-N are equipped with their own pipeline providing extracted and wavelength-calibrated spectra, as well as RV measurements and other data products such as cross-correlation functions (CCFs) and their bisector profiles.
Most of the observations were carried out using the Fabry Pérot interferometer (FP) as simultaneous calibration. The FP offers the possibility of monitoring the instrumental drift with a precision of 10 cms$^{-1}$ without the risk of contamination of the stellar spectra by the ThAr saturated lines [@Wildi2010]. While this is not usually a problem in G and K stars, the small amount of light collected in the blue part of the spectra of M dwarfs might compromise the quality of the measurement of the Ca II H&K flux. The FP allows a precision of $\sim$ 1 ms$^{-1}$ in the determination of the radial velocities of the spectra with highest signal-to-noise ratios while assuring the quality of the spectroscopic indicators even in those spectra with low signal-to-noise ratios. Measurements taken before the availability of the FP were taken without simultaneous reference.
We also use the photometric data on GJ 536 provided by the All Sky Automated Survey (ASAS) public database. ASAS [@Pojmanski1997] is an all-sky survey in the $V$ and $I$ bands running since 1998 at Las Campanas Observatory, Chile. The best photometric results are achieved for stars with V $\sim$8-14, but this range can be extended by implementing some quality control on the data. ASAS has produced light curves for around $10^{7}$ stars at $\delta < 28^{\circ}$. The ASAS catalogue supplies ready-to-use light curves with flags indicating the quality of the data. For this analysis we relied only on good quality data (grade A and B in the internal flags). Even after this quality control, there are still some high dispersion measurements which cannot be explained by a regular stellar behaviour. We reject those measurements by de-trending the series and eliminating points deviating more than three times the standard deviation from the median seasonal value. We are left with 359 photometric observations taken over 8.6 yr with a typical uncertainty of 9.6 mmag per exposure.
Determining stellar activity indicators and radial velocities
=============================================================
Activity indicators
-------------------
For the activity analysis we use the extracted order-by-order wavelength-calibrated spectra produced by the HARPS and HARPS-N pipelines. For a given star, the change in atmospheric transparency from day to day causes variations in the flux distribution of the recorded spectra that are particularly relevant in the blue where we intend to measure Ca II lines. In order to minimize the effects related to these atmospheric changes we create a spectral template for each star by de-blazing and co-adding every available spectrum; we use the co-added spectrum to correct the order-by-order flux of each individual spectrum. We also correct each spectrum for the Earth’s barycentric radial velocity and the radial velocity of the star using the measurements given by the standard pipeline, and re-binned the spectra into a wavelength-constant step. Using this HARPS dataset, we expect to have high-quality spectroscopic indicators to monitor tiny stellar activity variations with high accuracy.
S$_{MW}$ index {#s_mw-index .unnumbered}
--------------
We calculate the Mount Wilson $S$ index and the $\log_{10}(R'_{HK})$ by using the original @Noyes1984 procedure, following @Lovis2011 and @Masca2015. We define two triangular passbands with full width half maximum (FWHM) of 1.09 [Å]{} centred at 3968.470 [Å]{} and 3933.664 [Å]{} for the Ca II H&K line cores, and for the continuum we use two 20 [Å]{} wide bands centred at 3901.070 [Å]{} (V) and 4001.070 [Å]{}(R), as shown in Fig. \[Sindex\].
![Ca II H&K filter of the spectrum of the star GJ536 with the same shape as the Mount Wilson Ca II H&K passband.[]{data-label="Sindex"}](S_index_filter_S){width="9.0cm"}
Then the S-index is defined as
$$S=\alpha {{\tilde{N}_{H}+\tilde{N}_{K}}\over{\tilde{N}_{R}+\tilde{N}_{V}}} + \beta,$$
where $\tilde{N}_{H},\tilde{N}_{K},\tilde{N}_{R}$, and $\tilde{N}_{V}$ are the mean fluxes in each passband, while $\alpha$ and $\beta$ are calibration constants fixed as $\alpha = 1.111$ and $\beta = 0.0153$ . The S index serves as a measurement of the Ca II H&K core flux normalized to the neighbour continuum. As a normalized index to compare it to other stars, we compute the $\log_{10}(R'_\textrm{HK})$ following @Masca2015.
H$_{\alpha}$ index {#h_alpha-index .unnumbered}
------------------
We also use the H$\alpha$ index, with a simpler passband following [@GomesdaSilva2011]. It consists of a rectangular bandpass with a width of 1.6 [Å]{} and centred at 6562.808 [Å]{} (core), and two continuum bands of 10.75 [Å]{} and 8.75 [Å]{} in width centred at 6550.87 [Å]{} (L) and 6580.31 [Å]{} (R), respectively, as seen in Figure \[halpha\].
![Spectrum of the M-type star GJ536 showing the H$\alpha$ filter passband and continuum bands.[]{data-label="halpha"}](Ha_index_filter_S){width="9.0cm"}
Thus, the H$\alpha$ index is defined as
$$H\alpha_{\rm Index}={{H\alpha_{\rm core} }\over{H\alpha_{L} +H\alpha_{R}}}.$$
Radial velocities
-----------------
The radial-velocity measurements in the HARPS standard pipeline is determined by a Gaussian fit of the CCF of the spectrum with a binary stellar template [@Baranne1996; @PepeMayor2000]. In the case of M dwarfs, owing to the huge number of line blends, the CCF is not Gaussian and results in a less precise Gaussian fit which might cause distortions in the radial-velocity measurements and FWHM. To deal with this issue we tried two different approaches.
The first approach consisted in using a slightly more complex model for the CCF fitting, a Gaussian function plus a second-order polynomial (Fig. \[ccf\_plot\]) using only the central region of the CCF function. We use a 15 Km s$^{-1}$ window centred at the minimum of the CCF. This configuration provides the best stability of the measurements. Along with the measurements of the radial velocity we obtain the FWHM of the cross-correlation function, which we also use to track variations in the activity level of the star. The second approach to the problem was to recompute the radial velocities using a template matching algorithm with a high signal-to-noise stellar spectral template [@AstudilloDefru2015]. Every spectrum is corrected from both barycentric and stellar radial velocity to align it to the frame of the solar system barycentre. The radial velocities are computed by minimizing the $\chi^2$ of the residuals between the observed spectra and shifted versions of the stellar template, with all the elements contaminated by telluric lines masked. All radial-velocity measurements are corrected from the secular acceleration of the star.
![Cross-correlation function for GJ 536. Upper panels show the CCF with the Gaussian fit (left) and our Gaussian plus polynomial fit (right). Lower panels show the residuals after the fit for the Gaussian fit (left) and for our fit. Blue lines show the fit (upper panels) and the zero line (lower panels).[]{data-label="ccf_plot"}](ccf_plot){width="9.0cm"}
For the bisector span measurement we rely on the pipeline results as it does not depend on the fit, but on the CCF itself. The bisector has been a standard activity diagnostic tool for solar-type stars for more than 10 years. Unfortunately, its behaviour in slow rotating stars is not as informative as it is for fast rotators [@Saar1997; @Bonfils2007]. We report the measurements of the bisector span (BIS) for each radial-velocity measurement, but we do not find any meaningful information in its analysis.
Quality control of the data
---------------------------
As the sampling rate of our data is not well suited for modelling fast events, such as flares, and their effect on the radial velocity is not well understood, we identify and reject points likely affected by flares by searching for abnormal behaviour of the activity indicators [@Reiners2009]. The process rejected six spectra that correspond to flare events of the star with obvious activity enhancement and line distortion. That leaves us with 140 HARPS spectroscopic observations taken over 10.7 years, with most of the measurements taking place after 2013, with a typical exposure of 900 s and an average signal-to-noise ratio of 56 at $5500$ [Å]{}. We do not apply the quality control procedure to the HARPS-N data as the number of spectra is not high enough.
Stellar activity analysis
=========================
In order to properly understand the behaviour of the star, our first step is to analyse the different modulations present in the photometric and spectroscopic time-series.
We search for periodic variability compatible with both stellar rotation and long-term magnetic cycles. We compute the power spectrum using a generalised Lomb-Scargle periodogram [@Zechmeister2009] and if there is any significant periodicity we fit the detected period using a sinusoidal model, or a double sinusoidal model to account for the asymmetry of some signals [@BerdyuginaJarvinen2005] with the MPFIT routine [@Markwardt2009].
The significance of the periodogram peak is evaluated using the @Cumming2004 modification of the @HorneBaliunas1986 formula to obtain the spectral density thresholds for the desired false alarm probability (FAP) levels and the bootstrap randomization [@Endl2001] of the data.
Figure \[GJ\_ac\_data\] shows the time series for the photometry (top panel) and the three activity proxies (bottom panels) used for this analysis. The periodograms of both the photometric and FWHM time series show significant signals at $\sim$40 days, compatible with the typical rotation periods of low-activity M1 stars [@Masca2016a; @Newton2016]. On the other hand, the periodograms of the S$_{MW}$ and H$_{\alpha}$ indexes show long-term and short-term significant signals. The short period signal is again at $\sim$40 days, while the long-term signal is close to $\sim$1000 days.
![From top to bottom: Time series of the m$_{V}$, S$_{MW}$ index, H$_{\alpha}$ index, and FWHM time series. Grey dots show HARPS-S data; black asterisks show HARPS-N data.[]{data-label="GJ_ac_data"}](GJ536_data){width="9.0cm"}
![From top to bottom: Periodograms of the m$_{V}$, S$_{MW}$ index, H$_{\alpha}$ index, and FWHM time series. Horizontal lines show the different levels of false alarm probability. Red dotted line for a 10% false alarm probability, green dashed line for 1%, and blue thick line for 0.1%. Several peaks arise with significances better than 0.1%.[]{data-label="LT_period"}](GJ536_LTM_data){width="9.0cm"}
Long-term magnetic cycle
------------------------
Analysing the S$_{MW}$ and H$_{\alpha}$ indexes time series we find the presence of a long-term magnetic cycle of $\sim$3 years. Figure \[LT\_period\] shows the periodograms of the time series of both indexes. We see a well-defined peak in the S$_{MW}$ index periodogram at $\sim$806 d and several peaks going from $\sim$600 d to 1100 d in the H$_{\alpha}$ index periodogram implying that the shape of the cycle is still not well defined within our observations. Table \[periodicities\] shows the periods of the best fits for both time series using least-squares minimization with the period corresponding to the highest peak of the periodogram as the initial guess. Figure \[LT\_fit\] shows the phase folded curves using these periods. The two estimates differ significantly. This might be because of a sub-optimal sampling to detect signals of long periods. The detected periodicities might not be the true periodicities, but apparent periodicities close to the real one caused by the sampling. This also makes us think that the uncertainties in the cycle length are underestimated. The length of the signal is shorter than the typical magnetic cycles measured in solar-type stars, but is within the range of known magnetic cycles in M-type stars [@Masca2016a]. In the S$_{MW}$ and the H$_{\alpha}$ indexes it seems that the cycle shape shows a quick rise followed by a slow decline, as is the case in the Sun and many other main sequence stars [@Waldmeier1961; @Baliunas1995]. Unfortunately this cycle is not well covered in phase, making it difficult to properly characterise it. More observations are needed in order to better constrain its period.
{width="9.0cm"}
{width="9.0cm"}
Rotation
--------
The other activity signal expected in our data is the rotational modulation of the star. It shows up at $\sim$43 d with a false alarm probability close to or smaller than 1% in the four time series (Fig. \[LT\_period\]) that grow in significance after removing the long-term effects.
In the photometric light curve we measure a modulation of 43.33 $\pm$ 0.06 d with an amplitude of 5.21 $\pm$ 0.68 mmag. For the S$_{MW}$ index we find a signal of 43.84 $\pm$ 0.01 d with an amplitude of 0.0628 $\pm$ 0.0010 when doing a simultaneous fit with the 824-day signal from Table \[LT\_fit\]. In the case of the H$_{\alpha}$ index we find a signal 42.58 $\pm$ 0.08 d with an amplitude of 0.0042 $\pm$ 0.0010, also when doing a simultaneous fit with the $\sim$1075 days signal. The time series of the FWHM show a linear increase with time of $\sim$ 2 ms$^{-1}$yr$^{-1}$, which might be related to a slow focus drift of HARPS. After subtracting the linear trend we again find a periodicity of 44.47 $\pm$ 0.03 d period with an amplitude of 4.56 $\pm$ 0.31 ms$^{-1}$. Figure \[rot\_phase\] shows the phase folded fits of all the signals. The S$_{MW}$ index and FWHM signals seem to be in phase, while the photometric signal is shifted by a quarter phase. The uncertainty in the H$_{\alpha}$ long-term fit makes it difficult to give it a unique phase to the rotation signal. Table \[periodicities\] shows the parameters for the four signals.
Our measurement of 45.39 d strengthens the previous estimation of @Masca2015. Having such a clear detection of the rotational modulation in that many indicators over so many years supports the idea that activity regions in at least some M-type stars are stable over long time spans [@Robertson2015].
{width="9.0cm"}
{width="9.0cm"}
{width="9.0cm"}
{width="9.0cm"}
[ l l l l l l l l l l l l ]{} Series & Period (d) & Amplitude & FAP (%)\
S$_{MW ~ \rm Cyc}$ & 824.9 $\pm$ 1.7 & 0.0684 $\pm$ 0.0011 & $\textless$ 0.1\
H$_{\alpha ~ \rm Cyc}$ & 1075.8 $\pm$ 36.1 & 0.0046 $\pm$ 0.0011 & $\textless$ 0.1\
\
m$_{V ~ \rm Rot}$ & 43.33 $\pm$ 0.06 & 5.21 $\pm$ 0.68 mmag & $\textless$ 1\
S$_{MW ~ \rm Rot}$ & 43.84 $\pm$ 0.01 & 0.0628 $\pm$ 0.0010 & $\textless$ 0.1\
H$_{\alpha ~ \rm Rot}$ & 42.58 $\pm$ 0.08 & 0.0042 $\pm$ 0.0010 & $\textless$ 0.1\
FWHM$_{\rm Rot} $ & 44.47 $\pm$ 0.03 & 4.56 $\pm$ 0.31 ms$^{-1}$ & $\textless$ 1\
\
**$\textless$ Rot. $\textgreater$** & **43.87 $\pm$ 0.80**\
\[periodicities\]
The mean value is the weighted mean of all the individual measurements. The error of the mean is the standard deviation of the individual measurements divided by the square root of the number of measurements.
Radial-velocity analysis
========================
Our 152 radial-velocity measurements have a median error of 1.33 ms$^{-1}$ which includes both photon noise, calibration, and telescope related errors. We measure a total systematic radial velocity of -25.622 Kms$^{-1}$ with a dispersion of 3.28 ms$^{-1}$. Figure \[rv\_ts0\] shows the measured radial velocities. An F-test [@Zechmeister2009b] returns a negligible probability (smaller than the 0.1%) that the internal errors explain the measured dispersion.
![Radial-velocity time series. Grey dots show HARPS-S data; black asterisks show HARPS-N data.[]{data-label="rv_ts0"}](GJ536_data0){width="9.0cm"}
To search for periodic radial-velocity signals in our time-series we follow a similar procedure to the one explained in section 3.1. We search for periodic signals using a generalised Lomb-Scargle periodogram, and if there is any significant periodicity we fit the detected signal using the RVLIN package [@WrightHoward2012]. We sequentially find the dominant components in the time series and remove them until no significant signal remains.
Following this procedure we identify one signal with a false alarm probability much higher than 0.1%, using both the bootstrap and the @Cumming2004 estimates, corresponding to a period of 8.7 d with a semi-amplitude of 2.47 ms$^{-1}$ consistent with circular (Fig. \[rv\_ts\] shows the periodogram). Removing this signal leaves a 43.9 d signal with a semi-amplitude of 2.86 ms$^{-1}$ and an eccentricity of 0.57, with a false alarm probability better than 0.1%. No further significant signals are found after removing these two (Fig. \[rv\_ts\]). Fig. \[rv\_phase\] shows the phase folded fits of both the 8.7 d and the 43.9 d signals.
We tested the available dataset for the three ways of calculating the radial velocity, obtaining virtually the same results in every case. Results are shown for the Gaussian + polynomial fit of the cross-correlation function.
![Periodograms of the radial velocity. The upper panel shows the raw periodogram, the middle panel the periodogram of the residuals after subtracting the 8.7 d signal, and the lower panel the periodogram of the residuals after subtracting the 43 d signal present in the middle panel. Red regions show the periods of the measured rotation and magnetic cycle. Red dotted line for a 10% false alarm probability; green dashed line for 1%, and blue thick line for 0.1%.[]{data-label="rv_ts"}](GJ536_RV_data){width="9.0cm"}
![**Top panel:** Phase folded curve of the radial velocity using the 8.7 d period. Grey dots are the raw radial-velocity measurements after subtracting the mean value and the 43.9 d signal. **Bottom panel:** Phase folded curve of the radial velocity using the 43.9 d period using a double-harmonic sine curve. Grey dots and black asterisks are the raw radial-velocity measurements after subtracting the mean value and the 8.7 d signal. Red dots are the same points binned in phase with a bin size of 0.1. The error bar of a given bin is estimated using the weighted standard deviation of binned measurements divided by the square root of the number of measurements included in this bin. This estimation of the bin error bars assumes white noise, which is justified by the binning in phase and which regroups points that are uncorrelated in time.[]{data-label="rv_phase"}](plan_RV_fit_final "fig:"){width="9.0cm"} ![**Top panel:** Phase folded curve of the radial velocity using the 8.7 d period. Grey dots are the raw radial-velocity measurements after subtracting the mean value and the 43.9 d signal. **Bottom panel:** Phase folded curve of the radial velocity using the 43.9 d period using a double-harmonic sine curve. Grey dots and black asterisks are the raw radial-velocity measurements after subtracting the mean value and the 8.7 d signal. Red dots are the same points binned in phase with a bin size of 0.1. The error bar of a given bin is estimated using the weighted standard deviation of binned measurements divided by the square root of the number of measurements included in this bin. This estimation of the bin error bars assumes white noise, which is justified by the binning in phase and which regroups points that are uncorrelated in time.[]{data-label="rv_phase"}](rot_RV_fit_final "fig:"){width="9.0cm"}
Origin of the periodic radial-velocity signals
----------------------------------------------
Stellar activity can induce radial-velocity signals similar to those of Keplerian origin. The inhomogeneities in the surface of the star cause radial-velocity shifts due to the distortion of the spectral line shapes which can, in some cases, create a radial-velocity signal with a periodicity close to the stellar rotation and its first harmonic.
For this star we have a rotation period of 45.39 $\pm$ 1.33 d, and two radial-velocity signals of 8.7 d and 43.9 d. The second signal matches almost perfectly the rotation period of the star. On the other hand we do not see in the time series of activity indicators any signal close to the 8.7 d. This is the first evidence of the stellar origin of the 43.9 d signal, and the planetary origin of the 8.7 d signal.
As a second test we measured the Spearman correlation coefficient between the S$_{MW}$, the H$_{\alpha}$ index, the FWHM, and the radial velocities. We find a significant correlation between all the indexes and the raw radial velocity, which almost disappears when we isolate the 8.7 d signal, and slightly increases when isolating the 43.9 d signal (see Table. \[rv\_corre\]). This constitutes a second piece of evidence of the stellar origin of the 43.9 d signal, and of the planetary origin of the 8.7 d signal. Following this idea, we subtract the linear correlation between the radial velocity and each of the three activity diagnostic indexes. By doing this we see that the strength of the 8.7 d signal remains constant, or even increases, while the significance of the 43.9 d is reduced in all cases (see Fig. \[rv\_nocorr\]), even getting buried in the noise after correcting for the correlation with the H$_{\alpha}$ index.
Keplerian signals are deterministic and consistent in time. When measuring one signal, we expect to find that the significance of the detection increases steadily with the number of observations, and that the measured period is stable over time. However, in the case of an activity related signal this is not necessarily the case. As the stellar surface is not static, and the configuration of active regions may change in time, changes in the phase of the modulation and in the detected period are expected. Even the disappearance of the signal at certain seasons is possible. Fig. \[rv\_signals\] shows the evolution of the false alarm probability of the detection of both isolated signals, as well as the measurement of the most prominent period when isolating them. The 8.7 d signal increases steadily with time, and once it becomes the most significant signal it never moves again. On the other hand, the behaviour of the 43.9 d is more erratic, losing significance during the last observations.
Of the two significant radial-velocity signals detected in our data it seems clear that the one at 8.7 d has a planetary origin, while the one at 43.9 has a stellar activity origin.
The shape of the activity induced radial-velocity signal present in our data is evidently not sinusoidal. A double harmonic sinusoidal, as in the case of the activity signals, is the best fit model and the only one that does not create ghost signals after subtracting it. The rotation induced signal is not in phase with the rotation signals in the activity indicators. It appears to be shifted by $\sim$45$^\circ$ from the signal in the S$_{MW}$ index and FWHM time series as seen in @Bonfils2007 and @Santos2014. The uncertainty in the phase H$_\alpha$ time series makes it difficult to measure a reliable phase difference.
Parameter Raw data 8.7 d signal 43.9 d signal
------------------------- ---------------------------------- ------------------------------ ---------------------------------- -- -- -- -- -- -- -- --
S$_{MW}$ vs V$_{R}$ 0.292 ($\textgreater$ 3$\sigma$) 0.069 ($\textless$1$\sigma$) 0.345 ($\textgreater$ 3$\sigma$)
H$_{\alpha}$ vs V$_{R}$ 0.338 ($\textgreater$3$\sigma$) 0.113 (1$\sigma$) 0.321 ($\textgreater$3$\sigma$)
FWHM vs V$_{R}$ 0.356 ($\textgreater$ 3$\sigma$) 0.164 (1$\sigma$) 0.340 ($\textgreater$ 3$\sigma$)
\[rv\_corre\]
: Activity - Radial-velocity correlations \[tab:rv\_corre\]
Long-term variations of activity indicators have been subtracted. The parenthesis value indicates the significance of the correlation given by the bootstrapping process.
![Periodograms for the radial velocity after removing the correlation with the different activity diagnostic tools. From left to right there is the periodogram for the original data, the periodogram after detrending against the S$_{MW}$ index, against the H$_{\alpha}$ index, and against the FWHM.[]{data-label="rv_nocorr"}](GJ536_RV_nocorr){width="9cm"}
![Evolution of the false alarm probability of the detections (upper panel) for the isolated signals, and stability of the detections (lower panel). Blue thick line shows the behaviour for the 8.7 d signals and red dashed line for the 43.9 d signal.[]{data-label="rv_signals"}](GJ536_RV_signals){width="9cm"}
Finally, an analysis of the spectral window rules out that the peaks in the periodogram are artefacts of the time sampling alone. No features appear at 8.7 or 43.9 days even after masking the oversaturated regions of the power spectrum. Following @Rajpaul2016 we tried to re-create the 8.7-day by injecting the P$_{\rm Rot}$ signal along with a second signal at P$_{\rm Rot}$/2 at 1000 randomized phase shifts with a white noise model. We were never able to generate a signal at 8.7 days, or any significant signal at periods close to 8.7 days. It seems very unlikely that any of the signals are artefacts of the sampling.
GJ 536 b
--------
The analysis of the radial-velocity time series and of the activity indicators leads us to conclude that the best explanation of the observed data is the existence of a planet orbiting the star GJ 536 with a period of 8.7 d, with a semi-amplitude of $\sim$ 2.5 ms$^{-1}$. The best solution comes from a super-Earth with a minimum mass of 5.3 M$_{\oplus}$ orbiting at 0.067 AU of its star.
MCMC analysis of the radial-velocity time series {#mcmc-analysis-of-the-radial-velocity-time-series .unnumbered}
------------------------------------------------
In order to quantify the uncertainties of the orbital parameters of the planet, we perform a Bayesian analysis using the code [ExoFit]{} [@Balan2009]. This code follows the Bayesian method described in @Gregory2005 [@Ford2005] and @FordGregory2007. A single planet can be modelled using the following formula: $$v_i = \gamma - K [\sin( \theta(t_i+\chi P) + \omega ) + e \sin \omega ]
,$$where $\gamma$ is system radial velocity; $K$ is the velocity semi-amplitude equal to $2\pi P^{-1} (1-e^2)^{-1/2} a \sin i$; $P$ is the orbital period; $a$ is the semi-major axis of the orbit; $e$ is the orbital eccentricity; $i$ is the inclination of the orbit; $\omega$ is the longitude of periastron; $\chi$ is the fraction of an orbit, prior to the start of data taking, at which periastron occurs (thus, $\chi P$ equals the number of days prior to $t_i=0$ that the star was at periastron, for an orbital period of $P$ days); and $\theta(t_i + \chi P)$ is the angle of the star in its orbit relative to periastron at time $t_i$, also called the true anomaly.
To fit the previous equation to the data we need to specify the six model parameters, $P$, $K$, $\gamma$, $e$, $\omega$, and $\chi$. Observed radial-velocity data, $d_i$, can be modelled by the equation $d_i = v_i + \epsilon_i + \delta$ [@Gregory2005], where $v_i$ is the modelled radial velocity of the star and $\epsilon_i$ is the uncertainty component arising from accountable but unequal measurement error, which are assumed to be normally distributed. The term $\delta$ explains any unknown measurement error. Any noise component that cannot be modelled is described by the term $\delta$. The probability distribution of $\delta$ is chosen to be a Gaussian distribution with finite variance $s^2$. Therefore, the combination of uncertainties $\epsilon_i + \delta$ has a Gaussian distribution with a variance equal to $\sigma_i^2 + s^2$ [see @Balan2009 for more details].
The parameter estimation in the Bayesian analysis needs a choice of priors. We choose the priors following the studies by @FordGregory2007 [@Balan2009].The mathematical form of the prior is given in Table 1 and/or 4 of @Balan2009. In Table \[mcmc\_par\], we provide the parameter boundaries explored in the MCMC Bayesian analysis. [ExoFit]{} performs 100 chains of 10000 iterations each resulting in a final chain of 19600 sets of global-fit parameters.
We want to simultaneously model the stellar rotation and planetary signals. For that we use the [ExoFit]{} to model two RV signals and for the rotation signal we also leave the eccentricity as a free parameter. The posterior distribution of the eccentricity parameter for the rotation signal (not shown in Fig. \[mcmc\]) gives a value of $0.47 \pm 0.26$. In Fig. \[mcmc\] we depict the posterior distribution of the model parameters; the six fitted parameters; the semi-amplitude velocity, $K_{\rm rot}$, and the period, $P_{\rm rot}$, of the rotation signal; the derived mass of the planet, $m_p \sin i$; and the RV noise given by the $s$ parameter. Most of the parameters show symmetric density profiles except for the eccentricity, $e$; the longitude of periastron, $\omega$; and the fraction $\chi$ of the orbit at which the periastron occurs. We note that the density profile of the rotation period displays a tail towards slightly lower values although the rotation period is well defined.
In Table \[mcmc\_par\] we show the final parameters and uncertainties obtained with the MCMC Bayesian analysis with the code [ExoFit]{}.
{width="18cm"}
Parameter Value Upper error Lower error Prior
---------------------------------- ---------- ------------- ------------- -----------------
$P_{\rm planet}$ \[d\] 8.7076 $+$0.0022 $-$0.0025 8.3 - 9.0
$\gamma$ \[ms$^{-1}-25625$\] 1.17 $+$0.20 $-$0.20 $-$5.0 - $+$5.0
$e$ 0.08 $+$0.09 $-$0.06 0.0 - 0.99
$\omega$ \[deg\] 288.7 $+$42.5 $-$50.6 0.0 - 360.0
$\chi$ 0.88 $+$0.08 $-$0.12 0.0 - 0.99
$K_{\rm planet}$ \[ms$^{-1}$\] 2.60 $+$0.33 $-$0.30 0.0 - 5.0
$a$ \[AU\] 0.066610 $+$0.000011 $-$0.000013 –
$m_p \sin i$ \[M$_{\rm Earth}$\] 5.36 $+$0.69 $-$0.62 –
$P_{\rm rot}$ \[d\] 43.88 $+$0.03 $-$0.10 42.5 - 45.0
$K_{\rm rot}$ \[ms$^{-1}$\] 2.26 $+$0.92 $-$0.46 0.0 - 7.0
RV noise \[ms$^{-1}$\] 1.81 $+$0.18 $-$0.17 0.0 - 5.0
Discussion
==========
We detect the presence of a planet with a semi-amplitude of 2.60 m s$^{-1}$, which – given the stellar mass of 0.52 M$_{\odot}$ – converts to m sin $i$ of 5.36 M$_{\oplus}$, orbiting with a period of 8.7 d around GJ 536, an M-type star of 0.52 M$_{\odot}$ with a rotation period of 43.9 d that shows an additional activity signal compatible with an activity cycle shorter than 3 yr.
The planet is a small super-Earth with an equilibrium temperature 344 K for a Bond albedo A = 0.75 and 487 K for A=0. Following @Kasting1993 and @Selsis2007, we perform a simple estimation of the habitable zone (HZ) of this star. The HZ would go from 0.2048 to 0.3975 AU in the narrowest case (cloud free model), and from 0.1044 to 0.5470 AU in the broadest case (fully clouded model). This corresponds to orbital periods ranging from 46 to 126 days in the narrowest case, and from 17 to 204 days in the broadest one.
GJ 536 b is in the lower part of the mass vs period diagram of known planets around M-dwarf stars (Fig. \[period\_mass\]). The planet is too close to the star to be considered habitable. For this star the orbital periods of the habitable zone would be from $\sim$20 days to $\sim$ 40 days.
GJ 536 is a quiet early M dwarf, with a rotation period at the upper end of the stars of its kind [@Newton2016; @Masca2016a]. Its rotation induced radial-velocity signal has a semi-amplitude of 2.26 ms$^{-1}$ and seems to be stable enough to allow for a clean enough periodogram and to be correctly characterized. The phase of the rotation induced signal seems to be advanced by $\sim$45$^\circ$ with respect to the signals in the S$_{MW}$ index and FWHM time series. There is a hint of an activity cycle shorter than 3 yr, which would put it at the lower end of the stars of its kind [@Masca2016a], and whose amplitude is so small that would need further follow-up to be properly characterized. The radial-velocity signal induced by this cycle at this point is beyond our detection capabilities.
Given the rms of the residuals there is still room for the detection of more planets in this system, especially at orbital periods longer than the rotation period. Fig. \[period\_mass\] shows the upper limits to the mass of those hypothetical companions. The stability of its rotation signals and the low amplitude of the radial-velocity signals with a magnetic origin makes this star a good candidate to search for longer period planets of moderate mass. A rough estimate of the detection limits tells us there is still room for Earth-like planets ($\sim$ 1 M$_{\oplus}$) at orbits smaller than 10, super-Earths ($\textless$ 10 M$_{\oplus}$) at orbits going from 10 to 400 days, and even for a Neptune-mass planet ($\textless$ 20 M$_{\oplus}$) at periods longer than $\sim$3 yr. Giant planets, on the other hand, are discarded except for those with extremely long orbital periods. The time-span of the observations and the RMS of the residuals completely rules out the presence of any planet bigger than twice the mass of Neptune with an orbital period shorter than $\sim$ 20 years.
{width="18cm"}
Conclusions
===========
We have analysed 152 high-resolution spectra and 359 photometric observations to study the planetary companions around the M-dwarf star GJ 536 and its stellar activity. We detected two significant radial-velocity signals at periods of 8.7 and 43.8 days, respectively.
From the available photometric and spectroscopic information we conclude that the 8.7 d signal is caused by a 5.3 M$_{\oplus}$ planet with semi-major axis of 0.067 AU and equilibrium temperature lower than 500K. The short period of the planet makes it a potential transiting candidate. Detecting the transits would give a new constraining point to the mass-radius diagram.
The second radial-velocity signal of period 43.8 d and semi-amplitude of 1.6 ms$^{-1}$ is a magnetic activity induced signal related to the rotation of the star. We also found a magnetic cycle shorter than 3 yr which would place this star among those with the shortest reported magnetic cycles.
We have studied and set limits to the presence of other planetary companions taking into account the rms of the residuals after fitting both the planet and the rotation induced signal. The system still has room for other low-mass companions, but planets more massive than Neptune are discarded except at extremely long orbital periods beyond the habitable zone of the star.
Acknowledgements {#acknowledgements .unnumbered}
================
This work has been financed by the Spanish Ministry project MINECO AYA2014-56359-P. J.I.G.H. acknowledges financial support from the Spanish MINECO under the 2013 Ramón y Cajal program MINECO RYC-2013-14875. X.B., X.D., T.F., and F.M. acknowledge the support of the French Agence Nationale de la Recherche (ANR) under the programme ANR-12-BS05- 0012 Exo-atmos. X.B. and A.W. acknowledge funding from the European Research Council under the ERC Grant Agreement No. 337591-ExTrA. This work was supported by Fundação para a Ciência e a Tecnologia (FCT) within projects reference PTDC/FIS-AST/1526/2014 (POCI-01-0145-FEDER-016886) and UID/FIS/04434/2013 (POCI-01-0145-FEDER-007672). N.C.S. acknowledges support through Investigador FCT contract of reference IF/00169/2012, and POPH/FSE (EC) by FEDER funding through the program “Programa Operacional de Factores de Competitividade - COMPETE”. This work is based on data obtained via the HARPS public database at the European Southern Observatory (ESO). This research has made extensive use of the SIMBAD database, operated at CDS, Strasbourg, France, and NASA’s Astrophysics Data System. We are grateful to all the observers of the following ESO projects, whose data we are using: 072.C-0488, 085.C-0019, 183.C-0972, and 191.C-087.
Full Dataset {#append_a}
============
[@llccccccccccc@]{}
\
\
\
BJD - 2450000 & V$_{r}$ & $\sigma$ V$_{r}$ & FWHM & BIS Span & S$_{MW}$ index & $\sigma$ S$_{MW}$ & H$_{\alpha}$ index & $\sigma$ H$_{\alpha}$ index & Flag\
(d) & (ms$^{-1}$) & (ms$^{-1}$) & (Kms$^{-1}$) & (ms$^{-1}$)\
3202.5590 & -25616.5610 & 1.3488 & 4444.8091 & -5.6623 & 1.0772 & 0.0077 & 0.4960 & 0.0005\
3579.4972 & -25622.4158 & 1.3573 & 4448.4550 & -5.0540 & 1.2490 & 0.0073 & 0.5058 & 0.0005\
3811.8370 & -25620.2605 & 1.4564 & 4438.6871 & -9.9863 & 1.1320 & 0.0078 & 0.5043 & 0.0006\
3813.8047 & -25621.7093 & 1.2228 & 4433.0342 & -8.1469 & 1.1526 & 0.0053 & 0.5047 & 0.0004\
4196.7394 & -25621.5343 & 1.4416 & 4443.0970 & -6.2410 & 1.0422 & 0.0074 & 0.5001 & 0.0006\
4202.7156 & -25621.6073 & 1.2951 & 4425.1466 & -6.9079 & 1.0538 & 0.0059 & 0.5039 & 0.0005\
4340.4836 & -25623.6979 & 1.3024 & 4429.4257 & -7.9120 & 1.1163 & 0.0070 & 0.5076 & 0.0005\
4525.8756 & -25618.1517 & 1.2817 & 4433.4956 & -11.6982 & 1.1585 & 0.0063 & 0.5069 & 0.0005\
4528.8393 & -25622.4852 & 1.3126 & 4433.8369 & -5.2346 & 1.1304 & 0.0064 & 0.5044 & 0.0005\
4591.7914 & -25625.2134 & 1.5565 & 4442.0880 & -7.2876 & 0.9757 & 0.0090 & 0.4940 & 0.0007\
4703.4993 & -25621.8483 & 1.6802 & 4433.6780 & -10.1303 & 1.1269 & 0.0107 & 0.5032 & 0.0008\
5226.8854 & -25621.1796 & 1.4632 & 4446.4284 & -10.0405 & 1.2068 & 0.0081 & 0.4996 & 0.0006\
5281.7491 & -25623.1861 & 1.3854 & 4428.1838 & -8.6441 & 1.0451 & 0.0068 & 0.4889 & 0.0006\
5305.7265 & -25617.9212 & 1.3201 & 4427.4995 & -9.9302 & 1.0898 & 0.0084 & 0.5030 & 0.0005\
5306.7140 & -25617.4308 & 1.3665 & 4430.6156 & -8.9673 & 1.1605 & 0.0092 & 0.5063 & 0.0006\
5307.7196 & -25616.7602 & 1.2588 & 4436.3939 & -10.4465 & 1.2272 & 0.0087 & 0.5109 & 0.0004\
5308.7013 & -25614.7745 & 1.3191 & 4446.0321 & -8.0523 & 1.1629 & 0.0088 & 0.5039 & 0.0005\
5309.6925 & -25615.3506 & 1.3031 & 4442.9390 & -8.8002 & 1.2196 & 0.0092 & 0.5080 & 0.0005\
6385.6469 & -25626.3858 & 1.3821 & 4428.1616 & -3.0622 & 0.8991 & 0.0101 & 0.4978 & 0.0007\
6386.7448 & -25624.9118 & 1.2594 & 4425.8662 & -9.9115 & 0.8084 & 0.0078 & 0.4935 & 0.0007\
6387.7815 & -25624.7334 & 1.1527 & 4424.2671 & -8.2783 & 0.8489 & 0.0073 & 0.4952 & 0.0006\
6388.7254 & -25623.3896 & 1.3647 & 4416.3803 & -7.3567 & 0.8284 & 0.0087 & 0.4964 & 0.0007\
6389.7264 & -25621.3622 & 1.3006 & 4421.5546 & -11.3118 & 0.8655 & 0.0085 & 0.5048 & 0.0007\
6390.7371 & -25620.4734 & 1.6898 & 4419.3411 & -9.8647 & 0.8066 & 0.0105 & 0.4943 & 0.0008\
6391.7497 & -25620.8694 & 1.3286 & 4424.2490 & -11.7541 & 0.9001 & 0.0085 & 0.5003 & 0.0007\
6393.7913 & -25623.1815 & 1.2910 & 4415.1907 & -7.5922 & 0.8531 & 0.0089 & 0.4978 & 0.0007\
6394.7750 & -25624.4828 & 1.2851 & 4429.5659 & -9.1806 & 0.8915 & 0.0088 & 0.4949 & 0.0007\
6395.7000 & -25626.0196 & 1.2787 & 4417.1293 & -9.7255 & 0.9155 & 0.0089 & 0.4957 & 0.0007\
6396.7103 & -25623.1147 & 1.4092 & 4424.7395 & -12.7127 & 0.9218 & 0.0094 & 0.4968 & 0.0008\
6397.6863 & -25621.0933 & 1.3000 & 4423.8606 & -9.5686 & 0.9836 & 0.0095 & 0.5064 & 0.0007\
6398.6799 & -25618.8329 & 1.4313 & 4426.1457 & -7.6371 & 0.9339 & 0.0099 & 0.4992 & 0.0008\
6399.6958 & -25619.5072 & 1.3309 & 4440.7943 & -9.3043 & 1.0242 & 0.0097 & 0.5024 & 0.0008\
6400.6899 & -25617.6148 & 1.3276 & 4426.9125 & -10.7879 & 0.9848 & 0.0094 & 0.5022 & 0.0008\
6401.6532 & -25624.7094 & 1.2229 & 4430.4209 & -8.8857 & 1.0296 & 0.0087 & 0.5028 & 0.0007\
6402.6436 & -25623.9620 & 1.5248 & 4430.2528 & -9.1200 & 1.0116 & 0.0108 & 0.5027 & 0.0008\
6403.6245 & -25622.1027 & 1.3310 & 4443.0225 & -9.7811 & 1.0437 & 0.0098 & 0.5032 & 0.0007\
6404.6425 & -25623.0456 & 1.4532 & 4434.0050 & -10.1067 & 1.0945 & 0.0109 & 0.5106 & 0.0008\
6410.6262 & -25626.0065 & 1.5706 & 4448.4592 & -8.4057 & 0.9966 & 0.0115 & 0.5039 & 0.0008\
6414.6393 & -25623.0607 & 1.4853 & 4439.0697 & -6.4884 & 1.0073 & 0.0104 & 0.5072 & 0.0008\
6415.5922 & -25622.4846 & 2.8093 & 4434.8839 & -17.9962 & 0.8264 & 0.0184 & 0.5053 & 0.0014\
6415.7332 & -25621.0645 & 1.6543 & 4445.0213 & -9.2692 & 1.0473 & 0.0124 & 0.5114 & 0.0009\
6416.6954 & -25620.1184 & 1.2661 & 4431.5240 & -4.3249 & 0.9507 & 0.0089 & 0.5034 & 0.0007\
6451.5800 & -25616.3891 & 1.3405 & 4433.3694 & -16.8458 & 1.0419 & 0.0096 & 0.5063 & 0.0007\
6452.5545 & -25618.6586 & 1.3803 & 4434.6138 & -8.7892 & 1.0683 & 0.0081 & 0.5099 & 0.0007\
6454.5556 & -25623.8325 & 1.3937 & 4431.7379 & -9.5482 & 1.0701 & 0.0103 & 0.5075 & 0.0008\
6455.5374 & -25628.9792 & 1.4965 & 4422.2962 & -7.9232 & 0.9761 & 0.0106 & 0.5020 & 0.0008\
6458.5877 & -25622.2235 & 1.2855 & 4423.7546 & -11.6472 & 1.0516 & 0.0093 & 0.5101 & 0.0007\
6460.5668 & -25621.7813 & 1.6473 & 4431.8028 & -10.6556 & 0.9902 & 0.0115 & 0.5048 & 0.0008\
6481.4839 & -25619.0212 & 1.3733 & 4419.7151 & -8.8134 & 0.9295 & 0.0089 & 0.5034 & 0.0007\
6508.4718 & -25627.1324 & 1.8094 & 4441.7103 & -10.4168 & 0.8926 & 0.0119 & 0.5119 & 0.0010\
6514.4694 & -25623.2844 & 1.3027 & 4424.7201 & -6.1677 & 0.8424 & 0.0087 & 0.5033 & 0.0007\
6521.4589 & -25619.8433 & 1.2696 & 4431.0487 & -10.7298 & 0.8505 & 0.0081 & 0.5007 & 0.0007\
6690.8780 & -25624.1045 & 1.1926 & 4422.6617 & -9.2806 & 0.9930 & 0.0082 & 0.5063 & 0.0006\
6691.8339 & -25624.6074 & 1.3498 & 4426.5396 & -8.4985 & 1.1321 & 0.0107 & 0.5168 & 0.0007\
6692.8139 & -25625.0787 & 1.2900 & 4431.9081 & -10.8969 & 1.0794 & 0.0099 & 0.5048 & 0.0006\
6694.8640 & -25624.0179 & 1.1386 & 4433.0913 & -9.0442 & 1.0415 & 0.0081 & 0.5084 & 0.0006\
6695.8790 & -25622.2008 & 1.1906 & 4429.7997 & -8.1605 & 1.0384 & 0.0085 & 0.5108 & 0.0006\
6696.8539 & -25622.8668 & 1.3253 & 4428.3856 & -10.0882 & 1.2174 & 0.0107 & 0.5241 & 0.0007\
6697.7981 & -25625.2516 & 1.3466 & 4425.1099 & -6.5195 & 1.0566 & 0.0105 & 0.5067 & 0.0007\
6712.8127 & -25613.5661 & 1.3395 & 4437.9343 & -10.5794 & 1.5124 & 0.0112 & 0.5544 & 0.0008 & Rejected\
6713.8033 & -25613.2583 & 1.3276 & 4450.3946 & -9.6570 & 1.2983 & 0.0109 & 0.5214 & 0.0007\
6715.7953 & -25620.9555 & 1.3577 & 4436.1206 & -6.6456 & 1.2225 & 0.0106 & 0.5213 & 0.0007\
6720.8502 & -25616.0066 & 1.2598 & 4432.3554 & -9.5264 & 1.1898 & 0.0094 & 0.5123 & 0.0006\
6723.8540 & -25622.0386 & 1.2931 & 4439.7320 & -11.2245 & 1.2372 & 0.0103 & 0.5198 & 0.0007\
6724.7853 & -25624.4497 & 1.2257 & 4440.9264 & -10.5082 & 1.2255 & 0.0097 & 0.5205 & 0.0006\
6725.7743 & -25624.1936 & 1.2975 & 4439.1990 & -10.9028 & 1.1683 & 0.0103 & 0.5152 & 0.0007\
6725.8844 & -25626.5410 & 1.3803 & 4440.6907 & -7.2889 & 1.1309 & 0.0105 & 0.5138 & 0.0007\
6726.7959 & -25625.8960 & 1.1504 & 4442.3973 & -7.2719 & 1.0898 & 0.0084 & 0.5101 & 0.0006\
6727.8296 & -25620.9225 & 1.1367 & 4430.0904 & -10.2483 & 1.0773 & 0.0079 & 0.5083 & 0.0006\
6728.8039 & -25621.7110 & 1.1250 & 4434.7157 & -8.6990 & 1.0786 & 0.0080 & 0.5062 & 0.0005\
6729.7718 & -25616.4237 & 1.4276 & 4437.5013 & -12.1011 & 1.0759 & 0.0085 & 0.5096 & 0.0006\
6730.8216 & -25619.5981 & 1.3086 & 4433.6524 & -10.6152 & 1.0424 & 0.0095 & 0.5111 & 0.0007\
6732.7980 & -25622.9502 & 1.4351 & 4421.7672 & -4.3220 & 1.0392 & 0.0101 & 0.5106 & 0.0008\
6737.8572 & -25620.5708 & 1.3554 & 4428.3888 & -5.4130 & 1.0060 & 0.0093 & 0.5102 & 0.0007\
6738.8726 & -25622.5574 & 1.2338 & 4431.8640 & -7.8770 & 0.9898 & 0.0086 & 0.5047 & 0.0007\
6739.8058 & -25622.1678 & 1.1512 & 4425.2586 & -9.5216 & 1.0135 & 0.0078 & 0.5075 & 0.0006\
6740.8311 & -25624.4843 & 1.0987 & 4428.1852 & -6.4714 & 0.9862 & 0.0073 & 0.5061 & 0.0005\
6741.7462 & -25623.9200 & 1.1678 & 4429.0187 & -9.0739 & 1.0092 & 0.0081 & 0.5038 & 0.0006\
6742.8207 & -25623.1636 & 1.1028 & 4426.1406 & -9.8389 & 1.0343 & 0.0076 & 0.5094 & 0.0005\
6743.7632 & -25622.7840 & 1.2011 & 4432.8256 & -6.7370 & 1.0186 & 0.0084 & 0.5069 & 0.0006\
6745.7321 & -25617.2597 & 1.1588 & 4423.0781 & -11.4489 & 1.0638 & 0.0081 & 0.5124 & 0.0007\
6746.8203 & -25613.0541 & 1.2944 & 4428.6487 & -8.3582 & 1.0296 & 0.0090 & 0.5158 & 0.0007\
6752.8315 & -25621.4802 & 1.3848 & 4432.6911 & -7.1023 & 1.0633 & 0.0106 & 0.5210 & 0.0008\
6754.8603 & -25617.1072 & 2.0773 & 4434.8424 & -14.7497 & 1.1186 & 0.0161 & 0.5213 & 0.0011\
6755.8430 & -25617.0817 & 2.0606 & 4444.2810 & -9.2983 & 1.0480 & 0.0157 & 0.5182 & 0.0011\
6755.8530 & -25614.7730 & 1.9338 & 4449.5330 & -8.2994 & 1.1561 & 0.0154 & 0.5220 & 0.0010\
6756.8521 & -25616.6832 & 1.0939 & 4440.9586 & -9.8160 & 1.1383 & 0.0085 & 0.5169 & 0.0005\
6757.8085 & -25617.4805 & 1.4014 & 4441.8165 & -9.7412 & 1.1327 & 0.0112 & 0.5181 & 0.0007\
6758.8266 & -25622.4824 & 2.2098 & 4438.0807 & -15.1975 & 1.0999 & 0.0174 & 0.5266 & 0.0012\
6759.8277 & -25621.4838 & 1.3814 & 4438.3218 & -7.9763 & 1.1324 & 0.0111 & 0.5192 & 0.0007\
6760.8142 & -25620.2626 & 1.3119 & 4448.0559 & -8.2829 & 1.5703 & 0.0125 & 0.5642 & 0.0007 & Rejected\
6763.7243 & -25620.4110 & 1.0500 & 4442.6357 & -10.9225 & 1.1538 & 0.0073 & 0.5130 & 0.0005\
6764.7765 & -25618.1639 & 1.2340 & 4444.4456 & -9.9255 & 1.1184 & 0.0092 & 0.5128 & 0.0006\
6765.7208 & -25619.0905 & 1.3628 & 4437.2341 & -6.5105 & 1.1466 & 0.0078 & 0.5131 & 0.0006\
6766.7265 & -25621.2292 & 1.0945 & 4439.1904 & -9.8447 & 1.2270 & 0.0081 & 0.5229 & 0.0005\
6767.6534 & -25623.9784 & 1.5745 & 4433.4716 & -10.2058 & 1.1048 & 0.0115 & 0.5113 & 0.0008\
6768.6678 & -25625.3559 & 1.2107 & 4429.1243 & -7.0854 & 1.0659 & 0.0087 & 0.5089 & 0.0006\
6778.6271 & -25624.2948 & 1.3732 & 4424.6235 & -8.2880 & 0.9981 & 0.0077 & 0.5003 & 0.0005\
6779.7560 & -25623.0220 & 1.5571 & 4433.6733 & -7.6866 & 0.9779 & 0.0094 & 0.5036 & 0.0007\
6781.6011 & -25621.5230 & 1.5651 & 4434.4575 & -9.5425 & 0.9537 & 0.0096 & 0.5022 & 0.0007\
6782.6156 & -25621.6172 & 1.3793 & 4433.3780 & -4.7485 & 1.1337 & 0.0085 & 0.5159 & 0.0005\
6784.6137 & -25625.6796 & 1.2493 & 4430.6030 & -6.3606 & 0.9865 & 0.0089 & 0.5079 & 0.0006\
6785.5546 & -25623.9276 & 1.5364 & 4411.9798 & -7.3440 & 0.9995 & 0.0118 & 0.5097 & 0.0008\
6786.6679 & -25628.0817 & 1.1219 & 4420.8351 & -8.7321 & 0.9403 & 0.0073 & 0.5038 & 0.0005\
6814.7183 & -25618.4375 & 1.3822 & 4442.1718 & -6.9710 & 1.0105 & 0.0109 & 0.5029 & 0.0007\
6822.5823 & -25625.6969 & 1.6325 & 4427.8930 & -14.3619 & 1.0863 & 0.0119 & 0.5063 & 0.0009\
6823.5834 & -25627.6799 & 1.3314 & 4432.6413 & -9.5886 & 1.0479 & 0.0097 & 0.5004 & 0.0007\
6824.5777 & -25623.2701 & 1.4221 & 4430.6359 & -9.1015 & 1.0257 & 0.0097 & 0.4996 & 0.0007\
6825.6520 & -25622.3060 & 1.3593 & 4435.5327 & -7.2742 & 1.0811 & 0.0105 & 0.5055 & 0.0007\
6826.5764 & -25621.9304 & 1.2330 & 4428.6392 & -4.9112 & 1.3187 & 0.0094 & 0.5315 & 0.0007 & Rejected\
6827.5754 & -25624.3321 & 1.1535 & 4432.4085 & -6.5820 & 1.0300 & 0.0082 & 0.5024 & 0.0006\
6828.6006 & -25624.5384 & 1.1901 & 4429.2042 & -10.7163 & 1.1152 & 0.0086 & 0.5094 & 0.0006\
6838.5568 & -25626.1804 & 1.1653 & 4430.8116 & -10.7542 & 1.0408 & 0.0083 & 0.5163 & 0.0006\
6839.5704 & -25622.4896 & 1.4164 & 4425.8922 & -8.3993 & 0.9890 & 0.0105 & 0.5122 & 0.0006\
6840.5286 & -25620.4583 & 1.4235 & 4421.9119 & -7.9951 & 0.9953 & 0.0097 & 0.5107 & 0.0007\
6841.6035 & -25614.8073 & 2.5202 & 4438.8469 & -2.9369 & 1.1631 & 0.0220 & 0.5331 & 0.0013 & Rejected\
6842.4896 & -25617.9822 & 1.3193 & 4436.3770 & -9.2662 & 1.0293 & 0.0094 & 0.5150 & 0.0007\
6857.5388 & -25625.2710 & 1.7446 & 4421.9564 & -10.6323 & 1.0678 & 0.0125 & 0.5036 & 0.0009\
6858.5182 & -25622.6933 & 1.4250 & 4432.2060 & -9.9027 & 1.0034 & 0.0099 & 0.4929 & 0.0007\
6863.5169 & -25622.6876 & 1.3877 & 4434.8193 & -7.8466 & 1.0297 & 0.0103 & 0.5056 & 0.0007\
6864.5176 & -25624.8996 & 1.1102 & 4424.7214 & -8.4713 & 0.9533 & 0.0077 & 0.4994 & 0.0005\
6874.4791 & -25620.6138 & 1.6580 & 4440.2229 & -5.6786 & 1.0463 & 0.0121 & 0.5086 & 0.0009\
7047.8603 & -25624.7545 & 1.2191 & 4434.2142 & -8.8214 & 1.0933 & 0.0090 & 0.5089 & 0.0006\
7053.8561 & -25621.7439 & 1.4441 & 4446.4398 & -7.8204 & 1.0895 & 0.0103 & 0.5060 & 0.0007\
7057.8269 & -25622.0109 & 1.4588 & 4452.3988 & -8.1500 & 1.5121 & 0.0123 & 0.5507 & 0.0008 & Rejected\
7058.8515 & -25617.6495 & 1.8982 & 4442.7452 & -17.1716 & 1.0415 & 0.0136 & 0.5044 & 0.0010\
7079.8236 & -25619.7617 & 1.2772 & 4436.6205 & -11.9584 & 1.1289 & 0.0094 & 0.5148 & 0.0007\
7080.8500 & -25621.6358 & 1.4059 & 4427.7199 & -5.2545 & 1.0410 & 0.0100 & 0.5063 & 0.0007\
7082.8651 & -25623.5924 & 1.1403 & 4434.4908 & -7.9841 & 1.0022 & 0.0078 & 0.5025 & 0.0006\
7085.7333 & -25621.7323 & 1.2833 & 4430.9624 & -8.7143 & 1.0902 & 0.0098 & 0.5114 & 0.0006\
7114.8209 & -25625.5802 & 1.4614 & 4410.2902 & -14.3900 & 0.9577 & 0.0102 & 0.5028 & 0.0008\
7115.7150 & -25627.5907 & 1.3311 & 4428.5220 & -8.7149 & 1.0657 & 0.0092 & 0.5075 & 0.0007\
7116.7852 & -25627.3237 & 1.2820 & 4424.2680 & -7.4215 & 0.9395 & 0.0083 & 0.5014 & 0.0007\
7142.7719 & -25625.3593 & 1.2290 & 4428.5104 & -8.7045 & 0.9850 & 0.0083 & 0.5065 & 0.0006\
7147.7808 & -25616.3909 & 1.3538 & 4441.0481 & -9.6390 & 0.9792 & 0.0103 & 0.5038 & 0.0007\
7148.7468 & -25620.8954 & 1.2741 & 4429.2924 & -10.0502 & 0.9918 & 0.0094 & 0.5001 & 0.0006\
7202.5939 & -25616.1481 & 1.6565 & 4441.1922 & 5.6209 & 0.9188 & 0.0097 & 0.4982 & 0.0008\
7204.6007 & -25623.6987 & 1.4010 & 4452.3792 & 2.2721 & 0.9045 & 0.0081 & 0.4978 & 0.0006\
7211.5712 & -25624.5994 & 1.3200 & 4447.6138 & 2.6096 & 0.8998 & 0.0075 & 0.4998 & 0.0005\
7212.6084 & -25624.9280 & 5.6243 & 4437.5059 & -12.1776 & 0.4064 & 0.0213 & 0.5013 & 0.0024\
7214.5883 & -25625.9395 & 2.0090 & 4454.7852 & 5.6477 & 0.8619 & 0.0135 & 0.5093 & 0.0010\
7238.5220 & -25620.5551 & 1.4908 & 4444.8883 & 4.9186 & 0.9594 & 0.0104 & 0.5031 & 0.0006\
7249.4828 & -25623.8234 & 1.8006 & 4447.2116 & 4.2354 & 0.8713 & 0.0112 & 0.4972 & 0.0008\
7448.8620 & -25628.4475 & 1.4117 & 4457.4186 & 6.7720 & 0.9810 & 0.0103 & 0.5065 & 0.0007\
7473.8467 & -25621.6222 & 1.1138 & 4455.5312 & 0.5160 & 0.9392 & 0.0082 & 0.5075 & 0.0005\
7476.8649 & -25621.8943 & 1.1800 & 4459.2948 & 3.4832 & 0.9172 & 0.0083 & 0.5028 & 0.0006\
7508.4799 & -25624.9467 & 1.0813 & 4467.1596 & -5.9534 & 1.0357 & 0.0107 & 0.5091 & 0.0005 & HARPS-N\
7508.5698 & -25621.9227 & 1.0769 & 4464.6067 & -7.1361 & 1.2192 & 0.0108 & 0.5287 & 0.0005 & HARPS-N\
7509.4759 & -25625.1223 & 1.2681 & 4460.0000 & -9.6134 & 0.9486 & 0.0111 & 0.4974 & 0.0005 & HARPS-N\
7509.5684 & -25626.8176 & 1.4120 & 4462.4128 & -8.3783 & 0.9415 & 0.0128 & 0.4969 & 0.0006 & HARPS-N\
7510.4709 & -25621.9484 & 1.3119 & 4461.3527 & -5.7705 & 0.9028 & 0.0117 & 0.4927 & 0.0006 & HARPS-N\
7510.5488 & -25620.3344 & 1.2794 & 4467.5365 & -5.4813 & 1.0562 & 0.0124 & 0.5053 & 0.0006 & HARPS-N\
7535.4250 & -25625.4364 & 2.1712 & 4464.4853 & -2.7902 & 0.9369 & 0.0239 & 0.5095 & 0.0012 & HARPS-N\
7536.4320 & -25624.6383 & 1.1442 & 4463.1872 & -8.4287 & 0.9829 & 0.0134 & 0.5065 & 0.0006 & HARPS-N\
7537.4339 & -25619.0095 & 1.3012 & 4461.0889 & -10.0375 & 0.9686 & 0.0148 & 0.5083 & 0.0007 & HARPS-N\
7537.5223 & -25618.7979 & 1.1950 & 4464.6137 & -6.3194 & 1.0328 & 0.0149 & 0.5105 & 0.0007 & HARPS-N\
7538.4145 & -25618.5006 & 1.0871 & 4459.3290 & -6.8755 & 0.9491 & 0.0118 & 0.5029 & 0.0006 & HARPS-N\
7538.5184 & -25615.3666 & 1.0356 & 4470.6070 & -7.3075 & 0.9111 & 0.0099 & 0.5017 & 0.0005 & HARPS-N\
\[lastpage\]
[^1]: The data used in this paper (Table \[tab\_a1\]) is available in electronic form at the CDS via anonymous ftp to cdsarc.u-strasbg.fr (130.79.128.5) or via http://cdsweb.u-strasbg.fr/cgi-bin/qcat?J/A+A/
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
[**Mohamed Ben Ayed** ]{}\
[*Department de Mathematiques*]{},\
[*Faculte des Sciences de Sfax*]{}, [*Route Soukra, Sfax, Tunisia*]{}\
[*e-mail: [email protected]*]{}\
[**Khalil El Mehdi** ]{}\
[*Centre de Math[é]{}matique* ]{},\
[*Ecole Polytechnique*]{}, [*91128 Palaiseau, France*]{}\
[*e-mail: [email protected]*]{}\
[*and* ]{}\
[*Faculte des Sciences et techniques* ]{},\
[*Universite de Nouakchott, Mauritanie*]{}\
[**Mohameden Ould Ahmedou**]{}\
[*Mathematisches Institut*]{},\
[*Beringstrasse 4*]{},\
[*D-53115 Bonn, Germany* ]{}\
[*e-mail: [email protected]*]{}
date:
title:
---
[**Abstract** ]{}
[*1991 Mathematics Subject Classification*]{}. [35J60, 53C21, 58G30 ]{}.\
[*Key words*]{}. [Variational problems, Lack of compactness, Scalar curvature, Conformal invariance, Critical points at infinity ]{}
Introduction and the Main Results
==================================
In this paper we study some nonlinear problem arising from conformal geometry. Precisely, consider a Riemannian manifold with boundary $(M^n,g)$ of dimension $n\geq 3$ and take $\tilde{g}=u^{4/(n-2)}g$, be a conformal metric to $g$, where $u$ is a smooth positive function, then the following equations relate the scalar curvatures $R_g$, $R_{\tilde{g}}$ and the mean curvatures of the boundary $h_g$, $h_{\tilde{g}}$, with respect to $g$ and $\tilde{g}$ respectively. $$\begin{aligned}
(P_1)\quad \left\{
\begin{array}{ccccc}
-c_n{\Delta }_gu+R_gu&=&R_{\tilde{g}}u^{\frac{n+2}{n-2}}&\mbox{ in }& M\\
\frac{2}{n-2}\frac{\partial u}{\partial \nu}+h_g u &=&
h_{\tilde{g}}u^\frac{n}{n-2}&
\mbox{ on }& \partial M
\end{array}
\right.\end{aligned}$$ where $c_n=4(n-1)/(n-2)$ and $\nu$ denotes the outward normal vector with respect to the metric $g$.\
In view of the above equations, a natural question is whether it is possible to prescribe both the scalar curvature and the boundary mean curvature, that is : given two functions $K: M\to {\mathbb{R}}$ and $H: \partial M\to {\mathbb{R}}$, does exists a metric $\tilde{g}$ conformally equivalent to $g$ such that $R_{\tilde{g}}=K$ and $h_{\tilde{g}}=H$?\
According to equations $ (P_1)$, the problem is equivalent to finding a smooth positive solution $u$ of the following equation $$\begin{aligned}
(P_2)\quad \left\{
\begin{array}{ccccc}
-c_n{\Delta }_gu+R_gu&=& K u^{\frac{n+2}{n-2}}&\mbox{ in }& M \\
\frac{2}{n-2}\frac{\partial u}{\partial \nu}+h_g u &= &
H u^\frac{n}{n-2}&
\mbox{ on }& \partial M.
\end{array}
\right.\end{aligned}$$ Such a problem was studied in [@ALM] [@C],[@DMA] [@E1], [@E2], [@HL1], [@HL2] [@L1] . Yanyan Li [@L1], and Djadli-Malchiodi-Ould Ahmedou [@DMA] studied this problem when the manifold is the three dimensional standard half sphere. Their approach involves a fine blow up analysis of some subcritical approximations and the use of the topological degree tools.
Regarding the above problem it is well known that the most interesting case is the so called positive one, that is when the quadratic part of the associated Euler functional is positive definite. Another interesting case is when a noncompact group of conformal transformations acts on the equation leading to topological obstructions. The half sphere represents the simplest case where such a noncompactness occurs, and in this paper we consider the case of the standard half sphere under minimal boundary conditions:
More precisely, let $$S_+^n=\{x\in {\mathbb{R}}^{n+1}\, /\, |x|=1,\, x_{n+1}>0\}$$ $n\geq 3$ . Given a $C^2$ function $K$ on $\overline{S_+^n}$, we look for conditions on $K$ to ensure the existence of a positive solution of the problem $$\begin{aligned}
(1)\quad \left\{
\begin{array}{ccccc}
-{\Delta }_gu+ \frac{n(n-2)}{4}u&=& K u^{\frac{n+2}{n-2}}&\mbox{ in }& S_+^n\\
\frac{\partial u}{\partial \nu}&=&0&
\mbox{ on }& \partial S_+^n
\end{array}
\right.\end{aligned}$$ where $g$ is the standard metric of $S_+^n$.\
Problem (1) is in some sense related to the well known scalar curvature problem on $S^n$ $$\begin{aligned}
(2)\qquad-{\Delta }_gu+ \frac{n(n-2)}{4}u= K u^{\frac{n+2}{n-2}}\quad \mbox{ in } S^n\end{aligned}$$ to which much work has been devoted. For details please see [@B1],[@BC1] , [@BCCH],[@BCH], [@CY], [@CGY], [@H], [@L2], [@L3],[@SZ] and the references therein.\
As for (2), there are topological obstructions of Kazdan-Warner type to solve (1) (see [@BP]) and so a naturel question arises: under which conditions on $K$, (1) has a positive solution. We propose to handle such a question, using some topological and dynamical tools of the theory of critical points at infinity, see Bahri [@B1] , [@B2].\
Our approch follows closely the ideas developped in Aubin-Bahri [@AB], Bahri [@B1] and Ben Ayed-Chtioui-Hammami [@BCH] where the problem of prescribing the scalar curvature on closed manifolds was studied using some algebraic topological tools. The main idea is to use the difference of topology between the level sets of the function $K$ to produce a critical point of the Euler functional $J$ associated to (1) and the main issue is under which conditions on $K$, a topological accident between the level sets of $K$ induces a topological accident between the level sets of $J$. Such an accident is sufficient to prove the existence of a critical point when some compactness conditions are satisfied. However our problem presents a lack of compactness due to the presence of critical points at infinity, that is noncompact orbits for the gradient of $J$ along which $J$ is bounded and its gradient goes to zero. Therefore a careful study of such noncompact orbits is necessary, in order to take into account their contribution to the difference of topology between the level sets of $J$.\
In order to state our main results, we need to introduce the assumptions that we are using in our results.\
${\bf (A_1)}$0.3cm We assume that $K_1=K_{|\partial S_+^n}$ has only nondegenerate critical points $y_0,...,y_s$, where $y_0$ is the absolute maximum, such that $$K(y_0)\geq K(y_1)\geq ...\geq K(y_l)> K(y_{l+1}\geq ...\geq K(y_s)$$ with $$\frac{\partial K}{\partial \nu}(y_i)>0,\, \mbox{ for } 0\leq i\leq l, \qquad
\frac{\partial K}{\partial \nu}(y_i)\leq 0, \, \mbox{ for } l+1\leq i\leq s.$$ ${\bf (A_2)}$0.3cm Assume that there exists $c$ a postive constant such that $c < K(y_l)$, and every $y$ critical point of $K$, $K(y) < c $.\
${\bf (A_3)}$ 0.3cm Let $Z$ be a pseudogradient of $K_1$, of Morse-Smale type (that is the intersections of the stable and the unstable manifolds of the critical points of $K_1$ are transverse.)\
Set $$X=\overline{\cup_{0\leq i\leq l}W_s(y_i)}$$ where $W_s(y_i)$ is the stable manifold of $y_i$ for $Z$.\
We assume that $X$ is not contractible and denote by $m$ the dimension of the first nontrivial reduced homological group.\
${\bf (A_4)}$ 0.3cm Assume that $X$ is contractible in $K^c=\{x\in S_+^n\, /\, K(x)\geq c\}$, where $c$ is defined in the assumption $(A_2)$.\
Now, we are able to state our first main results.
\[t:11\] Assume that $n\geq 4$. Then, under the assumptions $(A_1)$, $(A_2)$, $(A_3)$, $(A_4)$, there exists a constant $c_0$ independent of $K$ such that if $K(y_0)/c\leq 1+c_0$, then (1) has a solution.
\[c:12\] The solution obtained in Theorem \[t:11\] has an augmented Morse\
index $\geq m$.
Next, we state another kind of existence results for problem (1) based on a “ topological invariant” for some Yamabe type problems introduced by Bahri see [@B1].\
To state these results, we need to introduce the assumptions that we will be using and some notations.\
${\bf (H_1)}$0.3cm We assume that $K_1$ has only nondegenerate critical points and we assume that there exists $y_0 \in \partial
S^n_+$ such that $y_0$ is the absolute maximum of $K_1$ and $(\partial K/\partial\nu )(y_0) > 0$.\
${\bf (H_2)}$0.3cm $W_s (y_i)\cap W_u(y_j) = \emptyset $ for any $i$ such that $(\partial K/\partial\nu )(y_i) > 0$ and for any $j$ such that $(\partial K/\partial\nu )(y_j) < 0$.\
For $ k \in \{1, 2, ..., n-1 \} $, we define $X$ as $$X=\overline{W_s(y_{i_0})}$$ where $y_{i_0}$ satisfies $$K_1(y_{i_0})= \max \{K_1(y_i), \, /\, \mbox{ind}(y_i)=n-1-k, \,
({\partial K}/{\partial \nu})(y_i)> 0\}$$ (Here ind$(y_i)$ denotes the Morse index of $y_i$ for the function $K_1$).\
${\bf (H_3)}$0.3cm We assume that $X$ is without boundary.\
We denote by $C_{y_0}(X)$ the following set $$C_{y_0}(X)=\{{\alpha }{\delta }_{y_0}+(1-{\alpha }){\delta }_x \, / \, {\alpha }\in [0,1],\, x\in X \}.$$ For ${\lambda }$ large enough, we introduce a map $f_{\lambda }: C_{y_0}(X)\to \Sigma ^+$, defined by $$C_{y_0}(X) \ni ({\alpha }{\delta }_{y_0}+(1-{\alpha }){\delta }_x) \longrightarrow
\frac{{\alpha }{\delta }_{(y_0,{\lambda })}
+(1-{\alpha }){\delta }_{(x,{\lambda })}}{|{\alpha }{\delta }_{(y_0,{\lambda })}+(1-{\alpha }){\delta }_{(x,{\lambda })}|}\in
\Sigma ^+.$$ Then $ C_{y_0}(X)$ and $f_{\lambda }( C_{y_0}(X))$ are manifolds in dimension $k+1$, that is, their singularities arise in dimension $k-1$ and lower, see [@B1]. Observe that $ C_{y_0}(X)$ and $f_{\lambda }( C_{y_0}(X))$ are contractible while $X$ is not contractible.\
For ${\lambda }$ large enough, we also define the intersection number(modulo 2) of $f_{\lambda }(C_{y_0}(X))$ with $ W_s(y_0,y_{i_0})_\infty$ $$\mu (y_0)=f_{\lambda }(C_{y_0}(X)). W_s(y_0,y_{i_0})_\infty$$ where $W_s(y_0,y_{i_0})_\infty$ is the stable manifold of $(y_0,y_{i_0})_\infty$ for a decreasing pseudogradient $V$ for $J$ which is transverse to $f_{\lambda }(C_{y_0}(X))$. Thus this number is well defined [@M].\
We then have the following result:
\[t:13\] Assume that $n\geq 4$. Under assumptions $(H_1)$, $(H_2)$ and $(H_3)$, if $\mu (y_0)=0$ then (1) has a solution of index $k$ or $k+1$.
Now, we state a statement more general than Theorem \[t:13\]. For this we define $X$ to be $$X=\overline{\cup_{y_i\in B_k} W_s(y_i)}, \quad \mbox{ with } B_k=\{y_i \, /
\, \mbox{ind}(y_i)=n-1-k \mbox{ and } ({\partial K}/{\partial \nu})(y_i)> 0\}.$$ For $y_i\in B_k$, we define, for ${\lambda }$ large enough, the intersection number(modulo 2) $$\mu_i (y_0)=f_{\lambda }(C_{y_0}(X)). W_s(y_0,y_i)_\infty.$$ By the above arguments, this number is well defined [@M].\
Then we have the following theorem
\[t:14\] Assume that $n\geq 4$. Under assumptions $(H_1)$, $(H_2)$ and $(H_3)$, if $\mu_i=0$ for each $y_i\in B_k$, then (1) has a solution of index $k$ or $k+1$.
The remainder of the present paper is organized as follows. In section 2, we set up the variational structure and recall some preliminaries. In section 3, we perform an expansion of the Euler functional associated to (1) and its gradient near the potential critical points at infinity, then we prove a Morse Lemma at infinity in section 4. In section 5, we provide the proof of Theorem \[t:11\] and Corollary \[c:12\], while section 6 is devoted to the proof of Theorems \[t:13\] and \[t:14\].
Variational Structure and Preliminaries
=========================================
In this section we recall the functional setting and the variational problem and its main features. Problem (1) has a variational structure. The functional is $$J(u)= \frac {\int_{S_+^n}|{\nabla }u|^2+\frac {n(n-2)}{4}\int_{S_+^n}u^2}
{\left(\int_{S_+^n}Ku^{\frac{2n}{n-2}}\right)^{\frac {n-2}{n}}}$$ defined on $H^1(\overline{S_+^n},{\mathbb{R}})$ equiped with the norm $$||u||^2=\int_{S_+^n}|{\nabla }u|^2+ \frac{n(n-2)}{4} \int_{S_+^n}u^2.$$ We denote by $\Sigma$ the unit sphere of $H^1(\overline{S_+^n},{\mathbb{R}})$ and we set $\Sigma ^+=\{u \in \Sigma \, / \, u\geq 0\}$.\
The Palais-Smale condition fails to be satisfied for $J$ on $\Sigma ^+$. Its failure has been studied by various authors (see Brezis-Coron [@BrC], Lions [@L], Struwe [@S]).\
In order to characterize the sequences failing the Palais-Smale condition, we need to introduce some notations.\
For $a\in \overline{S_+^n}$ and ${\lambda }>0$, let $$\tilde{\delta }_{a,{\lambda }}(x)= c_0\frac {{\lambda }^{\frac{n-2}{2}}}{({\lambda }^2+1+({\lambda }^2-1)
\cos d(a,x))^{\frac{n-2}{2}}}$$ where $d$ is the geodesic distance on $(S_+^n,g)$ and $c_0$ is chosen so that $$-{\Delta }\tilde{\delta }_{a,{\lambda }} + \frac{n(n-2)}{4}\tilde{\delta }_{a,{\lambda }}=
\tilde{\delta }_{a,{\lambda }}^{\frac{n+2}{n-2}}, \quad \mbox{ in } S_+^n.$$ For ${\varepsilon }>0$ and $p\in {\mathbb{N}}^*$, let us define $$\begin{aligned}
V(p,{\varepsilon })=& \{u\in \Sigma ^+/\exists \, a_1,...,a_p \in \overline{S_+^n}, \exists \,
{\lambda }_1,...,{\lambda }_p >0, \exists \, {\alpha }_1,...,{\alpha }_p>0 \\
& \mbox{ s.t. } ||u-\sum_{i=1}^p{\alpha }_i\tilde{\delta }_i||<{\varepsilon }\mbox{ and } |\frac
{{\alpha }_i ^{\frac{4}{n-2}}K(a_i)}{{\alpha }_j ^{\frac{4}{n-2}}K(a_j)}-1|<{\varepsilon },\\
& {\lambda }_i>{\varepsilon }^{-1}, {\varepsilon }_{ij}<{\varepsilon }\mbox{ and }{\lambda }_id_i<{\varepsilon }\mbox{ or }
{\lambda }_id_i >{\varepsilon }^{-1}\}\end{aligned}$$ where $\tilde{\delta }_i=\tilde{\delta }_{a_i,{\lambda }_i}$, $d_i=d(a_i,\partial S_+^n)$ and ${\varepsilon }_{ij}=({\lambda }_i/{\lambda }_j +{\lambda }_j/{\lambda }_i + {\lambda }_i{\lambda }_jd^2(a_i,a_j))^{\frac{2-n}{2}}$. The failure of Palais-Smale condition can be described as follows:
\[p:21\] (see [@BC2], [@L] and [@S]) Assume that $J$ has no critical point in $\Sigma ^+$ and let $(u_k)\in
\Sigma ^+$ be a sequence such that $J(u_k)$ is bounded and ${\nabla }J(u_k)\to 0$. Then, there exist an integer $p\in {\mathbb{N}}^*$, a sequence ${\varepsilon }_k>0$ (${\varepsilon }_k\to 0$) and an extracted subsequence of $u_k$, again denoted $(u_k)$, such that $u_k\in V(p,{\varepsilon }_k )$.
Now, we consider the following subset of $V(p,{\varepsilon })$ $$V_b(p,{\varepsilon })=\{u\in V(p,{\varepsilon })\, /\, {\lambda }_id_i<{\varepsilon }\}.$$ The following lemma defines a parametrization of the set $V_b(p,{\varepsilon })$.
\[l:22\] (see [@B2], [@BC2], [@R]) There is ${\varepsilon }_0>0$ such that if ${\varepsilon }<{\varepsilon }_0$ and $u\in V_b(p,{\varepsilon })$, then the problem $$\min\{||u-\sum_{i=1}^p{\alpha }_i\tilde{\delta }_i||,\, {\alpha }_i>0,\, {\lambda }_i>0,\, a_i\in
\partial S_+^n\}$$ has a unique solution (up to permutation). In particular, we can write $u\in V_b(p,{\varepsilon })$ as follows $$u=\sum_{i=1}^p\bar{{\alpha }}_i\tilde{\delta }_{\bar{a}_i,\bar{{\lambda }}_i}+v,$$ where $(\bar{{\alpha }}_1,...,\bar{{\alpha }}_p,\bar{a}_1,...,\bar{a}_p,\bar{{\lambda }}_1,...,
\bar{{\lambda }}_p)$ is the solution of the minimization problem and where $v\in H^1(S_+^n)$ such that for each $i=1,...,p$ $$(v,\tilde{\delta }_i)=(v,\partial \tilde{\delta }_i/\partial {\lambda }_i)=(v,
\partial \tilde{\delta }_i/\partial a _i)=0.$$ Here $(.,.)$ denoted the scalar inner defined on $H^1(\overline{S_+^n})$ by $$(u,v)=\int_{S_+^n}{\nabla }u {\nabla }v + \frac{n(n-2)}{4} \int_{S_+^n}uv.$$
Before ending this section, we mention that it will be convenient to perform some stereographic projection in order to reduce our problem to ${\mathbb{R}}^n_+$. Let $D^{1,2}({\mathbb{R}}^n_+)$ denote the completion of $C^\infty
_c (\bar{{\mathbb{R}}^n_+})$ with respect to Dirichlet norm. The stereographic projection $\pi _a$ throught a point $a \in \partial S^n_+$ induces an isometry $i : H^1(S^n_+) \to D^{1,2}({\mathbb{R}}^n_+) $ according to the following formula $$(iv)(x)= \left(\frac{2}{1+|x|^2}\right)^{(n-2)/2}v(\pi _a^{-1}(x)),
\qquad v\in H^1(S^n_+), \, x\in {\mathbb{R}}^n.$$ In particular, one can check that the following holds true, for every $v\in H^1(S^n_+)$ $$\int _{S^n_+}(|{\nabla }v|^2 + \frac{n(n-2)}{4}v^2) = \int _{{\mathbb{R}}^n_+}|\n
(iv)|^2 \qquad \mbox{and } \int _{S^n_+}|v|^{\frac{2n}{n-2}}= \int
_{{\mathbb{R}}^n_+}|iv|^{\frac{2n}{n-2}}.$$ In the sequel, we will identify the function $K$ and its composition with the stereographic projection $\pi _a$. We will also identify a point $b$ of $S^n_+$ and its image by $\pi _a$. These facts will be assumed as understood in the sequel.
Expansion of $J$ and its gradient at infinity
===============================================
This section is devoted to an useful expansion of $J$ and its gradient near a potential boundary critical point at infinity consisting of two masses.
\[p:31\] For ${\varepsilon }>0$ small enough and $u=\sum_{i=1}^2{\alpha }_i\tilde{\delta }_{a_i,{\lambda }_i}+v\in V_b(2,{\varepsilon })$, we have the following expansion $$\begin{aligned}
J(u)=& \frac{({\alpha }_1^2+{\alpha }_2^2)(S_n/2)^{(2/n)}}{({\alpha }_1^{\frac{2n}{n-2}}K(a_1)+
{\alpha }_2^{\frac{2n}{n-2}}K(a_2))^{\frac{n-2}{n}}}\left[1+\frac{4(n-2)}{n\b}c_1
\sum_{i=1}^2\frac{{\alpha }_i ^{\frac{2n}{n-2}}}{{\lambda }_i}\frac{\partial K}{\partial
\nu}(a_i)\right.\\
& -2c_2{\alpha }_1{\alpha }_2{\varepsilon }_{12}\left(-\frac{1}{\gamma}+\frac{1}{{\beta }}(\sum_{i=1}^2
{\alpha }_i ^{\frac{4}{n-2}}K(a_i))\right)+f(v)+Q(v,v)\\
& \left.+O\left({\varepsilon }_{12}^{\frac{n}{n-2}}log({\varepsilon }_{12}^{-1})+\sum(\frac{1}
{{\lambda }_i ^2}+\frac{{\varepsilon }_{12}}{{\lambda }_i}(log({\varepsilon }_{12}^{-1}))^{\frac{n-2}{n}})+
||v||^{\inf(3,\frac{2n}{n-2})}\right)\right]\end{aligned}$$ where $$\begin{aligned}
Q(v,v) = & \frac{1}{\gamma}||v||^2-\frac{n+2}{n-2}\frac{1}{\b} \int_{S_+^n}K({\alpha }_1
\tilde{\delta }_1+{\alpha }_2 \tilde{\delta }_2)^{\frac{4}{n-2}}v^2, \\
f(v) = & -\frac{2}{{\beta }}\int_{S_+^n}K({\alpha }_1\tilde{\delta }_1+{\alpha }_2
\tilde{\delta }_2)^{\frac{n+2}{n-2}}v,\qquad S_n = c_0^{\frac{2n}{n-2}}
\int_{{\mathbb{R}}^n}\frac{dx}{(1+|x|^2)^n},\\
{\beta }= &\frac{ S_n}{2}({\alpha }_1^{\frac{2n}{n-2}}K(a_1)+{\alpha }_2^{\frac{2n}{n-2}}
K(a_2)), \qquad \gamma = \frac{S_n}{2}({\alpha }_1^2+{\alpha }_2^2),\\
c_1 = & c_0^{\frac{2n}{n-2}}\int_{{\mathbb{R}}_+^n}\frac{x_ndx}{(1+|x|^2)^n},\qquad
\qquad \qquad c_2=c_0^{\frac{2n}{n-2}}\int_{{\mathbb{R}}^n}\frac{dx}{(1+
|x|^2)^{\frac{n+2}{2}}.}\end{aligned}$$
We need to estimate $$N(u)= ||u||^2 \mbox{ and } D^{\frac{n}{n-2}}=\int_{S_+^n} K(x)
u^{\frac{2n}{n-2}}.$$ In order to simplify the notations, in the remainder, we write $\tilde{\delta }_i$ instead of $\tilde{\delta }_{a_i, {\lambda }_i}$.\
We now have $$N(u)= {\alpha }_1^2||\tilde{\delta }_1||^2+{\alpha }_2^2||\tilde{\delta }_2||^2+||v||^2+2{\alpha }_1
{\alpha }_2(\int_{S_+^n}{\nabla }\tilde{\delta }_1{\nabla }\tilde{\delta }_2 +\frac{n(n-2)}{4}
\int_{S_+^n}\tilde{\delta }_1\tilde{\delta }_2)$$ Observe that $$\begin{aligned}
\label{e:o1}
||\tilde{\delta }||^2=\int_{{\mathbb{R}}_+^n}|{\nabla }{\delta }|^2 = \frac{S_n}{2}\end{aligned}$$ and $$\int_{S_+^n}{\nabla }\tilde{\delta }_1{\nabla }\tilde{\delta }_2 +\frac{n(n-2)}{4}
\int_{S_+^n}\tilde{\delta }_1\tilde{\delta }_2=
\int_{{\mathbb{R}}_+^n}{\nabla }{\delta }_1{\nabla }{\delta }_2 =\int_{{\mathbb{R}}_+^n}{\delta }_1^{\frac{n+2}{n-2}} {\delta }_2$$ where ${\delta }_i$ denotes $ {\delta }_{a_i,{\lambda }_i }$ and, for $a \in {\mathbb{R}}^n$ and ${\lambda }>
0$, $ {\delta }_{a,{\lambda }}$ denotes the family of solutions of Yamabe problem on $ {\mathbb{R}}^n$ defined by $${\delta }_{a,{\lambda }}(x)=c_0\frac{{\lambda }^{\frac{n-2}{2}}}{(1+{\lambda }^2|x-a|^2)^{\frac{n-2}{2}}}.$$ A computation similar to the one performed in [@B2]) shows that $$\begin{aligned}
\label{e:o2}
\int_{{\mathbb{R}}_+^n}{\delta }_1^{\frac{n+2}{n-2}} {\delta }_2= \frac{1}{2}c_2{\varepsilon }_{12} +
O({\varepsilon }_{12}^{\frac{n}{n-2}}log({\varepsilon }_{12}^{-1})).\end{aligned}$$ Thus $$N=\gamma + {\alpha }_1{\alpha }_2c_2{\varepsilon }_{12} + ||v||^2+
O({\varepsilon }_{12}^{\frac{n}{n-2}}log({\varepsilon }_{12}^{-1}))$$ For the denominator, we write $$\begin{aligned}
D^{\frac{n}{n-2}}&=&\int_{S_+^n}K ({\alpha }_1\tilde{\delta }_1+{\alpha }_2
\tilde{\delta }_2)^{\frac{2n}{n-2}}+
\frac{2n}{n-2} \int_{S_+^n}K ({\alpha }_1\tilde{\delta }_1+{\alpha }_2
\tilde{\delta }_2)^{\frac{n+2}{n-2}}v\\
& + & \frac{n(n+2)}{(n-2)^2}\int_{S_+^n}K({\alpha }_1\tilde{\delta }_1+{\alpha }_2
\tilde{\delta }_2)^{\frac{4}{n-2}}v^2\\
&+& O\left(\int_{S_+^n}({\alpha }_1\tilde{\delta }_1+{\alpha }_2
\tilde{\delta }_2)^{\frac{4}{n-2}-1}\inf(({\alpha }_1\tilde{\delta }_1+{\alpha }_2
\tilde{\delta }_2),|v|)^3+\int |v|^{\frac{2n}{n-2}}\right).\end{aligned}$$ We also write $$\begin{aligned}
\int_{S_+^n}K ({\alpha }_1\tilde{\delta }_1+{\alpha }_2\tilde{\delta }_2)^{\frac{2n}{n-2}} & =
\int_{S_+^n}K ({\alpha }_1\tilde{\delta }_1)^{\frac{2n}{n-2}}+
\int_{S_+^n}K ({\alpha }_2\tilde{\delta }_2)^{\frac{2n}{n-2}}+\frac{2n}{n-2}
\int_{S_+^n}K ({\alpha }_1\tilde{\delta }_1)^{\frac{n+2}{n-2}}{\alpha }_2\tilde{\delta }_2\\
& + \frac{2n}{n-2}\int_{S_+^n}K ({\alpha }_2\tilde{\delta }_2)^{\frac{n+2}{n-2}}
{\alpha }_1\tilde{\delta }_1 +O\left(\int_{S_+^n}\sup(\tilde{\delta }_1,
\tilde{\delta }_2)^{\frac{4}{n-2}} \inf(\tilde{\delta }_1,\tilde{\delta }_2)^2\right).\end{aligned}$$ Expansions of $K$ around $a_1$ and $a_2$ give $$\begin{aligned}
\label{e:o3}
\int_{S_+^n}K (\tilde{\delta }_i)^{\frac{2n}{n-2}}=K(a_i)\frac{S_n}{2}-
\frac{2c_1}{{\lambda }_i}\frac{\partial K}{\partial \nu}(a_i)+O(\frac{1}{\l
_i ^2})\end{aligned}$$ $$\begin{aligned}
\label{e:o4}
\int_{S_+^n}K (\tilde{\delta }_i)^{\frac{n+2}{n-2}}\tilde{\delta }_j=K(a_i)
\frac{c_2}{2}{\varepsilon }_{12}+O\left({\varepsilon }_{12}^{\frac{n}{n-2}}log({\varepsilon }_{12}^{-1})+
\frac{{\varepsilon }_{12}}{{\lambda }_i}(log({\varepsilon }_{12}^{-1}))^{\frac{n-2}{n}}\right)\end{aligned}$$ It easy to check $$\begin{aligned}
\label{e:o5}
\int_{S_+^n}\sup^{\frac{4}{n-2}}(\tilde{\delta }_1,\tilde{\delta }_2)\inf^2
(\tilde{\delta }_1,\tilde{\delta }_2)=
O({\varepsilon }_{12}^{\frac{n}{n-2}}log({\varepsilon }_{12}^{-1})) \, \mbox{ if } n\geq 4\end{aligned}$$ and $$\begin{aligned}
\label{e:o6}
\int_{S_+^n}({\alpha }_1\tilde{\delta }_1+{\alpha }_2
\tilde{\delta }_2)^{\frac{4}{n-2}-1}\inf(({\alpha }_1\tilde{\delta }_1+{\alpha }_2
\tilde{\delta }_2),|v|)^3+\int |v|^{\frac{2n}{n-2}}=O\left(||v||^{\inf(3,
\frac{2n}{n-2})}\right)\end{aligned}$$ Combining ,, , , and , we easily derive our proposition. $\Box$\
A natural improvement of Proposition \[p:31\] is obtained by taking care of the $v$-part, in order to show that it can be neglected with respect to the concentration phenomenon.\
Set $$E_{\varepsilon }=\{v\in H^1(S_+^n) \, /\, ||v||\leq {\varepsilon }\mbox{ and } v
\mbox{ satisfies } (V_0)\}$$ where $(V_0)$ is the following condition $$(V_0)\qquad (v,\tilde{\delta }_i)=(v,\partial \tilde{\delta }_i/\partial {\lambda }_i)=
(v,\partial \tilde{\delta }_i/\partial a_i)=0, \, \mbox{ for } i=1,2.$$ Notice that, one can prove arguing as in [@B2] (see also [@R]), that for ${\varepsilon }$ small enough, there exists $\rho >0$ such that for all $v\in E_{\varepsilon }$ $$Q(v,v) \geq \rho ||v||^2.$$ It follows the following lemma whose proof is similar , up to minor modifications to corresponding statements in [@B2] (see also [@R]).
\[l:32\] There exists a $C^1$-map which, to each $({\alpha }, a, {\lambda })$ such that $ {\alpha }_1\tilde{\delta }_1+{\alpha }_2\tilde{\delta }_2 \in V_b(2,{\varepsilon })$ with small ${\varepsilon }$, associates $\overline{v}=\overline{v}_{({\alpha },a,{\lambda })}$ satisfying $$J( {\alpha }_1\tilde{\delta }_1+{\alpha }_2\tilde{\delta }_2 +\overline{v})= \min\{
J( {\alpha }_1\tilde{\delta }_1+{\alpha }_2\tilde{\delta }_2 +v) , \, v \mbox{ satisfies } (V_0)\}.$$ Moreover, there exists $c>0 $ such that the following holds $$||\overline{v}||\leq c \left(\frac{1}{{\lambda }_1}+\frac{1}{{\lambda }_2}+{\varepsilon }_{12}^{\frac
{n+2}{2(n-2)}}log({\varepsilon }_{12}^{-1})+(\mbox{ if } n\leq 5) {\varepsilon }_{12}
(log({\varepsilon }_{12}^{-1}))^{\frac{n-2}{n}}\right).$$
\[p:33\] Let $n\geq 4$, for $ u={\alpha }_1\tilde{\delta }_1+{\alpha }_2\tilde{\delta }_2 \in V_b(2,{\varepsilon })$, we have the following expansion $$\begin{aligned}
({\nabla }J(u),{\lambda }_1\partial\tilde{\delta }_1/\partial{\lambda }_1)= & 2J(u)\left[\frac{c_2}
{2}{\alpha }_2 {\lambda }_1\frac{\partial {\varepsilon }_{12}}{\partial {\lambda }_1}(1-J(u)^\frac{n}{n-2}
({\alpha }_1^\frac{4}{n-2}K(a_1)+{\alpha }_2^\frac{4}{n-2}K(a_2)))\right.\\
& \left.-2J(u)^\frac{n}{n-2}{\alpha }_1^\frac{n+2}{n-2}\frac{c_3}{{\lambda }_1}\frac{\partial
K}{\partial \nu}(a_1)\right]+O(\frac{1}{{\lambda }_1^2})\\
& +O\left({\varepsilon }_{12}^{\frac{n}{n-2}}log({\varepsilon }_{12}^{-1})+{\varepsilon }_{12}(log(
{\varepsilon }_{12}^{-1}))^{\frac{n-2}{n}}(\frac{1}{{\lambda }_1}+\frac{1}{{\lambda }_2})\right).\end{aligned}$$ where $c_3=\frac{n-2}{2}c_0^{2n/(n-2)}\int_{{\mathbb{R}}_+^n}\frac{x_n(|x|^2-1)}
{(1+|x|^2)^{n+1}}dx $
We have $$\begin{aligned}
\label{e:31}
({\nabla }J(u),h)= 2J(u)\left[\int_{S_+^n}{\nabla }u{\nabla }h+\frac{n(n-2)}{4}
\int_{S_+^n}uh-J(u)^\frac{n}{n-2}\int_{S_+^n}Ku^\frac{n+2}{n-2}h\right]\end{aligned}$$ Observe that(see [@B2]) $$\begin{aligned}
\label{e:32}
\int _{{\mathbb{R}}^n_+}{\nabla }{\delta }_1 {\nabla }({\lambda }_1 \frac{\partial {\delta }_1}{\partial \l
_1})
= \int _{{\mathbb{R}}^n_+}{\delta }_1^{\frac{n+2}{n-2}}{\lambda }_1\frac{\partial {\delta }_1}{\partial \l
_1} = 0\end{aligned}$$ $$\begin{aligned}
\label{e:33}
\int _{{\mathbb{R}}^n_+}{\nabla }{\delta }_2 {\nabla }({\lambda }_1 \frac{\partial {\delta }_1}{\partial \l
_1})
&= \int _{{\mathbb{R}}^n_+}{\delta }_2^{\frac{n+2}{n-2}}{\lambda }_1\frac{\partial {\delta }_1}{\partial \l
_1}\nonumber \\
&=\frac{1}{2} c_2 {\lambda }_1 \frac{\partial {\varepsilon }_{12}}{\partial {\lambda }_1} +
O\left( {\varepsilon }_{12}^{\frac{n}{n-2}}log({\varepsilon }_{12}^{-1})\right) \end{aligned}$$ $$\begin{aligned}
\label{e:34}
\int _{{\mathbb{R}}^n_+}K {\delta }_1 ^{\frac{n+2}{n-2}} {\lambda }_1 \frac{\partial {\delta }_1}{\partial {\lambda }_1}
&= 2 {\nabla }K(a_1)\int _{{\mathbb{R}}^n_+}{\delta }_1 ^{\frac{n+2}{n-2}}{\lambda }_1 \frac{\partial
{\delta }_1}{\partial {\lambda }_1} (x-a_1) +O\left(\frac{1}{{\lambda }_1
^2}\right)\nonumber\\
&=- \frac{2c_3}{{\lambda }_1}{\nabla }K(a) e_n + O\left(\frac{1}{{\lambda }_1 ^2}\right),\end{aligned}$$ $$\begin{aligned}
\label{e:35}
\int _{{\mathbb{R}}^n_+}K{\delta }_2 ^{\frac{n+2}{n-2}}{\lambda }_1 \frac{\partial {\delta }_1}{\partial {\lambda }_1}& = K(a_2) \frac{1}{2}c_2 {\lambda }_1 \frac{\partial
{\varepsilon }_{12}}{\partial {\lambda }_1} +
O\left(\frac{1}{{\lambda }_2}{\varepsilon }_{12} (Log({\varepsilon }_{12}
^{-1}))^{\frac{n-2}{n}}\right)\nonumber\\
& + O\left( {\varepsilon }_{12}^{\frac{n}{n-2}}log({\varepsilon }_{12}^{-1})\right),\end{aligned}$$ $$\begin{aligned}
\label{e:36}
\frac{n+2}{n-2}\int _{{\mathbb{R}}^n_+}K{\delta }_2{\delta }_1 ^{\frac{4}{n-2}}{\lambda }_1 \frac{\partial {\delta }_1}{\partial {\lambda }_1}& = K(a_1) \frac{1}{2}c_2 {\lambda }_1 \frac{\partial
{\varepsilon }_{12}}{\partial {\lambda }_1} +
O\left(\frac{1}{{\lambda }_1}{\varepsilon }_{12} (Log({\varepsilon }_{12}
^{-1}))^{\frac{n-2}{n}}\right)\nonumber\\
& + O\left({\varepsilon }_{12}^{\frac{n}{n-2}}log({\varepsilon }_{12}^{-1})\right).\end{aligned}$$ Combining , , , , and , we easily derive our proposition. $\Box$\
\[p:34\] Let $n\geq 4$. For $ u = \sum {\alpha }_i \tilde{\delta }_i \in V_b (2, {\varepsilon })$, we have the following expansion: $$\begin{aligned}
\left({\nabla }J(u), \frac{1}{{\lambda }_1}\frac{\partial\tilde{\delta }_1}{\partial a_1}\right)
&= 2J(u){\alpha }_1 e_n \left[ c_4\left(1-
J(u)^{\frac{n}{n-2}}{\alpha }_1 ^{\frac{4}{n-2}}K(a_1)\right)+
J(u)^{\frac{n}{n-2}}{\alpha }_1 ^{\frac{4}{n-2}}\frac{c_5}{\l
_1}\frac{\partial K}{\partial\nu}(a_1)\right]\\
&-J(u){\alpha }_2 c_2 \frac{1}{\l
_1}\frac{\partial{\varepsilon }_{12}}{\partial a_1}\left(-1 +
J(u)^{\frac{n}{n-2}}\sum {\alpha }_i ^{\frac{4}{n-2}}K(a_i)\right)\\
&-4J(u)^{\frac{2(n-1)}{n-2}}{\alpha }_1 ^{\frac{n+2}{n-2}}\frac{2c_5}{{\lambda }_1}\n
_T K(a_1) + O\left({\varepsilon }_{12}^{\frac{n}{n-2}}Log({\varepsilon }_{12}^{-1})+ \e
_{12}^{\frac{n+1}{n-2}}{\lambda }_2 |a_1-a_2|\right)\\
&+O\left({\varepsilon }_{12}\left(Log({\varepsilon }_{12}^{-1})\right)^{\frac{n-2}{n}}\sum
\frac{1}{{\lambda }_k}\right)+ O\left(\frac{1}{{\lambda }_1^2}\right)\end{aligned}$$ where $$c_4 = (n-2)c_0^{\frac{2n}{n-2}}\int
_{{\mathbb{R}}^n_+}\frac{x_n}{(1+|x|^2)^{n+1}}dx \quad \mbox{and }
c_5=\frac{n-2}{2n}c_0^{\frac{2n}{n-2}}\int
_{{\mathbb{R}}^n}\frac{x_n^2}{(1+|x|^2)^{n+1}}dx$$
An easy computation shows $$\begin{aligned}
\label{e:37}
\int _{{\mathbb{R}}^n_+}\n{\delta }_1\n\left(\frac{1}{{\lambda }_1}\frac{\partial{\delta }_1}{\partial a_1}\right)=\int _{{\mathbb{R}}^n_+}{\delta }_1^{\frac{n+2}{n-2}}\frac{1}{{\lambda }_1}\frac{\partial{\delta }_1}{\partial
a_1} = c_4e_n,\end{aligned}$$ $$\begin{aligned}
\label{e:38}
\int _{{\mathbb{R}}^n_+}\n{\delta }_2\n\left(\frac{1}{{\lambda }_1}\frac{\partial{\delta }_1}{\partial a_1}\right)
&=\int _{{\mathbb{R}}^n_+}{\delta }_2^{\frac{n+2}{n-2}}\frac{1}{{\lambda }_1}\frac{\partial{\delta }_1}{\partial
a_1}\nonumber\\
&= \frac{1}{2}\frac{c_2}{{\lambda }_1}\frac{\partial{\varepsilon }_{12}}{\partial a_1} + O\left(\e
_{12}^{\frac{n}{n-2}}Log({\varepsilon }_{12}^{-1}) +
{\varepsilon }_{12}^{\frac{n+1}{n-2}}{\lambda }_2 |a_1-a_2|\right)\end{aligned}$$ $$\begin{aligned}
\label{e:39}
\int _{{\mathbb{R}}^n_+}K{\delta }_1^{\frac{n+2}{n-2}}\frac{1}{{\lambda }_1}\frac{\partial{\delta }_1}{\partial
a_1} =K(a_1) c_4e_n + 2\frac{c_5}{{\lambda }_1}{\nabla }K(a_1)
+O\left(\frac{1}{{\lambda }_1^2}\right),\end{aligned}$$ $$\begin{aligned}
\label{e:310}
\int _{{\mathbb{R}}^n_+}K{\delta }_2^{\frac{n+2}{n-2}}\frac{1}{{\lambda }_1}\frac{\partial{\delta }_1}{\partial
a_1} &=K(a_2)\frac{1}{2}c_2\frac{1}{{\lambda }_1}\frac{\partial\e
_{12}}{\partial a_1}+ O\left({\varepsilon }_{12}^{\frac{n+1}{n-2}}\l
_2|a_1-a_2|\right)\nonumber\\
&+O\left({\varepsilon }_{12}^{\frac{n}{n-2}}Log(\e
_{12}^{-1})\right)+O\left(\frac{1}{{\lambda }_2}{\varepsilon }_{12}\left(Log(\e
_{12}^{-1})\right)^{\frac{n-2}{n}}\right),\end{aligned}$$ $$\begin{aligned}
\label{e:311}
\frac{n+2}{n-2}\int _{{\mathbb{R}}^n_+}K{\delta }_1^{\frac{4}{n-2}}{\delta }_2 \frac{1}{{\lambda }_1}\frac{\partial{\delta }_1}{\partial
a_1} &=K(a_1)\frac{1}{2}c_2\frac{1}{{\lambda }_1}\frac{\partial\e
_{12}}{\partial a_1}+ O\left({\varepsilon }_{12}^{\frac{n+1}{n-2}}\l
_2|a_1-a_2|\right),\nonumber\\
&+O\left({\varepsilon }_{12}^{\frac{n}{n-2}}Log(\e
_{12}^{-1})\right)+O\left(\frac{1}{{\lambda }_1}{\varepsilon }_{12}\left(Log(\e
_{12}^{-1})\right)^{\frac{n-2}{n}}\right).\end{aligned}$$ Using , , , , and , our proposition follows. $\Box$\
\[p:35\] For $ u = \sum {\alpha }_i \tilde{\delta }_i \in V_b (2, {\varepsilon })$, we have the following expansion: $$\begin{aligned}
\left({\nabla }J(u),\tilde{\delta }_1\right)&=2J(u) {\alpha }_1
\frac{S_n}{2}\left(1-J(u)^{\frac{n}{n-2}}{\alpha }_1^{\frac{4}{n-2}}K(a_1)\right)
+2J(u)^{\frac{2(n-1)}{n-2}}{\alpha }_1^{\frac{n+2}{n-2}}\frac{2c_6}{{\lambda }_1}
\frac{\partial K}{\partial \nu}(a_1)\\
&-J(u)c_2{\varepsilon }_{12}{\alpha }_2 \left(-1 +
J(u)^{\frac{n}{n-2}}\left(\frac{n+2}{n-2}\a
_1^{\frac{4}{n-2}}K(a_1)+{\alpha }_2^{\frac{4}{n-2}}K(a_2)\right)\right)\\
&+O\left(\frac{1}{{\lambda }_1^2}+ {\varepsilon }_{12}^{\frac{n}{n-2}}Log({\varepsilon }_{12}^{-1})
+ \frac{1}{{\lambda }_2}{\varepsilon }_{12}\left(Log({\varepsilon }_{12}^{-1})\right)^{\frac{n-2}{n}}\right)\end{aligned}$$ where $c_6= c_0 \int _{{\mathbb{R}}^n_+}\frac{x_n}{(1+|x|^2)^n}dx$ and where $S_n$ is defined in Proposition \[p:31\].
Using estimates , , and , we easily derive our proposition. $\Box$\
Before ending this section, we state the above results in the case where we have only one mass instead of two masses.
\[p:36\] For ${\varepsilon }> 0$ small enough and $ u=\a\tilde{\delta }_{a,{\lambda }} + v \in V_b(1,\e
) $, we have the following expansion: $$\begin{aligned}
J(u)=&
\frac{(S_n/2)^{(2/n)}}{(K(a))^{\frac{n-2}{n}}}\left[1+\frac{8c_1(n-2)}
{nK(a)S_n \l}
\frac{\partial K}{\partial
\nu}(a) +f(v)\right.\\
&\left. +Q(v,v) +O\left(\frac{1}{{\lambda }^2}\right)+ O\left(
||v||^{\inf(3,\frac{2n}{n-2})}\right)\right]\end{aligned}$$ where $$\begin{aligned}
Q(v,v) = & \frac{2}{{\alpha }^2 S_n}\left[||v||^2-\frac{n+2}{(n-2) K(a)} \int_{S_+^n}K(\tilde{\delta }_{a,{\lambda }})^{\frac{4}{n-2}}v^2\right] \\
f(v) = & -\frac{4}{{\alpha }K(a)S_n }\int_{S_+^n}K(\tilde{\delta }_{a,{\lambda }})^{\frac{n+2}{n-2}}v.\end{aligned}$$
\[p:37\] For $u={\alpha }\tilde{\delta }_{a,\l} \in V_b(1,\e)$, we have the following expansion: $$\begin{aligned}
\left({\nabla }J(u), \l
\frac{\partial\tilde{\delta }}{\partial\l}\right)=-4J(u)^{\frac{2(n-1)}{n-2}}\a
^{\frac{n+2}{n-2}}\frac{c_3}{\l}\frac{\partial K}{\partial\nu}(a) +
O\left(\frac{1}{{\lambda }^2}\right)\end{aligned}$$
\[p:38\] For $u=u={\alpha }\tilde{\delta }_{a,\l} \in V_b(1,\e)$, we have the following expansion: $$\begin{aligned}
\left({\nabla }J(u),\frac{1}{ \l}
\frac{\partial\tilde{\delta }}{\partial a}\right)&=2J(u)\a
e_n\left[c_4\left(1-J(u)^{\frac{n}{n-2}}\a
^{\frac{4}{n-2}}K(a)\right)+2c_5J(u)^{\frac{n}{n-2}}\frac{\a
^{\frac{4}{n-2}}}{\l}\frac{\partial K}{\partial\nu}(a)\right]\\
&-4J(u)^{\frac{2(n-1)}{n-2}}{\alpha }^{\frac{n+2}{n-2}}c_5 \frac{{\nabla }_T K(a)}{\l}
+ O\left(\frac{1}{{\lambda }^2}\right).\end{aligned}$$
Morse Lemma at Infinity
=======================
In this section, we consider the case where we only have one mass and we perform a Morse lemma at infinity for J, which completely gets rid of the v-contribution and shows that the functional behaves, at infinity, as $J({\alpha }\tilde{\delta }_{\tilde{a},\tilde{\l}}) + |V|^2$, where $V$ is a variable completely independent of $\tilde{a}$, $\tilde{\l}$. Namely, we prove the following proposition.
\[p:41\] For ${\varepsilon }> 0$ small enough, there is a diffeomorphism $${\alpha }\tilde{\delta }_{a,\l} + v \longmapsto {\alpha }\tilde{\delta }_{\tilde{a},\tilde\l
}\in V_b(1,{\varepsilon }' )$$ for some ${\varepsilon }'$ such that $$J({\alpha }\tilde{\delta }_{a,\l}+v)= J({\alpha }\tilde{\delta }_{\tilde{a}, \tilde\l})+ |V|^2$$ where $V$ belongs to a neighborhood of zero in a fixed Hilbert space.
The above Morse Lemma can be improved when the concentration point is near a critical point $y$ of $K_1=K_{/\partial S^n_+}$ with $\frac{\partial K}{\partial \nu}(y) > 0 $ , leading to the following normal form:
\[p:42\] For $u={\alpha }\tilde{\delta }_{\tilde{a},\tilde\l}\in V_b(1,{\varepsilon })$ such that $\tilde{a} \in \mathcal{V}(y,\rho )$, $ \frac{\partial
K}{\partial\nu}(y) > 0$, $\rho > 0$ and $y$ is a critical point of $K_1$, there is another change of variable $(\tilde{\tilde{a}},
\tilde{\tilde\l} )$ such that $$\begin{aligned}
J(u)&=\psi (\tilde{\tilde{a}}, \tilde{\tilde{\l}})\\
&:= \frac{(S_n/2)^{\frac{2}{n}}}{(K(\tilde{\tilde{a}}))^{\frac{n-2}{n}}}\left[1+(c-\eta
)\frac{1}{\tilde{{\tilde{\l}}}} \frac{\partial K}{\partial\nu} (\tilde{\tilde{a}}) \, \right]\end{aligned}$$ where $\eta$ is a small positive real.\
Here and in the sequel, $\mathcal{V}(y,\rho )$ denotes a neighborhood of $y$.
The proof of Propositions \[p:41\] and \[p:42\] can be easily deduced from the following lemma, arguing as in [@B1] and [@BCCH] .
\[l:43\] There exists a pseudogradient $Z$ so that the following holds.\
There is a constant $c>0$ independent of $u=\a\tilde{\delta }_{a,{\lambda }}$ in $V_b(1,{\varepsilon })$ such that\
[**i.**]{}$-\left({\nabla }J(u), Z\right) \geq \frac{c}{\l}$\
[**ii.**]{}$-\left({\nabla }J(u+\bar{v}), Z +\frac{\partial
\bar{v}}{\partial ({\alpha }, a, {\lambda })}(Z)\right) \geq \frac{c}{\l}$\
[**iii.**]{} $ Z$ is bounded\
[**iv.**]{}the only region where $\l$ increases along $Z$ is the region where $a\in \mathcal{V}(y,\rho )$, where $y$ is a critical point of $K_1$ such that $\frac{\partial K}{\partial\nu}(y) > 0. $
Before giving the proof of Lemma \[l:43\], we notice that combining Proposition \[p:42\] and Lemma \[l:43\], one can easily derive the following corollary.
\[c:44\] Assume that $J$ does not have any critical point. Then, the only critical points at infinity of $J$ in $V_b(1,{\varepsilon })$, for ${\varepsilon }$ small enough, correspond to $\tilde{\delta }_{y,\infty}$, where $y$ is a critical point of $K_1 = K_{/\partial S^n_+}$ such that $\frac{\partial
K}{\partial\nu}(y) > 0 $.\
Moreover such a critical point at infinity has a Morse index equal to $(n-1-index(K_1,y))$.
[**Lemma \[l:43\]**]{} Let $u={\alpha }\tilde{\delta }_{a,\l} \in V_b(1,{\varepsilon })$. We divide $V_b(1,{\varepsilon })$ in three regions.\
[**$1^{st}$ region**]{}. $a\notin \cup _{0\leq i\leq s} \mathcal{V}(y_i,
\rho )$, where $\rho < \frac{1}{2} \min _{i\not= j}d(y_i,y_j)$.\
Set $Z_1=\frac{1}{{\lambda }}\frac{\partial\tilde{\delta }}{\partial a}{\nabla }_T K(a)$, from Proposition \[p:38\], we have $$\begin{aligned}
-({\nabla }J(u), Z_1)&=c\frac{|{\nabla }_TK(a)|^2}{\l} + O\left(\frac{1}{{\lambda }^2}\right)\\
&\geq \frac{c}{\l}\end{aligned}$$ [**$2^{nd}$ region**]{}. $a\in \cup _{0\leq i\leq l} \mathcal{V}(y_i,
2\rho )$\
We set $$Z_2 = {\lambda }\frac{\partial\tilde{\delta }}{\partial\l} +
\frac{1}{\l}\frac{\partial\tilde{\delta }}{\partial a} {\nabla }_TK(a).$$ Using Propositions \[p:37\] and \[p:38\], we obtain\
$$\begin{aligned}
-({\nabla }J(u),Z_2) &\geq c \frac{|\n
_TK(a)|^2}{\l}+\frac{c}{\l}\frac{\partial K}{\partial\nu}(a)+
O\left(\frac{1}{{\lambda }^2}\right)\\
&\geq \frac{c}{\l}.\end{aligned}$$ [**$3^{rd}$ region**]{}. $a\in \cup _{l+1\leq i\leq s} \mathcal{V}(y_i,
2\rho ).$\
We set $$Z_3 = -{\lambda }\frac{\partial\tilde{\delta }}{\partial\l}.$$ Using Proposition \[p:37\], we deduce that $$\begin{aligned}
-({\nabla }J(u), Z_3) &\geq -\frac{c}{\l}\frac{\partial K}{\partial\nu}(a) +
O\left(\frac{1}{{\lambda }^2}\right)\\
&\geq \frac{c}{\l}.\end{aligned}$$ Hence our global vector field will be built using a convex combination of $Z_1$, $Z_2$ and $Z_3$ and will satisfy obviously i., iii. and iv. Regarding the estimate ii., it can be obtained once we have i. arguing as in [@B1] and [@BCCH].
Proof of Theorem \[t:11\]
=========================
Our proof follows the algebraic topological arguments introduced in [@AB]. Arguing by contradiction, we suppose that $J$ has no critical points. It follows from Corollary \[c:44\], that under the assumptions of Theorem \[t:11\], the critical points at infinity of $J$ under the level $c_1= (S_n/2)^{\frac{2}{n}}(K(y_l))^{\frac{2-n}{n}} + {\varepsilon }$ , for $\e$ small enough, are in one to one correspondance with the critical points of $K_1$ $y_0$, $y_1$, ..., $y_l$. The unstable manifold at infinity of such critical points at infinity, $W_u(y_0)_\infty$, ..., $W_u(y_l)_\infty$ can be described, using Proposition \[p:42\], as the product of $W_s(y_0)$, ..., $W_s(y_l)$ (for a pseudogradient of $K$ ) by $[A, +\infty [$ domaine of the variable $\l$, for some positive number $A$ large enough.\
Since $J$ has no critical points, it follows that $ J_{c_1}=\{u\in
\sum ^+ / J(u) \leq c_1 \}$ retracts by deformation on $X_\infty =
\cup _{0\leq j\leq l}W_u(y_j)_\infty$ (see Sections 7 and 8 of [@BR]) which can be parametrized as we said before by $X \times
[A, +\infty[$.\
From another part, we have $X_\infty$ is contractible in $J_{c_2+\e}$, where $c_2=(S_n/2)^{\frac{2}{n}}c^{\frac{2-n}{n}}$. Indeed from $(A_4)$, it follows that there exists a contraction $ h :[0,1] \times
X \to K^c$, $h$ continuous such that for any $a\in X \quad h(0,a)=a$ and $h(1,a)=a_0$ a point of $X$. Such a contraction gives rise to the following contraction $\tilde{h} : X_\infty \to \sum ^+$ defined by $$[0,1] \times X \times \left[0,\right.+\infty\left[ \right.\ni (t,a_1,{\lambda }_1 ) \longmapsto \tilde{\delta }_{(h(t,a_1),{\lambda })}
+ \bar{v} \in \Sigma ^+, \quad a_1 \in X, \, \, {\lambda }_1 \geq A$$ For $t=0$, $\tilde{\delta }_{(h(0,a_1),{\lambda }_1)}+\bar{v} = \tilde{\delta }_{a_1, {\lambda }_1}
+\bar{v} \in X_\infty$. $\tilde{h}$ is continuous and $\tilde{h}(1,a_1,{\lambda }_1)= \tilde{\delta }_{a_0,{\lambda }_1} +\bar{v}$, hence our claim follows.\
Now, using Proposition \[p:36\], we deduce that $$J(\tilde{\delta }_{h(t,a_1), {\lambda }_1} + \bar{v}) \sim
(\frac{S_n}{2})^{\frac{2}{n}}(K(h(t,a_1)))^{\frac{2-n}{n}}\left(1+O(A
^{-2})\right)$$ where $K(h(t,a_1)) \geq c $ by construction.\
Therefore such a contraction is performed under $c_2 +\e$, for $A$ large enough, so $X_\infty$ is contractible in $J_{c_2+{\varepsilon }}$.\
In addition, choosing $c_0$ small enough, $J_{c_2+{\varepsilon }}$ retracts by deformation on $J_{c_1}$, which retracts by deformation on $X_\infty$, therfore $X_\infty$ is contractible leading to the contractibility of $X$, which is in contradiction with our assumption. Hence our theorem follows.\
\
Before ending this section, we give the proof of Corollary \[c:12\].\
[**Corollary \[c:12\]**]{} Arguing by contradiction, we may assume that the Morse index of the solution provided by Theorem \[t:11\] is $\leq m-1$.\
Perturbing, if necessary $J$, we may assume that all the critical points of $J$ are nondegenerate and have their Morse index $\leq
m-1$. Such critical points do not change the homological group in dimension $m$ of level sets of $J$.\
Since $X_\infty$ defines a homological class in dimension $m$ which is nontrivial in $J_{c_1}$, but trivial in $J_{c_2+\e}$, our result follows.
Proof of Theorems \[t:13\] and \[t:14\]
=======================================
First, we start by proving the following main results
\[p:61\] Let $y_0$ be defined in $(H_1)$. Then $(y_0,y_0)_\infty$ is not a critical point at infinity for $J$, that is, there exists a decreasing pseudogradient $W$ for $J$ satisfying Palais-Smaile in the neighborhood of $(y_0,y_0)_\infty$.
For ${\varepsilon }_0 > 0$ small enough, we set $$C_{{\varepsilon }_0}=\{u={\alpha }_1 \tilde{\delta }_{a_1, {\lambda }_1} + {\alpha }_2 \tilde{\delta }_{a_2,\l
_2} \in V_b (2,{\varepsilon }_0 )/ a_1, a_2 \in \mathcal{V}(y_0)\cap \partial
S^n_+ \}.$$ Our goal is to build a pseudogradient vector field $W$ for $J$ satisfying the Palais-Smale condition in $C_{{\varepsilon }_0}$ such that for $u\in C_{{\varepsilon }_0}$, we have\
[**i.**]{}$-\left({\nabla }J(u), W\right) \geq \gamma \left(\sum\frac{1}{\l
_i^2}+ \sum (1-J(u)^{\frac{n}{n-2}}a
_i^{\frac{4}{n-2}}K(a_i)\right) + c{\varepsilon }_{12}^{\frac{n-\frac{1}{2}}{n-2}},$\
[**ii.**]{}$-\left({\nabla }J(u+\bar{v}), W +\frac{\partial
\bar{v}}{\partial ({\alpha }, a, {\lambda })}(W)\right) \geq
\frac{\gamma}{2} \left(\sum\frac{1}{\l
_i^2}+ \sum (1-J(u)^{\frac{n}{n-2}}a
_i^{\frac{4}{n-2}}K(a_i)\right) + c{\varepsilon }_{12}^{\frac{n-\frac{1}{2}}{n-2}}.$\
[**iii.**]{} $ W$ is bounded\
[**iv.**]{} $\dot{\l}_{max} \leq 0$.\
where $\gamma$ is a positive constant large enough.\
We can assume, without loss of generality, that ${\lambda }_1 \leq {\lambda }_2$. We devide $C_{{\varepsilon }_0}$ in three principal regions.\
[**$1^{st}$ region.**]{}$M{\lambda }_1 \leq {\lambda }_2$ and $\forall i |1- J(u)^{\frac{n}{n-2}}a
_i^{\frac{4}{n-2}}K(a_i)| \leq \frac{2C'}{{\lambda }_i}$, where $M$ and $C'$ are postive constants large enough.\
We set $$W_1 = {\lambda }_1 \frac{\partial\tilde{\delta }_1}{\partial{\lambda }_1}- {\lambda }_2
\frac{\partial\tilde{\delta }_2}{\partial{\lambda }_2}\sqrt{M}.$$ From Proposition \[p:33\], we derive $$\begin{aligned}
-\left({\nabla }J(u), W_1\right)&\geq c\left(\frac{1}{{\lambda }_1} + O(\e
_{12})\right) + \sqrt{M}\left[-\frac{c}{{\lambda }_2} + c {\varepsilon }_{12}\right]
+O\left(\sum \frac{1}{{\lambda }_k^2}\right)\\
&+O\left({\varepsilon }_{12}^{\frac{n}{n-2}}Log({\varepsilon }_{12}^{-1}) + {\varepsilon }_{12}\left(Log(\e
_{12}^{-1})\right)^{\frac{n-2}{n}}\sum \frac{1}{{\lambda }_k}\right)\\
&\geq \frac{c}{2}\left(\frac{1}{{\lambda }_1}+ \frac{M}{{\lambda }_2}\right) +
\frac{c\sqrt{M}}{2}{\varepsilon }_{12}\\
&\geq C\left(\sum\frac{1}{\l
_i^2}+ \sum (1-J(u)^{\frac{n}{n-2}}a
_i^{\frac{4}{n-2}}K(a_i) + {\varepsilon }_{12}\right).\end{aligned}$$ [**$2^{nd}$ region.**]{}$2M{\lambda }_1 \geq {\lambda }_2$ and, $\forall i |1- J(u)^{\frac{n}{n-2}}a
_i^{\frac{4}{n-2}}K(a_i)| \leq \frac{2C'}{{\lambda }_i}$\
In this region, two cases may occur.\
[**Case 2.1**]{}${\varepsilon }_{12}^{\frac{n-\frac{1}{2}}{n-2}} \geq
\frac{d_0}{{\lambda }_1}$, where $d_0=\max (d(y_0,a_1), d(y_0,a_2),
d(a_1,a_2)).$\
We set $$W_2=\frac{1}{{\lambda }_1}\left[{\alpha }_1 \frac{\partial\tilde{\delta }_1}{\partial
a_1}\left(\frac{a_2-a_1}{d(a_2,a_1)}\right) -{\alpha }_2
\frac{\partial\tilde{\delta }_2}{\partial
a_2}\left(\frac{a_2-a_1}{d(a_2,a_1)}\right)\right].$$ From Proposition \[p:34\] and the fact that $\frac{\partial \e
_{12}}{\partial a_2} = -\frac{\partial \e
_{12}}{\partial a_1} $, we obtain $$\begin{aligned}
-\left({\nabla }J(u), W_2\right)&=\frac{1}{{\lambda }_1}\left({\alpha }_1{\alpha }_2 c_2
2J(u) \frac{\partial{\varepsilon }_{12}}{\partial
a_1}\right)\left(-1 +J(u)^{\frac{n}{n-2}}\sum \a
_k^{\frac{4}{n-2}}K(a_k)\right)2\left(\frac{a_2-a_1}{d(a_2,a_1)}\right)\\
&+O\left(\sum \left(\frac{|{\nabla }_TK(a_k)|}{{\lambda }_k} + \frac{1}{{\lambda }_k^2}\right)\right)\\
&+O\left({\varepsilon }_{12}^{\frac{n}{n-2}}Log({\varepsilon }_{12}^{-1}) + \sum \frac{1}{\l
_k} \e
_{12}\left(Log({\varepsilon }_{12}^{-1})\right)^{\frac{n-2}{n}}+
d(a_1,a_2){\varepsilon }_{12}^{\frac{n+1}{n-2}} \sum {\lambda }_k \right).\end{aligned}$$ Observe that $$\begin{aligned}
{\varepsilon }_{12} &\sim \left({\lambda }_1{\lambda }_2 d(a_1,a_2)^2 \right)^{\frac{2-n}{2}}
\quad \mbox.{Indeed } {\lambda }_1 \mbox{and } {\lambda }_2 \mbox{ are the same order}, and\\
\frac{\partial {\varepsilon }_{12}}{\partial a_1}&= -(n-2){\lambda }_1{\lambda }_2 (a_1-a_2)\e
_{12}^{\frac{n}{n-2}}.\end{aligned}$$ Thus $$\begin{aligned}
\label{e:41}
-\left({\nabla }J(u), W_2\right)&=\frac{4}{{\lambda }_1}{\alpha }_1{\alpha }_2 c_2 (n-2)
J(u){\lambda }_1{\lambda }_2 d(a_1,a_2)\e
_{12}^{\frac{n}{n-2}}(1+o(1)) + R\nonumber\\
&\geq \frac{c\e_{12}}{{\lambda }_1 d(a_1,a_2)} +R\nonumber\\
&\geq c{\varepsilon }_{12}^{\frac{n-1}{n-2}} +R\end{aligned}$$ where $$\begin{aligned}
R&= +O\left(\frac{d(a_1,y_0)}{{\lambda }_1}+ \frac{d(a_2,y_0)}{{\lambda }_2} +\sum
\frac{1}{{\lambda }_k^2} \right)\\
&+O\left({\varepsilon }_{12}^{\frac{n}{n-2}}Log({\varepsilon }_{12}^{-1}) + \sum \frac{1}{\l
_k} \e
_{12}\left(Log({\varepsilon }_{12}^{-1})\right)^{\frac{n-2}{n}}+
d(a_1,a_2){\varepsilon }_{12}^{\frac{n+1}{n-2}} \sum {\lambda }_k\right).\end{aligned}$$ We also observe that $$\begin{aligned}
\label{e:42}
\frac{d_0}{{\lambda }_1} \leq {\varepsilon }_{12}^{\frac{n-\frac{1}{2}}{n-2}}= o\left(\e
_{12}^{\frac{n-1}{n-2}}\right),\end{aligned}$$ $$\begin{aligned}
\label{e:43}
\frac{1}{{\lambda }_1d(a_1,a_2)}\geq \left(\frac{1}{{\lambda }_1{\lambda }_2
d(a_1,a_2)^2}\right)^{\frac{1}{2}} \sim {\varepsilon }_{12}^{\frac{1}{n-2}},\end{aligned}$$ $$\begin{aligned}
\label{e:44}
{\lambda }_id(a_1,a_2){\varepsilon }_{12}^{(n+1)/(n-2)}= {\varepsilon }_{12}^{n/(n-2)}\left(\frac{\l
_i}{{\lambda }_j}\right)^{1/2}=o({\varepsilon }_{12}^{(n-1)/(n-2)}),\end{aligned}$$ $$\begin{aligned}
\label{e:45}
\frac{{\varepsilon }_{12}}{{\lambda }_1}\left(Log({\varepsilon }_{12}^{-1})\right)^{(n-2)/n}&=
\frac{{\varepsilon }_{12}}{\sqrt{{\lambda }_1d_0}}\frac{\sqrt{d_0}}{\sqrt{\l
_1}}\left(Log({\varepsilon }_{12}^{-1})\right)^{(n-2)/n}\nonumber\\
&=O\left(\frac{d_0}{{\lambda }_1}+ \frac{{\varepsilon }_{12}^2\left(Log(\e
_{12}^{-1})\right)^{2(n-2)/n} }{{\lambda }_1 d_0}\right)\nonumber\\
&=o\left({\varepsilon }_{12}^{(n-1)/(n-2)}\right)+o\left(\frac{{\varepsilon }_{12}}{{\lambda }_1
d_0}\right).\end{aligned}$$ In the same way, we have $$\begin{aligned}
\label{e:46}
\frac{{\varepsilon }_{12}}{{\lambda }_2}\left(Log({\varepsilon }_{12}^{-1})\right)^{(n-2)/n}=
o\left({\varepsilon }_{12}^{(n-1)/(n-2)}\right)+o\left(\frac{{\varepsilon }_{12}}{{\lambda }_2
d_0}\right)\end{aligned}$$ We also have, since ${\lambda }_1 |a_1-a_2| \to +\infty $, $$\begin{aligned}
\label{e:47}
{\lambda }_1^{-2}= o\left(d_0 {\lambda }_1^{-1}\right)= o\left(\e
_{12}^{(n-1)/(n-2)}\right).\end{aligned}$$ Similary, we have $$\begin{aligned}
\label{e:48}
{\lambda }_2^{-2}= o\left(\e
_{12}^{(n-1)/(n-2)}\right).\end{aligned}$$ Using , , , , , , and , we find $$\begin{aligned}
-\left({\nabla }J(u), W_2\right) &\geq C \left({\varepsilon }_{12}^{(n-1)/(n-2)}
+\frac{d_0}{{\lambda }_1}+ \frac{d_o}{{\lambda }_2}\right)\\
&\gamma \left({\varepsilon }_{12}^{\frac{n-\frac{1}{2}}{n-2}}+\sum \frac{1}{\l
_i^2}+ \sum |1-J(u)^{\frac{n}{n-2}}{\alpha }_i^{\frac{4}{n-2}}K(a_i)^2|\right)\end{aligned}$$ where $\gamma$ is a large constant.\
[**Case 2.2**]{}${\varepsilon }_{12}^{\frac{n-\frac{1}{2}}{n-2}} \leq
2\frac{d_0}{{\lambda }_1}$\
In this case, we set $$W_3=\frac{1}{{\lambda }_1}\left[{\alpha }_1 \frac{\partial\tilde{\delta }_1}{\partial
a_1}\left(\frac{y_0-a_1}{d_o}\right) + {\alpha }_2
\frac{\partial\tilde{\delta }_2}{\partial
a_2}\left(\frac{y_0-a_2}{d_0}\right)\right].$$ Using Proposition \[p:34\], we obtain $$\begin{aligned}
-\left({\nabla }J(u), W_3\right)&= J(u)2c_2 \frac{{\alpha }_1\a
_2}{{\lambda }_1}\frac{\partial{\varepsilon }_{12}}{\partial
a_1}\left(\frac{a_2-a_1}{d_0}\right)(1+o(1))\\
&+J(u)^{\frac{2(n-1)}{n-2}}\frac{c_5}{{\lambda }_1}\sum \a
_k^{\frac{2n}{n-2}}{\nabla }_TK(a_k)\frac{y_0-a_k}{d_0} +R_1\end{aligned}$$ where $$\begin{aligned}
R_1=& O\left(
\frac{1}{{\lambda }_1^2}+ \frac{1}{{\lambda }_2^2} + {\varepsilon }_{12}^{\frac{n}{n-2}}Log(\e
_{12}^{-1}) \right)\\
&+O\left(\sum \frac{1}{{\lambda }_k} {\varepsilon }_{12}\left(Log({\varepsilon }_{12}^{-1})\right)^{\frac{n-2}{n}}+
d(a_1,a_2){\varepsilon }_{12}^{\frac{n+1}{n-2}} \sum {\lambda }_k\right).\end{aligned}$$ Thus, we find $$\begin{aligned}
\label{e:49}
-\left({\nabla }J(u), W_3\right) & \geq \frac{c}{{\lambda }_1}{\lambda }_1{\lambda }_2
\frac{d(a_1,a_2)^2}{d_0}{\varepsilon }_{12}^{\frac{n}{n-2}}+\frac{c}{\l
_1d_0}\left(d(a_1,y_0)^2+d(a_2,y_0)^2\right)+R_1\nonumber\\
&\geq \frac{c}{{\lambda }_1d_0}{\varepsilon }_{12}+\frac{c}{{\lambda }_1}d_0 + R_1.\end{aligned}$$ Observe that $$\begin{aligned}
\label{e:410}
{\varepsilon }_{12}^{\frac{n}{n-2}}Log({\varepsilon }_{12}^{-1}) = o\left(\e
_{12}^{\frac{n-\frac{1}{2}}{n-2}}\right)= o\left(d_0 \l
_1^{-1}\right)\end{aligned}$$ $$\begin{aligned}
\label{e:411}
{\lambda }_k^{-2}=o\left(d_0 {\lambda }_k^{-1}\right)\qquad \mbox{for
}k=1, 2,\end{aligned}$$ $$\begin{aligned}
\label{e:412}
{\varepsilon }_{12}\left(Log({\varepsilon }_{12}^{-1})\right)^{\frac{n-2}{n}}\frac{1}{\l
_1}&= \frac{{\varepsilon }_{12}\left(Log(\e
_{12}^{-1})\right)^{\frac{n-2}{n}}}{\sqrt{{\lambda }_1d_0}}
\frac{\sqrt{d_0}}{\sqrt{{\lambda }_1}}\nonumber\\
&=o\left(\frac{d_0}{{\lambda }_1}\right)+O\left( \frac{{\varepsilon }_{12}^2\left(Log(\e
_{12}^{-1})\right)^{\frac{2(n-2)}{n}}}{{\lambda }_1d_0}\right)\nonumber\\
&=o\left(\frac{d_0}{{\lambda }_1} + \frac{{\varepsilon }_{12}}{{\lambda }_1d_0}\right).\end{aligned}$$ Using ,, and ,we obtain $$\begin{aligned}
-\left({\nabla }J(u), W_3\right) & \geq C\left(\frac{d_0}{\l
_1}+\frac{d_0}{{\lambda }_2}+\frac{{\varepsilon }_{12}}{{\lambda }_1d_0}\right)\\
&\geq C\left(\frac{d_0}{{\lambda }_1}+\frac{d_0}{{\lambda }_2}+ \e
_{12}^{(n-\frac{1}{2})/(n-2)}\right)\\
&\left(\sum\frac{1}{\l
_i^2}+ \sum (1-J(u)^{\frac{n}{n-2}}a
_i^{\frac{4}{n-2}}K(a_i)\right) + C{\varepsilon }_{12}^{\frac{n-\frac{1}{2}}{n-2}}.\end{aligned}$$ [**$3^{rd}$ region.**]{} $\exists i\in \{1,2\}$ such that $|1-J(u)^{\frac{n}{n-2}}a _i^{\frac{4}{n-2}}K(a_i)|\geq \frac{C'}{\l
_i}.$\
In this case, we set $$Z=-sign
\left(1-J(u)^{\frac{n}{n-2}}a_i^{\frac{4}{n-2}}K(a_i)\right)\tilde{{\delta }}
_i.$$ Using Proposition \[p:35\], we obtain $$\begin{aligned}
-\left({\nabla }J(u), Z\right)&\geq
C|1-J(u)^{\frac{n}{n-2}}a_i^{\frac{4}{n-2}}K(a_i)| +
O\left(\frac{1}{{\lambda }_i}\right)+O({\varepsilon }_{12})\\
&\geq \frac{C}{2}|1-J(u)^{\frac{n}{n-2}}a_i^{\frac{4}{n-2}}K(a_i)| +
\frac{C'}{4{\lambda }_i}+O({\varepsilon }_{12}).\end{aligned}$$ We also set $$Z_1=
\begin{cases}
W_1 \qquad \mbox{if} \qquad M{\lambda }_1 \leq {\lambda }_2 \\
-{\lambda }_2 \frac{\partial\tilde{\delta }_2}{\partial{\lambda }_2} \qquad
\mbox{if}\qquad 2M{\lambda }_1 \geq {\lambda }_2.
\end{cases}$$ Setting $W_4 = Z + Z_1 \sqrt{C'}$, we derive $$\begin{aligned}
-\left({\nabla }J(u), W_4\right)&\geq
C\left(|1-J(u)^{\frac{n}{n-2}}a_i^{\frac{4}{n-2}}K(a_i)| +
\sum \frac{1}{{\lambda }_k}+{\varepsilon }_{12}\right)\\
&\geq \gamma \left(\sum |1-J(u)^{\frac{n}{n-2}}a_k^{\frac{4}{n-2}}K(a_k)|^2 +
\sum \frac{1}{{\lambda }_k^2}+{\varepsilon }_{12}^{\frac{n-\frac{1}{2}}{n-2}}\right).\end{aligned}$$ Hence our global vector field will be built using a convex combination of $W_1$, $W_2$, $W_3$ and $W_4$ and will satisfy obviously i., iii. and iv. Next, we give the proof of ii. As in [@B1] and [@BCCH], it is easy to prove that $$\begin{aligned}
-\left({\nabla }J(u+\bar{v}), \frac{\partial{\bar{v}}}{\partial ({\alpha }_i, a_i,
{\lambda }_i)}(W)\right)&\leq c|\bar{v}| |{\nabla }J(u+\bar{v})|\\
&=O\left(|\bar{v}|^2 + |{\nabla }J(u+\bar{v})|^2\right).\end{aligned}$$ By Propositions \[p:33\] and \[p:34\], we can prove that $$|{\nabla }J(u+\bar{v})|= O\left(\sum |1-J(u)^{\frac{n}{n-2}}a_k^{\frac{4}{n-2}}K(a_k)| +\sum \frac{1}{{\lambda }_k}+{\varepsilon }_{12} +|\bar{v}|\right).$$ Using now the estimate of $\bar{v}$, we easily derive ii. Thus our proposition follows. $\Box$\
Next, we state the following result whose proof is similar to corresponding statements in Lemma A1.1 [@B1]
\[l:B\](Lemma A1.1 [@B1]) Let $u={\alpha }_{i_0}{\delta }_{(x_{i_0}, {\lambda }_{i_0})}+{\alpha }_{j_0}{\delta }_{(x_{j_0}, \l
_{j_0})}$, where\
(i)${\varepsilon }_{i_0 j_0} \geq {\delta }_1$, ${\delta }_1$ a given constant\
(ii)$B^2 \geq {\lambda }_{i_0}$, ${\lambda }_{j_0} \geq B$\
If, given ${\delta }_1$, $B$ is large enough, there is a pseudogradient vector field of $J$, built with the Yamabe gradient on $u$, which leads functions such as $u$ in the neighborhood of functions of the type ${\alpha }{\delta }_{(y,{\lambda })}+v$, where $y$ is close to $\frac{1}{2}(x_{i_0}+x_{j_0})$ up to $O(\frac{1}{B})$, ${\lambda }\geq cB$ ($c$ is a universal constant) and $||v||=o(1)$.
Now, we will use the above Lemma in the proof of the next main result.
\[p:62\] Let ${\varepsilon }_0 > 0$ small enough. There exists a vector field $Z_0$ defined in $$W_{{\varepsilon }_0}=\{{\alpha }_1 \tilde{\delta }_{x,{\lambda }_1} +{\alpha }_2 \tilde{\delta }_{y_0, {\lambda }_2}/
{\alpha }_i \geq 0, {\alpha }_1+{\alpha }_2 =1, x\in X, {\lambda }_i > {\varepsilon }_0^{-1}, {\varepsilon }_{12}< \e
_0\}$$ which can be extended to $$W(2,{\varepsilon }_0)=\{{\alpha }_1 \tilde{\delta }_{a_1,{\lambda }_1} +{\alpha }_2 \tilde{\delta }_{a_2, {\lambda }_2}
+ v \in V_b(2,{\varepsilon }_0)/ a_1, a_2 \in \overline{S^n_+}\}$$ so that the following holds:\
$f_{\lambda }(C_{y_0}(X))$ retracts by deformation on $X\cup W_u(y_0,
y_{i_0})_\infty \cup D$, where $D \subset \sigma $ is a stratified set (in the topological sense, that is, $D\in \Sigma_k(S^n_+)$, the group of chains of dimensions k) and where $\sigma = \cup _{y_i \in X\diagdown
\{y_{i_0}, y_0\}}W_u(y_0, y_i)_\infty$ is a manifold in dimension at most $k-1$.\
Here $W_u$ denotes the unstable manifold for $Z_0$.
First, we notice that assumption $(H_2)$ implies that any critical point $y$ of $K_1$ such that $y\in X$ satisfies $(\partial
K/\partial\nu ) > 0$. Now, we distinguish five cases\
[**Case 1.**]{} There exists $i$ such that ${\alpha }_i$ is far away from $J(u)^{\frac{-n}{4}}K(a_i)^{\frac{2-n}{4}}$($i=1,2$).\
We set $$Z_1=\tilde{\delta }_{a_i,{\lambda }_i} \dot{{\alpha }_i}\quad \mbox{with } \dot{{\alpha }_i}
=\begin{cases}
1 \quad \mbox{if } {\alpha }_i >
J(u)^{\frac{-n}{4}}K(a_i)^{\frac{2-n}{4}}+\eta\\
-1 \quad \mbox{if } {\alpha }_i <
J(u)^{\frac{-n}{4}}K(a_i)^{\frac{2-n}{4}}-\eta
\end{cases}$$ where $\eta$ is a positive constant.\
Using Proposition \[p:35\], we obtain $$\begin{aligned}
-\left({\nabla }J(u), Z_1\right)&= c + O\left({\lambda }_i^{-1}\right)+ O({\varepsilon }_{12})\\
&\geq C > 0.\end{aligned}$$ [**Case 2.**]{} For each $i \in \{1,2\}$, $\a
_i=J(u)^{\frac{-n}{4}}K(a_i)^{\frac{2-n}{4}}$ and $x\notin
\mathcal{V}(y_i,\rho)$, where\
$\rho < \frac{1}{2}\min _{i\not=
j}d(y_i,y_j)$ and $y_i$ is any critical point of $K_1= K_{/\partial S^n_+}$.\
In this case, we have $d(x,y_0) \geq c$, thus $\e
_{12}=o\left(\frac{1}{{\lambda }_i}\right)$ for $i=1, 2$.\
Two subcases may occur\
If ${\lambda }_1 \leq C_1 {\lambda }_2$, where $C_1$ is a large enough positive constant, we set $$Z_{21}= \frac{1}{{\lambda }_1}\frac{\partial\tilde{\delta }_1}{\partial a_1} {\nabla }_TK.$$ If ${\lambda }_1 \geq C_1 {\lambda }_2$, we set $$Z_{22}= Z_{21}+ {\lambda }_2 \frac{\partial\tilde{\delta }_2}{\partial{\lambda }_2}.$$ Using Propositions \[p:33\] and \[p:34\], we derive $$-\left({\nabla }J(u), Z_{2i}\right) \geq c {\lambda }_1^{-1} + c {\lambda }_2^{-1} +
{\varepsilon }_{12}\qquad \mbox{for }
i= 1, 2.$$ [**Case 3.**]{} For each $i \in \{1,2\}$, $\a
_i=J(u)^{\frac{-n}{4}}K(a_i)^{\frac{2-n}{4}}$ and $x\in
\mathcal{V}(y_i,2\rho)$, where\
$y_i$ is any critical point of $K_1$ such that $y_i\not= y_0$.\
Since $x\in X$, $y_i \in X$ and therefore $(\partial K/\partial\nu
)(y_i) > 0$. Now, we set $$Z_3 = {\lambda }_1 \frac{\partial\tilde{\delta }_1}{\partial{\lambda }_1} +{\lambda }_2
\frac{\partial\tilde{\delta }_2}{\partial{\lambda }_2}.$$ Using Proposition \[p:33\], we obtain $$-\left({\nabla }J(u), Z_{3}\right) \geq c {\lambda }_1^{-1}+ c {\lambda }_2^{-1}+ {\varepsilon }_{12}.$$ [**Case 4.**]{} For each $i \in \{1,2\}$, $\a
_i=J(u)^{\frac{-n}{4}}K(a_i)^{\frac{2-n}{4}}$ and $x\in
\mathcal{V}(y_0,2\rho)$\
In this case, we use the vector field defined in the proof of Proposition \[p:61\] which we combine with the vector field defined in Lemma \[l:B\].\
[**Case 5.**]{} ${\alpha }_1 = 0$ or ${\alpha }_2 = 0.$\
In this case, we only have one mass and we use the vector field defined in Lemma \[l:43\].\
\
Our global vector field $Z_0$ will be built using a convex combination of vector fields defined in cases 1- 5.\
Now, let $u= {\alpha }{\delta }_{x,\l} + (1-{\alpha }){\delta }_{y_0,\l} \in f_\l
(C_{y_0}(X))$.\
The action of the flow of the pseudogradient $Z_0$ is described as follow.\
If ${\alpha }< 1/2$, the flow of $Z_0$ brings $\a$ to zero, and thus in this case $u$ goes to $\overline{W_u((y_0)_\infty)} = \{y_0\}$.\
If ${\alpha }> 1/2$, the flow of $Z_0$ brings $\a$ to $1$, and thus $u$ goes, in this case, to $\overline{W_u((y_{i_0})_\infty)} =X$.\
If ${\alpha }= (1-{\alpha }) = 1/2$, we have an action on $x\in X=
\overline{W_s(y_{i_0})}$. In this case, $u$ goes to $W_s(y_i)$, where $y_i$ is a critical point of $K_1$ dominated by $y_{i_0}$ and two cases may occur :\
In the first case $y_i \not= y_0$, then $x$ goes to $W_u(y_0,
y_i)_\infty$.\
In the second case $y_i=y_0$, $u$ goes to $W_u(y_0)_\infty$ by the vector field defined in Lemma \[l:B\].\
Then our result follows. $\Box$\
\
We now prove our theorems.
[**Theorem \[t:13\]**]{} We argue by contradiction. Assume that (1) has no solution. The strong retract defined in Proposition \[p:62\] does not intersect $W_u(y_0,y_{i_0}))_\infty$ and thus it is contained in $X\cup D$ (see Proposition \[p:62\]). Therefore $H_*(X\cup D) = 0$, for all $*\in
{\mathbb{N}}^*$, since $f_{\lambda }(C_{y_0}(X))$ is a contractible set.\
Using the exact homology sequence of $(X\cup D, X)$, we have $$...\to H_{k+1}(X\cup D) \to ^{\pi} H_{k+1}(X\cup D, D) \to
^{\partial} H_k(X) \to ^{i} H_k(X\cup D) \to ...$$ Since $H_*(X\cup D) = 0$, for all $*\in {\mathbb{N}}^*$, then $H_k(X)=H_{k+1}(X\cup D, X)$.\
In addition, $(X\cup D, X)$ is a stratified set of dimension at most $k$, then $H_{k+1}(X\cup D,X) = 0$, and therefore $H_k(X)=0$. This yields a contradiction since $X$ is a manifold in dimension $k$ without boundary. Then our theorem follows.
\
[**Theorem \[t:14\]**]{} Assume that (1) has no solution. By the above arguments, $X\cup \left(\cup
_{y_i\in B_k}W_u(y_0, y_i)\right)\cup D$ is a strong retract of $f_\l
(C_{y_0}(X))$, where $D\subset \sigma$ is a stratified set and where $\sigma = \cup _{y_i\in X\diagdown \left(B_k \cup \{y_{i_0}\}\right)}W_u(y_0, y_i)_\infty$ is a manifold in dimension at most $k$.\
Since $\mu _i (y_0)=0$ for each $y_i \in B_k$, $f_{\lambda }(C_{y_0}(X))$ retracts by deformation on $X\cup D$, and therefore $H_*(X\cup D)=0$, for all $*\in {\mathbb{N}}^*$. Using the exact homology sequence of $(X\cup D,
D)$, we obtain $H_{k+1}(X\cup D, X) = H_k(X) = 0$, a contradiction, and therefore our result follows.
[99]{}
A. Ambrosetti , Y.Y. Li , A. Malchiodi, *Yamabe and Scalar Curvature Problems under boundary conditions*, preprint S.I.S.S.A., ref. 52/2000/M. Preliminary note on C.R.A.S., S[é]{}rie 1 330 (2000), 1013-1018. T. Aubin and A. Bahri, *Methodes de topologie algebrique pour le probl[è]{}me de la courbure scalaire prescrite,* J. Math. Pures et Appl. **76** (1997), 525–549. A. Bahri, *An invarient for Yamabe-type flows with applications to scalar curvature problems in high dimension,* A celebration of J. F. Nash Jr., Duke Math. J. **81** (1996), 323-466. A. Bahri, *Critical point at infinity in some variational problems,* Pitman Res. Notes Math, Ser **182**, Longman Sci. Tech. Harlow 1989. A. Bahri and J. M. Coron, *The scalar curvature problem on the standard three dimensional spheres,* J. Funct. Anal. **95** (1991), 106-172. A. Bahri and J.M. Coron, *On a nonlinear elliptic equation involving the critical Sobolev exponent : the effect of topology of the domain ,* Comm. Pure Appl. Math. **41**(1988), 255–294. A. Bahri and P. Rabinowitz, *Periodic orbits of hamiltonian systems of three body type,* Ann. Inst. H. Poincar[é]{} Anal. Non lin[é]{}aire **8** (1991), 561-649. M. Ben Ayed, Y. Chen, H. Chtioui and M. Hammami, *On the prescribed scalar curvature problem on 4-manifolds,* Duke Math. J. **84** (1996), 633-677. Ben Ayed, H. Chtioui and M. Hammami, *The scalar curvature problem on higher dimensional spheres,* Duke Math. J. **93** (1998), 379-424. G. Bianchi and X. B. Pan, *Yamabe equations on half spheres,* Nonlinear Anal. **37** (1999), 161-186. H. Brezis and J. M. Coron, *Convergence of solutions of H-systems or how to blow bubbles,* Arch. Rational Mech. Anal. **89** (1985), 21-56. S. A. Chang, M. J. Gursky and P. Yang, *The scalar curvature equation on 2 and 3 spheres,* Calc. Var. **1** (1993), 205-229. S. A. Chang and P. Yang, *A perturbation result in prescribing scalar curvature on $S^n$,* Duke Math. J. **64** (1991), 27-69. P. Cherrier, *Probl[è]{}mes de Neumann non lin[é]{}aires sur les vari[é]{}t[é]{}s Riemaniennes,* J. Funct. Anal. **57** (1984), 154-207. Z. Djadli, A. Malchiodi and M. Ould Ahmedou, *Prescribing the scalar and the boundary mean curvature on the three dimensional half sphere,* Preprint **** (2001). J. Escobar, *Conformal deformation of Riemannian metric to scalar flat metric with constant mean curvature on the boundary,* Ann. of Math. **136** (1992), 1-50. J. Escobar, *Conformal metrics with prescribed mean curvature on the boundary,* Cal. Var. **4** (1996), 559-592. Z. C. Han and Y.Y. Li, *The Yamabe problem on manifolds with boundaries : existence and compactness results,* Duke Math. J. **99** (1999), 489-542. Z. C. Han and Y.Y. Li, *The existence of conformal metrics with constant scalar curvature and constant boundary mean curvature,* Comm. Anal. Geom. **8** (2000), 809-869. E. Hebey, *Changements de metriques conformes sur la sphere, le probl[è]{}me de Nirenberg,* Bull. Sci. Math. **114** (1990), 215-242. Y.Y. Li, *The Nirenberg problem in a domain with boundary,* Top. Meth. Nonlin. Anal. **6** (1995), 309-329. Y.Y. Li, *Prescribing scalar curvature on $S^n$ and related topics, Part I,* J. Diff. Eq. **120** (1995), 319-410. Y.Y. Li, *Prescribing scalar curvature on $S^n$ and related topics, Part II : existence and compactness,* Comm. Pure Appl. Math. **49** (1996), 437-477. P. L. Lions, *The concentration compactness principle in the calculus of variations. The limit case,* Rev. Mat. Iberoamericana **1** (1985), I: 165-201; II: 45-121. J. Milnor, *Lecturess on h-Cobordism Theorem,* Princeton University Press, Princeton **** 1965. O. Rey, *Boundary effect for an elliptic Neumann problem with critical nonlinearity,* Comm. Partial Diff. Eq. **22** (1997), 1055-1139. R. Schoen and D. Zhang, *Prescribed scalar curvature on the n-sphere,* Calculus of Variations and Partial Differential Equations, **4** (1996), 1-25. M. Struwe, *A global compactness result for elliptic boundary value problems involving nonlinearities,* Math. Z. **187** (1984), 511-517.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We revisit experimental studies performed by Ekman on dead-water [@bib:Ekman1904] using modern techniques in order to present new insights on this peculiar phenomenon. We extend its description to more general situations such as a three-layer fluid or a linearly stratified fluid in presence of a pycnocline, showing the robustness of dead-water phenomenon. We observe large amplitude nonlinear internal waves which are coupled to the boat dynamics, and we emphasize that the modeling of the wave-induced drag requires more analysis, taking into account nonlinear effects.'
author:
- 'Matthieu J. Mercier'
- Romain Vasseur
- Thierry Dauxois
bibliography:
- 'biblio\_IW.bib'
title: 'Resurrecting Dead-water Phenomenon'
---
*Dedicated to Fridtjöf Nansen born 150 years ago (10 October 1861).*
Introduction
============
For sailors, the dead-water phenomenon is a well-known peculiar phenomenon, when a boat evolving on a two-layer fluid feels an extra drag due to waves being generated at the interface between the two layers whereas the free surface remains still. Interestingly, one finds reports of similar phenomena in the Latin literature when Tacitus described a flat sea on which one could not row a boat, North of Scotland and of Germany, in the Agricola [@bib:Tacitus1] and in the Germania [@bib:Tacitus2].
This effect is only observed when the upper part of the fluid is composed of layers of different densities, due to variations in salt concentration or temperature. An important loss of steering power and speed is experienced by the boat, which can even undergo an oscillatory motion when the motors are stopped.
In this paper, we present detailed experimental results on the dead-water phenomenon as shown in the video by [@bib:videoSanAntonio]. The material is organized as follows. In the remaining of this section, we briefly review the different studies of this phenomenon, either directly related to Ekman’s work or only partially connected to it. Section \[experimentalsetup\] presents the experimental set-up. The case of a two-layer fluid is addressed in Sec. \[twolayer\], followed by the case with a three-layer fluid in Sec. \[threelayer\]. The more realistic stratification with a pycnocline above a linearly stratified fluid is finally discussed in Sec. \[pycnocline\]. Our conclusions, and suggestions for future work are presented in section \[sct:conclusion\].
Ekman’s PhD Thesis {#sec:EkmanPhD}
------------------
V. W. Ekman was the first researcher to study in detail the origin of the dead-water phenomenon. His work as a PhD student [@bib:Ekman1904] was motivated by the well-documented report from the Norvegian Artic explorer Fridtjöf Nansen who experienced it while sailing on the [*Fram*]{} near “Nordeskiöld" islands in 1893 [@bib:Nansen1897].
Several aspects of the phenomenon have been described by Ekman, who did experiments with different types of boat evolving on a two-layer fluid. We note $\rho_1$ and $h_1$ the density and depth of the upper layer, and $\rho_2$ and $h_2$ those of the lower layer.
i\) First of all, the drag experienced by the boat evolving on the stratified fluid is much stronger than the one associated with an homogeneous fluid. This difference is due to wave generation at the interface between the two layers of fluid, pumping energy from the boat.
(8.5,3.75) (0.25,-1)[![Drag-speed relations given by Ekman for a boat dragged on a layer of fresh water ($\rho_1=1.000$ g cm$^3$) of depth $h_1=2.0$ cm, resting above a salt layer ($\rho_2=1.030$ g cm$^3$), and compared to the homogeneous case “deep water" with a water level of $23$ cm, and “shallow water" corresponding to a smaller depth 5 cm, and 2.5 cm (taken from [Fig. 8, Pl. VI @bib:Ekman1904]). Experimental points are crosses and continuous lines are models.[]{data-label="fig:Fv_Ekman"}](PL_VI_Fig_8_refresh3.eps "fig:"){width="8\unitlength"}]{}
This effect is the strongest when the boat’s speed is smaller than the maximum wave speed [@bib:Gill1982], defined as $$c_{\phi}^{m}=\sqrt{g\frac{\rho_2-\rho_1}{\rho_2}\frac{h_1 h_2}{h_1+h_2}}\,,
\label{eq:cphi_2layer}$$ as can be seen in Fig. \[fig:Fv\_Ekman\] where the value of $c_\phi^m$ is indicated by the circle. A typical evolution of the drag versus speed obtained by Ekman is shown in this figure. The experimental points are crosses and are compared to linear theory of viscous drag in steady state (continuous lines). The local maximum of the drag for the stratified case is reached for a speed slightly smaller than $c_\phi^m$. At high speeds compared to $c_\phi^m$, the drag is similar to the quadratic law for homogeneous fluids, dominated by viscous drag.
This behavior has been reproduced since then, and similar results were obtained by [@bib:Milohetal93] or [@bib:Vosperetal99] for instance.
ii\) Another contribution of Ekman concerns the description of the interfacial waves generated at the rear of the boat. Two types of waves, transverse and divergent, could be observed. It must be noted that only interfacial waves can be easily observed since the surface waves associated are of very weak amplitude, their amplitude being related to the interfacial waves ones with a ratio equals to ($\rho_{a}-\rho_1)/(\rho_2-\rho_1)\simeq1/500$ (where $\rho_a$ is the density of air). The free surface remains still at the laboratory scale.
These waves are generated by a depression that develops at the rear of the boat while moving. The transverse waves can reach large amplitudes up to breaking. But they tend to disappear when the speed of the boat is greater than $c_\phi^m$, where only divergent waves remain.
Visualizations of these waves from numerical simulations can also be observed in [@bib:Milohetal93; @bib:Yeung99]. Their properties are set by the Froude number only, which compares the mean speed of the moving object to $c_\phi^m$. This will help us in our description later on.
iii\) Furthermore, a solitary wave at the bow of the boat can also be observed. This structure, a spatially localized bump is reminiscent of solutions of the Korteweg-DeVries (KdV) equation: it can evolve freely, conserving its shape when the boat is stopped. Otherwise, it remains trapped to the boat.
iv\) From these observations, Ekman gave an interpretation of the hysteretic behavior the speed of a boat can experience when evolving in dead-water. The analysis can be explained with the help of Fig. \[fig:Fv\_Ekman\] showing the drag-speed relation established by Ekman from linear theory. If the force moving the boat increases such that the boat accelerates from rest, the speed of the boat can jump from $6$ to $15$ cm.s$^{-1}$ when the boat overcomes the maximum drag. Similarly, when the force diminishes such that the boat decelerates from a speed larger than $c_\phi^m$, a sudden decrease from $11$ to $4$ cm.s$^{-1}$ occurs. The range of values between $6$ and $11$ cm.s$^{-1}$ is thus an unstable branch inaccessible to the system, where dead-water phenomenon occurs.
However it is important to emphasize that this analysis implies changing the moving force, hence not imposing a constant one. This observation is different from another remark made by Ekman which relates explicitly to the apparent unsteady behavior of the boat while towed by a constant force [@bib:Ekman1904 p. 67]. He observed oscillations of the speed of the boat that could be of large amplitude compared to the mean value. He noticed that they occur when the boat is evolving at speeds smaller than $c_\phi^m$, while amplitudes and periods of the oscillations depend on the towing force and the properties of the stratification. We emphasize here that this last property is not included in analytical approaches considering linear waves, such as [@bib:Ekman1904; @bib:Milohetal93; @bib:Yeung99], and seems to be an important characteristic of the dead-water phenomenon.
Other dead-water related works
------------------------------
[@bib:Hughes78] took advantage of the dead-water effect to study the effects of interfacial waves on wind induced surface waves. The study relates the statistical properties of surface waves to the currents induced by internal waves.
[@bib:Maas06] investigated if the dead-water effect could also be experienced by swimmers in a thermally stratified pool, offering a plausible explanation for unexplained drownings of experienced swimmers in lakes during the summer season, but found no effect. One can argue that the stratification considered might have been inadequate for swimmers to generate waves and led to mixing of the thermocline mainly. An energetic budget is given in a more detailed and idealized study [@bib:Maasetal09] where the authors also observed some retarding effects on the swimmers.
In a slightly different perspective, [@bib:Nicolaou95] demonstrated that an object accelerating in a stratified fluid generates oblique and transverse internal waves, the latter can be decomposed as a sum of baroclinic modes with the lowest mode always present. [@bib:Shishkina02] further showed through experiments that in such a dynamical evolution, the baroclinic modes generated propagate independently of each other, although nonlinear effects must become important when the amplitude of the internal waves is increasing.
Steady motion, body moving at constant speed
--------------------------------------------
Finally, numerous studies were focused on bodies evolving at constant speed [*within*]{} a stratified fluid. Some results can help the understanding of the dead-water experiments, especially the internal waves at the rear of the boat and the wave-induced drag on the boat.
In the case of a two-layer fluid, the drag on the boat is maximal when the Froude number, defined as the ratio of the boat speed $U$ to the maximum wave speed given in Eq. (\[eq:cphi\_2layer\]), is slightly less than 1 [@bib:Milohetal93; @bib:Motygin97]: this is the subcritical regime. The structure of the internal waves generated and their coupling with surface waves confirm that the dead-water regime is due to baroclinic waves only [@bib:Yeung99].
These results obtained in a linear case can be extended when considering weakly nonlinear effects [@bib:Baines1995]. Fully nonlinear calculations are needed when the amplitude of the waves reaches about 0.4 times the depth of the thinner layer [@bib:Grueetal97; @bib:Grueetal99].
In the case of a linearly stratified fluid, the drag is again maximal for slightly subcritical values of the Froude number [@bib:Greenslade00]. Nevertheless, the internal waves emitted can be very different depending on the regime considered for the Froude number [@bib:Chomazetal93], the location of the object being at the surface [@bib:RottmanetalAPS2004] or fully immersed [@bib:Hopfingeretal91; @bib:Meunieretal06]. It is interesting to emphasize that in experiments done at constant speed, difficulties are often encountered to reach a steady state, as noticed by [@bib:Vosperetal99] for instance.
Nevertheless, the dead-water phenomenon does not correspond to a constant speed evolution. By imposing a constant force to move the boat, or using a motor at constant power, the speed of the boat is free to evolve. The dynamical study of the problem is thus much richer than in the steady state case.
Experimental setup {#experimentalsetup}
==================
Nowadays technologies give us the opportunity to gain more insight into the interactions between the interfacial waves and the boat, and improve our knowledge of the dead-water phenomenon. The experimental setup is described in Figure \[fig:eauxmortes\_montage\].
We drag a plastic Playmobil© boat of width $10$ cm (and with a fisherman and fishes to modulate its weight) with a falling weight in a $3$-m long plexiglass tank of width $10.5$ cm filled with a stratified fluid. A belt with fixed tension is used to move the boat with a constant horizontal force. The tension of the belt is set to a constant value throughout all the experiments. The falling weight of mass $m$ fixed to the belt sets the boat into motion, such that the constant force used is gravity $mg$. One must notice that the weights used are paper clips of few milligrams, since a very small force is required to propel the boat within the interesting regime. A magnet fixed to the boat is used to release it in a systematic way thanks to an electro-magnet outside the tank.
(10,2.5) (0.25,-0.5)[![Experimental setup of a boat dragged by a falling weight. The vertical force is converted horizontally through pulleys and an horizontal belt of fixed tension.[]{data-label="fig:eauxmortes_montage"}](schema_principe_eauxmortes "fig:"){width="8\unitlength"}]{} (1,2.5)[$z$]{} (2.2,1.3)[$x$]{} (1.05,1.35)[(1,0)[1]{}]{} (1.05,1.35)[(0,1)[1]{}]{}
Stratification
--------------
Different types of stratification have been used. The density profiles presented here are obtained using a conductivity and temperature probe (CT-probe) from PME©.
1\) The two-layer fluid is composed of a layer of fresh water (density $\rho_1$) colored with red food dye resting above a transparent layer of salt water (density $\rho_2>\rho_1$). By siphoning the interface, the density jump extension can be reduced to a few millimeters but through successive experiments, diffusion and mixing make it widen with time, up to a few centimeters as can be seen in Fig. \[fig:bicouche\_profile\].
(10,4.5) (0.5,0)[![Density profile $\rho$ and Brunt-Väisälä frequency $N$ for a two-layer fluid after a series of experiments. The density jump is approximately $4$ cm wide as observed in the zoomed window close to the free surface.[]{data-label="fig:bicouche_profile"}](2layer_exp2_rhoN_noaxes "fig:"){width="8\unitlength"}]{} (0,1.9) (2,-0.5)[$\rho$ (kg.m$^{-3}$)]{} (6,-0.5)[$N$ (rad.s$^{-1}$)]{}
ii\) The three-layer fluid is obtained similarly, by adding another layer of salt fluid (density $\rho_3>\rho_2$) colored with green food dye from the bottom.
iii\) Finally, a continuously stratified fluid composed of an homogeneous layer above a linearly stratified part can be obtained from adding a fresh layer after filling the tank with the classic “two-bucket" method [@bib:Hill02]. An example of such stratification with pycnocline is shown in Fig. \[fig:eauxmortes\_linear\_profil\], along with the Brunt-Väisälä frequency associated with it.
(10,4) (0.5,0)[![Density profile $\rho$ and Brunt-Väisälä frequency $N$ obtained experimentally in the case of a linear stratification with a pycnocline.[]{data-label="fig:eauxmortes_linear_profil"}](strat2_profile13_rhoN_noaxes "fig:"){width="8\unitlength"}]{} (0,1.9) (2,-0.5)[$\rho$ (kg.m$^{-3}$)]{} (6,-0.5)[$N$ (rad.s$^{-1}$)]{}
Techniques of image analysis
----------------------------
We record the dynamics of the system (waves + boat) using a black and white camera. Depending on the stratification considered, two different techniques are used to extract the dynamics. For the two and three-layer cases, the different layers are identified using food dye and the position of the interfaces $\eta(x,t)$ are extracted from highly contrasted images, as can be seen in Fig. \[fig:eauxmortes\_technique1\]. The boat position with time $x(t)$ and the free surface evolution are also obtained from these images. Several hypothesis must be highlighted when dealing with information obtained with this technique, that we called technique $1$:
- we consider the interface between two layers as infinitely thin,
- we neglect diffusion of salt and dye, which is a slow phenomenon compared to the experiments (typically a day versus a few minutes),
- we neglect the small scale evolutions of the interface, more specifically mixing events.
(10,5) (0.4,2.6)[![(a) Gray scale image during an experiment, converted into a (b) two-level black and white image with technique 1.[]{data-label="fig:eauxmortes_technique1"}](2406_08_image70_raw_resized "fig:"){width="8.01\unitlength"}]{} (0.5,4.4)[(a)]{} (0.4,0)[![(a) Gray scale image during an experiment, converted into a (b) two-level black and white image with technique 1.[]{data-label="fig:eauxmortes_technique1"}](2406_08_image70_NB "fig:"){width="8\unitlength"}]{} (0.5,1.8)[(b)]{} (0,0.75)[$z$]{} (1.2,-0.45)[$x$]{} (0.05,-0.4)[(1,0)[1]{}]{} (0.05,-0.4)[(0,1)[1]{}]{}
The second technique used, technique $2$, is based on synthetic schlieren [@bib:Dalzieletal00; @bib:Dalzieletal07] and can be realized simultaneously with technique $1$. By computing correlations between images in the stratified case and a reference image with homogeneous water, one can measure the complete density gradient and not only the one associated with the internal waves field. In the domain considered, one can thus access $\partial_x\rho(x,z,t)$ and $\partial_z\rho(x,z,t)$. In the absence of any motion, the second quantities have been integrated to measure the density profile and it was compared successfully to the one obtained with the CT-probe. Allowing to quantify the large oscillations of strong density gradient, this technique has been especially considered in the case of a continuous stratification with a pycnocline.
We have checked that technique $1$ gives a good indication of the position of the interface by recording synchronously a two-layer case with both techniques. As can be seen in Fig. \[fig:eauxmortes\_technique2\] where the image (a) is recorded with technique $1$ and image (b) with technique $2$, the position of the interface (black line) extracted from the image (a) follows the main evolution of the density jump in (b) where the vertical density gradient is the strongest.
(10,5.2) (0.52,-0.1)[![Technique 2 for a two-layer fluid experiment: (a) color image (taken with camera) and (b) gray scale image over which is superposed the vertical gradient of density $\partial_z\rho$ in g cm$^{-4}$, obtained using synthetic schlieren with a reference image with homogeneous water. The black line corresponds to the interface position extracted from image (a) and added to image (b) while the pink dashed line is the free surface.[]{data-label="fig:eauxmortes_technique2"}](2layer_exp2_6tr_AS_im7 "fig:"){width="5.975\unitlength"}]{} (0,0.65) (2.75,-0.5)[$x$ (m)]{} (6.5,0.9)[$\partial_z\rho$ (g cm$^{-4}$)]{} (1,2.2)[(b)]{} (0.5,3)[![Technique 2 for a two-layer fluid experiment: (a) color image (taken with camera) and (b) gray scale image over which is superposed the vertical gradient of density $\partial_z\rho$ in g cm$^{-4}$, obtained using synthetic schlieren with a reference image with homogeneous water. The black line corresponds to the interface position extracted from image (a) and added to image (b) while the pink dashed line is the free surface.[]{data-label="fig:eauxmortes_technique2"}](2layer_exp2_6tr_canon_im7 "fig:"){width="8\unitlength"}]{} (0,3.5) (4,2.6)[$x$ (m)]{} (1,4.9)[(a)]{}
Although technique $2$ can be used for all experiments, we will use technique $1$ to analyze the two and three-layer cases since its computational cost is much less.
Revisiting Ekman’s work {#twolayer}
=======================
Our first experimental investigation is dedicated to the two-layer case, in order to reproduce the results obtained by Ekman. The main parameters are summarized in Table \[tab:params\_eauxmortes\].
[llccl]{} & symbols & values & units\
\
Tank & dimensions & $L\,\times\,H\,\times\,W$ & $300\,\times\,50\,\times\,10.5$ & cm$^{3}$\
\
& belt tension & $T$ & $\sim 0.35$ & N\
& force & $F_t$ & $0.011 - 0.035$ & N\
\
& dimensions & $L_b\,\times\,h_b\,\times\,w_b$ & $20.0\,\times\,5.0\,\times\,10.0$ & cm$^{3}$\
& immersed section & $S_b$ & $12.0 - 24.0$ & cm$^{2}$\
& mass & $M$ & $171 - 343$ & g\
\
& density & $\rho_1$ & $0.999 - 1.005$ & g cm$^{-3}$\
& depth & $h_1$ & $2.0 - 5.0$ & cm\
\
& density & $\rho_2$ & $1.010 - 1.030$ & g cm$^{-3}$\
& depth & $h_2$ & $5.0 - 15.0$ & cm\
\
& mean density & $\bar{\rho}=\frac{\rho_2+\rho_1}{2}$ & $\sim 1.01$ & g cm$^{-3}$\
& density jump & $\Delta\rho=\frac{\rho_2-\rho_1}{\bar{\rho}}$ & $0.01 - 0.1$ &\
& maximum phase speed & $c_{\phi}^m=\sqrt{\Delta\rho g\frac{h_1 h_2}{h_1+h_2}}$ & $0.03-0.2$ & m s$^{-1}$\
& Froude number & $Fr=\frac{U}{c_{\phi}^m}$ & $0.2-2$ &\
& Reynolds number & $Re=\frac{U h_1}{\nu}$ & $400-10000$ &\
Dynamics of the boat
--------------------
As can be seen in Table \[tab:params\_eauxmortes\], we have used several sets of parameters in order to verify that the dead-water phenomenon is indeed a function of the relative density jump $\Delta\rho=({\rho_2-\rho_1})/{\bar{\rho}}$, the fresh layer depth $h_1$ compared to the salted one $h_2$ along with the waterline $h_b$ on the boat (changing $h_b$ implies a different immersed cross-section and thus a different viscous drag). In the following, we will refer to the immersed cross-section $S_b$ for varying geometry of the boat, and which corresponds to $w_b\times h_b$.
For instance, the dead-water phenomenon is easy to reproduce with the following parameters: $h_1=5$ cm, $h_2=14$ cm, $h_b\simeq 2.4$ cm, $\Delta\rho=0.0244$ and the falling weight corresponding to a force $F_t=16.3$ mN (also studied in Sec. \[sec:2layerXTdiagram\], Fig. \[fig:eauxmortes\_oscillations\]). The time evolution of the boat $x(t)$ and its corresponding oscillating speed $v(t)=dx/dt$ normalized by $c_{\phi}^m$, associated with this experiment, are shown in blue in Fig. \[fig:xdxdt\_deadwater\]. A similar case but with a force $F_t=18.8$ mN (also studied in Sec. \[sec:2layerXTdiagram\], Fig. \[fig:Fr\_o1\_reflab\]) is shown in green in Fig. \[fig:xdxdt\_deadwater\].
(10,5.2) (0.5,2.75)[![(a) Position of the boat $x(t)$ (in m) and (b) its speed compared to the maximum phase speed $v(t)/c_{\phi}^m$ for experiments revealing the dead-water phenomenon (blue and green) and for an experiment without oscillation (red).[]{data-label="fig:xdxdt_deadwater"}](2506_11-10-09_xt_boat "fig:"){width="8\unitlength"}]{} (0,3.85) (1.2,4.85)[(a)]{} (0.5,0)[![(a) Position of the boat $x(t)$ (in m) and (b) its speed compared to the maximum phase speed $v(t)/c_{\phi}^m$ for experiments revealing the dead-water phenomenon (blue and green) and for an experiment without oscillation (red).[]{data-label="fig:xdxdt_deadwater"}](2506_11-10-09_dxdt_boat "fig:"){width="8\unitlength"}]{} (0,1.1) (4.25,-0.5)[$t$ (s)]{} (1.2,2.1)[(b)]{}
Ekman already commented on the fact that the speed can have large fluctuations $|v_{max}-v_{min}|\sim 0.027$ m.s$^{-1}$ compared to its mean value $<v>\simeq 0.031$ m s$^{-1}$ (blue case), referring to the capricious nature of the phenomenon. Oscillations of the order of $85\%$ of the mean value make it difficult to interpret the dead-water regime as stationary. The force imposed to move the boat being constant, such a temporal evolution implies a time-evolving drag due to the interfacial waves, and that is due to generation and breaking of waves at the rear of the boat. This will be confirmed later on when we present the spatio-temporal diagram associated with the interface position.
We also show in red in Fig. \[fig:xdxdt\_deadwater\] the time evolution of $x(t)$ and $v(t)/c_{\phi}^m$ for the same parameters except $F_t=21$ mN (also studied in Sec. \[sec:2layerXTdiagram\], Fig. \[fig:Fr\_o1\_reflab\]). Although only reached at the end of the recording, the boat tends to a steady evolution at constant speed, $v_\ell\simeq0.14$ m s$^{-1}$, after a transient state (for $t\leq15$ s). The comparison between the red and green case also reveals the sensitivity of the system regarding the drag force. The temporal evolutions are initially very similar (for the first five seconds) before behaving differently. For sufficiently large speeds, the hypothesis of constant velocity of the boat seems reliable and can lead to a good interpretation of the evolution of the boat.
Analogously to Ekman’s results presented in Fig. \[fig:Fv\_Ekman\], and in order to compare our results with his, we summarize in Figs. \[fig:Fv\_3h1\] and. \[fig:Fv\_3boats\] several time evolutions for the different stratifications and boat configurations considered in a speed-force diagram. However, two comments must be made with respect to Ekman’s presentation of his experimental results. When the boat is in an oscillating regime, Ekman takes the mean value as its speed and considers that the drag equals the moving force. This simplified representation of the data leads to cleaner diagrams although some strong hypotheses are hidden.
In our case, we consider the moving force $F_t$ instead of the drag force since both forces are equal only when the boat evolution is steady. We represent the speed by its mean value (symbols) and the range of values (lines) in which it fluctuates. Figure \[fig:Fv\_3h1\] combines the speed-force diagram associated with three different stratifications and one boat configuration ($S_b=12$ cm$^2$), whereas Fig. \[fig:Fv\_3boats\] gathers three boat configurations and one stratification.
(10,3.25) (0.5,0)[![Speed-Force diagram for three stratifications with a fresh layer $h_1$ being respectively $(\ast)0.06$ m, $(\circ)0.05$ m and $(\lhd)0.04$ m; with a salt layer $h_2=0.14$ m and $\Delta\rho=0.0247$, for a boat configuration being $S_b=12$ cm$^2$. For the unsteady regime, horizontal lines are drawn corresponding to the range of values of the speed in which it fluctuates. The continuous line corresponds to a homogeneous case of depth $0.15$ m. The vertical dashed lines gives the values of $c_\phi^m$ associated with $h_1$, (dotted line) 0.10 m s$^{-1}$, (dash-dotted line) 0.095 m s$^{-1}$ and (dashed line) 0.087 m s$^{-1}$.[]{data-label="fig:Fv_3h1"}](Fv_3h1 "fig:"){width="8\unitlength"}]{} (4,-0.5)[$v_\ell$ (m s$^{-1}$)]{} (0,1.2)
(10,3.25) (0.5,0)[![Speed-Force diagram for three boat configurations, $(\ast)-S_b=12$, $(\lhd)-S_b=18$, and $(\circ)-S_b=24$ cm$^2$ in a stratification with $h_1=0.06$ m, $h_2=0.10$ m and $\Delta\rho=0.0145$. For the unsteady regime, horizontal lines are drawn corresponding to the range of values of the speed in which it fluctuates. The continuous line corresponds to a homogeneous case of depth $0.15$ m with $S_b=18$ cm$^2$. The vertical dashed line gives the value of $c_\phi^m$[]{data-label="fig:Fv_3boats"}](Fv_3bateaux "fig:"){width="8\unitlength"}]{} (4,-0.5)[$v_\ell$ (m s$^{-1}$)]{} (0,1.2)
These two diagrams are in agreement with the established results by Ekman that the wave-induced drag increases along with the fresh water depth $h_1$ when a steady state exists (Fig. \[fig:Fv\_3h1\]), and depends also on the boat geometry (Fig. \[fig:Fv\_3boats\]). As the immersed section of the boat is larger, the drag gets stronger. The best coupling occurs for $h_b\simeq h_1/2$ where the wave induced drag is maximum and the boat is not deep enough to generate mixing at the interface. Furthermore, both diagrams show that when the speed of the boat is oscillating, it never reaches values larger than $c_\phi^m$. It is important to notice that the dynamical regime where the velocity of the boat is closed to $c_\phi^m$ is difficult to scan since the tiniest change of the moving force leads to a completely different evolution of the boat. One can thus argue there exists a range of “unattainable" values for the velocity of the boat, as discussed in Sec. \[sec:EkmanPhD\], § iv. However, since the moving force is kept constant in our case, this aspect is different from Ekman’s presentation of the phenomenon. Similar comparisons have been made for the influence of the density jump $\Delta\rho$ and the salt layer depth $h_2$ but are not shown here for the sake of concision.
Finally, several general comments are in order
- the wave-induced drag is always stronger than the one in homogeneous water (viscous drag only),
- a regime of low velocities exists with a non constant drag and large fluctuations in the boat speed (no steady state can be reached),
- this regime is always associated with velocities smaller than $c_\phi^m$.
From these remarks, we can discriminate the two regimes described previously according to the Froude number $Fr$ which is defined as the ratio of the speed of the boat (that can be its mean value $<v>$ or its limit $v_\ell$) to the maximum phase speed of the interfacial waves. It is important to realize here that the Froude number is intrinsically a time evolving parameter, but we will use characteristic values to identify the regime considered.
We now focus on the waves according to this classification.
Interfacial waves {#sec:2layerXTdiagram}
-----------------
We present here observations of the interfacial position $\eta(x,t)$ generated by the boat in two different frames, the laboratory reference frame and the one moving with the boat. The latter is not Galilean since the boat does not evolve at constant speed, but offers a better understanding of the coupled dynamics of the waves with the boat.
Visualizations of the interfacial positions are represented in spatio-temporal diagrams where elevations of the interface are in red and depressions in blue (see Figs. \[fig:Fr\_O1\_reflab\] and \[fig:Fr\_o1\_reflab\] for instance). The bow and stern of the boat are superimposed on these diagrams (as black lines).
Depending on the Froude number being larger or smaller than $1$, the wave dynamics is very different. We illustrate each case with an example corresponding to a stratification with $h_1=5.0$ cm, $h_2=14.0$ cm, $\rho_1=0.9980$ g cm$^{-3}$, $\rho_2=1.0227$ g cm$^{-3}$, leading to $c_\phi^m=0.094$ m s$^{-1}$. The boat used corresponds to an immersed section $S_b=24$ cm$^2$.
### In the laboratory frame {#sect:eauxmortes_bicouche_reflabo}
Let us first consider the case where a steady state is reached, illustrated in Fig. \[fig:Fr\_O1\_reflab\]. With a moving force $F_t=21$ mN, the limit speed reached by the boat is $v_\ell\simeq0.15$ m s$^{-1}$ which gives $Fr=1.6$. In this supercritical case, one can observe that the boat escapes from the wave train it generates at start-up. A depression (the blue region) remains attached below the boat and propagates with it at the speed of the boat. For times larger than $40$ s, the boat meets with the end of the tank and stops (around $x=2.5$ m). The depression below it is thus expelled and propagates in the other direction after reflecting on the side of the tank. This case is in the ballistic regime.
(10,5.25) (0.5,-0.1)[![[*Ballistic regime*]{} ([*$Fr>1$*]{}). Spatio-temporal diagram of interfacial displacements $\eta(x,t)$ (in m). The solid black lines are the bow and stern of the boat. Experimental parameters: $h_1=5.0$ cm, $h_2=14.0$ cm, $\rho_1=0.9980$ g cm$^{-3}$, $\rho_2=1.0227$ g cm$^{-3}$, $S_b=24$ cm$^2$, $F_t=21$ mN.[]{data-label="fig:Fr_O1_reflab"}](2506_11_xt "fig:"){width="8\unitlength"}]{} (3.9,-0.5)[$x$ (m)]{} (0,2.25) (7,5.2)[$\eta(x,t)$ (m)]{}
The oscillating regime, more atypical, with oscillations of the boat is illustrated in Fig. \[fig:Fr\_o1\_reflab\] and corresponds to $F_t=18.8$ mN. The mean speed of the boat is $<v>\simeq0.035$ m s$^{-1}$, which gives a Froude number of $0.4$. As the boat evolves, the amplitude of the interfacial waves grows up to a maximum value where the waves catch up with the boat and break on its hull, around $t=20$ s. This corresponds to the moment where the drag force is maximal since the boat is almost standing still. The vessel can speed up again after the first wave breaks and the same dynamics is reproduced.
(10,5.25) (0.5,-0.1)[![[*Oscillating regime*]{} ([*$Fr< 1$*]{}). Spatio-temporal diagram of interfacial displacements $\eta(x,t)$ (in m). The solid black lines are the bow and stern of the boat. Experimental parameters: $h_1=5.0$ cm, $h_2=14.0$ cm, $\rho_1=0.9980$ g cm$^{-3}$, $\rho_2=1.0227$ g cm$^{-3}$, $S_b=24$ cm$^2$, $F_t=18.8$ mN.[]{data-label="fig:Fr_o1_reflab"}](2506_10_xt "fig:"){width="8\unitlength"}]{} (3.9,-0.5)[$x$ (m)]{} (0,2.25) (7,5.2)[$\eta(x,t)$ (m)]{}
As noted earlier, a depression located below the boat evolves with the boat when it is accelerating. It evolves freely after the boat is stopped ($t>20$ s) and propagates to the right first and to the left for $t>50$ s after reflection at the side of the tank. This depression even bounces a second time on the left side of the tank and its shape remains almost unchanged ($t>75$ s). Although not presented here, we have verified that this solitary wave propagating upstream is well described by a Korteweg-de Vries model since its amplitude is always smaller than $0.4~h_1$ [@bib:Grueetal99]. When the boat starts again, the process is repeated and a new depression is generated.
### In the frame of the boat
More information can be obtained when following the dynamics in the frame associated with the boat. We are more specifically interested in what sets the amplitude and frequency of the oscillations of the boat when $Fr<1$.
In Fig. \[fig:eauxmortes\_raidissement\], we have superposed the interfacial displacements in the frame of the boat (represented as a black rectangle) for different times extracted from Fig. \[fig:Fr\_o1\_reflab\]. We can observe that the waves get closer to the stern of the boat as their amplitude grows. This is mainly due to the decrease of speed of the boat since we can see in Fig. \[fig:Fr\_o1\_reflab\] that the wave crests evolve with a constant speed. The nonlinear nature of the interfacial waves is also clearly visible as the wave fronts steepen, the limit of growth being set here by the time when the waves hit the hull ($27<t<32$ s). These downstream features are fully nonlinear (amplitudes larger than $0.6~h_1$) and have similar aspects to observations made in the transcritical regime by [@bib:MelvilleHelfrich87] and reproduced numerically by [@bib:Grueetal97], although corresponding here to waves in a non-galilean frame. It is moreover surprising since as observed on the green curve in Fig. \[fig:xdxdt\_deadwater\], the Froude number does not exceed 0.7. These observations add to the comments made previously concerning the inadequacy of models based on steady object in stratified fluid. Here again we see upstream the solitary depression below the boat escaping at the bow as a single oscillation.
(10,3.5) (0.5,0)[![[*$Fr<1$*]{}. Interface evolution $\eta(x,t)$ (in m) from its rest position $h_1$ for several times during the slow down of the boat, in the frame of the boat which is represented by the black rectangular box. Same parameters as in Fig. \[fig:Fr\_o1\_reflab\].[]{data-label="fig:eauxmortes_raidissement"}](2506_10_raidissement "fig:"){width="8\unitlength"}]{} (4,-0.5)[$x_b$ (m)]{} (0,1)
For times $t\geq32$ s, one can observe that the decapitated first crest disappears below the boat, the next wave depression takes position below the boat and the initially second elevation of the packet becomes the closest to the stern. Hence, during an oscillation of the boat, the wave train slides of one wavelength compared to the boat, repeating itself. No change in the dominant wavelength of the wave train is visible.
We can now estimate the frequency of the oscillations of the boat to be $$f_b=\frac{c_g}{\lambda}\label{eq:freq_oscillations}$$ with $c_g$ the group velocity of the wave train at the rear of the boat and $\lambda$ its dominant wavelength, corresponding approximately to the distance between the first two wave crests.
The experiment presented in Fig. \[fig:Fr\_o1\_reflab\] allows an observation of the steepening process but does not present many oscillations. In order to confirm the validity of formula (\[eq:freq\_oscillations\]), we consider another experiment with $F_t=16.3$ mN in Fig. \[fig:eauxmortes\_oscillations\]. From Fig. \[fig:eauxmortes\_oscillations\] (a), one can estimate the frequency of the oscillations of the boat to be $f_b\sim0.07-0.08$ Hz. The group speed and wavelength of the wave train obtained from the diagram in Fig. \[fig:eauxmortes\_oscillations\] (b), corresponds to $c_g\sim0.035-0.040$ m s$^{-1}$ and $\lambda\sim0.5$ m. Indeed, we obtain $c_g/\lambda\sim0.07-0.08$ Hz showing very good agreement with $f_b$.
(8,6.75) (0.5,3.9)[![[*Oscillating regime*]{} ([*$Fr<1$*]{}). Panel (a) presents the speed of the boat versus time, while panels (b) and (c) shows the spatio-temporal diagrams of the interfacial displacements in the laboratory frame and in the one of the boat respectively. The black lines represent the bow and stern. Experimental parameters: $h_1=5.0$ cm, $h_2=14.0$ cm, $\rho_1=0.9980$ g cm$^{-3}$, $\rho_2=1.0157$ g cm$^{-3}$, $S_b=24$ cm$^2$, $F_t=16.3$ mN.[]{data-label="fig:eauxmortes_oscillations"}](2506_09_reprocessed_dxdt_deadwater "fig:"){width="7.5\unitlength"}]{} (4.1,6.5)[(a)]{} (0,4.75) (4,3.6)[$t$ (s)]{} (0.4,-0.25)[![[*Oscillating regime*]{} ([*$Fr<1$*]{}). Panel (a) presents the speed of the boat versus time, while panels (b) and (c) shows the spatio-temporal diagrams of the interfacial displacements in the laboratory frame and in the one of the boat respectively. The black lines represent the bow and stern. Experimental parameters: $h_1=5.0$ cm, $h_2=14.0$ cm, $\rho_1=0.9980$ g cm$^{-3}$, $\rho_2=1.0157$ g cm$^{-3}$, $S_b=24$ cm$^2$, $F_t=16.3$ mN.[]{data-label="fig:eauxmortes_oscillations"}](2506_09_reprocessed_xbt "fig:"){width="8\unitlength"}]{} (2,-0.55)[$x$ (m)]{} (6.2,-0.55)[$x_b$ (m)]{} (2.2,3.1)[(b)]{} (6.3,3.1)[(c)]{} (0,1.2) (7.2,3.25)[$\eta(x,t)$ (m)]{}
After revisiting experiments similar to the ones performed by Ekman, we have considered new situations. Indeed, dead-water phenomenon can be studied in more complex (and realistic) stratifications, when the fluid has more than two layers of different densities. We now turn to new experiments with a three-layer fluid or with a continuously stratified fluid with a pycnocline, where the complex dynamics has also been observed.
Experiments with a three-layer fluid {#threelayer}
====================================
We consider a three-layer fluid of depth $h_i$ and density $\rho_i$, $i={1,2,3}$. Two interfaces must be considered now, $\eta_{ij}(x,t)$ corresponding to the interfacial displacements between layers $i$ and $j$. A picture summarizes the setup in Fig. \[fig:eauxmortes\_3couches\_exemple\] and the main parameters are given in Table \[tab:eauxmortes\_3couches\_params\]. Only one given stratification will be discussed here.
[llccl]{} & symbols & values & units\
\
& density & $\rho_1$ & $0.9967$ & g cm$^{-3}$\
& depth & $h_1$ & $5.0$ & cm\
\
& density & $\rho_2$ & $1.0079$ & g cm$^{-3}$\
& depth & $h_2$ & $3.0$ & cm\
\
& density & $\rho_3$ & $1.0201$ & g cm$^{-3}$\
& depth & $h_3$ & $5.5$ & cm\
\
& mean density & $\bar{\rho_{12}}=\frac{\rho_2+\rho_1}{2}$ & $1.0023$ & g cm$^{-3}$\
& density jump & $\Delta\rho_{12}=\frac{\rho_2-\rho_1}{\bar{\rho_{12}}}$ & $0.0112$ &\
& maximum phase speed & $c_{\phi,12}^m=\sqrt{\Delta\rho_{12} g\frac{h_1 h_2}{h_1+h_2}}$ & $0.0147$ & m s$^{-1}$\
& Froude number & $Fr_{12}=\frac{U}{c_{\phi,12}^m}$ & $4.1-7.5$ &\
& mean density & $\bar{\rho_{23}}=\frac{\rho_3+\rho_2}{2}$ & $1.014$ & g cm$^{-3}$\
& density jump & $\Delta\rho_{23}=\frac{\rho_3-\rho_2}{\bar{\rho_{23}}}$ & $0.012$ &\
& maximum phase speed & $c_{\phi,23}^m=\sqrt{\Delta\rho_{23} g\frac{h_2 h_3}{h_2+h_3}}$ & $0.0150$ & m s$^{-1}$\
& Froude number & $Fr_{23}=\frac{U}{c_{\phi,23}^m}$ & $4.0-7.3$ &\
& maximum phase speed & $c_{s/a}^m$ & $0.073/0.044$ & m s$^{-1}$\
& Froude number & $Fr_{s/a}=\frac{U}{c_{s/a}^m}$ & $1.5/2.4 - 0.6/0.9$ &\
(18,2.5) (0.25,-0.25)[{width="15.5\unitlength"}]{} (0,0.75)[$z$]{} (1.2,-0.45)[$x$]{} (0.05,-0.4)[(1,0)[1]{}]{} (0.05,-0.4)[(0,1)[1]{}]{} (1.75,0.7)[$\rho_1$]{} (2.75,0.4)[$\rho_2$]{} (3.75,0)[$\rho_3$]{} (14.2,0.45)[(1,0)[2.1]{}]{} (16.4,0.6)[$\eta_{12}(x,t)$]{} (14.2,0.15)[(1,0)[2.1]{}]{} (16.4,0)[$\eta_{23}(x,t)$]{}
We use technique $1$ again to extract the positions of the two interfaces $\eta_{12}$ et $\eta_{23}$, along with the position of the boat.
Analytics
---------
Although the three-layer case is a simple extension of the well documented two-layer one discussed in previous section, we here reproduce explicit expressions for the phase speed of the two eigenmode solutions of this problem, which can also be find established in the most general form by [@bib:RusasGrue02] or [@bib:Xiao-Gang06] for example.
Perturbations of a three-layer fluid can be described by two types of harmonic interfacial waves, verifying the dispersion relation expressed as the following determinant $$\left| \begin{array}{cc}
\alpha_{1} & \beta \\
\beta & \alpha_{2}
\end{array} \right| = 0\,,\label{eq:det}$$ with, for $i={1,2}$ $$\begin{aligned}
\alpha_{i} &=& g(\rho_i-\rho_{i+1})+\frac{\omega^2}{k}\left(\frac{\rho_i}{\tanh (kh_i)} + \frac{\rho_{i+1}}{\tanh (kh_{i+1})} \right),\\
\beta &=& -\frac{\rho_{2}\omega^2}{k\, \sinh (kh_{2})}\,.\end{aligned}$$ The fourth order polynomial in $\omega$ associated with (\[eq:det\]) leads to two types of interfacial waves. Using the notations introduced in Table \[tab:eauxmortes\_3couches\_params\], and considering the long-wave limit ($\tanh(kh_i)\simeq\sinh(kh_i)\simeq kh_i$) along with the limit of small density jumps ($\Delta\rho_{ij}\ll\rho_i,\rho_j$ and $\rho_i\sim\bar{\rho}\,, \forall i,j$), we can express the two phase velocities associated with each solution $$\begin{aligned}
c_s^m & = & \left[g\frac{\Delta\rho_{12}h_1(h_2+h_3)+\Delta\rho_{23}h_3(h_1+h_2)+ \tilde{\Delta}}{h_1+h_2+h_3} \right]^{1/2}\,,\label{eq:eauxmortes_mode1}\\
c_a^m & = & \left[g\frac{\Delta\rho_{12}h_1(h_2+h_3)+\Delta\rho_{23}h_3(h_1+h_2)- \tilde{\Delta}}{h_1+h_2+h_3}\right]^{1/2}\,, \label{eq:eauxmortes_mode2}\end{aligned}$$ with $$\begin{aligned}
\tilde{\Delta} & = & g\left[\Delta\rho_{12}^2 h_1^2(h_2+h_3)^2+\Delta\rho_{23}^2h_3^2(h_1+h_2)^2 \right. \nonumber\\
& & \left. -2\Delta\rho_{12}\Delta\rho_{23}h_1h_2h_3(h_1+h_2+h_3-\frac{h_1h_3}{h_2})\right]^{1/2}\,.\end{aligned}$$
These two phase speeds are associated with symmetric and anti-symmetric oscillations of the interfaces, $$\begin{aligned}
\eta_s (x,t) &=& \frac{1}{2}\left( \eta_{12}(x,t) + \eta_{23}(x,t)\right)\,,\\
\eta_a (x,t) &=& \frac{1}{2}\left( \eta_{12}(x,t) - \eta_{23}(x,t)\right)\,.\end{aligned}$$ We will refer to mode-$s$ and mode-$a$ for $\eta_s$ and $\eta_a$ respectively. We also define two Froude numbers $Fr_{s/a}=v/c_{s/a}^m$ which are the appropriate non-dimensional numbers to delimitate the oscillating regime as we will see in the following. Numerical values from the experiments are summarized in Table \[tab:eauxmortes\_3couches\_params\].
(17,7) (0,0)[{width="17\unitlength"}]{} (4,-0.5)[$x$ (m)]{} (12.5,-0.5)[$x$ (m)]{} (0.25,3) (6.75,6.75)[$\eta_{s}(x,t)$ (m)]{} (15.25,6.75)[$\eta_{a}(x,t)$ (m)]{}
To emphasize that the two interfaces cannot be seen as independent oscillatory structures, since they are coupled by the intermediate layer, we give in table \[tab:eauxmortes\_3couches\_params\] the values of the Froude numbers for the two-layer equivalent solutions, $Fr_{12}$ and $Fr_{23}$. The values associated with the experiments do not allow the sub/supercritical classification.
Case $Fr_s>1$ and $Fr_a>1$
--------------------------
We first consider the case where the vessel is going faster than the fastest wave. The experimental parameters associated with the results shown in Fig. \[fig:Frs\_O1\_reflabo\] are $S_b=24$ cm$^2$ and $F_t=23.5$ mN. The limit speed of the boat is $v_\ell\simeq0.11$ m s$^{-1}$, which gives $Fr_s=1.5$ and $Fr_a=2.4$. We only present the interfacial displacement as $\eta_s$ and $\eta_a$ because the spatio-temporal diagrams for $\eta_{12}$ and $\eta_{23}$ are very similar and do not clearly exhibit the link between the waves and the boat dynamics.
Both modes are generated initially but the amplitude of mode-$s$ is more than two times larger than that of mode-$a$ which is at the limit of detection. Symmetric oscillations form a clear wave train whose front propagates at a speed of $0.0575$ m s$^{-1}$, while showing a similar spreading of the wave crests as in the two-layer case: this is an indication of dispersion effects. Speed of propagation of mode-$a$ is $0.028$ m s$^{-1}$.
(18,7) (0,0)[{width="17\unitlength"}]{} (4,-0.5)[$x$ (m)]{} (12.5,-0.5)[$x$ (m)]{} (0.25,3) (6.75,6.75)[$\eta_{s}(x,t)$ (m)]{} (15.25,6.75)[$\eta_{a}(x,t)$ (m)]{}
Case $Fr_s<1$ and $Fr_a \simeq 1$
---------------------------------
We now consider a subcritical case. It is however difficult to define clearly this domain with respect to both modes. The experimental parameters associated with the results shown in Fig. \[fig:Frs\_o1\_reflabo\] are $S_b=24$ cm$^2$ and $F_t=20.6$ mN. The mean speed of the boat is $<v>\simeq0.06$ m s$^{-1}$ which gives $Fr_s=0.6$ and $Fr_a=0.9$, with oscillations of the order of $0.04$ m s$^{-1}$.
As observed previously, both modes are generated when the boat starts. Mode-$s$ remains two times larger than mode-$a$ but they are of noticeable amplitude. The symmetric mode evolves with the boat and reproduces the amplification and steepening processes observed in the two-layer case. The mode-$s$ breaks on the boat when its amplitude is the largest, and a depression (symmetric for both interfaces) is expelled at the bow. The characteristics of the mode-$s$, wavelength $\lambda_s\simeq0.4$ m and group velocity $c_{g,s}\simeq0.06$ m s$^{-1}$ lead to an oscillating frequency of $0.15$ Hz, too slow for this visualization.
Concerning mode-$a$, its spatial structure is close to a solitary wave train, with a group velocity $c_{g,a}\simeq0.0325$ m s$^{-1}$, close to $c_a^m$ (being $0.044$ m s$^{-1}$). Each time it is generated, a strong acceleration of the boat occurs.
Discussion {#sec:discussion_3layer}
----------
The unsteady behavior associated with dead-water is still observed in the three-layer fluid, with strong analogy to the two-layer case. The stratification considered being more complex, we must consider two baroclinic modes associated with symmetric and anti-symmetric oscillations of the interfaces and that are also referred to as mode-$1$ and mode-$2$ in the literature.
[@bib:RusasGrue02] present solutions of the nonlinear equations that have strong similarities with our observations. More specifically, the spatio-temporal diagram of mode-$a$ exhibit solitary waves of mode-$2$ with oscillatory short mode-$1$ waves superimposed. This is in agreement with close values of the experiments with the numerical calculations (Boussinesq limit, $h_1/h_3\simeq1$ and $h_1/h_2\simeq1.7$).
Perturbations generated by the boat give birth to both modes, especially when the acceleration of the boat is important. This result is consistent with the study of [@bib:Nicolaou95], verified experimentally by [@bib:Robey97], stating that an accelerated object in a continuously stratified fluid, with a Brunt–Väisälä frequency $N(z)$, excites a continuum of modes whose vertical profile $w(z)$ is described by $$\frac{d^2 w}{dz^2}(z) + k_x^2\left(\frac{N^2(z)}{\omega^2}-1\right)w(z)=0\,,
\label{eq:modes}$$ along with the boundary conditions for the vertical velocity to be zero at the top and bottom.
Finally, we have observed that the mode-$1$ is strongly coupled to the dynamics of the boat, which corresponds to the fastest wave propagating in this stratification. A weak mode-$2$ is associated to noticeable acceleration of the boat, but evolves freely from the other mode.
Continuously stratified fluid with a pycnocline {#pycnocline}
===============================================
We have observed that several waves are generated when the boat evolves in a complex stratification. In the case of a linearly stratified fluid, an infinite number of modes can propagate. We have actually considered the case of a linearly stratified fluid with a pycnocline (see Fig. \[fig:eauxmortes\_linear\_profil\]). Several reasons can be invoked. It keeps the stratification undisrupted when the boat evolves in the top layer, it allows larger vertical displacements at the density jump leading to larger amplitudes, it can be modeled as a coupling between interfacial and internal waves, and it corresponds to a more realistic setup in comparison with observations made in natural environment.
Experimental parameters associated with the experiments presented are in Table \[tab:eauxmortes\_Nz\_params\]. We use technique-$2$ in order to observe both interfacial oscillations and internal waves.
[llccl]{} & symbols & values & units\
\
& density & $\rho_1$ & $0.9978$ & g cm$^{-3}$\
& depth & $h_1$ & $5.0$ & cm\
\
& density & $\rho(z)$ & $1.007-1.028$ & g cm$^{-3}$\
& depth & $h_2$ & $20.0$ & cm\
& BV frequency & $N$ & $0.9$ & rad s$^{-1}$\
\
& mean density & $\bar{\rho_{12}}=\frac{\rho_2+\rho_1}{2}$ & $1.0024$ & g cm$^{-3}$\
& density jump & $\Delta\rho_{12}=\frac{\rho_2-\rho_1}{\bar{\rho_{12}}}$ & $0.0092$ &\
& maximum phase speed & $c_{\phi,12}^m=\sqrt{\Delta\rho_{12} g\frac{h_1 h_2}{h_1+h_2}}$ & $0.06$ & m s$^{-1}$\
& Froude number & $Fr_{12}=\frac{U}{c_{\phi,12}^m}$ & $0.6 - 2.3$ &\
\
& maximum phase speed & $\frac{Nh}{\pi}$ & $0.072$ & m s$^{-1}$\
& Froude number & $Fr_h=\frac{U\pi}{Nh}$ & $0.5 - 1.9$ &\
Similarly to the three-layer case, the complex stratification considered here allows different definition for the Froude number. We know from the previous study that the one associated with the internal waves and based on the phase speed of the first vertical mode, $Fr_h=U\pi/Nh$, is the adequate one although its value is actually not so different from the one obtained from the equivalent two-layer fluid.
Modal decomposition
-------------------
As discussed in Sec. \[sec:discussion\_3layer\], we expect the non-stationary evolution of the boat to generate several modes. We anticipate here by presenting the expected mode structures associated with the stratification (Fig. \[fig:eauxmortes\_linear\_profil\]) and obtained from solving Eq. (\[eq:modes\]). From the vertical structure of the vertical component of the velocity obtained from (\[eq:modes\]), we can deduce the corresponding density gradients, since $$\rho^\prime_n(z)=\mathi\rho^\star \frac{N(z)^2}{\omega}w_n(z)\,,$$ then $$\partial_x\rho^\prime_n(z)=-\rho^\star \frac{k_x N(z)^2}{\omega}w_n(z)\,,$$ where $\rho^\star$ is a reference density $\omega$ the frequency of the mode and $k_x$ its horizontal wavenumber.
Figure \[fig:eauxmortes\_Nz\_modes\] gives the vertical profiles (of the horizontal density gradient) of the first three modes $\rho_n$ for two specific frequencies, $\omega_1=0.25$ rad s$^{-1}$ in (a) and $\omega_2=1.1$ rad s$^{-1}$ in (b). The former corresponds to a mode propagating in the total depth of the fluid (since $\omega_1<N$) whereas the latter is trapped in the pycnocline. ($\omega_2\in[0.9;1.8]$ rad s$^{-1}$).
(10,3.25) (0.35,0)[![Normalized horizontal density gradient of the first three vertical modes of stratification in Fig. \[fig:eauxmortes\_linear\_profil\] with (a) $\omega_1=0.25$ rad s$^{-1}$ and (b) $\omega_2=1.1$ rad s$^{-1}$.[]{data-label="fig:eauxmortes_Nz_modes"}](strat2_density_profile_13_density_modes1234_39mHz "fig:"){width="4\unitlength"}]{} (4.45,0)[![Normalized horizontal density gradient of the first three vertical modes of stratification in Fig. \[fig:eauxmortes\_linear\_profil\] with (a) $\omega_1=0.25$ rad s$^{-1}$ and (b) $\omega_2=1.1$ rad s$^{-1}$.[]{data-label="fig:eauxmortes_Nz_modes"}](strat2_density_profile_13_density_modes1234_175mHz "fig:"){width="4\unitlength"}]{} (2.4,3.15)[(a)]{} (6.5,3.15)[(b)]{} (0,1.3) (2.2,-0.5)[$\partial_x\rho^\prime_n$]{} (6.3,-0.5)[$\partial_x\rho^\prime_n$]{}
The modal decomposition depends on the frequency of study. In our case, the dead-water phenomenon is a non-periodic state and we expect perturbations that are not harmonic modes. However, from previous observations, it is reasonable to expect nonlinear evolution of modes evolving freely from the others. Based on this assumption, we expect the following projection to catch the key features of the dynamics.
The computed modes in Fig. \[fig:eauxmortes\_Nz\_modes\] will help us identify dominant structures associated with specific frequency bands, but it is important not to apply too selective time filtering to the data in order to conserve information about their complex dynamics. Thus we extract the mode amplitudes $\hat{a}_n(x,t)$ after the two following steps done at each $x$-location:
1. a large band filtering centered around $\omega$ with a bandwidth $\Delta\omega$,
2. a projection of the vertical data onto the selected basis associated with the frequency band chosen $\omega$ $$\hat{a}_n(x,t)=\frac{1}{H}\int_{-H}^{0} \partial_x\rho^\prime_n(z).\partial_x\rho^\prime(x,z,t)\,dz \,.$$
Case $Fr_h>1$
-------------
The first experiment considered corresponds to the experimental parameters in Table \[tab:eauxmortes\_Nz\_params\] with $S_b=12$ cm$^2$ and $F_t=21.3$ mN. The limit speed reached by the boat is $v_\ell=0.14$ m s$^{-1}$ which corresponds to a supercritical case ($Fr_h=1.9$).
Figure \[fig:eauxmortes\_Frh\_01\_drho\] presents time series of the density gradient extracted on a vertical cut at $x_0=0.05$ m where the origin is taken at the bow of the boat at $t=0$. Interfacial displacements obtained from the evolution of the maximum of the vertical density gradient are given in white and are of less than $0.01$ m. The fast evolution of the boat perturbs weakly the stratification and the signal associated with the internal waves field, for $z\in[-0.25,-0.05]$ m, is weak.
(10,7.5) (0.4,3.75)[![[*$Fr_h>1$*]{}. Time series of (a) $\partial_z\rho(x_0,z,t)$ and (b) $\partial_x\rho(x_0,z,t)$ extracted at $x_0=0.5$ m, in g/cm$^{4}$. []{data-label="fig:eauxmortes_Frh_01_drho"}](20100219_strat2_exp13_6tr_x22_5_drhoz_kg "fig:"){width="8\unitlength"}]{} (0.4,0)[![[*$Fr_h>1$*]{}. Time series of (a) $\partial_z\rho(x_0,z,t)$ and (b) $\partial_x\rho(x_0,z,t)$ extracted at $x_0=0.5$ m, in g/cm$^{4}$. []{data-label="fig:eauxmortes_Frh_01_drho"}](20100219_strat2_exp13_6tr_x22_5_drhox_kg "fig:"){width="8\unitlength"}]{} (0.5,3.5)[(b)]{} (0.5,7.25)[(a)]{} (4,-0.5)[$t$ (s)]{} (0,5) (0,1.25)
Two types of internal waves can still be discriminated. Oscillations of the full water depth are triggered when the boat passes by at $t\simeq6$ s. Later on, some turbulent fluctuations in the pycnocline lead to small scales forcing radiating waves downwards ($t\in[10;20]$ s). The modal projection leads to no clear signal whatever the frequency considered and are not given here.
Case $Fr_h<1$
-------------
The same configuration is used but with a force $F_t=11.0$ mN. The oscillatory dynamics of the boat around a mean speed $<v>\simeq0.036$ m s$^{-1}$ corresponds to a subcritical case with $Fr_h=0.5$.
We represent the time series extracted at the location, $x=0.67$ m (Fig. \[fig:eauxmortes\_Frh\_o1\_drho\_x40\]). The vertical cut corresponds to a region close to where the boat will stop and its speed is already decreasing. The stopping point of the boat is located at $x\simeq1.5$ m and occurs at $t\simeq50$ s.
(10,7.5) (0.4,3.75)[![[*$Fr_h<1$*]{}. Time series of (a) $\partial_z\rho(x,z,t)$ and (b) $\partial_x\rho(x,z,t)$ extracted at $x=0.67$ m, in g/cm$^{4}$.[]{data-label="fig:eauxmortes_Frh_o1_drho_x40"}](20100219_strat2_exp11_3tr_x40_drhoz_kg "fig:"){width="8\unitlength"}]{} (0.4,0)[![[*$Fr_h<1$*]{}. Time series of (a) $\partial_z\rho(x,z,t)$ and (b) $\partial_x\rho(x,z,t)$ extracted at $x=0.67$ m, in g/cm$^{4}$.[]{data-label="fig:eauxmortes_Frh_o1_drho_x40"}](20100219_strat2_exp11_3tr_x40_drhox_kg "fig:"){width="8\unitlength"}]{} (0.5,3.5)[(b)]{} (0.5,7.25)[(a)]{} (4,-0.5)[$t$ (s)]{} (0,5) (0,1.25)
In Fig. \[fig:eauxmortes\_Frh\_o1\_drho\_x40\], the displacements of the pycnocline is of the order of $0.01$ m as in the supercritical case. Nevertheless, the internal waves field is much more intense and the amplitude of the waves generated are larger as the boat evolves, suggesting the same nonlinear evolution of the waves as for interfacial waves.
(10,2.75) (0.3,0)[![[*$Fr_h<1$*]{}. Instantaneous horizontal density gradients $\partial_x\rho(x,z,t_0)$ at times (a) $t_0=32$ s and (b) $t_0=61$ s, in g/cm$^{4}$.[]{data-label="fig:eauxmortes_Frh_o1_dxrho"}](20100219_strat2_exp11_3trdxrho_im253_kg "fig:"){height="2.39\unitlength"}]{} (4.5,0)[![[*$Fr_h<1$*]{}. Instantaneous horizontal density gradients $\partial_x\rho(x,z,t_0)$ at times (a) $t_0=32$ s and (b) $t_0=61$ s, in g/cm$^{4}$.[]{data-label="fig:eauxmortes_Frh_o1_dxrho"}](20100219_strat2_exp11_3trdxrho_im430_kg "fig:"){height="2.47\unitlength"}]{} (2.2,2.5)[(a)]{} (6.4,2.5)[(b)]{} (2,-0.5)[$x$ (m)]{} (6.2,-0.5)[$x$ (m)]{} (0,1)
The spatial structure seems to be dominated by a mode-$1$ shape close to the boat, which is even more obvious in instantaneous visualizations of the density gradients as shown in Fig. \[fig:eauxmortes\_Frh\_o1\_dxrho\] (a). Other modes are also present in the wave field but appear later in time and are hard to discriminate from the radiated waves generated by turbulent patches at the pycnocline (Fig. \[fig:eauxmortes\_Frh\_o1\_dxrho\] (b)).
In order to quantify in more detail the amplitude of each generated mode, we project the time series of the vertical structures extracted at different $x$-locations on the modal basis described before and computed at five different frequencies, $\omega=[0.25, 0.50, 0.75, 1.0, 1.25]$ rad s$^{-1}$. Large band filtering in time is done, with $\Delta\omega=0.25$ rad s$^{-1}$, so as to keep track of the nonlinear evolution of the amplitude of the modes with time. One can observe in Fig. \[fig:Frh\_o1\_modes\] (a) that the amplitude of the mode-$1$ structure is indeed dominant since it is emitted for all frequencies considered, whereas the mode-$2$ in Fig. \[fig:Frh\_o1\_modes\] (b) is of comparable amplitude at the lowest frequencies ($\omega=0.25$ and $0.50$ rad s$^{-1}$) but is absent for larger values of $\omega$. Higher modes ($n\geq3$) are even weaker in amplitudes and only present at low frequencies (Fig. \[fig:Frh\_o1\_modes\] (c)).
We notice that the initial perturbations for these low modes are generated below the boat since the black rectangles in Fig. \[fig:Frh\_o1\_modes\] (a) and (b) are associated with a ramping amplitude of modes $1$ and $2$. The longer the rectangle is, the smaller the speed of the boat is, leading to larger initial perturbations which are very similar to what has been observed in the two-layer case (see Fig. \[fig:eauxmortes\_raidissement\]).
(10,14.5) (0,9.5)[![[*$Fr_h<1$*]{}. Amplitudes $\hat{a}_n(x_0,t)$ with time of the projected modal structure at several $x_0$-locations for modes $n=1$ (a), $2$ (b) and $3$ (c) of the horizontal density gradient in arbitrary units. Different colors correspond to different frequencies associated with the modal basis. The tilted dashed lines corresponds to the characteristic speed of propagation $c_\phi^m=7.2$ cm/s. The black rectangle represents the boat passing through a vertical cross-section at each $x_0$-location studied.[]{data-label="fig:Frh_o1_modes"}](20100219_strat2_exp11_3tr_mode1_5om "fig:"){width="8.5\unitlength"}]{} (4.5,14.25)[(a)]{} (0,4.5)[![[*$Fr_h<1$*]{}. Amplitudes $\hat{a}_n(x_0,t)$ with time of the projected modal structure at several $x_0$-locations for modes $n=1$ (a), $2$ (b) and $3$ (c) of the horizontal density gradient in arbitrary units. Different colors correspond to different frequencies associated with the modal basis. The tilted dashed lines corresponds to the characteristic speed of propagation $c_\phi^m=7.2$ cm/s. The black rectangle represents the boat passing through a vertical cross-section at each $x_0$-location studied.[]{data-label="fig:Frh_o1_modes"}](20100219_strat2_exp11_3tr_mode2_5om "fig:"){width="8.5\unitlength"}]{} (4.5,9.25)[(b)]{} (0,-0.5)[![[*$Fr_h<1$*]{}. Amplitudes $\hat{a}_n(x_0,t)$ with time of the projected modal structure at several $x_0$-locations for modes $n=1$ (a), $2$ (b) and $3$ (c) of the horizontal density gradient in arbitrary units. Different colors correspond to different frequencies associated with the modal basis. The tilted dashed lines corresponds to the characteristic speed of propagation $c_\phi^m=7.2$ cm/s. The black rectangle represents the boat passing through a vertical cross-section at each $x_0$-location studied.[]{data-label="fig:Frh_o1_modes"}](20100219_strat2_exp11_3tr_mode3_5om "fig:"){width="8.5\unitlength"}]{} (4.5,4.25)[(c)]{}
Furthermore, the apparently cnoïdal (sharp crests and flat troughs) oscillations associated with the mode-$1$ internal waves observed in Fig. \[fig:Frh\_o1\_modes\] (a) as the modes propagate, are a signature of the nonlinear dynamics of the internal waves. Finally, by giving in all images the maximum phase speed of the waves with the tilted lines in Fig. \[fig:Frh\_o1\_modes\], we can verify that the waves crests propagate at a smaller speed than $7.2$ cm s$^{-1}$ but it is difficult to extract a constant speed with propagation for each mode.
Discussion {#discussion}
----------
The dead-water phenomenon has been observed in the more general case of a linearly stratified fluid with a pycnocline. The complex dynamics of the boat is coupled to the first mode of the stratification with frequencies below the maximum value of $N(z)$.
Although the wave field associated with the subcritical regime is analogous to the previous layered stratifications considered, the supercritical regime is different in nature since it consists mainly of radiated waves from turbulent perturbations in the pycnocline. This is similar to observations of the radiated wave field associated with an object moving at constant speed in a stratified fluid [@bib:RottmanetalAPS2004].
\[sct:conclusion\]
By revisiting the historical experiments of Ekman’s PhD Thesis and extending it to more general stratifications, we have shown the robustness of the dead-water phenomenon. The experimental techniques used have revealed new insights on the century-old problem.
One important characteristic of the dead-water phenomenon is that the dynamics of the boat is coupled to the fastest mode of the stratification considered, although several modes are generated at each acceleration of the boat. Furthermore, the nonlinear features of the phenomenon must be considered in order to describe analytically its unsteady nature. Classical models such as presented by @bib:Ekman1904 or @bib:Milohetal93 are not sufficient.
It would be of great interest to provide an analytical description of the coupled dynamics of the boat and the waves it generates in order to take into account unsteady behaviors.
An extended study of the problem in a three-dimensional setup is in preparation but we are still developing a collaboration with Playmobil©.
The authors are very thankful to Leo Maas for introducing the topic of dead-water to them, for many discussions and for his great knowledge of the historical background. TD thanks Peter Morgan and Christina Kraus for mentioning similar reports in the Latin literature.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The best way to search for new physics is by using a diverse set of probes - not just experiments at the energy and the cosmic frontiers, but also the low-energy measurements relying on high precision and high luminosity. One example of such ultra-precision experiments is the MOLLER experiment planned at JLab, which will measure the parity-violating electron-electron scattering asymmetry and allow a determination of the weak mixing angle with a factor of five improvement in precision over its predecessor, E-158. At this precision, any inconsistency with the Standard Model should signal new physics. The paper will explore how new physics particles enter at the next-to-leading order (one-loop) level. For MOLLER we analyze the effects of dark Z’-boson on the total calculated asymmetry, and show how this new physics interaction carriers may influence the analysis of the future experimental results.'
author:
- 'A. Aleksejevs'
- 'S. Barkanova'
- 'S. Wu'
- 'V. Zykunov'
title: 'New Physics Search with Precision Experiments: Theory Input '
---
Precision Parity Violating Physics
==================================
The fact of existence of the Dark Matter [@Zwicky-Dark-Matter] is one of the most striking evidences that the Standard Model (SM) is incomplete. The further investigation into possible extensions of SM with new physics particles became one of the main goal of both theoretical and experimental particle physics. Searches for physics beyond SM can be summarized into three major directions: energy, cosmic and precision frontiers. The energy frontier is concentrated on the direct production of the new physics particles, which might be accessible at high-energy colliders. In case of the cosmic frontier, direct searches for new physics are coming from underground experiments, ground and space telescopes. The precision frontier is driven by the indirect searches, where new physics particles could impact various observables in SM and hence cause small deviations from original SM predictions. This can be studied by using very precise measurements with intense particle beams. In this paper, we address one of the specific processes used at precision frontier, namely a test of SM using the parity-violating Møller ($e+e\rightarrow e+e$) scattering. The most recent parity-violating Møller scattering experiment, E-158 [@E158], measured parity-violating right-left asymmetry defined as $$\begin{gathered}
{\displaystyle A_{PV}=\frac{\sigma_{R}-\sigma_{L}}{\sigma_{R}+\sigma_{L}}},\label{eq:1a}\end{gathered}$$ and reported the value of $A_{PV}=(-131\pm14\text{\ensuremath{\pm}}10)\cdot10^{-9}$, which is resulted in the effective weak mixing angle of $\sin^{2}\theta_{W}^{eff}(Q^{2}=0.026\, GeV^{2})=0.2397\pm0.0010\pm0.0008.$ The reported result is found to be consistent with the SM predictions (in the $\overline{MS}$ scheme): $\sin^{2}\theta_{W}^{\overline{MS}}(Q^{2}=0.026\, GeV^{2})=0.2381\pm0.0006$ [@MSbar-Czarnecki; @and; @Marciano; @PDG2004]. In order to put more stringent bounds on the parity violating tests of SM, the MOLLER experiment planned at the Thomas Jefferson National Accelerator Facility (Jefferson Lab for short, or JLab) [@MOLLER], will measure $A_{PV}(Q^{2}=0.0056\, GeV^{2})$ at the level of the $\delta(A_{PV})=0.75$ ppb, which translates to the factor of five improvement in precision for the measurement of the effective mixing angle compared to the E-158 experiment. At this level of precision, the new physics signal may be experimentally detectable, so it is essential to study the potential impact of the new-physics degrees of freedom on the parity-violating cross section asymmetry in the Møller scattering.
Dark Photon and Z Bosons
========================
In our analysis we choose the simplest extension of SM by the additional $U(1)'$ symmetry proposed in [@Holdom].
Here, the mixing of $B_{\mu}(U(1)_{Y})$ and $A'_{\mu}(U(1)')$ fields is expressed through the kinetic mixing Lagrangian (see Fig.\[fig1\]): $$\begin{aligned}
\mathfrak{L}_{kin}=-\frac{1}{4}B_{\mu\nu}B^{\mu\nu}+\frac{1}{2}\frac{\epsilon}{\cos\theta_{W}}B_{\mu\nu}A'^{\mu\nu}-\frac{1}{4}A'_{\mu\nu}A'^{\mu\nu},\label{eq:1}\end{aligned}$$ where $B_{\mu\nu}=\partial_{\mu}B_{\nu}-\partial_{\nu}B_{\mu}$, $B_{\mu}=\cos\theta_{W}A_{\mu}-\sin\theta_{W}Z_{\mu}$ and $\epsilon$ is the $(B_{\mu}-A'_{\mu})$ mixing parameter. If we assume the simplest scenario for the Higgs sector, which is the SM Higgs doublet plus the Higgs singlet (used for breaking the $U(1)'$ symmetry and giving mass to $A'_{\mu}$), a Lagrangian describing interaction between the SM fermions and the dark vector boson $A'_{\mu}$, photon $V_{\mu}$ and weak $Z_{\mu}$ fields has the following form:
\_[int]{}= & -eQ\_[f]{}|[f]{}\_f(V\^+A’\^)-\
\
& |[f]{}(c\_[V]{}\^[f]{}\_+c\_[A]{}\^[f]{}\_\_[5]{})fZ\^.\[eq:2\]
Here, $Q_{f}$ is the charge of the fermion in units of $e$. Vector and axial-vector coupling strengths are defined as follows:
c\_[V]{}\^[f]{} & =T\_[3f]{}-Q\_[f]{}\^[2]{}\_[W]{}\
c\_[A]{}\^[f]{} & =-T\_[3f]{},\[eq:3\]
with $T_{3f}$ defined as fermion’s third component of the weak isospin. The Lagrangian in Eq.\[eq:2\] has only vector-type coupling of dark $A'_{\mu}$ to fermions, which is coming from the non-zero kinetic mixing of $V_{\mu}$ and $A'_{\mu}$ fields. At the leading order, the kinetic mixing term between $Z_{\mu}$ and $A'_{\mu}$ fields cancels out with their mass mixing term, so as a result $A'_{\mu}$ does not have the axial-vector type of coupling to fermions in Eq.\[eq:2\]. Hence, $A'_{\mu}$ is called a dark photon $V'_{\mu}$ $(A'_{\mu}\equiv V'_{\mu})$, which resembles a massive photon with the coupling weighted by the mixing parameter $\epsilon$:
\_\^[|[f]{}-V’-f]{}=-i eQ\_[f]{}\_.\[eq:4\]
A possible extension with non-vanishing mixing between dark $A'_{\mu}$ and weak $Z_{\mu}$ was explored in [@DLM-Dark-Z] with an additional mass mixing term described by the mixing parameter $\epsilon_{Z'}=\frac{m_{z'}}{m_{z}}\delta$. Here, $m_{Z'}$ is the mass of the dark $Z_{\mu}'$ boson and $\delta$ is an arbitrary model-dependent parameter. In this scenario, the interaction Lagrangian is given by
\_[int]{}= & -eQ\_[f]{}|[f]{}\_f(V\^+A’\_)-\
\
& |[f]{}(c\_[V]{}\^[f]{}\_+c\_[A]{}\^[f]{}\_\_[5]{})f(Z\^+\_[Z’]{}A’\_),\[eq:5\]
and, as we can see from above, the dark $A'_{\mu}$ couples to fermions through both vector and axial-vector interactions, which is similar to the weak $Z_{\mu}$ coupling. Hence, that type of the dark $A'_{\mu}$ in [@DLM-Dark-Z] is called the dark $Z_{\mu}'$ boson ($A'_{\mu}\equiv Z'_{\mu}$). As a result, the coupling $\bar{f}-Z'_{\mu}-f$ is written in the following form:
\_\^[|[f]{}-Z’-f]{}= & -ie(S’\_[V]{}\_+S’\_[A]{}\_\_[5]{}),\
\
S’\_[V]{}= & Q\_[f]{}+,\
\
S’\_[A]{}= & .\[eq:6\]
In the case when $\epsilon_{Z'}$ goes to zero, the dark $Z'_{\mu}$ becomes the dark photon $V'_{\mu}$. The coupling in Eq.\[eq:6\] is parity-violating by its nature. In our analysis we use left/right handed (chiral) notation which reflects the nature of the parity-violating interaction by the simple condition of $g_{L}\ne g_{R}$. Accordingly, in the chiral basis,
\_\^[|[f]{}-Z’-f]{}= & -ie(S’\_[L]{}g\_[L]{}\_\_[-]{}+S’\_[R]{}g\_[R]{}\_\_[+]{}),\[eq:7\]
where $\omega_{\pm}=\frac{1\pm\gamma_{5}}{2}$ are chirality projectors, and $g_{\{R,L\}}=c_{V}^{f}\pm c_{A}^{f}$ are the usual SM right- and left-handed coupling strengths. The scaling parameters $S'_{\{L,R\}}$ can now be expressed the through mixing parameters as:
S’\_[L]{}= & (Q\_[f]{}+)\
\
S’\_[R]{}= & (Q\_[f]{}+),\[eq:8\]
and the condition for the dark $Z'_{\mu}$ becoming the dark photon $V'_{\mu}$ is given by $S'_{R}g_{R}=S'_{L}g_{L}$, which is satisfied if either $\delta\rightarrow0$ or $m_{Z'}\ll m_{Z}$. Also, if $S'_{R}=S'_{L}=S'$, dark $Z'_{\mu}$ boson becomes the “usual” SM weak $Z_{\mu}$ boson with modified mass and scaled coupling by $\epsilon_{Z'}=\frac{m_{Z'}}{m_{Z}}\delta$. The condition $S'_{R}=S'_{L}=S'$ is satisfied if $\epsilon\rightarrow0$.
In this work, we have evaluated the parity-violating asymmetry up to one-loop level with the dark photon or dark $Z'_{\mu}$ appearing at the tree level and in the box, vertex, and self-energy diagrams. Representative diagrams for one loop are shown in Fig.\[fig2\].
![Representative one-loop diagrams for the Møller process with the new-physics (labeled as NP) vector boson in the loops. The label SM stands for the Standard Model vector bosons. In the actual calculations, the diagrams with vertex corrections to the lower electron current and the diagrams for the u-channel are taken into account as well. We also include the gauge fixing terms in the diagrams with $W^{\pm}$ in the vertex and self-energy graphs (not shown here).[]{data-label="fig2"}](graphs)
The diagrams shown in Fig.\[fig2\] do not contain the Higgs boson because we do not include the coupling of dark vector $A'_{\mu}$ to the Higgs field, assuming that the diagrams with the Higgs boson would give a small contribution to the asymmetry. However, for the sake of completeness, we plan to include this interaction in our next work. Using on-shell renormalization scheme for SM and NP fields we have calculated PV asymmetry up to one loop level and included soft-photon bremsstrahlung when treating infrared divergences. For the SM parameters we used last-year PDG values. For the cut on energy of the soft-photons, we choose $\Delta E=0.05\, E_{cms}$ with $E_{cms}=0.106\,\text{GeV}$.
Results and Conclusion
======================
Our calculation strategy basically consist of the following steps. First, we evaluate the PV asymmetry including one-loop diagrams for the SM particles. This will determine the SM central value. Then we proceed with calculations of the PV asymmetry with the new-physics particles included up to one-loop and construct exclusion plots for 1%, 2% and 3% deviations from the SM central value. Since the MOLLER experiment is mostly sensitive to the parity-violating interaction, which is enhanced through the interference term $\sim2\text{Re[}M_{\gamma}M_{Z}]$ in the numerator of Eq.\[eq:1a\], we concentrate our attention on the analysis of dark $Z'_{\mu}$. The exclusion plots for MOLLER for the case of new physics represented by dark $Z'_{\mu}$ are shown in Fig.\[fig3\].
![Exclusion plots for the dark $Z'_{\mu}$ for the MOLLER experiment with calculations including one-loop in the on-shell renormalization scheme, shown against exclusion plot from [@New-Physics-Plot-reference]. We use $\delta^{2}=3\cdot10^{-5}$. The blue dot-dashed, green dashed and red dotted graphs correspond to 1%, 2% and 3% the PV asymmetry deviations from the SM prediction, respectively. []{data-label="fig3"}](DarkZ_D)
In the case if the MOLLER experiment does not detect any significant deviations from the SM predictions, then this measurement will exclude everything that is above the corresponding 1%, 2% or 3% lines. Essentially, if MOLLER does not see the dark $Z'_{\mu}$, it will exclude the entire region which would explain the $g-2$ anomaly with the light $Z'_{\mu}$ dark boson. A larger value of the $\delta$ mixing parameter would increase the measurement sensitivity to $Z'_{\mu}$ and push the exclusion lines down. Clearly, as one can see from on Fig.\[fig3\], the MOLLER experiment is very competitive with the DarkLight [@DarkLight], APEX [@APEX], MAMI [@MAMI], KLOE [@KLOE] and HPS [@HPS].
Fig.\[fig4\] shows the exclusion regions for the fixed masses of $Z'_{\mu}$ in the space of $\epsilon$ and $\delta$ mixing parameters.
In the region of the small $Z'_{\mu}$ mass (left plot on Fig.\[fig4\]), the overall sensitivity to the variation of $\epsilon$ and $\delta$ is quite high but decreases significantly in the region of the higher mass of $Z'_{\mu}$ (middle plot of Fig.\[fig4\]). That is mostly related to the suppression coming from the dark $Z'_{\mu}$ propagator. If we assume the scenario of the heavy $Z'_{\mu}$, we observe that the sensitivity to $\epsilon$ and $\delta$ is enhanced at the leading order by the term $\sim\frac{\delta}{m_{Z}^{2}}$ and loop contribution from $Z'_{\mu}$. A detailed analysis of the one-loop contributions of the dark vector to the PV asymmetry will be addressed in our next work. In the limit when $\delta\rightarrow0$ (the dark photon), the sensitivity is weak for all masses of $Z'_{\mu}$. Thus, it is important to have a non-zero (although possibly small) mixing parameter $\delta$ when it comes to the low-momentum transfer PV experiments such as MOLLER. In the case of $\epsilon\rightarrow0$ (the “usual” $Z_{\mu}$ boson with the modified mass and scaled coupling), we also observe the reduced sensitivity for the lower masses of $Z'_{\mu}$, so $\epsilon$ should be non-zero in order to satisfy the constrain $|\delta|<1$ (see [@DLM-Dark-Z]). For the higher mass of $Z'_{\mu}$ (right plot of Fig.\[fig4\]) and the limit when $\epsilon\rightarrow0$, if no significant discrepancy between the measurement and the SM prediction is found, we will be able to say that $\delta^{2}\lesssim5\cdot10^{-6}$. As we can see, for the low-energy frontier, the probability of finding physics beyond the SM is primarily determined by the level of experimental precision. Therefore advancing that type of experiments in the precision domain could actually open a link to our understanding of the nature of Dark Matter.
ACKNOWLEDGMENTS
===============
This work has been supported by the Natural Sciences and Engineering Research Council of Canada (NSERC). We are grateful to W. Marciano and J. Erler for the useful discussions and encouragement during the MITP workshop on “Low-energy precision physics” in Mainz in 2013. Also, many thanks to our undergraduate student research assistants M. Bluteau, C. Griebler and J. Strickland for testing the first versions of the code in the summer of 2013. AA and SB thank JLab Theory Group for hospitality during their stay in 2014.
[10]{} F. Zwicky, Helv. Phys. Acta 6:11027 (1933).
P. L. Anthony et al., (SLAC E-158 Collaboration), Phys. Rev. Lett. 95 081601 (2005).
A. Czarnecki and W. J. Marciano, Phys. Rev. D 53, 1066 (1996).
J. Benesch et al., The MOLLER Experiment at http://http://hallaweb.jlab.org/12GeV/Moller/, (2014).
S. Eidelman et al., (Particle Data Group), Phys. Lett. B 592, 1 (2004).
B. Holdom, Phys. Lett. B166, 196 (1986).
H. Davoudiasl, H. Lee, W. Marciano, arXiv:1203.2947v2, Phys. Rev. D 85, 115019 (2012).
R. McKeown, arXiv:1109.4855v2 (2011).
J. Balweski et al., “Dark Light Proposal” at http://dmtpc.mit.edu/DarkLight/, (2012).
S. Abrahamyan et al., Phys. Rev. Lett. 107, 191804 (2011).
H. Merkel et al, arXiv:1101.4091v2, (2011).
P. Franzini, M. Moulson, Ann.Rev.Nucl.Part.Sci.56:207-251 (2006).
A. Grillo et al., “HPS: Heavy Photon Search Proposal” at https://www.jlab.org/exp\_prog/proposals/11/PR12-11-006.pdf, (2010).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We consider a modification to the standard cosmological history consisting of introducing a new species $\phi$ whose energy density red-shifts with the scale factor $a$ like $\rho_\phi \propto a^{-(4+n)}$. For $n>0$, such a red-shift is faster than radiation, hence the new species dominates the energy budget of the universe at early times while it is completely negligible at late times. If equality with the radiation energy density is achieved at low enough temperatures, dark matter can be produced as a thermal relic during the new cosmological phase. Dark matter freeze-out then occurs at higher temperatures compared to the standard case, implying that reproducing the observed abundance requires significantly larger annihilation rates. Here, we point out a completely new phenomenon, which we refer to as [*relentless*]{} dark matter: for large enough $n$, unlike the standard case where annihilation ends shortly after the departure from thermal equilibrium, dark matter particles keep annihilating long after leaving chemical equilibrium, with a significant depletion of the final relic abundance. Relentless annihilation occurs for $n \geq 2$ and $n \geq 4$ for $s$-wave and $p$-wave annihilation, respectively, and it thus occurs in well motivated scenarios such as a quintessence with a kination phase. We discuss a few microscopic realizations for the new cosmological component and highlight the phenomenological consequences of our calculations for dark matter searches.'
author:
- 'Francesco D’Eramo'
- Nicolas Fernandez
- Stefano Profumo
bibliography:
- 'EFCpaper1.bib'
date: 'March 13, 2017'
title: 'When the Universe Expands Too Fast: Relentless Dark Matter'
---
Introduction
============
Decades after the first observational evidences, the origin and composition of the dark matter (DM) is still among the most urgent open questions in particle physics [@Jungman:1995df; @Bertone:2004pz; @Feng:2010gw]. Weakly Interacting Massive Particles (WIMPs) are motivated particle candidates for DM, with a thermal relic abundance naturally close to the observed cosmological abundance of DM. A standard calculation [@Lee:1977ua; @Scherrer:1985zt; @Srednicki:1988ce] shows that this thermal relic WIMP abundance scales as the inverse annihilation cross section, and is mildly dependent on the particle mass. The observed DM abundance is reproduced for \_[th]{} v\_[rel]{} 3 10\^[-26]{} \^3 \^[-1]{} , \[eq:sigmavrelthermal\] where the brackets denote a thermal average and $v_{\rm rel}$ is the Møller velocity (for details see Ref. [@Gondolo:1990dk]). The cross section needed for a thermal relic is thus that typical of weak interactions. This phenomenal coincidence, combined with the expectation of new degrees of freedom at the weak scale for independent reasons such as the hierarchy problem, is referred to as the “WIMP miracle”.
The numerical value in [Eq. ]{} has been an important benchmark for WIMP searches. It is worth keeping in mind that it relies on a crucial assumption: at the time of DM genesis, the energy budget of the universe was dominated by its radiation content. We know from Big Bang Nucleosynthesis (BBN) that this is definitely the case at temperatures around and below $T_{\rm BBN} \simeq \text{few} \; {\rm MeV}$ [@Kawasaki:2000en; @Ichikawa:2005vw]. However, we have no direct information about the energy budget of the universe at higher temperatures. The WIMP DM thermal relic abundance may differ by orders of magnitude if deviations from a standard cosmological history are considered [@McDonald:1989jd; @Kamionkowski:1990ni; @Chung:1998rq; @Giudice:2000ex; @Moroi:1999zb; @Allahverdi:2002nb; @Allahverdi:2002pu; @Acharya:2009zt; @Monteux:2015qqa; @Co:2015pka; @Davoudiasl:2015vba; @Co:2016vsi; @Co:2016fln].
In this work we consider DM genesis for a broad class of alternative cosmological histories. We assume the presence of another species $\phi$, whose energy density red-shifts with the scale factor $a$ as follows \_a\^[-(4+n)]{} , n > 0 . \[eq:phiScaling\] The standard case of radiation follows the behavior above for $n=0$. Here, we always consider $n > 0$, which implies that the $\phi$ energy density dominates over radiation at early enough times. The equality between the energy density of $\phi$ and radiation must happen at a temperature $T_r \gtrsim T_{\rm BBN}$ (we will be more quantitative about this point in Sec. \[sec:BBN\]). If such an equality is achieved after the time of DM production, the standard relic calculation is significantly affected, as is the thermal relic abundance of the DM. We survey the options for DM genesis when the universe is dominated by a fluid red-shifting as in [Eq. ]{} in Sec. \[sec:FO\].
The two-dimensional parameter space $(T_r, n)$ fully describes the possible cosmological backgrounds in our setup. The two parameters cannot be arbitrary, since for low enough $T_r$ we must ensure not to spoil the success of BBN. This set of cosmological backgrounds are described in Sec. \[sec:background\], where we provide an expression for the Hubble parameter as a function of the radiation bath temperature $T$. For each temperature value $T > T_r$, the Hubble parameter is always larger than what it would be for a standard cosmological history. For this reason, the universe expands [*faster*]{} than in the standard case when dominated by $\phi$.
A significant energy density of $\phi$ around the time of BBN mimics the role of additional neutrino species. Light element abundances put bounds on $N_\nu$ [@Cyburt:2015mya], which can be used to exclude part of the $(T_r, n)$ plane. We discuss these bounds in Sec. \[sec:BBN\]. Interestingly, the energy density of $\phi$ is completely subdominant at the time of the decoupling of the Cosmic Microwave Background (CMB). The number of effective neutrinos at $T_{\rm CMB} \simeq 1 \, {\rm eV}$ is also constrained [@Ade:2015xua], but our framework does not predict any deviation from the SM value. This work focuses on freeze-out DM production. We assume the DM particles to achieve thermal equilibrium with the primordial plasma at high temperature, and decouple once the temperature drops below its mass. The faster expansion rate raises however an important question: does the DM [*ever*]{}= thermalize? This is inspected in App. \[app:DMthermalization\], where we quantify the conditions we need in order to have the DM in thermal equilibrium at early times. The answer to this question sets the stage for the DM relic density calculations in Sec. \[sec:FO\]. These calculations are performed by parameterizing the new cosmological phase by $(n, T_r)$, without specifying the microscopic origin of the new species $\phi$. At large enough $n$, we find a very peculiar behavior for the DM number density evolution, that had never been recognized before: The different Hubble scaling with the temperature allows significant DM annihilations long after the decoupling from the thermal bath. For a DM annihilating through an $s$-($p$-)wave process, this happens for $n \geq 2 (4)$. Remarkably, the red-shift with $n = 2$ arises from motivated theories of quintessence attempting to explain the current acceleration of our universe [@Caldwell:1997ii; @Sahni:1999gb]. We call relic particles freezing-out during this phase [*relentless*]{} dark matter, due to their obstinate struggle to get back to thermal equilibrium. This behavior, which we find in our numerical results shown in Figs. \[fig:FreezeOut\] and \[fig:FreezeOut\_p\], is easily understood with the semi-analytical results given in App. \[app:analytical\]. Notably, the phenomenon of relentless dark matter leads to [*significant numerical differences*]{} in the calculation of the thermal relic density for example in the case of kination-domination phases from previous studies (see e.g. [@Salati:2002md; @Profumo:2003hq]).
The faster expansion rate implies an earlier freeze-out. Since we are dealing with cold relic, reproducing the observed DM density requires couplings significantly larger than in the standard case. This opens up the possibility of having cross section substantially larger than the thermal value in [Eq. ]{}, in contrast with the case of an early matter-dominated epoch providing dilution where smaller values of the cross section are required [@McDonald:1989jd; @Kamionkowski:1990ni; @Chung:1998rq; @Giudice:2000ex], and consequently weaker signals in DM searches. We quantify how much annihilation cross sections can be enhanced in Figs. \[fig:Cross\_Section\_s\] and \[fig:Cross\_Section\_p\].
Finally, we address the question of the origin of the new cosmological component $\phi$ in Sec. \[sec:EFC\], where we provide one explicit example of a microscopic theory leading to the behavior in [Eq. ]{}. We summarize our results in Sec. \[sec:Conclusions\], where we also discuss future work addressing the implications of our analysis for dark matter searches.
A faster expansion {#sec:background}
==================
The expansion rate of the universe, quantified by the Hubble parameter $H$, is controlled by its energy density through Friedmann’s equations. We consider cosmological histories where two different species populate the early universe, radiation and $\phi$, with a total energy density $\rho = \rho_{\rm rad} + \rho_\phi$.
The contribution from radiation, the only one present for a standard cosmological history, can be expressed in terms of its temperature as follows \_[rad]{}(T) = g\_\*(T) T\^4 , \[eq:rhorad\] where $g_*(T)$ is the number of effective relativistic degrees of freedom. We find it useful to express $\rho_\phi$ as a function of the radiation temperature $T$. All we know is its red-shift behavior given in Eq. , hence we need to connect $a$ with $T$. This is achieved by assuming and imposing entropy conservation in a comoving volume $S = s a^3 = {\rm const}$, where the entropy density reads s(T) = g\_[\*s]{}(T) T\^3 . Here, $g_{*s}$ is the effective relativistic degrees of freedom contributing to the entropy density. Entropy conservation ensures $g_{*s}(T)^{1/3} T a = {\rm const}$, and the scaling in Eq. can be re-expressed as follows \_(T) = \_(T\_r) ()\^[(4+n)/3]{} ()\^[(4+n)]{} . Here, $T_r$ is some reference temperature set by the boundary conditions. We choose $T_r$ as the temperature where the two fluids have equal energy densities. The full energy density at any temperature reads
& (T) = \_[rad]{}(T) + \_(T) =\
& \_[rad]{}(T) , \[eq:rhototal\]
where we find it convenient to factor out the energy density of the radiation bath as given in [Eq. ]{}. From this expression it is manifest that the energy budget of the universe is dominated by $\phi$ for temperatures $T \gtrsim T_r$.
With [Eq. ]{} in hand, we can evaluate the Hubble parameter as a function of the temperature H = , \[eq:HubbleRate\] where the reduced Planck mass is $M_{\rm Pl} = (8 \pi G)^{-1/2} = 2.4 \times 10^{18} \, {\rm GeV}$. At temperatures larger than $T_r$, and setting for simplicity $g_*(T) = \overline{g}_* = {\rm const}$, the Hubble rate approximately is H(T) ()\^[n/2]{} , (T T\_r) , \[eq:HubbleEarlyTimes\] The full standard model (SM) degrees of freedom would lead $\overline{g}_* = g_{*{\rm SM}} =106.75$. The expression in [Eq. ]{} is the Hubble rate at the time of DM genesis for the cosmological histories considered in this work. This result manifestly shows how the expansion rate at a given temperature $T$ is always larger than the correspondent value for a standard cosmological history. In our numerical analysis we use the complete expression for $H$, including the full temperature dependence of $g_*(T)$.
BBN Constraints {#sec:BBN}
===============
The successful predictions of light element abundances give us a quantitative test of the energy content of the universe when it was few seconds old. Before we consider freeze-out in the cosmological background described in Sec. \[sec:background\], we have to ensure that we do not spoil this remarkable agreement between theoretical predictions and observations.
A potential issue with BBN arises if $T_r$ is not too far away from the ${\rm MeV}$ scale, where light elements begin to form. If this is the case, the universe expands faster than the usual case around the BBN time, and the theoretical prediction for BBN abundances may be altered.
We parameterize the effect of the field $\phi$ by an effective number of relativistic degrees of freedom (T) = g\^[eff]{}\_\*(T) T\^4 , \[eq:rhototal2\] where we define g\^[eff]{}\_\*(T) = g\_\*(T) + g\_\*\^(T) . Here, $g_*(T)$ is the standard contribution from radiation, whereas $\Delta g_*^\phi(T)$ accounts for the energy density of $\phi$. The expression for the total energy density in Eq. define unambiguously the latter. A historical and widely used way to parameterize this effect is to describe the presence of $\phi$ as the the number of effective neutrinos. Within this convention, the total number of relativistic degrees of freedom appearing in [Eq. ]{} reads g\^[eff]{}\_\*(T) = 2 + 4 + 2 N\_ . \[eq:geffdef\] Here, we account for photons and positrons as well as neutrinos. In the absence of physics beyond the SM, the number of neutrino flavors at temperatures $T > 1 \, {\rm MeV}$ is $N_\nu^{({\rm SM})} = 3$. [^1] By a comparison between the last two expressions, we compute $\Delta N_\nu \equiv N_\nu - N_\nu^{({\rm SM})} = 4 \Delta g_*^\phi / 7$. We find N\_= g\_\*(T\_r) ( )\^[(4+n)/3]{} ()\^[n]{} . This is the general expression for the temperature dependent contribution to the number of additional neutrinos. The temperature $T_r$ cannot be much lower than $T_{\rm BBN} \simeq 1 \, {\rm MeV}$, therefore this contribution is vanishing at the time of CMB formation. If we consider $T_r$ around the BBN time, the expression takes the simpler form N\_ ()\^[n]{} 6.14 ()\^[n]{} . \[eq:DeltaNu\]
We impose the recent bound on $\Delta N_\nu$ from Ref. [@Cyburt:2015mya], where the authors considered an effective number of relativistic species as in [Eq. ]{}, with $N_\nu$ constant over the different temperature range probed by BBN. Our case is different, since we have a temperature dependent $\Delta N_\nu$. As manifestly shown in [Eq. ]{}, such a correction to the number of SM neutrinos increases with the temperature. In order to put the most conservative limits, we evaluate $\Delta N_\nu$ at a time slightly before neutron freeze-out for temperature $T \simeq 1 \, {\rm MeV}$. At such a temperature, neutrons and protons are still in chemical equilibrium in the entire range for the parameters $(T_r, n)$ under consideration, as explicitly computed in App. \[app:BBN\]. In this regard, our bounds are very conservative. Ref. [@Cyburt:2015mya] found the range $2.3 \leq N_\nu \leq 3.4$ at $95 \%$ CL ($2 \sigma$). The contribution in [Eq. ]{} is always positive, so the BBN bounds only allow the region in the $(T_r, n)$ where T\_r (15.4)\^[1/n]{} . \[eq:BBNbound\]
Dark Matter Freeze-Out {#sec:FO}
======================
In this section we analyze dark matter freeze-out in the cosmological background introduced in Sec. \[sec:background\]. The underlying assumption here is that DM particles achieve thermal equilibrium in the early universe. The conditions needed to satisfy these requirement are given in App. \[app:DMthermalization\]. A DM particle interacting through a light mediator (i.e. lighter than the TeV scale) and with coupling strength at least as big as weak gauge interactions thermalizes at temperatures above the TeV scale. In this regime, DM is produced through freeze-out. We first present the Boltzmann equation describing the DM number density evolution. All the results presented in this Section are obtained by numerically solving this equation. In order to understand the qualitative features of the solutions we found, the semi-analytical solution presented in App. \[app:analytical\] is very useful. In particular, this solution allows us to estimate the freeze-out temperature and understand the relentless behavior of relics. This regime where DM particles keep annihilating until $T \simeq T_r$ is entered for $n \geq 2$ ($n \geq 4$) if DM annihilations are $s$-($p$-)wave processes. We present explicit solutions for the number density as a function of the temperature, and we quantify the enhancement in the cross section we need with respect to the standard calculation.
Finally, we investigate the relic density dependence on the DM mass. As is well known, the thermal relic density for WIMPs in a standard cosmology depends on the DM mass very weakly (logarithmic, see App. \[app:analytical\]). The quantity that sets the final abundance is the annihilation cross section. We find that this is not the case anymore for a fast expanding universe, since there is a new scale, the temperature $T_r$. The relative hierarchy between the DM mass and $T_r$ determines whether freeze-out happens before or after the epoch of $\phi$ domination. The final relic density differ enormously in the two cases, as we discuss extensively in this Section.
Boltzmann Equation
------------------
From now on, we denote $\chi$ the DM particle, and we assume it to be a Majorana fermion. The DM number density is governed by + 3 H n\_= -v\_[rel]{} (n\_\^[2]{}-n\_\^[[eq]{} 2]{}) . \[eq:BoltzmannEq\] Here, $n_\chi^{{\rm eq}}$ and $\langle \sigma v\rangle$ are the equilibrium number density distribution and the thermally averaged cross section, respectively. This is the same as the standard case [@Lee:1977ua; @Scherrer:1985zt; @Srednicki:1988ce; @Gondolo:1990dk], with one important difference: the Hubble parameter $H$ is different. Assuming $m_\chi \gg T_r$, the energy density at the freeze-out epoch is dominated by $\phi$ and the Hubble parameters in this regime is given in [Eq. ]{}.
As usual, it is convenient to write the [Eq. ]{} in terms of the comoving number density, $Y_\chi = n_\chi / s$, and to use $x = m_\chi / T$ as the “time variable” =- (1 - ) (Y\_\^[2]{} - Y\_\^[[eq]{} 2]{}) . \[eq:BoltzmannEq2\] The expression for the comoving equilibrium number density for a Maxwell-Boltzmann distribution is Y\_\^[[eq]{}]{}(x) = x\^[2]{} , \[eq:Ychieq\] where $g_\chi = 2$ for a Majorana fermion and $\BesselK{2}{x}$ is the modified Bessel function. At late times the comoving $Y_\chi(x)$ reaches a constant value $Y_\chi(\infty)$, since the actual number density only changes because of the expansion. The present DM density is $\rho_\chi(T_0) = m_\chi \, Y_\chi(\infty) \, s(T_0)$, where $T_0$ is the current temperature of the Cosmic Microwave Background (CMB) photons.
We expand the annihilation cross section times the relative velocity in partial waves v\_[rel]{} = \_s + \_p x\^[-1]{} + (x\^[-2]{}) , \[eq:sigmavpartial\] keeping only the leading $s$- and $p$-wave contributions. We present numerical results for both cases.
An Earlier Freeze-Out
---------------------
Before looking at the explicit numerical solution, we examine the qualitative features we expect to find in the solutions. First, and not surprisingly, freeze-out happens earlier than for the case of a radiation background. This is due to the Hubble parameter during the phase of $\phi$ domination, which for a given temperature is always larger than the associated value in a radiation background. A faster Hubble rate makes it harder for the DM to stay in thermal equilibrium, and freeze-out happens at higher temperatures.
We provide semi-analytical expressions for the freeze-out temperature in Eqs. and for the case of radiation and modified cosmology, respectively. Keeping the DM mass and the annihilation cross section fixed, and focusing for the purpose of this illustration on $s$-wave processes, the freeze-out temperatures are related by T\_[f [rad]{}]{}\^[1/2]{} e\^[- m\_/ T\_[f [rad]{}]{}]{} = T\_f\^[1/2]{} e\^[- m\_/ T\_f]{} ( )\^[n/2]{} . Here, $T_r$ and $T_{f \, {\rm rad}}$ are the freeze-out temperature in a generic $(T_r, n)$ and the radiation background, respectively. For freeze-out happening during the $\phi$ dominated epoch, $T_f > T_r$, the freeze-out temperature is larger than the one for the case of a radiation background, $T_f > T_{f \, {\rm rad}}$. Even if the numerical difference between the two temperatures is a factor of a few, the consequent modification of the relic density are significant, since freeze-out happens on the exponential tail of the Maxwell-Boltzmann distribution.
Relentless Relics
-----------------
We point out here a very peculiar and previously unrecognized behavior of the number density evolution once $n$ gets large. In order to understand the physics underlying this feature, it is useful to start the discussion by reviewing what happens right after freeze-out for a standard radiation background. DM particles depart from thermal equilibrium when the interaction rate, $\Gamma \simeq n_\chi \langle \sigma v_{\rm rel} \rangle$ is of the order of the Hubble rate, $H_{\rm rad} \simeq T^2 / M_{\rm Pl}$. Immediately after freeze-out, DM particles can still annihilate occasionally, just not enough to stay in thermal equilibrium. The post freeze-out annihilation rate scales as $\Gamma \propto T^3 (T^4)$ for $s$-wave ($p$-wave) annihilations, due to the dilution of the DM particles from the expansion of the universe. This is not enough for the annihilation rate to compete with the Hubble rate, and post freeze-out annihilations do not change the density significantly. This can be observed in our numerical solutions, and it can also be understood analytically (see [Eq. ]{}).
We can repeat the same analysis for the set of modified cosmologies considered here. The argument goes along the same lines, with one important difference: the Hubble parameter now scales as $H \propto T^{2 + n/2}$. Thus there is a critical value of $n$ above which the post freeze-out annihilation rate scales with a power of temperature lower than the one for the Hubble rate. For $s$-wave annihilation, this happens for $n \geq 2$. Interestingly, the case $n = 2$ corresponds to motivated theories of quintessence [@Caldwell:1997ii; @Sahni:1999gb]. For $p$-wave annihilation, the condition for this to be the case is $n \geq 4$.
What are the consequences of this relative scaling? For $s$-wave annihilating DM and $n \geq 2$ cosmologies, the annihilation rate red-shifts slower than the Hubble rate. The effects of post freeze-out annihilations is then substantial: DM particles keep annihilating, [*relentlessly*]{} trying to get to the equilibrium thermal distribution; thermal equilibrium, however, is always unaccessible due to the temperature being low enough for the equilibrium number density to be deeply in the exponential tail. The older the age of the universe, the lower the temperature, and the harder it is for DM particles to get to the equilibrium distribution. The process of depletion goes on until temperatures of the order $T_r$, when the expansion is driven by the radiation bath, and the usual scaling applies.
Number Density Evolution
------------------------
We now show results for the full numerical solutions to the Boltzmann equation in Fig. \[fig:FreezeOut\] and \[fig:FreezeOut\_p\] for the case of $s$- and $p$-wave annihilation, respectively. We take a DM mass $m_\chi = 100 \, {\rm GeV}$ (we discuss the very important dependence on mass below) and we fix the annihilation cross section in such a way that we reproduce the observed DM abundance for the case of a standard cosmology (red lines). The solutions for the other cosmological histories are obtained by fixing $T_r = 20 \, {\rm MeV}$ and $n$ as described in the figure caption.
The comoving number density $Y$ in Fig. \[fig:FreezeOut\] do not change significantly after freeze-out for the radiation (red line) and $n = 1$ (green line) cases. This is expected and consistent with the qualitative analysis above. However, an important difference is already clearly visible: freeze-out happens [*earlier*]{} for $n=1$, than for the $n=0$ standard case, and as a consequence the asymptotic comoving density is higher.
The phenomenon of [*relentless*]{} annihilation is visible in Fig. \[fig:FreezeOut\] already for the $n = 2$ (thick dashed blue line), as also expected from the discussion above: DM particles continue to find each other to annihilate much later than freeze-out, since the Hubble rate and the red-shifting annihilation rate feature the same scaling with temperature, until the universe becomes radiation dominated and eventually $H\gg\Gamma$. The number density evolution in this regime can be understood analytically (see Eqs. and ), and it is closely approximated by the expression Y\_(x) \^[-1]{} . Here, $x_f$ and $x_r$ are the freeze-out temperature and $T_r$ expressed in terms of the dimensionless variable $x = m_\chi / T$, respectively. The slow logarithmic decrease of the number density is the result of the [*relentless*]{} attempt of the DM to go back to thermal equilibrium. This behavior persists until $T \simeq T_r$, after which the DM comoving number density reaches a constant value.
This post freeze-out annihilation are even more pronounced for $n > 2$, as we can see from the orange and the magenta lines in Fig. \[fig:FreezeOut\]. In this regime for $n$, the comoving number density is approximated by the following expression Y\_(x) \^[-1]{} . The decrease of $Y_\chi$ is even faster, with a power law instead of the logarithmic dependence appearing for the marginal case of $n=2$. As before, the number density keeps decreasing with the behavior described above, until radiation takes over.
The discussion for the $p$-wave solutions in Fig. \[fig:FreezeOut\_p\] is analogous: Freeze-out happens earlier and earlier for higher and higher $n$, and the resulting number density is correspondingly larger. The only difference is that the transition to [*relentless*]{} relics sets in at $n = 4$, as correctly estimated above.
Enhancement in the Relic Density
--------------------------------
One of the central results of the number density evolution analysis is that freeze-out abundances are in general larger than in the standard case: The red lines is below all the other ones in Figs. \[fig:FreezeOut\] and \[fig:FreezeOut\_p\], for fixed values of the annihilation cross section. One can turn the argument around, and state that [*larger cross sections*]{} are therefore needed, with the cosmological setup we consider here, to reproduce the observed DM density. This is quite remarkable, as large cross sections translate into larger couplings and therefore larger signals in DM searches, especially in the context of indirect detection. This thus begs the question: How large can the annihilation cross section be, consistently with BBN bounds?
The two dimensional parameter space $(T_r, n)$ entirely fixes the cosmological background in the present setup. At large values of $T_r$, larger than the DM mass, the standard freeze-out calculation holds, and there is no enhancement to the cross section. The lower $T_r$, the larger the enhancement; However, we cannot take $T_r$ arbitrarily small, as we have to satisfy the BBN bounds in [Eq. ]{}.
The results for $s$-wave annihilation are shown in Fig. \[fig:Cross\_Section\_s\], where we fix the DM mass to $m_\chi = 100 \, {\rm GeV}$ and we calculate for each point in the $(T_r, n)$ the cross section needed to produce the right amount of thermal relic DM, normalized to $\sigma_s = 1.7\times10^{-9} \,\gev^{-2}$, the value producing the “correct” thermal relic density for a radiation background. We checked numerically that within better than 20% accuracy, [*the contour lines also correspond to the enhancement to the thermal relic abundance*]{} for a [*fixed*]{} value of the pair-annihilation cross section, in Fig. \[fig:Cross\_Section\_s\] $\sigma_s = 1.7\times10^{-9} \,\gev^{-2}$. In the bottom left corner of the figure we shade in grey the region excluded by BBN.
The figure importantly also indicates the [*“boost factors”*]{} expected in indirect detection signals, compared to a standard cosmological setup. The key message is that for the $s$-wave case enhancements beyond $\sim 10^3$ are possible.
The analogous analysis for $p$-wave annihilation is presented in Fig. \[fig:Cross\_Section\_p\]. As a result of the temperature dependence of the cross section, [*larger enhancement factors are possible*]{}, up to $\sim 10^6$ and above. In the case of $p$-wave annihilation, however, indirect signals are suppressed by the DM velocity in the late universe, thus the enhancement to indirect signals is both smaller than the enhancement factors to the relic density, and dependent on environment.
As a side note, we point out that the effect of relentless annihilation produces significantly [*smaller enhancements than what previously calculated in the literature*]{} (see e.g. [@Salati:2002md; @Profumo:2003hq]). This is presumably traced back to the previous calculations having assumed a constant value for the comoving number density after a certain effective freeze-out temperature, thus neglecting the [*relentless*]{} annihilation potentially affecting the relic density over several decades in temperature.
Dependence on the DM mass
-------------------------
All the results presented so far assumed the fixed DM mass benchmark value $m_\chi = 100 \, {\rm GeV}$. For a cold relic in a standard cosmology, the value of the DM mass has a weak impact on the final abundance, which is controlled by the annihilation cross section. We conclude this Section by pointing out one more interesting feature than the framework discussed in this work: the relic density has a strong dependence on the DM mass.
The reason why this is the case is the presence of the critical temperature $T_r$. If the freeze-out temperature is below $T_r$, there is no change with respect to the standard story. In the opposite case, the precise value of the DM mass is important. Freeze-out happens at temperatures $T_f \simeq m_\chi / 10$, thus the larger the DM mass, the longer the DM particle [*relentlessly*]{} reduce its comoving number density through residual annihilations. Again, this means that compared to previous calculations the larger the ratio of $m_\chi/T_r$, the larger the effect and the larger the suppression of the calculated enhancement to the thermal relic density.
To quantitatively study this effect, we fix a few benchmark cosmologies and show contours of fixed relic density in the $(m_\chi, \sigma_{s,p})$ plane. The results are shown in Figs. \[fig:Mass\_CrossSection\_s\] and \[fig:Mass\_CrossSection\_p\] for $s$-wave and $p$-wave annihilation cross sections, respectively. At low values of the DM mass, corresponding to a freeze-out temperature below $T_r$, these lines are close to horizontal: This is expected, as in the standard case the relic density depends only on the cross section. The mild dependence on the mass comes from two factors: (i) the logarithmic mass dependence of the freeze-out temperature, and (ii) the different value of $g_*$ at the freeze-out. However, for larger DM mass we see that the relic density strongly depends on the mass, since the larger the DM mass, the longer the phase of [*relentless*]{} annihilation, and the ensuing suppression of the relic density. In the figure we also indicate, in the top-right corners, regions in conflict with perturbative unitarity [@Griest:1989wd]. Comparing Figs. \[fig:Mass\_CrossSection\_s\] and \[fig:Mass\_CrossSection\_p\] one can also appreciate the steeper dependence on mass in the $p$-wave case. This arises because of the steeper dependence of $\Gamma$ on temperature in the $p$-wave case, and is already reflected in the larger enhancements we find, e.g., in Fig. \[fig:Cross\_Section\_p\] compared to Fig. \[fig:Cross\_Section\_s\].
Ultra Stiff Fluids {#sec:EFC}
==================
A virtue of the freeze-out analysis performed in the previous Section is its generality and model independence. Any DM particle thermalized in the cosmological background of [Eq. ]{} is produced through freeze out as described in Sec. \[sec:FO\]. The analysis only assumes our knowledge of the two parameters $(T_r, n)$, without the need of specifying any further property of the new species $\phi$. In this last part of the paper, we provide explicit microscopic realizations for $\phi$, reproducing the red-shift behavior in [Eq. ]{}.
All the examples we consider are theories of a single real scalar field $\phi$ minimally coupled to gravity S= d\^[4]{}x (- g\^ \_ \_ -V() ) . \[eq:Sphi\] For the remaining of this Section, we set $M_{\rm Pl} = 1$. The energy density and pressure for this fluid read p\_&=& ( )\^2 - V() ,\
\_&=& ( )\^2 + V() , leading to the equation of state w\_ = = . \[eq:wphidef\] For such an equation of state, the energy density red-shifts as $\rho_\phi \propto a^{-3 (1 + w_\phi)}$, which allows us to connect n = 3 w\_- 1 , \[eq:nvsw\] where $n$ is the index defined through the red-shift behavior in [Eq. ]{}. For a positive scalar potential, the allowed values of $w_\phi$ are in the range $(-1, +1)$. Equivalently, the range for $n$ is between $-4$ and $+2$. The highest $n$ is achieved during a kination phase, where the energy density of $\phi$ is mostly kinetic. In order to get values larger than $n = 2$, we need to consider negative scalar potentials. In what follows, we first describe examples of fluids with $n= 2 $ and then we show how the $n > 2$ domain can be accessed.
Quintessence ($n = 2$)
----------------------
Examples of theories with $n=2$ are quintessence fluids motivated by the discover of the accelerated expansion of the universe [@Caldwell:1997ii; @Sahni:1999gb]. The energy density of this type of fluid red-shifts as $\rho_\phi \propto a^{-6}$ in the kination regime, i.e. when the kinetic energy density dominates over the potential energy . One possible scalar potential leading to this behavior is the exponential form [@Ratra:1987rm; @Wetterich:1987fm] V() = . The role of quintessence for neutralino dark matter freeze-out was studied in Refs. [@Salati:2002md; @Profumo:2003hq]. Alternatives to quintessence, still with the same red-shift behavior, are Chaplygin gas [@Kamenshchik:2001cp] or a perfect fluid described by a polytropic equation of state [@Chavanis:2014lra].
Faster than Quintessence ($n > 2$)
----------------------------------
We provide here example theories where $n > 2$. We assume the energy density of the universe to be entirely dominated by $\phi$, with red-shift as in [Eq. ]{}. The scale factor vs time relation can be derived from the Friedmann equation a(t)= a\_i ( )\^[2/(n+4)]{} , where we define $a_i$ to be the value of the scale factor at $t = t_i$. The time derivative of the Hubble parameter reads = - (\_+ p\_) = - ( )\^2 . \[eq:dHdt\] By comparing this expression with the one resulting from direct calculation, $dH / dt = -2/[(n+4)t^{2}]$, we find the time evolution of the field (t)=\_i+() . \[eq:phivst\]
We go back to [Eq. ]{}, and if we assume that $w_\phi = {\rm const}$ we can solve for the scalar potential V() = - () ( )\^2 , where we have traded $w_\phi$ with $n$ by using [Eq. ]{}. The time derivative of the field $\phi$ is related to the one of the Hubble parameter, as shown in [Eq. ]{}. We know how the Hubble parameter scales with time in this background with $w_\phi = {\rm const}$, therefore we can find an expression for the potential as a function of time. Once this is done, we use [Eq. ]{} to trade the time variable with $\phi$. The output of this procedure is the scalar potential as a function of the field V() = V\_i e\^[- ]{} . The overall constant reads V\_i = - e\^[\_i ]{} , and it is negative for $n > 2$. It is straightforward to check that the solution in [Eq. ]{} with the potential above satisfies the equation of motion $\ddot{\phi}+3H\dot{\phi}+dV(\phi)/d\phi=0$, as it should. This class of potentials have been used in the context of ekpyrotic scenario [@Khoury:2001wf]. A dynamical $w_{\phi} > 1$ can be obtained also with periodic potentials [@Choi:1999xn; @Gardner:2004in].
No superluminal propagation
---------------------------
We conclude this Section with one important comment. The regime $w_\phi > 1$ implies $p_\phi > \rho_\phi$, and there may be concerns about superluminal propagation. However, the speed of sound for a canonical scalar field with action as in [Eq. ]{} is always $c_{s}^{2}=1$ [@ArmendarizPicon:1999rj; @Christopherson:2008ry]. Consequently, causality is not violated.
Conclusions {#sec:Conclusions}
===========
We analyzed DM freeze-out for non-standard cosmological histories which include a faster-than-usual expansion at early times, driven by a new cosmological species $\phi$. We gave a full description of the cosmological backgrounds in Sec. \[sec:background\]. We then parameterized the possible cosmological histories by the values of $n$ and $T_r$, i.e., respectively, the index appearing in [Eq. ]{} and the temperature when the energy densities of $\phi$ and radiation are equal. Light element abundances exclude part of this two-dimensional parameter space, and this BBN bound is summarized by [Eq. ]{} of Sec. \[sec:BBN\].
In calculating the DM density evolution we identified two distinct possibilities: For $n$ not too large, the behavior is quite similar to the one for standard freeze-out, where shortly after chemical decoupling the comoving number density approaches its asymptotic value. For large $n$, however, we found a new domain where post freeze-out annihilation substantially dilute the DM density. This is explained by the different scaling of the Hubble parameter with temperature, $H \propto T^{2 + n/2}$; we called this possibility [*relentless*]{} dark matter. The critical values of $n$ dividing the two regimes are $n = 2$ and $n = 4$ for $s$-wave and $p$-wave annihilation, respectively.
A central result of our analysis is that DM particles which freeze out in the cosmological era dominated by the new species $\phi$ must have cross sections way larger than the thermal value in [Eq. ]{} if DM is to be a thermal relic. We plan to study in the future the implications for dark matter searches of such a large annihilation cross section, such as CMB spectral distortion [@Slatyer:2015jla] and bounds from gamma rays [@Ackermann:2015zua].
The underlying assumption of our study of DM genesis was an early time thermalization. As discussed in App. \[app:DMthermalization\], this is not necessarily the case, and the faster expansion makes things even harder. If our assumption is not satisfied, DM production would be non-thermal. Assuming production from a decay and/or scattering of particles in the thermal bath, the comoving density produced at a given temperature $T$ approximately reads Y\_(T) (T) H(T)\^[-1]{} (T) M\_[Pl]{} T\^[- 2 - n/2]{} . If the rate is mediated by a higher dimensional operator of mass dimension $d$, it would scale as $\Gamma(T) \propto T^{2 d - 7}$. Thus the comoving density at a given temperature scales as $Y_\chi(T) \propto T^{2 d - n/2 - 9}$. We see that the relative size of $d$ and $n$ establishes where most of DM particles are produced. If the dimension is not too large, $d < n/4 + 4.5$, the production is dominated at lower temperature, of the size of the decaying/scattering bath particles. This type of “IR production” is known as freeze-in [@Hall:2009bx]. In the opposite case, $d > n/4 + 4.5$, the production is dominated by scattering at high temperatures, similarly to the UV production of axinos or gravitinos [@Rychkov:2007uq; @Strumia:2010aa]. This latter case is especially interesting, because it requires the knowledge of how the universe entered the $\phi$ domination phase after inflation. We will study both possibilities in a forthcoming analysis.
Acknowledgments {#acknowledgments .unnumbered}
===============
This work was supported by the U.S. Department of Energy grant number DE-SC0010107. We are grateful to Josué De Santiago and Simone Ferraro for discussions.
Dark matter thermalization {#app:DMthermalization}
==========================
The DM production mechanism depends on whether the DM ever reaches thermal equilibrium at early times. Thermalization is achieved by collisions, therefore a faster expanding universe makes it harder for the DM to thermalize. This is what we investigate in this Appendix, checking whether the interaction rate between DM and the radiation bath was ever larger than the expansion rate at high temperatures. If this was the case, then DM reaches thermal equilibrium and it is produced through thermal freeze-out. In the opposite case, the production mechanism must be non-thermal.
For temperatures much larger than the DM mass, the scattering rate can be parameterized as follows $$\begin{aligned}
\label{eq:interactionrate}
\Gamma_{\rm scatt}(T) \simeq & \, n_{\rm DM} \sigma_{\rm scatt} v_{\rm rel} \simeq \\ & \nonumber
\frac{3}{2} \frac{\zeta(3)}{\pi^2} T^3 \, \frac{\lambda^4}{32\pi} \frac{T^2}{(T^2 + M_*^2)^2} \ .\end{aligned}$$ Here, we use the number density for a Majorana fermion in the relativistic regime, and the scattering is assumed to be mediated by a particle with mass $M_*$ that couples to DM and radiation with strength $\lambda$. In what follows, we explore two different possibilities for $M_*$.
Massless Mediator
-----------------
The first case we study is a massless mediator, $M_* = 0$. Strictly speaking, this analysis is valid also for the case of a massive mediator with mass smaller than the temperatures under consideration. For example, any mediator lighter than the DM particle would fall within this category. The scattering rate in this case reads \_[scatt]{}(T) T , (M\_\* T) . \[eq:interactionrateMassless\] This linear scaling with the temperature has to be contrasted with the Hubble rate proportional to $T^{2 + n/2}$ (with $n>0$, see Eq. ). At high enough temperatures the expansion rate wins, and interactions become more effective as the universe expands and cools down.
A comparison between the Hubble rate in [Eq. ]{} for different values of $n$ and the scattering rate in [Eq. ]{} is shown in Fig. \[fig:DMthermalization1\], where we plot both these quantities as a function of the inverse temperature. The Hubble rate is obtained by fixing $T_r = 20 \; {\rm MeV}$ in order to have the faster expanding phase to last as long as possible, but still consistent with the BBN bounds discussed in Sec. \[sec:BBN\]. The red line corresponds to the standard cosmological history, the other colored line represent the faster expansion rate, with $n$ the index appearing in the exponent of Eq. . The rate is computed by setting the size of the coupling $\lambda \simeq 1$. DM thermalizes at a temperature $T_{\rm th}$ defined to satisfy the condition $H(T_{\rm th}) = \Gamma_{\rm scatt}(T_{\rm th})$. In other words, this temperature can be obtained by finding the intersection between the black lined and the colored line under consideration in Fig. \[fig:DMthermalization1\]. This value depends on $n$, and it falls within the range $T_{\rm th} \simeq (10^3, 10^9) \, {\rm GeV}$ as we vary $n$ from $1$ to $4$. DM particles always achieve thermal equilibrium at temperatures higher than the weak scale, even in the extreme case $n = 4$.
The above conclusion would be altered if we considered smaller values for the coupling $\lambda$. We find it useful to write down an analytical expression for $T_{\rm th}$, which can be obtained by using the approximate expression for the Hubble rate in Eq. . The thermalization temperature approximately reads T\_[th]{} ( M\_[Pl]{} T\_r\^[n/ 2]{})\^[2 / (n+2)]{} . \[eq:Tthcase1\] It scales as $\lambda^{8 / (n+2)}$, so taking a smaller $\lambda$ would affect less the cases of larger $n$. By taking a weak interaction coupling $\lambda \simeq 0.3$, the thermalization temperature is in the range $T_{\rm th} \simeq (10^3, 10^8) \, {\rm GeV}$. The numerical solution for the thermalization temperature as a function of $\lambda$ is shown in Fig. \[fig:DMthermalization2\]. For small couplings, $\lambda \lesssim 10^{-3}$, the thermalization temperature is below the weak scale. For weak scale DM this implies that thermal equilibrium is never achieved, and the production mechanism must necessarily be non-thermal.
Heavy Mediator
--------------
We consider here the case of heavy mediators. At temperatures below $M_*$, the scattering rate in approximately \_[scatt]{}(T) , (T M\_\*) . \[eq:interactionrateMassive\] Unlike the case discussed above, the interaction rate now scales with a higher temperature power than the Hubble parameter. This means that at very early times interactions are effective, and as the temperature drops the expansion takes over. We illustrate this case in Fig. \[fig:DMthermalization3\], where we compare again the rates as a function of the (inverse) temperature. We consider three masses for the mediator. We define $m_\Lambda \simeq 10^{10} \, {\rm GeV}$, the scale where the SM Higgs quartic vanishes [@Buttazzo:2013uya], hinting for possible new physics [@Hall:2013eko; @Hall:2014vga; @Fox:2014moa; @D'Eramo:2015ssa]. We also consider the unification scale for the gauge couplings ($M_{\rm GUT} \simeq 2 \times 10^{16} \, {\rm GeV}$) and the reduced Planck mass ($M_{\rm Pl})$. For an order one coupling, $\lambda \simeq 1$, thermalization is never achieved for $n>0$. This conclusion is unchanged even if we badly break perturbation theory, $\lambda \simeq 4 \pi$, and it is only strengthened if we consider smaller couplings. We conclude that for a heavy mediator, as heavy as at least $10^{10} \, {\rm GeV}$, DM never equilibrates with the thermal plasma.
Neutron Freeze-Out and BBN {#app:BBN}
==========================
The neutron freeze-out temperature for the cosmological background studied in this work can be found by using the analytical results of Ref. [@Mukhanov:2003xs]. The neutron abundance in conveniently expressed in terms of the neutron fraction $X_n \equiv n_n / (n_n + n_p)$, where $n_{n(p)}$ is the neutron (proton) abundance. The time evolution is described by the Boltzmann equation = - \_[n p]{} ( 1 + e\^[-Q/T]{} ) ( X\_n - X\_n\^[eq]{} ) . Here, we introduce the neutron-proton mass difference Q m\_n - m\_p = 1.293 , and the equilibrium neutron fraction reads X\_n\^[eq]{} = . For temperatures above the electrons mass, the neutron to proton conversion rates can be approximated by the analytical expression \_[n p]{} 2 1.63 ( )\^3 ( + 0.25 )\^2 [sec]{}\^[-1]{} . This simple expression reproduces the full result within the accuracy of a few percent.
The Boltzmann equation for the neutron fraction can be solved as an asymptotic series X\_n = X\^[eq]{}\_n . This expression is valid as long as the second term in the parenthesis is smaller than the first, namely if we are close to the equilibrium value. We define the neutron freeze-out as the temperature when the two are equal | |\_[T = T\_n\^[[FO]{}]{}]{} 1 . The only missing information to solve this equation is the time vs temperature relation, which differs from the one for a standard cosmology due to the presence of $\phi$. This can be derived by imposing conservation of the total entropy. Since we focus on temperatures around the ${\rm MeV}$ scale, we neglect the $g_{*s}$ temperature dependence and the final equation for the freeze-out temperature reads | H(T) |\_[T = T\_n\^[[FO]{}]{}]{} 1 . \[eq:neutronFO\]
The Hubble rate $H$ as defined in [Eq. ]{} contains both the energy density of $\phi$ and radiation. If we only account for radiation and we solve [Eq. ]{} we find $T_n^{{\rm FO}} \simeq 0.76 \, {\rm MeV}$, valid for a standard cosmology [@Mukhanov:2003xs]. If we also account for the $\phi$ energy density, we find that this result is not changed by much as long as we consider $T_r \geq 1 \, {\rm MeV}$. More specifically, in the extreme case $n = 0$ (where there is no temperature dependence in $N_\nu$) and $T_r = 1 \, {\rm MeV}$ we find $T_n^{{\rm FO}} \simeq 0.83 \, {\rm MeV}$. In order to put the most conservative bounds, we evaluate $\Delta N_\nu$ as defined in [Eq. ]{} for $T = 1 \, {\rm MeV}$.
Semi-Analytical Freeze-Out {#app:analytical}
==========================
In this Appendix we derive semi-analytical solutions for freeze-out relic density. We start with a review of the standard calculation for DM production in a background of radiation, then we extend it to the modified cosmological histories considered in this work.
Standard Cosmology
------------------
In order to connect with the new freeze-out scenarios studied in this paper, we review the Lee-Weinberg calculation for cold relics [@Lee:1977ua]. We simplify the Boltzmann equation in [Eq. ]{} by setting $g_{*} = g_{*s} = {\rm const}$. Furthermore, we Taylor expand the equilibrium density in [Eq. ]{} for temperatures lower than the DM mass Y\_\^[[eq]{}]{}(x) = x\^[3/2]{} e\^[-x]{} + …(x 1) . \[eq:YchieqNR\] The Boltzmann equation can be written as follows = - A (Y\_\^[2]{} - Y\_\^[[eq]{} 2]{}) , \[eq:BEsimple\] where we define the constant [^2] A = = g\_\*\^[1/2]{} m\_M\_[Pl]{} . \[eq:defA\]
We identify two distinct regimes for the solution. At early times, DM annihilations are efficient and $Y_\chi$ closely tracks the equilibrium distribution. At late times, the expansion takes over and the density freezes-out. We solve the Boltzmann equation in these two regimes and then match the two solutions at some intermediate point. We perform the matching at the temperature where $Y_\chi$ moves away from its equilibrium expression, a point known as the DM freeze-out.
We find it convenient to write the Boltzmann equation for $\Delta_\chi \equiv Y_\chi - Y_\chi^{\rm eq}$, which is obtained by plugging its definition into [Eq. ]{}. We find = - A \_(2 Y\_\^[eq]{} + \_) - . \[eq:BEsimple2\] At times much earlier than freeze-out, the departure from thermal equilibrium is minimal and we can neglect terms quadratic in $\Delta_\chi$ and its derivative. As a consequence, the DM number density can be approximated by Y\_(x) Y\_\^[eq]{}(x) + (1 < x < x\_f) . \[eq:Ychisemiearly\] In the opposite regime, we neglect the equilibrium distribution in the Boltzmann equation , which can be integrated to find the solution Y\_(x) \^[-1]{} (x > x\_f) . \[eq:Ychisemilate\] Here, we define the annihilation integral J(x) \_[x\_f]{}\^x dx .
The term equal proportional to the inverse comoving density at the freeze-out in [Eq. ]{} is important to ensure that our solution is continuous. However, it is numerically subdominant, unless we consider values $x \simeq x_f$. This can be explicitly checked for the partial wave expansion of [Eq. ]{}, for which the annihilation integral now reads: J(x) \_s ( - ) + ( - ) . The comoving number density after freeze-out reads Y\_(x) = {
[cccc]{} & & &\
& & &
. . \[eq:YchisemilateWAVES\] The above equation illustrates how the comoving number density quickly approaches a constant values after freeze-out. This is only valid for the standard case of a radiation background. In the cosmological histories discussed in this work, we find that DM particles keep annihilating well after the number density has departed from its equilibrium value.
The current DM abundance is evaluated from the asymptotic value ($x \gg x_f$) of the comoving number density. This can be obtained by extrapolating [Eq. ]{} to very large values of $x$, and we find Y\_\^= = . The first equality is general, whereas the second assumes the solutions in [Eq. ]{} for a partial wave expansion. The asymptotic number density scales as the inverse DM mass. In units of the critical density, the DM density results in \_h\^2 = . This quantity depends on the DM mass only through the value of $x_f$.
Finally, we determine the value of the freeze-out temperature. This is the point where we match the two solutions in Eqs. and . We define the freeze-out as temperature $x_f$ by imposing $\Delta_\chi(x_f) = c \,Y_\chi^{\rm eq}(x_{f})$, where $c$ is an order one coefficient. We plug this definition into the Boltzmann equation (\[eq:BEsimple2\]), and the freeze-out condition is expressed as follows . |\_[x = x\_f]{} = m\_ M\_[Pl]{} , \[eq:FO\] where we also restore the definition for $A$ as in [Eq. ]{}. We remind that the thermally averaged cross section can depend on $x$, as in the case of $p$-wave annihilation.
Non-Standard Cosmology Freeze-Out
---------------------------------
For the modified cosmological backgrounds considered here, the DM number density evolution is still described by [Eq. ]{}. However, the temperature dependence of the Hubble parameter is different. We introduce the quantity $x_r \equiv m_\chi / T_r$, where $T_r$ was defined as the temperature where the energy of the radiation bath reaches the one of $\phi$. The Boltzmann equation now reads = - A (Y\_\^[2]{} - Y\_\^[[eq]{} 2]{}) , \[eq:BEsimpleNS\] where we use again the parameter $A$ defined in [Eq. ]{}.
We assume that freeze-out happens during the time of $\phi$ domination, namely $x_f \ll x_r$. At the freeze-out time, the Boltzmann equation can then be approximated by - A (Y\_\^[2]{} - Y\_\^[[eq]{} 2]{}) . \[eq:BEsimpleNSFO\] We solve again before and after freeze-out by using the convenient variable $\Delta_\chi$. At earlier times we neglect terms quadratic in $\Delta_\chi$ and its derivative Y\_(x) Y\_\^[eq]{}(x) + (1 < x < x\_f) . \[eq:Ychisemiearly2\] After freeze-out, the solution takes the same form Y\_(x) \^[-1]{} (x\_f < x < x\_r) . \[eq:Ychisemilate2\] This looks analogous to [Eq. ]{}, but with the crucial difference that the annihilation integral reads J\_(x) \_[x\_f]{}\^x dx . We can perform the integral for partial wave expansion, and we find the expressions J\^[([s]{})]{}\_(x) = {
[cccl]{} & & & n 2\
(x / x\_f) & & & n = 2
. , \[eq:JEFCs\] and J\^[([p]{})]{}\_(x) = {
[cccl]{} & & & n 4\
(x / x\_f) & & & n = 4
. , for $s$- and $p$-wave, respectively.
The solution in [Eq. ]{} can only be extrapolated up to $x = x_r$. Once the radiation bath dominates the energy density, we perform an additional matching, analogous to the one for standard freeze-out (see [Eq. ]{}). The subsequent evolution is described by Y\_(x) \^[-1]{} (x > x\_r) , \[eq:Ychisemiverylate2\] where define the annihilation integral now reads J\_[rad]{}(x) \_[x\_r]{}\^x dx .
The final DM density is $\rho_\chi(t_0) = m_\chi Y_\chi^\infty \, s(T_0)$, where the asymptotic value of the comoving density can be extracted by [Eq. ]{}.
We conclude with the evaluation of the freeze-out temperature, defined as before by the condition $\Delta_\chi(x_f) = c \,Y_\chi^{\rm eq}(x_{f})$. We find . ()\^[n/2]{} |\_[x = x\_f]{} = m\_ M\_[Pl]{} . \[eq:FO2\] This relation is very similar to [Eq. ]{} with the important difference of a $(x_r/x)^{n/2}$ factor, which significantly enhances the left-hand side since we consider freeze-out during the $\phi$ domination phase ($x_f \ll x_r$). If we fix the DM mass and annihilation cross section, freeze-out must happen earlier with respect to the standard case.
[^1]: At lower temperatures, neutrinos decouple from the thermal bath, and after $e^+ e^-$ pair annihilations their temperature is lower than the photons, $T_\nu = (4/11)^{1/3} T_\gamma$. Furthermore, corrections due to non-instantaneous neutrino decoupling lead to a SM effective number of neutrino light flavors $N_{\rm eff}^{({\rm SM})} = 3.04$ [@Dodelson:1992km].
[^2]: $H_{\rm rad}(x)$ is Hubble parameter obtained by plugging only the energy density of the radiation bath. This is obviously the case for standard freeze-out. We find this definition useful also for the case when the energy density is dominated by $\phi$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
The nonlinear dynamics of an obliquely oriented wave packet at sea surface is studied both analytically and numerically for various initial parameters of the packet, in connection with the problem of oceanic rogue waves. In the framework of Gaussian variational ansatz applied to the corresponding (1+2D) hyperbolic nonlinear Schrödinger equation, a simplified Lagrangian system of differential equations is derived, which determines the evolution of coefficients of the real and imaginary quadratic forms appearing in the Gaussian. This model provides a semi-quantitative description for the process of nonlinear spatio-temporal focusing, which is one of the most probable mechanisms of rogue wave formation in random wave fields. The system is integrated in quadratures, which fact allows us to understand qualitative differences between the linear and nonlinear regimes of the focusing of wave packet. Comparison of the Gaussian model predictions with results of direct numerical simulation of fully nonlinear long-crested water waves is carried out.
[Key words: oceanic rogue waves, nonlinear focusing, variational approximation.]{}
author:
- 'V. P. Ruban'
title: 'Gaussian variational ansatz in the problem of anomalous sea waves: Comparison with direct numerical simulations'
---
Introduction
============
Anomalous waves at the ocean surface and in many other physical systems (known also as giant waves, rogue waves, freak waves) have been a popular subject of present-day scientific research (see, e.g., reviews [@Kharif-Pelinovsky; @DKM2008; @ORBMF-2013], and references therein). The most frequently used mathematical model for this phenomenon is the (1+1D) nonlinear Schrödinger equation (NLSE) with the focusing nonlinearity. In particular, such equation describes the complex amplitude of the main harmonic of a quasi-monochromatic weakly nonlinear surface wave over plane potential flows of an ideal fluid [@Zakharov68]. Due to nonlinear self-focusing, the modulation instability develops in a sufficiently lengthy and high wave group [@Zakharov68; @BF]. As the result, a one-dimensional anomalous wave arises. Such a scenario is confirmed by numerical and laboratory experiments [@ZDV2002; @DZ2005Pisma; @ZDP2006; @R2012; @CHA2011; @CHOA2012]. The full integrability makes the (1+1D) NLSE attractive for analytical studies and provides many exact solutions which describe some important properties of real rogue waves (see, e.g., [@AEK85; @AK86; @AEK87; @AAS-C2009; @EHKFAD2011]).
In the Nature however sea waves are rather far from the one-dimensional model (see, e.g., [@OOS2006; @OWT2009; @R2006PRE; @R2007PRL; @Peregrine; @LP2006; @ShS2014]). When on the free fluid surface there are disturbances depending on both horizontal coordinates, the wave dynamics becomes more complicated already in the framework of the corresponding NLSE, not to speak about strongly nonlinear regimes. First, (1+2D) NLSE is not integrable. Second, in the case of deep-water gravity waves, the spatial differential operator in it appears hyperbolic: $$2i\psi_t+\psi_{xx}-\psi_{yy}+|\psi|^2\psi=0.
\label{NLS}$$ Here the non-dimensional variables are in use: $k_0 A^*\rightarrow \psi$, $\omega_0 t \rightarrow t$, $2k_0 x\rightarrow x$, $\sqrt{2}k_0 y\rightarrow y$, where $A(x,y,t)$ is the complex envelope of the main harmonic, $k_0$ is the wave number, $\omega_0=2\pi/T_0=\sqrt{gk_0}$ is the frequency of the carrier wave (the reference frame is moving with the group velocity $v_{\rm gr}=(1/2)\sqrt{g/k_0}$ along $x$ axis). Although Eq.(\[NLS\]) has some disadvantages, which are not discussed here, but it is more suitable for approximate analytical studies (because in all its modified variants, some pseudo-differential operators appear inevitably, difficult to calculate them “by hands”).
Thus, in the transverse direction the nonlinearity acts in the defocusing manner, while in the longitudinal direction it works for focusing of a wave packet. In addition, the linear dispersion makes its own defocusing contribution. Owing to the indicated factors, the behavior of a nonlinear group of waves on the two-dimensional free surface of the fluid is more diverse as compared to the one-dimensional case. Depending on initial conditions, the nonlinearity can in some cases reinforce the linear focusing, but as well it can destroy the focusing in other cases. In the case of reinforcement, a rogue wave arises. Besides that, anomalous waves can appear as a result of interaction between previously formed coherent structures [@R2007PRL]. However, for the usual oceanic wave fields where the presence of coherent structures is hardly probable, more actual seems the scenario when anomalous waves rise due to occasional spatio-temporal focusing (see [@FGD2007; @R2013], and references therein).
In general, the question about optimal conditions for the nonlinear focusing of water waves is still far from being clear. It is only obvious that, similarly to the 1D case, the initial wave group should be sufficiently high and/or extensive. But an effect of geometric shape of the packet, as well as phase modulation, on the development of anomalous wave is poorly studied. In the absence of exact (1+2D) NLSE solutions, which could describe quantitatively all such processes, it seems reasonable to carry out approximate consideration of the dynamics of an idealized solitary wave packet characterized by a few time-dependent parameters. Thereby, the above qualitative reasons could be concretized, and a reference point could be put along the direction of development of future more accurate theory of three-dimensional rogue waves.
The purpose of this work is to investigate the nonlinear dynamics of a wave packet in the framework of the full Gaussian variational ansatz, i.e. in the case when the quadratic form under the exponential has an off-diagonal part. Such a packet has an elliptical shape, with main axes oriented at some time-dependent angle with respect to the coordinate axes. Contrary to the more widely known diagonal ansatz (see, e.g., [@DAL1991; @B1994; @BR1996; @BRKST1996; @B1998; @BEC1; @BEC2; @CD1996; @H2002]), the full Gaussian ansatz was previously applied to the elliptic NLSE only [@DBDK2010; @ADO2011]. The present work fills this gap in the theory. Here a system of variational equations will be derived for parameters of the ellipse, and its full integrability will be demonstrated. The fact of integrability takes place, first, because the third power of the NLSE nonlinearity is accorded with the two spatial dimensions, and second, because Eq.(\[NLS\]) has a special integral of motion, namely the hyperbolic analog of the angular momentum, $$\int i[y(\psi_x\psi^*-\psi\psi^*_x)+x(\psi_y\psi^*-\psi\psi^*_y)]dx dy =const.$$
The knowledge of general structure of analytical solutions of the variational model will allow us to understand qualitative differences between the linear and nonlinear regimes of the wave packet focusing. Besides that, with the purpose of immediate testing of the Gaussian model, a comparison between its predictions and the results of numerical simulations of fully nonlinear long-crested waves will be carried out. We shall see that the agreement can be rather good even for strongly nonlinear regimes, when the very NLSE is not applicable already.
Variational ansatz
==================
Let us first remind that in the simplest — diagonal — variant, the Gaussian ansatz describes an elliptical wave packet having the symmetry axes coinciding with the coordinate axes (see, e.g., [@DAL1991; @B1994; @BR1996; @BRKST1996; @B1998; @BEC1]): $$\psi=\sqrt{\frac{4N}{XY}}\exp\Big[-\frac{x^2}{2X^2}-\frac{y^2}{2Y^2}
+i\frac{Ux^2}{2X} -i\frac{Vy^2}{2Y} +i\phi\Big].
\label{Gaussian}$$ Besides the longitudinal size $X(t)$ and transverse size $Y(t)$, such a packet is characterized by four more real quantities: $U$, $V$, $N$, and $\phi$. Parameter $N$ does not depend on time because $4\pi N=\int |\psi|^2 dx dy$ is an exact integral of motion of NLSE, while the conjugate variable $\phi$ is cyclic. Parameters $U$ and $V$ describe a phase modulation, with positive/negative values of $U$ or $V$ corresponding to defocusing/focusing configurations respectively in $x$ or $y$ direction. Here it is necessary to note that for applicability of NLSE, the condition of narrow spectral width of the wave packet should be satisfied. In our case it practically means the following: $(X,Y)\gtrsim 10$, $(U,V)\lesssim 0.1$. Equations of motion determining the temporal behavior of the unknown quantities $X$, $Y$, $U$, and $V$, are derived by substitution of the variational ansatz (\[Gaussian\]) into the Lagrangian functional of the NLSE: $${\cal L}=\int(i\psi_t\psi^*-i\psi\psi^*_t-|\psi_x|^2+|\psi_y|^2+|\psi|^4/2) dx dy.
\label{L_NLSE}$$ In this way we obtain a reduced Lagrangian $\tilde{\cal L}$ depending on $N$, $X(t)$, $Y(t)$, $U(t)$, and $V(t)$. The action integral $\int\tilde{\cal L}dt$ should be then variated. The main outcome of this standard procedure is the following. Variation of the reduced action gives in particular that $U=\dot X$, $V=\dot Y$. Finally we arrive at the following system of differential equations of the Newtonian type: $$\ddot X=\frac{1}{X^3}-\frac{N}{X^2Y}, \qquad \ddot Y=\frac{1}{Y^3}+\frac{N}{Y^2X}.
\label{XY_dyn}$$ Note that equations (\[XY\_dyn\]) constitute a Lagrangian system with the Lagrangian functional $L\propto \tilde{\cal L}/N$, where $$2L=\dot X^2-\dot Y^2 -\frac{1}{X^2} +\frac{1}{Y^2}+\frac{2N}{XY},
\label{L_XY}$$ so the kinetic energy is not positively-definite (the second “particle” has a negative mass).
It should be said that equations of the type (\[XY\_dyn\]) and their generalizations are actively used in plasma physics, in nonlinear optics, and in atomic physics to investigate behavior of NLSE and Gross-Pitaevskii equation solutions in two and three spatial dimensions (see papers [@DAL1991; @B1994; @BR1996; @BRKST1996; @B1998; @BEC1; @BEC2; @CD1996; @H2002], and references therein). In particular in work [@B1994], analytical and some numerical solutions were presented namely for the hyperbolic (1+2D) NLSE, as in our case, but with such parameter values that are typical for nonlinear optics, not for water waves. In relation to the problem of anomalous sea wave focusing, this simplified variational ansatz was applied only very recently in the author’s work [@R2014].
Let us now turn our attention to the full Gaussian ansatz, i.e. when off-diagonal elements of the quadratic forms are involved. For further consideration, it is convenient to introduce linear combinations of the spatial coordinates, corresponding to a hyperbolic rotation with the parameter $\chi$ (analogously to the elliptic NLSE case, but with the difference that there the rotation was usual trigonometric): $$\bar x= x\cosh \chi+ y\sinh\chi,\qquad \bar y=x\sinh\chi +y\cosh\chi.$$ Instead of expression (\[Gaussian\]) we now write $$\psi=\sqrt{\frac{4N}{XY}}\exp\Big[-\frac{\bar x^2}{2X^2}-\frac{\bar y^2}{2Y^2}
+i\frac{U\bar x^2}{2X} -i\frac{V\bar y^2}{2Y} -i\gamma \bar x\bar y +i\phi\Big].
\label{Gaussian_full}$$ Strictly speaking, with $\chi\not= 0$ the parameters $X$ and $Y$ are no longer the longitudinal and transverse sizes of the packet, but we shall continue to call them so for brevity. Parameter $\chi$ cannot be too large, otherwise the wave packet becomes spectrally wide and goes beyond the NLSE applicability domain. Practically, as it will be seen from the results of direct numerical simulation of nonlinear waves with realistic parameters, the inequality $|\chi|\lesssim 0.5$ should take place.
Note that Gaussian form (\[Gaussian\_full\]) is one of exact solutions of the corresponding [*linear*]{} Schrödinger equation, with appropriate temporal dependencies of the parameters appearing there. Consequently, at small $N$ the approximation (\[Gaussian\_full\]) is certainly valid. In practice, values $N\approx$ 2–4 are of interest. In this case, generally speaking, only qualitative agreement of the results of variational model (\[Gaussian\_full\]) with the solutions of the NLSE (and, the more so, with the completely nonlinear dynamics of waves on water) can be expected. In more detail the comparison between the variational and numerical solutions will be discussed some later. An undoubted benefit of this approximation is that it provides a semiquantitative description of the process of spatio-temporal focusing, which is one of the most probable mechanisms of the formation of rogue waves under real conditions (see [@FGD2007; @R2013], and references therein).
Having substituted the trial function (\[Gaussian\_full\]) and its partial derivatives (expressed it terms of $\bar x$ and $\bar y$) into the Lagrangian (\[L\_NLSE\]), after straightforward calculations we obtain that variable $\chi$ is cyclic, while the corresponding conserved quantity is (taking into account that $N=const$) $$M=\gamma(X^2+Y^2)=const.$$ Herewith, the temporal dependence of parameter $\chi$ is determined by the equation $$\dot\chi=M(Y^2-X^2)/(X^2+Y^2)^2.
\label{d_chi}$$ The conserved quantity $M$ is a hyperbolic analog of the angular momentum. As previously, we have $U=\dot X$, $V=\dot Y$, but the dynamics of unknown functions $X(t)$ and $Y(t)$ is described now by the following Lagrangian: $$2L_M=\dot X^2-\dot Y^2 -\frac{1}{X^2} +\frac{1}{Y^2}+\frac{2N}{XY}
+\frac{M^2(X^2-Y^2)}{(X^2+Y^2)^2}.
\label{LM_XY}$$ This Lagrangian is a hyperbolic analog of the elliptic variational problem considered in works [@DBDK2010; @ADO2011]. Equations of motion that follow from here, have the form $$\ddot X=\frac{1}{X^3}-\frac{N}{X^2Y} +\frac{M^2X}{(X^2+Y^2)^2}\Bigg[
1-2\frac{(X^2-Y^2)}{(X^2+Y^2)}\Bigg],
\label{ddX_M}$$ $$\ddot Y=\frac{1}{Y^3}+\frac{N}{Y^2X} +\frac{M^2Y}{(X^2+Y^2)^2}\Bigg[
1-2\frac{(Y^2-X^2)}{(X^2+Y^2)}\Bigg].
\label{ddY_M}$$ The finding solutions of this system and the comparison of the corresponding evolution of wave packet with a direct numerical modeling of nonlinear waves is the essence of the present work.
Integrability of the model
==========================
An important advantage of system (\[LM\_XY\]) is an exact integrability, because here in analogy with the elliptic case, the separation of variables turns out to be possible. Integrability of the variational approximation is typical namely for NLSE in 2D space, because in 3D the reduced Lagrangian instead of $2N/(XY)$ would contain $2\tilde N/(XYZ)$, and it would destroy the homogeneity of the equations. It is well known from the analytical mechanics that for integration of the above system one should use the hyperbolic coordinates: $$X=Q\sinh \xi, \qquad Y=Q\cosh \xi.$$ These coordinates describe sector $X<Y$ which is most interesting in application to the problem of oceanic rogue waves, because a length of their crest usually 3–10 times exceeds the wave length $\lambda_0=2\pi/k_0$, while the longitudinal packet size in the minimum reaches approximately one $\lambda_0$ (that corresponds to dimensionless values $X_{\rm min}\sim$ 6–10). The Lagrangian (\[LM\_XY\]) takes the form $$2L_M=Q^2\dot \xi^2-\dot Q^2 -\frac{F(\xi)}{Q^2},
\label{L_Qxi}$$ where function $F(\xi)$ is defined by the formula $$F(\xi)=\frac{4}{\sinh^2(2\xi)}-\frac{4N}{\sinh(2\xi)}+\frac{M^2}{\cosh^2(2\xi)}.
\label{F}$$ The corresponding equations of motion are $$\begin{aligned}
\ddot Q+Q\dot\xi^2+\frac{F(\xi)}{Q^3}&=&0,
\label{ddQ}
\\
2Q\dot Q\dot \xi +Q^2\ddot\xi +\frac{F'(\xi)}{2Q^2}&=&0.
\label{dd_xi}\end{aligned}$$ After multiplying Eq.(\[dd\_xi\]) by $2Q^2\dot\xi$, we obtain in its left hand side a total time-derivative, so the integral of motion follows: $$Q^4\dot\xi^2+F(\xi)=C=const.
\label{C}$$ Looking at the expression (\[F\]), it is easy to understand that at sufficiently small negative values of the integration constant $C$, equation $F(\xi)=C$ determining “turn points”, has two real positive roots ($\xi_1$ and $\xi_2$), with the motion of variable $\xi$ taking place between them. At $C\ge 0$ the “turn point” is sole.
Eq.(\[ddQ\]) is simplified to $\ddot Q=-C/Q^3$. Its general solution is written in explicit form as $$Q(t)=\sqrt{I(t-t_*)^2-C/I},
\label{Q}$$ where $t_*$ and $I$ are two more integration constants. It is easy to check that $I=-2E$, where $E$ is the energy integral of the system: $$E=(Q^2\dot \xi^2-\dot Q^2 +F(\xi)/Q^2)/2.$$ With taking into account Eq.(\[Q\]), it follows from Eq.(\[C\]) that $$\int_0^{t}\frac{dt}{I(t-t_*)^2-C/I}=\pm\int_{\xi_0}^{\xi}\frac{d\xi}{\sqrt{C-F(\xi)}},
\label{t_xi}$$ where $\xi_0$ is the fourth — the last — integration constant. Formulas (\[Q\]) and (\[t\_xi\]) provide the full solution of the problem. In particular, with $I>0$, $C<0$, the dependence (\[t\_xi\]) can be represented in the following form, $$\sinh 2\xi(t)=
P\Big(a_0+\frac{1}{\sqrt{-C}}\arctan\Big[\frac{I(t-t_*)}{\sqrt{-C}}\Big]\Big),
\label{sh2xi}$$ where an even $a$-periodic function $P(a)$ (having a minimum at $a=0$) is the inverse with respect to elliptic integral (here the integration variable $z=\sinh 2\xi$): $$P\Big(\pm\int_{s_1}^s\frac{z dz}{2\sqrt{(1+z^2)(Cz^2+4Nz-4)-M^2z^2}}\Big)\equiv s,
\label{P_def}$$ and $s_1$ is the smaller root of the equation $(1+s^2)(Cs^2+4Ns-4)-M^2s^2=0$. The amplitude of wave packet equals to $\sqrt{4N/XY}$, and it is given by the formula $$A(t)=\sqrt{8N}/\sqrt{[I(t-t_*)^2 -C/I]\sinh 2\xi(t)}.$$
Selection of the solutions
==========================
In fact, we shall not use the analytical solutions, expressed in terms of special functions, for finding the dependencies $X(t)$ and $Y(t)$ with some given initial values. Practically it is much simpler and faster to carry out highly accurate numerical simulation of the variational equations (\[ddX\_M\]) and (\[ddY\_M\]) immediately. One should also remember that the absolute accuracy in solving approximate variational problem is hardly necessary. The knowledge of the analytical structure of the solutions will be needed for a general understanding of properties of the dynamical system under consideration.
We are interested mainly in those solutions from the whole diversity (\[Q\]) and (\[t\_xi\]), which result in essentially higher wave amplitude as compared to the maximally possible linear value $A_{\rm max\, lin}=\sqrt{4N U_{-\infty} V_{-\infty}}$, where $U_{-\infty}$ and $V_{-\infty}$ are the asymptotic values of the focusing parameters at large negative times. In essence, only such events are deserving the terrible title “rogue waves”. What is important, the most interesting solutions correspond to values of the parameters $C\approx 0$, $M\approx 0$. The condition $C=0$ means that $Q^2\equiv Y^2-X^2=I(t-t_*)^2$, i. e. the plots of functions $X(t)$ and $Y(t)$ touch each other at $t=t_*$. It is easy to make sure of the above assertion by numerical simulations of Eqs.(\[ddX\_M\])-(\[ddY\_M\]) \[with fixed $N$, $U_0$, $V_0$, $Y_0$, but different $X_0$ and $M$\], and subsequent comparison of the maximal amplitudes. Typical dependencies $X(t)$ and $Y(t)$ for such a case are shown in Fig.\[N3XYt\]. Let us also note that the minimal value of the ratio $(X/Y)_{\rm min}=\tanh\xi_1$, which is reached closely to the moment of maximal amplitude, at $M=0$ and $C=0$ can be found from the simple condition $\sinh 2\xi_1=1/N$, as it is clear in view of formula (\[P\_def\]).
It is worth noting that at $M=0$ the cyclic variable $\chi$ remains constant, as it follows from Eq.(\[d\_chi\]). As we shall see soon, the physical wave fields corresponding to different values of $\chi$, strongly differ between each other. In particular, at $\chi=0$ the wave picture is symmetric with respect to sign change of $y$ coordinate, while at $\chi\not =0$ there is no any symmetry (see Figs.\[N3chi00XY\]-\[N3chi03XY\]). But in the framework of NLSE, the envelope of one solution coincides with the envelope of the other solution after a hyperbolic rotation of the dimensionless coordinates. The presence of this approximate symmetry significantly clarifies the question about optimal conditions for the formation of anomalous waves.
\
\
\
Comparison with numerical experiments
=====================================
It is clear that any simplified model without verification remains just a toy. In order to make the status of the variational approximation more distinct, it is necessary to compare it with a more accurate model. In our case, numerical experiments were carried out in the framework of the fully nonlinear long-crested wave model which is described in detail in Refs.[@RD2005; @R2010]. The computational domain was a square of the size 5 km, with the periodic boundary conditions. The length of the carrier wave was chosen $\lambda_0=125$ m. The simulations were carried out for different sets of initial parameters of the packet.
In Fig.2a and Fig.3a, examples are shown of how Gaussian initial conditions look. The “portraits” of anomalous waves developing from those conditions, are presented in Fig.2b and Fig.3b. In general, deviations from the Gaussian shape are clearly seen there. Despite the non-Gaussianity of the arising rogue waves, Figs.4-5 demonstrate that there is a qualitative agreement between the variational theory and the numerical experiment in the important aspect as the temporal dependence of the maximal amplitude of the wave packet. It is seen that Gaussian model somewhat overestimates the height of anomalous wave (especially as parameter $\chi$ increases to values about unity, when the approximate symmetry of the hyperbolic rotation becomes invalid), and it moves to the future the moment of the maximal rise. However, it is hardly reasonable to expect a detailed accordance from the simplest variational approximation. In any case, the time period of existence of anomalous wave is predicted by the Gaussian theory rather well. Also another important effect is confirmed, the increase of the transverse packet size as compared to the linear theory (see Fig.1). This effect contributes to the subsequent more fast decrease of the amplitude in comparison with the linear case.
Conclusions
===========
Thus, based on the variational analysis and on the comparison with the results of direct numerical simulations, it is natural to put forward the following thesis. In the main approximation, a typical long-crested three-dimensional oceanic rogue wave at the nonlinear stage is, from the mathematical point of view, a dynamical system with just two degrees of freedom: $X(t)$ and $Y(t)$. This system contains two constant parameters ($N$ and $M$), which are related to the integrals of motion of 1+2D NLSE and characterize a given group of waves. The system is integrable. Real sea states where one can expect anomalous waves, require values $N\approx$ 2–4. The most favorable focusing parameters for the rogue wave formation are: 1) $M\approx 0$; 2) on the initial stage of (occasional) focusing at some time moment $t_*$ the relations $X_*\approx Y_*\sim 20-40$, $-\dot X_*\approx -\dot Y_* \gtrsim$ 0.05 should take place. To estimate probability of realization of such conditions in random weakly nonlinear wave fields is the task on the future.
[99]{}
C. Kharif and E. Pelinovsky, Eur. J. Mech. B/Fluids [**22**]{}, 603 (2003).
K. Dysthe, H. E. Krogstad, and P. Müller, Annu. Rev. Fluid Mech. [**40**]{}, 287 (2008).
M. Onorato, S. Residori, U. Bortolozzo, A. Montina, and F. T. Arecchi, Physics Reports [**528**]{}, 47 (2013).
V.E.Zakharov,J.Appl.Mech.Tech.Phys.[**9**]{},190(1968).
T. B. Benjamin and J. E. Feir, J. Fluid Mech. [**27**]{}, 417 (1967).
V.E. Zakharov, A.I. Dyachenko, and O.A. Vasilyev, Eur. J. Mech. B/Fluids [**21**]{}, 283 (2002).
A.I. Dyachenko and V.E. Zakharov, JETP Letters [**81**]{}, 255 (2005).
V.E. Zakharov, A. I. Dyachenko, and O. A. Prokofiev, Eur. J. Mech. B/Fluids [**25**]{}, 677 (2006).
V. P. Ruban, JETP Letters [**95**]{}, 486 (2012).
A. Chabchoub, N. P. Hoffmann, and N. Akhmediev Phys. Rev. Lett. [**106**]{}, 204502 (2011).
A. Chabchoub, N. Hoffmann, M. Onorato, and N. Akhmediev, Phys. Rev. X [**2**]{}, 011015 (2012).
N. Akhmediev, V.M. Eleonskii, and N.E. Kulagin, Sov. Phys. JETP [**62**]{}, 894 (1985).
N. Akhmediev and V. Korneev, Theor. Math. Phys. [**69**]{}, 1089 (1986).
N.N. Akhmediev, V.M. Eleonskii, and N.E. Kulagin, Theor. Math. Phys. [**72**]{}, 809 (1987).
N. Akhmediev, A. Ankiewicz, and J. M. Soto-Crespo Phys. Rev. E [**80**]{}, 026601 (2009)
M. Erkintalo, K. Hammani, B. Kibler, C. Finot [*et al.*]{}, Phys. Rev. Lett. [**107**]{}, 253901 (2011).
M. Onorato, A.R. Osborne, and M. Serio, Phys. Rev. Lett. [**96**]{}, 014503 (2006).
M. Onorato, T. Waseda, A. Toffoli, L. Cavaleri [*et al.*]{}, Phys. Rev. Lett. [**102**]{}, 114502 (2009).
V.P. Ruban, Phys. Rev. E [**74**]{}, 036305 (2006).
V.P. Ruban, Phys. Rev. Lett. [**99**]{}, 044502 (2007).
D. H. Peregrine, Adv. Appl. Mech. [**16**]{}, 9 (1976).
I.V. Lavrenov and A.V. Porubov, Eur. J. Mech. B/Fluids [**25**]{}, 574 (2006).
V. I. Shrira and A. V. Slunyaev, Phys. Rev. E [**89**]{}, 041002(R) (2014).
C. Fochesato, S. Grilli, and F. Dias, Wave Motion [**44**]{}, 395 (2007).
V. P. Ruban, JETP Letters [**97**]{}, 686 (2013).
M. Desaix, D. Anderson, and M. Lisak, J. Opt. Soc. Am. B [**8**]{}, 2082 (1991).
L. Bergé, Phys. Lett. A [**189**]{}, 290 (1994).
L. Bergé and J. J. Rasmussen, Phys. Rev. E [**53**]{}, 4476 (1996).
L. Bergé, J. J. Rasmussen, E. A. Kuznetsov, E. G. Shapiro, and S. K. Turitsyn, J. Opt. Soc. Am. B [**13**]{}, 1879 (1996).
L. Berge, Phys. Reports [**303**]{}, 259 (1998).
V. M. Pérez-Garcia, H. Michinel, J. I. Cirac, M. Lewenstein, and P. Zoller, Phys. Rev. Lett. [**77**]{}, 5320 (1996); Phys. Rev. A [**56**]{}, 1424 (1997).
Yu. Kagan, E. L. Surkov, and G. V. Shlyapnikov, Phys. Rev. A [**55**]{}, R18 (1997).
Y. Castin and R. Dum, Phys. Rev. Lett. [**77**]{}, 5315 (1996).
F. Haas, Phys. Rev. A [**65**]{}, 033603 (2002).
A. S. Desyatnikov, D. Buccoliero, M. R. Dennis, and Yu. S. Kivshar, Phys. Rev. Lett. [**104**]{}, 053902 (2010).
J. Abdullaev, A. S. Desyatnikov, and E. A. Ostrovskaya, J. Opt. [**13**]{}, 064023 (2011).
V. P. Ruban, JETP Letters [**100**]{}, 751 (2014).
V. P. Ruban and J. Dreher, Phys. Rev. E [**72**]{}, 066303 (2005).
V. P. Ruban, Eur. Phys. J. Special Topics [**185**]{}, 17 (2010).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We demonstrate slowing and longitudinal cooling of a supersonic beam of CaF molecules using counter-propagating laser light resonant with a closed rotational and almost closed vibrational transition. A group of molecules are decelerated by about 20m/s by applying light of a fixed frequency for 1.8ms. Their velocity spread is reduced, corresponding to a final temperature of about 300mK. The velocity is further reduced by chirping the frequency of the light to keep it in resonance as the molecules slow down.'
author:
- 'V. Zhelyazkova'
- 'A. Cournol'
- 'T. E. Wall'
- 'A. Matsushima'
- 'J. J. Hudson'
- 'E. A. Hinds'
- 'M. R. Tarbutt'
- 'B. E. Sauer'
title: Laser cooling and slowing of CaF molecules
---
There is currently great interest in cooling and controlling molecules, motivated by a wide range of applications [@Carr(1)09]. Polar molecules interact through strong, long-range, anisotropic, and controllable dipole-dipole interactions, and so a gas of molecules, at low temperature and under precise control, could be used as a simulator of strongly-interacting quantum systems [@Goral02; @Micheli06], or for quantum computation [@DeMille02; @Andre06]. Molecules are already used in several tests of fundamental physics, such as the measurement of the electron’s electric dipole moment [@Hudson(1)11; @Eckel(1)13; @Vutha(1)10], searches for changes in fundamental constants [@Hudson(1)06; @Truppe(1)13], and tests of parity violation [@DeMille(1)08; @Darquie(1)10]. The precision of these measurements can be improved by cooling the molecules to low temperatures [@Tarbutt(1)09; @Tar13a]. The availability of cold molecules will also open many new possibilities to study and control chemistry at low temperatures [@Balakrishnan01; @Krems08]. Ultracold molecules have been produced by binding together ultracold atoms, either by photoassociation [@Sage(1)05] or magnetoassociation [@Ni(1)08; @Danzl(1)10]. Other molecules have been cooled to about 1K using a cold buffer gas [@Weinstein(1)98], and beams of molecules have been decelerated and trapped using electric [@Bethlem(1)99; @Bethlem(1)00; @vanVeldhoven(1)05], magnetic [@Vanhaecke(1)07; @Narevicius(1)08; @Sawyer(1)07] and optical [@Fulton(1)06] fields. Once trapped, evaporative cooling to lower temperatures has been demonstrated [@Stuhl(1)12], and sympathetic cooling has been studied [@Parazzoli(1)11; @Tokunaga(1)11]. Following an initial demonstration of the radiative force acting on SrF [@Shuman(1)09], transverse laser cooling was applied to beams of SrF [@Shuman(1)10] and YO [@Hummon(1)13] molecules. Radiation pressure has been used to slow down beams of SrF [@Barry(1)12], but longitudinal cooling of a molecular beam has not previously been shown. Here we report longitudinal laser cooling and slowing of a supersonic beam of CaF molecules.
Figure \[hyperfineScheme\] shows the cooling scheme. We use $v$ to label the vibrational quantum number, and $N$, $J$ and $F$ for the rotational, total electronic, and total angular momentum quantum numbers. We excite the lowest-lying vibrational and rotational state of positive parity in the first electronically excited state, $\text{A} ^{2}\Pi_{1/2} (v'=0,J'=1/2)$, whose lifetime is 19.2ns [@WallFC]. Here, the hyperfine interaction results in two levels with $F=0$ and 1, separated by approximately $4.8$MHz [@WallFC], too small to be resolved in the experiment. Due to the angular momentum and parity selection rules for electric dipole transitions, the population can only decay to the states $\text{X} ^{2}\Sigma^{+} (v'', N''=1)$. There is no restriction on the allowed values of $v''$, but the relative probabilities are governed by the Franck-Condon factors which are approximately 97% for $v''=0$, 3% for $v''=1$, 0.08% for $v''=2$, and negligibly small for all $v''>2$ (see [@WallFC; @pelegrini] and the present paper). The spin-rotation and hyperfine interactions split each of these states into four components, separated by radio-frequency intervals, as shown in Fig.\[hyperfineScheme\]. We use two lasers, each modulated to provide the four frequencies needed to drive the transitions from $v''=0$ and $v''=1$, labelled $v_{00}$ and $v_{10}$. From this laser light, a molecule scatters 1000-2000 photons on average, before it decays to $v''=2$. Since each photon absorption reduces the velocity by $h/(M \lambda)$=0.011 m/s ($M$ is the molecular mass, $\lambda$ the laser wavelength), substantial changes in velocity are possible.
Figure \[fig:schematic\] shows the experimental setup. A pulsed beam of CaF molecules is produced by laser ablation of a Ca target into a supersonically expanding pulsed jet of Ar and SF$_6$ [@WallFC; @Tarbutt(1)02]. The target is at $z=0$, and the ablation laser fires at $t=0$. The pulses have a mean velocity of 600m/s and a translational temperature of 3K. The beam passes through a 2mm diameter skimmer at $z=70$mm. At $z=L=1.63$m, molecules in any of the $\text{X} ^{2}\Sigma^{+} (v''=0,1,2; N''=1)$ states can be detected by laser-induced fluorescence using a single-frequency probe laser beam that intersects the molecular beam at right angles. The fluorescence signal is recorded with a time resolution of 10$\mu$s, giving the time-of-flight (ToF) profile of each molecular pulse. Between pulses, the probe laser frequency is stepped by approximately 0.6MHz, so that over a sequence of pulses the frequency is scanned over the four hyperfine components of the P(1) rotational line of the $\text{A} ^{2}\Pi_{1/2} (v) - \text{X} ^{2}\Sigma^{+} (v)$ transition, with $v=0$, 1 or 2 selected by the choice of laser wavelength. We refer to this transition as $\text{A} - \text{X} (v-v)$
![(Color online) Schematic of the experimental setup.[]{data-label="fig:schematic"}](drawing2.eps)
![(Color online) Spectrum showing the hyperfine structure of the $\text{A} - \text{X} (0-0)$ transition both with (solid blue line) and without (dashed red line) the cooling lasers applied. Dots are experimental data and the lines are fits to four Gaussians. The scheme for generating the frequency sidebands using acousto- and electro-optic modulators is shown.[]{data-label="spectrum"}](spectrum.eps){width="45.00000%"}
Two cw dye lasers generate the laser cooling light, which counter-propagates to the molecular beam. The rf sidebands addressing the various hyperfine transitions are generated using a combination of an acousto-optic modulator (AOM) and an electro-optic modulator (EOM) as shown in Fig.\[spectrum\]. An additional AOM used as a fast switch turns on the $v_{00}$ cooling light at $t_{\text{start}}=200$$\mu$s for a variable duration, $\tau$. The laser beams are then spatially overlapped and double-pass through a final AOM which applies a frequency chirp, $\beta$, to all frequencies simultaneously. This chirp is used to compensate for the changing Doppler shift of the molecules as they slow down. The overlapped beams are shaped to an approximately Gaussian intensity distribution with $1/e^2$ radii of 3.81mm at the position of the detector and 0.78mm at the skimmer. The total power is 32mW, divided approximately equally between the eight frequencies. A mechanical shutter blocks the $v_{00}$ beam on alternate shots of the experiment, so that successive shots record the change in the ToF profile when the cooling is applied, minimizing sensitivity to slow drifts in the molecular flux. The $v_{10}$ light is always applied continuously so that population in $v''=1$ is transferred to $v''=0$. To avoid optically pumping molecules into Zeeman sub-states that do not couple with the linearly polarized light, we apply a $\approx$10G magnetic field orthogonal to the molecular beam and at 45$^\circ$ to the laser polarization [@Berkeland2002; @Shuman(1)09]. We study the slowing and cooling of the molecules as a function of cooling time, $\tau$, and frequency chirp, $\beta$.
![(Color online) Effect of laser cooling, without frequency chirp, for various durations, $\tau=0.1,0.5,1.0,1.4$ and 1.8ms. (a) Experimental ToF profiles with cooling off (red) and on (blue). Dashed purple lines show the best-fit ToF profiles predicted from the known initial velocity distribution and the simple model for the velocity-dependent force discussed in the text. (b) Velocity profiles inferred from the measured ToF profiles in (a) using this same simple model, with cooling off (solid red) and on (dashed purple). (c) Simulated ToF profiles and (d) velocity profiles, with cooling off (solid red) and on (dashed purple). All profiles are summed over $v''=0$ and $v''=2$ populations.[]{data-label="VariableCooling"}](variableCoolingTime.eps){width="50.00000%"}
Figure \[spectrum\] shows how the $\text{A} - \text{X} (0-0)$ spectrum changes when laser cooling is applied, with $\tau=1.8$ms and $\beta=0$. Some signal is lost because the light optically pumps molecules into $v''=2$, but this reappears in the equivalent $\text{A} - \text{X} (2-2)$ spectrum. In addition, the relative peak heights change because the light re-distributes the population amongst the hyperfine states. To obtain the ToF profile we integrate over this spectrum, applying a different weighting factor to each hyperfine component that accounts for their differing photon yields. These weighting factors are calculated from the spectrum measured without cooling, where we assume that each hyperfine level is populated according to its degeneracy. This procedure ensures that the ToF profiles are not complicated by the effects of velocity-selective hyperfine pumping.
Figure \[VariableCooling\](a) shows the ToF profiles when the lasers are tuned to maximize the scattering rate from molecules moving at 600m/s. The symmetrical (red) curves are profiles measured without laser cooling ($\tau=0$), while the other (blue) solid curves are for various nonzero values of $\tau$. In all cases, we sum the populations measured in $v''=0$ and $v''=2$. No population remains in $v''=1$ because the $v_{10}$ light, applied continuously, pumps these molecules to $v''=0$. These profiles show that molecules removed from the original distribution are both slowed and cooled, to produce a peak that is both later and narrower. The width of the peak together with the relation $v=L/t$, gives an immediate estimate of the velocity distribution and hence of the temperature. This is not accurate because the molecules are decelerated, but it does provide an upper limit on the temperature, which we find to be 430mK for the ToF peak at $\tau=1.8\,$ms. The actual temperature is lower, as we now discuss.
Because the source is spatially compact, the ToF profile measured without laser cooling is a direct measure of the velocity distribution produced by the source. Using a simple model for the light force, we can easily convert this initial velocity distribution into the ToF profile expected when the cooling is applied. We approximate the light force as a function of velocity $v$ by $F\exp(-(\tfrac{v-v_0}{\Delta})^{2})$. A single choice of the fitting parameters $F$, $\Delta$ and $v_0$ reproduces all the profiles measured, as shown by the dashed lines in Fig. \[VariableCooling\](a). With $F$, $\Delta$ and $v_0$ now determined, we use the same model to convert the measured ToF profiles into final velocity distributions. These are plotted in Fig. \[VariableCooling\](b). We see that the peaks in the ToF profiles do indeed correspond to a slowing and a narrowing of the velocity distribution. On fitting the $\tau=1.8$ms curve in Fig. \[VariableCooling\](b) to the sum of two Gaussians, we find that the final speed of the cooled bunch is $583\pm2\,$m/s and its temperature is $330\pm70\,$mK. The exact form of the light force in our model is not important: any peak-shaped function will do, for example a Lorentzian gives virtually the same final velocity and temperature as the Gaussian. We are therefore confident of our conclusion that these molecules are both decelerated and cooled.
One should consider whether our TOF profiles could be produced by a position-dependent force that longitudinally focusses the molecules onto the detector, rather than a velocity-dependent force that cools the beam. We have ruled this out in two ways. We have modelled the focussing effect of various, suitably-contrived, position-dependent forces, constrained only by the maximum possible scattering rate. Even in the most extreme cases, we find the effect to be far too weak to produce the narrowing we observe. We have also checked experimentally for any position-dependence of the scattering rate by measuring the optical pumping rate into $v''=1$ when the $v_{00}$ light is pulsed on for a short period. These data shows that the scattering rate has exactly the value and velocity-dependence we would expect, and has no significant position dependence. We use a camera to ensure that the $v_{10}$ and $v_{00}$ beams have the same spatial profiles and are well overlapped.
The best fit parameters in our model have the values $v_0=599\pm3$m/s, $F=(0.012\pm0.003) h \Gamma/ \lambda$, and $\Delta = 21\,^{+10}_{-4}$m/s, all of which are reasonable. The value of $v_0$ is exactly as expected, this being the target value in the experiment. We expect the force to be $P_{\textrm{ex}} h \Gamma/ \lambda$, where $P_{\textrm{ex}} $ is the probability of being in the excited state. Using a simple formula for excitation in this multi-level system [@Tar13a], taking an average value for the laser intensity, and accounting for the small detunings of some frequencies from their ideal values, we find $P_{\textrm{ex}} \simeq 0.04$. However, a substantial fraction of the decelerated molecules are pumped into the $v''=2$ state, at which point the force turns off. This is responsible for the lower average force that we infer, as discussed further below. The value for $\Delta$ is somewhat higher than we estimate from the power broadened linewidth of the transition. This can also be caused by optical pumping into $v''=2$ because there is an effective broadening of the transition associated with the saturation of the scattered photon number.
In order to understand the dynamics of the laser cooled molecules in more detail, we have modelled the experiment using a set of 33 coupled rate equations. For each molecule, we solve these equations numerically to follow the populations of the 24 ground-state Zeeman sub-levels and the 4 excited sub-levels, the axial and radial positions and speeds, and the total number of scattered photons. The excitation rate, the damping rate, and the formulae for calculating the branching ratios, are taken from [@WallFC]. The excitation rate is summed over the laser frequency components, each with its appropriate intensity and detuning. The laser beam has a Gaussian intensity distribution with parameters as measured in the experiment. The magnetic field, $B$, results in a rotation of the state vector about the field direction at the rate $\omega_{B}=g \mu_{B} B /\hbar$, but this coherent evolution is strongly damped in the experiment because the ground states are coupled to the short-lived excited state. To model this, the rate equations include terms that damp out population difference between states mixed by $B$ at the characteristic rate $\omega_{B}$. These are states of the same $J$ and $F$ that differ in $M_F$ by $\pm 1$. The Franck-Condon factors, $Z_{v'',v'}$ are set to $Z_{0,0}=0.972$ and $Z_{1,0}=1-Z_{0,0}$. The simulation continues until the molecule decays to $v''=2$. This happens after scattering $n$ photons where, for each molecule, $n$ is chosen at random from the probability distribution $Z_{2,0}(1-Z_{2,0})^n$, with $Z_{2,0}=7.84\times 10^{-4}$. We chose $Z_{2,0}$ so that the simulated fraction of molecules pumped into $v''=2$ best matches the experimental observations. The source emits a Gaussian distribution of forward speeds with a mean of 600m/s and a temperature of 3.1K, while the initial distribution along $z$ is Gaussian with a full width at half maximum (FWHM) of 20mm. The initial radial position and speed distributions have widths of 1.2mm and 24m/s FWHM. The simulation results are insensitive to these values since only molecules with small radial displacements and speeds are detected.
![(Color online) ToF profiles for molecules in $v''=0$ (solid purple line) and $v''=2$ (dashed orange line) after 1.8ms of cooling time without a frequency chirp. Dotted line shows the central arrival time without cooling. (a) Experiment. (b) Simulation.[]{data-label="PerState"}](perState.eps){width="50.00000%"}
Figure \[VariableCooling\](c) shows the simulated ToF profiles for our experimental parameters. The arrival time of the cooled peak matches experiment very well in all cases, indicating that the mean force is correctly described by the simulation. The simulated peaks are narrower than in the experiment, and are more prominent than in the experiment for the lower values of $\tau$, suggesting that the cooling is not as strong as the simulation predicts. The same conclusion follows from a comparison of the simulated velocity distributions, shown in (d), with the distributions inferred from our measurements shown in (b). In the simulation for $\tau=1.8$ms, the decelerated peak has a final speed of 583m/s, in agreement with experiment, while the final temperature is 85mK - somewhat lower than the 330mK that we infer from our data.
![(Color online) Effect of laser cooling with various frequency chirps, $\beta=0,10,20$ and 30MHz/ms, when $\tau=1.8$ms. Dashed red curves: cooling off. Solid blue curves: cooling on. (a) Experimental ToF profiles. (b) Simulated ToF profiles. All profiles are summed over $v''=0,1$ and 2 populations.[]{data-label="chirping"}](chirpingData.eps){width="45.00000%"}
We have investigated the pumping of molecules into $v''=2$ by the cooling lasers. In the experiment, this fraction gradually increases from 6% of the total when $\tau=0.1$ms to 33% when $\tau=1.8$ms. Figure \[PerState\] shows measured and simulated ToF profiles for molecules in $v''=0$ and $v''=2$ when $\tau=1.8$ms. The strong dip in the $v''=0$ profile, close to 2.7ms, is due to the interaction with the light. Some of these missing molecules have been decelerated and appear at later arrival times, while the rest have been optically pumped and appear in the $v''=2$ profile. The latter have also been slowed on average, though not by as much as the ones that remain in the cooling cycle for the entire period. All these details are reproduced by the simulation. The dip in the simulated profile is slightly deeper than we measure, and this difference is more pronounced for shorter values of $\tau$ (not shown), however, the arrival times and amplitudes of the peaks in both the $v''=0$ and $v''=2$ profiles match well for all $\tau$.
After $\tau=1.8$ms of deceleration, the Doppler shift is so large (28MHz or $3.4\Gamma$) that the cooling ceases. This shift can be compensated by chirping the frequencies of the lasers. Figures \[chirping\](a) and (b) show the measured and simulated ToF profiles for various chirp rates, $\beta$, with $\tau=1.8$ms. (Here, we include the population in $v''=1$ as well as 0 and 2 because the chirp reduces the optical pumping efficiency of the $v_{10}$ laser, particularly at early arrival times). We see that the chirp does indeed slow the molecules further, as intended, and that the delayed arrival times agree well with the simulation results. However, the measured peak height does not agree with the simulations, being noticeably smaller when the chirp rate is high. We do not yet know what causes this difference. A chirp rate in excess of 30MHz/ms produces little further deceleration. This is not surprising because, for $P_{\textrm{ex}} =0.04$, the critical chirp $P_{\textrm{ex}} h \Gamma /(M\lambda^2)$ that exactly matches the changing Doppler shift is 38MHz/ms.
In summary, we have shown that the scattering of laser light is able to slow and cool a molecular beam of CaF molecules. Without chirping, the molecules were slowed by up to 17m/s and cooled to 330mK. With chirping the deceleration was roughly doubled. Comparison with a detailed numerical model shows that the deceleration is understood, as is the optical pumping into the $v''=2$ state, while the cooling is not quite as strong as predicted. An additional laser could be used to close the leak to $v''=2$, and then we expect to cool the molecules to the Doppler temperature or below [@Shuman(1)09; @Tar13a]. Recent advances in doubled fiber laser and diode laser technology [@Bal13a] will make this easier to achieve. The cooling presented here would increase the phase-space density of CaF molecules slowed in a Stark decelerator [@Wall(1)11], or a travelling-wave decelerator [@Bulleid(1)12] by a factor of 1000. The recent development of cryogenic sources can provide molecular beams with mean speeds of about 50m/s [@Lu11a], which could then be laser-slowed to rest and captured in an optical molasses or magneto-optical trap [@Stuhl(1)08]. With its favourable Franck-Condon factors and large electric and magnetic dipole moments, CaF is ideal for exploring the physics of strongly-interacting many-body quantum systems.
This work was supported by the EPSRC and the Royal Society.
[99]{}
L. D. Carr, D. DeMille, R. V. Krems and J. Ye, New. J. Phys. **11**, 055049 (2009).
K. Goral, L. Santos, and M. Lewenstein, Phys. Rev. Lett. **88**, 170406 (2002).
A. Micheli, G. Brennen, and P. Zoller, Nature Phys. **2**, 341 (2006).
D. DeMille, Phys. Rev. Lett. **88**, 067901 (2002).
A. André. [*et al.*]{}, Nature Phys. **2**, 636 (2006).
J. J. Hudson, D. M. Kara, I. J. Smallman, B. E. Sauer, M. R. Tarbutt and E. A. Hinds, Nature **473**, 493 (2011).
S. Eckel, P. Hamilton, E. Kirilov, H. W. Smith and D. DeMille, arXiv:1303.3075 (2013).
A. C. Vutha, W. C. Campbell, Y. V. Gurevich, N. R. Hutzler, M. Parsons, D. Patterson, E. Petrik, B. Spaun, J. M. Doyle, G. Gabrielse and D. DeMille, J. Phys. B **43**, 074007 (2010).
E. R. Hudson, H. J. Lewandowski, B. C. Sawyer, and J. Ye, Phys. Rev. Lett. **96**, 143004 (2006).
S. Truppe, R. J. Hendricks, S. K. Tokunaga, H. J. Lewandowski, M. G. Kozlov, C. Henkel, E. A. Hinds and M. R. Tarbutt, Submitted (2013).
D. DeMille, S.B. Cahn, D. Murphree, D.A. Rahmlow, M.G. Kozlov, Phys. Rev. Lett. 100, 023003 (2008).
B. Darquié, *et al.*, Chirality **22**, 870 (2010).
M. R. Tarbutt, J. J. Hudson, B. E. Sauer and E. A. Hinds, Faraday Discuss. **142**, 37-56 (2009).
M. R. Tarbutt, B. E. Sauer, J. J. Hudson and E. A. Hinds, New J. Phys. **15**, 053034 (2013).
N. Balakrishnan and A. Dalgarno, Chem. Phys. Lett. **341**, 652 (2001).
R. V. Krems, Phys. Chem. Chem. Phys. **10**, 4079 (2008).
J. M. Sage, S. Sainis, T. Bergeman and D. DeMille, Phys. Rev. Lett. **94**, 203001 (2005).
K.-K. Ni, S. Ospelkaus, M. H. G. de Miranda, A. Pe’er, B. Neyenhuis, J. J. Zirbel, S. Kotochigova, P. S. Julienne, D. S. Jin and J. Ye, Science **322**, 231 (2008).
J. G. Danzl, M. J. Mark, E. Haller, M. Gustavsson, R. Hart, J. Aldegunde, J. M. Hutson, H.-C. Nägerl, Nature Phys. **6**, 265 (2010).
J. D. Weinstein, R. deCarvalho, T. Guillet, B. Friedrich and J. M. Doyle, Nature **395**, 148 (1998).
H.L. Bethlem, G. Berden and G. Meijer, Phys. Rev. Lett. **83**, 1558 (1999).
H.L. Bethlem, G. Berden, F.M.H. Crompvoets, R.T. Jongma, A.J.A. van Roij and G. Meijer, Nature **406**, 491 (2000).
J. van Veldhoven, H.L. Bethlem and G. Meijer, Phys. Rev. Lett. **94**, 083001 (2005).
N. Vanhaecke, U. Meier, M. Andrist, B. H. Meier and F. Merkt, Phys. Rev. A **75**, 031402(R) (2007).
E. Narevicius et al., Phys. Rev. A **77**, 051401(R) (2008).
B.C. Sawyer, B.L. Lev, E.R. Hudson, B.K. Stuhl, M. Lara, J.L. Bohn and J. Ye, Phys. Rev. Lett. **98**, 253002 (2007).
R. Fulton, A. I. Bishop, M. N. Shneider and P. F. Barker, Nature Physics 2, 465 (2006).
B. K. Stuhl, M. T. Hummon, M. Yeo, G. Quemeńer, J. L. Bohn, and J. Ye, Nature **492**, 396 (2012).
L. P. Parazzoli, N. J. Fitch, P. S. [Ż]{}uchowski, J.M. Hutson and H.J. Lewandowski (2011), arXiv:1101.2886
S. K. Tokunaga, W. Skomorowski, P. S. [Ż]{}uchowski, R. Moszynski, J. M. Hutson, E. A. Hinds and M. R. Tarbutt, Eur. Phys. J. D **65**, 141 (2011).
E. S. Shuman, J. F. Barry, D. R. Glenn, and D. DeMille, Phys. Rev. Lett. **103**, 223001 (2009).
E. S. Shuman, J. F. Barry, and D. DeMille, Nature **467**, 820 (2010).
M. T. Hummon, M. Yeo, B. K. Stuhl, A. L. Collopy, Y. Xia, and J. Ye, Phys. Rev. Lett. [**110**]{}, 143001 (2013).
J. F. Barry, E. S. Shuman, E .B. Norrgard, and D. DeMille Phys. Rev. Lett. **108**, 103002 (2012).
T.E Wall, J.F. Kanem, J.J. Hudson, B.E. Sauer, D. Cho, M.G. Boshier, E.A. Hinds, and M.R. Tarbutt Physical Review A [**78**]{}, 062509-062519 (2008).
M. Pelegrini, C.S. Vivacqua, O. Roberto-Neto, F.R. Ornellas, and F.B.C Machado, Brazilian Journal of Physics [**35**]{}, 950-956 (2005).
M. R. Tarbutt, J. J. Hudson, B. E. Sauer, E. A. Hinds, V. A. Ryzhov, V. L. Ryabov, V. F. Ezhov, J. Phys. B **35**, 5013 (2002). D. J. Berkeland and M. G. Boshier, Phys. Rev. A **65**, 033413 (2002).
H. Ball, M. W. Lee, S. D. Gensemer and M. J. Biercuk, Rev. Sci. Inst. [**84**]{}, 063107 (2013).
T. E. Wall, J. F. Kanem, J. M. Dyne, J. J. Hudson, B. E. Sauer, E. A. Hinds and M. R. Tarbutt, Phys. Chem. Chem. Phys. **13**, 18991 (2011).
N. E. Bulleid, R. J. Hendricks, E. A. Hinds, S. A. Meek, G. Meijer, A. Osterwalder, and M. R. Tarbutt, Phys. Rev. A **86**, 021404(R) (2012).
H-I Lu, J. Rasmussen, M. Wright, D. Patterson and J. M. Doyle, Phys. Chem. Chem. Phys. [**13**]{}, 18986 (2011).
B. K. Stuhl, B.C. Sawyer, D. Wang, and J. Ye, Phys. Rev. Lett. [**101**]{}, 243002 (2008).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
Let $T$ be a smooth homogeneous Calderón-Zygmund singular integral operator in ${{\mathbb R}^n}$. In this paper we study the problem of controlling the maximal singular integral $T^{\star}f$ by the singular integral $Tf$. The most basic form of control one may consider is the estimate of the $L^2({{\mathbb R}^n})$ norm of $T^{\star}f$ by a constant times the $L^2({{\mathbb R}^n})$ norm of $Tf$. We show that if $T$ is an even higher order Riesz transform, then one has the stronger pointwise inequality $T^{\star}f(x) \leq C \, M(Tf)(x)$, where $C$ is a constant and $M$ is the Hardy-Littlewood maximal operator. We prove that the $L^2$ estimate of $T^{\star}$ by $T$ is equivalent, for even smooth homogeneous Calderón-Zygmund operators, to the pointwise inequality between $T^{\star}$ and $M(T)$. Our main result characterizes the $L^2$ and pointwise inequalities in terms of an algebraic condition expressed in terms of the kernel $\frac{\Omega(x)}{|x|^n}$ of $T$, where $\Omega$ is an even homogeneous function of degree $0$, of class $C^\infty(S^{n-1})$ and with zero integral on the unit sphere $S^{n-1}$. Let $\Omega= \sum
P_j$ the expansion of $\Omega$ in spherical harmonics $P_j$ of degree $j$. Let $A$ stand for the algebra generated by the identity and the smooth homogeneous Calderón-Zygmund operators. Then our characterizing condition states that $T$ is of the form $R\circ U$, where $U$ is an invertible operator in $A$ and $R$ is a higher order Riesz transform associated with a homogeneous harmonic polynomial $P$ which divides each $P_j$ in the ring of polynomials in $n$ variables with real coefficients.
author:
-
title: '[**Estimates for the maximal singular integral in terms of the singular integral: the case of even kernels**]{}'
---
Introduction
============
Let $T$ be a smooth homogeneous Calderón-Zygmund singular integral operator on ${{\mathbb R}^n}$ with kernel $$\label{eq1}
K(x)=\frac{\Omega(x)}{|x|^n},\quad x \in {{\mathbb R}^n}\setminus \{0\}\,,$$ where $\Omega$ is a (real valued) homogeneous function of degree $0$ whose restriction to the unit sphere $S^{n-1}$ is of class $C^\infty(S^{n-1})$ and satisfies the cancellation property $$\int_{|x|=1} \Omega(x)\,d\sigma(x)=0\,,$$ $\sigma$ being the normalized surface measure on $S^{n-1}$. Recall that $Tf$ is the principal value convolution operator $$\label{eq2}
Tf(x)= P.V. \int f(x-y)\,K(y)\,dy \equiv \lim_{\epsilon\rightarrow
0} T^\epsilon f(x)\,,$$ where $T^\epsilon$ is the truncation at level $\epsilon$ defined by $$T^{\epsilon}f(x)= \int_{| y-x| > \epsilon} f(x-y) K(y) \,dy\,.$$ As we know, the limit in exists for almost all $x$ for $f$ in $L^p({{\mathbb R}^n})$, $1 \leq p < \infty$.
Let $T^{\star}$ be the maximal singular integral $$T^{\star}f(x)= \sup_{\epsilon > 0} | T^{\epsilon}f(x)|, \quad x \in
{{\mathbb R}^n}\,.$$
In this paper we consider the problem of controlling $T^{\star} f$ by $Tf$. The most basic form of control one may think of is the $L^2$ estimate $$\label{eq3}
\| T^{\star} f \|_2 \leq C \| T f \|_2,\quad f \in L^2({{\mathbb R}^n}) \,.$$ Another way of saying that $T^{\star} f$ is dominated by $Tf$, apparently much stronger, is provided by the pointwise inequality $$\label{eq4}
T^{\star}f(x) \leq C \, M(Tf)(x),\quad x \in {{\mathbb R}^n}\,,$$ where $M$ denotes the Hardy-Littlewood maximal operator. Notice that may be viewed as an improved version of classical Cotlar’s inequality $$\label{}
T^{\star}f(x) \leq C \left(M(Tf)(x) + Mf (x)\right),\quad x \in
{{\mathbb R}^n}\,,$$ because the term involving $Mf$ is missing in the right hand side of .
We prove that if $T$ is an even higher order Riesz transform, then holds. Recall that $T$ is a higher order Riesz transform if its kernel is given by a function $\Omega$ of the form $$\Omega(x)=\frac{P(x)}{|x|^d}, \quad x \in {{\mathbb R}^n}\setminus \{0\}\,,$$ with $P$ a homogeneous harmonic polynomial of degree $d \geq 1$. If $P(x)= x_j$, then one obtains the $j$-th Riesz transform $R_j$. If the homogeneous polynomial $P$ is not required to be harmonic, but has still zero integral on the unit sphere, then we call $T$ a polynomial operator.
Thus, if $T=R$ is an even higher order Riesz transform, one has the weak $L^1$ type inequality $$\label{eq5}
\|R^\star f \|_{1,\infty} \leq C\| Rf \|_1 \,,$$ which combined with the classical weak $L^1$ type estimate $$\|R^\star f \|_{1,\infty} \leq C\| f \|_1 \,,$$ yields the sharp inequality $$\| R^{\star} f \|_{1,\infty} \leq C \min \{\| f \|_1, \| Rf
\|_1\}\,.$$ In [@MV2] one proved for the Beurling transform in the plane and also that fails for the Riesz transforms $R_j$. Therefore the assumption that the operator is even is crucial.
The question of estimating $T^{\star}f$ by $Tf$ was first raised in [@MV2]. The problem originated in an attempt to gain a better understanding of how one can obtain a. e. existence of principal values of truncated singular integrals from $L^2$ boundedness, for underlying measures more general than the Lebesgue measure in ${{\mathbb R}^n}$. This in turn is motivated by a problem of David and Semmes [@DS], which consists in deriving uniform rectifiability of a $d$-dimensional Ahlfors regular subset of ${{\mathbb R}^n}$ from the $L^2$ boundedness of the Riesz kernel of homogeneity $-d$ with respect to $d$-dimensional Hausdorff measure on the set . For more details on that see the last section of [@MaV].
Our main result states that for even operators inequalities and are equivalent to an algebraic condition involving the expansion of $\Omega$ in spherical harmonics. This condition may be very easily checked in practice and so, in particular, we can produce extremely simple examples of even polynomial operators for which and fail. For these operators no control of $T^{\star}f$ by $Tf$ seems to be known. To state our main result we need to introduce a piece of notation.
Recall that $\Omega$ has an expansion in spherical harmonics, that is, $$\label{eq6}
\Omega(x) = \sum_{j=1}^\infty P_j(x), \quad x \in S^{n-1}\,,$$ where $P_j$ is a homogeneous harmonic polynomial of degree $j$. If $\Omega$ is even, then only the $P_j$ of even degree $j$ may be non-zero.
An important role in this paper will be played by the algebra $A$ consisting of the bounded operators on $L^2({{\mathbb R}^n})$ of the form $$\lambda I + S\,,$$ where $\lambda$ is a real number and $S$ a smooth homogeneous Calderón-Zygmund operator.
Our main result reads as follows.
\[T\] Let $T$ be an even smooth homogeneous Calderón-Zygmund operator with kernel and assume that $\Omega$ has the expansion . Then the following are equivalent.
1. $$T^{\star}f(x) \leq C \, M(Tf)(x),\quad x \in {{\mathbb R}^n}\,.$$
2. $$\| T^{\star} f \|_2 \leq C \| T f \|_2,\quad f \in L^2({{\mathbb R}^n}) \,.$$
3. The operator $T$ can be factorized as $T=R\circ U$, where $U$ is an invertible operator in the algebra $A$ and $R$ is a higher order Riesz transform associated with a harmonic homogeneous polynomial $P$ which divides each $P_j$ in the ring of polynomials in $n$ variables with real coefficients.
Two remarks are in order.
Observe that condition $(iii)$ is algebraic in nature. This is one of the reasons that makes the proof difficult. Condition $(iii)$ can be reformulated in a more concrete fashion as follows. Assume that the expansion of $\Omega$ in spherical harmonics is $$\Omega(x) = \sum_{j=j_0}^\infty P_{2j}(x),\quad P_{2j_0} \neq 0\,.$$ Then $(iii)$ is equivalent to the following
1. [*For each $j$ there exists a homogeneous polynomial $Q_{2j-2j_0}$ of degree $2j-2j_0$ such that $P_{2j}=
P_{2j_0}\,Q_{2j-2j_0}$ and $\sum_{j=j_0}^\infty \gamma_{2j}\,
Q_{2j-2j_0}(\xi) \neq 0, \quad \xi \in S^{n-1}$.*]{}
Here for a positive integer $j$ we have set $$\label{eq7}
\gamma_j = i^{-j}\, \pi^{\frac{n}{2}}
\frac{\Gamma(\frac{j}{2})}{\Gamma ( \frac{n+j}{2})}\,.$$ The quantities $\gamma_j$ appear in the computation of the Fourier multiplier of the higher order Riesz transform $R$ with kernel given by a homogeneous harmonic polynomial $P$ of degree $j$. One has (see [@St p. 73]) $$\widehat{Rf}(\xi)= \gamma_j\frac{P(\xi)}{|\xi|^j} \,
\hat{f}(\xi),\quad f \in L^2({{\mathbb R}^n})\,.$$ As we will show later, the series $\sum_{j=j_o}^\infty
\gamma_{2j}\,Q_{2j-2j_0}(x)$ is convergent in $C^\infty(S^{n-1})$. If $U$ is defined as the operator in the algebra $A$ whose Fourier multiplier is $\gamma_{2j_0}^{-1}$ times the sum of the preceding series, then $T=R \circ U$ for the higher order Riesz transform $R$ given by the polynomial $P_{2j_0}$. This shows that $(iii)$ follows from $(iv)$.
To show that $(iii)$ implies $(iv)$ we prove that $P=
\lambda\,P_{2j_0}$ for some real number $\lambda \neq 0$. Since $P$ divides $P_{2j_0}$ by assumption, we only need to show that the degree $d$ of $P$ must be $2 j_0$. Now, let $\mu(\xi)$ denote the Fourier multiplier of $U$, so that $\mu$ is a smooth function with no zeros on the sphere. The Fourier multiplier of $T$ is $$\sum_{j=j_0}^\infty \gamma_{2j}\, P_{2j}(\xi) =
\gamma_{d}\,P(\xi)\,\mu(\xi),\quad \xi \in S^{n-1}\,.$$ If $d$ is less than $2j_0$, then $P$ is orthogonal to all $P_{2j}$ [@St p. 69] and so $$\int_{|\xi|=1} P(\xi)^2 \,\mu(\xi)\,d\xi = 0\,,$$ which yields $P(\xi)=0$, $|\xi|=1$, a contradiction.
Condition $(iii)$ is rather easy to check in practice. For instance, take $n=2$ and consider the polynomial of fourth degree $$P(x,y)= xy+ x^4+y^4-6\, x^2y^2\,.$$ The polynomial operator associated to $P$ does not satisfy $(i)$ nor $(ii)$, because the definition of $P$ above is also the spherical harmonics expansion of $P$ and $xy$ clearly does not divide $
x^4+y^4-6\,x^2 y^2$. Section 7 contains other examples of polynomial operators that do not satisfy $(i)$ nor $(ii)$.
On the other hand, the polynomial operator associated with $$P(x,y)= xy+ x^3 y-y^3 x\,,$$ does satisfy $(i)$ and $(ii)$, but this is not the case for the operator determined by $$P(x,y)= xy+ 2 (x^3 y-y^3 x)\,,$$ although $xy$ obviously divides $x^3 y-y^3 x$. See section 8 for the details.
Thus the condition on $\Omega$ so that $T$ satisfies $(i)$ or $(ii)$ is rather subtle.
Having clarified the statement of the Theorem and some of its implications, we say now a few words on the proofs and the organization of the paper.
We devote sections 2, 3 and 4 to the proof of “$(iii)$ implies $(i)$", which we call the sufficient condition. In section 2 we prove that the even higher order Riesz transforms satisfy $(i)$. Section 3 is devoted to the proof of the sufficient condition for polynomial operators. The argument is an extension of that used in the previous section. The drawback is that we loose control on the dependence of the constants on the degree of the polynomial. The main difficulty we have to overcome in section 4 to complete the proof of the sufficient condition in the general case, is to find a second approach to the polynomial case which gives some estimates with constants independent of the degree of the polynomial. This allows us to use a compactness argument to finish the proof. It is an intriguing fact that the approach in section 3 cannot be dispensed with, because it provides certain properties which are vital for the final argument and do not follow otherwise.
In sections 5 and 6 we prove the necessary condition, that is, “$(ii)$ implies $(iii)$”. In section 5 we deal with the polynomial case. Analysing the inequality $(ii)$ via Plancherel at the frequency side we obtain various inclusion relations among zero sets of certain polynomials. This requires a considerable combinatorial effort for reasons that will become clear later on. We found in Maple a formidable ally in formulating the right identities which were needed, which we proved rigorously afterwards. In a second step we solve the division problem which leads us to $(iii)$ by a recurrent argument with some algebraic geometry ingredients, the Hilbert’s Nullstellensatz in particular. The question of independence on the degree of the polynomial appears again, this time related to the coefficients of certain expansions. We deal with this problem in section 6. Section 7 is devoted to the proof of the intricate combinatorial lemmas used in the previous sections. In section 8 we discuss some examples and we ask a couple of questions that we have not been able to answer.
Our methods are a combination of classical Fourier analysis techniques and Calderón-Zygmund theory with potential theoretic ideas coming from our previous work [@MO], [@MPV], [@MV1], [@MNOV], [@Ve1] and [@Ve2].
As it was discovered in [@MV2], there is remarkable difference between the odd and even cases for the problem we consider. To keep at a reasonable size the length of this article we decided to only deal here with the even case, which is more difficult, because one needs special $L^\infty$ estimates for singular integrals, which hold only in the even case. The results for the odd case will be published elsewhere [@MOPV].
The $L^\infty$ estimates mentioned above are not obvious even for the simplest even homogeneous Calderón-Zygmund operator, the Beurling transform, which plays an important role in planar quasiconformal mapping theory. An application of our estimates to planar quasiconformal mappings is given in [@MOV].
Even higher order Riesz transforms
==================================
In this section we prove that if $T$ is an even higher order Riesz transform, then $$\label{eq8}
T^{\star}f(x) \leq C \, M(Tf)(x),\quad x \in {{\mathbb R}^n}\,.$$
Let $B$ be the open ball of center $0$ and radius $1$, $\partial B$ its boundary and $\overline B$ its closure. In proving we will encounter the following situation. We are given a function $\varphi$ defined by different formulae in $B$ and ${{\mathbb R}^n}\setminus \overline B$, which is differentiable up to order $N$ on $B \cup ({{{\mathbb R}^n}\setminus \overline B})$ and whose derivatives up to order $N-1$ extend continuously up to $\partial B$. The question is to compare the distributional derivatives of order $N$ with the expressions one gets on $B$ and ${{{\mathbb R}^n}\setminus \overline B}$ by taking ordinary derivatives. The next simple lemma is a sample of what we need.
\[L1\] Let $\varphi$ be a continuously differentiable function on $B \cup
({{{\mathbb R}^n}\setminus \overline B}) $ which extends continuously to $\partial B$. Then we have the identity $$\partial_j \varphi = \partial_j \varphi(x) \chi_B(x) + \partial_j
\varphi(x)\chi_{{{{\mathbb R}^n}\setminus \overline B}}(x)\,,$$ where the left hand side is the $j$-th distributional derivative of $\varphi$.
Let $\psi$ be a test function. Then $$\langle \partial_j \varphi, \psi \rangle = - \int \varphi\,
\partial_j \psi = - \int_B \varphi \,
\partial_j \psi - \int_{{{{\mathbb R}^n}\setminus \overline B}} \varphi \,
\partial_j \psi \,.$$ Now apply Green-Stokes’ theorem to the domains $B$ and ${{{\mathbb R}^n}\setminus \overline B}$ to move the derivatives from $\psi$ to $\varphi$. The boundary terms cancel precisely because of the continuity of $\varphi$ on $\partial B$, and we get $$\langle \partial_j \varphi, \psi \rangle = \int (\chi_B \,
\partial_j \varphi + \chi_{{{{\mathbb R}^n}\setminus \overline B}} \,
\partial_j \varphi ) \psi \,dx\,.$$
We need an analog of the previous statement for second order derivatives and radial functions, which is the case we take up in the next corollary.
\[C1\] Assume that $\varphi$ is a radial function of the form $$\varphi(x)= \varphi_1(|x|) \chi_B(x) + \varphi_2(|x|)
\chi_{{{{\mathbb R}^n}\setminus \overline B}}(x)\,,$$ where $\varphi_1$ is continuously differentiable on $[0,1)$ and $\varphi_2$ on $(1,\infty)$. Let $L$ be a second order differential operator with constant coefficients. Then the distribution $L
\varphi$ satisfies $$L\varphi= L\varphi(x) \chi_B(x) + L\varphi(x) \chi_{{{{\mathbb R}^n}\setminus \overline B}}(x)\,,$$ provided $\varphi_1$, $\varphi_1'$, $\varphi_2$ and $\varphi_2'$ extend continuously to the point $1$ and the two conditions $$\varphi_1(1)= \varphi_2(1), \quad \varphi_1'(1) = \varphi_2'(1)\,,$$ are satisfied.
The proof reduces to applying Lemma \[L1\] twice. Before the second application one should remark that the hypothesis $
\varphi_1'(1) = \varphi_2'(1)\,$ gives the continuity of all first order partial derivatives of $\varphi$.
We proceed now to describe in detail the main argument for the proof of . By translating and dilating one reduces the proof of to $$\label{eq9}
|T^1 f(0)| \leq C \, M(Tf)(0)\,,$$ where $$T^1 f(0)= \int_{| y| > 1} f(y) K(y) \,dy$$ is the truncated integral at level $1$. Recall that the kernel of our singular integral is $$K(x)= \frac{\Omega(x)}{|x|^{n}}= \frac{P(x)}{|x|^{n+d}}\,,$$ where $P$ is an even homogeneous harmonic polynomial of degree $d\geq 2$. The idea is to obtain an identity of the form $$\label{eq10}
K(x)\chi_{{{{\mathbb R}^n}\setminus \overline B}}(x)= T(b)(x)\,,$$ for some measurable bounded function $b$ supported on $B$. Once is at our disposition we get, for $f$ in some $
L^p({{\mathbb R}^n})$, $1\leq p <\infty$, $$\begin{split}
T^1 f(0) & = \int \chi_{{{{\mathbb R}^n}\setminus \overline B}}(y)\,K(y)\, f(y) \,dy \\
& = \int T(b)(y)\,f(y)\,dy \\
&= \int_B b(y) \, Tf(y)\,dy\,,
\end{split}$$ and so follows with $C= V_n \, \|b \|_\infty$, $V_n$ being the volume of the unit ball of ${{\mathbb R}^n}$.
Let us turn our attention to the proof of . Set $d=2N$ and let $E$ be the standard fundamental solution of the $N$-th power $\triangle^N$ of the Laplacean. Consider the function $$\label{eq11}
\varphi(x)= E(x)\,\chi_{{{{\mathbb R}^n}\setminus \overline B}}(x) + (A_0+A_1\,|x|^2 +\dotsb+
A_{d-1}\,|x|^{2d-2})\,\chi_B(x)\,,$$ where the constants $A_0, A_1,\dotsc, A_{d-1}$ are chosen as follows. Since $\varphi(x)$ is radial, the same is true of $\triangle^j \varphi$ for each positive integer $j$. Thus, in order to apply the preceding corollary $N$ times one needs $2N=d$ conditions, which (uniquely) determine $A_0, A_1,\dotsc,
A_{d-1}$. Therefore, for some constants $\alpha_1, \alpha_2,\dotsc,
\alpha_{N-1}$, $$\label{eq11bis}
\triangle^N \,\varphi = (\alpha_0+ \alpha_1 |x|^2 +\dotsb+
\alpha_{N-1} |x|^{2(N-1)})\chi_B(x) = b(x)\,,$$ where the last identity is a definition of $b$. Since $$\varphi = E \star \triangle^N\,\varphi \,,$$ taking derivatives of both sides we obtain $$\label{eq12}
P(\partial)\,\varphi = P(\partial)\, E \star \triangle^N \,
\varphi\,.$$ To compute $P(\partial)E$ we take the Fourier transform $$\widehat{P(\partial)E}(\xi)= P(i\xi)\, \hat{E}(\xi)=
\frac{P(\xi)}{|\xi|^d}\,.$$ On the other hand, as it is well known ([@St p. 73], $$\widehat{P.V.\frac{P(x)}{|x|^{n+d}}}\,(\xi)= \gamma_d
\frac{P(\xi)}{|\xi|^d}\,.$$ See for the precise value of $\gamma_d$, which is not important now. We conclude that, for some constant $c_d$ depending on $d$, $$P(\partial)E = c_d \,P.V.\frac{P(x)}{|x|^{n+d}}\,.$$ Thus $$P(\partial)\varphi =
c_d\,P.V.\frac{P(x)}{|x|^{n+d}}\star\,\triangle^N\,\varphi =
c_d\,T(b)\,.$$
The only thing left is the computation of $P(\partial)\,\varphi$. We have, by Corollary \[C1\], $$P(\partial)\,\varphi = c_d\, K(x)\,\chi_{{{{\mathbb R}^n}\setminus \overline B}} +
P(\partial)(A_0+A_1\,|x|^2+\dotsb+
A_{d-1}\,|x|^{2d-2})(x)\,\chi_B(x)\,,$$ and so, to complete the proof of , we only have to show that $$\label{eq13}
P(\partial)(|x|^{2j})= 0,\quad 1 \leq j \leq d-1\,.$$ Notice that the degree of $P$ may be much smaller than the degree of $|x|^{2j}$ and so the previous identity is not obvious. Taking the Fourier transform we obtain $$\widehat{P(\partial)(|x|^{2j})}= c_j \,P(\xi)\,\triangle^j
\,\delta\,,$$ where $\delta$ is the Dirac delta at the origin and $c_j$ a constant depending on $j$. Let $\psi$ be a test function. Then, since $P$ is harmonic, $$\begin{split}
\langle P(\xi)\,\triangle^j \,\delta, \varphi \rangle &= \langle
\triangle^j \,\delta ,\, P(\xi)\, \varphi(\xi)\rangle \\
& =\langle \triangle^{j-1} \,\delta, \,2\, \nabla P(\xi)\cdot \nabla
\varphi(\xi)+ P(\xi)\,\triangle\varphi(\xi) \rangle\,.
\end{split}$$ Iterating the previous computation we obtain that $$\langle P(\xi)\,\triangle^j \,\delta,\, \varphi \rangle = \langle
\delta, D(\xi)\rangle= D(0)\,,$$ where $D$ is a linear combination of products of the form $\partial^\alpha \,\varphi(\xi)\,\partial^\beta\,P(\xi)$, with multi-indeces $\beta$ of length $|\beta| \leq j \leq d-1$. Therefore $\partial^\beta\,P(\xi)$ is a homogeneous polynomial of degree at least $d-j \geq 1$, and so $\partial^\beta\,P(0)=0$. This yields $D(0)=0$ and completes the proof of and, thus, of .
In fact follows immediately from an identity of Lyons and Zumbrun [@LZ] which will be discussed in the next section, but we prefer to present here the above independent natural argument for the reader’s convenience.
Proof of the sufficient condition: the polynomial case
======================================================
In this section we assume that $T$ is an even polynomial operator. This amounts to say that for some even integer $2N$, $N \geq 1$, the function $ |x|^{2N}\,\Omega(x)$ is a homogeneous polynomial of degree $2N$. Such a polynomial may be written as [@St p. 69] $$|x|^{2N}\,\Omega(x)= P_2(x)|x|^{2N-2}+ \dotsb+
P_{2j}(x)|x|^{2N-2j}+\dotsb+ P_{2N}(x)\,,$$ where $P_{2j}$ is a homogeneous harmonic polynomial of degree $2j$, $1\leq j \leq N $. In other words, the expansion of $\Omega(x)$ in spherical harmonics is $$\Omega(x)= P_2(x)+P_4(x)+\dotsb+P_{2N}(x),\quad |x|=1\,.$$
As in the previous section, we want to obtain an expression for the kernel $K(x)$ off the unit ball $B$. For this we need the differential operator $Q(\partial)$ defined by the polynomial $$Q(x)= \gamma_2 \, P_2(x)|x|^{2N-2}+\dotsb+ \gamma_{2j}\,
P_{2j}(x)|x|^{2N-2j}+\dotsb+ \gamma_{2N}\,P_{2N}(x) \,.$$ If $E$ is the standard fundamental solution of $\Delta^N$, then $$Q(\partial)E = P.V. \,K(x)\,,$$ which may be easily verified by taking the Fourier $E$ transform of both sides.
Take now the function $\varphi$ of the previous section. We have $\varphi = E \star \triangle^N\,\varphi$ and thus $$Q(\partial)\varphi = Q(\partial)E \star \triangle^N\,\varphi =
P.V.\, K(x) \star b = T(b)\,,$$ where $b$ is defined as $ \triangle^N\,\varphi$. On the other hand, by Corollary 2 $$\label{eq14}
Q(\partial)\,\varphi = K(x)\,\chi_{{{{\mathbb R}^n}\setminus \overline B}} +
Q(\partial)(A_0+A_1\,|x|^2+\dotsb+
A_{2N-1}\,|x|^{4N-2})(x)\,\chi_B(x)\,.$$ Contrary to what happened in the previous section, the term $$S(x) : = - Q(\partial)(A_0+A_1\,|x|^2+\dotsb+
A_{2N-1}\,|x|^{4N-2})(x)$$ does not necessarily vanish, the reason being that now $Q$ does not need to be harmonic.
Our goal is to find a function $\beta \in L^\infty({{\mathbb R}^n})$, satisfying the decay estimate $$\label{eq15}
|\beta(x)| \leq \frac{C}{|x|^{n+1}},\quad |x|\geq 2\,,$$ and $$\label{eq16}
S(x) \chi_B(x) = T(\beta)(x)\,.$$ Once this is achieved the proof of $(i)$ is just a variation of the argument presented in section 2, which we now explain. By , the definition of $S(x)$ and , we get $$\label{eq17}
K(x)\chi_{{{{\mathbb R}^n}\setminus \overline B}}(x)= T(b)(x)+ T(\beta)(x)\,.$$ Set $ \gamma = b+ \beta$. We show by arguing as follows. For $f$ in any $ L^p({{\mathbb R}^n})$, $1\leq p <\infty$, we have $$\begin{split}
T^1 f(0) & = \int \chi_{{{{\mathbb R}^n}\setminus \overline B}}(y)\,K(y)\, f(y) \,dy \\
& = \int T(\gamma)(y)\,f(y)\,dy \\
& = \int \gamma(y)\, Tf(y)\,dy \\
&= \int_{2B} \gamma(y) \, Tf(y)\,dy + \int_{{{\mathbb R}^n}\setminus 2B} \gamma(y) \,
Tf(y)\,dy
\end{split}$$ and thus, by the decay inequality with $\beta$ replaced by $\gamma$, $$\begin{split}
|T^1 f(0)| & \leq C \left( \|\gamma\|_\infty
\frac{1}{|2B|}\int_{2B}|Tf(y)|\,dy
+ \int_{{{\mathbb R}^n}\setminus 2B} \frac{|Tf(y)|}{|y|^{n+1}}\,dy \right)\\*[5pt]
& \leq C\,M(Tf)(0)\,.
\end{split}$$
To construct $\beta$ satisfying and we resort to our hypothesis, condition $(iii)$ in the Theorem, which says that $T = R \circ U$, where $R$ is a higher order Riesz transform, $U$ is an invertible operator in the algebra $A$ and the polynomial $P$ which determines $R$ divides $P_{2j}$, $1\leq j \leq
N $, in the ring of polynomials in $n$ variables with real coefficients. The construction of $\beta$ is performed in two steps.
The first step consists in proving that there exists a function $\beta_1$ in $L^\infty(B)$, satisfying a Lipschitz condition of order $1$ on $B$, $\int \beta_1(x)\,dx =0$ and such that $$\label{eq18}
S(x) \chi_B(x) = R(\beta_1)(x)\,.$$ It will become clear later on how the Lipschitz condition on $\beta_1$ is used. To prove we need an explicit formula for $S(x)$ and for that we will make use of the following formula of Lyons and Zumbrun [@LZ].
\[L2\] Let $L$ be a homogeneous polynomial of degree $l$ and let $f$ be a smooth function of one variable. Then $$L(\partial)f(r)= \sum_{\nu \geq 0} \frac{1}{2^\nu\,\nu !}\,
\Delta^\nu L(x)\left(\frac{1}{r}\frac{\partial}{\partial
r}\right)^{l-\nu} f(r), \quad r=|x|\,.$$
An immediate consequence of Lemma 3 is
\[L3\] Let $P_{2j}$ a homogeneous harmonic polynomial of degree $2j$ and let $k$ be a non-negative integer. Then $$P_{2j}(\partial)(|x|^{2k}) = 2^{2j}\frac{k!}{(k-2j)!}
\,P_{2j}(x)\,|x|^{2(k-2j)}\quad \text{if }\, 2j\leq k\,,$$ and $$P_{2j}(\partial)(|x|^{2k}) = 0,\quad \text{if }\, 2j > k\,.$$
On the other hand, a routine computation gives $$\label{eq19}
\triangle^{j}(|x|^{2k}) = 4^j\frac{j! \,k!
}{(k-j)!}\binom{\frac{n}{2}+k-1}{j} \,|x|^{2(k-j)},\quad k \geq j\,,$$ and $$\label{eq20}
\triangle^{j}(|x|^{2k}) = 0 ,\quad k < j \,.$$ By Lemma 4, and we get that for some constants $c_{jk}$ one has, in view of the definitions of $Q(x)$ and $S(x)$, $$\label{eq21}
S(x)= \sum_{j=1}^{N-1} \sum_{k=j}^{N-1}
c_{jk}\,P_{2j}(x)\,|x|^{2(k-j)}\,.$$ Therefore it suffices to prove with $S(x)$ replaced by $P_{2j}(x)\,|x|^{2k}$, for $1 \leq j \leq N$ and each non-negative integer $k$. The idea is to look for an appropriate function $\psi$ such that $$\label{eq22}
P(\partial)\psi (x) = P_{2j}(x)\,|x|^{2k}\,\chi_B(x)\,.$$ Indeed, if holds and $2d$ is the degree of $P$, then $$\psi = E \star \Delta^d \psi\,,$$ provided $E$ is the fundamental solution of $\triangle^d$ and $\psi$ is good enough. Hence $$P(\partial)\psi = P(\partial)E \star \Delta^d \psi = c\,
P.V.\,\frac{P(x)}{|x|^{n+2d}} \star \Delta^d \psi= R(\beta_1)\,,$$ if $\beta_1 = c\, \Delta^d \psi$. The conclusion is that we have to solve in such a way that $\Delta^d \psi$ is supported on $B$, is a Lipschitz function on $B$ and has zero integral.
Taking Fourier transforms in we get $$\label{eq23}
(-1)^d P(\xi)\, \widehat{\psi} (\xi) =
(-1)^{j+k}\,P_{2j}(\partial)\,\triangle^k
\left(\widehat{\chi_B}(\xi)\right)\,.$$ Recall that for $m=n/ 2$ one has [@Gr A-10] $$\widehat{\chi_B}(\xi)= \frac{J_m(\xi)}{|\xi|^m},\quad \xi \in {{\mathbb R}^n}\,,$$ where $J_m$ is the Bessel function of order $m$. Set $$G_\lambda(\xi)= \frac{J_\lambda(\xi)}{|\xi|^\lambda},\quad \xi \in
{{\mathbb R}^n}, \quad \lambda>0\,.$$ In computing the right hand side of we apply Lemma 3 to $L(x)= P_{2j}(x) \,|x|^{2k}$ and $f(r)= G_m(r)$ and we get $$P(\xi)\, \widehat{\psi} (\xi) = (-1)^{j+k+d} \sum_{\nu \geq 0}
\frac{(-1)^\nu}{2^\nu\,\nu !}\,\triangle^\nu
\left(P_{2j}(\xi)\,|\xi|^{2k}\right)\,G_{m+2j+2k-\nu}(\xi)\,,$$ owing to the well known formula, e.g. [@Gr A-6], $$\frac{1}{r}\,\frac{d}{dr}\,G_{\lambda}(r) = -
G_{\lambda+1}(r)\,,\quad r>0,\quad \lambda>0\,.$$ Since $P_{2j}(\xi)$ is homogeneous of degree $2j$, $\nabla
P_{2j}(\xi)\cdot \xi = 2j\,P_{2j}(\xi)$, and hence one may readily show by an inductive argument that $$\triangle^\nu \left(P_{2j}(\xi)\,|\xi|^{2k}\right)=
a_{jk\nu}\,P_{2j}(\xi)\,|\xi|^{2(k-\nu)}\,,$$ for some constants $a_{jk\nu}$. Thus, for some other constants $
a_{jk\nu}$, we get $$\label{eq24}
P(\xi)\, \widehat{\psi} (\xi) = \sum_{\nu \geq 0}
a_{jk\nu}\,P_{2j}(\xi)\,|\xi|^{2(k-\nu)}\,G_{m+2j+2k-\nu}(\xi)\,.$$ By hypothesis $P$ divides $P_{2j}$ in the ring of polynomials in $n$ variables and so $$P_{2j}(\xi)= P(\xi)\,Q_{2j-2d}(\xi)\,,$$ for some homogeneous polynomial $Q_{2j-2d}$ of degree $2j-2d$. Cancelling out the factor $P(\xi)$ in we conclude that $$\widehat{\psi} (\xi) = Q_{2j-2d}(\xi)\sum_{\nu = 0}^k
a_{jk\nu}\,|\xi|^{2(k-\nu)}\,G_{m+2j+2k-\nu}(\xi)\,.$$ Since [@Gr A-10] $$\widehat{\left((1-|x|^2)^\lambda \,\chi_B(x)\right)}(\xi) =
c_\lambda\,
G_{m+\lambda}(\xi)\,,$$ we finally obtain $$\psi(x) = Q_{2j-2d}(\partial)\sum_{\nu = 0}^k
a_{jk\nu}\,\triangle^{k-\nu}\left( (1-|x|^2)
^{2j+2k-\nu}\,\chi_B(x)\right)\,.$$ Observe that $\psi$ restricted to $B$ is a polynomial which vanishes on $\partial B$ up to order $2d$ and $\psi$ is zero off $B$. Therefore $
\triangle^d \psi$ is supported on $B$ and its restriction to $B$ is a polynomial with zero integral. This completes the first step of the construction of $\beta$.
The second step proceeds as follows. Since by hypothesis $T=R\circ U
$, with $U$ invertible in the algebra $A$, we have $$R(\beta_1)= T(U^{-1}\beta_1)\,.$$ Setting $$\label{eq24bis}
\beta=U^{-1}\beta_1\,,$$ we are only left with the task of showing that $$\label{eq25}
\beta \in L^\infty({{\mathbb R}^n})$$ and that, for some positive constant $C$, $$\label{eq26}
| \beta (x)| \leq \frac{C}{|x|^{n+1}},\quad |x| \geq 2\,.$$ Since $U^{-1} \in A$, for some real number $\lambda$ and some smooth homogeneous Calderón-Zygmund operator $V$, $$U^{-1} = \lambda\,I + V \,.$$ Thus $$\beta = \lambda\,\beta_1 + V (\beta_1)\,.$$ Now $\beta_1$ is supported on $B$ and has zero integral on $B$ and this is enough to insure the decay estimate . Indeed, let $L(x)$ be the kernel of $V$ and assume that $|x| \geq 2$. Then $$\label{eq27}
\begin{split}
V(\beta_1)(x) &= \int L(x-y) \,\beta_1(y)\,dy \\
&= \int \left(L(x-y) - L(x)
\right)\,\beta_1(y)\,dy\,,
\end{split}$$ and so $$\label{eq28}
\begin{split}
|V(\beta_1)(x)| & \leq \int |\left(L(x-y) - L(x)
\right)|\,|\beta_1(y)|\,dy,\\
& \leq C\, \int \frac{|y|}{|x|^{n+1}} \,|\beta_1(y)|\,dy,\\
& = \frac{C}{|x|^{n+1}}\,.
\end{split}$$
The boundedness of $\beta$ is a more delicate issue. It follows immediately from the next lemma applied to the operator $V$ and the function $\beta_1$. Is precisely here where we use the fact that $\beta_1$ satisfies a Lipschitz condition.
The constant of the kernel $K(x)= \Omega(x)/|x|^n$ of the smooth homogeneous Calderón-Zygmund operator $T$ is $$\label{eq29}
\|T\|_{CZ} \equiv \|K\|_{CZ}= \|\Omega \|_\infty+ \| |x|\, \nabla
\Omega(x) \|_\infty\,.$$
We adopt the standard notation for the minimal Lipschitz constant of a Lipschitz function $f$ on $B$, namely $$\|f\|_{\operatorname{Lip}(1,B)} = \sup\left \{
\frac{|f(x)-f(y)|}{|x-y|}: x,y \in B, \, x \neq y\right \} <
\infty\,.$$
\[Lip\] Let $T$ be the homogeneous singular integral operator with kernel $K(x)= \frac{\Omega(x)}{|x|^n}$, where $\Omega$ is an even homogeneous function of degree $0$, continuously differentiable and with zero integral on the unit sphere. Then $$\|T(f\,\chi_B)\|_{L^\infty({{\mathbb R}^n})} \leq C \, \|K\|_{CZ} \left(
\|f\|_{L^\infty(B)} + \|f\|_{\operatorname{Lip}(1,B)}\right)\,,$$ where $C$ is a positive constant which depends only on $n$.
We start by examining the behaviour of $T(f\,\chi_B)$ on the unit sphere. We claim that $$|T_\epsilon(f\,\chi_B)(a)| \leq C\, \|K\|_{CZ} \left(
\|f\|_{L^\infty(B)} + \|f\|_{\operatorname{Lip}(1,B)}\right),\quad
|a|=1,\quad \epsilon>0\,.$$ Indeed, if one follows in detail the proof of the claim, which we discuss below, one will realize that the principal value integral $T(f\,\chi_B)(a)$ exists for all $a$ in the sphere and satisfies the desired estimate.
We have $$\begin{split}
T_\epsilon(f\,\chi_B)(a)&= \int_{\epsilon < |x-a|< 1/2}
\chi_B(x)\,f(x)\, K(a-x)\,dx + \int_{1/2 < |x-a|} \dotsi\\*[5pt] &=
I_\epsilon + II \,.
\end{split}$$ Clearly, $$|II| \leq \int_{1/2 < |x-a|} \chi_B(x)\,|f(x)|\,
\frac{|\Omega(x-a)|}{|x-a|^n}\,dx \leq 2^n\,|B|\,\|\Omega\|_\infty\,
\|f\|_{L^\infty(B)} \,.$$ To deal with the term $I_\epsilon$ we write $$\begin{split}
I_\epsilon &= \int_{\epsilon < |x-a|< 1/2} \chi_B(x)\,(f(x)-f(a))\,
K(a-x)\,dx\\*[5pt] &\quad+ f(a)\int_{\epsilon < |x-a|< 1/2}
\chi_B(x)\, K(a-x)\,dx \\*[5pt] & = III_\epsilon+ f(a)\,IV_\epsilon
\,,
\end{split}$$ and we remark that $III_\epsilon$ can easily be estimated as follows $$|III_\epsilon| \leq \|f\|_{\operatorname{Lip}(1,B)}\,\int_B |x-a|
|K(a-x)|\,dx \leq C\,\|\Omega\|_\infty\,\|f
\|_{\operatorname{Lip}(1,B)}\,.$$ Taking care of $IV_\epsilon$ is not so easy. Take spherical coordinates centered at the point $a$, $x=a+r\, \omega$ with $0 \leq
r$ and $|\omega|=1$. Then $$\label{eq30}
IV_\epsilon = \int_\epsilon^{1/2} \left( \int_{A(r)}
\Omega(\omega)\,d\sigma(\omega)\right)\frac{dr}{r} \,,$$ where $$A(r)= \{ \omega : |\omega| =1 \text{ and }|a+r \omega| < 1 \}\,.$$ Let $H$ be the tangent hiperplane to $S= \{\omega : |\omega|=1\}$ at the point $a$. Call $V$ the half space with boundary $H$ containing the origin. Clearly $A(r) \subset S \cap V$. Since $\Omega$ is even, $$0= \int_S \Omega(\omega)\,d\sigma(\omega) = 2\, \int_{S \cap V}
\Omega(\omega)\,d\sigma(\omega)\,.$$ Thus $$\int_{A(r)} \Omega(\omega)\,d\sigma(\omega) = - \int_{(S \cap V)
\setminus A(r)} \Omega(\omega)\,d\sigma(\omega)\,,$$ and so $$\left|\int_{A(r)} \Omega(\omega)\,d\sigma(\omega)\right| \leq \|
\Omega\|_\infty \, \sigma((S \cap V) \setminus A(r))\,.$$ Since $H$ is tangent to $S$ at the point $a$, we obtain $$\sigma((S \cap V) \setminus A(r)) \leq C\,r\,,$$ which yields, by ,
$$|IV_\epsilon| \leq C\,\|\Omega \|_\infty\,.$$
\[lhoritzo\]
Assume now that $|a| < 1$. Proceeding as before we estimate in the same way the terms $II$ and $III_\epsilon$, so that we are again left with $IV_\epsilon$. Let $\epsilon_0$ stand for the distance from $a$ to the boundary of $B$. In estimating $IV_\epsilon$ we can assume, without loss of generality, that $\epsilon_0 \leq 1/4$. .Set $a_0 = a/|a|$, $$A= \{x \in B : \epsilon_0 < |x-a| < 1/2 \}$$ and $$A_0= \{x \in B : \epsilon_0 < |x-a_0| < 1/2 \}\,.$$ We compare $IV_\epsilon$ to the expression we get replacing $a$ by $a_0$ and $\epsilon$ by $\epsilon_0$ in the definition of $IV_\epsilon$. For $\epsilon \leq \epsilon_0$ we have $$\int_{\epsilon < |x-a|< 1/2} \chi_B(x)\, K(a-x)\,dx =
\int_{\epsilon_0 < |x-a|< 1/2} \chi_B(x)\, K(a-x)\,dx$$ and then $$\begin{split}
\biggl|\int_{\epsilon < |x-a|< 1/2} &\chi_B(x)\, K(a-x)\,dx -
\int_{\epsilon_0 < |x-a_0|< 1/2} \chi_B(x)\, K(a_0-x)\,dx \biggr|
\\*[7pt] &= \left|\int_{A} K(a-x)\,dx - \int_{A_0} K(a_0-x)\,dx
\right| \\*[7pt] &\leq \int_{A\cap A_0}
|K(a-x)-K(a_{0}-x)|\,dx\\*[7pt] &\quad+ \left|\int_{A \setminus A_0}
\!\chi_B(x)\, K(a-x)\,dx \right| +
\left| \int_{A_0 \setminus A} \chi_B(x)\, K(a_0-x)\,dx \right| \\*[7pt]
&= J_1+J_2+J_3\,.
\end{split}$$ If $x \in A \cap A_0$, then $$|K(a-x)-K(a_0-x)| \leq C \,\|K\|_{CZ}\,
\frac{|a-a_0|}{|x-a|^{n+1}}\,.$$ Hence $$J_1 \leq C \,\|K\|_{CZ}\, |a-a_0|\, \int_{|x-a|> \epsilon_0}
\frac{dx}{|x-a|^{n+1}} \leq C\,\|K\|_{CZ}\,.$$ To estimate $J_2$ observe that $$A \setminus A_0 = \left(A \cap B(a_0,\epsilon_0)\right) \cup \left(A
\cap ({{\mathbb R}^n}\setminus B(a_0,1/2))\right)\,.$$ Now, it is obvious that if $|x-a_0|\geq 1/2$, then $|x-a| \geq 1/4$, and so $$J_2 \leq \|\Omega\|_\infty \left( \int_{|x-a_0|< \epsilon_0}
\frac{dx}{\epsilon_0^n} + \int_B 4^n \,dx \right) \leq
C\,\|\Omega\|_\infty \,.$$ A similar argument does the job for $J_3$.
The case $|a|>1$ is treated in a completely analogous way.
The construction of $\beta$ is then completed and the Theorem is proved for polynomial operators.
We remark that a variant of Lemma \[Lip\] holds, with the same proof, replacing $B$ by ${{{\mathbb R}^n}\setminus \overline B}$. To control the term $II$ we have to assume, in addition to the hypothesis of Lemma \[Lip\], that $f$ satisfies a decay inequality of the type $$|f(x)| \le \frac{\|f \|_{L^\infty({{{\mathbb R}^n}\setminus \overline B})}}{|x|^\eta}, \quad |x| \geq
1\,.$$ Then we conclude that $$\|T(f\,\chi_{{{{\mathbb R}^n}\setminus \overline B}})\|_{L^\infty({{\mathbb R}^n})} \leq C \, \|K\|_{CZ} \left(
\|f\|_{L^\infty({{{\mathbb R}^n}\setminus \overline B})} + \|f\|_{\operatorname{Lip}(1,{{{\mathbb R}^n}\setminus \overline B})}\right)\,,$$ where $C$ depends on $n$ and $\eta$. We will use later on this variant of Lemma \[Lip\] with $f(x)= K(x)$ on ${{{\mathbb R}^n}\setminus \overline B}$, so that $\eta=n$ and the constant $C$ will depend only on $n$.
We mention another straightforward extension of Lemma \[Lip\] that will not be used in this paper. The function $f$ may be assumed to be in $\operatorname{Lip}(\alpha,B)$ $0<\alpha \leq 1$, and the unit ball may be replaced by a domain with boundary of class $C^{1+\epsilon}$.
After the paper was completed we learned from Stephen Semmes that Lemma \[Lip\] is known in dimension $2$ [@Ch p.52] or [@BM p.348] and that was used to prove global regularity of vortex patches for incompressible perfect fluids.
Proof of the sufficient condition: the general case
===================================================
We start this section by clarifying several facts about the convergence of the series . Let us then assume that $\Omega$ is a function in $C^\infty(S^{n-1})$ with zero integral. Then $\Omega$ has an expansion in spherical harmonics. For each positive integer $r$, one has the identity [@St p. 70] $$\label{eq30bis}
\sum_{j\geq 1} \left(j(j+n-2)\right)^r\,\|P_j\|^2_2 =
(-1)^r\int_{S^{n-1}} \triangle^r_S \Omega \,\Omega \,d\sigma \,,$$ where $\Delta_S$ stands for the spherical Laplacean. Then $$\sum_{j\geq 1} \left(j(j+n-2)\right)^r\,\|P_j\|^2_{2} \leq \|
\triangle^r_S \Omega \|_2 \,\|\Omega \|_2 \,,$$ where the $L^2$ norm is taken with respect to $d\sigma$. Thus, by Schwarz’s inequality, for each positive integer $M$ $$\label{eq31}
\sum_{j\geq 1} j^M \,\|P_j\|_{2} < \infty\,.$$ We want to see that we also have $$\label{eq32}
\sum_{j\geq 1} j^M \,\|P_j\|_{\infty} < \infty\,,$$ where the supremum norm is taken on $S^{n-1}$. This follows immediately from the next lemma, whose proof was indicated to us by Fulvio Ricci.
\[Hom\] For all homogeneous polynomials of degree $q$ $$\| Q\|_\infty \leq C\,q^{\frac{n-1}{2}}\,\|Q \|_2\,,$$ where $C$ is a positive constant which depends only on $n$.
Take an orthonormal base $Q_1,\dotsc,Q_d$, $d=d_q$, of the subspace of $L^2(d\sigma)$ consisting of the restrictions to $S^{n-1}$ of all homogeneous polynomials of degree $q$. Consider the function $$S(x)=\sum_{j=1}^{d} Q_j(x)^2,\quad x \in S^{n-1}\,.$$ We claim that $S$ is rotation invariant, and, hence, constant. Since $d\sigma$ is a probability measure this constant must be $
\sum_{j=1}^{d} \int Q_j(x)^2\,d\sigma(x) = d$. Now let $Q$ be a homogeneous polynomial of degree $q$ and set $Q= \sum_{j=1}^d
\lambda_j\,Q_j$. Then $$|Q(x)| \leq \left( \sum_{j=1}^d \lambda_j^2 \right)^{\frac{1}{2}}
\left( \sum_{j=1}^d Q_j(x)^2 \right)^{\frac{1}{2}} = \|Q
\|_2\,d^{\frac{1}{2}}, \quad x \in S^{n-1}\,,$$ which proves the lemma because $d = \binom{n+q-1}{q} \simeq q^{n-1}$ ([@SW p. 139]).
To show the claim take a rotation $\rho$. Then we have $$Q_j(\varrho(x))= \sum_{j=1}^d a_{jk}\,Q_k(x)\,,$$ for some matrix $(a_{jk})$ which is orthogonal, because the polynomials $Q_j(\rho(x))$ form also an orthonormal basis due to the rotation invariance of $\sigma$. Hence $$\sum_{j=1}^{d} Q_j(\varrho(x))^2 = \sum_{j=1}^{d} Q_j(x)^2 ,\quad x
\in S^{n-1}\,.$$
Let us return now to the context of the Theorem. Thus $T$ is an even smooth homogeneous Calderón-Zygmund operator with kernel $K(x) = \Omega(x) /
|x|^n$, and the expansion of $\Omega$ in spherical harmonics is $$\label{eq33}
\Omega(x) = \sum_{j \geq 1}^\infty \,P_{2j}(x)\,.$$ By hypothesis there is a homogeneous harmonic polynomial $P$ of degree $2d$ which divides each $P_{2j}$. In other words, $P_{2j} = P
\,Q_{2j-2d}$, where $Q_{2j-2d}$ is a homogeneous polynomial of degree $2j-2d$. We want to show that the series $ \sum_{j}
Q_{2j-2d}(x)$ is convergent in $C^\infty(S^{n-1})$, that is, that for each positive integer $M$ $$\label{eq34}
\sum_{j\geq d} j^M \,\|Q_{2j-2d}\|_{\infty} < \infty\,.$$ The next lemma states that when one divides two homogeneous polynomials, then the supremum norm (on $S^{n-1}$) of the quotient is controlled by the supremum norm of the dividend.
\[Div\] Let $P$ be a homogeneous polynomial non identically zero. Then there exists a positive $\epsilon$ and a positive constant $C=C(n,P)$ such that $$\| Q\|_\infty \leq C \,q^{2(n-1)/ \epsilon}\,\|P\,Q \|_\infty\,,$$ for each homogeneous polynomial $Q$ of degree $q$.
Assume that we can prove that for some positive $\epsilon$ $$\label{eq35}
\int_{|x|=1} \frac{1}{|P(x)|^\epsilon} \,d\sigma(x) < \infty \,.$$ Then, by Lemma 6 and Schwarz’s inequality, $$\begin{split}
\|Q\|_ \infty & \leq \,C \, q^{(n-1)/2}\, \|Q\|_2 \\*[5pt]
& \leq C \, q^{(n-1)/2} \left(\int_{|x|=1} \frac{1}{|P(x)|^\epsilon} \,d\sigma(x)\right)^{1/4}
\left(\int_{|x|=1} |P(x)|^\epsilon \, |Q|^4 \,
d\sigma(x)\right)^{1/4} \\*[5pt] & \leq C \, q^{(n-1)/2}\,
\|P\,Q\|_\infty^{\epsilon/4} \left(\int_{|x|=1} |Q|^{4-\epsilon} \,
d\sigma(x)\right)^{1/4} \\*[5pt]
& \leq C \, q^{(n-1)/2}\, \|P\,Q\|_\infty^{\epsilon/4} \,
\|Q\|_\infty^{1-\epsilon/4}\,,
\end{split}$$ which completes the proof of the lemma.
Let us prove . Let $d$ be the degree of $P$. By a well-known result of Ricci and Stein [@RS], $|P(x)|$ is a weight in the class $A^\infty$. Indeed, if $\epsilon \,d < 1$, then $$\int_{ |x| < 1} \frac{1}{|P(x)|^\epsilon} \,dx \le C(\epsilon,d)
\left(\int_{|x| < 1} |P(x)| \,dx \right)^{-\epsilon} < \infty \,.$$ Since $P$ is an homogeneous polynomial, follows by changing to spherical coordinates.
Now may be proved readily from Lemma 7 and . Indeed, setting $M_0 = 2(n-1)/\epsilon$, we have $$\|Q_{2j-2d}\|_{\infty} \leq \,C(n,P)\, (2j)^{M_0}
\,\|P_{2j}\|_{\infty}\,,$$ and $$\sum_{j\geq d} j^M \,\|Q_{2j-2d}\|_{\infty} \leq \,C(n,P)
\sum_{j\geq 1} (2j)^{M+M_0} \,\|P_{2j}\|_{\infty} < \infty\,.$$
The scheme for the proof of the sufficient condition in the general case is as follows. Taking a large partial sum of the series we pass to a polynomial operator $T_N$ (associated to a polynomial of degree $2N$), which still satisfies the hypothesis $(iii)$ of the Theorem. Then we may apply the construction of section 3 to $T_N$ and get functions $b_N$ and $\beta_N$. Unfortunately what was done in section 3 does not give any uniform estimate in $N$, which is precisely what we need to try a compactness argument. The rest of the section is devoted to get the appropriate uniform estimates and to describe the final compactness argument.
By hypothesis, $T=R\circ U$, where $R$ is the higher order Riesz transform associated to the harmonic polynomial $P$ of degree $2d$ that divides all $P_{2j}$, and $U$ is invertible in the algebra $A$. The Fourier multiplier of $T$ is $$\sum_{j=d}^\infty \gamma_{2j}\, \frac{P_{2j}(\xi)}{|\xi|^{2j}} =
\gamma_{2d}\,\frac{P(\xi)}{|\xi|^{2d}}\,\sum_{j\geq d}
\frac{\gamma_{2j}}
{\gamma_{2d}}\,\frac{Q_{2j-2d}(\xi)}{|\xi|^{2j-2d}},\quad \xi \in
{{\mathbb R}^n}\setminus \{0\}\,.$$ Therefore the Fourier multiplier of $U$ is $$\label{eq36}
\mu(\xi) = \gamma_{2d}^{-1}\, \sum_{j\geq d} \gamma_{2j}
\,\frac{Q_{2j-2d}(\xi)}{|\xi|^{2j-2d}} \,,$$ and the series is convergent in $C^\infty(S^{n-1})$ because $\gamma_{2j}\simeq (2j)^{-n/2}$ [@SW p. 226]. Set, for $N\geq
d$, $$\label{eq37}
\mu_N(\xi) = \gamma_{2d}^{-1}\, \sum_{j=d}^N \gamma_{2j}
\,\frac{Q_{2j-2d}(\xi)}{|\xi|^{2j-2d}},\quad \xi \in {{\mathbb R}^n}\setminus
\{0\}\,.$$ If $$K_N (x) = \sum_{j=d}^N \frac{P_{2j}(x)}{|x|^{2j+n}},\quad x \in {{\mathbb R}^n}\setminus \{0\}\,,$$ and $T_N$ is the polynomial operator with kernel $K_N$, then $T_N =
R \circ U_N$, where $U_N$ is the operator in the algebra $A$ with Fourier multiplier $\mu_N(\xi)$. From now on $N$ is assumed to be big enough so that $\mu_N(\xi)$ does not vanish on $S^{n-1}$. In fact, we will need later on the inequality $$\label{eq38}
| \partial^\alpha \mu^{-1}_N(\xi) | \leq C,\quad |\xi|= 1,\quad
0 \leq |\alpha| \leq 2(n+3)\,,$$ which may be taken for granted owing to the convergence in $C^\infty(S^{n-1})$ of the series . In $C$ is a positive constant depending only on the dimension $n$ and $\mu$.
Notice that $T_N$ satisfies condition $(iii)$ in the Theorem (with $T$ replaced by $T_N$), because $\mu_N(\xi) \neq 0$, $|\xi| = 1$, and so we can apply the results of section 3. In particular, $$K_N (x) \chi_{{{{\mathbb R}^n}\setminus \overline B}}(x) = T_N(b_N)(x) + T_N(\beta_N)(x)\,,$$ where $b_N$ and $\beta_N$ are respectively the functions $b$ and $\beta$ defined in . It is important to remark that $b_N$ does not depend on $T$. As shows, the function $b_N$ depends on $N$ only through the fundamental solution of the operator $\triangle^N$. The uniform estimate we need on $b_N$ is given by part (i) of the next lemma. The polynomial estimates in $N$ of (ii) and (iii) are also basic for the compactness argument we are looking for.
\[bN\] There exist a constant $C$ depending only on $n$ such that
1. $$|\widehat{b_N}(\xi)| \leq C,\quad \xi \in {{\mathbb R}^n}\,,$$
2. $$\| b_N \|_{L^\infty(B)} \leq C \,(2N)^{2n+2}\,,$$ and
3. $$\|\nabla b_N \|_{L^\infty(B)} \leq C \,(2N)^{2n+4}\,.$$
We first prove (i). Let $h_1,\dotsc,h_d$ be an orthonormal basis of the subspace of $L^2(d\sigma)$ consisting of all homogeneous harmonic polynomials of degree $2N$. As in the proof of Lemma 6 we have $h_1^2 +\dotsb+h_d^2 = d$, on $S^{n-1}$. Set $$H_j(x)= \frac{1}{\gamma_{2N} \sqrt{d}}\,h_j(x), \quad x \in {{\mathbb R}^n}\,,$$ and let $S_j$ be the higher order Riesz transform with kernel $K_j(x) = H_j(x)/|x|^{2N+n}$. The Fourier multiplier of $S_j^2$ is $$\frac{1}{d}\,\frac{h_j(\xi)^2}{|\xi|^{4N}}, \quad 0 \neq \xi \in
{{\mathbb R}^n}\,,$$ and thus $$\sum_{j=1}^d S_j^2 = I \,.$$ By , we get $$K_j(x)\,\chi_{{{{\mathbb R}^n}\setminus \overline B}}(x)= S_j(b_N)(x),\quad x \in {{\mathbb R}^n},\quad 1 \leq j
\leq d\,,$$ and so $$\label{eq39}
b_N = \sum_{j=1}^d S_j\left( K_j(x) \,\chi_{{{{\mathbb R}^n}\setminus \overline B}}(x)\right)\,.$$ We now appeal to a lemma of Calderón and Zygmund ([@CZ]; see [@LS] for a simpler proof), which can be stated as follows.
If $K$ is the kernel of a higher order Riesz transform, then, for some constant $C$ depending only on $n$, $$|\widehat{(K(x)\,\chi_{{{{\mathbb R}^n}\setminus \overline B}}(x))}(\xi)| \leq C \,
|\widehat{\left(P.V. \,K(x) \right)}(\xi)|, \quad \xi \in {{\mathbb R}^n}\setminus \{0\}\,.$$
By and the preceding lemma, we get $$\begin{split}
|\widehat{b_N}(\xi)| & \leq \sum_{j=1}^{d}
|\widehat{P.V.\,\,K_j(x)}(\xi)|\,|\widehat{(K_j(x)\,\chi_{{{{\mathbb R}^n}\setminus \overline B}}(x))}(\xi)| \\
& \leq C \, \sum_{j=1}^{d} |\widehat{P.V.\,\,K_j(x)}(\xi)|^2 \\
& = C\,.
\end{split}$$
We now turn to the proof of (ii) in Lemma \[bN\]. In view of the expression for $b_N$, we apply Lemma 5 to the operators $S_j$ and the functions $K_j(x)$, which satisfy a Lipschitz condition on ${{{\mathbb R}^n}\setminus \overline B}$. We obtain $$\label{eq40}
\|b_N\|_\infty \leq C\, d\, \max_{1 \leq j \leq d} \|K_j\|_{CZ}\,
(\|K_j\|_{L^\infty({{{\mathbb R}^n}\setminus \overline B})} + \|K_j\|_{\operatorname{Lip}(1,{{{\mathbb R}^n}\setminus \overline B})}) \,.$$ As it is well known, $d \simeq (2N)^{n-2}$ [@SW p. 140]. On the other hand $$\|K_j \|_{CZ} \leq \|H_j\|_\infty+\|\nabla H_j\|_\infty\,,$$ where the supremum norms are taken on $S^{n-1}$. Clearly $$\|H_j\|_\infty = \frac{1}{\gamma_{2N}}
\|\frac{h_j}{\sqrt{d}}\|_\infty \leq \frac{1}{\gamma_{2N}} \simeq
(2N)^{n/2}\,.$$ For the estimate of the gradient of $H_j$ we use the inequality [@St p. 276] $$\label{eq40bis}
\| \nabla H_j\|_\infty \leq C\,(2N)^{n/2+1}\,\|H_j\|_2\,,$$ where the $L^2$ norm is taken with respect to $d\sigma$. Since the $h_j$ are an orthonormal system, $$\|H_j\|_2 = \frac{1}{\sqrt{d}\,\gamma_{2N}} \simeq
\frac{(2N)^{n/2}}{(2N)^{(n-2)/2}} \simeq 2N\,.$$ Gathering the above inequalities we get $$\|K_j\|_{CZ} \leq C\,(2N)^{n/2+2}\,.$$ On the other hand, a straightforward computation yields $$\|K_j\|_{L^\infty({{{\mathbb R}^n}\setminus \overline B})} + \|K_j\|_{\operatorname{Lip}(1,{{{\mathbb R}^n}\setminus \overline B})} \leq
C\, N \|H_j\|_\infty + \|\nabla H_j \|_\infty \leq C\,
(2N)^{n/2+2}\,,$$ and therefore $$\| b_N \|_{L^\infty(B)} \leq C \,(2N)^{n-2}\,(2N)^{n/2+2} \,
(2N)^{n/2+2} = C\,(2N)^{2n+2}\,.$$ We are only left with the proof of (iii) in Lemma \[bN\]. Recalling the definition of $b$ in we see that $b_N$ has the form $$b_N(x) = \alpha_0 + \alpha_1 \, |x|^2 + \dotsb + \alpha_{N-1}
\,|x|^{2N-2} ,\quad |x| < 1 \,,$$ for some real coefficients $\alpha_j$, $0 \leq j \leq N-1$. Define the polynomial $p(t)$ of the real variable $t$ as $$p(t)= \alpha_0 + \alpha_1 \, t^2 +\dotsb + \alpha_{N-1} \,
t^{2N-2}\,,$$ so that $b_N(x)= p(|x|)$, $|x| < 1$. By part (ii) of the lemma $$\sup_{0 \leq t \leq 1} |p(t)| \leq C\,(2N)^{2n+2}\,,$$ and thus, appealing to Markov’s inequality [@Lo p. 40], $$\sup_{0 \leq t \leq 1} |p'(t)| \leq (2N-2)^2 \sup_{0 \leq t \leq 1}
|p(t)| \leq C\,(2N)^{2n+4}\,.$$ Now (iii) follows from the obvious identity $\frac{\partial
b_N}{\partial x_j} = p'(|x|)\, \frac{\partial |x|}{\partial x_j}$, which gives $ |\nabla b_N (x)| \leq p'(|x|)$, $|x| < 1$.
Our goal is now to show that under condition $(iii)$ of the Theorem we can find a function $\gamma$ in $L^\infty({{\mathbb R}^n})$ such that $$\label{eq42}
K(x)\chi_{{{{\mathbb R}^n}\setminus \overline B}}(x)= T(\gamma)(x),\quad x \in {{\mathbb R}^n}\,.$$ If $T$ is a polynomial operator this was proven in the preceding section for a $\gamma$ of the form $b+\beta$ (see ). The approach we take up now has the advantage that when applied to $T_N$ gives a uniform bound on $\gamma_N= b_N+\beta_N$.
Since $\Omega$ has the expansion in spherical harmonics, we have $$\begin{split}
K(x)\chi_{{{{\mathbb R}^n}\setminus \overline B}}(x)& = \sum_{j\geq 1}
\frac{P_{2j}(x)}{|x|^{2j+n}}\,\chi_{{{{\mathbb R}^n}\setminus \overline B}}(x)\\*[5pt]
& = \sum_{j\geq 1} T_j(b_{j})(x)\,,
\end{split}$$ where $T_j$ is the higher order Riesz transform with kernel $
P_{2j}(x) / |x|^{2j+n}$ and $b_{j}$ is the function constructed in section 2 (see and ). The Fourier multiplier of $T_j$ is $$\gamma_{2j}\, \frac{P_{2j}(\xi)}{|\xi|^{2j}} =
\gamma_{2d}\,\frac{P(\xi)}{|\xi|^{2d}}\, \frac{\gamma_{2j}}
{\gamma_{2d}}\,\frac{Q_{2j-2d}(\xi)}{|\xi|^{2j-2d}},\quad \xi \in
{{\mathbb R}^n}\setminus \{0\}\,.$$ Let $S_j$ be the operator whose Fourier multiplier is $$\label{eq42bis}
\frac{\gamma_{2j}}
{\gamma_{2d}}\,\frac{Q_{2j-2d}(\xi)}{|\xi|^{2j-2d}},\quad \xi \in
{{\mathbb R}^n}\setminus \{0\}\,,$$ so that $T_j = R \circ S_j $. Then $$\begin{split}
K(x)\chi_{{{{\mathbb R}^n}\setminus \overline B}}(x)& = \sum_{j\geq d} (R \circ S_j )(b_{j})\\*[5pt]
&= \sum_{j\geq d} T \left((U^{-1} \circ S_j )(b_{j})\right)\\*[5pt]
& = T\left(\sum_{j\geq d} (U^{-1}\circ S_j)(b_j)\right)\,.
\end{split}$$ The latest identity is justified by the absolute convergence of the series $\sum_{j\geq d} (U^{-1}\circ S_j)(b_j)$ in $L^2({{\mathbb R}^n})$, which follows from the estimate $$\begin{split}
\sum_{j\geq d} \|(U^{-1}\circ S_j)(b_j)\|_2 & \leq C\, \sum_{j\geq
d} \|Q_{2j-2d}\|_\infty \,\|b_j\|_{L^2({{\mathbb R}^n})}\\*[4pt] & \leq C\,
\sum_{j\geq d} \|Q_{2j-2d}\|_\infty \,\|b_j\|_{L^\infty(B)}\\*[5pt]
& \leq C\, \sum_{j\geq d} (2N)^{2n+2} \, \|Q_{2j-2d}\|_\infty \ <
\infty\,.
\end{split}$$ We claim now that the series $\sum_{j\geq d} (U^{-1}\circ S_j)(b_j)$ converges uniformly on ${{\mathbb R}^n}$ to a function $\gamma$, which will prove . Observe that the operator $U^{-1}\circ S_j \in
A$ is not necessarily a Calderón-Zygmund operator because the integral on the sphere of its multiplier does not need to vanish. However it can be written as $ U^{-1}\circ S_j = c_j I + V_j$, where $$c_j = \frac{\gamma_{2j}} {\gamma_{2d}}\,\int_{S^{n-1}} \mu(\xi)^{-1}
\, Q_{2j-2d}(\xi)\,d\sigma(\xi)$$ and $V_j$ is the Calderón-Zygmund operator with multiplier $$\label{eq43}
\mu(\xi)^{-1} \frac{\gamma_{2j}}
{\gamma_{2d}}\,\frac{Q_{2j-2d}(\xi)}{|\xi|^{2j-2d}} - c_j \,.$$ Now $$\sum_{j\geq d} (U^{-1}\circ S_j)(b_j) = \sum_{j\geq d} c_j\,b_j +
\sum_{j\geq d} V_j(b_j)$$ and the first series offers no difficulties because, by Lemma 8 (ii) and $$\sum_{j\geq d} |c_j|\, \|b_j\|_{L^\infty(B)} \leq C\, \sum_{j\geq d}
(2j)^{-n/2} (2j)^{2n+2} \|Q_{2j-2d} \|_\infty < \infty\,.$$ The second series is more difficult to treat. By Lemma 5 and Lemma 8 (ii) and (iii), $$\begin{split}
\|V_j(b_j) \|_{L^\infty({{\mathbb R}^n})} & \leq C\, \| V_j \|_{CZ} \left(\|b_j\|_{L^\infty(B)}+ \| \nabla b_j\|_{L^\infty(B)}\right) \\*[3pt]
& \leq C\, (2j)^{2n+4} \, \| V_j\|_{CZ} \,.
\end{split}$$ Estimating the Calderón-Zygmund constant of the kernel of the operator $V_j$ is not an easy task, because we do not have an explicit expression for the kernel. We do know, however, the multiplier of $V_j$. We need a way of estimating the constant of the kernel in terms of the multiplier and this is what the next lemma supplies.
Let $V$ be a smooth homogeneous Calderón-Zygmund operator withFourier multiplier $m$. Then for some constant $C$ depending only on $n$, $$\| V \|_{CZ} \leq C\, \|\triangle_{S}^{n+3} m \|_2^{1/2}\, \| m
\|_2^{1/2}\,,$$ where $\triangle_S$ is the spherical Laplacean and the $L^2$ norm is taken with respect to $d\sigma$.
Let $\omega(x)/|x|^n$ be the kernel of $V$, so that $\omega$ is a homogeneous function of degree zero, of class $C^\infty(S^{n-1})$ and with zero integral on the sphere. Consider the expansion of $\omega$ in spherical harmonics $\omega (x)= \sum_{j \geq 1}
p_j(x)$, $|x|=1$, so that the kernel of $V$ is $\sum_{j \geq 1}
p_j(x)/|x|^{j+n}$, $x \in {{\mathbb R}^n}\setminus\{0\}$ and its Fourier multiplier is $m(\xi) = \sum_{j \geq 1} \gamma_j\, p_j(\xi)$, $|\xi|=1..$ By the definition of the constant of the kernel of a Calderón-Zygmund operator we have $$\| V \|_{CZ} \leq C\, \sum_{j \geq 1} \left(j\,\|p_j\|_\infty +
\|\nabla p_j \|_\infty \right)\,,$$ where the supremum is taken on $S^{n-1}$. By with $H_j$ replaced by $p_j$, and Lemma 6 $$\begin{split}
\| V \|_{CZ}& \leq C\, \sum_{j \geq 1}
\left(j^{1+(n-1)/2}\,\|p_j\|_2 + j^{n/2+1}\,\| p_j \|_2\right)
\\*[5pt] & \leq C\, \sum_{j\geq 1} \left(j^{n/2+2}\, \| p_j \|_2
\right)\,.
\end{split}$$ Since $\gamma_j \simeq j^{-n/2}$, the above sum can be estimated, using Schwarz’s inequality and with $\Omega$ replaced by $m$ and $P_j$ by $\gamma_j \, p_j$, by $$\begin{split}
\sum_{j\geq 1} j^{n+2}\, \| \gamma_j\, p_j \|_2 & \leq
C\,\left(\sum_{j\geq 1} j^{2n+6}\, \| \gamma_j\, p_j \|_2^2
\right)^{1/2} \\*[5pt] & \leq C\,\left(\sum_{j\geq 1}
\left(j(j+n-2)\right)^{n+3}\, \|\gamma_j\, p_j\|^2_2\right)^{1/2}
\\*[5pt] & = C\,\left((-1)^{n}\,\int_{S^{n-1}} \triangle^{n+3}_S m
\, m \,d\sigma \right)^{1/2}\\*[5pt] & \leq C\,
\|\triangle_{S}^{n+3} m \|_2^{1/2}\, \| m \|_2^{1/2}\,.
\end{split}$$
Since the multiplier of $V_j$ is given by and $\mu^{-1}$ is in $C^\infty(S^{n-1})$, Lemma 9 reduces the estimate of $\|V_j\|_{CZ}$ to the estimate of the $L^2(d\sigma)$ norm of $\nabla^k Q_{2j-2d}$, for $0 \leq k \leq 2(n+3)$. Let us consider first the case $k=1$.
Since $P_{2j}= P \,\,Q_{2j-2d}$, we have $$\nabla P_{2j} = \nabla P \,Q_{2j-2d} +P \, \nabla Q_{2j-2d}\,,$$ and so, by Lemma 7 and with $H_j$ replaced by $P_{2j}$, there is a large positive integer $M = M(n,P)$ such that $$\begin{split}
\|\nabla Q_{2j-2d}\|_\infty & \leq C\,\,j^M\, \|P\,\nabla
Q_{2j-2d}\|_\infty \\*[5pt] & \leq C \,j^M\,\left( \|\nabla P_{2j}
\|_\infty + \|Q_{2j-2d}\|_\infty \right) \\*[5pt] & \leq C\,j^M
\,\left( C\,j^{n/2+1}\,\|P_{2j} \|_2 + C\,j^M \,\|P_{2j} \|_\infty
\right)\\*[5pt] & \leq C\,j^M \,\|P_{2j} \|_2 \,,
\end{split}$$ where in the latest inequality $M$ has been increased without changing the notation.
By induction we get, for some large integer $M=M(n,P)$, $$\|\nabla^k Q_{2j-2d}\|_\infty \leq C\,j^M\,\|P_{2j}\|_2, \quad 0 \leq
k \leq 2(n+3)\,.$$ Therefore, the estimate we finally obtain for the constant of the kernel of $V_j$ is $$\| V_j \|_{CZ} \leq C\,j^M \,\|P_{2j}\|_2\,,$$ and thus $$\|V_j(b_j) \|_{L^\infty({{\mathbb R}^n})} \leq C\,j^M \,\|P_{2j}\|_2\,,$$ where again $M=M(n,P)$ is a positive integer. Hence the series $\sum_{j\geq d} (U^{-1}\circ S_j)(b_j)$ converges uniformly on ${{\mathbb R}^n}$ and the proof of is complete.
We are now ready for the discussion of the final compactness argument that will complete the proof of the sufficient condition. The reader is invited to review the definitions of the operators $T_N$ (with kernel $K_N$) and $U_N$ given in this section just before Lemma 8. We know from section 3 (see ) that $$\label{eq45}
K_N(x)\chi_{{{{\mathbb R}^n}\setminus \overline B}}(x)= T_N(b_N)(x)+ T_N(\beta_N)(x)\,.$$ On the other hand, by the construction of the function $\gamma$ we have just described, we also have $$\label{eq46}
K_N(x)\chi_{{{{\mathbb R}^n}\setminus \overline B}}(x)= T_N(\gamma_N)(x), \quad \gamma_N = \sum_{j\geq
d}^N (U_N^{-1}\circ S_j)(b_j)\,.$$ Notice that guaranties that the estimate of the supremum norm of $\gamma$ on the whole of ${{\mathbb R}^n}$ is applicable to the operator $T_N$, and thus we get an estimate for $\|\gamma_N\|_{L^\infty({{\mathbb R}^n})}$ which is uniform in $N$. Since $T_N$ is injective, and imply $$\label{eq46bis}
b_N+\beta_N = \gamma_N$$ and, in particular, we conclude that the functions $b_N+\beta_N$ are uniformly bounded in $L^\infty({{\mathbb R}^n})$, a fact that cannot be derived from the work done in section 3. It is worth mentioning that numerical computations indicate that $b_N$, and thus $\beta_N$, are not uniformly bounded. On the other hand, section 3 tells us that $\gamma_N$ satisfies the decay estimate with $\beta$ replaced by $\gamma_N$, which we cannot infer from the preceding construction of $\gamma$. The advantages of both approaches will be combined now to get both the boundedness and decay property for $\gamma$.
In view of and the expressions of the multipliers of $U_N$ and $S_j$ (see ), $$\widehat{\gamma_N}(\xi) = \sum_{j=d}^N
\frac{1}{\mu_N(\xi)}\,\frac{\gamma_{2j}}{\gamma_{2d}}\,\frac{Q_{2j-2d}(\xi)}{|\xi|^{2j-2d}}\,
\widehat{b_j}(\xi)\,,$$ which yields, by Lemma 8 and for $M=0$, $$\label{eq48}
\begin{split}
\|\widehat{\gamma_N} \|_{L^\infty({{\mathbb R}^n})} & \leq C\,\sum_{j= d}^N \|Q_{2j-2d}
\|_{\infty}\\*[5pt]
& \leq C\, \sum_{j= d}^\infty \|Q_{2j-2d}\|_{\infty} \\*[5pt]
& \leq C\,,
\end{split}$$ where $C$ does not depend on $N$. Recall that, from in section 3, we have $$\beta_N =U_N^{-1}(\beta_{1,N})\,,$$ with $\beta_{1,N}$ a bounded function supported on $B$ satisfying $\int \beta_{1,N}(x) \,dx =0$. Since $$\widehat{\beta_{1,N}} = \mu_N \, \widehat{\beta_{N}} = \mu_N
\,(\widehat{\gamma_{N}} - \widehat{b_{N}})\,,$$ we have, again by Lemma 8, $$\|\widehat{\beta_{1,N}} \|_{L^\infty({{\mathbb R}^n})} \leq C\,.$$ Therefore, passing to a subsequence, we may assume that, as $N$ goes to $\infty$, $$\widehat{b_{N}}\longrightarrow a_0 \quad\quad\quad
{\text{and}}\quad\quad\quad \widehat{\beta_{1,N}} \longrightarrow
a_1\,,$$ weak $\star$ in $L^\infty({{\mathbb R}^n})$. Hence $$b_{N}\longrightarrow \Phi_0 = \mathcal{F}^{-1}{a_0} \quad\quad\quad
{\text{and}}\quad\quad\quad \beta_{1,N} \longrightarrow \Phi_1 =
\mathcal{F}^{-1}{a_1}\,,$$ in the weak $\star$ topology of tempered distributions, $\mathcal{F}^{-1}$ being the inverse Fourier transform. In particular, $\Phi_0$ and $ \Phi_1$ are distributions supported on $\overline B$ and $$\label{eq49}
\langle\Phi_1,1\rangle = \lim_{N \rightarrow \infty} \int
\beta_{1,N}(x)\,dx = 0\,.$$
We would like now to understand the convergence properties of the sequence of the $\beta_N$’s . Since $$\widehat{\beta_{N}}(\xi) = \mu_N^{-1}(\xi) \,
\widehat{\beta_{1,N}}(\xi)\,,$$ and we have pointwise bounded convergence of $\mu_N^{-1}(\xi)$ towards $\mu^{-1}(\xi)$ on ${{\mathbb R}^n}\setminus\{0\}$, we get that $\widehat{\beta_{N}} \rightarrow \mu^{-1}\,a_1$, in the weak $\star$ topology of $L^\infty({{\mathbb R}^n})$. Thus $\beta_{N} \rightarrow
U^{-1}(\Phi_1)$ in the weak $\star$ topology of tempered distributions. Letting $N \rightarrow \infty$ in we obtain $$\Phi_0 + U^{-1}(\Phi_1) = \gamma\,.$$ We come now to the last key point of the proof, namely, that one has decay estimate $$\label{eq50}
|\gamma(x)| \leq \frac{C}{|x|^{n+1}},\quad |x|\geq 2\,.$$ Since $\Phi_0$ and $\Phi_1$ are supported on $\overline B$ and $U^{-1}(\Phi_1) = \lambda\,\Phi_1 + V(\Phi_1 )$, where $\lambda$ is a real number and $V$ a smooth homogeneous Calderón-Zygmund operator, it is enough to show that $V(\Phi_1 )$ has the appropriate behavior off the ball $B(0,2)$. Let $L$ be the kernel of $V$. Regularizing $\Phi_1$ one checks that, for a fixed $x$ satisfying $|x|\geq 2$, $$\label{eq51}
\begin{split}
V(\Phi_1 )(x) & = \langle \Phi_1, L(x-y)\rangle \\*[3pt] & =
\langle\Phi_1, L(x-y)-L(x)\rangle \,,
\end{split}$$ where the latest identity follows from . Since $\Phi_1 $ is a distribution supported on $\overline B$ there exists a positive integer $\nu$ and a constant $C$ such that $$\label{eq52}
|\langle\Phi_1,\varphi\rangle | \leq C\sup_{|\alpha|\leq \nu}
\sup_{|y|\leq 3/2} |\partial^\alpha \varphi(y)|\,,$$ for each infinitely differentiable function $\varphi$ on ${{\mathbb R}^n}$. The kernel $L$ satisfies $$|\frac{\partial^\alpha}{{\partial y}^\alpha} \left(
L(x-y)-L(x)\right)| \leq \frac{C_\alpha}{|x|^{n+1+|\alpha|}},\quad
|y| \leq 3/2\,,$$ and hence by and $$|V(\Phi_1)(x)| \leq \frac{C}{|x|^{n+1}},\quad |x|\geq 2\,,$$ which proves and then completes the proof of the sufficient condition in the general case.
Proof of the necessary condition: the polynomial case
======================================================
We assume in this section that $T$ is a polynomial operator with kernel $$K(x)=\frac{\Omega(x)}{|x|^n}=
\frac{P_2(x)}{|x|^{2+n}}+\frac{P_4(x)}{|x|^{4+n}}+\dotsb+\frac{P_{2N}(x)}{|x|^{2N+n}},\quad
x \neq 0\,,$$ where $P_{2j}$ is a homogeneous harmonic polynomials of degree $2j$. Let $Q$ be the homogeneous polynomial of degree $2N$ defined by $$Q(x)= \gamma_2 \, P_2(x)|x|^{2N-2}+ \dotsb+ \gamma_{2j}\,
P_{2j}(x)|x|^{2N-2j}+\dotsb+ \gamma_{2N}\,P_{2N}(x) \,.$$ Then $$\widehat{P.V.K}(\xi) = \frac{Q(\xi)}{|\xi|^{2N}},\quad \xi \neq 0\,.$$ Our assumption is now the $L^2$ estimate between $T^{\star}$ and $T$ (see $(ii)$ in the statement of the Theorem). Since the truncated operator $T^1$ at level $1$ is obviously dominated by $T^{\star}$, we have $$\int (T^1f)^2 (x)\,dx \leq \int (T^{\star}f)^2 (x)\,dx \leq C\,\int
(Tf)^2 (x)\,dx \,.$$ The kernel of $T^1$ is (see ) $$\label{eq53}
K(x)\,\chi_{{{{\mathbb R}^n}\setminus \overline B}}(x)= T(b)(x) + S(x)\,\chi_{B}(x)\,,$$ where $b$ is given in equation and $$-S(x)= Q(\partial)(A_0+A_1\,|x|^2+\dotsb+
A_{2N-1}\,|x|^{4N-2})(x),\quad x \in {{\mathbb R}^n}\,.$$ The reader may consult the beginning of section 3 to review the context of the definition of $S$. In view of we have, for each $f \in L^2({{\mathbb R}^n})$, $$\begin{split}
\|S\,\chi_{B} \star f \|_2 & \leq C\, \|T^1 f \|_2 + \|b \star Tf \|_2 \\*[3pt]
& \leq C\, (\|Tf \|_2 +\|\widehat{b} \|_\infty \|Tf \|_2) \\*[3pt] &
= C\,\|Tf\|_2 \,.
\end{split}$$ By Plancherel, the above $L^2$ inequality translates into a pointwise inequality between the Fourier multipliers, namely, $$\label{eq54}
|\widehat{S\,\chi_B}(\xi)| \leq C\, |\widehat{P.V. K}(\xi)| = C\,
\frac{|Q(\xi)|}{|\xi|^{2N}}\,.$$ Our next goal is to show that provides interesting relations between the zero sets of $Q$ and the $P_{2j}$. For each function $f$ on ${{\mathbb R}^n}$ set $Z(f)= \{x \in {{\mathbb R}^n}: f(x)=0 \}$.
$$Z(Q) \subset Z(P_{2j}), \quad 1 \leq j \leq N \,.$$
We know that $S$ has an expression of the form (see ) $$\label{eq54bis}
S(x)= \sum_{l=N+1}^{2N-1} \sum_{j=1}^{l-N}
c_{lj}\,P_{2j}(x)\,|x|^{2(l-N-j)}\,.$$ Since $\widehat{\chi_B} = G_m$, $m=n/2$, Lemma 3 yields $$\label{eq54tris}
\begin{split}
\widehat{S\,\chi_{B}}(\xi) & = S(\imath\, \partial)\,
\widehat{\chi_B}(\xi)\\*[5pt] & = \sum_{l=N+1}^{2N-1}
\sum_{j=1}^{l-N}
c_{lj}\,(-1)^{l-N}\,\,P_{2j}(\partial)\,\,\triangle^{l-N-j}\,G_{m}(\xi)\\*[5pt]
& = \sum_{l=N+1}^{2N-1} \sum_{j=1}^{l-N} \sum_{k=0}^{l-N-j}
c_{\,l,\,j,\,k}\,P_{2j}(\xi)\,|\xi|^{2(l-N-j-k)}
G_{m+2l-2N-k}(\xi)\,.
\end{split}$$ The function $G_p(\xi)$ is, for each $p \geq 0$, a radial function which is the restriction to the real positive axis of an entire function [@Gr A-8]. Set $\xi= r\, \xi_0$, $|\xi_0|=1$, $r\geq0$. Then $$\label{eq55}
\widehat{S \chi_B}(r\xi_0)= \sum_{p=1}^{\infty}
a_{2p}(\xi_0)\,r^{2p} \,,$$ and the power series has infinite radius of convergence for each $\xi_0$. Assume now that $Q(\xi_0) = 0$. Then, by , $\widehat{S \chi_B}(r\xi_0)= 0$ for each $r \geq 0$, and hence $a_{2p}(\xi_0) = 0$, for each $p \geq 1$. For $p=1$ one has $a_{2}(\xi_0) = P_2(\xi_0)\,C_2$, where $$C_2 = \sum_{l=N+1}^{2N-1} c_{\,l,\,1,\,l-N-1}\,G_{m+l-N+1}(0)\,.$$ It will be shown later that $C_2 \neq 0$, and then we get $P_2(\xi_0) =0 $. Let us make the inductive hypothesis that $P_2(\xi_0)= \dotsb=P_{2(j-1)}(\xi_0)=0$. Then we obtain, if $j \leq
N-1$, $a_{2j}(\xi_0) = P_{2j}(\xi_0)\,C_{2j}$, where $$\label{eq55bis}
C_{2j} = \sum_{l=N+j}^{2N-1} c_{\,l,\,j,\,l-N-j}\,G_{m+l-N+j}(0)\,.$$
Since we will show that $C_{2j} \neq 0$, $P_{2j}(\xi_0) =0$, $1
\leq j \leq N-1$. We have $$0= Q(\xi_0)= \sum_{j=1}^N \gamma_{2j} \, P_{2j}(\xi_0)\,,$$ and so we also get $P_{2N}(\xi_0)=0$. Therefore the zero sets Lemma is completely proved provided we have at our disposition the following.
\[L10\] $$C_{2j} = \frac{- \pi^{\frac{n}{2}}}{V_{n}\, 2^{\frac{n}{2}}\, \Gamma({\frac{n}{2}}+1)} \;
\frac{(-1)^j }{j\, 4^{j} \, \Gamma (2j+{\frac{n}{2}})},\quad 1 \le j \le N-1
\,.$$
The proof of Lemma 10 is lengthy and rather complicated from the computational point of view, and so we postpone it to section 7.
Notice that, although the constants $C_{2j}$ are non-zero, they become rapidly small as the index $j$ increases.
The reason why Lemma 10 is involved is that one has to trace back the exact values of the constants $C_{2j}$ from the very beginning of our proof of . This forces us to take into account the exact values of various other constants. For instance, those which appear in the expression of the fundamental solution of $\triangle^N$ and the constants $A_0, A_1,\dotsc, A_{2N-1}$ in formula . Finally, we need to prove some new identities involving a triple sum of combinatorial numbers, in the spirit of those that can be found in the book of R. Graham D. Knuth and O.Patashnik [@GKP].
The following is elementary folklore, but is proved here for the reader’s sake.
If $f$ is a real valued continuous function on ${{\mathbb R}^n}$which changes sign, then $H^{n-1}(Z(f)) > 0$, $H^{n-1}$ being Hausdorff measure in dimension $n-1$. In particular, the Hausdorff dimension of $Z(f)$ is at least $n-1$.
Assume, without loss of generality, that $f(0)> 0$ and $f(p)< 0$, where $p=(0,\dotsc,0,1)$. For $\epsilon>0$ small enough we have $f(x) > 0$ if $|x|< \epsilon$ and $f(x)< 0$ if $|x -p|< \epsilon$. Set $B= \{x \in {{\mathbb R}^n}: x_n=0\quad {\text{and}}\quad |x|<\epsilon \}$. Bolzano’s theorem tells us that, for each $x \in B$, $f$ vanishes at some point of the segment $(x_1,\dotsc,x_{n-1},t)$, $0 \leq t \leq
1$. Hence the orthogonal projection of the set $Z(f)$ onto the hyperplane $\{x:x_n =0\}$ contains $B$ and so $H^{n-1}(Z(f)) \geq
H^{n-1}(B)
>0$.
We turn now our attention to an algebraic lemma which plays a key role in obtaining the necessary condition we are looking for.
\[divisio\] Let $F$ and $G$ be polynomials in $\mathbb{R}[x_1,\dotsc,x_n]$. Assume that $G$ is irreducible and that $H^{n-1}(Z(F)\cap Z(G)) >
0$. Then there exists a polynomial $H$ in $\mathbb{R}[x_1,\dotsc,x_n]$ such that $F=G\,H$.
Denote by $V(P)$ the complex hyper-surface $ \{z \in \mathbb{C}^n :
P(z) = 0 \}$ of a polynomial $P$. By hypothesis $V(F)\cap V(G)$ is not empty. If $V(G)$ is not contained in $V(F)$ then the complex dimension of $V(G)\cap V(F)$ is not greater than $n-2$ [@K 3.2 p. 131]. Since the real dimension of a variety is less than or equal to the complex dimension, we conclude that $Z(G)\cap Z(F)$ has real dimension not greater than $n-2$, which contradicts the fact that it has positive $n-1$-dimensional Hausdorff measure. Thus $V(G)
\subset V(F)$, and therefore $F=G\,H$ for some polynomial $H$ in $\mathbb{C}[x_1,\dotsc,x_n]$. Since $F$ and $G$ have real coefficients, the same happens to $H$.
We proceed to the proof of the necessary condition.
Let $j_0$ be the first positive index such that $P_{2j_0}$ does not vanish identically. We want to show that $P_{2j_0}$ divides $P_{2j}$ for $j_0 \leq j \leq N$.
Since $\mathbb{R}[x_1,\dotsc,x_n]$ is a unique factorization domain we can express $P_{2j_0}$ as a product of irreducible factors, say $R_k$, $1 \leq k \leq M$, which are also homogeneous. Clearly $Z(P_{2j_0}) = \cup_k Z(R_k)$ and so $$Z(Q) = \bigcup_k (Z(Q)\cap Z(R_k))\,.$$ Since the integral of $Q$ on the sphere is $0$, $Q$ changes sign and thus by the Dimension Lemma there is at least a $k$ such that $H^{n-1}(Z(Q)\cap Z(R_k)) > 0$. Change notation if necessary so that $k=1$. Then $R_1$ divides $Q$, by the Division Lemma. We may also apply the Division Lemma to $R_1$ and $P_{2j}$ for each $j_0
\leq j \leq N$, because $Z(Q)\cap Z(R_1) \subset Z(P_{2j})\cap
Z(R_1) $ by the Zero Sets Lemma. Hence $R_1$ also divides $P_{2j}$, for $j_0 \leq j \leq N$. Set $$Q= R_1\,Q_1 \qquad {\text{and}} \qquad P_{2j} = R_1\,
P_{2j,1},\quad j_0 \leq j \leq N\,,$$ for certain homogeneous polynomials $Q_1$ and $P_{2j,1}$.
If $M=1$ we are done. Otherwise our intention is to repeat as many times as we can the above division process. With this in mind we use to rewrite inequality in the form $$\label{eq56}
\left| \sum_{p=1}^{\infty} a_{2p}(\xi_0)\,r^{2p}\right| \leq C\,
|Q(\xi_0)| ,\quad 0< r \,.$$ The definition of the coefficients $a_{2p}$ and show that there exist real numbers $\mu_j(p)$ such that $$a_{2p}(\xi_0) = \sum_{j=j_0}^{N-1} \mu_j(p)\,P_{2j}(\xi_0)$$ and so $$\label{eq57}
\begin{split}
a_{2p}(\xi_0) & = R_1(\xi_0) \,\sum_{j=j_0}^{N-1}
\mu_j(p)\,P_{2j,1}(\xi_0) \\
& = R_1(\xi_0)\, a_{2p,1}(\xi_0) \,,
\end{split}$$ where the last identity provides the definition of the numbers $
a_{2p,1}(\xi_0)$. We can simplify the common factor $R_1(\xi_0)$ in and get $$\label{eq58}
\left| \sum_{p=1}^{\infty} a_{2p,1}(\xi_0)\,r^{2p}\right| \leq C\,
|Q_1(\xi_0)| ,\quad 0< r \,.$$ Equipped with , we are ready to begin the second step in the division process. If $Q_1(\xi_0)=0 $, then $a_{2p,1}(\xi_0) =
0$, for each $p \geq 1$. Hence, as in the proof of the Zero Sets Lemma, $$Z(Q_1) \subset Z(P_{2j,1}), \quad j_0 \leq j \leq N \,.$$ To apply the Division Lemma we need to ascertain that the zero set of $Q_1$ is big enough and for that it suffices to show, by the Dimension Lemma, that $Q_1$ changes sign. As we are assuming that $M$ is greater than $1$, the degree of $R_1$ is less than $2j_0$. Considering the expansions of $R_1$ and $Q$ in spherical harmonics, we see that they are orthogonal in $L^2(d\sigma)$ [@St p. 69]. Hence $$0= \int R_1(\xi)\,Q(\xi) \,d\sigma(\xi) = \int R_1^2(\xi)\,Q_1(\xi)
\,d\sigma(\xi)\,,$$ which tells us that $Q_1$ changes sign.
Since $P_{2j_0,1} = \prod_{k=2}^M R_k $, we conclude that one of the $R_k$, say $R_2$, divides the $P_{2j,1}$, $j_0 \leq j \leq N$. An inductive argument gives that $P_{2j_0}$ divides the $P_{2j}$, $j_0
\leq j \leq N$. At the $k$-th step one should observe that $Q=
\prod_{l=1}^k R_l \,Q_k$ and $$0= \int \prod_{l=1}^{k}R_l(\xi)\,Q(\xi) \,d\sigma(\xi) = \int
\prod_{l=1}^{k}R_l^2(\xi)\,Q_k(\xi) \,d\sigma(\xi)\,,$$ so that $Q_k$ changes sign. It is also important to remark that we have $$a_{2p,k}(\xi_0) = \sum_{j=j_0}^{N-1} \mu_j(p)\,P_{2j,k}(\xi_0),\quad
1 \leq k \leq M\,.$$ Thus at the $M$-th step we get for $p=j_0$ $$\label{eq58bis}
a_{2j_0,M}(\xi_0) = C_{2j_0} \neq 0\,.$$ We have $P_{2j}= P_{2j_0}\,Q_{2j-2j_0}$ for some homogeneous polynomials $Q_{2j-2j_0}$ of degree $2j-2j_0$ and so $$\begin{split}
Q(\xi) & = \sum_{j=j_0}^N \gamma_{2j}\,
P_{2j}(\xi)|\xi|^{2N-2j} \\
& = P_{2j_0}(\xi)\, \sum_{j=j_0}^N \gamma_{2j}\,
Q_{2j-2j_0}(\xi)|\xi|^{2N-2j} \,.
\end{split}$$ By and the definition of the coefficients $a_{2p,M}(\xi_0)$, for $|\xi_0|=1$ and $0<r$, we get $$\left| \sum_{p=j_0}^{\infty} a_{2p,M}(\xi_0)\,r^{2p}\right| \leq C
\left| \sum_{j=j_0}^N \gamma_{2j}\,
Q_{2j-2j_0}(\xi_0)\right|,\quad 0< r \,.$$ Taking into account we conclude that $$\sum_{j=j_0}^N \gamma_{2j}\, Q_{2j-2j_0}(\xi_0) \neq 0,\quad
|\xi_0|= 1 \,,$$ which completes the proof of the necessary condition in the polynomial case.
Proof of the necessary condition: the general case
===================================================
In this section the kernel of our operator has the general form $K(x)=\Omega(x)/|x|^n$ with $\Omega$ a homogeneous function of degree $0$, with vanishing integral on the sphere and of class $C^\infty(S^{n-1})$. Then $\Omega(x) = \sum_{j\geq 1}^\infty
P_{2j}(x)/|x|^{2j}$ with $P_{2j}$ a homogeneous harmonic polynomial of degree $2j$. The strategy consists in passing to the polynomial case by looking at a partial sum of the series above. Set, for each $N \geq 1$, $K_N(x)=\Omega_N(x)/|x|^n$, where $\Omega_N(x) =
\sum_{j= 1}^N P_{2j}(x)/|x|^{2j}$, and let $T_N$ be the operator with kernel $K_N$. The difficulty is now that there is no obvious way of obtaining the inequality $$\label{eq58tris}
\| T_N^{\star} f \|_2 \leq C \| T_N f \|_2,\quad f \in L^2({{\mathbb R}^n})
\,,$$ from our hypothesis, namely, $$\| T^{\star} f \|_2 \leq C \| T f \|_2,\quad f \in L^2({{\mathbb R}^n}) \,.$$ Instead we try to get with $\| T_N f \|_2$ replaced by $\| Tf \|_2$ in the right hand side plus an additional term which becomes small as $N$ tends to $\infty$. We start as follows. $$\begin{split}
\| T^1_Nf\|_2 & \leq \| T^1f\|_2 + \| T^1f -T^1_N f\|_2 \\*[5pt]
& \leq C\, \|Tf\|_2 +\| \sum_{j>N} \frac{P_{2j}(x)}{|x|^{2j+n}}\,
\chi_{{{{\mathbb R}^n}\setminus \overline B}}* f\|_2 \,.
\end{split}$$ By there exists a bounded function $b_j$ supported on $B$ such that $$\frac{P_{2j}(x)}{|x|^{2j+n}}\, \chi_{{{{\mathbb R}^n}\setminus \overline B}}(x) = P.V.
\frac{P_{2j}(x)}{|x|^{2j+n}}* b_j\,.$$ By Lemma 8 (i) $\|\widehat{b_j}\|_{L^\infty({{\mathbb R}^n})}$ is bounded uniformly in $j$ and then an application of Plancherel yields $$\begin{split}
\left \| \sum_{j>N} \frac{P_{2j}(x)}{|x|^{2j+n}}\, \chi_{{{{\mathbb R}^n}\setminus \overline B}}*
f\right\|_2 & = \left \| \sum_{j>N} P.V.
\frac{P_{2j}(x)}{|x|^{2j+n}}* b_j * f\right\|_2 \\*[5pt] & \leq
C\, \left( \sum_{j>N} \|P_{2j}\|_\infty \right) \,\|f\|_2\,,
\end{split}$$ where the supremum norm is taken on the sphere. By applied to $T_N$ $$T^1_N f=K_N\, \chi_{{{{\mathbb R}^n}\setminus \overline B}}* f= K_N * b_N * f + S_N\, \chi_B * f \,.$$ Hence, for each $f$ in $L^2({{\mathbb R}^n})$, $$\begin{split}
\| S_N\chi_B *f\|_2 & \le \|T^1_N f\|_2 + \|K_N* f * b_N \|_2 \\*[5pt]
& \le C\| Tf\|_2 + C\, \left( \sum_{j>N} \|P_{2j}\|_\infty \right) \,\|f\|_2 + \| K_ N * f * b_N\|_2 \\*[5pt]
& \le C\| Tf\|_2+ \|Tf * b_N
\|_2 + C \,\left( \sum_{j>N} \|P_{2j}\|_\infty \right) \,\|f\|_2 \\*[5pt]
& \le C\| Tf\|_2 + C \,\left( \sum_{j>N} \|P_{2j}\|_\infty \right) \,\|f\|_2 \,,
\end{split}$$ where in the latest inequality Lemma 8 (i) was used. The above $L^2$ inequality translates, via Plancherel, into the pointwise estimate $$\label{eq59}
| \widehat{S_N \chi_B} (\xi)| \le C| \widehat{P.V. K} (\xi) |+ C\,
\left( \sum_{j>N} \|P_{2j}\|_\infty \right), \quad \xi \neq 0 \,.$$ The idea is now to take limits, as N goes to $\infty$, in the preceding inequality. The remainder of the convergent series will disappear and we will get a useful analog of . The first task is to clarify how the left hand side converges.
Set $\xi= r\,\xi_0$, with $|\xi_0| = 1$ and $ r>0$. Rewrite with $S$ replaced by $S_N$ and $a_{2p}$ by $a_{2p}^N$: $$\widehat{S_N \chi_B}(r\xi_0)= \sum_{p=1}^{\infty} a_{2p}^N
(\xi_0)\,r^{2p} \,.$$ It is a remarkable key fact that for a fixed $p$ the sequence of the $a_{2p}^N$ stabilizes for $N$ large. This fact depends on a laborious computation of various constants and will be proved in section 7 in the following form.
\[L11\] If $p+1 \leq N $, then $a_{2p}^N = a_{2p}^{p+1}$.
If $p\geq1$ and $p+1 \leq N $ we set $a_{2p}= a_{2p}^N$. We need an estimate for the $a_{2p}^N$, which will be proved as well in section 7.
\[L12\] We have, for a constant $C$ depending only on $n$, $$\label{eq59bis}
|a_{2p}| \leq \frac{C}{(p-1)!\,4^p} \,\sum_{j=1}^p
\|P_{2j}\|_\infty, \quad 1 \le p \le N-1\,,$$ and $$\label{eq59tris}
|a_{2p}^N | \le \frac{C}{4^p} \, \binom{\frac{n}{2}+N-1}
{N-1}\,\sum_{j=1}^{N-1} \|P_{2j}\|_\infty , \quad 1 < N \le p\,.$$
Let us prove that for each $\xi_0$ in the sphere the sequence $S_N\,\chi_B (r\,\xi_0)$ converges uniformly on $0 \le r \le 1$. For $1 \le N \le M$ $$\begin{split}
\left| \widehat{S_{N}\, \chi_{B}}(r \xi_{0}) - \widehat{S_{M}\, \chi_{B}}(r\xi_{0}) \right|
& \le \sum_{p \ge N} | a_{2p}^{N}| r^{2p} + \sum_{p= N}^{M-1} | a_{2p}| r^{2p} + \sum_{p\ge M} | a_{2p}^{M}| r^{2p} \\*[5pt]
& \le C \left( \frac{1}{4^N} \, \binom{\frac{n}{2}+N-1}
{N-1} + \sum_{p\ge N} \frac{1}{(p-1)!\,4^p}
\right)\sum_{j=1}^{\infty} \|P_{2j}\|_\infty \,,
\end{split}$$ which clearly tends to $0$ as $N$ goes to $\infty$. Letting $N$ go to $\infty$ in we get $$\label{eq60}
\left |\sum_{p=1}^{\infty} a_{2p} (\xi_0)\,r^{2p}\right| \le C\,
|\widehat{P.V. K} (\xi_0)|, \quad 0 \leq r \leq 1, \quad |\xi_0| =
1\,.$$ At this point we may repeat almost verbatim the proof we presented in the previous section, because the coefficients $ a_{2p}^{N}$ stabilize. This allows us to argue as in the polynomial case. The only difference lays in the fact that now we are dealing with infinite sums. However, no convergence problems will really arise because of .
Proof of the combinatorial Lemmata
==================================
This section will be devoted to prove Lemmas 10, 11 and 12 stated and used in the preceding sections.
For the proof of Lemma 10 (see section 5) we need to carefully trace back the path that led us to the constants $C_{2j}$. To begin with we need a formula for the coefficients $A_l$ in and for that it is essential to have the expression for the fundamental solution $E_N= E_N^n$ of $\triangle^N$ in ${{\mathbb R}^n}$. One has [@ACL] $$E_N(x)=\frac{1}{|x|^{n-2N}}(\alpha(n,N)+\beta(n,N)\log|x|^2)\,,$$ where $\alpha$ and $\beta$ are constants that depend on $n$ and $N$. To write in close form $\alpha$ and $\beta$ we consider different cases. Let $\omega_n$ be the surface measure of $S^{n-1}$.
[*Case 1:*]{} $n$ is odd. Then $$\alpha(n,N)=
\frac{\Gamma(2-\frac{n}{2})}{4^{N-1}\,(N-1)!\,\Gamma(N+1-\frac{n}{2})\,(2-n)\,\omega_n}$$ and $$\beta(n, N)=0\,.$$
[*Case 2:*]{} $n$ is even, $n \ne 2$ and $N\le
\frac{n}{2}-1$. Then $$\alpha(n,N)=\frac{(-1)^{N-1}\,(\frac{n}{2}-N-1)!}{4^{N-1}(N-1)!\,(\frac{n}{2}-2)!\,(2-n)\,\omega_n}$$ and $$\beta(n, N)=0\,.$$
[*Case 3:*]{} $n$ is even, $n \ne 2$ and $N\ge
\frac{n}{2}$. Then $$\beta(n,N)=\frac{1}{(-1)^{\frac{n}{2}+1}(N-\frac{n}{2})!\,4^{N-1}\,(N-1)!\,(\frac{n}{2}-2)!\,(2-n)\,\omega_n}$$ and $$\alpha(n,N)=2\,\beta(n,N)\,S_{N-\frac{n}{2}}\,,$$ where $S_0=0$ and $$S_L= \sum_{k=1}^{L} \frac{1}{2k}+ \sum_{k={\frac{n}{2}}}^{L+{\frac{n}{2}}-1}\frac{1}{2k},
\quad 1 \le L\,.$$
[*Case 4:*]{} $n=2.$ $$\beta(2,N)= \frac{1}{2}\,\frac{1}{ 4^{N-1}(N-1)!^2\,\omega_2}$$ and $$\alpha(2,N)=2\,\beta(2,N)\,S_{N-1}\,.$$
Recall that the constants $A_0, A_1,\dotsc, A_{2N-1}$ are chosen so that the function (see ) $$\varphi(x)= E(x)\,\chi_{{{{\mathbb R}^n}\setminus \overline B}}(x) + (A_0+A_1\,|x|^2 +\dotsb+
A_{2N-1}\,|x|^{4N-2})\,\chi_B(x)\,,$$ and all its partial derivatives of order not greater than $2N-1$ extend continuously up to $\partial B$.
\[L13\] For $L=N+1, \dotsc, 2N-1$ we have $$A_{L}= \frac{(-1)^{N+L}\displaystyle\binom{N+{\frac{n}{2}}-1}{N-1}}{V_n\, 4^N
\, (L+{\frac{n}{2}}-N)\,(2N-L-1)!\,L!}\,,$$ where $V_n$ is the volume of the unit ball.
Set $t=|x|^2$, so that $$\label{eq69}
E_N^n(x)\equiv E(t)=t^{N-{\frac{n}{2}}}(\alpha+\beta\log(t))\,.$$
Let $P(t)$ be the polynomial $\sum_{L=0}^{2N-1}A_Lt^L.$ By Corollary 2 in section 2 we need that $$P^{k)}(1)=E^{k)}(1) , \quad 0 \le k \le 2N-1\,.$$
By Taylor’s expansion we have that $P(t)=\sum_{i=0}^{2N-1}\frac{E^{i)}(1)}{i!}(t-1)^{i},$ and hence, by the binomial formula applied to $(t-1)^i$, $$A_L=\sum_{i=L}^{2N-1}\frac{E^{i)}(1)}{i!}(-1)^{i-L}\binom{i}{L},
\quad 0 \le L \le 2N-1\,.$$ Now we want to compute $E^{i)}(1)$. Clearly $$\left( \frac{d}{dt}\right)^i (t^{N-{\frac{n}{2}}})=\left(N-{\frac{n}{2}}\right) \cdots
\left(N-{\frac{n}{2}}-i+1\right)t^{N-{\frac{n}{2}}-i}.$$ Notice that the logarithmic term in only appears when the dimension $n$ is even. In this case, for each $i\geq N+1$ $$\left( \frac{d}{dt}\right)^i(t^{N-{\frac{n}{2}}}\log t )=
\left(N-{\frac{n}{2}}\right)!(-1)^{i-N+{\frac{n}{2}}-1}\left(i-N+{\frac{n}{2}}-1\right)!\;t^{-i+N-{\frac{n}{2}}}\,.$$ Hence, for $i\geq N+1$, we obtain $$\begin{gathered}
E^{i)}(1)=\alpha(n,N)\left(N-{\frac{n}{2}}\right)\cdots\left(n-{\frac{n}{2}}-i+1\right)\\*[5pt]
+\beta(n,N)\left(N-{\frac{n}{2}}\right)!(-1)^{i-N+{\frac{n}{2}}-1}\left(i-N+{\frac{n}{2}}-1\right)!\end{gathered}$$ Consequently, $$\label{eq70}
\begin{split}
A_L & =(-1)^L\alpha(n,N)\sum_{i=L}^{2N-1}\left(N-{\frac{n}{2}}\right)\cdots
\left(N-{\frac{n}{2}}-i+1\right)\frac{(-1)^i}{i!}\binom{i}{L} \\*[5pt] &\quad+
(-1)^{L-N+{\frac{n}{2}}-1}\beta(n,N)\left(N-{\frac{n}{2}}\right)!
\sum_{i=L}^{2N-1}\left(i-N+{\frac{n}{2}}-1\right)!\frac{\binom{i}{L}}{i!}\,.
\end{split}$$ Let’s remark that for the cases $n=2$ or $n$ even and $N\geq{\frac{n}{2}}$ the first term in is zero, while for the cases $n$ odd or $n$ even and $N\leq {\frac{n}{2}}-1$ the second term is zero because $\beta(n,N) = 0.$ This explains why we compute below the two terms separately.
For the first term we show that $$\label{eq71}
\sum_{i=L}^{2N-1}\!\left(N\!-\!{\frac{n}{2}}\right)\cdots
\left(N-{\frac{n}{2}}-i+1\right)\frac{(-1)^i}{i!} \binom{i} {L}\! =\! (-1)^L
\!\binom{N-{\frac{n}{2}}}{L}\!\binom{{\frac{n}{2}}+N-1}{2N-1-L}\,.\!\!$$ Indeed, the left hand side of is, setting $k=i-L$, $$\begin{split}
\frac{1}{L!}\sum_{k=0}^{2N-1-L}&\left(N-{\frac{n}{2}}\right)\cdots\left(N-{\frac{n}{2}}-L-k+1\right)\frac{(-1)^{L+k}}{k!}\\*[7pt]
& = (-1)^L\binom{N-{\frac{n}{2}}}{L}
\sum_{k=0}^{2N-1-L}\binom{{\frac{n}{2}}+L-N+k-1}{k}\\*[7pt] & = (-1)^L
\binom{N-{\frac{n}{2}}}{L}\binom{N+{\frac{n}{2}}-1}{2N-1-L}\,,
\end{split}$$ where the last identity comes from ([@GKP (5.9), p. 159]).
To compute the second term we first show that $$\label{eq72}
\sum_{i=L}^{2N-1}\left(i-N+{\frac{n}{2}}-1\right)!\frac{1}{i!}\binom{i}{L}=
\frac{(L-N+{\frac{n}{2}}-1)!}{L!}\binom{N+{\frac{n}{2}}-1}{2N-1-L}\,.$$ As before, setting $k=i-L$ and applying [@GKP (5.9), p. 159], we see that the left hand side of is $$\begin{split}
\frac{1}{L!}\sum_{k=0}^{2N-1-L}&\left(L+k-N+{\frac{n}{2}}-1\right)!\frac{1}{k!} \\*[5pt]
&= \left(L-N+{\frac{n}{2}}-1\right)! \sum_{k=0}^{2N-1-L}\binom{{\frac{n}{2}}+L-N+k-1}{k}
\\*[5pt] &= \frac{(L-N+{\frac{n}{2}}-1)!}{L!}\binom{N+{\frac{n}{2}}-1}{2N-1-L}\,.
\end{split}$$
We are now ready to complete the proof of the lemma distinguishing $4$ cases.
[*Case 1:*]{} $n$ odd.
Since $\beta(n,N)=0$, replacing in $\alpha(n,N)$ by its value and using we get, by elementary arithmetics and the identity $n\,V_n = \omega_n$, $$\begin{split}
A_L&=
\frac{(-1)^L
\Gamma(2-{\frac{n}{2}})}{4^{N-1}(N-1)!\Gamma(N+1-{\frac{n}{2}})(2-n)\omega_n}(-1)^L
\binom{N-{\frac{n}{2}}}{L} \binom{{\frac{n}{2}}+N-1}{2N-1-L} \\*[7pt] &=
\frac{(-1)^{N+L}(N+{\frac{n}{2}}-1)\cdots ({\frac{n}{2}}+1)}{4^N \,V_n\,
L!\,(2N-1-L)!\,(L+{\frac{n}{2}}-N)(N-1)!} \\*[7pt] &=
\frac{(-1)^{N+L}\binom{N+{\frac{n}{2}}-1}{N-1}}{4^N \,V_n\,
L!\,(2N-1-L)!\,(L+{\frac{n}{2}}-N)}\,.
\end{split}$$
[*Case 2:*]{} $n$ even, $n\neq 2$ and $N\le {\frac{n}{2}}-1$.
As in case 1 $\beta(n,N)=0$, and we proceed similarly using to obtain $$\begin{split}
A_L&=
\frac{(-1)^L(-1)^{N-1}({\frac{n}{2}}-N-1)!}{4^N(N-1)!({\frac{n}{2}}-2)!(2-n)\omega_n}(-1)^L
\binom{N-{\frac{n}{2}}}{L}\binom{N+{\frac{n}{2}}-1}{2N-1-L}\\*[7pt] &=
\frac{(-1)^{N+1}\binom{N+{\frac{n}{2}}-1}{N-1}{\frac{n}{2}}!({\frac{n}{2}}-N-1)!\{(N-{\frac{n}{2}})\cdots(N-{\frac{n}{2}}-L+1)\}}{(2N-1-L)!\,
L!\,4^{N-1}\,(2-n)\,\omega_n \,({\frac{n}{2}}-2)!\,({\frac{n}{2}}-N+L)!}\\*[7pt] &=
\frac{(-1)^N\binom{N+{\frac{n}{2}}-1}{N-1}}{4^N \,V_n\,
L!\,(2N-1-L)!}\,\frac{(-1)L({\frac{n}{2}}-N+L-1)!}{({\frac{n}{2}}-N+L)!}\\*[7pt] &=
\frac{(-1)^{N+L}\binom{N+{\frac{n}{2}}-1}{N-1}}{4^N \,V_n\,
L!\,(2N-1-L)!\,(L+{\frac{n}{2}}-N)}\,.
\end{split}$$
[*Case 3:*]{} $n$ even, $n\neq 2$ and $N\ge {\frac{n}{2}}.$
Replacing in $\alpha(n,N)$ and $\beta(n,N)$ by their values and using we get, by elementary arithmetics and the identity $n\,V_n = \omega_n$, $$\begin{split}
A_L&=\beta(n,N)(N-{\frac{n}{2}})!(-1)^{N+{\frac{n}{2}}-1+L}\sum_{i=L}^{2N-1}(i-N+{\frac{n}{2}}-1)!\frac{1}{i!}
\binom{i}{L} \\*[7pt]
&=\frac{(-1)^{N+L}(L-N+{\frac{n}{2}}-1)!}{L!4^{N-1}({\frac{n}{2}}-2)!(N-1)!(2-n)\omega_n}\binom{N+{\frac{n}{2}}-1}{2N-1-L}\\*[7pt]
&= \frac{(-1)^{L+N}{\frac{n}{2}}({\frac{n}{2}}-1)}{4^{N-1}(2-n)\omega_n
L!(2N-1-L)!({\frac{n}{2}}-N+L)}\binom{N+{\frac{n}{2}}-1} {N-1} \\*[7pt] &=
\frac{(-1)^{L+N}\binom{N+{\frac{n}{2}}-1}{N-1}}{4^NL!V_n(2N-1-L)!({\frac{n}{2}}-N+L)}\,.
\end{split}$$
[*Case 4:*]{} $n=2.$
Proceeding as in case 3 and we obtain $$\begin{split}
A_L&=\beta(n,N)(N-{\frac{n}{2}})!(-1)^{N+{\frac{n}{2}}-1+L}\sum_{i=L}^{2N-1}(i-N+{\frac{n}{2}}-1)!\frac{1}{i!}
\binom{i}{L}\\*[7pt] &=
\frac{(-1)^{N+L}N!}{2\omega_24^{N-1}L!(N-1)!(2N-1-L)!(L+1-N)}\\*[7pt]
&= \frac{(-1)^{L+N}\binom{N}{N-1}}{V_24^NL!(2n-1-L)!(L+1-N)}\,.
\end{split}$$
Recall that (see ) $$C_{2j} = \sum_{l=N+j}^{2N-1} c_{\,l,\,j,\,l-N-j}\,G_{{\frac{n}{2}}+l-N+j}(0)\,.$$ Thus, we have to compute the constants $c_{l,j,k}$ appearing in the expression for $\widehat{S\,\chi_{B}}(\xi)$. For that we need the constants $c_{l,j}$ appearing in the formula for $S(x)$. We start by computing $P_{2j}(\partial)\Delta^{N-j} (|x|^{2l})$. Using and Lemma 4 one gets $$\begin{split}
P_{2j}(\partial)\Delta^{N-j} (|x|^{2l}) = \frac{4^N l!
(N-j)!}{(l-N-j)!}\binom{l-1+{\frac{n}{2}}}{N-j} P_{2j}(x) |x|^{2(l-N-j)}
\end{split}$$ if $l-N-j\ge 0$ (and $=0$ if $l-N-j < 0$).
As in (Section 3), we want to compute $P_{2j}(\partial)\Delta^{l-N-j}G_{{\frac{n}{2}}}(\xi)$ by using Lemma \[L2\] applied to $f(r)=G_{\frac{n}{2}}(r)$ and the homogeneous polynomial $L(x)=
P_{2j}(x)\,|x|^{2(l-N-j)}$. We obtain $$\begin{split}
P_{2j}(\partial)\Delta^{l-N-j}G_{{\frac{n}{2}}}(\xi) & =
(-1)^{2(l-N)}\sum_{k\ge 0}\frac{(-1)^k}{2^k k!}\Delta^{k} (
P_{2j}(x) | x|^{2(l-N-j} )\, G_{{\frac{n}{2}}+2(l-N)-k} (\xi)\\*[7pt]
& = \sum_{k=0}^{l-N-j} \frac{(-1)^k}{2^k k!} 4^{k}\frac{ (l-N-j) !}{(l-N-j-k)!} k!
\binom{{\frac{n}{2}}+j+l-N-1}{k} \\*[7pt]
& \quad \times P_{2j}(\xi) | \xi|^{2(l-N-j-k)} \, G_{{\frac{n}{2}}+2(l-N)-k}(\xi) \,.
\end{split}$$
In view of the definitions of $Q(x)$ and $S(x)$, $$\begin{split}
S(x) & = -Q(\partial)\left ( \sum_{l=0}^{2N-1}A_{l}|x|^{2l} \right
) = - \sum_{l=0}^{2N-1}A_{l} \sum_{j=1}^N \gamma_{2j} P_{2j }
(\partial) \Delta^{N-j}(|x|^{2l}) \\*[7pt] & = - \sum_{l=N+1}^{2N-1}
\sum_{j=1}^{l-N} A_{l} \gamma_{2j} \frac{4^N l!
(N-j)!}{(l-N-j)!}\binom{l-1+{\frac{n}{2}}}{N-j} P_{2j}(x)
|x|^{2(l-N-j)}\\*[7pt] & = \sum_{l=N+1}^{2N-1} \sum_{j=1}^{l-N}
c_{l,j}P_{2j}(x) |x|^{2(l-N-j)}\,,
\end{split}$$ where the last identity defines the $c_{l,j}$. In we set $$\begin{split}
\widehat{S\,\chi_{B}}(\xi) & = S(\imath\, \partial)\,
\widehat{\chi_B}(\xi)\\*[7pt] & = \sum_{l=N+1}^{2N-1}
\sum_{j=1}^{l-N}
c_{lj}\,(-1)^{l-N}\,\,P_{2j}(\partial)\,\triangle^{l-N-j}\,G_{{\frac{n}{2}}}(\xi)\\*[7pt]
& = \sum_{l=N+1}^{2N-1} \sum_{j=1}^{l-N} \sum_{k=0}^{l-N-j}
c_{\,l,\,j,\,k}\,\,P_{2j}(\xi)\,|\xi|^{2(l-N-j-k)}
G_{{\frac{n}{2}}+2l-2N-k}(\xi)\,.
\end{split}$$ Consequently, $$\begin{split}
c_{l,j,k} & =c_{l,j} (-1)^{l-N} \frac{(-1)^k}{2^k} 4^{k}\frac{
(l-N-j) !}{(l-N-j-k)!} \binom{{\frac{n}{2}}+j+l-N-1}{k} \\*[7pt] & =
-(-1)^{l+k+N}A_{l} \gamma_{2j} \frac{4^N l!
(N-j)!}{(l-N-j)!}\binom{l-1+{\frac{n}{2}}}{N-j}\\*[7pt] &\quad\times
2^{k}\frac{ (l-N-j) !}{(l-N-j-k)!} \binom{{\frac{n}{2}}+j+l-N-1}{k}\,.
\end{split}$$ Replacing $A_l$ by the formula given in lemma \[L13\] and performing some easy arithmetics we get $$\label{eq74}
c_{l,j,k}= - \frac{(-1)^{k}}{V_{n}}\frac{\gamma_{2j}\!\displaystyle
\binom{N+{\frac{n}{2}}-1}{N-1} 2^k (N-j)! \binom{l-1+{\frac{n}{2}}}{N-j}\!
\binom{{\frac{n}{2}}+j+l-N-1} {k} } {(l+{\frac{n}{2}}-N)(2N-l-1)! (l-N-j-k)! } \, .\!$$
The final computation of the $C_{2j}$ is as follows. [$$\begin{aligned}
C_{2j} & = \sum_{l=N+j}^{2N-1} c_{\,l,\,j,\,l-N-j}\,G_{{\frac{n}{2}}+l-N+j}(0)\\[9pt]
& = \hspace{1.5cm} \text{[by the explicit value of $G_p(0)$ given in
\eqref{eq75} below]}\\[9pt]
& = \sum_{l=N+j}^{2N-1} c_{\,l,\,j,\,l-N-j}
\frac{1}{2^{{\frac{n}{2}}+l-N+j}\Gamma
({\frac{n}{2}}+l-N+j+1)}\\[9pt]
& = \hspace{1.5cm} \text{[by \eqref{eq74}]}\\[9pt]
& =- \sum_{l=N+j}^{2N-1}\! \frac{(-1)^{l-N-j}}{V_{n}}
\frac{\gamma_{2j}\!\displaystyle\binom{N\!+\!{\frac{n}{2}}\!-\!1}{N\!-\!1} 2^{l-N-j} (N\!-\!j)! \binom{l\!-\!1\!+\!{\frac{n}{2}}}{N-j} \binom{{\frac{n}{2}}\!+\!j\!+\!l\!-\!N\!-\!1}{l-N-j} } {(l+{\frac{n}{2}}-N)(2N-l-1)! 2^{{\frac{n}{2}}+l-N+j}\Gamma({\frac{n}{2}}+l-N+j+1)} \\[9pt]
&= \hspace{1.5cm} \text{[substituting the value given in
\eqref{eq7} in
$\gamma_{2j}$]}\\[9pt]
& =- \frac{ \pi^{\frac{n}{2}}\Gamma(j) \displaystyle\binom{N+{\frac{n}{2}}-1}{N-1} (N-j)! }{ V_{n} \Gamma(j+{\frac{n}{2}}) 2^{{\frac{n}{2}}+2j} }\\[9pt]
&\hspace{1.5cm}\times
\sum_{l=N+1}^{2N-1}\frac{(-1)^{l+N}\displaystyle \binom{l-1+{\frac{n}{2}}}{N-j}
\binom{{\frac{n}{2}}+j+l-N-1}{l-N-j} } {(l+{\frac{n}{2}}-N)(2N-l-1)! \Gamma({\frac{n}{2}}+l-N+j+1)}\\[9pt]
&= \hspace{1.5cm} [\text{setting } l=i+N+j] \\[9pt]
& =- \frac{ \pi^{\frac{n}{2}}\Gamma(j) \displaystyle\binom{N+{\frac{n}{2}}-1}{N-1} (N-j)! }{ V_{n} \Gamma(j+{\frac{n}{2}}) 2^{{\frac{n}{2}}+2j} } \\[9pt]
& \hspace{1.5cm}\times \sum_{i=0}^{N-1-j}\frac{(-1)^{i+j}\displaystyle \binom{i+N+j-1+{\frac{n}{2}}}{N-j} \binom{{\frac{n}{2}}+2j+i-1}{i} } {(i+j+{\frac{n}{2}})(N-i-j-1)! \Gamma({\frac{n}{2}}+i+2j+1)}\\[9pt]
& = \hspace{1.5cm} [\text{because } \Gamma({\frac{n}{2}}+i+2j+1)= \Gamma(2j +{\frac{n}{2}}) \prod_{k=0}^{i} ({\frac{n}{2}}+2j+i-k) ]\\[9pt]
& =- \frac{ (-1)^{j}\pi^{\frac{n}{2}}\Gamma(j) \displaystyle\binom{N+{\frac{n}{2}}-1}{N-1} (N-j)! }{ V_{n} \Gamma(j+{\frac{n}{2}}) 2^{{\frac{n}{2}}+2j} \Gamma(2j +{\frac{n}{2}}) }\\[9pt]
& \hspace{1.5cm} \times \sum_{i=0}^{N-1-j}\frac{(-1)^{i}
\displaystyle \binom{i+N+j-1+{\frac{n}{2}}}{N-j} \binom{{\frac{n}{2}}+2j+i-1}{i} }
{(i+j+{\frac{n}{2}})(N-i-j-1)! \prod_{k=0}^{i} ({\frac{n}{2}}+2j+i-k)}\\[9pt]
& = \hspace{1.5cm} [\text{using Lemma \ref{sublema2} below}]\\[9pt]
&=- \frac{ (-1)^{j}\pi^{\frac{n}{2}}\Gamma(j) \displaystyle\binom{N+{\frac{n}{2}}-1}{N-1} (N-j)! }{ V_{n} \Gamma(j+{\frac{n}{2}}) 2^{{\frac{n}{2}}+2j} \Gamma(2j +{\frac{n}{2}}) } \binom{N-1}{j-1}\frac{\Gamma (j+{\frac{n}{2}})}{j\, \Gamma (N+{\frac{n}{2}})}\\[9pt]
& = \frac{- \pi^{\frac{n}{2}}}{V_{n} 2^{\frac{n}{2}}\Gamma({\frac{n}{2}}+1)} \; \frac{(-1)^j }{j\,
4^{j} \Gamma (2j+ {\frac{n}{2}})} \,.
$$]{}
\[sublema2\] For each $j=1, \dotsc , N-1$ $$\sum_{i=0}^{N-1-j}\frac{(-1)^{i}\displaystyle
\binom{i+N+j-1+{\frac{n}{2}}}{N-j} \binom{{\frac{n}{2}}+2j+i-1} {i} } {(i+j+{\frac{n}{2}})(N-i-j-1)!
\prod_{k=0}^{i} ({\frac{n}{2}}+2j+i-k)} = \binom{N-1}{j-1}\frac{\Gamma
(j+{\frac{n}{2}})}{j \Gamma (N+{\frac{n}{2}})}\,.$$
Divide the left hand side by the right hand side and denote the quotient by $A$. We have to prove that $A=1$. Using elementary arithmetics $$\begin{gathered}
\frac{\displaystyle \binom{i+N+j-1+{\frac{n}{2}}}{N-j}
\binom{{\frac{n}{2}}+2j+i-1}{i}\Gamma(N+{\frac{n}{2}}) }{(i+j+{\frac{n}{2}}) \prod_{k=0}^{i}
({\frac{n}{2}}+2j+i-k)\; \Gamma (j+{\frac{n}{2}})} \\*[9pt] =
\binom{N+i+j+{\frac{n}{2}}-1}{N-j-1}\binom{N+{\frac{n}{2}}-1}{N-i-j-1}\frac{(N-i-j-1)!}{N-j}
\binom{{\frac{n}{2}}+i+j-1} {i}\,,\end{gathered}$$ and so $$\begin{split}
A &= \frac{j}{\binom{N-1}{j-1}(N-j)}\! \sum_{i=0}^{N-1-j}\!
(-1)^{i}\!
\binom{N+i+j+{\frac{n}{2}}-1}{N-j-1}\!\binom{N+{\frac{n}{2}}-1}{N-i-j-1}\!\binom{{\frac{n}{2}}+i+j-1}{i}\\*[7pt]
& = \hspace{1.5cm} \left[\text{because } \binom{a+i}{i}=
(-1)^{i}\binom{-a-1}{i} \right]\\*[7pt] &=
\frac{j}{\binom{N-1}{j-1}(N-j)} \sum_{i=0}^{N-1-j}
\binom{N+i+j+{\frac{n}{2}}-1}{N-j-1}\binom{N+{\frac{n}{2}}-1}{N-i-j-1}\binom{-{\frac{n}{2}}-j}{i}\\*[7pt]
& = [\text{by the triple-binomial identity (5.28) of
(\cite[p.~171]{GKP}), see below}]\\*[7pt] & =
\frac{j}{\binom{N-1}{j-1}(N-j)}
\binom{{\frac{n}{2}}+N+j-1}{0}\binom{N-1}{N-j-1} =1 \,.
\end{split}$$ For the reader’s convenience and later reference we state the triple-binomial identity [@GKP (5.28), p. 171]: $$\label{eq74bis}
\sum _{k=0}^{n} \binom{m-r+s}{k}\binom{n+r-s}{n-k}\binom{r+k}{m+n} =
\binom{r}{m}\binom{s}{n}\,,$$ where $m$ and $n$ are non-negative integers.
Our next task is to prove Lemma \[L11\] and Lemma \[L12\]. Setting $\xi= r\,\xi_0$, $|\xi_{0}|=1$, in we obtain $$\begin{split}
\widehat{S_{N}\,\chi_{B}}(r\xi_{0}) & = \sum_{l=N+1}^{2N-1}
\sum_{j=1}^{l-N} \sum_{k=0}^{l-N-j} c_{\,
l,\,j,\,k}\,\,P_{2j}(r\xi_{0})\,\,|r\xi_{0}|^{2(l-N-j-k)}
G_{{\frac{n}{2}}+2l-2N-k}(r\xi_{0})\\*[5pt] & \hspace{1.5 cm} [\text{make the
change of indexes $l=N+s$}]\\*[5pt] & = \sum_{s=1}^{N-1}
\sum_{j=1}^{s} \sum_{k=0}^{s-j}
c_{\,N+s,\,j,\,k}\,P_{2j}(\xi_{0})\,r^{2(s-k)}
G_{{\frac{n}{2}}+2s-k}(r)\\*[5pt] & = \sum_{j=1}^{N-1} \sum_{s=j}^{N-1}
\sum_{k=0}^{s-j} c_{\,N+s,\,j,\,k}\,P_{2j}(\xi_{0})\,r^{2(s-k)}
G_{{\frac{n}{2}}+2s-k}(r)\\*[5pt] & := \sum_{p=1}^{\infty} a_{2p}^{N}(\xi_{0})
r^{2p}\,.
\end{split}$$
In order to compute the coefficients $ a_{2p}^{N}(\xi_{0})$ we substitute the power series expansion of $G_{q}(r)$ [@Gr A-8], namely, $$\label{eq75}
G_{q}(r)=\sum _{i=0}^{\infty}\frac{(-1)^{i}}{i!\, \Gamma(q+i+1)}
\frac{r^{2i}}{2^{2i+q}}\,,$$ in the last triple sum above.
We are assuming that $1 \le p \le N-1$. It is important to remark that, for this range of $p$, after introducing in the triple sum above, only the values of the index $j$ satisfying $1
\le j \le p$ are involved in the expression for $a_{2p}^N$. Once has been introduced in the triple sum one should sum, in principle, on the four indexes $i$, $j$, $s$ and $k$. But since we are looking at the coefficient of $r^{2p}$ we have the relation $2(s-k)+2i=2p$, which actually leaves us with three indexes. The range of each of these indexes is easy to determine and one gets $$a_{2p}^N = \sum_{j=1}^{p}P_{2j}(\xi_{0}) \sum_{i=0}^{p-j}
\sum_{s=p-i}^{N-1} c_{\,N+s,\,j,\,s-(p-i)}
\times \text{coefficient of $r^{2i}$ from $G_{{\frac{n}{2}}+s+p-i}(r)$}\,.$$ In view of $$\begin{split}
a_{2p}^N &= \sum_{j=1}^{p}P_{2j}(\xi_{0}) \sum_{i=0}^{p-j}
\sum_{s=p-i}^{N-1} c_{\,N+s,\,j,\,s-(p-i)} \frac{(-1)^{i}}{i!
2^{i+{\frac{n}{2}}+s+p}\Gamma({\frac{n}{2}}+s+p+1)}\\*[7pt] & = \hspace{1.5cm}[\text{by
the expression \eqref{eq74} for $c_{l,j,k}$}]\\*[7pt] & =
\sum_{j=1}^{p}P_{2j}(\xi_{0}) \sum_{i=0}^{p-j} \sum_{s=p-i}^{N-1}
\frac{(-1)^{i+1}}{i! 2^{i+{\frac{n}{2}}+s+p}\Gamma({\frac{n}{2}}+s+p+1)}
\frac{(-1)^{s-(p-i)}}{V_{n}} \\*[7pt] & \hspace{1.5cm}\times
\frac{\gamma_{2j}\displaystyle \binom{N+{\frac{n}{2}}-1}{N-1} 2^{s-(p-i)}
(N-j)! \binom{N+s-1+{\frac{n}{2}}}{N-j} \binom{{\frac{n}{2}}+j+s-1}{s-(p-i)} }
{(s+{\frac{n}{2}})(N-s-1)! (p-i-j)! }\\*[7pt] & =
(-1)^{p+1}\sum_{j=1}^{p}P_{2j}(\xi_{0}) \frac{(-1)^{j} \pi^{\frac{n}{2}}\Gamma(j) (N-j)!}{V_{n} \Gamma({\frac{n}{2}}+j)2^{{\frac{n}{2}}+2p}}\sum_{i=0}^{p-j}
\frac{\displaystyle \binom{N+{\frac{n}{2}}-1}{N-1} }{i! (p-i-j)! }\\*[7pt] &
\hspace{1.5cm}\times \sum_{s=p-i}^{N-1} \frac{(-1)^s\displaystyle
\binom{N+s-1+{\frac{n}{2}}}{N-j} \binom{{\frac{n}{2}}+j+s-1}{ s-(p-i)} } {(s+{\frac{n}{2}})(N-s-1)!
\Gamma({\frac{n}{2}}+s+p+1)}\,.
\end{split}$$
In Lemma \[L15\] below we give a useful compact form for the last sum. Using it we obtain $$\begin{gathered}
a_{2p}^N = (-1)^{p+1}\sum_{j=1}^{p}P_{2j}(\xi_{0}) \frac{(-1)^{j}
\pi^{\frac{n}{2}}\Gamma(j) (N-j)!}{V_{n} \Gamma({\frac{n}{2}}+j)2^{{\frac{n}{2}}+2p}}
\sum_{i=0}^{p-j} \frac{\displaystyle \binom{N+{\frac{n}{2}}-1}{N-1} }{i!
(p-i-j)! }\\*[7pt] \times \frac{ (-1)^{p-i}(N-p-1)!
\displaystyle\binom{N-1}{p}}{(N-j)! \Gamma(N+{\frac{n}{2}})
j!\binom{{\frac{n}{2}}+p-i+j-1}{j}}\,.\end{gathered}$$ Easy arithmetics with binomial coefficients gives $$\binom{N+{\frac{n}{2}}-1}{N-1}(N-p-1)!\binom{N-1}{p}\frac{1}{\Gamma(N+{\frac{n}{2}})}=
\frac{1}{p! \, \Gamma({\frac{n}{2}}+1)}\,.$$ We finally get the extremely surprising identity $$\label{eq78}
a_{2p}^N \!=\! \frac{-\pi^{\frac{n}{2}}}{V_{n}
2^{{\frac{n}{2}}+2p}p!\Gamma({\frac{n}{2}}+1)} \!\sum_{j=1}^{p}\! \frac{(-1)^j \Gamma(j)
P_{2j}(\xi_{0})} {\Gamma({\frac{n}{2}}+j) } \sum_{i=0}^{p-j} \frac{(-1)^{i}}{
i! (p-i-j)! j! \binom{{\frac{n}{2}}+p-i+j-1}{j}}\,,$$ in which $N$ has miraculously disappeared. Thus Lemma \[L11\] is proved.
We start by proving the inequality , so that $1\le p
\le N-1$. We roughly estimate $a_{2p}= a_{2p}^N$ by putting the absolute value inside the sums in . The absolute value of each term in the innermost sum in is obviously not greater than $1$ and there are at most $p$ terms. The factor in front of $P_{2j}(\xi_0)$ is again not greater than $1$ in absolute value. Denoting by $C$ the terms that depend only on $n$ we obtain the desired inequality .
We turn now to the proof of inequality . Recall that $$\widehat{S_{N}\,\chi_{B}}(r\xi_{0}) = \sum_{p=1}^{\infty}
a_{2p}^{N}(\xi_{0}) r^{2p}
= \sum_{j=1}^{N-1} \sum_{s=j}^{N-1} \sum_{k=0}^{s-j}
c_{\,N+s,\,j,\,k}\,\,P_{2j}(\xi_{0})\,\,r^{2(s-k)} G_{{\frac{n}{2}}+2s-k}(r)\,.$$ Replacing $G_{{\frac{n}{2}}+2s-k}(r)$ by the expression given by we obtain, as before, a sum with four indexes. Now we eliminate the index $i$ of using $s-k+i=p$. Hence $$\begin{split}
a_{2p}^N & = \sum_{j=1}^{N-1}P_{2j}(\xi_{0}) \sum_{s=j}^{N-1}
\sum_{k=0}^{s-j} c_{\,N+s,\,j,\, k}
\times \text{coefficient of $r^{2(p-s+k)}$ from $G_{{\frac{n}{2}}+2s-k}(r)$}\\*[7pt]
&=\sum_{j=1}^{N-1}P_{2j}(\xi_{0}) \sum_{s=j}^{N-1} \sum_{k=0}^{s-j}
c_{\,N+s,\,j,\, k}
\frac{(-1)^{p-s+k}}{(p-s+k)!\Gamma({\frac{n}{2}}+p+s-1)2^{2p+{\frac{n}{2}}+k}}\\*[7pt]
&=\sum_{j=1}^{N-1}P_{2j}(\xi_{0}) \sum_{k=0}^{N-1-j}
\sum_{s=j+k}^{N-1} c_{\,N+s,\,j,\, k}
\frac{(-1)^{p-s+k}}{(p-s+k)!\Gamma({\frac{n}{2}}+p+s-1)2^{2p+{\frac{n}{2}}+k}}\\*[7pt] &=
\frac{(-1)^{p}}{V_{n}4^{p}2^{\frac{n}{2}}}\binom{N+{\frac{n}{2}}-1}{N-1}\sum_{j=1}^{N-1}
\gamma_{2j}P_{2j}(\xi_{0}) \sum_{k=0}^{N-1-j} \sum_{s=j+k}^{N-1}
\\*[7pt] &\quad\times \frac{(-1)^s (N-j)!
\displaystyle\binom{N+s-1+{\frac{n}{2}}}{N-j}\binom{{\frac{n}{2}}+j+s-1}{k}}{(p-s+k)!
\Gamma({\frac{n}{2}}+p+s+1) (s+{\frac{n}{2}}) (N-s-1)! (s-j-k)!}\,.
\end{split}$$ The second identity is just . The third is a change of the order of summation and the latest follows from the formula for the constants $c_{l,j,k}$.
In view of the elementary fact that $$(N-j)! \displaystyle\binom{N+s-1+{\frac{n}{2}}}{N-j}\binom{{\frac{n}{2}}+j+s-1}{k} =
\frac{\Gamma(s+{\frac{n}{2}}+N)}{k! \Gamma(s+{\frac{n}{2}}+j-k)}$$ we get $$\begin{split}
&\left| \sum_{k=0}^{N-1-j} \sum_{s=j+k}^{N-1} \frac{(-1)^s (N-j)!
\displaystyle\binom{N+s-1+{\frac{n}{2}}}{N-j}\binom{{\frac{n}{2}}+j+s-1}{k}}{(p-s+k)!
\Gamma({\frac{n}{2}}+p+s+1) (s+{\frac{n}{2}}) (N-s-1)! (s-j-k)!} \right | \\*[7pt] &\le
\sum_{k=0}^{N-1-j}\frac{1}{k!}\\*[7pt]
&\quad\times\sum_{s=j+k}^{N-1}\frac{1}{\Gamma(s+{\frac{n}{2}}+j-k)(p-s+k)!
({\frac{n}{2}}+p+s) (s+{\frac{n}{2}}) (N-s-1)! (s-j-k)!}\\*[7pt] &\le \sum_{k=0}^{N-1-j}
\frac{1}{k!}\sum_{s=j+k}^{N-1}\frac{1}{ (s-j-k)!} \le e^2 \,,
\end{split}$$ where in the first inequality we used that, since $N \le p$, $$\frac{\Gamma(s+{\frac{n}{2}}+N)}{\Gamma({\frac{n}{2}}+p+s+1)} \le \frac{1}{{\frac{n}{2}}+p+s}\,.$$ The proof of is complete.
\[L15\] Let $N-1\ge L\ge j\ge k\ge 0$ be integers. Then $$\begin{gathered}
\sum_{s=j}^{N-1} \frac{(-1)^s\binom{{\frac{n}{2}}+N+s-1}{N-k}\binom{{\frac{n}{2}}+k+s-1}
{s-j}}{(s+{\frac{n}{2}})(N-s-1)!\Gamma({\frac{n}{2}}+s+L+1)} \\*[7pt] =
(-1)^{j}\frac{(N-L-1)!}{(N-k)!}\binom{N-1}{L}\frac{1}{k!\,
\Gamma({\frac{n}{2}}+N)}\frac{1}{\binom{{\frac{n}{2}}+k+j-1}{k}}\,.\end{gathered}$$
To simplify notation set $$A = \sum_{s=j}^{N-1}
\frac{(-1)^s\binom{{\frac{n}{2}}+N+s-1}{N-k}\binom{{\frac{n}{2}}+k+s-1}
{s-j}}{(s+{\frac{n}{2}})(N-s-1)!\Gamma({\frac{n}{2}}+s+L+1)}\,.$$ Making the change of index of summation $i=s-j$ and using repeatedly the identity $\Gamma(x+1)=x\,\Gamma(x)$ we have $$\begin{split}
A &= \frac{(-1)^{j}}{\Gamma({\frac{n}{2}}+j+L)}\sum_{i=0}^{N-1-j}
\frac{(-1)^{i}\binom{{\frac{n}{2}}+N+i+j-1}{N-k}\binom{{\frac{n}{2}}+k+i+j-1
}{i}}{(i+j+{\frac{n}{2}})(N-i-j-1)!\{\prod_{p=0}^{i}({\frac{n}{2}}+i+j+L-p)\}}\\*[7pt]
&= \frac{(-1)^{j}}{\Gamma({\frac{n}{2}}+j+L)}\; B \,,
\end{split}$$ where the last identity defines $B$. To compute $B$ we need to rewrite the terms in a more convenient way so that we may apply well known equalities, among which the triple-binomial identity . Notice that $$\begin{gathered}
\frac{\binom{{\frac{n}{2}}+N+i+j-1}{N-k}\binom{{\frac{n}{2}}+k+i+j-1}{i}}{\{\prod_{p=0}^{i}({\frac{n}{2}}+i+j+L-p)\}}
\\*[7pt]
=\frac{(N-L-1)!(L-k)!}{i!(N-k)!} \binom{{\frac{n}{2}}+N+i+j-1}{N-L-1}\binom{{\frac{n}{2}}+j+L-1}{L-k}\,.\end{gathered}$$ Hence $$B\!=\!\sum_{i=0}^{N-1-j}\!\frac{(-1)^i(N\!-\!L\!-\!1)!(L-k)!}{i!(N\!-\!k)!(N\!-\!i\!-\!j\!-\!1)!(i+j+{\frac{n}{2}})}\binom{{\frac{n}{2}}+N+i+j-1}{N-L-1}\!\binom{{\frac{n}{2}}+j+L-1}{L-k}\,.$$ Since clearly $$\frac{\Gamma({\frac{n}{2}}+N)}{\Gamma({\frac{n}{2}}+j)} \frac{1}{(i+j+{\frac{n}{2}})(N-i-j-1)!}
= \binom{{\frac{n}{2}}+N-1}{N-j-i-1}\binom{{\frac{n}{2}}+j+i-1}{i}i! \,,$$ $B$ can be written as $$\begin{split}
B&=
\frac{\Gamma({\frac{n}{2}}+j)(N-L-1)!(L-k)!}{\Gamma({\frac{n}{2}}+N)(N-k)!}\binom{{\frac{n}{2}}+j+L-1}{L-k}
\\*[7pt]
&\quad\times\sum_{i=0}^{N-1-j}(-1)^i\binom{{\frac{n}{2}}+N-1}{N-j-i-1}\binom{{\frac{n}{2}}+j+i-1}{i}\binom{{\frac{n}{2}}+N+i+j-1}
{N-L-1} \\*[7pt] &=
\frac{\Gamma({\frac{n}{2}}+j)(N-L-1)!(L-k)!}{\Gamma({\frac{n}{2}}+N)(N-k)!}\binom{{\frac{n}{2}}+j+L-1}
{L-k}\;C \,,
\end{split}$$ where the last identity defines $C$.
To simplify notation set $L=m+j,$ where $0\le m\le L-j.$ Using the elementary identity $$\binom{{\frac{n}{2}}+N-1}{N-j-i-1}\binom{{\frac{n}{2}}+i+j-1}{i}
=\frac{(N-j)!}{i!(N-i-j-1)!({\frac{n}{2}}+i+j)}\binom{{\frac{n}{2}}+N-1}{N-j}\,,$$ we get $$C =\binom{{\frac{n}{2}}+N-1}{N-j}(N-j)!\sum_{i=0}^{N-1-j}\frac{(-1)^i
\binom{{\frac{n}{2}}+N+i+j-1}{N-j-m-1}}{i!\,(N-i-j-1)!\,({\frac{n}{2}}+i+j)}\,.$$ The only task left is the computation of the sum $$D(j,m)= \sum_{i=0}^{N-1-j}\frac{(-1)^i
\binom{{\frac{n}{2}}+N+i+j-1}{N-j-m-1}}{i!\,(N-i-j-1)!\,({\frac{n}{2}}+i+j)}\,.$$ The identity $$\frac{1}{{\frac{n}{2}}+i+j}=\frac{1}{{\frac{n}{2}}+j}\left(1-\frac{i}{{\frac{n}{2}}+i+j}\right)\,,$$ yields the expression $$\begin{split}
D(j,m)&
=\sum_{i=0}^{N-1-j}\frac{(-1)^i}{({\frac{n}{2}}+j)i!(N-i-j-1)!}\binom{{\frac{n}{2}}+i+j+N-1}{N-j-m-1}
\\*[7pt] &\quad -\sum_{i=0}^{N-1-j}\frac{(-1)^i
i}{i!({\frac{n}{2}}+j)({\frac{n}{2}}+j+i)(N-i-j-1)!}\binom{{\frac{n}{2}}+i+j+N-1}{N-j-m-1}\,.
\end{split}$$ The first sum in the above expression for $D(j,m)$ turns out to vanish for $m \ge 1$. This is because $$\begin{split}
\sum_{i=0}^{N-1-j}&\frac{(-1)^i}{({\frac{n}{2}}+j)i!(N-i-j-1)!}\binom{{\frac{n}{2}}+i+j+N-1}
{N-j-m-1} \\*[7pt]
&= \frac{1}{(N-j-1)!\;({\frac{n}{2}}+j)}\sum_{i=0}^{N-1-j}(-1)^i\binom{N-j-1}
{i}\binom{{\frac{n}{2}}+i+j+N-1}{N-j-m-1}\\*[7pt]
& = (-1)^{N-j-1} \binom{{\frac{n}{2}}+j+N-1}{-m}\\*[7pt]
& =0 \,,
\end{split}$$ where the next to the last equality follows from an identity proven in [@GKP (5.24), p. 169] and the last equality follows from the fact that $\displaystyle\binom{s}{k}=0$ provided $k$ is a negative integer. Hence, setting $s=i-1$, $$\begin{split}
D(j,m) &=-\sum_{i=0}^{N-1-j}\frac{(-1)^i\;
i}{i!({\frac{n}{2}}+j)({\frac{n}{2}}+j+i)(N-i-j-1)!}\binom{{\frac{n}{2}}+i+j+N-1}{N-j-m-1}\\*[7pt]
& =\frac{1}{{\frac{n}{2}}+j}\sum_{s=0}^{N-(j+1)-1}\frac{(-1)^s}{s!({\frac{n}{2}}+s+(j+1))}\frac{\binom{{\frac{n}{2}}+(j+1)+N+s-1}{N-(j+1)-(m-1)-1}}{(N-(j+1)-s-1)!} \\*[7pt]
& = \frac{1}{({\frac{n}{2}}+j)}D(j+1,m-1)\,.
\end{split}$$ Repeating the above argument $m$ times we obtain that $$D(j,m)=\frac{1}{({\frac{n}{2}}+j)({\frac{n}{2}}+j+1)\cdots ({\frac{n}{2}}+j+m-1)}\;D(L,0)\,.$$ To compute $D(L,0)$ we use the elementary identity $$\frac{1}{({\frac{n}{2}}+L+s)}=\frac{s!(N-L-s-1)!\Gamma({\frac{n}{2}}+L)}{\Gamma({\frac{n}{2}}+N)}\binom{{\frac{n}{2}}+L+s-1}
{s}\binom{{\frac{n}{2}}+N-1}{N-L-s-1}\,,$$ from which we get $$\begin{split}
D(L,0)&=\sum_{s=0}^{N-1-L}\!(-1)^s\frac{\Gamma({\frac{n}{2}}+L)}{\Gamma({\frac{n}{2}}+N)}\!\binom{{\frac{n}{2}}+L+s-1}{s}\!\binom{{\frac{n}{2}}+N-1}{N\!-\!L\!-\!s\!-\!1}\!\binom{{\frac{n}{2}}\!+\!L\!+\!N\!+\!s\!-\!1}{N-L-1}\\*[7pt]
&
=\frac{\Gamma({\frac{n}{2}}+L)}{\Gamma({\frac{n}{2}}+N)}\sum_{s=0}^{N-1-L}\binom{-{\frac{n}{2}}-L}{s}\binom{{\frac{n}{2}}+N-1}{N-L-s-1}\binom{{\frac{n}{2}}+L+N+s-1}{N-L-1}\\*[7pt]
&=\frac{\Gamma({\frac{n}{2}}+L)}{\Gamma({\frac{n}{2}}+N)}\binom{N-1}{L}\,,
\end{split}$$ where in the second identity we applied [@GKP (5.14), p. 164] and the latest equality is consequence of the triple-binomial identity [@GKP (5.28), p. 171] (for $ n=N-L-1$, $m=0$, $r={\frac{n}{2}}+N+L-1$ and $s=N-1$). Consequently, $$\begin{split}
D(j,m)& =\frac{\Gamma({\frac{n}{2}}+L)}{\Gamma({\frac{n}{2}}+N)}\binom{N-1}{L}\frac{1}{({\frac{n}{2}}+j)\cdots ({\frac{n}{2}}+L-1)}\\*[7pt]
&=\frac{\Gamma({\frac{n}{2}}+j)}{\Gamma({\frac{n}{2}}+N)}\binom{N-1}{L}\,.
\end{split}$$ Hence $$C=\binom{{\frac{n}{2}}+N-1}{N-j}\;(N-j)!\;
\frac{\Gamma({\frac{n}{2}}+j)}{\Gamma({\frac{n}{2}}+N)}\binom{N-1}{L}\,,$$ and $$\begin{gathered}
B=
\frac{\Gamma({\frac{n}{2}}+j)}{\Gamma({\frac{n}{2}}+N)}\frac{(N-L-1)!(L-k)!}{(N-k)!}\binom{{\frac{n}{2}}+j+L-1}{L-k}\binom{{\frac{n}{2}}+N-1}{N-j}\\*[7pt]
\times \binom{{\frac{n}{2}}+N-1}{N-j}\;(N-j)!\;\frac{\Gamma({\frac{n}{2}}+j)}{\Gamma({\frac{n}{2}}+N)}\binom{N-1}{L}\,.\end{gathered}$$ Finally, after appropriate simplifications, $$\begin{split}
A&= \frac{(-1)^{j}}{\Gamma({\frac{n}{2}}+j+L)}\;B \\*[7pt] &=
(-1)^{j}\frac{(N-L-1)!}{(N-k)!}\frac{\binom{N-1}{L}}{\Gamma({\frac{n}{2}}+N)k!\binom{{\frac{n}{2}}+j+k-1}
{k}}\,,
\end{split}$$ which completes the proof of the lemma.
Examples and questions
======================
Consider the polynomial operator $T=T_\lambda$ in $\mathbb{R}^2$ determined by $$\Omega(x,y) = \frac{xy}{|z|^2}+\lambda \frac{x^3y-x y^3}{|z|^4}\,,$$ where $z= x+i y$. We claim that the inequality $
\| T_\lambda^{\star} f \|_2 \leq C \| T_\lambda f \|_2$, $f \in L^2(\mathbb{R}^2)$, holds if and only if $|\lambda|<2$. This follows from the Theorem because the multiplier of the operator $T_\lambda$ is $$\frac{\xi\eta}{\xi^2 + \eta^2}\left(-\pi +\lambda\frac{\pi}{2} \frac{(\xi^2
-\eta^2)}{\xi^2 + \eta^2}\right)\,,$$ $\frac{\xi\eta}{\xi^2 + \eta^2} $ is the multiplier of a second order Riesz Transform and the function $ -\pi +\lambda\frac{\pi}{2}
\frac{(\xi^2 -\eta^2)}{\xi^2 + \eta^2}$ has no zeroes on the unit circle if and only if $|\lambda| < 2$.
Let $B$ stand for the Beurling transform and consider the operator $$T_\lambda = B-\lambda\,B^2, \quad \lambda \in {\mathbb{C}}\,,$$ which corresponds to $$\Omega(z)= -\frac{1}{\pi} \left(\frac{\overline{z}^2 |z|^2 +
2\,\lambda \, \overline{z}^4}{|z|^4}\right) \,.$$ In this case the multiplier is $$\frac{\overline{\xi}^3 \xi \,-\lambda\,\overline{\xi}^4}{|\xi|^4}=
\frac{\overline{\xi}}{\xi} \left( 1 - \lambda \,
\frac{\overline{\xi}}{\xi}\right),\quad \xi\in \mathbb C \, ,$$ which is the multiplier of $B$ times a function that vanishes on the unit circle if and only if $|\lambda|=1.$ Then the Theorem tells us that $ \| T_\lambda^{\star} f \|_2 \leq C \| T_\lambda f \|_2$, $f \in L^2(\mathbb{R}^2)$, if and only if $|\lambda|\ne 1$. Indeed, strictly speaking, the necessary condition of the Theorem, i.e.(iii) implies (i), applies only to *real* homogeneous polynomials. However, in the case at hand the factorization results one needs can be checked by direct inspection.
We give an example of a polynomial operator $T$ which is of the form $T=R\circ U$, where $R$ is an even higher order Riesz Transform and $U$ is invertible, but the $L^2$ inequality $ \| T^{\star} f
\|_2 \leq C \| T f \|_2$, $f \in L^2({{\mathbb R}^n})$, does not hold. The operator $T$ is associated with the homogenous polynomial of degree $8$ $$P(x,y)=\frac{1}{\gamma_2}\,P_2(x,y)(x^2+y^2)^3+\varepsilon \left(
\frac{1}{\gamma_4}\,P_4(x,y)(x^2+y^2)^2
-\frac{1}{\gamma_8}\,P_8(x,y)\right)\,,$$ where $$\begin{split}
P_2(x,y) &=xy \,,\\
P_4(x,y)&=x^4 -6x^2y^2 +y^4\,,\\
\intertext{and}
P_8(x,y)&=x^8 +y^8 -28x^6 y^2 -28 x^2y^6+70x^4 y^4
\end{split}$$ are harmonic polynomials and $\varepsilon$ is small enough. Notice that $P_2$ does not divide $P_4$ nor $P_8$. Therefore, by the Theorem, there is no control of $T^{\star}$ in terms of $T$. On the other hand, $P_4$ and $P_8$ have been chosen so that $P_4(\xi_1,\xi_2)|\xi|^4 -P_8(\xi_1,\xi_2)$ is divisible by $P_2$, and so the multiplier of $T$ is $$\frac{P_2(\xi_1,\xi_2)}{|\xi|^2}+\varepsilon\left(
\frac{P_4(\xi_1,\xi_2)}{|\xi|^4}
-\frac{P_8(\xi_1,\xi_2)}{|\xi|^8}\right) =
\frac{P_2(\xi_1,\xi_2)}{|\xi|^2} \left(1 +\varepsilon
\frac{Q(\xi_1,\xi_2)}{|\xi|^6} \right)\,,$$ where $Q$ is a homogeneous polynomial of degree $6$. Define $R$ as the higher order Riesz transform whose multiplier is $\frac{P_2(\xi_1,\xi_2)}{|\xi|^2}$ and $U$ as the operator whose multiplier is $1 +\varepsilon \frac{Q(\xi_1,\xi_2)}{|\xi|^6}$. If $\varepsilon$ is small enough, then $U$ is invertible.
We show here that condition (iii) in the Theorem is very restrictive. Take $n=2$. Since $\Omega(e^{i\theta})$ is an even function it has, modulo a rotation, a Fourier series expansion of the type $$\label{eq80}
\Omega(e^{i\theta}) = \sum_{j\geq 1} a_j\, \sin (2 j \theta)\,.$$ Thus the harmonic polynomials $P_{2j}$ are $$P_{2j}(z) = a_j \frac{z^{2j}-\overline{z}^{2j}}{2i}\,.$$ If $a_j \ne 0$, then $P_{2j}$ vanishes exactly on $2j$ straight lines through the origin uniformly distributed and containing the two axis. Each such line is determined by a pair of opposed $4j$-th roots of the unity. Assume now that the operator determined by $\Omega (e^{i\theta})$ satisfies condition $(iii)$ of the Theorem. Let $j_0$ be the first positive integer with $a_{j_0} \ne
0$. Then only the $a_j$ with $j$ a multiple of $j_0$ may be non-zero, owing to the particular structure of the zero set of $P_{2j}$. In other words, the first non-zero $P_{2j}$ determines all the others, modulo the constants $a_j$.
We show now a method to construct even kernels in $\mathbb{R}^3$ that satisfy condition $(iii)$ in the Theorem. It can be easily adapted to any dimension. The kernel is determined by the function $$\Omega(x,y,z)= xy \,\sum_{j\geq 0} \epsilon_j\,Q_{2j}(x,y,z)$$ where the sequence $(\epsilon_j)$ is chosen so that $\Omega$ is in $C^\infty(S^{2})$ and the $Q_{2j}$ are defined by $$Q_{2j}(x,y,z)= \sum_{k=0}^{2j} c_k\,y^{2k}\,z^{2j-2k}\,.$$ The $c_k$ are determined by a recurrent formula obtained by requiring that$ xy\,Q_{2j}(x,y,z)$ is harmonic. Computing its Laplacean we get the recurrent condition $$c_k= - c_{k-1} \frac{2k (2k+1)}{(2j-2k+1)\,(2j-2k+2)},\quad 1\leq k
\leq j\,,$$ where $c_0$ may be freely chosen.
We would like to close the paper by asking a couple of questions which we have not been able to answer.
Since our methods are very much dependent on the Fourier transform we do not know whether either the weak type inequality $$\|T^\star f \|_{1,\infty} \leq C\| Tf \|_1 \,,$$ or the $L^p$ inequality with $ 1<p< \infty$, $p \ne 2$, $$\| T^{\star} f \|_p \leq C \| T f \|_p ,\quad f \in L^p({{\mathbb R}^n})\,,$$ imply the $L^2$ inequality $$\| T^{\star} f \|_2 \leq C \| T f \|_2,\quad f \in L^2({{\mathbb R}^n})\,.$$ As suggested by Carlos Pérez, this might be related to interpolation results for couples of sub-linear operators.
How far may the smoothness assumption on $\Omega$ be weakened ? More concretely, does the Theorem still hold true for $\Omega$ of class $C^m(S^{n-1})$ for some positive integer $m$ ?
The authors were partially supported by grants2009SGR420 (Generalitat de Catalunya) and MTM2010-15657 (Ministerio de Ciencia e Innovación).
The authors are indebted to F. Cedó and X. Xarles for several illuminating conversations on the Division Lemma.
[MOPV]{}
N. Aronszajn, T. Creese and L. Lipkin, [*Polyharmonic functions*]{}, Oxford Mathematical Monographs, Oxford University Press, New York, (1983).
A.L. Bertozzi and A.J. Majda, [*Vorticity and Incompressible Flow*]{}, Cambridge texts in applied Mathematics, Cambridge University Press, Cambridge, (2002).
A.P. Calderón and A. Zygmund, [*On a problem of Mihlin*]{}, Trans. Amer. Math. Soc. Math. [**78**]{} (1955), 209–224.
J.Y. Chemin, [*Fluides parfaits incompressibles*]{}, Astérisque [**230**]{}, Société Mathématique de France (1995).
G. David and S. Semmes, [*Singular integrals and rectifiable sets in ${{\mathbb R}^n}$: Au-delàs des graphes lipschitziens*]{}, Astérisque [**193**]{}, Société Mathématique de France (1991).
L. Grafakos, [*Classical and modern Fourier analysis*]{}, Pearson, New Jersey, (2004).
R. Graham, D. Knuth and O. Patashnik, [*Concrete mathematics. A foundation for computer science*]{}, Second edition. Addison-Wesley Publishing Company, Reading, MA, (1994).
E. Kunz, [*Introduction to commutative algebra and algebraic geometry,*]{} Birkhäuser, Basel, (1985).
L. Lorch and P. Szego, [*A singular integral whose kernel involves a Bessel function*]{}, Duke Math. J. [**22**]{} (1955), 410–418.
G. G. Lorentz, [*Approximation of functions,*]{} Chelsea Publishing Company, New York, (1986).
R. Lyons and K. Zumbrun, [*Homogeneous partial derivatives of radial functions*]{}, Proc. Amer. Math. Soc. [**121**]{}(1) (1994), 315–316.
J. Mateu, Y. Netrusov, J. Orobitg and J. Verdera, [*BMO and Lipschitz approximation by solutions of elliptic equations*]{}, Ann. Inst. Fourier (Grenoble) **46**(4) (1996), 1057–1081.
J. Mateu and J. Orobitg, [*Lipschitz approximation by harmonic functions and some applications to spectral synthesis*]{}, Indiana Univ. Math. J. **39**(3) (1990), 703–736.
J. Mateu, J. Orobitg and J. Verdera, [*Extra cancellation of even Calderón-Zygmund operators and quasiconformal mappings*]{}, J. Math. Pures Appl. **91** (4)(2009), 402–-431.
J. Mateu, J. Orobitg, C. Pérez and J. Verdera, [*New Estimates for the Maximal Singular Integral*]{}, Int Math Res Notices **2010**(19)(2010), 3658–3722.
J. Mateu, L. Prat and J. Verdera, [*The capacity associated to signed Riesz kernels, and Wolff potentials*]{}, J. Reine Angew. Math. **578** (2005), 201–223.
J. Mateu and J. Verdera, [*BMO harmonic approximation and spectral synthesis for Hardy-Sobolev spaces*]{}, Rev. Mat. Iberoamericana **4**(2) (1988), 291–318.
J. Mateu and J. Verdera, [*$L^p$ and weak $L^1$ estimates for the maximal Riesz transform and the maximal Beurling transform*]{}, Math. Res. Lett. [**13**]{}(5-6) (2006), 957–966.
P. Mattila and J. Verdera, [*Convergence of singular integrals with general measures*]{}, J. Eur. Math. Soc. **11** (2)(2009), 257–-271.
F. Ricci and E.M. Stein, [*Harmonic analysis on nilpoptent groups and singular integrals I: Oscillatory integrals*]{}, J. Funct. Anal. [**78**]{}(1987), 179–194.
E. M. Stein, [*Singular integrals and differentiability properties of functions,*]{} Princeton University Press, Princeton, (1970).
E. M. Stein and G. Weiss, [*Introduction to Fourier Analysis on Euclidean Spaces,*]{} Princeton University Press, Princeton, (1971).
J. Verdera, [*$C\sp m$ approximation by solutions of elliptic equations, and Calderón-Zygmund operators*]{}, Duke Math. J. **55**(1) (1987), 157–187.
J. Verdera, [*$L^2$ boundedness of the Cauchy Integral and Menger curvature,*]{} Contemporary Mathematics [**277**]{} (2001), 139–158.
[l]{} Joan Mateu\
Departament de Matemàtiques\
Universitat Autònoma de Barcelona\
08193 Bellaterra, Barcelona, Catalonia\
[*E-mail:*]{} [[email protected]]{}\
\
Joan Orobitg\
Departament de Matemàtiques\
Universitat Autònoma de Barcelona\
08193 Bellaterra, Barcelona, Catalonia\
[*E-mail:*]{} [[email protected]]{}\
\
Joan Verdera\
Departament de Matemàtiques\
Universitat Autònoma de Barcelona\
08193 Bellaterra, Barcelona, Catalonia\
[*E-mail:*]{} [[email protected]]{}
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The Ernst–like matrix representation of the multidimensional Einstein–Kalb–Ramond theory is developed and the O(d,d)–symmetry is presented in the matrix–valued $SL(2,R)$–form. The analogy with the Einstein and Einstein–Maxwell–Dilaton–Axion theories is discussed.'
address:
- |
Joint Institute for Nuclear Research,\
Dubna, Moscow Region 141980, RUSSIA,\
e-mail: [email protected]
- |
Nuclear Physics Institute,\
Moscow State University,\
Moscow 119899, RUSSIA,\
e-mail: [email protected]
author:
- Alfredo Herrera
- and
- Oleg Kechkin
date: December 1996
title: |
O(d, d)–Symmetry and Ernst Formulation for\
Einstein–Kalb–Ramond Theory in Three Dimensions
---
-10mm
Introduction
============
Recently much attention had been attracted to the study of symmetries of the dimensionally reduced low energy effective string theory. Such a theory describes the gravitational, dilaton, Kalb–Ramond and a set of $n$ Abelian vector fields in $D + d$ dimensions. It becomes $O(d, d+n)$–invariant after the compactification of $d$ dimensions on a torus [@ms] and allows the $O(d, d+n)/O(d) \times O(d+n)$ coset matrix formulation in the case of $D = 3$ [@s1]. The transformations of the $O(d, d+n)$ group were explored for the generation of new solutions in this theory as well as for the prediction of the features of exact string excitations [@s2]–[@tc].
The four–dimensional Einstein–Maxwell–Dilaton–Axion (EMDA) theory, being the simplest model of this type, admits instead of the orthogonal representation the symplectic one [@gk1]. It was established that this system possesses the $Sp(4,R)$ symmetry group and allows the coset $Sp(4, R)/U(2)$ matrix representation in the stationary case. In [@gk2] the Kähler formulation for this model, which generalizes the well known Ernst formalism for the stationary Einstein system, was established. Finally, in [@ky1] it was established the complete formal analogy between EMDA and vacuum Einstein theories in the stationary and stationary axisymmetric cases.
In this paper we show that the $O(d, d)$ symmetric ($n=0$) effective heterotic string theory, which describes the Einstein and Kalb–Ramond (EKR) fields, admits classical procedures well known for the stationary Einstein system. Namely, the EKR relations directly map to the Einstein ones under the change of the symmetric matrix $G$, constructed on the additional Kaluza–Klein metric components, by the real part of the vacuum Ernst potential $f$, and the antisymmetric Kalb–Ramond matrix $B$ by its imaginary part $i\chi$, correspondingly. We construct the matrix scalar–vector Lagrange representation of the model which relates with the original metric (non–target space) formulation of the Einstein theory.
Following Maharana and Schwarz we unite the metric and Kalb–Ramond matrices into the $d \times d$ matrix $X = G + B$ which provides the Ernst–like formulation of the problem. It is shown that in terms of this matrix the target space duality group becomes the matrix–valued $SL(2, R)$ one. Its three subgroups are identified as the matrix generalizations of the gauge shift, rescaling and Ehlers transformations for the vacuum Einstein theory. The consequences of the imposition of the additional axisymmetric property will be presented in the following publications.
Chiral Matrix Formulation
=========================
In this paper we discuss the system with the action $${\cal S} = \int d^{3+d}x {\mid {\cal G} \mid}^{\frac {1}{2}} \left\{ - {\cal R} +
\frac {1}{12} {\cal H}^2 \right\},$$ where ${\cal R}$ is the Ricci scalar for the metric ${\cal G}_{M N}$, $(M = 0, ..., 3 + d)$ and $${\cal H}_{MNL} = \partial _{M} {\cal B}_{NL}
+ {\rm cyc.\,\, perms.}$$ Such a system arises in the low energy limit of heterotic string theory or in the frames of $N=1$ supergravity. It does not contain the dilaton and gauge vector fields, so that this theory coincides with the $(3 + d)$–dimensional Einstein–Kalb–Ramond one.
Following Maharana and Schwarz [@ms], one can extract $d$ dimensions using the Kaluza–Klein compactification on a torus. Doing so one obtains the $O(d, d)$–symmetric $\sigma$–model action describing 3–dimensional gravity coupled with a set of scalar and Abelian vector fields [@s1]. (The Sen’s formulae can be transformed into the our ones using the replacement $7 \rightarrow d$ and $16 \rightarrow n$.) In this work we study the case when the vector fields are not present, i.e., we suggest that the metric and Kalb–Ramond field components with mixed indeces are equal to zero: $${\cal G}_{\mu, n + 2} = {\cal B}_{\mu, n + 2} = 0,$$ $\mu = 0, ..., 2; \, n = 1, ..., d$. It is easy to see that such a restriction does not provide any constraints on the remainder variables and can be considered as a non–trivial ansatz for the EKR theory.
In [@ms] and [@s1] it is shown that the result of the Kaluza–Klein compactification can be represented using the following 3–dimensional effective action: $$\begin{aligned}
^3 S =
\int d^3 x {\mid g \mid}^{\frac {1}{2}} \left\{ - ^3 R +
\frac {1}{8} Tr\left[(J^M)^2\right]
\right\},\end{aligned}$$ where $J^M = \nabla M \, M^{-1}$. Here the curvature scalar $^3 R$ is constructed on the 3–metric $g_{\mu \nu} = {\cal G}_{\mu \nu}$ and the chiral matrix $M$ is defined as $$\begin{aligned}
M = \left (\begin{array}{crc}
G^{-1} &\quad & - G^{-1}B\\
BG^{-1} &\quad & G - BG^{-1}B\\
\end{array}\right ),\end{aligned}$$ with $G_{mn} = {\cal G}_{m + 2, n + 2}$ and $B_{mn} =
{\cal B}_{m + 2, n + 2}$ (thus $G^T = G, \, B^T = - B$). This matrix satisfies the relations $$\begin{aligned}
M^T = M, \qquad M \eta M = \eta, \qquad {\rm where} \qquad
\eta = \left (\begin{array}{crc}
0 & I_d\\
I_d & 0\\
\end{array}\right ),\end{aligned}$$ i. e., belongs to the coset $O(d, d)/O(d) \times O(d)$ [@ms].
It is easy to see that the form of the Gauss decomposition (5) is very similar to the ones of Einstein and Einstein–Maxwell–Dilaton–Axion theories [@ky1]. Moreover, the case of an arbitrary value of $d$ corresponds to the general case of the $Sp(2d, R)/U(d)$ chiral matrix [@ky2]. The only difference is connected with the sign “-” in the upper right part of (5) and with the antisymmetric character of the matrix $B$. As it will be shown below, the analogy between EKR theory and symplectic systems permits to perform all the classical procedures well known for the stationary Einstein theory.
The action (4) allows the symmetry transformations belonging to the group $O(d, d)$. Namely, any matrix $C \in O(d, d)$, i. e. satisfying the relation $C^T \eta C = \eta$, defines an automorphism $$M \rightarrow C^T M C$$ for the target space of the problem. The explicit form of a group matrix which can be continuously transformed to the unit one is: $$\begin{aligned}
C = \left (\begin{array}{crc}
(S^T)^{-1} & \quad & - (S^T)^{-1} R\\
- L (S^T)^{-1} & \quad & S + L(S^T)^{-1}R\\
\end{array}\right ),\end{aligned}$$ where $R^T = - R$ and $L^T = - L$. As an example of $O(d, d)$ transformation which can not be parametrized according to (8), one can take the transformation defined by the matrix $\eta$. It corresponds to the recently discussed strong–weak coupling duality transformation $M \rightarrow M^{-1}$ ([@s1] and [@s3]) which formally exists for any chiral matrix.
It must be noted that the EKR theory with ${\cal G}_{\mu, 2 + m} \neq 0$ and ${\cal B}_{\mu, 2 + m} \neq 0$ is also $O(d, d)$–symmetric [@ms], but only the ansatz under consideration allows the formalism developed below.
Ernst Matrix Potential
======================
The set of Euler–Lagrange equations corresponding to the 3–action (4) $$\nabla J^M = 0,$$ $$^3R_{\mu \nu} = \frac {1}{8} Tr\left[J^M_{\mu}J^M_{\nu}\right],$$ can also be written in terms of the matrices $G$ and $B$: $$\nabla J^B - J^G J^B = 0,$$ $$\nabla J^G - (J^B)^2 = 0,$$ $$^3R_{mn} = \frac {1}{4}Tr\left[J^G_{\mu}J^G_{\nu} - J^B_{\mu}J^B_{\nu}\right],$$ where $J^G = \nabla G \, G^{-1}$ and $J^B = \nabla B \, G^{-1}$. One can prove that these equations correspond to the action $$\begin{aligned}
^3 S =
\int d^3x {\mid g \mid}^{\frac {1}{2}} \left\{ - ^3 R +
\frac {1}{4} Tr\left[(J^G)^2 - (J^B)^2\right]
\right\}.\end{aligned}$$
This action, as well as Eqs. (11)–(13), can be directly transformed into the stationary Einstein ones by the replacement $$G \rightarrow f, \qquad B \rightarrow i\chi,$$ where $f$ is the $g_{tt}$ metric coefficient and $\chi$ is the rotational potential. It is well known that the Ernst potential $E = f + i\chi$ [@e] plays an important role in the Einstein theory. It provides the Kähler formulation for the stationary Einstein system [@m], and allows to represent the stationary axisymmetric field configurations in compact form. Using the substitution (15) one can introduce the analogy of the Ernst potential of the Einstein theory for the EKR one: $$X = G + B,$$ which firstly was considered in [@ms].
It can easily be verified that Eqs. (11)–(13) can be rewritten as $$\nabla J^X = J^X\left(J^X + J^{X^T}\right),$$ $$^3 R_{\mu \nu} = Tr\left(J^X_{(\mu}J^{X^T}_{\nu)}\right),$$ where $J^X = \nabla X (X + X^T)^{-1}$. This system is very similar to the Ernst one [@e]–[@ky2] for the symplectic theories and so, the matrix function $X$ can be called “matrix Ernst potential” by natural reasons.
The corresponding to Eqs. (17) and (18) action has the form $$^3 S = \int d^3x {\mid g \mid}^{\frac {1}{2}}\left\{ - ^3 R +
Tr(J^X J^{X^T})\right\}.$$ One can see that the complex conjugation in the Einstein case transforms into the transposition in the EKR one. The symmetries of this action can be directly obtained from the formulae (5), (7) and (8) for the chiral matrix $M$. As result one obtains that the transformation $$X \rightarrow S^T[X^{-1} + L]^{-1}S + R$$ has the same sense that Eq. (7). Thus, the complete set of isometry transformations of EKR theory has a matrix–valued $SL(2, R)$ form. This allows to understand the subgroups determined by matrices $R$, $S$ and $L$ as the matrix analogies of gauge shift, rescaling and Ehlers transformations, correspondingly [@k]–[@eh].
Dualization Procedure
=====================
Now let us establish an alternative Lagrange formulation of the problem based on the use of non–target space variables. One can see that Eq. (11), being rewritten as $$\nabla [G^{-1}(\nabla B)G^{-1}] = 0,$$ ensures the compatibility conditions for the relation $$\nabla \times \vec \Omega = G^{-1}(\nabla B)G^{-1}$$ which defines the antisymmetric vector matrix $\vec \Omega$. This matrix provides another representation of the EKR system corresponding to the original metric one of the case of Einstein theory [@iw] and [@k], i. e., the formulation with the metric components $\omega _i = {g_{tt}}^{-1}g_{ti}$.
It is easy to see from Eq. (22) that the matrices $G$ and $\vec \Omega$ satisfy the relation which also can be rewritten using the matrix current $J^{\vec \Omega} =
G \,\,\nabla \times \vec \Omega$ as $$\nabla \times J^{\vec \Omega} - J^{\vec \Omega}\times J^G = 0.$$ This relation together with Eqs. (12) and (13), expressed in terms of the matrices $G$ and $\vec \Omega$ $$\nabla J^G - \left(J^{\vec \Omega}\right)^2 = 0,$$ $$^3 R_{\mu \nu} =\frac {1}{4} Tr\left(J^G_{\mu} J^G_{\nu} -
J^{\vec \Omega}_{\mu}J^{\vec \Omega}_{\nu}\right)$$ form the complete set of the Euler–Lagrange equations for the action $$^3 S = \int d^3x {\mid h \mid}^{\frac {1}{2}}\left\{- ^3 R + \frac {1}{4}
Tr\left[(J^G)^2 + (J^{\vec \Omega})^2\right]\right\};$$ thus these matrices provide an alternative Lagrange formulation of the theory under consideration. A similar procedure for the symplectic $Sp(2d, R)/U(d)$ theories has been performed in [@ky2].
At the end of the paper we would like to remark that the low energy limit of heterotic string theory with dilaton and $n$ Abelian vector fields, being reduced to three dimensions [@s1], also allows a dualization procedure. As it can be shown, this procedure is more similar to the Einstein–Maxwell one, than to discussed here Einstein–like construction. The reason of such a remarkable circumstance is related with the closed analogy between the chiral matrix of the string system with the coset matrix of Eistein–Maxwell theory [@gek].
Conclusion
==========
It is easy to see that Einstein–Kalb–Ramond system allows all the classical procedures established for the Einstein and Einstein–Maxwell–Dilaton–Axion theories in the stationary axisymmetric case. Namely, in the forthcoming publications it will be shown that this system, being reduced to two dimensions, also allows the Kramer–Neugebauer transformation ([@kn], [@ky1] and [@ky2]), the alternative (Belinsky–Zakharov–like [@bz]) chiral matrix formulation and, moreover, the construction of the Geroch group [@g1]–[@g2] and the O(d, d) matrix analogy of the Hauser–Ernst formalism [@he1]–[@wg]. We would like to thank our colleagues from the JINR and NPI for an encouraging relation to our work. One of the authors (A. H.) would like to thank CONACYT and SEP for partial financial support.
J. Maharana and J.H. Schwarz, Nucl. Phys. [**B390**]{} (1993) 3. A. Sen, Nucl. Phys. [**B434**]{} (1995) 179. T. Horowitz and A. Sen, Phys. Rev. [**D53**]{} (1996) 808. M. Cvetic and D. Youm, Nucl. Phys. [**B476**]{} (1996) 118. M. Cvetic and A. Tseytlin, Phys. Rev. [**D53**]{} (1996) 5619. D.V. Gal’tsov and O.V. Kechkin, Phys. Rev. [**D54**]{} (1996) 1656. D.V. Gal’tsov and O.V. Kechkin, Phys. Lett. [**B361**]{} (1995) 52. O. Kechkin and M. Yurova, Phys. Rev. [**D54**]{} (1996) 6132. O. Kechkin and M. Yurova, “Symplectic gravity models in four,\
three and two dimensions”, hep-th/9610222. A. Sen, Int. J. Mod. Phys. [**A9**]{} (1993) 3707. F.J. Ernst, Phys. Rev. [**167**]{} (1968) 1175. P.O. Mazur, Acta Phys. Pol. [**14**]{} (1983) 219. W. Kinnersley, J. Math. Phys. [**18**]{} (1977) 529. J. Ehlers, in “Les Theories de la Gravitation” (CNRS, Paris, 1959). W. Israel and G. A. Wilson, J. Math. Phys. [**13**]{} (1972) 865. A. Eris, A. Karasu and M. Gurses, J. Math. Phys. [**25**]{} (1984) 1489. G. Neugebauer and D. Kramer, Ann. Phys. (NY) [**24**]{} (1969) 62. V.A. Belinsky and V.E. Zakharov, Zh. Exp. Teor. Fiz. [**75**]{}\
(1978) 1955 \[Sov. Phys. JETP [**48**]{} (1978) 985\]. R. Geroch, J. Math. Phys. [**12**]{} (1971) 918. R. Geroch, J. Math. Phys. [**13**]{} (1972) 394. I. Hauser and F.J. Ernst, Phys. Rev. [**D20**]{} (1979) 362. I. Hauser and F.J. Ernst, J. Math. Phys. [**21**]{} (1980) 1126. Y.–S. Wu and M.–L. Ge, J. Math. Phys. [**24**]{} (1983) 1187.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Quantum money is a cryptographic protocol in which a mint can produce a quantum state, no one else can copy the state, and anyone (with a quantum computer) can verify that the state came from the mint. We present a concrete quantum money scheme based on superpositions of diagrams that encode oriented links with the same Alexander polynomial. We expect our scheme to be secure against computationally bounded adversaries.'
author:
- |
Edward Farhi, David Gosset, Avinatan Hassidim,\
Andrew Lutomirski, and Peter Shor
bibliography:
- 'knot-bib.bib'
date: 'April 28, 2010'
title: Quantum money from knots
---
(\#1,\#2) (\#1,\#2)[ (\#1-0.1,\#2-0.1) rectangle (\#1+0.1,\#2+0.1) ]{} \#1\#2
\#1;\#2
(\#1,\#2,\#3)[(\#1+0.1,\#3) – (\#2-0.1,\#3);]{} (\#1,\#2,\#3)[(\#1,\#2+0.1) – (\#1,\#3-0.1);]{}
(\#1,\#2)[(\#1-0.45,\#2) – (\#1-0.50,\#2);]{} (\#1,\#2)[(\#1+0.45,\#2) – (\#1+0.50,\#2);]{} (\#1,\#2)[(\#1,\#2+0.45) – (\#1,\#2+0.50);]{} (\#1,\#2)[(\#1,\#2-0.45) – (\#1,\#2-0.50);]{}
\_common(\#1,\#2)[(\#1-0.3,\#2-0.3) rectangle (\#1+0.3,\#2+0.3);]{} (\#1,\#2)[\_common(\#1,\#2) (\#1-0.2,\#2-0.2) – (\#1+0.2,\#2+0.2) (\#1-0.2,\#2+0.2) – (\#1+0.2,\#2-0.2);]{} (\#1,\#2)[\_common(\#1,\#2) (\#1,\#2) circle (0.2);]{}
(\#1,\#2,\#3)[(\#1+0.1,\#3-0.02) rectangle (\#2-0.1,\#3+0.02);]{} (\#1,\#2,\#3)[(\#1+0.1,\#3-0.02) rectangle (\#2-0.1,\#3+0.02);]{} (\#1,\#2,\#3)[(\#1-0.02,\#2+0.1) rectangle (\#1+0.02,\#3-0.1);]{} (\#1,\#2,\#3)[(\#1-0.02,\#2+0.1) rectangle (\#1+0.02,\#3-0.1);]{}
(\#1,\#2)[node at (\#1+0.5,0) [ \#2]{};]{} (\#1,\#2)[node at (0,\#1+0.5,0) [\#2]{};]{}
(\#1,\#2,\#3,\#4)[(\#1+0.1,\#2+0.1) rectangle (\#3-0.1,\#4-0.1);]{} (\#1,\#2,\#3,\#4)[(\#1-0.04,\#2-0.04) rectangle (\#3+0.04,\#4+0.04);]{}
Introduction
============
In this paper, we present quantum money based on knots (see Figure \[fig:dollar-sign\]). The purported security of our quantum money scheme is based on the assumption that given two different looking but equivalent knots, it is difficult to explicitly find a transformation that takes one to the other.
![Quantum money from knots\[fig:dollar-sign\]](braidmoney)
One of the problems in classical security is that information can be copied: passwords can be stolen, songs can be pirated, and when you email an attachment, you still have the original. One implication is that E-commerce requires communicating with a server (e.g. the credit card company or PayPal) whenever one makes a transaction. One could hope that the no-cloning theorem would help circumvent this and enable a physical quantum state to function like money. Such money could be used in transactions both in person and on a future “Quantum Internet”, not requiring contact with a central authority. As we will see, quantum money is harder to forge than the paper bills in our wallets. In order for a quantum state to function as money, we require that:
1. The mint can produce it.
2. Anyone can verify it. That is, there is an efficient measurement that anyone (with a quantum device) can perform on a quantum money state that accepts genuine money with high probability and without significantly damaging the money.
3. No one can forge it. That is, no one but the mint can efficiently produce states that are accepted by the verifier with better than exponentially small probability. In particular, it should not be possible to copy a bill.
A quantum money scheme has two components: pieces of quantum money and an algorithm $M$ that verifies quantum money. A piece of quantum money consists of a classical serial number $p$ (authenticated as coming from the mint) along with an associated quantum state $|\$_{p}\rangle$ on $n$ qubits. The verification algorithm $M$ takes as input a quantum state $|\phi\rangle$ and a serial number $q$ and then decides whether or not the pair $\left(q,|\phi\rangle\right)$ is a piece of quantum money. If the outcome is “good money” then the verifier also returns the state $|\phi\rangle$ undamaged so it can be used again. We now formalize each of the above requirements:
1. \[enu:There-is-an\]There is a polynomial-time algorithm that produces both a quantum money state $|\$_{p}\rangle$ and an associated serial number $p$.
2. Running the verification algorithm $M$ with inputs $p$ and $|\$_{p}\rangle$ returns “good money” and does not damage $|\$_{p}\rangle$. Furthermore, anyone with access to a quantum computer (for example a merchant) can run the verification algorithm $M$.
3. Given one piece of quantum money $\left(p,|\$_{p}\rangle\right)$, it is hard to generate a quantum state $|\psi\rangle$ on $2n$ qubits such that each part of $|\psi\rangle$ (along with the original serial number $p$) passes the verification algorithm.
What stops anyone other than the mint, using the same algorithm as the mint, from producing counterfeit money states? This is why the serial number needs to be authenticated. When the mint does a production run it produces a set of pairs $\left(p,|\$_{p}\rangle\right)$. In our quantum money scheme, the mint does not in advance choose the serial numbers, but rather they are produced by a random process. A rogue mint running the same algorithm as the mint can produce a new set of money pairs, but (with high probability) none of the serial numbers will match those that the mint originally produced. A simple way to ensure security is for the mint to publish the list of valid serial numbers and for the merchant to check against this list to see if the serial number from the tendered money is authentic.
An alternative to publishing a list of serial numbers is for the money state to come with a digital signature of the serial number. There are known classical digital signature protocols that allow anyone who knows the mint’s “public key” to verify that a serial number is one that the mint did indeed produce. Still, the authenticated description of the serial number can easily be copied and this is where the quantum security comes into play: knowledge of the serial number $p$ does not allow one to copy the associated quantum state $|\$_{p}\rangle$. In the remainder of this paper, we assume that anyone verifying quantum money will also check the authentication of the serial number.
We can imagine two different ways to use quantum money for commerce. If we had the technology to print a quantum state onto a piece of paper, then we could use quantum money protocols to enhance the security of paper money against forgery. Alternatively, people could use quantum money states on their personal quantum computers to conduct business either in person or over a quantum Internet. If small portable quantum computers were available (imagine quantum smart phones or quantum debit cards), then it would be easy to buy things with quantum money instead of paper money.
For these uses, the “quantum money” seen by an end-user would either be a file on a quantum computer or a physical piece of money with a quantum state somehow attached.
Prior work
----------
The idea of quantum money was introduced by Wiesner in 1969, where he proposed using the no cloning theorem to generate bills which can not be copied [@BBBW83; @Wie83]. However, in Wiesner’s original definition the only entity that could verify the quantum money state was the mint that produced it.
There is recent interest in designing quantum money which can be verified by anyone who has a quantum computer (also called “public-key quantum money”). Such money cannot be information-theoretically secure; instead, it must rely on computational assumptions. Aaronson showed that public-key quantum money exists relative to a quantum oracle and proposed a concrete scheme without an oracle [@aaronson-quantum-money]. Aaronson’s concrete scheme was broken in [@LAF10].
A different approach to quantum money was taken by Mosca and Stebila [@mosca-2009; @MS06; @MS07; @Ste09], which is somewhat based on ideas of Tokunaga, Okamoto, and Imoto [@TOI03]. This line of research adds the additional requirement that all quantum money states be identical and all share the same verification procedure. This additional requirement is an abstraction of coins (which are identical) as opposed to bills (which can have different serial numbers). In their works, Mosca and Stebila give a way to generate quantum coins using a specialized random quantum oracle. They also show how to generate quantum coins using blind quantum computation. This requires the mint to be involved in every transaction.
The main open question is how to design any secure quantum money (whether of the quantum coin variety or otherwise) that does not require an oracle and in which the mint’s participation is not needed to use the money.
A natural approach would be for the mint to choose some classical secret and, using the secret, generate the money state and the serial number. It should be hard to copy the quantum money state without knowing the secret. One such proposal for a quantum money scheme is based on product states. The mint first chooses a random product state $|\$\rangle$ on $n$ qubits and then constructs a set of local projectors $|\psi_{i}\rangle\langle\psi_{i}|$ for $i\in\{1,...m\}$ such that $|\$\rangle$ is a zero eigenvector of each projector (and $m$ is large enough that $|\$\rangle$ is the only state which is in the common zero eigenspace). The serial number associated with the state $|\$\rangle$ is a bit string which describes each of the projectors. To verify a state $|\phi\rangle,$ one measures each projector $|\psi_{i}\rangle\langle\psi_{i}|$ and accepts the state if and only if the outcomes are all $0$. Unfortunately this money scheme is insecure. It was shown in [@state-restoration] that given a piece of product state quantum money $|\$\rangle$ and the associated serial number it is possible to learn the classical secret (the description of the product state). A generalized version of this attack was also given in [@state-restoration].
Due to the possibility of this type of attack, it was proposed in [@LAF10] to use quantum money which is not based on a classical secret, but rather on the hardness of generating a known superposition. However, [@LAF10] did not present a proposal for a quantum money scheme, only a blueprint.
A blueprint for quantum money
-----------------------------
We present a scheme for quantum money which is not based on quantum oracles and does not require communicating with the mint for verification. The overall structure of the scheme is based on ideas from [@LAF10].
The mint begins by generating a uniform superposition
$$|\text{initial\ensuremath{\rangle}=}\frac{1}{\sqrt{|\mathcal{B}|}}\sum_{e\in\mathcal{B}}|e\rangle|0\rangle$$ where $\mathcal{B}$ is a big set. It then applies some function $f$ to $e$, obtaining the state
$$\frac{1}{\sqrt{|\mathcal{B}|}}\sum_{e\in\mathcal{B}}|e\rangle|f(e)\rangle.$$ Finally, the mint measures the value of $f$. If the result of the measurement is $v$, then the residual state is
$$|\$_{v}\rangle|v\rangle=\frac{1}{\sqrt{N_{v}}}\sum_{\substack{e\in\mathcal{B}\\
f(e)=v}
}|e\rangle|v\rangle.$$ The quantum money is the classical serial number $v$ and the quantum state $|\$_{v}\rangle$.
To verify the money state, the verifier first measures the value of $f$ on the state she is given to confirm that it has the value $v$. Then she verifies that the state is in a uniform superposition of states with the same value $v$ for $f$. It is not obvious in general how to verify this uniform superposition and in our scheme we will see that this is one of the main challenges.
For such a scheme to work, the following properties are required:
- Measuring $f$ twice on two unentangled copies of the initial state should give different values with probability exponentially close to 1. Otherwise, the forger can just repeat the procedure performed by the mint, and obtain a money state with non negligible probability.
- For “most” values $v$, there should be exponentially many initial states $e$ which have $f(e)=v$. Moreover, given one state $|e\rangle$ with $f(e)=v$ it should be hard to generate the uniform superposition over all such states.
- It should be possible for the verifier to verify that he has a uniform superposition over states with the same value of $f$.
Quantum money from knots
------------------------
To implement this kind of quantum money, we use ideas from knot theory. The initial state generated by the mint is a uniform superposition over planar grid diagrams, which are compact representations of oriented links (an oriented knot is a knot with a preferred direction around it, and an oriented link is a collection of possibly intertwined oriented knots). The function $f$ which the mint measures is the Alexander polynomial of the oriented link represented by a given planar grid diagram. Finally, the verification procedure is based on Reidemeister moves, which do not change the value of the polynomial. (Reidemeister moves turn a knot into another equivalent knot, and the Alexander polynomials for equivalent knots are the same.)
Knots, links, and grid diagrams
===============================
In this section we briefly review the standard concepts of knots and links and how they are represented using diagrams. The same knot (or link) can be represented as a diagram in many different ways; the Reidemeister moves are a set of local changes to a diagram that do not change the underlying knot or link. We will review how to compute the Alexander polynomial for a given oriented link diagram in polynomial time. The Alexander polynomial is a link invariant in the sense that if we compute its value on diagrams which represent the same oriented link we get the same polynomial. Finally, we review planar grid diagrams and grid moves, which implement Reidemeister moves on grid diagrams.
Knots and links\[sub:Knots-and-Links\]
--------------------------------------
We can think of a knot as a loop of string in 3 dimensions, that is, a map from $S^{1}$ into $\mathbb{R}^{3}$. Since it is hard to draw knots in 3 dimensions, usually a knot is represented by a projection of the 3 dimensional object into 2 dimensions where, at each crossing, it is indicated which strand passes above and which below. This is called a knot diagram. In this paper we will be interested in links, which correspond to one or more loops of string (called the components of the link). An oriented link is a link which has a direction associated with each component. An example of an oriented link diagram is shown in figure \[Flo:linkdiagram\].
\#1\#2
\#1; \#2
Two links (or knots) are equivalent if one can be smoothly morphed into the other without cutting the string. If unoriented links $K_{1}$ and $K_{2}$ are equivalent and they are represented by diagrams $D_{1}$ and $D_{2}$ respectively, then diagram $D_{1}$ can be transformed into $D_{2}$ (and vice versa) using the Reidemeister moves pictured in figure \[Flo:Reidemeister\]. (For oriented links the Reidemeister moves can be drawn with the same pictures as in figure \[Flo:Reidemeister\] but with orientations along the edges which have to be consistent before and after applying the move). Looking at these moves, one sees that if two diagrams can be brought into the same form by using the moves then the diagrams represent equivalent links. The converse of this statement is a theorem.
(0,0.86603) arc (60:-60:1cm); (1.5,0.86603) arc (120:240:1cm); (2,0) – (3,0);
(1,-1) rectangle (5,1) (3,0) circle (1.1cm) circle (0.9cm); (4.2,0.86603) arc (120:240:1cm);
(3.5,0.86603) arc (60:-60:1cm);
(0,0.86603) arc (60:-60:1cm);
(1,0) – (2,0);
(0.0000,0.8660) .. controls (0.1527,0.7691) and (0.2824,0.6374) .. (0.3573,0.4993) .. controls (0.4321,0.3616) and (0.4398,-0.0042) .. (0.5584,-0.1603) .. controls (0.6769,-0.3164) and (0.8713,-0.2029) .. (0.8713,0.0000);
(-1,-1) rectangle (1,1) (0.4688,-0.2188) .. controls (0.3874,-0.1116) and (0.3703,0.0256) .. (0.3438,0.1562) .. controls (0.3172,0.2869) and (0.3016,0.4000) .. (0.2812,0.4375) – (0.4375,0.5625) .. controls (0.4920,0.4623) and (0.5057,0.3132) .. (0.5312,0.1875) .. controls (0.5568,0.0618) and (0.5879,-0.0448) .. (0.6250,-0.0938) – (0.4688,-0.2188) – cycle; (0.8713,0.0000) .. controls (0.8713,0.2029) and (0.6769,0.3164) .. (0.5584,0.1603) .. controls (0.4398,0.0042) and (0.4321,-0.3616) .. (0.3573,-0.4993) .. controls (0.2821,-0.6374) and (0.1527,-0.7601) .. (0.0000,-0.8660);
\#1\#2
\#1; \#2
(0,0.86603) – (1,-0.86603);
(2.25,0) – (3.25,0);
\#1\#2
\#1; \#2
(0,0.86603) – (1,-0.86603);
The Alexander polynomial of an oriented link\[sub:The-Alexander-polynomial\]
----------------------------------------------------------------------------
The Alexander polynomial is a polynomial $\Delta(x)$ which can be computed from a given oriented link diagram and which is invariant under the Reidemeister moves. In this section, following Alexander [@1928], we describe how to compute this polynomial. It will be clear from our discussion that the computation of $\Delta(x)$ can be done in polynomial time in the number of crossings of the diagram.
Suppose we are given a diagram of an oriented link $L$. If the diagram is disconnected, apply the first Reidemeister move in Figure \[Flo:Reidemeister\] to connect the diagram. Let us write $a$ for the number of crosses in the diagram. The curve of the diagram then divides the two dimensional plane into $a+2$ regions including one infinite region (this follows from Euler’s formula). The following recipe can be used to calculate $\Delta(x)$:
1. For each region $i\in\{1,...,a+2\}$, associate a variable $r_{i}$.
2. For each of the $a$ crossings, write down an equation$$xr_{j}-xr_{k}+r_{l}-r_{m}=0,$$ where $\{r_{j},r_{k},r_{l},r_{m}\}$ are the variables associated with the regions adjacent to the crossing, in the order pictured in figure \[Flo:labelingregions\].
3. Write the set of equations as a matrix equation $$\mathcal{M}\left(\begin{array}{c}
r_{1}\\
r_{2}\\
\vdots\\
r_{m+2}\end{array}\right)=0.$$ This defines the matrix $\mathcal{M}$ which has $a$ rows and $a+2$ columns. The entries of $\mathcal{M}$ are elements of the set $\{\pm1,\pm x,1+x,1-x,-1+x,-1-x\}.$
4. Delete two columns of $\mathcal{M}$ which correspond to adjacent regions in the link diagram. This gives an $a\times a$ matrix $\mathcal{M}_{0}$.
5. Take the determinant of $\mathcal{M}_{0}$. This is a polynomial in $x$. Divide this polynomial by the factor $\pm x^{q}$ chosen to make the lowest degree term a positive constant. The resulting polynomial is the Alexander polynomial $\Delta\left(x\right)$ for the given link.
When we use the Alexander polynomial in this paper, we are referring to the list of coefficients, not the value of the polynomial evaluated at some $x$.
(-1,0) – (-0.1,0); (0.1,0) – (1,0); (0,-1) – (0,1); at (-0.5,0.5) [$r_k$]{}; at (0.5,0.5) [$r_j$]{}; at (-0.5,-0.5) [$r_l$]{}; at (0.5,-0.5) [$r_m$]{};
(0,-1) – (0,-0.1); (0,0.1) – (0,1); (-1,0) – (1,0); at (-0.5,0.5) [$r_j$]{}; at (0.5,0.5) [$r_m$]{}; at (-0.5,-0.5) [$r_k$]{}; at (0.5,-0.5) [$r_l$]{};
Grid diagrams\[sub:Grid-diagrams\]
----------------------------------
It is convenient to represent knots as *planar grid diagrams*. A planar grid diagram is a $d\times d$ grid on which we draw $d$ X’s and $d$ O’s. There must be exactly one X and one O in each row and in each column, and there may never be an X and an O in the same cell. We draw a horizontal line in each row between the O and the X in that row and a vertical line in each column between the X and the O in that column. Where horizontal lines and vertical lines intersect, the vertical line always crosses above the horizontal line.
Knots (or links) in a grid diagram carry an implicit orientation: each vertical edge goes from an X to an O, and each horizontal edge goes from an O to an X. Figure \[fig:simple-grid-diag\] shows an example of a $d=4$ planar grid diagram.
A planar grid diagram $G$ can be specified by two disjoint permutations $\pi_{X},\pi_{O}\in S_{d}$, in which case the X’s have coordinates $\{(i,\pi_{X}(i))\}$ and the O’s have coordinates $\{(i,\pi_{O}(i))\}$ for $i\in\{1,...,d\}$. Two permutations are said to be disjoint if for all $i$, $\pi_{X}(i)\neq\pi_{O}(i)$. Any two disjoint permutations $\pi_{X},\pi_{O}\in S_{d}$ thus define a planar grid diagram $G=(\pi_{X},\pi_{O})$. Every link can be represented by many different grid diagrams.
We can define grid moves, which are three types of moves (that is, transformations) on planar grid diagrams which, like the Reidemeister moves for link diagrams, are sufficient to generate all planar grid diagrams of the same oriented link.
- The first type of move is a cyclic permutation of either the rows or the columns. Figure \[fig:grid-cyclic-perm\] shows an example of this type of move on columns. We can think of these moves as grabbing both markers in the rightmost column, pulling them behind of the page, and putting them back down on the left. These moves are always legal. There are equivalent moves on rows.
- The second type of move is transposition of two adjacent rows or columns. This can be done only when no strand would be broken. In Figure \[fig:grid-transposition\] we show examples of when this move is allowed. The legality of this move depends only on the position of the markers in the rows or columns being transposed.[^1]
- The third type of move adds one row and one column (this is called stabilization) or deletes one row and one column (this is called destabilization), as shown in Figure \[fig:grid-stabilization\]. Destabilization selects three markers forming an “L” shape with sides of length 1 and collapses them into one. That is, it deletes one row and one column such that all three markers are removed. The inverse move selects any marker, adds a row and a column adjacent to that marker, and replaces that marker with three new markers. Any X or O can always be legally stabilized and any three markers forming an “L” shape with sides of length 1 can be destabilized unless they form a box (i.e. a $2\times2$ square with a marker in all four positions).
In the remainder of this paper we will represent links exclusively with planar grid diagrams.
(4,4) (0, 0) grid (4, 4); (0,$0$) (1,$1$) (2,$2$) (3,$3$) (0,$0$) (1,$1$) (2,$2$) (3,$3$) (0,0,2) (1,1,3) (2,0,2) (3,1,3) (0,1) (1,2) (2,1) (3,2) (1,2) (2,3) (1,0) (2,1) (0,2,0) (1,3,3) (0,0) (1,1) (2,2) (3,3) (0,2) (2,0) (1,3) (3,1)
(5,6)
(0,0, 5,6) (0,1, 1,5) (4,1, 5,5) (0, 0) grid (5, 6);
(0, 1) grid (1, 5); (4, 1) grid (5, 5);
(0,$0$) (4,$d-1$)
(0,3,1) (0,3,3) (1,4,2) (1,4,4) (4,2,4) (4,3) (0,1,3) (0,2) (0,1) (0,3) (4,2) (4,4)
(5.5,3) – (6.5,3);
(0,0, 5,6) (0,1, 2,5) (0, 0) grid (5, 6);
(0, 1) grid (2, 5);
(0,$0$) (1,$1$)
(1,4,1) (1,4,3) (2,0,4)
(0,0) rectangle (5,6) (1-0.1,2-0.1) rectangle (1+0.1,2+0.1); (2,0,2)
(0,2,4) (0,3) (1,1,3) (1,2) (1,1) (1,3) (0,2) (0,4)
(4,6)
(0,0, 4,6) (1,1, 3,5, 0,0, 4,6) (0,0) grid (4,6);
(1,1) grid (3,5);
(1,$e$) (2,$e+1$)
(0,1,1) (0,1,4) (1,1,4) (1,2) (2,2,3) (2,2) (1,4) (1,1) (2,2) (2,3)
(4.5,3) – (5.5,3);
(4,6)
(0,0, 4,6) (1,1, 3,5, 0,0, 4,6) (0,0) grid (4,6);
(1,1) grid (3,5);
(1,$e$) (2,$e+1$)
(0,2,1) (0,2,4) (2,1,4) (2,2) (1,2,3) (1,2) (0,1,3) (0,1,2) (2,4) (2,1) (1,2) (1,3)
(4,6)
(0,0, 4,6) (1,1, 3,5, 0,0, 4,6) (0,0) grid (4,6);
(1,1) grid (3,5);
(1,$e$) (2,$e+1$)
(0,1,1) (1,1,3) (1,2) (2,2,4) (2,3) (2,3,2) (0,2,4) (1,3) (1,1) (2,2) (2,4)
(4.5,3) – (5.5,3);
(0,0, 4,6) (1,1, 3,5, 0,0, 4,6) (0,0) grid (4,6);
(1,1) grid (3,5);
(1,$e$) (2,$e+1$)
(0,2,1) (2,1,3) (2,2) (1,2,4) (1,3) (0,1,4) (2,3,3) (2,3) (2,1) (1,2) (1,4)
(-0.5,0) – (10.5,6) (10.5,0) – (-0.5,6);
(4,6)
(0,0, 4,6) (1,1, 3,5, 0,0, 4,6) (0,0) grid (4,6);
(1,1) grid (3,5);
(1,$e$) (2,$e+1$)
(0,1,1) (1,1,2) (1,2) (2,3,4) (2,3) (1,3,2) (0,2,4) (2,3,3) (1,2) (1,1) (2,3) (2,4)
(4.5,3) – (5.5,3);
(4,6)
(0,0, 4,6) (1,1, 3,5, 0,0, 4,6) (0,0) grid (4,6);
(1,1) grid (3,5);
(1,$e$) (2,$e+1$)
(0,2,1) (2,1,2) (2,2) (1,3,4) (1,3) (2,3,2) (0,1,4) (1,3,3) (2,2) (2,1) (1,3) (1,4)
(4,4)
(0,0, 4,4) (1,1, 3,3) (0,0) grid (4,4);
(1,1) grid (3,3);
(1,$x$) (2,$x+1$) (1,$y$) (2,$y+1$)
(1,0,2) (1,2,2) (2,2) (2,1,2) (2,1) (1,2) (2,1) (2,2)
(4.5,1.5) – (5.5,1.5);
(3,3)
(0,0, 3,3) (1,1, 2,2) (0,0) grid (3,3);
(1,1) grid (2,2);
(1,$x$) (1,$y$)
(1,0,1) (0,1,1) (1,1)
(4,4)
(0,0, 4,4) (1,1, 3,3) (0,0) grid (4,4);
(1,1) grid (3,3);
(1,$x$) (2,$x+1$) (1,$y$) (2,$y+1$)
(1,2,3) (1,2,2) (2,2) (2,1,2) (2,1) (0,2,1) (1,2) (2,1) (2,2)
(4.5,1.5) – (5.5,1.5);
(3,3)
(0,0, 3,3) (1,1, 2,2) (0,0) grid (3,3);
(1,1) grid (2,2);
(1,$x$) (1,$y$)
(1,1,2) (0,1,1) (1,1)
Quantum money\[sec:Quantum-money\]
==================================
In this section we describe our quantum money scheme in full detail. In section \[sub:Minting-Quantum-Money\] we describe how the mint can make quantum money efficiently. In section \[sub:Verifying-quantum-money\] we describe the verification procedure which can be used by anyone (that is, anyone who owns a quantum computer) to ensure that a given quantum bill is legitimate. In section \[sub:Discussion\] we discuss the security of the verification procedure. In section \[sub:Why-not-Quantum\] we contrast our quantum money scheme based on knots with a similar but insecure scheme based on graphs.
Minting quantum money\[sub:Minting-Quantum-Money\]
--------------------------------------------------
The basis vectors of our Hilbert space are grid diagrams $|G\rangle=|\pi_{X},\pi_{O}\rangle$, where each grid diagram $G$ is encoded by the disjoint permutations $\pi_{X},\pi_{O}$. The dimension $d(G)$ is the number of elements in the permutations.[^2] The number of disjoint pairs of permutations on $d$ elements is equal to $d!$ times the number of permutations that have no fixed points. The latter quantity is called the number of derangements on $d$ elements, and it is given by $\left[\frac{d!}{e}\right]$ where the brackets indicate the nearest integer function.
To generate quantum money, the mint first chooses a security parameter $\bar{D}$ and defines an unnormalized distribution $$y\left(d\right)=\begin{cases}
\frac{1}{d!\left[\frac{d!}{e}\right]}\exp\frac{-\left(d-\bar{D}\right)^{2}}{2\bar{D}} & \text{ if }2\le d\le2\bar{D}\\
0 & \text{ otherwise}.\end{cases}$$ Define an integer valued function $$q(d)=\lceil\frac{y(d)}{y_{min}}\rceil,$$ where $\lceil\cdot\rceil$ means round up and where $y_{min}$ is the minimum value of $y(d)$ for $d\in{2,...,2\bar{D}}$ (we need an integer valued distribution in what follows). The mint then uses the algorithm in [@grover-2002] to prepare the state (up to normalization) $$\sum_{d=2}^{2\bar{D}}d!\sqrt{q(d)}|d\rangle.$$ Using a straightforward unitary transformation acting on this state and an additional register, the mint produces $$\sum_{d=2}^{2\bar{D}}d!\sqrt{q\left(d\right)}|d\rangle\bigg(\frac{1}{d!}\sum_{\pi_{X},\pi_{O}\in S_{d}}|\pi_{X},\pi_{O}\rangle\bigg),$$ and then measures whether or not the two permutations $\pi_{X}$ and $\pi_{O}$ are disjoint. (They are disjoint with probability very close to $\nicefrac{1}{e}$.) If the mint obtains the measurement outcome “disjoint”, it uncomputes the dimension $d$ in the first register to obtain the state $|\text{initial}\rangle$ on the last two registers, where $$|\text{initial}\rangle=\frac{1}{\sqrt{N}}\sum_{\text{grid diagrams }G}\sqrt{q\left(d\left(G\right)\right)}|G\rangle\label{eq:initialstate}$$ and $N$ is a normalizing constant. If the measurement says “not distinct”, the mint starts over.
The distribution $q\left(d\right)$ is chosen so that if one were to measure $d(G)$ on $|\text{initial}\rangle$, then the distribution of results would be extremely close to Gaussian with mean $\bar{D}$ restricted to integers in the interval $\left[2,2\bar{D}\right]$. As $\bar{D}$ becomes large, the missing weight in the tails falls rapidly to zero.
From the state $|\text{initial}\rangle$, the mint computes the Alexander polynomial $A(G)$ into another register and then measures this register, obtaining the polynomial $p$. The resulting state is $|\text{\$}_{p}\rangle$, the weighted superposition of all grid diagrams (up to size $2\bar{D}$) with Alexander polynomial $p$: $$|\text{\$}_{p}\rangle=\frac{1}{\sqrt{N}}\sum_{G:A\left(G\right)=p}\sqrt{q(d(G))}|G\rangle,\label{eq:moneystate}$$ where $N$ takes care of the normalization. The quantum money consists of the state $|\$_{p}\rangle$, and the serial number is the polynomial $p$, represented by its list of coefficients. If the polynomial $p$ is zero, the mint should reject that state and start over. Splittable links have Alexander polynomial equal to zero, and we wish to avoid this case.
Verifying quantum money\[sub:Verifying-quantum-money\]
------------------------------------------------------
Suppose you are a merchant. Someone hands you a quantum state $|\phi\rangle$ and a serial number which corresponds to a polynomial $p$ and claims that this is good quantum money. To check that indeed this is the case, you would use the following algorithm.
$\hphantom{}$
1. Verify that $|\phi\rangle$ is a superposition of basis vectors that validly encode grid diagrams. If this is the case then move on to step \[enu:Measure-the-Alexander\], otherwise reject.
2. \[enu:Measure-the-Alexander\]Measure the Alexander polynomial on the state $|\phi\rangle$. If this is measured to be $p$ then continue on to step \[enu:Measure-the-projector\]. Otherwise, reject.
3. \[enu:Measure-the-projector\]Measure the projector onto grid diagrams with dimensions in the range $[\frac{\bar{D}}{2},\frac{3\bar{D}}{2}].$ If you obtain $+1$ then continue on to step \[enu:Apply-the-Markov\]. Otherwise, reject. For valid money, you will obtain the outcome $+1$ with high probability and cause little damage to the state. (We discuss the rationale for this step in section \[sub:Discussion\].)
4. \[enu:Apply-the-Markov\]Apply the Markov chain verification algorithm described in section \[sub:A-Quantum-Verification\]. If the state passes this step, accept the state. Otherwise, reject.
If the money passes steps 0 and \[enu:Measure-the-Alexander\] then $|\phi\rangle$ is a superposition of grid diagrams with the correct Alexander polynomial $p$. Now passing steps \[enu:Measure-the-projector\] and \[enu:Apply-the-Markov\] will (approximately) confirm that $|\phi\rangle$ is the *correct* superposition of grid diagrams. This procedure will accept genuine quantum money states with high probability, but, as we discuss below, there will be other states that also pass verification. We believe that these states are hard to manufacture.
We now discuss the quantum verification scheme used in step \[enu:Apply-the-Markov\]. We begin by defining a classical Markov chain.
### A classical markov chain\[sub:A-Markov-Chain\]
The Markov chain is chosen to have uniform limiting distribution over pairs $(G,i)$ where $G$ is equivalent to (and reachable without exceeding grid dimension $2\bar{D}$ from) the starting configuration, and where $i\in\{1,...,q(d(G))\}.$ Therefore, in the limiting distribution (starting from a grid diagram $\tilde{G}$) the probability of finding a grid diagram $G$ (which is equivalent to $\tilde{G}$) is proportional to $q(d(G)).$ We use the extra label $i$ in our implementation of the verifier.
There are two types of update rules for the Markov chain. The first type is an update that changes $i$ while leaving $G$ unchanged (the new value of $i$ is chosen uniformly). The second type is an update that changes $G$ to a new diagram $G^{\prime}$ while leaving $i$ alone (this type of update is only allowed if $i\leq q(d(G^{\prime}))$). For the moves which leave $i$ unchanged, our Markov chain selects a grid move at random and proposes to update the current grid diagram by applying that move. The move $G\rightarrow G^{\prime}$ is accepted if the value $i\leq q(d(G^{\prime})).$
Recall (from section \[sub:Grid-diagrams\]) that there is a set of grid moves on planar grid diagrams which can be used to transition from one grid diagram to another. These moves only allow transitions between grid diagrams which represent equivalent links, and some of these moves change the dimension of the grid diagram. In general, any two grid diagrams representing equivalent links can be connected by a sequence of these moves (although finding this sequence is not easy). However, this sequence of moves may also pass through grid diagrams which have dimension greater than $2\bar{D}$. In this case, the two equivalent diagrams will not mix, due to the cutoff we defined.
In the appendix we define this Markov chain in detail. This Markov chain does not mix between nonequivalent grid diagrams. As such, it has many stationary distributions. In the next section, we use this Markov chain to produce a quantum verification procedure.
### A quantum verification procedure based on a classical Markov chain \[sub:A-Quantum-Verification\]
Let $B$ denote the Markov matrix. As shown in the appendix, $B$ is a doubly stochastic (the sum of the elements in each row and each column is $1$) matrix which can be written as $$B=\frac{1}{\left|\mathcal{S}\right|}\sum_{s\in\mathcal{S}}P_{s},\label{eq:B_matrix}$$ where each $P_{s}$ is a permutation on the configuration space of pairs $(G,i)$ such that $i\leq q(d(G))$, and $\mathcal{S}$ is a set of moves that the Markov chain can make at each iteration. We can view the same matrices as quantum operators acting on states in a Hilbert space where basis vectors are of the form $|G\rangle|i\rangle$. For clarity we will refer to the quantum operators as $\hat{B}$ and $\hat{P}_{s}$.
Our quantum money states $|\$_{p}\rangle$ live in a Hilbert space where basis vectors are grid diagrams $|G\rangle.$ To enable our verification procedure, we define a related state $|\$_{p}^{\prime}\rangle$ on two registers, where the second register holds integers $i\in\{0,...,q_{max}\}$ (here $q_{max}$ is the maximum that $q(d)$ can reach). $|\$'_{p}\rangle$ is unitarily related to $|\$_{p}\rangle$, so verifying the former is equivalent to verifying the latter. Define a unitary $U$ which acts on basis vectors in the expanded Hilbert space as$$U\left(|G\rangle|0\rangle\right)=|G\rangle\frac{1}{\sqrt{q(d(G))}}\sum_{i=1}^{q(d(G))}|i\rangle.$$ To obtain $|\$_{p}^{\prime}\rangle,$ take the quantum money state $|\$_{p}\rangle$ and adjoin the ancilla register initialized in the state $|0\rangle$. Then apply $U$ to both registers to produce the state $$\begin{aligned}
|\$_{p}^{\prime}\rangle & = & U\left(|\$_{p}\rangle|0\rangle\right)\\
& = & \frac{1}{\sqrt{N}}\sum_{G:A(G)=p}\sum_{i=1}^{q(d(G))}|G\rangle|i\rangle.\end{aligned}$$ Note that whereas $|\$_{p}\rangle$ is a weighted superposition in the original Hilbert space (see equation \[eq:moneystate\]), $|\$_{p}^{\prime}\rangle$ is a uniform superposition in the expanded Hilbert space. However both $|\$_{p}\rangle$ and $|\$_{p}^{\prime}\rangle$ give rise to the same probability distribution over grid diagrams $G$.
The subroutine in step \[enu:Apply-the-Markov\] of the verification procedure of Section \[sub:Verifying-quantum-money\] starts with the input state $|\phi\rangle$ and applies the following algorithm.[^3]
$\hphantom{}$
1. \[enu:Prepare-the-state\]Take the input state $|\phi\rangle$, append an ancilla register which holds integers $i\in\{0,...,q_{max}\}$ and prepare the state $$|\phi^{\prime}\rangle=U\left(|\phi\rangle|0\rangle\right)$$ using the unitary $U$ defined above.
2. \[enu:Adjoin-two-ancilla\]Adjoin an ancilla register which holds integers $s\in\{1,...|\mathcal{S}|\}$ and initialize the ancilla register to the state $\sum_{s}\frac{1}{\sqrt{\left|\mathcal{S}\right|}}|s\rangle$. This produces the state $$|\phi^{\prime}\rangle\sum_{s}\frac{1}{\sqrt{\left|\mathcal{S}\right|}}|s\rangle$$ on three registers (one which holds grid diagrams $G$, one which holds integers $i\in\{0,...,q_{max}\}$ and one which holds integers $s\in\{1,...,|\mathcal{S}|\}).$
3. \[enu:Now-apply-the\]Repeat $r=\mathop\mathrm{poly}(\bar{D})$ times:
1. \[enu:Apply-the-unitary\]Apply the unitary $V$, where $$V=\sum_{s}\hat{P}_{s}\otimes|s\rangle\langle s|.$$ This operator applies the permutation $\hat{P}_{s}$ to the first two registers, conditioned on the value $s$ in the third register.
2. \[enu:step\_b\]Measure the projector $$Q=I\otimes I\otimes\left(\sum_{s,s^{\prime}}\frac{1}{\left|\mathcal{S}\right|}|s\rangle\langle s^{\prime}|\right).$$
4. \[Accept/reject \]If you obtained the outcome 1 in each of the $r$ repetitions of step \[enu:Now-apply-the\], then accept the money. In this case apply $U^{\dagger}$ and then remove the second and third registers. The final state of the first register is the output quantum money state. If you did not obtain the outcome $1$ in each of the $r$ iterations in step \[enu:Now-apply-the\], then reject the money.
For a quantum money state $|\$_{p}\rangle$ the associated state $|\$_{p}^{\prime}\rangle\sum_{s}\frac{1}{\sqrt{\left|\mathcal{S}\right|}}|s\rangle$ prepared in steps \[enu:Prepare-the-state\] and \[enu:Adjoin-two-ancilla\] is unchanged by applying the unitary $V$ in step \[enu:Apply-the-unitary\]. Measuring the projector $Q$ in step \[enu:step\_b\] on this state gives $+1$ with certainty. So for good quantum money, all the measurement outcomes obtained in step \[Accept/reject \] are $+1$. We can see that good quantum money passes verification.
Now let us consider the result of applying the above procedure to a general state $|\phi\rangle$. The first step of the algorithm is to prepare $|\phi'\rangle=U\left(|\phi\rangle|0\rangle\right)$. If a single iteration of the loop in step \[enu:Now-apply-the\] results in the outcome 1 then the final state of the first two registers is $$\frac{\frac{1}{\left|\mathcal{S}\right|}\sum_{s}\hat{P}_{s}|\phi^{\prime}\rangle}{\left\Vert \frac{1}{\left|\mathcal{S}\right|}\sum_{s,t}\hat{P}_{s}|\phi^{\prime}\rangle\right\Vert }.$$ This occurs with probability $$\left\Vert \frac{1}{\left|\mathcal{S}\right|}\sum_{s}\hat{P}_{s}|\phi^{\prime}\rangle\right\Vert ^{2}.\label{eq:prob}$$ The entire procedure repeats the loop $r$ times and the probability of obtaining all outcomes $1$ is $$\left\Vert \left(\frac{1}{\left|\mathcal{S}\right|}\sum_{s}\hat{P}_{s}\right)^{r}|\phi^{\prime}\rangle\right\Vert ^{2},$$ in which case the final state (on the first two registers) is $$\frac{\left(\frac{1}{\left|\mathcal{S}\right|}\sum_{s}\hat{P}_{s}\right)^{r}|\phi^{\prime}\rangle}{\left\Vert \left(\frac{1}{\left|\mathcal{S}\right|}\sum_{s}\hat{P}_{s}\right)^{r}|\phi^{\prime}\rangle\right\Vert }.$$
To get some intuition about what states might pass even the first iteration of this test with reasonable probability (\[eq:prob\]), note that the state $\frac{1}{|\mathcal{S}|}\sum_{s}\hat{P}_{s}|\phi^{\prime}\rangle$ can only have norm close to 1 if most of its terms add coherently. In other words, most of the $\hat{P}_{s}|\phi^{\prime}\rangle$ must be nearly the same (they are all exactly the same for good quantum money).
Since $\frac{1}{\left|\mathcal{S}\right|}\sum_{s}\hat{P}_{s}=\hat{B}$ is our Markov matrix, the set of states that pass all of the rounds is directly related to the mixing properties of the Markov chain—these states correspond to eigenvectors of $\hat{B}$ with eigenvalues close to 1.
Security of the money scheme\[sub:Discussion\]
----------------------------------------------
We have shown that the quantum money states $|\$_{p}\rangle$ can be efficiently generated and pass verification with high probability. In this section we discuss why we believe this quantum money is hard to forge. We consider four possible types of attack against our quantum money.
First, an attacker might measure a valid quantum money state $|\$_{p}\rangle$ to learn a grid diagram with Alexander polynomial $p$ and then generate a superposition containing that diagram that passes verification. One such state is the (correctly weighted) superposition over grid diagrams equivalent to the measured diagram. If an attacker could do this, the attacker’s algorithm could be used to solve grid diagram equivalence, i.e. the problem of telling whether or not two grid diagrams represent the same link. This is believed to be a hard problem on average, even for quantum computers. In fact, even deciding whether or not a grid diagram represents the unknot is conjectured to be hard.
Second, there are likely to exist grid diagrams of dimension $2\bar{D}$ (our maximum grid dimension) where no allowed grid move reduces the dimension (this follows from the fact that every link has a minimum dimension of grid diagrams that represent it). Because we disallow moves which increase the dimension above $2\bar{D}$, the Markov chain starting from one of these grid diagrams with dimension $2\bar{D}$ will only mix over a small set of diagrams with the same dimension. Uniform superpositions over these sets will pass step \[enu:Apply-the-Markov\] of the verification procedure. Step \[enu:Measure-the-projector\] of the verification procedure is designed to reject such superpositions.
Third, if the Markov chain does not mix well, then there will be eigenstates of the matrix $\hat{B}$ with eigenvalues near +1. For such eigenstates there may be counterfeit quantum money states which will pass verification with nonnegligible probability. We do not know if these states exist, and even if they do we do not know how to make them. In fact even the simpler question, of finding two equivalent diagrams which require a super-polynomial number of grid moves to go from one to the other (or proving such diagrams do not exist) seems hard.[^4]
Fourth, the attacker could use a valid money state $|\$_{p}\rangle$ and attempt to generate $|\$_{p}\rangle\otimes|\$_{p}\rangle$ (or some entangled state on two registers where each register would pass verification). Such an attack worked to forge product state quantum money using quantum state restoration [@state-restoration]. However, in that case the attack works because the product state money has an embedded classical secret that the attacker can learn to forge the state. In the present work, there is no obvious secret hidden in the money or in the verification algorithm that an attacker could use.
The list above comprises all the lines of attack we were able to think of.
Why not quantum money from graphs?\[sub:Why-not-Quantum\]
---------------------------------------------------------
We now discuss an alternative quantum money scheme that is similar to the one presented in this paper, but which is based on graphs rather than knots. In practice, it is easy to find the isomorphism relating most pairs of isomorphic graphs; this is enough to render graph based quantum money unusable. The knot based quantum money is an analog of this scheme (translating various concepts from graphs to knots) but we believe it is secure since knot equivalence is believed to be difficult on average.
In the graph based quantum money scheme, the Hilbert space has basis vectors which encode adjacency matrices of graphs on $n$ vertices. The mint generates the state$$\sum_{\text{Adjacency matrices \ensuremath{A}}}|A\rangle|0\rangle.$$ The mint then computes the eigenvalues of the adjacency matrix into the second register and measures the second register to obtain a particular spectrum $R=\left\{ \lambda_{1},...\lambda_{n}\right\} .$ The resulting state is $$|\$_{R}\rangle|R\rangle=\sum_{A\text{ with spectrum }R}|A\rangle|R\rangle.$$ The quantum money is the state $|\$_{R}\rangle$ and the serial number encodes the spectrum $R$. The verification procedure is based on a classical Markov chain that, starting from a given adjacency matrix $A$, mixes to the uniform distribution over adjacency matrices $A^{\prime}$ that represent graphs isomorphic to $A$.
Given two adjacency matrices $A_{0}$ and $\sigma A_{0}\sigma^{-1}$ that are related by the permutation $\sigma\in S_{n}$, the forger can usually efficiently find the permutation $\sigma$ (we assume for simplicity that $A_{0}$ has no automorphisms so this permutation is unique). We now show how this allows a counterfeiter to forge the graph based quantum money. The forger first measures the state $|\$_{R}\rangle$ in the computational basis, obtaining a particular adjacency matrix $A$ with the spectrum $R$. Starting from this state, the forger adjoins two additional registers (one which holds permutations and one which holds adjacency matrices) and, applying a unitary transformation, the forger can produce the state
$$\sum_{\pi\in S_{n}}|A\rangle|\pi\rangle|\pi A\pi^{-1}\rangle.$$ Now the forger can use the procedure which solves graph isomorphism to uncompute the permutation which is held in the second register, producing the state $$|A\rangle|0\rangle\sum_{\pi\in S_{n}}|\pi A\pi^{-1}\rangle.$$ The state of the third register will pass verification. The forger can repeat this procedure using the same adjacency matrix $A$ to produce many copies of this state which pass the verification as quantum money.
Return to the knot based quantum money. Given two grid diagrams $G_{1}$ and $G_{2}$ which represent the same link, we believe it is hard (on average) to find a sequence of grid moves which transform $G_{1}$ into $G_{2}$. Even given an oracle for the decision problem of determining whether two links are equivalent, we do not know how to find a sequence of Reidemeister moves relating the links. We hope this discussion has motivated some of the choices we have made for our knot based quantum money.
Conclusions\[sec:Conclusions\]
==============================
Forge it if you can.
Acknowledgments
===============
We thank Louis Kauffman, Haynes Miller and Tom Mrowka for valuable discussions on knot theory. This work was supported in part by funds provided by the U.S. Department of Energy under cooperative research agreement DE-FG02-94ER40818, the W. M. Keck Foundation Center for Extreme Quantum Information Theory, the U.S. Army Research Laboratory’s Army Research Office through grant number W911NF-09-1-0438, the National Science Foundation through grant number CCF-0829421, the DoD through the NDSEG Fellowship Program, the Natural Sciences and Engineering Research Council of Canada, and Microsoft Research.
Details of the Markov chain on planar grid diagrams
===================================================
The Markov chain in section \[sub:A-Markov-Chain\] acts on the configuration space of pairs $(G,i)$ where $1\leq i\leq q(d(G))$. Here we describe an algorithm to implement this Markov chain. For each update of the state of the Markov chain, we first choose several uniformly random parameters: $$\begin{aligned}
j: & \quad\text{an integer from 1 to 8}\\
w: & \quad\text{an integer from 0 to \ensuremath{q_{max}^{2}}}\\
x,y: & \quad\text{integers from 0 to \ensuremath{2\bar{D}-1}}\\
k: & \quad\text{an integer from 0 to 3 }\end{aligned}$$ where $q_{max}$ is the maximum value attained by the function $q(d)$ for $d\in\{2,...,2\bar{D}\}.$ We then update the configuration $(G,i)$ in a manner that depends on the values of these variables, where $j$ picks the type of the update being performed, and the other variables are used as parameters:
- If $j=1$, set $$i\leftarrow\left(i+w\right)\text{ mod \ensuremath{q(d(G))}}$$ Leave $G$ unchanged. This is the only move which changes $i$, and it is used to mix between different labels.
- If $j=2$, cyclically permute the columns of $G$ by moving each column to the right by one. Leave $i$ unchanged.
- If $j=3$, cyclically permute columns of $G$ by moving each column to the left by one. Leave $i$ unchanged.
- If $j=4$, cyclically permute the rows of $G$ by moving each row up by one. Leave $i$ unchanged.
- If $j=5$, cyclically permute the rows of $G$ by moving each row down by one. Leave $i$ unchanged.
- If $j=6$ and $x+1<d\left(G\right)$, then check whether transposing columns $x$ and $x+1$ is a legal move (as defined in section \[sub:Grid-diagrams\]). If so, transpose them; otherwise, do nothing. Leave $i$ unchanged.
- If $j=7$ and $y+1<d\left(G\right)$, then check whether transposing rows $y$ and $y+1$ is a legal move (as defined in section \[sub:Grid-diagrams\]). If so, transpose them; otherwise, do nothing. Leave $i$ unchanged.
- If $j=8$, $k=0$, $x,y<d\left(G\right)$, and there is a marker (X or O) at position $\left(x,y\right)$, then consider stabilizing by adding markers forming an L to the upper right of position $(x,y)$. In this case, construct $G^{\prime}$ from $G$ by adding a new column to the right of $x$ and a new row above $y$, deleting the original marker, adding markers of the same type at $\left(x+1,y\right)$ and $\left(x,y+1\right)$, and adding a marker of the opposite type at $\left(x+1,y+1\right)$. Then if $i\leq q(d(G^{\prime})),$ set $(G,i)\leftarrow(G^{\prime},i)$. If not, leave the configuration unchanged.
- If $j=8$, $k=0,$ $x,y<d\left(G\right)-1$, there is no marker at $(x,y)$, and there are markers at positions $\left(x+1,y+1\right)$, $\left(x+1,y\right)$, $\left(x,y+1\right)$, then consider destabilizing by removing markers forming an L to the upper right of position $\left(x,y\right)$. In this case, construct $G^{\prime}$ from $G$ by removing column $x+1$ and row $y+1$ (thus removing those three markers) and adding a new marker of the appropriate type at $\left(x,y\right)$ (measured after deletion of the row and column). If $i\leq q(d(G^{\prime})),$ set $(G,i)\leftarrow(G^{\prime},i)$. If not, leave the configuration unchanged.
- If $j=8$, and $k\in{1,2,3}$, rotate the grid diagram by $90k$ degrees counterclockwise and then apply the update rule corresponding to $j=8,k=0$ with the same $(x,y)$. After applying the update, rotate the diagram back by the same amount.
The parameter $j$ determines the type of move. The first move ($j=1)$ is used to permute between different labels of the same graph. It’s the only move which changes the labels. Moves 2–5 are cyclic permutations of rows and columns, while moves 6 and 7 are transpositions. Finally, $j=8$ stabilizes or destabilizes an L in a direction ((1,0) – (1,1) – (0,1); (1,1) – (1,0) – (0,0); (0,1) – (0,0) – (1,0); or (0,0) – (0,1) – (1,1);) depending on $k$ (0, 1, 2, or 3 respectively). In all the moves $x$ is treated as column index (which describes where to apply the move), and $y$ is treated as a row index.
For fixed $\left(j,w,x,y,k\right)$, the update of the state is an easily computable, invertible function (in other words it is a permutation of the configuration space). This is easy to see for any of the moves with $j\in\{1,...,7\}$. One can check that each move with $j=8$ is its own inverse. This justifies our assertion that the Markov matrix $B$ can be written in the form of equation \[eq:B\_matrix\].
In the notation of Section \[sub:Verifying-quantum-money\], $s=\left(j,w,x,y,k\right)$ and $$\mathcal{S}=\{1,\ldots,8\}\times\{0,\ldots,q_{\text{max}}^{2}\}\times\{0,\ldots,2\bar{D}-1\}\times\{0,\ldots,2\bar{D}-1\}\times\{0,1,2,3\}.$$
[^1]: For the case of columns $i$ and $i+1$, the precise condition is that either the two intervals $\left[\min\left(x_{i},o_{i}\right),\max\left(x_{i},o_{i}\right)\right]$ and $\left[\min\left(x_{i+1},o_{i+1}\right),\max\left(x_{i+1},o_{i+1}\right)\right]$ do not overlap or one of the intervals contains the other.
[^2]: In practice, we will encode $d$, $\pi_{X}$ and $\pi_{O}$ as bit strings. Any such encoding will also contain extraneous bit strings that do not describe diagrams; the verification algorithm will ensure that these do not occur.
[^3]: Using a construction from [@aharonov-2003], it is possible to associate a Hamiltonian with a Markov chain such as ours. It may also be possible to construct a verifier using phase estimation on this Hamiltonian.
[^4]: Hass and Nowik [@hass-unknot] recently gave the first example of a family of knots that are equivalent to the unknot and require a quadratic number of Reidemeister moves to untangle (previous lower bounds were only linear).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Cotunneling is an important error process in the application of single electron tunneling devices for metrological and electronic applications. Here we present an experimental investigation of the theory for adiabatic enhancement of cotunneling by coherent microwaves. The dependence is investigated as function of temperature, gate voltage, frequency, and applied microwave power. At low temperatures and applied power levels, the results are consistent with theory, using only the unknown damping in the microwave line as a free parameter. However, the results indicate that the effects of temperature, frequency and microwave power are not independent, contrary to what is suggested by theory.'
author:
- Martin Manscher
- 'Marko T. Savolainen'
- Jesper Mygind
bibliography:
- 'PRB.bib'
title: Experimental Investigation of Microwave Enhanced Cotunneling in SET Transistors
---
Introdution\[sec:intro\]
========================
The unique properties of single electron tunneling (SET) devices has made them the subject of extensive research. This research includes applications such as current standards,[@Kautz] capacitance standards,[@Keller] electrometers,[@Aassime] quantum computing,[@Vion] and thermometry.[@Kauppinen; @Bergsten] The first-order behaviour is adequately described by the well-established [*orthodox theory*]{}.[@AverinLikharev; @IngoldNazarov] The so-called [*cotunneling*]{}, a second-order phenomenon in which one electron tunnels through each of the SET junctions at the same time, has also received some attention, as this constitutes an important error process in many SET applications. The dependence of the cotunneling current on voltage and temperature has been studied theoretically [@AverinNazarov] as well as experimentally.[@Geerlings; @Eiles; @Pasquier; @Paul] Furthermore, Covington et [*al.*]{}[@Covington] have studied the frequency dependence for 4- and 6-junction pumps. This paper investigates experimentally the theoretical prediction by Flensberg [@Flensberg] that the cotunneling current should depend not only on temperature and voltage, but also on the amplitude and frequency of an applied oscillating field.
The paper is organized as follows: The theory for cotunneling derived by Flensberg and others is briefly summarized in section \[sec:theory\]. The measurement setup is thoroughly described in section \[sec:setup\]. The experimental results on cotunneling are presented in section \[sec:results\], and in section \[sec:discussion\] the results are discussed. Finally, section \[sec:conclusion\] summarizes our conclusions.
Theory\[sec:theory\]
====================
[*Cotunneling*]{} (also referred to as [*macroscopic quantum tunneling of electric charge*]{} or ) is a second-order process by which an electron tunnels through the junctions of the SET transistor via an intermediate virtual state.[@AverinOdintsov; @AverinNazarov] The inelastic cotunneling current in a single SET transistor for low temperature ($k_\mathrm{B} T_\mathrm{e} \ll \Delta^\pm$) and bias voltage ($eV_\mathrm{DC} \ll \Delta^\pm$) was derived by Averin and Nazarov as[@AverinNazarov] $$\label{eq:DCcotun}
I_\mathrm{cot}^\mathrm{(3)} = \frac{R_\mathrm{K}}{24 \pi^2 R_1 R_2} \left( \frac{1}{\Delta^+} + \frac{1}{\Delta^-} \right)^2
\left[ \left(2 \pi k_\mathrm{B} T_\mathrm{e}\right)^2 + \left(eV_\mathrm{DC}\right)^2\right] V_\mathrm{DC}$$Here, $R_{1,2}$ are the tunneling resistances of the left and right electrodes, respectively, $\Delta^\pm = (e/C_\Sigma)(e/2 \mp C_\mathrm{G} V_\mathrm{G}) \; (\mathrm{mod}\;(e^2/C_\Sigma))$ are the energies to add/remove one electron to/from the island, $T_\mathrm{e}$ is the temperature of the electron system, and $R_\mathrm{K} = h/e^2$. In the maximum blockade state, $\Delta^\pm$ reduce to the charging energy $E_\mathrm{C} = e^2/2C_\Sigma$, where $C_\Sigma = C_1 + C_2 + C_\mathrm{G}$ is the total capacitance of the island (the ground capacitance is assumed negligible). The validity of Eq. \[eq:DCcotun\] has been verified experimentally in metallic systems by several researchers: Geerlings et [*al.*]{}[@Geerlings] first reported the observation of cotunneling current and its scaling with the conductances and voltage. The scaling of the current, rather than the quantitative value, made the results clearly distinguishable from thermally enhanced sequential tunneling. Later, the experiments by Eiles et [*al.*]{}[@Eiles] confirmed the temperature and gate bias dependence quantitatively. Furthermore, Pasquier et [*al.*]{}[@Pasquier] and other groups have investigated cotunneling in 2DEG systems. Since it seems well-confirmed by experiments, Eq. \[eq:DCcotun\] will be assumed valid here.
Flensberg [@Flensberg] has extended the analysis by Averin & Nazarov to the case where a harmonically varying signal $V_\mathrm{AC} \cos (2\pi f t)$ is applied on top of the DC bias. His result for the cotunneling current, obtained by expansion to the third order in the energies, becomes in the adiabatic (low-frequency, low-temperature, low-amplitude) limit $$\begin{gathered}
\label{eq:ACcotun3rdOrder}
I_\mathrm{cot}^\mathrm{(3)} = \frac{R_\mathrm{K}}{24 \pi^2 R_1 R_2} \left( \frac{1}{\Delta^+} + \frac{1}{\Delta^-} \right)^2 \\
\times \left[ \left(2 \pi k_\mathrm{B} T_\mathrm{e}\right)^2 + \left(eV_\mathrm{DC}\right)^2 + \tfrac{3}{2} \left(eV_\mathrm{AC}\right)^2 \right] V_\mathrm{DC}\end{gathered}$$ In restating his equations we assume (as is the case for our experiments) that the alternating bias is applied to the left lead, while the right is kept at a constant potential.
Eq. \[eq:ACcotun3rdOrder\] is not surprising, as this is what one would obtain from Eq. \[eq:DCcotun\] by making the substitution $V_\mathrm{DC} \rightarrow V_\mathrm{DC} + V_\mathrm{AC} \cos (2\pi f t)$ and averaging over time. However, because of the power expansion approach, Flensberg was able to derive also the next leading order correction to Eq. \[eq:ACcotun3rdOrder\] by an expansion to fourth order in the energies: $$\begin{gathered}
\label{eq:ACcotun4thOrder}
I_\mathrm{cot}^\mathrm{(4)} = \frac{R_\mathrm{K}}{48 \pi^2 R_1 R_2} \left( \frac{1}{\Delta^+} + \frac{1}{\Delta^-}\right) ^2 \left( \frac{1}{\Delta^-} - \frac{1}{\Delta^+}\right) \\
\times \left[ \left(2 \pi k_\mathrm{B} T_\mathrm{e}\right)^2 + \tfrac{3}{4} \left(eV_\mathrm{AC}\right)^2 + \left(h f\right)^2 \right] \left(eV_\mathrm{AC}\right)^2\end{gathered}$$ The temperature, amplitude and frequency are assumed low enough that an expansion in powers of these is appropriate. Note that in the case of maximum Coulomb blockade, $I_\mathrm{cot}^\mathrm{(4)}$ vanishes, since then $\Delta^\pm = E_\mathrm{C}$ and thus the third multiplicative term becomes zero.
The aim of the work on cotunneling presented here is to experimentally verify or disprove Eq. \[eq:ACcotun3rdOrder\]. We do this by applying microwaves to the device and measuring the differential conductance by a lock-in technique in zero DC bias. This ensures that the current predicted by Eq. \[eq:ACcotun4thOrder\] does not contribute to the result. Measurements to verify Eq. \[eq:ACcotun4thOrder\] are planned for the near future.
Using the familiar Master Equation approach for the orthodox theory, we have calculated the curve shown in Fig. \[fig:Simulation\] for a SET transistor with the parameters measured for the sample reported on here (JYU NM3). The figure shows the sequential tunneling and cotunneling conductances in zero bias and maximum Coulomb blockade, as function of temperature. It is seen that the conductances have the same order of magnitude, which means that sequential tunneling should be taken into account when interpreting the results.
Experimental Setup\[sec:setup\]
===============================
The samples are fabricated on 500 $\mu$m thick oxidized silicon. Gold pads and leads are made by a standard UV lithography technique. Before fabrication, the wafers are cleaned by ultrasound. A standard PMMA(-MAA) two-layer resist system is then applied and exposed by a pattern designed for 50 nm line with and 100 nm overlap. The line width becomes a little larger; 70-100 nm due to the exposure equipment and development. After exposure, the pattern is developed, and cleaned by reactive ion etching. Two layers of aluminium (40 and 30 nm thick) are evaporated at an angle, with an intermediate oxidation in 2 mbar O$_2$ for 3 minutes. After the last metal evaporation, the sample is oxidized again in 2 mbar O$_2$ for several minutes. Finally, the result is lifted off in acetone.
The samples were measured in the KelvinOx dilution refrigerator at the Institute of Physics, DTU. The plastic mixing chamber has been replaced by a metal one, and the sample is mounted on a cold finger extending into a superconducting magnet, which, in order to suppress the superconductivity of the aluminium, was set to one Tesla in all measurements presented here. The dilution refrigerator is set up in a shielded room, with the mechanical pumps outside the shielding. The temperature of the mixing chamber could be measured with calibrated germanium and ruthenium-oxide thermometers. The loaded base temperature is 50 mK. The samples were biased symmetrically using locally fabricated low-noise electronics. The electronics is mounted in an RF tight, compact metal box for shielding. For bias, an input amplifier adds two incoming signals; in this case a DC bias and a small low-frequency modulation provided by a lock-in amplifier. The added signal is then symmetrized by by an inverting and a non-inverting amplifier. Finally, the signal is voltage divided on each side to a suitable bias level $V_\mathrm{B}$, and fed to the sample through two large resistors $R_\mathrm{B}$. These were set to 10 M$\Omega$ in all measurements presented. A FET input, low-noise voltage preamplifier with a gain of 1000 was used to measure the resulting voltage $V_\mathrm{DC}$ across the sample, and the voltage signal is further amplified 10 times by an external PAR113 low-noise amplifier. The current through the sample is then calculated as $(V_\mathrm{B}-V_\mathrm{DC})/2R_\mathrm{B}$. A small-amplitude modulation, with frequency 2 Hz, was applied from the output of a SR850 lock-in amplifier on top of the DC bias to measure dynamic resistances. This results in a current modulation across the device of about 10 pA. After amplification in the FET preamplifier, the voltage response was fed back to the lock-in amplifier for measurement. With the measured zero-bias conductances, a 10 pA excitation corresponds in the worst case (2 $\mathrm{\mu S}$ differential conductance) to 50 aW heat input, and the voltage magnitude of the excitation is 5 $\mathrm{\mu V}$, corresponding to about 3% of $e/C_\Sigma$. From the measured noise level at 2 Hz, combined with the equivalent noise bandwidth of the lock-in amplifier, we estimate the RMS noise to be at most 2.5% of the signal responses obtained.
There are 20 DC connections from the room temperature electronics to the sample at the base temperature. Starting from the 300 K top flange, there are first 20 Thermocoax cables (length approx. 1.5 m) connect from room temperature to the 1.2 K level, then 20 superconducting wires in a ribbon to the mixing chamber (for thermal isolation), and finally 20 Thermocoax cables from the mixing chamber to the sample (length approx. 25 cm). The 20 Thermocoax cablesand the method of wiring minimize cross-talk and external noise input. The cables are thermally anchored at all temperature levels. The Thermocoax cables provide filtering of the room temperature thermal radiation;[@ZorinCoax] for $f$ = 10 GHz (corresponding to $h f/k_\mathrm{B}=0.5$ K) the attenuation of the cable is about 140 dB/m. The attenuation in dB increases as the square root of the frequency. Using a thermally conductive paste, the sample is mounted inside a thick-walled copper cavity and contacted electrically by “buckling wires” which tread on the sample. The 25 cm Thermocoax cables and the sample holder are enclosed in a copper shield at the mixing chamber temperature, which shields them from the surrounding 4.2 K radiation.
The individual shielding of the measurement leads also decreased the cross-capacitances considerably, which in turn minimizes cross-talk. The total capacitance to cryostat ground was measured to be 840 pF per lead at room temperature, which is in agreement with Zorin’s result.[@ZorinCoax] It is worth noticing that the leakage resistance in these Thermocoax cables is quite low when the cryostat is at room temperature (down to tens of M$\Omega$), but increases to many G$\Omega$ when cooled down. We conjecture that this may be due to water vapor being absorbed in the MgO insulation powder at room temperature and pressure, but being cryopumped and/or frozen out when cooling the cryostat. This is important because the resistance of the SET transistor can be very high in the Coulomb blockade state. A matrix connector at room temperature and a “switchboard” (20 connectors which can be connected to any of the 20 slots connecting to the sample) at the mixing chamber further enables us to choose the very best cables for the critical connections. Cross-talk can only take place in the room temperature electronics, the matrix board, the superconducting ribbon, in the “switchboard” at the mixing chamber, and in the on-chip wiring.
To avoid microphonic pick-up from mechanic vibrations in the building (pumps etc.), the whole cryostat can be suspended using three inflated rubber tubes. Furthermore, lateral vibrations are damped with a fourth rubber tube concentric with the cryostat. This arrangement proved very efficient; components deriving from the asynchronous pumps (around 49 Hz) and other components at around 30 Hz vanish upon inflation of the rubber tubes.
The high capacitances in the filters and Thermocoax cables force us to use quite low modulation frequencies. For our symmetric biasing setup, the cut-off frequency is about $(G_\mathrm{D}+1/2R_\mathrm{B})/(2 \pi C_\ell)$, where $R_\mathrm{B}$ is the value of the the bias resistor, $G_\mathrm{D}$ is the differential conductance of the device, and $C_\ell$ is the total capacitance per lead (the 4-point biasing means that there are two leads contributing to the capacitance on each side). Because of the high internal cut-off frequency of the device, we find it reasonable to assume a simple resistive behaviour for the SET device at these frequencies. In the worst case, $G_\mathrm{D}$ is smaller than $1/2R_\mathrm{B}$, and in this case the cut-off will be at $1/ (4 \pi R_\mathrm{B} C_\ell)$. We have used $R_\mathrm{B}$ = 10 M$\Omega$ in all experiments reported here, giving a worst-case cut-off at 8 Hz. The modulation frequency was chosen to be 2 Hz, from which we have calculated the worst-case amplitude error due to the filters to be a few percent at the lowest conductances (about 2 $\mathrm{\mu S}$).
To apply microwaves to the device, there is a separate coaxial connection to the mixing chamber. At room temperature, a vacuum tight feedtrough connects an SMA connector to a standard 0.05” 50 $\mathrm{\Omega}$ stainless steel coaxial cable inside the cryostat. A DC break is inserted before the vacuum feedthrough at room temperature, as well as before the thermal anchoring at 4.2 K. From this thermal anchor, the signal is fed through a short length of Thermocoax cable to another thermal anchor at the 1.2 K level. The Thermocoax cable serves as a cold attenuator (about 20 dB) for the microwave signal as well as unwanted signals such as room temperature radiation. The signal is then connected to the mixing chamber through a 20 m long superconducting 0.05” 50 $\mathrm{\Omega}$ Nb coax cable, which is thermally anchored at the 0.6 K level. The superconductivity of the Nb provides a microwave connection while retaining thermal isolation of the mixing chamber. Before connecting to the mixing chamber, a third DC break is inserted in the line. All DC breaks are “inside-outside” breaks, meaning that both the inner and outer conductors are interrupted. The final microwave connection to the samples is provided by one of the 25 cm Thermocoax cables also used for the DC connections. This gives further attenuation of the signal.
As a consequence of the distributed attenuation and possible resonances in the microwave transmission line and wiring to the SET device, the actual power applied to the sample is unknown. It would me meaningless to make a throughput measurement at room temperature, since the damping characteristics are different at low temperature, and the coupling to the sample is unknown. Therefore, all power values are measured at the leveled output of the microwave synthesizer. The attenuation of the line at each frequency will be a fitting parameter in the data. For a given frequency, it is assumed that the actual power delivered to the SET device is a fixed fraction of the power from the high-frequency source.
Pressures and temperatures in the circulation system are continuously monitored by a computer, which also controls the biasing of the sample and reads the results measured by the voltage meters etc. through a GPIB bus. The GPIB bus is separated electrically from all measurement electronics by an optical link, which galvanically separates the computer from the measurement, and also enables the computer to be moved out of the shielded room if necessary.
Results\[sec:results\]
======================
Device Characterization\[sec:char\]
-----------------------------------
For the particular SET transistor reported on here, the DC $I$-$V_\mathrm{DC}$ curve (Fig. \[fig:I-V\]) yield the device parameters and . From the $V_\mathrm{DC}$-$V_\mathrm{G}$ curves (not shown) we get . The set of $V_\mathrm{DC}$-$V_\mathrm{G}$ curves at different current biases also provided evidence that the device is symmetric, which we will assume in the following. The particular sample was chosen from a batch for its charging energy, which also means a higher tunneling resistance (for fixed $RC$ product). Our simulations show that a high charging energy is more important than a low tunneling resistance in obtaining a high cotunneling/sequential tunneling ratio.
Since we want to first compare with the theory for zero current bias and maximum blockade, it is important to make sure that this is actually the case in the measurements. To do this, we use the following procedure: First, the maximum Coulomb blockade is found by sweeping across a peak in the $V_\mathrm{DC}$-$V_\mathrm{G}$ curve, and fitting to find the exact location of the peak. Then, the zero DC bias point was found in a similar manner by recording and $I$-$V_\mathrm{DC}$ curve around zero DC bias and fitting the dynamic resistance to find the peak. This procedure was carried out before each measurement.
The microwave connection obviously exhibits resonances (see section \[sec:setup\]), at which the signal received by the SET is higher. It is advantageous for us to utilise these resonance frequencies for the microwave bias, and to determine them we used the SET device as a self-detector. It was assumed that the frequency dependence of the device response was small in the relatively narrow frequency range probed. The resonances are determined by measuring the dynamic resistance at zero DC bias and maximum blockade (using the method described above) in small steps. The method is based on the assumption of a response both from the sequential tunneling and cotunneling contributions, as neither of these are linear around zero bias. The result is shown in Fig. \[fig:Resonances\], where some clear peaks in the dynamic resistance are seen at various frequencies. The results shown in this paper are measured at the frequencies marked by arrows.
Temperature of the measurements {#sec:Temperature}
-------------------------------
An important question relating to any measurement at low temperature is: What is the correct temperature? First of all, the sample may be far away from the cooling source; in our case the mixing chamber of the dilution refrigerator. This means that any heat delivered to the sample by Joule heating or radiation may give a higher temperature at the sample than at the mixing chamber. Second, there is the question of whether the electron system is in thermal equilibrium with the phonon system.[@Wellstood] This suggests that it is not correct to rely on the temperature measured by the thermometers at the mixing chamber, even when these are well-calibrated.
To get an estimate of the real temperature temperature of the electron system in the SET transistor, we use the results of the differential conductance measurements at maximum blockade and zero bias. Using both the cotunneling theory (Eq. \[eq:ACcotun3rdOrder\]) and the orthodox theory, the temperature has been fitted to make the differential conductance $G_\mathrm{D}$ agree. The junction parameters were fixed at the values determined in the previous section, leaving $T_\mathrm{e}$ as the only free parameter. The results of this temperature fitting is shown in Fig. \[fig:Temperatures\] as function of the mixing chamber temperature $T_\mathrm{ph}$ (which we, following Ref. , assume equal to the phonon temperature, hence the symbol). The temperature has been fitted both to the sequential tunneling conductance alone (squares), which will give an upper bound on the temperature, and to the tunneling conductance with cotunneling included (circles), which will give a lower bound under the assumption that other effects can be neglected. Thus in the last case we have implicitly assumed that we can use the theory for cotunneling by Averin & Nazarov at low temperatures, which has been confirmed by several researchers.[@Geerlings; @Eiles] Also shown is the “strong coupling” curve $T_\mathrm{e}=T_\mathrm{ph}$ (dashed line), which can be considered a “hard” lower bound.
It is seen that the fitted electron temperatures indeed saturate at low mixing chamber temperatures, as suggested in the work by Wellstood et [*al.*]{}[@Wellstood] However, attempts to make the results fit to the $T_\mathrm{e} = \sqrt[5]{\smash[b]{T_0^5+T_\mathrm{ph}^5}}$ form, which applies for uniform heating in a thin film, seem to fail for our SET structure ($T_0$ is the temperature at which $T_\mathrm{e}$ saturates for $T_\mathrm{ph} \rightarrow 0$). A much more reasonable fit is provided by the form $T_\mathrm{e} = \sqrt[2.5]{\smash[b]{T_0^{2.5}+T_\mathrm{ph}^{2.5}}}$ (solid line). This form respects the lower bound as well as the limit at $T_\mathrm{ph} \rightarrow 0$. Such a dependence is not entirely unreasonable, considering that the heating is localized in the SET transistor, and that heat transfer is provided only by the leads, which should probably be considered between one- and two-dimensional. Indeed the experiments of Wellstood et [*al.*]{} on local heating of a SQUID with cooling fins gave a $T_\mathrm{e}^{2.7}$ form.
We will adopt the $T_\mathrm{e} = \sqrt[2.5]{\smash[b]{T_0^{2.5}+T_\mathrm{ph}^{2.5}}}$ form as our best estimate of the temperature in the following. We make no claims that this is the accurate temperature, only that it seems to be a reasonable approximation which makes a sensible transition from the low-temperature to the high-temperature regime.
$V_\mathrm{RF}^2$ dependence\[sec:Vac\]
---------------------------------------
In this section, we present measurements at maximum blockade, which means that $\Delta^\pm$ reduce to the charging energy $E_\mathrm{C}$. Also, since there is no asymmetry in the device at zero bias and maximum blockade (no tunneling direction is favored), there should be no frequency dependence in the differential conductance, which is also predicted by theory. The absence of frequency dependence enables us to find the relative damping of the line at the individual frequencies. Without frequency dependence, the zero bias conductance $G_\mathrm{D}$ as function of the microwave amplitude $V_\mathrm{RF}$ should be the same at all frequencies. This means that the relative damping can be found by minimizing the mean square difference between the $G_\mathrm{D}$-$V_\mathrm{RF}$ curves at the individual frequencies, using the damping as free parameter. This procedure provided us with the relative damping values 0.0, 3.8, 9.1, 11.5, 14.8 17.3, and 17.6 dB for the frequencies marked in Fig. \[fig:Resonances\]. As expected from Fig. \[fig:Resonances\], the damping increases with frequency. However, it should be noted that the optimal values vary slightly with temperature, suggesting that there might be a co-dependence on temperature and frequency.
The additional damping (i.e. the damping that should be added to the relative damping to get the absolute damping) should be the same at all frequencies and temperatures. The absolute damping is unfortunately unknown. All we can do is to find a damping value that is consistent with theory. In Fig. \[fig:Gd-Vac37.9dB\], the resulting theoretical prediction of the $G_\mathrm{D}$-$V_\mathrm{RF}$ curves at different temperatures are shown along with the measured values, assuming an additional damping of 37.9 dB. It is seen that the curves from different frequencies have indeed collapsed into one, especially at the lowest temperature. At higher temperatures, where the change in tunneling is more modest, the picture is not as perfect. Also shown is the predicted excess conductance with and without cotunneling at the same temperature added to the zero-power conductance. It is seen that the measured conductance fits the predicted one quite well, especially at low temperatures.
The results just presented could lead to the hasty conclusion that the theory for coherent photon assisted cotunneling is correct. However, there are several points which cause concern. First, the additional damping required to make the curves fit changes with temperature, which it should not, unless the change from about 200 mK to about 500 mK really increases the damping (which seems very improbable). There could be a range of other explanations for this result. For example, the change could be an artefact of the fitting procedure: If the excess cotunneling conductance becomes smaller than predicted by theory at higher temperatures (which seems reasonable), this will show up in the fit as a higher damping. A related problem is that the quality of the fit should be the same at all temperatures if the temperature used for the theoretical prediction is correct (i.e. equal to the actual electron temperature). This is because the temperature does not (in the theory) contribute to the [*excess*]{} cotunneling conductance; it only gives a constant contribution (compare Eq. \[eq:ACcotun3rdOrder\]). However, the fit is best at low temperatures.
One ‘sanity check’ that should be performed is whether the orthodox theory alone can explain the results. Using the value 36.2 dB for the additional damping gives the result in Fig. \[fig:Gd-Vac36.2dB\]. It is seen that the results cross the prediction for the orthodox tunneling at this damping value. At higher damping, the conductance would be too high at high amplitudes, while a lower damping the conductance would be too low at low amplitudes. Thus the results can not be explained by the orthodox theory, even if one assumes a higher electron temperature and another damping. The electron temperature would also have to be substantially higher than at the mixing chamber at all temperatures, as seen in Fig. \[fig:Temperatures\]. Note that although it would seem that we are applying a double standard here, since the horizontal scale is different in Fig. \[fig:Gd-Vac37.9dB\] and \[fig:Gd-Vac36.2dB\]. However, the orthodox theory should be valid for all voltages and thus amplitudes, while the cotunneling theory only claims validity for $e V_\mathrm{RF} \ll \Delta^\pm$.
The $\Delta^\pm$ dependence
---------------------------
The dependence of the cotunneling on $\Delta^\pm$ has been investigated by measuring the dynamic conductance at zero bias as function of $V_\mathrm{G}$. Zero bias was found as described previously, and the device was biased within one $e$-period of the gate bias. The result is shown in Fig. \[fig:Delta\], along with the predictions of the excess conductance with and without cotunneling added to the maximum-blockade result. For the prediction including cotunneling, the assumed electron temperature is used (solid line in Fig. \[fig:Temperatures\]); for the sequential tunneling the temperature fitting the maximum-blockade point (squares in Fig. \[fig:Temperatures\]). As expected, the cotunneling prediction fits reasonably well near maximum blockade, but diverges quickly when approaching minimum blockade, where $\Delta^\pm \rightarrow 0$. Also shown is the sequential tunneling contribution to the current; it is seen that if the temperature used is correct, the sequential tunneling more or less ‘takes over’ as $\Delta^\pm$ approaches zero. The sequential tunneling alone seems to fit surprisingly well. One explanation for the good fit is that as $\Delta^\pm$ approaches zero, the threshold voltage becomes smaller,[@Eiles] and thus the orthodox theory will account for most of the current.
Photon assisted cotunneling away from blockade
----------------------------------------------
The theory derived by Flensberg [@Flensberg] applies for $e V_\mathrm{DC},\:k_\mathrm{B} T_\mathrm{e},\:e V_\mathrm{RF},\:hf\ll \Delta^\pm$. Thus, when one of $\Delta^\pm$ becomes smaller, the voltage/temperature/amplitude interval where this assumption is valid becomes smaller. Measurements similar to those presented in the previous sections have been performed also at $C_\mathrm{G} V_\mathrm{G}/e = 0.25,0.50,0.75$, i.e. $\Delta^\pm/E_\mathrm{C} = 0.5,0.0,0.5$. Thus the bounds of validity for $V_\mathrm{DC}$, $T_\mathrm{e}$ and $V_\mathrm{RF}$ are diminished at $C_\mathrm{G} V_\mathrm{G}/e = 0.25,0.75$ (slopes in the $V_\mathrm{DC}$-$V_\mathrm{G}$ curve), while the condition is impossible to satisfy at $C_\mathrm{G} V_\mathrm{G}/e = 0.50$ (minimum blockade), where $\Delta^\pm = 0$. Another way of stating this is that Eq. \[eq:ACcotun3rdOrder\] diverges at $\Delta^\pm \rightarrow 0$.
The results are presented in Fig. \[fig:NonBlockade\]. It is seen that on the slopes, the observed conductance falls between the predictions with and without cotunneling, which is expected since the cotunneling should be overestimated at these values of $\Delta^\pm$. At minimum blockade the conductance approaches the prediction from the orthodox theory alone, which is reasonable for a SET transistor in the minimum blockade, where the sequential tunneling dominates (this is similar to approaching the charging voltage at maximum blockade).
[c]{}\
\
\
Discussion\[sec:discussion\]
============================
The interpretation of the results involves several explicit and implicit assumptions. To determine the lowest temperatures, it is assumed that both the orthodox theory for sequential tunneling and the original cotunneling theory by Averin & Nazarov [@AverinNazarov] were correct within the given bounds of validity. The orthodox theory has been extensively confirmed since the first theories were published, and also the original cotunneling has been verified.[@Geerlings; @Eiles] It is further assumed that the coupling from the electron system to the mixing chamber through the phonon system was strong enough at high temperatures to assume a convergence towards the mixing chamber temperature here.
The largest problem in the interpretation of the results is that both the electron temperature of the device and the damping in the microwave line were not known exactly. Since it is certainly insufficient to use the mxing chamber temperature, which is not in equilibrium with the electron temperature, it is necessary to rely on existing theories and some assumptions to get an estimate of the electron temperature. Similarly, it is necessary to rely on the absence of frequency dependence in the device to determine the relative damping values at the different frequencies. The additional damping is then an adjustable parameter. Here it is implicitly assumed that the damping does not depend on amplitude. Ideally, it should also have been checked thoroughly that the differential conductance measurement was independent of the modulation amplitude. However, this was only done sporadically, i.e. by halving and doubling the amplitude and checking that the response followed.
Another problem is that the temperatures did not range to below about $T_\mathrm{C}/4$ (another way of saying this is that the tunnel capacitances were somewhat too high). Thus it is difficult to say with conviction that the condition $k_\mathrm{B} T_\mathrm{e} \ll \Delta^\pm$ is satisfied, even at the lowest temperatures.
Among the other assumptions are that the the SET transistor exhibited a resistive behaviour at low frequencies (should be reasonable since $R_\Sigma C_\Sigma$ corresponds to about 1.7 GHz), that the damping in the microwave line is linear, that the leads are equal (small correction), and that other effects, such as noise and the environment, can be neglected. Under the given assumptions, the damping could be chosen such that results are consistent with the theory at low temperatures. Furthermore, it is shown that sequential tunneling alone could not explain the results. Thus, the results speak in favor of the cotunneling theory. At least we can say that more than the orthodox theory is needed to explain the results, and that cotunneling is a resonable explanation. However, due to the dicrepancies observed, it is clear that a more complete theory is needed.
There is still the unsettling matter of the required additional damping given by the fitting procedure is changing with temperature. A number of explanations can be offered for this. It could be the coupling that changes, but it is hard to see how such a small change in temperature (from 200 mK to 500 mK) would cause a change of 10 dB in the damping. A more reasonable explanation could be that the enhancement of cotunneling by a coherent source must be temperature dependent (i.e. if the real temperature contribution becomes smaller than predicted by the theory, this will in the fitting be compensated by increasing the damping, making the microwave contribution smaller instead), which would require the assumed electron temperatures to be approximately correct. For example, we may conjecture that at higher temperatures, the limit of validity is not far away; i.e. that we should require something like $(2\pi k_\mathrm{B} T_\mathrm{e})^2 + (eV_\mathrm{DC})^2 + \frac{3}{2}(e V_\mathrm{RF})^2 \ll (\Delta^\pm)^2$ in Eq. \[eq:ACcotun3rdOrder\]. Another possibility is that the temperatures used are too low, with the error increasing with temperature. This would make the orthodox prediction too high, which would then be compensated in the fit by a higher damping. However, the orthodox theory prediction approaches the conjectured electron temperatures at high temperatures, making the error margin upwards smaller.
Also the relative damping varied with temperature, hinting that also the frequency might have a different effect on the conductance at different temperatures, again contrary to the theoretical prediction. Indeed the charging frequency of the device is $f_\mathrm{C}=18.8$ GHz, which means that the frequencies used range from $0.11f_\mathrm{C}$ to $0.28f_\mathrm{C}$. Similar arguments as above could be applied to this matter.
Conclusion and Outlook\[sec:conclusion\]
========================================
The enhancement of cotunneling by a coherent microwave source was investigated experimentally and compared with theory. The results on cotunneling seem to confirm the theory for adiabatic enhancement of cotunneling by a coherent microwave source. However, the presence of assumptions and adjustable parameters should inspire modesty in the strength of the conclusions drawn. Furthermore, the results seem to suggest that the effect of microwave amplitude and temperature are not independent, contrary to what is suggested by theory. Thus a more complete theory is needed. Experimentalists would very much appreciate a theory which gives a method of actually calculating the cotunneling current at all voltages, temperatures etc. (e.g. something like the Master Equation approach). In any case, the presence of microwave induced cotunneling enhances the importance of taking cotunneling into account in the calculation of SET devices.
It is vital in future measurements to have samples with a high charging energy to make the charging temperature higher. Also one could attempt to lower the electron temperature, e.g. by using cooling fins on the sample. Cold resistors and cross-correlation voltage measurements should be used to improve the noise. A new voltage bias electronics, which will reduce the noise and eliminate the problem with resistive biasing of a high-impedance sample, is under construction. The measurements should focus on verifying the photon-assisted current in Eq. \[eq:ACcotun4thOrder\], e.g. by biasing in zero voltage and observing the current as function of amplitude and frequency. A careful determination of the device parameters and damping of the microwave line is essential, as there will be adiabatic contributions from both sequential tunneling and cotunneling. It should futher be checked that the differential conductance result is idependent of the amplitude of the lock-in modulation. Also the measurements should be performed on many more samples to rule out sample-specific effects.
The authors would like to thank the SET groups at PTB for many fruitful discussions. Mikkel Ejrnæs is gratefully acknowledged for his work on improving the cryostat and the computer measurements. Karsten Flensberg is thanked for clarifying discussions and thoughtful comments. The project was supported in parts by the Danish Natural Science Foundation, the Hartmann Foundation, and EU (contract IST-1999-10673). MS acknowledges a grant from Tekniikan Edistämissäätiö (TES).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
The [$\gamma$-rays ]{}from [$\gamma$-ray ]{}bursts (GRBs) are believed to be produced by internal shocks driven by small timescale, $\sim1$ ms, variation in the GRB outflows, and a pair-production spectral cutoff is generally expected around the GeV range. However, the observed optical flashes accompanying GRBs suggest that the delayed residual collisions due to large timescale variation continue to accelerate electrons. We show here that the inverse-Compton (IC) scattering of the prompt [$\gamma$-rays ]{}by these residual internal shock electrons leads to a high energy emission beyond the previously thought spectral cutoff, in agreement with the previous detections of GeV photons by EGRET in several GRBs in conjunction with MeV emission. We expect a spectral break due to the transition from the primary to residual internal shock emission at the previously thought spectral cutoff, and expect systematic time delays of high energy photons relative to MeV emission, the discovery of which would provide stringent constraint on the outflow properties, but requires large enough collection of high energy photons by, e.g., Fermi and AGILE satellites.
The recent Fermi-detected bright GRB 080916c unambiguously shows the shifting of the prompt emission toward later times as the photon energy increases. The second-scale shifting at >100 MeV is much longer than the MeV variability time, as predicted in the residual collision model. The observations imply that there should be emission above 70 GeV in the source frame, which may not be produced by primary internal shocks but by IC emission in residual collisions. With the method involving time delays of high energy emission, the bulk Lorentz factor of GRB 080916c is determined to be $\Gamma\sim300$.
author:
- Zhuo Li
title: 'Prompt GeV Emission from Residual Collisions in Gamma-Ray Burst Outflows: Evidence from Fermi Observations of GRB 080916c'
---
Introduction
============
The prompt MeV [$\gamma$-rays ]{}from a GRB are well believed to be produced by a unsteady outflow which causes internal collisions between different parts with various velocities, leading to kinetic energy dissipation [@PX94; @RM94][see, however, @LB03; @NK08]. The internal shock model can naturally explain both the non-thermal spectra and the complicated light curves of GRBs. The observed temporal variabilities of [$\gamma$-ray ]{}emission are believed to be reflecting the activities of the central engine [@SP97; @Kobayashi97]. The internal shocks are expected to generate/amplify magnetic field and accelerate electrons, leading to MeV [$\gamma$-rays ]{}by synchrotron emission [see, e.g. @Waxman; @rev for a review].
GRBs are mainly observed in MeV range, the properties of high energy, say, $>100$MeV, emission are not well understood, which, however, might be very helpful in constraining the physics of the GRB emission region. For example, the observed $>100$ MeV photons in conjunction with MeV emission by EGRET in several GRBs, suggesting that they can avoid the $\gamma\gamma$ absorption, have leaded to the conclusion that the GRB outflow must be relativistically expanding with a Lorentz factor of $\Gamma\ga100$ [e.g. @Baring97][see also @Waxman; @rev]. As the development in high energy [$\gamma$-ray ]{}observations, e.g. AGILE and Fermi are being well operated, there are great interests on detecting the high energy pair-production spectral cutoff in order to constrain the size and Lorentz factor of GRB emission region (e.g. ; and recent detailed consideration by [@Gupta07; @Gupta08; @Granot08; @Murase08]).
However, a simple spectral cutoff may not exist. @LW08 [hereafter LW08] had noticed that the frequently observed prompt optical emission [@Akerlof99; @Blake05; @Vestrand05; @Vestrand06; @Yost07] is above the expected synchrotron-self absorption limit, and suggests a relatively large size of optical emission region, compared to that of MeV emission. Actually in the context of internal shock models one would expect delayed collisions in the outflow following the small radius collisions driven by smallest timescale variation of the outflow properties, and these delayed collisions can naturally account for the relatively bright optical emission (LW08). The recently detected optical flash from the naked-eye GRB 080319b [@Bloom08; @DElia08; @Racusin08; @319bapj] appears to vary rapidly in times and its temporal profile is correlated to the MeV emission in second scales [@080319b_delay], supporting that the prompt optical emission in GRBs is produced by internal shocks within the outflow, i.e. synchrotron emission from residual collisions (Li & Waxman, in preparation). The electrons in residual collisions mainly cool by IC scattering off the MeV photons, which produces high energy emission at large radii where the optical depth due to pair production is reduced (LW08). We consider here this high energy emission and show that it may “smear” the previously thought pair-production spectral cutoff. But a spectral turnover is still expected, which may be observed by Fermi and AGILE although more difficult to detect than a simple cutoff. The high energy emission from residual collisions is also expected to be delayed relative to MeV [$\gamma$-rays ]{}. It should be noticed that we only focus on the [*prompt*]{} high energy emission which appears simultaneously with the MeV [$\gamma$-ray ]{}emission.
We show first in §2 the strong $\gamma\gamma$ absorption during the MeV [$\gamma$-ray ]{}emitting phase, next discuss in §3 the high energy emission from residual collisions, then the model is applied to the recent Fermi detection of GRB 080916c [@Fermi916c] in §4, which might have provided evidences of the model, finally a general discussion on observations is given in §5.
$\gamma\gamma$ absorption at small radii
========================================
Consider a highly relativistic outflow with a bulk Lorentz factor $\Gamma=10^{2.5}\Gamma_{2.5}$. The small timescale variation will lead to strong internal shocks, which produce [$\gamma$-ray ]{}emission. Let us denote the radius where [$\gamma$-rays ]{}arise by $R_\gamma$. Due to geometry effect, the observed fastest variability timescale $t_{\rm var}\la10^{-2}$s [e.g. @Woods95] in [$\gamma$-ray ]{}light curves suggests that the size of [$\gamma$-ray ]{}emission region is limited to $R_\gamma\la2\Gamma^2ct_{\rm var}$. If the [$\gamma$-ray ]{}emitting electrons are fast cooling (with cooling time shorter than the dynamical time of the outflow), we should take the equality, thus, $R_\gamma\approx6\times10^{13}\Gamma_{2.5}^2t_{\rm var,-2}$cm, where $t_{\rm var}=10^{-2}t_{\rm var,-2}$s. Actually, in the context of internal shock models, we do not expect the magnetic field is generated with energy density much higher than that of accelerated electrons. In order to have synchrotron emission peaking at $\eps_b\sim1$ MeV, as observed in GRBs, the radius of the [$\gamma$-ray ]{}emitting region should be small (LW08), $R_\gamma\la10^{13}L_{\rm bol,52}^{1/2}(\eps_b/1~\rm MeV)^{-1}$cm, where $L_{\rm bol}=10^{52}L_{\rm bol,52}\rm erg~s^{-1}$ is the bolometric [$\gamma$-ray ]{}luminosity.
Let us consider the $\gamma\gamma$ absorption due to pair production inside the GRB source. For a photon of high energy $\varepsilon$ the optical depth to pair-production within the GRB source is given by the product of the pair-production rate, $1/t'_{\gamma\gamma}(\varepsilon)$, and the dynamical time, the time required for significant expansion of the plasma, $t'_d\simeq R_\gamma/\Gamma c$ (primes denote quantities measured in the outflow rest frame), $\tau_{\gamma\gamma}(\varepsilon)\simeq R_\gamma/\Gamma
ct'_{\gamma\gamma}(\varepsilon)$. $t'_{\gamma\gamma}(\varepsilon)$ depends on the energy density and on the spectrum of the radiation. The (outflow rest frame) radiation energy density is approximately given by $U'_\gamma=L/4\pi
R_\gamma^2c\Gamma^2$. The GRB spectrum can typically be described as a broken power law, $dn/d\eps\propto\eps^{-\beta}$, with $\beta\approx-1$ at low energy, $\eps<\eps_b\sim1$ MeV, and $\beta\approx-2$ at $\eps>\eps_b$ [@Band93]. High energy photons with energy $\varepsilon'$ exceeding $\varepsilon'_b\equiv2(m_ec^2)^2/\eps'_b$, may produce pairs in interactions with photons of energy exceeding $\eps'=2(m_ec^2)^2/\varepsilon'<\eps'_b$ (the rest frame photon energy $\varepsilon'$ is related to the observed energy by $\varepsilon=\Gamma\varepsilon'$). For $\eps'<\eps'_b$ we have $dn/d\eps'\propto\eps^{\prime-1}$, which implies that the number density of photons with energy exceeding $\eps'$ depends only weakly on energy. Thus, $t_{\gamma\gamma}$ is nearly independent of energy for $\varepsilon'>\varepsilon'_b$, $t_{\gamma\gamma}^{\prime -1}\approx(\sigma_T/16)cU'_\gamma/\eps'_b$, which gives $$\label{eq:tau}
\tau_{\gamma\gamma}(\varepsilon>\varepsilon_b)\simeq
1.1\times10^2\frac{L_{52}}{t_{\rm var,-2}\Gamma_{2.5}^4} {\left(\frac{\epsilon_b}{\rm 1~MeV}\right)}^{-1}.$$ Note, we have approximated the $\gamma\gamma$ cross section above the pair-production threshold as $3\sigma_T/16$ [@Waxman; @rev]. Also note that as the energy density $U_\gamma'$ (and hence the related luminosity $L$) depends on the energy band considered, the one used in calculating $t_{\gamma\gamma}^{\prime}$ is the energy density below $2\times\eps_b'$. Hereafter, without special emphasis the luminosity $L=10^{52}L_{52}\rm erg\,s^{-1}$ is the so-called MeV luminosity, only corresponding to emission at $<2\times\eps_b$, i.e., $L\equiv\int_0^{2\eps_b}L_{\eps}d\eps$.
Photons of lower energy, $\varepsilon<\varepsilon_b$, interact to produce pairs only with photons of energy $\eps'>2(m_ec^2)^2/\varepsilon'>\eps'_b$. Since the number density of these photons drops like $1/\eps'$, we have $\tau_{\gamma\gamma}(\varepsilon<\varepsilon_b)\approx
(\varepsilon/\varepsilon_b)\tau_{\gamma\gamma}(\varepsilon>\varepsilon_b)$, i.e. $$\tau_{\gamma\gamma}(\varepsilon<\varepsilon_b)\simeq
2.2\times10^{-3}\frac{L_{52}}{t_{\rm var,-2}\Gamma_{2.5}^6}\frac{\varepsilon}{1~\rm MeV}.$$ The optical depth increases as photon energy increases. Photons with high enough energy are absorbed in the emission region. A spectral cutoff is defined by $\tau_{\gamma\gamma}=1$,[^1] $$\label{eq:cut1}
\varepsilon_{\rm cut}^{(1)}\simeq0.46\frac{t_{\rm
var,-2}\Gamma_{2.5}^6}{L_{52}} \rm ~GeV.$$ A comment on the approximation of the $\gamma\gamma$ cross section should be made here. Since both the cross section and the GRB photon spectrum decrease rapidly with photon energy, the approximation is excellent- for a GRB spectrum with $\beta\approx-2$, it precisely produces the optical depth within 2%, compared to a calculation with full cross section.
There is another restriction for the cutoff besides eq. (\[eq:cut1\]). In deriving the cutoff, eq. (\[eq:cut1\]), the optical depth is not self-consistently calculated since a GRB spectrum extending to infinity without a high energy cutoff is assumed. Given the two factors that GRB spectrum usually appears to be a steep slope, with the photon number dominated by low energy photons, and that the high ($\varepsilon'>2^{1/2}m_ec^2$) and low ($\varepsilon'<2^{1/2}m_ec^2$) energy photons annihilate each other one by one, we only expect high energy photons might be totally attenuated by low energy ones, other than the other way around. So the cutoff should be $\varepsilon_{\rm
cut}'>2^{1/2}m_ec^2$, which is not assured by $\varepsilon_{\rm
cut}^{(1)}$ in eq. (\[eq:cut1\]) ($\varepsilon_{\rm
cut}^{(1)}/\Gamma<2^{1/2}m_ec^2$ might happen). We need to define, in the GRB source frame, $$\varepsilon_{\rm cut}^{(2)}=2^{1/2}\Gamma m_ec^2\simeq0.23\Gamma_{2.5} \rm
~GeV.$$ Note, this condition is also forgotten by many other authors who calculate the cutoff energy assuming a no-cutoff GRB spectrum. The GRB spectral cutoff is, therefore, expected at the maximum between $\varepsilon_{\rm cut}^{(1)}$ and $\varepsilon_{\rm cut}^{(2)}$ [@LS04], $$\label{eq:cut}
\varepsilon_0\equiv\varepsilon_{\rm cut}(R_\gamma)
\simeq\max\left[0.46\frac{t_{\rm var,-2}\Gamma_{2.5}^6}{L_{52}},~0.23\Gamma_{2.5}\right]\rm
GeV.$$ A critical Lorentz factor where $\varepsilon_{\rm cut}^{(1)}=\varepsilon_{\rm
cut}^{(2)}$ is function of $R_\gamma$ (hence $t_{\rm var}$), $$\Gamma_c\simeq280(L_{52}/t_{\rm var,-2})^{1/5}.$$ $\varepsilon_0$ sensitively depends on $\Gamma$ thus detection of the spectral cutoff may be very useful in constraining $\Gamma$, $$\label{eq:Gamma_cut}
\Gamma=\min\left[360\left(\frac{L_{52}}{t_{\rm var,-2}^{\rm ob}}\frac{\varepsilon_0^{\rm ob}}{\rm
1~GeV}\right)^{1/6}(1+z)^{1/3},~
1390\frac{\varepsilon_0^{\rm ob}}{\rm 1~GeV}(1+z)\right].$$ Here the redshift factor has been included, i.e. $t^{\rm
ob}=t(1+z)$ and $\varepsilon^{\rm ob}=\varepsilon/(1+z)$.
The extension of GRB spectra to $\ga100$ MeV and the characteristic variability time, $t_{\rm var,-2}\la1$, have implied $\Gamma_{2.5}\gtrsim1$, assuming the $\ga100$ MeV photons are produced in the same time and place as the MeV photons. Since thermal pressure acceleration in the initial fireball can not lead to much larger Lorentz factors, and in internal shock model much larger Lorentz factors would lead to synchrotron emission peaking below MeV band (see discussion in the first paragraph of §2), $\Gamma_{2.5}\approx1$ is typically adopted [e.g @Waxman; @rev]. The exact values of $\Gamma$ would be determined by the detection of the high energy cutoffs in GRB spectra by, e.g. Fermi. We show below that the situation may be complicated by the delayed large-radius emission from the outflow.
Large-radius emission from residual collisions
==============================================
After the initial strong collisions at small radii driven by the small timescale, $\sim1$ ms, variation in the outflow, there are residual collisions continue to occur when the outflow is expanding to large radii. As the velocity fluctuation in the flow is being smoothed out by the on-going collisions, the delayed collisions become weaker, and the postshock electron energy and magnetic field are smaller, which give rise to synchrotron emission at longer wavelengths. LW08 has well discussed the dynamics of the residual collisions and naturally explained the optical flashes from GRBs by this delayed residual emission. As mentioned in LW08, the energy density in the delayed collisions is dominated by the primary [$\gamma$-ray ]{}emission, so that the residual emission is dominated by IC scattering of the prompt [$\gamma$-ray ]{}photons. In what follows consider the IC emission.
Dynamics
--------
Let us approximate the outflow by a sequence of $N\gg1$ individual shells. LW08 has considered the simplified case: the shell masses are equal; the shells are initially separated by a fixed distance $ct_{\rm var}$; the extent of the shells is much smaller than the radius of the outflow; and the Lorentz factors of individual shells are drawn from a random distribution with an average $\Gamma$ and initial variance $\sigma_{\Gamma,0}^2<\Gamma^2$. The model may be more complicated, by adding more degrees of freedom, however LW08 has shown that this simple case naturally accounts for the observed properties of GRB optical flashes. Here we will consider this simplified case. Moreover, we adopt the assumption that two shells merge into a bigger shell after a collision, i.e. the full inelastic collision where the internal energy generated is fully radiated. If the postshock electrons carry an energy close to equipartition and the electrons cool fast, the internal energy is always radiated significantly. In the case of a significant fraction of the internal energy being dissipated in each collision, the dynamical evolution of the outflow has been proven to be similar to the full inelastic case (LW08).
In the simple case under consideration, the variance of the velocities of individual shells (in the outflow rest frame) evolves with the outflow radius $R$ as $\sigma_v\propto R^{-1/3}$. The outflow energy that is associated with the fluctuation of shell velocities (in the outflow rest frame) and hence may be dissipated decreases as $$E_{\rm fluc}\propto\Gamma\sigma_v^2\propto R^{-2/3}.$$
In general, it is naturally expected that there might be a wide range of variability timescale, $\rm \sim1~ms-10~s$, in the flow properties. Large timescale variabilities might lead to more energy dissipated at large radii. Thus, the slope should be flatter than $-2/3$. If a power-law description, $E_{\rm
fluc}\propto R^{-q}$, is still available, one may have $0<q<2/3$. We carry a monte carlo simulation to demonstrate this point in the appendix.
High energy emission
--------------------
Based on the dynamical evolution, the emission from the residual collisions can be further predicted. Taking the common assumptions that in internal shocks the postshock electrons and magnetic fields carry fixed fractions, $\eps_e$ and $\eps_B$, respectively, of the postshock internal energy, the characteristic Lorentz factor of postshock electrons (in the outflow comoving frame) scales as $\gamma_i\propto\eps_e\sigma_v^2\propto R^{-2/3}$, and the postshock magnetic field scales as $B^2\propto\eps_B\sigma_v^2n_e\propto R^{-8/3}$ (the particle number density scales as $n_e\propto R^{-2}$).
We demonstrate that the electrons in residual collisions lose most of their energy by IC cooling. If the prompt [$\gamma$-rays ]{}last a duration $T$ (observer frame), the plasma is overlapped with these [$\gamma$-rays ]{}until the outflow expands to $R\ga2\Gamma^2cT\simeq(T/t_{\rm
var})R_\gamma$. LW08 showed that when the synchrotron emission lies in the optical band, the radius is $R_{\rm
opt}\simeq10^2R_\gamma$. For typical observed values $t_{\rm
var}\la10^{-2}$s and $T\sim10$s, the optical radius is still relatively small, $R_{\rm opt}<2\Gamma^2cT$. Therefore during the phase of late residual collisions that we concern, the plasma is immersed in the radiation bath of the prompt [$\gamma$-rays ]{}. Both the photon energy density $U_\gamma$ and the particle number density $n_e$ drop as $\propto R^{-2}$ hence the ratio $y=U_\gamma/(B^2/8\pi)\propto\sigma_v^{-2}\propto R^{2/3}$ increases with $R$. Because $y\sim1$ in the [$\gamma$-ray ]{}producing phase, we have $y>1$ in residual collision phase, so the radiation energy density dominates that of the magnetic field. Let us consider the properties of IC emission.
### Spectrum
Consider first the energy band into which energy is radiated. At radius $R$, the prompt [$\gamma$-ray ]{}photons with typical energy $\eps_b$ are up-scattered by electrons with characteristic Lorentz factor $\gamma_i$ to energy $\varepsilon_{\rm IC}\simeq
\lambda\gamma_i^2\eps_b\simeq
\lambda\epsilon_e^2(m_p/m_e)^2(R/R_\gamma)^{-4/3}\eps_b$. Here we assume $\gamma_{i,0}\sim \epsilon_e(m_p/m_e)$ as the electron Lorentz factor emitting MeV [$\gamma$-rays ]{}, and $\lambda$ accounts for the correction due to uncertain geometry effect. It will be shown in appendix that the correction factor $\lambda$ is order unity even in the case that the prompt MeV photons are strongly beamed in the rest frame of the outflow at $R$. The characteristic scattered photon energy is $$\label{eq:ICph_energy}
\varepsilon_{\rm IC}^{\rm ob}\simeq9\lambda\epsilon_e^2\frac{\epsilon_b}{\rm 1~MeV}{\left(\frac{R}{10^2R_\gamma}\right)}^{-4/3}(1+z)^{-1}\rm GeV.$$ The scattering might take place within slight Klein-Nishina regime, $\gamma_i\epsilon_b'/m_ec^2\sim$ a few $>1$, at small radii $R\sim R_\gamma$, where the energy of scattered photons is instead $\varepsilon_{\rm
IC}=\Gamma\gamma_im_ec^2=\epsilon_e\Gamma(R/R_\gamma)^{-2/3}m_pc^2\simeq3\times10^2\epsilon_e(R/R_\gamma)^{-2/3}$GeV.
Next consider the $\gamma\gamma$ absorption effect on the late residual emission. For GRB outflow with $\Gamma>\Gamma_c$, the initial cutoff energy for the primary emission is determined by the first term in the bracket of eq. (\[eq:cut\]), $\varepsilon_0=\varepsilon_{\rm cut}^{(1)}$ . Eq. (\[eq:cut1\]) implies that the spectral cutoff energy scales as $\varepsilon_{\rm cut}\propto R\Gamma^4L^{-1}$, so the cutoff energy increases with $R$ for fixed $L$. We have, for late residual collisions, $\varepsilon_{\rm cut}\simeq\varepsilon_0
R/R_\gamma$. In the case of GRB outflow with lower Lorentz factor $\Gamma<\Gamma_c$, the cutoff energy is initially a constant, $\varepsilon_0=\varepsilon_{\rm cut}^{(2)}$ (eq. \[eq:cut\]), until the outflow expands to a radius, $$R_m=3\times10^{13}L_{52}\Gamma_{2.5}^{-3}\rm cm$$ (note $R_m>R_\gamma$). At $R>R_m$, the cutoff energy turns to increase with $R$, $\varepsilon_{\rm cut}\simeq\varepsilon_0 R/R_m$. In both cases of $\Gamma>\Gamma_c$ and $\Gamma<\Gamma_c$ the cutoff energy at $R>\max[R_\gamma,\,R_m]$ (i.e, $\varepsilon_{\rm cut}>\varepsilon_0$) follows the same expression, $$\label{eq:cut evolution}
\varepsilon_{\rm cut}^{\rm ob}\simeq50\frac{t_{\rm var,-2}^{\rm ob}\Gamma_{2.5}^6}{L_{52}}{\left(\frac{R}{10^2R_\gamma}\right)}(1+z)^{-2}\rm GeV.$$ The evolution of $\varepsilon_{\rm cut}$ versus $R$ for fixed $L$ is plotted in fig \[fig:cutvsR\].
![Schematic plot of the attenuated energy evolving with radius. There are two regimes. If the primary collisions that produce MeV emission occur in the regime of $R_\gamma>R_m$ (i.e. $\Gamma>\Gamma_c$), $\varepsilon_0=\varepsilon_{\rm cut}^{(1)}$ and $\varepsilon_{\rm cut}\propto R$ later. If $R_\gamma<R_m$ ($\Gamma<\Gamma_c$), in the beginning the attenuated energy is a constant, $\varepsilon_{\rm cut}=\varepsilon_{\rm cut}^{(2)}$ at $R<R_m$, and turns to be $\varepsilon_{\rm cut}\propto R$ at $R>R_m$.[]{data-label="fig:cutvsR"}](fig1.EPS){width="\columnwidth"}
Comparing $\varepsilon_{\rm IC}$ and $\varepsilon_{\rm cut}$ it can be found that typically $\varepsilon_{\rm IC}>\varepsilon_{\rm
cut}$ at small radii, $R\la30R_\gamma$. In this case the bulk IC radiation is absorbed in the source, leading to electromagnetic cascades, and the photons escape until their energies decay to $\varepsilon_{\rm cut}$. Therefore the bulk high energy radiation is just re-emitted at $\varepsilon_{\rm cut}$. On the other hand, $\varepsilon_{\rm IC}<\varepsilon_{\rm cut}$ at large radii $R\ga30R_\gamma$, where only the high energy emission below $\varepsilon_{\rm cut}$ appears. The emission above $\varepsilon_{\rm cut}$ is truncated, and undergoes electromagnetic cascades, but does not affect much the apparent spectrum since the photon spectrum decreases rapidly with energy.
Finally consider the emission flux. It is straightforward to show that the cooling time of the electrons is short compared to the dynamical time during the late residual collision phase, up to radii $R\sim10^3R_\gamma$. We therefore assume that electrons radiate away all their energy. When $\varepsilon_{\rm
IC}>\varepsilon_{\rm cut}$ at small radii, the total electron energy appears at $\varepsilon_{\rm cut}\simeq\varepsilon_0$ if $R_\gamma<R<R_m$ or $\varepsilon_{\rm cut}\propto R$ if $R>\max[R_\gamma,R_m]$. The observed (time-integrated) IC spectrum at energy $\varepsilon>\varepsilon_0$ would be $\nu F_\nu\propto
E_{\rm fluc}|_{\nu_{\rm cut}=\nu}\propto R^{-2/3}|_{\nu_{\rm
cut}=\nu}\propto\nu^{-2/3}$ (Here $\nu=\varepsilon/h$).
When the outflow expands to large radii where $\varepsilon_{\rm
IC}<\varepsilon_{\rm cut}$, we need to consider the electrons accelerated to Lorentz factors larger than the characteristic Lorentz factor $\gamma_i$. Shock acceleration is expected to generate a power-law energy distribution of electrons $dn_e/d\gamma_e\propto\gamma_e^{-p}$ at $\gamma_e>\gamma_i$ with $p\simeq2$ [@BE87]. This flat-electron energy distribution, $\gamma_e^2dn_e/d\gamma_e\propto\gamma_e^0$, generates equal amounts of IC energy in logarithmic photon energy intervals, $\nu
F_\nu\propto \nu^0$ for $\nu>\nu_{\rm IC}$. So $\nu F_\nu(\nu_{\rm
cut})\simeq\nu F_\nu(\nu_{\rm IC})\propto E_{\rm fluc}$. The emission at low energy would be covered by earlier emission, while only the emission at high energy end, i.e. around the cutoff, shows up and interests us. The observed (time-integrated) spectrum would be similar to the $\varepsilon_{\rm IC}>\varepsilon_{\rm
cut}$ case, i.e., $\nu F_\nu\propto \nu^{-2/3}$.
Thus we expect the observed (time-integrated) prompt spectrum, above the spectral cutoff energy in the prompt [$\gamma$-ray ]{}emitting phase, $\varepsilon_0$ (eq. \[eq:cut\]), to be $$\label{eq:spec}
\nu F_\nu\propto\nu^{-2/3}, ~~h\nu>\varepsilon_0~~~(\rm simplified ~~case).$$ This fluence spectrum is resulted from summing up all emission components from different radii and times. A schematic plot of the prompt GRB spectrum is shown in fig. \[fig:spectrum\]. Note, the spectral slope $\nu^{-2/3}$ here is derived from the simple case, which has been confirmed by recent numerical calculation by [@Aoi09]. If in general $E_{\rm fluc}\propto R^{-q}$ we would expect the slope to be $\nu F_\nu\propto \nu^{-q}$.
Below $\varepsilon_0$ is the observed prompt MeV [$\gamma$-ray ]{}spectrum, i.e., typically $\nu F_\nu\propto\nu$ below $\eps_b$ and $\nu
F_\nu\propto\nu^0$ between $\eps_b$ and $\varepsilon_0$. Note, the transition of the emission from primary to residual collisions at $\varepsilon_0$ is smooth if $R_\gamma>R_m$, as shown by the dashed dot line. The transition for the case of $R_\gamma<R_m$ is discontinuous as shown by the thick solid line. The power law described by eq. (\[eq:spec\]) starts with a flux lower than the primary emission at $\varepsilon_0$ by a factor of $(R_m/R_\gamma)^{2/3}$.
Some comments should be made here. The spectral form described in eq. (\[eq:spec\]) holds only on average, especially for the high energy range. In individual GRB events the flux may differ significantly, because, for a small number of shells (and collisions) at large radii, large variations in the late residual collisions should be expected.
It should also be noticed that we have assumed the initial variance $\sigma_{\Gamma,0}<\Gamma$, whereas initial condition with $\sigma_{\Gamma,0}>\Gamma$ may lead to more efficient [$\gamma$-ray ]{}production [e.g. @Blbrdv00] around $R_\gamma$, in which case the ratio between fluxes of primary and residual emission at $\varepsilon_0$ should be larger by a factor of a few[^2], leading to more abrupt transition between the primary and residual emission.
One may worry about that the IC emission may be reduced as the seed photons are not isotropic in the shock frame of residual collisions. However, this kind of geometry effects do not play an important role even if the photons are completely collinear [@Wang06]. As usually assumed by many authors, suppose that the electrons accelerated in residual collision shocks are isotropic in the rest frame of the outflow, since the tangled magnetic fields in the shock might sufficiently isotropize electrons. Thus an electron is changing its angle $\theta'$ with respect to the photon beam and cooling fast. The IC power of an electron averaged over its cooling time is not different from interacting with isotropic photons. As long as the jet effect is not important to prompt GRB emission ($\theta_{\rm
jet}\gg1/\Gamma$), we can furthermore regard the GRB explosion as isotropic, and hence the observer at different angles will observe the same IC emission due to spherical symmetry. Consider both cases of isotropic and anisotropic scatterings, the radiated energy can be assumed to be the same because it is determined by the total electron energy if electrons radiate all their energy rapidly. The outside observers would observe the same fluence in both cases, otherwise one can simply ask where the electron energy have gone, given the same total electron energy. Thus, the IC fluence is not reduced by this geometry effect if electrons are isotropic distributed in the rest frame and radiate all their energy within a dynamic time. Nevertheless, this effect changes the angular distribution of IC emission. In the rest frame the IC power becomes $P_{\rm IC}'\propto(1-\cos\theta')$, although not much different from isotropic distribution. Correspondingly, in the lab frame the “image” of the IC emission is different from the isotropic case, i.e. the anglular dependence of the brightness is different, but the angular-integrated fluence is the same.
### Time delay
At energy $\varepsilon<\varepsilon_0$, the emission is mainly contributed by primary collisions at small radius, and arrives at detectors simultaneously with MeV emission. However, for higher energy $\varepsilon>\varepsilon_0$, the emission is produced at relatively large radii, and should have a time delay relative to the primary MeV emission. Since the spectral cutoff energy, where the high energy photons emerge, increases as the outflow expands, $\varepsilon_{\rm cut}\propto R$ for $\varepsilon_{\rm
cut}>\varepsilon_0$ (eq. \[eq:cut evolution\]), the time delay increases with observed photon energy. For energy $\varepsilon=\varepsilon_{\rm cut}(R)>\varepsilon_0$, the typical radius where photons emerge is $R\simeq[\varepsilon/\varepsilon_0]\max[R_\gamma,\,R_m]=[\varepsilon/\varepsilon_{\rm
cut}^{(1)}]R_\gamma$, therefore the related time delay relative to MeV [$\gamma$-ray ]{}emission, $\tau_{\rm delay}\simeq
(R-R_\gamma)/2\Gamma^2c$, is $$\label{eq:delay}
\tau_{\rm delay}^{\rm ob}\simeq\frac{\varepsilon}{\varepsilon_{\rm
cut}^{(1)}}t_{\rm
var}=2.2L_{52}\Gamma_{2.5}^{-6}\frac{\varepsilon^{\rm ob}}{\rm
10^2~GeV}(1+z)^2\rm
s~~(for ~\varepsilon^{\rm ob}>\varepsilon_0^{\rm ob}).$$ Thus the emission at $\varepsilon>\varepsilon_0$ is delayed later as $\varepsilon$ increases. A comment is made here that the target photons for scattering may be beamed with respect to electrons, which changes the angular distribution of IC emission, i.e. the maximum IC power may come from a certain angle other than $\theta=0$. This leads to an additional time delay. However, as implied by eq. (\[eq:max\_angle\]), the angle where the maximum power is emitted is smaller than $\pi/2$ in the comoving frame and hence $\theta<1/\Gamma$. The produced time delay is smaller than $\tau_{\rm delay}$, $R\theta^2/2c<R/2\Gamma^2c\sim \tau_{\rm
delay}$. Thus we neglect this additional time delay.
Eq.(\[eq:delay\]) implies that the detection of time delay $\tau_{\rm delay}$ at $\varepsilon$ helps to determine the Lorentz factor, $$\label{eq:Gamma_delay}
\Gamma=167L_{52}^{1/6}{\left(\frac{\varepsilon^{\rm ob}}{1\rm\,GeV}\right)}^{1/6}{\left(\frac{\tau_{\rm delay}^{\rm ob}|_{\varepsilon^{\rm ob}}}{1\,\rm
s}\right)}^{-1/6}(1+z)^{1/3}.$$ This determination by time delay should be consistent with that by detection of the spectral transition $\varepsilon_0$ between primary and residual emission, eq.(\[eq:Gamma\_cut\]).
![Schematic plot of the predicted $\nu F_\nu$ spectrum of prompt high energy emission from a GRB. The [*thin solid*]{} line shows the observed MeV [$\gamma$-ray ]{}emission, a broken power law with a break energy at $\eps_b\sim$ MeV, above which the spectrum goes as $\nu F_\nu\propto\nu^0$. The [*dot*]{} lines mark the break energy $\eps_b$ and the previously thought pair-production spectral cutoff, $\varepsilon_0\sim$ GeV (eq. \[\[eq:cut\]\]). The residual collisions at large radii contribute beyond $\varepsilon_0$. Summing up all emission components (the [*dashed*]{} lines) from different radii and times lead to a spectral slope $\nu
F_\nu\propto\nu^{-q}$ ($q$ is the index of the random energy evolution, $E_{\rm fluc}\propto R^{-q}$, and $0<q\leq2/3$). In the simplified case, $q=2/3$ (see the text). Here the [*thick solid*]{} line corresponds to the outflow satisfying $R_\gamma<R_m$ (or $\varepsilon_0\simeq\Gamma m_ec^2$), while the [*thick dashed dot*]{} line corresponds to $R_\gamma>R_m$ ($\varepsilon_0\ga\Gamma m_ec^2$). The $\nu F_\nu$ values in these two are different by a factor of $(R_m/R_\gamma)^{2/3}$. The spectral slope holds up to TeV range, but not higher (see the discussion section). []{data-label="fig:spectrum"}](fig2.EPS){width="\columnwidth"}
Application: Fermi-LAT GRB 080916c
==================================
As the Fermi observational data showed up after the first version of this paper was posted on the archive,[^3] we now apply the model to the observations.
[@Fermi916c] reports the measurements of the bright GRB 080916c by Fermi GBM and LAT. The redshift of this burst is $z=4.35$ [@afterglow916c], which implies, with flucence (10 keV$-$10 GeV) $\approx2.4\times10^{-4}\rm ~erg\,cm^{-2}$, the largest reported isotropic [$\gamma$-ray ]{}energy release, $E_{\rm
iso}\simeq9\times10^{54}$erg. The observed GRB duration is $T\approx50$ s, so the bolometric isotropic-equivalent luminosity is $L_{\rm bol}=E_{\rm iso}(1+z)/T\approx10^{54}\rm erg\,s^{-1}$. As the observed peak energy is $\eps_b/(1+z)\sim1$ MeV and the high energy slope is $\beta\approx-2$, the MeV luminosity, defined as the luminosity at $\la2\eps_b$, is $L\approx L_{\rm
bol}/\ln(10\,\rm GeV/1\,MeV)=10^{53}\rm erg\,s^{-1}$. The LAT detected 145 photons at $>100$ MeV, within which 14 are beyond 1 GeV, during the first 100 s after the trigger. The brightness may have enough statistics for spectral and temporal analysis of the high energy properties.
There are several interesting properties in the high energy emission of this GRB.
#### Time delay
[*The multi-band light curves unambiguously show that the bulk of the emission of the second light-curve peak is moving toward later times as the energy increases [see time bin b in Fig 1 and its inset panels in @Fermi916c], and the time delay of $100$-MeV emission is about 1 s relative to MeV emission, much larger than the MeV variability timescale, $\la100$ ms [@afterglow916c].*]{}[^4] First of all, these are qualitatively consistent with our prediction that the higher energy photons can only arise when the plasma expands to larger size in later time where the $\gamma\gamma$ optical depth reduces to below unity, and that the size of high energy emission can be much larger than MeV emission.
Let us consider the data quantitatively, and constrain the model parameters. The LAT >100 MeV detection consists of 145 photons which mainly come up in a single light-curve peak, therefore we have enough statistics for the time analysis of >100 MeV emission. It is obviously seen that the >100 MeV light curve peak has a time delay $\tau_{\rm delay}^{\rm ob}=\tau_{\rm
delay}(1+z)\sim1$ s relative to that of 250 keV-5 MeV (The script “ob” denotes quantities measured in the observer frame, with redshift effect taken into account). This implies $\varepsilon_0^{\rm ob}=\varepsilon_0/(1+z)<100$ MeV. Substituting the observed values of $L=10^{53}\rm erg\,s^{-1}$, $\tau_{\rm
delay}\simeq1/5.35$ s and $\varepsilon=100\times5.35=535$ MeV, with redshift $z=4.35$ taken, into eq.(\[eq:delay\]), we obtain the bulk Lorentz factor of GRB 080916c outflow, $$\label{eq:916c_LF}
\Gamma\simeq290{\left(\frac{L}{10^{53}\rm erg\,s^{-1}}\right)}^{1/6}{\left(\frac{\tau_{\rm delay}^{\rm
ob}|_{\rm 100\,MeV}}{1~\rm s}\right)}^{-1/6}.$$ This result is similar to those determined in other GRBs, $\Gamma\approx100-400$, through observations of the rising of optical afterglows [@LFopt1; @LFopt2; @LFopt4; @LFopt3], and the thermal components in the prompt emission [@peer07].
The determination of $\Gamma$ can be double checked by the location of $\varepsilon_0$. Using the result of eq.(\[eq:916c\_LF\]), we obtain $$\varepsilon_{\rm cut}^{(1)\rm
obs}\simeq10(t_{\rm var}^{\rm ob}/10^2\rm ms)(\tau_{\rm
delay}^{\rm ob}|_{\rm 100\,MeV}/1\,\rm s)^{-1}\rm MeV$$ and $$\varepsilon_{\rm cut}^{(2)\rm ob}\simeq40(L/10^{53}\rm
erg\,s^{-1})^{1/6}(\tau_{\rm delay}^{\rm ob}|_{\rm
100\,MeV}/1\,\rm s)^{-1/6}\rm MeV.$$ The observed MeV variability timescale is $t_{\rm var}^{\rm ob}\la$100 ms based on the INTEGRAL observation [@afterglow916c]. Thus $\varepsilon_0^{\rm
ob}\sim40$ MeV, consistent with requirement $\varepsilon_0^{\rm
ob}<100$ MeV. In addition, the broad light curve peak in the no-energy-selection band of LAT is consistent with, or a little delayed from, that of the GBM (260 keV-5 MeV) light curve, and is ahead of the >100 MeV peak. Thus $\varepsilon_0^{\rm ob}$ should be located in the LAT energy window (no selection) and below 100 MeV, consistent with the result $\varepsilon_0^{\rm
ob}\sim40$ MeV.
By our model the >1 GeV emission should be even 10 times longer delayed than the >100 MeV one, i.e., $\tau_{\rm delay}^{\rm ob}|_{1\,\rm GeV}\sim10$ s. The much fewer photons above 1 GeV prevent us from analyzing the temporal properties with high confidence. However the LAT >1 GeV light curve does agree with a longer delay by $\sim10$ s.
It should be noticed that other authors also constrain the bulk Lorentz factor of this GRB and obtain much larger lower limit [@afterglow916c; @Fermi916c]. Essentially, the difference is due to different models considered; they consider the GeV emission produced in the same time and place as the MeV emission, whereas in our model the GeV emission comes from delayed residual collisions at large radii, therefore our model looses the constraint on $\Gamma$. In addition, we consider that the cutoff energy should not locate below $2^{1/2}\Gamma m_ec^2$, which is ignored in [@afterglow916c] and [@Fermi916c].
#### Lack of the first LAT light-curve peak
The low energy GBM light curve shows two peaks, however the LAT observations only show one peak related to the second GBM peak and there is a paucity of emission in the first $\sim4$ s after the trigger. Note, some people call this as a “delayed onset” of high energy emission. In principle, one of the explanations could be that there is a spectral cutoff at $\sim10$ MeV for the prompt emission from primary collisions in the first 4 s, and the residual-collision emission at >10 MeV is $\ga4$ s delayed, longer than the second peak. If so, the properties of the ejecta emitting the first GBM peak are different from the later ejecta, which also suggests that there might be long-timescale, a-few-second ($\gg
t_{\rm var}\sim1$ ms), variabilities in the outflow of this GRB.
#### Time-integrated spectrum
The joint GBM-LAT spectrum of GRB 080916c can be fit by Band function [@Band93], with peak energy around 1 MeV, $\alpha\approx-1.0$ and $\beta\approx-2.2$, except for the first 4 s [@Fermi916c]. Because the time intervals used to construct the spectra are much longer than the MeV variability time, the resulted spectra are all time-integrated ones. Since the synchrotron self-Compton (SSC) model for GRBs, where the MeV peak is from the IC scattering by soft photon emitting electrons, predicts a bright GeV-TeV component due to the second order up-scattering [@Piran08], no evidence for high energy bump up to 10 GeV in observations does not favor SSC but synchrotron model [@Wang09]. In addition, the narrow $\nu F_\nu$ spectral peak of GRB spectra favor more synchrotron emission mechanism over SSC, since SSC usually has a much broader spectral bump [@Baring04].
In the framework of synchrotron internal shocks, our residual collision model predicts a slight spectral softening at high energies. However, due to the small detected GeV-photon number the Poisson scatter of low statistic still allows a slight softening at tens MeV to fit the data. Furthermore, if there are large timescale, $\gg1$ ms, variabilities in the outflow so that the residual emission spectral slope $q$ is larger than $2/3$ (see §\[section:simulation\]), the high-energy spectrum is less steepened and is closer to a single power law. Finally, the slope of eq.(\[eq:spec\]) holds on average, while the later residual collisions occur between smaller number of shells, thus there might be fluctuation from this average slope.
#### Highest energy emission
The highest energy photon is detected with 13.2 GeV only 17 s after the GRB trigger. With the redshift $z=4.35$ this suggests GRB 080916c produces radiation up to 71 GeV in the source frame. Moreover, LAT detects 145 photons with $>100$ MeV, within which 14 with $>1$ GeV and especially only one with $>10$ GeV, consistent with a power law spectrum with photon index $\beta\approx-2$ up to $\sim10$ GeV scale. There might be emission extending to higher energy, say, beyond tens of GeV, from GRB 080916c following the same slope, but the detection rate is less than one, i.e., no photon would be detected at this energy. Thus the observations actually suggest that the high energy spectral cutoff (or steep drop), if there exists, is more likely to be far above the energy of the only observed highest energy photon, $\varepsilon_{\max}^{\rm ob}=13.2\times g$ GeV with $g\gg1$. If the high energy emission beyond 13.2 GeV is produced by internal shocks, we argue here that it may not be produced by primary internal collisions that emit MeV [$\gamma$-rays ]{}but produced in other regions, e.g., by residual collisions.
As said above no high energy spectral component in GRB 080916c does not favor SSC model but synchrotron model. Now calculate the maximum synchrotron photon energy. If $B$ is the magnetic field strength in the internal shock, the Larmor time of an electron with Lorentz factor $\gamma$ is $t_L'=\gamma m_ec/eB$. The typical particle acceleration time can be scaled by Larmor time as $t_a'=ft_L'$ [e.g. @Hillas84] where $f$ is a correction factor accounting for the uncertainty of shock acceleration. It might be that $f\ga$ a few [e.g. @Lemoine06]. In the same time the electron suffers synchrotron cooling in a typical timescale $t_c'=3m_ec/4\sigma_T\gamma(B^2/8\pi)$ (We neglect the IC cooling as the IC scattering usually occurs in deep Klein-Nishina regime for the most energetic electrons). The competition between acceleration and cooling results in a maximum Lorentz factor of accelerated electrons, $\gamma_{\max} =(6\pi
e/f\sigma_TB)^{1/2}$. The relevant synchrotron photon frequency is a constant,[^5] $$\nu_{\max}'=0.3\gamma_{\max}^2eB/2\pi
m_ec=0.9e^2/f\sigma_Tm_ec.$$ The coefficient accounts for the fact that the synchrotron power per unit frequency of an electron peaks at the frequency a fraction 0.3 of the common characteristic frequency. The maximum synchrotron photon energy $\varepsilon_{\max}^{\rm ob}=h\nu_{\max}'\Gamma/(1+z)$ is, therefore, $$\varepsilon_{\max}^{\rm ob}=15\Gamma_{2.5}f^{-1}(1+z)^{-1}\rm
~GeV.$$
Comparing the predicted maximum synchrotron energy with that implied by the observation, $\varepsilon_{\max}^{\rm
ob}=13.2g$ GeV, we have a lower bound, $$\Gamma=1.5\times10^4f_{0.5}g_{0.5},$$ where the conservative values, $f=10^{0.5}f_{0.5}$ and $g=10^{0.5}g_{0.5}$, have been taken. This bulk Lorentz factor is too large for fireball-shock model, because it faces several problems. First, the large $\Gamma$ leads to (primary) internal shock radius larger than the deceleration radius of GRB outflow [e.g. @Lazzati99]. The deceleration radius is $R_d\approx(E_k/4nm_pc^2\Gamma^2)^{1/3}$. If $R_\gamma<R_d$, an upper limit is obtained, $$\Gamma<7\times10^3(E_{k,55}/n_0)^{1/8}t_{\rm var,-2}^{-3/8},$$ where $E_k=10^{55}E_{k,55}$erg and $n=1n_0\rm cm^{-3}$ are the outflow kinetic energy and medium density, respectively. Second, the large $\Gamma$ raises problem of low energy conversion efficiency due to slow cooling of accelerated electrons [e.g., @Derishev01]. If the synchrotron cooling time of electrons with typical postshock Lorentz factor $\gamma_m\sim m_p/m_e\sim10^3\gamma_3$, is required to be smaller than the dynamical time of the outflow, $t_c'(\gamma_m)<t_d'\simeq
R_\gamma/\Gamma c$, we have $$\Gamma<5\times10^3\gamma_3^{1/5}(L_{\rm bol}/10^{54}{\rm erg\,s^{-1}})^{1/5}t_{\rm
var,-2}^{-1/5}.$$ In this calculation we have assumed that the postshock magnetic field is limited by observed emission, $B^2/8\pi\leq
U_\gamma=L_{\rm bol}/4\pi R_\gamma^2\Gamma^2c$. Third, in synchrotron internal shock models the large $\Gamma$ leads to large collision radius $R_\gamma\approx2\Gamma^2ct_{\rm var}$, where the magnetic field $B$ is too small to give rise high energy synchrotron photon energy. Using $\eps_b\approx\Gamma\hbar\gamma_m^2eB/m_ec$ and the limit $B^2/8\pi< U_\gamma$, the restriction to obtain synchrotron emission peaking at MeV range is $$\Gamma<0.4\times10^3\gamma_3(L_{\rm bol}/10^{54}{\rm erg\,s^{-1}})^{1/4}t_{\rm
var,-2}^{-1/2}(\eps_b/\rm 1\,MeV)^{-1/2}.$$ Finally, the thermal pressure of the initial fireball is not expected to accelerate the loaded baryons to very large Lorentz factor with most energy kept as the kinetic energy of baryons. The final Lorentz factor is limited to be $$\Gamma<3\times10^3(L_0/10^{54}{\rm erg\,s^{-1}})^{1/4}r_{0,7}^{-1/4},$$ where $L_0$ is the rate at which the central source emits energy, and $r_0=10^7r_{0,7}$ cm is the source size [see, e.g @Waxman; @rev].
The above upper limits to $\Gamma$ imply that it may be impossible that $\Gamma\gg10^3$. This appear not to match the value suggested by the highest energy band observation, $\Gamma=1.5\times10^4f_{0.5}g_{0.5}$, unless $f\approx1$ and $g\approx1$ are satisfied at the same time: the shock acceleration must operate at the Bohm limit; furthermore the observed highest energy photon happens to be at the maximum synchrotron energy. Thus the observations imply there is emission much higher than 13.2 GeV, which cannot be originated from synchrotron emission in the primary internal shocks. Actually, this high energy emission can be produced by IC emission in residual collision shocks, as discussed in present study. So the observation of highest energy emission supports the residual collision model.
In conclusions: (1) the time delay of high energy emission and the spectral feature of highest energy emission in GRB 080916c might have provide evidences for the residual collision model; (2) its spectrum is not inconsistent with the residual emission; (3) the time delay of $>100$ MeV emission constrains the bulk Lorentz factor to be $\Gamma\sim300$, a typical value usually taken. It appears to be an applicable method to determine $\Gamma$ of GRB outflows by measuring the time delays of LAT light curves. If internal shocks also work in short GRBs, we expect similar delayed, prompt high energy emission in short GRBs. We also caution more careful spectral analysis to find the transition between primary and residual emission.
Discussion
==========
We have considered the high energy emission in the prompt GRB spectrum, which is dominated by the IC emission from the residual collisions. Instead of a exponential spectral cutoff, a steeper, compared to the prompt MeV emission, power-law slope $\nu
F_\nu\propto \nu^{-q}$ is expected beyond the previously thought cutoff, $\varepsilon_0$ (eq. \[\[eq:cut\]\]). Here $q$ is corresponding to the dynamical evolution of the random energy in the outflow, $E_{\rm fluc}\propto R^{-q}$. In the simplified case (see §3.1), which is consistent with optical flash observations (LW08), we take $q=2/3$, while $0< q\leq 2/3$ in general. The extended emission makes it complicated to detect the “cutoff energy” in the goal to constrain the GRB emission region.
Indeed, EGRET had detected prompt high energy photons past GeV in several brightest BATSE GRBs occurring in its field of view (e.g., GRB 930131, [@930131]; GRB 940217, [@Hurley94]), which suggest that the other faint GRBs may produce prompt GeV photons as well [@Dingus95]. There is also no sign of cutoff in the spectra [@Dingus95], which, if there is, should be far exceeding $\sim1$ GeV. These EGRET results are consistent with the predicted extension of prompt emission beyond GeV. However, the cutoff is not ruled out. Given the sensitive dependence of the cutoff on $\Gamma$ (eq.\[eq:cut\]), a slight variation of $\Gamma$ in individual GRB events may lead to much higher cutoff energy, $\gg1$ GeV, explaining the prompt GeV emission in EGRET-detected GRBs. Two properties may help to discriminate our residual-collision emission model from a very high energy cutoff model. The first is the steepening turnover in the spectrum. For typical Lorentz factors, $\Gamma\approx10^2-10^3$, the expected spectral turnover is $\rm \sim100~MeV-1~TeV$, well located in the windows of Fermi and AGILE. The second is the time delay of high energy emission. One may expect systematic time delay of the high energy photons in the residual emission model, while no delay is expected in the very high energy cutoff model. However, the task is not easy given that for a typical event with fluence $\sim10^{-6}\rm erg~cm^{-2}$ the observed GeV photon number is only a few. In order to have enough statistics for the spectral and temporal analyses, very bright events are needed, or one may integrate many events to obtain an average burst.
We have discussed that the recent Fermi detected bright GRB 080916c might have presented a good sample. Observations do not show a simple spectral cutoff, but a spectral tail up to 70 GeV in GRB frame. More than one hundred of photons detected above 100 MeV makes it obviously showing a time delay about 1 sec, which can be explained by the residual IC emission and results in the determination of a typical Lorentz factor value, $\Gamma\sim300$. The features of GRB 080916c support the residual-collision emission model as opposed to the very-high-energy-cutoff model.
The high energy emission would not extend to very high energy. There are several effects that lead to a drop in TeV range. First, when the plasma expands to very large radius, $R\ga2\Gamma^2cT\sim10^{17}\Gamma_{2.5}^2(T/10{\rm~s})$ cm ($T$ is the MeV [$\gamma$-ray ]{}duration), there would be no overlapping between the plasma and the MeV [$\gamma$-rays ]{}, and hence no scattering is expected. At $R\simeq2\Gamma^2cT$ the cutoff energy increases to $\varepsilon_{\rm cut}\simeq1(T/10{\rm~s})t_{\rm var,-2}^{-1}$ TeV (from eq.\[eq:cut evolution\]). A lack of $\ga1$ TeV photons would be expected for GRBs with duration $T\la10t_{\rm var,-2}$ s, although it should not be an exponential cutoff. Second, in interaction with $\sim1$ MeV photons, the Klein-Nishina limit becomes important for electrons with $\gamma_e\ga\Gamma$, giving rise to IC photons up to $\varepsilon_{\rm
IC}\simeq0.1\Gamma_{2.5}^2$TeV, beyond which the spectrum gradually turns below the low energy slope. Finally the cosmic infrared background would absorb the $>0.1$ TeV photons that arrive from GRBs at redshift $z\ga0.5$. Except for low redshift events, the observed prompt GRB emission may not extend far beyond TeV range.
It should be commented here that in the framework of the synchrotron internal shock model [@Waxman; @rev] we do not expect a bright high energy component, say $\ga 1$ GeV, in the prompt emission, other than the synchrotron self-Compton model [e.g. @Piran08], therefore the residual high energy emission will be dominant. A detection of high energy component in the prompt emission will be an evidence against the synchrotron model for GRBs, and vice verse. The recent Fermi observations of several GRBs do not support a high energy component in GRB spectrum, since the $>100$ MeV fluences are all less than those in MeV range [@Fermi916c].
Our residual collision model is not expected to produce much longer delayed high energy emission which is not apparently overlapped with the primary MeV [$\gamma$-rays ]{}in times. There are more and more observations showing that GRBs produce delayed $>100$ MeV emission even after the prompt [$\gamma$-rays ]{}end, lasting tens or even $\sim10^4$ seconds [@Hurley94; @dermer03; @Giuliani08]. This may require some long-lasting central engine activities or external productions [e.g. @Wang06 and references therein].
This work was partly supported by the National Natural Science Foundation of China through grants 10473010 and 10843007.
Dynamics with multi-timescale variabilities {#section:simulation}
===========================================
In order to show the effect of multi-timescale variabilities on the dynamics of the outflow, we carry simulations for both cases of single- and multi-timescale variabilities for comparison. We consider a series of individual material shells $i=1,2...,N$, with total shell number $N=3000$, released in a duration of $T=3$s, so that the interval between two nearby shells is $1$ ms. The shells have equal masses but different energies, with the bulk Lorentz factor of each shell following $$\log\Gamma_i=2+\xi_i\log9+\phi_i,$$ where $\xi_i$ is a random number between zero and unity and $\phi_i$ can be taken as the following forms, $$\phi_i=\left\{\begin{array}{ll}
0 & \rm (Single);\\
\sum_k \sin(\frac{2\pi Ti}{P_kN})\log A_k & \rm
(Multi).
\end{array}
\right.\label{eq:multi-scale}$$ For single-timescale case, we should take $\phi_i=0$ (case S), then the outflow has only one variability timescale of $10^{-3}$s. The Lorentz factors randomly and uniformly distributed in logarithmical scale between 100 and 900. For multi-timescale effect we consider 3 extra timescales besides the smallest timescale: $P_k=10^{-2},10^{-1},1$ s for $k=1,2,3$, with relevant $A_k$ values being $A_k=2,1.5,1.2$ (case M1) or $A_k=1.2,1.5,2$ (case M2) for $k=1,2,3$. $A_k$ decreases with $P_k$ in case M1 but increase in case M2, which means case M2 has larger fluctuations at larger timescales. Larger timescale fluctuations tend to produce strong collisions at larger radii so that larger fraction of energy is dissipated at larger radii. We further consider that in each two-shell collision, $1/3$ of the generated internal energy is emitted by radiation, because only the energy of shocked electrons is assumed to be emitted rapidly and the electron equipartition parameter is $\epsilon_e=1/3$. The two shells are considered to separate again after collision and share equally the remained internal energy (in the center-of-momentum frame of the two shells).
![The fraction of emission energy as function of radius in three simulations. The line marked with [*squares*]{}: case S, with only the 1-ms variability; [*circles*]{}. The other two are multi-timescale cases (M1 and M2), with three more scales of $P_k=10^{-2}, 10^{-1}$ and 1 s, for $k=1, 2,$ and 3, respectively. The line marked with [*circles*]{}: case M1, with $A_k=2, 1.5,$ and 1.2 for $k=1,2,$ and 3 (eq.\[\[eq:multi-scale\]\]); [*triangles*]{}: case M2, with $A_k=1.2, 1.5,$ and 2 for $k=1, 2,$ and 3 (eq.\[\[eq:multi-scale\]\]). The straight line show the $-2/3$ slop to guide eyes. See the relevant text for more details.[]{data-label="fig:simulation"}](fig3.EPS){width="\columnwidth"}
In Fig \[fig:simulation\] we show the fraction of emission energy $E_{\rm em}(>R)/E_{\rm em,tot}$, that is emitted beyond radius $R$. We see that case S show a slope close to the analytical resolution $\propto R^{-2/3}$ for single-timescale case by LW08. However, the multi timescales lead to flatter slopes, implying more fraction of emission energy tends to be emitted at larger radii. Case M2 has even flatter slope than case M1, since case M2 has relatively larger fraction of energy that is dissipated at larger radii. The steep drop at the end in both cases M1 and M2 means no more strong collisions later on. This is because there is a largest timescale of 1 s in our simulations. If there is still variabilities with timescale larger than 1 s then the slope will continue to even larger radii and show even later drop at the end. In summary, the simulations demonstrate that multi-timescale variabilities lead to a flatter slope $q<2/3$, and the $q=2/3$ slope is only for single-timescale case.
Anisotropic IC emission
=======================
Let us discuss at which energy the IC emission is emitted in the observer frame, taking into account the fact that the seed photons are beamed. Consider the extremely anisotropic case, where the photons are collinear in the comoving frame of the outflow. In this frame the electrons, as argued, is reasonably assumed to be isotropically distributed. For simplicity, we consider mono-energetic photons, since the photon number rapidly decreases with energy. Thus the IC power per unity solid angle in the comoving frame is angular dependent, $$\frac{\d P'}{\d \Omega'}\propto(1-\mu')^2,$$ where $\mu'=\cos\theta'$ with $\theta'$ the direction with respect to photon beam, and we have taken the velocity of the electron to be $\beta_e'\approx1$. Hereinafter prime denotes quantities in the comoving frame of the outflow, while non-prime denotes observer frame. Using the Lorentz transformation, $\mu'=(\mu-\beta_\Gamma)/(1-\beta_\Gamma\mu)$, where $\beta_\Gamma=(1-1/\Gamma^2)^{1/2}$, we have $$1-\mu'=(1+\beta_\Gamma)\frac{1-\mu}{1-\beta_\Gamma\mu},$$ then the angular distribution of IC power in observer frame is $$\frac{\d P}{\d \Omega}=\frac 1{\Gamma^4(1-\beta_\Gamma\mu)^3}\frac{\d P'}{\d \Omega'}
\propto\frac{(1-\mu)^2}{(1-\beta_\Gamma\mu)^5}.$$ This is not like the simple cone-like distribution of the isotropic-photon case. The maximum power per solid angle is emitted at angle with $$\mu_{\max}=\frac{5\beta_\Gamma-2}{3\beta_\Gamma}.$$ The corresponding angle in the comoving frame is given by $$\label{eq:max_angle}
1-\mu_{\max}'=\frac{2(1+\beta_\Gamma)}{5\beta_\Gamma}\approx\frac45.$$ The scattered photon energy in the comoving frame is given by $\varepsilon_{\rm IC}'\approx\gamma_e^2\eps'(1-\mu')$, with $\gamma_e$ the electron Lorentz factor and $\eps'$ the photon energy both in the comoving frame. The photon energy (in observer frame) emitted at angel $\mu_{\max}$, where the IC power is maximum, is then $$\varepsilon_{\rm IC}(\mu_{\max})=\frac{\varepsilon_{\rm
IC}'(\mu_{\max})}{\Gamma(1-\beta_\Gamma\mu_{\max})}=\frac{3\gamma_e^2\eps'}{5\Gamma(1-\beta_\Gamma)}
\approx\frac35\gamma_e^2\Gamma\eps'(1+\beta_\Gamma)\approx\frac35\gamma_e^2\eps,$$ where in the last equality $\eps=(1+\beta_\Gamma)\Gamma\eps'$ is taken for collinear photons. This is the observer-frame photon energy around which the IC emission is mainly emitted. We see that the anisotropic correction factor $\lambda$ in eq. (\[eq:ICph\_energy\]) is only order unity, implying that taking $\lambda\sim1$ is a good approximation even for highly beamed seed photons.
Abdo, A.A., et al. 2009, Science, in press
Akerlof, C., et al.1999, , 398, 400
Aoi, J., Murase, K., Takahashi, K., Ioka, K., & Nagataki, S. 2009, arXiv:0904.4878
Band, D., et al. 1993, , 413, 281
Baring, M. G. 2000, American Institute of Physics Conference Series, 515, 238
Baring, M. G., & Braby, M. L. 2004, , 613, 460
Baring, M. G., & Harding, A. K. 1997, , 491, 663
Beloborodov, A. M. 2000, , 539, L25
Beskin, G., Karpov, S., Bondar, S., Guarnieri, A., Bartolini, C., Greco, G., & Piccioni, A. 2009, arXiv:0905.4431
Blake, C. H., et al.2005, , 435, 181
Blandford, R., & Eichler, D. 1987, , 154, 1
Bloom, J. S., et al.2008, arXiv:0803.3215
D’Elia, V., et al.2008, arXiv:0804.2141
Derishev, E. V., Kocharovsky, V. V., & Kocharovsky, V. V. 2001, , 372, 1071
Dingus, B. L. 1995, , 231, 187
Giuliani, A., et al.2008, arXiv:0809.1230
Gonz[á]{}lez, M. M., Dingus, B. L., Kaneko, Y., Preece, R. D., Dermer, C. D., & Briggs, M. S.2003, , 424, 749
Granot, J., Cohen-Tanugi, J., & do Couto e Silva, E. 2008, , 677, 92
Greiner, J., et al.2008, arXiv:0811.4291
Greiner, J., et al.2009, arXiv:0902.0761
Gupta, N., & Zhang, B. 2007, , 380, 78
Gupta, N., & Zhang, B. 2008, , 384, L11
Hillas, A. M. 1984, , 22, 425
Hurley, K., et al.1994, , 372, 652
Kobayashi, S., Piran, T., & Sari, R. 1997, , 490, 92
Kr[ü]{}hler, T., et al. 2008, , 685, 376
Kr[ü]{}hler, T., et al. 2009, , submitted
Lazzati, D., Ghisellini, G., & Celotti, A. 1999, , 309, L13
Lemoine, M., & Revenu, B. 2006, , 366, 635
Li, Z., & Song, L. M. 2004, , 608, L17
Li, Z., & Waxman, E. 2007, arXiv:0711.4969
Li, Z., & Waxman, E. 2008, , 674, L65 (LW08)
Lithwick, Y., & Sari, R. 2001, , 555, 540
Lyutikov, M., & Blandford, R. 2003, arXiv:astro-ph/0312347
Molinari, E., et al. 2007, , 469, L13
Murase, K., & Ioka, K. 2008, , 676, 1123
Narayan, R., & Kumar, P. 2008, arXiv:0812.0018
Paczynski, B., & Xu, G. 1994, , 427, 708
Pe’er, A., Ryde, F., Wijers, R. A. M. J., M[é]{}sz[á]{}ros, P., & Rees, M. J. 2007, , 664, L1
Piran, T., Sari, R., & Zou, Y.-C. 2008, arXiv:0807.3954
Racusin, J. L., et al.2008, , 455, 183
Razzaque, S., M[é]{}sz[á]{}ros, P., & Zhang, B. 2004, , 613, 1072
Rees, M. J., & Meszaros, P. 1994, , 430, L93
Sari, R., & Piran, T. 1997, , 485, 270
Sommer, M., et al. 1994, , 422, L63
Vestrand, W. T., et al. 2005, , 435, 178
Vestrand, W. T., et al. 2006, , 442, 172
Wang, X.-Y., Li, Z., & M[é]{}sz[á]{}ros, P. 2006, , 641, L89
Wang, X.-Y., Li, Z., Dai, Z.-G., & Meszaros, P. 2009, arXiv:0903.2086
Waxman, E. 2003, Supernovae and Gamma-Ray Bursters, 598, 393
Woods, E., & Loeb, A.1995, , 453, 583
Wozniak, P. R., Vestrand, W. T., Panaitescu, A. D., Wren, J. A., Davis, H. R., & White, R. R. 2008, arXiv:0810.2481
Yost, S. A., et al. 2007, , 669, 1107
[^1]: Because a low energy turnover at $\eps_a\sim1$ keV is expected in GRB spectra due to synchrotron self absorption, very high energy photons with $\varepsilon\ga10^{16}L_{52}t_{\rm
var,-2}^{-1}\Gamma_{2.5}^{-2}(\eps_a/1\rm ~keV)^{-1}$eV still can escape from the GRB source (e.g. [@LW07]; see also [@razza04]).
[^2]: It is not expected that the initial variance of Lorentz factors is far exceeding the mean, $\sigma_{\Gamma,0}\gg\Gamma$.
[^3]: http://arxiv.org/abs/0810.2932
[^4]: Note, the time-delay issue here is different from what is called “delayed onset” by other authors. We concern indeed the delayed peaking time of high-energy emission, related to the delayed arrival of the bulk of high-energy emission.
[^5]: This upper bound for synchrotron energy is robust for any acceleration mechanisms involving electromagnetic processes, because the acceleration limit with $f=1$ is robust not only to Fermi shock accelerations but also to any particle accelerations through electromagnetic processes. Therefore this bound might be valid not only to internal shock models but also to electromagnetic-dominated models.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present a combined analysis of neutron scattering and photoemission measurements on superconducting [FeSe$_{0.5}$Te$_{0.5}$]{}. The low-energy magnetic excitations disperse only in the direction transverse to the characteristic wave vector $(\frac12,0,0)$, whereas the electronic Fermi surface near $(\frac12,0,0)$ appears to consist of four incommensurate pockets. While the spin resonance occurs at an incommensurate wave vector compatible with nesting, neither spin-wave nor Fermi-surface-nesting models can describe the magnetic dispersion. We propose that a coupling of spin and orbital correlations is key to explaining this behavior. If correct, it follows that these nematic fluctuations are involved in the resonance and could be relevant to the pairing mechanism.'
author:
- 'S.-H. Lee'
- Guangyong Xu
- 'W. Ku'
- 'J. S. Wen'
- 'C. C. Lee'
- 'N. Katayama'
- 'Z. J. Xu'
- 'S. Ji'
- 'Z. W. Lin'
- 'G. D. Gu'
- 'H.-B. Yang'
- 'P. D. Johnson'
- 'Z.-H. Pan'
- 'T. Valla'
- 'M. Fujita'
- 'T. J. Sato'
- 'S. Chang'
- 'K. Yamada'
- 'J. M. Tranquada'
title: 'Coupling of spin and orbital excitations in the iron-based superconductor FeSe$_{0.5}$Te$_{0.5}$'
---
The quest to understand the mechanism of high temperature superconductivity gained new momentum with the recent discovery of Fe-based superconductors.[@hoso09] Experimental studies of these materials have shown that by chemically tuning the carrier density one can obtain both magnetically ordered and superconducting phases.[@ishi09] The superconducting transition temperature is maximized for conditions close to where magnetic order is suppressed, leading to predictions that magnetic fluctuations are important for electron pairing and motivating comparisons with copper-oxide superconductors. Initial theories have treated the magnetic excitations as independent of the atomic orbital character of the conduction electron states,[@mazi08; @kuro08] an assumption that seems to work well in other metallic magnetic systems; however, there have been recent proposals that coupled spin and orbital order occur in the antiferromagnetic state.[@krug09; @lee09; @lv09; @turn09; @cric09; @chen09]
In this paper, we present a study of low-energy magnetic and electronic excitations in superconducting [FeSe$_{0.5}$Te$_{0.5}$]{}. Particularly striking is the anomalous anisotropic dispersion of magnetic excitations, which is quite distinct from the spin-wave-like excitations typically seen in magnetic metals. Angle-resolved photoemission spectroscopy (ARPES) reveals an unexpected splitting of larger Fermi-surface pockets into smaller pockets with distinct orbital characters. A simple nesting picture does not appear to be compatible with the magnetic dispersion. We propose, instead, that a strong coupling of spin and orbital correlations can lead to the observed anisotropy. Such a coupling would suggest that electronic nematic fluctuations may be present in the normal state and may participate in the superconductivity. This would be a rather different way to connect with the cuprates.[@frad09]
Before continuing, let us first resolve our choice of coordinates. The structure of [FeSe$_{0.5}$Te$_{0.5}$]{} contains layers of Fe atoms forming a square lattice, with Se/Te atoms centered above or below these squares, alternating in a checkerboard fashion \[Fig \[fg:latt\](a)\]. The c-axis displacements of the Se/Te atoms break the translational symmetry, doubling the unit cell size; however, as we will show, this effect is rather weak for the low-energy electronic structure involving Fe-derived orbitals. Hence, we choose to work with a unit cell containing one Fe atom ($a=b=2.69$ Å and $c = 6.27$ Å). Reciprocal lattice vectors will be specified in units of $(2\pi/a,2\pi/a, 2\pi/c)$.
It is already experimentally established that the low energy magnetic excitations in superconducting [FeSe$_x$Te$_{1-x}$]{} occur near the in-plane wave vector ${\bf Q}=(\frac12,0)$,[@qiu09] which is the same as the magnetic ordering wave vector found in $R$FeAsO ($R=$ La, Ce, Pr, Nd) and $A$Fe$_2$As$_2$ ($A=$ Ca, Sr, Ba),[@lynn09] where it corresponds to columnar antiferromagnetic order \[Fig. \[fg:latt\](b)\]. (Note that the magnetic ordering in Fe$_{1+\delta}$Te is different.[@bao09; @li09]) Several recent theoretical analyses have proposed that orbital ordering, involving the Fe $3d_{xz}$ and $3d_{yz}$ states, is an essential feature of the magnetically-ordered state.[@krug09; @lee09; @lv09; @turn09; @cric09; @chen09] This is consistent with the fact that a symmetry-lowering structural transition, which breaks the degeneracy of the $d_{xz}$ and $d_{yz}$ states, always occurs before magnetic ordering.[@lynn09] Very recent experiments emphasize the electronic anisotropy of this state,[@fern09; @nand10; @chu09] demonstrating its nematic character [@fang08; @chua10].
For the experimental studies, single crystals of [FeSe$_{0.5}$Te$_{0.5}$]{} were grown at Brookhaven by a unidirectional solidification method. The nominal composition had no excess Fe, and $T_c$, measured by magnetic susceptibility, is 14 K.[@wen09] The ARPES spectra were recorded on beamline U13UB at the National Synchrotron Light Sounce using a Scienta SES2002 electron spectrometer. The incident photon energy was 17 eV, while the photoelectron energy and angular resolution was 15 meV and 0.1$^\circ$, respectively. Samples (small pieces from the crystal studied with neutrons) were cleaved [*in situ*]{} (base pressure of $5\times 10^{-11}$ Torr) and held at $T=15$ K for the measurements. Complementary ARPES data were collected on beamline 12.0.1 at the Advanced Light Source (ALS), using 50-eV photons (energy resolution of 25 meV) and $T=15$ K.
The neutron scattering experiments were performed at the cold-neutron triple-axis spectrometer SPINS located at the National Institute of Standards and Technology Center for Neutron Research (NCNR). The 9-g crystal was aligned in the $(hk0)$ scattering plane; a mosaic width of $<0.6^\circ$ (limited by resolution) was measured. The incident neutron energy was selected with a vertically-focusing pyrolytic graphite (PG) (002) monochromator. Scattered neutrons were analyzed with five 2.1 cm $\times$ 15 cm PG(002) flat analyzer blades that reflected neutrons with $E_f = 5.0$ meV onto a $^3$He proportional counter. A cooled Be filter was placed between the sample and analyzer to suppress higher order neutron contamination.
Let us consider the ARPES results first. The electronic spectral weight within 0–10 meV of the Fermi energy, $E_F$, is plotted in Fig. \[fg:ph\](a); the positions of identified band crossings, determined from energy-momentum cuts as in Fig. \[fg:ph\](b), are indicated by thick red lines. For comparison, our local density calculation of the Fermi surface is shown in Fig. 2(c) and the back-folded band structure in Fig. 2(d). Several features are obvious. First of all, we can see that the effect of the unit cell doubling due to the Se/Te positions is small. To appreciate this, note from Fig. \[fg:latt\](c) that for the doubled unit cell the $\Gamma$ and M points correspond to the same crystal momentum. While the band dispersions must be the same at both points, the spectral weights of the individual bands need not be,[@ku10] and that is certainly the case here. Indeed, in Fig. \[fg:ph\](d) one can barely recognize the relationship between M and $\Gamma$ points, reflecting weak Umklapp effects on the Fe-orbitals.
Secondly, experiment indicates the presence of small, incommensurate pockets about the X and Y points, in contrast to the larger commensurate pockets predicted by our calculation. The spectral intensities of these pockets are quite sensitive to the orientation of the photon polarization. We can understand this dependence as follows. The theoretical pockets about X (and Y) have distinct orbital character, as indicated in Fig. \[fg:ph\](c) and (d). Fe $d_{yz}$ character (red) dominates near $(\frac12,\pm k_0,0)$, $d_{xz}$ (blue) near $(\pm h_0,\frac12,0)$, and $d_{xy}$ around $(\frac12-h_0,0,0)$ and $(0,\frac12-k,0)$. Due to the broken cubic symmetry at the Fe sites and the hybridization with the neighboring Te/Se $p$ orbitals, $d_{yz}$/$d_{xz}$/$d_{xy}$ Wannier orbitals actually contain $p_x$/$p_y$/$p_z$ character (see Ref. for a related analysis of Wannier functions). Assuming that the ARPES matrix elements are primarily $p\rightarrow s$ (which is reasonable for low photon energies), the association between orbital character and polarization direction is: $d_{yz}$ : \[100\], $d_{xz}$ : \[010\], $d_{xy}$ : \[001\]. From the observed intensities, we conclude that the pockets at $(\frac12,\pm k_0,0)$ are $d_{yz}$-like, those at $(\pm h_0,\frac12,0)$ are $d_{xz}$-like, and those at $(0,\frac12\pm k_0,0)$ \[and, by symmetry, at $(\frac12\pm h_0,0,0)$\] are $d_{xy}$-like. The orbital character of the spectral weight at $E_F$ is qualitatively compatible with calculations.[@ragh08; @lee09] At the same time, the observed small hole pockets are different from the larger Fermi surfaces of the local density calculation, probably due to the inability of this mean-field technique to properly account for dynamical magnetic correlations.
Now let us turn to the neutron scattering results. As already mentioned, a spin resonance has been observed to develop below $T_c$ in [FeSe$_x$Te$_{1-x}$]{} with $x=0.4$ at the in-plane wave vector ${\bf Q}=(\frac12,0)$ and an energy of 6.5 meV.[@qiu09] We have confirmed the presence of a resonance in our $x=0.5$ sample. What we wish to focus on here is the [**Q**]{} dependence of the magnetic response in the vicinity of the resonance. In the superconducting phase ($T=1.5$ K $<<T_c$), transverse scans along $(0.5,k)$ \[scan A in Fig. \[fg:latt\](d)\] exhibit pairs of peaks at finite $|k|$ that disperse with energy above 5 meV, as shown in Fig. \[fg:ns2\](a). The color-coded map of intensity vs. [**Q**]{} at the resonance energy of 6.5 meV, Fig. \[fg:ns2\](b), demonstrates an intriguing anisotropy: the transverse peaks are not reproduced along the longitudinal $(0.5+h,0)$ direction as one would expect from Fermi surface nesting between a pocket at $\Gamma$ and the four pockets about X.
In the normal state ($T=20$ K), the spectral weight of the resonance is moved to lower energies, and the peaks appear to remain split down to 4 meV \[blue symbols in Fig. \[fg:ns2\](a)\]; the asymmetry between transverse and longitudinal directions remains the same as below $T_c$. The dispersion of the peaks along the transverse direction is plotted in Fig. \[fg:ns2\](c), with points at 10 meV and above from Lumsden [*et al.*]{}[@lums10] (For similar recent results, see Refs. .) The line through the points is a fit to $\hbar\omega = E_0 \sin (\pi (k- k_0))$ with $E_0 = 121$ meV for $k<k_0 = -0.09$. If these excitations were like spin waves, as in CaFe$_2$As$_2$,[@dial09; @zhao09] we would expect to see cone-shaped dispersions coming out of $(0.5,\pm k_0)$, which is clearly not the case. The dispersion of isolated intensity peaks along a single direction is quite unusual and requires consideration of factors beyond the degree of electronic correlation.
We propose that the solution to this puzzle involves the coupling of spin and orbital correlations. As previously mentioned, such a coupling has already been proposed for the antiferromagnetic phase.[@krug09; @lee09; @lv09; @turn09; @cric09; @chen09] In fact, it has been shown[@lee09] that the reduction of lattice symmetry can be associated with long-range orbital ordering, and the resulting anisotropy of the superexchange parameters is consistent with the observed spin-wave dispersions.[@zhao09; @dial09] In the present case, the orbital degeneracy of $d_{xz}$ and $d_{yz}$ orbitals re-activates the orbital freedom in the absence of long-range orbital order. This changes entirely the nature of the low-energy spin excitations.
Spin correlations characterized by ${\bf Q}\approx(\frac12,0)$, as observed here, correspond to the columnar pattern of Fig. \[fg:latt\](b). As shown in Ref. , such correlations are directly tied to the spin polarization of the $d_{yz}$ orbital and its ferro-orbital correlation. \[Similarly, spin correlations at ${\bf Q} \approx(0,\frac12)$ are tied to the $d_{xz}$ orbital.\] We propose that the spin-flip excitations in the present case are constrained by a strong coupling to orbital correlations, with two leading contributions illustrated in Fig. \[fg:ff\]. First, if we consider flipping a single spin, we find that the orbital also needs to flip from $d_{yz}$ to $d_{xz}$ \[[*c.f.*]{} Fig. \[fg:ff\](a)\], in order to better utilize the superexchange energy. (If the orbital remains unchanged, a local spin flip would have closed almost all the superexchange paths within the $d_{yz}$ and $d_{xz}$ subspace.) This “inter-orbital spin/orbital flip” process introduces a strongly anisotropic form factor that goes to zero along the longitudinal direction, as shown in Fig. \[fg:ff\](b). This is consistent with the strong anisotropy of our neutron data and the previous observation at high energies.[@lums10] Another relevant process, one associated with itinerancy, is the “inter-site spin flip” process shown in Fig. \[fg:ff\](c); the latter has a large form factor around $(\frac12, 0)$ \[[*c.f.*]{} Fig. \[fg:ff\](d)\]. Interestingly, due to the columnar AF arrangement of the short-range spin correlations together with the Pauli exclusion principle, our proposed inter-site spin-flip process can only propagate along the $y$-direction at low-energy, in good agreement with the observed anomalous dispersion. A quantitative analysis of the dispersion (beyond the scope of this paper), will require evaluating the excitation energy as a function of wave vector, taking into account both processes. In any case, the essence of our proposal is that the spin excitations are hybrids of the magnons and orbitons that have been postulated previously.[@krug09; @lee09]
In conclusion, we have presented experimental evidence that the electronic and magnetic excitations of superconducting [FeSe$_{0.5}$Te$_{0.5}$]{} are different from those of common band-structure and spin-wave models. We have proposed that the intensity anisotropy of the magnetic excitations may be explained by considering coupled spin and orbital excitations. If correct, this would imply that nematic excitations are a key feature of the normal state from which the superconductivity develops. While the details of the magnetism are quite different, the nematic correlations would provide an intriguing connection to the physics of cuprate superconductors.[@frad09]
We are grateful to S. A. Kivelson for helpful comments and to A. V. Fedorov for experimental assistance. Work at the University of Virginia was supported by the Office of Science, US Department of Energy (DOE) through DE-FG02-07ER46384. Work at Brookhaven is supported by the US DOE under Contract No. DE-AC02-98CH10886. PDJ and JMT are supported in part by the Center for Emergent Superconductivity, an Energy Frontier Research Center funded by the US DOE, Office of Basic Energy Sciences. SPINS at NCNR is supported by the National Science Foundation under Agreement No. DMR-0454672. ALS is operated by the US DOE under Contract No. DE-AC03-76SF00098
[28]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
, ****, ().
, , , ****, ().
, , , , ****, ().
, , , , , , , ****, ().
, , , , ****, ().
, , , ****, ().
, , , ****, ().
, , , ****, ().
, , , .
, , , , , ****, ().
, , , , , .
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, .
, ****, ().
, .
, , , , , ****, ().
, ****, (). , ****, ().
, , , .
, , , , , ****, ().
, ****, ().
, .
, .
, ****, ().
, ****, ().
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper we prove Homological Projective Duality for categorical resolutions of several classes of linear determinantal varieties. By this we mean varieties that are cut out by the minors of a given rank of a $m\times n$ matrix of linear forms on a given projective space. As applications, we obtain pairs of derived-equivalent Calabi-Yau manifolds, and address a question by A. Bondal asking whether the derived category of any smooth projective variety can be fully faithfully embedded in the derived category of a smooth Fano variety. Moreover we discuss the relation between rationality and categorical representability in codimension two for determinantal varieties.'
address:
- |
Institut de Mathématiques de Toulouse\
Université Paul Sabatier\
118 route de Narbonne\
31062 Toulouse Cedex 9\
France
- |
Institut Montpellierain Alexander Grothendieck\
Université de Montpellier\
Case Courrier 051 - Place Eugène Bataillon\
34095 Montpellier Cedex 5\
France
- |
Université de Bourgogne\
Institut de Mathématiques de Bourgogne\
UMR CNRS 5584\
UFR Sciences et Techniques – Bâtiment Mirande – Bureau 310\
9 Avenue Alain Savary\
BP 47870 21078 Dijon Cedex\
France
author:
- Marcello Bernardara
- Michele Bolognesi
- Daniele Faenzi
title: Homological projective duality for determinantal varieties
---
[^1]
Introduction
============
Homological Projective Duality (HPD) is one of the most exciting recent breakthroughs in homological algebra and algebraic geometry. It was introduced by A. Kuznetsov in [@kuznetsov:hpd] and its goal is to generalize classical projective duality to a homological framework. One of the important features of HPD is that it offers a very important tool to study the bounded derived category of a projective variety together with its linear sections, providing interesting semiorthogonal decompositions as well as derived equivalences, cf. [@kuznetsov:v14; @kuz:4fold; @kuznetsov:quadrics; @auel-berna-bolo; @kuz:icm].
Roughly speaking, two (smooth) varieties $X$ and $Y$ are HP-dual if $X$ has an ample line bundle ${{{\mathscr}O}}_X(1)$ giving a map $X \to {\mathbb{P}}W$, $Y$ has an ample line bundle ${{{\mathscr}O}}_Y(1)$ giving a map $Y \to {\mathbb{P}}W{^{\vee}}$, and $X$ and $Y$ have dual semiorthogonal decompositions (called *Lefschetz decompositions) compatible with the projective embedding. In this case, given a generic linear subspace $L \subset W$ and its orthogonal $L^\perp \subset W{^{\vee}}$, one can consider the linear sections $X_L$ and $Y_L$ of $X$ and $Y$ respectively. Kuznetsov shows the existence of a category ${\begin{bf}C\end{bf}}_L$ which is admissible both in ${{\rm D}^{\rm b}}(X_L)$ and in ${{\rm D}^{\rm b}}(Y_L)$, and whose orthogonal complement is given by some of the components of the Lefschetz decompositions of ${{\rm D}^{\rm b}}(X)$ and ${{\rm D}^{\rm b}}(Y)$ respectively. That is, both ${{\rm D}^{\rm b}}(X_L)$ and ${{\rm D}^{\rm b}}(Y_L)$ admit a semiorthogonal decomposition by a “Lefschetz” component, obtained via iterated hyperplane sections, and a common “nontrivial” or “primitive” component.*
HPD is closely related to classical projective duality: [@kuznetsov:hpd Theorem 7.9] states that the critical locus of the map $Y \to {\mathbb{P}}W{^{\vee}}$ coincides with the classical projective dual of $X$. The main technical issue of this fact is that one has to take into account singular varieties, since the projective dual of a smooth variety is seldom smooth - *e.g. the dual of certain Grassmannians are singular Pfaffian varieties [@boris-cald:pfaff]. On the other hand, derived (dg-enhanced) categories should provide a so-called *categorical or *non-commutative resolution of singularities ([@kuznet:singul; @vdb:non-comm-resos]). Roughly speaking, one needs to find a sheaf of ${{{\mathscr}O}}_Y$-(dg)-algebras ${{{\mathscr}R}}$ such that the category ${{\rm D}^{\rm b}}(Y,{{{\mathscr}R}})$ of bounded complexes of coherent ${{{\mathscr}R}}$-modules is proper, smooth and ${{{\mathscr}R}}$ is locally Morita-equivalent to some matrix algebra over ${{{\mathscr}O}}_Y$ (this latter condition translates the fact that the resolution is birational). In the case where $Y$ is singular, one of the most difficult tasks in proving HPD is to provide such a resolution with the required Lefschetz decomposition (for example, see [@kuz:icm §4.7]). On the other hand, given a non-smooth variety, it is a very interesting question to provide such resolutions and study their properties such as crepancy, minimality and so forth.***
The main application of HPD is that it is a direct method to produce semiorthogonal decompositions for projective varieties with non-trivial canonical sheaf, and derived equivalences for Calabi-Yau varieties. The importance of this application is due to the fact that determining whether a given variety admits or not a semiorthogonal decomposition is a very hard problem in general. Notice that there are cases where it is known that the answer to this question is negative, for example if $X$ has trivial canonical bundle [@bridg-equiv-and-FM Ex. 3.2], or if $X$ is a curve of positive genus [@okawa-curves]. On the other hand, if $X$ is Fano, then any line bundle is exceptional and gives then a semiorthogonal decomposition. Almost all the known cases of semiorthogonal decompositions of Fano varieties described in the literature (see, *e.g., [@kuznetsov:v14; @kuz:fano; @kuz:4fold; @bolo_berna:conic; @auel-berna-bolo]) can be obtained via HPD or its relative version.*
Derived equivalences of Calabi-Yau (CY for short) varieties have deep geometrical insight. First of all, it was shown by Bridgeland that birational CY-threefolds are derived equivalent [@bridg-flop]. The converse in not true: the first example - that has been shown to be also a consequence of HPD in [@kuznet-grass] - was displayed by Borisov and Caldararu in [@boris-cald:pfaff].
Besides their geometric relevance, derived equivalences between CY varieties play an important role in theoretical physics. First of all, Kontsevich’s homological mirror symmetry conjectures an equivalence between the bounded derived category of a CY-threefold $X$ and the Fukaya category of its mirror. More recently, it has been conjectured that homological projective duality should be realized physically as phases of abelian *gauged linear sigma models (GLSM) (see [@herbst-hori-page:phases] and [@viennacircle:git]).*
As an example, denote by $X$ and $Y$ the pair of equivalent CY–threefolds considered by Borisov and Caldararu. Rødland [@rodlando] argued that the families of $X$’s and $Y$’s (letting the linear section move in the ambient space) seem to have the same mirror variety $Z$ (a more string theoretical argument has been given recently by Hori and Tong [@hori-tong]). The equivalence between $X$ and $Y$ would then fit Kontsevich’s Homological Mirror Symmetry conjecture via the Fukaya category of $Z$. It is thus fair to say that HPD plays an important role in understanding these questions and potentially providing new examples. Notice in particular that some determinantal cases were considered in [@vivalafisica].
In this paper, we describe new families of HP Dual varieties. We consider two vector spaces $U$ and $V$ of dimension $m$ and $n$ respectively with $m \leq n$. Let ${\mathbb{G}}={\mathbb{G}}(U,r)$ denote the Grassmannian of $r$-dimensional quotients of $U$, set ${{{\mathscr}Q}}$ and ${{{\mathscr}U}}$ for the universal quotient and sub-bundle respectively. Let ${{{\mathscr}X}}:= {\mathbb{P}}(V
{\otimes}{{{\mathscr}Q}})$ and ${{{\mathscr}Y}}:= {\mathbb{P}}(V {^{\vee}}{\otimes}{{{\mathscr}U}}^\vee)$, for any $0 < r < m$. Let $p:{{{\mathscr}X}}\to {\mathbb{G}}$ and $q:{{{\mathscr}Y}}\to {\mathbb{G}}$ be the natural projections. Set $H_X$ and $H_Y$ for the relatively ample tautological divisors on ${{{\mathscr}X}}$ and ${{{\mathscr}Y}}$. Orlov’s result [@orlov:proj_bundles] provides semiorthogonal decompositions $$\label{eq:the-dual-lefsch-decos}
\begin{array}{l}
{{\rm D}^{\rm b}}({{{\mathscr}X}}) = \langle p^*{{\rm D}^{\rm b}}({\mathbb{G}}), \ldots, p^*{{\rm D}^{\rm b}}({\mathbb{G}}) \otimes {{{\mathscr}O}}_{{{\mathscr}X}}((r n-1)H_X) \rangle, \\
\\
{{\rm D}^{\rm b}}({{{\mathscr}Y}}) = \langle q^*{{\rm D}^{\rm b}}({\mathbb{G}}) \otimes {{{\mathscr}O}}_{{{\mathscr}Y}}(((r-m)n+1)H_Y), \ldots, q^*{{\rm D}^{\rm b}}({\mathbb{G}}) \rangle.
\end{array}$$
In the previous notation, ${{{\mathscr}X}}$ and ${{{\mathscr}Y}}$ with Lefschetz decompositions are HP-dual.
The proof of the previous result is a consequence of Kuznetsov’s HPD for projective bundles generated by global sections (see [@kuznetsov:hpd §8]). Here, the spaces of global sections of ${{{\mathscr}O}}_{{{\mathscr}X}}(H_X)$ and ${{{\mathscr}O}}_{{{\mathscr}Y}}(H_Y)$ sheaves are, respectively, $W=V \otimes U$ and $W{^{\vee}}=V{^{\vee}}\otimes U{^{\vee}}$.
The main interest of Theorem \[thm:main1\] is that ${{{\mathscr}X}}$ is known to be the resolution of the variety ${{{\mathscr}Z}}^r$ of $m\times n$ matrices of rank at most $r$. Write such a matrix as $M: U \to V{^{\vee}}$. Then ${{{\mathscr}Z}}^r$ is naturally a subvariety of ${\mathbb{P}}W$, which is singular in general, with resolution $f: {{{\mathscr}X}}\to {{{\mathscr}Z}}^r$. Dually, $g:{{{\mathscr}Y}}\to
{{{\mathscr}Z}}^{m-r}$ is a desingularization of the variety of $m\times n$ matrices of corank at least $r$. Theorem \[thm:main1\] provides the categorical framework to describe HPD between the classical projectively dual varieties ${{{\mathscr}Z}}^r$ and ${{{\mathscr}Z}}^{m-r}$ (see, *e.g., [@weyman-book]).*
In the affine case, categorical resolutions for determinantal varieties have been constructed by Buchweitz, Leuschke and van den Bergh [@buch-leu-vdbergh1; @buch-leu-vdbergh2]. Such resolution is crepant if $m=n$ (that is, in the case where ${{{\mathscr}Z}}^r$ has Gorenstein singularities). The starting point is Kapranov’s construction of a full strong exceptional collection on Grassmannians [@kapra-grassa]. One can use the decompositions in exceptional objects to produce a sheaf of algebras ${{{\mathscr}R}}'$ and a categorical resolution of singularities ${{\rm D}^{\rm b}}({{{\mathscr}Z}}^r,{{{\mathscr}R}}') \simeq {{\rm D}^{\rm b}}({{{\mathscr}X}})$. For simplicity, we will denote by ${{{\mathscr}R}}'$ the algebra on any of the determinantal varieties ${{{\mathscr}Z}}^r$ (forgetting about the dependence of ${{{\mathscr}R}}'$ on the rank $r$). This gives a geometrically deeper version of Theorem \[thm:main1\].
In the previous notations, ${{{\mathscr}Z}}^r$ admits a categorical resolution of singularities ${{\rm D}^{\rm b}}({{{\mathscr}Z}}^r,{{{\mathscr}R}}')$, which is crepant if $m=n$. Moreover, ${{\rm D}^{\rm b}}({{{\mathscr}Z}}^r,{{{\mathscr}R}}')$ and ${{\rm D}^{\rm b}}({{{\mathscr}Z}}^{m-r},{{{\mathscr}R}}')$ are HP-dual.
Once the equivalence ${{\rm D}^{\rm b}}({{{\mathscr}Z}}^r,{{{\mathscr}R}}') \simeq {{\rm D}^{\rm b}}({{{\mathscr}X}})$ constructed, Theorem \[thm:main2\] is proved by applying directly Theorem \[thm:main1\]. However, the geometric relevance of Theorem \[thm:main2\], and its difference with Theorem \[thm:main1\], is that it shows HPD directly on noncommutative structures over the determinantal varieties ${{{\mathscr}Z}}^r$ and ${{{\mathscr}Z}}^{m-r}$ with respect to their natural embedding in ${\mathbb{P}}W$ and ${\mathbb{P}}W ^\vee$ respectively. That is, these natural smooth and proper noncommutative scheme structures are well-behaved with respect to projective duality and hyperplane sections. Finally, notice that whenever we pick a smooth linear section $Z_L$ of ${{{\mathscr}Z}}^r$ (or a smooth section $Z^L$ of ${{{\mathscr}Z}}^{m-r}$), the restriction to $Z_L$ of the sheaf ${{{\mathscr}R}}'$ is Morita-equivalent to ${{{\mathscr}O}}_{Z_L}$, so that we get the derived category ${{\rm D}^{\rm b}}(Z_L)$ of the section itself.
As a consequence, given a matrix of linear forms on some projective space, one can see the locus $Z$ where the matrix has rank at most $r$ as a linear section of ${{{\mathscr}Z}}^r$. Assuming $Z$ to have expected dimension, Theorem \[thm:main2\] gives a categorical resolution of singularities ${{\rm D}^{\rm b}}(Z,{{{\mathscr}R}}')$ of $Z$ and a semiorthogonal decomposition of this category involving the dual linear section of ${{{\mathscr}Z}}^{m-r}$.
Our construction of Homological Projective Duality allows us to recover some Calabi-Yau equivalences appeared in [@vivalafisica] and many more (see Corollary \[cor:functors\]).
A special case is obtained by setting $r=1$. In this case ${{{\mathscr}X}}$ is a Segre variety and ${{{\mathscr}Y}}$ is the variety of degenerate matrices rank.
As an application of this new instance of Homological Projective Duality, we try to address a fascinating question, asked by A. Bondal in Tokyo in 2011. Since any Fano variety admits semiorthogonal decompositions, it is natural to ask whether the derived category of any variety can be realized as a component of a semiorthogonal decomposition of a Fano variety. Under this perspective, considering Fano varieties will be enough to study all “geometric” triangulated categories.
\[qu:bondal\] Let $X$ be a smooth and projective variety. Is there any smooth Fano variety $Y$ together with a full and faithful functor ${{\rm D}^{\rm b}}(X) \to {{\rm D}^{\rm b}}(Y)$?
We will say that $X$ is *Fano-visitor if Question \[qu:bondal\] has a positive answer (see Definition \[host\]).*
On the other hand, an interesting geometrical insight of semiorthogonal decompositions is to provide a conjectural obstruction to rationality of a given variety $X$. In [@bolognesi_bernardara:representability], the first and second named authors introduced , based on existence of semiorthogonal decompositions, the notion of *categorical representability of a variety $X$ (see Definition \[def-cat-rep\]). This notion allows to formulate a natural question about categorical obstructions to rationality.*
\[qu:catrep=ration\] Is a rational projective variety always categorically representable in codimension at least 2?
The motivating ideas of question \[qu:catrep=ration\] can be traced back to the work of Bondal and Orlov, and to their address at the 2002 ICM [@bondal_orlov:ICM2002], and to Kuznetsov’s remarkable contributions (e.g. [@kuz:4fold] or [@kuz:rationality-report]). Notice that a projective space is representable in dimension 0. Roughly speaking, the idea supporting Question \[qu:catrep=ration\] is based on a motivic argument which let us suppose that birational transformations should not add components representable codimension 1 or less (see also [@bolognesi_bernardara:representability]).
Several examples seem to suggest that Question \[qu:catrep=ration\] may have a positive answer. Let us mention conic bundles over minimal surfaces [@bolo_berna:conic], fibrations in intersections of quadrics [@auel-berna-bolo], or some classes of cubic fourfolds [@kuz:4fold]. Moreover, Question \[qu:catrep=ration\] is equivalent to one implication of Kuznetsov Conjecture on the rationality of a cubic fourfold [@kuz:4fold], which was proved to coincide with Hodge theoretical expectations for a general cubic fourfold by Addington and Thomas [@add-thomas].
As consequences of Theorems \[thm:main1\] and \[thm:main2\], we can show that (the categorical resolution of singularities of) any determinantal hypersurface of general type is Fano visitor (§\[sect:segre-det\]), and that (the categorical resolution of singularities of) a rational determinantal variety is categorically representable in codimension at least two (§\[sect:fano\]). Hence we provide a large family of varieties for which Questions \[qu:bondal\] and \[qu:catrep=ration\] have positive answer. As an example, we easily get the following corollary (compare with Example \[ex:plane-cves\]).
A smooth plane curve is Fano visitor.
Acknowledgments {#acknowledgments .unnumbered}
---------------
We acknowledge A. Bondal for asking Question \[qu:bondal\] at the conference “Derived Categories 2011” in Tokyo, which has been a source of inspiration for this work. We thank J. Rennemo for pointing out a mistake in the first version of this paper, and the anonymous referee for useful suggestions and questioning. We are grateful to A. Kuznetsov, N. Addington and E. Segal for useful advises and exchange of ideas.
Preliminaries
=============
Notation
--------
We work over an algebraically closed field of characteristic zero $k$. A vector space will be denoted by a capital letter $W$; the dual vector space is denoted by $W^\vee$. Suppose ${{\rm dim}\,}(W)=N$, then the projective space of $W$ is denoted by ${\mathbb{P}}W$ or simply by ${\mathbb{P}}^{N-1}$. We follow Grothendieck’s convention, so that ${\mathbb{P}}W$ is the set of hyperplanes through the origin of $W$. The dual projective space is denoted by ${\mathbb{P}}W ^\vee$ or by $({\mathbb{P}}^{N-1})^\vee$.
We assume the reader to be familiar with the theory of semiorthogonal decompositions and exceptional objects (see [@bondal_orlov:semiorthogonal; @huybrechts:libro; @kuz:icm]). Recall that the summands of a semiorthogonal decomposition of a triangulated category $\mathbf{T}$, by definition are full triangulated subcategories of $\mathbf{T}$ which are admissible, i. e. such that the inclusion admits a left and right adjoint.
Categorical resolutions of singularities
----------------------------------------
By a [*noncommutative scheme*]{} we mean (following Kuznetsov [@kuznetsov:quadrics §2.1]) a scheme $X$ together with a coherent ${{{\mathscr}O}}_{X}$-algebra ${{{\mathscr}A}}$. Morphisms are defined accordingly. By definition, a noncommutative scheme $(X,{{{\mathscr}A}})$ has ${{\mathsf{Coh}}}(X, {{{\mathscr}A}})$, the category of coherent ${{{\mathscr}A}}$-modules, as category of coherent sheaves and ${{\rm D}^{\rm b}}(X,{{{\mathscr}A}})$ as bounded derived category.
Following Bondal–Orlov [@bondal_orlov:ICM2002 §5], a *categorical (or *noncommutative) *resolution of singularities $(X,{{{\mathscr}A}})$ of a possibly singular proper scheme $X$ is a torsion free ${{{\mathscr}O}}_X$-algebra ${{{\mathscr}A}}$ of finite rank such that ${{\mathsf{Coh}}}(X,{{{\mathscr}A}})$ has finite homological dimension (*i.e., is smooth in the noncommutative sense).****
Let $X$ be a scheme. An object $T$ of ${{\rm D}^{\rm b}}(X)$ is called a [*compact generator*]{} if $T$ is perfect and, for any object $S$ of ${{\rm D}^{\rm b}}(X)$, we have that the fact that ${{\rm Hom}}_{{{\rm D}^{\rm b}}(X)}(S,T[i])=0$ for all integers $i$ is equivalent to $S=0$. Notice that, if $X$ is smooth and proper, the natural inclusion $\mathrm{Perf}(X)
\subset {{\rm D}^{\rm b}}(X)$ of perfect complexes into ${{\rm D}^{\rm b}}(X)$ is an equivalence. Hence any object in ${{\rm D}^{\rm b}}(X)$ is perfect.
In the case where $X$ admits a full exceptional collection, there is an explicit compact generator $T$.
Suppose that $X$ is smooth and proper, and that ${{\rm D}^{\rm b}}(X)$ is generated by a full exceptional sequence ${{\rm D}^{\rm b}}(X) = \langle E_1, \ldots, E_s \rangle$. Then $E= \oplus_{i=1}^s E_i$ is a compact generator. In particular, consider the dg-$k$-algebra $\mathrm{End}(E)$. Then there is an equivalence of triangulated categories ${{\rm D}^{\rm b}}(X) \simeq {{\rm D}^{\rm b}}(\mathrm{End}(E))$.
Homological Projective Duality
------------------------------
Homological Projective Duality (HPD) was introduced by Kuznetsov [@kuznetsov:hpd] in order to study derived categories of hyperplane sections (see also [@kuznetsov:hyp-sections]).
Let us first recall the basic notion of HPD from [@kuznetsov:hpd]. Let $X$ be a projective scheme together with a base-point-free line bundle ${{{\mathscr}O}}_X(H)$.
A [*Lefschetz decomposition*]{} of ${{\rm D}^{\rm b}}(X)$ with respect to ${{{\mathscr}O}}_X(H)$ is a semiorthogonal decomposition $$\label{eq:def-of-lefschetz}
{{\rm D}^{\rm b}}(X) = \langle {\begin{bf}A\end{bf}}_0, {\begin{bf}A\end{bf}}_1(H), \dotsc, {\begin{bf}A\end{bf}}_{i-1}((i-1)H) \rangle,$$ with $$0 \subset {\begin{bf}A\end{bf}}_{i-1} \subset \dotsc \subset {\begin{bf}A\end{bf}}_0,$$ Such a decomposition is said to be *rectangular if ${\begin{bf}A\end{bf}}_0= \ldots = {\begin{bf}A\end{bf}}_{i-1}$.*
Let $W := H^0(X,{{{\mathscr}O}}_X(H))$, and $f: X \to {\mathbb{P}}W$ the map given by the linear system associated with ${{{\mathscr}O}}_X(H)$, so that $f^*
{{{\mathscr}O}}_{{\mathbb{P}}W}(1) {\cong}{{{\mathscr}O}}_X(H)$. We denote by ${{\mathcal}X} \subset X \times
{\mathbb{P}}W{^{\vee}}$ the universal hyperplane section of $X$ $${\mathcal}X:= \{(x,H)\in X\times {\mathbb{P}}W^\vee | x \in H\}.$$
Let $f: X \to {\mathbb{P}}W $ be a smooth projective scheme with a base-point-free line bundle ${{{\mathscr}O}}_X(H)$ and a Lefschetz decomposition as above. A scheme $Y$ with a map $g: Y \to {\mathbb{P}}W^\vee$ is called [*homologically projectively dual*]{} (or the [*HP-dual*]{}) to $f: X \to {\mathbb{P}}W$ with respect to the Lefschetz decomposition , if there exists a fully faithful functor $\Phi: {{\rm D}^{\rm b}}(Y) \to {{\rm D}^{\rm b}}({\mathcal}X)$ giving the semiorthogonal decomposition $${{\rm D}^{\rm b}}({{\mathcal}X}) = \langle \Phi ({{\rm D}^{\rm b}}(Y), {\begin{bf}A\end{bf}}_1(1) \boxtimes {{\rm D}^{\rm b}}({\mathbb{P}}W{^{\vee}}), \dotsc,
{\begin{bf}A\end{bf}}_{i-1}(i-1) \boxtimes {{\rm D}^{\rm b}}({\mathbb{P}}W{^{\vee}}) \rangle.$$
Let $N = {{\rm dim}\,}(W)$ and let $c\le N$ be an integer. Given a $c$-codimensional linear subspace $L \subset W$, we define the linear subspace ${\mathbb{P}}_L \subset {\mathbb{P}}W$ of codimension $c$ as ${\mathbb{P}}(W/L)$. Dually, we have a linear subspace ${\mathbb{P}}^L={\mathbb{P}}L^\perp$ of dimension $c-1$ in ${\mathbb{P}}W{^{\vee}}$, whose defining equations are the elements of $L^\perp
\subset W{^{\vee}}$. We define the varieties: $$X_L = X \times_{{\mathbb{P}}W} {\mathbb{P}}_L, \qquad Y_L = Y \times_{{\mathbb{P}}W^\vee} {\mathbb{P}}^L.$$
\[thm:HPD\] Let $X$ be a smooth projective variety with a map $f: X \to {\mathbb{P}}W$, and a Lefschetz decomposition with respect to ${{{\mathscr}O}}_X(H)$. If $Y$ is HP-dual to $X$, then:
- $Y$ is smooth projective and admits a [*dual*]{} Lefschetz decomposition $${{\rm D}^{\rm b}}(Y) = \langle {\begin{bf}B\end{bf}}_{j-1}(1-j), \dotsc, {\begin{bf}B\end{bf}}_1 (-1),
{\begin{bf}B\end{bf}}_0 \rangle, \quad\quad {\begin{bf}B\end{bf}}_{j-1} \subset \dotsc
\subset {\begin{bf}B\end{bf}}_1 \subset {\begin{bf}B\end{bf}}_0$$ with respect to the line bundle ${{{\mathscr}O}}_Y(H)=g^*{{{\mathscr}O}}_{{\mathbb{P}}W^\vee}(1)$.
- if $L$ is admissible, i.e. if $${{\rm dim}\,}X_L = {{\rm dim}\,}X - c, \quad \mbox{and}\quad {{\rm dim}\,}Y_L = {{\rm dim}\,}Y + c - N,$$ then there exist a triangulated category ${\begin{bf}C\end{bf}}_L$ and semiorthogonal decompositions: $$\begin{aligned}
{{\rm D}^{\rm b}}(X_L) & = \langle {\begin{bf}C\end{bf}}_L, {\begin{bf}A\end{bf}}_c(1), \dotsc, {\begin{bf}A\end{bf}}_{i-1}(i-c) \rangle,\\
{{\rm D}^{\rm b}}(Y_L) & = \langle {\begin{bf}B\end{bf}}_{j-1} (N-c-j), \dotsc, {\begin{bf}B\end{bf}}_{N-c}(-1), {\begin{bf}C\end{bf}}_L \rangle.\end{aligned}$$
In general, HPD involves non smooth varieties. Indeed, as shown by Kuznetsov [@kuznetsov:hpd Theorem 7.9] the critical locus of the map $g: Y \to {\mathbb{P}}W^\vee$ is the classical projective dual $X^\vee$ of $X$, which is rarely smooth even if $X$ is smooth. If $X$ (resp. $Y$) is singular, then we have to replace ${{\rm D}^{\rm b}}(X)$ (resp. ${{\rm D}^{\rm b}}(Y)$) by a categorical resolution of singularities ${{\rm D}^{\rm b}}(X,{{{\mathscr}A}})$ (resp. ${{\rm D}^{\rm b}}(Y,{{{\mathscr}B}})$) in all the statements and definitions of this section. Theorem \[thm:HPD\] holds in this more general framework, where we have to consider ${{\rm D}^{\rm b}}(X_L,{{{\mathscr}A}}_L)$ (resp. ${{\rm D}^{\rm b}}(Y_L,{{{\mathscr}B}}_L)$) for ${{{\mathscr}A}}_L$ (resp. ${{{\mathscr}B}}_L$) the restriction of ${{{\mathscr}A}}$ to $X_L$ (resp. of ${{{\mathscr}B}}$ to $Y_L$) in item (ii).
Categorical representability and Fano visitors
----------------------------------------------
First, let us recall the definition of categorical representability for a variety.
A triangulated category ${\begin{bf}T\end{bf}}$ is [*representable in dimension $j$*]{} if it admits a semiorthogonal decomposition $${\begin{bf}T\end{bf}} = \langle {\begin{bf}A\end{bf}}_1, \ldots, {\begin{bf}A\end{bf}}_l \rangle,$$ and for all $i=1,\ldots,l$ there exists a smooth projective connected variety $Y_i$ with ${{\rm dim}\,}Y_i \leq j$, such that ${\begin{bf}A\end{bf}}_i$ is equivalent to an admissible subcategory of ${{\rm D}^{\rm b}}(Y_i)$.
\[def-cat-rep\] Let $X$ be a projective variety. We say that $X$ is [*categorically representable*]{} in dimension $j$ (or equivalently in codimension ${{\rm dim}\,}(X)-j$) if there exists a categorical resolution of singularities of ${{\rm D}^{\rm b}}(X)$ representable in dimension $j$.
Based on Bondal’s Question \[qu:bondal\], we introduce the following definition.
\[host\] A triangulated category ${\begin{bf}T\end{bf}}$ is *Fano-visitor if there exists a smooth Fano variety $F$ and a fully faithful functor ${\begin{bf}T\end{bf}} \to {{\rm D}^{\rm b}}(F)$ such that ${{\rm D}^{\rm b}}(F) = \langle {\begin{bf}T\end{bf}},{\begin{bf}T\end{bf}}^\perp \rangle$. A smooth projective variety $X$ is said to be a *Fano-visitor if its derived category ${{\rm D}^{\rm b}}(X)$ is Fano-visitor.**
We remark that, having a fully faithful functor ${{\rm D}^{\rm b}}(X) \to {{\rm D}^{\rm b}}(F)$ is enough to have the required semiorthogonal decomposition [@bondal:representations]. Relaxing slightly the hypotheses on the smoothness of the Fano variety we get the following weaker definition.
\[whost\] A triangulated category ${\begin{bf}T\end{bf}}$ is *weakly Fano-visitor if there exists a (possibly singular) Fano variety $F$, a categorical crepant resolution of singularities ${\begin{bf}DF\end{bf}}$ of $F$ and a fully faithful functor ${\begin{bf}T\end{bf}} \to {\begin{bf}DF\end{bf}}$ such that ${\begin{bf}DF\end{bf}} = \langle {\begin{bf}T\end{bf}},{\begin{bf}T\end{bf}}^\perp \rangle$. Notice that this implies that the functor ${\begin{bf}T\end{bf}} \to {\begin{bf}DF\end{bf}}$ has a right and left adjoint by definition of semiorthogonal decomposition. As before, if ${\begin{bf}T\end{bf}}\cong {{\rm D}^{\rm b}}(X)$ for a smooth projective variety $X$, then $X$ itself is said to be *weakly Fano-visitor .**
Homological Projective Duality for determinantal varieties
==========================================================
We describe here homological projective duality for determinantal varieties in terms of the Springer resolution of the space of $n \times m$ matrices of rank at most $r$ and in terms of categorical resolution of singularities.
The Springer resolution of the space of matrices of bounded rank {#desingo}
----------------------------------------------------------------
Let us introduce the variety ${{{{\mathscr}Z}}^r_{m,n}}$ of $n \times m$ matrices over our base field, having rank at most $r$. Let $U$, $V$ be vector spaces, with ${{\rm dim}\,}U=m$, ${{\rm dim}\,}V=n$, and assume $n \ge m$. Set $W = U {\otimes}V$. Let $r$ be an integer in the range $1 \le r \le m-1$. We define ${{{\mathscr}Z}}^r={{{{\mathscr}Z}}^r_{m,n}}$ to be the variety of matrices $M : V
\to U{^{\vee}}$ in ${\mathbb{P}}W$ cut by the minors of size $r+1$ of the matrix of indeterminates: $$\psi = \begin{pmatrix}
x_{1,1} & \ldots & x_{m,1} \\
\vdots & \ddots & \vdots\\
x_{m,n} & \ldots & x_{m,n}
\end{pmatrix}$$
### Springer resolution and projective bundles
Consider the Grassmann variety ${\mathbb{G}}(U,r)$ of $r$-dimensional quotient spaces of $U$, the tautological sub-bundle and the quotient bundle over ${\mathbb{G}}(U,r)$, denoted respectively by ${{{\mathscr}U}}$ and ${{{\mathscr}Q}}$, respectively of rank $m-r$ and $r$. We will write ${\mathbb{G}}$ for ${\mathbb{G}}(U,r)$.
The tautological (or Euler) exact sequence reads: $$\label{tautological}
0 \to {{{\mathscr}U}}\to U {\otimes}{{{\mathscr}O}}_{{\mathbb{G}}} \to {{{\mathscr}Q}}\to 0.$$
We will use the following notation: $${{{{\mathscr}X}}^r_{m,n}}= {\mathbb{P}}(V {\otimes}{{{\mathscr}Q}}).$$ However, the dependency on $m,n,r$ will often be omitted.
The manifold ${{{\mathscr}X}}={{{{\mathscr}X}}^r_{m,n}}$ has dimension $r(n+m-r)-1$. It is the resolution of singularities of the variety of $m \times n$ matrices of rank at most $r$, in a sense that we will now review. Denote by $p$ the natural projection ${{{\mathscr}X}}\to {\mathbb{G}}$. The space $H^0({\mathbb{G}},{{{\mathscr}Q}})$ is naturally identified with $U$.
Let us denote by ${{{\mathscr}O}}_{{{{\mathscr}X}}}(H_X)$ the relatively ample tautological line bundle on ${{{\mathscr}X}}$. We will often write simply $H$ for $H_X$. We get natural isomorphisms: $$H^0({\mathbb{G}},V {\otimes}{{{\mathscr}Q}}) \simeq H^0({{{\mathscr}X}},{{{\mathscr}O}}_{{{{\mathscr}X}}}(H)) \simeq W = U {\otimes}V.$$ Therefore, the map $f$ associated with the linear system ${{{\mathscr}O}}_{{{{\mathscr}X}}}(H)$ maps ${{{\mathscr}X}}$ to ${\mathbb{P}}W$, and clearly ${{{\mathscr}O}}_{{{{\mathscr}X}}}(H) \simeq f^*({{{\mathscr}O}}_{{\mathbb{P}}W}(1))$. This is summarized by the diagram: $$\xymatrix@-2ex{
{{{\mathscr}X}}\ar^-{f}[r] \ar_-{p}[d] & {\mathbb{P}}W = {\mathbb{P}}(U {\otimes}V) \\
{\mathbb{G}}}$$
On the other hand, we will denote by $P$ the pull-back to ${\mathbb{P}}(V\otimes {{{\mathscr}Q}})$ of the first Chern class $c_1({{{\mathscr}Q}})$ on ${\mathbb{G}}$. Hence we have that $c_1(V\otimes
{{{\mathscr}Q}})$ pulls-back to $n P$ and $\omega_{{\mathbb{G}}}$ to $-m P$. The Picard group of ${{{\mathscr}X}}$ is generated by $P$ and $H$.
Notice that giving a rank-1 quotient of $W=U {\otimes}V$ corresponds to the choice of a linear map $M : V \to U{^{\vee}}$, so an element of ${\mathbb{P}}W$ can be considered as (the proportionality class of) the linear map $M$. On the other hand, the map $f$ sends a rank-1 quotient of $V
{\otimes}{{{\mathscr}Q}}$ over a point $\lambda \in {\mathbb{G}}$ to the quotient of $W$ obtained by composition with the obvious quotient $U \to {{{\mathscr}Q}}_\lambda$.
Therefore, the matrix $M$ lies in the image of $f$ if and only if $M$ factors through $V \to {{{\mathscr}Q}}_\lambda{^{\vee}}$, for some $\lambda \in
{\mathbb{G}}$, i.e., if and only if ${{\rm rk}}(M) \le r$. Clearly, if $M$ has precisely rank $r$ then it determines $\lambda$ and the associated quotient of $U \to {{{\mathscr}Q}}_\lambda$. Since this happens for a general matrix $M$ of ${{{\mathscr}Z}}^r={{{{\mathscr}Z}}^r_{m,n}}$, the map $f:
{{{\mathscr}X}}\to {{{\mathscr}Z}}^r$ is birational. This map is in fact a desingularization, called the [*Springer resolution*]{}, of ${{{\mathscr}Z}}^r$. It is an isomorphism above the locus of matrices of rank exactly $r$.
In a more concrete way, given $\lambda \in {\mathbb{G}}$ we let $\pi_\lambda$ be the linear projection from $U{^{\vee}}$ to $U{^{\vee}}/{{{\mathscr}Q}}{^{\vee}}_\lambda$. Then, the variety ${{{\mathscr}X}}$ can be thought of as: $${{{\mathscr}X}}=\{(\lambda,M) \in {\mathbb{G}}\times {{{\mathscr}Z}}^r \mid \pi_\lambda\circ M =0 \}.$$
This way, the maps $p$ and $f$ are just the projections from ${{{\mathscr}X}}$ onto the two factors.
Let us now look at the dual picture. We consider the projective bundle: $${{{{\mathscr}Y}}^r_{m,n}}= {\mathbb{P}}(V{^{\vee}}{\otimes}{{{\mathscr}U}}{^{\vee}}).$$
Write ${{{\mathscr}Y}}={{{{\mathscr}Y}}^r_{m,n}}$ for short. Denote by $q$ the projection ${{{\mathscr}Y}}\to
{\mathbb{G}}$. We will denote by $H_Y$ (or sometimes just by $H$) the tautological ample line bundle on ${{{\mathscr}Y}}$. This time, since $H^0({\mathbb{G}},{{{\mathscr}U}}{^{\vee}}) \simeq U{^{\vee}}$, the linear system associated with ${{{\mathscr}O}}_{{{{\mathscr}Y}}}(H)$ sends ${{{\mathscr}Y}}$ to ${\mathbb{P}}W^\vee \simeq {\mathbb{P}}(V{^{\vee}}{\otimes}U{^{\vee}})$ via a map that we call $g$. By the same argument as above, $g$ is a desingularization of the variety ${{{\mathscr}W}}^{r}$ of matrices $V^\vee \to U$ in ${\mathbb{P}}W^\vee$ of corank at least $r$. There exists an obvious isomorphism ${{{\mathscr}Z}}^{m-r} \cong {{{\mathscr}W}}^{r}$, which we will use without further mention.
The spaces ${\mathbb{P}}W$ and ${\mathbb{P}}W^\vee$ are equipped with tautological morphisms of sheaves, which are both identified by the the matrix $\psi$, corresponding to the identity in $W {\otimes}W{^{\vee}}= U \otimes V \otimes
U{^{\vee}}\otimes V{^{\vee}}$: $$\begin{aligned}
\label{tauto2} &V \otimes {{{\mathscr}O}}_{{\mathbb{P}}W}(-1) \stackrel{\psi}{\to} U{^{\vee}}\otimes {{{\mathscr}O}}_{{\mathbb{P}}W}, \\
\label{tauto} &V{^{\vee}}\otimes {{{\mathscr}O}}_{{\mathbb{P}}W ^\vee}(-1) \stackrel{\psi}{\to}
U \otimes {{{\mathscr}O}}_{{\mathbb{P}}W^\vee}.\end{aligned}$$
We will denote by ${{{\mathscr}F}}$ and ${{{\mathscr}E}}$, the cokernel of the tautological map appearing in Eq. , respectively Eq. .
\[un peu penible\] We have isomorphisms ${{{\mathscr}X}}\simeq {\mathbb{G}}({{{\mathscr}F}},m-r)$ and ${{{\mathscr}Y}}\simeq {\mathbb{G}}({{{\mathscr}E}},r)$.
We work out the proof for ${{{\mathscr}Y}}$, the argument for ${{{\mathscr}X}}$ being identical. Given a scheme $S$ over our field, an $S$-valued point $[e]$ of ${\mathbb{G}}({{{\mathscr}E}},r)$ is given by a morphism $s : S \to {\mathbb{P}}W^\vee$ and the equivalence class of an epimorphism $e : s^* {{{\mathscr}E}}\to {{{\mathscr}V}}$, where ${{{\mathscr}V}}$ is locally free of rank $r$ on $S$. On the other hand, an $S$-point $[y]$ of ${{{\mathscr}Y}}$ corresponds to a morphism $t : S \to {\mathbb{G}}$ together with the class of a quotient $y : V{^{\vee}}{\otimes}t^* {{{\mathscr}U}}{^{\vee}}\to {{{\mathscr}L}}$, with ${{{\mathscr}L}}$ invertible on $S$. In turn, $t$ is given by a locally free sheaf of rank $r$ on $S$ and a surjection from $U {\otimes}{{{\mathscr}O}}_S$ onto this sheaf.
Given the point $[e]$, we compose $e$ with the surjection $U {\otimes}{{{\mathscr}O}}_S \to s^* {{{\mathscr}E}}$ and denote by $t_e$ the resulting map $U {\otimes}{{{\mathscr}O}}_S \to {{{\mathscr}V}}$. This way, $t_e$ provides the required morphism $t : S \to
{\mathbb{G}}$, and clearly $t^* {{{\mathscr}Q}}\simeq {{{\mathscr}V}}$, so the kernel of $U
{\otimes}{{{\mathscr}O}}_S \to {{{\mathscr}V}}$ is just $t^* {{{\mathscr}U}}$. Clearly, we have $t_e \circ s^* \psi = 0$ so that $s^* \psi$ factors through a map $V{^{\vee}}{\otimes}{{{\mathscr}O}}_S(-1) \to t^* {{{\mathscr}U}}$. Giving this last map is equivalent to the choice of a map $V{^{\vee}}{\otimes}t^*{{{\mathscr}U}}{^{\vee}}\to {{{\mathscr}O}}_S(1)$, which we define to be the point $[y]$ associated with $[e]$.
Conversely, let $t$ be represented by a locally free sheaf ${{{\mathscr}V}}=t^* {{{\mathscr}Q}}$ of rank $r$ on $S$ and by a quotient $U {\otimes}{{{\mathscr}O}}_S \to {{{\mathscr}V}}$, whose kernel is $t^* {{{\mathscr}U}}$. Then, given point $[y]$ and the quotient $y$, we consider the composition of $y$ and $U{^{\vee}}{\otimes}{{{\mathscr}O}}_S \to {{{\mathscr}U}}{^{\vee}}$ to obtain a quotient $s_y : V{^{\vee}}{\otimes}U{^{\vee}}\to {{{\mathscr}L}}$. This gives the desired morphism $s
: S \to {\mathbb{P}}W^\vee$. Moreover, the map $V{^{\vee}}{\otimes}{{{\mathscr}O}}_S \to t^* {{{\mathscr}U}}{\otimes}{{{\mathscr}L}}$ associated with $y$ can be composed with the injection $t^* {{{\mathscr}U}}{\otimes}{{{\mathscr}L}}\to U {\otimes}{{{\mathscr}L}}$ to get a map $V{^{\vee}}{\otimes}{{{\mathscr}O}}_S \to U {\otimes}{{{\mathscr}L}}$, or equivalently $V{^{\vee}}{\otimes}{{{\mathscr}L}}{^{\vee}}\to U {\otimes}{{{\mathscr}O}}_S$, and this map is nothing but $s^* \psi$. Of course, composing this map with the projection $U {\otimes}{{{\mathscr}O}}_S
\to t^*{{{\mathscr}Q}}= {{{\mathscr}V}}$ we get zero, so there is an induced surjective map $s^* {{{\mathscr}E}}\to {{{\mathscr}V}}$. We define the class of this map to be the point $[e]$ associated with $[y]$.
We have defined two maps from the sets of $S$-valued points of our two schemes, which are inverse to each other by construction. The lemma is thus proved.
### Linear sections and projectivized sheaves {#proshe}
Let now $c$ be an integer in the range $1 \le c \le m n$, and suppose we a have $c$-dimensional vector subspace $L$ of $W$: $$L \subset U {\otimes}V = W.$$
We have thus the linear subspace ${\mathbb{P}}_L \subset {\mathbb{P}}W$ of codimension $c$, defined by ${\mathbb{P}}_L={\mathbb{P}}(W/L)$. Dually, we have a linear subspace ${\mathbb{P}}^L={\mathbb{P}}L^\perp$ of dimension $c-1$ in ${\mathbb{P}}W{^{\vee}}$, whose defining equations are the elements of $L^\perp
\subset W{^{\vee}}$. We define the varieties: $$X^r_L = {{{{\mathscr}X}}^r_{m,n}}\times_{{\mathbb{P}}W} {\mathbb{P}}_L, \qquad Y^r_L = {{{{\mathscr}Y}}^r_{m,n}}\times_{{\mathbb{P}}W^\vee} {\mathbb{P}}^L.$$
We also write: $$Z_L^r = {{{\mathscr}Z}}^{r}_{m,n} \cap {\mathbb{P}}_L, \qquad Z_r^L = {{{\mathscr}Z}}^{m-r}_{m,n} \cap {\mathbb{P}}^L.$$ We will drop $r$, $n$ and/or $m$ from the notation when no confusion is possible. We will always assume that $L \subset W$ is an [*admissible subspace*]{} in the sense of [@kuznetsov:hpd], which amounts to ask that $X_L$ and $Y_L$ have expected dimension. This means that we have: $$\begin{aligned}
{{\rm dim}\,}Z_L = {{\rm dim}\,}X_L & = {{\rm dim}\,}{{{{\mathscr}X}}^r_{m,n}}- c = r(n+m-r)-c-1 \\
{{\rm dim}\,}Z^L = {{\rm dim}\,}Y_L & = {{\rm dim}\,}{{{{\mathscr}Y}}^r_{m,n}}- (mn-c) = r(m-n-r)+c-1.\end{aligned}$$
Let us now give another interpretation of the choice of our linear subspace $L \subset W$. To this purpose we consider the Grassmann variety ${\mathbb{G}}(V,r)$ with the its tautological rank-$r$ quotient bundle which we denote by ${{{\mathscr}T}}$. Dually, we consider ${\mathbb{G}}(V^\vee,m-r)$ and denote by ${{{\mathscr}S}}^\vee$ the tautological quotient bundle of rank $m-r$. Observe that there are natural isomorphisms: $$\begin{aligned}
L{^{\vee}}{\otimes}W = L{^{\vee}}{\otimes}U {\otimes}V &
\simeq {{\rm Hom}}(L {\otimes}{{{\mathscr}O}}_{{\mathbb{G}}}, V {\otimes}{{{\mathscr}Q}}) \simeq \\
& \simeq L{^{\vee}}{\otimes}H^0({{{\mathscr}X}},{{{\mathscr}O}}_{{{{\mathscr}X}}}(H) )\simeq \\
& \simeq {{\rm Hom}}(L \otimes {{{\mathscr}O}}_{{\mathbb{G}}(V,r)}, U\otimes {{{\mathscr}T}}) \nonumber.\end{aligned}$$ There are similar isomorphisms for ${\mathbb{G}}(V^\vee,m-r)$. We denote by $s_L$ the global section of $L{^{\vee}}{\otimes}H^0({{{\mathscr}X}},{{{\mathscr}O}}_{{{{\mathscr}X}}}(H))$ corresponding to $L \subset W$ via these isomorphisms. The subspace $L$ corresponds also to morphisms of bundles on the Grassmann varieties, which we write as: $$M_L : L {\otimes}{{{\mathscr}O}}_{{\mathbb{G}}} \to V {\otimes}{{{\mathscr}Q}}, \qquad
N_L : L \otimes {{{\mathscr}O}}_{{\mathbb{G}}(V,r)} \to U\otimes {{{\mathscr}T}}$$ We also write: $$M^{L} : L^\perp {\otimes}{{{\mathscr}O}}_{{\mathbb{G}}} \to V{^{\vee}}{\otimes}{{{\mathscr}U}}{^{\vee}}, \qquad
N^{L} : L^\perp \otimes {{{\mathscr}O}}_{{\mathbb{G}}(V^\vee,m-r)} \to U{^{\vee}}\otimes {{{\mathscr}S}}{^{\vee}}$$ for the morphisms corresponding to $L^\perp \subset U{^{\vee}}{\otimes}V{^{\vee}}$.
\[mi sa che bisogna dimostrarla\] We have the following equivalent descriptions of $X_L$:
(i) \[e uno\] the vanishing locus ${\mathbb{V}}(s_L)$ of the section $s_L \in L{^{\vee}}{\otimes}H^0({{{\mathscr}X}},{{{\mathscr}O}}_{{{{\mathscr}X}}}(H))$;
(ii) \[e due\] the projectivization of ${{\rm coker}}(M_L)$;
(iii) \[e tre\] the projectivization of ${{\rm coker}}(N_L)$;
(iv) \[e quattro\] the Grassmann bundle ${\mathbb{G}}({{{\mathscr}F}}|_{{\mathbb{P}}_L},m-r)$.
Dually, the variety $Y_L$ is:
(i) the vanishing locus of the section $s^L \in
(W/L) {\otimes}H^0({{{\mathscr}X}},{{{\mathscr}O}}_{{{{\mathscr}X}}}(H))$;
(ii) the projectivization of ${{\rm coker}}(M^L)$;
(iii) the projectivization of ${{\rm coker}}(N^L)$;
(iv) the Grassmann bundle ${\mathbb{G}}({{{\mathscr}E}}|_{{\mathbb{P}}^L},r)$.
We work out the proof for $X_L$, the dual case $Y_L$ being analogous. First recall that the map ${{{\mathscr}X}}\to {\mathbb{P}}W$ is defined by the linear system ${{{\mathscr}O}}_{{{{\mathscr}X}}}(H)$, while the inclusion ${\mathbb{P}}_L \subset
{\mathbb{P}}W$ corresponds to the projection $W \to W/L$. Hence the fibre product defining $X_L$ is given by the vanishing of the global sections in $H^0({{{\mathscr}X}},{{{\mathscr}O}}_{{{{\mathscr}X}}}(H))$ which actually lie in $L$, *i.e. by the vanishing of $s_L$, so is clear.*
For we use essentially the same proof of Lemma \[un peu penible\]. Indeed, given a scheme $S$ over our field, an $S$-valued point of ${\mathbb{P}}({{\rm coker}}(M_L))$ is defined by a morphism $t : S \to {\mathbb{G}}$ together with the isomorphism class of a quotient $y : t^* ({{\rm coker}}(M_L)) \to {{{\mathscr}L}}$, with ${{{\mathscr}L}}$ invertible on $S$. On the other hand, an $S$-valued point of $X_L$ is given by a morphism $s : S
\to X_L$. Once given $s$, composing with $X_L \to {{{\mathscr}X}}\to {\mathbb{G}}$ we obtain the morphism $t$. By the definition of ${{{\mathscr}X}}$ as projective bundle, together with $t$ we get a map $V {\otimes}t^*{{{\mathscr}Q}}\to {{{\mathscr}L}}$ with ${{{\mathscr}L}}$ invertible on $S$. This map composes to zero with $t^*(M_L) : L
{\otimes}{{{\mathscr}O}}_S \to V {\otimes}t^*{{{\mathscr}Q}}$ since the image of $s$ is contained in $X_L$, hence in the vanishing locus of the linear section $s_L$. Therefore this map factors through $t^*({{\rm coker}}(M_L))$ and provides the quotient $y$. It is not hard to check that this procedure can be reversed, which finally proves .
The statement is proved in a similar fashion, while is just Lemma \[un peu penible\], restricted to ${\mathbb{P}}_L$.
The noncommutative desingularization
------------------------------------
In [@buch-leu-vdbergh1; @buch-leu-vdbergh2], noncommutative resolutions of singularities for the affine cone over ${{{\mathscr}Z}}^r={{{\mathscr}Z}}^{r}_{m,n}$ are constructed. This is done by considering the vector bundles $V \otimes {{{\mathscr}Q}}$ instead of their projectivization, and Kapranov’s strong exceptional collection on the Grassmannian [@kapra-grassa] (for the details see [@buch-leu-vdbergh2]). Here we carry on this construction to the projectivized determinantal varieties.
Consider ${{{\mathscr}X}}={{{{\mathscr}X}}^r_{m,n}}$ as rank$-(r n-1)$ projective bundle $p:{{{\mathscr}X}}\to {\mathbb{G}}$. Orlov [@orlov:proj_bundles] gives a semiorthogonal decomposition $$\label{eq:lefschetz-for-Xtondo}
{{\rm D}^{\rm b}}({{{\mathscr}X}})= \langle p^*{{\rm D}^{\rm b}}({\mathbb{G}}), \ldots, p^*{{\rm D}^{\rm b}}({\mathbb{G}})((r n-1)H)\rangle.$$ On the other hand, Kapranov shows that ${\mathbb{G}}$ has a full strong exceptional collection [@kapra-grassa] consisting of vector bundles. We obtain then an exceptional collection on ${{{\mathscr}X}}$ consisting of vector bundles, and hence a tilting bundle $E$ as the direct sum of the bundles from the exceptional collection. Let us consider $M:=Rf_*E$, and let ${{{\mathscr}R}}:={{{\mathscr}{E}nd}}(E)$ and ${{{\mathscr}R}}':={{{\mathscr}{E}nd}}(M)$ (where ${{{\mathscr}{E}nd}}$ denotes the sheaf of endomorphisms).
\[prop:noncomm-resol-general\] The endomorphism algebra ${{{\mathscr}{E}nd}}(M)$ is a coherent ${{{\mathscr}O}}_{{{{\mathscr}Z}}^r}$-algebra Morita-equivalent to ${{{\mathscr}R}}$. In particular, ${{\rm D}^{\rm b}}({{{\mathscr}Z}}^r,{{{\mathscr}R}}) \simeq {{\rm D}^{\rm b}}({{{\mathscr}X}})$ is a categorical resolution of singularities, which is crepant if $m=n$.
First of all, since ${\mathbb{G}}$ has a strong full exceptional collection, we have a tilting bundle $G$ over it. A cohomological calculation, together with the semiorthogonal decomposition provides a tilting bundle $E= \bigoplus_{i=0}^{nr-1} p^* G \otimes {{{\mathscr}O}}_{{{{\mathscr}X}}}(iH)$ over ${{{\mathscr}X}}$. We have thus: $${{\rm D}^{\rm b}}({{{\mathscr}X}}) \simeq {{\rm D}^{\rm b}}(\mathrm{End} (E)).$$ Since the exceptional locus of $f$ has codimension greater than one, [@vdb:flops Lemma 4.2.1] implies that $f_* {{{\mathscr}R}}$ is reflexive. There is a natural map $f_*{{{\mathscr}R}}\to {{{\mathscr}R}}'$ of reflexive sheaves which explicitly reads: $$f_*{{{\mathscr}R}}= \bigoplus_{i,j = 0}^{nr-1} f_*{{{\mathscr}{E}nd}}(p^*G)(i-j) \to
\bigoplus_{i,j = 0}^{nr-1} {{{\mathscr}{E}nd}}(f_*p^*G)(i-j) = {{{\mathscr}R}}'.$$ Again, since the exceptional locus of $f$ has codimension greater than one the locus where $f_*{{{\mathscr}R}}$ and ${{{\mathscr}R}}'$ may be non-isomorphic has codimension at least 2. Since both sheaves are reflexive, we obtain $f_*{{{\mathscr}R}}\cong {{{\mathscr}R}}'$ (compare with [@buch-leu-vdbergh1 Proposition 6.5]). Moreover we know from [@buch-leu-vdbergh2 Proposition 3.4] that $R^kf_* {{{\mathscr}R}}= 0$ for $k>0$ so we actually have: $$Rf_*{{{\mathscr}R}}\cong {{{\mathscr}R}}'.$$ Therefore: $$\mathrm{End}(E) \simeq H^\bullet({{{\mathscr}R}}) \simeq H^\bullet(Rf_* {{{\mathscr}R}}) \simeq H^\bullet({{{\mathscr}R}}').$$ We have now proved: $${{\rm D}^{\rm b}}({{{\mathscr}X}}) \simeq {{\rm D}^{\rm b}}(\mathrm{End}(E)) \simeq {{\rm D}^{\rm b}}(H^\bullet({{{\mathscr}R}}'))
\simeq {{\rm D}^{\rm b}}({{{\mathscr}Z}}^r,{{{\mathscr}R}}').$$
Finally, ${{{\mathscr}R}}'$ is maximally Cohen-Macaulay by [@buch-leu-vdbergh2 Proposition 3.4] (as this property is local) and has finite global dimension since it is Morita-equivalent to the endomorphism algebra ${{{\mathscr}R}}$, which is defined over a smooth variety. If $m=n$, the variety ${{{\mathscr}Z}}^r$ has Gorenstein singularities and $f$ is a crepant resolution, so that the noncommutative resolution is also crepant (compare with [@buch-leu-vdbergh1]).
Homological projective duality for matrices of bounded rank
-----------------------------------------------------------
With this in mind, we can prove our main result directly from Kuznetsov’s HPD for the projective bundles ${{{{\mathscr}X}}^r_{m,n}}={{{\mathscr}X}}$ and ${{{{\mathscr}Y}}^r_{m,n}}={{{\mathscr}Y}}$. We consider the rectangular Lefschetz decomposition for ${{{\mathscr}X}}$ with respect to ${{{\mathscr}O}}_{{{\mathscr}X}}(H)$.
\[thm:main1\] The morphism $g :
{{{\mathscr}Y}}\to {\mathbb{P}}W^\vee$ is the homological projective dual of $f : {{{\mathscr}X}}\to {\mathbb{P}}W$, relatively over ${\mathbb{G}}$, with respect to the rectangular Lefschetz decomposition induced by ${{{\mathscr}O}}_{{{\mathscr}X}}(H)$, generated by $n r
{m\choose r}$ exceptional bundles.
Given the setup of §\[desingo\], we consider the vector bundles $V
{\otimes}{{{\mathscr}Q}}$ and $V^\vee {\otimes}{{{\mathscr}U}}^\vee$ over ${\mathbb{G}}$ and recall that ${{{\mathscr}X}}={\mathbb{P}}(V
{\otimes}{{{\mathscr}Q}})$ and ${{{\mathscr}Y}}={\mathbb{P}}(V^\vee {\otimes}{{{\mathscr}U}}^\vee)$.
Set ${\begin{bf}A\end{bf}}= p^*({{\rm D}^{\rm b}}({\mathbb{G}}))$. The decomposition of the projective bundle ${{{\mathscr}X}}\to {\mathbb{G}}$ then reads: $${{\rm D}^{\rm b}}({{{\mathscr}X}}) = \langle {\begin{bf}A\end{bf}}, {\begin{bf}A\end{bf}}(H), \ldots, {\begin{bf}A\end{bf}}((r n-1)H) \rangle.$$ This is a rectangular Lefschetz decomposition with respect to ${{{\mathscr}O}}_{{{{\mathscr}X}}} (H)$, generated by $nr$ copies of Kapranov’s exceptional collection on ${\mathbb{G}}$, hence by $n r
{m\choose r}$ exceptional bundles.
Clearly the vector bundles $V
{\otimes}{{{\mathscr}Q}}$ and $V^\vee {\otimes}{{{\mathscr}U}}^\vee$ are generated by their global sections, so we may apply apply [@kuznetsov:hpd Corollary 8.3] to their projectivization (actually we use the Grothendieck’s notation for projectivized bundles rather than the usual notation as in [@kuznetsov:hpd], but this does affect the result). The evaluation map of global sections of $V {\otimes}{{{\mathscr}Q}}$ gives tensored with the identity over $V$ i.e.: $$0 \to V {\otimes}{{{\mathscr}U}}\to W \to V {\otimes}{{{\mathscr}Q}}\to 0.$$
This says that $V^\vee {\otimes}{{{\mathscr}U}}^\vee$ is the orthogonal in Kuznetsov’s sense of $V {\otimes}{{{\mathscr}Q}}$. Also, the morphism associated with the tautological line bundle $H_X$ over ${{{\mathscr}X}}$ is $f$, while $g$ is associated with $H_Y$ over ${{{\mathscr}Y}}$. Therefore [@kuznetsov:hpd Corollary 8.3] applies and gives the result.
Note that ${{\rm D}^{\rm b}}({{{\mathscr}Y}})$ is generated by $n (m-r) {m \choose r}$ exceptional vector bundles.
We can rephrase this in terms of categorical resolutions, as a consequence of Proposition \[prop:noncomm-resol-general\]. In this way, one can state HPD as a duality between categorical resolutions of determinantal varieties given by matrices of fixed rank and corank. This leads us to prove our second main Theorem.
\[thm:main2\] There is a ${{{\mathscr}O}}_{{{{\mathscr}Z}}^r}$-algebra ${{{\mathscr}R}}'$ such that $({{{\mathscr}Z}}^r, {{{\mathscr}R}}')$ is a categorical resolution of singularities of ${{{\mathscr}Z}}^r$. Moreover, ${{\rm D}^{\rm b}}({{{\mathscr}Z}}^{r}, {{{\mathscr}R}}') \simeq {{\rm D}^{\rm b}}({{{\mathscr}X}})$ so that $({{{\mathscr}Z}}^{r}, {{{\mathscr}R}}')$ is HP-dual to $({{{\mathscr}Z}}^{m-r},{{{\mathscr}R}}')$.
Recall that ${{{\mathscr}X}}$ is a projective bundle over a Grassmann variety, and hence has a full exceptional sequence. By applying Proposition \[prop:noncomm-resol-general\] to the full exceptional sequence on ${{{\mathscr}X}}$, we get the first statement. The second statement is now straightforward from Theorem \[thm:main1\], together with the isomorphism ${{{\mathscr}X}}^{m-r}\simeq {{{\mathscr}Y}}^r$.
Semiorthogonal decompositions for linear sections {#dimensions}
-------------------------------------------------
Let $L$ be a dimension $c$ subspace of $U{\otimes}V=W$, given by the choice of an element $t \in L{^{\vee}}\otimes W$. Recall that we assume that the subspace $L \subset W$ is [*admissible*]{} in the sense of [@kuznetsov:hpd]. This happens if $L$ is general enough in $W$.
Moreover, again if $L$ is general enough, the singularities of $Z_L=Z_L^r$ appear precisely along $Z_L^{r-1}$. Also, the map $f$, for the rank $r$ locus, is an isomorphism when restricted to $Z_L \setminus Z_L^{r-1}$. Furthermore, we recall from the preceding section that $Z_L$ is a determinantal variety inside ${\mathbb{P}}^{mn-c-1}$ given by a $m \times n$ matrix of linear forms and ${{\rm D}^{\rm b}}(Z_L, {{{\mathscr}R}}'_{{\mathbb{P}}_L})$ is a categorical resolution of singularities of $Z_L$, where ${{{\mathscr}R}}'_{{\mathbb{P}}_L}$ is the pull-back of ${{{\mathscr}R}}'$ from ${{{\mathscr}Z}}_{m,n}^r$ to $Z_L$ under the natural restriction map.
Notice that if $Z_L$ is smooth, then ${{\rm D}^{\rm b}}(Z_L) \simeq {{\rm D}^{\rm b}}(X_L)$, in fact, $Z_L \simeq X_L$ in this case. Similarly, if $Z^L=Z^L_{r}$ is smooth, then ${{\rm D}^{\rm b}}(Z^L) \simeq {{\rm D}^{\rm b}}(Y_L)$ as again $Z^L \simeq Y_L$ in this case. In particular, in the smooth case, the sheaves of algebras ${{{\mathscr}R}}'_{{\mathbb{P}}_L}$ are Morita-equivalent to the structure sheaf.
Our goal now is to draw consequences from the homological projective duality that we have displayed. Notably we will give in several examples a positive answer to the questions asked in the introduction, *i.e. Bondal’s Question \[qu:bondal\] and question \[qu:catrep=ration\] concerning rationality and categorical representability. Remember that ${{{\mathscr}X}}$ (respectively ${{{\mathscr}Y}}$) is the projectivization of a vector bundle of rank $nr$ (resp. $n(m-r)$) over ${\mathbb{G}}$. Hence, by Orlov’s result ([@orlov:proj_bundles]) on the semiorthogonal decompositions for projective bundles we have: $$\begin{aligned}
{{\rm D}^{\rm b}}({{{\mathscr}X}})& =& \langle {\begin{bf}A\end{bf}}, {\begin{bf}A\end{bf}}(H), \ldots, {\begin{bf}A\end{bf}}((nr-1)(H) \rangle;\nonumber\\
\nonumber {{\rm D}^{\rm b}}({{{\mathscr}Y}}) & = & \langle {\begin{bf}B\end{bf}}((1-nm+nr)H), \ldots, {\begin{bf}B\end{bf}}(-H),{\begin{bf}B\end{bf}} \rangle,\end{aligned}$$ where ${\begin{bf}A\end{bf}}$ and ${\begin{bf}B\end{bf}}$ are the respective pull-backs of ${{\rm D}^{\rm b}}({\mathbb{G}})$ to the projective bundles. This in turn implies that, via HPD, when we intersect ${{{\mathscr}X}}$ with ${\mathbb{P}}_L$ and ${{{\mathscr}Y}}$ with ${\mathbb{P}}^L$, we have the following*
$$\begin{aligned}
{{\rm D}^{\rm b}}(X_L)& =& \langle {\begin{bf}C\end{bf}}_L,{\begin{bf}A\end{bf}}(H), \ldots, {\begin{bf}A\end{bf}}(nr-c)(H) \rangle;\nonumber\\
\nonumber {{\rm D}^{\rm b}}(Y_L) & = & \langle {\begin{bf}B\end{bf}}((-c+nr)H), \ldots, {\begin{bf}B\end{bf}}(-H),{\begin{bf}C\end{bf}}_L
\rangle. \end{aligned}$$
Recalling that ${{\rm D}^{\rm b}}(X_L) \simeq {{\rm D}^{\rm b}}(Z_L^{r},{{{\mathscr}R}}'_{{\mathbb{P}}_L})$ and ${{\rm D}^{\rm b}}(Y_L)={{\rm D}^{\rm b}}(Z^L_r,{{{\mathscr}R}}'_{{\mathbb{P}}^L})$ are categorical resolutions of singularities of dual determinantal varieties, we get: $$\begin{aligned}
{{\rm D}^{\rm b}}(Z^{r}_L,{{{\mathscr}R}}'_{{\mathbb{P}}_L})& =& \langle {\begin{bf}C\end{bf}}_L,{\begin{bf}A\end{bf}}(H), \ldots, {\begin{bf}A\end{bf}}(nr-c)(H) \rangle;\nonumber\\
\nonumber {{\rm D}^{\rm b}}(Z_r^L,{{{\mathscr}R}}'_{{\mathbb{P}}^L}) & = & \langle {\begin{bf}B\end{bf}}((-c+nr)H), \ldots, {\begin{bf}B\end{bf}}(-H),{\begin{bf}C\end{bf}}_L \rangle.\end{aligned}$$
Finally, the categories ${\begin{bf}A\end{bf}}$ and ${\begin{bf}B\end{bf}}$ are both generated by ${m \choose r}$ exceptional objects.
\[cor:functors\] Suppose that $L \subset W$ is admissible of dimension $c$.
(i) If $c>nr$, there is a fully faithful functor $${{\rm D}^{\rm b}}(Z_L,{{{\mathscr}R}}'_{{\mathbb{P}}_L}) \simeq {{\rm D}^{\rm b}}(X_L) \longrightarrow {{\rm D}^{\rm b}}(Y_L) \simeq {{\rm D}^{\rm b}}(Z_{r}^L, {{{\mathscr}R}}'_{{\mathbb{P}}^L})$$ whose orthogonal complement is given by $c-nr$ copies of ${{\rm D}^{\rm b}}({\mathbb{G}})$, and is then generated by $(c-nr){m \choose r} $ exceptional objects.
(ii) If $nr=c$, there is an equivalence $${{\rm D}^{\rm b}}(Z_L,{{{\mathscr}R}}'_{{\mathbb{P}}_L}) \simeq {{\rm D}^{\rm b}}(X_L) \simeq {{\rm D}^{\rm b}}(Y_L) \simeq {{\rm D}^{\rm b}}(Z_{r}^L, {{{\mathscr}R}}'_{{\mathbb{P}}^L})$$
(iii) If $c<nr$, there is a fully faithful functor $${{\rm D}^{\rm b}}(Z_{r}^L,{{{\mathscr}R}}'_{{\mathbb{P}}^L}) \simeq {{\rm D}^{\rm b}}(Y_L) \longrightarrow {{\rm D}^{\rm b}}(X_L) \simeq {{\rm D}^{\rm b}}(Z_L,{{{\mathscr}R}}'_{{\mathbb{P}}_L})$$ whose orthogonal complement is given by $nr-c$ copies of ${{\rm D}^{\rm b}}({\mathbb{G}})$, and is then generated by ${m \choose r}(nr-c)$ exceptional objects.
The statement is obtained applying Kuznetsov’s Theorem \[thm:HPD\] to the pair of HPDual varieties from Theorem \[thm:main1\], and using the resolutions of singularities described in Theorem \[thm:main2\]. The functors involved can be explicitly described as Fourier–Mukai with kernels in ${{\rm D}^{\rm b}}(X_L \times Y_L)$ (see the detailed description in the original Kuznetsov’s paper [@kuznetsov:hpd §5]).
Using the notation introduced in Section \[desingo\] for the generators of the Picard group, we have the following formula for the canonical bundle of ${{{\mathscr}X}}$: $$\omega_{{{{\mathscr}X}}} \simeq {{{\mathscr}O}}_{{{{\mathscr}X}}}(-nr H +(n-m)P).$$
A consequence of this formula is the following lemma. Call $\phi_K$ the canonical map of $X_L$, i. e. the rational map associated with the linear system $|\omega_{X_L}|$. Write also $\phi_{-K}$ for the map associated with $|\omega_{X_L}^\vee|$ (i.e. the anticanonical map).
\[adjuncion\] The canonical bundle of the linear section $X_L$ is: $$\omega_{X_L} \simeq {{{\mathscr}O}}_{X_L}((c-nr) H + (n-m)P).$$
i) \[when CY\] The variety $X_L$ is Calabi-Yau if and only if $m=n$ and $c=nr$.
ii) \[when positive\] If $c>nr$, or if $c=nr$ and $n > m$, $\phi_K$ is a birational morphism onto its image.
iii) \[when negative\] If $c < nr$ and $m=n$, $\phi_{-K}$ is a birational morphism onto its image. If moreover $X_L^{r-1}=\emptyset$, $\phi_{-K}$ is an embedding and $X_L$ is Fano.
The formula for $\omega_{X_L}$ is obvious by adjunction. By this formula, $\omega_{X_L}\simeq {{{\mathscr}O}}_{X_L}$ whenever $m=n$ and $c=nr$. Conversely, remark that $X_L$ is connected, so if $X_L$ is CY, then there is no nontrivial semiorthogonal decomposition of ${{\rm D}^{\rm b}}(X_L)$. Corollary \[cor:functors\] forces then $c \geq nr$.
Suppose $c > nr$, or $c=nr$ and $n>m$. Notice first that both $P$ and $H$ are nef. Then canonical divisor is a linear combination of nef divisors with positive coefficients, which is in turn nef. On the other hand, we have that $\omega_{X_L}$ is ${{{\mathscr}O}}_{X_L}$ if $c=nr$ and $m=n$, so using $c \geq nr$ we conclude the proof of .
For , by definition $H$ and $P$ are base-point-free and $H$ is very ample away from the exceptional locus of $f$, so the statement follows directly from the formula for $\omega_{X_L}$. A similar argument proves .
\[cor:adjunction\] We have the following formulas for the canonical bundles $$\begin{aligned}
&\omega_{{{\mathscr}Y}}\simeq {{{\mathscr}O}}_{{{\mathscr}Y}}(-n(m-r) H + (n-m)Q), \\
&\omega_{Y_L} \simeq {{{\mathscr}O}}_{Y_L}((nr-c) H + (n-m)Q).\end{aligned}$$
In particular, $Y_L$ is Calabi-Yau if and only if $m=n$ and $c=nr$. If $c<nr$, or if $c=nr$ and $n > m$, then the canonical map of $Y_L$ is a birational morphism onto its image. If $c > nr$, $m=n$ and $Y_L^{r-1}=\emptyset$ then $Y_L$ is Fano.
Everything follows from the isomorphism ${{{\mathscr}Y}}^{r} \simeq
{{{\mathscr}X}}^{m-r}$. Indeed, we find $Y^r_L \simeq X_{L^\perp}^{m-r}$: and, since ${{\rm dim}\,}L^\perp + {{\rm dim}\,}L = {{\rm dim}\,}(W) = nm$, we get the formula for $\omega_{Y_L}$ from Lemma \[adjuncion\], recalling that the relative hyperplane section is identified with $Q$ in this case. The other statements follow as in Lemma \[adjuncion\].
We resume in Table \[table:all-r\] the results of this section. The functor mentioned there is the HPD functor.
$c<nr$ $c=nr$ $c>nr$
------------- ----------------------------------------------------- ---------------------------- -----------------------------------------------------
HPD Functor ${{\rm D}^{\rm b}}(Y_L) \to {{\rm D}^{\rm b}}(X_L)$ equivalence ${{\rm D}^{\rm b}}(X_L) \to {{\rm D}^{\rm b}}(Y_L)$
nef canonical nef canonical if $n\neq m$
Fano visitor if $n=m$ CY if $n=m$ Fano if $n=m$
nef canonical if $n\neq m$ nef canonical
Fano if $n=m$ CY if $n=m$ Fano visitor if $n=m$
: Behaviour of HPD functors according on $c$ and $nr$.[]{data-label="table:all-r"}
Birational and equivalent linear sections {#sect:CY}
=========================================
As explained in Corollary \[cor:functors\] and then displayed in Table \[table:all-r\], the condition $c=nr$ guarantees that HPD gives an equivalence of categories. Hence our construction gives examples of derived equivalences of Calabi-Yau manifolds for any $n=m$. One first example was produced in [@vivalafisica]. In fact the authors of [@vivalafisica] take $n=m=4$, $r=2$, the self dual orbit of rank 2, $4\times 4$ matrices and consider the codimension eight threefolds obtained by taking orthogonal linear sections in ${\mathbb{P}}^{15}$. In fact, our construction shows that these two Calabi-Yau are derived equivalent. On the other hand it is very likely that they are one the flop of the other. We can show indeed that $X_L$ and $Y_L$ are birational whenever $c=nr$.
Assume now that $c=nr$. Remark that the two vector bundles appearing in the map $M_L$ of Proposition \[mi sa che bisogna dimostrarla\] have the same rank, namely $nr$. Let us denote by $D_L$ the hypersurface in ${\mathbb{G}}$ defined by the vanishing of determinant of $M_L$: $$M_L : L {\otimes}{{{\mathscr}O}}_{{\mathbb{G}}} \to V {\otimes}{{{\mathscr}Q}}.$$
The degree of $D_L$ is $n$. Dually, we write $D^L$ the hypersurface in ${\mathbb{G}}$ whose equation is the determinant of: $$M^{L} : L^\perp {\otimes}{{{\mathscr}O}}_{{\mathbb{G}}} \to V{^{\vee}}{\otimes}{{{\mathscr}U}}{^{\vee}}.$$
If $c=nr$ then $D^L=D_L$, and $X_L$ is birational to $Y_L$.
To see this, we write the following exact commutative diagram: $$\xymatrix@-3ex{
&& 0 \ar[d] & 0 \ar[d] & \\
&& V{^{\vee}}{\otimes}{{{\mathscr}Q}}{^{\vee}}\ar[d] \ar@{=}[r] & V{^{\vee}}{\otimes}{{{\mathscr}Q}}{^{\vee}}\ar^-{(M_L)^*}[d]\\
0 \ar[r] &L^\perp {\otimes}{{{\mathscr}O}}_{{\mathbb{G}}} \ar@{=}[d] \ar[r] & U{^{\vee}}{\otimes}V{^{\vee}}{\otimes}{{{\mathscr}O}}_{{\mathbb{G}}}
\ar[d] \ar[r] & L{^{\vee}}{\otimes}{{{\mathscr}O}}_{{\mathbb{G}}} \ar[r] \ar[d] & 0\\
& L^\perp {\otimes}{{{\mathscr}O}}_{{\mathbb{G}}} \ar^-{M^L}[r] & V{^{\vee}}{\otimes}{{{\mathscr}U}}{^{\vee}}\ar[d] \ar[r] & {{{\mathscr}K}}\ar[d] \ar[r] & 0\\
& & 0 & 0
}$$ Here, ${{{\mathscr}K}}$ is the cokernel both of $M^L$ and of $(M_L)^*$. This says that: $$D^L={\mathbb{V}}(\det(M^L))={\mathbb{V}}(\det(M_L{^{\vee}}))={\mathbb{V}}(\det(M_L))=D_L.$$
Now let us look at $X_L$ and $Y_L$. The sheaf ${{{\mathscr}K}}$ is supported on $D=D^L$, and is actually of the form $\iota_*({{{\mathscr}K}}_r)$, where ${{{\mathscr}K}}_r$ is a reflexive sheaf of rank $1$ on $D$ and $\iota : D \to {\mathbb{G}}$ is the natural embedding. The cokernel of $M_L$ is also of the form $\iota_*({{{\mathscr}K}}^r)$, with ${{{\mathscr}K}}^r$ reflexive of rank $1$ on $D$. By Grothendieck duality, since $D$ has degree $n$, the previous diagram says that ${{{\mathscr}K}}^r \simeq {{{\mathscr}K}}_r^\vee(n)$. On the (open and dense) locus of $D$ where ${{{\mathscr}K}}^r$ and ${{{\mathscr}K}}_r$ are locally free, the variety $D$ coincides with $X_L$ and $Y_L$. Therefore, by Proposition \[mi sa che bisogna dimostrarla\], these varieties are both birational to $D$.
A priori, $X_L$ is not isomorphic to $Y_L$, as the projectivization of the two sheaves ${{{\mathscr}K}}_r$ and ${{{\mathscr}K}}_r^\vee$ gives in principle non-isomorphic varieties (cf. Example \[noniso\] below). This does not happen if ${{{\mathscr}K}}_r$ is locally free of rank $1$ on $D$, which in turn is the case if $D$ is smooth. Also, when the singularities of $D$ are isolated points, then in order for ${\mathbb{P}}({{{\mathscr}K}}_r)$ to be isomorphic to ${\mathbb{P}}({{{\mathscr}K}}_r^\vee)$, it suffices to check that the rank of ${{{\mathscr}K}}_r^\vee$ and ${{{\mathscr}K}}_r$ is the same at those points, and this is of course true. Then we have:
Suppose that $D$ is smooth or has isolated singularities, then $X_L$ is isomorphic to $Y_L$.
If we assume that $X_L$ is Calabi-Yau, then $m=n$ and $c=nr$ so we are in a sub-case of our description above, and birationality still holds. Thus, in dimension 3, the derived equivalences would follow also from the work of Bridgeland [@bridg-flop].
\[noniso\] Let us describe an example of two determinantal varieties $X_L$ and $Y_L$ which are derived equivalent, birational, but not isomorphic. Actually one can describe infinitely many examples this way, all of dimension at least $5$. In all of them the canonical system is birational onto a hypersurface of general type in ${\mathbb{G}}$.
Take $(r,m,n)=(3,5,7)$, $c=21$ and consider a general subspace $L
\subset W$. Then $X_L$ and $Y_L$ are both smooth projective $5$-folds. The Picard group ${{\rm Pic}}(X_L)$ is isomorphic to ${\mathbb{Z}}^2$, generated by (the restriction of) $H_X$ and $P$, while ${{\rm Pic}}(Y_L)$ is also isomorphic to ${\mathbb{Z}}^2$, generated by $H_Y$ and $Q$. Note that $\omega_{X_L} \simeq {{{\mathscr}O}}_{X_L}(2P)$ while $\omega_{Y_L} \simeq {{{\mathscr}O}}_{Y_L}(2Q)$.
We claim that $X_L$ and $Y_L$ are not isomorphic. Indeed, if there was an isomorphism $f:X_L
\to Y_L$, we should have $f^*(Q) = P$ because of the expression of the canonical bundle. Since $(f^*(Q),f^*(H_{Y_L}))$ should form a ${\mathbb{Z}}$-basis of ${{\rm Pic}}(X_L)$, we have $f^*(H_{Y_L})=H_{X_L}+aP$, for some $a \in {\mathbb{Z}}$. But a straightforward computation shows that $(H_{X_L}+aP)^5$ is never equal to $H_{Y_L}^5$, for any choice of $a$.
So the $5$-folds $X_L$ and $Y_L$ are not isomorphic. They are however derived equivalent via HPD and both birational by projection to a determinantal hypersurface $D$ of degree $7$ in ${\mathbb{G}}(5,3)$. The canonical bundle of this hypersurface is ${{{\mathscr}O}}_{D}(2)$. The determinantal model of $X_L$ (respectively, of $Y_L$) is the fivefold of degree $490$ (respectively, $1176$) cut in ${\mathbb{P}}^{13}$ (respectively, in ${\mathbb{P}}^{20}$) by the $4\times 4$ minors (respectively, the $3\times 3$ minors) of a sufficiently general $5\times 7$ matrix of linear forms.
Concerning rationality of determinantal varieties, we have the following result.
\[razionalita\] The variety $X_L$ is rational if $nr > c$; $Y_L$ is rational if $c>nr$.
By Proposition \[mi sa che bisogna dimostrarla\], $Y_L$ is the projectivization of the cokernel sheaf of the map $M^L$. Recall that ${{\rm dim}\,}(L^\perp)=nm-c$ and that $V^\vee \otimes {{{\mathscr}U}}^\vee$ has rank $n(m-r)$.
So if $c>nr$, i.e. if $n(m-r)>nm-c$, there is a Zariski dense open subset of ${\mathbb{G}}(U,r)$ where $M^L$ has constant rank $mn-c$. Hence an open piece of $Y_L$ is the projectivization of a locally free sheaf over a rational variety, so $Y_L$ is rational.
The same argument works for $X_L$.
A side remark is that, using $N^L$ instead of $M^L$ we would rationality of $Y_L$ if $c>m(n-m+r)$. However, one immediately proves that $m(n-m+r) \ge nr$.
The Segre-determinantal duality {#sect:segre-det}
===============================
In this section, we give a more detailed description of the case $r=1$ (we suppress $r$ from our notation for this section). In this case ${{{\mathscr}X}}\simeq {\mathbb{P}}^{n-1}\times{\mathbb{P}}^{m-1}$ is just a product of two projective spaces and $X_L$ is a linear section of a Segre variety. On the other hand, ${{{\mathscr}Y}}$ is the Springer desingularization of the space of degenerate matrices.
For this section and the following ones, we make use of the standard notation ${{{\mathscr}O}}_{X_L}(a,b)$ for the restriction to $X_L$ of ${{{\mathscr}O}}_{{\mathbb{P}}^{n-1}}(a)
\boxtimes {{{\mathscr}O}}_{{\mathbb{P}}^{m-1}}(b)$, so that ${{{\mathscr}O}}_{X_L}(1,1)={{{\mathscr}O}}_{X_L}(H)$ and ${{{\mathscr}O}}_{X_L}(0,1)={{{\mathscr}O}}_{X_L}(P)$. Proposition \[mi sa che bisogna dimostrarla\] and Lemma \[adjuncion\] become:
\[descrXL\] The variety $X_L$ can be described in two following ways:
(i) as the projectivization of the cokernel of: $$L \otimes {{{\mathscr}O}}_{{\mathbb{P}}U}(-1) \to V \otimes {{{\mathscr}O}}_{{\mathbb{P}}U};$$
(ii) as the projectivization of the cokernel of: $$L \otimes {{{\mathscr}O}}_{{\mathbb{P}}V}(-1) \to U \otimes {{{\mathscr}O}}_{{\mathbb{P}}V}.$$
Also, we have the formulas for the canonical bundle: $$\begin{aligned}
\omega_{X_L} & = {{{\mathscr}O}}_{X_L}(c-n,c-m),\end{aligned}$$ In particular $X_L$ is Fano if and only if $c < m$, and rational for $c<n$.
Since $X_L^0=\emptyset$ the condition for $X_L$ to be Fano descends directly for the formula for the canonical bundle. The statement on rationality is Proposition \[razionalita\].
The variety $Y_L$ is itself a linear section of a Segre variety, by Proposition \[mi sa che bisogna dimostrarla\], as the folloing Lemma shows.
The variety $Y_L$ is isomorphic to the complete intersection of $n$ hyperplanes in ${\mathbb{P}}U \times {\mathbb{P}}^L$ determined by $L \subset W$. So the canonical bundle $\omega_{Y_L}$ equals ${{{\mathscr}O}}_{Y_L}(n-m,n-c)$. Moreover, for generic $L\subset W$, the determinantal variety $Z^L$ is smooth if and only if $c<2n-2m+5$.
The first statement follows from the very last item of Proposition \[mi sa che bisogna dimostrarla\]. Indeed, since $r=1$, $Y_L$ the projectivization of the sheaf ${{{\mathscr}E}}$, restricted to ${\mathbb{P}}^L$. Therefore, just as in the proof of Proposition \[mi sa che bisogna dimostrarla\], $Y_L$ is the vanishing locus of the global section of ${{{\mathscr}O}}_{
{\mathbb{P}}U\times {\mathbb{P}}^L}(1,1)$ determined by the subspace $L^\perp \subset W^\vee$, i.e. by $L \subset W$.
Note that ${{{\mathscr}O}}_{Y_L}(0,1)\simeq {{{\mathscr}O}}_{Y_L}(H)$ and ${{{\mathscr}O}}_{Y_L}(1,0)={{{\mathscr}O}}_{Y_L}(Q)$. The canonical bundle formula follows by adjunction and agrees with Corollary \[cor:adjunction\].
The codimension in ${\mathbb{P}}^L$ of the singular locus of $Z^L$ is $2n-2m+4$ for a general choice of $L\subset W$. So $Z^L$ is smooth if and only if $c < 2n-2m+5$, which gives the last statement.
Here, since ${{{\mathscr}U}}{^{\vee}}=T_{{\mathbb{P}}U}(-1)$. By Proposition \[mi sa che bisogna dimostrarla\], the variety $Y_L$ can also be described as the projectivization of the cokernel sheaf of $$\label{pino}
L^{\perp} \otimes {{{\mathscr}O}}_{{\mathbb{P}}U} \to V{^{\vee}}\otimes T_{{\mathbb{P}}U}(-1),$$
The map appearing in (\[pino\]) in the remark above, corresponds once again to the choice of $L \subset W$. Dually, for $Y_L$, Proposition \[razionalita\] gives:
The variety $Y_L$ is rational if $c>n$.
This is just Proposition \[razionalita\].
Thanks to the constructions of section \[sect:CY\], we obtain the following corollary.
If $c=n$, then $X_L$ and $Y_L$ are birational $(m-2)$-folds. If $m=n$ they are Calabi-Yau and have nef canonical divisor otherwise.
We resume the results of this section in Table \[table:r=1\].
Fano and rational varieties {#sect:fano}
===========================
Representability into Fano varieties
------------------------------------
In this section, we consider question \[qu:bondal\]. We start by stating a straightforward consequence of Corollary \[cor:functors\] and Lemma \[adjuncion\] (see also Table \[table:all-r\]), which provides a large class of examples of weakly Fano-visitor (see Def. \[whost\]) varieties, up to categorical resolutions of singularities.
\[Fano-visitor-allr\] Suppose that $n=m$. If $c<rn$, then $Y_L^r$ and $(Z^L_r,{{{\mathscr}R}}'_{{\mathbb{P}}^L})$ are weakly Fano visitor. If $c>nr$, then $X_L^r$ and $(Z^r_L,{{{\mathscr}R}}'_{{\mathbb{P}}_L})$ are weakly Fano visitor.
If $r=1$ we have an interpretation of Proposition \[Fano-visitor-allr\] for determinantal varieties.
\[cor:determinantal-Fanovis\] Let $Z \subset {\mathbb{P}}^k$ be a determinantal variety associated with a generic $m \times n$ matrix. If $k<m-1$ then the categorical resolution of singularities of $Z$ is Fano visitor.
The determinantal variety $Z$ is $Z^L_{m-1}=
{{{\mathscr}Z}}_{m,n}^{m-1}\cap {\mathbb{P}}^L$ for a subspace $L\subset U\otimes V$ of codimension $k+1$. Then we use results from Table \[table:r=1\] and conclude.
Corollary \[cor:determinantal-Fanovis\] gives a positive answer to Question \[qu:bondal\] for almost every curve.
\[ex:plane-cves\] Let $C \subset {\mathbb{P}}^2$ be a plane curve of degree $d \ge 4$. Then, it is well known (see [@bove:det §3]) that $C$ can be written as the determinant of a $d \times d$ matrix of linear forms. In other words, we put $m=n=d$, $k=2$ and the inequality of Corollary \[cor:determinantal-Fanovis\] is respected. Hence any plane curve of degree at least four is a Fano-visitor, up to resolution of singularities.
On the other hand, one can check that the blow-up of ${\mathbb{P}}^3$ along a plane cubic is Fano (see, [*e.g.*]{}, [@blanc-lamy Proposition 3.1, (i)]). Hence any plane curve of positive genus is a Fano-visitor.
Determinantal varieties with $n\neq m$ provide a wealth of examples of (even non plane) curves of general type that are Fano-visitor.
Let us make the case where ${{\rm dim}\,}(Y_L^1) = {{\rm dim}\,}(Z^L)= 1$ explicit. We have $c=n-m+3$. From Table \[table:r=1\] it is straightforward to see that $Y_L^1$ is an elliptic curve (the Calabi-Yau case) if $m=n=c=3$; this yields indeed a plane cubic. On the other hand, we see that if $m=2$ then the curve is rational for any value of $n$ since $c=n+1$, and if $m>3$ it is forced to be a curve of general type in ${\mathbb{P}}^{c-1}$, which is Fano visitor if $c<m$.
The dual $X_L$ is a smooth variety of dimension $2m-5$. If $m=3$, we have that $Z_L$ is an elliptic curve. If $m > 3$, we have ${{\rm dim}\,}Z_L \geq 3$. This gives quite a lot of examples of space curves of general type that are Fano visitors. Take for example $c=4,\ n=6$ and $m=5$. This gives a curve of genus 4 in ${\mathbb{P}}^3$, complete intersection of two degree 5 determinantal hypersurfaces, whose derived category is fully faithfully embedded in the derived category of a rational Fano 5-fold in ${\mathbb{P}}^{25}$.
Rationality and categorical representability
--------------------------------------------
In this subsection, we consider Question \[qu:catrep=ration\]. The second consequence of Corollary \[cor:functors\] is a large class of examples of rational varieties which are categorically representable in codimension at least 2. For simplicity, let us assume that $r=1$, so that we already discussed in section \[sect:segre-det\] the rationality of the sections. We state the following Proposition in terms of Segre and determinantal varieties.
The categorical resolution of a rational determinantal variety is categorically representable in codimension at least 2. A rational linear section of the Segre variety ${\mathbb{P}}^{n-1} \times {\mathbb{P}}^{m-1} \subset {\mathbb{P}}^{nm-1}$ is categorically representable in codimension at least 2.
First we observe that the Segre linear section $X_L$ is rational for $c<n$ and the determinantal linear section $Y_L$ for $c>n$ by Table \[table:r=1\]. Then we recall from Corollary \[cor:functors\] that, assuming $r=1$, it is exactly in these ranges that we have the required functors and semiorthogonal decompositions. A computation of the dimensions of the linear sections, following the formulas in section \[proshe\], proves the claim.
Categorical resolution of the residual category of a determinantal Fano hypersurface
------------------------------------------------------------------------------------
The Segre-determinantal HPD involves categorical resolutions for determinantal varieties, which is crepant if $n=m$. In this subsection we consider the cases where such resolution gives a crepant categorical resolution for nontrivial components of a semiorthogonal decomposition. For simplicity, we will consider only determinantal *hypersurfaces, hence we need to assume $r=1$ and $m=n$. We will drop all the useless indexes.*
Let $F$ be a smooth Fano variety such that ${{\rm Pic}}(F) = {\mathbb{Z}}[{{{\mathscr}O}}_F(1)]$. The index of $F$ is the integer $i$ such that $\omega_F={{{\mathscr}O}}_F(-i)$. Kuznetsov observed that this kind of varieties have a Lefschetz-type semiorthogonal decomposition.
[@kuz:fano Lemma 3.4]\[uno\] Let $F$ be a smooth Fano variety of index $i$, then the collection ${{{\mathscr}O}}_F(-i+1), \dots, {{{\mathscr}O}}_F$ in ${{\rm D}^{\rm b}}(F)$ is exceptional.
[@kuz:fano Corollary 3.5] For any smooth Fano variety $F$ of Picard rank $1$ and index $i$ we have the following semiorthogonal decomposition $$\label{fanodeco}
{{\rm D}^{\rm b}}(F)= \langle {{{\mathscr}O}}_F(-i+1), \ldots, {{{\mathscr}O}}_F,{\begin{bf}T\end{bf}}_F \rangle,$$ where ${\begin{bf}T\end{bf}}_F= \{ E \in {{\rm D}^{\rm b}}(V)| H^\bullet(V,E(-k))=0\ for\ all\ 0\leq k \leq i-1 \}$.
The main technical tools used in the proof of Lemma \[uno\] are Kodaira vanishing Theorem and Serre duality. Before we proceed, we first need to broaden slightly the class of varieties for which the semiorthogonal decomposition (\[fanodeco\]) holds. In fact, we recall that Kodaira vanishing holds also for varieties with rational singularities (for example, see [@lazarsfeld-book I, Example 4.3.13]), and the well-known fact that the canonical divisor of a Gorenstein variety is Cartier.
\[propo-deco-for-singular\] Let $F$ be a projective Gorenstein variety with rational singularities. Suppose that ${{\rm Pic}}(F) = {\mathbb{Z}}$, ${{{\mathscr}O}}_F(1)$ is its (ample) generator and $K_F = {{{\mathscr}O}}_F(-i)$, with $i>0$. Then there is a semiorthogonal decomposition $${{\rm D}^{\rm b}}(F) = \langle {{{\mathscr}O}}_F(-i+1), \ldots, {{{\mathscr}O}}_F, {\begin{bf}T\end{bf}}_F \rangle.$$
This holds in particular if $F \subset {\mathbb{P}}^k$ is an hypersurface of degree $d < k$ with rational singularities (in which case, $i=k-d$).
It is straightforward to check that the line bundle ${{{\mathscr}O}}_F(i)$ is exceptional for any $i$. To show the semiorthogonality, we use a vanishing theorem for varieties with rational singularities (see [@lazarsfeld-book I, Example 4.3.13]), which states that $$\mathrm{Ext}^{j}({{{\mathscr}O}}_F(s),{{{\mathscr}O}}_F(t))\simeq \mathrm{Ext}^{j}({{{\mathscr}O}}_F,{{{\mathscr}O}}_F(t-s))\simeq H^j(F,{{{\mathscr}O}}_F(t-s))$$
vanishes for $j < \mathrm{dim}(F)$, and $s > t$. Thanks to Serre duality $$\mathrm{Ext}^{\mathrm{dim}(F)}({{{\mathscr}O}}_F(s),{{{\mathscr}O}}_F(t))\simeq H^{\mathrm{dim}(F)}(F,{{{\mathscr}O}}_F(t-s)) \simeq H^0(F,{{{\mathscr}O}}_F(s+i-t))$$ and the latter group vanishes if $s+i-t < 0$,
Homological Projective Duality allows us to describe a resolution of singularities of ${\begin{bf}T\end{bf}}_F$ in the case where $F$ is determinantal. This means that we consider $Z^L \subset {\mathbb{P}}^L$ for some integers $m=n$ and for some linear subspace $L\subset U \otimes V$ of Fano type (that is, of degree $d<k+1$). The Springer resolution of $Z^L$ is then $Y_L$ and the dual section of the Segre variety is $X_L$. Let us fix $L$, and drop it from the notations from now on. We want to describe a categorical resolution of the category ${\begin{bf}T\end{bf}}_Z$ described in Proposition \[propo-deco-for-singular\].
We constructed a crepant categorical resolution of singularities ${{\rm D}^{\rm b}}(Z,{{{\mathscr}R}}')$ of $Z$. The category ${{\rm D}^{\rm b}}(Z,{{{\mathscr}R}}')$ is equivalent to ${{\rm D}^{\rm b}}(Y)$, for $Y$ the corresponding fibre product of the linear section of the Springer resolution (see Theorem \[thm:main2\]). In particular, $Y$ is a (the fibre product over a) linear section of a projective bundle over ${\mathbb{P}}^{d-1}$, since $d=n=m$ is the degree of $Z$. Let us denote by $X$ the dual linear section of the Segre variety (notice in fact that $X$ is smooth). Numerical computations provide a semiorthogonal decomposition $${{\rm D}^{\rm b}}(Z,{{{\mathscr}R}}') \simeq {{\rm D}^{\rm b}}(Y) = \langle k-d+1 \text{ copies of } {{\rm D}^{\rm b}}({\mathbb{P}}^{d-1}), {{\rm D}^{\rm b}}(X) \rangle.$$
Hence ${{\rm D}^{\rm b}}(Z,{{{\mathscr}R}}')$ is generated by by $d(k-d+1)$ exceptional objects and ${{\rm D}^{\rm b}}(X)$.
More precisely, the $j$-th occurrence of ${{\rm D}^{\rm b}}({\mathbb{P}}^{d-1})$ can be generated by the exceptional sequence $({{{\mathscr}O}}_{Y}(j,1), \ldots, {{{\mathscr}O}}_{Y}(j,d))$, where we use the same notation ${{{\mathscr}O}}_{Y}(a,b)$ as in Section \[sect:segre-det\], *i.e. ${{{\mathscr}O}}_{Y_L}(0,1)\simeq {{{\mathscr}O}}_{Y_L}(H)$ and ${{{\mathscr}O}}_{Y_L}(1,0)={{{\mathscr}O}}_{Y_L}(Q)$.*
This allows one to calculate a categorical resolution of singularities of ${\begin{bf}T\end{bf}}_{Z}$ which is decomposed into ${{\rm D}^{\rm b}}(X)$ and exceptional objects.
Let $Z$ be a Fano determinantal hypersurface of ${\mathbb{P}}^k$, and $X$ the dual section of the Segre variety. There is a strongly crepant categorical resolution $\widetilde{{\begin{bf}T\end{bf}}}_Z$ of ${\begin{bf}T\end{bf}}_{Z}$, admitting a semiorthogonal decomposition by ${{\rm D}^{\rm b}}(X)$ and $(d-1)(k-d+1)$ exceptional objects.
Consider the resolution $p:Y \to Z$, and denote by $D$ its exceptional divisor. We have proved that ${{\rm D}^{\rm b}}(Y) \simeq {{\rm D}^{\rm b}}(Z,{{{\mathscr}R}}')$ is a categorical resolution of singularities of ${{\rm D}^{\rm b}}(Z)$. In particular (see [@kuznet:singul]), this comes equipped with a functor $p^\vee : \mathrm{Perf}(Z) \to
{{\rm D}^{\rm b}}(X)$ admitting a right adjoint. Indeed, according to [@kuznet:singul], to get such a pair for a variety $M$ with rational singularities, one needs to consider a desingularization $q: N \to M$ with exceptional divisor $E$, such that ${{\rm D}^{\rm b}}(E)$ admits a Lefschetz decomposition with respect to the conormal bundle. In our case, we can just consider the Lefschetz decomposition with one component ${\begin{bf}B\end{bf}}_0 = {{\rm D}^{\rm b}}(D)$. Now we will check that all the hypotheses of [@kuznet:singul Theorem 1] for the existence of such a categorical resolution are satisfied by the category generated by ${{\rm D}^{\rm b}}(X)$ and the exceptional objects. So, in order to get a categorical resolution of singularities for ${\begin{bf}T\end{bf}}_{Z}$, let us consider the functor $p^*$ introduced above and its action on the semiorthogonal decomposition from Proposition \[propo-deco-for-singular\].
Let ${\mathbb{P}}^L\simeq {\mathbb{P}}^k$. There is a commutative diagram:
$$\xymatrix{
Y \ar[d]^p \ar[dr]^f \\
Z \ar[r]^g & {\mathbb{P}}^k,
}$$
where the map $f$ is given by the restriction linear system $\vert {{{\mathscr}O}}_{Y}(1,1) \vert $, and the map $g$ is defined by $\vert {{{\mathscr}O}}_{Z} (1) \vert$. It follows that $p^*{{{\mathscr}O}}_{Z}(k)
= {{{\mathscr}O}}_{Y}(k,k)$, so that the exceptional sequence ${{{\mathscr}O}}_{Z}(-k+d),\ldots,{{{\mathscr}O}}_{Z}$ pulls back to the exceptional sequence ${{{\mathscr}O}}_{Y}(-k+d,-k+d),\ldots,{{{\mathscr}O}}_{Y}$.
Now recall that ${{{\mathscr}Y}}^{1}_{d,d}$ is a projective bundle $s: {{{\mathscr}Y}}^{1}_{d,d} \simeq {\mathbb{P}}(V{^{\vee}}\otimes T_{{\mathbb{P}}(U)}(-1)) \to {\mathbb{P}}U$. The Lefschetz decomposition of ${{\rm D}^{\rm b}}({{{\mathscr}Y}}^{1}_{d,d})$ giving the HP-duality of Theorem \[thm:main1\] is: $${{\rm D}^{\rm b}}({{{\mathscr}Y}}^{1}_{d,d}) = \langle {\begin{bf}A\end{bf}}_{-j}\otimes {{{\mathscr}O}}_{{\mathbb{P}}(V\otimes {{{\mathscr}Q}})}(-j), \ldots, {\begin{bf}A\end{bf}}_0 \rangle,$$ with $-j=1-d^2+d$, where ${\begin{bf}A\end{bf}}_0= \ldots = {\begin{bf}A\end{bf}}_j = s^* {{\rm D}^{\rm b}}({\mathbb{P}}U)$. In particular, we can choose, for each occurrence of $s^* {{\rm D}^{\rm b}}({\mathbb{P}}U)$, an appropriate exceptional collection generating ${{\rm D}^{\rm b}}({\mathbb{P}}U)$ in order to get, after taking the linear sections (recall that $Y:=Y^1_L$, and $X:=X_L^1$):
$$\begin{array}{rl}
{{\rm D}^{\rm b}}(Y) =& \langle {{{\mathscr}O}}_{Y}(-k+d,-k+d), \ldots, {{{\mathscr}O}}_{Y}(-k+d,-k+2d-1), \\
& {{{\mathscr}O}}_{Y}(-k+d+1,-k+d+1), \ldots, {{{\mathscr}O}}_{Y}(-k+d+1,-k+2d), \\
& \ldots \\
& {{{\mathscr}O}}_{Y}(0,0), \ldots, {{{\mathscr}O}}_{Y}(0,d-1), {{\rm D}^{\rm b}}(X) \rangle.
\end{array}$$
Now we can mutate all the exceptional objects which are not of the form ${{{\mathscr}O}}_{Y}(-t,-t)$, for some $t$, to the right until we get
$$\begin{array}{rl}
{{\rm D}^{\rm b}}(Y) =& \langle {{{\mathscr}O}}_{Y}(-k+d,-k+d), \ldots, {{{\mathscr}O}}_{Y}(-1,-1), {{{\mathscr}O}}_{Y}, \\
& E_1, \ldots, E_{(d-1)(k-d+1)}, {{\rm D}^{\rm b}}(X) \rangle,
\end{array}$$ where the $E_i$ are the exceptional objects resulting from the mutations. Hence, the first block is the pull-back from $Z$ of the exceptional sequence $({{{\mathscr}O}}_{Z}(-k+d), \ldots, {{{\mathscr}O}}_{Z})$, then by definition we get that the second block is the categorical resolution of singularities for ${\begin{bf}T\end{bf}}_{Z}$.
A particular and interesting case is given by determinantal cubics in ${\mathbb{P}}^4$ and ${\mathbb{P}}^5$. In both cases, the dual linear section $X$ is empty. So, the numeric values give explicitly:
- If $Z$ is a determinantal cubic threefold, then the category ${\begin{bf}T\end{bf}}_{Z}$ admits a crepant categorical resolution of singularities generated by 4 exceptional objects.
- If $Z$ is a determinantal cubic fourfold, then the category ${\begin{bf}T\end{bf}}_{Z}$ admits a crepant categorical resolution of singularities generated by 6 exceptional objects.
In the case of cubic threefolds and fourfolds with only one node, categorical resolution of singularities of ${\begin{bf}T\end{bf}}_Z$ are described (see resp. [@bolognesi_bernardara:representability] and [@kuz:4fold]). One should expect that these geometric descriptions carry over to the more degenerate case of determinantal cubics - which are all singular. We haven’t developed the (very long) calculations, but nevertheless we outline expectations about the geometrical nature of these categorical resolutions.
In the 3-dimensional case, the 4 exceptional objects should correspond to a disjoint union of two rational curves, arising as the geometrical resolution of singularities of the discriminant locus of a projection $Z \to {\mathbb{P}}^3$ from one of the six singular points. This discriminant locus is composed by two twisted cubics intersecting in five points, and turns out to be a degeneration of the $(3,2)$ complete intersection curve appearing in the one-node case (see [@bolognesi_bernardara:representability Proposition 4.6]).
In the 4-dimensional case, the 6 exceptional objects should correspond to a disjoint union of two Veronese-embedded planes (isomorphically projected to ${\mathbb{P}}^4$), arising as the geometrical resolution of singularities of the discriminant locus of a projection $Z \to {\mathbb{P}}^4$ from one of the singular points. This discriminant locus is composed by two cubic scrolls intersecting along a quintic elliptic curve, and turns out to be a degeneration of the degree 6 K3 surface appearing in the one-node case (see [@kuz:4fold §5]).
[10]{}
N. Addington, and R. Thomas, *Hodge theory and derived categories of cubic fourfolds*, Duke Math. J. [**163**]{} (2014), no. 10, 1885-1927.
M. Ballard, D. Favero, and L. Katzarkov, *Variation of geometric invariant theory quotients and derived categories*, math.AG/1203.6643. J. Reine Angew. Math., to appear.
A. Auel, M. Bernardara and M. Bolognesi, *Fibrations in complete intersections of quadrics, Clifford algebras, derived categories, and rationality problems*, J. Math. Pures Appl. (9) [**102**]{} (2014), no. 1, 249-291.
A. Beauville, *Determinantal hypersurfaces*, Michigan Math. J. [**48**]{} (2000), 39-64.
M. Bernardara and M. Bolognesi, *Categorical representability and intermediate Jacobians of Fano threefolds*, in *Derived Categories in Algebraic Geometry*, EMS Series of Congress Reports, 1-25 (2012).
M. Bernardara and M. Bolognesi, *Derived categories and rationality of conic bundles*, Compositio Math. [**149**]{} (2013), no. 11, 1789-1817.
J. Blanc and S. Lamy, *Weak Fano threefolds obtained by blowing-up a space curve and construction of Sarkisov links*, Proc. London Math. Soc. (3) [**105**]{} (2012), no. 5, 1047-1075.
A. I. Bondal, *Representations of associative algebras and coherent sheaves*, Izv. Akad. Nauk SSSR Ser. Mat. [**53**]{} (1989), no. 1, 25-44.
A. I. Bondal and M. M. Kapranov, *Representable functors, [S]{}erre functors, and reconstructions*, Izv. Akad. Nauk SSSR Ser. Mat. [**53**]{} (1989), no. 6, 1183-1205, 1337.
A. I. Bondal and D. O. Orlov, *Semiorthogonal decomposition for algebraic varieties*, arXiv math.AG/9506012.
A. I. Bondal and D. O. Orlov, *Derived categories of coherent sheaves*, Proceedings of the International Congress of Mathematicians, Vol. II (Beijing, 2002), 47-56, Higher Ed. Press, Beijing, 2002.
L. Borisov, A. Caldararu, *The Pfaffian-Grassmannian derived equivalence*, J. Algebraic Geom. [**18**]{} (2009), no. 2, 201-222.
T. Bridgeland, *Equivalences of triangulated categories and [F]{}ourier–[M]{}ukai transforms*, Bull. London Math. Soc. [**31**]{} (1999), no.1, 25-34.
T. Bridgeland, *Flops and derived categories*, Invent. Math. [**147**]{} (2002), no. 3, 613-632.
R. Buchweitz, G. J. Leuschke, and M. Van den Bergh, *Non-commutative desingularization of determinantal varieties I*, Invent. Math. [**182**]{} (2010), 47-115.
R. Buchweitz, G. J. Leuschke, and M. Van den Bergh, *Non-commutative desingularization of determinantal varieties II: arbitrary minors*, arXiv math.AG/1106.1833.
M. Herbst, K. Hori, D. Page, *B-type D-branes in toric Calabi-Yau varieties.* Homological mirror symmetry, 27-44, Lecture Notes in Phys., 757, Springer, Berlin, 2009.
K.Hori, D.Tong, *Aspects of non-abelian gauge dynamics in two-dimensional $\mathcal{N}=(2,2)$ theories*, J. High Energy Phys. (2007), no. 5, 079, 41 pp.
D. Huybrechts, *Fourier-Mukai transforms in algebraic geometry*, Oxford Mathematical Monographs. The Clarendon Press, Oxford University Press, Oxford, 2006.
H. Jockers, V. Kumar, J. M. Lapan, D. Morrison and M. Romo, *Nonabelian 2[D]{} gauge theories for determinantal [C]{}alabi-[Y]{}au varieties*, J. High Energy Phys., (2012), no. 11, 166, frontmatter + 46pp.
M. M. Kapranov, *Derived category of coherent sheaves on Grassmann manifolds* Math. USSR, Izv. [**24**]{}, (1985) 183-192; translation from Izv. Akad. Nauk SSSR, Ser. Mat. [**48**]{}, (1984), no.1, 192-202.
A. Kuznetsov, *Derived category of cubic and ${V}_{14}$ threefold*, Proc. V. A. Steklov Inst. Math. **246** (2004), 183-207.
A. Kuznetsov, *Derived categories of cubic fourfolds*, Cohomological and geometric approaches to rationality problems, Progr. Math., vol. 282, Birkhäuser Boston, 2010, 219-243.
A. Kuznetsov, [*Hyperplane sections and derived categories*]{}, Izv. Math. [**70**]{} (2006), no. 3, 447-547; translation from Izv. Ross. Akad. Nauk, Ser. Mat. [**70**]{} (2006), no. 3, 23-128.
A. Kuznetsov, *Homological projective duality for Grassmannians of lines*, arxiv math.AG/0610957.
A. Kuznetsov, *Homological projective duality*, Publ. Math. Inst. Hautes [É]{}tudes Sci. (2007), no. 105, 157-220.
A. Kuznetsov, *Semiorthogonal decompositions in algebraic geometry*, arxiv math.AG/1404.3143.
A. Kuznetsov, *Derived categories of quadric fibrations and intersections of quadrics*, Adv. Math. **218** (2008), no. 5, 1340-1369.
A. Kuznetsov, *Derived categories of Fano threefolds*, Proc. Steklov Inst. Math. [**264**]{} (2009), no. 1, 110-122.
A. Kuznetsov, *Lefschetz decompositions and categorical resolutions of singularities*, Sel. Math., New Ser., [**13**]{} (2007), no. 4 , 661-696.
A. Kuznetsov, *Derived categories view on rationality problems*, preprint arXiv:1509.09115.
R. Lazarsfeld, *Positivity in Algebraic Geometry*, Ergebnisse der Mathematik 48, Springer, 2004.
V. Lunts and D. Orlov, [*Uniqueness of enhancement for triangulated categories*]{}. J. Amer. Math. Soc. [**23**]{} (2010), 853-908.
S. Okawa, *semiorthogonal decomposability of the derived category of a curve*, Adv. Math. [**228**]{} (2011), no. 5, 2869-2873.
D. O. Orlov, *Projective bundles, monoidal transformations and derived categories of coherent sheaves*, Russian Math. Izv. [**41**]{} (1993), 133-141.
E. A. Rødland *The Pfaffian Calabi-Yau, its Mirror, and their link to the Grassmannian ${\mathbb{G}}(2,7)$*, Compositio Math. [**122**]{} (2000), no. 2, 135-149,
M. Van den Bergh, *Non-commutative crepant resolutions*, in *The legacy of Niels Henrik Abel*, 749-770, Springer, Berlin, 2004.
M. Van den Bergh *Three-dimensional flops and noncommutative rings*, Duke Math. J. [**122**]{} (2004), no. 3, 423-455.
J. Weyman, *Cohomology of vector bundles and Syzygies*, Cambridge Tracts in Mathematics (149), Cambridge University Press, 2003.
[^1]: D. F. partially supported by GEOLMI ANR-11-BS03-0011
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We analyze two-particle pseudorapidity correlations in a simple model, where strings of fluctuating length are attached to wounded nucleons. The obtained straightforward formulas allow us to understand the anatomy of the correlations, i.e., to identify the component due to the fluctuation of the number of wounded nucleons and the contribution from the string length fluctuations. Our results reproduce qualitatively and semiquantitatively the basic features of the recent correlation measurements at the LHC.'
address:
- '$^{1}$Institute of Physics, Jan Kochanowski University, PL-25406 Kielce, Poland'
- '$^{2}$The H. Niewodniczański Institute of Nuclear Physics, Polish Academy of Sciences, PL-31342 Cracow, Poland'
- '$^{3}$AGH University of Science and Technology, Faculty of Physics and Applied Computer Science, PL-30059 Cracow, Poland'
author:
- 'Wojciech Broniowski$^{1,2}$ and Piotr Bożek$^3$'
bibliography:
- 'hydr.bib'
title: 'Rapidity fluctuations in the initial state[^1] [^2]'
---
PACS: 5.75.-q, 25.75Gz, 25.75.Ld
In this talk (for details see Ref. [@Broniowski:2015oif]) we analyze the recently measured two-particle pseudorapidity correlations [@ATLAS:2015kla; @ATLAS:anm] in a simple model with wounded nucleons pulling strings of fluctuating length. The charges pulling the strings can be interpreted as the wounded nucleons [@Bialas:1976ed] (cf. Fig. \[fig:tubes\]). Importantly, the longitudinal position of the other end-point of the string is randomly distributed over the available space-time rapidity range. The assumption that the initial entropy distribution is generated by strings whose end-points are randomly distributed is similar to the mechanism of Ref. [@Brodsky:1977de]. It was also used in modeling the particle production in the fragmentation region [@Bialas:2004kt].
The two-particle correlation is defined as $$\begin{aligned}
C(\eta_1, \eta_2) &=& \frac{ {\langle \rho(\eta_1,\eta_2)\rangle} }{ {\langle \rho(\eta_1)\rangle} {\langle \rho(\eta_2)\rangle}}, \label{eq:C}\end{aligned}$$ where $\rho(\eta_1,\eta_2)$ and $\rho(\eta)$ denote the distribution functions for pairs and single particles and ${\langle .\rangle}$ stands for averaging over events. To minimize spurious effects, the ATLAS Collaboration [@ATLAS:2015kla] uses a measure obtained from $C(\eta_1, \eta_2)$ by dividing it with its marginal projections, namely $$\begin{aligned}
C_N(\eta_1, \eta_2) = \frac{C(\eta_1, \eta_2) }{C_p(\eta_1)C_p(\eta_2)}, \label{eq:CN}\end{aligned}$$ where $C_p(\eta_{1,2})=\int_{-Y}^Y d\eta_{2,1} \, C(\eta_1, \eta_2)$ and $Y=2.4$ determines the experimental acceptance range for $\eta_{1,2}$. Moreover, the correlation functions are conventionally normalized to unity, $$\begin{aligned}
\overline{C}(\eta_1, \eta_2) = \frac{C(\eta_1, \eta_2)}{\int_{-Y}^Y \! d\eta_1 \int_{-Y}^Y \! d\eta_2 \,C(\eta_1, \eta_2) }, \label{eq:Cbar}\end{aligned}$$ and analogously for $\overline{C}_N(\eta_1,\eta_2)$.
[![Correlation function $\overline{C}_N(\eta_1,\eta_2)$ for Pb-Pb collisions at 2.76 TeV for $c=30-40\%$. The flat (lighter color) sheet corresponds to the calculation without the length fluctuations. The sheet with an elongated ridge (darker color) corresponds to the case with the string length fluctuations. \[fig:c3040\]](C3040Nx.pdf "fig:"){width="65.00000%"}]{}
The details of the derivation of the formulas for the correlation function listed below are presented in Ref. [@Broniowski:2015oif]. The number of wounded nucleons in the two colliding nuclei is denoted as $N_A$ and $N_B$, and $N_\pm=N_A\pm N_B$. The cases with $r=1$ and $r=0$ correspond to, respectively, present or absent string length fluctuations. We introduce the scaled rapidities $$u_{1,2}=\eta_{1,2}/y_b, \label{eq:u}$$ with $y_b$ denoting the rapidity of the beam. Then in the general case $$\begin{aligned}
&& \hspace{-5mm} C(\eta_1, \eta_2) = 1 + \frac{1}{\left[ {\langle N_+\rangle} + {\langle N_-\rangle}u_1 \right] \left[ {\langle N_+\rangle}+ {\langle N_-\rangle}u_2 \right]} \label{eq:gen2} \times \\
&& \hspace{-2mm} \Big\{ {\langle N_+\rangle} \left [ r (1 - {u_1 u_2} - |u_1-u_2|) +s(\omega)(1+r + (1-r)u_1 u_2 - r |u_1-u_2| ) \right ] \nonumber \\
&& \hspace{-3mm} + {\langle N_-\rangle} s(\omega) (u_1+u_2) + {\rm var}(N_+) + {\rm var}(N_-) u_1 u_2 + {\rm cov}(N_+,N_-)(u_1+u_2) \Big\}, \nonumber\end{aligned}$$ whereas for symmetric collisions ($A=B$) the formula simplifies into $$\begin{aligned}
&& \hspace{-5mm} C(\eta_1, \eta_2) = 1 + \frac{1}{{\langle N_+\rangle}^2} \label{eq:gensym2} \times \\
&& \Big\{ {\langle N_+\rangle} \left [ r( 1 - {u_1 u_2} - |u_1-u_2|) +s(\omega)(1+r + (1-r)u_1 u_2 - r |u_1-u_2| ) \right ] \nonumber \\
&& + {\rm var}(N_+) + {\rm var}(N_-) u_1 u_2 \Big\}. \nonumber\end{aligned}$$ The quantity $s(\omega)$ stands for the square of the standard deviation of the overlaid distribution of strength of the sources.
As originally noticed in Ref. [@Bzdak:2012tp], fluctuation in the number of wounded nucleons alone ($r=0$, $s(\omega)=0$) generates non-trivial longitudinal correlations. Our formula shows, however, that a significant (and long-range in rapidity) part comes from the length fluctuations. The results are depicted in Fig. \[fig:c3040\]. Results of a similar study for the asymmetric case of p-Pb collisions are shown in Fig. \[fig:pPb\].
Analytic expressions may be obtained [@Broniowski:2015oif] for the $a_{nm}$ coefficients of the expansion in a set of orthonormal polynomials [@Bzdak:2012tp; @Jia:2015jga]
[![Correlation functions $\overline{C}(\eta_1,\eta_2)$ (lighter color) and $\overline{C}_N(\eta_1,\eta_2)$ (darker color) for p-Pb collisions at 5.02 TeV. \[fig:pPb\]](CpPb.pdf "fig:"){width="65.00000%"}]{}
The features found in our simple model are manifest in advanced models implementing string decays in the early phase of the high-energy collisions [@Andersson:1983ia; @Wang:1991hta; @Lin:2004en; @Monnai:2015sca]. Our formulas provides an intuitive understanding for these mechanisms.
We cordially wish Janek Pluta all the best on the occasion of his anniversary. Let the successful Cracow-Warsaw collaboration, animated by Janek long ago, continue for many years to come.
[^1]: e-mail: [email protected], [email protected]
[^2]: Research supported by the National Science Centre grants DEC-2012/05/B/ST2/ 02528 and DEC-2012/06/A/ST2/00390.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We propose a general model, Multi-Average Ion Collisional-Radiative Model (MAICRM), to rapid simulate the ionization and population distributions of hot dense plasmas. In MAICRM, the orbital occupation numbers of ions at the same charge stage are averaged and determined by the excitation and de-excitation processes; the populations of the average ions are determined by the ionization and recombination processes with the fixed orbital average occupation numbers in each ion. The calculated mean ionizations and charge state distributions of MAICRM are in general agreement with the other theoretical and experimental results especially for the mid- and high-density plasmas. Since MAICRM considers more detailed transitions and ionization balances than the average atom model and is faster than DCA/SCA models, this model has the advantage to be combined into hydrodynamic simulations.'
author:
- Xiaoying Han
- Lingxiao Li
- Zhensheng Dai
- Wudi Zheng
- Peijun Gu
- Zeqing Wu
nocite: '[@*]'
title: 'MAICRM : A general model for rapid simulation of hot dense plasmas'
---
Introduction
============
The charge state distribution (CSD) of hot dense plasmas is of primary interest of a great deal of experimental and theoretical attention since they are the key values for the hydrodynamic simulations, spectroscopic diagnostics and laser plasma interaction (LPI) simulations in the applications such as inertial confinement fusion (ICF) [@Lindl1995] and z-pinch plasmas[@Whitney2001]. The CSD of plasma is governed by a balance between processes such as ionization, recombination, excitation and radiative decay, as well as two-body processes such as autoionization and dielectronic recombination. For the high density plasma, the plasma approaches local thermodynamic equilibrium (LTE) and the CSD and the occupation numbers of the orbitals satisfy the Saha-Boltzmann equations. In the low-density limit, coronal equilibrium (CE) can be assumed and the CSD is evaluated by the CE equations. For an arbitrary density plasma the CSD and the orbital occupations are determined by the solution of a set of coupled collisional-radiative (CR) rate equations involving excitation/de-excitation and photoionization/recombination atomic processes.
As demonstrated in the 9th NLTE code comparison workshop[@NLTE9-2017], according to the coarseness of the statistical treatment of the atomic levels, all the models can be roughly classified into a few categories including average-atom (AA) codes, detailed configuration accounting (DCA) and/or Super-transition-arrays (STA) codes, detailed level accounting (DLA) codes and hybrid codes. In the detailed atomic models such as DCA[@UTA1988], STA[@STA1989] and DLA[@Ralchenko2001], it is crucial to consider a sufficiently large number of energy levels in each ionization stage; this leads to sometimes unmanageable thus various approximations, e.g., neglecting multi-excited or inner-shell excited states, are adopted. In addition to the completion, the computation time is another key point especially for the time dependent cases. In general the main computation time in the models spends on the rate and absorption/emission coefficient calculations, which is proportional to $(N_{CS}\times N_{C}\times N_{O})$. Here $N_{CS}$, $N_{C}$ and $N_{O}$ are the numbers of the considered charge states, configurations in each charge state and the orbital bases. Thus AAM is of big advantage since $(N_{CS}\times N_{C})^{\textsc{AAM}}=1$ while in other detailed models $(N_{CS}\times N_{C})=(Z+1)\times N_{C}$ and $N_{C}$ may be equal to tens or hundreds if including the single, double and multi-excited states. So the radiative hydrodynamic codes usually choose AAM as the combined inline atomic model despite its coarseness treatment of the atomic levels.With the progress of the experiment and simulation abilities some more detailed models such as SCA[@SCRAM2007] and DCA[@DCA1990] are carried out in the hydrodynamic simulations. Ref.[@JONES2017] shows the differences of the simulated emissivity and capsule bang time by using different inline atomic methods. In their comparisons, many mixed factors such as the numbers and the detail degrees of the involved energy levels are attributed to the differences. The influences of different fixed factors are difficult to be distinguished.
Considering the ionization/recombination processes are usually slower than the excitation/de-excitation processes if the electron density is not too low, the two kinds of processes can be decoupled. Here we propose a general method, i.e., Multi-Average Ion Collisional-Rdiative Model (MAICRM), to simulate the hot dense plasmas by separately solving the occupation numbers of the orbitals and the fraction populations of the ionic charge states. More specifically, in MAICRM, all the configurations at the same ionic stage, namely having the same number of bound electrons, are averaged and represented by an average ion and for a single element plasma, only (Z+1) average ions are taken into account. In this model, the average occupation numbers of each average ion are determined by the excitation and de-excitation processes and the fraction populations of all the average ion are determined by the ionization and recombination processes with the fixed non-integral average occupation numbers. Because of $(N_{CS}\times N_{C})^{\textsc{MAICRM}}=(Z+1)\times 1$ the computation time of MAICRM is shorter than those of DCA/SCA calculations by $N_{C}$ times, which is usually equal to tens or hundreds. On the other hand, in MAICRM all the configurations at one ionic stage are described by an average ion but the detailed ionization balances between the different ionic stage are considered, which is more detailed than AAM in which all the configurations in all ionic stages are averaged and represented by an average atom. Thus because of its shorter time consuming and more detailed treatment of the atomic levels the proposed MAICRM has the advantage to be combined into the hydrodynamic simulations. This paper is organized as following: Section II introduces the theoretical method and Section III shows the comparisons of the mean ionization and CSDs of Fe, Xe, Au plasmas between the results of MAICRM with the other theoretical and experimental results.
Theoretical method
==================
Average ion
-----------
In the present MAICRM, 65 single orbital bases $nlj$ or $nl$ are chosen and listed in table \[tab:orbitalbases\] where $n, l$ and $j$ are the principal quantum number, angular momentum and total angular momentum. For $n\leq5$, the orbital bases are relativistic and for $5<n\leq10$, the orbital bases are non-relativistic. The energy levels of the orbital bases and their transition matrix elements are calculated by RSCF (relativistic self-consistent-field) method[@TXM1995] and tabulated as an input database. For the non-relativistic orbitals the values are averaged from their related relativistic orbitals by their static weights $(2j+1)$. For each charge state, an average ion $\mathbf{\Lambda^{\theta}}$ is labeled as $$\begin{aligned}
{\mathbf{\Lambda^{\theta}}\equiv (n_{1}l_{1}j_{1})^{\mathbf{\Omega}^{\theta}_{1}}(n_{2}l_{2}j_{2})^{\mathbf{\Omega}^{\theta}_{2}} \cdots (n_{i^{\theta}_{max}}l_{i^{\theta}_{max}})^{\mathbf{\Omega}^{\theta}_{i^{\theta}_{max}}},}\end{aligned}$$ where $\theta$ is the number of bound electrons in the average ion and $i^{\theta}_{max}$ is the max number of the single orbital bases for average ion $\mathbf{\Lambda^{\theta}}$. $\mathbf{\Omega}^{\theta}_{i}$ is the electron average occupying number of the $i$th orbital in $\mathbf{\Lambda^{\theta}}$, which is non-integral and satisfies $ 0 \leq \mathbf{\Omega}^{\theta}_{i} \leq \mathbf{g}_{i} $; here $\mathbf{g}_{i}$ is the statistic weight of the $i$th orbital. We define a vector $\mathbf{\Xi^{\theta}}$ to present the average ion $\mathbf{\Lambda^{\theta}}$ $$\begin{aligned}
{\mathbf{\Xi^{\theta}}\equiv \{\mathbf{\Omega}^{\theta}_{1},\mathbf{\Omega}^{\theta}_{2},\ldots,\mathbf{\Omega}^{\theta}_{i^{\theta}_{max}}\}.}\end{aligned}$$
-------------------------------------------------------------------------------------------- --
$1s_{1/2}$
$2s_{1/2}, 2p_{1/2}, 2p_{3/2}$
$3s_{1/2}, 3p_{1/2}, 3p_{3/2}, 3d_{3/2}, 3d_{5/2}$
$4s_{1/2}, 4p_{1/2}, 4p_{3/2}, 4d_{3/2}, 4d_{5/2}, 4f_{5/2}, 4f_{7/2}$
$5s_{1/2}, 5p_{1/2}, 5p_{3/2}, 5d_{3/2}, 5d_{5/2}, 5f_{5/2}, 5f_{7/2}, 5g_{7/2}, 5g_{9/2}$
$6s, 6p, 6d, 6f, 6g, 6h$
$7s, 7p, 7d, 7f, 7g, 7h, 7i$
$8s, 8p, 8d, 8f, 8g, 8h, 8i, 8j$
$9s, 9p, 9d, 9f, 9g, 9h, 9i, 9j, 9k$
$10s, 10p, 10d, 10f, 10g, 10h, 10i, 10j, 10k$
-------------------------------------------------------------------------------------------- --
: \[tab:orbitalbases\]The single orbital bases
Rate coefficients
-----------------
In our model, the rate coefficients of ten kinds of atomic processes, including photon excitation, electron collisional excitation, photon ionization, electron collisional ionization, autoionization and their reverse processes, use the same calculation formulas in Ref.[@Florido2009] with the integer occupation numbers of orbitals in each configuration replaced by the average occupation number $\mathbf{\Omega}^{\theta}_{i}$ in each average ion $\mathbf{\Lambda^{\theta}}$.
Plasma effect
-------------
In the dense plasma, screening effects due to neighboring electrons and ions modify the energy levels. The effect on the ionization potentials of bound states and level populations leads to the phenomenon of pressure ionization. In despite of the importance of pressure ionization to calculate the ionic abundances and level populations, most CR models take into account plasma effects in an approximate way via an effective lowering of the ionization potential or continuum lowering (CL). In our model the ionization potential $\mathbf{I}_{\theta}$ is lowered a quantity $\Delta \mathbf{I}_{\theta}$ to be $\mathbf{I}'_{\theta}=\mathbf{I}_{\theta}-\Delta \mathbf{I}_{\theta}$. Here we apply the formulation proposed by Stewart and Pyatt[@Stewart1966] and use the following formula
$$\begin{aligned}
{\Delta \mathbf{I}_{\theta}=\frac{3}{2}\frac{I_{H}a_{0}}{\mathbf{r}^{\theta}}(\langle Z\rangle-\theta+1)\{[1+(\frac{\mathbf{D}}{\mathbf{r}^{\theta}})^{3}]^{\frac{2}{3}}-(\frac{\mathbf{D}}{\mathbf{r}^{\theta}})^{2}\},}\label{eq:eq18}\end{aligned}$$
where $\mathbf{r^{\theta}}=[3(\langle Z\rangle-\theta+1)/(4\pi n_{e})]^{1/3}$ is the ion-sphere radius assuming the plasma composed of ions with $\theta$ electrons only, $\mathbf{D}=[4\pi (\langle Z\rangle+\langle Z^{2}\rangle)n_{ion}/T_{e}]^{-1/2}$ is the Debye length, $\langle Z\rangle$ is the average ionization of plasma, $\langle Z^{2}\rangle$ is the second-order moment of the population distribution. Once the lowered ionization potential $\mathbf{I}'_{\theta}$ are calculated for each charge state, the $i^{\theta}_{max}$ for each charge state are determined with the orbital energy $\mathbf{E^{\theta}_{\emph{i}}}<\mathbf{I}'_{\theta}$. Then a new set of $\mathbf{I}'_{\theta}$ is obtained based on the updated $i^{\theta}_{max}$. Thus $i^{\theta}_{max}$ and $\mathbf{I}'_{\theta}$ are calculated iteratively. The CL correction may result in a significant reduction of the $i^{\theta}_{max}$ for some charge states if the density of the plasma is high.
Average occupation number and charge state distribution
-------------------------------------------------------
In our model, the average occupying numbers of the orbitals and the populations of ionic charge states are calculated by two steps iteratively. Firstly, we calculate the average orbital occupation numbers $\mathbf{\Omega}^{\theta}_{i} (i=1,\ldots,i^{\theta}_{max})$ of every average ion $\mathbf{\Lambda}^{\theta}$ by solving a set of $i^{\theta}_{max}$ rate equations which are only relative to the excitation and de-excitation atomic processes in the ion. The coupled $i^{\theta}_{max}$ rate equations of $\mathbf{\Omega}^{\theta}_{i}$ for each average ion are
$$\begin{aligned}
{\frac{d\mathbf{\Omega}^{\theta}_{i}}{d\mathbf{t}}=-\mathbf{\Omega}^{\theta}_{i}\sum_{j=1}^{i^{\theta}_{max}}(\mathbf{g}_{j}-\mathbf{\Omega}^{\theta}_{j})\mathbf{R^{E/D}_{\theta,\emph{i}\rightarrow \emph{j}}}+(\mathbf{g}_{i}-\mathbf{\Omega}^{\theta}_{i})\sum_{j=1}^{i^{\theta}_{max}}\mathbf{\Omega}^{\theta}_{j}\mathbf{R^{E/D}_{\theta,\emph{j}\rightarrow \emph{i}}},}\label{eq:rate1-1}\end{aligned}$$
where $\mathbf{R^{E/D}_{\theta,\emph{i}\rightarrow \emph{j}}}$ are the excitation or de-excitation rate coefficients for the transition of $i$th orbital to $j$th orbital in the average ion $\mathbf{\Lambda}^{\theta}$. $\mathbf{\Omega}^{\theta}_{i}$ should fulfill the condition of charge conservation $$\begin{aligned}
{\sum_{i=1}^{i^{\theta}_{max}}\mathbf{\Omega}^{\theta}_{i}=\theta~~(\theta = 0,\ldots,Z).}\label{eq:rate1-2}\end{aligned}$$ Combining Eqs. (\[eq:rate1-1\]) and (\[eq:rate1-2\]), a set of converged solutions $\{\Xi^{\theta},(\theta = 0,\ldots,Z)\}$ for all considered average ions are obtained.
Secondly, based on the set of fixed vectors $\{\Xi^{\theta},(\theta = 0,\ldots,Z)\}$, the coupled rate equations of the $(Z+1)$ average ions can be built as following
$$\begin{aligned}
{\frac{d \mathbf{P_{\theta}}}{dt}=-\mathbf{P_{\theta}}\sum_{\tilde{\theta}=\theta\pm 1}\mathbf{R^{I/R}_{\theta\rightarrow \tilde{\theta}}}+\sum_{\tilde{\theta}=\theta\pm 1}\mathbf{P}_{\tilde{\theta}}\mathbf{R^{I/R}_{\tilde{\theta}\rightarrow \theta}} ~~~(\tilde{\theta}, \theta = 0,\ldots,Z) ,}\label{eq:rate2-1}\end{aligned}$$
where $\mathbf{P}_{\theta}$ is the population of average ion $\mathbf{\Lambda^{\theta}}$. $\mathbf{R^{I/R}_{\theta\rightarrow \tilde{\theta}}}$ are the ionization or recombination rate coefficients of the average ion $\mathbf{\Lambda^{\theta}}$ to the average ion $\mathbf{\Lambda^{\tilde{\theta}}}$. $\mathbf{P}_{\theta}$ should fulfill the normalization condition $$\begin{aligned}
{\sum_{\theta=0}^{Z}\mathbf{P}_{\theta}=1.}\label{eq:rate2-2}\end{aligned}$$ Combining Eqs. (\[eq:rate2-1\]) and (\[eq:rate2-2\]), a set of $\{ \mathbf{P}_{\theta} (\theta = 0, \ldots, Z) \}$ are obtained. Then a new electron density value $n'_{e}=n_{i}\sum_{\theta=0}^{\mathbf{Z}}\mathbf{P}_{\theta}\cdot(Z-\theta)$ is calculated. With the new $n'_{e}$ the rate coefficients are updated and putted into Eq. (\[eq:rate1-1\]), then a new set of $\{\Xi'^{\theta},\theta = 0, \ldots, Z\}$ are obtained. With the updated $\{\Xi'^{\theta}\}$ a new set of $\{\mathbf{P}'_{\theta}\}$ are calculated by Eqs. (\[eq:rate2-1\]) and (\[eq:rate2-2\]). The iterative procedure of updating $\{\Xi^{\theta}, {\mathbf{P}_{\theta}} (\theta = 0, \ldots, Z) \}$ is stopped when a converged $n_{e}$ and mean ionization $\langle Z\rangle$ reach.
Results and discussions
=======================
Using MAICRM the mean ion charge $\langle Z\rangle$ and CSD of mid- and high-Z plasmas in LTE and NLTE conditions are calculated and compared with other theoretical and experimental results.
Fig.\[fig:FeLTE\] shows the results of Fe plasma in LTE condition. In panel (a), for the plasma of the condition $\rho$=0.0081 g/cm$^{3}$, $T$=25eV, our calculated $\langle Z\rangle$ and CSD agree with the results of AAM[@Faussurier1997] and Multi-AIM[@Kiyokawa2014]. The other two models assume a non-integer orbital occupation determined by the Fermi distribution. The CSD of AAM is obtained from minor manipulations of the grand canonical partition function[@Faussurier1997]. The CSD of Multi-AIM is obtained self-consistently by means of minimizing the free energy of the whole system established by the finite temperature density functional theory[@Kiyokawa2014].
The panels (b), (c) and (d) of Fig.\[fig:FeLTE\] show the CSDs of Fe plasma in condition of $\rho$=0.008 g/cm$^{3}$ and $T$=25, 50, 100 eV respectively. In general, our calculated CSDs agree with those of AAM. At 25eV the small difference between the most populated charge states Fe$^{7+}\& $ Fe$^{8+}$ should result from the different atomic data such at the ionization energies and oscillator strengths. At 100eV the bigger difference of the fractions of Fe$^{16+}\&$ Fe$^{17+}$ may be from the critical situation that the L shell is opening, which is sensitive to the ionization and recombination rate coefficients of the near full occupation L shell. Our high fraction of Fe$^{17+}$ means the L shell is opened earlier than AAM.
Fig.\[fig:FeNLTE\] shows the mean ion charge $\langle Z\rangle$ of Fe plasma in NLTE condition of $n_{e}=10^{14}, 10^{19}, 10^{22}, 10^{24}$ cm$^{-3}$ respectively. Our results are compared with the data from the various NLTE codes submitted in the 9th NLTE code comparison workshop[@NLTE9-2017], which include a few categories, including AAM codes, configuration and/or Superconfiguration codes, detailed level accounting codes, and hybrid codes. Panel (a) shows at the very low density $n_{e}=10^{14}$ cm$^{-3}$ our calculated mean ionization is generally lower than the others, while at higher electron densities $n_{e}=10^{19}, 10^{22}, 10^{24}$ cm$^{-3}$ our calculated $\langle Z\rangle$ agree with the other results as shown in panels (b),(c) and (d). In MAICRM, the mean occupation numbers are mainly determined by the competition between the excitation and de-excitation channels, which assumes the change of $\mathbf{\Omega}^{\theta}_{i}$ due to the ionization and recombination is a small perturbation and negligible. It is known that without the radiation field, the excitation channels are dominated by the collisional excitation (CE) and the de-excitation channels are dominated by spontaneous emission (SE) as the electron density is low. At very low electron density, the CE rates are much smaller than the SE rates, so the electrons are accumulated in the lowest orbitals and the excited orbitals are nearly empty. Because of the near empty occupation of the excited orbitals, the recombination processes, especially the resonant EC process, are prominent and their contribution to $\mathbf{\Omega}^{\theta}_{i}$ for the excited orbitals is not negligible. If considering the contribution of the recombination, $\mathbf{\Omega}^{\theta}_{i}$ of the excited orbitals should increase a small but non-negligible value comparing with their tiny absolute values, which will result in an obvious increase of AI rates. So the mean ionization of MAICRM at low density is lower than other models. How to treat the very low density situation reasonably deserves further study and will be reported elsewhere. At the mid- and high-densities, the collisional excitation/de-excitation channels become important and much excited states are produced so $\mathbf{\Omega}^{\theta}_{i}$ is a good estimation for the most populated configurations in one charge state, which is illustrated by the good agreement between the MAICRM calculations and the other code results shown in panels (b),(c) and (d).
 The comparison of the charge state distributions (CSDs) of Fe Plasma in LTE condition between MAICMR codes and the other calculations[@Faussurier1997; @Kiyokawa2014]. ](FeLTE)
 The comparison of the mean ionization of Fe Plasma as a function of free-electron temperature $T_{e}$ in NLTE condition between MAICRM results (blue dot) and the results of 14 codes and models submitted in the 3th NLTE code comparison workshop[@NLTE3-2006]. ](FeNLTE)
Fig.\[fig:Xe\] shows the mean ionization of Xe NLTE plasma at $n_{i}=4.75\times10^{18}$ cm$^{-3}$ as a function of free-electron temperature $T_{e}$ and the CSD at $T_{e}$=415eV. Panel (a) shows the MAICRM mean ionizations are in the spread range of 14 codes submitted in the 3th NLTE code comparison workshop[@NLTE3-2006] from superconfigurations to detailed level accounting. Panel (b) shows the CSD of Xe plasmas at $T_{e}=415eV$ and $n_{i}=4.75\times10^{18}$ cm$^{-3}$. $\langle Z\rangle =27.3$ from the MAICRM calculation agrees better with the experimental $\langle Z\rangle =27.4\pm 1.5$[@Popovics2002] than the ion-by-ion spin-orbit-split (SOSA) approach[@SOSA1991] and the AVERROES calculations based on the superconfiguration and STA concepts[@Peyrusse2000]. The CSD of MAICRM is in the spread range of the results gathered in the 3th NLTE code comparison workshop[@NLTE3-2006].
 (a) The comparison of the mean ionization of Xe plasma as a function of free-electron temperature $T_{e}$ in NLTE condition between MAICRM results (blue dot) and the results of 14 codes and models submitted in the 3th NLTE code comparison workshop [@NLTE3-2006] from superconfigurations to detailed level accounting. (b) Charge state distribution of Xe plasmas at $T_{e}=415eV$ and $n_{i}=4.75\times10^{18}$ cm$^{-3}$. The experimental value is $\langle Z\rangle =27.4\pm 1.5$[@Popovics2002] and the data in parenthesis are the experimental uncertainties; (red dot): the ion-by-ion spin-orbit-split (SOSA) approach[@SOSA1991]; (green dot): the AVERROES calculations based on the superconfiguration and STA concepts[@Peyrusse2000]. The other data (gray) are the results of 14 codes and models submitted in the 3th NLTE code comparison workshop [@NLTE3-2006].](Xe)
In Fig.\[fig:Au1\] the panels (a)-(f) show the mean ionizations of Au plasma at $n_{e}=10^{19}, 10^{20}, 10^{21}, 10^{22}, 10^{23}, 10^{24}$ cm$^{-3}$ as a function of free-electron temperature $T_{e}$. All the mean ionization curves of MAICRM are in the spread range of the submission data from 14 different codes in the 3th NLTE code comparison workshop[@NLTE3-2006]. (g) and (h) panels show the CSDs of Au plasma at $T_{e}$=2.2keV $n_{e}=6\times 10^{20}$ cm$^{-3}$and $T_{e}$=2.5keV $n_{e}=10^{19}$ cm$^{-3}$ respectively. As shown in panel (g) the CSD of MAICRM agree well with experimental data and the theoretical values of RIGEL including the two-electron AI/EC processes[@Foord2000], where RIGEL is a superconfiguration-based collisional-radiative code constructed using hydrogenic supershells. Panel (h) shows the mean ionization of MAICRM is lower than the experimental value while in the spread rang of the theoretical data in in the 3th NLTE workshop.
![(Color online) The comparisons of the mean ion charge $\langle Z\rangle$ and CSDs of Au plasma between MAICRM and the data from the 9th NLTE code comparison workshop[@NLTE3-2006] and Ref.[@Foord2000]. Note that in panels (g) and (h) the data in parenthesis of the experimental $\langle Z\rangle$ are the experimental uncertainties. ](Au1 "fig:")\[fig:Au1\]
Fig.\[fig:Au2\] shows the comparisons of the MAICRM mean ionization and CSDs of Au plasma with the recent experimental values[@HeeterPRL2007]. In panels (a)-(c) the MAICRM CSDs generally agree with the experimental values. In panel (d) the difference between the MAICRM results and the experimental data is due to a population fraction of 0.15 for the ions with Z$\leqslant$ 40 not modeled in the experimental fitting procedure. Panel (e) shows the comparisons of the mean ionizations at 10 different conditions between the MAICRM calculations and the experimental values[@HeeterPRL2007]. As a whole the MAICRM results agree with experimental values within the experimental uncertainties, which illustrates the credibility of MAICRM for high-Z hot dense plasma.
![(Color online) The comparisons of the mean ion charge $\langle Z\rangle$ and CSDs of Au plasma between MAICRM and the experimental results[@HeeterPRL2007]. Note that in panels (a), (b), (c) and (d) the data in parenthesis of the experimental $\langle Z\rangle$ are the experimental uncertainties and in panel (e) the units of electron temperature $T_{e}$, radiation temperature $T_{r}$ and electron density $n_{e}$ are keV, eV and $10^{20}$ cm$^{-3}$ respective.](Au2 "fig:")\[fig:Au2\]
Conclusion
==========
A general model MAICRM for the simulation of hot dense plasmas is proposed. In this model, all the energy levels at the same charge stage are averaged and represented by an average ion $\mathbf{\Lambda}^{\theta}$ and (Z+1) average ions are used to describe the plasma. The average occupation numbers $\mathbf{\Omega}^{\theta}_{i}$ are determined by the competition between the excitation and de-excitation processes in the average ion $\mathbf{\Lambda^{\theta}}$ and the populations $\mathbf{P}_{\theta}$ for each average ion are determined by the ionization and recombination processes between the average ions. After iteratively solving the two sets of rate equations a set of converged $\{ \mathbf{\Omega}^{\theta}_{i},\mathbf{P}_{\theta},i\in(1, i^{\theta}_{max}),\theta\in(0,Z)\}$ are obtained. The calculation results of MAICRM at mid- and high-density conditions agree with the other more detailed theoretical results and the experimental results. For the very low density condition the mean ionization of MAICRM is lower than the other theoretical results since the recombination rate coefficients are larger than the excitation and de-excitation rates, which situation deserves further study. Considering the calculation time, since the $(N_{CS}\times N_{C})^{\textsc{MAICRM}} = (Z+1)$, and $(N_{CS}\times N_{C})^{\textsc{AAM}}=1$, MAICRM is slower than AAM, while MAICRM is faster than the other detailed models such as DCA or SCA because of $(N_{CS}\times N_{C/SC})^{\textsc{DCA/SCA}}=(Z+1)\times N_{C/SC}$. On the other hand MAICRM can treat more detailed transitions and ionization balances than AAM. Therefore, MAICRM has the advantage to be combined into 1D, 2D and 3D radiative hydrodynamic calculations of ICF applications with more detailed treatments than AAM and shorter time-consuming than DCA/SCA models.
This work is partly supported by the National Key R&D Program of China Under Grant No. 2017YFA0402300.
[9]{} J. D. Lindl, Phys. Plasmas **2**, 3933 (1995). K. G. Whitney *et al.*, Rev. Sci. Instrum. **8**, 3708 (2001). R. Piron, F. Gilleron, Y. Aglitskiy, H.-K. Chung, C. J. Fontes, S. B. Hansen, O. Marchuk, H. A. Scott, E. Stambulchik, Yu. Ralchenko, High Energy Density Phys. **23**, 38 (2017). C. Bauche-Arnoult, J. Bauche, and M. Klapisch, Adv. At. Mol. Phys. **23**, 131 (1988). A. Bar-Shalom, J. Oreg, W. H. Goldstein, D. Shvarts, and A. Zigler, Phys. Rev. A **40**, 3183 (1989). Yu. Ralchenko, J. Quant. Spectrosc. Radiat. Transf. **71**, 609-621 (2001). S. S. B. Hansen, J. Bauche, C. Bauche-Arnoult, and M. F. Gu, High Energy Density Phys. **3**, 109 (2007). B. Wilson, J. Albritton, and D. Liberman, “Detailed configuration account opacity calculations using a Monte Carlo technique”, LLNL Report No. UCRL-JC-014252 (1990). O. S. Jones, L. J. Suter, H. A. Scott, M. A. Barrios, W. A. Farmer, S. B. Hansen, D. A. Liedahl, C. W. Mauche, A. S. Moore, M. D. Rosen, J. D. Salmonson, D. J. Strozzi, C. A. Thomas, and D. P. Turnbull, Physics of Plasmas **24**, 056312 (2017). X. M. Tong, Y. Zou and J. M. Li, Chin. Phys. Lett. **12**, 351 (1995). F. Florido *et al.*, Phys. Rev. E **80**, 056402 (2009). J. C. Stewart and K. D. Pyatt, Astrophys. J. **144**, 1203 (1966). G. Faussurier, C. Blancard, and B. A. Decoster, Phys. Rev. E. **56**, 3474 (1997). Shuji Kiyokawa, High Energy Density Phys. **13**, 40 (2014). C. Chenais-Popovics, J. C. Gauthier, J. C. Gary, O. Peyrusse, M. Rabec-Le Gloahec *et al.*, Phys. Rev. E **65**, 046418 (2002). J. Bauche, C. Bauche-Arnoult, and M. Klapisch, J. Phys. B **24**, 1 (1991). O. Peyrusse, J. Phys. B **33**, 4303 (2000). C. Bowen, R. W. Lee, and Yu. Ralchenko, JQSRT **99**, 102-119 (2006). M. E. Foord, S. H. Glenzer, R. S. Thoe, K. L Wong, K. B. Fournier, B. G. Wilson, and P. T. Springer, Phys. Rev. Lett. **85**, 992 (2000). R. F. Heeter, S. B. Hansen, K. B. Fournier, M. E. Foord, D. H. Froula, A. J. Mackinnon, M. J. May, M. B. Schneider, and B. K. F. Young, Phys. Rev. Lett. **99**, 195001 (2007).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The stability of a system of $N$ equal sized mutually gravitating spheres resting on each other in a straight line and rotating in inertial space is considered. This is a generalization of the “Euler Resting” configurations previously analyzed in the finite density 3 and 4 body problems. Specific questions for the general case are how rapidly the system must spin for the configuration to stabilize, how rapidly it can spin before the components separate from each other, and how these results change as a function of $N$. This paper shows that the Euler Resting configuration can only be stable for up to 5 bodies, and that for 6 or more bodies the configuration can never be stable. This places an ideal limit of 5:1 on the aspect ratio of a rubble pile body’s shape.'
author:
- 'D.J. Scheeres'
bibliography:
- '../../bibliographies/biblio\_article.bib'
- '../../bibliographies/biblio\_books.bib'
- '../../bibliographies/biblio\_misc.bib'
date: 'Received: date / Accepted: date'
title: 'Stability of the Euler Resting N-Body Relative Equlilbria'
---
[example.eps]{} gsave newpath 20 20 moveto 20 220 lineto 220 220 lineto 220 20 lineto closepath 2 setlinewidth gsave .4 setgray fill grestore stroke grestore
Introduction {#intro}
============
When gravitating bodies are modeled as rigid bodies with finite density, their gravitational dynamics becomes more rich and can include many additional stable and unstable relative equilibria that do not exist in the classical point-mass $N$ body problem. The study of the gravitational interaction between two rigid bodies is quite classical, tracing back to Lagrange [@lagrange1780theorie], and more recently reconsidered by a wide range of authors from the mathematical, scientific and engineering community. A partial sampling includes Duboshin [@duboshin1958differential], Wang et al. [@wang], Maciejewski [@maciejewski], Beck and Hall [@beck1998relative], Scheeres [@scheeres_F2BP] and Moeckel [@moeckel2017minimal], to give an incomplete list. More recently the problem has been reconsidered, motivated in part by the discovery that small asteroids are often best modeled as a collection of self-gravitating, rigid components resting on each other and rotating [@fujiwara_science]. This generalizes the study of interactions between $N$ rigid bodies to include cases where the bodies can be in physical contact, and have been studied in a series of papers by Scheeres [@scheeres_reconfiguration; @scheeres_minE; @F3BP_scheeres; @F4BP_chapter]. Key results from these studies include the identification of minimum energy configurations at any level of angular momentum, usually with components resting on each other, and conditions under which these systems can undergo dramatic reconfigurations including fission of components. A recent theoretical advancement was a proof by Moeckel that an orbiting system of rigid bodies cannot be energetically stable for $N \ge 3$ [@moeckel2017minimal], implying that minimum energy configurations for $N\ge3$ rigid bodies must consist of at least some of the bodies being at rest on each other.
This paper performs an in-depth study of one class of these relative equilibria, the case where the bodies are modeled as equal-sized spheres which rest on each other in a straight-line, reminiscent of the orbital version of the $N$-body Euler solutions from classical celestial mechanics [@moulton1920periodic]. Here the questions are different, and are specifically focused on what the minimum spin rate of this configuration is in order to stabilize the system, and what the maximum spin rate of the configuration is when gravitational forces can no longer hold it together and it fissions. For the case of equal masses this problem has been solved for $N=3$ and $N=4$ [@scheeres_minE; @F4BP_chapter], and for the case of unequal sized bodies it has been solved for $N=3$ [@F3BP_scheeres], however it has not been explored at higher numbers of components. A specific question of interest is whether there is a maximum number of bodies beyond which this configuration can never be stable. That question is answered in the affirmative in this paper, with the answer being 5.
While the study of this system is somewhat abstract, the results are instructive in understanding the limits of a natural system to arrange itself in a stable configuration. In particular, it shows that a rubble pile body can never achieve an aspect ratio beyond 5:1, at least when it consists of equal-sized, spherical components. This places an important limit on this otherwise general problem, and motivates variations of it to be studied that involve non-equal sized bodies or non-spherical bodies.
This paper is organized as follows. First, the general problem statement is given along with the necessary theory for its study. Then the Euler Resting configuration is defined along with its degrees of freedom. Following this the conditions for this system to be in equilibrium and exist are established. Finally, the stability of the configuration is explored and compared with the existence conditions to find the limiting constraint on the number of bodies. Following this, the stability of all equal mass Euler Resting configurations are reviewed.
Problem Statement {#sec:1}
=================
Consider a system of $N$ spherical bodies of equal diameter $D$ and common density $\rho$. The mass of each body is then ${\cal M} = (\pi/6) \rho D^3$ and their moment of inertia $I = {\cal M} D^2/10$. We will only consider planar configurations of this system, and thus relative to an absolute frame each body has 3 degrees of freedom, two that define its location in the plane and one that rotates the body about an axis perpendicular to the plane. Thus, the degrees of freedom of the system relative to an absolute frame is $3N$.
The Fundamental Functions
-------------------------
The potential energy, total (planar) moment of inertia about the system center of mass and angular momentum of the system is $$\begin{aligned}
{\cal U} & = & - {\cal G} {\cal M}^2 \sum_{i<j} \frac{1}{d_{ij}} \\
I_P & = & N \frac{1}{10} {\cal M} D^2 + \frac{m}{N} \sum_{i<j} d_{ij}^2 \\
{\mbox{\boldmath{$H$}}} & = & \sum_{i=1}^N I \omega_i + \frac{{\cal M}}{N} \sum_{i<j} {\mbox{\boldmath{$d$}}}_{ij}\times\dot{{\mbox{\boldmath{$d$}}}}_{ij}\end{aligned}$$ where ${\mbox{\boldmath{$d$}}}_{ij}$ denotes the relative position vector between two bodies with constraint on the magnitude $d_{ij} \ge D$, and $\omega_i$ is the spin rate of each body. Due to the symmetry of spherical bodies the gravitational potential does not depend on their relative rotational orientation.
Given these quantities, we can define the amended potential of the (planar) system as $$\begin{aligned}
{\cal E} & = & \frac{H^2}{2 I_P} + {\cal U}\end{aligned}$$ where $H$ is the magnitude of the total angular momentum of the system, which is treated as a constant parameter. Note that the amended potential is only a function of the relative positions of the bodies, and thus is independent of an absolute frame.
The existence of relative equilibria and their energetic stability for this system can be completely studied through analysis of the amended potential, and is outlined in general in [@smaleI; @smaleII], and more specifically in [@scheeres_minE; @F3BP_scheeres] for the finite density case where bodies can rest on each other. A key aspect of this problem is that minimum energy configurations exist at any level of angular momentum, but may consist of bodies resting on each other. Our current focus is finding the level of angular momentum for when these resting configurations exist and are stable.
The conditions for a system to be in equilibrium can be established through study of the amended potential [@scheeres_minE]. For the constrained degrees of freedom in this system, denoted as those degrees of freedom at a specific limit $d = d^*$ and only allowed one sided deviations $\delta d \ge 0$, the condition for that degree of freedom to be in equilibrium and to be stable is that $\left.\partial{\cal E}/\partial d\right|_* \ge 0$. For unconstrained degrees of freedom, denoted as $\theta_i$, the condition for a given configuration of the system to be in equilibrium is that $\left.\partial{\cal E}/\partial\theta_i\right|_* = 0$ for all of the unconstrained degrees of freedom evaluated at specific values $\theta_j^*$. Stability in this case is achieved when the Hessian of the amended potential with respect to the unconstrained degrees of freedom is positive definite when evaluated at the equilibrium configuration, $[\partial^2{\cal E}/\partial\theta_i\partial\theta_j]_* > 0$.
The Euler Resting Configuration
-------------------------------
Now define the nominal configuration for this system, shown in Fig. \[fig:definition\]. It is instructive to first review the degrees of freedom of the system under various constraints. The full system has $3N$ degrees of freedom, planar position and rotation for each body. Moving to a relative frame removes the absolute location of the system and the absolute rotation of the system, a total of $3$ degrees of freedom, leaving $3(N-1)$ degrees of freedom. A convenient way to imagine this is to take one body as a reference with a frame embedded in it. Then each of the other $N-1$ body’s configuration relative to this one body are specified by their relative location and relative orientation to the body 1 frame, giving the $3(N-1)$ additional degrees of freedom.
![Geometry of the Euler Resting configuration and the variations in its degrees of freedom. []{data-label="fig:definition"}](definition){width="1.\textwidth"}
Next, assume that every body is in contact with another body. This is a non-trivial step, as for $N$ bodies there are many different separate topologies for how these components can be attached to each other. For a useful example in the $4$ body problem see [@F4BP_chapter]. The current paper only looks at a very specific topology, which has each body in contact with two other bodies, except for the two bodies at the “ends” of the chain. Viewing this as a sequential process, starting with body 1, the first constraint is that body 2 rests on it, and then body 3 rests on body 2, etc., yielding $N-1$ constraints on the relative distances between our bodies. This reduces the degrees of freedom by $N-1$, leaving $2(N-1)$ degrees of freedom. This can again be imagined, starting with the body 1 frame. Body 2 then has two degrees of freedom of orientation to body 1, its angle of contact relative to body 1, and the rotation of body 2 relative to body 1. Then, body 3 has the same two degrees of freedom relative to body 2, etc. It is important to note that not all relative degrees of freedom are possible for the $i$th body attached to body $(i-1)$, in particular the angle between body $i-2$, $i-1$ and $i$ can never be less than $60^\circ$. However with this nominal topology one can always find an open set of relative orientations such that there are no local constraints on the angular placement of the bodies.
The final reduction is the no-slip assumption between bodies in contact. This ties the angular location of body $i$ relative to body $i-1$ to the orientation of body $i$ to body $i-1$. Again, an initial “topology” must be assumed, which gives the initial orientation of one body relative to another, however once this is chosen then there is only one degree of freedom orienting each sphere relative to the previous one. This adds an additional $N-1$ constraints, yielding the final system with $N-1$ degrees of freedom, when in contact.
For the nominal configuration, the bodies all rest on a neighbor and their centers of mass all lie along a line. Thus the distance between any two bodies equals the sum of their radii plus the diameters between them. Assume that the bodies are numbered in order, increasing to the right, such that body 1 only has contact on the right by body 2, but that body 2 has contact on both sides (1 and 3), etc., up to body $N$ which only has a contact on the left ($N-1$) and no contact on the right. Then given two bodies $i$ and $j$ chosen such that $i < j$, the distance between them equals $$\begin{aligned}
d_{ij} = (j-i) D \end{aligned}$$ Thus, when in this configuration the potential energy and moment of inertia are $$\begin{aligned}
{\cal U}^* & = & - \frac{{\cal G} {\cal M}^2}{D} \sum_{i<j} \frac{1}{j-i} \\
I_P^* & = & \frac{1}{10} N {\cal M} D^2 + \frac{{\cal M} D^2}{N} \sum_{i<j} (j-i)^2\end{aligned}$$ where the $()^*$ signifies that the system is in the nominal Euler Resting configuration.
Variations About the Euler Resting Configuration
------------------------------------------------
There are two different types of variations from the nominal configuration that must be considered. These are shown in Fig. \[fig:definition\] and consist of an introduced separation between each body and the rolling of each body on its neighbor.
The first variation assumes the straight line configuration and introduces positive variations in the distance between any two grains otherwise in contact. As there are $N-1$ contacts between grains, this will consist of $N-1$ possible variations in the state of the system. Due to symmetry this can be reduced in half for an even number of bodies, and by (N-1)/2 for an odd number, as the separation of a set of one bodies to the left is equivalent to the separation of an equal number of bodies to the right. In terms of notation, denote the separation between bodies $i$ and $i+1$ to be $d_{i,i+1} = D + d_i$, $i = 1, N-1$ and $d_i \ge 0$.
The second variation will maintain contact between the bodies but allow them to roll relative to each other, assuming the no-slip condition. While this apparently requires $N-1$ degrees of freedom to fully specify, the symmetry of the problem allows one of these rotational degrees of freedom to be expressed as a combination of all of the others, reducing the number to $N-2$. This can be understood simply with a few examples. First, for $N=2$ the formula states that there are 2 degrees of freedom, however within the amended potential the rolling motion of the two bodies does not appear given the symmetry of the masses. Thus there is only a single degree of freedom, which is the separation between the two spheres. For $N=3$ there are 4 degrees of freedom, however under symmetry there are only 3 degrees of freedom that are apparent, the separation between bodies 1 and 2 ($d_{12}$), 2 and 3 ($d_{23}$), and the angle between body 1 and 3. Then for every additional body just add 2 more degrees of freedom, the distance between $N-1$ and $N$ ($d_{N-1,N}$) and the angle measured by the deflection of body $N$ relative to the original placement of bodies $N-2$ and $N-1$. In general, denote the angle that body $i+1$ makes by rolling on body $i$ as $\theta_i$, $i = 1, N-1$ (see Fig. \[fig:definition\] for the specific geometry). When the bodies are in contact, the above definitions can develop the general distance between any two bodies, $j$ and $i$, assuming $i<j$, by temporarily introducing a vector system such that the $x$-axis is along the nominal straight line configuration defined by the centers of mass of the bodies, and the $y$-axis is perpendicular to this. Then the coordinates of a body $j$ relative to body 1 in this frame (assuming contact between all neighboring bodies) is $$\begin{aligned}
{\mbox{\boldmath{$d$}}}_{1j} & = & D \left[ \begin{array}{c}
\sum_{k=1}^{j-1} \cos\theta_{k} \hat{{\mbox{\boldmath{$x$}}}} \\
\\
\sum_{k=1}^{j-1} \sin\theta_k \hat{{\mbox{\boldmath{$y$}}}}
\end{array} \right] \end{aligned}$$ and the relative position between two bodies $j$ and $i$ with $i<j$ is found by subtracting the previous formula, ${\mbox{\boldmath{$d$}}}_{ij} = {\mbox{\boldmath{$d$}}}_{1j} - {\mbox{\boldmath{$d$}}}_{1i}$, to find $$\begin{aligned}
{\mbox{\boldmath{$d$}}}_{ij} & = & D \left[ \begin{array}{c}
\sum_{k=i}^{j-1} \cos\theta_{k} \hat{{\mbox{\boldmath{$x$}}}} \\
\\
\sum_{k=i}^{j-1} \sin\theta_k \hat{{\mbox{\boldmath{$y$}}}}
\end{array} \right] \end{aligned}$$ Finally the relative distance is found by taking the magnitude, which also removes the explicit coordinate system dependance, $$\begin{aligned}
d_{ij}^2 & = & D^2 \left[ \left( \sum_{k=i}^{j-1} \cos\theta_{k} \right)^2 + \left( \sum_{k=i}^{j-1} \sin\theta_{k} \right)^2 \right] \end{aligned}$$ It is easy to verify that these distances reduce to the nominal case when $\theta_k = 0$, or indeed whenever all of the angles equal the same value.
Equilibrium Conditions
======================
Having the Euler Resting configuration properly specified, the conditions under which this configuration can exist as an equilibrium can be explored. First consider the simpler case of showing that the system as stated will be in equilibrium when $\theta_i = 0$, assuming contact between all components. Then the conditions for when the system can rest on each other and thus be in equilibrium will be considered.
Angular Equilibrium Conditions
------------------------------
Under the assumption that all of the bodies are resting on each other (to be reviewed in the following subsection), it can be shown that the condition when all of the bodies are aligned is in fact an equilibrium condition. To do this one must show that the gradient of the amended potential with respect to each angular degree of freedom is in fact equal to zero when evaluated at the nominal, or $\left. \partial{\cal E}/\partial\theta_m\right|_{\theta_n=0} = 0$ for all $m$ and $n$.
Evaluating this condition in the amended potential yields $$\begin{aligned}
\frac{\partial {\cal E}}{\partial \theta_m} & = & - \frac{H^2}{2 I_P^2} \frac{\partial I_P}{\partial\theta_m} + \frac{\partial {\cal U}}{\partial\theta_m}
\label{eq:E_partial} \\
\frac{\partial I_P}{\partial\theta_m} & = & \frac{{\cal M}}{N} \sum_{i<j} \frac{\partial d_{ij}^2}{\partial\theta_m}
\label{eq:IP_partial} \\
\frac{\partial {\cal U}}{\partial\theta_m} & = & \frac{1}{2}{\cal G}{\cal M}^2 \sum_{i<j} \frac{1}{d_{ij}^3} \frac{\partial d_{ij}^2}{\partial\theta_m}
\label{eq:U_partial} \end{aligned}$$ Thus, the key component is the gradient of $d_{ij}^2$ with respect to a single angle, $\theta_m$. $$\begin{aligned}
\frac{\partial d_{ij}^2}{\partial \theta_m} & = & \left\{ \begin{array}{lc}
0 & m \le i-1 \\
2 D^2 \left[ - \sin\theta_m \sum_{k=i}^{j-1}\cos\theta_k + \cos\theta_m \sum_{k=i}^{j-1}\sin\theta_k \right] \ &
\ i \le m \le j-1 \\
0 & m \ge j
\end{array} \right. \end{aligned}$$
Given this, it is simple to note that when substituting in the value of zero for all nominal angles $\theta_k$ or $\theta_m$ that the gradient is zero. Thus, the Euler Resting configuration is in equilibrium for the case when $\theta_n = 0$. It is interesting to note that the equilibrium condition is also satisfied if $\theta_n = \theta$, i.e., when each angle is at the same value. Then the two terms in the bracket are equal and opposite and cancel each other. This, in fact, represents the additional symmetry in this system noted earlier, which can be removed by requiring one of the angles to be fixed at zero.
Resting Equilibrium Conditions
------------------------------
The equilibrium condition for a constrained degree of freedom $d_i = 0$ chosen such that $\delta d_i \ge 0$ is that $\left.\partial{\cal E}/\partial d_i\right|_* \ge 0$. To evaluate this condition assume that the system is in its unconstrained equilibrium with $\theta_n = 0$ and that each of the distance variations is tested in turn.
To analyze the constrained degree of freedom for the contact between any two neighboring bodies, consider the effect of an increase between the distance between two neighboring bodies, $m$ and $m+1$, defined earlier as $d_{m}$. Given this deviation, the general distance between two particles $i$ and $j$, with $i < j$, will be denoted as $(j-i) D + \delta_m d_{i,j}$ where $$\begin{aligned}
\delta_{m} d_{i,j} & = & \left\{ \begin{array}{lc}
0 & m-1 \le i \\
d_{m} \ & \
i \le m \le j-1 \\
0 & m \ge j
\end{array} \right. \end{aligned}$$ More generally, the partial of $d_{ij}$ with respect to the distance $d_{m}$ is $$\begin{aligned}
\frac{\partial d_{ij}}{\partial d_{m}} & = & \left\{ \begin{array}{lc}
0 & m \le i-1 \\
1 \ & \
i \le m \le j-1 \\
0 & m \ge j
\end{array} \right. \end{aligned}$$
Take the partial of the amended potential with respect to $d_m$ and evaluate it at $d_i = 0$ to find $$\begin{aligned}
\left. \frac{\partial {\cal E}}{\partial d_{m}} \right|_* & = & - \frac{H^2}{2 I_P^2} \left. \frac{\partial I_P}{\partial d_{m}}\right|_* + \left. \frac{\partial {\cal U}}{\partial d_{m}}\right|_* \\
\frac{\partial I_P}{\partial d_{m}} & = & \frac{2 {\cal M}}{N} \sum_{i<j} d_{ij} \left. \frac{\partial d_{ij}}{\partial d_{m}}\right|_* \\
\frac{\partial {\cal U}}{\partial d_{m}} & = & {\cal G}{\cal M}^2 \sum_{i<j} \frac{1}{d_{ij}^2} \left. \frac{\partial d_{ij}}{\partial d_{m}}\right|_* \end{aligned}$$ When evaluated at the nominal resting position the value of $d_{ij} = D (j-i)$. Thus the general summation will be of the form $$\begin{aligned}
\sum_{i<j} f(d_{ij}) \frac{\partial d_{ij}}{\partial d_{m}} & = & \sum_{i=1}^{N-1} \sum_{j=i+1}^N f(j-i) \frac{\partial d_{ij}}{\partial d_{m}} \\
& = & \sum_{i=1}^{m} \sum_{j=m+1}^{N} f(j-i) \\
& = & \sum_{k=1}^{N-1} f(k) \min[ k, N-k, m, N-m ]\end{aligned}$$ There are two separate summations which must be evaluated, one of which can be done in closed form and the other must in general be computed numerically. The gradient of the moment of inertia yields $f(d_{ij}) = D(j-i)$ and the summation $$\begin{aligned}
\frac{\partial I_P}{\partial d_{m}} & = & \frac{2 {\cal M} D}{N} \sum_{k=1}^{N-1} k \min[ k, N-k, m, N-m ] \\
& = & {\cal M} D m (N-m) \end{aligned}$$ The gradient of the gravitational potential yields $f(d_{ij}) = \frac{1}{D^2 (j-i)^2}$ and the summation $$\begin{aligned}
\frac{\partial {\cal U}}{\partial d_{m}} & = & \frac{{\cal G}{\cal M}^2}{D^2} \sum_{k=1}^{N-1} \frac{1}{k^2} \min[ k, N-k, m, N-m ]\end{aligned}$$ which does not have a closed form and must be summed using an algorithm. The general form of this summation can be denoted as $g_s(m,N) = \sum_{k=1}^{N-1} k^{-s} \min[ k, N-k, m, N-m ]$, where for the above the function is $g_2(m,N)$. Note that the function can be solved in closed form when $g_{-s}(m,N)$, such as for the moment of inertia term, and that $g_s(1,\infty) = \zeta(s)$, the Riemann zeta function. Finally, note the symmetry $g_s(m,N) = g_s(N-m,N)$.
With these results, the original condition can now be explicitly expressed as $$\begin{aligned}
\left. \frac{\partial {\cal E}}{\partial d_{m}}\right|_* & = &
- \frac{H^2}{2 I_P^2} {\cal M} D m (N-m) + \frac{{\cal G}{\cal M}^2}{D^2} g_2(m,N) \end{aligned}$$ and the condition $\left. \partial{\cal E}/\partial d_{m}\right|_* \ge 0$ yields a constraint on the angular momentum of the system for the equilibrium to exist $$\begin{aligned}
\left( \frac{H}{I_P}\right)^2 \frac{D^3}{{\cal G}{\cal M}} \le \frac{2}{m(N-m)} g_2(m,N)\end{aligned}$$ for $m = 1, 2, \ldots, N-1$. However, given the symmetry obtained in replacing $m$ by $N-m$ one only needs consider $m=1, 2, \ldots, N/2$ for $N$ even and up to $(N-1)/2$ for $N$ odd.
Note that the ratio $H/I_P$ equals the total spin rate of the relative equilibrium, thus this can be interpreted directly as a constraint on the total spin rate as well. The expression tells us the upper limit on the system spin rate before the system will separate between the body $m-1$ and $m$. Since this condition must hold for all pairs of bodies, the final constraint is the minimum of this limit for all values of $m$. For the leading term, it is simple to note that that choosing $m \sim N/2$ will in general minimize the leading portion, however the $g_2$ function is maximized at this point. Combining the two expressions we find that the minimum occurs at $m = N/2$ for $N$ even or $m = (N\pm1)/2$ for $N$ odd. This then defines the spin rate at which the Euler Resting configuration will fission, and we note that this occurs by the body splitting “in half”. This can be predicted from first principles analysis as well, as the stress from centrifugal accelerations will be the highest at the mid-point of the spinning system and hence will be the first point at which it becomes positive and the system separates.
Thus the fission spin limit for the Euler Resting configuration is $$\begin{aligned}
\bar{\Omega}^2_F & = & \left\{
\begin{array} {lr}
\frac{2}{P^2} g_2(P,2P) & \ N = 2P \\
& \\
\frac{2}{P(P+1)} g_2(P,2P+1) & \ N = 2P+1
\end{array} \right.
\label{eq:fission}\end{aligned}$$ where $\bar{\Omega}$ denotes a non-dimensional spin rate of the system, $H/I_P$, normalized by $\sqrt{{\cal G}{\cal M}/D^3}$.
Stability of the Euler Resting Configuration
============================================
The previous analysis establishes the existence of the Euler Resting configuration for a particular alignment and for spin rates less than some value. For the constrained degrees of freedom, they are stable whenever they exist, and thus need not be analyzed further. The same is not true for the unconstrained equilibria defined by $\theta_n = 0$. The condition for these equilibria to be stable is that the Hessian of the amended potential evaluated at the equilibrium position be positive definite, or that the matrix $[ \left. \partial^2{\cal E} / \partial\theta_m \partial\theta_n \right|_*] > 0$.
Evaluating the Hessian
----------------------
Starting from Eqn. \[eq:E\_partial\] one finds $$\begin{aligned}
\left. \frac{\partial^2 {\cal E}}{\partial \theta_m\partial\theta_n}\right|_* & = & - \frac{H^2}{2 I_P^2} \left. \frac{\partial^2 I_P}{\partial\theta_m\partial\theta_n} \right|_*
+ \frac{H^2}{I_P^3} \left. \frac{\partial I_P}{\partial\theta_m}\right|_* \left. \frac{\partial I_P}{\partial\theta_n} \right|_*
+ \left. \frac{\partial^2 {\cal U}}{\partial\theta_m\partial\theta_n} \right|_*\end{aligned}$$ however evaluating this at the equilibrium condition removes the second term as $\left. \frac{\partial I_P}{\partial\theta_m}\right|_* =0$. Evaluating the second partials of the moment of inertia and gravitational potential yields $$\begin{aligned}
\left. \frac{\partial^2 I_P}{\partial\theta_m\partial\theta_n}\right|_* & = & \frac{{\cal M}}{N} \sum_{i<j} \left. \frac{\partial^2 d_{ij}^2}{\partial\theta_m\partial\theta_n}\right|_* \\
\left. \frac{\partial^2 {\cal U}}{\partial\theta_m\partial\theta_n}\right|_* & = & \frac{1}{2}{\cal G}{\cal M}^2 \sum_{i<j}\left. \frac{1}{d_{ij}^3} \frac{\partial^2 d_{ij}^2}{\partial\theta_m\partial\theta_n}\right|_*\end{aligned}$$ where for the potential energy term recall that the first partial of $d_{ij}^2$ is zero when evaluated at the equilibrium, leaving only the term involving the second partial.
Then, assuming that $m\le n$, the second partial is $$\begin{aligned}
\frac{\partial^2 d_{ij}^2}{\partial \theta_m\partial\theta_n} & = & \left\{ \begin{array}{lc}
0 & m \le i-1 \\
2 D^2 \left[ 1 - \sum_{k=i}^{j-1}\left[ \cos\theta_k\cos\theta_m + \sin\theta_k\sin\theta_m \right] \right] \ &
\ i \le m = n \le j-1 \\
2 D^2 \left[ \sin\theta_m \sin\theta_n + \cos\theta_m \cos\theta_n \right] \ &
\ i \le m < n \le j-1\\
0 & n \ge j
\end{array} \right. \end{aligned}$$ and evaluating this at the equilibrium condition yields $$\begin{aligned}
\left. \frac{\partial^2 d_{ij}^2}{\partial \theta_m\partial\theta_n}\right|_* & = & 2 D^2 \left\{ \begin{array}{lc}
0 & m \le i-1 \\
\left[ 1 - (j-i) \right] \ &
\ i \le m = n \le j-1 \\
1 \ &
\ i \le m < n \le j-1 \\
0 & n \ge j
\end{array} \right. \end{aligned}$$
The second partial of the moment of inertia is then $$\begin{aligned}
\left. \frac{\partial^2 I_P}{\partial\theta_m\partial\theta_n}\right|_* & = & \frac{{\cal M}}{N} \sum_{i<j} \frac{\partial^2 d_{ij}^2}{\partial\theta_m\partial\theta_n} {\partial\theta_m\partial\theta_n} \\
& = & \frac{2{\cal M} D^2}{N} \sum_{i=1}^m \sum_{j=n+1}^N
\left\{ \begin{array}{lr}
1 - (j - i) & m = n \\
1 & m < n
\end{array} \right. \\
& = & \frac{2{\cal M} D^2}{N}
\left\{ \begin{array}{lr}
- \frac{1}{2} m (N-m) (N-2) & m = n \\
m(N-n) & m < n
\end{array} \right. \end{aligned}$$ The second partial of the potential energy is then $$\begin{aligned}
\left. \frac{\partial^2 {\cal U}}{\partial\theta_m\partial\theta_n}\right|_* & = & \frac{{\cal G}{\cal M}^2}{2D^3} \sum_{i<j} \frac{1}{(j-i)^3} \frac{\partial^2 d_{ij}^2}{\partial\theta_m\partial\theta_n} {\partial\theta_m\partial\theta_n} \\
& = & \frac{{\cal G}{\cal M}^2}{D} \sum_{i=1}^m \sum_{j=n+1}^N
\left\{ \begin{array}{lr}
\frac{1 - (j - i)}{(j-i)^3} & m = n \\
\frac{1}{(j-i)^3} & m < n
\end{array} \right. \end{aligned}$$ Introduce a generalization of the previous summation formula $$\begin{aligned}
\sum_{i=1}^m \sum_{j=n+1}^N f(j-i) & = & \sum_{k=n-m+1}^{N-1} f(k) \min[ k-(m-n), N-k, m, N-n ]\end{aligned}$$ and the generalized summation function $$\begin{aligned}
h_s(m,n,N) & = & \sum_{k=n-m+1}^{N-1} \frac{1}{k^s} \min[ k-(n-m), N-k, m, N-n ]\end{aligned}$$ where $m\le n$ and $h_s(m,m,N) = g_s(m,N)$. Then the potential energy partials can be expressed as $$\begin{aligned}
\left. \frac{\partial^2 {\cal U}}{\partial\theta_m\partial\theta_n}\right|_* & = & \frac{{\cal G}{\cal M}^2}{D}
\left\{ \begin{array}{lr}
g_3(m,N) - g_2(m,N) & m = n \\
h_3(m,n,N) & m < n
\end{array} \right. \end{aligned}$$
Thus, there are two possible forms for the second partial of the amended potential, denoted in short hand as ${\cal E}_{mn}$. $$\begin{aligned}
{\cal E}_{mm} & = & \frac{H^2}{2 I_P^2} {{\cal M} D^2} \frac{m (N-m) (N-2)}{N}
- \frac{{\cal G}{\cal M}^2}{D} \left[ g_2(m,N) - g_3(m,N) \right] \end{aligned}$$ where $g_2 > g_3$ in general. For the off-diagonal terms ${\cal E}_{mn} = {\cal E}_{nm}$ and, for $m < n$ $$\begin{aligned}
{\cal E}_{mn} & = & - \frac{H^2}{I_P^2} {\cal M} D^2\frac{m(N-n)}{N}
+ \frac{{\cal G}{\cal M}^2}{D} h_3(m,n,N) \end{aligned}$$
The condition for the system to be stable is that the matrix $[{\cal E}_{ij}]$ be positive definite. There are a number of ways to set up the necessary and sufficient conditions. Two that we use here are to determine that the principal minors are all positive, or that the eigenvalues of the matrix are all positive. To aid in these discussions rewrite the Hessian in a scaled version where $\bar{\cal E} = {\cal E} / ({\cal G}{\cal M}^2/D)$ and $\bar{\Omega} = (H/I_P) / \sqrt{{\cal G}{\cal M}/D^3}$. Then $$\begin{aligned}
\bar{\cal E}_{mm} & = & \bar{\Omega}^2 \frac{m (N-m) (N-2)}{2 N}
- \left[ g_2(m,N) - g_3(m,N) \right] \\
\bar{\cal E}_{mn} & = & - \bar{\Omega}^2 \frac{m(N-n)}{N}
+ h_3(m,n,N) \end{aligned}$$
Necessary Conditions
--------------------
A useful necessary condition is that all of the diagonals of the matrix be positive, ${\cal E}_{mm} > 0$, as this is a precondition for the principal minors to be positive. This yields a simple condition on the normalized spin rate of the system, $$\begin{aligned}
\bar{\Omega}^2 & > & \frac{2N}{m(N-m)(N-2)} \left[ g_2(m,N) - g_3(m,N) \right] \label{eq:nec}\end{aligned}$$ with the additional restriction that $N \ge 3$, and that $1\le m \le N-1$. From this it is clear that $H^2$ can always be chosen small enough so that the Euler Resting configuration is definitely unstable. Similarly, the angular momentum must be greater than the maximum of the right-hand side for stability to be a possibility.
This condition is in direct competition for the existence condition for the Euler Resting configuration, summarized in Eqn. \[eq:fission\]. Comparing Eqn. \[eq:nec\] with Eqn. \[eq:fission\] one finds that the angular momentum for the necessary condition to be satisfied exceeds the existence condition limit for $N \ge 9$. Thus this sets a limit on the size of an Euler Resting configuration, requiring a more detailed analysis only up to this value. Of course, sharper necessary conditions can be tested, such as the trace of the matrix, however these become progressively more complex to analyze, making the jump directly to the sufficiency conditions reasonable.
Sufficient Conditions
---------------------
To check the sufficient condition, the positive definiteness of the amended potential Hessian must be checked across a range of angular momenta between the necessary condition for stability up to the fission limit. To make this computation definitive, the sufficiency condition can be evaluated at the fission limit value given in Eqn. \[eq:fission\], to check whether the stability condition is satisfied up to this level. Carrying out the computation shows that for $N=5$ the sufficiency condition is definitely satisfied when the system is evaluated at the fission condition, however at $N=6$ and higher it is violated with at least one negative eigenvalue.
Note that the case $N=2$ is trivially stable when it exists, and that the cases $N=3,4$ have been studied in detail previously. Thus, in the current paper a complete analysis is only carried out for the case $N=5$, which is also the only remaining case to be studied that can be stable.
Stability limits of the $N=5$ case
----------------------------------
Per the earlier discussion on the number of free degrees of freedom, only the variation of $5-2 = 3$ degrees of freedom need to be considered, which can be arbitrarily chosen out of the set $\theta_i, i=1,2,3,4$. For notational ease we consider the first three, giving the specific entries for the matrix as follows: $$\begin{aligned}
\bar{\cal E}_{11} & = & \bar{\Omega}^2 \frac{6}{5}
- \left[ g_2(1,5) - g_3(1,5) \right] \\
\bar{\cal E}_{22} & = & \bar{\Omega}^2 \frac{9}{5}
- \left[ g_2(2,5) - g_3(2,5) \right] \\
\bar{\cal E}_{33} & = & \bar{\Omega}^2 \frac{9}{5}
- \left[ g_2(3,5) - g_3(3,5) \right] \\
\bar{\cal E}_{12} & = & - \bar{\Omega}^2 \frac{3}{5}
+ h_3(1,2,5) \\
\bar{\cal E}_{13} & = & - \bar{\Omega}^2 \frac{2}{5}
+ h_3(1,3,5) \\
\bar{\cal E}_{23} & = & - \bar{\Omega}^2 \frac{4}{5}
+ h_3(2,3,5) \end{aligned}$$ Where the values of the $g_s$ and $h_s$ functions are given in Table \[tab:functions\].
-------------- ------------------
$g_2(1,5)$ 1.42361111193895
$g_2(2,5)$ 1.78472222387791
$g_2(3,5)$ 1.78472222387791
$g_3(1,5)$ 1.17766203731298
$g_3(2,5)$ 1.33969907462597
$g_3(3,5)$ 1.33969907462597
$h_3(1,2,5)$ 0.17766203731298
$h_3(1,3,5)$ 0.05266203731298
$h_3(2,3,5)$ 0.21469907462597
-------------- ------------------
: Specific values of the functions $g_s$ and $h_s$ for $N=5$.[]{data-label="tab:functions"}
The range of values to be explored go up to $\bar{\Omega}_F^2 = g_2(2,5) / 3$ or $\bar{\Omega}_F = 0.7713024\ldots$, which is the fission spin rate for $N=5$. Through computation we find that the determinant is the defining condition for the stability of this system, which is to be expected as it is the product of the eigenvalues. Evaluating the determinant as the controlling condition, we have a cubic equation in $\bar{\Omega}^2$ to be considered. This can be iteratively solved to find the zero crossing, which occurs at a value of $\bar{\Omega}_S / \bar{\Omega}_F = 0.674064\ldots$, or at a normalized value of $\bar{\Omega}_S = 0.5199071\ldots$. Thus, the Euler Resting Configuration for $N=5$ is stable when the normalized spin rate lies in the interval $$\begin{aligned}
0.5199071\ldots \le \bar{\Omega} \le 0.7713024\ldots\end{aligned}$$ At lower values of angular momentum the system is unstable relative to the angular variations. At higher values of angular momentum the system will fission into two components, one with two connected bodies and the other with three.
Discussion
==========
It is instructive to compare the different spin limits for the different Euler resting configurations. The comparable spin rate at which the system becomes stabilized, as a fraction of the fission spin rate, is zero for the $N=2$ case, is $\sqrt{0.3} = 0.5477$ for the $N=3$ case, and is $0.6343$ for the $N=4$ case. Thus, the $N=5$ case has the smallest interval of normalized spin rate for stability, as measured by the fission spin rate. The complete fission and stabilization non-dimensional spin rates as a function of $N$ is given in Table \[tab:fission\].
To bring these results back to the physical domain, the dimensional value of these limiting spin rates will be considered for typical body densities. First, note that the normalizing spin rate is only a function of the density of the bodies, as $\sqrt{ {\cal G}{\cal M}/D^3} = \sqrt{{\cal G}\pi \rho / 6}$, where ${\cal G} = 6.6702\times10^{-11}$ m$^3$/ kg / s$^2$ and where $\rho$ is the body density measured in units of kg/m$^3$. Typical values of density for natural solar system bodies can range from 500 for cometary bodies up to 5000 for boulders with low porosity. This corresponds to a range of normalizing frequencies from $1.3215\times10^{-4} \rightarrow 4.1788\times10^{-4}$ rad/s, or equivalently normalizing rotation periods from $13.2076 \rightarrow 4.1766$ hr.
$N$ $\bar{\Omega}_F$ $\bar{\Omega}_S$
----- ------------------ ------------------
2 $1.4142$ 0
3 $1.1180$ 0.6123
4 0.8975 0.5693
5 0.7713 0.5199
: Non-dimensional fission and stabilization and fission spin rates. To account for densities ranging from $500\rightarrow5000$ kg/m$^3$ multiply by $1.3215\times10^{-4} \rightarrow 4.1788\times10^{-4}$ rad/s, respectively.[]{data-label="tab:fission"}
That the Euler Resting configuration is only stable up to $N=5$ places a limit on the elongation that a collection of resting particles can have at 5:1. While such an elongation in a natural body is extreme, it is interesting to note that the first interstellar object, 1I/‘Oumuamua, was found to have an elongation commensurate with this level. A fundamental question of interest for future studies is whether a similar elongation limit holds when systems of more bodies are considered, arranged to have additional components stacked on each other in directions transverse to the long axis.
This research was supported by NASA grant NNX14AL16G. The author declares no conflict of interests.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Click-through rate prediction is an essential task in industrial applications, such as online advertising. Recently deep learning based models have been proposed, which follow a similar Embedding&MLP paradigm. In these methods large scale sparse input features are first mapped into low dimensional embedding vectors, and then transformed into fixed-length vectors in a group-wise manner, finally concatenated together to fed into a multilayer perceptron (MLP) to learn the nonlinear relations among features. In this way, user features are compressed into a fixed-length representation vector, in regardless of what candidate ads are. The use of fixed-length vector will be a bottleneck, which brings difficulty for Embedding&MLP methods to capture user’s diverse interests effectively from rich historical behaviors. In this paper, we propose a novel model: Deep Interest Network (DIN) which tackles this challenge by designing a local activation unit to adaptively learn the representation of user interests from historical behaviors with respect to a certain ad. This representation vector varies over different ads, improving the expressive ability of model greatly. Besides, we develop two techniques: mini-batch aware regularization and data adaptive activation function which can help training industrial deep networks with hundreds of millions of parameters. Experiments on two public datasets as well as an Alibaba real production dataset with over 2 billion samples demonstrate the effectiveness of proposed approaches, which achieve superior performance compared with state-of-the-art methods. DIN now has been successfully deployed in the online display advertising system in Alibaba, serving the main traffic.'
author:
- |
Guorui Zhou, Chengru Song, Xiaoqiang Zhu\
Ying Fan, Han Zhu, Xiao Ma, Yanghui Yan, Junqi Jin, Han Li, Kun Gai
bibliography:
- 'DIN.bib'
title: 'Deep Interest Network for Click-Through Rate Prediction'
---
<ccs2012> <concept> <concept\_id>10002951.10003260.10003272.10003275</concept\_id> <concept\_desc>Information systems Display advertising</concept\_desc> <concept\_significance>500</concept\_significance> </concept> <concept> <concept\_id>10002951.10003317.10003347.10003350</concept\_id> <concept\_desc>Information systems Recommender systems</concept\_desc> <concept\_significance>500</concept\_significance> </concept> </ccs2012>
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'It is shown that the barbell distribution of a gas of relativistic molecules above its critical temperature, can be interpreted as an antifragile response to the relativistic constraint of subluminal propagation.'
author:
- 'Sauro Succi [^1]'
title: Relativistic antifragility
---
Introduction {#intro}
============
In the recent years, the notion of anti-fragility, as introduced by Nassim Taleb [@TALEB], has gained a boost of popularity across most walks of science and society. Although to a physicist the term “anti-fragility” sounds a bit like a clever renaming of an old wine in a new bottle, the value of extending and applying the notion beyond the context of the natural sciences, finance first, up to a philosophy of life, should not be underestimated.
In this short note, we wish to point out that relativistic kinetic theory, namely the statistical theory of particles which obey Einstein’s (special) relativity, shows distinct signatures of anti-fragility, most notably the emergence of extreme (bell-bar) statistics in-lieu of standard “Mediocristian” gaussian distribution, in response to increasing thermal load. The reason for such transition from Mediocristan to Extremistan, to borrow from Taleb’s terminology is adamant: material particles cannot move faster than light.
Maxwell-Boltzmann statistics {#sec:1}
============================
It is well known that a gas of non-relativistic molecules at equilibrium obeys the Maxwell-Boltzmann (MB) distribution (one spatial dimension for simplicity) [@BOL]: $$\label{MB}
f_{MB}(v) = \frac{n}{(2\pi)^{1/2} v_T} e^{-\frac{(v-u)^2}{2v_T^2}}$$ In the above $n$ is the gas number density (number of molecules per unit volume), $u$ is the macroscopic gas velocity and $v_T = \sqrt{\frac{k_BT}{m}}$ is the thermal speed, $T$ being the temperature and $m$ the mass of the particle. According to kinetic theory, the macroscopic gas velocity $u$ coincides with the mean molecular velocity, namely: $$u = <v> = \frac{1}{n} \int_{-\infty}^{+\infty} f v dv$$ Note that the integral in velocity space runs from minus infinity to plus infinity since classical mechanics sets no restrictions on the magnitude of the particle velocity. Likewise, $mv_T^2/2=k_BT/2$ is the kinetic energy contained in the molecular fluctuations $$k_BT= <m(v-u)^2> = \frac{1}{n} \int f mv^2 dv$$ In other words, the kinetic energy of the gas splits into the sum of a macroscopic (mechanical) and a microscopic (thermal) components $$E_K = \frac{mu^2}{2} + \frac{m v_T^2}{2}$$ In classical kinetic theory both terms are potentially unlimited.
A few comments are in order.
[*First*]{}, the mean velocity $u$ is also the most probable one, in the sense that the MB distribution attains its peak value precisely at $v=u$. This is typical conformistic-behaviour, most molecules “go with the flow”, they move at the same speed as the average.
[*Second*]{}, such conformistic behaviour is fairly intolerant of outliers, namely particles which move much faster or slower than the average are exponentially suppressed. For instance, molecules moving at three thermal speed faster than average (FTA) are suppressed by more than 1:1000, and at five FTA their number goes down to about one in a million! Extreme behaviour is suppressed in Mediocristan. Hence, “socially” speaking, the MB distribution speaks for a comfortable and stable world in which most individuals behave like the average and those who don’t are exponentially suppressed. For good or for worst, this is the most stable statistics a gas of classical molecules can achieve, as sharply pinpointed by Boltzmann’s H-theorem, showing that any different distribution is bound to converge to $f_{MB}$ in order to maximize its entropy, basically a micrioscipic underpinning of the second principle of Thermodynamics.
[*Third*]{}, the MB distribution encodes the non-relativistic principle of Galilean invariance, i.e. the statistics does not depend on the absolute molecular velocity $v$, but on its speed relative to the average, $v-u$, also known as peculiar velocity, the one characterizing microscopic fluctuations.
Conservation constraints {#sec:2}
========================
The MB distribution maximizes entropy under the constraint of mass, momentum and energy conservation, i.e $$\begin{aligned}
\int_{-\infty}^{+\infty} f_{MB} dv = n\\
\int_{-\infty}^{+\infty} f_{MB} v dv = nu\\
\int_{-\infty}^{+\infty} f_{MB} v^2 dv = nu^2 + n v_T^2\end{aligned}$$
In Taleb’s parlance, constraints are the “stressors”, sources of stress, since they constrain the freedom of the system.
A few numbers won’t hurt. For air in standard conditions, 300 Kelvin degrees, the thermal speed is close to $300$ meters per second (basically the sound speed). If the gas is macroscopically at rest, $u=0$, the probability of finding a molecule moving faster than one thermal speed, in either direction, is a sizeable 16 percent. At two thermal speeds, the number goes down to about 2 percent, and at above three thermal speeds, we are left with just one in thousands. That means that in a sample of thousands molecules, on average, only one moves faster than $3 \times 300=900$ m/s. By iterating the game, numbers get rapidly ridiculously small, at five thermal speeds, ($1500$ m/s) only one in a billion is left. This speaks clearly for outlier suppression.
Now suppose that we set the gas in motion, with a substantial macroscopic speed, say $u=30$ m/s (100 Km/h, a pretty strong wind). How does the “molecular society” adjust to such stressor? The answer is fairly straightforward: by simply shifting the MB distribution towards positive values so that the new peak locates precisely at $v=u$. This breaks the left/right symmetry, the probability of finding a particle moving right at a given speed $+v$ is higher than the probability of moving left with the same but opposite speed $-v$. The probability of finding a molecule moving at $300$ m/s is $20/100$, slightly larger than $16/100$ and the probability of moving at $-300$ m/s is $12/100$, slightly smaller than $16/100$, but the distribution keeps the same bell-shaped form, symmetric around $u=30$ m/s. The gas of molecules adjusts to the constraint without compromising any of the three basic features described above, Mediocristan still rules.
Now suppose you keep the gas at rest, but increase its temperature instead, say $600$ Kelvin instead of $300$. How does the gas adjust to this thermal stressor?
Again the policy is adamant; by simply “broadening” its distribution, so that the average kinetic energy in the fluctuations is doubled, while the mean velocity is left unchanged. This means that now there are sixteen out of hundred molecules which travel at $600$ m/s instead of $300$ m/s. This is how the system manages to increase its random energy content, but again, none of three distinctive features above are broken. In particular, while the high velocity region gets more populated, the most probable molecules remain those that move at mean speed, in this case zero. The outliers are less suppressed, but they don’t take over the conformists.
At this point, it is worth noting that the conformists contribute a little precious nothing to the temperature constraint, since they carry no energy at all! But they do contribute to the constraint that the gas should not move on average, so they still have a role in this business. This argument is flawed, but in classical mechanics the flaw remains silent, as we shall see shortly.
In classical kinetic theory you can play the game ad libitum, heat to the point of making the sound speed equal to the speed of light, and the most probable molecules will still be the ones at rest, although by an hardly appreciable extent, since the MB distribution becomes utterly flat. The outliers are no longer such, granted, but the “conformist” are still there.
Since the outliers can afford unlimited speed, the system can absorb virtually any amount of thermal energy, by simply increasing their population at expense of the conformists. Yet, the thermal constraint can [*always*]{} be matched without destroying the conformist in the process, even though they contribute nothing to the energy budget!
The availability of unlimited speed provides the tolerance towards the zero-speed population of molecules.
Relativistic kinetic theory {#sec:3}
===========================
Next we move to relativistic molecules.
The relativistic analogue of the MB distribution, known as Maxwell-Juettner (MJ) distribution,reads as follows [@JUT; @CERCI]: $$f_{MJ}(v) = n A(z) \gamma_v^{3} e^{-z \gamma_v \gamma_u (1-uv)}$$ where $\gamma_v = (1-v^2/c^2)^{-1/2}$ and $\gamma_u = (1-u^2/c^2)^{-1/2}$ are the molecular and gas Lorentz factors, $z=mc^2/kT$ is the rest energy in thermal units and $A(z)$ a normalization prefactor [@CERCI]. The MJ distribution follows directly from the expression of the relativistic energy, $E^2=m^2 c^4 + p^2 c^2$, which in turn encodes Lorentz rather than Galilean in invariance. This means that the MJ depends no longer on the peculiar speed $v-u$, but on the (scalar in multi-dimensions) product $vu$.
In addition, it obeys the boundary condition $f(v \to c) \to 0$, in compliance with the relativistic constraint $v \le c$.
Crucial to the MJ statistics is the term $\gamma_v^{3}$, which stems from the metric transformation from the momentum distribution $f(p)$ to the velocity distribution $f(v)$. Such factor exhibits a cubic divergence in the limit $v/c \to 1$, thereby promoting depletion of the conformist population in favour of the outliers. This divergence, in turn, results from the fact that while relativistic velocities are bound, the corresponding momenta are not, which
This divergence is tamed by the exponential term $e^{-\gamma_v}$ which enforces the boundary condition $f(v=c)=0$. The MJ distribution reflects the basic competition between these two terms, whose outcome is strongly sensitive to temperature via the parameter $z$.
In the non-relativistic limits $v/c \ll 1$ and $1/z \to 0$, the MJ reduces to the MB distribution, hence it reacts to flow and thermal constraints in a similar way.
In the genuinely relativistic limit $\gamma_v, \gamma_u \gg 1$, however, the response is completely different. To appreciate the point, let’s consider again the case at rest, $u=0$. It can be readily checked that upon increasing the temperature, i.e. $z \to 0$, the MJ undergoes a depletion of the conformist region in favor of the emergence of a barbell distribution, with two distinct peaks away from the origin. More precisely, the transition from a MB-like unimodal to the bimodal barbell distribution occurs sligthly above a critical temperature $kT_{crit} \sim mc^2$, namely below $z \sim 1$. This stands in sharp contrast with the reaction of the MB distribution, which broadens indefinitely in response to the constraint of increasing temperature, without ever developing any double-humped structure.
Why such a different strategy? Simply because the relativistic molecules cannot afford infinite-speed, the fast runners must nevertheless stop at $v=\pm c$. Hence, the only chance for the system to meet an increasing thermal constraint is to enhance the fastest-running populations as much as possible. However, since the outliers can no longer afford infinite speed, a point comes where the conformists are no longer sustainable, and above the critical temperature they begin to be suppressed.
But, didn’t we say that the conformists are needed to comply with the net flow constraint, i.e. zero net motion for the case in point?
That is true, but now the system “realizes” that the same constraint can be fulfilled without the conformists, by simply keeping an exact balance between the left and right movers. In other words, the relativistic barrier $v \le c$ exposes the “uselessness” of the comformists and the system gets rid of them: that’s Extremistan in full action!
But how about the case of a net macroscopic motion, say $u>0$?
Detailed inspection of the MJ distribution [@NUNO], shows that this is accomodated by simply developing a positive bias on the right-moving hump and a negative one on the left-moving one.
This mechanism, known as “skweness”, reflects Lorentz invariance, as opposed to the rigid shift of the non-relativistic case, which reflects Galilean invariance instead.
Quantitative analysis permits to compute the exact location of the humps and their width as a function of the macroscopic parameters $u/c$ and $z$, but does not add anything substantial to the essence of the story. And the essence is that, above a critical temperature, the relativistic constraint $v \le c$ makes the conformist simply unsustainable.
The emergence of the barbell distribution then appears as quintessential antifragility, i.e an innovative survival strategy which was left silent in the non-relativistic case, just due to the availability of unlimited-speed runners.
“Social” relativity {#sec:4}
===================
The “biological” interpretation of the MJ distribution is adamant: a survival strategy against thermal constraints in a finite-resource (velocity) environment. The social one, as usual, is a bit less straightforward. The collapse of the conformist in relativistic gas is strongly conducive to the economic nosedive of the middle class in the modern “the winner takes it all” global economy.
With a big pinch of imagination, one might even posit that this is due to the mind-boggling acceleration of financial transactions, a sort of analogue of the relativistic limit $v \to c$. In a world where data and algorithms can evaporate a lifetime savings at keystroke speed, quick-witted “influencers” and fake news, win hands down over “slow-paced” engineers, spelling doom for the real economy against the virtual one.
The second consideration is for the conformists, which relativity reveals in their true colors, basically as a disposable population. In human society, some individuals don’t move because they don’t want to, but some others don’t because they really can’t. The antifragile policy is fair for the former but not for the latter.
Acknowledgments {#acknowledgments .unnumbered}
===============
This research leading has received funding from the European Research Council under the European Union’s Horizon 2020 Framework Programme (No. FP/2014-2020)/ERC Grant Agreement No. 739964 (COPMAT).
[99]{}
N. Taleb, Antifragile: Things That Gain From Disorder, Random House, NY (2012)
L. Boltzmann, Lectures on Gas Theory, University of California Press, Berkeley, (1964)
F. Juettner, Annalen der Physik. 339 (5): 856-882 (1911), C. Cercignani and G.M. Kremer, The Relativistic Boltzmann Equation: Theory and Applications, Birkhauser, Berlin (2002).
M. Mendoza, N. Arajuno, H. Herrmann and S. Succi, Sci. Rep, 2, 611, (2012)
[^1]: Electronic address: `[email protected]`; Corresponding author
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We consider a class of nonsmooth optimization problems over Stefiel manifold, which are ubiquitous in engineering applications but still largely unexplored. We study this type of nonconvex optimization problems under the settings that the function is weakly convex in Euclidean space and locally Lipschitz continuous, where we propose to address these problems using a family of Riemannian subgradient methods. First, we show iteration complexity $\calO(\varepsilon^{-4})$ for these algorithms driving a natural stationary measure to be smaller than $\varepsilon$. Moreover, local linear convergence can be achieved for Riemannian subgradient and incremental subgradient methods if the optimization problem further satisfies the sharpness property and the algorithms are initialized close to the set of weak sharp minima. As a result, we provide the first convergence rate guarantees for a family of Riemannian subgradient methods utilized to optimize nonsmooth functions over Stiefel manifold, under reasonable regularities of the functions. The fundamental ingredient for establishing the aforementioned convergence results is that any weakly convex function in Euclidean space admits an important property holding uniformly over Stiefel manifold which we name *Riemannian subgradient inequality*. We then extend our convergence results to a broader class of compact Riemannian manifolds embedded in Euclidean space. Finally, we discuss the sharpness property for robust subspace recovery and orthogonal dictionary learning, and demonstrate the established convergence performance of our algorithms on both problems via numerical simulations.'
author:
- |
Xiao Li, Shixiang Chen, Zengde Deng,\
Qing Qu, Zhihui Zhu and Anthony Man Cho So [^1]
bibliography:
- 'nonconvex.bib'
- 'manifold.bib'
title: |
Nonsmooth Optimization over Stiefel Manifold:\
Riemannian Subgradient Methods
---
[**[Key words —]{}** Manifold optimization, nonconvex optimization, orthogonal constraint, iteration complexity, linear convergence, robust subspace recovery, dictionary learning.]{}
Introduction {#sec:introduction}
============
Preliminaries and Algorithms
============================
Riemannian Subgradient Inequality over Stiefel Manifold {#sec:Riemannian subgradient inequality}
=======================================================
Global Convergence {#sec:global convergence}
==================
Local Linear Convergence for Sharp Functions \[sec:linear convergence\]
=======================================================================
Extension {#sec:extension}
=========
Applications and Experimental Results {#sec:application}
=====================================
Conclusion
==========
Acknowledgement {#acknowledgement .unnumbered}
===============
XL would like to acknowledge the support by Grant CUHK14210617 from the Hong Kong Research Grants Council. QQ thanks the support of Moore-Sloan fellowship. ZZ was partly supported by NSF Grant 1704458.
[^1]: *The first and second authors contributed equally to this paper*. Xiao Li is with Electronic Engineering, The Chinese University of Hong Kong (e-mail: [email protected]). Shixiang Chen is with Industrial and Systems Engineering, Texas A$\&$M University (e-mail: [email protected]). Zengde Deng and Anthony Man Cho So are with Systems Engineering and Engineering Management, The Chinese University of Hong Kong (e-mail: {zddeng, manchoso}@se.cuhk.edu.hk). Qing Qu is with Center for Data Science, New York University (e-mail: [email protected]). Zhihui Zhu is with Mathematical Institute for Data Science, Johns Hopkins University (e-mail: [email protected]).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The response functions for the unpolarized $(e,e''p)$ and polarized $(\vec{e},e''p)$ reaction are calculated for medium-heavy nuclei under quasifree conditions. The formalism presented here incorporates two-body currents related to meson-exchange and the $\Delta (1232)$ excitation. The final-state interaction of the outgoing nucleon with the residual nucleus is handled in a HF-RPA formalism. The sensitivity of the results to the two-body currents is discussed for the five structure functions in quasielastic $(\vec{e},e''p)$ scattering off the target nuclei $^{16}O$ and $^{40}Ca$. A selective sensitivity to the two-body currents is obtained in the longitudinal-transverse interference term $W_{LT}$ where two-body currents can explain part of the discrepancy between the impulse-approximation calculations and the data.'
author:
- |
V. Van der Sluys, J.Ryckebusch[^1] and M.Waroquier[^2]\
[*Laboratory for Nuclear Physics*]{}\
[*Laboratory for Theoretical Physics*]{}\
[*Proeftuinstraat 86*]{}\
[*9000 Gent, Belgium*]{}
title: 'Effects of meson-exchange currents on the $(\protect\vec{e},e''p)$ structure functions'
---
=-1cm =0.5cm =0.5cm
The coincidence $(e,e'p)$ reaction when performed under quasifree conditions has proven to be an excellent tool in the study of single-particle properties of the nucleus [@frul]. The analysis of the quasielastic $(e,e'p)$ reaction yields information on the single-particle wave functions, spectroscopic factors and strength distributions. Generally, the quasielastic $(e,e'p)$ results have been analyzed within the framework of a relativistic [@jin][@pick] or non-relativistic [@rep][@jan2][@bern] distorted wave impulse approximation (DWIA) approach. In these models the quasielastic electromagnetic response is expected to be dominated by one-body interactions. Hence, the nuclear current is handled in the impulse approximation (IA) in which it will be regarded as the sum of the one-body currents from the individual nucleons. The final-state interaction of the outgoing proton with the residual nucleus is treated in a distorted wave approximation either within an optical potential model [@rep] or within a microscopic HF-RPA approach [@jan]. On the average, the $(e,e'p)$ cross sect ions can be reasonably well reproduced within this DWIA approach.
At present, improved experimental techniques have become available and measureme nts have been carried out to extract more detailed information on the response to th e longitudinal and transverse polarization states of the virtual photon. Assuming the one-photon exchange approximation, the unpolarized $(e,e'p)$ cross section can be written in terms of four nuclear structure functions each multiplied with a different kinematical factor. In case of polarized incoming electrons, an additional helicity $h(\pm 1)$ dependent structure function $W_{LT}'$ can be extracted e[ras]{}[@pick2]. The angular distribution of the polarized $(\vec{e},e'p)$ cross section is determined by the following expression $$\begin{aligned}
\label{cross}
\frac{{\rm d}^{5}\sigma}{{\rm d}E_{f}{\rm d}\Omega_{E_{f}}{\rm
d}\Omega_{p_{a}}}(\vec{e},e'p) &=&
C \left\{v_{L}W_{L} + v_{T}W_{T} + v_{TT}W_{TT}\cos 2\varphi_{a}
\right.\nonumber \\ && \qquad\left.+ v_{LT}W_{LT}\cos\varphi_{a} + h v_{LT}'
W_{LT}' \sin \varphi_{a}
\right\}.\end{aligned}$$ All five structure functions $W$ depend on the momentum and energy transfer $(\vec{q},\omega$) of the virtual photon, the proton momentum $p_{a}$ and proton angle $\theta_a$. The angle between the scattering plane and the reaction plane is denoted by $\varphi_{a}$. The various structure functions $W$ are independent of the electron kinematics and are sensitive in a particular way to a different aspect of the re action mechanism. A complete experimental determination of the structure functions $W$ under suitable electron kinematics, could yield additional information on the reaction mechanism and might impose a constraint on the theoretical model. Determination of all five structure functions asks for polarized electrons and out-of plane experiments which are generally difficult to realize.
Recently, new results of in-plane measurements have become available for $^{16}O(e,e'p)$ [@saclay][@chiara][@wies] and $^{40}Ca(e,e'p)$ [@wies][@kramer]. These experiments yield information o n three nuclear structure functions $W_{L}+q^{2}/(2q_{\mu}q^{\mu}) W_{TT}$, $W_{T}$ and $W_{LT}$ . In ref.[@saclay] the $^{16}O(e,e'p)$ data have been compared with the results of the non-relativistic and relativistic DWIA calculations of Van Orden [*et al.*]{}. Both theoretical models are found to produce comparable results pointing towards relativity playing a rather unimportant role in quasielastic kinematics. The DWIA result reproduces the shape of the total cross section and the structure functions. However, a consistent descr iption of both the cross section and structure functions could neither be obtained within the relativistic nor the non-relativistic approach since the fitted reduction factors for the $^{16}O(e,e'p)^{15}N(1/2^{-},g.s)$ and the $^{16}O(e,e'p)^{15}N(3/2^{-},6.32 $ MeV) cross sections considerably deviated from corresponding $W_{LT}$ terms. Other $(e,e'p)$ separation measurements on $^{16}O$ and $^{40}Ca$ have been performed at NIKHEF -K [@chiara][@wies]. These data have been confronte d with the non-relativistic DWIA model of the Pavia group [@rep]. For the $(e,e'p)$ processes feeding the residual (A-1) nucleus in its ground state ($1p_{1/2}$ knockout in $^{16}O$ and $1d_{3/2}$ knockout in $^{40}Ca$) a fair agreement between theory and experiment is observed. For knockout from the corresponding spin-orbit partners ($1p_{3/2}$ in $^{16}O$ and $1d_{5/2}$ in $^{40}Ca$) the absolute $W_{LT}$ is considerably larger than what the theory predicts. This means that the quasielastic electron excitation of the nucleus is more comp lex than it was generally believed. [*The aim of the present work is to go beyond the DWIA by including two-body contributions in the nuclear current from meson-exchange and intermediate delta excitations and to investigate the effect of these two-body currents on the structure functions*]{}.
In order to determine the structure functions the following transition matrix elements of the nuclear current $$\begin{aligned}
\label{matr}
<J_{R}M_{R};\vec{p}_{a},1/2m_{{s}_{a}}\mid J_{\mu}(q)\mid J_{i}M_{i}>\end{aligned}$$ have to be calculated. The final state in this matrix element refers to the residual nucleus in a state $\mid \! J_{R}M_{R}\! >$ and an escaping particle with momentum $\vec{p}_{a}$ an d spin projection $m_{{s}_{a}}$. Throughout this work the residual nucleus is considered to remain in a pure hole state relative to the ground state $\mid \!J_{i}M_{i}\!>$ of the target nucleus. The spectroscopic factor, extracted from a least square fit of the calculated cross section to the data, reflects the amount of hole strength in the final state. The wave function for the escaping particle and the residual nucleus is obtained in the continuum RPA formalism as described in ref. [@jan]. The RPA formalism involves a multipole expansion in terms of linear combinations of particle-hole and hole-particle excitations out of a correlated ground state. As such we account for the multi-step processes of the type depicted in Fig. 1(a). Bound and continuum single-particle states are taken to be eigenstates of the HF mean-field potential obtained with an effective interaction of the Skyrme type (SKE2) [@war]. I n this way, we preserve the orthogonality between the bound and the continuum states.
The nuclear current in the matrix element of eq. (\[matr\]) is taken to be the sum of a one-body operator and a two-body operator. The nucleonic one-body term consists of the well-known convection and magnetization current. The two-body current is taken from a non-relativistic reduction (retaining only terms up to the order $1/M^{2}$ in the nucleon mass) of the lowest order Feynman diagrams with one exchanged pion and intermediate delta excitation. We assume pseudovector coupling of the pion to the nucleon. This procedure gives rise to the seagull terms (Fig. 1(b)), the pion-in-flight term (Fig. 1(c)) and terms with a $\Delta (1232)$ excitation in the intermediate state (Fig. 1(d)). In this non-relativistic approach the nuclear charge operator is not affected by two-body contributions. The explicit expressions for the two-body current in momentum space can be found in ref. [@riska].
To account for the composite structure of the $\gamma N$, $\gamma \pi$ and $\gamma \Delta$ vertices, electromagnetic form factors have to be introduced. For the $\gamma N$ form factor we use the common dipole form [@gal]. Current conservation with the one-pion exchange potential is me rely satisfied for the seagull and the pion-in-flight current in case that the pion ($f_{\gamma\pi}$) and nucleon ($f_{\gamma N}$) form factor coincide. In all further calculations we h ave adopted the $f_{\gamma\pi}$ extracted from the vector dominance model [@eric]. The delta current is divergenceless and can be multiplied with an arbitrary form factor without violating the charge-current conservation rules. For simplicity, we assumed that $f_{\gamma\Delta}=f_{\gamma N}$ in all calculations presented here. The short-range structure of the $\pi NN$ and $\pi N
\Delta$ vertices is implemented in a phenomenological way by introducing hadronic form factors. A s is usually done the monopole form is adopted. For both types of vertices the same pion cut-off mass $\Lambda_{\pi}(=\Lambda_{\pi NN}=\Lambda_{\pi \Delta N}$) is used.
The final state wave function in the matrix element of eq. (\[matr\]) is evaluated using a multipole expansion in t erms of distorted waves. So, the nuclear current operator is decomposed in the well-known electric and m agnetic transition operators $T^{el}_{JM}$ and $T^{mag}_{JM}$ [@for]. If we restrict ourselves to the evaluation of diagrams of the type depicted in Fig. 1, reduced matrix elements of the following type $$\begin{aligned}
\label{trans}
<0^{+}\mid\mid T_{J}^{(1)}(q)+T_{J}^{(2)}(q)\mid\mid (ph^{-1});J>\end{aligned}$$ remain to be calculated. In this expression $T_{J}^{(1)}$ and $T_{J}^{(2)}$ refer to the IA and the pionic contribution. The two-body part of the transition operators is handled exactly and involves tw o active nucleons in the absorption process. Hence, in the evaluation of (\[trans\]), the two-body part has been expressed in terms of two-body matrix elements. The explicit expression for these two-body matrix elements for the diagrams of Fig. 1 can be found in ref. [@jan3].
All results presented here are obtained under the kinematical conditions of the $^{16}O(e,e'p)$ [@chiara][@wies] and $^{40}Ca(e,e'p)$ [@wies][@kramer] NIKHEF -K experiments. These expe riments were performed under perpendicular kinematics at ($\omega=96$ MeV, $q=460$ MeV) ($^{16}O$) and ($\omega=116$ MeV, $q=446$ MeV) ($^{40}Ca$). The structure functions are plotted as a function of the missing momentum $p_{m}=\mid \vec{p}_{a}-\vec{q} \mid$.
Some theoretical uncertainties exist with respect to the pion cut-off mass $\Lambda_{\pi}$ in the hadronic form factor and to the electromagnetic pion form factor. Before confronting our theoretical approach with the data, we investigate the sensitivity of the results to these parameters. Whereas the Bonn potential leads to a cut-off mass $\Lambda_{\pi} = 1200$ MeV [@mach], recent studies on the triton system [@sasa] seem to prefer a smaller value ($\Lambda_{\pi} = 810$ MeV). We have performed calculations including all diagrams depicted in Fig. 1 with different values of $\Lambda_{\pi}$. As a representative example we display in Fig. 2(a) the longitudinal-transverse interference structure function $W_{LT}$ for proton knockout out of the $1p_{3/2}$ orbit in $^{16}O(e,e'p)$. Results are displayed for a pion cut-off mass that should be considered as a lower limit ($\Lambda_{\pi} = 650$ MeV) and an upper limit ($
\Lambda_{\pi} = 1200$ MeV). Also shown is the $W_{LT}$ structure function as obtained within the IA . The uncertainty of the results due to the theoretical ambiguity in $\Lambda_{\pi}$ should be estimated around 20% of the total two-body contribution. In all furth er calculations we used a cut-off mass $\Lambda_{\pi}=800$ MeV . With this lower cut-off value heavier mesons such as the $\rho$-meson are partially taken into account [@riska].
The sensitivity of the present approach to the $f_{\gamma\pi}$ form factor in the pion-in-flight current is investigated in Fig. 2(b). In the forthcoming discussion it will be shown that, of all structure functions, the $W_{T}$ exhibits the largest sensitivity to the pion-in -flight term. We have plotted $W_{T}$ for proton knockout out of the $1p_{3/2}$ orbit in $^{16}O$. The results including the full nuclear current are compared with the IA predictions. Two different $f_{\gamma\pi}$ form factors are considered in the calculations. Firstly, the pion form factor is set equal to the nucleon form factor and, as a consequence, current conservation is satisfied. Secondly, we adopt the $f_{\gamma\pi}$ form factor as derived from the vector dominance model. Clearly, the results are rather insensitive to the choice of the pion form factor.
The impact of the different components in the nuclear current on the structure functions in the $^{16}O(\vec{e},e'p)$ reaction is illustrated in Figs. 3 and 4. In Fig. 3 we display the transverse and interference structure functions $W_{T
}$ and $W_{LT}$ for $^{16}O(e,e'p)^{15}N(1p_{1/2}^{-1},g.s.)$ and $^{16}O(e,e'p)^{15}N(1p_{3/2}^{-1},6.32 MeV)$. All curves are multiplied with a spectroscopic factor extracted from a best fit of the calculated cross section to the data. In conformity with the results of refs. [@rep][@wies] we do not arrive at a [*simultaneous*]{} descript ion of the total cross section and the $W_{LT}$ structure function in the IA. Furthermore, a spin-orbit dependence of the results is observed: whereas the IA seems to work reasonably well for the $1p_{1/2}$ state, the absolute value of $W_{LT}$ is severely underestimated for the $1p_{3/2}$ orbit. From Fig. 3 it is clear that for both single-particle states the seagull and pion-in-flight current exhibit the same characteristics. Whereas the seagull contribution enhances the transverse and charge-current interference response functions, the pion-in-flight current has the opposite effect. Generally, the pion-in-flight term has a smaller influence on the structure functions than the seagull current. The spin-dependent behaviour of the $W_{LT}$ term for the two states originates from the contribution of the $\Delta(1232)$ current. For the $1p_{1/2}$ state a further quenching is observed in contrast with the $1p_{3/2}$ results where the deviation from the IA approach becomes more pronounced. As can be noticed, the discrepancy between experiment and the DWIA results for $W_{LT}$ of the $1p_{3/2}$ state can be partially ascribed to the two-body contributions in the nuclear current.
In contrast to the cross section for which the effect of the pionic currents is hardly visible, two-body contributions in the nuclear current cannot be discarded in order to obtain a complete description of the structure functions.
In Fig. 4 we display the fifth structure function $W'_{LT}$ as can be extracted from the polarized $^{16}O(\vec{e},e'p)$ reaction for the same orbits. The calcu lations are performed using the same kinematics as before. To our knowledge, no data of this response function are available as yet. In the near future a large number of polarization experiments will be carried out at MIT-Bates [@pap]. From a theoretical standpoint, this additional measurable quantity $W_{LT}'$ has as its main advantage that it vanishes identic ally in the plane wave impulse approximation (the escaping proton is described by a plane wave) [@ras] . So, the shape of the calculated $W_{LT}'$ term is [*completely determined*]{} by the fin al-state interaction and the different contributions in the nuclear current. In comparison with the results for the $W_{LT}$ term, we observe a similar spin-dependent behaviour with respect to the two-body nuclear current. For both single-particle orbits, the shape remains almost unaffected compared to the DWIA calculation. Whereas a small quenching is observed for the $1p_{1/2}$ state, pionic contributions seem to enhance the fifth structure function for the $1p_{3/2}$ state. Generally, the effect of the two-body contributions is more apparent in the $W_{LT}$ term than in the corresponding $W_{LT}'$ structure function.
Until now, all conclusions drawn referred to $^{16}O$ results. Similar calculations have been performed for proton ejection from the target nucleus $^{40}Ca$. We have calculated cross sections and structure functions for the spin-orbit partners $1d_{3/2}$ and $1d_{5/2}$. The results for the interference structure function $W_{LT}$ are displayed in Fig. 5. Once again, meson contributions in the current are not negligible and can accoun t for part of the discrepancy between the IA calculations and the data. Due to the spin-dependent behaviour of the $\Delta$ current, the $W_{LT}$ function for the $1d_{3/2}$ state and the $1d_{5/2}$ state is modif ied in a different way. This conclusion is in conformity with the results for the spin-orbit partners $1
p_{1/2}$ and $1p_{3/2}$ in $^{16}O$.
Summarizing, the results presented here indicate the importance of two-body contributions in the nuclear current in order to reach a complete descr iption of the $(\vec{e},e'p)$ cross section and structure functions under quasifree con ditions. Going beyond the impulse approximation, we have accounted for two-body contributions in the current related to one-pion exchange and intermediate $\De
lta(1232)$ creation. Calculations were performed for proton knockout from the target nuclei $^{16}O$ and $^{40}Ca$. The results were shown to be rather insensitive to the model assumptions with respect to the pion and hadronic form factor. The charge-current interference structure function $W_{LT}$ is found to be strongly affected by the two-body cur rents. A rather good agreement with the data could be obtained within a model that accounts for the FSI within an HF-RPA model and in which one and two-body photoabsorption mechanisms are included. Rather than the cross section the separated structure functions are sensitive to the different aspects of the reaction mechanism. In this sense, a further exploration of the separate structure functions opens good pers pectives to obtain a better insight into the nature of the $(e,e'p)$ reaction mechanism.
[**Acknowledgement**]{}
The authors are grateful to K.Heyde for fruitful discussions and suggestions. This work has been supported by the Inter-University Institute for Nuclear Sciences (IIKW) and the National Fund for Scientific Research (NFWO).
[99]{} S.Frullani and J.Mougey, Adv.Nucl.Phys.[**14**]{}(1984)1 Y.Jin, K.Zhang, D.S.Onley and L.E.Wright, Phys.Rev.[**C47**]{}(1993)2 024 A.Picklesimer, J.W.Van Orden and S.J. Wallace, Phys. Rev.[**C32**]{}(1985)1312 S.Boffi, C.Giusti and F.D.Pacati, Phys.Rep.[**226**]{}(1993)1 J.Ryckebusch, K.Heyde, D.Van Neck and M. Waroquier, Phys.Lett.[**B216**]{}(1988)252 M.Bernheim [*et al.*]{}, Nucl.Phys.[**A375**]{}(1982)381 J.Ryckebusch, M.Waroquier, K.Heyde, J.Moreau and D.Ryckbosch,\
Nucl.Phys.[**A476**]{}(1988)237 A.S.Raskin and T.W.Donnelly, Ann. Phys. [**191**]{}(1989)78 A.Picklesimer and J.W.Van Orden, Phys.Rev.[**C40**]{}(1989)290 L.Chinitz [*et al.*]{}, Phys.Rev.Lett [**67**]{}(1991)568 Annual report 1992, NIKHEF (Amsterdam) L.Lapikás, Nucl.Phys.[**A553**]{}(1993)297c H.Blok, private communication M.Waroquier, J.Ryckebusch, J.Moreau, K.Heyde, N.Blasi, S.Y. van der Werf and G.Wenes, Phys.Rep.[**148**]{}(1987)249 D.O.Riska, Phys.Rep.[**181**]{}(1989)207 S.Galster [*et al.*]{}, Nucl.Phys.[**B32**]{}(1971)221 T.Ericson and W.Weise, in [*Pions and Nuclei*]{} (Oxford Science Publications, Oxford,1988) T.deForest and J.D.Walecka, Adv. in Phys.[**15**]{}(1966)1 J.Ryckebusch, L.Machenil, M.Vanderhaeghen and M. Waroquier, to be published R.Machleidt, K.Holinde and Ch. Elster, Phys.Rep.[**149**]{}(1987)1 T. Sasakawa, S. Ishikawa, Y. Wu and T-Y. Saito, Phys. Rev. Lett. [**68**]{} (1992) 3503. C.Papanicolas, talk given at the 6th Workshop on Perspectives in Nuclear Physics at Intermediate Energies, Trieste 1993, eds. S.Boffi,C.Ciofi degli Atti and M.M.Giannini, to be published
[^1]: Postdoctoral research fellow NFWO
[^2]: Research Director NFWO
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We propose a degree–based coarse graining approach that not just accelerates the evaluation of dynamics on complex networks, but also satisfies the consistency conditions for both equilibrium statistical distributions and nonequilibrium dynamical flows. For the Ising model and Susceptible-Infected-Susceptible epidemic model, we introduce these required conditions explicitly and further prove that they are satisfied by our coarse-grained network construction within the annealed network approximation. Finally, we numerically show that the phase transitions and fluctuations on the coarse-grained network are all in good agreements with those on the original one.'
author:
- Hanshuang Chen
- Zhonghuai Hou
- Houwen Xin
- YiJing Yan
title: 'Statistically consistent coarse-grained simulations for critical phenomena in complex networks'
---
Introduction {#sec1}
============
Complex network has been one of the most active research topics in statistical physics and many other disciplines [@RMP02000047; @AIP02001079; @SIR03000167; @PRP06000175; @PRP08000093]. It describes not only the pattern discovered ubiquitously in real world but also a unified theoretical framework to understand the inherent complexity in nature. However, the investigation of large networks, such as human brain that composes of about $10^{11}$ neurons and $10^{14}$ synapses [@RMP06001213], requires tremendous time-demanding efforts. The phenomenological description may capture certain properties of system, but always neglects microscopic information. A promising alternative is to develop coarse-grained (CG) methods, aiming at significant reducing the degree of freedom while proper preserving the microscopic information of interest.
Several CG schemes have been proposed. Renormalization transformation has been used to simplify self–similar networks, and the reduced networks often preserve some topological properties of the original ones [@PRL04016701; @NTR05000392; @PRL06018701; @PRL08148701]. Spectral coarse graining technique has been proposed, in which the eigenvalues of Laplace matrix of network are almost unchanged, such that the dynamics of random walk and synchronization are preserved [@PRL07038701; @PRL08174104]. Equation–free multiscale computational framework has been developed to accelerate simulation using a coarse time stepper [@CMS03000715]. This approach has been applied to study the CG dynamics of oscillators network [@PRL06144101], gene regulatory network [@JCP06084106], and adaptive epidemic network [@EPL08038004]. However, to our best knowledge, no attempt has been made for developing CG simulation method to study critical phenomena (usually described by stochastic models) on complex networks. The size–dependent and scaling behaviors in these systems are studied so far mainly by such as Monte Carlo (MC) and kinetic MC (KMC) simulations [@RMP08001275; @AIP00000815]. Apparently, these microscopic approaches are often too expensive. It is noticed that the CG stochastic models have been proposed to study reaction–diffusion processes on regular lattices [@JCP03000250]. However, the existing methods are largely inapplicable to critical phenomena on complex networks with diversified heterogeneity. Moreover, the crucial issue concerning the methodology development as to what criteria should be met to make the CG model statistically consistent with the microscopic one is yet to be addressed.
In this paper, we propose the degree-based CG ([$d$-CG ]{}for short hereafter) be a statistically consistent scheme, within the annealed network approximation (ANA) [@RMP08001275; @PRE03036112; @PRL02258702]. It may therefore be an efficient and reliable CG method for evaluating the stationary properties of phase transitions and studying size effect on complex networks. We put forward the conditions of statistical consistency (CSC) on both equilibrium and nonequilibrium properties, exemplified with the Ising model and Susceptible–Infected–Susceptible (SIS) model, respectively. We show that the [$d$-CG ]{}approach that merges together the nodes of similar degrees does warrant the CSC within ANA. The stochastic Ising spin–flip and epidemic spreading dynamics can therefore be evaluated faithfully and efficiently with the [$d$-CG ]{}networks. The calculated phase diagrams, fluctuation dynamics, and system size–scaling behaviors are all shown in excellent agreements with the corresponding microscopic MC and KMC results.
The rest of this paper is organized as follows. In Sec.\[sec2\], we present a general scheme for coarse graining network and [$d$-CG ]{}approach, and further prove that this approach satisfies statistical consistency. Extensive numerical demonstrations of the CG approach are performed on diverse networks in Sec.\[sec3\]. At last, main conclusions and discussion are addressed in Sec.\[sec4\].
Coarse Graining Procedure {#sec2}
=========================
Network Coarse Graining {#sec2.1}
-----------------------
Let us start with the basic ingredients of network coarse graining. Consider a network consisted of $N$ nodes whose connectivity is given by the adjacency matrix ${\bm A}$, in which $A_{ij}= 1$ if nodes $i$ and $j$ are connected and $A_{ij}=0$ otherwise. Merging $q_\mu$ nodes together into a CG-node (denoted by $C_\mu$) leads to a CG–network with $N^c$ CG-nodes. We adopt the mean–field definition of the CG connectivity between $C_\mu$ and $C_\nu$, as the average number of links connecting any two nodes inside $C_\mu$ and $C_\nu$. The adjacency matrix ${\bm A}^c$ of the CG network is then $$A_{\mu\nu}^c =
\begin{cases}
\frac{2}{q_\mu(q_\mu-1)}\sum\limits_{i,j \in C_\mu; i<j} A_{ij}
& \text{if $\mu = \nu$}, \\
\frac{1}{q_\mu q_\nu}\sum\limits_{i \in C_\mu,j \in C_\nu} A_{ij}
& \text{if $\mu \ne \nu$}.
\end{cases}
\label{eq1}$$ For illustration, Fig.\[fig1\] depicts an example of coarse graining a network with six nodes into a weighted CG network with three CG-nodes.
![(color online) (a) A schematic example of coarse graining network in which nodes of the same shape are merged; (b) the adjacency matrix ${\bm A}^c$ of CG network by Eq.(\[eq1\]). Note that the CG network may include self–connections; for example, the CG-node III where the connected nodes are merged. \[fig1\]](fig1.eps){width="0.8\columnwidth"}
CSC {#sec2.2}
---
We address the issue on the statistical consistency of a CG scheme with the microscopic network, in terms of both the equilibrium distribution and the nonequilibrium flow. We exploit the Ising model and the SIS model as the paradigmatic examples for equilibrium and nonequilibrium systems, respectively.
For the Ising model defined on a network, the Hamiltonian is given by $H=-J\sum\nolimits_{i < j}{A_{ij}s_i s_j}$, with the spin variable $s_i=\pm 1$, and the ferromagnetic interaction parameter $J>0$. The probability of a given microscopic spin configuration $\{s_i\}$ is given by canonical distribution $e^{-\beta H}/Z$, where $\beta$ is the inverse temperature and $Z$ is the partition function. The corresponding CG Hamiltonian assumes the sum of pairwise interactions inside and between the CG-nodes, i.e. $H^c=H_1^c+H_2^c$, $$\begin{aligned}
H_1^c
&= -\frac{J}{2}\sum\limits_\mu {A_{\mu\mu}^c}
\big[n^+_\mu(n^+_\mu - 1) + n^-_\mu(n^-_\mu - 1) - 2n^+_\mu n^-_\mu\big]
\nonumber \\
&= -\frac{J}{2}\sum\limits_\mu {A_{\mu\mu}^c}
\big(\eta^2_\mu - q_\mu \big),
\label{eq2} \\
H_2^c
&= -J\sum\limits_{\mu<\nu} A^c_{\mu\nu}
\big(n^+_\mu n^+_\nu + n^-_\mu n^-_\nu - n^+_\mu n^-_\nu - n^-_\mu n^+_\nu\big)
\nonumber \\
&= - J\sum\limits_{\mu<\nu} A^c_{\mu\nu} \eta_\mu \eta_\nu . \label{eq3}\end{aligned}$$ Here, $\eta_\mu=\sum\nolimits_{i \in C_\mu }{s_i}$ is the CG spin variable, and $n_\mu^\pm={{\left( {q_\mu
\pm \eta_\mu }\right)}\mathord{\left/ {\vphantom{{\left( {q_\mu \pm \eta_\mu}\right)}2}}\right.
\kern-\nulldelimiterspace}2}$ is the number of up/down spins inside $C_\mu$. Note that $H^c=H^c_1+H^c_2$ above is a closure expression at the CG level that depends only on ${\bm A}^c$ and $\{\eta_\mu\}$. To make the CG model consistent with the microscopic one, it demands that the probability of any given CG configuration $\{\eta_\mu\}$ in equilibrium be the sum of the probabilities of all microscopic configurations that contribute to it. That is $$g(\{\eta _\mu\}) e^{-\beta H^c}
= \sum_{\{s_i\}} \prod_\mu
\delta\Big(\eta_\mu - \sum_{i \in C_\mu }\!\!s_i\Big) e^{-\beta H} ,
\label{eq4}$$ where $g(\{\eta_\mu\})= \prod_\mu {q_\mu}! \big/ ({n^+_\mu}! {n^-_\mu}!)$ is the number of microscopic configurations corresponding to $\{\eta_\mu\}$.
We now consider the nonequilibrium scenario, exemplified by the SIS network for epidemic spreading dynamics. At the microscopic level, the SIS network nodes represent individuals being either susceptible ($\sigma=0$) or infected ($\sigma=1$). A susceptible individual $j$ can get infected at rate $\tilde r_j=\sum\nolimits_i f_{i \to j}$, where $f_{i\to j}=\lambda A_{ij}\sigma_i(1-\sigma_j)$ is the spreading flow from an infected individual $i$ to $j$. On the other hand, an infected node can recover to susceptible state at rate $r_i=\gamma\sigma_i$. Without loss of generality, we set $\gamma=1$ hereafter to scale the infection parameter $\lambda$. The SIS model exhibits a nonequilibrium dynamical phase transition at $\lambda=\lambda_c$ from an absorbing state (all recovered) to an active state (disease spreading persistently) [@AIP00000815; @JRS05000295]. For the CG–SIS network, we define the CG variables as the number of infected individuals inside a CG-node $C_\mu$ by $\sigma^c_\mu= \sum_{i \in C_\mu}{\sigma_i}$. Written in a closure form, the CG recovery rate is $r^c_\mu =\sigma^c_\mu$, and the CG spreading flow between two CG-nodes is $f^c_{\mu \to \nu}=\lambda A^c_{\mu\nu}\sigma^c_\mu ({q_\nu-\sigma^c_\nu})$. For this nonequilibrium system, the CG model is consistent with the microscopic one if the CG flow matches the microscopic flows: $$f_{\mu \to \nu}^c = \sum_{i \in C_\mu, j \in C_\nu} f_{i \to j} \label{eq5}$$ Note that the CG recovery rate of $r^c_\mu = \sum_{i \in C_\mu} {r_i}$ holds trivially.
Degree-based CG Scheme {#sec2.3}
----------------------
In our approach, we merge the nodes with similar degrees together. We shall show the resulting [$d$-CG ]{}scheme does satisfy the CSC, as defined by Eq.(\[eq4\]) for the Ising model and Eq.(\[eq5\]) for the SIS model, within the ANA for the ensemble averaged dynamics [@RMP08001275; @PRE03036112; @PRL02258702]. In many previous studies [@PLA02000166; @PRE02035108], ANA has been extensively confirmed to be a useful tool for describing quenched networks, as in case of the present paper. Although ANA is just an approximation, it still gives a reasonable description of the average behavior of nodes of the same degree. Moreover, one can consider that the ANA is a statistical characterization of large number of quenched networks with the same degree distributions. Nevertheless, in Ref.[@PRE09051127] it has been pointed out that there exists some discrepancies between considering annealed approximation for quenched networks and considering annealed network models by themselves. According to the ANA, one can replace the dynamics on a given network of $N$ nodes by that on a weighted graph of the full connectivity $A_{ij}=k_i k_j/(N{\langle k \rangle})$, where $k_i$ and $k_j$ are the degrees of nodes $i$ and $j$, respectively, and $\left\langle k \right\rangle$ is the mean degree. In an ideal [$d$-CG ]{}scheme, the microscopic nodes in a single CG-node are of the same degree: $k_i|_{i \in C_\mu}= K_\mu$. The CG connectivity is then $A^c_{\mu\nu}=\frac{K_\mu K_\nu}{\langle
k \rangle N}$, for the CG–network dynamics treated at the ANA level. As results, Eqs.(\[eq2\]) and (\[eq3\]) become, respectively (noting that $s^2_j = 1$) $$\begin{aligned}
H_1^c &= -\frac{J}{2N \langle k\rangle}
\sum_\mu K_\mu^2 \Big[\sum_{i,j \in C_\mu} s_i s_j -q_\mu \Big]
\nonumber \\
&= -\frac{J}{N \langle k \rangle}
\sum_\mu {K_\mu^2} \sum_{i,j \in C_\mu; i < j} s_i s_j ,
\label{eq6}\\
H_2^c &= - \frac{J}{N \langle k \rangle}
\sum_{\mu<\nu} K_\mu K_\nu
\sum_{i\in C_\mu, j\in C_\nu} s_i s_j .
\label{eq7}\end{aligned}$$ Their sum can be written as $$H^c = - \frac{J}{N\langle k \rangle} \sum_{\mu,\nu} K_\mu K_\nu
\sum_{i\in C_\mu, j\in C_\nu; i < j} s_i s_j .
\label{eq8}$$ It is identical to the microscopic Hamiltonian at the ANA level. In other words, the Hamiltonian of any [$d$-CG ]{}configuration equals to the collection of its contributing microscopic configurations. The pre-exponential terms in two sides of Eq.(\[eq4\]) are both the degeneracy of CG configuration, as well as the identical energy factors ($H^c=H$). We have thus proved that the ideal [$d$-CG ]{}approach to the Ising model obeys the CSC of Eq.(\[eq4\]) exactly. It is also easy to prove that the nonequilibrium CSC of Eq.(\[eq5\]) is true for the SIS model in consideration; both sides there equal to $\lambda\frac{K_\mu K_\nu}{N\langle k
\rangle}{\sigma^c_\mu}(q_\nu - \sigma^c_\nu)$. Therefore, the [$d$-CG ]{}approach satisfies the CSC for both equilibrium probability distributions and nonequilibrium dynamical flows. Certainly it is anticipated that the CSC holds approximately if merged together are the nodes of similar degrees rather than the exactly same ones.
Numerical Demonstrations {#sec3}
========================
CG-MC and CG-KMC simulations {#sec3.1}
----------------------------
The MC simulation with the Metropolis dynamics [@Lan2000] and the KMC simulation [@JCP77002340] are applied to the Ising model and SIS model, respectively, at both the microscopic and the CG levels. In the CG-MC simulation, a CG-node $C_\mu$ is randomly chosen, followed by the Metropolis try for spin-flip process. The probabilities of flipping a up/down–spin is $n^{\pm}_\mu W(\beta,
\Delta E_{\uparrow/\downarrow})$, where $W(\beta, \Delta
E)=\min(e^{-\beta \Delta E}, 1)$, and $\Delta
E_{\uparrow/\downarrow}$ is the change of CG Hamiltonian resulting from the flip of a up/down spin. It is easy to confirm that the CG-MC simulation obeys the detailed balance condition. In the CG-KMC simulation, following the Gillespie algorithm [@JCP77002340], a CG process to be executed is randomly selected based on the transition rates of all processes. The configuration and transition rates are updated for executing the next CG process.
![Typical time evolutions of the magnetization $m$ in Ising model at $T=10$ (in unit of $J/k_B$) and the density of infected nodes $\rho$ in SIS model at $\lambda=0.1$ for both microscopic and CG levels. Other parameters are $N=1024$ and $N^c=16$. \[fig2\]](fig2.eps){width="0.8\columnwidth"}
Scale-free networks {#sec3.2}
-------------------
We first consider the Barabási–Albert (BA) scale-free network [@SCI99000509], with the degree distribution follows a power-law $P(k)\sim k^{-\gamma}$ with scaling exponent $\gamma=3$. Figure \[fig2\] plots typical time evolutions of the magnetization $m=\sum\nolimits_i {{s_i}}/N$ in Ising model at $T=10$ (in unit of $J/k_B$) and the density of infected nodes $\rho=\sum\nolimits_i
{{\sigma_i}}/N$ in SIS model at $\lambda=0.1$, where $N=1024$ and $N^c=16$ are used. For both the microscopic and CG simulations, the systems attain the steady states associated with fluctuating noise after transient time. It is clear that there are in good agreement in the steady-state values of $m$ and $\rho$, as well as their fluctuating amplitudes for both simulations cases.
![(a)-(b) $|m|$ and $\chi$ as a function of $T$. Symbols and lines correspond to results by simulating the CG model and by direct microscopic MC simulation, respectively. (c) $T_c$ as a function of $\ln N$ obtained by the CG model with $N^c=64$. All dotted lines are the results of randomly merging models with $N=16384$ and $N^c=64$. All the networks have fixed mean degree $\langle k \rangle=20.$ The error bars (not shown) are smaller than the symbol size. \[fig3\]](fig3.eps){width="0.8\columnwidth"}
![(a)-(b) $\rho$ and and $\Delta \rho^2$ as a function of $\lambda$. Symbols and lines correspond to results by simulating the CG model and by direct microscopic KMC simulation, respectively. (c) $\lambda_c$ as a function of $1/\ln N$ obtained by the CG model with $N^c=64$. All dotted line are the results of randomly merging models with $N=16384$ and $N^c=64$. All networks have fixed mean degree $\langle k \rangle =20$. The error bars (not shown) are smaller than the symbol size. \[fig4\]](fig4.eps){width="0.8\columnwidth"}
For the Ising model, we construct initial configurations by preparing each node with a random spin value $s_i=+1$ or $-1$ with an equal probability. As the simulation proceeds, the system quickly relaxes to an equilibrium state. With the same temperature $T$ we run at least $100$ times of simulations corresponding to different initial configurations and network realizations. In each simulation, $2\times10^3$ MC steps (MCS: each spin is attempted to flip once on average during each MCS) are performed and the last $10^3$ MCS are used to investigate the system’s behavior. As $T$ decreases the value of the magnetization $m$ undergoes a transition from zero to nonzero at the critical temperature $T_c$. Below $T_c$ we notice that due to finite-size effects the system can switch between two stable states via a nucleation mechanism, resulting in the oscillations of $m$ [@PHA02000260]. Above $T_c$, $m$ fluctuates around zero, and the susceptibility $\chi$ per node has a maximum at phase transition, which can be used to determine $T_c$ as we shall show in Fig.\[fig3\](b). The susceptibility is related to the magnetization fluctuation via the fluctuation–dissipation theorem. To avoid the offset of $m$ due to its oscillations in simulations, we will use instead the absolute value of $m$ in the following. $|m|$ and $\chi$ are shown in Fig.\[fig3\](a) and (b), respectively, as a function of the temperature $T$ for different $N$ and $N^c$. Apparently, the [$d$-CG ]{}results (reported by symbols) are in excellent agreements with the microscopic–level counterparts (by solid and dashed lines). As contrast, we also report the results of random-merging (RM) CG scheme (by dotted lines) in Fig.\[fig3\](a) and (b), where each CG–node includes $q_\mu\equiv N/N^c$ ($N=1024$, $N^c=64$) nodes selected randomly from the whole network. Evidently the random scheme fails badly in reproducing the microscopic behaviors. We also used the same $q_\mu$ in the random scheme as in the case of [$d$-CG ]{}scheme, and found that the two random schemes produce the same results.
It is important to merge nodes with similar degrees together, as already shown both analytically and numerically. Strikingly, even when the original network is reduced to one with only $16$ CG-nodes, the [$d$-CG ]{}scheme still faithfully reproduces the phase transition curves and fluctuations properties. Since $N^c$ is largely reduced compared to $N$, a considerable speed–up of CPU time, about a factor of 40, is realized for $N=16384$. A significantly higher gain can be expected as the network gets larger, allowing the computational study of network size effect very affordable.
Using the [$d$-CG ]{}approach with a fixed size of CG–networks $N^c=64$, we calculate the dependence of phase–transition critical $T_c$ on the network size $N$, as reported in Fig.\[fig3\](c). It had been well established that the Ising model with ferromagnetic interactions on BA scale-free network undergoes a phase transition from ferromagnetism to paramagnetism at a critical temperature $T_c$ that increases as the logarithm of network size [@PHA02000260; @PLA02000166; @EPB02000191; @PRE02016104]. Our result is consistent with the theoretical expression [@PLA02000166] that ${T_c}=\frac{\langle k \rangle}{4}\ln N$, where ${\left\langle
k \right\rangle }=20$ for the present case of study. Note that the critical phenomenon disappears when $N \to \infty$.
For the SIS model there is an epidemic threshold $\lambda _c$ on a finite size BA scale-free network network [@PRL01003200; @PRE02035108]. Our numerical simulation starts from a random configuration with about half nodes being infected. After an initial transient regime, the system will evolve into a steady state with a constant average density of infected nodes. The steady density of infected nodes $\rho$ is computed by averaging over at least $50$ different initial configurations and at least $10$ different network realizations with the same parameter $\lambda$. The epidemic threshold $\lambda _c$ occurs at $\rho=0$ (absorbing state) if $\lambda<\lambda _c$ and $\rho>0$ (active state) if $\lambda>\lambda _c$ [@PRL01003200]. Due to finite size effects, the fluctuation can drive the system to the absorbing state, especially in the vicinity of $\lambda_c$. Once the absorbing state is arrived, the system will never leave it. Based on the consideration, we use, in practice, a nonzero tolerance in $\rho$ (with the order of $N^{ - 1}$) as the boundary of the phase transition point. Reported in Fig.\[fig4\](a) and (b) are the calculated results of $\rho$ and its fluctuation $\Delta {\rho ^2} =
\left\langle {{\rho ^2}} \right\rangle - {\left\langle \rho
\right\rangle ^2}$, respectively, as a function of $\lambda$, obtained by the CG model and the microscopic model. Again, while the RM scheme fails when the relative infection rate parameter $\lambda<0.1$, the agreement between the [$d$-CG ]{}and the microscopic results remains excellent. Fixing $N^c=64$ in the [$d$-CG ]{}approach, the resulting $\lambda_c$ shown in Fig.\[fig4\](c) is proportional to $1/\ln N$, also consistent with the theoretical prediction [@PRE02035108].
Figure \[fig5\] show that the simulation results of the microscopic and CG levels on scale-free networks with other two scaling exponents, $\gamma=2$ and $\gamma=3.5$. For larger $\gamma$, $T_c$ becomes lower, while $\lambda_c$ gets larger. It is clearly observed that our [$d$-CG ]{}approach is still applicable. In addition, many other types of networks such as random network and small-world network are also tested, and all results show that the validity of our [$d$-CG ]{}approach does not depend on network topology.
![Comparison of microscopic and CG simulations on scale-free networks with the scaling exponent $\gamma=2.0$ and $\gamma=3.5$. Left panel: $|m|$ and $\chi$ as a function of $T$ for Ising model. Right panel: $\rho$ and and $\Delta \rho^2$ as a function of $\lambda$ for SIS model. All networks have fixed size of $N=1024$ and mean degree of $\langle k \rangle =20$. \[fig5\]](fig5.eps){width="1.0\columnwidth"}
Degree correlated networks {#sec3.3}
--------------------------
It is worthy noting that the above numerical demonstrations are carried out on degree uncorrelated networks. We will show that our [$d$-CG ]{}approach is valid to reproduce critical behaviors on degree–degree correlated networks as well. It has been witnessed that many real networks display different degree–mixing patterns [@PRL02208701]. To measure the degree of the correlation, in Ref.[@PRL02208701] Newman introduced a degree-mixing coefficient: $r_k = ( {\left\langle {k_i k_j } \right\rangle -
\left\langle {k_i } \right\rangle \left\langle {k_j } \right\rangle
} )/( {\left\langle {k_i^2 } \right\rangle - \left\langle {k_i }
\right\rangle ^2 } )$, where $k_i$ and $k_j$ are the remaining degrees at the two ends of a link and $\left\langle \bullet
\right\rangle$ means the average over all links. $r_k=0$ indicates that there is no degree correlation, while $r_k>0$ ($<0$) indicates that a network is assortatively (disassortatively) mixed by degree. Previous studies have revealed that degree–mixing pattern plays an important role in dynamical behaviors on networks, such as percolation [@PRL02208701], epidemic spreading [@LNP03000127], synchronization [@PRE06066107]. To generate different degree–mixing networks, we employ a algorithm proposed in [@PRE04066102]. At each elementary step, two links in a given network with four different nodes are randomly selected. To get an assortative network, the links are rewired in such a way that one link connects the two nodes with the smaller degrees and the other connects the two nodes with the larger degrees. Multiple connections are forbidden in this process. Repeat this operation until an assortative network is generated without changing the node degrees of the original network. Similarly, a disassortative network can be produced with the rewiring operation in the mirror method. We start from BA scale-free networks with a neutrally degree-mixing pattern, and produce some groups of degree-mixing networks by performing the above algorithm. Figure \[fig6\] displays the results of Ising model and SIS model for three different values of $r_k$. For each $r_k$, the simulation results of CG models agree well with those of microscopic ones. In Ising model, $T_c$ shifts to right and $\chi$ at $T_c$ becomes smaller as $r_k$ increases. In SIS model, both $\lambda_c$ and the fluctuation of $\rho$ at $\lambda_c$ decrease with $r_k$.
![Comparison of microscopic and CG simulations on correlated networks for different degree-mixing coefficient $r_k$. Left panel: $|m|$ and $\chi$ as a function of $T$ for Ising model. Right panel: $\rho$ and and $\Delta \rho^2$ as a function of $\lambda$ for SIS model. All networks have power-law degree distributions with the scaling exponent $\gamma=3$, and fixed size of $N=1024$ and mean degree of $\langle k \rangle =20$. \[fig6\]](fig6.eps){width="1.0\columnwidth"}
Conclusions and Discussion {#sec4}
==========================
In summary, we propose an approach for coarse graining the phase transition dynamics on complex networks described by stochastic models for both equilibrium and nonequilibrium systems. The [$d$-CG ]{}approaches via degree-based merging scheme are feasible since the reliable microscopic information such as phase transition behaviors and fluctuations are preserved. We have verified that our [$d$-CG ]{}approach supports, in consistent with the microscopic models, the equilibrium distributions for Ising model and the nonequilibrium dynamical flows for SIS model. Stochastic description, as exemplified here, is ubiquitously important in the study of phase transition dynamics and complex networks for a wide range of realistic systems. Moreover, this work also suggests the development of other promising CG statistical models satisfying CSC.
It is interesting to compare our [$d$-CG ]{}approach with heterogeneous mean-field theory (HMFT) that successfully predicts $T_c$ and $\lambda_c$ on heterogeneous scale-free networks [@PRE02016104; @PRL01003200]. Based on the ansatz that nodes with the same degree share the same dynamical properties, HMFT derives a series of coupled mean-field equations for degree-dependent quantities. Recently, Langevin approach together with the HMFT has been developed in Refs.[@PRE09036110; @JSM0910004], which have confirmed that such an approach is responsible for the effect of fluctuations in a finite-size network that often play important roles in the vicinity of phase transitions. With regard to the present study, we develop a reliable CG simulation approach, that not only can correctly predict the critical phenomena and fluctuations information, but also can be applied to diverse networks. Especially, it is shown that the validity of our [$d$-CG ]{}approach on the application of correlated networks; however, in this case the HMFT becomes, in general, very difficult to deal with. This is because that specific formulation of degree–degree correlation are unknown for most of correlated networks, such as $P(k'|k)$, that is the conditional probability of a node of degree $k$ being connected to a node of degree $k'$. On the other hand, the CSC discussed in this work may provide a solid understanding of the physical mechanism behind the basic assumption of HMFT. Our analysis here may lead to the advancement in efficient and consistent CG approaches for dynamics on surfaces and soft lattices.
Note that the present study is limited to the case of quenched networks, that is, the connectivity of networks is frozen in time. While for the case of annealed networks, i.e. the networks themselves are dynamical objects, our CG approach will encounter some difficulties in application. In this case, since the network connections are frequently reshuffled, the adjacency matrix of the resulting CG–network and CG variables should be accordingly updated. This will lead to the very inefficiency of our CG approach in simulations. However, an important advancement in studying critical phenomena on annealed networks has been made in Refs.[@PRE09056115; @PRE09051127] by means of finite-size scaling theory. Developing coarse-grained simulation methods on annealed networks and adaptive networks [@JRS08000259] deserve further investigations.
This work was supported by the National Science Foundation of China under Grants No. 20933006 and No. 20873130.
[40]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
, ****, ().
, ****, ().
, ****, ().
, , , , , ****, ().
, , , , , ****, ().
, , , , ****, ().
, ****, ().
, , , ****, ().
, , , , ****, ().
, , , , ****, ().
, ****, ().
, ****, ().
, , , , , , ****, ().
, , , ****, ().
, , , , ****, ().
, ****, ().
, , , ****, ().
, ****, ().
, , , ****, ().
, ****, ().
, , , , ****, ().
, ****, ().
, ****, ().
, ****, ().
, , , , , ****, ().
, ** (, , ).
, ****, ().
, ****, ().
, , , ****, ().
, , , , ****, ().
, , , ****, ().
, ****, ().
, ****, ().
, , , ****, ().
, ****, ().
, ****, ().
, , , ****, ().
, ****, ().
, ****, ().
, ****, ().
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Given a compact Riemannian spin manifold with positive scalar curvature, we find a family of connections $\nabla^{A_t}$ for $t\in[0,1]$ on a trivial vector bundle of sufficiently high rank, such that the first eigenvalue of the twisted Dirac operator $D_{A_t}$ is nonzero and becomes arbitrarily small as $t\to1$. However, if one restricts the class of twisting connections considered, then nonzero lower bounds do exist. We illustrate this fact by establishing a nonzero lower bound for the Dirac operator twisted by Hermitian-Einstein connections over Riemann surfaces.'
author:
- |
Marcos Jardim\
IMECC - UNICAMP\
Department of Mathematics\
C. P. 6065\
13083-970 Campinas-SP, Brazil\
and\
Rafael F. Leão\
Federal University of Paraná\
Department of Mathematics\
C. P. 019081\
81531-990 Curitiba-PR, Brazil
title: On the eigenvalues of the twisted Dirac operator
---
Introduction
============
Given a ${\mbox{Spin}}$ manifold $(M,g)$, one can consider the spinor bundle ${\mathbb}{S}$ and the associated Dirac operator $D$. There is a vast literature about this operator, in particular concerning the behavior of its spectrum. One of the most studied problems is to find lower bounds for the eigenvalues of $D$. The most well-known result was obtained by Friedrich [@Fr], and can be stated as follows: if $(M,g)$ is a compact Riemannian ${\mbox{Spin}}$ manifold, with positive scalar curvature $R$, then the eigenvalues of the associated Dirac operator satisfy the inequality $$\lambda^2 \geq \frac{1}{4} \frac{n}{n-1} R_0$$ where $R_0$ is the minimum of the scalar curvature and $n$ is the dimension of $M$. If the equality is satisfied, then the scalar curvature $R$ is constant and $M$ is an Einstein manifold. In addition, if further geometric structures in $(M,g)$ are conjecture, the above lower bound can be improved. For example, when $(M,g)$ is a compact Kähler manifold, then Kirchberg [@Ki], proved that the eigenvalues of the Dirac operator satisfy the inequality $$\lambda^2 \geq
\left\{
\begin{matrix}
\frac{1}{4} \frac{k+1}{k} R_0 & \mbox{if $k = \dim_{{\mathbb{C}}} M$ is odd,} \\
\/ & \/ \\
\frac{1}{4} \frac{k}{k-1} R_0 & \mbox{if $k = \dim_{{\mathbb{C}}} M$ is even,}
\end{matrix}
\right.$$ The influence on the eigenvalues of the Dirac operator of other geometric structures on $M$ are found in [@Kr1; @Kr2]; see also the survey [@JL1]. Kirchberg has also considered estimates in terms of other curvature tensors, see [@Ki2] and the references therein.
However, this is not the only Dirac operator one can define on a ${\mbox{Spin}}$ manifold $(M,g)$. For instance, consider a Hermitian vector bundle with a compatible connection $(E,\nabla^A)$ over $(M,g)$. Using the connection $\nabla^A$ on $E$, one can define the twisted Dirac operator $D_A$ on ${\mathbb}{S} \otimes E$. This operator is extremely important in classical field theory, for it describes particles like electrons and neutrinos coupled to external gauge fields. Very little is known about the behavior of the eigenvalues of $D_A$ in terms of the coupling connection $\nabla^A$.
More precisely, fix the base manifold and geometry $(M,g)$, and a hermitian vector bundle $E\to M$. Let ${\mathcal}{A}$ be the affine space of compatible connections on $E$, and consider the functional $$\label{functional}
\begin{split}
\lambda &: {\mathcal}{A} \rightarrow {\mathbb{R}}^+ \\
\nabla^A &\mapsto \mid \lambda_1(\nabla^A) \mid
\end{split}$$ that associates to each connection $\nabla^A\in{\mathcal}{A}$ the absolute value of the first nonzero eigenvalue of the associated Dirac operator $D_A$. In analogy with the untwisted case, it would be interesting to determine whether this functional possesses a nonzero lower bound, and whether it is bounded above.
In [@VW], Vafa and Witten have proved that the functional $\lambda$ does admit an universal upper bound which depends only in the geometry of the compact base manifold $(M,g)$ but not on the twisting bundle $E$; see also [@At] for a clear geometrical proof. On the other hand, lower bounds are known only for very special cases, see for instance [@Mi], and rely on strong restrictions both on the base manifold $(M,g)$ and especially on the connection $\nabla^A$.
Despite these particular results, the problem of how the bounds for eigenvalues of the twisted Dirac operator, and consequently the spectrum, changes with the twisting connection is not well understood. The first natural question is whether the functional $\lambda$ has a nonzero lower bound.
In this article, we construct a general example showing that the first eigenvalue of the twisted Dirac operator depends strongly on the twisting connection. More precisely, if $(M,g)$ is a compact Riemannian ${\mbox{Spin}}$ manifold with positive scalar curvature $R>0$, then we find a trivial bundle $\underline{{\mathbb{C}}}^N\to M$, for some $N$ large enough, and a family of connections $\nabla^{A_t}$ on $\underline{{\mathbb{C}}}^N$ for $t\in[0,1]$, such that the first eigenvalue of the twisted Dirac operator $D_{A_t}$, is nonzero and becomes arbitrarily small as $t\to1$. In other words, a Vafa-Witten type of result for an universal lower bound for the first eigenvalue of the twisted Dirac operator is impossible; lower bounds for the first nonzero eigenvalue necessarily depends on the topology and the geometry of the twisting bundle.
However, if one restricts the class of twisting connections considered, then lower bounds for the nonzero eigenvalues do exist. We illustrate this fact by establishing a lower bound for the Dirac operator twisted by Hermitian-Einstein connections over Riemann surfaces. More generally, we believe that interesting lower bounds for the twisted Dirac operator can be obtained whenever the twisting connection satisfies some classical field equation, like the anti-self-duality Yang-Mills equation.
#### Acknowledgments.
M.J. is partially supported by the CNPQ grant number 305464/2007-8 and the FAPESP grant number 2005/04558-0. R.F.L.’s research was supported by a CNPQ doctoral grant.
Arbitrarily Small Eigenvalues
=============================
Let $(M,g)$ be a compact Riemannian ${\mbox{Spin}}$ manifold with positive scalar curvature $R$. In this Section, we will construct a 1-parameter family of connections $\nabla^{A_t}$ for $t\in[0,1]$ on a trivial bundle $\underline{{\mathbb{C}}}^N={{\mathbb{C}}}^N\times M$ for which the first eigenvalue of the associated twisted Dirac operator $D_{A_t}$, is nonzero and arbitrarily small. The main idea is to show that the trivial bundle $\underline{{\mathbb{C}}}^N$ admits two connections, $\nabla^0$ and $\nabla^1$, such that the Dirac operator twisted by $\nabla^0$ has trivial kernel while the Dirac operator twisted by $\nabla^1$ has non trivial kernel.
The difficult part is to find the right trivial bundle $\underline{{\mathbb{C}}}^N$ and the connection $\nabla^1$, after constructing this connection the connection $\nabla^0$ is easily find. To do this, let us first assume that the manifold $(M,g)$ is even dimensional; let $E\to M$ be a hermitian vector bundle over $M$, and $\nabla^A$ be a compatible connection on $E$. In this case, the spinor bundle splits as follows $${\mathbb}{S} = {\mathbb}{S}^+ \oplus {\mathbb}{S}^-~,$$ which implies that the twisted Dirac operator also splits $D_A=D_A^+\oplus D_A^-$, where $$\begin{split}
D_A^+:{\mathbb}{S}^+ \otimes E \rightarrow {\mathbb}{S}^- \otimes E \\
D_A^-:{\mathbb}{S}^- \otimes E \rightarrow {\mathbb}{S}^+ \otimes E
\end{split}$$
The index for the operator $D_A^+$ can be topologically calculated with the following formula $${\mbox{ind}}( D_A^+ ) = (-1)^n \int_M ch(E) \wedge \hat{A}(M) \label{ind_aco}$$ where $\hat{A}$ is the so-called Â-genus of $M$, a characteristic class that can be written in terms of the Pontrjagin classes, and $ch(E)$ is the Chern character of the bundle $E$. This expression, together with the analytical index $${\mbox{ind}}( D_A^+ ) = \dim \ker D_A^+ - \dim \ker D_A^-$$ and the fact that both ways to calculate the index are equal by the Atiyah-Singer index theorem, allows us to conclude that we can find a hermitian vector bundle $E$ with connection $\nabla^A$ such that $D_A$ has nontrivial kernel. Indeed, if the index of $D_A^+$ is not zero, then $D_A$ has a nontrivial kernel. In other words, if the Dirac operator does not have kernel, then the index of $D_A^+$ must vanish.
If $\dim M=2n$, take $E\to M$ to be the pullback of the generating bundle $H$ over $S^{2n}$ by a map $f:M\to S^n$ of degree 1. Then the only non vanishing Chern class of $E$ is the top Chern class $c_n(E)$, which can be represented, in de Rham cohomology, by a generator of $H^{2n}(M)$. Using this construction and the assumption that the original Dirac operator $D$ does not have kernel it is easy to see, using the topological index formula (\[ind\_aco\]), that $D_A^+$ has nonzero index, hence $D_A$ must have kernel for any connection $\nabla^{A}$ we choose in $E$.
Since $M$ is compact, we can find a bundle $E^{\prime}$ such that $$E \oplus E^{\prime} \simeq \underline{{\mathbb{C}}}^N$$ for some large enough $N$. This is the desired trivial bundle. Choosing any connection $\nabla^B$ in $E^{\prime}$, set $\tilde{A}=A\oplus B$, which is a connection on $E \oplus E^{\prime} \simeq \underline{{\mathbb{C}}}^N$. Next, consider the twisted Dirac operator $D_{\tilde{A}} = D_{A}\oplus D_B$; the fact that $D_A$ has nontrivial kernel in $E$ implies that $D_{\tilde{A}}$ also does.
If the dimension of $M$ is odd then we can consider the even dimensional manifold $M \times S^1$ and use the fact that the spectrum of the Dirac of this product manifold is given by $$\pm \sqrt{\lambda_j^2 + \beta_k^2}$$ where $\lambda_j$ are the eigenvalues of the Dirac operator on $M$ and $\beta_k$ the eigenvalues of the Dirac operator on $S^1$. But the first eigenvalue of the Dirac operator on $S^1$ is zero, so the proof for the odd dimensional case follows as above.
The construction of the connection $\nabla^0$ is trivial. Note that sections of the trivial bundle ${\mathbb}{S}\otimes\underline{{\mathbb{C}}}^N$ can be written in the form $$\psi =
\begin{pmatrix}
\psi_1 \\
\psi_2 \\
\vdots \\
\psi_N
\end{pmatrix}$$ which implies that, for the trivial connection $\underline{d}$ on $\underline{{\mathbb{C}}}$, the associated twisted Dirac operator $D_0$ can be written as $$D_0 \psi =
\begin{pmatrix}
D \psi_1 \\
D \psi_2 \\
\vdots \\
D \psi_N
\end{pmatrix}$$ where $D$ is the free, untwisted Dirac operator associated to $(M,g)$. In particular, this means that the only effect of the coupling with the trivial connection is to change the multiplicity of the eigenvalues, leaving the spectrum unchanged. Since we assumed that the scalar curvature of $(M,g)$ is positive, the Weitzenböck formula $$D^2 = \Delta + \frac{1}{4}R$$ ensures that the Dirac operator associated to $(M,g)$ does not have kernel and, consequently, that the twisted operator $D_0$ also has trivial kernel.
Summing up, if $(M,g)$ is a Riemannian ${\mbox{Spin}}$ manifold with positive scalar curvature, then one can choose a connection, the trivial one, on the trivial vector bundle $\underline{{\mathbb{C}}}^N$ such that the twisted Dirac operator does not have a kernel.
Now we have in $\underline{{\mathbb{C}}}^N$ the two connections with the desired properties: one is the trivial connection $\underline{d}$ and the other is $\nabla^{\tilde{A}} = \nabla^{A} \oplus \nabla^B$. The Dirac operator associated to the first has trivial kernel, while the one associated to the latter has nontrivial kernel. Since the space of connections over a fixed vector bundle is an affine space modeled in the space of 1-forms over $M$ with values in the endomorphism bundle, we can write $$\nabla^{\tilde A} = \underline{d} + \alpha$$ where $\alpha$ is an 1-form over $M$ with values in the endomorphism bundle. This can be used to define the family of connections $$\nabla^{A_t} = \underline{d} + t \alpha$$ and the family of twisted Dirac operators $D_{A_t}$ in the obvious way. Clearly, $D_{A_0}=D_0$ and $D_{A_1}=D_{\tilde A}$, and by construction, $D_{A_0}$ does not have kernel while $D_{A_1}$ does. Since the functional $\lambda$ defined in (\[functional\]) is continuous (see [@At]), we conclude that $\lambda(D_{A_t})$ becomes arbitrarily small as $t\to 1$, as desired.
Uniform Bound for Riemann surfaces
==================================
In this section we show that for suitable conditions on the base manifold and on the twisting connection $\nabla^A$, it is possible to find lower bounds for the first nonzero eigenvalue of $D_A$.
First of all, note that the topological index formula (\[ind\_aco\]) always implies that the twisted operator $D_A$, over Riemann surfaces, has non vanishing index, which implies that for Riemann surfaces $D_A$ always have non trivial kernel. Because of this fact, lower bounds only makes sense for non null eigenvalues.
Riemann surfaces are naturally Kähler manifolds, with the Kähler form given by the volume form $\omega = i \xi \wedge \bar{\xi}$. We can use this to make the connection $\nabla^A$ satisfies a compatibility condition known as the Hermitian-Einstein condition. A connection $\nabla^A$ is called a Hermitian-Einstein connection if $$\omega \lrcorner F_A = -ic {\mathbb}{I}$$ where $F_A$ denotes the curvature 2-form of $\nabla^A$ and $\omega \lrcorner$ is the contraction by the Kähler form. For background on the importance of the Hermitian-Einstein condition, see [@Ko1].
Another important feature of a Riemann surface $M$ is that the spinor bundle can be explicitly described in terms of forms; more precisely, it is well-known that the spinor bundle associated with the complex structure is ${{\mathbb}{S}_{{\mathbb}{C}}}\simeq \wedge^{(0,0)} \oplus \wedge^{(0,1)}$, so the usual spinor bundle is $$\label{spins}
{\mathbb}{S} \simeq {{\mathbb}{S}_{{\mathbb}{C}}}\otimes K_M^{\frac{1}{2}} \simeq \left( \wedge^{(0,0)} \oplus \wedge^{(0,1)} \right) \otimes k_M^{\frac{1}{2}}$$ where $K_M$ is the canonical bundle of $M$. Furthermore, the complex Dirac operator $D$ coincides with a twisted real Dirac $\cal D_S$, where $S$ is the connection on $K_M^{-\frac{1}{2}}$ induced by the Chern connection on $M$.
Now let $E\to M$ be a holomorphic vector bundle of negative degree on $M$ that admits a Hermitian-Einstein connection $\nabla^A$. In [@JL2], the authors have shown that the eigenvalues of the twisted complex Dirac operator satisfy the following lower bound: $$\label{lb-d}
\lambda^2 \geq - \frac{4 \pi \deg(E)}{{\mbox{rk}}(E) {\mbox{vol}}(M)}~.$$ This lower bound for the eigenvalues of the twisted complex Dirac operator can be translated into a lower bound for the eigenvalues of the twisted real Dirac operator in the following manner. First, consider $L=K_{M}^{\frac{1}{2}}\otimes E$, so that $\deg(L) = \deg(E) - {\mbox{rk}}(E)(1-g)$, where $g$ is the genus of $M$. Let $\nabla^{\tilde A}$ be the tensor connection $\nabla^S \otimes {\mathbb}{I} + {\mathbb}{I} \otimes \nabla^A$. Then the twisted complex Dirac operator $D_{\tilde A}$, on ${{\mathbb}{S}_{{\mathbb}{C}}}\otimes L$, coincides with the twisted real Dirac operator ${\cal D}_A$, on ${\mathbb}{S} \otimes E = {{\mathbb}{S}_{{\mathbb}{C}}}\otimes K_{M}^{-\frac{1}{2}}\otimes E$. Thus applying the lower bound (\[lb-d\]) to the bundle $L$, we obtain $$\label{lb-d2}
\lambda^2 \geq \frac{4\pi(1-g)}{{\mbox{vol}}(M)}- \frac{4 \pi \deg(E)}{{\mbox{rk}}(E) {\mbox{vol}}(M)}~,$$ which is a lower bound for the eigenvalues of the twisted real Dirac operator on $M$. Of course, this lower bound is only significant provided $\deg(E)<{\mbox{rk}}(E)(1-g)$.
The Gauss-Bonnet formula can be used to re-write (\[lb-d2\]) in the following manner: $$\lambda^2 \geq \frac{R_0}{2} \left( 1 - \frac{\deg(E)}{(1-g){\mbox{rk}}(E)} \right)~~ {\rm if} ~~ g\ne1 ~,$$ $$\lambda^2 \geq - \frac{4 \pi \deg(E)}{{\mbox{rk}}(E) {\mbox{vol}}(M)} ~~ {\rm if} ~~ g=1 ~.$$ Comparing with the results in [@Al] (see Theorems 5.3, 5.10 and 5.22 there), we conclude that the above estimates are actually attained in the case where $M$ has constant scalar curvature, $E\to M$ is a line bundle and the connection on $E$ ha constant curvature.
Conclusion
==========
We have constructed a general example showing that the first eigenvalue of the twisted Dirac operator on a vector bundle over an arbitrary compact Riemannian spin manifold $M$ with positive scalar curvature depends strongly on the twisting connection. This was done by providing a 1-parameter family of connections $\nabla^{A_t}$ on a trivial vector bundle over $M$ of sufficiently high rank such that the first eigenvalue of the twisted Dirac operator $D_{A_t}$, is nonzero and becomes arbitrarily small as $t\to1$. Therefore, one cannot expect absolute lower bounds for the nonzero eigenvalues of a Dirac operator twisted by arbitrary connections.
However, if one imposes conditions on the class of twisting connections, e.g. the connection is satisfy some classical field equation like the anti-self-duality Yang-Mills equation of the Hermitian-Einstein condition, then lower bounds do exist. We illustrate this fact by establishing a sharp, nonzero lower bound for the Dirac operator twisted by Hermitian-Einstein connections over Riemann surfaces.
[c]{}
A. L. Almorox & C. T. Prieto, [*Holomorphic spectrum of twisted Dirac operators on compact Riemann surfaces*]{}, J. Geom. Phys. [**56**]{} (2006) 2069-2091.
M. Atiyah, [*Eigenvalues of the Dirac Operator*]{}, Lecture Notes in Math. 1111 (1985), 251-260.
T. Friedrich, [*Dirac Operator in Riemannian Geometry*]{}, AMS 2000.
M. Jardim, R. F. Leão, Survey on eigenvalues of the Dirac operator and geometric structures, Int. Math. Forum [**3**]{} (2008), 49-67.
M. Jardim, R. F. Leão, On the spectrum of the twisted Dolbeault Laplacian over Kähler manifolds, Preprint math/0706.0878.
K. D. Kirchberg, [*An Estimation for the First Eigenvalue of the Dirac Operator on Closed Kähler Manifolds of Positive Scalar Curvature*]{}, Ann. Global Anal. Geom. [**4**]{} (1986), 291-325.
K. D. Kirchberg, Curvature dependent lower bounds for the first eigenvalue of the Dirac operator. J. Geom. Phys. [**50**]{} (2004), 205-222.
S. Kobayashi, [*Differential Geometry of Complex Vector Bundles*]{}, Princeton University Press 1987.
W. Krammer, U. Semmelmann & G. Weingart, [it The First Eigenvalue of the Dirac Operator on Quaternionic Kähler Manifold]{}, Comm. Math. Phys. [**199**]{} (1998), 327-349.
W. Krammer, U. Semmelmann & G. Weingart, [*Eigenvalues estimates for the Dirac operator on quaternionic Kähler manifolds*]{}, Math. Z. [**230**]{} (1999), 727-751.
R. J. Miatello & R. A. Podestá, [*The Spectrum of Twisted Dirac Operators on Compact Flat Manifolds*]{}, Trans. Am. Math. Soc. [**358**]{} (2006), 4569-4603.
C. Vafa & E. Witten, [*Eigenvalues Inequalities for Fermions in Gauge Theories*]{}, Commun. in Math. Phys. [**95**]{} (1984), 257-276.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We propose a new method of estimation of the black hole masses in AGN based on the normalized excess variance, . We derive a relation between , the length of the observation, , the light curve bin size, $\Delta t$, and the black hole mass, assuming that (i) the power spectrum above the high frequency break, , has a slope of $-2$, (ii) the high frequency break scales with black hole mass, (iii) the power spectrum amplitude (in $frequency \times power$ space) is universal and (iv) is calculated from observations of length $T < 1/$. Values of black hole masses in AGN obtained with this method are consistent with estimates based on other techniques such as reverberation mapping or the -stellar velocity dispersion relation. The method is formally equivalent to methods based on power spectrum scaling with mass but the use of has the big advantage of being applicable to relatively low quality data.'
title: 'Black hole mass estimation from X-ray variability measurements in AGN'
---
\[firstpage\]
galaxies: active – galaxies: Seyfert – X-rays: galaxies
Introduction
============
The X-ray emission of active galactic nuclei (AGN) displays variations over a wide range of time scales. The first convincing demonstration of this phenomenon came with the [*EXOSAT*]{} ‘long looks’ (Lawrence 1987; McHardy & Czerny 1987). The data showed no characteristic time-scales, and the power spectral density function (PSD) showed a ‘red-noise’ shape. At the same time, it was also noticed that more luminous sources show slower variations. Various methods have been used in the past in order to determine the X-ray variability amplitude of AGN: determination of the ‘two-folding’ time-scale (i.e. the time-scale for the emitted flux to change by a factor of two), calculation of the PSD amplitude at a given frequency, and estimation of the so called ‘normalized excess variance’ (, i.e. the variance of a light curve normalized by its mean squared after correcting for the experimental noise). In all cases, these quantities appear to anti-correlate with the source luminosity (Barr & Mushotzky 1986; Lawrence & Papadakis 1993; Green, McHardy & Lehto 1993; Nandra 1997; Turner 1999; Leighly 1999; Markowitz & Edelson 2001). One possible explanation for the observed anti-correlation between the X-ray variability amplitude and X-ray luminosity, , is that more luminous AGN have larger black hole masses () as well. In this case, as increases, so does the size of the X-ray source, and a change of the source luminosity by a constant fraction takes relatively longer; equivalently, variability amplitude measured at a fixed timescale should decrease with mass. The significant progress in the measurement of the black hole (BH) mass in the centers of nearby galaxies has made it possible to actually directly test the dependence of the X-ray variability amplitude on in AGN. The data show that measured at a certain timescale is indeed anti-correlated with (Lu & Yu 2001; Bian & Zao 2003; Papadakis 2004).
If there is an intrinsic correlation between X-ray variability and , then X-ray variability measurements could be used in order to measure the central black hole mass in these objects. This possibility has already been studied in the past. Hayashida (1998) and Czerny (2001) have used the PSD normalized by the square of the mean flux as a measure of the X-ray variability of a source. By calculating the ratio of the frequencies at which the PSD$\times frequency$ has a certain value in AGN and Cyg X-1, they were able to estimate the in AGN. Recently, long, well-sampled light curves have been used in order to accurately estimate the X-ray PSD of AGN. One of the main results from these studies has been the unambiguous detection of a characteristic ‘break frequency’, , at which the power spectrum changes its slope from a value of $-2$ to $-1$ above and below , respectively (Uttley 2002; Markowitz 2003). This break frequency is analogous to the high frequency break in galactic sources and should not be confused with the low frequency break, where the spectrum changes slope from $-1$ above to $0$ below. This high frequency break appears to correlate well with , in the sense that $1/\nu_{\rm bf} \propto {\rm M_{BH}}$ (Markowitz 2003). Thus another way to estimate is through the determination of in the X-ray PSD of a source. However, this method requires rather long, high quality data.
We propose a new method to estimate in AGN, which may be used even for relatively low quality and/or short data coverage. The method is based on the use of as a measure of their X-ray variability amplitude, and takes Cyg X-1 as a reference point. Using recent, archival and light curves of 21 AGN, we estimate their , and we compare our results with estimates which are based on other methods. We find that the accuracy of the estimates using our new method is roughly similar to the accuracy of the estimates with well established methods like reverberation mapping. Furthermore, the estimation of an X-ray light curve is rather easy and straightforward. Consequently, as long as typical AGN observations performed by present day satellites (like [*XMM-Newton*]{} and [*CHANDRA*]{}) detect significant variations, they can be used to estimate and hence , using the method which we present in this work. It is possible then that the use of our method can provide estimates for a large number of AGN in the near future.
The relation between BH mass and excess variance
================================================
The scaling of the PSD with mass is most conveniently discussed after multiplying the normalized PSD by the frequency $\nu$. We assume here that the break frequency, , scales with the mass: $\nu_{\rm bf} \propto {\rm
M_{BH}}^{-1}$, but the amplitude of the normalized PSD $\times \nu$ part at the frequency break does not depend on mass. We specifically define this normalization at and call it the PSD amplitude,
$$PSD_{\rm amp} = P(\nu_{\rm bf}) \times \nu_{\rm bf},$$
where $P(\nu)$ is the power spectrum normalized to the mean squared. Recently, Papadakis (2004) has showed that this quantity is indeed similar in a few well studied Seyfert 1 galaxies.
Let us suppose that an AGN is observed for a period $T<1/$, and that the resulting light curve is binned using a bin size of $\Delta t$. The excess variance of the light curve is equal to the integral of the PSD,
$$\sigma^{2}_{\rm nxs}=\int_{1/T}^{1/2\Delta t}P(\nu)d\nu$$
Let us also suppose that $P(\nu)=A(\nu/$$)^{-2}$, where $A=$/ and is constant, and that = $B$/, where $B$ is another constant. Under these assumptions,
$$\sigma^{2}_{\rm nxs} = PSD_{\rm amp} \frac{B}{{\rm M}_{\rm
BH}}(T-2\Delta t).$$
The equation above can be re-written as follows, $${\rm M}_{\rm BH} = C(T-2\Delta t)/\sigma^{2}_{\rm nxs},
\label{bhm}$$ where $C=$ $B$. Eq. (\[bhm\]) can be used to estimate provided the of a source has been estimated, and $C$ is known. In order to determine $C$, we used the results from the PSD analysis of a Galactic black hole candidate, namely Cyg X-1.
The X-ray PSD of this source has been extensively studied the last twenty years or so (Kemp et al. 1978; Gierliński et al. 1999; Nowak et al. 1999; Churazov et al. 2001; Pottschmidt et al. 2003). In its low/hard state, the PSD of the source can be roughly represented as a power law with a high frequency slope of $-2$ , which breaks to a slope of $-1$ below $2-3$ Hz, and then to a slope of zero below $\sim 0.2-0.02$ Hz (Nowak et al. 1999; Pottschmidt 2003). Assuming that Eq. (\[bhm\]) holds for Cyg X-1 as well, and adopting the value of 10 M$_{\odot}$ for its black hole mass (Nowak et al. 1999; Gierliński et al. 1999 and references therein), we determined the value of $C$ as follows.
We used the results from the PSD analysis of 130 *RXTE* observations of Cyg X-1 reported by Pottschmidt et al. 2003. They calculated the PSD for each observation and fitted them using Lorentzian functions. The RXTE observations cover a long period of time (from December 1997 to July 2001) in which Cyg X-1 was mainly in its hard state. We chose 68 observations, carefully avoiding data from soft or failed transition states. We calculated the excess variance using the relation =$\int_{10 Hz}^{+\infty} P(\nu) d\nu$, where the lower frequency limit was chosen in such a way that it will always be higher than in this source. Using Eq. (4), with =10 M$_{\odot}$, $\Delta t=0$, and $T=1/(10$ Hz)$=0.1$ sec, we find that $C$ varies between $\sim 0.7$ and $\sim 1.3$, with the arithmetic mean being $C=0.96\pm0.02$. We accept this value for the normalization constant $C$ in Eq. (\[bhm\]).
New estimates for 21 AGN
=========================
In order to investigate the accuracy of our method, we selected for analysis several AGN with known black hole masses and recently obtained or light curves. First, we estimated the of each light curve using the relation (Nandra et al., 1997),
$$\sigma^{2}_{\rm nxs}= \frac{1}{N_{\rm data}\,\bar{x}^{2}}
\sum_{i=1}^{N_{\rm data}} [(x_{i}-\bar{x})^{2}-\sigma_{i}^{2}] \,,
\label{rms}$$
where $N_{\rm data}$ is the number of data points, and $\bar{x},
\sigma_{i}$ are the unweighted, arithmetic mean and error of the $x_{i}$, respectively. Then, using Eq. (\[bhm\]) we estimated , and compared our estimates with the existing values from literature based on the reverberation mapping technique or the measurement of stellar velocity dispersion (Section \[res\], Table 1).
\[tab.mbh\]
---------------------- ------------------------------ ------------------------------ ---------------
Name Proposal
($\times 10^{7}$M$_{\odot})$ ($\times 10^{7}$M$_{\odot})$ number
3C120 $6.3$ (1,2) $14.5$ R/P30404(3)
3C390.3 $35.5$ (1) $26.8$ R/P10340(7)
Akn120 $18.6$ (1) $5.6$ R/P30232(1)
IC4329A $0.6$ (1) $12.3$ R/P40153(2)
R/P20315(2)
R/P10313(4)
R/P50706(1)
IC5063 (S2) $5.5$ (2) $(0.19)$ R/P10337(1)
3.8
Mrk348 (S2) $1.6$ (2) $7.5$ R/P10326(3)
Mrk509 $7.5$ (1) $7.0$ R/P10311(8)
Mrk766 ($\star$) $0.35$ (2) $(0.0066)$ R/P60135(23)
0.132
NGC3227 $4.1$ (1,3) $2.2$ R/P40151(6)
R/P10292(3)
NGC3516 $1.7$ (3) $1.7$ R/P50159(1)
R/P20316(4)
R/P30224(2)
NGC3783 $1.0$ (1) $0.83$ R/P10297(2,1)
R/P30227(1)
NGC4051 ($\star$) $0.05$ (5) $(0.0014)$ R/P50153(128)
0.028
NGC4151 $1.4$ (1) $1.9$ R/P50155(2)
R/P00022(3)
R/P00024(4)
NGC4395 $0.007$(6) $0.022$ A/76006000(2)
NGC4593 $0.66$ (3) $0.60$ A/71024000(3)
A/75023010(1)
NGC5506 ($\star$) $8.8$ (4) $(0.32)$ R/P20318(10)
6.4
NGC5548 $9.2$ (1,7) $12.3$ R/P30220(2)
R/P10297(3)
R/P30218(3)
NGC7469 $0.7$ (1) $1.3$ R/P10315(20)
PG0052+251 $26.1$(1) $10.0$ R/P40157(1)
R/P20338(1)
PG0804+761 $7.6$ (1) $5.6$ R/P40157(3)
PG1211+143 ($\star$) $3.2$ (1) $(0.40)$ A/70025000(1)
8.1
---------------------- ------------------------------ ------------------------------ ---------------
: Previous and new estimates for 21 AGN
Col. (1) lists the object name and type: (S2) denote Seyfert 2 objects, ($\star$) - Narrow Line Seyfert 1. All other objects are classified as Seyfert 1 or quasars. Col. (2) lists the values taken from literature. The numbers in parentheses in Col. (2) correspond to the following references: (1) Kaspi (2000), (2) Woo & Urry (2002), (3) Onken (2003), (4) Papadakis (2004), (5) Shemmer (2003), (6) Filippenko & Ho (2003), (7) Peterson & Wandel (2000). Col. (3) lists the values obtained in this work. For NLS1s, the values in parentheses are the values found using Eq. (3), while the values listed below are multiplied by a factor of 20 (see text for details). Col. (4) lists the observation details. The first letter refers to the satellite (R – RXTE, A – ASCA), the number that follows refers to proposal number, and we list the number of light curve parts, $N_{\rm p}$, which we used in order to determine , in parentheses.
Remarks about preparation of the data {#prescr}
-------------------------------------
has regularly observed a large number of AGN over the last few years, providing long light curves with good signal to noise (S/N). For almost all the objects listed in Table 1 we have retrieved $2-9$ keV PCA background subtracted light curves from the standard data products (‘StdProds’) archive of (detailed information about the data in this archive is given at the mission web site). There are no archival data for 3 AGN with previously estimated (namely PG1211+143, NGC4395 and NGC4593). For these sources we retrieved $2-10$ keV, ASCA light curves from the TARTARUS database. We used background subtracted GIS and SIS light curves which we combined in all cases, except from the first observation of NGC4593. In this case we used the GIS data only, as there were many missing points in the SIS light curve.
The character of the obtained light curves depended mostly on the type of the monitoring. A number of the light curves came from sparse extensive monitoring. The source was observed typically for 10 ksec once per day, for several days. For example, 3C390.3 was observed usually once per day, for 2 months, with a single observation lasting about 3 ksec. Such light curves were binned to 5400 sec (roughly the orbital period of the satellite). If the resulting light curve consisted predominantly of single points separated by a day, then the remaining few consecutive data points were averaged to create a uniformly covered light curve.
Other light curves came from intensive monitoring. The source was observed continuously for 40 ksec or more, with gaps only caused by earth occultation. For example, NGC3516 was observed for over one day. The observational campaign was sometimes repeated or supplemented by sparse extensive monitoring. NGC7469 was almost continuously observed for about 1 month: 10 to 20 ksec exposures were made almost every day. Such light curves were also binned to 5400 s.
Several sources were occasionally observed for $10-40$ ksec. The bin size used for these light curves typically varied from 512 s to 5400 s, depending on the length of the data and the S/N ratio.
The light curve of PG0052+251 was binned to 10800 s in order to increase the S/N ratio. The light curve from continuous monitoring of NGC4593 was binned to 2048 s and divided into separate sets of 30 ksec duration in order to estimate in more than 2 sub-parts. Even shorter bin size was requested for NLS1 sources, as discussed below.
The long light curves were divided into separate shorter strings of length , as required by the condition $T<1/$ (see Section 2). The value of for each source was estimated from the previous estimates, using the 1/- relation of Markowitz et al. (2003). The PSD break frequencies of 3 NLS1 objects (NGC4051, NGC4253/Mrk766, and NGC5506) are high (see McHardy et al. 2003; Vaughan & Fabian 2003 and Uttley et al. 2002). Consequently, their light curves were divided into parts with $T=$1.25, 2, and 10 ksec, respectively. In order to have as many points as possible in them, we used a bin size of 16 s for the NGC4051 and Mrk766 light curves, and a bin size of 128 s in the case of NGC5506. In some cases, even if $T<1/$, long, well sampled light curves were also divided into shorter parts in order to increase the number of estimates.
The values were calculated for each short light curve or each substring of a long light curve separately. The number of light curve sub-parts ($N_{\rm p}$) used for the estimation of each source is listed in column (4) of Table 1. In the case when $N_{\rm p}>1$, we used the average value to derive the black hole mass from Eq. (\[bhm\]).
Results {#res}
-------
The values of the black hole masses obtained from Eq. (\[bhm\]) are given in Table 1.
Column (1) in this table lists the names of AGN. Four objects are classified as ‘Narrow Line Seyfert 1’ (NLS1), while the others are classical broad line Seyfert 1 (S1)/quasars (QSO), and Seyfert 2 (S2) objects. Column (2) gives the estimates taken mostly from Kaspi (2000), and they correspond to the mean value of their ‘mean’ and ‘rms’ values. In cases where we found more than one estimate for an object, the value listed in Table 1 corresponds to the mean of the respective estimates. Column (3) gives our mass determination based on X-ray variability.
Fig. 1 shows a comparison of the two mass determinations. Generally, the agreement between and is quite good for the classical Seyferts (filled squares), but not in the case of NLS1s (filled triangles). Our estimates for the latter are significantly smaller than . Excluding NLS1s, we find that on average, $\langle$/$\rangle = 2.44 \pm 1.16$. Two other sources with the largest differences between and are IC5063 and IC4329A; excluding also those sources from our sample we find that on average, $\langle$/$\rangle = 1.41 \pm 0.31$. The results for NLS1 galaxies can be reconciled with the previous mass determination if we multiply our NLS1 estimates by a factor of 20 (open triangles in Fig. 1). Interestingly, when we multiply our IC5063 estimate by the same factor 20 as for other NLS1, the new value becomes very similar to for that object. This source is classified as S2 with broad polarized Balmer lines (V[' e]{}ron-Cetty & V[' e]{}ron, 2001). Colina et al. (1991) found strong evidence for the presence of a hidden, luminous source in the nucleus of IC5063. Perhaps then, this S2 galaxy may host a NLS1 rather than a S1 nucleus, like NGC5506 (Nagar 2002). In this light we suggest that IC5063 is a hidden NLS1.
The source of the problems with IC4329A is more difficult to understand. If we use the stellar velocity dispersion measurement of 225 km s$^{-1}$ (Oliva et al., 1999) for this object, together with the -stellar velocity dispersion relation of Tremaine et al. (2002), we obtain $\sim 2.2 \times 10^{8}$ M$_{\odot}$, entirely consistent with our estimate for this source. Marziani et al. (1992) found that this source has large extinction and they gave $E(B-V)
\simeq 0.8$. After correcting the optical luminosity, $\nu L_{\nu}$, at 5100 Å for the large extinction, Nikolajuk (2004) finds = 3.9$\times 10^8$ M$_{\odot}$, using the accretion disk method. This value is again consistent with the estimate listed in Table 1. Therefore, the value given by Kaspi et al. (2000) may be too small.
When we exclude IC4329A from our sample but we consider S1 mass estimates, together with the scaled NLS1 and IC5063 estimates, then $\langle$/$\rangle = 1.30 \pm 0.25$. This result suggests that the new method we propose for the estimation in AGN provides estimates which are consistent with the estimates based on the reverberation mapping technique or the -stellar velocity dispersion relation.
Discussion
==========
In this work we propose a new method for the estimation of black hole mass in AGN. The method is based on the use of the ‘normalized excess variance’, , as estimated from X-ray light (Eq. \[rms\]). We find a simple expression between and (Eq. 4), if we assume that the power spectrum in AGN has a ‘universal’ form: (1) there is a break frequency which scales with , (2) the PSD slope above the break frequency is $-2$, and (3) the PSD amplitude below the break frequency is roughly constant in all objects, independent of (see Section 2). Note that the method is applicable to $2-10$ keV light curves, as the PSD may be different in other energy bands (e.g. Nandra & Papadakis 2001; Vaughan & Fabian 2003; Vaughan et al. 2003).
Using our method on archival , and , $2-10$ keV light curves, we estimated in 21 AGN, with previously known estimates. The comparison between the new and past estimates indicates that the method works very well for ‘classical’ AGN, i.e. S1, S2 galaxies and quasars. The values that we find with the proposed X-ray variability method are comparable to the values determined with other methods. However, there seems to be a significant disagreement for NLS1 objects. Their mass estimates, determined from the X-ray variability method, are significantly smaller than the estimates which are based on other methods. The basic issue, that there is a shift in the X-ray variability amplitude of NLS1 and ‘classical’ AGN of the same mass, was already noticed by Czerny et al. (2001). Furthermore, Papadakis (2004) and McHardy et al. (2003) have showed that the break frequency of NLS1 is probably higher than the respective frequency in the power spectra of ‘classical’ AGN with the same black hole mass. If $B^{\rm NLS1} \sim
20 \times B^{\rm class}$ (Papadakis 2004), where $B$ is the constant defined in Section (2), and is the same for both classes of objects, then since $C=$$B$, we should expect that $C^{\rm
NLS1} \sim 20 \times C^{\rm class}$. If we adopt this view, then we obtain an excellent agreement between the estimates which are based on the present X-ray variability method and the previous estimates for all objects, both ‘classical’ AGN and NLS1s.
The method we present here is based on the assumption of a ‘universal’ power spectrum shape in AGN. Consequently, it is formally equivalent to other methods which directly use the scaling of a certain power spectrum characteristic (like the break frequency, or a fixed PSD value) with (e.g. Hayashida et al. 1998; Czerny et al. 2001; Markowitz et al. 2003). However, as our method requires the estimation of only, it should be applicable for many more cases than methods which require the estimation of the PSD, as the latter is much more complicated.
Present day satellites like and [*CHANDRA*]{} will soon provide a large number of AGN light curves with a typical exposure time of $30-50$ ksec. Due to the low internal background of their detectors, most of these light curves should be of high signal to noise in the $2-10$ keV band. Furthermore, because of their long orbital periods, light curves (as long as $\sim 1.5$ days) will be evenly sampled. Provided that significant variations are observed, they could be used to estimate , and hence through Eq. (4).
The choice of the bin size, $\Delta t$, is not crucial and should depend mainly on the average count rate. It needs to be large enough in order to minimize the effects of the experimental Poisson noise, but at the same time the resulting number of the points in the light curve, $N_{\rm data}$, should also be large (at least larger than 10).
The choice of the length of a single data string, $T$, must satisfy the requirement $1/T >$ . If for a given object is not known in advance, this condition should be checked *a posteriori*. For NLS1, the length should be $\sim 20$ times smaller than the appropriate $T$ value for classical AGN, according to the results of the present work.
Apart from observational errors, the determination of the is biased by intrinsic statistical error connected with the specific shape of the power spectrum (Vaughan et al. 2003). Averaging over several independent variance measurements performed at various epochs reduces the error. If we have a single long light curve the error reduction is less efficient but can be assessed through simulations. We performed such simulations for NGC 5506 light curve (6 times longer than 1/) with many gaps, and with typical duration of a single string of $10^4$ s. Systematic ’leak’ of the power from long to shorter timescales appeared to be negligible, and the intrinsic statistical scatter was $\Delta \log$ $ = ^{+0.20}_{-0.19}$, at 90 per cent confidence level. Finally, we determined the constant $C$ in Eq. (\[bhm\]) assuming that = 10 M$_{\odot}$ for Cyg X-1. However, if a higher value of $\sim 20$ M$_{\odot}$ is more appropriate for this object (e.g. Ziółkowski 2004), then $C$ and the estimates listed in Table 1 should be increased by a factor of $\sim 2$. Such a change would, however, actually introduce a discrepancy between the masses estimated from the present X-ray variability and from other methods. Although the mass determination by the reverberation method may also contain a systematical error of the order of a factor $2-3$ (see Krolik 2001), this possibility underlines the importance of the accurate determination of the properties of Cyg X-1 as this object is routinely used as a reference source.
Conclusions
===========
- We propose a new method based on measurements of of X-ray variability of AGN to estimates. The estimates of black hole masses in AGN obtained from our method () are consistent with the estimates based on other techniques such as the reverberation mapping technique or the -stellar velocity dispersion relation.
- The constant $C$ in Eq. (\[bhm\]) is equal to 0.96 for ‘classical’ AGN and 19.2 (i.e. 20 times larger) for NLS1
- We suggest that IC5063, classified as S2 galaxy, is in fact a hidden NLS1.
Acknowledgments {#acknowledgments .unnumbered}
===============
This work was partially supported by grants 2P03D 00322 and PBZ-KBN-054/P03/2001. This research has made use of the TARTARUS database, which is supported by Jane Turner and Kirpal Nandra under NASA grants NAG5-7385 and NAG5-7067. IEP thanks CAMK for their hospitality.
Barr P., Mushotzky R.F., 1986, Nat, 320, 421 Bian W., Zao Y., 2003, MNRAS, 343, 164 Churazov E., Gilfanov M., Revnivtsev M., 2001, MNRAS, 321, 759 Colina L, Sparks W.B., Macchetto F., 1991, ApJ, 370, 102 Czerny B., Niko[ł]{}ajuk M., Piasecki M., Kuraszkiewicz J., 2001, MNRAS, 325, 865 Filippenko A.V., Ho L.C., 2003, ApJ, 588, L13 Gierli[' n]{}ski M., Zdziarski A.A., Poutanen J., Coppi P.S., Ebisawa K., Johnson W.N., 1999, MNRAS, 309, 496 Green A.R., McHardy I.M., Lehto H.J., 1993, MNRAS, 265, 664 Hayashida K., Miyamoto S., Kitamoto S., Negoro H., Inoue H., 1998, ApJ, 500, 642 Kaspi S., Smith P.S., Netzer H., Maoz D., Buell T.J., Giveon U., 2000, ApJ, 533, 631 Kemp J.C., Herman L.C., Barbour M.S., 1978, AJ, 83, 962 Krolik J.H., 2001, ApJ, 551, 72 Lawrence A., Papadakis I.E., 1993, ApJ, 414, L85 Lawrence A., Watson M.G., Pounds K.A., Elvis M., 1987, Nat, 325, 694 Leighly K.M., 1999, ApJS, 125, 297 Lu Y., Yu Q., 2001, MNRAS, 324, 653 Markowitz, A., Edelson, R., 2001, ApJ, 547, 684 Markowitz, A. , 2003, ApJ, 593, 96 Marziani P., Calvani M., Sulentic J.W., 1992, ApJ, 393, 658 McHardy I., Czerny B., 1987, Nat, 325, 696 McHardy I., Papadakis I.E., Uttley P., Page M., Mason K., 2003, MNRAS, submitted (astro-ph/0311220) Nagar N.M., Oliva E., Marconi A., Maiolino R., 2002, A&A, 391, L21 Nandra K., George I.M., Mushotzky R.F., Turner T.J., Yaqoob T., 1997, ApJ, 476, 70 Nandra K., Papadakis, I. E., 2001, ApJ, 554, 710 Nikolajuk M., 2004, A&A, in preparation Nowak M.A., Vaughan B.A., Wilms J., Dove J.B., Begelman M.C., 1999, ApJ, 510, 874 Oliva E., Origlia L., Maiolino R., Moorwood A.F.M., 1999, A&A, 350, 90 Onken C.A., Peterson B.M., Dietrich M., Robinson A., Salamanca I.M., 2003, ApJ, 585, 120 Papadakis I.E., 2004, MNRAS, 348, 207 Peterson B.M., Wandel A., 2000, ApJ, 540, L13 Pottschmidt K. , 2003, A&A, 407, 1039 Shemmer O., Uttley P., Netzer H., McHardy I.M., 2003, MNRAS, 343, 1341 Tremaine S., 2002, ApJ, 574, 740 Turner T.J., George I.M., Nandra K., Turcan D., 1999, ApJ, 524, 667 Uttley P., McHardy I., Papadakis I.E., 2002, MNRAS, 332, 231 Vaughan S., Edelson R., Warwick R.S., Uttley P., 2003, MNRAS, 345, 1271 Vaughan S., Fabian A.C., 2003, MNRAS, 341, 496 Vaughan S., Fabian A.C., Nandra K., 2003, MNRAS, 339, 1237 V[' e]{}ron-Cetty M.-P., V[' e]{}ron P., 2001, A&A, 374, 92 Woo J.H.,Urry M.C., 2002, ApJ, 579, 530 Ziółkowski, 2004, MNRAS, submitted
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The relationship between galaxies and the state/chemical enrichment of the warm-hot intergalactic medium (WHIM) expected to dominate the baryon budget at low-$z$ provides sensitive constraints on structure formation and galaxy evolution models. We present a deep redshift survey in the field of 1ES1553$+$113, a blazar with a unique combination of UV$+$X-ray spectra for surveys of the circum-/intergalactic medium (CGM/IGM). [@Nicastro:2018] reported the detection of two OVII WHIM absorbers at $z=0.4339$ and $0.3551$ in its spectrum, suggesting that the WHIM is metal-rich and sufficient to close the missing baryons problem. Our survey indicates that the blazar is a member of a $z=0.433$ group and that the higher-$z$ OVII candidate arises from its intragroup medium. The resulting bias precludes its use in baryon censuses. The $z=0.3551$ candidate occurs in an isolated environment $630$ kpc from the nearest galaxy (with stellar mass $\log M_*/M_\odot \approx 9.7$) which we show is unexpected for the WHIM. Finally, we characterize the galactic environments of broad HI Ly$\alpha$ absorbers (Doppler widths of $b=40-80$ [${\rm km\,s}^{-1}$]{}; $T\lesssim4\times10^5$ K) which provide metallicity independent WHIM probes. On average, broad [Ly$\alpha$]{} absorbers are ${\approx}2\times$ closer to the nearest luminous ($L>0.25 L_*$) galaxy (700 kpc) than narrow ($b<30$ [${\rm km\,s}^{-1}$]{}; $T\lesssim4\times10^5$ K) ones (1300 kpc) but ${\approx}2\times$ further than OVI absorbers (350 kpc). These observations suggest that gravitational collapse heats portions of the IGM to form the WHIM but with feedback that does not enrich the IGM far beyond galaxy/group halos to levels currently observable in UV/X-ray metal lines.'
author:
- 'Sean D. Johnson'
- 'John S. Mulchaey'
- 'Hsiao-Wen Chen'
- 'Nastasha A. Wijers'
- Thomas Connor
- Sowgat Muzahid
- Joop Schaye
- Renyue Cen
- 'Scott G. Carlsten'
- Jane Charlton
- 'Maria R. Drout'
- 'Andy D. Goulding'
- 'Terese T. Hansen'
- 'Gregory L. Walth'
title: 'The Physical Origins of the Identified and Still Missing Components of the Warm-Hot Intergalactic Medium: Insights from Deep Surveys in the Field of Blazar 1ES1553+113'
---
Introduction {#section:intro}
============
Cosmological simulations predict that gravitational shocks associated with structure formation will heat a large fraction of the cool ($T\approx 10^4$ K) intergalactic medium (IGM) that dominates the baryon budget in the early Universe to form a Warm-Hot Intergalactic Medium (WHIM; $T\approx 10^5 - 10^7$ K) at $z\lesssim1$ [e.g. @Cen:1999]. The predicted physical state and enrichment levels of the WHIM depend sensitively on stellar and black hole feedback which provide additional heating and chemical enrichment [e.g. @Rahmati:2016; @Nelson:2018; @Wijers:2019]. Observations of the WHIM and its relationship to galaxies can, therefore, serve as a check of our cosmological paradigm and as a laboratory for studying galaxy evolution.
While observationally elusive, the WHIM can be detected via absorption spectroscopy through ionic transitions [ in the UV and X-ray]{} as well as through metallicity independent probes such as broad HI [Ly$\alpha$]{} absorption [e.g. @Danforth:2010], the Sunyaev-Zel’dovich (SZ) effect [e.g. @de-Graaff:2019], and the dispersion measure of fast radio bursts [FRBs; e.g. @Bannister:2019; @Ravi:2019]. [ Surveys of the highly ionized phases of the CGM/IGM traced by OVI [e.g. @Danforth:2016], NeVIII [e.g. @Pachat:2017; @Frank:2018], and MgX [@Qu:2016] ]{}with the Cosmic Origins Spectrograph [@Green:2012] on the [*Hubble Space Telescope*]{} (HST) can account for a large fraction of the baryons expected in the WHIM but leave $\sim30\%$ missing [e.g. @Shull:2012a] and potentially in a chemically pristine or more highly ionized phase.
Surveys of CGM/IGM [ around galaxies]{} find that metal ion absorption is common in the CGM at projected distances ($d$) less than the estimated galaxy host halo virial radii ($R_{\rm h}$) but comparatively rare at larger distances [e.g. @Liang:2014; @Turner:2014; @Johnson:2015a; @Johnson:2017; @Burchett:2019]. These observations suggest that feedback may be ineffective at enriching the IGM far beyond galaxy halos. Indeed, the statistical detection of SZ signal from the filaments between massive galaxies [@de-Graaff:2019] can potentially account for the remaining missing baryons, suggesting that a substantial portion of the WHIM exhibits low metallicities [$<\frac{1}{10}$ solar; @Liang:2014; @Johnson:2015a] or high temperatures ($T>6\times10^{5}$ K) not traced in the UV.
New insights into chemical enrichment mechanisms and the physical state of the CGM/IGM require deep galaxy surveys in fields with UV and X-ray absorption spectra. Blazars are ideal for such studies because of their high UV/X-ray flux levels. Recently, [@Nicastro:2018] obtained a 1.7 Msec [*XMM-Newton*]{} X-ray spectrum of the blazar [${\rm 1ES}\,1553{+}113$]{}, reaching the S/N levels required to detect hot CGM/IGM absorbers individually over a large redshift pathlength for the first time. [ The X-ray spectrum revealed two candidate OVII absorption systems at $z=0.4339$ and $0.3551$, each with statistical significance of $\approx3-4\sigma$, though we note that systematic/non-Gaussian errors [e.g. @Nevalainen:2019] and contamination [e.g. @Nicastro:2016] have led to past controversies over X-ray absorbers. Nevertheless, taken together, the two OVII absorbers reported by [@Nicastro:2018] suggest that the hot phase of the CGM/IGM is metal-rich and accounts for $10-70\%$ of the baryon budget.]{} However, the combination of a poorly constrained blazar redshift (due to a featureless spectrum) and limited complementary galaxy surveys complicates the interpretation of [ absorbers]{} toward [${\rm 1ES}\,1553{+}113$]{}.
Here, we present a deep and highly complete galaxy redshift survey in the field of [${\rm 1ES}\,1553{+}113$]{}. When combined with UV absorption spectra, the survey enables a precise measurement of the redshift of [${\rm 1ES}\,1553{+}113$]{} and provides insights into the origins of intervening IGM/CGM systems. The letter proceeds as follows: In Section \[section:data\], we describe the galaxy survey and UV [ spectroscopy]{}. In Section \[section:redshift\], we combine these datasets to infer the blazar redshift. In Section \[section:environment\], we characterize the galactic environments of the candidate WHIM absorbers and draw insights into their origins.
[ We]{} adopt a flat $\Lambda$ cosmology with $\Omega_{\rm m}=0.3$, $\Omega_\Lambda = 0.7$, and $H_0=70\ {\rm km\,s^{-1}\,Mpc^{-1}}$. All magnitudes are in the AB system. [ We define the knee in the galaxy luminosity function, $L_*$, as $M_r=-21.5$ [@Loveday:2012]]{}.
Observations and Data {#section:data}
=====================
{width="\textwidth"}
Galaxy survey data
------------------
To [ study]{} the relationship between galaxies and the IGM, we conducted a deep and highly complete redshift survey targeting galaxies of $m_r<23.5$ mag in the field of [${\rm 1ES}\,1553{+}113$]{} with [ multi-slit spectrographs]{} on the Magellan Telescopes. We acquired deep $g$-, $r$-, and $i$-band images with [ MOSAIC on the Mayall]{} telescope with 1800 sec of exposure in each filter under $1''$ [ seeing]{} (PI: Johnson; PID: 2015A-0187) and an [*HST*]{} image with the ACS$+$F814W filter and 1200 sec of exposure (PI: Mulchaey; PID: 13024). We processed the data as described in [@Chen:2009] and [@Johnson:2015a]. In total, we measured spectroscopic redshifts for 921 galaxies at angular distances of $\Delta \theta < 14'$ from the blazar sightline and also include 25 redshifts from [@Prochaska:2011] and one from [@Keeney:2018]. [ Redshift]{} histograms and completeness levels for galaxies of $L \gtrsim 0.25 L_*$ ($>90\%$ at projected distances of $d<500$ kpc and $z<0.5$) are shown in Figure \[figure:completeness\].
The [ survey]{} results are summarized in Table \[table:galaxies\] which reports coordinates, apparent magnitudes ($m_g, m_r, m_i$), redshift quality (“g” for secure redshifts and “s” for single-line redshifts), redshift ($z_{\rm gal}$), rest-frame $M_g - M_r$ color, absolute rest-frame $r$-band magnitude ($M_r$), stellar mass ($\log M_*/M_\odot$), and projected angular & physical separations from the blazar sightline ($\Delta \theta$ & $d$). The absolute magnitudes include $k$-corrections, and the stellar masses are [ estimated as in]{} [@Johnson:2015a] assuming a [@Chabrier:2003] IMF. Typical uncertainties in the redshifts, magnitudes, and stellar masses are $60$ [${\rm km\,s}^{-1}$]{}, $0.1$ mags, and $0.2$ dex respectively. Table \[table:galaxies\] is separated into sections by redshift within $\pm1000$ [${\rm km\,s}^{-1}$]{} of the two candidate OVII absorbers ($z=0.4291-0.4387$; $0.3506-0.3596$), those at higher redshift ($z>0.4387$), and all other redshifts. Figure \[figure:field\_image\] displays an image of the field with galaxy redshifts labeled.
UV absorption spectroscopy
--------------------------
The COS GTO team acquired [ G130M$+$G160M]{} spectra of [${\rm 1ES}\,1553{+}113$]{} (PI: Green; PID: 11520, 12025) which are useful for inferring the redshift of the blazar based on the presence/absence of [Ly$\alpha$]{} forest absorption. We retrieved all available [ COS spectra]{} for [${\rm 1ES}\,1553{+}113$]{} from the [*HST*]{} archive and combined them into a single coadded spectrum as described in [@Johnson:2013].
[ccccccccrrrr]{}
R.A. & Decl. & $m_g$ & $m_r$ & $m_i$ & quality & $z_{\rm gal}$ & $M_g - M_r$ & $M_r$ & $\log M_*/M_\odot$ & &\
(J2000) & (J2000) & (AB) & (AB) & (AB) & & (AB) & (AB) & & & &\
\
15:55:44.20 & +11:11:04.6 & 23.1 & 22.1 & 21.4 & g & 0.5721 & 0.4 & $-21.3$ & 10.3 & 26.1 & 170\
15:55:43.91 & +11:11:54.4 & 22.9 & 21.6 & 20.8 & g & 0.5963 & 0.5 & $-21.9$ & 10.7 & 32.7 & 218\
15:55:45.22 & +11:11:20.7 & 23.3 & 21.9 & 21.1 & g & 0.6626 & 0.4 & $-22.2$ & 10.6 & 32.3 & 226\
15:55:45.38 & +11:11:09.9 & 23.6 & 23.1 & 22.9 & s & 1.1130 & 0.2 & $-22.5$ & 10.3 & 37.2 & 305\
15:55:42.08 & +11:10:32.1 & 23.1 & 22.2 & 22.0 & g & 0.5244 & 0.2 & $-20.2$ & 9.4 & 54.1 & 338\
15:55:44.62 & +11:12:05.5 & 23.7 & 23.2 & 22.8 & g & 0.7753 & 0.2 & $-20.4$ & 9.5 & 47.3 & 351\
15:55:40.28 & +11:10:55.3 & 23.9 & 23.1 & 22.3 & g & 0.8778 & 0.4 & $-21.9$ & 10.5 & 49.9 & 386\
15:55:47.24 & +11:11:11.6 & 23.7 & 23.1 & 22.3 & g & 0.7270 & 0.2 & $-20.8$ & 9.7 & 63.0 & 457\
15:55:37.80 & +11:11:43.4 & 23.0 & 21.4 & 20.9 & g & 0.4695 & 0.4 & $-21.2$ & 10.3 & 79.5 & 469\
15:55:38.87 & +11:11:54.8 & 22.5 & 21.6 & 21.1 & g & 0.6935 & 0.2 & $-21.9$ & 10.1 & 68.5 & 488\
15:55:37.26 & +11:11:23.9 & 23.2 & 21.5 & 20.9 & g & 0.4699 & 0.5 & $-21.3$ & 10.5 & 85.2 & 503\
& & & & & & & & & &\
\
15:55:43.51 & +11:11:13.0 & 23.9 & 22.7 & 22.5 & g & 0.4300 & 0.5 & $-20.8$ & 10.2 & 13.2 & 74\
15:55:43.19 & +11:11:02.8 & 23.6 & 22.2 & 21.5 & g & 0.4347 & 0.6 & $-20.5$ & 10.3 & 21.6 & 122\
15:55:41.45 & +11:11:29.1 & 25.4 & 23.3 & 22.9 & s & 0.4297 & 0.4 & $-19.1$ & 9.5 & 23.9 & 134\
15:55:42.76 & +11:11:56.4 & 24.9 & 23.1 & 22.6 & g & 0.4344 & 0.7 & $-19.6$ & 9.9 & 32.3 & 183\
15:55:43.95 & +11:11:56.7 & 21.4 & 19.6 & 18.9 & g & 0.4343 & 0.8 & $-23.2$ & 11.4 & 35.0 & 198\
15:55:42.94 & +11:10:47.6 & 23.6 & 22.1 & 21.5 & g & 0.4329 & 0.6 & $-20.4$ & 10.2 & 36.8 & 207\
15:55:39.87 & +11:11:46.8 & 24.3 & 23.3 & 23.4 & g & 0.4347 & 0.0 & $-18.5$ & 8.7 & 51.9 & 293\
15:55:46.15 & +11:10:54.1 & 22.5 & 21.1 & 20.6 & g & 0.4328 & 0.5 & $-21.3$ & 10.5 & 54.8 & 308\
15:55:46.69 & +11:11:07.7 & 23.1 & 22.3 & 22.2 & g & 0.4332 & 0.0 & $-19.5$ & 9.1 & 56.1 & 316\
15:55:47.25 & +11:11:16.0 & 22.3 & 20.6 & 20.0 & g & 0.4328 & 0.7 & $-22.1$ & 11.0 & 62.4 & 351\
15:55:42.66 & +11:12:29.8 & 23.4 & 22.1 & 21.9 & g & 0.4335 & 0.3 & $-20.0$ & 9.6 & 65.7 & 370\
15:55:42.53 & +11:12:30.2 & 24.5 & 22.8 & 22.4 & g & 0.4339 & 0.5 & $-19.7$ & 9.8 & 66.3 & 374\
15:55:37.79 & +11:10:51.4 & 24.3 & 23.0 & 22.8 & g & 0.4344 & 0.3 & $-19.2$ & 9.3 & 84.0 & 474\
15:55:44.56 & +11:09:57.9 & 24.4 & 23.0 & 22.7 & g & 0.4327 & 0.4 & $-19.5$ & 9.5 & 89.3 & 503\
& & & & & & & & & & &\
\
15:55:44.49 & +11:09:19.3 & 22.8 & 21.7 & 21.4 & g & 0.3531 & 0.4 & $-19.9$ & 9.7 & 126.8 & 630\
15:56:02.12 & +11:12:44.1 & 22.7 & 21.6 & 21.3 & g & 0.3537 & 0.4 & $-20.0$ & 9.8 & 291.8 & 1451\
15:55:29.10 & +11:05:37.7 & 24.3 & 23.3 & 23.2 & g & 0.3547 & 0.1 & $-18.3$ & 8.6 & 402.8 & 2007\
15:55:47.81 & +11:04:41.8 & 22.8 & 21.9 & 21.6 & g & 0.3596 & 0.4 & $-19.8$ & 9.6 & 408.6 & 2054\
15:55:41.14 & +11:03:51.8 & 21.2 & 19.5 & 18.8 & g & 0.3580 & 0.8 & $-22.7$ & 11.2 & 453.4 & 2273\
15:56:13.90 & +11:09:06.0 & 24.1 & 23.3 & 23.4 & g & 0.3590 & 0.0 & $-18.1$ & 8.5 & 474.8 & 2384\
15:55:24.61 & +11:18:04.5 & 22.2 & 20.7 & 20.3 & g & 0.3538 & 0.6 & $-21.1$ & 10.4 & 483.4 & 2405\
15:55:12.93 & +11:17:48.8 & 21.9 & 20.9 & 20.6 & g & 0.3542 & 0.4 & $-20.7$ & 10.0 & 586.6 & 2920\
15:55:52.15 & +11:21:42.4 & 20.6 & 19.4 & 18.9 & g & 0.3531 & 0.6 & $-22.5$ & 11.0 & 632.4 & 3142\
& & & & & & & & & & &\
\
15:55:07.77 & +11:01:42.3 & 21.1 & 19.8 & 19.3 & g & 0.0017 & 1.2 & $-9.6$ & 5.8 & 780.0 & 28\
15:55:46.11 & +11:11:49.4 & 23.5 & 22.9 & 22.8 & g & 0.1022 & 0.3 & $-15.6$ & 7.8 & 51.6 & 97\
15:55:44.01 & +11:11:09.1 & 23.1 & 22.6 & 22.3 & g & 0.3892 & 0.3 & $-19.2$ & 9.3 & 20.8 & 110\
& & & & & & & & & & &\
{width="\textwidth"}
Discovery and redshift of the group hosting [${\rm 1ES}\,1553{+}113$]{} {#section:redshift}
=======================================================================
[ Optical$-$X-ray spectra of [${\rm 1ES}\,1553{+}113$]{} exhibit no detected emission lines]{}, preventing systemic redshift measurements [@Landoni:2014]. The lack of a precise redshift measurement complicates the interpretation of absorption features in the spectrum of [${\rm 1ES}\,1553{+}113$]{} due to an inability to differentiate intervening IGM/CGM systems from [ associated absorption]{}. Previous estimates of the redshift of [${\rm 1ES}\,1553{+}113$]{} based on the detection of intervening HI [Ly$\alpha$]{} absorption [e.g. @Danforth:2010] and the shape of its $\gamma$-ray spectrum [e.g. @Abramowski:2015] imply $0.413 \lesssim z_{\rm sys} \lesssim 0.6$.
Blazars are typically hosted by luminous elliptical galaxies [e.g. @Urry:2000] in massive groups [e.g. @Wurtz:1997]. Moreover, [${\rm 1ES}\,1553{+}113$]{} is a high energy peaked blazar which are thought to arise from beamed FR-I radio galaxies [e.g. @Rector:2000]. [${\rm 1ES}\,1553{+}113$]{} exhibits a complex, one-sided radio-jet morphology [see Figure \[figure:field\_image\]; @Rector:2003], indicating disturbance by a hot intragroup or intracluster medium.
[ Identification]{} of the blazar’s host group therefore represents a precise means of inferring its redshift [e.g. @Rovero:2016; @Farina:2016].
To identify the host group of [${\rm 1ES}\,1553{+}113$]{}, the top panel of Figure \[figure:zhist\] displays the redshift histogram for galaxies of $L>0.25 L_*$ from our survey at $d<500$ and $<1000$ proper kpc (pkpc) from the blazar sightline. [ With high completeness levels of $100\%$, ${>}90\%$, and ${>}80$% for galaxies of $L>1.0, 0.5,$ and $0.25$ $L_*$ respectively at $d<500$ pkpc and $z<0.6$, our redshift survey is sensitive to galaxy groups over the full range of possible systemic redshifts for [${\rm 1ES}\,1553{+}113$]{}. The only significant overdensity with multiple luminous galaxies near the blazar sightline is at $z\approx0.433$, a strong indication that the blazar is a member of the $z=0.433$ group.]{}
[ The blazar host group consists of $7$ ($14$) members of $L>0.25 L_*$ at $d<500$ (1000) pkpc from the blazar and exhibits a light-weighted redshift of $z_{\rm group}=0.433$. Not including the blazar host, the total stellar mass (luminosity) of the group is $\approx 8\times10^{11} M_\odot$ ($18\ L_*$) with $\approx 60\%$ ($50\%$) coming from three massive, quiescent galaxies of $\log M_*/M_\odot > 11.0$.]{} The measured line-of-sight velocity dispersion of the group is $\sigma_{\rm group} \approx 300$ [${\rm km\,s}^{-1}$]{}, which corresponds to an estimated dynamical mass of $M_{\rm dyn}\sim 2{-}5\times10^{13}\ M_\odot$. [ Such a massive group is consistent with expectations for the environment of blazars like [${\rm 1ES}\,1553{+}113$]{}. Assuming that the luminosity of the blazar host galaxy is $L=6 L_*$ [@Urry:2000], the group light-weighted center is $13''$ E. (70 pkpc) and $7''$ N. (40 pkpc) of the blazar position.]{}
{width="\textwidth"}
The presence/absence of HI [Ly$\alpha$]{} absorption in the blazar spectrum as a function of redshift can be used for an independent estimate of the blazar redshift. The archival COS spectrum of [${\rm 1ES}\,1553{+}113$]{} enables searches for HI [Ly$\alpha$]{} absorption at $\lambda <1796.7$ Å which corresponds to a maximum redshift of $z_{\rm Ly\alpha}=0.478$. In this wavelength range, [ the $S/N$ is sufficient]{} to detect absorbers of $W_{\rm r} > 0.03$ Å at $3\sigma$ significance. The spectrum reveals $7$ [Ly$\alpha$]{} absorbers at $z=0.350-0.413$ implying $z_{\rm sys}\gtrsim 0.413$ but none over the similar interval of $z=0.413-0.478$ (bottom left panel of Figure \[figure:zhist\]).
To quantify the redshift constraint on [${\rm 1ES}\,1553{+}113$]{} from the [Ly$\alpha$]{} forest with objects of similar luminosity, we identified $59$ available QSOs with measured systemic redshifts, archival COS spectra, and IGM absorption line identifications from [@Danforth:2016]. For each QSO, we computed the difference between the systemic redshift and that of the highest redshift HI Ly$\alpha$ line with $W_r>0.03$ Å in the spectrum, $\Delta z = z_{\rm sys} - {\rm max}(z_{\rm Ly\alpha})$. The resulting $\Delta z$ distribution is shown in the bottom right panel of Figure \[figure:zhist\]. When combined with the highest redshift [Ly$\alpha$]{} line in the spectrum of [${\rm 1ES}\,1553{+}113$]{} at $z=0.413$ ([ 50 pMpc from $z=0.433$ where the UV background dominates]{}), this distribution implies a 95% confidence interval for the redshift of [${\rm 1ES}\,1553{+}113$]{} of $z_{\rm sys}=0.411-0.435$, consistent with membership of the $z=0.433$ galaxy group. [ This constraint assumes that blazars and QSOs reside in similar intergalactic environments and is subject to small number statistics in the wings of the distribution. It will be further tested with new [*HST*]{} NUV spectra (PI: Muzahid, PID: 15835) for improved [Ly$\alpha$]{} searches]{}.
Implications for the WHIM {#section:environment}
=========================
![Metallicity independent equilibrium ion fraction of OVI (blue), OVII (red), and OVIII (black) as a function of distance from the blazar for gas with a temperature of $T = 10^6$ K and with a density of $n_{\rm H}=10^{-5}$ (top), $10^{-4}$ (middle), and $10^{-3}$ (bottom) $\rm cm^{-3}$.[]{data-label="figure:PI_CI"}](PI_CI_distance.pdf){width="\columnwidth"}
The origins of candidate WHIM X-ray absorbers
---------------------------------------------
[ [@Nicastro:2018] identified two candidate WHIM OVII absorbers at $z=0.4339$ and $0.3551$ in the [*XMM*]{} spectrum of [${\rm 1ES}\,1553{+}113$]{}]{}, suggesting that the hot phase of the IGM is metal-rich and potentially closing the missing baryon problem. Neither OVII candidate is detected in OVI, OVIII, or lower ionization metal ions, similar to recent non-detections in an X-ray emitting cosmic filament [@Connor:2019]. Here, we discuss the origins of these WHIM candidates based on our redshift survey.
As discussed in Section \[section:redshift\], [${\rm 1ES}\,1553{+}113$]{} is most likely a member of a galaxy group at $z=0.433$. The identified OVII system at $z=0.4339$ is therefore associated with the blazar environment and cannot be used in cosmic baryon [ censuses]{}.
The blazar host group is part of a larger scale overdensity consisting of three additional groups at $\approx 1.5$ pMpc S.E., $\approx1.5$ pMpc N.W., and $\approx 2.5$ pMpc E.S.E. from the blazar. [ The]{} OVII candidate could be due to WHIM from this overdensity, but photoionization from the blazar and UV background would be important. To evaluate the feasibility of the WHIM interpretation under these circumstances, we ran a series of Cloudy [@Ferland:2017] models to calculate the equilibrium OVI, OVII, and OVIII ion fractions as a function of distance from the blazar for gas with $n_{\rm H}=10^{-5}{-}10^{-3}\ {\rm cm}^{-3}$ and temperatures of $T = 10^5-10^7$ K including photoionization from the blazar ($\lambda L_{\lambda}=10^{46}\ {\rm erg\,s^{-1}}$ at $1$ Rydberg and UV spectral slope of $\alpha = -1.4$ based the COS spectrum) and UV background [@Khaire:2019] as shown in Figure \[figure:PI\_CI\].
[@Nicastro:2018] demonstrated that the OVII detection and OVI/OVIII non-detections at $z=0.4339$ require a gas temperature of $T \approx 10^6$ K with little contribution from photoionization. This rules out WHIM gas with $n_{\rm H}<10^{-4}\ {\rm cm^{-3}}$ at any distance from the blazar because photoionization by the UV background is significant at such low densities [see Figure \[figure:PI\_CI\]; @Wijers:2019]. Denser hot gas of $n_{\rm H}=10^{-4}$ ($10^{-3}$) $\rm cm^{-3}$ can reproduce the absorber properties but only at $>10$ ($1$) pMpc from the blazar (see Figure \[figure:PI\_CI\]). The $z=0.4339$ OVII candidate is, therefore, unlikely to be due to low-density WHIM but may arise from hot CGM or intragroup medium [@Mulchaey:1996] in the [ blazar environment]{}.
The $z=0.3551$ [ OVII candidate]{} resides in a comparatively isolated region with no galaxies at $d<500$ pkpc from the sightline within $\Delta v = \pm 1000$ [${\rm km\,s}^{-1}$]{} from the absorber redshift despite $100\%$ completeness levels for galaxies of $L>0.1 L_*$. The nearest galaxy to the sightline is a star-forming galaxy of $\log M_*/M_\odot = 9.7$ at $z=0.3531$ and $d=630$ pkpc or $\approx 5\times$ [ its virial radius]{} (estimated with the stellar-to-halo mass relation from [@Kravtsov:2018] and virial radius definition from [@Bryan:1998]). The nearest massive galaxy has a stellar mass of $\log M_*/M_\odot = 11.2$ and is at $d=2273$ pkpc or $\approx 5\times$ [ its]{} virial radius. The $z=0.3551$ candidate could be due to an undetected dwarf in principle, but surveys of the CGM/IGM around dwarfs [@Johnson:2017] indicate that metal absorption systems are rare beyond the virial radius, and dwarfs are not expected to maintain a hot halo [e.g. @Correa:2018].
To determine whether strong OVII systems are expected from the WHIM in isolated environments, we calculated the fraction of predicted strong OVII absorbers as a function of environment using WHIM predictions [@Wijers:2019] from the EAGLE cosmological hydrodynamical simulations [@Schaye:2015; @Crain:2015; @McAlpine:2016]. We calculated column densities within $2000\ {\rm km\,s}^{-1}$ simulation slices and cross-correlated with galaxies as a function stellar mass and projected distance. In total, only $1-3$% of the predicted, comparably strong OVII ($\log N({\rm O\,VII})/{\rm cm^{-2}}=15.6$) systems occur in similarly isolated environments ($d>630$ pkpc to the nearest galaxy of $\log M_*/M_\odot > 9.7$). While the model predictions are subject to non-negligible uncertainties [ due to]{} treatment of peculiar velocities and feedback, we nevertheless conclude that strong OVII WHIM systems are not expected to be common in isolated environments. Moreover, [@Bonamente:2018] estimated a 4% probability that the $z=0.3551$ OVII candidate arises from noise fluctuations.
[ We conclude]{} that neither of the two candidate OVII absorbers in the spectrum of [${\rm 1ES}\,1553{+}113$]{} are of confident and unbiased intergalactic origin. This implies a 95% upper limit on the number of WHIM OVII absorbers with $W_{\rm r} \gtrsim 6$ mÅ per unit redshift of $\frac{dN}{dz}<8$. The lack of strong WHIM X-ray absorption systems suggests that metal enrichment is primarily confined to galaxy halos and their immediate outskirts. This is consistent with the EAGLE simulations which predict that most strong [ OVII systems arise]{} from metal rich ($>0.5 Z_\odot$) gas at over-densities of $\delta \gtrsim 100$ [see @Wijers:2019]. Further exploration of the relationship between the WHIM and galaxies [ requires]{} metallicity independent probes.
The origins of broad HI [Ly$\alpha$]{} systems
----------------------------------------------
While metallicity independent probes of the hot IGM are not currently available (except via stacking), broad HI absorbers ($b>40$ [${\rm km\,s}^{-1}$]{}) can be used to trace metal poor, warm IGM. [While temperature measurements are not possible for most broad HI systems due to lack of detected metals, [@Savage:2014] found that $78^{+7}_{-12}\%$ of broad HI absorbers with well aligned OVI detections exhibit warm-hot temperatures of $\log T/{\rm K}=5-6$.]{} [@Danforth:2010] identified 12 broad HI [ absorbers]{} in the COS spectrum of [${\rm 1ES}\,1553{+}113$]{} at $\Delta v <-10,000$ [${\rm km\,s}^{-1}$]{} from the blazar redshift. [ None]{} of the broad HI absorbers are coincident with detected galaxies at $d<R_{\rm h}$. However, [ all]{} are coincident with at least one luminous galaxy of $L>0.25 L_*$ within $\Delta v = \pm 1000$ [${\rm km\,s}^{-1}$]{} with a median projected distance to the closest one of $700$ pkpc. In contrast, narrow ($b<30$ [${\rm km\,s}^{-1}$]{}; $T<5\times10^4$ K) HI absorption systems detected toward [${\rm 1ES}\,1553{+}113$]{} are further from luminous galaxies on average with a median distance to the nearest one of $1300$ pkpc while OVI absorbers are closer to luminous galaxies [$350$ pkpc; @Johnson:2013; @Pratt:2018].
Summary and conclusions
-----------------------
Based on deep and highly complete redshift surveys in the field of [${\rm 1ES}\,1553{+}113$]{} we found that:
1. Neither of the two candidate OVII WHIM systems reported toward the sightline [@Nicastro:2018] are of confident and unbiased intergalactic origin. [ The origins, state, and cosmological mass density of the hot IGM therefore remain uncertain]{}.
2. Low metallicity warm IGM traced by broad HI [Ly$\alpha$]{} absorbers occur $\approx 2\times$ further from luminous ($L>0.25 L_*$) galaxies than OVI absorbers on average, but $2\times$ closer than [ cool IGM traced by narrow [Ly$\alpha$]{}]{}.
Our findings are consistent with gravitational collapse heating portions of the IGM to form the WHIM. However, they also suggest that feedback is ineffective at enriching the low-$z$ IGM far beyond galaxy/group halos to levels currently observable in UV and X-ray metal ions. Indeed, [@Liang:2014] and [@Johnson:2015a] placed upper limits on the mean metallicity of the IGM of $<0.1 Z_\odot$ and pristine ($Z<0.01 Z_\odot$) gas can be found even around massive galaxies [@Chen:2019]. These observations highlight the need for [ a variety of WHIM probes coupled with deep galaxy surveys]{}.
Acknowledgements {#acknowledgements .unnumbered}
================
We are grateful to J. Nevalainen, F. Nicastro, and M. Petropoulou for insightful comments. SDJ is supported by a NASA Hubble Fellowship (HST-HF2-51375.001-A). MRD acknowledges support from the Dunlap Institute at the University of Toronto and the Canadian Institute for Advanced Research (CIFAR). J.C.C. acknowledges support by the National Science Foundation under Grant No. AST-1517816. Based on observations from the Magellan, the NOAO Mayall, and NASA/ESA [*Hubble*]{} Telescopes. The authors are honored to conduct research on Iolkam Duág (Kitt Peak), a mountain with particular significance to the Tohono Oódham. We made use of the NASA Astrophysics Data System.
natexlab\#1[\#1]{}\[1\][[\#1](#1)]{} \[1\][doi: [](http://doi.org/#1)]{} \[1\][[](http://ascl.net/#1)]{} \[1\][[](https://arxiv.org/abs/#1)]{}
, A., [Aharonian]{}, F., [Ait Benkhali]{}, F., [et al.]{} 2015, , 802, 65,
, K. W., [Deller]{}, A. T., [Phillips]{}, C., [et al.]{} 2019, Science, 365, 565.
, R. H., [White]{}, R. L., & [Helfand]{}, D. J. 1995, , 450, 559,
, M. 2018, arXiv e-prints.
, G. L., & [Norman]{}, M. L. 1998, , 495, 80,
, J. N., [Tripp]{}, T. M., [Prochaska]{}, J. X., [et al.]{} 2019, , 877, L20,
, R., & [Ostriker]{}, J. P. 1999, , 514, 1,
, G. 2003, , 115, 763,
, H.-W., & [Mulchaey]{}, J. S. 2009, , 701, 1219,
, H.-W., [Johnson]{}, S. D., [Straka]{}, L. A., [et al.]{} 2019, , 484, 431,
, T., [Zahedy]{}, F. S., [Chen]{}, H.-W., [et al.]{} 2019, arXiv e-prints, arXiv:1909.10518.
, C. A., [Schaye]{}, J., [Wyithe]{}, J. S. B., [et al.]{} 2018, , 473, 538,
, R. A., [Schaye]{}, J., [Bower]{}, R. G., [et al.]{} 2015, , 450, 1937,
, C. W., [Keeney]{}, B. A., [Stocke]{}, J. T., [Shull]{}, J. M., & [Yao]{}, Y. 2010, , 720, 976,
, C. W., [Keeney]{}, B. A., [Tilton]{}, E. M., [et al.]{} 2016, , 817, 111,
, A., [Cai]{}, Y.-C., [Heymans]{}, C., & [Peacock]{}, J. A. 2019, , 624, A48,
, E. P., [Fumagalli]{}, M., [Decarli]{}, R., & [Fanidakis]{}, N. 2016, , 455, 618,
, G. J., [Chatzikos]{}, M., [Guzm[á]{}n]{}, F., [et al.]{} 2017, , 53, 385.
, S., [Pieri]{}, M. M., [Mathur]{}, S., [Danforth]{}, C. W., & [Shull]{}, J. M. 2018, , 476, 1356,
, J. C., [Froning]{}, C. S., [Osterman]{}, S., [et al.]{} 2012, , 744, 60,
, S. D., [Chen]{}, H.-W., & [Mulchaey]{}, J. S. 2013, , 434, 1765,
—. 2015, , 449, 3263,
, S. D., [Chen]{}, H.-W., [Mulchaey]{}, J. S., [Schaye]{}, J., & [Straka]{}, L. A. 2017, , 850, L10,
, B. A., [Stocke]{}, J. T., [Pratt]{}, C. T., [et al.]{} 2018, , 237, 11,
, V., & [Srianand]{}, R. 2019, , 484, 4174,
, A. V., [Vikhlinin]{}, A. A., & [Meshcheryakov]{}, A. V. 2018, Astronomy Letters, 44, 8,
, M., [Falomo]{}, R., [Treves]{}, A., & [Sbarufatti]{}, B. 2014, , 570, A126,
, C. J., & [Chen]{}, H.-W. 2014, , 445, 2061,
, J., [Norberg]{}, P., [Baldry]{}, I. K., [et al.]{} 2012, , 420, 1239,
, S., [Helly]{}, J. C., [Schaller]{}, M., [et al.]{} 2016, Astronomy and Computing, 15, 72,
, J. S., [Mushotzky]{}, R. F., [Burstein]{}, D., & [Davis]{}, D. S. 1996, , 456, L5,
, D., [Kauffmann]{}, G., [Pillepich]{}, A., [et al.]{} 2018, , 477, 450,
, J., [Tempel]{}, E., [Ahoranta]{}, J., [et al.]{} 2019, , 621, A88,
, F., [Senatore]{}, F., [Gupta]{}, A., [et al.]{} 2016, , 458, L123,
, F., [Kaastra]{}, J., [Krongold]{}, Y., [et al.]{} 2018, , 558, 406,
, S., [Narayanan]{}, A., [Khaire]{}, V., [et al.]{} 2017, , 471, 792,
, C. T., [Stocke]{}, J. T., [Keeney]{}, B. A., & [Danforth]{}, C. W. 2018, , 855, 18,
, J. X., [Weiner]{}, B., [Chen]{}, H.-W., [Cooksey]{}, K. L., & [Mulchaey]{}, J. S. 2011, , 193, 28,
, Z., & [Bregman]{}, J. N. 2016, , 832, 189,
, A., [Schaye]{}, J., [Crain]{}, R. A., [et al.]{} 2016, , 459, 310,
, V., [Catha]{}, M., [D’Addario]{}, L., [et al.]{} 2019, , 572, 352.
, T. A., [Gabuzda]{}, D. C., & [Stocke]{}, J. T. 2003, , 125, 1060,
, T. A., [Stocke]{}, J. T., [Perlman]{}, E. S., [Morris]{}, S. L., & [Gioia]{}, I. M. 2000, , 120, 1626,
, A. C., [Muriel]{}, H., [Donzelli]{}, C., & [Pichel]{}, A. 2016, , 589, A92,
, B. D., [Kim]{}, T.-S., [Wakker]{}, B. P., [et al.]{} 2014, , 212, 8,
, J., [Crain]{}, R. A., [Bower]{}, R. G., [et al.]{} 2015, , 446, 521,
, J. M., [Smith]{}, B. D., & [Danforth]{}, C. W. 2012, , 759, 23,
, M. L., [Schaye]{}, J., [Steidel]{}, C. C., [Rudie]{}, G. C., & [Strom]{}, A. L. 2014, , 445, 794,
, C. M., [Scarpa]{}, R., [O’Dowd]{}, M., [et al.]{} 2000, , 532, 816,
, N. A., [Schaye]{}, J., [Oppenheimer]{}, B. D., [Crain]{}, R. A., & [Nicastro]{}, F. 2019, , 488, 2947.
, R., [Stocke]{}, J. T., [Ellingson]{}, E., & [Yee]{}, H. K. C. 1997, , 480, 547,
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The reachability problem in cooperating systems is known to be PSPACE-complete. We show here that this problem remains PSPACE-complete when we restrict the communication structure between the subsystems in various ways. For this purpose we introduce two basic and incomparable subclasses of cooperating systems that occur often in practice and provide respective reductions. The subclasses we consider consist of cooperating systems the communication structure of which forms a line respectively a star.'
author:
- 'Mila Majster-Cederbaum Nils Semmelrock[^1]'
bibliography:
- 'biblio.bib'
nocite: '[@*]'
title: 'Reachability in Cooperating Systems with Architectural Constraints is PSPACE-Complete'
---
Introduction
============
Cooperating systems are systems that consist of subsystems which cooperate by a *glue-code*. The reachable state space of such systems can be exponentially large in the number of subsystems what is referred to as the *state space explosion problem*. Moreover, there are PSPACE-completeness results for the reachability problem in various formalisms that model cooperating systems, e.g., see [@everythingIsPSPACE] for interaction systems, [@JonesLL77] for results in $1$-conservative Petri nets and [@esparza] for results in $1$-save Petri nets. Clearly, all methods that rely on the exploration of the reachable state space of cooperating systems suffer from these results. Particularly formal verification techniques as LTL or CTL model checking have a runtime that is exponential in the number of subsystems. There are various ways to cope with this problem. One approach is to identify subclasses for which an analysis can be achieved in polynomial time. Hence, the question arises whether there are relevant subclasses of cooperating systems where the reachability problem is decidable in polynomial time. Popular decision problems that are complete in NP or even in PSPACE are decidable in polynomial or linear time in certain subclasses of instances. Maybe the most popular example is the Boolean satisfiability problem where $3$SAT is NP-complete, $2$SAT is decidable in polynomial time and HORNSAT (the problem of deciding whether a given set of propositional Horn clauses is satisfiable) is even decidable in linear time. Similarly, the quantified $3$SAT problem is PSPACE-complete, whereas the quantified $2$SAT problem is also decidable in polynomial time (see [@complexGuide] for descriptions and more examples).
There are various starting points to specify subclasses of cooperating systems.
1. Restrictions regarding the behavior of the subsystems.
2. The degree of synchronization among the subsystems as systems with a very high degree of synchronization tend to display a smaller reachable state space.
3. The glue-code, i.e., the structure of the interaction among the subsystems.
Here, our concern lies on the latter.
As a formal model we consider here interaction systems [@interactionSystems], a very general formalism for modeling cooperating systems that allows for multiway interactions between subsystems. The results in this paper can be easily applied to other formalisms that model cooperating systems. This can be achieved by either adapting the results, e.g., the formalism of interface automata [@deAlfaro01] comes close to interaction systems, or by using a mapping among formalisms, e.g., a mapping between interaction systems and $1$-save Petri nets can be found in [@christoph08].
Deciding the reachability problem in general interaction systems is PSPACE-complete [@everythingIsPSPACE]. Here we strengthen this result by showing that the reachability problem remains PSPACE-complete in subclasses consisting of interaction systems the communication structure of which forms a star or a linear sequence of components. As star structures appear in practice in, e.g., client/server systems as banking or booking systems and linear structures appear in, e.g., pipeline systems as instruction pipelines or general queue based algorithms, it is important to know that even for such “simply” structured systems there is no general efficient analysis method. Our results justify investigations that search for sufficient conditions that can check and guarantee reachability in polynomial time. Also approaches that guarantee correctness by construction, i.e., modeling rules that ensure certain system properties, and are based on structural restrictions become justified by our results. See for example [@treeLike; @HJK08; @Lambertz09; @Lambertz11; @hoare; @roscoeArch; @bernardoArch] for approaches that treat these topics.
The paper is organized as follows. Section \[is\] contains the definitions. In Section \[lin\] we introduce a reduction from the acceptance problem in linear bounded Turing machines to the reachability problem in linear interaction systems. A reduction from the reachability problem in general interaction systems to star-like interaction systems is introduced in Section \[star\]. Section \[con\] concludes this paper.
Interaction Systems {#is}
===================
Interaction systems have been proposed by Sifakis and Gössler in [@interactionSystems] to model cooperating systems. The model was studied, e.g., in [@treeLike; @sifakis10; @goessler11; @moe09; @GGM+07]. An interaction system consists of components which cooperate through so called interactions. An interaction specifies a multiway cooperation among components by connecting different interfaces (called ports) of different components. The model is defined in two layers. The first layer, the interaction model, specifies the components, their interfaces and the communication between them. The second layer, the interaction system, describes the behavior of the components by labeled transition systems. In contrast to [@interactionSystems] we allow an interaction to be contained in another interaction and do not consider complete interactions.
\[defIM\] Let $K$ be a set of [**components**]{} and $\{A_i\}_{i\in K}$ a family of pairwise disjunct sets of [**ports**]{} of the components in $K$. In the following we assume that $K=\{1,2,\dots,n\}$. An [**interaction**]{} $\alpha$ is a nonempty set of ports from different components, i.e., $\alpha\subseteq\bigcup_{i\in K}A_i$ and for all $i\in K$ $|\alpha\cap A_i|\leq 1$ holds.
An interaction $\alpha_i=\{a_{i_1},a_{i_2},\dots,a_{i_k}\}$ with $a_{i_j}\in A_{i_j}$ ($j\in\{1,2,\dots,k\}$) denotes a possible cooperation among the components $i_1,\dots,i_k$ via their respective ports. A set $Int$ of interactions is called [**interaction set (for $K$)**]{}, if each port appears in at least one interaction in $Int$, i.e., $\bigcup_{i\in K}A_i=\bigcup_{\alpha\in Int}\alpha$. The tuple $IM=(K,\{A_i\}_{i\in K},Int)$ is called [**interaction model**]{} if $Int$ is an interaction set for $K$.
\[ex1\] Let $r>0$ be a natural number and $K=\{S,c_1,c_2,\dots,c_r\}$ a set of components. $S$ models a server with a set of ports $A_S=\{connect,disconnect\}$ where $connect$ models the connection of a client to this server and $disconnect$ models the disconnection. For $1\leq i\leq r$ component $c_i$ models a client with a set of ports $A_{c_i}=\{connect_i,disconnect_i\}$. $connect_i$ models the connection of client $i$ to the server and $disconnect_i$ the disconnection.
For $1\leq i\leq r$ the interaction $connect\_S\_c_i=\{connect,connect_i\}$ models a connection from client $i$ to the server and the interaction $disconnect\_S\_c_i=\{disconnect,disconnect_i\}$ models the disconnection. Let $${Int}= \{connect\_S\_c_i,disconnect\_S\_c_i|1\leq i\leq r\}$$ be a set of interactions. Note that ${Int}$ is an interaction set for $K$, i.e., ${IM}=(K,\{A_i\}_{i\in K},{Int})$ is a well defined interaction model.
Let $IM=(K,\{A_i\}_{i\in K},Int)$ be an interaction model. $Sys=(IM,\{T_i\}_{i\in K})$ is called [**interaction system**]{} where $T_i=(Q_i,A_i,\to_i,q_i^0)$ for $i\in K$ is a labeled transition systems that models the behavior of component $i\in K$. $Q_i$ is a finite state space, $\to_i\subseteq Q_i\times A_i\times Q_i$ a transition relation and $q_i^0\in Q_i$ an initial state. We refer to $T_i$ for $i\in K$ as the [**local behavior**]{} of component $i$ and we denote $q_i{{\xrightarrow{a_i}}}_iq_i'$ instead of $(q_i,a_i,q_i')\in \to_i$.
We say a state $q_i\in Q_i$ [**enables**]{} $a_i$ if there is $q_i'\in Q_i$ with $q_i{{\xrightarrow{a_i}}}_iq_i'$. We denote the set of enabled ports of a state $q_i\in Q_i$ by $en(q_i)$. This is, $en(q_i) = \{a_i\in A_i|\exists_{q_i'\in Q_i}q_i{{\xrightarrow{a_i}}}_iq_i'\}$.
\[ex2\] Let ${IM}=(K,\{A_i\}_{i\in K},{Int})$ be the interaction model from example \[ex1\]. Figure \[clse\] depicts possible local behavior $T_i$ for $i\in K$. This is, the tuple $Sys=(IM,\{T_i\}_{i\in K})$ is a well defined interaction system. We mark initial states by an incoming arrow.
The behavior of an interaction system is defined as follows.
Let $Sys=(IM,\{T_i\}_{i\in K})$ be an interaction system where the interaction model is given by $IM=(K,\{A_i\}_{i\in K},Int)$. The [**global behavior**]{} of $Sys$ is the transition system $T=(Q,Int,\to,q^0)$ where
- the Cartesian product $Q=\prod_{i\in K}Q_i$ is the [**global state space**]{} which we assume to be order independent,
- $q^0=(q_1^0,\dots,q_n^0)$ is the [**global initial state**]{} and
- $\to\subseteq Q\times Int\times Q$ is the [**global transition relation**]{} with $q{{\xrightarrow{\alpha}}}q'$ if for all $i\in K$:
- $q_i{{\xrightarrow{a_i}}}_iq_i'$ if $\alpha\cap A_i=\{a_i\}$ and
- $q_i=q_i'$ if $\alpha\cap A_i=\emptyset$.
A state $q\in Q$ is called a [**global state**]{}. Globally, a transition $q{{\xrightarrow{\alpha}}}q'$ can be performed if each port in $\alpha$ can be performed in the state of the local behavior of its respective component.
Let $Sys$ be an interaction system and $T=(Q,{Int},\to,q^0)$ the associated global transition system. A global state $q\in Q$ is called [**reachable**]{} iff there is a path in $T$ that leads from the initial state $q^0$ to $q$. Given an interaction system ${Sys}$ and a global state $q$, the [**reachability problem**]{} consists of deciding whether or not $q$ is reachable in the global behavior of ${Sys}$.
In order to define subclasses of interaction systems we study architectural constraints with respect to the communication structure between components, i.e., our constraints are defined on the interaction model and are independent from the behavior of the components. The communication structure is defined by an undirected graph the nodes of which are components that are connected by an edge if these components are able to interact.
Let ${IM}=(K,\{A_i\}_{i\in K},{Int})$ be an interaction model with $|K|=n$. The [**interaction graph**]{} $G=(K,E)$ of ${IM}$ is an undirected graph with $\{i,j\}\in E$ ($i\not=j$) iff there is an interaction $\alpha\in{Int}$ with $\alpha\cap A_i\not=\emptyset$ and $\alpha\cap A_j\not=\emptyset$, i.e., if there is an interaction in which both components participate.
An interaction model ${IM}$ is called [**star-like**]{} iff $G$ is a star, i.e., exactly one node is of degree $n-1$ and all other nodes are of degree $1$. $IM$ is called [**linear**]{} iff $G$ is connected, two nodes are of degree $1$ and any other node is of degree $2$. An interaction system ${Sys}$ is called star-like respectively linear if the interaction model of ${Sys}$ is star-like respectively linear.
Note that star-like and linear interaction systems with a set ${Int}$ of interactions imply that for all $\alpha\in {Int}$ $|\alpha|\leq 2$.
A star-like or linear interaction system can be seen as a system with a simple hierarchical communication structure, e.g., the simple client/server system in Example \[ex1\]. Of course, such systems can be far more complex and thus exhibit a highly branched communication structure. This is, a PSPACE-completeness result for deciding the reachability problem in the subclass of star-like or linear interaction systems implies the PSPACE-completeness of deciding the reachability problem in systems with a hierarchical communication structure.
As deciding the reachability problem in general interaction systems is in PSPACE it follows that the same holds for the classes of linear and star-like systems.
\[ex3\] The interaction graph $G$ of the interaction model ${IM}=(K,\{A_i\}_{i\in K},{Int})$ from Example \[ex1\] is depicted in Figure \[igcs\]. The interaction graph is a star, i.e., ${IM}$ is star-like and thus, every interaction system, particularly ${Sys}$ in Example \[ex2\], that contains ${IM}$ is star-like.
(3.5,1.25) (A)(1.75,1.0)[$S$]{} (B)(0.0,0.0)[$c_1$]{} (C)(1.0,0.0)[$c_2$]{} (D)(2.0,0.0)[$c_3$]{} (2.55,0.0)[…]{} (E)(3.5,0.0)[$c_r$]{} (A,B) (A,C) (A,D) (A,E)
\[exline\] This example illustrates a linear interaction system. We consider a simple communication pipeline consisting of $n$ stations. Station one initiates passing a message to station two, station two passes the message to station three and so on. If the message arrives at station $n$ then station $n$ passes an acknowledge message, on the same way, back to station one.
Let ${IM}=(K,\{A_i\}_{i\in K},{Int})$ be the interaction model with components $K=\{s_1,s_2,\dots,s_n\}$ for $n\geq 2$ where $s_i$ models station $i$ for $1\leq i\leq n$. A station $s_i$ with $1<i<n$ can receive a message ($rec\_m_i$), pass the message forward ($send\_m_i$), receive an acknowledge ($rec\_a_i$) and pass the acknowledge forward ($send\_a_i$). Station $s_1$ can only send the initial message and receive the acknowledge and station $s_n$ can only receive a message and send an acknowledge. This is, the port sets of the components are defined as follows. $$\begin{array}{lcl}
A_{s_1} &=& \{send\_m_1,rec\_a_1\} \\
A_{s_i} &=& \{rec\_m_i,send\_m_i,rec\_a_i,send\_a_i\}, \; 1<i<n \\
A_{s_n} &=& \{rec\_m_n,send\_a_n\}
\end{array}$$ The interaction set ${Int}$ is given by the following interactions. $$\begin{array}{lcl}
send\_message_i &=& \{send\_m_i,rec\_m_{i+1}\}, \; 1\leq i<n \\
send\_acknowledge_i &=& \{send\_a_i,rec\_a_{i-1}\}, \; 1<i\leq n
\end{array}$$ Let $Sys=(IM,\{T_i\}_{i\in K})$ be the interaction system with local behavior depicted in Figure \[pipebeh\].
The interaction graph $G$ of ${IM}$ is depicted in Figure \[igline\]. $G$ forms a line of components. Thus, ${IM}$ is a linear interaction model and ${Sys}$ is a linear interaction system.
(5.0,0.75) (A)(0.0,0.0)[$s_1$]{} (B)(1.5,0.0)[$s_2$]{} (C)(3.0,0.0)[$s_3$]{} (D)(5.0,0.0)[$s_n$]{} (A,B) (B,C) (C,D)
PSPACE-completeness of Reachability in Linear Systems {#lin}
=====================================================
In the following we give a reduction from the accepting problem in linear bounded Turing machines to the reachability problem in linear interaction systems. We use the following syntax for a Turing machine but we refrain from repeating the well known semantics (see [@complexGuide] for details).
A $4$-tuple $M=(\Gamma,\Sigma,P,\delta)$ is called [**deterministic Turing machine**]{} ([**DTM**]{}) where
- $\Gamma$ is a finite set of [**tape symbols**]{},
- $\Sigma\subseteq\Gamma$ is a set of [**input symbols**]{} with a distinguished [**blank symbol**]{} $b\in\Gamma\setminus\Sigma$,
- $P$ is a finite set of [**states**]{}, including an [**initial state**]{} $p^0$ and two [**halt states**]{} $p^Y$ and $p^N$ and
- $\delta$ is the [**transition function**]{} with $\delta:(P\setminus\{p^Y,p^N\})\times\Gamma\to P\times\Gamma\times\{-1,+1\}$.
We consider a both-sided infinite tape with cells labeled by integers. Given an input $x\in\Sigma^*$ written on the cells labeled $1$ through $|x|$ we assume $M$ to be initially in the initial state $p^0$ and the tape head pointing at cell $1$. For a string $x\in\Sigma^*$ with $|x|=n$ we denote the $i$th letter in $x$ by $x^i$ for $1\leq i\leq n$.
A DTM $M$ is called [**linear bounded**]{} if no computation on $M$ uses more than $n+1$ tape cells, where $n$ is the length of the input string. A [**configuration**]{} of a bounded DTM $M$ is denoted by $(p;\gamma_0,\dots,\underline{\gamma_i},\dots,\gamma_{n+1})$ where $M$ is in state $p$, $\gamma_j$ is the tape symbol in cell $0\leq j\leq n+1$ and the tape head is on cell $i$.
The problem [**linear space acceptance**]{} ([**LSA**]{}) has as input a linear bounded DTM $M$ and a finite string $x$ over the input alphabet of $M$. The question is whether $M$ accepts $x$, i.e., does $M$ halt in the state $p^Y$. It is well known that LSA is PSPACE-complete [@complexGuide].
The idea for our reduction is to model the cells of a DTM $M$ by components of an interaction system ${Sys}_M$ and the transition function of $M$ by interactions such that a path in the global behavior of ${Sys}_M$ corresponds to an execution of $M$. In order to calculate the next configuration of $M$ we need the current tape head position, the current tape symbol in the respective cell and the current state of $M$. We model all these informations in each cell, i.e., in order to model the calculation of the next configuration we need interactions between the component that models the cell with the tape head and the respective components that model the neighboring cells.
Let $M=(\Gamma,\Sigma,P,\delta)$ be a linear bounded DTM and $x\in\Sigma^*$ an input with $|x|=n$. Let ${Sys}_M=({IM}_M,\{T_i\}_{i\in {K}})$ be an interaction system with interaction model ${IM}=({K},\{A_i\}_{i\in {K}},{Int})$ such that ${K}=\{0,\dots,n+1\}$.
The set of ports $A_i$ for a component $i\in K$ with $1\leq i\leq n$ is given by $$A_i = \{(p,\gamma)_i^1,(p,\gamma)_i^2|p\in P\setminus\{p^Y,p^N\},\gamma\in\Gamma\}.$$ $(p,\gamma)_i^1$ models that the tape head moves away from cell $i$ where $\gamma$ is the current tape symbol in this cell and $M$ is in state $p$. Analogously, $(p,\gamma)_i^2$ models that the tape head moves onto cell $i$ where $\gamma$ is written and $M$ is in state $p$.
Because of $M$ being linear bounded, we know that $\delta$ does not move the tape head from cell $0$ to the left respectively from cell $n+1$ to the right. Thus, we can omit ports in $A_0$ and $A_{n+1}$ that model a head movement from or onto cell $-1$ and $n+2$. $A_0$ is given by $$\begin{aligned}
A_0 &=& \{(p,\gamma)_0^1|p\in P\setminus\{p^Y,p^N\},\gamma\in\Gamma,\neg\exists_{p',\gamma'}\delta(p,\gamma)=(p',\gamma',-1)\} \cup \\
&& \{(p,\gamma)_0^2|p\in P\setminus\{p^Y,p^N\},\gamma\in\Gamma,\neg\exists_{p',\gamma'}\delta(p,\gamma)=(p',\gamma',1)\}.
\end{aligned}$$ $A_{n+1}$ is defined analogously. The set of interactions is given by $${Int}=\{\{(p,\gamma)_i^1,(p,\gamma)_{i+T}^2\}|\exists_{p',\gamma'}\delta(p,\gamma)=(p',\gamma',T),0\leq i+T\leq n+1\}.$$
For $i\in {K}$ let $T_i=(Q_i,A_i,\to_i,q_i^0)$ be the local behavior of component $i$ with $Q_i=\{(p,\gamma)|p\in P\cup\{s\},\gamma\in\Gamma\}$ where $s$ is an auxiliary symbol that is not included in $P$. $(p,\gamma)\in Q_i$ with $p\not=s$ models that the tape head is currently on cell $i$ and the current tape symbol in this cell is $\gamma$. $(s,\gamma)$ models that $\gamma$ is the content of cell $i$ and the tape head is not on this cell. The local initial states are derived from the initial word on the tape, i.e., $q_0^0=(s,b)$, $q_1^0=(p^0,x^1)$, $q_i^0=(s,x^i)$ for $2\leq i\leq n$ and $q_{n+1}^0=(s,b)$. For $i\in {K}$ let $\to_i$ be the union of the following transitions.
1. \[onCell\]For all $\gamma,\gamma'\in\Gamma$ and $p\in P\setminus\{p^Y,p^N\}$ let $(p,\gamma)\xrightarrow{(p,\gamma)_i^1}_i(s,\gamma')$ if there are $p'\in P$ and $T\in\{-1,1\}$ such that $\delta(p,\gamma)=(p',\gamma',T)$.
2. \[toCell\]For all $\gamma,\tilde{\gamma}\in\Gamma$, $p\in P\setminus\{p^Y,p^N\}$ and $p'\in P$ let $(s,\tilde{\gamma})\xrightarrow{(p,\gamma)_i^2}_i(p',\tilde{\gamma})$ if there are $\gamma'\in\Gamma$ and $T\in\{-1,1\}$ such that $\delta(p,\gamma)=(p',\gamma',T)$.
The transitions described in a) model the impact of $\delta$ on cell $i$ if the tape head is on this cell. Let $M$ be in state $p$ and the tape head on cell $i$ reading $\gamma$, i.e., $T_i$ is in the state $(p,\gamma)$. If $\delta(p,\gamma)=(p',\gamma',T)$ then $\gamma'$ is written and the tape head moves to a neighboring cell, i.e., $T_i$ moves to the state $(s,\gamma')$. On the other hand, the transitions described in b) model a head movement onto cell $i$. Let $\tilde{\gamma}$ be the current tape symbol on cell $i$, i.e., $T_i$ is in state $(s,\tilde{\gamma})$ before the head moves. After the movement let $M$ change its state to $p'$, i.e., $T_i$ moves to the state $(p',\tilde{\gamma})$.
${Sys}_M$ satisfies the conditions of an interaction system: every port of a component occurs in at least one interaction. Let $i\in {K}$, $(p,\gamma)^1_i\in A_i$ and $\delta(p,\gamma)=(p',\gamma',T)$ then $0\leq i+T\leq n+1$ and $\{(p,\gamma)^1_i,(p,\gamma)^2_{i+T}\}\in {Int}$. For $(p,\gamma)^2_i\in A_i$ is $0\leq i-T\leq n+1$ and $\{(p,\gamma)^1_{i-T},(p,\gamma)^2_i\}\in {Int}$.
${Sys}_M$ has a linear communication structure because every component $1\leq i\leq n$ only interacts with its neighboring components $i-1$ and $i+1$.
The reduction is polynomial in the size of an underlying DTM $M=(\Gamma,\Sigma,P,\delta)$, since $|{Int}|\leq |P|\cdot|\Gamma|$ and for all $i\in {K}$ $|A_i|\leq 2\cdot|P|\cdot|\Gamma|$ and $|Q_i|\leq (|P|+1)\cdot|\Gamma|$.
Let $M=(\Gamma,\Sigma,P,\delta)$ be a linear bounded DTM, $x\in\Sigma^*$ with $|x|=n$ an input for $M$ and ${Sys}_M$ the associated linear interaction system. $M$ accepts $x$ iff a global state $q=(q_0,\dots,q_{n+1})$ is reachable in ${Sys}$ such that there is $i\in\{0,\dots,n+1\}$ with $q_i=(p^Y,\gamma)$ for a tape symbol $\gamma\in\Gamma$.
We prove this theorem by giving an isomorphism, with respect to transitions in ${Sys}_M$ and transitions among configurations in $M$, between global states of ${Sys}_M$ and configurations of $M$. The statement of the theorem then follows by induction as the isomorphism maps the initial configuration of $M$ to the initial state of ${Sys}_M$.
Let $R$ be the set of configurations of $M$. We map $(p;\gamma_0,\dots,\underline{\gamma_i},\dots,\gamma_{n+1})\in R$ to a global state $q=(q_0,\dots,q_{n+1})$ such that $q_i=(p,\gamma_i)$ and $q_j=(s,\gamma_j)$ for $j\not=i$. Let $Q'$ be the set of global states that correspond to the configurations in $R$. It is clear that this mapping is a bijection between $R$ and $Q'$.
Let $(p;\gamma_0,\dots,\underline{\gamma_i},\dots,\gamma_{n+1})\in R$ and $q=(q_0,\dots,q_{n+1})\in Q'$ be the associated state in ${Sys}_M$. Let $\delta(p,\gamma_i)=(p',\gamma_i',T)$, i.e., the next configuration in $M$ is $(p';\gamma_0,\dots,\gamma_i',\underline{\gamma_{i+1}},\dots,\gamma_{n+1})\in R$ if $T=1$ (the case $T=-1$ is treated analogously). The only enabled port in component $i$ is $(p,\gamma_i)_i^1$, then the only enabled interaction in $q$ is $\{(p,\gamma_i)_i^1,(p,\gamma_i)_{i+T}^2\}$. Thus, component $i$ reaches the state $(s,\gamma_i')$ and component $i+T$ the state $(p',\gamma_{i+T})$. The resulting global state $q'$ corresponds to the respective configuration in $M$. The fact that the inverse of the mapping is also a homomorphism can be shown analogously.
An instance of the reachability problem is an interaction system ${Sys}$ and a global state $q$. The interaction system ${Sys}_M$ for a linear bounded DTM $M$ and an input $x$ can be extended such that a distinguished global state is reached if $M$ halts on $x$. This can be achieved by a technique that is used in [@orgtreelike] for tree-like interaction systems. The idea is to invoke, starting from the component that reached $(p^Y,\gamma)$, that each component shall reach a distinguished state. This invocation can be propagated through neighboring components.
PSPACE-completeness of Reachability in Star-Like Systems {#star}
========================================================
Here we show that deciding the reachability problem in the class of star-like interaction systems is PSPACE-complete by providing a reduction from a general interaction systems ${Sys}$ to a star-like systems ${Sys}'$. The idea of the reduction is to construct a “control component” $cc$ that forms the center of the star structure in ${Sys}'$ and is surrounded by the components of ${Sys}$. An interaction in ${Sys}$ is modeled by multiple interactions in ${Sys}'$. The execution of an interaction in ${Sys}$ then corresponds to the execution of a sequence of interactions in ${Sys}'$ that is coordinated by $cc$ and achieved in two steps. Let $\alpha$ be an interaction in ${Sys}$.
1. In a first step $cc$ interacts with each component that participates in $\alpha$ and checks whether the respective port in $\alpha$ is enabled without changing the local states of the components. If this check fails then $cc$ returns to its initial state.
2. If the check succeeds then $cc$ interacts with each respective component on the ports in $\alpha$, i.e., a global transition in ${Sys}$ that is labeled by $\alpha$ is simulated.
Let $Q=\prod_{i\in {K}}Q_i$ be the global state space of ${Sys}$ then we have a global state space $\prod_{i\in {K}\cup\{cc\}}Q_i$ for ${Sys}'$ with the property that $q\in Q$ is reachable in ${Sys}$ iff a state $q'$ is reachable in ${Sys}'$ such that $q'$ equals $q$ up to the local state of the component $cc$. Since reachability in general interaction systems is PSPACE-complete, the consequence of this transformation is the PSPACE-completeness of reachability in star-like interaction systems.
Let ${Sys}=({IM},\{T_i\}_{i\in {K}})$ be an interaction system with interaction model ${IM}=({K},\{A_i\}_{i\in {K}},{Int})$ and ${Sys}'=({IM}',\{T_i'\}_{i\in {K}'})$ be an interaction system with interaction model ${IM}=({K}',\{A_i'\}_{i\in {K}'},{Int}')$.
Let ${K}'={K}\cup\{cc\}$, where $cc$ is a control component that coordinates sequences of interactions in ${Int}'$ that correspond to interactions in ${Int}$. For $i\in K$ let $A_i'=A_i\cup\{a_i^{ok},a_i^{\neg ok}|a_i\in A_i\}$. $a_i^{ok}$ respectively $a_i^{\neg ok}$ models that component $i$ is in a local state that enables respectively does not enable the port $a_i\in A_i$. The set of ports $A_{cc}$ of component $cc$ is given by $$A_{cc}=\{a\_i_{cc}^{ok},a\_i_{cc}^{\neg ok},a\_i_{fire}^{cc}|i=1,\dots,n,a_i\in A_i\}\cup \{\alpha_{cc}|\alpha\in {Int}\}.$$ Let $i\in K$ and $a_i\in A_i$ a port in $i$ then $a\_i_{cc}^{ok}$ models that component $i$ currently enables $a_i$ and $a\_i_{cc}^{\neg ok}$ models that $a_i$ is currently not enabled by $i$. $a\_i_{fire}^{cc}$ models that component $i$ performs a transition labeled by $a_i$. For an interaction $\alpha\in{Int}$ the port $\alpha_{cc}$ models the initiation of a process that checks whether $\alpha$ is enabled by the respective components and, if applicable, coordinates that all ports in $\alpha$ interact one after another.
The set of interactions ${Int}'$ is given by $$\begin{aligned}
{Int}' &=& \{\{a_i^{ok},a\_i_{cc}^{ok}\},\{a_i^{\neg ok},a\_i_{cc}^{\neg ok}\},\{a_i,a\_i_{cc}^{fire}\}|a_i\in A_i,i=1,\dots,n\}\cup \{\{\alpha_{cc}\}|\alpha\in {Int}\}.
\end{aligned}$$
The local behavior of $i\in {K}$ is given by $T_i'=(Q_i,A_i',\to_i',q_i^0)$ with $$\begin{array}{ll}
\to_i'= & \to_i\cup \{(q_i,a_i^{ok},q_i)|q_i\in Q_i\wedge a_i\in en(q_i)\}\cup \{(q_i,a_i^{\neg ok},q_i)|q_i\in Q_i\wedge a_i\notin en(q_i)\}.
\end{array}$$
$T_i'$ extends $T_i$ such that for each port $a_i\in A_i$ there is a loop on each state $q_i\in Q_i$ that is labeled by $a_i^{ok}$ if $q_i$ enables $a_i$ and by $a_i^{\neg ok}$ otherwise. These transitions are used to check whether or not each port of an interaction $\alpha\in{Int}$ is enabled in a global state of ${Sys}'$ without changing the local state of the respective components.
Let $\alpha^j=\{a_{j_1},\dots,a_{j_{|\alpha^j|}}\}\in{Int}$. Figure \[cc\] depicts the part of the local behavior $T_{cc}=(Q_{cc},A_{cc},\to_{cc},q_{cc}^0)$ of component $cc$ that coordinates a test that checks whether each port in $\alpha^j$ is enabled in ${Sys}'$ and, if applicable, enables ports that can interact with each port in $\alpha^j$. $q_{cc}^0$ is marked by an incoming arrow.
(1.0,4.75) (A)(1.0,4.0) (A) (B)(1.0,3.0) (C)(1.0,2.0) (D)(1.0,1.0) (E)(1.0,0.0) (F)(2.5,0.0) (G)(2.5,1.0) (H)(2.5,2.0) (I)(2.5,3.0) (J)(2.5,4.5)
(A,B) [$\alpha^j_{cc}$]{} (B,C) [$a\_{j_1}_{cc}^{ok}$]{} (C,D) [$a\_{j_2}_{cc}^{ok}$]{} (D,E) (E,F) [$a\_{j_{|\alpha^j|}}_{cc}^{ok}$]{}
(B,180,1.0,A,180,1.0)[$a\_{j_1}_{cc}^{\neg ok}$]{} (C,180,2.0,A,170,1.0)[$a\_{j_2}_{cc}^{\neg ok}$]{} (D,180,3.0,A,160,1.0)[$a\_{j_3}_{cc}^{\neg ok}$]{} (E,180,4.0,A,150,1.0)[$a\_{j_{|\alpha^j|}}_{cc}^{\neg ok}$]{}
(F,G) [$a\_{j_1}_{cc}^{fire}$]{} (G,H) [$a\_{j_2}_{cc}^{fire}$]{} (H,I) (I,A)[$a\_{j_{|\alpha^j|}}_{cc}^{fire}$]{}
(A,J) [$\alpha^k_{cc}$, $k\not=j$]{}
Each port of ${Sys}'$ occurs in at least one interaction, i.e., ${Sys}'$ satisfies the conditions of an interaction system. It is clear that ${Sys}'$ is star-like because each component that originated from ${Sys}$ interacts only with the control component $cc$.
Furthermore, the size of ${Sys}'$ is polynomial in the size of ${Sys}$. $|K'|=|K|+1$, $|{Int}'|=|{Int}|+\sum_{i\in K}3\cdot|A_i|$ and for $i\in K$ holds $|A_i'|=3\cdot|A_i|$ and $|\to_i'|=|\to_i|+|Q_i|\cdot|A_i|$. For $cc\in K'$ holds $|A_{cc}|=|{Int}|+\sum_{i\in K}3\cdot|A_i|$, $|Q_{cc}|=1+\sum_{\alpha\in{Int}}2\cdot|\alpha|$ and $|\to_{cc}|=\sum_{\alpha\in{Int}}(3\cdot|\alpha|+1)$.
Let ${Sys}$ be an interaction system with components $K$ and ${Sys}'$ the associated star-like interaction system. A global state $q$ is reachable in ${Sys}$ iff a global state $q'$ is reachable in ${Sys}'$ such that $q_i=q_i'$ for $i\in K$ and $q_{cc}'=q_{cc}^0$.
Let $q$ be a state in the global behavior $T$ of ${Sys}$ and $q'$ be the state in the global behavior $T'$ of ${Sys}'$ where $q_i=q_i'$ for $i\in K$ and $q_{cc}'=q_{cc}^0$, i.e., component $cc$ is in its initial state. Consider $\alpha^j=\{a_{j_1},\dots,a_{j_{|\alpha^j|}}\}\in {Int}$ such that each port in $\alpha^j$ is enabled in $q$, i.e., all local states $q_l'$, $l=j_1,\dots,j_{|\alpha^j|}$ in $q'$ enable the ports $a_l^{ok}$ and $a_l$ and do not enable $a_l^{\neg ok}$. $q'$ enables the interaction $\{\alpha_{cc}^j\}$. If this interaction is performed then the only possible sequence of interactions results in a state $\tilde{q}'$ with $\tilde{q}_i=\tilde{q}_i'$ for $i\in K$ and $\tilde{q}_{cc}'=q_{cc}^0$. Let there be a port in $\alpha^j$ that is not enabled in $q$, e.g., $q_l$ with $l\in\{j_1,\dots,j_{|\alpha^j|}\}$ does not enable $a_l$ then $q_l'$ does enable $a_l^{\neg ok}$ and not $a_l^{ok}$. If $\{\alpha_{cc}^j\}$ performed in $q'$ then the only possible sequence of interactions in ${Sys}'$ leads back to state $q'$. For the global initial states $q^0$ of ${Sys}$ and ${q^0}'$ of ${Sys}'$ holds that $q^0_i={q^0_i}'$ for $i\in K$ and ${q^0_{cc}}'$ is the initial state of the local behavior of component $cc$. The “if” part follows by induction over paths in the global behavior of ${Sys}$. The “and only if” part follows analogously.
Conclusion and Related Work {#con}
===========================
We investigated complexity issues for classes of interaction systems that are relevant in various applications. One with a linear the other with a star-like communication pattern. We showed that even for these simply structured systems deciding the reachability problem is PSPACE-complete. These results strengthen PSPACE-completeness results of the reachability problem in general interaction systems [@everythingIsPSPACE]. The formalism of interaction systems is very basic, and thus our results are easily applicable to other formalisms that model cooperating systems. Our results justify techniques that are based on a sufficient condition and establish reachability or reachability dependent system properties in subclasses of cooperating systems that are defined by a restricted communication structure that forms a star or a line or in respective superclasses, which are sketched in the following.
[@bernardoArch] examined a process algebra based on an architectural description language called *PADL* and considers deadlock-freedom in systems with a tree-like communication pattern (a proper superclass of systems with a star-like or linear pattern). The technique is based on a compatibility condition that is tested among pairs of cooperating subsystems, i.e., the composite behavior of two subsystems is weak bisimilar to the behavior of one of the components. An efficient technique based on a sufficient conditions for establishing deadlock-freedom in interaction systems with a star-like communication pattern is introduced in [@Lambertz09] where, similar to [@bernardoArch], a compatibility condition based on branching bisimilarity is tested. A sufficient condition for establishing deadlock-freedom for the subclass of tree-like interaction systems is described in [@treeLike] where a condition is tested on the reachable state spaces of pairs of interacting components. In [@Lambertz11] the condition in [@treeLike] is extended such that deadlock-freedom can be established in a proper superclass of tree-like interaction systems. Hennicker et al. proposed in [@BHH+06; @HJK08] a technique to construct so called *observable* behavior of a cooperating system with an acyclic communication pattern which can be used to establish certain system properties. [@roscoeArch] describes a general communication graph for CSP models and shows how tree structures can be constructed by merging several processes. *Communicating Sequential Processes* are introduced in [@hoare] where a directed communication structure based on input/output communication is considered. It is argued that communicating processes, if a directed input/output communication structure forms a rooted tree, can not deadlock.
[^1]: Corresponding author - Phone: (49) 621-181 2564 - Fax: (49) 621-181 2560
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Birefringent metasurfaces are two-dimensional structures capable of independently controlling the amplitude, phase and polarization of orthogonally polarized incident waves. In this work, we propose a in-depth discussion on the mathematical synthesis of such metasurfaces. We compare two methods, one that is rigorous and based on the exact electromagnetic fields involved in the transformation and one that is based on approximate reflection and transmission coefficients. We next validate the synthesis technique in metasurfaces performing the operations of half- and quarter-wave plates, polarization beam splitting and orbital angular momentum multiplexing.'
author:
- 'Karim Achouri , Guillaume Lavigne, and Christophe Caloz, [^1]'
bibliography:
- 'LIB.bib'
title: |
Synthesis and Applications\
of Birefringent Metasurfaces
---
Introduction
============
Birefringence, also called double refraction, is the physical property of an anisotropic medium to exhibit an angle dependent refractive index [@saleh2007fundamentals]. This phenomenon, first observed in crystals more than 300 years ago [@Bartho], has already lead to the realization of several major optical components and applications. More recently, metasurfaces [@yu2014flat; @lin2014dielectric; @Pfeiffer2013a; @023902; @kim2014optical; @achouri2014general], the two dimensional counterparts of metamaterials, have seen an important rise of interest due to their rich potential in the transformation of electromagnetic fields. Combined with the orthogonality property of $x$ and $y$ polarized waves, birefringent structures, and especially metasurfaces, have the capability to independently control the amplitude, phase and polarization of two orthogonal electromagnetic waves, leading to a wealth of possible applications at optical and microwave frequencies.
In this work, we propose an in-depth discussion on the synthesis of birefringent metasurfaces. This discussion is based on the general bianisotropic metasurface synthesis technique developed in [@achouri2014general]. We compare two different synthesis methods, one that is rigorous and one that is based on paraxial approximate transmission and reflection coefficients. We also report the implementation details of several metasurfaces that were only briefly presented in [@art1; @AchouriEPJAM]. These metasurfaces perform the operations of half-wave plate [@ding2015broadband; @pors2013broadband; @lin2014dielectric], quarter-wave plate [@yu2012broadband; @zhao2011manipulating], polarization beam splitting [@pors2013gap; @farmahini2013birefringent; @lee2014semiconductor] and orbital angular momentum [@PhysRevApplied.2.044012; @karimi2014generating; @capasso1; @chen2015creating; @wang2015ultra] multiplexing.
The paper is organized as follows. In the second section, the mathematical synthesis as well as the physical realization of metasurfaces are addressed. Two different approaches for the mathematical synthesis are discussed and compared. In the third section, we present the implementation of the four metasurfaces introduced above.
Metasurface Design
==================
Mathematical Synthesis
----------------------
The metasurface synthesis technique in [@achouri2014general] stems from the continuity relations, initially derived by Idemen [@Idemen1973], which express the discontinuities of the electromagnetic fields in the presence of a spatial discontinuity, such as a metasurface. In the simple case of a monoanisotropic (zero magnetoelectric coupling coefficients) birefringent metasurface given in terms of its transverse diagonal susceptibility tensors, ${\overline{\overline{\chi}}}_\text{ee}$ and ${\overline{\overline{\chi}}}_\text{mm}$, the continuity relations read
\[eq:BC\_plane\] $$\begin{aligned}
\hat{z}\times\Delta{\boldsymbol{H}}
&=j\omega\epsilon_0{\overline{\overline{\chi}}}_\text{ee}\cdot{\boldsymbol{E}}_\text{av},\label{eq:BC_plane_1}\\
\Delta{\boldsymbol{E}}\times\hat{z}
&=j\omega\mu_0 {\overline{\overline{\chi}}}_\text{mm}\cdot{\boldsymbol{H}}_\text{av},\label{eq:BC_plane_2}\end{aligned}$$
where it is assumed that the metasurface lies in the $x-y$ plane at $z=0$. In , the differences of the electric and magnetic fields between both sides of the metasurface (expressed by the operator $\Delta$) are related to the metasurface susceptibilities excited by the average electric and magnetic fields (denoted by the subscript ‘av’). The system of can be easily solved to express the susceptibilities in terms of the specified fields. Due to the orthogonality between $x$ and $y$ polarizations, the solutions split into two independent sets which are respectively given by
\[eq:X\_diag\] $$\chi_{\text{ee}}^{xx}=\frac{-\Delta H_{y}}{j\omega\epsilon_0 E_{x,\text{av}}},\qquad
\chi_{\text{mm}}^{yy}=\frac{-\Delta E_{x}}{j\omega\mu_0 H_{y,\text{av}}},\label{eq:Xx}$$ $$\chi_{\text{ee}}^{yy}=\frac{\Delta H_{x}}{j\omega\epsilon_0 E_{y,\text{av}}},\qquad
\chi_{\text{mm}}^{xx}=\frac{\Delta E_{y}}{j\omega\mu_0 H_{x,\text{av}}}.\label{eq:Xy}$$
At this stage, the metasurface is completely defined by the susceptibilities in and performs the required transformation between the incident, reflected and transmitted waves [@achouri2014general; @AchouriEPJAM]. The birefringent operation leverages the property of orthogonality between and .
In order to realize the metasurface, one has to find the appropriate shapes of the scattering particles. Therefore, it is convenient to find the transmission and reflection coefficients of the metasurface which could then be related to the transmission and reflection coefficients of each scattering particle obtained via full-wave simulations. To obtain these coefficients, it is assumed that the metasurface is illuminated by a normally incident plane wave and that it reflects and transmits normally propagating plane waves (either $x$ or $y$ polarized). The corresponding electric and magnetic fields are thus, respectively, given by $E_\text{i}=e^{-jkz}, E_\text{t}=Te^{-jkz}$ and $E_\text{r}=Re^{jkz}$, and $H_\text{i}=e^{-jkz}/\eta_0, H_\text{t}=Te^{-jkz}/\eta_0$ and $H_\text{r}=-Re^{jkz}/\eta_0$. These fields are then substituted into relations , which are then solved for the coefficients $T$ and $R$, which read [@achouri2014general]
\[eq:RTcoef\] $$\begin{aligned}
T&=\frac{4+\chi_{\text{ee}}\chi_{\text{mm}}k_0^2}{(2+jk_0\chi_{\text{ee}})(2+jk_0\chi_{\text{mm}})},\\
R&=\frac{2jk_0\left(\chi_{\text{mm}}-\chi_{\text{ee}}\right)}{\left(2+jk_0\chi_{\text{ee}}\right)\left(2+jk_0\chi_{\text{mm}}\right)},
\label{eq:RTcoef2}\end{aligned}$$
where $k_0$ is the free space wave number. Using the susceptibilities from or will yield the transmission and reflection coefficients for $x$ or $y$ polarizations, respectively. Since a monoanisotropic metasurface is necessarily symmetric with respect to the $z$ direction [@asadchy2016metasurfaces], the reflection coefficients of our metasurface are the same from both sides of the structure.
Relations can be reversed to express the susceptibilities in terms of the transmission and reflection coefficients as
\[eq:chi\_Sparam\] $$\begin{aligned}
\chi_{\text{ee}}&=\frac{2j\left(T+R-1\right)}{k_0\left(T+R+1\right)},\\
\chi_{\text{mm}}&=\frac{2j\left(T-R-1\right)}{k_0\left(T-R+1\right)},\end{aligned}$$
where $\chi_{\text{ee}}^{xx}$ and $\chi_{\text{mm}}^{yy}$ are found assuming that $T$ and $R$ are the coefficients of $x$-polarized waves. Alternatively, $\chi_{\text{ee}}^{yy}$ and $\chi_{\text{mm}}^{xx}$ are found when $T$ and $R$ are the coefficients of $y$-polarized waves. Although Eqs. can be rigorously used to synthesize metasurfaces, relations suggest an alternative synthesis technique which would consists in specifying the required transformation in terms of transmission and reflection coefficients instead of the tangential electromagnetic fields, as would be done in .\
In what follows, we will compute the responses of metasurfaces synthesized using the methods based on relations and and compare them. Let us consider a reflection-less metasurface that refracts at $45^\circ$ a normally incident $x$-polarized plane wave. The electromagnetic fields of the incident wave are, at $z=0$, given by $E_{i}^x=1, H_{i}^y=1/\eta_0$, while the fields of the transmitted wave are $E_{t}^x=\sqrt{2}/2e^{-jk_0\sqrt{2}/2x}, H_{t}^y=e^{-jk_0\sqrt{2}/2x}/\eta_0$. The first metasurface synthesis method consists in substituting these fields into , which results in the following susceptibilities
\[eq:Sin1\] $$\begin{aligned}
\chi_{\text{ee}}^{xx}&=\frac{4j}{k_0}\left (\frac{e^{-jk_0\sqrt{2}/2x}-1}{2+\sqrt{2}e^{-jk_0\sqrt{2}/2x}}\right ),\\
\chi_{\text{mm}}^{xx}&=\frac{j}{k_0}\left (\frac{\sqrt{2}e^{-jk_0\sqrt{2}/2x}-2}{1+e^{-jk_0\sqrt{2}/2x}}\right ).\end{aligned}$$
The second synthesis method, based on relations , seems a priori unsuitable for such a kind of transformation (i.e. refraction) since relations and were obtained assuming that all waves impinging on or scattered by the metasurface are propagating normally to the structure, which is obviously contradictory with the concept of refraction. Indeed, this second synthesis technique rigorously applies *only* to normally propagating waves, but it may also be used as an approximation to synthesize refractive metasurfaces in the case of small refraction angles, i.e. paraxial approximation, as will be shown next. In fact, this metasurface synthesis technique allows one to obtain the material properties, through relations , from the transmission and reflection coefficients that would be initially defined using the complex amplitude transmittance method [@saleh2007fundamentals] (in the case of zero reflection) or, more generally, the momentum transformation technique introduced in [@Salem2013c].
In the case of a metasurface illuminated by a normally impinging incident wave, the condition of zero reflection, that may be deduced from with $R=0$, is that $\chi_{\text{ee}}=\chi_{\text{mm}}=\chi$. This reduces equations and to $$T=\frac{2-jk_0\chi}{2+jk_0\chi}$$ and $$\label{eq:X_T}
\chi=\frac{2j\left(T-1\right)}{k_0\left(T+1\right)},$$ respectively. Because this metasurface is assumed to be reflection-less, the complex amplitude transmittance method can be used to define the parameter $T$ in , that is to say $$\label{eq:T_opt}
T =\frac{\Psi_t}{\Psi_i},$$ where $\Psi_i$ and $\Psi_t$ are the phase profiles, projected on the metasurface plane, of the incident and the transmitted waves, respectively. Since the incident wave is normally impinging on the metasurface, we have that $\Psi_i = 1$. The transmission coefficient is then simply defined by $T =\Psi_t=e^{-jk_0\sqrt{2}/2x}$, which transforms to $$\label{eq:X_approx}
\chi=\frac{2}{k_0}\tan{\left(\frac{k_0 x}{2\sqrt{2}}\right)}.$$ Now, let us compare the susceptibilities in , obtained from the first synthesis method, which is rigorous, and the susceptibilities in , derived from the second method, which is approximate. The real and imaginary parts of the electric and magnetic susceptibilities are plotted at the top and bottom of Fig. \[Fig:X\], respectively. The left-hand side of the figure corresponds to the first method while the right-hand side corresponds to the second method. They are two main differences between the two methods. Firstly, the electric and magnetic susceptibilities are different in whereas they are equal in . Secondly, the susceptibilities in are complex with negative imaginary parts corresponding to absorption whereas the susceptibilities in are purely real and thus correspond to a lossless and passive structure.
![Susceptibilities of a metasurface refracting a normally incident plane wave under $45^\circ$. The solid blue lines correspond to real parts while the dashed red lines correspond to imaginary parts. The susceptibilities in the left are obtained with relations , while those in the right are obtained from .[]{data-label="Fig:X"}](X){width="1\columnwidth"}
The reason for the imaginary parts in , and the resulting absorption, is the unequal normal power flow between the incident wave and the transmitted wave [@7274678; @asadchy2016metasurfaces]. Indeed, the transmitted wave, which propagates under $45^\circ$, has a lower normal transmitted power than the normally incident wave. This translates into a reduced transmission efficiency where the excess energy of the incident wave is absorbed by the metasurface.
Another interesting comparison to establish between the two methods is the differences of the transmission and reflection coefficients in and . Substituting the susceptibilities found in into yields the transmission and reflection coefficients, for the first synthesis method, that are plotted on the left-hand side of Fig. \[Fig:TR\] in solid blue lines and in dashed red lines, respectively. The transmission and reflection coefficients of the second synthesis method are simply $T=e^{-jk_0\sqrt{2}/2x}$ and $R=0$, and are plotted on the right-hand side of Fig. \[Fig:TR\].
![Transmission (blue solid line) and reflection (red dashed line) coefficients for the metasurfaces given by the susceptibilities in Fig. \[Fig:X\]. The top and bottom plots correspond to the amplitude and phase of these coefficients, respectively. The plots on the left are obtained with relations , while the ones on the right are obtained from .[]{data-label="Fig:TR"}](TR){width="1\columnwidth"}
It appears that the metasurface synthesized with the first method may be seen as an equivalent amplitude and phase grating in transmission and reflection, while the other metasurface is a simple transmission phase gradient structure. The non-zero reflection coefficient that is plotted in Fig. \[Fig:TR\] (top-left) seems, a priori, contradictory with the prescription of zero reflection specified to obtain relations . In fact, no propagating reflected wave is produced by the metasurface because the $k$-vector of the reflection phase (bottom-left in Fig. \[Fig:TR\]), defined as $k_r = 2\pi/P_r$, where $P_r$ is the period of the reflection phase, is larger than the free space wave number, $k_0$. This means that the reflected wave is an evanescent wave and thus does not propagate. Moreover, the non ideal transmission efficiency that was discussed above and that is responsible for the imaginary parts of the susceptibilities in also contributes to the non-zero reflection coefficient in Fig. \[Fig:TR\] and the loss evidenced by the fact that $|T|^2+|R|^2<1$.
![Full-wave simulated real part of the $H_z\eta_0$ component. The left side corresponds to the first synthesis method with susceptibilities as in while the right side corresponds to the second synthesis method with susceptibilities as in . The results in the top row were obtained using an FDFD code and, in the bottom row, using COMSOL.[]{data-label="Fig:Sims"}](Sims){width="1\columnwidth"}
Finally, let us see how these two synthesis techniques compare by performing full-wave simulations. We have made two slightly different kinds of simulations. One using an home made Finite-Difference Frequency Domain (FDFD) code [@vahabzadeh2016simulation] that simulates an exactly zero-thickness metasurface. The other simulations were made using COMSOL by assuming a thin metasurface of thickness $d=\lambda_0/100$. Note that, in COMSOL, the susceptibilities were introduced by assigning relative permittivity and permeability to a slab which are respectively defined by ${\overline{\overline{\epsilon}}}_r={\overline{\overline{I}}}+{\overline{\overline{\chi}}}_\text{ee}/d$ and ${\overline{\overline{\mu}}}_r={\overline{\overline{I}}}+{\overline{\overline{\chi}}}_\text{mm}/d$ where the division by $d$ dilutes the effect of the susceptibilities over the thickness of the slab [@Idemen1973]. Simulation results showing the real part of $H_z$ are plotted in Fig. \[Fig:Sims\], where the top and bottom rows correspond to the FDFD simulations and COMSOL simulations, respectively. The left- and the right-hand sides correspond to the first and second synthesis techniques, respectively.
As can be seen, the best result (Fig. \[Fig:Sims\] top left) was obtained with the FDFD code and using the rigorous relations in (first synthesis technique). This is not surprising because, in this simulation, the metasurface has exactly zero-thickness and, therefore, it can be rigorously described by the continuity equations . The diffraction efficiency, defined as the transmitted power density at $45^\circ$ divided by the incident power density, is about $99\%$ for the first method and $97\%$ for the second one which, surprisingly, works fairly well except for undesired reflection. The simulations with COMSOL give worse results due to the thickness of the metasurface. For both synthesis techniques, several diffraction orders appear, either in reflection or in transmission. Moreover, due to the thickness, some modes are trapped (guided modes) inside the metasurface, which contributes to further lower the diffraction efficiency into the desired direction to $26\%$ and $40\%$ for the first and second methods, respectively. The rigorous method gives, in that case, worse results than the approximate method which might be explained by the fact that the susceptibilities in are lossy and thus the wave is more attenuated by propagating through the thickness of the metasurface whereas the metasurface obtained with the second method is not lossy at all.
To conclude this section, it must be noted that, while the first synthesis technique is the most rigorous one and gives the best results in FDFD simulation, it remains much more complicated to implement than the second method. This is because the physical realization of these metasurfaces necessarily requires a mapping between the susceptibilities, given in Fig. \[Fig:X\], and the scattering parameters, given in Fig. \[Fig:TR\]. And, as can be seen from the scattering parameters, the realization of the metasurface synthesized with the first method would require implementing non-uniform reflection and transmission coefficients that present different phase gradients, moreover this metasurface would also be lossy. Compare this now to the metasurface synthesized with the second method and that presents a uniform unity transmission coefficient and a phase gradient, it is clear that this second method is much easier to realize and considering the excellent results in FDFD simulation, it is the one that is usually preferred for the realization of most metasurfaces. The metasurfaces presented in the following section of this paper are synthesized based on the second synthesis technique.
Physical Realization
--------------------
In order to fabricate the metasurfaces, the required scattering parameters are discretized with subwavelength resolution. At each lattice site, a scattering particle (or unit cell) is realized such that it exhibits the required scattering behavior. The unit cells are simulated in a commercial software and assuming periodic boundary conditions. The shape of the unit cells are optimized such that the scattering matrices obtained by simulation correspond to the transmission and reflection coefficients given in .
To implement each unit cell, we have used a cascade of three metallic layers (with identical outer layers) held together by two dielectric spacers. This type of unit cells has been shown to present full transmission (assuming lossless material) and a complete $360^\circ$-phase coverage [@PhysRevApplied.2.044011].
Each metallic layer of the unit cell consists in a Jerusalem cross, as shown in Fig. \[Fig:metallicUnitCell\] with all its variable dimensions. All the metasurfaces discussed thereafter have been realized with unit cells of size $d = 6$ mm which corresponds to $\lambda_0/5$ at 10 GHz. The dielectric substrates used are Rogers RO3003 ($\epsilon_r = 3,~\tan{\delta}=0.001$) with a thickness of 1.52 mm for each spacer leading to a total metasurface thickness of 3.04 mm ($\approx\lambda_0/10$).
![Generic metallic layer used to realize the metasurface unit cells.[]{data-label="Fig:metallicUnitCell"}](Jcross){width="0.8\columnwidth"}
Applications
============
In this section, several types of birefringent metasurfaces are discussed and demonstrated experimentally. The realized metasurface are, in order of appearance, a half-wave plate, a quarter-wave plate, a polarization beam splitter and an orbital angular momentum generator. These four operations are illustrated in Fig. \[Fig:Blend\].
![Representations of the operation of the four realized metasurfaces.[]{data-label="Fig:Blend"}](blend){width="1\columnwidth"}
Electromagnetic Wave Plates
---------------------------
Electromagnetic wave plates are birefringent structures that exhibit specific transmission phase difference between $x$- and $y$-polarizations defined as $\Delta\phi= |\phi_x - \phi_y|$. Here, we present the two most common wave plates: a half-wave plate and a quarter-wave plate which, respectively, correspond to $\Delta\phi = \pi$ and $\Delta\phi = \pi/2$. The half-wave plate performs a $90^\circ$ rotation of polarization for linear polarization or changes the handedness of circular polarization. The quarter-wave plate transforms linear polarization into circular polarization and vice-versa.
For such electromagnetic transformations, the metasurface is uniform since there is no variation in the direction of propagation of the waves and, therefore, no phase gradient is required. This make these metasurfaces very easy to design since only one unit cell has to be implemented and repeated periodically to form the metasurface.\
### Half-Wave Plate
The fabricated metasurface is shown in Fig. \[Fig:HWP\_fab\] while the dimensions of its unique unit cell are given in Table \[Table1\]. The metasurface is made of $24\times24$ unit cells corresponding to an aperture of about 5$\lambda_0\times5\lambda_0$. Note that the two holes on both sides of the metasurface are used to attach it to the measurement setup.
![Fabricated half-wave plate metasurface.[]{data-label="Fig:HWP_fab"}](HWP){width="0.6\columnwidth"}
Lx Ly Wx Wy Ax Ay Bx By
---- ---- ------ ----- ------ ------ ------ ------ ------ --
OL 4 4.75 0.5 0.5 0.5 0.5 3.25 2.25
ML 5 5 0.5 0.25 0.25 0.75 2.25 3.5
: Geometrical dimensions (in mm) for the unit cell of the metasurface in Fig. \[Fig:HWP\_fab\]. OL denotes the outer layers and ML the middle layer.[]{data-label="Table1"}
The metasurface has then been measured. Two horn antennas, placed on both sides of the metasurface, have been used to measure the normal transmission from a normally incident wave. The measured transmissions for $x$ and $y$ polarizations are, respectively, plotted in Figs. \[Fig:HWP\_T1\] and \[Fig:HWP\_T2\], where the red solid lines correspond to measurements with the metasurface and the blue dashed lines are the reference line of sight measurements of the horn antennas. The phase difference, $\Delta \phi$, is plotted in Fig \[Fig:HWP\_T3\].
\
\
As can be seen, the metasurface transmission is almost unity around the specified frequency of operation of 10 GHz. The ideal phase shift difference of $\Delta\phi = \pi$ is obtained at $f=10.2$ GHz. In order to verify the $90^\circ$ rotation of polarization capability of this structure, cross-polarization measurements with and without the metasurface have been performed and the result is shown in Fig. \[Fig:HWP\_P\] only for the metasurface transmission case for convenience.
\[Fig:susc1\] ![Measured and normalized cross-polarized transmission and bandwidth of the half-wave plate.[]{data-label="Fig:HWP_P"}](HWP_xy "fig:"){width="0.8\linewidth"}
The result in Fig. \[Fig:HWP\_P\] confirms that the metasurface behaves almost as a perfect half-wave plate with a power transmission efficiency of $95\%$ at 10.2 GHz and a -3-dB bandwidth of about $10\%$.\
### Quarter-Wave Plate
The quarter-wave plate metasurface was designed and realized following exactly the same procedure as that of the half-wave plate metasurface. The fabricated metasurface is shown in Fig. \[Fig:QWP\_fab\] and the dimensions of its unit cell are given in Table \[Table2\].
![Fabricated quarter-wave plate metasurface.[]{data-label="Fig:QWP_fab"}](QWP){width="0.6\columnwidth"}
Lx Ly Wx Wy Ax Ay Bx By
---- ------ ---- ----- ------ ------ ----- ------ ------ --
OL 5 5 0.5 0.75 0.5 0.5 3 2.75
ML 2.75 5 0.5 0.75 0.25 0.5 2.75 1.75
: Geometrical dimensions (in mm) for the unit cell of the metasurface in Fig. \[Fig:QWP\_fab\]. OL denotes the outer layers and ML the middle layer.[]{data-label="Table2"}
As was done for the half-wave plate metasurface, the measurements of the quarter-wave plate metasurface, corresponding to $x$ polarization transmission, $y$ polarization transmission and phase difference, are plotted in Figs. \[Fig:QWP\_T1\], \[Fig:QWP\_T2\] and \[Fig:QWP\_T3\], respectively.
\
\
The metasurface exhibits very good transmission (near unity) for both $x$ and $y$ polarization around the frequency of operation. The phase difference reaches the required value of $\Delta\phi=\pi/2$ at the specified frequency of 10 GHz. Finally, the linear-to-circular conversion efficiency has been estimated from the $x$ and $y$ polarization amplitude and phase measurements and has been plotted in Fig. \[Fig:QWP\_P\].
\[Fig:susc1\] ![Measured linear-to-circular power conversion and bandwidth of the quarter-wave plate.[]{data-label="Fig:QWP_P"}](QWP_xy "fig:"){width="0.8\linewidth"}
As can be seen, the linear-to-circular conversion efficiency is very good reaching about $94\%$ at 10 GHz with a -3-dB bandwidth of about $12\%$.\
Generalized Birefringent Reflection and Refraction
--------------------------------------------------
The concept of generalized birefringent reflection and refraction consists in controlling independently the reflection and transmission angles and amplitudes of orthogonally polarized waves. To illustrate this concept, we have realized a polarization beam splitting (PBS) reflection-less metasurface that refracts in opposite directions normally incident $x$ and $y$ polarized waves. The synthesis of this metasurface essentially follows the procedure elaborated in the introduction of this paper and which corresponds to the second synthesis technique that was discussed. Accordingly, the transmission coefficients for $x$ and $y$ polarization are, respectively, given by $T_x(x,y)=e^{-jk_0\sin{\theta_\text{t}}x}$ and $T_y(x,y)=e^{jk_0\sin{\theta_\text{t}}x}$, where $\theta_\text{t}$ is the specified transmission angle. Note that separation of both polarizations was initially specified to be along the $x$ direction. The phase gradient, corresponding to $T_x$ and $T_y$, has a period that is given by $P=\lambda_0/\sin{\theta_\text{t}}$. For the realization of that metasurface, we decided to sample the period $P$ with 8 unit cells of lateral size $\lambda_0/5$. Consequently, the transmission angle is determined by the unit cell size and the number of unit cells and is, thus, given by $\theta_\text{t}=\arcsin{(5/8)}\approx 38.7^\circ$.
Each unit cell has then been implemented to realize a specific phase shift for $x$ and $y$ polarizations, respectively, $\phi_x$ and $\phi_y$. The transmission phases, for each unit cell, are given in Table \[Table3\]. Note that the absolute phase shift of a single unit cell is irrelevant, only the phase shift difference between adjacent unit cells (here $45^\circ$) matters.
1 2 3 4 5 6 7 8
---------- ----- ----- ----- ----- ----- ----- ----- ----- --
$\phi_x$ 0 45 90 135 180 225 270 315
$\phi_y$ 315 270 225 180 135 90 45 0
: Transmission phase shifts (in degrees) for $x$ and $y$ polarization of the 8 unit cells forming the PBS metasurface.[]{data-label="Table3"}
As can be seen in Table \[Table3\], the supercell (formed by the 8 unit cells) has an asymmetric phase progression meaning that the unit cells number 5, 6, 7 and 8 have opposite $x$ and $y$ phase shifts as the unit cells number 4, 3, 2 and 1, respectively. This means that the 4 lasts unit cells are simply the $90^\circ$ rotated version of the first 4 unit cells. Consequently, the realization of this metasurface is greatly simplified since only 4 unit cells need to be implemented.
After designing the supercell, we performed full-wave simulations to verify the beam splitting behavior of the metasurface. The $x$ polarization refraction yielded good result but, unfortunately, the $y$ polarization transformation was not good, which can be explained by the presence of spurious coupling between adjacent unit cells. While the coupling affected both polarizations, it turns out that it was more damaging to the $y$ polarization than the $x$ polarization. It has to be noted that the metasurface is non-uniform only in the $x$ direction while being perfectly uniform in the $y$ direction, this asymmetry in the structure was hypothesized to be the cause of the different behavior of the two polarizations. To overcome this difficulty, we modified the metasurface such that the same non-uniformity was present in both $x$ and $y$ directions. Consequently, the supercell is now made of $8\times8$ unit cells instead of $8\times1$. The realized metasurface is shown in Fig. \[Fig:PBS\_fab\], note the sinusoidally varying pattern in the diagonal direction indicating the direction of the phase gradients. The metasurface is made was a repetition of 9 supercells.
![Fabricated polarization beam splitting metasurface. The supercell made of $8\times8$ unit cells is highlighted by the black square.[]{data-label="Fig:PBS_fab"}](PBS){width="0.6\columnwidth"}
The dimensions of each of the first 4 unit cells are given in Table \[Table4\]. As said above, the 4 remaining unit cell are just the rotated version of the first 4 ones.
Lx Ly Wx Wy Ax Ay Bx By
-- ---- ------ ------ ------- ------- ----- ----- ------- ------- --
OL 5.5 4 0.5 0.625 0.5 0.5 4.875 2.25
ML 5.5 5.5 0.375 1.125 0.5 0.5 1.375 2.875
OL 5.25 4.25 0.625 0.5 0.5 0.5 4.125 3
ML 5.75 3.5 0.5 0.375 0.5 0.5 0.5 1.375
OL 4.25 4.75 0.625 0.25 0.5 0.5 4.25 3
ML 3.75 5.25 0.5 0.375 0.5 0.5 2.25 4.5
OL 3.75 3.5 0.125 0.125 0.5 0.5 2.25 4.375
ML 5.5 5.25 0.375 1.125 0.5 0.5 4.75 0.125
: Geometrical dimensions (in mm) for the first four unit cells of the metasurface in Fig. \[Fig:PBS\_fab\]. OL denotes the outer layers and ML the middle layers.[]{data-label="Table4"}
Because the metasurface has now a period in the diagonal direction, the dimension of the phase gradient period is reduced to $P=8\lambda_0/(5\sqrt{2})$. This changes the transmission angle to $\theta_\text{t}\approx 62.1^\circ$ at 10 GHz.
The metasurface has then been measured. A horn antenna was used to generate the normally incident waves while a probe was scanning the near-field over a plane parallel to the metasurface in the transmission region. Near-field to far-field transformation [@balanis2016antenna] was then used to evaluate the transmission response of the metasurface. The measured $x$ and $y$ polarization transmissions, in the diagonal plane of the metasurface, are plotted in Fig. \[Fig:PBS\_FF\] as a dashed blue line and a solid red line, respectively. Note that the curves have been normalized with respect to the $y$-polarized transmission.
![Measured normalized transmitted power for $x$-polarization (dashed blue curve) and $y$-polarization.[]{data-label="Fig:PBS_FF"}](PBS_PolardB){width="0.8\linewidth"}
As can be seen, the metasurface effectively separates the two polarizations which are refracted, with almost identical amplitude, at about $+60^\circ$ and $-60^\circ$ from broadside, respectively. The frequency corresponding to the results in Fig. \[Fig:PBS\_FF\] is about 10.4 GHz and the transmission efficiency, defined has the ratio between the transmitted electric field and the incident electric field, is about $80\%$. The efficiency of the metasurface versus frequency is plotted in Fig. \[Fig:PBS\_Eff\]. The reasons for which the metasurface efficiency does not exceed $80\%$ can be explained partly by the presence of loss in the dielectric layers but mostly from undesired refraction orders (either in reflection or in transmission) that are due to the spurious coupling of the unit cells. For instance, zeroth diffraction orders are clearly visible in the measurements shown in Fig. \[Fig:PBS\_FF\].
![PBS metasurface transmission efficiency.[]{data-label="Fig:PBS_Eff"}](Teff_PBS){width="0.8\linewidth"}
Orbital Angular Momentum Multiplexing
-------------------------------------
The last metasurface that was realized is a structure that generates waves possessing orbital angular momentum (OAM) of different topological charges depending on the polarization of the incident wave. The OAM wave that we have chosen as the transmitted wave is a Hypergeometric Gaussian (HyG) wave that corresponds to the solution of the paraxial Maxwell equations in cylindrical coordinates. The reason why the HyG wave has been chosen over the more common Bessel wave [@recami2009localized] is that the HyG has the advantage of being linearly polarized whereas the Bessel wave is either radially of longitudinally polarized making the multiplexing of two OAM waves with a single metasurface more difficult for Bessel waves.
The metasurface is thus required to transform an $x$-polarized normally incident wave into an HyG wave of topological charge $m=+1$ and $y$-polarized normally incident wave into an HyG wave of topological charge $m=-1$. The electric field of an HyG wave reads [@karimiHYG] $$\begin{split}
\label{eq:HYG}
E(\rho,\phi,z) =& \frac{\Gamma \left ( 1+|m|+\frac{p}{2} \right )}{\Gamma (|m|+1)}\frac{i^{|m|+1}\zeta^{|m|/2}\xi^{p/2}}{[\xi+i]^{1+|m|/2+p/2}}e^{im\phi-i\zeta}\\
&\times~_{1}F_1 \left ( -\frac{p}{2},|m|+1;\frac{\zeta[\xi + i]}{\xi[\xi - i]} \right ),
\end{split}$$ where $_{1}F_1(a,b;x)$ is the confluent hypergeometric function, $\Gamma(x)$ is the gamma function, $m$ is the OAM topological charge, $p\geqslant -|m|$ is a real parameter, and where $\zeta = \rho^2/(w_0^2[\xi+i])$, $\xi=z/z_R$ $w_0$, with $w_0$ being the beam waist and $z_r$ the Rayleigh range given by $z_r=\pi w_0^2/\lambda$.
The amplitude and phase of the HyG wave, for $m=-1$, are plotted in Figs. \[Fig:Hyg1\] and \[Fig:Hyg2\], respectively. It is clear that this kind of wave has a non-periodic phase pattern, as evidenced in Fig. \[Fig:Hyg2\], contrary to the oblique transmitted plane waves that were specified for the polarization beam splitting metasurface in the previous section. This means that a larger number of unit cells has to be implemented because of the aperiodicity of the transformation. Moreover, the amplitude of the HyG wave is non-uniform which further complicates the realization of the unit cells. However, these difficulties may by overcome by assuming that the amplitude of the $x$ and $y$ transmission coefficients are $|T_x(x,y)| = |T_y(x,y)|=1$ instead of following the profile in Fig. \[Fig:Hyg1\]. Despite the fact that this approximation might a priori seem extreme, it turns out that the main properties of the HyG wave may be obtained by only implementing its phase evolution. For instance, the null amplitude at the center of the wave is achieved by destructive interferences due to the phase rotation around the center. Moreover, the orbital angular momentum information is contained not in the amplitude but rather in the phase of the wave. These considerations justify the assumption that only the phase profile of the transmitted waves should be implemented while their respective magnitude can be assumed to be uniform and equal to 1. Additionally, the phase profiles of the two OAM waves were discretized by four phase samples each. Consequently, the metasurface is made of $24\times 24$ unit cells with phase shifts for $x$ and $y$ polarizations as given in Figs. \[Fig:Hyg3\] and \[Fig:Hyg4\], respectively.
\
When combining together the phase shifts in Figs. \[Fig:Hyg3\] and \[Fig:Hyg4\], it turns out that the total number of different unit cells composing the metasurface is 16. The Fig. \[Fig:Hyg\_comb\] represents the $24\times 24$ unit cells of the metasurface. In that figure, each color corresponds to a specific unit cell having unique phase shift for $x$ and $y$ polarizations. Interestingly, the unit cells in the highlighted regions 1, 2 and 3 are quarter-wave plates, half-wave plates and isotropic wave plates (where $\phi_x=\phi_y=\phi$), respectively. Consequently, 6 out of the 10 correspond to the same unit cells but rotated in the plane of the metasurface with respect to each other.
![Representation of the $24\times 24$ unit cells of the OAM multiplexing metasurface. Each color represents a unit cell with specific phase shifts for $x$ and $y$ polarizations. There is a total number of 16 different unit cells where, notably, the ones in regions 1 are quarter-wave plates, the ones in regions 2 are half-wave plates and the ones in regions 3 are isotropic.[]{data-label="Fig:Hyg_comb"}](Hyg_comb){width="0.7\linewidth"}
The dimensions for the 10 remaining unit cells are given in Table \[Table5\] and the fabricated metasurface is shown in Fig. \[Fig:HYG\_fab\].
Lx Ly Wx Wy Ax Ay Bx By
-- ---- ------ ------ ------ ------ ------ ------ ------ ------ --
OL 5 5 0.5 0.5 0.5 0.5 3.5 3.5
ML 4.75 4.75 0.75 0.75 0.5 0.5 3.5 3.5
OL 5 4.75 0.5 0.5 0.5 0.5 3.25 3.25
ML 4.75 4.75 0.25 1 1 0.5 2.5 3.75
OL 4.75 5 0.75 0.5 0.25 0.25 2.75 4.25
ML 4.5 2.75 0.25 0.5 0.5 0.75 1.75 3.25
OL 4.75 4 0.5 0.5 0.5 0.5 2.25 3.75
ML 4.75 5 0.25 0.5 0.5 0.5 3 2.75
OL 4.75 4.75 0.5 0.5 0.5 0.5 3.25 3.25
ML 4.75 4.75 0.5 0.5 0.75 0.75 2.75 2.75
OL 4.75 5 0.5 0.75 0.5 0.5 3 3
ML 2.75 4.5 0.5 0.75 0.25 0.75 3.25 1.75
OL 4 4.75 0.5 0.5 0.5 0.5 3.25 2.25
ML 5 4.75 0.5 0.25 0.5 0.75 2.5 3
OL 4.75 4.75 0.75 0.75 0.5 0.5 3.25 3.25
ML 2.75 2.75 0.75 0.75 0.25 0.25 1.75 1.75
OL 4 4.75 0.5 0.5 0.5 0.5 2.75 1.5
ML 5 2.75 0.75 0.25 0.5 0.25 1.75 4.25
OL 4 4 0.5 0.5 0.5 0.5 2.25 2.25
ML 5 5 0.25 0.25 0.5 0.5 3 3
: Geometrical dimensions (in mm) for 10 of the unit cells of the metasurface in Fig. \[Fig:HYG\_fab\]. OL denotes the outer layers and ML the middle layers. The numbers in the first column correspond to the phase shifts $(\phi_x,\phi_y)$ for $x$ and $y$ polarizations, respectively.[]{data-label="Table5"}
![Fabricated orbital angular momentum multiplexing metasurface. It is made of 16 uniquely different unit cells arranged according to the pattern in Fig. \[Fig:Hyg\_comb\], the metasurface contains a total of $24\times24$ unit cells.[]{data-label="Fig:HYG_fab"}](HYG){width="0.6\columnwidth"}
The metasurface has then been simulated and measured and the results are reported in Figs. \[Fig:Hyg\_aclk\] and \[Fig:Hyg\_clk\] which respectively correspond to $x$ and $y$ polarizations. In these two figures, plots (a) and (b) are the simulated amplitude and phase transmissions. These simulations were obtained by first measuring the radiated reference field of the exciting horn antenna at the position of the metasurface. Then, the expected scattered field of the metasurface was calculated using the antenna reference field and assuming ideal transmission of flat and unity amplitude and phase profiles as in Figs. \[Fig:Hyg3\] and \[Fig:Hyg4\]. This simulation technique allows fair comparison between the expected scattered fields and the measured scattered fields from the metasurface that are shown in plots (c) and (d).
\
\
The measured results are in good agreements with the expected simulated results. The topological charges of $m=+1$ and $m=-1$ are achieved with a transmission efficiency near $90\%$ at 10 GHz. Finally, the transmission efficiency of the metasurface was evaluated for a frequency band between 8 and 12 GHz. The result is reported in Fig. \[Fig:Hyg\_eff\].
![Orbital angular momentum multiplexing metasurface transmission efficiency.[]{data-label="Fig:Hyg_eff"}](Hyg_eff){width="0.8\linewidth"}
Conclusion
==========
In the first part of this work, we have discussed and elaborated two different approaches for the synthesis of birefringent metasurfaces. The two methods yield the metasurface electric and magnetic susceptibilities either when the exact electromagnetic fields on both sides on the metasurface are specified or when the transmission and reflection coefficients are specified. The first synthesis technique is rigorous while the second one is an approximation. However, we have seen that both methods lead to similar results in terms of the metasurface scattering response. One of the main differences between the two techniques lies in the difficulty of the physical realization of the metasurfaces. While the first technique requires the implementation of complicated non-uniform amplitude and phase transmission and reflection coefficients, the second usually only requires the implementation of non-uniform phase profiles. Consequently, although less rigorous, the second synthesis method is generally preferred over the first one.
In the second part, we have presented the synthesis and realization of four different birefringent metasurfaces performing the operation of half-wave plate, quarter-wave plate, polarization beam splitting and orbital angular momentum multiplexing, respectively. These metasurfaces were synthesized based on the second synthesis technique and their measurements were found in good agreement with the expected scattering responses.
[^1]: K. Achouri, G. Lavigne and C. Caloz are with the Department of Electrical Engineering, $\acute{\mathrm{E}}$cole polytechnique de Montr$\acute{\mathrm{e}}$al, Montr$\acute{\mathrm{e}}$al, QC, H3T 1J4 Canada (e-mail: [email protected]).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The electricity consumption forecasting is a critical component of the intelligent power system. And accurate monthly electricity consumption forecasting, as one of the the medium and long term electricity consumption forecasting problems, plays an important role in dispatching and management for electric power systems. Although there are many studies for this problem, large sample data set is generally required to obtain higher prediction accuracy, and the prediction performance become worse when only a little data is available. However, in practical, mostly we experience the problem of insufficient sample data and how to accurately forecast the monthly electricity consumption with limited sample data is a challenge task. The Holt-Winters exponential smoothing method often used to forecast periodic series due to low demand for training data and high accuracy for forecasting. In this paper, based on Holt-Winters exponential smoothing method, we propose a hybrid forecasting model named FOA-MHW. The main idea is that, we use fruit fly optimization algorithm to select smoothing parameters for Holt-Winters exponential smoothing method. Besides, electricity consumption data of a city in China is used to comprehensively evaluate the forecasting performance of the proposed model. The results indicate that our model can significantly improve the accuracy of monthly electricity consumption forecasting even in the case that only a small number of training data is available.'
address:
- 'Chongqing University, Chongqing 400044, China.'
- 'Southern University of Science and Technology, Shenzhen 518055, China'
author:
- Weiheng Jiang
- Xiaogang Wu
- Yi Gong
- Wanxin Yu
- Xinhui Zhong
bibliography:
- 'bibfile.bib'
title: 'Monthly electricity consumption forecasting by the fruit fly optimization algorithm enhanced Holt-Winters smoothing method'
---
Monthly electricity consumption forecasting ,Holt-Winters exponential smoothing method ,Fruit fly optimization algorithm
Introduction {#Introduction}
============
Electricity consumption forecasting is a critical component for the modern management of the electric power systems, which has attracted more and more attentions from both the academic and the industry society [@DEBNATH2018297]. And highly-accurate mid-term and long-term electricity consumption forecasting plays an important role in dispatching and planning of electric power systems. However, due to the fact that the mid-term or long-term electricity consumption has complex and non-linear relationships with lots of external factors, such as the political environment, economic policy, human activities, irregular behaviors and other non-linear factors [@Li2013A], which then makes it is difficult to forecast accurately. In order to improve the accuracy of mid-term and long-term electricity consumption forecasting, various prediction methods and models are proposed these years.
From the perspective of forecasting model, the mid-term or long-term electricity consumption forecasting methods can be classified into two types, i.e., the stand-alone based model and the hybrid model [@DEBNATH2018297] [@Shao2017A]. In addition, based on the used techniques, the stand-alone model is further divided into three categories, i.e., statistical model [@Tratar2016The] [@Aydin2014Modeling], computation intelligence (CI) model [@KIRAN201293] [@Zhang2018A], and the mathematical programming (MP) based model [@forouzanfar2010modeling]. While for the hybrid methods, they are further divided into four categories, i.e., the statistical-statistical [@tan2010day], statistical-CI [@da2019bottom], CI-CI [@Ju2013Application] [@Chen2018Forecasting], and statistical-MP methods [@forouzanfar2010modeling]. The stand-alone methods can also be divided into linear and nonlinear categories [@Hernandez2014A].
In particular, for the stand-alone model, lots of work have been done over these different categories of models. Firstly, for the statistical model, the main methods include exponential smoothing method [@Tratar2016The] [@RENDONSANCHEZ2019916], regression analysis method [@Aydin2014Modeling] [@Tsekouras2007A], and the time series method [@DEOLIVEIRA2018776] [@Nawaz2014Modelling]. In detail, Tsekouras et. al [@Tsekouras2007A] presented a non-linear multi-variable regression method for midterm energy forecasting of the power systems in annually [@HE2019565]. The author of [@Nawaz2014Modelling] estimated Pakistan’s electricity demand by applying STAR (Smooth Transition Auto-Regressive) model. Secondly, the computational intelligence methods include metaheuristic method [@KIRAN201293] [@Amjadi2010Estimation], machine learning method [@Zhang2018A] [@Taylor2002Neural], knowledge-based method [@Kandil2002Long] [@TANG20191144] and the uncertainty method [@Wang2012Optimization] [@Ali2016Long]. More specifically, the regressive convolution neural network (RCNN) is used to extract features from data and then the regressive support vector machine (SVR) trained with features is adopted to predict the electricity consumption, which achieves a low predicting error in [@Zhang2018A]. In addition, instead of point forecasting, based on fuzzy Bayesian theory and expert prediction, a novel long-term probability forecasting model is proposed to predict the Chinese per-capita electricity consumption (PEC) and its variation interval over the period 2010-2030 [@TANG20191144].
Comparing with the stand-alone model, the hybrid model is more popularly studied due to its excellent representation ability of nonlinear factors and random factors. Concretely, Ju et. al employed chaotic gravitational search algorithm to determine the three free parameters in the support vector regression (SVR) model, and which can significantly improve the performance of the SVR [@Ju2013Application]. In [@Chen2018Forecasting], a novel approach for monthly electricity demands forecasting by the integration of both the wavelet transform and the neuro-fuzzy system is proposed, and the results confirm that this model can provide an accurate forecasting. He et. al proposed a method of probability density forecasting based on least absolute shrinkage and selection operator-quantile regression neural network (LASSO-QRNN), and the prediction accuracy of the proposed model is evaluated through the empirical analyses with the Guangdong province dataset in China and the California dataset in the United States [@HE2019565]. In [@da2019bottom], a methodology that combines the bottom-up approach with hierarchical linear models for long term electricity consumption forecasting of a particular industrial sector considering energy efficiency scenarios is proposed, and the model was used to generate long term point and probability distribution forecasts for the period ranging from 2015 until 2050.
Till now, though lots of work have been done for the mid-term electricity consumption forecasting and different models have been presented. It is noted that, all these proposed models can only handle the case with large number of sample data, and have low forecasting accuracy if we only have a small number of training sample data, especially for the computational intelligence method and the machine learning based method. However, in practical, as we know that, in China, the development of the informatization for the electrical companies in some areas only beginning recently, thus the aggregate historical electricity consumption data is limited. Thus, how to forecast the electricity consumption accurately with insufficient sample data is a challenge problem.
In this paper, a new model based on the integration of the fruit fly optimization algorithm (FOA) and the Holt-Winters exponential smoothing (HW) is proposed for the mid-term electricity consumption forecasting, to handle the situation that we have only a small number of aggregated historical electricity consumption data. On the one hand, HW is known as a common approach and used to forecast the seasonal time series [@Holt2004Forecasting] [@Winters1960Forecasting]. However, the performance of HW is mainly depended on the selection of appropriate smoothing parameters. On the other hand, we know that the FOA is an excellent tool in pursuing the global optimization for the parameter optimization problems [@Pan2012A] [@Li2013A]. Therefore, herein, we introduce the FOA to assist the smoothing parameter selection for the HW and which formulates the proposed FOA-MHW model. To the best knowledge of the authors, this is the first enhanced HW model, and the real data set based test shows that it can obtain incredible performance even in the situation of insufficient sample data.
The rest of this paper is organized as follows: Section 2 introduces the Holt-Winters exponential smoothing method and the FOA, and then a hybrid forecasting model constructed by the Holt-Winters exponential smoothing method and FOA is proposed and discussed. In Section 3, the sample data set used in this paper is illustrated, and then based on this sample data set, the performance of the proposed model are evaluated and compared under different length of training data sets. At last, we conclude this paper in Section 4.
Methodology of FOA-MHW model
============================
In this section, at first, we introduce the Holt-Winters (HW) exponential smoothing model, and then the theory of fruit fly optimization algorithm is presented. At last, the process of the proposed FOA-MHW model is described in details. \[Model\]
Holt-Winters exponential smoothing method {#ModelHW}
-----------------------------------------
Exponential smoothing is an important time series forecasting method, its basic idea is to preprocess the original data to eliminate the randomness in the time series, and improve the importance of the recent data in predicting the collected data. The processed data is called “smoothing value”, and then the forecasting model is constructed according to the smoothing value, and the future target value is forecasted by the model [@ForeastBook]. However, the simple exponential smoothing method can’t overcome the randomness of time series enough, to this end, the Holt-Winters exponential smoothing method is then developed by Holt and Winters [@Holt2004Forecasting] [@Winters1960Forecasting]. Previous analysis shows that, even with a small number of the sample data, the Holt-Winters exponential smoothing method still can achieve a good forecasting result [@ForeastBook]. While the Holt-Winters exponential smoothing method can be classified into four types: 1) Multiplicative model (MHW), 2) Additive model (AHW), 3) Modified model (MoHW), 4) Extended model (EHW). The differences among these models come from the methods used in calculating the seasonal indices [@Tratar2016The]. In which, the MHW is the most popular Holt-Winters exponential smoothing method dues to the fact that, the calculation method of seasonal indices is more suitable for the actual situation. Hence, we choose MHW as our primary forecasting model.
The multiplicative Holt-Winters (MHW) model forecasts trend, stationarity and seasonal components of the time series at the same time. The basic form of the Holt-Winters exponential smoothing forecasting model is as follows: $$\hat{y}_{t+m} = (T_t + b_tm)S_{t+m-L}
\label{eq:1}$$ In which, $\hat{y}_{t+m}$ means the $m$th forecasting data, $t$ is the length of training data and herein, it denotes the length of the period. Furthermore, the forecasting model includes three components of time series: stationarity $T_t$, trend $b_t$ and seasonality $S_t$. $T_t$ is used to correct time series to rule out seasonal factors in time series, $b_t$ is used to correct trend values to eliminate seasonal interference in the time series, and $S_t$ is used to forecast seasonal indices to exclude random interference [@ForeastBook]. The recursive process of the three variable are as follows. $$T_t = \alpha \frac{y_t}{S_{t-L}} + (1- \alpha)(T_{t-1}+b_{t-1})
\label{eq:2}$$
$$b_t = \beta (T_t - T_{t-1}) + (1 - \beta)b_{t-1}
\label{eq:3}$$
$$S_t = \gamma \frac{y_t}{T_t} + (1 - \gamma)S_{t-L}
\label{eq:4}$$
Herein, $\alpha$, $\beta$ and $\gamma$ are three smoothing parameters of the Holt-Winters model, and all of them take value from $[0,1]$. In addition, $y_t$ denotes the observation data. Based on the analysis of the first order exponential smoothing method [@ForeastBook], it is found that the larger value of $\alpha$, $\beta$ and $\gamma$, the smaller influence of the long-term historical electricity quantity on the prediction, and vice versa. Hence, these smoothing parameters play a critical role in the final predicting results and how to determine the value of these parameters is the focus of this paper.
The initialization method of the equation (2) - (4) is as follows. $$T_0 = \frac{1}{n} \sum_{t=1}^n y_t
\label{eq:5}$$
$$S_{0k} = \frac{\bar{y_k}}{T_0}
\label{eq:6}$$
$$b_0 = 0
\label{eq:6}$$
where, $$\bar{y_k} = \frac{1}{n/L}\sum_{t=1 \wedge t = jL+k}^n y_t, (k = 1, ..., L, j = 1,2, ..., n/L-1)
\label{eq:6}$$ Herein, $T_0$ is the average value of the training data, $S_{0k}$ ($k = 1, ..., L$) is the initial value of the seasonal index, and the initial value of $b_t$ is set to 0.
Fruit fly optimization algorithm
--------------------------------
Fruit fly optimization algorithm (FOA) is an intelligent optimization algorithm developed by Pan in [@Pan2012A]. And it is often used to search the global optimal solution for parameter optimization problem based on the food finding behavior of the fruit fly. The fruit fly itself is superior to other species in sensing and perception, especially in osphresis and vision. The flow chart of the evolution of fruit fly population is shown in Figure \[FOA\], and the details of the algorithm are summarized as follows [@Cao2016Support].
![$~$The evolution process of fruit fly population [@Li2013A].[]{data-label="FOA"}](FOA.eps)
- Parameter setting: The population size ($sizepop$), the maximum iteration number ($maxgen$), and the random flight range ($FR$), the initial fruit fly swarm location ($x_0$,$y_0$).
- Initialization: Give the random direction and distance for the search of food using osphresis by an individual fruit fly. $$X_i = x_0 + rand(FR),
\label{eq:1}$$ $$Y_i = y_0 + rand(FR),
\label{eq:2}$$
- Calculate the smell concentration judgment value: Calculate the distance between the fruit fly individual to the origin ($Dist_i$), then calculate the smell concentration judgment value ($S_i$). $$Dist_i = \sqrt{X_i^2+Y_i^2},
\label{eq:3}$$ $$S_i = 1/Dist_i;
\label{eq:4}$$
- Calculate the smell concentration: Feed the $S_i$ into the fitness function, thus obtain the smell concentration ($Smell_i$) of the individual fruit fly location, i.e., the equation (5). $$Smell_i = Fuction(S_i),
\label{eq:5}$$
- Find the best individual: Find out the fruit fly with maximal smell concentration (finding the maximal value) among the fruit fly swarm. $$[bestSmell, bestIndex] = max(Smell_i),
\label{eq:6}$$ Where the $bestSmell$ represent the highest smell concentration among the current fruit fly swarm, and the $bestIndex$ represents the fruit fly which has the highest smell concentration among the current fruit fly swarm.
- Fruit fly swarm movement: Select the best fruit fly individual $bestIndex$, keep the best direction, and at this moment, the fruit fly swarm will use vision to fly towards that best location. $$smellBest = bestSmell
\label{eq:7}$$ $$x_0 = X_{bestIndex}
\label{eq:8}$$ $$y_0 = Y_{bestIndex}
\label{eq:9}$$
- Population evolution: Perform the iterative optimization and repeat the implementations from step 2 to step 6, when the smell concentration is not better than the previous iteration results any more or the number of iteration reaches the $maxgen$, the algorithm is ended.
Fruit fly optimization algorithm for parameter selection of the Holt-Winters exponential smoothing method {#FOA-HW}
---------------------------------------------------------------------------------------------------------
As we have mentioned in section \[ModelHW\], the forecasting performance of the Holt-Winters exponential smoothing method is mainly determined by the value of the three smoothing parameters, i.e., $\alpha$, $\beta$, and $\gamma$. Therefore, in this paper, the FOA is used to search the value of these smoothing parameters for the Holt-Winters exponential smoothing model and which will significantly improve the accuracy of the monthly electricity consumption forecasting.
![$~$The flowchart of FOA selecting smoothing parameters for the Holt-Winters exponential smoothing model.[]{data-label="FOAHW"}](FOAHW.eps)
The flowchart of the FOA based smoothing parameter selecting for the Holt-Winters exponential smoothing model (abbreviated as FOA-MHW) is presented in Figure \[FOAHW\]. And the details of the algorithm are summarized as follows.
- Data sets splitting: Divide the sample data set into training data set and test data set. In addition, the last period of training data set is defined as validation data set, and which is used to conduct the optimal smoothing parameter selection in the FOA for Holt-Winters exponential smoothing model.
- Parameter initialization: Initialize the population size ($sizepop$), the maximum iteration number ($maxgen$), the random flight range ($FX^\alpha$, $FY^\alpha$), ($FX^\beta$, $FY^\beta$), ($FX^\gamma$, $FY^\gamma$) and the initial fruit fly swarm location ($x_0^\alpha$, $y_0^\alpha$), ($x_0^\beta$, $y_0^\beta$) and ($x_0^\gamma$,$y_0^\gamma$).
Where $(x_0^\alpha$, $y_0^\alpha)$ is the initial fruit fly swarm location of parameter $\alpha$, $(x_0^\beta$, $y_0^\beta)$ denotes the initial fruit fly swarm location of parameter $\beta$ and $(x_0^\gamma$, $y_0^\gamma)$ characterizes the initial fruit fly swarm location of parameter $\gamma$. Hence, there are three fruit flies searching their optimal smoothing parameters for the Holt-Winters exponential smoothing model.
- Fruit fly swarm flight: Each of the three fruit fly randomly searches of the food at a predetermined distance, and each of fruit fly’s location of these three fruit fly swarms are $(X_i^\alpha, Y_i^\alpha)$, $(X_i^\beta, Y_i^\beta)$ and $(X_i^\gamma, Y_i^\gamma)$, respectively.
Where, $$X_i^\alpha = x_0^\alpha + FX^\alpha, Y_i^\alpha = y_0^\alpha + FY^\alpha
\label{eq:9}$$ $$X_i^\beta = x_0^\beta + FX^\beta, Y_i^\beta = y_0^\beta + FY^\beta
\label{eq:9}$$ $$X_i^\gamma = x_0^\gamma + FX^\gamma, Y_i^\gamma = y_0^\gamma + FY^\gamma
\label{eq:9}$$
- Calculate the smell concentration judgment value: the smell concentration judgment value ($S_i$) of fruit fly is calculated as follows. $$S_i^\alpha = \alpha = \frac{1}{D_i^\alpha}
\label{eq:9}$$ $$S_i^\beta = \beta = \frac{1}{D_i^\beta}
\label{eq:9}$$ $$S_i^\gamma = \gamma = \frac{1}{D_i^\gamma}
\label{eq:9}$$
Where, $$D_i^\alpha = 1/\sqrt{X_i^{\alpha2} + Y_i^{\alpha2}}
\label{eq:9}$$ $$D_i^\beta = 1/\sqrt{X_i^{\beta2} + Y_i^{\beta2}}
\label{eq:9}$$ $$D_i^\gamma = 1/\sqrt{X_i^{\gamma2} + Y_i^{\gamma2}}
\label{eq:9}$$ Herein, $D_i^\alpha$, $D_i^\beta$ and $D_i^\gamma$ denote the distance between the fruit fly individual and the origin. We feed the smell concentration judgment value (use as $\alpha$, $\beta$ and $\gamma$) and training data into the Holt-Winters model to obtain a forecasting result, then we calculate the root mean square error (RMSE) between the forecasting results and the validation data set.
- Calculate the smell concentration: we choose the RMSE as the smell concentration ($Smell_i$), and it is calculated as follows. $$RMSE = \sqrt{\sum_m (y_m-\hat{y_m})^2}
\label{eq:9}$$ Where $y_m$ is the data in validation data set, and $\hat{y_m}$ is the data of forecasting results.
- Find the best individual: Find out the three fruit flies with maximal smell concentration (finding the maximal value) among the fruit fly swarm, as equation (14).
- Fruit fly swarm movement: Select the best three fruit fly individuals $bestIndex$, and save the smell concentration judgment value. Keep the best direction, and at this moment, the fruit fly swarm will use vision to fly towards that best location, as follows. $$x_0^\alpha = X_{bestIndex}^\alpha, y_0^\alpha = Y_{bestIndex}^\alpha,
\label{eq:9}$$ $$x_0^\beta = X_{bestIndex}^\beta, y_0^\beta = Y_{bestIndex}^\beta,
\label{eq:9}$$ $$x_0^\gamma = X_{bestIndex}^\gamma, y_0^\gamma = Y_{bestIndex}^\gamma,
\label{eq:9}$$
- Population evolution: Repeat the implementations of step 3 to step 7, when the smell concentration is not better than the previous iteration results any more or the number of iteration reaches the $maxgen$, then we get the best smell concentration judgment value, i.e. the optimal smoothing parameters $\alpha_{opt}$, $\beta_{opt}$ and $\gamma_{opt}$ and go to step 9.
- Forecasting: Feed the optimal smoothing parameters and the training data set into the Holt-Winters Model to obtain the final forecasting result.
Numerical example {#Example}
=================
Two examples are selected to verify the effectiveness of the proposed FOA-MHW model in this paper, which are the monthly electricity consumption forecasting of a southern city in China and the telecommunications and television industry in this city. In order to testify the performance of the proposed algorithms, three typical algorithms are introduced herein as the benchmarks, they are the seasonal index (SI) model [@WANG2012109], the MHW model with default parameters [@ForeastBook] and the GASVR model [@OYEHAN201885]. In addition, to verify the performance gain of our FOA-MHW model under small sample data set, we evaluate the performance of FOA-MHW model over different lengthes of training data set. The performance metric used herein is the mean absolute percentage error (MAPE) and it is defined as below. $$MAPE = \frac{1}{N}\sum_{t=1}^N \left| \frac{\hat{y_t}-y_t}{y_t} \right|
\label{eq:9}$$ Where $N$ is the number of forecasting data and it takes the value of 6 in our simulation. $y_t$ is real value of the electricity consumption, and the $\hat{y_t}$ is forecasting result.
The experiment¡¯s environment includes MATLAB R2016b, self-written MATLAB program and the computer with the Intel(R) Core(TM) i5-6500 3.2 GHz CPU, 16 GB RAM and Windows 7 professional system.
Data set {#Dataset}
--------
![$~$The details of the sample data set.[]{data-label="dataset"}](dataset.eps)
The sample data set is provided by a Chinese energy supply company, which includes the monthly electricity consumption of a southern city in China from Jan. 2010 to Dec. 2018 and the monthly electricity consumption of the telecommunications and television industry in this city from Jan. 2010 to Dec. 2018. Since the meters of the users collect and report the electricity consumption every two months to the power management system, thus we only have 6 data points per year. Therefore, there are 54 data points in the sample data set. We select the data from Jan. 2010 to Dec. 2017 as training data set, i.e., the first 48 data points, and the data from Jan. 2018 to Dec. 2018 is used as the test data set, i.e., the last 6 points. Furthermore, we divide the training data set into optimizational training data set and validation data set, the optimizational training data set contains the first 42 data points of the training data set, and the validation data set contains the last 6 data points of the training data set. The optimizational training data set and the validation data set are used to conduct the FOA to search the optimal smoothing parameters for the Holt-Winters model, while the test data set is used to evaluate the performance of the finally forecasting results. The details of the sample data set is as Figure \[dataset\].
The forecasting results of considered city {#Province}
------------------------------------------
The real monthly electricity consumption data of the considered city from Jan. 2010 to Dec. 2018 shows in Figure \[ProvinceMES\], and the forecasting process is as follows.
![$~$The monthly electricity consumption of considered city.[]{data-label="ProvinceMES"}](ProvinceMES.eps)
- Data set processing: According to the method presented in \[Dataset\], we get the training data set and test data.
- Forecasting: Implement the process of FOA-MHW model which described in \[FOA-HW\]. Specifically, initialize $maxgen$ and $sizepop$ by 20 and 50, respectively, $FX^\eta$ and $FY^\eta$ follow an independent uniform distribution of $[5, 10]$, $\eta\in{\alpha, \beta, \gamma}$, and the initial locations of these fruit fly swarms follow an independent uniform distribution over $[0, 1]$, i.e., $x_0^\eta, y_0^\eta \in[0,1],\eta\in{\alpha, \beta, \gamma}$.
- Comparative analysis: We implement the seasonal index model, the HW model with default parameters and GASVR model in the same data set, where the smoothing parameters of Holt-winters exponential smoothing method is set to a default parameters, $\alpha = 0.2$, $\beta = 0.1$ and $\gamma = 0.6$ [@ForeastBook].
![$~$The search route of fruit fly swarm for $\alpha$, $\beta$ and $\gamma$ for the considered city.[]{data-label="ProvinceRoute"}](ProvinceRoute.eps)
![$~$The update process of optimal smoothing parameters in considered city.[]{data-label="ProvinceParameters"}](ProvinceParameters.eps)
Therefore, by performing the forecasting process described above, we have the search route of the three fruit fly swarms as shown in Figure \[ProvinceRoute\], and the update process of the optimal smoothing parameters is shown in Figure \[ProvinceParameters\]. One can note that the three fruit fly swarms can find the food quickly, i.e., the iteration of the smoothing parameters is finally converged. The optimal smoothing parameters selected by FOA are $\alpha = 0.5562$, $\beta = 0.2022$ and $\gamma = 0.3590$, however it’s very difficult to obtain the optimal smoothing parameters by manually adjusting.
![$~$The forecasting results of different models in considered city.[]{data-label="ProvinceResult"}](ProvinceResult.eps)
The forecasting results of different models are shown in Figure \[ProvinceResult\], meanwhile, we have compared the values of relative errors every month and the total MAPE for different models and they are listed in Table \[tab:MAPEProvince\]. Firstly, we can note that, the proposed FOA-MHW model has the smallest MAPE, i.e., 3.65%, and both the SI model and the MHW model with default parameters can achieve a better performance than the GASVR model. That is, though the time series of the monthly electricity consumption is highly randomness, leverage the optimal parameter selection of the fruit fly optimization algorithm (FOA), the FOA-MHW model can overcome this randomness at an acceptable level. However, since the order of the SI model and the HW model with default parameters are both lower than the FOA-MHW model, which then make them obtain worse performance than the FOA-MHW.
Month SI model MHW-default GASVR FOA-MHW
------------ ---------- ------------- ------- ---------
Feb. 2018 5.40 7.00 3.07 7.20
Apr. 2018 2.16 1.84 2.63 3.29
July. 2018 1.52 3.28 5.59 1.54
Aug. 2018 4.82 6.82 11.58 1.66
Oct. 2018 8.15 6.35 13.80 0.39
Dec. 2018 4.26 1.81 5.35 7.84
MAPE 4.38 4.52 7.00 3.65
: \[tab:MAPEProvince\]The relative errors and MAPE of different models in considered city (%)
The forecasting results of the telecommunications and television industry for the considered city
-------------------------------------------------------------------------------------------------
In order to further evaluate the performance of our model and understand its potential insight, we conduct another simulation but use a particular industry electricity consumption data, i.e., the telecommunications and television industry and which is shown in Figure \[IndustryMES\]. One can note that, now, the time series of this industry have a better periodicity than the monthly electricity consumption of the considered city, i.e., the Figuie \[ProvinceMES\]. The forecasting process and the parameter settings are the same as that described in section \[Province\].
The search route of the three fruit fly swarms are shown in Figure \[IndustryRoute\], and the update process of the optimal smoothing parameters is shown in Figure \[IndustryParameters\]. Similarly, the three fruit fly swarms can find the optimal smoothing parameters very quickly. The optimal smoothing parameters selected by FOA is $\alpha = 0.7992$, $\beta = 0.3556$ and $\gamma = 0.9893$.
![$~$The monthly electricity consumption of the telecommunications and television industry for the considered city.[]{data-label="IndustryMES"}](IndustryMES.eps)
![$~$The search route of fruit fly swarm for $\alpha$, $\beta$ and $\gamma$ of telecommunications and television industry.[]{data-label="IndustryRoute"}](IndustryRoute.eps)
![$~$The update process of optimal smoothing parameters of telecommunications and television industry.[]{data-label="IndustryParameters"}](IndustryParameters.eps)
![$~$The forecasting results of telecommunications and television industry.[]{data-label="IndustryResult"}](IndustryResult.eps)
From the Figure \[IndustryResult\] and the Table \[MAPEIndustry\], we can observe that, the FOA-MHW model still obtains the best performance over the other benchmark models in this particular industry. In specifically, the MAPE of our FOA-MHW model is 1.89%, and the MAPE of the GASVR model and the SI model are 2.71% and 9.05%, respectively. And this result confirms that the FOA-MHW model can always get better performance than these three models for different scenarios.
Month SI model MHW-default GASVR FOA-MHW
------------ ---------- ------------- ------- ---------
Feb. 2018 7.68 2.60 1.04 2.33
Apr. 2018 11.35 5.48 4.13 0.41
July. 2018 10.22 5.26 1.38 0.06
Aug. 2018 6.15 4.19 1.04 4.23
Oct. 2018 7.86 1.19 4.92 1.96
Dec. 2018 11.04 0.37 3.76 2.34
MAPE 9.05 3.18 2.71 1.89
: \[MAPEIndustry\]The relative errors and MAPE of different models in telecommunications and television industry(%)
At last, in order to evaluate the performance of our model in the case of small sample data set, the monthly electricity consumption with different lengthes of the training data set are forecasted. We use at least 3 years training data, i.e. 18 data points, and no more than 8 years training data, i.e. 48 data points. The MAPE of different models are shown in Figure \[IndustryYear\], we can find that, with the decreasing of the length of training data sets, the MAPE of the GA-SVR model and MHW model with default parameters increasing. In more detail, the MAPE of HW model with default parameters and the GASVR model increase linearly as the training data decreases, but the MAPE of our FOA-MHW model increase little as the training data decreases, and the MAPE of FOA-MHW model is 3.58% even though there are only 3 years training data. One can note that, the MAPE of SI model is smallest when the training data less than 5 years, the reason is that the SI model can achieve a perfect performance when the data set has a regular periodicity, and the data from 2014 to 2018 is quite periodic in Figure \[IndustryMES\], so the SI model can get a best result, but the data from 2010-2013 isn’t periodic enough in Figure \[IndustryMES\], which make the result of SI model is worse than other model. Besides, the MAPE of FOA-MHW model is smaller than the HW model with default parameters and the GASVR model all the time. The reason is that the machine learning method GASVR model need a lot of training data to get a good result, but the FOA-MHW model has a smaller data requirement. Thus, the FOA-MHW model applies to the situations of little training data available, where the machine learning method can’t deal with.
![$~$The MAPE of different models in different lengthes of training data in telecommunications and television industry.[]{data-label="IndustryYear"}](IndustryYear.eps)
Conclusion {#Conclusion}
==========
In this paper, we deal with the problem of accurate monthly electricity consumption forecasting. In more detail, a new hybrid forecasting model named FOA-MHW is proposed herein. And the original idea is that, the fruit fly optimization algorithm is adopted for the parameters selection of the Holt-Winters exponential smoothing method. Based on the collected real electricity consumption data, the performance of the proposed model are testified over different scenarios. From which, we can conclude that, our proposed FOA-MHW model always achieve the best performance over other benchmark models, i.e., SI model, GASVR model and the MHW model with default parameters. In particular, our further experiments indicate that the proposed FOA-MHW model can obtain an excellent forecasting performance even with only a small number of sample data, e.g., the MAPE of FOA-MHW model is 3.58% even though there are only 3 years training data. And now, our proposed FOA-MHW model has been deployed in a Chinese energy supply company to help them to forecast the monthly electricity consumption and its performance is well proved.
Reference
=========
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In Higgsless models, new vector resonances appear to restore the unitarity of the $W_L W_L$ scattering amplitude without the Higgs boson. In the ideal delocalized three site Higgsless model, one of large production cross section of the neutral vector resonance ($Z^{\prime}$) at the Large Hadron Collider is the $W$-associated production, $pp \to Z^{\prime}W^{\pm} \to W^{\mp}W^{\pm}W^{\pm}$. Although the dileptonic decay channel, $l \nu l^{\prime} \nu^{\prime} jj$, is experimentally clean to search for the $Z''$ signals, it is difficult to reconstruct the $Z^{\prime}$ invariant mass due to the two neutrinos in the final state. We study collider signatures of $Z^{\prime}$ using the $M_{T2}$-Assisted On-Shell (MAOS) reconstruction of the missing neutrino momenta. We show the prospect of the $Z^{\prime}$ mass determination in the channel, $l \nu l^{\prime} \nu^{\prime} jj$, at the Large Hadron Collider.'
---
TU-874
[ **Measurements of neutral vector resonance in Higgsless models at the LHC**]{}\
\
$^a$[*Department of Physics, Tohoku University, Sendai 980-8578, Japan*]{}\
$^b$[*IIAIR, Tohoku University, Sendai 980-8578, Japan*]{}\
Introduction {#sect:intro}
==============
The Large Hadron Collider (LHC) experiments are now operating. At the LHC, a new electroweak signal will be discovered because the unitarity of the $W_L W_L$ scattering amplitude is violated at the scale higher around 1 TeV due to the $E^2$ dependence. One of the candidates is a signal from a scalar $SU(2)_L$ doublet, called the Higgs boson. Due to the contribution from the Higgs boson, the $E^2$ dependence of the $W_L W_L$ scattering amplitude vanishes in the standard model (SM).
On the other hand, it is also possible to maintain the unitarity without the Higgs scalar in Higgsless models [@Csaki:2003dt; @Csaki:2003zu]. In Higgsless models, vector resonances which are responsible for the restoration of the unitarity of the $W_L W_L$ scattering amplitude will be produced at the LHC. The masses are less than around 1 TeV to avoid the unitarity violation, but such particles are strictly constrained by direct observation if their couplings to SM fermions are similar to the SM ones [@Amsler:2008zzb]. Also the coupling induces a large correction to electroweak precision measurements at the tree level. To avoid the constraints, the couplings should be suppressed and it is realized by the fermion delocalization [@Cacciapaglia:2004rb; @Foadi:2004ps; @Chivukula:2005bn; @Casalbuoni:2005rs; @Chivukula:2005xm; @Chivukula:2005cc]. The vector resonances therefore have only small coupling to the SM fermions in Higgsless models.
Fermiophobic vector resonances can be produced mainly by the weak gauge boson associated process or the vector boson fusion process and mainly decay into the weak gauge bosons at the LHC. The measurements of charged vector resonance have been studied [@Birkedal:2004au; @He:2007ge; @Alves:2008up; @Bian:2009kf; @Han:2009qr]. However, in a case that the $W$-associated production is dominant, the mass determination of the neutral vector resonance ($Z^{\prime}$) is very difficult because there are three $W$ bosons in the final state. A way using $W^{\pm}W^{\pm}W^{\mp} \to l^{\pm} \nu l^{\prime \pm} \nu^{\prime} jj$ mode is proposed by Tao Han et.al.[@Han:2009qr], and they showed that the mass can be read off by an endpoint of the $m(l j j)$ distribution from the parton level study.
In this paper, we show that neutral vector resonance can be reconstructed even in $W W W \to l \nu l^{\prime} \nu^{\prime} jj$ final state using $M_{T2}$ Assisted On-Shell (MAOS) reconstruction of missing momenta [@Cho:2008tj; @Choi:2009hn; @Choi:2010dw]. Since MAOS momenta equal true neutrino four momenta at the $M_{T2}$ [@Lester:1999tx] endpoint of $W W \to l \nu l^{\prime} \nu^{\prime}$ system, the $Z^{\prime}$ mass will be determined by the peak of the invariant mass of reconstructed $W^{\pm} W^{\mp}$ bosons.
This article is organized as follows. In the next section, we briefly review the Higgsless models. In our study, we focus on the three site Higgsless model [@Chivukula:2006cg] as a benchmark model. In section 3, we briefly review the MAOS momenta. The Monte Carlo simulation study for measurements of neutral vector resonance at the LHC are discussed in section 4, where we show the mass will be measured with an integrated luminosity of $100$ ${\rm fb}^{-1}$ at $\sqrt{s} = 14 {\rm TeV}$. Section 5 is devoted to summary and discussion.
Higgsless models {#sect:model}
==================
In continuum five-dimensional gauge theory, the cancellations of $E^2$ and $E^4$ dependence of $W_L W_L$ scattering amplitude are guaranteed by following sum rules [@Csaki:2003dt; @SekharChivukula:2008mj]: $$\begin{aligned}
\sum^{\infty}_{i=1} g_{Z_i WW}^2
&=& g_{WWWW} - g_{ZWW}^2 - g_{\gamma WW}^2,
\\ \nonumber
3 \sum^{\infty}_{i=1} g_{Z_i WW}^2 M_{Z_i}^2
&=& 4 g_{WWWW} M_W^2 - 3 g_{ZWW}^2 M_Z^2,
\label{fig:sum_rule_1}\end{aligned}$$ where $g_{WWWW}$, $g_{ZWW}$ and $g_{\gamma WW}$ are the SM $WWWW$, $WWZ$ and $\gamma WW$ coupling constants, respectively. The $Z_i$ represents the $i$-th KK excitation and $Z_1$ is also denoted $Z^{\prime}$ in this paper. The $Z^{\prime}$ mass should be less than the unitarity violation scale $\sqrt{8 \pi} v \sim $ 1.2 TeV if there is no Higgs boson nor other vector resonances.
In this paper, we study the collider signatures of the three site Higgsless model [@Chivukula:2006cg] as a benchmark model of Higgsless models. The model contains many essential ingredients of Higgsless models and a gauge invariant four-dimensional effective theory of Higgsless models. We briefly review the three site Higgsless model.
The three site Higgsless model is a deconstructed Higgsless model with only three sites [^1]. The model is based on two nonlinear $( SU(2) \times SU(2) )/SU(2)$ sigma models. The non-linear sigma fields, $U_1$ and $U_2$, are given as $$\begin{aligned}
U_i &=& e^{i \pi_i^a \tau^a / f_i} \quad {\rm for } \ i=1,2,
\label{fig:U_i}\end{aligned}$$ where $\pi_i^a$ are Nambu-Goldstone bosons and $\tau^a$ are the Pauli matrices. The model incorporates $SU(2) \times SU(2) \times U(1)$ gauge symmetry with gauge coupling strengths $g_0$, $g_1$ and $g_2$, respectively. The vacuum expectation values (VEVs) $f_1$ and $f_2$ break the $SU(2) \times SU(2) \times U(1)$ gauge symmetry to $U(1)_{em}$: $$\begin{aligned}
\mathcal{L}_{gauge}
&=& \sum_{i=1}^2 \frac{f_i^2}{4}
{\rm Tr} \left[ (D_{\mu}U_i)^{\dagger} (D^{\mu}U_i) \right]
- \sum_{i=0}^2 \frac{1}{2}
{\rm Tr} \left[ {\bf V}_{i \mu \nu} {\bf V}_i^{\mu \nu} \right],
\\ \nonumber
D_{\mu}U_i
&\equiv& \partial_{\mu} U_i
+ i g_{(i-1)} {\bf V}_{(i-1) \mu} U_i - i g_i U_i {\bf V}_{i \mu},
\qquad
{\bf V}_{i \mu} = \sum_{a=1,2,3} \frac{\tau^a}{2} V_{i \mu}^a
\quad {\rm for } \ i=0,1,
\\ \nonumber
{\bf V}_{i \mu \nu}
&\equiv& \partial_{\mu} {\bf V}_{i \nu}
- \partial_{\nu} {\bf V}_{i \mu}
+ i g_i \left[ {\bf V}_{i \mu}, {\bf V}_{i \nu}\right],
\qquad
{\bf V}_{2 \mu} = \frac{\tau^3}{2} V_{2 \mu}^3,
\label{fig:L_gauge}\end{aligned}$$ where the $V_{0(1)}^a$ is the gauge field for the $SU(2)$ gauge group at the site 0(1) and $V_2^3$ is the gauge field for the $U(1)$ gauge group at the site 2 which is embedded as the $\tau_3$ generator of $SU(2)$. Therefore, this model contains charged and neutral gauge bosons, $W^{\prime}$ and $Z^{\prime}$, in addition to the SM $W$, $Z$ and photon. The combinations of $V_{0,1}^{1,2}$ correspond to the $W^{\prime}$ and $W$. On the other hand, the combination of $V_{0,1,2}^3$ are the $Z^{\prime}$, $Z$ and photon, respectively. For simplicity, we take $f_1 = f_2 = \sqrt{2} v$ in this study.
The left-handed fermions $\psi_{L0}$ and $\psi_{L1}$ are $SU(2)$ doublets coupling to the groups at the sites 0 and 1, respectively. On the other hand, the right handed fermions $\psi_{R1}$ are SU(2) doublets coupling to the group at site 1 and $u_{R2}$, $d_{R2}$ and $e_{R2}$ are singlets coupling to the group at site 2. The quantum numbers of the fermions are shown in Table \[tab:quantum\_num\].
The Yukawa couplings are written by $$\begin{aligned}
\mathcal{L}_{Yukawa}
&=& \lambda f_1 \bar{\psi}_{L0} U_1 \psi_{R1}
+ f_2 \bar{\psi}_{L1} U_2
\begin{pmatrix}
\lambda^{\prime}_u & 0 \\
0 & \lambda^{\prime}_d
\end{pmatrix}
\begin{pmatrix}
u_{R2} \\ d_{R2}
\end{pmatrix}
+ M \bar{\psi}_{L1}\psi_{R1}
+ {\rm h.c.}
\\ \nonumber
&=&
\begin{pmatrix}
\bar{\psi}_{L0} & \bar{\psi}_{L1}
\end{pmatrix}
M
\begin{pmatrix}
\varepsilon_L & 0 \\
1 & \varepsilon_{uR,dR}
\end{pmatrix}
\begin{pmatrix}
\psi_{R1} \\ u_{R2},d_{R2}
\end{pmatrix}
+ {\rm h.c.} ,
\label{fig:L_Yukawa}\end{aligned}$$ where $M \varepsilon_L = \lambda f_1 \equiv m$ and $M \varepsilon_{uR,dR} = \lambda^{\prime}_{u,d} f_2 \equiv m^{\prime}_{u,d}$. In this paper, we assume that the $m$ and $M$ are universal for the generation and quark-lepton.
In the limit, $g_0/g_1 \ll 1$ and $g_2/g_1 \ll 1$, the gauge boson masses are given by $$\begin{aligned}
M_W
&\sim& \frac{g_0^2}{4} v^2
\left[ 1 - \frac{1}{4} \left( \frac{g_0}{g_1} \right)^2 \right],
\quad
M_{W^{\prime}}
\sim g_1^2 v^2
\left[ 1 + \frac{1}{4} \left( \frac{g_0}{g_1} \right)^2 \right],
\\ \nonumber
M_Z
&\sim& \frac{g_0^2}{4 c^2} v^2
\left[ 1 - \frac{(c^2 - s^2)^2}{4c^2}
\left( \frac{g_0}{g_1} \right)^2 \right],
\quad
M_{Z^{\prime}}
\sim g_1^2 v^2
\left[ 1 + \frac{1}{4 c^2} \left( \frac{g_0}{g_1} \right)^2 \right].
\label{fig:mass_gauge}\end{aligned}$$ Here $s=\sin\theta$, $c=\cos\theta$, and $\tan\theta=g_2/g_0$. In the $\varepsilon_L \ll 1$ limit, fermion masses are written as the following: $$\begin{aligned}
m_{t,b}
&\sim& \frac{ m m_{t,b}^{\prime} }{ \sqrt{M^2 + m_{t,b}^{\prime 2}} },
\quad
M_{T,B} \sim \sqrt{M^2 + m_{t,b}^{\prime 2}},
\label{fig:mass_fermion}\end{aligned}$$ where we show the masses of the third-generation quarks as an example. The $M_T(M_B)$ is the new heavy top(bottom) quark mass.
The couplings between the light fermions and the heavy gauge bosons are constrained by direct heavy gauge boson search and the electroweak precision measurements. The coupling can be small by delocalizing fermion and $\varepsilon_L$ is a parameter which denotes the degree of the delocalization. In the case, $\varepsilon_L = (1 + \varepsilon_{fR}^2)^2
\{(g_0/g_1)^2/2 + \mathcal{O}((g_0/g_1)^4) + \cdots\}$, the coupling constants are given by $$\begin{aligned}
g_{ffW^{\prime}}
&=& 0.
\label{fig:couping_ffZ}\end{aligned}$$ This case is called ideal delocalization in which the precision electroweak corrections are minimized at tree level [@Chivukula:2005xm] [^2]. In the ideal delocalization case, the lower bound of $W^\prime$ mass, $380$ GeV, is given by the bound on the triple gauge vertex from the LEP-II experiments [@Chivukula:2006cg].
MAOS momentum {#sect:MAOS}
===============
In this section, we briefly review the MAOS momentum [@Cho:2008tj; @Choi:2009hn; @Choi:2010dw]. The MAOS momentum is defined using the $M_{T2}$ formula [@Lester:1999tx]. In the $WW \to l(p) \nu(k) l^{\prime}(p^{\prime}) \nu^{\prime}(k^{\prime}) $ system, the $M_{T2}$ valuable is written by $$\begin{aligned}
M_{T2}^2
&=& \min_{ {\bf k}_T + {\bf k}_T^{\prime} = {\mathbf{p} \hspace{-0.5 em}/}_T }
\left[ \max \left\{ M_T^2 ({\bf p}_T,{\bf k}_T),
M_T^2 ({\bf p}_T^{\prime},{\bf k}_T^{\prime})\right\} \right],
\label{fig:MT2_1}\end{aligned}$$ where ${\mathbf{p} \hspace{-0.5 em}/}_T$ is the missing transverse momentum and the transverse mass, $M_T$, is defined by $$\begin{aligned}
M_T^2 ({\bf p}_T,{\bf k}_T)
&=& 2 \left( |{\bf p}_T||{\bf k}_T|
- {\bf p}_T \cdot {\bf k}_T
\right),
\label{fig:MT2_2}\end{aligned}$$ where the lepton masses are neglected.
The MAOS momenta, $k^{MAOS}$ and $k^{\prime MAOS}$, are defined by the following. Assuming ${\mathbf{p} \hspace{-0.5 em}/}_T = - ( {\bf p}_T + {\bf p}_T^{\prime} )$, the transverse momenta of $k^{MAOS}$ and $k^{\prime MAOS}$ are defined by $$\begin{aligned}
{\bf k}_T^{MAOS} &=& - {\bf p}_T^{\prime},
\quad
{\bf k}_T^{\prime MAOS} = - {\bf p}_T,
\label{fig:MAOS_2}\end{aligned}$$ and $$\begin{aligned}
M_{T2}( {\bf p}_T,{\bf p}_T^{\prime},{\mathbf{p} \hspace{-0.5 em}/}_T)
&=& \sqrt{2}\sqrt{ \left( |{\bf p}_T||{\bf k}_T^{MAOS}|
- {\bf p}_T \cdot {\bf k}_T^{MAOS}
\right)}, \\ \nonumber
&=& \sqrt{2}\sqrt{ \left( |{\bf p}_T^{\prime}||{\bf k}_T^{\prime MAOS}|
- {\bf p}_T^{\prime} \cdot {\bf k}_T^{\prime MAOS}
\right)}.
\label{fig:MAOS_1}\end{aligned}$$ The MAOS momenta are also required to satisfy the following conditions in the $WW \to l(p) \nu(k) l^{\prime}(p^{\prime}) \nu^{\prime}(k^{\prime}) $ system: $$\begin{aligned}
\left( k^{MAOS} \right)^2
&=& \left( k^{\prime MAOS} \right)^2 = 0, \\ \nonumber
\left( p + k^{MAOS} \right)^2
&=& \left( p^{\prime} + k^{\prime MAOS} \right)^2 = M_{T2}^2.
\label{fig:MAOS_3}\end{aligned}$$ From these equations, the longitudinal momenta of $k^{MAOS}$ and $k^{\prime MAOS}$ are also determined by $$\begin{aligned}
k_L^{MAOS} (\pm)
&=& \frac{|{\bf k}_T^{MAOS}|}{|{\bf p}_T|} p_L,
\quad
k_L^{\prime MAOS} (\pm)
= \frac{|{\bf k}_T^{\prime MAOS}|}{|{\bf p}_T^{\prime}|} p_L^{\prime}.
\label{fig:MAOS_4}\end{aligned}$$ At the $M_{T2}$ endpoint, the MAOS momenta become equal to the final state neutrino momenta $k, k^{\prime}$, respectively.
In the next section, we show that the fermiophobic $Z^{\prime}$ can be reconstructed using MAOS momenta of neutrinos at the LHC through Monte Carlo simulations. The $Z^{\prime}$ mass can be measured by the invariant mass peak of the reconstructed $WW$.
Monte Carlo Simulation {#sect:simulation}
========================
We study the possibility of $Z^{\prime}$ reconstruction through Monte Carlo simulations in this section. We investigate the ideal delocalized three site Higgsless model as a benchmark model and the parameters at the representative points are summarized in Table \[tab:representative points\].
For the Monte Carlo simulation, we have produced the signal parton events by Madgraph/Madevent [@Alwall:2007st; @Christensen:2009jx] and they have been hadronized by PYTHIA [@Sjostrand:2003wg]. For the SM backgrounds, we have generated the events using ALPGEN [@Mangano:2002ea] and HERWIG [@Corcella:2000bw; @Corcella:2002jc]. Our detector simulation is based on ACERDET [@RichterWas:2002ch]. Hereafter we assume an integrated luminosity of ${\cal L}=100$ fb$^{-1}$.
In our analysis, we concentrate on the $W$-associated production depicted in Fig. \[fig:wwz\], especially in the following dilepton mode, $$\begin{aligned}
pp\to W^\pm Z' \to \left\{
\begin{array}{ll}
W^\pm\ (W^+ W^-) \to (l^\pm \nu) ((l^\pm\nu) (jj))~~(\mathrm{same~sign}),\\
W^\pm\ (W^+ W^-) \to (jj) ((l^\pm \nu) (l^\mp\nu))~~(\mathrm{opposite~sign}),
\end{array}
\right.\end{aligned}$$ because the large SM backgrounds can be reduced in the dilepton modes. In particular, it is shown that the $t\bar{t}$ background is significantly reduced for the same sign dilepton mode [@Han:2009qr].
![(Left) Feynman diagram for the same sign dilepton mode. (Right) Feynman diagram for the opposite sign dilepton mode. []{data-label="fig:wwz"}](Fig_WWZ_same.eps "fig:"){width="6cm"} ![(Left) Feynman diagram for the same sign dilepton mode. (Right) Feynman diagram for the opposite sign dilepton mode. []{data-label="fig:wwz"}](Fig_WWZ_odd.eps "fig:"){width="6cm"}
In the signal mode, there are three $W$ bosons; two $W$ bosons decay leptonically and one $W$ decays hadronically. In order to reconstruct the hadronically decaying $W$, we require the following cuts:
- at least two and less than three hard jets with $p_T>20$ GeV,
- 65 GeV $< m_{jj} < $ 95 GeV,
where the $m_{jj}$ is the invariant mass of two hard jets. We also impose $b$-jet veto to reduce $t \bar{t}$ events.
To reconstruct the two $W$ bosons which decay leptonically, we use neutrino MAOS momenta which are defined by a $M_{T2}$ valuable using two leptons and missing momenta. In the same (opposite) sign charge mode, we require the following:
- two leptons with the same (opposite) charge with $p_T > 10$ GeV,
- ${\mathrm{p} \hspace{-0.5 em}/_T}> 50$ GeV.
In the same sign dilepton mode, the requirement of the same charge lepton reduce the $t \bar{t}$ events significantly. On the other hand, huge $t \bar{t}$ events still remain for the opposite sign dilepton mode. To reduce the $t \bar{t}$ events, we also define another $M_{T2}$ valuable, $M_{T2}^{jj}$, in which we define the “effective missing $p_T$”, ${{\mathbf{P} \hspace{-0.5 em}/}}^{\rm eff}_T = {\mathbf{p} \hspace{-0.5 em}/}_T + {\bf p}_{T l1} + {\bf p}_{Tl2}$ and the mass of missing particle, $m_X = m_W$, in the opposite charge mode. The $M_{T2}^{jj}$ valuable should have a maximum value at the top quark mass taking the missing particle mass, $m_X = m_W$, because we deal with two $W$ bosons decaying leptonically as missing particles. In Fig. \[fig:Mt2jj\], we show the $M_{T2}^{jj}(m_X = m_W)$ distributions for the signal with $m_{Z'}=380$ GeV and the $t\overline{t}$ events after the jet and lepton cuts. We can see a sharp endpoint around the top quark mass for the $t\overline{t}$ events while the signal events give larger $M_{T2}^{jj}$. Therefore, we impose $M_{T2}^{jj}(m_X = m_W) > 200$ GeV to cut the $t \bar{t}$ events. The procedure reduces the combinatorial ambiguity of the jets and the leptons than a procedure using $M_{T2}^{jljl}$ which should have a maximum value at the top quark mass taking the missing particle mass, $m_X = 0 ~\rm{GeV}$.
![The $M_{T2}^{jj}$ distribution in the opposite sign dilepton mode for $m_{Z^{\prime}} = 380$ GeV (white) and the $t\overline{t}$ backgrounds (hatched). Here the integrated luminosity is $100$ fb$^{-1}$ and the mass of the missing particle is set to $m_W$.[]{data-label="fig:Mt2jj"}](ttbarMT2.eps){width="8cm"}
In addition, we also impose other cuts, the value of $M_\mathrm{eff}$, which is defined as a scalar sum of the visible and the missing momenta [@Hinchliffe:1996iu], and invariant mass of two leptons, $m_{ll}$, listed in Table \[tab:CUTs\]. The $m_{ll}$ cuts reduce the $Z$ boson and $t \bar{t}$ backgrounds because a lepton from a top quark decay is softer than a lepton in the signal events. In Table \[table:cut\_same\] (\[table:cut\_odd\]), we show the number of events after the selection cuts for the same (opposite) sign mode.
From the MAOS momenta, we can reconstruct the invariant mass of the $Z'$ as follows: $$\begin{aligned}
m^2_{Z'}= \left\{
\begin{array}{ll}
(p+k^{MAOS}+p_{j_1}+p_{j_2})^2~~~(\mathrm{same~sign}), \\
(p+k^{MAOS}+p'+k'^{MAOS})^2~~(\mathrm{opposite~sign}).
\end{array}
\right.\end{aligned}$$ In both modes, there is a two-fold ambiguity to reconstruct the $Z^\prime$ invariant mass from two $W$ bosons. In the same sign mode, there are two choices to combine the lepton momenta with the hadron momenta. We choose the combination which gives the smaller $m_{Z'}$ in our analysis. On the other hand, we reconstruct $Z^\prime$ invariant mass from two $W$ bosons which decay leptonically in order to confirm the charge of the $Z^\prime$ in the opposite sign mode.
In Figs. \[fig:mZp380\] and \[fig:mZp380\_odd\], we show the invariant mass distributions of $Z'$ in the same and opposite sign dilepton mode for $m_{Z'}=380$ GeV, respectively. The hatched histograms are the SM backgrounds. Although the MAOS momenta become equal to the true neutrino momenta only at the $M_{T2}$ endpoint, $M_{T2}\simeq 80$ GeV, we use the MAOS momenta in wide $M_{T2}$ regions and plot the invariant mass distributions of $Z'$ for $M_{T2}>60, 40, 20, 0$ GeV. [^3] In Fig. \[fig:mZp380\] we can see the peaks around the true $Z'$ mass regions for all the cases. The shape becomes sharper when we use the MAOS momenta in the higher $M_{T2}$ regions. However, the number of the events becomes smaller. Thanks to the $M_{T2}^{jj}$ cut, we can see clear signal peaks around the true $m_{Z'}$ value even in the opposite sign mode. Although the shapes of the invariant distributions are slightly irregular compared with the same sign mode, we can determine the $Z'$ mass from the peak.
In Figs. \[fig:mZp500\] and \[fig:mZp500\_odd\], we show the invariant mass distributions of $Z'$ in the same and opposite sign dilepton mode for $m_{Z'}=500$ GeV, respectively. Since the production cross section is smaller compared with the $m_{Z'}=380$ GeV case, the numbers of events are limited if we use the tight $M_{T2}$ cut. In the opposite sign mode, we cannot see clear peaks for the $M_{T2}>20, 40, 60$ GeV cases. However, we can see the peaks around the true $m_{Z'}$ region even for $m_{Z'}=500$ GeV in the both modes.
-------- -------------------------------------------- ---------------- -------- ------------ --
Selection cut Signal top BG W boson BG
(380, 500 GeV)
jet \# of jets ($P_T > 20$ GeV)=2 (4844, 1451) 774624 5862
65 GeV $< m_{jj} < $ 95 GeV (1116, 313) 124142 1485
b veto (1069, 299) 44009 1410
lepton 2 same charge leptons ($P_T > 10$ GeV) (284, 96) 1234 91
${\mathrm{p} \hspace{-0.5 em}/_T}> 50$ GeV (257, 90) 777 64
other $M_\mathrm{eff}>500$ GeV (207, 85) 21 9
$m_{ll}>130$ GeV (167, 74) 0 5
-------- -------------------------------------------- ---------------- -------- ------------ --
: Number of events after selection cuts for the same sign mode with an integrated luminosity ${\cal L}=100^{-1}$ fb.[]{data-label="table:cut_same"}
--------------- -------------------------------------------- ---------------- --------- ------------ --
Selection cut Signal top BG W boson BG
(380, 500 GeV)
jet \# of jets ($P_T > 20$ GeV)=2 (4844, 1451) 774624 5862
65 GeV $< m_{jj} < $ 95 GeV (1116, 313) 124142 1485
b veto (1069, 299) 44009 1410
lepton 2 opposite charge leptons ($P_T > 10$ GeV) (508, 141) 15845 993
${\mathrm{p} \hspace{-0.5 em}/_T}> 50$ GeV (461, 130) 10216 533
$M_{T2}^{jj}$ $M_{T2}^{jj}$ (289, 99) 125 44
other $M_\mathrm{eff}> (600, 700)$ GeV (254, 83) (25, 9) (16, 12)
$m_{ll}>110$ GeV (206, 67) (9, 1) (11, 8)
--------------- -------------------------------------------- ---------------- --------- ------------ --
: Number of events after selection cuts for the opposite sign mode with an integrated luminosity ${\cal L}=100^{-1}$ fb.[]{data-label="table:cut_odd"}
![The invariant mass distributions of $Z'$ in the same sign dilepton mode for $m_{Z^{\prime}} = 380$ GeV with an integrated luminosity ${\cal L}=100^{-1}$ fb. Here we impose the cuts listed in Table \[table:cut\_same\] and $M_{T2} >$ 60(top left), 40(top right), 20(bottom left), 0(bottom right) GeV. The hatched histograms are the SM backgrounds.[]{data-label="fig:mZp380"}](mZp380.eps "fig:"){width="5cm"} ![The invariant mass distributions of $Z'$ in the same sign dilepton mode for $m_{Z^{\prime}} = 380$ GeV with an integrated luminosity ${\cal L}=100^{-1}$ fb. Here we impose the cuts listed in Table \[table:cut\_same\] and $M_{T2} >$ 60(top left), 40(top right), 20(bottom left), 0(bottom right) GeV. The hatched histograms are the SM backgrounds.[]{data-label="fig:mZp380"}](mZp380_cut40.eps "fig:"){width="5cm"}\
![The invariant mass distributions of $Z'$ in the same sign dilepton mode for $m_{Z^{\prime}} = 380$ GeV with an integrated luminosity ${\cal L}=100^{-1}$ fb. Here we impose the cuts listed in Table \[table:cut\_same\] and $M_{T2} >$ 60(top left), 40(top right), 20(bottom left), 0(bottom right) GeV. The hatched histograms are the SM backgrounds.[]{data-label="fig:mZp380"}](mZp380_cut20.eps "fig:"){width="5cm"} ![The invariant mass distributions of $Z'$ in the same sign dilepton mode for $m_{Z^{\prime}} = 380$ GeV with an integrated luminosity ${\cal L}=100^{-1}$ fb. Here we impose the cuts listed in Table \[table:cut\_same\] and $M_{T2} >$ 60(top left), 40(top right), 20(bottom left), 0(bottom right) GeV. The hatched histograms are the SM backgrounds.[]{data-label="fig:mZp380"}](mZp380_full.eps "fig:"){width="5cm"}
![The invariant mass distributions in the opposite sign dilepton mode for $m_{Z^{\prime}} = 380$ GeV with an integrated luminosity ${\cal L}=100^{-1}$ fb. Here we impose the cuts listed in Table \[table:cut\_odd\] and $M_{T2} >$ 60(top left), 40(top right), 20(bottom left), 0(bottom right) GeV. The hatched histograms are the SM backgrounds.[]{data-label="fig:mZp380_odd"}](mZp380_odd.eps "fig:"){width="5cm"} ![The invariant mass distributions in the opposite sign dilepton mode for $m_{Z^{\prime}} = 380$ GeV with an integrated luminosity ${\cal L}=100^{-1}$ fb. Here we impose the cuts listed in Table \[table:cut\_odd\] and $M_{T2} >$ 60(top left), 40(top right), 20(bottom left), 0(bottom right) GeV. The hatched histograms are the SM backgrounds.[]{data-label="fig:mZp380_odd"}](mZp380_cut40_odd.eps "fig:"){width="5cm"}\
![The invariant mass distributions in the opposite sign dilepton mode for $m_{Z^{\prime}} = 380$ GeV with an integrated luminosity ${\cal L}=100^{-1}$ fb. Here we impose the cuts listed in Table \[table:cut\_odd\] and $M_{T2} >$ 60(top left), 40(top right), 20(bottom left), 0(bottom right) GeV. The hatched histograms are the SM backgrounds.[]{data-label="fig:mZp380_odd"}](mZp380_cut20_odd.eps "fig:"){width="5cm"} ![The invariant mass distributions in the opposite sign dilepton mode for $m_{Z^{\prime}} = 380$ GeV with an integrated luminosity ${\cal L}=100^{-1}$ fb. Here we impose the cuts listed in Table \[table:cut\_odd\] and $M_{T2} >$ 60(top left), 40(top right), 20(bottom left), 0(bottom right) GeV. The hatched histograms are the SM backgrounds.[]{data-label="fig:mZp380_odd"}](mZp380_full_odd.eps "fig:"){width="5cm"}
![The invariant mass distributions in the same sign dilepton mode for $m_{Z^{\prime}} = 500$ GeV with an integrated luminosity ${\cal L}=100^{-1}$ fb. Here we impose the cuts listed in Table \[table:cut\_same\] and $M_{T2} >$ 60(top left), 40(top right), 20(bottom left), 0(bottom right) GeV. The hatched histograms are the SM backgrounds.[]{data-label="fig:mZp500"}](mZp500.eps "fig:"){width="5cm"} ![The invariant mass distributions in the same sign dilepton mode for $m_{Z^{\prime}} = 500$ GeV with an integrated luminosity ${\cal L}=100^{-1}$ fb. Here we impose the cuts listed in Table \[table:cut\_same\] and $M_{T2} >$ 60(top left), 40(top right), 20(bottom left), 0(bottom right) GeV. The hatched histograms are the SM backgrounds.[]{data-label="fig:mZp500"}](mZp500_cut40.eps "fig:"){width="5cm"}\
![The invariant mass distributions in the same sign dilepton mode for $m_{Z^{\prime}} = 500$ GeV with an integrated luminosity ${\cal L}=100^{-1}$ fb. Here we impose the cuts listed in Table \[table:cut\_same\] and $M_{T2} >$ 60(top left), 40(top right), 20(bottom left), 0(bottom right) GeV. The hatched histograms are the SM backgrounds.[]{data-label="fig:mZp500"}](mZp500_cut20.eps "fig:"){width="5cm"} ![The invariant mass distributions in the same sign dilepton mode for $m_{Z^{\prime}} = 500$ GeV with an integrated luminosity ${\cal L}=100^{-1}$ fb. Here we impose the cuts listed in Table \[table:cut\_same\] and $M_{T2} >$ 60(top left), 40(top right), 20(bottom left), 0(bottom right) GeV. The hatched histograms are the SM backgrounds.[]{data-label="fig:mZp500"}](mZp500_full.eps "fig:"){width="5cm"}
![The invariant mass distributions in the opposite sign dilepton mode for $m_{Z^{\prime}} = 500$ GeV with an integrated luminosity ${\cal L}=100^{-1}$ fb. Here we take the cuts listed in Table \[table:cut\_odd\] and $M_{T2} >0$ GeV. The hatched histograms are the SM backgrounds.[]{data-label="fig:mZp500_odd"}](mZp500_full_odd.eps){width="5cm"}
Summary and Discussion {#sect:Summary}
=======================
In this paper, we have studied the search for the neutral vector resonance, focusing on a fermiophobic case at the LHC. Such fermiophobic vector bosons are predicted by perturbative TeV scale scenarios without the Higgs boson, and can be produced in the $W$-associated processes at the LHC. We have studied the dilepton decay modes, $pp \to W^{\pm}Z^{\prime} \to W^{\pm}W^{\pm}W^{\mp}
\to l \nu l^{\prime} \nu^{\prime} jj$, since the SM model backgrounds are suppressed. There are two neutrinos in the dilepton modes and we cannot measure each neutrino momentum directly. In order to determine the neutrino momenta we use the MAOS momenta. From the MAOS momenta we can reconstruct the invariant mass of the neutral vector boson and the mass of the neutral vector resonance can be determined by the peak of the invariant mass distribution.
We have applied the method to both the same and the opposite sign dilepton modes. We have found that the $M_{T2}^{jj}$ cut is very effective to reduce the $t\bar{t}$ background, especially in the opposite sign mode. We can obtain clear invariant mass distributions of the vector resonance and the shape of the invariant mass distributions changes depending on the $M_{T2}$ cut. Although the shape becomes broader if we use a looser $M_{T2}$ cut, we have found that the peak positions of the invariant mass distributions do not change drastically. To measure the mass more accurately, we should optimize the $M_{T2}$ cut analysing detailed Monte Carlo simulations.
Let us discuss advantages of our method compared to other analyses on the $Z'$ measurements at the LHC. Since we have studied the dilepton mode in the $Z'$ decay, the SM backgrounds are reduced and we can see the clear peak of the $Z'$ resonance. From the charge of the dileptons, we can confirm the charge of the resonance in the opposite sign mode. In the same sign mode, we can also infer that the charge is neutral unless double charged particles exist. If we consider one lepton mode, $Z^\prime W \to l \nu jjjj$ or $Z^\prime \to W^+W^- \to l \nu jj$, which is produced via the vector boson fusion or Drell-Yan production process, the $Z^{\prime}$ and $W^{\prime}$ resonances may overlap each other due to difficulty of the $W$($Z$) identification at the LHC. However, it is possible to study the $Z'$ peak separately in our method. In the three site Higgsless model, the $Z^\prime$ can also be measured via Drell-Yan production process in the semileptonic decay mode because the couplings with SM fermions are not zero even if we take the ideal delocalization [@Ohl:2008ri; @Alves:2009aa; @Bach:2009zz]. Using MAOS momenta, the $Z^\prime$ may also be measured in the leptonic decay mode, $Z^\prime \to WW \to l \nu l^\prime \nu^\prime $, without $W^\prime$ contamination. The similar signal, ${\rm Higgs} \to WW \to l \nu l^\prime \nu^\prime $, has been studied by using MAOS momenta [@Choi:2009hn; @Choi:2010dw]. Combining the results from the $W$-associated and Drell-Yan production process study with MAOS momenta, we may also determine the $Z^\prime$ couplings with weak gauge bosons and SM fermions.
Although we focus on the $Z^{\prime}$ measurement in the three-site Higgsless model in this paper, the method will be also useful for the mass measurement of general fermiophobic neutral vector resonances.
Acknowledgments {#acknowledgments .unnumbered}
=================
We would like to thank Masafumi Kurachi for discussions and comments. This work is supported in part by the Grant-in-Aid for the Global COE Program Weaving Science Web beyond Particle-matter Hierarchy from the Ministry of Education, Culture, Sports, Science and Technology of Japan (M. A.).
[99]{} C. Csaki, C. Grojean, H. Murayama, L. Pilo and J. Terning, Phys. Rev. D [**69**]{} (2004) 055006 \[arXiv:hep-ph/0305237\]. C. Csaki, C. Grojean, L. Pilo and J. Terning, Phys. Rev. Lett. [**92**]{} (2004) 101802 \[arXiv:hep-ph/0308038\]. C. Amsler [*et al.*]{} \[Particle Data Group\], Phys. Lett. B [**667**]{} (2008) 1. G. Cacciapaglia, C. Csaki, C. Grojean and J. Terning, Phys. Rev. D [**71**]{} (2005) 035015 \[arXiv:hep-ph/0409126\]. R. Foadi, S. Gopalakrishna and C. Schmidt, Phys. Lett. B [**606**]{} (2005) 157 \[arXiv:hep-ph/0409266\]. R. S. Chivukula, E. H. Simmons, H. J. He, M. Kurachi and M. Tanabashi, Phys. Rev. D [**71**]{} (2005) 115001 \[arXiv:hep-ph/0502162\]. R. Casalbuoni, S. De Curtis, D. Dolce and D. Dominici, Phys. Rev. D [**71**]{} (2005) 075015 \[arXiv:hep-ph/0502209\]. R. S. Chivukula, E. H. Simmons, H. J. He, M. Kurachi and M. Tanabashi, Phys. Rev. D [**72**]{} (2005) 015008 \[arXiv:hep-ph/0504114\]. R. S. Chivukula, E. H. Simmons, H. J. He, M. Kurachi and M. Tanabashi, Phys. Rev. D [**72**]{} (2005) 095013 \[arXiv:hep-ph/0509110\].
A. Birkedal, K. Matchev and M. Perelstein, Phys. Rev. Lett. [**94**]{} (2005) 191803 \[arXiv:hep-ph/0412278\]. H. J. He [*et al.*]{}, Phys. Rev. D [**78**]{} (2008) 031701 \[arXiv:0708.2588 \[hep-ph\]\].
A. Alves, O. J. P. Eboli, M. C. Gonzalez-Garcia and J. K. Mizukoshi, Phys. Rev. D [**79**]{} (2009) 035009 \[arXiv:0810.1952 \[hep-ph\]\]. J. G. Bian [*et al.*]{}, Nucl. Phys. B [**819**]{} (2009) 201 \[arXiv:0905.2336 \[hep-ex\]\].
T. Han, H. S. Liu, M. x. Luo, K. Wang and W. Wu, Phys. Rev. D [**80**]{} (2009) 095010 \[arXiv:0908.2186 \[hep-ph\]\]. W. S. Cho, K. Choi, Y. G. Kim and C. B. Park, Phys. Rev. D [**79**]{} (2009) 031701 \[arXiv:0810.4853 \[hep-ph\]\]. K. Choi, S. Choi, J. S. Lee and C. B. Park, Phys. Rev. D [**80**]{} (2009) 073010 \[arXiv:0908.0079 \[hep-ph\]\]. K. Choi, J. S. Lee and C. B. Park, arXiv:1008.2690 \[hep-ph\].
C. G. Lester and D. J. Summers, Phys. Lett. B [**463**]{} (1999) 99 \[arXiv:hep-ph/9906349\]. R. S. Chivukula, B. Coleppa, S. Di Chiara, E. H. Simmons, H. J. He, M. Kurachi and M. Tanabashi, Phys. Rev. D [**74**]{} (2006) 075011 \[arXiv:hep-ph/0607124\].
R. S. Chivukula, H. J. He, M. Kurachi, E. H. Simmons and M. Tanabashi, Phys. Rev. D [**78**]{} (2008) 095003 \[arXiv:0808.1682 \[hep-ph\]\]. R. Casalbuoni, S. De Curtis, D. Dominici and R. Gatto, Phys. Lett. B [**155**]{} (1985) 95. R. Casalbuoni, A. Deandrea, S. De Curtis, D. Dominici, R. Gatto and M. Grazzini, Phys. Rev. D [**53**]{} (1996) 5201 \[arXiv:hep-ph/9510431\]. T. Abe, S. Matsuzaki and M. Tanabashi, Phys. Rev. D [**78**]{} (2008) 055020 \[arXiv:0807.2298 \[hep-ph\]\].
J. Alwall [*et al.*]{}, JHEP [**0709**]{} (2007) 028 \[arXiv:0706.2334 \[hep-ph\]\]. N. D. Christensen [*et al.*]{}, arXiv:0906.2474 \[hep-ph\].
T. Sjostrand, L. Lonnblad, S. Mrenna and P. Z. Skands, arXiv:hep-ph/0308153. M. L. Mangano, M. Moretti, F. Piccinini, R. Pittau and A. D. Polosa, JHEP [**0307**]{} (2003) 001 \[arXiv:hep-ph/0206293\]. G. Corcella [*et al.*]{}, JHEP [**0101**]{} (2001) 010 \[arXiv:hep-ph/0011363\]. G. Corcella [*et al.*]{}, arXiv:hep-ph/0210213. E. Richter-Was, arXiv:hep-ph/0207355.
I. Hinchliffe, F. E. Paige, M. D. Shapiro, J. Soderqvist and W. Yao, Phys. Rev. D [**55**]{}, 5520 (1997) \[arXiv:hep-ph/9610544\].
T. Ohl and C. Speckner, Phys. Rev. D [**78**]{} (2008) 095008 \[arXiv:0809.0023 \[hep-ph\]\].
A. Alves, O. J. P. Eboli, D. Goncalves, M. C. Gonzalez-Garcia and J. K. Mizukoshi, Phys. Rev. D [**80**]{} (2009) 073011 \[arXiv:0907.2915 \[hep-ph\]\].
F. Bach, Master theis, [SPIRES entry](http://www.slac.stanford.edu/spires/find/hep/www?irn=8773432)
[^1]: The gauge sector is equivalent to the Breaking Electroweak Symmetry Strongly (BESS) model [@Casalbuoni:1985kq; @Casalbuoni:1995qt] one.
[^2]: At the 1-loop level, the parameter does not satisfy the electroweak precision measurements [@Abe:2008hb]. But we take the parameter in the following analysis since the correction to the signal in our collider study is small even if we take a parameter which satisfies the electroweak precision measurements at the 1-loop level.
[^3]: When the $Z^\prime$ is heavy enough to decay to two on-shell $W$ bosons, there is the other definition of the MAOS momenta [@Cho:2008tj]. Even if we take the MAOS momenta, the distributions are not drastically changed in our study with $100$ fb$^{-1}$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In a quasi-two-dimensional $ p $-wave superconductor there find six Cooper pairing states which are degenerate within the weak-coupling approach. We show that this degeneracy can be lifted by feedback effect favoring the so-called chiral $ p $-wave state. This effect is based on the anomalous coupling between charge and current in a system with broken time reversal symmetry and parity.'
address: 'Yukawa Institute for Theoretical Physics, Kyoto University, Kyoto 606-8502, Japan'
author:
- 'J. Goryo and M. Sigrist'
---
\#1\#2
The discovery of odd-parity Cooper pairing in Sr$_2$RuO$_4$ led to the renewed interest in so-called p-wave (spin-triplet) superconductivity.[@MAENO; @ISHIDA] There is a number of other systems, such as the heavy Fermion superconductors UPt$_3$ and UBe$_{13}$ or the organic superconductor (TMTSF)$_2$PF$_6$, where very likely odd-parity pairing is realized.[@TOU; @CHAIKIN] The spin-1 degree of freedom of the spin-triplet Cooper pairs provides a considerably wider space of potential pairing states than in the even-parity (spin singlet) case.[@REV] The symmetry and effective dimensionality of the electronic band structure plays an important role in determining the possible p-wave pairing states. The examples listed above cover many of the possibilities: the organic superconductor is quasi-one-dimensional, the heavy Fermion compounds are three-dimensional, while Sr$_2$RuO$_4$ represents the case of quasi-two-dimensional system.
The large number of possible p-wave states makes their identification for each material a difficult task. A simple weak-coupling BCS type of approach can give a first guess on the most stable state. Because the condensation energy is in this case directly connected with the presence of the energy gap in the quasiparticle spectrum, the state with the least nodes in the gap would be most favorable. In one and three dimensions the most stable state is unique up to spin rotation. Assuming parabolic band structure for the corresponding dimension and using the d-vector notation one finds $$\begin{array}{ll}
{\bf d} ({\bf k}) = \hat{{\bf x}} k_x & \qquad \mbox{for 1d }\\
{\bf d} ({\bf k}) = \hat{{\bf x}} k_x + \hat{{\bf y}} k_y +
\hat{{\bf z}} k_z & \qquad \mbox{for 3 d}
\end{array}$$ where the gap matrix is defined as $ \hat{\Delta}_{{\bf k}} = i
\sigma_{2} {\boldmath \sigma} \cdot {\bf d}({\bf k}) $ and the quasiparticle gap is $ \frac{1}{2} tr [ \hat{\Delta}^{+}_{{\bf k}}
\hat{\Delta}_{{\bf k}} ] = |{\bf d}({\bf k}) |^2 $. Obviously the two states are nodeless on the corresponding Fermi surfaces. Note that the example for three dimensions corresponds to the Balian-Werthammer state or the B-phase of superfluid $^3$He.[@LEGGETT]
We now consider the case of as quasi-two-dimensional system which is characterized by the fact that the Fermi surface is open in one direction, the $ z$-axis. In such a system the weak coupling approach does not lead to a unique state, but we can find six degenerate states with the same nodeless gap. In Sr$_2$RuO$_4$ their degeneracy is lifted by spin-orbit coupling and the corresponding states labeled according to the representation of the tetragonal crystal point group of this compound is given in Table 1.[@RICE-SIG; @NG-SIG] We can distinguish two types of states here: those which have $ {\bf d} $-vector that changes orientation for different points on the Fermi surface belonging to the one-dimensional representation $ A_{1u}, A_{2u}, B_{1u} $ and $ B_{2u}
$ and those which have a fixed $ {\bf d} $-vector orientation but a finite orbital angular momentum, belonging to the two-dimensional $
E_u $ representation.[@RICE-SIG; @NG-SIG] Note that the latter is the chiral state, i.e. it breaks time reversal symmetry and parity. Since all these states are degenerate in the spin rotation symmetric case beyond simple spin rotation transformation, the question arises which among them is stable. For Sr$_2$RuO$_4$ experiments suggest the chiral state with $ {\bf d}
\parallel \hat{{\bf z}} $.[@ISHIDA; @LUKE]
[@cc]{} $\Gamma$ & ${\bf d}({\bf k})$\
A$_{1u}$ & $\hat{\bf x} \hat{k}_{x} + \hat{\bf y} \hat{k}_{y}$\
A$_{2u}$ & $\hat{\bf x} \hat{k}_{y} - \hat{\bf y} \hat{k}_{x}$\
B$_{1u}$ & $\hat{\bf x} \hat{k}_{x} - \hat{\bf y} \hat{k}_{y}$\
B$_{1u}$ & $\hat{\bf x} \hat{k}_{y} + \hat{\bf y} \hat{k}_{x}$\
E$_{u}$ (chiral states) & $\hat{\bf z} (\hat{k}_{x} \pm i \hat{k}_{y})$\
In Sr$_2$RuO$_4$ the loss of spin rotation symmetry by spin-orbit coupling carries the main responsibility in picking the stable state. In this letter, however, we assume that the spin rotation symmetry is preserved in the normal state so that all states listed in Table 1 have the same transition temperature $ T_c $ as a solution of the linearized weak coupling gap equation. In this case the degeneracy must be lifted in a higher order process. A well-known concept introduced by Anderson and Brinkman is the spin fluctuation feedback mechanism.[@ANDERSON; @VOLLHARDT] If paramagnon exchange plays a dominant role in the pairing interaction, the modification of the spin fluctuation spectrum by the superconducting condensation also alters the pairing interaction. It was shown that this mechanism works in favor of the so-called AMB-state or A-phase in $^3$He.[@LEGGETT] This mechanism applied to the 2D situation turns out to stabilize the time reversal breaking state which is indeed the analogue to the A-phase.[@SEOUL; @HERA]
Here we would like to introduce an additional feedback mechanism which does not exist in a neutral Fermi liquid such as $^3$He. It is based on the presence of chirality in the orbital part of pairing state and we will call it, therefore, [*chiral feedback mechanism*]{}. It was shown that in the state $ {\bf d}({\bf k})=
{\bf n} (k_x \pm i k_y) $ a Chern-Simons-like term, $$\epsilon_{ij} (A_{0} \partial_{i} A_{j} + A_{i} \partial_{j} A_{0})$$ appears in the effective low-energy field theory of a static quasi-two-dimensional system ($ i,j=x,y$ ). [@VOLOVIK; @GORYO-ISHI] Consequently, charge fluctuations generate local magnetic field distributions ($z$-axis oriented) and current fluctuations lead to transverse electric field distributions, whose orientation depends on the chirality. This property yields an additional (anomalous) pairing interaction in the superconducting state which has selective power for chirality. This can be seen by the following simple picture. The magnetic field induced by the charge of a quasiparticle acts via the Lorentz force on a passing-by quasiparticle.[@MORINARI] This force is either attractive or repulsive depending on which side the quasiparticle trajectory is located (Fig. 1). In this way the interaction is attractive for one choice of Cooper pair angular momentum (chirality) and repulsive for the other. The attractive interaction appears for the same chirality realized in the pairs of the condensate. Hence this leads to a positive feedback for the chiral state. This effect does not exist for the other states.
We will now discuss the effect by explicitly calculating its contribution to condensation energy immediately below the transition temperature $ T_c $. We represent the $
{\bf d} $-vector by $ {\bf d}({\bf p})= {\bf n}_{\gamma} d_{\gamma
i} k_{i}/k_{\rm F} $ where $ d_{\gamma i} $ is a complex order parameter and $ k_{\rm F} $ inverse of the Fermi wavelength. The band structure is simply parabolic $ \epsilon_{{\bf k}} = \hbar^2 (k_x^2 +
k_y^2 - k_{\rm F}^2) / 2m_e $ without any dispersion along the $ z $-axis leading to a cylindrical (open) Fermi surface. We assume that the system has layered structure as Sr$_2$RuO$_4$ with an interlayer spacing $d$ leading to the density of states, $N(0) = m_e/2 \pi \hbar^2
d $, at the Fermi level. The anomalous pairing interaction appearing in the superconducting phase is connected with the density-current correlation function which for the 2D electrons has the form (in the unit $\hbar=c=1$) $$\begin{aligned}
\pi_{0j}(i \nu_{n}, {\bf q})&=& \int d \tau d^3 x e^{i
\nu_n \tau} e^{ i {\bf q} \cdot {\bf x}} \langle \hat{\rho}({\bf x}, \tau)
\hat{j}_j (0,0) \rangle
\nonumber \\
& = & k_B T \sum_{m}
\int\frac{d^{3}k}{(2 \pi)^{3}}
\frac{-(2 k_j + q_j)}{2m_e}
{\rm Tr}[{\cal{G}}(i \omega_{m} + i
\nu_{n}, {\bf k} + {\bf q})
{\cal{G}}(i \omega_{m},{\bf k})
-
\nonumber \\
&&{\cal{F}}^{\dagger}(i \omega_{m} + i
\nu_{n}, {\bf k} + {\bf q})
{\cal{F}}(i \omega_{m},{\bf k})
] ~~(j=1,2),
\label{den-cur}\end{aligned}$$ and the Green’s functions are $${\cal {G}}(i \omega_{m},{\bf k})=\frac{i \omega_{m} + \epsilon({\bf
k})}{\omega_{m}^{2} + E ({\bf k})^{2}} \quad \mbox{and} \quad
{\cal {F}}(i \omega_{m},{\bf k})=-\frac{i \Delta({\bf
k})}{\omega_{m}^{2} + E({\bf k})^{2}}$$ with $ E({\bf p})=\pm\sqrt{\epsilon({\bf k})^{2} + \Delta ({\bf
k})^2}$ and $\omega_{m}=(2 m + 1) \pi k_B T$ and $\nu_{n}=2 n \pi k_B T$ as the fermionic and bosonic Matsubara frequencies, respectively. We express $ \pi_{0j}(i \nu_{n}, {\bf q}) = i \epsilon_{ij} q_{i}
f(i \nu_{n},{\bf q}) + \nu_n q_j \pi(i \nu_n, {\bf q})$ due to the gauge invariance. $f(i \nu_n, {\bf q})$ comes from the chirality and is written as $$f(i \nu_n, {\bf q}) = - \frac{i}{2!} \epsilon_{ij} \frac{\partial
\pi_{0j}(i \nu_n, {\bf q})}{\partial q_i}.$$ Close to $ T_c $ we can restrict ourselves to the leading contributions in $ d_{\gamma i} $ and obtain, $$f(i \nu_{n},{\bf q}) = \frac{e^{2} k_B T_c}{2 m_e k^2_{F}}
\sum_{m} \int \frac{d^{3} k}{(2 \pi)^{3}}
\frac{- i \epsilon_{ij} d_{\gamma l} d_{\gamma l'}^{*}
k_{l} k_{l'}}
{\{(\omega_{m} + \nu_{n})^{2} + \epsilon({\bf k}+ {\bf q})^{2}\}
\{ \omega_{m}^{2} + \epsilon({\bf k})^{2} \}}
\label{f-dyna}$$ where we sum over repeated indices. Note that this expression is only finite for a chiral p-wave state. We restrict ourselves to static contributions and comment on the dynamical part later.
We now consider the $q$-dependence. We approximate $ \epsilon({\bf k} + {\bf q}) = \epsilon({\bf k}) + {\bf
v}_{\rm F} \cdot {\bf q} $ where $ {\bf v}_{F} $ is the Fermi velocity. In order to evaluate the integral we introduce cylindrical coordinates and perform first the integration over the radial part and $ z$-component of $k$, $$f({\bf q}) = \frac{i e^2 k_B T_c}{4 m_e} N(0) \pi (\epsilon_{ij}
d_{\gamma i} d_{\gamma l}^* ) \sum_m \int \frac{d
\theta}{2 \pi} \frac{\hat{k}_l \hat{k}_j }{|\omega_m|(|\omega_m|^2 +
\frac{1}{4} ({\bf v}_{\rm F}\cdot {\bf q})^2)},$$ where $\hat{k}_i= k_i / |{\bf k}|$. For small $ q $ we may expand $ f ({\bf q}) $ as $$f({\bf q}) \approx \frac{i e^2 N(0) \pi}{4 m_e (\pi k_B T_c)^2}
\epsilon_{ij}
d_{\gamma i} d^*_{\gamma j} \left[ \frac{7}{4} \zeta(3) -
\frac{31}{32} \zeta(5) \xi^2 q_{\perp}^2 + \cdots \right]$$ where $ q_{\perp}^2 = q_x^2 + q_y^2 $, $ \xi = v_{\rm F}/ 2 \pi
k_B T_c $ defines the coherence length and $ \zeta(n) $ is the zeta-function. The behavior of $ f({\bf q}) $ for $ \xi q_{\perp} \gg
1 $ is dominated by the regions of the $ \theta $-integral for which $
| {\bf v}_{\rm F} \cdot {\bf q} | \ll 2 \pi k_B T_c $ and leads to $$f({\bf q}) \approx \frac{i e^2}{4 m_e} \frac{N(0)}{(\pi k_B T_c)^2}
\frac{\epsilon_{ij}
d_{\gamma i} d^*_{\gamma j}}{\xi q_{\perp}}.$$ Matching the limiting behaviors together we can approximate $ f({\bf
q}) $ by the following form, $$f({\bf q}) \approx
\frac{i e^2}{4 m_e} \frac{N(0) \pi}{(\pi k_B T_c)^2}
\frac{\epsilon_{ij}
d_{\gamma i} d^*_{\gamma j}}{\sqrt{1 + \gamma \xi^2 q_{\perp}^2}}, ~
\gamma={\cal {O}}(1),$$ which represents the form factor of parity and time reversal symmetry breaking part in $ \pi_{0j}(0, {\bf q}) $.
The current-charge density interaction introduced via $ \pi_{0j} (0,
{\bf q}) $ gives an additional contribution to the pairing interaction below the superconducting transition. As a feedback effect this appears in the GL free energy in a fourth-order correction expressed by $$\Delta F_{{\rm fb}} = k_B T_c \int \frac{d^{3} q}{(2 \pi)^3} D_{00}({\bf q})
\pi_{0i} ({\bf q}) D_{ij}({\bf q}) \pi_{j0}({\bf q}),
\label{fd-c}$$ following the diagram in Fig.2. Here $D_{00}$, $D_{ij} ~(i,j=1,2)$ is the gauge field propagators which in Coulomb gauge are, $$D_{00}({\bf q})=\frac{1}{q^{2} + l_{\rm TF}^{-2}} \quad \mbox{and}
\quad D_{ij}({\bf q})= \frac{-\delta_{ij}}{q^{2}},$$ These propagators contain all renormalizations, i.e. Thomas-Fermi screening for the scalar potential with the screening length $ l_{\rm
TF} $. Since $ T \approx T_c $ London screening of the superconductor can be neglected. We ignore here also the dynamical part for simplicity, as it would give the same contributions for all competing states.
If we separate the $ q$-integration in $q_z $- and $ q_{\perp} $-part, we obtain, $$\begin{aligned}
\Delta F_{\rm fb}& = & \left\{ \frac{ e^2 N(0) \pi}{4 m_e (\pi k_B T_c)^2}
\epsilon_{ij} d_{\gamma i} d_{\gamma j}^* \right\}^2 k_B T_c l_{\rm
TF}^2
\nonumber \\
&& \quad \times \int \frac{d^2 q_{\perp}}{(2 \pi)^2} \int \frac{d q_z}{2 \pi}
\frac{q_{\perp}^2}{1 + \gamma \xi^2 q_{\perp}^2} \frac{1}{q_{\perp}^2 +
q_z^2} \frac{1}{1 + l_{\rm TF}^2 (q_{\perp}^2 + q_z^2)} \end{aligned}$$ After performing the $ q_z $-integration the remaining $ q_{\perp}
$-integral has a cutoff $ q_{\perp} \sim l_{\rm TF}^{-1} $. This leads to the free energy correction, $$\Delta F_{\rm fb}
%& \approx & \left( \frac{e^2}{2 h c} \right)^2
%\frac{\hbar^2 }{2 \pi m_ek_B T_c} \frac{l_{\rm TF}}{\xi^2 d}
%\frac{N(0)}{(\pi k_B T_c)^2} (\epsilon_{ij} d_{\gamma i} d_{\gamma
%j}^*)^2 \nonumber \\
%& = &
\approx \frac{8 \alpha^2}{\pi} \frac{T_c}{T_{\rm F}} \frac{l_{\rm TF}}{d}
\frac{N(0)}{(\pi k_B T_c)^2} (\epsilon_{ij} d_{\gamma i} d_{\gamma
j}^*)^2$$ where we give an expression formally close to the conventional fourth-order terms in order to give a comparison of its magnitude. Here we recover the constants $\hbar$ and $c$, the factor $ \alpha = e^2 / \hbar c $ is the fine structure constant and the ratio $ T_c / T_{\rm F} $ indicates the strong coupling nature of the correction term, similar to the spin fluctuation feedback mechanism. [@VOLLHARDT]
For the chiral $ p $-wave state $ \epsilon_{ij} d_{\gamma i} d_{\gamma
j}^* = i 2 \chi |\Delta | $ and zero for all other states ($ \chi =
\pm 1 $ denotes the chirality). Thus, the correction to the fourth order term is negative definite and favors the chiral $p$-wave state. The ratio between this correction and the usual fourth order coefficient is $$\frac{\delta \beta_{\rm fb}}{\beta} \sim \frac{\alpha^2}{\pi}
\frac{T_c}{T_{\rm F}} \frac{l_{\rm TF}}{d}$$ which for Sr$_2$RuO$_4$ is of the order $ 10^{-6} $.
It is easy to see that the dynamical contributions, taking into account $ \nu_n \neq 0 $, does not change the result qualitatively. The corresponding coefficient in the free energy, however, increases. We have verified numerically that an increase of one order of magnitude is possible. It is clear that other mechanisms, such as the spin fluctuation feedback or spin-orbit coupling, would dominate over the chiral feedback effect in stabilizing the chiral $ p $-wave state. We would like to emphasize, however, that our analysis shows that for a quasi-two-dimensional $ p $-wave superconductor the chiral feedback effect, based on the anomalous coupling between charge and current, supports the chiral superconducting phase and, thus, works in the same direction that the spin fluctuation feedback mechanism.
The authors are grateful to A. Furusaki, K. Ishikawa and T. Morinari for many helpful discussions. This work has been financially supported by a Grant-in-Aid of the Japanese Ministry of Education, Science, Culture and Sports.
References {#references .unnumbered}
==========
[99]{}
Y. Maeno et al.: Nature [**372**]{} (1994) 532. Physica C [**282-287**]{} (1997) 206. T.M. Rice: Nature [**396**]{} (1998) 627.
K. Ishida et al.: Nature [**396**]{} (1998) 658.
H. Tou et al., Phys. Rev. Lett. [**77**]{}, 1374 (1996).
I. J. Lee, D. S. Chow, W. G. Clark, J. Strouse, M. J. Naughton, P. M. Chaikin and S. E. Brown, cond-mat/0001332.
M. Sigrist and K. Ueda, Rev. Mod. Phys. [**63**]{}, 239 (1991).
A.J. Leggett, Rev. Mod. Phys. [**47**]{}, 331 (1975).
T. M. Rice and M. Sigrist, J. Phys.: Condensed Matter [**7**]{} (1995) L643.
K. K. Ng and M. Sigrist, Europhys. Lett. [**49**]{}, 473 (2000).
G.M. Luke et al.: Nature [**394**]{} (1998) 558.
P. W. Anderson and W. F. Brinkman, Phys. Rev. Lett. [**30**]{}, 1108 (1973); W. F. Brinkman, J. W. Serene and P. W. Anderson, Phys. Rev. A [**10**]{}, 2386 (1974). Y. Kuroda, Prog. Theor. Phys. [**51**]{}, 1269 (1974), [*ibid*]{} [**53**]{}, 349 (1975).
D. Vollhardt and P. Wölfle, [*The Superfluid Phases of Helium 3*]{}, Taylor & Francis, London (1990).
M. Sigrist and T.M. Rice, [*Current Topics in Physics*]{} edited by Y.M Cho et al. World Scientific 1998, p. 273 (1998).
M. Sigrist [*et. al.*]{}, Physica C 317-318, 134 (1999).
G. E. Volovik, Sov. Phys. JETP [**67**]{}, 1804 (1988).
J. Goryo and K. Ishikawa, Phys. Lett. A [**246**]{} 549 (1998); [*ibid*]{} [**260**]{}, 294 (1999).
M. Greiter, X. G. Wen and F. Wilczek, Nucl. Phys. B [**374**]{}, 567 (1992); T. Morinari, Phys. Rev. Lett. [**81**]{}, 3741 (1998).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The black-hole binary, V404 Cygni, went into outburst in June 2015, after 26 years of X-ray quiescence. We observed the outburst with the Neil Gehrels [*Swift*]{} observatory. We present optical/UV observations taken with the [*Swift*]{} Ultra-violet Optical Telescope, and compare them with the X-ray observations obtained with the [*Swift*]{} X-ray Telescope. We find that dust extinction affecting the optical/UV, does not correlate with absorption due to neutral hydrogen that affects the X-ray emission. We suggest there is a small inhomogeneous high density absorber containing a negligible amount of dust, close to the black hole. Overall, temporal variations in the optical/UV appear to trace those in the X-rays. During some epochs we observe an optical time-lag of $(15 - 35)$s. For both the optical/UV and X-rays, the amplitude of the variations correlates with flux, but this correlation is less significant in the optical/UV. The variability in the light curves may be produced by a complex combination of processes. Some of the X-ray variability may be due to the presence of a local, inhomogeneous and dust-free absorber, while variability visible in both the X-ray and optical/UV may instead be driven by the accretion flow: the X-rays are produced in the inner accretion disc, some of which are reprocessed to the optical/UV; and/or the X-ray and optical/UV emission is produced within the jet.'
author:
- |
S. R. Oates$^{1,2}$, S. Motta$^{3}$, A. P. Beardmore$^{4}$, D. M. Russell$^{5}$, P. Gandhi$^{6}$, N. P. M. Kuin$^{7}$, M. De Pasquale$^{7,8}$, D. Altamirano$^{6}$, A. A. Breeveld$^{7}$, A. J. Castro-Tirado$^{2,9}$, C. Knigge$^{6}$, M. J. Page$^{7}$, D. Steeghs$^{1}$\
$^{1}$ Department of Physics, University of Warwick, Coventry, CV4 7AL, UK; [email protected]\
$^{2}$ Instituto de Astrofísica de Andalucía (IAA-CSIC), Glorieta de la Astronomía s/n, E-18008, Granada, Spain\
$^{3}$ University of Oxford, Department of Physics, Astrophysics, Denys Wilkinson Building, Keble Road, OX1 3RH, Oxford, United Kingdom\
$^{4}$ X-ray and Observational Astronomy Group, Department of Physics & Astronomy, University of Leicester, LE1 7RH, UK\
$^{5}$ New York University Abu Dhabi, P.O. Box 129188, Abu Dhabi, UAE\
$^{6}$ School of Physics and Astronomy, University of Southampton, Southampton, SO17 1BJ, UK\
$^{7}$ Mullard Space Science Laboratory, University College London, Holmbury St. Mary, Dorking, Surrey, RH5 6NT, UK\
$^{8}$ Department of Astronomy and Space Sciences, Istanbul University, Beyaz[i]{}t, 34119, Istanbul, Turkey\
$^{9}$ Unidad Asociada Departamento de Ingeniería de Sistemas y Automática, E.T.S. de Ingenieros Industriales, Universidad de Málaga, Spain\
bibliography:
- 'V404cyg\_UVOT.bib'
date: 'Accepted...Received...'
title: '[*Swift*]{} UVOT observations of the 2015 outburst of V404 Cygni'
---
\[firstpage\]
X-rays: binaries - X-rays: bursts – X-rays: individual: V404 Cyg.
Introduction {#intro}
============
Low-mass X-ray binaries (LMXBs) consist of a compact object, neutron star or black hole, and a lower mass companion star. Matter is stripped off the donor star, usually via Roche-lobe overflow, and is accreted onto the compact object via an accretion disk. Black hole LMXBs are mostly transient systems, i.e., they spend most of their time in a quiescent state, where accretion activity is low making them faint in the X-ray band. During this state, the optical luminosity is also low, which often allows measurements to be performed of the donor star that can determine the orbital period and the mass of the compact accretor [see @lop19 and references therein]. Black hole LMXBs sporadically enter an outburst phase that can last weeks to years, during which their luminosity increases by several orders of magnitude and dramatic changes in spectral and fast-time variability properties can be observed.
In June 2015, the LMXB V404 Cygni, also known as GS 2023+338, went into outburst [@GCN:17929] after 26 years of quiescence. This prototypical black hole binary (BHB) was the first system with a conclusive mass function [@cas92]. It hosts a black hole with mass $M_{BH}\sim9\;{\rm M_{solar}}$ and a donor star with mass $\sim0.7\,{\rm M_{solar}}$ of spectral type K3 III [@kha10]. The orbit is $6.473\pm0.001$ days [@cas92] and the inclination is 67$^{+3}_{-1}$ degrees [@kha10]. V404 Cygni was first discovered as an X-ray source by the Ginga satellite in 1989, during an outburst [@IAU4782]. This detection prompted examination of photographic plates that showed V404 Cygni had previously been in outburst in 1938 and 1956 [@ric89; @wag91 see also references therein].
Compared to previous outbursts, the June 2015 outburst is unique in terms of the exceptional coverage both in time and wavelength by a multitude of space and ground-based facilities. The outburst was first detected by the Burst Alert Telescope [BAT; 15-350keV @bar05] on-board the NASA Neil Gehrels [*Swift*]{} satellite, hereafter referred to as [*Swift*]{}, and shortly afterwards by the [*FERMI*]{}/GBM [@GCN:17932]. It was also detected at high energies by MAXI [@ATel:7646] and by [*INTEGRAL*]{} [@ATel:7693]. A large number of optical/IR observatories took photometric and spectroscopic measurements of the June 2015 outburst. [@kim16] alone report 85000 individual data points in the $B$, $V$, $R$ and $I$ bands. Radio and sub-mm observations were also reported [@tet17; @cha17]. As such, V404 Cygni has been extensively studied over a broad range of wavelengths.
Using observations taken in the X-ray band ($0.5-10$keV) with [*Swift*]{}, [@motta17] reported highly variable absorption local to the source. This was interpreted as the effect of a cold outflow ejected by the source itself, that could sustain an unstable, radiation dominated thick disk because of the high accretion rates reached. *INTEGRAL*/IBIS-ISGRI observations of the system during the June 2015 outburst suggest that the X-ray spectrum of V404 Cygni is not intrinsically very different from the spectrum of a standard BHB when the effects of heavy absorption are accounted for (i.e., at energies above 25keV; @san17).
In optical spectra, taken over the entire active phase, [@mun16] detected P-Cyg profiles, suggesting the presence of a sustained outer accretion disc wind in V404 Cygni. A disc wind had previously been observed in BHBs at X-ray wavelengths, e.g., GRO J1655−40 [@mil06], but had never been observed before at optical wavelengths, although a disk wind has been recently noticed in spectra of V4641 Sgr [see @mun18 and references therein]. P-Cyg profiles were also observed by [*Chandra*]{} in X-ray spectra of V404 Cygni [@kin15]. [@kim16] reported the observations of optical oscillations on time-scales of 100s to 2.5 hours from V404 Cygni during the June 2015 outburst, while [@gan16] presented observations of rapid (sub-second) optical flux variability. [@gan16] state that their data initially showed steep power spectra dominated by slow variations on $\sim 100-1000$s, while near the peak of the outburst (June 26), persistent sub-second optical flaring also appeared, close in time to giant radio and X-ray flaring. The fast flares were stronger in the red, and could be explained as optically-thin synchrotron emission from a compact jet. Radio data collected at different frequencies revealed striking variability [see e.g., @tet17; @fen19; @Mil19], characterised by hundreds of flares visible at different frequencies over a two-week period.
With its fast slewing capabilities, [*Swift*]{} was able to provide immediate follow-up using its two narrow field telescopes: X-ray Telescope [XRT; @bur05] and Ultra-violet and Optical Telescope [UVOT; @roming]. Detailed spectral analysis of the [*Swift*]{}/XRT observations during the brightest phase of the outburst of V404 Cygni have already been presented in [@motta17]. In this paper, we focus on the [*Swift*]{}/UVOT observations and their comparison with the X-ray behaviour observed simultaneously by the [*Swift*]{}/XRT. V404 Cygni also had a secondary outburst in December 2015. Since this later outburst has limited sampling we focus on the June 2015 outburst, but for completeness in the Appendix we give a short description of the data from the second outburst and present the photometry. This paper is organized as follows. In § \[reduction\] we discuss the data reduction and analysis. The main results are presented in § \[results\]. Discussion and conclusions follow in § \[discussion\] and \[conclusions\], respectively. All uncertainties throughout this paper are quoted at 1$\sigma$, unless otherwise stated. Throughout this paper, we use a distance to V404 Cygni of $2.39\pm0.14$kpc as determined through parallax measurements by [@mil09].
Observations & Data Reduction {#reduction}
=============================
[*Swift*]{}/BAT first triggered on V404 Cygni at 18:31:38 UT on the 15th June 2015 [MJD =57188.771976 UTC; @GCN:17929] and this was followed by a series of secondary triggers at 2.26, 2.65, 2.78, 4.79 and 10.76 days after the first trigger [equivalent to $200$ks, $229$ks, $240$ks, $414$ks and $930$ks after the first trigger; @GCN:17944; @GCN:17945; @GCN:17946; @GCN:17949; @GCN:17963]. To prevent further triggers, V404 Cygni was entered into the on-board catalogue on the 22nd June 2015, which set a minimum trigger threshold for this source. However, the trigger threshold was exceeded at 10.76 days, resulting in another automatic slew to the source and an automatic increase in the BAT trigger threshold. Additional follow-up of V404 Cygni by [*Swift*]{} was therefore performed in response to a number of target of opportunity (ToO) requests (PIs: Altamirano, Beardmore, Cadolle Bel, Gandhi, Heinz, Kennea, Kuulkers, Middleton, Motta, Plotkin, Rodriguez, Sivakoff and Vasilopoulos).
For the following analysis we correct all times to that at the barycentre. This results in a barycentric corrected, BAT trigger time $\rm T_0$ = 57188.773986 MJD (TDB), which we consider as the time the outburst began.
As [*Swift*]{} slews to a number of targets in a given 95 min orbit to maximise its observing efficiency, a UVOT and XRT observation of a specific source on a given day is broken up into one or more ‘snapshots’ of continuous exposure. Therefore the observed light curves are intermittent on long time scales, e.g several minutes to hours, but within this have periods of continuous observations e.g., $\sim 0.2 - 1.8\,{\rm ks}$ at few second resolution depending on the instrument, signal to noise and observing mode. Therefore, [*Swift*]{} observations can provide information on the overall behaviour of the outburst and some of the fast variability, but will have gaps in the observations during the main outburst when the source was highly variable.
[*Swift*]{} UV/Optical Telescope {#UVOTreduction}
--------------------------------
[*Swift*]{}/UVOT began settled observations of V404 Cygni 184s after the initial BAT trigger [@GCN:17929]. [*Swift*]{}/UVOT monitored the outburst for the next 70 days ($\sim 6000$ks), until it returned to quiescence. V404 Cygni was initially detected in 5 of the 7 UVOT filters ([*white*]{}, [*v*]{}, [*b*]{}, [*u*]{}, [*uvw1*]{}) at $>2\sigma$, but as the source brightened it also became detectable in the [*uvm2*]{} and [*uvw2*]{} filters.
[*Swift*]{}/UVOT also observed the field of V404 Cygni a number of times prior to 2015: observations were taken in the three optical and three UV filters on 26th April 2009; several observations were performed in the [*uvw1*]{} filter between the 3rd July 2012 and 16th September 2012; and on the 14th October 2013 and 2nd December 2013 observations were taken in the [*u*]{} filter.
UVOT observations of V404 Cygni were taken in image and event modes[^1]. Before extracting count rates from the event lists, the astrometry was refined following the methodology in [@oates09]. In order to be compatible with the UVOT calibration [@poole; @bre11], we used an aperture of 5$\arcsec$ to obtain the source counts from both the event and image mode data. However, since it is recommended to use a smaller aperture when the count rate is low [@poole], we used a 3$\arcsec$ radius below a threshold of 0.5 counts per second. In this case, the source count rates were corrected to a 5$\arcsec$ radius aperture using a table of aperture correction factors contained within the calibration. The background counts were obtained from a number of circular regions positioned in blank regions of the sky situated near to the source position. The count rates were obtained from the event and image lists using the [*Swift*]{} tools <span style="font-variant:small-caps;">uvotevtlc</span> and <span style="font-variant:small-caps;">uvotsource</span>, respectively. The UVOT data were processed using the [*Swift*]{} software, version 4.4 (from HEASOFT release 6.19), using the UVOT calibration files (20160321) released on 2016-Mar-21.
It has recently come to light that the UVOT detector is less sensitive in a few small patches[^2] for which a correction has not yet been determined. Therefore we checked to see if the source falls on any of these spots in any of our images and excluded 3 individual [*uvw1*]{} exposures for this reason. We also excluded [*white*]{} filter data between 5.8 and 9.2 days after the initial BAT trigger due to saturation.
A nearby star is reported to be 1.4" away from the position of V404 Cygni [@uda91]. This star may not be physically associated with the system, though [@mai17] recently proposed it may be an F2 star forming a triple system with V404 Cygni. The brightness of this star is $20.59\pm0.05$ and $18.90\pm0.02$ magnitudes in the B and V bands (Vega system), respectively, and was observed resolved with V404 Cygni in quiescence, when the flux of V404 Cygni was $20.63\pm0.05$ and $18.42\pm0.02$ magnitudes in the B and V bands (Vega system), respectively. In UVOT observations, this nearby star and V404 Cygni are not resolved, the emission from both combining to form a single point source. The photometry of V404 Cygni is therefore contaminated by flux from this nearby star. Since V404 Cygni is many times brighter in outburst than the quiescent phase, the nearby star will not strongly affect the photometry while in outburst. However, the contribution from this nearby star to the flux measurements will increase as V404 Cygni decreases in brightness. In quiescence, it is expected that this nearby star and V404 Cygni each will contribute about 50 per cent of the flux within the UVOT aperture (in the V and B filters). We therefore determine the nearby star’s contribution to the photometry as V404 Cygni returns to quiescence, taking the lowest pre-2015 UVOT magnitudes of V404 Cygni as an upper limit to the flux from this contaminating star, namely an AB magnitude of $u=22.66\pm0.51$. We calculate that the nearby star contributes approximately $\leq4\%$, $\leq10\%$, $\leq25\%$ of the measured flux at ${\rm T_0}+$ 11.6, 18 and 23 days (${\rm T}_0+$ $1000$ks, $1600$ks and $2000$ks), to the $u$-band. In addition, the nearby star will be fainter in the UV bands than in the $u$-band since it has an F-type spectrum.
The count rates from both the image and event mode data were converted to AB magnitudes using the UVOT zero points and to flux using the AB flux conversion factors [@bre11]. The UVOT observations are displayed in Fig. \[v404lc\_multipanel\]. To correct for interstellar extinction we use an $A_v=4.0\pm0.4$ [@cas93; @hyn09] assuming the Milky Way extinction law given in [@pei92]. We note that more recent measurements of extinction along the line of sight has been performed by [@ito17], with $3.0 < {\rm A_V} < 3.6$ and [@rah17] who measure ${\rm A_V} = 3.82\pm0.36$. These measurements are slightly lower than that measured by [@cas93] and [@hyn09]. If we had chosen to use a lower $\rm{A_V}$ value, it would have systematically decreased our measured magnitudes and luminosities, but it would not have any strong influence on our conclusions. Finally, the extinction corrected fluxes were converted to luminosity.
[*Swift*]{} X-ray Telescope {#XRTreduction}
---------------------------
Following the initial [*Swift*]{}/BAT trigger [@GCN:17929], the XRT observed V404 Cygni for a total of 171.4 ks over 70 days. 114.4 ks of data were collected in Photon Counting (PC) mode ($\rm{T_0}+\,<$23ks and $\rm{T_0}+\,>127$ks) and 57.0 ks in Windowed Timing (WT) mode, with the latter used when the source was bright ($\rm{T_0}+\,$19ks until $\rm{T_0}+\,140$ks). In the analysis that follows the XRT data were processed using the [*Swift*]{} software, version 4.4 (from HEASOFT release 6.16), using the latest XRT gain calibration files for this epoch (released on 2015-Jul-21).
The analysis of the central point source is made more challenging by the presence of a sometimes significant dust scattered X-ray halo [@vas16; @hei16; @bea16]. The effect of the halo is most notable when a strong X-ray flare is followed by a drop in the source count rate to below $\sim 100\ {\rm count\, s^{-1}}$ [see appendix A of @motta17].
A $0.5-10.0$keV count rate light curve was created from the XRT observations. The PC mode data were extracted at per snapshot binning, using grade $0-4$ events. The PC data taken after day 14 were background subtracted using events contained within an annular region located inside the innermost dust scattering ring. WT light curves were extracted at 1s binning from grade 0 events (as this grade selection lessens the effects of pile-up, which becomes important above $\sim 100-150\,{\rm count\,s^{-1}}$ in this mode). When required, the halo contribution was estimated from the WT data using events obtained within a $1-2$ arcminute-wide strip chosen to lie as close to the central point source as possible when the halo dominated. Appropriate corrections for (sometimes severe) pile-up were also made to the data when necessary. This was achieved using suitably sized annular extraction regions to excise the piled-up core of the point spread profile, then correcting for the expected loss of flux. The extraction regions thus ranged from circles of radius 23.6 arcseconds (when the central source was faint compared to the halo) to annular regions of size $47.1-94.3$ arcseconds (when the central source was extremely piled-up). Use of the $0.5-10.0$keV count rate light curve will be discussed in §\[highres\].
To convert the X-ray count rate to luminosity we need to determine the count rate to flux conversion factors. This requires analysis of a large number of X-ray spectra, such as that performed by [@motta17]. [@motta17] computed X-ray fluxes in the energy range, $2-10$keV (note the narrower energy range compared to our $0.5-10.0$keV count rate light curve) using spectra extracted with 16s duration or long enough to contain about 1600 counts, resulting in 1054 data points with variable exposure. To compute these fluxes, the SEDs were fitted with an absorbed power law combined with a narrow Gaussian line around the 6.4keV Fe-K energy and interstellar absorption, when statistically required. A neutral absorber partially covering the source was also included when the fit showed improvement; see [@motta17] for further details. We use the $2-10$keV flux light curve from [@motta17] from this point onwards to produce the X-ray luminosity light curve, however, as [@motta17] only focused on the active part of the outburst (18th June - 26th June 2015), we also add data observed by [*Swift*]{}/XRT prior to and after this time, converting to flux following the same method. The XRT flux light curve, in the $2-10$keV energy range, is displayed in Fig. \[v404lc\_multipanel\]. We convert these 1080 flux points to luminosity. These luminosity measurements have been corrected for Galactic absorption, but they have not been corrected for intrinsic absorption as there may be some degeneracy between the absorption and reflection parameters in the $2-10$keV energy band [see @motta17], which if corrected for may give intrinsic luminosity values that may be significantly lower than the true values.
High time resolution Analysis {#highres}
-----------------------------
### Light curves
UVOT observations taken in event mode can be subdivided into smaller time bins in order to get a higher temporal resolution. Many of the UVOT $u$-band observations were observed in event mode. We extracted photometry from the UVOT $u$-band event mode data with a resolution of 5s. Since the time resolution of the X-ray light curves can also be resolved to a better time resolution than individual exposures and is observed at the same time, we can directly compare the behaviour in the X-rays with that in the $u$-band. We extract the $0.5-10$keV X-ray count rate in 1s intervals and then rebin the X-ray light curve so that it has the same time bins as the $u$-band data points. There are $\sim 4$k individual data points during the brightest phase of the outburst, i.e., ${\rm T_0}+\,240$ks $-$ 970ks (${\rm T_0}+\,2.8-11.2$ days). We display a selection of light curve segments from the resulting $u$-band and $0.5-10$keV X-ray count rate light curves in Fig. \[eventlcXU\]. The full sample comparing the $u$-band with the X-ray $0.5-10$keV during the brightest phase of the outburst can be found in the on-line material, Fig. S.1, and also displayed in Fig. S.2. In Fig. S.1 we also provide the results from the X-ray spectral analysis by [@motta17] and as discussed in §\[XRTreduction\], displaying the column density $N_H$, covering fraction and spectral index, $\Gamma$.
The X-ray count rate light curves have better temporal resolution compared with the X-ray luminosity light curves. In the following we therefore use the X-ray and $u$-band count rate light curves for the high time resolution temporal comparison with the $u$-band data. However, a caveat should be placed on the interpretation of the count rates, for instance, an increase in X-ray count rate does not necessarily mean an increase in X-ray flux at all X-ray energies. However, the X-ray count rate is a consistent photometric measure of the flux of X-rays through a fixed passband, and using this rather than luminosity, means that we do not introduce any systematics related to the different, longer time bin over which the spectra have been accumulated.
### Timing
To determine if there is any time lag between the optical/UV and X-ray light curves, we calculate the cross-correlation for each continuous section of the 5s binned light curves. We determine the cross-correlation using the function [*correlate*]{} in the python numpy package. An example of the analysis performed on one section of the $u$-band and X-ray light curve is given in Fig. \[cross\_corr\], while the results for the entire outburst can be found in the on-line material Fig. S.2. To determine if a given lag coefficient could be achieved due to chance, we perform a Monte Carlo simulation. For each section of light curve we simulate an X-ray light curve with the same power spectrum as the actual data. We perform a Fourier transform of the X-ray light curve and create a new X-ray light curve by randomising the Fourier phases, but maintaining the Fourier amplitude[^3]. With the simulated X-ray light curve we perform a cross correlation between those points and the corresponding, observed, $u$-band data. We repeated this $10^5$ times and store the resulting correlation coefficients for each lag time, for each iteration. Once all iterations are complete, for each lag time we extract values at $0.5, 2.5, 97.5$ and $99.5$ per cent from the resulting cross correlation distribution.
To explore how the amplitude of the variability in the $u$-band and X-ray light curves changes with time, we measure the excess root mean square (rms) following the method outlined in [@vau03] and taking into account the measurement errors on the data. This method essentially takes a segment of uniformly binned data and determines the variance as the integral of the power spectral density (PSD) between two frequencies, subtracting the Poissonian noise from both the power spectra. We make use of the IDL routine [<span style="font-variant:small-caps;"></span>]{} and [<span style="font-variant:small-caps;"></span>]{} written by S. Vaughan[^4]. We adapt the output so that the measurement errors are subtracted from the resulting variance and taken into account when determining the error on the variance. We use the full 1s X-ray ($0.5-10$keV) light curve (e.g., all observations, not only those overlapping with the $u$-band), rebinned to have a bin width of 5s, and the 5s $u$-band light curves. We calculate the PSD, and hence the excess rms, every 128 consecutive data points (i.e., 640s). The resulting light curve displaying the excess rms can be observed in the bottom panel of Fig. \[v404lc\_multipanel\]; we only display values where an excess could be determined above the error on the data, i.e., positive values of the excess.
![The UVOT $u$-band event mode observations (blue triangles) of V404 Cygni compared to the X-ray ($0.5-10$keV; red circles). In these panels, the X-ray data has been binned to have the same bin width as the $u$-band event mode data, which is binned uniformly with a bin width of 5s. These three panels were selected as examples of the type of temporal behaviour observed within V404 Cygni and how the behaviour compares in the two bands. The full sample is provided in the online material S.1.[]{data-label="eventlcXU"}](./Images/CountRate_574093_575013.pdf "fig:") ![The UVOT $u$-band event mode observations (blue triangles) of V404 Cygni compared to the X-ray ($0.5-10$keV; red circles). In these panels, the X-ray data has been binned to have the same bin width as the $u$-band event mode data, which is binned uniformly with a bin width of 5s. These three panels were selected as examples of the type of temporal behaviour observed within V404 Cygni and how the behaviour compares in the two bands. The full sample is provided in the online material S.1.[]{data-label="eventlcXU"}](./Images/CountRate_672007_672977.pdf "fig:") ![The UVOT $u$-band event mode observations (blue triangles) of V404 Cygni compared to the X-ray ($0.5-10$keV; red circles). In these panels, the X-ray data has been binned to have the same bin width as the $u$-band event mode data, which is binned uniformly with a bin width of 5s. These three panels were selected as examples of the type of temporal behaviour observed within V404 Cygni and how the behaviour compares in the two bands. The full sample is provided in the online material S.1.[]{data-label="eventlcXU"}](./Images/CountRate_814381_815746.pdf "fig:")
Results
=======
The UVOT optical/UV image mode light curves for the June 2015 outburst of V404 Cygni are shown in Fig. \[v404lc\_multipanel\]. The temporal behaviour in the different optical/UV filters is similar. Over the course of observations V404 Cygni fluctuated in brightness by over 8 magnitudes. By the time UVOT began observing, V404 Cygni was already at a flux level $\sim 4$ magnitudes brighter than the level observed in quiescence during the pre-2015 observations. The optical/UV light curves rose to the highest level between $\rm T_0+\,3$ $-$ ${\rm T_0}+11$ days ($57191.5 - 57199.5$ MJD). At ${\rm T_0}+11$ days, the flux dropped rapidly, decreasing by 5 magnitudes within a day. Over the subsequent days the brightness continued to decrease, but at a slower rate, until it reached the level measured by [*Swift*]{}/UVOT prior to 2015.
In Fig. \[v404lc\_multipanel\] we also display observations taken by [*Swift*]{}/XRT in the soft $2-10$keV X-ray band as well as those in the hard 25-60keV X-ray band performed by [*INTEGRAL*]{} IBIS/ISGRI [@jou17]. The IBIS/ISGRI light curve has been adaptively rebinned to achieve S/N $>8$ in each bin[^5] (see, e.g., @ATel:7758 [@ATel:8512; @jou17; @san17], for details on the [*INTEGRAL*]{} data). The IBIS/ISGRI data has better sampling than the [*Swift*]{} optical/UV and X-ray observations, although from the three panels, it is apparent that the overall behaviour is the same in optical/UV through to hard X-rays. In particular, the drop back down to quiescence is observed in all of the top three panels of Fig. \[v404lc\_multipanel\] [see, e.g., @san17]. The soft/hard X-rays appear to vary strongly, flaring on time scales of minutes to hours. Due to the UVOT filter rotation, there are only a couple of images per optical/UV filter taken during these X-ray flares, but the optical/UV also seems to flare at similar times.
Optical/UV Colour versus X-ray Column Density & Photon Index
------------------------------------------------------------
To determine if there are changes in the optical/UV spectral shape during the outburst, we display the colour light curves for the UVOT filters in Fig \[v404lc\_multipanel\]. Since the UVOT observations in the different filters are not simultaneous, we take the difference in extinction-corrected AB magnitudes of the data points closest in time to each other. The typical difference in the start and stop time of each bin[^6] is 306s, but ranges from 44s to 1210s. At several points during observations the light curves show rapid changes in colour, for instance at $\rm{T}_0$+6.72 days ($\rm{T}_0$+581ks). These rapid changes are likely due to short timescale temporal evolution between the two filter visits since they are not strictly simultaneous. However, between $\rm{T}_0$+2 days and $\rm{T}_0$+10 days, the colour light curves are generally flat indicating no strong spectral changes throughout the main outburst.
In Fig \[colour-NH\] we display UVOT colour versus the column density of neutral hydrogen, $N_H$, determined from the X-ray spectra as detailed in §\[XRTreduction\]. The figure consists of four panels, each displaying the difference in magnitude between a different pair of UV filters versus $N_H$. Since the colour points span a larger time interval than the interval typically used to extract the X-ray spectra, we take the average $N_H$ value of the spectra that fall within the start and stop time of the UVOT colour data point. In the determination of the average $N_H$ value, we do not include spectra for which the $N_H$ could not be constrained (247/1080), which may weight the average $N_H$ value to higher values. In Fig \[colour-NH\], we also show the relationship between extinction between the two filters, $E(\lambda_1-\lambda_2)$, and the column density of neutral hydrogen, $N_H$, as expected for the Milky Way. Since extinction has a stronger effect in the UV, the colour difference between an optical and a UV filter is expected to be larger than the $E(B-V)$ for a given value of $N_H$. While the $N_H$ value changes in time, clustering into high and low $N_H$ groups, the UVOT colour does not become redder with increasing $N_H$.
In Fig. \[colour-gamma\], we display the same UVOT colours against X-ray photon index, $\Gamma$, in order to determine if a change in the X-ray spectrum corresponds to a change in the optical spectrum. Since the colour points span a larger time interval than the interval typically used to extract the X-ray spectra, we take the average $\Gamma$ value of the spectra that fall within the start and stop time of the UVOT colour data point. In the determination of the average $\Gamma$ value, we do not include spectra for which the $\Gamma$ could not be constrained (52/1080), which may weight the average $\Gamma$ value to softer values. [@motta17] note that spectra with $\Gamma <1$ are not consistent with the spectral slope observed by INTEGRAL, $\Gamma>1.4$ [@san17]. [@motta17] find the spectra become harder when a reflection component is included.
In the top two panels, displaying $uvw1-u$ and $uvm2-u$, the UVOT colours are relatively stable at $\sim 0.7$ while the X-ray $\Gamma$ varies from $\sim 0.5-2.5$ over the brightest phase of the outburst. The bottom two panels show a much wider range in colour, but the uncertainty is typically larger than the upper panels. In any case, in all 4 panels there is no apparent correlation between optical/UV colour and $\Gamma$.
![Top panel: time sliced observation of the $u$-band (blue triangles) and X-ray (0.5-10keV; red circles) observations of V404 Cygni at ${\rm T_0}$+969ks with 5s binning. Middle panel: the corresponding cross-correlation determined for a range of optical lag times. The blue and red dotted lines indicate the $0.5, 2.5, 97.5$ and $99.5$ per cent percentile boundaries. Bottom Panel: The power spectral density (PSD) for the ${\rm T_0}$+969ks segment displaying the PSD for both the X-ray (red line) and $u$-band (blue line) light curves. The dotted lines indicating the expected constant Poissonian white noise contribution. Above the top right corner of the middle panel we give the optical lag time and coefficient at the peak of the cross correlation function and the probability of it occurring by chance considering the multiple cross correlations performed for that segment. The optical and X-ray light curves in the top panel appear to show similar behaviour. This is confirmed by the cross correlation, which indicates the optical and X-ray are correlated with a coefficient of $\rho = 0.43$, but that the optical lags the X-ray emission by 35s. The power spectral density shows that the variability in both bands is due to a combination of red noise, producing the power-law behaviour and Poisson noise, dominating the signal above $>10^{-2}$Hz. Similar figures for the full sample are provided in the online material S.2.[]{data-label="cross_corr"}](./Images/CountRate_969535_970950.pdf "fig:") ![Top panel: time sliced observation of the $u$-band (blue triangles) and X-ray (0.5-10keV; red circles) observations of V404 Cygni at ${\rm T_0}$+969ks with 5s binning. Middle panel: the corresponding cross-correlation determined for a range of optical lag times. The blue and red dotted lines indicate the $0.5, 2.5, 97.5$ and $99.5$ per cent percentile boundaries. Bottom Panel: The power spectral density (PSD) for the ${\rm T_0}$+969ks segment displaying the PSD for both the X-ray (red line) and $u$-band (blue line) light curves. The dotted lines indicating the expected constant Poissonian white noise contribution. Above the top right corner of the middle panel we give the optical lag time and coefficient at the peak of the cross correlation function and the probability of it occurring by chance considering the multiple cross correlations performed for that segment. The optical and X-ray light curves in the top panel appear to show similar behaviour. This is confirmed by the cross correlation, which indicates the optical and X-ray are correlated with a coefficient of $\rho = 0.43$, but that the optical lags the X-ray emission by 35s. The power spectral density shows that the variability in both bands is due to a combination of red noise, producing the power-law behaviour and Poisson noise, dominating the signal above $>10^{-2}$Hz. Similar figures for the full sample are provided in the online material S.2.[]{data-label="cross_corr"}](./Images/CrossCorrelation_969535_970950.pdf "fig:") ![Top panel: time sliced observation of the $u$-band (blue triangles) and X-ray (0.5-10keV; red circles) observations of V404 Cygni at ${\rm T_0}$+969ks with 5s binning. Middle panel: the corresponding cross-correlation determined for a range of optical lag times. The blue and red dotted lines indicate the $0.5, 2.5, 97.5$ and $99.5$ per cent percentile boundaries. Bottom Panel: The power spectral density (PSD) for the ${\rm T_0}$+969ks segment displaying the PSD for both the X-ray (red line) and $u$-band (blue line) light curves. The dotted lines indicating the expected constant Poissonian white noise contribution. Above the top right corner of the middle panel we give the optical lag time and coefficient at the peak of the cross correlation function and the probability of it occurring by chance considering the multiple cross correlations performed for that segment. The optical and X-ray light curves in the top panel appear to show similar behaviour. This is confirmed by the cross correlation, which indicates the optical and X-ray are correlated with a coefficient of $\rho = 0.43$, but that the optical lags the X-ray emission by 35s. The power spectral density shows that the variability in both bands is due to a combination of red noise, producing the power-law behaviour and Poisson noise, dominating the signal above $>10^{-2}$Hz. Similar figures for the full sample are provided in the online material S.2.[]{data-label="cross_corr"}](./Images/PSD_969535_970950.pdf "fig:")
High Time Resolution {#shortvar}
--------------------
In Fig. \[eventlcXU\], we show examples of the $u$-band in comparison to the $0.5-10$keV X-ray light curves. The full sample can be observed in the on-line material, Fig. S.1. We also show the data again in Fig S.2, but with each panel having the same y-axis and the same time range, allowing an easy comparison of how the behaviour changes from segment to segment. From these figures it is clear that the X-ray base flux changes strongly between segments, while the optical remains fairly constant. The segments cover the period where the outburst is in the brightest phase, i.e., $\rm T_0+\,240$ks $-$ 970ks ($\rm T_0+\,2.8-11.2$ days). Within each segment the X-rays in general are more variable than the optical/UV: the count rate of the X-ray light curves varies by a factor of up to $\sim$100, whereas the optical count rate changes less drastically by a factor of a few. There are, however, some segments where the optical/UV variability is larger than the X-ray, but the factor is $\lesssim 2$.
On the whole, the $u$-band behaves similarly to the X-ray light curve, but the changes are less prominent and tend to be smoother. Some light curve segments are flat with small variations of a few per cent for both the X-ray and $u$-band light curves, e.g., ${\rm T_0}+579$ks and ${\rm T_0}+734$ks. Other segments vary more strongly either in one or both bands, with variations on 10s and 100s time scales, which often appear oscillatory, e.g., segments ${\rm T_0}+$729ks, 780ks, 792ks and 798ks. In addition, there are some segments, e.g., ${\rm T_0}+$672ks, 820ks and 877ks, where the X-ray and $u$-band show larger scale variations over the course of a few hundred seconds, but the behaviour is more pronounced in one of the bands.
We computed the cross correlation for each of the continuous sections of the $u$-band and X-ray light curves. An example is displayed in Fig. \[cross\_corr\] and the entire sample is given in the on-line material Fig. S.2. In these panels, we also show the coefficient values found at the $0.5, 2.5, 97.5$ and $99.5$ per cent level from the simulated coefficient distributions. We find 11 segments to have coefficients which lie outside of either the 0.5 or 99.5 per cent percentile boundaries. Of the 23 segments we tested for a correlation, we would expect by chance that $<1$ of these segments will have a peak coefficient outside of this percentile boundary. However, this statistic should also take in to account that within each segment we are performing multiple cross correlations, one at each lag. Therefore for each segment, we take the chance probability associated with the peak of the cross correlation to be the fraction of simulations that have absolute values that exceed the absolute value of this coefficient at [*any*]{} lag time. Of the 23 segments, we find 2 segments (segments ${\rm T_0}+240$ks and $+969$ks given in Fig. S.2) have $<1\%$ chance of the peak cross correlation coefficient being obtained by chance. These segments are a subset of the 11 which have cross correlation coefficients which lie outside of either the 0.5 or 99.5 per cent percentile boundaries. These 2 segments have coefficients that are positive with values of $\rho =0.43$ and 0.64, suggesting the two segments are correlated with slightly different degrees of strength. The optical time lag of these peak coefficient values are 35 and 15s, respectively. In Fig. \[cross\_corr\] and Fig. S.2, in the top right above each optical lag panel we give the optical lag time and coefficient at the peak of the cross correlation function and the probability of it occurring by chance considering the multiple cross correlations performed for that segment. The probability is given only when the value is less than 10 per cent.
![UVOT colour versus column density of neutral hydrogen, $N_H$. We give 4 panels, comparing four different colour terms for the UVOT with $N_H$. The colour scale represents the time each point was determined. The black dotted line represents the expected relationship between $N_H$ and $E(\lambda_1-\lambda_2)$ for the Milky Way, where $\lambda_1$ and $\lambda_2$ represent the 2 different filters. Since extinction has a strong effect in the UV, the expected colour difference in the UV for a given $N_H$ will depend on the exact two filters. However while we observe $N_H$ to change significantly, and cluster into high and low $N_H$ groups, the UVOT colour does not become redder with increasing $N_H$. This suggests strong changes in absorption do not correspond to changes in optical/UV extinction.[]{data-label="colour-NH"}](./Images/OX_nH.pdf)
### Power Spectra
For segments of the X-ray and $u$-band light curves of duration $>500$s (100 individual data points) we compute the power spectral densities (PSD) from the modulus-square of the discrete Fourier transforms, with the constant of the PSD determined using the usual rms-squared standardisation [e.g., @gan10; @utt14]. One example can be observed in the last panel of Fig. \[cross\_corr\], while the full sample can be observed in the on-line material, Fig. S.2. In the panels displaying the power spectra, we also provide lines indicating the expected constant Poissonian white noise contribution for both the X-ray and $u$-band, which were computed following the procedure in [@vau03] and [@gan10]. The power spectra appear to be featureless except for a constant and a power-law component. For many of the power spectra, the Poisson noise dominates the signal above $>10^{-2}$Hz. Red noise, identified by the power-law slope of $\sim-2$, is the other strong component of the power spectra. However the strength of this component varies and is not apparent in all light curve segments. Steep red noise power spectra were also reported during the outburst of V404 Cygni by [@ATel:7677], [@ATel:7688] and [@gan16].
### RMS variability
To quantify the variability, we computed the excess rms for the high time resolution $u$-band and $0.5-10$keV X-ray light curves (see § \[timing\]). The results are shown in the bottom panel of Fig. \[v404lc\_multipanel\]. The excess rms for both the optical/UV and X-ray light curves changes with brightness, increasing as V404 Cygni becomes brighter, with the rms being a factor of 100 larger for the X-ray light curves in comparison to the optical at any given time. In the top panel of Fig. \[rms\_flux\], we display the square root of the excess rms (rms amplitude) against the mean count rate[^7]. Both for the $u$-band and X-ray $0.5-10$keV data, the excess rms increases with increasing count rate. The dependence on the flux is removed by normalizing the excess rms by the count rate. This results in the fractional variability, which is shown in the bottom panel of Fig. \[rms\_flux\].
To quantify the correlation between excess rms and count rate, a function should be fitted to the data, however, we do not expect to obtain a good $\chi^2$ due to the intrinsic scatter of the data. Instead we use a Spearman rank correlation to determine if the logarithms of the excess rms and count rate are correlated. For the X-ray observations we find evidence for a strong correlation with a correlation coefficient of 0.94 and reject the null hypothesis with $p\ll 10^{-5}$. However, for the optical/UV the correlation coefficient indicates a weak correlation, much weaker than for the X-ray. The coefficient is 0.7 and we can not reject the null hypothesis ($p = 0.03$). Furthermore, there are some X-ray data points that imply over 100% fractional variability. These coincide with segments where the variability spans the entire segment.
### X-ray and $u$-band count rate comparison {#OXcomparison}
As discussed in §\[shortvar\], the variability in the $u$-band is less pronounced compared with the X-ray. We can explore this behaviour further by displaying all the high time resolution $u$-band and X-ray data together. In the top panel of Fig. \[OXcorr\], we show the $u$-band count rate versus the X-ray count rate between ${\rm T_0}$ and ${\rm T_0}+70$ days (0 to $\sim 6000$ks). Neither the $u$-band nor the X-ray count rates have been corrected for extinction or absorption and caution should be used at low optical count rate because of the contaminating star (see § \[UVOTreduction\]). At $0.1$ count per second, this nearby star may contribute up to $50\%$ of the total counts, but this fraction decreases as V404 Cygni brightens. The $u$-band and X-ray observations in general increase together, as also shown in Fig. \[v404lc\_multipanel\], but as V404 Cygni brightens the distribution becomes broader in the X-rays compared to the optical/UV.
We can look at time-resolved versions of Fig. \[OXcorr\], comparing the $u$-band and X-ray count rates of individual light curve segments. These are given in the bottom left sub-panels of the on-line material Fig. S.2. Their behaviour can be divided into two groups: 1) randomly distributed data points, suggesting no strong correlation; these tend to be sections of the light curve where there is no strong variability in either the X-ray or $u$-band; 2) correlated count rates, increasing and decreasing together, although the scatter tends to be large. For two segments the behaviour differs from these two categories: in the ${\rm T_0}$+877ks segment the X-ray and $u$-band count rates change independently producing vertical and horizontal lines, while in the ${\rm T_0}$+820ks segment similar behaviour is observed, but a hysteresis-like pattern is produced. Large variations in X-ray flux, but no pronounced optical response was also reported by [@hyn19].
We can examine the relationship between optical/UV and X-rays in luminosity space. The luminosity ratio is shown in the bottom panel of Fig. \[OXcorr\], with the colour of each point representing the time of the observation. In this panel, due to the need to create X-ray spectra with enough signal so that spectra could be fitted and a count rate to flux conversion factor could be determined (see §\[XRTreduction\]), the data points are coarser in time in comparison to the top panel. In Figure \[OXcorr\], the data points are distributed either side of the line of equality, suggesting that there are periods where the $u$-band is more luminous than the X-ray and vice versa. The colour of the points represent the time since the trigger and the colours roughly divide the figure in to horizontal bands. The data at the earliest and latest epochs form roughly horizontal bands at lower optical luminosity. There is weak indication that at low X-ray luminosity ($<$ a few $10^{36}$erg s$^{-1}$) the optical and X-ray may increase at a different rate than at higher X-ray count rate, perhaps resembling the track given in the ${\rm T_0}$+820ks or ${\rm T_0}$+877ks segments in the on-line material Fig. S.2.
![UVOT colour versus the X-ray photon index, $\Gamma$. We give 4 panels, comparing four different colour terms for the UVOT with $\Gamma$. The colour scale represents the average X-ray flux during the time each colour point was determined. The bottom two panels show much wider range in colour, but the uncertainty is typically larger than the upper panels. However, for all 4 panels there is no apparent correlation between optical/UV colour and $\Gamma$, suggesting that changes in the X-ray spectrum do not correspond to a changes in the optical spectrum.[]{data-label="colour-gamma"}](./Images/OX_gamma.pdf)
### X-ray and Optical/UV Count Rate Distributions {#OXdistributions}
Another view of the behaviour of the optical and X-ray data can be obtained by examining the distribution of the count rates for each segment of the light curve between $\rm T_0+\,240$ks and 970ks (${\rm T_0}+\,2.8$ days and ${\rm T_0}+11.2$ days) and for this portion of the light curve as a whole. We compute the logarithmic count rate distribution in the optical/UV and X-ray bands for each segment, binning them dynamically with 20 data points per bin, and express it as a probability density function (PDF), following the methodology of [@utt05]. We then checked the consistency of the PDFs with a Gaussian distribution. For 5 separate optical/UV and X-ray segments we discard their PDFs as they contain 3 bins or less. For the 18 remaining segments, we find that 15 and 16 optical and X-ray segments respectively, can be well fitted by a Gaussian distribution, rejecting the null hypothesis for the remaining 3 optical ($\rm T_0+\,672$ks, $820$ks and $877$ks) and 2 X-ray ($\rm T_0+\,797$ks and $877$ks) segments with $p<10^{-3}$ from the $\chi^2$-test.
We can also examine the logarithmic count rate distribution during the entire period from $\rm T_0+\,240$ks $-$ 970ks (${\rm T_0}+\,2.8$ to ${\rm T_0}+11.2$ days), scaling individual X-ray and optical/UV segments by their mean count rate, before computing the PDF. We bin them dynamically with 100 data points per bin, following the methodology of [@utt05]. We first explore whether the distributions are consistent with a single Gaussian. We do not find the optical/UV or X-ray PDFs to be consistent with a Gaussian ($\chi_{O}^2/dof=$ 570/18 and $\chi_{X}^2/dof=$ 679/18), but the distributions are singularly peaked. We therefore also attempted to fit a model with 2 Gaussians distributions. This model substantially improved the fit for both the optical/UV and X-ray distributions, resulting in $\chi^2/dof=$ 16/15 and 24/15 for optical/UV and X-ray, respectively.
In Fig. \[flux\_dist\], we display the lognormal normalized count rate distributions for the $u$-band and X-ray observations, comparing the single Gaussian with the best fit 2 Gaussian models for each data set.
![Top Panel: Square root of the excess variance (RMS amplitude) versus the average count rate in log-log space. The dotted line represents the line of equality. The inset panel shows the data up to a count rate of 1000 cts/s in linear space. Bottom Panel: fractional root mean square of the excess variance (fractional RMS) versus the average count rate. Red Circles: $0.5-10$keV X-rays, Blue Triangles: $u$-band. All values are determined from 128 consecutive data points of duration 5s. The top panel suggests that for both the $u$-band and X-ray $0.5-10$keV data, the amplitude of the variability increases with brightness. This is confirmed in the bottom panel by normalizing the amplitude by brightness and therefore removing this trend.[]{data-label="rms_flux"}](./Images/RMS_flux.pdf)
Discussion
==========
We have explored and compared the optical/UV and X-ray behaviour of the June 2015 outburst of V404 Cygni. Overall the optical/UV filter light curves and the soft and hard X-ray light curves show similar behaviour. The main outburst consists of a series of large flares on time-scales of a few hours, but strong variations can also be observed at shorter time-scales from a few seconds to hundreds of seconds. The variations are most prominent in the X-rays, while in the optical/UV, they tend to be less prominent and smoother. The X-ray and optical/UV rms-flux ratio indicates that the strength of the variability in both bands is correlated with brightness, but less strongly in the optical. We have shown that strong changes in absorption, as inferred from X-ray spectral fits, do not correspond to changes in optical/UV extinction.
In the following we will first discuss and compare the extinction and absorption, then discuss the rms-flux relation and finally the origin of the variability in the X-ray and optical/UV light curves. We will draw upon all these aspects to provide an overall picture describing the variability observed within V404 Cygni by [*Swift*]{}.
![Comparison of the optical and X-ray brightness during the June 2015 outburst. The top panel displays the $0.5-10$keV X-ray count rate versus the $u$ band count rate from ${\rm T_0}$ to ${\rm T_0}+70$ days (0 - 6000ks). The bottom panel displays the $2-10$keV X-ray luminosity together with the UVOT [*u*]{}-band luminosity. The luminosities have been corrected for Galactic extinction and absorption. The colour represents the time of observation in days since trigger. The luminosity panel displays only X-ray data taken in WT mode since the signal to noise of the PC mode spectra is insufficient to constrain the flux conversion factor. In both panels, the black lines indicate where the $u$-band and X-ray count rate or luminosity are linearly proportional (dashed), and the square root of the X-ray is proportional to the optical (dotted). In the top panel, the $u$-band and X-ray count rates in general increase together, but as V404 Cygni brightens, the distribution becomes broader in the X-rays compared to the optical/UV. This is also observed in luminosity space given in the bottom panel.[]{data-label="OXcorr"}](./Images/OX_counts.pdf "fig:") ![Comparison of the optical and X-ray brightness during the June 2015 outburst. The top panel displays the $0.5-10$keV X-ray count rate versus the $u$ band count rate from ${\rm T_0}$ to ${\rm T_0}+70$ days (0 - 6000ks). The bottom panel displays the $2-10$keV X-ray luminosity together with the UVOT [*u*]{}-band luminosity. The luminosities have been corrected for Galactic extinction and absorption. The colour represents the time of observation in days since trigger. The luminosity panel displays only X-ray data taken in WT mode since the signal to noise of the PC mode spectra is insufficient to constrain the flux conversion factor. In both panels, the black lines indicate where the $u$-band and X-ray count rate or luminosity are linearly proportional (dashed), and the square root of the X-ray is proportional to the optical (dotted). In the top panel, the $u$-band and X-ray count rates in general increase together, but as V404 Cygni brightens, the distribution becomes broader in the X-rays compared to the optical/UV. This is also observed in luminosity space given in the bottom panel.[]{data-label="OXcorr"}](./Images/OX_lum.pdf "fig:")
Extinction and absorption {#extincabsorb}
-------------------------
The absorption, as expressed by the column depth $N_H$ observed in the X-ray spectra, changes by over two orders of magnitude. If the X-ray absorbing material contained a significant amount of dust affecting the optical/UV emission, we would expect to see a colour change of $>1.5$ magnitudes. If the absorption and extinction were correlated as found in the Milky Way, the data would folow the dashed black line in Fig. \[colour-NH\]. However, in Fig \[colour-NH\] there appears to be no correlation between colour and $N_H$.
[@motta17] suggest that there may be two absorbers: an inhomogeneous high density absorber and a uniform tenuous absorber, which either coexist or the high density rarefies into the lower, more uniform absorber (although rarefaction is less likely due to the fast transitions between the two phases). The separation of the two absorbers cannot be determined through the X-ray observations alone. From our observations, we can rule out the possibility that the optical/UV is affected by the inhomogeneous high density absorber due to the lack of corresponding colour changes and lack of correlated variability during the increase in the X-ray emission associated with sharp drops in the $N_H$, the latter implying the absorber must not be optically thick. Colour evolution was also not observed during a 20 min observation of an optical polarization flare, which ruled out intrinsic extinction as the cause [@sha16].
There are several possibilities why large extinction is not observed to be associated with the inhomogeneous high density absorber. It may be that the absorber either resides in-between the regions emitting the optical/UV and X-ray emission regions, or it is not large enough to cover the entire optical emitting region. Alternatively (or in addition), the absorber may be dust-free/have a very low dust-to-gas ratio, or have a flat distribution of dust grain size that results in a flat extinction curve. The dust-to-gas ratio, $E(B-V)/N_H$ must be of order $10^4$ smaller than that typically observed in the Milky Way, in order to not show significant change in colour as the absorption increases by a factor of 1000. To be dust-free the absorber must be within or close to the dust sublimation radius. Taking the quiescent X-ray luminosity for V404 Cygni of $8\times 10^{32}{\rm erg\,s}^{-1}$ [@plo17; @ber14], a dust sublimation temperature of 1500K and applying the formalism for the sublimation radius of AGN [@nen08a; @nen08], we obtain a sublimation radius of V404 Cygni in quiescence of $\sim10^{12}\,{\rm cm}$, consistent with the size of the accretion disk [@zyc99]. Absorption systems entering this radius from the interstellar medium or formed within this radius are likely to have their dust destroyed or are unable to form dust in the first place. The absorber is therefore unlikely to be dusty and thus we do not expect a correlation between absorption and extinction. Additionally, the second, thin absorber, is not ionized so it could correspond to the neutral out-flowing material launched as a disk wind $\sim10^{10}$cm from the central black hole [@mun16]. Since this region is also within the dust sublimation zone, this wind is unlikely to be dusty and therefore would not strongly affect the optical/UV emission.
Since the absorbers are not dusty, we may be able to gain some understanding of where the high density absorber lies relative to the regions producing the X-ray and optical/UV emission, assuming the optical/UV emission is produced from the reprocessed X-rays (see the following section for discussion of the origin of the optical/UV emission). We can examine two types of behaviour. The $\rm{T}_0+820$ks segment shows strong correlated changes in the optical/UV and X-ray light curves and has time resolved X-ray spectra. In this epoch, an X-ray flare of duration $\sim100$s is observed on top of longer term undulating behaviour. A flare is not observed in the optical/UV. The X-ray flare is associated with a rapid decrease in $N_H$, and the covering fraction during this period initially decreases before changing to 100 per cent coverage. In another instance, the two sections ${\rm T}_0+574$ks and ${\rm T_0}+579$ks, have a constant optical/UV flux level, but the X-ray flux changes by over an order of magnitude. In these instances a decrease in column density appears to be correlated with increase in X-ray flux. Both these types of behaviour suggest that the X-ray source is uncovered and is visible along our line of sight. If the absorber lies in-between the X-ray emitting region and the region where the X-rays are reprocessed to optical/UV, we would expect to see corresponding changes in the optical/UV brightness. While we do find optical and X-ray light curves at times to be correlated, we don’t observe the overall optical brightness to change between the ${\rm T}_0+574$ks and ${\rm T}_0+579$ks epochs. We suggest instead, that the high density absorber is large and lies at a significant distance away from the disk (so that the X-rays are only absorbed after the reprocessing of X-rays to optical/UV; but at this distance, it may already be cool enough for dust to exist - which would lead to extinction), or more likely that the absorber is small and close to the black hole, so that while the central object is uncovered along our line of sight, no additional X-ray flux is transferred to the region producing the optical emission. This is consistent with the schematic view of V404 Cygni given in [@motta17].
![The logarithmic count rate distributions of the outburst between $\rm T_0+\,240$ks and 970ks (${\rm T_0}+\,2.8$ days and 11.2 days) for the X-ray (red circles) and optical/UV (blue triangles), expressed as probability density functions. The data has been dynamically binned with 100 data points per bin. The dotted lines show the best fit Gaussian distributions ($u$-band: light blue, X-ray: pink). The solid lines show the best fit 2-Gaussian function for each data set (X-ray: purple, $u$-band: green). The count rates from each segment included in the distribution have been normalized by the mean count rate in that segment. The error on each bin is given as the $\sqrt{N}$, where N is the number of count rates in the bin. The $\chi^2$ over the degrees of freedom (dof) for the Gaussian fits are given in the top right corner of each panel.[]{data-label="flux_dist"}](./Images/Flux_distribution_total.pdf)
RMS-flux relation
-----------------
The magnitude of variability and the count rate of both the X-ray and $u$-band light curves scale linearly. Similar behaviour was found in X-ray observations of both GX 339-4 during outburst [@mun11] and Cyg X-1 [@gle04], and was initially explored in X-ray binaries by [@utt01]. Linear behaviour was also observed in the rms-flux correlations in both the X-ray and optical for a sample of 4 X-ray binaries [@gan09]. For GX 339-4, the rms was observed to increase with brightness, however, a continuous hysteresis pattern was observed forming a slanted P shape in the rms-flux (log-log) space [@mun11]. [@mun11] associated the different phases of the rms-flux diagram with specific spectral phases. The spine of the P shape is associated with the hard state identified on hardness intensity diagrams (HIDs) [e.g., see @bel16 for a recent review]. The curve of the P shape corresponded to the period where a thermal disk black body component was found in some of the spectral fits of GX 339-4 indicating that the XRB had entered the soft state. For Cyg X-1, the distinction between states in the rms-flux diagram was less prominent. [@gle04] showed that the soft and hard states both provided linear rms-flux relations, but the slope was slightly smaller for the soft state compared to the hard state. For V404 Cygni, there is a lot of scatter in the rms-flux relation. It cannot be determined if the observation is a combination of linear relations, but the rms-flux behaviour observed in V404 Cygni is most similar to that found for Cyg X-1 and the XRBs reported by [@gan09].
Information on the state of the outburst can be determined using the X-ray spectra and rms amplitude of V404 Cygni. In XRBs, the hard state is normally characterized by a power-law-shaped energy spectrum with a photon index of $\sim1.6$ ($2-20$keV), and a high level of aperioic variability with an rms amplitude above ∼30 per cent [@mun11]. The hard state would be characterised by a thermal disk black body component on the left-hand side of the HID [@bel16] and almost no variability is seen [@mun11]. For V404 Cygni, none of the X-ray spectra require disk black body components [@motta17]. They are best fitted during the main outburst by a power-law with photon index $\lesssim 2$ ($0.6-10$keV) and the fractional rms, in the X-ray, on the whole ranges between 10 and 60 per cent. V404 Cygni is therefore consistent with being in the hard state during the main outburst, with the observed rms fluctuations and rms-flux correlation associated with the power-law emission. Two common explanations for X-ray emission in the hard state are that it arises from a ‘corona’ of hot electrons, where seed photons from an optically thin accretion disc are up-Comptonized, or that it originates from a jet [@mun11]. The outburst of V404 Cygni therefore never likely reached the disk-dominated soft state [see also @motta17]. The near Eddington luminosities could be explained by accretion at Eddington/Super-Eddington rates, which may have sustained a surrounding slim disk, but the flux was partly or completely obscured by the inflated disk and its outflow. In this case, the largest flares produced by the source might not be accretion-driven events, but instead the effects of the unveiling of the extremely bright source hidden within the system [@motta17].
In the lower panel of Fig. \[rms\_flux\], we have 3 X-ray data points with 100 per cent fractional variability. At these epochs the X-ray light curve is dominated by large undulations in the baseline flux level, changes in flux of order 100. In one case (${\rm T_0}+579$ks) the undulation is a large flare[^8], which corresponds to a strong decrease in $N_H$, likely associated with the uncovering of the central object. For the other two points with 100 per cent fractional variability, strong changes in $N_H$ are observed, but do not seem to be the only factor driving the strong changes in X-ray flux. Exploring the rest of the rms values (e.g., those with fractional variability $< 100$ per cent), we find strong changes in $N_H$ are only observed during 2 further data points. This indicates that varying $N_H$ is not solely responsible for the observed rms-flux relation. The optical/UV emission has a lower fractional rms, typically of a few to 20 per cent and only weak evidence for a correlation between rms and flux. This suggests that the mechanism producing the X-ray emission does not directly produce the optical emission, and that, perhaps the X-ray reprocessing is modulating the optical/UV variability.
[@utt05] discuss the rms-flux relation as being produced by stationary data. For a light curve to be considered as having stationary variability, the mean and variance should not change with time and thus the flux distribution should be consistent with lognormal. In §\[OXdistributions\], we examined the count rate distributions for each light curve segment, finding that a lognormal distribution is consistent with what we observe for the majority of the individual X-ray and optical light curve segments (see §\[OXdistributions\],) between $\rm T_0+\,240$ks and 970ks ($\rm T_0+\,2.8$ days and 11.2 days). This suggests that the fluctuations for these segments are not additive and are unlikely to be caused by many independent emitting regions or due to shot noise [@utt05]. Since most of the light curve segments are consistent with having a lognormal count rate distribution, it is likely that the PSDs used to determine the rms values (1 or 2 per light curve section) are in general measuring stable variability. However, there are 3 and 2 optical/UV and X-ray light curve segments, respectively, where the count rate distributions are inconsistent with a lognormal distribution, suggesting non-stationary behaviour for at least some portion of the light curve. Non-stationary behaviour was also shown by [@gan17] for V404 Cygni, in cross correlation of optical and X-ray light curves. We are also unable to state if the variability is stable on much shorter timescales due to the resolution of the data. Furthermore, examining the normalized count rate distribution of the entire time period, we note that the logarithmic X-ray and optical/UV distributions are not consistent with a single Gaussian. This may suggest that the emission process does not result in a logarithmic flux distribution that is Gaussian, or that the emission in the two bands may be produced by more than one process [see also @als19].
X-ray and optical variability
-----------------------------
Within the [*Swift*]{} high time resolution observations the variability shows three main characteristics: 1) fluctuations on tens of seconds up to a couple of hundred seconds time scales, observed in both bands, but more prominent in the X-ray; 2) larger variations changing the flux level over hundreds of seconds (e.g., ${\rm T_0}+877$ks, see supplementary material Fig. S1), observed in both bands, although more strongly in one (typically the X-ray) compared to the other; 3) the base-line count rate of both bands change between orbits/segments of data, e.g., the ratio between the optical and X-ray count rate varies from epoch to epoch, typically the X-ray varies more strongly than the optical/UV.
The range in behaviour between the optical/UV and X-rays suggests that the variability may not be due to a single mechanism. One possible explanation for the production of the optical/UV variability is reprocessing of the X-ray emission by the disk [@gan16; @kim16; @rod15]. The X-ray emission originates from the inner regions near to the black hole. Modelling of spectral energy distributions (SEDs) performed by [@kim16] suggests reprocessing is present. The high degree of correlation between the $u$-band and X-ray light curves observed in two of our light curve segments supports reprocessing as the dominant source of emission at least for this portion of the outburst. The light curve segment at ${\rm T_0}+$820ks shows large scale variability in the X-ray and similar emission in the optical, but that is broader in time compared with the X-ray. This is a classic expectation for reprocessing, where by the lag and blurring of the optical light curve depends on the reprocessing area [e.g. @obr02; @hyn06]. Therefore, when the optical/UV variability is broader than the X-rays, we can assume that the optical/UV emitting region is larger than the X-ray. [@motta17], using [*Swift*]{}/XRT and INTEGRAL observations, constrain the X-ray emission to within 100 gravitational radii, with softer X-rays being produced within 10 gravitational radii. [@wal17] with NuSTAR observations also suggest that the X-ray emission, at least during the flares, is produced close to the black hole within ∼10 gravitational radii. With a black hole mass of 9 $M_\odot$, it has been estimated that the the separation between the black hole and the companion is $\sim2.2\times10^{12}$ cm or $\sim75$ light seconds [@rod15; @gan17]. The short time lag observed for the segments of our optical and X-ray light curves, similar to that reported by [@ATel:7727], [@kim16] and [@alf18], is consistent with this picture. However, our typical measured optical lag of $<35$ sec, may suggest that the $u$-band emission from reprocessing does not originate at the very outskirts of the disk, but at a radius of $\lesssim1\times10^{12}$ cm. Standard disk reprocessing is expected to produce optical variations that follow an $L_X^{1/2}$ trend [@van94]. This relationship is shown in Fig. \[OXcorr\], but it is not consistent with the overall behaviour observed in either count rate or luminosity space. In Fig. \[OXcorr\] horizontal track behaviour is observed with the X-ray luminosity appearing to increase without significant change in the optical/UV, this is also observed in the equivalent count rate figure for the ${\rm T_0}+$877 segment shown in Fig. S.2. This horizontal behaviour could suggest that either the X-ray and optical/UV emission are produced independent or that the optical/UV emitting region is unable to respond to the increase in the X-rays [see also @hyn19]. However, the optical/UV count rate versus X-ray count rate figure of the ${\rm T_0}+$877 segment also displays a vertical line, which shows the optical/UV count rate changing while the X-ray count rate remaining constant. This could suggest a scenario where by the production of the optical/UV emission is delayed in comparison to the X-ray perhaps consistent with reprocessing. However, we can not draw strong conclusions from these figures since the X-ray count rates and luminosities are not corrected for intrinsic absorption, which may be responsible in part for the observed horizontal track behaviour in Fig. \[OXcorr\].
There is considerable evidence suggesting that reprocessing is not solely responsible for producing the optical/UV and X-ray variability. Lack of X-ray variations coincident with optical variations suggest an alternative mechanism to reprocessing for producing at least some of the optical variations [@dal17]. In addition, long lags discovered between the optical and INTEGRAL data by [@rod15] from 0s up to 20-30min [see also @alf18] also cannot readily be explained by reprocessing given the size of the binary. Within the [*Swift*]{} XRT/UVOT observations, we also find that reprocessing may not be responsible for all the observed temporal fluctuations. For instance, in the ${\rm T_0}+$671 $-$ ${\rm T_0}$+672ks segment the dip in the optical/UV count rate is stronger than that observed in the X-ray.
Variable absorption has also been proposed as the cause of at least some of the X-ray variations [@kin15; @motta17]. [@motta17] identified strong spectral variations in the time resolved $0.5-10$keV X-ray spectra of V404 Cygni, which they identify as due to variations in the layer of material covering the source. They find that changes to the column density cause moderate changes to the flux on timescales of minutes, but that larger flux variations are due to sudden drops in the local column density and simultaneous disappearance of partial covering. These changes in column density are proposed as the covering/uncovering of a bright central source causing the large scale flux variations. Indeed, strong change in $N_H$ is observed within the ${\rm T_0}+$820ks and ${\rm T_0}+$877ks segments and between the ${\rm T_0}+$574ks and ${\rm T_0}+$579ks segments. We do not observe flux variations in the optical/UV that are coincident with changes in the X-ray flux due to varying absorption. This is to be expected as the optical emission is unlikely to be affected by the local absorbers since the dust is likely sublimated (see §\[extincabsorb\]). While changing absorption is a cause of X-ray variability in V404 Cygni, it is unlikely to be the sole cause since we observe correlated variable behaviour between the optical and X-ray light curves, even when rapid, short lived changes in flux due to changing $N_H$ are observed in the X-ray, e.g., ${\rm T_0}+820$ks and ${\rm T_0}+877$ks segments.
An alternative mechanism for producing the observed X-ray and optical/UV emission and associated variability is a non-thermal jet. Evidence for jet emission in V404 Cygni has been shown in the hard X-rays through to radio wavelength observations [see also @mai17 for a list of references for jet emission in the 2015 outburst of V404 Cygni]. The earliest evidence for a jet was observed by INTEGRAL between ${\rm T_0}$+2 to ${\rm T_0}$+5 days ($\sim \rm T_0+\,170$ks - 430ks), with hard emission due to Compton-upscattering of seed photons likely originating in a synchrotron driven jet [@nat15; @roq15]. At optical/UV wavelengths, [@gan16; @gan17] propose that the jet was responsible for producing the sub-second optical/UV variability observed by ULTRACAM during the brightest phase of the outburst, while [@mai17], propose that the slow variability is also a result of synchrotron emission from the jet because they see no optical spectral changes over a factor of 20 change in optical flux. They propose that the sub-second variability and slow variability may originate from two different jet components: an optically thick component that produces the slower variations and an optically thin component that exhibits red and rapid variability. [@dal17] also find evidence for optical variability being produced by the jet; they focus on a sharp drop in flux in X-ray - IR bands at MJD 57198.23, which lasts $\sim 100$s. They exclude a thermal origin for the emission during this decline, ruling out X-ray reprocessing as the source of the optical/IR emission. They instead conclude synchrotron cooling process is the dominant source of the optical/IR emission. [@gan17] also find, coincident with an increase in radio brightness, that the optical lags the X-ray by $\sim 0.1$s. They conclude that the X-ray and optical emission is attributed to synchrotron emission from the jet, with the optical being produced at a distance $\simeq 10^3$ gravitational radii from the X-ray emission.
It is likely that at least some of the optical/UV variability observed by the UVOT results from the jet, potentially contributing to the optical emission from ${\rm T_0}$+2 days ($\sim \rm T_0+\,170$ks) onwards. The dip observed between the ${\rm T_0}+671$ks and ${\rm T_0}+672$ks segments may be consistent with a picture of varying accretion causing variability within the jet. Examining the colour terms in Fig. \[v404lc\_multipanel\], we can derive a spectral slope from the $uvw1-u$ colour term, which remains approximately the same during the main outburst, resulting in a spectral index $\beta \sim 0.3$; this value is approximate and is dependent on uncertainties in the reddening correction and flux conversion factor. However, this spectral index is consistent with the spectral index between $V$ and $Ic$ filters observed by [@mai17] and the $u'$ and $g'$ filters observed by [@gan16], although the high time resolution spectral energy distributions (SEDs) of [@gan16] show that the spectral slope changes significantly, typically between 0-0.5 during their observations on 4 different nights. A spectral index of $1/3$ is consistent with what is expected for a steady-state disk with radiation due to viscous dissipation, but [@mai17] favour an optically thick jet, due to the persistence of the spectrum through the large variability and because typical X-ray transients show steeper optical spectra [@hyn05].
Overall, the discrepancies from standard reprocessing and the evidence of a jet, suggest that the variability we observed is built up from a complex amalgamation of processes and conditions that dominate at different times during the outburst. The lack of strong cross-correlation coefficients and a dominating lag time supports our conclusion requiring the outburst to be a result of multiple processes that may be contributing at different levels during different segments of light curve. This will likely reduce or at least complicate any potential cross-correlation signal.
Conclusions
===========
We have presented observations of the 2015 outburst of V404 Cygni, performed by [*Swift*]{}. We examined observations taken simultaneously by the UVOT and XRT in the optical/UV and X-ray bands, respectively, comparing the properties of the source in these two bands.
We examined the relationship between extinction and (neutral) absorption measured from the optical/UV and X-ray data, respectively, finding no correlation between the two. Due to the lack of corresponding colour changes in the optical/UV band we excluded the possibility that (in contrast to the X- ray emission) the optical is affected by an inhomogeneous high density absorber. We conclude that the high density absorber is likely small, contains a negligible amount of dust and resides close to the black hole.
We also analysed the temporal behaviour of the optical/UV and the X-ray emission. We found that strong variations can be observed on tens of seconds to hourly time-scales. The variations are most prominent in the X-rays, as opposed to those observed in the optical/UV, which are less prominent and smoother. In both bands the amplitude of the variations are correlated with flux, but to a lesser extent in the optical/UV. We found that within a few outburst phases the optical/UV and X-rays are correlated, with an optical lag typically $15-35$s. We conclude that the variability in the optical/UV and X-ray bands is due to a complex combination of processes. Variability visible in both the X-ray and optical/UV bands may be accretion-driven, with either the X-rays produced in the inner accretion disc, which is then reprocessed to the optical/UV, or with both the X-rays and optical/UV produced within the jet. In addition, some of the variability that is observed in the X-ray band only, may be due to the presence of a local, inhomogeneous and dust-free absorber.
Acknowledgements
================
We thank the referee for their constructive comments and suggestions. This research has made use of data obtained from the High Energy Astrophysics Science Archive Research Center (HEASARC) and the Leicester Database and Archive Service (LEDAS), provided by NASA’s Goddard Space Flight Center and the Department of Physics and Astronomy, Leicester University, UK, respectively. SRO gratefully acknowledges the support of the Leverhulme Trust Early Career Fellowship. AJCT and SRO acknowledge support from the Spanish Ministry, Project Number AYA2012-39727-C03-01. SEM acknowledges the Violette and Samuel Glasstone Research Fellowship program and the UK Science and Technology Facilities Council (STFC) for financial support. PG thanks STFC and UKIERI for support. This work was also supported by the UK Space Agency. We also thank J. Lyman for useful discussion.
\[appendix\]
\[Dec\_app\]
December 2015 Outburst
======================
At ${\rm T_0}\sim190$ days, [*Swift*]{}/BAT detected a secondary outburst [@GCN:18716]. At this time, [*Swift*]{} UVOT and XRT resumed observing and continued to monitor V404 Cygni for an additional 30 days, although the sampling was less frequent compared with the June 2015 outburst. For these observations the optical/UV and X-ray data were reduced following the method outlined in §\[reduction\].
The light curves from the UVOT, XRT and INTEGRAL can be observed in Fig. \[Dec2015\]. At this time, the optical/UV flux level was already above the quiescent level observed prior to 2015 and at the end of the June 2015 outburst. The brightest flux values occurred between $\rm T_0$+199 and ${\rm T_0}+200$ days. After peak brightness, V404 Cygni rapidly decayed, within a day, down towards the pre-2015 level, which it fluctuated around until the end of observations. The X-ray light curves observed by XRT and INTEGRAL behave in a similar way. This mimics the behaviour observed in the June 2015 outburst and the duration of the outburst is similar. However, the outburst is weaker at all frequencies. In the $u$-band the peak magnitude is $\sim 3$ orders of magnitude fainter than the peak magnitude of the June 2015 outburst. This is in agreement with the behaviour observed for the June and December outbursts in i, R, V and B bands reported by [@kim17]. Examination of the colour light curves, for which the best sampled for the December outburst is $u-uvw1$, suggests the spectrum becomes bluer from the start of the outburst through to quiescence, although we note that, except for the flare at ${\rm T_0} \sim 198$days, the $uvw1$ band is relatively flat in comparison with the $u$ band and the colour change is driven by a change in the $u$-band flux.
[^1]: Image Mode: the data are recorded as an image accumulated over a fixed period of time; Event Mode: the arrival time and position are recorded for every photon detected.
[^2]: https://heasarc.gsfc.nasa.gov/docs/heasarc/caldb/swift/docs/\
uvot/uvotcaldb\_sss\_01.pdf
[^3]: We also tried varying the Fourier amplitude, using a Gaussian random sampling, taking the width of the Gaussian to be 10% of the amplitude. We did not find any substantial changes to the 0.5, 2.5, 97.5 or 99.5 per cent values of the cross correlation distribution at a given lag time.
[^4]: both IDL scripts are available at http://www.star.le.ac.uk/sav2/idl.html.
[^5]: http://www.isdc.unige.ch/integral/analysis\#QLAsources
[^6]: where the start time is the start time of the earliest of the two exposures and the end time is the stop time of the latest exposure.
[^7]: Since the PSDs are a result of a single time series. The measured PSD will therefore not be the true PSD, but will be scattered around it [see e.g @vau03]. As a result, the fractional rms values are typically averaged, but due to the highly variable X-ray absorption, the same flux values do not necessarily imply the same spectral properties. Therefore we have chosen not to provide the average fractional rms.
[^8]: this X-ray flare, peaking at ${\rm T_0}+578.8$ks, is not displayed in online material S.1. as there is no overlapping $u$-band event mode data
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Non-perturbative constraints on many body physics–such as the famous Lieb-Schultz-Mattis theorem–are valuable tools for studying strongly correlated systems. To this end, we present a number of non-perturbative results that constrain the low-energy physics of systems having conserved dipole moments. We find that for these systems, a unique translationally invariant gapped ground state is only possible if the polarization of the system is integer. Furthermore, if a lattice system also has $U(1)$ subsystem charge conservation symmetry, a unique gapped ground state is only possible if the particle filling along these subsystems is integer. We also apply these methods to spin systems, and determine criteria for the existence of a new type of magnetic response plateau in the presence of a non-uniform magnetic field. Finally, we formulate a version of Luttinger’s theorem for 1D systems consisting of dipoles.'
author:
- Oleg Dubinkin
- 'Julian May-Mann'
- 'Taylor L. Hughes'
bibliography:
- 'References.bib'
title: 'Lieb Schultz Mattis-Type Theorems and Other Non-perturbative Results for Strongly Correlated Systems with Conserved Dipole Moments'
---
Introduction {#sec:intro}
============
Strongly correlated quantum systems are often the source of striking phenomena, yet they remain one of the most challenging to analyze. If perturbative approaches fail, one’s only recourse is numerical simulation unless non-perturbative methods or results can be applied. One of the most celebrated non-perturbative results is the Lieb-Schultz-Mattis (LSM) theorem, which from very little information, i.e., the number of spin-$1/2$ degrees of freedom per unit cell, can draw conclusions about the low-energy properties of 1D spin chains[@lieb1961]. This result has had wide-ranging applicability in quantum systems, and has been extended to higher dimensions[@oshikawa2000A; @hastings2004; @hastings2005; @nachtergaele2007].
A key feature of the proof of the LSM theorem is a twist operator that slowly rotates the spins across a spin-chain. In the work of Oshikawa[@oshikawa2000A], which applied and generalized the LSM result to lattice systems with conserved particle number, a related twist operator is used $$U_{X}=\exp\left[ \frac{2\pi i\hat{X}}{L_x}\right],
\label{eqn:twist_op}$$where $\hat{X}=\sum_{{\bf{x}}} x \hat{n}_{\bf{x}}$ is the many-body position operator. Indeed, Oshikawa and collaborators also used this operator to provide a non-perturbative understanding of Luttinger’s theorem[@luttinger1960; @oshikawa1997], determine the Fermi surface properties of the Kondo lattice[@oshikawa1997; @oshikawa2000B; @ueda1994; @otsuki2009], and more recently to calculate filling-enforced constraints on the quantum Hall conductivity in lattice systems[@lu2017filling]. Remarkably, this operator $U_{X}$ has had parallel uses in the theory of electronic polarization where it was introduced by Resta[@resta1998]. In this context, the complex phase of the ground state expectation value of $U_{X}$ is determined by the electronic polarization[@resta1998], and the magnitude is determined by the electron localization length[@resta1999; @aligia1999; @souza2000].
In recent work, new twist operators have been proposed whose ground state expectation values can be used to calculate higher multipole moments[@wheeler2018many; @gil2018; @dubinkinmaymannhughes2019; @you2019]. The simplest multipole generalization of $U_{X}$ is the operator $$U_{XY}=\exp\left[\frac{2\pi i\widehat{XY}}{L_x L_y}\right]
\label{eqn:multipole_twist}$$where $\widehat{XY}=\sum_{{\bf{x}}}x y \hat{n}_{\bf{x}}$ is the many-body quadrupole operator. In light of this development it is natural to use these multipole operators to try to derive non-perturbative results analogous to the previous work on the LSM theorem[@oshikawa2000A; @lu2017lieb], Luttinger’s theorem[@oshikawa1997; @oshikawa2000B], magnetization plateaus[@affleck1989; @oshikawa1997mag], and filling-enforced Hall conductivity constraints[@lu2017filling]. In this article we focus on higher multipole generalizations of some of these results, derived through the application of operators related to $U_{XY}$ in each context. Our goal is to recast the original results that apply to particles/charges to apply to dipoles. Notably, we study generalizations of the LSM theorem for systems that conserve dipole moments (Section \[sec:dip\_LSM\]), and then apply these results to study magnetization (gradient) plateaus in spin systems (Section \[sec:plateau\]), and an extension of Luttinger’s theorem to dipole conserving systems (Section \[sec:luttinger\]). We also extend our results on the LSM-type theorems to systems with $U(1)$ subsystem symmetry, e.g., symmetries enforcing charge conservation along rows or columns in 2D[@batista2005; @you2018subsystem]. These results are applicable to some fracton systems, and it is possible they may eventually be adapted in some form to systems with broken subsystem symmetry[@you2018higher; @you2019] (and broken microscopic dipole conservation), e.g., higher order multipole band insulators of fermions or bosons[@benalcazar2017quantized; @you2018higher; @dubinkin2018higher], though we leave those developments to future work.
LSM-Type Theorems for Dipole Conserving Systems {#sec:dip_LSM}
===============================================
Let us briefly recount the concept behind Oshikawa’s proof[@oshikawa2000A] of the LSM theorem that we wish to generalize. His starting point was a system with $U(1)$ particle conservation, and he utilized a twist operator $U_X$ that achieves two different things simultaneously: when applied to the ground state of a periodic insulating system, it extracts the charge polarization of the ground state, *and* it performs a large gauge transformation on this state, i.e., it effectively inserts or removes a single unit of magnetic flux through the periodic loop running along the $\hat{x}$-direction. To see how this works, consider a nearest-neighbor lattice model of free fermions in 1D with $N_x$ sites along $\hat{x}:$ $H=-t\sum_{j=1}^{N_x}(c^{\dagger}_{j+1}c^{\phantom{\dagger}}_j+h.c.)$. Applying $U_X$ to the Hamiltonian modifies every fermionic hopping term by attaching a phase factor of $\text{e}^{2\pi i/N_x},$ which is equivalent to introducing an external gauge field $A_x$ with a $2\pi$ circulation along the loop going around the periodic $\hat{x}$-direction. If $U_X$ is applied to the ground state $\vert\Psi_0\rangle$ of this free-fermion system, for example, then, in the thermodynamic limit, a set of filled bands will return back to themselves up to a phase proportional to the polarization, while a partially filled band will change momentum and be orthogonal to the original state. This is a very simple application of Oshikawa’s results[@oshikawa2000A] that indicates that partially filled bands cannot support an excitation gap since the energy of the state $U_x\vert\Psi_0\rangle$ will nominally approach that of $\vert \Psi_0\rangle$ itself in the thermodynamic limit, and the two states are orthogonal if there are partially filled bands because they have different momentum quantum numbers.
Dipole Conserving Systems {#sec:dip_LSM_A}
-------------------------
To make progress toward an LSM-type theorem for systems with charge and dipole conservation, let us consider a system defined on a rectangular periodic $L_x \times L_y=N_x a\times N_y a,$ lattice where $a$ is the lattice constant. We will work with Hamiltonians $H$ that are translationally invariant and conserve both global particle and dipole number. Hence, the Hamiltonian is invariant when the charged operators are changed by constant phase transformations $\text{e}^{i\alpha},$ and phase transformations with linear coordinate dependence $\text{e}^{i{\boldsymbol{\alpha}}\cdot{\bf{x}}},$ respectively. The latter condition also automatically implies that $H$ commutes with the twist operators $U_X$ and $U_Y$, i.e., the total polarization is a fixed quantum number for eigenstates of $H.$ We note that it has been shown[@pretko18gauge] that systems with dipole conservation can be coupled to a rank-2 gauge field $A_{ij}$. The rank-2 gauge field transforms as $A_{ij} \rightarrow A_{ij} + \partial_{i}\partial_j \lambda$ under a gauge transformation, where $\lambda$ is a generic function.
Similar to Oshikawa’s work, we will be considering the action of generalized $U_{XY}$ twist operators on the ground states of insulating systems: $U_{XY}(\alpha)=\text{e}^{i\alpha\widehat{XY}}$. For $\alpha=\frac{2\pi}{L_x L_y},$ and for systems with open boundaries, this operator was shown to be successful in extracting the quadrupolar polarization[@wheeler2018many; @kang2018many; @dubinkinmaymannhughes2019] and, when applied to systems that explicitly conserve dipole moment, it introduces a constant rank-2 gauge field $A_{xy}$ across the lattice. However, if periodic boundary conditions are introduced, $U_{XY}(\frac{2\pi}{L_x L_y})$ exhibits problematic behavior. For instance, while the complex *phase* of its ground-state expectation value $\langle U_{XY}(\frac{2\pi}{L_x L_y})\rangle_0,$ computed for the ground states of free-fermion tight-binding models, correctly captures the quadrupolar polarization, the absolute value of the expectation value vanishes when the thermodynamic limit is taken because of fluctuations of the dipole moments. Even if we mitigate this issue by restricting ourselves to Hamiltonians with manifest dipole conservation, as we shall do in this article, the dipole-conserving terms in the Hamiltonian that cross the periodic boundary pick up an additional phase factor under the action of $U_{XY}(\frac{2\pi}{L_x L_y}).$ To see this, one can act on dipole-conserving terms in the Hamiltonian, e.g., ring-exchange terms $$\begin{split}
U^{-1}_{XY}(\alpha)\Big(c^{\dagger}_{x,y}c^{\phantom{\dagger}}_{x+a,y}&c^{\dagger}_{x+a,y+a}c^{\phantom{\dagger}}_{x,y+a}\Big)U_{XY}(\alpha)\nonumber\\=&\text{e}^{i\alpha}c^{\dagger}_{x,y}c^{\phantom{\dagger}}_{x+a,y}c^{\dagger}_{x+a,y+a}c^{\phantom{\dagger}}_{x,y+a}.\nonumber
\end{split}$$ For terms that cross a periodic boundary additional phase factors are generated, e.g., $\text{e}^{i\alpha L_x a}$ or $\text{e}^{i\alpha L_y a}.$ Indeed, one can check that in order to have $U_{XY}(\alpha)$ insert a constant rank-2 gauge field $A_{xy}$ for a system with periodic boundary conditions then $\alpha L_x a =2\pi\mathbb{Z}$ and $\alpha L_y a=2\pi \mathbb{Z}.$ Hence, in order to be consistent with periodic boundary conditions, we will choose $\alpha=\frac{2\pi}{a^2 \gcd(N_x, N_y)}$ where $a$ is the lattice constant in the $x$ and $y$ directions.
To proceed from this setup, let us consider adiabatically turning on a constant gauge field configuration of $A_{xy}$ over a time $T$ having the form $$A_{xy} = \frac{2\pi }{a^2\gcd(N_x,N_y)}\frac{t}{T}.$$ Let us label the Hamiltonian as a function of time as $H(t),$ and its instantaneous ground state as $\ket{\Psi(t)}$. Since the initial system is translationally invariant, $\ket{\Psi(0)}$ is an eigenstate of the many-body translation operators $T_x$ and $T_y$ that send each particle coordinate $(x,y)$ to $(x+a,y)$ or $(x,y+a)$ respectively. We will take the $T_x$ eigenvalue to be $\text{e}^{iP_{x0}}$ and the $T_y$ eigenvalue to be $\text{e}^{iP_{y0}}$. Since $H(t)$ is translationally invariant at all times, $\ket{\Psi(t)}$ will remain an eigenstate of $T_x$ and $T_y$ with the same eigenvalues $\text{e}^{iP_{x0}}$ and $e^{iP_{y0}}$ at all times. Similarly, since $H(t)$ commutes with both $U_X$ and $U_Y$ at all times, $\ket{\Psi(t)}$ will also remain an eigenstate of $U_X$ and $U_Y$ with eigenvalues $\text{e}^{2\pi i X_0/L_x}$ and $\text{e}^{ 2\pi i Y_0/L_y}$ at all times. At $t = T$, we have $A_{xy} = \frac{2\pi }{a^2\gcd(N_x,N_y)},$ which is equivalent to a (large) gauge transformation[@you2019] that can be removed by applying the many-body twist operator $U_{XY}(\alpha_g)$ where $\alpha_g = \frac{2\pi}{a^2\gcd(N_x,N_y)}$. As a result $$H(0) = U_{XY}^{-1}(\alpha_g)H(T)U_{XY}(\alpha_g),$$ and so $U_{XY}^{-1}(\alpha_g)\ket{\Psi(T)}$ is also a ground state of $H(0)$.
The next key step for an LSM-type theorem is to determine if $U_{XY}^{-1}(\alpha_g)\ket{\Psi(T)}$ and $\ket{\Psi(0)}$ are orthogonal. To do this, we will compare the eigenvalues of the translation operator for $U_{XY}^{-1}(\alpha_g)\ket{\Psi(T)}$ and $\ket{\Psi(0)}$. A simple calculation shows that $$\begin{split}
T_x U^{-1}_{XY}&(\alpha_g)\ket{\Psi(T)}\\
&= \exp\left[\frac{2\pi i \hat{Y}}{a\gcd(N_x,N_y)}\right] e^{i P_{x0}} U^{-1}_{XY}(\alpha_g) \ket{\Psi(T)}.
\end{split}$$ We already know that $T_x\ket{\Psi(0)}=\text{e}^{iP_{x0}}\ket{\Psi(0)},$ so the ground state is unique only if $\exp\left[\frac{2\pi i \hat{Y}}{a\gcd(N_x,N_y)}\right]\ket{\Psi(T)} = \ket{\Psi(T)}$. If we define $n=L_y/(a\gcd(N_x,N_y))$, then $\exp\left[\frac{2\pi i \hat{Y}}{a \gcd(N_x,N_y)}\right] = (U_Y)^n$, and $$\exp\left[\frac{2\pi i \hat{Y}}{a\gcd(N_x,N_y)}\right]\ket{\Psi(T)} = \text{e}^{2\pi i n Y_0/L_y}\ket{\Psi(T)},$$ where we have used that $U_Y \ket{\Psi(T)} = \text{e}^{2\pi i Y_0/L_y}\ket{\Psi(T)}$. Since $(U_Y)^{N_y} = 1$, $\text{e}^{2\pi i Y_0/L_y}$ must be an $N_y^{\text{th}}$ root of unity, and $Y_0/L_y \equiv
\frac{1}{2\pi i} \log(e^{2\pi i Y_0/L_y})$ must be a rational number. We can define $Y_0/L_y = p/q$, where $p$ and $q$ are co-prime. So $U^{-1}_{XY}(\alpha_g)\ket{\Psi(T)}$ is an eigenstate of $T_x$ with eigenvalue $\exp(iP_{x0} + 2\pi i n p/q)$. If $n$ and $q$ are co-prime $U^{-1}_{XY}(\alpha_g)\ket{\Psi(T)}$ must be orthogonal to $\ket{\Psi(0)}$, and thus the system is either gapless, or there must be at least $q$ degenerate ground states if the system is gapped.
This statement relies on $n$ and $q$ being co-prime, but if $n$ is an integer multiple of $q$ nothing can be said about the degeneracy of the ground state. A similar issue was remarked upon in Oshikawa’s proof of the LSM theorem[@oshikawa2000A]. Here, we can avoid this issue by requiring that the thermodynamic limit is taken such that $N_x = N_y=N$. Since ground state properties should be independent of how the thermodynamic limit is taken, we can assume that the the thermodynamic limit is taken in this way. From this we can conclude that the ground state of the system is unique only if $\text{e}^{2\pi i Y_0/Na} = 1$. Using the same logic, i.e., by acting with the translation operator $T_y,$ we can also conclude that the ground state is unique only if $\text{e}^{2\pi i X_0/Na} = 1$ as well. In other words, for the ground state to be unique, we must require that the components of the polarization in both directions vanish up to a polarization quantum. The bulk of a locally electrically neutral system must carry a uniform polarization, which allows us to relate a microscopic dipole moment stretching between the pair of neighboring unit cells to a macroscopic polarization of the system picked up by the phase of the unitary operator $\text{e}^{2\pi i \hat{X}_j/N_ja}$. Therefore, we see that the pair of conditions $\text{e}^{2\pi i Y_0/Na}=\text{e}^{2\pi i X_0/Na}=1$ is equivalent to requiring that the microscopic dipole moments stretching between every pair of neighboring unit cells must be an integer times $ea$. Thus we conclude that filling factor for $x$ and $y$ dipoles must be an integer, analogous to the requirement that the charge filling factor be integer in the conventional LSM theorem.
Subsystem Symmetric Systems {#sec:subs_LSM}
---------------------------
We can construct a stronger version of the above dipole LSM theorem for Hamiltonians with dipole conservation arising from $U(1)$ subsystem symmetry. Let us first provide a brief background discussion on subsystem symmetries. To give an example, consider a 2-dimensional $L_x\times L_y$ rectangular lattice. The subsystem symmetry operator corresponding to $U(1)$ charge conservation along a single column with $x=x_0$ is given by: $$U_{0,x_0}(\alpha) = \exp\left(i\alpha \sum^{L_y}_{y=1} \hat{n}_{x_0,y}\right).$$ Such operators rotate the phase of all electrons along a single column in the lattice. In other words, $U_{0,x_0}$ can be thought of as a restriction of the global $U(1)$ symmetry operator $U_0(\alpha)=\exp\left(i\alpha \sum^{L_x,L_y}_{x,y=1} \hat{n}_{x,y}\right)$ to a particular subsystem. Similarly, we can define subsystem symmetries $U_{0,y_0}$ that impose charge conservation along every single row $y_0$ of the lattice. For the purposes of our work, by an $n$-dimensional subsystem in a $d$-dimensional Bravais lattice, we will understand an $n$-dimensional lattice subspace spanned by any $n$ linearly independent lattice basis vectors. Now, taking a collection of such “parallel" subspaces that cover the entire lattice, we can impose $U(1)$ charge conservation along *each* of the subspaces individually. This restriction leads to a conservation of *all* multipole moments[@gromov2018multipole] in a $(d-n)$-dimensional subspace orthogonal to these subsystems.
Coming back to our two-dimensional lattice example, take a collection of parallel lattice lines that cover the whole lattice. For concreteness, let us take a collection of lattice rows that are parallel to $\hat{x}$. Imposing charge conservation along each row is equivalent to fixing the total charge at each position along $\hat{y},$ which is the normal vector to these subsystems. Thus, for any arbitrary function $f(y)$, the conservation of the quantity $$Q=\sum_{x,y} f(y)q(x,y),$$ where $q(x,y)$ is the charge at a site with coordinates $(x,y)$, is guaranteed by the $U(1)$ subsystem charge conservation. For example, by taking $f(y)=y^m$, we can see that all multipole moments along $\hat{y}$, such as the $P_y$ component of the dipole moment, the $Q_{yy}$ component of the quadrupole moment, etc., are conserved. Similarly, imposing charge conservation along every row of sites parallel to the $\hat{x}$-axis in a 3D lattice leads to the conservation of all multipole moments in the $yz$-plane, e.g., $P_y$, $P_z$, $Q_{yy}$, $Q_{yz}$, $Q_{zz}$, etc.
Furthermore, we can impose subsystem charge conservation along two different families of subsystems simultaneously, e.g., rows and columns in 2D. For a two-dimensional lattice this leads to a conservation of both components of the dipole moment as well as all higher multipole moments diagonal in either $x$ or $y$ coordinates. However, these subsystem symmetries do not guarantee conservation of all multipole moments with components along the subsystem. In particular, conservation of the off-diagonal component of the quadrupole moment $Q_{xy}$ is not achieved by imposing charge conservation along rows parallel to $\hat{x}$ and columns parallel to $\hat{y}$. Instead, one would have to impose charge conservation along rows of sites that are perpendicular to the $xy$-plane. In 2D, for example, this translates into requiring charge conservation at each individual site of the lattice, which trivially leads to a conservation of *all* multipole moments of such system. For the majority of this article we will be focused on the simple case of 1D and 2D systems having subsystem symmetries along the rows and/or columns.
After that brief discussion let us develop an LSM-type theorem for systems with subsystem symmetry. First, let us consider a periodic $L_x\times L_y$ rectangular lattice with a Hamiltonian $H$ that conserves $U(1)$ charge along every row and every column of the lattice, i.e., we have $[H,U_{0,x_0}(\alpha)]=[H,U_{0,y_0}(\alpha)]=0$ for every $x_0$, $y_0,$ and $\alpha$. Particle and dipole conserving Hamiltonians satisfying these criteria can be built from subsystem-symmetric ring-exchange type terms[@dubinkinmaymannhughes2019; @xu2007; @pretko2019emergent].
Now, consider the following twist operator acting along a single column of sites with fixed coordinate $x=x_0$: $$U_{Y,x_0}=\exp\left(\frac{2\pi i}{L_y}\sum_{y}y\hat{n}_{x_0,y}\right).
\label{eqn:1d_multipole_twist}$$ For a periodic lattice and a dipole-conserving Hamiltonian built from local ring-exchange type interactions (say, of fermions) such as $$\begin{split}
H=-t_{\square}\sum_{\textbf{r}}\left(c^\dagger_{\textbf{r}+\hat{x}}c^{\ }_{\textbf{r}} c^{\dagger}_{\textbf{r}+\hat{y}}c^{\ }_{\textbf{r}+\hat{x}+\hat{y}}+h.c.\right),
\end{split}
\label{eqn:ring-exchange}$$ we can calculate the energy difference between the ground state $\vert\Psi_0\rangle$, and a twisted “variational" state $U_{Y,x_0}\vert\Psi_0\rangle$: $$\begin{split}
&\langle \Psi_0|U_{Y,x_0}^{-1} H U_{Y,x_0} - H|\Psi_0 \rangle\\
=&-t_{\square}\bigg[\left(\text{e}^{2\pi i/N_y}-1\right)\sum_{y}\langle c^\dagger_{x_0+a,y}c_{x_0,y}c^\dagger_{x_0,y+a}c_{x_0+a,y+a}\rangle\\
&+\left(\text{e}^{-2\pi i/N_y}-1\right)\sum_{y}\langle c^\dagger_{x_0,y}c_{x_0-a,y}c^\dagger_{x_0-a,y+a}c_{x_0,y+a}\rangle\bigg]\\
&+h.c.
\end{split}
\label{eqn:ham_difference}$$ We can follow analogous calculations to Refs. by expanding the exponents in the powers of $1/N_y,$ and assuming that the ground state preserves at least one of the reflection symmetries. The constant term in the Taylor expansion of the exponential factors immediately cancels, but we we need to check that the next order term also vanishes so that the first non-vanishing term is actually the second-order term from the exponential, which would imply that the energy difference is ultimately of the order $O(1/N_y)$. To see this, after expanding the exponential, consider a pair of plaquettes related by mirror $\hat{M}_y: y\to -y$ and consider the following sum of two of their ring-exchange terms from Eq. \[eqn:ham\_difference\]: $$\begin{split}
\frac{2\pi i}{N_y}&\langle c^\dagger_{x_0+a,y}c_{x_0,y}c^\dagger_{x_0,y+a}c_{x_0+a,y+a}\rangle\\
&-\frac{2\pi i}{N_y}\langle c^\dagger_{x_0+a,-y}c_{x_0,-y}c^\dagger_{x_0,-y-a} c_{x_0+a,-y-a}\rangle.
\end{split}
\label{eqn:sum_reflection}$$ We note that the second term, having an opposite sign, comes from the hermitian conjugate part of the overall Hamiltonian (\[eqn:ring-exchange\]). If the ground state is an eigenstate of the reflection symmetry, $\hat{M}_y\ket{\Psi_0}=\pm\ket{\Psi_0}$, we can rewrite the second term as: $$\begin{split}
\langle&\Psi_0|c^\dagger_{x_0+a,-y}c_{x_0,-y}c^\dagger_{x_0,-y-a} c_{x_0+a,-y-a}|\Psi_0\rangle \\
&=\langle\Psi_0|\hat{M}^{-1}_y c^\dagger_{x_0+a,-y}c_{x_0,-y}c^\dagger_{x_0,-y-a} c_{x_0+a,-y-a}\hat{M}_y|\Psi_0\rangle\\
&=\langle\Psi_0| c^\dagger_{x_0+a,y}c_{x_0,y}c^\dagger_{x_0,y+a}c_{x_0+a,y+a}|\Psi_0\rangle,
\end{split}$$ and so we see that the sum in Eq. \[eqn:sum\_reflection\] exactly vanishes. The same analysis is applicable to every other pair of plaquettes that are related by $\hat{M}_y,$ and so we conclude that the first non-vanishing term in Eq. \[eqn:ham\_difference\] is of the order $O(1/N_y)$. Therefore, in the thermodynamic limit, where $N_y\to\infty$, the state $\ket{\tilde{\Psi}_0}=U_{Y,x_0}\ket{\Psi_0}$ has either exactly the same energy as the ground state $\ket{\Psi_0},$ or it is an excited state with an infinitesimally small energy. Using this result, if we can now show that $\ket{\Psi_0}$ and $\ket{\tilde{\Psi}_0}$ are orthogonal, then the system must necessarily be gapless (or at least have ground state degeneracy). To this end, let us assume that the ground state does not spontaneously break the translational symmetry, so $\ket{\Psi_0}$ will be an eigenstate of the translation operator $T_x$. Using $T_x\ket{\Psi_0}=\text{e}^{i P_{x0}}\ket{\Psi_0}$ it follows that $$\begin{split}
T_x\ket{\tilde{\Psi}_0}&=T_xU_{Y,x_0}T_x^{-1}T_x\ket{\Psi_0}\\
&=\text{e}^{iP_{x0}+\frac{2\pi i}{L_y}\sum_{y=a}^{L_y}y(\hat{n}_{x_0+a,y}-\hat{n}_{x_0,y})}\ket{\tilde{\Psi}_0}.
\end{split}$$ With the subsystem charge conservation in place, we can introduce a subsystem polarization associated with a subsystem $\mathfrak{s}$ as follows: $$\mathcal{P}^j_{\mathfrak{s}}=\frac{e}{2\pi}\text{Im}\log\langle\Psi_0|U_{j,\mathfrak{s}}|\Psi_0\rangle
\label{eqn:subsystem_pol}$$ where $$U_{j,\mathfrak{s}}=\exp\left(\frac{2\pi i}{L_j}\sum_{\textbf{r}\in \mathfrak{s}}x_j \hat{n}_{\textbf{r}}\right),$$ and where $j=x,y.$ Now the condition on the ground state momentum shift to be an integer times $2\pi$ can be understood as a condition on the polarizations of two neighboring subsystems: $$\begin{split}
\text{e}^{\frac{2\pi i}{L_y}\sum_{y=a}^{L_y}y(\hat{n}_{x_0+a,y}-\hat{n}_{x_0,y})}\ket{\tilde{\Psi}_0}&=\ket{\tilde{\Psi}_0}\\
\Leftrightarrow \tfrac{1}{e}\left(\mathcal{P}^y_{x=x_0+a}\right.&-\left.\mathcal{P}^y_{x=x_0}\right)\in\mathbb{Z},
\end{split}
\label{eqn:subs_pol_diff}$$ where $\mathcal{P}^y_{x=x_0}$ is the polarization (\[eqn:subsystem\_pol\]) along the column with fixed coordinate $x=x_0$. In other words, the $\hat{x}$ lattice derivative of the $y$-polarization must be an integer. In general, since we can translate by any number of lattice constants in the $x$-direction, each column must have $\mathcal{P}^{y}_{x=x_0}$ that differ from each other at most by an integer. While this may seem at odds with translation symmetry, the integer ambiguity in the polarization for periodic systems implies that the subsystem polarizations can differ by an integer (or more generally can differ by a polarization quantum) and still maintain translation symmetry. Now, recall the statement from the previous subsection, that, in order for dipole-conserving system on an $N\times N$ lattice to support a gap, we require the total polarization to vanish. Combining this result with Eq. \[eqn:subs\_pol\_diff\] we find that the polarization of each subsystem can take only the following possible of values if the system is gapped: $$\mathcal{P}^j_{\mathfrak{s}}=\frac{n^j_{\mathfrak{s}}}{N}e\mod e,\ n^j_{\mathfrak{s}}\in\mathbb{Z}.$$
Next, we can derive a condition using the fact that the ground state $\ket{\Psi_0}$ will also be an eigenstate of the translation operator $T_y$: $T_y\ket{\Psi_0}=\text{e}^{i P_{y0}}\ket{\Psi_0}.$ Hence for the twisted state $\ket{\tilde{\Psi}_0}$ we find: $$T_y\ket{\tilde{\Psi}_0}=T_yU_{Y,x_0}T_y^{-1}T_y\ket{\Psi_0}=\text{e}^{iP_{y0}+\frac{2\pi i}{N_y}\sum_{y=1}^{N_y}\hat{n}_{x_0,y}}\ket{\tilde{\Psi}_0}.$$ Thus, $\ket{\tilde{\Psi}_0}$ is an eigenstate of $T_y$ with eigenvalue $\exp(i P_{y0} + 2\pi i\nu^{x_0}),$ where $\nu^{x_0}$ is the filling factor of the subsystem $x=x_0$. Hence, the states $\ket{\tilde{\Psi}_0}$ and $\ket{\Psi_0}$ are orthogonal unless each column has integer filling $\nu^{x_0}$. So the system will either have degenerate ground states or gapless excitations if $\nu^{x_0} \notin \mathbb{Z}$ for any column. We can go back and repeat this analysis for subsystem symmetries at fixed $y_0.$ We will find that $T_x$ will require that the subsystem filling factors $\nu^{y_0}$ must all be integers for the ground state to be gapped/non-degenerate. Analogously, $T_y$ will require that the subsystem polarizations $\mathcal{P}^{x}_{y=y_0}$ must all be the same up to an integer (polarization quantum) for the ground state to be gapped/non-degenerate. Although we have chosen a specific Hamiltonian and imposed reflection symmetry to illustrate that the twisted state has low energy, we expect that the result is much more general[@tasaki2018lieb], just as the original LSM result is.
One consequence of these results is that the regular LSM theorem derived for systems with global $U(1)$ charge conservation symmetry[@oshikawa2000A] must be satisfied along every subsystem individually. In the original case, the LSM theorem states that for the ground state to be unique, the particle number per unit cell calculated across the whole lattice must be an integer. Our arguments invoking subsystem symmetry and translation can be straightforwardly applied to $d$-dimensional Hamiltonians conserving total $U(1)$ charge across $n$-dimensional subsystems. We can use this logic to formulate a generalization of the LSM theorem to classes of Hamiltonians that posses not only global $U(1)$ symmetry but also subsystem $U(1)$ symmetries:
*Consider a $d$-dimensional periodic lattice with a short-range Hamiltonian that respects $U(1)$ subsystem charge conservation across a family of parallel $n$-dimensional subspaces. Assume that the ground state does not spontaneously break translational symmetry along the subsystem. For an arbitrary lattice there is at least one low-energy state degenerate, or infinitesimally close in energy, to the ground state, if the particle number per unit cell $\nu^{\mathfrak{s}}$ in any particular $n$-dimensional subsystem $\mathfrak{s}$ is not an integer.* Furthermore, the low-energy state has a crystal momentum in the $j$-direction (associated to a lattice translation operator $T_j$ that leaves the subsystem invariant) differing from the ground state by an amount $\Delta P_j =2\pi\frac{V^{\mathfrak{s}}}{L_j}\nu^{\mathfrak{s}}$, where $V^{\mathfrak{s}}$ is the volume of an individual subsystem, and $L_j$ is the lattice size along the $\hat{x}_j$.
To provide a proof for this general statement, we can use a suitably adapted version of the argument in Ref. . Let us consider a general (bosonic or fermionic) Hamiltonian $H$ defined on a periodic $d$-dimensional lattice that conserves the total $U(1)$ charge across an $n$-dimensional subspace spanned by a collection of $n$ linearly independent primitive lattice vectors $\{\vec{a}_1,...,\vec{a}_n\},$ and their integer linear combinations. We will label this subspace as $\mathfrak{s}$. There are a family of subsystems “parallel" to $\mathfrak{s}$, but for now let us focus on this one subsystem. The $U(1)$ symmetry operator for this subsystem is given by $$U_{0,\mathfrak{s}}(\alpha) = \exp \left(i \alpha \sum_{\textbf{r} \in \mathfrak{s}} \hat{n}_{\textbf{r}}\right),$$ where $\textbf{r}$ is a lattice vector.
The most general lowest-order Hamiltonian that acts on $\mathfrak{s}$ while commuting with $U_{0,\mathfrak{s}}(\alpha)$ takes the following form $$H=\sum_{\textbf{r}_1,\textbf{r}_2,j} J_{\textbf{r}_1,\textbf{r}_2,j}\ c^\dagger_{\textbf{r}_1}c_{\textbf{r}_2}\hat{\mathcal{O}}_j + h.c.
\label{eqn:Hamiltonian}$$ where $J_{\textbf{r}_1,\textbf{r}_2,j}$ is a set of coupling constants, $\textbf{r}_1,\textbf{r}_2 \in \mathfrak{s},$ and $\hat{\mathcal{O}}_j$ are products of particle creation and annihilation operators that have no support on $\mathfrak{s}.$ The $\hat{\mathcal{O}}_j$ may be further constrained by subsystem symmetries for other subspaces $\mathfrak{s}'$, but we will treat them as arbitrary. If we want to gauge the subsystem $U(1)$ symmetry it requires the introduction of an $(n+1)$-dimensional vector-potential $A^{\mathfrak{s}}_i$, $i=0...n$ associated with subsystem $\mathfrak{s}$. Individual terms in the Hamiltonian Eq. (\[eqn:Hamiltonian\]) couple to the lattice gauge field, which modifies the overall Hamiltonian with a subsystem Peierls factor: $$H(A^{\mathfrak{s}})=\sum_{\textbf{r}_1,\textbf{r}_2,j}\left( J_{\textbf{r}_1,\textbf{r}_2,j}\ \text{e}^{iA^{\mathfrak{s}}_{\textbf{r}_1,\textbf{r}_2}}c^\dagger_{\textbf{r}_1}c_{\textbf{r}_2}\hat{\mathcal{O}}_j +h.c.\right).$$
Now we will invoke the momentum counting argument for the subsystem $\mathfrak{s}$ following the presentation of Ref. , with the main difference being that the gauge field is now associated with subsystem $U(1)$ charge conservation instead of a regular global $U(1)$ symmetry. Let us define subsystem magnetic flux quantum insertion operators $\mathcal{F}^{\mathfrak{s}}_j$, $j=1...n$ that adiabatically, over time period $T$, evolves the gauge field $A^{\mathfrak{s}}$ between two configurations that differ by a large gauge transformation performed along the direction spanned by the $\vec{a}_j$-th primitive vector. Adiabaticity of the process is required so that the gap never closes during the time evolution and that, by starting with a ground state $\ket{\Psi_0}$, we are guaranteed to end up with a (possibly different) state $\ket{\tilde{\Psi}_0}$ which lies in the ground state subspace, after the time evolution is finished. Let us define the time evolution of the Hamiltonian as: $$H_j(t)=H\left(A^{\mathfrak{s}}_j=\frac{2\pi}{N_j a}\frac{t}{T}\right),$$ where $N_j$ is the lattice period along the primitive vector $\vec{a}_j$ having lattice constant $a$, and the subsystem lattice gauge field is picked to be uniform in space. The corresponding time evolution operator is then given by $$\mathcal{F}^{\mathfrak{s}}_j=\mathcal{T}\exp\left(-i\int_0^T dt\ H_j(t)\right),$$ where $\mathcal{T}$ denotes time-ordering. The final Hamiltonian $H_j(T)$ differs from the initial one $H_j(0)$ by a large gauge transformation, which can be implemented by the following operator: $$U_{X_j,\mathfrak{s}}=\exp\left(\frac{2\pi i}{L_j}\sum_{\textbf{r}\in \mathfrak{s}}r_j \hat{n}_{\textbf{r}}\right),$$ and one can directly verify that: $$H_j(T)=U_{X_j,\mathfrak{s}}^{-1}H_j(0)U_{X_j,\mathfrak{s}}.$$ We can now define the subsystem flux insertion-removal operator that leaves the Hamiltonian invariant: $$\tilde{\mathcal{F}}^{\mathfrak{s}}_j\equiv U_{X_j,\mathfrak{s}}^{-1}\mathcal{F}^{\mathfrak{s}}_j.$$
Importantly, the initial ground state $\ket{\Psi_0}$, in general, might be different from the state $\ket{\tilde{\Psi}_0}\equiv\tilde{\mathcal{F}}^{\mathfrak{s}}_j\ket{\Psi_0}$ obtained after the subsystem flux insertion and removal procedure. To see this explicitly, let us consider the action of the translation operator $T_{j}$ on the states $\ket{\Psi_0}$ and $\ket{\tilde{\Psi}_0}$ (we note that these translations preserve the subsystem $\mathfrak{s}$). Provided that the translational invariance is not spontaneously broken, we must have: $$T_j\ket{\Psi_0}=\text{e}^{iP_{j0}}\ket{\Psi_0}$$ where $P_{j0}$ is the many-body momentum along $\vec{a}_j$, and is a good quantum number modulo $2\pi$. To see how $T_j$ acts on $\ket{\tilde{\Psi}_0}$ we first note that, since the subsystem gauge field evolution implemented by $\mathcal{F}^{\mathfrak{s}}_j$ does not break translational invariance, we have: $$[\mathcal{F}^{\mathfrak{s}}_j,T_j]=0.$$ However, $T_j$ acts non-trivially on the flux removal unitary: $$T_j U_{X_j,\mathfrak{s}} T_j^{-1}=\text{e}^{-\frac{2\pi i}{N_j}\sum_{\textbf{r}\in\mathfrak{s}}\hat{n}_{\textbf{r}}}U_{X_j,\mathfrak{s}}.$$ Therefore, the $T_j$ eigenvalue of the final state is: $$\begin{split}
T_j\ket{\tilde{\Psi}_0}&=\mathcal{F}^{\mathfrak{s}}_jT_j U_{X_j,\mathfrak{s}}T_j^{-1}T_j\ket{\Psi_0}\\
&=\exp\left(iP_{j0}+2\pi i\frac{V^{\mathfrak{s}}}{N_j}\nu^{\mathfrak{s}}\right)\ket{\tilde{\Psi}_0},
\end{split}$$ where $V^{\mathfrak{s}}$ is the total number of unit cells inside the subsystem $\mathfrak{s},$ and $\nu^{\mathfrak{s}}$ is the filling fraction of said subsystem $\nu^{\mathfrak{s}}=\frac{\mathcal{N}^{\mathfrak{s}}}{V^{\mathfrak{s}}}$, where $\mathcal{N}^{\mathfrak{s}}$ is the total particle number in $\mathfrak{s}$. We thus see that the momentum along $\vec{a}_j$ of the two states $\ket{\Psi_0}$ and $\ket{\tilde{\Psi}_0}$ differs by: $$\Delta P_{j}=2\pi \frac{V^{\mathfrak{s}}}{N_j}\nu^{\mathfrak{s}}.$$ Hence, whenever $\frac{V^{\mathfrak{s}}}{N_j}\nu^{\mathfrak{s}}$ is not an integer, the two states $\ket{\Psi_0}$ and $\ket{\tilde{\Psi_0}}$ must be orthogonal to each other.
The same argument can be applied to the flux insertion-removal procedure for every direction along the subsystem $\mathfrak{s}$. Therefore, for the ground state to be unique, we must require that: $$\frac{V^{\mathfrak{s}}}{N_j}\nu^{\mathfrak{s}}\in\mathbb{Z},\ j=1...n.
\label{eqn:conditions}$$ Since each $N_j$ is a divisor of $V^{\mathfrak{s}}$ the ratios $V^{\mathfrak{s}}/N_j$ are all integers. However, for this set of conditions to be satisfied for arbitrary lattice sizes, we must require $\nu^{\mathfrak{s}}$ itself to be integer. For example, in the case where all $N_i$ are co-prime with each other, the only way to satisfy all of the conditions (\[eqn:conditions\]) is to require that $$\nu^{\mathfrak{s}}\in\mathbb{Z}.$$ We thereby arrive at the theorem stated at the end of the Section \[sec:subs\_LSM\].
We can also provide a lower bound on the ground state degeneracy in the case when $\nu^{\mathfrak{s}}\notin\mathbb{Z}$. Assuming that $\frac{V^{\mathfrak{s}}}{N_j}\nu^{\mathfrak{s}}=\frac{p^{\mathfrak{s}}_j}{q^{\mathfrak{s}}_j}$, where the pair of integers $p^{\mathfrak{s}}_j$ and $q^{\mathfrak{s}}_j$ are co-prime for all $j=1...n$: $$GSD\geq \prod_{j=1}^n q^{\mathfrak{s}}_j.$$ If we now consider subsystem charge conservation along $m$ subsystems $\mathfrak{s}_i$, $i=1,\ldots m,$ which are all translationally invariant in the $j$-th direction (e.g., parallel rows), we can combine different twist operators $U_{X_j,\mathfrak{s_i}}$, $i=1,...,m$ to generate low-lying states where the many-body momentum is shifted when compared to the ground state by: $$\Delta P_{j}=2\pi \frac{V^{\mathfrak{s}}}{N_j}\sum_{i=1}^m n_i\nu^{\mathfrak{s_i}},$$ where $n_i$ are integers and $V^{\mathfrak{s}}$ is the total number of unit cells in every subsystem. The total number of inequivalent values (mod $2\pi$) that $\Delta P_{j}$ can take is the least common multiple of corresponding $q^{\mathfrak{s}_i}_j$ integers. Thus, the ground state degeneracy associated with translations $T_j$ is bounded below by $$GSD_j\ge \text{lcm}\left(q^{\mathfrak{s}_1}_j,q^{\mathfrak{s}_2}_j,...,q^{\mathfrak{s}_m}_j\right).$$ Hence, the total ground state degeneracy for a $d$-dimensional lattice with $m$ subsystem symmetries is bounded below by: $$GSD\ge \prod_{j=1}^d \text{lcm}\left(q^{\mathfrak{s}_1}_j,q^{\mathfrak{s}_2}_j,...,q^{\mathfrak{s}_m}_j\right)$$ where we set $q^{\mathfrak{s}_i}_j=1$ if a particular subsystem $\mathfrak{s}_i$ is not left invariant under the action of $T_j$.
To give an example, imagine a two-dimensional lattice where, on top of the regular global $U(1)$ symmetry we impose independent subsystem $U(1)$ symmetries along two rows $\mathfrak{r}_1$ and $\mathfrak{r}_2$ defined by the equations $y=y_1$ and $y=y_2$ respectively. Both $\mathfrak{r}_1$ and $\mathfrak{r}_2$ are invariant under the action of $T_x$ operator. As an example, let us now choose the fillings of these two rows to satisfy: $$\frac{V^{\mathfrak{r}_1}}{N_x}\nu^{\mathfrak{r}_1}=\frac12\equiv\frac{1}{q_x^{\mathfrak{r}_1}},\quad \frac{V^{\mathfrak{r}_2}}{N_x}\nu^{\mathfrak{r}_2}=\frac14\equiv\frac{1}{q_x^{\mathfrak{r}_2}}.$$ Applying the unitary operator $U_{X,\mathfrak{r}_1}$, we can conclude that there are at least two states in the ground state subspace which are eigenstates of the translation operator $T_x$ with eigenvalues $\text{e}^{iP_{x0}}$ and $\text{e}^{iP_{x0}+i\pi}$. Similarly, applying unitary $U_{X,\mathfrak{r}_2}$ we find at least four translationally invariant states with eigenvalues $\text{e}^{iP_{x0}}$, $\text{e}^{iP_{x0}+i\pi/2}$, $\text{e}^{iP_{x0}+i\pi}$, and $\text{e}^{iP_{x0}+i3\pi/2}.$ Since two of these values have already appeared when we used $U_{X,\mathfrak{r}_1}$, and we cannot easily distinguish two twisted states with the same momentum, we conclude that the total ground state degeneracy is at least $4=\text{lcm}(q_x^{\mathfrak{r}_1},q_x^{\mathfrak{r}_2})$.
As another example let us consider a pair of subsystem symmetries imposed along two columns $\mathfrak{c}_1$ and $\mathfrak{c}_2$ which are defined by the equations $x=x_1$ and $x=x_2$ with fillings $$\frac{V^{\mathfrak{c}_1}}{N_y}\nu^{\mathfrak{c}_1}=\frac12\equiv\frac{1}{q_y^{\mathfrak{c}_1}},\quad \frac{V^{\mathfrak{c}_2}}{N_y}\nu^{\mathfrak{c}_2}=\frac13\equiv\frac{1}{q_y^{\mathfrak{c}_2}}.$$ Analogous to the previous paragraph, these subsystems are invariant under $T_y,$ and acting on the ground state with unitary operators $U_{Y,\mathfrak{c}_1}$ and $U_{Y,\mathfrak{c}_2}$ generates low-lying states with translation eigenvalues $\text{e}^{iP_{y0}}$ and $\text{e}^{iP_{y0}+i\pi}$ for the first operator, and $\text{e}^{iP_{y0}}$, $\text{e}^{iP_{y0}+i2\pi/3}$, and $\text{e}^{iP_{y0}+i4\pi/3}$ for the second one. Additionally, we can combine $U_{Y,\mathfrak{c}_1}$ with $U_{Y,\mathfrak{c}_2}$ to obtain a low-energy state with the translation eigenvalue equal to $\text{e}^{iP_{y0}+i\pi/3}$ and $\text{e}^{iP_{y0}+i5\pi/3}$. Thus, we conclude that the total number of low-lying states that can be generated by the column twist operators is $6=\text{lcm}(q_y^{\mathfrak{c}_1},q_y^{\mathfrak{c}_2})$.
For a two-dimensional system that is translationally invariant simultaneously along both $\hat{x}$ and $\hat{y}$, we expect the fillings of all rows to be the same if the ground state does not spontaneously break translation symmetry, therefore $q_x=q^{\mathfrak{r}_1}_{x}=q^{\mathfrak{r}_2}_{x}=...$, and so the ground state degeneracy associated with translations $T_x$ is $$GSD_x\ge \text{lcm}\left(q^{\mathfrak{r}_1}_x,q^{\mathfrak{r}_2}_x,...\right)=q_x.$$ Similarly, the fillings of all columns must also be the same giving us $q_y=q^{\mathfrak{c}_1}_{y}=q^{\mathfrak{c}_2}_{y}=...$, leading to the ground state degeneracy associated with translations along $\hat{y}$ to be $$GSD_y\ge \text{lcm}\left(q^{\mathfrak{c}_1}_y,q^{\mathfrak{c}_2}_y,...\right)=q_y.$$ The total ground state degeneracy of such system is then bounded from below by the product of the two factors: $$GSD\ge q_x q_y.$$
Application: Plateaus in Magnetization and Magnetization Gradients {#sec:plateau}
==================================================================
Let us take these concepts and apply them to spin systems with an aim toward making physical predictions. In the work of Oshikawa, Yamanaka, and Affleck[@oshikawa1997mag], bosonic spin counterparts of the twist operators (\[eqn:twist\_op\]) were successfully used to derive conditions for the appearance of magnetization plateaus as a function of applied external magnetic field in, e.g., spin chains. Here we will define bosonic spin counterparts of the multipole twist operators Eqs. (\[eqn:multipole\_twist\]), (\[eqn:1d\_multipole\_twist\]) to derive conditions in spin ladder systems for the appearance of plateaus of the *gradient* of magnetization as a function of an applied magnetic field *gradient* placed across the ladder. Explicitly, we imagine tuning the magnetic field so that the system is on a conventional magnetization plateau, and then test how the gradient of the magnetization responds to a magnetic field gradient around a uniform background field. We will want to distinguish between two cases: (i) the system has a non-constant magnetization gradient response, or (ii) the system exhibits a plateau in the magnetization gradient. These physical phenomena are closely related to the recent work on dipole insulators[@dubinkinmaymannhughes2019]. In the language of Ref. , since the system is tuned to a conventional magnetization plateau, the analogous charge system would be a charge insulator. However, case (i) would be a charge insulator but a dipole metal, while (ii) would be both a charge insulator and a dipole insulator.
To illustrate these two possibilities we will consider systems with axial spin rotation symmetry along subspaces, i.e., models with $U(1)$ subsystem symmetry corresponding to a conservation of the total $S^z$ along each subspace. A unitary operator corresponding to a $U(1)$ subsystem symmetry associated to a particular subsystem $\mathfrak{s}$ reads: $$U_{0,\mathfrak{s}}(\alpha) = \exp\left(i\alpha \sum_{\textbf{r}\in\mathfrak{s}} \hat{S}^{z}_{\textbf{r}}\right).$$ This operator rotates all spins belonging to $\mathfrak{s}$ around the $z$-axis by the same amount. The corresponding conserved quantity is the total $S^z$ magnetization on $\mathfrak{s}$.
One-dimensional Spin Ladder Model
---------------------------------
As an explicit test system let us first consider a two-leg spin-$S$ ladder that is stretched along the $\hat{x}$-axis with periodic boundary conditions in the $x$-direction. We will also enforce two $U(1)$ subsystem symmetries, one of which implies conservation of the total magnetization on the top leg (which we label with ‘$\uparrow$’), and the other which implies conservation of the total magnetization of the bottom leg (which we label with ‘$\downarrow$’). Let us assume that the system’s ground state does not break translational symmetry, and has a fixed total magnetization for some range of values of an external magnetic field $h_0 < B^z < h_1$, i.e., the state of the system is at a magnetization plateau. Applying the conventional magnetization plateau argument[@oshikawa1997mag] to a two-leg ladder spin system we find the magnetization *per spin* in a two-leg ladder of the size $L_x\times L_y = N_x a\times 2a$: $$M^z\equiv\frac{1}{2N_x}\sum_{x=1}^{N_x}(S^z_{x,\uparrow}+S^z_{x,\downarrow}) \label{eq:magperspin}$$ takes half-integer values, i.e., $M^z=0,\pm \frac12,\pm 1,$ etc.
To preserve the subsystem symmetries we can build a Hamiltonian from spin ring-exchange terms, e.g., nearest neighbor ring exchanges: $$H=J_{\square}\sum_{x}\left(S^+_{x,\uparrow}S^-_{x,\downarrow}S^+_{x+1,\downarrow}S^-_{x+1,\uparrow}+h.c.\right).
\label{eqn:ham_bos_ring-exchange}$$ Similar to the ring-exchange model studied in the previous section, where such terms tunneled charge dipoles, here they can be interpreted as tunneling terms for magnetic quadrupole moments (spins are already magnetic dipoles so separating opposite spins by a distance to create a “dipole of spins" creates a magnetic quadrupole). Now, consider the following unitary twist operator acting along one of the legs of the ladder: $$U_{X,\uparrow}=\exp\left(\frac{2\pi i}{L_x}\sum_{x}x S^z_{x,\uparrow}\right).
\label{eqn:magnetic_unitary}$$ Under the action of the operator $U_{X,\uparrow}$ each term in $H$ is modified as: $$\begin{split}
U^{-1}_{X,\uparrow}S^+_{x,\uparrow}S^-_{x,\downarrow}&S^+_{x+1,\downarrow}S^-_{x+1,\uparrow}U_{X,\uparrow}\\
&=\text{e}^{\frac{2\pi i}{N_x}}S^+_{x,\uparrow}S^-_{x,\downarrow}S^+_{x+1,\downarrow}S^-_{x+1,\uparrow}.
\end{split}$$ Therefore, we can show that for the ground state $\ket{\Psi_0},$ which we assume preserves translation and reflection symmetry along $\hat{x},$ we have: $$\langle\Psi_0 |U_{X,\uparrow}^{-1} H U_{X,\uparrow} - H |\Psi_0\rangle=O\left(\frac{1}{N_x}\right).$$ And so, similar to the previous section, we see that in the thermodynamic limit $N_x\to \infty$ the state $\ket{\tilde{\Psi}_0}=U_{X,\uparrow}\ket{\Psi_0}$ lies in, or infinitesimally near, the ground state subspace.
Now let us check if $\ket{\tilde{\Psi}_0}=U_{X,\uparrow}\ket{\Psi_0}$ is orthogonal to $\vert\Psi_0\rangle.$ Following logic that should now be apparent, we can compute the commutation relation between the translation operator and $U_{X,\uparrow}$ to find: $$T_x U_{X,\uparrow} T^{-1}_{x}=U_{X,\uparrow} \text{e}^{2\pi i S^z_{1,\uparrow}-\frac{2\pi i}{N_x}\sum_{x=1}^{N_x}S^z_{x,\uparrow}}.$$ Therefore, starting from a ground state $\ket{\Psi_0}$ having a well-defined many-body momentum $T_x\ket{\Psi_{0}}=\text{e}^{iP_{x0}}\ket{\Psi_{0}}$, we find the state $\ket{\tilde{\Psi}_0}$ has the eigenvalue: $$\begin{split}
T_x\ket{\tilde{\Psi}_0}&=T_xU_{X,\uparrow}T_x^{-1}T_x\ket{\Psi_0}\\
&=\text{e}^{iP_{x0}+2\pi i S^z_{1,\uparrow}-\frac{2\pi i}{N_x}\sum_{x=1}^{N_x}S^z_{x,\uparrow}}\ket{\tilde{\Psi}_0}.
\end{split}$$ A notable difference from the previous section is the appearance of the extra term $2\pi S_{1,\uparrow}$ in the phase factor which can be integer or half-integer depending on the spin model of interest. From this analysis we conclude that the two states are orthogonal unless $S_{\uparrow}-m_{\uparrow}\in\mathbb{Z},$ where $m_{\uparrow}=\tfrac{1}{N_x}\sum_{x}S^{z}_{x,\uparrow}.$ We can obtain a similar condition by considering the unitary operator $U_{X,\downarrow}$ that acts on the bottom leg. Thus, the ground state can be unique only if the spin minus the average magnetization $m^z_{\uparrow/\downarrow}$ of the top row or bottom row of spins are both integers: $$S_{\uparrow/\downarrow}-m^z_{\uparrow/\downarrow}\in\mathbb{Z}.$$
Let us analyze these conditions in more detail. We can actually make a more physically intuitive statement by noticing that the sum of average magnetizations of both rows must be an integer as well: $$m^z_\uparrow + m^z_\downarrow=2M^z\in \mathbb{Z},$$where we have used the assumption that our system is tuned to a conventional magnetization plateau, and the fact that the magnetization per spin (Eq. \[eq:magperspin\]) must be a multiple of $1/2$ on the plateau. We can rewrite magnetizations on both legs as: $$\begin{split}
m^z_\downarrow=&2M^z-m^z_\uparrow,\ m^z_\uparrow=n-S,\ n\in\mathbb{Z}.
\end{split}$$ Combining these statements we end up with the following condition for the *magnetization gradient* in the direction transverse to the legs of the ladder: $$\Delta_y m^z \equiv (m^z_\uparrow - m^z_\downarrow)=2(n-S-M^z).
\label{eq:MGradDef}$$ In this equation the spin $S$ (in a unit cell on a single leg) is fixed, and since we are tuned to a magnetization plateau $M^z$ is a multiple of $1/2.$ Thus, we expect the magnetization gradient to have plateaus at only even or only odd integer values where the parity is determined by whether the sum of the total spin $S$ and magnetization $M^z$ is integer, or half-integer respectively. Alternatively, since the total magnetization of the ladder is vanishing, we can recast the magnetization gradient as a magnetic quadrupole moment $Q^{M}_{yz}$ where $z$-oriented magnetizations are separated along the $y$-direction. Our results imply that the system has a plateau of $Q^{M}_{yz}$ as a function of magnetic field gradient. If the total magnetization were on a plateau, but non-vanishing then the conversion to a magnetic quadrupole moment would depend on our choice of coordinate origin. In all of the examples here the total magnetization vanishes so this issue does not arise.
We corroborate our results using the numerics presented in Fig. \[fig:plateaus-noplateaus\]. In this figure we compare the magnetic responses of two types of spin-1/2 ladders. In Fig. \[fig:plateaus-noplateaus\]a,c we show results for a two-leg spin-ladder with nearest neighbor XY couplings, while in Fig. \[fig:plateaus-noplateaus\]b,d we show results for a two-leg spin ladder with ring-exchange terms (c.f. Eq. \[eqn:ham\_bos\_ring-exchange\]). For each system we first show their response to a uniform magnetic field. For the XY ladder (Fig. \[fig:plateaus-noplateaus\]a) we see magnetization plateaus at $M^z=0, \pm 1/2$ which matches the expected results since $2(S-M^z)\in\mathbb{Z}$ is the condition for a plateau for a two-leg system. We also find that the ring exchange model exhibits magnetization plateaus at $M^z=0,\pm 1/2$ (see Fig. \[fig:plateaus-noplateaus\]b), in accordance to our expectations, since, as was mentioned above, for a two-leg ladder we must have $2(S-M^z)\in\mathbb{Z}$. In Fig. \[fig:plateaus-noplateaus\]a we also overlayed a dashed red line showing the magnetization response of a single, nearest-neighbor spin chain coupled via $(2S^x_iS^x_{i+1}+S^y_iS^y_{i+1})$ interactions. These interactions explicitly break the axial $U(1)$ symmetry corresponding to spin rotations around the $\hat{z}$ axis, and hence the system does not exhibit discrete magnetization plateaus, but instead smoothly interpolates between the two configurations where the average magnetization saturates. We chose to compare the XY ladder with an XY spin chain having broken $U(1)$ spin symmetry to make an analogy to the comparison between Figs. \[fig:plateaus-noplateaus\]c,d between the XY ladder and ring-exchange ladder, where the former has broken $U(1)$ subsystem symmetry.
Now we want to examine the magnetization gradient response of the XY ladder and the ring-exchange ladder. Let us consider applying a magnetization gradient centered around zero uniform applied magnetic field, i.e., we apply a $B^z=+h$ to the top leg of the ladder and $B^z=-h$ to the bottom one. The total magnetization of both systems stays at a magnetization plateau with $M^z=0$. For the XY ladder that respects only a global axial $U(1)$ symmetry, but not a subsystem $U(1)$, we see a smooth interpolation of the magnetization *gradient* between the saturation points at $\Delta_y m^z=-1$ and $\Delta_y m^z=+1$. This is quite similar to the behavior of the *magnetization* of the XY spin chain shown in Fig. \[fig:plateaus-noplateaus\]a that does not respect the global axial $U(1)$ symmetry.
For the ring exchange model, which respects axial subsystem $U(1)$ symmetry, we find that $\Delta_y m^z$ exhibits a series of plateaus. The two most stable ones are located exactly at $\Delta_y m^z=\pm 1$ as we expect from Eq. \[eq:MGradDef\]. In our numerical simulations, we also see a plateau at $\Delta_y m^z=0$, however, as we show in the inset plot in Fig. \[fig:plateaus-noplateaus\]d, this plateau is shrinking rapidly as we increase the system size, and it is not clear if it will survive or not in the thermodynamic limit.
at (0,0) [![Magnetization **(a, b)** and magnetization gradient **(c, d)** responses of a two-leg spin-1/2 ladder to an applied external magnetic field $B^z$ and magnetic field gradient $\Delta_y B^z$ respectively. In plots **(a)** and **(c)**, the two-leg ladder is coupled via spin-anisotropic XY interactions that do not preserve subsystem symmetry, while the ladder in plots **(b)** and **(d)** is coupled via ring-exchange interactions that have subsystem symmetry. Additionally, for we include the magnetization response to a constant $B^z$ of a chain that does not respect *global* $U(1)$ $S^z$ rotation symmetry, which is depicted as a dashed red line in **(a)**. We superimposed numerical data for ladders with 4, 6, 8, 10, and 12 rungs. We clearly see that both magnetization and its gradient experience plateaus for the ring-exchange Hamiltonian, while there are no plateaus of magnetization gradient in the data for the XY-coupled ladder. In the magnetization gradient data for the ring-exchange model we see a small plateau at $\Delta_y m^z=0$ that monotonically shrinks with increasing system size as shown by the inset plot. It is possible that this plateau will not survive the thermodynamic limit.[]{data-label="fig:plateaus-noplateaus"}](mplh2.pdf "fig:"){width="21.00000%"}]{}; at (0.5,2.3) [$M^z$]{}; at (3.5,1.4) [$ B^z$]{}; at (-0.05,0.15) ; at (-0.05,1.2) ; at (-0.05,2.2) ; at (0.4,1) ; at (1.1,1) ; at (1.9,1) ; at (2.65,1) ; at (3.4,1) ; at (0.2,2.6) ;
at (0,0) [![Magnetization **(a, b)** and magnetization gradient **(c, d)** responses of a two-leg spin-1/2 ladder to an applied external magnetic field $B^z$ and magnetic field gradient $\Delta_y B^z$ respectively. In plots **(a)** and **(c)**, the two-leg ladder is coupled via spin-anisotropic XY interactions that do not preserve subsystem symmetry, while the ladder in plots **(b)** and **(d)** is coupled via ring-exchange interactions that have subsystem symmetry. Additionally, for we include the magnetization response to a constant $B^z$ of a chain that does not respect *global* $U(1)$ $S^z$ rotation symmetry, which is depicted as a dashed red line in **(a)**. We superimposed numerical data for ladders with 4, 6, 8, 10, and 12 rungs. We clearly see that both magnetization and its gradient experience plateaus for the ring-exchange Hamiltonian, while there are no plateaus of magnetization gradient in the data for the XY-coupled ladder. In the magnetization gradient data for the ring-exchange model we see a small plateau at $\Delta_y m^z=0$ that monotonically shrinks with increasing system size as shown by the inset plot. It is possible that this plateau will not survive the thermodynamic limit.[]{data-label="fig:plateaus-noplateaus"}](mpl.pdf "fig:"){width="21.00000%"}]{}; at (0.5,2.3) [$M^z$]{}; at (3.48,1.4) [$ B^z$]{}; at (-0.05,0.15) ; at (-0.05,1.2) ; at (-0.05,2.2) ; at (1,1) ; at (1.95,1) ; at (2.8,1) ; at (0.2,2.6) ;
\
at (0,0) [![Magnetization **(a, b)** and magnetization gradient **(c, d)** responses of a two-leg spin-1/2 ladder to an applied external magnetic field $B^z$ and magnetic field gradient $\Delta_y B^z$ respectively. In plots **(a)** and **(c)**, the two-leg ladder is coupled via spin-anisotropic XY interactions that do not preserve subsystem symmetry, while the ladder in plots **(b)** and **(d)** is coupled via ring-exchange interactions that have subsystem symmetry. Additionally, for we include the magnetization response to a constant $B^z$ of a chain that does not respect *global* $U(1)$ $S^z$ rotation symmetry, which is depicted as a dashed red line in **(a)**. We superimposed numerical data for ladders with 4, 6, 8, 10, and 12 rungs. We clearly see that both magnetization and its gradient experience plateaus for the ring-exchange Hamiltonian, while there are no plateaus of magnetization gradient in the data for the XY-coupled ladder. In the magnetization gradient data for the ring-exchange model we see a small plateau at $\Delta_y m^z=0$ that monotonically shrinks with increasing system size as shown by the inset plot. It is possible that this plateau will not survive the thermodynamic limit.[]{data-label="fig:plateaus-noplateaus"}](mgplh.pdf "fig:"){width="21.00000%"}]{}; at (0.7,2.3) [$\Delta_y m^z$]{}; at (3.3,1.4) [$\Delta_y B^z$]{}; at (-0.05,0.15) ; at (-0.05,1.2) ; at (-0.05,2.2) ; at (1,1) ; at (1.95,1) ; at (2.8,1) ; at (3.65,1) ; at (0.2,2.6) ; (0.43+,1.27+) – (1.75+,1.27+) – (1.75+,2.1+) – (0.43+,2.1+) – (0.43+,1.27+);
at (0,0) [![Magnetization **(a, b)** and magnetization gradient **(c, d)** responses of a two-leg spin-1/2 ladder to an applied external magnetic field $B^z$ and magnetic field gradient $\Delta_y B^z$ respectively. In plots **(a)** and **(c)**, the two-leg ladder is coupled via spin-anisotropic XY interactions that do not preserve subsystem symmetry, while the ladder in plots **(b)** and **(d)** is coupled via ring-exchange interactions that have subsystem symmetry. Additionally, for we include the magnetization response to a constant $B^z$ of a chain that does not respect *global* $U(1)$ $S^z$ rotation symmetry, which is depicted as a dashed red line in **(a)**. We superimposed numerical data for ladders with 4, 6, 8, 10, and 12 rungs. We clearly see that both magnetization and its gradient experience plateaus for the ring-exchange Hamiltonian, while there are no plateaus of magnetization gradient in the data for the XY-coupled ladder. In the magnetization gradient data for the ring-exchange model we see a small plateau at $\Delta_y m^z=0$ that monotonically shrinks with increasing system size as shown by the inset plot. It is possible that this plateau will not survive the thermodynamic limit.[]{data-label="fig:plateaus-noplateaus"}](mgpl.pdf "fig:"){width="21.00000%"}]{}; at (0.7,2.3) [$\Delta_y m^z$]{}; at (3.4,1.4) [$\Delta_y B^z$]{}; at (-0.05,0.15) ; at (-0.05,1.2) ; at (-0.05,2.2) ; at (1,1) ; at (1.95,1) ; at (2.8,1) ; at (3.6,1) ; at (0.2,2.6) ; at (0.4+,1.2+) [![Magnetization **(a, b)** and magnetization gradient **(c, d)** responses of a two-leg spin-1/2 ladder to an applied external magnetic field $B^z$ and magnetic field gradient $\Delta_y B^z$ respectively. In plots **(a)** and **(c)**, the two-leg ladder is coupled via spin-anisotropic XY interactions that do not preserve subsystem symmetry, while the ladder in plots **(b)** and **(d)** is coupled via ring-exchange interactions that have subsystem symmetry. Additionally, for we include the magnetization response to a constant $B^z$ of a chain that does not respect *global* $U(1)$ $S^z$ rotation symmetry, which is depicted as a dashed red line in **(a)**. We superimposed numerical data for ladders with 4, 6, 8, 10, and 12 rungs. We clearly see that both magnetization and its gradient experience plateaus for the ring-exchange Hamiltonian, while there are no plateaus of magnetization gradient in the data for the XY-coupled ladder. In the magnetization gradient data for the ring-exchange model we see a small plateau at $\Delta_y m^z=0$ that monotonically shrinks with increasing system size as shown by the inset plot. It is possible that this plateau will not survive the thermodynamic limit.[]{data-label="fig:plateaus-noplateaus"}](mid_plat.pdf "fig:"){width="7.50000%"}]{}; (0.43+,1.27+) – (1.75+,1.27+) – (1.75+,2.1+) – (0.43+,2.1+) – (0.43+,1.27+); at (0.52+,1.88+) ; at (0.77+,1.63+) ; at (1.07+,1.5+) ; at (1.4+,1.43+) ; at (1.63+,1.64+) ;
at (0.35+,1.97+) ; (0.44+,1.97+) – (0.47+,1.97+); at (0.35+,1.62+) ; (0.44+,1.62+) – (0.47+,1.62+); at (0.35+,1.3+) ;
at (0.0+,1.8+) ; at (0.15+,1.6+) ;
### Ising-Coupled Spin Ladder
As a brief aside, we can further illustrate the physics of magnetization gradient plateaus in one-dimensional ladders by connecting them to the ordinary magnetization plateaus in an effective single spin chain. To see this, consider a spin-1/2 ladder of length $L_x$ in a magnetic field with ring-exchange, Ising, and Zeeman couplings: $$\begin{split}
H = \sum_{x} [(J_{\square}S^+_{x,\uparrow} S^-_{x+1,\uparrow} S^+_{x+1,\downarrow} S^-_{x,\downarrow} + h.c.) + \lambda S^z_{x,\uparrow} S^z_{x,\downarrow}\\ + h_\uparrow S^z_{x,\uparrow} + h_\downarrow S^z_{x,\downarrow}],
\label{eq:XYLadder}
\end{split}$$ where ${\bf {S}}_{x,\uparrow/\downarrow}$ is the spin$-1/2$ operator on rung $x$ and the top ($\uparrow$) or bottom ($\downarrow$) leg. This Hamiltonian commutes with the total magnetization operator of the entire ladder ($\sum_x [S_{x,\uparrow}+ S_{x,\downarrow}]$), and the individual magnetization operators of each leg ($\sum_x S_{x,\uparrow }$ and $\sum_x S_{x,\downarrow }$). The former is a $U(1)$ global symmetry, while the latter are a pair of $U(1)$ subsystem symmetries.
Here, we will be interested in the limit $\lambda \gg J_{\square}, |h_{\uparrow/\downarrow}|$, where every rung on the ladder will be pinned such that $\langle S^z_{x,\uparrow} S^z_{x,\downarrow} \rangle = -\frac{1}{4}$. In this limit, the total magnetization of the system is fixed to be an integer $M^z = 0$, and the system is on a magnetization plateau. There are two configurations that satisfy this constraint: $\langle S^z_{x,\uparrow} \rangle = \pm \frac{1}{2}$, $\langle S^z_{x,\downarrow} \rangle = \mp \frac{1}{2}$. Let us define a new effective spin degree of freedom on each rung, $\tilde{S}_r$ such that $\tilde{S}^z_x = S^z_{x,\uparrow} = -S^z_{x,\downarrow}$. Using this, we can also define: $\tilde{S}^+_x = S^+_{x,\uparrow} S^-_{x+1,\uparrow}$, $\tilde{S}^-_x = S^-_{x,\downarrow} S^+_{x+1,\downarrow}$. Combined with $\tilde{S}^z_x$, these operators satisfy the spin-1/2 algebra, and the spin ladder becomes: $$\begin{split}
H = J_{\square} \sum_{x}\tilde{S}^+_x\tilde{S}^-_{x+1} + (h_\uparrow-h_\downarrow) \tilde{S}^z_x,
\end{split}
\label{eq:XYChain}$$ which is the Hamiltonian for a single $XY$ spin *chain* in an effective magnetic field given by $h_\uparrow-h_\downarrow$.
Now let us consider the response of this system to external magnetic fields. When this system is placed in a uniform physical magnetic field ($h_1 = h_2$), the effective magnetic field vanishes so the system does not develop a magnetization as might be expected. However, if we instead consider a physical magnetic field gradient parallel to the ladder rungs (e.g., $h_1 = - h_2$), the effective magnetic field is non-vanishing and the system can develop an effective magnetization. The key point is that the magnetization of the effective spins in Eq. \[eq:XYChain\] is equal to half the magnetization gradient of the original spin ladder: $$\frac{1}{L_x}\sum_x \tilde{S}^z_x =\frac{1}{2}\frac{1}{L_x}\sum_x [S^z_{x,1} - S^z_{x,2}].$$ In conclusion, the effective magnetic field and magnetization associated to Eq. \[eq:XYChain\] are respectively the physical magnetic field gradient and magnetization gradient of the spin ladder Eq. \[eq:XYLadder\]. A magnetization plateau for the effective spins in Eq. \[eq:XYChain\], is thereby equivalent to a magnetization gradient plateau for the physical spin in Eq. \[eq:XYLadder\].
It is well known that the $XY$ spin chain has magnetization plateaus when the magnetization is equal to $\pm \frac{1}{2}$[@lieb1961]. From this, we can conclude that that the spin ladder Eq. \[eq:XYLadder\] is at a plateau in the gradient of its magnetization when $\Delta_y m^z = \sum_x (S^z_{x,1} - S^z_{x,2})/L_x = 2 \sum_x \tilde{S}^z_x/L_x = \pm 1$. Since this model is also at an ordinary magnetization plateau at $M^z = 0$, this result agrees with Eq. \[eq:MGradDef\]. It is worth noting that this result is only true in the limit $\lambda \gg |h_{\uparrow/\downarrow}|$. In the opposite limit, $|h_{\uparrow/\downarrow}| \gg \lambda, J$, the system will be at ordinary magnetic plateaus where the spins are aligned parallel to the magnetic field.
Two-dimensional Spin Systems
----------------------------
Now let us consider two-dimensional spin models. We will first derive the spin analog of the dipole LSM theorem from Sec. \[sec:dip\_LSM\_A\]. Working on a square periodic $L\times L=N a\times N a$ lattice, we will consider spin Hamiltonians which possess global $U(1)$ symmetry that acts by rotating all spins around the $\hat{z}$-axis by the same amount. The corresponding conserved quantity is the total magnetization $M^z$ of the system. We will additionally impose conservation of two components of the magnetic quadrupole moment $Q^M_{xz}$ and $Q^M_{yz}$ which is the analog of the conservation of the $x$ and $y$ components of the dipole moment for particles whose charge under the global $U(1)$ symmetry is itself a magnetic dipole moment pointing in the $z$-direction.
This setup is entirely analogous to the one considered in Sec. \[sec:dip\_LSM\_A\]. It is natural then to consider Hamiltonians where the lowest-order dynamical terms are built of bosonic spin ring-exchange terms, as in (\[eqn:ham\_bos\_ring-exchange\]). Such systems were recently discussed in the literature[@you2019] where it was shown, that they naturally couple to the background symmetric rank-2 gauge field $A_{xy}$ with a Peierls phase factor. To derive an LSM-type theorem we will briefly recount the argument already discussed in detail in the context of dipole-conserving systems in Sec. \[sec:dip\_LSM\_A\]. We start by adiabatically driving the value of the background field $A_{xy}$ from $0$ to $2\pi/Na^2$ over time period $T$. This evolves the ground state of the system from $\ket{\Psi(0)}$ to $\ket{\Psi(T)}$. As this process is performed uniformly across the lattice, without breaking translational symmetry, $\ket{\Psi(T)}$ will remain an eigenstate of $T_x$ and $T_y$, provided that the translational invariance of the initial Hamiltonian was not spontaneously broken in its ground state $\ket{\Psi(0)}$. Then, we apply the unitary twist operator $$U_{XY}=\exp\left(\frac{2\pi i}{aL}\sum_{\textbf{r}}xy\ S^z_{\textbf{r}}\right)$$ which removes the change in $A_{xy}$ and brings the Hamiltonian back to its original form. The resulting eigenstate $U_{XY}^{-1}\ket{\Psi(T)}$ has an energy infinitesimally close to the ground state in the thermodynamic limit, and it may be different from the original ground state $\ket{\Psi(0)}$. If we consider the commutation relation between translations in the $\hat{x}$-direction and the twist operator, we obtain an additional phase factor: $$T_x U^{-1}_{XY} T^{-1}_{x}=U^{-1}_{XY}\text{e}^{\frac{2\pi i }{L}\sum_{\textbf{r}}y S^z_{\textbf{r}}}e^{-2\pi i\sum^{N}_{y=1} y S^z_{x=1,y}}
\label{eqn:spin_extra_phase}$$ where the first extra factor on the RHS of the equation contains the total magnetic quadrupole moment $Q^{M,tot}_{yz}$ of the system in the exponential. The phase in the second factor can take either integer or half-integer multiples of $2\pi$ depending on whether the spin $S$ is integer or half-integer, and whether the value of $N(N+1)/2$ is even or odd. For eigenvalues of the translation operator $T_x$ to be the same for $U_{XY}^{-1}\ket{\Psi(T)}$ and $\ket{\Psi(0)}$ the extra phase factors appearing in (\[eqn:spin\_extra\_phase\]) must be trivial. This yields the following condition for the uniqueness of the ground state: $$\begin{split}
\frac{N(N+1)}{2}S+\sum_{\textbf{r}}\frac{y S^z_{\textbf{r}}}{Na} \in\mathbb{Z},
\end{split}
\label{eqn:spin_LSM}$$ which is very similar to the condition obtained in Sec. \[sec:dip\_LSM\_A\]. For instance, consider integer $S$. The condition (\[eqn:spin\_LSM\]) then requires that, for the ground state to be unique, the total magnetic quadrupole moment $Q^M_{yz}$ must be an integer. We can repeat this calculation using translation in the $\hat{y}$-direction to derive $$\begin{split}
\frac{N(N+1)}{2}S+\sum_{\textbf{r}}\frac{x S^z_{\textbf{r}}}{Na} \in\mathbb{Z},
\end{split}
\label{eqn:spin_LSM2}$$which gives a similar condition but for $Q^M_{xz}.$
Now, let us move on to spin systems with $U(1)$ subsystem symmetry where the $S^z$ spin component is conserved on rows and columns of the lattice. We can again consider a twist operator that acts along a single column of spins with fixed coordinate $x=x_0$: $$U_{Y,x_0}=\exp\left(\frac{2\pi i}{L_y}\sum_{y}y S^{z}_{x_0,y}\right).$$ For a periodic lattice and a subsystem symmetric Hamiltonian built from local ring-exchange terms such as Eq. (\[eqn:ham\_bos\_ring-exchange\]), we can compare the energy of a twisted state with the original ground state $\ket{\Psi_0}.$ If we assume that $\ket{\Psi_0}$ does not spontaneously break the translational invariance and preserves a reflection symmetry $\hat{M}_y: y\to -y$, similar to (\[eqn:ham\_difference\]) we have: $$\langle\Psi_0 | U_{Y,x_0}^{-1} H U_{Y,x_0} - H |\Psi_0\rangle=O\left(\frac{1}{N_y}\right).$$ Therefore, in the thermodynamic limit, where $N_y\to\infty$, the state $\ket{\tilde{\Psi}_0}=U_{Y,x_0}\ket{\Psi_0}$ has either exactly the same energy as the ground state $\ket{\Psi_0}$, or it is an excited state with an energy infinitesimally close to the ground state.
We now want to see if $\ket{\Psi_0}$ and $\ket{\tilde{\Psi}_0}$ are orthogonal. Assuming that the ground state does not spontaneously break the translational symmetry, i.e., $T_x\ket{\Psi_0}=\text{e}^{iP_{x0}}\ket{\Psi_0}$, we can show that the translation eigenvalue for $\ket{\tilde{\Psi}_0}$ may take a dstinct value: $$\begin{split}
T_x\ket{\tilde{\Psi}_0}&=T_x U_{Y,x_0} T^{-1}_{x}T_x\ket{\Psi_0}\\
&=\text{e}^{iP_{x0}+\frac{2\pi i}{L_y}\sum_{y} y(S^z_{x_0+1,y}-S^z_{x_0,y})}\ket{\tilde{\Psi}_0}.
\end{split}$$ Similar to the subsystem polarization introduced in Sec. \[sec:dip\_LSM\_A\], we introduce an analogous notion for spin systems – a subsystem quadrupole polarization: $$\mathcal{Q}^M_{j z}(\mathfrak{s})=\frac{1}{2\pi}\text{Im}\log\langle\Psi_0|U_{j,\mathfrak{s}}|\Psi_0\rangle,
\label{eqn:subsystem_qpol}$$ where $$U_{j,\mathfrak{s}}=\exp\left(\frac{2\pi i}{L_j}\sum_{\textbf{r}\in \mathfrak{s}}x_j S^z_{\textbf{r}}\right),$$ and where $j=x,y$. Therefore, for the ground state to be unique we must have that the pair of magnetic quadrupole moments $\mathcal{Q}^M_{yz}$ computed along neighboring subsystems must differ by an integer number: $$\begin{split}
\text{e}^{\frac{2\pi i}{L_y}\sum_{y}y(S^z_{x_0+1,y}-S^z_{x_0,y})}\ket{\tilde{\Psi}_0}&=\ket{\tilde{\Psi}_0}\\
\Leftrightarrow \mathcal{Q}^M_{yz}(x=x_0+1)&-\mathcal{Q}^M_{yz}(x=x_0)\in\mathbb{Z},
\end{split}
\label{eqn:spin_subsystem_rel}$$ where $\mathcal{Q}^M_{yz}(x=x_0)$ is the magnetic quadrupolar polarization (\[eqn:subsystem\_qpol\]) along the column with fixed coordinate $x=x_0$. Therefore, for the states $\ket{\Psi_0}$ and $\ket{\tilde{\Psi}_0}$ to have the same eigenvalue of the translation operator $T_x$ we need to require that the difference between the subsystem magnetic $\mathcal{Q}^M_{yz}$ quadrupole moments computed along two the adjacent rows of spins is an integer number. In general, since we can translate by any number of lattice constants in the $x$-direction, each column must have $\mathcal{Q}^M_{yz}$ that differ at most by an integer if we want to preserve translation symmetry and have a unique ground state. Noting that the total magnetic quadrupole moment $\mathcal{Q}^M_{yz}$ should satisfy the Eq. \[eqn:spin\_LSM2\] we can add in the relationship between subsystem quadrupole moments (\[eqn:spin\_subsystem\_rel\]) to see that on a $N\times N$ lattice subsystem quadrupolarization must take the following set of values: $$\mathcal{Q}^M_{yz}(x=x_0)=\frac{n}{N}-\frac{N+1}{2}S,\ n\in\mathbb{Z}$$ and similarly for $\mathcal{Q}^M_{xz}(y=y_0)$ on every column with fixed coordinate $y=y_0$.
Now, considering translations along $\hat{y}$ we find: $$T_y\ket{\tilde{\Psi}_0}=
%T_y U_{Y,x_0} T^{-1}_{y}T_y\ket{\Psi_0}=
\text{e}^{iP_{y0}+2\pi i S^z_{x_0,1}-\frac{2\pi i}{N_y}\sum_{y=1}^{N_y} S^z_{x_0,y}}\ket{\tilde{\Psi}_0},$$ which means that for the ground state $\ket{\Psi_0}$ to be unique we need the average magnetization $m^z_{x=x_0}$ of a single column at $x=x_0$ to satisfy: $(S-m^z_{x=x_0})\in\mathbb{Z}$, with $S$ being the total spin per unit cell of a subsystem. Hence, the states $\ket{\Psi_0}$ and $\ket{\tilde{\Psi}_0}$ are orthogonal unless the average magnetization of each subsystem $\mathfrak{s}$ satisfies $$S-m^z_{\mathfrak{s}}\in\mathbb{Z}.$$
The physical consequences of these results are more subtle than the ladder case. We have found that in order for a system with magnetic quadrupole conservation to have a unique ground state the spin and magnetic quadrupolarization must satisfy an integer constraint. Furthermore, if the system has subsystem spin-rotation symmetry then each subsystem has to be on a magnetization plateau for the ground state to be unique. Thus, in the latter case, if we apply a spatially varying magnetic field that is constant along a family of subsystems, and weak enough not to drive any subsystem off its plateau, then the system will have a constant magnetization plateau response even to a *spatially varying* magnetic field. For example, if we have subsystem symmetry in 2D along rows and columns parallel to $x$ and $y$ respectively, then our system will have a non-varying response to magnetic fields having only $x$ or only $y$ dependence as long as the field applied to any given subsystem is not strong enough to drive it off its plateau. In the former case without subsystem symmetry the system can exhibit a plateau in the magnetic quadrupolarization in the presence of a pure magnetic field gradient, i.e., a non-uniform magnetic field configuration that can have, at most, linear dependence on the spatial coordinates. Both of these possibilities suggest that for magnetic systems tuned to magnetization plateaus there can be a refinement of the magnetic response characterization based on how the system responds to non-uniform fields. Indeed, the systems we studied here can exhibit additional types of magnetic response plateaus when they have unique ground states.
Luttinger-like theorem for dipoles {#sec:luttinger}
==================================
The LSM theorem has been used in non-perturbative arguments supporting Luttinger’s theorem[@oshikawa1997; @oshikawa2000B]. At the heart of these arguments is the connection between the momentum of a low-energy excitation and the particle filling. For a Fermi-liquid this relates the Fermi-surface, where low-energy excitations are created (having momentum of order $k_F$), to the electron filling, even in an interacting system. In this section we will apply similar arguments to show that some systems having $U(1)$ particle number and dipole conservation can support low-energy excitations having momentum determined by the filling of dipoles. Here we will just provide an example, and leave a full discussion, and generalization to higher dimensions to future work.
Let us consider a two-leg fermion ladder model parallel to the $x$-direction (let the lattice constant $a=1$). We will use the Hamiltonian[@dubinkinmaymannhughes2019] $$H=\frac{J}{2}\sum_{i=1}^N(d^{\dagger}_i d^{\phantom{\dagger}}_{i+1}+h.c.)+U\sum_{i=1}^N n_{i\uparrow}n_{i\downarrow},
\label{eqn:dipole_metal}$$ where $\uparrow/\downarrow$ label the two legs of the ladder, $d_i\equiv c^\dagger_{i\downarrow}c_{i\uparrow}$ is a dipole annihilation operator for a dipole parallel to $y$, i.e., along the rungs of the ladder, $c^{\dagger}_{i\downarrow/\uparrow}$ is a fermion creation operator on the lower/upper legs respectively at site $i$, and the $n_{i\downarrow/\uparrow}$ are the fermion density operators on each leg. In Ref. , it was shown that when this system is at half-filling ($N_F=N$), and $U\gg J,$ then the dipole operators effectively become hard-core bosons having onsite anticommutation relations $$\{d^\dagger_i,d_i\}=1,\ \{d^\dagger_i,d^\dagger_i\}=\{d_i,d_i\}=0,
\label{eqn:dipole_anticommute2}$$ and commuting off site. Thus, in this limit this model becomes a hopping model for $y$-oriented dipoles that behave as hardcore bosons. We can identify up-dipoles (down-dipoles) with a configuration where, at a particular unit cell $i$ there is a fermion on the upper leg (lower leg) and no fermion on the lower leg (upper leg). Based on this, we can define the total dipole number operator as $N_D = \sum^N_{i=1} [n_{i\uparrow} - n_{i\downarrow}]$, and the $y$ component of the polarization as $p^y = N_D/N$. It is clear that the Hamiltonian in Eq. \[eqn:dipole\_metal\] conserves the dipole number $N_D$. Since the total fermion number ($N_F = \sum^N_{i=1} [n_{i\uparrow} + n_{i\downarrow}]$) is also conserved, the fermion number on each leg of the ladder ($N_{\uparrow}=\sum^N_{i=1} n_{i\uparrow}$ and $N_{\downarrow}=\sum^N_{i=1}n_{i\downarrow}$) is conserved as well.
To prove a Luttinger-like theorem we want to show that the low-energy modes of this model at some filling of dipoles have momentum related to that dipole filling. Let us take the ground state of the system to be $\ket{\Psi_0}$. We will consider the twisted variational state $\ket{\tilde{\Psi}_0} = \exp(2\pi i\sum^N_{j =1}j n_{j,\uparrow}/N)\ket{\Psi_0}$. A calculation analogous to what we have presented in detail above for the ring exchange model shows that the energy of this state is within $\mathcal{O}(1/N)$ of the ground state energy. We can now calculate the momentum of this state. If $\ket{\Psi_0}$ is an eigenstate of the lattice translation operator $T_x$ with eigenvalue $e^{i P_{x0}}$, then $\ket{\tilde{\Psi}_0}$ will have an eigenvalue $e^{iP_{x0} + 2\pi i\sum^N_{i =1} n_{i,\uparrow}/N}$. Using the relation $\sum^N_{i =1} n_{i,\uparrow}/N = \frac{1}{2N} (N_D + N_F)$, (and also that $N_F=N$ since the fermions are half-filled), there must be a low energy mode with momentum $[P_{x0} + \pi(p^y + 1)]$, where we recall $p_y$ is the $y$-component of the charge polarization. Similarly, if we twist the ground state with the inverse of the operator above we will find another low energy mode with momentum $[P_{x0} - \pi(p^y + 1)]$. We recall that these modes are only guaranteed to be orthogonal to the untwisted ground state if the polarization $p^y$ is not an integer. We will argue below that these points form an analog of a Fermi surface for dipoles with Fermi wavevector $$k^{(dipole)}_F=\pi\left(p^y+1\right).
\label{eqn:Luttinger_dipole}$$
Alternatively, we can derive these results with an explicit solution of this model. If the dipoles are effectively hard-core bosons, this model can be transformed into a spin-1/2 XY model using: $$S^\alpha_i=\frac12 \vec{c}_{i}^{\ \dagger} \sigma^\alpha \vec{c}_{i},\ \text{where}\ \vec{c}_{i}=(c_{i,\uparrow},c_{i\downarrow})^T,
\label{eqn:fermions_to_spin}$$ so that $$d^\dagger_i=2S^+_i,\ d_i=2S^-_i.
\label{eqn:dipoles_to_spin}$$ The resulting spin Hamiltonian is $$H=2J\sum_{i=1}^N \left(S^+_iS^-_{i+1}+S^-_{i}S^+_{i+1}\right).
\label{eqn:heisenberg_model}$$ It is well-known that such an XY model is exactly solvable in 1D via a Jordan-Wigner transformation: $$S^+_i=\text{e}^{i\pi\sum_{j=1}^{i-1}f^\dagger_j f_j}f^\dagger_i,\ S^-_i=\text{e}^{-i\pi\sum_{j=1}^{i-1}f^\dagger_j f_j}f_i,
\label{eqn:jw-map}$$ and the resulting transformed Hamiltonian is: $$H=2J\left(\sum_{i=1}^{N-1}f^\dagger_i f_{i+1}+\text{e}^{i\pi\sum_{j=1}^{N}f^\dagger_j f_j}f^\dagger_N f_1\right)+h.c.
\label{eqn:fermion_ham}$$where $f_{i}$ is the annihilation operator for a Jordan-Wigner fermion on site $i.$
Using these mappings we can identify the low-energy excitations of the dipole model with the Fermi-surface excitations of the Jordan-Wigner fermions. These excitations occur at momentum $\pm k^{(dipole)}_{F}$ which is directly proportional to the density of Jordan-Wigner fermions, and through the mappings above to the density of $y$-dipoles. Precisely we have $2k^{(dipole)}_F=2\pi\nu$, where $\nu$ is the fraction of up-dipoles in the system. Alternatively we can rewrite dipole density as: $$\nu=\frac{N_\uparrow}{N_\uparrow+N_\downarrow}=\frac12\frac{N_\uparrow-N_\downarrow}{N_\uparrow+N_\downarrow}+\frac{1}{2}=\frac12(p^y+1),
\label{eqn:fractional_dipole-polarization}$$ where $p^y$ is the charge polarization in the $y$-direction. Thus we can relate the area enclosed by a Fermi surface in 1D to the polarization of the dipole chain in the transverse direction and we recover Eq. \[eqn:Luttinger\_dipole\].
We expect that results like this can apply beyond one-dimensional ladders. As an example, we could consider a model for a 2D *dipole metal* recently discussed in Ref. . This model is built by stacking dipole ladder models (which are parallel to $\hat{x}$) into the $y$-direction and introducing dipole hopping terms between the nearest-neighbor rungs of two neighboring ladders. Effectively, the model describes a system of free $y$-dipoles that can move across a rectangular lattice. It was shown that this model can be Jordan-Wigner transformed to a fermionic tight-binding model which has a well-defined Fermi surface. This transformation translates number operators for $y$-dipoles into ordinary number operators for the Jordan-Wigner fermions. Hence Luttinger’s theorem for a two-dimensional fermionic model, when translated to a dipole language, once again relates an area enclosed by a Fermi surface to the density of $y$-dipoles in the lattice, or, in other words, to the $\hat{y}$ polarization of the ground state.
We note a possible connection to the recent work in Ref. , where elementary dipole particles having a fixed dipole moment were considered. In comparison, however, the statistics of those particles was taken to be fermionic (in our case they are hard-core bosons), and the interactions between particles were taken into account. It was then argued that this system develops a stable interacting Fermi liquid with a Fermi surface elongated in the direction of the dipole moment. A Luttinger theorem for such fermionic dipoles, which we do not prove here, would also necessarily relate an area enclosed by a Fermi surface to the density of fermionic dipoles, i.e., the polarization of the system in the dipole moment direction.
Conclusions
===========
In this paper we derived several non-perturbative results for dipole-conserving Hamiltonians and their spin counterparts. We provided a generalization of the LSM theorem to multipole-conserving systems, and find that for dipole conserving systems, a unique gapped, translationally invariant ground state is possible only if the bulk polarization is integer (integer filling of dipoles). A rational polarization of $p/q$ implies that there are at least $q$ degenerate ground states. Furthermore, if the system both conserves polarization and has a $U(1)$ charge conservation symmetry along subsystems, a unique gapped, translationally invariant ground state is possible only if the filling in each subystem is an integer. A rational filling implies either a gapless system or a ground state degeneracy. We also provided the spin counterpart of this theorem, that applies to spin systems that have conserved magnetic quadrupole moments and possibly preserve spin-rotation symmetry on subsystems. These systems can experience plateaus in the magnetic response in some types of non-uniform fields. Finally we have also discussed a possible extension of a Luttinger-like theorem to dipole systems.
From these results, we have been able to place strong constraints on the low energy physics of systems having conserved multipole moments. As with the famous results of Lieb, Schultz and Mattis, these results can be used to study strongly correlated systems, where normal perturbative methods fail. Much is still unknown about multipole conserving systems on lattices[@vijay2016; @nandkishore2019; @pai2019scar], and in the continuum [@pretko2018; @gromov2019; @may2019], and our results may prove useful in these contexts. These results also hint at possible exotic gapped phases that have fractional polarization/dipole moment, in analogy to topologically ordered systems having fractional charge. The new types of magnetic response plateaus we predicted may also belong to topological phases, analogous to the Haldane phase in SPT spin chains. Experimentally, our results can be tested in cold-atom systems, where dipole conserving systems can be constructed[@buchler2005; @li2005; @dai2017]. These cold-atom systems may be an interesting place to look for the aforementioned exotic phases. While we have focused primarily on 1D and 2D we expect the results can be extended straightforwardly to higher dimensions, and with a variety of conserved types of multipole moments. Finally, it could prove useful exploring possible connections between our LSM-type theorems and similar results recently acquired in the context of systems with higher-form symmetries[@Kobayashi19].
[**[Note:]{}**]{} During the preparation of this manuscript we became aware of a recent work titled “Lieb-Schultz-Mattis type constraints on Fractonic Matter" by Huan He, Yizhi You, and Abhinav Prem[@yizhiLSM]. Our work has some overlapping concepts and results with this article, but both were carried out independently.
We thank Yizhi for useful discussions, and for pointing us to their recent article. OD and TLH acknowledge support from the US National Science Foundation under grant DMR 1351895-CAR, and the MRSEC program under NSF Award Number DMR-1720633 (SuperSEED)for support. JMM is supported by the National Science Foundation Graduate Research Fellowship Program under Grant No. DGE – 1746047.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Using methods of differential geometry, a discrete analog of the Yang-Mills equations in Minkowski space is constructed. The gauge transformation law in a discrete formulation is given and gauge invariance of discrete Yang-Mills equations is studied. Difference self-dual and anti-self-dual equations with respect to the Lorentz metric are presented.'
address: 'Department of Mathematics, Technical University of Koszalin, Sniadeckich 2, 75-453 Koszalin, Poland; Pidstrygach Institute for Applied Problems of Mechanics and Mathematics, Lviv, Ukraine '
author:
- Volodymyr Sushch
date: Octovber 2004
title: 'Discrete model of Yang-Mills equations in Minkowski space'
---
Introduction
============
The main goal of this paper is to construct a gauge-invariant discrete model of Yang-Mills equations in Minkowski space. Based on the formalism described in [@D1] by Dezin, we consider some intrinsically defined geometric discrete model. A simple two-dimensional discrete model of the classical Yang-Mills equations has been constructed and studied in [@D2]. However, this discrete model is the lacking of gauge invariance. Some another approaches are proposed in [@SM; @SD]. In [@SM] a gauge-invariant discrete analog of the Yang-Mills equations is constructed in Euclidean space $\Bbb{R}^n$. We try to define what gauge invariance is in the case of discrete models. The method described in [@SD] is applicable for obtaining a discrete model of the Yang-Mills equations on the 2-dimensional sphere.
In this paper we concerned with two problems related to discretization in Minkowski space. First we must determine a combinatorial pseudo-Euclidean space and define a discrete analog of the Lorentz metric. Note that in this case to define discrete analogs of the differential and the exterior multiplication can be used the results of [@SM]. As in the continual case these operations do not depend on a metric. Secondly, given a discrete analog of the connection 1-form, a discrete covariant derivative must be defined. We always try to be as close to continual Yang-Mills theory as possible. Nevertheless, gauge invariance of the discrete Yang-Mills equations is obtained under some additional conditions (Theorem 1).
It is known that Yang-Mills theory can be regarded as a non-linear generalization of Hodge theory in the 4-dimensional case (see [@FU]). In Section 5, we construct an operator formally adjoint to the discrete covariant differentiation operator. Then we show how to obtain a discrete analog of the generalized Laplace type operator with respect to the Lorentz metric.
By analogy with the continual case, one of the questions to be studied is discrete analogs of the self-dual and anti-self-dual equations. In Section 6, difference self-dual and anti-self-dual equations are presented as a system of non-linear matrix equations.
Preliminaries
=============
Let $M^4=\Bbb{R}^{1,3}$ be the Minkowski space-time manifold. Suppose that $M^4$ has the Lorentz metric $g_{\mu\nu}=diag(-+++)$. Consider the trivial bundle ${P=M^4\times
SU(2)}$. Let $T^\ast P$ be the cotangent bundle of $P$. It is known (see [@NS]) that a connection can be shown to arise from a certain 1-form $\omega$ belonging to $T^\ast P$, where $\omega$ is required to have values in the Lie algebra $su(2)$. Let $(x,g)$, $x\in M^4$, $g\in SU(2)$, be local coordinates of the bundle $P$. Then $\omega$ is given by $$\omega=g^{-1}dg+g^{-1}Ag,\eqno (1)$$ where $$A=\sum_{\alpha,\mu}A_{\mu}^{\alpha}(x)\lambda_{\alpha}
dx^\mu.\eqno (2)$$ Here we take as a basis for $su(2)$ the set $\{\lambda_\alpha=\frac{\sigma_\alpha}{2i}, \alpha=1,2,3\}$, where $\sigma_\alpha$ are the standard Pauli matrices. The $su(2)$-valued 1-form $A$ is called the connection form and the functions $A_{\mu}^{\alpha}(x)$, connections.
Let the coordinates of $P$ change (locally) from $(x,g)$ to $(x^\prime, g^\prime)$. Let us only make a change of fibre coordinates, i.e. $x=x^\prime$ and $g^\prime$ is given by $$g^\prime=hg, \qquad h\in SU(2).\eqno (3)$$ So invariance of $\omega$ means that $$g^{-1}dg+g^{-1}Ag=(g^\prime)^{-1}dg^\prime+(g^\prime)^{-1}A^\prime g^\prime.$$ Under the change of coordinates (3) the invariant 1-form $\omega$ induces a certain transformation law for the connection form $A$. Taking into account the fact that $dg^\prime=dhg+hdg$ and $dhh^{-1}+hdh^{-1}=0$, we obtain $$A^\prime=hdh^{-1}+hAh^{-1}.\eqno (4)$$ In Yang-Mills theory this transformation law is called the gauge transformation law.
The curvature 2-form $F$ can be defined as follows $$F=dA+A\wedge A. \eqno (5)$$ We have the tensorial law $$F^\prime=hFh^{-1}$$ for the change of $F$ under the gauge transformation (4).
Define the covariant exterior differential operator $d_A$ by $$d_A\Omega=d\Omega+A\wedge \Omega+(-1)^{r+1}\Omega\wedge A,\eqno (6)$$ where $\Omega$ is a $su(2)$-valued $r$-form.
Consider the equations $$d_A F=0,\eqno (7)$$ $$d_A\ast F=0, \eqno (8)$$ where $\ast$ is the metric adjoint operation (Hodge star). Equations (7), (8) are called the Yang-Mills equations [@FU]. Equation (7) is known as the Bianchi identity.
Let $\Phi, \Psi$ be $su(2)$-valued $r$-forms on $M^4$. The “inner product” can be defined as $$(\Phi, \Psi)=-tr\int_{M^4}\Phi\wedge\ast\Psi,\eqno (9)$$ where $tr$ is the trace operator. Note that $M^4$ is non-compact. So all forms referring to the inner product have compact support by assumption.
Then the adjoint operator to $d_A$ can be expressed in the form $$\delta_A=\ast^{-1}d_A\ast,$$ where $\ast^{-1}$ is the inverse operation to $\ast$ ($\ast\ast^{-1}=1$). Combining the latter with Equation (8) we get $$\delta_A F=0.$$ By virtue of (7), this equation is similar to the criterion for a scalar differential form to be harmonic [@W]. Thus, if $F$ is a solution of the Yang-Mills equations, then the following Laplace-Beltrami type equation $$(d_A\delta_A+\delta_Ad_A)F=0 \eqno (10)$$ holds immediately on $M^4$ with respect to the Lorentz metric.
In Minkowski space the self-dual or anti-self-dual equations can be written as follows $$\ast F=\mp iF. \eqno (11)$$ Since $F$ is $su(2)$-valued, so therefore is $\ast F$, then we must have ${su(2)=isu(2)}$. However, this condition is not satisfied for $su(2)$ because it is not satisfied for the Lie algebra of any compact Lie groups [@NS]. In this case, to study the instanton problems one must choose some non-compact groups instead $SU(2)$ such as $SL(2,\Bbb{C})$ or $GL(2,\Bbb{C})$ say.
Combinatorial model of Minkowski space
======================================
Following \[1\], let the tensor product $C(4)=C\otimes C\otimes C\otimes C$ of a 1-dimensional complex be a combinatorial model of Euclidean space $\Bbb{R}^4$. The 1-dimensional complex $C$ is defined in the following way. Let $C^0$ denotes the real linear space of 0-dimensional chains generated by basis elements $x_\kappa$ (points), $\kappa\in \Bbb{Z}$. It is convenient to introduce the shift operators $\tau,\sigma$ in the set of indices by $$\tau\kappa=\kappa+1, \qquad \sigma\kappa=\kappa-1.$$ We denote the open interval $(x_\kappa, x_{\tau\kappa})$ by $e_\kappa$. One can regard the set $\{e_{\kappa}\}$ as a set of basis elements of the real linear space $C^1$. Suppose that $C^1$ is the space of 1-dimensional chains. Then the 1-dimensional complex (combinatorial real line) is the direct sum of the introduced spaces $C=C^0\oplus C^1$. The boundary operator $\partial$ in $C$ is given by $$\partial x_\kappa=0, \qquad \partial e_\kappa=x_{\tau\kappa}-x_\kappa.$$ The definition is extended to arbitrary chains by linearity.
Multiplying the basis elements $x_\kappa, e_\kappa$ in various way we obtain basis elements of $C(4)$. If $c_p, c_q$ are chains of the indicated dimension, belonging to the complexes being multiplied, then $$\partial(c_p\otimes c_q)=\partial c_p\otimes c_q+(-1)^pc_p\otimes\partial c_q. \eqno (12)$$ Relation (12) defines the boundary operator in $C(4)$.
We suppose that the combinatorial model of Minkowski space has the same structure as $C(4)$. We denote only the basis elements corresponding to the time coordinate of $M^4$ by $\bar x_\kappa$, $\bar e_\kappa$. So, for example, the 1-dimensional basis elements of $C(4)$ can be written as $$e_k^1=\bar e_{k_1}\otimes x_{k_2}\otimes x_{k_3}\otimes x_{k_4}, \qquad
e_k^2=\bar x_{k_1}\otimes e_{k_2}\otimes x_{k_3}\otimes x_{k_4},$$ $$e_k^3=\bar x_{k_1}\otimes x_{k_2}\otimes e_{k_3}\otimes x_{k_4}, \qquad
e_k^4=\bar x_{k_1}\otimes x_{k_2}\otimes x_{k_3}\otimes e_{k_4}, \eqno(13)$$ where $k=(k_1, k_2, k_3, k_4)$ is multiindex, $k_j\in\Bbb{Z}$, $j=1,2,3,4$.
Let us now consider a dual complex to $C(4)$. We define its as the complex of cochains $K(4)$ with coefficients belonging to $su(2)$ The complex $K(4)$ has a similar structure, namely ${K(4)=K\otimes K\otimes K\otimes K}$, where $K$ is a dual complex to the 1-dimensional complex $C$. Basis elements of $K$ can be written as $\{x^\kappa\}, \{e^\kappa\}$. Then an arbitrary basis element of $K(4)$ is given by ${s^k=\bar s^{k_1}\otimes s^{k_2}\otimes s^{k_3}\otimes s^{k_4}}$, where $s^{k_j}$ is either $x^{k_j}$ or $e^{k_j}$.
As in [@D2], we define the pairing operation for arbitrary basis elements $\varepsilon_k\in C(4)$, $s^k\in K(4)$ by the rule $$<\varepsilon_k, as^k>=\left\{\begin{array}{l}0,\ \varepsilon_k\ne s_k\\
a,\ \varepsilon_k=s_k,\ a\in su(2).
\end{array}\right. \eqno (14)$$ The operation (14) is linearly extended to cochains. We will call cochains forms, emphasizing their relationship with the corresponding continual objects, differential forms.
The coboundary operator $d^c$ is defined by $$<\partial\varepsilon_k, as^k>=<\varepsilon_k, ad^cs^k>. \eqno (15)$$ The operator $d^c$ is an analog of the exterior differentiation operator.
Let us now introduce in $K(4)$ a multiplication which is an analog of the exterior multiplication for differential forms. First we introduce the $r$-dimensional complex $K(r)$, ${r=1,2,3}$, in an obvious notation. Let $s_{(p)}^k$ be an arbitrary $p$-dimensional basis element of $K(r)$, i.e. the following product ${s_{(p)}^k=\bar s^{k_1}\otimes...\otimes s^{k_r}}$ contains exactly $p$ of the 1-dimensional basis elements $e^{k_{j}}$ and $r-p$ of the 0-dimensional basis elements $x^{k_{j}}$, $k_j\in\Bbb{Z}$, $j=1,...r$. It should be noted that the whole requisite information about the number and situation of “components” is contained in the symbol $(p)$. Then, supposing that the $\cup$-multiplication in $K(r)$ has been defined, we introduce it for basis elements of $K(r+1)$ by the rule $$(s^k_{(p)}\otimes s^\kappa)\cup(s^k_{(q)}\otimes s^\mu)=
Q(\kappa,q)(s^k_{(p)}\cup s^k_{(q)})\otimes(s^\kappa\cup s^\mu),
\eqno (16)$$ where $s^k_{(p)}, s^k_{(q)}\in K(r)$, $s^\kappa(s^\mu)$ is either $x^\kappa(x^\mu)$ or $e^\kappa(e^\mu)$, $\kappa, \mu\in\Bbb{Z}$, and the signum function $Q(\kappa, q)$ is equal to $-1$ if the dimension of both elements $s^\kappa$, $s_{(q)}^k$ is odd and to $+1$ otherwise (see [@D1]). For the basis elements of $K$ the $\cup$-multiplication is defined as follows $$x^\kappa\cup
x^\kappa=x^\kappa, \quad e^\kappa\cup x^{\tau\kappa}=e^\kappa,
\quad x^\kappa\cup e^\kappa=e^\kappa, \quad \kappa\in\Bbb{Z},$$ supposing the product to be zero in all other case. To arbitrary forms the $\cup$-multiplication can be extended linearly. Coefficients of forms multiply as matrices.
[Let $\varphi$ and $\psi$ be arbitrary forms of $K(4)$. Then $$d^c(\varphi\cup\psi)=d^c\varphi\cup\psi+(-1)^p\varphi\cup
d^c\psi, \eqno (17)$$ where $p$ is the dimension of a form $\varphi$.]{}
The proof of Proposition 1 is totally analogous to one in [@D1 p.147] for the case of discrete forms with real coefficients.
By definition, the coboundary operator $d^c$ and the $\cup$-multiplication do not depend on a metric. So they have the same structure in $K(4)$ as in the case of the combinatorial Euclidean space [@SM]. At the same time, to define a discrete analog of the operation $\ast$ we must take into account the structure of the Lorentz metric on $K(4)$. In this case it is convenient to write the basis elements of the complex $K(4)$ in the form $\bar\mu^\kappa\otimes s^k$, where $s^k$ is a basis element of $K(3)$ and $\bar\mu^\kappa$ is either $\bar x^\kappa$ or $\bar e^\kappa$, $\kappa\in\Bbb{Z}$.
Then we define the operation $\ast$ as follows $$\bar\mu^\kappa\otimes s^k\cup\ast(\bar\mu^\kappa\otimes s^k)=
Q(\mu)\bar e^\kappa\otimes e^{k_1}\otimes e^{k_2}\otimes e^{k_3},\eqno (18)$$ where $Q(\mu)$ is equal to $+1$ if $\bar\mu^\kappa=\bar x^\kappa$ and to $-1$ if $\bar\mu^\kappa=\bar e^\kappa$. Relation (18) describes the structure of the Lorentz metric in the discrete model.
Discrete Yang-Mills equations
=============================
The discrete analog of the connection 1-form (2) can be written as $$A=\sum_{j=1}^4\sum_kA_k^je_j^k, \eqno (19)$$ where $e_j^k$ is the 1-dimensional basis element of $K(4)$ and $A_k^j\in su(2)$, $k=(k_1,k_2,k_3,k_4), \ k_j\in\Bbb Z$.
Consider the discrete form $$h=\sum_kh_kx^k, \eqno (20)$$ where $x^k$ is the 0-dimensional basis element of $K(4)$ and $h_k\in SU(2)$. Note that the 0-form (20) does not belong to the complex $K(4)$. But, since $x^k\in K(4)$, the $\cup$-multiplication and the coboundary operator $d^c$ are generalized on the forms (20) in an obvious way.
Then the discrete analog of the gauge transformation (3), (4) can be written as $$g^\prime=h\cup g, \qquad A^\prime=h\cup d^ch^{-1}+h\cup A\cup h^{-1}, \eqno(21)$$ where $h, h^{-1}, g$ are forms of the type (20). Here we denote by $h^{-1}$ the form whose coefficients (matrices) are inverse to coefficients of $h$. If $e$ is the 0-form (20) all of whose coefficients are unit elements of the group $SU(2)$, then we have $$h\cup h^{-1}=h^{-1}\cup h=e.$$ It should be noted that the 0-forms defined by (20) generate a group by respect to the $\cup$-multiplication.
Given the discrete analog of the 1-form (1) by the formula $$\omega=g^{-1}\cup d^cg+g^{-1}\cup A\cup g,$$ it is easy to proof that $\omega$ is invariant ($\omega=\omega^\prime$) under the transformation (21) (see [@SM]).
Now consider the 2-form $$F=\sum_{j=1}^6\sum_kF_k^j\varepsilon_j^k, \eqno (22)$$ where $F_k^j\in su(2)$ and $\varepsilon_j^k$ is the 2-dimensional basis element of $K(4)$. The 2-dimensional basis elements of $K(4)$ can be written as follows $$\varepsilon_1^k=\bar e^{k_1}\otimes e^{k_2}\otimes x^{k_3}\otimes x^{k_4}, \qquad
\varepsilon_2^k=\bar e^{k_1}\otimes x^{k_2}\otimes e^{k_3}\otimes x^{k_4},$$ $$\varepsilon_3^k=\bar e^{k_1}\otimes x^{k_2}\otimes x^{k_3}\otimes e^{k_4}, \qquad
\varepsilon_4^k=\bar x^{k_1}\otimes e^{k_2}\otimes e^{k_3}\otimes x^{k_4},$$ $$\varepsilon_5^k=\bar x^{k_1}\otimes e^{k_2}\otimes x^{k_3}\otimes e^{k_4}, \qquad
\varepsilon_6^k=\bar x^{k_1}\otimes x^{k_2}\otimes e^{k_3}\otimes e^{k_4},$$ where $k_i\in\Bbb{Z}$, $i=1,2,3,4.$
We define the discrete analog of the curvature 2-form by the formula $$F=d^cA+A\cup A. \eqno (23)$$
Under the gauge transformation (21) the curvature form (23) changes as $$F^\prime=h\cup F\cup h^{-1}.$$
The proof closely follows the proof Theorem 2 of [@SM]. Using (21) and (17) we compute $$d^c A^\prime=d^ch\cup
d^ch^{-1}+d^ch\cup A\cup h^{-1}+ h\cup d^cA\cup h^{-1}-h\cup A\cup
d^ch^{-1}.$$ Since $d^ce=0$ by definition of $d^c$, we have $$d^c(h\cup h^{-1})=d^ch\cup h^{-1}+h\cup d^ch^{-1}=0$$ and so $$d^ch\cup h^{-1}=-h\cup d^ch^{-1}.\eqno (24)$$ Taking into account (24), we obtain $$\begin{array}{l}
A^\prime\cup A^\prime=(h\cup d^ch^{-1}+h\cup A\cup h^{-1})\cup
(h\cup d^ch^{-1}+h\cup A\cup h^{-1})\\
=-d^ch\cup d^ch^{-1}+h\cup A\cup d^ch^{-1}-d^ch\cup A\cup h^{-1}+
h\cup A\cup A\cup h^{-1}.
\end{array}$$ Then we finally have $$\begin{array}{l}
F^\prime=d^cA^\prime+A^\prime\cup A^\prime=h\cup d^c A\cup h^{-1}
+h\cup A\cup A\cup h^{-1}\\
\quad {}=h\cup(d^c A+A\cup A)\cup h^{-1}.
\end{array}$$
From the definition (23) one easily derives that the curvature form $F$ satisfies the identity $$d^cF+A\cup F-F\cup A=0. \eqno (25)$$ The comparison of (25) with (7) yields a discrete analog of the Bianchi identity. Define now the discrete analog of the exterior covariant differentiation operator by setting $$d_A^c\Omega=d^c\Omega+A\cup\Omega+(-1)^{r+1}\Omega\cup A,$$ where $\Omega$ is an arbitrary $r$-form of $K(4)$. Then Identity (25) can be rewritten as $$d^c_AF=0.$$ In similar manner, we obtain the discrete analog of Equation (8) $$d_A^c\ast F\equiv d^c\ast F+A\cup\ast F-\ast F\cup A=0. \eqno (26)$$
Let $\tau_{ij}$ ($\sigma_{ij}$), $i,j=1,2,3,4,\ i\ne j$, be the shift operator acting as the operator $\tau$ ($\sigma$) by the $i$-th, $j$-th components of the multiindex $k=(k_1,k_2,k_3,k_4)$. For example, $$\tau_{12}k=(\tau k_1,\tau k_2,k_3,k_4),\qquad
\sigma_{23}k=(k_1,\sigma k_2,\sigma k_3,k_4).$$ Using the definition (18) we compute $$\ast F=\sum_k (F_{\sigma_{34}k}^6\varepsilon_1^k-
F_{\sigma_{24}k}^5\varepsilon_2^k +
F_{\sigma_{23}k}^4\varepsilon_3^k
- F_{\sigma_{14}k}^3\varepsilon_4^k+
F_{\sigma_{13}k}^2\varepsilon_5^k -
F_{\sigma_{12}k}^1\varepsilon_6^k).
\eqno (27)$$
Let $h$ be a discrete 0-form. Then we have $$\ast(h\cup
f)=h\cup\ast f \eqno (28)$$ for an arbitrary $p$-form $f\in
K(4)$.
Any $p$-form $f\in K(4)$ can be expressed as $$f=\sum_k f_k^{(p)}s_{(p)}^k,$$ where $f_k^{(p)}\in su(2)$ and $s_{(p)}^k$ is the $p$-dimensional basis element of $K(4)$. By definition, we have $x^k\cup s_{(p)}^k=s_{(p)}^k$ for an arbitrary 0-dimensional basis element $x^k$ of $K(4)$. Hence, $$h\cup f=\left(\sum_k h_kx^k\right)\cup\left(\sum_k f_k^{(p)}s^k_{(p)}\right)
=\sum_k h_kf_k^{(p)}s^k_{(p)}.$$ Then we obtain $$\ast(h\cup f)=\sum_k h_kf_k^{(p)}\ast s^k_{(p)}=h\cup\sum_k f_k^{(p)}\ast s^k_{(p)}=
h\cup\ast f.$$
We have $$\ast(f\cup h)=\ast f\cup h \eqno (29)$$ for an arbitrary 2-form $f\in K(4)$ if and only if coefficients of a 0-form $h$ satisfy the following conditions $$h_k=h_{\sigma_{ij}k} \eqno
(30)$$ for all $i,j=1,2,3,4, i\ne j$.
From the definition (16) one easily derives that in the form $f\cup h$ two indices in the coefficients $h_k$ are shifted and we have $$f\cup
h=\sum_k(f_k^1h_{\tau_{12}k}\varepsilon_1^k+f_k^2h_{\tau_{13}k}\varepsilon_2^k+
f_k^3h_{\tau_{14}k}\varepsilon_3^k+f_k^4h_{\tau_{23}k}\varepsilon_4^k+
f_k^5h_{\tau_{24}k}\varepsilon_5^k+f_k^6h_{\tau_{34}k}\varepsilon_6^k).$$ Since $$\ast\varepsilon_1^k=-\varepsilon_6^{\tau_{12}k},
\qquad \ast\varepsilon_2^k=\varepsilon_5^{\tau_{13}k}, \qquad
\ast\varepsilon_3^k=-\varepsilon_4^{\tau_{14}k},$$ $$\ast\varepsilon_4^k=\varepsilon_3^{\tau_{23}k}, \qquad
\ast\varepsilon_5^k=-\varepsilon_2^{\tau_{24}k}, \qquad
\ast\varepsilon_6^k=\varepsilon_1^{\tau_{34}k},$$ we obtain $$\ast(f\cup
h)=\sum_k(-f_k^1h_{\tau_{12}k}\varepsilon_6^{\tau_{12}k}
+f_k^2h_{\tau_{13}k}\varepsilon_5^{\tau_{13}k}-
f_k^3h_{\tau_{14}k}\varepsilon_4^{\tau_{14}k}$$ $$\qquad\qquad\qquad +f_k^4h_{\tau_{23}k}\varepsilon_3^{\tau_{23}k}-
f_k^5h_{\tau_{24}k}\varepsilon_2^{\tau_{24}k}+
f_k^6h_{\tau_{34}k}\varepsilon_1^{\tau_{34}k}).\eqno (31)$$ Taking into account the relation $$\sum_kh_kx^k=\sum_kh_{\tau
k}x^{\tau k},$$ we compute $$\ast f\cup h=\sum_k(-f_k^1h_{\tau
k}\varepsilon_6^{\tau_{12}k}+ f_k^2h_{\tau
k}\varepsilon_5^{\tau_{13}k}- f_k^3h_{\tau
k}\varepsilon_4^{\tau_{14}k}$$ $$\qquad\qquad\qquad
+f_k^4h_{\tau k}\varepsilon_3^{\tau_{23}k}- f_k^5h_{\tau
k}\varepsilon_2^{\tau_{24}k}+ f_k^6h_{\tau
k}\varepsilon_1^{\tau_{34}k}),\eqno (32)$$ where $\tau k=(\tau
k_1,\tau k_2,\tau k_3,\tau k_4)$.
Inserting (31), (32) into (29), we get $$h_{\tau k}=h_{\tau_{12}k}=h_{\tau_{13}k}=h_{\tau_{14}k}=h_{\tau_{23}k}
=h_{\tau_{24}k}=h_{\tau_{34}k}$$ for an arbitrary $k$. Clearly, these relations imply (30).
On other hand, Conditions (30) we can rewritten as follows $h_{\tau k}=h_{\tau_{ij}k}$ for all $i,j=1,2,3,4,\ i\ne j$. Substituting the latter into (31) and comparing (31) and (32), we obtain (29).
It should be noted that in the Lemmas we can taken the 0-form $h$ either as an element of $K(4)$ or as a form of the type (20).
Conditions (30) mean that the “diagonal components” of the 0-form $h$ are equal in all plans as shown in Fig.1. Remind that we regard $h$ as the function over the points $x_k$ and ${h|_{x_k}=<x_k, h>=h_k}$.
(300,180) (89,40)[.]{} (89,90)[.]{} (139,40)[.]{} (139,140)[.]{} (189,90)[.]{} (139,90)[.]{} (189,140)[.]{} (90,40)[(0,1)[60]{}]{} (90,40)[(1,0)[60]{}]{} (90,40)[(0,-1)[10]{}]{} (90,40)[(-1,0)[10]{}]{} (90,90)[(1,0)[110]{}]{} (140,40)[(0,1)[110]{}]{} (90,90)[(-1,0)[10]{}]{} (140,40)[(0,-1)[10]{}]{} (140,140)[(1,0)[60]{}]{} (190,90)[(0,1)[60]{}]{} (140,140)[(-1,0)[10]{}]{} (190,90)[(0,-1)[10]{}]{} (143,95)[$h_k$]{} (193,95)[$h_{\tau_ik}$]{} (193,145)[$h_{\tau_{ij}k}$]{} (143,145)[$h_{\tau_jk}$]{} (143,45)[$h_{\sigma_jk}$]{} (93,45)[$h_{\sigma_{ij}k}$]{} (93,95)[$h_{\sigma_ik}$]{} (110,5)[Fig. 1. (in the plan ($k_i,k_j$)).]{}
The set of 0-forms (20) satisfying Conditions (30) is a group under $\cup$-multiplication.
The claim is obvious. By the definition (16), the product of any 0-forms is a 0-form and indices in coefficients do not shift. From this the result follows at once.
Under Conditions (30) the Yang-Mills equation (26) is gauge invariant.
Here gauge invariance is understood as follows. If $A(F)$ is a solution of Equation (26), then $A^\prime(F^\prime)$ is also a solution of (26).
By Proposition 3, the form $h^{-1}$ satisfies Conditions (30). Using Proposition 2 from Lemma 1 and 2 we have $$\ast F^\prime=h\cup\ast F\cup h^{-1}.$$ Now express $d^c_{A^\prime}\ast F^\prime$ in terms of $F, A$. Applying (17) we compute $$d^c\ast F^\prime=d^ch\cup\ast F\cup h^{-1}+h\cup d^c\ast F\cup h^{-1}+
h\cup\ast F\cup d^ch^{-1}.$$ Taking into account (21) and (24), we obtain $$A^\prime\cup\ast F^\prime=-d^ch\cup\ast F\cup h^{-1}+
h\cup A\cup\ast F\cup h^{-1}$$ and $$\ast F^\prime\cup A^\prime=h\cup\ast F\cup d^ch^{-1}+
h\cup\ast F\cup A\cup h^{-1}.$$ Thus, $$d^c_{A^\prime}\ast F^\prime=h\cup d_A^c\ast F\cup h^{-1}.$$
The operator formally adjoint to $d_A^c$
========================================
Let $V\subset C(4)$ be some fixed “domain” of the complex $C(4)$. We can written $V$ as follows $$V=\sum_kV_k, \qquad k=(k_1,k_2,k_3,k_4),\quad k_i=1,2, ...,N_i, \eqno (33)$$ where $V_k=\bar e_{k_1}\otimes e_{k_2}\otimes e_{k_3}\otimes e_{k_4}$ is the 4-dimensional basis element of $C(4)$. We agree that in what follows the subscripts $k_i,\ i=1,2,3,4$, always run the set of values indicated in (33). In this section we suppose that coefficients of the discrete forms are vanished on $C(4)\setminus V$. Then the “inner product” for forms $\Phi, \Psi\in K(4)$ of the same degree is defined by the relation $$(\Phi, \Psi)_V=-tr<V, \Phi\cup\ast\Psi>. \eqno (34)$$ For the forms of different degrees the product (34) is set equal to zero.
The definition imitates correctly the continual case (Relation (9)). It follows from (18) that for the basis elements $e_i^k$ and $\varepsilon_j^k$ we have $e_1^k\cup\ast e_1^k=-V^k$, $e_i^k\cup\ast e_i^k=V^k$ for $i=2,3,4$, $\varepsilon_j^k\cup\ast\varepsilon_j^k=-V^k$ for $j=1,2,3$ and $\varepsilon_j^k\cup\ast\varepsilon_j^k=V^k$ for $j=4,5,6$. Then we obtain $$(A, A)_V=-tr\sum_k\left[-(A_k^1)^2+(A_k^2)^2+(A_k^3)^2+(A_k^4)^2\right]$$ and $$(F, F)_V=-tr\sum_k\left[-(F_k^1)^2-(F_k^2)^2-(F_k^3)^2+(F_k^4)^2+(F_k^5)^2+
(F_k^6)^2\right],$$ where $A$ is an 1-form (19) and $F$ is a 2-form (22).
Let $\Phi\in K(4)$ be an 1-form and $\Psi\in K(4)$ be a 2-form. Then we have $$(d^c\Phi, \Psi)_V=(\Phi, \delta^c\Psi)_V,$$ where $$\delta^c\Psi=\ast^{-1}d^c\ast\Psi \eqno(35)$$ is the operator formally adjoint to $d^c$.
From (17) and (34) we obtain $$\begin{array}{rcl} (d^c\Phi,
\Psi)_V&=&-tr<V, d^c\Phi\cup\ast\Psi)>\\ &=&-tr<V,
d^c(\Phi\cup\ast\Psi)>-tr<V, \Phi\cup d^c\ast\Psi>\\
&=&-tr<\partial V, \Phi\cup\ast\Psi>-
tr<V,\Phi\cup\ast(\ast^{-1}d^c\ast\Psi)>\\ &=&-tr<\partial V,
\Phi\cup\ast\Psi>+ (\Phi,\ \ast^{-1}d^c\ast\Psi)_V,
\end{array}$$ where we used $\ast\ast^{-1}=1$.
Let $\tilde e_k^j, \ j=1,2,3,4,$ denote the 3-dimensional basis element of $C(4)$. Using (12) we derive that $$\partial V=\sum_k (\tilde e_{\tau N_1,k_2,k_3,k_4}^1-
\tilde e_{1,k_2,k_3,k_4}^1-\tilde e_{k_1,\tau N_2,k_3,k_4}^2+
\tilde e_{k_1,1,k_3,k_4}^2$$ $$\qquad\qquad +\tilde e_{k_1,k_2,\tau N_3,k_4}^3-
\tilde e_{k_1,k_2,1,k_4}^3-\tilde e_{k_1,k_2,k_3,\tau N_4}^4+
\tilde e_{k_1,k_2,k_3,1}^4),$$ where $$\tilde e_k^1=\bar x_{k_1}\otimes e_{k_2}\otimes e_{k_3}\otimes e_{k_4}, \qquad
\tilde e_k^2=\bar e_{k_1}\otimes x_{k_2}\otimes e_{k_3}\otimes e_{k_4},$$ $$\tilde e_k^3=\bar e_{k_1}\otimes e_{k_2}\otimes x_{k_3}\otimes e_{k_4}, \qquad
\tilde e_k^4=\bar e_{k_1}\otimes e_{k_2}\otimes e_{k_3}\otimes x_{k_4}.$$ Computing the “boundary components” of the form $\Phi\cup\ast\Psi$ we obtain the linear combination of the following products: $$\Phi_{k_1...\tau N_i...k_4}^j\cdot\Psi_{k_1...N_i...k_4}^r \
\mbox{and}\
\Phi_{k_1...0...k_4}^j\cdot\Psi_{k_1...0...k_4}^r,\ i,j=1,2,3,4,\
r=1,2,...,6.$$ Since we have ${\Phi_{k_1...\tau N_i...k_4}^j=\Psi_{k_1...0...k_4}^r=0}$ for all $i,j,r$ by assumption, it follows that ${<\partial V,\Phi\cup\ast\Psi>=0}$.
For the 2-form $F$ using (27) and the definition of $d^c$ we can rewritten (35) in the form $$\begin{array}{r}
\delta^cF=\sum_k[(\Delta_{k_2}F_{\sigma_2k}^1+
\Delta_{k_3}F_{\sigma_3k}^2+\Delta_{k_4}F_{\sigma_4k}^3)e_1^k\\
+(\Delta_{k_1}F_{\sigma_1k}^1+\Delta_{k_3}F_{\sigma_3k}^4+
\Delta_{k_4}F_{\sigma_4k}^5)e_2^k\\
+(\Delta_{k_1}F_{\sigma_1k}^2-\Delta_{k_2}F_{\sigma_2k}^4+
\Delta_{k_4}F_{\sigma_4k}^6)e_3^k\\
+(\Delta_{k_1}F_{\sigma_1k}^3-\Delta_{k_2}F_{\sigma_2k}^5-
\Delta_{k_3}F_{\sigma_3k}^6)e_4^k].
\end{array}$$ Here we denote by $\Delta_{k_i}F_k^j$ the difference $F_{\tau_ik}^j-F_k^j,\ j=1,2,...,6,$ and $\tau_ik={k_1...\tau k_i...k_4}$, $\sigma_ik={k_1...\sigma k_i...k_4}$, $i=1,2,3,4$.
For any 1-form $\Phi\in K(4)$ and 3-form $\Psi\in K(4)$ the following relation holds $$tr<V, \Phi\cup\Psi>=-tr<V,
\Psi\cup\ast\ast\Phi>.\eqno(36)$$
The forms $\Phi$ and $\Psi$ can be expressed as $$\Phi=\sum_{i=1}^4\sum_k\Phi_k^ie_i^k, \qquad
\Psi=\sum_{i=1}^4\sum_k\Psi_k^i\tilde e_i^k,$$ where $\Phi_k^i,\
\Psi_k^i\in su(2)$ and $\tilde e_i^k$ is the 3-dimensional basis element of $K(4)$.
Using (14), (16) we compute $$tr<V, \Phi\cup\Psi>=tr\sum_k(\Phi_{\sigma_1k}^1\cdot\Psi_k^1-
\Phi_{\sigma_2k}^2\cdot\Psi_k^2+\Phi_{\sigma_3k}^3\cdot\Psi_k^3
-\Phi_{\sigma_4k}^4\cdot\Psi_k^4),$$ where $\sigma_ik=(k_1...\sigma k_i...k_4)$.
On the other hand, since $$\ast\ast\Phi=\sum_{i=1}^4\sum_k\Phi_k^ie_i^{\tau k}, \quad
\tau k=(\tau k_1,\tau k_2,\tau k_3,\tau k_4),\eqno (37)$$ we have $$tr<V, \Psi\cup\ast\ast\Phi>=tr\sum_k(-\Psi_k^1\cdot\Phi_{\sigma_1k}^1+
\Psi_k^2\cdot\Phi_{\sigma_2k}^2-\Psi_k^3\cdot
\Phi_{\sigma_3k}^3+\Psi_k^4\cdot\Phi_{\sigma_4k}^4)$$ $$\qquad\qquad\qquad\qquad=tr\sum_k(-\Phi_{\sigma_1k}^1\cdot
\Psi_k^1+\Phi_{\sigma_2k}^2\cdot\Psi_k^2-\Phi_{\sigma_3k}^3\cdot
\Psi_k^3+\Phi_{\sigma_4k}^4\cdot\Psi_k^4),$$ where we used $tr(\Phi_k^i\cdot\Psi_k^i)=tr(\Psi_k^i\cdot\Phi_k^i)$. From this the result follows at once.
It should be noted that in the continual case we have the equality $$tr(\varphi\wedge\psi)=(-1)^{pq}tr(\psi\wedge\varphi),$$ where $\varphi$ and $\psi$ are matrix-valued differential forms of degree $p,q$, respectively. Unfortunately, this equality has not an exact analog in our formalism.
For any 2-form $F\in K(4)$ the formal adjoint operator to $d_A^c$ acts as follows $$\delta_A^cF=\ast^{-1}(d^c\ast F-\ast
F\cup\ast\ast A+A\cup\ast F).\eqno (38)$$
We will compute the operator $\delta_A^c$ defined by the relation $$(d_A^c\Phi, F)_V=(\Phi, \delta_A^cF)_V,$$ where $\Phi$ is an 1-form.
Using (36) and (34) we have $$\begin{array}{l}
(d_A^c\Phi,\ F)_V=-tr<V,\ d_A^c\Phi\cup\ast F>\\=-tr<V,\ d^c\Phi\cup\ast F>
-tr<V,\ (A\cup\Phi\cup\ast F+ \Phi\cup A\cup\ast F)>\\
=-tr<V,\ d^c\Phi\cup\ast F>+tr<V,\ (\Phi\cup\ast F\cup\ast\ast A-
\Phi\cup A\cup\ast F)>\\
=(d^c\Phi,\ F)_V+tr<V,\ (\Phi\cup\ast\ast^{-1}(\ast F\cup\ast\ast A)-
\Phi\cup\ast\ast^{-1}(A\cup\ast F))>\\
=(\Phi,\ \delta^c F)_V-(\Phi,\ \ast^{-1}(\ast F\cup\ast\ast A))_V+
(\Phi,\ \ast^{-1}(A\cup\ast F))_V\\
=(\Phi,\ \delta^cF-\ast^{-1}(\ast F\cup\ast\ast A)+\ast^{-1}(A\cup\ast F))_V.
\end{array}$$
In the continual case, if we choose the Lorentz metric, then $$\ast\ast A=A$$ for an arbitrary differential 1-form $A$.
Hence the Yang-Mills equation (8) can be rewritten as follows $$d\ast F+A\wedge\ast F-\ast F\wedge\ast\ast A=0.\eqno (39)$$ It follows that a discrete analog of Equation (39) (or (8)) can be given by $$d_A^c\ast F\equiv d^c\ast F+A\cup\ast F-\ast F\cup\ast\ast A=0.\eqno (40)$$ Comparing the latter and (38) we obtain $$\delta_A^cF=\ast^{-1}d_A^c\ast F.$$ Thus, if the discrete curvature 2-form $F$ is a solution of Equation (40), then the Laplace type equation $$(d_A^c\delta_A^c+\delta_A^cd_A^c)F=0$$ holds immediately. This equation we call a discrete analog of Equation (10).
It should be noted that in our discrete model the operation $\ast^2=\ast\ast$ is equivalent to a shift with corresponding sign (see (37)). So, unfortunately, Equation (40) differs from Equation (26). The possibility of involute $(\ast\ast=1)$ definition of $\ast$ is discussed in [@SM2].
Discrete models of the self-dual and anti-self-dual equations
==============================================================
In this section we will construct a difference analog of Equations (9). For this reason we take the group $SL(2,\Bbb C)$ instead $SU(2)$. Let the components $F_k^j$ of the curvature form $F$ be belonging to $sl(2,\Bbb C)$. Combining (22) with (27) the discrete self-dual equation $\ast F=iF$ can be written as follows $$F_{\sigma_{34}k}^6=iF_k^1, \qquad -F_{\sigma_{24}k}^5=iF_k^2, \qquad
F_{\sigma_{23}k}^4=iF_k^3,$$ $$-F_{\sigma_{14}k}^3=iF_k^4, \qquad F_{\sigma_{13}k}^2=iF_k^5, \qquad
-F_{\sigma_{12}k}^1=iF_k^6$$ for all $k=(k_1,k_2,k_3,k_4), \ k_r\in\Bbb Z, \ r=1,2,3,4$. From the latter we obtain $$F_{\sigma k}^6=iF_{\sigma_{12}k}^1=-i^2F_k^6=F_k^6, \qquad
F_{\sigma k}^5=-iF_{\sigma_{13}k}^2=-i^2F_k^5=F_k^5$$ and similarly for any other components $F_k^j, \ j=1,2,...,6$.
So we have $$F_k^j=F_{\sigma k}^j. \eqno (41)$$ We call Equations (41) difference self-dual equations.
In similar manner, we obtain the difference anti-self-dual equations $$F_k^j=-F_{\sigma_k}^j. \eqno (42)$$
For any 2-form $F$ such that $F_k^j=\pm F_{\sigma k}^j$ we have $$\ast\ast F=\mp F.$$
$$\ast\ast F=\ast(\pm iF)=\pm i\ast F=\pm i^2F=\mp F.$$
Using (23) Equations (41) can be rewritten as $$\Delta_{k_i}A_k^r-\Delta_{k_r}A_k^i+A_k^i\cdot A_{\tau_ik}^r
-A_k^r\cdot A_{\tau_rk}^i=
\Delta_{k_i}A_{\sigma k}^r-\Delta_{k_r}A_{\sigma k}^i+
A_{\sigma k}^i\cdot A_{\sigma\tau_ik}^r
-A_{\sigma k}^r\cdot A_{\sigma\tau_rk}^i,$$ where $\sigma\tau_ik=(\sigma k_1...k_i...\sigma k_4)$, $i<r, \ r=1,2,3,4$ and $A_k^r\in sl(2,\Bbb C)$ is a component of the connection form $A$. Similarly, Equations (42) become $$\Delta_{k_i}A_k^r-\Delta_{k_r}A_k^i+A_k^i\cdot A_{\tau_ik}^r
-A_k^r\cdot A_{\tau_rk}^i=\Delta_{k_r}A_{\sigma k}^i
-\Delta_{k_i}A_{\sigma k}^r-
A_{\sigma k}^i\cdot A_{\sigma\tau_ik}^r
+A_{\sigma k}^r\cdot A_{\sigma\tau_rk}^i.$$
Let $F\in K(4)$ be a 2-form with compact support. Then the discrete self-dual (anti-self-dual) equations have the unique solution $F=0$.
Since Equations (41) (Equations (42)) hold for all $k=(k_1,k_2,k_3,k_4)$, $k_r\in\Bbb Z$, then the assertion is obvious.
In the continual case we must write the self-dual and anti-self-dual equations in the form (11) because we have $\ast\ast F=-F$ for the Lorentz metric. In the case of the discrete model it is easy to check that in $K(4)$ we have $$\ast\ast F=-\sum_{j=1}^6\sum_k
F_{\sigma k}^j\varepsilon_j^k.$$ It should now be clear that a discrete model of Equation (11) can be defined as follows $\ast
F=\pm F$. Then we obtain the following difference equations $$F_k^j=\mp F_{\sigma k}^j$$ for all $j=1,2, ...,6, \
k=(k_1,k_2,k_3,k_4), \ k_r\in \Bbb Z$. Therefore, on opposite to the continual case we can study these equations for the group $SU(2)$, i. e. for $F$ with components $F_k^j\in su(2)$.
[99]{} Dezin A. A. Multidimensional Analysis and Discrete Models. [*CRC Press*]{}. Boca Raton. 1995. Dezin A. A. Models generated by the Yang-Mills equations. [*Differentsial’nye Uravneniya*]{}. Vol. 29. No 5. P. 846–851. 1993; English transl.in [*Differential Equations*]{}. 1993. D. Fried, K. Uhlenbek. Instantons and Four-Manifolds. [*Springer–Verlag*]{}. 1984. C. Nash, S. Sen. Toplogy and Geometry for Physicists. [*Acad. Press*]{}. London. 1989. Sushch V. N. Gauge-invariant discrete models of Yang-Mills equations. [*Mat. Zametki*]{}. Vol. 61. No 5. P. 742–754. 1997; English transl. in [*Mathematical Notes*]{}. Vol. 61. No 5. P. 621–631. 1997. Sushch V. N. Discrete models on the 2-sphere. [*Dop. NAN Ukrainy*]{}. No 2. P. 27–32. 2000. Sushch V. N. On some difference analog of the first order invariant hyperbolic systems. [*Differentsial’nye Uravneniya*]{}. Vol. 35. No 3. P. 1–7. 1999; English transl in [*Differential Equations*]{}. 1999. Sushch V. N. Discrete models of Yang-Mills equations on the sphere. [*Mathematical Notes*]{}. (to appear). C. von Westenholz. Differential Forms in Mathematical Physics. (Studies in Mathematics and its Applications). Vol. 3. [*North-Holland Publishing Company*]{}. Amsterdam. 1981.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We describe a prototype of a new experimental GeoGebra command and tool `Discover` that analyzes geometric figures for salient patterns, properties, and theorems. This tool is a basic implementation of automated discovery in elementary planar geometry. The paper focuses on the mathematical background of the implementation, as well as methods to avoid combinatorial explosion when storing the interesting properties of a geometric figure.'
author:
- Zoltán Kovács
- 'Jonathan H. Yu'
bibliography:
- 'kovzol.bib'
- 'external.bib'
title: |
Towards Automated Discovery of\
Geometrical Theorems in GeoGebra
---
Introduction
============
In this technical paper we introduce a new GeoGebra command and tool `Discover` that is available in a development GitHub repository [@geogebra-discovery]. This research is closely related to a former project [@ag] (see [@adg-ag; @aisc-ag; @LNAI11110-ag] for further details).
Given a Euclidean geometry construction drawn in GeoGebra, suppose a user wants to know if a given object $O$ has some “interesting features,” such as relevant theorems or properties. This object can be a point, a line, a circle, or something else, although in the current implementation $O$ will always be a point. Without any further user input, the `Discover` command will then analyze $O$ for its interesting and relevant features, and present them to the user as both a list of formulas and graphics outputs.
![Initial setup for a discovery[]{data-label="midline1"}](midline1.png)
For example, let $ABC$ an arbitrary triangle, and let $D$ and $E$ be the midpoints of $BC$ and $AC$, respectively (Fig. \[midline1\]). Has $D$ some interesting features? Yes: $DE$ is parallel to $AB$, independent of the position of $A$, $B$ and $C$. Indeed, the command `Discover(D)` confirms this observation with the output shown in Fig. \[rel-midline\]; GeoGebra adds lines $DE$ and $AB$ in the same color (Fig. \[midline2\]). (Note, however, that the current implementation of GeoGebra does not report that $2\cdot|DE|=|AB|$.) Also, the software reports the somewhat trivial finding that the segments $BD$ and $CD$ are congruent, with $BD$ and $CD$ highlighted in the same color. This output can be obtained by selecting the Discover tool in GeoGebra’s toolbox:
{height="0.8cm"}
and then clicking on the point $D$. This functionality is implemented in both GeoGebra Classic 5 and 6, available as an experimental software package called *GeoGebra Discovery*, at <http://github.com/kovzol/geogebra-discovery>.
![Further output of the `Discover` command[]{data-label="midline2"}](midline2.png)
What strategy is used in the background? First, all points are analyzed to determine whether they are the same as another point. Then, all possible point triplets are examined for collinearity. Next, all possible subsets containing four points on the figure are checked for concyclicity. With knowledge of the collinear points, separate lines can be uniquely defined, in order find whether they are parallel. Finally, considering the pairs of all possible point pairs, congruent segments can be identified. This strategy is a result of a combination of numerical and symbolic processes. Our second example shows a more complicated setup. A regular hexagon $ABCDEF$ is given in Fig. \[hexagon1\]. Point $G$ is defined as the intersection of $AD$ and $BE$, and, in addition, $H=BE\cap CF$, $I=AD\cap CF$. The points $G$, $H$ and $I$ may have trivial differences in their numerical representations, but in the geometrical sense they should be equal. In the figure rounding was set to 13 digits to emphasize that GeoGebra computes objects numerically by default. Note that while the $y$-coordinates of $H$ and $I$ numerically differ, the final calculations to prove that they are identical will be symbolic and exact.
Now we are about to learn if point $F$ has some interesting features, so the command `Discover(F)` will be issued. GeoGebra reports a set of properties in a message box (Fig. \[rel-hexagon\]) and adds some additional outputs to the initial setup (Fig. \[hexagon2\]).
![Further output of discovery[]{data-label="hexagon2"}](6-gon2.png)
Here, we see that concyclic points are reported as a single item and not as separate data. Also, parallel lines are classified into five different sets. Finally, there are three sets of congruent segments. This approach in computation and reporting helps avoid combinatorial explosion.
Mathematical background {#sec2}
=======================
The above mentioned strategies have some similiarities to the ones introduced in [@song], but here we focus on minimizing the number of objects that have to be compared in the process that practically compares all objects with all other objects.
Our current implementation deals with *points*, *lines*, *circles* and *parallel lines* (or *directions*) and *congruent segments*.
A *geometric point* $P$ is a GeoGebra object, described by the `GeoPoint` class (see GeoGebra’s source code at [github.com/geogebra/geogebra](github.com/geogebra/geogebra) for more details). While we will not provide a detailed definition of a geometric point, generally speaking it is an object with a very complex structure containing two real coordinates, several style settings (including size and color, for example) and other technical details that are used in the application. Some geometric points are dependent of other geometric points or other geometric objects—this hierarchy is stored in the set of `GeoPoint`s, too.
Independent of the detailed definition of a geometric point, we can still define the notion of *point* in our context.
A set of geometric points ${\cal P}=\{P_1,P_2,\ldots,P_n\}$ is called a point if for all different $P,Q\in {\cal P}$ the points $P$ and $Q$ are identical in general.
Henceforth, unless otherwise mentioned, we will consider points according to the definition above, not as geometric points.
Here, we do not precisely define when two points are identical *in general*. Instead, we will illustrate the concept of point identicality with the following example. Consider geometric points $P_1$, $P_2$, $P_3$ and $P_4$ that form a parallelogram. Now define $P_5$ and $P_6$ as the midpoint of $P_1$ and $P_3$, and $P_2$ and $P_4$, respectively. This setting implies that $P_5$ and $P_6$ are identical, because the diagonals of a parallelogram always bisect each other. In a dynamic geometry setting like GeoGebra, this simply means that by changing some points of the set $\{P_1,P_2,P_3,P_4\}$, the points $P_5$ and $P_6$ will still share the same position in the plane. (See Fig. \[parallelogram1\]. Here the construction is controlled by the points $P_1$, $P_2$ and $P_3$ only: they can be freely chosen, and based on them, the point $P_4$ is already dependent and uniquely defined as the intersection of the two parallel lines to $P_1P_2$ and $P_2P_3$, respectively, through $P_3$ and $P_1$.)
![Points $P_5$ and $P_6$ are defined as midpoints of opposite vertices of parallelogram $P_1P_2P_3P_4$[]{data-label="parallelogram1"}](parallelogram1)
In fact, general truth includes statements that are not always true, but just “in most cases”—here we can think of some degeneracies that can occur in some constructions when some objects are degenerate. For example, altitudes of a triangle generally meet at a point—but not always, since a degenerate triangle “usually” has three parallel “altitudes”; unless two (or even three!) vertices of the triangle coincide. (See [@Chou_1987] for more details on the concept of general truth and degeneracies.)
A set of points $\ell=\{P_1,P_2,\ldots,P_n\}$ is called a line if for all different $P,Q,R\in \ell$ the points $P$, $Q$ and $R$ are collinear in general.
For example, the set $\ell=\{C,F,G\}$ in Fig. \[hexagon2\] forms a line.
A set of points ${\cal C}=\{P_1,P_2,P_3,\ldots,P_n\}$ is called a circle if for all different $P,Q,R,S\in {\cal C}$ the points $P$, $Q$, $R$ and $S$ are concyclic in general.
A set of lines $\vec{D}=\{\ell_1,\ldots,\ell_n\}$ is called parallel lines (or a direction) if for all different $\ell,m\in \vec{D}$ the lines $\ell$ and $m$ are parallel in general.
A set $\overline{s}=\{P,Q\}$ of two points is called a segment.
A set of segments $s=\{\overline{s_1},\ldots,\overline{s_n}\}$ is called equal length segments (or congruent segments) if for all different $\overline{s_1},\overline{s_2}\in s$ the segments $\overline{s_1}$ and $\overline{s_1}$ are equally long in general.
In fact, GeoGebra Discovery uses a more general concept of being identical: it allows two points (or two objects) to have a kind of relationship also if it is true just *on parts* (see [@rmc-top] for more details). The main idea of storing the objects is that points, lines, circles, directions and equally long segments designate equivalence classes, that is:
Let $\ell$ and $m$ be lines. Then, for all different points $P,Q,R\in m$, if $\{P, Q\}\subset \ell$, then $R\in \ell$; that is, $\ell = m$.
In Euclidean geometry two points always designate a unique line.
Let ${\cal C}$ and ${\cal D}$ be circles. Then, for all different points $P,Q,R,S\in{\cal D}$, if $\{P,Q,R\}\subset {\cal C}$, then $S\in{\cal C}$; that is, ${\cal C}={\cal D}$.
In Euclidean geometry three non-collinear points always designate a unique circle.
Let $\vec{D}$ and $\vec{E}$ be directions. Let $\ell\in \vec{D}$ and $m\in\vec{E}$. If $\ell\parallel m$ in general, then $\vec{D}=\vec{E}$.
This follows immediately from the transitive property of parallelism.
Let $s$ and $t$ be segments. Let $\overline{u}\in s$ and $\overline{v}\in t$. If $|\overline{u}|=|\overline{v}|$ in general, then $s_1=s_2$.
This is an immediate consequence of the transitive property of equality of lengths.
By using these theorems we can maintain a minimal set of objects during discovery.
Fig. \[hexagon-pool\] shows the objects identified during the command `Discover(F)` for the input in Fig. \[hexagon1\]. The set of lines are not listed in the figure separately, but as a single entry at the bottom list of equally long segments. Also, some of the outputs are not particularly interesting features, such as lines with only point, circles with only three points, directions with only one line, or isolated equal length segments.
![The list of objects as shown in IntelliJ IDEA, a popular integrated development environment for Java[]{data-label="hexagon-pool"}](hexagon-pool)
Examples of Discover with selected theorems
===========================================
GeoGebra is a well-known and widely used software tool in education, with meaningful potential for using geometric discovery and exploration to teach elementary geometry. Even so, the range of mathematical knowledge is broad, including secondary school topics, international math competitions, and higher level mathematics. Below we examine selected theorems confirmed in the current implementation of Discover.
The diagonals of a parallelogram bisect each other
--------------------------------------------------
We already mentioned this simple theorem. The problem is shown in Fig. \[parallelogram1\]. With discovery on point $P_5$, the applicable theorems are reported in Fig. \[parallelogram2\].
![Output of the command `Discover(P_5)`[]{data-label="parallelogram2"}](parallelogram2)
Euler line
----------
The Euler line is a line determined from any triangle that is not regular. It passes through the orthocenter, the circumcenter and the centroid. The problem is shown in Fig. \[EulerLine1\].
![Euler line[]{data-label="EulerLine1"}](EulerLine1)
With discovery on point $P$, the relevant theorems are listed in Fig. \[EulerLine2\].
![Output of the command `Discover(P)`[]{data-label="EulerLine2"}](EulerLine2)
The Euler line theorem implicitly includes several simple theorems, including concurrency of the medians of a triangle ($J=K=L$, the generated points being the pairwise intersections of the medians), concurrency of the altitudes ($G=H=I$, these points being the pairwise intersections of the altitudes), and concurrency of the perpendicular bisectors of the altitudes ($P=Q=R$, pairwise intersections as above).
Nine-point circle
-----------------
The nine-point circle passes through nine significant points of an arbitrary triangle, namely:
- the midpoint of each side of the triangle,
- the foot point of each altitude,
- the midpoint of the line segment from each vertex of the triangle to the orthocenter.
The problem setting is shown in Fig. \[9pc1\].
![Problem setting for the nine-point circle[]{data-label="9pc1"}](9pc1)
With discovery on point $D$, the appropriate theorems are reported in Fig. \[9pc2\].
![Output of the command `Discover(D)`[]{data-label="9pc2"}](9pc3)
The nine-point circle theorem implicitly includes several other simple theorems. In addition, the graphical result suggests further theorems: segments $DJ$, $EK$, and $FL$ are congruent and concurrent; these three segments are also diameters of the nine-point circle; and their intersection designates the center of the nine-point circle. By using another discovery this can be confirmed.
A contest problem
-----------------
In 2010, at the 51st International Mathematics Olympiad in Astana, Kazakhstan, the following shortlisted problem was proposed by United Kingdom:
> Let $ABC$ be an acute triangle with $D$, $E$, $F$ the feet of the altitudes lying on $BC$, $CA$, $AB$ respectively. One of the intersection points of the line $EF$ and the circumcircle is $P$. The lines $BP$ and $DF$ meet at point $Q$. Prove that $AP=AQ$.
After constructing the according figure with GeoGebra Discovery (Fig. \[contest1\]),
![Problem setting for a shortlisted contest problem at IMO 2010[]{data-label="contest1"}](contest1)
we start discovery on point $Q$. The discovered theorems appear in Fig. \[contest2\]. We learn a few unexpected properties: $DP\parallel EQ$, and the concyclicity of points $C$, $D$, $P$, $Q$, and $A$, $F$, $P$, $Q$.
![Output of the command `Discover(Q)`[]{data-label="contest2"}](contest2)
Pappus’s hexagon theorem
------------------------
Consider two sets of collinear triplets $A$, $B$, $E$; and $C$, $D$, $F$. The intersection points $G=AD\cap BC$, $H=AF\cap CE$, $I=BF\cap DE$ are created. Pappus’s hexagon theorem (Fig. \[pappus1\]) claims that the points $G$, $H$ and $I$ are collinear (in general, after assuming certain non-degeneracy conditions).
![Problem setting for Pappus’s hexagon theoreom[]{data-label="pappus1"}](Pappus1)
With discovery on point $G$, the theorem is reported in Fig. \[pappus2\]. This final example is more commonly discussed at the university level, rather than in secondary school.
![Output of the command `Discover(G)`[]{data-label="pappus2"}](Pappus2)
\*
As a final note we highlight that the user interface for the geometric discovery is designed to be easy for non-experts as well. One does not need to use anything else but the mouse pointer to obtain all the information.
Discussion {#sec4}
==========
Trivial statements and theorems
-------------------------------
In Fig. \[rel-midline\] the collinearity of points $B$, $C$ and $D$ and of points $A$, $C$ and $E$ were not reported. This is intentional: by defining $D$ as the midpoint of $BC$ we implicitly assumed this collinearity, so it does not make any sense to reiterate this. Therefore, it seems useful to make a distinction between *trivial statements* and *theorems*.
The question of which properties are considered trivial or not is at some level a judgment call. In Fig. \[rel-midline\] most users may regard the information $BD=CD$ as trivial, with $D$ being the midpoint of $BC$. On the other hand, for beginners this information may still be useful.
At the moment GeoGebra Discovery maintains some background information if the obtained theorem is to be displayed or not. For example, in Fig. \[parallelogram2\], the collinearity of $P_2$, $P_4$ and $P_6$ is considered trivial and not displayed, but the fact that $P_5=P_6$ is presented as non-trivial. By considering both of these ideas, the collinearity of $P_2$, $P_4$ and $P_5$ could be considered either trivial or non-trivial—currently it is considered as trivial and not shown. The decision process for such questions should be clarified in the future.
Combinatorial explosion and computational complexity
----------------------------------------------------
Despite the large number of possible statements, the combinatorial complexity is still polynomial, because from a given set of input objects $P_1,P_2,\ldots,P_n$ we need to select just at most four objects (four objects are required to confirm concyclicity.) On the other hand, by using the classes of the equivalence relations, the number of statements to be checked can be decreased significantly.
For each possible statement, a numerical check is first performed. We assume that this is always successful when a generally true statement is about to check. Unfortunately, in reality this is not always the case, because for some exotic coordinates, the numerical check can be completely misleading. For example, some very large numbers can result in numerically unstable computations. Regardless, if a numerical check is positive, then the statement is added to the list of conjectures, but if it is negative, no conjecture is registered. As a consequence, while our implementation may miss some true statements (due to numerical errors), it will not output false statements.
For each conjecture, a symbolic check will be performed. If the symbolic check is positive, then the statement will be saved as a theorem. If the symbolic check is negative, then the statement will be removed from the list of conjectures. If the symbolic check cannot decide if a conjecture is true or false, the conjecture is removed from the list.
A special case of a conjecture is $P_1=P_2$ for each two geometric points. If this conjecture cannot be proven or disproven symbolically, then the discovery process will be completely stopped and the user will be notified that the construction must be redrawn in a different way—otherwise no output can be produced. This exception is required to keep the internal data consistent.
Symbolic checks usually require more time than numerical verifications. The underlying computation uses Gröbner bases that require at most double exponential time of the number of variables [@mayrmeyer82] according to the given figure. Usually, the number of variables are double the number of geometric points in the figure (since there are two coordinates for each).
GeoGebra internally sets 5 seconds for the maximal execution time of each symbolic check. After timeout the result of the symbolic check will be undecided.
User interface enhancements
---------------------------
GeoGebra is designed with a straightforward user interface that asks the users no questions if possible. However, its usability could be improved for situation when the user wishes to limit the output by filtering or excluding certain relationships.
Currently only points can be investigated. In a future version a set of points, segments, lines, circles or a set of these should be permitted as input.
Currently the computation process cannot be interrupted by the user. Given a large number of points in the figure, the calculation can be time consuming. For example, investigating the relationships of a regular 20-gon may require about 4 minutes on a modern personal computer (in our test a Lenovo ThinkPad E480 with 8$\times$i7, 16 GB RAM, Ubuntu Linux 18.04, was used). See Fig. \[20gon\] for the output.
![Output for `Discover(A)` in a regular 20-gon[]{data-label="20gon"}](20gon2 "fig:")\
The version that is based on GeoGebra Classic 5 performs better than the one on Classic 6—the latter is a web implementation of the GeoGebra application and uses a WebAssembly compilation of the computer algebra system Giac. Even if the code is reasonably fast as embedded code in a web page, this latter version underperforms the native technology: the same hardware is unable to handle the input of the regular 20-gon, and the browser tab crashes after 12 minutes of computation. (Google Chrome 83 was used for testing.)
Colors
------
At the moment a limited set of colors is used to highlight parallelism and congruence. In the future a pre-defined sequence of distinguishable colors should be added to GeoGebra Discovery—for example, at the moment in Fig. \[rel-hexagon\] the same black color is used to highlight different sets of parallel lines.
Perpendicular lines
-------------------
Perpendicular lines play an important role in elementary planar geometry. Their detection and presentation are not yet implemented in GeoGebra Discovery. Here we mention that the relationship of perpendicularity is *not* an equivalence, in contrast to the previous relationships defined in Section \[sec2\]. On the other hand, if $\vec{D}$ and $\vec{E}$ are directions, if $\ell\in\vec{D}$ and $m\in\vec{E}$, the relationship $\ell\perp m$ implies perpendicularity for all $\ell'\in\vec{D}$ and $m'\in\vec{E}$, that is, $\ell'\perp m'$.
It seems convenient to color perpendicular lines with the same color. So a rectangular grid can be observed for each pair of directions $\vec{D}$ and $\vec{E}$ whose representative lines are perpendicular, accordingly. Fig. \[grid\] shows an example that includes four rectangular grids for the parallel diagonals of a regular octagon.
![Four rectangular grids describing the parallel diagonals of a regular octagon[]{data-label="grid"}](grid "fig:")\
Angles
------
In a complex algebraic geometry setting, the study of angles is not as straightforward as investigating other objects. For a future version, however, this feature would be an important improvement.
By combining algebraic and pure geometric observations, however, simple theorems on angle equality could be easily detected. For example, Fig. \[contest2\] states that points $A$, $F$, $P$, $Q$ are concyclic. The inscribed angle theorem automatically implies $\angle QAP=\angle QFP$, among others.
Stepwise suggestions
--------------------
Prior research (see [@matematech p. 46]) proposed that collecting the interesting new objects in a figure could be done stepwise, similarly to GeoGebra’s former feature “special objects.” For our midline theorem example (Fig. \[midline1\]), this meant that after constructing the triangle $ABC$, and then midpoint $D$, the segments $BD$ and $CD$ were automatically shown by the system. The user could then accept these newly generated segments or remove them from the system. Then, by creating midpoint $E$, the system could show lines $AB$ and $DE$ to visualize parallelism.
Actually, the “special objects” feature was recently removed from GeoGebra after some negative feedback from the community—many users found this feature confusing. As a consequence, adding stepwise suggestions in GeoGebra Discovery remains a question for future research.
Benchmarks
----------
There is no benchmarking suite for the `Discover` command yet. This should be addressed in the next phase of the development.
Related work
============
We now discuss several projects that share some similarity to GeoGebra Discover but differ significantly in meaningful ways.
First of all, GeoGebra Discovery is not the first tool that systematically displays confirmed theorems in a geometric figure. We refer the reader to
- Zlatan Magajna’s *OK Geometry* [@Magajna2011] (available at [www.ok-geometry.com](www.ok-geometry.com)) and
- Jacques Gressier’s *Géométrix* (available at [geometrix.free.fr](geometrix.free.fr)).
These systems are available free of charge, but without the source code. On the other hand, GeoGebra Discovery focuses on an intuitive user interface and proofs in the most mathematical sense.
Second, we highlight that there is a growing interest in creating algorithms related to success completion of secondary school or undergraduate mathematics entrance exams. (See [@robot-china], [@text-diagram], [@japan-todai], among others.) Sometimes these projects rely significantly on techniques used in the underlying computational methods. Also, these projects are often related to artificial intelligence and Big Data rather than to computational mathematics.
Third, we mention a theoretical issue. The idea to store a geometric point only once if it is identical to another one was previously described in Kortenkamp’s work [@ulli99 9.3.1]. This concept is a main design element in the dynamic geometry software Cinderella, which never stores a geometric point twice if the two variants are identical in general.
GeoGebra has a different design concept by allowing the user an arbitrary number of identical points to be defined. From the theorem prover’s point of view, GeoGebra’s concept is more difficult to handle, and a kind of translation is required to have a different data structure by using the concepts from Section \[sec2\].
Also, we note that GeoGebra Discovery proves the truth in a different manner from Cinderella, with Cinderella using a probabilistic method, and GeoGebra Discovery literally *proving* all the deduced facts.
Conclusion
==========
We described a prototype of the `Discover` command that is available in an experimental version of GeoGebra, called GeoGebra Discovery. Our current implementation can be directly downloaded from <https://github.com/kovzol/geogebra/releases/tag/v5.0.591.0-2020Jul16>.
Our work is still in progress, as noted with the issues listed in Section \[sec4\].
Acknowledgments
===============
The `Discover` command is a result of a long collaboration of several researchers. The project was initiated by Tomás Recio in 2010, and several other researchers joined, including Francisco Botana and M. Pilar Vélez, to name just the most prominent collaborators. The development and research work was continuously monitored and supported by the GeoGebra Team. Special thanks to Markus Hohenwarter, project director of GeoGebra.
The work was partially supported by a grant MTM2017-88796-P from the Spanish MINECO (Ministerio de Economia y Competitividad) and the ERDF (European Regional Development Fund).
| {
"pile_set_name": "ArXiv"
} |
---
address: 'Massachusetts Institute of Technology, Cambridge, MA 02139'
author:
- 'Maurice H.P.M. van Putten'
title: 'Nonlinear wave equations for numerical relativity: towards the computation of gravitational wave forms of black hole binaries'
---
\#1\#2\#3\#4[[\#1]{} [**\#2**]{}, \#3 (\#4)]{}
Introduction
============
The advanced stages of construction of LIGO and VIRGO has stimulated some new developments in hyperbolic formulations for numerical relativity[@mvp1; @Abr]. These developments describe nonlinear gravitational wave motion in the late stages of spiral in of compact binaries of black holes and neutron stars. Binary black hole and neutron star systems remain today the prime candidates for astrophysical sources of gravitational radiation, and statistical calculations indicate their event rates to be more than a few per year at the level of the sensitivity of the initial LIGO [@ph; @na]. Other known burst sources with potentially detectable levels of gravitational radiation are the collapse of white dwarfs into neutron stars[@usov], and secularly unstable single neutron stars[@lai].
$\dot{f}(f)-$diagram for burst sources
======================================
The time evolution of the above mentioned three classes of burst sources of gravitational radiation can be illustrated (at the present level of our understanding) in an $\dot{f}(f)-$diagram (Figure 1). Recall that compact binaries show an initial spiral in set by the Newtonian relationship $$\begin{aligned}
\dot{f}\sim f^{\frac{11}{3}},\end{aligned}$$ where $\dot{f}=\mbox{d}f/\mbox{d}t$. The frequency $f$ refers to the base frequency of the gravitational wave-form, initially of the “chirp," and finally of the quasi-normal mode ring down. During the intermediate merger phase, $f$ and $\dot{f}$ either follow by interpolation, or from $f=\frac{1}{\pi}\dot{\phi}$ with $\phi$ the orbital phase. Numerical relativity on black hole binaries specifically targets the merger phase, where post-Newtonian calculations break down. The present challenges are found to be the computation of stable, multiple orbits and the treatment of horizon boundary conditions.
[**FIGURE 1.**]{} [ Three cartoons of burst sources of gravitational waves in an $\dot{f}(f)$-diagram. The [**top**]{} diagram shows the three phases of binary coalescence: the chirp (with the Newtonian result $\dot{f}(f)\sim
f^{11/3}$), the merger phase, and the ring down phase with quasi-normal mode frequency $f_{QNR}\approx865 \frac{37M}{M}$ Hz (see, $e.g.$, Flanagan & Hughes[@fla]). The merger phase of two black holes and two neutron stars is expected to be distinct, but in as yet unknown ways. Sketched is the expected trend with increasing spin $S$ of the binary objects. Order of magnitude estimates[@fla; @tho; @sch] of the radiated energy in the merger phase is also indicated. (A possible break up and disk formation in neutron star mergers has been by Zhuge[@zhu], and could also be included here.) The [**middle**]{} diagram sketches the expected gravitational wave form resulting from a dynamical instability in an (accretion induced) collapse scenario, such as the collapse of a white dwarf to a neutron star. The arrow indicates the direction of time in this trajectory. The first stage shows a rapid sweep of increasing frequency, followed by a phase of rapid damping due to gravitational radiation. This scenario assumes the ratio $\beta$ of kinetic energy to gravitational energy to exceed the value which triggers a dynamical instability ($\beta>0.27$). It has been suggested as a model for GRBs by Usov[@usov]. The [**lower**]{} diagram sketches a translation of results on the secular instability of single neutron stars, $\beta_{\mbox{sec}}(=0.1375)
<\beta<\beta_{\mbox{dyn}}(=0.27)$, of Lai & Shapiro[@lai] into the $\dot{f}(f)-$diagram. The secular instability describes gravitational radiation driven transitions from unstable Jacobi ($J_u$) ellipsoids to either stable Mclaurin spheroids ($M_s$) or stable Dedekind ellipsoids ($D_s$). The Jacobi ellipsoids are rotation $(\Omega)$ dominated (rel. vorticity), and the Dedekind ellipsoids are vorticity ($\omega$) dominated (rel. rotation) stars. The trajectories are extended to the left, where the shape is in counter rotation relative to the fluid’s vorticity. Note that the end point of the trajectory in the middle diagram connects to $J_u$ of the lower diagram.]{}
Nonlinear wave equations for relativity
=======================================
Fully covariant, nonlinear wave equations have recently been obtained general relativity in the tetrad approach. They are Yang-Mills type equations for the Utiyama-connections $\omega_{a\mu\nu}=(e_\mu)_c\nabla_a(e_\nu)^c$ (for a tetrad field $\{(e_\mu)_a\}_{\mu=1}^4$) in the Lorentz gauge $\nabla^a\omega_{a\mu\nu}$. In vacuo, they reduce to $$\begin{aligned}
%\hat{\nabla}_c\hat{\nabla}^c
\hat{\Box}
\omega_{a\mu\nu}-
[\omega^c,\nabla_a\omega_{c}]_{\mu\nu}=0,\end{aligned}$$ were $\hat{\Box}=
\hat{\nabla}_c\hat{\nabla}^c$ is the $SO(3,1)$ gauged wave-operator. The full set of equations further involves the equations of structure, which can be integrated in the form of $$\begin{aligned}
(e_\mu)_a(\tau)=\Lambda_\mu^{\hskip.1in\nu}(\tau,\tau_0)(e_\mu)_c(\tau_0)
+\int_{\tau_0}^\tau \Lambda_\mu^{\hskip.1in\nu}(\tau,s)\hat{\partial}_b
N_\nu(s)\mbox{d}s,\end{aligned}$$ where the spatial coordinates have been suppressed. Here, the $$\begin{aligned}
N_\mu=(e_\mu)_\tau\end{aligned}$$ are the [*tetrad lapse functions*]{}, which are algebraically equivalent to the Hamiltonian lapse and shift functions. Strict hyperbolicity is maintained with arbitrary $N_\mu$, provided they define space-like foliations. There is full generality in this regard. By comparison, the otherwise closely related hyperbolic formulation of Abrahams $et$ $al.$[@Abr] shows some restrictions on the Hamiltonian lapse function for maintaining strict hyperbolicity. (This difference was left unnoticed in the comparison by Finn[@Finn]).
A numerical implementation is given in van Putten,[@mvp2] and applied to a nonlinear Gowdy-wave in a Gowdy $T^3$-cosmology. Stable and accurate results are obtained with proper second-order convergence by comparison with analytical solutions. No indication of any “coordinate shocks" as reported by Alcubierre & Masso [@Alc] have been observed in these numerical experiments.
References {#references .unnumbered}
==========
[99]{}
E.S. Phinney,. R. Narayan, T. Piran & A. Shemi, . K.S. Thorne, gr-qc/9506086 (1995). B.F. Shutz,. X. Zhuge, J.M. Centrella & L.W. MacMillan, gr-qc/9610039 (1996). D. Lai & S.L.Shapiro,. E.E. Flanagan & S.A. Hughes, gr-qc/9701039 (1997). V.V. Usov,. M.H.P.M. van Putten & D.M. Eardley, . M.H.P.M. van Putten, . L.S. Finn, gr-qc/9603004, (1996). A. Abrahams, A. Anderson, Y. Choquet-Bruhat & J.W. York, ; in these proceedings. M. Alcubierre, J. Masso, gr-qc/9709024, (1997).
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'Michael Blaser$^{*}$, Tristan Rivière[^1]'
title: A Minimality property for entropic solutions to scalar conservation laws in $1+1$ dimensions
---
Introduction
============
We consider solutions to the following equation $$\label{scl}
\left .
\begin{array}{rclcl}
\partial_{t}u+\partial_{x}f(u)& =&0&\mbox{in}&\mathbb{R}_+\times\mathbb{R}\,, \\
u(x,0) & =&u_{0}(x)\,,
\end{array}
\right \}$$ where the flux $f$ is strictly convex ( $f''\ge c>0$ )and the initial date $u_0\in L^{\infty}$. It is well known, that, even for smooth initial data, the classical solution can cease to exist in finite time, due to the possible formation of shocks (see Chapter 4.2 in [@Da]). Therefore one has to consider weak solutions of (\[scl\]), i.e. solutions, which satisfy (\[scl\]) in the distributional sense. However it turned out, that, for a given initial data, the space of weak solutions is huge (see Chapter 4.4 in [@Da]). Therefore additional conditions have to be imposed to single out the physical relevant weak solutions in some models.
In 1957 Oleinik proved in [@Ol] uniqueness of bounded weak solutions, which satisfy almost everywhere her ’*E-condition*’ $$\label{e_condition}
u(y,t)-u(x,t)\leq \frac{y-x}{ct}\,,\quad\mbox{for}\quad x<y\,,t>0\,,$$ where $c=\inf f''$. A immediate consequence of this condition (\[e\_condition\]) is a spectacular regularization phenomena. Oleinik proved, that for bounded measurable initial data, the weak solution satisfying almost everywhere (\[e\_condition\]) becomes immediately locally BV in space and locally in space-time in the complement of the initial line .
A more powerful approach was given by Kruzhkov in [@Kr], where he replaces condition (\[e\_condition\]) by a family of integral inequalities. This approach covers also cases, where $f$ is non-convex and the space dimension is bigger than one. However in the case of convex fluxes one can show, that his *entropy condition* is equivalent to Oleinik’s *E-condition* (see Chapter 8.5 in [@Da]). More precisely for $u_0\in L^{\infty}$ he proved existence and uniqueness of weak solutions satisfying the *entropy condition*: He considers the family of convex entropy flux pairs $(\eta_a, \xi_a)_{a\in\mathbb{R}}$, where $$\label{kruzhkov_family}
\eta_{a}(u)=(u-a)^+\quad\mbox{and}\quad \xi_a(u)=\operatorname{sign}(u-a)^+(f(u)-f(a))\,,$$ and $w^+$ stands for $\max\{w, 0\}$. Then an entropy solution is a bounded function $u$, which satisfies (\[scl\]) in the sense of distributions and $$\label{e_inequality}
\partial_t\eta_a(u)+\partial_x\xi_a(u)\leq0\,.$$ Equivalently one can replace the one parameter family $(\eta_a, \xi_a)_{a\in\mathbb{R}}$ and assume, that (\[e\_inequality\]) is fulfilled for all convex $\eta$ with corresponding entropy flux $\xi$, which is defined by $\xi=\int \eta'f'$. As a consequence of this one can show, if the initial data $u_0$ is in BV, that $u$ is in BV for all later times.
Let $a\wedge b$ denote $\min\{a,b\}$. Let $u\in L^{\infty}(\mathbb{R}\times[0,T))$ be a weak solution of (\[scl\]), such that $$m(x,t,a)=\partial_{t}(u\wedge a)+\partial_x f(u\wedge a)\in\mathcal{M}_{loc}(\mathbb{R}\times\mathbb{R}_+\times\mathbb{R})\,$$ where $\mathcal{M}$ denotes the space of Radon measures. One can define the absolute value of the entropy production over a set $\Omega\subset\mathbb{R}\times\mathbb{R}_+$ as being $$\label{entropy_production}
EP=\int_{\mathbb{R}}|m|(\Omega,a)\,da\,.$$ In the case of $u$ being an entropy solution and hence in BV, the measure $m(x,t,a)$ and therefore the entropy production of $u$ simplifies to $$\label{ep_entropy_solution}
EP=\int_{\Omega}\Delta(u^{+},u^{-})\mathcal{H}^1\res J_u\,,$$ where $J_u$ denotes the rectifiable set of jump points of $u$, $u_+$ and $u_-$ are respectively the left and right approximate limits of $u$ for some orientation of $J_u$ and $$\label{Delta}
\Delta(a,b)=\frac{(a-b)^2\left[\frac{f(a)+f(b)}{2}\right]-(a-b)\int_{a}^{b}f(s)\,ds}{\left[(a-b)^2+(f(a)-f(b))^2\right]^{\frac12}}\,.$$ It is natural to compare the different entropic productions of the weak solutions to (\[scl\]) - $BV$ or not $BV$ ! - and to ask the following questions : [*does there exists a weak solution which minimizes the entropy production and, if so, what properties does a minimizer of (\[entropy\_production\]) have*]{}.
In this work we provide a partial answer to this question. We show a weak solution of (\[scl\]) whose entropy production increases in time less, than any other weak solution’s entropy production, has to be the entropy solution. Precisely
Let $\mathcal{W}$ denote the set of defect measures induced by a weak solution of (\[scl\]), i.e. $$\mathcal{W}:=\left \{
\begin{array}{c}
m(x,t,a)\in \mathcal{M}_{loc} ~\text{s.t.}~m(x,t,a)=\partial_{t}(u\wedge a)+\partial_{x}f(u\wedge a),\\
\text{where}~u\in L^{\infty}~\text{is a weak sol. of (\ref{scl}).}\end{array}\right\}$$ Our main result in the present work is the following.
\[minimality\_theorem\] Let $f\in C^{2}(\mathbb{R})$ such that $f''\geq c>0$ and $$\label{infinity_behaviour}
\lim_{|x|\rightarrow\infty}f(x)=\infty\,.$$ Moreover let $u_{0}\in L^{\infty}(\mathbb{R})$ be compactly supported. Let $u\in L^{\infty}(\mathbb{R}\times [0,T))$ be an arbitrary weak solution of (\[scl\]), such that $m(x,t,a)=\partial_t(u\wedge a)+\partial_x f(u\wedge a)$ is locally a Radon measure in ${\R}\times [0,T)\times {\R}$ . Assume the ”entropy production” $m$ satisfies $$\label{minimality}
\int_{\mathbb{R}}\left|m\right|\left(\mathbb{R}\times(0,\bar{t}),a\right)\,da\leq \int_{\mathbb{R}}\left|q\right|\left(\mathbb{R}\times(0,\bar{t}),a\right)\,da~\forall~q\in\mathcal{W}~\text{and}~\forall \bar{t}\in(0,T)\,.$$ Then $u$ is the entropy solution, i.e. satisfies (\[e\_condition\]) and equivalently (\[e\_inequality\]).
A similar criteria in a more restrictive setting is considered by Dafermos in Chapter 9.7 of [@Da]. He considers weak solutions $u$ of (\[scl\]) with initial data $$\label{riemann}
u_0(x)=\left\{\begin{array}{lcl}
u_l&\mbox{if}&x<0\,,\\
u_r&\mbox{if}&x>0\,.
\end{array}\right.$$ Since the conservation law is invariant under Galilean transformations it is reasonable in this case to consider weak solutions of the form$$u(x,t)=v\left(\frac{x}{t}\right)\,.$$ One can then define $\omega=\frac{x}{t}$ and consider $v$ as a function only dependent of $\omega$, i.e. $v=v(\omega)$. Then $v(\omega)$ satisfies the ordinary differential equation $$\frac{d}{d\omega}\left(f(v(\omega))-\omega v(\omega)\right)+v(\omega)=0$$ in the sense of distributions and has prescribed end states $$\lim_{\omega\rightarrow-\infty}v(\omega)=u_l\quad\mbox{and}\quad\lim_{\omega\rightarrow\infty}v(\omega)=u_r\,.$$ Furthermore it is assumed that $v$ is in $BV$ and denotes $J_v$ the set of jump points $\omega$ for $v$. For a given entropy-entropy flux pair $(\eta(u),\xi(u))$ C. Dafermos defines the combined entropy of the shocks in $v$ by $$\label{combined_entropy}
\mathcal{P}_{v}=\sum_{\omega\in J_v}\left\{\xi(v(\omega+))-\xi(v(\omega-))-\omega\left[\eta(v(\omega+))-\eta(v(\omega-))\right]\right\}\,.$$ Furthermore he introduces the rate of change of the total entropy production $$\dot{\mathcal{H}}_v=\frac{d}{dt}\int_{-\infty}^{\infty}\eta(u(x,t))\,dx=\int_{-\infty}^{\infty}\eta(v(\omega))\,d\omega\,,$$ for entropy-entropy flux pairs $(\eta,\xi)$ such that $\eta(u_l)=\eta(u_r)=0$.
He shows that in this simple case the rate of change of the total entropy and the entropy productions are related to each other by $$\dot{\mathcal{H}}_v=\mathcal{P}_v+\xi(u_l)-\xi(u_r)\,.$$ We can now relate the combined entropy $\mathcal{P}_v$ to our entropy productions (\[entropy\_production\]). To do so one notices, that for a $T>0$, $\psi\in C^{\infty}_{c}(\mathbb{R}\times(0,T))$ and an entropy-entropy flux pair $(\eta,\xi)$ we get after a change of variable $$\label{relation_entropy_production}
\begin{split}
\int_{\mathbb{R}\times[0,T]}&\eta(u(x,t))\partial_t\psi+\xi(u(x,t))\partial_x\psi\,dx\,dt\\
&=\int_{\mathbb{R}}\xi(v(\omega))\frac{d}{d\omega}\phi(\omega)-\eta(\omega)\frac{d}{d\omega}\left[\omega\cdot\phi(\omega)\right]\,d\omega\\
&=\sum_{\omega\in J_v}\phi(\omega)\left\{\xi(v(\omega+))-\xi(v(\omega-))-\omega\left[\eta(v(\omega+))-\eta(v(\omega-))\right]\right\}\,,
\end{split}$$ where $$\label{def_phi}
\phi(\omega)=\int_0^T\psi(\omega t,t)\,dt\,.$$ For a jump point $\omega$ we write $v_+=v(\omega+)$ and $v_-=v(\omega-)$, then taking the particular entropy-entropy flux pair $(\eta_a,\xi_a)$, defined in (\[kruzhkov\_family\]) and using identity (\[relation\_entropy\_production\]) gives $$\begin{gathered}
\label{relation_entropy_production1}
\int_{\mathbb{R}}\int_{\mathbb{R}\times[0,T]}\psi(x,t)dm(x,t,a)\,da\\
=-\int_{\mathbb{R}}\sum_{\omega\in J_v}\phi(\omega)\left\{\xi_a(v_+)-\xi_a(v_-))-\omega\left[\eta_a(v_+)-\eta_a(v_-)\right]\right\}\,da\\\end{gathered}$$ A short calculation reveals $$\begin{gathered}
\label{relation_entropy_production2}
\int_{\mathbb{R}}\xi_a(v_+)-\xi_a(v_-)-\omega\left[\eta_a(v_+)-\eta_a(v_-)\right]\,da\\=-\frac12\omega\left(v_-^2-v_+^2\right)-\int_{v_-}^{v_+}sf'(s)\,ds\,,\end{gathered}$$ where we used the Rankine-Hugoniot condition for self similar solutions: $$f(v_+)-f(v_-)=\omega(v_+-v_-)\,.$$ Applying (\[relation\_entropy\_production2\]) in (\[relation\_entropy\_production1\]) gives $$\begin{gathered}
\label{relation_entropy_production3}
\int_{\mathbb{R}}\int_{\mathbb{R}\times[0,T]}\psi(x,t)dm(x,t,a)\,da\\
=\sum_{\omega\in J_v}\phi(\omega)\left\{\xi(v(\omega+))-\xi(v(\omega-))-\omega\left[\eta(v(\omega+))-\eta(v(\omega-))\right]\right\}\,da\,,\end{gathered}$$ where $(\eta,\xi)$ is the entropy-entropy flux pair $$\label{part_entropy_pair}
\eta(v)=\frac12v^2\quad\mbox{and}\quad \xi(v)=\int_0^v sf'(s)\,ds\,.$$ From (\[relation\_entropy\_production3\]) we deduce with (\[def\_phi\]) $$\label{relation_entropy_production4}
\frac{1}{T}\int_{\mathbb{R}}m(\mathbb R\times[0,T],a)\,da=\mathcal{P}_v\,$$ where the combined entropy production $\mathcal{P}_v$ is taken for the entropy-entropy flux pair defined in (\[part\_entropy\_pair\]). Since $T>0$ is arbitrary and $\mathcal{P}_v$ independent of $T$ it follows from (\[relation\_entropy\_production4\]) $$\label{relation_entropy_production5}
\frac{d}{dt}\int_{\mathbb{R}}m(\mathbb R\times[0,t],a)\,da=\mathcal{P}_v\quad\mbox{for all}\quad t>0\,,$$ which finally relates (\[combined\_entropy\]) to (\[entropy\_production\]).
Then a weak solution $u=v\left(\frac{x}{t}\right)$ of (\[scl\]) with initial data (\[riemann\]) is said to satisfy the *entropy rate admissibility criterion* if it satisfies the following optimality criterion of the entropy production $$\mathcal{P}_v\leq\mathcal{P}_{\tilde{v}}$$ or equivalently $$\dot{\mathcal{H}}_v\leq\dot{\mathcal{H}}_{\tilde{v}}$$ holds, for any other weak solution $\tilde{u}=\tilde{v}\left(\frac{x}{t}\right)$ of (\[scl\]) with initial condition (\[riemann\]).
Using (\[relation\_entropy\_production5\]) one can express the entropy rate admissibility criterion for the particular entropy-entropy flux pair in (\[part\_entropy\_pair\]) in terms of the entropy production (\[entropy\_production\]): A solution $u=v\left(\frac{x}{t}\right)$ with initial data (\[riemann\]) and defect measure $m(x,t,a)$ satisfies *entropy rate admissibility criterion* if $$\label{entropy_rate}
\frac{d}{dt}\int_{\mathbb{R}}m(\mathbb R\times[0,t],a)\,da\leq \frac{d}{dt}\int_{\mathbb{R}}\tilde{m}(\mathbb R\times[0,t],a)\,da\quad\mbox{for all}\quad t>0\,$$ for any other weak solution $\tilde{u}=\tilde{v}\left(\frac{x}{t}\right)$ of (\[scl\]) with initial condition (\[riemann\]) and defect measure $\tilde{m}(x,t,a)$. One can also integrate (\[entropy\_rate\]) and obtains the equivalent condition $$\label{entropy_rate2}
\int_{\mathbb{R}}m(\mathbb R\times[0,t],a)\,da\leq \int_{\mathbb{R}}\tilde{m}(\mathbb R\times[0,t],a)\,da\quad\mbox{for all}\quad t>0\,.$$
Therefore (\[entropy\_rate\]) and (\[entropy\_rate2\]) show, that the entropy rate admissibility criterion can be interpreted as a growth condition of the entropy production (\[entropy\_production\]), which is similar to the growth condition (\[minimality\]) in Theorem \[minimality\_theorem\]. In Chapter 9.5 of [@Da] it is proved:
[@Da] A weak solution $u$ of (\[scl\]) with initial data (\[riemann\]) satisfies the entropy rate admissibility criterion for an entropy-entropy flux pair $(\eta,\xi)$ if and only if $u$ satisfies the E-condition (\[e\_condition\]).
Again by (\[entropy\_rate\]) and (\[entropy\_rate2\]) one sees, that this Theorem establishes, similar as in Theorem \[minimality\_theorem\], a connection between growth rate of the entropy production (\[entropy\_production\]) and entropy admissibility conditions (\[e\_condition\]) and (\[e\_inequality\]). In Chapter 9.5 there is also an extension of this theorem in the case of strictly hyperbolic systems.
Another results relating an optimality criterion to entropic solution is given by A. Poliakovsky in [@Po]. For $u:\mathbb{R}^n\times[0,T]\to\mathbb{R}^k$ he considers a family of energy functionals $$\label{energy1}
I_{\varepsilon, f}(u)=\frac12\int_0^T\int_{\mathbb{R}^n}\left(\varepsilon|\nabla_x u|^2+\frac{1}{\varepsilon}|\nabla_x H|^2\right)\,dx\,dt+\frac12\int_{\mathbb{R}^n}|u(x,T)|^2\,dx\,$$ where $$\Delta_x H_u=\partial_t u+\operatorname{div}_x f(u)\,.$$ Under certain assumptions on the flux $f$ he shows, that there exists a minimizer to $$\inf \left\{ I_{\varepsilon,f}(u):~u(x,0)=u_0(x)\right\}$$ and this minimizer satisfies $$\left.\begin{array}{rcll}
\partial_t u+\operatorname{div}_x f(u)&=&\varepsilon\Delta_x H_u&\forall(x,t)\in\mathbb{R}^n\times(0,T)\,,\\
u(x,0)&=&u_0(x)&\forall x\in\mathbb{R}^n\,.
\end{array}\right\}$$ In the particular case $k=1$, he calculates the $\Gamma$-limit of (\[energy1\]) as $\varepsilon\to0^+$ and finds an alternative variational formulation of the admissibility criterion for the particular solutions to the scalar conservation laws that can be achieved by this relaxation procedure.
The result of A.Poliakovsky has been inspired by previous works establishing a link between some variational optimality condition of a relaxed problem and the entropy condition at the limit. Among these works we can quote [@RS1], [@RS2] and [@ALR]. Let us describe the results established in this 3 works here :
We consider for a bounded domain $\Omega\subset\mathbb{R}^2$ the space $\mathcal{M}_{div}(\Omega)$, which consists of unit vectorfields $u$ such that $u=e^{i\varphi}$ for a $\phi\in L^{\infty}(\Omega,\mathbb{R})$ and $\operatorname{div}e^{i\varphi\wedge a}$ is a Radon measure over $\Omega\times {\R}$. This space $\mathcal{M}_{div}$ was introduced by S. Serfaty and the second author in [@RS1] and [@RS2] in connection to a problem related to micromagnetism. We give here a brief description. Let $\Omega$ be a bounded and simply connected domain, for $u\in W^{1,2}(\Omega,\mathbb{S}^1)$ and a $\varepsilon>0$ we consider $$\label{energy}
E_{\varepsilon}(u)=\varepsilon\int_{\Omega}|\nabla u|^2 + \frac{1}{\varepsilon}\int_{\mathbb{R}^2}|H|^2\,,$$ where $H=\nabla(G*\hat{u})\,, \hat{u}=u$ on $\Omega$ and $\hat{u}=0$ in $\Omega^c$ and $G$ is the kernel of the Laplacian on $\mathbb{R}^2$.
It was proved in [@RS1], [@RS2] that from any sequence $u_{\varepsilon_n}\in W^{1,2}(\Omega,\mathbb{S}^1)$ such that $\varepsilon\to0$ and $E_{\varepsilon_n}(u_{\varepsilon_n})<C$ one can extract a subsequence $u_{\varepsilon_{n'}}$ such that $\varphi_{\varepsilon_{n'}}$ converges strongly in $L^p(\Omega)$ for any $p<\infty$ to a limit $\varphi$ such that $e^{i\varphi}=u\in \mathcal{M}_{div}(\Omega)$. Furthermore the authors are conjecturing that the $\Gamma$-Limit should be given by the following functional $E_0$ over ${\mathcal M}_{div}(\Omega)$ : $$E_0(u):=2\int_{a\in\mathbb{R}}\left|\operatorname{div}\left(e^{i\varphi\wedge a}\right)\right|(\Omega)\,da$$ Part of the $\Gamma-$convergence has been proved as they established in one hand the following inequality $$E_0(u):=2\int_{a\in\mathbb{R}}\left|\operatorname{div}\left(e^{i\varphi\wedge a}\right)\right|(\Omega)\,da\leq\liminf E_{\varepsilon_{n'}}(u_{\varepsilon_{n'}})$$ and in the other hand that $$\label{inf_energy}
\lim_{\varepsilon\to0}\inf_{u\in W^{1,2}} E_{\varepsilon}(u)=2\inf_{u\in\mathcal{M}_{div}(\Omega)}\int_{a\in\mathbb{R}}\left|\operatorname{div}\left(e^{i\varphi\wedge a}\right)\right|(\Omega)\,da=2|\partial\Omega|\,,$$ where $|\partial\Omega|$ is the perimeter of the set $\Omega$. One can prove (see [@RS1]), that the infimum on the right hand side is achieved by $u=-\nabla^{\perp}\operatorname{dist}(\cdot,\partial\Omega)\in\mathcal{M}_{div}(\Omega)$. The function $g=\nabla^{\perp}\operatorname{dist}(\cdot,\partial\Omega)$ is the viscosity solution of $$\label{eikonal}
\left.\begin{array}{rclcl}
|\nabla g|-1&=&0&\mbox{on}&\Omega\,,\\
g&=&0&\mbox{on}&\partial\Omega\,.
\end{array}\right\}$$ A question, which was left open in [@RS1] and [@RS2] was to describe the possible limits $u$ of minimizing sequence of (\[energy\]). It was conjectured that $u=\pm\nabla^{\perp}dist(\cdot,\partial\Omega)$ are the only possible limits of sequences of minimizers. A positive answer to this conjecture has been given in [@ALR]. Precisely, in [@RS2] it is proved that the limit $u$ of a minimizing sequence of (\[energy\]) satisfies the [*entropy condition*]{} $$\label{positive_limit}
\operatorname{div}e^{i\varphi\wedge a}\geq0\quad\mbox{for all}\quad a\in\mathbb{R}$$ or $\operatorname{div}e^{i\varphi\wedge a}\leq0\quad\mbox{for all}\quad a\in\mathbb{R}$. Then in [@ALR] the following result is established
[@ALR] Let $u=-\nabla^{\perp}g$ be a divergence free unit vector-field in the space $\mathcal{M}_{div}(\Omega)$. The entropy condition (\[positive\_limit\]) holds if and only if $g$ is a viscosity solution of (\[eikonal\]) and therefore $g$ is locally semiconcave in $\Omega$ and $u\in BV_{loc}(\Omega,\mathbb{S}^1)$.
Therefore, as a conclusion, one deduces the following equivalences for this particular problem $$\begin{array}{c}
\displaystyle\mbox{viscosity solution to (\ref{eikonal})}\quad\Longleftrightarrow\quad \mbox{ entropy condition (\ref{positive_limit}) }\\[5mm]
\Longleftrightarrow \mbox{ minimality of the entropy production (\ref{inf_energy})}\quad .
\end{array}$$
The paper is organized as follows: First, in section 2, we establish some technical preliminary results. Then in Section \[section\_blow\_up\] we will show, that the measure $$\int_{\mathbb{R}}m(x,t,a)\,da$$ has no points with strictly negative density, outside possibly a set of 1-dimensional measure $0$, i.e. we claim $$\label{uuu}
\lim_{r\rightarrow 0^+}\frac{1}{r}\int_{\mathbb{R}}m\left(B_{r}((x_{0},t_{0})),a\right)\,da\geq 0\quad\mbox{ for }{\mathcal H}^1\mbox{ a.e. }~ (x_{0},t_{0})\in\mathbb{R}\times(0,T).$$ In the last section, using an argument similar to the one used to prove the main result in [@ALR], we deduce that the non negativity condition (\[uuu\]) implies that $u$ is entropic.
Preliminary results {#preliminary}
-------------------
In this section we define a notion of weak entropy solutions (see Definition \[def\_entropy\_solution\]) of scalar conservation laws on domain of trapezoidal shape (see (\[definition\_boundary\]). Afterward we will prove Lemma \[properties\_of\_entropy\_solutions\], which roughly says that for that kind of entropy solutions the same properties hold as in the classical case. We will use this results in Section 2.2 and Section 2.3.
For $0<t_{1}<t_{2}<T$ and a $\delta>0$ we define the set $$\label{the_set_gamma}
\Gamma^{t_{2}}_{t_{1}}:=\{ (x,t) |~t_{2}> t> \gamma(x,t_{1})\}$$ where $$\label{definition_boundary}
\gamma(x,t):=\left\{\begin{array}{ll} t-\hat{\lambda}(x+\delta))&\text{if}~x\leq -\delta\,, \\
t&\text{if}~|x|\leq\delta\,,\\
t+\hat{\lambda}(x-\delta))&\text{if}~x\geq\delta\,.\end{array}\right.$$ for a constant $0<\hat{\lambda}\leq1$. Further we set $$\Lambda_{t_{1}}^{t_{2}}:=\{(x,t)|~(x,t)=(x,\gamma(x,t_{1}))~\text{and}~t_{1}\leq t<t_{2} \}\,.$$
\[gamma\]
As for mentioned we define now a notion of weak respective entropy solution on the domain $\Gamma_{t_{1}}^{t_{2}}$
\[def\_entropy\_solution\] For a $v_{1}\in L^{\infty}(\Lambda_{t_{1}}^{t_{2}})$ we say that $v\in L^{\infty}(\Gamma_{t_{1}}^{t_{2}})$ is weak solution of $$\label{modified_scl}
\left.\begin{array}{rclcl}
\partial_{t}v+\partial_{x}f(x)&=&0&\mbox{in}&\Gamma_{t_{1}}^{t_{2}}\,,\\
v&=&v_{1}&\mbox{on}&\Lambda_{t_{1}}^{t_{2}}\,,
\end{array}\right\}$$ if for all $\psi\in C_{c}^{\infty}(\mathbb{R}\times [0,t_{2}))$ $$\label{weak_formulation}
\int_{\Gamma_{t_{1}}^{t_{2}}}v\partial_{t}\psi+f(v)\partial_{x}\psi\,dx\,dt+\int_{\Lambda_{t_{1}}^{t_{2}}}
\psi\begin{pmatrix} v_1 \\
-f(v_1)\end{pmatrix}\cdot\tau\,d\sigma=0$$ holds, where $\tau$ is the unit tangent vector of $\Lambda_{t_{1}}^{t_{2}}$. Furthermore we say that $v\in L^{\infty}(\Gamma_{t_{1}}^{t_{2}})$ is an entropy solution of (\[modified\_scl\]), if $v$ additionally satisfies $$q(x,t,a):=\partial_{t}v\wedge a+\partial_{x}f(v\wedge a)\in\mathcal{M}_{loc}\quad\mbox{and}\quad q(x,t,a)\geq0\,.$$
A priory it is unclear if, for an arbitrary boundary condition $v_{1}\in L^{\infty}(\Lambda_{t_{1}}^{t_{2}})$, the conservation law (\[modified\_scl\]) possess a weak solution or not. We can however prove the following proposition.
\[existence\] Let $v_1\in L^{\infty}(\mathbb{R}\times[0,T))$ be a weak solution of (\[scl\]). Then for all $0<\hat{\lambda}\leq1$ and for almost every $t_{1}\in(0,T)$ and all $t_2\in(t_1,T)$ the problem $$\left.\begin{array}{rclcl}
\partial_{t}v+\partial_{x}f(v)&=&0&\mbox{in}&\Gamma_{t_{1}}^{t_{2}}\,,\\
v&=&v_1&\mbox{on}&\Lambda_{t_{1}}^{t_{2}}\,,
\end{array}\right\}$$ has an entropy solution in the sense of Definition \[def\_entropy\_solution\].
The basic idea for proving Proposition \[existence\] is to use the correspondence between weak solutions of (\[scl\]) and viscosity subsolutions of $$\label{hamilton_jacobi}
\left. \begin{array}{rcl}
\partial_t g+f(\partial_x g)&=&0\,,\\
g(x,0)&=&g_0(x)\,.
\end{array}\right\}$$ Before we are going to prove our assertion, we briefly repeat the definitions of viscosity sub- and supersolutions. We say that $g$ is a viscosity solution of (\[hamilton\_jacobi\]), if for any point $(x_0,t_0)\in\mathbb{R}\times(0,T)$ and for any $\psi \in C^1(\mathbb{R}^2)$ such that $g-\psi$ attains its maximum in $(x_0,t_0)$ the following inequality holds $$\partial_t \psi(x_0,t_0)+f(\partial_x\psi(x_0,t_0))\leq0\,.$$ Similarly we say, that $g$ is a viscosity supersolution of (\[hamilton\_jacobi\]), if for any point $(x_0,t_0)\in\mathbb{R}\times(0,T)$ and for any for any $\psi \in C^1(\mathbb{R}^2)$ such that $g-\psi$ attains its minimum in $(x_0,t_0)$ the following inequality holds $$\partial_t \psi(x_0,t_0)+f(\partial_x\psi(x_0,t_0))\geq0\,.$$ We say that $g$ is a viscosity solution of (\[hamilton\_jacobi\]), if $g$ is both a sub- and supersolution. Theorem 2 in [@CH] establishes a correspondence between weak solutions of (\[scl\]) and viscosity subsolutions of (\[hamilton\_jacobi\]).
\[correspondence\] Let $u\in L^{\infty}(\mathbb{R}\times[0,t))$ be a weak solution of (\[scl\]). Then there exists a $g\in W^{1,\infty}(\mathbb{R}\times[0,T))$ which satisfies (\[hamilton\_jacobi\]) almost everywhere and is such that $u(x,t)=\partial_x g(x,t)$ and $u_0=\partial_x g(x,0)$ for almost every $x\in\mathbb{R}$.
Let $v_1\in L^{\infty}(\mathbb{R}\times[0,T))$ be a weak solution of (\[scl\]); then, according to Theorem \[correspondence\], there exists $g_1\in W^{1,\infty}(\mathbb{R}\times[0,T))$, which solves (\[hamilton\_jacobi\]) almost everywhere. By Fubini’s Theorem we can choose $t_1$ such that both $\partial_t g_1$ and $\partial_x g_1$ are in $L^\infty(\Lambda_{t_1}^{t_2})$ and such that $$\label{choice_of_t1}
\int_{\Lambda_{t_1}^{t_2}}\partial_{t}g_1+f(\partial_x g_1)\,d\sigma=0\quad\mbox{and}\quad v_1=\partial_x g_1\quad\mbox{a.e. on}\quad\Lambda_{t_1}^{t_2}\,.$$ For $t_1<t_2<T$ we want to show, that there exists a viscosity solution $g$ of $$\label{viscosity_problem}
\left.\begin{array}{rclcl}
\partial_t g+f(\partial_x g)&=&0&\mbox{in}&\Gamma_{t_1}^{t_2}\,,\\
g&=&g_1&\mbox{on}&\Lambda_{t_1}^{t_2}\,.
\end{array}\right\}$$ Then we claim, that $v=\partial_x g$ is an entropy solution of (\[modified\_scl\]), in the sense of Definition \[def\_entropy\_solution\]. The existence of such a viscosity solution $g$ will be guaranteed by the existence result of Ishi (see Theorem 3.1 in [@Is]). In order to be able to apply that theorem we must find a viscosity subsolution $\underline{g}$ and a viscosity supersolution $\overline{g}$ of (\[viscosity\_problem\]), which satisfy pointwise $\underline{g}=\overline{g}=g_1$ on $\Lambda_{t_1}^{t_2}$ and $\underline{g}\leq\overline{g}$ in $\Gamma_{t_1}^{t_2}$. According to Proposition 5.1 on page 77 in [@BC], the fact that $g_1$ satisfies (\[hamilton\_jacobi\]) almost everywhere implies, that $g_1$ is a viscosity subsolution of (\[hamilton\_jacobi\]). Thus we can put $\underline{g}=g_1$ and it remains to find a viscosity supersolution $\overline{g}$ such that $\overline{g}\geq g_1$ and $\overline{g}=g_1$ on $\Lambda_{t_{1}}^{t_2}$. For two positive constants $A$, $B$ we consider the function $$\overline{g}_y(x,t)=g_1(y,\gamma(y,t_1))+A|x-y|+B|t-\gamma(y)|\,.$$ We calculate for $(x,t)\in\Gamma_{t_1}^{t_2}$ $$\partial_t\overline{g}_y(x,t)+f(\partial_x \overline{g}_y(x,t))=B\operatorname{sign}(t-\gamma(y))+f(a \operatorname{sign}(x-y))\,.$$ By (\[infinity\_behaviour\]) this is positive, if we choose $A$ large enough. Thus $$\partial_t\overline{g}_y(x,t)+f(\partial_x \overline{g}_y(x,t))>0\quad\mbox{for}\quad (x,t)\in\Gamma_{t_1}^{t_2}\,.$$ Proposition 5.1 on page 77 and Proposition 5.4 on page 78 in [@BC] imply, that $\overline{g}$ is a viscosity supersolution. Further we notice, since $g_1\in W^{1,\infty}(\mathbb{R}\times[0,T))$, that for all $y$ and suitable choices of $A$ and $B$ $$g_1(x,t)\leq g(y,\gamma(y))+A|x-y|+B|t-\gamma(y)|\,.$$ By Proposition 2.11 on page 302 in [@BC] $$\overline{g}(x,t)=\inf_{y}\overline{g}_y(x,t)$$ is still a supersolution. Furthermore $\overline{w}$ satisfies by construction $\overline{g}=g_1$ on $\Lambda_{t_{1}}^{t_{2}}$ and $\overline{g}\geq g$ in $\Gamma_{t_{1}}^{t_{2}}$. Hence all assumptions of the existence result (Theorem 3.1) in [@Is] are fulfilled. Therefore there exists a viscosity solution $g$ of (\[viscosity\_problem\]) such that $g_1\leq g\leq\overline{g}$. By Example 1 in [@Is], the viscosity solution is Lipschitz continuous, i.e. $g\in W^{1,\infty}(\Gamma_{t_1}^{t_2})$. For $(x,t)\in\Gamma_{t_1}^{t_2}$ and $(y,s)\in\Lambda_{t_1}^{t_2}$ we notice that $$g_1(x,t)-\overline{g}(y,s)\leq g(x,t)-g(y,s)\leq \overline{g}(x,t)-g_1(y,s)\,.$$ Using the fact that $g_1$ is Lipschitz continuous and the construction of $\overline{g}$ we deduce from the previous line $$-\|(x,t)-(y,s)\|C_1\leq g(x,t)-g(y,s)\leq C_2\|(x,t)-(y,s)\|\,,$$ which means $g\in W^{1,\infty}(\Gamma_{t_1}^{t_2}\cup \Lambda_{t_1}^{t_2})$.
Next we are going to show, that $v=\partial_x g$ is a weak solution of (\[modified\_scl\]) in $\Gamma_{t_1}^{t_2}$ in the sense of Definition \[def\_entropy\_solution\]. Since $g$ satisfies (\[viscosity\_problem\]) almost everywhere, it follows for a $\psi\in C^{\infty}_c(\mathbb{R}\times[0,t_2))$ $$\label{int1}
\int_{\Gamma_{t_1}^{t_2}}\partial_x\psi\partial_{t}g+f(\partial_xg)\partial_{x}\psi\,dx\,dt=0\,.$$ We denote the outer unit normal vector of $\Gamma_{t_1}^{t_2}$ by $n$. Integrating (\[int1\]) twice by parts gives $$\label{int2}\begin{split}
\int_{\Gamma_{t_1}^{t_2}}\partial_x\psi\partial_{t}g\,dx\,dt&=\int_{\partial\Gamma_{t_1}^{t_2}}g\begin{pmatrix}-\partial_t\psi\\\partial_x\psi\end{pmatrix}\cdot n\,d\sigma+ \int_{\Gamma_{t_1}^{t_2}}\partial_t\psi\partial_{x}g\,dx\,dt\,.
\end{split}$$ Rewriting the boundary term in (\[int2\]) and using the fact that $\psi(x,t_2)=0$ leads to $$\begin{gathered}
\label{int3}
\int_{\partial\Gamma_{t_1}^{t_2}}g\begin{pmatrix}-\partial_t\psi\\\partial_x\psi\end{pmatrix}\cdot n\,d\sigma=\int_{\Lambda_{t_1}^{t_2}}g_1\begin{pmatrix}-\partial_t\psi\\\partial_x\psi\end{pmatrix}\cdot n\,d\sigma \\=\int_{s_1}^{s_2} g_1(s,\gamma(s,t_1))\left[\partial_t\psi(s,\gamma(s,t_1))\gamma'(s,t_1)+\partial_x\psi(s,\gamma(s,t_{1}))\right]\,ds\,.\end{gathered}$$ Integrating the right-hand side of (\[int3\]) by parts leads to $$\begin{gathered}
\label{int4}
\int_{s_1}^{s_2} g_1(s,\gamma(s,t_1))\left[\partial_t\psi(s,\gamma(s,t_1))\partial\gamma(s,t_1)+\partial_x\psi(s,\gamma(s,t_{1}))\right]\,ds\\
=\int_{s_1}^{s_2} g_1(s,\gamma(s,t_1))\frac{d}{ds}\psi(s,\gamma(s,t_1))\,ds=-\int_{s_1}^{s_2}\frac{d}{ds}g_1(s,\gamma(s,t_1))\cdot\psi\,ds\,.\end{gathered}$$ Therefore, combining (\[int3\]) and (\[int4\]) we can rewrite the boundary term in (\[int2\]) $$\label{int5}
\int_{\partial\Gamma_{t_1}^{t_2}}g_1\begin{pmatrix}-\partial_t\psi\\\partial_x\phi\end{pmatrix}\cdot n\,d\sigma=\int_{s_1}^{s_2}\left[\partial_x g_1+\partial_{t}g_1\cdot \partial_s\gamma(s,t_1)\right]\psi\,ds$$ Using (\[choice\_of\_t1\]), the right-hand side of (\[int5\]) simplifies to $$\begin{aligned}
\int_{\partial\Gamma_{t_1}^{t_2}}g_1\begin{pmatrix}-\partial_t\psi\\\partial_x\psi\end{pmatrix}\cdot n\,d\sigma&=\int_{s_1}^{s_2}\left[\partial_x g_1-f(\partial_x g_1)\cdot \partial_s\gamma(s,t_1)\right]\psi\,ds\\
&=\int_{\Lambda_{t_1}^{t_2}}\psi\begin{pmatrix}\partial_x g_1\\-f(\partial_x g_1)\end{pmatrix}\cdot \tau\,d\sigma\,,\end{aligned}$$ where $\tau$ is the unit tangent vector of $\Lambda_{t_1}^{t_2}$. We replace now the boundary term in (\[int2\]) using the above identity $$\int_{\Gamma_{t_1}^{t_2}}\partial_x\psi\partial_{t}g\,dx\,dt=\int_{\Lambda_{t_1}^{t_2}}\psi\begin{pmatrix}\partial_x g_1\\-f(\partial_x g_1)\end{pmatrix}\cdot \tau\,d\sigma+ \int_{\Gamma_{t_1}^{t_2}}\partial_t\psi\partial_{x}g\,dx\,dt\,.$$ Finally, this together with (\[int1\]) gives $$\int_{\Lambda_{t_1}^{t_2}}\psi\begin{pmatrix}\partial_x g_1\\-f(\partial_x g_1)\end{pmatrix}\cdot \tau\,d\sigma+ \int_{\Gamma_{t_1}^{t_2}}\partial_t\psi\partial_{x}g+f(\partial_xg)\partial_{x}\psi\,dx\,dt=0\,.$$ Since $v_1=\partial_x g_1$ and by putting $v=\partial_x g$, we see, that $v$ is a solution of (\[modified\_scl\]) in the sense of Definition \[def\_entropy\_solution\]. Finally it remains to show, that $v$ is an entropy solution in the sense, that $$\partial_tv\wedge a+\partial_xf(v\wedge a)\geq0\,.$$ By Corollary 1.7.2 in [@CS] $v$ satisfies for all $(x,t),~(y,t)\in\Gamma_{t_1}^{t_2}$ such that $x<y$ $$v(y,t)-v(x,t)\leq \frac{y-x}{ct}\,.$$ This immediately implies $q(x,t,a)\geq0$ (see Section 8.5 in [@Da]).
Proposition \[existence\] being proved, we now establish some properties for entropy solutions to (\[modified\_scl\]) analogous to those in the classical case (see [@Da]). Precisely we are going to show
\[properties\_of\_entropy\_solutions\] Let $v_1\in L^{\infty}(\mathbb{R}\times(0,T))$ be a weak solution of (\[scl\]). Then there exists a constant $\lambda_0>0$, depending on $f$ and $\|v_1\|_{\infty}$, such that, for any domain $\Gamma_{t_1}^{t_2}$ satisfying $0<\hat{\lambda}\leq \lambda_0$, the entropy solution $v\in L^{\infty}(\Gamma_{t_{1}}^{t_{2}})$ of (\[modified\_scl\]) with boundary condition $v_{1}\in L^{\infty}(\Lambda_{t_{1}}^{t_{2}})$ satisfies $$\label{boundary_asymptotic}
\lim_{\varepsilon\to0^+}\int_{s_1}^{s_2-\varepsilon}\left|v(s,\gamma(s,t_1+\varepsilon))-v_1(s,\gamma(s,t_1))\right|\,ds=0\,,$$ where $$s_1=-\frac{t_2-t_1}{\hat{\lambda}}-\delta\quad\mbox{and}\quad s_2=\frac{t_2-t_1}{\hat{\lambda}}+\delta\,.$$ Moreover $$\label{Linfty_bound}
\|v\|_{\infty}\leq \|v_{1}\|_{\infty}\,$$ and there exists a constant $C>0$, depending only on $\|v\|_1$ and $\hat{\lambda}$, such that $$\label{control_of_the_defect_measure}
\int_{\Gamma_{t_{1}}^{t_{2}}}q(x,t,a)\, da\, dx\, dt\leq C(t_{2}-t_{1})\,.$$ Let now $w_1,\,w_2\in L^{\infty}(\mathbb{R}\times(0,T))$ be weak solutions of (\[scl\]). Then there there exists a constant $\lambda_1>0$ depending on $f$ and $\max\{\|w_1\|_{\infty},\|w_2\|_{\infty}\}$ such that, for any domain $\Gamma_{t_1}^{t_2}$ satisfying $0<\hat{\lambda}\leq \lambda_1$ and any choice of two entropy solutions respectively $v_1\in L^{\infty}(\Gamma_{t_{1}}^{t_{2}})$ with boundary condition $w_{1}\in L^{\infty}(\Lambda_{t_{1}}^{t_{2}})$ and $v_2\in L^{\infty}(\Gamma_{t_{1}}^{t_{2}})$ with boundary condition $w_2\in L^{\infty}(\Lambda_{t_{1}}^{t_{2}})$ the following holds : for any $t\in(t_1,t_2)$ and a constant $C>0$ depending on $\Gamma_{t_1}^{t_2}$ and $\max\{\|w_1\|_{\infty},\|w_2\|_{\infty}\}$: $$\label{contraction_inequality}
\int_{\theta^{-}(t)}^{\theta^{+}(t)}|v_1(x,t)-v_2(x,t)|\, dx\leq C\int_{\Lambda_{t_{1}}^{t_{2}}}|w_1-w_2|\, d\sigma\,,$$ where $$\theta^{\pm}(t)=\pm\frac{t-t_1}{\hat{\lambda}}\pm\delta\,.$$
Inequality (\[contraction\_inequality\]) implies in particular the uniqueness of the entropy solution for a given initial data $w$ on $\Lambda_{t_1}^{t_2}$ issued from a weak solution to (\[scl\]).
We start to prove (\[boundary\_asymptotic\]). Let $R>0$ such that $$R+f(\pm R)\geq0\,.$$ We choose $\lambda_0$ such that $$\label{def_lambda0}
\lambda_0^{-1}=\max\left\{\left|f'(R+1+\|v_1\|_{\infty})\right|,\left|f'(-R-1-\|v_1\|_{\infty})\right|\right\}\,.$$ We consider now a domain $\Gamma_{t_1}^{t_2}$ such that $\hat{\lambda}\leq\lambda_0$ and an entropy solution $v\in L^{\infty}(\Gamma_{t_1}^{t_2})$ of (\[modified\_scl\]) exists. From Example 1 in [@Is], we know, that $$\label{bound1}
\|v\|_{\infty}\leq R+1\,.$$ Let $\psi\in C^{\infty}_c(\mathbb{R}\times[0,t_2))$. From Theorem 1.3.4 in [@Da] we get for all $\varepsilon>0$ and sufficiently small ($t_1+\varepsilon<t_2$) $$\int_{\Gamma_{t_1+\varepsilon}^{t_2-\varepsilon}}v\partial_t\psi+f(v)\partial_x\psi\,dx\,dt=-\int_{\partial\Gamma_{t_1+\varepsilon}^{t_2-\varepsilon}}\begin{pmatrix}v\\ -f(v)\end{pmatrix}\cdot\tau\psi\,d\sigma\,.$$ Since $\psi(x,t_2)=0$ this implies $$\label{bound_asymp1}
\int_{\Gamma_{t_1+\varepsilon}^{t_2-\varepsilon}}v\partial_t\psi+f(v)\partial_x\psi\,dx\,dt=-\int_{\Lambda_{t_1+\varepsilon}^{t_2-\varepsilon}}\begin{pmatrix}v\\ -f(v)\end{pmatrix}\cdot\tau\psi\,d\sigma\,.$$ As $\varepsilon\to0^+$ the left-hand side of (\[bound\_asymp1\]) converges to $$\int_{\Gamma_{t_1}^{t_2}}v\partial_t\psi+f(v)\partial_x\psi\,dx\,dt\,.$$ Since $v$ is a weak solution of (\[modified\_scl\]) this later fact implies for the right-hand side of (\[bound\_asymp1\]) $$\label{bound_asymp2}
\lim_{\varepsilon\to0^+}\int_{\Lambda_{t_1+\varepsilon}^{t_2-\varepsilon}}\begin{pmatrix}v\\ -f(v)\end{pmatrix}\cdot\tau\psi\,d\sigma=\int_{\Lambda_{t_1}^{t_2}}\begin{pmatrix}v_1\\ -f(v_1)\end{pmatrix}\cdot\tau\psi\,d\sigma$$ In order to keep the notation simple we introduce $$\bar{\gamma}(s)=\begin{pmatrix}s\\\gamma(s,t_1)\end{pmatrix}\quad\mbox{and}\quad v_{\varepsilon}(x,t)=v(x,t+\varepsilon)\,.$$ From (\[bound\_asymp2\]) we deduce $$\label{bound_asymp3}
\lim_{\varepsilon\to0^+}\int_{s_1+\varepsilon}^{s_2-\varepsilon}\left\{v_{1}(\bar{\gamma}(s))-v_{\varepsilon}(\bar{\gamma}(s))+\hat{\lambda}\left[f(v_{\varepsilon}(\bar{\gamma}(s)))-f(v_1(\bar{\gamma}(s))) \right]\right\}\psi\,ds\,.$$ By (\[def\_lambda0\]) we obtain the existence of some constants $C,\,c>0$ for which the following holds $$c\leq 1\pm f(\alpha)\leq C\quad\mbox{for all}\quad\alpha\in\left(-R-1-\|v_1\|_{\infty},R+1+\|v\|_{\infty}\right)\,.$$ Therefore we get from (\[bound\_asymp3\]), that $$\label{bound_asymp4}
\lim_{\varepsilon\to0^+}v(s,\gamma(s,t_1+\varepsilon))=v_1(s,\gamma(s))\quad\mbox{for a.e.}\quad s\in[s_1,s_2]\,.$$ By dominated convergence, we deduce the claim (\[boundary\_asymptotic\]).
To prove the remaining claims of our lemma, we need to introduce the kinetic formulation of conservation laws, we recommend the introduction to this subject given in [@Pe]. However we need here a slight modified version of this formulation. We define for any $v\in\mathbb{R}$ $$\chi(v;a):=\mathbbm{1}_{a\leq v}\, ,$$ where $\mathbbm{1}_{a\leq v}$ is the characteristic function of the set $\{a\in\mathbb{R}:~a\leq v\}$. Then a weak solution $v\in L^{\infty}(\Gamma_{t_{1}}^{t_{2}})$ of (\[modified\_scl\]) satisfies in the distributional sense $$\label{kinetic_formulation}
\left.\begin{array}{rclcl}
\partial_{t}\chi(v(x,t);a)+f'(a)\partial_{x}\chi(v(x,t);a)&=&\partial_{a}q(x,t,a)&\mbox{in}&\Gamma_{t_1}^{t_2}\,,\\
\chi(v;a)&=&\chi(v_1;a)&\mbox{on}&\Lambda_{t_1}^{t_2}\,.
\end{array}\right\}$$ In other words this means, that for all $\psi\in C_c^{\infty}(\mathbb{R}\times[0,t_2)\times\mathbb{R})$ $$\begin{gathered}
\int_{\Gamma_{t_1}^{t_2}}\int_{\mathbb{R}}\chi(v;a)\partial_t\psi+f'(a)\chi(v;a)\partial_x\psi\,da\,dx\,dt\\=\int_{\Gamma_{t_1}^{t_2}}+\int_{\Lambda_{t_1}^{t_2}}\int_{\mathbb{R}}\psi\begin{pmatrix}\chi(v1;a)\\-f'(a)\chi(v_1;a)\end{pmatrix}\cdot\tau\,da\,d\sigma\,.\end{gathered}$$ In order to prove (\[contraction\_inequality\]), (\[Linfty\_bound\]) and (\[control\_of\_the\_defect\_measure\]) we need to regularize our kinetic equation (\[kinetic\_formulation\]). We choose $\varphi_{1}(x),~\varphi_{2}(t)\in C^{\infty}_{c}(\mathbb{R})$ non-negative functions such that $$\mbox{supp}~\varphi_1\subset(-1,1)\,,\quad\mbox{supp}~\varphi_{2}\subset[-1,0]$$and$$\int_{\mathbb{R}}\varphi_{2}\,dx=\int_{\mathbb{R}}\varphi_{1}\,dx=1\,.$$ We define the kernel $$\label{kernel}
\varphi_{\varepsilon}(x,t)=\frac{1}{\varepsilon^2}\varphi_1\left(\frac{x}{\varepsilon}\right)\varphi_2\left(\frac{t}{\varepsilon}\right)\,.$$ For a constant $C$ depending only from $\hat{\lambda}$ we have $$\operatorname{dist}((x,t),\partial\Gamma_{t_1}^{t_2})>\varepsilon\quad\mbox{for all}\quad(x,t)\in\Gamma_{t_1+C\varepsilon}^{t_2-C\varepsilon}\,.$$ Consequently for $(x,t)\in\Gamma_{t_{1}+C\varepsilon}^{t_{2}-C\varepsilon}$ $$\label{zero_on_boundary}
\varphi_{\varepsilon}(x-y,t-s)=0\quad\mbox{for}\quad(y,s)\in\partial\Gamma_{t_{1}}^{t_{2}}\,.$$ We define moreover the two mollified functions $$\chi_{\varepsilon}(x,t,a)=\int_{\Gamma_{t_{1}}^{t_{2}}}\varphi_{\varepsilon}(x-y,t-s)\chi(v(y,s);a)\,dy\,ds$$ and $$q_{\varepsilon}(x,t,a)=\int_{\Gamma_{t_{1}}^{t_{2}}}\varphi_{\varepsilon}(x-y,t-s)q(y,s,a)\,dy\,ds\,.$$ For $q_{\varepsilon}$ and $(x,t)\in\Gamma_{t_1+C\varepsilon}^{t_2-C\varepsilon}$ we compute $$\begin{aligned}
q_{\varepsilon}(x,t,a)- q_{\varepsilon}(x,t,b)&=\int_{\Gamma_{t_{1}}^{t_{2}}}\varphi_{\varepsilon}(x-y,t-s)\left[q(y,s,a)-q(y,s,b)\right]\,dy\,ds\\
&=\int_{\Gamma_{t_{1}}^{t_{2}}}\partial_{t}\varphi_{\varepsilon}(x-y,t-s)\left[v\wedge a-v\wedge b\right]\\
&\quad+\partial_{x}\varphi_{\varepsilon}(x-y,t-s)\left[f(v\wedge a)-f(v\wedge b)\right]\,dy\,ds\,,\\\end{aligned}$$ where we have made use of (\[zero\_on\_boundary\]). Since $$|v\wedge a-v\wedge b|\leq\|v\|_{\infty}|b-a|$$ it follows from the calculation above $$\begin{aligned}
\left|q_{\varepsilon}(x,t,a)- q_{\varepsilon}(x,t,b)\right|&\leq \int_{\Gamma_{t_{1}}^{t_{2}}}\left|\partial_{t}\varphi_{\varepsilon}\right|\cdot\|v\|_{\infty}|b-a|+C\left|\partial_{x}\varphi_\varepsilon\right|\cdot\|v\|_{\infty}|b-a|\,dy\,ds\\
&\leq C|b-a|\,.\end{aligned}$$ Therefore $q_{\varepsilon}$ is Lipschitz continuous with respect to the kinetic variable $a$ and we have for almost every $a\in\mathbb{R}$ in the classical sense $$\label{regularized_equation}
\partial_{t}\chi_{\varepsilon}+f'(a)\partial_{x}\chi_{\varepsilon}=\partial_{a}q_{\varepsilon}(x,t,a)\quad\mbox{in}\quad\Gamma_{t_1+C\varepsilon}^{t_2-C\varepsilon}\,.$$ Notice that due to the convolution with $\varphi_{\varepsilon}$ both $\chi_{\varepsilon}$ and $q_{\varepsilon}$ are smooth with respect to $(x,t)$. Furthermore for $(x,t)\in\Gamma_{t_1+C\varepsilon}^{t_2-C\varepsilon}$ the function $q_{\varepsilon}$ satisfies $$\label{compact_support_of_q}
q_{\varepsilon}(x,t,a)=0\quad\mbox{if}\quad |a|\geq \|v\|_{\infty}\,.$$ This follows from the classical fact, that $$q(x,t,a)=0\quad\mbox{for}\quad |a|\geq \|v\|_{\infty}\,.$$ Indeed for $|a|\geq \|v\|_{\infty}$ and $\psi\in C^{\infty}_c(\Gamma_{t_1}^{t_2})$ we compute $$\begin{aligned}
\int_{\Gamma_{t_{1}}^{t_{2}}}q(x,t,a)\psi(x,t)\,dx\,dt&=\int_{\Gamma_{t_{1}}^{t_{2}}}\left[\partial_{t}v(x,t)\wedge a+\partial_{x}f(v(x,t)\wedge a)\right]\psi(x,t)\,dy\,ds\\
&=\int_{\Gamma_{t_{1}}^{t_{2}}}\left[\partial_{t}v(x,t)+\partial_{x}f(v(x,t))\right]\psi(x,t)\,dx\,dt\\
&=-\int_{\Gamma_{t_{1}}^{t_{2}}}v(x,t)\partial_{t}\psi(x,t)+f(v(x,t))\partial_{x}\psi(x,t)\,dx\,dt\\
&=0\,.\end{aligned}$$ Consider now a convex function $\eta(a)$ in $C^1$, which satisfies $$\lim_{a\rightarrow-\infty}\eta(a)=0$$ and denote $$\xi(a)=\int \eta'(a)f'(a)\,da\,.$$ We claim that for all $(x,t)\in\Gamma_{t_1+C\varepsilon}^{t_2-C\varepsilon}$ the following holds $$\label{entropy_equality}
\partial_t \eta_{\varepsilon}(x,t)+\partial_x \xi_{\varepsilon}(x,t)=-\int_{\Gamma_{t_1}^{t_2}}\eta''(a)q_{\varepsilon}(x,t,a)\,da\,,$$ where $$\eta_{\varepsilon}(x,t)=\int_{\Gamma_{t_{1}}^{t_{2}}}\eta(v(y,s))\varphi_{\varepsilon}\left(x-y,t-s\right)\,dy\,ds$$ and $$\xi_{\varepsilon}(x,t)=\int_{\Gamma_{t_{1}}^{t_{2}}}\xi(v(y,s))\varphi_{\varepsilon}\left(x-y,t-s\right)\,dy\,ds\,.$$ Later will make special choices of $\eta$ in order to get (\[Linfty\_bound\]) and (\[control\_of\_the\_defect\_measure\]).
**Proof of claim (\[entropy\_equality\])**. We multiply (\[regularized\_equation\]) by $\eta'(a)$ $$\eta'(a) \partial_{t}\chi_{\varepsilon}+\eta'(a)f'(a)\partial_{x}\chi_{\varepsilon}=\partial_{a}q_{\varepsilon}(x,t,a)\,.$$ Then integrating this equation with respect to $a$ gives $$\label{zwischen1}
\int_{\mathbb{R}}\eta'(a)\partial_{t} \chi_{\varepsilon}+\eta'(a)f'(a)\partial_{x}\chi_{\varepsilon}\,da= \int_{\mathbb{R}}\eta'(a)\partial_{a}q_{\varepsilon}(x,t,a)\,da\,.$$ We compute for the left-hand side $$\begin{aligned}
\int_{\mathbb{R}}\eta'(a)\chi_{\varepsilon}\,da&=\int_{\Gamma_{t_{1}}^{t_{2}}}\int_{\mathbb{R}}\eta'(a)\chi(v(y,s);a)\varphi_{\varepsilon}\left(x-y,t-s\right)\,da\,dy\,ds\\
&=\int_{\Gamma_{t_{1}}^{t_{2}}}\eta(v(y,s))\varphi_{\varepsilon}\left(x-y,t-s\right)\,dy\,ds=\eta_{\varepsilon}(x,t)\end{aligned}$$ and similarly for the second term $$\int_{\mathbb{R}}\eta'(a)f'(a)\chi_{\varepsilon}\,da=\int_{\Gamma_{t_{1}}^{t_{2}}}\xi(v(y,s))\varphi_{\varepsilon}\left(x-y,t-s\right)\,dy\,ds=\xi_{\varepsilon}(x,t)\,.$$ Thus (\[zwischen1\]) reduces to $$\label{zwischen2}
\partial_{t}\eta_{\varepsilon}(x,t)+\partial_{x}\xi_{\varepsilon}(x,t)=\int_{\mathbb{R}}\eta'(a)\partial_{a}q_{\varepsilon}(x,t,a)\,da\,.$$ Integrating the right-hand side by parts gives $$\int_{\mathbb{R}}\eta'(a)\partial_{a}q_{\varepsilon}(x,t,a)\,da=-\int_{\mathbb{R}}\eta''(a)q(x,t,a)\,da\,,$$ where we have used the fact, that $q_{\varepsilon}$ is compactly supported in $a$. This gives the result (\[entropy\_equality\]) as claimed.
Next we integrate inequality (\[entropy\_equality\]) over the set $\Gamma_{t_1+C\varepsilon}^{\bar{t}}$, where $\bar{t}\in(t_1+C\varepsilon,t_2-C\varepsilon)$. We will abbreviate $t_1+C\varepsilon$ by $\bar{t}_1$. We have $$\label{inequality_for_contraction}
\int_{\Gamma_{\bar{t}_1}^{\bar{t}}}\partial_{t}\eta_{\varepsilon}(x,t)+\partial_{x}\xi_{\varepsilon}(x,t)\,dx\,dt=- \int_{\Gamma_{\bar{t}_1}^{\bar{t}}}\int_{\mathbb{R}}\eta''(a)q(x,t,a)\,da\,.$$ For the first term on the left-hand side of (\[inequality\_for\_contraction\]) we compute $$\begin{aligned}
\int_{\Gamma_{\bar{t}_1}^{\bar{t}}}\partial_{t}\eta_{\varepsilon}\,dx\,dt&=\int_{-\delta}^{\delta}\eta_{\varepsilon}(x,\bar{t})-\eta_{\varepsilon}(x,t_{1})\,dx\\
&\quad+\int_{\delta}^{\frac{(\bar{t}-\bar{t}_{1})}{\hat{\lambda}}+\delta}\int_{\hat{\lambda}(x-\delta)+\bar{t}_{1}}^{\bar{t}}\partial_{t}\eta_{\varepsilon}(x,t)\,dt\,dx\\
&\quad+\int^{-\delta}_{-\frac{\bar{t}-\bar{t}_{1}}{\hat{\lambda}}-\delta}\int_{-\hat{\lambda}(x+\delta)+\bar{t}_{1}}^{\bar{t}}\partial_{t}\eta_{\varepsilon}(x,t)\,dt\,dx\,.\end{aligned}$$ This gives $$\begin{aligned}
\int_{\Gamma_{t_1}^{\bar{t}}}\partial_{t}\eta_{\varepsilon}\,dx\,dt&=\int_{-\delta}^{\delta}\eta_{\varepsilon}(x,\bar{t})-\eta_{\varepsilon}(x,\bar{t}_{1})\,dx\\
&\quad+\int_{\delta}^{\frac{\bar{t}-\bar{t}_{1}}{\hat{\lambda}}+\delta}\eta_{\varepsilon}(x,\bar{t})-\eta_{\varepsilon}\left(x,\hat{\lambda}(x-\delta)+\bar{t}_{1}\right)\,dx\\
&\quad+\int^{-\delta}_{-\frac{\bar{t}-\bar{t}_{1}}{\hat{\lambda}}-\delta}\eta_{\varepsilon}(x,\bar{t})-\eta_{\varepsilon}\left(x,-\hat{\lambda}(x-\delta)+\bar{t}_{1}\right)\,dx\,.\end{aligned}$$ A regrouping of the terms together with a change of variable leads to $$\label{zwischen3}\begin{split}
\int_{\Gamma_{t_1}^{\bar{t}}}\partial_{t}\eta_{\varepsilon}\,dx\,dt&=\int_{\theta^{-}_{\varepsilon}(\bar{t})}^{\theta^{+}_{\varepsilon}(\bar{t})}\eta_{\varepsilon}(x,\bar{t})\,dx-\int_{-\delta}^{\delta}\eta_{\varepsilon}(x,\bar{t}_{1})\,dx\\
&\quad+\hat{\lambda}\int_{\bar{t}_{1}}^{\bar{t}}\eta_{\varepsilon}(\theta^{-}_{\varepsilon}(t),t)-\eta_{\varepsilon}(\theta^{+}_{\varepsilon}(t),t)\,dt\,,
\end{split}$$ where $$\label{theta}
\theta_{\varepsilon}^{\pm}(t)=\pm\frac{\bar{t}-\bar{t}_1}{\hat{\lambda}}\pm\delta\,.$$ Integrating now the second term on the left-hand side of (\[inequality\_for\_contraction\]) gives $$\label{zwischen4}
\int_{\Gamma_{\bar{t}_1}^{\bar{t}}}\partial_{x}\xi_{\varepsilon}(x,t)\,dx\,dt=\int_{t_{1}}^{\bar{t}}\xi_{\varepsilon}(\theta_{\varepsilon}^{+}(t),t)-\xi_{\varepsilon}(\theta_{\varepsilon}^{-}(t),t)\,dt\,.$$ Inserting (\[zwischen3\]) and (\[zwischen4\]) back in (\[inequality\_for\_contraction\]) leads to the identity $$\label{main_equality_lemma}\begin{split}
\int_{\theta_{\varepsilon}^{-}(\bar{t})}^{\theta^{+}_{\varepsilon}(\bar{t})}\eta_{\varepsilon}(x,\bar{t})\,dx&=\int_{\Lambda_{\bar{t}_1}^{\bar{t}}}\begin{pmatrix}\eta_{\varepsilon}\\-\xi_{\varepsilon}\end{pmatrix}\cdot\tau\,d\sigma- \int_{\Gamma_{\bar{t}_1}^{\bar{t}}}\int_{\mathbb{R}}\eta''(a)q(x,t,a)\,da\,.
\end{split}$$ For suitable choices of $\eta$ this equality (\[main\_equality\_lemma\]) will imply the first two claims of Lemma \[properties\_of\_entropy\_solutions\].
First we prove (\[Linfty\_bound\]). Let $a_0$ be a real number being fixed later in this proof. We choose $$\eta(a)=\left\{\begin{array}{cl}
(a-a_{0})&\mbox{if}~~a-a_{0}\geq0\,,\\
0 &\mbox{if}~~a-a_{0}\leq0\
\end{array}\right.$$ and we aim to deduce $$\label{positive_part_contraction}
\int_{\theta^{-}(\bar{t})}^{\theta^{+}(\bar{t})}|v(x,\bar{t})-a_{0}|^{+}\,d\sigma\leq C \int_{\Lambda_{t_{1}}^{\bar{t}}}|v_{0}(x)-a_{0}|^{+}\,d\sigma\,,$$ from equality (\[main\_equality\_lemma\]). The non-negativity of $\eta''(a)$ and $q_{\varepsilon}$ implies $$\int_{\Gamma_{\bar{t}_1}^{\bar{t}}}\int_{\mathbb{R}}\eta''(a)q(x,t,a)\,da\ge 0\quad .$$ Using this inequality in equality (\[main\_equality\_lemma\]), we obtain the estimate $$\int_{\theta_{\varepsilon}^{-}(\bar{t})}^{\theta^{+}_{\varepsilon}(\bar{t})}\eta_{\varepsilon}(x,\bar{t})\,dx\leq\int_{\Lambda_{\bar{t}_1}^{\bar{t}}}\begin{pmatrix}\eta_{\varepsilon}\\-\xi_{\varepsilon}\end{pmatrix}\cdot\tau\,d\sigma\,.$$ Letting $\varepsilon\to0^+$ we get $$\int_{\theta^{-}(\bar{t})}^{\theta^{+}(\bar{t})}\eta(x,\bar{t})\,dx\leq\int_{\Lambda_{t_1}^{\bar{t}}}\begin{pmatrix}\eta(v_1)\\-\xi(v_1)\end{pmatrix}\cdot\tau\,d\sigma\,.$$ We observe $$\label{bound_for_xi}
\left|\xi(a)\right|\leq \max_{|b|\leq\|v_1\|_{\infty}}|f'(b)|\cdot\eta(a)\,,$$ which implies $$\int_{\theta^{-}(\bar{t})}^{\theta^{+}(\bar{t})}\eta(x,\bar{t})\,dx\leq C\int_{\Lambda_{t_1}^{\bar{t}}}\eta(v_1)\,dx\,.$$ This is our desired result (\[positive\_part\_contraction\]) and choosing $a_0=\|v_1\|_{\infty}$ in (\[positive\_part\_contraction\]) gives $$\int_{\theta_{-}(\bar{t})}^{\theta^{+}(\bar{t})}|v(x,\bar{t})-a_{0}|^{+}\,d\sigma=0$$ and thus (\[Linfty\_bound\]) follows: $$|v(x,t)|\leq \|v_1\|_{\infty}\quad\mbox{a.e. in}\quad\Gamma_{t_1}^{t_2}\,.$$
In order to prove (\[control\_of\_the\_defect\_measure\]), we choose now $$\eta(a):=\left\{\begin{array}{ccl}
2a^{2}&\mbox{if}&a\geq-\|v\|_{\infty}\,,\\
(a+\|v\|_{\infty})+2\|v\|_{\infty}^{2}&\mbox{if}&-(\|v\|_{\infty}+2\|v\|_{\infty}^{2})\leq a\leq-\|v\|_{\infty}\,,\\
0 &\mbox{if}&a\leq-(\|v\|_{\infty}+2\|v\|_{\infty}^{2})\,.
\end{array}\right.$$ Since $\eta$ is non-negative, we deduce from (\[main\_equality\_lemma\]) $$\label{last_equation_for_mass_bound}\begin{split}
2\int_{\Gamma_{t_{1}+C\varepsilon}^{t_{2}-C\varepsilon}}\int_{\mathbb{R}}q_{\varepsilon}(x,t,a)\,da\,dx\,dt&\leq \int_{\Lambda_{\bar{t}_1}^{\bar{t}}}\begin{pmatrix}\eta_{\varepsilon}\\-\xi_{\varepsilon}\end{pmatrix}\cdot\tau\,d\sigma\quad .
\end{split}$$ Since $\eta(a)=2a^2$ for $a\in[-\|v\|_{\infty},\|v\|_{\infty}]$ we get $$|\xi(a)|=\int_{-\|v\|_{\infty}}^{a}|\eta'(b)f'(b)|\,db
\leq f'(\|v\|_{\infty}) \int_{-\|v\|_{\infty}}^{|a|}|\eta'(b)|\,db\,.$$ Hence, by letting $\varepsilon\to0$ in (\[last\_equation\_for\_mass\_bound\]), we obain $$\int_{\Gamma_{t_{1}}^{t_{2}}}\int_{\mathbb{R}}q(x,t,a)\,da\,dx\,dt\leq C(\delta+t_2-t_1)\,,$$ as announced in (\[control\_of\_the\_defect\_measure\]).
Finally we are going to prove (\[contraction\_inequality\]). We choose the domain $\Gamma_{t_1}^{t_2}$ in such a way that $$0<\hat{\lambda}\leq \lambda_1\,,$$ where $$\label{def_lambda1}\begin{array}{c}
\lambda_1=\left(\max\left\{f(-R-1-\alpha),f(R+1+\alpha)\right\}\right)^{-1}\\[5mm]
\mbox{and}\\[5mm]
\alpha=\max\{\|w_1\|_{\infty},\|w_2\|_{\infty}\}\,.\end{array}$$ For the two entropy solutions $v_{1},\, v_{2}$ with boundary conditions $w_{1}$ and $w_{2}$ we consider the kinetic equations $$\left.\begin{array}{rclcl}
\partial_{t}\chi_{i}+f'(a)\partial_{x}\chi_{i}&=&\partial_{a}q_{i}&\mbox{in}&\mathcal{D}'(\Gamma_{t_{1}}^{t_{2}}\times\mathbb{R})\\
\chi_{i}&=&\chi(w_{i};a)&\mbox{on}&\Lambda_{t_{1}}^{t_{2}}
\end{array}\right\}$$ where $\chi_i=\chi(v_i(x,t);a)$ for $i=1,\,2\,$. Then, as before, we can regularize our kinetic equations with the kernel defined in (\[kernel\]) $$\begin{array}{rclcl}
\partial_{t}\chi^{\varepsilon}_{i}+f'(a)\partial_{x}\chi^{\varepsilon}_{i}&=&\partial_{a}q^{\varepsilon}_{i}(x,t,a)&\mbox{in}&\Gamma_{t_{1}+C\varepsilon}^{t_{2}-C\varepsilon}\\
\end{array}$$ where $$\chi_{i}^{\varepsilon}(x,t,a)=\int_{\Gamma_{t_{1}}^{t_{2}}}\chi(v_{i}(x,t);a)\varphi_{\varepsilon}(x-y,t-s)\,dx\,dt\quad\mbox{for}\quad i=1,\,2$$ and $C>0$ is again chosen such that for $(x,t)\in \Gamma_{t_{1}+C\varepsilon}^{t_{2}-C\varepsilon}$ $$\varphi_{\varepsilon}(x-y,t-s)=0\quad\mbox{for}\quad (y,s)\in\partial\Gamma_{t_1}^{t_2}\,.$$ Then the function $\left(\chi_{1}^{\varepsilon}-\chi_{2}^{\varepsilon}\right)^{2}$ satisfies for $(x,t)\in\Gamma_{t_{1}+C\varepsilon}^{t_{2}-C\varepsilon}$ and almost every $a\in\mathbb{R}$ $$\label{regualrized_equation_contraction}
\partial_{t}\left(\chi_{1}^{\varepsilon}-\chi_{2}^{\varepsilon}\right)^{2}+f'(a)\partial_{x}\left(\chi_{1}^{\varepsilon}-\chi_{2}^{\varepsilon}\right)^{2}=\chi^{\varepsilon}_{1}\partial_{a}q_{1}^{\varepsilon}+\chi_{2}^{\varepsilon}\partial_{a}q_{2}^{\varepsilon}
-\chi_{2}^{\varepsilon}\partial_{a}q_{1}^{\varepsilon}-\chi_{1}^{\varepsilon}\partial_{a}q_{2}^{\varepsilon}\,.$$ We make use again of the following abbreviation: $t_1+C\varepsilon=\bar{t}_1$. Let $\bar{t}\in(\bar{t}_1,+t_2-C\varepsilon)$, then we integrate (\[regualrized\_equation\_contraction\]) in $\Gamma_{\bar{t}_1}^{\bar{t}}\times\mathbb{R}$, which leads to $$\begin{gathered}
\label{result1_for_contraction}
\int_{\Gamma_{\bar{t}_{1}}^{\bar{t}}}\int_{\mathbb{R}}\partial_{t}\left(\chi_{1}^{\varepsilon}-\chi_{2}^{\varepsilon}\right)^{2}+f'(a)\partial_{x}\left(\chi_{1}^{\varepsilon}-\chi_{2}^{\varepsilon}\right)^{2}\,da\,dx\,dt\\
=\int_{\Gamma_{\bar{t}_{1}}^{\bar{t}}}\int_{\mathbb{R}}\chi^{\varepsilon}_{1}\partial_{a}q_{1}^{\varepsilon}+\chi_{2}^{\varepsilon}\partial_{a}q_{2}^{\varepsilon}-\chi_{2}^{\varepsilon}\partial_{a}q_{1}^{\varepsilon}-\chi_{1}^{\varepsilon}\partial_{a}q_{2}^{\varepsilon}\,da\,dx\,dt\,.\end{gathered}$$ We recall, that $\chi(v;a)=\mathbbm{1}_{a\leq v}$ and $$q_{1}(x,t,a)=q_{2}(x,t,a)=0\quad\mbox{for}\quad|a|\geq \max\{\|v_{1}\|_{\infty},\|v_{2}\|_{\infty}\}\,.$$ Therefore we can calculate for $(x,t)\in \Gamma_{\bar{t}_1}^{\bar{t}}$ and $i,\,j\in\{1,\,2\}$ $$\begin{aligned}
\int_{\mathbb{R}}\chi_{i}^{\varepsilon}\partial_{a}q_{j}^{\varepsilon}\,da&=\int_{\mathbb{R}}\int_{\Gamma_{t_{1}}^{t_{2}}}\chi(v_{i}(y,s);a)\varphi_{\varepsilon}(x-y,t-s)q_{j}^{\varepsilon}(x,t,a)\,dy\,ds\,da\\
&=\int_{\Gamma_{t_{1}}^{t_{2}}}q_{j}^{\varepsilon}(x,t,v_{i}(y,s))\varphi_{\varepsilon}(x-y,t-s)\,dy\,ds\,.\end{aligned}$$ This implies, since $\varphi_{\varepsilon}$ and $q_{\varepsilon}$ are non-negative $$\int_{\mathbb{R}}\chi_{2}^{\varepsilon}\partial_{a}q_{1}^{\varepsilon}\,da\geq0\quad\mbox{and}\quad \int_{\mathbb{R}}\chi_{1}^{\varepsilon}\partial_{a}q_{2}^{\varepsilon}\,da\geq0\,,$$ which applied in (\[result1\_for\_contraction\]) leads to the inequality $$\begin{gathered}
\label{result2_for_contraction}
\int_{\Gamma_{\bar{t}_{1}}^{\bar{t}}}\int_{\mathbb{R}}\partial_{t}\left(\chi_{1}^{\varepsilon}-\chi_{2}^{\varepsilon}\right)^{2}+f'(a)\partial_{x}\left(\chi_{1}^{\varepsilon}-\chi_{2}^{\varepsilon}\right)^{2}\,da\,dx\,dt\\
\leq\int_{\Gamma_{\bar{t}_{1}}^{\bar{t}}}\int_{\mathbb{R}}\chi^{\varepsilon}_{1}\partial_{a}q_{1}^{\varepsilon}+\chi_{2}^{\varepsilon}\partial_{a}q_{2}^{\varepsilon}\,da\,dx\,dt\,.\end{gathered}$$ For the left hand-side of (\[result2\_for\_contraction\]) we compute $$\label{calculation0}\begin{split}
\int_{\mathbb{R}}\int_{\Gamma_{\bar{t}_{1}}^{\bar{t}}}&\partial_{t}\left(\chi_{1}^{\varepsilon}-\chi_{2}^{\varepsilon}\right)^{2}+f'(a)\partial_{x}\left(\chi_{1}^{\varepsilon}-\chi_{2}^{\varepsilon}\right)^{2}\,dx\,dt\,da\\
&=\int_{\mathbb{R}}\int_{\theta^{-}_{\varepsilon}(\bar{t})}^{\theta_{\varepsilon}^{+}(\bar{t})}\left(\chi_{1}^{\varepsilon}-\chi_{2}^{\varepsilon}\right)^{2}(x,\bar{t})\,dx\,da -\int_{\mathbb{R}}\int_{-\delta}^{\delta}\left(\chi_{1}^{\varepsilon}-\chi_{2}^{\varepsilon}\right)^{2}(x,\bar{t}_{1})\,dx\,da\\
&\quad+\hat{\lambda}\int_{\mathbb{R}}\int_{\bar{t}_{1}}^{\bar{t}}\left(\chi_{1}^{\varepsilon}-\chi_{2}^{\varepsilon}\right)^{2}(\theta^{+}_{\varepsilon}(t),t)-\left(\chi_{1}^{\varepsilon}-\chi_{2}^{\varepsilon}\right)^{2}(\theta_{\varepsilon}^{-}(t),t)\,dt\,da\\
&\quad+\int_{\mathbb{R}}\int_{\bar{t}_{1}}^{\bar{t}}f'(a)\left[\left(\chi_{1}^{\varepsilon}-\chi_{2}^{\varepsilon}\right)^{2}(\theta_{\varepsilon}^{-}(t),t)-\left(\chi_{1}^{\varepsilon}-\chi_{2}^{\varepsilon}\right)^{2}(\theta_{\varepsilon}^{+}(t),t)\right]\,dt\,da\,,
\end{split}$$ where $\theta_{\varepsilon}^{\pm}$ are defined in (\[theta\]). After a change of variable this expression simplifies to $$\begin{gathered}
\label{calculation1}
\int_{\mathbb{R}}\int_{\Gamma_{\bar{t}_{1}}^{\bar{t}}}\partial_{t}\left(\chi_{1}^{\varepsilon}-\chi_{2}^{\varepsilon}\right)^{2}+f'(a)\partial_{x}\left(\chi_{1}^{\varepsilon}-\chi_{2}^{\varepsilon}\right)^{2}\,dx\,dt\,da\\
=\int_{\mathbb{R}}\int_{\theta^{-}_{\varepsilon}(\bar{t})}^{\theta_{\varepsilon}^{+}(\bar{t})}\left(\chi_{1}^{\varepsilon}-\chi_{2}^{\varepsilon}\right)^{2}(x,\bar{t})\,dx\,da-\int_{\Lambda_{\bar{t}_1}^{\bar{t}}}\int_{\mathbb{R}}\begin{pmatrix}\left(\chi_{1}^{\varepsilon}-\chi_{2}^{\varepsilon}\right)^{2}\\f'(a)\left(\chi_{1}^{\varepsilon}-\chi_{2}^{\varepsilon}\right)^{2}\end{pmatrix}\cdot \tau\,da\,d\sigma\,.\end{gathered}$$ Using identity (\[calculation1\]) in (\[result2\_for\_contraction\]) gives $$\begin{gathered}
\label{result3_for_contraction}
\int_{\mathbb{R}}\int_{\theta_{\varepsilon}^{-}(\bar{t})}^{\theta^{+}_{\varepsilon}(\bar{t})}\left(\chi_{1}^{\varepsilon}-\chi_{2}^{\varepsilon}\right)^{2}(x,\bar{t})\,dx\,da\\
\leq\int_{\Gamma_{\bar{t}_{1}}^{\bar{t}}}\int_{\mathbb{R}}\chi^{\varepsilon}_{1}\partial_{a}q_{1}^{\varepsilon}+\chi_{2}^{\varepsilon}\partial_{a}q_{2}^{\varepsilon}\,da\,dx\,dt+\int_{\Lambda_{\bar{t}_1}^{\bar{t}}}\int_{\mathbb{R}}\begin{pmatrix}\left(\chi_{1}^{\varepsilon}-\chi_{2}^{\varepsilon}\right)^{2}\\f'(a)\left(\chi_{1}^{\varepsilon}-\chi_{2}^{\varepsilon}\right)^{2}\end{pmatrix}\cdot \tau\,da\,d\sigma\,.\end{gathered}$$ We claim $$\label{asymptotic_of_q}
\lim_{\varepsilon\to0^+}\int_{\Gamma_{t_{1}}^{\bar{t}}}\int_{\mathbb{R}}\chi^{\varepsilon}_{i}\partial_{a}q_{i}^{\varepsilon}\,da\,dx\,dt=0\quad\mbox{for}\quad i\in\{1,\,2\}\,.$$ **Proof of Claim (\[asymptotic\_of\_q\]).** We consider the function $\chi_{i}^{\varepsilon}-\left(\chi^{\varepsilon}_{i}\right)^{2}$ which satisfies satisfies pointwise for $(x,t)\in\Gamma_{t_1+C\varepsilon}^{t_2-C\varepsilon}$ and almost every $a\in\mathbb{R}$ $$\partial_{t}\left[\chi_{i}^{\varepsilon}-\left(\chi_{i}^{\varepsilon}\right)^{2}\right]+f'(a)\partial_{x}\left[\chi_{i}^{\varepsilon}-\left(\chi_{i}^{\varepsilon}\right)^{2}\right]=\partial_{a}q_{i}^{\varepsilon}+2\chi_{i}^{\varepsilon}\partial_{a}q_{i}^{\varepsilon}\,.$$ Integrating this in $\Gamma_{\bar{t}_{1}}^{\bar{t}}\times\mathbb{R}$ leads to $$\begin{gathered}
\label{asymptotic0}
\int_{\mathbb{R}}\int_{\Gamma_{\bar{t}_{1}}^{\bar{t}}}\partial_{t}\left[\chi_{i}^{\varepsilon}-\left(\chi_{i}^{\varepsilon}\right)^{2}\right]+f'(a)\partial_{x}\left[\chi_{i}^{\varepsilon}-\left(\chi_{i}^{\varepsilon}\right)^{2}\right]\,dx\,dt\,da\\=\int_{\mathbb{R}}\int_{\Gamma_{\bar{t}_{1}}^{\bar{t}}}2\chi_{i}^{\varepsilon}\partial_{a}q_{i}^{\varepsilon}\,dx\,dt\,da\,,\end{gathered}$$ where we made use of the fact, that $q_i^{\varepsilon}$ is compactly supported in $a$. For the left-hand side of (\[asymptotic0\]) one can compute following step by step (\[calculation0\]) and (\[calculation1\]) $$\begin{gathered}
\label{asymptotic1}
\int_{\mathbb{R}}\int_{\Gamma_{\bar{t}_{1}}^{\bar{t}}}\partial_{t}\left[\chi_{i}^{\varepsilon}-\left(\chi_{i}^{\varepsilon}\right)^{2}\right]+f'(a)\partial_{x}\left[\chi_{i}^{\varepsilon}-\left(\chi_{i}^{\varepsilon}\right)^{2}\right]\,dx\,dt\,da\\
=\int_{\mathbb{R}}\int_{\theta^{-}_{\varepsilon}(\bar{t})}^{\theta_{\varepsilon}^{+}(\bar{t})}\left[\chi_{i}^{\varepsilon}-\left(\chi_{i}^{\varepsilon}\right)^2\right](x,\bar{t})\,dx\,da-\int_{\Lambda_{\bar{t}_1}^{\bar{t}}}\int_{\mathbb{R}}\begin{pmatrix}\chi_{i}^{\varepsilon}-\left(\chi_{i}^{\varepsilon}\right)^2\\f'(a)\left[\chi_{i}^{\varepsilon}-\left(\chi_{i}^{\varepsilon}\right)^2\right]\end{pmatrix}\cdot \tau\,da\,d\sigma\,.\end{gathered}$$ For the right-hand side of (\[asymptotic1\]) we observe $$\begin{gathered}
\label{asymptotic2}
\lim_{\varepsilon\rightarrow0^+}\int_{\mathbb{R}}\int_{\theta^{-}_{\varepsilon}(\bar{t})}^{\theta_{\varepsilon}^{+}(\bar{t})}\left[\chi_{i}^{\varepsilon}-\left(\chi_{i}^{\varepsilon}\right)^2\right](x,\bar{t})\,dx\,da\\=\int_{\mathbb{R}}\int_{\theta^{-}(\bar{t})}^{\theta^{+}(\bar{t})}\left[\chi_{i}-\left(\chi_{i}\right)^2\right](x,\bar{t})\,dx\,da\end{gathered}$$ and $$\begin{gathered}
\label{asymptotic3}
\lim_{\varepsilon\rightarrow0^+}\int_{\Lambda_{\bar{t}_1}^{\bar{t}}}\int_{\mathbb{R}}\begin{pmatrix}\chi_{i}^{\varepsilon}-\left(\chi_{i}^{\varepsilon}\right)^2\\f'(a)\left[\chi_{i}^{\varepsilon}-\left(\chi_{i}^{\varepsilon}\right)^2\right]\end{pmatrix}\cdot \tau\,da\,d\sigma\\=\int_{\Lambda_{t_1}^{\bar{t}}}\int_{\mathbb{R}}\begin{pmatrix}\chi_{i}-\left(\chi_{i}\right)^2\\f'(a)\left[\chi_{i}-\left(\chi_{i}\right)^2\right]\end{pmatrix}\cdot \tau\,da\,d\sigma\,.\end{gathered}$$ Since $$\chi_i=\left(\chi_{i}\right)^2$$ the right-hand side of (\[asymptotic2\]) and (\[asymptotic3\]) are zero. Thus $$\lim_{\varepsilon\rightarrow0^+}\int_{\mathbb{R}}\int_{\theta^{-}_{\varepsilon}(\bar{t})}^{\theta_{\varepsilon}^{+}(\bar{t})}\left[\chi_{i}^{\varepsilon}-\left(\chi_{i}^{\varepsilon}\right)^2\right](x,\bar{t})\,dx\,da-\int_{\Lambda_{\bar{t}_1}^{\bar{t}}}\int_{\mathbb{R}}\begin{pmatrix}\chi_{i}^{\varepsilon}-\left(\chi_{i}^{\varepsilon}\right)^2\\f'(a)\left[\chi_{i}^{\varepsilon}-\left(\chi_{i}^{\varepsilon}\right)^2\right]\end{pmatrix}\cdot \tau\,da\,d\sigma=0\,.$$ With (\[asymptotic1\]) one concludes $$\lim_{\varepsilon\rightarrow0^+}\int_{\mathbb{R}}\int_{\Gamma_{\bar{t}_{1}}^{\bar{t}}}\partial_{t}\left[\chi_{i}^{\varepsilon}-\left(\chi_{i}^{\varepsilon}\right)^{2}\right]+f'(a)\partial_{x}\left[\chi_{i}^{\varepsilon}-\left(\chi_{i}^{\varepsilon}\right)^{2}\right]\,dx\,dt\,da=0\,.$$ Finally taking limits on both sides of (\[asymptotic0\]) we get $$\lim_{\varepsilon\to0^+}\int_{\Gamma_{t_{1}}^{\bar{t}}}\int_{\mathbb{R}}\chi^{\varepsilon}_{i}\partial_{a}q_{i}^{\varepsilon}\,da\,dx\,dt\quad\mbox{for}\quad i\in\{1,\,2\}\,,$$ as announced.
Letting $\varepsilon\to0^+$ in (\[result3\_for\_contraction\]) and using (\[asymptotic\_of\_q\]) leads to $$\label{last1}
\int_{\mathbb{R}}\int_{\theta^{-}(\bar{t})}^{\theta^{+}(\bar{t})}\left(\chi_{1}-\chi_{2}\right)^{2}(x,\bar{t})\,dx\,da\\
\leq\int_{\Lambda_{t_1}^{\bar{t}}}\int_{\mathbb{R}}\begin{pmatrix}\left(\chi_{1}-\chi_{2}\right)^{2}\\f'(a)\left(\chi_{1}-\chi_{2}\right)^{2}\end{pmatrix}\cdot \tau\,da\,d\sigma\,.$$ We compute $$\label{last2}
\int_{\mathbb{R}}\left(\chi(v_{1}(x,t);a)-\chi(v_{2}(x,t);a)\right)^{2}\,da=|v_{1}(x,t)-v_{2}(x,t)|\,,$$ and $$\label{last3}
\int_{\mathbb{R}}\begin{pmatrix}\left(\chi_{1}-\chi_{2}\right)^{2}\\f'(a)\left(\chi_{1}-\chi_{2}\right)^{2}\end{pmatrix}\cdot \tau\,da=\begin{pmatrix}|w_{1}-w_{2}|\\\operatorname{sign}(w_1-w_2)(f(w_1)-f(w_2))\end{pmatrix}\cdot \tau\,.$$ Applying (\[last2\]) and (\[last3\]) in (\[last1\]) gives $$\label{last4}
\int_{\theta^{-}(\bar{t})}^{\theta^{+}(\bar{t})}|v_{1}(x,\bar{t})-v_{2}(x,\bar{t})|\,dx
\leq\int_{\Lambda_{t_1}^{\bar{t}}}\begin{pmatrix}|w_{1}-w_{2}|\\\operatorname{sign}(w_1-w_2)(f(w_1)-f(w_2))\end{pmatrix}\cdot \tau\,d\sigma\,.$$ For the right hand side, we compute $$\begin{split}
\int_{\Lambda_{t_1}^{\bar{t}}}&\begin{pmatrix}|w_{1}-w_{2}|\\\operatorname{sign}(w_1-w_2)(f(w_1)-f(w_2))\end{pmatrix}\cdot \tau\,d\sigma\\
&=\int_{s_1}^{s_2}|w_1(s,\gamma(s,t_1))-w_2(s,\gamma(s))|\\
&\quad+\partial_s\gamma(s,t_1) \left[\operatorname{sign}(w_1-w_2)(f(w_1)-f(w_2))\right](s,\gamma(s,t_1)\,ds\\
&=\int_{s_1}^{s_2}|w_1(s,\gamma(s,t_1))-w_2(s,\gamma(s))|\cdot(1\pm\partial_s\gamma(s,t_1)f'(\alpha))\\
&\leq C \int_{s_1}^{s_2}|w_1(s,\gamma(s,t_1))-w_2(s,\gamma(s))|\,ds\,,
\end{split}$$ for a function $\alpha$. From (\[last4\]) we obtain $$\int_{\theta^{-}(\bar{t})}^{\theta^{+}(\bar{t})}|v_{1}(x,\bar{t})-v_{2}(x,\bar{t})|\,dx
\leq C\int_{\Lambda_{t_1}^{\bar{t}}}|w_1(s,\gamma(s,t_1))-w_2(s,\gamma(s))|\,d\sigma$$ as claimed.
Blow up at the points of negative density. {#section_blow_up}
------------------------------------------
In this section we aim to prove the following lemma
\[lemma\_positiv\_density\] Let $u\in L^{\infty}(\mathbb{R}\times[0,T))$ be a weak solution of (\[scl\]), which satisfies (\[minimality\]). Then for $\mathcal{H}^1$ almost every $(x_{0},t_{0})\in\mathbb{R}\times(0,T)$ $$\limsup_{r\rightarrow 0^+}\frac{1}{r}\int_{\mathbb{R}}m\left(B_r(x_0,t_0),a\right)\,da\geq 0\,.$$
A useful lemma that will be used to prove Lemma \[lemma\_positiv\_density\] is the following.
\[compactness\] Let $u\in L^{\infty}(\mathbb{R}\times(0,T))$ be a weak solution of (\[scl\]). Let $r_{n}\rightarrow0^+$. For $(x_{0},t_{0})\in\mathbb{R}\times(0,T)$ define $$\begin{array}{c}
u_{n}(x,t):=\left(D_n^{-1}\right)^{*}u(x,t)\\[5mm]
\mbox{and}\\[5mm]
\mu_{n}:=\displaystyle\frac{1}{r_n}\int_{\mathbb{R}}\left(D_n\right)_*m\,da\,,
\end{array}$$ where $$\label{dilation}
D_n(x,t)=\left(\frac{x-x_0}{r_n},\frac{t-t_0}{r_n}\right)\,.$$ Then there exists for $\mathcal{H}^1$ almost every $(x_{0},t_{0})\in\mathbb{R}\times(0,T)$ a subsequence $r_{k}$ such that $$u_{k}\rightarrow u_{\infty}\quad\mbox{in}\quad L^{1}_{loc}(\mathbb{R}^2)\,.$$ And furthermore $$\mu_{k}\rightharpoondown\mu_{\infty}\quad\mbox{in}\quad\mathcal{M}_{loc}(\mathbb{R}^{2})\,.$$ Which means in other words $$\int_{\mathbb{R}^2}\psi\,d\mu_k\rightarrow\int_{\mathbb{R}^2}\psi\,d\mu_{\infty}\quad\mbox{for all}\quad\psi\in C^{0}_c(\mathbb{R}^2)\,.$$
Lemma \[compactness\] will be a consequence of of the following proposition, which is proved in Appendix A of [@Le].
\[prop\_lecumberry\] For any constant $M\geq0$, for any bounded set $\Omega$, the set $$\left\{u\in L^{\infty}(\Omega):~\|u\|_{\infty}+\int_{\Omega}\int_{\mathbb{R}}|m(x,t,a)|\leq M\right\}$$ is compact in $L^{1}(\Omega)$ with respect to the strong topology.
By construction we already have $$\label{L_infinity_bound_for_u}
\|u_{n}\|_{\infty}\leq\|u\|_{\infty}\,.$$ For this reason it remains to show that for all $R>0$ $|\mu_{n}|(B_{R}(0,0))$ and for $\mathcal{H}^1$ almost every $(x_{0},t_{0})\in\mathbb{R}\times(0,T)$ there exists a constant $C>0$, such that $$\label{bound_measure}
\limsup_{n\rightarrow\infty}\mu_{n}(B_{r_n}(0,0))\leq C\,.$$ But this is a direct consequence of Theorem 2.56 in [@AFP]. Since (\[L\_infinity\_bound\_for\_u\]) and (\[bound\_measure\]) hold, the assumptions of Proposition \[prop\_lecumberry\] are fulfilled and we can extract a subsequence $r_{k'}$ such that $$u_{k'}\rightarrow u_{\infty}\quad\mbox{in}\quad L^{1}_{loc}(\mathbb{R}^2)\,.$$ Additionally we have by the weak$^{*}$ compactness of measures (see Theorem 1.59 in [@AFP]), that, possibly after extracting a further subsequence $r_{k}$, $$\mu_{k}\rightharpoondown\mu_{\infty}\quad\mbox{in}\quad\mathcal{M}\,,$$ Altogether we have for the sequence $r_{k}$ $$u_{k}\rightarrow u_{\infty}\quad\mbox{in}\quad L^{1}_{loc}(\mathbb{R}^2)\,$$ and $$\mu_{k}\rightharpoondown\mu_{\infty}\quad\mbox{in}\quad\mathcal{M}_{loc}(\mathbb{R}^{2})\,,$$ which is what we aimed to prove.
We argue by contradiction. Therefore we assume that there exists a point $(x_{0},t_{0})$ such that $$\label{contradiction}
\limsup_{r\rightarrow 0^+}\frac{1}{r}\int_{B_{r}((x_{0},t_{0}))}\int_{\mathbb{R}}m(x,t,a)\,da\,dx\,dt<0\,.$$ For a sequence $r_{n}\rightarrow0^{+}$ we define $$\begin{array}{c}
u_{n}(x,t):=\left(D_{n}^{-1}\right)^{*}u(x,t)\\[5mm]
\mbox{and}\\[5mm]
\mu_{n}:=\displaystyle\frac{1}{r_n}\int_{\mathbb{R}}\left(D_n\right)_*m\,da\,.
\end{array}$$ Let $u_{k}$ and $\mu_{k}$ be the subsequences given by Lemma \[compactness\], with limits $u_{\infty}$, $\mu_{\infty}$. Then we have by strong convergence, that $u_{\infty}$ is a weak solution of $$\partial_{t}u_{\infty}+\partial_{x}f(u_{\infty})=0\,.$$ Furthermore, by the uniqueness of the distributional limit, we conclude that $$\mu_{\infty}=\int_{\mathbb{R}}\partial_{t}(u_{\infty}\wedge a)+\partial_{x}f(u_{\infty}\wedge a)\,da\,.$$ From (\[contradiction\]) we want to conclude now, that $$\label{negative_of_limit}
\mu_{\infty}(B_R(0,0))<0\quad\mbox{for all}\quad R>0\,.$$
**Proof of (\[negative\_of\_limit\]).** For the sake of contradiction, we assume, that there exists a $R_0$ such that $$\mu_{\infty}(B_{R_0}(0,0))\geq0\,.$$ In [@Le] it is proved, that there exits a set $K$, which is either a line, or a half-line, or the empty set, such that $$\label{structure_of_measure}
\partial_{t}u_{\infty}\wedge a+\partial_{x}f(u_{\infty}\wedge a)=\left((X(u_{\infty}^{+}\wedge a)-X(u_{\infty}^{-}\wedge a)\right)\cdot\omega_{K}\,\mathcal{H}^{1}\res K\,,$$ where $$\label{vectorfield_X}
X(u)=\begin{pmatrix}f(u)\\ u\end{pmatrix}\quad\mbox{and}\quad \omega_{K}=\frac{|u_{\infty}^{+}-u_{\infty}^{-}|}{|X(u^{+}_{\infty})-X(u_{\infty}^{-})|}\begin{pmatrix}1\\ -\frac{f(u_{\infty}^{+})-f(u_{\infty}^{-})}{u_{\infty}^{+}-u_{\infty}^{-}}\end{pmatrix}\,.$$ Moreover therein it is proved, that $u_{\infty}$ is $\mathcal{H}^{1}$-a.e. approximately continuous in $K^{c}$ and has $\mathcal{H}^{1}$-a.e. constant approximate jump points $u_{\infty}^{\pm}$ on $K\,$.
A short calculation reveals $$\begin{gathered}
\label{sign_of_measure}
\int_{\mathbb{R}} \left((X(u_{\infty}^{+}\wedge a)-X(u_{\infty}^{-}\wedge a)\right)\cdot\omega_{K}\\=\operatorname{sign}(u^{-}_{\infty}-u^+_{\infty})\int_{\min\{u_{\infty}^{+},u_{\infty}^{-}\}}^{\max\{u_{\infty}^{+},u_{\infty}^{-}\}}\frac{f(u_{\infty}^{+})+f(u_{\infty}^{-})}{2}-f(a)\,da\,.\end{gathered}$$ By convexity of $f$ we get $$\frac{f(u_{\infty}^{+})+f(u_{\infty}^{-})}{2}>f(a)\quad\mbox{for}\quad a\in\left[\min\{u_{\infty}^{+},u_{\infty}^{-}\},\max\{u_{\infty}^{+},u_{\infty}^{-}\}\right]\,.$$ This and (\[sign\_of\_measure\]) imply, that sign of $\mu_{\infty}$ is completely determined by $\operatorname{sign}(u^{-}_{\infty}-u^{+}_{\infty})$. Henceforth $$\mu_{\infty}(B_{R_0}(0,0))\geq0$$ can only be fulfilled, if $$u^-_{\infty}\geq u^+_{\infty}\,.$$ But this implies that the measure is $\mu_{\infty}$ has a sign, i.e. $$\mu_{\infty}\geq0\,.$$ Let $\mu_{k}^{\pm}$ be the positive respective negative part of $\mu_{k}$, i.e. $\mu_{k}^{\pm}$ are non-negative measures such that $$\mu_{k}=\mu_k^+-\mu_k^-\,.$$ Then we can extract a further subsequence $k'$ such that $$\mu_{k'}^{+}\rightharpoondown\nu^{+}\quad\mbox{and}\quad\mu_{k'}^{-}\rightharpoondown\nu^{-}\quad\mbox{in}\quad\mathcal{M}_{loc}(\mathbb{R}^2)\,.$$ For $R>0$ and non-negative $\psi\in C^{\infty}_{c}(B_{R}(0,0))$ we get $$\int_{B_{R_0}(0,0)}\psi\,d\mu_{\infty}=\lim_{k'\rightarrow\infty}\int_{B_{R}(0,0)}\psi\,d\mu_{k'}=\int_{B_{R}(0,0)}\psi\,d\nu^+-\int_{B_{R}(0,0)}\psi\,d\nu^-\,.$$ Since $\mu_{\infty}$ is non-negative we get for all non-negative $\psi\in C^{\infty}_{c}(B_{R_0}(0,0))$ $$\int_{B_{R_0}(0,0)}\psi\,d\nu^-\leq\int_{B_{R}(0,0)}\psi\,d\nu^+\,.$$ Hence $$\label{inequality_for_negative_and_positive_part}
\nu^{-}(B_{R}(0,0))\leq \nu^{+}(B_{R}(0,0))$$ By Theorem 1.2 in [@Le] (see also Theorem 1.1 in [@AKLR]) we have for a rectifiable set $J_u$ and an ${\mathcal H}^1$ measurable function $h:~J_u\to\mathbb{R}$ $$\label{rectifiable}
\int_{\mathbb{R}}|m(x,t,a)|\,da=h\cdot\mathcal{H}^1\res J_u+\delta_u\,,$$ where $\delta_u$ satisfies $$\forall~B\quad\mbox{Borel}\quad \mathcal{H}^1(B)<\infty\Longrightarrow\delta_u(B)=0\,.$$ Therefore we can choose $R_1$, such that for all $k'$ $$\mu^{-}_{k'}(\partial B_{R_1}(0,0))\leq\frac{1}{r_{k'}}\int_{D_{k'}^{-1}(\partial B_{R_{1}}(0,0))}h\,d\mathcal{H}^1\res J_u=0\,.$$ Hence $$\nu^{-}(\partial B_{R_1}(0,0))= \lim_{k'\rightarrow\infty}\mu^{-}_{k'}(\partial B_{R_1}(0,0))=0\,.$$ This and (\[inequality\_for\_negative\_and\_positive\_part\]) imply $$\begin{aligned}
\limsup_{k'\rightarrow\infty}\mu_{k'}^-(B_{R_1}(0,0)&\leq\nu^-(\bar{B}_{R_1}(0,0))=\nu^-(B_{R_1}(0,0))\\
&\leq\nu^{+}(B_{R_1}(0,0))\leq\liminf_{k'\rightarrow\infty}\mu_{k'}^{+}(B_{R_1}(0,0))\,.\end{aligned}$$ $$\begin{aligned}
\limsup_{k'\rightarrow\infty}\mu_{k'}(B_{R_1}(0,0))\geq\liminf_{k'\rightarrow\infty}\mu_{k'}^+B_{R_1}(0,0))-\limsup_{k'\rightarrow\infty}\mu^-_{k'}(B_{R_1}(0,0))\geq0\,,\end{aligned}$$ which obviously contradicts (\[contradiction\]) and we get (\[negative\_of\_limit\]).
Inequality (\[negative\_of\_limit\]) implies, that the set $K$ in (\[structure\_of\_measure\]) is non-empty and $$\mu_{\infty}<0\,,$$ which gives, again from above considerations $$u_{\infty}^-<u_{\infty}^+\,.$$ Moreover the convexity of $f$ implies for every $a\in(u_{\infty}^{-},u_{\infty}^{+})$ $$\begin{gathered}
\partial_{t}u_{\infty}\wedge a+\partial_{x}f(u_{\infty}\wedge a)\\=\left(\frac{f(a)-f(u_{\infty}^{-})}{a-u_{\infty}^{-}}-\frac{f(u_{\infty}^{+})-f(u_{\infty}^{-})}{u^+_{\infty}-u^-_{\infty}}\right)\left(a-u_{\infty}^{-}\right)\mathcal{H}^{1}\res K\leq0\,.\end{gathered}$$ In other words, we get $$\partial_{t}u_{\infty}\wedge a+\partial_{x}f(u_{\infty}\wedge a)\leq0\,.$$ For $P=(x_{p},t_{p})\in\mathbb{R}^{2}$ let $K=P+\mathbb{R}\omega_{K}^{\perp}$ if $K$ is a line or $K=P+\mathbb{R}_{+}\omega_{K}^{\perp}$ if $K$ is a halfline. Define $$\begin{array}{c}
H^{+}:=\left\{(x,t):~((x,t)-P)\cdot\omega_{K}>0\right\}\\[5mm]
\mbox{and}\\[5mm]
H^{-}:=\left\{(x,t):~((x,t)-P)\cdot\omega_{K}<0\right\}
\end{array}$$ if $K$ is a line and $$\begin{array}{c}
H^{+}:=\left\{(x,t):~((x,t)-P)\cdot\omega_{K}>0~\mbox{and}~x>f'(u_{\infty}^{+})(t-t_{p})+x_{p}\right\}\\[5mm]
H^{-}:=\left\{(x,t):~((x,t)-P)\cdot\omega_{K}<0~\mbox{and}~x<f'(u_{\infty}^{-})(t-t_{p})+x_{p}\right\}\,,
\end{array}$$ if $K$ is a half-line. From the proof of Proposition 3.3 in [@Le] (see also Theorem 6.2 in [@AKLR] for a similar proof) we get, that $$u_{\infty}(x,t)=u_{\infty}^{-}\quad\mbox{on}\quad H^{-}\quad\mbox{and}\quad u_{\infty}(x,t)=u_{\infty}^{+}\quad\mbox{on}\quad H^{+}\,.$$ Now we choose $\bar{t}\in\mathbb{R}$ and $\delta>0$ in the definition of the sets $\Lambda_{\bar{t}}^{\bar{t}+1}$ and $\Gamma_{\bar{t}}^{\bar{t}+1}$ (see (\[the\_set\_gamma\])), in such a way that $$\left[-\frac{\delta}{2},\frac{\delta}{2}\right]\times\{t\}\cap K\neq\emptyset\quad\forall~t\in(\bar{t},\bar{t}+1)\,.$$ Furthermore $\Gamma_{\bar{t}}^{\bar{t}+1}$ is defined such that the conclusions of Lemma \[properties\_of\_entropy\_solutions\] applies to this trapeze. In particular the strong convergence of $u_{k}$ in $L^{1}_{loc}(\mathbb{R}^{2})$ implies $$u_{k}\rightarrow u_{\infty}\quad\text{in}\quad L^{1}\left(\Gamma_{\bar{t}}^{\bar{t}+1}\right)\,,$$ which directly implies by a change of variable $$\begin{aligned}
\int_{\bar{t}}^{\bar{t}+1}\int_{\Lambda_{t'}^{\bar{t}+1}}|u_{k}-u_{\infty}|\,d\sigma\,dt' \rightarrow 0\,.\end{aligned}$$ Thus for almost every $t_{1}\in(\bar{t},\bar{t}+1)$ we get $$\label{convergence_of_inital_data}
\int_{\Lambda_{t_{1}}^{\bar{t}+1}}|u_{k}-u_{\infty}|\,d\sigma\rightarrow0\,$$ and moreover by (\[rectifiable\]) $$\label{no_measure}
\mu_{k}(\Lambda_{t_1}^{\bar{t}+1})=\int_{D_k^{-1}(\Lambda_{t_1}^{\bar{t}+1})}h\,\mathcal{H}^1\res J_{u}=0\,.$$ We set $t_{2}:=\bar{t}+1$, then, according to Proposition \[existence\], we can choose a $t_1\in(\bar{t},\bar{t}+1)$ such that for all $k\in\mathbb{N}$ (\[convergence\_of\_inital\_data\]), (\[no\_measure\]) holds and for $k\in\mathbb{N}\cup\{\infty\}$ there exists an entropy solution $w_{k}$ of $$\label{problem_for_uk}
\left.\begin{array}{rclcl}
\partial_{t}w_{k}+\partial_{x}f(w_{k})&=&0&\mbox{in}&\Gamma_{t_{1}}^{t_{2}}\,,\\[2mm]
w_{k}&=&u_{k}&\mbox{on}&\Lambda_{t_{1}}^{t_{2}}\,.
\end{array}\right\}$$ By Lemma \[properties\_of\_entropy\_solutions\] we have for all $t_{1}\leq t<t_{2}$ $$\int_{\theta^{-}(t)}^{\theta^{+}(t)}|w_{k}(x,t)-w_{\infty}(x,t)|\,dx\leq\int_{\Lambda_{t_{1}}^{t_{2}}}|u_{k}-u_{\infty}|\,d\sigma\,.$$ This and (\[convergence\_of\_inital\_data\]) imply $$w_{k}\rightarrow w_{\infty}\quad\mbox{in}\quad L^{1}\left(\Gamma_{t_{1}}^{t_{2}}\right)\,.$$ By our choice of $t_{1}$, we have for an $x_{1}\in\left[-\frac{\delta}{2},\frac{\delta}{2}\right]$ $$u_{\infty}(x,t_{1})=\left\{\begin{array}{rl}
u_{\infty}^{-} &\mbox{if}~x<x_{1}\\
u_{\infty}^{+} &\mbox{if}~x>x_{1}\,.\\
\end{array}\right.$$ This structure of $u_{\infty}$ at the time $t_1$ allows us to compute $w_{\infty}$ explicitly. Since $u_{\infty}^{-}<u_{\infty}^{+}$, the two states $u_{\infty}^{-}$ and $u_{\infty}^{+}$ are connected by a rarefaction wave $$w_{\infty}(x,t):=\left\{\begin{array}{lcl}
u_{\infty}^{-}&\mbox{if}&x-x_{1}<f'(u_{\infty}^{-})(t-t_{1})\,,\\
\left(f'\right)^{-1}\left(\frac{x-x_{1}}{t-t_{1}}\right)&\mbox{if}&f'(u_{\infty}^{-})(t-t_{1})<x-x_{1}<f'(u_{\infty}^{+})(t-t_{1})\,,\\
u_{\infty}^{+}&\mbox{if}&x-x_{1}>f'(u_{\infty})^{+}(t-t_{1})\,.
\end{array}\right.$$ We observe, that $w_{\infty}$ is a Lipschitz function and this implies pointwise almost everywhere in $\Gamma_{t_{1}}^{t_{2}}$ $$\partial_{t}w_{\infty}+\partial_{x}f(w_{\infty})=0\,.$$ Hence $$\begin{aligned}
q_{\infty}(x,t,a)&=\partial_{t}(w_{\infty}\wedge a)+\partial_{x}f(w_{\infty}\wedge a)\\
&=\mathbbm{1}_{w\leq a}\left[\partial_{t}w_{\infty}+f'(w_{\infty}\wedge a)\partial_{x}w_{\infty}\right]=0\quad\mbox{in}\quad\Gamma_{t_1}^{t_2}\,.\end{aligned}$$ Furthermore the strong convergence of $w_{k}$ in $L^{1}(\Gamma_{t_{1}}^{t_{2}})$ implies $$q_{k}\rightharpoondown q_{\infty}\quad\mbox{in}\quad\mathcal{M}_{loc}(\mathbb{R}^2)\,,$$ where $$q_k=\partial_t w_k\wedge a+\partial_x f(w_k\wedge a)\,.$$ To simplify notations, we define $$\begin{array}{c}
\varGamma_{k}:=\left\{(x,t)\in\mathbb{R}\times(0,T):~D_k(x,t)\in\Gamma_{t_1}^{t_2}\right\}\\[5mm]
\mbox{and}\\[5mm]
\varLambda_k:=\left\{(x,t)\in\mathbb{R}\times(0,T):~D_k(x,t)\in\Lambda_{t_1}^{t_2}\right\}
\end{array}\,,$$ where the map $D_k$ is defined in (\[dilation\]). Then we define the rescaled function $$\tilde{w}_{k}(x,t)=\left\{
\begin{array}{lcl}
\left(D_k\right)^{*}w_{k}&\mbox{if}&(x,t)\in\varGamma_{k}\,,\\
u&\mbox{if}&(x,t)\in\mathbb{R}\times(0,t_0+r_{k}t_2)\backslash\varGamma_{k}\,.
\end{array}\right.$$ and claim, that $w_{k}\in L^{\infty}(\mathbb{R}\times(0,t_{0}+r_{k}t_{2}))$ is a weak solution of (\[scl\]) for all $k\in\mathbb{N}$. To do so, we first observe that $u_{k}$ itself is a weak solution of (\[problem\_for\_uk\]). With that knowledge we calculate. $$\begin{split}
\int_{\varGamma_{k}}\tilde{w}_{k}\partial_{t}\psi+f(\tilde{w}_k)\partial_{x}\psi\,dx\,dt&=r^{2}\int_{\Gamma_{t_1}^{t_2}}\tilde{w}_{k}\partial_{t}\psi+f(\tilde{w}_{k})\partial_{x}\psi\,dx\,dt\\
&=-r_k^2\int_{\Lambda_{t_1}^{t_2}}\psi\begin{pmatrix} u_{k} \\
-f(u_{k})\end{pmatrix}\cdot\tau\,d\sigma\\
&=r_k^{2}\int_{\Gamma_{t_1}^{t_2}}u_{k}\partial_{t}\psi+f(u_{k})\partial_{x}\psi\,dx\,dt\\
&=\int_{\varGamma_{k}}u\partial_{t}\psi+f(u)\partial_{x}\psi\,dx\,dt\,.
\end{split}$$ Using this equality we see $$\begin{aligned}
\int_{\mathbb{R}\times[0,t_{0}+r_{n}t_2]}\tilde{w}_{n}\partial_{t}\psi+f(\tilde{w}_{k})\partial_{x}\psi\,dx\,dt&= \int_{\Gamma_{k}}\tilde{w}_{k}\partial_{t}\psi+f(\tilde{w}_{k})\partial_{x}\psi\,dx\,dt\\
&\quad+\int_{\mathbb{R}\times(0,t_0+r_{k}t_2)\backslash\varGamma_{k}}u\partial_{t}\psi+f(u)\partial_{x}\psi\,dx\,dt\\
&=\int_{\varGamma_{k}}u\partial_{t}\psi+f(u)\partial_{x}\psi\,dx\,dt\\
&\quad+\int_{\mathbb{R}\times(0,t_0+r_{k}t_2)\backslash\varGamma_{k}}u\partial_{t}\psi+f(u)\partial_{x}\psi\,dx\,dt\\
&=\int_{\mathbb{R}\times[0,t_{0}+r_{n}t_2]}u\partial_{t}\psi+f(u)\partial_{x}\psi\,dx\,dt\\
&=\int_{\mathbb{R}}u_{0}(x)\psi(x,0)\,dx\,,\end{aligned}$$ which means, that $\tilde{w}_k$ is indeed a weak solution of (\[scl\]). Therefore the minimality condition (\[minimality\]) of $u$ applies and we deduce $$\int_{\mathbb{R}\times (0,t_{0}+r_{k}t_{2})\times\mathbb{R}}|m(x,t,a)|\,da\,dx\,dt\leq \int_{\mathbb{R}\times(0,t_{0}+r_{k}t_{2})\times\mathbb{R}}|\tilde{q}_{k}(x,t,a)|\,da\,dx\,dt\,.$$ But since $$m(x,t,a)=\tilde{q}_k(x,t,a)\quad\mbox{on}\quad\mathbb{R}\times(0,t_0+r_{k}t_2)\backslash\bar{\varGamma}_{k}$$ we get $$\label{comparison_measure}
\int_{\mathbb{R}}|m|(\varGamma_{k}\cup\varLambda_k,a)\,da\leq\int_{\mathbb{R}}|\tilde{q}_{k}|(\varGamma_{k}\cup\varLambda_k,a)\,da\,.$$ We claim now $$\label{no_measure_on_boundary}
|\tilde{q}_{k}|(\varLambda_k,a)=0\quad\mbox{for all}\quad k\in\mathbb{N}\,.$$ **Proof of (\[no\_measure\_on\_boundary\]).** We define the domain $\Lambda_{\varepsilon}$ such that $$\partial\Lambda_{\varepsilon}=\Lambda_{t_1+\varepsilon}^{t_2}\cup\Lambda_{t_1-\varepsilon}^{t_2}\cup I_l\cup I_r$$ and $$\Lambda_{t_1}^{t_2}\subset\Lambda_{\varepsilon}\,,$$ where $$I_l=\left[\frac{t_2-(t_1+\varepsilon)}{\hat{\lambda}}+\delta,\frac{t_2-(t_1-\varepsilon)}{\hat{\lambda}}\right]$$ and $$I_r=\left[-\frac{t_2-(t_1-\varepsilon)}{\hat{\lambda}}-\delta,-\frac{t_2-(t_1+\varepsilon)}{\hat{\lambda}}-\delta\right]$$ Then for $\varGamma^{\varepsilon}_{k}:=D_{k}^{-1}(\Lambda_{\varepsilon})$ and $\psi\in C_{c}^{\infty}(\mathbb{R}\times(0,t_0+t_2r_k))$ it follows by Theorem 1.3.4 in [@Da] $$\label{boundary1}
\int_{\varLambda^{\varepsilon}_k}\tilde{w}\wedge a\ \partial_t\psi+f(u\wedge a)\partial_x\psi\,dx\,dt=\int_{\partial\varLambda^{\varepsilon}_k}\begin{pmatrix}f(\tilde{w})\\\tilde{w}\end{pmatrix}\cdot n\psi\,d\sigma+\int_{\varLambda^{\varepsilon}_k}\psi\,d\tilde{q}(x,t,a)\,,$$ where $n$ is the outer unit normal of $\varLambda_{k}^{\varepsilon}$. The boundary term can be separated in three parts $$\label{bound2}\begin{split}
\int_{\partial\varLambda^{\varepsilon}_k}\begin{pmatrix}f(\tilde{w})\\\tilde{w}\end{pmatrix}\cdot n\psi\,d\sigma&=\int_{D_k ^{-1}(\Lambda_{t_1-\varepsilon}^{t_2})}\begin{pmatrix}f(u)\\\ u\end{pmatrix}\cdot n\psi\,d\sigma-\int_{D_k ^{-1}(\Lambda_{t_1+\varepsilon}^{t_2})}\begin{pmatrix}f(\tilde{w})\\\tilde{w}\end{pmatrix}\cdot n\psi\,d\sigma\\
&\quad+\int_{D_k^{-1}(I_l)}\tilde{w}(x,t_0+r_kt_2)\,dx+\int_{D_k^{-1}(I_r)}\tilde{w}(x,t_0+r_kt_2)\,dx
\end{split}$$ As $\varepsilon\to0^+$ the two last quantities in the right-hand side of in (\[bound2\]) vanish. For the first expression on the right hand side of (\[bound2\]) one concludes $$\label{bound3}
\lim_{\varepsilon\to0^+}\int_{D_k ^{-1}(\Lambda_{t_1-\varepsilon}^{t_2})}\begin{pmatrix}f(u)\\\ u\end{pmatrix}\cdot n\psi\,d\sigma=\int_{D_k ^{-1}(\Lambda_{t_1}^{t_2})}\begin{pmatrix}f(u)\\\ u\end{pmatrix}\cdot n\psi\,d\sigma\,.$$ With a change of variable and with Lemma \[properties\_of\_entropy\_solutions\] it follows $$\begin{split}\label{bound4}
\lim_{\varepsilon\to0^+}\int_{D_k ^{-1}(\Lambda_{t_1+\varepsilon}^{t_2})}\begin{pmatrix}f(\tilde{w})\\\tilde{w}\end{pmatrix}\cdot n\psi\,d\sigma&=\lim_{\varepsilon\to 0^+ }r_k\int_{\Lambda_{t_1+\varepsilon}^{t_2}}\begin{pmatrix}f(w)\\w\end{pmatrix}\cdot n\psi\,d\sigma\\
&=r_k\int_{\Lambda_{t_1+\varepsilon}^{t_2}}\begin{pmatrix}f(u_k)\\u_k\end{pmatrix}\cdot n\psi\,d\sigma\\
&=\int_{D_k ^{-1}(\Lambda_{t_1}^{t_2})}\begin{pmatrix}f(u)\\\ u\end{pmatrix}\cdot n\psi\,d\sigma\,.
\end{split}$$ From (\[bound2\]), (\[bound3\]) and (\[bound4\]) we conclude $$\lim_{\varepsilon\to0^+}\int_{\partial\varLambda^{\varepsilon}_k}\begin{pmatrix}f(\tilde{w})\\\tilde{w}\end{pmatrix}\cdot n\psi\,d\sigma=0\,.$$ Therefore we can conclude from (\[boundary1\]) $$\lim_{\varepsilon\to0^+}\int_{\Lambda^{\varepsilon}_{k}}\psi\ d\tilde{q}(x,t,a)=0\quad\mbox{for}\quad \psi\in C_c^{\infty}(\mathbb{R}\times(0,t_0+r_kt_2))\,.$$ From this it follows, $$|\tilde{q}_k|(\varLambda_k,a)=0$$ as claimed.
In a next step we show, that (\[no\_measure\_on\_boundary\]) induces $$\label{limit}
\lim_{k\rightarrow\infty}\int_{\mathbb{R}}q_{k}(\Gamma_{t_1}^{t_2},a)\,da=\int_{\mathbb{R}}q_{\infty}(\Gamma_{t_1}^{t_2},a)\,da\,.$$ **Proof of (\[limit\]).** Since $w_k$ is an entropy solution we deduce from (\[no\_measure\_on\_boundary\]) that $|\tilde{q}_k|(\partial \varGamma_k,a)=0$ and therefore $$\label{limit2}
\frac{1}{r_k}\int_{\mathbb{R}}\left(D_k\right)_{*}|\tilde{q}_k|(\partial\Gamma_{t_1}^{t_2},a)\,da=0\,.$$ Lemma \[properties\_of\_entropy\_solutions\] and (\[limit2\]) imply for a constant $C>0$ $$\begin{aligned}
\frac{1}{r_k}\int_{\mathbb{R}}\left(D_k\right)_{*}|\tilde{q}_k|(\bar{\Gamma}_{t_1}^{t_2},a)\,da&=\frac{1}{r_k}\int_{\mathbb{R}}\left(D_k\right)_{*}|\tilde{q}_k|(\partial\Gamma_{t_1}^{t_2},a)\,da+\int_{\mathbb{R}}q_k(\Gamma_{t_1}^{t_2},a)\,da\\
&=\int_{\mathbb{R}}q_k(\Gamma_{t_1}^{t_2},a)\,da<C\,.\end{aligned}$$ Hence one gets for a positive measure $\nu\in\mathcal{M}(\bar{\Gamma}_{t_1}^{t_1})$ after possibly extracting a subsequence $$\frac{1}{r_k}\int_{\mathbb{R}}\left(D_k\right)_{*}|\tilde{q}_k|\rightharpoondown\nu\quad\mbox{in}\quad\mathcal{M}(\bar{\Gamma}_{t_1}^{t_1})\,.$$ Then Proposition 1.62 in [@AFP] and (\[limit2\]) imply $$\label{limit3}
\lim_{k\rightarrow\infty}\frac{1}{r_k}\int_{\mathbb{R}}\left(D_k\right)_{*}|\tilde{q}_k|(\partial\Gamma_{t_1}^{t_2},a)\,da=\nu(\partial\Gamma_{t_1}^{t_2})=0\,.$$ But $\nu(\partial\Gamma_{t_1}^{t_2})=0$ and Proposition 1.62 in [@AFP] give again $$\lim_{k\rightarrow\infty}\frac{1}{r_k}\int_{\mathbb{R}}\left(D_k\right)_{*}|\tilde{q}_k|(\Gamma_{t_1}^{t_2},a)\,da=\lim_{k\rightarrow\infty}\int_{\mathbb{R}}q_{k}(\Gamma_{t_1}^{t_2},a)\,da=\int_{\mathbb{R}}q_{\infty}(\Gamma_{t_1}^{t_2},a)\,da\,.$$
Since (\[no\_measure\]) and (\[no\_measure\_on\_boundary\]) holds we deduce from (\[comparison\_measure\]) $$|\mu_k|(\Gamma_{t_1}^{t_2})\leq\int_{\mathbb{R}}q_k(\Gamma_{t_1}^{t_1},a)\,da\,.$$ Taking the limit on both sides and applying (\[limit\]) gives $$\begin{aligned}
|\mu_{\infty}|(\Gamma_{t_1}^{t_2})&\leq \liminf_{k\rightarrow+\infty}|\mu_{k}|(\Gamma_{t_1}^{t_2})\leq \liminf_{k\rightarrow+\infty}\int_{\mathbb{R}}q_{k}(\Gamma_{t_1}^{t_2},a)\,da\\
&=\int_{\mathbb{R}}q_{\infty}(\Gamma_{t_1}^{t_2},a)\,da=0\,.\end{aligned}$$ But $$|\mu_{\infty}|(\Gamma_{t_1}^{t_2})=0$$ is contradiction to (\[negative\_of\_limit\]). Therefore $$\limsup_{r\rightarrow 0^+}\frac{1}{r}\int_{\mathbb{R}}m(B_{r}((x_{0},t_{0})),a)\,da\geq 0\,,$$ which is, what we aimed to prove.
Proving that $u$ is entropic
----------------------------
In this last section we are going to prove
\[proving\_entropy\] Let $u\in L^{\infty}(\mathbb{R}\times [0,T)$ be a weak solution of (\[scl\]). Let $m(x,t,a)$ its entropy defect measure. If for $\mathcal{H}^1$ almost every $(x_{0},t_{0})\in\mathbb{R}\times(0,T)$ $$\label{positivity_assumption}
\limsup_{r\rightarrow0^+}\frac{1}{r}\int_{B_{r}(x_{0},t_{0})}\int_{\mathbb{R}}m(x,t,a)\,da\,dx\,dt\geq0\,,$$ then $u$ is the entropy solution of (\[scl\]).
We follow closely [@ALR]. Without loss of generality we can assume $f(0)=0$ and $f\geq0$. According to Theorem \[correspondence\] there exists a $g\in W^{1,\infty}(\mathbb{R}\times[0,T))$ such that $u=\partial_{x}g$ and it satisfies almost everywhere $$\label{HJ}
\left.
\begin{array}{rl}
\partial_{t}g+f(\partial_{x}g)&=0\,,\\
\partial_{x}g(x,0)&=u_{0}(x)\,.
\end{array}\right\}$$ We want to show, that $g$ is a viscosity solution of (\[HJ\]), i.e. we want to prove, that $g$ is a sub- and supersolution of (\[HJ\]). This immediately implies by Corollary 1.7.2 in [@ALR], that $u$ is an entropy solution. We already now, that $g$ satisfies (\[HJ\]) almost everywhere, then Proposition 5.1 in [@BC] implies, that $g$ is a subsolution. Therefore it remains to show, that $g$ is a supersolution of (\[HJ\]). Let $\psi\in C^{1}(\mathbb{R}\times\mathbb{R}_{+})$ such that $g-\psi$ has a local minimum in $(x_{0},t_{0})$. Without loss of generality we can assume $g(x_{0},t_{0})=\psi(x_{0},t_{0})$. We want to show that $$\partial_{t}\psi(x_{0},t_{0})+f\left(\partial_{x}\psi(x_{0},t_{0})\right)\geq0\,.$$ We argue by contradiction, therefore we assume $$\partial_{t}\psi(x_{0},t_{0})+f\left(\partial_{x}\psi(x_{0},t_{0})\right)<0\,.$$ Since $f\geq0$ this immediately implies $$\label{psi_is_negative}
\partial_t\psi(x_0,t_0)<0\,.$$ For a sequence $r_{n}\rightarrow0^{+}$ we introduce $$\begin{aligned}
u_{n}(x,t)&=u(x_{0}+r_{n}x,t_{0}+r_{n}t),\\
\psi_{n}(x,t)&=\frac{1}{r_{n}}\left(\psi(x_{0}+\lambda r_{n}x,t_{0}+r_{n}t)-\psi(x_{0},t_{0})\right),\\
g_{n}(x,t)&=\frac{1}{r_{n}}\left(g(x_{0}+r_{n}x,t_{0}+r_{n}t)-g(x_{0},t_{0})\right),\end{aligned}$$ where $0<\lambda<1$ is a constant, which we choose later. According to Lemma \[compactness\] we can extract a subsequence $r_{k}$ such that $$u_{k}\rightarrow u_{\infty}\quad\mbox{in}\quad L^{1}(B_{1})$$ Since $\partial_{x}g_{k}=u_{k}$ and $\partial_{t}g_{k}=f(u_{k})$ we have by Arzela-Ascoli, that $g_{k}$ converges uniformly to a Lipschitz function $g_{\infty}$ such that $\partial_{x}u_{\infty}=g_{\infty}$ and $g_{\infty}$ fulfills (\[HJ\]) almost everywhere. Furthermore we have for $\psi_{\infty}:=\nabla\psi(x_{0},t_{0})\cdot(\lambda x, t)^{\mathrm{T}}$ $$\lim_{k\rightarrow\infty}\psi_{k}(x,t)=\psi_{\infty}\,.$$ We notice, that for all $0<\lambda<1$ and for all $k$ the functions $g_{k}-\psi_{k}$ have a local minimum in $(0,0)\,.$ By uniform convergence the function $g_{\infty}-\psi_{\infty}$ admits also a local minimum in $(0,0)\,.$ Moreover $$\mu_{k}=\frac{1}{r_k}\int_{\mathbb{R}}\left(D_k\right)_{*}m\,da\rightharpoondown \mu_{\infty}\quad\mbox{in}\quad\mathcal{M}(B_{1})\,.$$ Similar as in Section \[section\_blow\_up\] from $$\lim_{k\rightarrow\infty}\int_{B_{1}(0,0)}\mu_{k}(B_1(0,0))\geq0\,,$$ we can conclude $$m_{\infty}(x,t,a):=\partial_{t}u_{\infty}\wedge a+\partial_{x}f(u_{\infty}\wedge a)\geq0\,.$$ Let $\delta>0$, then the function $$h_{\delta}(x,t):=g_{\infty}-\psi_{\infty}+\frac{\delta}{2}\left[(1-\lambda)x^2+t^2\right]$$ is defined on $B_{1}$ and has a strict minimum in $(0,0)$. Notice that $h_{\delta}(0,0)=0$ and $h\geq0$ in $B_{1}$. We claim that $$\label{gradient_positive}
|\nabla h_{\delta}|>0\quad\mbox{a.e. in}\quad B_{1}\,.$$ **Proof of (\[gradient\_positive\]).** Let $(x,t)\in B_{1}$ such that $h_{\delta}$ is differentiable in $(x,t)$ and $\nabla h_{\delta}(x,t)=0$. It follows since $g_{\infty}$ solves (\[HJ\]) $$\begin{aligned}
0&=\partial_t g_{\infty}+f(\partial_x g_{\infty})\\
&=\partial_t\psi(x_0,t_0)-\delta t+f(\lambda\partial_x\psi(x_0,t_0)+(1-\lambda)\delta x)\\
&\leq \partial_t\psi(x_0,t_0)+\lambda f(\partial_x\psi(x_0,t_0))+((1-\lambda)f(\delta x)-\delta t)\,.\end{aligned}$$ Since (\[psi\_is\_negative\]) holds, we can choose $\delta$ and $\lambda$ small enough the expression $$\partial_t\psi(x_0,t_0)+\lambda f(\partial_x\psi(x_0,t_0))+\delta(f(\delta x)-t)$$ becomes strictly negative, which is a contradiction. Therefore the claim (\[gradient\_positive\]) is proved.
Further we choose $\delta$ and $\lambda$ small enough such that $$\label{choice_of_delta_lambda}
|\partial_t\psi(x_0,t_0)|>\lambda\,\partial_x\psi(x_0,t_0)\cdot\sup_{s\in[-\|u\|_{\infty},\|u\|_{\infty}]}f'(s)+\delta((1-\lambda)x+t)\,.$$ By $\tau>0$ we denote the minimum of $h_{\delta}$ on $\partial B_{1}$ and by $\overline{a}$ the essential supremum of $u_{\infty}$ on $\{h_{\delta}<\tau\}$. If $\overline{a}>0$ let $\underline{a}$ be close to $\overline{a}$ such that $0<\underline{a}<\overline{a}$. Let $A:=\{h_{\delta}<\tau\}\cap \{\underline{a}<u_{\infty}\}$. The set $A$ has positive Lebesgue measure. Therefore by the Coarea Formula and by $|\nabla h_{\delta}|>0$ it follows for $E_{s}:=\{h_{\delta}=s\}$ $$0<\int_{A}|\nabla h_{\delta}(x,t)|\,dx\,dt=\int_{0}^{\tau}\mathcal{H}^{1}(A\cap E_{s})\,ds\,.$$ Hence the set $$S:=\left\{s\in(0,\tau): \mathcal{H}^{1}(\{\underline{a}<u_{\infty}\}\cap E_{s})>0,~\mathcal{H}^{1}(\{u_{\infty}>\overline{a}\}\cap E_{s})=0\right\}$$ has positive Lebesgue measure. For a vector $v=(v_{1},v_{2})$ we define $v^{\perp}:=(-v_{2},v_{1})$ and for a $s\in S$ the function $$s\to l(s):=\int_{E_{s}}\left[X(u_{\infty}\wedge a)-\nabla^{\perp}\psi_{\infty}+\delta((1-\lambda)x,t)^{\perp}\right]\cdot\nu\,$$ where $\nu=\frac{\nabla h_{\delta}}{|\nabla h_{\delta}|}$ and the $X$ is the vectorfield from (\[vectorfield\_X\]). We choose $s\in S$ such that $$\lim_{\varepsilon\rightarrow0}\frac{1}{\varepsilon}\int^{s}_{s-\varepsilon}l(s')ds'=l(s)\,.$$ We define $\zeta_{\varepsilon}(x,t):=1\wedge(s-h_{\delta})^{+}/\varepsilon$ and calculate $$\nabla \zeta_{\varepsilon}=\left\{\begin{array}{ll}
0&\mbox{if}~h_{\delta}>s~\mbox{or}~h_{\delta}<s-\varepsilon\\
-\frac{1}{\varepsilon}\nabla h_{\delta} &\mbox{if}~ s-\varepsilon<h_{\delta}<s\,.
\end{array}\right.$$ The choice of $s\in S$ an the Coarea Formula implies $$\begin{gathered}
\lim_{\varepsilon\rightarrow0}\int_{B_{1}}\left[X(u_{\infty}\wedge \underline{a})-\nabla^{\perp}\psi_{\infty}+\delta((1-\lambda)x,t)^{\perp}\right]\cdot\nabla\zeta_{\varepsilon}\\
=-\lim_{\varepsilon\rightarrow0}\frac{1}{\varepsilon}\int_{s-\varepsilon}^{s}l(s')ds'=l(s)\,.\end{gathered}$$ The sign of $m_{\infty}$ gives $$0\leq -\int_{B_{1}}\left[X(u_{\infty}\wedge \underline{a})-\nabla^{\perp}\psi_{\infty}+\delta((1-\lambda)x,t)^{\perp}\right]\cdot\nabla\zeta_{\varepsilon}\,.$$ As $\varepsilon\rightarrow0$ this implies $$0\leq\int_{E_{s}}\left[X(u_{\infty}\wedge \underline{a})-\nabla^{\perp}\psi_{\infty}+\delta((1-\lambda)x,t)^{\perp}\right]\cdot\nu\,.$$ Now define $E^{+}_{s}:=E_{s}\cap\{u_{\infty}>\underline{a}\}$ and $E^{-}_{s}:=E_{s}\cap\{u_{\infty}\leq\underline{a}\}$. For $(x,t)\in E^{-}_{s}$ we notice $$\left[X(u_{\infty}\wedge a)-\nabla^{\perp}\psi_{\infty}+\delta((1-\lambda)x,t)^{\perp}\right]\cdot\nu=\nabla^{\perp} h_{\delta}\cdot\nabla h_{\delta}=0\,.$$ Therefore it follows $$0\leq\int_{E_{s}^+}\left[X(\underline{a})-\nabla^{\perp}\psi_{\infty}+\delta((1-\lambda)x,t)^{\perp}\right]\cdot\nabla h_{\delta}\,.$$ In order to get a contradiction we claim $$\label{negativity_of_the_whole_expression}
\left(X(\underline{a})-\nabla^{\perp}\psi_{\infty}+\delta((1-\lambda)x,t)^{\perp}\right)\cdot\nabla h_{\delta}<0\,.$$ We rearrange terms $$\begin{gathered}
\label{rearanged_terms}
\left[X(\underline{a})-\nabla^{\perp}\psi_{\infty}+\delta((1-\lambda)x,t)^{\perp}\right]\cdot\nabla h_{\delta}\\
=X(\underline{a})\cdot\nabla g_{\infty}+\left(\nabla\psi_{\infty}-\delta((1-\lambda)x,t)\right)\left(\nabla^{\perp}g_{\infty}-X(\underline{a})\right)\,.\end{gathered}$$ We show (\[negativity\_of\_the\_whole\_expression\]), by proving that each term on the right hand side of (\[rearanged\_terms\]) is negative respectively strictly negative. Firstly we treat the first term and claim $$\label{negative_of_scalarproduct}
X(\underline{a})\cdot\nabla g_{\infty}<0\,.$$ A short calculation reveals $$\begin{aligned}
X(\underline{a})\cdot\nabla g_{\infty}&=f(\underline{a})u_{\infty}-f(u_{\infty})\underline{a}\\
&=f(\underline{a})(u_{\infty}-\underline{a})+(f(a)-f(u_{\infty}))\underline{a}\\
&=\underline{a}(u_{\infty}-\underline{a})\left(\frac{f(\underline{a})-f(0)}{\underline{a}}-\frac{f(u_{\infty})-f(\underline{a})}{u_{\infty}-\underline{a}}\right)\,.\end{aligned}$$ By convexity of $f$ we have in the case $\underline{a}<u_{\infty}<\overline{a}<0$ $$\frac{f(u_{\infty})-f(\underline{a})}{u_{\infty}-\underline{a}}<\frac{f(\overline{a})-f(\underline{a})}{\overline{a}-\underline{a}}<\frac{f(\underline{a})-f(0)}{\underline{a}}\,.$$ This implies $$\underline{a}\left(\frac{f(\underline{a})-f(0)}{\underline{a}}-\frac{f(u_{\infty})-f(\underline{a})}{u_{\infty}-\underline{a}}\right)\leq0$$ and henceforth (\[negative\_of\_scalarproduct\]), if $\overline{a}\leq0$. On the other hand if $0<\underline{a}<\overline{a}$, we get for $\xi\in(0,\underline{a})$, $\alpha\in(\underline{a},u_{\infty})$ $$\frac{f(\underline{a})-f(0)}{a}=f'(\xi)<f'(\underline{a})<f'(\alpha)=\frac{f(u_{\infty})-f(\underline{a})}{u_{\infty}-\underline{a}}\,,$$ which implies (\[negative\_of\_scalarproduct\]). Hence the first term of (\[rearanged\_terms\]) is non-positive and it remains to treat the second term. A short calculation gives $$\begin{gathered}
\left(\nabla\psi_{\infty}-\delta((1-\lambda)x,t)\right)\left(\nabla^{\perp}g_{\infty}-X(\underline{a})\right)\\
=(u_{\infty}-\underline{a})\left[\partial_t\psi(x_0,t_0)+\lambda\partial_x\psi(x_0,t_0)\frac{f(u_{\infty})-f(\underline{a})}{u_{\infty}-\underline{a}}+\delta((1-\lambda)x+t)\right]\,.\end{gathered}$$ Our choice of $\delta$ and $\lambda$ (see (\[choice\_of\_delta\_lambda\])) imply, that
$$\left(\nabla\psi_{\infty}-\delta((1-\lambda)x,t)\right)\left(\nabla^{\perp}g_{\infty}-X(\underline{a})\right)<0$$ and thus (\[negativity\_of\_the\_whole\_expression\]). Finally (\[negativity\_of\_the\_whole\_expression\]) implies $$\int_{E^{+}_{s}}\left(X(\underline{a})-\lambda\nabla^{\perp}\psi(x_{0},t_{0})+(1-\lambda)\delta(x,t)^{\perp}\right)\cdot\nabla h_{\delta}=0\,.$$ Since $$\left(X(\underline{a})-\lambda\nabla^{\perp}\psi(x_{0},t_{0})+(1-\lambda)\delta(x,t)^{\perp}\right)\cdot\nabla h_{\delta}<0$$ it follows $\mathcal{H}^{1}(E^{+}_{s})=0$, which is a contradiction to our choice of $s\in S$. Thus $$\partial_{t}\psi(x_{0},t_{0})+f\left(\partial_{x}\psi(x_{0},t_{0})\right)\geq0$$ as claimed. Henceforth $g$ is the viscosity solution of (\[HJ\]) and $u=\partial_x g$ the entropy solution of (\[scl\]) as claimed.
**Proof of Theorem 1** Thanks to Lemma \[lemma\_positiv\_density\] and Lemma \[proving\_entropy\] we can conclude the proof of Theorem \[minimality\_theorem\]. Indeed, we see that a weak solution $u\in L^{\infty}(\mathbb{R}\times[0,T))$ satisfying the assumptions of Theorem \[minimality\_theorem\], has by Lemma \[lemma\_positiv\_density\] ${\mathcal H}^1$-a.e. points of positive density, i.e. $$\limsup_{r\rightarrow0^+}\frac1r\int_{B_{r}(x_0,t_0)}\int_{\mathbb{R}}m(x,t,a)\,da\,dx\,dt\geq0\quad\mbox{for }{\mathcal H}^1\mbox{ a.e. }\quad (x_0,t_0)\in\mathbb{R}\times(0,T)\,.$$ By Lemma \[proving\_entropy\] we know then, that $u$ has to be entropic.
[99]{} Ambrosio, Luigi and Fusco, Nicola and Pallara, Diego, “Functions of bounded variation and free discontinuity problems”, *Oxford Mathematical Monographs*, New York, (2000). Ambrosio, Luigi and Lecumberry, Myriam and Rivi[è]{}re, Tristan, “A viscosity property of minimizing micromagnetic configurations”, *Comm. Pure Appl. Math., 56, No. 6*, (2003), 681-688. Ambrosio, Luigi and Kirchheim, Bernd and Lecumberry, Myriam and Rivi[è]{}re, Tristan, “On the rectifiability of defect measures arising in a micromagnetics model”, *Nonlinear problems in mathematical physics and related topics, II, Int. Math. Ser. (N. Y.), 2*, 29-60, Kluwer/Plenum, New York (2002). Bardi, Martino and Capuzzo-Dolcetta, Italo “Optimal control and viscosity solutions of Hamilton-Jacobi-Bellmann equations”, Birkhäuser Boston Inc, Boston MA 1997. Cannarsa, Piermarco and Sinestrari, Carlo, “Semiconcave functions, [H]{}amilton-[J]{}acobi equations, and optimal control”, *Progress in Nonlinear Differential Equations and their Applications, 58*, Birkhäuser Boston Inc., Boston, MA (2004). Dafermos, Constantine M. “Hyperbolic conservation laws in continuum physics” *Grundlehren der Mathematischen Wissenschaften 325*, Springer-Verlag, Berlin (2005). Ishii, Hitoshi, “Perron’s method for [H]{}amilton-[J]{}acobi equations”, *Duke Math. J., 55, No. 2*, (1987), 369-384. Kru[ž]{}kov, S. N., “First order quasilinear equations with several independent variables.”, *Mat. Sb. (N.S.), 81 (123)*, (1970), 228-255. Lecumberry, Myriam, “The rectifiability of shock waves for the solutions of genuinely non-linear scalar conservation laws in Structures geometriques des parois.”, *These de Doctorat de l’universite de Nantes*, (2003). Conway, E. D. and Hopf, E., “Hamilton’s theory and generalized solutions of the [H]{}amilton-[J]{}acobi equation”, *J. Math. Mech., 13*, (1964), 939-986. Ole[ĭ]{}nik, O. A., “On discontinuous solutions of non-linear differential equations”, *Dokl. Akad. Nauk SSSR (N.S.), 109*, (1956), 1098-1101. Perthame, Beno[î]{}t, “Kinetic formulation of conservation laws”, *Oxford Lecture Series in Mathematics and its Applications, 21*, (2002). Poliakovsky, Arkady, “On a variational approach to the method of vanishing viscosity for conservation laws”, *Adv. Math. Sci. Appl., 18, No. 2*, (2008), 429-451. Rivi[è]{}re, Tristan and Serfaty, Sylvia, “Limiting domain wall energy for a problem related to micromagnetics”’, *Comm. Pure Appl. Math., 54, No. 3 ,* (2001), 294-338. Rivi[è]{}re, Tristan, Serfaty, Sylvia “Compactness, kinetic formulation, and entropis for a problem related to micromagnetics”, *Comm. Partial Differential Equations, 28*, (2003), 249-269.
[^1]: Department of Mathematics, ETH Zentrum, 8093 Zürich, Switzerland.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We introduce a novel view to understand how dropout works, and propose a new neural network training method named *internal node bagging*, which explicitly forces a group of nodes to learn a certain feature in train time and combine those nodes to be one node in test time. It means we can use much more parameters to improve model’s fitting ability in train time while keeping model small in test time. We test our method on several benchmark datasets and find it can significantly improve test performance of small models.'
author:
- |
Shun Yi\
Department of Computer Science\
North China University of Technology\
`[email protected]`\
bibliography:
- 'reference.bib'
title: Internal node bagging
---
neural networks, ensemble learning, deep learning
Introduction
============
Neural network is a universal approximator, we can easily increase its fitting ability by adding more layers or more nodes each layer. As large labeled datasets are relatively easy to obtain now, neural network is widely used in computer vision, NLP and other domain. However, achieve state-of-the-art performance always need big models with regularization [@Goodfellow-et-al-2016], or even ensemble of several models, which limits the use of neural network especially on mobile device. Although some lightweight models have been proposed recently, like [@Howard2017MobileNets; @Zhang2017ShuffleNet], but they rely on well-designed structures and only focus on convolutional neural network.
Dropout [@Hinton2012Improving] is a famous regularization method in neural network training which randomly set the outputs of some hidden nodes or input nodes to zero in train time. It is commonly accepted that dropout training is similar to bagging [@Breiman1996Bagging]. For each training sample, dropout randomly deletes some nodes from network, and trains a thinner subnet, those subnets are trained on different samples and averaged in test time. Instead of making any real model average which will cost too much computing resource, a very simple approximate averaging method is applied by weight scaling. There are some empirical analyses show weight scaling works well in deep models [@Srivastava2014Dropout; @Wardefarley2014An; @Pham2014Dropout]. Dropout can indiscriminately and reliably yield a modest improvement in performance when applied to almost any type of model [@Goodfellow2013Maxout], but may not very efficiency on small model[@Srivastava2014Dropout].
In this paper, we introduce a novel view to understand how dropout works as a layer-wise ensemble learning method basing on several assumptions, and propose a new training method named *internal node bagging* according to our theory. We test our method on MINIST [@Lecun1998Gradient], CIFAR-10 [@Krizhevsky2009Learning] and SVHN [@Netzer2011Reading], with fully connected network and convolutional network, find it can significantly improve test performance of small models.
Motivation
==========
Consider a network that classify white horse and zebra in figure \[Fig:zebra\_net\]. This simple fully connect network has 4 input nodes which represent different features belong to horse and zebra. If we apply dropout on input layer and the only difference between white horse and zebra is black-white strip, then the network absolutely cannot work if the first node which represents black-white strip is dropped.
To avoid those features like black-white strip which are the key to classify different type of samples are dropped, network should learn to use more than one node to learn each of them. As we can see in figure \[Fig:grouped\_zebra\_net\] which is the multi nodes per feature edition, there are 4 groups of nodes, every group contains more than one node and all of those nodes represent same feature. If drop probability is 0.5 for every node and feature’s drop probability is less than 0.01, then every group should at least contain ${\log_{0.5} {0.01}}$ nodes.
Although what we mentioned above is too idealized because features learned by neural network are distributed, there are still some empirical evidences support our assumption, for example, dropout training always need bigger models [@Srivastava2014Dropout], shutting off a hidden neuron in dropout network can not simply remove features of input [@Bouthillier2015Dropout], and there is significant redundancy in the parameterization of several deep learning models [@Shakibi2013Predicting].
Consider a fully connected dropout network which every feature is represented by a group of nodes like network in figure \[Fig:grouped\_zebra\_net\], the $l-1$th layer contains $k$ groups and every group has $n$ nodes, $y_{ij}^{l-1}$ is the output of the $j$th node from $i$th group of layer $l-1$, $w_{ij}^l$ and $b^l$ are the corresponding weights and biases, $m_{ij}^{l-1}$ is drop mask, $f$ is activation function, $y^l$ is the output of layer $l$. Then the forward propagation is:
$$y^l=f(\sum_i^k\sum_j^nw_{ij}^l(m_{ij}^{l-1}y_{ij}^{l-1})+b^l) \label{eq:prime}$$
Because all nodes in a group represent same feature, so we can assume their weights will converge to similar value as the training process going. Let the value be $w_i^l$, so we can simplify \[eq:prime\] to:
$$y^l=f(\sum_i^kw_i^l\sum_j^nm_{ij}^{l-1}y_{ij}^{l-1}+b^l) \label{eq:grouped}$$
According to \[eq:grouped\], we can simplify neural network with dropout to network described in figure \[Fig:internal\_bagging\], which is different to network in figure \[Fig:grouped\_zebra\_net\] that every group only has one output sampled from corresponding nodes.
![*Internal node bagging* style network.[]{data-label="Fig:internal_bagging"}](figure/internal_bagging_net.pdf){width="0.5\linewidth"}
Network described in figure \[Fig:internal\_bagging\] is very similar to maxout [@Goodfellow2013Maxout], but instead choose the biggest output in a group, it randomly samples one. Let the sampled output be $s_i^l$, it computed by:
$$s_i^l=\sum_j^nm_{ij}^ly_{ij}^l \label{eq:sample}$$
Now we have a novel view to understand how dropout works as a layer-wise ensemble training method: For every feature in a layer, there are a group of nodes to learn it, next layer randomly samples a value from those nodes as the feature activation while training. In test time, weight scaling approximately let every group output the expected feature activation.
We consider that if we know which nodes in a layer represent same feature, then we may be able to combine those nodes to be one node in test time which will reduce tons of parameters and computation. It can also be interpreted from opposite perspective: for every internal node in a small network, we use a group of nodes to estimate its parameters in train time. We name this method *internal node bagging*.\[Sec:understanding\]
There are 2 problems we should resolve, the first is how do we know which nodes represent same feature, the second is how to combine nodes to be one node. For resolving the 2 problems, we apply follow 2 tricks:
- We manually assign nodes to different groups and force them learn same feature in train time. For every group, we initialize them with same initialization and periodically compute average weights and biases and assign it to all nodes.
- We use relu [@Glorot2012Deep] in all experiments[^1]. We assume the outputs of nodes in a group are similar as they represent same feature, so it is highly possible that those outputs distribute on the linear part of relu, and combining nodes in test time is feasible.
We outline detail model description in section \[Sec:motivation\], and experiment results in section \[Sec:experiments\].
Model description {#Sec:motivation}
=================
In this section, we first introduce how to combine nodes in a group to be one node and how to compute average weights, then we introduce 2 methods to sample a feature activation from a group used in our experiments.
Combine nodes {#Sec:comine_nodes}
-------------
For given input, the expected value sampled from a group is: $$\begin{aligned}
E_m(s_i^l) &= E_m(\sum_j^nm_{ij}^ly_{ij}^l)= \sum_j^nE_m(m_{ij}^l)y_{ij}^l \label{eq:expection_1}
\end{aligned}$$ For all $m_{ij}^l$ in a group, they obey same distribution, so, let: $$E_m(m_{ij}^l)=E_m(m_i^l)$$ Then, we can simplify \[eq:expection\_1\] to: $$E_m(s_i^l) = E_m(m_i^l)\sum_j^ny_{ij}^l \label{eq:expection_2}$$ Consider[^2]: $$y_{ij}^l = relu(w_{ij}^ly^{l-1}+b_{ij}^l)$$ Assume for all $y_{ij}^l$ in a group, they distribute on the linear part of relu, so: $$E_m(s_i^l) = relu((E_m(m_i^l)\sum_j^nw_{ij}^l)y^{l-1}+E_m(m_i^l)\sum_j^nb_{ij}^l) \label{eq:expection_4}$$ We combine nodes in a group to be one node in test time according to \[eq:expection\_4\], which is equal to weight scaling if every group only contain one node.
Compute average weights {#Sec:average weights}
-----------------------
Consider a layer $l$ and group $i$, average weights are $w^{avg}$, average biases are $b^{avg}$, then: $$nE_m(m_i^l)relu(w^{avg}y^{l-1}+b^{avg})=E_m(s_i^l)$$ So: $$\begin{aligned}
w^{avg}&=\frac{1}{n}\sum_j^nw_{ij}^l \\
b^{avg}&=\frac{1}{n}\sum_j^nb_{ij}^l
\end{aligned}$$ We periodically compute average weights and biases in a group, and assign it to all nodes in this group to force them learn same feature.
Sample methods
--------------
We propose 2 methods to sample an activation from a group:
- Method A: every node in a group has same probability to be sampled independently.
- Method B: only one node will be sampled every group.
If there is only one node every group, method A is equal to dropout, network apply method B is equal to standard network.
Experiments {#Sec:experiments}
===========
We evaluate our methods on MINIST [@Lecun1998Gradient], CIFAR-10 [@Krizhevsky2009Learning] and SVHN [@Netzer2011Reading]. MNIST dataset consists of 28\*28 pixel gray images of handwritten digits, with 60000 samples for training and 10000 samples for testing; CIFAR-10 dataset consists of 32\*32 RGB images in 10-classes with 50000 images for training and 10000 for testing; SVHN dataset consists of 32\*32 RGB image dataset of digits, with 73257 images for training and 26032 images for testing.
We implement our models using tensorflow, all source code is available in [www.github.com/Xiong-Da/internal\_node\_bagging\_V2](www.github.com/Xiong-Da/internal_node_bagging_V2). Settings shared in all experiments are listed below:
- We apply *internal node bagging* on all hidden layers.
- All sample probability used in method A is 0.5.
- We don’t use any other normalization method.
- We default apply weight average described in \[Sec:average weights\] every 10 epochs.
For experiments on MNIST, we use fully connected network with 2 same width hidden layers. For experiments on CIFAR-10 and SVHN, we use CNN described in table \[Tab:CNN\], which is modified from the “base model C” in [@Springenberg2014Striving], just remove last 1\*1 convolution layer, all stride in convolution layer is 1, all padding is “SAME” except last 3\*3 convolution layer.
We train all models with Adam optimization algorithm [@Kingma2014Adam]. For experiments on MNIST, we train first 100 epochs with learning rate 1e-3, and train another 100 epochs with learning rate 1e-4. For Experiments on CIFAR-10 and SVHN, we train models with initial learning rate 1e-3, and decay learning rate when validate error stop decrease untill models are converged.
32\*32 RGB image
-----------------------------------------------
2 layer of 3\*3 conv.64
3\*3 max-pooling stride 2
2 layer of 3\*3 conv.128
3\*3 max-pooling stride 2
3\*3 conv.192
1\*1 conv.192
global averaging over 6\*6 spatial dimensions
10-way softmax
: The architecture of CNN used for classification experiments on CIFAR-10 and SVHN.[]{data-label="Tab:CNN"}
Performance on models of different size {#Sec:performance}
---------------------------------------
In this section, we investigate the performance of our methods on models with variety size.
Figures in \[Fig:method\_A\] show the performance of method A on 3 datasets. As we can see in those figures, when model size is small, increasing group size can significantly improve test performance especially on CIFAR-10 and SVHN. But as model size increase, the performance improvement start to decrease, test error of models with big group size is even worse than dropout network on SVHN (method A with group size 1 is equal to dropout).
Figures in \[Fig:method\_B\] show the performance of method B. Compare to method A, method B is relatively more complex to analyze. On MNIST dataset, increasing group size can modestly improve performance both on small models and big models. On CIFAR-10 datatset, models with different group size have relatively similar performance, models with big group size perform slightly better. On SVHN dataset, method B can significantly improve performance especially on small models.
Effect of weight average
------------------------
Figure in \[Fig:256\_average\] shows the effect of weight average described in \[Sec:average weights\] on MNIST dataset with model width 256. “weight average frequency” mean train how many epochs and then apply weight average once, we only train 200 epochs on MNIST dataset, so frequency 200 means don’t apply it. In our experiments, method B seemed not sensitive to weight average frequency, but method A can’t converge well without moderate frequency, especially on models with large group size.
![Analyze the effect of weight average described in \[Sec:average weights\].[]{data-label="Fig:256_average"}](figure/256_average.pdf){width="0.9\linewidth"}
Convergence propertie
---------------------
Figure in \[Fig:256\_converge\] shows the convergence properties of our 2 methods on MNIST dataset with model width 256. As we can see in those figure, for both methods, models with big group size do converge slower, but not slow too much.
![Analyze the convergence properties.[]{data-label="Fig:256_converge"}](figure/256_converge.pdf){width="0.9\linewidth"}
Performance on different activation function
--------------------------------------------
![Analyze the performance of our methods with different activation function and group size.[]{data-label="Fig:activation_functions"}](figure/activation_functions.pdf){width="0.9\linewidth"}
In section \[Sec:comine\_nodes\], we assume outputs of nodes from a group is similar and distribute on the linear part of relu as they represent same feature, and this is why combining nodes in a group to be one node in test time is feasible. In this section, we analyze that if relu is inreplaceable to our method. Figure in \[Fig:activation\_functions\] shows the experiment results of our 2 methods with 3 different activation functions on MNIST dataset. For method A, all 3 activation functions perform better when increase group size to 2, but only relu’s performance keep improving when increase group size to 4. For method B, when increase group size to 2, relu and tanh perform better, but when increase group size to 4, both 2 activation function don’t perform good.
In our experiment, relu is not inreplaceable, but do perform slightly better.
Discussion
==========
We introduced a novel view to understand how dropout works as a layer-wise ensemble learning method, and proposed a new ensemble training algorithm named *internal node bagging*. We tested 2 sample methods in our experiments: method A can be seen as generalization of dropout, method B can be seen as generalization of standard network. For method A, increasing group size can significantly improve test performance on small models. For method B, increasing group size can moderately improve performance both on small models and big models, but it performs quite different on 3 different datasets. We also introduced 2 way to understand how *internal node bagging* works: the first one thinks our method is equivalent to simplify big redundant models to small models without redundancy, the second one thinks our method is equivalent to estimate the parameters of every internal node by multiple nodes in train time. It seems the second one is more reasonable basing on our experiments.
[^1]: except in the experiments of comparing the performance of different activation functions
[^2]: $w_{ij}^l$ here is different from it in \[eq:prime\], they just represent corresponding weights
| {
"pile_set_name": "ArXiv"
} |
[ **Deterministic and stochastic aspects of the transition to turbulence**]{}
Baofang Song$^{1,2}$ and Björn Hof$^{1,2}$\
[$^{1}$ IST Austria (Institute of Science and Technology Austria), 3400 Klosterneuburg, Austria\
$^{2}$ Max Planck Institute for Dynamics and Self-Organization, Bunsenstrasse 10, Göttingen, Germany]{}\
[**Abstract**]{}\
The purpose of this contribution is to summarize and discuss recent advances regarding the onset of turbulence in shear flows. The absence of a clear cut instability mechanism, the spatio-temporal intermittent character and extremely long lived transients are some of the major difficulties encountered in these flows and have hindered progress towards understanding the transition process. We will show for the case of pipe flow that concepts from nonlinear dynamics and statistical physics can help to explain the onset of turbulence. In particular the turbulent structures (’puffs’) observed close to onset are spatially localized chaotic transients and their lifetimes increase super exponentially with Reynolds number. At the same time fluctuations of individual turbulent puffs can (although very rarely) lead to the nucleation of new puffs. The competition between these two stochastic processes gives rise to a non-equilibrium phase transition where turbulence changes from a super-transient to a sustained state.\
Introduction
============
[\[sec:intro\]]{} How turbulence first arises in simple shear flows has remained an open question for well over a century. Osborne Reynolds [@reynolds1883] was the first to observe that this transition depends on a dimensionless group, i.e. the Reynolds number, as well as on the amplitude of disturbances present in the system. Some of the leading theorists at the time (e.g. Lord Kelvin, Lord Rayleigh, Arnold Sommerfeld, Werner Heisenberg, Hendrik Antoon Lorentz [@Eckert2010]) attempted to probe the stability of pipe and related shear flows (i.e. channel and Couette flow) with essentially linear methods. After many unsuccessful attempts it has become clear (e.g. see [@Drazin_Reid1981]) that the occurrence of turbulence is unrelated to the stability of the laminar state, as Reynolds had already concluded from his experimental observations many years earlier. While pipe and Couette flow are believed to be stable for all Reynolds numbers, plane Poiseuille (i.e. channel) flow becomes unstable at Re of about 5800. However in the latter case turbulence is typically already observed at Reynolds numbers a little above 1000 and to hold the flow laminar up to the linear stability threshold actually requires considerable effort in experiments.\
Transition in the type of flows discussed above is qualitatively different from the classical pictures for the transition to turbulence which goes back to Landau and to Ruelle and Takens [@ruelletakens]. In both scenarios turbulence arises following a sequence of instabilities of the base flow (which hence becomes linearly unstable). While the Ruelle Takens type transition has been observed in several closed flows (e.g. [@Gollub_Swinney1975]) it is noteworthy that even in this case the transition sequence only explains the onset of comparably low dimensional chaotic motion which dynamically is still far from the full spatio-temporal complexity encountered in turbulent flows. In open shear flows such as pipes all experimental observations show that the transition in contrast is rather abrupt directly from laminar to turbulent. The latter type of transition has turned out to be far more difficult to understand.\
In more recent years a new transition mechanism has been proposed based on dynamical systems concepts. Invariant solutions of the Navier Stokes equations such as periodic orbits or traveling waves are deemed to be ultimately responsible for the existence of the turbulent state. These new solutions [^1] arise as the Reynolds number is increased and importantly are entirely disconnected form the laminar flow. The proposition is then that chaotic and ultimately turbulent motion will arise following instabilities of these disconnected solutions (independent of the laminar state). Over the last two decades or so many traveling waves and periodic orbits have been found in direct numerical simulations for shear flows (for reviews see [@Kerswell2005; @Eckhardt2007; @Kawahara2012]. While coherent structures resembling travelling waves have also been observed in turbulent flows in experiments [@Hof2004; @Hof2005; @Lozar2012] a main discrepancy remains: close to onset turbulence is always confined to structures localized in the streamwise direction and surrounded by laminar flow, called ‘puffs’ for the case of pipe flow. Invariant solutions in this $Re$ range on the other hand are usually periodic in this direction. Only very recently the first streamwise localized invariant solutions were discovered for pipe flow[@Avila2013]. In addition it could be shown that chaotic motion indeed originates from one such localized periodic orbit. Albeit in this study the dynamics were limited to a symmetry subspace it nevertheless illustrates how turbulence can arise from a simple invariant solution, unrelated to the laminar state, as proposed above. Also this study could explain the origin of another property of localized puffs in pipe flow which is their transient nature. How (and if) turbulence develops from a transient to a sustained state has been subject to much recent debate. As we will argue below spatial aspects are crucial in the transition mechanism and eventually lead to a non equilibrium phase transition giving rise to sustained turbulence. In the following we will discuss this transition in more detail and briefly review important recent results.\
Discussion
==========
In pipe flow turbulent puffs (Fig. \[fig:puff\_spacetime\](top)) are typically observed in a Reynolds number regime of approximately $1700\lesssim Re\lesssim 2300$. They result from perturbations of finite amplitude and in experiments if no great care is taken they will often result from distortions the flow experiences directly at the inlet. If the inlet is designed carefully to avoid such disturbances and if the pipe is sufficiently straight and smooth, flows can be held laminar up to much higher $Re$ ( the record in experiments is currently at $Re$=100000). For controlled studies of transition it is desirable to start with a laminar flow and to study its response to a well controlled disturbance which for example can be a jet of fluid injected for a brief period through a small hole in the pipe wall. If the amplitude of such a perturbation is large enough a turbulent puff is created. While directly after the flow has been perturbed the dynamics depend on the nature of the disturbance (the puff first has to develop), after 100 to 150 advective time units (measured in D/U, where D is the diameter and U the mean velocity) the resulting puff is independent of the perturbation that triggered it.\
Curiously in this Reynolds number regime turbulence always remains localized. Even if an extended (axially) part of the pipe is disturbed the flow will always arrange itself in turbulent segments (i.e. puffs) which are about $5$ to $10 D$ long (the turbulent core excluding the leading edge) and interspersed by laminar fluid. Vorticity isosurfaces of a turbulent puff at $Re=2200$ are shown in Fig. \[fig:puff\_spacetime\](top). Fig. \[fig:puff\_spacetime\](bottom) shows energy levels in a space time plot (time from bottom to top) of a simulation of an initially (t=0) fully turbulent flow. Upon reduction of $Re$ to 2200 laminar regions (blue) appear and turbulence becomes confined to localized regions, i.e. puffs (red vertical stripes in Fig. \[fig:puff\_spacetime\](bottom)). The direct numerical simulations were carried out with a spectral code [@Willis2009]. Fourier modes in the axial and azimuthal direction and finite difference in the radial are employed and the resolution chosen here is $48\times(\pm2048)\times(\pm40)$ in radial, axial, and azimuthal directions. The domain is $180D$ in the axial direction ($D$ being the diameter) with periodic boundary conditions.
 Isosurfaces of the streamwise vorticity of a puff at $Re=2200$, taken from the segment denoted by a black horrizontal bar in the bottom figure. Flow is from left to right. (bottom) Space-time diagram for the reduction from $Re$=2800 to 2200 in a comoving frame with the mean flow. The local turbulence intensity $q(z)=\int\int(u_r^2+u_{\theta}^2)rdrd\theta$ is plotted versus aixal position in a logrithmic color scale. ](Re2200_180D_puff_spacetime.eps){width="0.8\linewidth"}
As pointed out in [@Hof2010] at Reynolds numbers close to 2000 turbulent puffs extract energy from the adjacent (actually upstream) laminar parabolic flow. The plug like turbulent profile in the central part of the puff is unable to sustain turbulence (or rather to feed turbulence in the downstream direction) consequently the turbulence intensity decreases along the leading edge of the puff in the downstream direction and the flow eventually relaminarises. Due to the action of viscosity the laminar profile now begins to recover its parabolic shape so that at sufficient distance a second puff can be sustained. If on the other hand the distance between puffs is too short the downstream puff will border onto fluid with a flatter plug like profile in the upstream direction. Consequently it cannot extract sufficient kinetic energy from the flow upstream and decays as shown in [@Hof2010]. The interaction distance between two puffs is approximately $20D$ [@Samanta2011]. As a result turbulent puffs have a minimum spacing of that same distance. While this argument qualitatively explains why fully turbulent flow cannot be sustained at these low Re, the energetic aspects of this process are not understood in full detail.\
A key attribute of puffs is their highly chaotic dynamics. This gives rise to a loss of memory and limits the prediction of the flows future evolution. To illustrate the sensitive dependence on initial conditions we simulated two puffs with velocity fields that only deviate by $10^{-10}$. As usual for chaotic systems this deviation grows exponentially and as can be seen form the energy time series shown in Fig. \[fig:Re1850\_sensitive\] the signals notably depart and become completely unrelated after about 200 advective time units $D/U$ (the time the puffs takes to travel 200 pipe diameters downstream). The loss of predictabiliy becomes especially clear when one puff suddenly decays (green curve in Fig. \[fig:Re1850\_sensitive\]) while the other continues unchanged (i.e. the average quantities remain unchanged).\
It is a typical feature of puffs that they live for very long times and decay suddenly (see Fig. \[fig:Re1850\_sensitive\]). Extensive statistical studies have shown that the survival probability is exponentially distributed and the decay is memoryless. This behaviour is in line with the escape from chaotic repellers observed in lower dimensional systems [@Grebogi1983; @Do2004]. Here a chaotic attractor turns into a chaotic saddle after an unstable periodic orbit within the attractor and one on the basin boundary collide (unstable-unstable pair bifurcation). Above the bifurcation point chaotic transients persist for very long times before they eventually decay.\
![\[fig:Re1850\_sensitive\]Sensitive dependence on initial conditions of a puff at $Re=1850$. These two lines are the time traces of the kinetic energy $E_{3D}$ of two runs starting with two very close initial conditions separated by $\sim 10^{-10}$. While one decays after about 300 time units, the other persists. Eventually the second one will also decay (not shown in the figure) due to the transient nature of puffs](Re1850_180D_sensitive.eps){width="0.8\linewidth"}
A similar scenario has recently been observed for pipe flow[@Avila2013]. In this numerical study the dynamics were confined to a symmetry subspace (imposing a mirror and a 2-fold rotation symmetry with respect to the pipe axis). While this somewhat simplifies the dynamics the flow still exhibits turbulent motion for large enough Re. By following the laminar turbulent boundary (the so called “edge state” [@Schneider2007], which in this case is a localised periodic orbit) to lower Re the saddle node bifurcation where the periodic orbits originates was reached (at $Re$=1430). The upper branch of this saddle node however is a stable periodic orbit(see Fig. \[fig:splitting\]) and again localised and its length is comparable to that of puffs.
![\[fig:orbit\]Stable periodic orbit in direct numerical simulations of pipe flow with an imposed 2-fold rotation and a reflection symmetry. (Note that this is necessarily also a numerical solution of the full Navier Stokes equation.) This orbit arises in a saddle node bifurcation and is originally stable in the sub-space. At higher Re a bifurcation sequence leads to chaos and transient turbulent puffs. The periodic orbit is shown at $Re=1490$ ( where $Re=UD/\nu$, $\nu$ is the kinematic viscosity, $D$ the pipe diameter and $U$ the mean velocity). Isosurfaces in red/blue show velocity deviations of $+/- 0.1U$ from the laminar parbolic profile. Yellow/ cyan mark isosurfaces of positive/negative streamwise (i.e.axial) vorticity. Upstream and downstream of the periodic orbit the velocity filed quickly approaches the laminar parabolic profile. ](periodicorbit.eps){width="0.8\linewidth"}
For increasing Re the orbit first undergoes a secondary Hopf bifurcation followed by the formation of a chaotic attractor. This is the point where chaotic motion originates in this subspace. The basin of the attractor increases rapidly with $Re$ until it reconnects with the unstable periodic orbit on the basin boundary. At this moment the chaotic dynamics turn into long lived transients (presumably following an unstable-unstable pair bifurcation). The memoryless nature of the decay and the loss of predictability is a direct consequence of the sensitive dependence on initial conditions characteristic for deterministic chaos.\
While the classical picture of turbulence is that of a chaotic attractor, dating back to the landmark paper of Ruelle and Takens, this approach only takes the temporal dynamics into account and neglects the spatial complexity which is however intrinsic to turbulent flows. The importance of spatial aspects and the spatio temporal intermittent character of the turbulence transition have been emphasized in studies of model system [@Kaneko1985; @Chate1987]. Also the chaotic attractor hypothesis has been questioned in 1980’s by Crutchfield and Kaneko [@Crutchfield] who propose that chaotic super transients are more relevant to turbulence. A first observation supporting this view came from direct numerical simulations of pipe flows where transients were observed [@Brosa1989]. A number of later studies were concerned with long lived transients at low Reynolds numbers in pipe flow [@Faisst2003; @Peixinho2006; @Hof2006; @Willis2007; @Hof2008; @Avila2010; @Kuik2010] and found that the decay is a memoryless process with a characteristic lifetime $\tau$ which is a function of Re. There was however no consensus if or if not the lifetimes of individual puffs became infinite or remained finite. A number of studies proposed that the turbulent flow decay rates (inverse characteristic lifetimes) scale as $\tau^{-1}\sim(Re_c-Re)$ [@Willis2009; @Faisst2003; @Peixinho2006] or at least as a power law $\tau^{-1}\sim(Re_c-Re)^n$ [@Willis2009]. This view was questioned by Hof [@Hof2006] who found that instead $\tau^{-1}\sim exp(Re)$ implying that turbulence remains transient. In a refined study for much longer observation times than in any previous study (spanning almost 8 orders of magnitude in time) $\tau^{-1}$ was found to scale super exponentially with $Re$ [@Hof2008] [^2]. This scaling was confirmed by another experiment [@Kuik2010] and in direct numerical simulations [@Avila2010]. It is remarkable that based on their studies of spatially coupled maps Crutchfield and Kaneko had not only proposed that fluid turbulence could evolve around spatially coupled transients but also that the lifetimes under certain conditions (type 2 supertransients) are memoryless and increase superexponentially with system size. While it has been argued that an increase in Re in turbulent flows is analogous to an increase in system size in coupled chaotic maps this correpondence is however not entirely clear. The lifetime studies in pipe flow were carried out for single turbulent puffs (not many spatially coupled ones) and the puff size hardly changes with $Re$. One may argue that the smallest scales of turbulence decrease with $Re$ and hence the system size based on this smallest scale increases. Nevertheless the Reynolds number range over which the lifetime increase is observed is relatively small ($1700<Re<2050$) and hence this size effect will be only very moderate.
A more direct analogy can be drawn to another model system of coupled chaotic maps which is motivated by bistable excitable media [@Barkley2011]. A key difference to the above mentioned map models is that here the susceptibility of a “laminar” site to perturbations from neighbouring chaotic sites (and hence the minimum perturbation amplitude to trigger chaotic dynamics at a laminar site), decreases with $R^{-1}$ (where $R$ is a the control parameter analogous to the Reynolds number). This model input reflects experimental observations of pipe flow where the minimum perturbation amplitude to trigger turbulence was found to scale with $Re^{-1}$ [@Hof2003; @Hof2004; @Mullin]. In the model just like in the experiments localized excited states, i.e. puffs, with transient lifetimes are observed and with an increase in the control parameter lifetimes of individual puffs scale faster than exponential and hence remain transient.\
Consequently in pipe flow the increase in the temporal complexity alone does not lead to sustained turbulence. As proposed by Moxey and Barkley [@Moxey2010] and later explicitly shown by Avila et al. [@Avila2011] turbulence becomes sustained by a spatial growth process called puff splitting. Puff splitting is commonly observed at Reynolds numbers of around 2300 [@Wygnanski1975; @Nishi2008] and while here turbulence is still confined to puffs typically $5$ to $10 D$ in length puff sizes fluctuate and can occasionally reach larger values. In these instances in a small number of cases a segment of turbulent fluid at the leading edge of the puff can escape further downstream beyond the puff-puff interaction distance and a new puff develops here (see Fig. \[fig:splitting\]). This splitting process leads to an increase in turbulent fraction.
![\[fig:splitting\]Puff splitting process. At the puff leading edge vortices are shed in the downstream direction. While normally they decay, occasionally in rare cases they manage to escape beyond the interaction distance of the upstream puff. The first puff and the vortex patch are now separated by a region of laminar fluid and the vortex patch grows to a new puff. The isosurfaces correspond to the axial vorticity component (positive in blue, negative in yellow). ](splitting.eps){width="0.8\linewidth"}
As shown by Avila et al. [@Avila2011] puff splitting is also intrinsically memoryless and can already be found at much lower Re as previously expected. The characteristic time for such an event to occur decreases super-exponentially with Re. The argument for turbulence to become sustained is now straightforward. If the characteristic time for turbulent puffs to decay is smaller than the time for new puffs to be created (i.e. by splitting) turbulence will eventually decay. If on the other hand new puffs are created faster than existing ones decay in the thermodynamic limit turbulence becomes sustained. The critical Reynolds number where turbulence changes from a transient to a sustained state can be estimated by the intersection point of the characteristic time scales of the two processes shown in Fig. \[fig:decay\_split\].\
This transition is analogous to non equilibrium phase transitions and as we will argue below, bears close resemblance to directed percolation (DP) and related contact processes. First speculations about a possible connection between transition in linearly stable shear flows and DP date back to Pomeau in 1986 [@Pomeau1986]. Just like in DP, pipe flow has a unique absorbing state which is the laminar flow. Once turbulence has decayed the flow cannot by itself return to turbulence unless it is disturbed from the outside. The recent studies of puff decay and splitting [@Hof2008; @Avila2011] also suggest that the interaction is only short range, which is another requirement for DP [@Hinrichsen2000]. No observations indicate that a localized puff would create a second one in a part of the pipe not adjacent to it (i.e. more than 25 D or so away). Equally it has never been observed that puffs would influence the lifetimes of other puffs that are sufficiently far away. These recent studies also infer that if such a realtion to DP exists a single unit (i.e. a lattice point in DP) must correspond to a turbulent puff and not for example to a single vortex. To explore this analogy further hence requires much larger system sizes allowing to follow the evolution of many turbulent puffs(/spots). While a number of earlier studies (e.g. [@bottin]) looked at such aspects retrospectively the system sizes used were too small.
![\[fig:decay\_split\]The intersection point between characteristic life- and splitting times determines the onset of sustained turbulence in pipe flow. To the right of the intersection point on average new puffs are created faster (by puff splitting) than existing puffs decay and hence (in the thermodynamic limit) spatio-temporally intermittent turbulence persists. The data is taken from [@Hof2008; @Avila2011]. Note that both the decay and the splitting process are memoryless and described by characteristic timescales (see original papers for details](decay_splitting.eps){width="0.65\linewidth"}
In a numerical investigation [@Shi2013] of plane Couette we consequently chose a narrow but very long domain so that a large number of turbulent stripes (analogous structure to puffs in pipes) could be accommodated. This study gave more direct evidence that the transition is indeed a non-equilibrium continuous phase transition. Here just like in pipe flow super-exponential lifetime and splitting statistics were observed for single stripes, and a critical point for the onset of sustained turbulence could be determined in the same manner described above. The much shorter time scales at the intersection point between decay and splitting curves allowed to resolve size distributions at this point. The distributions of laminar gaps exhibit scale invariance supporting the proposition of a continuous phase transition. Also here the same transition scenario between transient localized chaos and sustained spatio temporal intermittent chaos was found.\
Analogies to DP have also been explored in recent theoretical studies, e.g. [@Barkley2011; @Sipos2011; @Allhoff2012] In particular for the coupled map model presented in [@Barkley2011] close agreement was found to the experimental results for pipe flow: the turbulent state becomes sustained when the splitting outweighs the decay of individual puffs. In addition the critical exponent for the increase of the turbulent fraction above onset was found to agree well with the universal one for DP in $1+1$ dimension.\
While at present a final answer to the question if the laminar turbulence transition is a non-equilibrium phase transition in accordance with DP is outstanding experiments and numerical simulations to clarify this question are under way. One of the main challenges here is to resolve the extremely long time scales relevant in the vicinity of the transition point (note that characteristic splitting and decay times in pipes correspond to almost $10^8$ advective time units!). In this time puffs trave a distance correpsonding to $10^8$ pipe diameters. Equally an accuracy in the Reynolds number of about 0.1% is required setting a further challenge for experiments. A further open issue is the transition from spatially intermittent turbulence (i.e. puffs) to expanding space filling turbulent structures which takes place somewhere between Reynolds numbers 2300 and 3000.
[88]{}
Reynolds O, , 1883 [*Proc. R. Soc. London*]{} [**34**]{} 84-99
Drazin P G and Reid W H, , 1981 Cambridge University Press Ruelle D and Takens F, , 1971 [*Comm. Math. Phys.*]{} [**20**]{} 167-192
Gollub J P and Swinney H L, , 1975 [*Phys. Rev. Lett.*]{} [**35**]{} 14
Nagata M, , 1990 [*J. Fluid Mech.*]{} [**217**]{} 519-527
Kerswell R R, , 2005 [*Nonlinearity*]{} [**18**]{} R17
Eckhardt B, Schneider T M, Hof B, and Westerweel J, , 2007 [*Annu. Rev. Fluid Mech.*]{} [**39**]{} 447-68
Kawahara G, Uhlmann M, and van Veen L, , 2012 [*Annu. Rev. Fluid Mech.*]{} [**44**]{} 203-25
Hof B, van Doorne C W H, Westerweel J, Nieuwstadt F T M, Faisst H, Eckhardt B, Wedin H, Kerswell R R, and Waleffe F, , 2004 [*Science*]{} [**305**]{} 1594-1598
Hof B, van Doorne C W H, Westerweel J, and Nieuwstadt F T M , 2005 [*Phys. Rev. Lett.*]{} [**95**]{} 214502
de Lozar A, Mellibovsky F, Avila M, and Hof B, , 2012 [*Phys. Rev. Lett.*]{} [**108**]{} 214502
Avila M, Mellibovsky F, Roland N, and Hof B, , 2013 [*Phys. Rev. Lett.*]{} [**110**]{} 224502
Willis A P and Kerswell R R, , 2009 [*J. Fluid Mech.*]{} [**619**]{} 213-233
Hof B, de Lozar A, Avila M, Tu X, and Schneider T M, , 2010 [*Science*]{} [**327**]{} 1491-1494
Samanta D, de Lozar A, and Hof B, , 2011 [*J. Fluid Mech.*]{} [**681**]{} 193–204
Schneider T M, Eckhardt B, and Yorke J A, , 2007 [*Phys. Rev. Lett.*]{} [**99**]{} 034502
Kaneko K, , 1985 [*Prog. Theor. Phys.*]{} [**74**]{} 1033-1044
Crutchfied J.P., Kaneko K, , 1985 [*Phys. Rev. Lett.*]{} [**60**]{} 2715-2718
Chaté H and Manneville P, , 1987 [*Phys. Rev. Lett.*]{} [**58**]{} 112-115
Grebogi C, Ott E, and Yorke J A, , 1983 [*Phys. Rev. Lett.*]{} [**50**]{} 935-938
Do Y and Lai Y C , 2004 [*Europhys. Lett.*]{} [**67**]{} 914-920
Brosa U, , 1989 [*Journal of Statistical Physics*]{} [**55**]{} 1303-1312
Faisst H, , 2003 [*Ph.D., thesis*]{} Universität Marburg Peixinho J and Mullin T, , 2006 [*Phys. Rev. Lett.*]{} [**96**]{} 094501
Hof B, Westerweel J, Schneider T M, and Eckhardt B, , 2006 [*Nature*]{} [**443**]{} 59-62
Willis A P and Kerswell R R, , 2007 [*Phys. Rev. Lett.*]{} [**98**]{} 014 501
Hof B, de Lozar A, Kuik D J, and Westerweel J, , 2008 [*Phys. Rev. Lett.*]{} [**101**]{} 214501
Avila M, Willis A P, and Hof B, , 2010 [*J. Fluid Mech.*]{} [**646**]{} 127-136
KUIK D J, Poelma C, and Westerweel J, , 2010 [*J. Fluid Mech.*]{} [**645**]{} 529-539
Goldenfeld N, Guttenberg N, and Gioia G, , 2010 [*Phys. Rev. E*]{} [**81**]{} 035304
Barkley D, , 2011 [*Phys. Rev. E*]{} [**84**]{} 016309
Hof B, Juel A, and Mullin T, , 2003 [*Phys. Rev. Lett.*]{} [**91**]{} 244502
Hof B, , 2005 [In [*[ Laminar-Turbulent Transition and Finite Amplitude Solutions,]{}*]{} ed. T. Mullin, R.R. Kerslwell, Springer]{} 221-231
Moxey D and Barkley D, , 2010 [*Proc. Natl. Acad. Sci. USA* ]{} [**107**]{} 8091
K. Avila, D. Moxey, A. de Lozar, M. Avila, D. Barkley and B. Hof, (2011), , 2011 [*Science*]{} [**333**]{} 192-196
Wygnanski I J, Sokolov M, and Friedman D, , 1975 [*J. Fluid Mech.*]{} [**69**]{} 283-304
Nishi M, Ünsal B, Durst F, and Biswas G, , 2008 [*J. Fluid Mech.*]{} [**614**]{} 425-446
Pomeau Y, , 1986 [*Physica D*]{} [**23**]{} 3-11
Hinrichen H, , 2000 [*Adv. Phys.*]{} [**49:7**]{} 815-958
Bottin S, Chate H,, , 2000 [*Eur. Phys. J B*]{} [**6**]{} 143-155
Shi L, Avila M, and Hof B, , 2013 [*Phys. Rev. Lett.*]{} [**110**]{} 204502
Sipos M and Goldenfeld N, , 2011 [*Phys. Rev. E*]{} [**84**]{} 035304(R)
Allhoff K T and Eckhardt B, , 2012 [*Fluid Dyn. Res.*]{} [**44**]{} 031201
Eckert M, , 2010 [*Eur. Phys. J. H*]{} [**35**]{} 29-51
[^1]: The first such solution was discovered for Couette flow by Nagata in 1990 [@Nagata1990].
[^2]: The super exponential scaling has been related to extreme statistics theory [@Goldenfeld2010].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We investigate possibilities of solar coronal heating by acoustic waves generated not at the photosphere but in the corona, aiming at heating in the mid- to low-latitude corona where the low-speed wind is expected to come from. Acoustic waves of period $\tau\sim 100$s are triggered by chromospheric reconnection, one model of small scale magnetic reconnection events recently proposed by Sturrock. These waves having a finite amplitude eventually form shocks to shape sawtooth waves (N-waves), and directly heat the surrounding corona by dissipation of their wave energy. Outward propagation of the N-waves is treated based on the weak shock theory, so that the heating rate can be evaluated consistently with physical properties of the background coronal plasma without setting a dissipation length in an ad hoc manner. We construct coronal structures from the upper chromosphere to the outside of 1AU for various inputs of the acoustic waves having a range of energy flux of $F_{\rm w,0}=(1-20)\times 10^5$erg cm$^{-2}$ s$^{-1}$ and a period of $\tau=60-300$s. The heating by the N-wave dissipation effectively works in the inner corona and we find that the waves of $F_{\rm w,0}\ge 2\times 10^5$erg cm$^{-2}$s$^{-1}$ and $\tau \ge 60$s could maintain peak coronal temperature, $T_{\rm max} > 10^6$K. The model could also reproduce the density profile observed in the streamer region. However, due to its short dissipation length, the location of $T_{\rm max}$ is closer to the surface than the observation, and the resultant flow velocity of the solar wind is lower than the observed profile of the low-speed wind. The cooperations with other heating and acceleration sources with the larger dissipation length are inevitable to reproduce the real solar corona.'
author:
- Takeru Ken Suzuki
title: 'On the Heating of the Solar Corona and the Acceleration of the Low-Speed Solar Wind by Acoustic Waves Generated in the Corona'
---
Introduction
============
The heating of the solar is been still poorly understood, and it is one of the most challenging but interesting questions to be solved in astrophysics. The origin of energy that heats the corona is generally believed to lurk in the turbulent convective motions beneath the photosphere. A certain fraction of the kinetic energy of those turbulent motions is carried up to the corona in the shape of non-thermal energy, such as magnetic or wave energy, and thermalization of such energy in the corona results in the heating of the surrounding plasma. Granule motions of the surface convection simply expect wave generation at the photospheric level. Possibilities of coronal heating by those waves have been investigated by many researchers (Osterbrock 1961; Ulmschneider 1971; McWhirter, Thonemann, & Wilson 1975 ). To date, among various modes of waves, although an Alfvén wave has widely taken up as a convincing candidate in coronal heating and acceleration of the high-speed wind in the polar coronal holes through the ion-cyclotron damping mechanism (e.g. Cranmer, Field, & Kohl 1999; Hollweg 1999), acoustic waves have not been regarded as a major heating source of the corona because of their dissipative character. Acoustic waves with a finite amplitude inevitably steepen their wave fronts to form shocks when they travel. They are strongly damped through upward propagation in the chromosphere, consequently the only 0.01% of the initial wave energy could reach the corona (Stein & Schwartz 1972; hereafter SS). A role of acoustic waves in solar coronal heating has been buried in oblivion for a long time.
However, recent observations of dynamical structures of the solar corona (see Aschwanden, Poland, & Rabin 2001 for recent review) highlight a role of acoustic waves generated not at the photosphere but in the corona. A large number of flares and flare-like events (e.g. Tsuneta et al.1992; Tsuneta 1996) have been observed by various telescopes until today (e.g. Nishio et al.1997; see Bastian, Benz, & Gary 1998 for review) in a wide range of energy from $\sim 10^{24}$erg to $\sim 10^{32}$erg, and they are statistically argued in terms of supply of bulk energy to heat the global corona. Recent measurements of EUV frequency distribution, $\frac{dN}{dE}(E) \propto E^{-\alpha}$, of flare-like events gives a steeper power-law index ($\alpha=2.3 - 2.6$ by Krucker & Benz 1998; $\alpha=2.0 - 2.6$ by Parnell & Jupp 2000) on the lower energy side ($E=10^{24}-10^{25}$erg). This fact makes us infer that small-scale flare-like events, such as nano-flares, might be sufficient (Hudson 1991) for the required energy budget ($\gtrsim 10^{5.5}$erg cm$^{-2}$s$^{-1}$; Withbroe & Noyes 1977). Several models taking into account such small-scale events have been introduced (Sturrock 1999; hereafter S99; Sturrock, Roald, & Wolfson 2000; Roald, Sturrock, & Wolfson 2000; hereafter RSW; Tarbell, Ryutova, & Covington 1999; Sakai et al. 2000; Furusawa & Sakai 2000). Among these models, S99 proposed a model of chromospheric magnetic reconnection. In this model, a flux tube, newly formed as a result of reconnection events, oscillates vertically to excite acoustic waves in the corona. RSW further showed based on a simple kinetic model that this process possibly liberates the sufficient energy in the corona with the correct relevant parameters. A different mechanism also predicts generation of acoustic waves in the corona. For example, spicules in the open magnetic field region effectively transports energy of random motions at the photosphere into the corona and produce longitudinal (acoustic) waves there [@hol92; @ks99]. Thus, acoustic waves are supposed to be generated constantly in the corona by various dynamical processes. Such acoustic waves could heat the surrounding corona directly, unlike acoustic waves created at the photospheric level. They are expected to heat the inner corona effectively by the shock dissipation. Therefore, density at the coronal base might be large enough to creat dense wind flow observed as the streamer in the mid- to low-latitude region, which is believed to be connected to the low-speed solar wind [@hwf97]. Acoustic waves might become one of relevant processes working in the region generating the low-speed wind of which mechanisms have been poorly elucidated.
In the above models introducing production of acoustic waves in the corona, the authors concentrated on the amount of kinetic (wave) energy released in the corona. However, heating of coronal gas is accomplished by thermalization of wave energy through dissipation of such acoustic waves. Therefore, an appropriate treatment of propagation and dissipation of these waves is indispensable in understanding a problem of coronal heating. In this paper, we employ a formulation of a weak-shock theory constructed by SS, originally developed to study acoustic waves generated by convective motions at the photosphere, for propagation and dissipation of the waves produced in the corona. Then, we can explicitly determine the rate of heating by dissipation of the waves on given wave energy flux (or amplitude) and wave period without tuning other free parameters; we do not have to arbitrarily take an exponential type of heating of mechanical energy input, $F_{\rm m} \propto
\exp(\frac{-(r-R_{\odot})}{l_{\rm m}})$, for an assumed constant dissipation length, $l_{\rm m}$, which is poorly supported by fundamental physical processes, although this conventional shape of the heating law was adopted, for lack of an alternative in most previous models [@ko76; @wtb88; @sl94], to study the global coronal structure.
The paper is organized as follows: In §\[sc:mdl\], we present our model. We briefly summarize the generation of acoustic waves in the corona (§\[sc:cmr\]), and the weak-shock theory formulated by SS (§\[sc:wst\]). Basic equations describing the global coronal structure from $1 R_{\odot}$ to $\gtrsim$ 1AU are shown in §\[sc:be\], and we present our method to construct a unique coronal solution on given wave parameters. In §\[sc:rslt\] we show our model results. First, we discuss variation of coronal wind structures with respect to input parameters from a theoretical point of view (§\[sc:awdp\]). Second, we examine several characteristic properties of the resultant corona (§\[sc:rscp\]). Finally, we test whether our model is able to reproduce recent observed results for the low-speed wind in the mid- to low-latitude region (§\[sc:obs\]). In §\[sc:smds\], we summarize our results and discuss related topics.
Model {#sc:mdl}
=====
Generation of Acoustic Waves in the Corona {#sc:cmr}
------------------------------------------
As a new relevant process of coronal heating in the quiet-Sun region, S99 proposed a model of network-field magnetic reconnection at the chromospheric level. The chromosphere, specially at the minimum-temperature location, is quite a favorable site for reconnection to occur, since the magnetic resistivity is greatest, and the width of the current sheet, which is possibly scaled with the pressure scale height, is smallest there. Such reconnection events form a new closed magnetic flux tube, which is initially located far from the equilibrium state, and hence, will spring upward quickly (consult S99 for a schematic picture). The tube would oscillate about the equilibrium state normal to the solar surface with a period of $\tau = 2L/v_{\rm A}$, where $v_{\rm A}$ is Alfvén velocity and $L$ is the length of the tube. If we use the quoted values in S99 as a typical example of the tube, $L \sim 1.5\times 10^9$cm, a mean magnetic field, $\sim 100$G, and a mean density, $\sim 10^{-12}{\rm g\; cm^{-3}}$, the oscillation periods would be $\tau \sim 100$s. It could keep oscillating several tens of times before it is damped (S99). As a result, these perpendicular oscillations would excite longitudinal waves traveling in the vertical direction. Given that the configurations of the magnetic fields are perpendicular above the flux tube, such waves would propagate upwardly as acoustic waves in the rarefied atmosphere. They must contribute to the heating of the surrounding corona directly by the dissipation of the waves.
Using a simple kinetic model, RSW have estimated the energy liberated by the above mechanism as a function of a mean magnetic field strength at the photospheric level. They showed that the mean field of $\sim 10$G could generate the required heating rate ($\gtrsim 10^{5.5}$erg cm$^{-2}$s$^{-1}$) to explain the conditions in the quiet-Sun region. A sizable fraction of the released energy is supposed to be transported to the energy of the acoustic waves. In our model, we parameterize the input energy flux of the acoustic waves (not the total energy liberated by the reconnection) as $F_{\rm w,0}$ (erg cm$^{-2}$s$^{-1}$). In this paper, to study the physical processes clearly, we focus on the role of such acoustic waves in the coronal heating and construct the coronal structure on the input $F_{\rm w,0}$ and wave period $\tau$, though some of the released energy would be transformed into other MHD waves depending on complex configurations of magnetic fields [@trc99; @sky00].
We would like to remark that other mechanisms also predict production of longitudinal waves at the coronal height, although we have taken up the chromospheric reconnection model as a typical process in this paper. Generation of spicules in an open magnetic flux tube have been investigated by various authors [@hjg82; @hol92; @ks99]. They found that Alfvén waves excited by random motions at the photosphere [@ulr96] effectively transport their energy to the corona. The nonlinear effect of torsional Alfvén waves produces longitudinal waves along the vertical flux tube in the corona. A sizable fraction of the initial energy of the transverse waves at the photosphere is converted to energy of the longitudinal waves at the coronal height, and these waves could become acoustic waves propagating upwardly. They could heat the surrounding plasma in the very same way as those triggered by the reconnection events above. Thus, acoustic waves are expected to be universally generated by various mechanisms in the corona far above the photosphere, and therefore, it is quite worth studying their role in coronal heating.
Dissipation of Acoustic Waves in Corona {#sc:wst}
---------------------------------------
In this section, we describe our method of treating the outward propagation of acoustic waves, after they have been excited in the corona. We first estimate a distance acoustic waves travel before forming shocks on a plane-parallel geometry. Then, we derive an equation dealing with variation of N-wave amplitude on spherical geometry. For simplicity’s sake, in the following discussions, we neglect the effects of magnetic field on the propagation of the waves. This simplification is valid when the circumstantial magnetic configuration is perpendicular above the flux tube generating the waves.
### A Distance Acoustic Waves Travel Before Forming Shocks
Any acoustic wave having a finite amplitude inevitably changes its shape, makes the wave front steepen and eventually forms the shock front (e.g. Landau & Lifshitz 1959). The distance the acoustic waves travel before forming the shocks can be estimated on a given wave length, $\lambda$, and initial amplitude, $\delta v_0$. Consider the acoustic waves propagating in the upward direction, $z$, in isothermal atmosphere with density structure of $\rho = \rho_0 \exp(-z/H_{\rho})$. Assuming waves having initially sinusoidal velocity profiles, $\delta v=\delta v_0 \sin(2\pi Z/\lambda)$, the wave crest overtakes the preceding trough to form a shock front at $$\label{eq:shfm}
z-z_0=2H_{\rho} \ln(1+\frac{1}{4(\gamma +1)}\frac{\lambda}{H_{\rho}}
\frac{c_{\rm s}}{\delta v_0})\;$$ (SS), where $z_0$ is the position at which the waves are created, $\gamma=5/3$ is a ratio of specific heat, and $c_{\rm s}$ is a sound velocity.
In our calculations, waves of the initial amplitude, $\delta v_0/c_{\rm s} = 0.1 \sim 1$ are considered (§\[sc:awdp\]), and therefore, the second term in the logarithm in eq.(\[eq:shfm\]) will be bounded by an upper limit: $$\label{eq:shht}
\frac{1}{4(\gamma +1)}\frac{\lambda}{H_{\rho}}\frac{c_{\rm s}}{\delta v_0} \lesssim
\frac{\lambda}{H_p} = (\frac{{c_{\rm s}}^2}{\gamma g})^{-1}{c_{\rm s}
\tau}\simeq 0.25(\frac{\tau}{100{\rm s}})(\frac{c_{\rm s}}{2\times
10^7{\rm cm\:s^{-1}}})^{-1}$$ where $g$ is acceleration of gravity, and we have used $H_{\rho}\simeq H_p$ which is satisfied in the corona where temperature varies slowly within a scale of $\lambda(\sim 10^9{\rm cm})$. Then, eq.(\[eq:shfm\]) can be expanded to a first order as $$\label{eq:shfmex}
z-z_0\simeq \frac{\lambda}{2(\gamma+1)}\frac{c_{\rm s}}{\delta v_0}.$$ The factor $\frac{1}{2(\gamma+1)}\frac{c_{\rm s}}{\delta v_0}$ is an order of unity, indicating that after traveling one wave length, the waves begin to dissipate energy.
### Variation of N-wave amplitude
The initially sinusoidal acoustic waves are expected to be transported as N-waves after the formation of the shocks, provided that they are continuously generated from the lower corona. These N-waves are propagated outwardly in the rarefied atmosphere and dissipate their wave energy to heat the corona. For the purpose of giving a reasonable estimate of heating rate as a function of position, variations of amplitude, $\delta v_{\rm w}$, of the transported N-waves are treated based on the weak-shock theory, following the formulation presented by SS. In the discussions below, propagation of the wave train are considered in a flow tube with a cross-section of $A$, to keep the consistencies with a model for the global corona presented in §\[sc:be\]. $A$ is a function of $r$, a distance measured from the center of the Sun, and modeled in §\[sc:be\] with taking into account non-radial expansion of the flow tube. From now on, all the physical quantities are expressed as functions of $r$ only, unless explicitly declared.
An equation for the variation of wave amplitude normalized by ambient sound velocity, $\alpha_{\rm w}\equiv \frac{\delta v_{\rm w}}{c_{\rm s}}$, can be found from SS: $$\label{eq:wvdr}
\frac{1}{\alpha_{\rm w}}\frac{d\alpha_{\rm w}}{dr}=\frac{1}{2}(-\frac{1}{p}\frac{dp}{dr}+
\frac{1}{E_{\lambda}}\frac{dE_{\lambda}}{dr} -
\frac{1}{\lambda}\frac{d\lambda}{dr}).$$ where $p$ is gas pressure, and $E_{\lambda} = \frac{1}{3}\rho
(\delta v_{\rm w})^2 \lambda = \frac{1}{3}
\gamma p \alpha_{\rm w}^2\lambda$ is a wave energy per wave length, $\lambda$. The variation of the wave energy can be estimated from entropy generation by the weak shock (SS; §56 in Mihalas & Mihalas 1984) as $$\nabla\cdot{E_{\lambda}}=\frac{dE_{\lambda}}{dr}+\frac{E_{\lambda}}{A}
\frac{dA}{dr}=-E_{\lambda}\frac{2(\gamma +1) \alpha_{\rm w}}{\lambda}\simeq
-E_{\lambda}\frac{2(\gamma +1) \alpha_{\rm w}}{c_{\rm s}\tau} ,$$ where we have used a relation of $\lambda =c_{\rm s}(1+\frac{\gamma+1}{2}
\alpha_{\rm w})\tau \simeq c_{\rm s}\tau$ by assuming $ \alpha_{\rm w} <1$. In general, the period of waves traveling in different media remains as a constant, implying $
\frac{1}{\lambda}\frac{d\lambda}{dr}=\frac{1}{c_{\rm s}}\frac{dc_{\rm s}}{dr}.
$ Hence, eq.(\[eq:wvdr\]) is reduced to $$\label{eq:wvdrrd}
\frac{d\alpha_{\rm w}}{dr}=\frac{\alpha_{\rm w}}{2}(-\frac{1}{p}
\frac{dp}{dr}-\frac{2(\gamma +1) \alpha_{\rm w}}{c_{\rm s}\tau}
-\frac{1}{A}\frac{dA}{dr} -\frac{1}{c_{\rm s}}\frac{dc_{\rm s}}{dr}).$$ This equation determines the variation of the wave amplitude in the solar corona according to the physical properties of the background coronal plasma. The first term in the r.h.s. is positive in the density decreasing atmosphere, and the second term of the entropy generation (heating ) is negative. In the $(1-2)R_{\odot}$ region, where the dissipation is important to heat the corona, these two terms always dominate the other terms, the third term arising from the geometrical expansion and the fourth due to temperature variation. Using $\alpha_{\rm w}$ determined above as well as background physical quantities, $\rho$ and $c_{\rm s}$, wave energy flux, $F_{\rm w} ({\rm erg\; cm^{-2} s^{-1}})$, is derived as $$\label{eq:wvfx}
F_{\rm w} = \frac{1}{3}\rho c_{\rm s}^3(\alpha_{\rm w})^2 (1+
\frac{\gamma+1}{2}\alpha_{\rm w}),$$ if recalling that the wave crest moves at a speed of $c_{\rm s} (1 + \frac{\gamma+1}{2}\alpha_{\rm w})$.
Since the model described above is simple, we instead have several limitations that should be taken into account with great care. The first obvious limitation is that the wave amplitude should satisfy the assumption of weak-shock, $\alpha_{\rm w}<1$. Second, it does not take into account the effects of gravity. Third, it is constructed in a static medium, therefore we cannot apply it to cases of moving media, such as N-waves in the solar wind where the flow velocity, $v$, exceeds the ambient $c_{\rm s}$. Of the three limitations, the third one is most easily overcome, because in all the cases we calculate in this paper, at least 99% of the initial wave energy dissipates within $r<1.3 R_{\odot}$, the region where $v\ll c_{\rm s}$ is fulfilled (figs.\[fig:taudp\] & \[fig:fwdp\]). The second limitation also seems to have little effect if the wave period is small enough, since SS found that the weak-shock theory gives reasonable estimates for waves of $\tau < \frac{1}{2}\frac{2\pi}{\omega_{\rm ac}}$, where $\omega_{\rm ac}=\frac{\gamma g}{2 c_{\rm s}}$ is the acoustic cut-off frequency in a gravitationally stratified atmosphere, after comparing results of the weak-shock theory with those of fully non-linear calculations. As $\omega_{\rm ac}\simeq \frac{1}{800({\rm s})}(\frac{c_{\rm s}}
{2\times 10^7{\rm cm\: s^{-1}}})^{-1}$ in the corona, the weak-shock theory seems to be applicable to waves of $\tau<2000$s. To check whether the first limitation is overcome, we need to calculate a $\alpha_{\rm w}$ variation in the solar corona at first hand. Our results show that any waves we calculate give $\alpha_{\rm w} < 0.5$ in the entire corona (§\[sc:awdp\]). Thus, the weak-shock theory appears to be applicable to our model and to give a reasonable estimate of the heating rate as a function of $r$.
Basic Equations {#sc:be}
---------------
We here present basic equations to describe one-component coronal wind structure in a flow tube with a cross-section of $A$ under a steady state condition. Then, an equation of continuity becomes $$\label{eq:cntn}
\rho v A = {\rm const.}$$ An equation of momentum conservation is $$\label{eq:eqm}
v \frac{dv}{dr}=-\frac{G M_{\odot}}{r^2}-\frac{1}{\rho}\frac{dp}{dr}-
\frac{1}{\rho c_{\rm s}(1+\frac{\gamma+1}{2}\alpha_{\rm w})}
\nabla \cdot F_{\rm w}.$$ Pressure, $p$, is related to $\rho$ and temperature, $T$, by an equation of state for ideal gas: $$\label{eq:eqst}
p=\rho \frac{k_{\rm B}}{m_{\rm H}\mu}T ,$$ where $m_{\rm H}$ is hydrogen mass, $k_{\rm B}$ is Boltzmann constant, and $\mu$ is mean atomic weight of particles in unit of $m_{\rm H}$. The third term in eq.(\[eq:eqm\]) represents wave momentum deposition, neglecting the effects of reflection and refraction [@rv77]. Total energy equation is obtained as $$\label{eq:egcns}
\nabla \cdot[\rho v (\frac{1}{2} v^2 + \frac{\gamma}{\gamma-1}
\frac{k_{\rm B}}{m_{\rm H}\mu}T -
\frac{G M_{\odot}}{r}) + F_{\rm w} + F_{\rm c}] + q_{\rm R} = 0,$$ where we adopt the classical form of conductive flux for ionized gas, $$\label{eq:cdflx}
F_{\rm c}=-\kappa \frac{dT}{dr} = -\kappa_0 T^{5\over{2}}\frac{dT}{dr}\; ,$$ with $\kappa_0=1.0\times 10^{-6}$ in c.g.s unit [@aln73], and the radiative cooling term, $q_{\rm R}$(erg cm$^{-3}$s$^{-1}$), is derived from tabulated radiative loss function, $\Lambda({\rm erg\: cm^3 s^{-1}})$, [@LM90] for the optically thin plasma as $q_{\rm R}=n_{\rm e} n_{\rm p}\Lambda,$ where $n_{\rm e}$ & $n_{\rm p}$ are electron and H$^{+}$ density respectively. They are calculated by solving ionization of H and He (we ignore heavy elements) under LTE condition. The term, $\nabla \cdot F_{\rm w} (\le 0)$, in eq.(\[eq:egcns\]) indicates the heating by the dissipation of the waves, which is explicitly written as $$\label{eq:wvgr}
\nabla \cdot F_{\rm w}=\cases{ 0 & $r\le r_{\rm d}$ \cr \rho v [\frac{1}{3}
\alpha_{\rm w}^2 (1+\frac{\gamma+1}{2} \alpha_{\rm w})
\frac{d}{dr}(\frac{c_{\rm s}^3}{v})+(\frac{2}{3}\alpha_{\rm w}
+\frac{\gamma+1}{2}\alpha_{\rm w}^2)(\frac{c_{\rm s}^3}{v})
\frac{d\alpha_{\rm w}}{dr}]& $r> r_{\rm d}$ \cr}\; ,$$ where $r_{\rm d}$ is the position where the waves start to dissipate by shaping N-waves, and the height $h_{\rm d}\equiv r_{\rm d}-R_{\odot}$ from the photosphere corresponds to the sum of the height of the magnetic flux tube ($\sim 10^9$cm; the generation point of the waves) and the distance the acoustic waves travel before forming the shock fronts ($\sim \lambda
\sim 10^9$cm; eq.(\[eq:shfmex\])). We adopt $h_{\rm d}=2\times 10^9$cm as a standard value.
To take into account non-radial expansion of the flow tube due to configurations of magnetic field, cross-sectional area, $A$, is modeled as $$\label{eq:ftg}
A=r^2 \frac{f_{\rm max}e^{(r-r_1)/\sigma}+f_1}{e^{(r-r_1)/\sigma}+1}$$ where $$f_1=1-(f_{\rm max}-1)e^{(1-r_1)/\sigma}$$ [@ko76; @wtb88]. The cross section expands from unity to $f_{\rm max}$ most drastically between $r=r_1-\sigma$ and $r_1+\sigma$. Of three input parameters, $f_{\rm max}$ is the most important in determining solar wind structure. In this paper, we consider cases for $f_{\rm max}=5$ and $f_{\rm max}=1$ (pure radial expansion). As for the rests of the two parameters, we employ $r_1=1.25R_{\odot}$ and $\sigma=0.1R_{\odot}$, the same values adopted in a model for ’quiet corona’ in @wtb88.
Boundary Conditions and Computational Method {#sc:bc}
--------------------------------------------
Now we would like to explain the practical aspects of our method of constructing wind structure with respect to various input properties of the waves. In order to solve both heating of corona (energy transfer) and formation of solar wind (momentum transfer) consistently, our calculation is performed in a broad region from an inner boundary of the upper chromosphere where temperature, $T_{\rm ch}=10^4$K, at $r_{\rm ch}=R_{\odot}+h_{\rm ch}$, which is located at $h_{\rm ch}=2\times 10^8$cm (2000 km) above the photosphere [@aln73] to an arbitrary outer boundary at $r_{\rm out}=300 R_{\odot}$.
We set four boundary conditions to construct a unique solution for a given input wave energy flux $F_{{\rm w}, 0}$ as follows: $$\label{eq:bc1}
F_{\rm w}(r_{\rm ch})=F_{\rm w}(r_{\rm d})=F_{{\rm w}, 0},$$ $$\label{eq:bc2}
T(r_{\rm ch})=T_{\rm ch},$$ $$\label{eq:bc3}
|F_{\rm c}(r_{\rm ch})|(\simeq 0) \ll |F_{\rm c, max}|,$$ $$\label{eq:bc4}
\nabla \cdot F_{\rm c}(r_{\rm out})=0,$$ where $F_{\rm c, max}$ in eq.(\[eq:bc3\]) is a maximum value of downward conductive flux in the inner corona. The first condition denotes that wave energy flux must agree with a given value when the waves start to dissipate. The second condition is also straightforward: the temperature has to coincide with the fixed value at the inner boundary. The third condition is the requirement that the downward thermal conductive flux should become sufficiently small at the upper chromosphere ($T=10^4$K), diminishing from its enormous value at the coronal base ($T\sim 10^6$K). Practically, we continue calculations iteratively until $F_{\rm c}(r_{\rm ch})/F_{\rm c, max}<1\%$ is satisfied. The fourth condition corresponds to an ordinary requirement that no heat is conducted inward from infinity [@sl94]. Note that thanks to the third condition, coronal base density, which is poorly determined from the observations, does not have to be used as a boundary condition. As a result, the number of free parameters to be set in advance is reduced [@hm82a; @hm82b; @wtb88]. The density at the coronal base or the transition region (TR) is calculated as an output; larger input $F_{{\rm w}, 0}$ increases downward $F_{\rm c}$ in the lower corona, demanding larger density in the coronal base and TR to enhance radiative cooling to balance with the increased conductive heating.
For numerical integration of the momentum equation (eq.(\[eq:eqm\])) and the energy equation (eq.(\[eq:egcns\])), we respectively use $v$ and an isothermal sound velocity, $a$, defined as $$a^2=\frac{c_{\rm s}^2}{\gamma}=\frac{p}{\rho}=\frac{k_{\rm B}}{m_{\rm H}\mu}T.$$ To carry out the integration, the equations shown in the previous section need to be transformed into useful forms. First, an expression for velocity gradient can be written from eqs.(\[eq:cntn\]),(\[eq:eqm\]), & (\[eq:eqst\]): $$\label{eq:vlgr}
\frac{dv}{dr}=\frac{-\frac{G M_{\odot}}{r^2}+\frac{a^2}{A}\frac{dA}{dr}-
\frac{da^2}{dr}}{v-\frac{a^2}{v}}$$ Second, an expression for gradient of $a^2$ is derived from an integrated form of eq.(\[eq:egcns\]): $$\label{eq:svgr}
\frac{d a^2}{dr}=\frac{\rho v}{\kappa_0 a^5}(\frac{k_{\rm B}}{m_{\rm H}\mu})^{
\frac{7}{2}}(\frac{v^2}{2}+\frac{\gamma}{\gamma -1}a^2 -\frac{G M_{\odot}}{r}
+\frac{F_{\rm w}}{\rho v} + \int_{r_0}^{r} dr\frac{q_{\rm R}}{\rho v}
- E_{\rm tot})\; ,$$ where, $r_0$ is a certain reference point of integration that will be set later. $E_{\rm tot}$ is mathematically an integral constant, which must be conserved in the entire region of the calculation.
Only transonic solutions are allowed for the flow speed, $v$. The integration of three differential equations (\[eq:vlgr\]),(\[eq:svgr\]), and (\[eq:wvdrrd\]) is carried out simultaneously from the sonic point, $r=r_{\rm s}$ to both outward and inward directions by the fourth-order Runge-Kutta method, also deriving density from eq.(\[eq:cntn\]) at each integration. We have used variable sizes for the grid of the integration, setting the smaller mesh size in the region where physical values change rapidly. For instance, in the TR, we set a mesh as small as $10^3$ cm (0.01 km) for one grid. To start the integration, we have to set nine variables, $v_{\rm s}, (\frac{dv}{dr})_{\rm s},
a^2_{\rm s}, (\frac{da^2}{dr})_{\rm s}, \alpha_{\rm w,s},
(\frac{d \alpha_{\rm w}}{dr})_{\rm s},\rho_{\rm s}$, $E_{\rm tot}$ and $r_{\rm s} $ (subscript ’s’ denotes the sonic point). We can determine $v_{\rm s}$, $(\frac{dv}{dr})_{\rm s}$ and $r_{\rm s}$ for given $a^2_{\rm s}$ and $(\frac{da^2}{dr})_{\rm s}$ by a condition that both the numerator and the denominator of eq.(\[eq:vlgr\]) are zero at $r_{\rm s}$ [@pkr58]. $(\frac{d \alpha_{\rm w}}{dr})_{\rm s}$ is also derived from eq.(\[eq:wvdrrd\]) on a given $\alpha_{\rm w,s}$. Setting the reference point, $r_0=r_{\rm s}$, for the integration of radiative cooling function, we obtain $E_{\rm tot}$ as $$E_{\rm tot}=[\frac{1}{2} v^2+\frac{\gamma}{\gamma-1} a^2
-\frac{G M_{\odot}}{r}+\frac{F_{\rm w}}{\rho v}-(\frac{m_{\rm H}\mu}
{k_{\rm B}})^{\frac{7}{2}}\frac{1}{\rho v} \kappa_0 a^5
(\frac{da^2}{dr})]_{\rm s}\; ,$$ where all the variables are evaluated at the sonic point. Now we have four variables, $a^2_{\rm s}, (\frac{da^2}{dr})_{\rm s},
\alpha_{\rm w,s}$, and $\rho_{\rm s}$, remaining to be regulated by the four boundary conditions of eqs.(\[eq:bc1\])–(\[eq:bc4\]). Concrete procedures for finding a unique solution on a given $F_{{\rm w},0 }$ are described below.
1. [One makes an initial guess for $[a^2_{\rm s}, (\frac{da^2}{dr})_{\rm s},
\alpha_{\rm w,s}, \rho_{\rm s}]$.]{}
2. [The integration is performed in the outward direction from $r_{\rm s}$. Leaving $a^2_{\rm s}, \alpha_{\rm w,s}, \& \rho_{\rm s}$ unchanged, $(\frac{da^2}{dr})_{\rm s}$ are determined by carrying out the integration iteratively to satisfy the outer boundary condition of eq.(\[eq:bc4\]).]{}
3. [The integration is performed iteratively in the inward direction, improving $a^2_{\rm s}, \alpha_{\rm w,s}$, and $\rho_{\rm s}$ for the fixed $(\frac{da^2}{dr})_{\rm s}$ until they satisfy the three inner boundary conditions, eqs.(\[eq:bc1\])-(\[eq:bc3\]). Physically, the condition of the wave flux (eq.(\[eq:bc1\])) regulates $\alpha_{\rm w,s}$, that of temperature (eq.(\[eq:bc2\])) regulates $a^2_{\rm s}$, and that of the conductive flux (eq.(\[eq:bc3\])) regulates $\rho_{\rm s}$. These relations guide improvement of the respective initial guesses, though they are not independently approved. Unless the above inner boundary conditions are satisfied simultaneously, one returns to procedure 2, preparing a new set of $[a^2_{\rm s}, (\frac{da^2}{dr})_{\rm s}, \alpha_{\rm w,s}, \rho_{\rm s}]$. ]{}
4. [One can finally find a unique solution by iterating procedures 2 and 3.]{}
Results {#sc:rslt}
=======
Theoretical Interpretation of Resultant Coronal Structures {#sc:awdp}
----------------------------------------------------------
Figure \[fig:taudp\] demonstrates our result of coronal wind structures employing the same input $F_{\rm w,0}=7.8\times10^5$erg cm$^{-2}$s$^{-1}$, but different sets of wave periods and non-radial expansion factors, $(\tau({\rm s}),f_{\rm max})=(300,5), (60,5), (300,1)$. Figure \[fig:fwdp\] compares the wind structures adopting three different $F_{\rm w,0}=(3.2,5.9,10)\times10^5$erg cm$^{-2}$s$^{-1}$ for identical inputs of $\tau=120$s and $f_{\rm max}=5$. Distribution of temperature, flow velocity, and electron density from the inner boundary to 1AU ($=215 R_{\odot}$) are displayed from top to bottom on the left side of the both figures. On the right side, we present variation of wave amplitude, $\alpha_{\rm w}$, a dissipation length, $l_{\rm w}$, of N-waves, which is defined as $l_{\rm w}=\frac{F_{\rm w}}{\nabla \cdot F_{\rm w}} $, and heating per unit mass, $|\frac{1}{\rho v} \nabla \cdot F_{\rm w}|$, respectively. In tab.\[tab:mdr\], we tabulated several resultant properties of the corona and solar wind as well as the input parameters.
--------------- ------------------ ------------------- ---------------------- ------------------------------ ---------------------
[$p_{\rm tr}$]{} [$T_{\rm max}$]{} [$r_{T{\rm max}}$]{} [$(n_{\rm p}v)_{\rm 1AU}$]{} [$(v)_{\rm 1AU}$]{}
(7.8, 60, 5) 0.28 1.43 1.04 $3.8\times 10^6$ 222
(7.8, 300, 5) 0.24 1.50 1.08 $7.5\times 10^7$ 251
(7.8, 300, 1) 0.29 1.71 1.12 $3.8\times 10^8$ 222
(3.2, 120, 5) 0.13 1.16 1.05 $2.5\times 10^6$ 230
(5.9, 120, 5) 0.21 1.35 1.05 $1.3\times 10^7$ 240
(10, 120, 5) 0.32 1.52 1.05 $1.3\times 10^7$ 230
--------------- ------------------ ------------------- ---------------------- ------------------------------ ---------------------
: Input parameters and output wind properties of each model; $F_{\rm w,0}$ is in $10^5$erg cm$^{-2}$s$^{-1}$ and $\tau$ in second. $p_{\rm tr}$ is pressure in dyn cm$^{-2}$ at the TR where $T=10^5$K, $T_{\rm max}$ is peak coronal temperature in $10^6$K, $r_{T {\rm max}}$ is location of $T_{\rm max}$ in $R_{\odot}$, $(n_{\rm p}v)_{\rm 1AU}$ is proton flux in cm$^{-2}$s$^{-1}$ at 1AU, and $(v)_{\rm 1AU}$ is the flow velocity in km/s at 1AU.[]{data-label="tab:mdr"}
To begin with, we would like to emphasize that corona heats up to $>10^6$K in every case, because acoustic waves generated in the corona are able to heat the surrounding gas directly, unlike acoustic waves produced at the photosphere. However, N-waves are rapidly damped, so that the heating occurs only in inner region as seen in the lower right panels of figs.\[fig:taudp\] and \[fig:fwdp\]. Then, location of maximum temperature, $T_{\rm max}$, is quite close to the surface. In an outside region of $\gtrsim 1.5R_{\odot}$, heat is input only by outward thermal conduction, and the flow is accelerated mostly by thermal pressure. As a result, speed of the solar wind at 1AU is $\lesssim 300$km/s, which is slightly slower than the actual low-speed wind ($300 \sim 450$km/s).
The top right panel of fig.\[fig:fwdp\] interestingly illustrates that distributions of $\alpha_{\rm w}$ are almost identical in spite of very different inputs of $F_{\rm w,0}$. Particularly, initial N-wave amplitudes, $\alpha_{\rm w}(r_{\rm d})$, at $r_{\rm d}$, are within a range between 0.48 and 0.49. This is because $F_{\rm w,0}$ ($\sim p \alpha_{\rm w,0}^2 c_{\rm s}$; eq.(\[eq:wvfx\])) mostly owes its variation to change of ambient pressure (see §\[sc:tmxptr\]). Moreover, an upper right panel of fig.\[fig:taudp\] also indicates that initial $\alpha_{\rm w}(\simeq 0.5)$ is almost independent of $\tau$ and $f_{\rm max}$. We have found that $0.45 < \alpha_{\rm w}(r_{\rm d}) < 0.52$ within our parameter regions of $F_{\rm w,0}=(1-20)\times 10^5$erg cm$^{-2}$s$^{-1}$, $\tau=60-300$s, and $f_{\rm max}=1-5$. This proves that $\alpha_{\rm w} < 1$ is fulfilled in the entire region, which justifies the assumption of weak shock. Middle right panels of figs.\[fig:taudp\] and \[fig:fwdp\] show that the dissipation length is a drastically varying function on $r$. This implies that the assumption of a constant dissipation length usually taken in previous models for the global corona [@wtb88; @sl94] is very poor for our N-wave process.
In the following discussions, we examine dependences of the wind structures on the respective input parameters. First, we argue dependences on wave periods. As illustrated in fig.\[fig:taudp\], N-waves with smaller $\tau$ dissipate more quickly and the heating occurs in thinner region close to the surface. This simply leads to deposition of wave energy in denser region. Since radiative loss, $q_{\rm R}$(erg cm$^{-3}$s$^{-1}$), is in proportion to $\rho^2$ for optically thin plasma, a greater fraction of energy supplied in denser region goes into radiative escape. Consequently, a smaller amount of energy remains to heat the corona and accelerate the flow. The case adopting smaller $\tau$(=60s) gives lower temperature in the corona, and therefore, a smaller pressure scale height and a more rapid decrease of density, as shown in fig.\[fig:taudp\]. Lower temperature also takes the sonic point more distant from the solar surface, and then, mass flux of the solar wind becomes much smaller than that expected from the $\tau=300$s case (tab.\[tab:mdr\]).
Second, we study effects on areal expansion of the flow tube. Comparing results adopting the same $F_{\rm w,0}=7.8\times 10^5$erg cm$^{-2}$s$^{-1}$ and $\tau=300$s but different $f_{\rm max}=1$ and 5 in fig.\[fig:taudp\], one can notice significant change of density structure. The model considering the non-radial expansion gives more drastic decrease of density as a function of $r$ in spite of similar initial density at the inner boundary. Temperature in the inner corona is also lower in that model, since more fraction of the input energy is lost adiabatically due to geometrical expansion of the flow tube. On the other hand, decrease of temperature is slower and temperature in the outer region ($\gtrsim 8 R_{\odot}$) is higher. This is because the lower density reduces both radiative cooling and adiabatic loss. The higher temperature also leads to larger acceleration of the flow there, giving larger speed of the solar wind in the outer corona.
Third, we investigate dependencies on input wave energy flux. Models employing larger input $F_{\rm w,0}$ give larger density in the inner corona (fig.\[fig:fwdp\]), because radiative cooling should be raised to offset the increased heating. More accurately, larger input $F_{\rm w,0}$ results in larger downward conductive flux from the corona to the chromosphere, which needs larger radiative cooling to balance with enhanced conductive heating. Larger $F_{\rm w,0}$ also leads to higher temperature in the inner corona. However, more rapid decrease of temperature occurs because of enhanced radiative escape. A model employing $F_{\rm w,0}=10\times 10^5$erg cm$^{-2}$s$^{-1}$ yields lower temperature in a region of $r\gtrsim 1.5 R_{\odot}$ than that adopting $F_{\rm w,0}=5.9\times 10^5$erg cm$^{-2}$s$^{-1}$. Therefore, a larger input of $F_{\rm w,0}$ does not simply anticipate larger acceleration of the flow, and relations of $F_{\rm w,0}-(n_{\rm p}v)_{\rm 1AU}$ and $F_{\rm w,0}-(v)_{\rm 1AU}$ does not show simple positive correlations (tab.\[tab:mdr\], see also §\[sc:npvel\])
Some Characteristic Properties of Resultant Corona {#sc:rscp}
--------------------------------------------------
### Peak Temperature and Pressure at the TR {#sc:tmxptr}
We would like to inspect relations of peak coronal temperature, $T_{\rm max}$, and pressure, $p_{\rm tr}$, at the TR with respect to the input model parameters. In fig.\[fig:fwtmx\], we present $T_{\rm max}$ as a function of input $F_{\rm w,0}$ for different $\tau$ and $f_{\rm max}$. The figure shows that $T_{\rm max}>10^6$K is accomplished for $F_{\rm w,0}\ge 2\times 10^5$erg cm$^{-2}$s$^{-1}$, even if one chooses waves with a short period ($\tau=60$s) and a large expansion factor ($f_{\rm max}=5$). $T_{\rm max}$ is a monotonically increasing function of $F_{\rm w,0}$ on each set of $(\tau, f_{\rm max})$, and the relations can be fitted by power-law as $$\label{eq:factmx}
T_{\rm max} \propto (F_{\rm w,0})^{k}, \;\;\;\;
{\rm where} \;\;\;\;\; k=0.23-0.26.$$
Figure \[fig:fwptr\] displays relations between $F_{\rm w,0}$ and $p_{\rm tr}$. $p_{\rm tr}$ is evaluated at $T=10^5$K, whereas $p_{\rm tr}$ is almost a constant through the TR from the upper chromosphere ($T\simeq 10^4$K; inner boundary) to the coronal base ($T\simeq 5\times 10^5$K) because of its geometrically thin configuration. The figure indicates that $p_{\rm tr}$ weakly depends on $\tau$ and $f_{\rm max}$, and has a positive correlation with $F_{\rm w,0}$: $$\label{eq:fwptr}
p_{\rm tr} \propto (F_{\rm w,0})^{l}, \;\;\;\;\; {\rm where} \;\;\; l=0.75-0.79$$ A rise of $F_{\rm w,0}$ leads directly to an increase in downward conductive flux, which demands higher density (or pressure) in the lower corona and the TR to raise radiative cooling to balance with the enhanced conductive heating. Alternatively, it could be interpreted that injection of larger energy in the corona can heat more deeply to the higher density chromosphere by the downward thermal conduction.
A combination of the above two relations of eqs.(\[eq:factmx\]) and (\[eq:fwptr\]) roughly give $$T_{\rm max} \propto (p_{\rm tr})^{1/3},$$ on given $\tau$ and $f_{\rm max}$, which reminds us of the famous RTV scaling law [@rtv78], $T_{\rm max} \simeq 1400 (p_{\rm l} L_{\rm h})^{1/3}$, for closed magnetic loops, where $p_{\rm l}$(dyn cm$^{-2}$) is loop pressure and $L_{\rm h}$(cm) is loop height. Taking $r_{\rm Tmax}$ instead of $L_{\rm h}$ for our model, we can actually derive a relation between $T_{\rm max}$ and $(p_{\rm tr} r_{\rm Tmax})$ from our results as $$T_{\rm max} \simeq 3000 (p_{\rm tr} r_{\rm Tmax})^{0.30},$$ showing a form analogous to that of the original RTV law. This is because we consider the same energy balance among thermal conduction, radiative cooling, and heating with the same boundary condition that the conductive flux should become almost zero at the base (eq.(\[eq:bc3\])), although the configurations are quite different (closed loop for the RTV law and open flow tube for ours). The slight discrepancies of the prefactor and power-law index are caused by the fact that the RTV law was derived on the assumption of spatially uniform heating along the loop, while our heating function is determined by eq.(\[eq:wvgr\]), which is not uniform at all.
### Mass Flux and Coronal Energy Loss {#sc:npvel}
In fig.\[fig:fwmd\] we show anticipated proton flux, $(n_{\rm p}v)_{\rm 1AU}$, at 1AU as a function of $F_{\rm w,0}$ for different $\tau$ and $f_{\rm max}$, with observational constraints, $(n_{\rm p}v)_{\rm 1AU}=(3.8\pm 1.5)\times 10^8$cm$^{-2}$s$^{-1}$ (Shaded), compiled by @wtb88 as the empirical value for ’quiet corona’ which is supposed to corresponds to the mid- to low-latitude region generating the low-speed wind. Larger $\tau$ waves give a greater $(n_{\rm p}v)_{\rm 1AU}$ owing to the effective transport of dissipated energy to solar wind flow by avoiding radiative escape. Introduction of the non-radial expansion of the flow tube reduces $(n_{\rm p}v)_{\rm 1AU}$, because input energy per unit flow tube normalized at 1AU decreases on increasing $f_{\rm max}$, even though one inputs identical wave energy flux at the inner corona. Therefore, the larger areal expansion straightforwardly reduces mass flux of the solar wind. As to dependences on $F_{\rm w,0}$, $(n_{\rm p}v)_{\rm 1AU}$ has an upper limit for given $\tau$ and $f_{\rm max}$. (Even models of $\tau=300$s and $f_{\rm max}=1$ are supposed to have the upper limit in a $F_{\rm w,0}>2\times 10^6$erg cm$^{-2}$s$^{-1}$ area.) Figure \[fig:fwmd\] shows that only one case employing $(\tau,f_{\rm max})
=(300,1)$ can reproduce observed proton flux of the slow wind. Unfortunately more realistic cases considering the non-radial expansion of $f_{\rm max}=5$ cannot explain the observations, which implies that other mechanisms of heating (and acceleration) necessarily work cooperatively.
We would like to study in detail why $(n_{\rm p}v)_{\rm 1AU}$ has an upper limit on fixed $\tau$ and $f_{\rm max}$. Figure \[fig:fwfrc\] displays the ratios of the main three types of energy loss: downward thermal conduction from the coronal base, radiative escape in the corona, and total amount of energy converted to the flow (mass loss) at $r_{\rm out}$, respectively normalized by $F_{\rm w,0}$ for the cases adopting $\tau=120$s and $f_{\rm max}=5$. The values for downward thermal conduction in fig.\[fig:fwfrc\] are taken from maximum conductive flux $F_{\rm c, max}$ in lower corona (see fig.\[fig:egtr\]) and the values for the radiative escape are derived by the integration of the radiative cooling function from points with $F_{\rm c, max}$ to $r_{\rm out}$. The region not labeled between ’radiative escape’ and ’flow’ denotes energy carried out of $r_{\rm out}$ by thermal conduction. According to fig.\[fig:fwfrc\], the main source of coronal energy loss is downward thermal conduction within our range of $F_{\rm w,0}$ (strictly speaking, most of the conducted energy finally radiates away), whereas radiative escape comes to play a significant role for larger $F_{\rm w,0}$ since density in the corona becomes higher, being subject to eq.(\[eq:fwptr\]). Although a ratio of energy transfered to the flow is as small as $\lesssim 3\%$ throughout the range, it has a bimodal tendency. In $F_{\rm w,0}<6\times 10^5$erg cm$^{-2}$s$^{-1}$ it increases, which implies that the larger $F_{\rm w,0}$ is, the more effectively the energy is transfered to the flow. However, in $F_{\rm w,0}>6\times 10^5$erg cm$^{-2}$s$^{-1}$, it decreases because of the abrupt dominance of radiative cooling. As a result, the predicted $(n_{\rm p}
v)_{\rm 1AU}$ increases rapidly on increasing $F_{\rm w,0}$ at first and eventually decreases as seen in fig.\[fig:fwmd\].
To examine these differences in terms of energy transfer, we show variations of energy flux of four components, $${\rm wave}\; :\; f_{\rm w} = \frac{A(r)}{A(r_{\rm ch})}F_{\rm w},$$ $${\rm conduction} \; : \; f_{\rm c} = \frac{A(r)}{A(r_{\rm ch})}F_{\rm c},$$ $$\label{eq:flx3}
{\rm flow}\; :\; f_{\rm f} = \rho v \frac{A(r)}{A(r_{\rm ch})}
[(\frac{1}{2} v^2+\frac{\gamma}{\gamma-1} a^2-\frac{G M_{\odot}}{r})-
(\frac{1}{2} v^2+\frac{\gamma}{\gamma-1} a^2
-\frac{G M_{\odot}}{r})_{r_{\rm ch}}],$$ $$\label{eq:flx4}
{\rm radiation}\; :\; f_{\rm R} = \rho v \frac{A(r)}{A(r_{\rm ch})}
\int_{r_{\rm ch}}^{r}dr \frac{q_{\rm R}}{\rho v},$$ per flow tube with a cross section of $A=$1cm$^2$ at the inner boundary. The ’flow’ term of eq.(\[eq:flx3\]) contains three ingredients, kinetic energy of the solar wind, enthalpy, and gravitational energy. Note that, except the wave energy flux, the zero point of the energy is taken here at the inner boundary, to clarify what fraction of the input energy flux is transfered to the other components. In a left panel of fig.\[fig:egtr\], the status of energy transfer in the inner corona is displayed in a linear scale for both X and Y-axis, and in a right panel, that in the broader region is shown in $\log$ scale. The ratio of the downward thermal conduction, $f_{\rm c}/F_{\rm w,0}$, is smaller for larger $F_{\rm w,0}$, though the absolute value of $F_{\rm c}$ is increasing along with $F_{\rm w,0}$. At the TR, most of the heat flux by the downward conduction finally escapes as radiation, except for a tiny fraction transfered to the enthalpy to be used to heat the TR. This leads to a smaller ratio of the radiative loss, $f_{\rm R}/F_{\rm w,0}$, for larger $F_{\rm w,0}$ in the inner region of $<1.2 R_{\odot}$. However, the contribution from the radiation continues in a much more distant region in the model employing largest $F_{\rm w,0}$, and $f_{\rm R}/F_{\rm w,0}$ finally outdoes the other two cases. Consequently, the ratio of energy transfered to the solar wind becomes smaller than the model adopting smaller $F_{\rm w,0}(=5.9\times 10^5$erg cm$^{-2}$s$^{-1})$ as seen in the right panel. It can be concluded that if $F_{\rm w,0}$ is larger than a certain threshold, an increase of the input $F_{\rm w,0}$ does not lead to effective heating of the corona to accelerate the solar wind but results in deposition of the wave energy in the high density region to be wasted as radiative escape.
### Wave Amplitude in the Inner Corona
Ultraviolet and X-ray emission lines of the corona show non-thermal broadenings [@hrsh90; @edpw98], which are inferred to originate from wave motions. We investigate whether the obtained wave amplitudes are consistent with these observations. In tab.\[tab:dvwm\], we show our results of the wave rms velocities in the inner corona, which are calculated as $$<\delta_{\rm w}>=\frac{1}{\sqrt{3}}\delta_{\rm w}$$ for the N-shaped wave.
--------------- ------------------- -----------
$T=5\times 10^5$K $T=10^6$K
(7.8, 60, 5) 32 42
(7.8, 300, 5) 34 44
(7.8, 300, 1) 32 41
(3.2, 120, 5) 31 41
(5.9, 120, 5) 32 42
(10, 120, 5) 34 44
--------------- ------------------- -----------
: $<\delta_{\rm w}>$(km/s) in the inner corona[]{data-label="tab:dvwm"}
As the waves propagate upwardly from a location of $T=5\times 10^5$K to that of $T=10^6$K, the wave amplitude increases according as the ambient density decreases. Our models give the very similar results for different $F_{\rm w,0}$ (also $\tau$ and $f_{\rm max}$), since change of $F_{\rm w,0}(\sim \rho
<\delta_{\rm w}>^2 c_{\rm s})$ mostly reflects variation of the density (fig.\[fig:fwdp\] and §\[sc:awdp\]). The tabulated results are marginally consistent with non-thermal velocities of $20 \sim 40$km/s obtained from observations of inner coronal lines in the solar disks and limbs [@edpw98], if we assume the observed non-thermal components totally consist of the waves of the acoustic mode. However, this assumption may be too extreme, because other modes of the waves actually exist in the real solar corona.
Comparison of Coronal Wind Structure with Observed Low-Speed Wind {#sc:obs}
-----------------------------------------------------------------
In this section, we study a feasibility of the process of acoustic waves in the coronal heating by comparing our results with recent observations. Aiming at coronal heating and wind acceleration in a region where the low-speed wind is formed, we take observational data of the coronal streamer in the mid- to low-latitude region.
### Density Distribution
In fig.\[fig:obsne\], we display our results of electron density with observation of the streamer [@pbp00; @hvh01]. Our models adopt an identical $F_{\rm w,0} = 7.8\times 10^5$erg cm$^{-2}$s$^{-1}$ and three sets of $(\tau({\rm s}), f_{\rm max})=(300, 5), (60,5), (300, 1)$ (same as fig.\[fig:taudp\]). As for the observation, we show results derived from a line ratio of Si IX by CDS/SOHO for a region of $1.02 - 1.19 R_{\odot}$, a ratio of radiative and collisional intensities of O VI line by UVCS/SOHO for a region of $1.58 - 1.6R_{\odot}$ [@pbp00], and total brightness obtained from LASCO/SOHO for a region $1.5 R_{\odot} \lesssim r \lesssim 6 R_{\odot}$ [@hvh01]. With respect to the CDS and UVCS data, both results for equatorial and mid-latitude streamer regions are displayed.
Our results, adopting the same $F_{\rm w,0}$, show almost identical density in very inner part, being independent of $\tau$ and $f_{\rm max}$ (§\[sc:tmxptr\] and fig.\[fig:fwptr\]). The figure exhibits that the adopted value, $F_{\rm w,0}=7.8 \times 10^5$erg cm$^{-2}$s$^{-1}$, reproduces the CDS observation of the mid-latitude streamer well. To fit to the data of the equatorial streamer, slightly smaller input of $F_{\rm w,0}$ is favored. In our model, $\tau$ and $f_{\rm max}$ control slope of density distribution. $\tau$ influences temperature in an intermediate region of $2R_{\odot} \lesssim r \lesssim 10R_{\odot}$ (fig.\[fig:taudp\]), hence, regulates the density scale height (or decreasing slope of density) in that region. The data based on LASCO/SOHO indicates that models considering larger $\tau$(=300s) are more likely. $f_{\rm max}$ determines density decrease in a region of $1 \sim 2 R_{\odot}$. The data of CDS ($\lesssim 1.2
R_{\odot}$) and UVCS ($\sim 1.6 R_{\odot}$) exhibits the drastic decrease of density, which indicates that non-radial expansion is desired. Adjustment of the other parameters of the flow tube geometry ($r_1$ and $\sigma$; eq.(\[eq:ftg\])) would give the still better fit. Although we do not further search the best parameter set to fit to the observation, the figure indicates that our model could reproduce the observed density profile by the choice of the appropriate parameters ($F_{\rm w,0}\simeq (5-8)\times 10^5$erg cm$^{-2}$s$^{-1}$, $\tau\approx 300$s, and $f_{\rm max}\approx 5$).
### Temperature Distribution
Figure \[fig:obste\] compares our results of temperature distribution with observation in the streamer region. Our models employ the same parameter sets as in fig.\[fig:obsne\], and the data were electron temperature obtained from observations of a line ratio of Fe XIII/Fe X by the CDS and UVCS/SOHO [@pbp00]. Although the observed data are electron temperature, they are supposed to represent the plasma temperature because electron-ion equilibrium is attained in dense streamer region of $r\lesssim 2R_{\odot}$ [@rskn98]. Therefore, it is reasonable to compare them to the results of our model considering one-fluid coronal plasma. A case employing $(F_{\rm w,0}, \tau, f_{\rm max})=(7.8\times 10^5, 300, 5)$ gives reasonable peak temperature of $\simeq 1.5 \times 10^6$K. However, none of our models can reproduce the observed location of $T_{\rm max}$. While the location is observationally inferred to be between $1.2 R_{\odot}$ and $1.6 R_{\odot}$, all of our models give $r_{T{\rm max}}< 1.2 R_{\odot}$. This is because the dissipation length of the N-waves is essentially short, even though one considers long period-waves that are generated in the corona. We can summarize that acoustic waves excited in the corona could certainly heat the surrounding plasma to $T>10^6$K, however, they cannot maintain the high temperature till the sufficiently distant region by themselves. Therefore, the cooperations with other heating sources with larger dissipation length are necessary to explain the observed solar corona.
### Velocity Distribution
In fig.\[fig:obsvl\], we show the results of velocity distribution of the solar wind, with observational results in the low-latitude streamer (shaded). The observational data are from @she97, who determined velocity profile between 2 and 30 $R_{\odot}$ from measurements of about 65 moving objects in the streamer belt. They used two different technic in deriving the results, whereas the shaded area displayed in fig.\[fig:obsvl\] are based on the straight-line fit method (an upper panel of fig.6 in @she97; The shaded region is traced from that figure). Figure \[fig:obsvl\] indicates that the resultant velocity of our models is lower than the observed data in the whole region. The observation exhibits rapid acceleration in $3 \sim 5R_{\odot}$, while our results show gradual acceleration by thermal gas pressure, since the N-waves are damped in $<1.5 R_{\odot}$ and wave pressure cannot contribute to the acceleration of the wind flow, as shown in figs.\[fig:taudp\] and \[fig:fwdp\]. Other mechanisms are also required in acceleration of the low-speed wind.
Summary and Discussions {#sc:smds}
=======================
We have investigated the process of acoustic waves generated in the corona as a heating source especially in the mid- to low-latitude corona where the low-speed winds come from. We have found that the acoustic waves with $\tau\ge 60$s and $F_{\rm w,0}
\ge 2\times 10^5$erg cm$^{-2}$s$^{-1}$ could certainly heat up the ambient plasma to $T\ge 10^6$K by the dissipation of the N-waves even in a flow tube with the areal expansion of $f_{\rm max}=5$, by balancing with the losses of the radiative cooling, the downward thermal conduction, and the adiabatic loss by the solar wind. Due to its dissipative character, the dissipation of N-waves effectively works in the inner corona and reproduces the density profile observed in the streamer region. However, it cannot contribute to the heating of the outer corona, since most of the wave energy is damped within a region of few-tenth of the solar radius even if one considers the waves with a long period of $\tau=300$s. As a result, it is impossible to explain the observed temperature profile and flow velocity of the low-speed wind only by that process. Therefore, other mechanisms with the larger dissipation length should play a role cooperatively in the coronal heating and the acceleration of the low-speed wind.
Small-scale reconnection events and spicules also predict generation of fast shock waves [@lw00; @lee01; @hol82], though we have concentrated on the role of the N-waves (slow shocks along the magnetic field line) excited by those events in this paper. The dissipation length of the fast shocks must be larger than N-waves with the identical period, since their phase speed is $\gtrsim v_{\rm A}$, which is much larger than that of the N-waves ($\sim c_{\rm s}$) for low-$\beta$ coronal plasma. Hence, cooperation with the fast shocks would give a more distant location of $T_{\rm max}$ to match the observed temperature profile (fig.\[fig:obste\]), and would let the acceleration of the wind flow continue in the outer corona, as suggested by observation of the streamer belt (fig.\[fig:obsvl\]).
Finally, we had better remark an issue on anisotropic and selective heating of ions. According to recent observations [@ssp02], O VI ions in the streamer have high perpendicular kinetic temperature, though this is not so extreme as that observed in the high-speed wind. Our results show that the deposition of energy and momentum from the N-waves to the ambient gas is completed in the region of $r\lesssim 1.5 R_{\odot}$, where the particles are thermally well coupled. As a result, whatever anisotropies obtained by the N-wave heating would be wiped out and all the ions would have isotropic kinetic temperature. The process of the N-waves cannot explain the observed anisotropies, which also indicates that the other mechanisms causing the anisotropic heating have to operate simultaneously.
We thank K. Shibata, K. Ohki, K. Omukai, K. Tomisaka, H. Saio, S. Nitta, S. Inutsuka, T. Kudoh, Y. Yoshii, T. Kajino, and A. Tohsaki as well as members of DTAP in NAOJ for many valuable and critical comments and Y. Mclean for improvement of presentation in this paper. The author is supported by the JSPS Research Fellowship for Young Scientists, grant 5936.
Allen, C. W. 1973, Astrophysical Quantities(London:Athlone)
Aschwanden, M. J., Poland, A. I., & Rabin, D. M. 2001, , 39, 175
Bastian, T. S., Benz, A. O., & Gary, D. E. 1998, , 36, 131
Cranmer, S. R., Field, G. B., & Kohl, J. L. 1999, , 518, 937
Erdélyi, R., Doyle, J. G., Perez, M. E., & Wilhelm, K. 1998, , 337, 287
Furusawa, K. & Sakai, J. I. 2000, , 540, 1156
Habbal, S. R., Woo, R., Fineschi, S., O’neal, R., Kohl, J., Noci, G., & Korendyke, C. 1997, , 489, L103
Hammer, R. 1982a, , 259, 767
Hammer, R. 1982b, , 259, 779
Hassler, D. M., Rottman, G. J., Shoub, E. C., & Holzer, T. E. 1990, , 348, L77
Hayes, A. P., Vourlidas, A., & Howard, R. A. 2001, , 548, 1081
Hollweg, J. V. 1982, , 254, 806
Hollweg, J. V. 1992, , 389, 731
Hollweg, J. V. 1999, , 104, 24781
Hollweg, J. V., Jackson, S., & Galloway, D. 1982, Sol.Phys., 75, 35
Hudson, H. S. 1991, Sol.Phys., 133, 357
Kopp, R. A. & Orall, F. Q. 1976, , 53,363
Krucker, S. & Benz, A. O. 1998, , 501, L213
Kudoh, T. & Shibata, K. 1999, , 514, 493
Landau, L. D. & Lifshitz, E. M. 1959, ’Fluid Mechanics’
Landini, M. & Monsignori-Fossi, B. C. 1990, , 82, 229
Lee, L. C. & Wu, B. H. 2000, , 535, 1014
Lee, L. C. 2001, , 95, 95
McWhirter, R. W. P., Thonemann, P. C., & Wilson, R. 1975, , 40, 63
Mihalas, D. & Mihalas, B. W. 1984, ’Foundation of Radiation Hydrodynamics’, Oxford Unuversity Press
Nishio, M., Yaji, K., Kosugi, T., Nakajima, H., & Sakurai, T. 1997, , 489, 976
Osterbrock, D. E. 1961, , 134, 347
Parenti, S., Bromage, B. J. I., Poletto, G., Noci, G., Raymond, J. C., & Bromage, G. E. 2000, , 363, 800
Parker, E. N. 1958, , 128, 664
Parnell, C. E. & Jupp, P. E. 2000, , 529, 554
Raymond, J. C., Suleiman, R., Kohl, J. L., & Noci, G. 1998, , 85, 283
Roald, C. B., Sturrock, P. A., & Wolfson, R. 2000, , 538, 960 (RSW)
Rosner, R. & Vaiana, G. S. 1977, , 216, 141
Rosner, R., Tucker, W. H., & Vaiana, G. S. 1978, , 220, 643
Sakai, J.I., Kawata, T., Yoshida, K., Furusawa, K., & Cramer, N. F. 2000, , 537, 1063
Sandb[æ]{}k, Ø. & Leer, E. 1994, , 423, 500
Sheeley, N. R. Jr. et al. 1997, , 484, 472
Stein, R. F. & Schwartz, R. A. 1972, , 177, 807 (SS)
Strachan, L, Suleiman, R., Panasyuk, A. V., Biesecker, D. A., & Kohl, J. L. 2002, , 571, 1008
Sturrock, P. A. 1999, , 521, 451 (S99)
Sturrock, P. A., Roald, C. B., & Wolfson, R. 1999, , 524, L75
Tarbell, T., Ryutova, M., & Covington, J. 1999, , 514, L47
Tsuneta, S. et al. 1992, , 44, L63
Tsuneta, S. 1996, , 456, 840
Ulmschneider, P. 1971, , 12, 297
Ulrich, R. K. 1996, , 465, 436
Withbroe, G. L. & Noyes, R. W. 1977, , 15, 363
Withbroe, G. L. 1988, , 325, 442
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Chevalley’s theorem on the images of morphisms of schemes and the principle of quantifier elimination for the theory of algebraically closed fields are widely understood to be two perspectives on the same theorem. In this paper, we demonstrate that both results can easily be proven simultaneously, using a mixture of geometric and logical techniques. In doing so, we give logical meaning to geometric points of schemes and to finitely presented morphisms thereof, in a manner reminiscent of Spencer Breiner’s logical schemes.'
address: 'Merton College, Oxford OX1 4JD, United Kingdom'
author:
- 'L. Alexander Betts'
title: 'The first-order theory of geometric points of schemes: Chevalley’s theorem and quantifier elimination'
---
Introduction {#introduction .unnumbered}
============
In his PhD thesis [@breiner], Breiner develops the notion of a logical scheme. To a good approximation, this is a geometric object which is locally a first-order theory, and its points correspond to models of these theories.
We can realise a similar idea concretely in the case of (classical) schemes. To each ring $R$ we can associate a first-order theory ${\mathsf{ACF}}_R$, so that we have a canonical bijection$$\{\text{Geometric points of $\operatorname{Spec}(R)$}\}\longleftrightarrow\{\text{Models of ${\mathsf{ACF}}_R$}\}$$This correspondence induces a bijection between the $0$th Stone space of ${\mathsf{ACF}}_R$ and the underlying topological space of $\operatorname{Spec}(R)$, which is a homeomorphism when $|\operatorname{Spec}(R)|$ is endowed with the constructible topology.
Using this correspondence we can easily understand the images of (locally) finitely presented morphisms of schemes: they are just (locally) described by a sentence in the language of ${\mathsf{ACF}}_R$. Pursuing this idea naturally leads to a straightforward proof of Chevalley’s theorem (on the scheme-theoretic side) and quantifier elimination for the theories ${\mathsf{ACF}}_R$ (on the logical side), and the aim of this paper is to explain these basic applications of the theory.
Although our treatment involves heavy use of the underlying topological space of a scheme, for the purely algebraic applications of this theory it is easy to dispense with this and just work with the functor-of-points instead.
Geometric points of schemes as models of theories {#geometric-points-of-schemes-as-models-of-theories .unnumbered}
=================================================
The correspondence for geometric points is very simple. We form the language of ${\mathsf{ACF}}_R$ by taking the language of rings and adjoining constants $c_a$ for each element $a\in R$ (we will frequently write $c_a$ just as $a$, for ease of notation). The theory ${\mathsf{ACF}}_R$ is then formed by taking the theory of algebraically closed fields (in the language of rings) and adjoining axioms $c_1=1$ and, for each $a,b\in R$, $c_{a+b}=c_a+c_b$ and $c_{ab}=c_ac_b$.
Models of ${\mathsf{ACF}}_R$ are then algebraically closed fields $k$, together with named elements $c_a^k$ for each $a\in R$, so that the function $R\rightarrow k$ sending $a$ to $c_a^k$ is a homomorphism of rings. In other words, models of ${\mathsf{ACF}}_R$ can be naturally thought of as ring homomorphisms $x:R\rightarrow k$ from $R$ into an algebraically closed field (or more accurately, such a homomorphism makes $k$ into a model of ${\mathsf{ACF}}_R$ in a natural way – we will be happy to conflate these concepts).
Since $\operatorname{Spec}$ is a contravariant embedding, such homomorphisms are in bijection with scheme morphisms $\operatorname{Spec}(k)\rightarrow\operatorname{Spec}(R)$, which are precisely geometric points. Thus the correspondence between geometric points of $\operatorname{Spec}(R)$ and models of ${\mathsf{ACF}}_R$ is just transposition across the $\operatorname{Spec}$/global sections adjunction.
The image formula {#the-image-formula .unnumbered}
-----------------
With this correspondence between geometric points and models of ${\mathsf{ACF}}_R$ in mind, we make the crucial observation that the image of a finitely presented morphism of affine schemes is described by a particular first-order formula. It follows that the image will be constructible iff this formula can be rewritten in an equivalent, quantifier-free form. Of course, elimination of quantifiers for the theories ${\mathsf{ACF}}_R$ makes this immediate, but since we are aiming to prove this, we use more elementary methods.
\[prop:image\_formula\] Let $S\cong R[{\mathbf}t]/({\mathbf}f)$ be a finitely presented $R$-algebra, and $\operatorname{Spec}(S)\rightarrow\operatorname{Spec}(R)$ the structure map. Then a geometric point $x:\operatorname{Spec}(k)\rightarrow\operatorname{Spec}(R)$ lifts to $\operatorname{Spec}(S)$ iff, viewed as a model of ${\mathsf{ACF}}_R$, it satisfies the sentence$$(\exists{\mathbf}y)\left(\bigwedge_j(f_j({\mathbf}y)=0)\right)$$
We will refer to this formula as the *image formula* associated to this presentation of $S$ as an $R$-algebra.
The lifts of $x$ to $\operatorname{Spec}(S)$ correspond with homomorphisms $R[{\mathbf}t]/({\mathbf}f)\rightarrow k$ of $R$ algebras, and hence with tuples ${\mathbf}y\in k^n$ such that all $f_j({\mathbf}y)=0$. Thus $x$ lifts to $\operatorname{Spec}(S)$ iff such a tuple exists, which says precisely that $x:R\rightarrow k$ satisfies the image formula.
The preceding proposition gives logical meaning to the collection of geometric points in $\operatorname{Spec}(R)$ lifting to $\operatorname{Spec}(S)$, but we should check that this agrees with the usual topological notion of the image of $\operatorname{Spec}(S)\rightarrow\operatorname{Spec}(R)$. This is provided (in a strong sense) by the following easy proposition, which is also the sole algebraic input into the theory.
\[prop:geom\_image=top\_image\] Let $Y\rightarrow X$ be a locally finitely presented morphism of schemes, and $x:\operatorname{Spec}(k)\rightarrow X$ a geometric point with underlying topological point $P\in|X|$. Then $x$ lifts to $Y$ iff $P$ lies in the image of $Y\rightarrow X$. In particular, whether $x$ lifts to $Y$ depends only on its image in $X$.
Let $\kappa_P$ be the residue field at $P$, so that $x$ factors (uniquely) through $\operatorname{Spec}(\kappa_P)\rightarrow X$. We know that $P$ lies in the image of $Y\rightarrow X$ iff the fibre $\operatorname{Spec}(\kappa_P)\times_XY$ is not the empty scheme. Since $\kappa_P\rightarrow k$ is faithfully flat, this occurs iff $\operatorname{Spec}(k)\times_XY$ is not the empty scheme. Yet this is locally finitely presented over $\operatorname{Spec}(k)$, so the Nullstellensatz shows that $\operatorname{Spec}(k)\times_XY$ is non-empty iff its structure morphism to $\operatorname{Spec}(k)$ has a section. By the universal property of fibre products, such sections are in bijective correspondence with lifts of $x$ to $Y$, completing the proof.
If $R$ is any ring and $b\in R$ an element, then the (finitely presented) morphism $\operatorname{Spec}(R_b)\rightarrow\operatorname{Spec}(R)$ is just the inclusion of a standard Zariski-open affine. Under ${\mathsf{ACF}}_R$, its image formula is equivalent to the formula $(b\neq0)$, so that the image formula describing $D(b)$ is just the statement that $b$ does not vanish.
Chevalley’s theorem {#chevalleys-theorem .unnumbered}
===================
We are now in a position to prove Chevalley’s theorem on the images of finitely presented scheme morphisms. The argument is essentially a two-pass application of the compactness theorem from first-order logic, but the niceties can be cut out by recognising the argument as exactly the same one that proves the transfer lemma from model theory. For clarity, we recall this here before applying it to prove the affine case of Chevalley’s theorem.
\[lem:transfer\_lemma\] Let $T_0\subseteq{\mathsf{FO}}_0$ be some base theory (over some signature) and $F\subseteq{\mathsf{FO}}_0$ a fragment closed under $\vee$ and $\wedge$. Suppose that a sentence $\Phi$ has the following property: whenever $A$ and $B$ are two $T_0$-models such that $A\models\Phi$ and $F\cap{\mathrm{Th}}(A)\subseteq{\mathrm{Th}}(B)$, then $B\models\Phi$ also. Then $\Phi$ is equivalent under $T_0$ to a sentence in $F$.
\[thm:affine\_Chevalley\] Let $S$ be a finitely presented $R$-algebra. Then the image of the map $|\operatorname{Spec}(S)|\rightarrow|\operatorname{Spec}(R)|$ is constructible. Equivalently, the image formula $\Phi$ for $\operatorname{Spec}(S)\rightarrow\operatorname{Spec}(R)$ is equivalent under ${\mathsf{ACF}}_R$ to a quantifier-free sentence.
Translating proposition \[prop:geom\_image=top\_image\] into the language of model theory, we see that whether a model $x:R\rightarrow k$ of ${\mathsf{ACF}}_R$ satisfies the image formula $\Phi$ depends only on $\ker(x)\unlhd R$. In particular, whether a model satisfies $\Phi$ depends only on which quantifier-free sentences it satisfies. Thus $\Phi$ and the quantifier-free fragment of ${\mathsf{FO}}_0$ satisfy the conditions of the transfer lemma, where we take $T_0={\mathsf{ACF}}_R$. As a consequence, we see that $\Phi$ is equivalent under ${\mathsf{ACF}}_R$ to a quantifier-free sentence, as desired.
To pass back to constructibility is now easy. $\Phi$ is equivalent to a sentence of the form $\bigvee_i\left((b_i\neq0)\wedge\bigwedge_j(a_{ij}=0)\right)$. Translating into the language of scheme theory, a geometric point of $\operatorname{Spec}(R)$ lies in the image of $\operatorname{Spec}(S)\rightarrow\operatorname{Spec}(R)$ iff its underlying topological point lies in $\bigcup_i\left(D(b_i)\cap\bigcap_j V(a_{ij})\right)$, which is a constructible set.
Exactly the same method can be used to prove that finitely presented flat morphisms $f:\operatorname{Spec}(S)\rightarrow\operatorname{Spec}(R)$ of affine schemes have Zariski-open image. A standard result [@matsumura Theorem 5.D] in commutative algebra tells us that the topological image of $f$ is closed under generisation. In other words, given geometric points $x:R\rightarrow k$ and $y:R\rightarrow l$ such that $\ker(y)\leq\ker(x)$, if $x$ lies in the image of $f$ then so does $y$. Thus we may apply the transfer lemma immediately, with respect to the fragment of ${\mathsf{FO}}_0$ consisting of the quantifier-free *negative* sentences, to deduce that the image formula $\Phi$ of $f$ is equivalent under ${\mathsf{ACF}}_R$ to a quantifier-free negative sentence. In other words, the image is Zariski-open (indeed is a finite union of basic opens).
Let $f:Y\rightarrow X$ be a locally finitely presented, quasicompact morphism of schemes. Then the image of any locally constructible set in $Y$ is locally constructible in $X$.
We can immediately reduce to the case when $Y=\operatorname{Spec}(S)$ and $X=\operatorname{Spec}(R)$ are affine, so that the morphism is given by a finitely presented ring homomorphism $R\rightarrow S$. Any constructible set $Z$ in $\operatorname{Spec}(S)$ is the image of a finitely presented morphism $\operatorname{Spec}(T)\rightarrow\operatorname{Spec}(S)$ and thus $f(Z)$ is the image of the composite $\operatorname{Spec}(T)\rightarrow\operatorname{Spec}(R)$. It is hence constructible by theorem \[thm:affine\_Chevalley\].
\[cor:QE\_for\_ACF\_R\] The theory ${\mathsf{ACF}}_R$ has quantifier elimination. Every first-order formula in the language of ${\mathsf{ACF}}_R$ is equivalent to a quantifier-free formula.
Proceeding by structural induction, we just need to show that existential formulae are equivalent to quantifier-free formulae. It is easy to see that every quantifier-free formula $\Psi$ in the language of ${\mathsf{ACF}}_R$ is equivalent to one in the form$$\bigvee_i\left((b_i\neq0)\wedge\bigwedge_j(a_{ij}=0)\right)$$for some $R$-polynomials $b_i,a_{ij}$ in the free variables of $\Psi$.
If the formula $(\exists{\mathbf}x)\Psi$ has no free variables, it is equivalent to the image formula of some finitely presented morphism of affine schemes (associated to the $R$-algebra $\prod_iR[{\mathbf}t]_{b_i}/(a_{ij})_j$), so that we are done by theorem \[thm:affine\_Chevalley\]. In general, if the formula $(\exists{\mathbf}x)\Psi$ has free variables ${\mathbf}y$, we can view it as a sentence in the language of ${\mathsf{ACF}}_{R[{\mathbf}y]}$, whose models can be thought of as models of ${\mathsf{ACF}}_R$ with named values for the variables ${\mathbf}y$. By the previous case, under ${\mathsf{ACF}}_{R[{\mathbf}y]}$ it is equivalent to a quantifier-free sentence, i.e. under ${\mathsf{ACF}}_R$ it is equivalent to a quantifier-free formula with free variables ${\mathbf}y$. This concludes the proof.
Topological points {#topological-points .unnumbered}
------------------
To complete the picture, we will now show how our correspondence between models of ${\mathsf{ACF}}_R$ and geometric points descends to a bijection between elementary equivalence classes of models and topological points. Let $S_0({\mathsf{ACF}}_R)$ denote the Stone space of ${\mathsf{ACF}}_R$, with its usual topology. There is a canonical map $p:S_0({\mathsf{ACF}}_R)\rightarrow|\operatorname{Spec}(R)|$, taking a model $x:R\rightarrow k$ of ${\mathsf{ACF}}_R$ to the prime $\ker(x)$ of $R$. This is easily seen to be well-defined and continuous with respect to the constructible topology (e.g. the preimage of $D(f)$ is just $S_0({\mathsf{ACF}}_R\cup\{f\neq0\})$, which is clopen in the topology on $S_0({\mathsf{ACF}}_R)$).
We demonstrate that $p$ is a homeomorphism. That it is surjective is clear, since every topological point of $\operatorname{Spec}(R)$ arises from a geometric point. To prove injectivity, just note that if $x$ and $y$ are models of ${\mathsf{ACF}}_R$ such that $p(x)=p(y)$, then in particular $x$ and $y$ satisfy the same quantifier-free sentences in the language of ${\mathsf{ACF}}_R$. Quantifier elimination (corollary \[cor:QE\_for\_ACF\_R\]) tells us that they are elementarily equivalent, and so represent the same point of $S_0({\mathsf{ACF}}_R)$. Thus $p$ is a continuous bijection from a compact space to a Hausdorff space, and hence is a homeomorphism.
[9]{} Spencer Breiner. *Scheme representation for first-order logic.* PhD thesis, Carnegie Mellon University, 2014. arxiv: 1402.2600 Hideyuki Matsumura. *Commutative algebra.* The Benjamin/Cummings Publishing Company, second edition, 1980.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Recently, a new domino signal encryption algorithm (DSEA) was proposed for digital signal transmission, especially for digital images and videos. This paper analyzes the security of DSEA, and points out the following weaknesses: 1) its security against the brute-force attack was overestimated; 2) it is not sufficiently secure against ciphertext-only attacks, and only one ciphertext is enough to get some information about the plaintext and to break the value of a sub-key; 3) it is insecure against known/chosen-plaintext attacks, in the sense that the secret key can be recovered from a number of continuous bytes of only one known/chosen plaintext and the corresponding ciphertext. Experimental results are given to show the performance of the proposed attacks, and some countermeasures are discussed to improve DSEA.'
address:
- 'Department of Mathematics, Zhejiang University, Hangzhou 310027, China'
- 'Department of Electronic Engineering, City University of Hong Kong, 83 Tat Chee Avenue, Kowloon Tong, Hong Kong, China'
- 'Department of Electrical Engineering, Chung Cheng Institute of Technology, National Defense University, Taiwan, China'
- 'College of Computer Science, Zhejiang University, Hangzhou 310027, China'
author:
- Chengqing Li
- Shujun Li
- 'Der-Chyuan Lou'
- Dan Zhang
bibliography:
- 'DSEA.bib'
title: 'On the security of the Yen-Guo’s domino signal encryption algorithm (DSEA)'
---
[^1]
,
and
DSEA ,dominos ,cryptanalysis ,encryption ,ciphertext-only attack ,known-plaintext attack ,chosen-plaintext attack
Introduction
============
In today’s networked world, the security issues become more and more important, so various encryption algorithms have been developed to fulfill the needs of different applications [@Schneier:AppliedCryptography96]. In recent years, Yen and Guo et al. proposed a series of chaos-based[^2] signal/image encryption schemes [@Li:ChaosImageVideoEncryption:Handbook2004 Sec. 4.4.3], some of which have been broken according to the works reported in [@ShujunLi:AttackCKBA:ISCAS2002; @ShujunLi:AttackBRIE:ICIP2002; @Li:AttackingMES2004; @Li:AttackingCNN2004; @Li:AttackingRCES2004; @Li:AttackTDCEA2004]. The present paper gives the cryptanalysis results on a new Yen-Guo encryption scheme called DSEA [@Yen-Guo:DSEA:JCIEE2003], which has not been cryptanalyzed before.
DSEA encrypts the plaintext block by block, which is composed of multiple bytes. The first byte of each block is masked by part of the secret key, and other bytes are masked by the previous cipher-byte, under the control of a chaotic pseudo-random bit sequence (PRBS). That is to say, DSEA works like the dominos. This paper analyzes the security of DSEA, and points out the following defects: 1) its security against the brute-force attack was overestimated; 2) it is not sufficiently secure against ciphertext-only attacks, and only one ciphertext is enough to get some information about the plaintext and to break the value of a sub-key; 3) it is insecure against known/chosen-plaintext attacks, in the sense that the secret key can be recovered from a number of continuous bytes of only one known/chosen plaintext and the corresponding ciphertext.
The rest of this paper is organized as follows. At first, Sec. \[sec:DSEA\] gives a brief introduction to DSEA. Then, the cryptanalysis results are presented in detail in Sec. \[sec:Cryptanalysis\], with some experimental results. Section \[sec:ImprovingDSEA\] discusses how to improve DSEA. The last section concludes the paper.
Domino Signal Encryption Algorithm (DSEA) {#sec:DSEA}
=========================================
Assume that the plaintext is $g=\{g(n)\}_{n=0}^{M-1}$ and that the ciphertext is $g'=\{g'(n)\}_{n=0}^{M-1}$, where $g(n)$ and $g'(n)$ denote the $n$-th plain-byte and cipher-byte, respectively. Then, the encryption procedure of DSEA can be described as follows (see also Fig. \[figure:DSEA\]).
- *The secret key*: two integers, $L\in\{1,\cdots,M\}$, $initial\_key\in\{0,\cdots,255\}$, the control parameter $\mu$ and the initial condition $x(0)$ of the following chaotic Logistic map[@Devaney:Chaos; @HaoBailin:ChaoticDynamics]: $$x(k+1)=\mu\cdot x(k)\cdot(1-x(k)).$$
- *The initialization procedure*: under 8-bit finite computing precision, run the Logistic map from $x(0)$ to generate a chaotic sequence $\{x(k)\}_{k=0}^{\lceil M/8\rceil-1}$, and then extract the 8 significant bits of $x(k)$ to yield a PRBS $\{b(n)\}_{n=0}^{M-1}$, where $x(k)=\sum_{i=0}^7\left(b_{8k+i}\cdot
2^{-(i+1)}\right)=0.b_{8k+0}\cdots b_{8k+7}$.
- *The encryption procedure*: for $n=0\sim M-1$, do $$g'(n)=
\begin{cases}
g(n)\oplus true\_key, & b(n)=1,\\
g(n)\oplus \overline{true\_key}, & b(n)=0,
\end{cases}$$ where $$true\_key=
\begin{cases}
initial\_key, & n\bmod L=0,\\
g'(n-1), & n\bmod L\neq 0,
\end{cases}$$ and $\oplus$ denotes the bitwise XOR operation.
![The diagrammatic view of the encryption procedure of DSEA.[]{data-label="figure:DSEA"}](DSEA_Cipher){width="70.00000%"}
The decryption procedure is identical with the above encryption procedure, since XOR is an invertible operation.
Cryptanalysis {#sec:Cryptanalysis}
=============
Brute-force attack
------------------
The brute-force attack is the attack of exhaustively searching the secret key from the set of all possible keys [@Schneier:AppliedCryptography96]. Apparently, the attack complexity is determined by the size of the key space and the complexity of verifying each key. The secret key of DSEA is $(L,
initial\_key, \mu, x(0))$, which has $M\cdot 2^{3\cdot 8}=M\cdot
2^{24}$ possible values. Taking the complexity of verifying each key into consideration, the total complexity of searching for all possible keys is $O\left(2^{24}\cdot M^2\right)$. When the plaintext is selected as a typical image of size $256\times 256$, the complexity will be $O(2^{56})$, which is much smaller than $O(2^M
\cdot M)=O(2^{65552})$, the complexity claimed in [@Yen-Guo:DSEA:JCIEE2003]. Note that the real complexity is even smaller since not all values of $\mu$ can ensure the chaoticity of the Logistic map [@Devaney:Chaos; @HaoBailin:ChaoticDynamics]. That is, the security of DSEA against brute-force attacks was over-estimated much in [@Yen-Guo:DSEA:JCIEE2003]. In today’s digitized and networked world, the complexity of order $O(2^{128})$ is required for a cryptographically-strong cipher [@Schneier:AppliedCryptography96], which means DSEA is not practically secure.
Ciphertext-only attacks {#sec:CiphertextOnlyAttack}
-----------------------
Ciphertext-only attacks are such attacks in which one can access a set of ciphertexts [@Schneier:AppliedCryptography96]. Since the transmission channel is generally insecure, the security against ciphertext-only attacks are required for any ciphers. However, it is found that DSEA is not sufficiently secure against ciphertext-only attacks, since much information about the plaintext and the secret key can be found from even one ciphertext.
Given an observed ciphertext $g'$, generate two mask texts, $g_0^*$ and $g_1^*$, as follows: $g_0^*(0)=0$, $g_1^*(0)=0,
\forall\; n=1\sim M-1$, $g_0^*(n)=g'(n)\oplus \overline{g'(n-1)}$, $g_1^*(n)=g'(n)\oplus g'(n-1)$. From the encryption procedure of DESA, it can be easily verified that the following result is true when $n \bmod L\neq 0$: $$g(n)=\begin{cases}
g_0^*(n), & b(n)=0,\\
g_1^*(n), & b(n)=1,
\end{cases}$$ which means that $g(n)$ is equal to either $g_0^*(n)$ or $g_1^*(n)$. Assuming that each chaotic bit distributes uniformly over $\{0,1\}$, one can deduce that the percentage of right plain-pixels in $g_0^*$ and $g_1^*$ is not less than $\frac{L-1}{L}\cdot\frac{1}{2}=\frac{1}{2}-\frac{1}{2L}$. When $L$ is large, about half pixels in $g_0^*$ and $g_1^*$ are plain-pixels in $g$, and it is expected that some visual information of the plain-image can be distinguished from $g_0^*$ and $g_1^*$.
To verify the above idea, one $256\times 256$ image, “Lenna", has been encrypted to get $g_0^*$ and $g_1^*$, with the following secret parameters: $L=15$, $initial\_key=170$, $\mu=251/2^6\approx
3.9219$, $x(0)=69/2^8\approx 0.2695$. The experimental results are shown in Fig. \[figure:CiphertextOnlyAttack\]. In $g_0^*$ there are 27726 pixels that are identical with those in $g$, and in $g_1^*$ there are 33461 such pixels. Observing Figs. \[figure:CiphertextOnlyAttack\] c and d, one can see that the plain-image roughly emerges from both $g_0^*$ and $g_1^*$.
![A ciphertext-only attack to DSEA.[]{data-label="figure:CiphertextOnlyAttack"}](lenna "fig:"){width="\textwidth"} a) The plain-image $g$
![A ciphertext-only attack to DSEA.[]{data-label="figure:CiphertextOnlyAttack"}](lenna_e "fig:"){width="\textwidth"} b) The cipher-image $g'$
\
![A ciphertext-only attack to DSEA.[]{data-label="figure:CiphertextOnlyAttack"}](lenna_d0 "fig:"){width="\textwidth"} c) The mask image $g_0^*$
![A ciphertext-only attack to DSEA.[]{data-label="figure:CiphertextOnlyAttack"}](lenna_d1 "fig:"){width="\textwidth"} d) The mask image $g_1^*$
In addition, from either $g_0^*$ or $g_1^*$, it is possible to directly get the value of $L$, if there exists strong correlation between adjacent bytes of the plaintext (speeches and natural images are good examples). This is due to the probability difference existing between the following two kinds of plain-bytes:
- when $n\bmod L\neq 0$, $g_0^*(n)=g(n)$ and $g_1^*(n)=g(n)$ with a probability of $\frac{1}{2}$;
- when $n\bmod L=0$, $g_0^*(n)=g(n)$ and $g_1^*(n)=g(n)$ with a probability[^3] of $\frac{1}{256}$: $g_0^*(n)=g(n)$ if and only if $g'(n-1)=\overline{initial\_key}$; $g_1^*(n)=g(n)$ if and only if $g'(n-1)=initial\_key$.
When there exists strong correlation between adjacent bytes, the above probability difference implies that there exists strong discontinuity around each position satisfying $n\bmod L=0$ (with a high probability). The fixed occurrence period of such discontinuous bytes will generate periodically-occurring straight lines in the mask text when it is an image or displayed in 2-D mode, as shown in Figs. \[figure:CiphertextOnlyAttack\]c and d. Then, it is easy to determine the occurrence period, i.e., the value of $L$, by checking the horizontal distance between any two adjacent lines. To make the straight line clearer, one can calculate the differential images of $g_0^*$ and $g_1^*$, as shown in Fig. \[figure:DifferenceImages\], where the differential image of an image $g=\{g(n)\}_{n=0}^{M-1}$ is defined as follows: $g_d(0)=g(0)$ and $\forall\;n=1\sim M-1$, $g_d(n)=|g(n)-g(n-1)|$. Note that the two differential images of $g_0^*$ and $g_1^*$ are identical according to the following theorem, from which one can get that $|g_0^*(n)-g_0^*(n-1)|=|g'(n)\oplus\overline{g'(n-1)}-g'(n-1)\oplus\overline{g'(n-2)}|
=|g'(n)\oplus g'(n-1)-g'(n-1)\oplus g'(n-2)|=|g_1^*(n)-g_1^*(n-1)|$.
![The differential images of $g_0^*$ and $g_1^*$.[]{data-label="figure:DifferenceImages"}](lenna_d0_diff "fig:"){width="\textwidth"} a) $g_{d,0}^*$
![The differential images of $g_0^*$ and $g_1^*$.[]{data-label="figure:DifferenceImages"}](lenna_d1_diff "fig:"){width="\textwidth"} b) $g_{d,1}^*$
For any three $s$-bit integers, $a,b,c$, it is true that $|(a\oplus b)-(b\oplus c)|=|(a\oplus\bar{b})-(b\oplus\bar{c})|$.
Introduce four new variables, $A=a\oplus b$, $B=b\oplus c$, $A'=a\oplus\bar{b}$, $B'=b\oplus\bar{c}$. It can be easily verified that $A'=\overline{A}$ and $B'=\overline{B}$, since $a\oplus\bar{b}=a\oplus b\oplus b\oplus\bar{b}=a\oplus b\oplus
(2^s-1)=\overline{a\oplus b}$. That is, $(a\oplus b)-(b\oplus c)=A-B$ and $(a\oplus\bar{b})-(b\oplus\bar{c})=\overline{A}-\overline{B}$. Let $A=(A_0\cdots A_{s-1})_2=\sum_{i=0}^{s-1}A_i\cdot 2^i$, $B=(B_0\cdots B_{s-1})_2=\sum_{i=0}^{s-1}B_i\cdot 2^i$. Since $\forall\; A_i,B_i\in\{0,1\}$, $A_i-B_i=\bar{B_i}-\bar{A_i}$, it is obvious that $A-B=\sum_{i=0}^{s-1}(A_i-B_i)\cdot
2^i=\sum_{i=0}^{s-1}(\bar{B_i}-\bar{A_i})\cdot
2^i=\overline{B}-\overline{A}$. As a result, $|(a\oplus
b)-(b\oplus
c)|=|A-B|=|\overline{B}-\overline{A}|=|\overline{A}-\overline{B}|=|(a\oplus\bar{b})-(b\oplus\bar{c})|$, which completes the proof.
Known/chosen-plaintext attacks {#subsec:KnownPlaintextAttacks}
------------------------------
Known/chosen-plaintext attacks are such attacks in which one can access/choose a set of plaintexts and observe the corresponding ciphertexts [@Schneier:AppliedCryptography96]. In today’s networked world, such attacks occur more and more frequently. For a cipher with a high level of security, the security against both known-plaintext and chosen-plaintext attacks are required. Although it was claimed that DSEA can resist this kind of attacks [@Yen-Guo:DSEA:JCIEE2003 Sec. IV.B], we found this claim is not true: with a limited number of continuous plain-bytes of only one known/chosen plaintext, one can completely break the secret key to decrypt other unknown plain-bytes of the known/chosen plaintext and any new ciphertexts encrypted with the same key. Apparently, even when the secret key is changed for each plaintext (as mentioned in [@Yen-Guo:DSEA:JCIEE2003 Sec. IV.B]), DSEA is insecure against known/chosen-plaintext attacks. In the following, let us discuss how to break the four sub-keys, respectively.
*1) Breaking the sub-key $L$:* as mentioned above, once one gets a ciphertext, he can easily deduce the value of $L$ by observing the periodically-occurring straight lines in the two constructed mask texts, $g_0^*$ and $g_1^*$. Furthermore, since the plaintext is also known, it is possible to generate an enhanced differential image, $g_d^*$, as follows: $g_d^*(0)=0$, and $\forall\; n=1\sim M-1$, $$g_d^*(n)=\begin{cases} 0, & g(n)\in\{g_0^*(n), g_1^*(n)\},\\
255, & g(n)\not\in\{g_0^*(n), g_1^*(n)\}.
\end{cases}$$ See Fig. \[figure:EnhancedDifferenceImage\] for the enhanced differential image corresponding the cipher-image shown in Fig. \[figure:CiphertextOnlyAttack\]b. Compared with Fig. \[figure:DifferenceImages\], one can see that the straight lines become clearer.
![The enhanced differential image $g_d^*$.[]{data-label="figure:EnhancedDifferenceImage"}](diff){width="\figwidth"}
*2) Breaking the $initial\_key$:* for all values of $n$ that satisfy $n\bmod L=0$, it is obvious that $$\label{equation:GetInitialKey}
initial\_key=
\begin{cases}
g(n)\oplus g'(n), & b(n)=1,\\
\overline{g(n)\oplus g'(n)}, & b(n)=0.
\end{cases}$$
Note that it is possible to uniquely determine the value of $initial\_key$, when there may exist pixels satisfying $n\bmod
L=0$ and $g_d^*(n)=0$, i.e., $g(n)\in\{g_0^*(n),
g_1^*(n)\}=\left\{g'(n)\oplus \overline{g'(n-1)},g'(n)\oplus
g'(n-1)\right\}$. Considering $g'(n)=g(n)\oplus initial\_key$, one can immediately deduce that $$\label{equation:GetInitialKey}
initial\_key=
\begin{cases}
g'(n-1), & g(n)=g_1^*(n),\\
\overline{g'(n-1)}, & g(n)=g_0^*(n).
\end{cases}$$
*3) Breaking the chaotic PRBS and the other two sub-keys:* once $L$ and $initial\_key$ have been determined, the chaotic PRBS, $\{b(n)\}_{n=0}^{M-1}$, can be immediately derived as follows:
- when $n \bmod L\neq 0$: if $g(n)=g_0^*(n)$ then $b(n)=0$, else $b(n)=1$;
- when $n \bmod L=0$: if $initial\_key=g(n)\oplus g'(n)$ then $b(n)=1$, else $b(n)=0$.
Once $\{b(n)\}_{n=0}^{M-1}$ is uniquely determined, $x(0)=0.b(0)\cdots b(7)$ can be immediately recovered.
With 16 consecutive chaotic bits, $b(8k+0)\sim b(8k+15)$, one can further derive two consecutive chaotic states: $x(k)=0.b(8k+0)\cdots b(8k+7)$ and $x(k+1)=0.b(8k+8)\cdots
b(8k+15)$, and then derive an estimation of the sub-key $\mu$ as $$\widetilde{\mu}=\frac{x(k+1)}{x(k)\cdot(1-x(k))}.$$ Due to the quantization errors introduced in the finite-precision arithmetic, generally $x(k+1)\neq\mu\cdot x(k)\cdot(1-x(k))$, so $\widetilde{\mu}\neq\mu$. Fortunately, following the error analysis of $\widetilde{\mu}$ in [@Li:AttackingCNN2004 Sec. 3.2], the following result has been obtained: when $x(k+1)\geq 2^{-n}\;(n=1\sim
8)$, $|\widetilde{\mu}-\mu|<2^{n+3}\cdot 2^{-8}$. Specially, when $x(k+1)\geq 2^{-1}=0.5$, $|\widetilde{\mu}-\mu|<2^4\cdot 2^{-8}$, which means that one can exhaustively search for $2^4=16$ values in the neighborhood of $\widetilde{\mu}$ to find the right value of $\mu$. To verify which searched value is the right one, one can iterate the Logistic map from $x(k+1)$ for some times to get some new chaotic states and then check the coincidence between these chaotic states and corresponding recovered chaotic bits.
With the above steps, the whole secret key $(L,initial\_key,\mu,x(0))$ can be recovered, and then be used for decryption. For the plain-image “Lenna", a breaking result is shown in Fig. \[figure:KnownPlaintextAttack\]. It can be verified that the complexity of the known/chosen-plaintext attacks is only $O(M)$, which means a perfect breaking of DSEA.
![The recovered plain-image of “Lenna" in a known-plaintext attack.[]{data-label="figure:KnownPlaintextAttack"}](lenna_e_d){width="\figwidth"}
Improving DSEA {#sec:ImprovingDSEA}
==============
In this section, we study some possible remedies to DSEA to resist the proposed attacks. It is concluded that DSEA cannot be simply enhanced to resist known/chosen-plaintext attacks.
To ensure the complexity of the brute-force attack cryptographically large, the simplest idea is to increase the presentation precision of $x(0)$ and $\mu$. Binary presentations of $x(0)$ and $\mu$ with 64-bit (long integers) are suggested to provide a complexity not less than $O(2^{128})$ against the brute-force attack.
Apparently, the insecurity of DSEA against ciphertext-only and known/chosen-plaintext attacks is mainly due to the invertibility of XOR operations. This is actually the weakness of all XOR-based stream ciphers. To make DSEA securer, one has to change the encryption structure and/or the basic masking operations, in other words, one has to design a completely new cipher, instead of enhancing DSEA to design a modified cipher.
In addition, there exists a special flaw in DSEA. According to [@Li:Dissertation2003 Sec. 2.5], when a chaotic system is implemented in $s$-bit finite computing precision, each chaotic orbit will lead to a cycle whose length is smaller than $2^s$ (and generally much smaller than $2^s$). Figure \[figure:ChaoticBits\]a shows the pseudo-image of the chaotic PRBS recovered in a known-plaintext attack. It is found that the cycle of the chaotic PRBS is only $2^6=64$ and the period of the corresponding chaotic orbit is only $2^3=8$. Such a small period of the chaotic PRBS will make all attacks easier. To amend this defect, using a higher implementation precision or floating-point arithmetic is suggested. Figure \[figure:ChaoticBits\]b gives the pseudo-image of the chaotic PRBS when the chaotic states are calculated under double-precision floating-point arithmetic. It is obvious that the short-period effect of the chaotic PRBS is effectively avoided.
![The pseudo-image of the chaotic PRBS, under two different finite-precision arithmetics.[]{data-label="figure:ChaoticBits"}](bit "fig:"){width="\figwidth"} a) 8-bit fixed-point arithmetic
![The pseudo-image of the chaotic PRBS, under two different finite-precision arithmetics.[]{data-label="figure:ChaoticBits"}](bit_f "fig:"){width="\figwidth"} b) double-precision floating-point arithmetic
Conclusion
==========
In this paper, the security of a recently-proposed signal security system called DSEA [@Yen-Guo:DSEA:JCIEE2003] has been studied in detail. It is pointed out that DSEA is not secure enough against the following attacks: the brute-force attack, ciphertext-only attacks, and known/chosen-plaintext attacks. Experimental results are also given to support the theoretical analysis. Also, some remedies of enhancing the performance of DSEA are discussed. As a conclusion, DSEA is not suggested in serious applications requiring a high level of security.
Acknowledgements
================
This research was partially supported by the National Natural Science Foundation, China, under grant no. 60202002, and by the Applied R&D Centers of the City University of Hong Kong under grants nos. 9410011 and 9620004.
[^1]: This paper has been published in *Journal of Systems and Software*, vol. 79, no. 2, pp. 253-258, 2006.
[^2]: Chaos is a dynamical phenomenon demonstrated in many dynamical systems [@Devaney:Chaos; @HaoBailin:ChaoticDynamics]. Due to the tight relationship between chaos and cryptography, chaotic systems have been used to design encryption schemes since 1990s. For a survey of digital chaotic ciphers, see [@Li:Dissertation2003 Chap. 2].
[^3]: Without loss of generality, it is assumed that each cipher-byte distributes uniformly in $\{0,\cdots,255\}$.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- Siegfried Röser
- Elena Schilbach
bibliography:
- 'mybib.bib'
date: 'Received 20 March 2019; accepted 16 April 2019.'
title: 'Praesepe (NGC 2632) and its tidal tails.'
---
[Within a 400 pc sphere around the Sun, we search for Praesepe’s tidal tails in the Gaia DR2 dataset.]{} [We used a modified convergent-point method to find stars with space velocities close to the space velocity of the Praesepe cluster.]{} [We find a clear indication for the existence of Praesepe’s tidal tails, both extending up to 165 pc from the centre of the cluster. A total of 1393 stars populate the cluster and its tails, giving a total mass of 794 M$_\odot$. We determined a tidal radius of 10.77 pc for the cluster and a tidal mass of 483 M$_\odot$. The corresponding half-mass radius is 4.8 pc. We also found clear indication for mass segregation in the cluster. The tidal tails outside 2 tidal radii are populated by 389 stars. The total contamination of our sample by field stars lies between 50 to 100 stars or 3.6 to 7.2$\%$. We used an astrometrically and photometrically clean sub-sample of Gaia DR2 which makes our Praesepe sample incomplete beyond $M_G \approx$ 12.0 mag, which corresponds to about 0.25 M$_\odot$. A comparison with an N-body model of the cluster and its tails shows remarkably good coincidence. Three new white dwarfs are found in the tails.]{}
Introduction {#intro}
============
Tidal tails around open clusters have been predicted about a decade ago by, e.g. and . An introduction to this topic was given in our previous paper , hereafter called Paper I, and is not repeated here. The first detection of tidal tails of open clusters using the data from Gaia Data Release 2 were reported at the end of 2018 by and who revealed the tidal tails of the Hyades. This was followed by a number of publications where tidal tails have now been found around other open clusters. @2019arXiv190201404T and @2019arXiv190207216F recently reported about the detection of tidal tails of the Coma Berenices (Melotte 111) open cluster, and @2019arXiv190104253Y give hints for tidal tails from the very old, $\log \rm{t}$ = 9.33 , open cluster Ruprecht 147. So, tidal tails now seem to be a common feature of older open clusters and are revealed thanks to the high data quality of Gaia DR2.
NGC 2632 (Praesepe) is an old open cluster, $\log \rm{t}$ = 8.92 in MWSC, $\log \rm{t}$ = 8.85 according to . The latest major review of the stellar content of Praesepe [@2007AJ....134.2340K] dates back more than one decade. Over an area of 300 $\deg^2$ on the sky, they found 1010 candidate members in Praesepe, and also estimated the total mass to be 550 $\pm$ 40 M$_\odot$ and a tidal radius of 11.5 $\pm$ 0.3 pc. @2013MNRAS.434.3236K used 2MASS and SDSS photometry and proper motions from the PPMXL catalogue to determine the stellar content of Praesepe. They found a total cluster mass of about 630 M$_\odot$ and a 2-D half-mass radius of 3.90 pc. Using Gaia DR2 data, and published membership lists of stars for quite a few open clusters in the Solar neighbourhood including Praesepe. found 938 candidate members, and they also determined a mean parallax, mean proper motions, as well as age of, and reddening towards NGC 2632. For Praesepe, give membership probabilities for 719 stars as well as mean parallax and mean proper motions.
This paper primarily aimed at the detection of the tidal tails of Praesepe, if they existed, but we also publish some characteristic astrophysical data of the cluster itself. The paper is structured as follows: In Section \[detect\] we describe the steps we took to find Praesepe’s tidal tails. Section \[detect\] is divided into subsections describing the cuts to obtain an astrometrically and photometrically clean sample, the separation of over-densities in position and velocity space from the Galactic background, the identification of Praesepe and its tails, and the estimation of contamination and completeness. In Section \[CAMDs\] we discuss the cluster and its tails and present some characteristic astrophysical parameters. A summary concludes the paper.
{width="\textwidth"}
{width="\textwidth"}
\
Finding Praesepe’s tidal tails {#detect}
==============================
The N-body simulations of the dynamical evolution of a Galactic open cluster literally were tailored to model the case of the Hyades cluster at an age of $\log \rm{t}$ = 8.8 [@2018ApJ...856...40M; @2018ApJ...863...67G] and at about the galacto-centric distance of the Sun. As the galacto-centric distance and age of Praesepe are comparable to that of the Hyades, and as the clusters are also similar with regard to their stellar content, it is justified that we use the data set of the model from as our model cluster to compare with the observations of Praesepe.
As a starting point we determined the 6-D phase space coordinates of the centre of Praesepe on the basis of positions, proper motions, parallaxes, and radial velocities from Gaia DR2 and information on membership from . The coordinates are given in the solar system bary-centric Galactic Cartesian coordinate system where the $X$-axis points to the Galactic centre, the $Y$-axis in the direction of Galactic rotation, and $Z$-axis to the Galactic north pole. The corresponding velocity coordinates are $U, V, W$. We found mean values $$\begin{array}{lcl}
\vec{R_c} = (X_c,Y_c,Z_c) & = & (-141.75 , -69.35, 100.45)\,{\rm pc}, \\
\vec{V_c} = (U_c,V_c,W_c) & = & ( -43.17, -20.62, -9.60)\,{\rm km\,s^{-1}}.\label{COPO}
\end{array}$$ These mean values are based upon 136 stars from with measured radial velocities. The standard errors of the mean in a position coordinate are 0.25 pc, and those of a velocity coordinate are 0.08 km s$^{-1}$.
Preparatory work {#prep}
----------------
From the Gaia DR2 dataset we extracted all entries with a parallax greater than 2.4 mas, which gave 43,038,142 objects. For the further processing, we followed the procedures described in , Sect. 4.3 and Appendix C, Figs. C.1 and C.2, to obtain a stellar sample cleaned from possible artefacts with one small, but important change. Instead of the “unit weight error"-cut (UWE) we followed the new approach presented by Lindegren in the document `GAIA-C3-TN-LU-LL-124-01` which can be found on ESA’s webpage[^1]. In this concept UWE is scaled by a factor depending on the magnitude and colour of the source. The result is a re-normalised UWE, or RUWE, which needs only a single threshold of 1.4. Values for RUWE can be found at the service[^2] in table `gaiaDR2light`. All the other cuts are the same as in Paper I. We obtained an astrometrically and photometrically clean sample of 8,131,092 stars out of the original 43 million objects. The price for this cleaning is an incompleteness at the faint end of the magnitude distribution near $G$ = 18 mag and also near the parallax limit of 2.4 mas.
Constraining the range in space and velocity {#CVspace}
--------------------------------------------
In general, a cluster and its possible tidal tails reveal themselves as an over-density in position- and velocity-space. In the ideal case, we need observed space velocities for each star in a sample to identify stars co-moving with the Praesepe cluster. However, accurate radial velocities are lacking for the vast majority of stars. Therefore we had to rely on criteria that are solely based on their tangential velocities. This implies that we may detect stars that are highly probable co-moving, although they need final confirmation when the radial velocities are measured. We followed the formalism of the convergent-point (CP) method as described in , for instance, and transformed the Cartesian velocity vector of the cluster motion from Eq.\[COPO\] to give predicted velocities $V_{\parallel pred}$ and $V_{\bot pred}$ parallel and perpendicular to the direction to the CP for each star depending on its position on the sky. We note that $V_{\bot pred} \equiv$ 0. Also following , we similarly transformed the measured (observed) tangential velocities for each star, $\kappa\,\mu_{\alpha*}/\varpi$ and $\kappa\,\mu_{\delta}/\varpi$ into $V_{\parallel obs}$ and $V_{\bot obs}$. Here $\varpi$ is the measured trigonometric parallax in Gaia DR2 and $\kappa=4.74047$ is the transformation factor from 1 mas yr$^{-1}$ at 1 kpc to 1 km s$^{-1}$. We set in the following $\Delta V_{\parallel}$ = $V_{\parallel obs}$ - $V_{\parallel pred}$ and $\Delta V_{\bot}$ = $V_{\bot obs}$ - $V_{\bot pred}$. Note that the differences between the predicted and observed velocities ($\Delta V_{\parallel}$, $\Delta V_{\bot}$) will be equal to (0,0), when the space velocity of a star is identical to $\vec{V_c}$ in Eq.\[COPO\]. We also determined the covariance matrix for the velocities $V_{\parallel obs}$ and $V_{\bot obs}$ via error propagation from the covariance matrix of the $\mu_{\alpha*}, \mu_{\delta}$, and $\varpi$.
For further analysis we did not use the full sphere with $\varpi \leq$ 2.4 mas, but constrained the volume around Praesepe by a cut in the z-plane as $|{Z - Z_c}| \leq 50$ pc where $Z_c$ is given in Eq. \[COPO\]. This gives a volume of 4.867$\times 10^7~\rm{pc}^3$. We also introduced a restriction in the tangential velocity plane by requiring $|\Delta V_{\parallel}|\leq $7 km s$^{-1}$ and $|\Delta V_{\bot}| \leq$7 km s$^{-1}$, a total area of 196 ${(\rm{km~s}^{-1})}^2$. These cuts are ample compared to the predicted extent of the model tails, and they reduced our selection to a sub-sample containing 78,163 stars which populate this search volume.
As we are interested in Praesepe and its possible tails, we also transform the spacial coordinates $X,Y,Z$ to a coordinate system $X_P,Y_P,Z_P$ (the subscript $P$ refers to Praesepe) with origin at the centre of the Praesepe cluster, where the $X_P$-axis points to the Galactic centre, $Y_P$-axis to the direction of Galactic rotation and the $Z_P$-axis to the Galactic North Pole. To achieve this we first went from the solar system bary-centric Galactic Cartesian coordinates $X,Y,Z$ to the Galactocentric cylindrical coordinate system $R,\theta,Z$ with $R_\odot$ = 8300 pc [@2013IAUS..289...29G] . Then we transformed back into the cluster-centric Galactic Cartesian system $X_P,Y_P,Z$, and finally $Z_P$ = $Z -Z_c$. Note that this transformation is not necessary for the way we determine over-densities, it only puts Praesepe into the centre of the volume we consider.
A general method to find over-densities {#ITT}
---------------------------------------
In Paper I we cut a small rectangle in the tangential velocity plane where we suspected the signature of the tails, and then looked for over-densities in 3-D space. Here we follow a different, more general approach: we search simultaneously for over-densities in 5-D space. The main purpose is to separate the over-densities such as clusters or moving groups (signal) from the local Galactic background (noise). In other words, for each star in the sub-sample of 78,163 stars we want to decide, if it belongs to the general Galactic field, or to an over-density significantly above the general background. We make the decision by counting neighbours of this star in a 5-D neighbourhood and compare this with the expected background counts where the background is modelled by a Poisson distribution. We make the following neighbourhood definition: a star $j$ with 5-D coordinates $(X_{P_{j}},Y_{P_{j}},Z_{P_{j}},\Delta V_{\parallel_{j}},\Delta V_{\bot_{j}})$ is a neighbour to star $i$ with coordinates $(X_{P_{i}},Y_{P_{i}},Z_{P_{i}},\Delta V_{\parallel_{i}},\Delta V_{\bot_{i}})$ if $$(X_{P_{i}}-X_{P_{j}})^2 + (Y_{P_{i}}-Y_{P_{j}})^2 +(Z_{P_{i}}-Z_{P_{j}})^2 \leq {r_{lim}}^2,$$ and $$\frac{(\Delta V_{\parallel_{i}}-\Delta V_{\parallel_{j}})^2}{a^2} + \frac{(\Delta V_{\bot_{i}}-\Delta V_{\bot_{j}})^2}{b^2} \leq 1.$$ The free parameters $a,b$ and ${r_{lim}}$ can be specified according to the goal of the study. We chose an elliptical shape for the velocity condition since the error ellipses in the tangential velocity plane can be eccentric. Indeed, in the case of the Praesepe sub-sample, $\sigma_{V_{\parallel obs}}$ and $\sigma_{V_{\bot obs}}$ are quite different; the mean $\sigma_{V_{\parallel obs}}$ is 0.79 km s$^{-1}$, and the mean $\sigma_{V_{\bot obs}}$ is 0.12 km s$^{-1}$. Considering also an intrinsic velocity dispersion in the cluster and tails, we chose 1.2 km s$^{-1}$ as the semi-major axis $a$ of our selection ellipse and 0.5 km s$^{-1}$ as the semi-minor axis $b$. With these velocity conditions we found, on average, 1.6$\times 10^{-5}$ stars for each velocity ellipse and per 1 pc$^{-3}$. Further we chose r$_{lim}$ = 15 pc to get, on average, 0.2 stars per 5-D neighbourhood volume. Note that we can determine neighbourhood only for those stars, where the volumes specified for neighbourhood are fully contained in the subset given in Sec. \[CVspace\]. This holds for 54652 stars which we call the basic sample (BS).
--- -------- --------- ------- --------
0 0.761 41587.5 41586 1.000
1 0.208 11361.2 9952 1.142
2 0.028 1551.8 1571 0.988
3 0.003 141.3 252 0.561
4 1.8E-4 9.6 85 0.114
5 9.6E-6 0.5 36 0.015
6 4.4E-7 2.4E-2 42 5.7E-4
7 1.7E-8 9.4E-4 20 4.7E-5
--- -------- --------- ------- --------
: Distribution of the $k$-neighbourhoods of stars in the BS. For more information see text.[]{data-label="table:1"}
When a star $i$ is surrounded by $k$ neighbours, we call this case a “$k$-neighbourhood”. In the BS we found a total of 41586 zero-neighbourhoods and 9952 one-neighbourhoods, i.e. the vast majority of stars (94.3%) has less than 2 neighbours in the BS. With high probability, these are Galactic field stars (background noise). To determine an upper limit for the average density of the neighbours in the background we tentatively allowed all cases with up to 5 neighbours to be considered as background. These contain 97.9 per cent of the stars in the basic sample, which, on the other hand means that only the remaining about 2.1 per cent are signal. The average number of neighbours in the BS is 0.273 stars. So, the distribution of neighbours in the background should follow a Poisson distribution with the mean $\lambda$ = 0.273. We then calculated the quantity $N_{exp}(k) = P_\lambda(k) \times N_{BS}$ where $P_\lambda(k)$ is the probability mass function of the Poisson distribution representing the background. We give $P_\lambda(k)$ in the second and $N_{exp}(k)$ in the 3rd column of Table \[table:1\] for all k $\leq$ 7. The fourth column in Table \[table:1\] gives the observed number of $k$-neighbourhoods $N_{obs}(k)$ in the BS. Comparing columns 3 and 4 we found that $N_{obs}(k)$ follows $N_{exp}(k)$ quite well for cases $k$=0,1,2. For increasing $k$ the number of observed $k$-neighbourhoods $N_{obs}(k)$ rapidly exceeds the prediction $N_{exp}(k)$. If the case $k$=5 would belong to the background, the expected number of cases would be less than 1, observed are 36. So, we did certainly not underestimate the mean density $\lambda$ for the background distribution. In the case of Praesepe the highest value of $k$ is 405. The ratio $p_{cont}$ = $N_{exp}(k)$/$N_{obs}(k)$ given in column 5 of Table \[table:1\] is the probability that a star in a $k$-neighbourhood belongs to the background distribution. Hence, for each star, we can interpret $p_{cont}$ as the degree how much it is contaminated by background. According to the construction of these $k$-neighbourhoods a particular star from the BS can be found in the $k$-neighbourhood of several other stars, and we give it the contamination $p_{cont}$ of the neighbourhood with the highest $k$, where this star is found.
{width="\textwidth"}
{width="\textwidth"}
\
In Fig. \[Figure1\] we show the distribution of all stars from the BS coded by their individual $p_{cont}$. The left panel shows the distribution in the $\Delta V_{\parallel}$, $\Delta V_{\bot}$-plane, in the right panel we present the same set of stars, but in the $X_P,Y_P$-plane. The background is filled by the vast majority (97.2%) of stars from the BS having $p_{cont}$ $\approx$ 1 (cases $k$ = 0,1,2 in Table \[table:1\]), which cover the whole area. Stars having intermediate contamination (cases $k$ = 3,4) are clearly identifiable by their colour, and the remaining stars with $p_{cont} \approx 0$ are shown in black. The huge signal centred at $\Delta V_{\parallel}$, $\Delta V_{\bot}$ = (0, 0) in the left panel and $X_P,Y_P$ = (0, 0) in the right panel represents Praesepe and its tails. A minor over-density at about $\Delta V_{\parallel}$, $\Delta V_{\bot}$ = (2.5, 6.3) and $X_P,Y_P$ = (445, 220) is the open cluster Collinder 350 and has nothing to do with Praesepe. In the following we considered only cases with $k$ $\geq$ 3 which we regard as signal. This yields 1543 stars or 2.8% of the BS.
The selection of the parameters $(a,b)$ = (1.2 km s$^{-1}$, 0.5 km s$^{-1}$) in velocity space and ${r_{lim}}$ = 15 pc in positional space turned out to be a reasonable compromise to obtain a significant signal for Praesepe and its tails on one hand and a relatively small noise in the background on the other hand (see Fig.\[Figure1\]). We tentatively kept $a$ and $b$ constant and increased ${r_{lim}}$ up to 25 pc. This resulted in a slight increase of the length of the tails together with a raise of spurious over-densities in the surrounding background created at velocities inconsistent with that of Praesepe. A similar effect occurred when we kept ${r_{lim}}$ constant at 15 pc and increased $(a,b)$ to (1.5 km s$^{-1}$, 0.7 km s$^{-1}$).
We used TOPCAT [@2005ASPC..347...29T] to extract the stars belonging to Praesepe and its tails out of the 1543 stars with $k$ $\geq$ 3. Three manual cuts, first in the $X_P,Y_P$-plane, then in the $\Delta V_{\parallel}$, $\Delta V_{\bot}$-plane and cut in the $X_P,Z_P$-plane were necessary to give the final number (1393) of members (candidates) for Praesepe and its tails. Henceforth we call this set of 1393 stars the Praesepe Sample (PrS). By adding $p_{cont}$ of the stars from the PrS we found a total contamination of 47 stars, 40 of which come from the stars in a 3-neighbourhood alone.
In Fig. \[Figure2\] we show the 1393 stars from Praesepe and its tails in the $(Y_P,X_P)$-plane and the $(Y_P,Z_P)$-plane. Stars coded in grey are 1266 stars with $p_{cont}$ <0.015, in orange 53 stars with $p_{cont}$ $\approx$ 0.1 and in light green 74 stars with $p_{cont}$ $\approx$ 0.5. The tails extend up to 165 pc from the cluster centre. For comparison we also present the result of the N-body calculation for the Hyades cluster by , which we also transformed into the $X_P,Y_P,Z_P$ coordinate system of Praesepe. The model stars are shown as a beige density plot in the background. The coincidence between observations and model is remarkable, especially in the $(Y_P,X_P)$-plane. Such a good coincidence we already found in Paper I in the case of the Hyades. The present result could have been expected since the Hyades and Praesepe are very similar in age and in stellar content. Among the stars in Fig. \[Figure2\] which are slightly inconsistent with the model, we find that a majority of them has higher contamination (green and orange dots). A more rigorous cut in TOPCAT would have eliminated them, but it serves as an illustration how well our general selection method using the $k$-neighbourhood works. On the other hand, there are 5 stars with higher contamination in the trailing tail in Fig. \[Figure2\] that are quite consistent with the model (left panel) and they are those farthest away from the cluster centre. At the same time these are the stars that are most distant from the Sun (up to 300 pc). The restrictions from Sec. \[prep\] applied to the full Gaia DR2 data set to get an astrometrically and photometrically clean sample creates some incompleteness here, especially at the faint end of the stellar sample. Note that incompleteness in signal leads to higher contamination by noise.
The internal contamination by field stars is given by summing up the values of $p_{cont}$. We sub-divided the PrS stars into 3 subgroups: a) 725 stars within the tidal radius $r_t$ of 10.77 pc which is derived below, b) 279 stars between 1 and 2 tidal radii, and c) 389 stars outside 2 $r_t$ and in the tails. For these groups we found a contamination of 0.83 stars (group a), 0.69 stars (group b), and 46.1 stars or 11.8% (group c). These values refer only to the contamination within our 5-D approach. At this stage it is also appropriate to estimate the possible field star contamination that comes from the fact that we did not consider radial velocities for estimating membership, since radial velocity data are only measured for a small sub-set of stars in the BS. Nevertheless, for 245 stars in the PrS radial velocities are available in Gaia DR2. Only 8 of them (3.3%) have radial velocities inconsistent with Praesepe’s space velocity. The contamination increases from 1.2% within the tidal radius to 10% outside 2 $r_t$. Combining these 2 kinds of contamination we expect a total contamination between about 50 to 100 stars for our Praesepe Sample.
![Colour-absolute-magnitude diagram (CAMD) $M_G$ versus $G_{BP} - G_{RP}$ of the 1393 stars of Praesepe and its tails. Stars are colour-coded by their $p_{cont}$. In addition we plot a PARSEC isochrone (Z = 0.02) with $\log \rm{t}$ = 8.85, and $E(B-V)$ = 0.027 mag.[]{data-label="Figure3"}](Figure3.pdf){width="49.00000%"}
Astrophysical properties of Praesepe {#CAMDs}
====================================
In Fig. \[Figure3\] we show the Colour-absolute-magnitude diagram (CAMD) $M_G$ versus $G_{BP} - G_{RP}$ of the 1393 stars of the Praesepe Sample. The stars are colour-coded according to their contamination value $p_{cont}$. We also present a PARSEC-(version 1.2S)-isochrone [@2014MNRAS.444.2525C] with Z = 0.02 and $\log \rm{t}$ = 8.85, and $E(B-V)$ = 0.027 mag . The Gaia passbands from were used; we also applied the corrections to the Gaia $G$-magnitudes given in the same paper. We note here that the tails are spread over a wide part of space and, hence, the assumption of a unique reddening is not fully appropriate for tail stars. Nevertheless, over a wide range of $G_{BP} - G_{RP}$, the isochrone gives a reasonable fit to the observations except for the low-mass stars redder than about $G_{BP} - G_{RP}$ = 2.8 mag. Anyway, we used the isochrone shown in Fig. \[Figure3\] to roughly estimate individual masses via a mass-luminosity relation in the 3 Gaia photometric bands, neglecting binary issues. The incompleteness of our sample beyond $G$ $\approx$ 18 mag (see Sec. \[prep\]) creates incompleteness at $M_G$ $\approx$ 11.8 mag near the cluster centre, and brighter or fainter in the tails, depending on the individual parallax of the stars. Via our mass-luminosity relation this transfers to about 0.25 M$_\odot$.
The total mass of all stars from the PrS amounts to 794 M$_\odot$. We followed the recipes in and used the individual masses to estimate the tidal radius of the Praesepe cluster. In the galactic disk the sphere of influence of a cluster is given by $$x_L = \left(\frac{\rm{G} M_t}{4A(A-B)}\right)^\frac{1}{3}
\label{tidal}$$ where $x_L$ is the distance of the Lagrangian points from the centre, $M_t$ is the total mass inside a distance $x_L$ from the centre, G = $4.3 \times 10^{-3} \rm{pc/M_\odot (km/s)^2}$ is the gravitational constant, $A$ and $B$ are Oort’s constants at the position of the Sun. Here we use $A$ (15.3 $\rm{kms}^{-1}\rm{kpc}^{-1}$) and $B$ (-11.9 $\rm{kms}^{-1} \rm{kpc}^{-1}$) from @2017MNRAS.468L..63B. The distance $x_L$ is often referred to as the tidal radius $r_t$ of a cluster. The tidal radius $r_t$ separates, in general, stars gravitationally bound to a cluster from those that are unbound. Adding up the individual masses from the centre outwards we found $r_t$ = 10.77 pc; 725 stars are within $r_t$ and give a tidal mass $M_t$ = 483 M$_\odot$ for Praesepe. Considering our incompleteness at about 0.25 M$_\odot$, this tidal radius is a lower bound. But, even if a total of 100 M$_\odot$ in low-mass stars with $m$ < 0.25 M$_\odot$ would be contained in the inner part of the cluster, the tidal radius would only rise to 11.5 pc. In models of star clusters the Lagrangian radius for 50% of the cluster mass, plays an important role and is called half-mass radius $r_h$. Estimated from a tidal mass of 483 $\rm{M}_{\odot}$, the half-mass radius $r_h$ is 4.8 pc at the present state of evolution of Praesepe.
@2007AJ....134.2340K made an inventory of the Praesepe cluster about a decade ago. They used data from the SDSS, 2MASS, USNOB1.0, and UCAC-2.0 surveys, and analysed proper motions and photometry over 300 $\deg^2$ on the sky. They found 1010 stars in Praesepe being candidate members with probability > 80$\%$. These authors also estimated the total mass of the cluster to be 550 $\pm$ 40 M$_\odot$ and a tidal radius of 11.5 $\pm$ 0.3 pc. The small disagreement between our tidal radius and that from @2007AJ....134.2340K comes partly from their assumed values of $A$ = 14.4 $\rm{kms}^{-1}\rm{kpc}^{-1}$ and $B$ = -12.0 $\rm{kms}^{-1}\rm{kpc}^{-1}$. With the values from @2017MNRAS.468L..63B, @2007AJ....134.2340K would have got $r_t$ = 11.25 pc given that their derived total mass of 550 M$_\odot$ were inside 11.25 pc. Considering the fact that @2007AJ....134.2340K did not have available accurately measured trigonometric parallaxes, the coincidence of their tidal radius with ours is quite satisfactory. From their 1010 80$\%$-members we found 785 among our 1393 stars. From the remaining 225, only 77 passed the cuts in Sec.\[prep\], but were not selected as members when using our criteria.
Using the individual masses we determine the radial density and mass profiles of the cluster. In Fig. \[Figure4\] we present the number density $D$, the mass density $\rho$ , and the average mass per star as a function of the distance from the centre $r_{c}$. The bottom panel of Fig. \[Figure4\] reveals the mass segregation in Praesepe; the average mass per star ($\rho/D$) decreases from 1.17 $\rm{M}_{\odot}$ close to the centre to about 0.46 $\rm{M}_{\odot}$ at the tidal radius $r_t$. Again the apparent increase of the average mass per star towards larger distances from the centre is attributed to incompleteness at the low mass end. The mass density of Praesepe, shown in the middle part of Fig. \[Figure4\], is fitted to a Plummer model [@1915MNRAS..76..107P], where the mass density follows the equation $$\rho({r_{c}}) = \frac{3M_t}{4\pi{r_{co}}^3}\frac{1}{[1+(r_c/{r_{co}})^2]^{5/2}}.
\label{pluden}$$ Here $r_{co}$ is the so-called core radius of a cluster. Using the tidal mass of $M_t$ = 483 $\rm{M}_{\odot}$ the best fit to the observed density distribution (blue line) in Fig. \[Figure4\] is obtained with a core radius $r_{co}$ of 3.7 pc. In the two upper panels of Fig. \[Figure4\] the core radius is seen as the distance where the slope in the density and mass distributions changes. The corresponding Plummer model (dashed line in Fig. \[Figure4\]) shows good agreement with observations. The ratio of the half-mass radius ${r_h}$ to ${r_{co}}$ in Praesepe is 1.297 which is in remarkable coincidence with the theoretical ratio of 1.305 for a Plummer model [see @1975ApJ...200..339S].
Fig. \[Figure3\] shows the presence of white dwarfs (WD) in Praesepe and its tails. In total, we find 13 white dwarfs with 9 of them inside the tidal radius of 10.77 pc. We disregard here the two objects with $M_G \approx 14$ mag and $G_{BP} - G_{RP}$ near 1. Twelve out of 13 objects are known in SIMBAD as white dwarfs or blue proper motion stars. One star is not known as a WD in SIMBAD, but it fits perfectly to the loci of the other white dwarfs and has recently been rated as a member of Praesepe by @2019MNRAS.483.3098S. All WDs by @2019MNRAS.483.3098S are confirmed by us with the exception of which did not pass the quality cuts in Sec. \[prep\]. Three white dwarfs, , and in the tidal tails have not been mentioned before as being related to Praesepe, but now we can confirm their provenance from Praesepe.
![Number density $D$, mass density $\rho$, and average mass per star $\overline{m}$ as a function of the logarithm of the distance from the centre ${r_{c}}$. A bin size of 2 pc in steps of 1 pc is used within 10 pc; beyond this the bin size is increased to 4 pc to get a better signal-to-noise ratio. From top to bottom: the number density $D$ as a green curve, the observed mass density (blue curve) and a density distribution (black dashed line) from a Plummer model (Eq. \[pluden\]) with $M_t$ = 483 $\rm{M}_{\odot}$ and ${r_{co}}$ = 3.7 pc, the average mass per star ($\rho/D$) as a red line. []{data-label="Figure4"}](Figure4.pdf){width="49.00000%"}
A short comparison between Praesepe and the Hyades {#Praehya}
--------------------------------------------------
In Fig. \[Figure5\] we show the location on the sky of Praesepe (yellow pluses) and the Hyades (cyan pluses) as well as the tidal tails of both clusters. The background is the grey-scale density map of Gaia DR2 (©ESA/Gaia/DPAC). This composite image has been produced using the facilities of Aladin Desktop[^3].
![The tidal tails of Praesepe (yellow) and the Hyades (cyan) over-plotted to the Gaia DR2 sky using TOPCAT and Aladin facilities. []{data-label="Figure5"}](Figure5.jpg){width="49.00000%"}
A comparison with a theoretical model from N-body calculations for an 650 Myr old cluster (Hyades) showed very good coincidence with regard to the location of the tidal tails of Praesepe, too. After the findings in Paper I, this is another very satisfactory confirmation for both theory and observations, in general. In detail, we observe for both clusters that the preceding tail is more pronounced than the trailing one, though we have good reasons to suspect that this is due to different causes. In the case of Praesepe the trailing tail stretches to larger distances from the Sun and probably suffers from the incompleteness of our sample at fainter stars. Both Hyades tails are located at about the same distance from the Sun but the trailing tail of the Hyades shows an inhomogeneous structure and may have been shredded by shocks or collisions in the past. Moreover, the proportion $N_{2rt}/N_{pt}$ of stars within the central part of the cluster, inside 2$r_t$ ($N_{2rt}$), to stars in the preceding tail ($N_{pt}$) is considerably higher for Praesepe than for the Hyades, namely 5 and 2, respectively. This may indicate a later stage of dynamical evolution of the Hyades. They may be closer to dissolution than is Praesepe. However, this hypothesis must be supported by future investigations, e.g. N-body simulations.
Summary {#summ}
=======
Using the data from Gaia DR2, we searched for the presence of Praesepe’s (NGC 2632) tidal tails in a sphere of some 400 pc radius around the Sun. First, the Gaia DR2 data were cleaned according to the recipes given in and described in Sec. \[prep\] to obtain an astrometrically and photometrically clean sample. Within this volume we cut a slice of $\pm$ 50 pc around the $Z_c$ coordinate of Praesepe and a primary window of 7 km s$^{-1}$ by 7 km s$^{-1}$ in velocity space around the expected tangential velocity of Praesepe members predicted by the convergent point method. We developed a method to find over-densities in a 5-D phase space with 3 spatial coordinates and 2 in velocity space. By modelling the Galactic background distribution via a Poisson distribution we were able to distinguish between over-densities and background noise. The over-density related to Praesepe is populated by 1393 stars. The background contamination in 5-D phase space amounts to 47 stars, and the estimate of contamination due to unknown discordant radial velocities gives about the same number. So the total contamination is estimated to lie between 50 and 100 stars or between 3.6 and 7.2% of the Praesepe Sample. We used a 3-D ($M_G$, $M_{G_{BP}}, M_{G_{RP}}$) mass-luminosity relation to obtain individual masses for the stars from the PrS and found a total mass of 794 M$_\odot$. We determined the tidal radius ${r_t}$ of 10.77 pc and a tidal mass ${M_t}$ = 483 M$_\odot$ for Praesepe that is a lower limit since our sample gets incomplete at masses below 0.25 M$_\odot$. The half-mass radius ${r_h}$ is 4.8 pc. The shape of the cluster can be represented by a Plummer model with core radius ${r_{co}}$ = 3.7 pc. We observe a clear mass segregation of the cluster inside the tidal radius. The average mass per star drops from 1.17 $\rm{M}_{\odot}$ in the centre to about 0.46 $\rm{M}_{\odot}$ at the distance of the tidal radius ${r_t}$. We found 725 cluster members within one tidal radius; and a total of 1004 within two tidal radii. The remaining 389 stars from the Praesepe Sample populate two tidal tails. The tails extend up to about 165 pc on both sides along the $Y$-axis of Galactic rotation at the position of Praesepe.
The locations of the tidal tails of Praesepe in 3-D space are in very good coincidence with a theoretical model for tidal tails tailored to fit the 650 Myr old Hyades cluster. In detail, however, the ratio between the number of stars inside two tidal radii and outside (in the preceding tail) is much higher for Praesepe (factor 5) than for the Hyades (factor 2). This is a hint that the Hyades may already be in a later stage of evolution than is Praesepe.
We found 13 white dwarfs in our Praesepe Sample, 9 within one tidal radius and 4 outside. Three white dwarfs in the tail have not been assigned to Praesepe before, but this paper now identifies them as of Praesepe origin. We publish the data of all 1393 stars of the Praesepe Sample as on-line material.
We thank the referee, Carme Jordi, for her very constructive and useful comments that helped to improve the paper. This study was supported by Sonderforschungsbereich SFB 881 “The Milky Way System" (subproject B5) of the German *Deutsche Forschungsgemeinschaft, DFG*. This research has made use of the SIMBAD database and the VizieR catalogue access tool, operated at CDS, Strasbourg, France. 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.
[^1]: `https://www.cosmos.esa.int/web/gaia/``public-dpac-documents`
[^2]: `http://dc.zah.uni-heidelberg.de/tap`
[^3]: Aladin Desktop is a product of CDS, Strasbourg, France
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'This paper reviews the results obtained so far with the newly developed tomographic technique, which probes the velocity field across the atmosphere of long-period variable (LPV) stars. The method cross-correlates the optical spectrum with numerical masks constructed from synthetic spectra and probing layers of increasing depths. This technique reveals that the line doubling often observed in LPV stars around maximum light is the signature of the shock wave propagating in the atmosphere of these pulsating stars.'
author:
- Alain Jorissen
- Maya Dedecker
- Bertrand Plez
- Denis Gillet
- André Fokin
title: 'Tomography of the atmosphere of long-period variable stars'
---
Introduction
============
It is known since long that the brightness variations of long-period variable stars (LPVs) go along with spectral changes such as the doubling of the absorption lines around maximum light [@adams41; @alvarez00; @maehara68]. This line-doubling phenomenon appears to be shared by all LPVs [@alvarez01b], and occurs when the shock wave associated with the envelope pulsation propagates through the line-forming region [@alvarez00; @schwarzschild52]. Since it is almost impossible to study individual line profiles in the very crowded optical spectra of LPVs, it is necessary to resort to a technique that correlates the observed spectrum with a spectral template [@alvarez01a; @baranne79; @queloz95]. The shape of the resulting cross-correlation function (CCF) thus represents the average shape of those lines in the observed spectrum that match the template. By carefully designing spectral templates collecting all spectral lines forming at given depths in the atmosphere, it is even possible to probe the velocity field in pulsating variables. This is the general principle guiding the tomographic[^1] method described in this review. When applied to a temporal sequence of spectra, the method is even able to reveal in a spectacular way the [**outward propagation of the shock wave**]{}, as shown in Sect. \[Sect:RTCyg\] and Figs. \[Fig:RTCyg\]–\[Fig:RYCep\]!
Tomography of the atmosphere {#Sect:tomo}
============================
The study of the velocity field in the atmosphere of LPVs poses a special challenge, as the spectrum of these stars is extremely crowded, particularly in the optical domain. The cross-correlation technique provides a powerful tool to overcome this difficulty. The information relating to the line doubling (velocity shift and line shape) is in fact distributed among a large number of spectral lines, and can be summed up into an average profile, or more precisely into a cross-correlation function (CCF). If the correlation of the stellar spectrum with a mask involves many lines, it is possible to extract the relevant information from very crowded and/or low signal-to-noise spectra. The CCF writes $${\rm CCF}(\Delta\lambda) = \int_{\lambda_1}^{\lambda_2} s(\lambda - \Delta\lambda)\;
m(\lambda)\; {\rm d}\lambda,$$ where $s(\lambda)$ is the observed spectrum, $m(\lambda)$ is the template (a binary template has been adopted, being 0 around spectral lines of interest, and 1 elsewhere), and $\lambda_1$ and $\lambda_2$ are the boundaries of the spectral range covered by the observed spectrum. The radial velocity $V_r$ is then obtained from the wavelength shift $\Delta\lambda$(min CCF) where the CCF is minimum: $$V_r = \frac{\Delta\lambda({\rm min \; CCF}) } {0.5\; (\lambda_1 + \lambda_2)} \;c,$$ where $c$ is the speed of light in vacuum. We refer the reader to reference [@baranne79] for a detailed description of the CCF mathematical properties.
The tomographic method rests on our ability to construct reliable synthetic spectra of late-type giant stars [@plez92a; @plez92b; @plez99], and from those, to identify the depth of formation of any given spectral line. Rigorously, the contribution function (CF) to the flux depression [@albrow96] should be used to evaluate the geometrical depth at which a line forms. However, it would be a formidable task to compute the CF for each line appearing in the optical spectrum of LPV stars. For the sake of simplicity, the ‘depth function’ $x = x(\lambda)$ is used instead, which provides the geometrical depth corresponding to monochromatic optical depth $\tau_\lambda = 2/3$ at the considered wavelength $\lambda$. This function expresses the depth from which the emergent flux arises, in the Eddington-Barbier approximation, and should not differ much from the average depth of formation for sufficiently strong lines [@magain86].
Different masks $C_i$ are then constructed from the collection of $N_i$ lines $\lambda_{i,j} (1 \le j \le N_i)$ such that $x_i \le x(\lambda_{i,j}) < x_{i+1} = x_i + \Delta x$, where $\Delta x$ is some constant optimized to keep enough lines in any given mask without losing too much resolution in terms of geometrical depth. A natural limit to the resolution that can be achieved on the geometrical depth is provided by the width of the CF [@alvarez01a], and according to the Nyquist-Shannon theorem of elementary signal theory, there is no advantage in taking $\Delta x$ smaller than half the CF width of typical lines probed by the mask. Each mask $C_i$ should then probe lines forming at (geometrical) depths in the range $x_i, x_i + \Delta x$ in the atmosphere. These masks are used as templates to correlate with the observed spectra of the LPV stars. This procedure should provide the velocity field as a function of depth in the atmosphere (Fig. \[Fig:ZOph2\]), whereas for static atmospheres, all CCFs should yield the same radial velocity (Fig. \[Fig:constant\]).
A more detailed description of the method can be found in [@alvarez01a], while specific tomographic masks are available at [*http://www-astro.ulb.ac.be/Html/home.html\#tomography*]{}. The set of tomographic masks used in this paper was constructed from a synthetic spectrum at $T_{\rm eff}$ = 3500 K and $\log g = 0.9$ (see reference [@alvarez01a] for details). The properties of the masks (in terms of numbers of lines and depths probed) are listed in Table \[Tab:masks\]. The depth $x$ is in fact expressed in terms of a reference optical depth $x \equiv \log \tau_0$, where $\tau_0$ is the optical depth at the reference wavelength of 1.2 $\mu$m. There is a one-to-one correspondence between $\log \tau_0$ and the geometrical depth.
[lccc]{} Mask \# & depth probed &\
& $x \equiv \log \tau_{1.2 \mu m}$\
$C_i$ & $x_i, x_i + \Delta x$ & $N_i$\
$C_1$ (innermost) & $-2.00, -2.75$ & 777\
$C_2$ & $-2.75, -3.50$ & 610\
$C_3$ & $-3.50, -4.25$ & 433\
$C_4$ & $-4.25, -5.00$ & 321\
$C_5$ & $-5.00, -5.75$ & 168\
$C_6$ & $-5.75, -6.50$ & 167\
$C_7$ & $-6.50, -7.25$ & 94\
$C_8$ (outermost) & $-7.25, -8.00$ & 46\
\[Tab:masks\]
![ The CCFs of the Mira Z Oph at phase 0.08 obtained with the tomographic masks. Note how the shape of the CCFs evolve from the innermost layer (involving ascending matter only, hence C1 exhibits a single blue peak) to the outermost layer (involving mostly matter falling in, hence C8 exibits predominantly a red peak). This spatial sequence of line doubling reflects the presence of a shock wave in the line-forming region, with the shock front being centered on the layer probed by the mask C5[]{data-label="Fig:ZOph2"}](ZOph.ps){width="100.00000%"}
![Comparison of the $\lambda$ 6358.69 FeI line as seen in the spectrum of Z Oph at phases 0.08, 0.17, 1.04 and 1.21 (bottom row, from left to right), with the CCF obtained with the tomographic mask C5 (upper row), to which belongs the $\lambda$ 6358.69 FeI line. Z Oph is warm enough at maximum light (K3) for its spectrum to be not too crowded so that clean, almost unblended, lines as the one displayed here may be isolated[]{data-label="Fig:ZOph"}](ZOph_lines.ps){width="100.00000%"}
![Same as Fig. \[Fig:ZOph2\] for the non-Mira star $\mu$ Gem. Note how the CCFs broaden and become more contrasted in the upper atmospheric layers (C8 is the outermost mask), reflecting the fact that the strongest spectral lines form in the outermost layers. The radial velocity (in km/s) derived from each CCF is given in each panel. []{data-label="Fig:constant"}](mugem.ps){width="100.00000%"}
![The Schwarzschild scenario: temporal sequence followed by the intensity of the red and blue components of absorption lines close to maximum light, when the shock wave propagates through the photosphere[]{data-label="Fig:Schwarzschild"}](schwarz.eps){height="6cm"}
\[Fig:schwarz\]
An [*a posteriori*]{} validation of the method is provided by the fact that (i) the CCFs obtained with the tomographic masks for the non-pulsating M giant $\mu$ Gem (Lb variable with a variability range of only 0.11 mag in the Hipparcos catalogue) are all single-lined and yield almost identical radial velocities, as expected (Fig. \[Fig:constant\]), and (ii) the asymmetric CCFs obtained for the warm Mira variable Z Oph reflect the shape of the individual lines probed by any given mask (Fig. \[Fig:ZOph\]). The asymmetric, and sometime even double-peaked, CCFs observed for Mira variables thus correctly represent the underlying spectral lines, and are not artifacts of the correlation process. Moreover, around maximum light, the shape of CCFs in the successive tomographic masks follows a definite sequence: the CCF in the deepest layer exhibits a single blue component whereas a single red component is observed in the outermost CCF, and double-peaked CCFs are observed in between (Fig. \[Fig:ZOph2\]). This sequence, which is observed in almost all Mira around maximum light, is consistent with the scenario describing the effects of the passage of a shock wave through the atmosphere of a pulsating star (Sect. \[Sect:schw\]).
The scenario {#Sect:schw}
=============
The passage of a shock wave through a stellar atmosphere is expected to give rise to a very specific temporal evolution of the spectral line shape (Fig. \[Fig:schwarz\]), as first noted by Schwarzschild [@schwarzschild52] in relation with W Vir Cepheids. When the shock wave penetrates the line-forming region, the velocity discontinuity associated with the shock gives rise to the doubling of the spectral line: the red component of the line is formed in the matter falling in above the shock, whereas the blue component arises from the ascending matter lying behind the shock. The intensity of the blue and red components of a double line observed around maximum light in a LPV star should follow the temporal sequence illustrated in Fig. \[Fig:Schwarzschild\]. This is indeed the case, as revealed by the evolution of the CCFs of the Mira variables RT Cyg and RY Cep around maximum light (Sect. \[Sect:RTCyg\]). But the tomographic method described in Sect. \[Sect:tomo\] allows us to go one step further, by showing that the Schwarzschild scenario also reveals itself in terms of a definite [*spatial*]{} sequence of profiles (Fig. \[Fig:ZOph2\]). By combining those spatial sequences obtained at successive phases in the light cycle, the tomographic method reveals that [*line-doubling appears later in upper layers of the Mira atmosphere, thus translating the upward motion of the shock wave*]{} (Sect. \[Sect:RTCyg\]).
Application to the Mira variables RT Cyg and RY Cep {#Sect:RTCyg}
===================================================
A long-term monitoring of the Mira stars RTCyg ($P = 190$ d; $6.0 \le V
\le 13.1$) and RY Cep ($P = 145$ d; $8.6 \le V
\le 13.6$) has been performed with the fibre-fed echelle spectrograph ELODIE [@baranne96]. The spectrograph ELODIE is mounted on the 1.93-m telescope of the Observatoire de Haute Provence (France), and covers the full range from 3906 Å to 6811 Å in one exposure at a resolving power of 42000.
For RT Cyg, a monitoring in August-September 1999 covered phases $-0.20$ to 0.16 around maximum light, with 32 spectra obtained during this phase range (corresponding to an average resolution of $\Delta\phi = 0.01$; Fig. \[Fig:RTCyg\]). For RY Cep, the monitoring covered a [*full*]{} light cycle extending from August 2001 (phase -0.17) to February 2002 (phase 0.92), and 40 spectra were obtained (Fig. \[Fig:RYCep\]).
The sequences of CCFs presented in Figs. \[Fig:RTCyg\] and \[Fig:RYCep\] clearly obey the temporal evolution, with a single red component transforming progressively into a single blue component around maximum light. There are moreover clear phase lags between the different layers, this transformation occurring at later phases in outer layers. Thus, the scenario holds for Mira variables, and this conclusion definitely points towards the velocity stratification associated with the shock wave as the cause of the double absorption lines observed in Mira variables, as opposed to complex radiative processes (e.g., radiative release of thermal energy into the post-shock layer or temperature inversion) [@gillet85; @karp75].
C3 C4 C5 C7 phase
![Top panel: Sequence of cross-correlation profiles of RTCyg obtained with the tomographic masks C3, C4, C5 and C7 (from left to right) in August-September 1999, around maximum light. The labels beside each CCF refer to the corresponding visual phase based on the AAVSO ephemeris [@mattei99]\
Bottom panels: Radial velocity curves obtained with the tomographic masks C3 (leftmost panel - solid line - triangles), C4 (middle panel - long-dashed line - squares), and C7 (rightmost panel - short-dashed line - circles). The C4 radial-velocity curve (long-dashed line) has been duplicated in the left- and rightmost panels to allow an easy intercomparison []{data-label="Fig:RTCyg"}](suiviRTCyg_C3C4C5C7.ps "fig:"){height="0.7\textheight"} ![Top panel: Sequence of cross-correlation profiles of RTCyg obtained with the tomographic masks C3, C4, C5 and C7 (from left to right) in August-September 1999, around maximum light. The labels beside each CCF refer to the corresponding visual phase based on the AAVSO ephemeris [@mattei99]\
Bottom panels: Radial velocity curves obtained with the tomographic masks C3 (leftmost panel - solid line - triangles), C4 (middle panel - long-dashed line - squares), and C7 (rightmost panel - short-dashed line - circles). The C4 radial-velocity curve (long-dashed line) has been duplicated in the left- and rightmost panels to allow an easy intercomparison []{data-label="Fig:RTCyg"}](vrRTCyg.eps "fig:"){height="0.2\textheight" width="80.00000%"}
C3 C4 C5 C6 phase
![Top panel: Same as Fig. \[Fig:RTCyg\] for RY Cep with the tomographic masks C3 (inner), C4, C5 and C6 (outer), from left to right. The labels beside each CCF refer to the corresponding visual phase based on the AFOEV monitoring.\
Bottom panels: Radial velocity curves obtained with the tomographic masks C3 (leftmost panel - solid line - triangles), C4 (middle panel - long-dashed line - squares), and C6 (rightmost panel - short-dashed line - circles). The C4 radial-velocity curve (long-dashed line) has been duplicated in the left- and rightmost panels to allow an easy intercomparison []{data-label="Fig:RYCep"}](suiviRYCep.ps "fig:"){height="0.7\textheight"} ![Top panel: Same as Fig. \[Fig:RTCyg\] for RY Cep with the tomographic masks C3 (inner), C4, C5 and C6 (outer), from left to right. The labels beside each CCF refer to the corresponding visual phase based on the AFOEV monitoring.\
Bottom panels: Radial velocity curves obtained with the tomographic masks C3 (leftmost panel - solid line - triangles), C4 (middle panel - long-dashed line - squares), and C6 (rightmost panel - short-dashed line - circles). The C4 radial-velocity curve (long-dashed line) has been duplicated in the left- and rightmost panels to allow an easy intercomparison []{data-label="Fig:RYCep"}](vrRYCep.eps "fig:"){height="0.2\textheight" width="80.00000%"}
Radial velocities have been extracted from the CCFs of Figs. \[Fig:RTCyg\] and \[Fig:RYCep\] (top panels) to yield the curves displayed in the bottom panels of the same figures. We defer the detailed comparison between the radial-velocity curves of RY Cep and RT Cyg, and their interpretation, to a forthcoming paper. We just stress here noteworthy features on the radial-velocity curve of RY Cep: (i) the red peaks (corresponding to matter falling in) are observed at the same velocity in all three masks; they disappear at increasingly later phases as one considers layers closer and closer to the surface, since the outermost layers are the last ones to be penetrated by the shock wave which suppresses the infalling component (i.e., red peak); (ii) the maximum outward velocity is the same for the three masks ($\sim -247$ km/s), and is reached around phase $-0.1$. After that, matter will decelerate in the innermost layers first, as they lose the impetus provided by the shock wave which is moving away.
Conclusion and perspectives
===========================
The tomographic technique presented in this paper opens new perspectives for the study of the dynamics of LPV stars, as it allows to visualize the outward motion of the shock wave in the atmosphere. At this stage, however, the results remain qualitative as it is difficult to quantify the geometrical depth of the layers probed by the different masks. It is hoped that, in the future, this method may be combined with interferometric imaging in the different masks to derive the corresponding radii. This would give direct access to the velocity of the shock, for instance.\
[. We thank the Observing Program Committee of the Observatoire de Haute-Provence for a generous allocation of telescope time. The AFOEV is warmly thanked for positively answering our request to monitor RY Cep during the time of the spectroscopic monitoring. A.J. is Research Associate from the [*Fonds National de la Recherche Scientifique*]{} (Belgium). Financial support has been received from the [*Communauté Française de Belgique*]{} in the framework of a TOURNESOL programme.]{}
[99.]{}
W.S. Adams, 1941, ApJ, 93, 11
M. D. Albrow, P. L. Cottrell, 1996, MNRAS, 278, 337
R. Alvarez, A. Jorissen, B. Plez, D. Gillet, A. Fokin, 2000 A&A, 362, 655
R. Alvarez, A. Jorissen, B. Plez, D. Gillet, A. Fokin, M. Dedecker, 2001a A&A, 379, 288
R. Alvarez, A. Jorissen, B. Plez, D. Gillet, A. Fokin, M. Dedecker, 2001b A&A, 379, 305
A. Baranne, M. Mayor, J.L. Poncet, 1979, Vistas in Astronomy, 23, 279
A. Baranne, D. Queloz, M. Mayor et al., 1996, A&AS, 119, 373
D. Gillet, E. Maurice, P. Bouchet, R. Ferlet, 1985, A&A, 148, 155
A. H. Karp, 1975, ApJ, 201, 641
H. Maehara, 1968, PASJ, 20, 77
P. Magain, 1986, A&A, 163, 135
J. A. Mattei, 1999, private communication
B. Plez, 1992, A&AS, 94, 527
B. Plez, J.M. Brett, Å. Nordlund, 1992, A&A, 256, 551
B. Plez, 1999, in IAU Symp. 191, Asymptotic Giant Branch Stars, ed. T. Le Bertre, A. Lèbre, C. Waelkens, (Astron. Soc. Pacific), 75
D. Queloz, 1995 in IAU Symp. 167, New Developments in Array Technology and Applications, ed. A.G. Davis Philips, K. A. Janes & A. R. Upgren (Dordrecht: Kluwer), 221
M. Schwarzschild, 1952, in Transactions of the IAU, Vol. VIII, ed. P.T. Oosterhoff (Cambridge University Press), 811
[^1]: The word [*tomography*]{} is used here in its etymological sense ([*‘display cuts’*]{}), which differs somewhat from the broader sense in use within the astronomical community (reconstruction of a structure using projections taken under different angles).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present an illative system ${\ensuremath{{\cal I}}}_s$ of classical higher-order logic with subtyping and basic inductive types. The system ${\ensuremath{{\cal I}}}_s$ allows for direct definitions of partial and general recursive functions, and provides means for handling functions whose termination has not been proven. We give examples of how properties of some recursive functions may be established in our system. In a technical appendix to the paper we prove consistency of ${\ensuremath{{\cal I}}}_s$. The proof is by model construction. We then use this construction to show conservativity of ${\ensuremath{{\cal I}}}_s$ over classical first-order logic. Conservativity over higher-order logic is conjectured, but not proven.'
author:
- Łukasz Czajka
bibliography:
- 'biblio.bib'
title: 'Partiality and Recursion in Higher-Order Logic'
---
[**Note:**]{} This paper is an extended technical report based on a conference paper with the same title published by Springer-Verlag in the proceedings of FoSSaCS 2013. The final publication is available at .
Table of Contents {#table-of-contents .unnumbered}
=================
starttoc[toc]{}
Introduction
============
We present an illative $\lambda$-calculus system ${\ensuremath{{\cal I}}}_s$ of classical higher-order logic with subtyping and basic inductive types. Being illative means that the system is a combination of higher-order logic with the *untyped* $\lambda$-calculus. It therefore allows for unrestricted recursive definitions directly, including definitions of possibly non-terminating partial functions. We believe that this feature of ${\ensuremath{{\cal I}}}_s$ makes it potentially interesting as a logic for an interactive theorem prover intended to be used for program verification.
Most popular proof assistants allow only total functions, and totality must be ensured by the user, either by very precise specifications of function domains, restricting recursion in a way that guarantees termination, explicit well-foundedness proofs, or other means.
Obviously, there is a reason why most proof assistants do not handle partial functions directly. This is to ensure consistency of the system. Combining an expressive higher-order logic with unrestricted recursion is a non-trivial problem.
There are various indirect ways of dealing with general recursion in popular theorem provers based on total logics. There are also many non-standard logics allowing partial functions directly. We discuss some related work in Sect. \[sec\_related\].
In Sect. \[sec\_I\_s\] we introduce the system ${\ensuremath{{\cal I}}}_s$. Our approach builds on the old tradition of illative combinatory logic [@illat01; @Seldin2009; @Czajka2013Accepted; @Czajka2011]. This tradition dates back to early inconsistent systems of Shönfinkel, Church and Curry proposed in the 1920s and the 1930s [@Seldin2009]. However, after the discovery of paradoxes most logicians abandoned this approach. A notable exception was Haskell Curry and his school, but not much progress was made in establishing consistency of illative systems strong enough to interpret traditional logic. Only in the 1990s some first-order illative systems were shown consistent and complete for traditional first-order logic [@illat01; @illat02]. The system ${\ensuremath{{\cal I}}}_s$, in terms of the features it provides, may be considered an extension of the illative system ${\ensuremath{{\cal I}}}_\omega$ from [@Czajka2013Accepted]. We briefly discuss the relationship between ${\ensuremath{{\cal I}}}_s$ and ${\ensuremath{{\cal I}}}_\omega$ in Sect. \[sec\_related\].
Because ${\ensuremath{{\cal I}}}_s$ is based on the untyped $\lambda$-calculus, its consistency is obviously open to doubt. In an appendix we give a proof by model construction of consistency of ${\ensuremath{{\cal I}}}_s$. Unfortunately, the proof is too long to fit within the page limits of a conference paper. In Sect. \[sec\_consistent\] we give a general overview of the proof. The model construction is similar to the one from [@Czajka2013Accepted] for the traditional illative system ${\ensuremath{{\cal I}}}_\omega$. It is extended and adapted in a non-trivial way to account for additional features of ${\ensuremath{{\cal I}}}_s$. To our knowlege ${\ensuremath{{\cal I}}}_s$ is the first higher-order illative system featuring subtypes and some form of induction, for which there is a consistency proof.
In Sect. \[sec\_partiality\] we provide examples of proofs in ${\ensuremath{{\cal I}}}_s$ indicating possible applications of our approach to the problem of dealing with partiality, non-termination and general recursion in higher-order logic. We are mainly interested in partiality arising from non-termination of non-well-founded recursive definitions.
For lack of space we omit proofs of most of the lemmas and theorems we state. The proofs of non-trivial results may be found in technical appendices to this paper.
The Illative System {#sec_I_s}
===================
In this section we present the system ${\ensuremath{{\cal I}}}_s$ of illative classical higher-order logic with subtyping and derive some of its basic properties.
\[def\_I\_s\]
The system ${\ensuremath{{\cal I}}}_s$ consists of the following.
- A countably infinite set of variables $V_s = \{x, y, z, \ldots
\}$ and a set of constants $\Sigma_{s}$.
- The set of sorts ${\ensuremath{{\cal S}}}= \{{\ensuremath{\mathrm{Type}}}, {\ensuremath{\mathrm{Prop}}}\}$.
- The set of *basic inductive types* ${\ensuremath{{\cal T}}}_I$ is defined inductively by the rule: if $\iota_{1,1},\ldots,\iota_{1,n_1},\ldots,\iota_{m,1},\ldots,\iota_{m,n_m}
\in {\ensuremath{{\cal T}}}_I \cup \{\star\}$ then $$\mu(\langle \iota_{1,1},\ldots,\iota_{1,n_1} \rangle, \ldots,
\langle \iota_{m,1},\ldots,\iota_{m,n_m} \rangle) \in {\ensuremath{{\cal T}}}_I$$ where $m \in {\ensuremath{\mathbb{N}}}_+$ and $n_1,\ldots,n_m \in {\ensuremath{\mathbb{N}}}$.
- We define the sets of *constructors* ${\ensuremath{{\cal C}}}$, *destructors* ${\ensuremath{{\cal D}}}$, and *tests* ${\ensuremath{{\cal O}}}$ as follows. For each $\iota \in {\ensuremath{{\cal T}}}_I$ of the form $$\iota = \mu(\langle \iota_{1,1},\ldots,\iota_{1,n_1} \rangle,
\ldots, \langle \iota_{m,1},\ldots,\iota_{m,n_m} \rangle) \in
{\ensuremath{{\cal T}}}_I$$ where $\iota_{i,j} \in {\ensuremath{{\cal T}}}_I \cup \{\star\}$, the set ${\ensuremath{{\cal C}}}$ contains $m$ distinct constants $c_1^\iota,\ldots,c_m^\iota$. The number $n_i$ is called the *arity* of $c_i^\iota$, and $\langle \iota_{i,1},\ldots,\iota_{i,n_i} \rangle$ is its *signature*. With each $c_i^\iota \in {\ensuremath{{\cal C}}}$ of arity $n_i$ we associate $n_i$ distinct destructors $d_{i,1}^\iota,\ldots,d_{i,n_i}^\iota \in {\ensuremath{{\cal D}}}$ and one test $o_i^\iota \in {\ensuremath{{\cal O}}}$. When we use the symbols $c_i^\iota$, $o_i^\iota$ and $d_{i,j}^\iota$ we implicitly assume that they denote the constructors, tests and destructors associated with $\iota$. When it is clear from the context which type $\iota$ is meant, we use the notation $\iota_{i,j}^*$ for $\iota_{i,j}$ if $\iota_{i,j} \ne \star$, or for $\iota$ if $\iota_{i,j} = \star$.
- The set of *${\ensuremath{{\cal I}}}_s$-terms* ${\ensuremath{\mathbb T}}$ is defined by the following grammar. $$\begin{aligned}
{\ensuremath{\mathbb T}}&::=& V_s \;|\; \Sigma_{s} \;|\; {\ensuremath{{\cal S}}}\;|\; {\ensuremath{{\cal C}}}\;|\; {\ensuremath{{\cal D}}}\;|\; {\ensuremath{{\cal O}}}\;|\; {\ensuremath{{\cal T}}}_I \;|\; \lambda V_s \,.\, {\ensuremath{\mathbb T}}\;|\; ({\ensuremath{\mathbb T}}{\ensuremath{\mathbb T}})
\;|\; \mathrm{Is} \;|\; \mathrm{Subtype} \;|\; \mathrm{Fun}
\;|\; \\ && \forall \;|\; \vee \;|\; \bot \;|\; \epsilon \;|\;
\mathrm{Eq} \;|\; \mathrm{Cond}
\end{aligned}$$ We assume application associates to the left and omit spurious brackets.
- We identify $\alpha$-equivalent terms, i.e. terms differing only in the names of bound variables are considered identical. We use the symbol $\equiv$ for identity of terms up to $\alpha$-equivalence. We also assume without loss of generality that all bound variables in a term are distinct from the free variables, unless indicated otherwise.[^1]
- In what follows we use the abbreviations: $$\begin{aligned}
t_1 : t_2 &\equiv& \mathrm{Is}\, t_1\, t_2 \\
{\ensuremath{\{ x : \alpha \;|\; \varphi \}}} &\equiv& \mathrm{Subtype}\, \alpha\,
\lambda x \,.\, \varphi \\
\alpha \to \beta &\equiv& \mathrm{Fun}\, \alpha\, \beta \\
\forall x : \alpha \,.\, \varphi &\equiv& \forall\, \alpha\, \lambda x
\,.\, \varphi \\
\forall x_1,\ldots,x_n : \alpha \,.\, \varphi &\equiv& \forall x_1 :
\alpha \,.\, \ldots \forall x_n : \alpha \,.\, \varphi \\
\varphi \supset \psi &\equiv& \forall x : {\ensuremath{\{ y : {\ensuremath{\mathrm{Prop}}}\;|\; \varphi \}}}
\,.\, \psi \;\; \mathrm{where\ } x, y \notin FV(\varphi, \psi)
\\
\neg \varphi &\equiv& \varphi \supset \bot \\
\top &\equiv& \bot \supset \bot \\
\varphi \vee \psi &\equiv& \vee \varphi \psi \\
\varphi \wedge \psi &\equiv& \neg (\neg \varphi \vee \neg \psi) \\
\exists x : \alpha \,.\, \varphi &\equiv& \neg \forall x : \alpha
\,.\, \neg \varphi
\end{aligned}$$ We assume that $\neg$ has the highest precedence.
- The system ${\ensuremath{{\cal I}}}_s$ is given by the following rules and axioms, where $\Gamma$ is a finite set of terms, $t, \varphi, \psi,
\alpha, \beta$, etc. are arbitrary terms. The notation $\Gamma,
\varphi$ is a shorthand for $\Gamma \cup \{\varphi\}$. We use Greek letters $\varphi$, $\psi$, etc. to highlight that a term is to be intuitively interpreted as a proposition, and we use $\alpha$, $\beta$, etc. when it is to be interpreted as a type, but there is no a priori syntactic distinction. All judgements have the form $\Gamma {\ensuremath{\vdash}}t$ where $\Gamma$ is a set of terms and $t$ a term. In particular, $\Gamma {\ensuremath{\vdash}}t : \alpha$ is a shorthand for $\Gamma {\ensuremath{\vdash}}\mathrm{Is}\, t\, \alpha$, according to the convention from the previous point.
[**Axioms**]{}
- $\Gamma, \varphi {\ensuremath{\vdash}}\varphi$
- $\Gamma {\ensuremath{\vdash}}{\ensuremath{\mathrm{Eq}\,t\,t\,}}$
- $\Gamma {\ensuremath{\vdash}}{\ensuremath{\mathrm{Prop}}}: {\ensuremath{\mathrm{Type}}}$
- $\Gamma {\ensuremath{\vdash}}\iota : {\ensuremath{\mathrm{Type}}}$ for $\iota \in {\ensuremath{{\cal T}}}_I$
- $\Gamma {\ensuremath{\vdash}}o_i^\iota (c_i^\iota t_1 \ldots
t_{n_i})$ if $c_i^\iota \in {\ensuremath{{\cal C}}}$ has arity $n_i$
- $\Gamma {\ensuremath{\vdash}}\neg (o_i^\iota (c_j^\iota t_1 \ldots
t_{n_j}))$ if $i \ne j$ and $c_j^\iota \in {\ensuremath{{\cal C}}}$ has arity $n_j$
- $\Gamma {\ensuremath{\vdash}}{\ensuremath{\mathrm{Eq}\,(d_{i,k}^\iota (c_i^\iota t_1 \ldots
t_{n_i}))\,t_k\,}}$ for $k=1,\ldots,n_i$, if $c_i^\iota \in {\ensuremath{{\cal C}}}$ has arity $n_i$
- $\Gamma {\ensuremath{\vdash}}\bot : {\ensuremath{\mathrm{Prop}}}$
- $\Gamma {\ensuremath{\vdash}}\forall p : {\ensuremath{\mathrm{Prop}}}\,.\, p \lor \neg p$
- $\Gamma {\ensuremath{\vdash}}{\ensuremath{\mathrm{Eq}\,((\lambda x \,.\, t_1)
t_2)\,(t_1[x/t_2])\,}}$
[**Rules**]{}
[cc]{}\
&\
\
&\
\
&\
\
&\
\
&\
$
{\vee_{i1}:}\; \inferrule{\Gamma {\ensuremath{\vdash}}\varphi}
{\Gamma {\ensuremath{\vdash}}\varphi \vee \psi}
$ & $
{\vee_{i2}:}\; \inferrule{\Gamma {\ensuremath{\vdash}}\psi}
{\Gamma {\ensuremath{\vdash}}\varphi \vee \psi}
$\
&\
\
&\
\
&\
$
{\wedge_{e1}:}\;
\inferrule{
\Gamma {\ensuremath{\vdash}}\varphi \wedge \psi
}{
\Gamma {\ensuremath{\vdash}}\varphi
}
$ & $
{\wedge_{e2}:}\;
\inferrule{
\Gamma {\ensuremath{\vdash}}\varphi \wedge \psi
}{
\Gamma {\ensuremath{\vdash}}\psi
}
$\
&\
$
{\supset_{t2}:}\;
\inferrule{
\Gamma {\ensuremath{\vdash}}(\varphi \supset \psi) : {\ensuremath{\mathrm{Prop}}}}{
\Gamma {\ensuremath{\vdash}}\varphi : {\ensuremath{\mathrm{Prop}}}}
$ & $
{\bot_{e}:}\;
\inferrule{
\Gamma {\ensuremath{\vdash}}\bot
}{
\Gamma {\ensuremath{\vdash}}\varphi
}
$\
&\
\
&\
$
{\to_e:}\; \inferrule{\Gamma {\ensuremath{\vdash}}t_1 : \alpha \to \beta \\ \Gamma {\ensuremath{\vdash}}t_2 : \alpha}{\Gamma {\ensuremath{\vdash}}t_1
t_2 : \beta}
$ & $
{\to_t:}\; \inferrule{\Gamma {\ensuremath{\vdash}}\alpha : {\ensuremath{\mathrm{Type}}}\\ \Gamma
{\ensuremath{\vdash}}\beta : {\ensuremath{\mathrm{Type}}}}
{\Gamma {\ensuremath{\vdash}}(\alpha \to \beta) : {\ensuremath{\mathrm{Type}}}}
$\
&\
\
&\
$
{s_e:}\; \inferrule{\Gamma {\ensuremath{\vdash}}t : \{ x :
\alpha \;|\; \varphi \}}{\Gamma {\ensuremath{\vdash}}\varphi[x/t]}
$ & $
{s_{et}:}\; \inferrule{\Gamma {\ensuremath{\vdash}}t : \{ x :
\alpha \;|\; \varphi\}}{\Gamma {\ensuremath{\vdash}}t : \alpha}
$\
&\
\
&\
$
{\epsilon_i:}\; \inferrule{\Gamma {\ensuremath{\vdash}}\exists x : \alpha \,.\, \top}
{\Gamma {\ensuremath{\vdash}}(\epsilon \alpha) : \alpha}
$ & $
{p_i:}\; \inferrule{\Gamma {\ensuremath{\vdash}}\varphi}
{\Gamma {\ensuremath{\vdash}}\varphi : {\ensuremath{\mathrm{Prop}}}}
$\
&\
$
{c_1:}\; \inferrule{
\Gamma {\ensuremath{\vdash}}\varphi
}{
\Gamma {\ensuremath{\vdash}}{\ensuremath{\mathrm{Eq}\,({\ensuremath{\mathrm{Cond}\,\varphi\,t_1\,t_2\,}})\,t_1\,}}
}
$ & $
{c_2:}\; \inferrule{
\Gamma {\ensuremath{\vdash}}\neg\varphi
}{
\Gamma {\ensuremath{\vdash}}{\ensuremath{\mathrm{Eq}\,({\ensuremath{\mathrm{Cond}\,\varphi\,t_1\,t_2\,}})\,t_2\,}}
}
$\
&\
\
&\
\
&\
\
&\
$
{\mbox{eq}:}\; \inferrule{\Gamma {\ensuremath{\vdash}}\varphi \\ \Gamma {\ensuremath{\vdash}}{\ensuremath{\mathrm{Eq}\,\varphi\,\varphi'\,}}}{\Gamma {\ensuremath{\vdash}}\varphi'}
$ & $
{\mbox{eq-sym}:}\; \inferrule{\Gamma {\ensuremath{\vdash}}{\ensuremath{\mathrm{Eq}\,t_1\,t_2\,}}}{\Gamma {\ensuremath{\vdash}}{\ensuremath{\mathrm{Eq}\,t_2\,t_1\,}}}
$\
&\
\
&\
\
&\
$${i_i^\iota:}\; \inferrule{\Gamma, x_1 : \iota_{i,1}^*, \ldots, x_{n_i} : \iota_{i,n_i}^*, t
x_{j_{i,1}}, \ldots, t x_{j_{i,k_i}} {\ensuremath{\vdash}}t (c_i^\iota x_1 \ldots
x_{n_i}) \\ \mathrm{for\ } i=1,\ldots,m
}{
\Gamma {\ensuremath{\vdash}}\forall x : \iota \,.\, t x
}$$ where $x,x_1,\ldots,x_{n_i} \notin FV(\Gamma, t)$, $c_1^\iota,\ldots,c_m^\iota \in {\ensuremath{{\cal C}}}$ are all constructors associated with $\iota \in {\ensuremath{{\cal T}}}_I$, and $j_{i,1},\ldots,j_{i,k_i}$ is an increasing sequence of all indices $1 \le j \le n_i$ such that $\iota_{i,j} = \star$ $${i_t^{\iota,k}:}\; \inferrule{
\Gamma {\ensuremath{\vdash}}t_j : \iota_{k,j}^* \mathrm{\ for\ } j=1,\ldots,n_k
}{
\Gamma {\ensuremath{\vdash}}(c_k^\iota t_1 \ldots t_{n_k}) : \iota
}$$
For an arbitrary set of terms $\Gamma$, we write $\Gamma
{\ensuremath{\vdash}}_{{\ensuremath{{\cal I}}}_s} \varphi$ if there exists a finite subset $\Gamma'
\subseteq \Gamma$ such that $\Gamma' {\ensuremath{\vdash}}\varphi$ is derivable in the system ${\ensuremath{{\cal I}}}_s$. We drop the subscript when irrelevant or obvious from the context.
If $\Gamma {\ensuremath{\vdash}}\varphi$ then $\Gamma, \psi {\ensuremath{\vdash}}\varphi$.
\[lem\_I\_s\_subst\] If $\Gamma {\ensuremath{\vdash}}\varphi$ then $\Gamma[x/t] {\ensuremath{\vdash}}\varphi[x/t]$, where $\Gamma[x/t] = \{ \psi[x/t] \;|\; \psi \in \Gamma \}$.
Representing Logic
------------------
The inference rules of ${\ensuremath{{\cal I}}}_s$ may be intuitively justified by appealing to an informal many-valued semantics. A term $t$ may be true, false, or something entirely different (“undefined”, a program, a natural number, a type, …). By way of an example, we explain an informal meaning of some terms:
- $t : {\ensuremath{\mathrm{Prop}}}$ is true iff $t$ is true or false,
- $\alpha : {\ensuremath{\mathrm{Type}}}$ is true iff $\alpha$ is a type,
- $t : \alpha$ is true iff $t$ has type $\alpha$, assuming $\alpha$ is a type,
- $\forall x : \alpha . \varphi$ is true iff $\alpha$ is a type and for all $t$ of type $\alpha$, $\varphi[x/t]$ is true,
- $\forall x : \alpha . \varphi$ is false iff $\alpha$ is a type and there exists $t$ of type $\alpha$ such that $\varphi[x/t]$ is false,
- $t_1 \vee t_2$ is true iff $t_1$ is true or $t_2$ is true,
- $t_1 \vee t_2$ is false iff $t_1$ is false and $t_2$ is false,
- $t_1 \supset t_2$ is true iff $t_1$ is false or both $t_1$ and $t_2$ are true,
- $t_1 \supset t_2$ is false iff $t_1$ is true and $t_2$ is false,
- $\neg t$ is true iff $t$ is false,
- $\neg t$ is false iff $t$ is true.
Obviously, $\Gamma {\ensuremath{\vdash}}t$ is then (informally) interpreted as: for all possible substitution instances $\Gamma^*,t^*$ of $\Gamma,t$, [^2] if all terms in $\Gamma^*$ are true, then the term $t^*$ is also true.
Note that the logical connectives are “lazy”, e.g. for $t_1 \vee
t_2$ to be true it suffices that $t_1$ is true, but $t_2$ need not have a truth value at all – it may be something else: a program, a type, “undefined”, etc. This laziness allows us to omit many restrictions which would otherwise be needed in inference rules, and would thus make the system less similar to ordinary logic.
The following rules may be derived in ${\ensuremath{{\cal I}}}_s$.
------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------
$ $
\supset_i:\; \supset_e:\;
\inferrule{ \inferrule{
\Gamma {\ensuremath{\vdash}}\varphi : {\ensuremath{\mathrm{Prop}}}\\ \Gamma, \varphi {\ensuremath{\vdash}}\psi \Gamma {\ensuremath{\vdash}}\varphi \supset \psi \\ \Gamma {\ensuremath{\vdash}}\varphi
}{ }{
\Gamma {\ensuremath{\vdash}}\varphi \supset \psi \Gamma {\ensuremath{\vdash}}\psi
} }
$ $
$ $
\supset_t:\; {\wedge_{i}:}\;
\inferrule{ \inferrule{
\Gamma {\ensuremath{\vdash}}\varphi : {\ensuremath{\mathrm{Prop}}}\\ \Gamma, \varphi {\ensuremath{\vdash}}\psi : {\ensuremath{\mathrm{Prop}}}}{ \Gamma {\ensuremath{\vdash}}\varphi \\ \Gamma {\ensuremath{\vdash}}\psi
\Gamma {\ensuremath{\vdash}}(\varphi \supset \psi) : {\ensuremath{\mathrm{Prop}}}} }{
$ \Gamma {\ensuremath{\vdash}}\varphi \wedge \psi
}
$
------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------
Note that in general the elimination rules for $\wedge$ and the rules for $\exists$ cannot be derived from the rules for $\vee$ and $\forall$, because we would not be able to prove the premise $\varphi : {\ensuremath{\mathrm{Prop}}}$ when trying to apply the rule $\supset_i$. It is instructive to try to derive these rules and see where the proof breaks down.
In ${\ensuremath{{\cal I}}}_s$ the only non-standard restriction in the usual inference rules for logical connectives is the additional premise $\Gamma
{\ensuremath{\vdash}}\varphi : {\ensuremath{\mathrm{Prop}}}$ in the rule $\supset_i$. It is certainly unavoidable, as otherwise Curry’s paradox may be derived (see e.g. [@illat01; @Seldin2009]). However, we have standard classical higher-order logic if we restrict to terms of type ${\ensuremath{\mathrm{Prop}}}$, in the sense that the natural deduction rules then become identical to the rules of ordinary logic. This is made more precise in Sect. \[sec\_consistent\] where a sound translation from a traditional system of higher-order logic into ${\ensuremath{{\cal I}}}_s$ is described.
Note that we have the law of excluded middle only in the form $\forall
p : {\ensuremath{\mathrm{Prop}}}\,.\, p \lor \neg p$. Adding $\Gamma {\ensuremath{\vdash}}\varphi \vee
\neg\varphi$ as an axiom for an arbitrary term $\varphi$ gives an inconsistent system.[^3]
It is well-known (see e.g. [@SorensenUrzyczyn2006 Chapter 11]) that in higher-order logic all logical connectives may be defined from $\forall$ and $\supset$ as follows. $$\begin{aligned}
\bot &\equiv& \forall p : {\ensuremath{\mathrm{Prop}}}\,.\, p \\
\neg \varphi &\equiv& \varphi \supset \bot \\
\varphi \wedge \psi &\equiv& \forall p : {\ensuremath{\mathrm{Prop}}}\,.\, (\varphi \supset \psi \supset p) \supset p \\
\varphi \vee \psi &\equiv& \forall p : {\ensuremath{\mathrm{Prop}}}\,.\, (\varphi \supset p) \supset (\psi \supset p) \supset p \\
\exists x : \alpha \,.\, \varphi &\equiv& \forall p : {\ensuremath{\mathrm{Prop}}}\,.\, (\forall x : \alpha \,.\, \varphi \supset p) \supset p\end{aligned}$$ One may therefore wonder why we take $\vee$ and $\bot$ as primitive. The answer is that if we defined the connectives by the above equations, then the inference rules that could be derived for them would need to contain additional restrictions. For instance, we would be able to derive only the following variants of $\vee$-introduction.
------------------------------------------------------------------------------ ----------------------------------------------------------------------------------------------------------------------------------------
$ $
{\vee_{i1}':}\; \inferrule{\Gamma {\ensuremath{\vdash}}\varphi \\ \Gamma {\vee_{i2}':}\; \inferrule{\Gamma {\ensuremath{\vdash}}\psi \\ \Gamma {\ensuremath{\vdash}}\varphi : {\ensuremath{\mathrm{Prop}}}}
{\ensuremath{\vdash}}\psi : {\ensuremath{\mathrm{Prop}}}} {\Gamma {\ensuremath{\vdash}}\varphi \vee \psi}
{\Gamma {\ensuremath{\vdash}}\varphi \vee \psi} $
$
------------------------------------------------------------------------------ ----------------------------------------------------------------------------------------------------------------------------------------
Equality, Recursive Definitions and Extensionality
--------------------------------------------------
It is well-known (see e.g. [@Barendregt1984 Chapters 2, 6]) that since untyped $\lambda$-terms are available together with the axiom $\beta$ and usual rules for equality, any set of equations of the following form has a solution for $z_1,\ldots,z_n$, where the expressions $\Phi_i(z_1, \ldots, z_n, x_1, \ldots, x_m)$ are arbitrary terms with the free variables listed. $$\begin{aligned}
z_1 x_1 \ldots x_m &=& \Phi_1(z_1, \ldots, z_n, x_1, \ldots, x_m) \\
&\vdots& \\
z_n x_1 \ldots x_m &=& \Phi_n(z_1, \ldots, z_n, x_1, \ldots, x_m)\end{aligned}$$ In other words, for any such set of equations, there exist terms $t_1,
\ldots, t_n$ such that for any terms $s_1, \ldots, s_m$ we have ${\ensuremath{\vdash}}{\ensuremath{\mathrm{Eq}\,\left(t_i s_1 \ldots s_m\right)\,\left(\Phi_i(t_1,
\ldots, t_n, s_1, \ldots, s_m)\right)\,}}$ for each $i=1,\ldots,n$.
We will often define terms by such equations. In what follows we freely use the notation $t_1 = t_2$ for ${\ensuremath{\vdash}}{\ensuremath{\mathrm{Eq}\,t_1\,t_2\,}}$, or for $\Gamma {\ensuremath{\vdash}}{\ensuremath{\mathrm{Eq}\,t_1\,t_2\,}}$ when it is clear which context $\Gamma$ is meant. We use $t_1 = t_2 = \ldots = t_n$ to indicate that ${\ensuremath{\mathrm{Eq}\,t_i\,t_{i+1}\,}}$ may be derived for $i=1,\ldots,n-1$. We also sometimes write a term of the form ${\ensuremath{\mathrm{Eq}\,t_1\,t_2\,}}$ as $t_1 = t_2$.
It is worth stressing once again that there is no a priori syntactic distinction between terms, formulas, types, type assertions, etc. Formally, there are only terms, but some terms are intuitively interpreted as formulas, types, etc. In particular, the aforementioned method of defining terms by arbitrary recursive equations may be applied to define terms which could be intuitively considered to be formulas, e.g. we may define a term $\varphi$ such that $\varphi =
\neg \varphi$. Inconsistency is avoided, because it will not be actually possible to prove ${\ensuremath{\vdash}}_{{\ensuremath{{\cal I}}}_s} \varphi : {\ensuremath{\mathrm{Prop}}}$, hence $\varphi$ will not really be a formula. In ${\ensuremath{{\cal I}}}_s$ the inference rules serve the purpose of classifying terms into different categories. This classification is not enforced a priori, but instead it is a part of derivations in the logic.
In ${\ensuremath{{\cal I}}}_s$ there is no rule for typing the equality $\mathrm{Eq}$. One consequence is that ${\ensuremath{\vdash}}\neg({\ensuremath{\mathrm{Eq}\,t_1\,t_2\,}})$ cannot be derived for any terms $t_1,t_2$.[^4] For this reason $\mathrm{Eq}$ is more like a meta-level notion of equality.
Leibniz equality $\mathrm{Eql}$ is defined as: $$\mathrm{Eql} \equiv \lambda \alpha \lambda x \lambda y . \forall p :
\alpha \to {\ensuremath{\mathrm{Prop}}}\,.\, p x \supset p y$$
As with $=$, we will often write $t_1 =_\alpha t_2$ to denote ${\ensuremath{\vdash}}{\ensuremath{\mathrm{Eql}\,\alpha\,t_1\,t_2\,}}$ or $\Gamma {\ensuremath{\vdash}}{\ensuremath{\mathrm{Eql}\,\alpha\,t_1\,t_2\,}}$, or write $t_1 =_\alpha t_2$ instead of ${\ensuremath{\mathrm{Eql}\,\alpha\,t_1\,t_2\,}}$.
\[lem\_leibniz\] If $\Gamma {\ensuremath{\vdash}}\alpha : {\ensuremath{\mathrm{Type}}}$ then
- $\Gamma {\ensuremath{\vdash}}\forall x, y : \alpha \,.\, (x =_\alpha y) : {\ensuremath{\mathrm{Prop}}}$,
- $\Gamma {\ensuremath{\vdash}}\forall x : \alpha \,.\, (x =_\alpha x)$,
- $\Gamma {\ensuremath{\vdash}}\forall x, y : \alpha \,.\, (x =_\alpha y) \supset
(y =_\alpha x)$,
- $\Gamma {\ensuremath{\vdash}}\forall x, y, z : \alpha \,.\, (x =_\alpha y)
\wedge (y =_\alpha z) \supset (x =_\alpha z)$.
The system ${\ensuremath{{\cal I}}}_s$, as it is stated, is intensional with respect to Leibniz equality. We could add the rules $$e_f:\;
\inferrule{
\Gamma {\ensuremath{\vdash}}\alpha : {\ensuremath{\mathrm{Type}}}\\ \Gamma {\ensuremath{\vdash}}\beta : {\ensuremath{\mathrm{Type}}}}{
\forall f_1, f_2 : \alpha \to \beta \,.\, (\forall x
: \alpha \,.\, f_1 x =_\beta f_2 x) \supset (f_1 =_{\alpha\to\beta} f_2)
}$$ $$e_b:\;
\inferrule{
\Gamma {\ensuremath{\vdash}}\varphi_1 \supset \varphi_2 \\
\Gamma {\ensuremath{\vdash}}\varphi_2 \supset \varphi_1
}{
\Gamma {\ensuremath{\vdash}}\varphi_1 = \varphi_2
}$$ to obtain an extensional variant $e{\ensuremath{{\cal I}}}_s$ of ${\ensuremath{{\cal I}}}_s$. The system $e{\ensuremath{{\cal I}}}_s$ is still consistent – the model we construct for ${\ensuremath{{\cal I}}}_s$ validates the above rules.
${\ensuremath{\vdash}}_{e{\ensuremath{{\cal I}}}_s} \forall x, y : {\ensuremath{\mathrm{Prop}}}\,.\, (x =_{\ensuremath{\mathrm{Prop}}}y) \supset (x
= y)$
Induction and Natural Numbers {#sec_induction}
-----------------------------
The system ${\ensuremath{{\cal I}}}_s$ incorporates basic inductive types. In accordance with the terminology from [@BlanquiJounnaudOkada2002], an inductive type is basic if its constructors have no functional arguments. This class of inductive types includes most simple commonly used inductive types, e.g. natural numbers, lists, finite trees.
In our approach the types of constructors are encoded in the syntactic form of the inductive type. For instance, if $\iota_0, \iota_1 \in
{\ensuremath{{\cal T}}}_I$, then $\iota = \mu(\langle \rangle, \langle \iota_0, \iota_1,
\star, \star \rangle)$ is an inductive type with constructors: $c_1^\iota : \iota$ and $c_2^\iota : \iota_0 \to \iota_1 \to \iota \to
\iota \to \iota$.
If $c_i^\iota \in {\ensuremath{{\cal C}}}$ of arity $n_i$ has signature $\langle
\iota_1,\ldots,\iota_{n_i} \rangle$ then ${\ensuremath{\vdash}}_{{\ensuremath{{\cal I}}}_s} c_i^\iota :
\iota_1^* \to \ldots \to \iota_{n_i}^* \to \iota$.
${\ensuremath{\vdash}}_{{\ensuremath{{\cal I}}}_s} o_i^\iota : \iota \to {\ensuremath{\mathrm{Prop}}}$ and ${\ensuremath{\vdash}}_{{\ensuremath{{\cal I}}}_s}
\forall x : \iota \,.\, o_i^\iota x \supset (d_{i,j}^\iota x :
\iota_{i,j}^*)$
\[lem\_ind\_leibniz\_implies\_eq\] If $\iota \in {\ensuremath{{\cal T}}}_I$ then ${\ensuremath{\vdash}}_{{\ensuremath{{\cal I}}}_s} \forall x, y : \iota \,.\,
x =_\iota y \supset x = y$.
We may define the type of natural numbers by ${\ensuremath{\mathrm{Nat}}}\equiv \mu(\langle
\rangle, \langle \star \rangle)$. We use the abbreviations: $0 \equiv
c_1^{\ensuremath{\mathrm{Nat}}}$ (zero), ${\ensuremath{\mathfrak{0}}}\equiv o_1^{\ensuremath{\mathrm{Nat}}}$ (test for zero), ${\ensuremath{\mathfrak{s}}}\equiv c_2^{\ensuremath{\mathrm{Nat}}}$ (successor) and ${\ensuremath{\mathfrak{p}}}\equiv \lambda x \,.\,
{\ensuremath{\mathrm{Cond}\,({\ensuremath{\mathfrak{0}}}x)\,0\,(d_{2,1}^{\ensuremath{\mathrm{Nat}}}x)\,}}$ (predecessor). The rules $i_i^{\ensuremath{\mathrm{Nat}}}$ and $i_t^{{\ensuremath{\mathrm{Nat}}},k}$ become:
[lr]{}\
$
{n_t^1:}\; \inferrule{
}{
\Gamma {\ensuremath{\vdash}}0 : {\ensuremath{\mathrm{Nat}}}}
$ & $
{n_t^2:}\; \inferrule{
\Gamma {\ensuremath{\vdash}}t : {\ensuremath{\mathrm{Nat}}}}{
\Gamma {\ensuremath{\vdash}}({\ensuremath{\mathfrak{s}}}t) : {\ensuremath{\mathrm{Nat}}}}
$
To simplify the exposition, we discuss some properties of our formulation of inductive types using the example of natural numbers. Much of what we say applies to other basic inductive types, with appropriate modifications.
The rule $n_i$ is an induction principle for natural numbers. An important property of this induction principle is that it places no restrictions on $t$. This allows us to prove by induction on natural numbers properties of terms about which nothing is known beforehand. In particular, we do not need to know whether $t$ has a $\beta$-normal form in order to apply the rule $n_i$ to it. In contrast, an induction principle of the form e.g. $$n_i':\;\; \forall f : {\ensuremath{\mathrm{Nat}}}\to {\ensuremath{\mathrm{Prop}}}\,.\, \left(\left(f 0 \wedge (\forall
x : {\ensuremath{\mathrm{Nat}}}\,.\, f x \supset f ({\ensuremath{\mathfrak{s}}}x))\right) \supset \forall x : {\ensuremath{\mathrm{Nat}}}\,.\, f
x\right)$$ would be much less useful, because to apply it to a term $t$ we would have to prove $t : {\ensuremath{\mathrm{Nat}}}\to {\ensuremath{\mathrm{Prop}}}$ *beforehand*. Examples of the use of the rule $n_i$ for reasoning about possibly nonterminating general recursive programs are given in Sect. \[sec\_partiality\].
We may define a recursor $R$ for natural numbers in the following way: $$\begin{aligned}
R &=& \lambda g h x y \,.\, {\ensuremath{\mathrm{Cond}\,({\ensuremath{\mathfrak{0}}}y)\,(g x)\,(h\, x\, ({\ensuremath{\mathfrak{p}}}y)\,
(R\, g\, h\, x\, ({\ensuremath{\mathfrak{p}}}y)))\,}}.\end{aligned}$$ Note that we need the predecessor as a primitive, because otherwise a recursor would not be definable.
Now $+$, $-$, ${\ensuremath{\cdot}}$, $<$ and $\le$, usually used in infix notation, are defined as follows. $$\begin{aligned}
x + y &=& R (\lambda x \,.\, x) (\lambda x y z \,.\, {\ensuremath{\mathfrak{s}}}z) x y \\
x - y &=& R (\lambda x \,.\, x) (\lambda x y z \,.\, {\ensuremath{\mathfrak{p}}}z) x y \\
x {\ensuremath{\cdot}}y &=& R (\lambda x \,.\, 0) (\lambda x y z \,.\, x + z) x y \\
x \le y &=& {\ensuremath{\mathfrak{0}}}(x - y) \\
x < y &=& ({\ensuremath{\mathfrak{s}}}x) \le y\end{aligned}$$
\[lem\_nat\_op\_well\_defined\] The following terms are derivable in the system ${\ensuremath{{\cal I}}}_s$:
- $\forall x, y : {\ensuremath{\mathrm{Nat}}}\,.\, (x + y) : {\ensuremath{\mathrm{Nat}}}$, $\forall x, y :
{\ensuremath{\mathrm{Nat}}}\,.\, (x - y) : {\ensuremath{\mathrm{Nat}}}$, $\forall x, y : {\ensuremath{\mathrm{Nat}}}\,.\, (x {\ensuremath{\cdot}}y)
: {\ensuremath{\mathrm{Nat}}}$,
- $\forall x, y : {\ensuremath{\mathrm{Nat}}}\,.\, (x \le y) : {\ensuremath{\mathrm{Prop}}}$, $\forall x, y : {\ensuremath{\mathrm{Nat}}}\,.\, (x < y) : {\ensuremath{\mathrm{Prop}}}$.
\[lem\_le\_eq\] ${\ensuremath{\vdash}}_{{\ensuremath{{\cal I}}}_s} \forall x, y : {\ensuremath{\mathrm{Nat}}}\,.\, (x \ge y) \wedge (x \le y)
\supset (x =_{\ensuremath{\mathrm{Nat}}}y)$.
It is possible to derive Peano axioms for $+$ and ${\ensuremath{\cdot}}$ defined as above.
\[thm\_peano\] The following terms are derivable in the system ${\ensuremath{{\cal I}}}_s$:
- $\forall x, y : {\ensuremath{\mathrm{Nat}}}\,.\, \left({\ensuremath{\mathfrak{s}}}x =_{\ensuremath{\mathrm{Nat}}}{\ensuremath{\mathfrak{s}}}y\right)
\supset \left(x =_{\ensuremath{\mathrm{Nat}}}y\right)$,
- $\forall x : {\ensuremath{\mathrm{Nat}}}\,.\, \neg ({\ensuremath{\mathfrak{s}}}x =_{\ensuremath{\mathrm{Nat}}}0)$,
- $\forall x : {\ensuremath{\mathrm{Nat}}}\,.\, (x + 0 =_{\ensuremath{\mathrm{Nat}}}x)$,
- $\forall x, y : {\ensuremath{\mathrm{Nat}}}\,.\, (x + {\ensuremath{\mathfrak{s}}}y =_{\ensuremath{\mathrm{Nat}}}{\ensuremath{\mathfrak{s}}}(x + y))$,
- $\forall x : {\ensuremath{\mathrm{Nat}}}\,.\, (x {\ensuremath{\cdot}}0 =_{\ensuremath{\mathrm{Nat}}}0)$,
- $\forall x, y : {\ensuremath{\mathrm{Nat}}}\,.\, (x {\ensuremath{\cdot}}({\ensuremath{\mathfrak{s}}}y) =_{\ensuremath{\mathrm{Nat}}}(x {\ensuremath{\cdot}}y)
+ x)$.
The following theorem shows that any function for which there exists a measure on its arguments, which may be shown to decrease with every recursive call in each of a finite number of exhaustive cases, is typable in our system.
Suppose $\Gamma {\ensuremath{\vdash}}\forall x_1 : \alpha_1 \ldots \forall x_n :
\alpha_n \,.\, \varphi_1 \lor \ldots \lor \varphi_m$, $\Gamma
{\ensuremath{\vdash}}\alpha_j : {\ensuremath{\mathrm{Type}}}$ for $j = 1,\ldots,n$, and for $i=1,\ldots,m$: $\Gamma {\ensuremath{\vdash}}\forall x_1 : \alpha_1 \ldots
\forall x_n : \alpha_n \,.\, t_i : \beta \to \ldots \to \beta$ where $\beta$ occurs $k_i + 1$ times, $\Gamma {\ensuremath{\vdash}}\forall x_1 :
\alpha_1 \ldots \forall x_n : \alpha_n \,.\, t_{i,j,k} : \alpha_k$ for $j=1,\ldots,k_i$, $k=1,\ldots,n$, $x_1,\ldots,x_n \notin
FV(f,\alpha_1,\ldots,\alpha_n,\beta)$ and $$\begin{aligned}
\Gamma {\ensuremath{\vdash}}\forall x_1 : \alpha_1 \ldots \forall x_n :
\alpha_n \,.\, \varphi_i &\supset& (f x_1 \ldots x_n = \\ && \;t_i (f t_{i,1,1}
\ldots t_{i,1,n}) \ldots (f t_{i,k_i,1} \ldots t_{i,k_i,n})).
\end{aligned}$$ If there is a term $g$ such that $\Gamma {\ensuremath{\vdash}}g : \alpha_1 \to
\ldots \to \alpha_n \to {\ensuremath{\mathrm{Nat}}}$ and for $i=1,\ldots,m$ $$\begin{aligned}
\Gamma {\ensuremath{\vdash}}\forall x_1 : \alpha_1 \ldots \forall x_n : \alpha_n
\,.\, \varphi_i &\supset& (\left((f x_1 \ldots x_n) :
\beta\right) \lor \\ && \;((g t_{i,1,1} \ldots t_{i,1,n}) < (g x_1 \ldots
x_n) \land \ldots \land \\ && \;\;(g t_{i,k_i,1} \ldots t_{i,k_i,n}) < (g x_1
\ldots x_n)))
\end{aligned}$$ where $x_1,\ldots,x_n \notin FV(g)$, then $$\Gamma {\ensuremath{\vdash}}f : \alpha_1 \to \ldots \to \alpha_n \to \beta.$$
Conservativity and Consistency {#sec_consistent}
==============================
In this section we show a sound embedding of ordinary classical higher-order logic into ${\ensuremath{{\cal I}}}_s$, which we also conjecture to be complete. We have a completeness proof only for a restriction of this embedding to first-order logic. We also give a brief overview of the model construction used to establish consistency of ${\ensuremath{{\cal I}}}_s$.
First, let us define the system CPRED$\omega$ of classical higher-order logic.
- The *types* of CPRED$\omega$ are given by $${\ensuremath{{\cal T}}}\;\; ::= \;\; o \;|\; {\ensuremath{{\cal B}}}\;|\; {\ensuremath{{\cal T}}}\rightarrow {\ensuremath{{\cal T}}}$$ where ${\ensuremath{{\cal B}}}$ is a specific finite set of base types. The type $o$ is the type of propositions.
- The set of terms of CPRED$\omega$ of type $\tau$, denoted $T_\tau$, is defined as follows:
- $V_\tau, \Sigma_\tau \subseteq T_\tau$,
- if $t_1 \in T_{\sigma\to\tau}$ and $t_2 \in T_\sigma$ then $t_1 t_2 \in T_\tau$,
- if $x \in V_{\tau_1}$ and $t \in T_{\tau_2}$ then $\lambda x :
\tau_1 \,.\, t \in T_{\tau_1\to\tau_2}$,
- if $\varphi, \psi \in T_o$ then $\varphi \supset \psi \in
T_o$,
- if $x \in V_{\tau}$ and $\varphi \in T_o$ then $\forall x :
\tau \,.\, \varphi \in T_o$,
where for each type $\tau$ the set $V_\tau$ is a countable set of variables and $\Sigma_\tau$ is a countable set of constants. We assume that the sets $V_\tau$ and $\Sigma_\sigma$ are all pairwise disjoint. Terms of type $o$ are *formulas*. As usual, we omit spurious brackets and assume that application associates to the left. We identify $\alpha$-equivalent terms, i.e. terms differing only in the names of bound variables are considered identical.
- The system CPRED$\omega$ is given by the following rules and axioms, where $\Delta$ is a finite set of formulas, $\varphi, \psi$ are formulas. The notation $\Delta, \varphi$ is a shorthand for $\Delta \cup \{\varphi\}$.
[**Axioms**]{}
- $\Delta, \varphi {\ensuremath{\vdash}}\varphi$
- $\Delta {\ensuremath{\vdash}}\forall p : o \,.\, ((p \supset \bot) \supset
\bot) \supset p$ where $\bot \equiv \forall p : o \,.\, p$
[**Rules**]{}
[lr]{} $
{\supset_i^P:}\; \inferrule{\Delta, \varphi {\ensuremath{\vdash}}\psi}{\Delta {\ensuremath{\vdash}}\varphi \supset \psi}
$ & $
{\supset_e^P:}\; \inferrule{\Delta {\ensuremath{\vdash}}\varphi \supset
\psi \;\;\; \Delta {\ensuremath{\vdash}}\varphi}{\Delta {\ensuremath{\vdash}}\psi}
$\
&\
$
{\forall_i^P:}\; \inferrule{\Delta {\ensuremath{\vdash}}\varphi}{\Delta {\ensuremath{\vdash}}\forall x : \tau \,.\, \varphi} \; x
\notin FV(\Delta), x \in V_\tau
$ & $
{\forall_e^P:}\; \inferrule{\Delta {\ensuremath{\vdash}}\forall x : \tau \,.\, \varphi}{\Delta
{\ensuremath{\vdash}}\varphi[x/t]}\; t \in T_\tau
$\
&\
In CPRED$\omega$, we define Leibniz equality in type $\tau \in {\ensuremath{{\cal T}}}$ by $$t_1 =_\tau t_2 \equiv \forall p : \tau \to o \,.\, p t_1 \supset p t_2$$ The system CPRED$\omega$ is intensional. An extensional variant E-CPRED$\omega$ may be obtained by adding the following axioms for all $\tau, \sigma \in {\ensuremath{{\cal T}}}$: $$e_f^P: \forall f_1, f_2 : \tau \to \sigma \,.\, \left(\forall x : \tau
\,.\, f_1 x =_\sigma f_2 x \right) \supset (f_1 =_{\tau\to\sigma} f_2)$$ $$e_b^P: \forall \varphi_1, \varphi_2 : o \,.\, \left((\varphi_1 \supset
\varphi_2) \wedge (\varphi_2 \supset \varphi_1)\right) \supset
(\varphi_1 =_{o} \varphi_2)$$
For an arbitrary set of formulas $\Delta$ we write $\Delta {\ensuremath{\vdash}}_S
\varphi$ if $\varphi$ is derivable from a subset of $\Delta$ in system $S$.
We now define a mapping ${\ensuremath{\lceil-\rceil}}$ from types and terms of CPRED$\omega$ to terms of ${\ensuremath{{\cal I}}}_s$, and a mapping $\Gamma(-)$ from sets of terms of CPRED$\omega$ to sets of terms of ${\ensuremath{{\cal I}}}_s$ providing necessary context. We assume that ${\ensuremath{{\cal B}}}\subseteq \Sigma_s$ and $\Sigma_\tau \subseteq \Sigma_s$ for $\tau \in {\ensuremath{{\cal T}}}$, i.e. that all base types and all constants of CPRED$\omega$ occur as constants in ${\ensuremath{{\cal I}}}_s$, and also $V_\tau \subseteq V_s$ for $\tau \in {\ensuremath{{\cal T}}}$. The definition of ${\ensuremath{\lceil-\rceil}}$ is inductive:
- ${\ensuremath{\lceil\tau\rceil}} = \tau$ for $\tau \in {\ensuremath{{\cal B}}}$,
- ${\ensuremath{\lceilo\rceil}} = {\ensuremath{\mathrm{Prop}}}$,
- ${\ensuremath{\lceil\tau_1\to\tau_2\rceil}} = {\ensuremath{\lceil\tau_1\rceil}} \to {\ensuremath{\lceil\tau_2\rceil}}$ for $\tau_1,\tau_2 \in {\ensuremath{{\cal T}}}$,
- ${\ensuremath{\lceilc\rceil}} = c$ if $c \in \Sigma_\tau$ for some $\tau \in {\ensuremath{{\cal T}}}$,
- ${\ensuremath{\lceilx\rceil}} = x$ if $x \in V_\tau$ for some $\tau \in {\ensuremath{{\cal T}}}$,
- ${\ensuremath{\lceilt_1 t_2\rceil}} = {\ensuremath{\lceilt_1\rceil}} {\ensuremath{\lceilt_2\rceil}}$,
- ${\ensuremath{\lceil\lambda x : \tau \,.\, t\rceil}} = \lambda x \,.\,
{\ensuremath{\lceilt\rceil}}$,
- ${\ensuremath{\lceil\varphi \supset \psi\rceil}} = {\ensuremath{\lceil\varphi\rceil}} \supset
{\ensuremath{\lceil\psi\rceil}}$,
- ${\ensuremath{\lceil\forall x : \tau \,.\, \varphi\rceil}} = \forall x :
{\ensuremath{\lceil\tau\rceil}} \,.\, {\ensuremath{\lceil\varphi\rceil}}$.
If $\Delta$ is a set of formulas, then ${\ensuremath{\lceil\Delta\rceil}}$ denotes the image of ${\ensuremath{\lceil-\rceil}}$ on $\Delta$. The set $\Gamma(\Delta)$ is defined to contain the following:
- $x : {\ensuremath{\lceil\tau\rceil}}$ for all $\tau \in {\ensuremath{{\cal T}}}$ and all $x \in
FV(\Delta)$ such that $x \in V_\tau$,
- $c : {\ensuremath{\lceil\tau\rceil}}$ for all $\tau \in {\ensuremath{{\cal T}}}$ and all $c \in
\Sigma_\tau$,
- $\tau : {\ensuremath{\mathrm{Type}}}$ for all $\tau \in {\ensuremath{{\cal B}}}$,
- $y : \tau$ for all $\tau \in {\ensuremath{{\cal B}}}$ and some $y \in V_\tau$ such that $y \notin FV(\Delta)$.
The last point is needed because in ordinary logic one always assumes that the domains are non-empty.
\[thm\_sound\_hol\] If $\Delta {\ensuremath{\vdash}}_{\mathrm{CPRED}\omega} \varphi$ then $\Gamma(\Delta, \varphi), {\ensuremath{\lceil\Delta\rceil}} {\ensuremath{\vdash}}_{{\ensuremath{{\cal I}}}_s}
{\ensuremath{\lceil\varphi\rceil}}$. The same holds if we change CPRED$\omega$ to and ${\ensuremath{{\cal I}}}_s$ to $e{\ensuremath{{\cal I}}}_s$.
The above theorem shows that ${\ensuremath{{\cal I}}}_s$ may be considered an extension of ordinary higher-order logic. This extension is essentially obtained by relaxing typing requirements on allowable $\lambda$-terms. Type-checking is obviously undecidable in ${\ensuremath{{\cal I}}}_s$, but the purpose of types in illative systems is not to have a decidable method for syntactic correctness checks, but to provide general means for classifying terms into various categories. In practice, one might still want to have a decidable (and necessarily incomplete) method for checking correctness of some designated type assertions. Such a method may be obtained by employing any type-checking algorithm sound w.r.t. ${\ensuremath{{\cal I}}}_s$.[^5] However, the difference would be that a priori type-checks would not be enforced in every situation. Occasionally, with some more complex recursive functions, it might be convenient to forgo these checks and reason about the types of such functions explicitly, using the rules of ${\ensuremath{{\cal I}}}_s$.
\[conj\_conservative\_hol\] If $\Gamma(\Delta, \varphi), {\ensuremath{\lceil\Delta\rceil}} {\ensuremath{\vdash}}_{{\ensuremath{{\cal I}}}_s}
{\ensuremath{\lceil\varphi\rceil}}$ then $\Delta {\ensuremath{\vdash}}_{\mathrm{CPRED}\omega}
\varphi$. The same holds if we change CPRED$\omega$ to and ${\ensuremath{{\cal I}}}_s$ to $e{\ensuremath{{\cal I}}}_s$.
We were able to prove this conjecture only for first-order logic. The system of classical first-order logic (FOL) is obtained by restricting CPRED$\omega$ in obvious ways (leaving only one base type $\iota$, disallowing $\lambda$-abstraction, allowing quantification only over $\iota$, and constants only of types $\iota$, $\iota \to \ldots \to \iota \to \iota$ or $\iota \to \ldots \to \iota
\to o$).
\[thm\_embedding\_01\] If ${\ensuremath{{\cal I}}}= {\ensuremath{{\cal I}}}_s$ or ${\ensuremath{{\cal I}}}= e{\ensuremath{{\cal I}}}_s$ then $$\Delta {\ensuremath{\vdash}}_{\mathrm{FOL}} \varphi \mathrm{\ \ \ iff\ \ \ }
\Gamma(\Delta, \varphi), {\ensuremath{\lceil\Delta\rceil}} {\ensuremath{\vdash}}_{{\ensuremath{{\cal I}}}}
{\ensuremath{\lceil\varphi\rceil}}$$
\[thm\_consistent\_01\] The systems ${\ensuremath{{\cal I}}}_s$ and $e{\ensuremath{{\cal I}}}_s$ are consistent, i.e. $\not{\ensuremath{\vdash}}_{{\ensuremath{{\cal I}}}_s} \bot$ and $\not{\ensuremath{\vdash}}_{e{\ensuremath{{\cal I}}}_s} \bot$.
This follows from Theorem \[thm\_embedding\_01\], but we actually prove Theorem \[thm\_consistent\_01\] first by constructing a model, and then use this construction to show Theorem \[thm\_embedding\_01\].
We now give an informal overview of the model construction. To simplify the exposition we pretend ${\ensuremath{{\cal I}}}_s$ allows only function types. Inductive types and subtypes add some technicalities, but the general idea of the construction remains the same. This overview is necessarily very brief. An interested reader is advised to consult a technical appendix for more details.
An ${\ensuremath{{\cal I}}}_s$-model is defined essentially as a $\lambda$-model (see e.g. [@Barendregt1984 Chapter 5]) with designated elements interpreting the constants of ${\ensuremath{{\cal I}}}_s$, satisfying certain requirements. By ${\ensuremath{\llbrackett\rrbracket_{}^{{\ensuremath{{\cal M}}}}}}$ we denote the interpretation of the ${\ensuremath{{\cal I}}}_s$-term $t$ in a model ${\ensuremath{{\cal M}}}$. The conditions imposed on an ${\ensuremath{{\cal I}}}_s$-model express the meaning of each rule of ${\ensuremath{{\cal I}}}_s$ according to the intuitive semantics. For instance, we have the conditions:
- for $a \in {\ensuremath{{\cal M}}}$, if ${\ensuremath{\llbracket\mathrm{Is}\rrbracket_{}^{{\ensuremath{{\cal M}}}}}} \cdot a \cdot
{\ensuremath{\llbracket{\ensuremath{\mathrm{Type}}}\rrbracket_{}^{{\ensuremath{{\cal M}}}}}} = {\ensuremath{\llbracket\top\rrbracket_{}^{{\ensuremath{{\cal M}}}}}}$ and for all $c
\in {\ensuremath{{\cal M}}}$ such that ${\ensuremath{\llbracket\mathrm{Is}\rrbracket_{}^{{\ensuremath{{\cal M}}}}}} \cdot c \cdot a =
{\ensuremath{\llbracket\top\rrbracket_{}^{{\ensuremath{{\cal M}}}}}}$ we have $b \cdot c =
{\ensuremath{\llbracket\top\rrbracket_{}^{{\ensuremath{{\cal M}}}}}}$ then ${\ensuremath{\llbracket\forall\rrbracket_{}^{{\ensuremath{{\cal M}}}}}} \cdot a
\cdot b = {\ensuremath{\llbracket\top\rrbracket_{}^{{\ensuremath{{\cal M}}}}}}$,
- for $a,b \in {\ensuremath{{\cal M}}}$, if ${\ensuremath{\llbracket\forall\rrbracket_{}^{{\ensuremath{{\cal M}}}}}}
\cdot a \cdot b = {\ensuremath{\llbracket\top\rrbracket_{}^{{\ensuremath{{\cal M}}}}}}$ then for all $c \in A$ such that ${\ensuremath{\llbracket\mathrm{Is}\rrbracket_{}^{{\ensuremath{{\cal M}}}}}} \cdot c \cdot a =
{\ensuremath{\llbracket\top\rrbracket_{}^{{\ensuremath{{\cal M}}}}}}$ we have $b \cdot c =
{\ensuremath{\llbracket\top\rrbracket_{}^{{\ensuremath{{\cal M}}}}}}$.
Here $\cdot$ is the application operation in the model.
We show that the semantics based on ${\ensuremath{{\cal I}}}_s$-models is sound for ${\ensuremath{{\cal I}}}_s$. Then it suffices to construct a non-trivial (i.e. such that ${\ensuremath{\llbracket\top\rrbracket_{}^{}}} \ne {\ensuremath{\llbracket\bot\rrbracket_{}^{}}}$) ${\ensuremath{{\cal I}}}_s$-model to establish consistency of ${\ensuremath{{\cal I}}}_s$. The model will in fact satisfy additional conditions corresponding to the rules $e_f$ and $e_b$, so we obtain consistency of $e{\ensuremath{{\cal I}}}_s$ as well.
The model is constructed as the set of equivalence classes of a certain relation ${\ensuremath{\stackrel{*}{\Leftrightarrow}}}$ on the set of so called semantic terms. A semantic term is a well-founded tree whose leaves are labelled with variables or constants, and whose internal nodes are labelled with $\cdot$, $\lambda x$ or ${\ensuremath{\mathrm{A}}}\tau$. The intuitive interpretation of nodes labelled with $\cdot$ or $\lambda x$ is obvious. For semantic terms with the roots labelled with $\cdot$ and $\lambda x$ we use the abbreviations $t_1 t_2$ and $\lambda x . t$, respectively. A node labelled with ${\ensuremath{\mathrm{A}}}\tau$ represents universal quantification over a set of constants $\tau$, i.e. it “represents” the statement: for all $c \in \tau$, $t c$ is true. Such a node has one child for each $c \in
\tau$. The relation ${\ensuremath{\stackrel{*}{\Leftrightarrow}}}$ will be defined as the equivalence relation generated by a certain reduction relation ${\ensuremath{\Rightarrow}}$ on semantic terms. The relation ${\ensuremath{\Rightarrow}}$ will satisfy[^6]: $(\lambda x . t_1) t_2
{\ensuremath{\Rightarrow}}t_1[x/t_2]$, $\lor \top t {\ensuremath{\Rightarrow}}\top$, $\lor \bot \bot
{\ensuremath{\Rightarrow}}\bot$, etc. The question is how to define ${\ensuremath{\Rightarrow}}$ for $\forall t_1 t_2$ so that the resulting structure satisfies ($\forall_\top$). One could try closing ${\ensuremath{\Rightarrow}}$ under the rule:
- if $\mathrm{Is}\, t_1\, {\ensuremath{\mathrm{Type}}}{\ensuremath{\stackrel{*}{\Rightarrow}}}\top$ and for all semantic terms $t$ such that $\mathrm{Is}\, t\, t_1 {\ensuremath{\stackrel{*}{\Rightarrow}}}\top$ we have $t_2 t {\ensuremath{\stackrel{*}{\Rightarrow}}}\top$, then $\forall t_1 t_2 {\ensuremath{\Rightarrow}}\top$.
However, there is a negative reference to ${\ensuremath{\Rightarrow}}$ here, so the definition would not be monotone, and we would not necessarily reach a fixpoint. This is a major problem. We somehow need to know the range of all quantifiers beforehand. However, the range (i.e. the set of all semantic terms $t$ such that $t_1 t {\ensuremath{\stackrel{*}{\Rightarrow}}}\top$) depends on the definition of ${\ensuremath{\Rightarrow}}$, so it is not at all clear how to achieve this.
Fortunately, it is not so difficult to analyze a priori the form of types of ${\ensuremath{{\cal I}}}_s$. Informally, if $t : {\ensuremath{\mathrm{Type}}}$ is true, then $t$ corresponds to a set in ${\ensuremath{{\cal T}}}$, where ${\ensuremath{{\cal T}}}$ is defined as follows, ignoring subtypes and inductive types, but instead introducing a base type $\delta$ of individuals.
- $\delta, {\ensuremath{\mathrm{Bool}}}\in {\ensuremath{{\cal T}}}$ where ${\ensuremath{\mathrm{Bool}}}= \{\top,\bot\}$ and $\delta$ is an arbitrary set of fresh constants.
- If $\tau_1,\tau_2 \in {\ensuremath{{\cal T}}}$ then $\tau_2^{\tau_1} \in {\ensuremath{{\cal T}}}$, where $\tau_2^{\tau_1}$ is the set of all set-theoretic functions from $\tau_1$ to $\tau_2$.
We take the elements of ${\ensuremath{{\cal T}}}$ and $\bigcup {\ensuremath{{\cal T}}}\setminus {\ensuremath{\mathrm{Bool}}}$ as fresh constants, i.e. they may occur as constants in semantic terms. The elements of $\bigcup {\ensuremath{{\cal T}}}$ are *canonical constants*. If $c \in \tau_2^{\tau_1}$ and $c_1 \in \tau_1$ then we write ${\ensuremath{{\cal F}}}(c)(c_1)$ instead of $c(c_1)$ to avoid confusion with the semantic term $c c_1$. We then define a relation $\succ$ satisfying:
- $c \succ c$ for a canonical constant $c$,
- if $c \in \tau_2^{\tau_1}$ and for all $c_1 \in \tau_1$ there exists a semantic term $t'$ such that $t c_1 {\ensuremath{\stackrel{*}{\Rightarrow}}}t' \succ
{\ensuremath{{\cal F}}}(c)(c_1)$, then $t \succ c$.
Intuitively, $t \succ c \in \tau$ holds if $c$ “simulates” $t$ in type $\tau$, i.e. $t$ behaves exactly like $c$ in every context where a term of type $\tau$ is “expected”.
The relation ${\ensuremath{\Rightarrow}}$ is then defined by transfinite induction in a monotone way. It will satisfy e.g.:
- if $t \succ c \in \tau \in {\ensuremath{{\cal T}}}$ then $\mathrm{Is}\,t\,\tau
{\ensuremath{\Rightarrow}}\top$,
- if $t \succ c_1 \in \tau_1$ and $c \in \tau_2^{\tau_1}$ then $c
t {\ensuremath{\Rightarrow}}{\ensuremath{{\cal F}}}(c)(c_1)$,
- $\mathrm{Fun}\,\tau_1\,\tau_2 {\ensuremath{\Rightarrow}}\tau_2^{\tau_1}$,
- $\forall \tau t {\ensuremath{\Rightarrow}}t'$ where the label at the root of $t'$ is ${\ensuremath{\mathrm{A}}}\tau$, and for each $c \in \tau$, $t'$ has a child $t c$,
- $t {\ensuremath{\Rightarrow}}\top$ if the label of the root of $t$ is ${\ensuremath{\mathrm{A}}}\tau$, and all children of $t$ are labelled with $\top$,
- if $t_c {\ensuremath{\stackrel{*}{\Rightarrow}}}t_c'$ for all $c \in \tau \in {\ensuremath{{\cal T}}}$, the label of the root of $t$ is ${\ensuremath{\mathrm{A}}}\tau$, and $\{ t_c \;|\; c \in \tau \}$ is the set of children of $t$, then $t {\ensuremath{\Rightarrow}}t'$, where the label of the root of $t'$ is ${\ensuremath{\mathrm{A}}}\tau$ and $\{t_c' \;|\; c \in \tau\}$ is the set of children of $t'$.
We removed negative references to ${\ensuremath{\Rightarrow}}$, but it is not easy to show that the resulting model will satisfy the required conditions. Two key properties established in the correctness proof are:
1. ${\ensuremath{\Rightarrow}}$ has the Church-Rosser property,
2. if $t_2 \succ c$ and $t_1 c {\ensuremath{\stackrel{*}{\Rightarrow}}}d \in \{\top,\bot\}$ then $t_1 t_2 {\ensuremath{\stackrel{*}{\Rightarrow}}}d$.
The second property shows that quantifying over only canonical constants of type $\tau$ is in a sense equivalent to quantifying over all terms of type $\tau$. This is essential for establishing e.g. the condition ($\forall_\top$).
Both of these properties have rather intricate proofs. Essentially, the proofs show certain commutation and postponement properties for ${\ensuremath{\Rightarrow}}$, $\succ$ and other auxiliary relations. The proofs proceed by induction on lexicographic products of various ordinals and other parameters associated with the relations and terms involved.
Partiality and General Recursion {#sec_partiality}
================================
In this section we give some examples of proofs in ${\ensuremath{{\cal I}}}_s$ of properties of functions defined by recursion. For lack of space, we give only informal indications of how formal proofs may be obtained, assuming certain basic properties of operations on natural numbers. The transformation of the given informal arguments into formal proofs in ${\ensuremath{{\cal I}}}_s$ is not difficult. Mostly complete formal proofs may be found in a technical appendix.
Consider a term ${\ensuremath{\mathrm{subp}}}$ satisfying the following recursive equation: $${\ensuremath{\mathrm{subp}}}= \lambda i j \,.\, {\ensuremath{\mathrm{Cond}\,(i =_{\ensuremath{\mathrm{Nat}}}j)\,0\,\left(\left({\ensuremath{\mathrm{subp}}}\, i \, \left(j + 1\right)\right) + 1\right)\,}}.$$ If $i \ge j$ then ${\ensuremath{\mathrm{subp}}}\,i\,j = i - j$. If $i < j$ then ${\ensuremath{\mathrm{subp}}}\,i\,j$ does not terminate. An appropriate specification for ${\ensuremath{\mathrm{subp}}}$ is $\forall i, j : {\ensuremath{\mathrm{Nat}}}\,.\, (i \ge j) \supset ({\ensuremath{\mathrm{subp}}}\, i\,
j = i - j)$.
Let $\varphi(y) = \forall i : {\ensuremath{\mathrm{Nat}}}\,.\, \forall j : {\ensuremath{\mathrm{Nat}}}\,.\,
\left(i \ge j \supset y =_{\ensuremath{\mathrm{Nat}}}i - j \supset
{\ensuremath{\mathrm{subp}}}\,i\,j\,=\,i-j\right)$. We show by induction on $y$ that $\forall y : {\ensuremath{\mathrm{Nat}}}\,.\, \varphi(y)$.
First note that under the assumptions $y : {\ensuremath{\mathrm{Nat}}}$, $i : {\ensuremath{\mathrm{Nat}}}$, $j :
{\ensuremath{\mathrm{Nat}}}$ it follows from Lemma \[lem\_nat\_op\_well\_defined\] that $(i
\ge j) : {\ensuremath{\mathrm{Prop}}}$ and $(y =_{\ensuremath{\mathrm{Nat}}}i - j) : {\ensuremath{\mathrm{Prop}}}$. Hence, whenever $y :
{\ensuremath{\mathrm{Nat}}}$, to show $i \ge j \supset y =_{\ensuremath{\mathrm{Nat}}}i - j \supset
{\ensuremath{\mathrm{subp}}}\,i\,j\,=\,i-j$ it suffices to derive ${\ensuremath{\mathrm{subp}}}\,i\,j\,=\,i-j$ under the assumptions $i \ge j$ and $y =_{\ensuremath{\mathrm{Nat}}}i - j$. By Lemma \[lem\_ind\_leibniz\_implies\_eq\] the assumption $y =_{\ensuremath{\mathrm{Nat}}}i -
j$ may be weakened to $y = i - j$.
In the base step it thus suffices to show ${\ensuremath{\mathrm{subp}}}\,i\,j\,=\,i-j$ under the assumptions $i : {\ensuremath{\mathrm{Nat}}}$, $j : {\ensuremath{\mathrm{Nat}}}$, $i \ge j$, $i - j =
0$. From $i - j = 0$ we obtain ${\ensuremath{\mathfrak{0}}}(i - j)$, so $j \ge i$. From $i \ge j$ and $i \le j$ we derive $i =_{\ensuremath{\mathrm{Nat}}}j$ by Lemma \[lem\_le\_eq\]. Then ${\ensuremath{\mathrm{subp}}}\,i\,j\,=\,i-j$ follows by simple computation (i.e. by applying rules for and appropriate rules for the conditional).
In the inductive step we have $\varphi(y)$ for $y : {\ensuremath{\mathrm{Nat}}}$ and we need to obtain $\varphi({\ensuremath{\mathfrak{s}}}y)$. It suffices to show ${\ensuremath{\mathrm{subp}}}\,i\,j = i
- j$ under the assumptions $i : {\ensuremath{\mathrm{Nat}}}$, $j : {\ensuremath{\mathrm{Nat}}}$ and ${\ensuremath{\mathfrak{s}}}y = i -
j$. Because ${\ensuremath{\mathfrak{s}}}y \ne_{\ensuremath{\mathrm{Nat}}}0$ we have $i \ne_{\ensuremath{\mathrm{Nat}}}j$, hence ${\ensuremath{\mathrm{subp}}}\,i\,j = {\ensuremath{\mathfrak{s}}}({\ensuremath{\mathrm{subp}}}\,i\,({\ensuremath{\mathfrak{s}}}j))$ follows by computation. Using the inductive hypothesis we now conclude ${\ensuremath{\mathrm{subp}}}\,i\,({\ensuremath{\mathfrak{s}}}j) = i - ({\ensuremath{\mathfrak{s}}}j)$, and thus ${\ensuremath{\mathrm{subp}}}\,i\,({\ensuremath{\mathfrak{s}}}j) =_{\ensuremath{\mathrm{Nat}}}i - ({\ensuremath{\mathfrak{s}}}j)$ by reflexivity of $=_{\ensuremath{\mathrm{Nat}}}$ on natural numbers. Then it follows by properties of operations on natural numbers that ${\ensuremath{\mathfrak{s}}}({\ensuremath{\mathrm{subp}}}\,i\,({\ensuremath{\mathfrak{s}}}j)) =_{\ensuremath{\mathrm{Nat}}}i -
j$. By Lemma \[lem\_ind\_leibniz\_implies\_eq\] we obtain the thesis.
We have thus completed an inductive proof of $\forall y : {\ensuremath{\mathrm{Nat}}}\,.\,
\varphi(y)$. Now we use this formula to derive ${\ensuremath{\mathrm{subp}}}\,i\,j\,=\,i-j$ under the assumptions $i : {\ensuremath{\mathrm{Nat}}}$, $j : {\ensuremath{\mathrm{Nat}}}$, $i \ge j$. Then it remains to apply implication introduction and $\forall$-introduction twice.
In the logic of PVS [@RushbyOwreShankar1998] one may define ${\ensuremath{\mathrm{subp}}}$ by specifying its domain precisely using predicate subtypes and dependent types, somewhat similarly to what is done here. However, an important distinction is that we do not require a domain specification to be a part of the definition. In an interactive theorem prover based on our formalism no proof obligations would need to be generated to establish termination of ${\ensuremath{\mathrm{subp}}}$ on its domain.
Note that because domain specification is not part of the definition of ${\ensuremath{\mathrm{subp}}}$, we may easily derive $\varphi \equiv \forall i, j : {\ensuremath{\mathrm{Nat}}}\,.\, \left(\left({\ensuremath{\mathrm{subp}}}\, i\, j = i - j\right) \vee
\,\left({\ensuremath{\mathrm{subp}}}\,j\, i = j - i\right)\right)$. This is not possible in PVS because the formula $\varphi$ translated to PVS generates false proof obligations [@RushbyOwreShankar1998].
The next example is a well-known “challenge” posed by McCarthy: $$f(n) = {\ensuremath{\mathrm{Cond}\,(n > 100)\,(n - 10)\,(f(f(n + 11)))\,}}$$ For $n \le 101$ we have $f(n) = 91$, which fact may be proven by induction on $101 - n$. This function is interesting because of its use of nested recursion. Termination behavior of a nested recursive function may depend on its functional behavior, which makes reasoning about termination and function value interdependent. This creates problems for systems with definitional restrictions of possible forms of recursion. Below we give an indication of how a formal proof of $\forall n : {\ensuremath{\mathrm{Nat}}}\,.\, n \le 101 \supset f(n) = 91$ may be derived in ${\ensuremath{{\cal I}}}_s$. Lemma \[lem\_nat\_op\_well\_defined\] is used implicitly with implication introduction.
Let $\varphi(y) \equiv \forall n : {\ensuremath{\mathrm{Nat}}}\,.\, n \le 101 \supset 101 - n
\le y \supset f(n) = 91$. We prove $\forall y : {\ensuremath{\mathrm{Nat}}}\,.\, \varphi(y)$ by induction on $y$.
In the base step we need to prove $f(n) = 91$ under the assumptions $n : {\ensuremath{\mathrm{Nat}}}$, $n \le 101$ and $101 - n \le y = 0$. We have $n =_{\ensuremath{\mathrm{Nat}}}101$, hence $n = 101$, and the thesis follows by simple computation.
In the inductive step we distinguish three cases:
1. $n + 11 > 101$ and $n < 101$,
2. $n + 11 > 101$ and $n \ge 101$,
3. $n + 11 \le 101$.
We need to prove $f(n) = 91$ under the assumptions of the inductive hypothesis $y : {\ensuremath{\mathrm{Nat}}}, \forall m : {\ensuremath{\mathrm{Nat}}}\,.\, m \le 101 \supset 101 -
m \le y \supset f(m) = 91$, and of $n : {\ensuremath{\mathrm{Nat}}}$, $n \le 101$ and $101
- n \le ({\ensuremath{\mathfrak{s}}}y)$.
In the first case we have $f(n + 11) = n + 1$ and $n + 1 \le
101$. Hence by the inductive hypothesis we conclude $100 - n \le y
\supset f(n + 1) = 91$. From $101 - n \le {\ensuremath{\mathfrak{s}}}y$ we infer $100 - n
\le y$, and hence $f(n + 1) = 91$. Since $n \le 100$ it follows by computation that $f(n) = f(f(n + 11)) = f(n + 1) = 91$.
In the second case $n = 101$ and the thesis follows by simple computation.
In the third case, from $101 - n \le {\ensuremath{\mathfrak{s}}}y$ we infer $101 - (n + 11)
\le y$. Since $n + 11 \le 101$ we conclude by the inductive hypothesis that $f(n + 11) = 91$. Because $n + 11 \le 101$, so $n
\le 100$, and by definition we infer $f(n) = f(f(n + 11)) =
f(91)$. Now we simply compute $f(91) = f(f(102)) = f(92) = f(f(103))
= \ldots = f(100) = f(f(111)) = f(101) = 91$ (i.e. we apply rules for $\mbox{Eq}$ and rules for the conditional an appropriate number of times).
This concludes the inductive proof of $\forall y : {\ensuremath{\mathrm{Nat}}}\,.\, \forall n
: {\ensuremath{\mathrm{Nat}}}\,.\, n \le 101 \supset 101 - n \le y \supset f(n) = 91$. Having this it is not difficult to show $\forall n : {\ensuremath{\mathrm{Nat}}}\,.\, n \le 101
\supset f(n) = 91$.
Note that the computation of $f(91)$ in the inductive step relies on the fact that in our logic values of functions may always be computed for specific arguments, regardless of what we know about the function, regardless of whether it terminates in general.
Related Work {#sec_related}
============
In this section we discuss the relationship between ${\ensuremath{{\cal I}}}_s$ and the traditional illative system ${\ensuremath{{\cal I}}}_\omega$. We also briefly survey some approaches to dealing with partiality and general recursion in proof assistants. A general overview of the literature relevant to this problem may be found in [@BoveKraussSozeau2012].
Relationship with Systems of Illative Combinatory Logic
-------------------------------------------------------
In terms of the features provided, the system ${\ensuremath{{\cal I}}}_s$ may be considered an extension of ${\ensuremath{{\cal I}}}_\omega$ from [@Czajka2013Accepted]. However, there are some technical differences between ${\ensuremath{{\cal I}}}_s$ and traditional systems of illative combinatory logic. For one thing, traditional systems strive to use as few constants and rules as possible. For instance, ${\ensuremath{{\cal I}}}_\omega$ has only two primitive constants, disregarding constants representing base types. Because of this in ${\ensuremath{{\cal I}}}_\omega$ e.g. ${\ensuremath{\mathrm{Is}\,=\,\lambda\,}} xy \,.\, y x$ and ${\ensuremath{\mathrm{Prop}}}= \lambda x \,.\, {\ensuremath{\mathrm{Type}}}(\lambda y . x)$, using the notation of the present paper. Moreover, the names of the primitive constants and the notations employed when discussing traditional illative systems are not in common use today. We will not explain these technicalities in any more detail. The reader may consult [@Seldin2009; @illat01; @Czajka2013Accepted] for more information on illative combinatory logic.
The system ${\ensuremath{{\cal I}}}_\omega$ from [@Czajka2013Accepted] is a direct extension of ${\ensuremath{{\cal I}}}\Xi$ from [@illat01] to higher-order logic. The ideas behind ${\ensuremath{{\cal I}}}_\omega$ date back to [@Bunder1983], or even earlier as far as the general form of restrictions in inference rules is concerned.
Below we briefly describe a system ${\ensuremath{{\cal I}}}_\omega'$ which is a variant of ${\ensuremath{{\cal I}}}_\omega$ adapted to our notation. It differs somewhat from ${\ensuremath{{\cal I}}}_\omega$, mostly by taking more constants as primitive, and thus having more rules and axioms. However, we believe that despite these differences it is reasonably close to ${\ensuremath{{\cal I}}}_\omega$ and shares its essential properties.
The terms of ${\ensuremath{{\cal I}}}_\omega'$ are those of ${\ensuremath{{\cal I}}}_s$, except that we do not allow subtypes, inductive types, Eq, Cond, $\vee$, $\bot$ and $\epsilon$. There are also additional primitive constants: $\omega$ (the type of all terms), $\varepsilon$ (the empty type) and $\supset$. The axioms are: $\Gamma, \varphi {\ensuremath{\vdash}}\varphi$, $\Gamma {\ensuremath{\vdash}}{\ensuremath{\mathrm{Prop}}}: {\ensuremath{\mathrm{Type}}}$, $\Gamma {\ensuremath{\vdash}}\varepsilon : {\ensuremath{\mathrm{Type}}}$, $\Gamma {\ensuremath{\vdash}}\omega : {\ensuremath{\mathrm{Type}}}$, $\Gamma {\ensuremath{\vdash}}t : \omega$. The rules are: $\forall_i$, $\forall_e$, $\forall_t$, $\supset_i$, $\supset_e$, $\supset_t$, $\to_i$, $\to_e$, $\to_t$, $p_i$, and the rules:
[lr]{} $
\mathrm{conv}: \inferrule{
\Gamma {\ensuremath{\vdash}}\varphi \\ \varphi =_\beta \psi
}{
\Gamma {\ensuremath{\vdash}}\psi
}
$ & $
\varepsilon_\bot: \inferrule{
\Gamma {\ensuremath{\vdash}}t : \varepsilon
}{
\Gamma {\ensuremath{\vdash}}\bot
}
$\
&\
Here $\varphi =_\beta \psi$ is a meta-level side-condition expressing $\beta$-equivalence of the terms $\varphi$ and $\psi$.
Partiality and Recursion in Proof Assistants
--------------------------------------------
Perhaps the most common way of dealing with recursion in interactive theorem provers is to impose certain syntactic restrictions on the form of recursive definitions so as to guarantee well-foundedness. For instance, the [**fix**]{} construct in Coq allows for structurally recursive definitions whose well-foundedness must be checked by a built-in automatic syntactic termination checker. Some systems, e.g. ACL2 or PVS, pass the task of proving termination to the user. Such systems require that a well-founded relation or a measure be given with each recursive function definition. Then the system generates so called proof obligations, or termination conditions, which state that the recursive calls are made on smaller arguments. The user must solve, i.e. prove, these obligations.
The method of restricting possible forms of recursive definitions obviously works only for total functions. If a function does not in fact terminate on some elements of its specified domain, then it cannot be introduced by a well-founded definition. One solution is to use a rich type system, e.g. dependent types combined with predicate subtyping, to precisely specify function domains so as to rule out the arguments on which the function does not terminate. This approach is adopted by PVS [@RushbyOwreShankar1998]. A related method of introducing general recursive functions in constructive type theory is to first define a special inductive accessibility predicate which precisely characterises the domain [@BoveCapretta2005]. The function is then defined by structural recursion on the proof that the argument satisfies the accessibility predicate.
A different approach to dealing with partiality and general recursion is to use a special logic which allows partial functions directly. Systems adopting this approach are often based on variants of the logic of partial terms of Beeson [@Feferman1995], [@Beeson1986b]. For instance, the IMPS interactive theorem prover [@FarmerGuttmanThayer1993] uses Farmer’s logic PF of partial functions [@Farmer1990], which is essentially a variant of the logic of partial terms adapted to higher-order logic. In these logics there is an additional definedness predicate which enables direct reasoning about definedness of terms.
The above gives only a very brief overview. There are many approaches to the problem of partiality and general recursion in interactive theorem provers, most of which we didn’t mention. We do not attempt here to provide a detailed comparison with a multitude of existing approaches or give in-depth arguments in favor of our system. For such arguments to be entirely convincing, they would need to be backed up by extensive experimentation in proving properties of sizable programs using a proof assistant based on our logic. No such proof assistants yet exist and no such experimentation has been undertaken. In contrast, our interest is theoretical.
Conclusion
==========
We have presented a system ${\ensuremath{{\cal I}}}_s$ of classical higher-order illative $\lambda$-calculus with subtyping and basic inductive types. A distinguishing characteristic of ${\ensuremath{{\cal I}}}_s$ is that it is based on the untyped $\lambda$-calculus. Therefore, it allows recursive definitions of potentially non-terminating functions directly. The inference rules of ${\ensuremath{{\cal I}}}_s$ are formulated in a way that makes it possible to apply them even when some of the terms used in the premises have not been proven to belong to any type. Additionally, our system may be considered an extension of ordinary higher-order logic, obtained by relaxing the typing restrictions on allowable $\lambda$-terms. We believe these facts alone make it relevant to the problem of partiality and recursion in proof assistants, and the system at least deserves some attention.
Some open problems related to ${\ensuremath{{\cal I}}}_s$ are as follows.
1. Is ${\ensuremath{{\cal I}}}_s$ conservative over higher-order logic (Conjecture \[conj\_conservative\_hol\])?
2. How to best incorporate a broader class of inductive types than just basic inductive types, e.g. all strictly positive inductive types?
3. How far is it possible to broaden the class of allowed types and still have a consistency proof in ZFC? For instance, in our model construction we could try naively interpreting dependent types by set-theoretic cartesian products when constructing the set ${\ensuremath{{\cal T}}}$ (see the overview of the consistency proof in Sect. \[sec\_consistent\]), but we would run out of sets. We conjecture that our construction may be modified to incorporate dependent types in the way indicated if we work in ZFC with one strongly inaccessible cardinal. This modification should not pose any fundamental difficulties. Is it possible prove consistency of ${\ensuremath{{\cal I}}}_s$ with dependent types in plain ZFC?
4. Can the premises $\Gamma {\ensuremath{\vdash}}\varphi : {\ensuremath{\mathrm{Prop}}}$ in the rules $c_3$ and $c_4$ be removed?
5. Can the premise $\Gamma {\ensuremath{\vdash}}\varphi : {\ensuremath{\mathrm{Prop}}}$ in the rule $c_5$ be removed?
Note that the premises $\Gamma {\ensuremath{\vdash}}\varphi : {\ensuremath{\mathrm{Prop}}}$ cannot be removed in $c_3$, $c_4$ and $c_5$ simultaneously. Let $\varphi$ be such that $\varphi = \neg \varphi$. Since $\varphi {\ensuremath{\vdash}}\bot$ and $\neg\varphi {\ensuremath{\vdash}}\bot$, it is easy to see that we would have both ${\ensuremath{\vdash}}{\ensuremath{\mathrm{Cond}\,\varphi\,\top\,\bot\,}} = \top$ and ${\ensuremath{\vdash}}{\ensuremath{\mathrm{Cond}\,\varphi\,\top\,\bot\,}} = \bot$.
Derived Rules for Implication {#app_logic}
=============================
In all derivations in this and subsequent sections we omit certain steps and rule assumptions, simplify inferences, and generally only give sketches of completely formal proofs, omitting the parts which may be easily reconstructed by the reader.
\[lem\_01\] If $\Gamma, \varphi {\ensuremath{\vdash}}\psi$, where $x \notin FV(\Gamma,
\varphi, \psi)$ and $y \notin FV(\varphi)$, then $\Gamma, x : {\ensuremath{\{ y
: {\ensuremath{\mathrm{Prop}}}\;|\; \varphi \}}} {\ensuremath{\vdash}}\psi$.
Straightforward induction on the length of derivation, using rule $s_e$ to show that $\Gamma {\ensuremath{\vdash}}x : {\ensuremath{\{ y :
{\ensuremath{\mathrm{Prop}}}\;|\; \varphi \}}}$ implies $\Gamma {\ensuremath{\vdash}}\varphi$, if $x \notin
FV(\Gamma, \varphi)$ and $y \notin FV(\varphi)$.
Now the rules for $\supset$ are derived as follows.
The rule $$\supset_i:\;
\inferrule{
\Gamma {\ensuremath{\vdash}}\varphi : {\ensuremath{\mathrm{Prop}}}\\ \Gamma, \varphi {\ensuremath{\vdash}}\psi
}{
\Gamma {\ensuremath{\vdash}}\varphi \supset \psi
}$$ follows by $$\inferrule*{
\inferrule*{
}{\Gamma {\ensuremath{\vdash}}{\ensuremath{\mathrm{Prop}}}: {\ensuremath{\mathrm{Type}}}} \\
\inferrule*{
\Gamma {\ensuremath{\vdash}}\varphi : {\ensuremath{\mathrm{Prop}}}\\
y \notin FV(\Gamma, \varphi)
}{
\Gamma, y : {\ensuremath{\mathrm{Prop}}}{\ensuremath{\vdash}}\varphi : {\ensuremath{\mathrm{Prop}}}}
}{
\Gamma {\ensuremath{\vdash}}{\ensuremath{\{ y : {\ensuremath{\mathrm{Prop}}}\;|\; \varphi \}}} : {\ensuremath{\mathrm{Type}}}} (a)$$ $$\inferrule*{
(a) \\
\inferrule*{
\Gamma, \varphi {\ensuremath{\vdash}}\psi
}{
\Gamma, x : {\ensuremath{\{ y : {\ensuremath{\mathrm{Prop}}}\;|\; \varphi \}}} {\ensuremath{\vdash}}\psi
}
}{
\Gamma {\ensuremath{\vdash}}\forall x : {\ensuremath{\{ y : {\ensuremath{\mathrm{Prop}}}\;|\; \varphi \}}} \,.\, \psi
}$$
The rule $$\supset_e:\;
\inferrule{
\Gamma {\ensuremath{\vdash}}\varphi \supset \psi \\ \Gamma {\ensuremath{\vdash}}\varphi
}{
\Gamma {\ensuremath{\vdash}}\psi
}$$ follows by $$\inferrule*{
\inferrule*{
\inferrule*{
\inferrule*{
\Gamma {\ensuremath{\vdash}}\varphi
}{
\Gamma {\ensuremath{\vdash}}\varphi : {\ensuremath{\mathrm{Prop}}}}
}{
\ldots
}
}{\Gamma {\ensuremath{\vdash}}{\ensuremath{\{ y : {\ensuremath{\mathrm{Prop}}}\;|\; \varphi \}}} : {\ensuremath{\mathrm{Type}}}} \\
\inferrule*{
\ldots }{
\Gamma {\ensuremath{\vdash}}\bot : {\ensuremath{\mathrm{Prop}}}} \\
\Gamma {\ensuremath{\vdash}}\varphi \\
y \notin FV(\Gamma, \varphi)
}{
\Gamma {\ensuremath{\vdash}}\bot : {\ensuremath{\{ y : {\ensuremath{\mathrm{Prop}}}\;|\; \varphi \}}}
} (b)$$ $$\inferrule*{
\Gamma {\ensuremath{\vdash}}\forall x : {\ensuremath{\{ y : {\ensuremath{\mathrm{Prop}}}\;|\; \varphi \}}} \,.\, \psi \\ (b)
}{
\Gamma {\ensuremath{\vdash}}\psi[x / \bot] \;\; (\equiv \psi)
}$$
Finally, the rule $$\supset_t:\;
\inferrule{
\Gamma {\ensuremath{\vdash}}\varphi : {\ensuremath{\mathrm{Prop}}}\\ \Gamma, \varphi {\ensuremath{\vdash}}\psi : {\ensuremath{\mathrm{Prop}}}}{
\Gamma {\ensuremath{\vdash}}(\varphi \supset \psi) : {\ensuremath{\mathrm{Prop}}}}$$ follows by $$\inferrule*{
\inferrule*{
}{\Gamma {\ensuremath{\vdash}}{\ensuremath{\mathrm{Prop}}}: {\ensuremath{\mathrm{Type}}}} \\
\inferrule*{
\Gamma {\ensuremath{\vdash}}\varphi : {\ensuremath{\mathrm{Prop}}}\\
y \notin FV(\Gamma, \varphi)
}{
\Gamma, y : {\ensuremath{\mathrm{Prop}}}{\ensuremath{\vdash}}\varphi : {\ensuremath{\mathrm{Prop}}}}
}{
\Gamma {\ensuremath{\vdash}}{\ensuremath{\{ y : {\ensuremath{\mathrm{Prop}}}\;|\; \varphi \}}} : {\ensuremath{\mathrm{Type}}}} (c)$$ $$\inferrule*{
(c) \\
\inferrule*{
\Gamma, \varphi {\ensuremath{\vdash}}\psi : {\ensuremath{\mathrm{Prop}}}}{
\Gamma, x : {\ensuremath{\{ y : {\ensuremath{\mathrm{Prop}}}\;|\; \varphi \}}} {\ensuremath{\vdash}}\psi : {\ensuremath{\mathrm{Prop}}}}\;\raisebox{0.7em}{($\star$)}
}{
\Gamma {\ensuremath{\vdash}}(\forall x : {\ensuremath{\{ y : {\ensuremath{\mathrm{Prop}}}\;|\; \varphi \}}} \,.\, \psi) : {\ensuremath{\mathrm{Prop}}}}$$ where we use Lemma \[lem\_01\] to perform the inference ($\star$).
Proofs for Section \[sec\_induction\] {#sec_proofs_induction}
=====================================
\[lem\_constr\_type\] If $c_i^\iota \in {\ensuremath{{\cal C}}}$ of arity $n_i$ has signature $\langle
\iota_1,\ldots,\iota_{n_i} \rangle$ then ${\ensuremath{\vdash}}_{{\ensuremath{{\cal I}}}_s} c_i^\iota :
\iota_1^* \to \ldots \to \iota_{n_i}^* \to \iota$.
Follows directly from rules $i_t^{\iota,i}$ and $\to_i$, and from axiom 4 ($\Gamma {\ensuremath{\vdash}}\iota : {\ensuremath{\mathrm{Type}}}$ for $\iota \in {\ensuremath{{\cal T}}}_I$).
\[lem\_destr\_test\_type\] ${\ensuremath{\vdash}}_{{\ensuremath{{\cal I}}}_s} o_i^\iota : \iota \to {\ensuremath{\mathrm{Prop}}}$ and ${\ensuremath{\vdash}}_{{\ensuremath{{\cal I}}}_s}
\forall x : \iota \,.\, o_i^\iota x \supset (d_{i,j}^\iota x :
\iota_{i,j}^*)$
We first show $\forall x : \iota \,.\, o_i^\iota x : {\ensuremath{\mathrm{Prop}}}$ by induction. Once we have this, to derive ${\ensuremath{\vdash}}_{{\ensuremath{{\cal I}}}_s} o_i^\iota :
\iota \to {\ensuremath{\mathrm{Prop}}}$ it suffices to apply $\forall_e$ and $\to_i$. Because ${\ensuremath{\vdash}}o_i^\iota (c_i^\iota x_1 \ldots
x_{n_i})$ by axiom 5 and ${\ensuremath{\vdash}}\neg (o_i^\iota (c_j^\iota x_1
\ldots x_{n_j}))$ for $i \ne j$ by axiom 6, we have ${\ensuremath{\vdash}}(o_i^\iota (c_i^\iota x_1 \ldots x_{n_i})) : {\ensuremath{\mathrm{Prop}}}$ using rule $p_i$, and ${\ensuremath{\vdash}}(o_i^\iota (c_j^\iota x_1 \ldots x_{n_j}))
: {\ensuremath{\mathrm{Prop}}}$ for $i \ne j$, using rules $p_i$ and $\supset_{t2}$. It is then easy to see that the premises of the rule $i_i^\iota$ are derivable for $t \equiv \lambda x \,.\, o_i^\iota x : {\ensuremath{\mathrm{Prop}}}$.
We show the second claim also by induction. Let $
\Gamma_i \equiv x_1 : \iota_{i,1}^*, \ldots, x_{n_i} :
\iota_{i,n_i}^*.
$ Let $1 \le k \le m$, where $m$ is the number of constructors of $\iota$. If $i \ne k$ then $\Gamma_k {\ensuremath{\vdash}}\neg (o_i^\iota
(c_k^\iota x_1 \ldots x_{n_k}))$. Hence, $\Gamma_k, o_i^\iota
(c_k^\iota x_1 \ldots x_{n_k}) {\ensuremath{\vdash}}d_{i,j}^\iota (c_k^\iota x_1
\ldots x_{n_k}) : \iota_{i,j}^*$ by applying rules $\supset_e$ and $\bot_e$. Since ${\ensuremath{\vdash}}o_i^\iota : \iota \to {\ensuremath{\mathrm{Prop}}}$ has already been proven in the previous paragraph, and $\Gamma_k {\ensuremath{\vdash}}(c_k^\iota x_1 \ldots x_{n_k}) : \iota$ by Lemma \[lem\_constr\_type\], we obtain $\Gamma_j {\ensuremath{\vdash}}(o_i^\iota
(c_j^\iota x_1 \ldots x_{n_j})) : {\ensuremath{\mathrm{Prop}}}$. Thus $\Gamma_k {\ensuremath{\vdash}}o_i^\iota (c_k^\iota x_1 \ldots x_{n_k}) \supset (d_{i,j}^\iota
(c_k^\iota x_1 \ldots x_{n_k}) : \iota_{i,j}^*)$ by rule $\supset_i$. If $i = k$ then $d_{i,j} (c_k^\iota x_1 \ldots
x_{n_k}) = x_j$ and thus $\Gamma_k, o_i^\iota (c_k^\iota x_1 \ldots
x_{n_k}) {\ensuremath{\vdash}}(d_{i,j} (c_k^\iota x_1 \ldots x_{n_k})) :
\iota_{i,j}^*$ by axiom 1 and rule eq. We then have ${\ensuremath{\vdash}}o_i^\iota (c_k^\iota x_1 \ldots x_{n_k})$, hence ${\ensuremath{\vdash}}(o_i^\iota
(c_k^\iota x_1 \ldots x_{n_k})) : {\ensuremath{\mathrm{Prop}}}$ by $p_i$, and thus we may use $\supset_i$ to derive $\Gamma_k {\ensuremath{\vdash}}o_i^\iota (c_k^\iota x_1
\ldots x_{n_k}) \supset (d_{i,j} (c_k^\iota x_1 \ldots x_{n_k})) :
\iota_{i,j}^*$. Therefore, by applying weakening we derive the premises of rule $i_i^\iota$. Hence, we obtain our thesis by $i_i^\iota$.
\[lem\_ind\_eq\] If $\iota \in {\ensuremath{{\cal T}}}_I$ then ${\ensuremath{\vdash}}_{{\ensuremath{{\cal I}}}_s} \forall x, y : \iota \,.\,
x =_\iota y \supset x = y$.
We prove this lemma by induction in the meta-theory on the structure of $\iota \in {\ensuremath{{\cal T}}}_I$.
Let $\varphi(x) \equiv \forall y : \iota \,.\, x =_\iota y \supset x
= y$ and $$\Gamma_i^x \equiv x_1 : \iota_{i,1}^*, \ldots, x_{n_i} :
\iota_{i,n_i}^*, \varphi(x_{j_{i,1}}), \ldots, \varphi(x_{j_{i,k_i}})$$ where $j_{i,1},\ldots,j_{i,k_i}$ are all indices $j$ such that $\iota_{i,j} = \star$, as in the conditions for rule $i_i^\iota$. It suffices to show $\Gamma_i^x {\ensuremath{\vdash}}\varphi(c_i^\iota x_1 \ldots
x_{n_i})$ for $i=1,\ldots,m$. Then ${\ensuremath{\vdash}}\forall x, y : \iota
\,.\, x =_\iota y \supset x = y$ is obtained by applying rule $i_i^\iota$.
For $i=1,\ldots,m$, we prove $\Gamma_i^x {\ensuremath{\vdash}}\varphi(c_i^\iota
x_1 \ldots x_{n_i})$, i.e. $$\Gamma_i^x {\ensuremath{\vdash}}\forall y : \iota
\,.\, (c_i^\iota x_1 \ldots x_{n_i}) =_\iota y \supset (c_i^\iota
x_1 \ldots x_{n_i}) = y$$ by induction on $y$ (in the theory, i.e. applying rule $i_i^\iota$). Let $$\Gamma_{i,j}^y \equiv y_1 : \iota_{i,1}^*, \ldots, y_{n_i} :
\iota_{i,n_i}^*, \psi_i(y_{j_{i,1}}), \ldots, \psi_i(y_{j_{i,k_i}})$$ where $$\psi_i(y) \equiv (c_i^\iota x_1 \ldots x_{n_i}) =_\iota y \supset
(c_i^\iota x_1 \ldots x_{n_i}) = y.$$ It therefore suffices to show $$(\star)\;\;\;\;\;\Gamma_i^x, \Gamma_{i,j}^y {\ensuremath{\vdash}}(c_i^\iota x_1 \ldots x_{n_i})
=_\iota (c_j^\iota y_1 \ldots y_{n_j}) \supset (c_i^\iota x_1 \ldots
x_{n_i}) = (c_j^\iota y_1 \ldots y_{n_j})$$ for all $i=1,\ldots,m$, $j=1,\ldots,m$.
Thus assume $1 \le i \le m$ and $1 \le j \le m$. By Lemma \[lem\_constr\_type\] we have $
\Gamma_i^x {\ensuremath{\vdash}}(c_i^\iota x_1 \ldots x_{n_i}) : \iota
$ and $
\Gamma_i^x, \Gamma_{i,j}^y {\ensuremath{\vdash}}(c_j^\iota y_1 \ldots y_{n_j}) : \iota.
$ Thus by Lemma \[lem\_leibniz\] we obtain $$\begin{array}{lr}
(\star\star)\;\;\;\;\; & \Gamma_i^x, \Gamma_{i,j}^y {\ensuremath{\vdash}}((c_i^\iota x_1 \ldots
x_{n_i}) =_\iota (c_j^\iota y_1 \ldots y_{n_j})) : {\ensuremath{\mathrm{Prop}}}.
\end{array}$$ For $k=1,\ldots,n_i$ take $$f_k = \lambda x \,.\, {\ensuremath{\mathrm{Cond}\,(o_i^\iota x)\,(d_{i,k}^\iota x
=_{\iota_{i,k}^*} x_k)\,\bot\,}}.$$
Assume $1 \le k \le n_i$. Since $o_i^\iota : \iota \to {\ensuremath{\mathrm{Prop}}}$ by Lemma \[lem\_destr\_test\_type\], we obtain $x : \iota {\ensuremath{\vdash}}(o_i^\iota x) \lor \neg (o_i^\iota x)$ by $\to_e$ and the law of excluded middle. It is easy to derive $x : \iota, \neg (o_i^\iota x)
{\ensuremath{\vdash}}f_k x : {\ensuremath{\mathrm{Prop}}}$ using the rules for and . By Lemma \[lem\_destr\_test\_type\] we also have $x
: \iota, o_i^\iota x {\ensuremath{\vdash}}d_{i,k}^\iota x : \iota_{i,k}^*$. We have $\Gamma_i^x {\ensuremath{\vdash}}x_k : \iota_{i,k}^*$ by definition of $\Gamma_i^x$. Then $\Gamma_i^x, x : \iota, o_i^\iota x {\ensuremath{\vdash}}(d_{i,k}^\iota x =_{\iota_{i,k}^*} x_k) : {\ensuremath{\mathrm{Prop}}}$ follows from Lemma \[lem\_leibniz\]. Using rule $\vee_e$ it is now easy to derive $\Gamma_i^x, x : \iota {\ensuremath{\vdash}}f_k x : {\ensuremath{\mathrm{Prop}}}$. Since ${\ensuremath{\vdash}}\iota : {\ensuremath{\mathrm{Type}}}$, we obtain $\Gamma_i^x {\ensuremath{\vdash}}f_k : \iota \to {\ensuremath{\mathrm{Prop}}}$ by rule $\to_i$.
Suppose $i = j$. Assume $1 \le k \le n_i$. Because $\Gamma_i^x
{\ensuremath{\vdash}}f_k : \iota \to {\ensuremath{\mathrm{Prop}}}$, by definition of $=_\iota$ we have $$\Gamma_i^x,\Gamma_{i,i}^y,(c_i^\iota x_1 \ldots x_{n_i})
=_\iota (c_i^\iota y_1 \ldots y_{n_i}) {\ensuremath{\vdash}}(f_k (c_i^\iota x_1
\ldots x_{n_i})) \supset (f_k (c_i^\iota y_1 \ldots y_{n_i}))$$ We have $
\Gamma_i^x {\ensuremath{\vdash}}(f_k (c_i^\iota x_1
\ldots x_{n_i})) = (x_k =_\iota x_k)
$. Hence $ \Gamma_i^x {\ensuremath{\vdash}}f_k (c_i^\iota x_1 \ldots x_{n_i})
$. Because $$\Gamma_i^x, \Gamma_{i,i}^y {\ensuremath{\vdash}}(f_k (c_i^\iota y_1 \ldots
y_{n_i})) = (y_k =_{\iota_{i,k}^*} x_k)$$ we thus obtain $$\Gamma_i^x,\Gamma_{i,i}^y,(c_i^\iota x_1 \ldots x_{n_i}) =_\iota
(c_i^\iota y_1 \ldots y_{n_i}) {\ensuremath{\vdash}}y_k =_{\iota_{i,k}^*} x_k.$$ If $\iota_{i,k}^* = \iota$ then $\Gamma_i^x {\ensuremath{\vdash}}\forall y :
\iota \,.\, x_k =_\iota y \supset x_k = y$ and $\Gamma_i^x,
\Gamma_{i,i}^y {\ensuremath{\vdash}}y_k : \iota$. Hence $\Gamma_i^x,\Gamma_{i,i}^y {\ensuremath{\vdash}}x_k = y_k$. If $\iota_{i,k}^* =
\iota_{i,k}$ then by the inductive hypothesis in the meta-theory we obtain ${\ensuremath{\vdash}}\forall x, y : \iota_{i,k} \,.\, x =_{\iota_{i,k}} y
\supset x = y$, which again implies $\Gamma_i^x,\Gamma_{i,i}^y
{\ensuremath{\vdash}}x_k = y_k$, because $\Gamma_i^x {\ensuremath{\vdash}}x_k : \iota_{i,k}$ and $\Gamma_{i,i}^y {\ensuremath{\vdash}}y_k : \iota_{i,k}$. Since $1 \le k \le
n_i$ was arbitrary, we obtain $$\Gamma_i^x,\Gamma_{i,i}^y,(c_i^\iota x_1 \ldots x_{n_i}) =_\iota
(c_i^\iota y_1 \ldots y_{n_i}) {\ensuremath{\vdash}}(c_i^\iota x_1 \ldots x_{n_i}) =
(c_i^\iota y_1 \ldots y_{n_i}).$$ By $(\star\star)$ and $\supset_i$ we obtain $(\star)$ for $i=j$.
Suppose $i \ne j$. Assume $1 \le k \le n_i$. Because $\Gamma_i^x
{\ensuremath{\vdash}}f_k : \iota \to {\ensuremath{\mathrm{Prop}}}$, by definition of $=_\iota$ we have $$\Gamma_i^x,\Gamma_{i,j}^y,(c_i^\iota x_1 \ldots x_{n_i})
=_\iota (c_j^\iota y_1 \ldots y_{n_i}) {\ensuremath{\vdash}}(f_k (c_i^\iota x_1
\ldots x_{n_i})) \supset (f_k (c_j^\iota y_1 \ldots y_{n_j}))$$ As in the case $i = j$ we have $\Gamma_i^x {\ensuremath{\vdash}}(f_k (c_i^\iota x_1
\ldots x_{n_i}))$. Because $i \ne j$ we have $$\Gamma_i^x,\Gamma_{i,j}^\iota {\ensuremath{\vdash}}(f_k (c_j^\iota y_1 \ldots
y_{n_j})) = \bot.$$ Therefore $$\Gamma_i^x,\Gamma_{i,j}^y,(c_i^\iota x_1 \ldots x_{n_i})
=_\iota (c_j^\iota y_1 \ldots y_{n_i}) {\ensuremath{\vdash}}\bot.$$ By rule $\bot_e$ we conclude $$\Gamma_i^x,\Gamma_{i,j}^y,(c_i^\iota x_1 \ldots x_{n_i})
=_\iota (c_j^\iota y_1 \ldots y_{n_i}) {\ensuremath{\vdash}}(c_i^\iota x_1 \ldots x_{n_i}) =
(c_j^\iota y_1 \ldots y_{n_j})$$ By $(\star\star)$ and $\supset_i$ we obtain $(\star)$ for $i \ne
j$. This finishes the proof.
\[lem\_nat\_peano\_3\] ${\ensuremath{\vdash}}_{{\ensuremath{{\cal I}}}_s} \forall x : {\ensuremath{\mathrm{Nat}}}\,.\, \left(({\ensuremath{\mathfrak{0}}}x) \vee \exists y
: {\ensuremath{\mathrm{Nat}}}\,.\, x = ({\ensuremath{\mathfrak{s}}}y)\right)$.
Recall that $x = {\ensuremath{\mathfrak{s}}}y$ stands for ${\ensuremath{\mathrm{Eq}\,x\,({\ensuremath{\mathfrak{s}}}y)\,}}$. Let $\varphi(x)
\equiv ({\ensuremath{\mathfrak{0}}}x) \vee \exists y : {\ensuremath{\mathrm{Nat}}}\,.\, x = {\ensuremath{\mathfrak{s}}}y$. We have the following derivation. $$\inferrule*{
\inferrule*{
{\ensuremath{\vdash}}{\ensuremath{\mathfrak{0}}}0
}{
{\ensuremath{\vdash}}\varphi(0)
} \\
\inferrule*{
\inferrule*{
x : {\ensuremath{\mathrm{Nat}}}, \varphi(x) {\ensuremath{\vdash}}{\ensuremath{\mathfrak{s}}}x = {\ensuremath{\mathfrak{s}}}x
\\ x : {\ensuremath{\mathrm{Nat}}}, \varphi(x) {\ensuremath{\vdash}}x : {\ensuremath{\mathrm{Nat}}}}{
x : {\ensuremath{\mathrm{Nat}}}, \varphi(x) {\ensuremath{\vdash}}\exists y : {\ensuremath{\mathrm{Nat}}}\,.\, {\ensuremath{\mathfrak{s}}}x = {\ensuremath{\mathfrak{s}}}y
}
}{
x : {\ensuremath{\mathrm{Nat}}}, \varphi(x) {\ensuremath{\vdash}}\varphi({\ensuremath{\mathfrak{s}}}x)
}
}{
{\ensuremath{\vdash}}\forall x : {\ensuremath{\mathrm{Nat}}}\,.\, \left(({\ensuremath{\mathfrak{0}}}x) \vee \exists y : {\ensuremath{\mathrm{Nat}}}\,.\, x = {\ensuremath{\mathfrak{s}}}y\right)
}\;\raisebox{0.7em}{by $n_i$}$$
\[lem\_iszero\_eq\] ${\ensuremath{\vdash}}_{{\ensuremath{{\cal I}}}_s} \forall x : {\ensuremath{\mathrm{Nat}}}\,.\, ({\ensuremath{\mathfrak{0}}}x) \supset (x = 0)$.
Let $\varphi(x) \equiv ({\ensuremath{\mathfrak{0}}}x) \supset (x = 0)$. $$\inferrule*{
\inferrule*{
{\ensuremath{\vdash}}{\ensuremath{\mathfrak{0}}}0
}{
{\ensuremath{\vdash}}({\ensuremath{\mathfrak{0}}}0) : {\ensuremath{\mathrm{Prop}}}} \\
{\ensuremath{\mathfrak{0}}}0 {\ensuremath{\vdash}}0 = 0
}{
{\ensuremath{\vdash}}\varphi(0)
}(a)$$ $$\inferrule*{
\inferrule*{
\inferrule*{
x : {\ensuremath{\mathrm{Nat}}}, \varphi(x) {\ensuremath{\vdash}}\neg ({\ensuremath{\mathfrak{0}}}({\ensuremath{\mathfrak{s}}}x))
}{
x : {\ensuremath{\mathrm{Nat}}}, \varphi(x) {\ensuremath{\vdash}}\neg ({\ensuremath{\mathfrak{0}}}({\ensuremath{\mathfrak{s}}}x)) : {\ensuremath{\mathrm{Prop}}}}
}{
x : {\ensuremath{\mathrm{Nat}}}, \varphi(x) {\ensuremath{\vdash}}{\ensuremath{\mathfrak{0}}}({\ensuremath{\mathfrak{s}}}x) : {\ensuremath{\mathrm{Prop}}}}
\\
\inferrule*{
x : {\ensuremath{\mathrm{Nat}}}, \varphi(x), {\ensuremath{\mathfrak{0}}}({\ensuremath{\mathfrak{s}}}x) {\ensuremath{\vdash}}\bot
}{
x : {\ensuremath{\mathrm{Nat}}}, \varphi(x), {\ensuremath{\mathfrak{0}}}({\ensuremath{\mathfrak{s}}}x) {\ensuremath{\vdash}}{\ensuremath{\mathfrak{s}}}x = 0
}
}{
x : {\ensuremath{\mathrm{Nat}}}, \varphi(x) {\ensuremath{\vdash}}\varphi({\ensuremath{\mathfrak{s}}}x)
} (b)$$ $$\inferrule*{
(a)
\\
(b)
}{
{\ensuremath{\vdash}}\forall x : {\ensuremath{\mathrm{Nat}}}\,.\, ({\ensuremath{\mathfrak{0}}}x) \supset (x = 0)
}$$
The following terms are derivable in the system ${\ensuremath{{\cal I}}}_s$:
- $\forall x, y : {\ensuremath{\mathrm{Nat}}}\,.\, (x + y) : {\ensuremath{\mathrm{Nat}}}$,
- $\forall x, y : {\ensuremath{\mathrm{Nat}}}\,.\, (x - y) : {\ensuremath{\mathrm{Nat}}}$,
- $\forall x, y : {\ensuremath{\mathrm{Nat}}}\,.\, (x {\ensuremath{\cdot}}y) : {\ensuremath{\mathrm{Nat}}}$,
- $\forall x, y : {\ensuremath{\mathrm{Nat}}}\,.\, (x \le y) : {\ensuremath{\mathrm{Prop}}}$,
- $\forall x, y : {\ensuremath{\mathrm{Nat}}}\,.\, (x < y) : {\ensuremath{\mathrm{Prop}}}$.
We give a sample proof for the first term. $$\inferrule*{
\inferrule*{
x : {\ensuremath{\mathrm{Nat}}}{\ensuremath{\vdash}}(x + 0) = ((\lambda x \,.\, x) x)
}{
x : {\ensuremath{\mathrm{Nat}}}{\ensuremath{\vdash}}(x + 0) = x
} \\ x : {\ensuremath{\mathrm{Nat}}}{\ensuremath{\vdash}}x : {\ensuremath{\mathrm{Nat}}}}{
x : {\ensuremath{\mathrm{Nat}}}{\ensuremath{\vdash}}(x + 0) : {\ensuremath{\mathrm{Nat}}}} \; (a)$$ $$\inferrule*{
x : {\ensuremath{\mathrm{Nat}}}, y : {\ensuremath{\mathrm{Nat}}}, (x + y) : {\ensuremath{\mathrm{Nat}}}{\ensuremath{\vdash}}(x + y) : {\ensuremath{\mathrm{Nat}}}}{
x : {\ensuremath{\mathrm{Nat}}}, y : {\ensuremath{\mathrm{Nat}}}, (x + y) : {\ensuremath{\mathrm{Nat}}}{\ensuremath{\vdash}}({\ensuremath{\mathfrak{s}}}(x + y)) : {\ensuremath{\mathrm{Nat}}}} \; (b)$$ $$\inferrule*{
x : {\ensuremath{\mathrm{Nat}}}, y : {\ensuremath{\mathrm{Nat}}}, (x + y) : {\ensuremath{\mathrm{Nat}}}{\ensuremath{\vdash}}x + ({\ensuremath{\mathfrak{s}}}y) = {\ensuremath{\mathfrak{s}}}(x + y)
\\
(b)
}{
x : {\ensuremath{\mathrm{Nat}}}, y : {\ensuremath{\mathrm{Nat}}}, (x + y) : {\ensuremath{\mathrm{Nat}}}{\ensuremath{\vdash}}(x + ({\ensuremath{\mathfrak{s}}}y)) : {\ensuremath{\mathrm{Nat}}}} \; (c)$$ $$\inferrule*{
\inferrule*{
(a)
\\ (c)
}{
x : {\ensuremath{\mathrm{Nat}}}{\ensuremath{\vdash}}\forall y : {\ensuremath{\mathrm{Nat}}}\,.\, (x + y) : {\ensuremath{\mathrm{Nat}}}} \\ {\ensuremath{\vdash}}{\ensuremath{\mathrm{Nat}}}: {\ensuremath{\mathrm{Type}}}}{
{\ensuremath{\vdash}}\forall x, y : {\ensuremath{\mathrm{Nat}}}\,.\, (x + y) : {\ensuremath{\mathrm{Nat}}}}$$ The second and third terms are derived in a similar way. The fourth and fifth terms are derived from the second using Lemma \[lem\_constr\_type\].
\[lem\_s\_minus\] ${\ensuremath{\vdash}}_{{\ensuremath{{\cal I}}}_s} \forall x, y : {\ensuremath{\mathrm{Nat}}}\,.\, \left({\ensuremath{\mathfrak{s}}}x - {\ensuremath{\mathfrak{s}}}y = x -
y\right)$
We assume $x : {\ensuremath{\mathrm{Nat}}}$ and show $\forall y : {\ensuremath{\mathrm{Nat}}}\,.\, \left({\ensuremath{\mathfrak{s}}}x -
{\ensuremath{\mathfrak{s}}}y = x - y\right)$ by induction on $y$. For $y = 0$ we have ${\ensuremath{\mathfrak{s}}}x
- {\ensuremath{\mathfrak{s}}}0 = {\ensuremath{\mathfrak{p}}}({\ensuremath{\mathfrak{s}}}x - 0) = {\ensuremath{\mathfrak{p}}}({\ensuremath{\mathfrak{s}}}x) = x = x - 0$, and under the assumptions $x :{\ensuremath{\mathrm{Nat}}}, y : {\ensuremath{\mathrm{Nat}}}, \left({\ensuremath{\mathfrak{s}}}x - {\ensuremath{\mathfrak{s}}}y\right) = x - y$ we may derive ${\ensuremath{\mathfrak{s}}}x - {\ensuremath{\mathfrak{s}}}({\ensuremath{\mathfrak{s}}}y) = {\ensuremath{\mathfrak{p}}}({\ensuremath{\mathfrak{s}}}x - {\ensuremath{\mathfrak{s}}}y) = {\ensuremath{\mathfrak{p}}}(x - y) = x -
{\ensuremath{\mathfrak{s}}}y$. By rule $n_i$ we obtain the thesis.
\[lem\_le\_ge\_then\_eq\] ${\ensuremath{\vdash}}_{{\ensuremath{{\cal I}}}_s} \forall x, y : {\ensuremath{\mathrm{Nat}}}\,.\, (x \ge y) \wedge (x \le y)
\supset (x =_{\ensuremath{\mathrm{Nat}}}y)$.
Let $\varphi(x) \equiv \forall y : {\ensuremath{\mathrm{Nat}}}\,.\, (x \ge y) \wedge (x
\le y) \supset (x =_{\ensuremath{\mathrm{Nat}}}y)$. We proceed by induction on $x$.
In the base step we need to show $\varphi(0)$. We assume $y : {\ensuremath{\mathrm{Nat}}}$ and $(0 \ge y) \wedge (0 \le y)$. From $y \le 0$ we have ${\ensuremath{\mathfrak{0}}}(y
- 0)$, i.e. ${\ensuremath{\mathfrak{0}}}y$, and thus $y = 0$ by Lemma \[lem\_iszero\_eq\]. By rule $p_i$ we have $((0 \ge y) \wedge
(0 \le y)) : {\ensuremath{\mathrm{Prop}}}$. Hence, we may use implication introduction and then universal quantifier introduction to obtain $\varphi(0)$.
In the inductive step we need to prove $\varphi({\ensuremath{\mathfrak{s}}}x)$ under the assumptions $x : {\ensuremath{\mathrm{Nat}}}$ and $\varphi(x)$. We assume further $y :
{\ensuremath{\mathrm{Nat}}}$ and $({\ensuremath{\mathfrak{s}}}x \ge y) \wedge ({\ensuremath{\mathfrak{s}}}x \le y)$. By Lemma \[lem\_nat\_peano\_3\] there are two possibilities: $y = 0$ or $\exists z : {\ensuremath{\mathrm{Nat}}}\,.\, y = {\ensuremath{\mathfrak{s}}}z$. If $y = 0$ then we easily obtain ${\ensuremath{\mathfrak{s}}}x = 0$, which leads to a contradiction, from which we may derive $x = y$. If $y = {\ensuremath{\mathfrak{s}}}z$ then we have ${\ensuremath{\mathfrak{s}}}x - {\ensuremath{\mathfrak{s}}}z = 0$ and ${\ensuremath{\mathfrak{s}}}z - {\ensuremath{\mathfrak{s}}}x = 0$. By Lemma \[lem\_s\_minus\] we obtain $x - z = 0$ and $z - x
= 0$. From the inductive hypothesis we have $x =_{\ensuremath{\mathrm{Nat}}}z$, and hence $x = z$ by Lemma \[lem\_ind\_eq\]. Thus ${\ensuremath{\mathfrak{s}}}x = {\ensuremath{\mathfrak{s}}}z = y$. Since $x
: {\ensuremath{\mathrm{Nat}}}$ and $y : {\ensuremath{\mathrm{Nat}}}$, it is not difficult to show that $(({\ensuremath{\mathfrak{s}}}x \ge
y) \wedge ({\ensuremath{\mathfrak{s}}}x \le y)) : {\ensuremath{\mathrm{Prop}}}$. We may therefore use implication introduction and then universal quantifier introduction to obtain $\varphi({\ensuremath{\mathfrak{s}}}x)$.
The following terms are derivable in the system ${\ensuremath{{\cal I}}}_s$:
- $\forall x, y : {\ensuremath{\mathrm{Nat}}}\,.\, \left({\ensuremath{\mathfrak{s}}}x =_{\ensuremath{\mathrm{Nat}}}{\ensuremath{\mathfrak{s}}}y\right)
\supset \left(x =_{\ensuremath{\mathrm{Nat}}}y\right)$,
- $\forall x : {\ensuremath{\mathrm{Nat}}}\,.\, \neg ({\ensuremath{\mathfrak{s}}}x =_{\ensuremath{\mathrm{Nat}}}0)$,
- $\forall x : {\ensuremath{\mathrm{Nat}}}\,.\, (x + 0 =_{\ensuremath{\mathrm{Nat}}}x)$,
- $\forall x, y : {\ensuremath{\mathrm{Nat}}}\,.\, (x + {\ensuremath{\mathfrak{s}}}y =_{\ensuremath{\mathrm{Nat}}}{\ensuremath{\mathfrak{s}}}(x + y))$,
- $\forall x : {\ensuremath{\mathrm{Nat}}}\,.\, (x {\ensuremath{\cdot}}0 =_{\ensuremath{\mathrm{Nat}}}0)$,
- $\forall x, y : {\ensuremath{\mathrm{Nat}}}\,.\, (x {\ensuremath{\cdot}}({\ensuremath{\mathfrak{s}}}y) =_{\ensuremath{\mathrm{Nat}}}(x {\ensuremath{\cdot}}y)
+ x)$.
All proofs are easy. As an example we give an indication of how the first term may be derived. We assume $x : {\ensuremath{\mathrm{Nat}}}$, $y : {\ensuremath{\mathrm{Nat}}}$ and ${\ensuremath{\mathfrak{s}}}x =_{\ensuremath{\mathrm{Nat}}}{\ensuremath{\mathfrak{s}}}y$. From ${\ensuremath{\mathfrak{s}}}x =_{\ensuremath{\mathrm{Nat}}}{\ensuremath{\mathfrak{s}}}y$ we obtain ${\ensuremath{\mathfrak{s}}}x = {\ensuremath{\mathfrak{s}}}y$ by Lemma \[lem\_ind\_eq\]. We thus obtain $x = {\ensuremath{\mathfrak{p}}}({\ensuremath{\mathfrak{s}}}x) = {\ensuremath{\mathfrak{p}}}({\ensuremath{\mathfrak{s}}}y) =
y$. Since $x : {\ensuremath{\mathrm{Nat}}}$ and $y : {\ensuremath{\mathrm{Nat}}}$ we have $x =_{\ensuremath{\mathrm{Nat}}}y$ by Lemma \[lem\_leibniz\]. We also have $({\ensuremath{\mathfrak{s}}}x =_{\ensuremath{\mathrm{Nat}}}{\ensuremath{\mathfrak{s}}}y) : {\ensuremath{\mathrm{Prop}}}$ by Lemma \[lem\_constr\_type\] and Lemma \[lem\_nat\_op\_well\_defined\]. Therefore, we may use implication introduction, and then universal quantifier introduction twice, to show $\forall x, y : {\ensuremath{\mathrm{Nat}}}\,.\, \left({\ensuremath{\mathfrak{s}}}x =_{\ensuremath{\mathrm{Nat}}}{\ensuremath{\mathfrak{s}}}y\right) \supset \left(x =_{\ensuremath{\mathrm{Nat}}}y\right)$.
\[lem\_nat\_less\] $\Gamma {\ensuremath{\vdash}}_{{\ensuremath{{\cal I}}}_s} \forall x, y : {\ensuremath{\mathrm{Nat}}}\,.\, x < {\ensuremath{\mathfrak{s}}}y \supset (x
< y \lor x = y)$
Easy induction on $x$, recalling that $x < y \equiv {\ensuremath{\mathfrak{0}}}(x -
y)$, using Lemma \[lem\_nat\_peano\_3\] for the basis, and Lemma \[lem\_s\_minus\] in the inductive step.
\[lem\_nat\_ind\_2\] The following rule is admissible in ${\ensuremath{{\cal I}}}_s$. $${n_i':}\;\;\;\inferrule{
\Gamma {\ensuremath{\vdash}}t 0 \\ \Gamma, y : {\ensuremath{\mathrm{Nat}}}, \forall x : {\ensuremath{\mathrm{Nat}}}\,.\, x <
y \supset t x {\ensuremath{\vdash}}t y \\ x,y \notin FV(\Gamma, t)
}{
\Gamma {\ensuremath{\vdash}}\forall x : {\ensuremath{\mathrm{Nat}}}\,.\, t x
}$$
We show that if the premises are derivable, then so is $\Gamma
{\ensuremath{\vdash}}\forall y, x : {\ensuremath{\mathrm{Nat}}}\,.\, x < y \supset t x$. We achieve this by deriving the premises of the rule $n_i$ for $\psi(y) \equiv
\forall x : {\ensuremath{\mathrm{Nat}}}\,.\, x < y \supset t x$. Since $x : {\ensuremath{\mathrm{Nat}}}, x < 0
{\ensuremath{\vdash}}\bot$ and $x : {\ensuremath{\mathrm{Nat}}}{\ensuremath{\vdash}}(x < 0) : {\ensuremath{\mathrm{Prop}}}$, we have $\Gamma {\ensuremath{\vdash}}\psi(0)$. By our assumption that the second premise of $n_i'$ is derivable, we have $\Gamma, y : {\ensuremath{\mathrm{Nat}}}, \psi(y) {\ensuremath{\vdash}}t
y$. Let $\Gamma_1 \equiv \Gamma, y : {\ensuremath{\mathrm{Nat}}}, \psi(y), x : {\ensuremath{\mathrm{Nat}}}, x < {\ensuremath{\mathfrak{s}}}y$. We show $\Gamma_1 {\ensuremath{\vdash}}t x$. By Lemma \[lem\_nat\_less\] we obtain $\Gamma_1 {\ensuremath{\vdash}}x < y \lor x = y$. Since $\Gamma_1 {\ensuremath{\vdash}}\forall x : {\ensuremath{\mathrm{Nat}}}\,.\, x < y \supset t x$ we have $\Gamma_1, x < y
{\ensuremath{\vdash}}t x$. Since $\Gamma_1 {\ensuremath{\vdash}}t y$ we have $\Gamma_1, x = y
{\ensuremath{\vdash}}t x$. Using rule $\lor_e$ we obtain $\Gamma_1 {\ensuremath{\vdash}}t
x$. Since $\Gamma, y : {\ensuremath{\mathrm{Nat}}}, \psi(y), x : {\ensuremath{\mathrm{Nat}}}{\ensuremath{\vdash}}(x < {\ensuremath{\mathfrak{s}}}y) :
{\ensuremath{\mathrm{Prop}}}$ we have $\Gamma, y : {\ensuremath{\mathrm{Nat}}}, \psi(y) {\ensuremath{\vdash}}\forall x : {\ensuremath{\mathrm{Nat}}}\,.\, x < {\ensuremath{\mathfrak{s}}}y \supset t x$, i.e. $\Gamma, y : {\ensuremath{\mathrm{Nat}}}, \psi(y) {\ensuremath{\vdash}}\psi({\ensuremath{\mathfrak{s}}}y)$. By applying rule $n_i$ we thus obtain $\Gamma {\ensuremath{\vdash}}\forall y,x : {\ensuremath{\mathrm{Nat}}}\,.\, x < y \supset t x$. Now $\Gamma, x : {\ensuremath{\mathrm{Nat}}}{\ensuremath{\vdash}}{\ensuremath{\mathfrak{s}}}x : {\ensuremath{\mathrm{Nat}}}$ and $\Gamma, x : {\ensuremath{\mathrm{Nat}}}{\ensuremath{\vdash}}x < {\ensuremath{\mathfrak{s}}}x$, hence $\Gamma, x : {\ensuremath{\mathrm{Nat}}}{\ensuremath{\vdash}}t x$. Therefore $\Gamma {\ensuremath{\vdash}}\forall x
: {\ensuremath{\mathrm{Nat}}}\,.\, t x$.
Suppose $\Gamma {\ensuremath{\vdash}}\forall x_1 : \alpha_1 \ldots \forall x_n :
\alpha_n \,.\, \varphi_1 \lor \ldots \lor \varphi_m$, $\Gamma
{\ensuremath{\vdash}}\alpha_j : {\ensuremath{\mathrm{Type}}}$ for $j = 1,\ldots,n$, and for $i=1,\ldots,m$: $\Gamma {\ensuremath{\vdash}}\forall x_1 : \alpha_1 \ldots
\forall x_n : \alpha_n \,.\, t_i : \beta \to \ldots \to \beta$ where $\beta$ occurs $k_i + 1$ times, $\Gamma {\ensuremath{\vdash}}\forall x_1 :
\alpha_1 \ldots \forall x_n : \alpha_n \,.\, t_{i,j,k} : \alpha_k$ for $j=1,\ldots,k_i$, $k=1,\ldots,n$, $x_1,\ldots,x_n \notin
FV(f,\alpha_1,\ldots,\alpha_n,\beta)$ and $$\begin{aligned}
\Gamma {\ensuremath{\vdash}}\forall x_1 : \alpha_1 \ldots \forall x_n :
\alpha_n \,.\, \varphi_i &\supset& (f x_1 \ldots x_n = \\ && \;t_i (f t_{i,1,1}
\ldots t_{i,1,n}) \ldots (f t_{i,k_i,1} \ldots t_{i,k_i,n})).
\end{aligned}$$ If there is a term $g$ such that $\Gamma {\ensuremath{\vdash}}g : \alpha_1 \to
\ldots \to \alpha_n \to {\ensuremath{\mathrm{Nat}}}$ and for $i=1,\ldots,m$ $$\begin{aligned}
\Gamma {\ensuremath{\vdash}}\forall x_1 : \alpha_1 \ldots \forall x_n : \alpha_n
\,.\, \varphi_i &\supset& (\left((f x_1 \ldots x_n) :
\beta\right) \lor \\ && \;((g t_{i,1,1} \ldots t_{i,1,n}) < (g x_1 \ldots
x_n) \land \ldots \land \\ && \;\;(g t_{i,k_i,1} \ldots t_{i,k_i,n}) < (g x_1
\ldots x_n)))
\end{aligned}$$ where $x_1,\ldots,x_n \notin FV(g)$, then $$\Gamma {\ensuremath{\vdash}}f : \alpha_1 \to \ldots \to \alpha_n \to \beta.$$
We derive the premises of the rule $n_i'$ for $$\psi(x) \equiv \forall x_1 : \alpha_1 \ldots \forall x_n : \alpha_n
\,.\, x =_{\ensuremath{\mathrm{Nat}}}(g x_1 \ldots x_n) \supset f x_1 \ldots x_n : \beta.$$ We have $\Gamma, x_1 : \alpha_1,\ldots,x_n : \alpha_n {\ensuremath{\vdash}}\varphi_1 \lor \ldots \lor \varphi_m$. Let $\Gamma_i \equiv \Gamma,
x_1 : \alpha_1,\ldots,x_n : \alpha_n,\varphi_i$ for $i=1,\ldots,m$. We first prove $\Gamma_i, 0 =_{\ensuremath{\mathrm{Nat}}}(g x_1 \ldots
x_n) {\ensuremath{\vdash}}f x_1 \ldots x_n : \beta$. It suffices to show $$\begin{aligned}
\Gamma_i, 0 =_{\ensuremath{\mathrm{Nat}}}(g x_1 \ldots x_n), && ((g t_{i,1,1} \ldots
t_{i,1,n}) < (g x_1 \ldots x_n) \land \ldots \land \\ && \;(g
t_{i,k_i,1} \ldots t_{i,k_i,n}) < (g x_1 \ldots x_n)) {\ensuremath{\vdash}}f x_1
\ldots x_n : \beta.
\end{aligned}$$ Since ${\ensuremath{\vdash}}g : \alpha_1 \to \ldots \to \alpha_n \to {\ensuremath{\mathrm{Nat}}}$, by Lemma \[lem\_ind\_eq\] and ${\ensuremath{\vdash}}\neg ({\ensuremath{\mathfrak{0}}}({\ensuremath{\mathfrak{s}}}(g
t_{i,1,1} \ldots t_{i,1,n})))$ we have $$\begin{aligned}
\Gamma_i, 0 =_{\ensuremath{\mathrm{Nat}}}(g x_1 \ldots x_n), && ((g t_{i,1,1} \ldots
t_{i,1,n}) < (g x_1 \ldots x_n) \land \ldots \land \\ && \;(g t_{i,k_i,1}
\ldots t_{i,k_i,n}) < (g x_1 \ldots x_n)) {\ensuremath{\vdash}}\bot.
\end{aligned}$$ so our claim follows by $\bot_e$. Therefore $\Gamma_i {\ensuremath{\vdash}}0
=_{\ensuremath{\mathrm{Nat}}}(g x_1 \ldots x_n) \supset (f x_1 \ldots x_n : \beta)$ by ${\ensuremath{\vdash}}g : \alpha_1 \to \ldots \to \alpha_n \to {\ensuremath{\mathrm{Nat}}}$, Lemma \[lem\_nat\_op\_well\_defined\] and $\supset_i$. Since $$\Gamma, x_1 : \alpha_1, \ldots, x_n : \alpha_n {\ensuremath{\vdash}}\varphi_1
\lor \ldots \lor \varphi_m$$ by $m-1$ applications of $\lor_e$ we obtain $$\Gamma, x_1 : \alpha_1, \ldots, x_n : \alpha_n {\ensuremath{\vdash}}0 =_{\ensuremath{\mathrm{Nat}}}(g
x_1 \ldots x_n) \supset (f x_1 \ldots x_n : \beta).$$ Because $\Gamma {\ensuremath{\vdash}}\alpha_i : {\ensuremath{\mathrm{Type}}}$ for $i=1,\ldots,n$, we may apply $\forall$-introduction $n$ times to obtain $\Gamma {\ensuremath{\vdash}}\psi(0)$.
Let $$\Gamma' \equiv \Gamma, y : {\ensuremath{\mathrm{Nat}}}, \forall x : {\ensuremath{\mathrm{Nat}}}\,.\, x < y
\supset \psi(x)$$ and $$\Gamma_i' \equiv \Gamma', x_1 : \alpha_1, \ldots, x_n : \alpha_n,
\varphi_i.$$ Because $\Gamma', x_1 : \alpha, \ldots, x_n : \alpha_n {\ensuremath{\vdash}}\varphi_1 \lor \ldots \lor \varphi_m$ and $\Gamma' {\ensuremath{\vdash}}\alpha_i
: {\ensuremath{\mathrm{Type}}}$ for $i=1,\ldots,n$, to derive the second premise of $n_i'$ for $\psi$, it suffices to show $$\Gamma_i' {\ensuremath{\vdash}}y =_{\ensuremath{\mathrm{Nat}}}(g x_1 \ldots x_n) \supset f x_1 \ldots
x_n : \beta$$ for $i=1,\ldots,m$, apply $\lor$-elimination $m-1$ times, and then $\forall$-introduction $n$ times. Let $\Gamma_i'' \equiv \Gamma_i',
y =_{\ensuremath{\mathrm{Nat}}}(g x_1 \ldots x_n)$. Since $$\Gamma_i' {\ensuremath{\vdash}}(y =_{\ensuremath{\mathrm{Nat}}}(g x_1 \ldots x_n)) : {\ensuremath{\mathrm{Prop}}}$$ it actually suffices to prove $ \Gamma_i'' {\ensuremath{\vdash}}f x_1 \ldots x_n
: \beta $ and apply $\supset_i$. By the assumption on $g$ in the lemma, we have $$\begin{aligned}
\Gamma_i'' {\ensuremath{\vdash}}\left((f x_1 \ldots x_n) : \beta\right) &\lor& ((g
t_{i,1,1} \ldots t_{i,1,n}) < (g x_1 \ldots x_n) \land \ldots \land
\\ && \;(g t_{i,k_i,1} \ldots t_{i,k_i,n}) < (g x_1 \ldots x_n))
\end{aligned}$$ Let $$\Gamma_i^+ \equiv \Gamma_i'', (g t_{i,1,1} \ldots t_{i,1,n}) < (g
x_1 \ldots x_n) \land \ldots \land (g t_{i,k_i,1} \ldots
t_{i,k_i,n}) < (g x_1 \ldots x_n)$$ It thus suffices to show $\Gamma_i^+ {\ensuremath{\vdash}}f x_1 \ldots x_n :
\beta$ and apply rule $\lor_e$. Let $1 \le j \le k_i$. Since $\Gamma_i^+ {\ensuremath{\vdash}}t_{i,j,k} : \alpha_k$ for $k=1,\ldots,n$ and $\Gamma_i^+ {\ensuremath{\vdash}}g : \alpha_1 \to \ldots \to \alpha_n \to {\ensuremath{\mathrm{Nat}}}$, we have $\Gamma_i^+ {\ensuremath{\vdash}}(g t_{i,j,1} \ldots t_{i,j,n}) :
{\ensuremath{\mathrm{Nat}}}$. Thus $$\Gamma_i^+ {\ensuremath{\vdash}}(g t_{i,j,1} \ldots t_{i,j,n}) < y \supset \psi(g
t_{i,j,1} \ldots t_{i,j,n}).$$ But $\Gamma_i^+ {\ensuremath{\vdash}}y =_{\ensuremath{\mathrm{Nat}}}(g x_1 \ldots x_n)$, $\Gamma_i^+
{\ensuremath{\vdash}}y : {\ensuremath{\mathrm{Nat}}}$ and $\Gamma_i^+ {\ensuremath{\vdash}}(g x_1 \ldots x_n) :
{\ensuremath{\mathrm{Nat}}}$, so by Lemma \[lem\_ind\_eq\] we obtain $\Gamma_i^+ {\ensuremath{\vdash}}y
= (g x_1 \ldots x_n)$ and thus $$\Gamma_i^+ {\ensuremath{\vdash}}(g t_{i,j,1} \ldots t_{i,j,n}) < (g x_1 \ldots
x_n) \supset \psi(g t_{i,j,1} \ldots t_{i,j,n}).$$ By the rule $\supset_e$ we have $\Gamma_i^+ {\ensuremath{\vdash}}\psi(g t_{i,j,1}
\ldots t_{i,j,n})$, i.e. $$\Gamma_i^+ {\ensuremath{\vdash}}\forall x_1 : \alpha_1 \ldots \forall x_n : \alpha_n \,.\, (g t_{i,j,1}
\ldots t_{i,j,n}) =_{\ensuremath{\mathrm{Nat}}}(g x_1 \ldots x_n) \supset f x_1 \ldots x_n
: \beta.$$ Thus $
\Gamma_i^+ {\ensuremath{\vdash}}(f t_{i,j,1} \ldots t_{i,j,n}) : \beta
$ for $j=1,\ldots,k_i$. Because $\Gamma_i^+ {\ensuremath{\vdash}}t_i : \beta \to
\ldots \to \beta$ where $\beta$ occurs $k_i+1$ times, we have $$\Gamma_i^+ {\ensuremath{\vdash}}(t_i (f t_{i,1,1} \ldots
t_{i,1,n}) \ldots (f t_{i,k_i,1} \ldots t_{i,k_i,n})) : \beta.$$ But $$\Gamma_i^+ {\ensuremath{\vdash}}f x_1 \ldots x_n = t_i (f t_{i,1,1} \ldots
t_{i,1,n}) \ldots (f t_{i,k_i,1} \ldots t_{i,k_i,n})$$ so $\Gamma_i^+ {\ensuremath{\vdash}}(f x_1 \ldots x_n) : \beta$.
Therefore, we have derived the premises of $n_i'$ for $\psi$, hence by Lemma \[lem\_nat\_ind\_2\] we obtain $\Gamma {\ensuremath{\vdash}}\forall x :
{\ensuremath{\mathrm{Nat}}}\,.\, \psi(x)$, i.e. $$\Gamma {\ensuremath{\vdash}}\forall x : {\ensuremath{\mathrm{Nat}}}\forall x_1 : \alpha_1 \ldots
\forall x_n : \alpha_n \,.\, x =_{\ensuremath{\mathrm{Nat}}}(g x_1 \ldots x_n) \supset f
x_1 \ldots x_n : \beta.$$ Hence $$\Gamma, x_1 : \alpha_1, \ldots, x_n : \alpha_n, x : {\ensuremath{\mathrm{Nat}}}{\ensuremath{\vdash}}x
=_{\ensuremath{\mathrm{Nat}}}(g x_1 \ldots x_n) \supset f x_1 \ldots x_n : \beta$$ so $$\Gamma, x_1 : \alpha_1, \ldots, x_n : \alpha_n {\ensuremath{\vdash}}\forall x :
{\ensuremath{\mathrm{Nat}}}\,.\, x =_{\ensuremath{\mathrm{Nat}}}(g x_1 \ldots x_n) \supset f x_1 \ldots x_n :
\beta.$$ Because $\Gamma, x_1 : \alpha_1, \ldots, x_n : \alpha_n {\ensuremath{\vdash}}(g
x_1 \ldots x_n) : {\ensuremath{\mathrm{Nat}}}$, we easily obtain $$\Gamma, x_1 : \alpha_1, \ldots, x_n : \alpha_n {\ensuremath{\vdash}}f x_1 \ldots
x_n : \beta$$ and thus $\Gamma {\ensuremath{\vdash}}f : \alpha_1 \to \ldots \to \alpha_n \to
\beta$.
Soundness of the Translation of CPRED$\omega$ {#sec_cpred_soundness}
=============================================
In this section by $\tau$, $\tau_1$, $\tau_2$, $\sigma$, $\sigma_1$, etc. we denote types of , by $t$, $t_1$, $t_2$, etc. terms of , by $\varphi$, $\psi$, etc. formulas of , and by $s$, $s_1$, $s_2$, etc. terms of ${\ensuremath{{\cal I}}}_s$. By ${\ensuremath{{\cal T}}}$ we denote the set of types of . Below by $\Delta$ we denote an arbitrary set of formulas of .
If $t$ is a term of then we use the notation $\Gamma(t)$ for $\Gamma(\{t\})$. By $\Gamma(\Delta,\varphi)$ we denote $\Gamma(\Delta \cup \{\varphi\})$. Below $\Delta$ denotes an arbitrary set of formulas of .
Note that if $\Delta \subseteq \Delta'$ then $\Gamma(\Delta) \subseteq
\Gamma(\Delta')$, after possibly renaming some variables $y \in
FV(\Gamma(\Delta))$ such that $y \notin FV(\Delta)$. Because of Lemma \[lem\_I\_s\_subst\], we may assume without loss of generality that this implication always holds verbatim, and also that $\Gamma(t_1
t_2) = \Gamma(t_1,t_2)$, $\Gamma(\varphi \supset \psi) =
\Gamma(\varphi,\psi)$, etc.
\[lem\_transl\_is\_type\] $\Gamma(\Delta) {\ensuremath{\vdash}}_{{\ensuremath{{\cal I}}}_s} {\ensuremath{\lceil\tau\rceil}} : {\ensuremath{\mathrm{Type}}}$
Straightforward induction on the structure of $\tau$, using rule $\to_t$.
\[lem\_transl\_typed\] If $t \in T_\tau$ then $\Gamma(t) {\ensuremath{\vdash}}{\ensuremath{\lceilt\rceil}} :
{\ensuremath{\lceil\tau\rceil}}$.
Induction on the structure of $t$, using rules $\to_i$, $\to_e$, $\supset_t$ and $\forall_t$, and Lemma \[lem\_transl\_is\_type\] and weakening.
\[lem\_transl\_subst\] ${\ensuremath{\lceilt_1[x/t_2]\rceil}} = {\ensuremath{\lceilt_1\rceil}}[x/{\ensuremath{\lceilt_2\rceil}}]$
Straightforward induction on the structure of $t_1$.
\[lem\_transl\_eq\] If $t_1 \to_\beta t_2$ then $\Gamma(t_1,t_2) {\ensuremath{\vdash}}_{{\ensuremath{{\cal I}}}_s}
{\ensuremath{\lceilt_1\rceil}} = {\ensuremath{\lceilt_2\rceil}}$.
Induction on the structure of $t_1$, using the axiom $\beta$, the rules for Eq and Lemma \[lem\_transl\_subst\].
\[lem\_inhabited\] For every $\tau \in {\ensuremath{{\cal T}}}$ there is an ${\ensuremath{{\cal I}}}_s$-term $s$ such that $\Gamma(\Delta) {\ensuremath{\vdash}}_{{\ensuremath{{\cal I}}}_s} s : {\ensuremath{\lceil\tau\rceil}}$.
Induction on the structure of $\tau$. If $\tau \in {\ensuremath{{\cal B}}}$ then ${\ensuremath{\lceil\tau\rceil}} = \tau$ and there exists $y \in V_\tau$ such that $y
\notin FV(\Delta)$ and $(y : \tau) \in \Gamma(\Delta)$. Thus $\Gamma(\Delta) {\ensuremath{\vdash}}y : {\ensuremath{\lceil\tau\rceil}}$. If $\tau = o$ then ${\ensuremath{\lceil\tau\rceil}} = {\ensuremath{\mathrm{Prop}}}$ and e.g. $\Gamma(\Delta) {\ensuremath{\vdash}}\bot :
{\ensuremath{\lceil\tau\rceil}}$. If $\tau \notin {\ensuremath{{\cal B}}}\cup \{o\}$ then $\tau =
\tau_1\to\tau_2$ for some $\tau_1,\tau_2\in {\ensuremath{{\cal T}}}$. By the inductive hypothesis there exists an ${\ensuremath{{\cal I}}}_s$-term $s_2$ such that $\Gamma(\Delta) {\ensuremath{\vdash}}s_2 : {\ensuremath{\lceil\tau_2\rceil}}$. Suppose $x \notin
FV(s_2, \Gamma(\Delta))$. Then $\Gamma(\Delta), x : {\ensuremath{\lceil\tau_1\rceil}}
{\ensuremath{\vdash}}s_2 : {\ensuremath{\lceil\tau_2\rceil}}$. Since ${\ensuremath{\lceil\tau\rceil}} =
{\ensuremath{\lceil\tau_1\rceil}} \to {\ensuremath{\lceil\tau_2\rceil}}$ and $\Gamma(\Delta) {\ensuremath{\vdash}}{\ensuremath{\lceil\tau_1\rceil}} : {\ensuremath{\mathrm{Type}}}$ we obtain $\Gamma(\Delta) {\ensuremath{\vdash}}(\lambda
x \,.\, s_2) : {\ensuremath{\lceil\tau\rceil}}$ by Lemma \[lem\_transl\_is\_type\].
If $\Delta {\ensuremath{\vdash}}_{\mathrm{CPRED}\omega} \varphi$ then $\Gamma(\Delta, \varphi), {\ensuremath{\lceil\Delta\rceil}} {\ensuremath{\vdash}}_{{\ensuremath{{\cal I}}}_s}
{\ensuremath{\lceil\varphi\rceil}}$. The same holds if we change CPRED$\omega$ to and ${\ensuremath{{\cal I}}}_s$ to $e{\ensuremath{{\cal I}}}_s$.
Induction on the length of the derivation of $\Delta {\ensuremath{\vdash}}_{{{\ensuremath{{\mathrm{CPRED}\omega}}}}}
\varphi$.
If $\Delta {\ensuremath{\vdash}}_{{{\ensuremath{{\mathrm{CPRED}\omega}}}}} \varphi$ is an instance of the axiom $\Delta', \varphi {\ensuremath{\vdash}}_{{{\ensuremath{{\mathrm{CPRED}\omega}}}}} \varphi$, then obviously $\Gamma(\Delta, \varphi), {\ensuremath{\lceil\Delta\rceil}} {\ensuremath{\vdash}}_{{\ensuremath{{\cal I}}}_s}
{\ensuremath{\lceil\varphi\rceil}}$ by axiom 1 of ${\ensuremath{{\cal I}}}_s$. It is also easy to see that $$\Gamma(\Delta,\forall p : {\ensuremath{\mathrm{Prop}}}\,.\, ((p \supset \bot) \supset
\bot) \supset p), {\ensuremath{\lceil\Delta\rceil}} {\ensuremath{\vdash}}_{{\ensuremath{{\cal I}}}_s} {\ensuremath{\lceil\forall p :
{\ensuremath{\mathrm{Prop}}}\,.\, ((p \supset \bot) \supset \bot) \supset p\rceil}}.$$
Suppose $\Delta {\ensuremath{\vdash}}_{{{\ensuremath{{\mathrm{CPRED}\omega}}}}} \varphi_1 \supset \varphi_2$ is obtained from $\Delta, \varphi_1 {\ensuremath{\vdash}}_{{{\ensuremath{{\mathrm{CPRED}\omega}}}}} \varphi_2$ by the rule $\supset_i^P$. Then by the inductive hypothesis $\Gamma(\Delta,
\varphi_1, \varphi_2), {\ensuremath{\lceil\Delta\rceil}}, {\ensuremath{\lceil\varphi_1\rceil}} {\ensuremath{\vdash}}_{{\ensuremath{{\cal I}}}_s}
{\ensuremath{\lceil\varphi_2\rceil}}$. By Lemma \[lem\_transl\_typed\] we have $\Gamma(\varphi_1) {\ensuremath{\vdash}}_{{\ensuremath{{\cal I}}}_s} {\ensuremath{\lceil\varphi_1\rceil}} : {\ensuremath{\mathrm{Prop}}}$. Hence by weakening and $\supset_i$ we obtain $\Gamma(\Delta, \varphi_1,
\varphi_2), {\ensuremath{\lceil\Delta\rceil}} {\ensuremath{\vdash}}_{{\ensuremath{{\cal I}}}_s} {\ensuremath{\lceil\varphi_1\rceil}} \supset
{\ensuremath{\lceil\varphi_2\rceil}}$, i.e. $\Gamma(\Delta, \varphi_1 \supset
\varphi_2), {\ensuremath{\lceil\Delta\rceil}} {\ensuremath{\vdash}}_{{\ensuremath{{\cal I}}}_s} {\ensuremath{\lceil\varphi_1 \supset
\varphi_2\rceil}}$.
Suppose $\Delta {\ensuremath{\vdash}}_{{{\ensuremath{{\mathrm{CPRED}\omega}}}}} \varphi$ is a direct consequence of $\Delta {\ensuremath{\vdash}}_{{{\ensuremath{{\mathrm{CPRED}\omega}}}}} \psi \supset \varphi$ and $\Delta
{\ensuremath{\vdash}}_{{{\ensuremath{{\mathrm{CPRED}\omega}}}}} \psi$ by the rule $\supset_e^P$. By the IH and $\supset_e$ we have $\Gamma(\Delta,\psi,\varphi),
{\ensuremath{\lceil\Delta\rceil}} {\ensuremath{\vdash}}_{{\ensuremath{{\cal I}}}_s} {\ensuremath{\lceil\varphi\rceil}}$. Note that $\Gamma(\Delta,\psi,\varphi) = \Gamma(\Delta,\varphi), x_1 :
{\ensuremath{\lceil\tau_1\rceil}}, \ldots, x_n : {\ensuremath{\lceil\tau_n\rceil}}$ where $\{x_1,\ldots,x_n\} = FV(\psi) \setminus FV(\Delta, \varphi)$, $x_i
\in V_{\tau_i}$, and without loss of generality $x_i \notin
\Gamma(\Delta,\varphi)$. Since $x_i \notin FV(\Delta, \varphi)$ we have $x_i \notin FV({\ensuremath{\lceil\Delta\rceil}}, {\ensuremath{\lceil\varphi\rceil}})$. By Lemma \[lem\_inhabited\] there exist ${\ensuremath{{\cal I}}}_s$-terms $s_1, \ldots,
s_n$ such that $\Gamma(\Delta, \varphi) {\ensuremath{\vdash}}_{{\ensuremath{{\cal I}}}_s} s_i :
{\ensuremath{\lceil\tau_i\rceil}}$. By Lemma \[lem\_I\_s\_subst\] we have $$\Gamma(\Delta,\varphi), {\ensuremath{\lceil\Delta\rceil}}, s_1 : {\ensuremath{\lceil\tau_1\rceil}},
\ldots, s_n : {\ensuremath{\lceil\tau_n\rceil}} {\ensuremath{\vdash}}_{{\ensuremath{{\cal I}}}_s} {\ensuremath{\lceil\varphi\rceil}}.$$ By applying $p_i$, $\supset_i$ and $\supset_e$ successively we obtain $\Gamma(\Delta, \varphi), {\ensuremath{\lceil\Delta\rceil}} {\ensuremath{\vdash}}_{{\ensuremath{{\cal I}}}_s}
{\ensuremath{\lceil\varphi\rceil}}$.
Suppose $\Delta {\ensuremath{\vdash}}_{{{\ensuremath{{\mathrm{CPRED}\omega}}}}} \forall x : \tau \,.\, \varphi$ is obtained from $\Delta {\ensuremath{\vdash}}_{{{\ensuremath{{\mathrm{CPRED}\omega}}}}} \varphi$ by rule $\forall_i^P$. By the IH we have $\Gamma(\Delta,\varphi),
{\ensuremath{\lceil\Delta\rceil}} {\ensuremath{\vdash}}_{{\ensuremath{{\cal I}}}_s} {\ensuremath{\lceil\varphi\rceil}}$. Since $x \notin
FV(\Delta)$, we have $\Gamma(\Delta, \varphi) = \Gamma(\Delta,
\forall x : \tau \,.\, \varphi), x : {\ensuremath{\lceil\tau\rceil}}$. By Lemma \[lem\_transl\_is\_type\] and weakening we have $$\Gamma(\Delta, \forall x : \tau \,.\, \varphi), {\ensuremath{\lceil\Delta\rceil}}
{\ensuremath{\vdash}}_{{\ensuremath{{\cal I}}}_s} {\ensuremath{\lceil\tau\rceil}} : {\ensuremath{\mathrm{Type}}}.$$ Hence $\Gamma(\Delta, \forall x : \tau \,.\, \varphi),
{\ensuremath{\lceil\Delta\rceil}} {\ensuremath{\vdash}}_{{\ensuremath{{\cal I}}}_s} {\ensuremath{\lceil\forall x : \tau \,.\,
\varphi\rceil}}$ by $\forall_i$ and ${\ensuremath{\lceil\forall x : \tau \,.\,
\varphi\rceil}} = \forall x : {\ensuremath{\lceil\tau\rceil}} \,.\, {\ensuremath{\lceil\varphi\rceil}}$.
Suppose $\Delta {\ensuremath{\vdash}}_{{{\ensuremath{{\mathrm{CPRED}\omega}}}}} \varphi[x/t]$ is obtained from $\Delta
{\ensuremath{\vdash}}_{{{\ensuremath{{\mathrm{CPRED}\omega}}}}} \forall x : \tau \,.\, \varphi$ by rule $\forall_i^P$. We then have $x \notin FV(\Delta)$ and $t \in
T_\tau$. By the IH we have $$\Gamma(\Delta, \forall x : \tau \,.\, \varphi), {\ensuremath{\lceil\Delta\rceil}}
{\ensuremath{\vdash}}_{{\ensuremath{{\cal I}}}_s} \forall x : {\ensuremath{\lceil\tau\rceil}} \,.\, {\ensuremath{\lceil\varphi\rceil}}.$$ If $x \notin FV(\varphi)$ then $\varphi \equiv \varphi[x/t]$ and $\Gamma(\Delta, \forall x : \tau \,.\, \varphi) = \Gamma(\Delta,
\varphi[x/t])$. By Lemma \[lem\_inhabited\] there exists an ${\ensuremath{{\cal I}}}_s$-term $s$ such that $\Gamma(\Delta, \varphi[x/t])
{\ensuremath{\vdash}}_{{\ensuremath{{\cal I}}}_s} s : {\ensuremath{\lceil\tau\rceil}}$. By $\forall_e$ we obtain $\Gamma(\Delta, \varphi[x/t]), {\ensuremath{\lceil\Delta\rceil}} {\ensuremath{\vdash}}_{{\ensuremath{{\cal I}}}_s}
{\ensuremath{\lceil\varphi[x/t]\rceil}}$. If $x \in FV(\varphi)$ then $\Gamma(\Delta,\varphi[x/t]) = \Gamma(\Delta, \forall x : \tau \,.\,
\varphi, t)$. By Lemma \[lem\_transl\_typed\] we have $\Gamma(\Delta,\varphi[x/t]) {\ensuremath{\vdash}}{\ensuremath{\lceilt\rceil}} : {\ensuremath{\lceil\tau\rceil}}$, so by $\forall_e$ and Lemma \[lem\_transl\_subst\] we obtain $\Gamma(\Delta,\varphi[x/t]), {\ensuremath{\lceil\Delta\rceil}} {\ensuremath{\vdash}}{\ensuremath{\lceil\varphi[x/t]\rceil}}$.
Suppose $\Delta {\ensuremath{\vdash}}_{{{\ensuremath{{\mathrm{CPRED}\omega}}}}} \psi$ is obtained from $\Delta
{\ensuremath{\vdash}}_{{{\ensuremath{{\mathrm{CPRED}\omega}}}}} \varphi$ by rule $\mathrm{conv}^P$. Then $\varphi
=_\beta \psi$. By the IH we have $\Gamma(\Delta, \varphi),
{\ensuremath{\lceil\Delta\rceil}} {\ensuremath{\vdash}}_{{\ensuremath{{\cal I}}}_s} {\ensuremath{\lceil\varphi\rceil}}$. It suffices to show that if $\varphi \to_\beta \psi$ or $\psi \to_\beta \varphi$ then $\Gamma(\Delta, \psi), {\ensuremath{\lceil\Delta\rceil}} {\ensuremath{\vdash}}_{{\ensuremath{{\cal I}}}_s}
{\ensuremath{\lceil\psi\rceil}}$. If $\varphi \to_\beta \psi$ then $\Gamma(\psi)
\subset \Gamma(\varphi)$ and by Lemma \[lem\_transl\_eq\], rule eq and weakening we obtain $\Gamma(\Delta, \varphi), {\ensuremath{\lceil\Delta\rceil}}
{\ensuremath{\vdash}}{\ensuremath{\lceil\psi\rceil}}$. Note that $\Gamma(\Delta,\varphi) =
\Gamma(\Delta,\psi), x_1 : {\ensuremath{\lceil\tau_1\rceil}}, \ldots, x_n :
{\ensuremath{\lceil\tau_n\rceil}}$ where $\{x_1,\ldots,x_n\} = FV(\varphi) \setminus
FV(\Delta, \psi)$, $x_i \in V_{\tau_i}$. Hence, we may use the same argument as in the proof for the rule $\supset_e^P$ to obtain $\Gamma(\Delta,\psi), {\ensuremath{\lceil\Delta\rceil}} {\ensuremath{\vdash}}_{{\ensuremath{{\cal I}}}_s}
{\ensuremath{\lceil\psi\rceil}}$. If $\psi \to_\beta \varphi$ then $\Gamma(\varphi)
\subseteq \Gamma(\psi)$, and by Lemma \[lem\_transl\_eq\], rule eq and weakening we obtain $\Gamma(\Delta,\psi), {\ensuremath{\lceil\Delta\rceil}}
{\ensuremath{\vdash}}{\ensuremath{\lceil\psi\rceil}}$.
To show that if $\Delta {\ensuremath{\vdash}}_{\mathrm{E{-}CPRED}\omega} \varphi$ then $\Gamma(\Delta, \varphi), {\ensuremath{\lceil\Delta\rceil}} {\ensuremath{\vdash}}_{e{\ensuremath{{\cal I}}}_s}
{\ensuremath{\lceil\varphi\rceil}}$, it now suffices to prove that the translations of the axioms $e_f^P$ and $e_b^P$ are derivable in $e{\ensuremath{{\cal I}}}_s$. This is straightforward, using the rules $e_f$, $e_b$, eq, Lemma \[lem\_transl\_is\_type\] and Lemma \[lem\_leibniz\].
Proofs for Section \[sec\_partiality\] {#sec_proofs_examples}
======================================
In this section we give sketches of formal proofs for the examples in Sect. \[sec\_partiality\]. The formal derivations are slightly simplified, by omitting certain steps and rule assumptions, simplifying inferences, and generally omitting some parts which may be easily reconstructed by the reader. We also assume certain basic properties of operations on natural number and don’t derive them. These properties may be derived from Theorem \[thm\_peano\], Lemma \[lem\_nat\_op\_well\_defined\] and the definitions of $\le$ and $-$.
${\ensuremath{\vdash}}_{{\ensuremath{{\cal I}}}_s} \forall i,j : {\ensuremath{\mathrm{Nat}}}\,.\, (i \ge j) \supset
({\ensuremath{\mathrm{subp}}}\,i\,j = i - j)$
Let $\varphi(x) \equiv \forall i,j : {\ensuremath{\mathrm{Nat}}}\,.\, i \ge j \supset x
=_{\ensuremath{\mathrm{Nat}}}i - j \supset {\ensuremath{\mathrm{subp}}}\,i\,j = i - j$. $$(a0.0)
\inferrule*{
i : {\ensuremath{\mathrm{Nat}}}, j : {\ensuremath{\mathrm{Nat}}}{\ensuremath{\vdash}}i : {\ensuremath{\mathrm{Nat}}}\\
\inferrule*{
}{
{\ensuremath{\vdash}}\forall x, y : {\ensuremath{\mathrm{Nat}}}\,.\, (x - y) : {\ensuremath{\mathrm{Nat}}}}\;\raisebox{0.7em}{by~\ref{lem_nat_op_well_defined}}
}{
i : {\ensuremath{\mathrm{Nat}}}, j : {\ensuremath{\mathrm{Nat}}}{\ensuremath{\vdash}}\forall y : {\ensuremath{\mathrm{Nat}}}\,.\, (i - y) : {\ensuremath{\mathrm{Nat}}}}$$ $$(a0)
\inferrule*{
\inferrule*{
(a0.0)
\\
i : {\ensuremath{\mathrm{Nat}}}, j : {\ensuremath{\mathrm{Nat}}}{\ensuremath{\vdash}}j : {\ensuremath{\mathrm{Nat}}}}{
i : {\ensuremath{\mathrm{Nat}}}, j : {\ensuremath{\mathrm{Nat}}}{\ensuremath{\vdash}}(i - j) : {\ensuremath{\mathrm{Nat}}}}
\\
{\ensuremath{\vdash}}0 : {\ensuremath{\mathrm{Nat}}}}{
i : {\ensuremath{\mathrm{Nat}}}, j : {\ensuremath{\mathrm{Nat}}}{\ensuremath{\vdash}}(0 =_{\ensuremath{\mathrm{Nat}}}(i - j)) : {\ensuremath{\mathrm{Nat}}}}\;\raisebox{0.7em}{by~\ref{lem_leibniz}}$$ Completely analogously, we obtain $$(a1)\;\; i : {\ensuremath{\mathrm{Nat}}}, j : {\ensuremath{\mathrm{Nat}}}{\ensuremath{\vdash}}(i \ge j) : {\ensuremath{\mathrm{Nat}}}$$ Now let $\Gamma_0 \equiv i : {\ensuremath{\mathrm{Nat}}}, j : {\ensuremath{\mathrm{Nat}}}, i \ge j, 0 =_{\ensuremath{\mathrm{Nat}}}i -
j$. $$(a2)
\inferrule*{
\inferrule*{
\inferrule*{
\inferrule*{
\inferrule*{
\inferrule*{
}{
\Gamma_0 {\ensuremath{\vdash}}0 = i - j
}\;\raisebox{0.7em}{by~\ref{lem_ind_eq}}
}{
\Gamma_0 {\ensuremath{\vdash}}{\ensuremath{\mathfrak{0}}}(i - j)
}
\\
\Gamma_0 {\ensuremath{\vdash}}i \ge j
}{
\Gamma_0 {\ensuremath{\vdash}}i =_{\ensuremath{\mathrm{Nat}}}j
}\;\raisebox{0.7em}{by~\ref{lem_le_ge_then_eq}}
}{
\Gamma_0 {\ensuremath{\vdash}}{\ensuremath{\mathrm{subp}}}\,i\,j = 0
}
\\
\Gamma_0 {\ensuremath{\vdash}}0 = i - j
}{
\Gamma_0 {\ensuremath{\vdash}}{\ensuremath{\mathrm{subp}}}\,i\,j = i - j
}
\\
(a0)
}{
i : {\ensuremath{\mathrm{Nat}}}, j : {\ensuremath{\mathrm{Nat}}}, i \ge j {\ensuremath{\vdash}}0 =_{\ensuremath{\mathrm{Nat}}}i - j \supset {\ensuremath{\mathrm{subp}}}\,i\,j = i - j
}$$ $$(a)
\inferrule*{
\inferrule*{
(a1)
\\
(a2)
}{
i : {\ensuremath{\mathrm{Nat}}}, j : {\ensuremath{\mathrm{Nat}}}{\ensuremath{\vdash}}i \ge j \supset 0 =_{\ensuremath{\mathrm{Nat}}}i - j \supset {\ensuremath{\mathrm{subp}}}\,i\,j = i - j
}
\\
{\ensuremath{\vdash}}{\ensuremath{\mathrm{Nat}}}: {\ensuremath{\mathrm{Type}}}}{
{\ensuremath{\vdash}}\varphi(0)
}$$ Let $\Gamma_1 \equiv y : {\ensuremath{\mathrm{Nat}}}, \varphi(y), i : {\ensuremath{\mathrm{Nat}}}, j : {\ensuremath{\mathrm{Nat}}}, {\ensuremath{\mathfrak{s}}}y
=_{\ensuremath{\mathrm{Nat}}}i - j$. $$(b0.0)
\inferrule*{
\Gamma_1 {\ensuremath{\vdash}}i : {\ensuremath{\mathrm{Nat}}}\\ \Gamma_1 {\ensuremath{\vdash}}j : {\ensuremath{\mathrm{Nat}}}}{
\Gamma_1 {\ensuremath{\vdash}}(i =_{\ensuremath{\mathrm{Nat}}}j) : {\ensuremath{\mathrm{Prop}}}}\;\raisebox{0.7em}{by~\ref{lem_leibniz}}$$ $$(b0.1)
\inferrule*{
\inferrule*{
\inferrule*{
\inferrule*{
}{
\Gamma_1, i =_{\ensuremath{\mathrm{Nat}}}j {\ensuremath{\vdash}}i = j
}\;\raisebox{0.7em}{by~\ref{lem_ind_eq}}
\\
\Gamma_1 {\ensuremath{\vdash}}{\ensuremath{\mathfrak{s}}}y =_{\ensuremath{\mathrm{Nat}}}i - j
}{
\Gamma_1, i =_{\ensuremath{\mathrm{Nat}}}j {\ensuremath{\vdash}}{\ensuremath{\mathfrak{s}}}y =_{\ensuremath{\mathrm{Nat}}}i - i
}
}{
\Gamma_1, i =_{\ensuremath{\mathrm{Nat}}}j {\ensuremath{\vdash}}{\ensuremath{\mathfrak{s}}}y =_{\ensuremath{\mathrm{Nat}}}0
}
\\
\inferrule*{
\Gamma_1 {\ensuremath{\vdash}}y : {\ensuremath{\mathrm{Nat}}}}{
\Gamma_1 {\ensuremath{\vdash}}({\ensuremath{\mathfrak{s}}}y) : {\ensuremath{\mathrm{Nat}}}}
}{
\Gamma_1, i =_{\ensuremath{\mathrm{Nat}}}j {\ensuremath{\vdash}}{\ensuremath{\mathfrak{s}}}y = 0
}$$ $$(b0)
\inferrule*{
\inferrule*{
\inferrule*{
(b0.1)
\\
\Gamma_1, i =_{\ensuremath{\mathrm{Nat}}}j {\ensuremath{\vdash}}{\ensuremath{\mathfrak{0}}}0
}{
\Gamma_1, i =_{\ensuremath{\mathrm{Nat}}}j {\ensuremath{\vdash}}{\ensuremath{\mathfrak{0}}}({\ensuremath{\mathfrak{s}}}y)
}
\\
\Gamma_1, i =_{\ensuremath{\mathrm{Nat}}}j {\ensuremath{\vdash}}\neg ({\ensuremath{\mathfrak{0}}}({\ensuremath{\mathfrak{s}}}y))
}{
\Gamma_1, i =_{\ensuremath{\mathrm{Nat}}}j {\ensuremath{\vdash}}\bot
}
\\
(b0.0)
}{
\Gamma_1 {\ensuremath{\vdash}}\neg (i =_{\ensuremath{\mathrm{Nat}}}j)
}$$ $$(c0)
\inferrule*{
\Gamma_1 {\ensuremath{\vdash}}\varphi(y)
\\
\Gamma_1 {\ensuremath{\vdash}}i : {\ensuremath{\mathrm{Nat}}}\\
\inferrule*{
\Gamma_1 {\ensuremath{\vdash}}j : {\ensuremath{\mathrm{Nat}}}}{
\Gamma_1 {\ensuremath{\vdash}}({\ensuremath{\mathfrak{s}}}j) : {\ensuremath{\mathrm{Nat}}}}
}{
\Gamma_1 {\ensuremath{\vdash}}i \ge ({\ensuremath{\mathfrak{s}}}j) \supset y =_{\ensuremath{\mathrm{Nat}}}i - ({\ensuremath{\mathfrak{s}}}j) \supset
{\ensuremath{\mathrm{subp}}}\,i\,({\ensuremath{\mathfrak{s}}}j) = i - ({\ensuremath{\mathfrak{s}}}j)
}$$ $$(c1)
\inferrule*{
\inferrule*{
\Gamma_1 {\ensuremath{\vdash}}i \ge j \\ (b0)
}{
\Gamma_1 {\ensuremath{\vdash}}i \ge ({\ensuremath{\mathfrak{s}}}j)
}
\\
(c0)
}{
\Gamma_1 {\ensuremath{\vdash}}y =_{\ensuremath{\mathrm{Nat}}}i - ({\ensuremath{\mathfrak{s}}}j) \supset {\ensuremath{\mathrm{subp}}}\,i\,({\ensuremath{\mathfrak{s}}}j) = i - ({\ensuremath{\mathfrak{s}}}j)
}$$ $$(c2)
\inferrule*{
\Gamma_1 {\ensuremath{\vdash}}{\ensuremath{\mathfrak{s}}}y =_{\ensuremath{\mathrm{Nat}}}i - j
\\
\inferrule*{
\Gamma_1 {\ensuremath{\vdash}}i : {\ensuremath{\mathrm{Nat}}}\\ \Gamma_1 {\ensuremath{\vdash}}j : {\ensuremath{\mathrm{Nat}}}}{
\Gamma_1 {\ensuremath{\vdash}}(i - j) : {\ensuremath{\mathrm{Nat}}}}
\\
\inferrule*{
\Gamma_1 {\ensuremath{\vdash}}y : {\ensuremath{\mathrm{Nat}}}}{
\Gamma_1 {\ensuremath{\vdash}}({\ensuremath{\mathfrak{s}}}y) : {\ensuremath{\mathrm{Nat}}}}
}{
\Gamma_1 {\ensuremath{\vdash}}{\ensuremath{\mathfrak{s}}}y = i - j
}$$ $$(c3)
\inferrule*{
\inferrule*{
\inferrule*{
\Gamma_1 {\ensuremath{\vdash}}i \ge j \\ (b0)
}{
\Gamma_1 {\ensuremath{\vdash}}i > j
}
}{
\ldots
}
}{
\Gamma_1 {\ensuremath{\vdash}}{\ensuremath{\mathfrak{s}}}(i - ({\ensuremath{\mathfrak{s}}}j)) = i - j
}$$ $$(c4)
\inferrule*{
\inferrule*{
\inferrule*{
\inferrule*{
(c3)
\\
(c2)
}{
\Gamma_1 {\ensuremath{\vdash}}{\ensuremath{\mathfrak{s}}}(i - ({\ensuremath{\mathfrak{s}}}j)) = {\ensuremath{\mathfrak{s}}}y
}
}{
\Gamma_1 {\ensuremath{\vdash}}{\ensuremath{\mathfrak{p}}}({\ensuremath{\mathfrak{s}}}(i - ({\ensuremath{\mathfrak{s}}}j))) = {\ensuremath{\mathfrak{p}}}({\ensuremath{\mathfrak{s}}}y)
}
}{
\Gamma_1 {\ensuremath{\vdash}}i - ({\ensuremath{\mathfrak{s}}}j) = y
}
}{
\Gamma_1 {\ensuremath{\vdash}}y =_{\ensuremath{\mathrm{Nat}}}i - ({\ensuremath{\mathfrak{s}}}j)
}$$ $$(c)
\inferrule*{
\inferrule*{
\inferrule*{
\inferrule*{
(b0)
}{
\Gamma_1 {\ensuremath{\vdash}}{\ensuremath{\mathrm{subp}}}\,i\,j = {\ensuremath{\mathfrak{s}}}({\ensuremath{\mathrm{subp}}}\,i\,({\ensuremath{\mathfrak{s}}}j))
}
\\
\inferrule*{
\inferrule*{
(c1) \\ (c4)
}{
\Gamma_1 {\ensuremath{\vdash}}{\ensuremath{\mathrm{subp}}}\,i\,({\ensuremath{\mathfrak{s}}}j) = i - ({\ensuremath{\mathfrak{s}}}j)
}
}{
\Gamma_1 {\ensuremath{\vdash}}{\ensuremath{\mathfrak{s}}}({\ensuremath{\mathrm{subp}}}\,i\,({\ensuremath{\mathfrak{s}}}j)) = {\ensuremath{\mathfrak{s}}}(i - ({\ensuremath{\mathfrak{s}}}j))
}
}{
\Gamma_1 {\ensuremath{\vdash}}{\ensuremath{\mathrm{subp}}}\,i\,j = {\ensuremath{\mathfrak{s}}}(i - ({\ensuremath{\mathfrak{s}}}j)) \\ (c3)
}
}{
\Gamma_1 {\ensuremath{\vdash}}{\ensuremath{\mathrm{subp}}}\,i\,j = i - j
}
}{
y : {\ensuremath{\mathrm{Nat}}}, \varphi(y) {\ensuremath{\vdash}}\varphi({\ensuremath{\mathfrak{s}}}y)
}$$ $$(d)
\inferrule*{
(a) \\ (c)
}{
{\ensuremath{\vdash}}\forall y,i,j : {\ensuremath{\mathrm{Nat}}}\,.\, i \ge j \supset y =_{\ensuremath{\mathrm{Nat}}}i - j \supset {\ensuremath{\mathrm{subp}}}\,i\,j = i - j
}\;\raisebox{0.7em}{by~$n_i$}$$ $$(e)
\inferrule*{
\inferrule*{
\inferrule*{
\inferrule*{
\ldots
}{
i : {\ensuremath{\mathrm{Nat}}}, j : {\ensuremath{\mathrm{Nat}}}{\ensuremath{\vdash}}(i - j) : {\ensuremath{\mathrm{Nat}}}}
\\
(d)
}{
i : {\ensuremath{\mathrm{Nat}}}, j : {\ensuremath{\mathrm{Nat}}}{\ensuremath{\vdash}}\forall k,l : {\ensuremath{\mathrm{Nat}}}\,.\, k \ge l
\supset i - j =_{\ensuremath{\mathrm{Nat}}}k - l \supset {\ensuremath{\mathrm{subp}}}\,k\,l = k - l
}
}{
i : {\ensuremath{\mathrm{Nat}}}, j : {\ensuremath{\mathrm{Nat}}}{\ensuremath{\vdash}}i \ge j \supset i - j =_{\ensuremath{\mathrm{Nat}}}i - j \supset {\ensuremath{\mathrm{subp}}}\,i\,j = i - j
}
}{
i : {\ensuremath{\mathrm{Nat}}}, j : {\ensuremath{\mathrm{Nat}}}, i \ge j {\ensuremath{\vdash}}i - j =_{\ensuremath{\mathrm{Nat}}}i - j \supset {\ensuremath{\mathrm{subp}}}\,i\,j = i - j
}$$ $$(f)
\inferrule*{
(e)
\\
\inferrule*{
\inferrule*{
\ldots
}{
i : {\ensuremath{\mathrm{Nat}}}, j : {\ensuremath{\mathrm{Nat}}}{\ensuremath{\vdash}}(i - j) : {\ensuremath{\mathrm{Nat}}}}
}{
i : {\ensuremath{\mathrm{Nat}}}, j : {\ensuremath{\mathrm{Nat}}}{\ensuremath{\vdash}}i - j =_{\ensuremath{\mathrm{Nat}}}i - j
}
}{
i : {\ensuremath{\mathrm{Nat}}}, j : {\ensuremath{\mathrm{Nat}}}, i \ge j {\ensuremath{\vdash}}{\ensuremath{\mathrm{subp}}}\,i\,j = i - j
}$$ $$\inferrule*{
\inferrule*{
(f)
\\
\inferrule*{
\ldots
}{
i : {\ensuremath{\mathrm{Nat}}}, j : {\ensuremath{\mathrm{Nat}}}{\ensuremath{\vdash}}(i \ge j) : {\ensuremath{\mathrm{Prop}}}}
}{
i : {\ensuremath{\mathrm{Nat}}}, j : {\ensuremath{\mathrm{Nat}}}{\ensuremath{\vdash}}i \ge j \supset {\ensuremath{\mathrm{subp}}}\,i\,j = i - j
}
\\
{\ensuremath{\vdash}}{\ensuremath{\mathrm{Nat}}}: {\ensuremath{\mathrm{Type}}}}{
{\ensuremath{\vdash}}\forall i,j : {\ensuremath{\mathrm{Nat}}}\,.\, i \ge j \supset {\ensuremath{\mathrm{subp}}}\,i\,j = i - j
}$$
If $f$ is a term such that $${\ensuremath{\vdash}}_{{\ensuremath{{\cal I}}}_s} f(n) = {\ensuremath{\mathrm{Cond}\,(n > 100)\,(n - 10)\,(f(f(n + 11)))\,}}$$ then $${\ensuremath{\vdash}}_{{\ensuremath{{\cal I}}}_s} \forall n : {\ensuremath{\mathrm{Nat}}}\,.\, n \le 101 \supset f(n) = 91.$$
Let $\varphi(y) \equiv \forall n : {\ensuremath{\mathrm{Nat}}}\,.\, n \le 101 \supset 101
- n \le y \supset f(n) = 91$. $$(a0)
\inferrule*{
\inferrule*{
n : {\ensuremath{\mathrm{Nat}}}, n \le 101, 101 - n \le 0 {\ensuremath{\vdash}}101 \le n
}{
n : {\ensuremath{\mathrm{Nat}}}, n \le 101, 101 - n \le 0 {\ensuremath{\vdash}}n =_{\ensuremath{\mathrm{Nat}}}101
}\;\raisebox{0.7em}{by~\ref{lem_le_ge_then_eq}}
}{
n : {\ensuremath{\mathrm{Nat}}}, n \le 101, 101 - n \le 0 {\ensuremath{\vdash}}f(n) = f(101) = 101 -
10 = 91
}$$ $$(a)
\inferrule*{
\inferrule*{
\inferrule*{
\inferrule*{
n : {\ensuremath{\mathrm{Nat}}}{\ensuremath{\vdash}}(101 - n) : {\ensuremath{\mathrm{Nat}}}}{
n : {\ensuremath{\mathrm{Nat}}}{\ensuremath{\vdash}}(101 - n \le 0) : {\ensuremath{\mathrm{Prop}}}}
\\
(a0)
}{
n : {\ensuremath{\mathrm{Nat}}}, n \le 101 {\ensuremath{\vdash}}101 - n \le 0 \supset f(n) = 91
}
\\
n : {\ensuremath{\mathrm{Nat}}}{\ensuremath{\vdash}}(n \le 101) : {\ensuremath{\mathrm{Prop}}}}{
n : {\ensuremath{\mathrm{Nat}}}{\ensuremath{\vdash}}n \le 101 \supset 101 - n \le 0 \supset f(n) = 91
}
\\
{\ensuremath{\vdash}}{\ensuremath{\mathrm{Nat}}}: {\ensuremath{\mathrm{Type}}}}{
{\ensuremath{\vdash}}\varphi(0)
}$$ Let $\Gamma_0 \equiv y : {\ensuremath{\mathrm{Nat}}}, \varphi(y), n : {\ensuremath{\mathrm{Nat}}}, n \le 101, 101
- n \le {\ensuremath{\mathfrak{s}}}y$. Let $\Gamma_1 \equiv \Gamma_0, n + 11 > 101, n <
101$. $$(b0.0)
\inferrule*{
\inferrule*{
\Gamma_1 {\ensuremath{\vdash}}n : {\ensuremath{\mathrm{Nat}}}}{
\Gamma_1 {\ensuremath{\vdash}}(n + 1) : {\ensuremath{\mathrm{Nat}}}}
\\
\Gamma_1 {\ensuremath{\vdash}}\varphi(y)
}{
\Gamma_1 {\ensuremath{\vdash}}n + 1 \le 101 \supset 101 - (n + 1) \le y
\supset f(n + 1) = 91
}$$ $$(b0)
\inferrule*{
(b0.0)
\\
\Gamma_1 {\ensuremath{\vdash}}n + 1 \le 101
}{
\Gamma_1 {\ensuremath{\vdash}}101 - (n + 1) \le y \supset f(n + 1) = 91
}$$ $$(b1)
\inferrule*{
\inferrule*{
\inferrule*{
\Gamma_1 {\ensuremath{\vdash}}101 - n \le {\ensuremath{\mathfrak{s}}}y
\\
\inferrule*{
\Gamma_1 {\ensuremath{\vdash}}101 - ({\ensuremath{\mathfrak{s}}}n) = {\ensuremath{\mathfrak{p}}}(101 - n)
}{
\Gamma_1 {\ensuremath{\vdash}}{\ensuremath{\mathfrak{s}}}(101 - ({\ensuremath{\mathfrak{s}}}n)) = 101 - n
}
}{
\Gamma_1 {\ensuremath{\vdash}}{\ensuremath{\mathfrak{s}}}(101 - (n + 1)) \le {\ensuremath{\mathfrak{s}}}y
}
}{
\Gamma_1 {\ensuremath{\vdash}}101 - (n + 1) \le y
}
\\
(b0)
}{
\Gamma_1 {\ensuremath{\vdash}}f(n + 1) = 91
}$$ $$(b)
\inferrule*{
\inferrule*{
\inferrule*{
\Gamma_1 {\ensuremath{\vdash}}n + 11 > 100
}{
\Gamma_1 {\ensuremath{\vdash}}f(n + 11) = n + 11 - 10 = n + 1
}
\\
\inferrule*{
\Gamma_1 {\ensuremath{\vdash}}n < 101
}{
\Gamma_1 {\ensuremath{\vdash}}\neg (n > 100)
}
}{
\Gamma_1 {\ensuremath{\vdash}}f(n) = f(f(n + 11)) = f(n + 1)
}
\\
(b1)
}{
\Gamma_1 {\ensuremath{\vdash}}f(n) = 91
}$$ Let $\Gamma_2 \equiv \Gamma_0, n + 11 > 101, \neg (n < 101)$. $$(c)
\inferrule*{
\inferrule*{
\Gamma_2 {\ensuremath{\vdash}}n \le 101 \\ \Gamma_2 {\ensuremath{\vdash}}\neg (n < 101)
}{
\Gamma_2 {\ensuremath{\vdash}}n =_{\ensuremath{\mathrm{Nat}}}101
}
}{
\Gamma_2 {\ensuremath{\vdash}}f(n) = n - 10 = 101 - 10 = 91
}$$ $$(d)
\inferrule*{
\inferrule*{
\inferrule*{
\Gamma_0, n + 11 > 101 {\ensuremath{\vdash}}n : {\ensuremath{\mathrm{Nat}}}}{
\Gamma_0, n + 11 > 101 {\ensuremath{\vdash}}(n < 101) : {\ensuremath{\mathrm{Prop}}}}
}{
\Gamma_0, n + 11 > 101 {\ensuremath{\vdash}}n < 101 \lor \neg (n < 101)
}
\\
(b)
\\
(c)
}{
\Gamma_0, n + 11 > 101 {\ensuremath{\vdash}}f(n) = 91
}\;\raisebox{0.7em}{by~$\lor_e$}$$ Let $\Gamma_3 \equiv \Gamma_0, n + 11 \le 101$. $$(e0.1)
\inferrule*{
\Gamma_3 {\ensuremath{\vdash}}\varphi(y)
\\
\inferrule*{
\Gamma_3 {\ensuremath{\vdash}}n : {\ensuremath{\mathrm{Nat}}}\\ \Gamma_3 {\ensuremath{\vdash}}11 : {\ensuremath{\mathrm{Nat}}}}{
\Gamma_3 {\ensuremath{\vdash}}(n + 11) : {\ensuremath{\mathrm{Nat}}}}
}{
\Gamma_3 {\ensuremath{\vdash}}n + 11 \le 101 \supset 101 - (n + 11) \le y
\supset f(n + 11) = 91
}$$ $$(e0.2)
\inferrule*{
(e0.1)
\\
\Gamma_3 {\ensuremath{\vdash}}n + 11 \le 101
}{
\Gamma_3 {\ensuremath{\vdash}}101 - (n + 11) \le y \supset f(n + 11) = 91
}$$ $$(e0)
\inferrule*{
(e0.2)
\\
\inferrule*{
\Gamma_3 {\ensuremath{\vdash}}101 - n \le {\ensuremath{\mathfrak{s}}}y
}{
\Gamma_3 {\ensuremath{\vdash}}101 - (n + 11) \le y.
}
}{
\Gamma_3 {\ensuremath{\vdash}}f(n + 11) = 91
}$$ $$(e1)
\inferrule*{
\inferrule*{
\inferrule*{
\Gamma_3 {\ensuremath{\vdash}}n + 11 \le 101 \\ \Gamma_3 {\ensuremath{\vdash}}n : {\ensuremath{\mathrm{Nat}}}\\ \ldots
}{
\Gamma_3 {\ensuremath{\vdash}}n \le 100
}
}{
\Gamma_3 {\ensuremath{\vdash}}f(n) = f(f(n+11))
}
\\
(e0)
}{
\Gamma_3 {\ensuremath{\vdash}}f(n) = f(91)
}$$ $$(e2.1)
\inferrule*{
\Gamma_3 {\ensuremath{\vdash}}f(91) = f(f(102)) \\ \Gamma_3 {\ensuremath{\vdash}}f(102) = 92
}{
\Gamma_3 {\ensuremath{\vdash}}f(91) = f(92)
}$$ $$(e2.2)
\inferrule*{
\Gamma_3 {\ensuremath{\vdash}}f(92) = f(f(103)) \\ \Gamma_3 {\ensuremath{\vdash}}f(103) = 93
}{
\Gamma_3 {\ensuremath{\vdash}}f(92) = f(93)
}$$ $$\vdots$$ $$(e2.10)
\inferrule*{
\inferrule*{
\Gamma_3 {\ensuremath{\vdash}}f(100) = f(f(111)) \\ \Gamma_3 {\ensuremath{\vdash}}f(111) = 101
}{
\Gamma_3 {\ensuremath{\vdash}}f(100) = f(101)
}
\\
\Gamma_3 {\ensuremath{\vdash}}f(101) = 91
}{
\Gamma_3 {\ensuremath{\vdash}}f(100) = 91
}$$ $$(e2)
\inferrule*{
\inferrule*{
\inferrule*{
\inferrule*{
\inferrule*{
(e2.1)
\\
(e2.2)
}{
\Gamma_3 {\ensuremath{\vdash}}f(91) = f(93)
}
\\
(e2.3)
}{
\Gamma_3 {\ensuremath{\vdash}}f(91) = f(94)
}
\\
(e2.4)
}{
\vdots
}
\\
\vdots
}{
\Gamma_3 {\ensuremath{\vdash}}f(91) = f(100)
}
\\
(e2.10)
}{
\Gamma_3 {\ensuremath{\vdash}}f(91) = 91
}$$ $$(e)
\inferrule*{
(e1) \\ (e2)
}{
\Gamma_3 {\ensuremath{\vdash}}f(n) = 91
}$$ $$(f)
\inferrule*{
\inferrule*{
\inferrule*{
\Gamma_0 {\ensuremath{\vdash}}n : {\ensuremath{\mathrm{Nat}}}}{
\Gamma_0 {\ensuremath{\vdash}}(n + 11 > 101) : {\ensuremath{\mathrm{Prop}}}}
}{
\Gamma_0 {\ensuremath{\vdash}}n + 11 > 101 \lor n + 11 \le 101
}
\\
(d)
\\
(e)
}{
\Gamma_0 {\ensuremath{\vdash}}f(n) = 91
}\;\raisebox{0.7em}{by~$\lor_e$}$$ $$(g0)
\inferrule*{
y : {\ensuremath{\mathrm{Nat}}}, \varphi(y), n : {\ensuremath{\mathrm{Nat}}}{\ensuremath{\vdash}}(101 - n \le {\ensuremath{\mathfrak{s}}}y) : {\ensuremath{\mathrm{Prop}}}\\
(f)
}{
y : {\ensuremath{\mathrm{Nat}}}, \varphi(y), n : {\ensuremath{\mathrm{Nat}}}, n \le 101 {\ensuremath{\vdash}}101 - n \le {\ensuremath{\mathfrak{s}}}y \supset
f(n) = 91
}$$ $$(g1)
\inferrule*{
y : {\ensuremath{\mathrm{Nat}}}, \varphi(y), n : {\ensuremath{\mathrm{Nat}}}{\ensuremath{\vdash}}(n \le 101) : {\ensuremath{\mathrm{Prop}}}\\
(g0)
}{
y : {\ensuremath{\mathrm{Nat}}}, \varphi(y), n : {\ensuremath{\mathrm{Nat}}}{\ensuremath{\vdash}}n \le 101 \supset 101 - n
\le {\ensuremath{\mathfrak{s}}}y \supset f(n) = 91
}$$ $$(g)
\inferrule*{
{\ensuremath{\vdash}}{\ensuremath{\mathrm{Nat}}}: {\ensuremath{\mathrm{Type}}}\\
(g1)
}{
y : {\ensuremath{\mathrm{Nat}}}, \varphi(y) {\ensuremath{\vdash}}\varphi({\ensuremath{\mathfrak{s}}}y)
}$$ $$(h)
\inferrule*{
(a) \\ (g)
}{
{\ensuremath{\vdash}}\forall y, n : {\ensuremath{\mathrm{Nat}}}\,.\, n \le 101 \supset 101 - n \le y \supset
f(n) = 91
}\;\raisebox{0.7em}{by~$n_i$}$$ $$(i)
\inferrule*{
n : {\ensuremath{\mathrm{Nat}}}{\ensuremath{\vdash}}(101 - n) : {\ensuremath{\mathrm{Nat}}}\\
(h)
}{
n : {\ensuremath{\mathrm{Nat}}}{\ensuremath{\vdash}}\forall m : {\ensuremath{\mathrm{Nat}}}\,.\, m \le 101 \supset 101 - m
\le 101 - n \supset f(m) = 91
}$$ $$(j)
\inferrule*{
\inferrule*{
\inferrule*{
n : {\ensuremath{\mathrm{Nat}}}{\ensuremath{\vdash}}n : {\ensuremath{\mathrm{Nat}}}\\
(i)
}{
n : {\ensuremath{\mathrm{Nat}}}{\ensuremath{\vdash}}n \le 101 \supset 101 - n \le 101 - n \supset
f(n) = 91
}
}{
n : {\ensuremath{\mathrm{Nat}}}, n \le 101 {\ensuremath{\vdash}}101 - n \le 101 - n \supset f(n) = 91
}
}{
n : {\ensuremath{\mathrm{Nat}}}, n \le 101 {\ensuremath{\vdash}}f(n) = 91
}$$ $$\inferrule*{
\inferrule*{
n : {\ensuremath{\mathrm{Nat}}}{\ensuremath{\vdash}}(n \le 101) : {\ensuremath{\mathrm{Prop}}}\\
(j)
}{
n : {\ensuremath{\mathrm{Nat}}}{\ensuremath{\vdash}}n \le 101 \supset f(n) = 91
}
\\
{\ensuremath{\vdash}}{\ensuremath{\mathrm{Nat}}}: {\ensuremath{\mathrm{Type}}}}{
{\ensuremath{\vdash}}\forall n : {\ensuremath{\mathrm{Nat}}}\,.\, n \le 101 \supset f(n) = 91
}$$
Semantics {#sec_semantics}
=========
In this appendix we define a semantics for ${\ensuremath{{\cal I}}}_s$ and for $e{\ensuremath{{\cal I}}}_s$. This semantics will be used in Appendix F to show consistency of ${\ensuremath{{\cal I}}}_s$ and $e{\ensuremath{{\cal I}}}_s$.
\[def\_I\_s\_structure\] An *${\ensuremath{{\cal I}}}_s$-structure* is a triple ${\ensuremath{{\cal A}}}= \langle A, \cdot,
{\ensuremath{\llbracket\rrbracket_{}^{}}} \rangle$ where $A$ is the *domain* of ${\ensuremath{{\cal A}}}$, $\cdot$ is a binary operation on $A$, and the *interpretation* ${\ensuremath{\llbracket\rrbracket_{}^{}}} : {\ensuremath{\mathbb T}}\times A^V \to A$ is a function from ${\ensuremath{{\cal I}}}_s$-terms and valuations to $A$. We sometimes write $\cdot^{\ensuremath{{\cal A}}}$ and ${\ensuremath{\llbracket\rrbracket_{}^{{\ensuremath{{\cal A}}}}}}$ to indicate that these are components of ${\ensuremath{{\cal A}}}$.
A *valuation* $v$ is a function $v : V \to A$. We usually write ${\ensuremath{\llbrackett\rrbracket_{v}^{}}}$ instead of ${\ensuremath{\llbracket\rrbracket_{}^{}}}(t, v)$, and we drop the subscript when obvious or irrelevant. To stress that a valuation is associated with an ${\ensuremath{{\cal I}}}_s$-structure ${\ensuremath{{\cal A}}}$, we sometimes call it an ${\ensuremath{{\cal A}}}$-valuation. By $v[x/a]$ for $a \in A$ we denote a valuation $u$ such that $u(y) = v(y)$ for $y \ne x$ and $u(x) =
a$. We use the abbreviations ${\ensuremath{{\cal T}}}^{\ensuremath{{\cal A}}}= \{ a \in A \;|\;
{\ensuremath{\llbracketx : {\ensuremath{\mathrm{Type}}}\rrbracket_{\{x \mapsto a\}}^{{\ensuremath{{\cal A}}}}}} =
{\ensuremath{\llbracket\top\rrbracket_{}^{{\ensuremath{{\cal A}}}}}}\}$ and $\iota^{\ensuremath{{\cal A}}}= \{ a \in A \;|\;
{\ensuremath{\llbracketx : \iota\rrbracket_{\{x \mapsto a\}}^{{\ensuremath{{\cal A}}}}}} =
{\ensuremath{\llbracket\top\rrbracket_{}^{{\ensuremath{{\cal A}}}}}}\}$ for $\iota \in {\ensuremath{{\cal T}}}_I$. The symbols $a$, $a'$, $b$, $b'$, etc. denote elements of $A$, unless otherwise stated. We often confuse $\top$, $\bot$, $\mathrm{Is}$, etc. with ${\ensuremath{\llbracket\top\rrbracket_{}^{}}}$, ${\ensuremath{\llbracket\bot\rrbracket_{}^{}}}$, ${\ensuremath{\llbracket\mathrm{Is}\rrbracket_{}^{}}}$, etc., to avoid onerous notation. It is always clear from the context which interpretation is meant.
\[def\_I\_s\_model\] An ${\ensuremath{{\cal I}}}_s$-model is an ${\ensuremath{{\cal I}}}_s$-structure ${\ensuremath{{\cal A}}}$ satisfying the following requirements:
- ${\ensuremath{\llbracketx\rrbracket_{v}^{}}} = v(x)$ for $x \in V$,
- ${\ensuremath{\llbrackett_1 t_2\rrbracket_{v}^{}}} = {\ensuremath{\llbrackett_1\rrbracket_{v}^{}}} \cdot
{\ensuremath{\llbrackett_2\rrbracket_{v}^{}}}$,
- ${\ensuremath{\llbracket\lambda x \,.\, t\rrbracket_{v}^{}}} \cdot a =
{\ensuremath{\llbrackett\rrbracket_{v[x/a]}^{}}}$ for every $a \in A$,
- if ${\ensuremath{{v}_{\upharpoonrightFV(t)}}} = {\ensuremath{{w}_{\upharpoonrightFV(t)}}}$ then ${\ensuremath{\llbrackett\rrbracket_{v}^{}}} = {\ensuremath{\llbrackett\rrbracket_{w}^{}}}$,
- if for all $a \in A$ we have ${\ensuremath{\llbracket\lambda x
\,.\, t_1\rrbracket_{v}^{}}} \cdot a = {\ensuremath{\llbracket\lambda x \,.\, t_2\rrbracket_{v}^{}}} \cdot a$ then ${\ensuremath{\llbracket\lambda x \,.\, t_1\rrbracket_{v}^{}}} = {\ensuremath{\llbracket\lambda x
\,.\, t_2\rrbracket_{v}^{}}}$,
- $\mathrm{Is} \cdot a \cdot {\ensuremath{\mathrm{Prop}}}= \top$ iff $a \in
\{\top, \bot\}$,
- ${\ensuremath{\llbracket{\ensuremath{\mathrm{Prop}}}: {\ensuremath{\mathrm{Type}}}\rrbracket_{}^{}}} = \top$,
- ${\ensuremath{\llbracket\iota : {\ensuremath{\mathrm{Type}}}\rrbracket_{}^{}}} = \top$ for $\iota \in
{\ensuremath{{\cal T}}}_I$,
- if $a \in {\ensuremath{{\cal T}}}^{\ensuremath{{\cal A}}}$ and for all $c \in A$ such that $\mathrm{Is} \cdot c \cdot a = \top$ we have $b \cdot c = \top$ then $\forall \cdot a \cdot b = \top$,
- if $a \in {\ensuremath{{\cal T}}}^{\ensuremath{{\cal A}}}$ and there exists $c \in A$ such that $\mathrm{Is} \cdot c \cdot a = \top$ and $b \cdot c =
\bot$ then $\forall \cdot a \cdot b = \bot$,
- if $\forall \cdot a \cdot b = \top$ then for all $c \in A$ such that $\mathrm{Is} \cdot c \cdot a = \top$ we have $b \cdot c = \top$,
- if $\forall \cdot a \cdot b = \bot$ then there exists $c \in A$ such that $\mathrm{Is} \cdot c \cdot a = \top$ and $b \cdot c = \bot$,
- $\vee \cdot a \cdot b = \top$ iff $a = \top$ or $b
= \bot$,
- $\vee \cdot a \cdot b = \bot$ iff $a = \bot$ and $b = \bot$,
- if $\supset \cdot a \cdot b = \top$ then $a
\in \{\top,\bot\}$, where $\supset = {\ensuremath{\llbracket\lambda x y \,.\,
x \supset y\rrbracket_{}^{}}}$,
- if $\bot = \top$ then for all $a \in A$ we have $a =
\top$,
- if $f \in A$, $a \in {\ensuremath{{\cal T}}}^{\ensuremath{{\cal A}}}$ and for all $c \in A$ such that $\mathrm{Is} \cdot c \cdot a = \top$ we have $\mathrm{Is} \cdot (f \cdot c) \cdot b = \top$, then $\mathrm{Is}
\cdot f \cdot (\mathrm{Fun} \cdot a \cdot b) = \top$,
- if $\mathrm{Is} \cdot f \cdot (\mathrm{Fun} \cdot a
\cdot b) = \top$ and $\mathrm{Is} \cdot c \cdot a = \top$ then $\mathrm{Is} \cdot (f \cdot c) \cdot b = \top$,
- if $a, b \in {\ensuremath{{\cal T}}}^{\ensuremath{{\cal A}}}$ then $\mathrm{Fun} \cdot a \cdot
b \in {\ensuremath{{\cal T}}}^{\ensuremath{{\cal A}}}$,
- if $\mathrm{Subtype} \cdot a \cdot b \in {\ensuremath{{\cal T}}}^{\ensuremath{{\cal A}}}$, $\mathrm{Is} \cdot c \cdot a = \top$ and $b \cdot c = \top$, then $\mathrm{Is} \cdot c \cdot (\mathrm{Subtype} \cdot a \cdot b) =
\top$,
- if $\mathrm{Is} \cdot c \cdot (\mathrm{Subtype} \cdot a
\cdot b) = \top$ then $b \cdot c = \top$,
- if $\mathrm{Is} \cdot c \cdot (\mathrm{Subtype} \cdot a
\cdot b) = \top$ then $\mathrm{Is} \cdot c \cdot a = \top$,
- if $a \in {\ensuremath{{\cal T}}}^{\ensuremath{{\cal A}}}$ and for all $c \in A$ such that $\mathrm{Is} \cdot c \cdot a = \top$ we have $b \cdot c \in
\{\top, \bot\}$, then $\mathrm{Subtype} \cdot a \cdot b \in {\ensuremath{{\cal T}}}^{\ensuremath{{\cal A}}}$,
- $o_i^\iota \cdot (c_i^\iota \cdot a_1 \cdot \ldots \cdot
a_{n_i}) = \top$ where $o_i^\iota \in {\ensuremath{{\cal O}}}$ and $c_i^\iota \in {\ensuremath{{\cal C}}}$ has arity $n_i$,
- $o_i^\iota \cdot (c_j^\iota \cdot a_1 \cdot \ldots \cdot
a_{n_j}) = \bot$ where $i \ne j$, $o_i^\iota \in {\ensuremath{{\cal O}}}$, $c_j^\iota
\in {\ensuremath{{\cal C}}}$ has arity $n_j$,
- $d_{i,j}^\iota \cdot (c_i^\iota \cdot a_1 \cdot \ldots
\cdot a_{n_i}) = a_j$ where $d_{i,j}^\iota \in {\ensuremath{{\cal D}}}$ and $c_i^\iota
\in {\ensuremath{{\cal C}}}$ has arity $n_i$,
- if for all $i=1,\ldots,m$ we have:
- for all $b_{1},\ldots,b_{n_i}$ such that
- $b_{j} \in {\iota_{i,j}^*}^{\ensuremath{{\cal A}}}$ for $j=1,\ldots,n_i$, and
- $a \cdot b_{k} = \top$ for $1 \le k \le n_i$ such that $\iota_{i,k} = \star$
we have $a \cdot (c_i^\iota \cdot b_{1} \cdot \ldots \cdot
b_{n_i}) = \top$
then $\forall\cdot \iota^{\ensuremath{{\cal A}}}\cdot a = \top$, where $c_i^\iota \in
{\ensuremath{{\cal C}}}$ has arity $n_i$, $\iota \in {\ensuremath{{\cal T}}}_I$, $\iota_{i,j}^* =
\iota_{i,j}$ if $\iota_{i,j} \in {\ensuremath{{\cal T}}}_I$, $\iota_{i,j}^* = \iota$ if $\iota_{i,j} = \star$, and $\iota = \mu(\langle \iota_{1,1},
\ldots, \iota_{1,n_1} \rangle, \ldots, \langle \iota_{m,1},
\ldots, \iota_{m,m_i} \rangle)$,
- if $a_j \in {\iota_{i,j}^*}^{\ensuremath{{\cal A}}}$ for $i=1,\ldots,n_i$ then $c_i^\iota \cdot a_1 \cdot \ldots \cdot a_{n_i} \in
\iota^{\ensuremath{{\cal A}}}$, where $c_i^\iota \in {\ensuremath{{\cal C}}}$ has arity $n_i$, $\iota \in
{\ensuremath{{\cal T}}}_I$, $\iota_{i,j}^* = \iota_{i,j}$ if $\iota_{i,j} \in {\ensuremath{{\cal T}}}_I$, $\iota_{i,j}^* = \iota$ if $\iota_{i,j} = \star$, $\iota =
\mu(\langle \iota_{1,1}, \ldots, \iota_{1,n_1} \rangle, \ldots,
\langle \iota_{m,1}, \ldots, \iota_{m,m_i} \rangle)$,
- if for $a \in {\ensuremath{{\cal T}}}^{\ensuremath{{\cal A}}}$ there exists $b \in A$ such that $\mathrm{Is} \cdot b \cdot a = \top$ then $\mathrm{Is} \cdot
(\epsilon \cdot a) \cdot a = \top$,
- $\mathrm{Cond} \cdot \top \cdot a \cdot b = a$,
- $\mathrm{Cond} \cdot \bot \cdot a \cdot b = b$,
- $\mathrm{Eq} \cdot a \cdot b = \top$ iff $a = b$,
An *$e{\ensuremath{{\cal I}}}_s$-model* is an *${\ensuremath{{\cal I}}}_s$-model* which additionally satisfies:
- if $a, b \in {\ensuremath{{\cal T}}}^{\ensuremath{{\cal A}}}$, $\mathrm{Is} \cdot f \cdot
(\mathrm{Fun} \cdot a \cdot b) = \top$, $\mathrm{Is} \cdot g \cdot
(\mathrm{Fun} \cdot a \cdot b) = \top$, and for all $c \in A$ such that $\mathrm{Is} \cdot c \cdot a = \top$, and all $p \in A$ such that $\mathrm{Is} \cdot p \cdot (\mathrm{Fun} \cdot b \cdot {\ensuremath{\mathrm{Prop}}})
= \top$ we have $p \cdot (f \cdot c) = p \cdot (g \cdot c)$, then for all $p \in A$ such that $\mathrm{Is} \cdot p \cdot
(\mathrm{Fun} \cdot (\mathrm{Fun} \cdot a \cdot b) \cdot {\ensuremath{\mathrm{Prop}}}) =
\top$ we have $p \cdot f = p \cdot g$,
- if ${\supset} \cdot a \cdot b = {\supset} \cdot b \cdot
a = \top$ then $a = b$.
Here ${\supset} = {\ensuremath{\llbracket\lambda x y \,.\, x \supset y\rrbracket_{}^{}}}$.
For a term $\varphi$ and a valuation $u$ we write ${\ensuremath{{\cal A}}},u \models
\varphi$ if ${\ensuremath{\llbracket\varphi\rrbracket_{u}^{{\ensuremath{{\cal A}}}}}} = \top$. Given a set of terms $\Gamma$, we use the notation ${\ensuremath{{\cal A}}}, u \models \Gamma$ if ${\ensuremath{{\cal A}}},
u \models \varphi$ for all $\varphi \in \Gamma$. We drop the subscript ${\ensuremath{{\cal A}}}$ when obvious or irrelevant. We write $\Gamma
\models_{{\ensuremath{{\cal I}}}_s} \varphi$ if for every ${\ensuremath{{\cal I}}}_s$-model ${\ensuremath{{\cal A}}}$ and every valuation $u$, the condition ${\ensuremath{{\cal A}}}, u \models \Gamma$ implies ${\ensuremath{{\cal A}}}, u
\models \varphi$. The notation $\Gamma \models_{e{\ensuremath{{\cal I}}}_s} \varphi$ is defined analogously. The subscript is dropped when obvious from the context or irrelevant.
Note that every (nontrivial) ${\ensuremath{{\cal I}}}_s$-model is a $\lambda$-model. See e.g. [@Barendregt1984 Chapter 5] for a definition of a $\lambda$-model.
\[lem\_I\_s\_model\_conditions\] In every ${\ensuremath{{\cal I}}}_s$-model ${\ensuremath{{\cal A}}}$ the following conditions hold.
- ${\wedge} \cdot a \cdot b = \top$ iff $a = \top$ and $b = \top$,
- ${\wedge} \cdot a \cdot b = \bot$ iff $a = \bot$ or $b = \bot$,
- if $a \in {\ensuremath{{\cal T}}}^{\ensuremath{{\cal A}}}$ and there exists $c \in
A$ such that $\mathrm{Is} \cdot c \cdot a = \top$ and $b \cdot c =
\top$ then $\exists \cdot a \cdot b = \top$,
- if $a \in {\ensuremath{{\cal T}}}^{\ensuremath{{\cal A}}}$ and for all $c \in A$ such that $\mathrm{Is} \cdot c \cdot a = \top$ we have $b \cdot c
= \bot$ then $\exists \cdot a \cdot b = \bot$,
- if $\exists \cdot a \cdot b = \top$ then there exists $c \in A$ such that $\mathrm{Is} \cdot c \cdot a = \top$ and $b \cdot c = \top$,
- $\supset \cdot a \cdot b = \top$ iff $a =
\bot$, or $a = b =\top$,
- $\supset \cdot a \cdot b = \bot$ iff $a = \top$ and $b = \bot$,
- $\neg \cdot a = \bot$ iff $a = \top$,
- $\neg \cdot a = \top$ iff $a = \bot$.
Here $\supset = {\ensuremath{\llbracket\lambda x y \,.\, x \supset y\rrbracket_{}^{}}}$, $\neg = {\ensuremath{\llbracket\lambda x \,.\, \neg x\rrbracket_{}^{}}}$, $\wedge =
{\ensuremath{\llbracket\lambda x y \,.\, \neg ((\neg x) \vee (\neg y))\rrbracket_{}^{}}}$ and $\exists = {\ensuremath{\llbracket\lambda x y \,.\, \neg (\forall\, x\, \lambda
z \,.\, \neg(y z))\rrbracket_{}^{}}}$.
Easy.
\[lem\_I\_s\_ef\_condition\] In every ${\ensuremath{{\cal I}}}_s$-model ${\ensuremath{{\cal A}}}$, for any $a \in {\ensuremath{{\cal T}}}^{\ensuremath{{\cal A}}}$ and $b, c \in
{\ensuremath{{\cal A}}}$ such that $\mathrm{Is} \cdot b \cdot a = \top$ and $\mathrm{Is}
\cdot c \cdot a = \top$, the condition
- for all $p \in {\ensuremath{{\cal A}}}$ such that $\mathrm{Is} \cdot p \cdot
(\mathrm{Fun} \cdot a \cdot {\ensuremath{\mathrm{Prop}}}) = \top$ we have ${\supset}
\cdot (p \cdot b) \cdot (p \cdot c) = \top$
is equivalent to
- for all $p \in {\ensuremath{{\cal A}}}$ such that $\mathrm{Is} \cdot p \cdot
(\mathrm{Fun} \cdot a \cdot {\ensuremath{\mathrm{Prop}}}) = \top$ we have $p \cdot b = p
\cdot c$
where ${\supset} = {\ensuremath{\llbracket\lambda x y \,.\, x \supset y\rrbracket_{}^{}}}$.
It is obvious from the definition of ${\supset}$ and the conditions ($\to_e$) and (pr) that (2) implies (1). We show that (1) implies (2). Let $p \in {\ensuremath{{\cal A}}}$ be such that $\mathrm{Is} \cdot p \cdot
(\mathrm{Fun} \cdot a \cdot {\ensuremath{\mathrm{Prop}}}) = \top$ and let $b, c \in {\ensuremath{{\cal A}}}$ be such that $\mathrm{Is} \cdot b \cdot a = \top$ and $\mathrm{Is}
\cdot c \cdot a = \top$. By ($\to_e$) and (pr) we have $p \cdot b
\in \{\top,\bot\}$ and $p \cdot c \in \{\top,\bot\}$. If $p \cdot b
= \top$ then we must have $p \cdot c = \top$ as well, by definition of ${\supset}$. If $p \cdot b = \bot$ then it is easy to see that $p' = {\ensuremath{\llbracket\lambda x . \neg(y x)\rrbracket_{\{y \mapsto p\}}^{}}}$ also satisfies $\mathrm{Is} \cdot p' \cdot (\mathrm{Fun} \cdot a \cdot
{\ensuremath{\mathrm{Prop}}}) = \top$. We have $p' \cdot b = \top$, which implies $p' \cdot
c = \top$, which means that $p \cdot c = \bot$ by ($\neg_\bot$).
\[thm\_sound\] If $\Gamma {\ensuremath{\vdash}}_{{\ensuremath{{\cal I}}}} \varphi$ then $\Gamma \models_{\ensuremath{{\cal I}}}\varphi$, where ${\ensuremath{{\cal I}}}= {\ensuremath{{\cal I}}}_s$ or ${\ensuremath{{\cal I}}}= e{\ensuremath{{\cal I}}}_s$.
Induction on the length of derivation. All cases follow easily from appropriate conditions in the definition of an ${\ensuremath{{\cal I}}}_s$-model or from Lemma \[lem\_I\_s\_model\_conditions\]. With the rule $e_f$ we also need to use Lemma \[lem\_I\_s\_ef\_condition\].
For instance, suppose $\Gamma {\ensuremath{\vdash}}{\ensuremath{\mathrm{Eq}\,({\ensuremath{\mathrm{Cond}\,\varphi\,t_1\,t_2\,}})\,({\ensuremath{\mathrm{Cond}\,\varphi\,t_1'\,t_2\,}})\,}}$ was obtained by rule $c_3$. By the inductive hypothesis we have $\Gamma,
\varphi \models {\ensuremath{\mathrm{Eq}\,t_1\,t_1'\,}}$ and $\Gamma \models \varphi :
{\ensuremath{\mathrm{Prop}}}$. Assume ${\ensuremath{{\cal M}}}, u \models \Gamma$. We have ${\ensuremath{{\cal M}}}, u \models
\varphi : {\ensuremath{\mathrm{Prop}}}$, so ${\ensuremath{\llbracket\varphi : {\ensuremath{\mathrm{Prop}}}\rrbracket_{u}^{{\ensuremath{{\cal M}}}}}} =
\top$. By condition (pr) we obtain ${\ensuremath{\llbracket\varphi\rrbracket_{u}^{{\ensuremath{{\cal M}}}}}} \in
\{\top, \bot\}$. Suppose ${\ensuremath{\llbracket\varphi\rrbracket_{u}^{{\ensuremath{{\cal M}}}}}} = \top$. Then ${\ensuremath{{\cal M}}}, u \models \Gamma, \varphi$, so ${\ensuremath{\llbracket{\ensuremath{\mathrm{Eq}\,t_1\,t_1'\,}}\rrbracket_{u}^{{\ensuremath{{\cal M}}}}}} = \top$. Hence by conditions (app), (eq) and (c1), we have ${\ensuremath{\llbracket{\ensuremath{\mathrm{Cond}\,\varphi\,t_1\,t_2\,}}\rrbracket_{u}^{{\ensuremath{{\cal M}}}}}} =
{\ensuremath{\llbracket{\ensuremath{\mathrm{Cond}\,\varphi\,t_1'\,t_2\,}}\rrbracket_{u}^{{\ensuremath{{\cal M}}}}}}$. By conditions (app) and (eq) we obtain ${\ensuremath{{\cal M}}}, u \models
{\ensuremath{\mathrm{Eq}\,({\ensuremath{\mathrm{Cond}\,\varphi\,t_1\,t_2\,}})\,({\ensuremath{\mathrm{Cond}\,\varphi\,t_1'\,t_2\,}})\,}}$. So suppose ${\ensuremath{\llbracket\varphi\rrbracket_{u}^{{\ensuremath{{\cal M}}}}}} = \bot$. Then we obtain the thesis by applying conditions (app), (c2) and (eq).
Other cases are established in a similar way.
If $\Gamma \models_{e{\ensuremath{{\cal I}}}_s} \varphi$ then $\Gamma {\ensuremath{\vdash}}_{e{\ensuremath{{\cal I}}}_s}
\varphi$.
We do not attempt to prove the above conjecture in this paper, as it is not necessary for establishing consistency of $e{\ensuremath{{\cal I}}}_s$, which is our main concern here. We treat the semantics given merely as a technical device in the consistency proof.
Model Construction {#sec_construction}
==================
In this appendix we construct a nontrivial $e{\ensuremath{{\cal I}}}_s$-model, thus establishing consistency of the systems ${\ensuremath{{\cal I}}}_s$ and $e{\ensuremath{{\cal I}}}_s$. The construction is parametrized by a set of constants $\delta$. We will use this construction in the next appendix to show a complete translation of classical first-order logic into ${\ensuremath{{\cal I}}}_s$ and $e{\ensuremath{{\cal I}}}_s$.
The construction is an adaptation and extension of the one from [@Czajka2013Accepted] for the traditional illative system ${\ensuremath{{\cal I}}}_\omega$. The proof is perhaps a bit easier to understand, because we do not have to deal with certain oddities of traditional illative systems from [@Czajka2013Accepted], like e.g. the fact that $t : {\ensuremath{\mathrm{Prop}}}$ is defined as equivalent to $(\lambda x \,.\, t) :
{\ensuremath{\mathrm{Type}}}$ where $x \notin FV(t)$, using the notation of the present paper.
For two sets $\tau_1$ and $\tau_2$, we denote by $\tau_2^{\tau_1}$ the set of all (set-theoretical) functions from $\tau_1$ to $\tau_2$. Formally, $f \in \tau_2^{\tau_1}$ is a subset of $\tau_1
\times \tau_2$ such that for every $f_1 \in \tau_1$ there is exactly one $f_2 \in \tau_2$ such that ${\langlef_1,f_2\rangle} \in f$.
\[def\_canonical\] The subset ${\ensuremath{\mathbb T}}_I$ of ${\ensuremath{{\cal I}}}_s$-terms is defined inductively by the following rule:
- for every $c_i^\iota \in {\ensuremath{{\cal C}}}$ of arity $n_i$, if $t_1,\ldots,t_{n_i} \in {\ensuremath{\mathbb T}}_I$ then $c_i^\iota t_1 \ldots t_{n_i}
\in {\ensuremath{\mathbb T}}_I$.
The set ${\ensuremath{{\cal K}}}$ consists of unique fresh constants, one for each element of ${\ensuremath{\mathbb T}}_I$. By $\bar{c}_i^\iota(\bar{t}_1,\ldots,\bar{t}_{n_i}) \in {\ensuremath{{\cal K}}}$ we denote the constant corresponding to $c_i^\iota t_1 \ldots t_{n_i} \in
{\ensuremath{{\cal K}}}$. Each basic inductive type $\iota \in {\ensuremath{{\cal T}}}_I$ determines in an obvious way a subset of ${\ensuremath{\mathbb T}}_I$ (the set of all terms generated by the constructors of $\iota$, respecting argument types), and thus it determines a subset $\bar{\iota} \subseteq {\ensuremath{{\cal K}}}$. Let $\bar{{\ensuremath{{\cal T}}}}_I =
\{\bar{\iota} \;|\; \iota \in {\ensuremath{{\cal T}}}_I\}$. Note that $\bigcup
\bar{{\ensuremath{{\cal T}}}}_I = {\ensuremath{{\cal K}}}$ and $\bar{\iota_1} \cap \bar{\iota_2} =
\emptyset$ if $\iota_1 \ne \iota_2$.
Let ${\ensuremath{\mathrm{Bool}}}= \{{\ensuremath{\mathbf{t}}}, {\ensuremath{\mathbf{f}}}\}$ and ${\ensuremath{{\cal T}}}_0 = \{\delta, {\ensuremath{\mathrm{Bool}}},
\emptyset\} \cup \bar{{\ensuremath{{\cal T}}}}_I$, where ${\ensuremath{\mathbf{t}}}$ and ${\ensuremath{\mathbf{f}}}$ are fresh constants. We define ${\ensuremath{{\cal T}}}_{n+1}$ as follows.
- If $\tau \in {\ensuremath{{\cal T}}}_n$ and $\tau' \subseteq \tau$ then $\tau' \in
{\ensuremath{{\cal T}}}_{n+1}$.
- If $\tau_1, \tau_2 \in {\ensuremath{{\cal T}}}_n$ then $\tau_2^{\tau_1} \in
{\ensuremath{{\cal T}}}_{n+1}$.
The set of *types* is now defined by ${\ensuremath{{\cal T}}}=
\bigcup_{n\in{\ensuremath{\mathbb{N}}}}{\ensuremath{{\cal T}}}_n$.
We define a notation $\tau^{(\tau_1,\ldots,\tau_n)}$ inductively as follows:
- $\tau^{(\tau_1)} = \tau^{\tau_1}$,
- $\tau^{(\tau_1,\tau_2,\ldots,\tau_n)} =
(\tau^{(\tau_2,\ldots,\tau_n)})^{\tau_1}$.
We define the set of *canonical constants* as $\Sigma = \delta
\cup {\ensuremath{\mathrm{Bool}}}\cup {\ensuremath{{\cal K}}}\cup \Sigma_f$, where $\Sigma_f$ contains a unique fresh constant for each function in $\bigcup {\ensuremath{{\cal T}}}$. We denote the function corresponding to a constant $c \in \Sigma_f$ by ${\ensuremath{{\cal F}}}(c)$. To save on notation we often confuse constants $c \in
\Sigma_f$ with their corresponding functions, and write e.g. $c \in
\tau$ instead of ${\ensuremath{{\cal F}}}(c) \in \tau$, for $\tau \in {\ensuremath{{\cal T}}}$. It is always clear from the context what we mean. Note that if ${\ensuremath{{\cal F}}}(c) \in
\tau_2^{\tau_1}$ then $\tau_1$ is uniquely determined. This is because ${\ensuremath{{\cal F}}}(c)$ is a set-theoretical function, i.e. a set of pairs, so its domain is uniquely determined. Note also that if ${\ensuremath{{\cal F}}}(c) \in
\tau_2^{\tau_1}$ and ${\ensuremath{{\cal F}}}(c) \in {\tau_2'}^{\tau_1}$, then ${\ensuremath{{\cal F}}}(c)
\in {(\tau_2 \cap \tau_2')}^{\tau_1}$.
Let $v, w \in \Sigma^*$. We write $v \sqsubseteq w$ if $w = vu$ for some $u \in \Sigma^*$, i.e. when $v$ is a prefix of $w$. We use the notation $v \sqsubset w$ when $v \sqsubseteq w$ and $v \ne w$.
A *$\Sigma$-tree* $T$ is a set of strings over the alphabet $\Sigma$, i.e. a subset of $\Sigma^*$, satisfying the following conditions:
- if $w \in T$ and $v \sqsubseteq w$ then $v \in T$ (prefix-closedness),
- $\sqsubseteq^{-1}$ is well-founded on $T$ (no infinite branches).
A *node* of a $\Sigma$-tree $T$ is any $w \in T$. We say that a node $w \in T$ is a *leaf* if there is no $w' \in T$ such that $w \sqsubset w'$. If $w \in T$ is not a leaf, then it is an *internal node*. The *root* of a $\Sigma$-tree is the empty string $\epsilon$.
We say that $T_1$ is a *subtree* of $T_2$ if there exists $w
\in T_2$ such that $T_1 = \{v \in \Sigma^* \;|\; wv \in T_2\}$.
Note that a relation $\le$, defined by $T_1 \le T_2$ iff $T_1$ is a subtree of $T_2$, is a well-founded partial order, because $\Sigma$-trees have no infinite branches. This allows us to perform *induction on the structure of a $\Sigma$-tree*. We write $T_1
< T_2$ if $T_1 \le T_2$ and $T_1 \ne T_2$.
The *height* $h(T)$ of a $\Sigma$-tree $T$ is an ordinal defined by induction on the structure of $T$.
- If $T = \emptyset$ then $h(T) = 0$.
- If $T \ne \emptyset$ then $h(T) = \sup_{T' < T} (h(T') + 1)$.
The set of constants $\Sigma^+$ is defined as $$\Sigma^+ = \Sigma \cup {\ensuremath{{\cal T}}}\cup \{\forall, \vee, \mathrm{Is},
\mathrm{Subtype}, \mathrm{Fun}, \mathrm{Eq}, \mathrm{Cond},
\mathrm{Choice}, {\ensuremath{\mathrm{Type}}}\} \cup {\ensuremath{{\cal C}}}\cup {\ensuremath{{\cal D}}}\cup {\ensuremath{{\cal O}}}$$ where ${\ensuremath{{\cal C}}}$, ${\ensuremath{{\cal D}}}$ and ${\ensuremath{{\cal O}}}$ are the sets of constructors, destructors and tests from the definition of ${\ensuremath{{\cal I}}}_s$. Note that for each $\tau \in {\ensuremath{{\cal T}}}$ we have $\tau \in \Sigma^+$ as a constant.
We use $V^+$ to denote a set of variables of cardinality at least the cardinality of $\Sigma^*$.
The set of *operation symbols* ${\ensuremath{\mathrm{Op}}}$ is defined to contain the following:
- $\cdot \in {\ensuremath{\mathrm{Op}}}$,
- if $x \in V^+$ then $\lambda x \in {\ensuremath{\mathrm{Op}}}$,
- if $\tau \in {\ensuremath{{\cal T}}}$ and $\tau \ne \emptyset$ then ${\ensuremath{\mathrm{A}}}\tau \in
{\ensuremath{\mathrm{Op}}}$ and ${\ensuremath{\mathrm{S}}}\tau \in {\ensuremath{\mathrm{Op}}}$.
Intuitively, ${\ensuremath{\mathrm{A}}}\tau$ means “for all elements of $\tau$ satisfying …”, and ${\ensuremath{\mathrm{S}}}\tau$ means “a subtype of $\tau$ consisting of elements satisfying …”. These will appear as node labels in a $\Sigma$-tree representing a semantic term. A node labelled with e.g. ${\ensuremath{\mathrm{S}}}\tau$ will have a child corresponding to each element of $\tau$. The subtype represented by this node will consist of those elements for which the corresponding child reduces to ${\ensuremath{\mathbf{t}}}$.
A *semantic term* is a pair ${\langle{\mbox{\cal Pos}},\kappa\rangle}$, where ${\mbox{\cal Pos}}$ is a $\Sigma$-tree and $\kappa : {\mbox{\cal Pos}}\to {\ensuremath{\mathrm{Op}}}\cup \Sigma^+ \cup
V^+$ is a function such that:
- $w \in {\mbox{\cal Pos}}$ is a leaf iff $\kappa(w) \in \Sigma^+ \cup V^+$,
- if $\kappa(w) = \lambda x$ then $w0 \in {\mbox{\cal Pos}}$ and $wc \notin {\mbox{\cal Pos}}$ for $c \ne 0$,
- if $\kappa(w) = \cdot$ then $w0, w1 \in {\mbox{\cal Pos}}$ and $wc \notin
{\mbox{\cal Pos}}$ for $c \notin \{0,1\}$,
- if $\kappa(w) \in \{ {\ensuremath{\mathrm{A}}}\tau, {\ensuremath{\mathrm{S}}}\tau\}$ then $wc \in
{\mbox{\cal Pos}}$ iff $c \in \tau$.
In other words, semantic terms are possibly infinitely branching well-founded trees, whose internal nodes are labelled with operation symbols ${\ensuremath{\mathrm{Op}}}$, and leaves are labelled with constants from $\Sigma^+$ or variables from $V^+$.
We usually denote semantic terms by $t$, $t_1$, $t_2$, etc. We write ${\mbox{\cal Pos}}(t)$ for the underlying tree of $t$, and ${\ensuremath{{t}^{|p}}}$ instead of $\kappa(p)$.
The *height* of a semantic term $t$, denoted $h(t)$, is the height of its associated $\Sigma$-tree. When we say that we perform *induction on the structure of a semantic term* we mean induction on $h(t)$. By induction on an ordinal $\alpha$ and the structure of a semantic term we mean induction on pairs ${\langle\alpha,h(t)\rangle}$ ordered lexicographically.
A *position* in a semantic term $t$ is a string $w \in
{\mbox{\cal Pos}}(t)$. The *subterm* of $t$ at position $p \in {\mbox{\cal Pos}}(t)$, denoted ${\ensuremath{{t}_{|p}}}$, is a semantic term ${\langle{\mbox{\cal Pos}}',\kappa'\rangle}$ where:
- ${\mbox{\cal Pos}}' = \{ w \in \Sigma^* \;|\; pw \in {\mbox{\cal Pos}}(t) \}$,
- $\kappa'(w) = \kappa(pw)$.
A variable $x \in V^+$ is *free* in a semantic term $t$ if there exists $p \in {\mbox{\cal Pos}}(t)$ such that ${\ensuremath{{t}^{|p}}} = x$ and for no $p' \sqsubseteq p$ we have ${\ensuremath{{t}^{|p}}} = \lambda x$. A variable is *bound* if it is not free.
We identify $\alpha$-equivalent semantic terms, i.e. ones differing only in the names of bound variables. We assume that no semantic term contains some variable both free and bound. We use the symbol $\equiv$ for identity of semantic terms up to $\alpha$-equivalence.
Substitution $t[x/t']$ for semantic terms is defined in an obvious way, avoiding variable capture. In other words, we adopt the convention that whenever we write a term of the form $t[x/t']$ we assume that no free variables of $t'$ become bound in $t[x/t']$.
In this section, when we speak of terms we mean semantic terms, unless otherwise stated. We often use abbreviations for semantic terms of the form $\lambda x \,.\, t$, $t_1 t_2$, $\lambda x
. (\mathrm{Fun}\,{x}\,{(t_1 t_2)})$, etc. The meaning of these abbreviations is obvious.
A *rewriting system* $R$ is a set of pairs of semantic terms. We usually write ${\ensuremath{t_1\rightarrowt_2}} \in R$ instead of ${\langlet_1,t_2\rangle} \in R$. A term $t$ is said to *$R$-contract* to a term $t'$ at position $p$, denoted $t {\ensuremath{\rightarrow}}_R^p t'$, if $p \in
{\mbox{\cal Pos}}(t) \cap {\mbox{\cal Pos}}(t')$, the terms $t$ and $t'$ differ only in subterms at position $p$, and there exists ${\ensuremath{t_1\rightarrowt_2}} \in
R$ such that ${\ensuremath{{t}_{|p}}} \equiv t_1$ and ${\ensuremath{{t'}_{|p}}} \equiv
t_2$. We write $t {\ensuremath{\rightarrow}}_R t'$ if $t {\ensuremath{\rightarrow}}_R^p t'$ for some $p \in
{\mbox{\cal Pos}}(t)$.
For each ordinal $\alpha$, we define two relations ${\ensuremath{\Rightarrow}}_R^\alpha$ and $\succ_R^\alpha$ by induction on the ordinal $\alpha$ and the structure of $t$.
- If $c \in \Sigma$ then $c \succ_R^\alpha c$.
- If $c \in \tau_2^{\tau_1}$ and for all $c_1 \in \tau_1$ there exists $t'$ such that $t c_1 {\ensuremath{\stackrel{*\;<\alpha}{\Rightarrow_{R}\;\,}}} t'
\succ_R^{<\alpha} {\ensuremath{{\cal F}}}(c)(c_1)$, then $t \succ_R^\alpha c$.
- If $t \equiv t'$ or $t {\ensuremath{\rightarrow}}_R^\epsilon t'$ then $t
{\ensuremath{\Rightarrow}}_R^\alpha t'$.
- If $t_1 {\ensuremath{\Rightarrow}}_R^\alpha t_1'$ and $t_2
{\ensuremath{\Rightarrow}}_R^\alpha t_2'$ then $t_1 t_2 {\ensuremath{\Rightarrow}}_R^\alpha t_1'
t_2'$.
- If $t {\ensuremath{\Rightarrow}}_R^\alpha t'$ then $\lambda x \,.\, t
{\ensuremath{\Rightarrow}}_R^\alpha \lambda x \,.\, t'$.
- If $t_1 {\ensuremath{\Rightarrow}}_R^{\alpha} t_1'$ and $t_2
{\ensuremath{\Rightarrow}}_R^{\alpha} t_2'$ then $(\lambda x \,.\, t_1) t_2
{\ensuremath{\Rightarrow}}_R^\alpha t_1'[x/t_2']$.
- If $c \in \Sigma^+$, $n \in {\ensuremath{\mathbb{N}}}$, $c t_1' \ldots t_n'
{\ensuremath{\rightarrow}}_R^\epsilon t$ and $t_i {\ensuremath{\Rightarrow}}_R^\alpha t_i'$ for $i=1,\ldots,n$, then $c t_1 \ldots t_n {\ensuremath{\Rightarrow}}_R^\alpha t$.
- If $t \succ_R^{<\alpha} c_1 \in \tau_1$ and $c \in
\tau_2^{\tau_1}$, then $c t {\ensuremath{\Rightarrow}}_R^\alpha {\ensuremath{{\cal F}}}(c)(c_1)$.
- If $t \succ_R^{<\alpha} c$ for some $c \in \tau \in {\ensuremath{{\cal T}}}$ then ${\ensuremath{\mathrm{Is}\,t\,\tau\,}} {\ensuremath{\Rightarrow}}_R^\alpha {\ensuremath{\mathbf{t}}}$.
- If ${\ensuremath{{t}^{|\epsilon}}} \equiv {\ensuremath{{t'}^{|\epsilon}}} \equiv
{\ensuremath{\mathrm{A}}}\tau$ or ${\ensuremath{{t}^{|\epsilon}}} \equiv {\ensuremath{{t'}^{|\epsilon}}}
\equiv {\ensuremath{\mathrm{S}}}\tau$, and for all $c \in \tau$ there exists $t''$ such that ${\ensuremath{{t}_{|c}}} {\ensuremath{\Rightarrow}}_R^\alpha t'' {\ensuremath{\stackrel{*\;<\alpha}{\Rightarrow_{R}\;\,}}}
{\ensuremath{{t'}_{|c}}}$, then $t {\ensuremath{\Rightarrow}}_R^\alpha t'$.
The notation ${\ensuremath{\Rightarrow}}_R^{<\alpha}$ is an abbreviation for $\bigcup_{\gamma<\alpha}{\ensuremath{\Rightarrow}}_R^{\gamma}$, and ${\ensuremath{\stackrel{*\;<\alpha}{\Rightarrow_{R}\;\,}}}$ denotes the transitive-reflexive closure of ${\ensuremath{\Rightarrow}}_R^{<\alpha}$. The notation $\succ_R^{<\alpha}$ is an abbreviation for $\bigcup_{\gamma<\alpha}\succ_R^\gamma$.
We define the relations ${\ensuremath{\Rightarrow}}_R$ and $\succ_R$ as the smallest fixpoint of the above construction, i.e. by monotonicity of the definition there exists the least ordinal $\zeta$ such that ${\ensuremath{\Rightarrow}}_R^\zeta \,=\, {\ensuremath{\Rightarrow}}_R^{<\zeta}$ and $\succ_R^\zeta
\,=\, \succ_R^{<\zeta}$, and we take ${\ensuremath{\Rightarrow}}_R \,=\,
{\ensuremath{\Rightarrow}}_R^\zeta$, $\succ_R \,=\, \succ_R^\zeta$. Note that $\omega$ steps do not suffice to reach the fixpoint. In fact, the ordinal $\zeta$ will be quite large.
We denote by ${\ensuremath{\stackrel{*}{\Rightarrow}}}_R$ the transitive-reflexive closure of ${\ensuremath{\Rightarrow}}_R$, and by ${\ensuremath{\stackrel{*}{\Leftrightarrow}}}_R$ the transitive-reflexive-symmetric closure of ${\ensuremath{\Rightarrow}}_R$. The subscript is often dropped when obvious from the context.
Notice that the relation ${\ensuremath{\Rightarrow}}_R$ encompasses an analogon of $\beta$-reduction, regardless of what the rules of $R$ are. Intuitively, the relation ${\ensuremath{\Rightarrow}}_R$ is a kind of parallel reduction on semantic terms, parametrized by the rules of $R$.
If $R \subseteq R'$ then ${\ensuremath{\Rightarrow}}_R \,\subseteq\,{\ensuremath{\Rightarrow}}_{R'}$ and $\succ_R \,\subseteq\,\succ_{R'}$.
Intuitively, and very informally, $t \succ_R c$ is intended to hold if $c \in \tau \in {\ensuremath{{\cal T}}}$ is a “canonical” object which “simulates” $t$ in type $\tau$. By $c$ “simulating” $t$ in type $\tau$ we mean that $c$ behaves in essentially the same way as $t$, modulo ${\ensuremath{\Rightarrow}}_R$, whenever a term of type $\tau$ is “expected”. Let us give some examples to elucidate what we mean by this. For instance, let $c_1 \in {\ensuremath{\mathrm{Nat}}}^{\ensuremath{\mathrm{Nat}}}= \tau_1$, $c_2 \in
\delta^\delta = \tau_2$ be two constants such that ${\ensuremath{{\cal F}}}(c_1)(c) \equiv
c$ for all $c \in {\ensuremath{\mathrm{Nat}}}$ and ${\ensuremath{{\cal F}}}(c_2)(c) \equiv c$ for all $c \in
\delta$. Note that by condition (6) in the definition of ${\ensuremath{\Rightarrow}}_R$ and the fact that $c_1 \succ_R c_1$ and $c_2 \succ_R c_2$ we have $c_1
c {\ensuremath{\Rightarrow}}_R c$ for all $c \in {\ensuremath{\mathrm{Nat}}}$ and $c_2 c {\ensuremath{\Rightarrow}}_R c$ for all $c \in \delta$. Now we have both $\lambda x \,.\, x \succ_R c_1$ and $\lambda x \,.\, x \succ_R c_2$, because $\lambda x \,.\, x$ behaves exactly like $c_1$ when given arguments of type ${\ensuremath{\mathrm{Nat}}}$, and exactly like $c_2$ when given arguments of type $\delta$. The condition (6) ensures that $\lambda x . x$ and $c_1$ will be indistinguishable wherever a term of type $\tau_1$ is “expected”. For instance, if $d
\in \delta^{\tau_1}$ then we have $d (\lambda x \,.\, x) {\ensuremath{\Rightarrow}}_R
{\ensuremath{{\cal F}}}(d)(c_1)$. In fact, we will later prove that, for an appropriate rewriting system $R$, the conditions $t c {\ensuremath{\stackrel{*}{\Rightarrow}}}_R t' \succ_R c'$ and $r \succ_R c$ imply the existence of $t''$ such that $t r
{\ensuremath{\stackrel{*}{\Rightarrow}}}_R t'' \succ_R c'$, where $t$ is an arbitrary term.
Note that we may have $c_1 \succ_R c_2$ with $c_1, c_2 \in \Sigma$, $c_1 \not\equiv c_2$ (i.e. $c_1$ and $c_2$ being two distinct canonical constants), if there does not exist a single $\tau \in {\ensuremath{{\cal T}}}$ such that $c_1,c_2 \in \tau$. We will later show that this is not possible, for a rewriting system $R$ to be defined, if such a $\tau
\in {\ensuremath{{\cal T}}}$ does exist.
We will build our model from equivalence classes of ${\ensuremath{\stackrel{*}{\Leftrightarrow}}}_R$ on semantic terms, for a certain rewriting system $R$ to be defined below. One of the main problems in the model construction is to ensure that the condition $(\forall_e)$ of Definition \[def\_I\_s\_model\] holds. The problem is that condition $(\to_i)$ needs to be satisfied as well, which means that we cannot know a priori which terms $t$ should satisfy ${\ensuremath{\mathrm{Is}\,t\,\tau\,}}$ for a function type $\tau \in {\ensuremath{{\cal T}}}$, because this must depend on the definition of ${\ensuremath{\Rightarrow}}_R$ for $(\to_i)$ to hold. And we cannot use a conditional rule of the form
- if for all $t$ such that ${\ensuremath{\mathrm{Is}\,t\,\tau\,}} {\ensuremath{\stackrel{*}{\Rightarrow}}}_R {\ensuremath{\mathbf{t}}}$ we have $t_1 t {\ensuremath{\stackrel{*}{\Rightarrow}}}_R {\ensuremath{\mathbf{t}}}$ then $\forall\,\tau\,t_1 {\ensuremath{\Rightarrow}}_R
{\ensuremath{\mathbf{t}}}$
because the definition would not be monotone. Our solution is to restrict quantification to canonical constants only, and to define the relation ${\ensuremath{\Rightarrow}}_R$ in such a way as to ensure that for each term $t$ with ${\ensuremath{\mathrm{Is}\,t\,\tau\,}} {\ensuremath{\stackrel{*}{\Rightarrow}}}_R {\ensuremath{\mathbf{t}}}$ there exist a canonical constant $c \in \tau$ and a term $t'$ such that $t {\ensuremath{\stackrel{*}{\Rightarrow}}}_R t'
\succ c$.
Let $\chi$ be a choice function for the family of sets ${\ensuremath{{\cal T}}}\setminus \{\emptyset\}$. We define a rewriting system $R$ by the rules presented in Fig. \[fig\_1\].
$$\begin{aligned}
{\ensuremath{\mathrm{Eq}\,t_1\,t_2\,}} &\to& {\ensuremath{\mathbf{t}}}\;\;\; \mathrm{if\ } t_1 {\ensuremath{\stackrel{*}{\Leftrightarrow}}}_R t_2 \\
\forall \tau t &\to& t' \;\;\; \mathrm{where\ } \tau \in {\ensuremath{{\cal T}}},\,
\tau \ne \emptyset,\, {\ensuremath{{t'}^{|\epsilon}}} \equiv {\ensuremath{\mathrm{A}}}\tau \mathrm{\ and\ }
{\ensuremath{{t'}_{|c}}} \equiv t c \mathrm{\ for\ } c \in \tau \\
\mathrm{Subtype}\, \tau\, t &\to& t' \;\;\; \mathrm{where\ }
\tau \in {\ensuremath{{\cal T}}},\, \tau \ne \emptyset,\, {\ensuremath{{t'}^{|\epsilon}}} \equiv {\ensuremath{\mathrm{S}}}\tau
\mathrm{\ and\ } {\ensuremath{{t'}_{|c}}} \equiv t c \mathrm{\ for\ } c \in
\tau \\
\forall \emptyset t &\to& {\ensuremath{\mathbf{t}}}\\
\mathrm{Subtype}\, \emptyset\, t &\to& \emptyset \\
t &\to& {\ensuremath{\mathbf{t}}}\;\;\; \mathrm{if\ } {\ensuremath{{t}^{|\epsilon}}} \equiv {\ensuremath{\mathrm{A}}}\tau \mathrm{\ and\ for\ all\ } c \in \tau \mathrm{\ we\ have\ } {\ensuremath{{t}_{|c}}} \equiv {\ensuremath{\mathbf{t}}}\\
t &\to& {\ensuremath{\mathbf{f}}}\;\;\; \mathrm{if\ } {\ensuremath{{t}^{|\epsilon}}} \equiv {\ensuremath{\mathrm{A}}}\tau \mathrm{\ and\ there\ exists\ } c \in
\tau \mathrm{\ such\ that\ } {\ensuremath{{t}_{|c}}} \equiv {\ensuremath{\mathbf{f}}}\\
t &\to& \tau' \;\;\; \mathrm{if\ } {\ensuremath{{t}^{|\epsilon}}} \equiv {\ensuremath{\mathrm{S}}}\tau \mathrm{\ and\ for\ all\ } c \in \tau \mathrm{\ we\ have\ }
{\ensuremath{{t}_{|c}}} \in \{{\ensuremath{\mathbf{t}}},{\ensuremath{\mathbf{f}}}\}, \\ &&
\;\;\;\;\;\;\;\;\mathrm{\ and\ } \tau' = \{ c \in \tau \;|\; {\ensuremath{{t}_{|c}}} \equiv {\ensuremath{\mathbf{t}}}\} \\
\mathrm{Fun}\,\tau_1\, \tau_2 &\to& \tau_2^{\tau_1} \;\;\; \mathrm{for\ } \tau_1, \tau_2 \in {\ensuremath{{\cal T}}}\\
{\ensuremath{\mathrm{Cond}\,{\ensuremath{\mathbf{t}}}\,t_1\,t_2\,}} &\to& t_1 \\
{\ensuremath{\mathrm{Cond}\,{\ensuremath{\mathbf{f}}}\,t_1\,t_2\,}} &\to& t_2 \\
\vee {\ensuremath{\mathbf{t}}}t &\to& {\ensuremath{\mathbf{t}}}\\
\vee t {\ensuremath{\mathbf{t}}}&\to& {\ensuremath{\mathbf{t}}}\\
\vee {\ensuremath{\mathbf{f}}}{\ensuremath{\mathbf{f}}}&\to& {\ensuremath{\mathbf{f}}}\\
c_i^\iota \bar{t}_1 \ldots \bar{t}_{n_i} &\to& \bar{c}_i^\iota(\bar{t}_1, \ldots, \bar{t}_{n_i})
\;\;\; \mathrm{if\ } \bar{t}_1, \ldots, \bar{t}_{n_i} \in \bar{\iota} \in \bar{{\ensuremath{{\cal T}}}}_I \\
d_{i,j}^\iota (\bar{c}_i^\iota(\bar{t}_1,\ldots,\bar{t}_{n_i})) &\to& \bar{t}_j \\
d_{i,j}^\iota (c_i^\iota t_1 \ldots t_{n_i}) &\to& t_j \\
o_i^\iota (\bar{c}_i^\iota(\bar{t}_1, \ldots, \bar{t}_{n_i})) &\to& {\ensuremath{\mathbf{t}}}\\
o_i^\iota (\bar{c}_k^\iota(\bar{t}_1, \ldots, \bar{t}_{n_k})) &\to& {\ensuremath{\mathbf{f}}}\;\;\; \mathrm{if\ } i \ne k \\
o_i^\iota (c_i^\iota t_1 \ldots t_{n_i}) &\to& {\ensuremath{\mathbf{t}}}\\
o_i^\iota (c_k^\iota t_1 \ldots t_{n_k}) &\to& {\ensuremath{\mathbf{f}}}\;\;\; \mathrm{if\ } i \ne k \\
\mathrm{Choice}\, \tau &\to& \chi(\tau) \;\;\; \mathrm{if\ } \tau \in {\ensuremath{{\cal T}}},\, \tau \ne \emptyset \\
{\ensuremath{\mathrm{Is}\,\tau\,{\ensuremath{\mathrm{Type}}}\,}} &\to& {\ensuremath{\mathbf{t}}}\;\;\; \mathrm{for\ } \tau \in {\ensuremath{{\cal T}}}\end{aligned}$$
The above is a circular definition because the condition in the rule for $\mathrm{Eq}$ refers to the system $R$. Note, however, that this reference is positive. Formally, we may therefore define a progression of rewrite systems $R_\alpha$ consisting of the above rules, but each using as $R$ the system $R_{<\alpha} =
\bigcup_{\gamma<\alpha}R_\gamma$. We note that $R_\alpha \subseteq
R_\beta$ for $\alpha \le \beta$ and take $R$ to be the least fixpoint.
From now on $R$ refers to the rewriting system defined above, and the relations ${\ensuremath{\rightarrow}}$, ${\ensuremath{\Rightarrow}}$, ${\ensuremath{\stackrel{*}{\Rightarrow}}}$, etc. refer to ${\ensuremath{\rightarrow}}_R$, ${\ensuremath{\Rightarrow}}_R$, ${\ensuremath{\stackrel{*}{\Rightarrow}}}_R$, etc. We write $t_1
{\ensuremath{\rightarrow}}_\equiv^p t_2$ if $t_1 {\ensuremath{\rightarrow}}^p t_2$ or $t_1 \equiv t_2$.
The three simple lemmas below follow by an easy inspection of the definition of $R$ and of ${\ensuremath{\Rightarrow}}$.
If $c {\ensuremath{\Rightarrow}}t$ where $c \in \Sigma^+$, then $t \equiv c$ or $c
\in {\ensuremath{{\cal C}}}$ has arity $0$.
\[lem\_contr\_epsilon\] If $c \in \Sigma^+$, there exist $s_1,\ldots,s_n$ such that $c s_1
\ldots s_n$ is a left side of a rule in $R$, and $c t_1 \ldots t_n
{\ensuremath{\Rightarrow}}^\alpha t$, then there exist $t_1',\ldots,t_n'$ such that $t_i {\ensuremath{\Rightarrow}}^\alpha t_i'$ for $i=1,\ldots,n$ and $c t_1' \ldots
t_n' {\ensuremath{\rightarrow}}_\equiv^\epsilon t$.
\[lem\_orthogonal\] If $t {\ensuremath{\rightarrow}}^\epsilon t_1$ and $t {\ensuremath{\rightarrow}}^\epsilon t_2$ then $t_1
\equiv t_2$.
\[lem\_subst\_1\] The following conditions hold.
- If $t_1 {\ensuremath{\rightarrow}}_{R_\beta}^{\epsilon} t_1'$ then $t_1[x/t_2]
{\ensuremath{\rightarrow}}_{R_\beta}^{\epsilon} t_1'[x/t_2]$.
- If $t_1 {\ensuremath{\Rightarrow}}_{R_\beta}^{\alpha} t_1'$ then $t_1[x/t_2]
{\ensuremath{\Rightarrow}}_{R_\beta}^{\alpha} t_1'[x/t_2]$.
- If $t_1 \succ_{R_\beta}^\alpha c$ then $t_1[x/t_2]
\succ_{R_\beta}^\alpha c$.
Induction on triples $\langle \beta, \alpha, h(t_1) \rangle$ ordered lexicographically.
We first show that $t_1 {\ensuremath{\rightarrow}}_{R_\beta}^{\epsilon} t_1'$ implies $t_1[x/t_2] {\ensuremath{\rightarrow}}_{R_\beta}^{\epsilon} t_1'[x/t_2]$. The only non-obvious case is when $t_1 \equiv {\ensuremath{\mathrm{Eq}\,r_1\,r_2\,}}
{\ensuremath{\rightarrow}}_{R_\beta}^\epsilon {\ensuremath{\mathbf{t}}}\equiv t_1'$ by virtue of $r_1
{\ensuremath{\stackrel{*}{\Leftrightarrow}}}_{R_{<\beta}} r_2$. But then by the inductive hypothesis (for smaller $\beta$) we obtain $r_1[x/t_2] {\ensuremath{\stackrel{*}{\Leftrightarrow}}}_{R_{<\beta}}
r_2[x/t_2]$. This implies that $t_1[x/t_2] \equiv
{\ensuremath{\mathrm{Eq}\,r_1[x/t_2]\,r_2[x/t_2]\,}} {\ensuremath{\rightarrow}}_{R_\beta}^\epsilon {\ensuremath{\mathbf{t}}}\equiv
t_1'[x/t_2]$.
We show that $t_1 {\ensuremath{\Rightarrow}}_{R_\beta}^{\alpha} t_1'$ implies $t_1[x/t_2] {\ensuremath{\Rightarrow}}_{R_\beta}^{\alpha} t_1'[x/t_2]$. If $t_1 \equiv
t_1'$ then this is obvious. If $t_1 {\ensuremath{\rightarrow}}_{R_\beta}^\epsilon t_1'$ then this follows from the previous condition. If $t_1 \equiv c r_1
\ldots r_n {\ensuremath{\Rightarrow}}_{R_\beta}^\alpha t_1'$, $c \in \Sigma^+$, $c
r_1' \ldots r_n' {\ensuremath{\rightarrow}}_{R_\beta}^\epsilon t_1'$, $r_i
{\ensuremath{\Rightarrow}}_{R_\beta}^{\alpha} r_i'$, then $r_i[x/t_2]
{\ensuremath{\Rightarrow}}_{R_\beta}^\alpha r_i'[x/t_2]$ by the inductive hypothesis (because $h(r_i) < h(t_1)$). As in the previous paragraph, it also follows from the IH that $c r_1'[x/t_2] \ldots r_n'[x/t_2]
{\ensuremath{\rightarrow}}_{R_\beta}^\epsilon t_1'[x/t_2]$. Therefore $t_1[x/t_2] \equiv
c r_1[x/t_2] \ldots r_n[x/t_2] {\ensuremath{\Rightarrow}}_{R_\beta}^\alpha
t_1'[x/t_2]$. All other cases follow easily from the inductive hypothesis, and we omit them.
Now we prove that $t_1 \succ_{R_\beta}^\alpha c$ implies $t_1[x/t_2]
\succ_{R_\beta}^\alpha c$. If $t_1 \equiv c$ then this is obvious. Otherwise $c \in \tau_2^{\tau_1} \in {\ensuremath{{\cal T}}}$ and for all $c_1
\in \tau_1$ there exists $t_1'$ such that $t_1 c_1
{\ensuremath{\stackrel{*\;<\alpha}{\Rightarrow_{R_\beta}\;\,}}} t_1' \succ_{R_\beta}^{<\alpha}
{\ensuremath{{\cal F}}}(c)(c_1)$. But then by the IH we have $t_1[x/t_2] c_1
{\ensuremath{\stackrel{*\;<\alpha}{\Rightarrow_{R_\beta}\;\,}}} t_1'[x/t_2] \succ_{R_\beta}^{<\alpha}
{\ensuremath{{\cal F}}}(c)(c_1)$. Therefore $t_1[x/t_2] \succ_{R_\beta}^{\alpha} c$.
\[lem\_subst\_2\] If $t_2 {\ensuremath{\Rightarrow}}^{\alpha} t_2'$ then $t_1[x/t_2] {\ensuremath{\Rightarrow}}^{\alpha}
t_1[x/t_2']$.
Induction on the structure of $t_1$. If $t_1 \equiv x$ or $x \notin
FV(t_1)$ then this is obvious. If ${\ensuremath{{t_1}^{|\epsilon}}} \in \{{\ensuremath{\mathrm{A}}}\tau, {\ensuremath{\mathrm{S}}}\tau\}$ then by the inductive hypothesis ${\ensuremath{{t_1}_{|c}}}[x/t_2] {\ensuremath{\Rightarrow}}^{\alpha} {\ensuremath{{t_1}_{|c}}}[x/t_2']$ for $c
\in \tau$. Therefore $t_1[x/t_2] {\ensuremath{\Rightarrow}}^\alpha t_1[x/t_2']$. Other cases follow directly from the inductive hypothesis in a similar fashion.
\[lem\_subst\] If $t_1 {\ensuremath{\Rightarrow}}^\alpha t_1'$ and $t_2 {\ensuremath{\Rightarrow}}^\alpha t_2'$ then $t_1[x/t_2] {\ensuremath{\Rightarrow}}^{\alpha} t_1'[x/t_2']$.
Induction on $\alpha$ and the structure of $t_1$.
If $t_1 \equiv t_1'$ then the claim follows from Lemma \[lem\_subst\_2\]. If $t_1 {\ensuremath{\rightarrow}}^\epsilon t_1'$ then we consider possible forms of $t_1$. Suppose $t_1 \equiv {\ensuremath{\mathrm{Eq}\,r_1\,r_2\,}}
{\ensuremath{\rightarrow}}^\epsilon {\ensuremath{\mathbf{t}}}\equiv t_1'$ by virtue of $r_1 {\ensuremath{\stackrel{*}{\Leftrightarrow}}}r_2$. By Lemma \[lem\_subst\_1\] we have $r_1[x/t_2] {\ensuremath{\stackrel{*}{\Leftrightarrow}}}r_2[x/t_2]$. Hence $t_1[x/t_2] \equiv {\ensuremath{\mathrm{Eq}\,r_1[x/t_2]\,r_2[x/t_2]\,}}
{\ensuremath{\rightarrow}}^\epsilon {\ensuremath{\mathbf{t}}}\equiv t_1'[x/t_2']$. Suppose $t_1 \equiv
\forall \tau t {\ensuremath{\rightarrow}}^\epsilon t_1'$ where $\tau \in {\ensuremath{{\cal T}}}$, ${\ensuremath{{t_1'}^{|\epsilon}}} \equiv {\ensuremath{\mathrm{A}}}\tau$, and for all $c \in \tau$ we have ${\ensuremath{{t_1'}_{|c}}} \equiv t c$. Then $t[x/t_2] {\ensuremath{\Rightarrow}}^\alpha
t[x/t_2']$ by the inductive hypothesis, and $\forall \tau
(t[x/t_2']) {\ensuremath{\rightarrow}}^\epsilon t_1'[x/t_2']$ by Lemma \[lem\_subst\_1\]. Hence by condition (5) in the definition of ${\ensuremath{\Rightarrow}}^\alpha$ we obtain $t_1[x/t_2] \equiv \forall \tau
(t[x/t_2]) {\ensuremath{\Rightarrow}}^\alpha t_1'[x/t_2']$. Other cases are established in a similar manner.
If $t_1 \equiv (\lambda y \,.\, r_1) r_2 {\ensuremath{\Rightarrow}}^\alpha r_1'[y/r_2']
\equiv t_1'$ where $x \not\equiv y$, $r_1 {\ensuremath{\Rightarrow}}^\alpha r_1'$ and $r_2 {\ensuremath{\Rightarrow}}^\alpha r_2'$, then by the inductive hypothesis $r_1[x/t_2] {\ensuremath{\Rightarrow}}^\alpha r_1'[x/t_2']$ and $r_2[x/t_2]
{\ensuremath{\Rightarrow}}^\alpha r_2'[x/t_2']$. Recall that by our implicit assumption that in $t_1[x/t_2]$ no free variables of $t_2$ become bound, we have $y \notin FV(t_2)$, and hence $y \notin
FV(t_2')$. Thus $t_1[x/t_2] \equiv (\lambda y \,.\, r_1[x/t_2])
r_2[x/t_2] {\ensuremath{\Rightarrow}}^\alpha r_1'[x/t_2'][y/(r_2'[x/t_2'])] \equiv
(r_1'[y/r_2'])[x/t_2'] \equiv t_1'[x/t_2']$.
If $t_1 \equiv c r_1 \ldots r_n {\ensuremath{\Rightarrow}}^\alpha t_1'$ where $c \in
\Sigma^+$, $r_i {\ensuremath{\Rightarrow}}^\alpha r_i'$ for $i=1,\ldots,n$, and $c
r_1' \ldots r_n' {\ensuremath{\rightarrow}}^\epsilon t_1'$, then $c r_1'[x/t_2'] \ldots
r_n'[x/t_2'] {\ensuremath{\rightarrow}}^\epsilon t_1'[x/t_2']$ by Lemma \[lem\_subst\_1\] and $r_i[x/t_2] {\ensuremath{\Rightarrow}}^\alpha
r_i'[x/t_2']$ by the inductive hypothesis. We thus conclude $t_1[x/t_2] \equiv c r_1[x/t_2] \ldots r_n[x/t_2] {\ensuremath{\Rightarrow}}^\alpha
t_1'[x/t_2']$.
If $t_1 \equiv {\ensuremath{\mathrm{Is}\,t\,\tau\,}} {\ensuremath{\Rightarrow}}^\alpha {\ensuremath{\mathbf{t}}}\equiv t_1'$ where $t \succ^{<\alpha} c$ for some $c \in \tau$, then $t[x/t_2]
\succ^{<\alpha} c$ by Lemma \[lem\_subst\_1\]. Therefore $t_1[x/t_2]
\equiv {\ensuremath{\mathrm{Is}\,t[x/t_2]\,\tau\,}} {\ensuremath{\Rightarrow}}^\alpha {\ensuremath{\mathbf{t}}}\equiv
t_1'[x/t_2']$.
If ${\ensuremath{{t_1}^{|\epsilon}}} \equiv {\ensuremath{{t_1'}^{|\epsilon}}} \equiv
{\ensuremath{\mathrm{A}}}\tau$ and for all $c \in \tau$ there exists $t_c$ such that ${\ensuremath{{t_1}_{|c}}} {\ensuremath{\Rightarrow}}^\alpha t_c {\ensuremath{\stackrel{\;\;*\;<\alpha}{\Rightarrow\;\;}}}
{\ensuremath{{t_1'}_{|c}}}$, then ${\ensuremath{{t_1}_{|c}}}[x/t_2] {\ensuremath{\Rightarrow}}^\alpha
t_c[x/t_2']$ by the inductive hypothesis. By Lemma \[lem\_subst\_1\] we obtain $t_c[x/t_2'] {\ensuremath{\stackrel{\;\;*\;<\alpha}{\Rightarrow\;\;}}}
{\ensuremath{{t_1'}_{|c}}}[x/t_2']$. This implies that $t_1[x/t_2]
{\ensuremath{\Rightarrow}}^\alpha t_1'[x/t_2']$.
Other cases follow by analogous proofs.
\[lem\_contr\_stable\] If $c \in \Sigma^+$, $n \in {\ensuremath{\mathbb{N}}}$, $c t_1 \ldots t_n {\ensuremath{\rightarrow}}^\epsilon
t$ and $t_i {\ensuremath{\Rightarrow}}^\alpha t_i'$ for $i=1,\ldots,n$, then there exists $t'$ such that $c t_1' \ldots t_n' {\ensuremath{\rightarrow}}^\epsilon t'$ and $t
{\ensuremath{\Rightarrow}}^\alpha t'$.
If $c t_1 t_2 \equiv {\ensuremath{\mathrm{Eq}\,t_1\,t_2\,}} {\ensuremath{\rightarrow}}^\epsilon {\ensuremath{\mathbf{t}}}$ then $t_1
{\ensuremath{\stackrel{*}{\Leftrightarrow}}}t_2$. Since ${\ensuremath{\Rightarrow}}^\alpha \,\subseteq\, {\ensuremath{\Rightarrow}}$, we have $t_1' {\ensuremath{\stackrel{*}{\Leftrightarrow}}}t_2'$. Thus ${\ensuremath{\mathrm{Eq}\,t_1'\,t_2'\,}} {\ensuremath{\rightarrow}}^\epsilon
{\ensuremath{\mathbf{t}}}$.
Suppose $c t_1 t_2 \equiv \forall \tau t_2 {\ensuremath{\rightarrow}}^\epsilon t$ where $\tau \in {\ensuremath{{\cal T}}}$, ${\ensuremath{{t}^{|\epsilon}}} \equiv {\ensuremath{\mathrm{A}}}\tau$, ${\ensuremath{{t}_{|c}}} \equiv t_2 c$ for $c \in \tau$. We have $\forall \tau
t_2' {\ensuremath{\rightarrow}}^\epsilon t'$ where ${\ensuremath{{t'}^{|\epsilon}}} \equiv {\ensuremath{\mathrm{A}}}\tau$ and ${\ensuremath{{t'}_{|c}}} \equiv t_2' c$ for $c \in \tau$. Since $t_2 c
{\ensuremath{\Rightarrow}}^\alpha t_2' c$, by condition (8) in the definition of ${\ensuremath{\Rightarrow}}^\alpha$ we conclude that $t {\ensuremath{\Rightarrow}}^\alpha t'$.
If $c t_1 \equiv d_{i,j}^\iota (c_i^\iota r_1 \ldots r_{n_i})
{\ensuremath{\rightarrow}}^\epsilon r_j \equiv t$ then $c_i^\iota r_1 \ldots r_{n_i}
{\ensuremath{\Rightarrow}}^\alpha t_1'$. By Lemma \[lem\_contr\_epsilon\] there exist $r_1',\ldots,r_{n_i}'$ such that $r_k {\ensuremath{\Rightarrow}}^\alpha r_k'$ for $k=1,\ldots,n_i$ and $c_i^\iota r_1' \ldots r_{n_i}'
{\ensuremath{\rightarrow}}_\equiv^\epsilon t_1'$. If $c_i^\iota r_1' \ldots r_{n_i}'
\equiv t_1'$ then $c t_1' \equiv d_{i,j}^\iota (c_i^\iota r_1'
\ldots r_{n_i}') {\ensuremath{\rightarrow}}^\epsilon r_j'$ and we are done. Otherwise $r_1',\ldots,r_{n_i}'\in\bar{\iota}$ and $t_1' \equiv
\bar{c}_i^\iota(r_1',\ldots,r_{n_i}')$. Thus $c t_1' \equiv d_{i,j}
(\bar{c}_i^\iota(r_1',\ldots,r_{n_i}')) {\ensuremath{\rightarrow}}^\epsilon r_j'$.
Other cases are trivial or follow by a similar proof.
We say that two binary relations on terms $\to_1$ and $\to_2$ *commute* if $t_1 \to_1 t_1'$ and $t_2 \to_2 t_2'$ imply $t_1'
\to_2^\equiv t_3$ and $t_2' \to_1^\equiv t_3$ for some term $t_3$, where $\to_i^\equiv$ is the reflexive closure of $\to_i$.
\[lem\_hindley\_rosen\] If ${\ensuremath{\Rightarrow}}^{\alpha'}$ and ${\ensuremath{\Rightarrow}}^{\beta'}$ commute for all $\alpha' \circ_1 \alpha$ and $\beta' \circ_2 \beta$, then ${\ensuremath{\stackrel{\;\;*\;\circ_1\alpha}{\Rightarrow\;\;}}}$ and ${\ensuremath{\stackrel{\;\;*\;\circ_2\beta}{\Rightarrow\;\;}}}$, ${\ensuremath{\Rightarrow}}^{\circ_1\alpha}$ and ${\ensuremath{\stackrel{\;\;*\;\circ_2\beta}{\Rightarrow\;\;}}}$, as well as ${\ensuremath{\stackrel{\;\;*\;\circ_1\alpha}{\Rightarrow\;\;}}}$ and ${\ensuremath{\Rightarrow}}^{\circ_2\beta}$, commute. Here $\circ_1,\circ_2 \in \{<, \le\}$ and ${\ensuremath{\Rightarrow}}^{\le\gamma}\,=\,{\ensuremath{\Rightarrow}}^\gamma$ for $\gamma \in
\{\alpha,\beta\}$.
The proof is a simple tiling argument similar to the proof of the Hindley-Rosen lemma, see e.g. [@Barendregt1984 Chapter 3].
\[lem\_commute\] For all ordinals $\alpha$, $\beta$ the following conditions hold:
- ${\ensuremath{\Rightarrow}}^\alpha$ and ${\ensuremath{\Rightarrow}}^\beta$ commute,
- if $t \succ^\alpha c$ and $t {\ensuremath{\Rightarrow}}^\beta t'$ then $t' \succ^\alpha c$,
- if $t \succ^\alpha c_1$, $t \succ^\beta c_2$ and $c_1, c_2 \in \tau \in {\ensuremath{{\cal T}}}$ then $c_1 \equiv c_2$.
The proof is by induction on triples $\langle \alpha, \beta, h(t)
\rangle$ ordered lexicographically, where in condition (${\romannumeral 1}$) the term $t$ is such that $t {\ensuremath{\Rightarrow}}^\alpha t_1$ and $t
{\ensuremath{\Rightarrow}}^\beta t_2$ for some $t_1, t_2$. Together with condition (${\romannumeral 2}$) we also prove its dual, i.e. the condition with $\alpha$ and $\beta$ exchanged. We only give a proof for the original condition, but it is easy to see that the dual condition follows by exactly the same proof but with $\alpha$ and $\beta$ exchanged.
We first show condition (${\romannumeral 1}$). Assume $t {\ensuremath{\Rightarrow}}^\alpha
t_1$ and $t {\ensuremath{\Rightarrow}}^\beta t_2$. We need to show that there exists $t'$ such that $t_1 {\ensuremath{\Rightarrow}}^\beta t'$ and $t_2
{\ensuremath{\Rightarrow}}^\alpha t'$. It is clear that it suffices to consider only the situations when $t {\ensuremath{\Rightarrow}}^\alpha t_1$ follows by condition ($m$) and $t {\ensuremath{\Rightarrow}}^\beta t_2$ follows by condition ($n$) in the definition of ${\ensuremath{\Rightarrow}}$ for $m \le n$, provided that we never use the inductive hypothesis with $\beta$ increased, which is easily verified to be the case. Indeed, then we may use exactly the same proofs, but with $\alpha$ and $\beta$ exchanged, to handle the cases when $m > n$.
If $t \equiv t_1$ or $t \equiv t_2$ then the claim is obvious. Suppose $t {\ensuremath{\Rightarrow}}^\alpha t_1$ follows by condition (1) in the definition of ${\ensuremath{\Rightarrow}}^\alpha$. Then $t {\ensuremath{\rightarrow}}^\epsilon
t_1$. By Lemma \[lem\_orthogonal\] it is impossible that $t
{\ensuremath{\Rightarrow}}^\beta t_2$ follows by condition (1) in the definition of ${\ensuremath{\Rightarrow}}^\beta$, unless $t_1 \equiv t_2$. Suppose that $t
{\ensuremath{\Rightarrow}}^\beta t_2$ follows by condition (5). Then $t \equiv c r_1
\ldots r_n {\ensuremath{\rightarrow}}^\epsilon t_1$, $r_i {\ensuremath{\Rightarrow}}^\beta r_i'$ and $c
r_1' \ldots r_n' {\ensuremath{\rightarrow}}^\epsilon t_2$. By Lemma \[lem\_contr\_stable\] there exists $t'$ such that $t_1
{\ensuremath{\Rightarrow}}^\beta t'$ and $c r_1' \ldots r_n' {\ensuremath{\rightarrow}}^\epsilon t'$. But by Lemma \[lem\_orthogonal\] we have $t' \equiv t_2$. The only remaining possibility, when $t {\ensuremath{\Rightarrow}}^\alpha t_1$ follows by condtion (1), is that $t {\ensuremath{\Rightarrow}}^\beta t_2$ follows by condition (2). But then the claim follows from Lemma \[lem\_contr\_stable\].
Suppose $t \equiv r_1 r_2 {\ensuremath{\Rightarrow}}^\alpha r_1' r_2' \equiv t_1$ where $r_1 {\ensuremath{\Rightarrow}}^\alpha r_1'$ and $r_2 {\ensuremath{\Rightarrow}}^\alpha r_2'$. If $t {\ensuremath{\Rightarrow}}^\beta t_2$ follows by condition (2) then $t_2 \equiv
r_1'' r_2''$ where $r_1 {\ensuremath{\Rightarrow}}^\beta r_1''$ and $r_2
{\ensuremath{\Rightarrow}}^\beta r_2''$. By the inductive hypothesis (note that $h(r_1), h(r_2) < h(t)$) there exist $q_1$, $q_2$ such that $r_1'
{\ensuremath{\Rightarrow}}^\beta q_1$, $r_1'' {\ensuremath{\Rightarrow}}^\alpha q_1$, $r_2'
{\ensuremath{\Rightarrow}}^\beta q_2$ and $r_2'' {\ensuremath{\Rightarrow}}^\alpha q_2$. Thus $t_1
\equiv r_1' r_2' {\ensuremath{\Rightarrow}}^\beta q_1 q_2$ and $t_2 \equiv r_1'' r_2''
{\ensuremath{\Rightarrow}}^\alpha q_1 q_2$.
It is not possible that $t {\ensuremath{\Rightarrow}}^\beta t_2$ follows by condition (3). If it follows by condition (4) then $r_1 \equiv
\lambda x \,.\, s_1$, $r_1' \equiv \lambda x \,.\, s_1'$, $s_1
{\ensuremath{\Rightarrow}}^\alpha s_1'$, and $t_2 \equiv s_1''[x/r_2'']$ where $s_1
{\ensuremath{\Rightarrow}}^\beta s_1''$, $r_2 {\ensuremath{\Rightarrow}}^\beta r_2''$. By the inductive hypothesis there exist $q_1$ and $q_2$ such that $s_1'
{\ensuremath{\Rightarrow}}^\beta q_1$, $s_1'' {\ensuremath{\Rightarrow}}^\alpha q_1$, $r_2'
{\ensuremath{\Rightarrow}}^\beta q_2$ and $r_2'' {\ensuremath{\Rightarrow}}^\alpha q_2$. By condition (4) in the definition of ${\ensuremath{\Rightarrow}}^\beta$ we have $t_1
\equiv (\lambda x \,.\, s_1') r_2' {\ensuremath{\Rightarrow}}^\beta q_1[x/q_2]$. By Lemma \[lem\_subst\] we obtain $t_2 \equiv s_1''[x/r_2'']
{\ensuremath{\Rightarrow}}^\alpha q_1[x/q_2]$.
If $t {\ensuremath{\Rightarrow}}^\beta t_2$ follows by condition (5) then $r_1 \equiv
c s_1 \ldots s_n$, $s_i {\ensuremath{\Rightarrow}}^\beta s_i''$, $r_2 {\ensuremath{\Rightarrow}}^\beta
r_2''$ and $c s_1'' \ldots s_n'' r_2'' {\ensuremath{\rightarrow}}^\epsilon t_2$. By inspecting the definition of $R$ we see that in this case $c q_1
\ldots q_m {\ensuremath{\rightarrow}}^\epsilon r_1'$ is not possible for any $q_1,\ldots,q_m$ and any $m \le n$. By inspecting the definition of ${\ensuremath{\Rightarrow}}^\alpha$ we thus see that $r_1 {\ensuremath{\Rightarrow}}^\alpha r_1'$ is only possible when $r_1' \equiv c s_1' \ldots s_n'$ and $s_i
{\ensuremath{\Rightarrow}}^\alpha s_i'$. By the inductive hypothesis there exist $q_1,\ldots,q_{n+1}$ such that $s_i'' {\ensuremath{\Rightarrow}}^\alpha q_i$, $s_i'
{\ensuremath{\Rightarrow}}^\beta q_i$ for $i=1,\ldots,n$, and $r_2'' {\ensuremath{\Rightarrow}}^\alpha
q_{n+1}$, $r_2' {\ensuremath{\Rightarrow}}^\beta q_{n+1}$. Therefore $c s_1' \ldots
s_n' r_2' {\ensuremath{\Rightarrow}}^\beta c q_1 \ldots q_{n+1}$ and $c s_1'' \ldots
s_n'' r_2'' {\ensuremath{\Rightarrow}}^\alpha c q_1 \ldots q_{n+1}$. By Lemma \[lem\_contr\_stable\] there exists $t'$ such that $t_2
{\ensuremath{\Rightarrow}}^\alpha t'$ and $c q_1 \ldots q_{n+1} {\ensuremath{\rightarrow}}^\epsilon
t'$. Hence by condition (5) also $t_1 \equiv c s_1' \ldots s_n' r_2'
{\ensuremath{\Rightarrow}}^\beta t'$. See Fig. \[fig\_02\].
If $t {\ensuremath{\Rightarrow}}^\beta t_2$ follows by condition (6) then $r_1 \equiv
r_1' \equiv c \in \tau_2^{\tau_1}$, $r_2 \succ^{<\beta} c_1 \in
\tau_1$ and $t_2 \equiv {\ensuremath{{\cal F}}}(c)(c_1)$. By part (${\romannumeral 2}$) of the inductive hypothesis we conclude that $r_2' \succ^{<\beta}
c_1$. Therefore $t_1 \equiv c r_2' {\ensuremath{\Rightarrow}}^\beta {\ensuremath{{\cal F}}}(c)(c_1) \equiv
t_2$.
If $t {\ensuremath{\Rightarrow}}^\beta t_2$ follows by condition (7) then $r_1 \equiv
\mathrm{Is}\, s$, $r_2 \equiv \tau$, $r_1' \equiv \mathrm{Is}\, s'$, $s {\ensuremath{\Rightarrow}}^\alpha s'$, $t_2 \equiv {\ensuremath{\mathbf{t}}}$ and $s \succ^{<\beta} c$ for some $c \in \tau \in {\ensuremath{{\cal T}}}$. By part (${\romannumeral 2}$) of the inductive hypothesis we have $s' \succ^{<\beta} c$. Therefore $t_1
\equiv {\ensuremath{\mathrm{Is}\,s'\,r_2\,}} {\ensuremath{\Rightarrow}}^\beta {\ensuremath{\mathbf{t}}}\equiv t_2$. It is easy to see that it is impossible that $t {\ensuremath{\Rightarrow}}^\beta t_2$ follows by condition (8).
Now suppose that $t {\ensuremath{\Rightarrow}}^\alpha t_1$ follows by condition (3). Then $t \equiv \lambda x \,.\, r$ and $t_1 \equiv
\lambda x \,.\, r_1$ where $r {\ensuremath{\Rightarrow}}^\alpha r_1$. It is easy too see that the only possibility is that $t {\ensuremath{\Rightarrow}}^\beta t_2$ follows by condition (3) as well. Then $t_2 \equiv \lambda x \,.\, r_2$ where $r {\ensuremath{\Rightarrow}}^\beta r_2$. By the inductive hypothesis there exists $q$ such that $r_1 {\ensuremath{\Rightarrow}}^\beta q$ and $r_2 {\ensuremath{\Rightarrow}}^\alpha
q$. Therefore $t_1 \equiv \lambda x \,.\, r_1 {\ensuremath{\Rightarrow}}^\beta \lambda
x \,.\, q$ and $t_2 \equiv \lambda x \,.\, r_2 {\ensuremath{\Rightarrow}}^\alpha
\lambda x \,.\, q$.
Suppose that $t {\ensuremath{\Rightarrow}}^\alpha t_1$ follows by condition (4). Then $t \equiv (\lambda x \,.\, r_1) r_2$ and $t_1 \equiv r_1'[x/r_2']$ where $r_1 {\ensuremath{\Rightarrow}}^\alpha r_1'$ and $r_2 {\ensuremath{\Rightarrow}}^\alpha r_2'$. It is easy to see that the only possibility is that $t {\ensuremath{\Rightarrow}}^\beta t_2$ follows by condition (4) as well. Then $t_2 \equiv r_1''[x/r_2'']$ where $r_1 {\ensuremath{\Rightarrow}}^\beta r_1''$ and $r_2 {\ensuremath{\Rightarrow}}^\beta r_2''$. By the inductive hypothesis there exist $q_1$ and $q_2$ such that $r_1'
{\ensuremath{\Rightarrow}}^\beta q_1$, $r_1'' {\ensuremath{\Rightarrow}}^\alpha q_1$, $r_2'
{\ensuremath{\Rightarrow}}^\beta q_2$ and $r_2'' {\ensuremath{\Rightarrow}}^\alpha q_2$. Therefore by Lemma \[lem\_subst\] we obtain $t_1 \equiv r_1'[x/r_2']
{\ensuremath{\Rightarrow}}^\beta q_1[x/q_2]$ and $t_2 \equiv r_1''[x/r_2'']
{\ensuremath{\Rightarrow}}^\alpha q_1[x/q_2]$.
Suppose that $t {\ensuremath{\Rightarrow}}^\alpha t_1$ follows by condition (5). Then $t \equiv c r_1 \ldots r_n$, $r_i {\ensuremath{\Rightarrow}}^\alpha r_i'$ and $c r_1'
\ldots r_n' {\ensuremath{\rightarrow}}^\epsilon t_1$. If $t {\ensuremath{\Rightarrow}}^\beta t_2$ also follows by condition (5), then there exist $r_1'',\ldots,r_n''$ such that $r_i {\ensuremath{\Rightarrow}}^\beta r_i''$ and $c r_1'' \ldots r_n''
{\ensuremath{\rightarrow}}^\epsilon t_2$. By the inductive hypothesis there exist $q_1,\ldots,q_n$ such that $r_i' {\ensuremath{\Rightarrow}}^\beta q_i$ and $r_i''
{\ensuremath{\Rightarrow}}^\alpha q_i$. Therefore $c r_1' \ldots r_n' {\ensuremath{\Rightarrow}}^\beta c
q_1 \ldots q_n$ and $c r_1'' \ldots r_n'' {\ensuremath{\Rightarrow}}^\alpha c q_1
\ldots q_n$. By Lemma \[lem\_contr\_stable\] there exist $t_1'$ and $t_2'$ such that $t_1 {\ensuremath{\Rightarrow}}^\beta t_1'$, $t_2 {\ensuremath{\Rightarrow}}^\alpha
t_2'$, $c q_1 \ldots q_n {\ensuremath{\rightarrow}}^\epsilon t_1'$ and $c q_1 \ldots q_n
{\ensuremath{\rightarrow}}^\epsilon t_2'$. But by Lemma \[lem\_orthogonal\] we have $t_1' \equiv t_2'$. See Fig. \[fig\_03\].
It is easy to verify that it is not possible that $t {\ensuremath{\Rightarrow}}^\beta
t_2$ follows by condition (6). If $t {\ensuremath{\Rightarrow}}^\beta t_2$ follows by condition (7), then we must have $c \equiv \mathrm{Is}$, $t_1 \equiv
t_2 \equiv {\ensuremath{\mathbf{t}}}$. It is not possible that $t {\ensuremath{\Rightarrow}}^\beta t_2$ follows by condition (8).
Suppose $t {\ensuremath{\Rightarrow}}^\alpha t_1$ follows by condition (6). Then $t
\equiv c r {\ensuremath{\Rightarrow}}^\alpha {\ensuremath{{\cal F}}}(c)(c_1) \equiv t_2$ where $c \in
\tau_2^{\tau_1}$ and $r \succ^{<\alpha} c_1 \in \tau_1$. It is easily verified that the only possibility is when $t {\ensuremath{\Rightarrow}}^\beta
t_2$ follows by condition (6) as well. Then $t_2 \equiv
{\ensuremath{{\cal F}}}(c)(c_1')$ and $r \succ^{<\beta} c_1'$ for some $c_1' \in
\tau_1$. By part (${\romannumeral 3}$) of the inductive hypothesis we obtain $c_1' \equiv c_1$. Hence $t_1 \equiv t_2$.
Suppose $t {\ensuremath{\Rightarrow}}^\alpha t_1$ follows by condition (7). Then $t
\equiv {\ensuremath{\mathrm{Is}\,r\,\tau\,}}$, $t_1 \equiv {\ensuremath{\mathbf{t}}}$, and $t {\ensuremath{\Rightarrow}}^\beta t_2$ may only follow by condition (7). But then we have $t_2 \equiv {\ensuremath{\mathbf{t}}}\equiv t_1$.
Finally, suppose $t {\ensuremath{\Rightarrow}}^\alpha t_1$ follows by condition (8) and so does $t {\ensuremath{\Rightarrow}}^\beta t_2$. Then e.g. ${\ensuremath{{t}^{|\epsilon}}}
\equiv {\ensuremath{{t_1}^{|\epsilon}}} \equiv {\ensuremath{{t_2}^{|\epsilon}}} \equiv
{\ensuremath{\mathrm{A}}}\tau$, and for all $c \in \tau$ there exist $t_c$ and $t_c'$ such that ${\ensuremath{{t}_{|c}}} {\ensuremath{\Rightarrow}}^\alpha t_c {\ensuremath{\stackrel{\;\;*\;<\alpha}{\Rightarrow\;\;}}}
{\ensuremath{{t_1}_{|c}}}$ and ${\ensuremath{{t}_{|c}}} {\ensuremath{\Rightarrow}}^\beta t_c'
{\ensuremath{\stackrel{\;\;*\;<\beta}{\Rightarrow\;\;}}} {\ensuremath{{t_2}_{|c}}}$. By the inductive hypothesis there exists $r$ such that $t_c {\ensuremath{\Rightarrow}}^\beta r$ and $t_c'
{\ensuremath{\Rightarrow}}^\alpha r$. By the inductive hypothesis and Lemma \[lem\_hindley\_rosen\] there exist $q_1$ and $q_2$ such that ${\ensuremath{{t_1}_{|c}}} {\ensuremath{\Rightarrow}}^\beta q_1$, $r {\ensuremath{\stackrel{\;\;*\;<\alpha}{\Rightarrow\;\;}}} q_1$, ${\ensuremath{{t_2}_{|c}}} {\ensuremath{\Rightarrow}}^\alpha q_2$ and $r {\ensuremath{\stackrel{\;\;*\;<\beta}{\Rightarrow\;\;}}}
q_2$. Again, by the inductive hypothesis and Lemma \[lem\_hindley\_rosen\] there exists $q_c$ such that $q_1
{\ensuremath{\stackrel{\;\;*\;<\beta}{\Rightarrow\;\;}}} q_c$ and $q_2 {\ensuremath{\stackrel{\;\;*\;<\alpha}{\Rightarrow\;\;}}} q_c$. Hence for all $c \in \tau$ there exists $q_1$ such that ${\ensuremath{{t_1}_{|c}}}
{\ensuremath{\Rightarrow}}^\beta q_1 {\ensuremath{\stackrel{\;\;*\;<\beta}{\Rightarrow\;\;}}} q_c$, and for all $c \in
\tau$ there exists $q_2$ such that ${\ensuremath{{t_2}_{|c}}} {\ensuremath{\Rightarrow}}^\alpha q_2
{\ensuremath{\stackrel{\;\;*\;<\alpha}{\Rightarrow\;\;}}} q_c$. Let $q$ be such that ${\ensuremath{{q}^{|\epsilon}}}
\equiv {\ensuremath{\mathrm{A}}}\tau$ and ${\ensuremath{{q}_{|c}}} \equiv q_c$ for $c \in \tau$. By the above considerations we have $t_1 {\ensuremath{\Rightarrow}}^\beta q$ and $t_2
{\ensuremath{\Rightarrow}}^\alpha q$. See Fig. \[fig\_04\].
Now we show condition (${\romannumeral 2}$). Thus suppose $t \succ^\alpha c$ and $t {\ensuremath{\Rightarrow}}^\beta t'$. If $t \equiv c$ then $t' \equiv c$ and thus $t' \succ^\alpha c$. Otherwise $c \in \tau_2^{\tau_1}$ and for all $c_1 \in \tau_1$ there exists $r$ such that $t c_1
{\ensuremath{\stackrel{\;\;*\;<\alpha}{\Rightarrow\;\;}}} r \succ^{<\alpha} {\ensuremath{{\cal F}}}(c)(c_1)$. Then $t c_1
{\ensuremath{\Rightarrow}}^\beta t' c_1$ and we conclude by part (${\romannumeral 1}$) of the inductive hypothesis and Lemma \[lem\_hindley\_rosen\] that there exists $r'$ such that $r {\ensuremath{\Rightarrow}}^\beta r'$ and $t' c_1
{\ensuremath{\stackrel{\;\;*\;<\alpha}{\Rightarrow\;\;}}} r'$. By part (${\romannumeral 2}$) of the inductive hypothesis we obtain $r' \succ^{<\alpha} {\ensuremath{{\cal F}}}(c)(c_1)$. Thus for every $c_1 \in \tau_1$ there exists $r'$ such that $t' c_1
{\ensuremath{\stackrel{\;\;*\;<\alpha}{\Rightarrow\;\;}}} r' \succ^{<\alpha} {\ensuremath{{\cal F}}}(c)(c_1)$. Hence $t'
\succ^\alpha {\ensuremath{{\cal F}}}(c)(c_1)$. See Fig. \[fig\_05\].
It remains to show condition (${\romannumeral 3}$). Thus suppose $t
\succ^\alpha c_1$ and $t \succ^\beta c_2$ for $c_1,c_2 \in \tau \in
{\ensuremath{{\cal T}}}$. If $\tau \subseteq {\ensuremath{\mathrm{Bool}}}$, $\tau \subseteq \delta$ or $\tau
\subseteq \bar{\iota} \in \bar{{\ensuremath{{\cal T}}}}_I$ then $t \equiv c_1 \equiv
c_2$ because in this case $t \succ^\alpha c_1$ and $t \succ^\beta
c_2$ may only be obtained by condition (a) in the definition of $\succ$. Otherwise $\tau \subseteq \tau_2^{\tau_1}$ and e.g. $t
\succ^\alpha c_1$ is obtained by condition (b), hence $\alpha > 0$.
If $c_1 \not\equiv c_2$ then there exists $c \in \tau_1$ such that ${\ensuremath{{\cal F}}}(c_1)(c) \not\equiv {\ensuremath{{\cal F}}}(c_2)(c)$. There exists $t_1$ such that $t c {\ensuremath{\stackrel{\;\;*\;<\alpha}{\Rightarrow\;\;}}} t_1 \succ^{<\alpha} {\ensuremath{{\cal F}}}(c_1)(c)$. If $t
\equiv c_2$ then by inspecting the definitions we see that this is only possible when $c_2 c {\ensuremath{\Rightarrow}}^{\gamma} {\ensuremath{{\cal F}}}(c_2)(c')
\succ^{<\alpha} {\ensuremath{{\cal F}}}(c_1)(c)$ where $c \succ^{<\gamma} c' \in \tau_1$ and $\gamma < \alpha$. By condition (a) we have $c \succ^\beta
c$. Since $c, c' \in \tau_1$ and $\gamma < \alpha$ we conlcude by part (${\romannumeral 3}$) of the inductive hypothesis that $c \equiv
c'$. Thus ${\ensuremath{{\cal F}}}(c_2)(c) \succ^{<\alpha} {\ensuremath{{\cal F}}}(c_1)(c)$. Obviously ${\ensuremath{{\cal F}}}(c_2)(c) \succ^\beta {\ensuremath{{\cal F}}}(c_2)(c)$ and ${\ensuremath{{\cal F}}}(c_2)(c), {\ensuremath{{\cal F}}}(c_1)(c)
\in \tau_2$, so again by part (${\romannumeral 3}$) of the inductive hypothesis we obtain ${\ensuremath{{\cal F}}}(c_1)(c) \equiv
{\ensuremath{{\cal F}}}(c_2)(c)$. Contradiction.
Thus assume that $t \succ^\beta c_2$ also follows by condition (b) in the definition of $\succ$. Then there exists $t_2$ such that $t c
{\ensuremath{\stackrel{\;\;*\;<\beta}{\Rightarrow\;\;}}} t_2 \succ^{<\beta} {\ensuremath{{\cal F}}}(c_2)(c)$. By part (${\romannumeral 1}$) of the inductive hypothesis and Lemma \[lem\_hindley\_rosen\] there exists $r$ such that $t_2
{\ensuremath{\stackrel{\;\;*\;<\alpha}{\Rightarrow\;\;}}} r$ and $t_1 {\ensuremath{\stackrel{\;\;*\;<\beta}{\Rightarrow\;\;}}} r$. By part (${\romannumeral 2}$) of the inductive hypothesis we have $r
\succ^{<\alpha} {\ensuremath{{\cal F}}}(c_1)(c)$ and $r \succ^{<\beta} {\ensuremath{{\cal F}}}(c_2)(c)$. By part (${\romannumeral 3}$) of the inductive hypothesis we obtain ${\ensuremath{{\cal F}}}(c_1)(c) \equiv {\ensuremath{{\cal F}}}(c_2)(c)$. Contradiction. See Fig. \[fig\_06\].
\[cor\_cr\] The relation ${\ensuremath{\Rightarrow}}$ has the Church-Rosser property.
The *rank* of a type $\tau \in {\ensuremath{{\cal T}}}$, denoted ${\ensuremath{\mathrm{rank}}}(\tau)$, is the smallest $n \in {\ensuremath{\mathbb{N}}}$ such that $\tau \in {\ensuremath{{\cal T}}}_n$. The *canonical type* of a canonical constant $c \in \Sigma$, denoted $\tau(c)$, is defined as follows.
- If $c \in \delta$ then $\tau(c) = \delta$.
- If $c \in {\ensuremath{\mathrm{Bool}}}$ then $\tau(c) = {\ensuremath{\mathrm{Bool}}}$.
- If $c \in \bar{\iota} \in \bar{{\ensuremath{{\cal T}}}}_I$ then $\tau(c) = \iota$.
- Otherwise let $\tau_2^{\tau_1} \in {\ensuremath{{\cal T}}}$ be such that $c \in
\tau_2^{\tau_1}$ and ${\ensuremath{\mathrm{rank}}}(\tau_2) \le {\ensuremath{\mathrm{rank}}}(\tau_2')$ for every $\tau_2' \in {\ensuremath{{\cal T}}}$ such that $c \in {\tau_2'}^{\tau_1}$. Then $\tau(c) = \tau_2^{\tau_1}$. Note that there may be more than one $\tau_2$ satisfying the above condition. In this case we arbitrarily choose one of them, and it does not matter which.
The rank of a canonical constant $c$, denoted ${\ensuremath{\mathrm{rank}}}(c)$, is the rank of its canonical type.
\[lem\_smaller\_rank\] The following conditions hold.
- For all $\tau_1, \tau_2 \in {\ensuremath{{\cal T}}}$ we have ${\ensuremath{\mathrm{rank}}}(\tau_1),
{\ensuremath{\mathrm{rank}}}(\tau_2) < {\ensuremath{\mathrm{rank}}}(\tau_2^{\tau_1})$.
- If $c \in \tau$ then ${\ensuremath{\mathrm{rank}}}(c) \le {\ensuremath{\mathrm{rank}}}(\tau)$.
For the first condition, note that ${\ensuremath{\mathrm{rank}}}(\tau_2^{\tau_1}) > 0$ and if $\tau_2^{\tau_1} \in {\ensuremath{{\cal T}}}_{n+1}$ then $\tau_1,\tau_2 \in {\ensuremath{{\cal T}}}_n$.
If $\tau(c) \in \{\delta, {\ensuremath{\mathrm{Bool}}}\} \cup \bar{{\ensuremath{{\cal T}}}}_I$ then the second condition is obvious. Otherwise $\tau(c) = \tau_2^{\tau_1}$, and if $c \in \tau$ then $\tau \subseteq \tau_3^{\tau_1}$ with ${\ensuremath{\mathrm{rank}}}(\tau)
\ge {\ensuremath{\mathrm{rank}}}(\tau_3^{\tau_1})$ and ${\ensuremath{\mathrm{rank}}}(\tau_2) \le
{\ensuremath{\mathrm{rank}}}(\tau_3)$. Suppose ${\ensuremath{\mathrm{rank}}}(\tau_1) = n_1$, ${\ensuremath{\mathrm{rank}}}(\tau_2) =
n_2$, ${\ensuremath{\mathrm{rank}}}(\tau_3) = n_3$. Then $n_2 \le n_3$, ${\ensuremath{\mathrm{rank}}}(c) =
\max(n_1, n_2) + 1$ and ${\ensuremath{\mathrm{rank}}}(\tau) \ge \max(n_1, n_3) + 1$. Thus ${\ensuremath{\mathrm{rank}}}(c) \le {\ensuremath{\mathrm{rank}}}(\tau)$.
We write $t \succsim c$ if $c \in \tau_2^{\tau_1}$ and for every $c_1 \in \tau_1$ there exists $t_{c_1}$ such that for all $t_1$ with $t_1 \succ c_1$ we have $t t_1 {\ensuremath{\stackrel{*}{\Rightarrow}}}t_{c_1} \succ
{\ensuremath{{\cal F}}}(c)(c_1)$. If for some $c_1 \in \tau_1$ there is more than one term $t_{c_1}$ satisfying the above condition, then we fix one arbitrarily, but globally, i.e. given $t$ and $c$ such that $t
\succsim c$ we assume that $t_{c_1}$ is uniquely determined for each $c_1 \in \tau_1$, and it depends only on $t$, $c$ and $c_1$. Note that if $t \succsim c$ then $t \succ c$.
Let $t \succ c$. The *mutual rank* of $t$ and $c$, denoted ${\ensuremath{\mathrm{rank}}}(t,c)$, is defined by induction on ${\ensuremath{\mathrm{rank}}}(c)$. If $t \equiv c$ then ${\ensuremath{\mathrm{rank}}}(t,c) = 0$. If $t \not\succsim c$ then ${\ensuremath{\mathrm{rank}}}(t,c) =
{\ensuremath{\mathrm{rank}}}(c)$. If $t \succsim c$ but $t \not\equiv c$ then $c \in
\tau_2^{\tau_1}$ and ${\ensuremath{\mathrm{rank}}}(t,c)$ is defined by $${\ensuremath{\mathrm{rank}}}(t,c) = \sup_{c_1 \in \tau_1} {\ensuremath{\mathrm{rank}}}(t_{c_1},{\ensuremath{{\cal F}}}(c)(c_1))$$ where $t_{c_1}$ is the term required by the definition of $\succsim$, such that for all terms $t_1$ with $t_1 \succ c_1$ we have $t t_1 {\ensuremath{\stackrel{*}{\Rightarrow}}}t_{c_1} \succ {\ensuremath{{\cal F}}}(c)(c_1)$. Note that ${\ensuremath{\mathrm{rank}}}(t,c) \le {\ensuremath{\mathrm{rank}}}(c)$, and if $t \succsim c$ then ${\ensuremath{\mathrm{rank}}}(t,c) <
{\ensuremath{\mathrm{rank}}}(c)$.
Two positions $p_1,p_2 \in \Sigma^*$ are *parallel* if neither $p_1 \sqsubseteq p_2$ nor $p_2 \sqsubseteq p_1$. We write $t_1 \gg^n
t_2$ if there exists a set $P \subseteq {\mbox{\cal Pos}}(t_1) \cap {\mbox{\cal Pos}}(t_2)$ of pairwise parallel positions such that for $p \in P$ we have ${\ensuremath{{t_1}_{|p}}} \succ {\ensuremath{{t_2}_{|p}}}$ and ${\ensuremath{{t_1}_{|p}}} \not\equiv
{\ensuremath{{t_2}_{|p}}}$, no free variables of ${\ensuremath{{t_1}_{|p}}}$ become bound in $t_1$, for every $p \in {\mbox{\cal Pos}}(t_1) \setminus P$ we have $p \in
{\mbox{\cal Pos}}(t_2)$ and ${\ensuremath{{t_1}^{|p}}} \equiv {\ensuremath{{t_2}^{|p}}}$, and ${\ensuremath{\mathrm{rank}}}_P(t_1,t_2) \le n$, where ${\ensuremath{\mathrm{rank}}}_P(t_1,t_2) = \sup_{p\in P}
{\ensuremath{\mathrm{rank}}}({\ensuremath{{t_1}_{|p}}}, {\ensuremath{{t_2}_{|p}}})$. We write $t_1 \gg^{<n} t_2$ if the same conditions hold except that ${\ensuremath{\mathrm{rank}}}_P(t_1,t_2) < n$. We write $t_1 \gg t_2$ if $t_1 \gg^n t_2$ for some $n \in {\ensuremath{\mathbb{N}}}$.
\[lem\_gg\_subst\] If $t_1 \gg^n t_1'$, $t_2 \gg^n t_2'$ and $x \notin
FV({\ensuremath{{t_1}_{|p}}})$ for all $p \in P_1$, where $P_1$ is the set of positions required by the definition of $t_1 \gg^n t_1'$, then $t_1[x/t_2] \gg^n t_1'[x/t_2']$.
Let $P_2$ be the set of positions required by the definition of $t_2
\gg^n t_2'$. Take $$P = P_1 \cup \{ p \in {\mbox{\cal Pos}}(t_1[x/t_2]) \;|\; p = p_1p_2,\, {\ensuremath{{t_1}_{|p_1}}} \equiv x,\, p_2 \in P_2 \}$$ as the set of positions required by the definition of $t_1[x/t_2]
\gg^n t_1'[x/t_2']$.
\[lem\_gg\] The following conditions hold.
- If $t_1 \gg^n t_2 {\ensuremath{\Rightarrow}}_{R_\beta}^\alpha
t_2'$ then $t_1 {\ensuremath{\stackrel{*}{\Rightarrow}}}t_1' \gg^n t_2'$.
- If $t_1 \gg^n t_1'$ and $t_2
{\ensuremath{\Rightarrow}}_{R_\beta}^\alpha t_1'$ then there exists $t_2'$ such that $t_2' {\ensuremath{\stackrel{*}{\Rightarrow}}}t_1$ and $t_2' \gg^n t_2$.
- If $t_1 \gg^n t_2 \succ_{R_\beta}^\alpha c$ then $t_1 \succ c$.
Induction on tuples $\langle n, \beta, \alpha, h(t_2) \rangle$ ordered lexicographically.
We first show condition (${\romannumeral 1}$). Thus suppose $t_1 \gg^n t_2
{\ensuremath{\Rightarrow}}_{R_\beta}^\alpha t_2'$. We consider possible forms of $t_2$ according to the definition of ${\ensuremath{\Rightarrow}}$.
If $t_2 \equiv t_2'$ then the claim is obvious. If $t_2
{\ensuremath{\rightarrow}}_{R_\beta}^\epsilon t_2'$ and $t_2 {\ensuremath{\Rightarrow}}_{R_\beta}^\alpha
t_2'$ follows by condition (1) in the definition of ${\ensuremath{\Rightarrow}}$, then the only non-obvious case is when $t_2 \equiv {\ensuremath{\mathrm{Eq}\,r_1\,r_2\,}}
{\ensuremath{\rightarrow}}_{R_\beta}^\epsilon {\ensuremath{\mathbf{t}}}\equiv t_2'$. Then $r_1
{\ensuremath{\stackrel{*}{\Leftrightarrow}}}_{R_{<\beta}} r_2$ and $t_1 \equiv {\ensuremath{\mathrm{Eq}\,r_1'\,r_2'\,}}$ where $r_1' \gg^n r_1$ and $r_2' \gg^n r_2$. It follows from parts (${\romannumeral 1}$) and (${\romannumeral 2}$) of the inductive hypothesis that $r_1' {\ensuremath{\stackrel{*}{\Leftrightarrow}}}r_2'$. Thus $t_1 {\ensuremath{\Rightarrow}}{\ensuremath{\mathbf{t}}}\equiv t_2'$.
If $t_2 {\ensuremath{\Rightarrow}}_{R_\beta}^\alpha t_2'$ follows by condition (2) then $t_2 \equiv r_1 r_2$, $t_2' \equiv r_1' r_2'$, $r_1
{\ensuremath{\Rightarrow}}_{R_\beta}^\alpha r_1'$ and $r_2 {\ensuremath{\Rightarrow}}_{R_\beta}^\alpha
r_2'$. We must also have $t_1 \equiv q_1 q_2$ where $q_1 \gg^n r_1$ and $q_2 \gg^n r_2$. By the inductive hypothesis ($h(r_1), h(r_2) <
h(t_2)$) there exist $q_1'$ and $q_2'$ such that $q_1 {\ensuremath{\stackrel{*}{\Rightarrow}}}q_1'$, $q_2 {\ensuremath{\stackrel{*}{\Rightarrow}}}q_2'$, $q_1' \gg^n r_1'$, $q_2' \gg^n
r_2'$. Thus $t_1 \equiv q_1 q_2 {\ensuremath{\stackrel{*}{\Rightarrow}}}q_1' q_2' \gg^n r_1' r_2'
\equiv t_2'$. If $t_2 {\ensuremath{\Rightarrow}}_{R_\beta}^\alpha t_2'$ follows by condition (3) then the argument is analogous.
Suppose $t_2 {\ensuremath{\Rightarrow}}_{R_\beta}^\alpha t_2'$ follows by condition (4). Then $t_2 \equiv (\lambda x \,.\, r_1) r_2$ and $t_2'
\equiv r_1'[x/r_2']$ where $r_1 {\ensuremath{\Rightarrow}}_{R_\beta}^\alpha r_1'$ and $r_2 {\ensuremath{\Rightarrow}}_{R_\beta}^\alpha r_2'$. We must also have $t_1 \equiv
(\lambda x \,.\, q_1) q_2$ where $q_1 \gg^n r_1$ and $q_2 \gg^n r_2$. By the inductive hypothesis there exist $q_1'$ and $q_2'$ such that $q_1 {\ensuremath{\stackrel{*}{\Rightarrow}}}q_1'$, $q_2 {\ensuremath{\stackrel{*}{\Rightarrow}}}q_2'$, $q_1' \gg^n r_1'$, $q_2' \gg^n r_2'$. By Lemma \[lem\_gg\_subst\] we obtain $q_1'[x/q_2'] \gg^n r_1'[x/r_2']$. Thus $t_1 \equiv (\lambda x
\,.\, q_1) q_2 {\ensuremath{\stackrel{*}{\Rightarrow}}}q_1'[x/q_2'] \gg^n r_1'[x/r_2'] \equiv t_2'$.
Suppose $t_2 {\ensuremath{\Rightarrow}}_{R_\beta}^\alpha t_2'$ follows by condition (5). Then $t_2 \equiv c r_1 \ldots r_m$, $r_i
{\ensuremath{\Rightarrow}}_{R_\beta}^\alpha r_i'$, $c r_1' \ldots r_m'
{\ensuremath{\rightarrow}}_{R_\beta}^\epsilon t_2'$. By the definition of $R_\beta$, the constant $c$ is not a canonical constant. This implies that $t_1
\equiv c q_1 \ldots q_m$ where $q_i \gg^n r_i$. By the inductive hypothesis ($h(r_i) < h(t_2)$) there exist $q_1',\ldots,q_m'$ such that $q_i' \gg^n r_i'$ and $q_i {\ensuremath{\stackrel{*}{\Rightarrow}}}q_i'$. Thus $c q_1'
\ldots q_m' \gg^n c r_1' \ldots r_m' {\ensuremath{\rightarrow}}_{R_\beta}^\epsilon
t_2'$. But we have already verified in this inductive step that this implies that there exists $t_1'$ such that $c q_1' \ldots q_m'
{\ensuremath{\stackrel{*}{\Rightarrow}}}t_1' \gg^n t_2'$. Therefore $t_1 \equiv c q_1 \ldots q_m
{\ensuremath{\stackrel{*}{\Rightarrow}}}c q_1' \ldots q_m {\ensuremath{\stackrel{*}{\Rightarrow}}}t_1' \gg^n t_2'$.
Suppose $t_2 {\ensuremath{\Rightarrow}}_{R_\beta}^\alpha t_2'$ follows by condition (6). Then $t_2 \equiv c r_2'$, $t_2' \equiv c_2 \equiv
{\ensuremath{{\cal F}}}(c)(c_1)$, $r_2' \succ_{R_\beta}^{<\alpha} c_1 \in \tau_1$, $\tau(c) = \tau_2^{\tau_1}$ for some $\tau_2 \in {\ensuremath{{\cal T}}}$. We also have $t_1 \equiv r_1 r_2$ where $r_1 \gg^n c$, hence $r_1 \succ c$, and $r_2 \gg^n r_2' \succ_{R_\beta}^{<\alpha} c_1$. By part (${\romannumeral 3}$) of the inductive hypothesis we obtain $r_2 \succ
c_1$. First assume that $r_1 \equiv c$. Then $t_1 \equiv r_1 r_2
\equiv c r_2 {\ensuremath{\Rightarrow}}{\ensuremath{{\cal F}}}(c)(c_1) \equiv t_2'$ by virtue of $r_2
\succ c_1$, and we are done. So suppose $r_1 \not\equiv c$. If $r_1
\succsim c$ then let $q$ be the term required by the definition of $\succsim$, such that for every term $r$ with $r \succ c_1$ we have $r_1 r {\ensuremath{\stackrel{*}{\Rightarrow}}}q \succ c_2$. Then ${\ensuremath{\mathrm{rank}}}(q,c_2) \le
{\ensuremath{\mathrm{rank}}}(r_1,c) \le n$, and $t_1 \equiv r_1 r_2 {\ensuremath{\stackrel{*}{\Rightarrow}}}q \succ
c_2$, because $r_2 \succ c_1$. So $t_1 {\ensuremath{\stackrel{*}{\Rightarrow}}}q \gg^n c_2 \equiv
t_2'$, which is our claim. Therefore suppose $r_1 \not\equiv c$ and $r_1 \not\succsim c$. Then we have ${\ensuremath{\mathrm{rank}}}(r_2,c_1) \le {\ensuremath{\mathrm{rank}}}(c_1)
\le {\ensuremath{\mathrm{rank}}}(\tau_1) < {\ensuremath{\mathrm{rank}}}(c) = {\ensuremath{\mathrm{rank}}}(r_1,c) \le n$ by Lemma \[lem\_smaller\_rank\]. Thus $r_1 r_2 \gg^{<n} r_1 c_1$. By the fact that $r_1 \succ c$ there exists $t$ such that $r_1 c_1
{\ensuremath{\stackrel{*}{\Rightarrow}}}t \succ {\ensuremath{{\cal F}}}(c)(c_1)$. By part (${\romannumeral 1}$) of the inductive hypothesis there exists $t'$ such that $r_1 r_2 {\ensuremath{\stackrel{*}{\Rightarrow}}}t' \gg^{<n} t$. Let $c_2 \equiv {\ensuremath{{\cal F}}}(c)(c_1)$. We have $t' \gg^{<n} t
\succ c_2$, so by part (${\romannumeral 3}$) of the inductive hypothesis we obtain $t' \succ c_2$. Since ${\ensuremath{\mathrm{rank}}}(t',c_2) \le {\ensuremath{\mathrm{rank}}}(c_2) \le
{\ensuremath{\mathrm{rank}}}(\tau_2) < {\ensuremath{\mathrm{rank}}}(c) = {\ensuremath{\mathrm{rank}}}(r_1,c) \le n$, we conclude that $t_1 \equiv r_1 r_2 {\ensuremath{\stackrel{*}{\Rightarrow}}}t' \gg^n c_2 \equiv t_2'$.
Suppose $t_2 {\ensuremath{\Rightarrow}}_{R_\beta}^\alpha t_2'$ follows by condition (7). Then $t_2 \equiv {\ensuremath{\mathrm{Is}\,r_2\,\tau\,}}$, $t_2' \equiv {\ensuremath{\mathbf{t}}}$, $t_1 \equiv {\ensuremath{\mathrm{Is}\,r_1\,\tau\,}}$, $r_1 \gg^n r_2$, and $r_2
\succ_{R_\beta}^{<\alpha} c$ for some $c \in \tau \in {\ensuremath{{\cal T}}}$. By part (${\romannumeral 3}$) of the inductive hypothesis we have $r_1 \succ
c$. Therefore $t_1 \equiv {\ensuremath{\mathrm{Is}\,r_1\,\tau\,}} {\ensuremath{\Rightarrow}}{\ensuremath{\mathbf{t}}}\equiv t_2'$.
Finally, suppose $t_2 {\ensuremath{\Rightarrow}}_{R_\beta}^\alpha t_2'$ follows by condition (8). Then e.g. ${\ensuremath{{t_2}^{|\epsilon}}} \equiv
{\ensuremath{{t_2'}^{|\epsilon}}} \equiv {\ensuremath{{t_1}^{|\epsilon}}} \equiv {\ensuremath{\mathrm{A}}}\tau$, and for all $c \in \tau$ there exists $t_c$ such that ${\ensuremath{{t_1}_{|c}}}
\gg^n {\ensuremath{{t_2}_{|c}}} {\ensuremath{\Rightarrow}}_{R_\beta}^\alpha t_c
{\ensuremath{\stackrel{*\;<\alpha}{\Rightarrow_{R_{\beta}}\;\,}}} {\ensuremath{{t_2'}_{|c}}}$. By part (${\romannumeral 1}$) of the inductive hypothesis ($h({\ensuremath{{t_2}_{|c}}}) <
h(t_2)$) there exists $t_c'$ such that ${\ensuremath{{t_1}_{|c}}} {\ensuremath{\stackrel{*}{\Rightarrow}}}t_c'
\gg^n t_c$. Applying the inductive hypothesis again, we conclude that there exists $q_c$ such that $t_c' {\ensuremath{\stackrel{*}{\Rightarrow}}}q_c \gg^n
{\ensuremath{{t_2'}_{|c}}}$. Let $q$ be such that ${\ensuremath{{q}^{|\epsilon}}} \equiv
{\ensuremath{\mathrm{A}}}\tau$ and ${\ensuremath{{q}_{|c}}} \equiv q_c$ for $c \in \tau$. Then ${\ensuremath{{t_1}_{|c}}} {\ensuremath{\stackrel{*}{\Rightarrow}}}{\ensuremath{{q}_{|c}}}$ for all $c \in \tau$, and hence $t_1 {\ensuremath{\rightarrow}}q$. We also have $q \gg^n t_2'$, because ${\ensuremath{{q}_{|c}}}
\gg^n {\ensuremath{{t_2'}_{|c}}}$ for all $c \in \tau$.
We now show condition (${\romannumeral 2}$). Thus suppose $t_1 \gg^n t_1'$ and $t_2 {\ensuremath{\Rightarrow}}_{R_\beta}^\alpha t_1'$. We consider possible forms of $t_2$ according to the definition of $t_2
{\ensuremath{\Rightarrow}}_{R_\beta}^\alpha t_1'$.
If $t_2 \equiv t_1'$ then the claim is obvious. If $t_2
{\ensuremath{\Rightarrow}}_{R_\beta}^\alpha t_1'$ follows by condition (1) in the definition of ${\ensuremath{\Rightarrow}}_{R_\beta}^\alpha$, then $t_2
{\ensuremath{\rightarrow}}_{R_\beta}^\epsilon t_1'$ and the claim follows easily by inspecting the definition of $R_\beta$. If $t_2
{\ensuremath{\Rightarrow}}_{R_\beta}^\alpha t_1'$ follows by condition (2), then $t_2
\equiv r_1 r_2$, $t_1' \equiv r_1' r_2'$, $r_1
{\ensuremath{\Rightarrow}}_{R_\beta}^\alpha r_1'$, $r_2 {\ensuremath{\Rightarrow}}_{R_\beta}^\alpha
r_2'$, $t_1 \equiv q_1' q_2'$, $q_1' \gg^n r_1'$, $q_2' \gg^n
r_2'$. By the inductive hypothesis ($h(r_1), h(r_2) < h(t_2)$) there exist $q_1$, $q_2$ such that $q_1 {\ensuremath{\stackrel{*}{\Rightarrow}}}q_1'$, $q_2 {\ensuremath{\stackrel{*}{\Rightarrow}}}q_2'$, $q_1 \gg^n r_1$, $q_2 \gg^n r_2$. Thus $q_1 q_2 {\ensuremath{\stackrel{*}{\Rightarrow}}}t_1$ and $q_1 q_2 \gg^n t_2$. If $t_2 {\ensuremath{\Rightarrow}}_{R_\beta}^\alpha
t_1'$ follows by condition (3) or condition (4) then the argument is analogous.
If $t_2 {\ensuremath{\Rightarrow}}_{R_\beta}^\alpha t_1'$ follows by condition (5) then $t_2 \equiv c r_1 \ldots r_m$, $r_i {\ensuremath{\Rightarrow}}_{R_\beta}^\alpha
r_i'$, $c r_1' \ldots r_m' {\ensuremath{\rightarrow}}_{R_\beta}^\epsilon t_1'$, $t_1
\equiv c q_1' \ldots q_m'$, $q_i' \gg^n r_i'$. By the inductive hypothesis there exist $q_1,\ldots,q_m$ such that $q_i {\ensuremath{\stackrel{*}{\Rightarrow}}}q_i'$ and $q_i \gg^n r_i$. Let $q \equiv c q_1 \ldots q_m$. We have $q \gg^n t_2$ and $q {\ensuremath{\stackrel{*}{\Rightarrow}}}c q_1' \ldots q_m' \equiv t_1$.
If $t_2 {\ensuremath{\Rightarrow}}_{R_\beta}^\alpha t_1'$ follows by condition (6) then $t_2 \equiv c r$, $t_1' \equiv c_2 \equiv {\ensuremath{{\cal F}}}(c)(c_1)$, $c \in
\tau_2^{\tau_1}$, $r \succ_{R_\beta}^{<\alpha} c_1 \in \tau_1$, $t_1
\gg^n c_2$. If $t_1 \equiv c_2$ then the claim is obvious. Otherwise $t_1 \succ c_2$. Let $e \in (\mbox{\small Bool}^{\tau_1})^{\tau_1}$ be such that for $d_1, d_2 \in \tau_1$, we have ${\ensuremath{{\cal F}}}(e)(d_1)(d_2)
\equiv {\ensuremath{\mathbf{t}}}$ if $d_1 \equiv d_2$, and ${\ensuremath{{\cal F}}}(e)(d_1)(d_2) \equiv
{\ensuremath{\mathbf{f}}}$ if $d_1 \not\equiv d_2$. Let $q \equiv \lambda x . {\ensuremath{\mathrm{Cond}\,(e
c_1 x)\,t_1\,(c x)\,}}$. If $s \succ d \in \tau_1$ and $d \not\equiv
c_1$ then $q s {\ensuremath{\Rightarrow}}{\ensuremath{\mathrm{Cond}\,(e c_1 s)\,t_1\,(c s)\,}} {\ensuremath{\stackrel{*}{\Rightarrow}}}{\ensuremath{\mathrm{Cond}\,{\ensuremath{\mathbf{f}}}\,t_1\,{\ensuremath{{\cal F}}}(c)(d)\,}} {\ensuremath{\Rightarrow}}{\ensuremath{{\cal F}}}(c)(d)$. If $s \succ c_1$ then $q s {\ensuremath{\Rightarrow}}{\ensuremath{\mathrm{Cond}\,(e c_1 s)\,t_1\,(c s)\,}} {\ensuremath{\Rightarrow}}{\ensuremath{\mathrm{Cond}\,{\ensuremath{\mathbf{t}}}\,t_1\,(c s)\,}} {\ensuremath{\Rightarrow}}t_1 \succ {\ensuremath{{\cal F}}}(c)(c_1) \equiv
c_2$. Therefore $q \succsim c$, and ${\ensuremath{\mathrm{rank}}}(q,c) = {\ensuremath{\mathrm{rank}}}(t_1,c_2) \le
n$, since ${\ensuremath{\mathrm{rank}}}({\ensuremath{{\cal F}}}(c)(d), {\ensuremath{{\cal F}}}(c)(d)) = 0$. Thus $q r \gg^n c r
\equiv t_2$ and $q r {\ensuremath{\stackrel{*}{\Rightarrow}}}t_1$, because $r \succ c_1$.
If $t_2 {\ensuremath{\Rightarrow}}_{R_\beta}^\alpha t_1'$ follows by condition (7) then the claim is obvious. If $t_2 {\ensuremath{\Rightarrow}}_{R_\beta}^\alpha t_1'$ follows by condition (8) then e.g ${\ensuremath{{t_2}^{|\epsilon}}} \equiv
{\ensuremath{{t_1'}^{|\epsilon}}} \equiv {\ensuremath{{t_1}^{|\epsilon}}} \equiv {\ensuremath{\mathrm{A}}}\tau$ and for all $c \in \tau$ there exists $t_c$ such that ${\ensuremath{{t_2}_{|c}}}
{\ensuremath{\Rightarrow}}_{R_\beta}^\alpha t_c {\ensuremath{\stackrel{*}{\Rightarrow}}}_{R_\beta}^{<\alpha}
{\ensuremath{{t_1'}_{|c}}}$. We also have ${\ensuremath{{t_1}_{|c}}} \gg^n {\ensuremath{{t_1'}_{|c}}}$ for all $c \in \tau$. Therefore by the inductive hypothesis for every $c
\in \tau$ there exists $t_c'$ such that $t_c' \gg^n t_c$ and $t_c'
{\ensuremath{\stackrel{*}{\Rightarrow}}}{\ensuremath{{t_1}_{|c}}}$. Again, by the inductive hypothesis ($h({\ensuremath{{t_2}_{|c}}}) < h(t_2)$), for every $c \in \tau$ there exists $q_c$ such that $q_c \gg^n {\ensuremath{{t_2}_{|c}}}$ and $q_c {\ensuremath{\stackrel{*}{\Rightarrow}}}t_c' {\ensuremath{\stackrel{*}{\Rightarrow}}}{\ensuremath{{t_1}_{|c}}}$. Let $q$ be such that ${\ensuremath{{q}^{|\epsilon}}} \equiv {\ensuremath{\mathrm{A}}}\tau$ and ${\ensuremath{{q}_{|c}}} \equiv q_c$ for $c \in \tau$. Then $q \gg^n t_2$ and $q {\ensuremath{\Rightarrow}}t_1$.
It remains to show condition (${\romannumeral 3}$). Thus suppose $t_1 \gg^n
t_2 \succ_{R_\beta}^\alpha c$. If $t_2 \equiv c$ then the claim is obvious. Otherwise $c \in \tau_2^{\tau_1}$ and for every $c_1 \in
\tau_1$ there exists $q_{c_1}$ such that $t_2 c_1
{\ensuremath{\stackrel{*\;<\alpha}{\Rightarrow_{R_\beta}\;\,}}} q_{c_1} \succ_{R_\beta}^{<\alpha}
c$. Since $t_1 c_1 \gg^n t_2 c_1$, by part (${\romannumeral 1}$) of the inductive hypothesis for each $c_1 \in \tau_1$ there exists $q_{c_1}'$ such that $t_1 c_1 {\ensuremath{\stackrel{*}{\Rightarrow}}}q_{c_1}' \gg^n
q_{c_1} \succ_{R_\beta}^{<\alpha} c$. By part (${\romannumeral 3}$) of the inductive hypothesis we obtain $q_{c_1}' \succ c$. This implies that $t_1 \succ c$.
\[cor\_succ\_arg\] If $t_2 \succ c$ and $t_1 c {\ensuremath{\stackrel{*}{\Rightarrow}}}d \in {\ensuremath{\mathrm{Bool}}}$, then $t_1 t_2
{\ensuremath{\stackrel{*}{\Rightarrow}}}d$.
The above lemma and the ensuing corollary confirm our intuition about the meaning of $\succ$. This basically finishes the hard part of the proof. What remains are some relatively straightforward lemmas.
We write $t {\ensuremath{\vartriangleright}}\tau$ if $t {\ensuremath{\stackrel{*}{\Rightarrow}}}t' \succ c$ for some term $t'$ and some $c \in \tau$.
\[lem\_is\] We have ${\ensuremath{\mathrm{Is}\,t_1\,t_2\,}} {\ensuremath{\stackrel{*}{\Leftrightarrow}}}{\ensuremath{\mathbf{t}}}$ iff there exists $\tau \in
{\ensuremath{{\cal T}}}$ such that $t_1 {\ensuremath{\vartriangleright}}\tau$ and $t_2 {\ensuremath{\stackrel{*}{\Rightarrow}}}\tau$. Moreover, this $\tau \in {\ensuremath{{\cal T}}}$ is uniquely determined.
If $t_2 {\ensuremath{\stackrel{*}{\Rightarrow}}}\tau \in {\ensuremath{{\cal T}}}$ and $t_1 {\ensuremath{\stackrel{*}{\Rightarrow}}}t' \succ c$ for some $c \in \tau$ then ${\ensuremath{\mathrm{Is}\,t_1\,t_2\,}} {\ensuremath{\stackrel{*}{\Rightarrow}}}{\ensuremath{\mathrm{Is}\,t_1\,\tau\,}}
{\ensuremath{\stackrel{*}{\Rightarrow}}}{\ensuremath{\mathrm{Is}\,t'\,\tau\,}} {\ensuremath{\Rightarrow}}{\ensuremath{\mathbf{t}}}$ by condition (7) in the definition of ${\ensuremath{\Rightarrow}}$. If ${\ensuremath{\mathrm{Is}\,t_1\,t_2\,}} {\ensuremath{\stackrel{*}{\Leftrightarrow}}}{\ensuremath{\mathbf{t}}}$ then ${\ensuremath{\mathrm{Is}\,t_1\,t_2\,}} {\ensuremath{\stackrel{*}{\Rightarrow}}}{\ensuremath{\mathbf{t}}}$ by the Church-Rosser property of ${\ensuremath{\Rightarrow}}$ and the fact that ${\ensuremath{\mathbf{t}}}$ is in normal form. But this is only possible when ${\ensuremath{\mathrm{Is}\,t_1\,t_2\,}} {\ensuremath{\stackrel{*}{\Rightarrow}}}{\ensuremath{\mathrm{Is}\,t'\,\tau\,}}
{\ensuremath{\Rightarrow}}{\ensuremath{\mathbf{t}}}$ where $\tau \in {\ensuremath{{\cal T}}}$, $t_2 {\ensuremath{\stackrel{*}{\Rightarrow}}}\tau$ and $t_1
{\ensuremath{\stackrel{*}{\Rightarrow}}}t' \succ c$ for some $c \in \tau$, i.e. $t_1 {\ensuremath{\vartriangleright}}\tau$.
To see that $\tau$ is uniquely determined it suffices to notice that it is in normal form w.r.t. ${\ensuremath{\Rightarrow}}$ and ${\ensuremath{\Rightarrow}}$ has the Church-Rosser property.
\[lem\_forall\] The following conditions hold.
- If $\tau \in {\ensuremath{{\cal T}}}$ and for all $t_2$ such that $t_2 {\ensuremath{\vartriangleright}}\tau$ we have $t_1 t_2 {\ensuremath{\stackrel{*}{\Leftrightarrow}}}{\ensuremath{\mathbf{t}}}$, then $\forall \tau t_1 {\ensuremath{\stackrel{*}{\Leftrightarrow}}}{\ensuremath{\mathbf{t}}}$.
- If $\tau \in {\ensuremath{{\cal T}}}$ and there exists $t_2$ such that $t_2 {\ensuremath{\vartriangleright}}\tau$ and $t_1 t_2 {\ensuremath{\stackrel{*}{\Leftrightarrow}}}{\ensuremath{\mathbf{f}}}$, then $\forall \tau t_1
{\ensuremath{\stackrel{*}{\Leftrightarrow}}}{\ensuremath{\mathbf{f}}}$.
- If $\forall t_1 t_2 {\ensuremath{\stackrel{*}{\Leftrightarrow}}}{\ensuremath{\mathbf{t}}}$ then for all $t_3$ such that ${\ensuremath{\mathrm{Is}\,t_3\,t_1\,}} {\ensuremath{\stackrel{*}{\Leftrightarrow}}}{\ensuremath{\mathbf{t}}}$ we have $t_2 t_3 {\ensuremath{\stackrel{*}{\Leftrightarrow}}}{\ensuremath{\mathbf{t}}}$.
- If $\forall t_1 t_2 {\ensuremath{\stackrel{*}{\Leftrightarrow}}}{\ensuremath{\mathbf{f}}}$ then there exists $t_3$ such that ${\ensuremath{\mathrm{Is}\,t_3\,t_1\,}} {\ensuremath{\stackrel{*}{\Leftrightarrow}}}{\ensuremath{\mathbf{t}}}$ and $t_2 t_3
{\ensuremath{\stackrel{*}{\Leftrightarrow}}}{\ensuremath{\mathbf{f}}}$.
We show condition (a). Suppose $\tau \in {\ensuremath{{\cal T}}}$ and for all $t_2$ such that $t_2 {\ensuremath{\vartriangleright}}\tau$ we have $t_1 t_2 {\ensuremath{\stackrel{*}{\Leftrightarrow}}}{\ensuremath{\mathbf{t}}}$. If $\tau
\equiv \emptyset$ then $\forall \emptyset t_1 {\ensuremath{\rightarrow}}{\ensuremath{\mathbf{t}}}$ and the claim is obvious. Otherwise we have $\forall \tau t_1 {\ensuremath{\rightarrow}}t$ where ${\ensuremath{{t}^{|\epsilon}}} \equiv {\ensuremath{\mathrm{A}}}\tau$ and ${\ensuremath{{t}_{|c}}} \equiv
t_1 c$ for $c \in \tau$. Since $c \in \tau$ we have $c {\ensuremath{\vartriangleright}}\tau$, so $t_1 c {\ensuremath{\stackrel{*}{\Leftrightarrow}}}{\ensuremath{\mathbf{t}}}$, hence $t_1 c {\ensuremath{\stackrel{*}{\Rightarrow}}}{\ensuremath{\mathbf{t}}}$ by the Church-Rosser property. Therefore $t {\ensuremath{\stackrel{*}{\Rightarrow}}}{\ensuremath{\mathbf{t}}}$. Condition (b) is shown in a completely analogous way.
We show condition (c). Suppose $\forall t_1 t_2 {\ensuremath{\stackrel{*}{\Leftrightarrow}}}{\ensuremath{\mathbf{t}}}$. Then $\forall t_1 t_2 {\ensuremath{\stackrel{*}{\Rightarrow}}}{\ensuremath{\mathbf{t}}}$, which is only possible when $\forall \tau t_2 {\ensuremath{\stackrel{*}{\Rightarrow}}}{\ensuremath{\mathbf{t}}}$ and $t_1 {\ensuremath{\stackrel{*}{\Rightarrow}}}\tau$ for some $\tau \in {\ensuremath{{\cal T}}}$. Suppose $t_3$ is such that ${\ensuremath{\mathrm{Is}\,t_3\,t_1\,}} {\ensuremath{\stackrel{*}{\Leftrightarrow}}}{\ensuremath{\mathbf{t}}}$. Then ${\ensuremath{\mathrm{Is}\,t_3\,t_1\,}} {\ensuremath{\stackrel{*}{\Rightarrow}}}{\ensuremath{\mathbf{t}}}$, and we must have $t_1 {\ensuremath{\stackrel{*}{\Rightarrow}}}\tau' \in {\ensuremath{{\cal T}}}$ and $t_3 {\ensuremath{\vartriangleright}}\tau'$, by Lemma \[lem\_is\]. Because both $\tau$ and $\tau'$ are in normal form, we conclude by the Church-Rosser property that $\tau
\equiv \tau'$. Hence $t_3 {\ensuremath{\vartriangleright}}\tau$, i.e. $t_3 {\ensuremath{\stackrel{*}{\Rightarrow}}}t_3'
\succ c \in \tau$. Since $\forall \tau t_2 {\ensuremath{\stackrel{*}{\Rightarrow}}}{\ensuremath{\mathbf{t}}}$, it is easy to see by inspecting the definitions that $t_2 c {\ensuremath{\stackrel{*}{\Rightarrow}}}{\ensuremath{\mathbf{t}}}$. By Corollary \[cor\_succ\_arg\] we obtain $t_2 t_3 {\ensuremath{\stackrel{*}{\Rightarrow}}}t_2 t_3' {\ensuremath{\stackrel{*}{\Rightarrow}}}{\ensuremath{\mathbf{t}}}$.
Condition (c) follows easily from definitions, Lemma \[lem\_is\] and the Church-Rosser property.
\[lem\_fun\] The following conditions hold.
- If $\tau_1, \tau_2 \in {\ensuremath{{\cal T}}}$ and for all $t_2$ such that $t_2
{\ensuremath{\vartriangleright}}\tau_1$ we have $t_1 t_2 {\ensuremath{\vartriangleright}}\tau_2$, then $t_1 {\ensuremath{\vartriangleright}}\tau_2^{\tau_1}$.
- If $\tau_1, \tau_2 \in {\ensuremath{{\cal T}}}$, $t_1 {\ensuremath{\vartriangleright}}\tau_2^{\tau_1}$ and $t_2 {\ensuremath{\vartriangleright}}\tau_1$ then $t_1 t_2 {\ensuremath{\vartriangleright}}\tau_2$.
We show condition (a). Suppose $\tau_1, \tau_2 \in {\ensuremath{{\cal T}}}$ and for all $t_2$ such that $t_2 {\ensuremath{\vartriangleright}}\tau_1$ we have $t_1 t_2 {\ensuremath{\vartriangleright}}\tau_2$. Let $c \in \tau_1$. We obviously have $c {\ensuremath{\vartriangleright}}\tau_1$, so $t_1 c {\ensuremath{\vartriangleright}}\tau_2$, i.e. there exists a term $t_c$ and a constant $c' \tau_2$ such that $t_1 c {\ensuremath{\stackrel{*}{\Rightarrow}}}t_c \succ
c'$. Recall that $\tau_2^{\tau_1}$ consists of all set-theoretic functions from $\tau_1$ to $\tau_2$. In particular, there exists $d
\in \tau_2^{\tau_1}$ such that ${\ensuremath{{\cal F}}}(d)(c) \equiv c'$ for every $c
\in \tau_1$ and $c' \in \tau_2$ depending on $c$ as above. But then $t_1 \succ d$, and hence $t_1 {\ensuremath{\vartriangleright}}\tau_2^{\tau_1}$.
We show condition (b). Suppose $\tau_1, \tau_2 \in {\ensuremath{{\cal T}}}$, $t_1 {\ensuremath{\vartriangleright}}\tau_2^{\tau_1}$ and $t_2 {\ensuremath{\vartriangleright}}\tau_1$. Then $t_1 {\ensuremath{\stackrel{*}{\Rightarrow}}}t_1'
\succ c \in \tau_2^{\tau_1}$ and $t_2 {\ensuremath{\stackrel{*}{\Rightarrow}}}t_2' \succ c_1 \in
\tau_1$. By condition (6) in the definition of ${\ensuremath{\Rightarrow}}$ we obtain $c t_2' {\ensuremath{\Rightarrow}}{\ensuremath{{\cal F}}}(c)(c_1) \succ {\ensuremath{{\cal F}}}(c)(c_1) \in \tau_2$. If $t_1'
\equiv c$ then $t_1 t_2 {\ensuremath{\stackrel{*}{\Rightarrow}}}c t_2' {\ensuremath{\Rightarrow}}{\ensuremath{{\cal F}}}(c)(c_1) \in
\tau_2$, so $t_1 t_2 {\ensuremath{\vartriangleright}}\tau_2$. Otherwise $t_1' t_2' \gg c t_2'
{\ensuremath{\Rightarrow}}{\ensuremath{{\cal F}}}(c)(c_1)$, and by part (${\romannumeral 1}$) of Lemma \[lem\_gg\] there exists $t$ such that $t_1 t_2 {\ensuremath{\stackrel{*}{\Rightarrow}}}t_1' t_2' {\ensuremath{\stackrel{*}{\Rightarrow}}}t \succ {\ensuremath{{\cal F}}}(c)(c_1) \in \tau_2$. Hence $t_1 t_2
{\ensuremath{\vartriangleright}}\tau_2$.
\[lem\_subtype\] If $\mathrm{Subtype}\,t_1\,t_2 {\ensuremath{\stackrel{*}{\Leftrightarrow}}}\tau \in {\ensuremath{{\cal T}}}$ then there exists $\tau' \in {\ensuremath{{\cal T}}}$ such that $\tau \subseteq \tau'$, $t_1
{\ensuremath{\stackrel{*}{\Rightarrow}}}\tau'$, and for all terms $t_3$:
- if $t_3 {\ensuremath{\vartriangleright}}\tau'$ then $t_2 t_3 {\ensuremath{\vartriangleright}}{\ensuremath{\mathrm{Bool}}}$,
- $t_3 {\ensuremath{\vartriangleright}}\tau$ iff $t_3 {\ensuremath{\vartriangleright}}\tau'$ and $t_2 t_3
{\ensuremath{\stackrel{*}{\Leftrightarrow}}}{\ensuremath{\mathbf{t}}}$.
Suppose $\mathrm{Subtype}\,t_1\,t_2 {\ensuremath{\stackrel{*}{\Leftrightarrow}}}\tau \in
{\ensuremath{{\cal T}}}$. Since $\tau$ is in normal form, we conclude by the Church-Rosser property that $\mathrm{Subtype}\,t_1\,t_2 {\ensuremath{\stackrel{*}{\Rightarrow}}}\tau$. By inspecting the definition of $R$ we see that this is only possible when $t_1 {\ensuremath{\stackrel{*}{\Rightarrow}}}\tau' \in {\ensuremath{{\cal T}}}$. If $\tau' \equiv
\emptyset$ then $\mathrm{Subtype}\,t_1\,t_2 {\ensuremath{\stackrel{*}{\Rightarrow}}}\emptyset$ and by the Church-Rosser property we obtain $\tau \equiv
\emptyset$. If $\tau \equiv \emptyset$ then obviously $\tau
\subseteq \tau'$ and both conditions (a) and (b) are satisfied, because $t_3 \not{\ensuremath{\vartriangleright}}\emptyset$ for any $t_3$, as $t_3 {\ensuremath{\vartriangleright}}\emptyset$ would require the existence of some $c \in \emptyset$.
So suppose $\tau' \not\equiv \emptyset$ and $\tau \not\equiv
\emptyset$. Then $\mathrm{Subtype}\,\tau'\,t_2 {\ensuremath{\rightarrow}}t$ where ${\ensuremath{{t}^{|\epsilon}}} \equiv {\ensuremath{\mathrm{S}}}\tau'$ and ${\ensuremath{{t}_{|c}}} \equiv t_2 c$ for $c \in \tau'$. By the Church-Rosser property we have $t
{\ensuremath{\stackrel{*}{\Rightarrow}}}\tau$, and by inspecting the definitions we easily see that this is only possible when $t_2 c \equiv {\ensuremath{{t}_{|c}}} {\ensuremath{\stackrel{*}{\Rightarrow}}}d_c \in {\ensuremath{\mathrm{Bool}}}$ for every $c \in \tau'$ and $\tau = \{ c \in \tau'
\;|\; d_c \equiv {\ensuremath{\mathbf{t}}}\}$. But then obviously $\tau \subseteq
\tau'$. To show (a) suppose $t_3 {\ensuremath{\vartriangleright}}\tau'$, i.e. $t_3 {\ensuremath{\stackrel{*}{\Rightarrow}}}t_3' \succ c \in \tau'$. By Corollary \[cor\_succ\_arg\] we obtain $t_2 t_3 {\ensuremath{\stackrel{*}{\Rightarrow}}}d_c \in {\ensuremath{\mathrm{Bool}}}$. We now prove (b). Suppose $t_3
{\ensuremath{\vartriangleright}}\tau$, i.e. $t_3 {\ensuremath{\stackrel{*}{\Rightarrow}}}t_3' \succ c \in \tau$. Since $\tau
\subseteq \tau'$ we obviously have $t_3 {\ensuremath{\vartriangleright}}\tau'$. By Corollary \[cor\_succ\_arg\] we obtain $t_2 t_3 {\ensuremath{\stackrel{*}{\Rightarrow}}}d_c$. Since $c \in \tau$ we have $d_c \equiv {\ensuremath{\mathbf{t}}}$. For the other direction, assume $t_2 t_3 {\ensuremath{\stackrel{*}{\Rightarrow}}}{\ensuremath{\mathbf{t}}}$ and $t_3 {\ensuremath{\vartriangleright}}\tau'$, i.e. $t_3 {\ensuremath{\stackrel{*}{\Rightarrow}}}t_3' \succ c \in \tau'$. By Corollary \[cor\_succ\_arg\] we have $t_2 t_3 {\ensuremath{\stackrel{*}{\Rightarrow}}}d_c \in
{\ensuremath{\mathrm{Bool}}}$. By the Church-Rosser property we conclude $d_c \equiv
{\ensuremath{\mathbf{t}}}$. Hence $c \in \tau$ and $t_3 {\ensuremath{\vartriangleright}}\tau$.
\[lem\_or\_eq\_choice\]
- $\vee t_1 t_2 {\ensuremath{\stackrel{*}{\Leftrightarrow}}}{\ensuremath{\mathbf{t}}}$ iff $t_1 {\ensuremath{\stackrel{*}{\Leftrightarrow}}}{\ensuremath{\mathbf{t}}}$ or $t_2 {\ensuremath{\stackrel{*}{\Leftrightarrow}}}{\ensuremath{\mathbf{t}}}$.
- $\vee t_1 t_2 {\ensuremath{\stackrel{*}{\Leftrightarrow}}}{\ensuremath{\mathbf{f}}}$ iff $t_1 {\ensuremath{\stackrel{*}{\Leftrightarrow}}}{\ensuremath{\mathbf{f}}}$ and $t_2 {\ensuremath{\stackrel{*}{\Leftrightarrow}}}{\ensuremath{\mathbf{f}}}$.
- ${\ensuremath{\mathrm{Eq}\,t_1\,t_2\,}} {\ensuremath{\stackrel{*}{\Leftrightarrow}}}{\ensuremath{\mathbf{t}}}$ iff $t_1 {\ensuremath{\stackrel{*}{\Leftrightarrow}}}t_2$.
- If $\forall t \lambda x . {\ensuremath{\mathbf{f}}}{\ensuremath{\stackrel{*}{\Leftrightarrow}}}{\ensuremath{\mathbf{f}}}$ with $x \notin FV(t_2)$, then $\mathrm{Is}\, (\mathrm{Choice}\, t)\, t
{\ensuremath{\stackrel{*}{\Leftrightarrow}}}{\ensuremath{\mathbf{t}}}$.
Follows easily from definitions and the Church-Rosser property.
\[lem\_inductive\] Suppose $\iota \in {\ensuremath{{\cal T}}}_I$, $\iota = \mu(\langle \iota_{1,1}, \ldots,
\iota_{1,n_1}\rangle, \ldots, \langle
\iota_{m,1},\ldots,\iota_{m,n_m} \rangle)$, $c_i^\iota \in {\ensuremath{{\cal C}}}$ has arity $n_i$, $\iota_{i,j}^* = \iota_{i,j}$ if $\iota_{i,j} \in
{\ensuremath{{\cal T}}}_I$, $\iota_{i,j}^* = \iota$ if $\iota_{i,j} = \star$, and $\bar{\iota}_{i,j}^* \in \bar{{\ensuremath{{\cal T}}}}_I$ is the subset of ${\ensuremath{{\cal K}}}$ determined by $\iota_{i,j}^*$ (see Definition \[def\_canonical\]). Let $t$ be a term.
If for all $i=1,\ldots,m$ we have:
- for all $t_1,\ldots,t_{n_i}$ such that
- $t_j {\ensuremath{\vartriangleright}}\bar{\iota}_{i,j}^*$ for $j=1,\ldots,n_i$, and
- $t t_j {\ensuremath{\stackrel{*}{\Rightarrow}}}{\ensuremath{\mathbf{t}}}$ for all $1 \le j \le n_i$ such that $\iota_{i,j} = \star$
we have $t (c_i^\iota t_1 \ldots t_{n_i}) {\ensuremath{\stackrel{*}{\Rightarrow}}}{\ensuremath{\mathbf{t}}}$
then $\forall \bar{\iota} t {\ensuremath{\stackrel{*}{\Rightarrow}}}{\ensuremath{\mathbf{t}}}$.
Note that for every ${\ensuremath{{\cal I}}}_s$-term in ${\ensuremath{\mathbb T}}_I$ (see Definition \[def\_canonical\]) there exists a semantic term corresponding to it in the obvious way. Thus, we may assume without loss of generality that ${\ensuremath{\mathbb T}}_I$ is a set of semantic terms. Let $s
\in \bar{\iota}$. It follows by straightforward induction on the structure of the term $r \in {\ensuremath{\mathbb T}}_I$ corresponding to $s$ (see Definition \[def\_canonical\]) that $t r {\ensuremath{\stackrel{*}{\Rightarrow}}}{\ensuremath{\mathbf{t}}}$. It is easy to see from the definition of $R$ that $r {\ensuremath{\Rightarrow}}s$. Therefore $t s {\ensuremath{\stackrel{*}{\Rightarrow}}}{\ensuremath{\mathbf{t}}}$.
\[lem\_truth\_value\] $\forall \, (\mathrm{Subtype}\,{\ensuremath{\mathrm{Bool}}}\,(\lambda x . {\ensuremath{\mathbf{f}}})) \,
(\lambda x . {\ensuremath{\mathbf{f}}}) {\ensuremath{\Rightarrow}}{\ensuremath{\mathbf{t}}}$.
It is easy to see that $\mathrm{Subtype}\,{\ensuremath{\mathrm{Bool}}}\,(\lambda x \,.\,
{\ensuremath{\mathbf{f}}}) {\ensuremath{\Rightarrow}}\emptyset$, and by definition of $R$ we have $\forall\, \emptyset \lambda x \,.\, {\ensuremath{\mathbf{f}}}{\ensuremath{\rightarrow}}{\ensuremath{\mathbf{t}}}$.
\[lem\_supset\_t2\] If $\forall \, (\mathrm{Subtype}\,{\ensuremath{\mathrm{Bool}}}\,(\lambda x . t)) {\ensuremath{\stackrel{*}{\Rightarrow}}}{\ensuremath{\mathbf{t}}}$ and $x \notin FV(t)$ then $t {\ensuremath{\stackrel{*}{\Rightarrow}}}\top$ or $t
{\ensuremath{\stackrel{*}{\Rightarrow}}}\bot$.
Follows easily from definitions and the Church-Rosser property.
For $\tau \in {\ensuremath{{\cal T}}}$, we write $t_1 \sim_\tau t_2$ if for every term $t$ such that $t {\ensuremath{\vartriangleright}}{\ensuremath{\mathrm{Bool}}}^\tau$ we have $t t_1 {\ensuremath{\stackrel{*}{\Leftrightarrow}}}t
t_2$.
\[lem\_ext\_fun\] If $\tau_1, \tau_2 \in {\ensuremath{{\cal T}}}$, $\tau = \tau_2^{\tau_1}$, $t_1 {\ensuremath{\vartriangleright}}\tau$, $t_2 {\ensuremath{\vartriangleright}}\tau$ and for all $s$ such that $s {\ensuremath{\vartriangleright}}\tau_1$ we have $t_1 s \sim_{\tau_2} t_2 s$, then $t_1 \sim_{\tau} t_2$.
We have $t_1 {\ensuremath{\stackrel{*}{\Rightarrow}}}t_1' \succ c_1 \in \tau$ and $t_2 {\ensuremath{\stackrel{*}{\Rightarrow}}}t_2' \succ c_2 \in \tau$. Let $d_1 \in \tau_1$, $d_2 \in \tau_2$ and let $f \in {\ensuremath{\mathrm{Bool}}}^{\tau_2}$ be such that ${\ensuremath{{\cal F}}}(f)(d_2) \equiv
{\ensuremath{\mathbf{t}}}$ and ${\ensuremath{{\cal F}}}(f)(d) \equiv {\ensuremath{\mathbf{f}}}$ for $d \ne d_2$. Since $t_1'
\succ c_1$, there exists $s_1$ such that $t_1' d_1 {\ensuremath{\stackrel{*}{\Rightarrow}}}s_1
\succ {\ensuremath{{\cal F}}}(c_1)(d_1)$. Analogously $t_2' d_1 {\ensuremath{\stackrel{*}{\Rightarrow}}}s_2 \succ
{\ensuremath{{\cal F}}}(c_2)(d_1)$. Because $t_1 d_1 \sim_{\tau_2} t_2 d_1$, we have $f
(t_1 d_1) {\ensuremath{\stackrel{*}{\Leftrightarrow}}}f (t_2 d_1)$. We have $f (t_1 d_1) {\ensuremath{\stackrel{*}{\Rightarrow}}}f
(t_1' d_1) {\ensuremath{\stackrel{*}{\Rightarrow}}}f s_1$. Since $s_1 \succ {\ensuremath{{\cal F}}}(c_1)(d_1) \in
\tau_2$ and $f \in {\ensuremath{\mathrm{Bool}}}^{\tau_2}$, we obtain $f s_1 {\ensuremath{\Rightarrow}}{\ensuremath{{\cal F}}}(f)({\ensuremath{{\cal F}}}(c_1)(d_1))$ by condition (6) in the definition of ${\ensuremath{\Rightarrow}}$. Analogously $f (t_2 d_1) {\ensuremath{\stackrel{*}{\Rightarrow}}}{\ensuremath{{\cal F}}}(f)({\ensuremath{{\cal F}}}(c_2)(d_1))$. Hence, by the Church-Rosser property ${\ensuremath{{\cal F}}}(f)({\ensuremath{{\cal F}}}(c_1)(d_1)) \equiv {\ensuremath{{\cal F}}}(f)({\ensuremath{{\cal F}}}(c_2)(d_1))$. In other words, ${\ensuremath{{\cal F}}}(c_1)(d_1) \equiv d_2$ iff ${\ensuremath{{\cal F}}}(c_2)(d_1) \equiv
d_2$. Since $d_2 \in \tau_2$ was arbitrary, ${\ensuremath{{\cal F}}}(c_1)(d_1) \equiv
{\ensuremath{{\cal F}}}(c_2)(d_1)$. But $d_1 \in \tau_1$ was also arbitrary, so $c_1
\equiv c_2$.
Now, let $t$ be such that $t {\ensuremath{\vartriangleright}}{\ensuremath{\mathrm{Bool}}}^{\tau}$. Then $t {\ensuremath{\stackrel{*}{\Rightarrow}}}t' \succ p \in {\ensuremath{\mathrm{Bool}}}^{\tau}$. We have $t t_1 {\ensuremath{\stackrel{*}{\Rightarrow}}}t' t_1
{\ensuremath{\stackrel{*}{\Rightarrow}}}t' t_1' \gg p c_1 {\ensuremath{\Rightarrow}}{\ensuremath{{\cal F}}}(p)(c_1)$. By part (${\romannumeral 1}$) of Lemma \[lem\_gg\] we have $t' t_1' {\ensuremath{\stackrel{*}{\Rightarrow}}}r_1 \gg {\ensuremath{{\cal F}}}(p)(c_1)$ for some term $r_1$. But ${\ensuremath{{\cal F}}}(p)(c_1) \in
{\ensuremath{\mathrm{Bool}}}$, so in fact $r_1 \equiv {\ensuremath{{\cal F}}}(p)(c_1)$ Analogously $t t_2
{\ensuremath{\stackrel{*}{\Rightarrow}}}{\ensuremath{{\cal F}}}(p)(c_2) \equiv {\ensuremath{{\cal F}}}(p)(c_1)$. Therefore $t t_1
{\ensuremath{\stackrel{*}{\Leftrightarrow}}}t t_2$. Hence $t_1 \sim_\tau t_2$.
\[lem\_ext\_bool\] Define $t_1 \supset t_2$ by $t_1 \supset t_2 \equiv \forall
(\mathrm{Subtype}\, {\ensuremath{\mathrm{Bool}}}\, \lambda x . t_1) \lambda y . t_2$, where $x,y \notin FV(t_1,t_2)$. If $t_1 \supset t_2 {\ensuremath{\stackrel{*}{\Rightarrow}}}{\ensuremath{\mathbf{t}}}$ and $t_2 \supset t_1 {\ensuremath{\stackrel{*}{\Rightarrow}}}{\ensuremath{\mathbf{t}}}$, then $t_1 {\ensuremath{\stackrel{*}{\Leftrightarrow}}}t_2$.
Suppose $t_1 \supset t_2 {\ensuremath{\stackrel{*}{\Rightarrow}}}{\ensuremath{\mathbf{t}}}$ and $t_2 \supset t_1
{\ensuremath{\stackrel{*}{\Rightarrow}}}{\ensuremath{\mathbf{t}}}$. We thus have $$\forall (\mathrm{Subtype}\, {\ensuremath{\mathrm{Bool}}}\, \lambda x . t_1) \lambda y . t_2
{\ensuremath{\stackrel{*}{\Rightarrow}}}{\ensuremath{\mathbf{t}}}$$ which is only possible when $(\mathrm{Subtype}\, {\ensuremath{\mathrm{Bool}}}\, \lambda x
. t_1) {\ensuremath{\stackrel{*}{\Rightarrow}}}\tau \in {\ensuremath{{\cal T}}}$. But this in turn is only possible when $t_1 {\ensuremath{\stackrel{*}{\Rightarrow}}}t_1' \in \{{\ensuremath{\mathbf{t}}}, {\ensuremath{\mathbf{f}}}\}$. By an analogous argument $t_2 {\ensuremath{\stackrel{*}{\Rightarrow}}}t_2' \in \{{\ensuremath{\mathbf{t}}}, {\ensuremath{\mathbf{f}}}\}$. Now it is easy to see that if $t_1' \ne t_2'$ then $t_1 \supset t_2 {\ensuremath{\stackrel{*}{\Rightarrow}}}{\ensuremath{\mathbf{f}}}$ or $t_2 \supset t_1 {\ensuremath{\stackrel{*}{\Rightarrow}}}{\ensuremath{\mathbf{f}}}$.
From now on, to avoid confusion, we use $r$, $r_1$, $r_2$, etc. to denote terms of ${\ensuremath{{\cal I}}}_s$, and $t$, $t_1$, $t_2$, etc. to denote semantic terms.
\[def\_model\] We define an ${\ensuremath{{\cal I}}}_s$-structure ${\ensuremath{{\cal M}}}= \langle A, \cdot,
{\ensuremath{\llbracket\rrbracket_{}^{}}} \rangle$ as follows. As $A$ we take the set of equivalence classes of ${\ensuremath{\stackrel{*}{\Leftrightarrow}}}$ on semantic terms. We denote the equivalence class of a semantic term $t$ by $[t]$. We define $[t_1]
\cdot [t_1] = [t_1 t_2]$. This is well-defined because $t_1
{\ensuremath{\stackrel{*}{\Leftrightarrow}}}t_1'$ and $t_2 {\ensuremath{\stackrel{*}{\Leftrightarrow}}}t_2'$ imply $t_1 t_2 {\ensuremath{\stackrel{*}{\Leftrightarrow}}}t_1' t_2'$. To save on notation, we sometimes confuse $[t]$ with $t$, where it does not lead to ambiguities.
Let $v : V \to A$ be an ${\ensuremath{{\cal M}}}$-valuation. We define a function $u$ from $V$ to semantic terms by $u(x) = t$ where $t$ is an arbitrary but fixed semantic term such that $[t] = v(x)$. By $u[x/t]$ we denote a function $w$ from $V$ to semantic terms such that $w(y) =
u(y)$ for $y \ne x$, and $w(x) = t$. By induction on the structure of an ${\ensuremath{{\cal I}}}_s$-term $r$, we define a translation ${\ensuremath{\llceil r \rrceil}}^u$ from terms of ${\ensuremath{{\cal I}}}_s$ to semantic terms, parametrized by a function $u$ from $V$ to semantic terms:
- ${\ensuremath{\llceil \mathrm{Is} \rrceil}} = \mathrm{Is}$, ${\ensuremath{\llceil \mathrm{Subtype} \rrceil}} = \mathrm{Subtype}$, ${\ensuremath{\llceil \mathrm{Fun} \rrceil}} = \mathrm{Fun}$, ${\ensuremath{\llceil \forall \rrceil}} =
\forall$, ${\ensuremath{\llceil \vee \rrceil}} = \vee$, ${\ensuremath{\llceil \bot \rrceil}} = {\ensuremath{\mathbf{f}}}$, ${\ensuremath{\llceil \epsilon \rrceil}} = \mathrm{Choice}$, ${\ensuremath{\llceil \mathrm{Eq} \rrceil}} =
\mathrm{Eq}$, ${\ensuremath{\llceil \mathrm{Cond} \rrceil}} = \mathrm{Cond}$, ${\ensuremath{\llceil \mathrm{Type} \rrceil}} = \mathrm{Type}$, ${\ensuremath{\llceil \mathrm{Prop} \rrceil}} = \mathrm{Bool}$, ${\ensuremath{\llceil \iota \rrceil}} =
\bar{\iota}$ for $\iota \in {\ensuremath{{\cal T}}}_I$, ${\ensuremath{\llceil c_i^\iota \rrceil}} =
c_i^\iota$ for $c_i^\iota \in {\ensuremath{{\cal C}}}$, ${\ensuremath{\llceil d_{i,j}^\iota \rrceil}} =
d_{i,j}^\iota$ for $d_{i,j}^\iota \in {\ensuremath{{\cal D}}}$, ${\ensuremath{\llceil o_i^\iota \rrceil}} =
o_i^\iota$ for $o_i^\iota \in {\ensuremath{{\cal O}}}$,
- ${\ensuremath{\llceil x \rrceil}}^u = u(x)$ for $x \in V$,
- ${\ensuremath{\llceil r_1 r_2 \rrceil}}^u = {\ensuremath{\llceil r_1 \rrceil}}^u \cdot {\ensuremath{\llceil r_2 \rrceil}}^u$,
- ${\ensuremath{\llceil \lambda x \,.\, r \rrceil}}^u = \lambda y \,.\, {\ensuremath{\llceil r \rrceil}}^{u[x/y]}$ where $x \in V$ and $y \in V^+$ is a fresh variable.
Now the interpretation ${\ensuremath{\llbracket\rrbracket_{}^{}}}$ is defined by $${\ensuremath{\llbracketr\rrbracket_{v}^{}}} = [{\ensuremath{\llceil r \rrceil}}^u]$$ where $u$ is the function from $V$ to semantic terms, corresponding to $v$, as defined above.
\[thm\_consistent\] The system $e{\ensuremath{{\cal I}}}_s$ is consistent, i.e. $\not{\ensuremath{\vdash}}_{e{\ensuremath{{\cal I}}}_s} \bot$.
We show that ${\ensuremath{{\cal M}}}$ is an $e{\ensuremath{{\cal I}}}_s$-model. We need to check the conditions in Definition \[def\_I\_s\_model\]. Conditions (var) and (app) follow directly from the definition of ${\ensuremath{{\cal M}}}$. For condition ($\beta$) note that ${\ensuremath{\llbracket\lambda x \,.\, r\rrbracket_{v}^{}}}
\cdot [t] = [{\ensuremath{\llceil \lambda x \,.\, r \rrceil}}^u] \cdot [t] = [(\lambda y
\,.\, {\ensuremath{\llceil r \rrceil}}^{u[x/y]}) t] = [{\ensuremath{\llceil r \rrceil}}^{u[x/t]}] =
{\ensuremath{\llbracketr\rrbracket_{v[x/[t]]}^{}}}$ where $u$ is like in Definition \[def\_model\]. Condition (fv) is obvious from the definition of ${\ensuremath{\llbracket\rrbracket_{}^{}}}$. For condition ($\xi$) suppose ${\ensuremath{\llbracket\lambda x \,.\, r_1\rrbracket_{v}^{}}} \cdot [t] = {\ensuremath{\llbracket\lambda
x \,.\, r_2\rrbracket_{v}^{}}} \cdot [t]$ for every semantic term $t$. Then in particular this holds for any variable $y \in V^+$. We have ${\ensuremath{\llbracket\lambda x \,.\, r_1\rrbracket_{v}^{}}} \cdot [t] = [(\lambda y \,.\,
{\ensuremath{\llceil r_1 \rrceil}}^{u[x/y]}) y] = [{\ensuremath{\llceil r_1 \rrceil}}^{u[x/y]}]$ and ${\ensuremath{\llbracket\lambda x \,.\, r_2\rrbracket_{v}^{}}} \cdot [t] = [(\lambda y \,.\,
{\ensuremath{\llceil r_2 \rrceil}}^{u[x/y]}) y] = [{\ensuremath{\llceil r_2 \rrceil}}^{u[x/y]}]$, where $u$ is like in Definition \[def\_model\]. Hence $[{\ensuremath{\llceil r_1 \rrceil}}^{u[x/y]}] = [{\ensuremath{\llceil r_2 \rrceil}}^{u[x/y]}]$. But ${\ensuremath{\llbracket\lambda x \,.\, r_1\rrbracket_{v}^{}}} = [\lambda y \,.\,
{\ensuremath{\llceil r_1 \rrceil}}^{u[x/y]}]$ and ${\ensuremath{\llbracket\lambda x \,.\, r_2\rrbracket_{v}^{}}}
= [\lambda y \,.\, {\ensuremath{\llceil r_2 \rrceil}}^{u[x/y]}]$.
Condition (pr) follows from Lemma \[lem\_truth\_value\]. Condition (pt) follows directly from the definition of $R$. Conditions ($\forall_\top$), ($\forall_\bot$), ($\forall_e$) and ($\forall_e'$) follow from Lemma \[lem\_forall\] and Lemma \[lem\_is\]. Conditions ($\vee_1$) and ($\vee_2$) follow from Lemma \[lem\_or\_eq\_choice\]. Condition ($\supset_{t2}$) follows from Lemma \[lem\_supset\_t2\]. Condition ($\bot$) follows from Lemma \[lem\_truth\_value\] and the Church-Rosser property. Conditions ($\to_i$) and ($\to_e$) follow from Lemma \[lem\_fun\] and Lemma \[lem\_is\]. Condition ($\to_t$) follows from the definition of $R$ and the Church-Rosser propery. Conditions (s1)-(s3) follow from Lemma \[lem\_subtype\] and Lemma \[lem\_is\]. Condition (s4) follows easily from definitions and Lemma \[lem\_is\]. Conditions (o1), (o2), (d1) and (i2) follow directly from definitions. Condition (i1) follows from Lemma \[lem\_inductive\]. Conditions ($\epsilon$) and (eq) follow from Lemma \[lem\_or\_eq\_choice\]. Conditions (c1) and (c2) are obvious from the definition of $R$ and the Church-Rosser property. Condition (ef) follows from Lemma \[lem\_ext\_fun\]. Condition (eb) follows from Lemma \[lem\_ext\_bool\].
Now suppose ${\ensuremath{\vdash}}_{e{\ensuremath{{\cal I}}}_s} \bot$. Then by Theorem \[thm\_sound\] we have ${\ensuremath{{\cal M}}}\models \bot$, so ${\ensuremath{\llbracket\bot\rrbracket_{}^{}}} = [{\ensuremath{\mathbf{t}}}]$. But we have ${\ensuremath{\llbracket\bot\rrbracket_{}^{}}} = [{\ensuremath{\mathbf{f}}}]$, which implies ${\ensuremath{\mathbf{t}}}{\ensuremath{\stackrel{*}{\Leftrightarrow}}}{\ensuremath{\mathbf{f}}}$. This is impossible by the Church-Rosser property of ${\ensuremath{\Rightarrow}}$.
Complete Translation of First-Order Logic {#sec_embedding}
=========================================
In this appendix we show that the translation from Sect. \[sec\_consistent\] restricted to first-order logic is complete, i.e. we prove Theorem \[thm\_embedding\_01\]. The method of the proof is essentially the same as in [@Czajka2011], and it is a relatively simple application of the construction from the previous appendix.
First, let us precisely state the definition of the system $\mathrm{FO}$ of classical first-order logic.
- The *types* of ${{\ensuremath{{\mathrm{FO}}}}}$ are given by $${\ensuremath{{\cal T}}}\;\; ::= \;\; o \;|\; \delta \;|\; \delta \to {\ensuremath{{\cal T}}}$$
- The set of terms of ${{\ensuremath{{\mathrm{FO}}}}}$ of type $\tau$, denoted $T_\tau$, is defined as follows:
- $V_\delta \subseteq T_\delta$,
- $\Sigma_\tau \subseteq T_\tau$,
- if $t_1 \in T_{\delta\to\tau}$ and $t_2 \in T_\delta$ then $t_1 t_2 \in T_\tau$,
- $\bot \in T_o$,
- if $\varphi, \psi \in T_o$ then $\varphi \supset \psi \in
T_o$,
- if $x \in V_\delta$ and $\varphi \in T_o$ then $\forall x :
\delta \,.\, \varphi \in T_o$,
where $V_\delta$ is a countable set of variables and for each $\tau \in {\ensuremath{{\cal T}}}\setminus \{o\}$ the set $\Sigma_\tau$ is a countable set of constants. We assume that the sets $\Sigma_\tau$ are all pairwise disjoint, and disjoint with $V_\delta$. Terms of type $o$ are *formulas*. As usual, we omit spurious brackets and assume that application associates to the left. We identify $\alpha$-equivalent terms, i.e. terms differing only in the names of bound variables are considered identical.
- The system ${{\ensuremath{{\mathrm{FO}}}}}$ is given by the following rules and axioms, where $\Delta$ is a finite set of formulas, $\varphi, \psi$ are formulas. The notation $\Delta, \varphi$ is a shorthand for $\Delta \cup \{\varphi\}$.
[**Axioms**]{}
- $\Delta, \varphi {\ensuremath{\vdash}}\varphi$
- $\Delta {\ensuremath{\vdash}}((\varphi \supset \bot) \supset \bot)
\supset \varphi$
[**Rules**]{}
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ -------------------------------------------------------------------------------------------------------------------------------------------
$ $
{\supset_i^{{{\ensuremath{{\mathrm{FO}}}}}}:}\; \inferrule{\Delta, \varphi {\ensuremath{\vdash}}\psi}{\Delta {\ensuremath{\vdash}}\varphi \supset \psi} {\supset_e^{{{\ensuremath{{\mathrm{FO}}}}}}:}\; \inferrule{\Delta {\ensuremath{\vdash}}\varphi \supset
$ \psi \;\;\; \Delta {\ensuremath{\vdash}}\varphi}{\Delta {\ensuremath{\vdash}}\psi}
$
$ $
{\forall_i^{{{\ensuremath{{\mathrm{FO}}}}}}:}\; \inferrule{\Delta {\ensuremath{\vdash}}\varphi}{\Delta {\ensuremath{\vdash}}\forall x : \delta \,.\, \varphi} \; x \notin FV(\Delta) {\forall_e^{{{\ensuremath{{\mathrm{FO}}}}}}:}\; \inferrule{\Delta {\ensuremath{\vdash}}\forall x : \delta \,.\, \varphi}{\Delta
$ {\ensuremath{\vdash}}\varphi[x/t]}\; t \in T_\delta
$
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ -------------------------------------------------------------------------------------------------------------------------------------------
Note that the system ${{\ensuremath{{\mathrm{FO}}}}}$ is a restriction of .
A first-order structure ${\ensuremath{{\cal A}}}$ is a pair $\langle A, \{ f_c \;|\; c
\in \Sigma_\tau, \tau \in {\ensuremath{{\cal T}}}\} \rangle$ where $A$ is the non-empty universe and $f_c$ are interpretations of constants (functions or relations on $A$ of appropriate arity, or elements of $A$). By ${\ensuremath{\llbracketc\rrbracket_{}^{{\ensuremath{{\cal A}}}}}}$ we denote the interpretation of $c$ in ${\ensuremath{{\cal A}}}$. A first-order valuation $u$ is a function from variables to the universe of a first-order structure. We define the relations of satisfaction ${\ensuremath{{\cal A}}}\models_{{\ensuremath{{\mathrm{FO}}}}}\varphi$ and semantic consequence $\Delta \models_{{\ensuremath{{\mathrm{FO}}}}}\varphi$ in an obvious way.
The following is a well-known result from elementary logic.
$\Delta {\ensuremath{\vdash}}_{{\ensuremath{{\mathrm{FO}}}}}\varphi$ iff $\Delta \models_{{\ensuremath{{\mathrm{FO}}}}}\varphi$
The mappings ${\ensuremath{\lceil-\rceil}}$ and $\Gamma(-)$ are defined exactly as in Sect. \[sec\_consistent\], restricting to terms of ${{\ensuremath{{\mathrm{FO}}}}}$. Recall that we assume $V_\delta \subseteq V$ and $\Sigma_\tau \subseteq \Sigma_s$ for $\tau \in {\ensuremath{{\cal T}}}$.
If $\Delta {\ensuremath{\vdash}}_{{\ensuremath{{\mathrm{FO}}}}}\varphi$ then $\Gamma(\Delta, \varphi),
{\ensuremath{\lceil\Delta\rceil}} {\ensuremath{\vdash}}_{{\ensuremath{{\cal I}}}_s} {\ensuremath{\lceil\varphi\rceil}}$.
This is a special case of Theorem \[thm\_sound\_hol\].
In this section we use $s$, $s_1$, $s_2$, etc. for terms of ${\ensuremath{{\cal I}}}_s$. We assume that there is a fresh constant $\delta \in \Sigma_s$. This constant will represent the first-order universe. We use the notation $\tau^n\to\tau$ for $\tau\to\ldots\to\tau$ where $\tau$ occurs $n+1$ times, where $n > 0$.
If $\Gamma(\Delta, \varphi), {\ensuremath{\lceil\Delta\rceil}} {\ensuremath{\vdash}}_{e{\ensuremath{{\cal I}}}_s}
{\ensuremath{\lceil\varphi\rceil}}$ then $\Delta {\ensuremath{\vdash}}_{{\ensuremath{{\mathrm{FO}}}}}\varphi$.
Suppose $\Gamma(\Delta, \varphi), {\ensuremath{\lceil\Delta\rceil}} {\ensuremath{\vdash}}_{e{\ensuremath{{\cal I}}}_s}
{\ensuremath{\lceil\varphi\rceil}}$ but $\Delta \not{\ensuremath{\vdash}}_{{\ensuremath{{\mathrm{FO}}}}}\varphi$. Then $\Delta
\not\models_{{\ensuremath{{\mathrm{FO}}}}}\varphi$, so there exist a first-order structure ${\ensuremath{{\cal A}}}$ and a first-order valuation $u$ such that ${\ensuremath{{\cal A}}}, u
\models_{{\ensuremath{{\mathrm{FO}}}}}\Delta$, but ${\ensuremath{{\cal A}}}, u \not\models_{{\ensuremath{{\mathrm{FO}}}}}\varphi$.
We use the construction of Definition \[def\_model\] to transform ${\ensuremath{{\cal A}}}$ into an $e{\ensuremath{{\cal I}}}_s$-model ${\ensuremath{{\cal M}}}$, by taking the set of constants $\delta$ of ${\ensuremath{{\cal M}}}$ to consist of the elements of the universe of ${\ensuremath{{\cal A}}}$. We also need to extend the definition of the translation ${\ensuremath{\llceil - \rrceil}}$ from Definition \[def\_model\] to interpret the new constants that we added to the language of $e{\ensuremath{{\cal I}}}_s$. We set ${\ensuremath{\llceil \delta \rrceil}} = \delta$. If $c \in
\Sigma_{\delta}$ then we set ${\ensuremath{\llceil c \rrceil}} =
{\ensuremath{\llbracketc\rrbracket_{}^{{\ensuremath{{\cal A}}}}}}$. If $f \in \Sigma_{\delta^n\to\delta}$ then let $c_f \in \delta^{(\delta,\ldots,\delta)}$ be such that ${\ensuremath{{\cal F}}}(c_f)(a_1)(a_2)\ldots(a_n) = {\ensuremath{\llbracketf\rrbracket_{}^{{\ensuremath{{\cal A}}}}}}(a_1, \ldots,
a_n)$ for any $a_1,\ldots,a_n \in \delta$. We set ${\ensuremath{\llceil f \rrceil}} =
c_f$. Similarly, if $r \in \Sigma_{\delta^n\to o}$ then let $c_r \in
{\ensuremath{\mathrm{Bool}}}^{(\delta,\ldots,\delta)}$ be such that for all $a_1,\ldots,a_n
\in \delta$ we have ${\ensuremath{{\cal F}}}(c_r)(a_1)(a_2)\ldots(a_n) = {\ensuremath{\mathbf{t}}}$ iff ${\ensuremath{\llbracketr\rrbracket_{}^{{\ensuremath{{\cal A}}}}}}(a_1, \ldots, a_n)$ holds.
Note that in ${\ensuremath{{\cal M}}}$ we have $\delta = \{ a \in {\ensuremath{{\cal M}}}\;|\; \mathrm{Is}
\cdot a \cdot \delta = \top \}$. This follows directly from Lemma \[lem\_is\] and the fact that for $a \in \delta$ we have $a
\notin \tau_2^{\tau}$, and hence if $t {\ensuremath{\vartriangleright}}a$ then $t {\ensuremath{\stackrel{*}{\Leftrightarrow}}}a$. Note also that any first-order ${\ensuremath{{\cal A}}}$-valuation $v$ is also an ${\ensuremath{{\cal M}}}$-valuation, if for a variable $x$ we interpret $v(x)$ as an element of the set $\delta$ in ${\ensuremath{{\cal M}}}$.
Now it is easy to show by induction on the structure of a term $t
\in T_\tau$ that ${\ensuremath{\llbracket{\ensuremath{\lceilt\rceil}}\rrbracket_{v}^{{\ensuremath{{\cal M}}}}}} =
{\ensuremath{\llbrackett\rrbracket_{v}^{{\ensuremath{{\cal A}}}}}} \in \delta$ for any first order-valuation $v$, where $\tau \in {\ensuremath{{\cal T}}}\setminus \{o\}$. Using this we verify by straightforward induction on the structure of a first-order formula $\varphi$ that for any first-order valuation $v$ we have ${\ensuremath{\llbracket{\ensuremath{\lceil\varphi\rceil}}\rrbracket_{v}^{{\ensuremath{{\cal M}}}}}} = {\ensuremath{\mathbf{t}}}$ iff ${\ensuremath{{\cal A}}}, v \models
\varphi$, and ${\ensuremath{\llbracket{\ensuremath{\lceil\varphi\rceil}}\rrbracket_{v}^{{\ensuremath{{\cal M}}}}}} = {\ensuremath{\mathbf{f}}}$ iff ${\ensuremath{{\cal A}}}, v \not\models \varphi$. For instance, suppose $\varphi \equiv
\forall x : \delta \,.\, \psi$. Then ${\ensuremath{\lceil\varphi\rceil}} = \forall x :
\delta \,.\, {\ensuremath{\lceil\psi\rceil}}$. Since ${\ensuremath{\llbracket\delta :
{\ensuremath{\mathrm{Type}}}\rrbracket_{v}^{{\ensuremath{{\cal M}}}}}} = {\ensuremath{\mathbf{t}}}$, and $a \in {\ensuremath{{\cal A}}}$ iff $\mathrm{Is} \cdot a
\cdot \delta = {\ensuremath{\mathbf{t}}}$, we have ${\ensuremath{\llbracket{\ensuremath{\lceil\varphi\rceil}}\rrbracket_{v}^{{\ensuremath{{\cal M}}}}}}
= {\ensuremath{\mathbf{t}}}$ iff ${\ensuremath{\llbracket{\ensuremath{\lceil\psi\rceil}}\rrbracket_{v[x/a]}^{{\ensuremath{{\cal M}}}}}} = {\ensuremath{\mathbf{t}}}$ for every $a \in {\ensuremath{{\cal A}}}$ iff ${\ensuremath{{\cal A}}}, v[x/a] \models \psi$ for every $a \in {\ensuremath{{\cal A}}}$ iff ${\ensuremath{{\cal A}}}, v \models \forall x : \delta \,.\, \psi \equiv \varphi$, where we use the inductive hypothesis in the second equivalence.
It is easy to check that ${\ensuremath{{\cal M}}}, u \models_{e{\ensuremath{{\cal I}}}_s} \Gamma(\Delta,
\varphi)$. Hence ${\ensuremath{{\cal M}}}, u \models_{e{\ensuremath{{\cal I}}}_s} \Gamma(\Delta, \varphi),
{\ensuremath{\lceil\Delta\rceil}}$, because ${\ensuremath{{\cal A}}}, u \models_{{\ensuremath{{\mathrm{FO}}}}}\Delta$. Since ${\ensuremath{{\cal A}}}, u
\not\models_{{\ensuremath{{\mathrm{FO}}}}}\varphi$, we also have ${\ensuremath{{\cal M}}}, u \not\models_{e{\ensuremath{{\cal I}}}_s}
{\ensuremath{\lceil\varphi\rceil}}$. But by Theorem \[thm\_sound\] and $\Gamma(\Delta, \varphi), {\ensuremath{\lceil\Delta\rceil}} {\ensuremath{\vdash}}_{e{\ensuremath{{\cal I}}}_s}
{\ensuremath{\lceil\varphi\rceil}}$ we have ${\ensuremath{{\cal M}}}, u \models_{e{\ensuremath{{\cal I}}}_s}
{\ensuremath{\lceil\varphi\rceil}}$. Contradiction.
[^1]: So e.g. in the axiom $\beta$ the free variables of $t_2$ do not become bound in $t_1[x/t_2]$.
[^2]: To be more precise, for every possible substitution of terms for the free variables of $\Gamma,t$ we perform this substitution on $\Gamma,t$, denoting the result by $\Gamma^*,t^*$.
[^3]: By defining (see the next subsection) $\varphi = \neg \varphi$ one could then easily derive $\bot$ using the rule $\vee_e$ applied to $\varphi \vee \neg\varphi$.
[^4]: We mean this in a precise sense. This follows from our model construction.
[^5]: By soundness we mean that if the algorithm declares $t : \alpha$ correct (in a context $\Gamma$), then $\Gamma
{\ensuremath{\vdash}}_{{\ensuremath{{\cal I}}}_s} t : \alpha$. Completeness is the implication in the other direction. The point is that one would want some standard type-checking algorithms to be modified to work on terms of ${\ensuremath{{\cal I}}}_s$, by declaring incorrect all type assertions not conforming to the syntax of (a straightforward translation of) the assertions handled by the algorithm.
[^6]: Substitution is defined for semantic terms in an obvious way, avoiding variable capture.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Low-rank matrix completion is a problem of immense practical importance. Recent works on the subject often use nuclear norm as a convex surrogate of the rank function. Despite its solid theoretical foundation, the convex version of the problem often fails to work satisfactorily in real-life applications. Real data often suffer from very few observations, with support not meeting the randomness requirements, ubiquitous presence of noise and potentially gross corruptions, sometimes with these simultaneously occurring. This paper proposes a Proximal Alternating Robust Subspace Minimization (PARSuMi) method to tackle the three problems. The proximal alternating scheme explicitly exploits the rank constraint on the completed matrix and uses the $\ell_0$ pseudo-norm directly in the corruption recovery step. We show that the proposed method for the non-convex and non-smooth model converges to a stationary point. Although it is not guaranteed to find the global optimal solution, in practice we find that our algorithm can typically arrive at a good local minimizer when it is supplied with a reasonably good starting point based on convex optimization. Extensive experiments with challenging synthetic and real data demonstrate that our algorithm succeeds in a much larger range of practical problems where convex optimization fails, and it also outperforms various state-of-the-art algorithms.'
author:
- 'Yu-Xiang Wang'
- Choon Meng Lee
- 'Loong-Fah Cheong'
- 'Kim-Chuan Toh'
bibliography:
- 'ARSuMi.bib'
date: 'Received: 5 September 2013 / Accepted: 26 June 2014'
title: Practical Matrix Completion and Corruption Recovery using Proximal Alternating Robust Subspace Minimization
---
Introduction {#intro}
============
Completing a low-rank matrix from partially observed entries, also known as matrix completion, is a central task in many real-life applications. The same abstraction of this problem has appeared in diverse fields such as signal processing, communications, information retrieval, machine learning and computer vision. For instance, the missing data to be filled in may correspond to plausible movie recommendations [@koren2009MF; @funk2006netflix], occluded feature trajectories for rigid or non-rigid structure from motion, namely SfM [@hartley2003powerfactorization; @Damped_Newton_2005] and NRSfM [@Paladini_NRSfM], relative distances of wireless sensors [@oh2010sensor], pieces of uncollected measurements in DNA micro-array [@DNA06], just to name a few.
![ Sampling pattern of the Dinosaur sequence: 316 features are tracked over 36 frames. Dark area represents locations where no data is available; sparse highlights are injected gross corruptions. Middle stripe in grey are noisy observed data, occupying 23% of the full matrix. The task of this paper is to fill in the missing data and recover the corruptions. []{data-label="fig:DinosaurSampling"}](pics/data_matrix.png){width="0.9\linewidth" height="0.5\linewidth"}
The common difficulty of these applications lies in the scarcity of the observed data, uneven distribution of the support, noise, and more often than not, the presence of gross corruptions in some observed entries. For instance, in the movie rating database Netflix [@netflix], only less than 1% of the entries are observed and 90% of the observed entries correspond to 10% of the most popular movies. In photometric stereo, the missing data and corruptions (arising from shadow and specular highlight as modeled in @Wu_photometric) form contiguous blocks in images and are by no means random. In structure from motion, the observations fall into a diagonal band shape, and feature coordinates are often contaminated by tracking errors (see the illustration in Fig. \[fig:DinosaurSampling\]). Therefore, in order for any matrix completion algorithm to work in practice, these aforementioned difficulties need to be tackled altogether. We refer to this problem as **practical matrix completion**. Mathematically, the problem to be solved is the following: $$\boxed{
\begin{aligned}
& \underset{}{\text{Given}}
& & \Omega, \widehat{W}_{ij}\text{ for all }(i,j)\in\Omega, \\
& \underset{}{\text{find}}
& & W, \tilde{\Omega},\\
& \text{s.t.}
& & \mathrm{rank}(W)\text{ is small};\; \mathrm{card}(\tilde{\Omega})\text{ is small};\\
& & & |W_{ij}-\widehat{W}_{ij}| \text{ is small } \forall(i,j)\in\Omega|\tilde{\Omega}.
\end{aligned}
}$$ where $\Omega$ is the index set of observed entries whose locations are not necessarily selected at random, $\tilde{\Omega}\in\Omega$ represents the index set of corrupted data, $\widehat{W}\in \mathbb{R}^{m\times n}$ is the measurement matrix with only $\widehat{W}_{ij\in{\Omega}}$ known, i.e., its support is contained in $\Omega$. Furthermore, we define the projection $\mathcal{P}_{\Omega}:\mathbb{R}^{m\times n}\mapsto \mathbb{R}^{|\Omega|}$ so that $\mathcal{P}_{\Omega}(\widehat{W})$ denotes the vector of observed data. The adjoint of $\mathcal{P}_\Omega$ is denoted by $\mathcal{P}_\Omega^*$.
Extensive theories and algorithms have been developed to tackle some aspect of the challenges listed in the preceding paragraph, but those tackling the full set of challenges are far and few between, thus resulting in a dearth of practical algorithms. Two dominant classes of approaches are nuclear norm minimization, e.g. @Candes_ExactMC [@Candes2011_JACM; @CandesNoise; @chen2011erasures], and matrix factorization, e.g., @koren2009MF [@Damped_Newton_2005; @WibergL2; @Subspace_ChenPei_2008; @WibergL1]. Nuclear norm minimization methods minimize the convex relaxation of rank instead of the rank itself, and are supported by rigorous theoretical analysis and efficient numerical computation. However, the conditions under which they succeed are often too restrictive for it to work well in real-life applications (as reported in @shi2011limitations and @jain2012alt_min_global). In contrast, matrix factorization is widely used in practice and are considered very effective for problems such as movie recommendation [@koren2009MF] and structure from motion [@Kanade_Factorization; @Paladini_NRSfM] despite its lack of rigorous theoretical foundation. Indeed, as one factorizes matrix $W$ into $UV^T$, the formulation becomes bilinear and thus optimal solution is hard to obtain except in very specific cases (e.g., in @jain2012alt_min_global). A more comprehensive survey of the algorithms and review of the strengths and weaknesses will be given in the next section.
In this paper, we attempt to solve the practical matrix completion problem under the prevalent case where the rank of the matrix $W$ and the cardinality of $\tilde{\Omega}$ are upper bounded by some known parameters $r$ and $N_0$ via the following non-convex, non-smooth optimization model: $$\begin{aligned}
\label{eq:MC_Formulation_0}
\begin{array}{rl}
\underset{W,E}{\text{min}}\;
& \frac{1}{2}\|\mathcal{P}_{\Omega}(W-\widehat{W}+E)\|^2 + \frac{\epsilon}{2}
\norm{\mathcal{P}_{\overline{\Omega}}(W)}^2
\\[5pt]
\text{s.t.} \;
& \mathrm{rank}(W)\leq r, \; W\in \R^{m\times n}
\\[5pt]
& \|E\|_0\leq N_0, \; \norm{E} \leq K_E, \; E\in \R^{m\times n}_\Omega
\end{array}\end{aligned}$$ where $\R^{m\times n}_\Omega$ denotes the set of $m\times n$ matrices whose supports are subsets of $\Omega$ and $\norm{\cdot}$ is the Frobenius norm; $K_E$ is a finite constant introduced to facilitate the convergence proof. Note that the restriction of $E$ to $\R^{m\times n}_\Omega$ is natural since the role of $E$ is to capture the gross corruptions in the observed data $\widehat{W}_{ij\in\Omega}$. The bound constraint on $\norm{E}$ is natural in some problems when the true matrix $W$ is bounded (e.g., Given the typical movie ratings of 0-10, the gross outliers can only lie in \[-10, 10\]). In other problems, we simply choose $K_E$ to be some large multiple (say 20) of $\sqrt{N_0}\times {\rm median}(\mathcal{P}_\Omega(\widehat{W}))$, so that the constraint is essentially inactive and has no impact on the optimization. Note that without making any randomness assumption on the index set $\Omega$ or assuming that the problem has a unique solution $(W^*,E^*)$ such that the singular vector matrices of $W^*$ satisfy some inherent conditions like those in @Candes2011_JACM, the problem of practical matrix completion is generally ill-posed. This motivated us to include the Tikhonov regularization term $\frac{\epsilon}{2}
\norm{\mathcal{P}_{\overline{\Omega}}(W)}^2$ in , where $\overline{\Omega}$ denotes the complement of $\Omega$, and $0<\epsilon<1 $ is a small constant. Roughly speaking, what the regularization term does is to pick the solution $W$ which has the smallest $\norm{\mathcal{P}_{\overline{\Omega}}(W)}$ among all the candidates in the optimal solution set of the non-regularized problem. Notice that we only put a regularization on those elements of $W$ in $\overline{\Omega}$ as we do not wish to perturb those elements of $W$ in the fitting term. Finally, with the Tikhonov regularization and the bound constraint on $\norm{E}$, we can show that problem has a global minimizer.
By defining $H\in\R^{m\times n}$ to be the matrix such that $$\begin{aligned}
H_{ij} = \left\{ \begin{array}{ll}
1 & \mbox{if $(i,j)\in \Omega$} \\[5pt]
\sqrt{\epsilon} & \mbox{if $(i,j)\not\in \Omega$},
\end{array} \right.
\label{eq:H}\end{aligned}$$ and those elements of $E$ and $\widehat{W}$ in $\overline{\Omega}$ to be zero, we can rewrite the objective function in in a compact form, and the problem becomes: $$\begin{aligned}
\label{eq:MC_Formulation}
\begin{array}{rl}
\underset{W,E}{\text{min}}\;
& \frac{1}{2}\|H\circ(W+E-\widehat{W})\|^2
\\[5pt]
\text{s.t.}\;
& \mathrm{rank}(W)\leq r, \; W\in \R^{m\times n}
\\[5pt]
& \|E\|_0\leq N_0, \; \norm{E} \leq K_E, \; E\in \R^{m\times n}_\Omega.
\end{array}\end{aligned}$$ In the above, the notation “$\circ$" denotes the element-wise product between two matrices.
We propose PARSuMi, a proximal alternating minimization algorithm motivated by the algorithm in @attouch2010proximal to solve . This involves solving two subproblems each with an auxiliary proximal regularization term. It is important to emphasize that the subproblems in our case are non-convex and hence it is essential to design appropriate algorithms to solve the subproblems to global optimality, at least empirically. We develop essential reformulations of the subproblems and design novel techniques to efficiently solve each subproblem, provably achieving the global optimum for one, and empirically so for the other. We also prove that our algorithm is guaranteed to converge to a limit point, which is necessarily a stationary point of (\[eq:MC\_Formulation\]). We emphasize here that the convergence is established even though one of the subproblems may not be solved to global optimality. Together with the initialization schemes we have designed based on the convex relaxation of (\[eq:MC\_Formulation\]), our method is able to solve challenging real matrix completion problems with corruptions robustly and accurately. As we demonstrate in the experiments, PARSuMi is able to provide excellent reconstruction of unobserved feature trajectories in the classic Oxford Dinosaur sequence for SfM, despite structured (as opposed to random) observation pattern and data corruptions. It is also able to solve photometric stereo to high precision despite severe violations of the Lambertian model (which underlies the rank-3 factorization) due to shadow, highlight and facial expression difference. Compared to state-of-the-art methods such as GRASTA [@he2011grasta], Wiberg $\ell_1$ [@WibergL1] and BALM [@DelBue2012balm], our results are substantially better both qualitatively and quantitatively.
Note that in (\[eq:MC\_Formulation\]) we do not seek convex relaxation of any form, but rather constrain the rank and the corrupted entries’ cardinality directly in their original forms. While it is generally not possible to have an algorithm guaranteed to compute the global optimal solution, we demonstrate that with appropriate initializations, the faithful representation of the original problem often offers significant advantage over the convex relaxation approach in denoising and corruption recovery, and is thus more successful in solving real problems.
The rest of the paper is organized as follows. In Section \[sec:PrevWorks\], we provide a comprehensive review of the existing theories and algorithms for practical matrix completion, summarizing the strengths and weaknesses of nuclear norm minimization and matrix factorization. In Section \[sec:numerical\_MF\], we conduct numerical evaluations of predominant matrix factorization methods, revealing those algorithms that are less-likely to be trapped at local minima. Specifically, these features include parameterization on a subspace and second-order Newton-like iterations. Building upon these findings, we develop the PARSuMi scheme in Section \[sec:PARSuMi\] to simultaneously handle sparse corruptions, dense noise and missing data. The proof of convergence and a convex initialization scheme are also provided in this section. In Section \[sec:experiment\], the proposed method is evaluated on both synthetic and real data and is shown to outperform the current state-of-the-art algorithms for robust matrix completion.
A survey of results {#sec:PrevWorks}
===================
Matrix completion and corruption recovery via nuclear norm minimization {#sec:MC_theory}
-----------------------------------------------------------------------
MC [@Candes_ExactMC] RPCA [@Candes2011_JACM] NoisyMC [@CandesNoise] StableRPCA [@RPCA2010stable] RMC [@li2013compressed] RMC [@chen2011erasures]
-------------------------------- ---------------------- ------------------------- ------------------------ ------------------------------ ------------------------- -------------------------
Missing data Yes Yes Yes No Yes Yes
Corruptions No Yes No Yes Yes Yes
Noise No No Yes Yes No No
Deterministic $\Omega$ No No No No No Yes
Deterministic $\tilde{\Omega}$ No No No No No Yes
Recently, the most prominent approach for solving a matrix completion problem is via the following nuclear norm minimization: $$\label{eq:MC_nuc}
\underset{W}{\text{min}}\, \left\{
\|W\|_* \, \middle|\,
\mathcal{P}_{\Omega}(W-\widehat{W})=0 \right\},$$ in which $\mathrm{rank}(X)$ is replaced by the nuclear norm $\|X\|_*=\sum_i\sigma_i(X)$, where the latter is the tightest convex relaxation of $\mathrm{rank}$ over the unit (spectral norm) ball. @Candes_ExactMC showed that when sampling is uniformly random and sufficiently dense, and the underlying low-rank subspace is *incoherent* with respect to the standard bases, then the remaining entries of the matrix can be exactly recovered. The guarantee was later improved in @candes2010optimal [@recht2009simpler], and extended for noisy data in @CandesNoise [@negahban2012restricted] relaxed the equality constraint to $$\|\mathcal{P}_{\Omega}(W-\widehat{W})\|\leq\delta.$$ Using similar assumptions and arguments, @Candes2011_JACM and @Parrilo_RPCA concurrently proposed solution to the related problem of robust principal component analysis (RPCA) where the low-rank matrix can be recovered from sparse corruptions (with no missing data[^1]). This is formulated as $$\label{eq:RPCA_nuc}
\underset{W,E}{\text{min}}\, \left\{
\|W\|_*+\lambda\|E\|_1 \, \middle|\,
W+E=\widehat{W} \right\}.$$ Noisy extension and improvement of the guarantee for RPCA were provided by @RPCA2010stable and @ganesh2010dense respectively. @chen2011erasures and @li2013compressed combined and and provided guarantee for the following $$\label{eq:RMC_nuc}
\underset{W,E}{\text{min}} \, \left\{
\|W\|_*+\lambda\|E\|_1\,\middle|\,
\mathcal{P}_{\Omega}(W+E-\widehat{W})=0
\right\}.$$ In particular, the results in @chen2011erasures lifted the uniform random support assumptions in previous works by laying out the exact recovery condition for a class of deterministic sampling ($\Omega$) and corruptions ($\tilde{\Omega}$) patterns. We summarize the theoretical and algorithmic progress in practical matrix completion achieved by each method in Table \[tab:MC\_theory\]. It appears that researchers are moving towards analyzing all possible combinations of the problems; from past indication, it seems entirely plausible albeit tedious to show that the noisy extension $$\label{eq:RMCN_nuc}
\underset{W,E}{\text{min}} \left\{
\|W\|_*+\lambda\|E\|_1\,\middle|\,
\|\mathcal{P}_{\Omega}(W+E-\widehat{W})\|\leq \delta
\right\}$$ will return a solution stable around the desired $W$ and $E$ under appropriate assumptions. Wouldn’t that solve the practical matrix completion problem altogether?
The answer is unfortunately no. While this line of research have provided profound understanding of practical matrix completion itself, the actual performance of the convex surrogate on real problems (e.g., movie recommendation) is usually not competitive against nonconvex approaches such as matrix factorization. Although convex relaxation is amazingly equivalent to the original problem under certain conditions, those well versed in practical problems will know that those theoretical conditions are usually not satisfied by real data. Due to noise and model errors, real data are seldom truly low-rank (see the comments on Jester joke dataset in @keshavan2009comparison), nor are they as incoherent as randomly generated data. More importantly, observations are often structured (e.g., diagonal band shape in SfM) and hence do not satisfy the random sampling assumption needed for the tight convex relaxation approach. As a consequence of all these factors, the recovered $W$ and $E$ by convex optimization are often neither low-rank nor sparse in practical matrix completion. This can be further explained by the so-called “Robin Hood” attribute of $\ell_1$ norm (analogously, nuclear norm is the $\ell_1$ norm in the spectral domain), that is, it tends to steal from the rich and give it to the poor, decreasing the inequity of “wealth” distribution. Illustrations of the attribute will be given in Section \[sec:experiment\].
Nevertheless, the convex relaxation approach has the advantage that one can design *efficient* algorithms to find or approximately reach the *global* optimal solution of the given convex formulation. In this paper, we take advantage of the convex relaxation approach and use it to provide a powerful initialization for our algorithm to converge to the correct solution.
Matrix factorization and applications
-------------------------------------
Another widely-used method to estimate missing data in a low-rank matrix is matrix factorization (MF). It is at first considered as a special case of the weighted low-rank approximation problem with $\{0,1\}$ weight by @gabriel1979lower in [-@gabriel1979lower] and much later by @srebro2003weighted. The buzz of Netflix Prize further popularizes the missing data problem as a standalone topic of research. Matrix factorization turns out to be a robust and efficient realization of the idea that people’s preferences of movies are influenced by a small number of latent factors and has been used as a key component in almost all top-performing recommendation systems [@koren2009MF] including BellKor’s Pragmatic Chaos, the winner of the Netflix Prize [@koren2009bellkor].
In computer vision, matrix factorization with missing data is recognized as an important problem too. Tomasi-Kanade affine factorization [@Kanade_Factorization], Sturm-Triggs projective factorization [@sturm1996factorization], and many techniques in Non-Rigid SfM and motion tracking [@Paladini_NRSfM] can all be formulated as a matrix factorization problem. Missing data and corruptions emerge naturally due to occlusions and tracking errors. For a more exhaustive survey of computer vision problems that can be modelled by matrix factorization, we refer readers to @DelBue2012balm.
Regardless of its applications, the key idea is that when $W=UV^T$, one ensures that the required rank constraint is satisfied by restricting the factors $U$ and $V$ to be in $\R^{m\times r}$ and $\R^{n\times r}$ respectively. Since the $(U,V)$ parameterization has a much smaller degree of freedom than the dimension of $W$, completing the missing data becomes a better posed problem. This gives rise to the following optimization problem: $$\label{eq:MC_L2}
\begin{aligned}
& \underset{U,V}{\text{min}}
& & \frac{1}{2}\left \Vert \mathcal{P}_{\Omega}(UV^T - \widehat{W}) \right \Vert^2
\end{aligned}$$ or its equivalence reformulation $$\label{eq:MC_L22}
\underset{U}{\text{min}}\,
\left\{ \frac{1}{2}\left \Vert \mathcal{P}_{\Omega}(UV(U)^T - \widehat{W}) \right \Vert^2
\middle| U^T U = I_r \right\}$$ where the factor $V$ is now a function of $U$.
Unfortunately, is not a convex optimization problem. The quality of the solutions one may get by minimizing this objective function depends on specific algorithms and their initializations. Roughly speaking, the various algorithms for may be grouped into three categories: **alternating minimization**, **first order** gradient methods and **second order** Newton-like methods.
Simple approaches like alternating least squares (ALS) or equivalently PowerFactorization [@hartley2003powerfactorization] fall into the first category. They alternatingly fix one factor and minimize the objective over the other using least squares method. A more sophisticated algorithm is BALM [@DelBue2012balm], which uses the Augmented Lagrange Multiplier method to gradually impose additional problem-specific manifold constraints. The inner loop however is still alternating minimization. This category of methods has the reputation of reducing the objective value quickly in the first few iterations, but they usually take a large number of iterations to converge to a high quality solution [@Damped_Newton_2005].
First order gradient methods are efficient, easy to implement and they are able to scale up to million-by-million matrices if stochastic gradient descent is adopted. Therefore it is very popular for large-scale recommendation systems. Typical approaches include Simon Funk’s incremental SVD [@funk2006netflix], nonlinear conjugate gradient [@srebro2003weighted] and more sophisticatedly, gradient descent on the Grassmannian/Stiefel manifold, such as GROUSE [@balzano2010grouse] and OptManifold [@yin2013orthogonality]. These methods, however, as we will demonstrate later, easily get stuck in local minima[^2].
The best performing class of methods are the second order Newton-like algorithms, in that they demonstrate superior performance in both accuracy and the speed of convergence (though each iteration requires more computation); hence they are suitable for small to medium scale problems requiring high accuracy solutions (e.g., SfM and photometric stereo in computer vision). Representatives of these algorithms include the damped Newton method [@Damped_Newton_2005], Wiberg($\ell_2$) [@WibergL2], LM\_S and LM\_M of @Subspace_ChenPei_2008 and LM\_GN, which is a variant of LM\_M using Gauss-Newton (GN) to approximate the Hessian function.
As these methods are of special importance in developing our PARSuMi algorithm, we conduct extensive numerical evaluations of these algorithms in Section \[sec:numerical\_MF\] to understand their pros and cons as well as the key factors that lead to some of them finding global optimal solutions more often than others.
It is worthwhile to note some delightful recent efforts to scale the first two classes of MF methods to internet scale, e.g., parallel coordinate descent extension for ALS [@yu2012scalable] and stochastic gradient methods in “Hogwild” [@Recht2011Hogwild]. It will be an interesting area of research to see if the ideas in these papers can be used to make the second order methods more scalable.
In addition, there are a few other works in each category that take into account the corruption problem by changing the quadratic penalty term of into $\ell_1$-norm or Huber function $$\label{eq:MC_L1}
\begin{aligned}
& \underset{U,V}{\text{min}}
& & \left \Vert \mathcal{P}_{\Omega}(UV^T - \widehat{W}) \right \Vert_1\, ,
\end{aligned}$$ $$\label{eq:MC_huber}
\begin{aligned}
& \underset{U,V}{\text{min}}
& &
\sum_{(ij)\in\Omega} \mbox{Huber}\big((UV^T - \widehat{W})_{ij} \big).
\end{aligned}$$ Notable algorithms to solve these formulations include alternating linear programming (ALP) and alternating quadratic programming (AQP) in @ke2005L1, GRASTA [@he2011grasta] that extends GROUSE, as well as Wiberg $\ell_1$ [@WibergL1] that uses a second order Wiberg-like iteration. While it is well known that the $\ell_1$-norm or Huber penalty term can better handle outliers, and the models (\[eq:MC\_L1\]) and (\[eq:MC\_huber\]) are seen to be effective in some problems, there is not much reason for a “convex” relaxation of the $\ell_0$ pseudo-norm[^3], since the rank constraint imposed by matrix factorization is already highly non-convex. Empirically, we find that $\ell_1$-norm penalty offers poor denoising ability to dense noise and also suffers from “Robin Hood” attribute. Comparison with this class of methods will be given later in Section \[sec:experiment\], which shows that our method can better handle noise and corruptions.
The practical advantage of $\ell_0$ over $\ell_1$ penalty is well illustrated in @DRMF, where @DRMF proposed an $\ell_0$-based robust matrix factorization method which deals with corruptions and a given rank constraint. Our work is similar to @DRMF in that we both eschew the convex surrogate $\ell_1$-norm in favor of using the $\ell_0$-norm directly. However, our approach treats both corruptions and missing data. More importantly, our treatment of the problem is different and it results in a convergence guarantee that covers the algorithm of @DRMF as a special case; this will be further explained in Section \[sec:PARSuMi\].
{width="0.63\linewidth"} {width="0.35\linewidth"}\
Emerging theory for matrix factorization
----------------------------------------
As we mentioned earlier, a fundamental drawback of matrix factorization methods for low rank matrix completion is the lack of proper theoretical foundation. However, thanks to the better understanding of low-rank structures nowadays, some theoretical analysis of this problem slowly emerges. This class of methods are essentially designed for solving noisy matrix completion problem with an explicit rank constraint, i.e., $$\label{eq:MC_rank_constraint}
\underset{W}{\text{min}}
\left\{
\frac{1}{2}\left \Vert \mathcal{P}_{\Omega}(W - \widehat{W}) \right \Vert^2
\,\middle|\, \mathrm{rank}(W)\leq r
\right\}.$$ From a combinatorial-algebraic perspective, @kiraly2012algebraic provided a sufficient and necessary condition on the existence of an unique rank-$r$ solution to . It turns out that if the low-rank matrix is *generic*, then *unique completability* depends only on the support of the observations $\Omega$. This suggests that the incoherence and random sampling assumptions typically required by various nuclear norm minimization methods may limit the portion of problems solvable by the latter to only a small subset of those solvable by matrix factorization methods.
Around the same time, @wang2012stability studied the stability of matrix factorization under arbitrary noise. They obtained a stability bound for the optimal solution of around the ground truth, which turns out to be better than the corresponding bound for nuclear norm minimization in @CandesNoise by a scale of $\sqrt{\min{(m,n)}}$ (in Big-O sense). The study however bypassed the practical problem of how to obtain the global optimal solution for this non-convex problem.
This gap is partially closed by the recent work of @jain2012alt_min_global, in which the global minimum of can be obtained up to an accuracy $\epsilon$ with $O(\log{1/\epsilon})$ iterations using a slight variation of the ALS scheme. The guarantee requires the observation to be noiseless, sampled uniformly at random and the underlying subspace of $W$ needs to be incoherent—basically all assumptions in the convex approach—yet still requires slightly more observations than that for nuclear norm minimization. It does not however touch on when the algorithm is able to find the global optimal solution when the data is noisy. Despite not achieving stronger theoretical results nor under weaker assumptions than the convex relaxation approach, this is the first guarantee of its kind for matrix factorization. Given its more effective empirical performance, we believe that there is great room for improvement on the theoretical front. A secondary contribution of this paper is to find the potentially “right” algorithm or rather constituent elements of algorithm for theoreticians to look deeper into.
Numerical evaluation of matrix factorization methods {#sec:numerical_MF}
====================================================
To better understand the performance of different methods, we compare the following attributes quantitatively for all three categories of approaches that solve or [^4]:
**Sample complexity**
: Number of samples required for exact recovery of random uniformly sampled observations in random low-rank matrices, an index typically used to quantify the performance of nuclear norm based matrix completion.
**Hits on global optimal\[synthetic\]**
: The proportion of random initializations that lead to the global optimal solution on random low rank matrices with (a) increasing Gaussian noise, (b) exponentially decaying singular values.
**Hits on global optimal\[SfM\]**
: The proportion of random initializations that lead to the global optimal solution on the Oxford Dinosaur sequence [@Damped_Newton_2005] used in the SfM community.
![Percentage of hits on global optimal with **increasing level of noise**. Five rank-$4$ matrices are generated by multiplying two standard Gaussian matrices of dimension $40\times 4$ and $4\times 60$. 30% of entries are uniformly picked as observations with additive Gaussian noise $N(0,\sigma)$. 24 different random initialization are tested for each matrix. The “global optimal” is assumed to be the solution with lowest objective value across all testing algorithm and all initializations.[]{data-label="fig:mf_exp2a"}](pics/Exp2a_GlobalHits_vs_Noise.png "fig:"){width="\linewidth" height="0.67\linewidth"}\
![Percentage of hits on global optimal for **ill-conditioned low-rank matrices**. Data are generated in the same way as in Fig. \[fig:mf\_exp2a\] with $\sigma=0.05$, except that we further take SVD and rescale the $i^{th}$ singular value according to $1/\alpha^i$. The Frobenious norm is normalized to be the same as the original low-rank matrix. The exponent $\alpha$ is given on the horizontal axis.[]{data-label="fig:mf_exp2b"}](pics/Exp2b_GlobalHits_on_MatrixCondition.png "fig:"){width="\linewidth" height="0.67\linewidth"}\
![Cumulative histogram on the pixel RMSE for 100 randomly initialized runs conducted for each algorithm on Dinosaur sequence. The curve summarizes how many runs of each algorithm corresponds to the global optimal solution (with pixel RMSE 1.0847) on the horizontal axis. Note that the input pixel coordinates are normalized to between $[0,1]$ for the experiments, but to be comparable with @Damped_Newton_2005, the objective value is scaled back to the original size.[]{data-label="fig:mf_exp3"}](pics/Exp3_GlobalHits_dinosaur.png "fig:"){width="\linewidth"}\
The sample complexity experiment in Fig. \[fig:mf\_exp1\] shows that the best performing matrix factorization algorithm attains exact recovery with the number of observed entries at roughly 18%, while CVX for nuclear norm minimization needs roughly 36% (even worse for numerical solvers such as TFOCS). This seems to imply that the sample requirement for MF is fundamentally smaller than that of nuclear norm minimization. As MF assumes known rank of the underlying matrix while nuclear norm methods do not, the results we observe are quite reasonable. In addition, among different MF algorithms, some perform much better than others. The best few of them achieve something close to the lower bound[^5]. This corroborates our intuition that MF is probably a better choice for problems with known rank.
From Fig. \[fig:mf\_exp2a\] and \[fig:mf\_exp2b\], we observe that the following classes of algorithms, including LM\_X series [@Subspace_ChenPei_2008], Wiberg [@WibergL2], Non-linear Conjugate Gradient method (NLCG) [@srebro2003weighted] and the curvilinear search on Stiefel manifold\
(OptManifold [@yin2013orthogonality]) perform significantly better than others in reaching the global optimal solution despite their non-convexity. The percentage of global optimal hits from random initialization is promising even when the observations are highly noisy or when the condition number of the underlying matrix is very large[^6].
The common attribute of the four algorithms is that they are all based on the model which parameterizes the factor $V$ as a function of $U$ and then optimizes over $U$ alone. This parameterization essentially reduces the problem to finding the best subspace that fits the data. What is different between them is the way they update the solution in each iteration. OptManifold and NLCG adopt a Strong Wolfe line search that allows the algorithm to take large step sizes, while the second order methods approximate each local neighborhood with a convex quadratic function and jump directly to the minimum of the approximation. This difference appears to matter tremendously on the SfM experiment (see Fig. \[fig:mf\_exp3\]). We observe that only the second order methods achieve global optimal solution frequently, whereas the Strong Wolfe line search adopted by both OptManifold and NLCG does not seem to help much on the real data experiment like it did in simulation with randomly generated data. Indeed, neither approach reaches the global optimal solution even once in the hundred runs, though they are rather close in quite a few runs. Despite these close runs, we remark that in applications like SfM, it is important to actually reach the global optimal solution. Due to the large amount of missing data in the matrix, even slight errors in the sampled entries can cause the recovered missing entries to go totally haywire with a seemingly good local minimum (see Fig. \[fig:local\_vs\_global\_min\]). We thus refrain from giving any credit to local minima even if the $\mathrm{RMSE}_{\rm visible}$ error (defined in ) is very close to that of the global minimum. $$\label{eq:RMS_visible}
\mathrm{RMSE}_{\rm visible} := \frac{\Vert {\cal P}_{\Omega}(W_{\mathrm{recovered}} - \widehat{W}) \Vert}{\sqrt{|\Omega|}}.$$
Another observation is that LM\_GN seems to work substantially better than other second-order methods with subspace or manifold parameterization, reaching global minimum 93 times out of the 100 runs. Compared to LM\_S and LM\_M, the only difference is the use of Gauss-Newton approximation of the Hessian. According to the analysis in @chen2011hessian, the Gauss-Newton Hessian provides the only non-negative convex quadratic approximation that preserves the so-called “zero-on-$(n-1)$-D” structure of a class of nonlinear least squares problems, into which can be formulated. Compared to the Wiberg algorithm that also uses Gauss-Newton approximation, the advantage of LM\_GN is arguably the better global convergence due to the augmentation of the LM damping factor. Indeed, as we verify in the experiment, Wiberg algorithm fails to converge at all in most of its failure cases. The detailed comparisons of the second order methods and their running time on the Dinosaur sequence are summarized in Table \[tab:MC\_comparison\]. Part of the results replicate that in @Subspace_ChenPei_2008; however, Wiberg algorithm and LM\_GN have not been explicitly compared previously. It is clear from the Table that LM\_GN is not only better at reaching the optimal solution, but also computationally cheaper than other methods which require explicit computation of the Hessian[^7].
DN Wiberg LMS LMM LMGN
----------------------------------- -------------- --------------------- --------------- --------------- -------------- --
No. of hits at global min. 2 46 42 32 93
No. of hits on stopping condition 75 47 99 93 98
Average run time(sec) 324 837 147 126 40
No. of variables (m+n)r (m-r)r mr (m-r)r (m-r)r
Hessian Yes Gauss-Newton Yes Yes Gauss-Newton
LM/Trust Region Yes No Yes Yes Yes
Largest Linear system to solve $[(m+n)r]^2$ $|\Omega|\times mr$ $mr\times mr$ $[(m-r)r]^2 $ $[(m-r)r]^2$
To summarize the key findings of our experimental evaluation, we observe that: (a) the fixed-rank MF formulation requires less samples than nuclear norm minimization to achieve exact recovery; (b) the compact parameterization on the subspace, strong line search or second order update help MF algorithms in avoiding local minima in high noise, poorly conditioned matrix setting; (c) LM\_GN with Gauss-Newton update is able to reach the global minimum with a very high success rate on a challenging real SfM data sequence.
(Partially Majorized) Proximal Alternating Robust Subspace Minimization for {#sec:PARSuMi}
============================================================================
Our proposed PARSuMi method for problem works in two stages. It first obtains a good initialization from an efficient convex relaxation of , which will be described in Section \[sec:Convex\_relaxation\]. This is followed by the minimization of the low rank matrix $W$ and the sparse matrix $E$ alternatingly until convergence. The efficiency of our PARSuMi method depends on the fact that the two inner minimizations of $W$ and $E$ admit efficient solutions, which will be derived in Sections \[sec:MC\] and \[sec:outlier\_removal\] respectively. Specifically, in step $k$, we compute $W^{k+1}$ from either $$\label{eq:MC_W}
\begin{aligned}
& \underset{W}{\text{min}}
\; \frac{1}{2}\|H\circ(W-\widehat{W}+E^k)\|^2
+ \frac{\beta_1}{2}\norm{H\circ(W-W^k)}^2\\
& \text{subject to}
\quad \mathrm{rank}(W)\leq r,
\end{aligned}$$ or its quadratic majorization[^8], and $E^{k+1}$ from $$\label{eq:MC_E}
\begin{aligned}
& \underset{E}{\text{min}}
\; \frac{1}{2}\norm{H\circ(W^{k+1}-\widehat{W}+E)}^2 +
\frac{\beta_2}{2}\norm{E-E^k}^2\\
& \text{subject to}
\quad \|E\|_0\leq N_0, \; \norm{E} \leq K_E,\; E\in \R^{m\times n}_\Omega,
\end{aligned}$$ where $H$ is defined as in . Note that the above iteration is different from applying a direct alternating minimization of . We have added the proximal regularization terms $\norm{H\circ(W-W^k)}^2$ and $\|E-E^k\|^2$ to make the objective functions in the subproblems coercive and hence ensuring that $W^{k+1}$ and $E^{k+1}$ are well defined. As is shown in @attouch2010proximal, the proximal terms are critical to ensure the critical point convergence of the sequence. In addition, we have added a quadratic majorization safeguard step when computing $W^{k+1}$. This is to safeguard the convergence even if our computed $W^{k+1}$ fails to be a global minimizer of . Further details of the algorithm and the proof of its convergence are provided in the subsequent sections.
Computation of $W^{k+1}$ in {#sec:MC}
----------------------------
Our solution for consists of two steps. We first transform the rank-constrained minimization into an equivalent subspace fitting problem, then solve the new formulation using LM\_GN.
Motivated by the findings in Section \[sec:numerical\_MF\] where the most successful algorithms for solving are based on the formulation , we will now derive a similar equivalent reformulation of . Our reformulation of is motivated by the $N$-parametrization of due to @Subspace_ChenPei_2008, who considered the task of matrix completion as finding the best subspace to fit the partially observed data. In particular, Chen proposes to solve using $$\label{eq:Subspace_MC_objective_function}
\underset{N}{\text{min}} \left\{
\frac{1}{2}\sum\limits_i \hat{w}_i^T(I-\mathbb{P}_i)\hat{w}_i
\middle| N^TN=I \right\}$$ where $N$ is a $m\times r$ matrix whose column space is the underlying subspace to be reconstructed, $N_i$ is $N$ but with those rows corresponding to the missing entries in column $i$ removed. $\mathbb{P}_i=N_i N_i^+$ is the projection onto $\mathrm{span}(N_i)$ with $N_i^+$ being the Moore-Penrose pseudo inverse of $N_i$, and the objective function minimizes the sum of squares distance between $\hat{w}_i$ to $\mathrm{span}(N_i)$, where $\hat{w}_i$ is the vector of observed entries in the $i^{th}$ column of $\widehat{W}$.
### N-parameterization of the subproblem
First define the matrix $\overline{H}\in \R^{m\times n}$ as follows: $$\begin{aligned}
\label{eq:H_bar}
\overline{H}_{ij} = \left\{ \begin{array}{ll}
\sqrt{1+\beta_1} & \mbox{if $(i,j)\in \Omega$}
\\[5pt]
\sqrt{ \epsilon+\epsilon \beta_1} &\mbox{if $(i,j)\not\in \Omega$.}
\end{array}\right.\end{aligned}$$ Let $B^k\in \R^{m\times n}$ be the matrix defined by $$\begin{aligned}
\label{eq:Bk}
B^k_{ij} = \left\{ \begin{array}{ll}
\frac{1}{\sqrt{1+\beta_1}}(\widehat{W}_{ij} - E^k_{ij} + \beta_1 W^k_{ij}) & \mbox{if $(i,j)\in \Omega$}
\\[5pt]
\frac{\epsilon\beta_1}{\sqrt{\epsilon+\epsilon\beta_1}}\, W^k_{ij} &\mbox{if $(i,j)\not\in \Omega$.}
\end{array}\right.\end{aligned}$$ Define the diagonal matrices $\mathbb{D}_{i} \in \R^{m\times m}$ to be $$\begin{aligned}
\mathbb{D}_{i} = {\rm diag}(\overline{H}_i), \quad i=1,\dots,n\end{aligned}$$ where $\overline{H}_i$ is the $i$th column of $\overline{H}$. It turns out that the $N$-parameterization for the regularized problem has a similar form as , as shown below.
\[prop:equivalence\] Let $\mathbb{Q}_i(N) = \mathbb{D}_i N(N^T \mathbb{D}_i^2N)^{-1}N^T\mathbb{D}_i$, which is the $m\times m$ projection matrix onto the column space of $\mathbb{D}_i N$. The problem is equivalent to the following problem: $$\label{eq:Subspace_MC_new}
\begin{aligned}
& \underset{N}{\text{min}}
\quad f(N) := \frac{1}{2}\sum_{i=1}^n \norm{B^k_i -\mathbb{Q}_i(N) B^k_i}^2\\
& \text{subject to}
\quad N^TN=I, \; N\in \R^{m\times r}
\end{aligned}$$ where $B^k_i$ is the $i$th columns of $B^k$. If $N_*$ is an optimal solution of , then $W^{k+1}$, whose columns are defined by $$\begin{aligned}
\label{eq:W}
W^{k+1}_i = \mathbb{D}_i^{-1}\mathbb{Q}_i(N_*) B^k_i,\end{aligned}$$ is an optimal solution of .
We can show by some algebraic manipulations that the objective function in is equal to $$\begin{aligned}
\frac{1}{2}\norm{\overline{H}\circ W-B^k}^2
+ \mbox{constant}\end{aligned}$$ Now note that we have $$\begin{aligned}
&&\{ W \in \R^{m\times n}\mid \mbox{rank}(W) \leq r \} \nonumber \\
&&=
\{ NC \mid N\in \R^{m\times r}, C \in \R^{r\times n}, N^TN= I\}.\end{aligned}$$ Thus the problem is equivalent to $$\begin{aligned}
\min_N \{ f(N) \mid N^TN = I, N\in \R^{m\times r}\}\end{aligned}$$ where $$\begin{aligned}
f(N) :=
\min_{C} \frac{1}{2}\norm{\overline{H}\circ(NC)-B^k}^2.\end{aligned}$$ To derive from the above, we need to obtain $f(N)$ explicitly as a function of $N$. For a given $N$, the unconstrained minimization problem over $C$ in $f(N)$ has a strictly convex objective function in $C$, and hence the unique global minimizer satisfies the following optimality condition: $$\begin{aligned}
N^T ((\overline{H}\circ\overline{H})\circ(NC)) =
N^T (\overline{H}\circ B^k).\end{aligned}$$ By considering the $i$th column $C_i$ of $C$, we get $$\begin{aligned}
N^T \mathbb{D}_i^2 N C_i = N^T \mathbb{D}_i B^k_i, \quad i=1,\dots,n.\end{aligned}$$ Since $N$ has full column rank and $D^i$ is positive definite, the coefficient matrix in the above equation is nonsingular, and hence $$\begin{aligned}
C_i = (N^T \mathbb{D}_i^2 N)^{-1} N^T \mathbb{D}_i B^k_i.\end{aligned}$$ Now with the optimal $C_i$ above for the given $N$, we can show after some algebra manipulations that $f(N)$ is given as in .
We can see that when $\beta_1\downarrow 0$ in , then the problem reduces to , with the latter’s $\hat{w}_i$ appropriately modified to take into account of $E^k$. Also, from the above proof, we see that the $N$-parameterization reduces the feasible region of $W$ by restricting $W$ to only those potential optimal solutions among the set of $W$ satisfying the expression in . This seems to imply that it is not only equivalent but also advantageous to optimize over $N$ instead of $W$. While we have no theoretical justification of this conjecture, it is consistent with our experiments in Section \[sec:numerical\_MF\] which show the superior performance of those algorithms using subspace parameterization in finding global minima and vindicates the design motivations of the series of LM\_X algorithms in @Subspace_ChenPei_2008.
### LM\_GN updates
Now that we have shown how to handle the regularization term and validated the equivalence of the transformation, the steps to solve essentially generalize those of LM\_GN (available in Section 3.2 and Appendix A of @chen2011hessian) to account for the general mask $H$. The derivations of the key formulae and their meanings are given in this section.
In general, Levenberg-Marquadt solves the non-linear problem with the following sum-of-squares objective function $$\label{eq:nonlinear_LS}
\mathcal{L}(x)= \frac{1}{2}\sum_{i=1:n}\norm{y_i-f_i(x)}^2,$$ by iteratively updating $x$ as follows: $$x\leftarrow x+(J^TJ+\lambda I)^{-1}J^T\mathbf{r},$$ where $J=[J_1;\dots;J_n]$ is the Jacobian matrix and $J_i$ is the Jacobian matrix of $f_i$; $\mathbf{r}$ is the concatenated vector of residual $r_i := y_i-f_i(x)$ for all $i$, and $\lambda$ is the damping factor that interpolates between Gauss-Newton update and gradient descent. We may also interpret the iteration as a Damped Newton method with a first order approximation of the Hessian matrix using $J^TJ$. Note that the objective function of can be expressed in the form of by taking $x:=\mathrm{vec}(N)$, data $y_i:=B^k_i$, and function $$f_i(x:=\mathrm{vec}(N))=\QQ_i (N) B^k_i = \QQ_iy_i$$
\[prop:Jacobian\] Let $\mathcal{T} \in \R^{mr\times mr}$ be the permutation matrix such that ${\rm vec}(X^T) = \mathcal{T} {\rm vec}(X)$ for any $X\in\R^{m\times r}$. The Jacobian of $f_i(x) = \mathbb{Q}_i(N) y_i$ is given as follows: $$\begin{aligned}
J_i (x) =(\AA_i^T y_i)^T \otimes ((I-\mathbb{Q}_i)\DD_i)
+\; [ (\DD_i r_i)^T\otimes \AA_i] \mathcal{T}. \quad\;
\label{eq:Ji}\end{aligned}$$ Also $J^TJ = \sum_{i=1}^n J_i^TJ_i$, $J^Tr = \sum_{i=1}^n J_i^T r_i$, where $$\begin{aligned}
J_i^T J_i &=& (\AA_i^T y_iy_i^T \AA_i) \otimes (\DD_i(I-\QQ_i)\DD_i)
\nonumber \\[5pt]
&&+\, \mathcal{T}^T [ (\DD_i r_i r_i^T \DD_i) \otimes (\AA_i^T \AA_i)]
\mathcal{T} \label{eq:JTJ}
\\[5pt]
J_i^Tr_i &=& {\rm vec}(\DD_i r_i (\AA_i^T y_i)^T).
\label{eq:JTr}\end{aligned}$$ In the above, $\otimes$ denotes the Kronecker product.
Let $\AA_i = \DD_i N (N^T\DD_i^2 N)^{-1}$. Given sufficiently small $\dN$, we can show that the directional derivative of $f_i$ at $N$ along $\dN$ is given by $$\begin{aligned}
f_i^\prime (N+\dN) &= &
(I-\mathbb{Q}_i)\DD_i\dN \AA_i^T y_i
+ \AA_i \dN^T \DD_i r_i.\end{aligned}$$ By using the property that ${\rm vec}(AXB)=(B^T\otimes A){\rm vec}(X)$, we have $$\begin{aligned}
{\rm vec}(f_i^\prime (N+\dN)) &=&
[(\AA_i^T y_i)^T \otimes ((I-\mathbb{Q}_i)\DD_i)] {\rm vec}(\dN)
\\
&&+ [ (\DD_i r_i)^T\otimes \AA_i] {\rm vec}(\dN^T)\end{aligned}$$ From here, the required result in follows.
To prove , we make use of the following properties of Kronecker product: $(A\otimes B)(C\otimes D) = (AC)\otimes (BD)$ and $(A\otimes B)^T = A^T\otimes B^T$. By using these properties, we see that $J_i^T J_i$ has four terms, with two of the terms containing involving $\DD_i(I-\QQ_i)\AA_i$ or its transpose. But we can verify that $\QQ_i\AA_i=\AA_i$ and hence those two terms become $0$. The remaining two terms are those appearing in after using the fact that $(I-\QQ_i)^2 = I-\QQ_i$. Next we prove . We have $$\begin{aligned}
J_i^T r_i = {\rm vec}(\DD_i (I-\QQ_i)r_i (\AA_i^Ty_i)^T) +
\mathcal{T}^T {\rm vec}(\AA_i^T r_i r_i^T \DD_i).\end{aligned}$$ By noting that $\AA_i^T r_i = 0$ and $ \QQ_i r_i =0$, we get the required result in .
The complete procedure of solving is summarized in Algorithm \[alg:LM\_GN\]. In all our experiments, the initial $\lambda$ is chosen as $1e-6$ and $\rho=10$.
$\widehat{W}, E^k, W^k,\bar{H}$, objective function $\mathcal{L}(x)$ and initial $N^k$; numerical parameter $\lambda$, $\rho>1$. Compute $y_i=B_i^k$ for $i=1,...,n$, and $x^0=\mathrm{vec}(N^k)$, $j=0.$ Compute $J^T\mathbf{r}$ and $J^TJ$ using and. Compute $\Delta x=(J^TJ+\lambda I)^{-1}J^Tr$ $\lambda=\rho\lambda.$ $\Delta x=(J^TJ+\lambda I)^{-1}J^Tr.$ $\lambda=\lambda/\rho.$ Orthogonalize $N=\mathrm{orth}[\mathrm{reshape}(x^j+\Delta x)]$. Update $x^{j+1}=\mathrm{vec}(N)$. Iterate $j=j+1$ $N^{k+1}=N$, $W^{k+1}$ using with $N^{k+1}$ replacing $N_*$.
Quadratic majorization of {#sec:majorized_MC}
--------------------------
Recall that while the LM$\_$GN method may be highly successful in computing a global minimizer for empirically, $W^{k+1}$ may fail to be a global minimizer occasionally. Thus before deriving the update rule for $E^{k+1}$, we consider minimizing a quadratic majorization of as a safeguard step to ensure the convergence of the PARSuMi iterations. Recall that is equivalent to $$W^{k+1}={\mathop{\mathrm{argmin}}}_{W}\left\{\frac{1}{2}\|\overline{H}\circ (W-\hat{B}^k)\|^2\middle| \mathrm{rank}(W)\leq r\right\}$$ where $\hat{B}^k=\overline{H}^{-1}\circ B^k$ with $\overline{H}$ and $B^k$ defined as in and respectively.
For convenience, we denote the above objective function $F(W,\hat{B}^k)$. Suppose that we have positive vectors $p\in\R^m$ and $q\in \R^n$ such that $$\begin{aligned}
\bar{H}_{ij}^2 \leq p_i q_j\quad \forall\; i=1,\dots,m, \; j=1,\dots,n.
\label{eq-1}\end{aligned}$$ Note that the above inequality always holds if $p$, $q$ are chosen to be $$\begin{aligned}
p_i &=& \max\{ \bar{H}_{ij} \mid j=1,\dots,n\}, \quad i=1,\dots,m
\nonumber \\
q_j &=& \max\{ \bar{H}_{ij} \mid i=1,\dots,m\}, \quad j=1,\dots,n.
\label{eq-2}\end{aligned}$$ Let $G^k=\nabla_W F(W^k,\hat{B}^k)$. We majorize $F(W,\hat{B}^k)$ by bounding its Taylor expansion at $W^k$ $$\begin{aligned}
&F(W,\hat{B}^k) - F(W^k,\hat{B}^k)
= \inprod{G^k}{W-W^k} \nonumber\\
&+ \frac{1}{2}\inprod{W-\hat{B}^k}{(\bar{H}\circ\bar{H})\circ (W-\hat{B}^k)}
\nonumber \\\
\leq &\;
\inprod{G^k}{W-W^k} + \frac{1}{2}\inprod{W-W^k}{P(W-W^k)Q}
\nonumber \\
=&
\frac{1}{2}\norm{P^{1/2} W Q^{1/2} -U^k}^2
- \frac{1}{2}\norm{ P^{-1/2}G^k Q^{-1/2}}^2 \label{eq:F_majorization}\end{aligned}$$ where $P= \diag{p}$ and $Q = \diag{q}$, $U^k=P^{1/2}W^k Q^{1/2} - P^{-1/2}G^k Q^{-1/2}$.
\[prop:majorization\_update\] The minimizer of quadratic majorization function in is given in closed-form by $$\label{eq:majorization_update}
W_{\rm QM}^{k+1} \in P^{-1/2}\Pi_r(U^k)Q^{-1/2}.$$ Here $\Pi_r(U^k)$ denotes the set of best rank-r approximation of $U^k$.
The proof is simple and is given in the Appendix. Note that this is a nonconvex minimization, yet we have an efficient closed-form solution thanks to SVD.
As we shall see later in Algorithm \[alg:PAM\], the global minimizer $W^{k+1}_{\rm QM}$ in of the quadratic majorization function of $F(W,\hat{B}^k)$ is used as a safeguard when the computed solution $W^{k+1}$ from is inferior (which necessarily implies that $W^{k+1}$ is not a global optimal solution) to $W^{k+1}_{\rm QM}$. By doing so, the convergence of PARSuMi can be ensured, as we shall prove in Theorem \[thm:critical\_pt\].
Sparse corruption recovery step {#sec:outlier_removal}
--------------------------------
In the sparse corruption step, we need to solve the $\ell_0$-constrained least squares minimization . This problem is combinatorial in nature, but fortunately, for our problem, we show that a closed-form solution can be obtained. Let $x := \mathcal{P}_\Omega(E)$. Observe that can be expressed in the following equivalent form: $$\begin{aligned}
\min_x \Big\{ \norm{x- b}^2 \mid \norm{x}_0 \leq N_0, \; \norm{x}^2 -K_E^2\leq 0\Big\}
\label{eq:MC_E2}\end{aligned}$$ where $b = \mathcal{P}_\Omega(\widehat{W}-W^{k+1}+\beta_2 E^k)/(1+\beta_2)$.
\[prop:E\_update\] Let $I$ be the set of indices of the $N_0$ largest (in magnitude) component of $b$. Then the nonzero components of the optimal solution $x$ of is given by $$\begin{aligned}
x_I = \left\{ \begin{array}{ll}
K_E b_I/\norm{b_I} & \mbox{if $\norm{b_I} > K_E$}
\\[5pt]
b_I & \mbox{if $\norm{b_I}\leq K_E$.}
\end{array} \right.
\label{eq:MC_E2x}\end{aligned}$$
The proof (deferred to the Appendix) involves checking the optimality conditions of assuming known support set and finding the optimal support set in a decoupled fashion.
The procedure to obtain the optimal solution of is summarized in Algorithm \[alg:MC\_E\]. We remark that this is a very special case of $\ell_0$-constrained optimization; the availability of the exact closed form solution depends on both terms in being decomposable into individual $(i,j)$ term. In general, if we change the operator $M \rightarrow H\circ M$ in to a general linear transformation (e.g., a sensing matrix in compressive sensing), or change the norm $\|\cdot\|$ of the proximal term to some other norm such as spectral norm or nuclear norm, then the problem becomes NP-hard.
$\widehat{W}, W^{k+1},E^k, \Omega$. Compute $b$ using . Compute $x$ using . $E^{k+1} = \cP_\Omega^*(x)$.
Algorithm
---------
Our method is summarized in Algorithm \[alg:PARSuMi\].
Observed data $\widehat{W}$, sample mask ${\Omega}$, parameter $r, N_0$. Initialization $W^0$ and $E^0$ (typically by Algorithm \[alg:APG\] described in Section \[sec:Convex\_relaxation\]), $k=0$. Solve using Algorithm \[alg:LM\_GN\] with $W^k$,$E^k$,$N^k$, obtain updates $\tilde{W}^{k+1}$ and $\tilde{N}^{k+1}$ Evaluate with $W^k$,$E^k$ obtain updates $\hat{W}^{k+1}$. Assign $W^{k+1}={\mathop{\mathrm{argmin}}}_{W\in\{\tilde{W}^{k+1},\hat{W}^{k+1}\}} F(W,\hat{B}^k)$, and then assign the corresponding $N^{(k+1)}$. Solve using Algorithm \[alg:MC\_E\] with $W^{k+1}, E^k$; obtain updates $E^{k+1}$. Accumulation points $\overline{W}$ and $\overline{E}$
Note that we do not need to know the exact cardinality of the corrupted entries; $N_0$ can be taken as an upper bound of the allowable number of corruptions. As a rule of thumb, 10-15% of $|\Omega|$ is a reasonable size. The surplus in $N_0$ will only label a few noisy samples as corruptions, which should not affect the recovery of either $W$ or $E$, so long as the remaining $|\Omega|-N_0$ samples are still sufficient. The other parameter $r$ is typically given by the physical model of the problem. For those problems where $r$ is not known, choosing $r$ is analogous to choosing the regularization parameter as in other machine learning tasks. A large $r$ will lead to overfitting and poorly estimated missing data while an overly small $r$ will cause underfitting of the observed data.
Convergence to a critical point {#sec:convergence}
--------------------------------
In this section, we show the convergence of Algorithm \[alg:PARSuMi\] to a critical point. Note that due to the non-convex nature of the subproblem , Algorithm \[alg:LM\_GN\] is guaranteed to converge only to a local minimum. Therefore, the result in @attouch2010proximal that requires global optimal solutions in all subproblems cannot be directly applied in our case for the critical point convergence proof. Empirically, we cannot hope LM\_GN to always find the global optimal solution of either, as our experiments on LM\_GN in Section \[sec:numerical\_MF\] clearly demonstrated. As a result, we design the partial majorization (Step 1b) in Algorithm \[alg:PARSuMi\] to safeguard against the case when the computed solution from Step 1a is not a global optimal solution. The safeguard step is powerful in that we do not need to assume anything on the computed solution of the subproblem before we can prove the overall critical point convergence.
We start our convergence proof by first defining an equivalent formulation of in terms of closed, bounded sets. The convergence proof is then based on the indicator functions for these closed and bounded sets, which have the key lower semicontinuous property.
Let $K_W = 2\norm{\widehat{W}} + K_E$. Define the closed and bounded sets: $$\begin{aligned}
\mathcal{W} &=& \{ W \in \R^{m\times n} \mid {\rm rank}(W)\leq r, \norm{H\circ W} \leq
K_W \}
\\[5pt]
\mathcal{E} &=& \{ E\in \R^{m\times n}_\Omega \mid
\norm{E}_0 \leq N_0, \norm{E} \leq K_E\}.\end{aligned}$$ We will first show that is equivalent to the problem given in the next proposition.
Let $f(W,E):=\frac{1}{2}\norm{H\circ(W+E-\widehat{W})}^2$. The problem is equivalent to the following problem: $$\begin{aligned}
\min \big\{ f(W,E)
\mid
W\in\mathcal{W}, E\in \mathcal{E} \big\}.
\label{eq:MC_Formulation_2}\end{aligned}$$
Observe that the only difference between and is the inclusion of the bound constraint on $\norm{H\circ W}$ in . To show the equivalence, we only need to show that any minimizer $(\overline{W}, \overline{E})$ of must satisfy the bound constraint in $\mathcal{W}$. By definition, we know that $$f(\overline{W}, \overline{E}) \leq f(0,0) = \frac{1}{2}\norm{\widehat{W}}^2.$$ Now for any $(W,E)$ such that ${\rm rank}(W)\leq r$, $E\in \mathcal{E}$ and $\norm{H\circ W} > K_W$, we must have $$\begin{aligned}
& & \hspace{-0.7cm}
\norm{H\circ(W+E-\widehat{W})} \geq \norm{H\circ W} - \norm{H\circ(E-\widehat{W})}
\\[5pt]
& > & K_W - \norm{E}-\norm{\widehat{W}} \geq \norm{\widehat{W}}.\end{aligned}$$ Hence $ f(W,E) > \frac{1}{2} \norm{\widehat{W}}^2 = f(0,0).$ This implies that we must have $\norm{H\circ \overline(W)} \leq K_W$.
To establish the convergence of PARSuMi, it is more convenient for us to consider the following generic problem which includes as a special case. Let $\mathcal{X}$, $\mathcal{Y}$, $\mathcal{Z}$ be finite-dimensional inner product spaces, and $f:\mathcal{X} \rightarrow \mathbb{R} \cup \{\infty\}$, $g:\mathcal{Y} \rightarrow \mathbb{R} \cup \{\infty\}$ are lower semi-continuous functions. We consider the problem: $$\begin{aligned}
\label{eq:L}
\min_{x,y} \{L(x,y) := f(x) + g(y) + q(x,y)\}\end{aligned}$$ where $
q(x,y) = \frac{1}{2} \norm{Ax + By - c}^2
$ and $A:\mathcal{X}\rightarrow \mathcal{Z}$, $B: \mathcal{Y}\rightarrow \mathcal{Z}$ are given linear maps. For , we have $\mathcal{X}=\mathcal{Z}=\R^{m\times n}$, $\mathcal{Y}=\R^{m\times n}_\Omega$, $A(x) = H\circ x$, $B(y) = H\circ y$, $c=\widehat{W}$. and $f$, $g$ are the following indicator functions, $$\begin{aligned}
\label{eq:Sr}
f(x) = \left\{
\begin{array}{ll}
0 & \quad \mbox{if $x\in\mathcal{W}$}\\
\infty & \quad \mbox{otherwise}
\end{array} \right.
\quad
g(y)= \left\{
\begin{array}{ll}
0 & \quad \mbox{if $y \in\mathcal{E}$}\\
\infty & \quad \mbox{otherwise}
\end{array} \right.\end{aligned}$$ Note that in this case, $f$ are $g$ are lower semicontinuous since indicator functions of closed sets are lower semicontinuous [@Rudin].
To denote the corresponding majorization safeguard, we define, for a fixed $(\hx,\hy)$ and given $M \succ A^*A$ $$\begin{aligned}
Q(x; \hx,\hy) &:= q(\hx,\hy) + \inprod{\nabla_xq(\hx,\hy)}{x-\hx}
+ \frac{1}{2}
\norm{x-\hx}_M^2 \label{eq-Q}
\\
\widehat{L}(x;\hx,\hy) &:= Q(x; \hx,\hy) + f(x) + g(\hy) \label{eq-hL}\end{aligned}$$ where $\norm{\cdot}_M$ is defined in the last part of Algorithm \[alg:PAM\].
Then, we have that $$\begin{aligned}
q(x,\hy) &= Q(x;\hx,\hy) -\frac{1}{2}\norm{x-\hx}_{M-A^*A}^2\label{eq-qQ}
\\[5pt]
L(x,\hy) &= \widehat{L}(x;\hx,\hy) -\frac{1}{2}\norm{x-\hx}_{M-A^*A}^2\quad
\label{eq-LhL}\end{aligned}$$
Consider the partially majorized proximal alternating minimization (PMPAM) outlined in Algorithm \[alg:PAM\], which we have modified from @attouch2010proximal. The algorithm alternates between minimizing $x$ and $y$, but with the important addition of the quadratic Moreau-Yoshida regularization term (which is also known as the proximal term) in each step. The importance of Moreau-Yoshida regularization for convex matrix optimization problems has been demonstrated and studied in @Nuclear_PPA [@LogDet_PPA; @Knorm_PPA]. For our non-convex, non-smooth setting here, the importance of the proximal term will become clear when we prove the convergence of Algorithm \[alg:PAM\]. For our problem , the positive linear maps $S$ and $T$ in Algorithm \[alg:PAM\] correspond to $\beta_1 (H \circ H) \circ$ and $\beta_2 I$ respectively, where $\beta_1,\beta_2$ are given positive parameters. Our algorithm differs from that in @attouch2010proximal by having the safeguard step (in Step 1b and Step 2) to ensure that critical point convergence can be achieved even if the computed solution in Step 1a is not globally optimal. Observe that one can bypass Step 1a in Algorithm \[alg:PAM\] completely and always choose to use Step 1b. But the minimization in Step 1b based on quadratic majorization may not reduce the merit function $L(x,y_k)+\frac{1}{2}\norm{x-x_k}_S^2$ as quickly as the minimization in Step 1a. Thus it is necessary to have Step 1a to ensure that Algorithm 4 converges at a reasonable speed. We note that a similar safeguard step can be introduced for the subproblem in Step 3 if the global optimality of $y_{k+1}$ is not guaranteed.
$(x_0,y_0)\in \cX\times\cY$; positive linear operators $S$ and $T$. Choose $M$ such that $M\succ A^*A + S$ in . Compute $
\tx^{k+1} \in \mbox{argmin} \{ L(x,y_k) + \frac{1}{2}\norm{x-x_k}_S \}.
$ Compute $
\hx_{k+1} \in \mbox{argmin} \left\{ \begin{array}{l}
\widehat{L}(x;x_k,y_k)
\end{array}
\right\}
$ Consider condition $
\mbox{(I)} \quad L(\tx_{k+1},y_k)+\frac{1}{2}\norm{\tx_{k+1}-x_k}_S\leq
L(\hx_{k+1},y_k) +\frac{1}{2}\norm{\hx_{k+1}-x_k}_S.
$ Set $$\begin{aligned}
x_{k+1} = \left\{ \begin{array}{ll}
\tx_{k+1} & \mbox{if condition (I) holds}
\\[5pt]
\hx_{k+1} & \mbox{otherwise.}
\end{array}
\right.\end{aligned}$$ $y^{k+1} = {\mathop{\mathrm{argmin}}}\{L(x^{k+1},y) + \frac{\beta_2}{2}\|y-y^k\|_T^2\}$ Accumulation points $\overline{x}$ and $\overline{y}$
In the above, $S$ and $T$ are given positive definite linear maps, and $\norm{x-x^k}_S^2 = \inprod{x-x^k}{S(x-x^k)}$, $\norm{y-y^k}_T^2 = \inprod{y-y^k}{T(y-y^k)}$. Note that Step 1b is to safeguard against the possibility that the computed $\tilde{x}_{k+1}$ is not a global optimal solution of the subproblem. We assume that it is possible to compute the global optimal solution $\widehat{x}_{k+1}$ analytically.
Note that for our problem , the global minimizer of the nonconvex subproblem in Step 1b can be computed analytically as discussed in Section \[sec:majorized\_MC\]. Next we show that any limit point of $\{ (x_{k},y_{k})\}$ is a stationary point of $L$ even if $\tx_{k+1}$ computed in Step 1a is not a global minimizer of the subproblem.
\[thm:critical\_pt\] Let $\{(x_k,y_k)\}$ be the sequence generated by Algorithm \[alg:PAM\], and $(\tx_{k+1},\hx_{k+1})$ are the intermediate iterates at iteration $k$.\
(a) For all $k\geq 0$, we have that $$\begin{aligned}
& L(\hx_{k+1},y_k) + \frac{1}{2}\norm{\hx_{k+1}-x_k}_S^2
+\frac{1}{2}\norm{\hx_{k+1}-x_k}_{M-A^*A-S}^2\nonumber\\
&=\; \widehat{L}(\hx_{k+1};x_k,y_k)
\;\leq\; L(x_k,y_k),
\label{eq-thm1-1}
\\
& L(x_{k+1},y_k) + \frac{1}{2}\norm{x_{k+1}-x_k}_S^2 \leq L(\hx_{k+1},y_k)+ \frac{1}{2}\norm{\hx_{k+1}-x_k}_S^2.
\label{eq-thm1-2}\end{aligned}$$ (b) For all $k\geq 0$, we have that $$L(x_{k+1},y_{k+1}) + \frac{1}{2}\norm{x_{k+1}-x_k}_S^2 + \frac{1}{2}\norm{y_{k+1}-y_k}_T^2
\leq L(x_k,y_k).
\label{eq-thm1-3}$$ Hence $\sum_{k=0}^\infty \norm{x_{k+1}-x_k}_S^2 + \norm{y_{k+1}-y_k}_T^2 < \infty$ and $\displaystyle\lim_{k\rightarrow\infty} \norm{x_{k+1}-x_k} = 0 = \lim_{k\rightarrow\infty} \norm{y_{k+1}-y_k}$.\
(c) Let $\{(x_{k'},y_{k'})\}$ be any convergent subsequence of $\{(x_k,y_k)\}$ with limit $(\bar{x},\bar{y})$. Then $
\lim_{k\rightarrow\infty} L(x_{k},y_{k}) =\lim_{k\rightarrow\infty}
L(x_{k+1},y_{k}) = \lim_{k\rightarrow\infty}
\widehat{L}(\hx_{k+1};x_k,y_{k}) = L(\bar{x},\bar{y}) .
$ Furthermore $\lim_{k\rightarrow\infty} \norm{\hx_{k+1}-x_k}=0$.\
(d) Let $\{(x_{k'},y_{k'})\}$ be any convergent subsequence of $\{(x_k,y_k)\}$ with limit $(\bar{x},\bar{y})$. Then $(\bar{x},\bar{y})$ is a stationary point of $L$.
The full proof is given in the Appendix. Here we explain the four parts of Theorem \[thm:critical\_pt\]. Part(a) establishes the non-increasing monotonicity of the proximal regularized update. Leveraging on part(a), part(b) ensures the existence of the limits. Using Part(a), (b) and (c), (d) then shows the critical point convergence of Algorithm \[alg:PAM\].
Convex relaxation of as initialization {#sec:Convex_relaxation}
--------------------------------------
Due to the non-convexity of the rank and $\ell_0$ cardinality constraints, it is expected that the outcome of Algorithm \[alg:PARSuMi\] depends on initializations. A natural choice for the initialization of PARSuMi is the convex relaxation of both the rank and $\ell_0$ function: $$\label{eq:RIRM_convex}
\min\Big\{
f(W,E) + \lambda \norm{W}_* + \gamma \norm{E}_1 \mid
W\in \R^{m\times n}, E \in \R^{m\times n}_\Omega
\Big\}$$ where $f(W,E) = \frac{1}{2}\norm{H\circ(W + E - \widehat{W}) }^2$, $\norm{\cdot }_*$ is the nuclear norm, and $\lambda$ and $\gamma$ are regularization parameters.
Problem can be solved efficiently by the quadratic majorization-APG (accelerated proximal gradient) framework proposed by @APG_NN. At the $k$th iteration with iterate $(\bar{W}^k,\bar{E}^k)$, the majorization step replaces with a quadratic majorization of $f(W,E)$, so that $W$ and $E$ can be optimized independently, as we shall see shortly. Let $G^k = (H\circ H)\circ(\bar{W}^k+\bar{E}^k+\widehat{W})$. By some simple algebra, we have $$\begin{aligned}
&& f(W,E) - f(\bar{W}^k,\bar{E}^k) = \frac{1}{2} \norm{H\circ(W-\bar{W}^k+E-\bar{E}^k)}^2
\\
& &\qquad \qquad +\; \inprod{W-\bar{W}^k+E-\bar{E}^k}{G^k}
\\
&& \leq \; \norm{W-\bar{W}^k}^2+\norm{E-\bar{E}^k}^2 +
\inprod{W-\bar{W}^k+E-\bar{E}^k}{G^k}
\\[5pt]
&&= \norm{W-\widetilde{W}^k}^2 + \norm{E-\widetilde{E}^k}^2 +{\rm constant}\end{aligned}$$ where $\widetilde{W}^k = \bar{W}^k-G^k/2$ and $\widetilde{E}^k = \bar{E}^k-G^k/2$. At each step of the APG method, one minimizes with $f(W,E)$ replaced by the above quadratic majorization. As the resulting problem is separable in $W$ and $E$, we can minimize them separately, thus yielding the following two optimization problems: $$\begin{aligned}
W^{k+1} = \mbox{argmin} \; &\frac{1}{2} \norm{ W - \widetilde{W}^k}^2 +
\frac{\lambda}{2} \norm{W}_* \label{eq:RIRM_convex_majorized_W}\\
E^{k+1} = \mbox{argmin} \; & \frac{1}{2}\norm{ E - \widetilde{E}^k}^2 +
\frac{\gamma}{2} \norm{ E}_1 \label{eq:RIRM_convex_majorized_E}\end{aligned}$$ The main reason for performing the above majorization is because the solutions to and can readily be found with closed-form solutions. For , the minimizer is given by the Singular Value Thresholding (SVT) operator. For , the minimizer is given by the well-known soft thresholding operator [@donoho1995softthresh]. The APG algorithm, which is adapted from @APG_L1 and analogous to that in @APG_NN, is summarized below.
Initialize $W^0=\bar{W}^0=0$, $E^0=\bar{E}^0=0$, $t_0=1$, $k=0$ Compute $G^k = (H\circ H)\circ(\bar{W}^k+\bar{E}^k+\widehat{W})$, $\widetilde{W}^k$, $\widetilde{E}^k$. Update $W^{k+1}$ by applying the SVT on $\widetilde{W}^k$ in . Update $E^{k+1}$ by applying the soft-thresholding operator on $\widetilde{E}^k$ in . Update step size $t_{k+1} = \frac{1}{2}(1+\sqrt{1+4t_k^2})$. $(\bar{W}^{k+1},\bar{E}^{k+1}) = (W^{k+1},E^{k+1}) + \frac{t_k-1}{t_{k+1}}
(W^{k+1}-W^k,E^{k+1}-E^k) $ Accumulation points $\overline{W}$ and $\overline{E}$
As has already been proved in @APG_L1, the APG algorithm, including the one above, has a very nice worst case iteration complexity result in that for any given $\epsilon>0$, the APG algorithm needs at most $O(1/\sqrt{\epsilon})$ iterations to compute an $\epsilon$-optimal (in terms of function value) solution.
The tuning of the regularization parameters $\lambda$ and $\gamma$ in is fairly straightforward. For $\lambda$, we use the singular values of the converged $\overline{W}$ as a reference. Starting from a relatively large value of $\lambda$, we reduce it by a constant factor in each pass to obtain a $\overline{W}$ such that its singular values beyond the $r$th are much smaller than the first $r$ singular values. For $\gamma$, we use the suggested value of $1/\sqrt{\max(m,n)}$ from RPCA [@Candes2011_JACM]. In our experiments, we find that we only need a ballpark figure, without having to do a lot of tuning. Taking $\lambda=0.2$ and $\gamma=1/\sqrt{\max(m,n)}$ serve the purpose well.
Other heuristics {#sec:heuristics}
----------------
In practice, we design two heuristics to further boost the quality of the convex initialization. These are tricks that allow PARSuMi to detect corrupted entries better and are always recommended.
We refer to the first heuristic as “Huber Regression”. The idea is that the quadratic loss term in our matrix completion step is likely to result in a dense spread of estimation error across all measurements. There is no guarantee that those true corrupted measurements will hold larger errors comparing to the uncorrupted measurements. On the other hand, we note that the quality of the subspace $N^k$ obtained from LM\_GN is usually good despite noisy/corrupted measurements. This is especially true when the first LM\_GN step is initialized with Algorithm \[alg:APG\]. Intuitively, we should be better off with an intermediate step, using $N^{k+1}$ to detect the errors instead of $W^{k+1}$, that is, keeping $N^{k+1}$ as a fixed input and finding coefficient $C$ and $E$ simultaneously with $$\label{eq:use_N_E}
\begin{aligned}
& \min_{E,C}
\; \frac{1}{2}\norm{H\circ (N^{k+1}C-\widehat{W}+E)}^2\\
& \text{subject to}
\quad \|E\|_0\leq N_0.
\end{aligned}$$ To make it computationally tractable, we relax to $$\label{eq:use_N_E_convex}
\begin{aligned}
& \min_{E,C}
\; \frac{1}{2}\norm{H\circ (N^{k+1}C-\widehat{W}+E)}^2+\eta_0\|E\|_1\\
\end{aligned}$$ where $\eta_0 >0$ is a penalty parameter. Note that each column of the above problem can be decomposed into the following Huber loss regression problem ($E$ is absorbed into the Huber penalty) $$\label{eq:use_N_E_huber}
\begin{aligned}
& \min_{C_j} \sum_{i=1}^m
\mbox{Huber}_{\eta_0/H_{ij}}(H_{ij} ((N^{k+1} C_j)_i-\widehat{W}_{ij})). \\
\end{aligned}$$ Since $N^{k+1}$ is known, can be solved very efficiently using the APG algorithm, whose derivation is similar to that of Algorithm \[alg:APG\], with soft-thresholding operations on $C$ and $E$. To further reduce the Robin Hood effect (that haunts all $\ell_1$-like penalties) and enhance sparsity, we may optionally apply the iterative re-weighted Huber minimization (a slight variation of the method in @candes2008enhancing), that is, solving for $l_{max}$ iterations using an entrywise weighting factor inversely proportional to the previous iteration’s fitting residual. In the end, the optimal columns $C_j$’s are concatenated into the optimal solution matrix $C^*$ of , and we set $$W^{k+1}=N^{k+1}C^*.$$ With this intermediate step between the $W$ step and the $E$ step, it is much easier for the $E$ step to detect the support of the actual corrupted entries.
The above procedure can be used in conjunction with another heuristic that avoids adding false positives into the corruption set in the $E$ step when the subspace $N$ has not yet been accurately recovered. This is achieved by imposing a threshold $\eta$ on the minimum absolute value of $E^k$’s non-zero entries, and shrink this threshold by a factor (say 0.8) in each iteration. The “Huber regression” heuristic is used only when $\eta>\eta_0$, and hence only in a very small number of iteration before the support of $E$ has been reliably recovered. Afterwards the pure PARSuMi iterations (without the Huber step) will take over, correct the Robin Hood effect of Huber loss and then converge to a high quality solution.
Note that our critical point convergence guarantee in Section \[sec:convergence\] is not hampered at all by the two heuristics, since after a small number of iterations, $\eta\leq\eta_0$ and we come back to the pure PARSuMi.
Experiments and discussions {#sec:experiment}
===========================
In this section, we present the methodology and results of various experiments designed to evaluate the effectiveness of our proposed method. The experiments revolve around synthetic data and two real-life datasets: the Oxford Dinosaur sequence, which is representative of data matrices in SfM works, and the Extended YaleB face dataset [@lee2005extendedyaleb], which we use to demonstrate how PARSuMi works on photometric stereo problems.
In the synthetic data experiments, our method is compared with the state-of-the-art algorithms for the objective function in namely Wiberg $\ell_1$ [@WibergL1] and GRASTA [@he2011grasta]. ALP and AQP [@ke2005L1] are left out since they are shown to be inferior to Wiberg $\ell_1$ in @WibergL1. For the sake of comparison, we perform the experiment on recovery effectiveness using the same small matrices as in Section 5.1 of @WibergL1. Other synthetic experiments are conducted with more reasonably-sized matrices. Whenever appropriate, we also include a comparison to a variant of RPCA that handles missing data [@Wu_photometric] which solves using the augmented Lagrange multiplier (ALM) algorithm (we will call it ALM-RPCA from here onwards). This serves as a representative of the nuclear norm based methods. The real data from the SfM and photometric stereo problems contain many challenges typical in practical scenarios. They contain large contiguous areas of missing data, and potentially highly corrupted observations which may not be sparse too. For instance, in the YaleB face dataset, grazing illumination tends to produce large area of missing data (well over 50%) and often large number of outliers too (due to specular highlights). The PARSuMi method outperformed a variety of other methods in the experiments, even uncovering hitherto unknown corruptions inherent in the Dinosaur data from SfM. The results also corroborate those obtained in the synthetic data experiments, in that our method can handle a substantially larger fraction of missing data and corruptions, thus providing empirical evidence for the efficacy of PARSuMi under practical scenarios.
For a summary of the parameters used in the experiments, please refer to the Appendix.
Convex Relaxation as an Initialization Scheme
---------------------------------------------
We first investigate the results of our convex initialization scheme by testing on a randomly generated $100\times 100$ rank-4 matrix. A random selection of 70% and 10% of the entries are considered missing and corrupted respectively. Corruptions are generated by adding large uniform noise between $[-1,1]$. In addition, Gaussian noise $\mathcal{N}(0,\sigma)$ for $\sigma=0.01$ is added to all observed entries. From Fig. \[fig:E\_robinhood\], we see that the convex relaxation outlined in Section \[sec:Convex\_relaxation\] was able to recover the error support, but there is considerable difference in magnitude between the recovered error and the ground truth, owing to the “Robin Hood” attribute of $\ell_1$-norm as a convex proxy of $\ell_0$. Nuclear norm as a proxy of rank also suffers from the same woe. Similar observations can be made on the results of the Dinosaur experiments, which we will show later.
![The Robin Hood effect of Algorithm \[alg:APG\] on detected sparse corruptions $E_{\mathrm{Init}}$. **Left**: illustration of a random selection of detected E vs. true E. Note that the support is mostly detected, but the magnitude falls short. **Right**: scatter plot of the detected E against true E (perfect recovery falls on the $y=x$ line, false positives on the $y$-axis and false negatives on the $x$-axis).[]{data-label="fig:E_robinhood"}](pics/E_robinhood.png "fig:"){width="0.49\linewidth"} ![The Robin Hood effect of Algorithm \[alg:APG\] on detected sparse corruptions $E_{\mathrm{Init}}$. **Left**: illustration of a random selection of detected E vs. true E. Note that the support is mostly detected, but the magnitude falls short. **Right**: scatter plot of the detected E against true E (perfect recovery falls on the $y=x$ line, false positives on the $y$-axis and false negatives on the $x$-axis).[]{data-label="fig:E_robinhood"}](pics/E_robinhood2.png "fig:"){width="0.49\linewidth"}\
Despite the problems with the solution of the convex initialization, we find that it is a crucial step for PARSuMi to work well in practice. As can be seen from Fig. \[fig:E\_robinhood\], the detected error support can be quite accurate. This makes the $E$-step of PARSuMi more likely to identify the true locations of corrupted entries.
Impacts of poor initialization
------------------------------
When the convex initialization scheme fails to obtain the correct support of the error, the “Huber Regression” heuristic may help PARSuMi to identify the support of the corrupted entries. We illustrate the impact by intentionally mis-tuning the parameters of Algorithm \[alg:APG\] such that the initial $E$ bears little resemblance to the true injected corruptions. Specifically, we test the cases when the initialization fails to detect many of the corrupted entries (false negatives) and when many entries are wrongly detected as corruptions (false positives). From Fig. \[fig:poor\_initialization\], we see that PARSuMi is able to recover the corrupted entries to a level comparable to the magnitude of the injected Gaussian noise in both experiments[^9].
\
In most of our experiments, we find that PARSuMi is often able to detect the corruptions perfectly from a simple initializations with all zeros, even without the “Huber Regression” heuristic. This is especially true when the data are randomly generated with benign sampling pattern and well-conditioned singular values. However, in challenging applications such as SfM, a good convex initialization and the “Huber Regression” heuristic are always recommended.
Recovery effectiveness from sparse corruptions {#sec:Recovery_Effectiveness}
----------------------------------------------
For easy benchmarking, we use the same synthetic data in Section 5.1 of @WibergL1 to investigate the quantitative effectiveness of our proposed method. A total of 100 random low-rank matrices with missing data and corruptions are generated and tested using PARSuMi, Wiberg $\ell_1$ and GRASTA.
In accordance with @WibergL1, the ground truth low rank matrix $W \in \mathbb{R}^{m \times n},m=7,n=12,r=3,$ is generated as $W = UV^T$, where $U\in \mathbb{R}^{m \times r}, V\in \mathbb{R}^{n \times r}$ are generated using uniform distribution, in the range \[-1,1\]. 20% of the data are designated as missing, and 10% are added with corruptions, both at random locations. The magnitude of the corruptions follows a uniform distribution $[-5,5]$. Root mean square error (RMSE) is used to evaluate the recovery precision: $$\label{eq:RMS}
\mathrm{RMSE} := \frac{\Vert W_{\mathrm{recovered}} - W \Vert_F}{\sqrt{mn}}.$$ Out of the 100 independent experiments, the number of runs that returned RMSE values of less than 5 are 100 for PARSuMi, 78 and 58 for Wiberg $\ell_1$ (with two different initializations) and similarly 94 and 93 for GRASTA. These are summarized in Fig. \[fig:Synthetic\_hist\].
![[]{data-label="fig:Synthetic_hist"}](pics/histogram_exp.png){width="8cm" height="4.5cm"}
Recovery under varying level of corruptions, missing data and noise
-------------------------------------------------------------------
To gain a holistic understanding of our proposed method, we perform a series of systematically parameterized experiments on $40\times60$ rank-4 matrices (with the elements of the factors $U,V$ drawn independently from the uniform distribution on $[-1,1]$), with conditions ranging from 0-80% missing data, 0-20% corruptions of range \[-2,2\], and Gaussian noise with $\sigma$ in the range \[0,0.1\]. By fixing the Gaussian noise at a specific level, the results are rendered in terms of phase diagrams showing the recovery precision as a function of the missing data and outliers. The precision is quantified as the difference between the recovered RMSE and the oracle bound RMSE [^10]. As can be seen from Fig. \[fig:phase\_diagram\], our algorithm obtains near optimal performance at an impressively large range of missing data and outlier at $\sigma=0.01$[^11].
For comparison, we also displayed the results for closely related methods, e.g., ALM-RPCA [@Wu_photometric], GRASTA [@he2011grasta], DRMF [@DRMF], LM\_GN [@chen2011hessian] as well as Algorithm \[alg:APG\] (our initialization). Wiberg $\ell_1$ is omitted because it is too slow. Among all the methods we compared, PARSuMi is able to successfully reconstruct the largest range of matrices with almost optimal numerical accuracy. Also, the results for DRMF and LM\_GN are well-expected since they are not designed to handle both missing data and outliers.
SfM with missing and corrupted data on Dinosaur {#sec:expt_outlier}
-----------------------------------------------
In this section, we apply PARSuMi to the problem of SfM using the Dinosaur sequence and investigate how well the corrupted entries can be detected and recovered in real data. We have normalized image pixel dimensions (width and height) to be in the range \[0,1\]; all plots, unless otherwise noted, are shown in the normalized coordinates. To simulate data corruptions arising from wrong feature matches, we randomly add sparse error of the range \[-2,2\][^12] to 1% of the sampled entries. This is a more realistic (and much larger[^13]) definition of outliers for SfM compared to the \[-50,50\] pixel range used to evaluate Wiberg $\ell_1$ in @WibergL1.
We conducted the experiment 10 times each for PARSuMi, Wiberg $\ell_1$ (with SVD initialization) and GRASTA (random initialization as recommended in the original paper) and count the number of times they succeed. As there are no ground truth to compare against, we cannot use the RMSE to evaluate the quality of the filled-in entries. Instead, we plot the feature trajectory of the recovered data matrix for a qualitative judgement. As is noted in @Damped_Newton_2005, a correct recovery should consist of all elliptical trajectories. Therefore, if the recovered trajectories look like that in Fig. \[fig:Traj\_DN\_global\], we count the recovery as a success.
The results are summarized in Table \[tab:SummaryDinosaur\]. Notably, PARSuMi managed to correctly detect the corrupted entries and fill in the missing data in 9 runs while Wiberg $\ell_1$ and GRASTA failed on all 10 attempts. Typical feature trajectories recovered by each method are shown in Fig. \[fig:trajectory comparison\]. Note that only PARSuMi is able to recover the elliptical trajectories satisfactorily.
For comparison, we also include the input (partially observed trajectories) and the results of our convex initialization in Fig. \[fig:Traj\_Input\] and \[fig:Traj\_Convex\] respectively. An interesting and somewhat surprising finding is that the result of PARSuMi is even better than the global optimal solution for data containing supposedly no corruptions (and thus can be obtained with $\ell_2$ method) (see Fig. \[fig:Traj\_DN\_global\], which is obtained under no corruptions in the observed data)! In particular, the trajectories are now closed.
The reason becomes clear when we look at Fig. \[fig:Convex\_Recovery\_Residue\], which shows two large spikes in the vectorized difference between the artificially injected corruptions and the recovered corruptions by PARSuMi. This suggests that there are hitherto unknown corruptions inherent in the Dinosaur data. We trace the two large ones into the raw images, and find that they are indeed data corruptions corresponding to mismatched feature points from the original dataset; our method managed to recover the correct feature matches (left column of Fig. \[fig:Outliers\_in\_the\_Dinosaur\_data\]).
The result shows that PARSuMi recovered not only the artificially added errors, but also the intrinsic errors in the data set. In @Damped_Newton_2005, it was observed that there is a mysterious increase of the objective function value upon closing the trajectories by imposing orthogonality constraint on the factorized camera matrix. Our discovery of these intrinsic tracking errors explained this matter evidently. It is also the reason why the $\ell_2$-based algorithms (see Fig. \[fig:Traj\_DN\_global\]) find a global minimum solution that is of poorer quality (trajectories fail to close loop).
To complete the story, we generated the 3D point cloud of Dinosaur with the completed data matrix. The results viewed from different directions are shown in Fig. \[fig:Point cloud\].
![3D point cloud of the reconstructed Dinosaur.[]{data-label="fig:Point cloud"}](pics/PointCloud/DinosaurPointCloud2.png "fig:"){width="0.47\linewidth"} ![3D point cloud of the reconstructed Dinosaur.[]{data-label="fig:Point cloud"}](pics/PointCloud/DinosaurPointCloud6.png "fig:"){width="0.47\linewidth"}
Photometric Stereo
------------------
Another intuitive application for PARSuMi is photometric stereo, a problem of reconstructing the 3D shape of an object from images taken under different lighting conditions. In the most ideal case of Lambertian surface model (diffused reflection), the data matrix obtained by concatenating vectorized images together is of rank 3.
Real surfaces are of course never truly Lambertian. There are usually some localized specular regions appearing as highlights in the image. Moreover, since there is no way to obtain a negative pixel value, all negative inner products will be observed as zero. This is the so-called attached shadow. Images of non-convex object often also contain cast shadow, due to the blocking of light path. If these issues are teased out, then the seemingly naive Lambertian model is able to approximate many surfaces very well.
@Wu_photometric subscribed to this low-rank factorization model and proposed to model all dark regions as missing data, all highlights as sparse corruptions and then use a variant of RPCA (identical to ) to recover the full low-rank matrix. The solution however is only tested on noise-free synthetic data and toy-scale real examples. @DelBue2012balm applied their BALM on photometric stereo too, attempting on both synthetic and real data. Their contribution is to impose the normal constraint of each normal vector during the optimization.
We compare PARSuMi with the aforementioned two methods on several photometric stereo datasets. Quantitatively, we use the Caesar and Elephant data in @Wu_photometric and compare the reconstructed surface normal against the ground truth. The data is illustrated in Fig. \[fig:CaesarElephant\] and the comparison is detailed in Table \[tab:photometrc\]. As we can see, PARSuMi has the smallest reconstruction error among the three methods in all experiments.
![Illustration of the synthetic data and their surface normal. Note that there are specular regions and shadows.[]{data-label="fig:CaesarElephant"}](pics/PS/CaesarElephant "fig:"){width="\linewidth"}\
Dataset PARSuMi ALM-RPCA [@Wu_photometric] BALM [@DelBue2012balm] Oracle (a lower bound)
---------------------------------- ------------------------ ---------------------------- ------------------------ ------------------------
Elephant **7.13e-2** (16.7 min) 7.87e-2 (1.1 min) 3.55 (1.1 min) model error
Caesar **1.83e-1** (28.6 min) 2.71e-1 (7.2 min) 3.11 (5.2 min) model error
Elephant + $\mathcal{N}(0,0.05)$ **2.35** (28.3 min) 2.62 (1.5 min) 4.37 (1.1 min) 1.70 + model error
Caesar + $\mathcal{N}(0,0.05)$ **2.34** (99.2 min) 2.53 (8.3 min) 4.06 (6.6 min) 1.73 + model error
We also conducted a qualitative comparison of the methods on a real-life data using Subject 3 in the Extended YaleB dataset since it was initially used to evaluate BALM in @DelBue2012balm[^14]. As we do not have any ground truth, we can only compare the reconstruction qualitatively.
From Fig. \[fig:YaleB\_comparison\], we can clearly see that PARSuMi is able to recover the missing pixels in the image much better than the other two methods. In particular, Fig. \[subfig:missing\_data\_comparison\] and \[subfig:missing\_data\_comparison\_zoom-in\] shows that PARSuMi’s reconstruction (in the illuminated half of the face) has fewest artifacts. This can be seen from the unnatural grooves that the red arrows point to in Fig. \[subfig:missing\_data\_comparison\_zoom-in\]. Moreover, we know from the original image that the light comes from the right-hand-side of the subject; thus all the pixels on the left side of his face (e.g. the red ellipse area in Fig. \[subfig:missing\_data\_comparison\_zoom-in\]) should have negative filled-in values and therefore should be dark in the image. Neither BALM nor ALM-RPCA’s reconstructed images comply to this physical law.
To see this more clearly, we invert the pixel values of the reconstructed image in Fig. \[subfig:missing\_data\_comparison\_negative\]. This is equivalent to inverting the direction of lighting. From the tag of the image, we know that the original lighting is $-20^{\circ}$ from the subject’s right posterior and $40^{\circ}$ from the top, so the inverted light should illuminate the left half of his face from $20^{\circ}$ left frontal and $40^{\circ}$ from below. As is shown in the comparison, only PARSuMi’s result revealed what should be correctly seen with a light shining from this direction.
In addition, we reconstruct the 3D depth map with the classic method by @horn1990sfs. In Fig. \[subfig:shape\_comparison\], the shape from PARSuMi reveals much richer depth information than those from the other two algorithms, whose reconstructions appear flattened. This is a known low-frequency bias problem for photometric stereo and it is often caused by errors in the surface normal estimation @nehab2005efficiently. The fact that BALM and ALM-RPCA produces a flatter reconstruction is a strong indication that their estimations of the surface normal are noisier than that of PARSuMi.
From our experiments, we find that PARSuMi is able to successfully reconstruct the 3D face for all 38 subjects with little artifacts. As illustrated in Fig. \[fig:YaleB\_reconstruction\], our 3D reconstructions of the features seem to reveal the characteristic features of subjects across different ethnic groups. Moreover, due to the robust $\ell_0$ penalty, PARSuMi is able to effectively recover the input images from many different types of irregularities, e.g. specular regions, different facial expressions, or even image corruptions caused by hardware malfunctions (see Fig. \[fig:illustration\_of\_PARSuMi\_recovery\] and \[fig:comparison\_yale\_recovery\]). This makes it possible for PARSuMi to be integrated reliably into engineering systems that function with minimal human interactions [^15].
\
\
Speed
-----
The computational complexity of PARSuMi is cheap for some problems but not for others. Since PARSuMi uses LM\_GN for its matrix completion step, the numerical cost is dominated by either solving the linear system $(J^TJ+\lambda I)\delta=J\mathbf{r}$ which requires the Cholesky factorization of a potentially dense $mr\times mr$ matrix, or the computation of $J$ which requires solving a small linear system of normal equation involving the $m\times r$ matrix $N$ for $n$ times. As the overall complexity of $O(\max(m^3r^3, mnr^2))$ scales merely linearly with number of columns $n$ but cubic with $m$ and $r$, PARSuMi is computationally attractive when solving problems with small $m$ and $r$, and potentially large $n$, e.g., photometric stereo and SfM (since the number of images is usually much smaller than the number of pixels and feature points). However, for a typical million by million data matrix as in social networks and collaborative filtering, PARSuMi will take an unrealistic amount of time to run.
Experimentally, we compare the runtime between our algorithm and Wiberg $\ell_1$ method in our Dinosaur experiment in Section \[sec:expt\_outlier\]. Our Matlab implementation is run on a 64-bit Windows machine with a 1.6 GHz Core i7 processor and 4 GB of memory. We see from Table \[tab:SummaryDinosaur\] that there is a big gap between the speed performance. The near 2-hour runtime for Wiberg $\ell_1$ is discouragingly slow, whereas ours is vastly more efficient. On the other hand, as an online algorithm, GRASTA is inherently fast. Examples in @he2011grasta show that it works in real time for live video surveillance. However, our experiment suggests that it is probably not appropriate for applications such as SfM, which requires a higher numerical accuracy.
The runtime comparison for the photometric stereo problems is shown in Table \[tab:photometrc\]. We remark that PARSuMi is roughly ten times slower than other methods. The pattern is consistent for the YaleB face data too, where PARSuMi takes 23.4 minutes to converge while BALM and RPCA takes only 4.8 and 1.7 minutes respectively.
We note that PARSuMi is currently not optimized for computation. Speeding up the algorithm for application on large scale dataset would require further effort (such as parallelization) and could be a new topic of research. For instance, the computation of Jacobians $J_i$ and the evaluation of the objective function can be easily done in parallel and the Gauss-Newton update (a positive definite linear system of equations) can be solved using the conjugate gradient method; hence, we do not even need to store the matrix in memory. Furthermore, since PARSuMi seeks to find the best subspace, perhaps using only a small portion of the data columns is sufficient. If the subspace is correct, the rest of the columns can be recovered in linear time with our iterative reweighted Huber regression technique (see Section \[sec:heuristics\]). A good direction for future research is perhaps on how to choose the best subset of data to feed into PARSuMi.
Conclusion
==========
In this paper, we have presented a practical algorithm (PARSuMi) for low-rank matrix completion in the presence of dense noise and sparse corruptions. Despite the non-convex and non-smooth optimization formulation, we are able to derive a set of update rules under the proximal alternating scheme such that the convergence to a critical point can be guaranteed. The method was tested on both synthetic and real life data with challenging sampling and corruption patterns. The various experiments we have conducted show that our method is able to detect and remove gross corruptions, suppress noise and hence provide a faithful reconstruction of the missing entries. By virtue of the explicit constraints on both the matrix rank and cardinality, and the novel reformulation, design and implementation of appropriate algorithms for the non-convex and non-smooth model, our method works significantly better than the state-of-the-art algorithms in nuclear norm minimization, $\ell_2$ matrix factorization and $\ell_1$ robust matrix factorization in real life problems such as SfM and photometric stereo.
Moreover, we have provided a comprehensive review of the existing results pertaining to the “practical matrix completion” problem that we considered in this paper. The review covered the theory of matrix completion and corruption recovery, and the theory and algorithms for matrix factorization. In particular, we conducted extensive numerical experiments which reveals (a) the advantages of matrix factorization over nuclear norm minimization when the underlying rank is known, and (b) the two key factors that affect the chance of $\ell_2$-based factorization methods reaching global optimal solutions, namely “subspace parameterization” and “Gauss-Newton” update. These findings provided critical insights into this difficult problem, upon the basis which we developed PARSuMi as well as its convex initialization.
The strong empirical performance of our algorithm calls for further analysis. For instance, obtaining the theoretical conditions for the convex initialization to yield good support of the corruptions should be plausible (following the line of research discussed in Section \[sec:MC\_theory\]), and this in turn guarantees a good starting point for the algorithm proper. Characterizing how well the following non-convex algorithm works given such initialization and how many samples are required to guarantee high-confidence recovery of the matrix remain open questions for future study.
Other interesting topics include finding a cheaper but equally effective alternative to the LM\_GN solver for solving , parallel/distributed computation, incorporating additional structural constraints, selecting optimal subset of data for subspace learning and so on. Step by step, we hope this will eventually lead to a practically working robust matrix completion algorithm that can be confidently embedded in real-life applications.
Appendix
========
Proofs
------
Given a subset $I$ of $\{1,\dots,|\Omega|\}$ with cardinality at most $N_0$ such that $b_I\not = 0$. Let $J = \{1,\dots,|\Omega|\} \backslash I$. Consider the problem for $x\in \R^{|\Omega|}$ supported on $I$, we get the following: $$\begin{aligned}
v_I := \min_{x_I} \Big\{ \norm{x_I- b_I}^2 + \norm{b_J}^2 \mid \norm{x_I}^2 -K_E^2\leq 0\Big\},\end{aligned}$$ which is a convex minimization problem whose optimality conditions are given by $$\begin{aligned}
x_I - b_I + \mu \, x_I = 0, \; \mu (\norm{x_I}^2-K_E^2) =0, \; \mu \geq 0\end{aligned}$$ where $\mu$ is the Lagrange multiplier for the inequality constraint. First consider the case where $\mu > 0$. Then we get $x_I=K_E b_I/\norm{b_I}$, and $1+\mu = \norm{b_I}/K_E$ (hence $\norm{b_I} > K_E$). This implies that $
v_I = \norm{b}^2 +K_E^2 - 2 \norm{b_I}K_E.
$ On the other hand, if $\mu=0$, then we have $x_I = b_I$ and $v_I = \norm{b_J}^2 = \norm{b}^2-\norm{b_I}^2$. Hence $$\begin{aligned}
v_I = \left\{ \begin{array}{ll}
\norm{b}^2 +K_E^2 - 2 \norm{b_I}K_E & \mbox{if $\norm{b_I} > K_E$}
\\[5pt]
\norm{b}^2-\norm{b_I}^2 & \mbox{if $\norm{b_I}\leq K_E$.}
\end{array} \right.\end{aligned}$$ In both cases, it is clear that $v_I$ is minimized if $\norm{b_I}$ is maximized. Obviously $\norm{b_I}$ is maximized if $I$ is chosen to be the set of indices corresponding to the $N_0$ largest components of $b$.
\(a) The equality in follows directly from . By the minimal property of $\hx_{k+1}$, we have that $$\begin{aligned}
\widehat{L}(\hx_{k+1};x_k,y_k) \;\leq\; \widehat{L}(\xi;x_k,y_k) \quad \forall\;
\xi\in \cX.
\label{eq-thm1-a2}\end{aligned}$$ Thus when $\xi = x_k$, we get $\widehat{L}(\hx_{k+1};x_k,y_k)\leq \widehat{L}(x_k;x_k,y_k) = L(x_k,y_k)$, and the required inequality in follows. On the other hand, the inequality follows readily from the definition of $x_{k+1}$.\
(b) If $x_{k+1} = \tx_{k+1}$, then from the definition of $x_{k+1}$ and , we have that, $$\label{eq-thm1-b1}
\begin{aligned}
&L(x_{k+1},y_k) + \frac{1}{2}\norm{x_{k+1}-x_k}_S^2\\
\leq& L(\hx_{k+1},y_k) +\frac{1}{2}\norm{\hx_{k+1}-x_k}_S
\leq L(x_k,y_k).
\end{aligned}$$ On the other hand, if $x_{k+1} =\hx_{k+1}$, we have from that $$\begin{aligned}
L(x_{k+1},y_k) + \frac{1}{2}\norm{x_{k+1}-x_k}_S^2
& \leq & L(x_k,y_k).
\label{eq-thm1-b2}\end{aligned}$$ By the minimal property of $y_{k+1}$, we have that $ \forall\;\eta\in\cY$ $$\begin{aligned}
L(x_{k+1},y_{k+1}) + \frac{1}{2}\norm{y_{k+1}-y_k}_T^2
& \leq & L(x_{k+1},\eta) + \frac{1}{2}\norm{\eta-y_k}_T^2.\end{aligned}$$ In particular, when $\eta=y_k$, we get $$\begin{aligned}
L(x_{k+1},y_{k+1}) + \frac{1}{2}\norm{y_{k+1}-y_k}_T^2 \;\leq\;
L(x_{k+1},y_k).
\label{eq-thm1-c}\end{aligned}$$ By combining - and , we get the inequality .\
(c) Note that by using the result in part (b), we also have $\lim_{k'\rightarrow\infty} x_{k'+1} = \bar{x}$ and $\lim_{k'\rightarrow\infty} y_{k'+1} = \bar{y}$. From , and , we have $\forall\; k\geq 0,\; \xi\in\cX$ $$\begin{aligned}
L(x_{k+1},y_k) + \frac{1}{2}\norm{x_{k+1}-x_k}_S^2 \leq \widehat{L}(\xi;x_k,y_k).\end{aligned}$$ Thus $\forall\; \xi\in\cX$ $$\begin{aligned}
\limsup_{k'\rightarrow\infty} f(x_{k'+1}) + q(\bar{x},\bar{y}) \leq
f(\xi) + Q(\xi;\bar{x},\bar{y}).\end{aligned}$$ By taking $\xi = \bar{x}$, we get $$\limsup_{k'\rightarrow\infty} f(x_{k'+1}) \leq
f(\bar{x}) + Q(\bar{x};\bar{x},\bar{y}) - q(\bar{x},\bar{y})
= f(\bar{x}) .$$ On the other hand, since $f$ is lower semicontinuous, we have that $\liminf_{k'\rightarrow\infty} f(x_{k'+1}) \geq f(\bar{x})$. Thus $\lim_{k'\rightarrow\infty}$ $f(x_{k'+1}) = f(\bar{x})$. Similarly, we can show that $\lim_{k'\rightarrow\infty}$ $g(y_{k'+1}) = g(\bar{y})$. As a result, we have $$\begin{aligned}
\lim_{k'\rightarrow\infty} L(x_{k'+1},y_{k'+1}) = L(\bar{x},\bar{y}).\end{aligned}$$ Since $\{ L(x_k,y_k)\}$ is a nonincreasing sequence, the above result implies that $$\lim_{k\rightarrow\infty} L(x_k,y_k) = L(\bar{x},\bar{y}) = \inf_{k} L(x_k,y_k).$$ Also, - and implies that $$\lim_{k\rightarrow\infty} L(x_{k+1},y_k) = L(\bar{x},\bar{y}).$$ From and , we have $$\begin{aligned}
L(x_{k+1},y_k) + \frac{1}{2}\norm{x_{k+1}-x_k}_S^2\\
\;\leq\; \widehat{L}(\hx_{k+1};x_k,y_k)
\;\leq\; L(x_k,y_k).\end{aligned}$$ Thus $\lim_{k\rightarrow\infty} \widehat{L}(\hx_{k+1};x_k,y_k) = L(\bar{x},\bar{y})$.
Now by and again, we have $$\begin{aligned}
\frac{1}{2}\norm{x_{k+1}-x_k}_S^2 + \frac{1}{2}\norm{\hx_{k+1}-x_k}_{M-A^*A-S}^2\\
\leq \widehat{L}(\hx_{k+1};x_k,y_k) - L(x_{k+1},y_k).\end{aligned}$$ Thus $
\lim_{k\rightarrow\infty} \norm{\hx_{k+1}-x_k}_{M-A^*A-S}^2 =0$. Since $M - A^*A-S \succ 0$, we also get $\lim_{k\rightarrow\infty} \norm{\hx_{k+1}-x_k}^2 =0$.\
(d) From the optimality of $\hx_{k+1}$, we have that $$\begin{aligned}
0 &\in \partial \widehat{L}(\hx_{k+1}; x_k,y_k) \\
&= \partial f(\hx_{k+1}) + A^*(A x_{k}+By_k-c) +
M (\hx_{k+1}-x_k)
\\
&= \partial f(\hx_{k+1}) + A^*(A\hx_{k+1}+By_{k+1}-c)
- \Delta x_{k+1}\end{aligned}$$ where $\Delta x_{k+1} = -(M-A^*A) (\hx_{k+1}-x_k) - A^*B(y_k-y_{k+1})$. Thus $$\begin{aligned}
\Delta x_{k+1} \in \partial_x L(\hx_{k+1},y_{k+1}).\end{aligned}$$ From the optimality of $y_{k+1}$, we have that $$\begin{aligned}
0 &\in& \partial g(y_{k+1})
+ B^*(A x_{k+1}+By_{k+1}-c) + T (y_{k+1}-y_k)
\\[5pt]
&=& \partial_y L (\hx_{k+1},y_{k+1}) + T (y_{k+1}-y_k)
+ B^*A (x_{k+1}-\hx_{k+1})\end{aligned}$$ Hence $
\Delta y_{k+1} := - T (y_{k+1}-y_k)-B^*A (x_{k+1}-\hx_{k+1}) \in \partial_y L(\hx_{k+1},y_{k+1}).
$
From part (b) and (c), we have that $$\begin{aligned}
&\lim_{k'\rightarrow \infty} \norm{\hx_{k'+1}-x_{k'}} = 0,&&
\lim_{k'\rightarrow \infty}\norm{\hx_{k'+1}-x_{k'+1}} =0,\\
&\lim_{k'\rightarrow \infty}\hx_{k'+1} = \bar{x},&&
\lim_{k'\rightarrow \infty} y_{k'+1} = \bar{y}.\end{aligned}$$ Thus $$\lim_{k'\rightarrow \infty} \Delta x_{k'+1} = 0 = \lim_{k'\rightarrow \infty} \Delta y_{k'+1}.$$ By the closedness property of $\partial L$ [@clarke1990optimization Proposition 2.1.5], we get $$(0,0) \in \partial L(\bar{x},\bar{y}).$$ Thus $(\bar{x},\bar{y})$ is a stationary point of L.
Software/code used
------------------
The point cloud in Fig. \[fig:Point cloud\] are generated using VincentSfMToolbox [@vincentsSFMToolbox]. Source codes of BALM, GROUSE, GRASTA, Damped Newton, Wiberg, LM\_X used in the experiments are released by the corresponding author(s) of @DelBue2012balm [@balzano2010grouse; @he2011grasta; @Damped_Newton_2005; @WibergL2; @Wu_photometric] and @Subspace_ChenPei_2008. In particular, we are thankful that @balzano2010grouse and @Wu_photometric shared with us a customized version of GROUSE and ALM-RPCA that are not yet released online. For Wiberg $\ell_1$ [@WibergL1], we have optimized the computation for Jacobian and adopted the commercial LP solver: cplex. The optimized code performs identically to the released code in small scale problems, but it is beyond the scope for us to verify for larger scale problems. In addition, we implemented SimonFunk’s SVD ourselves. The ALS implementation is given in the released code package of LM\_X. For OptManifold, TFOCS and CVX, we use the generic optimization packages released by the author(s) of @yin2013orthogonality [@becker2011tfocs; @gb08cvx] and customize for the particular problem. For NLCG, we implement the derivations in @srebro2003weighted and used the generic NLCG package [@nlcg].
Additional experimental results
-------------------------------
Illustration of the decomposition on Subject 3 of Extended YaleB dataset is given in Fig. \[fig:YaleB\_decomposition\]. Additional qualitative comparisons on the recovery of the image is given in Fig. \[fig:comparison\_yale\_recovery\].
\
![Additional comparisons in the quality of face image recovery. From left to right, they are original image, missing data mask (in green), results for PARSuMi, BALM and missing-RPCA.[]{data-label="fig:comparison_yale_recovery"}](pics/PS/Subject2Comparison1 "fig:"){width="\linewidth"}\
![Additional comparisons in the quality of face image recovery. From left to right, they are original image, missing data mask (in green), results for PARSuMi, BALM and missing-RPCA.[]{data-label="fig:comparison_yale_recovery"}](pics/PS/Subject22Comparison1 "fig:"){width="\linewidth"}
The lower bounds in the experiments
-----------------------------------
- The lower bound in Fig. \[fig:mf\_exp1\]: the lower bound is obtained by the data set that contains less than $r$ data points per-column and per-row. It is clear from @kiraly2012algebraic that this is an easy-to-check necessary condition of recoverability.
- *The oracle RMSE for Phase Diagram*: We also adapt the oracle lower bound from @CandesNoise to represent the theoretical limit of recovery accuracy under noise. Our extended oracle bound under both sparse corruptions and Gaussian noise is: $$\label{eq:OracleBound}
\mathrm{RMSE}_{oracle}=\sigma\sqrt{\frac{(m+n-r)r}{p-e}},$$ This is used for benchmarking in our phase diagram experiments.
- *The oracle angular error in Table \[tab:photometrc\]:* For the Caesar and Elephant experiments, we use (ignoring corruptions by taking $e=0$) but transformed it by taking $$\arcsin \sqrt{1-(n\cdot \hat{n})^2},$$ where $\hat{n}$ is the surface normal obtained by an oracle projection of the noisily observed image.
Summary of parameters used in the experiments {#sec:parameters}
---------------------------------------------
- *Parameters in our formulation*: We assume $r$ (the underlying rank) to be known. $N_0$ is chosen to be an upper bound of the number of corrupted entries. In experiments, we use 120% of the actual number of corruptions. In practice, we should choose $N_0=0.1|\Omega|$ or $0.15|\Omega|$. $\epsilon=1e-10$ (almost negligible).$K_E=20\sqrt{N_0\times \mathrm{median}(\cP)_{\Omega}(\widehat{W})}$ (very large, negligible). In theory, we only need $\epsilon>0$ and $K_E<\infty$ to ensure the convergence. In practice, unless it is meaningful to choose an effective $K_E$, we will choose it large enough so that it has no impact on the optimization.
- *Parameters for PARSuMi*: $\beta_1=\beta_2=\frac{1e-3}{\sqrt{\max{m,n}}}$. For Algorithm \[alg:LM\_GN\], $\rho=10$ and initial $\lambda=1e-6$.
- *Parameters for APG*: $\gamma = \frac{1}{\sqrt{\max{m,n}}}$ $\lambda = 0.2$
[^1]: @Candes2011_JACM actually considered missing data too, but their guarantee (Theorem 1.2) for is only preliminary according to their own remarks. A stronger result is released later by the same group in @li2013compressed.
[^2]: Our experiment on synthetic data shows that the strong Wolfe line search adopted by @srebro2003weighted and @yin2013orthogonality somewhat ameliorates the issue, though it does not seem to help much on real data.
[^3]: The cardinality of non-zero entries, which strictly speaking is not a norm.
[^4]: As a reference, we also included nuclear norm minimization that solve where applicable.
[^5]: The lower bound is given by the percentage of randomly generated data that have at least one column or row having less than $r$ samples. Clearly, having at least $r$ samples for every column and row is a necessary condition for exact recovery.
[^6]: When $\alpha=3.5$ in Fig. \[fig:mf\_exp2b\], $r^{th}$ singular value is almost as small as the spectral norm of the input noise.
[^7]: Wiberg algoirthm takes longer time mainly because it sometimes does not converge and exhausts the maximum number of iterations.
[^8]: We will explain this further shortly.
[^9]: Note that a number of false positives persist in the second experiment. This is understandable because false positives often contaminate an entire column or row, making it impossible to recover that column/row in later iterations even if the subspace is correctly detected. To avoid such an undesirable situation, we prefer “false negatives” over “false positives” when tuning Algorithm \[alg:APG\]. In practice, it suffices to keep the initial $E$ relatively sparse.
[^10]: See the Appendix for details.
[^11]: The phase diagrams for other levels of noise look very much like Fig. \[fig:phase\_diagram\]; we therefore did not include them in the paper.
[^12]: In SfM data corruptions are typically matching failures. Depending on where true matches are, error induced by a matching failure can be arbitrarily large. If we constrain true match to be inside image frame $[0,1]$(which is often not the case), then the maximum error magnitude is $1$. We found it appropriate to at least double the size to account for general matching failures in SfM, hence $[-2,2]$.
[^13]: \[-50,50\] in pixel is only about \[-0.1,0.1\] in our normalized data, which could hardly be regarded as “gross” corruptions.
[^14]: The authors claimed that it is Subject 10 [@DelBue2012balm Figure 9], but careful examination of all faces shows that it is in fact Subject 3.
[^15]: For the best of our knowledge, all previous works that use this dataset for photometric 3D reconstruction manually removed a number of images of poor qualities, e.g.@DelBue2012balm
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Inspired by the experimental results of Cuevas [*et al.*]{} (Physical Review Letters [**102**]{}, 224101 (2009)), we consider theoretically the behaviour of a chain of planar rigid pendulums suspended in a uniform gravitational field and subjected to a horizontal periodic driving force applied to the pendulum pivots. We characterize the motion of a single pendulum, finding bistability near the fundamental resonance, and near the period-3 subharmonic resonance. We examine the development of modulational instability in a driven pendulum chain and find both a critical chain length and critical frequency for the appearance of the instability. We study the breather solutions and show their connection to the single pendulum dynamics, and extend our analysis to consider multi-frequency breathers connected to the period-3 periodic solution, showing also the possibility of stability in these breather states. Finally we examine the problem of breather generation and demonstrate a robust scheme for generation of on-site and off-site breathers.'
author:
- 'Y. Xu$^1$'
- 'T. J. Alexander$^{1}$'
- 'H. Sidhu$^1$'
- 'P. G. Kevrekidis$^2$'
title: Instability dynamics and breather formation in a horizontally shaken pendulum chain
---
Introduction {#intro}
============
Driving of the pivot point of a single pendulum has long been known to lead to parametric driving of the pendulum motion (see e.g. [@Nayfeh1979] for a textbook treatment). The majority of research has focused on vertical driving (see e.g. [@BlackburnAJP1992; @GrandyAJP1997] and references therein), where the well-known stabilization of the inverted pendulum may be observed [@StephensonMPMLPS1908]. Horizontal driving has received less attention, with early works considering the nature of the periodic solutions [@StrubleQJMAM1965] and the appearance of subharmonic excitations [@CheshankovJAMM1971]. More recently, interest in this problem has been revived, with recent results exploring the appearance of chaotic motion [@VanDoorenCSF1996], dynamic stabilization of two off-centre equilibrium points [@SchmittND1998] and complex bifurcation behaviour of the period-1 oscillation [@JeongJKPS1999].
In this work, motivated by the experimental setup of [@CuevasPRL2009], we consider a setup of pendula subject to a horizontal driving force. The progression of our study starts with the case of a single pendulum. There, we identify period-1 solutions near the nonlinear resonance and reveal the familiar foldover effect for a driven pendulum and the appearance of subharmonic, period-3, solutions. This is a building block of relevance towards the study of the pendulum chain, to which we then turn our attention. In the latter, we find similar results for the period-1 solutions, albeit with the appearance of instability at a critical chain length. We examine this instability in some detail, identifying the characteristic instability wavenumbers through modulational instability analysis. We then examine the breather solutions in the system, and in agreement with the results of [@CuevasPRL2009] we find families of breather solutions existing in the bistable region. We examine the instability dynamics and breather formation near this region, and provide a prescription for the generation of breather solutions. We also explore the possibility of multi-frequency breathers supported by the subharmonic pendulum response and show that these solutions may persist for long times.
Our presentation will be structured as follows. In section II, we present the relevant model in connection to the experiments of [@CuevasPRL2009], while in section III, we study the single pendulum dynamical features. Section IV extends this to multi-pendulum dynamics and the modulational stability, breather existence/stability, as well as multi-frequency generalizations that can arise. Finally, section V presents our conclusions, as well as some potential future challenges.
Model
=====
![Schematic of experimental arrangement for horizontally shaken pendulum chain showing chain in (a) side and (b) profile views.[]{data-label="fig1"}](fig1.eps){width="0.9\columnwidth"}
The Lagrangian in the absence of damping for a chain of $N$ pendulums subjected to a periodic horizontal displacement of the pivot point with frequency $\omega_d$ and amplitude $A$ (as shown in Fig. \[fig1\]) takes the form [@JoseSaletan]: $$\begin{aligned}
\label{lagrange}
\mathcal{L} &=\sum^{n}_{i=1} \frac{1}{2}(Ml+ m\frac{l}{2})\left[2A\omega_d \sin(\omega_d t)\cos \theta_i \dot{\theta_i} \right] \\ &+ \frac{1}{2}I\dot{\theta_i}^2 + (Ml+m\frac{l}{2}) g\cos\theta_i\nonumber \\
&-\frac{1}{2}\beta\left[(\theta_i-\theta_{i-1})^2+(\theta_i-\theta_{i+1})^2\right] \nonumber\end{aligned}$$ where $m$ is the mass of the thin rod of length $l$ supporting the pendulum bob of mass $M$, and $\theta_i$ and $\dot{\theta}_i$ are the angle relative to vertical and angular speed respectively for the $i$-th pendulum. The moment of inertia is given by $I = Ml^2 + ml^2/3$, $g$ is the acceleration due to gravity and $\beta$ is the linear coupling between pendulums through a torsion spring. We consider the case with end pendulums defined by taking $\theta_{N+1} = \theta_N$, $\dot{\theta}_{N+1} = \dot{\theta}_N$ and $\theta_0 = \theta_1$,$\dot{\theta}_0 = \dot{\theta}_1$.
As discussed in Ref. [@CuevasPRL2009] in a physical system there is also on-site damping due to velocity-dependent friction (due to air resistance) and intersite damping due to frictional loss in the torsion spring. The equations of motion therefore take the form [@CuevasPRL2009]: $$\begin{aligned}
\label{physmod}
\ddot{\theta}_n&- \frac{\beta}{I}\left(\theta_{n+1}+\theta_{n-1}-2\theta_{n}\right)+\frac{\tilde{\gamma}_1}{I}\dot{\theta}_n\nonumber\\
&-\frac{\tilde{\gamma}_2}{I}\left(\dot{\theta}_{n+1}+\dot{\theta}_{n-1}-2\dot{\theta}_n\right)+\omega_0^2\sin(\theta_n)\nonumber\\
&+f\omega_d^2\cos(\omega_d t)\cos(\theta_n)=0\end{aligned}$$ where the natural pendulum frequency is given by $\omega_0^2 = (mgl/2 + Mgl)/I$ and the dimensionless forcing coefficient is $f = A\omega_0^2/g$. We use parameter values consistent with the experimental setup of Ref. [@CuevasPRL2009]: $\beta = 0.0165$ Nm/rad, $m = 13$ g, $l = 25.4$ cm, $A = 1.12$ cm, $\tilde{\gamma}_1 = 284$ g cm$^2$/s and $\tilde{\gamma}_2 = 70$ g cm$^2$/s. The driving frequency $\omega_d$ is an experimental control parameter. In principle the pendulum bob mass can also be varied, however we take $M = 1.8$ g. The model can be further simplified by scaling time using the natural frequency, $t \rightarrow t/\omega_0$, to reach the dimensionless form: $$\begin{aligned}
\label{normmod}
\ddot{\theta}_n&- c\left(\theta_{n+1}+\theta_{n-1}-2\theta_{n}\right)+\gamma_1\dot{\theta}_n\nonumber\\
&-\gamma_2\left(\dot{\theta}_{n+1}+\dot{\theta}_{n-1}-2\dot{\theta}_n\right)\nonumber\\
&+\sin(\theta_n)+f\omega^2\cos(\omega t)\cos(\theta_n)=0\end{aligned}$$ where $c = \beta/(I\omega_0^2)$, $\omega = \omega_d/\omega_0$ and $\gamma_i = \tilde{\gamma_i}/(I\omega_0)$. In this dimensionless model the parameters take the values: $c=0.799$, $f=0.060$, $\gamma_1 = 0.010$, $\gamma_2=0.0024$. We consider the effect of varying the frequency ratio $\omega$, and to a lesser extent the forcing amplitude $f$ (corresponding to a variation of the lateral driving amplitude $A$ in the physical parameters).
Single pendulum dynamics
========================
Earlier analysis of a horizontally shaken single pendulum has revealed the possibility of chaotic dynamics [@VanDoorenCSF1996] and oscillations about a nonzero equilibrium point [@SchmittND1998]. We are primarily interested in exploring some of the dynamics possible in the experimental set-up of Ref. [@CuevasPRL2009], so we begin by examining the period-1 solutions within the experimentally accessible parameter space. To this end we consider Eqs. (\[normmod\]) in the limit of a single pendulum, i.e. $N = 1$, $c=\gamma_2 = 0$.
Period-1 solutions
------------------
In the presence of damping and low amplitude forcing we expect to find regular solutions oscillating with the period of the driving force. We search for these periodic solutions numerically using the classical shooting method (see, e.g., the discussion in Ref. [@VanDoorenCSF1996]), seeking a solution across one forcing period. The advantage of the shooting method is that the eigenvalues of the correction matrix (used in the shooting iterations) correspond to the Floquet multipliers of the periodic solutions and so determine the stability of the converged solution. A periodic solution whose Floquet multipliers have a magnitude less than or equal to 1 is stable.
Proceeding with our numerical method we find bistability in a region near the resonance, as is evident in Fig. \[singfam1\](a), where we see the familiar foldover effect for a driven sinusoidal pendulum (the resonance drifts to lower frequencies with higher amplitude, due to the softening nature of the sinusoidal nonlinearity). We see that in the region $\omega = [0.718, 0.9385 ]$, we have two stable solutions, one at high amplitude, the other at low amplitude. As evident in the inset in Fig. \[singfam1\](a) the low amplitude solution is almost exactly out of phase with the force, while the high amplitude solution is closer to being in phase. The phase here is calculated as: $$\label{phase_per1}
{\rm phase} = \frac{\omega}{2\pi}{\rm mod}(t_{\theta_{max}},2\pi/\omega)$$ where $t_{\theta_{max}}$ is a time at which the pendulum has maximum amplitude. This calculation for the phase thus indicates when the pendulum is at maximum amplitude relative to the forcing period (for instance a value of 0.5 means it is exactly out of phase with the force).
The stability of the solutions follows from the maximum absolute values of the Floquet multipliers, shown in Fig. \[singfam1\](b). Both the high and low amplitude solutions have magnitudes less than 1 and so are stable. The inset in Fig. \[singfam1\][b]{} shows details of the spectra at $\omega = 0.8$. In Fig. \[singfam1\](c) we show examples of the stable solutions at $\omega = 0.8$, with the phase and amplitude relationships evident.
![(color online). Families of period-1 solutions in the driven system at low forcing amplitude. $f=0.0597,~\gamma_1=0.01$. (a) The amplitudes of stable and unstable solutions (solid and dashed lines respectively) as a function of $\omega$ with the inset showing the corresponding phase; (b) the maximum absolute value of the corresponding Floquet multipliers and the inset showing the spectrum corresponding to the three different solutions at $\omega=0.8$ (triangles down: dashed line family; triangles to side: thin solid line family; triangles up: thick solid line family); (c) the appearance of the two stable solutions corresponding to, respectively, nearly in-phase (thin solid line family in (a) and (b)) and out-of-phase (thick solid line family in (a) and (b)). For comparison, force is shown as a dashed line.[]{data-label="singfam1"}](fig2.eps){width="0.95\columnwidth"}
While it is not a focus of our work, it is interesting to compare the resonance picture at large forcing amplitude $f = 1$, which has more in common with the analysis of [@VanDoorenCSF1996]. As can be seen in Fig. \[singfam2\](a) the amplitude response exhibits a cross-over point, and a significantly more complex spectrum (Fig. \[singfam2\](b)). The phase response (Fig. \[singfam2\](a)\[inset\]) is still much the same as that observed at low forcing amplitude. As a final point, we note that the high amplitude solution becomes unstable at higher frequency, unlike the low forcing case.
![(color online). Family of period-1 solutions in the driven system at large forcing amplitude, $f=1.0,~\gamma_1=0.01$. (a) The amplitudes of stable and unstable solutions (solid and dashed lines respectively) with the inset showing the corresponding phase; (b) the maximum absolute value of the stability matrix eigenvalues.[]{data-label="singfam2"}](fig3.eps){width="0.95\columnwidth"}
Approximate solution {#sec_approx}
--------------------
Looking ahead to our multi-pendulum analysis we seek an approximate solution for the period-1 response. We find that we can obtain a good prediction of this response by simply assuming the solution takes the harmonic form: $$\label{ansatz}
\theta = V_{c} \cos(\omega t+\phi),$$ where $V_c$ is the amplitude of the pendulum oscillation and $\phi$ is the phase offset from the driving phase $\omega t$. To proceed, we make two assumptions which allow us to simplify the problem. Firstly, we simplify our equation of motion Eq. (\[normmod\]) by replacing the trigonometric nonlinearities with their low order algebraic expansions, i.e. $\sin\theta \approx \theta-\frac{1}{6}\theta^{3}$ and $\cos\theta \approx 1-\frac{1}{2}\theta^{2}$. This leads to the new (approximate) equation of motion:
$$\ddot{\theta}+\gamma_1\dot{\theta}+(\theta-\frac{1}{6}\theta^{3})+f\omega^2\cos(\omega t)(1-\frac{1}{2}\theta^{2})=0
\label{simplifiedmodel}$$
Secondly, remembering the trigonometric identity $\cos^3(\omega t+\phi) = \frac{3}{4}\cos(\omega t + \phi) + \frac{1}{4}\cos(3(\omega t + \phi))$, we assume all contributions not at the frequency of the driving force are weak and can be neglected (the so-called rotating wave approximation). Incorporating this assumption into our simplified model (\[simplifiedmodel\]) we obtain two equations for $(V_c,\phi)$ by multiplying by $\cos(\omega t)$ and $\sin(\omega t)$ respectively and integrating out the time dependence (integration over $[0,2\pi/\omega]$):
$$\begin{aligned}
&-\frac{\pi V_c}{8\omega}\Bigg( 2\omega^2fV_c\cos(\phi)^2+f\omega^2V_c-8f\omega^2+8\gamma_1\omega \sin(\phi)\notag\\ &+8\omega^2\cos(\phi)-8\omega_0^2\cos(\phi)+V_c^2\omega_0^2\cos(\phi)\Bigg) =0
\label{eq1Vphi} \\
&\frac{\pi V_c}{8\omega}\Bigg( 8\omega^2\sin(\phi)-8\omega_0^2\sin(\phi)+2\omega^2 f V_c\cos(\phi)\notag\\ &+V_c^2\omega_0^2\sin(\phi)-8\gamma_1\omega\cos(\phi)\Bigg) =0
\label{eq2Vphi}\end{aligned}$$
We solve these equations using the symbolic mathematics package ${\rm Maple}^{\texttrademark}$, after setting values for $\gamma_1$ and $\omega$. The results are shown in Fig. \[comparison\] with $\gamma_1 = 0.01$ and $\omega$ varying, showing good qualitative agreement with the numerical results, even when $V_c$ is large.
![(color online). Comparison between single-pendulum period-1 amplitudes found numerically (lines), and using the ansatz (\[ansatz\]), (circles), for $f=0.0597,~\gamma_1=0.001$.[]{data-label="comparison"}](fig4.eps){width="0.9\columnwidth"}
Dynamical response
------------------
We turn now to the dynamical response of the single pendulum with the non-zero initial condition $\theta(0) = 0.5$ and $\dot{\theta}(0) = 0$. We represent the dynamical response as a bifurcation plot, plotting all values of $\theta$ sampled at the forcing frequency (between $t = 4000$ and $t = 5000$ to allow initial transients to die out), for a given forcing frequency $\omega$. A single value of $\theta$ indicates a period-1 solution. Multiple values for a given $\omega$ indicate a longer period solution. At the low forcing amplitude of $f = 0.0597$ we see in Fig. \[bif\](a) a relatively simple plot, with the generation of a single period solution from the given initial condition for most of the frequency values, except around $\omega= 3$ where it appears the dynamics has converged to a multi-period solution. Note that this has only appeared in the dynamics because we have chosen a non-zero initial condition. For small (or zero) initial amplitudes we find instead the dynamics converge to the low amplitude solution. The jump around $\omega = 0.9$ is due to the change in convergence from the low amplitude solution to the high amplitude solution (thus the jump indicates the presence of bistability).
By way of contrast we consider also the dynamics at large forcing amplitude $f = 1.0$ in Fig. \[bif\](b), where we see the appearance of aperiodic solutions, which we surmise to be chaotic, for large ranges of $\omega$. A deeper analysis of the nature of the dynamics in this regime, and the possible routes to chaos, is beyond the scope of this work.
![Bifurcation plot of $\theta$ vs $\omega$ constructed from a fixed initial condition $\theta = 0.5$, $\dot{\theta} = 0$, where $\theta(t)$ is plotted every $t = 2\pi/\omega$ from $t = 4000$ to $t = 5000$. (a) $f = 0.059$ showing appearance of period-1 and multi-period solutions and (b) $f = 1.0$ showing predominantly aperiodic dynamics.[]{data-label="bif"}](fig5.eps){width="0.95\columnwidth"}
Period-3 solutions
------------------
We find that the jump observed near $\omega = 3$ in Fig. \[bif\](a) is due to the existence of a stable three-period solution. This subharmonic response is well known in the pendulum system (see e.g. [@ChesterJIMA1975], or for a textbook treatment [@Nayfeh1979]). To examine in more detail this solution we turn again to our numerical shooting method but this time seek solutions with period $T=\frac{6\pi}{\omega}$. We find the existence of a saddle-node bifurcation near $\omega = 3$ and the appearance of lower amplitude (unstable) and higher amplitude (stable) solutions, as can be seen in Fig. \[tripsol\]. Perhaps somewhat surprisingly the stability region in driving frequency for the upper branch solution is large, extending to less than $\omega = 2$. The two solutions are of much larger amplitude than the period-1 solutions, and appear to form an isola (as seen in [@ChesterJIMA1975]), well separated from the low amplitude solution.
![(color online). Period-3 solutions (isola) compared with period-1 solutions (fold-over). The dashed and solid lines correspond to unstable and stable solutions respectively. \[Inset\] Zoom in showing the left-hand bifurcation point.[]{data-label="tripsol"}](fig6.eps){width="0.95\columnwidth"}
Multi-pendulum dynamics
=======================
We now turn to the multi-pendulum case, i.e. the system (\[normmod\]) with $N>1$. We begin by examining the fundamental oscillation mode (all pendulums synchronised) and find similar results to those of the single-pendulum case, albeit with the emergence of instability for $N \ge N_{cr}$. We then turn our attention to the nature of the breather solutions (localized energy states), examine their connection to the fundamental mode solution and explore methods for their generation.
Fundamental oscillation mode {#sec:numerical}
----------------------------
We find that the period-1 solutions mirror those found in the single-pendulum case, except for the appearance of instability in the large amplitude solution beyond a critical chain length. We show in Fig. \[multifam1\] the dependence of the period-1 solutions on $\omega$ for $N = 41$. The solid lines indicate stable solutions and the dashed lines indicate unstable ones. The instability in this case appears at $\omega = 1.07$. As can be seen in the inset of Fig. \[multifam1\](b) the instability appears through the collision of complex eigenvalues producing real eigenvalues with magnitude greater than 1.
![(color online). Family of period-1 solutions in the driven system at low forcing amplitude for $N = 41$. $f=0.0597,~\gamma_1=0.001, ~\gamma_2=0.020,~c=0.799$ showing (a) amplitude response and (b) associated maximum instability eigenvalue with \[inset\] spectrum on either side of the stability change ($\omega = 1.06$ and $\omega = 1.07$) for the high amplitude solution.[]{data-label="multifam1"}](fig7.eps){width="0.95\columnwidth"}
In Fig. \[instabilityregion\] we compare the results for chains with different numbers of pendulums, and find the critical chain length for the appearance of the instability is $N_{cr} = 4$. Also evident in Fig. \[instabilityregion\] is a suggestion of convergence to a maximum $\omega$ beyond which we have stability, even for longer chains. The instability threshold for $N = 20$ is $\omega = 1.07$, while for $N = 100$ it is $\omega=1.14$. We should note that this analysis is focused only on the solutions for which all pendulums behave identically. The full bifurcation picture is expected to be significantly more complicated (as for instance seen in a recent work considering coupling between two Duffing oscillators [@IkedaJCND2013]). The detailed bifurcation study is beyond the scope of this work, instead we study the nature of the instability development for the fundamental oscillation mode at long chain lengths, through modulational instability analysis.
![(color online). Appearance of maximum instability eigenvalue for the period-1 solution at different chain lengths, showing the appearance of a critical chain length for the appearance of instability. Also evident is a critical value of driving frequency $\omega$ beyond which the period-1 solution appears to be stable for all chain lengths.[]{data-label="instabilityregion"}](fig8.eps){width="0.95\columnwidth" height="0.20\textheight"}
Modulational Instability Analysis
---------------------------------
We look for the emergence of modulational instability (MI) in the oscillator chain, using standard MI analysis, as for instance carried out for chains of periodically-forced anharmonic oscillators [@BurlakovPRL1998].
We begin with the approximate solution $\theta_n = V_c \cos(\omega t + \phi)$, equivalent to the single pendulum solution calculated in Section \[sec\_approx\]. We then add a perturbation to this solution in the form: $$\begin{aligned}
\delta\theta(n) &= \frac{1}{2}\cos(qn)\Bigg( V_{p1}\exp[i(\omega-\Omega)t] \\\nonumber
&+V_{p2}\exp[i(-\omega-\Omega)t]\Bigg)
\nonumber
\\
&+ C.C.
\label{perturbation}\end{aligned}$$ where $V_{pj}$ is the complex amplitude, and $q$ and $\Omega$ are the vector and (in principle) complex frequency of the perturbation, respectively. Substituting $\theta_n + \delta\theta_n$ into (\[normmod\]) and keeping only terms to first order in the perturbation (\[perturbation\]), we obtain an equation for $V_{p1}$ and for $V_{p2}$. Multiplying these two equations together and integrating over a single forcing period $[0,2\pi/\omega]$ to remove the time dependence we end up with the following equation relating the instability wavenumber $q$ and growth rate $\Omega$: $$\begin{aligned}
\Bigg[ &-(\omega-\Omega)^2+ i\gamma_1(\omega-\Omega)-c(2\cos(q)-2)\\\notag
&+\omega_0^2(\omega-\Omega)-2i\gamma_2(\cos(q)-1)-\frac{1}{4}\omega_0^2 V_c^2 \Bigg]\\\notag
&\times \Bigg[ -(-\omega-\Omega)^2-i\gamma_1(\omega+\Omega)-c(2\cos(q)-2)\\\notag &+\omega_0^2+2i\gamma_2(\cos(q)-1)(\omega+\Omega)-\frac{1}{4}\omega_0^2 V_c^2 \Bigg]\\\notag &=\frac{1}{32} \Bigg( \omega_0^4 V_c^2+4\omega_0^2V_c f\omega^2\cos(\phi) \\\notag &+4f^2\omega^4+8f^2\omega^2\cos(\phi)^2 \Bigg)
\label{MIeq}\end{aligned}$$ To proceed, we substitute the $V_c$ and $\phi$ associated with the fundamental mode of interest at a given $\omega$ (found from solving Eqs.(\[eq1Vphi\]) and (\[eq2Vphi\])), and solve for $\Omega$ at a given $q$. We see from the nature of the perturbation (\[perturbation\]) that for any $\Omega$ with a positive imaginary part we will have growth of the perturbation, and a resultant instability with wavenumber given by $q$.
As we can see in Fig. \[qOm\_unstable\] for $\omega = 1.09$ (top panel) we find instability for a small range of wavenumbers around $q = 0.165\pi$. This suggests the instability development will progress with a characteristic periodicity corresponding to about 12 pendula. As we move deeper into the unstable region we see in Fig. \[qOm\_unstable\] for $\omega= 0.95$ (bottom panel) that the interval of unstable wavenumbers increases, as well as the instability growth rate. Examining the dependence of the maximum growth rate $\operatorname{Im}\Omega$ versus $\omega$ we see in Fig. \[mapleinstab\] some oscillations around the critical threshold, but extended stability above $\omega = 1.15$, agreeing well with the numerical prediction of $\omega = 1.14$ shown in Fig. \[instabilityregion\] for 100 pendulums.
![(color online). (a) The maximal growth rates associated with the modulational stability analysis when (a) $\omega=1.09$ and (b) $\omega=0.95$.[]{data-label="qOm_unstable"}](fig9.eps){width="0.9\columnwidth"}
![(color online). The maximum imaginary part of $\Omega$ variation as a function of the driving frequency $\omega$, showing extended stability beyond $\omega > 1.15$.[]{data-label="mapleinstab"}](fig10.eps){width="0.9\columnwidth"}
Instability dynamics
--------------------
We turn now to an investigation of the time-dependent response of the pendulum chain, i.e. evolution of system (\[normmod\]) given initial conditions for each pendulum amplitude and velocity. While there are some interesting questions concerning the basins of attraction of the various periodic solutions we have found, we shall consider here only the dynamics following from the initial conditions $\theta_n(0) = \dot{\theta}_n(0) = 0$.
Firstly, we see in Fig. \[mi\_dynom1p09\] for $\omega = 1.09$ the spontaneous symmetry breaking accompanying the appearance of modulational instability. We monitor the energy per pendulum to visualize the dynamics, with the associated energy density given by: $$\label{energy}
E_n = \frac{1}{2}\dot{\theta}_n^2 - \cos(\theta_n) + c\frac{1}{2}\left[(\theta_{n+1} - \theta_n)^2 + (\theta_{n} - \theta_{n-1})^2\right]$$ Fig. \[mi\_dynom1p09\](a) shows the energy per pendulum as a function of time $t$ and pendulum number $n$ with white corresponding to $E_n = -0.8$ and black $E_n = -1$. As can be seen after some initial transient time the symmetry is spontaneously broken through energy localization. We can see in Fig. \[mi\_dynom1p09\](b) that the periodicity of the emergent pattern is roughly 12 pendula, as was earlier predicted by our modulational instability analysis.
![(color online). Development of modulational instability at $\omega = 1.09$ for a chain of 41 pendulums. (a) Plot of total energy per pendulum (black: -1, white: -0.8), showing formation of regular high energy peaks. (b) Amplitude at dashed line in (a), given by $t = 4001$ (up triangles), $t = 4002$ (right triangles), $t = 4003$ (down triangles) and $t=4004$ (left triangles) showing larger oscillations at high energy areas. Lines through the pendulum amplitudes are shown to aid the eye. Initial conditions $\theta_n = 0$, $\dot{\theta}_n = 0$[]{data-label="mi_dynom1p09"}](fig11.eps){width="1.0\columnwidth"}
Our MI analysis also showed that decreasing the driving frequency increases both the instability growth rate and the range of unstable wavenumbers. To explore the effect of this in the dynamics we consider $\omega = 0.95$ with 101 pendulums. As can be seen in Fig. \[mi\_dynom0p95\] the instability appears much more rapidly (in comparison with Fig. \[mi\_dynom1p09\]), and without a clearly dominant wavenumber to the instability, as may be expected from the wide interval of unstable wavenumbers, including many with roughly similar growth rates. Instead, we see highly nonstationary dynamics and apparent energy localization. This, in turn, motivates us to study in Section \[breather\] the localized states spontaneously arising in the instability evolution.
![(color online). Rapid development of instability at $\omega = 0.95$ for a chain of 41 pendulums. (a) Plot of total energy per pendulum (black: -1, white: 7), showing formation of high energy excitations. (b) Amplitude at dashed line in (a), given by $t = 800$ (up triangles), $t = 801$ (right triangles), $t = 802$ (down triangles) and $t=803$ (left triangles) showing a transient localized excitation. Lines through the pendulum amplitudes are shown to aid the eye. Initial conditions $\theta_n = 0$, $\dot{\theta}_n = 0$[]{data-label="mi_dynom0p95"}](fig12.eps){width="1.0\columnwidth"}
Finally we consider the emergence of instability in short chain lengths, below the 12 pendulum periodicity predicted by our MI analysis. Our earlier analysis of the pendulum chain indicated that the large amplitude state becomes unstable when consisting of four pendulums, for frequencies below approximately $\omega = 0.9$ (see Fig. \[instabilityregion\]). In practice at this driving frequency we find the pendulum chain will converge to the stable low amplitude solution (which exists up until approximately $\omega = 0.938$). The instability region increases with chain length and in fact already a chain of five pendulums is unstable beyond the low amplitude existence region. We see in Fig. \[MInpen5\] the instability dynamics which emerge at $\omega = 0.94$. The instability pattern of the long chain is gone and instead we see that the energy appears to move through the chain becoming temporarily trapped at the ends. It is important to note that the modulational stability analysis above applies to the infinite chain, since the wavenumber parameter $q$ was taken to be a continuous variable. An interesting direction for future work would be to consider separately the case of short chains, with $q$ suitably quantized.
![(color online). Instability development in a short chain of five pendulums at $\omega = 0.94$. No stationary pattern is evident, instead energy appears to move through the chain and become temporarily trapped at the end pendulums. The initial conditions are $\theta_n = 0$, $\dot{\theta}_n = 0$. Colormap indicates total energy per pendulum (white: 1.76, black: -1)[]{data-label="MInpen5"}](fig13.eps){width="1.0\columnwidth"}
Period-1 Breather solutions {#breather}
---------------------------
The energy localization we have observed in our instability dynamics appears somewhat connected to the breather solutions observed in Ref. [@CuevasPRL2009]. Thus, we continue our analysis by examining this problem, namely the form and stability of the period-1 breather solutions. In agreement with the results of Ref. [@CuevasPRL2009] we find two families of breather solutions corresponding to on-site (Fig. \[breather\_onsite\]) and off-site (Fig. \[breather\_offsite\]) configurations. We find these solutions by beginning in the anti-continuous limit [@MarinN1996], with the central site(s) corresponding to the high amplitude single pendulum solution and outer sites the low amplitude solution. Both families are only stable for a limited, and (roughly) mutually exclusive, range of driving frequencies $\omega$, with the onsite state being stable at higher frequencies.
Examining the Floquet multipliers at the instability crossing we see that the on-site breather becomes unstable due to the growth of a complex conjugate pair (Fig. \[breather\_onsite\](c)), while the off-site breather becomes unstable through the growth of a purely real Floquet multiplier (Fig. \[breather\_offsite\](c)).
![(color online). Family of onsite breathers for $N = 41$. (a) amplitude of central site (blue) and tails (red); (b) maximum eigenvalue; (c) the eigenvalues from stable and unstable breathers in the complex plane.[]{data-label="breather_onsite"}](fig14.eps){width="0.95\columnwidth"}
![(color online). Family of off-site breathers for $N = 41$. (a) amplitude of central site (blue) and tails (red); (b) maximum eigenvalue; (c) the eigenvalues from stable and unstable breathers in the complex plane.[]{data-label="breather_offsite"}](fig15.eps){width="0.95\columnwidth"}
The small range of $\omega$ for which the solutions exist is directly related to the region of bistability in the single pendulum case. This can be most clearly seen by superimposing a plot of the amplitudes of the breather maximum and tails on the single pendulum amplitude response shown originally in Fig. \[singfam1\](a). We show this in Fig. \[breathervsfamily\] where the maximum amplitudes of the central site(s) for the two breather families (upper dashed lines, top: on-site, lower:off-site) is shown along with the regions over which these families are stable (upper circles) and the amplitudes of the breather tails (lower circles), superimposed on the pendulum chain results (lower solid and dashed lines). It is evident in this figure that the breathers only exist in the region of bistability for the period-1 solutions of the pendulum chain. Interestingly there is a region in $\omega$ approximately $[0.94, 1.12]$ where we find no stable solutions. We will exploit this fact in our breather generation scheme discussed in Section \[generation\].
![(color online). Breather solution for $N = 41$ compared with period-1 solutions. The green and magenta lines stand for the on-site and off-site breather respectively, and the circles mean stable states.[]{data-label="breathervsfamily"}](fig16.eps){width="0.9\columnwidth"}
We finish this section by briefly examining the instability dynamics of an unstable breather. As an example we consider the on-site breather at $\omega = 0.9$. For this driving frequency the instability growth rate is weak, and we expect still connected with complex conjugate Floquet multipliers. The resultant break-up dynamics of the breather shown in Fig. \[breathinstab\] is indeed slow, showing amplitude oscillations and gradual emission of energy into the surrounding pendulum chain. By $t= 4000$ the tails of the breather are showing significant amplitude oscillations (see Fig. \[breathinstab\](b)), yet the energy is still predominantly localized about the central site of the pendulum chain, indicating the weak nature of the instability.
![(color online). Dynamics of unstable on-site breather at $\omega = 0.9$ for a chain of 41 pendula. (a) Plot of total energy per pendulum (black: -1, white: 4.8), showing large amplitude oscillations of central sites. (b) Amplitude at dashed line in (a), given by $t = 4001$ (up triangles), $t = 4002$ (right triangles), $t = 4003$ (down triangles) and $t=4004$ (left triangles) showing presence of large amplitude oscillations in breather tails.[]{data-label="breathinstab"}](fig17.eps){width="0.95\columnwidth"}
Mixed-frequency breather solutions
----------------------------------
Inspired by the connection between the period-1 single-pendulum solutions and the nature of the period-1 breathers we also examine the possibility of breathers connected to the period-3 single-pendulum solutions. In particular, the possibility of breathers in the bistable region of low amplitude period-1 solutions and high amplitude period-3 solutions. We find that indeed breathers exist in this bistable region, with the tails connected to the period-1 solution and the central site(s) connected to the period-3 solution (see Fig. \[mperbreath\]).
![Mixed-frequency breather trajectories across one period ($T = 6\pi/\omega$) for every pendulum ($N = 101$) with $\omega=2.7$. (a) Central pendulums moving at three times the period of the driving force, and out of phase with pendulums at edges (weakly unstable); (b) as above, except central pendulums in phase with tails (strongly unstable).[]{data-label="mperbreath"}](fig18.eps){width="0.95\columnwidth"}
Due to their mixed frequency nature more breather families become possible. In particular, we find that the central sites (moving with one third the frequency of the driving force) may move either out of phase or in phase with the low amplitude tails (moving with the same frequency as the driving force). These two classes are shown in Fig. \[mperbreath\](a) and \[mperbreath\](b) respectively, where we have plotted $\theta(t)$ for each pendulum over three forcing periods, for $\omega = 2.7$ and $N = 101$. Such, so-called subharmonic breathers have also been found in the context of electrical lattices recently in [@lars2]. Examining the Floquet spectrum for these families we find that the in-phase state is highly unstable (dashed line in Fig. \[mperfam\]) while the out-of-phase state is only weakly unstable (solid line in Fig. \[mperfam\]).
![Absolute value of maximum Floquet multiplier for multi-frequency breather solutions in a chain of 101 pendula. Solid line corresponds to the out-of-phase solution, showing weak instability, while the dashed line corresponds to the in-phase solution showing strong instability.[]{data-label="mperfam"}](fig19.eps){width="1.0\columnwidth"}
While we find no stable examples of multi-frequency breathers for a chain of 101 pendulums, at shorter chain lengths windows of stability appear. We show in Fig. \[mperfamN41\] the dependence on driving frequency $\omega$ for the family of on-site out-of-phase mixed-frequency breathers in a chain of 41 pendulums. The central site amplitude (solid line) and edge site amplitude (dashed line) are shown in Fig. \[mperfamN41\](a), with the largest Floquet multiplier shown in Fig. \[mperfamN41\](b). Around $\omega = 2.7$ we see that this mixed-frequency breather is predicted to be stable. As we approach $\omega = 3$ the mixed-frequency breather becomes more and more delocalized, evident in the apparent collision of the central site and tail amplitudes. We also see some evidence of symmetry breaking in the tail oscillators, which is perhaps the origin of the small loss in continuity of the breather family amplitude near cut-off. The full analysis of the properties of this breather state are beyond the scope of this work, but would be an interesting direction for future work.
![Multi-frequency breather solutions for $N=41$. (a) Central site amplitude (solid line) and edge site amplitude (dashed line) versus $\omega$ for on-site, out-of-phase, breather solution; (b) Maximum absolute value of the instability eigenvalues associated with solutions in (a) showing small windows of stability near $\omega = 2.7$.[]{data-label="mperfamN41"}](fig20.eps){width="0.95\columnwidth"}
To confirm the stability results we turn to simulations of the dynamics. In Fig. \[mperdyn\](a) we see the effect of the weak instability on the on-site out-of-phase breather for $N=101$. After an extended period with little change the breather suddenly evaporates, with the system converging to the in-phase period-1 solution for all pendulums. In contrast, at $\omega = 2.7$ for the breather in a chain with $N=41$ we see persistence over long simulation times (see Fig. \[mperdyn\](b)) suggesting the state is stable, in agreement with the stability results shown in Fig. \[mperfamN41\].
![(color online). Dynamics of multi-frequency breathers for different chain lengths. (a) Weakly unstable case for $\omega = 2.7$ and $N = 101$ showing disappearance of breather (black: -1, white: 2.5); (b) Very long lived multi-frequency breather for $\omega = 2.7$ and $N = 41$ (black: -1, white: 1).[]{data-label="mperdyn"}](fig21.eps){width="0.94\columnwidth"}
Breather generation {#generation}
-------------------
An important issue concerning breathers is how to generate them in practice. We consider one possible scheme involving modulation of the driving frequency. We propose starting with all pendulums at rest and driving at a frequency at which modulational instability emerges with a well-defined instability wavenumber. In this way a regular pattern emerges spontaneously in the system. We then suggest lowering the frequency until it sits within the stable region of the target breather type. In Fig. \[breathgen\] we show generation of a stable on-site breather, through variation of $\omega$ from $\omega = 1.04$ down to $\omega = 0.84$ (as depicted in Fig. \[breathgen\](b)). If we lower the frequency further we sit in the band of stable off-site breathers, and we see in Fig. \[breathgenoff\] that similarly we can generate members of this family of breathers. In this latter case, we lower the frequency more rapidly from $\omega = 1.04$ to $\omega = 0.75$. As expected, when relying on instability dynamics, we find the final configuration depends on the initial period at the unstable frequency, and the speed of the switch. However, we find that the generation of breathers through this process is robust and occurs without any special choice of these parameters. We should note that we have been unsuccessful in using this method to generate mixed-frequency breathers. This appears to be due to the relatively small windows of stability (and presumably also corresponding basins of attraction) of the breathers for our choice of parameters. To achieve generation we have instead used initial conditions near the breather state. The general problem of the nature of the dynamics for different initial conditions is largely still an open one.
![(color online). On-site breather generation from an initially quiescent pendulum configuration ($\theta(0) = \dot{\theta}(0) = 0$), through modulation of the driving frequency. (a) Generation of single on-site breather following period of instability dynamics; (b) time dependence of driving frequency $\omega$.[]{data-label="breathgen"}](fig22.eps){width="0.95\columnwidth"}
![(color online). Off-site breather generation from an initially quiescent pendulum configuration ($\theta(0) = \dot{\theta}(0) = 0$), through modulation of the driving frequency. (a) Generation of single off-site breather following a period of unstable dynamics; (b) time dependence of driving frequency $\omega$.[]{data-label="breathgenoff"}](fig23.eps){width="0.95\columnwidth"}
Conclusions
===========
We have examined in detail the response of both a single pendulum and a pendulum chain to horizontal driving of the pendulum pivot point, in an experimentally accessible region of the system parameter space, focusing on small driving amplitude. We have characterized the single pendulum solutions, finding a region of bistability near the linear resonance of the pendulum natural frequency and driving frequency ($\omega = 1$), with one stable solution of large amplitude and roughly in phase with the driving force and the second stable solution of small amplitude and out of phase. We find also the existence of solutions three times the period of the driving force, near the subharmonic resonance at $\omega = 3$. These are large amplitude solutions, with no connection to the period-1 solutions. Turning to the pendulum chain we show that the period-1 higher amplitude solution exhibits a modulational instability below a certain critical frequency, and beyond a certain chain length. We characterize this modulational instability by studying the instability of a sinusoidal solution ansatz, for a low order expansion of the sinusoidal nonlinearity, and find good agreement with our numerical results. We examine the on-site and off-site localized breather solutions in the pendulum chain, since the breather waveforms appear to spontaneously form as a result of this instability. More specifically, we show their close connection to the single pendulum periodic solutions. In particular, the breathers only exist for parameters for which there is single pendulum bistability. Following this we examine the possibility of multi-frequency breathers, corresponding to the bistable region near the first subharmonic resonance. We find that for moderate chain lengths stable breathers exist in which the central sites are moving at one third of the frequency of the edge sites. For longer chain lengths however we find these breathers are weakly unstable. Finally we turn to the possibility of breather generation through dynamically tuning the driving frequency. We discuss a robust scheme which may be used to generate both on-site and off-site breathers. Natural directions for future work include further characterization of the subharmonic response, possibly in connection with shorter chain settings and the potential connection of this setting with experiments such as the mechanical ones of [@CuevasPRL2009], as well as the electrical ones of [@lars2]. On the other hand, understanding more systematically breather, as well as multi-breather states and a potential tuning of their existence, as well as stability regimes would be of particular interest in inducing (and optimizing) energy localization in this system.
[1]{}
A.H. Nayfeh and D.T. Mook, [*Nonlinear Oscillations*]{} John Wiley & Sons, Inc. (1979).
James A. Blackburn, H. J. T. Smith, and N. Grønbech Jensen. 903 (1992).
W. T. Grandy Jr. and Matthias Schöck, [American Journal of Physics]{} [**65**]{}, 376 (1997).
A. Stephenson. (1908).
R.A. Struble and J.A. Marlin, Quart. J. Mech. Appl. Math. [**18**]{}, 405 (1965).
B.I. Cheshankov, J. Appl. Math. & Mech. [**35**]{}, 301 (1971).
R. Van Dooren. , 77 (1996).
John M. Schmitt and Philip V. Bayly, , 1 (1998). J. Jeong and S.-Y. Kim. , 393 (1999).
J. Cuevas, L. Q. English, P. G. Kevrekidis, and M. Anderson. , 224101 (2009).
J.V. José and E.J. Saletan, [*Classical Dynamics, A Contemporary Approach*]{} Cambridge University Press (1998).
W. Chester, J. Inst. Maths Applics [**15**]{}, 289 (1975).
T. Ikeda, Y. Harata and K. Nishimura, J. Comput. Nonlinear Dynam. [**8**]{}, 041009 (2013).
V.M. Burlakov, Phys. Rev. Lett. [**80**]{}, 3988 (1998).
J.L. Marín and S. Aubry, Nonl. [**9**]{}, 1501 (1996).
L. Q. English, F. Palmero, P. Candiani, J. Cuevas, R. Carretero-González, P. G. Kevrekidis, and A. J. Sievers, Phys. Rev. Lett. [**108**]{}, 084101 (2012).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We prove under some assumptions that the Tate conjecture holds for products of Fermat varieties of different degrees. The method is to use a combinatorial property of eigenvalues of geometric Frobenius acting on $\ell$-adic étale cohomology.'
author:
- 'Rin Sugiyama[^1]'
title: Tate conjecture for products of Fermat varieties over finite fields
---
[Mathematics Subject Classification(2010)]{}: 14G15, 11G25, 14H52
[Keywords]{}: Tate conjecture, Chow group, finite fields
Introduction
============
Let $p$ be a prime number. Let $k$ be a finite field of characteristic $p$. Let $\overline{k}$ be a separable closure of $k$ and let $G_k$ be the Galois group of $\overline{k}/k$. Let $\ell$ be a prime number different from $p$. For a projective smooth and geometrically integral variety $X$ over $k$, ${{\rm CH}}^i(X)$ denotes the Chow group of algebraic cycles on $X$ of codimension $i$ modulo rational equivalence. We first state the following famous conjecture raised by Tate [@Ta1]:
The cycle class map $$\label{map}
\rho_X^i : {{\rm CH}}^i(X)\otimes {\mathbb}{Q}_{\ell}\longrightarrow H^{2i}(\overline{X},{\mathbb}{Q}_{\ell}(i))^{G_k}$$ is surjective.
Here $H^{2i}(\overline{X},{\mathbb}{Q}_{\ell}(i))^{G_k}$ denotes the $G_k$-invariant part of the $\ell$-adic cohomology $H^{2i}(\overline{X},{\mathbb}{Q}_{\ell}(i))$ (cf. Notation) of $\overline{X}:=X\times_k\overline{k}$. Tate [@Ta2] proved that ${\boldsymbol{T}}(X/k,1)$ holds in case where $X$ is an abelian variety or a product of curves. In case where $X$ is a product of curves, Soulé [@So] proved that the Tate conjecture holds for $i=0,1,\dim X-1, \dim X$. Spiess [@Sp] proved that the Tate conjecture holds for products of elliptic curves. For other known cases we refer to [@Ta3].
Beilinson [@Be] furthermore conjectured that $\rho_X^i$ is also injective. Therefore the map $\rho_X^i$ is conjectured to be bijective and we call the conjecture *the Tate-Beilinson conjecture*.
Let $K_i(X)$ be Quillen’s higher $K$-groups associated to the category of vector bundles on $X$. Let $K_i(X)_{{\mathbb}{Q}}=K_i(X)\otimes {\mathbb}{Q}$. The Tate conjecture can be reformulated in terms of $K$-groups, because there is an isomorphism $$K_0(X)^{(i)}_{{\mathbb}{Q}}\simeq {{\rm CH}}^i(X)\otimes {\mathbb}{Q}.$$ Here $K_0(X)^{(i)}_{{\mathbb}{Q}}$ denotes the subspace of the rational $K$-group $K_0(X)_{{\mathbb}{Q}}$ on which the Adams operator $\psi^m$ acts as the multiplication by $m^i$ for all $m\in {\mathbb}{N}$. For higher $K$-groups of projective smooth varieties over finite fields, there is a conjecture raised by Parshin:
For every projective smooth variety $X$ over a finite field and every integer $i>0$, $K_i(X)_{{\mathbb}{Q}}=0$.
By the Bass conjecture [@Ba] on finite generation of $K$-groups, these groups are expected to be finite. Geisser [@Ge] proved that if the Tate conjecture holds and numerical and rational equivalence over finite fields agree, then Parshin’s conjecture holds for all projective smooth varieties over finite fields. Kahn [@Ka] considered the class ${\mathcal}{B}_{tate}(k)$ of projective smooth varieties of Abelian type over $k$, for which the Tate conjecture holds. A projective smooth variety $X$ is called of Abelian type if the Chow motive of $X$ belongs to the subcategory of the category of Chow motives generated by Artin motives and Chow motives of abelian varieties. For example, products of curves and Fermat varieties are of Abelian type. For any variety $X$ in ${\mathcal}{B}_{tate}(k)$, Kahn proved that rational and numerical equivalence agree, namely the Tate-Beilinson conjecture holds, and that Parshin’s conjecture holds for $X$ using Geisser’s result [@Ge] and Kimura’s result [@Ki].
Parshin’s conjecture implies that for all $n\geq0$ and $i>0$, ${{\rm CH}}^n(X,i)_{{\mathbb}{Q}}=0$ by the isomorphism $\displaystyle K_i(X)_{{\mathbb}{Q}} \simeq \bigoplus _{n\ge0}{{\rm CH}}^n(X,i)_{{\mathbb}{Q}}$ ([@Bl]). Here ${{\rm CH}}^n(X,i)$ denotes Bloch’s higher Chow group ([@Bl]) and ${{\rm CH}}^n(X,i)_{{\mathbb}{Q}}:={{\rm CH}}^n(X,i)\otimes {\mathbb}{Q}$. By Deligne’s proof of the Weil conjecture [@De], we have $H^{2i-j}(\overline{X},{\mathbb}{Q}_{\ell}(i))^{G_k}=0$ for $j\neq0$. Therefore we can combine the Tate-Beilinson conjecture and the Parshin conjecture into a conjecture below. The results of Geisser and Kahn mentioned above imply that the following conjecture holds for any variety $X$ in ${\mathcal}{B}_{tate}(k)$:
\[cj\] The cycle map $$\begin{aligned}
{{\rm CH}}^i(X,j)\otimes {\mathbb}{Q}_{\ell}\longrightarrow H^{2i-j}(\overline{X},{\mathbb}{Q}_{\ell}(i))^{G_k}\end{aligned}$$ is bijective for all integers $i,j\geq0$.
For a non-negative integer $r$ and a positive integer $m$ prime to $p$, we define a variety $X_m^r=X_m^r(a_0,\dots,a_{r+1}) \subset {\mathbb}{P}^{r+1}$ of dimension $r$ and of degree $m$ by the equation $$a_0x_0^m+a_1x_1^m+\cdots +a_{r+1}x_{r+1}^m=0.$$ Here $a_i$ is a non-zero elements in $k$. In case $a_0=\cdots=a_{r+1}\neq0$, $X_m^r$ is called a Fermat variety and denoted by $V_m^r$. Shioda-Katsura [@SK; @Sh] proved that the Tate conjecture holds for a product $V_m^{r_1}\times\cdots\times V_m^{r_s}$ of Fermat varieties of same degree under an arithmetic condition(for example, $m$ is a prime and $p \equiv 1 \mod m$).
In this paper, we are concerned with products of $X_m^r$ of different degrees, and prove the following theorem:
\[t\] Let $k$ be a finite field of characteristic $p$. Let $m_1, \ldots, m_d$ be positive integers prime to $p$. Let $r_1,\dots,r_d$ be positive integers. Let $X_{m_j}^{r_j}=X_{m_j}^{r_j}({\bf a}_j)$ with coefficients ${\bf a}_j \in (k^{\times})^{r_j+2}$. Let $X$ be the product $X_{m_1}^{r_1}\times \cdots \times X_{m_d}^{r_d}$. Then ${\boldsymbol{T}}(X/k,i)$ is true for all $i$ in the following cases:
- In case that $r_j=1$ for all $j$, assume that one of the following two conditions holds:
- For each $1 \leq j\leq d$, there is at most one $j^{\prime} \neq j$ such that ${\rm gcd}(m_j,m_{j^{\prime}})> 2$.
- For every even integer $j$ with $4\leq j\leq d$ and for every $1\leq n_1<n_2<\dots<n_j\leq d$, there exists an integer $a$ with $1\leq a\leq j$ such that
- ${\rm gcd}(m_{n_a},m_{n_r})\leq 2$ for any $r \not =a$,
- the order of $p$ in the group $({\mathbb}{Z}/m_{n_a})^{\times}$ is odd.
In this case, moreover, we obtain that ${{\rm CH}}^i(X)$ is the direct sum of a free abelian group of finite rank and a group of finite exponent for all $i$.
- In case that $r_j$ is odd for all $j$, assume that the following conditions hold[:]{}
${\rm gcd}(m_j,m_{j^{\prime}})\leq 2$ for $j \neq j^{\prime}$,
the order of $p$ in the group $({\mathbb}{Z}/m_j)^{\times}$ is odd for all $j$.
- In case that $r_j$ is even for all $j$, assume that the following conditions hold[:]{}
${\rm gcd}(m_j,m_{j^{\prime}})\leq 2$ for $j \neq j^{\prime}$,
${\boldsymbol{T}}(X_{m_j}^{r_j}/L,r_j/2)$ is true for all $j$ and for a sufficiently large finite extension $L$ of $k$.
- In general case, assume that the following conditions hold[:]{}
${\rm gcd}(m_j,m_{j^{\prime}})\leq 2$ for $j \neq j^{\prime}$,
for $j$ with odd $r_j$, the order of $p$ in the group $({\mathbb}{Z}/m_j)^{\times}$ is odd,
for $j$ with even $r_j$, ${\boldsymbol{T}}(X_{m_j}^{r_j}/L,r_j/2)$ is true for a sufficiently large finite extension $L$ of $k$.
For a variety $X$ in (1)–(4) of Theorem \[t\], the vector space $H^{2i}(\overline{X},{\mathbb}{Q}_{\ell}(i))^{G_k}$ does not vanish. Moreover, if $X$ is a variety as in (2)–(4), then the dimension of $H^{2i}(\overline{X},{\mathbb}{Q}_{\ell}(i))^{G_k}$ over ${\mathbb}{Q}_{\ell}$ can be described. For example, let $X=X_{m_1}^{r_1}\times\cdots\times X_{m_d}^{r_d}$ be as in (2). Then we have $$\begin{aligned}
\dim_{{\mathbb}{Q}_{\ell}}&H^{2i}(\overline{X},{\mathbb}{Q}_{\ell}(i))^{G_k}\\
&=\sharp\{(i_1,\dots,i_d)\;|\; i_1+\cdots+i_d=i \;\text{and}\; 0\le i_j\le {\rm min}\{r_j,i\}\; \text{for all $j$}\}\end{aligned}$$ where $\sharp $ denotes the cardinality of a finite set. The above equation follows from the proof of Theorem \[t\] (see §3.1.).
Our proof of Theorem \[t\] is based on a combinatorial property of eigenvalues of Frobenius acting on $\ell$-adic étale cohomology of $X_m^r$. We will use an argument similar to that Spiess used in [@Sp]. For (1)(b), we use a result of Soulé [@So Théorème 3 i)].
From Theorem \[t\](3) and the fact that the Tate conjecture holds for $X_m^2/k$, we obtain the following:
\[cf\] Let the notation be as in Theorem \[t\]. Assume that $(m_j,m_{j^{'}})\leq 2$ for $j \neq j^{'}$. Then the Tate conjecture holds for $X_{m_1}^2\times X_{m_2}^2\times \cdots \times X_{m_d}^2/k$.
In case that all coefficients are same, that is, in case that $X_{m_j}^{r_j}=V_{m_j}^{r_j}$ for all $j$, this corollary also follows from Theorem \[t\](1) and “inductive structure" for Fermat varieties given by Shioda-Katsura [@SK] (cf. §3.2).
From Theorem \[t\], we see that the varieties as in Theorem \[t\] belong to $B_{tate}(k)$, and hence we obtain the following (cf. Corollary \[ca\]):
\[ct\] Let $X$ be a variety as in $(1)$$(4)$ of Theorem \[t\]. Then the following holds[:]{}
$(1)$ Conjecture \[cj\] holds for $X$,
$(2)$ The Lichtenbaum conjecture holds for $X$.
Furthermore if $X$ is a variety as in Theorem \[t\]$\,(1)\,(b)$, then the following holds[:]{}
${{\rm CH}}^2(X)$ is finitely generated.
$K_0(X)$ is the direct sum of a free abelian group of finite rank and a group of finite exponent.
Here the Lichtenbaum conjecture is the conjecture stated in [@Li §7].
Corollary \[ct\](1)(2) immediately follow from Theorem \[t\] and results of Geisser [@Ge] and Kahn [@Ka] mentioned above. Corollary \[ct\](4)(5) follow from Theorem \[t\](1)(b), a spectral sequence from higher Chow groups to algebraic $K$-theory [@FS] $$E_2^{p,q} ={{\rm CH}}^{-q}(X,-p-q) \Longrightarrow K_{-p-q}(X)$$ and the following theorem:
Let $X$ be a projective smooth and geometrically connected variety over a finite field, the subgroup ${{\rm CH}}^2(X)_{tor}$ of torsion elements of ${{\rm CH}}^2(X)$ is finite.
This paper is organized as follows: in the next section (§2), we prove some lemmas about eigenvalues of Frobenius acting on $\ell$-adic étale cohomology. In §3, we prove Theorem \[t\] using lemmas in §2. In §3.1, we calculate the dimension of $\ell$-adic étale cohomology for varieties of Theorem \[t\]. In §3.2, we give an alternative proof of Corollary \[cf\] using “inductive structure" for Fermat varieties. In the last section (§4), we give an application to a zeta value of products of four curves using a result of Kohmoto [@Ko].
Notation {#notation .unnumbered}
--------
For a field $k$, $k^{\times }$ denotes the multiplicative group. For an integer $m>1$, ${\mathbb}{Z}/m$ denotes the cokernel of the map ${\mathbb}{Z} \stackrel{\times m}{\rightarrow}{\mathbb}{Z}$ and $({\mathbb}{Z}/m)^{\times }$ denotes the group of invertible elements in ${\mathbb}{Z}/m$.
Unless indicated otherwise, all cohomology groups of schemes are taken for the $\acute {\text{e}}$tale topology. Let $X$ be a scheme. For a prime number $\ell$ which is invertible on $X$ and integers $i,j\geq 0$, we define $$\begin{aligned}
H^i(X,{\mathbb}{Z}_{\ell}(j))&:=H^i_{\rm cont}(X,{\mathbb}{Z}_{\ell}(j)),\\
H^i(X,{\mathbb}{Q}_{\ell}(j))&:=H^i(X,{\mathbb}{Z}_{\ell}(j))\otimes _{{\mathbb}{Z}_{\ell}}{\mathbb}{Q}_{\ell}.\end{aligned}$$ where $H^i_{\rm cont}(X,{\mathbb}{Z}_{\ell}(j))$ is the continuous $\ell$-adic étale cohomology defined by Jannsen [@Ja0]. We write $H^i(X,{\mathbb}{Q}_{\ell})$ for $H^i(X,{\mathbb}{Q}_{\ell}(0))$. For a scheme $X$ and a positive integer $m$ which is invertible on $X$, we write $\mu _m$ for the $\acute {\text{e}}$tale sheaf of $m$-th roots of unity on $X$. For all schemes $X$ considering in this paper, we have $$\begin{aligned}
H^i(X,{\mathbb}{Z}_{\ell}(j))&=\varprojlim _n H^i(X,\mu_{\ell^n}^{\otimes j}).\end{aligned}$$
Lemmas
======
In this section, we prove some lemmas about eigenvalues of Frobenius acting on $\ell$- adic étale cohomology, which is used in the proof of Theorem \[t\].
Let $k$ be a finite field ${\mathbb}{F}_q$. We consider a variety $X_m^r$ over $k$. Let $F \in G_k$ be the geometric Frobenius. In this paper, we call an eigenvalue $\alpha$ of $F$ acting on $H^{i}(\overline{X_m^r},{\mathbb}{Q}_{\ell})$ *Weil number* of weight $i$ for $X_m^r/k$. By Deligne’s theorem (Weil conjecture) [@De], $\alpha$ is an algebraic integer and the absolute value $|\alpha|$ of $\alpha$ is equal to $q^{i/2}$. We have known that for any $i\neq r$, $0\leq i \leq 2r$, $$\begin{aligned}
\label{c}
H^{i}(\overline{X_m^r},{\mathbb}{Q}_{\ell})=\left\{ \begin{array}{ll}
0 & \text{for $i$ odd} \\
\text{1-dimensional} & \text{for $i$ even}\\
\end{array} \right.\end{aligned}$$ For this fact, we refer to [@SGA5 Expose VII, Corollaire 7.5.]. In case $i$ is even, $H^{i}(\overline{X_m^r},{\mathbb}{Q}_{\ell})$ is generated by algebraic cycles, that is, the Tate conjecture is true. From we see that a Weil number for $X_m^r/k$ of even weight $i \neq r$ is $q^{i/2}$, and that we have no Weil number for $X_m^r/k$ of odd weight $i \neq r$.
Now assume that $q=p^f$ is the least power of $p$ such that $q \equiv 1 \mod m$. We then recall Weil’s result on describing Weil numbers of weight $r$ for $X_m^r/k$ in terms of Jacobi sums [@We]. A Weil number $\alpha$ of weight $r$ for $X_m^r/k$ is given by $$\alpha=(-1)^r\bar{\chi}(a_0)^{\gamma_0}\cdots\bar{\chi}(a_{r+1})^{\gamma_{r+1}}j(\gamma).$$ Here $\chi$ is a fixed character of order $m$ of $k^{\times}$ and $j(\gamma)$ is the Jacobi sum:
$$j(\gamma)=\sum_{\stackrel{1+v_1+\cdots+v_{r+1}=0}{ \ v_i \in k^{\times}}}\chi(v_1)^{\gamma_1}\cdots\chi(v_{r+1})^{\gamma_{r+1}},$$ where $\gamma$ is an element of the set $${\mathfrak}{D}_{m,r}=\bigl\{ (\gamma_0,\dots,\gamma_{r+1}) \ | \ \gamma_i \in {\mathbb}{Z}/m, \ \gamma_i \not \equiv 0, \ \gamma_0+\gamma_1+\cdots+\gamma_{r+1}\equiv 0\bigr\}.$$ Conversely, for any $\gamma \in {\mathfrak}{D}_{m,r}$, the number $(-1)^r\bar{\chi}(a_0)^{\gamma_0}\cdots\bar{\chi}(a_{r+1})^{\gamma_{r+1}}j(\gamma)$ is a Weil number of weight of $r$ for $X_m^r/k$.
\[l1\] Let $k$ be a finite field ${\mathbb}{F}_q$. Let $\alpha$ be a Weil number of weight $r$ for $X_m^r/k$. Then there is an integer $e>0$ such that $\alpha^e$ belongs to ${\mathbb}{Q}(\zeta_m)$, where $\zeta_m$ is a primitive $m$-th root of unity.
Let $L:={\mathbb}{F}_{q^e}$ be a finite extension of $k$ such that $q^e\equiv 1 \mod m$. Then $\alpha ^e$ is a Weil number of weight $r$ for $X_m^r/L$. By the above Weil’s result, $\alpha^e$ is a power of $(-1)^r\bar{\chi}(a_0)^{\gamma_0}\cdots\bar{\chi}(a_{r+1})^{\gamma_{r+1}}j(\gamma)$, and therefore belongs to ${\mathbb}{Q}(\zeta_m)$.
\[lsk\] Let $f$ be the residue degree of $p$ in ${\mathbb}{Q}(\zeta_m)$. Let $\alpha$ be a Weil number of weight $r$ for $X_m^r/{\mathbb}{F}_{p^f}$. Then the following statements are equivalent[:]{}
$(i)$ some power of $\alpha$ is a power of $p$[;]{}
$(ii)$ the valuations $v_{{\mathfrak}{p}}(\alpha)$ of $\alpha$ are independent of a prime ${\mathfrak}{p}$ in ${\mathbb}{Q}(\zeta_m)$ dividing $p$, and equal to $\frac{fr}{2}$.
\[l2\] Let $m_1,m_2,\dots,m_{2i}$ be positive integers prime to $p$. Let $\alpha_j$ be a Weil number of weight $1$ for $X_{m_j}^1/k$. Assume that for each $1 \leq j\leq 2i$, there is at most one $j^{\prime} \neq j$ such that ${\rm gcd}(m_j,m_{j^{\prime}})> 2$. If $\alpha_1\alpha_2\cdots\alpha_{2i}=q^i$, then$-$after renumbering the $\alpha_j$ if necessary$-$there exists an integer $N$ such that $$\begin{aligned}
(\alpha_1\alpha_2)^N=\cdots=(\alpha_{2i-1}\alpha_{2i})^N=q^N.\end{aligned}$$
From the assumption, after renumbering the $m_j$ if necessary, we have pairs $(m_1,m_2),\dots,(m_{2i-1},m_{2i})$ such that $${\rm gcd}({\rm lcm}(m_{2j-1},m_{2j}),{\rm lcm}(m_{2j^{\prime}-1},m_{2j^{\prime}}))\leq2 \ \text{ for all } \ j \neq j^{\prime}.$$ Put $L_j:={\rm lcm}(m_{2j-1},m_{2j})$ for $1\leq j\leq i$. From the equation $\alpha_1\alpha_2\cdots\alpha_{2i}=q^i$ and Lemma \[l1\], there is a positive integer $M$ such that $$\begin{aligned}
(\alpha_1\alpha_2)^M=q^{Mi}(\alpha_3\cdots\alpha_{2i})^{-M} \in {\mathbb}{Q}(\zeta_{L_1})\cap {\mathbb}{Q}(\{\zeta_{L_j},j\neq 1\}).\end{aligned}$$ Now we have $${\mathbb}{Q}(\zeta_{L_1})\cap {\mathbb}{Q}(\{\zeta_{L_j},j\neq 1\}) \subset {\mathbb}{Q}(\zeta_{L_1})\cap {\mathbb}{Q}(\zeta_L)={\mathbb}{Q}.$$ Here $L={\rm lcm}(L_j,j\neq 1)$. Hence $(\alpha_1\alpha_2)^M$ belongs to ${\mathbb}{Q}$ and therefore $(\alpha_1\alpha_2)^{2M}=|(\alpha_1\alpha_2)^M|^2=q^{2M}$, because $\alpha$ is a Weil number of weight $1$. The same argument shows that $(\alpha_{2i-1}\alpha_{2i})^{2M}=q^{2M}$ for $2\leq i\leq d$.
\[l3\] Let $m_1, \dots, m_d$ be positive integers prime to $p$. Let $r_1, \dots, r_d$ be positive odd integers. For $j=1,\dots,d$, let $\alpha_j$ be a Weil number of weight $i_j$ for $X_{m_j}^{r_j}/k$. Assume that there is an integer $j_0$ such that
$(1)$ $(m_{j_0},m_j)\leq 2$ for $j \neq j_0$,
$(2)$ $i_{j_0}=r_{i_{j_0}}$,
$(3)$ the order of $p$ in the group $({\mathbb}{Z}/m_{j_0})^{\times}$ is odd.\
Then the product $\alpha_1\alpha_2\cdots\alpha_d$ does not belong to ${\mathbb}{Q}$.
We may assume that $j_0 =1$. From Lemma \[l1\], there is a positive integer $N$ such that $\alpha_j^N \in {\mathbb}{Q}(\zeta_{m_j})$ for all $j$. If $\alpha_1\alpha_2\cdots\alpha_d=:b \in {\mathbb}{Q}$, then we have $$\begin{aligned}
\alpha_1^N = b^N(\alpha_2\cdots\alpha_d)^{-N} \in {\mathbb}{Q}(\zeta_{m_1})\cap{\mathbb}{Q}(\zeta_{m_j}, j\neq 1).\end{aligned}$$ From the assumption (1), ${\mathbb}{Q}(\zeta_{m_1})\cap{\mathbb}{Q}(\zeta_{m_j}, j\neq 1)={\mathbb}{Q}$. By the assumption (2), we have $|\alpha_1^N|=q^{Nr_1/2}$. Therefore we see that some power of $\alpha_1$ is a power of $p$.
On the other hand, by linear algebra, we see that some power of $\alpha_1$ is a power of a Weil number $\alpha$ of weight $r_1$ for $X_{m_1}^{r_1}/{\mathbb}{F}_{p^f}$. Here $f$ is the order of $p$ in $({\mathbb}{Z}/m_1)^{\times}$ which is odd by assumption (3). Since $\frac{fr_1}{2}$ is not an integer, the valuation $v_{{\mathfrak}{p}}(\alpha)$ is not equal to $\frac{fr_1}{2}$ for all primes ${\mathfrak}{p}$ in ${\mathbb}{Q}(\zeta_{m_1})$ dividing $p$. By Lemma \[lsk\], any power of $\alpha$ is not a power of $p$. Therefore any power of $\alpha_1$ is not a power of $p$, which is a contradiction.
\[l4\] Let $m_1, \dots, m_d$ be positive integers prime to $p$. Let $r_1, \dots, r_d$ be positive integers. Assume that ${\rm gcd}(m_j,m_{j^{\prime}})\leq 2$ for $j \neq j^{\prime}$. For $j=1,\dots,d$, let $\alpha_j$ be a Weil number of weight $2r_j$ for $X_{m_j}^{2r_j}/k$. Put $r:=r_1+\cdots+r_d$. If $\alpha_1\alpha_2\cdots\alpha_d=q^r$, then there exists an integer $N$ such that $$\begin{aligned}
\alpha_1^N=q^{Nr_1}, \cdots, \alpha_d^N=q^{Nr_d}.\end{aligned}$$
From the equation $\alpha_1\alpha_2\cdots\alpha_d=q^r$ and Lemma \[l1\], there is a positive integer $M$ such that $$\alpha_1^M=q^{Mr}(\alpha_2\cdots\alpha_d)^{-M} \in {\mathbb}{Q}(\zeta_{m_1})\cap {\mathbb}{Q}(\{\zeta_{m_j},j\neq 1\}).$$ By the assumption, we have $${\mathbb}{Q}(\zeta_{m_1})\cap {\mathbb}{Q}(\{\zeta_{m_j},j\neq 1\})={\mathbb}{Q}.$$ Hence $\alpha_1^M \in {\mathbb}{Q}$ and therefore $\alpha_1^{2M}=|\alpha_1^M|^2=q^{2Mr_1}$, because $\alpha_1$ is a Weil number of weight $r_1$. The same argument shows that $\alpha_j^{2M}=q^{2Mr_j}$ for $2\le j\le d$.
Proof of the theorem
====================
Before beginning to prove the theorem, we define a class ${\mathcal}{A}(k)$ constructed from projective smooth varieties as in Theorem \[t\] and prove that Conjecture \[cj\] in the introduction holds for varieties which belong to ${\mathcal}{A}(k)$. The argument is based on that of Soulé [@So].
Let $Y$ be a projective smooth variety over a finite field $k$ which is not necessarily geometrically integral over $k$. Put $\kappa :=\Gamma (Y,{\mathcal}{O}_Y)$. Then the scalar extension $Y\otimes _k\kappa $ is a disjoint union of copies of $Y$, and we have $$\begin{aligned}
{{\rm CH}}^i(Y)\otimes {\mathbb}{Q}_{\ell}&\simeq \bigl({{\rm CH}}^i(Y\otimes _k\kappa )\otimes {\mathbb}{Q}_{\ell}\bigr)^{{\rm Gal}(\kappa /k)}\\
&\simeq \bigl({{\rm CH}}^i(Y)\otimes {\mathbb}{Q}_{\ell}[{\rm Gal}(\kappa/k)]\bigr)^{{\rm Gal}(\kappa /k)}.\end{aligned}$$ Hence the bijectivity of the cycle class map for $Y/k$ is equivalent to that for $Y/\kappa $. For this reason we also consider the case that $Y$ is not geometrically integral.
\[ak\] For any finite field $k$, we define ${\mathcal}{A}(k)$ to be the smallest class of projective smooth varieties over $k$ satisfying the following properties:
\(1) Varieties which satisfy the assumption of Theorem \[t\] belong to ${\mathcal}{A}(k)$.
\(2) If $X$ and $Y$ belong to ${\mathcal}{A}(k)$, then the disjoint union $X\amalg Y$ of $X$ and $Y$ belongs to ${\mathcal}{A}(k)$.
\(3) If $X$ belongs to ${\mathcal}{A}(k)$, $Y$ is a projective smooth variety such that $\dim X = \dim Y$, $Y$ is a direct summand of $X$ as Chow motives with ${\mathbb}{Z}[\frac{1}{n}]$ coefficients for some $n\geq1$, then $Y$ belongs to ${\mathcal}{A}(k)$.
\(4) Let $X$ be a projective smooth variety over $k$ and $k^{\prime }$ be a finite extension of $k$. If $X\otimes k^{\prime }$ belongs to ${\mathcal}{A}(k^{\prime})$, then $X$ belongs to ${\mathcal}{A}(k)$.
\(5) If $X$ belongs to ${\mathcal}{A}(k)$ and $E$ is a vector bundle on $X$, then the projective bundle $P(E)$ on $X$ associated to $E$ belongs to ${\mathcal}{A}(k)$.
\(6) Let $X$ be a projective smooth variety over $k$ and $Y$ be a closed smooth subvariety of $X$. Let $W$ be the blowing up of $X$ along $Y$. Then $W$ belongs to ${\mathcal}{A}(k)$ if and only if $X$ and $Y$ belong to ${\mathcal}{A}(k)$.
\[ca\] Let $X$ be a variety which belong to ${\mathcal}{A}(k)$. Then we have the following[:]{}
$(1)$ Conjecture \[cj\] holds for $X$,
$(2)$ The Lichtenbaum conjecture holds for $X$.
By results of Kahn [@Ka] and Geisser [@Ge], it suffices to show that the Tate-Beilinson conjecture holds for all $i$ and for $X \in {\mathcal}{A}(k)$.
Let ${\mathcal}{A}^{\prime }(k)$ be a class of all projective smooth varieties over $k$ for which the Tate-Beilinson conjecture holds. Then by the smallness of ${\mathcal}{A}(k)$, it suffices to show that ${\mathcal}{A}^{\prime }(k)$ satisfies the above properties (1)–(6). For (1), the assertion follows from Theorem \[t\]. We only show that ${\mathcal}{A}^{\prime }(k)$ satisfies property (6). The other properties (2)–(5) can be checked by similar arguments. Let the notation be as in property (6). Since we have the following decomposition of the Chow motive $\widetilde{W}$ of $W$ (cf. [@So]) $$\begin{aligned}
\widetilde{W}\simeq \widetilde{X}\oplus \bigl(\bigoplus_{j=1}^{c-1}\widetilde{Y}\otimes {\mathbb}{L}^j\bigr),\end{aligned}$$ we have the following isomorphisms (cf. [@So Th$\acute{\text{e}}$or$\grave {\text{e}}$me 4.]): $$\begin{aligned}
{{\rm CH}}^i(W)&\simeq {{\rm CH}}^i(X)\oplus \big( \bigoplus_{j=1}^{c-1}{{\rm CH}}^{i-j}(Y)\big),\\
H^{2i}(\overline{W},{\mathbb}{Q}_{\ell}(i))&\simeq H^{2i}(\overline{X},{\mathbb}{Q}_{\ell}(i))\oplus \big( \bigoplus_{j=1}^{c-1}H^{2(i-j)}(\overline{Y},{\mathbb}{Q}_{\ell}(i-j))\big).\end{aligned}$$ Hence $W$ belongs to ${\mathcal}{A}^{\prime }(k)$ if and only if $X$ and $Y$ belong to ${\mathcal}{A}^{\prime }(k)$.
We prove Theorem \[t\] using Lemmas about Weil numbers for $X_m^r/k$ in §2. We can formulate the assertion of the Tate conjecture for Chow motives and simplify the argument. For the definition and basic properties of Chow motives, we refer to [@So]. Spiess [@Sp] used Chow motives in the proof of the Tate conjecture for products of elliptic curves. We use similar arguments of Spiess. We first give the following Lemma:
\[l5\] Let $L/k$ be a finite extension and $X$ be a projective smooth variety over $k$. Then ${\boldsymbol{T}}(X\times_kL/L,i)$ implies ${\boldsymbol{T}}(X/k,i)$.
For the proof of this lemma, see [@Sp Lemma 1].
Using the motivic decomposition of the Chow motive $X_m^1=1\oplus X_m^{+} \oplus {\mathbb}{L}$ (see [@So]), our task is reduced to show that ${\boldsymbol{T}}(X_{m_1}^{+}\otimes X_{m_2}^{+}\otimes \cdots\otimes X_{m_s}^{+}/k,i)$ holds for all $s$ and $i$. Since Soulé proved that $H^{2i}(\overline{X_{m_1}^{+}\otimes X_{m_2}^{+}\otimes \cdots\otimes X_{m_s}^{+}},{\mathbb}{Q}_{\ell}(i))=0$ for $s\neq 2i$, we may assume $s=2i$. We have $$\begin{aligned}
H^{2i}(\overline{X_{m_1}^{+}\otimes X_{m_2}^{+}\otimes \cdots\otimes X_{m_s}^{+}},{\mathbb}{Q}_{\ell}(i))^{G_k}\simeq \bigl(H_{1}\otimes H_{2}\otimes \cdots \otimes H_{2i}\bigr)^{F=q^i}.\end{aligned}$$ Here we set $H_{s}=H^1(\overline{X_{m_s}^1},{\mathbb}{Q}_{\ell})$ and $F \in G_k$ is the geometric Frobenius.
A basis of the vector space $\bigl(H_{1}\otimes H_{2}\otimes \cdots \otimes H_{2i}\bigr)^{F=q^i}$ corresponds to $2i$-tuples $(\alpha_1,\alpha_2,\dots,\alpha_{2i})$ such that $\alpha_1\alpha_2\cdots\alpha_{2i}=q^i$ (where $\alpha _s$ is a Weil number of weight $1$ for $X_{m_s}^1/k$).
From the assumption (a) in Theorem \[t\] and Lemma \[l2\], there is a positive integer $N$ such that for every such tuples $(\alpha_1,\alpha_2,\dots,\alpha_{2i})$, after renumbering the $\alpha_j$ if necessary, we have $$\begin{aligned}
(\alpha_1\alpha_2)^N=\cdots=(\alpha_{2i-1}\alpha_{2i})^N=q^N.\end{aligned}$$ From this we obtain that the map $$\begin{aligned}
\bigoplus_{\sigma}\bigotimes _{j=1}^iH^{2}(\overline{X_{m_{\sigma(2j-1)}}^{+}\otimes X_{m_{\sigma(2j)}}^{+}}&,{\mathbb}{Q}_{\ell}(1))^{G_L} \\
&\longrightarrow H^{2i}(\overline{X_{m_1}^{+}\otimes X_{m_2}^{+}\otimes \cdots\otimes X_{m_{2i}}^{+}},{\mathbb}{Q}_{\ell}(i))^{G_L}\end{aligned}$$ is surjective, where $\sigma$ runs through the set of permutations of $\{1,2,\dots,2i\}$ and $L$ is a finite extension of $k$ with $[L:k]=N$. Now the assertion follows from Lemma \[l5\], ${\boldsymbol{T}}(X_m^1\times X_{m^{\prime}}^1\times_kL/L,1)$(it is true) and the commutative diagram $$\begin{aligned}
{\scriptsize
\xymatrix{\displaystyle
\bigoplus_{\sigma}\bigotimes _{j=1}^i{{\rm CH}}^1(X_{m_{\sigma(2j-1)}}^{+}\otimes X_{m_{\sigma(2j)}}^{+}\otimes L)\otimes {\mathbb}{Q}_{\ell} \ar[r] \ar[d] &\displaystyle\bigoplus_{\sigma}\bigotimes _{j=1}^iH^{2}(\overline{X_{m_{\sigma(2j-1)}}^{+}\otimes X_{m_{\sigma(2j)}}^{+}},{\mathbb}{Q}_{\ell}(1))^{G_L} \ar@{>>}[d]\\
{{\rm CH}}^i(X_{m_1}^{+}\otimes X_{m_2}^{+}\otimes \cdots\otimes X_{m_{2i}}^{+}\otimes L)\otimes{\mathbb}{Q}_{\ell}\ar[r]&H^{2i}(\overline{X_{m_1}^{+}\otimes X_{m_2}^{+}\otimes \cdots\otimes X_{m_{2i}}^{+}},{\mathbb}{Q}_{\ell}(i))^{G_L}.
}}\end{aligned}$$
We first recall a result of Soulé ([@So Theorem 3]).
Let $C_1,\dots,C_d$ be projective, smooth and geometrically irreducible curves over $k={\mathbb}{F}_{q}$, and $X=C_1\times \cdots\times C_d$ be the product. We consider the following condition on $X/k$ and integer $0\leq i\leq d$:\
[**Condition $\boldsymbol{(*)_i}$**]{} Let $j$ be an even integer with $4\leq j\leq \inf(2i,2d-2i)$. For any $1\leq n_1<n_2<\dots<n_j\leq d$, if $\alpha _{n_1},\dots,\alpha _{n_j}$ are eigenvalues of $F$ acting on $H^1(\overline{C}_{n_1},{\mathbb}{Q}_{\ell}),\dots,H^1(\overline{C}_{n_j},{\mathbb}{Q}_{\ell})$, then the product $\alpha_{n_1} \alpha_{n_2}\cdots \alpha_{n_j} $ is not equal to $q^{j/2}$.
\[ts\] Let $i$ be a positive integer with $0\leq i\leq d$. If $X$ satisfies the condition $(*)_i$ over $k$, then ${{\rm CH}}^i(X)$ is the direct sum of a free abelian group of finite rank and a group of finite exponent. Moreover the cycle map $\rho_X^i$ is bijective.
The condition $(*)_i$ holds for $i=0,1,d-1,d$. Therefore if $d\leq 3$, the Tate-Beilinson conjecture holds by Theorem \[ts\].
From Theorem \[ts\], our task is reduced to show that $X$ satisfies the condition $(*)_i$ for all $i$. In our case, from the assumption (b) in Theorem \[t\] and Lemma \[l3\], we see that the product $\alpha_{n_1} \alpha_{n_2}\cdots \alpha_{n_j}$ does not belong to ${\mathbb}{Q}$. Hence $X$ satisfies the condition $(*)$ for all $i$.
From Künneth formula, we have $$\begin{aligned}
H^{2i}(&\overline{X_{m_1}^{r_1}\times X_{m_2}^{r_2}\times \cdots \times X_{m_d}^{r_d}},{\mathbb}{Q}_{\ell}(i))^{G_k}\simeq \bigoplus_{i_1+\cdots+i_d=2i} W(i_1,\dots,i_d),
\end{aligned}$$ where $$W(i_1,\dots,i_d):=\bigl(H^{i_1}(\overline{X_{m_1}^{r_1}},{\mathbb}{Q}_{\ell})\otimes \cdots \otimes H^{i_d}(\overline{X_{m_d}^{r_d}},{\mathbb}{Q}_{\ell})\bigr)^{F=q^i}.$$ From , Lemma \[l3\] and the assumption of the theorem, we obtain that if $i_j$ is odd for some $j$, then $W(i_1,\dots,i_d)=0$.
In the remaining case where $i_j$ is even for all $j$, from we have $$\begin{aligned}
\bigotimes _{j=1}^dH^{i_j}(\overline{X_{m_j}^{r_j}},{\mathbb}{Q}_{\ell}(i_j/2))^{G_k} \simeq W(i_1,\dots,i_d).\end{aligned}$$ We know that if $r$ is odd, then the Tate conjecture ${\boldsymbol{T}}(X_m^{r}/k, i)$ is true for all $i$. Therefore the assertion follows from a similar argument in the proof of Theorem \[t\](1)(a).
Let $i$ be an integer with $0\leq i\leq r$, where $r=r_1+\cdots+r_d$. From Künneth formula and isomorphisms , we have $$\begin{aligned}
H^{2i}(&\overline{X_{m_1}^{r_1}\times X_{m_2}^{r_2}\times \cdots \times X_{m_d}^{r_d}},{\mathbb}{Q}_{\ell}(i))^{G_k}\simeq \bigoplus_{i_1+\cdots+i_{d}=i} V(i_1,\dots,i_{d})^{G_k},
\end{aligned}$$ where $$V(i_1,\dots,i_{d}):=\bigl(H^{2i_1}(\overline{X_{m_1}^{r_1}},{\mathbb}{Q}_{\ell}(i_1))\otimes \cdots \otimes H^{2i_{d}}(\overline{X_{m_{d}}^{r_{d}}},{\mathbb}{Q}_{\ell}(i_{d}))\bigr).$$ Similarly to the above proof of (1)(a), a basis of $V(i_1,\dots,i_{d})^{G_k}$ corresponds to $d$-tuples $(\alpha_1,\alpha_2,\dots,\alpha_d)$ such that $\alpha_1\alpha_2\cdots\alpha_d=q^i$ (where $\alpha _j$ is a Weil number of weight $2i_j$ for $X_{m_j}^{r_j}/k$).
From assumption (i) of the theorem, Lemma \[l4\] and isomorphism , there is a positive integer $N$ such that for every such $d$-tuples $(\alpha_1,\alpha_2,\dots,\alpha_d)$, we have $$\begin{aligned}
\alpha_1^N=q^{Ni_1}, \cdots, \alpha_d^N=q^{Ni_d}.\end{aligned}$$ From this we obtain that the map $$\begin{aligned}
\bigotimes _{j=1}^{d}H^{2i_j}(\overline{X_{m_j}^{r_j}},{\mathbb}{Q}_{\ell}(i_j))^{G_L} \longrightarrow V(i_1,\dots,i_{d})^{G_L}\end{aligned}$$ is surjective, where $L$ is a finite extension of $k$ with $[L:k]=N$. The assertion now follows from assumption (ii) of the theorem and the commutative diagram $$\begin{aligned}
{\footnotesize
\xymatrix{\displaystyle
\bigoplus_{i_1+\cdots+i_{d}=i \ \ }\bigotimes _{j=1}^{d}{{\rm CH}}^{i_j}(X_{m_j}^{r_j}\times L)\otimes {\mathbb}{Q}_{\ell} \ar[r] \ar[d] &\displaystyle\bigoplus_{i_1+\cdots+i_{d}=i \ \ }\bigotimes _{j=1}^dH^{2i_j}(\overline{X_{m_j}^{r_j}},{\mathbb}{Q}_{\ell}(i_j))^{G_L} \ar@{>>}[d]\\
{{\rm CH}}^i(X_{m_1}^{r_1}\times X_{m_2}^{r_2}\times \cdots\times X_{m_{d}}^{r_{d}}\times L)\otimes{\mathbb}{Q}_{\ell}\ar[r]&H^{2i}(\overline{X_{m_1}^{r_1}\times X_{m_2}^{r_2}\times \cdots \times X_{m_{d}}^{r_{d^{\prime}}}},{\mathbb}{Q}_{\ell}(i))^{G_L}.
}}\end{aligned}$$
Let $Y$ be the product of all factors of $X$ of odd dimension and let $Z$ be the product of all factors of $X$ of even dimension. From Künneth formula and isomorphisms , we have $$\begin{aligned}
H^{2i}(\overline{X},{\mathbb}{Q}_{\ell}) \simeq \bigoplus_{i_1+i_2=i} H^{2i_1}(\overline{Y},{\mathbb}{Q}_{\ell})\otimes H^{2i_2}(\overline{Z},{\mathbb}{Q}_{\ell}).\end{aligned}$$
A basis of $\bigl(H^{2i_1}(\overline{Y},{\mathbb}{Q}_{\ell}(i_1))\otimes H^{2i_2}(\overline{Z},{\mathbb}{Q}_{\ell}(i_2))\bigr)^{G_k}$ corresponds to pairs $(\alpha,\beta)$ such that $\alpha\beta=q^i$ where $\alpha$ and $\beta$ are an eigenvalue of geometric Frobenius acting on $H^{2i_1}(\overline{Y},{\mathbb}{Q}_{\ell})$ and $H^{2i_2}(\overline{Z},{\mathbb}{Q}_{\ell})$ respectively.
Let $m$ (resp. $n$) be the least common multiple of $\{m_j\;|\; \text{$r_j$ is odd (resp. even)}\}$. From assumption (i) of the theorem, ${\rm gcd}(m,n)\leq2$. From Künneth formula and Lemma \[l1\], there is a positive integer $N$ such that for every such pairs $(\alpha,\beta)$, we have $$\begin{aligned}
\alpha^N=q^i\beta^{-N} \in {\mathbb}{Q}(\zeta_m)\cap {\mathbb}{Q}(\zeta_n)={\mathbb}{Q}.\end{aligned}$$ From this we obtain that the map $$\begin{aligned}
\bigoplus_{i_1+i_2=i} H^{2i_1}(\overline{Y},{\mathbb}{Q}_{\ell}(i_1))^{G_E}\otimes H^{2i_2}(\overline{Z},{\mathbb}{Q}_{\ell}(i_2))^{G_E}\longrightarrow H^{2i}(\overline{X},{\mathbb}{Q}_{\ell}(i))^{G_E}\end{aligned}$$ is surjective, where $E/k$ is a finite extension of degree $N$. Hence the assertion follows from Theorem \[t\](2)(3) and a commutative diagram $$\begin{aligned}
{\footnotesize
\xymatrix{\displaystyle
\bigoplus_{i_1+i_2=i}{{\rm CH}}^{i_1}(Y\times E)\otimes {{\rm CH}}^{i_2}(Z\times E)\otimes {\mathbb}{Q}_{\ell} \ar[r] \ar[d] &\displaystyle\bigoplus_{i_1+i_2=i \ \ }H^{2i_1}(\overline{Y},i_1)^{G_E}\otimes H^{2i_2}(\overline{Z},i_2)^{G_E}\ar@{>>}[d]\\
{{\rm CH}}^i(X\times E)\otimes{\mathbb}{Q}_{\ell}\ar[r]&H^{2i}(\overline{X},{\mathbb}{Q}_{\ell}(i))^{G_E}.
}}\end{aligned}$$ Here $H^{2i_1}(\overline{Y},i_1)$ and $H^{2i_2}(\overline{Z},i_2)$ denote $H^{2i_1}(\overline{Y},{\mathbb}{Q}_{\ell}(i_1))$ and $H^{2i_2}(\overline{Z},{\mathbb}{Q}_{\ell}(i_2))$ respectively.
Dimension of $\ell$-adic cohomology
-----------------------------------
For a variety $X$ as in Theorem \[t\], the ${\mathbb}{Q}_{\ell}$-vector space $H^{2i}(\overline{X},{\mathbb}{Q}_{\ell}(i))^{G_k}$ does not vanish. In case where $X$ is a variety as in (2) of Theorem \[t\], $H^{2i}(\overline{X},{\mathbb}{Q}_{\ell}(i))^{G_k}$ comes from tensor products of Lefschetz motives. In other cases, $H^{2i}(\overline{X},{\mathbb}{Q}_{\ell}(i))^{G_k}$ may contain a subspace which does not comes from tensor products of Lefschetz motives. For example, $H^{2}(\overline{X_m^1\times X_m^1},{\mathbb}{Q}_{\ell}(1))^{G_k}$ contains the subspace which is isomorphic to ${\rm Hom}_{{\mathbb}{Q}_{\ell}}(V_{\ell}(J_m),V_{\ell}(J_m))^{G_k}$. Here $J_m$ is the Jacobian variety of $X_m^1$ over $k$ and $V_{\ell}(J_m)$ is defined as follows: let $J_m[n]$ denote the group of elements $x \in J_m(\overline{k})$ such that $nx=0$. We define $T_{\ell}(J_m)$ to be the projective limit of the groups $J_m[\ell^n]$ with respect to the maps $J_m[\ell^{n+1}]\stackrel{\times \ell}{\longrightarrow} J_m[\ell^n]$. Then we define $V_{\ell}(J_m)={\mathbb}{Q}_{\ell}\otimes_{{\mathbb}{Z}_{\ell}}T_{\ell}(J_m)$. It is well known that $T_{\ell}(J_m)$ is a free module over ${\mathbb}{Z}_{\ell}$ of rank $(m-1)(m-2)$. Therefore $V_{\ell}(J_m)$ is a ${\mathbb}{Q}_{\ell}$-vector space of dimension $(m-1)(m-2)$.
The dimension of $H^{2i}(\overline{X},{\mathbb}{Q}_{\ell}(i))^{G_k}$ over ${\mathbb}{Q}_{\ell}$ can be computed for a variety $X$ as in (2)–(4) of Theorem \[t\].
First let $X=X_{m_1}^{r_1}\times\cdots\times X_{m_d}^{r_d}$ be a variety as in (2). By the proof of Theorem \[t\](2), we have $$\begin{aligned}
H^{2i}(&\overline{X},{\mathbb}{Q}_{\ell}(i))^{G_k}\\
&=\bigoplus_{(i_1,\dots,i_d)\in I(i)}H^{2i_1}(\overline{X_{m_1}^{r_1}},{\mathbb}{Q}_{\ell}(i_1))^{G_k}\otimes \cdots \otimes H^{2i_d}(\overline{X_{m_d}^{r_d}},{\mathbb}{Q}_{\ell}(i_d))^{G_k}\end{aligned}$$ where $$I(i)=\Big\{(i_1,\dots,i_d)\;\Big|\; i_1+\cdots+i_d=i \;\text{and}\; 0\le i_j\le {\rm min}\{r_j,i\}\; \text{for all $j$}\Big\}.$$ The direct summand of right hand side is isomorphic to ${\mathbb}{Q}_{\ell}$. Hence we have $$\begin{aligned}
\label{d1}
\dim_{{\mathbb}{Q}_{\ell}}H^{2i}(\overline{X},{\mathbb}{Q}_{\ell}(i))^{G_k}=\sharp I(i)\end{aligned}$$ where $\sharp $ denotes the cardinality of a finite set.
Next let $X=X_{m_1}^{r_1}\times\cdots\times X_{m_d}^{r_d}$ be a variety as in (3). By the proof of Theorem \[t\](3), we have $$\begin{aligned}
H^{2i}(\overline{X},{\mathbb}{Q}_{\ell}(i)&)^{G_k}\\
=\bigoplus_{(i_1,\dots,i_{d})\in I^{\prime}(i)}&H^{2i_1}(\overline{X_{m_1}^{r_1}},{\mathbb}{Q}_{\ell}(i_1))^{G_k}\otimes \cdots \otimes H^{2i_{d}}(\overline{X_{m_{d}}^{r_{d}}},{\mathbb}{Q}_{\ell}(i_{d}))^{G_k}\\
&\oplus \bigoplus _{(i_1,\dots,i_{d})\in I^{''}(i)} H^{2i_1}(\overline{X_{m_1}^{r_1}},{\mathbb}{Q}_{\ell}(i_1))^{G_k}\otimes \cdots \otimes H^{2i_{d}}(\overline{X_{m_{d}}^{r_{d}}},{\mathbb}{Q}_{\ell}(i_{d}))^{G_k}\end{aligned}$$ where $$\displaystyle I^{\prime}(i)=\Big\{(i_1,\dots,i_{d})\;\Big|\; \sum_{j=1}^{d}i_j=i,\; 0\le i_j\le {\rm min}\{r_j,i\}\;\text{and}\; 2i_j\neq r_j\; \text{for all $j$}\Big\}$$ and $$\begin{aligned}
I^{''}(i)=\Bigg\{(i_1,\dots,i_{d})\;\Bigg|
\begin{array}{l}
\displaystyle \sum_{j=1}^{d}i_j=i,\; 0\le i_j\le {\rm min}\{r_j,i\}\;\text{for all $j$,}\\
2i_j= r_j\; \text{for some $j$}
\end{array}\Bigg\}\end{aligned}$$ The direct summand of the first part of the right hand side is isomorphic to ${\mathbb}{Q}_{\ell}$. Hence we have $$\begin{aligned}
\label{d2}
\dim_{{\mathbb}{Q}_{\ell}}H^{2i}(\overline{X},{\mathbb}{Q}_{\ell}(i))^{G_k}=\sharp I^{\prime}(i)+\sum_{(i_1,\dots,i_{d^{\prime}})\in I^{''}(i)}\,\prod_{2i_j=r_j} \dim_{{\mathbb}{Q}_{\ell}}H^{r_j}(\overline{X_{m_j}^{r_j}},{\mathbb}{Q}_{\ell}(i_j))^{G_k}.\end{aligned}$$ The description of $\dim_{{\mathbb}{Q}_{\ell}}H^{r_j}(\overline{X_{m_j}^{r_j}},{\mathbb}{Q}_{\ell}(r_j/2))^{G_k}$ is known (cf. [@Sh3 p. 125]). We recall it here. We use the notation in §2.1. Let $m$ be a positive integer prime to $p$ and let $r$ be a positive even integer. We denote the cardinality of $k$ by $q$. Then we have $$\begin{aligned}
\dim_{{\mathbb}{Q}_{\ell}}H^{r}(\overline{X_{m}^{r}},{\mathbb}{Q}_{\ell}(r/2))^{G_k}=1+\sharp {\mathfrak}{B}_{m,r,q},\end{aligned}$$ where $${\mathfrak}{B}_{m,r,q}=\big\{\gamma \in {\mathfrak}{D}_{m,r}\;|\;j(\gamma)=q^{r/2}\big\}.$$ To give another description of ${\mathfrak}{B}_{m,r,q}$, we introduce some notation. For $\gamma=(\gamma_0,\dots,\gamma_{r+1}) \in {\mathfrak}{D}_{m,r}$, we define $$\parallel \gamma \parallel=\sum_{i=1}^{r+1}\big\langle \frac{\gamma_i}{m}\big\rangle-1,$$ where $\langle x\rangle$ is the fractional part of $x\in {\mathbb}{Q}/{\mathbb}{Z}$. We write $H$ for the subgroup of $({\mathbb}{Z}/m)^{\times}$ generated by $p$, and $f$ for the order of $H$. Then for a sufficiently large $q$, the set ${\mathfrak}{B}_{m,r,q}$ is equal to the following set $$\Big\{\gamma \in {\mathfrak}{D}_{m,r}\;\Big|\;\sum_{h \in H}\parallel ht\gamma \parallel=rf/2, \;{}^{\forall}t \in ({\mathbb}{Z}/m)^{\times}\Big\}.$$
Lastly let $X$ be a variety as in (4). Let $Y$ and $Z$ be the varieties as in the proof of Theorem \[t\] (4) (i.e. $X=Y \times Z$). By the proof, we have $$H^{2i}(\overline{X},{\mathbb}{Q}_{\ell}(i))^{G_k}\simeq \bigoplus_{i_1+i_2=i} H^{2i_1}(\overline{Y},{\mathbb}{Q}_{\ell}(i_1))^{G_k}\otimes H^{2i_2}(\overline{Z},{\mathbb}{Q}_{\ell}(i_2))^{G_k}.$$ Therefore the dimension of $H^{2i}(\overline{X},{\mathbb}{Q}_{\ell}(i))^{G_k}$ can be computed using the above descriptions .
Alternative proof of Corollary \[cf\]
-------------------------------------
Let $k$ be a finite field of characteristic $p$ and $m$ be a positive integer prime to $p$. For an integer $r\geq 0$, let $V_m^r \subset {\mathbb}{P}_{k}^{r+1}$ be the Fermat variety of dimension $r$ and of degree $m$ defined by the following equation $$\begin{aligned}
x_0^m+x_1^m+x_2^m+\cdots +x_{r+1}^m=0.\end{aligned}$$ Shioda and Katsura [@SK] gave the “inductive structure" of Fermat varieties of a common degree and of various dimensions, and proved that the Tate conjecture holds for Fermat surfaces using this structure. We here give an alternative proof of Corollary \[cf\] using the inductive structure of Fermat varieties and Theorem \[t\](1).
\[isf\] Let $m$ be a positive integer prime to $p$ and $k$ be a finite field with $\zeta_{2m} \in k$. Let $r$ and $s$ be positive integers. Then there is a commutative diagram[:]{} $$\begin{aligned}
\xymatrix{
& Z_m^{r,s} \ar[d]^{\beta } \ar[r]^{\pi } &Z_m^{r,s}/\mu_m \ar[d]^{\overline{\psi} }\\
V_m^{r-1}\times V_m^{s-1} \ar@{^{(}->}[r]^{ \ j}&V_m^r\times V_m^s \ar@{.>}[r]^{\varphi } &V_m^{r+s} & V_m^{r-1}\amalg V_m^{s-1}\ar@{_{(}->}[l]_{i \ \ }.
}\end{aligned}$$ Here $\mu_m$ is the group of $m$-th roots of unity and the above maps are defined as follows[:]{}
$\varphi :$ the rational map of degree $m$ defined by
$\varphi ((x_0:\cdots:x_{r+1}),(y_0:\cdots:y_{s+1}))$
$=(x_0y_{s+1}:\cdots:x_{r+1}y_{s+1}:\zeta_{2m}x_{r+1}y_0:\cdots:\zeta_{2m}x_{r+1}y_{s+1})$,
$j : (((x_0:\cdots:x_r),(y_0:\cdots:y_s)))\mapsto ((x_0:\cdots:x_r:0),(y_0:\cdots:y_s:0))$,
$i=1_1\amalg i_2 :$
$\begin{cases}
i_1 : (x_0:\cdots:x_r)\longmapsto (x_0\cdots:x_r:0:\cdots:0)\\
i_2 : (y_0:\cdots:y_s)\longmapsto (0:\cdots:0:y_0:\cdots:y_s),
\end{cases}$
$\beta :$ the blowing up of $V_m^r\times V_m^s$ along $V_m^{r-1}\times V_m^{s-1}$,
$\pi :$ the quotient map of degree $m$,
$\overline{\psi } :$ the blowing up of $V_m^{r+s}$ along $V_m^{r-1}\amalg V_m^{s-1}$.\
The action of $\mu_m$ on $V_m^r\times V_m^s$ is defined by
$((x_0:\cdots:x_{r+1}),(y_0:\cdots:y_{s+1}))$
$\mapsto ((x_0:\cdots:x_r:\zeta_mx_{r+1}),(y_0:\cdots:y_s:\zeta_my_{s+1}))$,\
and this $\mu_m$-action naturally extends to that on $Z^{r,s}_m$.
We can prove Corollary \[cf\] by induction on dimension from the following lemma:
Let $m$ be a positive integer prime to $p$ and $k$ be a finite field with $\zeta_{2m} \in k$. Let $W$ be a projective smooth variety over $k$. If ${\boldsymbol{T}}(W/k,i-1)$ and ${\boldsymbol{T}}(V_m^1\times V_m^1\times W/k,i)$ hold, then ${\boldsymbol{T}}(V_m^2\times W/k,i)$ also holds.
For a variety $V$, put ${{\rm CH}}^i(V)_{{\mathbb}{Q}_{\ell}}:={{\rm CH}}^i(V)\otimes _{{\mathbb}{Z}}{\mathbb}{Q}_{\ell}$. Let $X:=V_m^2\times W$. By Theorem \[isf\], we have the diagram [$$\begin{aligned}
\xymatrix{
& Z_m^{1,1}\times W \ar[d]^{\beta } \ar[r]^{\pi } &Z_m^{1,1}/\mu_m\times W \ar[d]^{\overline{\psi} }\\
V_m^0\times V_m^0\times W \ar@{^{(}->}[r]^{j}&V_m^1\times V_m^1\times W \ar@{.>}[r]^{ \ \ \varphi } &X & (V_m^0\amalg V_m^0)\times W \ar@{_{(}->}[l]_{i \ \ \ \ \ }.
}\end{aligned}$$]{} Here the above maps are similar to that in Theorem \[isf\]. From this diagram, we have the following isomorphisms (cf. [@SK $\S$2], [@Fu Example 1.7.6]) $$\begin{aligned}
\nonumber
H^{2i}(\overline{X},{\mathbb}{Q}_{\ell}(i))&\oplus H^{2(i-1)}(\overline{W},{\mathbb}{Q}_{\ell}(i-1))^{\oplus 2m} \\\label{i1}
\simeq & H^{2i}(\overline{V_m^1\times V_m^1\times W},{\mathbb}{Q}_{\ell}(i))^{\mu_m}\oplus H^{2(i-1)}(\overline{W},{\mathbb}{Q}_{\ell}(i-1))^{\oplus m^2}
\\\label{i2}
{{\rm CH}}^i(X)_{{\mathbb}{Q}_{\ell}}\oplus {{\rm CH}}^{i-1}&(W)_{{\mathbb}{Q}_{\ell}}^{\oplus 2m} \simeq {{\rm CH}}^i(V_m^1\times V_m^1\times W)^{\mu_m}_{{\mathbb}{Q}_{\ell}}\oplus {{\rm CH}}^{i-1}(W)_{{\mathbb}{Q}_{\ell}}^{\oplus m^2}.\end{aligned}$$ Here $H^{2i}(\overline{V_m^1\times V_m^1\times W},{\mathbb}{Q}_{\ell}(i))^{\mu_m}$ and ${{\rm CH}}^i(V_m^1\times V_m^1\times W)^{\mu_m}_{{\mathbb}{Q}_{\ell}}$ are the $\mu_m$-invariant subspace of $H^{2i}(\overline{V_m^1\times V_m^1\times W},{\mathbb}{Q}_{\ell}(i))$, ${{\rm CH}}^i(V_m^1\times V_m^1\times W)_{{\mathbb}{Q}_{\ell}}$ respectively. The $G_k$-action and the $\mu_m$-action commute, and the isomorphisms are compatible with the cycle class map . Therefore if ${\boldsymbol{T}}(W/k,i-1)$ and ${\boldsymbol{T}}(V_m^1\times V_m^1\times W/k,i)$ hold, then ${\boldsymbol{T}}(V_m^2\times W/k,i)$ also holds.
A zeta value of products of four curves
=======================================
Let $X$ be a projective smooth fourfold over a finite field. In his paper [@Ko] Theorem 2, Kohmoto gave a formula of the special value of the zeta function of $X$ at $s=2$ assuming the Tate-Beilinson conjecture $\boldsymbol{TB}^2(X)$ in the sense of [@Ko] and that ${{\rm CH}}^2(X)$ is finitely generated. His work is based on formulae of zeta values established by Bayer, Neukirch, Schneider, Milne and Kahn ([@BN], [@Sc], [@Mi] and [@Ka]). Now let $X/k$ be as in Theorem \[t\](1)(b). The Galois group $G_k$ of $\overline{k}/k$ acts semisimply on $H^i(\overline{X},{\mathbb}{Q}_{\ell})$ by a theorem of Tate [@Ta2]. Hence $\boldsymbol{TB}^2(X)$ holds for $X/k$ by Theorem \[t\](1)(b) and Corollary \[ct\](1)(4), and we obtain the following corollaries using the result of Kohmoto mentioned above. This means that we give a new example which satisfy the assumptions above.
We introduce some notation. Let $W_{\nu }\Omega _{X,{\rm log}}^n$ be the $\acute {\text{e}}$tale sheaf of the logarithmic part of the Hodge-Witt sheaf $W_{\nu }\Omega _{X}^n$ ([@Il]). We define ${\mathbb}{Z}/m(2)$ and ${\mathbb}{Q}/{\mathbb}{Z}(2)$ as follows: $$\begin{aligned}
{\mathbb}{Z}/m(2):=&\left\{ \begin{array}{ll}
\mu_m^{\otimes 2} & \text{if } (m,p)=1 \\
\mu_{m^{\prime }}^{\otimes 2}\oplus W_{\nu }\Omega _{X,{\rm log}}^2[-2] &\text{if } m=p^{\nu}m^{\prime } \text{ for }\nu>0 \text{ and }(m^{\prime },p)=1
\end{array}\right.\\
{\mathbb}{Q}/{\mathbb}{Z}(2):=& \displaystyle \varinjlim _m {\mathbb}{Z}/m(2).\end{aligned}$$ We define ${\mathbb}{Z}(2)$ as the $\acute{\text{e}}$tale sheafification on $X$ of the presheaf of cochain complexes $$\begin{aligned}
U\longmapsto z^2(U,*)[-4],\end{aligned}$$ where $z^2(U,*)$ denotes Bloch’s cycle complex ([@Bl]) $$\begin{aligned}
\cdots \longrightarrow \ z^2(U,r) \ \stackrel{d_r}{\longrightarrow } \ z^2(U,r-1) \ \stackrel{d_{r-1}}{\longrightarrow } \cdots \stackrel{d_1}{\longrightarrow } \ z^2(U,0).\end{aligned}$$ Here $z^2(U,r)$ is placed in degree $-r$. We define the unramified cohomology group $H_{ur}^3(k(X),{\mathbb}{Q}/{\mathbb}{Z}(2))$ as the kernel of the boundary map of a localization sequence $$\begin{aligned}
H^3({\rm Spec}(k(X)),{\mathbb}{Q}/{\mathbb}{Z}(2))\longrightarrow \bigoplus_{x \in X^{(1)}}H_x^4({\rm Spec}({\mathcal}{O}_{X,x}),{\mathbb}{Q}/{\mathbb}{Z}(2)),\end{aligned}$$ where $k(X)$ is the function field of $X$, and $X^{(1)}$ is the set of points of $X$ of codimension one.
By Bloch [@Bl] and Geisser$-$Levine [@GL], we know that the complex ${\mathbb}{Z}(2)\otimes{\mathbb}{Z}/m$ is isomorphic to ${\mathbb}{Z}/m(2)$ defined above in the derived category of complexes of étale sheaves.
By Theorem \[t\] and results of Kohmoto [@Ko Theorem 1, Theorem 2], we obtain the following corollaries:
\[ft\] Let $X/k$ be as in Theorem \[t\](1)(b) and assume $\dim X=4$.\
The unramified cohomology group $H_{ur}^3(k(X),{\mathbb}{Q}/{\mathbb}{Z}(2))$ is finite.\
$H^5(X,{\mathbb}{Z}(2))$ is finite.\
The intersection pairing $$\begin{aligned}
{{\rm CH}}^2(X)\times {{\rm CH}}^2(X)\longrightarrow {{\rm CH}}^4(X)\simeq {{\rm CH}}_0(X)\stackrel{\rm{deg}}{\longrightarrow }{\mathbb}{Z}\end{aligned}$$ is non-degenerate when tensored with ${\mathbb}{Q}$.\
The subgroup ${{\rm CH}}^2(X,i)_{tor}$ of torsion elements of the higher Chow group ${{\rm CH}}^2(X,i)$ is finite for $i=1,2,3,$ and zero for $i\geq 4$.
Corollary \[ft\](4) holds for arbitrary projective smooth fourfolds over finite fields by [@Ko Theorem 1(d)]
For a smooth projective variety $X$ over a finite field $k={\mathbb}{F}_q$, the zeta function of $X$ is defined as follows: $$\begin{aligned}
\zeta (X,s):=Z(X/k,q^{-s}) \ \text{ with } \ Z(X/k,t):={\rm exp}\Bigl( \sum_{n\geq 1} \frac{\sharp X({\mathbb}{F}_{q^n})}{n} t^n\Bigr),\end{aligned}$$ where $\sharp $ denotes the cardinality of a finite set. We define the special value of $\zeta (X,s)$ at $s=2$ as $$\begin{aligned}
\zeta (X,2)^{*}:= \lim _{s\rightarrow 2} \zeta (X,s)(1-q^{2-s})^{-\rho _2},\end{aligned}$$ where $\rho _2$ is the order of $\zeta (X,s)$ at $s=2$.
Let $X/k$ be as in Theorem \[t\](1)(b) and assume $\dim X=4$. Then the following formula holds[:]{} $$\begin{aligned}
&\zeta (X,2)^{*}=\\
&(-1)^{S(2)}\cdot q^{\chi (X,{\mathcal}{O}_X,2)}\cdot \frac{\mid H_{ur}^3(k(X),{\mathbb}{Q}/{\mathbb}{Z}(2))\mid ^2}{\mid H^5(X,{\mathbb}{Z}(2))\mid\cdot R_1 }\cdot \prod_{i=0}^3\mid {{\rm CH}}^2(X,i)_{tor}\mid ^{2\cdot (-1)^i}. \end{aligned}$$ Here $R_1$ is the order of the cokernel of the map $$\begin{aligned}
{{\rm CH}}^2(X)\longrightarrow {\rm Hom}({{\rm CH}}^2(X),{\mathbb}{Z})\end{aligned}$$ induced by the intersection pairing, $S(2)$ and $\chi (X,{\mathcal}{O}_X,2)$ are defined as $$\begin{aligned}
&S(2) :=\sum_{a>4}\rho _{\frac{a}{2}} \ \ \ \bigl( \rho _{\frac{a}{2}}:=\rm{ord}_{s=\frac{a}{2}}\zeta (X,s)\bigr),\\
\chi (X,{\mathcal}{O}_X,2):=\sum_{i,j}&(-1)^{i+j}(2-i)\dim_kH^j_{Zar}(X,\Omega _X^i) \ (0\leq i\leq 2, 0\leq j \leq 4).\end{aligned}$$
**Acknowledgements** The author expresses his gratitude to Professors Kanetomo Sato and Thomas Geisser for many helpful suggestions and comments. He also thank Professors Takao Yamazaki and Noriyuki Otsubo for valuable comments.
[00]{} Bass, H.: Some Problems in “classical" algebraic $K$-theory, *“Classical" Algebraic K-Theory, and Connections with Arithmetic* (Proc. Conf. , Seattle, Battelle Memorial Inst., 1972) Lecture Notes in Math. **342**, Berlin, Springer, 1973, pp. 1–73 B$\acute {\text{a}}$yer, P., Neukirch, J.: On values of zeta functions and $\ell$-adic Euler characteristics. Invent. Math. **50**, 35-64 (1978/79) Beilinson, A. A.: Height pairings between algebraic cycles. In: Yu. I. Manin (ed.), $K$-theory, *Arithmetic and Geometry*, Lecture Notes in Math. **1289**, Berlin, Springer, 1987, pp. 1–27 Bloch, S.: Algebraic cycles and higher $K$-theory. Adv. Math. **64**, 267–304 (1986) Colliot-Th$\acute{\text{e}}$l$\grave{\text{e}}$ne, J.-L., Sansuc, J.-J., Soul$\acute{\text{e}}$, C.: Torsion dans le groupe de Chow de codimension deux. Duke Math. J. **50**, 763–801 (1983) Deligne, P.: La conjecture de Weil I, Publ. Math. Inst. Hautes Étude Sci. [**43**]{} 273–308 (1974) Fulton, W.: *Intersection Theory, 2nd ed*. Ergeb. Math. Grenzgeb. (3)2, Springer, Berlin (1998) Friedlander, E. M., Suslin, A.: The spectral sequence relating algebraic K-theory to motivic cohomology. Ann. Sci. $\acute{\text{E}}$cole Norm. Sup. (4) **35**, no. 6, 773–875 (2002) Geisser, T.: Tate’s conjecture, algebraic cycles and rational $K$-theory in characteristic $p$. $K$-theory **13**, 109–122 (1998) Geisser, T., Levine, M.: The $K$-theory of fields in characteristic $p$, Invent. Math. [**139**]{}, 459–493 (2000) Grothendieck, A.: *Cohomologie $\ell$-adique et Fonctions $L$*. Lecture Notes in Math. [**589**]{}, Berlin, Springer, (1977) Jannsen, U.: Continuous Étale Cohomology, Math. Ann. [**280**]{}, 207–245 (1988) Illusie, L.: Complexe de De Rham-Witt et cohomologie cristalline. Ann. Sci. $\acute{\text{E}}$cole Norm. Sup. (4) **12**, 501–661 (1979) Kahn, B.: $\acute{\text{E}}$quivalences rationnelle et num$\acute{\text{e}}$rique sur certaines vari$\acute{\text{e}}$t$\acute{\text{e}}$s de type Ab$\acute{\text{e}}$lien sur un corps fini. Ann. Sci. $\acute{\text{E}}$cole. Norm. Sup. (4) **36**, 977–1002 (2003) Kimura, S.: Chow groups are finite dimensional, in some sense. Math. Ann. [**331**]{}, 173–201 (2005) Kohmoto, D.: A generalization of the Artin-Tate formula for fourfolds. J. Math. Sci. Univ. Tokyo [**17**]{} (2010), no. 4, 419–453 (2011). Lichtenbaum, S.: Values of zeta functions at non-negative integers. In: Number theory, Noordwijkerhout 1983, Lecture Notes in Math. **1068**, Springer,Berlin, 1984, pp. 127–138 Milne, J. S.: Étale cohomology, Princeton University Press, Princeton (1980). Milne, J. S.: Values of zeta functions of varieties over finite fields. Amer. J. Math. **108**, 297–360 (1986) Shioda, T., Katsura, T.: On Fermat varieties. T$\hat {\text{o}}$hoku J. Math. **31**, 97–115 (1979) Shioda, T.: The Hodge Conjecture and the Tate Conjecture for Fermat Varieties. *Proc. Japan Acad.* **55**, 111–114 (1979) Shioda, T.: The Hodge Conjecture for Fermat Varieties. Math. Ann. **245**, 175–184 (1979) Shioda, T.: Some observations on Jacobi sums. Advanced Studies in Pure Math., **12**, North Holland, Kinokuniya, pp. 119–135 (1987) Schneider, P.: On the values of the zeta function of a variety over a finite field. Comp. Math. **46**, 133–143 (1982) Soul$\acute{\text{e}}$, C.: Groupes de Chow et $K$- th$\acute{\text{e}}$orie de vari$\acute{\text{e}}$t$\acute{\text{e}}$s sur un corps fini. Math. Ann. **268**, 317–345 (1984) Spiess, M.: Proof of the Tate conjecture for products of elliptic curves over finite fields. Math. Ann. [**314**]{}, 285–290 (1999) Sugiyama, R.: Tate conjecture for products of Fermat varieties over finite fields, arXiv. Tate, J.: Algebraic cycles and poles of zeta-functions. *Arithmetical Algebraic Geometry*. Harper and Row, New York, 93–110 (1965) Tate, J.: Endomorphisms of abelian varieties over finite fields. Invent. Math. [**2**]{}, 134–144 (1966) Tate, J.: Conjectures on Algebraic Cycles in $\ell$-adic Cohomology. Proc. Symp. Pure Math. **55.1**, 71–83 (1994) Weil, A.: Number of solutions of equations in finite fields. Bull. Amer. Math. Soc., **55**, 497–508 (1949)
Rin Sugiyama\
Graduate School of Mathematics, Nagoya University\
Furo-cho, Chikusa-ku, Nagoya 464-8602, Japan\
e-mail: [email protected]
[^1]: The author is supported by JSPS Research Fellowships for Young Scientists.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We consider a system of two species of bosons of equal mass, with interactions $U^{a}(|{\mathbf{x}}|)$ and $U^{x}(|{\mathbf{x}}|)$ for bosons of the same and different species respectively. We present a rigorous proof–valid when the Hamiltonian does not include a species switching term–showing that, when $U^{x}(|{\mathbf{x}}|)>U^{a}(|{\mathbf{x}}|)$, the ground state is fully “polarized" (consists of atoms of one kind only). In the unpolarized phase the low energy excitation spectrum corresponds to two linearly dispersing modes that are even and odd under species exchange. The polarization instability is signaled by the vanishing of the velocity of the odd modes.'
address: ' Department of Physics, The University of Michigan, Ann Arbor, Michigan 48109-1120, USA'
author:
- 'A. G. Rojo'
title: 'Instabilities in a Two-Component, Species Conserving Condensate'
---
The experimental observation of Bose–Einstein condensation (BEC) in dilute atomic systems[@exp_bec] has triggered a very intense theoretical activity[@rmp]. Attention has broadened to include condensates with internal degrees of freedom, or multi–species BEC, which were realized for trapped rubidium[@myatt] and sodium[@stamper]. Early theoretical analysis of $m$–species condensates focuses on the large $m$ limit for Hamiltonians invariant under $U(m)$ transformations[@halperin], superfluid helim mixtures[@khalatinov] and spin–polarized hydrogen[@siggia]. In the context of BEC much of the theoretical attention concentrates on spinor condensates[@spinor] in which the internal degrees of freedom correspond to the different Zeeman states of the of a particular hyperfine manifold such as the $f=1$ manifold in sodium. In these cases the two–body interaction is invariant under under rotation in species space. On the other hand, for rubidium one has two non–degenerate internal states corresponding to two different manyfolds, and the two–body interaction is not invariant under $SU(2)$ rotations in species space[@esry]. The ground state and excitation spectrum of a two–species ($a$ and $b$) condensate of this kind was studied in mean field in Ref.[@meystre], where it was shown that the quasiparticle energy can become imaginary, signaling an instability when the inter–species repulsion $U^{x}$ is larger than the intra–species repulsion $U^{a}$. This kind of treatment follows–as do the majority of theoretical approaches to BEC–the Gross–Pitaevskii (GP) mean field equation[@gross]. On the other hand, rigorous results for BEC (and many body problems in general) are scarce and at the same time useful in providing control for approximate solutions. With this motivation, in this paper we consider a two–species system of interacting bosons, and show rigorously that the instability mentioned above corresponds to the tendency of the system to “polarize", the true ground state consisting of only one species of bosons when $U^{x}>U^{a}$. We also discuss the low–energy excitation spectrum and show that the instability is signaled by a divergence of the “compressibility" associated with exchanging particles from one species to the other at fixed total particle number.
The Hamiltonian for our two-species system ($a$ and $b$) of bosons of equal mass $m$ is given by ($\hbar =1$) $$H=K+U_{a}+U_{b}+U_{ab},$$ with $$\begin{aligned}
K&=&\sum_{\mathbf{k}}{ k^2\over 2m}\left[
\psi_a^{\dagger }({\mathbf{k}})%
\psi_a^{}({\mathbf{k}})+
\psi_b^{\dagger }({\mathbf{k}})%
\psi_b^{}({\mathbf{k}})
\right]
\\
U_{a}+U_{b}&=&\frac{1}{2}
\int d^3x\,d^3y\, U^a(|{\bf x}-{\bf y}|) \left[ \rho_a({\bf x})\rho_a({\bf y})
+\rho_b({\bf x})\rho_b({\bf y})
\right]
\\U_{ab} &=&\int d^3x\,d^3y\, U^x(|{\bf x}-{\bf y}|) \rho_a({\bf x})\rho_b({\bf y}).
\label{h1}\end{aligned}$$
In the above equations $K$ is the kinetic term, and the terms $(U_{a}+U_{b})$ and $U_{ab}$ correspond to the interaction between bosons of the same and different species respectively. The operators $\psi_i^{}({\mathbf{x}})$ destroy a boson of species $i$ ($i=a,b$) at position ${\mathbf{x}}$, and obey the following commutation relations $[\psi_i^{\dagger}({\mathbf{x}}),
\psi_j^{}({\mathbf{x'}})]=\delta_{i,j}\delta ({\mathbf{x}}-{\mathbf{x}'})$. Also, $\rho_i({\bf x})=\psi^{\dagger}_i({\bf x})\psi_i({\bf x})$, and $\psi_i^{}({\mathbf{k}})=V^{-1/2}\int d^3x \, \exp(i{\bf k}\cdot{\bf x})
\psi_i^{}({\mathbf{x}})$ with $V$ the total volume.
Using a variational argument we will prove that, for potentials satisfying $U^x(|{\mathbf{x}}|)>U^a(|{\mathbf{x}}|)$ the ground state of the above Hamiltonian is completely polarized. By “polarized" we mean that either of the two situations is realized: $\left\langle \hat{N}_{a}\right\rangle =0,\left\langle \hat{N}_{b}\right\rangle =N$; or $%
\left\langle \hat{N}_{a}\right\rangle =N,\left\langle \hat{N}_{b}\right\rangle =0,$ with $N$ the total particle number and $\hat{N}_{i}=\int d^3x \, \rho_i({{\bf{x}}})$.
We note that, since the masses are the same for both species, the kinetic term of Hamiltonian (\[h1\]) commutes with the “rising” operator defined as: $${\cal{O}}_{R}=\int d^3x \,
\psi_a^{\dagger}({\mathbf{x}})
\psi_b^{}({\mathbf{x}})
\equiv \sum_{\bf k}
\psi_a^{\dagger}({\mathbf{k}})\psi_b^{}({\mathbf{k}})$$ which conserves the total number of particles but converts particles of type $b$ into particles of type $a.$ More specifically, for the case of different masses for each species we have that $[{\cal{O}}_{R},K]=\left({1\over 2m_b}-{1\over 2m_a}\right)\sum_{\bf k}k^2
\psi_a^{\dagger}({\mathbf{k}})\psi_b^{}({\mathbf{k}})$.
We have in mind the alkali atoms, which have a hard core interaction, meaning that the exact wave functions vanish when the coordinates of two atoms of either species coincide.
Since the Hamiltonian conserves the particle number for each species, we can start with the normalized ground state wave function $|\Psi_0(N_a,N_b)\rangle$ in the subspace of $N_a$ ($N_b$) particles of species $a$($b$). The unpolarized situation corresponds to $N_a=N_b=N/2$. Now let us consider the normalized variational wave function $|\Psi_v(N_a+N_b,0)\rangle$ obtained by the action of the rising operator on $|\Psi_0(N_a,N_b)\rangle$ $N_b$ times: $$|\Psi_v(N_a+N_b,0)\rangle={1\over\sqrt{N_b!}}({\cal{O}_{R}})^{N_b}|\Psi_0(N_a,N_b)\rangle.
\label{psiv}$$
Note that $\Psi_v$ represents a completely polarized wave function, with particles of species $a$ only. Since $[{\cal{O}_{R}},K]=0$, we have
$$\langle \Psi_v(N_a+N_b,0)|K |\Psi_v(N_a+N_b,0)\rangle
=\langle \Psi_0(N_a,N_b)|K |\Psi_0(N_a,N_b)\rangle,$$
meaning that the completely polarized variational wave function and the ground state of the subspace $(N_a,N_b)$ have the same expectation value of the kinetic energy. We stress that the function $|\Psi_v(N_a+N_b,0)\rangle$ as defined in Eq. (\[psiv\]) is normalized only because the exact wave function $|\Psi_0(N_a,N_b)\rangle$ vanishes when any two coordinates coincide. Otherwise we would have to worry about permutation factors whenever coordinates coincide.
In order to compute the change in the potential energy we write the expectation values of $U^a$ and $U^x$ as an integral over all the multiparticle configurations $\Gamma\equiv(\Gamma_a,\Gamma_b)$ with coordinates $\{{\bf X}\}_\Gamma\equiv \{{\bf{x}}_{1,\Gamma_a},\cdots, {\bf{x}}_{N_a,\Gamma_a}
;{\bf{x}}_{1,\Gamma_b},\cdots, {\bf{x}}_{N_b,\Gamma_b}\}$. For each configuration $\Gamma$ let us regard the particle coordinates $\{{\bf X}\}_\Gamma$ as nodes of a graph. There are $N_a(N_a-1)/2$ and $N_b(N_b-1)/2$ edges connecting pairs of particles of species $a$ and $b$ respectively, and $N_aN_b$ edges connecting a particles of different species. The contribution to the expectation value of the potential energy $U_0=\langle U_{a}\rangle
+\langle U_{b}\rangle+\langle U_{ab}\rangle$ from this configuration is a function of the length of the edges of the graph, which can be classified in three sets: $\{ {\bf{\ell}}^{a}_{1,\Gamma_a},\cdots, {\bf{\ell}}^{a}_{N_a(N_a-1)/2,\Gamma_a}\}$, $\{ {\bf{\ell}}^{b}_{1,\Gamma_b},\cdots, {\bf{\ell}}^{b}_{N_b(N_b-1)/2,\Gamma_b}\}$, and $\{ {\bf{\ell}}^{ab}_{1,\Gamma},\cdots, {\bf{\ell}}^{ab}_{N_aN_b,\Gamma}\}
$, where $ {\bf{\ell}}^{a}_{i,\Gamma_a}$ is one of the possible lenghts $|{\bf{x}}_{k,\Gamma_a}-{\bf{x}}_{l,\Gamma_a}| $, etc. The potential energy is therefore given by
$$U_0=\int d^{3N} X_\Gamma \,|\Psi_0({\bf X}_\Gamma)|^2 \left[
\sum_{i=1}^{{N_a(N_a-1)/2}}U^{a}({{\bf{\ell}}^{a}_{i,\Gamma_a}})
+
\sum_{i=1}^{{N_b(N_b-1)/2}}U^{a}({{\bf{\ell}}^{b}_{i,\Gamma_b}})
+
\sum_{i=1}^{{N_aN_b}}U^{x}({{\bf{\ell}}^{ab}_{i,\Gamma}})
\right],$$
with $\Psi_0({\bf X}_\Gamma)$ the ground state wave–function in first quantization.
In the variational wave function all the edges of type $b$ and $ab$ are both changed to edges of type $a$. Therefore the contribution to the potential energy of each edge of configuration $\Gamma$ changes according to $U^{a}({{\bf{\ell}}^{b}_{i,\Gamma_b}})\rightarrow
U^{a}({{\bf{\ell}}^{b}_{i,\Gamma_b}})$, $U^{x}({{\bf{\ell}}^{ab}_{i,\Gamma}})\rightarrow
U^{a}({{\bf{\ell}}^{ab}_{i,\Gamma}})$. If we call $U_v$ the expectation value of the potential energy in the variational wave function, and $E_v$ the variational wave function, we obtain that $\Delta E=E_v -E_0=U_v-U_0$ is given by $$\Delta E=
\int d^N X_\Gamma \,|\Psi_0({\bf X}_\Gamma)|^2
\sum_{i=1}^{N_aN_b} \left[ U^{a}({{\bf{\ell}}^{ab}_{i,\Gamma}})
-U^{x}({{\bf{\ell}}^{ab}_{i,\Gamma}})\right].$$
We see that when the interactions satisfy $U^x(|{\mathbf{x}}|) > U^a(|{\mathbf{x}}|)$ for all values of ${\bf{x}}$, $\Delta E$ is a sum of negative terms. In general the potentials have a repulsive short-range term and a long-range attractive tail. The condition for validity of our proof is that there are no “crossings” of the potentials $U^x(|{\mathbf{x}}|)$ and $U^a(|{\mathbf{x}}|)$ as a function of the coordinate. The simplest approximation will be to take the attractive components of both potentials as equivalent (of the form $-C_6/R^6$) and differing short range components. This dependence is consistent with calculations for ultra-cold Na collisions[@julienne].
Since the completely polarized variational wave function has lower energy than the true ground state of the partially polarized subspace, the gound state will be completely polarized. It is also evident that the proof is also valid in the case different intra–species interactions $U^a({\mathbf{x}})$ and $U^b({\mathbf{x}})$. As long as $U^{x} >
U^{a},U^{b}$, the ground state is polarized with particles of type $a$ ($b$) when $U^a(|{\mathbf{x}}|) < U^b(|{\mathbf{x}}|)$ ($
(U^b(|{\mathbf{x}}|) < U^a(|{\mathbf{x}}|)$). This means that the difference $U^b(|{\mathbf{x}}|) - U^a(|{\mathbf{x}}|)$ plays the role of a “symmetry breaking field". Also, note that the above proof is also valid in the presence of an external potential $U_e({\mathbf{x}})$ that is equal for both species, meaning that the polarization transition also occurs for trapped atoms.
We note that the bosonic nature of the particles is crucial for establishing our rigorous proof. If the atoms $a$ and $b$ were fermions, the variational argument ceases to be valid: since the Hamiltonian conserves species, atoms belonging to different species can be considered distinguishable, and the wave function does not change sign if we exchange any two atoms $a$ and $b$ following a path $P$. If we convert an atom $a$ to an atom $b$, the wave function has to change sign under the particle exchange following the same path $P$, implying that we have to introduce an additional node in the wave function. Formally this means that ${\cal{O}}_{R}|\Psi_0(N_a,N_b)\rangle=0$ for fermions.
We now discuss the low energy excitation spectrum, which in the symmetric case ($U^{a} = U^{b}$) can be computed using longitudinal sum rules[@huang1; @huang2]. In the unpolarized case the excitations correspond to two phonon branches with wave functions $
\rho ^{\pm}_{\bf k} =
(\rho ^{a}_{\bf k}\pm\rho ^{b}_{\bf k})|0\rangle$. The operators $\rho ^{i}_{\bf k}=V^{-1/2}\int d^3 x e^{i{\bf k}\cdot {\bf x}}\rho_{{\bf x},i}
$ are the Fourier transforms of the density operators for each species. In other words, since the Hamiltonian is invariant under exchange of species, the excitations are either even or odd in the species index.
The excitation energies are given by $$\omega^{\pm}_{\bf k}=
{\langle 0|\rho_{\bf -k}^{\pm}H\rho_{\bf k}^{\pm}|0\rangle
\over
\langle 0|\rho_{\bf -k}^{\pm}\rho_{\bf k}^{\pm}|0\rangle},$$ where we have shifted the origin of energies ($H \rightarrow H-E_0$; $H|0\rangle =0$). We will first show that $\langle 0|\rho_{\bf -k}^{\pm}H\rho_{\bf k}^{\pm}|0\rangle
=nk^2/2m,$ with $n=N/V$. Consider $$\begin{aligned}
\langle 0|\rho_{\bf -k}^{a}H\rho_{\bf k}^{b}|0\rangle&=&
{1\over 2}
\langle 0|\left\{[ \rho_{\bf -k}^{a},H]\rho_{\bf k}^{b} +
\rho_{\bf -k}^{a} [\rho_{\bf k}^{b},H]
\right\}|0\rangle \label{ab1} \\
&=& {\hbar^2\over 2m}\sum_{{\bf q},{\bf q}'}
{\bf k}\cdot{\bf q}\langle 0|\left[
\psi_a^{\dagger}({{\bf q}'-{\bf k}})
\psi_a^{}({{\bf q}'})
\psi_b^{\dagger}({{\bf q}+{\bf k}})
\psi_b^{}({{\bf q}}) \right. \nonumber \\
&& \left.
-\psi_a^{\dagger}({{\bf q}+{\bf k}})
\psi_a^{}({{\bf q}})
\psi_b^{\dagger}({{\bf q}'-{\bf k}})
\psi_b^{}({{\bf q}}')\right]|0\rangle,
\label{ab2}
\end{aligned}$$ with Eq. (\[ab2\]) following from the direct evaluation of the commutators in Eq. (\[ab1\]). Since the ground state is symmetric under species exchange, and since the operators $b$ and $a$ commute, Eq. (\[ab2\]) implies that $\langle 0|\rho_{\bf -k}^{a}H\rho_{\bf k}^{b}|0\rangle=
\langle 0|\rho_{\bf -k}^{b}H\rho_{\bf k}^{a}|0\rangle=0$. On the other hand, if in Eq. (\[ab2\]) we replace $b$ with $a$ we obtain the usual sum rule $
\langle 0|\rho_{\bf -k}^{a}H\rho_{\bf k}^{a}|0\rangle=(\hbar^2 k^2/2m) V^{-1}\langle 0|
\sum_{\bf q} \psi_a^{\dagger}({{\bf q}}) \psi_a^{}({{\bf q}})
|0\rangle$, which implies that for the even and odd modes the exact $f$–sum rule applies [@nozieres]: $\langle 0|\rho_{\bf -k}^{\pm}H\rho_{\bf k}^{\pm}|0\rangle
=nk^2/2m$, or, which is equivalent: $$n\int d \omega \, \omega S^{\pm}(k,\omega)={nk^2\over 2m},
\label{sr1}$$ with
$$S^{\pm}(k,\omega)=n^{-1}\sum_{\nu} |\langle \nu |\rho_{\bf k}^{\pm}|0\rangle |^2
\delta (\omega - \omega_\nu)$$
the dynamic structure factor of the even and odd modes. The excitation energies are therefore given by $$\omega ^{\pm}_{\bf k}= {k^2\over 2mS^{\pm}(k)},$$ with $S^{\pm}(k)=\int d\omega S^{\pm}(k,\omega)$ the corresponding static structure factor.
We can establish compressibility sum rules for $S^{\pm}(k,\omega)$ by considering the response of the system to an external perturbation $H'_{\pm}={\lambda \over 2}
\sum_{\bf k} (\rho_{\bf k}^{\pm}+\rho_{-\bf k}^{\pm})$ that couples to either the even or odd modes. We follow the analysis of Ref. [@huang2] and assume that in the long wave length limit the perturbed wave function is locally the same as the unperturbed one with a modulated density (even or odd for each mode). We obtain $$\lim_{k \rightarrow0}\int d\omega {S^{\pm}(k,\omega)\over \omega}={1\over 2m v_{\pm}^2},
\label{sr2}$$ with $$v^2_{+} = {n\over m} {\partial ^2 \epsilon_0(n,n_-)\over \partial n^2},\;\;\;\;
v^2_{-} = {n\over m} {\partial ^2 \epsilon_0(n,n_-)\over \partial n_-^2},$$ and $\epsilon_0(n,n_-)$ the ground state energy per unit volume written as a function of the total density $n=n_a+n_b$ and the density difference $n_-=n_a-n_b$. Since in the long wave–length limit the sum rules are exhausted by the above quasiparticles[@nozieres] \[$\lim_{k\rightarrow 0} S^{\pm}(k,\omega)
=S^{\pm}(k)\delta(\omega -
\omega^{\pm}_{\bf k})$\], the sum rules (\[sr1\]) and (\[sr2\]) imply that two branches have energies $ \omega^{\pm}_{\bf k}=v_{\pm}k$, with the corresponding structure factors given by $S^{\pm}(k)=k/2mv_{\pm}$. The low energy spectrum therefore consists of two linear modes, corresponding to modulations of the density with the two species “in–phase"(even mode) and “out–of phase"(odd mode) respectively. The odd mode has total density $n$ constant in all space, and in the spinor language (where the species index is treated as a spin $1/2$) corresponds to a spin density wave. We can see qualitatively that the odd modes have lower frequency by perturbing around low values of the inter–species interaction. For $U^{x}=0$ the even and odd modes are degenerate. If we turn on $U^{x}$ the odd mode will have lower expectation value $\langle U^{x} \rangle$ since locally $\langle \rho_a({\bf x})\rho_b({\bf x}) \rangle_+
>\langle \rho_a({\bf x})\rho_b({\bf x}) \rangle_-$. This is reasonable as long as the repulsive component of $U^{x}$ is dominant. If we now increase $U^{x}$ we will reach the instability discussed above: when $U^{x}=U^{a}$ the ground state is multiply degenerate with all the polarized wave functions $|\Psi_0[(N-M)/2,(N+M)/2]\rangle$ ($M=-N,\cdots,N$) having the same energy. This implies that $v_{-}\rightarrow 0$, since $ \epsilon_0(n,n_-)=\epsilon_0(n,0)$. Therefore the instability towards a polarized state is signaled by a vanishing velocity of the odd modes, or, equivalently, by a divergence of the “compressibility” $\kappa_- \propto
[\partial ^2 \epsilon_0(n,n_-)/ \partial n_-^2]^{-1}$ associated with changes in species polarization at constant particle number.
Finally, we note that our proof is not valid in the presence of a field coupling the two species of the form ${\cal R}\int d^3x\, (\psi_a^{\dagger}({\bf x}) \psi_b({\bf x}) +
{\rm h.c.})$. This term corresponds (in the spinor language) to a magnetic field tilted with respect to the direction of the otherwise fully polarized system. If the system is prepared in the unstable regime with fixed $N_a$ and $N_b$ and no mechanism of interconversion is allowed, our proof indicates that the system will separate into two phases. This kind of phase separation was discussed within the Bogoliubov model by Nepomnyaschii[@nepo]. An infinitesimal coupling with a heat bath that allows the system to equilibrate will equate the chemical potentials of both species and induce a transition to a state of one species-only. The precise kinetics and the time scale to reach equilibrium are beyond the scope of our paper.
When ${\cal R}\neq 0$ the solutions will not be completely polarized even in the case where $U^{x}>U^{a}$, but that the ground state will be “rotated" with respect to the quantization axis in species space. The detailed mean field analysis of this problem is the subject of a forthcoming paper[@search].
We thank Chris Search, Paul Berman and Paul Julienne for very valuable discussions.
[9]{}
M.H. Anderson [*et al.*]{} Science [**269**]{}, 198 (1995); K. B. Davis [*et al.*]{}, Phys. Rev. Lett. [**75**]{}, 3969 (1995); C. C. Bradley [*et al.*]{}, [*ibid.*]{} [**78**]{}, 985 (1997).
See for example F. Dalfovo, S. Giorgini, L. P. Pitaevskii, and S. Stringari, Rev. Mod. Phys. [**71**]{}, 463 (1999) and references therein.
C. J. Myatt [*et al.*]{}, Phys. Rev. Lett.[**78**]{}, 586 (1997).
D. M. Stamper-Kurn [*et al.*]{}, Phys. Rev. Lett.[**80**]{}, 2027 (1998).
B. I. Halperin, Phys. Rev. B [**11**]{}, 178 (1975).
I. M. Khalatinov, Zh. Eksp. Teor. Fiz. [**32**]{} 653 (1957) \[Sov. Phys. JETP [**5**]{}, 542 (1957)\].
E. D. Siggia and A. E. Ruckenstein, Phys. Rev. Lett.[**44**]{}, 1423 (1980).
T. L. Ho, Phys. Rev. Lett.[**81**]{}, 742 (1998); T. Ohmi and K. Machida, J. Phys. Soc. Jap. [**67**]{} 1822 (1998); M. Koashi and M. Ueda Phys. Rev. Lett.[**84**]{}, 1066 (2000); C. K Law, H. Pu, and N. Bigelow, Phys. Rev. Lett.[**81**]{}, 5257 (1998); T. L. Ho and S. K. Yip, Phys. Rev. Lett.[**84**]{}, 4031 (2000); M. Ueda, preprint cond-mat/0008189.
For a detailed mean field treatment of the experiments of Ref. [@myatt], see B. D. Esry [*et al.*]{}, Phys. Rev. Lett. [**78**]{}, 3594 (1997).
E. V. Goldstein and P. Meystre, Phys. Rev. A [**55**]{}, 2935 (1997).
E. M. Lifshitz and L. P. Pitaevskii, [*Statistical Physics, Part 2*]{} (Pergamon, New York, 1980).
See P. S. Julienne, “Ultra-Cold Collisions of Atoms and Molecules", to appear as Chapter 2.6.3 om [*Scattering*]{}, edited by P. Sabatier and E. R. Pike, Academic Press; and J. Weiner [*et al.*]{}, Rev. Mod. Phys. [**71**]{}, 1 (1999).
K. Huang and A. Klein, Ann. Phys. (N.Y.) [**30**]{}, 203 (1964).
K. Huang, [*Statistical Mechanics*]{} (John Wiley & Sons, New York, 1987), pp. 325-328 and Appendix A.4.
P. Nozieres and D. Pines, [*The Qheory of Quantum Liquids, Vol. 2*]{} (Addison-Wesley, New York, 1990), p.26.
Y. A. Nepomnyaschii, Theor. Mat. Phys. [**20**]{}, 904 (1974) \[Teor. Mat. Fiz. [**20**]{}, 399 (1974)\].
C. P. Search, A. G. Rojo and P. R. Berman, Phys. Rev. A (in press), and preprint cond-mat/0012010.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The recent data for Bose-Einstein Correlations (BEC) of three-charged particles obtained by NA44 Collaboration have been analysed using theoretical formula with Coulomb wave functions. It has been recently proposed by Alt et al. It turns out that there are discrepancies between these data and the respective theoretical values. To resolve this problem we seek a possibly modified theoretical formulation of this problem by introducing the degree of coherence for the exchange effect due to the BEC between two-identical bosons. As a result we obtain a modified formulation for the BEC of three-charged particles showing good agreement with the data. Moreover, we investigate physical connection between our modified formulation and the core-halo model proposed by Csörgő et al. Our study indicates that the interaction region estimated by the BEC of three-charged particles in the S + Pb collisions at 200 GeV/c per nucleon is equal to about 1.5 fm$\sim$1.8 fm.'
address:
- 'Toba National College of Maritime Technology, Toba 517-8501, Japan'
- 'Department of Physics, Faculty of Science, Shinshu University, Matsumoto 390-8621, Japan'
author:
- 'T. Mizoguchi,'
- 'M. Biyajima'
title: 'Possible formulations for three-charged particles correlations in terms of Coulomb wave functions'
---
Bose-Einstein Correlation, three-charged particles, Coulomb wave functions, high energy heavy-ion collisions
Introduction
============
One of the most interesting subjects in high energy heavy-ion collisions is study of the higher order Bose-Einstein Correlation (BEC) effect [@biya90; @cram91; @hein97; @heis97; @alt99; @huma99] (known also as the HBT or the GGLP effect, or as the hadron interferometry [@hanb56; @gold59; @gyul79; @boal90]). From data on BEC we can (in principle) infer the size of the interaction region and therefore estimate the energy densities reached in high energy collisions. Such work is a necessary task in the search for the quark-gluon plasma [@bert89; @weid99] - a new, hypothetical form of matter.
To get more precise sizes of the interaction regions, we have to take into account the final state interactions among the charged particles [@prat86; @biya94]. A great advance in this direction for the BEC of the three-charged particles has been recently made by Alt et. al. [@alt99]. They have derived a correction formula for the raw data introducing distribution functions of the charged particles. Their formulation is based on the plane wave functions and on the Coulomb wave functions, assuming that produced hadrons are already in the asymptotic region of the Coulomb interactions where the strong interaction already vanishes [@merk78; @brau89]. It amounts in the following correction factor $K_{Coul}$ due to the Coulomb effect for identical three-charged particles: [^1] $$K_{Coul} = \frac{N_{Coul}}{D_{plane}}\ .
\label{eq1}$$ The denominator $D_{plane}$ is given by ($\rho ({{\rm {\bf x}}}_i)$ are distribution functions of charged particles): $$\begin{aligned}
D_{plane} & \cong & \frac 16 \int
d^3 {{\rm {\bf x}}}_1\rho ({{\rm {\bf x}}}_1)
d^3 {{\rm {\bf x}}}_2\rho ({{\rm {\bf x}}}_2)
d^3 {{\rm {\bf x}}}_3\rho ({{\rm {\bf x}}}_3)\nonumber\\
&&\quad \cdot \left|
e^{i({{\rm {\bf k}}}_1\cdot {{\rm {\bf x}}}_1+{{\rm {\bf k}}}_2\cdot {{\rm {\bf x}}}_2+{{\rm {\bf k}}}_3\cdot
{{\rm {\bf x}}}_3)}
+ e^{i({{\rm {\bf k}}}_1\cdot {{\rm {\bf x}}}_2+{{\rm {\bf k}}}_2\cdot {{\rm {\bf x}}}_1+{{\rm {\bf k}}}_3\cdot
{{\rm {\bf x}}}_3)}\right .\nonumber\\
&&\quad +
e^{i({{\rm {\bf k}}}_1\cdot {{\rm {\bf x}}}_2+{{\rm {\bf k}}}_2\cdot {{\rm {\bf x}}}_3+{{\rm {\bf k}}}_3\cdot
{{\rm {\bf x}}}_1)}
+ e^{i({{\rm {\bf k}}}_1\cdot {{\rm {\bf x}}}_1+{{\rm {\bf k}}}_2\cdot {{\rm {\bf x}}}_3+{{\rm {\bf k}}}_3\cdot
{{\rm {\bf x}}}_2)}\nonumber\\
&&\quad \left . + e^{i({{\rm {\bf k}}}_1\cdot {{\rm {\bf x}}}_3+{{\rm {\bf k}}}_2\cdot {{\rm {\bf x}}}_1 +
{{\rm {\bf k}}}_3\cdot {{\rm {\bf x}}}_2)}
+ e^{i({{\rm {\bf k}}}_1\cdot {{\rm {\bf x}}}_3+{{\rm {\bf k}}}_2\cdot {{\rm {\bf x}}}_2+{{\rm {\bf k}}}_3\cdot
{{\rm {\bf x}}}_1)}\right|^2\ ,
\label{eq2}\end{aligned}$$ The numerator $N_{Coul}$ has the following form: $$\begin{aligned}
N_{Coul} & \cong & \frac 16 \int
d^3 {{\rm {\bf x}}}_1\rho ({{\rm {\bf x}}}_1)
d^3 {{\rm {\bf x}}}_2\rho ({{\rm {\bf x}}}_2)
d^3 {{\rm {\bf x}}}_3\rho ({{\rm {\bf x}}}_3)\nonumber\\
&&\quad \cdot \left|
\psi_{{{\rm {\bf k}}}_1{{\rm {\bf k}}}_2}^C({{\rm {\bf x}}}_1,\ {{\rm {\bf x}}}_2)
\psi_{{{\rm {\bf k}}}_2{{\rm {\bf k}}}_3}^C({{\rm {\bf x}}}_2,\ {{\rm {\bf x}}}_3)
\psi_{{{\rm {\bf k}}}_3{{\rm {\bf k}}}_1}^C({{\rm {\bf x}}}_3,\ {{\rm {\bf x}}}_1)\right .\nonumber\\
&&\quad +
\psi_{{{\rm {\bf k}}}_1{{\rm {\bf k}}}_2}^C({{\rm {\bf x}}}_1,\ {{\rm {\bf x}}}_3)
\psi_{{{\rm {\bf k}}}_2{{\rm {\bf k}}}_3}^C({{\rm {\bf x}}}_3,\ {{\rm {\bf x}}}_2)
\psi_{{{\rm {\bf k}}}_3{{\rm {\bf k}}}_1}^C({{\rm {\bf x}}}_2,\ {{\rm {\bf x}}}_1)\nonumber\\
&&\quad +
\psi_{{{\rm {\bf k}}}_1{{\rm {\bf k}}}_2}^C({{\rm {\bf x}}}_2,\ {{\rm {\bf x}}}_1)
\psi_{{{\rm {\bf k}}}_2{{\rm {\bf k}}}_3}^C({{\rm {\bf x}}}_1,\ {{\rm {\bf x}}}_3)
\psi_{{{\rm {\bf k}}}_3{{\rm {\bf k}}}_1}^C({{\rm {\bf x}}}_3,\ {{\rm {\bf x}}}_2)\nonumber\\
&&\quad +
\psi_{{{\rm {\bf k}}}_1{{\rm {\bf k}}}_2}^C({{\rm {\bf x}}}_2,\ {{\rm {\bf x}}}_3)
\psi_{{{\rm {\bf k}}}_2{{\rm {\bf k}}}_3}^C({{\rm {\bf x}}}_3,\ {{\rm {\bf x}}}_1)
\psi_{{{\rm {\bf k}}}_3{{\rm {\bf k}}}_1}^C({{\rm {\bf x}}}_1,\ {{\rm {\bf x}}}_2)\nonumber\\
&&\quad +
\psi_{{{\rm {\bf k}}}_1{{\rm {\bf k}}}_2}^C({{\rm {\bf x}}}_3,\ {{\rm {\bf x}}}_1)
\psi_{{{\rm {\bf k}}}_2{{\rm {\bf k}}}_3}^C({{\rm {\bf x}}}_1,\ {{\rm {\bf x}}}_2)
\psi_{{{\rm {\bf k}}}_3{{\rm {\bf k}}}_1}^C({{\rm {\bf x}}}_2,\ {{\rm {\bf x}}}_3)\nonumber\\
&&\quad \left . +
\psi_{{{\rm {\bf k}}}_1{{\rm {\bf k}}}_2}^C({{\rm {\bf x}}}_3,\ {{\rm {\bf x}}}_2)
\psi_{{{\rm {\bf k}}}_2{{\rm {\bf k}}}_3}^C({{\rm {\bf x}}}_2,\ {{\rm {\bf x}}}_1)
\psi_{{{\rm {\bf k}}}_3{{\rm {\bf k}}}_1}^C({{\rm {\bf x}}}_1,\ {{\rm {\bf x}}}_3)\right|^2\ .
\label{eq3}\end{aligned}$$ Here $\psi_{{{\rm {\bf k}}}_i{{\rm {\bf k}}}_j}^C({{\rm {\bf x}}}_i,\ {{\rm {\bf x}}}_j)$ are the Coulomb wave functions of the respective 2-body collision expressed as, $$\psi_{{{\rm {\bf k}}}_i {{\rm {\bf k}}}_j}^C({{\rm {\bf x}}}_i,\ {{\rm {\bf x}}}_j) = \Gamma(1 + i\eta_{ij})
e^{-\pi \eta_{ij}/2} e^{ i {{\rm {\bf k}}}_{ij} \cdot {{\rm {\bf r}}}_{ij} }
F[- i \eta_{ij},\ 1;\ i ( k_{ij} r_{ij} - {{\rm {\bf k}}}_{ij} \cdot {{\rm {\bf r}}}_{ij} )],
\label{eq4}$$ with ${{\rm {\bf r}}}_{ij} = ({{\rm {\bf x}}}_i - {{\rm {\bf x}}}_j)$, ${{\rm {\bf k}}}_{ij} = ({{\rm {\bf k}}}_i -
{{\rm {\bf k}}}_j)/2$, $r_{ij} = |{{\rm {\bf r}}}_{ij}|$, $k_{ij} = |{{\rm {\bf k}}}_{ij}|$ and $\eta_{ij} = m\alpha/k_{ij}$. $F[a,\ b;\ x]$ and $\Gamma(x)$ are the confluent hypergeometric function and the Gamma function, respectively. In order to use Eqs. (\[eq1\]), (\[eq2\]) and (\[eq3\]), one has to assume first some shapes and sizes for the source functions. In fact, this is the procedure already used in Ref. [@na4499] by NA44 Collaboration: $${\rm Corrected\ data} = ({\rm raw\ data}) \times K_{\rm spc} \times
K_{\rm acceptance} \times K_{Coul}\ ,$$ where $K_{spc}$ and $K_{acceptance}$ denote the effect of multiparticle production in the single particle spectra and the acceptance effect in the experiment.
In this paper, we would like to adopt a different point of view for Eq. (\[eq3\]). As is seen in Ref. [@biya94], the BEC of identical two-charged pions can also be analysed by the Coulomb wave functions. It is therefore reasonable to expect that the numerator $N_{Coul}$ is the main theoretical ingredient in analysis of the BEC of three-charged particles. We argue therefore that $$N^{(3+\ {\rm or }\ 3-)}/N^{BG} \equiv \mbox{$C$}\times N_{Coul}\ ,
\label{eq5}$$ where we have introduced the normalization factor $C$, which corresponds to the asymptotic value of the BEC. Using Eq. (\[eq5\]) we can now (with the help of the CERN-MINUIT program) analyse data of Ref. [@na4499] using Gaussian source distributions of radii $R$, $\rho({{\rm {\bf x}}})=
\frac{1}{(2\pi R^2)^{3/2}} \exp\left[-\frac{{{{\rm {\bf x}}}}^2}{2R^2}\right]$ [^2]. In the next paragraph, we analyse the data of NA44 Collaboration [@na4499] by Eq. (\[eq5\]). In the third paragraph we shall derive a modefied theoretical formula for 3-particle BEC introducing the degree of coherence parameter into Eq. (\[eq5\]). This formula will be then used in the 4th paragraph for the re-analyses of the experimental data [@na4499]. In the 5th paragraph, we investigate whether or not there is physical connection between our study and the core-halo model [@csorgo99]. Concluding remarks are given in the final paragraph.
Application of Eq. (\[eq5\]) to the data by NA44 Collaboration
==============================================================
Here we analyse the data by Eq. (\[eq5\]). As can be seen in Fig. \[fig1\] and Table \[table1\], there are some discrepancies between the data points and theoretical values calculated by means of Eq. (\[eq5\]). The minimum $\chi^2$ value, 17.6 in Table \[table1\], seems to be big, as the number of the data points are considered. Thus we would like to know why this equation cannot explain the data [@na4499]. We know that there are several possible reasons due to effects of the partial coherence, the contamination [@cramer96] and the long-lived resonances [@csorgo99]. At present, we consider the effect of the possible partial coherent of produced pions. In fact, authors of Ref. [@na4499] have used not the equivalence of Eq. (\[eq5\]) but the following formula instead (cf., Ref. [@gyul79; @deut78]): $$\frac{N^{(3+)}}{N^{BG}} = C\left(1+\lambda_3 e^{-R_3^2Q_3^2}\right)\ .
\label{eq6}$$ It contains one more parameter, $\lambda_3$, which can be regarded as a kind of effective degree of coherence and which, in our opinion, should therefore occur also somehow in Eq. (\[eq5\]).
Formulas $C$ $R$ \[fm\] $\lambda_3$ $\chi^2/N_{dof}$
--------------- ----------------- --------------- --------------- ------------------
Eq. (\[eq5\]) 0.941$\pm$0.026 2.47$\pm$0.14 — 17.6/16
Eq. (\[eq6\]) 0.986$\pm$0.028 2.36$\pm$0.26 1.37$\pm$0.19 7.8/15
: Estimated values for the data [@na4499] by Eqs. (\[eq5\]) and (\[eq6\]) using CERN-MINUIT program.[]{data-label="table1"}
Diagram Decomposition of Eq. (\[eq5\])
======================================
First of all, we have to find a possible way for the introduction of the degree of coherence parameter $\lambda$ into Eq. (\[eq5\]). Let us therefore examine the plane wave (PW) approximations of the Coulomb wave functions:
\[eq7\] A(1) &=& \_[k\_1k\_2]{}\^C(x\_1, x\_2) \_[k\_2k\_3]{}\^C(x\_2, x\_3) \_[k\_3k\_1]{}\^C(x\_3, x\_1)\
&& e\^[ i k\_[12]{} r\_[12]{}]{} e\^[ i k\_[23]{} r\_[23]{}]{} e\^[ i k\_[31]{} r\_[31]{}]{} = e\^[ (3/2)i (k\_1 x\_1 + k\_2 x\_2 + k\_3 x\_3)]{} ,\
A(2) &=& \_[k\_1k\_2]{}\^C(x\_1, x\_3) \_[k\_2k\_3]{}\^C(x\_3, x\_2) \_[k\_3k\_1]{}\^C(x\_2, x\_1)\
&& e\^[ i k\_[12]{} r\_[13]{}]{} e\^[ i k\_[23]{} r\_[32]{}]{} e\^[ i k\_[31]{} r\_[21]{}]{} = e\^[ (3/2)i (k\_1 x\_1 + k\_2 x\_3 + k\_3 x\_2)]{} ,\
A(3) &=& \_[k\_1k\_2]{}\^C(x\_2, x\_1) \_[k\_2k\_3]{}\^C(x\_1, x\_3) \_[k\_3k\_1]{}\^C(x\_3, x\_2)\
&& e\^[ i k\_[12]{} r\_[21]{}]{} e\^[ i k\_[23]{} r\_[13]{}]{} e\^[ i k\_[31]{} r\_[32]{}]{} = e\^[ (3/2)i (k\_1 x\_2 + k\_2 x\_1 + k\_3 x\_3)]{} ,\
A(4) &=& \_[k\_1k\_2]{}\^C(x\_2, x\_3) \_[k\_2k\_3]{}\^C(x\_3, x\_1) \_[k\_3k\_1]{}\^C(x\_1, x\_2)\
&& e\^[ i k\_[12]{} r\_[23]{}]{} e\^[ i k\_[23]{} r\_[31]{}]{} e\^[ i k\_[31]{} r\_[12]{}]{} = e\^[ (3/2)i (k\_1 x\_2 + k\_2 x\_3 + k\_3 x\_1)]{} ,\
A(5) &=& \_[k\_1k\_2]{}\^C(x\_3, x\_1) \_[k\_2k\_3]{}\^C(x\_1, x\_2) \_[k\_3k\_1]{}\^C(x\_2, x\_3)\
&& e\^[ i k\_[12]{} r\_[31]{}]{} e\^[ i k\_[23]{} r\_[12]{}]{} e\^[ i k\_[31]{} r\_[23]{}]{} = e\^[ (3/2)i (k\_1 x\_3 + k\_2 x\_1 + k\_3 x\_2)]{} ,\
A(6) &=& \_[k\_1k\_2]{}\^C(x\_3, x\_2) \_[k\_2k\_3]{}\^C(x\_2, x\_1) \_[k\_3k\_1]{}\^C(x\_1, x\_3)\
&& e\^[ i k\_[12]{} r\_[32]{}]{} e\^[ i k\_[23]{} r\_[21]{}]{} e\^[ i k\_[31]{} r\_[13]{}]{} = e\^[ (3/2)i (k\_1 x\_3 + k\_2 x\_2 + k\_3 x\_1)]{} .
Notice that, except for the factor $3/2$, exponential functions are the same expressions as those present in the integrand of Eq. (\[eq2\]). This difference is attributed to the fact that Coulomb wave function used here describes two-charged particles collisions, therefore factor $3/2$ appears because there are relevant two-particle three combinations among three-charged particles.
Combining Eqs. (\[eq7\]) and Figs. \[fig2\], we obtain the following three sets of equations:
\[eq8\] F\_1 &=& 16 \_[i=1]{}\^6 A(i)A\^\*(i) 1 ,\
F\_2 &=& 16 \[ A(1)A\^\*(2) + A(1)A\^\*(3) + A(1)A\^\*(6) + A(2)A\^\*(4) + A(2)A\^\*(5)\
&& + A(3)A\^\*(4) + A(3)A\^\*(5) + A(4)A\^\*(6) + A(5)A\^\*(6) + [ c. c.]{}\]\
&&\
&& ,\
F\_3 &=& 16 \[ A(1)A\^\*(4) + A(1)A\^\*(5) + A(2)A\^\*(3) + A(2)A\^\*(6) + A(3)A\^\*(6)\
&& + A(4)A\^\*(5) + [c. c.]{}\]\
&&\
&& .
Combining now Eqs. (\[eq8\]) and the concept of partial coherence for the BEC [@na4499; @deut78], we can introduce a coherence parameter $\sqrt{\lambda}$ for the single mark ([$\times$]{}) in Fig. \[fig2\]. The $\lambda = 1$ corresponds to the totally chaotic source, which is the assumption behind Eq. (\[eq5\]). Taking into account the strength of the degree of coherence $\lambda$ between two-identical bosons and $\lambda^{3/2}$ among three-identical bosons in Figs. \[fig2\], we can finally express the BEC for three identical charged particles as: $$\frac{N^{(3+\ {\rm or }\ 3-)}}{N^{BG}} \cong C \int
d^3 {{\rm {\bf x}}}_1\rho ({{\rm {\bf x}}}_1)
d^3 {{\rm {\bf x}}}_2\rho ({{\rm {\bf x}}}_2)
d^3 {{\rm {\bf x}}}_3\rho ({{\rm {\bf x}}}_3)
[F_1 + \lambda F_2 + \lambda^{3/2} F_3]\ .
\label{eq9}$$ Equation (\[eq9\]) is the modified theoretical formula we were looking for. It differs from Eq. (\[eq5\]) originally proposed by Alt et. al. in [@alt99] by the presence of the degree of coherence $\lambda$ and in the limit of $\eta_{ij} \to 0$ it becomes $${\rm Eq.\ (\ref{eq9})} \stackrel{\eta_{ij} \to 0}{\longrightarrow}
C\left(1 + 3\lambda e^{-\frac 34 R^2Q_3^2} + 2\lambda^{3/2}
e^{-\frac 98 R^2Q_3^2}\right)\ ,
\label{eq10}$$ which is the extended formula proposed some time ago by Deutschmann et al. [@deut78].
It should be noticed that Eq. (\[eq9\]) can be applied to data corrected only by the Gamow factor $G(\eta_{12})G(\eta_{23})G(\eta_{31})$ in an ideal case [@biya96], because Eq. (\[eq9\]) is described by the Coulomb wave functions including the Gamow factors (see Ref. [@liu86]) [^3].
Reanalyses of NA44 Collaboration data by means of Eq. (\[eq9\])
===============================================================
At present we have no data corrected only by the Gamow factors, therefore we apply Eq. (\[eq9\]) to the analysis of NA44 Collaboration data [@na4499] using the CERN-MINUIT program. Our results are shown in Fig. \[fig3\] and Table \[table2\]. Comparing them with those of Table \[table1\], it can be said that the $\chi^2$-value becomes smaller, i.e., the agreement is now better. The range of interaction becomes also smaller. For the sake of reference we present in Table \[table2\] also results obtained by using Eq. (\[eq10\]) .
Formula $C$ $R$ \[fm\] $\lambda$ $\chi^2/N_{dof}$
---------------- ----------------- --------------- --------------- ------------------
Eq. (\[eq9\]) 0.917$\pm$0.032 1.53$\pm$0.20 0.55$\pm$0.07 6.7/15
Eq. (\[eq10\]) 0.984$\pm$0.029 2.60$\pm$0.28 0.33$\pm$0.04 7.7/15
: Reanalyses of 3$\pi^+$ BEC in S + Pb collision [@na4499] by Eqs. (\[eq9\]) and (\[eq10\]).[]{data-label="table2"}
From results of Table \[table2\], we see that $R = 1.53$ fm by Eq. (\[eq9\]) is small. The reason is attributed to the fraction of partially coherent effect ($\lambda$). From comparisons between results by Eqs. (\[eq9\]) and (\[eq10\]), it can be seen the interaction region becomes smaller, and the degree of coherence does conversely bigger due to the Coulomb interaction.
Possible interpretation of $\lambda$ by core-halo model
=======================================================
The core-halo model has been proposed by Csörgő et al. [@csorgo99]. We study whether or not there is physical connection between our previous formulation and theirs [@csorgo99]. To apply their model to Eq. (\[eq8\]), first of all we use the following source functions $$\rho (x_i) = \rho_c (x_i) + \rho_{halo} (x_i)\ ,
\label{eq11}$$ where $\rho_c (x_i) = \frac{1}{(2\pi R^2)^{3/2}}
\exp \left[-\frac{{{{\rm {\bf x}}}_i}^2}{2R^2}\right]$ and $\rho_{halo} (x_i) =
\frac{1}{(2\pi R_h^2)^{3/2}} \exp\left[-\frac{{{{\rm {\bf x}}}_i}^2}{2R_h^2}\right]$.
In concrete analyses, the radius of the halo becomes to be infinite, due to the effect of long-lived resonances ($R_h \to \infty$). The exchange function of the halo part due to the BEC, $E_{halo} = \exp(-Q^2R_h^2) \to 0$. Then we can interpret $\lambda$ in Eq. (\[eq9\]) by their parameter, i. e., fraction of multiplicity from the core part, $f_c = \langle n_{core}\rangle/\langle
n_{tot}\rangle$, as $\lambda = f_c^2$. In this case, it should be stressed that their parameter $p_c = \langle n_{co}\rangle/\langle n_{core}\rangle$, the fraction of the coherently produced multiplicity from the core part is zero.
Moreover, we can assume the laser optical approach for the core part [@csorgo99], introducing a parameter $p = \langle n_{chao}\rangle/\langle
n_{core}\rangle$ [@biya90]. Notice that $p = 1 - p_c$ [@csorgo99]. For the cross mark ([$\times$]{}) in Fig. \[fig2\], we assume two components of coherently ($1 - p$) and chaotically produced particles ($p$). Then we obtain the following expression $$\begin{aligned}
\frac{N^{(3+\ {\rm or }\ 3-)}}{N^{BG}} &\cong& C \left[\ \int
d^3 {{\rm {\bf x}}}_1\rho_c ({{\rm {\bf x}}}_1)
d^3 {{\rm {\bf x}}}_2\rho_c ({{\rm {\bf x}}}_2)
d^3 {{\rm {\bf x}}}_3\rho_c ({{\rm {\bf x}}}_3)
(F_1 + f_c^2p^2 F_2 + f_c^3p^3 F_3)\right .\nonumber\\
&&\qquad + \int
d^3 {{\rm {\bf x}}}_1\rho_c ({{\rm {\bf x}}}_1) \cdot \delta^3 ({{\rm {\bf x}}}_1)
d^3 {{\rm {\bf x}}}_2\rho_c ({{\rm {\bf x}}}_2)
d^3 {{\rm {\bf x}}}_3\rho_c ({{\rm {\bf x}}}_3) \nonumber\\
&&\qquad\qquad \left . \cdot \left ( 2f_c^2p(1-p) F_2 + 3f_c^3p^2(1-p) F_3
\right )\ \right ] \ .
\label{eq12}\end{aligned}$$ The effective degree of coherence, $\lambda_3^*$, the intercept at smallest $Q_3$ is given as $$\lambda_3^* = f_c^2(p^2+2p(1-p)) + f_c^3(p^3+3p^2(1-p))\ .
\label{eq13}$$ In Eq. (\[eq12\]), as $p = 1$, we obtain Eq. (\[eq9\]) with $\lambda =
f_c^2$. On the contrary, as $f_c = 1$, we obtain an expression of laser optical approach [@biya90]. By making use of Eq. (\[eq12\]) and an expression for the two-charged particles of the BEC, we can analyse the data by NA44 Collaboration.
For the two-charged particles of the BEC by means of the core-halo model with laser optical approach, we have the following equation, $$\begin{aligned}
\frac{N^{(2+\ {\rm or }\ 2-)}}{N^{BG}} &\cong& C \left[\int
d^3 {{\rm {\bf x}}}_1\rho ({{\rm {\bf x}}}_1)
d^3 {{\rm {\bf x}}}_2\rho ({{\rm {\bf x}}}_2) (G_1 + f_c^2p^2 G_2)\right .\nonumber\\
&&\left . + \int
d^3 {{\rm {\bf x}}}_1\rho ({{\rm {\bf x}}}_1)
d^3 {{\rm {\bf x}}}_2\rho ({{\rm {\bf x}}}_2) \cdot \delta^3 ({{\rm {\bf x}}}_2) 2f_c^2p(1-p)G_2
\right ]\ ,
\label{eq14}\end{aligned}$$ where $G_1 = \frac 12 \left (\left |\psi_{{{\rm {\bf k}}}_1 {{\rm {\bf k}}}_2}^C({{\rm {\bf x}}}_1,\
{{\rm {\bf x}}}_2)\right |^2 + \left |\psi_{{{\rm {\bf k}}}_1 {{\rm {\bf k}}}_2}^C({{\rm {\bf x}}}_2,\
{{\rm {\bf x}}}_1)\right |^2\right )$ and $G_2 = {\rm Re}\left (\psi_{{{\rm {\bf k}}}_1
{{\rm {\bf k}}}_2}^{C*}({{\rm {\bf x}}}_1,\ {{\rm {\bf x}}}_2)\right .$ $\left .\cdot \psi_{{{\rm {\bf k}}}_1
{{\rm {\bf k}}}_2}^C({{\rm {\bf x}}}_2,\ {{\rm {\bf x}}}_1)\right )$. The effective degree of coherence $\lambda_2^*$ is given as $$\lambda_2^* = f_c^2(p^2+2p(1-p))\ .
\label{eq15}$$ Results of our analyses are shown in Table \[table3\] and Figure \[fig5\]. As is seen in them, the interaction ranges of $R$(core part) are estimated in the ranges of $1.5\ {\rm fm} < R{\rm (core)} < 1.8\ {\rm fm}$, and $4.7\
{\rm fm} < R{\rm (core)} < 5.4\ {\rm fm}$, for the BEC of 3$\pi$ and $3\pi\to
2\pi$, respectively.
[cccc]{}\
$p$ & 1.0 & 0.8 & 0.6\
$f_c$ & $0.743\pm 0.050$ & $0.779\pm 0.055$ & $0.856\pm 0.061$\
$\lambda_3^*$ & 0.964 & 1.007 & 1.022\
$R$ (fm) & $1.53\pm 0.21$ & $1.72\pm 0.25$ & $1.87\pm 0.27$\
$\chi^2/N_{dof}$ & 6.7/15 & 6.6/15 & 6.6/15\
\
$p$ & 1.0 & 0.8 & 0.6\
$f_c$ & $0.633\pm 0.028$ & $0.647\pm 0.029$ & $0.688\pm 0.032$\
$\lambda_2^*$ & 0.400 & 0.402 & 0.398\
$R$ (fm) & $4.69\pm 0.45$ & $5.34\pm 0.54$ & $5.85\pm 0.59$\
$\chi^2/N_{dof}$ & 14.6/17 & 14.9/17 & 14.9/17\
The common region between results of the BEC of three-pion and two-pion is roughly described by $0.6 {\raisebox{-0.55ex}{$\stackrel{\displaystyle <}{\sim}$}}p {\raisebox{-0.55ex}{$\stackrel{\displaystyle <}{\sim}$}}1.0$ and $f_c \sim 0.7$.
Concluding remarks
==================
We have derived the theoretical formula for the BEC of three-charged identical particle using both the Coulomb wave functions and the notion of the degree of coherence and compared it with the experimental data [^4]. Historically the degree of coherence in the BEC of the two-identical bosons has been introduced by experimentalist [@deut78], and theoretical works in this direction have been performed in Ref. [@gyul79].
Our first analyses suggest that the degree of coherence $\lambda$ is a necessary ingredient also for the BEC of three-charged particles, in the same way as it was for the BEC for two-charged identical particles. This fact means that the source producing finally observed particles is not purely chaotic. However the investigation in 5th paragraph elucidates that our modified formulation can be interpreted by the core-halo model. Our degree of coherence $\lambda$ is equal to $f_c^2$ \[introduced in Ref. [@csorgo99]\], provided that particles are chaotically produced; $\lambda = f_c^2 (= (\langle n_{core}
\rangle/\langle n_{tot}\rangle)^2)$. Moreover, if we can assume the laser-optical approach for the cross mark ([$\times$]{}) in Fig. \[fig2\], we obtain Eqs. (\[eq12\]) $\sim$ (\[eq15\]). By making use of them, we obtain Table \[table3\] and Fig. \[fig5\]. There is common region among results from analyses of the BEC of $3\pi$ and $3\pi\to 2\pi$ processes, $0.6 {\raisebox{-0.55ex}{$\stackrel{\displaystyle <}{\sim}$}}p {\raisebox{-0.55ex}{$\stackrel{\displaystyle <}{\sim}$}}1.0$ and $f_c \sim 0.7$. However, the magnitude of the interaction regions estimated by the BEC of $3\pi$ and $3\pi\to 2\pi$ process are different. This problem should be considered in the future.
It should be noticed that also recent data on the BEC of $3\pi^-$ reported by OPAL Collaboration [@opal98] suggest the necessity of introduction of some degree of coherence or $f_c$ in Ref. [@csorgo99] [^5].
Acknowledgements {#acknowledgements .unnumbered}
================
Authors would like to thank Y. Nambu, S. Oryu and E. O. Alt for their kind suggestions and useful information. They are also indebted to G. Wilk for reading the manuscript. Our numerical calculations were partially carried out at RCNP of Osaka University. One of author (M. B.) is partially indebted to Japanese Grant-in-Aid for Education, Science, Sports and Culture (No. 09440103).
[18]{}
M. Biyajima, A. Bartl, T. Mizoguchi, O. Terazawa and N. Suzuki, Prog. Theor. Phys. 84 (1990) 931; Addendum-[*ibid*]{}. 88 (1992) 157.
J. G. Cramer, Phys. Rev. C43 (1991) 2798.
U. Heinz and Q. H. Zhang, Phys. Rev. C56 (1997) 426.
H. Heiselberg, A. P. Vischer, [*The Phase in Three-Pion Correlations*]{}, NBI-96-32, Jul 1997 (nucl-th/9707036).
E. O. Alt, T. Csörgő, B. Lörstad and J. Schmidt-Sorensen, Phys. Lett. B458 (1999) 407; See also, E. O. Alt, T. Csörgő, B. Lörstad and J. Schmidt-Sorensen, Eur. Phys. J. C13 (2000) 663.
T. J. Humanic, Phys. Rev. C60 (1999) 014901; See also, H. Nakamura and R. Seki, Phys. Rev. C60 (1999) 064904.
R. Hanbury-Brown and R. Q. Twiss, Nature 178 (1956) 1046.
G. Goldhaber et al., Phys. Rev. Lett. 3 (1959) 181.
M. Gyulassy, S. K. Kauffmann and L. W. Wilson, Phys. Rev. C20 (1979) 2267; See also, M. Biyajima, Phys. Lett. B 92 (1980) 193.
D. H. Boal, C. K. Gelbke and B. K. Jennings, Rev. Mod. Phys. 62 (1990) 553.
G. F. Bertsch, Nucl. Phys. A 498 (1989) 173c.
U. A. Wiedemann and U. Heinz, Phys. Rept. 319 (1999) 145; See also, R. M. Weiner, Phys. Rept. 327 (2000) 249.
S. Pratt, Phys. Rev. D 33 (1986) 72; See also, S. Pratt, T. Csörgő and J. Zimanyi, Phys. Rev. C 42 (1990) 2646.
M. Biyajima and T. Mizoguchi, preprint SULDP-1994-9 (Dec. 1994); See also, M. Biyajima , T. Mizoguchi, T. Osada and G. Wilk, Phys. Lett. B 353 (1995) 340.
S. P. Merkurev, Theor. Math. Phys. 32 (1978) 680.
M. Brauner, J. S. Briggs and H. J. Klar, J. Phys. B 22 (1989) 2265.
H. Bøggild et al., NA44 Collaboration, Phys. Lett. B 455 (1999) 77.
T. Csörgő, B. Lörstad, J. Schmidt-Sorensen and A. Ster Eur. Phys. J. C9 (1999) 275.
J. G. Cramer, K. Kadija, Phys. Rev. C 53 (1996) 908.
M. Deutschmann et al., Nucl. Phys. B 204 (1982) 333; cf. also: T.Peitzmann, Z. Phys. C 55 (1992) 485 and C 59 (1993) 127.
M. Biyajima, T. Mizoguchi, T. Osada and G. Wilk, Phys. Lett. B 366 (1996) 394; See also, “Final State Interactions in Bose-Einstein Correlations”, in Proc. of the 7th International Workshop on Multiparticle Production: Correlations and Fluctuations (Nijmegen, Netherlands, 1996), eds. R. C. Hwa, W. Kittel, W. J. Metzger and D. J. Schotanus (World Scientific, Singapore, 1997), p. 68.
Y.M. Liu et al., Phys. Rev. C 34 (1986) 1667.
K. Ackerstaff et al., OPAL Collaboration, Eur. Phys. J. C 5 (1998) 239.
M. Biyajima and T. Mizoguchi, in preparation.
Addendum {#addendum .unnumbered}
========
Because we had no sufficient information on the raw data and corrected data in 2001, we utilized the corrected data with $ K_{\rm Coulomb} $ in our analyses. Of course we mentioned that our Eqs. (9), (12), and (14) are available for the corrected data including $K_{\rm SPC} $ (single particle correction (SPC)) and $K_{\rm acceptance}$, explicitly. Referring to Refs. [@Schmidt-Sorensen:nd; @Adams:2003vd], we have examined the methods of correction used by NA44 Collaboration \[17\]. In this addendum, thus we can present the data including $K_{\rm SPS}$ and $K_{\rm acceptance}$ and analyze them by means of Eqs.(9), (12) and (14). Before concrete analyses, we categorize raw and corrected data in Table \[table4\].
--------------------------------- -----------------------------------------------------------------------------------------
1\) Raw data raw data
2\) Quasi-corrected data (Q-CD) $({\rm raw\ data}) \times K_{\rm SPC} \times K_{\rm acceptance}$
3\) Corrected data with Gamow $({\rm raw\ data}) \times K_{\rm SPC} \times K_{\rm acceptance} \times K_{\rm Gamow}$
4\) Corrected data with Coulomb $({\rm raw\ data}) \times K_{\rm SPC} \times K_{\rm acceptance} \times K_{\rm Coulomb}$
--------------------------------- -----------------------------------------------------------------------------------------
: Category of data. Notice that $R =$ input is necessary for $K_{\rm Coulomb}$, where $R = 5$ fm is used in Refs. \[17,25,26\].[]{data-label="table4"}
Using the data in Ref. \[17\] with $K_{\rm Coulomb}$ or $K_{\rm Gamow}$, we can obtain the quasi-corrected data (Q-CD) shown in Fig. \[fig6\], which are available for our purpose. Using Eqs. (5) and (9), we obtain the results shown in Table 5. As seen in Fig. 6, the our formulation including the degree of coherence $\lambda$, i.e., Eq. (9), seems to be available for analyses of the charged $\pi$ BEC.
Formulas $C$ $R$ \[fm\] $\lambda$ $\chi^2/N_{dof}$
---------- --------------- --------------- --------------- ------------------
Eq. (5) 0.94$\pm$0.03 5.83$\pm$0.40 — 30/16
Eq. (9) 0.91$\pm$0.03 2.89$\pm$0.39 0.45$\pm$0.05 6.7/15
: Estimated values for the data \[17\] by Eqs. (5) and (9) using CERN-MINUIT program.[]{data-label="table5"}
To examine availability of the core-halo model described by the Coulomb wave functions, i.e., Eqs. (12) and (14), we analyze the quasi-corrected data (Q-CD). Our results are shown in Fig. \[fig7\] and Table \[table6\]. Coincidences between $3\pi$ BEC and $3\pi \to 2\pi$ BEC in Fig. \[fig7\] ( $f_c$ vs. $p$) seem to be more enlarged than those of Fig. 5.
In this addendum, we have analyzed the quasi-corrected data (Q-CD) named in Table 4, utilizing Eqs. (5), (6), (12) and (14). It can be said that our formulation works well for the description of the charged $3\pi$ BEC.
[cccc]{} $p$ & 1.0 & 0.8 & 0.6\
\
$f_c$ & 0.67$\pm$0.04 & 0.70$\pm$0.04 & 0.76$\pm$0.05\
$\lambda_3^*$ & 0.75 & 0.77 & 0.78\
$R$ (fm) & 2.89$\pm$0.39 & 3.22$\pm$0.46 & 3.47$\pm$0.51\
$R_{\rm plane}$ (fm) & 4.33$\pm$0.58 & 4.83$\pm$0.69 & 5.21$\pm$0.76\
$\chi^2/N_{dof}$ & 6.7/15 & 6.8/15 & 6.8/15\
\
$f_c$ & 0.67$\pm$0.03 & 0.69$\pm$0.03 & 0.73$\pm$0.03\
$\lambda_2^*$ & 0.45 & 0.46 & 0.45\
$R$ (fm) & 4.42$\pm$0.38 & 5.01$\pm$0.45 & 5.49$\pm$0.50\
$\chi^2/N_{dof}$ & 14.5/17 & 14.7/17 & 14.7/17\
As seen in the explanation above, Fig. 1(a) and the upper line (Eq. (5)) of Table 1 , and Fig. 3(a) and the upper line (Eq. (9)) of Table 2 should be replaced by Fig. 6 and Table 5. Moreover, Fig. 5 and Table 4 should be replaced by Fig. \[fig7\] and Table \[table6\].
One of authors (M. B.) would like to thank H. Bøggild, T. Csörgő and B. Lörstad for useful conversations.
[99]{} J. Schmidt-Sorensen, Ph. D. Thesis, UMI-18-1281, 1999.
J. Adams [*et al.*]{} \[STAR Collaboration\], arXiv:nucl-ex/0306028; C. Adler [*et al.*]{} \[STAR Collaboration\], Phys. Rev. Lett. [**87**]{} (2001) 082301
[^1]: The correlation functions for two and three-identical particles are given as usual by $$\frac{N^{(2+\ {\rm or }\ 2-)}}{N^{BG}} = \frac{P({{\rm {\bf k}}}_1,\ {{\rm {\bf k}}}_2)}
{P({{\rm {\bf k}}}_1)P({{\rm {\bf k}}}_2)} \quad{\rm and}\quad
\frac{N^{(3+\ {\rm or }\ 3-)}}{N^{BG}} = \frac{P({{\rm {\bf k}}}_1,\ {{\rm {\bf k}}}_2,\
{{\rm {\bf k}}}_3)}{P({{\rm {\bf k}}}_1)P({{\rm {\bf k}}}_2)P({{\rm {\bf k}}}_3)}\ ,$$ where ${{\rm {\bf k}}}_i$ is the momentum of particle $i$, and $P({{\rm {\bf k}}}_1,\
{{\rm {\bf k}}}_2)$ and $P({{\rm {\bf k}}}_1,\ {{\rm {\bf k}}}_2,\ {{\rm {\bf k}}}_3)$ are two and three particles probability densities, respectively. The probability densities for two-identical particles case can be written as, $$\int \int \left|\psi_{{{\rm {\bf k}}}_1{{\rm {\bf k}}}_2}^{\rm BE}({{\rm {\bf x}}}_1,\
{{\rm {\bf x}}}_2)\right|^2\rho ({{\rm {\bf x}}}_1)\rho ({{\rm {\bf x}}}_2) d^3 {{\rm {\bf x}}}_1d^3
{{\rm {\bf x}}}_2\ ,$$ where $$\psi_{{{\rm {\bf k}}}_1{{\rm {\bf k}}}_2}^{\rm BE}({{\rm {\bf x}}}_1,\ {{\rm {\bf x}}}_2) =
\frac 1{\sqrt 2}[\psi_{{{\rm {\bf k}}}_1{{\rm {\bf k}}}_2}({{\rm {\bf x}}}_1,\ {{\rm {\bf x}}}_2) +
\psi_{{{\rm {\bf k}}}_1{{\rm {\bf k}}}_2}({{\rm {\bf x}}}_2,\ {{\rm {\bf x}}}_1)]\ ,$$ $\rho ({{\rm {\bf x}}}_i)$ stand for the source functions of particle $i$.
[^2]: It should be remembered that NA44 Collaboration data are for the variable $$Q_3^2(4D) = (k_1 - k_2)^2 + (k_2 - k_3)^2 + (k_3 - k_1)^2$$ where $k_i$ are four-momentum of charged particles. $Q_3 = \sqrt{Q_3^2}$. However, in our calculations we assume that $q_{0,ij}^2 = ({\rm k}_{0i} -
{\rm k}_{0j})^2 \approx 0$ and use, instead, $$Q_3^2(3D) = ({{\rm {\bf k}}}_1 - {{\rm {\bf k}}}_2)^2 + ({{\rm {\bf k}}}_2 - {{\rm {\bf k}}}_3)^2 +
({{\rm {\bf k}}}_3 - {{\rm {\bf k}}}_1)^2\ .$$ There is an approximation between them $$Q_3(4D) \cong Q_3(3D) - \sum q_{0,ij}^2/2Q_3(3D)\ .$$ $Q_3(4D)$ depends on sum of squared energy differences.
[^3]: In other words, ideal data sets for Eq. (\[eq9\]) are of the form of $${\rm Corrected\ data} = ({\rm raw\ data}) \times K_{\rm spc} \times
K_{\rm acceptance} \times K_{\rm Gamow}\ ,$$ where $K_{\rm Gamow} = 1/(G(\eta_{12})G(\eta_{23})G(\eta_{31}))$.
[^4]: For the numerical calculations of Eqs. (\[eq5\]) and (\[eq9\]) (in order to save the CPU-time), we have first calculated 10$^3$ k values of the Coulomb wave functions, which were then used together with some interpolation procedure during the concrete calculations. In this way we could make use of the CERN-MINUIT program in our analyses.
[^5]: Analysis of these data will be reported elsewhere [@biya00].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Convolutional Neural Networks (CNNs) exhibit remarkable performance in various machine learning tasks. As sensor-equipped Internet of Things (IoT) devices permeate into every aspect of modern life, the ability to execute CNN inference, a computationally intensive application, on resource constrained devices has become increasingly important. In this context, we present Cappuccino, a framework for synthesis of efficient inference software targeting mobile System-on-Chips (SoCs). We propose techniques for efficient parallelization of CNN inference targeting mobile SoCs, and explore the underlying tradeoffs. Experiments with different CNNs on three mobile devices demonstrate the effectiveness of our approach.'
author:
- |
Mohammad Motamedi, Daniel Fong, and Soheil Ghiasi\
Electrical and Computer Engineering Department, University of California, Davis, USA.
bibliography:
- 'IEEEabrv.bib'
- 'ESL.bib'
title: 'Cappuccino: Efficient Inference Software Synthesis for Mobile System-on-Chips'
---
[Shell : Bare Demo of IEEEtran.cls for IEEE Journals]{}
Introduction
============
Neural Networks (CNNs) have proven to be one of the most effective approaches to feature extraction [@krizhevsky2012imagenet; @szegedy2015going]. While frameworks such as Caffe [@jia2014caffe] or Torch [@collobert2011torch7] are commonly used for training CNN models, inference using trained CNNs on resource-constrained platforms remains a challenge. We hypothesize that platforms based on mobile system-on-chips (SoCs) will be a major player in the emerging IoT landscape due to their rich feature set and market forces, and thus, we contend that efficient CNN inference on such platforms is increasingly essential.
Forward evaluation of a trained CNN, also known as inference, is computationally intensive. The research community has put forth a number of solutions for accelerating CNN inference on different platforms, including the design of a customized ASIC chip ([@chen2014dadiannao; @jouppi2016google]), FPGA-based accelerator design ([@zhang2015optimizing; @motamedi2016design; @qiu2016going]), and parallelization on server-grade graphics processing units (GPUs). Latifi *et al.* offered a library for parallel execution of CNNs on mobile devices [@latifi2016cnndroid].
We present Cappuccino, a tool for automatic synthesis of efficient CNN inference software targeting mobile SoCs. In addition to the software synthesis capability, Cappuccino features a novel approach to zero-overhead utilization of vector instructions. Furthermore, it considers the effect of inexact computing on classification accuracy, and leverages imprecise arithmetic to further optimize the computation.
Convolutional Neural Networks {#SEC:CNN}
=============================
Modern Convolutional Neural Networks (CNNs) have millions of parameters, whose values are obtained during training. Each CNN has multiple convolutional layers, which use 3D filter banks for feature extraction. The convolution result of kernels of a filter bank with Input Feature Maps (IFMs) is accumulated to create Output Feature Maps (OFMs). The number of IFMs, the number of OFMs, and the output size are $N$, $M$, and $Wout \times Hout$, respectively. The convolution operation is visualized in Figure \[fig:conv\]. Thus, a CNN layer has $M \times N$ kernels ($M$ filter banks with $N$ kernels each). Kernels have dimension of $K \times K$. Each pixel in an OFM is the sum of convolutions between kernels and the corresponding pixels in IFMs. To generate adjacent pixels in an OFM, the kernel bank is slid across IFMs by a stride of $S$. A simplified pseudo-code for a convolution operation is shown in Figure \[fig:conv\_code\]. The vast majority of CNN inference execution time is spent in convolutional layers [@cong2014minimizing], and thus, we restrict our discussion to them.
![A convolutional layer. $M$ filter banks are convolved with $N$ IFMs to generate $M$ OFMs.[]{data-label="fig:conv"}](conv){width="\columnwidth"}
![A simplified pseudo-code for a convolutional layer. The computation involves six nested loops, which results in a high computational complexity.[]{data-label="fig:conv_code"}](conv_code){width="7cm"}
Cappuccino {#SEC:CAP}
==========
{width="\textwidth"}
\[fig:Cappuccino\]
In order to use a CNN for inference on mobile devices, one has to evaluate its forward path with known parameter values. Cappuccino serves this very purpose in that, it synthesizes an optimized SoC-based inference software for a given CNN description. Our current embodiment of Cappuccino synthesizes the CNN in form of an optimized RenderScript program, which exploits the available processing resources on a mobile SoC to execute the computation. Depending on the target SoC, the generated program will be typically launched on multiple CPU cores, the mobile GPU, and the mobile DSP.
As Figure \[fig:Cappuccino\] illustrates, Cappuccino requires three inputs. The first is a network description file that contains the CNN architectural information such as number, size, and type of its layers. The second input is a model file, which contains the weight and bias parameter values. Cappuccino reorders CNN parameters to improve the performance of vectorized operations. Parameter reordering does not change the model size, and occurs during compile-time. The third input is the validation dataset that was originally used during training of the CNN. Using this dataset, Cappuccino analyzes the impact of optimizations, such as inexact computing on the given CNN, to determine the suitability of utilizing imprecise arithmetic for the given CNN. The result of this analysis guides the corresponding decision during software generation.
Inference Optimization Strategies {#SEC:ACC}
=================================
Thread Workload Allocation
--------------------------
Convolutional layers contain three main sources of parallelism: Kernel-Level Parallelism (KLP), Filter bank-Level Parallelism (FLP), and Output-Level Parallelism (OLP). In accelerating a CNN, one or more of these types of parallelism should be used for workload allocation to threads.
### Kernel-Level Parallelism (KLP)
In KLP, parallelization is obtained by executing the computations for convolving a kernel with corresponding IFM pixels in parallel. Hence, each thread computes one multiplication, and the final result is generated via accumulation by an eventual reduction operation.
### Filter bank-Level Parallelism (FLP)
In FLP, filter banks exploit parallelism by allocating kernel computation to separate threads. In this case, each thread computes the convolution of an entire kernel. Subsequently, a reduction addition yields the final result.
### Output-Level Parallelism (OLP)
In OLP, computation of output pixels are carried out in parallel. That is, each software thread computes the 3D convolution of an entire filter bank of $N$ kernels and corresponding pixels in IFMs.
An advantage of OLP is that a kernel loaded by one thread can be reused by other threads that are responsible for generating another pixel in the same OFM. In SoCs with efficient cache systems, it is possible to load each kernel once and use it $Wout \times Hout$ times. In contrast, data loaded in KLP and FLP cannot be reused as efficiently by the same or other threads. Moreover, in KLP and FLP the required reduction incurs additional overhead for thread synchronization and inter-thread data transfer. As such, Cappuccino uses OLP as its primary workload allocation policy at the thread level. Furthermore, it utilizes vector processing to exploit KLP and FLP within each thread.
![Workload allocated to a thread. Values of $w, h$, and $m$ can be computed using the thread Id.[]{data-label="fig:code1"}](code1){width="7cm"}
The output of each convolutional layer is a 3D data structure which includes $\alpha = M \times Wout \times Hout$ elements (also referred to as pixels). These elements can be uniquely identified using three variables: the OFM ($m$), the column ($w$), and the row ($h$) number. Each element is the result of a convolution between the corresponding window of an IFM and a specific filter bank (Figure \[fig:code1\]). Each thread is identified using a unique index $x$, where $x \in [0,~\alpha)$. The identifier index is used to compute values of $w$, $h$, and $m$.
Data Reordering for Vector Processing
-------------------------------------
Cappuccino uses vector processing to further optimize intra-thread workload execution. Before executing a vector instruction, it is necessary to load all of the operands. In most SoCs with vector processing support, the memory bus is wide enough to load multiple words of a contiguous block in one memory access. To utilize this feature, known as memory access locality, model parameter values have to be shuffled around. Conventionally, IFMs and kernel parameters are stored in either row- or column- major order. Therefore, data elements stored in the adjacent memory addresses are either the next element from the same row/column or the first element of the next row/column. If we represent an element’s address using (Layer, Row, Column) format, the data stored in a row major format reads: $$\footnotesize (0, 0, 0), (0, 0, 1), \cdots, (0, 1, 0), (0, 1, 1), \cdots, (0, 2, 0), (0, 2, 1), \cdots
\label{EQ: row}$$ In a $u$-way vector processor, one wants to load at least $u$ operands with a single memory access. Cappucino reorders the model data to achieve this goal. In particular, we propose to store the model data in a map major order, as opposed to row or column major, so that a thread can apply vector instructions to corresponding elements of different maps. Absent of this optimization, vector processing would incur significant overhead at the boundaries of a kernel. For example, assuming $u = 4$, we reorder the model data in the following order (\[EQ: vectorized\]): $$\label{EQ: vectorized}
\footnotesize
\begin{split}
(0, 0, 0), (1, 0, 0), (2, 0, 0), (3, 0, 0), (0, 0, 1), (1, 0, 1), (2, 0, 1), \\ (3, 0, 1), \cdots, (4, 0, 0), (5, 0, 0), (6, 0, 0), (7, 0, 0), \cdots
\end{split}$$
![Efficient vector processing enabled by data reordering from either row or column major to map major. Groups of $u$ Elements, in either gray or blue, in the same spatial location from consecutive feature maps form a vector (assuming $u = 4$).[]{data-label="fig:reshape"}](datareshape){width="7cm"}
A 3D representation of this transform is shown in Figure \[fig:reshape\]. When model data is reordered, Cappuccino reads IFMs as super-words (vectors), performs vectorized convolution, and accumulates the result. The optimized computation is shown in the algorithm of Figure \[fig:code2\].
![Cappuccino optimizes data ordering to enable efficient vector processing. $u$ words are loaded per access. Subsequently, a vectorized MAC operation performs partial convolution on $2u$ operand elements in parallel.[]{data-label="fig:code2"}](code2){width="7cm"}
### Zero-Overhead Dynamic Reordering of OFMs
![Zero overhead reordering of OFMs: Cappuccino directly stores data in the map major, in lieu of row major format. The case of $u = 4$ is illustrated.[]{data-label="fig:data_reshape2"}](datareshape2){width="5cm"}
Note that model data can be reordered and written to a new model file without any overhead as it happens statically at compile-time. However, reordering the input to an intermediate CNN layer is not as straight forward. In CNNs, the output of a layer becomes the input to the next layer. It follows that the output of a layer has to be reordered to allow the use of vectorized operations in computing the next CNN layer. This process has to happen dynamically, and thus, is expected to incur time and energy overhead.
Cappuccino avoids the dynamic data reordering overhead by directly storing elements of the OFMs in map major order as they are computed. Parameters $w$, $h$, and $m$ are used to determine the location of the output element that thread $x$ generates. To store OFMs in map major format, one has to swap the priorities associated with these parameters. For example, the result of computations by the second thread ($x = 1$) is by default stored in the second location of the output memory. After reordering, however, the second element of the output memory must contain $(m = 1, h = 0, w = 0)$. Such an output can be directly used as the input to the next layer without any overhead. Figure \[fig:data\_reshape2\] illustrates the idea.
To create the output in the reordered map major format, we generate indexes for stacks of $u$ layers, instead of a single layer (Figure \[fig:data\_reshape2\]). That is, we start indexing the second row only after all first rows of all $u$ layers are indexed. Equations (\[EQ: w\_new\]) and (\[EQ: h\_new\]) map a thread id to $w$ and $h$, respectively. For computing the value of $m$ (map index), it is required to see which stack and layer a particular output belongs to (Figures \[fig:reshape\] and Figure \[fig:data\_reshape2\]). Equation (\[EQ: m\_new\]) computes the value of $m$. $$\footnotesize w = \lfloor x / u\rfloor~\%~Wout
\label{EQ: w_new}$$ $$\footnotesize h = \lfloor\frac{x}{u \times Wout}\rfloor~\%~Hout
\label{EQ: h_new}$$ $$\footnotesize m = (x~\%~u) + \lfloor\frac{x}{u \times Wout \times Hout}\rfloor \times u
\label{EQ: m_new}
\vspace{-0.4cm}$$
Inexact Computing
-----------------
Modern mobile SoCs tend to support a number of predefined imprecise computing modes that are likely to results in faster or more energy efficient execution [@mitra2013use]. On such platforms, the target processing mode has to strike a balance between the implementation metrics, e.g., runtime or energy dissipation, and the inference classification accuracy.
For example, RenderScript offers two imprecise computing modes for applications that do not need a strict implementation of the IEEE 754 standard, called relaxed and imprecise computing modes. In both modes, the implementation of floating point arithmetic is not fully compliant with the IEEE standard 754 for handling denormalized numbers. The imprecise computing mode is more efficient, but has a lower arithmetic accuracy. In this mode, operations resulting in -0.0 can return +0.0, and operations on INF and NAN are unsupported. Perhaps more importantly, vector processing is only available under imprecise computing modes, in current version of RenderScript. Vector processing under the RenderScript precise computing mode would result in sequential processing of vector elements.
Cappuccino analyzes the given CNN layer by layer to determine the best matching computing mode for every layer of the CNN. In every layer, it utilizes the validation dataset to measure the classification accuracy under different processing modes. Subsequently, Cappuccino determines which layers of a CNN can be processed using inexact arithmetic and which ones demand a precise implementation. The goal is to execute as many CNN layers as possible in inexact modes, under user specified constraints in terms of acceptable degradation in classification accuracy.
In our discussions, the term accuracy arises in reference to either arithmetic accuracy or classification accuracy. The former measures the numerical difference between values computed in exact vs. inexact arithmetic. The latter indicates the inference classification performance of CNNs, for example by measuring the percentage of true positive predictions.
Experimental Results {#SEC:EXP}
====================
Setup
-----
\[TBL:EXE\]
We used Cappuccino to implement three modern CNNs: AlexNet [@krizhevsky2012imagenet], GoogLeNet [@szegedy2015going], and SqueezeNet [@iandola2016squeezenet]. Subsequently, the parallelized implementations are evaluated on three different smartphones with different generations of Qualcomm Snapdragon SoCs. In order to increase the precision in measurements, all experiments have been repeated 100 times, the minimum and maximum observations are omitted, and the average of the remaining 98 observations are reported. In all of the experiments, the smartphones were put in airplane mode, their screen brightness were fully dimmed, and their background processes were stopped to the extent possible.
Runtime and Energy Efficiency
-----------------------------
\[TBL:ENG\]
### Speedup
We executed the synthesized programs on the platforms and measured the execution time. Table \[TBL:EXE\] summarizes the results. Programs synthesized by Cappuccino offer a speedup of at least 31.95X (GoogLeNet on Galaxy S7) and at most 272.03X (SqueezeNet on Nexus 5) compared to the baseline implementation of single-threaded Java. Moreover, the execution time in all but one case is below a second.
### Effect of Inexact Computing
To determine the best inexact computing mode, we use Cappuccino to measure the classification accuracy of the aforementioned CNNs in computing modes supported by target platforms. This analysis is performed on 5000 random images of ILSVRC 2012 validation dataset [@ILSVRC15]. The classification accuracy in imprecise mode turns out to be identical to the exact mode. Hence, Cappuccino recommends utilization of imprecise computing in all layers.
Table \[TBL:EXE\] demonstrates the effect of imprecise computing on execution time. In our experiments, use of imprecise computing mode offers up to 8X speedup compared to the same implementation under exact arithmetic. Note that RenderScript incarnation of the imprecise computing mode enables vector processing in addition to other optimizations, such as using a rapid exception handling for denormalized numbers.
### Comparison with Related Work
Table \[TBL:Comparison\] compares the performance of software synthesized by Cappuccino with the state-of-the-art work [@latifi2016cnndroid]. The proposed solution under exact arithmetic improves the execution time by 1.38X. In addition, when the synthesized software is both parallel and imprecise, it shows up to 11.47X speedup compared to CNNDroid [@latifi2016cnndroid].
### Energy Consumption
Cappuccino invokes many threads, which increases the instantaneous power consumption compared to a sequential program. However, software synthesized by Cappuccino runs drastically faster than a sequential equivalent. This results in reduction of energy consumption. Table \[TBL:ENG\] compares the energy consumption for running SqueezeNet on Nexus 5. Reported numbers are computed by running each program 1000 times, and calculating the average. Measurements are performed twice to showcase repeatability (2000 runs total).
Conclusion {#SEC:CON}
==========
\[TBL:Comparison\]
In this paper we presented Cappuccino, a platform for efficient synthesis of CNN inference software targeting mobile SoCs. Cappuccino leverages RenderScripts via which, it utilizes CPUs, the GPU and the DSP that commonly exist on a mobile SoC to execute a CNN efficiently. Cappuccino performs an assessment on the impact of inexact computing on execution time and classification accuracy. Subsequently, it selects an inexact computing mode that best fits a layer of a CNN. Compared to sequential implementations, programs synthesized by Cappuccino achieve a speedup of at least 31.95X and at most 272.03X, and improve the energy consumption by 7.81X.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'N. Peretto'
- 'G. A Fuller'
- 'A. A Zijlstra'
- 'N. A. Patel'
title: The massive expanding molecular torus in the planetary nebula NGC 6302
---
[We measure the mass and kinematics of the massive molecular torus in the planetary nebula NGC 6302. The nebula is the proto-typical butterfly nebula. The origin of the wing-like morphology is disputed: determining the mass-loss history of the confining torus is an important step in understanding the formation of this structure. ]{} [We performed submillimeter observations with JCMT and the SMA interferometer. The continuum emission as well as the J=2–1 and 3–2 transitions of $^{12}$CO and $^{13}$CO are analysed at arcsecond resolution. ]{} [The CO emission indicates a mass of the torus of $\sim 2$ M$_\odot \pm 1$M$_\odot$. The $^{12}$CO and $^{13}$CO emission matches the dark lane seen in absorption in the H${\alpha}$ image of the object. The CO torus is expanding with a velocity of $\sim8$kms$^{-1}$, centred at $V_{\rm lsr}=-31.5\rm \,km\,s^{-1}$. The size and expansion velocity of the torus indicates that the torus was ejected from $\sim7500$ yr to $2900$yr ago, with a mass-loss rate of $5\times10^{-4}\,\rm
M_\odot\,yr^{-1}$. We also see a ballistic component in the CO images with a velocity gradient of 140kms$^{-1}$pc$^{-1}$. ]{} [The derived mass-loss history of the torus favours binary interaction as the cause of the ejection of the torus. We predict the existence of a companion with an orbital period $P\lsim 1\,$month. ]{}
Introduction
============
Asymptotic Giant Branch (AGB) stars lose between 50% and 80% of their stellar mass at mass-loss rates up to $10^{-4}\,\rm
M_\odot\,yr^{-1}$. About 80% of planetary nebulae (NP), i.e. the visible part of the ionized gas ejected by AGB stars, show elliptical and/or bipolar morphologies (e.g. Manchado 1997). Bipolar PN are believed to form when a fast wind from the now post-AGB star ploughs into the earlier slow AGB wind, amplifying an initial equatorial density enhancement (Kwok et al. 1978, Balick et al. 1987; Frank & Mellema 1994, Zijlstra et al. 2001). Some 15% of PNe show strong morphological evidence for such an equatorial enhancement of cold gas around the central star (Corradi & Schwarz 1995). Detailed studies of these equatorial structures can then give direct insights on the formation of bipolar PNe. Related to bipolar nebulae, butterfly PNe do not show well collimated outflows, but have thicker expanding equatorial torii. The complex morphology of their outflows has been attributed to an interaction between the fast wind and a warped disk (Icke 2003). NGC 6302 is the proto-typical ‘butterfly’ nebula. It is located at a distance $d\sim 1$ kpc (Meaburn et al., 2005). A broad central absorption lane at its centre is suggestive of a massive circumstellar torus. The nebula is driven by one of the highest temperature PN central stars known (Casassus et al. 2000), although due to the high extinction of the absorption lane, it has not been directly detected. Previous CO observations suggest a mass for the torus of $\sim 0.1$ M$_{\odot}$ (Gomez et al. 1989; Huggins & Healy 1989). However, recently Matsuura et al. (2005a) measured a total gas mass of $\sim3$ M$_{\odot}$ from observations of the submillimetre dust continuum emission. This extremely high mass, especially when compared to the $\sim0.7$ M$_\odot$ estimated for the central star (Casassus et al. 2000), prompted us to reexamine the CO emission from this source and study of the kinematics of the circumstellar gas.
Observations
============
JCMT observations
-----------------
NGC6302 was observed with the James Clark Maxwell Telescope (JCMT)[^1] in Hawaii as observing program S04AU08 on 26 May, 18 and 19 June 2004 using the facility 230 GHz and 345 GHz receivers. We observed and , as well as the corresponding transitions of . The data were taken with a velocity resolution of $\sim 0.1$ kms$^{-1}$ for all the transitions, but was rebinned to 0.4 kms$^{-1}$ for analysis.
We raster-mapped the region around NGC 6302 in the and transitions of . Extended emission components were seen at some velocities, but the main emission component corresponding to NGC6302 was found to be unresolved at 230GHz and only slightly resolved at 345GHz. We subsequently took spectra pointing towards the $^{12}$CO emission peak position, taken to be $\alpha=17^{\rm h}13^{\rm m}43.9^{\rm s}$, $\delta=-37^\circ06^{\prime}11^{\prime\prime}$ (J2000).
All observations were made by position-switching to an off position (0,+300) from this central position. This off position was carefully checked from emission which may have contaminated the source observations, but no emission with a peak temperature of greater than 0.2K (with a line width of 1 km/s or more) was found at the frequency of either the or transitions. The telescope pointing was regularly checked during the observations and was found to be constant to within $\pm2$. A linear baseline has been removed from each of the spectra. The main beam efficiency of the JCMT is 0.69 and 0.63 at the frequency of the and respectively. The observational parameters are summarised in Table \[obs\].
------------ ----------- ----------- -------------- -----------------
Transition Frequency Telescope Resolution Beam
(GHz) (kms$^{-1}$) (arcsec)
220.398 JCMT 0.4 20
SMA 0.3 $6.2\times 3.2$
230.538 JCMT 0.4 20
SMA 4.4 $6.2\times 3.2$
345.796 JCMT 0.5 13
330.587 JCMT 0.5 13
------------ ----------- ----------- -------------- -----------------
: Observed transitions[]{data-label="obs"}
SMA observations
----------------
NGC 6302 was observed at a wavelength of 1.3mm with the SMA[^2] on 26 June 2005. We used all 8 antennas in the compact configuration, with a maximum of baseline of 70m. The six shortest baselines range in length from 16.5 to 32 m. The phase center was RA(2000)=17h$^h$13$^m$44.2$^s$ Dec(2000)=-370615.9. The central frequency of the upper side band was set to 230.538 GHz with a total bandwidth of $\sim
2$ GHz (i.e. covering the frequency range $\sim 229.5$ to $\sim
231.5$ GHz). The lower side band was centered at $\sim 220.5$ GHz (i.e. covering $\sim 219.5$ to $\sim 221.5$ GHz). The correlator was configured to provide a velocity resolution of $\sim
0.26$ kms$^{-1}$ for the CO line and $\sim
4.4$ kms$^{-1}$ for the $^{13}$CO line. Weather conditions were good for 230 GHz observations with a relative humidity of 20%, atmospheric opacity at zenith $\tau_{225} = 0.2$ and $T_{sys,
DSB}\approx 250$ K.
The phase and amplitude were calibrated by observing SgrA$^{*}$ for 5 minutes around every 20 minutes of integration on NGC 6302. The angular separation between SgrA$^{*}$ and NGC 6302 is $\sim 11\degr$. The passband calibration was done using Callisto, 3C279 and 3C454.3. The flux calibration was performed using Callisto assuming a brightness temperature of 120 K at 229.45 GHz. The absolute astrometric positional accuracy in the observations is found to be better than 0.3based on observations of the quasars 1911$-$201 and 1924$-$292.
The visibility data were reduced using the SMA version of the Miriad software package. Applying natural weighting, the final maps have a synthesized beam of 6.2$\times$3.2with P.A.=6.9. The rms noise in the continuum image obtained by averaging the LSB and USB data is $\sigma=5.9$ mJy. The rms noise in the CO channel maps, at 1 kms$^{-1}$ resolution, is $\sim 125$ mJy.
![ H${\alpha}$ image from Matsuura et al. (2005a) (colour scale) overlaid with our SMA 1.3 mm continuum map (white contours).The contours are 5$\%$ and then from 10 to 90$\%$ in steps of 20$\%$ of the 1.3 mm continuum emission peak. The emission peak is 0.62 Jy/beam. \[cont\_map\]](ha_cont.eps){width="7.5cm"}
Continuum emission
==================
The SMA 1.3 mm continuum map is shown in Fig. \[cont\_map\]. To avoid molecular line contamination of our continuum data, the CO emission, on which the the correlator bandwidth was centred, has been removed. This continuum map reveals a single source elongated in the North-South direction. The angular resolution provided by the SMA spatially resolves the structure. The position, size and flux are given in Table \[cont\], together with the integrated flux measured inside the 5$\sigma$ contour ($\sim30$ mJy/beam).
The 1.3 mm continuum flux contains contributions from dust emission and thermal free–free emission. The free-free contribution is $\sim50$% of the total flux (see Fig. 13 of Matsuura et al. 2005a), leaving $0.65$ Jy from the dust continuum emission. Assuming a mean dust temperature ranging from 30 K to 120 K (Kemper et al. 2002), a dust to H$_2$ mass ratio of 1$\%$, a typical dust opacity for protostellar disks of 2 cm$^2$g$^{-1}$ (Beckwith et al. 1990), and a distance $d\sim 1$ kpc (Meaburn et al., 2005), we infer a total (dust plus gas) mass ranging from 0.8 M$_{\odot}$ to 3.9 M$_{\odot}$. However, the temperature range given by Kemper et al. (2002) corresponds to two different components, a cold one with T$_k =$ 30 to 60 K, and a hot one with T$_k =$ 100 to 120 K. As discussed later (see Sect. 5.3), the bulk of the mass comes from the cold component, meaning that 0.8 M$_{\odot}$ is rather a lower limit to the mass. Taking 60 K as the average temperature, we find a mass of 1.8 M$_{\odot}$. Uncertainties due to temperature and dust opacity are at least a factor of 2 in each way. Comparing this mass estimate to the one from Matsuura et al. (2005a) obtained from single dish observations, we conclude that there is almost no missing dust continuum flux in our SMA data, as confirmed later in Sec. 5.1 through our molecular line data set. Because the SMA is not sensitive to extended emission larger than approximately 20 the lack of missing flux confirms then the compact nature of the NGC 6302 torus.
The peak of the millimetre map is slightly offset, i.e. $\sim
2$, from the centre of the dark lane. Two effects explain this offset. First, the contribution from the free-free emission seen at 6 cm (Matsuura et al. 2005a) shifts the 1.3 mm peak toward the location of the free-free emission peak. Second, the dust absorption is tracing only the foreground part of the dusty torus, whilst the 1.3 mm emission traces it entirely. A slight tilt of the torus results in the 1.3 mm emission appearing offset from the absorption.
------------- ---------------------------- ------------ ----------- ----------------- --
Tracer Coordinates S$_{peak}$ S$_{int}$ FWHM
(J2000) (Jy/beam) (Jy) (arcsec)
Cont. 1.3mm 17:13:44.45 $-$37:06:11.1 0.62 1.26 $6.0\times4.6$
CO 17:13:44.47 $-$37:06:08.0 $11.3\times4.2$
------------- ---------------------------- ------------ ----------- ----------------- --
Line data
=========
Averaged spectrum
-----------------
The JCMT data show a compact CO structure approximately coincident with the continuum emission, which traces the molecular content. The CO spectra for all four transitions, at the central position, are shown in Fig. \[spec\_jcmt\]. The spectra show several broad components, with emission ranging between $-15$ and $-65\,\rm km \,s^{-1}$, using local standard of rest (LSR) velocities. In addition, several narrow components are seen in only, and are better seen in than in . These components are at $-38$, $-30$ and $-10\,\rm km \,s^{-1}$. The relative strengths of these features suggest they trace gas which has a high / ratio, and may be spatially extended.
The SMA has both higher resolution, and as an interferometer it is not sensitive to very extended emission. This allows us to identify interstellar emission components, which will be absent from the SMA maps. The averaged obtained with the SMA is shown in Fig. \[spec\_sma\] [^3]. Two of the narrow components are absent, but the one at $-38\,\rm km \,s^{-1}$ is seen, suggesting this gas is associated with the molecular torus. Its weakness at in the JCMT data can be attributed to an off-centre position combined with the smaller beam at this frequency. The narrow components at $-30$ and $-10\,\rm km \,s^{-1}$, non observed on our SMA data, may be unrelated interstellar gas, or possibly associated with the extended (8 arcmin) emission of the planetary nebula.
Based especially on the averaged spectrum, we define three broad velocity components, at $\sim-25$ , $\sim-40$ and a feature/wing at $\sim-50$ to $-60$ . These features are present in all four spectra, but are best defined at . The same three components are also seen in the CO of Gomez et al. (1989). The low velocity feature is much stronger in the SMA integrated spectrum than in the JCMT spectrum, indicating it is located off-centre. It is absent from the JCMT spectrum.
![$^{12}$CO and $^{13}$CO spectra obtained with the JCMT at the position RA(2000)= 17$^h$13$^m$43.9$^s$ Dec(2000)=-370611. Note the presence of three velocity components in each line. \[spec\_jcmt\]](cospectra_Tmb.eps){width="8cm"}
![Averaged $^{12}$CO spectrum obtained with the SMA, integrated over the full spatial extent of the source. The dashed lines mark the velocity range of each of the velocity components. The stellar velocity, estimated at $-31.5 \,\rm km\,s^{-1}$ (cf Sec. 6.2), is approximately the boundary between the medium and high velocity components. \[spec\_sma\]](sumspec.eps){width="4.5cm"}
SMA $^{12}$CO and $^{13}$CO maps
--------------------------------
The SMA channel maps allow us to define the precise velocity ranges for each of the three broad components. These range from $-64$ to $-48$ for the low velocity component, from $-48$ to $-30$ for the medium velocity component, and from $-30$ to $-15$ for the high velocity component. The relative intensity of these components varies from position to position, resulting from a complex velocity structure in the source.
The $^{13}$CO transition is optically thinner, and then is a better tracer of the mass distribution and kinematics than the $^{12}$CO . In the following we then discuss further the $^{13}$CO emission. Figure \[velcompco21\] displays the full integrated $^{12}$CO (colour scale) and $^{13}$CO (contours) emission observed with the SMA. We can see that the $^{13}$CO emission is more structured and elongated than the $^{12}$CO one. Comparing with the continuum emission, we see that the line emission is more extended, and peaks three arcseconds to the North of the continuum peak (see Table \[cont\]). The lower contour levels also show extensions to the South and East which are not as strong in the continuum. Part of the difference may be caused by the free-free contribution to the continuum.
![ Superposition of the $^{12}$CO integrated intensity map (colour scale) with the $^{13}$CO integrated intensity map (contours), both integrated over the full velocity range, i.e. from -64 to -15 km s$^{-1}$. The peak value of the $^{12}$CO emission is 190 K km s$^{-1}$, while for the $^{13}$CO emission the contours go from 20 to 90 $\%$ in steps of 10$\%$, with a peak at 78 K km s$^{-1}$. \[velcompco21\]](12co_13co_int.eps){width="8cm"}
Fig. \[velcomp13co21\] shows the low velocity interval of the (a) and (b), as well as the medium and high intervals of the line (b and d). The low velocity component (Fig. \[velcomp13co21\]a and b) differs significantly from one line to another. The $^{13}$CO shows only one clear component, which is offset to the South-East compared to the brighter (and closest) component. The missing emission peak in the $^{13}$CO is likely too weak to have been detected in these observations which have a one $\sigma$ noise level of $\simeq2$ K . Its weakness indicates a lower opacity for this clump, which is not apparent in $^{12}$CO where both clumps are optically thick. The cause for the position shift is less clear, but it may indicate a difference between the location of the temperature peak, traced by the optically thick peak, and the column density peak, traced by the optically thin peak.
The medium velocity component (Fig. \[velcomp13co21\]c) matches the dark lane seen in the H${\alpha}$ map better than the $^{12}$CO, and leaves little doubt about its association with the dark lane. We note also a slightly curved shape of this component. The high velocity component (Fig. \[velcomp13co21\]d) is more extended than its $^{12}$CO counterpart, and is curved in the opposite direction to the medium velocity component. The lack of evidence for dust extinction associated with this strong CO component suggests it is located behind the H${\alpha}$ emitting material. The medium and high velocity peaks are situated on either side of the continuum peak, but shifted North by a few arcseconds. The location and velocity of these two components suggests they belong to an expanding, toroidal structure, but the northward shift suggests the torus may be weaker or incomplete towards the South.
{width="12cm"}
In the next section we will discuss the mass of this torus based on CO observations
Mass determination {#sec:mass}
==================
------------------- ----------------- ---------------------------------- ---------------------------------- ----------------------------------- -----------------------------------
Source Name Velcocity range $I_{\rm int}^{^{12}\rm CO(3-2)}$ $I_{\rm int}^{^{12}\rm CO(2-1)}$ $I_{\rm int}^{ ^{13}\rm CO(3-2)}$ $I_{\rm int}^{ ^{13}\rm CO(2-1)}$
(kms$^{-1}$) (Kkms$^{-1}$) (Kkms$^{-1}$) (Kkms$^{-1}$) (Kkms$^{-1}$)
low vel. comp. -64 to -48 20.0 (2.4) 12.5 (1.4) 1.8 (0.3) 2.6 (0.2)
medium vel. comp. -48 to -30 56.1 (2.6) 27.3 (1.5) 17.4 (0.3) 8.7 (0.2)
high vel. comp. -30 to -15 40.7 (2.4) 15.4 (1.4) 6.2 (0.3) 3.0 (0.2)
------------------- ----------------- ---------------------------------- ---------------------------------- ----------------------------------- -----------------------------------
In order to determine the total molecular mass of the torus we estimate the excitation temperature and opacity of the CO lines using two methods. These both rely on a good knowledge of the observed $^{12}$CO and $^{13}$CO line ratios.
Since our JCMT $^{12}$CO observations only barely resolve the torus, even in the transition, we assume that the central JCMT spectra contains all the emission from the torus and so we can use these spectra alone to determine the total molecular mass. However, note that this is not true for the low velocity component which is not part of the torus, as this component peaks at 7 South-East.
Table \[co\_prop\] gives the JCMT main beam integrated intensities of the four transitions, $^{12}$CO , $^{12}$CO , $^{13}$CO , and $^{13}$CO , integrated over the three velocity components identified in our SMA data. As noted above the CO intensity is likely underestimated for the low velocity component. The JCMT beam filling factors can be calculated quite accurately for each velocity component using the source size measured in our SMA observations. These beam filling factors are listed in Table \[fil\_fac\] assuming that the size of the source is the same for both CO species and both transitions. The filling factor corrected line ratios are given in Table \[ratio\]. Also, the comparison of the JCMT and SMA $^{12}$CO integrated flux shows that the SMA observations recover $\sim 80\%$ of the JCMT flux, demonstrating once again the compact nature of NGC 6302.
Source Name BFF CO BFF CO
------------------- -------- --------
low vel. comp. 0.62 0.93
medium vel. comp. 0.43 0.64
high vel. comp. 0.31 0.57
: Beam filling factors (BFF)[]{data-label="fil_fac"}
Source Name $I_{\rm int}^{^{12}\rm CO(3-2)}$/$I_{\rm int}^{^{12}\rm CO(2-1)}$ $I_{\rm int}^{^{13}\rm CO(3-2)}$/$I_{\rm int}^{^{13}\rm CO(2-1)}$ $I_{\rm int}^{^{12}\rm CO(3-2)}$/$I_{\rm int}^{^{13}\rm CO(3-2)}$ $I_{\rm int}^{ ^{12}\rm CO(2-1)}$/$I_{\rm int}^{ ^{13}\rm CO(2-1)}$
------------------- ------------------------------------------------------------------- ------------------------------------------------------------------- ------------------------------------------------------------------- ---------------------------------------------------------------------
low vel. comp. 1.07 (0.25) 0.46 (0.11) 11.11 (3.18) 4.81 (0.91)
medium vel. comp. 1.38 (0.14) 1.34 (0.05) 3.22 (0.21) 3.14 (0.24)
high vel. comp. 1.35 (0.22) 1.12 (0.13) 6.56 (0.70) 5.13 (0.81)
Analytical estimates of the kinetic temperature and opacity
-----------------------------------------------------------
The ratio of two different transitions of the same molecule constrains the excitation temperature, as given by Eq.(3) of Levreault (1988), assuming that the beam filling factors and the opacities of the transitions are known. For simplicity, we assume that the excitation temperature is uniform along the line-of-sight and that the transitions are thermalized, which allows us to infer the opacity of one transition given the opacity of the other one.
For the NGC 6302 case, we know that the $^{13}$CO lines are not optically thick (i.e. $\tau < 1$). So assuming an opacity of 0.5 for the $^{13}$CO line we find a kinetic temperature of $T_{\rm k} = 25\pm5$K for the high velocity component and $T_{\rm
k} = 35\pm5$K for the medium velocity component. For the low velocity component we find $T_{\rm k} \le 10$K, but because of the missed flux in the $^{13}$CO spectrum (Sec. \[sec:mass\]) we consider this value underestimated by a factor of $\sim2$ and believe that kinetic temperature of $T_{\rm k} \simeq
20$K is more realistic for this component.
With this range of kinetic temperatures inferred from the $^{13}$CO ratio, it is then possible to infer the opacity of the $^{12}$CO lines. We find a maximum opacity for the $^{12}$CO transition of unity for every velocity component, taking into account the uncertainties on the observed ratio.\
The opacity and kinetic temperature can also be inferred from the line ratio of two isotopic substituted molecules for the same transition. This method does not require the knowledge of the beam filling factor, if the filling factor is the same for both species, but it does require knowledge of the relative abundance of the species used. Assuming that the beam filling factor is the same for both transitions, that the excitation temperature is the same for both transitions, and also that the linewidth is the same in both transitions, the ratio of the integrated intensities is given by Eq.(3) of Myers et al. (1983).
Although the terrestrial value of $A_b$, the abundance ratio $X[^{12}\rm CO]$/$X[^{13}\rm CO]$, is $90$, for PNe $A_b < 30$ (Balser et al. 2002), and 15 is a lower limit for NGC 6302 (Sec. \[sec:twocomp\]). For $A_b=15$ the observed $^{12}$CO(2-1)/$^{13}$CO(2-1) ratios lead to a $^{12}$CO line opacity of 3 for the high and low velocity components and 5.5 for the medium velocity component, with a typical uncertainty of $\pm0.5$. The corresponding opacities for the $^{12}$CO transitions are then $3\pm0.5$ and $7.5\pm0.5$ for the high and medium velocity components, respectively. The opacity inferred for the low velocity component is $\sim 1\pm1.5$, well below the one inferred from the transitions. This is likely due to the issue with the $^{13}$CO already mentioned in the previous section. The corresponding opacities in the $^{13}$CO transitions are thus 15 times smaller, that is $\sim0.1$ for the low velocity component, 0.2 for the high velocity component, and $\sim0.5$ for the medium velocity component, with a typical uncertainty of $\pm0.05$. For an abundance ratio $A_b=30$ rather than 15, the opacities are approximately doubled.
The $^{12}$CO opacities inferred through this method are 3 to 15 times higher than the ones inferred using two transistions of the same species. Such differences in the opacity estimates lead to a similar dispersion in the mass estimates and indicate the need for a more complete model.
![$^{12}$CO(2-1)/$^{13}$CO(2-1) ratio from the SMA data. The star symbol shows the assumed position of the central star (Matsuura et al. 2005a). **a)** High velocity component. **b)** Medium velocity component. \[ratio\_map\]](ratio.eps){width="6.5cm"}
A two phase model: A hot inner edge and a cold outer part {#sec:twocomp}
---------------------------------------------------------
The most likely explanation for the discrepancy between the two analytical methods discussed above is the presence of opacity and temperature gradients within the beam (Levreault 1988). Given the presence of a luminous, hot stellar remnant in the centre of NGC 6302, such gradients would not be unexpected.
Figure \[ratio\_map\] shows maps of the ratio $^{12}$CO(2-1)/$^{13}$CO(2-1) for the medium and high velocity components. For both components this ratio is far from uniform over the structure. The high velocity component shows higher values of this ratio than does the medium velocity component; the peak is located in the central part, close to the proposed location of the central star (Matsuura et al. 2005a). Direct or indirect energy input by the central star appears to be responsible for a temperature gradient within the molecular structure, making the gas hotter in the inner part and colder in the outer part.
Figure \[ratio\_map\] suggests an inner, marginally optically thin, warm component, seen directly through the high velocity component, and an outer, optically thicker, cold part, obscuring the warm component associated with the medium velocity component. This is consistent with the interpretation above that the high velocity component is located behind the star, while the medium velocity component is in front of star (Sec. 4.2).
The highest value of the $^{12}$CO(2-1)/$^{13}$CO(2-1) ratio, found for the low velocity component, gives a lower limit to the abundance ratio $A_b$ of 15.
To account for the observed non-uniform physical conditions we have modelled the CO emission assuming it arises from two phases with different temperatures. Using the 1D radiative transfer code RADEX (Schöier et al. 2005) we have constructed a grid of models varying the kinetic temperature and the column density of the CO. The volume density, the linewidth and the background temperature were fixed to $1\times10^5$ cm$^{-3}$, 10 and 2.73 K, respectively. We used two different values, 15 and 30, for the abundance ratio between $^{12}$CO and $^{13}$CO. Four grids of models were constructed, one for each transition, $^{12}$CO , $^{12}$CO , $^{13}$CO , and $^{13}$CO . Each individual model within these grids gives the resulting radiation temperature for a medium with a different column density and kinetic temperature. The $^{12}$CO column density was varied from $5\times10^{16}$ cm$^{-2}$ to $5\times10^{18}$ cm$^{-2}$ in steps of $5\times10^{16}$ cm$^{-2}$, and the kinetic temperature was varied from 10 K to 300 K in steps of 5 K. In order to simulate a two-phase medium we mixed each point on each grid with every other point, adopting a contribution of each phase to the observed spectra ranging from 10$\%$ to 90$\%$ in steps of 10$\%$.
This procedure generated almost $6\times10^7$ different two component models. We then built for each model, the four line ratios, i.e. $^{12}$CO(3-2)/$^{12}$CO(2-1), $^{13}$CO(3-2)/$^{13}$CO(2-1), $^{12}$CO(3-2)/$^{13}$CO(3-2), and $^{12}$CO(2-1)/$^{13}$CO(2-1). A $\chi^2$ analysis on these grids yields the models which best fit the observed ratios for each of the three velocity components.
Figure \[model\_high\] shows the results of the radiative transfer modelling for high velocity component. These plots only shows the model ratios fall within the uncertainties of the observed ratios (Table \[ratio\]), for the preferred abundance ratio of 15. We can clearly see on Fig \[model\_high\]a that two different phases are needed to reproduce the observed ratios, one cold phase, with a kinetic temperature $\sim20$ K and a warmer phase, with a kinetic temperature ranging from $\sim$80 K to 300 K, depending on the model.
Figure \[model\_high\]b shows the mass distribution of these models. For each mass bin we show in red and blue the relative percentage of hot and cold material, respectively. Here, hot means a kinetic temperature of $T_{\rm k}>80$K and cold means $T_{\rm k}<80$K. The most likely mass for the high velocity component is $\sim 0.3$M$_{\odot}$ and this mass is largely dominated by hot molecular gas. This is consistent with the fact that the high velocity component exhibits the hot inner part of the torus.
Figure \[model\_med\] shows the same plots for the medium velocity component. We see that the temperature distribution is quite similar to that of the high velocity component, but the mass distribution is very different. The most likely mass for the medium velocity component is $\sim 1.1$ M$_{\odot}$ and this mass is this time largely dominated by cold material. This is consistent with our picture of NGC 6302 since the medium velocity component shows the cold outer part of the torus. Using an abundance of 30 increases the best fit masses by roughly a factor of 2, but we did not find any model for the medium velocity component fitting within the uncertainties of the four line ratios, suggesting 15 is a better choice value for the abundance ratio.
The derived mass of the medium velocity component is almost four times higher than the mass of the high velocity component. This can be explained by the optical depth effects: the bulk of the cold phase of the high velocity component is hidden by the hot optically thick material in front of it, and thus, does not appear in our mass estimate of the high velocity component. However this cold component actually dominates the total mass of the toroidal structure. This discrepancy shows the limits of our simple two phase model. Although the total mass of the molecular content resulting from our models is 1.4 M$_{\odot}$, geometrical symmetry suggests that the actual mass is closer to 2 M$_{\odot}$, with an estimated uncertainty of 1 M$_{\odot}$ given the mass distribution of the fit models (Fig. \[model\_high\]b and \[model\_med\]b). This value is in very good agreement with the mass estimate from the dust continuum emission (see Sect. 3).
We tried to estimate the mass of the low velocity component in the same way, but excluding the somewhat unreliable $^{13}$CO data for this component did not allow us to identify a clear mass limit. For this component we have estimated a lower limit on the mass by assuming a temperature T$_k = 20$ K, LTE, and using the opacity of the $^{13}$CO line, $\sim0.35$ (obtained in a same way as for the $^{12}$CO opacity; cf Sect. 5.1). Combining equations (9) and (19) from Goldsmith & Langer (1999), and using the $^{13}$CO integrated intensity in Table \[co\_prop\] we find a lower limit of 0.06 M$_{\odot}$ for the mass of the low velocity component. All these mass estimates are summarized in Table \[mass\].
----------------- ----------------------- ----------------------- ---------------
Component N$_{^{13}CO}$ N$_{H_2}$ Mass
(10$^{16}$ cm$^{-2}$) (10$^{21}$ cm$^{-2}$) (M$_{\odot}$)
high velocity 2.9 4.4 0.3
medium velocity 10.7 16.1 1.1
low velocity 0.3 0.5 0.06
----------------- ----------------------- ----------------------- ---------------
: Column densities and mass. The mass of the low velocity component is considered a lower limit. The mass of the high velocity components is only likely underestimated[]{data-label="mass"}
![Results of radiative transfer modelling of a two phase medium for the high velocity component. **a)** Temperature distribution of the models fitting the four observed line ratios within the uncertainties. These models use an abundance ratio $X$\[$^{12}$CO\]/$X$\[$^{13}$CO\] of 15. **b)** Mass distributions of the models fitting the four observed line ratios within the uncertainties. The fraction of hot material (i.e. with $T_{\rm k} > 80$ K) is represented in red, while the fraction of cold material (i.e. with $T_{\rm k} < 80$ K) is represented in blue. \[model\_high\]](model_high_temp_15.eps){width="5cm"}
![Results of radiative transfer modelling of a two phase medium for the high velocity component. **a)** Temperature distribution of the models fitting the four observed line ratios within the uncertainties. These models use an abundance ratio $X$\[$^{12}$CO\]/$X$\[$^{13}$CO\] of 15. **b)** Mass distributions of the models fitting the four observed line ratios within the uncertainties. The fraction of hot material (i.e. with $T_{\rm k} > 80$ K) is represented in red, while the fraction of cold material (i.e. with $T_{\rm k} < 80$ K) is represented in blue. \[model\_high\]](high_mass_coldhot.eps){width="5cm"}
![Same as for Fig. \[model\_high\] but for the medium velocity component. \[model\_med\]](model_med_temp_15.eps){width="5cm"}
![Same as for Fig. \[model\_high\] but for the medium velocity component. \[model\_med\]](med_mass_coldhot.eps){width="5cm"}
Kinematics
==========
The molecular structure surrounding the central star of NGC 6302 is elongated in the North-South direction (see Fig. \[velcompco21\] and \[velcomp13co21\]) and is coincident with the elongated structure previously seen in absorption on the H${\alpha}$ image (Matsuura et al. 2005a). When we overlay the high and medium velocity components observed in CO with the SMA (see Fig. \[ellipse\]a and c), we see that these two compoments appear to form the farside and nearside parts of a thick disk-like shaped structure surrounding the central star. Based on the angular size observed in $^{13}$CO (see Fig. \[ellipse\]c) we can estimate the physical diameter and thickness of this structure. Considering that NGC 6302 is located at $d\simeq1$ kpc (Meaburn et al. 2005), we infer a diameter for the disk-like structure of D$\simeq 0.11$ pc, and a thickness $T\simeq
0.04$ pc. However although the morphology is similar to a disk, the velocity structure is clearly not.
The torus velocity structure
----------------------------
Considering the medium and high velocity CO components as tracing a disk-like structure, the velocity channel marking the limit between the two should correspond to the systemic velocity of NGC 6302. This velocity is found to be about $V_{\rm lsr} = -30\pm 1$ kms$^{-1}$. This is in remarkable agreement with the value of $-30.4$ kms$^{-1}$ derived by Meaburn et al. (2005) given the very different methods used.
For the $^{13}$CO transition, Fig. \[ellipse\]b displays the position-velocity (PV) diagram constructed along a North-South axis (dashed lines) passing through the center of NGC 6302. In this PV diagram we see the presence of a ring-shape structure centered at the systemic velocity, $V_{\rm lsr} = -30$ kms$^{-1}$, and at the position $\delta$(2000)=-37$^d$0612.5. A similar PV diagram is observed in $^{12}$ CO. The low velocity component on the left part of the $^{13}$CO PV diagrams as a fainter peak. The ring-shape structure observed in the PV diagrams does not suggest rotation as it would be the case for a rotating disk. Compared to the systemic velocity the high velocity component, the rear part of the structure, is red shifted, while the medium velocity component is blue shifted. This configuration clearly suggests expanding motion. Rather than a disk, the molecular structure seen around the central star of NGC 6302 is more like an expanding molecular torus.
For a complete, uniform expanding torus seen edge-on, the PV diagram constructed along the torus should ideally reveal an ellipse whose extent along the velocity axis directly gives its expansion velocity. Actually, the shape and overlap of the front and back parts of the torus (see Fig. \[ellipse\]a), as well as the previous measurements of the NGC 6302 outflow angle compared to the plane of the sky (i.e. 18; Meaburn et al. 2005) show that this torus is seen nearly, but not exactly, edge-on. In Fig. \[ellipse\]b we plot an ellipse which delimitates the molecular emission. This ellipse fits quite well the ring-like shape structure, although we note that there is emission missing in the southern part of the red shifted emission (i.e. the high velocity component is less extended than the medium one; cf Fig. \[ellipse\]a). From these ellipses we can infer the mean projected expansion velocity of 9 kms$^{-1}$. Since we know the angle of the outflow axis to the plane of the sky, 18, we can, assuming that the torus axis has the same angle, correct the expansion velocity from the projection. But since this correction corresponds to only 0.5 kms$^{-1}$ which is comparable to the uncertainties, we adopt V$_{\rm exp} = 9$ kms$^{-1}$. If there was some rotation (or overall velocity gradient) associated with the torus the ellipse in the PV diagram would be inclined and possibly twisted. Although difficult to accurately constrain without a detailed model, Fig. \[ellipse\]b indicates that any such gradient has a magnitude of $< 20$ kms$^{-1}$pc$^{-1}$.
The analysis of the lowest velocity CO component is more problematic. The emission from this component is less similar in the $^{12}$CO and $^{13}$CO transitions than is the case for the torus emission. However its greater velocity from the systemic velocity of the sources leads us to interpret this component as associated with the outflow from the source. Interestingly this component is projected onto the source in the region where the torus emission is weakest, perhaps suggesting this may material being accelerated out of the torus.
![**a)** H$\alpha$ map (grey scale) overlaid with the SMA $^{13}$CO medium (solid blue lines) and high (dashed red lines) components. The black dashed line corresponds to the axis along which we constructed the position-velocity diagram shown in (b). **b)** $^{13}$CO position-velocity diagram constructed along the black dashed axis shown in (a). The ellipse delimitates the molecular emission coming from the torus. The 0 position corresponds to $\delta$(2000)=-37$^d$0615.9. \[ellipse\]](torus_13co.eps){width="5.5cm"}
![**a)** H$\alpha$ map (grey scale) overlaid with the SMA $^{13}$CO medium (solid blue lines) and high (dashed red lines) components. The black dashed line corresponds to the axis along which we constructed the position-velocity diagram shown in (b). **b)** $^{13}$CO position-velocity diagram constructed along the black dashed axis shown in (a). The ellipse delimitates the molecular emission coming from the torus. The 0 position corresponds to $\delta$(2000)=-37$^d$0615.9. \[ellipse\]](pv_ellipse_center_13co.eps){width="6.5cm"}
Position-offset diagrams: ballistic flow
----------------------------------------
To uncover the general velocity pattern, we determine for each velocity channel the position and flux of the brightest pixel. The radial offset of this pixel from the stellar position is measured. We assume the star to be located at the position proposed by Matsuura et al. (2005a; cf Fig.\[ratio\_map\]). This technique reduces drastically the amount of information contained in the images, but has been found to provide a powerful tool, used, e.g., in maser studies of circumstellar envelopes.
{width="16cm"}
We apply this method to the SMA $^{12}$CO data, which shows best signal to noise ratio. The result is shown in Fig. \[velpos\]. We select all channels with peak flux above 1 mJy/beam. The left panel shows the positions of these peaks. The stellar position is taken as (0,0). Different symbols indicate the velocity ranges. Larger symbols are used for stronger peak fluxes. The distribution is reasonably well fitted with a tilted ring, with a radius of 5.5, an inclination of 18 (Meaburn et al. 2005), rotated to a position angle of $-5$.
The right panel of Fig. \[velpos\] shows the velocity versus the angular separation between the peak and the assumed stellar position. Such diagrams can distinguish between expanding shells and polar flows (Zijlstra et al. 2001). Both are present here. The solid line is our model for the ring, where the best fit is found for a central velocity $V_{\rm lsr}=-31.5\,\rm km\,s^{-1}$ and an expansion velocity $V_{\rm
exp}=8\,\rm km\,s^{-1}$. These differ slightly from the earlier fits but are within the uncertainties. The symbols clustering around the drawn line shows that part of the CO emission is well derscribed by an expanding ring. In the following sections we consider the values inferred from the best fit model as the reference ones.
On Fig. \[velpos\] (right) we also notice that at higher and lower velocity two linear features appear. These show velocity linearly increasing with distance from the star. This is indicative of polar flows. The ballistic nature of these can be understood as due to the interaction of a slow and a fast wind, where the swept-up interface moves at a constant but direction-dependent velocity. This gives rise to a ballistic flow pattern (Zijlstra et al. 2001), seen in bipolar OH/IR stars.
The velocity range of this component is small in NGC 6302, compared to typical for bipolar OH/IR stars. The gradient is $\sim 140 \,\rm
km\,s^{-1}\,pc^{-1}$. But it shows a considerable spatial extent, 2-3 times the size of the ring. The linear flows are seen on either side of the star, and both approaching and receding components with respect to the systemic velocity, are present.
Discussion
==========
Comparison with other objects
-----------------------------
Molecular gas is relatively common among planetary nebulae. It is seen among the youngest objects where the high density material traps the dissociation and ionization fronts within the nebulae. As the nebula expands and dilutes, the molecular regions will be overrun (Huggins et al. 1996, Woods et al. 2005). In the most recent and the most sensitive survey, Huggins et al. (2005) detect CO emission for 40 nebula out of a sample of 110, although only a few are as bright as NGC6302.
CO torii are known for a number of objects in a similar evolutionary phase. M1-16 (Huggins et al. 2000) has an extensive envelope of 0.12M$_\odot$, around a very compact ionized region. NGC7027 also has a high-mass ($\sim 1\,\rm M_\odot$) CO shell, but its ionized nebula is elliptical and lacks the butterfly shape. A closer optical morphological analogy is NGC 2346, which is also a butterfly nebula: it shows CO distributed in a clumped expanding ring with an estimated mass of 0.1M$_\odot$ (Bachiller et al. 1989).
During the preceding phase of the post-AGB nebulae, where the star has ceased its mass loss and left the AGB, but the star is still too cool to cause significant ionization, CO is always present. Such objects tend to show strong bipolarity; CO velocities up to 200 kms$^{-1}$ are seen (Bujarrabal et al. 2001) although the spectra are dominated by more sedate material. In some cases a Keplerian disk is present (e.g. Bujarrabal et al. 2005). Woods et al. (2005) show that two distinct classes of objects exist in this phase, one rich in different molecules and the other only showing a small subset, but both types show CO emission. The molecule rich class (objects such as CRL618, OH231.8+4.2) shows the highest density torii, similar to that in NGC6302.
However, NGC6302 differs from the other objects in the class, in that it is more evolved, with a much hotter central star and a high mass for the molecular torus. The full line width of 50kms$^{-1}$ is typical for planetary nebulae, althought about half the line width is due to the bipolar, ballistic flow first detected in the SMA data.
Torus properties
----------------
The CO emission in NGC6302 comes predominantly from a compact, expanding torus centred on the exciting star. This CO emission requires that there are at least two different temperature components of the CO. These components are similar in temperature to those identified in the dust emission from this source by Kemper et al. (2002) suggesting that both dust components are associated with the torus rather than the outflow lobes as Kemper et al. suggested for the warmer dust. The properties of the CO emission have a number of important implications for the origin and evolution of this system.
The best estimate of the mass of the torus is $\sim2$ assuming a $^{13}$CO to $^{12}$CO abundance ratio of 15. This mass is similar to that inferred from the submillimetre continuum observations of Matsuura et al. (2005a), but significantly higher than those estimated from previous observations of CO, e.g. Huggins & Healy (1989). This discrepancy probably results from a combination of previous underestimates of both the integrated intensity of the CO line and opacity of the transition. In addition to the molecular material traced by the CO, previous observations have identified 0.05 of ionised HI (Gomez et al. 1989), and 0.25 in low excitation atomic gas traced by CII (Castro-Carrizo et al. 2001) (all scaled to the assumed 1kpc distance) associated with the source. The CO torus therefore completely dominates the mass of the circumstellar material associated with NGC6302.
The high mass of the CO torus means that NGC6302 has the highest mass of inner circumstellar material of any PN or proto-planetary nebula (PPN) known. The survey of PPNe, likely models for the progenitor to NGC6302, by Bujarrabal et al. (2001) found only two objects with masses of circumstellar comparable or larger than found in NGC6302, but both these objects, AFGL2343 and IRAS+10420, are classified as yellow hypergiants with broad CO lines without a clear line core-wing separation, unlike the PPN sources. However it is has been suggested that that the distance to these sources have been overestimated, consequently with much lower masses of circumstellar material (Josselin & Lèbre 2001).
As mentioned above, the lowest velocity CO component is not part of the expanding torus. Spatially it appears close to the region where the torus emission is incomplete, suggestive of it being part of the torus which has been distrupted and accelerated by the outflow from the central star. Evidence that this may be occurring comes from the high excitation lines detected by Feibelman (2001). These lines peak at velocities between $-54$ km/s and $-64$ which would associate them with this most blueshifted CO material. Meaburn et al. (2005) also finds significantly blueshifted forbidden N[ii]{} emission at close to the position of the lowest velocity CO component. This clump of CO material therefore appears to the source of at least some of the material seen in the ionised outflow. The velocity of these UV and optical lines is quite different from the $-23$ to $-34$ velocities of the IR recombination lines and low excitation atomic lines (Casassus et al. 2000; Castro-Carrizo et al. 2001), which in light of the CO torus would appear kinematically to originate from within the CO torus.
Age
---
The size and velocity of the CO torus can be used to infer a number of timescales associated with the system. The radius of the outer edge of the torus is 13, corresponding to $1.3\times10^4$ AU. With the best fit model expansion velocity of $8$ this implies that the event which initiated the expansion of the torus occurred $\sim$7500 years ago. From the apparent thickness of the torus, i.e. 8000 AU, the event which produced the torus lasted for about $\sim$4600 years; the event ended $\sim2900$ years ago. These timescales assume that the limits of the current CO emission trace the full extent of the original material of the torus and that the apparent inner edge marks the end of the ejection event.
The ballistic flow cannot be dated as its inclination angle with respect to the plane of the sky is not known. If we assume it dates from the time after the torus ejection, we find from the velocity gradient (140kms$^{-1}$pc$^{-1}$) that $i>30$ degrees, where $i$ is the angle with the line of sight.
Meaburn et al. (2005) find a younger age from the expansion profile of the north-western optical lobe, 1.7 arcmin from the star. They date the lobe to an age of 1900yr. Within the uncertainties, this may be similar to the end of the torus ejection (as seen in NGC6537, a similar nebula to NGC6302: Matsuura et al. 2005b). However, the bipolar lobes may also have formed later.
Progenitor star
---------------
Combining the total mass of the circumstellar material currently associated with the nebula, $\sim2$ , with the estimated current mass of the central star, 0.7-0.8 (Casassus et al. 2000) gives an absolute lower limit on the mass of the progenitor of the system of $\sim3$. The upper mass limit is 8, based on the fact that stars more massive than this do not become post-AGB stars, but instead ignite their carbon core. This mass limit is consistent with the $\sim4$ progenitor required to produce the high temperature, high-mass stellar remnant. Intriguingly, this means that the bulk of the inital mass of the system is still present close to the star.
Bipolar planetary nebulae have on average higher-mass progenitors, based on their smaller Galactic scale height of bipolar nebulae (Corradi & Schwarz 1995; Stanghellini et al. 2002). This is consistent with the high derived progenitor mass. The formation of bipolar nebulae has been linked to binary progenitors (Soker 1998a) and the angular momentum in such a system could easily drive the mass loss needed to generate the torus. Based on IUE observations, Feibelman (2001) has claimed that NGC 6302 is a binary system with a G[V]{} secondary, however the extensive circumstellar material and the extremely high extinction towards the central star makes this questionable. The UV flux more likely represents scattered light from the central star (see the discussion in Meaburn et al. 2005).
An infrared excess seen at the position of the star (Matsuura et al. 2005a) possibly represents a very compact circumstellar nebula. The excess can be free-free emission (as in Be stars) or very hot dust. The presence of circumstellar material located within the central cavity can be related to binarity (as in post-AGB stars with disks: van Winckel 2002). However, this is indicative only, and the binary nature of the central star remains suspected but unproven.
Casassus et al. (2000) noted that NGC6302 does not show any evidence for a hot wind or wind blown cavity.This is surprising given the evidence for the ballistic flow, and the much faster flows in the ionized gas (Meaburn et al. 2005). The kinematics of the nebula makes it likely that a fast stellar wind did exist in the past.
Mass Loss and Evolution
-----------------------
For the progenitor to eject 2 of material over period of 4600 years (based on the thickness of the torus) implies an average mass-loss rate of $5\times10^{-4}$, concentrated towards the equatorial plane. Even adopting the outer radius only of the torus to estimate the timescale and reducing the mass to minimum measured, 1.4 only reduces the mass-loss rate to $2\times10^{-4}$. This mass loss rate is consistent with those measured towards PPNe, although a factor of a few larger than typical (Bujarrabal et al. 2001). Interestingly the mass-loss rate for NGC6302 is very similar to that infered from the similarly massive, but more spatially extended, CO shell around the hot PN NGC7027 (Masson et al. 1985), although its shell is much less equatorially condensed.
Models show that mass-loss rates can only approach these values during the late AGB stage of evolution. However values as high as that seen for NGC6302 are difficult to produce in the models, even for short periods. The NGC6302 mass-loss rate exceeds any value obtained in models by Vassiliadis & Wood (1993) and is barely reached at the very tip of the distribution of the models by Bloecker (1995).
Mass loss on the AGB is assumed to be driven by radiation pressure on the dust. The achievable rates are limited by the available momentum in the stellar radiation field: this limit can be exceeded by a factor of two if multiple scattering of photons is taking into account (Vassiliadis & Wood 1993). Using the lower limit to the mass-loss rate, the expansion velocity derived above and a stellar luminosity $L=10^4\,\rm L_\odot$, we find that the ratio between the momentum in the torus and the one available in the stellar radiation, is in NGC 6302 :
$$\beta = \frac{\dot M V_{\rm exp}c}{L} = 7.8$$
Considering that the mass loss is concentrated towards the equatorial plane, while the stellar radiation is isotropic, worsens the discrepancy. We conclude that, despite the low $V_{\rm exp}$, the mass-loss history of the torus is difficult to achieve by standard radiation-pressure-driven mass loss. A similar argument, but with much larger discrepancies, is presented by Bujarrabal et al. (2001). On the other hand, the low expansion velocity is as expected from a dust–driven wind at very high $\rm \dot M$ (Habing et al. 1994).
In contrast, the energy requirements are relatively easily satisfied using a hypothetical binary companion. If we assume a 1M$_\odot$ companion in a 1AU orbit, and that the envelope of a 3M$_\odot$ primary star is at the same radius at the onset of mass loss, a reduction of the secondary orbital radius to $\sim 0.1$AU, whilst the mass of the primary reduces from 3 to 1M$_\odot$, will suffice to provide the energy to eject the torus. In the final configuration, the orbital period of the hypothetical binary will be $P\sim
15$days. We conclude that binary interaction can be a possible explanation for the ejection of the torus. A prediction of this would be the existence of a companion with an orbital period $P\lsim
1\,$month.
The only butterfly-type nebula with a known binary nucleus is the above-mentioned NGC2346, with a period of 16 days. This object also has a CO ring, albeit with considerable lower mass than seen in NGC6302. It has an unusal period: almost all other known planetary nebulae with close binary central stars show much shorter periods, requiring common envelope evolution (Zijlstra 2006). A period of 16days requires binary interaction (as the final orbit is smaller than the AGB star) whilst avoiding common envelope evolution which would lead to shorter orbits. However, whether this is common for the butterfly nebulae is not known.
Summary
=======
High angular resolution observations of $^{12}$CO and $^{13}$CO have detected a massive torus centred on the exciting star. The observations indicate a lower limit of 15 for the abundance ratio $^{12}$CO/$^{13}$CO and so by implication, $^{12}$C/$^{13}$C. A detailed analysis of the emission in the and transitions of $^{12}$CO and $^{13}$CO indicates that the emission arises from material at two different temperatures. In total the torus contains $\sim2$ M$_\odot \pm
1$ M$_{\odot}$ of material, a very similar mass to that previously derived from observations of the dust continuum emission from the source. From our best fit model, this torus is expanding with a velocity of 8 which leads to a dynamical age of the torus of 7500 yr, with the inner edge of the torus being ejected $\sim2900$ years ago. This ejection may have coincided with the eruptive event proposed to have formed the systems bipolar lobes. We speculate that the ejection of the inner edge of the torus also terminated the fast CO outflow which, while typical in the PPN progenitors of PN, is not detected in NGC 6302. The current mass within the PN implies that the progenitor of the system had a mass of $>3$M$_\odot$ and during the formation of the torus sustained a mass loss rate of $5\times10^{-4}\,\rm M_\odot\,yr^{-1}$
The derived mass-loss history is difficult to fit using mass loss driven by radiation pressure. We find that binary interaction is a plausible mechanism leading to the ejection of the torus.
Ultimately the torus is likely to become photoionised by the central star, as is happening to the CO shell around NGC7027. The photon dominated region where the UV from the star penetrates the CO torus, which is traced by the warm CO component observed here and the low excitation atomic lines observed by Castro-Carrizo et al. (2001), is likely to support a rich and complex chemistry, making NGC6302 an interesting target for further molecular line observations.
Bachiller, R., Planesas, P., Martin-Pintado, J., Bujarrabal, V., Tafalla, M. 1989. , 210, 366
Balick, B. 1987, , 94, 671 Balick, B., Preston, H. L., & Icke, V. 1987, , 94, 1641 Balser, D. S., McMullin, J. P., & Wilson, T. L. 2002, , 572, 326
Beckwith, S. V. W., Sargent, A. I., Chini, R. S., & Guesten, R. 1990, , 99, 924
Bloecker, T. 1995, , 297, 727
Bujarrabal, V., Castro-Carrizo, A., Alcolea, J., & S[á]{}nchez Contreras, C. 2001, , 377, 868
Bujarrabal, V., Castro-Carrizo, A., Alcolea, J., & Neri, R. 2005, , 441, 1031
Casassus, S., Roche, P. F., & Barlow, M. J. 2000, , 314, 657
Castro-Carrizo, A., Bujarrabal, V., Fong, D., Meixner, M., Tielens, A. G. G. M., Latter, W. B., & Barlow, M. J. 2001, , 367, 674
Corradi, R. L. M., & Schwarz, H. E. 1995, , 293, 871
Goldsmith, P. F., & Langer, W. D. 1999, , 517, 209
Feibelman, W. A. 2001, , 550, 785 Frank, A., & Mellema, G. 1994, , 289, 937 Gomez, Y., Rodriguez, L. F., Moran, J. M., & Garay, G. 1989, , 345, 862
Habing, H. J., Tignon, J., Tielens, A. G. G. M., 1994, A&A, 286, 523
Huggins, P. J., & Healy, A. P. 1989, , 346, 201
Huggins, P. J., Bachiller, R., Cox, P., & Forveille, T. 1996, , 315, 284
Huggins, P. J., Forveille, T., Bachiller, R., & Cox, P., & 2000, , 544, 889
Huggins, P. J., Bachiller, R., Planesas, P.;, Forveille, T., & Cox, P., 2005, ApJS, 160, 272
Icke V., 2003, A&A, 405, L11
Josselin, E., & L[è]{}bre, A. 2001, , 367, 826
Kemper, F., Molster, F. J., J[ä]{}ger, C., & Waters, L. B. F. M. 2002, , 394, 679 Kwok, S., Purton, C. R., & Fitzgerald, P. M. 1978, , 219, L125
Levreault, R. M. 1988, , 67, 283
Manchado, A. 1997, IAU Symp. 180: Planetary Nebulae, 180, 184
Masson, C. R., et al. , 292, 464
Matsuura, M., Zijlstra, A. A., Molster, F. J., Waters, L. B. F. M., Nomura, H., Sahai, R., & Hoare, M. G. 2005a, , 359, 383
Matsuura, M., Zijlstra, A. A., Gray, M. D., Molster, F. J., Waters, L. B. F. M., 2005b, , 363, 628
Meaburn, J., L[ó]{}pez, J. A., Steffen, W., Graham, M. F., & Holloway, A. J. 2005, , 130, 2303
Myers, P. C., Linke, R. A., & Benson, P. J. 1983, , 264, 517
Pauldrach, A., Puls, J., Kudritzki, R. P., Mendez, R. H., & Heap, S. R. 1988, , 207, 123
Payne, H. E., Phillips, J. A., & Terzian, Y. 1988, , 326, 368
Siess, L., & Livio, M. 1999, , 304, 925
Stanghellini, L., Villaver, E., Manchado, A., & Guerrero, M. A. 2002, , 576, 285
Schöier, F.L., van der Tak, F.F.S., van Dishoeck, E.F., & Black, J.H. 2005, A&A 432, 369
Soker, N. 1998a, , 496, 833
Soker, N. 1998b, , 299, 1242
van Winckel, H., 2002, ARA&A, 41, 391
Vassiliadis, E., & Wood, P. R. 1993, , 413, 641
Woods, P. M., Nyman, L.-�., Schöier, F. L., Zijlstra, A. A., Millar, T. J., & Olofsson, H., 2005, A&A, 429, 977
Zijlstra, A. A., Chapman, J. M., te Lintel Hekkert, P., Likkel, L., Comeron, F., Norris, R. P., Molster, F. J., & Cohen, R. J. 2001, , 322, 280
Zijlstra A. A., 2006, in ’Evolution and chemistry of symbiotic stars and related objects’, Baltic Astronomy, in press (astro-ph/0610558)
[^1]: The James Clerk Maxwell Telescope is operated by The Joint Astronomy Centre on behalf of the Particle Physics and Astronomy Research Council of the United Kingdom, the Netherlands Organisation for Scientific Research, and the National Research Council of Canada.
[^2]: The Submillimeter Array is a joint project between the Smithsonian Astrophysical Observatory and the Academia Sinica Institute of Astronomy and Astrophysics, and is funded by the Smithsonian Institution and the Academia Sinica.
[^3]: We have applied a conversation factor of 0.8 K/Jy, which takes into account the frequency and the size of the beam, to convert the SMA line observations to a main beam temperature scale.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We consider the Cauchy problem for the evolutive discrete $p$-Laplacian in infinite graphs, with initial data decaying at infinity. We prove optimal sup and gradient bounds for nonnegative solutions, when the initial data has finite mass, and also sharp evaluation for the confinement of mass, i.e., the effective speed of propagation. We provide estimates for some moments of the solution, defined using the distance from a given vertex.
Our technique relies on suitable inequalities of Faber-Krahn type, and looks at the local theory of continuous nonlinear partial differential equations. As it is known, however, not all of this approach can have a direct counterpart in graphs. A basic tool here is a result connecting the supremum of the solution at a given positive time with the measure of its level sets at previous times.
We also consider the case of slowly decaying initial data, where the total mass is infinite.
address:
- |
Department of Basic and Applied Sciences for Engineering\
Sapienza University of Rome, Italy
- |
South Mathematical Institute of VSC RAS\
Vladikavkaz, Russian Federation
author:
- Daniele Andreucci
- 'Anatoli F. Tedeev'
bibliography:
- 'paraboli.bib'
- 'pubblicazioni\_andreucci.bib'
title: |
Asymptotic estimates for the $p$-Laplacian on infinite graphs\
with decaying initial data
---
\[1994/06/01\]
[^1]
[^2]
Introduction {#s:intro}
============
We consider nonnegative solutions to the Cauchy problem for discrete degenerate parabolic equations $$\begin{aligned}
{2}
\label{eq:pde}
{\frac{\partial {u}}{\partial t}}(x,t)
-
\operatorname{\Delta}_{p}
{u}(x,t)
&=
0
\,,
&\qquad&
x\in V\,, t>0
\,,
\\
\label{eq:init}
{u}(x,0)
&=
{u}_{0}(x)
\ge 0
\,,
&\qquad&
x\in V
\,.\end{aligned}$$ Here $V$ is the set of vertices of the graph $G(V,E)$ with edge set $E\subset V\times V$ and weight ${\omega}$, and $$\operatorname{\Delta}_{p} u(x,t)
=
\frac{1}{{d_{{\omega}}}(x)}
\sum_{y\in V}
{\lvert{u}(y)-{u}(x)\rvert}^{p-2}
({u}(y)-{u}(x))
{\omega}(x,y)
\,.$$ We assume that the graph $G$ is simple, undirected, infinite, connected with locally finite degree $${d_{{\omega}}}(x)
=
\sum_{y\sim x}
{\omega}(x,y)
\,,$$ where we write $y\sim x$ if and only if $\{x,y\}\in E$. Here the weight ${\omega}:V\times V\to [0,+\infty)$ is symmetric, i.e., ${\omega}(x,y)={\omega}(y,x)$, and is strictly positive if and only if $y\sim x$; then ${\omega}(x,x)=0$ for $x\in V$.
We assume also that $p>2$ and that ${u}_{0}$ is nonnegative; further assumptions on ${u}_{0}$ will be stated below.
We prove sharp sup bounds for large times of solutions corresponding to finite mass initial data; in order to prove the bound from below we find an optimal estimate for the effective speed of propagation of mass. We also determine the stabilization rate for data exhibiting slow decay ‘at infinity’, in a suitable sense. To the best of our knowledge such results are new in the framework of discrete nonlinear diffusion equations on graphs.\
We apply an alternative approach, more local than the one in [@Mugnolo:2013], [@Hua:Mugnolo:2015] where the global arguments of semigroup theory are extended to graphs, actually in a more general setting which is out of the scope of this paper. We comment below in the Introduction on the inherent difficulty and even partial unfeasibility of a local approach in graphs. It is therefore an interesting and not trivial problem to understand how much of this body of techniques can be used in this environment. This paper can be seen as a cross section of this effort; specifically we look at the interplay between spread of mass and sup estimates, following ideas coming from the theory of continuous partial differential equations, with the differences required by the discrete character of graphs.
We recall the following notation: for any $R\in {{\boldsymbol}{N}}$, we let $$B_{R}(x_{0})
=
\{
x\in V
\mid
d(x,x_{0})
\le R
\}
\,.$$ Here $d$ is the standard combinatorial distance in $G$ so that $d$ only takes integral values. For any $f:V\to {{\boldsymbol}{R}}$ we set for all $q\ge 1$, $U\subset V$ $$\begin{gathered}
{{\lVertf\rVert}_{\ell^{q}(U)}}^{q}
=
\sum_{x\in U}
{\lvertf(x)\rvert}^{q}
{d_{{\omega}}}(x)
\,,
\quad
{{\lVertf\rVert}_{\ell^{\infty}(U)}}
=
\sup_{x\in U} {\lvertf(x)\rvert}
\,,
\\
{\mu_{{\omega}}}(U)
=
\sum_{x\in U}
{d_{{\omega}}}(x)
\,.\end{gathered}$$ All the infinite sums in this paper are absolutely convergent. In the following we always assume, unless explicitly noted, that all balls are centered at a given fixed $x_{0}\in V$ and we write $B_{R}(x_{0})=B_{R}$. We denote generic constants depending on the parameters of the problem by $\gamma$ (large constants), $\gamma_{0}$ (small constants). We also set for all $A\subset V$ $$\chi_{A}(x)
=
1
\,,
\quad
x\in A
\,;
\qquad
\chi_{A}(x)
=
0
\,,
\quad
x\not\in A
\,.$$
\[d:fk\] We say that $G$ satisfies a global Faber-Krahn inequality for a given $p>1$ and function ${\Lambda_p}:(0,+\infty)\to(0,+\infty)$ if for any $v>0$ and any finite subset $U\subset V$ with ${\mu_{{\omega}}}(U)=v$ we have $$\label{eq:fk}
{\Lambda_p}(v)
\sum_{x\in U}
{\lvertf(x)\rvert}^{p}
{d_{{\omega}}}(x)
\le
\sum_{x,y\in (U)_{1}}
{\lvertf(y)-f(x)\rvert}^{p}
{\omega}(x,y)
\,,$$ for all $f:V\to {{\boldsymbol}{R}}$ such that $f(x)=0$ if $x\not \in U$; here $$(U)_{1}
=
\{
x\in V
\mid
d(x,U)
\le 1
\}
\,.$$
We assume throughout that ${\Lambda_p}\in C(0,+\infty)$ is decreasing and that two suitable positive constants ${N}$, $\omega$ exists such that $$\begin{aligned}
\label{eq:fkf_nd}
v&\mapsto
{\Lambda_p}(v)^{-1}
v^{-\frac{p}{{N}}}
\,,
\quad
v>0
\,,
\quad
\text{is nondecreasing;}
\\
\label{eq:fkf_ni}
v&\mapsto
{\Lambda_p}(v)^{-1}
v^{-\omega}
\,,
\quad
v>0
\,,
\quad
\text{is nonincreasing.}\end{aligned}$$ An important class of functions in our approach is given by $$\label{eq:dcf_def}
{\psi}_{r}(s)
=
s^{\frac{p-2}{r}}
{\Lambda_p}(s^{-1})
\,,
\qquad
s>0
\,,$$ for each fixed $r\ge 1$. They, or more exactly their inverses, give the correct time-space scaling for the equation , see for example Theorems \[t:l1\] and \[p:bbl\] below.
If we make the additional assumption that for some constant $c>0$ $$\label{eq:fkf_bbl}
{\Lambda_p}(v)
\ge
c
{\mathcal{R}}(v)^{-p}
\,,
\qquad
v>0
\,,$$ where ${\mathcal{R}}:(0,+\infty)\to (0,+\infty)$ is such that ${\mu_{{\omega}}}(B_{{\mathcal{R}}(v)})=v$, we may connect ${\psi}_{1}$ to the measure of a ball in $G$. This in turn allows us to prove sharpness of our $\ell^{1}$–$\ell^{\infty}$ estimate. Property is rather natural. For instance it is known to hold for the explicit examples in Subsection \[s:examples\], to which we refer for implementations of our results in some concrete relevant cases.
\[r:spdim\] The constant ${N}$ in has no intrinsic meaning in this paper, and it is employed here only with the purpose of making easier the comparison with the case of standard regular graphs ${{\boldsymbol}{Z}}^{{N}}$, where ${\Lambda_p}(v)=\gamma_{0}v^{-p/{N}}$, see Subsection \[s:examples\].
\[r:fk\_below\] Let $x\in V$ and choose $U=\{x\}$, $f=\chi_{U}$ in , which then yields $$\label{eq:rem_fk}
{\Lambda_p}({d_{{\omega}}}(x))
{d_{{\omega}}}(x)
\le
2{d_{{\omega}}}(x)
\,.$$ Since ${\Lambda_p}$ is decreasing by assumption we infer $$\label{eq:rem_fk_bound}
{d_{{\omega}}}(x)\ge {\Lambda_p}^{(-1)}(2)
\,.$$ A remark in this connection is perhaps in order: clearly according to its definition the Faber-Krahn function ${\Lambda_p}(v)$ is defined for uniformly positive $v$ according to , so that , should be assumed for such $v$. Aiming at a technically streamlined framework, we extend ${\Lambda_p}$ to all $v>0$, while easily preserving the latter assumptions. However one can check that for large times, ${\Lambda_p}$ is evaluated at large arguments in our results, which are thus independent of this extension.
\[r:lp\_scale\] A consequence of is that any bound in $\ell^{q}(V)$ yields immediately a uniform pointwise bound: if ${v}\in\ell^{q}(V)$, $$\label{eq:lp_linf}
{\lvert{v}(z)\rvert}^{q}
\le
{\lvert{v}(z)\rvert}^{q}\frac{{d_{{\omega}}}(z)}{{\Lambda_p}^{(-1)}(2)}
\le
\frac{1}{{\Lambda_p}^{(-1)}(2)}
{{\lVert{v}\rVert}_{\ell^{q}(V)}}^{q}
\,,
\quad
z\in V
\,.$$ In turn this implies that $\ell^{p}(V)\subset\ell^{q}(V)$ if $p<q$, since $$\sum_{x\in V}
{\lvertf(x)\rvert}^{q}
{d_{{\omega}}}(x)
\le
M^{q-p}
\sum_{x\in V}
{\lvertf(x)\rvert}^{p}
{d_{{\omega}}}(x)
\,,$$ for a suitable $M$ as in .
\[d:sol\] We say that ${u}\in L^{\infty}(0,T; \ell^{r}(V))$ is a solution to if ${u}(x)\in C^{1}([0,T])$ for every $x\in V$ and ${u}$ satisfies in the classical pointwise sense.\
A solution to – also is required to take the initial data prescribed by , for each $x\in V$.
We refer the reader to [@Hua:Mugnolo:2015] for existence and uniqueness of solutions. To make this paper more self-contained however we sketch in Section \[s:prelim\] a proof of existence in Proposition \[p:ex\] (in $\ell^{q}$, $q>1$, see Theorem \[t:l1\] for $q=1$), and of uniqueness via comparison in Proposition \[p:compare\].
Our first two results are typical of the local approach we pursue. All solutions we consider below are nonnegative.
\[p:linf\_meas\] Let ${u}:V\to {{\boldsymbol}{R}}$ be a solution to , with ${u}\in L^{\infty}(0,T;\ell^{r}(V))$ for some $r\ge 1$. Then for all $x\in V$, $0<t<T$ $$\label{eq:linf_meas_n}
{u}(x,t)
\le
k
\,,$$ provided $k>0$ satisfies for a suitable $\gamma_{0}(p,{N})$ $$\label{eq:linf_meas_nn}
k^{-1}
t^{-\frac{1}{p-2}}
{\Lambda_p}\Big(\sup_{\frac{t}{4}<\tau<t}{\mu_{{\omega}}}(\{x\in V\mid {u}(x,\tau)> {k}/{2}\})\Big)^{-\frac{1}{p-2}}
\le
\gamma_{0}
\,.$$
\[co:linf\_int\] Under the assumptions in Proposition \[p:linf\_meas\], we have $$\label{eq:linf_int_m}
{u}(x,t)
\le
\gamma
\sup_{0<\tau<t}{{\lVert{u}(\tau)\rVert}_{\ell^{r}(V)}}
\big[{\psi}_{r}^{(-1)}
\big(
t^{-1}
\sup_{0<\tau<t}{{\lVert{u}(\tau)\rVert}_{\ell^{r}(V)}}^{-(p-2)}
\big)
\big]^{\frac{1}{r}}
\,,$$ for all $x\in V$, $0<t<T$. Here ${\psi}_{r}^{(-1)}$ is the inverse function of ${\psi}_{r}$ as defined in .
\[r:dcf\] One can check easily using the fact that ${\Lambda_p}$ is nonincreasing that $$a\mapsto a{\psi}_{r}^{(-1)}(sa^{-(p-2)})^{\frac{1}{r}}$$ is nondecreasing in $a>0$ for each fixed $s>0$.
Next Theorem follows directly from the estimates we stated above. Note that conservation of mass in was proved also in [@Hua:Mugnolo:2015], while the other estimates are new, as far as we know.
\[t:l1\] Let ${u}_{0}\in\ell^{1}(V)$, ${u}_{0}\ge 0$. Then problem – has a unique solution satisfying for all $t>0$ $$\begin{aligned}
\label{eq:l1_n}
{{\lVert{u}(t)\rVert}_{\ell^{1}(V)}}
&=
{{\lVert{u}_{0}\rVert}_{\ell^{1}(V)}}
\,,
\\
\label{eq:l1_nn}
{{\lVert{u}(t)\rVert}_{\ell^{\infty}(V)}}
&\le
\gamma
{{\lVert{u}_{0}\rVert}_{\ell^{1}(V)}}
{\psi}_{1}^{(-1)}
\big(
t^{-1}
{{\lVert{u}_{0}\rVert}_{\ell^{1}(V)}}^{-(p-2)}
\big)
\,.
\end{aligned}$$ In addition ${u}$ satisfies $$\begin{gathered}
\label{eq:l1_nnn}
\int_{0}^{t}
\sum_{x,y\in V}
{\lvert{D}_{y}{u}(x,\tau)\rvert}^{p-1}
{\omega}(x,y)
{\,\textup{\textmd{d}}}\tau
\\
\le
\gamma
t^{\frac{1}{p}}
{{\lVert{u}_{0}\rVert}_{\ell^{1}(V)}}^{\frac{2(p-1)}{p}}
{\psi}_{1}^{(-1)}(t^{-1}{{\lVert{u}_{0}\rVert}_{\ell^{1}(V)}}^{-(p-2)})^{\frac{p-2}{p}}
\,.
\end{gathered}$$
\[r:lp\_est\] We notice that one could exploit , to derive trivially a bound of the integral in . This is due of course to the fact that the $p$-laplacian in our setting is discrete, and it would not be possible in the framework of continuous partial differential equations.
Such a bound however is not sharp, and for example could not be used in the proof of Theorem \[p:bbl\].
In other instances where optimality is not needed we exploit a device similar to the one just described, relying on Remark \[r:lp\_scale\]; see for example the proof of Lemma \[l:cacc2\].
So far our extension to graphs of methods and results of continuous differential equations has been successful. However, in the latter setting a standard device to prove optimality of the bound in relies on the property of finite speed of propagation (i.e., solutions with initially bounded support keep this feature for all $t>0$). In the setting of graphs this property strikingly fails, as shown in [@Hua:Mugnolo:2015]. As a technical but perhaps worthwile side remark, we note that all the main ingredients in the proof of finite speed of propagation (see [@Andreucci:Tedeev:1999], [@Andreucci:Tedeev:2000]) seem to be available in graphs too: embeddings as in [@Ostrovskii:2005], Caccioppoli inequalities as in Lemma \[l:cacc\] below, and of course iterative techniques as the one displayed in the proof of Proposition \[p:linf\_meas\]. The key exception in this regard is the fact that full localization via an infinite sequence of nested shrinking balls is clearly prohibited by the discrete metric at hand. This is a point of marked difference with the continuous setting.
Still we can prove sharpness of our $\ell^{1}$–$\ell^{\infty}$ bound by means of the following result of confinement of mass. By the same argument we can estimate also a suitable moment of the solution, which is also a new result for nonlinear diffusion in graphs, see Section \[s:bbl\].
\[p:bbl\] Let ${u}_{0}\ge 0$ be finitely supported. Then for every $1>{\varepsilon}>0$ there exists a $\varGamma>0$ such that $$\label{eq:bbl_n}
{{\lVert{u}(t)\rVert}_{\ell^{1}(B_{R})}}
\ge
(1-{\varepsilon})
{{\lVert{u}_{0}\rVert}_{\ell^{1}(V)}}
\,,
\qquad
t>0
\,,$$ provided $B_{{\lfloorR/2\rfloor}}$ contains the support of ${u}_{0}$, and $R$ is chosen so that $$\label{eq:bbl_nn}
R
\ge
\varGamma
t^{\frac{1}{p}}
{{\lVert{u}_{0}\rVert}_{\ell^{1}(V)}}^{\frac{p-2}{p}}
{\psi}_{1}^{(-1)}(t^{-1}{{\lVert{u}_{0}\rVert}_{\ell^{1}(V)}}^{-(p-2)})^{\frac{p-2}{p}}
\ge 8
\,.$$ In addition, provided $R$ is chosen as in , for ${\varepsilon}=1/2$, and $\alpha\in (0,1)$, $$\label{eq:bbl_p}
\sum_{x\in V}
d(x,x_{0})^{\alpha}
{u}(x,t)
{d_{{\omega}}}(x)
\le
\gamma
R^{\alpha}
{{\lVert{u}_{0}\rVert}_{\ell^{1}(V)}}
\,,
\qquad
t>0
\,.$$
Next we exploit the estimate – in order to show that up to a change in the constant we can reverse the inequality in , proving at once the optimality of both results.
\[p:bbl2\] Under the assumptions in Theorem \[p:bbl\], let in addition ${\Lambda_p}$ satisfy . Then $$\label{eq:bbl_nnn}
{{\lVert{u}(t)\rVert}_{\ell^{\infty}(V)}}
\ge
\frac{{{\lVert{u}_{0}\rVert}_{\ell^{1}(V)}}}{2{\mu_{{\omega}}}(B_{R})}
\ge
\gamma_{0}
{{\lVert{u}_{0}\rVert}_{\ell^{1}(V)}}
{\psi}_{1}^{(-1)}
\big(
t^{-1}
{{\lVert{u}_{0}\rVert}_{\ell^{1}(V)}}^{-(p-2)}
\big)
\,,$$ where $R$ is as in , for ${\varepsilon}=1/2$.
Clearly, owing to the comparison principle of Proposition \[p:compare\], results like those in and may be proved even dropping the assumption that ${u}_{0}$ is finitely supported; for the sake of brevity we omit the details.
In order to state our last result we need to introduce the following function, which essentially gives the correct scaling between time and space in the case of slow decay initial data: for ${u}_{0}\in\ell^{q}(V)\setminus\ell^{1}(V)$ for some $q>1$ set $$\label{eq:decay_fn}
{T}_{{u}_{0}}(R,x_{0})
=
\Big[
\frac{{{\lVert{u}_{0}\rVert}_{\ell^{1}(B_{R}(x_{0}))}}}{{{\lVert{u}_{0}\rVert}_{\ell^{q}(V\setminus B_{R}(x_{0}))}}^{q}}
\Big]^{\frac{p-2}{q-1}}
\,
{\Lambda_p}\bigg(
\Big(
\frac{
{{\lVert{u}_{0}\rVert}_{\ell^{1}(B_{R}(x_{0}))}}
}{
{{\lVert{u}_{0}\rVert}_{\ell^{q}(V\setminus B_{R}(x_{0}))}}
}
\Big)^{\frac{q}{q-1}}
\bigg)^{-1}
\,,$$ for $R\in {{\boldsymbol}{N}}$, $x_{0}\in V$. Clearly for each fixed $x_{0}$ the function ${T}_{{u}_{0}}$ is nondecreasing in $R$ and ${T}_{{u}_{0}}(R,x_{0})\to +\infty$ as $R\to\infty$. Conversely, ${T}_{{u}_{0}}(0,x_{0})$ may be positive. However it can be easily seen that for any given ${\varepsilon}>0$ there exists $x_{0}$ such that ${T}_{{u}_{0}}(0,x_{0})<{\varepsilon}$.
\[t:decay\] Let ${u}_{0}\in\ell^{q}(V)\setminus\ell^{1}(V)$ for some $q>1$. Then for all $t>0$, $x_{0}\in V$ $$\label{eq:decay_n}
{{\lVert{u}(t)\rVert}_{\ell^{\infty}(V)}}
\le
\gamma
{{\lVert{u}_{0}\rVert}_{\ell^{1}(B_{R}(x_{0}))}}
{\psi}_{1}^{(-1)}\big(t^{-1} {{\lVert{u}_{0}\rVert}_{\ell^{1}(B_{R}(x_{0}))}}^{-(p-2)}\big)
\,,$$ provided $R$ is chosen so that $$\label{eq:decay_nn}
t\le
{T}_{{u}_{0}}(R,x_{0})
\,,$$ the optimal choice being of course the minimum $R=R(t)$ such that holds true.
Let us comment briefly on the existing literature on the non-linear $p$-Laplacian in graphs. The papers [@Mugnolo:2013], [@Hua:Mugnolo:2015], deal with the Cauchy problem applying techniques inspired from the theory of semigroups of continuous differential operators. They consider a more general variety of weighted graphs and operators than we do here, dealing e.g., with existence, uniqueness, time regularity, possible extinction in a finite time. However our results do not seem to be easily reached by this approach. We also quote [@Keller:Mugnolo:2016] where a connection between Cheeger constants and the eigenvalues of the $p$-laplacian is drawn in a very flexible setting.\
Boundary problems on finite subgraphs are also considered in several papers dealing with features like blow up or extinction; we quote only [@Chung:Choi:2014] and [@Chung:Park:2017].
The case of the discrete linear Laplacian where $p=2$ is more classical, also for its connections with probability theory (see e.g., [@Andres:etal:2013] and references therein), and is often attacked by means of suitable parallels with the theory of heat kernels in manifolds. We quote [@Coulhon:Grigoryan:1998], [@Barlow:Coulhon:Grigoryan:2001] where a connection is drawn between properties of heat kernels, of graphs and Faber-Krahn functions.\
In [@Lin:Wu:2017] heat kernels are used to study the blow up of solutions to the Cauchy problem for a semilinear equation on a possibly infinite graph.
The subject of diffusion in graphs is popular also owing to its applicative interest. We refer the reader to [@Mugnolo:2013], [@Elmoataz:Toutain:Tenbrinck:2015] and to the references therein for more on this point.
Finally we recall the papers [@Bakry:etal:1995a], [@Ostrovskii:2005] and books [@Chung:SGT], [@Grigoryan:AG] for basic information on functional analysis on graphs and manifolds.
We mention that in our setting it is still valid the argument in [@Bonforte:Grillo:2007] showing that optimal decay rates imply suitable embeddings.
Here we look essentially at the approach of [@DiBenedetto:Herrero:1989] and [@Andreucci:Tedeev:2015].
The paper is organized as follows: Section \[s:prelim\] is devoted to preliminary material. Proposition \[p:linf\_meas\] and its Corollary \[co:linf\_int\] are proved in Section \[s:linf\], while Section \[s:l1\] contains the proof of Theorem \[t:l1\] and Section \[s:bbl\] deals with Theorem \[p:bbl\] and Corollary \[p:bbl2\]. Finally Theorem \[t:decay\] is proved in Section \[s:decay\].
Examples {#s:examples}
--------
1\) As a first example we consider the case of the standard lattice $G={{\boldsymbol}{Z}}^{{N}}$, where one can take ${\Lambda_p}(v)=\gamma_{0}v^{-p/{N}}$, according to the results of [@Wang:Wang:1977], [@Ostrovskii:2005]. This is the case where comparison with the Cauchy problem for the continuous $p$-Laplacian is more straightforward. In this case $$\label{eq:dcf_euc}
{\psi}_{r}(s)
=
\gamma_{0}
s^{\frac{{N}(p-2)+pr}{{N}r}}
\,,
\qquad
s>0
\,,$$ and for example estimate becomes $$\label{eq:l1_nn_grid}
{{\lVert{u}(t)\rVert}_{\ell^{\infty}(V)}}
\le
\gamma
{{\lVert{u}_{0}\rVert}_{\ell^{1}(V)}}^{\frac{p}{{N}(p-2)+p}}
t^{-\frac{{N}}{{N}(p-2)+p}}
\,,$$ while the critical radius for expansion of mass in amounts to $$\label{eq:bbl_nn_grid}
R\ge
\gamma
{{\lVert{u}_{0}\rVert}_{\ell^{1}(V)}}^{\frac{p-2}{{N}(p-2)+p}}
t^{\frac{1}{{N}(p-2)+p}}
\,.$$ We remark that both results formally coincide with the corresponding ones for the continuous $p$-Laplacian in ${{{\boldsymbol}{R}}^{{N}}}$, see [@DiBenedetto:Herrero:1989].\
Next we apply Theorem \[t:decay\] to the following initial data: for $x=(x_{1},\dots,x_{{N}})\in{{\boldsymbol}{Z}}^{{N}}$ set ${u}_{0}(x)=({\lvertx_{1}\rvert}+\dots+{\lvertx_{{N}}\rvert})^{-\alpha}$ for a given $0<\alpha<{N}$. Let us write here $a(s)\simeq b(s)$ if $\gamma_{0}a(s)\le b(s)\le \gamma a(s)$ for two constants independent of $s$. One can see that $${{\lVert{u}_{0}\rVert}_{\ell^{1}(B_{R}(0))}}
\simeq
R^{{N}-\alpha}
\,;
\qquad
{{\lVert{u}_{0}\rVert}_{\ell^{q}(V\setminus B_{R}(0))}}
\simeq
R^{{N}-\alpha q}
\,,$$ for all $q>{N}/\alpha$. Therefore in this case $${T}_{{u}_{0}}(R,0)
\simeq
R^{\alpha(p-2)+p}
\,,$$ and the estimate in essentially amounts to the decay rate $t^{-\alpha/(\alpha(p-2)+p)}$, which is the expected one in view of the results of [@Tedeev:1991].
2\) One can treat also other examples of product graphs; for instance if $H$ is a finite connected graph we let $G=H\times {{\boldsymbol}{Z}}^{{N}}$ and recover results similar to the ones of the previous example.
3\) All examples where the Faber-Krahn function is estimated for $p=2$ yield also examples in our case of $p>2$, as it follows from applying Hölder’s inequality; see e.g., [@Coulhon:Grigoryan:1998], [@Barlow:Coulhon:Grigoryan:2001].
Preliminary material {#s:prelim}
====================
We use for $f:V\to {{\boldsymbol}{R}}$ the notation $${D}_{y}f(x)
=
f(y)-f(x)
=
-
{D}_{x}f(y)
\,,
\qquad
x\,,y\in V
\,.$$
Caccioppoli type inequalities {#s:prelim_cacc}
-----------------------------
\[l:monot\] Let $q>0$, $p>2$, $h\ge 0$, ${u}$, ${v}:V\to{{\boldsymbol}{R}}$. Then for all $x$, $y\in V$ $$\begin{gathered}
\label{eq:monot_n}
\big(
{\lvert{D}_{y}{u}(x)\rvert}^{p-2}
{D}_{y}{u}(x)
-
{\lvert{D}_{y}{v}(x)\rvert}^{p-2}
{D}_{y}{v}(x)
\big)
{D}_{y}{{\ifthenelse{\equal{*}{*}}{({u}(x)-{v}(x)-h)_{+}}{\left({u}(x)-{v}(x)-h\right)_{+}}}}^{q}
\\
\ge
\gamma_{0}
{\left|
{D}_{y}{{\ifthenelse{\equal{*}{*}}{({u}(x)-{v}(x)-h)_{+}}{\left({u}(x)-{v}(x)-h\right)_{+}}}}^{\frac{q-1+p}{p}}
\right|}^{p}
\,.
\end{gathered}$$
First we remark that we may assume $h=0$, by renaming $\tilde{v}={v}+h$. The corresponding version of clearly holds true if ${D}_{y}{u}(x)={D}_{y}{v}(x)$.
If ${D}_{y}{u}(x)\not={D}_{y}{v}(x)$ the left hand side of with $h=0$ can be written as, on appealing also to a classical elementary result in monotone operators, see [@DiB:dpe], $$\begin{gathered}
\label{eq:monot_i}
\big(
{\lvert{D}_{y}{u}(x)\rvert}^{p-2}
{D}_{y}{u}(x)
-
{\lvert{D}_{y}{v}(x)\rvert}^{p-2}
{D}_{y}{v}(x)
\big)
{D}_{y}({u}(x)-{v}(x))
\,
\mathcal{A}
\\
\ge
\gamma_{0}(p)
{\lvert{D}_{y}({u}(x)-{v}(x))\rvert}^{p}
\,
\mathcal{A}
\,,
\end{gathered}$$ where we define $$\mathcal{A}
=
\frac{
{D}_{y}{{\ifthenelse{\equal{*}{*}}{({u}(x)-{v}(x))_{+}}{\left({u}(x)-{v}(x)\right)_{+}}}}^{q}
}{
{D}_{y}({u}(x)-{v}(x))
}
\ge 0
\,.$$
On the other hand, we write the right hand side of with $h=0$ as $$\label{eq:monot_ii}
{\lvert{D}_{y} ({u}(x)-{v}(x))\rvert}^{p}
\,
\mathcal{B}
\,,
\qquad
\mathcal{B}
:=
{\left|
\frac{
{D}_{y}{{\ifthenelse{\equal{*}{*}}{({u}(x)-{v}(x))_{+}}{\left({u}(x)-{v}(x)\right)_{+}}}}^{\frac{q-1+p}{p}}
}{
{D}_{y}({u}(x)-{v}(x))
}
\right|}^{p}
\,.$$ Therefore we have only to prove that $\mathcal{A}\ge \gamma_{0}\mathcal{B}$. Clearly in doing so we may assume without loss of generality that $${u}(y)
-
{v}(y)
>
{u}(x)
-
{v}(x)
\,.$$ Hence it is left to prove that $$\begin{gathered}
\label{eq:monot_iii}
\big[
{u}(y)-{v}(y)
-
({u}(x)-{v}(x))
\big]^{p-1}
\big[
{{\ifthenelse{\equal{*}{*}}{({u}(y)-{v}(y))_{+}}{\left({u}(y)-{v}(y)\right)_{+}}}}^{q}
-
{{\ifthenelse{\equal{*}{*}}{({u}(x)-{v}(x))_{+}}{\left({u}(x)-{v}(x)\right)_{+}}}}^{q}
\big]
\\
\ge
\gamma_{0}
\Big[
{{\ifthenelse{\equal{*}{*}}{({u}(y)-{v}(y))_{+}}{\left({u}(y)-{v}(y)\right)_{+}}}}^{\frac{q-1+p}{p}}
-
{{\ifthenelse{\equal{*}{*}}{({u}(x)-{v}(x))_{+}}{\left({u}(x)-{v}(x)\right)_{+}}}}^{\frac{q-1+p}{p}}
\Big]^{p}
\,.
\end{gathered}$$ Denote $$a
=
{u}(y)
-
{v}(y)
\,,
\qquad
b
=
{u}(x)
-
{v}(x)
\,.$$ If $b\le 0$, is obviously satisfied with $\gamma_{0}=1$. If $b>0$, by Hölder’s inequality we have $$\begin{gathered}
\label{eq:monot_iv}
\big[
a^{\frac{q-1+p}{p}}
-
b^{\frac{q-1+p}{p}}
\big]^{p}
=
\Big[
\frac{q-1+p}{p}
\int_{b}^{a}
s^{\frac{q-1}{p}}
{\,\textup{\textmd{d}}}s
\Big]^{p}
\\
\le
\gamma(q,p)
\Big[
\int_{b}^{a}
s^{q-1}
{\,\textup{\textmd{d}}}s
\Big]
\Big[
\int_{b}^{a}
{\,\textup{\textmd{d}}}s
\Big]^{p-1}
\le
\gamma(q,p)
(a^{q}-b^{q})
(a-b)^{p-1}
\,,
\end{gathered}$$ proving and concluding the proof.
In the following all radii of balls in $G$ will be assumed to be natural numbers. Let $R_{2}\ge R_{1}+1$, $R_{1}$, $R_{2}>0$; we define the cutoff function $\zeta$ in $B_{R_{2}}(x_{0})$ by means of $$\begin{aligned}
{2}
\zeta(x)
&=
1
\,,
&\qquad&
x\in B_{R_{1}}(x_{0})
\,,
\\
\zeta(x)
&=
\frac{
R_{2}
-
d(x,x_{0})
}{
R_{2}
-
R_{1}
}
\,,
&\qquad&
x\in B_{R_{2}}\setminus B_{R_{1}}(x_{0})
\,,
\\
\zeta(x)
&=
0
\,,
&\qquad&
x\not \in B_{R_{2}}(x_{0})
\,.\end{aligned}$$ The function $\zeta$ is chosen so that $${\lvert{D}_{y}\zeta(x)\rvert}
=
{\lvert
\zeta(y)
-
\zeta(x)
\rvert}
\le
\frac{1}{R_{2}-R_{1}}
\,,
\qquad
x\sim y
\,.$$ For $\tau_{1}>\tau_{2}>0$ we also define the standard nonnegative cutoff function $\eta\in C^{1}({{\boldsymbol}{R}})$ such that $$\eta(t)
=0
\,,
\,\,\,
t\ge \tau_{1}
\,;
\quad
\eta(t)
=
0
\,,
\,\,\,
t\le \tau_{2}
\,;
\quad
0\le\eta'(t)\le \frac{2}{\tau_{1}-\tau_{2}}
\,,
\,\,\,
t\in{{\boldsymbol}{R}}\,.$$
Our next Lemma is not used in the sequel; we present it here to substantiate our claim made in the Introduction that suitable local Caccioppoli type inequalities are available in the nonlinear setting, and also for its possible independent interest. The proof is somehow more complex than in the continuous case.
\[l:cacc\] Let ${u}$ be a solution of in $V\times (0,T)$, $x_{0}\in V$. Then for $T>\tau_{1}>\tau_{2}>0$, $R_{2}>R_{1}+1$, $R_{1}>0$, $h>k>0$, $1>\theta>0$ we have $$\label{eq:cacc_n}
\begin{split}
&\sup_{\tau_{1}<\tau<t}
\sum_{x\in B_{R_{1}}(x_{0})}
{{\ifthenelse{\equal{*}{*}}{({u}(x,\tau)-h)_{+}}{\left({u}(x,\tau)-h\right)_{+}}}}^{\theta+1}
\zeta(x)^{p}
{d_{{\omega}}}(x)
\\
&\quad+
\int_{\tau_{1}}^{t}
\sum_{x\in B_{R_{1}}(x_{0}),y\in V}
{\left|
{D}_{y}
{{\ifthenelse{\equal{*}{*}}{({u}(x,\tau)-h)_{+}}{\left({u}(x,\tau)-h\right)_{+}}}}^{\frac{p+\theta-1}{p}}
\right|}^{p}
{\omega}(x,y)
{\,\textup{\textmd{d}}}\tau
\\
&\qquad\le
\frac{\gamma}{\tau_{1}-\tau_{2}}
\int_{\tau_{2}}^{t}
\sum_{x\in B_{R_{2}}(x_{0})}
{{\ifthenelse{\equal{*}{*}}{({u}(x,\tau)-h)_{+}}{\left({u}(x,\tau)-h\right)_{+}}}}^{\theta+1}
{d_{{\omega}}}(x)
{\,\textup{\textmd{d}}}\tau
+
\gamma
A^{\frac{1}{p}}
B^{\frac{p-1}{p}}
+
\gamma
A
\,,
\end{split}$$ where $$\begin{aligned}
A&=
\frac{1}{(R_{2}-R_{1})^{p}}
\int_{\tau_{2}}^{t}
\sum_{x\in B_{R_{2}}(x_{0})}
{{\ifthenelse{\equal{*}{*}}{({u}(x,\tau)-k)_{+}}{\left({u}(x,\tau)-k\right)_{+}}}}^{p+\theta-1}
{d_{{\omega}}}(x)
{\,\textup{\textmd{d}}}\tau
\,,
\\
B&=
h^{p}
(h-k)^{\theta-1}
\int_{\tau_{2}}^{t}
{\mu_{{\omega}}}(B_{R_{2}}(x_{0})\cap\{2h\ge {u}(x,\tau)>h\})
{\,\textup{\textmd{d}}}\tau
\,.
\end{aligned}$$
\[r:caccio\] The term $A^{1/p}B^{(p-1)/p}$ in can be reduced to one containing only $A$ by means of Young’s and Chebychev’s inequalities.
We multiply against $\zeta(x)^{p}\eta(t)^{p}{{\ifthenelse{\equal{*}{*}}{({u}(x,t)-h)_{+}}{\left({u}(x,t)-h\right)_{+}}}}^{\theta}$ and apply the well known formula of integration by parts $$\begin{gathered}
\sum_{x,y\in V}
{\lvert{D}_{y}{u}(x)\rvert}^{p-2}
{D}_{y}{u}(x)
f(x)
{\omega}(x,y)
\\=
-
\frac{1}{2}
\sum_{x,y\in V}
{\lvert{D}_{y}{u}(x)\rvert}^{p-2}
{D}_{y}{u}(x)
{D}_{y}f(x)
{\omega}(x,y)
\,,
\end{gathered}$$ where $f:V\to {{\boldsymbol}{R}}$ has finite support. Below we denote $B_{R}(x_{0})=B_{R}$ for simplicity of notation.
We obtain $$\begin{gathered}
\label{eq:cacc_i}
J_{1}+J_{2}
:=
\frac{1}{\theta+1}
\sum_{x\in B_{R_{2}}}
{{\ifthenelse{\equal{*}{*}}{({u}(x,t)-h)_{+}}{\left({u}(x,t)-h\right)_{+}}}}^{\theta+1}
\zeta(x)^{p}
\eta(t)^{p}
{d_{{\omega}}}(x)
\\
+
\frac{1}{2}
\int_{0}^{t}
\sum_{x,y\in V}
{\lvert{D}_{y}{u}(x,\tau)\rvert}^{p-2}
{D}_{y}{u}(x,\tau)
{D}_{y}[
{{\ifthenelse{\equal{*}{*}}{({u}(x,\tau)-h)_{+}}{\left({u}(x,\tau)-h\right)_{+}}}}^{\theta}
\zeta(x)^{p}
]
{\omega}(x,y)
\eta(\tau)^{p}
{\,\textup{\textmd{d}}}\tau
\\
=
\frac{p}{\theta+1}
\int_{0}^{t}
\sum_{x\in B_{R_{2}}}
{{\ifthenelse{\equal{*}{*}}{({u}(x,\tau)-h)_{+}}{\left({u}(x,\tau)-h\right)_{+}}}}^{\theta+1}
\zeta(x)^{p}
\eta(\tau)^{p-1}
\eta'(\tau)
{d_{{\omega}}}(x)
{\,\textup{\textmd{d}}}\tau
=:J_{3}
\,.
\end{gathered}$$ We split $J_{2}$ according to the equality $${D}_{y}[
{{\ifthenelse{\equal{*}{*}}{({u}(x,\tau)-h)_{+}}{\left({u}(x,\tau)-h\right)_{+}}}}^{\theta}
\zeta(x)^{p}
]
=
\zeta(y)^{p}
{D}_{y}
{{\ifthenelse{\equal{*}{*}}{({u}(x,\tau)-h)_{+}}{\left({u}(x,\tau)-h\right)_{+}}}}^{\theta}
+
{{\ifthenelse{\equal{*}{*}}{({u}(x,\tau)-h)_{+}}{\left({u}(x,\tau)-h\right)_{+}}}}^{\theta}
{D}_{y} \zeta(x)^{p}
\,.$$ Next we appeal to Lemma \[l:monot\] with ${v}=0$ to get $$\label{eq:cacc_iii}
{\lvert{D}_{y}{u}(x,\tau)\rvert}^{p-2}
{D}_{y}{u}(x,\tau)
{D}_{y}[
{{\ifthenelse{\equal{*}{*}}{({u}(x,\tau)-h)_{+}}{\left({u}(x,\tau)-h\right)_{+}}}}^{\theta}
]
\ge
\gamma_{0}
{\left|
{D}_{y}
{{\ifthenelse{\equal{*}{*}}{({u}(x,\tau)-h)_{+}}{\left({u}(x,\tau)-h\right)_{+}}}}^{\frac{p+\theta-1}{p}}
\right|}^{p}
\,.$$ Thus from we infer the bound $$\label{eq:cacc_iv}
J_{1}
+
J_{21}
+
J_{22}
\le
J_{3}
+
J_{23}
\,,$$ where $$\begin{aligned}
J_{21}
&=
\gamma_{0}
\int_{0}^{t}
\sum_{x,y\in V}
{\left|
{D}_{y}
{{\ifthenelse{\equal{*}{*}}{({u}(x,\tau)-h)_{+}}{\left({u}(x,\tau)-h\right)_{+}}}}^{\frac{p+\theta-1}{p}}
\right|}^{p}
\zeta(y)^{p}
{\omega}(x,y)
\eta(\tau)^{p}
{\,\textup{\textmd{d}}}\tau
\,,
\\
J_{22}
&=
\frac{1}{4}
\int_{0}^{t}
\sum_{x,y\in V}
{\lvert{D}_{y}{u}(x,\tau)\rvert}^{p-2}
{D}_{y}{u}(x,\tau)
{D}_{y}[
{{\ifthenelse{\equal{*}{*}}{({u}(x,\tau)-h)_{+}}{\left({u}(x,\tau)-h\right)_{+}}}}^{\theta}
]
\zeta(y)^{p}
{\omega}(x,y)
\eta(\tau)^{p}
{\,\textup{\textmd{d}}}\tau
\,,
\\
J_{23}
&=
\frac{1}{2}
\int_{0}^{t}
\sum_{x,y\in V}
{\lvert{D}_{y}{u}(x,\tau)\rvert}^{p-1}
{\lvert{D}_{y}\zeta(x)^{p}\rvert}
{{\ifthenelse{\equal{*}{*}}{({u}(x,\tau)-h)_{+}}{\left({u}(x,\tau)-h\right)_{+}}}}^{\theta}
\eta(\tau)^{p}
{\omega}(x,y)
{\,\textup{\textmd{d}}}\tau
\,.
\end{aligned}$$ The reason to preserve the fraction $J_{22}$ of $J_{2}$ (rather than treating it as in $J_{21}$) will become apparent presently. Let us introduce the functions $$\begin{gathered}
H(x,y;r)
=
\max[
{{\ifthenelse{\equal{*}{*}}{({u}(x,\tau)-r)_{+}}{\left({u}(x,\tau)-r\right)_{+}}}}
,
{{\ifthenelse{\equal{*}{*}}{({u}(y,\tau)-r)_{+}}{\left({u}(y,\tau)-r\right)_{+}}}}
]
\,,
\\
\chi_{x,y}
=
1
\,,
\quad
\text{if $H(x,y;h)>0$;}
\qquad
\chi_{x,y}
=
0
\,,
\quad
\text{if $H(x,y;h)=0$.}
\end{gathered}$$ Note that $r>0$ is arbitrary in the definition of $H$ but we fix $r=h$ in the definition of $\chi_{x,y}$. Next we select $0<k<h$; by elementary calculations and Young’s inequality we get $$\begin{split}
J_{23}
&\le
\frac{p}{2}
\int_{0}^{t}
\sum_{x,y\in V}
{\lvert{D}_{y} {u}(x,\tau)\rvert}^{p-1}
{\lvert{D}_{y} \zeta(x)\rvert}
(\zeta(x)+\zeta(y))^{p-1}
H(x,y;k)^{\theta}
\chi_{x,y}
{\omega}(x,y)
\eta(\tau)^{p}
{\,\textup{\textmd{d}}}\tau
\\
&\le
{\varepsilon}\int_{0}^{t}
\sum_{x,y\in V}
{\lvert{D}_{y}{u}(x,\tau)\rvert}^{p}
(\zeta(x)^{p}+\zeta(y)^{p})
H(x,y;k)^{\theta-1}
\chi_{x,y}
{\omega}(x,y)
\eta(\tau)^{p}
{\,\textup{\textmd{d}}}\tau
\\
&\quad+
\gamma {\varepsilon}^{1-p}
\int_{0}^{t}
\sum_{x,y\in V}
{\lvert{D}_{y}\zeta(x)\rvert}^{p}
H(x,y;k)^{p+\theta-1}
{\omega}(x,y)
\eta(\tau)^{p}
{\,\textup{\textmd{d}}}\tau
=:
J_{231}
+
J_{232}
\,.
\end{split}$$ We want to absorb partially the term $J_{231}$ into $J_{22}$, for a suitable choice of ${\varepsilon}$. To this end we observe that by a change of variables we have $$\begin{split}
J_{22}
&=
\frac{1}{4}
\int_{0}^{t}
\sum_{x,y\in V}
{\lvert{D}_{x}{u}(y,\tau)\rvert}^{p-1}
{\lvert{D}_{x}{{\ifthenelse{\equal{*}{*}}{({u}(y,\tau)-h)_{+}}{\left({u}(y,\tau)-h\right)_{+}}}}^{\theta}\rvert}
\zeta(x)^{p}
{\omega}(y,x)
\eta(\tau)^{p}
{\,\textup{\textmd{d}}}\tau
\\
&=
\frac{1}{4}
\int_{0}^{t}
\sum_{x,y\in V}
{\lvert{D}_{y}{u}(x,\tau)\rvert}^{p-1}
{\lvert{D}_{y}{{\ifthenelse{\equal{*}{*}}{({u}(x,\tau)-h)_{+}}{\left({u}(x,\tau)-h\right)_{+}}}}^{\theta}\rvert}
\zeta(x)^{p}
{\omega}(x,y)
\eta(\tau)^{p}
{\,\textup{\textmd{d}}}\tau
\\
&=
\frac{1}{8}
\int_{0}^{t}
\sum_{x,y\in V}
{\lvert{D}_{y}{u}(x,\tau)\rvert}^{p-1}
{\lvert{D}_{y}{{\ifthenelse{\equal{*}{*}}{({u}(x,\tau)-h)_{+}}{\left({u}(x,\tau)-h\right)_{+}}}}^{\theta}\rvert}
(\zeta(x)^{p}+\zeta(y)^{p})
{\omega}(x,y)
\eta(\tau)^{p}
{\,\textup{\textmd{d}}}\tau
\,.
\end{split}$$ Then by elementary calculus $$\begin{gathered}
\label{eq:cacc_j}
\chi_{x,y}
{\lvert{D}_{y}{{\ifthenelse{\equal{*}{*}}{({u}(x,\tau)-h)_{+}}{\left({u}(x,\tau)-h\right)_{+}}}}^{\theta}\rvert}
\ge
\chi_{x,y}
\theta
{\lvert{D}_{y}{{\ifthenelse{\equal{*}{*}}{({u}(x,\tau)-h)_{+}}{\left({u}(x,\tau)-h\right)_{+}}}}\rvert}
H(x,y;h)^{\theta-1}
\\
\ge
\chi_{x,y}
\theta
{\lvert{D}_{y}{{\ifthenelse{\equal{*}{*}}{({u}(x,\tau)-h)_{+}}{\left({u}(x,\tau)-h\right)_{+}}}}\rvert}
H(x,y;k)^{\theta-1}
\,.
\end{gathered}$$ Next we discriminate three cases in , aggregating equivalent symmetric cases: i) ${u}(x,\tau)>h$, ${u}(y,\tau)>h$. In this case clearly $${\lvert{D}_{y}{{\ifthenelse{\equal{*}{*}}{({u}(x,\tau)-h)_{+}}{\left({u}(x,\tau)-h\right)_{+}}}}\rvert}
=
{\lvert{D}_{y}{u}(x,\tau)\rvert}
\,.$$ ii) ${u}(x,\tau)>2h$, $h\ge {u}(y,\tau)$. Then $${\lvert{D}_{y}{{\ifthenelse{\equal{*}{*}}{({u}(x,\tau)-h)_{+}}{\left({u}(x,\tau)-h\right)_{+}}}}\rvert}
\ge
\frac{{u}(x,\tau)}{2}
\ge
\frac{1}{2}
{\lvert{D}_{y}{u}(x,\tau)\rvert}
\,.$$ iii) $2h \ge {u}(x,\tau)>h\ge {u}(y,\tau)$. In this case $J_{22}$ does not offer any help. We rather bound directly this part of $J_{231}$ as shown below.
Collecting the estimates above we see that, provided ${\varepsilon}\le 1/16$, $$\begin{gathered}
J_{231}
\le
J_{22}
+
{\varepsilon}2^{p+2}
h^{p}
(h-k)^{\theta-1}
\int_{\tau_{2}}^{t}
{\mu_{{\omega}}}(B_{R_{2}}\cap\{2h\ge {u}(x,\tau)>h\})
{\,\textup{\textmd{d}}}\tau
\,.
\end{gathered}$$ Hence we have transformed into $$\label{eq:cacc_jj}
J_{1}+J_{21}
\le
J_{3}
+
\gamma {\varepsilon}B
+
\gamma {\varepsilon}^{1-p}
A
\,,$$ where $A$ and $B$ are as in the statement.
Finally we check whether the root ${\varepsilon}$ of ${\varepsilon}B={\varepsilon}^{1-p}A$ is less than $1/16$; on distinguishing the cases ${\varepsilon}\le 1/16$, ${\varepsilon}>1/16$ we get the inequality in .
\[l:cacc2\] Let ${u}\in L^{\infty}(0,T;\ell^{q}(V))$, for a given $q> 1$, be a solution of in $V\times(0,T)$. Then for all $T>\tau_{1}>\tau_{2}>0$, $h\ge0$, we have for all $0<t<T$ $$\begin{gathered}
\label{eq:cacc2_n}
\sup_{\tau_{1}<\tau<t}
\sum_{x\in V}
{{\ifthenelse{\equal{*}{*}}{({u}(x,\tau)-h)_{+}}{\left({u}(x,\tau)-h\right)_{+}}}}^{q}
{d_{{\omega}}}(x)
+
\int_{\tau_{1}}^{t}
\sum_{x,y\in V}
{\left|{D}_{y}{{\ifthenelse{\equal{*}{*}}{({u}(x,\tau)-h)_{+}}{\left({u}(x,\tau)-h\right)_{+}}}}^{\frac{p+q-2}{p}}\right|}^{p}
{\omega}(x,y)
{\,\textup{\textmd{d}}}\tau
\\
\le
\frac{\gamma}{\tau_{1}-\tau_{2}}
\int_{\tau_{2}}^{t}
\sum_{x\in V}
{{\ifthenelse{\equal{*}{*}}{({u}(x,\tau)-h)_{+}}{\left({u}(x,\tau)-h\right)_{+}}}}^{q}
{d_{{\omega}}}(x)
{\,\textup{\textmd{d}}}\tau
\,.
\end{gathered}$$ We have also, if condition is satisfied, $$\begin{gathered}
\label{eq:cacc2_nn}
\sup_{0<\tau<t}
\sum_{x\in V}
{{\ifthenelse{\equal{*}{*}}{({u}(x,\tau)-h)_{+}}{\left({u}(x,\tau)-h\right)_{+}}}}^{q}
{d_{{\omega}}}(x)
+
\int_{0}^{t}
\sum_{x,y\in V}
{\left|{D}_{y}{{\ifthenelse{\equal{*}{*}}{({u}(x,\tau)-h)_{+}}{\left({u}(x,\tau)-h\right)_{+}}}}^{\frac{p+q-2}{p}}\right|}^{p}
{\omega}(x,y)
{\,\textup{\textmd{d}}}\tau
\\
\le
\gamma
\int_{0}^{t}
\sum_{x\in V}
{{\ifthenelse{\equal{*}{*}}{({u}_{0}(x)-h)_{+}}{\left({u}_{0}(x)-h\right)_{+}}}}^{q}
{d_{{\omega}}}(x)
{\,\textup{\textmd{d}}}\tau
\,.
\end{gathered}$$
Let us prove ; the inequality is proved similarly.
We multiply against $\zeta(x)\eta(t){{\ifthenelse{\equal{*}{*}}{({u}(x,t)-h)_{+}}{\left({u}(x,t)-h\right)_{+}}}}^{q-1}$; on integrating by parts as in the proof of Lemma \[l:cacc\] we obtain $$\label{eq:cacc2_i}
\begin{split}
&\frac{1}{q}
\sum_{x\in V}
\zeta(x)
{{\ifthenelse{\equal{*}{*}}{({u}(x,t)-h)_{+}}{\left({u}(x,t)-h\right)_{+}}}}^{q}
{d_{{\omega}}}(x)
\eta(t)
\\
&\quad+
\int_{0}^{t}
\sum_{x,y\in V}
{\lvert{D}_{y}{u}(x,\tau)\rvert}^{p-2}
{D}_{y}{u}(x,\tau)
\zeta(y)
{D}_{y}{{\ifthenelse{\equal{*}{*}}{({u}(x,\tau)-h)_{+}}{\left({u}(x,\tau)-h\right)_{+}}}}^{q-1}
{\omega}(x,y)
\eta(\tau)
{\,\textup{\textmd{d}}}\tau
\\
&\quad+
\int_{0}^{t}
\sum_{x,y\in V}
{\lvert{D}_{y}{u}(x,\tau)\rvert}^{p-2}
{D}_{y}{u}(x,\tau)
{D}_{y}\zeta(x)
{{\ifthenelse{\equal{*}{*}}{({u}(x,\tau)-h)_{+}}{\left({u}(x,\tau)-h\right)_{+}}}}^{q-1}
{\omega}(x,y)
\eta(\tau)
{\,\textup{\textmd{d}}}\tau
\\
&\qquad=
\frac{1}{q}
\int_{0}^{t}
\sum_{x\in V}
\zeta(x)
{{\ifthenelse{\equal{*}{*}}{({u}(x,\tau)-h)_{+}}{\left({u}(x,\tau)-h\right)_{+}}}}^{q}
{d_{{\omega}}}(x)
\eta'(\tau)
{\,\textup{\textmd{d}}}\tau
\,.
\end{split}$$ We estimate next the second integral in . The absolute value of the integrand is bounded from above by $$\begin{gathered}
\frac{1}{R_{2}-R_{1}}
\sum_{x,y\in B_{R_{2}+1}}
{\lvert{D}_{y}{u}(x,\tau)\rvert}^{p-1}
{{\ifthenelse{\equal{*}{*}}{({u}(x,\tau)-h)_{+}}{\left({u}(x,\tau)-h\right)_{+}}}}^{q-1}
{\omega}(x,y)
\\
\le
\frac{1}{R_{2}-R_{1}}
\sum_{x,y\in B_{R_{2}+1}}
\big(
{u}(x,\tau)^{p+q-2}
+
{u}(y,\tau)^{p-1}
{u}(x,\tau)^{q-1}
\big)
{\omega}(x,y)
\le
\frac{C_{u}}{R_{2}-R_{1}}
\,,
\end{gathered}$$ where $C_{u}$ is independent of $R_{i}$. Owing to $p+q-2>q$ and to Remark \[r:lp\_scale\], to this end it is only left to observe that $$\begin{gathered}
\sum_{x,y\in B_{R_{2}+1}}
{u}(y,\tau)^{p-1}
{u}(x,\tau)^{q-1}
{\omega}(x,y)
\\
\le
\Big(
\sum_{y\in V}
{u}(y,\tau)^{(p-1)q}
{d_{{\omega}}}(y)
\Big)^{\frac{1}{q}}
\Big(
\sum_{x\in V}
{u}(x,\tau)^{q}
{d_{{\omega}}}(x)
\Big)^{\frac{q-1}{q}}
\,,
\end{gathered}$$ and to use once more Remark \[r:lp\_scale\], since $(p-1)q>q$.
The sought after estimates follows immediately upon applying Lemma \[l:monot\] with ${v}=0$ and then letting first $R_{2}\to\infty$ and then $R_{1}\to\infty$.
\[r:prelim\_diff\] Lemma \[l:cacc2\] is still in force if ${u}$ is the difference of two solutions to . The proof is the same, when we start from the difference of the two equations and recall Lemma \[l:monot\].
Existence and comparison {#s:prelim_ex}
------------------------
\[p:ex\] Let ${u}_{0}\in\ell^{q}(V)$, $q>1$. Then – has a solution in $L^{\infty}(0,+\infty;\ell^{q}(V))$. If ${u}_{0}\ge 0$ then ${u}\ge 0$.
Let ${u}_{0}\in\ell^{q}(V)$, $q>1$. Define for $n\ge 1$ ${u}_{n}$ as the solution to $$\begin{aligned}
{2}
\label{eq:exist_pde_n}
{\frac{\partial {u}_{n}}{\partial t}}(x,t)
&=
\operatorname{\Delta}_{p}{u}_{n}(x,t)
\,,
&\qquad&
x\in B_{n}
\,,
t>0
\,,
\\
\label{eq:exist_init_n}
{u}_{n}(x,0)
&=
{u}_{0}(x)
\,,
&\qquad&
x\in B_{n}
\,,
\\
\label{eq:exist_dir_n}
{u}_{n}(x,t)
&=
0
\,,
&\qquad&
x\not\in B_{n}
\,,
t\ge 0
\,.
\end{aligned}$$ In practice this is a finite system of ordinary differential equations, uniquely solvable in the class $C^{1}(0,T)$ at least as long as the solution stays bounded over $(0,T)$.
In this connection, we rewrite , as $${u}_{n}(x,t)^{q-1}
{\frac{\partial {u}_{n}}{\partial t}}(x,t)
=
{u}_{n}(x,t)^{q-1}
\operatorname{\Delta}_{p}{u}_{n}(x,t)
\,,
\qquad
x\in V
\,,
t>0
\,,$$ where we stress that the equality holds for all $x\in V$. In this Subsection we denote $s^{q-1}={\lverts\rvert}^{q-1}\textup{sign}(s)$ for all $s\in {{\boldsymbol}{R}}$. Thus, summing over $x\in V$ and integrating by parts both in $t$ and in $x$ (in the suitable sense) we see that the elliptic part of the equation yields a nonnegative contribution, so that $$\label{eq:exist_energy_n}
\sum_{x\in V}
{\lvert{u}_{n}(x,t)\rvert}^{q}
{d_{{\omega}}}(x)
\le
\sum_{x\in B_{n}}
{\lvert{u}_{0}(x)\rvert}^{q}
{d_{{\omega}}}(x)
\le
{{\lVert{u}_{0}\rVert}_{\ell^{q}(V)}}^{q}
\,.$$ In turn, as explained in Remark \[r:lp\_scale\], this implies stable sup bounds for ${u}_{n}$ which, together with the discrete character of the $p$-laplacian and with the equation , also give stable sup bounds for the time derivative $\partial {u}_{n}/\partial t$, for each fixed $x$. However $V$ is countable, so that this is enough to enable us to extract a subsequence, still denoted by ${u}_{n}$ such that $$\label{eq:exist_unif_conv}
{u}_{n}(x,t)
\to
{u}(x,t)
\,,
\quad
{\frac{\partial {u}_{n}}{\partial t}}(x,t)
\to
{\frac{\partial {u}}{\partial t}}(x,t)$$ for each $x\in V$, uniformly for $t\in [0,T]$, where we have made use of the equation again to obtain convergence for the time derivative. Finally owing to we have $$\label{eq:exist_energy_nj}
\sum_{x\in V}
{\lvert{u}(x,t)\rvert}^{q}
{d_{{\omega}}}(x)
\le
{{\lVert{u}_{0}\rVert}_{\ell^{q}(V)}}^{q}
\,,
\qquad
t>0
\,.$$ It is easily seen that ${u}\in L^{\infty}(0,+\infty;\ell^{q}(V))$ is a solution to –. If ${u}_{0}\ge 0$, we appeal to our next result to prove that ${u}\ge 0$.
\[p:compare\] If ${u}_{1}$, ${u}_{2}\in L^{\infty}(0,T; \ell^{q}(V))$ solve – with ${u}_{01}$, ${u}_{02}\in\ell^{q}(V)$, ${u}_{01}\ge {u}_{02}$, then ${u}_{1}\ge{u}_{2}$.
According to Remark \[r:lp\_scale\] and to Definition \[d:sol\], we may assume $q>1$. Define ${w}={u}_{2}-{u}_{1}$. Then ${w}$ does not solve , but we may still apply (with $h=0$) to it, see Remark \[r:prelim\_diff\]. This proves ${{\ifthenelse{\equal{*}{*}}{({w})_{+}}{\left({w}\right)_{+}}}}=0$ and thus the statement.
Elementary inequalities {#s:prelim_elem}
-----------------------
We record for future use two immediate consequences of , : $$\begin{aligned}
2
\label{eq:fkf_above}
{\Lambda_p}(sa)^{-1}
&\le
s^{\omega}
{\Lambda_p}(a)^{-1}
\,,
&\qquad&
s\ge 1
\,,
a>0
\,;
\\
\label{eq:fkf_below}
{\Lambda_p}(\sigma a)^{-1}
&\le
\sigma^{\frac{p}{{N}}}
{\Lambda_p}(a)^{-1}
\,,
&\qquad&
0<\sigma\le 1
\,,
a>0
\,.\end{aligned}$$ Also the following Lemma relies on and will be used in a context where it is important that $\nu<1/(p-1)$.
\[l:dcf\] Let $\nu={N}(p-2)/[({N}(p-2)+p)(p-1)]$ and $b>0$. Then the function $$\tau\mapsto \tau^{\nu}{\psi}_{1}^{(-1)}(\tau^{-1}b)^{\frac{p-2}{p-1}}
\,,
\qquad
\tau>0
\,,$$ is nondecreasing.
Equivalently we show that $$r\mapsto r^{-\alpha}{\psi}_{1}^{(-1)}(r)^{p-2}$$ is nonincreasing for $\alpha=\nu(p-1)$. Set $s={\psi}_{1}^{(-1)}(r)$, so that by definition of ${\psi}_{1}$ $$r^{-\alpha}{\psi}_{1}^{(-1)}(r)^{p-2}
=
s^{(1-\alpha)(p-2)}{\Lambda_p}(s^{-1})^{-\alpha}
=
[s^{-\frac{p}{{N}}}{\Lambda_p}(s^{-1})]^{-\alpha}
\,.$$ By assumption the latter quantity is indeed nonincreasing in $s$ which however is a nondecreasing function of $r$.
\[l:bbl\] Under assumption we have that if $R$, $s>0$, $c\ge 1$ and $$\label{eq:bbl_m}
R^{p}
=
cs
{\psi}_{1}^{(-1)}(s^{-1})^{p-2}
\,,$$ then $$\label{eq:bbl_mm}
{\mu_{{\omega}}}(B_{{\lfloorR\rfloor}})
\le
\gamma(c)
\psi_{1}^{(-1)}
(s^{-1})^{-1}
\,.$$
Let $\tau>0$ be such that $s^{-1}={\psi}_{1}(\tau)=\tau^{p-2}{\Lambda_p}(\tau^{-1})$. Then $$c^{-1}R^{p}
=
{\Lambda_p}(\tau^{-1})^{-1}
\,.$$ On the other hand, on setting $v={\mu_{{\omega}}}(B_{{\lfloorR\rfloor}})$ and invoking we get $$c^{-1}R^{p}
\ge
c^{-1}{\mathcal{R}}(v)^{p}
\ge
c^{-1}
\gamma_{0}
{\Lambda_p}(v)^{-1}
\ge
{\Lambda_p}((\gamma_{0}c^{-1})^{\frac{{N}}{p}}v)^{-1}
\,,$$ where we also used . Since ${\Lambda_p}$ is nonincreasing, the result follows.
Proofs of Proposition \[p:linf\_meas\] and Corollary \[co:linf\_int\] {#s:linf}
=====================================================================
By assumption, and by Remark \[r:lp\_scale\], ${u}\in L^{\infty}(0,T;\ell^{q}(V))$ for some $q> 1$; then for all $k>0$ the cut function ${{\ifthenelse{\equal{*}{*}}{({u}(t)-k)_{+}}{\left({u}(t)-k\right)_{+}}}}$ is finitely supported. For given $0<\sigma_{1}<\sigma_{2}<1/2$, $k>0$, $0<t<T$ define the decreasing sequences $$\begin{gathered}
k_{i}
=
k[
1-\sigma_{2}
+
2^{-i}
(\sigma_{2}-\sigma_{1})
]
\,,
\qquad
i=0\,,1\,,2\,,\dots
\\
t_{i}
=
\frac{t}{2}
[
1-\sigma_{2}
+
2^{-i}
(\sigma_{2}-\sigma_{1})
]
\,,
\qquad
i=0\,,1\,,2\,,\dots
\end{gathered}$$ and let $f_{i}(x,\tau)={{\ifthenelse{\equal{*}{*}}{({u}(x,\tau)-k_{i})_{+}}{\left({u}(x,\tau)-k_{i}\right)_{+}}}}^{\nu}$, where $\nu=(p+q-2)/p$. Let also $$\begin{aligned}
{m}_{i}(\tau)
&=
{\mu_{{\omega}}}(
\{
x\in V
\mid
{u}(x,\tau)>k_{i}
\}
)
\,,
\quad
{M}_{i}
=
\sup_{t_{i}<\tau<t}
{m}_{i}(\tau)
\,,
\\
{{D}}_{i}(\tau)
&=
\sum_{x,y\in V}
{\lvert{D}_{y}f_{i}(x,\tau)\rvert}^{p}
{\omega}(x,y)
\,.
\end{aligned}$$ Since $b:=q/\nu<p$, it follows from Faber-Krahn inequality and Hölder’s and Young’s inequalities that $$\begin{gathered}
\label{eq:fkb}
\sum_{x\in V}
f_{i+1}(x,\tau)^{b}
{d_{{\omega}}}(x)
\le
{m}_{i+1}(\tau)^{1-\frac{b}{p}}
{\Lambda_p}({m}_{i+1}(\tau))^{-\frac{b}{p}}
{{D}}_{i+1}(\tau)^{\frac{b}{p}}
\\
\le
{\varepsilon}^{\frac{p}{b}}
{{D}}_{i+1}(\tau)
+
{\varepsilon}^{-\frac{p}{p-b}}
{\Lambda_p}({m}_{i+1}(\tau))^{-\frac{b}{p-b}}
{m}_{i+1}(\tau)
\,.
\end{gathered}$$ Here ${\varepsilon}>0$ is arbitrary and will be selected below.
We integrate over $(t_{i+1},t)$ to find $$\label{eq:linf_i}
\begin{split}
&\int_{t_{i+1}}^{t}
\sum_{x\in V}
f_{i+1}(x,\tau)^{b}
{d_{{\omega}}}(x)
{\,\textup{\textmd{d}}}\tau
\le
{\varepsilon}^{\frac{p}{b}}
\int_{t_{i+1}}^{t}
{{D}}_{i+1}(\tau)
{\,\textup{\textmd{d}}}\tau
\\
&\qquad+
{\varepsilon}^{-\frac{p}{p-b}}
\int_{t_{i+1}}^{t}
{\Lambda_p}({m}_{i+1}(\tau))^{-\frac{b}{p-b}}
{m}_{i+1}(\tau)
{\,\textup{\textmd{d}}}\tau
\\
&\quad\le
{\varepsilon}^{\frac{p}{b}}
\int_{t_{i+1}}^{t}
{{D}}_{i+1}(\tau)
{\,\textup{\textmd{d}}}\tau
+
{\varepsilon}^{-\frac{p}{p-b}}
t
{\Lambda_p}({M}_{i+1})^{-\frac{b}{p-b}}
{M}_{i+1}
\,.
\end{split}$$ Next we invoke Lemma \[l:cacc2\] with $\tau_{1}=t_{i}$, $\tau_{2}=t_{i+1}$, $h=k_{i}$, to infer $$\label{eq:linf_ii}
\begin{split}
&
L_{i}:=
\sup_{t_{i}<\tau<t}
\sum_{x\in V}
f_{i}(x,\tau)^{b}
{d_{{\omega}}}(x)
+
\int_{t_{i}}^{t}
{{D}}_{i}(\tau)
{\,\textup{\textmd{d}}}\tau
\\
&\quad
\le
\frac{\gamma 2^{i}}{t(\sigma_{2}-\sigma_{1})}
\int_{t_{i+1}}^{t}
\sum_{x\in V}
f_{i+1}(x,\tau)^{b}
{d_{{\omega}}}(x)
{\,\textup{\textmd{d}}}\tau
\\
&\quad\le
\frac{\gamma 2^{i}}{t(\sigma_{2}-\sigma_{1})}
{\varepsilon}^{\frac{p}{b}}
\int_{t_{i+1}}^{t}
{{D}}_{i+1}(\tau)
{\,\textup{\textmd{d}}}\tau
\\
&\qquad
+
\frac{\gamma 2^{i}}{\sigma_{2}-\sigma_{1}}
{\varepsilon}^{-\frac{p}{p-b}}
{\Lambda_p}({M}_{i+1})^{-\frac{b}{p-b}}
{M}_{i+1}
\,,
\end{split}$$ where the second inequality follows of course from . For a $\delta>0$ to be chosen, select ($\gamma$ denotes here the constant in ) $$\frac{\gamma 2^{i}}{t(\sigma_{2}-\sigma_{1})}
{\varepsilon}^{\frac{p}{b}}
=
\delta
\qquad
\text{i.e.,}
\qquad
{\varepsilon}=
\gamma_{0}
\delta^{\frac{b}{p}}
t^{\frac{b}{p}}
(\sigma_{2}-\sigma_{1})^{\frac{b}{p}}
2^{-\frac{b}{p}i}
\,.$$ On substituting this choice of ${\varepsilon}$ in we arrive at an estimate which can be successfully iterated, that is $$\label{eq:linf_iii}
L_{i}
\le
\delta L_{i+1}
+
\frac{\gamma 2^{\frac{pi}{p-b}}}{(\sigma_{2}-\sigma_{1})^{\frac{p}{p-b}}}
\delta^{-\frac{b}{p-b}}
t^{-\frac{b}{p-b}}
{\Lambda_p}({M}_{\infty})^{-\frac{b}{p-b}}
{M}_{\infty}
\,.$$ Here we set $$\begin{gathered}
\label{eq:linf_ij}
t_{\infty}
=
\lim_{i\to\infty}
t_{i}
=
\frac{t}{2}(1-\sigma_{2})
\,,
\qquad
k_{\infty}
=
\lim_{i\to\infty} k_{i}
=
k(1-\sigma_{2})
\,,
\\
\label{eq:linf_ijj}
M_{\infty}
=
\sup_{t_{\infty}<\tau<t}
{\mu_{{\omega}}}(
\{
x\in V
\mid
{u}(x,\tau)>k_{\infty}
\}
)
\,.
\end{gathered}$$ On iterating we infer $$L_{0}
\le
\delta^{j}
L_{j}
+
\Big(
\sum_{i=0}^{j}
\delta^{i}
2^{\frac{pi}{p-b}}
\Big)
\frac{\gamma}{(\sigma_{2}-\sigma_{1})^{\frac{p}{p-b}}}
t^{-\frac{b}{p-b}}
{\Lambda_p}({M}_{\infty})^{-\frac{b}{p-b}}
{M}_{\infty}
\,,$$ which yields as $j\to\infty$, provided we select $\delta<2^{-p/(p-b)}$, $$\label{eq:linf_j}
\begin{split}
&
\sup_{t(1-\sigma_{1})/2<\tau<t}
\sum_{x\in V}
{{\ifthenelse{\equal{*}{*}}{({u}(x,\tau)-k(1-\sigma_{1}))_{+}}{\left({u}(x,\tau)-k(1-\sigma_{1})\right)_{+}}}}^{q}
{d_{{\omega}}}(x)
\le L_{0}
\\
&\quad
\le
\frac{\gamma}{(\sigma_{2}-\sigma_{1})^{\frac{q}{p-2}}}
t^{-\frac{q}{p-2}}
{\Lambda_p}({M}_{\infty})^{-\frac{q}{p-2}}
{M}_{\infty}
\,,
\end{split}$$ for ${M}_{\infty}$ as in , owing also to $b/(p-b)=q/(p-2)$.
The proof will be concluded by a second process of iteration, built on . Let $1/2>\sigma>0$ and $k>0$, and define the increasing sequences $$\begin{gathered}
\tau_{n}
=
\frac{t}{2}(1-\sigma 2^{-n})
\,,
\qquad
h_{n}
=
k(1-\sigma 2^{-n})
\,,
\\
\bar h_{n}
=
\frac{h_{n}+h_{n+1}}{2}
=
k(1-3\sigma 2^{-n-2})
\,,
\qquad
n\ge 0
\,,
\end{gathered}$$ as well as the decreasing one $$Y_{n}
=
\sup_{\tau_{n}<\tau<t}
{\mu_{{\omega}}}(\{x\in V\mid {u}(x,\tau)> h_{n}\})
\,.$$ Next we apply Chebychev’s inequality to find $$\label{eq:linf_jj}
Y_{n+1}
\le
2^{(n+2)q}
\sigma^{-q}
k^{-q}
\sup_{\tau_{n+1}<\tau<t}
\sum_{x\in V}
{{\ifthenelse{\equal{*}{*}}{({u}(x,\tau)-\bar h_{n})_{+}}{\left({u}(x,\tau)-\bar h_{n}\right)_{+}}}}
{d_{{\omega}}}(x)
\,.$$ The right hand side of is then majorized by appealing to , where we select $$\sigma_{1}
=
3\sigma
2^{-n-2}
\,,
\quad
\sigma_{2}
=
\sigma
2^{-n}
\,,$$ in order to obtain $$\label{eq:linf_jjj}
Y_{n+1}
\le
\gamma
\sigma^{-\frac{q(p-1)}{p-2}}
2^{\frac{n(p-2+q)}{p-2}}
t^{-\frac{q}{p-2}}
k^{-q}
{\Lambda_p}(Y_{n})^{-\frac{q}{p-2}}
Y_{n}
\,.$$ In turn, on invoking our assumption , we transform into $$\label{eq:linf_k}
Y_{n+1}
\le
\gamma
\sigma^{-\frac{q(p-1)}{p-2}}
2^{\frac{n(p-2+q)}{p-2}}
t^{-\frac{q}{p-2}}
k^{-q}
{\Lambda_p}(Y_{0})^{-\frac{q}{p-2}}
Y_{0}^{-\frac{p}{{N}}\,\frac{q}{p-2}}
Y_{n}^{1+\frac{p}{{N}}\,\frac{q}{p-2}}
\,.$$ This inequality yields $Y_{n}\to0$ as $n\to\infty$ provided we choose $k$ so that (see [@LSU Lemma 5.6 Ch. II]) $$\label{eq:linf_kk}
k^{-1}
t^{-\frac{1}{p-2}}
{\Lambda_p}(Y_{0})^{-\frac{1}{p-2}}
\le
\gamma_{0}(q,p,{N})
\,.$$ In this connection we may assume e.g., $\sigma=1/4$. The proof is concluded when we remark that $Y_{n}\to0$ immediately implies $${u}(x,t)
\le
k
\,,
\qquad
x\in V
\,.$$
\[r:cacc\] We note that the proof of Proposition \[p:linf\_meas\] makes use of the differential equation only thru inequality . This fact will be used below.
We remark on using Chebychev’s inequality once more that in $$Y_{0}
\le
2^{r}k^{-r}
\sup_{\frac{t}{4}<\tau<t}
\sum_{x\in V}
{u}(x,\tau)^{r}
{d_{{\omega}}}(x)
\,.$$ Let us set $$E_{r}
=
\sup_{0<\tau<t}
\sum_{x\in V}
{u}(x,\tau)^{r}
{d_{{\omega}}}(x)
\,.$$ Then is certainly fulfilled if $$\label{eq:linf_int_i}
k^{-1}
t^{-\frac{1}{p-2}}
{\Lambda_p}(
k^{-r}
E_{r}
)^{-\frac{1}{p-2}}
=
\gamma_{0}
\,,$$ where we also used . On the other hand, if we set $${\psi}_{r}(s)
=
s^{\frac{p-2}{r}}
{\Lambda_p}(s^{-1})
\,,
\qquad
s>0
\,,$$ then amounts to $$\label{eq:linf_int_ii}
k
=
E_{r}^{\frac{1}{r}}
\big[{\psi}_{r}^{(-1)}
\big(\gamma t^{-1}E_{r}^{-\frac{p-2}{r}}\big)
\big]^{\frac{1}{r}}
\le
\gamma
E_{r}^{\frac{1}{r}}
\big[{\psi}_{r}^{(-1)}
\big(t^{-1}E_{r}^{-\frac{p-2}{r}}\big)
\big]^{\frac{1}{r}}
\,,$$ where we have made use of the definition of ${\psi}_{r}$ and of .
Proof of Theorem \[t:l1\] {#s:l1}
=========================
Let ${u}_{0}\in\ell^{1}(V)$, ${u}_{0}\ge 0$. Then we have also ${u}_{0}\in\ell^{2}(V)$ as noted in the Introduction, and we may consider the solution ${u}\ge 0$ constructed according to Subsection \[s:prelim\_ex\]. First we bound the $\ell^{1}(V)$ norm of the solution from above. We multiply the equation against $\varTheta({u}(x,\tau))\zeta(x)$ where $\zeta$ is as in Section \[s:prelim\], $$\varTheta({u})
=
\frac{
{{\ifthenelse{\equal{*}{*}}{({u}-h)_{+}}{\left({u}-h\right)_{+}}}}
}{
{u}+{\varepsilon}}
\,,$$ for any given $h>0$, ${\varepsilon}>0$, and integrate by parts. The purpose of the cut at level $h$ is to ease technically the argument. Since $\varTheta$ is a nondecreasing function, reasoning as in the proof of Lemma \[l:cacc2\] we easily obtain $$\begin{gathered}
\sum_{x\in V}
\int_{0}^{{u}(x,t)}
\frac{
{{\ifthenelse{\equal{*}{*}}{(s-h)_{+}}{\left(s-h\right)_{+}}}}
}{
s +{\varepsilon}}
\zeta(x)
{d_{{\omega}}}(x)
\le
\sum_{x\in V}
\int_{0}^{{u}_{0}(x)}
\frac{
{{\ifthenelse{\equal{*}{*}}{(s-h)_{+}}{\left(s-h\right)_{+}}}}
}{
s +{\varepsilon}}
{d_{{\omega}}}(x)
+
K_{1}
\\
\le
\sum_{x\in V}
{u}_{0}(x)
{d_{{\omega}}}(x)
+
K_{1}
\,,\end{gathered}$$ where $$\begin{gathered}
K_{1}
=
\frac{1}{R_{2}-R_{1}}
\int_{0}^{t}
\sum_{x\in V}
{\lvert{D}_{y}{u}(x,\tau)\rvert}^{p-1}
\chi_{\{{u}(\tau)>h\}}(x)
{\omega}(x,y)
{\,\textup{\textmd{d}}}\tau
\\
\le
\frac{h^{-1}}{R_{2}-R_{1}}
\int_{0}^{t}
\sum_{x\in V}
{\lvert{D}_{y}{u}(x,\tau)\rvert}^{p-1}
{u}(x,\tau)
{\omega}(x,y)
{\,\textup{\textmd{d}}}\tau
\,.\end{gathered}$$ Then we may proceed as in the proof of Lemma \[l:cacc2\] with $q=2$ and let $R_{2}\to \infty$ and then $R_{1}\to \infty$ to make $K_{1}$ vanish. Finally we let first ${\varepsilon}\to 0$ and then $h\to 0$: on invoking the monotone convergence theorem we get $$\label{eq:l1_above}
{{\lVert{u}(t)\rVert}_{\ell^{1}(V)}}
\le
{{\lVert{u}_{0}\rVert}_{\ell^{1}(V)}}
\,.$$ Therefore from Corollary \[co:linf\_int\] and Remark \[r:dcf\] we infer that is satisfied.
In order to prove we proceed as follows. We multiply the equation against $\zeta(x)$ as above and integrate by parts obtaining $$\label{eq:l1_j}
\sum_{x\in V}
{u}(x,t)
\zeta(x)
{d_{{\omega}}}(x)
+
K_{2}
=
\sum_{x\in V}
{u}_{0}(x)
\zeta(x)
{d_{{\omega}}}(x)
\,,$$ where $$\label{eq:l1_entropy}
\begin{split}
{\lvertK_{2}\rvert}
&=
{\left|
\int_{0}^{t}
\sum_{x\in V}
{\lvert{D}_{y}{u}(x,\tau)\rvert}^{p-2}
{D}_{y}{u}(x,\tau)
{D}_{y}\zeta(x)
{\omega}(x,y)
{\,\textup{\textmd{d}}}\tau
\right|}
\\
&\le
\frac{1}{R_{2}-R_{1}}
\int_{0}^{t}
\sum_{x\in B_{R_{2}}+1}
{\lvert{D}_{y}{u}(x,\tau)\rvert}^{p-1}
{\omega}(x,y)
{\,\textup{\textmd{d}}}\tau
\\
&\le
\frac{2t}{(R_{2}-R_{1}){\Lambda_p}^{(-1)}(2)^{p-2}}
{{\lVert{u}_{0}\rVert}_{\ell^{1}(V)}}^{p-1}
\,.
\end{split}$$ Here we reasoned as in (with $q=1$), exploiting $p>2$ and the already proved bound . Then we rewrite as $${{\lVert{u}(t)\rVert}_{\ell^{1}(V)}}
+
K_{2}
\ge
\sum_{x\in B_{R_{1}}}
{u}_{0}(x)
{d_{{\omega}}}(x)
\,,$$ and let first $R_{2}\to \infty$ then $R_{1}\to\infty$ to obtain the converse to .
Finally we prove the entropy estimate . First we invoke Hölder’s inequality to bound $$\label{eq:l1_k}
\begin{split}
&I:=
\int_{0}^{t}
\sum_{x,y\in V}
{\lvert{D}_{y}{u}(x,\tau)\rvert}^{p-1}
{\omega}(x,y)
{\,\textup{\textmd{d}}}\tau
\\
&\quad\le
\Big(
\int_{0}^{t}
\sum_{x,y\in V}
\tau^{-\delta(p-1)}
({u}(x,\tau)+{u}(y,\tau))^{(2-\theta)(p-1)}
{\omega}(x,y)
{\,\textup{\textmd{d}}}\tau
\Big)^{\frac{1}{p}}
\\
&\qquad\times
\Big(
\int_{0}^{t}
\sum_{x,y\in V}
\tau^{\delta}
{\lvert{D}_{y}{u}(x,\tau)\rvert}^{p}
({u}(x,\tau)+{u}(y,\tau))^{\theta-2}
{\omega}(x,y)
{\,\textup{\textmd{d}}}\tau
\Big)^{\frac{p-1}{p}}
\\
&\quad=:
K_{3}^{\frac{1}{p}}
K_{4}^{\frac{p-1}{p}}
\,.
\end{split}$$ Here $\delta>0$ is to be chosen and we select $$\theta
=
\frac{2p-3}{p-1}
\in(1,2)
\,,
\quad
\text{so that}
\quad
(2-\theta)(p-1)
=
1
\,.$$ Thus $$\label{eq:l1_kk}
K_{3}
\le
2
\int_{0}^{t}
\tau^{-\delta(p-1)}
{{\lVert{u}(\tau)\rVert}_{\ell^{1}(V)}}
{\,\textup{\textmd{d}}}\tau
\le
\gamma
{{\lVert{u}_{0}\rVert}_{\ell^{1}(V)}}
t^{1- \delta(p-1)}
\,,$$ provided $$\label{eq:l1_kkk}
\delta(p-1)<1
\,.$$ In order to bound $K_{4}$ we multiply the differential equation against $\tau^{\delta}{u}^{\theta-1}$ and integrate by parts. After dropping a positive contribution from the left hand side of the resulting equality we obtain $$\label{eq:l1_kj}
\begin{split}
K_{4}
&\le
\gamma
\int_{0}^{t}
\sum_{x\in V}
\tau^{\delta -1}
{u}(x,\tau)^{\frac{p-2}{p-1}+1}
{d_{{\omega}}}(x)
{\,\textup{\textmd{d}}}\tau
\\
&\le
\gamma
{{\lVert{u}_{0}\rVert}_{\ell^{1}(V)}}
\int_{0}^{t}
\tau^{\delta-1}
{{\lVert{u}(\tau)\rVert}_{\ell^{\infty}(V)}}^{\frac{p-2}{p-1}}
{\,\textup{\textmd{d}}}\tau
\\
&\le
\gamma
{{\lVert{u}_{0}\rVert}_{\ell^{1}(V)}}^{\frac{p-2}{p-1}+1}
\int_{0}^{t}
\tau^{\delta-1}
{\psi}_{1}^{(-1)}(\tau^{-1}{{\lVert{u}_{0}\rVert}_{\ell^{1}(V)}}^{2-p})^{\frac{p-2}{p-1}}
{\,\textup{\textmd{d}}}\tau
\,.
\end{split}$$ Select now $\nu<\delta<1/(p-1)$, where $\nu$ is the constant defined in Lemma \[l:dcf\]. Accordingly, the last integral above is bounded by $$\begin{gathered}
\label{eq:l1_kjj}
\int_{0}^{t}
\tau^{\delta-\nu-1}
\tau^{\nu}
{\psi}_{1}^{(-1)}(\tau^{-1}{{\lVert{u}_{0}\rVert}_{\ell^{1}(V)}}^{2-p})^{\frac{p-2}{p-1}}
{\,\textup{\textmd{d}}}\tau
\\
\le
t^{\nu}
{\psi}_{1}^{(-1)}(t^{-1}{{\lVert{u}_{0}\rVert}_{\ell^{1}(V)}}^{2-p})^{\frac{p-2}{p-1}}
(\delta-\nu)^{-1}
t^{\delta-\nu}
\,.\end{gathered}$$ Collecting all the estimates in –, we finally arrive at $$\label{eq:l1_kkj}
I
\le
\gamma
t^{\frac{1}{p}}
{{\lVert{u}_{0}\rVert}_{\ell^{1}(V)}}^{\frac{2(p-1)}{p}}
{\psi}_{1}^{(-1)}(t^{-1}{{\lVert{u}_{0}\rVert}_{\ell^{1}(V)}}^{2-p})^{\frac{p-2}{p}}
\,.$$
Proof of Theorem \[p:bbl\] and Corollary \[p:bbl2\] {#s:bbl}
===================================================
Let ${u}$ be as in the statement of Theorem \[p:bbl\]. For all $t>0$, $R\in{{\boldsymbol}{N}}$ we write $${{\lVert{u}(t)\rVert}_{\ell^{1}(V)}}
=
{{\lVert{u}(t)\rVert}_{\ell^{1}(B_{R})}}
+
{{\lVert{u}(t)\rVert}_{\ell^{1}(V\setminus B_{R})}}
\,.$$ Here we denote for a fixed $x_{0}\in V$ $$B_{R}
=
B_{R}(x_{0})
\,,
\quad
{\lvertx\rvert}
=
d(x,x_{0})
\,,
\quad
x\in V
\,.$$ For the sake of clarity let us denote by $\zeta_{R_{1},R_{2}}$ the cutoff function defined in Section \[s:prelim\]. Let $\rho>4R$, $R\ge 4$, $\rho$, $R\in {{\boldsymbol}{N}}$ and $\phi=1-\zeta_{R,2R}$. We use ${\lvertx\rvert}^{\alpha}\phi(x)\zeta_{\rho,2\rho}(x)$ as a testing function in , for a fixed $0<\alpha<1$. We obtain, assuming in addition that $R$ is so large as ${u}_{0}(x)=0$ for $x\not\in B_{R}$, $$\begin{gathered}
\sum_{x\in V}
{\lvertx\rvert}^{\alpha}
\phi(x)
\zeta_{\rho,2\rho}(x)
{u}(x,t)
{d_{{\omega}}}(x)
\\
=
-
\int_{0}^{t}
\sum_{x\in V}
{\lvert{D}_{y}{u}(x,\tau)\rvert}^{p-2}
{D}_{y}{u}(x,\tau)
{D}_{y}[\phi(x)\zeta_{\rho,2\rho}(x){\lvertx\rvert}^{\alpha}]
{\omega}(x,y)
{\,\textup{\textmd{d}}}\tau
\,.
\end{gathered}$$ In last integral, the term originating from ${D}_{y} \zeta_{\rho,2\rho}$ is seen to become vanishingly small as $\rho\to\infty$, since $\alpha<1$, similarly to what we did to bound $K_{2}$ in Section \[s:l1\]. Thus in the limit $\rho\to\infty$ we get $$\begin{split}
&\sum_{x\not\in B_{2R}}
{\lvertx\rvert}^{\alpha}
{u}(x,t)
{d_{{\omega}}}(x)
\\
&\quad\le
-
\int_{0}^{t}
\sum_{x\in V}
{\lvert{D}_{y}{u}(x,\tau)\rvert}^{p-2}
{D}_{y}{u}(x,\tau)
{D}_{y}[\phi(x){\lvertx\rvert}^{\alpha}]
{\omega}(x,y)
{\,\textup{\textmd{d}}}\tau
\\
&\quad=
\int_{0}^{t}
\sum_{x\in V}
{\lvert{D}_{y}{u}(x,\tau)\rvert}^{p-2}
{D}_{y}{u}(x,\tau)
{D}_{y}\zeta_{R,2R}(x)
{\lverty\rvert}^{\alpha}
{\omega}(x,y)
{\,\textup{\textmd{d}}}\tau
\\
&\qquad-
\int_{0}^{t}
\sum_{x\in V}
{\lvert{D}_{y}{u}(x,\tau)\rvert}^{p-2}
{D}_{y}{u}(x,\tau)
{D}_{y}{\lvertx\rvert}^{\alpha}
\phi(x)
{\omega}(x,y)
{\,\textup{\textmd{d}}}\tau
=:
Q_{1}+Q_{2}
\,.
\end{split}$$ Since if $x\sim y$, $x\not\in B_{R}$, $${\lvert{D}_{y}{\lvertx\rvert}^{\alpha}\rvert}
\le
\alpha
\min({\lvertx\rvert},{\lverty\rvert})^{\alpha-1}
\le
\gamma
R^{\alpha-1}
\,,$$ we have $${\lvertQ_{1}\rvert}
+
{\lvertQ_{2}\rvert}
\le
\gamma
R^{\alpha-1}
\int_{0}^{t}
\sum_{x\in V}
{\lvert{D}_{y}{u}(x,\tau)\rvert}^{p-1}
{\omega}(x,y)
{\,\textup{\textmd{d}}}\tau
\,.$$ We bound the last integral by means of , concluding as follows: $$\begin{gathered}
\label{eq:bbl_ik}
\sum_{x\not\in B_{2R}}
{u}(x,t)
{d_{{\omega}}}(x)
\le
R^{-\alpha}
\sum_{x\not\in B_{2R}}
{\lvertx\rvert}^{\alpha}
{u}(x,t)
{d_{{\omega}}}(x)
\\
\le
\gamma
R^{-1}
t^{\frac{1}{p}}
{{\lVert{u}_{0}\rVert}_{\ell^{1}(V)}}^{\frac{2(p-1)}{p}}
{\psi}_{1}^{(-1)}(t^{-1}{{\lVert{u}_{0}\rVert}_{\ell^{1}(V)}}^{-(p-2)})^{\frac{p-2}{p}}
\le
\gamma \varGamma^{-1}
{{\lVert{u}_{0}\rVert}_{\ell^{1}(V)}}
\,,
\end{gathered}$$ where we have selected $$\label{eq:bbl_jk}
R
\ge
R_{p}({u}_{0},t)
:=
\varGamma
t^{\frac{1}{p}}
{{\lVert{u}_{0}\rVert}_{\ell^{1}(V)}}^{\frac{p-2}{p}}
{\psi}_{1}^{(-1)}(t^{-1}{{\lVert{u}_{0}\rVert}_{\ell^{1}(V)}}^{-(p-2)})^{\frac{p-2}{p}}
\,,$$ for a $\varGamma>0$. This together with conservation of mass proves , upon an unessential redefinition of $R$.
In order to prove we remark that from the argument above it follows that for $R$ as in , $$\begin{gathered}
\sum_{x\in V}
{\lvertx\rvert}^{\alpha}
{u}(x,t)
{d_{{\omega}}}(x)
\le
(2R)^{\alpha}
\sum_{x\in B_{2R}}
{u}(x,t)
{d_{{\omega}}}(x)
\\
+
\sum_{x\not\in B_{2R}}
{\lvertx\rvert}^{\alpha}
{u}(x,t)
{d_{{\omega}}}(x)
\le
\gamma
R^{\alpha}
{{\lVert{u}_{0}\rVert}_{\ell^{1}(V)}}
\,,
\end{gathered}$$ where we have used conservation of mass again.
For a suitable choice of $\varGamma$, setting $R=2R_{p}$, $R_{p}$ as in , we have from $$\label{eq:bbl_ikk}
{{\lVert{u}(t)\rVert}_{\ell^{\infty}(V)}}
{\mu_{{\omega}}}(B_{R}(t))
\ge
{{\lVert{u}(t)\rVert}_{\ell^{1}(B_{R}(t))}}
\ge
\frac{1}{2}
{{\lVert{u}_{0}\rVert}_{\ell^{1}(V)}}
\,.$$ The statement in then follows, if is assumed, on invoking Lemma \[l:bbl\].
Proof of Theorem \[t:decay\] {#s:decay}
============================
We follow here ideas from [@Andreucci:1997], [@Andreucci:Cirmi:Leonardi:Tedeev:2001], [@Afanaseva:Tedeev:2004]. Let ${u}_{R}$ be the solution to with initial data $${u}_{R}(x,0)
=
{u}_{0}(x)
\chi_{B_{R}(x_{0})}(x)
\,,
\qquad
x\in V
\,.$$ Then mass conservation and with $r=1$ imply $$\label{eq:decay_i}
{{\lVert{u}_{R}(t)\rVert}_{\ell^{\infty}(V)}}
\le
\gamma
m_{R}
{\psi}_{1}^{(-1)}\big(t^{-1}m_{R}^{-(p-2)}\big)
\,,
\qquad
t>0
\,,$$ where $$m_{R}
=
\sum_{x\in B_{R}(x_{0})}
{u}_{0}(x)
{d_{{\omega}}}(x)
\,.$$ Let us also define ${w}_{R}={u}-{u}_{R}$; note that ${w}_{R}\ge 0$ by Proposition \[p:compare\]. In spite of the fact that ${w}_{R}$ does not solve we may still prove the following inequality for $h\ge 0$, $t>\tau_{1}>\tau_{2}>0$, also by appealing to Lemma \[l:monot\]: $$\begin{gathered}
\label{eq:decay_ii}
\sup_{\tau_{1}<\tau<t}
\sum_{x\in V}
{{\ifthenelse{\equal{*}{*}}{({w}_{R}(x,\tau)-h)_{+}}{\left({w}_{R}(x,\tau)-h\right)_{+}}}}^{q}
{d_{{\omega}}}(x)
+
\int_{\tau_{1}}^{t}
\sum_{x,y\in V}
{\left|{D}_{y}{{\ifthenelse{\equal{*}{*}}{({w}_{R}(x,\tau)-h)_{+}}{\left({w}_{R}(x,\tau)-h\right)_{+}}}}^{\frac{p+q-2}{p}}\right|}^{p}
{\omega}(x,y)
{\,\textup{\textmd{d}}}\tau
\\
\le
\frac{\gamma}{\tau_{1}-\tau_{2}}
\int_{\tau_{2}}^{t}
\sum_{x\in V}
{{\ifthenelse{\equal{*}{*}}{({w}_{R}(x,\tau)-h)_{+}}{\left({w}_{R}(x,\tau)-h\right)_{+}}}}^{q}
{d_{{\omega}}}(x)
{\,\textup{\textmd{d}}}\tau
\,.\end{gathered}$$ As already observed in Remark \[r:cacc\] this is enough for us to apply Proposition \[p:linf\_meas\] and thus Corollary \[co:linf\_int\] to ${w}_{R}$, and get $$\label{eq:decay_iii}
{{\lVert{w}_{R}\rVert}_{\ell^{\infty}(V)}}
\le
\gamma
E_{q}^{\frac{1}{q}}
\big[{\psi}_{q}^{(-1)}
\big(
t^{-1}
E_{q}^{-\frac{p-2}{q}}
\big)
\big]^{\frac{1}{q}}
\le
\gamma
E_{q0}^{\frac{1}{q}}
\big[{\psi}_{q}^{(-1)}
\big(
t^{-1}
E_{q0}^{-\frac{p-2}{q}}
\big)
\big]^{\frac{1}{q}}
\,,$$ where by invoking a simple variant of with $h=0$ we find $$E_{q}
:=
\sup_{0<\tau<t}
\sum_{x\in V}
{w}_{R}(x,\tau)^{q}
\le
E_{q0}
:=
\sum_{x\not\in B_{R}(x_{0})}
{u}_{0}(x)^{q}
{d_{{\omega}}}(x)
\,.$$ We use here also Remark \[r:dcf\].
Thus we have that for all $R>0$, since ${u}={u}_{R}+{w}_{R}$, $$\label{eq:decay_j}
{{\lVert{u}(t)\rVert}_{\ell^{\infty}(V)}}
\le
\gamma
\Big\{
m_{R}
{\psi}_{1}^{(-1)}\big(t^{-1}m_{R}^{-(p-2)}\big)
+
E_{q0}^{\frac{1}{q}}
\big[{\psi}_{q}^{(-1)}
\big(
t^{-1}
E_{q0}^{-\frac{p-2}{q}}
\big)
\big]^{\frac{1}{q}}
\Big\}
\,.$$ The first term on the right hand side of is increasing in $R$, while the second one is decreasing. We aim at making them equal, but this is in general impossible in the discrete setting of graphs. We instead select $R$ as any number (optimally the minimum one) such that $$\label{eq:decay_kk}
m_{R}
{\psi}_{1}^{(-1)}\big(t^{-1}m_{R}^{-(p-2)}\big)
\ge
E_{q0}^{\frac{1}{q}}
\big[{\psi}_{q}^{(-1)}
\big(
t^{-1}
E_{q0}^{-\frac{p-2}{q}}
\big)
\big]^{\frac{1}{q}}
\,.$$ Then is proved under assumption .
We need to make explicit. First, we define $$X_{1}
=
{\psi}_{1}^{(-1)}(t^{-1}m_{R}^{-(p-2)})
\,,
\qquad
X_{q}
=
{\psi}_{q}^{(-1)}(t^{-1}E_{q0}^{-\frac{p-2}{q}})
\,,$$ so that from the definition of ${\psi}_{r}$, we get $$\label{eq:decay_k}
X_{1}
=
t^{-\frac{1}{p-2}}
m_{R}^{-1}
{\Lambda_p}(X_{1}^{-1})^{-\frac{1}{p-2}}
\,,
\qquad
X_{q}^{\frac{1}{q}}
=
t^{-\frac{1}{p-2}}
E_{q0}^{-\frac{1}{q}}
{\Lambda_p}(X_{q}^{-1})^{-\frac{1}{p-2}}
\,.$$ Therefore can be written as $$\label{eq:decay_kkk}
{\Lambda_p}(X_{1}^{-1})
\le
{\Lambda_p}(X_{q}^{-1})
\,,
\quad
\text{that is}
\quad
X_{1}
\le
X_{q}
\,.$$ We apply ${\psi}_{1}$ and write the last inequality in the form $$\begin{split}
(tm_{R}^{p-2})^{-1}
&\le
{\psi}_{1}(X_{q})
=
X_{q}^{p-2}
{\Lambda_p}(X_{q}^{-1})
\\
&=
X_{q}^{\frac{(p-2)(q-1)}{q}}
{\psi}_{q}(X_{q})
=
X_{q}^{\frac{(p-2)(q-1)}{q}}
(tE_{q0}^{\frac{p-2}{q}})^{-1}
\,.
\end{split}$$ From here we immediately get, on recalling the definition of ${\psi}_{q}$, $$\frac{1}{t}
\ge
\Big[
\frac{E_{q0}}{m_{R}}
\Big]^{\frac{p-2}{q-1}}
{\Lambda_p}\Big(
{m_{R}^{\frac{q}{q-1}}}
{E_{q0}^{-\frac{1}{q-1}}}
\Big)
\,.$$ This amounts to concluding the proof.
[^1]: The first author is member of the Gruppo Nazionale per la Fisica Matematica (GNFM) of the Istituto Nazionale di Alta Matematica (INdAM)
[^2]: The second author was supported by Sapienza Grant C26V17KBT3
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We analyze the electric power load in the Czech Republic (CR) which exhibits a seasonality as well as other oscillations typical for European countries. Moreover, we detect 1/f noise property of electrical power load with extra additional peaks that allows to separate it into a deterministic and stochastic part. We then focus on the analysis of the stochastic part using improved Multi-fractal Detrended Fluctuation Analysis method (MFDFA) to investigate power load datasets with a minute resolution. Extracting the noise part of the signal by using Fourier transform allows us to apply this method to obtain the fluctuation function and to estimate the generalized Hurst exponent together with the correlated Hurst exponent, its improvement for the non-Gaussian datasets. The results exhibit a strong presence of persistent and the dataset is characterized by a non-Gaussian skewed distribution. There are also indications for the presence of the probability distribution that has heavier tail than the Gaussian distribution.'
address:
- 'Charles University in Prague, Faculty of Social Sciences, Institute of Economic Studies, Opletalova 26. CZ-11000 Prague 1, Czech Republic'
- 'Czech Technical University in Prague, Faculty of Nuclear Sciences and Physical Engineering, Department of Physics, Břehová 7, CZ-11519 Prague 1, Czech Republic'
- 'Bogolyubov Laboratory of Theoretical Physics, Joint Institute of Nuclear Research, RU-141980 Dubna, Russia'
author:
- Jiří Kracík
- Hynek Lavička
title: Fluctuation analysis of high frequency electric power load in the Czech Republic
---
MFDFA, electric power load, Hurst exponent, persistent process, 1/f noise, non-Gaussian distribution
Introduction
============
The responsibility for the safe and reliable operation is one of the basic duties of the national Transmission System Operator (TSO). The gradual liberalization of the European electricity market led to a necessity of the integration of mutually uncoordinated transmission systems. The enhancements of these transmission systems are very intensive in terms of both the time as well as capital investments and due to this the current energy networks are reaching their technical limits. That is mostly obvious in case of a massive increase of the offshore wind power plant installations located in the distant parts, hundreds of kilometers far from the end consumer. The electricity, which cannot pass through the under-dimensioned transmission lines or so called congestions, flows through the surrounding system which must accommodate these unscheduled flows. Unfortunately, the market with electricity and its mechanisms do not reflect this fact. In our work, we analyze high frequency data of electricity consumption in the Czech Republic and we also determine the degree of the uncertainty of the behavior of the consumers.
The analysis of the electricity prices and loads has been discussed by R. Weron [@key02]. He stated that the electricity loads, which are non-stationary time series, are combinations of both the trends and the periodic cycles with a random component. It is known from literature that electricity loads are correlated with the weather (e.g., the temperature, see [@key02; @Peirson1994235; @Lee2011896]) as well as with socio-economical changes and processes.
The first method (R/S method) for a non-stationary time series analysis was invented by H.E. Hurst [@key29]. Since its introduction the method has been tested on various datasets and also implemented very effectively on computer [@key08; @key16; @key17; @key20]. The method estimates the Hurst exponent of dataset that is related to the exponent of the autocorrelation function from the theory of fractional Brownian motion [@key32; @key24]. A modern alternative of the Hurst exponent estimation for series with local trends is the Detrended Fluctuation Analysis (DFA) which was introduced in [@key14; @key15] and used for economy datasets [@key03], heart rate dynamics [@key23; @key25], DNA sequences [@key26; @key15; @key14], long-time weather records [@key13], electricity prices time series [@key16; @key17] and wind speed records [@key37]. Recently, the DFA was improved to quantify the fluctuation function of datasets using different metrics [@key06; @key07]. The MFDFA is able to estimate the exponent of the autocorrelation function and also the exponent of the probability distribution function. In recent years, there has been a considerable focus on the investigation of multifractal cross-correlation between a pair of synchronized datasets [@key40].
There is a broad literature of modelling and forecasting methods of both price and/or load time series. It usually incorportes the Autoregressive Moving Average processes (ARMA), the Vector Autoregression (VAR), the Vector Error Correction (VECM), machine learning, an adaptive neuro-fuzzy network and a customers segmentation. Fixed mean, restricted variance and normally distributed error term represent basic assumptions for finding the best linear unbiased estimation, for a summary see Ref [@key42; @key45; @key46].
In this paper we study a dataset of electric power load in the Czech Republic since $2008$ till $2011$ with a one-minute time step. We focus on the properties of the fluctuation function where the first periodic part of the signal is filtered from the dataset and then the MFDFA is used. Our main aim is to determine the Hurst exponent which provides information about the autocorrelation function as well as the probability distribution. We also validate the assumptions of the normal (Gaussian) noise distribution and the short-range correlations.
The paper is organized as follows: In section 2 we describe the methodology of data processing. We first describe the Fourier filtering method and then the MFDFA. In section 3 we analyze the dataset using the methodology from section 2. Finally, in section 4, we draw the conclusions of our study.
![The electric power load in Czech Republic between January 1$^{st}$ 2008 and December 31$^{st}$ 2011 (top) and the stochastic part $P^{stoch}$ obtained by the filtration of the signal (bottom).[]{data-label="fig:Load"}](load_2008-2011)
![The power spectrum of the electric power load between 1$^{st}$ of January 2008 and 31$^{st}$ of December 2011 spanning $4$ years in total showing the 1/f property with the extra peaks. The blue points show one-year (Y), one-week (W), one-day (D) and 12 hour periods (12H). In the inset of the plot, there is the dependence of the RMSE on $\beta$ for the different exponents of filtration (parameter $\alpha$). The vertical black line shows the actual location of the parameter $\beta$ used for the filtration procedure.[]{data-label="fig:Power_spectrum"}](power_spectrum_2008-2011)
Methodology
===========
Human behavior datasets typically exhibit the oscillations with the periods related to the units of calendar [@key02; @key13] and the same applies to the electric power load. The one-year and one-week oscillations are clearly visible in Fig. \[fig:Load\] but the presence of other frequencies is not so easily observable. To obtain the information regarding the strength of the oscillations we employ the Power spectrum which is shown in Fig. \[fig:Power\_spectrum\]. It depicts additional periods with the lengths of one day and $12$ hours beside the others. Moreover, since the periods of the power loads do not follow harmonic functions, we can also observe peaks at the positions of the integer multiples of a typical trend. The reconstruction of the original load on the basis of these most significant trend components is influenced by randomness which is represented by less significant components of the Power spectrum.
Motivation
----------
In our study, we focus on the properties of the random part and we use the MFDFA which is popular among scientists [@key16; @key14; @key15; @key23; @key25; @key03; @key06; @key07; @key11] as an effective tool for extracting the properties of a long-range memory within the time series.
Since time series generally might be non-stationary, polynomial trends may still govern them. The basic idea of the DFA is to strip off the trends and use the residues for the further analysis. In the MFDFA, we are looking for typical patterns, which govern the time series manifesting a self-affine property defined by $X\left(c\cdot t\right)=c^{H}\cdot X\left(t\right)$. The generalized Hurst exponent H, determined by the method, is the measure of the long term memory in the time series and it is directly related to the non-integer fractal dimension D.
The disadvantage of this method is that the periodic trends disturb the estimation of the Hurst exponent [@key09] and therefore, before we employ the method, we have to filter out the oscillations from the signal.
We use the Fourier transform to execute the filtration. The MFDFA itself then removes the polynomial trends. The resulting signal is decomposed as
$$\begin{aligned}
P\left(t\right) & = & P^{stoch}\left(t\right)+P^{deter}\left(t\right),\label{eq:Decomposition_of_signal}\end{aligned}$$
where $P^{deter}$ describes the periodic behavior of the system, while $P^{stoch}$ stands for the random part.
We used a regression model with dummy variables indicating holidays and we perform the method described below. We observed negligible differencies for the low orders of the MFDFA but the observable differences for the higher orders of the MFDFA. However, the widths of the multifractal spectrums are negligible in both cases.
Mathematical description
------------------------
We execute our analysis in three steps. First, we perform the Fourier transform to separate the signal into the stochastic and the deterministic part by the Fourier transform. In the next step, we execute the MFDFA. Finally, in the last step, we calculate the correlated Hurst exponent, which requires shuffling of the original-time series. It is an improvement of the typically used generalized Hurst exponent, exploited in cases, where we have an assumption of the non-normally distributed time series.
### Fourier transform filtering
We deal with a signal in the discrete time-steps $P\left(t_{n}\right)$ where $t_{n}=t_{1}+n\cdot\Delta t$ and $n\in M\equiv\{1,\ldots,N\}$. Since the Discrete Fourier Transform of the signal is $\widehat{P}\left(m\right)=\frac{1}{\sqrt{N}}\sum_{n\in M}\exp\left(-\frac{2\pi\boldsymbol{i}\cdot n\cdot m}{M}\right)P\left(t_{n}\right)$, and the related Power spectrum $S\left(m\right)=\widehat{P}\left(m\right)\cdot\widehat{P}^{*}\left(m\right)$, where $x^{*}$ stands for conjugation). The Power spectrum, see Fig. \[fig:Power\_spectrum\], of the signal $P\left(t_{n}\right)$ exhibits a power law-like shape with extra peaks and each coefficient of the Fourier transform is separated into two parts according to the threshold $\beta\cdot m^{-\alpha}$:
- discrete significant coefficients in the Power spectrum for certain frequencies above the threshold forms $\vert\widehat{P^{deter}}\left(m\right)\vert$;
- coefficients below the threshold forms $\widehat{P^{stoch}}\left(m\right)$ ;
where $\alpha$ and $\beta$ are the parameters set with regard to the chosen RMSE level. We also note that if $\widehat{P^{deter}}\left(m\right)\neq0$ then we define $\arg\widehat{P^{deter}}\left(m\right)=\arg\widehat{P^{stoch}}\left(m\right)=\arg\widehat{P}\left(m\right)$. Otherwise $\arg\widehat{P^{deter}}\left(m\right)$ is not defined. The Fourier transform of the sub-signals $\widehat{P^{deter}}\left(m\right)$ and $\widehat{P^{stoch}}\left(m\right)$ then follow $\widehat{P}\left(m\right)=\widehat{P^{deter}}\left(m\right)+\widehat{P^{stoch}}\left(m\right)$, which is the Fourier transform of . \[eq:Decomposition\_of\_signal\]. By executing the the inverse Fourier transform $P\left(t_{n}\right)=\frac{1}{\sqrt{N}}\sum_{n\in M}\exp\left(\frac{2\pi\boldsymbol{i}\cdot n\cdot m}{M}\right)\widehat{P}\left(m\right)$ we obtain a deterministic part $P^{deter}\left(t_{n}\right)$ from $\widehat{P^{deter}}\left(m\right)$ . The later part $\widehat{P^{stoch}}\left(m\right)$ is transformed to $P^{stoch}\left(t\right)$.
To measure the quality of the filter we use a root mean square error, see inset of Fig. \[fig:Power\_spectrum\], defined as follows:
$$RMSE=\frac{\sqrt{N\sum_{i=1}^{N}\left(P^{deter}\left(t_{i}\right)-P\left(t_{i}\right)\right)^{2}}}{\sum_{i=1}^{N}P\left(t_{i}\right)}.\label{eq:RMSE}$$
The level of the error was determined both to decrease the RMSE and to prevent $P^{stoch}$ from incorporating a periodic function that produces the artificial behavior of the fluctuation function.
### Multi-fractal Detrended Fluctuation Analysis
We employ the Multi-fractal Detrended Fluctuation Analysis (MFDFA) for analyzing the filtered signal $P^{stoch}\left(t_{i}\right)$. The method is employed as an effective tool to avoid the artificial (see Ref. [@key22]) in the autocorrelation function or in the Power spectrum due to the oscillation of the electric power loads and the presence of the peaks in the Power spectrum, see Fig. \[fig:Power\_spectrum\].
Each element $\left\{ x_{i}\equiv P^{stoch}\left(t_{i}\right)\right\} $ of the dataset is indexed by $i\in M$. The application of the MFDFA consists of five steps:
Step 1. Integration of the dataset to produce the dataset $X_{j}=\sum_{i=1}^{j}x_{i}$. The “double” integration of the dataset $\tilde{X}_{j}=\sum_{i=1}^{j}X_{i}$ is also performed.
Step 2. Division of the dataset $X_{i}$ into $L_{s}\equiv\left\lfloor \frac{N}{t}\right\rfloor $ overlapping segments $X_{j,k}$ with length $s$ and $j\in\left\{ 1,\ldots,s\right\} $.
Step 3. Use of a standard (least-square) regression method of fixed order $M$ on each segment $X_{j,k}$ to obtain the local trend $T_{k}\left(x\right)$ in the region $x\in\left[1,s\right]$.
Step 4. Calculation of the sample variance for each of the $L_{s}$ segments of the original dataset $$V\left(k\right)\equiv\frac{1}{s}\sum_{j=1}^{s}\left(X_{j,k}-T_{k}\left(j\right)\right)^{2}.$$
Step 5. Averaging over all the segments of the original dataset to obtain the multi-fractal fluctuation function
$$F_{q}\left(s\right)\equiv\begin{cases}
\left(\frac{1}{L_{s}}\sum_{k=1}^{L_{s}}V^{\frac{q}{2}}\left(k\right)\right)^{\frac{1}{q}} & \textrm{if}\ q\neq0\\
\exp\left(\frac{1}{2\cdot L_{s}}\sum_{k=1}^{L_{s}}\ln V\left(k\right)\right) & \textrm{if}\ q=0
\end{cases}.\label{eq:Fluctuation_function}$$
In the analysis we investigate the properties of the fluctuation function $F_{q}\left(s\right)$ on the window of the size $s$ and on the measure $q$. Generally, $F_{q}\left(s\right)$ grows with increasing $s$ for all $q$ (see Fig. \[fig:MF-DFA-2\] or follow original literature [@key06; @key07; @key09; @key15; @key14; @key23; @key27]), following the power law
$$\begin{aligned}
F_{q}\left(s\right) & \sim & s^{H\left(q\right)+1}.\label{eq:Definition of Hurst exponent}\end{aligned}$$
The exponent $H\left(q\right)$ is called the Hurst exponent, see Ref. [@key29]. Generally, it is related to the long-term autocorrelation or the heavy-tailed distribution of the governing process, see Ref. [@key06; @key07]. We also note that $+1$ in Eq. \[eq:Definition of Hurst exponent\] stands due to the application of the double integration instead of the single integration of dataset, for discussion, please, see Ref. [@key06].
We exploit a fractal spectrum to analyze whether the dataset is governed by a single exponent or by a set of exponents. We define a scaling function by formula:
$$\tau\left(q\right)=q\cdot H\left(q\right)-1.\label{eq:scaling_function}$$
We define a fractal spectrum as the Legendre transform of $\tau\left(q\right)$ using the definition of a new variable $\pi=\frac{d\tau}{dq}$:
$$f\left(\pi\right)=q\cdot\pi-\tau.\label{eq:multifractal_spectrum}$$
Generally, the fractal spectrum allows to distinguish mono- and multifractal processes. The width of the fractal spectrum is defined by the formula $\Delta\pi=\max_{q\in\mathbb{{R}}}\pi-\min_{q\in\mathbb{{R}}}\pi$. The value of $\pi$ in peak of $f\left(\pi\right)$ denoted by $\pi^{max}$ represents the most frequent value of the exponent. As the width of the fractal spectrum goes wider, the number of admitted exponents increases and the monofractality shifts to the multifractality.
### Shuffling of the stochastic part of the time series
Generally, if a stochastic process generates the time series following a non-normal (non-Gaussian) distribution, the generalized Hurst exponent $H\left(q\right)$ combines the information about the autocorrelation function influenced by the properties of its probability distribution. We extract the correlation Hurst exponent $H^{cor}\left(q\right)$ that separates the generalized Hurst exponents calculated using the original time series and calculated using the shuffled one[^1].
While executing the shuffling procedure, we destroy the autocorrelations (if present) within the sample. Then we use a standard MFDFA described in previous section to the calculate shuffled fluctuation function:
$$F_{q}^{shuf}(s)=\overline{F_{q}\left(\left\{ x_{i}\right\} ^{shuf}\right)(s)},$$ where $\overline{x}$ stands for the averaging samples of shuffling and $\left\{ x_{i}\right\} ^{shuf}$ means shuffling of the time serie $x_{i}$. Finally, we estimate the generalized Hurst exponent of the shuffled time serie $H^{shuf}\left(q\right)$. As it was noted in the previous paragraph, the correlation Hurst exponent is then defined by following formula:
$$H^{cor}(q)=H(q)-H^{shuf}(q).\label{eq:CorrelatedHurstexponent}$$
Analogically to the generalized Hurst exponent $H\left(q\right)$ we can define the correlation fractal spectrum $f^{cor}\left(\alpha\right)$ related to $H^{cor}\left(q\right)$ by the formulas \[eq:scaling\_function\] and \[eq:multifractal\_spectrum\].
Implementation of the method
----------------------------
We used a multi-threaded implementation of the MFDFA with Zarja library [@key01] [^2] which can effectively run on multi-core cluster computers. We also compared the results with the implementations used in [@key27; @key26; @key23]. The filtration of dataset was executed in the Python using the NumPy and SciPy modules [@key30; @Key31].
Analysis of dataset
===================
![The quantile diagram of the probability density function generated from $P^{stoch}$ and its counterpart generated from it by the regression model with dummy variables $P^{stoch,dummy}$ and their comparison with the normal distribution with the same mean $\mu$ and variance $\sigma^{2}$. In the insets, we show the comparison of the probability density functions of $P^{stoch}$ with the appropriate normal distribution (the normal plot at the bottom, the log-normal scale at the top).[]{data-label="fig:Quantile diagram"}](quantile_diagram_2008-2011)
![The fluctuation function $F_{q}\left(s\right)$ of the signal $P^{stoch}$ obtained using the MFDFA of the order $4$ for various $q$s. We present $q\in\left\{ -10,-5,-2,0,2,5,10\right\} $ from the bottom to the top, respectively. Each plot is multiplied by factor $10$ from its predecessor. []{data-label="fig:MF-DFA-2"}](DFA_comparison_2008-2011)
![The correlation Hurst exponent $H^{cor}\left(q\right)$ estimated using the MFDFA of orders $2-6$. In the inset, we show the generalized Hurst exponent $H\left(q\right)$ for the same MFDFA orders. We used the dataset obtained by the regression model with dummy variables indicating the holidays. The dataset without use of the method follows the analogous pattern.[]{data-label="fig:Hurst exponent"}](hurst_exponent_2008-2011)
![The multifractal spectrum $f\left(\pi\right)$ of $P^{stoch}$ for various orders of the MFDFA method and initial detreding with parameter $\alpha$ is shown on the top. In the middle we present the correlation and shuffled multifractal spectrum, respectively.[]{data-label="fig:multifractal_spectrum"}](multifractal_spectra_2008-2011)
![The generalized Hurst exponents (top) and the multifractal spectrums (middle, bottom) of the surrogate datasets which underwent phase randomization. The middle figure show the multifractal spectrums for the surogate dataset and bottom one illustrates dependence on the subset of the dataset (each color shows the different subset). The top and middle figure plots dependence on the order of the method. The bottom subplot is for order $4$ of the method.[]{data-label="fig:surrogate_multifractal_spectra"}](surrogate_multifractal_spectra_2008-2011)
Data description
----------------
Our dataset describes the electric power load of the Czech Republic which is monitored by national Transmission System Operator (TSO), ČEPS a.s. It was calculated with high frequency from the stored data using the formula:
$$P\left(t\right)=\sum_{i\in1}^{M}T\left(t,i\right)-E\left(t\right)+I\left(t\right)+P_{u}\left(t\right),\label{eq:Electric power load}$$
where $T\left(t,i\right)$ stands for $i$-th turbo-generator output of the total number $M$. The turbo-generators are directly measured from their minimal value of $100\ kW$ of installed capacity. $E\left(t\right)$ and $I\left(t\right)$ are the exports and imports, respectively. Generally, they are a kind of bottlenecks because there are only few direct transmission lines between the Czech Republic and the neighboring countries. Finally, $P_{u}\left(t\right)$ stands for the balance of the pumped-storage hydroelectricity[^3].
The dataset is calculated in real time from various sources and the datalinks are not generally completely error-proof. Each datapoint is thus accompanied with the confidence flag indicating the credibility of the source. Some datapoints are calculated, using Eq. \[eq:Electric power load\], others are interpolated.
Our dataset consists of $N=2,103,840$ datapoints and it spans $4$ years since $2008$ till $2011$ with a one-minute time step. In our analysis, we neglect the confidence flag and we use the electric power load measured in $\mathrm{MW}$ only.
Results of Fourier filtering
----------------------------
The electric power load dataset of the Czech Republic is depicted in Fig. \[fig:Load\], where the Power spectrum $S\left(\omega\right)$ exhibits the power law with extra significant peaks, see Fig. \[fig:Power\_spectrum\], and therefore we first execute the Fourier filtering of the dataset where we assume $\vert\widehat{P^{stoch}}\left(m\right)\vert=\beta\cdot m^{-\alpha}$ with parameters $\alpha$ and $\beta$ yet to be determined. In our study we mainly choose $\text{\ensuremath{\alpha}}=0.7$ as an approximation of the best fit of this exponent and in order to the prove robustness of the method, we also plot the $RMSE$ for the two other values close to the chosen value of the parameter $\alpha$, see inset of Fig. \[fig:Power\_spectrum\]. The extensive test of the dependence of the fractal spectra on the exponent $\alpha$ is shown in Fig. \[fig:multifractal\_spectrum\].
Then we construct the dependence of the $RMSE$ on the parameter $\alpha$ and we choose the break-point of this dependence as an $\alpha$ value. The $RMSE$ is defined by Eq. \[eq:RMSE\] and at $\beta=7\cdot10^{8}\ \mathrm{MW}$ (we note that it is the equivalent of $S\left(m\right)\sim m^{-2\alpha}$). The $P^{stoch}$ does not exhibit large periodic fluctuations (see the bottom part of Fig. \[fig:Load\]) and the quantile diagram as well as the probability density distribution around the mean behave close to the normal (Gaussian) distribution (Fig. \[fig:Quantile diagram\]). We note that the choice of $\beta=2\cdot10^{9}\ \mathrm{MW}$ leads to both the significant deviation from normal distribution in its center part as well as to the increase of the periodicity in the stochastic part. The filtered signal is shown at the bottom of Fig. \[fig:Load\] and it is then more analyzed.
Results of application of MFDFA
-------------------------------
Firstly, we investigate the probability distribution function of the time series $P^{stoch}$ despite of the fact that there can still be temporary trends, see Fig. \[fig:Quantile diagram\]. The comparison of the quantile diagram, the mean and the variance of $P^{stoch}$ with quantiles of the normal (Gaussian) distribution is presented in Fig. \[fig:Quantile diagram\]. It clearly shows the deviations for the small values of the power load. In the lower right inset in Fig. \[fig:Quantile diagram\], the comparison of the histogram of $P^{stoch}$ with the appropriate normal distribution exhibits a good approximation about the average. In the upper left inset in Fig. \[fig:Quantile diagram\], we can observe the deviations of the small values of the power load from the normal distribution in the semi-logarithmic scale.
In the next step, we perform the MFDFA to calculate the fluctuation function $F_{q}\left(s\right)$ and we estimate the generalized Hurst exponent $H\left(q\right)$, see the inset of Fig. \[fig:Hurst exponent\] in range $[2\cdot10^{3},2\cdot10^{5}]$. The generalized Hurst exponent depends on $q$ we expect presence of multifractality. To get valuable information about the autocorrelation function, we shuffle the dataset to calculate the fluctuation function $F_{q}^{shuf}\left(s\right)$. The ratio of the original fluctuation function $F_{q}\left(s\right)$ against the fluctuation function of the shuffled dataset $F_{q}^{shuf}\left(s\right)$ formulated as $F_{q}^{cor}\left(s\right)=\frac{F_{q}\left(s\right)}{F_{q}^{shuf}\left(s\right)}$ follows the power law similarily as $F_{q}\left(s\right)$ see Fig. \[fig:MF-DFA-2\]. Then the calculation of the correlation Hurst exponent $H_{q}^{cor}\left(s\right)$ is performed using the formula \[eq:CorrelatedHurstexponent\]. We show $H_{q}^{cor}\left(s\right)$ in Fig. \[fig:Hurst exponent\] and the exponent stands between the values of $0.55$ till $0.8$ (in contradiction to the calculation of the generalized Hurst exponent based on the normally distributed time series), showing a strong persistence. Additionally we note that the estimation of the Hurst exponents is stable with regard to the orders of the MFDFA.
In Fig. \[fig:multifractal\_spectrum\], the fractal spectrum $f\left(\pi\right)$, the correlation fractal spectrum $f^{cor}\left(\pi\right)$ and also the shuffled fractal spectrum $f^{shuf}\left(\pi\right)$ of the stochastic part $P^{stoch}$ are not concentrated at single $\pi$ but they are broadly spread among the wide range of $\pi$s. conclude that the processes are multifractal in the distribution as well as in the correlation function. However, multifractality of the correlation function is stronger $\Delta\pi^{cor}\cong0.3$ in contrast to the multifractality of the distribution function $\Delta\pi^{shuf}=0.15$ for the same order of the method.
Tests of stability of the results
---------------------------------
The above mentioned results of the analysis may depend on additional factors. To address the factors we execute additional tests to show the invariance of the conclusions.
### Stability of results with respect to the filter
As a test of the stability of the results, we performed multiple calculations of MFDFA for different values of the parameters $\alpha$ and $\beta$. The generalized Hurst exponent as well as the fractal spectrum depend on a particular value of $\alpha$ and it is independent on the order of the method, see Fig. \[fig:multifractal\_spectrum\]. The change of the order does not significantly imply the change of the width of the fractal spectra. On the other hand, the shuffled fractal spectrum is independent on the value of $\alpha$ and it is localized around $\frac{1}{2}$ – the value of the Gaussian distribution. The persistence of the time series is conserved in the proximity of $\alpha=0.7$, see the middle of the Fig. \[fig:multifractal\_spectrum\].
### Surrogate data test
Generally, there are usually two reasons of the multifractality in time series:
- long range correlations of small and large fluctuations within the time serie;
- heavy-tailed probability distribution function (not necessarily the Lévy $\alpha$-stable distribution, see Ref. [@key55]).
The long-range correlation property and the fat-tailed probability distribution are investigated by shuffling and by a phase randomization. Shuffling destroys the correlations within the time series but it preserves the probability distribution. On the other hand, the phase randomization preserves the correlation function but weakens both the non-Gaussian and non-linear properties of the time serie. The procedures were firstly proposed in Ref. [@key50] and a review of its use can be found in Ref. [@key51]. We note that this method was initially used in the context of the MFDFA in Ref. [@key52].
We practically performed the test on $50$ samples of the surrogate datasets and we present the results in Fig. \[fig:surrogate\_multifractal\_spectra\]. In the graph in the top we can see similar results of $H$ as in the inset in the Fig. \[fig:Hurst exponent\]. In the middle graphs there is the result comparable with the top graphs in the Fig. \[fig:multifractal\_spectrum\]. We obtained the width of the fractal spectra $\Delta\pi\cong0.3$ and the location of the maximum is around $\pi^{max}\cong1.2$. We conclude that the multifractality is not caused by non-linearity and beside that there are the indications of the presence of a distribution with the tail heavier than the Gaussian distribution possess. From theory of the stable distributions and the stochastic processes, Refs. [@key55; @key60], the Gaussian distribution possess $H\left(2\right)=\frac{1}{2}$ and the Lévy $\alpha$-stable distribution $H\left(2\right)=\frac{1}{\omega}$ where $\omega$ is the exponent of the tail (for the Gaussian distribution we have $\omega=2$). We obtained for the shuffled multifractal spectra, where shuffling erases the autocorrelations with in the time series, see the bottom of Fig. \[fig:multifractal\_spectrum\], wide peak around $\pi^{shuf}=\frac{1}{2}$. Based on the assumption that the probability distribution is stable we admit presence the Lévy $\alpha$-stable distribution with the exponents $\omega$ close to the values of the Gaussian distribution. We also note that the result is independent of the set up of the initial filtering method. Additionally the Lévy $\alpha$-stable distribution must be skewed due to indications in Fig. . We also tested the influence of using a regression model with dummy variables for the decrease of the effect of holidays. As you can see on Fig. \[fig:Quantile diagram\], the result is not significant.
### Problems of stationarity and deficient random generators
We applied the Augmented Durbin-Watson test on the $P_{stoch}$ and we rejected the null hypothesis of non-stationarity at the $5\%$ significance level.
As a test of the stability of the results we separated the original dataset into $8$ sets with equal size and we executed the proposed method for each segment. The results of the method are at the bottom of Fig. \[fig:surrogate\_multifractal\_spectra\] where the curves representing the surrogate fractal spectra show the overlap with the width of fractal spectra $\Delta\pi^{sur}\cong0.3$ and $\pi^{max}\cong1.3$. These values are approximately equal to the results of the complete dataset. Thus, we conclude, that the results of the method are stable with the respect to the change of the scale.
Conclusions and Outlook
=======================
The main contribution of this paper is an analysis of the high-frequency electric power loads dataset of the Czech Republic using the improved MFDFA methodology. We discovered that the power spectrum of the signal exhibits 1/f noise property with the additional peaks that are caused by a periodic behavior of the electricity consumption. Based on that fact, we first separated the noise from modulating signal and then we applied the MFDFA without dealing with an artificial behavior of the fluctuation function, see Ref. [@key09]. After that we exploited the MFDFA for the analysis of the dataset to obtain information about the autocorrelation function. The major part of the power load is governed by oscillations. Beside that we report a strong persistence of the power loads where the distribution function exhibits non-Gaussian properties. The fractal spectra of both the distribution as well as the autocorrelation function indicate the presence of multifractality. We also performed a test using surrogate datasets as well as a test of stationarity to validate the strength of our conclusions. The analysis suggests the presence of the probability distribution with the tails heavier than the Gaussian distribution.
Some of our results are in contradiction with the previously published work analyzing electricity consumption and also with the assumptions of electricity load prediction models [@NowickaZagrajek20021903; @key02; @Pappas20081353; @Pappas2010256]. First, our analysis indicates that the stochastic part of the signal is not normally distributed, second, the distribution function is skewed and it may even have infinite moments of the probability distribution and third, the autocorrelation function is persistent. We also conclude that the estimations of risks based on traditional forecasting methods using the Gaussian distribution and short-range correlations are not usable due to both the long-range autocorrelation and the probability distribution’s extremes. The main part of the load constituting approximately $95\%$ of the signal was filtered out and it is systematically driven by external factors. Modeling by means of a regime-switching model makes a good sense to us.
The Czech transmission system is sufficiently dimensioned to cope with electricity consumption fluctuations contained in the dataset we had at our disposal. The problem that attracts actual attention of the TSO is dealing with the unexpected flows from north to south of Europe through the Czech Republic, see Ref. [@key28], which are caused by inhomogeneity of sources generating electricity and consumption of electricity in Europe. The presented approach might also be applied to solve a more complex problem, where in addition to the uncertainty of the electricity consumption, we may also consider the uncertainty caused by real power inflows and outflows (imported and exported electricity) or the uncertainty due to differences between cross-border trading and real electricity flows (obeying Kirchhoff’s laws). The level of uncertainty is expressed as a deviation from foreseeable behavior described by polynomial trends and periodic oscillations.
Author contributions {#author-contributions .unnumbered}
====================
J.K. obtained and prepared the dataset. H.L. prepared the tool for analysis and performed the analysis. J.K. and H.L. contributed to the writing of the manuscript. The work described in this paper will be used in J.K.’s Ph.D. thesis.
Acknowledgement {#acknowledgement .unnumbered}
===============
This article was supported by Czech Ministry of Education RVO68407700 and it also was written with the support of SVV project Strengthenning Doctoral Research in Economics and Finance. We thank for the fruitful discussion to P. Jizba, J. Lavička, A.M. Povolotsky, V.B. Priezzhev, E. Lutz, T. Kiss, G. Alber and H.E.Stanley.
Parameters and symbols of the methodology {#parameters-and-symbols-of-the-methodology .unnumbered}
=========================================
Value Symbol Unit
--------------------------------------- --------------------------- ----------------
Electric load $P$ $\mathrm{MW}$
Stochastic part of electric load $P^{stoch}$ $\mathrm{MW}$
Window size $s$ $\mathrm{min}$
Multifractal measure (parameter) $q$ $1$
Multifractal fluctuation function $F_{q}\left(s\right)$ $\mathrm{MW}$
Generalized Hurst exponent $H\left(q\right)$ $1$
Hurst exponent $H\equiv H\left(2\right)$ $1$
Correlation Hurst exponent $H^{cor}\left(q\right)$ $1$
Hurst exponent of shuffled time serie $H^{shuf}\left(q\right)$ $1$
Scaling exponent $\tau\left(q\right)$ $1$
[10]{} url \#1[`#1`]{}urlprefixhref \#1\#2[\#2]{} \#1[\#1]{}
R. Weron, Modeling and Forecasting Electricity Loads and Prices: A Statistical Approach, John Wiley and Sons Ltd, 2006.
J. Peirson, A. Henley, Electricity load and temperature: Issues in dynamic specification, Energy Economics 16 (4) (1994) 235 – 243.
C.-C. Lee, Y.-B. Chiu, Electricity demand elasticities and temperature: Evidence from panel smooth transition regression with instrumental variable approach, Energy Economics 33 (5) (2011) 896 – 902.
H. E. Hurst, Long term storage capacity of reservoirs, Trans. Am. Soc. Civ. Eng. 116 (770).
T. Preis, P. Virnau, W. Paul, J. J. Schneider, Accelerated fluctuation analysis by graphic cards and complex pattern formation in financial markets, New J. Phys. 11 (2009) 093024.
R. Weron, A. Przybylowicz, Hurst analysis of electricity price dynamics, Physica A 283.
R. Weron, Energy price risk management, Physica A 285 (2000) 127.
T.Preis, W. Paul, J. J. Schneider, Fluctuation patterns in high-frequency financial asset returns, Europhys. Lett. 82 (2008) 68005.
B. B. Mandelbrot, J. W. [Van Ness]{}, Fractional brownian motions, fractional noises and applications, SIAM Review 10 (4) (1968) 422–437.
H. A.Makse, S. Havlin, M. Schwartz, H. E. Stanley, Method for generating long-range correlations for large systems, Phys. Rev. E 53 (1996) 5445–5449.
C.-K. Peng, S. V. Buldyrev, S. Havlin, M. Simons, H. E. Stanley, A. L. Goldberger, Mosaic organization of [DNA]{} nucleotides, Phys. Rev. E 49 (1994) 1685–1689.
C.-K. Peng, S. V. Buldyrev, A. L. Goldberger, S. Havlin, F. Sciortino, M. Simons, H. E. Stanley, Long-range correlations in nucleotide sequence, Nature 356 (1992) 168–170.
R. Mantegna, H. Stanley, An introduction to econophysics: correlations and complexity in finance, Cambridge University Press, New York, NY, USA, 2000.
C.-K. Peng, S. Havlin, H. E. Stanley, A. L.Goldberger, Quantification of scaling exponents and crossover phenomena in nonstationary heartbeat time series, Chaos 5 (1995) 82–87.
J. W. Kantelhardt, Y. Askenazy, P. C. Ivanov, A. Bunde, S. Havlin, T. Penzel, J.-H. Peter, H. E. Stanley, Characterization of sleep stages by correlations in the magnitude and sign of heartbeat increments, Phys. Rev. E 051908.
A. L. Goldberger, L. A. N. Amaral, L. Glass, J. M. Hausdorff, P. C. Ivanov, R. G. Mark, J. E. Mietus, G. B. Moody, C.-K. Peng, H. E. Stanley, [PhysioToolkit]{}, and [PhysioNet]{}: Components of a new research resource for complex physiologic signals, Circulation 101 (2000) e215–e220.
P. Talkner, R. O. Weber, Power spectrum and detrended fluctaution analysis: Application to daily temperatures, Phys. Rev. E 64 (2000) 150–160.
K. Koçak, Examination of persistence properties of wind speed records using detrended fluctuation analysis, Energy 34 (2009) 1980–1985.
J. Kantelhardt, S. Zschiegner, E. Koscielny-Bunde, S. Havlin, A. Bunde, H. Stanley, Multifractal detrended fluctuation analysis of nonstationary time series, Physica A 316 (2002) 87.
J. Kantelhardt, E. Koscielny-Bunde, H. Rego, S. Havlin, A. Bunde, Detecting long-range correlations with detrended fluctuation analysis, Physica A 295 (2001) 441.
D. Horvatic, H. Staley, B. Podobnik, Detrended cross-correlation analysis for non-stationary time series with periodic trends, Europhys. Lett. 94 (18007).
S. Chan, K. Tsui, H. Wu, Y. Hou, Y.-C. Wu, F. Wu, Load/price forecasting and managing demand response for smart grids: Methodologies and challenges, IEEE Signal Processing Magazine (2012) 68 – 85.
T. Hong, P. Pinson, S. Fan, Global energy forecasting competition 2012, Int. J. Forecasting 30 (2) (2014) 357–363.
R. Weron, Electricity price forecasting: A review of the state-of-the-art with a look into the future, Int. J. Forecasting 30 (2014) 1030–1081.
A. Bunde, S. Havlin, J. W. Kanterhardt, T. Penzel, J. H. Peter, K. Voigt, Correlated and uncorrelated regions in heart-rate fluctuations during sleep, Phys. Rev. Lett. 85 (2000) 3736–3739.
K. Hu, P. C. Ivanov, Z. Chen, P. Carpena, , H. E. Stanley, Effect of trends on detrended fluctuation analysis, Phys. Rev. E 64 (2001) 011114.
S. Akselrod, D. Gordon, F. A.Ubel, D. C. Shannon, A. C.Barger, R. J.Cohen, Science 213 (1981) 220.
C.-K. Peng, J. Mietus, J. M. Hausdorff, S. Havlin, H. E. Stanley, A. L. Goldberger, Long-range anticorrelations and non-gaussian behavior of the heartbeat, Phys. Rev. Lett 70 (9) (1993) 1343–1346.
H. Lavička, Simulations of Agents on Social Network, LAP Lambert Academic Publishing, 2010.
H. P. Langtangen, A Primer on Scientific Programming with Python, Springer, 2009.
I. Idris, NumPy 1.5 Beginner’s Guid, Packt Publishing, 2011.
K.-I. Sato, Lévy Processes and Infinitely Divisible Distributions, Cambridge University Press, 1999.
J. Theiler, S. Eubank, A. Longtin, B. Galdrikian, J. Farmer, Testing for nonnonlinear in time series: the method of surrogate data, Physica D 58 (1992) 77–94.
T. Schreiber, A. Schmitz, Surrogate time series, Physica D 142 (2000) 346–382.
P. Norouzzadeh, W. Dullaert, B. Rahmani, Anti-correlation and multifractal features of spain electricity spot market, Physica A 380 (2007) 333–342.
G. Samorodnitsky, M. S. Taqqu, Stable Non-Gaussian Random Processes: Stochastic Models with Infinite Variance., Chapman and Hall, New York, 1994.
J. Nowicka-Zagrajek, R. Weron, Modeling electricity loads in [California]{}: [ARMA]{} models with hyperbolic noise, Signal Processing 82 (12) (2002) 1903 – 1915.
S. S. Pappas, L. Ekonomou, D. C. Karamousantas, G. E. Chatzarakis, S. K. Katsikas, P. Liatsis, Electricity demand loads modeling using autoregressive moving average [(ARMA)]{} models, Energy 33 (9) (2008) 1353 – 1360.
S. S. Pappas, L. Ekonomou, P. Karampelas, D. C. Karamousantas, S. K. Katsikas, G. E. Chatzarakis, P. D. Skafidas, Electricity demand load forecasting of the hellenic power system using an [ARMA]{} model, Electric Power Systems Research 80 (3) (2010) 256 – 264.
Z. Boldi[š]{}, Czech electricity grid challenged by [German]{} wind, Europhys. News 44 (4) (2013) 16–18. [](http://dx.doi.org/http://dx.doi.org/10.1051/epn/2013401).
[^1]: To shuffle the dataset we utilized Fisher-Yates algorithm that is effective even in the case of large dataset. In our case, we used the average of 100 samples of shuffling.
[^2]: Source code can be found at <http://zarja.sourceforge.net>.
[^3]: There are three of them - Dlouhé stráně $600\ MW$, Dale¨ice $450\ MW$ and ¦těchovice with $48\ MW$ of installed capacity.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We explore from a theoretical point of view pump and probe (P&P) analysis for fission of metal cluster where probe pulses are generalized to allow for scanning various frequencies. We show that it is possible to measure the time the system needs to develop to scission. This is achieved by a proper choice of both delay and frequency of the probe pulse. A more detailed analysis even allows to access the various intermediate stages of the fission process.'
author:
- 'P. M. Dinh$^1$, P.-G. Reinhard$^2$, E. Suraud$^1,^2$'
title: Time resolved fission in metal clusters
---
Pump and probe (P&P) experiments have become a widely used, versatile and powerful tool for time resolved studies of dynamical reactions in all sorts of electronic physics from molecules [@Zew94; @Gar95] to bulk [@Hoe97]. Experimental and theoretical investigations of P&P analysis for clusters as intermediate size systems have started more recently and are mostly still on an exploratory level [@Har98b; @Lei99; @Hei00], see also [@Rei03a] for an overview. Clusters are here a great challenge because they are much more complex than molecules and there is thus a much larger variety of possible scenarios. This calls for theoretical explorations and there exist already several of them, see [@Rei03a]. In the spirit of the studies in [@And02; @And03a], we propose to exploit the Mie plasmon resonance as a handle to track the time evolution of the cluster shape. A most dramatic evolution of shapes is experienced in cluster fission [@Bre94b; @Yan02], see also [@Nae97] for a review. The various shapes along the fission path are related to characteristic resonance spectra [@Rei97b] which provides a unique opportunity for P&P analysis of the associated time scales. As we shall see below a standard P&P scenario, augmented by a frequency scan, indeed allows to access a fully time resolved analysis of fission dynamics. This is an interesting task as such and it carries farther reaching information. The time the system needs to attain scission (namely to actually split into two pieces) provides an indirect but crucial clue to the viscosity of clusters [@Fro97a]. Fission is also a well known property in another class of finite fermion systems, namely atomic nuclei. It is actually by studying fission dynamics in atomic nuclei that the first experimental estimate of nuclear viscosity could be attained [@Dur00].
The dipole plasmon plays a key role in the dynamics of metal clusters, both in the linear and non-linear domains [@Kre93; @Rei03a]. It will again provide here the handle for the P&P analysis. The theoretical modeling needs to account properly for both, electronic and ionic degrees of freedom. This is achieved by the Time Dependent Local Density Approximation (TDLDA) for electrons, coupled to Molecular Dynamics (MD) for ions, the coupling being achieved by a local pseudo-potential [@Kue99]. Absorbing boundary conditions are used throughout, which allows a proper description of ionization. We use a cylindrically averaged description of the electrons to simplify the extensive calculations [@Cal00]. Details on the TDLDA-MD approach can be found in [@Cal00; @Rei03a].
The laser field is described in the dipole approximation as a time-dependent external potential $V_{\rm las} =
{\bf E}_0\!\cdot\!\hat{\bf d} f_{\rm las}(t) \cos(\omega t)$, with the dipole operator $\hat{\bf d}$, the laser polarization and amplitude ${\bf E}_0$, and $f_{\rm las}(t)$ the temporal profile of the laser pulse, here chosen as a sin$^2$ profile. Pulses are short (FWHM typically less than 50 fs) as required for a P&P scenario. Laser intensities are kept moderate and tuned to provide the proper pump excitation and the proper probe analysis in terms of ionization. Pump as well as probe pulses have the same polarization but different frequencies and FWHM (see below).
Both electronic and ionic observables are recorded as a function of time. The two key observables for the P&P scenario are ionization and the electronic dipole moment with respect to the ionic center of mass, which provides the plasmon response by Fourier transformation into the frequency domain, [@Cal00; @Yab96]. The ionization is computed as a function of time, by recording the number of electrons lost through the absorbing boundaries ($N_{\rm esc}$). Ions are treated as classical particles by standard MD propagation of positions and momenta. The P&P analysis will thus provide a way, with optical techniques, to access this ionic dynamics experimentally.
We consider the cluster Na$_{14}$ initially excited by a pump laser to a 3$^+$ charge state. The laser is linearly polarized along the symmetry axis of the cluster. The ${\rm Na_{14}}^{+++}$ is unstable with respect to fission [@Nae97] and thus starts to elongate until it breaks up. With the pump laser pulse used here (intensity: $5\times 10^{10}$ W/cm$^2$, frequency: 2.3 eV, FWHM: 36 fs), the average fission path leads to $\rm {Na_{14}}^{+++} \longrightarrow
{Na_8}^+ + {Na_6}^{++}$. A word of caution is in place here: TDLDA-MD provides an average description and does not give access to all the possible fission channels. Still, even with that limitation TDLDA-MD suffices for a first exploration of P&P analysis of cluster fission. The fission scenario which we have in mind allows a selective identification of either the fissioning system or the fission fragments in terms of optical response. It is indeed typical that the plasmon frequencies of the fragments lie safely above the plasmon frequency of the elongated fissioning cluster. The key mechanism which makes the analysis possible is the fact that the ionization yield is strongly correlated to the plasmon resonance: The closer the frequency of the probe laser to the plasmon frequency, the larger the ionization and [*vice versa*]{} [@Ull98a]. The plasmon response, in turn, is usually concentrated in a well defined frequency domain and provides a direct fingerprint of the underlying ionic structure. Scanning the probe ionization thus provides a strong clue to the ionic structure and its time evolution. The position (and structure) of the plasmon resonance changes in a characteristic manner along a fission path [@Rei97b]: the elongation leads to a progressive red shift of the plasmon resonance mode along the fission axis. However, as soon as the system has fissioned, the now more compact fission products both lead to a significantly blue shifted response. This is sketched in figure \[fig:scheme\]. Because of the expected strong changes of the resonance landscape, the probe pulses have to hit with the proper frequency as well as the proper delay time. We will thus consider a generalization of the standard P&P scenario to scanning probe frequencies.
First let us start with a standard P&P analysis by fixing the probe frequency $\omega_{\rm probe}$ and by varying its time delay. For the present test case, the Mie frequencies of the two fission fragments are about 0.7 to 1.5 eV above that of the elongated $\rm
{Na_{14}}^{+++}$ which lies around 1.9 eV. By choosing a probe frequency at about 3 eV, we should thus be able to track down the scission time by scanning the dynamics in terms of time delay. Before scission the fragmenting system should appear more “red” while afterwards the system will appear more “blue”. Indeed the observable here is the additional ionization due to the probe laser with respect to the mere pump ionization, defined as $
\Delta N_{\rm esc} = N_{\rm esc}[{\rm probe}] - N_{\rm esc}[{\rm
pump}].
$ That quantity depends slightly on time for a given delay. We consider here an “asymptotic” value, recorded 500 fs after the probe pulse. The $\Delta N_{\rm esc}$ reflects the closeness of the probe frequency to that of the plasmon. For delays lower than the scission time, $\omega_{\rm probe}$ remains far away from the fragment plasmon and this off-resonance gives vanishing or low extra ionizations. However as we increase the probe delay steadily, a sudden increase of $\Delta
N_{\rm esc}$ should occur, corresponding to the first resonance of $\omega_{\rm probe}$ with one or both fragment plasmons. Thus the first delay at which we observe a significant ionization after probe laser gives us an estimate of the scission time. This is indeed what we obtained for $\omega_{\rm probe}=3.13$ eV, see the full curve in figure \[fig:delay\]: the obtained extra ionization, as a function of delay, is insignificant up to 1 ps and jumps around 1.25 ps.
An alternative strategy is to use a low probe frequency, close to that of the $\rm {Na_{14}}^{+++}$ compound system, and to check the disappearance of $\Delta N_{\rm esc}$ with increased delay time. The picture can be complemented by tracking also an intermediate frequency. These two scenarios are shown together with the high frequency calculations in figure \[fig:delay\]. The low frequency of 1.77 eV (long dashes) attaches to the elongated and still connected cluster while the intermediate frequency (dotted curve) attaches the final stages of pre-scission. As expected the three signals are nicely complementing each other and allow to conclude on a scission time of order 1.2 ps for the given test case.
The full richness of resonance spectra at a given stage of fission can be studied in even more detail by fixing the time delay and scanning a dense variety of probe laser frequencies. This way of analysis is illustrated in figure \[fig:freq\], which demonstrates that the plasmon peak position (see vertical lines at the maximum of each spectrum for guidance) and its fragmentation pattern change dramatically on the way to scission. In turn, a probe pulse with fixed frequency will come into resonance only during a rather well defined delay window, as was seen in figure \[fig:delay\]. A world of interesting detailed information is thus contained in figure \[fig:freq\] when scanning the probe frequency. Such multi-color P&P setups are, however, an enormous experimental challenge. The trend of the plasmon peak sketched in figure \[fig:scheme\] complies with figure \[fig:freq\]. For 500 fs delay, the spectrum is too fragmented to define unambigously a plasmon peak. This holds for the transitional interval 300–800 fs, as hinted by a dotted regime in figure \[fig:scheme\].
Thus tuning probe delay time and frequency complement each other to analyze the fission process in great detail. A summary analysis can be performed in terms of moments of the resonance spectra [@Boh79]. To that end, we compute weighted averages over the ionization spectra $
E_n =
\sqrt{{\int\textrm d\omega\,\Delta N_{\rm esc}\,\omega^n}\Big/
{\int \textrm d\omega \, \Delta N_{\rm esc}\,\omega^{n-2}}}.
%\nonumber
$ Good reference averages are $E_1$ and $E_3$ which put moderate emphasis on the lower ($E_1$) and the upper ($E_3$) side of the spectrum. A well peaked resonance response is signaled by $E_3 \simeq
E_1$. The difference between the two values indicates the width of the spectrum. The results for $E_1$ and $E_3$ are shown in the bottom panel of figure \[fig:plasmon\]. As can also be seen from figure \[fig:freq\], the spectrum is spread up to 1 ps, which corresponds to a large difference between $E_1$ and $E_3$. Above that delay time, both averages coincide. At short times, an initial red shift is observed, whereas above 1 ps, a sudden blue shift occurs up to 1.3 ps, followed by a final red shift: this panel, compared with figure \[fig:scheme\], shows that the results are once again consistent with the expected time evolution of the plasmon. Even if well defined values of $E_1$ and $E_3$ are hard to assign in the 300–800 fs delay range, one can see that the average values oscillate with a period of a few hundred fs which shrinks somewhat with time going on. We expect that the mean plasmon frequency depends on the ionic structure. Thus we check the time evolution of the transverse and the total mean square radii, respectively $\langle {R_\perp}^2 \rangle$ and $\langle R^2 \rangle$, see middle and top panels of figure \[fig:plasmon\]. Note that after 1.2 ps, the radii are calculated for the two separate fission fragments relative to each separate center of mass. Before scission time, $\langle R^2 \rangle$ shows a breathing although partially masked by the global extension of the cluster along its symmetry axis. The transversal radius $\langle
{R_\perp}^2 \rangle$ also oscillates but is not exactly in phase with $\langle R^2 \rangle$. Both radii present periods that shrink with increasing time. We note a relatively good correspondence between the minima of $\omega_{\rm pl}$ and the maxima of the radii, and [*vice versa*]{} for the maxima of $\omega_{\rm pl}$ and the minima of the radii. The oscillations of the plasmon frequency are not dominated by either the transverse ionic motion or the longitudinal one, but seem to be a mixture of both. After scission, the plasmon evolution is clearly related to the radial oscillations since the extrema of $\omega_{\rm pl}$ and $\langle {R_\perp}^2 \rangle$ perfectly coincide. Thus this “measurement” of the plasmon frequency through the resonant ionization of the cluster gives access to the time evolution of the ionic structure, and to some extent to the potential energy surface along the fission path.
In this paper we have presented a scenario for pump and probe (P&P) analysis of fission of metal clusters. The scheme exploits the marked response of the Mie surface plasmon to laser fields and the fact that the plasmon resonance spectrum changes dramatically along the fission path. The response is quantified in a measurable way in terms of ionization induced by the laser pulse. With a single-frequency P&P setup, one can measure the scission time (that is, the final break up into fragments) when properly tuning the frequency. Scanning the frequency of the probe pulses would allow to resolve more details of the ionic geometries along the fission path. Finally, we ought to mention that a proper orientation of the cluster is required during the time span of the measurement time, [*i.e.*]{} for a few ps. This can be achieved by preparing preliminary a sample of clusters of same orientation, for instance, with the laser burning technique [@Wen99; @Sei00]; and by cooling the beam to a temperature of about 10 K, so that the orientation is conserved for at least 2 ps.
Acknowledgments. The authors thank the French-German exchange program PROCOPE, the Alexander von Humboldt foundation, the French ministry of education, the CNRS program “Matériaux” and the Institut Universitaire de France for financial support during the realization of this work.
[10]{}
B.M. Garraway and K.-A. Suominen. , 58:365, 1995.
A.H. Zewail. . World Scientific, Singapore, 1994.
U. H[ö]{}fer, et al. , 277:1480, 1997.
M. Hartmann, et al. , 108:3096, 1998.
R. Heinicke and J. Grotemeyer. , 71:419, 2000.
T. Leisner, S. Vajda, S. Wolf, and L. Wöste. , 111:1017, 1999.
P.-G. Reinhard and E. Suraud. . Wiley, New York, 2003.
K. Andrae, P.-G. Reinhard, and E. Suraud. , 35:4203, 2002.
K. Andrae, P.-G. Reinhard, and E. Suraud. , 92:173402, 2004.
C. Bréchignac, et al. , 72:1636, 1994.
C. Yannouleas, et al. , 89:173403, 2002.
U. Näher, et al. , 285:245, 1997.
P.-G. Reinhard, F. Calvayrac, and E. Suraud. , 41:151, 1997.
P. Froebrich. , 56:6450, 1997.
D. Durand, E. Suraud, and B. Tamain. . Institute of Physics, London, 2000.
U. Kreibig and M. Vollmer. , volume 25. Springer Series in Materials Science, 1993.
S. Kümmel, M. Brack, and P.-G. Reinhard. , 9:149, 1999.
F. Calvayrac, P.-G. Reinhard, E. Suraud, and C.A. Ullrich. , 337:493, 2000.
K. Yabana and G.F. Bertsch. , 54:4484, 1996.
C.A. Ullrich, P.-G. Reinhard, and E. Suraud. , 57:1938, 1998.
O. Bohigas, A.M. Lane, and J. Martorell. , 51:267, 1979.
T. Wenzel, J. Bosbach, A. Goldmann, and F. Träger. , 69:513, 1999.
G. Seifert, M. Kaempfe, K.-J. Berg, and H. Graener. , 71:795, 2000.
| {
"pile_set_name": "ArXiv"
} |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.